diff --git a/libvpx/Android.mk b/libvpx/Android.mk
new file mode 100644
index 0000000..d0fe10a
--- /dev/null
+++ b/libvpx/Android.mk
@@ -0,0 +1,9 @@
+LOCAL_PATH := $(call my-dir)
+include $(CLEAR_VARS)
+
+# libvpx
+# if ARMv7 + NEON etc blah blah
+include external/libvpx/libvpx.mk
+
+# libwebm
+include external/libvpx/libwebm.mk
diff --git a/libvpx/MODULE_LICENSE_BSD b/libvpx/MODULE_LICENSE_BSD
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/libvpx/MODULE_LICENSE_BSD
diff --git a/libvpx/NOTICE b/libvpx/NOTICE
new file mode 100644
index 0000000..1ce4434
--- /dev/null
+++ b/libvpx/NOTICE
@@ -0,0 +1,31 @@
+Copyright (c) 2010, The WebM Project authors. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+  * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+
+  * Redistributions in binary form must reproduce the above copyright
+    notice, this list of conditions and the following disclaimer in
+    the documentation and/or other materials provided with the
+    distribution.
+
+  * Neither the name of Google, nor the WebM Project, nor the names
+    of its contributors may be used to endorse or promote products
+    derived from this software without specific prior written
+    permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
diff --git a/libvpx/README.android b/libvpx/README.android
new file mode 100644
index 0000000..176c159
--- /dev/null
+++ b/libvpx/README.android
@@ -0,0 +1,56 @@
+Name: libvpx
+URL: http://www.webmproject.org
+Version: v1.6.0
+License: BSD
+License File: libvpx/LICENSE
+
+Date: Wednesday July 20 2016
+Branch: khakicampbell
+Commit: 042572177b5c58404836fce3fc221fd077dcd896
+
+Description:
+Contains the sources used to compile libvpx.
+
+The libvpx source is from webmproject.org:
+  https://chromium.googlesource.com/webm/libvpx
+
+Notes on updating libvpx source code:
+
+Please follow these steps to update libvpx source code:
+
+1. Update libvpx source tree. Look for "Current HEAD: <hash>" output to update
+   README.android. Look for "git log from upstream: <git log>" output to add
+   to the commit message.
+
+   ./update_libvpx.sh [branch name]
+
+2. Generate updated .gypi and config files.
+
+   ./generate_config.sh
+
+3. Update this file with any Version, Date, Branch, or Commit changes. The
+   version is in the file source/config/vpx_version.h
+
+4. Commit the changes. The commit message should look like this and is printed
+   by update_libvpx.sh:
+   libvpx: Pull from upstream
+
+   Current HEAD: <hash>
+
+   git log from upstream:
+   a6b2070 <git commit message 1>
+   08dabbc <git commit message 2>
+   c29fb02 <git commit message 3>
+
+Tools needed to build libvpx:
+
+- generate_config.sh
+
+Generate config files that contain the source list for each platform.
+Configuration for the build is taken from vpx_config.h for each platform.
+
+- lint_config.sh
+
+A tool to verify vpx_config.h and vpx_config.asm are matched. This also
+prints the final configuration after checking.
+
diff --git a/libvpx/README.version b/libvpx/README.version
new file mode 100644
index 0000000..bb8471d
--- /dev/null
+++ b/libvpx/README.version
@@ -0,0 +1,4 @@
+URL: https://storage.googleapis.com/downloads.webmproject.org/releases/webm/libvpx-1.6.0.tar.bz2
+Version: 1.6.0
+BugComponent: 42195
+Owners: johannkoenig
diff --git a/libvpx/config.arm.mk b/libvpx/config.arm.mk
new file mode 100644
index 0000000..6ae58b8
--- /dev/null
+++ b/libvpx/config.arm.mk
@@ -0,0 +1,24 @@
+# Output variables:
+# libvpx_config_dir_arm
+# libvpx_codec_srcs_c_arm
+# libvpx_codec_srcs_asm_arm
+
+ifeq ($(ARCH_ARM_HAVE_NEON),true)
+libvpx_target := config/arm-neon
+else ifeq ($(ARCH_ARM_HAVE_ARMV7A),true)
+libvpx_target := config/arm
+else
+libvpx_target := config/generic
+endif
+
+LOCAL_ARM_MODE := arm
+
+libvpx_config_dir_arm := $(LOCAL_PATH)/$(libvpx_target)
+libvpx_codec_srcs := $(sort $(shell cat $(libvpx_config_dir_arm)/libvpx_srcs.txt))
+
+# vpx_config.c is an auto-generated file in $(libvpx_target).
+libvpx_codec_srcs_c_arm := $(addprefix libvpx/, $(filter-out vpx_config.c, \
+    $(filter %.c, $(libvpx_codec_srcs)))) \
+    $(libvpx_target)/vpx_config.c
+
+libvpx_codec_srcs_asm_arm := $(filter %.asm, $(libvpx_codec_srcs))
diff --git a/libvpx/config.arm64.mk b/libvpx/config.arm64.mk
new file mode 100644
index 0000000..0740fd9
--- /dev/null
+++ b/libvpx/config.arm64.mk
@@ -0,0 +1,18 @@
+# Output variables:
+# libvpx_config_dir_arm64
+# libvpx_codec_srcs_c_arm64
+# libvpx_codec_srcs_asm_arm64
+
+libvpx_target := config/arm64
+
+LOCAL_ARM_MODE := arm
+
+libvpx_config_dir_arm64 := $(LOCAL_PATH)/$(libvpx_target)
+libvpx_codec_srcs := $(sort $(shell cat $(libvpx_config_dir_arm64)/libvpx_srcs.txt))
+
+# vpx_config.c is an auto-generated file in $(libvpx_target).
+libvpx_codec_srcs_c_arm64 := $(addprefix libvpx/, $(filter-out vpx_config.c, \
+    $(filter %.c, $(libvpx_codec_srcs)))) \
+    $(libvpx_target)/vpx_config.c
+
+libvpx_codec_srcs_asm_arm64 := $(filter %.asm.s, $(libvpx_codec_srcs))
diff --git a/libvpx/config.mips.mk b/libvpx/config.mips.mk
new file mode 100644
index 0000000..80b61e3
--- /dev/null
+++ b/libvpx/config.mips.mk
@@ -0,0 +1,25 @@
+# Output variables:
+# libvpx_config_dir_mips
+# libvpx_codec_srcs_c_mips
+
+ifneq ($(ARCH_HAS_BIGENDIAN),true)
+  ifeq ($(ARCH_MIPS_DSP_REV),2)
+    libvpx_target := config/mips32-dspr2
+  else
+    ifeq ($(ARCH_MIPS_HAS_MSA),true)
+      libvpx_target := config/mips32-msa
+    else
+      libvpx_target := config/mips32
+    endif
+  endif
+else
+  libvpx_target := config/generic
+endif
+
+libvpx_config_dir_mips := $(LOCAL_PATH)/$(libvpx_target)
+libvpx_codec_srcs := $(sort $(shell cat $(libvpx_config_dir_mips)/libvpx_srcs.txt))
+
+# vpx_config.c is an auto-generated file in $(libvpx_target).
+libvpx_codec_srcs_c_mips := $(addprefix libvpx/, $(filter-out vpx_config.c, \
+    $(filter %.c, $(libvpx_codec_srcs)))) \
+    $(libvpx_target)/vpx_config.c
diff --git a/libvpx/config.mips64.mk b/libvpx/config.mips64.mk
new file mode 100644
index 0000000..f258b10
--- /dev/null
+++ b/libvpx/config.mips64.mk
@@ -0,0 +1,21 @@
+# Output variables:
+# libvpx_config_dir_mips64
+# libvpx_codec_srcs_c_mips64
+
+ifneq ($(ARCH_HAS_BIGENDIAN),true)
+  ifeq ($(ARCH_MIPS_HAS_MSA),true)
+    libvpx_target := config/mips64-msa
+  else
+    libvpx_target := config/mips64
+  endif
+else
+  libvpx_target := config/generic
+endif
+
+libvpx_config_dir_mips64 := $(LOCAL_PATH)/$(libvpx_target)
+libvpx_codec_srcs := $(sort $(shell cat $(libvpx_config_dir_mips64)/libvpx_srcs.txt))
+
+# vpx_config.c is an auto-generated file in $(libvpx_target).
+libvpx_codec_srcs_c_mips64 := $(addprefix libvpx/, $(filter-out vpx_config.c, \
+    $(filter %.c, $(libvpx_codec_srcs)))) \
+    $(libvpx_target)/vpx_config.c
diff --git a/libvpx/config.x86.mk b/libvpx/config.x86.mk
new file mode 100644
index 0000000..3ebb313
--- /dev/null
+++ b/libvpx/config.x86.mk
@@ -0,0 +1,16 @@
+# Output variables:
+# libvpx_config_dir_x86
+# libvpx_codec_srcs_c_x86
+
+libvpx_target := config/x86
+
+libvpx_config_dir_x86 := $(LOCAL_PATH)/$(libvpx_target)
+libvpx_codec_srcs := $(sort $(shell cat $(libvpx_config_dir_x86)/libvpx_srcs.txt))
+
+# vpx_config.c is an auto-generated file in $(libvpx_target).
+libvpx_codec_srcs_c_x86 := $(addprefix libvpx/, $(filter-out vpx_config.c, \
+    $(filter %.c, $(libvpx_codec_srcs)))) \
+    $(libvpx_target)/vpx_config.c
+
+# X86 asm files are processed by the system and sent to yasm
+libvpx_codec_srcs_c_x86 += $(addprefix libvpx/, $(filter %.asm, $(libvpx_codec_srcs)))
diff --git a/libvpx/config.x86_64.mk b/libvpx/config.x86_64.mk
new file mode 100644
index 0000000..fb9ceda
--- /dev/null
+++ b/libvpx/config.x86_64.mk
@@ -0,0 +1,17 @@
+# Output variables:
+# libvpx_config_dir_x86_64
+# libvpx_codec_srcs_c_x86_64
+# libvpx_codec_srcs_asm_x86_64
+
+libvpx_target := config/generic
+
+libvpx_config_dir_x86_64 := $(LOCAL_PATH)/$(libvpx_target)
+libvpx_codec_srcs := $(sort $(shell cat $(libvpx_config_dir_x86_64)/libvpx_srcs.txt))
+
+# vpx_config.c is an auto-generated file in $(libvpx_target).
+libvpx_codec_srcs_c_x86_64 := $(addprefix libvpx/, $(filter-out vpx_config.c, \
+    $(filter %.c, $(libvpx_codec_srcs)))) \
+    $(libvpx_target)/vpx_config.c
+
+# X86_64 asm files are processed by the system and sent to yasm
+libvpx_codec_srcs_c_x86_64 += $(addprefix libvpx/, $(filter %.asm, $(libvpx_codec_srcs)))
diff --git a/libvpx/config/arm-neon/libvpx_srcs.txt b/libvpx/config/arm-neon/libvpx_srcs.txt
new file mode 100644
index 0000000..8a85a83
--- /dev/null
+++ b/libvpx/config/arm-neon/libvpx_srcs.txt
@@ -0,0 +1,420 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/arm/armv6/bilinearfilter_v6.asm
+vp8/common/arm/armv6/copymem16x16_v6.asm
+vp8/common/arm/armv6/copymem8x4_v6.asm
+vp8/common/arm/armv6/copymem8x8_v6.asm
+vp8/common/arm/armv6/dc_only_idct_add_v6.asm
+vp8/common/arm/armv6/dequant_idct_v6.asm
+vp8/common/arm/armv6/dequantize_v6.asm
+vp8/common/arm/armv6/filter_v6.asm
+vp8/common/arm/armv6/idct_blk_v6.c
+vp8/common/arm/armv6/idct_v6.asm
+vp8/common/arm/armv6/iwalsh_v6.asm
+vp8/common/arm/armv6/loopfilter_v6.asm
+vp8/common/arm/armv6/simpleloopfilter_v6.asm
+vp8/common/arm/armv6/sixtappredict8x4_v6.asm
+vp8/common/arm/bilinearfilter_arm.c
+vp8/common/arm/bilinearfilter_arm.h
+vp8/common/arm/dequantize_arm.c
+vp8/common/arm/filter_arm.c
+vp8/common/arm/loopfilter_arm.c
+vp8/common/arm/neon/bilinearpredict_neon.c
+vp8/common/arm/neon/copymem_neon.c
+vp8/common/arm/neon/dc_only_idct_add_neon.c
+vp8/common/arm/neon/dequant_idct_neon.c
+vp8/common/arm/neon/dequantizeb_neon.c
+vp8/common/arm/neon/idct_blk_neon.c
+vp8/common/arm/neon/idct_dequant_0_2x_neon.c
+vp8/common/arm/neon/idct_dequant_full_2x_neon.c
+vp8/common/arm/neon/iwalsh_neon.c
+vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.c
+vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.c
+vp8/common/arm/neon/mbloopfilter_neon.c
+vp8/common/arm/neon/shortidct4x4llm_neon.c
+vp8/common/arm/neon/sixtappredict_neon.c
+vp8/common/arm/neon/vp8_loopfilter_neon.c
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/arm/armv6/vp8_short_fdct4x4_armv6.asm
+vp8/encoder/arm/armv6/walsh_v6.asm
+vp8/encoder/arm/dct_arm.c
+vp8/encoder/arm/neon/denoising_neon.c
+vp8/encoder/arm/neon/fastquantizeb_neon.c
+vp8/encoder/arm/neon/shortfdct_neon.c
+vp8/encoder/arm/neon/vp8_shortwalsh4x4_neon.c
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8cx_arm.mk
+vp8/vp8dx.mk
+vp9/common/arm/neon/vp9_iht4x4_add_neon.c
+vp9/common/arm/neon/vp9_iht8x8_add_neon.c
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/arm/neon/vp9_dct_neon.c
+vp9/encoder/arm/neon/vp9_error_neon.c
+vp9/encoder/arm/neon/vp9_quantize_neon.c
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/arm/avg_neon.c
+vpx_dsp/arm/bilinear_filter_media.asm
+vpx_dsp/arm/fwd_txfm_neon.c
+vpx_dsp/arm/hadamard_neon.c
+vpx_dsp/arm/idct16x16_1_add_neon.asm
+vpx_dsp/arm/idct16x16_add_neon.asm
+vpx_dsp/arm/idct16x16_neon.c
+vpx_dsp/arm/idct32x32_1_add_neon.asm
+vpx_dsp/arm/idct32x32_add_neon.asm
+vpx_dsp/arm/idct4x4_1_add_neon.asm
+vpx_dsp/arm/idct4x4_add_neon.asm
+vpx_dsp/arm/idct8x8_1_add_neon.asm
+vpx_dsp/arm/idct8x8_add_neon.asm
+vpx_dsp/arm/intrapred_neon.c
+vpx_dsp/arm/intrapred_neon_asm.asm
+vpx_dsp/arm/loopfilter_16_neon.asm
+vpx_dsp/arm/loopfilter_4_neon.asm
+vpx_dsp/arm/loopfilter_8_neon.asm
+vpx_dsp/arm/loopfilter_mb_neon.asm
+vpx_dsp/arm/loopfilter_neon.c
+vpx_dsp/arm/sad4d_neon.c
+vpx_dsp/arm/sad_media.asm
+vpx_dsp/arm/sad_neon.c
+vpx_dsp/arm/save_reg_neon.asm
+vpx_dsp/arm/subpel_variance_media.c
+vpx_dsp/arm/subpel_variance_neon.c
+vpx_dsp/arm/subtract_neon.c
+vpx_dsp/arm/variance_halfpixvar16x16_h_media.asm
+vpx_dsp/arm/variance_halfpixvar16x16_hv_media.asm
+vpx_dsp/arm/variance_halfpixvar16x16_v_media.asm
+vpx_dsp/arm/variance_media.asm
+vpx_dsp/arm/variance_neon.c
+vpx_dsp/arm/vpx_convolve8_avg_neon_asm.asm
+vpx_dsp/arm/vpx_convolve8_neon_asm.asm
+vpx_dsp/arm/vpx_convolve_avg_neon_asm.asm
+vpx_dsp/arm/vpx_convolve_copy_neon_asm.asm
+vpx_dsp/arm/vpx_convolve_neon.c
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/arm.h
+vpx_ports/arm_cpudetect.c
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/arm-neon/vp8_rtcd.h b/libvpx/config/arm-neon/vp8_rtcd.h
new file mode 100644
index 0000000..6fd2dac
--- /dev/null
+++ b/libvpx/config/arm-neon/vp8_rtcd.h
@@ -0,0 +1,230 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_neon
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict4x4_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_armv6
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x4_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x4_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_neon
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_neon
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_c
+
+void vp8_clear_system_state_c();
+#define vp8_clear_system_state vp8_clear_system_state_c
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_v6(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_neon(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_neon
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x4_v6(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x4_neon(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_neon
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x8_v6(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x8_neon(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_neon
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+void vp8_dc_only_idct_add_v6(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+void vp8_dc_only_idct_add_neon(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_neon
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_neon
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_uv_neon(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_neon
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+void vp8_dequant_idct_add_v6(short *input, short *dq, unsigned char *output, int stride);
+void vp8_dequant_idct_add_neon(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_neon
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_v6(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_neon(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_neon
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_v6(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_neon(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_neon
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+void vp8_dequantize_b_v6(struct blockd*, short *dqc);
+void vp8_dequantize_b_neon(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_neon
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sad_c
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+void vp8_fast_quantize_b_neon(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_neon
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sad_c
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_armv6(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_neon(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_neon
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_armv6(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_neon(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_neon
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_armv6(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_neon(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_neon
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_armv6(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_neon(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_neon
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_armv6(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_neon(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_neon
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_armv6(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_neon(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_neon
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_horizontal_edge_armv6(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_mbhs_neon(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_mbhs_neon
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_vertical_edge_armv6(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_mbvs_neon(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_mbvs_neon
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_c
+
+int vp8_mbuverror_c(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_c
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sad_c
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_c
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_armv6(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_neon(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_neon
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_armv6(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_neon(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_neon
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+void vp8_short_idct4x4llm_v6_dual(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+void vp8_short_idct4x4llm_neon(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_neon
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+void vp8_short_inv_walsh4x4_v6(short *input, short *output);
+void vp8_short_inv_walsh4x4_neon(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_neon
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+void vp8_short_walsh4x4_armv6(short *input, short *output, int pitch);
+void vp8_short_walsh4x4_neon(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_neon
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_neon
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict4x4_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_armv6
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_neon
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_neon
+
+void vp8_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm-neon/vp9_rtcd.h b/libvpx/config/arm-neon/vp9_rtcd.h
new file mode 100644
index 0000000..cf2f106
--- /dev/null
+++ b/libvpx/config/arm-neon/vp9_rtcd.h
@@ -0,0 +1,104 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_c
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+int64_t vp9_block_error_fp_neon(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_neon
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_fdct8x8_quant_neon(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_neon
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_c
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_c
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_c
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sad_c
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_c
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_c
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_neon
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_neon
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_neon
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_c
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
+
+void vp9_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm-neon/vpx_config.asm b/libvpx/config/arm-neon/vpx_config.asm
new file mode 100644
index 0000000..8f663f6
--- /dev/null
+++ b/libvpx/config/arm-neon/vpx_config.asm
@@ -0,0 +1,86 @@
+@ This file was created from a .asm file
+@  using the ads2gas.pl script.
+	.equ DO1STROUNDING, 0
+.equ ARCH_ARM ,  1
+.equ ARCH_MIPS ,  0
+.equ ARCH_X86 ,  0
+.equ ARCH_X86_64 ,  0
+.equ HAVE_EDSP ,  0
+.equ HAVE_MEDIA ,  1
+.equ HAVE_NEON ,  1
+.equ HAVE_NEON_ASM ,  1
+.equ HAVE_MIPS32 ,  0
+.equ HAVE_DSPR2 ,  0
+.equ HAVE_MSA ,  0
+.equ HAVE_MIPS64 ,  0
+.equ HAVE_MMX ,  0
+.equ HAVE_SSE ,  0
+.equ HAVE_SSE2 ,  0
+.equ HAVE_SSE3 ,  0
+.equ HAVE_SSSE3 ,  0
+.equ HAVE_SSE4_1 ,  0
+.equ HAVE_AVX ,  0
+.equ HAVE_AVX2 ,  0
+.equ HAVE_VPX_PORTS ,  1
+.equ HAVE_PTHREAD_H ,  1
+.equ HAVE_UNISTD_H ,  1
+.equ CONFIG_DEPENDENCY_TRACKING ,  1
+.equ CONFIG_EXTERNAL_BUILD ,  1
+.equ CONFIG_INSTALL_DOCS ,  0
+.equ CONFIG_INSTALL_BINS ,  1
+.equ CONFIG_INSTALL_LIBS ,  1
+.equ CONFIG_INSTALL_SRCS ,  0
+.equ CONFIG_USE_X86INC ,  0
+.equ CONFIG_DEBUG ,  0
+.equ CONFIG_GPROF ,  0
+.equ CONFIG_GCOV ,  0
+.equ CONFIG_RVCT ,  0
+.equ CONFIG_GCC ,  1
+.equ CONFIG_MSVS ,  0
+.equ CONFIG_PIC ,  1
+.equ CONFIG_BIG_ENDIAN ,  0
+.equ CONFIG_CODEC_SRCS ,  0
+.equ CONFIG_DEBUG_LIBS ,  0
+.equ CONFIG_DEQUANT_TOKENS ,  0
+.equ CONFIG_DC_RECON ,  0
+.equ CONFIG_RUNTIME_CPU_DETECT ,  0
+.equ CONFIG_POSTPROC ,  0
+.equ CONFIG_VP9_POSTPROC ,  0
+.equ CONFIG_MULTITHREAD ,  1
+.equ CONFIG_INTERNAL_STATS ,  0
+.equ CONFIG_VP8_ENCODER ,  1
+.equ CONFIG_VP8_DECODER ,  1
+.equ CONFIG_VP9_ENCODER ,  1
+.equ CONFIG_VP9_DECODER ,  1
+.equ CONFIG_VP8 ,  1
+.equ CONFIG_VP9 ,  1
+.equ CONFIG_ENCODERS ,  1
+.equ CONFIG_DECODERS ,  1
+.equ CONFIG_STATIC_MSVCRT ,  0
+.equ CONFIG_SPATIAL_RESAMPLING ,  1
+.equ CONFIG_REALTIME_ONLY ,  1
+.equ CONFIG_ONTHEFLY_BITPACKING ,  0
+.equ CONFIG_ERROR_CONCEALMENT ,  0
+.equ CONFIG_SHARED ,  0
+.equ CONFIG_STATIC ,  1
+.equ CONFIG_SMALL ,  0
+.equ CONFIG_POSTPROC_VISUALIZER ,  0
+.equ CONFIG_OS_SUPPORT ,  1
+.equ CONFIG_UNIT_TESTS ,  1
+.equ CONFIG_WEBM_IO ,  1
+.equ CONFIG_LIBYUV ,  1
+.equ CONFIG_DECODE_PERF_TESTS ,  0
+.equ CONFIG_ENCODE_PERF_TESTS ,  0
+.equ CONFIG_MULTI_RES_ENCODING ,  0
+.equ CONFIG_TEMPORAL_DENOISING ,  1
+.equ CONFIG_VP9_TEMPORAL_DENOISING ,  0
+.equ CONFIG_COEFFICIENT_RANGE_CHECKING ,  0
+.equ CONFIG_VP9_HIGHBITDEPTH ,  0
+.equ CONFIG_BETTER_HW_COMPATIBILITY ,  0
+.equ CONFIG_EXPERIMENTAL ,  0
+.equ CONFIG_SIZE_LIMIT ,  0
+.equ CONFIG_SPATIAL_SVC ,  0
+.equ CONFIG_FP_MB_STATS ,  0
+.equ CONFIG_EMULATE_HARDWARE ,  0
+.equ CONFIG_MISC_FIXES ,  0
+	.section	.note.GNU-stack,"",%progbits
diff --git a/libvpx/config/arm-neon/vpx_config.c b/libvpx/config/arm-neon/vpx_config.c
new file mode 100644
index 0000000..48dd606
--- /dev/null
+++ b/libvpx/config/arm-neon/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--target=armv7-linux-gcc --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/arm-neon/vpx_config.h b/libvpx/config/arm-neon/vpx_config.h
new file mode 100644
index 0000000..b4e2097
--- /dev/null
+++ b/libvpx/config/arm-neon/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 1
+#define ARCH_MIPS 0
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 1
+#define HAVE_NEON 1
+#define HAVE_NEON_ASM 1
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 0
+#define CONFIG_DC_RECON 0
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 0
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/arm-neon/vpx_dsp_rtcd.h b/libvpx/config/arm-neon/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..637da71
--- /dev/null
+++ b/libvpx/config/arm-neon/vpx_dsp_rtcd.h
@@ -0,0 +1,854 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+unsigned int vpx_avg_4x4_neon(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_neon
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+unsigned int vpx_avg_8x8_neon(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_neon
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_neon
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_neon
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_neon
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_neon
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_neon
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_neon
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_avg_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_neon
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_copy_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_neon
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_neon
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_c
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_c
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_c
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_c
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_c
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_c
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_c
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_c
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_neon
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_c
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_neon
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_neon
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_c
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_c
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_c
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_c
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_neon
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_neon
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_neon
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_neon
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_neon
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_neon
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_neon
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_neon
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_neon
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_neon
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_neon
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_neon
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_neon
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_neon
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_neon
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_neon
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_c
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_c
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_c
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_c
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_c
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_c
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_c
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_neon(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_neon
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_1_neon(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_neon
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get16x16var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_neon
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+unsigned int vpx_get4x4sse_cs_neon(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_neon
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get8x8var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_neon
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_c
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_neon
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_neon
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_neon
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_neon
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void vpx_hadamard_16x16_neon(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_neon
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void vpx_hadamard_8x8_neon(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_neon
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_10_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_neon
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_neon
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_256_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_neon
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_neon
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_1024_add_neon
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_neon
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_1024_add_neon
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_neon
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_neon
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_12_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_neon
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_neon
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_neon
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+int16_t vpx_int_pro_col_neon(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_neon
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+void vpx_int_pro_row_neon(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_neon
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_c
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_neon
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_neon
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_neon
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_8_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_neon
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_16_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_neon
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_neon
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_neon
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_neon
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_neon
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_neon
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_neon
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_8_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_neon
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void vpx_minmax_8x8_neon(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_neon
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x16_media(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_neon
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_c
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_c
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_c
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_c
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x16_media(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_neon
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_c
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_c
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_neon
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_c
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_c
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_c
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_c
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_neon
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_c
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_c
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_c
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_c
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_c
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_c
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_c
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_neon
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_c
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_c
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_neon
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_c
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_c
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_c
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_c
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_neon
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_c
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_c
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_c
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_c
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_c
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_c
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_c
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_c
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_c
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_c
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_c
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_neon
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_c
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_c
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_neon
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_c
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_neon
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_c
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_c
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_c
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_c
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_c
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_c
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_c
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_c
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_neon
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_c
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_c
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_c
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_c
+
+int vpx_satd_c(const int16_t *coeff, int length);
+int vpx_satd_neon(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_neon
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_c
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_c
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_c
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_c
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_c
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_c
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_c
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_c
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_c
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_c
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_c
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_c
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_c
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_c
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x16_media(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_neon
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_c
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_c
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_c
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_neon
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_c
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_c
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_c
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_c
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_neon
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_c
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_c
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x8_media(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_neon
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void vpx_subtract_block_neon(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_neon
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_neon
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_neon
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_neon
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_neon
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_neon
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_neon
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_neon
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_neon
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x16_media(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_neon
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_c
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_neon
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_c
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_neon
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_neon
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_c
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_c
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_neon
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_neon
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_neon
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_c
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x8_media(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_neon
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_h_media(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_media
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_hv_media(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_media
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_v_media(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_media
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+int vpx_vector_var_neon(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_neon
+
+void vpx_dsp_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm-neon/vpx_scale_rtcd.h b/libvpx/config/arm-neon/vpx_scale_rtcd.h
new file mode 100644
index 0000000..a1564b7
--- /dev/null
+++ b/libvpx/config/arm-neon/vpx_scale_rtcd.h
@@ -0,0 +1,71 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_c
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm-neon/vpx_version.h b/libvpx/config/arm-neon/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/arm-neon/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/config/arm/libvpx_srcs.txt b/libvpx/config/arm/libvpx_srcs.txt
new file mode 100644
index 0000000..f20898b
--- /dev/null
+++ b/libvpx/config/arm/libvpx_srcs.txt
@@ -0,0 +1,366 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/arm/armv6/bilinearfilter_v6.asm
+vp8/common/arm/armv6/copymem16x16_v6.asm
+vp8/common/arm/armv6/copymem8x4_v6.asm
+vp8/common/arm/armv6/copymem8x8_v6.asm
+vp8/common/arm/armv6/dc_only_idct_add_v6.asm
+vp8/common/arm/armv6/dequant_idct_v6.asm
+vp8/common/arm/armv6/dequantize_v6.asm
+vp8/common/arm/armv6/filter_v6.asm
+vp8/common/arm/armv6/idct_blk_v6.c
+vp8/common/arm/armv6/idct_v6.asm
+vp8/common/arm/armv6/iwalsh_v6.asm
+vp8/common/arm/armv6/loopfilter_v6.asm
+vp8/common/arm/armv6/simpleloopfilter_v6.asm
+vp8/common/arm/armv6/sixtappredict8x4_v6.asm
+vp8/common/arm/bilinearfilter_arm.c
+vp8/common/arm/bilinearfilter_arm.h
+vp8/common/arm/dequantize_arm.c
+vp8/common/arm/filter_arm.c
+vp8/common/arm/loopfilter_arm.c
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/arm/armv6/vp8_short_fdct4x4_armv6.asm
+vp8/encoder/arm/armv6/walsh_v6.asm
+vp8/encoder/arm/dct_arm.c
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8cx_arm.mk
+vp8/vp8dx.mk
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/arm/bilinear_filter_media.asm
+vpx_dsp/arm/sad_media.asm
+vpx_dsp/arm/subpel_variance_media.c
+vpx_dsp/arm/variance_halfpixvar16x16_h_media.asm
+vpx_dsp/arm/variance_halfpixvar16x16_hv_media.asm
+vpx_dsp/arm/variance_halfpixvar16x16_v_media.asm
+vpx_dsp/arm/variance_media.asm
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/arm.h
+vpx_ports/arm_cpudetect.c
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/arm/vp8_rtcd.h b/libvpx/config/arm/vp8_rtcd.h
new file mode 100644
index 0000000..f7287a5
--- /dev/null
+++ b/libvpx/config/arm/vp8_rtcd.h
@@ -0,0 +1,200 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_armv6
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict4x4_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_armv6
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x4_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_armv6
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_armv6
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_c
+
+void vp8_clear_system_state_c();
+#define vp8_clear_system_state vp8_clear_system_state_c
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_v6(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_v6
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x4_v6(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_v6
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x8_v6(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_v6
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+void vp8_dc_only_idct_add_v6(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_v6
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_c
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_c
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+void vp8_dequant_idct_add_v6(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_v6
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_v6(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_v6
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_v6(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_v6
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+void vp8_dequantize_b_v6(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_v6
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sad_c
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_c
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sad_c
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_armv6(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_armv6
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_armv6(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_armv6
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_armv6(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_armv6
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_armv6(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_armv6
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_armv6(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_armv6
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_armv6(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_armv6
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_horizontal_edge_armv6(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_armv6
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_vertical_edge_armv6(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_armv6
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_c
+
+int vp8_mbuverror_c(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_c
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sad_c
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_c
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_armv6(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_armv6
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_armv6(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_armv6
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+void vp8_short_idct4x4llm_v6_dual(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_v6_dual
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+void vp8_short_inv_walsh4x4_v6(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_v6
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+void vp8_short_walsh4x4_armv6(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_armv6
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_armv6
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict4x4_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_armv6
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_armv6
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_armv6(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_armv6
+
+void vp8_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm/vp9_rtcd.h b/libvpx/config/arm/vp9_rtcd.h
new file mode 100644
index 0000000..56ca1fe
--- /dev/null
+++ b/libvpx/config/arm/vp9_rtcd.h
@@ -0,0 +1,99 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_c
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_c
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_c
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_c
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_c
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_c
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sad_c
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_c
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_c
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_c
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_c
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_c
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
+
+void vp9_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm/vpx_config.asm b/libvpx/config/arm/vpx_config.asm
new file mode 100644
index 0000000..94b1d0d
--- /dev/null
+++ b/libvpx/config/arm/vpx_config.asm
@@ -0,0 +1,86 @@
+@ This file was created from a .asm file
+@  using the ads2gas.pl script.
+	.equ DO1STROUNDING, 0
+.equ ARCH_ARM ,  1
+.equ ARCH_MIPS ,  0
+.equ ARCH_X86 ,  0
+.equ ARCH_X86_64 ,  0
+.equ HAVE_EDSP ,  0
+.equ HAVE_MEDIA ,  1
+.equ HAVE_NEON ,  0
+.equ HAVE_NEON_ASM ,  0
+.equ HAVE_MIPS32 ,  0
+.equ HAVE_DSPR2 ,  0
+.equ HAVE_MSA ,  0
+.equ HAVE_MIPS64 ,  0
+.equ HAVE_MMX ,  0
+.equ HAVE_SSE ,  0
+.equ HAVE_SSE2 ,  0
+.equ HAVE_SSE3 ,  0
+.equ HAVE_SSSE3 ,  0
+.equ HAVE_SSE4_1 ,  0
+.equ HAVE_AVX ,  0
+.equ HAVE_AVX2 ,  0
+.equ HAVE_VPX_PORTS ,  1
+.equ HAVE_PTHREAD_H ,  1
+.equ HAVE_UNISTD_H ,  1
+.equ CONFIG_DEPENDENCY_TRACKING ,  1
+.equ CONFIG_EXTERNAL_BUILD ,  1
+.equ CONFIG_INSTALL_DOCS ,  0
+.equ CONFIG_INSTALL_BINS ,  1
+.equ CONFIG_INSTALL_LIBS ,  1
+.equ CONFIG_INSTALL_SRCS ,  0
+.equ CONFIG_USE_X86INC ,  0
+.equ CONFIG_DEBUG ,  0
+.equ CONFIG_GPROF ,  0
+.equ CONFIG_GCOV ,  0
+.equ CONFIG_RVCT ,  0
+.equ CONFIG_GCC ,  1
+.equ CONFIG_MSVS ,  0
+.equ CONFIG_PIC ,  1
+.equ CONFIG_BIG_ENDIAN ,  0
+.equ CONFIG_CODEC_SRCS ,  0
+.equ CONFIG_DEBUG_LIBS ,  0
+.equ CONFIG_DEQUANT_TOKENS ,  0
+.equ CONFIG_DC_RECON ,  0
+.equ CONFIG_RUNTIME_CPU_DETECT ,  0
+.equ CONFIG_POSTPROC ,  0
+.equ CONFIG_VP9_POSTPROC ,  0
+.equ CONFIG_MULTITHREAD ,  1
+.equ CONFIG_INTERNAL_STATS ,  0
+.equ CONFIG_VP8_ENCODER ,  1
+.equ CONFIG_VP8_DECODER ,  1
+.equ CONFIG_VP9_ENCODER ,  1
+.equ CONFIG_VP9_DECODER ,  1
+.equ CONFIG_VP8 ,  1
+.equ CONFIG_VP9 ,  1
+.equ CONFIG_ENCODERS ,  1
+.equ CONFIG_DECODERS ,  1
+.equ CONFIG_STATIC_MSVCRT ,  0
+.equ CONFIG_SPATIAL_RESAMPLING ,  1
+.equ CONFIG_REALTIME_ONLY ,  1
+.equ CONFIG_ONTHEFLY_BITPACKING ,  0
+.equ CONFIG_ERROR_CONCEALMENT ,  0
+.equ CONFIG_SHARED ,  0
+.equ CONFIG_STATIC ,  1
+.equ CONFIG_SMALL ,  0
+.equ CONFIG_POSTPROC_VISUALIZER ,  0
+.equ CONFIG_OS_SUPPORT ,  1
+.equ CONFIG_UNIT_TESTS ,  1
+.equ CONFIG_WEBM_IO ,  1
+.equ CONFIG_LIBYUV ,  1
+.equ CONFIG_DECODE_PERF_TESTS ,  0
+.equ CONFIG_ENCODE_PERF_TESTS ,  0
+.equ CONFIG_MULTI_RES_ENCODING ,  0
+.equ CONFIG_TEMPORAL_DENOISING ,  1
+.equ CONFIG_VP9_TEMPORAL_DENOISING ,  0
+.equ CONFIG_COEFFICIENT_RANGE_CHECKING ,  0
+.equ CONFIG_VP9_HIGHBITDEPTH ,  0
+.equ CONFIG_BETTER_HW_COMPATIBILITY ,  0
+.equ CONFIG_EXPERIMENTAL ,  0
+.equ CONFIG_SIZE_LIMIT ,  0
+.equ CONFIG_SPATIAL_SVC ,  0
+.equ CONFIG_FP_MB_STATS ,  0
+.equ CONFIG_EMULATE_HARDWARE ,  0
+.equ CONFIG_MISC_FIXES ,  0
+	.section	.note.GNU-stack,"",%progbits
diff --git a/libvpx/config/arm/vpx_config.c b/libvpx/config/arm/vpx_config.c
new file mode 100644
index 0000000..a45c0d2
--- /dev/null
+++ b/libvpx/config/arm/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--target=armv6-linux-gcc --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/arm/vpx_config.h b/libvpx/config/arm/vpx_config.h
new file mode 100644
index 0000000..3937bf2
--- /dev/null
+++ b/libvpx/config/arm/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 1
+#define ARCH_MIPS 0
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 1
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 0
+#define CONFIG_DC_RECON 0
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 0
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/arm/vpx_dsp_rtcd.h b/libvpx/config/arm/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..d2cc47d
--- /dev/null
+++ b/libvpx/config/arm/vpx_dsp_rtcd.h
@@ -0,0 +1,752 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_c
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_c
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_c
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_c
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_c
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_c
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_c
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_c
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_c
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_c
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_c
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_c
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_c
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_c
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_c
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_c
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_c
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_c
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_c
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_c
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_c
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_c
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_c
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_c
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_c
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_c
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_c
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_c
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_c
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_c
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_c
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_c
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_c
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_c
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_c
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_c
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_c
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_c
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_c
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_c
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_c
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_c
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_c
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_c
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_c
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_c
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_c
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_c
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_c
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_c
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_c
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_c
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_c
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_c
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_c
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_c
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_c
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_c
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_c
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_c
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_c
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_c
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_c
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_c
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_c
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_135_add_c
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_c
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_34_add_c
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_c
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_c
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_c
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_c
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_c
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_c
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_c
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_c
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_c
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_c
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_c
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_c
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_c
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_c
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_c
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_c
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_c
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_c
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_c
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_c
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_c
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x16_media(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_media
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_c
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_c
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_c
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_c
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x16_media(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_media
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_c
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_c
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_c
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_c
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_c
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_c
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_c
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_c
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_c
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_c
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_c
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_c
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_c
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_c
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_c
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_c
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_c
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_c
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_c
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_c
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_c
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_c
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_c
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_c
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_c
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_c
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_c
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_c
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_c
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_c
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_c
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_c
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_c
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_c
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_c
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_c
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_c
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_c
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_c
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_c
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_c
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_c
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_c
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_c
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_c
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_c
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_c
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_c
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_c
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_c
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_c
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_c
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_c
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_c
+
+int vpx_satd_c(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_c
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_c
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_c
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_c
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_c
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_c
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_c
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_c
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_c
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_c
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_c
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_c
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_c
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_c
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_c
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x16_media(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_media
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_c
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_c
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_c
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_c
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_c
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_c
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_c
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_c
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_c
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_c
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_c
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x8_media(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_media
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_c
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_c
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_c
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_c
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_c
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_c
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_c
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_c
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_c
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x16_media(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_media
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_c
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_c
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_c
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_c
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_c
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_c
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_c
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_c
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_c
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_c
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_c
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x8_media(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_media
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_h_media(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_media
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_hv_media(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_media
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_v_media(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_media
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_c
+
+void vpx_dsp_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm/vpx_scale_rtcd.h b/libvpx/config/arm/vpx_scale_rtcd.h
new file mode 100644
index 0000000..a1564b7
--- /dev/null
+++ b/libvpx/config/arm/vpx_scale_rtcd.h
@@ -0,0 +1,71 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_c
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm/vpx_version.h b/libvpx/config/arm/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/arm/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/config/arm64/libvpx_srcs.txt b/libvpx/config/arm64/libvpx_srcs.txt
new file mode 100644
index 0000000..4178248
--- /dev/null
+++ b/libvpx/config/arm64/libvpx_srcs.txt
@@ -0,0 +1,392 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/arm/dequantize_arm.c
+vp8/common/arm/filter_arm.c
+vp8/common/arm/loopfilter_arm.c
+vp8/common/arm/neon/bilinearpredict_neon.c
+vp8/common/arm/neon/copymem_neon.c
+vp8/common/arm/neon/dc_only_idct_add_neon.c
+vp8/common/arm/neon/dequant_idct_neon.c
+vp8/common/arm/neon/dequantizeb_neon.c
+vp8/common/arm/neon/idct_blk_neon.c
+vp8/common/arm/neon/idct_dequant_0_2x_neon.c
+vp8/common/arm/neon/idct_dequant_full_2x_neon.c
+vp8/common/arm/neon/iwalsh_neon.c
+vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.c
+vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.c
+vp8/common/arm/neon/mbloopfilter_neon.c
+vp8/common/arm/neon/shortidct4x4llm_neon.c
+vp8/common/arm/neon/sixtappredict_neon.c
+vp8/common/arm/neon/vp8_loopfilter_neon.c
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/arm/dct_arm.c
+vp8/encoder/arm/neon/denoising_neon.c
+vp8/encoder/arm/neon/fastquantizeb_neon.c
+vp8/encoder/arm/neon/shortfdct_neon.c
+vp8/encoder/arm/neon/vp8_shortwalsh4x4_neon.c
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8cx_arm.mk
+vp8/vp8dx.mk
+vp9/common/arm/neon/vp9_iht4x4_add_neon.c
+vp9/common/arm/neon/vp9_iht8x8_add_neon.c
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/arm/neon/vp9_dct_neon.c
+vp9/encoder/arm/neon/vp9_error_neon.c
+vp9/encoder/arm/neon/vp9_quantize_neon.c
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/arm/avg_neon.c
+vpx_dsp/arm/fwd_txfm_neon.c
+vpx_dsp/arm/hadamard_neon.c
+vpx_dsp/arm/idct16x16_1_add_neon.c
+vpx_dsp/arm/idct16x16_add_neon.c
+vpx_dsp/arm/idct16x16_neon.c
+vpx_dsp/arm/idct32x32_1_add_neon.c
+vpx_dsp/arm/idct32x32_add_neon.c
+vpx_dsp/arm/idct4x4_1_add_neon.c
+vpx_dsp/arm/idct4x4_add_neon.c
+vpx_dsp/arm/idct8x8_1_add_neon.c
+vpx_dsp/arm/idct8x8_add_neon.c
+vpx_dsp/arm/intrapred_neon.c
+vpx_dsp/arm/loopfilter_16_neon.c
+vpx_dsp/arm/loopfilter_4_neon.c
+vpx_dsp/arm/loopfilter_8_neon.c
+vpx_dsp/arm/loopfilter_neon.c
+vpx_dsp/arm/sad4d_neon.c
+vpx_dsp/arm/sad_neon.c
+vpx_dsp/arm/subpel_variance_neon.c
+vpx_dsp/arm/subtract_neon.c
+vpx_dsp/arm/variance_neon.c
+vpx_dsp/arm/vpx_convolve8_avg_neon.c
+vpx_dsp/arm/vpx_convolve8_neon.c
+vpx_dsp/arm/vpx_convolve_avg_neon.c
+vpx_dsp/arm/vpx_convolve_copy_neon.c
+vpx_dsp/arm/vpx_convolve_neon.c
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/arm.h
+vpx_ports/arm_cpudetect.c
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/arm64/vp8_rtcd.h b/libvpx/config/arm64/vp8_rtcd.h
new file mode 100644
index 0000000..5ab06f4
--- /dev/null
+++ b/libvpx/config/arm64/vp8_rtcd.h
@@ -0,0 +1,201 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_neon
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_c
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x4_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_neon
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_neon
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_c
+
+void vp8_clear_system_state_c();
+#define vp8_clear_system_state vp8_clear_system_state_c
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_neon(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_neon
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x4_neon(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_neon
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x8_neon(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_neon
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+void vp8_dc_only_idct_add_neon(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_neon
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_neon
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_uv_neon(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_neon
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+void vp8_dequant_idct_add_neon(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_neon
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_neon(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_neon
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_neon(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_neon
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+void vp8_dequantize_b_neon(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_neon
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sad_c
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+void vp8_fast_quantize_b_neon(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_neon
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sad_c
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_neon(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_neon
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_neon(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_neon
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_neon(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_neon
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_neon(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_neon
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_neon(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_neon
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_neon(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_neon
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_mbhs_neon(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_mbhs_neon
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_mbvs_neon(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_mbvs_neon
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_c
+
+int vp8_mbuverror_c(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_c
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sad_c
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_c
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_neon(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_neon
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_neon(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_neon
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+void vp8_short_idct4x4llm_neon(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_neon
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+void vp8_short_inv_walsh4x4_neon(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_neon
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+void vp8_short_walsh4x4_neon(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_neon
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_neon
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_c
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_neon
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_neon(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_neon
+
+void vp8_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm64/vp9_rtcd.h b/libvpx/config/arm64/vp9_rtcd.h
new file mode 100644
index 0000000..cf2f106
--- /dev/null
+++ b/libvpx/config/arm64/vp9_rtcd.h
@@ -0,0 +1,104 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_c
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+int64_t vp9_block_error_fp_neon(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_neon
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_fdct8x8_quant_neon(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_neon
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_c
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_c
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_c
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sad_c
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_c
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_c
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_neon
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_neon
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_fp_neon(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_neon
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_c
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
+
+void vp9_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm64/vpx_config.asm b/libvpx/config/arm64/vpx_config.asm
new file mode 100644
index 0000000..da8c843
--- /dev/null
+++ b/libvpx/config/arm64/vpx_config.asm
@@ -0,0 +1,86 @@
+@ This file was created from a .asm file
+@  using the ads2gas.pl script.
+	.equ DO1STROUNDING, 0
+.equ ARCH_ARM ,  1
+.equ ARCH_MIPS ,  0
+.equ ARCH_X86 ,  0
+.equ ARCH_X86_64 ,  0
+.equ HAVE_EDSP ,  0
+.equ HAVE_MEDIA ,  0
+.equ HAVE_NEON ,  1
+.equ HAVE_NEON_ASM ,  0
+.equ HAVE_MIPS32 ,  0
+.equ HAVE_DSPR2 ,  0
+.equ HAVE_MSA ,  0
+.equ HAVE_MIPS64 ,  0
+.equ HAVE_MMX ,  0
+.equ HAVE_SSE ,  0
+.equ HAVE_SSE2 ,  0
+.equ HAVE_SSE3 ,  0
+.equ HAVE_SSSE3 ,  0
+.equ HAVE_SSE4_1 ,  0
+.equ HAVE_AVX ,  0
+.equ HAVE_AVX2 ,  0
+.equ HAVE_VPX_PORTS ,  1
+.equ HAVE_PTHREAD_H ,  1
+.equ HAVE_UNISTD_H ,  1
+.equ CONFIG_DEPENDENCY_TRACKING ,  1
+.equ CONFIG_EXTERNAL_BUILD ,  1
+.equ CONFIG_INSTALL_DOCS ,  0
+.equ CONFIG_INSTALL_BINS ,  1
+.equ CONFIG_INSTALL_LIBS ,  1
+.equ CONFIG_INSTALL_SRCS ,  0
+.equ CONFIG_USE_X86INC ,  0
+.equ CONFIG_DEBUG ,  0
+.equ CONFIG_GPROF ,  0
+.equ CONFIG_GCOV ,  0
+.equ CONFIG_RVCT ,  0
+.equ CONFIG_GCC ,  1
+.equ CONFIG_MSVS ,  0
+.equ CONFIG_PIC ,  1
+.equ CONFIG_BIG_ENDIAN ,  0
+.equ CONFIG_CODEC_SRCS ,  0
+.equ CONFIG_DEBUG_LIBS ,  0
+.equ CONFIG_DEQUANT_TOKENS ,  0
+.equ CONFIG_DC_RECON ,  0
+.equ CONFIG_RUNTIME_CPU_DETECT ,  0
+.equ CONFIG_POSTPROC ,  0
+.equ CONFIG_VP9_POSTPROC ,  0
+.equ CONFIG_MULTITHREAD ,  1
+.equ CONFIG_INTERNAL_STATS ,  0
+.equ CONFIG_VP8_ENCODER ,  1
+.equ CONFIG_VP8_DECODER ,  1
+.equ CONFIG_VP9_ENCODER ,  1
+.equ CONFIG_VP9_DECODER ,  1
+.equ CONFIG_VP8 ,  1
+.equ CONFIG_VP9 ,  1
+.equ CONFIG_ENCODERS ,  1
+.equ CONFIG_DECODERS ,  1
+.equ CONFIG_STATIC_MSVCRT ,  0
+.equ CONFIG_SPATIAL_RESAMPLING ,  1
+.equ CONFIG_REALTIME_ONLY ,  1
+.equ CONFIG_ONTHEFLY_BITPACKING ,  0
+.equ CONFIG_ERROR_CONCEALMENT ,  0
+.equ CONFIG_SHARED ,  0
+.equ CONFIG_STATIC ,  1
+.equ CONFIG_SMALL ,  0
+.equ CONFIG_POSTPROC_VISUALIZER ,  0
+.equ CONFIG_OS_SUPPORT ,  1
+.equ CONFIG_UNIT_TESTS ,  1
+.equ CONFIG_WEBM_IO ,  1
+.equ CONFIG_LIBYUV ,  1
+.equ CONFIG_DECODE_PERF_TESTS ,  0
+.equ CONFIG_ENCODE_PERF_TESTS ,  0
+.equ CONFIG_MULTI_RES_ENCODING ,  0
+.equ CONFIG_TEMPORAL_DENOISING ,  1
+.equ CONFIG_VP9_TEMPORAL_DENOISING ,  0
+.equ CONFIG_COEFFICIENT_RANGE_CHECKING ,  0
+.equ CONFIG_VP9_HIGHBITDEPTH ,  0
+.equ CONFIG_BETTER_HW_COMPATIBILITY ,  0
+.equ CONFIG_EXPERIMENTAL ,  0
+.equ CONFIG_SIZE_LIMIT ,  0
+.equ CONFIG_SPATIAL_SVC ,  0
+.equ CONFIG_FP_MB_STATS ,  0
+.equ CONFIG_EMULATE_HARDWARE ,  0
+.equ CONFIG_MISC_FIXES ,  0
+	.section	.note.GNU-stack,"",%progbits
diff --git a/libvpx/config/arm64/vpx_config.c b/libvpx/config/arm64/vpx_config.c
new file mode 100644
index 0000000..50e1a4b
--- /dev/null
+++ b/libvpx/config/arm64/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--force-target=armv8-linux-gcc --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/arm64/vpx_config.h b/libvpx/config/arm64/vpx_config.h
new file mode 100644
index 0000000..b922a5e
--- /dev/null
+++ b/libvpx/config/arm64/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 1
+#define ARCH_MIPS 0
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 1
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 0
+#define CONFIG_DC_RECON 0
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 0
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/arm64/vpx_dsp_rtcd.h b/libvpx/config/arm64/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..6dfe3d9
--- /dev/null
+++ b/libvpx/config/arm64/vpx_dsp_rtcd.h
@@ -0,0 +1,839 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+unsigned int vpx_avg_4x4_neon(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_neon
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+unsigned int vpx_avg_8x8_neon(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_neon
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_neon
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_neon
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_neon
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_neon
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_neon
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_neon
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_avg_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_neon
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_copy_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_neon
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_neon
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_c
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_c
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_c
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_c
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_c
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_c
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_c
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_c
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_neon
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_c
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_neon
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_neon
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_c
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_c
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_c
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_c
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_neon
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_neon
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_neon
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_neon
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_neon
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_neon
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_neon
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_neon
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_neon
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_neon
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_neon
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_neon
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_neon
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_neon
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_neon
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_neon
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_c
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_c
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_c
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_c
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_c
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_c
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_c
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_neon(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_neon
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_1_neon(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_neon
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get16x16var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_neon
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+unsigned int vpx_get4x4sse_cs_neon(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_neon
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get8x8var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_neon
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_c
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_neon
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_neon
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_neon
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_neon
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void vpx_hadamard_16x16_neon(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_neon
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void vpx_hadamard_8x8_neon(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_neon
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_10_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_neon
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_neon
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_256_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_neon
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_neon
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_1024_add_neon
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_neon
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_1024_add_neon
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_neon
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_neon
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_12_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_neon
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_1_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_neon
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_64_add_neon(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_neon
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+int16_t vpx_int_pro_col_neon(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_neon
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+void vpx_int_pro_row_neon(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_neon
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_c
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_neon
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_neon
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_neon
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_c
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_c
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_c
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_c
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_c
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_4_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_neon
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_4_dual_neon(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_neon
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_8_neon(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_neon
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_c
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void vpx_minmax_8x8_neon(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_neon
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x16_neon(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_neon
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_c
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_c
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_c
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_c
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_neon
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_c
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_c
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_neon
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_c
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_c
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_c
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_c
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_neon
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_c
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_c
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_c
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_c
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_c
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_c
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_c
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x32_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_neon
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_c
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_c
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x32x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_neon
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_c
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_c
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_c
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_c
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad4x4_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_neon
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_c
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_c
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_c
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_c
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_c
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_c
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_c
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_c
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_c
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_c
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_c
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad64x64_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_neon
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_c
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_c
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad64x64x4d_neon(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_neon
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_c
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x16_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_neon
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_c
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_c
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_c
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_c
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_c
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_c
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_c
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_c
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x8_neon(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_neon
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_c
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_c
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_c
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_c
+
+int vpx_satd_c(const int16_t *coeff, int length);
+int vpx_satd_neon(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_neon
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_c
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_c
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_c
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_c
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_c
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_c
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_c
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_c
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_c
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_c
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_c
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_c
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_c
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_c
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x16_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_neon
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_c
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_c
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_c
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x32_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_neon
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_c
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_c
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_c
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_c
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x64_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_neon
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_c
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_c
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x8_neon(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_neon
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void vpx_subtract_block_neon(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_neon
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_neon
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_neon
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_neon
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_neon
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_neon
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_neon
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_neon
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_neon
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_neon
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_c
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_neon
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_c
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_neon
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_neon
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_c
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_c
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x32_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_neon
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x64_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_neon
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x16_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_neon
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_c
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x8_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_neon
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_c
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_c
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_c
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+int vpx_vector_var_neon(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_neon
+
+void vpx_dsp_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm64/vpx_scale_rtcd.h b/libvpx/config/arm64/vpx_scale_rtcd.h
new file mode 100644
index 0000000..a1564b7
--- /dev/null
+++ b/libvpx/config/arm64/vpx_scale_rtcd.h
@@ -0,0 +1,71 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_c
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/arm64/vpx_version.h b/libvpx/config/arm64/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/arm64/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/config/generic/libvpx_srcs.txt b/libvpx/config/generic/libvpx_srcs.txt
new file mode 100644
index 0000000..cdc83db
--- /dev/null
+++ b/libvpx/config/generic/libvpx_srcs.txt
@@ -0,0 +1,334 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8dx.mk
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/generic/vp8_rtcd.h b/libvpx/config/generic/vp8_rtcd.h
new file mode 100644
index 0000000..bad72a3
--- /dev/null
+++ b/libvpx/config/generic/vp8_rtcd.h
@@ -0,0 +1,166 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_c
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_c
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_c
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_c
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_c
+
+void vp8_clear_system_state_c();
+#define vp8_clear_system_state vp8_clear_system_state_c
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_c
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_c
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_c
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_c
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_c
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_c
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_c
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_c
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_c
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_c
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sad_c
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_c
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sad_c
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_c
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_c
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_c
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_c
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_c
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_c
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_c
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_c
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_c
+
+int vp8_mbuverror_c(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_c
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sad_c
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_c
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_c
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_c
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_c
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_c
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_c
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_c
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_c
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_c
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_c
+
+void vp8_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/generic/vp9_rtcd.h b/libvpx/config/generic/vp9_rtcd.h
new file mode 100644
index 0000000..c1a84a0
--- /dev/null
+++ b/libvpx/config/generic/vp9_rtcd.h
@@ -0,0 +1,94 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_c
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_c
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_c
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_c
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_c
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_c
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sad_c
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_c
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_c
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_c
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_c
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_c
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
+
+void vp9_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/generic/vpx_config.asm b/libvpx/config/generic/vpx_config.asm
new file mode 100644
index 0000000..3fe4feb
--- /dev/null
+++ b/libvpx/config/generic/vpx_config.asm
@@ -0,0 +1,86 @@
+@ This file was created from a .asm file
+@  using the ads2gas.pl script.
+	.equ DO1STROUNDING, 0
+.equ ARCH_ARM ,  0
+.equ ARCH_MIPS ,  0
+.equ ARCH_X86 ,  0
+.equ ARCH_X86_64 ,  0
+.equ HAVE_EDSP ,  0
+.equ HAVE_MEDIA ,  0
+.equ HAVE_NEON ,  0
+.equ HAVE_NEON_ASM ,  0
+.equ HAVE_MIPS32 ,  0
+.equ HAVE_DSPR2 ,  0
+.equ HAVE_MSA ,  0
+.equ HAVE_MIPS64 ,  0
+.equ HAVE_MMX ,  0
+.equ HAVE_SSE ,  0
+.equ HAVE_SSE2 ,  0
+.equ HAVE_SSE3 ,  0
+.equ HAVE_SSSE3 ,  0
+.equ HAVE_SSE4_1 ,  0
+.equ HAVE_AVX ,  0
+.equ HAVE_AVX2 ,  0
+.equ HAVE_VPX_PORTS ,  1
+.equ HAVE_PTHREAD_H ,  1
+.equ HAVE_UNISTD_H ,  1
+.equ CONFIG_DEPENDENCY_TRACKING ,  1
+.equ CONFIG_EXTERNAL_BUILD ,  1
+.equ CONFIG_INSTALL_DOCS ,  0
+.equ CONFIG_INSTALL_BINS ,  1
+.equ CONFIG_INSTALL_LIBS ,  1
+.equ CONFIG_INSTALL_SRCS ,  0
+.equ CONFIG_USE_X86INC ,  0
+.equ CONFIG_DEBUG ,  0
+.equ CONFIG_GPROF ,  0
+.equ CONFIG_GCOV ,  0
+.equ CONFIG_RVCT ,  0
+.equ CONFIG_GCC ,  1
+.equ CONFIG_MSVS ,  0
+.equ CONFIG_PIC ,  1
+.equ CONFIG_BIG_ENDIAN ,  0
+.equ CONFIG_CODEC_SRCS ,  0
+.equ CONFIG_DEBUG_LIBS ,  0
+.equ CONFIG_DEQUANT_TOKENS ,  0
+.equ CONFIG_DC_RECON ,  0
+.equ CONFIG_RUNTIME_CPU_DETECT ,  0
+.equ CONFIG_POSTPROC ,  0
+.equ CONFIG_VP9_POSTPROC ,  0
+.equ CONFIG_MULTITHREAD ,  1
+.equ CONFIG_INTERNAL_STATS ,  0
+.equ CONFIG_VP8_ENCODER ,  1
+.equ CONFIG_VP8_DECODER ,  1
+.equ CONFIG_VP9_ENCODER ,  1
+.equ CONFIG_VP9_DECODER ,  1
+.equ CONFIG_VP8 ,  1
+.equ CONFIG_VP9 ,  1
+.equ CONFIG_ENCODERS ,  1
+.equ CONFIG_DECODERS ,  1
+.equ CONFIG_STATIC_MSVCRT ,  0
+.equ CONFIG_SPATIAL_RESAMPLING ,  1
+.equ CONFIG_REALTIME_ONLY ,  1
+.equ CONFIG_ONTHEFLY_BITPACKING ,  0
+.equ CONFIG_ERROR_CONCEALMENT ,  0
+.equ CONFIG_SHARED ,  0
+.equ CONFIG_STATIC ,  1
+.equ CONFIG_SMALL ,  0
+.equ CONFIG_POSTPROC_VISUALIZER ,  0
+.equ CONFIG_OS_SUPPORT ,  1
+.equ CONFIG_UNIT_TESTS ,  1
+.equ CONFIG_WEBM_IO ,  1
+.equ CONFIG_LIBYUV ,  1
+.equ CONFIG_DECODE_PERF_TESTS ,  0
+.equ CONFIG_ENCODE_PERF_TESTS ,  0
+.equ CONFIG_MULTI_RES_ENCODING ,  0
+.equ CONFIG_TEMPORAL_DENOISING ,  1
+.equ CONFIG_VP9_TEMPORAL_DENOISING ,  0
+.equ CONFIG_COEFFICIENT_RANGE_CHECKING ,  0
+.equ CONFIG_VP9_HIGHBITDEPTH ,  0
+.equ CONFIG_BETTER_HW_COMPATIBILITY ,  0
+.equ CONFIG_EXPERIMENTAL ,  0
+.equ CONFIG_SIZE_LIMIT ,  0
+.equ CONFIG_SPATIAL_SVC ,  0
+.equ CONFIG_FP_MB_STATS ,  0
+.equ CONFIG_EMULATE_HARDWARE ,  0
+.equ CONFIG_MISC_FIXES ,  0
+	.section	.note.GNU-stack,"",%progbits
diff --git a/libvpx/config/generic/vpx_config.c b/libvpx/config/generic/vpx_config.c
new file mode 100644
index 0000000..269eb82
--- /dev/null
+++ b/libvpx/config/generic/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--target=generic-gnu --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/generic/vpx_config.h b/libvpx/config/generic/vpx_config.h
new file mode 100644
index 0000000..dd011ae
--- /dev/null
+++ b/libvpx/config/generic/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 0
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 0
+#define CONFIG_DC_RECON 0
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 0
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/generic/vpx_dsp_rtcd.h b/libvpx/config/generic/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..c941e34
--- /dev/null
+++ b/libvpx/config/generic/vpx_dsp_rtcd.h
@@ -0,0 +1,738 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_c
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_c
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_c
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_c
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_c
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_c
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_c
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_c
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_c
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_c
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_c
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_c
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_c
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_c
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_c
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_c
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_c
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_c
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_c
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_c
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_c
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_c
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_c
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_c
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_c
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_c
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_c
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_c
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_c
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_c
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_c
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_c
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_c
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_c
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_c
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_c
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_c
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_c
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_c
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_c
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_c
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_c
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_c
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_c
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_c
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_c
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_c
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_c
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_c
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_c
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_c
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_c
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_c
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_c
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_c
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_c
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_c
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_c
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_c
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_c
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_c
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_c
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_c
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_c
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_c
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_135_add_c
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_c
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_34_add_c
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_c
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_c
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_c
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_c
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_c
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_c
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_c
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_c
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_c
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_c
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_c
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_c
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_c
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_c
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_c
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_c
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_c
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_c
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_c
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_c
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_c
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_c
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_c
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_c
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_c
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_c
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_c
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_c
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_c
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_c
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_c
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_c
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_c
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_c
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_c
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_c
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_c
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_c
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_c
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_c
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_c
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_c
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_c
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_c
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_c
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_c
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_c
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_c
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_c
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_c
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_c
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_c
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_c
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_c
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_c
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_c
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_c
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_c
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_c
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_c
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_c
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_c
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_c
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_c
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_c
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_c
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_c
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_c
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_c
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_c
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_c
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_c
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_c
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_c
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_c
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_c
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_c
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_c
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_c
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_c
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_c
+
+int vpx_satd_c(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_c
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_c
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_c
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_c
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_c
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_c
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_c
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_c
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_c
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_c
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_c
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_c
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_c
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_c
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_c
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_c
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_c
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_c
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_c
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_c
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_c
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_c
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_c
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_c
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_c
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_c
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_c
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_c
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_c
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_c
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_c
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_c
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_c
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_c
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_c
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_c
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_c
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_c
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_c
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_c
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_c
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_c
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_c
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_c
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_c
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_c
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_c
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_c
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_c
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_c
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_c
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_c
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_c
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_c
+
+void vpx_dsp_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/generic/vpx_scale_rtcd.h b/libvpx/config/generic/vpx_scale_rtcd.h
new file mode 100644
index 0000000..f419cc7
--- /dev/null
+++ b/libvpx/config/generic/vpx_scale_rtcd.h
@@ -0,0 +1,66 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_c
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/generic/vpx_version.h b/libvpx/config/generic/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/generic/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/config/mips32-dspr2/libvpx_srcs.txt b/libvpx/config/mips32-dspr2/libvpx_srcs.txt
new file mode 100644
index 0000000..c1fc5ee
--- /dev/null
+++ b/libvpx/config/mips32-dspr2/libvpx_srcs.txt
@@ -0,0 +1,373 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/mips/dspr2/dequantize_dspr2.c
+vp8/common/mips/dspr2/filter_dspr2.c
+vp8/common/mips/dspr2/idct_blk_dspr2.c
+vp8/common/mips/dspr2/idctllm_dspr2.c
+vp8/common/mips/dspr2/reconinter_dspr2.c
+vp8/common/mips/dspr2/vp8_loopfilter_filters_dspr2.c
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8dx.mk
+vp9/common/mips/dspr2/vp9_itrans16_dspr2.c
+vp9/common/mips/dspr2/vp9_itrans4_dspr2.c
+vp9/common/mips/dspr2/vp9_itrans8_dspr2.c
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/mips/common_dspr2.c
+vpx_dsp/mips/common_dspr2.h
+vpx_dsp/mips/convolve2_avg_dspr2.c
+vpx_dsp/mips/convolve2_avg_horiz_dspr2.c
+vpx_dsp/mips/convolve2_dspr2.c
+vpx_dsp/mips/convolve2_horiz_dspr2.c
+vpx_dsp/mips/convolve2_vert_dspr2.c
+vpx_dsp/mips/convolve8_avg_dspr2.c
+vpx_dsp/mips/convolve8_avg_horiz_dspr2.c
+vpx_dsp/mips/convolve8_dspr2.c
+vpx_dsp/mips/convolve8_horiz_dspr2.c
+vpx_dsp/mips/convolve8_vert_dspr2.c
+vpx_dsp/mips/convolve_common_dspr2.h
+vpx_dsp/mips/intrapred16_dspr2.c
+vpx_dsp/mips/intrapred4_dspr2.c
+vpx_dsp/mips/intrapred8_dspr2.c
+vpx_dsp/mips/inv_txfm_dspr2.h
+vpx_dsp/mips/itrans16_dspr2.c
+vpx_dsp/mips/itrans32_cols_dspr2.c
+vpx_dsp/mips/itrans32_dspr2.c
+vpx_dsp/mips/itrans4_dspr2.c
+vpx_dsp/mips/itrans8_dspr2.c
+vpx_dsp/mips/loopfilter_filters_dspr2.c
+vpx_dsp/mips/loopfilter_filters_dspr2.h
+vpx_dsp/mips/loopfilter_macros_dspr2.h
+vpx_dsp/mips/loopfilter_masks_dspr2.h
+vpx_dsp/mips/loopfilter_mb_dspr2.c
+vpx_dsp/mips/loopfilter_mb_horiz_dspr2.c
+vpx_dsp/mips/loopfilter_mb_vert_dspr2.c
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/mips/dspr2/yv12extend_dspr2.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/mips32-dspr2/vp8_rtcd.h b/libvpx/config/mips32-dspr2/vp8_rtcd.h
new file mode 100644
index 0000000..03d3f0c
--- /dev/null
+++ b/libvpx/config/mips32-dspr2/vp8_rtcd.h
@@ -0,0 +1,195 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_c
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_c
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_c
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_c
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_c
+
+void vp8_clear_system_state_c();
+#define vp8_clear_system_state vp8_clear_system_state_c
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_dspr2(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_dspr2
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x4_dspr2(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_dspr2
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x8_dspr2(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_dspr2
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+void vp8_dc_only_idct_add_dspr2(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_dspr2
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_c
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_c
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+void vp8_dequant_idct_add_dspr2(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_dspr2
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_dspr2(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_dspr2
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_dspr2(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_dspr2
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_c
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sad_c
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_c
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sad_c
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_dspr2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_dspr2
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_dspr2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_dspr2
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_dspr2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_dspr2
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_dspr2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_dspr2
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_c
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_c
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_c
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_c
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_c
+
+int vp8_mbuverror_c(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_c
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sad_c
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_c
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_c
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_c
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+void vp8_short_idct4x4llm_dspr2(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_dspr2
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+void vp8_short_inv_walsh4x4_dspr2(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_dspr2
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+void vp8_short_inv_walsh4x4_1_dspr2(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_dspr2
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_c
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_dspr2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_dspr2
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict4x4_dspr2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_dspr2
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_dspr2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_dspr2
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_dspr2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_dspr2
+
+void vp8_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32-dspr2/vp9_rtcd.h b/libvpx/config/mips32-dspr2/vp9_rtcd.h
new file mode 100644
index 0000000..a91d894
--- /dev/null
+++ b/libvpx/config/mips32-dspr2/vp9_rtcd.h
@@ -0,0 +1,108 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_c
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_c
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_c
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_c
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_c
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_c
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sad_c
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_c
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+void vp9_iht16x16_256_add_dspr2(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_dspr2
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht4x4_16_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_dspr2
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht8x8_64_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_dspr2
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_c
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
+
+void vp9_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32-dspr2/vpx_config.c b/libvpx/config/mips32-dspr2/vpx_config.c
new file mode 100644
index 0000000..b2f14a3
--- /dev/null
+++ b/libvpx/config/mips32-dspr2/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--target=mips32-linux-gcc --enable-dspr2 --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/mips32-dspr2/vpx_config.h b/libvpx/config/mips32-dspr2/vpx_config.h
new file mode 100644
index 0000000..cbb9721
--- /dev/null
+++ b/libvpx/config/mips32-dspr2/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 1
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 1
+#define HAVE_DSPR2 1
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 1
+#define CONFIG_DC_RECON 1
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 0
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/mips32-dspr2/vpx_dsp_rtcd.h b/libvpx/config/mips32-dspr2/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..37baf0e
--- /dev/null
+++ b/libvpx/config/mips32-dspr2/vpx_dsp_rtcd.h
@@ -0,0 +1,789 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_c
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_c
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_dspr2
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_dspr2
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_dspr2
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_dspr2
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_dspr2
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_dspr2
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_avg_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_dspr2
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_copy_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_dspr2
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_c
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_c
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_c
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_c
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_c
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_c
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_c
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_c
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_c
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_c
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_c
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_c
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_c
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_c
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_c
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_c
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_c
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_c
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_c
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_c
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_c
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_c
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_c
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_c
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_dspr2
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_c
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_dspr2
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_dspr2
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_c
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_c
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_c
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_c
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_c
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_c
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_c
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_c
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_c
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_c
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_c
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_c
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_c
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_c
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_c
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_c
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_c
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_dspr2
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_c
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_dspr2
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_dspr2
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_c
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_c
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_10_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_dspr2
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_1_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_dspr2
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_256_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_dspr2
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_dspr2
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_1024_add_dspr2
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_dspr2
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_34_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_34_add_dspr2
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_16_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_dspr2
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_1_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_dspr2
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_12_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_dspr2
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_1_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_dspr2
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_64_add_dspr2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_dspr2
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_c
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_c
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_c
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_4_dspr2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_dspr2
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_4_dual_dspr2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_dspr2
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_8_dspr2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_dspr2
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_8_dual_dspr2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_dspr2
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_16_dspr2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_dspr2
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_8_dspr2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_dspr2
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_dspr2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_dspr2
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_dual_dspr2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_dspr2
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_4_dspr2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_dspr2
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_4_dual_dspr2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_dspr2
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_8_dspr2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_dspr2
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_8_dual_dspr2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_dspr2
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_c
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_c
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_c
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_c
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_c
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_c
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_c
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_c
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_c
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_c
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_c
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_c
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_c
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_c
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_c
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_c
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_c
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_c
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_c
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_c
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_c
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_c
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_c
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_c
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_c
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_c
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_c
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_c
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_c
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_c
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_c
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_c
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_c
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_c
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_c
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_c
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_c
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_c
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_c
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_c
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_c
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_c
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_c
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_c
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_c
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_c
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_c
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_c
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_c
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_c
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_c
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_c
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_c
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_c
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_c
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_c
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_c
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_c
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_c
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_c
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_c
+
+int vpx_satd_c(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_c
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_c
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_c
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_c
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_c
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_c
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_c
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_c
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_c
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_c
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_c
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_c
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_c
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_c
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_c
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_c
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_c
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_c
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_c
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_c
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_c
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_c
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_c
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_c
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_c
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_c
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_c
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_c
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_c
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_c
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_c
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_dspr2
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_dspr2
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_c
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_c
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_c
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_c
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_c
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_c
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_c
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_c
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_c
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_c
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_c
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_c
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_c
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_c
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_c
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_c
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_c
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_c
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_c
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_c
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_c
+
+void vpx_dsp_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32-dspr2/vpx_scale_rtcd.h b/libvpx/config/mips32-dspr2/vpx_scale_rtcd.h
new file mode 100644
index 0000000..15b1b5a
--- /dev/null
+++ b/libvpx/config/mips32-dspr2/vpx_scale_rtcd.h
@@ -0,0 +1,79 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+void vpx_extend_frame_borders_dspr2(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_dspr2
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+void vpx_extend_frame_inner_borders_dspr2(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_dspr2
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32-dspr2/vpx_version.h b/libvpx/config/mips32-dspr2/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/mips32-dspr2/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/config/mips32-msa/libvpx_srcs.txt b/libvpx/config/mips32-msa/libvpx_srcs.txt
new file mode 100644
index 0000000..2ae46f1
--- /dev/null
+++ b/libvpx/config/mips32-msa/libvpx_srcs.txt
@@ -0,0 +1,382 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/mips/msa/bilinear_filter_msa.c
+vp8/common/mips/msa/copymem_msa.c
+vp8/common/mips/msa/idct_msa.c
+vp8/common/mips/msa/loopfilter_filters_msa.c
+vp8/common/mips/msa/sixtap_filter_msa.c
+vp8/common/mips/msa/vp8_macros_msa.h
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/mips/msa/dct_msa.c
+vp8/encoder/mips/msa/denoising_msa.c
+vp8/encoder/mips/msa/encodeopt_msa.c
+vp8/encoder/mips/msa/quantize_msa.c
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8dx.mk
+vp9/common/mips/msa/vp9_idct16x16_msa.c
+vp9/common/mips/msa/vp9_idct4x4_msa.c
+vp9/common/mips/msa/vp9_idct8x8_msa.c
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/mips/msa/vp9_error_msa.c
+vp9/encoder/mips/msa/vp9_fdct16x16_msa.c
+vp9/encoder/mips/msa/vp9_fdct4x4_msa.c
+vp9/encoder/mips/msa/vp9_fdct8x8_msa.c
+vp9/encoder/mips/msa/vp9_fdct_msa.h
+vp9/encoder/mips/msa/vp9_temporal_filter_msa.c
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/mips/avg_msa.c
+vpx_dsp/mips/fwd_dct32x32_msa.c
+vpx_dsp/mips/fwd_txfm_msa.c
+vpx_dsp/mips/fwd_txfm_msa.h
+vpx_dsp/mips/idct16x16_msa.c
+vpx_dsp/mips/idct32x32_msa.c
+vpx_dsp/mips/idct4x4_msa.c
+vpx_dsp/mips/idct8x8_msa.c
+vpx_dsp/mips/intrapred_msa.c
+vpx_dsp/mips/inv_txfm_msa.h
+vpx_dsp/mips/loopfilter_16_msa.c
+vpx_dsp/mips/loopfilter_4_msa.c
+vpx_dsp/mips/loopfilter_8_msa.c
+vpx_dsp/mips/loopfilter_msa.h
+vpx_dsp/mips/macros_msa.h
+vpx_dsp/mips/sad_msa.c
+vpx_dsp/mips/sub_pixel_variance_msa.c
+vpx_dsp/mips/subtract_msa.c
+vpx_dsp/mips/txfm_macros_msa.h
+vpx_dsp/mips/variance_msa.c
+vpx_dsp/mips/vpx_convolve8_avg_horiz_msa.c
+vpx_dsp/mips/vpx_convolve8_avg_msa.c
+vpx_dsp/mips/vpx_convolve8_avg_vert_msa.c
+vpx_dsp/mips/vpx_convolve8_horiz_msa.c
+vpx_dsp/mips/vpx_convolve8_msa.c
+vpx_dsp/mips/vpx_convolve8_vert_msa.c
+vpx_dsp/mips/vpx_convolve_avg_msa.c
+vpx_dsp/mips/vpx_convolve_copy_msa.c
+vpx_dsp/mips/vpx_convolve_msa.h
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/mips32-msa/vp8_rtcd.h b/libvpx/config/mips32-msa/vp8_rtcd.h
new file mode 100644
index 0000000..802a14e
--- /dev/null
+++ b/libvpx/config/mips32-msa/vp8_rtcd.h
@@ -0,0 +1,213 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_msa
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict4x4_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_msa
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x4_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_msa
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_msa
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+int vp8_block_error_msa(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_msa
+
+void vp8_clear_system_state_c();
+#define vp8_clear_system_state vp8_clear_system_state_c
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_msa(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_msa
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x4_msa(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_msa
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x8_msa(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_msa
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+void vp8_dc_only_idct_add_msa(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_msa
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_msa(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_msa
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_uv_msa(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_msa
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+void vp8_dequant_idct_add_msa(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_msa
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_msa(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_msa
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_msa(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_msa
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+void vp8_dequantize_b_msa(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_msa
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sad_c
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+void vp8_fast_quantize_b_msa(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_msa
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sad_c
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_msa(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_msa
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_msa(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_msa
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_msa(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_msa
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_msa(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_msa
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_msa(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_msa
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_msa(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_msa
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_horizontal_edge_msa(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_msa
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_vertical_edge_msa(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_msa
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+int vp8_mbblock_error_msa(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_msa
+
+int vp8_mbuverror_c(struct macroblock *mb);
+int vp8_mbuverror_msa(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_msa
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sad_c
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+void vp8_regular_quantize_b_msa(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_msa
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_msa(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_msa
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_msa(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_msa
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+void vp8_short_idct4x4llm_msa(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_msa
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+void vp8_short_inv_walsh4x4_msa(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_msa
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+void vp8_short_walsh4x4_msa(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_msa
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_msa
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict4x4_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_msa
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_msa
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_msa
+
+void vp8_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32-msa/vp9_rtcd.h b/libvpx/config/mips32-msa/vp9_rtcd.h
new file mode 100644
index 0000000..c51dcc7
--- /dev/null
+++ b/libvpx/config/mips32-msa/vp9_rtcd.h
@@ -0,0 +1,114 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t vp9_block_error_msa(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_msa
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_c
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_c
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht16x16_msa(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_msa
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht4x4_msa(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_msa
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht8x8_msa(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_msa
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sad_c
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void vp9_fwht4x4_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_msa
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+void vp9_iht16x16_256_add_msa(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_msa
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht4x4_16_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_msa
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht8x8_64_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_msa
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_c
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_msa(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_msa
+
+void vp9_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32-msa/vpx_config.c b/libvpx/config/mips32-msa/vpx_config.c
new file mode 100644
index 0000000..51a7c06
--- /dev/null
+++ b/libvpx/config/mips32-msa/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--target=mips32-linux-gcc --enable-msa --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/mips32-msa/vpx_config.h b/libvpx/config/mips32-msa/vpx_config.h
new file mode 100644
index 0000000..614ff4d
--- /dev/null
+++ b/libvpx/config/mips32-msa/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 1
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 1
+#define HAVE_DSPR2 0
+#define HAVE_MSA 1
+#define HAVE_MIPS64 0
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 1
+#define CONFIG_DC_RECON 1
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 0
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/mips32-msa/vpx_dsp_rtcd.h b/libvpx/config/mips32-msa/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..52dadf7
--- /dev/null
+++ b/libvpx/config/mips32-msa/vpx_dsp_rtcd.h
@@ -0,0 +1,924 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+unsigned int vpx_avg_4x4_msa(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_msa
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+unsigned int vpx_avg_8x8_msa(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_msa
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_msa
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_msa
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_horiz_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_msa
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_vert_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_msa
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_horiz_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_msa
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_vert_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_msa
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_avg_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_msa
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_copy_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_msa
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_c
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_c
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_c
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_c
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_c
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_c
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_c
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_c
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_c
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_c
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_c
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_c
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_c
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_c
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_c
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_c
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_msa
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_msa
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_msa
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_msa
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_msa
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_msa
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_msa
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_msa
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_msa
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_msa
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_msa
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_msa
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_msa
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_msa
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_msa
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_msa
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct16x16_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_msa
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct16x16_1_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_msa
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_msa
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_1_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_msa
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_rd_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_msa
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct4x4_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_msa
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_c
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_msa
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_1_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_msa
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get16x16var_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_msa
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+unsigned int vpx_get4x4sse_cs_msa(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_msa
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get8x8var_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_msa
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+unsigned int vpx_get_mb_ss_msa(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_msa
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_msa
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_msa
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_msa
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_msa
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_c
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_c
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_10_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_msa
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_1_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_msa
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_256_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_msa
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_msa
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_1024_add_msa
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_msa
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_34_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_34_add_msa
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_16_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_msa
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_1_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_msa
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_12_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_msa
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_1_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_msa
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_64_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_msa
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_c
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_c
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_iwht4x4_16_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_msa
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_iwht4x4_1_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_msa
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_4_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_msa
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_4_dual_msa(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_msa
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_8_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_msa
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_8_dual_msa(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_msa
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_16_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_msa
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_8_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_msa
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_msa
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_dual_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_msa
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_4_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_msa
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_4_dual_msa(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_msa
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_8_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_msa
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_8_dual_msa(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_msa
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_c
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x16_msa(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_msa
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x8_msa(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_msa
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse8x16_msa(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_msa
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse8x8_msa(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_msa
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_c
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x16_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_msa
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x16_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_msa
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_msa
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_msa
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_msa
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x32_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_msa
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x32_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_msa
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x32x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_msa
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_msa
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x8_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_msa
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_msa
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_msa
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_msa
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x16_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_msa
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x16_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_msa
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x16x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_msa
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x32_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_msa
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x32_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_msa
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad32x32x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_msa
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x32x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_msa
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad32x32x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_msa
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x64_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_msa
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x64_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_msa
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x64x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_msa
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad4x4_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_msa
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad4x4_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_msa
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad4x4x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_msa
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad4x4x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_msa
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad4x4x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_msa
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad4x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_msa
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad4x8_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_msa
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad4x8x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_msa
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad4x8x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_msa
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad64x32_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_msa
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad64x32_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_msa
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad64x32x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_msa
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad64x64_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_msa
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad64x64_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_msa
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad64x64x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_msa
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad64x64x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_msa
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad64x64x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_msa
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x16_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_msa
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x16_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_msa
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x16x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_msa
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x16x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_msa
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x16x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_msa
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x4_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_msa
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x4_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_msa
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x4x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_msa
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x4x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_msa
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_msa
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x8_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_msa
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x8x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_msa
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x8x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_msa
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x8x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_msa
+
+int vpx_satd_c(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_c
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_c
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_msa
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_msa
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_msa
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_msa
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_msa
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x64_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_msa
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x4_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_msa
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_msa
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_msa
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x64_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_msa
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_msa
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x4_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_msa
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_msa
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_msa
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_msa
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_msa
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_msa
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_msa
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x64_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_msa
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x4_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_msa
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_msa
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_msa
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x64_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_msa
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_msa
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x4_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_msa
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_msa
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void vpx_subtract_block_msa(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_msa
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_msa
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_msa
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_msa
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_msa
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_msa
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_msa
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_msa
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_msa
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x16_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_msa
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x32_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_msa
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x8_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_msa
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x16_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_msa
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x32_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_msa
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x64_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_msa
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance4x4_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_msa
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance4x8_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_msa
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x32_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_msa
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x64_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_msa
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x16_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_msa
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x4_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_msa
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x8_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_msa
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_c
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_c
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_c
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_c
+
+void vpx_dsp_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32-msa/vpx_scale_rtcd.h b/libvpx/config/mips32-msa/vpx_scale_rtcd.h
new file mode 100644
index 0000000..ea70efc
--- /dev/null
+++ b/libvpx/config/mips32-msa/vpx_scale_rtcd.h
@@ -0,0 +1,77 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_c
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32-msa/vpx_version.h b/libvpx/config/mips32-msa/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/mips32-msa/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/config/mips32/libvpx_srcs.txt b/libvpx/config/mips32/libvpx_srcs.txt
new file mode 100644
index 0000000..cdc83db
--- /dev/null
+++ b/libvpx/config/mips32/libvpx_srcs.txt
@@ -0,0 +1,334 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8dx.mk
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/mips32/vp8_rtcd.h b/libvpx/config/mips32/vp8_rtcd.h
new file mode 100644
index 0000000..791c155
--- /dev/null
+++ b/libvpx/config/mips32/vp8_rtcd.h
@@ -0,0 +1,177 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_c
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_c
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_c
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_c
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_c
+
+void vp8_clear_system_state_c();
+#define vp8_clear_system_state vp8_clear_system_state_c
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_c
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_c
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_c
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_c
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_c
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_c
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_c
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_c
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_c
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_c
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sad_c
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_c
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sad_c
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_c
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_c
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_c
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_c
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_c
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_c
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_c
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_c
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_c
+
+int vp8_mbuverror_c(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_c
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sad_c
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_c
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_c
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_c
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_c
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_c
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_c
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_c
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_c
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_c
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_c
+
+void vp8_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32/vp9_rtcd.h b/libvpx/config/mips32/vp9_rtcd.h
new file mode 100644
index 0000000..784d4a4
--- /dev/null
+++ b/libvpx/config/mips32/vp9_rtcd.h
@@ -0,0 +1,105 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_c
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_c
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_c
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_c
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_c
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_c
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sad_c
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_c
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_c
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_c
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_c
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_c
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
+
+void vp9_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32/vpx_config.c b/libvpx/config/mips32/vpx_config.c
new file mode 100644
index 0000000..c993eb6
--- /dev/null
+++ b/libvpx/config/mips32/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--target=mips32-linux-gcc --disable-dspr2 --disable-msa --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/mips32/vpx_config.h b/libvpx/config/mips32/vpx_config.h
new file mode 100644
index 0000000..bb25008
--- /dev/null
+++ b/libvpx/config/mips32/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 1
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 1
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 1
+#define CONFIG_DC_RECON 1
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 0
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/mips32/vpx_dsp_rtcd.h b/libvpx/config/mips32/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..280695b
--- /dev/null
+++ b/libvpx/config/mips32/vpx_dsp_rtcd.h
@@ -0,0 +1,749 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_c
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_c
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_c
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_c
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_c
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_c
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_c
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_c
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_c
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_c
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_c
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_c
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_c
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_c
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_c
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_c
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_c
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_c
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_c
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_c
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_c
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_c
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_c
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_c
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_c
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_c
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_c
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_c
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_c
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_c
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_c
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_c
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_c
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_c
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_c
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_c
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_c
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_c
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_c
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_c
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_c
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_c
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_c
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_c
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_c
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_c
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_c
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_c
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_c
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_c
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_c
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_c
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_c
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_c
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_c
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_c
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_c
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_c
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_c
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_c
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_c
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_c
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_c
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_c
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_c
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_135_add_c
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_c
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_34_add_c
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_c
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_c
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_c
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_c
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_c
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_c
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_c
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_c
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_c
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_c
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_c
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_c
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_c
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_c
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_c
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_c
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_c
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_c
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_c
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_c
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_c
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_c
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_c
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_c
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_c
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_c
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_c
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_c
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_c
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_c
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_c
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_c
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_c
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_c
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_c
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_c
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_c
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_c
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_c
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_c
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_c
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_c
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_c
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_c
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_c
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_c
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_c
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_c
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_c
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_c
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_c
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_c
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_c
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_c
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_c
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_c
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_c
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_c
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_c
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_c
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_c
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_c
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_c
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_c
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_c
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_c
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_c
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_c
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_c
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_c
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_c
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_c
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_c
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_c
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_c
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_c
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_c
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_c
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_c
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_c
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_c
+
+int vpx_satd_c(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_c
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_c
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_c
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_c
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_c
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_c
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_c
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_c
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_c
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_c
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_c
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_c
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_c
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_c
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_c
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_c
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_c
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_c
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_c
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_c
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_c
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_c
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_c
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_c
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_c
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_c
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_c
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_c
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_c
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_c
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_c
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_c
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_c
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_c
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_c
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_c
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_c
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_c
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_c
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_c
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_c
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_c
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_c
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_c
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_c
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_c
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_c
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_c
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_c
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_c
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_c
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_c
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_c
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_c
+
+void vpx_dsp_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32/vpx_scale_rtcd.h b/libvpx/config/mips32/vpx_scale_rtcd.h
new file mode 100644
index 0000000..ea70efc
--- /dev/null
+++ b/libvpx/config/mips32/vpx_scale_rtcd.h
@@ -0,0 +1,77 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_c
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips32/vpx_version.h b/libvpx/config/mips32/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/mips32/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/config/mips64-msa/libvpx_srcs.txt b/libvpx/config/mips64-msa/libvpx_srcs.txt
new file mode 100644
index 0000000..2ae46f1
--- /dev/null
+++ b/libvpx/config/mips64-msa/libvpx_srcs.txt
@@ -0,0 +1,382 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/mips/msa/bilinear_filter_msa.c
+vp8/common/mips/msa/copymem_msa.c
+vp8/common/mips/msa/idct_msa.c
+vp8/common/mips/msa/loopfilter_filters_msa.c
+vp8/common/mips/msa/sixtap_filter_msa.c
+vp8/common/mips/msa/vp8_macros_msa.h
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/mips/msa/dct_msa.c
+vp8/encoder/mips/msa/denoising_msa.c
+vp8/encoder/mips/msa/encodeopt_msa.c
+vp8/encoder/mips/msa/quantize_msa.c
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8dx.mk
+vp9/common/mips/msa/vp9_idct16x16_msa.c
+vp9/common/mips/msa/vp9_idct4x4_msa.c
+vp9/common/mips/msa/vp9_idct8x8_msa.c
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/mips/msa/vp9_error_msa.c
+vp9/encoder/mips/msa/vp9_fdct16x16_msa.c
+vp9/encoder/mips/msa/vp9_fdct4x4_msa.c
+vp9/encoder/mips/msa/vp9_fdct8x8_msa.c
+vp9/encoder/mips/msa/vp9_fdct_msa.h
+vp9/encoder/mips/msa/vp9_temporal_filter_msa.c
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/mips/avg_msa.c
+vpx_dsp/mips/fwd_dct32x32_msa.c
+vpx_dsp/mips/fwd_txfm_msa.c
+vpx_dsp/mips/fwd_txfm_msa.h
+vpx_dsp/mips/idct16x16_msa.c
+vpx_dsp/mips/idct32x32_msa.c
+vpx_dsp/mips/idct4x4_msa.c
+vpx_dsp/mips/idct8x8_msa.c
+vpx_dsp/mips/intrapred_msa.c
+vpx_dsp/mips/inv_txfm_msa.h
+vpx_dsp/mips/loopfilter_16_msa.c
+vpx_dsp/mips/loopfilter_4_msa.c
+vpx_dsp/mips/loopfilter_8_msa.c
+vpx_dsp/mips/loopfilter_msa.h
+vpx_dsp/mips/macros_msa.h
+vpx_dsp/mips/sad_msa.c
+vpx_dsp/mips/sub_pixel_variance_msa.c
+vpx_dsp/mips/subtract_msa.c
+vpx_dsp/mips/txfm_macros_msa.h
+vpx_dsp/mips/variance_msa.c
+vpx_dsp/mips/vpx_convolve8_avg_horiz_msa.c
+vpx_dsp/mips/vpx_convolve8_avg_msa.c
+vpx_dsp/mips/vpx_convolve8_avg_vert_msa.c
+vpx_dsp/mips/vpx_convolve8_horiz_msa.c
+vpx_dsp/mips/vpx_convolve8_msa.c
+vpx_dsp/mips/vpx_convolve8_vert_msa.c
+vpx_dsp/mips/vpx_convolve_avg_msa.c
+vpx_dsp/mips/vpx_convolve_copy_msa.c
+vpx_dsp/mips/vpx_convolve_msa.h
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/mips64-msa/vp8_rtcd.h b/libvpx/config/mips64-msa/vp8_rtcd.h
new file mode 100644
index 0000000..802a14e
--- /dev/null
+++ b/libvpx/config/mips64-msa/vp8_rtcd.h
@@ -0,0 +1,213 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_msa
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict4x4_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_msa
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x4_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_msa
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_msa
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+int vp8_block_error_msa(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_msa
+
+void vp8_clear_system_state_c();
+#define vp8_clear_system_state vp8_clear_system_state_c
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_msa(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_msa
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x4_msa(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_msa
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x8_msa(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_msa
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+void vp8_dc_only_idct_add_msa(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_msa
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_msa(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_msa
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_uv_msa(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_msa
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+void vp8_dequant_idct_add_msa(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_msa
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_msa(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_msa
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_msa(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_msa
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+void vp8_dequantize_b_msa(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_msa
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sad_c
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+void vp8_fast_quantize_b_msa(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_msa
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sad_c
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_msa(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_msa
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_msa(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_msa
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_msa(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_msa
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_msa(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_msa
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_msa(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_msa
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_msa(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_msa
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_horizontal_edge_msa(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_msa
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_vertical_edge_msa(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_msa
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+int vp8_mbblock_error_msa(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_msa
+
+int vp8_mbuverror_c(struct macroblock *mb);
+int vp8_mbuverror_msa(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_msa
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sad_c
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+void vp8_regular_quantize_b_msa(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_msa
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_msa(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_msa
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_msa(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_msa
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+void vp8_short_idct4x4llm_msa(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_msa
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+void vp8_short_inv_walsh4x4_msa(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_msa
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+void vp8_short_walsh4x4_msa(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_msa
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_msa
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict4x4_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_msa
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_msa
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_msa(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_msa
+
+void vp8_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips64-msa/vp9_rtcd.h b/libvpx/config/mips64-msa/vp9_rtcd.h
new file mode 100644
index 0000000..c51dcc7
--- /dev/null
+++ b/libvpx/config/mips64-msa/vp9_rtcd.h
@@ -0,0 +1,114 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t vp9_block_error_msa(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_msa
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_c
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_c
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht16x16_msa(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_msa
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht4x4_msa(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_msa
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht8x8_msa(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_msa
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sad_c
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void vp9_fwht4x4_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_msa
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+void vp9_iht16x16_256_add_msa(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_msa
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht4x4_16_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_msa
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht8x8_64_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_msa
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_c
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_msa(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_msa
+
+void vp9_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips64-msa/vpx_config.c b/libvpx/config/mips64-msa/vpx_config.c
new file mode 100644
index 0000000..e0cc82a
--- /dev/null
+++ b/libvpx/config/mips64-msa/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--target=mips64-linux-gcc --enable-msa --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/mips64-msa/vpx_config.h b/libvpx/config/mips64-msa/vpx_config.h
new file mode 100644
index 0000000..8e426fa
--- /dev/null
+++ b/libvpx/config/mips64-msa/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 1
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 1
+#define HAVE_MIPS64 1
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 1
+#define CONFIG_DC_RECON 1
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 0
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/mips64-msa/vpx_dsp_rtcd.h b/libvpx/config/mips64-msa/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..52dadf7
--- /dev/null
+++ b/libvpx/config/mips64-msa/vpx_dsp_rtcd.h
@@ -0,0 +1,924 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+unsigned int vpx_avg_4x4_msa(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_msa
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+unsigned int vpx_avg_8x8_msa(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_msa
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_msa
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_msa
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_horiz_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_msa
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_vert_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_msa
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_horiz_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_msa
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_vert_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_msa
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_avg_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_msa
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_copy_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_msa
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_c
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_c
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_c
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_c
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_c
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_c
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_c
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_c
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_c
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_c
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_c
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_c
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_c
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_c
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_c
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_c
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_msa
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_msa
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_msa
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_msa
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_msa
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_msa
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_msa
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_msa
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_msa
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_msa
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_msa
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_msa
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_msa
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_msa
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_msa
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_msa
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct16x16_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_msa
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct16x16_1_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_msa
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_msa
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_1_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_msa
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_rd_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_msa
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct4x4_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_msa
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_c
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_msa
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_1_msa(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_msa
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get16x16var_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_msa
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+unsigned int vpx_get4x4sse_cs_msa(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_msa
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get8x8var_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_msa
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+unsigned int vpx_get_mb_ss_msa(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_msa
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_msa
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_msa
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_msa
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_msa
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_c
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_c
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_10_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_msa
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_1_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_msa
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_256_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_msa
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_msa
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_1024_add_msa
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_msa
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_34_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_34_add_msa
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_16_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_msa
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_1_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_msa
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_12_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_msa
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_1_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_msa
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_64_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_msa
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_c
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_c
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_iwht4x4_16_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_msa
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_iwht4x4_1_add_msa(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_msa
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_4_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_msa
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_4_dual_msa(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_msa
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_8_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_msa
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_8_dual_msa(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_msa
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_16_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_msa
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_8_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_msa
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_msa
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_dual_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_msa
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_4_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_msa
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_4_dual_msa(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_msa
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_8_msa(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_msa
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_8_dual_msa(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_msa
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_c
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x16_msa(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_msa
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x8_msa(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_msa
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse8x16_msa(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_msa
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse8x8_msa(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_msa
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_c
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x16_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_msa
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x16_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_msa
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_msa
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_msa
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_msa
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x32_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_msa
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x32_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_msa
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x32x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_msa
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_msa
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x8_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_msa
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_msa
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_msa
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_msa
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x16_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_msa
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x16_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_msa
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x16x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_msa
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x32_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_msa
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x32_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_msa
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad32x32x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_msa
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x32x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_msa
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad32x32x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_msa
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x64_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_msa
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x64_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_msa
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x64x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_msa
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad4x4_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_msa
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad4x4_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_msa
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad4x4x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_msa
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad4x4x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_msa
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad4x4x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_msa
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad4x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_msa
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad4x8_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_msa
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad4x8x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_msa
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad4x8x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_msa
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad64x32_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_msa
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad64x32_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_msa
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad64x32x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_msa
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad64x64_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_msa
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad64x64_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_msa
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad64x64x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_msa
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad64x64x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_msa
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad64x64x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_msa
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x16_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_msa
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x16_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_msa
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x16x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_msa
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x16x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_msa
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x16x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_msa
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x4_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_msa
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x4_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_msa
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x4x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_msa
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x4x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_msa
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_msa
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x8_avg_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_msa
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x8x3_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_msa
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x8x4d_msa(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_msa
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x8x8_msa(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_msa
+
+int vpx_satd_c(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_c
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_c
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_msa
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_msa
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_msa
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_msa
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_msa
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x64_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_msa
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x4_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_msa
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_msa
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_msa
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x64_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_msa
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_msa
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x4_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_msa
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_msa
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_msa
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_msa
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_msa
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_msa
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_msa
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x64_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_msa
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x4_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_msa
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_msa
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x32_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_msa
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x64_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_msa
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x16_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_msa
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x4_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_msa
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x8_msa(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_msa
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void vpx_subtract_block_msa(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_msa
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_msa
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_msa
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_msa
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_msa
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_msa
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_msa
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_msa
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_msa
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x16_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_msa
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x32_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_msa
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x8_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_msa
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x16_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_msa
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x32_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_msa
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x64_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_msa
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance4x4_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_msa
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance4x8_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_msa
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x32_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_msa
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x64_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_msa
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x16_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_msa
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x4_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_msa
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x8_msa(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_msa
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_c
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_c
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_c
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_c
+
+void vpx_dsp_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips64-msa/vpx_scale_rtcd.h b/libvpx/config/mips64-msa/vpx_scale_rtcd.h
new file mode 100644
index 0000000..ea70efc
--- /dev/null
+++ b/libvpx/config/mips64-msa/vpx_scale_rtcd.h
@@ -0,0 +1,77 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_c
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips64-msa/vpx_version.h b/libvpx/config/mips64-msa/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/mips64-msa/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/config/mips64/libvpx_srcs.txt b/libvpx/config/mips64/libvpx_srcs.txt
new file mode 100644
index 0000000..cdc83db
--- /dev/null
+++ b/libvpx/config/mips64/libvpx_srcs.txt
@@ -0,0 +1,334 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8dx.mk
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/mips64/vp8_rtcd.h b/libvpx/config/mips64/vp8_rtcd.h
new file mode 100644
index 0000000..791c155
--- /dev/null
+++ b/libvpx/config/mips64/vp8_rtcd.h
@@ -0,0 +1,177 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_c
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_c
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_c
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_c
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_c
+
+void vp8_clear_system_state_c();
+#define vp8_clear_system_state vp8_clear_system_state_c
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_c
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_c
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_c
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_c
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_c
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_c
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_c
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_c
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_c
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_c
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sad_c
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_c
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sad_c
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_c
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_c
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_c
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_c
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_c
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_c
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_c
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_c
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_c
+
+int vp8_mbuverror_c(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_c
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sad_c
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_c
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_c
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_c
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_c
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_c
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_c
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_c
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_c
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_c
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_c
+
+void vp8_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips64/vp9_rtcd.h b/libvpx/config/mips64/vp9_rtcd.h
new file mode 100644
index 0000000..784d4a4
--- /dev/null
+++ b/libvpx/config/mips64/vp9_rtcd.h
@@ -0,0 +1,105 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_c
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_c
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_c
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_c
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_c
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_c
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sad_c
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_c
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_c
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_c
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_c
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_c
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_c
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_c
+
+void vp9_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips64/vpx_config.c b/libvpx/config/mips64/vpx_config.c
new file mode 100644
index 0000000..3750612
--- /dev/null
+++ b/libvpx/config/mips64/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--target=mips64-linux-gcc --disable-msa --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/mips64/vpx_config.h b/libvpx/config/mips64/vpx_config.h
new file mode 100644
index 0000000..4aadcf8
--- /dev/null
+++ b/libvpx/config/mips64/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 1
+#define ARCH_X86 0
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 1
+#define HAVE_MMX 0
+#define HAVE_SSE 0
+#define HAVE_SSE2 0
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 0
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 1
+#define CONFIG_DC_RECON 1
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 0
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/mips64/vpx_dsp_rtcd.h b/libvpx/config/mips64/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..280695b
--- /dev/null
+++ b/libvpx/config/mips64/vpx_dsp_rtcd.h
@@ -0,0 +1,749 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_c
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_c
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_c
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_c
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_c
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_c
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_c
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_c
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_c
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_c
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_c
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_c
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_c
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_c
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_c
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_c
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_c
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_c
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_c
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_c
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_c
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_c
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_c
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_c
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_c
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_c
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_c
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_c
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_c
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_c
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_c
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_c
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_c
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_c
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_c
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_c
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_c
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_c
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_c
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_c
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_c
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_c
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_c
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_c
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_c
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_c
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_c
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_c
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_c
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_c
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_c
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_c
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_c
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_c
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_c
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_c
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_c
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_c
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_c
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_c
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_c
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_c
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_c
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_c
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_c
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_135_add_c
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_c
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_34_add_c
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_c
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_c
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_c
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_c
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_c
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_c
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_c
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_c
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_c
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_c
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_c
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_c
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_c
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_c
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_c
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_c
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_c
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_c
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_c
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_c
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_c
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_c
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_c
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_c
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_c
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_c
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_c
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_c
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_c
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_c
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_c
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_c
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_c
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_c
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_c
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_c
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_c
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_c
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_c
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_c
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_c
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_c
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_c
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_c
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_c
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_c
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_c
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_c
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_c
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_c
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_c
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_c
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_c
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_c
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_c
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_c
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_c
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_c
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_c
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_c
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_c
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_c
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_c
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_c
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_c
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_c
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_c
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_c
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_c
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_c
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_c
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_c
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_c
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_c
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_c
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_c
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_c
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_c
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_c
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_c
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_c
+
+int vpx_satd_c(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_c
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_c
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_c
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_c
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_c
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_c
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_c
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_c
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_c
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_c
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_c
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_c
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_c
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_c
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_c
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_c
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_c
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_c
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_c
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_c
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_c
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_c
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_c
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_c
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_c
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_c
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_c
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_c
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_c
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_c
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_c
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_c
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_c
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_c
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_c
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_c
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_c
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_c
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_c
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_c
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_c
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_c
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_c
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_c
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_c
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_c
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_c
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_c
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_c
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_c
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_c
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_c
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_c
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_c
+
+void vpx_dsp_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips64/vpx_scale_rtcd.h b/libvpx/config/mips64/vpx_scale_rtcd.h
new file mode 100644
index 0000000..ea70efc
--- /dev/null
+++ b/libvpx/config/mips64/vpx_scale_rtcd.h
@@ -0,0 +1,77 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_c
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/mips64/vpx_version.h b/libvpx/config/mips64/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/mips64/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/config/x86/libvpx_srcs.txt b/libvpx/config/x86/libvpx_srcs.txt
new file mode 100644
index 0000000..b0f438e
--- /dev/null
+++ b/libvpx/config/x86/libvpx_srcs.txt
@@ -0,0 +1,413 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/mfqe.c
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/postproc.c
+vp8/common/postproc.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/common/x86/copy_sse2.asm
+vp8/common/x86/copy_sse3.asm
+vp8/common/x86/dequantize_mmx.asm
+vp8/common/x86/filter_x86.c
+vp8/common/x86/filter_x86.h
+vp8/common/x86/idct_blk_mmx.c
+vp8/common/x86/idct_blk_sse2.c
+vp8/common/x86/idctllm_mmx.asm
+vp8/common/x86/idctllm_sse2.asm
+vp8/common/x86/iwalsh_mmx.asm
+vp8/common/x86/iwalsh_sse2.asm
+vp8/common/x86/loopfilter_sse2.asm
+vp8/common/x86/loopfilter_x86.c
+vp8/common/x86/mfqe_sse2.asm
+vp8/common/x86/postproc_mmx.asm
+vp8/common/x86/postproc_sse2.asm
+vp8/common/x86/recon_mmx.asm
+vp8/common/x86/recon_sse2.asm
+vp8/common/x86/subpixel_mmx.asm
+vp8/common/x86/subpixel_sse2.asm
+vp8/common/x86/subpixel_ssse3.asm
+vp8/common/x86/vp8_asm_stubs.c
+vp8/common/x86/vp8_loopfilter_mmx.asm
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/encoder/x86/dct_mmx.asm
+vp8/encoder/x86/dct_sse2.asm
+vp8/encoder/x86/denoising_sse2.c
+vp8/encoder/x86/encodeopt.asm
+vp8/encoder/x86/fwalsh_sse2.asm
+vp8/encoder/x86/quantize_mmx.asm
+vp8/encoder/x86/quantize_ssse3.c
+vp8/encoder/x86/vp8_enc_stubs_mmx.c
+vp8/encoder/x86/vp8_enc_stubs_sse2.c
+vp8/encoder/x86/vp8_quantize_sse2.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8dx.mk
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/common/x86/vp9_idct_intrin_sse2.c
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/encoder/x86/vp9_dct_intrin_sse2.c
+vp9/encoder/x86/vp9_dct_sse2.asm
+vp9/encoder/x86/vp9_dct_ssse3.c
+vp9/encoder/x86/vp9_error_sse2.asm
+vp9/encoder/x86/vp9_frame_scale_ssse3.c
+vp9/encoder/x86/vp9_quantize_sse2.c
+vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/add_noise.c
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_dsp/x86/add_noise_sse2.asm
+vpx_dsp/x86/avg_intrin_sse2.c
+vpx_dsp/x86/convolve.h
+vpx_dsp/x86/fwd_dct32x32_impl_sse2.h
+vpx_dsp/x86/fwd_txfm_impl_sse2.h
+vpx_dsp/x86/fwd_txfm_sse2.c
+vpx_dsp/x86/fwd_txfm_sse2.h
+vpx_dsp/x86/halfpix_variance_impl_sse2.asm
+vpx_dsp/x86/halfpix_variance_sse2.c
+vpx_dsp/x86/intrapred_sse2.asm
+vpx_dsp/x86/intrapred_ssse3.asm
+vpx_dsp/x86/inv_txfm_sse2.c
+vpx_dsp/x86/inv_txfm_sse2.h
+vpx_dsp/x86/inv_wht_sse2.asm
+vpx_dsp/x86/loopfilter_sse2.c
+vpx_dsp/x86/quantize_sse2.c
+vpx_dsp/x86/sad4d_sse2.asm
+vpx_dsp/x86/sad_sse2.asm
+vpx_dsp/x86/sad_sse3.asm
+vpx_dsp/x86/sad_ssse3.asm
+vpx_dsp/x86/subpel_variance_sse2.asm
+vpx_dsp/x86/subtract_sse2.asm
+vpx_dsp/x86/txfm_common_sse2.h
+vpx_dsp/x86/variance_sse2.c
+vpx_dsp/x86/vpx_asm_stubs.c
+vpx_dsp/x86/vpx_convolve_copy_sse2.asm
+vpx_dsp/x86/vpx_subpixel_8t_intrin_ssse3.c
+vpx_dsp/x86/vpx_subpixel_8t_sse2.asm
+vpx_dsp/x86/vpx_subpixel_8t_ssse3.asm
+vpx_dsp/x86/vpx_subpixel_bilinear_sse2.asm
+vpx_dsp/x86/vpx_subpixel_bilinear_ssse3.asm
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/emms.asm
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_ports/x86.h
+vpx_ports/x86_abi_support.asm
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/x86/vp8_rtcd.h b/libvpx/config/x86/vp8_rtcd.h
new file mode 100644
index 0000000..cb3d248
--- /dev/null
+++ b/libvpx/config/x86/vp8_rtcd.h
@@ -0,0 +1,276 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_ssse3
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_mmx
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_mmx
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_ssse3
+
+void vp8_blend_b_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp8_blend_b vp8_blend_b_c
+
+void vp8_blend_mb_inner_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp8_blend_mb_inner vp8_blend_mb_inner_c
+
+void vp8_blend_mb_outer_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp8_blend_mb_outer vp8_blend_mb_outer_c
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+int vp8_block_error_mmx(short *coeff, short *dqcoeff);
+int vp8_block_error_xmm(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_xmm
+
+void vp8_clear_system_state_c();
+void vpx_reset_mmx_state();
+#define vp8_clear_system_state vpx_reset_mmx_state
+
+void vp8_copy32xn_c(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
+void vp8_copy32xn_sse2(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
+void vp8_copy32xn_sse3(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
+#define vp8_copy32xn vp8_copy32xn_sse3
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_sse2(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_sse2
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x4_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_mmx
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x8_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_mmx
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+void vp8_dc_only_idct_add_mmx(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_mmx
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_sse2
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_sse2
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+void vp8_dequant_idct_add_mmx(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_mmx
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_mmx(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_sse2(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_sse2
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_mmx(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_sse2(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_sse2
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+void vp8_dequantize_b_mmx(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_mmx
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+int vp8_diamond_search_sadx4(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sadx4
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+void vp8_fast_quantize_b_sse2(struct block *, struct blockd *);
+void vp8_fast_quantize_b_ssse3(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_ssse3
+
+void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+void vp8_filter_by_weight16x16_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+#define vp8_filter_by_weight16x16 vp8_filter_by_weight16x16_sse2
+
+void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+#define vp8_filter_by_weight4x4 vp8_filter_by_weight4x4_c
+
+void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+void vp8_filter_by_weight8x8_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+#define vp8_filter_by_weight8x8 vp8_filter_by_weight8x8_sse2
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+int vp8_full_search_sadx3(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sadx3
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_sse2
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_sse2
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_sse2
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_sse2
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_sse2
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_sse2
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_horizontal_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_horizontal_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_sse2
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_vertical_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_vertical_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_sse2
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+int vp8_mbblock_error_mmx(struct macroblock *mb, int dc);
+int vp8_mbblock_error_xmm(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_xmm
+
+void vp8_mbpost_proc_across_ip_c(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+void vp8_mbpost_proc_across_ip_xmm(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+#define vp8_mbpost_proc_across_ip vp8_mbpost_proc_across_ip_xmm
+
+void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+void vp8_mbpost_proc_down_mmx(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+void vp8_mbpost_proc_down_xmm(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+#define vp8_mbpost_proc_down vp8_mbpost_proc_down_xmm
+
+int vp8_mbuverror_c(struct macroblock *mb);
+int vp8_mbuverror_mmx(struct macroblock *mb);
+int vp8_mbuverror_xmm(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_xmm
+
+void vp8_post_proc_down_and_across_mb_row_c(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size);
+void vp8_post_proc_down_and_across_mb_row_sse2(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size);
+#define vp8_post_proc_down_and_across_mb_row vp8_post_proc_down_and_across_mb_row_sse2
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+int vp8_refining_search_sadx4(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sadx4
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+void vp8_regular_quantize_b_sse2(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_sse2
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_mmx(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_sse2(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_sse2
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_mmx(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_sse2(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_sse2
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+void vp8_short_idct4x4llm_mmx(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_mmx
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+void vp8_short_inv_walsh4x4_mmx(short *input, short *output);
+void vp8_short_inv_walsh4x4_sse2(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_sse2
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_sse2
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_ssse3
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict4x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_ssse3
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_ssse3
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_ssse3
+
+void vp8_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/x86/vp9_rtcd.h b/libvpx/config/x86/vp9_rtcd.h
new file mode 100644
index 0000000..159c85b
--- /dev/null
+++ b/libvpx/config/x86/vp9_rtcd.h
@@ -0,0 +1,112 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t vp9_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_sse2
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+int64_t vp9_block_error_fp_sse2(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_sse2
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_fdct8x8_quant_sse2(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_ssse3
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_sse2
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_sse2
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_sse2
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+int vp9_full_search_sadx3(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sadx3
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void vp9_fwht4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_sse2
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+void vp9_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_sse2
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_sse2
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_sse2
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_sse2
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_c
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+void vp9_scale_and_extend_frame_ssse3(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_ssse3
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_sse2(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_sse2
+
+void vp9_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/x86/vpx_config.asm b/libvpx/config/x86/vpx_config.asm
new file mode 100644
index 0000000..4042a8a
--- /dev/null
+++ b/libvpx/config/x86/vpx_config.asm
@@ -0,0 +1,82 @@
+%define ARCH_ARM 0
+%define ARCH_MIPS 0
+%define ARCH_X86 1
+%define ARCH_X86_64 0
+%define HAVE_EDSP 0
+%define HAVE_MEDIA 0
+%define HAVE_NEON 0
+%define HAVE_NEON_ASM 0
+%define HAVE_MIPS32 0
+%define HAVE_DSPR2 0
+%define HAVE_MSA 0
+%define HAVE_MIPS64 0
+%define HAVE_MMX 1
+%define HAVE_SSE 1
+%define HAVE_SSE2 1
+%define HAVE_SSE3 1
+%define HAVE_SSSE3 1
+%define HAVE_SSE4_1 0
+%define HAVE_AVX 0
+%define HAVE_AVX2 0
+%define HAVE_VPX_PORTS 1
+%define HAVE_PTHREAD_H 1
+%define HAVE_UNISTD_H 1
+%define CONFIG_DEPENDENCY_TRACKING 1
+%define CONFIG_EXTERNAL_BUILD 1
+%define CONFIG_INSTALL_DOCS 0
+%define CONFIG_INSTALL_BINS 1
+%define CONFIG_INSTALL_LIBS 1
+%define CONFIG_INSTALL_SRCS 0
+%define CONFIG_USE_X86INC 1
+%define CONFIG_DEBUG 0
+%define CONFIG_GPROF 0
+%define CONFIG_GCOV 0
+%define CONFIG_RVCT 0
+%define CONFIG_GCC 1
+%define CONFIG_MSVS 0
+%define CONFIG_PIC 1
+%define CONFIG_BIG_ENDIAN 0
+%define CONFIG_CODEC_SRCS 0
+%define CONFIG_DEBUG_LIBS 0
+%define CONFIG_DEQUANT_TOKENS 0
+%define CONFIG_DC_RECON 0
+%define CONFIG_RUNTIME_CPU_DETECT 0
+%define CONFIG_POSTPROC 1
+%define CONFIG_VP9_POSTPROC 0
+%define CONFIG_MULTITHREAD 1
+%define CONFIG_INTERNAL_STATS 0
+%define CONFIG_VP8_ENCODER 1
+%define CONFIG_VP8_DECODER 1
+%define CONFIG_VP9_ENCODER 1
+%define CONFIG_VP9_DECODER 1
+%define CONFIG_VP8 1
+%define CONFIG_VP9 1
+%define CONFIG_ENCODERS 1
+%define CONFIG_DECODERS 1
+%define CONFIG_STATIC_MSVCRT 0
+%define CONFIG_SPATIAL_RESAMPLING 1
+%define CONFIG_REALTIME_ONLY 1
+%define CONFIG_ONTHEFLY_BITPACKING 0
+%define CONFIG_ERROR_CONCEALMENT 0
+%define CONFIG_SHARED 0
+%define CONFIG_STATIC 1
+%define CONFIG_SMALL 0
+%define CONFIG_POSTPROC_VISUALIZER 0
+%define CONFIG_OS_SUPPORT 1
+%define CONFIG_UNIT_TESTS 1
+%define CONFIG_WEBM_IO 1
+%define CONFIG_LIBYUV 1
+%define CONFIG_DECODE_PERF_TESTS 0
+%define CONFIG_ENCODE_PERF_TESTS 0
+%define CONFIG_MULTI_RES_ENCODING 0
+%define CONFIG_TEMPORAL_DENOISING 1
+%define CONFIG_VP9_TEMPORAL_DENOISING 0
+%define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+%define CONFIG_VP9_HIGHBITDEPTH 0
+%define CONFIG_BETTER_HW_COMPATIBILITY 0
+%define CONFIG_EXPERIMENTAL 0
+%define CONFIG_SIZE_LIMIT 0
+%define CONFIG_SPATIAL_SVC 0
+%define CONFIG_FP_MB_STATS 0
+%define CONFIG_EMULATE_HARDWARE 0
+%define CONFIG_MISC_FIXES 0
diff --git a/libvpx/config/x86/vpx_config.c b/libvpx/config/x86/vpx_config.c
new file mode 100644
index 0000000..4b7c558
--- /dev/null
+++ b/libvpx/config/x86/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--target=x86-linux-gcc --disable-sse4_1 --disable-avx --disable-avx2 --as=yasm --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/x86/vpx_config.h b/libvpx/config/x86/vpx_config.h
new file mode 100644
index 0000000..19ee330
--- /dev/null
+++ b/libvpx/config/x86/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 0
+#define ARCH_X86 1
+#define ARCH_X86_64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 1
+#define HAVE_SSE 1
+#define HAVE_SSE2 1
+#define HAVE_SSE3 1
+#define HAVE_SSSE3 1
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 1
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 0
+#define CONFIG_DC_RECON 0
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 1
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/x86/vpx_dsp_rtcd.h b/libvpx/config/x86/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..4d3742a
--- /dev/null
+++ b/libvpx/config/x86/vpx_dsp_rtcd.h
@@ -0,0 +1,970 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+unsigned int vpx_avg_4x4_sse2(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_sse2
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+unsigned int vpx_avg_8x8_sse2(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_sse2
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_ssse3
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_ssse3
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_ssse3
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_ssse3
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_ssse3
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_ssse3
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_sse2
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_sse2
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d153_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_ssse3
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d153_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_ssse3
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d153_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_ssse3
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d153_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_ssse3
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d207_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_ssse3
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d207_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_ssse3
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d207_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_sse2
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d207_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_ssse3
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_ssse3
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_ssse3
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_sse2
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_sse2
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d63_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_ssse3
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d63_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_ssse3
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d63_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_ssse3
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d63_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_ssse3
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_sse2
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_sse2
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_sse2
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_sse2
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_sse2
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_sse2
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_sse2
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_sse2
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_sse2
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_sse2
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_sse2
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_sse2
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_sse2
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_sse2
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_sse2
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_sse2
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_sse2
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_sse2
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_sse2
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_sse2
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_rd_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_sse2
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_sse2
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_sse2
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_sse2
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_sse2
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get16x16var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_sse2
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_c
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get8x8var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_sse2
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+unsigned int vpx_get_mb_ss_sse2(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_sse2
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_sse2
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_sse2
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_sse2
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_sse2
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void vpx_hadamard_16x16_sse2(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_sse2
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void vpx_hadamard_8x8_sse2(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_sse2
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_sse2
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_sse2
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_sse2
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_sse2
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_1024_add_sse2
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_sse2
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_34_add_sse2
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_sse2
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_sse2
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_sse2
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_sse2
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_sse2
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+int16_t vpx_int_pro_col_sse2(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_sse2
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+void vpx_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_sse2
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_iwht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_sse2
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_sse2
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_sse2
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_sse2
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_sse2
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_sse2
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_sse2
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_sse2
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_sse2
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_sse2
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_sse2
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_sse2
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_sse2
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void vpx_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_sse2
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_sse2
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_sse2
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_sse2
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_sse2
+
+void vpx_plane_add_noise_c(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
+void vpx_plane_add_noise_sse2(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
+#define vpx_plane_add_noise vpx_plane_add_noise_sse2
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_sse2
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_c
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_sse2
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_sse2
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x3_ssse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_ssse3
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_sse2
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_c
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_sse2
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_sse2
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_sse2
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_sse2
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_sse2
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x3_ssse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_ssse3
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_sse2
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_c
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_sse2
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_sse2
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_sse2
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_sse2
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_sse2
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_c
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_sse2
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_c
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_sse2
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_sse2
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_sse2
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_sse2
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_sse2
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad4x4x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_sse3
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_sse2
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_c
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_sse2
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_sse2
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_sse2
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_c
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_sse2
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_sse2
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_sse2
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_sse2
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_sse2
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_c
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_sse2
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_c
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_sse2
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_sse2
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x16x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_sse3
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_sse2
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_c
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_sse2
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_sse2
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_sse2
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_c
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_sse2
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_sse2
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x8x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_sse3
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_sse2
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_c
+
+int vpx_satd_c(const int16_t *coeff, int length);
+int vpx_satd_sse2(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_sse2
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_ssse3
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_ssse3
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_ssse3
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_ssse3
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_ssse3
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_ssse3
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_ssse3
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_ssse3
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_ssse3
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_ssse3
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_ssse3
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_ssse3
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_ssse3
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_ssse3
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_ssse3
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void vpx_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_sse2
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_sse2
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_sse2
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_sse2
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_sse2
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_sse2
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_sse2
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_sse2
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_sse2
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_sse2
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_sse2
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_sse2
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_sse2
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_sse2
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_sse2
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_sse2
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_sse2
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_sse2
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_sse2
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_sse2
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_sse2
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_sse2
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_h_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_sse2
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_hv_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_sse2
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_v_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_sse2
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+int vpx_vector_var_sse2(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_sse2
+
+void vpx_dsp_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/x86/vpx_scale_rtcd.h b/libvpx/config/x86/vpx_scale_rtcd.h
new file mode 100644
index 0000000..ddf7d01
--- /dev/null
+++ b/libvpx/config/x86/vpx_scale_rtcd.h
@@ -0,0 +1,69 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_c
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/x86/vpx_version.h b/libvpx/config/x86/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/x86/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/config/x86_64/libvpx_srcs.txt b/libvpx/config/x86_64/libvpx_srcs.txt
new file mode 100644
index 0000000..8ae4cff
--- /dev/null
+++ b/libvpx/config/x86_64/libvpx_srcs.txt
@@ -0,0 +1,420 @@
+CHANGELOG
+build/make/rtcd.pl
+build/make/version.sh
+libs.mk
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/copy_c.c
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.h
+vp8/common/loopfilter_filters.c
+vp8/common/mbpitch.c
+vp8/common/mfqe.c
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyx.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/postproc.c
+vp8/common/postproc.h
+vp8/common/ppflags.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra.c
+vp8/common/reconintra.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/vp8_loopfilter.c
+vp8/common/x86/copy_sse2.asm
+vp8/common/x86/copy_sse3.asm
+vp8/common/x86/dequantize_mmx.asm
+vp8/common/x86/filter_x86.c
+vp8/common/x86/filter_x86.h
+vp8/common/x86/idct_blk_mmx.c
+vp8/common/x86/idct_blk_sse2.c
+vp8/common/x86/idctllm_mmx.asm
+vp8/common/x86/idctllm_sse2.asm
+vp8/common/x86/iwalsh_mmx.asm
+vp8/common/x86/iwalsh_sse2.asm
+vp8/common/x86/loopfilter_block_sse2_x86_64.asm
+vp8/common/x86/loopfilter_sse2.asm
+vp8/common/x86/loopfilter_x86.c
+vp8/common/x86/mfqe_sse2.asm
+vp8/common/x86/postproc_mmx.asm
+vp8/common/x86/postproc_sse2.asm
+vp8/common/x86/recon_mmx.asm
+vp8/common/x86/recon_sse2.asm
+vp8/common/x86/subpixel_mmx.asm
+vp8/common/x86/subpixel_sse2.asm
+vp8/common/x86/subpixel_ssse3.asm
+vp8/common/x86/vp8_asm_stubs.c
+vp8/common/x86/vp8_loopfilter_mmx.asm
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_quantize.c
+vp8/encoder/x86/dct_mmx.asm
+vp8/encoder/x86/dct_sse2.asm
+vp8/encoder/x86/denoising_sse2.c
+vp8/encoder/x86/encodeopt.asm
+vp8/encoder/x86/fwalsh_sse2.asm
+vp8/encoder/x86/quantize_mmx.asm
+vp8/encoder/x86/quantize_ssse3.c
+vp8/encoder/x86/vp8_enc_stubs_mmx.c
+vp8/encoder/x86/vp8_enc_stubs_sse2.c
+vp8/encoder/x86/vp8_quantize_sse2.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8_dx_iface.c
+vp8/vp8cx.mk
+vp8/vp8dx.mk
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_thread_common.c
+vp9/common/vp9_thread_common.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/common/x86/vp9_idct_intrin_sse2.c
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/encoder/vp9_aq_360.c
+vp9/encoder/vp9_aq_360.h
+vp9/encoder/vp9_aq_complexity.c
+vp9/encoder/vp9_aq_complexity.h
+vp9/encoder/vp9_aq_cyclicrefresh.c
+vp9/encoder/vp9_aq_cyclicrefresh.h
+vp9/encoder/vp9_aq_variance.c
+vp9/encoder/vp9_aq_variance.h
+vp9/encoder/vp9_bitstream.c
+vp9/encoder/vp9_bitstream.h
+vp9/encoder/vp9_block.h
+vp9/encoder/vp9_context_tree.c
+vp9/encoder/vp9_context_tree.h
+vp9/encoder/vp9_cost.c
+vp9/encoder/vp9_cost.h
+vp9/encoder/vp9_dct.c
+vp9/encoder/vp9_encodeframe.c
+vp9/encoder/vp9_encodeframe.h
+vp9/encoder/vp9_encodemb.c
+vp9/encoder/vp9_encodemb.h
+vp9/encoder/vp9_encodemv.c
+vp9/encoder/vp9_encodemv.h
+vp9/encoder/vp9_encoder.c
+vp9/encoder/vp9_encoder.h
+vp9/encoder/vp9_ethread.c
+vp9/encoder/vp9_ethread.h
+vp9/encoder/vp9_extend.c
+vp9/encoder/vp9_extend.h
+vp9/encoder/vp9_firstpass.c
+vp9/encoder/vp9_firstpass.h
+vp9/encoder/vp9_lookahead.c
+vp9/encoder/vp9_lookahead.h
+vp9/encoder/vp9_mbgraph.c
+vp9/encoder/vp9_mbgraph.h
+vp9/encoder/vp9_mcomp.c
+vp9/encoder/vp9_mcomp.h
+vp9/encoder/vp9_noise_estimate.c
+vp9/encoder/vp9_noise_estimate.h
+vp9/encoder/vp9_picklpf.c
+vp9/encoder/vp9_picklpf.h
+vp9/encoder/vp9_pickmode.c
+vp9/encoder/vp9_pickmode.h
+vp9/encoder/vp9_quantize.c
+vp9/encoder/vp9_quantize.h
+vp9/encoder/vp9_ratectrl.c
+vp9/encoder/vp9_ratectrl.h
+vp9/encoder/vp9_rd.c
+vp9/encoder/vp9_rd.h
+vp9/encoder/vp9_rdopt.c
+vp9/encoder/vp9_rdopt.h
+vp9/encoder/vp9_resize.c
+vp9/encoder/vp9_resize.h
+vp9/encoder/vp9_segmentation.c
+vp9/encoder/vp9_segmentation.h
+vp9/encoder/vp9_skin_detection.c
+vp9/encoder/vp9_skin_detection.h
+vp9/encoder/vp9_speed_features.c
+vp9/encoder/vp9_speed_features.h
+vp9/encoder/vp9_subexp.c
+vp9/encoder/vp9_subexp.h
+vp9/encoder/vp9_svc_layercontext.c
+vp9/encoder/vp9_svc_layercontext.h
+vp9/encoder/vp9_temporal_filter.c
+vp9/encoder/vp9_temporal_filter.h
+vp9/encoder/vp9_tokenize.c
+vp9/encoder/vp9_tokenize.h
+vp9/encoder/vp9_treewriter.c
+vp9/encoder/vp9_treewriter.h
+vp9/encoder/x86/vp9_dct_intrin_sse2.c
+vp9/encoder/x86/vp9_dct_sse2.asm
+vp9/encoder/x86/vp9_dct_ssse3.c
+vp9/encoder/x86/vp9_error_sse2.asm
+vp9/encoder/x86/vp9_frame_scale_ssse3.c
+vp9/encoder/x86/vp9_quantize_sse2.c
+vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
+vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
+vp9/vp9_common.mk
+vp9/vp9_cx_iface.c
+vp9/vp9_dx_iface.c
+vp9/vp9_dx_iface.h
+vp9/vp9_iface_common.h
+vp9/vp9cx.mk
+vp9/vp9dx.mk
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8.h
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
+vpx_config.c
+vpx_dsp/add_noise.c
+vpx_dsp/avg.c
+vpx_dsp/bitreader.c
+vpx_dsp/bitreader.h
+vpx_dsp/bitreader_buffer.c
+vpx_dsp/bitreader_buffer.h
+vpx_dsp/bitwriter.c
+vpx_dsp/bitwriter.h
+vpx_dsp/bitwriter_buffer.c
+vpx_dsp/bitwriter_buffer.h
+vpx_dsp/fwd_txfm.c
+vpx_dsp/fwd_txfm.h
+vpx_dsp/intrapred.c
+vpx_dsp/inv_txfm.c
+vpx_dsp/inv_txfm.h
+vpx_dsp/loopfilter.c
+vpx_dsp/prob.c
+vpx_dsp/prob.h
+vpx_dsp/quantize.c
+vpx_dsp/quantize.h
+vpx_dsp/sad.c
+vpx_dsp/subtract.c
+vpx_dsp/txfm_common.h
+vpx_dsp/variance.c
+vpx_dsp/variance.h
+vpx_dsp/vpx_convolve.c
+vpx_dsp/vpx_convolve.h
+vpx_dsp/vpx_dsp.mk
+vpx_dsp/vpx_dsp_common.h
+vpx_dsp/vpx_dsp_rtcd.c
+vpx_dsp/vpx_dsp_rtcd_defs.pl
+vpx_dsp/vpx_filter.h
+vpx_dsp/x86/add_noise_sse2.asm
+vpx_dsp/x86/avg_intrin_sse2.c
+vpx_dsp/x86/avg_ssse3_x86_64.asm
+vpx_dsp/x86/convolve.h
+vpx_dsp/x86/fwd_dct32x32_impl_sse2.h
+vpx_dsp/x86/fwd_txfm_impl_sse2.h
+vpx_dsp/x86/fwd_txfm_sse2.c
+vpx_dsp/x86/fwd_txfm_sse2.h
+vpx_dsp/x86/fwd_txfm_ssse3_x86_64.asm
+vpx_dsp/x86/halfpix_variance_impl_sse2.asm
+vpx_dsp/x86/halfpix_variance_sse2.c
+vpx_dsp/x86/intrapred_sse2.asm
+vpx_dsp/x86/intrapred_ssse3.asm
+vpx_dsp/x86/inv_txfm_sse2.c
+vpx_dsp/x86/inv_txfm_sse2.h
+vpx_dsp/x86/inv_txfm_ssse3_x86_64.asm
+vpx_dsp/x86/inv_wht_sse2.asm
+vpx_dsp/x86/loopfilter_sse2.c
+vpx_dsp/x86/quantize_sse2.c
+vpx_dsp/x86/quantize_ssse3_x86_64.asm
+vpx_dsp/x86/sad4d_sse2.asm
+vpx_dsp/x86/sad_sse2.asm
+vpx_dsp/x86/sad_sse3.asm
+vpx_dsp/x86/sad_ssse3.asm
+vpx_dsp/x86/ssim_opt_x86_64.asm
+vpx_dsp/x86/subpel_variance_sse2.asm
+vpx_dsp/x86/subtract_sse2.asm
+vpx_dsp/x86/txfm_common_sse2.h
+vpx_dsp/x86/variance_sse2.c
+vpx_dsp/x86/vpx_asm_stubs.c
+vpx_dsp/x86/vpx_convolve_copy_sse2.asm
+vpx_dsp/x86/vpx_subpixel_8t_intrin_ssse3.c
+vpx_dsp/x86/vpx_subpixel_8t_sse2.asm
+vpx_dsp/x86/vpx_subpixel_8t_ssse3.asm
+vpx_dsp/x86/vpx_subpixel_bilinear_sse2.asm
+vpx_dsp/x86/vpx_subpixel_bilinear_ssse3.asm
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/bitops.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/emms.asm
+vpx_ports/mem.h
+vpx_ports/mem_ops.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/msvc.h
+vpx_ports/system_state.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_ports/x86.h
+vpx_ports/x86_abi_support.asm
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx_util/endian_inl.h
+vpx_util/vpx_thread.c
+vpx_util/vpx_thread.h
+vpx_util/vpx_util.mk
diff --git a/libvpx/config/x86_64/vp8_rtcd.h b/libvpx/config/x86_64/vp8_rtcd.h
new file mode 100644
index 0000000..cb3d248
--- /dev/null
+++ b/libvpx/config/x86_64/vp8_rtcd.h
@@ -0,0 +1,276 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_ssse3
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_mmx
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_mmx
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_ssse3
+
+void vp8_blend_b_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp8_blend_b vp8_blend_b_c
+
+void vp8_blend_mb_inner_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp8_blend_mb_inner vp8_blend_mb_inner_c
+
+void vp8_blend_mb_outer_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp8_blend_mb_outer vp8_blend_mb_outer_c
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+int vp8_block_error_mmx(short *coeff, short *dqcoeff);
+int vp8_block_error_xmm(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_xmm
+
+void vp8_clear_system_state_c();
+void vpx_reset_mmx_state();
+#define vp8_clear_system_state vpx_reset_mmx_state
+
+void vp8_copy32xn_c(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
+void vp8_copy32xn_sse2(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
+void vp8_copy32xn_sse3(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
+#define vp8_copy32xn vp8_copy32xn_sse3
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_sse2(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_sse2
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x4_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_mmx
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x8_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_mmx
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+void vp8_dc_only_idct_add_mmx(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_mmx
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter vp8_denoiser_filter_sse2
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
+#define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_sse2
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+void vp8_dequant_idct_add_mmx(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_mmx
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_mmx(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_sse2(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_sse2
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_mmx(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_sse2(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_sse2
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+void vp8_dequantize_b_mmx(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_mmx
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+int vp8_diamond_search_sadx4(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sadx4
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+void vp8_fast_quantize_b_sse2(struct block *, struct blockd *);
+void vp8_fast_quantize_b_ssse3(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_ssse3
+
+void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+void vp8_filter_by_weight16x16_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+#define vp8_filter_by_weight16x16 vp8_filter_by_weight16x16_sse2
+
+void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+#define vp8_filter_by_weight4x4 vp8_filter_by_weight4x4_c
+
+void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+void vp8_filter_by_weight8x8_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+#define vp8_filter_by_weight8x8 vp8_filter_by_weight8x8_sse2
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+int vp8_full_search_sadx3(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sadx3
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_sse2
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_sse2
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_sse2
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_sse2
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_sse2
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_sse2
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_horizontal_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_horizontal_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_sse2
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_vertical_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_vertical_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_sse2
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+int vp8_mbblock_error_mmx(struct macroblock *mb, int dc);
+int vp8_mbblock_error_xmm(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_xmm
+
+void vp8_mbpost_proc_across_ip_c(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+void vp8_mbpost_proc_across_ip_xmm(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+#define vp8_mbpost_proc_across_ip vp8_mbpost_proc_across_ip_xmm
+
+void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+void vp8_mbpost_proc_down_mmx(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+void vp8_mbpost_proc_down_xmm(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+#define vp8_mbpost_proc_down vp8_mbpost_proc_down_xmm
+
+int vp8_mbuverror_c(struct macroblock *mb);
+int vp8_mbuverror_mmx(struct macroblock *mb);
+int vp8_mbuverror_xmm(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_xmm
+
+void vp8_post_proc_down_and_across_mb_row_c(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size);
+void vp8_post_proc_down_and_across_mb_row_sse2(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size);
+#define vp8_post_proc_down_and_across_mb_row vp8_post_proc_down_and_across_mb_row_sse2
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+int vp8_refining_search_sadx4(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sadx4
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+void vp8_regular_quantize_b_sse2(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_sse2
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_mmx(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_sse2(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_sse2
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_mmx(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_sse2(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_sse2
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+void vp8_short_idct4x4llm_mmx(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_mmx
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+void vp8_short_inv_walsh4x4_mmx(short *input, short *output);
+void vp8_short_inv_walsh4x4_sse2(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_sse2
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_sse2
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_ssse3
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict4x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_ssse3
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_ssse3
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_ssse3
+
+void vp8_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/x86_64/vp9_rtcd.h b/libvpx/config/x86_64/vp9_rtcd.h
new file mode 100644
index 0000000..12f3549
--- /dev/null
+++ b/libvpx/config/x86_64/vp9_rtcd.h
@@ -0,0 +1,114 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t vp9_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_sse2
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+int64_t vp9_block_error_fp_sse2(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
+#define vp9_block_error_fp vp9_block_error_fp_sse2
+
+int vp9_diamond_search_sad_c(const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv);
+#define vp9_diamond_search_sad vp9_diamond_search_sad_c
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_fdct8x8_quant_sse2(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_fdct8x8_quant vp9_fdct8x8_quant_ssse3
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht16x16 vp9_fht16x16_sse2
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht4x4 vp9_fht4x4_sse2
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void vp9_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+#define vp9_fht8x8 vp9_fht8x8_sse2
+
+int vp9_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+int vp9_full_search_sadx3(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
+#define vp9_full_search_sad vp9_full_search_sadx3
+
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void vp9_fwht4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_sse2
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+void vp9_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_sse2
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_sse2
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_sse2
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_fp_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp vp9_quantize_fp_ssse3
+
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_fp_32x32_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_fp_32x32 vp9_quantize_fp_32x32_ssse3
+
+void vp9_scale_and_extend_frame_c(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+void vp9_scale_and_extend_frame_ssse3(const struct yv12_buffer_config *src, struct yv12_buffer_config *dst);
+#define vp9_scale_and_extend_frame vp9_scale_and_extend_frame_ssse3
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_sse2(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_sse2
+
+void vp9_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/x86_64/vpx_config.asm b/libvpx/config/x86_64/vpx_config.asm
new file mode 100644
index 0000000..6adb207
--- /dev/null
+++ b/libvpx/config/x86_64/vpx_config.asm
@@ -0,0 +1,82 @@
+%define ARCH_ARM 0
+%define ARCH_MIPS 0
+%define ARCH_X86 0
+%define ARCH_X86_64 1
+%define HAVE_EDSP 0
+%define HAVE_MEDIA 0
+%define HAVE_NEON 0
+%define HAVE_NEON_ASM 0
+%define HAVE_MIPS32 0
+%define HAVE_DSPR2 0
+%define HAVE_MSA 0
+%define HAVE_MIPS64 0
+%define HAVE_MMX 1
+%define HAVE_SSE 1
+%define HAVE_SSE2 1
+%define HAVE_SSE3 1
+%define HAVE_SSSE3 1
+%define HAVE_SSE4_1 0
+%define HAVE_AVX 0
+%define HAVE_AVX2 0
+%define HAVE_VPX_PORTS 1
+%define HAVE_PTHREAD_H 1
+%define HAVE_UNISTD_H 1
+%define CONFIG_DEPENDENCY_TRACKING 1
+%define CONFIG_EXTERNAL_BUILD 1
+%define CONFIG_INSTALL_DOCS 0
+%define CONFIG_INSTALL_BINS 1
+%define CONFIG_INSTALL_LIBS 1
+%define CONFIG_INSTALL_SRCS 0
+%define CONFIG_USE_X86INC 1
+%define CONFIG_DEBUG 0
+%define CONFIG_GPROF 0
+%define CONFIG_GCOV 0
+%define CONFIG_RVCT 0
+%define CONFIG_GCC 1
+%define CONFIG_MSVS 0
+%define CONFIG_PIC 1
+%define CONFIG_BIG_ENDIAN 0
+%define CONFIG_CODEC_SRCS 0
+%define CONFIG_DEBUG_LIBS 0
+%define CONFIG_DEQUANT_TOKENS 0
+%define CONFIG_DC_RECON 0
+%define CONFIG_RUNTIME_CPU_DETECT 0
+%define CONFIG_POSTPROC 1
+%define CONFIG_VP9_POSTPROC 0
+%define CONFIG_MULTITHREAD 1
+%define CONFIG_INTERNAL_STATS 0
+%define CONFIG_VP8_ENCODER 1
+%define CONFIG_VP8_DECODER 1
+%define CONFIG_VP9_ENCODER 1
+%define CONFIG_VP9_DECODER 1
+%define CONFIG_VP8 1
+%define CONFIG_VP9 1
+%define CONFIG_ENCODERS 1
+%define CONFIG_DECODERS 1
+%define CONFIG_STATIC_MSVCRT 0
+%define CONFIG_SPATIAL_RESAMPLING 1
+%define CONFIG_REALTIME_ONLY 1
+%define CONFIG_ONTHEFLY_BITPACKING 0
+%define CONFIG_ERROR_CONCEALMENT 0
+%define CONFIG_SHARED 0
+%define CONFIG_STATIC 1
+%define CONFIG_SMALL 0
+%define CONFIG_POSTPROC_VISUALIZER 0
+%define CONFIG_OS_SUPPORT 1
+%define CONFIG_UNIT_TESTS 1
+%define CONFIG_WEBM_IO 1
+%define CONFIG_LIBYUV 1
+%define CONFIG_DECODE_PERF_TESTS 0
+%define CONFIG_ENCODE_PERF_TESTS 0
+%define CONFIG_MULTI_RES_ENCODING 0
+%define CONFIG_TEMPORAL_DENOISING 1
+%define CONFIG_VP9_TEMPORAL_DENOISING 0
+%define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+%define CONFIG_VP9_HIGHBITDEPTH 0
+%define CONFIG_BETTER_HW_COMPATIBILITY 0
+%define CONFIG_EXPERIMENTAL 0
+%define CONFIG_SIZE_LIMIT 0
+%define CONFIG_SPATIAL_SVC 0
+%define CONFIG_FP_MB_STATS 0
+%define CONFIG_EMULATE_HARDWARE 0
+%define CONFIG_MISC_FIXES 0
diff --git a/libvpx/config/x86_64/vpx_config.c b/libvpx/config/x86_64/vpx_config.c
new file mode 100644
index 0000000..2561717
--- /dev/null
+++ b/libvpx/config/x86_64/vpx_config.c
@@ -0,0 +1,10 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "--target=x86_64-linux-gcc --disable-sse4_1 --disable-avx --disable-avx2 --as=yasm --enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/libvpx/config/x86_64/vpx_config.h b/libvpx/config/x86_64/vpx_config.h
new file mode 100644
index 0000000..93bec08
--- /dev/null
+++ b/libvpx/config/x86_64/vpx_config.h
@@ -0,0 +1,95 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      inline
+#define ARCH_ARM 0
+#define ARCH_MIPS 0
+#define ARCH_X86 0
+#define ARCH_X86_64 1
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_NEON_ASM 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MSA 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 1
+#define HAVE_SSE 1
+#define HAVE_SSE2 1
+#define HAVE_SSE3 1
+#define HAVE_SSSE3 1
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_PTHREAD_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_DEPENDENCY_TRACKING 1
+#define CONFIG_EXTERNAL_BUILD 1
+#define CONFIG_INSTALL_DOCS 0
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 1
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_DEQUANT_TOKENS 0
+#define CONFIG_DC_RECON 0
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 1
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 1
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 1
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 1
+#define CONFIG_WEBM_IO 1
+#define CONFIG_LIBYUV 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_VP9_TEMPORAL_DENOISING 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_VP9_HIGHBITDEPTH 0
+#define CONFIG_BETTER_HW_COMPATIBILITY 0
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_SVC 0
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_EMULATE_HARDWARE 0
+#define CONFIG_MISC_FIXES 0
+#endif /* VPX_CONFIG_H */
diff --git a/libvpx/config/x86_64/vpx_dsp_rtcd.h b/libvpx/config/x86_64/vpx_dsp_rtcd.h
new file mode 100644
index 0000000..4fa1aff
--- /dev/null
+++ b/libvpx/config/x86_64/vpx_dsp_rtcd.h
@@ -0,0 +1,979 @@
+#ifndef VPX_DSP_RTCD_H_
+#define VPX_DSP_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vpx_avg_4x4_c(const uint8_t *, int p);
+unsigned int vpx_avg_4x4_sse2(const uint8_t *, int p);
+#define vpx_avg_4x4 vpx_avg_4x4_sse2
+
+unsigned int vpx_avg_8x8_c(const uint8_t *, int p);
+unsigned int vpx_avg_8x8_sse2(const uint8_t *, int p);
+#define vpx_avg_8x8 vpx_avg_8x8_sse2
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+#define vpx_comp_avg_pred vpx_comp_avg_pred_c
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8 vpx_convolve8_ssse3
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg vpx_convolve8_avg_ssse3
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_horiz vpx_convolve8_avg_horiz_ssse3
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_avg_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_avg_vert vpx_convolve8_avg_vert_ssse3
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_horiz vpx_convolve8_horiz_ssse3
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve8_vert vpx_convolve8_vert_ssse3
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_avg vpx_convolve_avg_sse2
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_convolve_copy vpx_convolve_copy_sse2
+
+void vpx_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_16x16 vpx_d117_predictor_16x16_c
+
+void vpx_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_32x32 vpx_d117_predictor_32x32_c
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_4x4 vpx_d117_predictor_4x4_c
+
+void vpx_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d117_predictor_8x8 vpx_d117_predictor_8x8_c
+
+void vpx_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_16x16 vpx_d135_predictor_16x16_c
+
+void vpx_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_32x32 vpx_d135_predictor_32x32_c
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_4x4 vpx_d135_predictor_4x4_c
+
+void vpx_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d135_predictor_8x8 vpx_d135_predictor_8x8_c
+
+void vpx_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d153_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_16x16 vpx_d153_predictor_16x16_ssse3
+
+void vpx_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d153_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_32x32 vpx_d153_predictor_32x32_ssse3
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d153_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_4x4 vpx_d153_predictor_4x4_ssse3
+
+void vpx_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d153_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d153_predictor_8x8 vpx_d153_predictor_8x8_ssse3
+
+void vpx_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d207_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_16x16 vpx_d207_predictor_16x16_ssse3
+
+void vpx_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d207_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_32x32 vpx_d207_predictor_32x32_ssse3
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d207_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_4x4 vpx_d207_predictor_4x4_sse2
+
+void vpx_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d207_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207_predictor_8x8 vpx_d207_predictor_8x8_ssse3
+
+void vpx_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_16x16 vpx_d207e_predictor_16x16_c
+
+void vpx_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_32x32 vpx_d207e_predictor_32x32_c
+
+void vpx_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_4x4 vpx_d207e_predictor_4x4_c
+
+void vpx_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d207e_predictor_8x8 vpx_d207e_predictor_8x8_c
+
+void vpx_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_16x16 vpx_d45_predictor_16x16_ssse3
+
+void vpx_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_32x32 vpx_d45_predictor_32x32_ssse3
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_4x4 vpx_d45_predictor_4x4_sse2
+
+void vpx_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d45_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45_predictor_8x8 vpx_d45_predictor_8x8_sse2
+
+void vpx_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_16x16 vpx_d45e_predictor_16x16_c
+
+void vpx_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_32x32 vpx_d45e_predictor_32x32_c
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_4x4 vpx_d45e_predictor_4x4_c
+
+void vpx_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d45e_predictor_8x8 vpx_d45e_predictor_8x8_c
+
+void vpx_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d63_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_16x16 vpx_d63_predictor_16x16_ssse3
+
+void vpx_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d63_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_32x32 vpx_d63_predictor_32x32_ssse3
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d63_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_4x4 vpx_d63_predictor_4x4_ssse3
+
+void vpx_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_d63_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63_predictor_8x8 vpx_d63_predictor_8x8_ssse3
+
+void vpx_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_16x16 vpx_d63e_predictor_16x16_c
+
+void vpx_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_32x32 vpx_d63e_predictor_32x32_c
+
+void vpx_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_4x4 vpx_d63e_predictor_4x4_c
+
+void vpx_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63e_predictor_8x8 vpx_d63e_predictor_8x8_c
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_d63f_predictor_4x4 vpx_d63f_predictor_4x4_c
+
+void vpx_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_16x16 vpx_dc_128_predictor_16x16_sse2
+
+void vpx_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_32x32 vpx_dc_128_predictor_32x32_sse2
+
+void vpx_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_4x4 vpx_dc_128_predictor_4x4_sse2
+
+void vpx_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_128_predictor_8x8 vpx_dc_128_predictor_8x8_sse2
+
+void vpx_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_16x16 vpx_dc_left_predictor_16x16_sse2
+
+void vpx_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_32x32 vpx_dc_left_predictor_32x32_sse2
+
+void vpx_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_4x4 vpx_dc_left_predictor_4x4_sse2
+
+void vpx_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_left_predictor_8x8 vpx_dc_left_predictor_8x8_sse2
+
+void vpx_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_16x16 vpx_dc_predictor_16x16_sse2
+
+void vpx_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_32x32 vpx_dc_predictor_32x32_sse2
+
+void vpx_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_4x4 vpx_dc_predictor_4x4_sse2
+
+void vpx_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_predictor_8x8 vpx_dc_predictor_8x8_sse2
+
+void vpx_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_16x16 vpx_dc_top_predictor_16x16_sse2
+
+void vpx_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_32x32 vpx_dc_top_predictor_32x32_sse2
+
+void vpx_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_4x4 vpx_dc_top_predictor_4x4_sse2
+
+void vpx_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_dc_top_predictor_8x8 vpx_dc_top_predictor_8x8_sse2
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16 vpx_fdct16x16_sse2
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct16x16_1 vpx_fdct16x16_1_sse2
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32 vpx_fdct32x32_sse2
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_1 vpx_fdct32x32_1_sse2
+
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct32x32_rd_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct32x32_rd vpx_fdct32x32_rd_sse2
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4 vpx_fdct4x4_sse2
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct4x4_1 vpx_fdct4x4_1_sse2
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_ssse3(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8 vpx_fdct8x8_ssse3
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
+void vpx_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride);
+#define vpx_fdct8x8_1 vpx_fdct8x8_1_sse2
+
+void vpx_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get16x16var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get16x16var vpx_get16x16var_sse2
+
+unsigned int vpx_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vpx_get4x4sse_cs vpx_get4x4sse_cs_c
+
+void vpx_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vpx_get8x8var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vpx_get8x8var vpx_get8x8var_sse2
+
+unsigned int vpx_get_mb_ss_c(const int16_t *);
+unsigned int vpx_get_mb_ss_sse2(const int16_t *);
+#define vpx_get_mb_ss vpx_get_mb_ss_sse2
+
+void vpx_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_16x16 vpx_h_predictor_16x16_sse2
+
+void vpx_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_32x32 vpx_h_predictor_32x32_sse2
+
+void vpx_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_4x4 vpx_h_predictor_4x4_sse2
+
+void vpx_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_h_predictor_8x8 vpx_h_predictor_8x8_sse2
+
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void vpx_hadamard_16x16_sse2(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_16x16 vpx_hadamard_16x16_sse2
+
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void vpx_hadamard_8x8_sse2(const int16_t *src_diff, int src_stride, int16_t *coeff);
+void vpx_hadamard_8x8_ssse3(const int16_t *src_diff, int src_stride, int16_t *coeff);
+#define vpx_hadamard_8x8 vpx_hadamard_8x8_ssse3
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_he_predictor_4x4 vpx_he_predictor_4x4_c
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_10_add vpx_idct16x16_10_add_sse2
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_1_add vpx_idct16x16_1_add_sse2
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct16x16_256_add vpx_idct16x16_256_add_sse2
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1024_add vpx_idct32x32_1024_add_ssse3
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_135_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_135_add vpx_idct32x32_135_add_ssse3
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_1_add vpx_idct32x32_1_add_sse2
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct32x32_34_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct32x32_34_add vpx_idct32x32_34_add_ssse3
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_16_add vpx_idct4x4_16_add_sse2
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct4x4_1_add vpx_idct4x4_1_add_sse2
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_12_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_12_add vpx_idct8x8_12_add_ssse3
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_1_add vpx_idct8x8_1_add_sse2
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_idct8x8_64_add_ssse3(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_idct8x8_64_add vpx_idct8x8_64_add_ssse3
+
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width);
+int16_t vpx_int_pro_col_sse2(const uint8_t *ref, const int width);
+#define vpx_int_pro_col vpx_int_pro_col_sse2
+
+void vpx_int_pro_row_c(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+void vpx_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height);
+#define vpx_int_pro_row vpx_int_pro_row_sse2
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+void vpx_iwht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_16_add vpx_iwht4x4_16_add_sse2
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
+#define vpx_iwht4x4_1_add vpx_iwht4x4_1_add_c
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_4 vpx_lpf_horizontal_4_sse2
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_4_dual vpx_lpf_horizontal_4_dual_sse2
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_8 vpx_lpf_horizontal_8_sse2
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_horizontal_8_dual vpx_lpf_horizontal_8_dual_sse2
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_16 vpx_lpf_horizontal_edge_16_sse2
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_horizontal_edge_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_horizontal_edge_8 vpx_lpf_horizontal_edge_8_sse2
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16 vpx_lpf_vertical_16_sse2
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_16_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_16_dual vpx_lpf_vertical_16_dual_sse2
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_4 vpx_lpf_vertical_4_sse2
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_4_dual vpx_lpf_vertical_4_dual_sse2
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vpx_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vpx_lpf_vertical_8 vpx_lpf_vertical_8_sse2
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vpx_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vpx_lpf_vertical_8_dual vpx_lpf_vertical_8_dual_sse2
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+void vpx_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
+#define vpx_minmax_8x8 vpx_minmax_8x8_sse2
+
+unsigned int vpx_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x16 vpx_mse16x16_sse2
+
+unsigned int vpx_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse16x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse16x8 vpx_mse16x8_sse2
+
+unsigned int vpx_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse8x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x16 vpx_mse8x16_sse2
+
+unsigned int vpx_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vpx_mse8x8_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vpx_mse8x8 vpx_mse8x8_sse2
+
+void vpx_plane_add_noise_c(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
+void vpx_plane_add_noise_sse2(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
+#define vpx_plane_add_noise vpx_plane_add_noise_sse2
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vpx_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vpx_quantize_b_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b vpx_quantize_b_ssse3
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vpx_quantize_b_32x32_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vpx_quantize_b_32x32 vpx_quantize_b_32x32_ssse3
+
+unsigned int vpx_sad16x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x16 vpx_sad16x16_sse2
+
+unsigned int vpx_sad16x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x16_avg vpx_sad16x16_avg_sse2
+
+void vpx_sad16x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x3_ssse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x3 vpx_sad16x16x3_ssse3
+
+void vpx_sad16x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x4d vpx_sad16x16x4d_sse2
+
+void vpx_sad16x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x16x8 vpx_sad16x16x8_c
+
+unsigned int vpx_sad16x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x32 vpx_sad16x32_sse2
+
+unsigned int vpx_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x32_avg vpx_sad16x32_avg_sse2
+
+void vpx_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x32x4d vpx_sad16x32x4d_sse2
+
+unsigned int vpx_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad16x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad16x8 vpx_sad16x8_sse2
+
+unsigned int vpx_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad16x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad16x8_avg vpx_sad16x8_avg_sse2
+
+void vpx_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x3_ssse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x3 vpx_sad16x8x3_ssse3
+
+void vpx_sad16x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad16x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x4d vpx_sad16x8x4d_sse2
+
+void vpx_sad16x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad16x8x8 vpx_sad16x8x8_c
+
+unsigned int vpx_sad32x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x16 vpx_sad32x16_sse2
+
+unsigned int vpx_sad32x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x16_avg vpx_sad32x16_avg_sse2
+
+void vpx_sad32x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x16x4d vpx_sad32x16x4d_sse2
+
+unsigned int vpx_sad32x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x32 vpx_sad32x32_sse2
+
+unsigned int vpx_sad32x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x32_avg vpx_sad32x32_avg_sse2
+
+void vpx_sad32x32x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x3 vpx_sad32x32x3_c
+
+void vpx_sad32x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x4d vpx_sad32x32x4d_sse2
+
+void vpx_sad32x32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x32x8 vpx_sad32x32x8_c
+
+unsigned int vpx_sad32x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad32x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad32x64 vpx_sad32x64_sse2
+
+unsigned int vpx_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad32x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad32x64_avg vpx_sad32x64_avg_sse2
+
+void vpx_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad32x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad32x64x4d vpx_sad32x64x4d_sse2
+
+unsigned int vpx_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad4x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x4 vpx_sad4x4_sse2
+
+unsigned int vpx_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad4x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x4_avg vpx_sad4x4_avg_sse2
+
+void vpx_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad4x4x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x3 vpx_sad4x4x3_sse3
+
+void vpx_sad4x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad4x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x4d vpx_sad4x4x4d_sse2
+
+void vpx_sad4x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x4x8 vpx_sad4x4x8_c
+
+unsigned int vpx_sad4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad4x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad4x8 vpx_sad4x8_sse2
+
+unsigned int vpx_sad4x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad4x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad4x8_avg vpx_sad4x8_avg_sse2
+
+void vpx_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad4x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x4d vpx_sad4x8x4d_sse2
+
+void vpx_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad4x8x8 vpx_sad4x8x8_c
+
+unsigned int vpx_sad64x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad64x32_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x32 vpx_sad64x32_sse2
+
+unsigned int vpx_sad64x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad64x32_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x32_avg vpx_sad64x32_avg_sse2
+
+void vpx_sad64x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad64x32x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x32x4d vpx_sad64x32x4d_sse2
+
+unsigned int vpx_sad64x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad64x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad64x64 vpx_sad64x64_sse2
+
+unsigned int vpx_sad64x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad64x64_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad64x64_avg vpx_sad64x64_avg_sse2
+
+void vpx_sad64x64x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x3 vpx_sad64x64x3_c
+
+void vpx_sad64x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad64x64x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x4d vpx_sad64x64x4d_sse2
+
+void vpx_sad64x64x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad64x64x8 vpx_sad64x64x8_c
+
+unsigned int vpx_sad8x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x16_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x16 vpx_sad8x16_sse2
+
+unsigned int vpx_sad8x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x16_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x16_avg vpx_sad8x16_avg_sse2
+
+void vpx_sad8x16x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x16x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x3 vpx_sad8x16x3_sse3
+
+void vpx_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x16x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x4d vpx_sad8x16x4d_sse2
+
+void vpx_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x16x8 vpx_sad8x16x8_c
+
+unsigned int vpx_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x4_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x4 vpx_sad8x4_sse2
+
+unsigned int vpx_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x4_avg vpx_sad8x4_avg_sse2
+
+void vpx_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x4d vpx_sad8x4x4d_sse2
+
+void vpx_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x4x8 vpx_sad8x4x8_c
+
+unsigned int vpx_sad8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vpx_sad8x8_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define vpx_sad8x8 vpx_sad8x8_sse2
+
+unsigned int vpx_sad8x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+unsigned int vpx_sad8x8_avg_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define vpx_sad8x8_avg vpx_sad8x8_avg_sse2
+
+void vpx_sad8x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+void vpx_sad8x8x3_sse3(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x3 vpx_sad8x8x3_sse3
+
+void vpx_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+void vpx_sad8x8x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x4d vpx_sad8x8x4d_sse2
+
+void vpx_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vpx_sad8x8x8 vpx_sad8x8x8_c
+
+int vpx_satd_c(const int16_t *coeff, int length);
+int vpx_satd_sse2(const int16_t *coeff, int length);
+#define vpx_satd vpx_satd_sse2
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void vpx_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_2d vpx_scaled_2d_ssse3
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_2d vpx_scaled_avg_2d_c
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_horiz vpx_scaled_avg_horiz_c
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_avg_vert vpx_scaled_avg_vert_c
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_horiz vpx_scaled_horiz_c
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define vpx_scaled_vert vpx_scaled_vert_c
+
+uint32_t vpx_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x16 vpx_sub_pixel_avg_variance16x16_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x32 vpx_sub_pixel_avg_variance16x32_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance16x8 vpx_sub_pixel_avg_variance16x8_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x16 vpx_sub_pixel_avg_variance32x16_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x32 vpx_sub_pixel_avg_variance32x32_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance32x64 vpx_sub_pixel_avg_variance32x64_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x4 vpx_sub_pixel_avg_variance4x4_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance4x8 vpx_sub_pixel_avg_variance4x8_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x32 vpx_sub_pixel_avg_variance64x32_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance64x64 vpx_sub_pixel_avg_variance64x64_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x16 vpx_sub_pixel_avg_variance8x16_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x4 vpx_sub_pixel_avg_variance8x4_ssse3
+
+uint32_t vpx_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+uint32_t vpx_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define vpx_sub_pixel_avg_variance8x8 vpx_sub_pixel_avg_variance8x8_ssse3
+
+uint32_t vpx_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x16 vpx_sub_pixel_variance16x16_ssse3
+
+uint32_t vpx_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x32 vpx_sub_pixel_variance16x32_ssse3
+
+uint32_t vpx_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance16x8 vpx_sub_pixel_variance16x8_ssse3
+
+uint32_t vpx_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x16 vpx_sub_pixel_variance32x16_ssse3
+
+uint32_t vpx_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x32 vpx_sub_pixel_variance32x32_ssse3
+
+uint32_t vpx_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance32x64 vpx_sub_pixel_variance32x64_ssse3
+
+uint32_t vpx_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x4 vpx_sub_pixel_variance4x4_ssse3
+
+uint32_t vpx_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance4x8 vpx_sub_pixel_variance4x8_ssse3
+
+uint32_t vpx_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x32 vpx_sub_pixel_variance64x32_ssse3
+
+uint32_t vpx_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance64x64 vpx_sub_pixel_variance64x64_ssse3
+
+uint32_t vpx_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x16 vpx_sub_pixel_variance8x16_ssse3
+
+uint32_t vpx_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x4 vpx_sub_pixel_variance8x4_ssse3
+
+uint32_t vpx_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+uint32_t vpx_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define vpx_sub_pixel_variance8x8 vpx_sub_pixel_variance8x8_ssse3
+
+void vpx_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+void vpx_subtract_block_sse2(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+#define vpx_subtract_block vpx_subtract_block_sse2
+
+void vpx_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_16x16 vpx_tm_predictor_16x16_sse2
+
+void vpx_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_32x32 vpx_tm_predictor_32x32_sse2
+
+void vpx_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_4x4 vpx_tm_predictor_4x4_sse2
+
+void vpx_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_tm_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_tm_predictor_8x8 vpx_tm_predictor_8x8_sse2
+
+void vpx_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_16x16 vpx_v_predictor_16x16_sse2
+
+void vpx_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_32x32 vpx_v_predictor_32x32_sse2
+
+void vpx_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_4x4 vpx_v_predictor_4x4_sse2
+
+void vpx_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vpx_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_v_predictor_8x8 vpx_v_predictor_8x8_sse2
+
+unsigned int vpx_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x16 vpx_variance16x16_sse2
+
+unsigned int vpx_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x32 vpx_variance16x32_sse2
+
+unsigned int vpx_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance16x8 vpx_variance16x8_sse2
+
+unsigned int vpx_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x16 vpx_variance32x16_sse2
+
+unsigned int vpx_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x32 vpx_variance32x32_sse2
+
+unsigned int vpx_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance32x64 vpx_variance32x64_sse2
+
+unsigned int vpx_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x4 vpx_variance4x4_sse2
+
+unsigned int vpx_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance4x8 vpx_variance4x8_sse2
+
+unsigned int vpx_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x32 vpx_variance64x32_sse2
+
+unsigned int vpx_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance64x64 vpx_variance64x64_sse2
+
+unsigned int vpx_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x16 vpx_variance8x16_sse2
+
+unsigned int vpx_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x4 vpx_variance8x4_sse2
+
+unsigned int vpx_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vpx_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vpx_variance8x8 vpx_variance8x8_sse2
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_h_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_h vpx_variance_halfpixvar16x16_h_sse2
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_hv_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_hv vpx_variance_halfpixvar16x16_hv_sse2
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+uint32_t vpx_variance_halfpixvar16x16_v_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
+#define vpx_variance_halfpixvar16x16_v vpx_variance_halfpixvar16x16_v_sse2
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vpx_ve_predictor_4x4 vpx_ve_predictor_4x4_c
+
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src, const int bwl);
+int vpx_vector_var_sse2(const int16_t *ref, const int16_t *src, const int bwl);
+#define vpx_vector_var vpx_vector_var_sse2
+
+void vpx_dsp_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/x86_64/vpx_scale_rtcd.h b/libvpx/config/x86_64/vpx_scale_rtcd.h
new file mode 100644
index 0000000..ddf7d01
--- /dev/null
+++ b/libvpx/config/x86_64/vpx_scale_rtcd.h
@@ -0,0 +1,69 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vpx_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_borders vpx_extend_frame_borders_c
+
+void vpx_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vpx_extend_frame_inner_borders vpx_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/libvpx/config/x86_64/vpx_version.h b/libvpx/config/x86_64/vpx_version.h
new file mode 100644
index 0000000..5cff3b4
--- /dev/null
+++ b/libvpx/config/x86_64/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  6
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.6.0"
+#define VERSION_STRING      " v1.6.0"
diff --git a/libvpx/generate_config.sh b/libvpx/generate_config.sh
new file mode 100755
index 0000000..c9a57cc
--- /dev/null
+++ b/libvpx/generate_config.sh
@@ -0,0 +1,272 @@
+#!/bin/bash -e
+#
+# Copyright (c) 2012 The Chromium Authors. All rights reserved.
+# Use of this source code is governed by a BSD-style license that can be
+# found in the LICENSE file.
+
+# This script is used to generate files in the <platform> directories needed to
+# build libvpx. Every time libvpx source code is updated run this script.
+#
+# For example:
+# $ ./generate_config.sh
+#
+# And this will update all the config files needed.
+
+export LC_ALL=C
+BASE_DIR=$(pwd)
+LIBVPX_SRC_DIR="libvpx"
+LIBVPX_CONFIG_DIR="config"
+
+# Clean files from previous make.
+function make_clean {
+  make clean > /dev/null
+  rm -f libvpx_srcs.txt
+}
+
+# Lint a pair of vpx_config.h and vpx_config.asm to make sure they match.
+# $1 - Header file directory.
+function lint_config {
+  # mips does not contain any assembly so the header does not need to be
+  # compared to the asm.
+  if [[ "$1" != *mips* ]]; then
+    $BASE_DIR/lint_config.sh \
+      -h $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vpx_config.h \
+      -a $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vpx_config.asm
+  fi
+}
+
+# Print the configuration.
+# $1 - Header file directory.
+function print_config {
+  $BASE_DIR/lint_config.sh -p \
+    -h $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vpx_config.h \
+    -a $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vpx_config.asm
+}
+
+# Print the configuration from Header file.
+# This function is an abridged version of print_config which does not use
+# lint_config and it does not require existence of vpx_config.asm.
+# $1 - Header file directory.
+function print_config_basic {
+  combined_config="$(cat $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vpx_config.h \
+                   | grep -E ' +[01] *$')"
+  combined_config="$(echo "$combined_config" | grep -v DO1STROUNDING)"
+  combined_config="$(echo "$combined_config" | sed 's/[ \t]//g')"
+  combined_config="$(echo "$combined_config" | sed 's/.*define//')"
+  combined_config="$(echo "$combined_config" | sed 's/0$/=no/')"
+  combined_config="$(echo "$combined_config" | sed 's/1$/=yes/')"
+  echo "$combined_config" | sort | uniq
+}
+
+# Generate *_rtcd.h files.
+# $1 - Header file directory.
+# $2 - Architecture.
+function gen_rtcd_header {
+  echo "Generate $LIBVPX_CONFIG_DIR/$1/*_rtcd.h files."
+
+  # We don't properly persist the config options specificed on the configure
+  # line. Until that is fixed, force them here.
+  DISABLE_CONFIG="--disable-sse4_1 --disable-avx --disable-avx2"
+
+  rm -rf $BASE_DIR/$TEMP_DIR/libvpx.config
+  if [[ "$2" == *mips* ]]; then
+    print_config_basic $1 > $BASE_DIR/$TEMP_DIR/libvpx.config
+  else
+    $BASE_DIR/lint_config.sh -p \
+      -h $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vpx_config.h \
+      -a $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vpx_config.asm \
+      -o $BASE_DIR/$TEMP_DIR/libvpx.config
+  fi
+
+  $BASE_DIR/$LIBVPX_SRC_DIR/build/make/rtcd.pl \
+    --arch=$2 \
+    --sym=vp8_rtcd \
+    $DISABLE_CONFIG \
+    --config=$BASE_DIR/$TEMP_DIR/libvpx.config \
+    $BASE_DIR/$LIBVPX_SRC_DIR/vp8/common/rtcd_defs.pl \
+    > $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vp8_rtcd.h
+
+  $BASE_DIR/$LIBVPX_SRC_DIR/build/make/rtcd.pl \
+    --arch=$2 \
+    --sym=vp9_rtcd \
+    --config=$BASE_DIR/$TEMP_DIR/libvpx.config \
+    $DISABLE_CONFIG \
+    $BASE_DIR/$LIBVPX_SRC_DIR/vp9/common/vp9_rtcd_defs.pl \
+    > $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vp9_rtcd.h
+
+  $BASE_DIR/$LIBVPX_SRC_DIR/build/make/rtcd.pl \
+    --arch=$2 \
+    --sym=vpx_scale_rtcd \
+    --config=$BASE_DIR/$TEMP_DIR/libvpx.config \
+    $DISABLE_CONFIG \
+    $BASE_DIR/$LIBVPX_SRC_DIR/vpx_scale/vpx_scale_rtcd.pl \
+    > $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vpx_scale_rtcd.h
+
+  $BASE_DIR/$LIBVPX_SRC_DIR/build/make/rtcd.pl \
+    --arch=$2 \
+    --sym=vpx_dsp_rtcd \
+    --config=$BASE_DIR/$TEMP_DIR/libvpx.config \
+    $DISABLE_CONFIG \
+    $BASE_DIR/$LIBVPX_SRC_DIR/vpx_dsp/vpx_dsp_rtcd_defs.pl \
+    > $BASE_DIR/$LIBVPX_CONFIG_DIR/$1/vpx_dsp_rtcd.h
+
+  rm -rf $BASE_DIR/$TEMP_DIR/libvpx.config
+}
+
+# Generate Config files. "--enable-external-build" must be set to skip
+# detection of capabilities on specific targets.
+# $1 - Header file directory.
+# $2 - Config command line.
+function gen_config_files {
+  ./configure $2 > /dev/null
+
+  # Generate vpx_config.asm. Do not create one for mips.
+  if [[ "$1" != *mips* ]]; then
+    if [[ "$1" == *x86* ]]; then
+      egrep "#define [A-Z0-9_]+ [01]" vpx_config.h \
+        | awk '{print "%define " $2 " " $3}' > vpx_config.asm
+    else
+      egrep "#define [A-Z0-9_]+ [01]" vpx_config.h \
+        | awk '{print $2 " EQU " $3}' \
+        | perl $BASE_DIR/$LIBVPX_SRC_DIR/build/make/ads2gas.pl > vpx_config.asm
+    fi
+  fi
+
+  # Generate vpx_version.h
+  $BASE_DIR/$LIBVPX_SRC_DIR/build/make/version.sh "$BASE_DIR/$LIBVPX_SRC_DIR" vpx_version.h
+
+  cp vpx_config.* vpx_version.h $BASE_DIR/$LIBVPX_CONFIG_DIR/$1
+  make_clean
+  rm -rf vpx_config.* vpx_version.h
+}
+
+echo "Create temporary directory."
+TEMP_DIR="$LIBVPX_SRC_DIR.temp"
+rm -rf $TEMP_DIR
+cp -R $LIBVPX_SRC_DIR $TEMP_DIR
+cd $TEMP_DIR
+
+echo "Generate config files."
+all_platforms="--enable-external-build --enable-realtime-only --enable-pic --disable-runtime-cpu-detect"
+intel="--disable-sse4_1 --disable-avx --disable-avx2 --as=yasm"
+gen_config_files x86 "--target=x86-linux-gcc ${intel} ${all_platforms}"
+gen_config_files x86_64 "--target=x86_64-linux-gcc ${intel} ${all_platforms}"
+gen_config_files arm "--target=armv6-linux-gcc  ${all_platforms}"
+gen_config_files arm-neon "--target=armv7-linux-gcc ${all_platforms}"
+gen_config_files arm64 "--force-target=armv8-linux-gcc ${all_platforms}"
+gen_config_files mips32 "--target=mips32-linux-gcc --disable-dspr2 --disable-msa ${all_platforms}"
+gen_config_files mips32-dspr2 "--target=mips32-linux-gcc --enable-dspr2 ${all_platforms}"
+gen_config_files mips32-msa "--target=mips32-linux-gcc --enable-msa ${all_platforms}"
+gen_config_files mips64 "--target=mips64-linux-gcc --disable-msa ${all_platforms}"
+gen_config_files mips64-msa "--target=mips64-linux-gcc --enable-msa ${all_platforms}"
+gen_config_files generic "--target=generic-gnu ${all_platforms}"
+
+echo "Remove temporary directory."
+cd $BASE_DIR
+rm -rf $TEMP_DIR
+
+echo "Lint libvpx configuration."
+lint_config x86
+lint_config x86_64
+lint_config arm
+lint_config arm-neon
+lint_config arm64
+lint_config mips32
+lint_config mips32-dspr2
+lint_config mips32-msa
+lint_config mips64
+lint_config mips64-msa
+lint_config generic
+
+echo "Create temporary directory."
+TEMP_DIR="$LIBVPX_SRC_DIR.temp"
+rm -rf $TEMP_DIR
+cp -R $LIBVPX_SRC_DIR $TEMP_DIR
+cd $TEMP_DIR
+
+gen_rtcd_header x86 x86
+gen_rtcd_header x86_64 x86_64
+gen_rtcd_header arm armv6
+gen_rtcd_header arm-neon armv7
+gen_rtcd_header arm64 armv8
+gen_rtcd_header mips32 mips32
+gen_rtcd_header mips32-dspr2 mips32
+gen_rtcd_header mips32-msa mips32
+gen_rtcd_header mips64 mips64
+gen_rtcd_header mips64-msa mips64
+gen_rtcd_header generic generic
+
+echo "Prepare Makefile."
+./configure --target=generic-gnu > /dev/null
+make_clean
+
+echo "Generate X86 source list."
+config=$(print_config x86)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/x86/
+
+echo "Generate X86_64 source list."
+config=$(print_config x86_64)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/x86_64/
+
+echo "Generate ARM source list."
+config=$(print_config arm)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/arm/
+
+echo "Generate ARM NEON source list."
+config=$(print_config arm-neon)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/arm-neon/
+
+echo "Generate ARM64 source list."
+config=$(print_config arm64)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/arm64/
+
+echo "Generate MIPS source list."
+config=$(print_config_basic mips32)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/mips32/
+
+echo "Generate MIPS DSPR2 source list."
+config=$(print_config_basic mips32-dspr2)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/mips32-dspr2/
+
+echo "Generate MIPS MSA source list."
+config=$(print_config_basic mips32-msa)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/mips32-msa/
+
+echo "Generate MIPS64 source list."
+config=$(print_config_basic mips64)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/mips64/
+
+echo "Generate MIPS64 MSA source list."
+config=$(print_config_basic mips64-msa)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/mips64-msa/
+
+echo "Generate GENERIC source list."
+config=$(print_config_basic generic)
+make_clean
+make libvpx_srcs.txt target=libs $config > /dev/null
+cp libvpx_srcs.txt $BASE_DIR/$LIBVPX_CONFIG_DIR/generic/
+
+
+echo "Remove temporary directory."
+cd $BASE_DIR
+rm -rf $TEMP_DIR
diff --git a/libvpx/libvpx-asm-translation.mk b/libvpx/libvpx-asm-translation.mk
new file mode 100644
index 0000000..0b4b58d
--- /dev/null
+++ b/libvpx/libvpx-asm-translation.mk
@@ -0,0 +1,26 @@
+# Rules to generate GAS compatible assembly from RVCT syntax files.
+# Input variables:
+#   libvpx_2nd_arch
+#   libvpx_source_dir
+#   libvpx_codec_srcs_asm_<arch>
+#
+# Output variables:
+#   LOCAL_GENERATED_SOURCES_<arch>
+
+ifneq ($(strip $(libvpx_codec_srcs_asm_$(TARGET_$(libvpx_2nd_arch)ARCH))),)
+libvpx_intermediates := $(call local-intermediates-dir,,$(libvpx_2nd_arch))
+# This step is only required for ARM. MIPS uses intrinsics exclusively and x86
+# requires 'yasm' to pre-process its assembly files.
+# The ARM assembly sources must be converted from ADS to GAS compatible format.
+VPX_ASM := $(addprefix $(libvpx_intermediates)/, $(libvpx_codec_srcs_asm_$(TARGET_$(libvpx_2nd_arch)ARCH)))
+# The build system will only accept arm assembly which ends in '.s'
+VPX_GEN := $(addsuffix .s, $(VPX_ASM))
+$(VPX_GEN) : PRIVATE_SOURCE_DIR := $(libvpx_source_dir)
+$(VPX_GEN) : PRIVATE_CUSTOM_TOOL = cat $< | perl $(PRIVATE_SOURCE_DIR)/build/make/ads2gas.pl > $@
+$(VPX_GEN) : $(libvpx_intermediates)/%.s : $(libvpx_source_dir)/% \
+		$(libvpx_source_dir)/build/make/ads2gas.pl \
+		$(libvpx_source_dir)/build/make/thumb.pm
+	$(transform-generated-source)
+
+LOCAL_GENERATED_SOURCES_$(TARGET_$(libvpx_2nd_arch)ARCH) += $(VPX_GEN)
+endif
diff --git a/libvpx/libvpx.mk b/libvpx/libvpx.mk
new file mode 100644
index 0000000..b48d8e7
--- /dev/null
+++ b/libvpx/libvpx.mk
@@ -0,0 +1,45 @@
+LOCAL_PATH := $(call my-dir)
+include $(CLEAR_VARS)
+
+# Clang arm assembler cannot compile libvpx .s files yet.
+LOCAL_CLANG_ASFLAGS_arm += -no-integrated-as
+
+libvpx_source_dir := $(LOCAL_PATH)/libvpx
+
+## Arch-common settings
+LOCAL_MODULE := libvpx
+LOCAL_MODULE_CLASS := STATIC_LIBRARIES
+
+LOCAL_CFLAGS := -DHAVE_CONFIG_H=vpx_config.h
+
+# Want arm, not thumb, optimized
+LOCAL_ARM_MODE := arm
+LOCAL_CFLAGS += -O3
+
+LOCAL_CFLAGS += -Wno-unused-parameter
+
+LOCAL_C_INCLUDES := $(libvpx_source_dir)
+
+# Load the arch-specific settings
+include $(LOCAL_PATH)/config.$(TARGET_ARCH).mk
+LOCAL_SRC_FILES_$(TARGET_ARCH) := $(libvpx_codec_srcs_c_$(TARGET_ARCH))
+LOCAL_C_INCLUDES_$(TARGET_ARCH) := $(libvpx_config_dir_$(TARGET_ARCH))
+libvpx_2nd_arch :=
+include $(LOCAL_PATH)/libvpx-asm-translation.mk
+
+ifdef TARGET_2ND_ARCH
+include $(LOCAL_PATH)/config.$(TARGET_2ND_ARCH).mk
+LOCAL_SRC_FILES_$(TARGET_2ND_ARCH) := $(libvpx_codec_srcs_c_$(TARGET_2ND_ARCH))
+LOCAL_C_INCLUDES_$(TARGET_2ND_ARCH) := $(libvpx_config_dir_$(TARGET_2ND_ARCH))
+libvpx_2nd_arch := $(TARGET_2ND_ARCH_VAR_PREFIX)
+include $(LOCAL_PATH)/libvpx-asm-translation.mk
+libvpx_2nd_arch :=
+endif
+
+libvpx_target :=
+libvpx_source_dir :=
+libvpx_intermediates :=
+libvpx_asm_offsets_intermediates :=
+libvpx_asm_offsets_files :=
+
+include $(BUILD_STATIC_LIBRARY)
diff --git a/libvpx/libvpx/.mailmap b/libvpx/libvpx/.mailmap
new file mode 100644
index 0000000..94cb1ec
--- /dev/null
+++ b/libvpx/libvpx/.mailmap
@@ -0,0 +1,37 @@
+Adrian Grange <agrange@google.com>
+Aℓex Converse <aconverse@google.com>
+Aℓex Converse <aconverse@google.com> <alex.converse@gmail.com>
+Alexis Ballier <aballier@gentoo.org> <alexis.ballier@gmail.com>
+Alpha Lam <hclam@google.com> <hclam@chromium.org>
+Daniele Castagna <dcastagna@chromium.org> <dcastagna@google.com>
+Deb Mukherjee <debargha@google.com>
+Erik Niemeyer <erik.a.niemeyer@intel.com> <erik.a.niemeyer@gmail.com>
+Guillaume Martres <gmartres@google.com> <smarter3@gmail.com>
+Hangyu Kuang <hkuang@google.com>
+Hui Su <huisu@google.com>
+Jacky Chen <jackychen@google.com>
+Jim Bankoski <jimbankoski@google.com>
+Johann Koenig <johannkoenig@google.com>
+Johann Koenig <johannkoenig@google.com> <johann.koenig@duck.com>
+Johann Koenig <johannkoenig@google.com> <johann.koenig@gmail.com>
+Johann Koenig <johannkoenig@google.com> <johannkoenig@chromium.org>
+John Koleszar <jkoleszar@google.com>
+Joshua Litt <joshualitt@google.com> <joshualitt@chromium.org>
+Marco Paniconi <marpan@google.com>
+Marco Paniconi <marpan@google.com> <marpan@chromium.org>
+Pascal Massimino <pascal.massimino@gmail.com>
+Paul Wilkins <paulwilkins@google.com>
+Peter de Rivaz <peter.derivaz@gmail.com>
+Peter de Rivaz <peter.derivaz@gmail.com> <peter.derivaz@argondesign.com>
+Ralph Giles <giles@xiph.org> <giles@entropywave.com>
+Ralph Giles <giles@xiph.org> <giles@mozilla.com>
+Ronald S. Bultje <rsbultje@gmail.com> <rbultje@google.com>
+Sami Pietilä <samipietila@google.com>
+Tamar Levy <tamar.levy@intel.com>
+Tamar Levy <tamar.levy@intel.com> <levytamar82@gmail.com>
+Tero Rintaluoma <teror@google.com> <tero.rintaluoma@on2.com>
+Timothy B. Terriberry <tterribe@xiph.org> <tterriberry@mozilla.com>
+Tom Finegan <tomfinegan@google.com>
+Tom Finegan <tomfinegan@google.com> <tomfinegan@chromium.org>
+Yaowu Xu <yaowu@google.com> <yaowu@xuyaowu.com>
+Yaowu Xu <yaowu@google.com> <Yaowu Xu>
diff --git a/libvpx/libvpx/AUTHORS b/libvpx/libvpx/AUTHORS
new file mode 100644
index 0000000..fcd5c53
--- /dev/null
+++ b/libvpx/libvpx/AUTHORS
@@ -0,0 +1,142 @@
+# This file is automatically generated from the git commit history
+# by tools/gen_authors.sh.
+
+Aaron Watry <awatry@gmail.com>
+Abo Talib Mahfoodh <ab.mahfoodh@gmail.com>
+Adam Xu <adam@xuyaowu.com>
+Adrian Grange <agrange@google.com>
+Aℓex Converse <aconverse@google.com>
+Ahmad Sharif <asharif@google.com>
+Alexander Voronov <avoronov@graphics.cs.msu.ru>
+Alexis Ballier <aballier@gentoo.org>
+Alok Ahuja <waveletcoeff@gmail.com>
+Alpha Lam <hclam@google.com>
+A.Mahfoodh <ab.mahfoodh@gmail.com>
+Ami Fischman <fischman@chromium.org>
+Andoni Morales Alastruey <ylatuya@gmail.com>
+Andres Mejia <mcitadel@gmail.com>
+Andrew Russell <anrussell@google.com>
+Angie Chiang <angiebird@google.com>
+Aron Rosenberg <arosenberg@logitech.com>
+Attila Nagy <attilanagy@google.com>
+Brion Vibber <bvibber@wikimedia.org>
+changjun.yang <changjun.yang@intel.com>
+Charles 'Buck' Krasic <ckrasic@google.com>
+chm <chm@rock-chips.com>
+Christian Duvivier <cduvivier@google.com>
+Daniele Castagna <dcastagna@chromium.org>
+Daniel Kang <ddkang@google.com>
+Deb Mukherjee <debargha@google.com>
+Dim Temp <dimtemp0@gmail.com>
+Dmitry Kovalev <dkovalev@google.com>
+Dragan Mrdjan <dmrdjan@mips.com>
+Ed Baker <edward.baker@intel.com>
+Ehsan Akhgari <ehsan.akhgari@gmail.com>
+Erik Niemeyer <erik.a.niemeyer@intel.com>
+Fabio Pedretti <fabio.ped@libero.it>
+Frank Galligan <fgalligan@google.com>
+Fredrik Söderquist <fs@opera.com>
+Fritz Koenig <frkoenig@google.com>
+Gaute Strokkenes <gaute.strokkenes@broadcom.com>
+Geza Lore <gezalore@gmail.com>
+Ghislain MARY <ghislainmary2@gmail.com>
+Giuseppe Scrivano <gscrivano@gnu.org>
+Gordana Cmiljanovic <gordana.cmiljanovic@imgtec.com>
+Guillaume Martres <gmartres@google.com>
+Guillermo Ballester Valor <gbvalor@gmail.com>
+Hangyu Kuang <hkuang@google.com>
+Hanno Böck <hanno@hboeck.de>
+Henrik Lundin <hlundin@google.com>
+Hui Su <huisu@google.com>
+Ivan Maltz <ivanmaltz@google.com>
+Jacek Caban <cjacek@gmail.com>
+Jacky Chen <jackychen@google.com>
+James Berry <jamesberry@google.com>
+James Yu <james.yu@linaro.org>
+James Zern <jzern@google.com>
+Jan Gerber <j@mailb.org>
+Jan Kratochvil <jan.kratochvil@redhat.com>
+Janne Salonen <jsalonen@google.com>
+Jean-Yves Avenard <jyavenard@mozilla.com>
+Jeff Faust <jfaust@google.com>
+Jeff Muizelaar <jmuizelaar@mozilla.com>
+Jeff Petkau <jpet@chromium.org>
+Jia Jia <jia.jia@linaro.org>
+Jian Zhou <zhoujian@google.com>
+Jim Bankoski <jimbankoski@google.com>
+Jingning Han <jingning@google.com>
+Joey Parrish <joeyparrish@google.com>
+Johann Koenig <johannkoenig@google.com>
+John Koleszar <jkoleszar@google.com>
+Johnny Klonaris <google@jawknee.com>
+John Stark <jhnstrk@gmail.com>
+Joshua Bleecher Snyder <josh@treelinelabs.com>
+Joshua Litt <joshualitt@google.com>
+Julia Robson <juliamrobson@gmail.com>
+Justin Clift <justin@salasaga.org>
+Justin Lebar <justin.lebar@gmail.com>
+KO Myung-Hun <komh@chollian.net>
+Lawrence Velázquez <larryv@macports.org>
+Linfeng Zhang <linfengz@google.com>
+Lou Quillio <louquillio@google.com>
+Luca Barbato <lu_zero@gentoo.org>
+Makoto Kato <makoto.kt@gmail.com>
+Mans Rullgard <mans@mansr.com>
+Marco Paniconi <marpan@google.com>
+Mark Mentovai <mark@chromium.org>
+Martin Ettl <ettl.martin78@googlemail.com>
+Martin Storsjo <martin@martin.st>
+Matthew Heaney <matthewjheaney@chromium.org>
+Michael Kohler <michaelkohler@live.com>
+Mike Frysinger <vapier@chromium.org>
+Mike Hommey <mhommey@mozilla.com>
+Mikhal Shemer <mikhal@google.com>
+Minghai Shang <minghai@google.com>
+Morton Jonuschat <yabawock@gmail.com>
+Nico Weber <thakis@chromium.org>
+Parag Salasakar <img.mips1@gmail.com>
+Pascal Massimino <pascal.massimino@gmail.com>
+Patrik Westin <patrik.westin@gmail.com>
+Paul Wilkins <paulwilkins@google.com>
+Pavol Rusnak <stick@gk2.sk>
+Paweł Hajdan <phajdan@google.com>
+Pengchong Jin <pengchong@google.com>
+Peter de Rivaz <peter.derivaz@gmail.com>
+Philip Jägenstedt <philipj@opera.com>
+Priit Laes <plaes@plaes.org>
+Rafael Ávila de Espíndola <rafael.espindola@gmail.com>
+Rafaël Carré <funman@videolan.org>
+Ralph Giles <giles@xiph.org>
+Rob Bradford <rob@linux.intel.com>
+Ronald S. Bultje <rsbultje@gmail.com>
+Rui Ueyama <ruiu@google.com>
+Sami Pietilä <samipietila@google.com>
+Sasi Inguva <isasi@google.com>
+Scott Graham <scottmg@chromium.org>
+Scott LaVarnway <slavarnway@google.com>
+Sean McGovern <gseanmcg@gmail.com>
+Sergey Kolomenkin <kolomenkin@gmail.com>
+Sergey Ulanov <sergeyu@chromium.org>
+Shimon Doodkin <helpmepro1@gmail.com>
+Shunyao Li <shunyaoli@google.com>
+Stefan Holmer <holmer@google.com>
+Suman Sunkara <sunkaras@google.com>
+Taekhyun Kim <takim@nvidia.com>
+Takanori MATSUURA <t.matsuu@gmail.com>
+Tamar Levy <tamar.levy@intel.com>
+Tao Bai <michaelbai@chromium.org>
+Tero Rintaluoma <teror@google.com>
+Thijs Vermeir <thijsvermeir@gmail.com>
+Tim Kopp <tkopp@google.com>
+Timothy B. Terriberry <tterribe@xiph.org>
+Tom Finegan <tomfinegan@google.com>
+Vignesh Venkatasubramanian <vigneshv@google.com>
+Yaowu Xu <yaowu@google.com>
+Yi Luo <luoyi@google.com>
+Yongzhe Wang <yongzhe@google.com>
+Yunqing Wang <yunqingwang@google.com>
+Yury Gitman <yuryg@google.com>
+Zoe Liu <zoeliu@google.com>
+Google Inc.
+The Mozilla Foundation
+The Xiph.Org Foundation
diff --git a/libvpx/libvpx/CHANGELOG b/libvpx/libvpx/CHANGELOG
new file mode 100644
index 0000000..795d395
--- /dev/null
+++ b/libvpx/libvpx/CHANGELOG
@@ -0,0 +1,654 @@
+2016-07-20 v1.6.0 "Khaki Campbell Duck"
+  This release improves upon the VP9 encoder and speeds up the encoding and
+  decoding processes.
+
+  - Upgrading:
+    This release is ABI incompatible with 1.5.0 due to a new 'color_range' enum
+    in vpx_image and some minor changes to the VP8_COMP structure.
+
+    The default key frame interval for VP9 has changed from 128 to 9999.
+
+  - Enhancement:
+    A core focus has been performance for low end Intel processors. SSSE3
+    instructions such as 'pshufb' have been avoided and instructions have been
+    reordered to better accommodate the more constrained pipelines.
+
+    As a result, devices based on Celeron processors have seen substantial
+    decoding improvements. From Indian Runner Duck to Javan Whistling Duck,
+    decoding speed improved between 10 and 30%. Between Javan Whistling Duck
+    and Khaki Campbell Duck, it improved another 10 to 15%.
+
+    While Celeron benefited most, Core-i5 also improved 5% and 10% between the
+    respective releases.
+
+    Realtime performance for WebRTC for both speed and quality has received a
+    lot of attention.
+
+  - Bug Fixes:
+    A number of fuzzing issues, found variously by Mozilla, Chromium and others,
+    have been fixed and we strongly recommend updating.
+
+2015-11-09 v1.5.0 "Javan Whistling Duck"
+  This release improves upon the VP9 encoder and speeds up the encoding and
+  decoding processes.
+
+  - Upgrading:
+    This release is ABI incompatible with 1.4.0. It drops deprecated VP8
+    controls and adds a variety of VP9 controls for testing.
+
+    The vpxenc utility now prefers VP9 by default.
+
+  - Enhancements:
+    Faster VP9 encoding and decoding
+    Smaller library size by combining functions used by VP8 and VP9
+
+  - Bug Fixes:
+    A variety of fuzzing issues
+
+2015-04-03 v1.4.0 "Indian Runner Duck"
+  This release includes significant improvements to the VP9 codec.
+
+  - Upgrading:
+    This release is ABI incompatible with 1.3.0. It drops the compatibility
+    layer, requiring VPX_IMG_FMT_* instead of IMG_FMT_*, and adds several codec
+    controls for VP9.
+
+  - Enhancements:
+    Faster VP9 encoding and decoding
+    Multithreaded VP9 decoding (tile and frame-based)
+    Multithreaded VP9 encoding - on by default
+    YUV 4:2:2 and 4:4:4 support in VP9
+    10 and 12bit support in VP9
+    64bit ARM support by replacing ARM assembly with intrinsics
+
+  - Bug Fixes:
+    Fixes a VP9 bitstream issue in Profile 1. This only affected non-YUV 4:2:0
+    files.
+
+  - Known Issues:
+    Frame Parallel decoding fails for segmented and non-420 files.
+
+2013-11-15 v1.3.0 "Forest"
+  This release introduces the VP9 codec in a backward-compatible way.
+  All existing users of VP8 can continue to use the library without
+  modification. However, some VP8 options do not map to VP9 in the same manner.
+
+  The VP9 encoder in this release is not feature complete. Users interested in
+  the encoder are advised to use the git master branch and discuss issues on
+  libvpx mailing lists.
+
+  - Upgrading:
+    This release is ABI and API compatible with Duclair (v1.0.0). Users
+    of older releases should refer to the Upgrading notes in this document
+    for that release.
+
+  - Enhancements:
+      Get rid of bashisms in the main build scripts
+      Added usage info on command line options
+      Add lossless compression mode
+      Dll build of libvpx
+      Add additional Mac OS X targets: 10.7, 10.8 and 10.9 (darwin11-13)
+      Add option to disable documentation
+      configure: add --enable-external-build support
+      make: support V=1 as short form of verbose=yes
+      configure: support mingw-w64
+      configure: support hardfloat armv7 CHOSTS
+      configure: add support for android x86
+      Add estimated completion time to vpxenc
+      Don't exit on decode errors in vpxenc
+      vpxenc: support scaling prior to encoding
+      vpxdec: support scaling output
+      vpxenc: improve progress indicators with --skip
+      msvs: Don't link to winmm.lib
+      Add a new script for producing vcxproj files
+      Produce Visual Studio 10 and 11 project files
+      Produce Windows Phone project files
+      msvs-build: use msbuild for vs >= 2005
+      configure: default configure log to config.log
+      Add encoding option --static-thresh
+
+  - Speed:
+      Miscellaneous speed optimizations for VP8 and VP9.
+
+  - Quality:
+      In general, quality is consistent with the Eider release.
+
+  - Bug Fixes:
+      This release represents approximately a year of engineering effort,
+      and contains multiple bug fixes. Please refer to git history for details.
+
+
+2012-12-21 v1.2.0
+  This release acts as a checkpoint for a large amount of internal refactoring
+  and testing. It also contains a number of small bugfixes, so all users are
+  encouraged to upgrade.
+
+  - Upgrading:
+    This release is ABI and API compatible with Duclair (v1.0.0). Users
+    of older releases should refer to the Upgrading notes in this
+    document for that release.
+
+  - Enhancements:
+      VP8 optimizations for MIPS dspr2
+      vpxenc: add -quiet option
+
+  - Speed:
+      Encoder and decoder speed is consistent with the Eider release.
+
+  - Quality:
+      In general, quality is consistent with the Eider release.
+
+      Minor tweaks to ARNR filtering
+      Minor improvements to real time encoding with multiple temporal layers
+
+  - Bug Fixes:
+      Fixes multithreaded encoder race condition in loopfilter
+      Fixes multi-resolution threaded encoding
+      Fix potential encoder dead-lock after picture resize
+
+
+2012-05-09 v1.1.0 "Eider"
+  This introduces a number of enhancements, mostly focused on real-time
+  encoding. In addition, it fixes a decoder bug (first introduced in
+  Duclair) so all users of that release are encouraged to upgrade.
+
+  - Upgrading:
+    This release is ABI and API compatible with Duclair (v1.0.0). Users
+    of older releases should refer to the Upgrading notes in this
+    document for that release.
+
+    This release introduces a new temporal denoiser, controlled by the
+    VP8E_SET_NOISE_SENSITIVITY control. The temporal denoiser does not
+    currently take a strength parameter, so the control is effectively
+    a boolean - zero (off) or non-zero (on). For compatibility with
+    existing applications, the values accepted are the same as those
+    for the spatial denoiser (0-6). The temporal denoiser is enabled
+    by default, and the older spatial denoiser may be restored by
+    configuring with --disable-temporal-denoising. The temporal denoiser
+    is more computationally intensive than the spatial one.
+
+    This release removes support for a legacy, decode only API that was
+    supported, but deprecated, at the initial release of libvpx
+    (v0.9.0). This is not expected to have any impact. If you are
+    impacted, you can apply a reversion to commit 2bf8fb58 locally.
+    Please update to the latest libvpx API if you are affected.
+
+  - Enhancements:
+      Adds a motion compensated temporal denoiser to the encoder, which
+      gives higher quality than the older spatial denoiser. (See above
+      for notes on upgrading).
+
+      In addition, support for new compilers and platforms were added,
+      including:
+        improved support for XCode
+        Android x86 NDK build
+        OS/2 support
+        SunCC support
+
+      Changing resolution with vpx_codec_enc_config_set() is now
+      supported. Previously, reinitializing the codec was required to
+      change the input resolution.
+
+      The vpxenc application has initial support for producing multiple
+      encodes from the same input in one call. Resizing is not yet
+      supported, but varying other codec parameters is. Use -- to
+      delineate output streams. Options persist from one stream to the
+      next.
+
+      Also, the vpxenc application will now use a keyframe interval of
+      5 seconds by default. Use the --kf-max-dist option to override.
+
+  - Speed:
+      Decoder performance improved 2.5% versus Duclair. Encoder speed is
+      consistent with Duclair for most material. Two pass encoding of
+      slideshow-like material will see significant improvements.
+
+      Large realtime encoding speed gains at a small quality expense are
+      possible by configuring the on-the-fly bitpacking experiment with
+      --enable-onthefly-bitpacking. Realtime encoder can be up to 13%
+      faster (ARM) depending on the number of threads and bitrate
+      settings. This technique sees constant gain over the 5-16 speed
+      range. For VC style input the loss seen is up to 0.2dB. See commit
+      52cf4dca for further details.
+
+  - Quality:
+      On the whole, quality is consistent with the Duclair release. Some
+      tweaks:
+
+        Reduced blockiness in easy sections by applying a penalty to
+        intra modes.
+
+        Improved quality of static sections (like slideshows) with
+        two pass encoding.
+
+        Improved keyframe sizing with multiple temporal layers
+
+  - Bug Fixes:
+      Corrected alt-ref contribution to frame rate for visible updates
+      to the alt-ref buffer. This affected applications making manual
+      usage of the frame reference flags, or temporal layers.
+
+      Additional constraints were added to disable multi-frame quality
+      enhancement (MFQE) in sections of the frame where there is motion.
+      (#392)
+
+      Fixed corruption issues when vpx_codec_enc_config_set() was called
+      with spatial resampling enabled.
+
+      Fixed a decoder error introduced in Duclair where the segmentation
+      map was not being reinitialized on keyframes (#378)
+
+
+2012-01-27 v1.0.0 "Duclair"
+  Our fourth named release, focused on performance and features related to
+  real-time encoding. It also fixes a decoder crash bug introduced in
+  v0.9.7, so all users of that release are encouraged to upgrade.
+
+  - Upgrading:
+      This release is ABI incompatible with prior releases of libvpx, so the
+      "major" version number has been bumped to 1. You must recompile your
+      applications against the latest version of the libvpx headers. The
+      API remains compatible, and this should not require code changes in most
+      applications.
+
+  - Enhancements:
+      This release introduces several substantial new features to the encoder,
+      of particular interest to real time streaming applications.
+
+      Temporal scalability allows the encoder to produce a stream that can
+      be decimated to different frame rates, with independent rate targetting
+      for each substream.
+
+      Multiframe quality enhancement postprocessing can make visual quality
+      more consistent in the presence of frames that are substantially
+      different quality than the surrounding frames, as in the temporal
+      scalability case and in some forced keyframe scenarios.
+
+      Multiple-resolution encoding support allows the encoding of the
+      same content at different resolutions faster than encoding them
+      separately.
+
+  - Speed:
+      Optimization targets for this release included the decoder and the real-
+      time modes of the encoder. Decoder speed on x86 has improved 10.5% with
+      this release. Encoder improvements followed a curve where speeds 1-3
+      improved 4.0%-1.5%, speeds 4-8 improved <1%, and speeds 9-16 improved
+      1.5% to 10.5%, respectively. "Best" mode speed is consistent with the
+      Cayuga release.
+
+  - Quality:
+      Encoder quality in the single stream case is consistent with the Cayuga
+      release.
+
+  - Bug Fixes:
+      This release fixes an OOB read decoder crash bug present in v0.9.7
+      related to the clamping of motion vectors in SPLITMV blocks. This
+      behavior could be triggered by corrupt input or by starting
+      decoding from a P-frame.
+
+
+2011-08-15 v0.9.7-p1 "Cayuga" patch 1
+  This is an incremental bugfix release against Cayuga. All users of that
+  release are strongly encouraged to upgrade.
+
+    - Fix potential OOB reads (cdae03a)
+
+          An unbounded out of bounds read was discovered when the
+          decoder was requested to perform error concealment (new in
+          Cayuga) given a frame with corrupt partition sizes.
+
+          A bounded out of bounds read was discovered affecting all
+          versions of libvpx. Given an multipartition input frame that
+          is truncated between the mode/mv partition and the first
+          residiual paritition (in the block of partition offsets), up
+          to 3 extra bytes could have been read from the source buffer.
+          The code will not take any action regardless of the contents
+          of these undefined bytes, as the truncated buffer is detected
+          immediately following the read based on the calculated
+          starting position of the coefficient partition.
+
+    - Fix potential error concealment crash when the very first frame
+      is missing or corrupt (a609be5)
+
+    - Fix significant artifacts in error concealment (a4c2211, 99d870a)
+
+    - Revert 1-pass CBR rate control changes (e961317)
+      Further testing showed this change produced undesirable visual
+      artifacts, rolling back for now.
+
+
+2011-08-02 v0.9.7 "Cayuga"
+  Our third named release, focused on a faster, higher quality, encoder.
+
+  - Upgrading:
+    This release is backwards compatible with Aylesbury (v0.9.5) and
+    Bali (v0.9.6). Users of older releases should refer to the Upgrading
+    notes in this document for that release.
+
+  - Enhancements:
+          Stereo 3D format support for vpxenc
+          Runtime detection of available processor cores.
+          Allow specifying --end-usage by enum name
+          vpxdec: test for frame corruption
+          vpxenc: add quantizer histogram display
+          vpxenc: add rate histogram display
+          Set VPX_FRAME_IS_DROPPABLE
+          update configure for ios sdk 4.3
+          Avoid text relocations in ARM vp8 decoder
+          Generate a vpx.pc file for pkg-config.
+          New ways of passing encoded data between encoder and decoder.
+
+  - Speed:
+      This release includes across-the-board speed improvements to the
+      encoder. On x86, these measure at approximately 11.5% in Best mode,
+      21.5% in Good mode (speed 0), and 22.5% in Realtime mode (speed 6).
+      On ARM Cortex A9 with Neon extensions, real-time encoding of video
+      telephony content is 35% faster than Bali on single core and 48%
+      faster on multi-core. On the NVidia Tegra2 platform, real time
+      encoding is 40% faster than Bali.
+
+      Decoder speed was not a priority for this release, but improved
+      approximately 8.4% on x86.
+
+          Reduce motion vector search on alt-ref frame.
+          Encoder loopfilter running in its own thread
+          Reworked loopfilter to precalculate more parameters
+          SSE2/SSSE3 optimizations for build_predictors_mbuv{,_s}().
+          Make hor UV predict ~2x faster (73 vs 132 cycles) using SSSE3.
+          Removed redundant checks
+          Reduced structure sizes
+          utilize preload in ARMv6 MC/LPF/Copy routines
+          ARM optimized quantization, dfct, variance, subtract
+          Increase chrow row alignment to 16 bytes.
+          disable trellis optimization for first pass
+          Write SSSE3 sub-pixel filter function
+          Improve SSE2 half-pixel filter funtions
+          Add vp8_sub_pixel_variance16x8_ssse3 function
+          Reduce unnecessary distortion computation
+          Use diamond search to replace full search
+          Preload reference area in sub-pixel motion search (real-time mode)
+
+  - Quality:
+      This release focused primarily on one-pass use cases, including
+      video conferencing. Low latency data rate control was significantly
+      improved, improving streamability over bandwidth constrained links.
+      Added support for error concealment, allowing frames to maintain
+      visual quality in the presence of substantial packet loss.
+
+          Add rc_max_intra_bitrate_pct control
+          Limit size of initial keyframe in one-pass.
+          Improve framerate adaptation
+          Improved 1-pass CBR rate control
+          Improved KF insertion after fades to still.
+          Improved key frame detection.
+          Improved activity masking (lower PSNR impact for same SSIM boost)
+          Improved interaction between GF and ARFs
+          Adding error-concealment to the decoder.
+          Adding support for independent partitions
+          Adjusted rate-distortion constants
+
+
+  - Bug Fixes:
+          Removed firstpass motion map
+          Fix parallel make install
+          Fix multithreaded encoding for 1 MB wide frame
+          Fixed iwalsh_neon build problems with RVDS4.1
+          Fix semaphore emulation, spin-wait intrinsics on Windows
+          Fix build with xcode4 and simplify GLOBAL.
+          Mark ARM asm objects as allowing a non-executable stack.
+          Fix vpxenc encoding incorrect webm file header on big endian
+
+
+2011-03-07 v0.9.6 "Bali"
+  Our second named release, focused on a faster, higher quality, encoder.
+
+  - Upgrading:
+    This release is backwards compatible with Aylesbury (v0.9.5). Users
+    of older releases should refer to the Upgrading notes in this
+    document for that release.
+
+  - Enhancements:
+      vpxenc --psnr shows a summary when encode completes
+      --tune=ssim option to enable activity masking
+      improved postproc visualizations for development
+      updated support for Apple iOS to SDK 4.2
+      query decoder to determine which reference frames were updated
+      implemented error tracking in the decoder
+      fix pipe support on windows
+
+  - Speed:
+      Primary focus was on good quality mode, speed 0. Average improvement
+      on x86 about 40%, up to 100% on user-generated content at that speed.
+      Best quality mode speed improved 35%, and realtime speed 10-20%. This
+      release also saw significant improvement in realtime encoding speed
+      on ARM platforms.
+
+        Improved encoder threading
+        Dont pick encoder filter level when loopfilter is disabled.
+        Avoid double copying of key frames into alt and golden buffer
+        FDCT optimizations.
+        x86 sse2 temporal filter
+        SSSE3 version of fast quantizer
+        vp8_rd_pick_best_mbsegmentation code restructure
+        Adjusted breakout RD for SPLITMV
+        Changed segmentation check order
+        Improved rd_pick_intra4x4block
+        Adds armv6 optimized variance calculation
+        ARMv6 optimized sad16x16
+        ARMv6 optimized half pixel variance calculations
+        Full search SAD function optimization in SSE4.1
+        Improve MV prediction accuracy to achieve performance gain
+        Improve MV prediction in vp8_pick_inter_mode() for speed>3
+
+  - Quality:
+      Best quality mode improved PSNR 6.3%, and SSIM 6.1%. This release
+      also includes support for "activity masking," which greatly improves
+      SSIM at the expense of PSNR. For now, this feature is available with
+      the --tune=ssim option. Further experimentation in this area
+      is ongoing. This release also introduces a new rate control mode
+      called "CQ," which changes the allocation of bits within a clip to
+      the sections where they will have the most visual impact.
+
+        Tuning for the more exact quantizer.
+        Relax rate control for last few frames
+        CQ Mode
+        Limit key frame quantizer for forced key frames.
+        KF/GF Pulsing
+        Add simple version of activity masking.
+        make rdmult adaptive for intra in quantizer RDO
+        cap the best quantizer for 2nd order DC
+        change the threshold of DC check for encode breakout
+
+  - Bug Fixes:
+      Fix crash on Sparc Solaris.
+      Fix counter of fixed keyframe distance
+      ARNR filter pointer update bug fix
+      Fixed use of motion percentage in KF/GF group calc
+      Changed condition for using RD in Intra Mode
+      Fix encoder real-time only configuration.
+      Fix ARM encoder crash with multiple token partitions
+      Fixed bug first cluster timecode of webm file is wrong.
+      Fixed various encoder bugs with odd-sized images
+      vp8e_get_preview fixed when spatial resampling enabled
+      quantizer: fix assertion in fast quantizer path
+      Allocate source buffers to be multiples of 16
+      Fix for manual Golden frame frequency
+      Fix drastic undershoot in long form content
+
+
+2010-10-28 v0.9.5 "Aylesbury"
+  Our first named release, focused on a faster decoder, and a better encoder.
+
+  - Upgrading:
+    This release incorporates backwards-incompatible changes to the
+    ivfenc and ivfdec tools. These tools are now called vpxenc and vpxdec.
+
+    vpxdec
+      * the -q (quiet) option has been removed, and replaced with
+        -v (verbose). the output is quiet by default. Use -v to see
+        the version number of the binary.
+
+      * The default behavior is now to write output to a single file
+        instead of individual frames. The -y option has been removed.
+        Y4M output is the default.
+
+      * For raw I420/YV12 output instead of Y4M, the --i420 or --yv12
+        options must be specified.
+
+          $ ivfdec -o OUTPUT INPUT
+          $ vpxdec --i420 -o OUTPUT INPUT
+
+      * If an output file is not specified, the default is to write
+        Y4M to stdout. This makes piping more natural.
+
+          $ ivfdec -y -o - INPUT | ...
+          $ vpxdec INPUT | ...
+
+      * The output file has additional flexibility for formatting the
+        filename. It supports escape characters for constructing a
+        filename from the width, height, and sequence number. This
+        replaces the -p option. To get the equivalent:
+
+          $ ivfdec -p frame INPUT
+          $ vpxdec --i420 -o frame-%wx%h-%4.i420 INPUT
+
+    vpxenc
+      * The output file must be specified with -o, rather than as the
+        last argument.
+
+          $ ivfenc <options> INPUT OUTPUT
+          $ vpxenc <options> -o OUTPUT INPUT
+
+      * The output defaults to webm. To get IVF output, use the --ivf
+        option.
+
+          $ ivfenc <options> INPUT OUTPUT.ivf
+          $ vpxenc <options> -o OUTPUT.ivf --ivf INPUT
+
+
+  - Enhancements:
+      ivfenc and ivfdec have been renamed to vpxenc, vpxdec.
+      vpxdec supports .webm input
+      vpxdec writes .y4m by default
+      vpxenc writes .webm output by default
+      vpxenc --psnr now shows the average/overall PSNR at the end
+      ARM platforms now support runtime cpu detection
+      vpxdec visualizations added for motion vectors, block modes, references
+      vpxdec now silent by default
+      vpxdec --progress shows frame-by-frame timing information
+      vpxenc supports the distinction between --fps and --timebase
+      NASM is now a supported assembler
+      configure: enable PIC for shared libs by default
+      configure: add --enable-small
+      configure: support for ppc32-linux-gcc
+      configure: support for sparc-solaris-gcc
+
+  - Bugs:
+      Improve handling of invalid frames
+      Fix valgrind errors in the NEON loop filters.
+      Fix loopfilter delta zero transitions
+      Fix valgrind errors in vp8_sixtap_predict8x4_armv6().
+      Build fixes for darwin-icc
+
+  - Speed:
+      20-40% (average 28%) improvement in libvpx decoder speed,
+      including:
+        Rewrite vp8_short_walsh4x4_sse2()
+        Optimizations on the loopfilters.
+        Miscellaneous improvements for Atom
+        Add 4-tap version of 2nd-pass ARMv6 MC filter.
+        Improved multithread utilization
+        Better instruction choices on x86
+        reorder data to use wider instructions
+        Update NEON wide idcts
+        Make block access to frame buffer sequential
+        Improved subset block search
+        Bilinear subpixel optimizations for ssse3.
+        Decrease memory footprint
+
+      Encoder speed improvements (percentage gain not measured):
+        Skip unnecessary search of identical frames
+        Add SSE2 subtract functions
+        Improve bounds checking in vp8_diamond_search_sadx4()
+        Added vp8_fast_quantize_b_sse2
+
+  - Quality:
+      Over 7% overall PSNR improvement (6.3% SSIM) in "best" quality
+      encoding mode, and up to 60% improvement on very noisy, still
+      or slow moving source video
+
+        Motion compensated temporal filter for Alt-Ref Noise Reduction
+        Improved use of trellis quantization on 2nd order Y blocks
+        Tune effect of motion on KF/GF boost in two pass
+        Allow coefficient optimization for good quality speed 0.
+        Improved control of active min quantizer for two pass.
+        Enable ARFs for non-lagged compress
+
+2010-09-02 v0.9.2
+  - Enhancements:
+      Disable frame dropping by default
+      Improved multithreaded performance
+      Improved Force Key Frame Behaviour
+      Increased rate control buffer level precision
+      Fix bug in 1st pass motion compensation
+      ivfenc: correct fixed kf interval, --disable-kf
+  - Speed:
+      Changed above and left context data layout
+      Rework idct calling structure.
+      Removed unnecessary MB_MODE_INFO copies
+      x86: SSSE3 sixtap prediction
+      Reworked IDCT to include reconstruction (add) step
+      Swap alt/gold/new/last frame buffer ptrs instead of copying.
+      Improve SSE2 loopfilter functions
+      Change bitreader to use a larger window.
+      Avoid loopfilter reinitialization when possible
+  - Quality:
+      Normalize quantizer's zero bin and rounding factors
+      Add trellis quantization.
+      Make the quantizer exact.
+      Updates to ARNR filtering algorithm
+      Fix breakout thresh computation for golden & AltRef frames
+      Redo the forward 4x4 dct
+      Improve the accuracy of forward walsh-hadamard transform
+      Further adjustment of RD behaviour with Q and Zbin.
+  - Build System:
+      Allow linking of libs built with MinGW to MSVC
+      Fix target auto-detection on mingw32
+      Allow --cpu= to work for x86.
+      configure: pass original arguments through to make dist
+      Fix builds without runtime CPU detection
+      msvs: fix install of codec sources
+      msvs: Change devenv.com command line for better msys support
+      msvs: Add vs9 targets.
+      Add x86_64-linux-icc target
+  - Bugs:
+      Potential crashes on older MinGW builds
+      Fix two-pass framrate for Y4M input.
+      Fixed simple loop filter, other crashes on ARM v6
+      arm: fix missing dependency with --enable-shared
+      configure: support directories containing .o
+      Replace pinsrw (SSE) with MMX instructions
+      apple: include proper mach primatives
+      Fixed rate control bug with long key frame interval.
+      Fix DSO link errors on x86-64 when not using a version script
+      Fixed buffer selection for UV in AltRef filtering
+
+
+2010-06-17 v0.9.1
+  - Enhancements:
+      * ivfenc/ivfdec now support YUV4MPEG2 input and pipe I/O
+      * Speed optimizations
+  - Bugfixes:
+      * Rate control
+      * Prevent out-of-bounds accesses on invalid data
+  - Build system updates:
+      * Detect toolchain to be used automatically for native builds
+      * Support building shared libraries
+      * Better autotools emulation (--prefix, --libdir, DESTDIR)
+  - Updated LICENSE
+      * http://webmproject.blogspot.com/2010/06/changes-to-webm-open-source-license.html
+
+
+2010-05-18 v0.9.0
+  - Initial open source release. Welcome to WebM and VP8!
+
diff --git a/libvpx/libvpx/LICENSE b/libvpx/libvpx/LICENSE
new file mode 100644
index 0000000..1ce4434
--- /dev/null
+++ b/libvpx/libvpx/LICENSE
@@ -0,0 +1,31 @@
+Copyright (c) 2010, The WebM Project authors. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+  * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+
+  * Redistributions in binary form must reproduce the above copyright
+    notice, this list of conditions and the following disclaimer in
+    the documentation and/or other materials provided with the
+    distribution.
+
+  * Neither the name of Google, nor the WebM Project, nor the names
+    of its contributors may be used to endorse or promote products
+    derived from this software without specific prior written
+    permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
diff --git a/libvpx/libvpx/PATENTS b/libvpx/libvpx/PATENTS
new file mode 100644
index 0000000..caedf60
--- /dev/null
+++ b/libvpx/libvpx/PATENTS
@@ -0,0 +1,23 @@
+Additional IP Rights Grant (Patents)
+------------------------------------
+
+"These implementations" means the copyrightable works that implement the WebM
+codecs distributed by Google as part of the WebM Project.
+
+Google hereby grants to you a perpetual, worldwide, non-exclusive, no-charge,
+royalty-free, irrevocable (except as stated in this section) patent license to
+make, have made, use, offer to sell, sell, import, transfer, and otherwise
+run, modify and propagate the contents of these implementations of WebM, where
+such license applies only to those patent claims, both currently owned by
+Google and acquired in the future, licensable by Google that are necessarily
+infringed by these implementations of WebM. This grant does not include claims
+that would be infringed only as a consequence of further modification of these
+implementations. If you or your agent or exclusive licensee institute or order
+or agree to the institution of patent litigation or any other patent
+enforcement activity against any entity (including a cross-claim or
+counterclaim in a lawsuit) alleging that any of these implementations of WebM
+or any code incorporated within any of these implementations of WebM
+constitute direct or contributory patent infringement, or inducement of
+patent infringement, then any patent rights granted to you under this License
+for these implementations of WebM shall terminate as of the date such
+litigation is filed.
diff --git a/libvpx/libvpx/README b/libvpx/libvpx/README
new file mode 100644
index 0000000..a8e6aeb
--- /dev/null
+++ b/libvpx/libvpx/README
@@ -0,0 +1,138 @@
+README - 20 July 2016
+
+Welcome to the WebM VP8/VP9 Codec SDK!
+
+COMPILING THE APPLICATIONS/LIBRARIES:
+  The build system used is similar to autotools. Building generally consists of
+  "configuring" with your desired build options, then using GNU make to build
+  the application.
+
+  1. Prerequisites
+
+    * All x86 targets require the Yasm[1] assembler be installed.
+    * All Windows builds require that Cygwin[2] be installed.
+    * Building the documentation requires Doxygen[3]. If you do not
+      have this package, the install-docs option will be disabled.
+    * Downloading the data for the unit tests requires curl[4] and sha1sum.
+      sha1sum is provided via the GNU coreutils, installed by default on
+      many *nix platforms, as well as MinGW and Cygwin. If coreutils is not
+      available, a compatible version of sha1sum can be built from
+      source[5]. These requirements are optional if not running the unit
+      tests.
+
+    [1]: http://www.tortall.net/projects/yasm
+    [2]: http://www.cygwin.com
+    [3]: http://www.doxygen.org
+    [4]: http://curl.haxx.se
+    [5]: http://www.microbrew.org/tools/md5sha1sum/
+
+  2. Out-of-tree builds
+  Out of tree builds are a supported method of building the application. For
+  an out of tree build, the source tree is kept separate from the object
+  files produced during compilation. For instance:
+
+    $ mkdir build
+    $ cd build
+    $ ../libvpx/configure <options>
+    $ make
+
+  3. Configuration options
+  The 'configure' script supports a number of options. The --help option can be
+  used to get a list of supported options:
+    $ ../libvpx/configure --help
+
+  4. Cross development
+  For cross development, the most notable option is the --target option. The
+  most up-to-date list of supported targets can be found at the bottom of the
+  --help output of the configure script. As of this writing, the list of
+  available targets is:
+
+    armv6-linux-rvct
+    armv6-linux-gcc
+    armv6-none-rvct
+    arm64-darwin-gcc
+    armv7-android-gcc
+    armv7-darwin-gcc
+    armv7-linux-rvct
+    armv7-linux-gcc
+    armv7-none-rvct
+    armv7-win32-vs11
+    armv7-win32-vs12
+    armv7-win32-vs14
+    armv7s-darwin-gcc
+    mips32-linux-gcc
+    mips64-linux-gcc
+    sparc-solaris-gcc
+    x86-android-gcc
+    x86-darwin8-gcc
+    x86-darwin8-icc
+    x86-darwin9-gcc
+    x86-darwin9-icc
+    x86-darwin10-gcc
+    x86-darwin11-gcc
+    x86-darwin12-gcc
+    x86-darwin13-gcc
+    x86-darwin14-gcc
+    x86-iphonesimulator-gcc
+    x86-linux-gcc
+    x86-linux-icc
+    x86-os2-gcc
+    x86-solaris-gcc
+    x86-win32-gcc
+    x86-win32-vs7
+    x86-win32-vs8
+    x86-win32-vs9
+    x86-win32-vs10
+    x86-win32-vs11
+    x86-win32-vs12
+    x86-win32-vs14
+    x86_64-android-gcc
+    x86_64-darwin9-gcc
+    x86_64-darwin10-gcc
+    x86_64-darwin11-gcc
+    x86_64-darwin12-gcc
+    x86_64-darwin13-gcc
+    x86_64-darwin14-gcc
+    x86_64-iphonesimulator-gcc
+    x86_64-linux-gcc
+    x86_64-linux-icc
+    x86_64-solaris-gcc
+    x86_64-win64-gcc
+    x86_64-win64-vs8
+    x86_64-win64-vs9
+    x86_64-win64-vs10
+    x86_64-win64-vs11
+    x86_64-win64-vs12
+    x86_64-win64-vs14
+    generic-gnu
+
+  The generic-gnu target, in conjunction with the CROSS environment variable,
+  can be used to cross compile architectures that aren't explicitly listed, if
+  the toolchain is a cross GNU (gcc/binutils) toolchain. Other POSIX toolchains
+  will likely work as well. For instance, to build using the mipsel-linux-uclibc
+  toolchain, the following command could be used (note, POSIX SH syntax, adapt
+  to your shell as necessary):
+
+    $ CROSS=mipsel-linux-uclibc- ../libvpx/configure
+
+  In addition, the executables to be invoked can be overridden by specifying the
+  environment variables: CC, AR, LD, AS, STRIP, NM. Additional flags can be
+  passed to these executables with CFLAGS, LDFLAGS, and ASFLAGS.
+
+  5. Configuration errors
+  If the configuration step fails, the first step is to look in the error log.
+  This defaults to config.log. This should give a good indication of what went
+  wrong. If not, contact us for support.
+
+VP8/VP9 TEST VECTORS:
+  The test vectors can be downloaded and verified using the build system after
+  running configure. To specify an alternate directory the
+  LIBVPX_TEST_DATA_PATH environment variable can be used.
+
+  $ ./configure --enable-unit-tests
+  $ LIBVPX_TEST_DATA_PATH=../libvpx-test-data make testdata
+
+SUPPORT
+  This library is an open source project supported by its community. Please
+  please email webm-discuss@webmproject.org for help.
+
diff --git a/libvpx/libvpx/args.c b/libvpx/libvpx/args.c
new file mode 100644
index 0000000..14b0310
--- /dev/null
+++ b/libvpx/libvpx/args.c
@@ -0,0 +1,236 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+#include "args.h"
+
+#include "vpx_ports/msvc.h"
+
+#if defined(__GNUC__) && __GNUC__
+extern void die(const char *fmt, ...) __attribute__((noreturn));
+#else
+extern void die(const char *fmt, ...);
+#endif
+
+
+struct arg arg_init(char **argv) {
+  struct arg a;
+
+  a.argv      = argv;
+  a.argv_step = 1;
+  a.name      = NULL;
+  a.val       = NULL;
+  a.def       = NULL;
+  return a;
+}
+
+int arg_match(struct arg *arg_, const struct arg_def *def, char **argv) {
+  struct arg arg;
+
+  if (!argv[0] || argv[0][0] != '-')
+    return 0;
+
+  arg = arg_init(argv);
+
+  if (def->short_name
+      && strlen(arg.argv[0]) == strlen(def->short_name) + 1
+      && !strcmp(arg.argv[0] + 1, def->short_name)) {
+
+    arg.name = arg.argv[0] + 1;
+    arg.val = def->has_val ? arg.argv[1] : NULL;
+    arg.argv_step = def->has_val ? 2 : 1;
+  } else if (def->long_name) {
+    const size_t name_len = strlen(def->long_name);
+
+    if (strlen(arg.argv[0]) >= name_len + 2
+        && arg.argv[0][1] == '-'
+        && !strncmp(arg.argv[0] + 2, def->long_name, name_len)
+        && (arg.argv[0][name_len + 2] == '='
+            || arg.argv[0][name_len + 2] == '\0')) {
+
+      arg.name = arg.argv[0] + 2;
+      arg.val = arg.name[name_len] == '=' ? arg.name + name_len + 1 : NULL;
+      arg.argv_step = 1;
+    }
+  }
+
+  if (arg.name && !arg.val && def->has_val)
+    die("Error: option %s requires argument.\n", arg.name);
+
+  if (arg.name && arg.val && !def->has_val)
+    die("Error: option %s requires no argument.\n", arg.name);
+
+  if (arg.name
+      && (arg.val || !def->has_val)) {
+    arg.def = def;
+    *arg_ = arg;
+    return 1;
+  }
+
+  return 0;
+}
+
+
+const char *arg_next(struct arg *arg) {
+  if (arg->argv[0])
+    arg->argv += arg->argv_step;
+
+  return *arg->argv;
+}
+
+
+char **argv_dup(int argc, const char **argv) {
+  char **new_argv = malloc((argc + 1) * sizeof(*argv));
+
+  memcpy(new_argv, argv, argc * sizeof(*argv));
+  new_argv[argc] = NULL;
+  return new_argv;
+}
+
+
+void arg_show_usage(FILE *fp, const struct arg_def *const *defs) {
+  char option_text[40] = {0};
+
+  for (; *defs; defs++) {
+    const struct arg_def *def = *defs;
+    char *short_val = def->has_val ? " <arg>" : "";
+    char *long_val = def->has_val ? "=<arg>" : "";
+
+    if (def->short_name && def->long_name) {
+      char *comma = def->has_val ? "," : ",      ";
+
+      snprintf(option_text, 37, "-%s%s%s --%s%6s",
+               def->short_name, short_val, comma,
+               def->long_name, long_val);
+    } else if (def->short_name)
+      snprintf(option_text, 37, "-%s%s",
+               def->short_name, short_val);
+    else if (def->long_name)
+      snprintf(option_text, 37, "          --%s%s",
+               def->long_name, long_val);
+
+    fprintf(fp, "  %-37s\t%s\n", option_text, def->desc);
+
+    if (def->enums) {
+      const struct arg_enum_list *listptr;
+
+      fprintf(fp, "  %-37s\t  ", "");
+
+      for (listptr = def->enums; listptr->name; listptr++)
+        fprintf(fp, "%s%s", listptr->name,
+                listptr[1].name ? ", " : "\n");
+    }
+  }
+}
+
+
+unsigned int arg_parse_uint(const struct arg *arg) {
+  long int   rawval;
+  char      *endptr;
+
+  rawval = strtol(arg->val, &endptr, 10);
+
+  if (arg->val[0] != '\0' && endptr[0] == '\0') {
+    if (rawval >= 0 && rawval <= UINT_MAX)
+      return rawval;
+
+    die("Option %s: Value %ld out of range for unsigned int\n",
+        arg->name, rawval);
+  }
+
+  die("Option %s: Invalid character '%c'\n", arg->name, *endptr);
+  return 0;
+}
+
+
+int arg_parse_int(const struct arg *arg) {
+  long int   rawval;
+  char      *endptr;
+
+  rawval = strtol(arg->val, &endptr, 10);
+
+  if (arg->val[0] != '\0' && endptr[0] == '\0') {
+    if (rawval >= INT_MIN && rawval <= INT_MAX)
+      return rawval;
+
+    die("Option %s: Value %ld out of range for signed int\n",
+        arg->name, rawval);
+  }
+
+  die("Option %s: Invalid character '%c'\n", arg->name, *endptr);
+  return 0;
+}
+
+
+struct vpx_rational {
+  int num; /**< fraction numerator */
+  int den; /**< fraction denominator */
+};
+struct vpx_rational arg_parse_rational(const struct arg *arg) {
+  long int             rawval;
+  char                *endptr;
+  struct vpx_rational  rat;
+
+  /* parse numerator */
+  rawval = strtol(arg->val, &endptr, 10);
+
+  if (arg->val[0] != '\0' && endptr[0] == '/') {
+    if (rawval >= INT_MIN && rawval <= INT_MAX)
+      rat.num = rawval;
+    else die("Option %s: Value %ld out of range for signed int\n",
+               arg->name, rawval);
+  } else die("Option %s: Expected / at '%c'\n", arg->name, *endptr);
+
+  /* parse denominator */
+  rawval = strtol(endptr + 1, &endptr, 10);
+
+  if (arg->val[0] != '\0' && endptr[0] == '\0') {
+    if (rawval >= INT_MIN && rawval <= INT_MAX)
+      rat.den = rawval;
+    else die("Option %s: Value %ld out of range for signed int\n",
+               arg->name, rawval);
+  } else die("Option %s: Invalid character '%c'\n", arg->name, *endptr);
+
+  return rat;
+}
+
+
+int arg_parse_enum(const struct arg *arg) {
+  const struct arg_enum_list *listptr;
+  long int                    rawval;
+  char                       *endptr;
+
+  /* First see if the value can be parsed as a raw value */
+  rawval = strtol(arg->val, &endptr, 10);
+  if (arg->val[0] != '\0' && endptr[0] == '\0') {
+    /* Got a raw value, make sure it's valid */
+    for (listptr = arg->def->enums; listptr->name; listptr++)
+      if (listptr->val == rawval)
+        return rawval;
+  }
+
+  /* Next see if it can be parsed as a string */
+  for (listptr = arg->def->enums; listptr->name; listptr++)
+    if (!strcmp(arg->val, listptr->name))
+      return listptr->val;
+
+  die("Option %s: Invalid value '%s'\n", arg->name, arg->val);
+  return 0;
+}
+
+
+int arg_parse_enum_or_int(const struct arg *arg) {
+  if (arg->def->enums)
+    return arg_parse_enum(arg);
+  return arg_parse_int(arg);
+}
diff --git a/libvpx/libvpx/args.h b/libvpx/libvpx/args.h
new file mode 100644
index 0000000..1f37151
--- /dev/null
+++ b/libvpx/libvpx/args.h
@@ -0,0 +1,60 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef ARGS_H_
+#define ARGS_H_
+#include <stdio.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct arg {
+  char                 **argv;
+  const char            *name;
+  const char            *val;
+  unsigned int           argv_step;
+  const struct arg_def  *def;
+};
+
+struct arg_enum_list {
+  const char *name;
+  int         val;
+};
+#define ARG_ENUM_LIST_END {0}
+
+typedef struct arg_def {
+  const char *short_name;
+  const char *long_name;
+  int         has_val;
+  const char *desc;
+  const struct arg_enum_list *enums;
+} arg_def_t;
+#define ARG_DEF(s,l,v,d) {s,l,v,d, NULL}
+#define ARG_DEF_ENUM(s,l,v,d,e) {s,l,v,d,e}
+#define ARG_DEF_LIST_END {0}
+
+struct arg arg_init(char **argv);
+int arg_match(struct arg *arg_, const struct arg_def *def, char **argv);
+const char *arg_next(struct arg *arg);
+void arg_show_usage(FILE *fp, const struct arg_def *const *defs);
+char **argv_dup(int argc, const char **argv);
+
+unsigned int arg_parse_uint(const struct arg *arg);
+int arg_parse_int(const struct arg *arg);
+struct vpx_rational arg_parse_rational(const struct arg *arg);
+int arg_parse_enum(const struct arg *arg);
+int arg_parse_enum_or_int(const struct arg *arg);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // ARGS_H_
diff --git a/libvpx/libvpx/build/make/Android.mk b/libvpx/libvpx/build/make/Android.mk
new file mode 100644
index 0000000..9eb6dd2
--- /dev/null
+++ b/libvpx/libvpx/build/make/Android.mk
@@ -0,0 +1,202 @@
+##
+##  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+#
+# This file is to be used for compiling libvpx for Android using the NDK.
+# In an Android project place a libvpx checkout in the jni directory.
+# Run the configure script from the jni directory.  Base libvpx
+# encoder/decoder configuration will look similar to:
+# ./libvpx/configure --target=armv7-android-gcc --disable-examples \
+#                    --sdk-path=/opt/android-ndk-r6b/
+#
+# When targeting Android, realtime-only is enabled by default.  This can
+# be overridden by adding the command line flag:
+#  --disable-realtime-only
+#
+# This will create .mk files that contain variables that contain the
+# source files to compile.
+#
+# Place an Android.mk file in the jni directory that references the
+# Android.mk file in the libvpx directory:
+# LOCAL_PATH := $(call my-dir)
+# include $(CLEAR_VARS)
+# include jni/libvpx/build/make/Android.mk
+#
+# There are currently two TARGET_ARCH_ABI targets for ARM.
+# armeabi and armeabi-v7a.  armeabi-v7a is selected by creating an
+# Application.mk in the jni directory that contains:
+# APP_ABI := armeabi-v7a
+#
+# By default libvpx will detect at runtime the existance of NEON extension.
+# For this we import the 'cpufeatures' module from the NDK sources.
+# libvpx can also be configured without this runtime detection method.
+# Configuring with --disable-runtime-cpu-detect will assume presence of NEON.
+# Configuring with --disable-runtime-cpu-detect --disable-neon \
+#     --disable-neon-asm
+# will remove any NEON dependency.
+
+# To change to building armeabi, run ./libvpx/configure again, but with
+# --target=armv6-android-gcc and modify the Application.mk file to
+# set APP_ABI := armeabi
+#
+# Running ndk-build will build libvpx and include it in your project.
+#
+
+CONFIG_DIR := $(LOCAL_PATH)/
+LIBVPX_PATH := $(LOCAL_PATH)/libvpx
+ASM_CNV_PATH_LOCAL := $(TARGET_ARCH_ABI)/ads2gas
+ASM_CNV_PATH := $(LOCAL_PATH)/$(ASM_CNV_PATH_LOCAL)
+
+# Use the makefiles generated by upstream configure to determine which files to
+# build. Also set any architecture-specific flags.
+ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+  include $(CONFIG_DIR)libs-armv7-android-gcc.mk
+  LOCAL_ARM_MODE := arm
+else ifeq  ($(TARGET_ARCH_ABI),armeabi)
+  include $(CONFIG_DIR)libs-armv6-android-gcc.mk
+  LOCAL_ARM_MODE := arm
+else ifeq  ($(TARGET_ARCH_ABI),arm64-v8a)
+  include $(CONFIG_DIR)libs-armv8-android-gcc.mk
+  LOCAL_ARM_MODE := arm
+else ifeq ($(TARGET_ARCH_ABI),x86)
+  include $(CONFIG_DIR)libs-x86-android-gcc.mk
+else ifeq ($(TARGET_ARCH_ABI),x86_64)
+  include $(CONFIG_DIR)libs-x86_64-android-gcc.mk
+else ifeq ($(TARGET_ARCH_ABI),mips)
+  include $(CONFIG_DIR)libs-mips-android-gcc.mk
+else
+  $(error Not a supported TARGET_ARCH_ABI: $(TARGET_ARCH_ABI))
+endif
+
+# Rule that is normally in Makefile created by libvpx
+# configure.  Used to filter out source files based on configuration.
+enabled=$(filter-out $($(1)-no),$($(1)-yes))
+
+# Override the relative path that is defined by the libvpx
+# configure process
+SRC_PATH_BARE := $(LIBVPX_PATH)
+
+# Include the list of files to be built
+include $(LIBVPX_PATH)/libs.mk
+
+# Optimise the code. May want to revisit this setting in the future.
+LOCAL_CFLAGS := -O3
+
+# For x86, include the source code in the search path so it will find files
+# like x86inc.asm and x86_abi_support.asm
+LOCAL_ASMFLAGS := -I$(LIBVPX_PATH)
+
+.PRECIOUS: %.asm.s
+$(ASM_CNV_PATH)/libvpx/%.asm.s: $(LIBVPX_PATH)/%.asm
+	@mkdir -p $(dir $@)
+	@$(CONFIG_DIR)$(ASM_CONVERSION) <$< > $@
+
+# For building *_rtcd.h, which have rules in libs.mk
+TGT_ISA:=$(word 1, $(subst -, ,$(TOOLCHAIN)))
+target := libs
+
+LOCAL_SRC_FILES += vpx_config.c
+
+# Remove duplicate entries
+CODEC_SRCS_UNIQUE = $(sort $(CODEC_SRCS))
+
+# Pull out C files.  vpx_config.c is in the immediate directory and
+# so it does not need libvpx/ prefixed like the rest of the source files.
+# The neon files with intrinsics need to have .neon appended so the proper
+# flags are applied.
+CODEC_SRCS_C = $(filter %.c, $(CODEC_SRCS_UNIQUE))
+LOCAL_NEON_SRCS_C = $(filter %_neon.c, $(CODEC_SRCS_C))
+LOCAL_CODEC_SRCS_C = $(filter-out vpx_config.c %_neon.c, $(CODEC_SRCS_C))
+
+LOCAL_SRC_FILES += $(foreach file, $(LOCAL_CODEC_SRCS_C), libvpx/$(file))
+ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+  LOCAL_SRC_FILES += $(foreach file, $(LOCAL_NEON_SRCS_C), libvpx/$(file).neon)
+else # If there are neon sources then we are building for arm64 and do not need to specify .neon
+  LOCAL_SRC_FILES += $(foreach file, $(LOCAL_NEON_SRCS_C), libvpx/$(file))
+endif
+
+# Pull out assembly files, splitting NEON from the rest.  This is
+# done to specify that the NEON assembly files use NEON assembler flags.
+# x86 assembly matches %.asm, arm matches %.asm.s
+
+# x86:
+
+CODEC_SRCS_ASM_X86 = $(filter %.asm, $(CODEC_SRCS_UNIQUE))
+LOCAL_SRC_FILES += $(foreach file, $(CODEC_SRCS_ASM_X86), libvpx/$(file))
+
+# arm:
+CODEC_SRCS_ASM_ARM_ALL = $(filter %.asm.s, $(CODEC_SRCS_UNIQUE))
+CODEC_SRCS_ASM_ARM = $(foreach v, \
+                     $(CODEC_SRCS_ASM_ARM_ALL), \
+                     $(if $(findstring neon,$(v)),,$(v)))
+CODEC_SRCS_ASM_ADS2GAS = $(patsubst %.s, \
+                         $(ASM_CNV_PATH_LOCAL)/libvpx/%.s, \
+                         $(CODEC_SRCS_ASM_ARM))
+LOCAL_SRC_FILES += $(CODEC_SRCS_ASM_ADS2GAS)
+
+ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+  CODEC_SRCS_ASM_NEON = $(foreach v, \
+                        $(CODEC_SRCS_ASM_ARM_ALL),\
+                        $(if $(findstring neon,$(v)),$(v),))
+  CODEC_SRCS_ASM_NEON_ADS2GAS = $(patsubst %.s, \
+                                $(ASM_CNV_PATH_LOCAL)/libvpx/%.s, \
+                                $(CODEC_SRCS_ASM_NEON))
+  LOCAL_SRC_FILES += $(patsubst %.s, \
+                     %.s.neon, \
+                     $(CODEC_SRCS_ASM_NEON_ADS2GAS))
+endif
+
+LOCAL_CFLAGS += \
+    -DHAVE_CONFIG_H=vpx_config.h \
+    -I$(LIBVPX_PATH) \
+    -I$(ASM_CNV_PATH)
+
+LOCAL_MODULE := libvpx
+
+ifeq ($(CONFIG_RUNTIME_CPU_DETECT),yes)
+  LOCAL_STATIC_LIBRARIES := cpufeatures
+endif
+
+# Add a dependency to force generation of the RTCD files.
+define rtcd_dep_template
+rtcd_dep_template_SRCS := $(addprefix $(LOCAL_PATH)/, $(LOCAL_SRC_FILES))
+rtcd_dep_template_SRCS := $$(rtcd_dep_template_SRCS:.neon=)
+ifeq ($(CONFIG_VP8), yes)
+$$(rtcd_dep_template_SRCS): vp8_rtcd.h
+endif
+ifeq ($(CONFIG_VP9), yes)
+$$(rtcd_dep_template_SRCS): vp9_rtcd.h
+endif
+$$(rtcd_dep_template_SRCS): vpx_scale_rtcd.h
+$$(rtcd_dep_template_SRCS): vpx_dsp_rtcd.h
+
+ifneq ($(findstring $(TARGET_ARCH_ABI),x86 x86_64),)
+$$(rtcd_dep_template_SRCS): vpx_config.asm
+endif
+endef
+
+$(eval $(call rtcd_dep_template))
+
+.PHONY: clean
+clean:
+	@echo "Clean: ads2gas files [$(TARGET_ARCH_ABI)]"
+	@$(RM) $(CODEC_SRCS_ASM_ADS2GAS) $(CODEC_SRCS_ASM_NEON_ADS2GAS)
+	@$(RM) -r $(ASM_CNV_PATH)
+	@$(RM) $(CLEAN-OBJS)
+
+ifeq ($(ENABLE_SHARED),1)
+  include $(BUILD_SHARED_LIBRARY)
+else
+  include $(BUILD_STATIC_LIBRARY)
+endif
+
+ifeq ($(CONFIG_RUNTIME_CPU_DETECT),yes)
+$(call import-module,cpufeatures)
+endif
diff --git a/libvpx/libvpx/build/make/Makefile b/libvpx/libvpx/build/make/Makefile
new file mode 100644
index 0000000..3e8c024
--- /dev/null
+++ b/libvpx/libvpx/build/make/Makefile
@@ -0,0 +1,451 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+include config.mk
+quiet?=true
+ifeq ($(target),)
+# If a target wasn't specified, invoke for all enabled targets.
+.DEFAULT:
+	@for t in $(ALL_TARGETS); do \
+	     $(MAKE) --no-print-directory target=$$t $(MAKECMDGOALS) || exit $$?;\
+        done
+all: .DEFAULT
+clean:: .DEFAULT
+exampletest: .DEFAULT
+install:: .DEFAULT
+test:: .DEFAULT
+test-no-data-check:: .DEFAULT
+testdata:: .DEFAULT
+utiltest: .DEFAULT
+exampletest-no-data-check utiltest-no-data-check: .DEFAULT
+
+
+# Note: md5sum is not installed on OS X, but openssl is. Openssl may not be
+# installed on cygwin, so we need to autodetect here.
+md5sum := $(firstword $(wildcard \
+          $(foreach e,md5sum openssl,\
+          $(foreach p,$(subst :, ,$(PATH)),$(p)/$(e)*))\
+          ))
+md5sum := $(if $(filter %openssl,$(md5sum)),$(md5sum) dgst -md5,$(md5sum))
+
+TGT_CC:=$(word 3, $(subst -, ,$(TOOLCHAIN)))
+dist:
+	@for t in $(ALL_TARGETS); do \
+	     $(MAKE) --no-print-directory target=$$t $(MAKECMDGOALS) || exit $$?;\
+        done
+        # Run configure for the user with the current toolchain.
+	@if [ -d "$(DIST_DIR)/src" ]; then \
+            mkdir -p "$(DIST_DIR)/build"; \
+            cd "$(DIST_DIR)/build"; \
+            echo "Rerunning configure $(CONFIGURE_ARGS)"; \
+            ../src/configure $(CONFIGURE_ARGS); \
+            $(if $(filter vs%,$(TGT_CC)),make NO_LAUNCH_DEVENV=1;) \
+        fi
+	@if [ -d "$(DIST_DIR)" ]; then \
+            echo "    [MD5SUM] $(DIST_DIR)"; \
+	    cd $(DIST_DIR) && \
+	    $(md5sum) `find . -name md5sums.txt -prune -o -type f -print` \
+                | sed -e 's/MD5(\(.*\))= \([0-9a-f]\{32\}\)/\2  \1/' \
+                > md5sums.txt;\
+        fi
+endif
+
+# Since we invoke make recursively for multiple targets we need to include the
+# .mk file for the correct target, but only when $(target) is non-empty.
+ifneq ($(target),)
+include $(target)-$(TOOLCHAIN).mk
+endif
+BUILD_ROOT?=.
+VPATH=$(SRC_PATH_BARE)
+CFLAGS+=-I$(BUILD_PFX)$(BUILD_ROOT) -I$(SRC_PATH)
+CXXFLAGS+=-I$(BUILD_PFX)$(BUILD_ROOT) -I$(SRC_PATH)
+ASFLAGS+=-I$(BUILD_PFX)$(BUILD_ROOT)/ -I$(SRC_PATH)/
+DIST_DIR?=dist
+HOSTCC?=gcc
+TGT_ISA:=$(word 1, $(subst -, ,$(TOOLCHAIN)))
+TGT_OS:=$(word 2, $(subst -, ,$(TOOLCHAIN)))
+TGT_CC:=$(word 3, $(subst -, ,$(TOOLCHAIN)))
+quiet:=$(if $(or $(verbose), $(V)),, yes)
+qexec=$(if $(quiet),@)
+
+# Cancel built-in implicit rules
+%: %.o
+%.asm:
+%.a:
+%: %.cc
+
+#
+# Common rules"
+#
+.PHONY: all
+all:
+
+.PHONY: clean
+clean::
+	rm -f $(OBJS-yes) $(OBJS-yes:.o=.d) $(OBJS-yes:.asm.s.o=.asm.s)
+	rm -f $(CLEAN-OBJS)
+
+.PHONY: clean
+distclean: clean
+	if [ -z "$(target)" ]; then \
+      rm -f Makefile; \
+      rm -f config.log config.mk; \
+      rm -f vpx_config.[hc] vpx_config.asm; \
+    else \
+      rm -f $(target)-$(TOOLCHAIN).mk; \
+    fi
+
+.PHONY: dist
+dist:
+.PHONY: exampletest
+exampletest:
+.PHONY: install
+install::
+.PHONY: test
+test::
+.PHONY: testdata
+testdata::
+.PHONY: utiltest
+utiltest:
+.PHONY: test-no-data-check exampletest-no-data-check utiltest-no-data-check
+test-no-data-check::
+exampletest-no-data-check utiltest-no-data-check:
+
+# Force to realign stack always on OS/2
+ifeq ($(TOOLCHAIN), x86-os2-gcc)
+CFLAGS += -mstackrealign
+endif
+
+$(BUILD_PFX)%_mmx.c.d: CFLAGS += -mmmx
+$(BUILD_PFX)%_mmx.c.o: CFLAGS += -mmmx
+$(BUILD_PFX)%_sse2.c.d: CFLAGS += -msse2
+$(BUILD_PFX)%_sse2.c.o: CFLAGS += -msse2
+$(BUILD_PFX)%_sse3.c.d: CFLAGS += -msse3
+$(BUILD_PFX)%_sse3.c.o: CFLAGS += -msse3
+$(BUILD_PFX)%_ssse3.c.d: CFLAGS += -mssse3
+$(BUILD_PFX)%_ssse3.c.o: CFLAGS += -mssse3
+$(BUILD_PFX)%_sse4.c.d: CFLAGS += -msse4.1
+$(BUILD_PFX)%_sse4.c.o: CFLAGS += -msse4.1
+$(BUILD_PFX)%_avx.c.d: CFLAGS += -mavx
+$(BUILD_PFX)%_avx.c.o: CFLAGS += -mavx
+$(BUILD_PFX)%_avx2.c.d: CFLAGS += -mavx2
+$(BUILD_PFX)%_avx2.c.o: CFLAGS += -mavx2
+
+$(BUILD_PFX)%.c.d: %.c
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(CC) $(INTERNAL_CFLAGS) $(CFLAGS) -M $< | $(fmt_deps) > $@
+
+$(BUILD_PFX)%.c.o: %.c
+	$(if $(quiet),@echo "    [CC] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(CC) $(INTERNAL_CFLAGS) $(CFLAGS) -c -o $@ $<
+
+$(BUILD_PFX)%.cc.d: %.cc
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -M $< | $(fmt_deps) > $@
+
+$(BUILD_PFX)%.cc.o: %.cc
+	$(if $(quiet),@echo "    [CXX] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -c -o $@ $<
+
+$(BUILD_PFX)%.cpp.d: %.cpp
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -M $< | $(fmt_deps) > $@
+
+$(BUILD_PFX)%.cpp.o: %.cpp
+	$(if $(quiet),@echo "    [CXX] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -c -o $@ $<
+
+$(BUILD_PFX)%.asm.d: %.asm
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(SRC_PATH_BARE)/build/make/gen_asm_deps.sh \
+            --build-pfx=$(BUILD_PFX) --depfile=$@ $(ASFLAGS) $< > $@
+
+$(BUILD_PFX)%.asm.o: %.asm
+	$(if $(quiet),@echo "    [AS] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(AS) $(ASFLAGS) -o $@ $<
+
+$(BUILD_PFX)%.s.d: %.s
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(SRC_PATH_BARE)/build/make/gen_asm_deps.sh \
+            --build-pfx=$(BUILD_PFX) --depfile=$@ $(ASFLAGS) $< > $@
+
+$(BUILD_PFX)%.s.o: %.s
+	$(if $(quiet),@echo "    [AS] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(AS) $(ASFLAGS) -o $@ $<
+
+.PRECIOUS: %.c.S
+%.c.S: CFLAGS += -DINLINE_ASM
+$(BUILD_PFX)%.c.S: %.c
+	$(if $(quiet),@echo "    [GEN] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(CC) -S $(CFLAGS) -o $@ $<
+
+.PRECIOUS: %.asm.s
+$(BUILD_PFX)%.asm.s: %.asm
+	$(if $(quiet),@echo "    [ASM CONVERSION] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(ASM_CONVERSION) <$< >$@
+
+# If we're in debug mode, pretend we don't have GNU strip, to fall back to
+# the copy implementation
+HAVE_GNU_STRIP := $(if $(CONFIG_DEBUG),,$(HAVE_GNU_STRIP))
+ifeq ($(HAVE_GNU_STRIP),yes)
+# Older binutils strip global symbols not needed for relocation processing
+# when given --strip-unneeded. Using nm and awk to identify globals and
+# keep them caused command line length issues under mingw and segfaults in
+# test_libvpx were observed under OS/2: simply use --strip-debug.
+%.a: %_g.a
+	$(if $(quiet),@echo "    [STRIP] $@ < $<")
+	$(qexec)$(STRIP) --strip-debug \
+          -o $@ $<
+else
+%.a: %_g.a
+	$(if $(quiet),@echo "    [CP] $@ < $<")
+	$(qexec)cp $< $@
+endif
+
+#
+# Utility functions
+#
+pairmap=$(if $(strip $(2)),\
+    $(call $(1),$(word 1,$(2)),$(word 2,$(2)))\
+    $(call pairmap,$(1),$(wordlist 3,$(words $(2)),$(2)))\
+)
+
+enabled=$(filter-out $($(1)-no),$($(1)-yes))
+cond_enabled=$(if $(filter yes,$($(1))), $(call enabled,$(2)))
+
+find_file1=$(word 1,$(wildcard $(subst //,/,$(addsuffix /$(1),$(2)))))
+find_file=$(foreach f,$(1),$(call find_file1,$(strip $(f)),$(strip $(2))) )
+obj_pats=.c=.c.o $(AS_SFX)=$(AS_SFX).o .cc=.cc.o .cpp=.cpp.o
+objs=$(addprefix $(BUILD_PFX),$(foreach p,$(obj_pats),$(filter %.o,$(1:$(p))) ))
+
+install_map_templates=$(eval $(call install_map_template,$(1),$(2)))
+
+not=$(subst yes,no,$(1))
+
+ifeq ($(CONFIG_MSVS),yes)
+lib_file_name=$(1).lib
+else
+lib_file_name=lib$(1).a
+endif
+#
+# Rule Templates
+#
+define linker_template
+$(1): $(filter-out -%,$(2))
+$(1):
+	$(if $(quiet),@echo    "    [LD] $$@")
+	$(qexec)$$(LD) $$(strip $$(INTERNAL_LDFLAGS) $$(LDFLAGS) -o $$@ $(2) $(3) $$(extralibs))
+endef
+define linkerxx_template
+$(1): $(filter-out -%,$(2))
+$(1):
+	$(if $(quiet),@echo    "    [LD] $$@")
+	$(qexec)$$(CXX) $$(strip $$(INTERNAL_LDFLAGS) $$(LDFLAGS) -o $$@ $(2) $(3) $$(extralibs))
+endef
+# make-3.80 has a bug with expanding large input strings to the eval function,
+# which was triggered in some cases by the following component of
+# linker_template:
+#   $(1): $$(call find_file, $(patsubst -l%,lib%.a,$(filter -l%,$(2))),\
+#                           $$(patsubst -L%,%,$$(filter -L%,$$(LDFLAGS) $(2))))
+# This may be useful to revisit in the future (it tries to locate libraries
+# in a search path and add them as prerequisites
+
+define install_map_template
+$(DIST_DIR)/$(1): $(2)
+	$(if $(quiet),@echo "    [INSTALL] $$@")
+	$(qexec)mkdir -p $$(dir $$@)
+	$(qexec)cp -p $$< $$@
+endef
+
+define archive_template
+# Not using a pattern rule here because we don't want to generate empty
+# archives when they are listed as a dependency in files not responsible
+# for creating them.
+$(1):
+	$(if $(quiet),@echo "    [AR] $$@")
+	$(qexec)$$(AR) $$(ARFLAGS) $$@ $$^
+endef
+
+define so_template
+# Not using a pattern rule here because we don't want to generate empty
+# archives when they are listed as a dependency in files not responsible
+# for creating them.
+#
+# This needs further abstraction for dealing with non-GNU linkers.
+$(1):
+	$(if $(quiet),@echo "    [LD] $$@")
+	$(qexec)$$(LD) -shared $$(LDFLAGS) \
+            -Wl,--no-undefined -Wl,-soname,$$(SONAME) \
+            -Wl,--version-script,$$(EXPORTS_FILE) -o $$@ \
+            $$(filter %.o,$$^) $$(extralibs)
+endef
+
+define dl_template
+# Not using a pattern rule here because we don't want to generate empty
+# archives when they are listed as a dependency in files not responsible
+# for creating them.
+$(1):
+	$(if $(quiet),@echo "    [LD] $$@")
+	$(qexec)$$(LD) -dynamiclib $$(LDFLAGS) \
+	    -exported_symbols_list $$(EXPORTS_FILE) \
+        -Wl,-headerpad_max_install_names,-compatibility_version,1.0,-current_version,$$(VERSION_MAJOR) \
+        -o $$@ \
+        $$(filter %.o,$$^) $$(extralibs)
+endef
+
+define dll_template
+# Not using a pattern rule here because we don't want to generate empty
+# archives when they are listed as a dependency in files not responsible
+# for creating them.
+$(1):
+	$(if $(quiet),@echo "    [LD] $$@")
+	$(qexec)$$(LD) -Zdll $$(LDFLAGS) \
+        -o $$@ \
+        $$(filter %.o,$$^) $$(extralibs) $$(EXPORTS_FILE)
+endef
+
+
+#
+# Get current configuration
+#
+ifneq ($(target),)
+include $(SRC_PATH_BARE)/$(target:-$(TOOLCHAIN)=).mk
+endif
+
+skip_deps := $(filter %clean,$(MAKECMDGOALS))
+skip_deps += $(findstring testdata,$(MAKECMDGOALS))
+ifeq ($(strip $(skip_deps)),)
+  ifeq ($(CONFIG_DEPENDENCY_TRACKING),yes)
+    # Older versions of make don't like -include directives with no arguments
+    ifneq ($(filter %.d,$(OBJS-yes:.o=.d)),)
+      -include $(filter %.d,$(OBJS-yes:.o=.d))
+    endif
+  endif
+endif
+
+#
+# Configuration dependent rules
+#
+$(call pairmap,install_map_templates,$(INSTALL_MAPS))
+
+DOCS=$(call cond_enabled,CONFIG_INSTALL_DOCS,DOCS)
+.docs: $(DOCS)
+	@touch $@
+
+INSTALL-DOCS=$(call cond_enabled,CONFIG_INSTALL_DOCS,INSTALL-DOCS)
+ifeq ($(MAKECMDGOALS),dist)
+INSTALL-DOCS+=$(call cond_enabled,CONFIG_INSTALL_DOCS,DIST-DOCS)
+endif
+.install-docs: .docs $(addprefix $(DIST_DIR)/,$(INSTALL-DOCS))
+	@touch $@
+
+clean::
+	rm -f .docs .install-docs $(DOCS)
+
+BINS=$(call enabled,BINS)
+.bins: $(BINS)
+	@touch $@
+
+INSTALL-BINS=$(call cond_enabled,CONFIG_INSTALL_BINS,INSTALL-BINS)
+ifeq ($(MAKECMDGOALS),dist)
+INSTALL-BINS+=$(call cond_enabled,CONFIG_INSTALL_BINS,DIST-BINS)
+endif
+.install-bins: .bins $(addprefix $(DIST_DIR)/,$(INSTALL-BINS))
+	@touch $@
+
+clean::
+	rm -f .bins .install-bins $(BINS)
+
+LIBS=$(call enabled,LIBS)
+.libs: $(LIBS)
+	@touch $@
+$(foreach lib,$(filter %_g.a,$(LIBS)),$(eval $(call archive_template,$(lib))))
+$(foreach lib,$(filter %so.$(SO_VERSION_MAJOR).$(SO_VERSION_MINOR).$(SO_VERSION_PATCH),$(LIBS)),$(eval $(call so_template,$(lib))))
+$(foreach lib,$(filter %$(SO_VERSION_MAJOR).dylib,$(LIBS)),$(eval $(call dl_template,$(lib))))
+$(foreach lib,$(filter %$(SO_VERSION_MAJOR).dll,$(LIBS)),$(eval $(call dll_template,$(lib))))
+
+INSTALL-LIBS=$(call cond_enabled,CONFIG_INSTALL_LIBS,INSTALL-LIBS)
+ifeq ($(MAKECMDGOALS),dist)
+INSTALL-LIBS+=$(call cond_enabled,CONFIG_INSTALL_LIBS,DIST-LIBS)
+endif
+.install-libs: .libs $(addprefix $(DIST_DIR)/,$(INSTALL-LIBS))
+	@touch $@
+
+clean::
+	rm -f .libs .install-libs $(LIBS)
+
+ifeq ($(CONFIG_EXTERNAL_BUILD),yes)
+PROJECTS=$(call enabled,PROJECTS)
+.projects: $(PROJECTS)
+	@touch $@
+
+INSTALL-PROJECTS=$(call cond_enabled,CONFIG_INSTALL_PROJECTS,INSTALL-PROJECTS)
+ifeq ($(MAKECMDGOALS),dist)
+INSTALL-PROJECTS+=$(call cond_enabled,CONFIG_INSTALL_PROJECTS,DIST-PROJECTS)
+endif
+.install-projects: .projects $(addprefix $(DIST_DIR)/,$(INSTALL-PROJECTS))
+	@touch $@
+
+clean::
+	rm -f .projects .install-projects $(PROJECTS)
+endif
+
+# If there are any source files to be distributed, then include the build
+# system too.
+ifneq ($(call enabled,DIST-SRCS),)
+    DIST-SRCS-yes            += configure
+    DIST-SRCS-yes            += build/make/configure.sh
+    DIST-SRCS-yes            += build/make/gen_asm_deps.sh
+    DIST-SRCS-yes            += build/make/Makefile
+    DIST-SRCS-$(CONFIG_MSVS)  += build/make/gen_msvs_def.sh
+    DIST-SRCS-$(CONFIG_MSVS)  += build/make/gen_msvs_proj.sh
+    DIST-SRCS-$(CONFIG_MSVS)  += build/make/gen_msvs_sln.sh
+    DIST-SRCS-$(CONFIG_MSVS)  += build/make/gen_msvs_vcxproj.sh
+    DIST-SRCS-$(CONFIG_MSVS)  += build/make/msvs_common.sh
+    DIST-SRCS-$(CONFIG_RVCT) += build/make/armlink_adapter.sh
+    DIST-SRCS-$(ARCH_ARM)    += build/make/ads2gas.pl
+    DIST-SRCS-$(ARCH_ARM)    += build/make/ads2gas_apple.pl
+    DIST-SRCS-$(ARCH_ARM)    += build/make/ads2armasm_ms.pl
+    DIST-SRCS-$(ARCH_ARM)    += build/make/thumb.pm
+    DIST-SRCS-yes            += $(target:-$(TOOLCHAIN)=).mk
+endif
+INSTALL-SRCS := $(call cond_enabled,CONFIG_INSTALL_SRCS,INSTALL-SRCS)
+ifeq ($(MAKECMDGOALS),dist)
+INSTALL-SRCS += $(call cond_enabled,CONFIG_INSTALL_SRCS,DIST-SRCS)
+endif
+.install-srcs: $(addprefix $(DIST_DIR)/src/,$(INSTALL-SRCS))
+	@touch $@
+
+clean::
+	rm -f .install-srcs
+
+ifeq ($(CONFIG_EXTERNAL_BUILD),yes)
+    BUILD_TARGETS += .projects
+    INSTALL_TARGETS += .install-projects
+endif
+BUILD_TARGETS += .docs .libs .bins
+INSTALL_TARGETS += .install-docs .install-srcs .install-libs .install-bins
+all: $(BUILD_TARGETS)
+install:: $(INSTALL_TARGETS)
+dist: $(INSTALL_TARGETS)
+test::
diff --git a/libvpx/libvpx/build/make/ads2armasm_ms.pl b/libvpx/libvpx/build/make/ads2armasm_ms.pl
new file mode 100755
index 0000000..2a2c470
--- /dev/null
+++ b/libvpx/libvpx/build/make/ads2armasm_ms.pl
@@ -0,0 +1,39 @@
+#!/usr/bin/env perl
+##
+##  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+use FindBin;
+use lib $FindBin::Bin;
+use thumb;
+
+print "; This file was created from a .asm file\n";
+print ";  using the ads2armasm_ms.pl script.\n";
+
+while (<STDIN>)
+{
+    undef $comment;
+    undef $line;
+
+    s/REQUIRE8//;
+    s/PRESERVE8//;
+    s/^\s*ARM\s*$//;
+    s/AREA\s+\|\|(.*)\|\|/AREA |$1|/;
+    s/qsubaddx/qsax/i;
+    s/qaddsubx/qasx/i;
+
+    thumb::FixThumbInstructions($_, 1);
+
+    s/ldrneb/ldrbne/i;
+    s/ldrneh/ldrhne/i;
+    s/^(\s*)ENDP.*/$&\n$1ALIGN 4/;
+
+    print;
+}
+
diff --git a/libvpx/libvpx/build/make/ads2gas.pl b/libvpx/libvpx/build/make/ads2gas.pl
new file mode 100755
index 0000000..7272424
--- /dev/null
+++ b/libvpx/libvpx/build/make/ads2gas.pl
@@ -0,0 +1,236 @@
+#!/usr/bin/env perl
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+# ads2gas.pl
+# Author: Eric Fung (efung (at) acm.org)
+#
+# Convert ARM Developer Suite 1.0.1 syntax assembly source to GNU as format
+#
+# Usage: cat inputfile | perl ads2gas.pl > outputfile
+#
+
+use FindBin;
+use lib $FindBin::Bin;
+use thumb;
+
+my $thumb = 0;
+
+foreach my $arg (@ARGV) {
+    $thumb = 1 if ($arg eq "-thumb");
+}
+
+print "@ This file was created from a .asm file\n";
+print "@  using the ads2gas.pl script.\n";
+print "\t.equ DO1STROUNDING, 0\n";
+if ($thumb) {
+    print "\t.syntax unified\n";
+    print "\t.thumb\n";
+}
+
+# Stack of procedure names.
+@proc_stack = ();
+
+while (<STDIN>)
+{
+    undef $comment;
+    undef $line;
+    $comment_char = ";";
+    $comment_sub = "@";
+
+    # Handle comments.
+    if (/$comment_char/)
+    {
+      $comment = "";
+      ($line, $comment) = /(.*?)$comment_char(.*)/;
+      $_ = $line;
+    }
+
+    # Load and store alignment
+    s/@/,:/g;
+
+    # Hexadecimal constants prefaced by 0x
+    s/#&/#0x/g;
+
+    # Convert :OR: to |
+    s/:OR:/ | /g;
+
+    # Convert :AND: to &
+    s/:AND:/ & /g;
+
+    # Convert :NOT: to ~
+    s/:NOT:/ ~ /g;
+
+    # Convert :SHL: to <<
+    s/:SHL:/ << /g;
+
+    # Convert :SHR: to >>
+    s/:SHR:/ >> /g;
+
+    # Convert ELSE to .else
+    s/\bELSE\b/.else/g;
+
+    # Convert ENDIF to .endif
+    s/\bENDIF\b/.endif/g;
+
+    # Convert ELSEIF to .elseif
+    s/\bELSEIF\b/.elseif/g;
+
+    # Convert LTORG to .ltorg
+    s/\bLTORG\b/.ltorg/g;
+
+    # Convert endfunc to nothing.
+    s/\bendfunc\b//ig;
+
+    # Convert FUNCTION to nothing.
+    s/\bFUNCTION\b//g;
+    s/\bfunction\b//g;
+
+    s/\bENTRY\b//g;
+    s/\bMSARMASM\b/0/g;
+    s/^\s+end\s+$//g;
+
+    # Convert IF :DEF:to .if
+    # gcc doesn't have the ability to do a conditional
+    # if defined variable that is set by IF :DEF: on
+    # armasm, so convert it to a normal .if and then
+    # make sure to define a value elesewhere
+    if (s/\bIF :DEF:\b/.if /g)
+    {
+        s/=/==/g;
+    }
+
+    # Convert IF to .if
+    if (s/\bIF\b/.if/g)
+    {
+        s/=+/==/g;
+    }
+
+    # Convert INCLUDE to .INCLUDE "file"
+    s/INCLUDE(\s*)(.*)$/.include $1\"$2\"/;
+
+    # Code directive (ARM vs Thumb)
+    s/CODE([0-9][0-9])/.code $1/;
+
+    # No AREA required
+    # But ALIGNs in AREA must be obeyed
+    s/^\s*AREA.*ALIGN=([0-9])$/.text\n.p2align $1/;
+    # If no ALIGN, strip the AREA and align to 4 bytes
+    s/^\s*AREA.*$/.text\n.p2align 2/;
+
+    # DCD to .word
+    # This one is for incoming symbols
+    s/DCD\s+\|(\w*)\|/.long $1/;
+
+    # DCW to .short
+    s/DCW\s+\|(\w*)\|/.short $1/;
+    s/DCW(.*)/.short $1/;
+
+    # Constants defined in scope
+    s/DCD(.*)/.long $1/;
+    s/DCB(.*)/.byte $1/;
+
+    # RN to .req
+    if (s/RN\s+([Rr]\d+|lr)/.req $1/)
+    {
+        print;
+        print "$comment_sub$comment\n" if defined $comment;
+        next;
+    }
+
+    # Make function visible to linker, and make additional symbol with
+    # prepended underscore
+    s/EXPORT\s+\|([\$\w]*)\|/.global $1 \n\t.type $1, function/;
+    s/IMPORT\s+\|([\$\w]*)\|/.global $1/;
+
+    s/EXPORT\s+([\$\w]*)/.global $1/;
+    s/export\s+([\$\w]*)/.global $1/;
+
+    # No vertical bars required; make additional symbol with prepended
+    # underscore
+    s/^\|(\$?\w+)\|/_$1\n\t$1:/g;
+
+    # Labels need trailing colon
+#   s/^(\w+)/$1:/ if !/EQU/;
+    # put the colon at the end of the line in the macro
+    s/^([a-zA-Z_0-9\$]+)/$1:/ if !/EQU/;
+
+    # ALIGN directive
+    s/\bALIGN\b/.balign/g;
+
+    if ($thumb) {
+        # ARM code - we force everything to thumb with the declaration in the header
+        s/\sARM//g;
+    } else {
+        # ARM code
+        s/\sARM/.arm/g;
+    }
+
+    # push/pop
+    s/(push\s+)(r\d+)/stmdb sp\!, \{$2\}/g;
+    s/(pop\s+)(r\d+)/ldmia sp\!, \{$2\}/g;
+
+    # NEON code
+    s/(vld1.\d+\s+)(q\d+)/$1\{$2\}/g;
+    s/(vtbl.\d+\s+[^,]+),([^,]+)/$1,\{$2\}/g;
+
+    if ($thumb) {
+        thumb::FixThumbInstructions($_, 0);
+    }
+
+    # eabi_attributes numerical equivalents can be found in the
+    # "ARM IHI 0045C" document.
+
+    # REQUIRE8 Stack is required to be 8-byte aligned
+    s/\sREQUIRE8/.eabi_attribute 24, 1 \@Tag_ABI_align_needed/g;
+
+    # PRESERVE8 Stack 8-byte align is preserved
+    s/\sPRESERVE8/.eabi_attribute 25, 1 \@Tag_ABI_align_preserved/g;
+
+    # Use PROC and ENDP to give the symbols a .size directive.
+    # This makes them show up properly in debugging tools like gdb and valgrind.
+    if (/\bPROC\b/)
+    {
+        my $proc;
+        /^_([\.0-9A-Z_a-z]\w+)\b/;
+        $proc = $1;
+        push(@proc_stack, $proc) if ($proc);
+        s/\bPROC\b/@ $&/;
+    }
+    if (/\bENDP\b/)
+    {
+        my $proc;
+        s/\bENDP\b/@ $&/;
+        $proc = pop(@proc_stack);
+        $_ = "\t.size $proc, .-$proc".$_ if ($proc);
+    }
+
+    # EQU directive
+    s/(\S+\s+)EQU(\s+\S+)/.equ $1, $2/;
+
+    # Begin macro definition
+    if (/\bMACRO\b/) {
+        $_ = <STDIN>;
+        s/^/.macro/;
+        s/\$//g;                # remove formal param reference
+        s/;/@/g;                # change comment characters
+    }
+
+    # For macros, use \ to reference formal params
+    s/\$/\\/g;                  # End macro definition
+    s/\bMEND\b/.endm/;              # No need to tell it where to stop assembling
+    next if /^\s*END\s*$/;
+    print;
+    print "$comment_sub$comment\n" if defined $comment;
+}
+
+# Mark that this object doesn't need an executable stack.
+printf ("\t.section\t.note.GNU-stack,\"\",\%\%progbits\n");
diff --git a/libvpx/libvpx/build/make/ads2gas_apple.pl b/libvpx/libvpx/build/make/ads2gas_apple.pl
new file mode 100755
index 0000000..a82f3eb
--- /dev/null
+++ b/libvpx/libvpx/build/make/ads2gas_apple.pl
@@ -0,0 +1,235 @@
+#!/usr/bin/env perl
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+# ads2gas_apple.pl
+# Author: Eric Fung (efung (at) acm.org)
+#
+# Convert ARM Developer Suite 1.0.1 syntax assembly source to GNU as format
+#
+# Usage: cat inputfile | perl ads2gas_apple.pl > outputfile
+#
+
+my $chromium = 0;
+
+foreach my $arg (@ARGV) {
+    $chromium = 1 if ($arg eq "-chromium");
+}
+
+print "@ This file was created from a .asm file\n";
+print "@  using the ads2gas_apple.pl script.\n\n";
+print "\t.set WIDE_REFERENCE, 0\n";
+print "\t.set ARCHITECTURE, 5\n";
+print "\t.set DO1STROUNDING, 0\n";
+
+my %register_aliases;
+my %macro_aliases;
+
+my @mapping_list = ("\$0", "\$1", "\$2", "\$3", "\$4", "\$5", "\$6", "\$7", "\$8", "\$9");
+
+my @incoming_array;
+
+my @imported_functions;
+
+# Perl trim function to remove whitespace from the start and end of the string
+sub trim($)
+{
+    my $string = shift;
+    $string =~ s/^\s+//;
+    $string =~ s/\s+$//;
+    return $string;
+}
+
+while (<STDIN>)
+{
+    # Load and store alignment
+    s/@/,:/g;
+
+    # Comment character
+    s/;/ @/g;
+
+    # Hexadecimal constants prefaced by 0x
+    s/#&/#0x/g;
+
+    # Convert :OR: to |
+    s/:OR:/ | /g;
+
+    # Convert :AND: to &
+    s/:AND:/ & /g;
+
+    # Convert :NOT: to ~
+    s/:NOT:/ ~ /g;
+
+    # Convert :SHL: to <<
+    s/:SHL:/ << /g;
+
+    # Convert :SHR: to >>
+    s/:SHR:/ >> /g;
+
+    # Convert ELSE to .else
+    s/\bELSE\b/.else/g;
+
+    # Convert ENDIF to .endif
+    s/\bENDIF\b/.endif/g;
+
+    # Convert ELSEIF to .elseif
+    s/\bELSEIF\b/.elseif/g;
+
+    # Convert LTORG to .ltorg
+    s/\bLTORG\b/.ltorg/g;
+
+    # Convert IF :DEF:to .if
+    # gcc doesn't have the ability to do a conditional
+    # if defined variable that is set by IF :DEF: on
+    # armasm, so convert it to a normal .if and then
+    # make sure to define a value elesewhere
+    if (s/\bIF :DEF:\b/.if /g)
+    {
+        s/=/==/g;
+    }
+
+    # Convert IF to .if
+    if (s/\bIF\b/.if/g)
+    {
+        s/=/==/g;
+    }
+
+    # Convert INCLUDE to .INCLUDE "file"
+    s/INCLUDE(\s*)(.*)$/.include $1\"$2\"/;
+
+    # Code directive (ARM vs Thumb)
+    s/CODE([0-9][0-9])/.code $1/;
+
+    # No AREA required
+    # But ALIGNs in AREA must be obeyed
+    s/^\s*AREA.*ALIGN=([0-9])$/.text\n.p2align $1/;
+    # If no ALIGN, strip the AREA and align to 4 bytes
+    s/^\s*AREA.*$/.text\n.p2align 2/;
+
+    # DCD to .word
+    # This one is for incoming symbols
+    s/DCD\s+\|(\w*)\|/.long $1/;
+
+    # DCW to .short
+    s/DCW\s+\|(\w*)\|/.short $1/;
+    s/DCW(.*)/.short $1/;
+
+    # Constants defined in scope
+    s/DCD(.*)/.long $1/;
+    s/DCB(.*)/.byte $1/;
+
+    # Build a hash of all the register - alias pairs.
+    if (s/(.*)RN(.*)/$1 .req $2/g)
+    {
+        $register_aliases{trim($1)} = trim($2);
+        next;
+    }
+
+    while (($key, $value) = each(%register_aliases))
+    {
+        s/\b$key\b/$value/g;
+    }
+
+    # Make function visible to linker, and make additional symbol with
+    # prepended underscore
+    s/EXPORT\s+\|([\$\w]*)\|/.globl _$1\n\t.globl $1/;
+
+    # Prepend imported functions with _
+    if (s/IMPORT\s+\|([\$\w]*)\|/.globl $1/)
+    {
+        $function = trim($1);
+        push(@imported_functions, $function);
+    }
+
+    foreach $function (@imported_functions)
+    {
+        s/$function/_$function/;
+    }
+
+    # No vertical bars required; make additional symbol with prepended
+    # underscore
+    s/^\|(\$?\w+)\|/_$1\n\t$1:/g;
+
+    # Labels need trailing colon
+#   s/^(\w+)/$1:/ if !/EQU/;
+    # put the colon at the end of the line in the macro
+    s/^([a-zA-Z_0-9\$]+)/$1:/ if !/EQU/;
+
+    # ALIGN directive
+    s/\bALIGN\b/.balign/g;
+
+    # Strip ARM
+    s/\sARM/@ ARM/g;
+
+    # Strip REQUIRE8
+    #s/\sREQUIRE8/@ REQUIRE8/g;
+    s/\sREQUIRE8/@ /g;
+
+    # Strip PRESERVE8
+    s/\sPRESERVE8/@ PRESERVE8/g;
+
+    # Strip PROC and ENDPROC
+    s/\bPROC\b/@/g;
+    s/\bENDP\b/@/g;
+
+    # EQU directive
+    s/(.*)EQU(.*)/.set $1, $2/;
+
+    # Begin macro definition
+    if (/\bMACRO\b/)
+    {
+        # Process next line down, which will be the macro definition
+        $_ = <STDIN>;
+
+        $trimmed = trim($_);
+
+        # remove commas that are separating list
+        $trimmed =~ s/,//g;
+
+        # string to array
+        @incoming_array = split(/\s+/, $trimmed);
+
+        print ".macro @incoming_array[0]\n";
+
+        # remove the first element, as that is the name of the macro
+        shift (@incoming_array);
+
+        @macro_aliases{@incoming_array} = @mapping_list;
+
+        next;
+    }
+
+    while (($key, $value) = each(%macro_aliases))
+    {
+        $key =~ s/\$/\\\$/;
+        s/$key\b/$value/g;
+    }
+
+    # For macros, use \ to reference formal params
+#   s/\$/\\/g;                  # End macro definition
+    s/\bMEND\b/.endm/;              # No need to tell it where to stop assembling
+    next if /^\s*END\s*$/;
+
+    # Clang used by Chromium differs slightly from clang in XCode in what it
+    # will accept in the assembly.
+    if ($chromium) {
+        s/qsubaddx/qsax/i;
+        s/qaddsubx/qasx/i;
+        s/ldrneb/ldrbne/i;
+        s/ldrneh/ldrhne/i;
+        s/(vqshrun\.s16 .*, \#)0$/${1}8/i;
+
+        # http://llvm.org/bugs/show_bug.cgi?id=16022
+        s/\.include/#include/;
+    }
+
+    print;
+}
diff --git a/libvpx/libvpx/build/make/armlink_adapter.sh b/libvpx/libvpx/build/make/armlink_adapter.sh
new file mode 100755
index 0000000..75c342e
--- /dev/null
+++ b/libvpx/libvpx/build/make/armlink_adapter.sh
@@ -0,0 +1,54 @@
+#!/bin/sh
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+verbose=0
+set -- $*
+for i; do
+    if [ "$i" = "-o" ]; then
+        on_of=1
+    elif [ "$i" = "-v" ]; then
+        verbose=1
+    elif [ "$i" = "-g" ]; then
+        args="${args} --debug"
+    elif [ "$on_of" = "1" ]; then
+        outfile=$i
+        on_of=0
+    elif [ -f "$i" ]; then
+        infiles="$infiles $i"
+    elif [ "${i#-l}" != "$i" ]; then
+        libs="$libs ${i#-l}"
+    elif [ "${i#-L}" != "$i" ]; then
+        libpaths="${libpaths} ${i#-L}"
+    else
+        args="${args} ${i}"
+    fi
+    shift
+done
+
+# Absolutize library file names
+for f in $libs; do
+    found=0
+    for d in $libpaths; do
+        [ -f "$d/$f" ] && infiles="$infiles $d/$f" && found=1 && break
+        [ -f "$d/lib${f}.so" ] && infiles="$infiles $d/lib${f}.so" && found=1 && break
+        [ -f "$d/lib${f}.a" ] && infiles="$infiles $d/lib${f}.a" && found=1 && break
+    done
+    [ $found -eq 0 ] && infiles="$infiles $f"
+done
+for d in $libpaths; do
+    [ -n "$libsearchpath" ] && libsearchpath="${libsearchpath},"
+    libsearchpath="${libsearchpath}$d"
+done
+
+cmd="armlink $args --userlibpath=$libsearchpath --output=$outfile $infiles"
+[ $verbose -eq 1 ] && echo $cmd
+$cmd
diff --git a/libvpx/libvpx/build/make/configure.sh b/libvpx/libvpx/build/make/configure.sh
new file mode 100755
index 0000000..4f0071b
--- /dev/null
+++ b/libvpx/libvpx/build/make/configure.sh
@@ -0,0 +1,1560 @@
+#!/bin/sh
+##
+##  configure.sh
+##
+##  This script is sourced by the main configure script and contains
+##  utility functions and other common bits that aren't strictly libvpx
+##  related.
+##
+##  This build system is based in part on the FFmpeg configure script.
+##
+
+
+#
+# Logging / Output Functions
+#
+die_unknown(){
+  echo "Unknown option \"$1\"."
+  echo "See $0 --help for available options."
+  clean_temp_files
+  exit 1
+}
+
+die() {
+  echo "$@"
+  echo
+  echo "Configuration failed. This could reflect a misconfiguration of your"
+  echo "toolchains, improper options selected, or another problem. If you"
+  echo "don't see any useful error messages above, the next step is to look"
+  echo "at the configure error log file ($logfile) to determine what"
+  echo "configure was trying to do when it died."
+  clean_temp_files
+  exit 1
+}
+
+log(){
+  echo "$@" >>$logfile
+}
+
+log_file(){
+  log BEGIN $1
+  cat -n $1 >>$logfile
+  log END $1
+}
+
+log_echo() {
+  echo "$@"
+  log "$@"
+}
+
+fwrite () {
+  outfile=$1
+  shift
+  echo "$@" >> ${outfile}
+}
+
+show_help_pre(){
+  for opt in ${CMDLINE_SELECT}; do
+    opt2=`echo $opt | sed -e 's;_;-;g'`
+    if enabled $opt; then
+      eval "toggle_${opt}=\"--disable-${opt2}\""
+    else
+      eval "toggle_${opt}=\"--enable-${opt2} \""
+    fi
+  done
+
+  cat <<EOF
+Usage: configure [options]
+Options:
+
+Build options:
+  --help                      print this message
+  --log=yes|no|FILE           file configure log is written to [config.log]
+  --target=TARGET             target platform tuple [generic-gnu]
+  --cpu=CPU                   optimize for a specific cpu rather than a family
+  --extra-cflags=ECFLAGS      add ECFLAGS to CFLAGS [$CFLAGS]
+  --extra-cxxflags=ECXXFLAGS  add ECXXFLAGS to CXXFLAGS [$CXXFLAGS]
+  ${toggle_extra_warnings}    emit harmless warnings (always non-fatal)
+  ${toggle_werror}            treat warnings as errors, if possible
+                              (not available with all compilers)
+  ${toggle_optimizations}     turn on/off compiler optimization flags
+  ${toggle_pic}               turn on/off Position Independent Code
+  ${toggle_ccache}            turn on/off compiler cache
+  ${toggle_debug}             enable/disable debug mode
+  ${toggle_gprof}             enable/disable gprof profiling instrumentation
+  ${toggle_gcov}              enable/disable gcov coverage instrumentation
+  ${toggle_thumb}             enable/disable building arm assembly in thumb mode
+  ${toggle_dependency_tracking}
+                              disable to speed up one-time build
+
+Install options:
+  ${toggle_install_docs}      control whether docs are installed
+  ${toggle_install_bins}      control whether binaries are installed
+  ${toggle_install_libs}      control whether libraries are installed
+  ${toggle_install_srcs}      control whether sources are installed
+
+
+EOF
+}
+
+show_help_post(){
+  cat <<EOF
+
+
+NOTES:
+    Object files are built at the place where configure is launched.
+
+    All boolean options can be negated. The default value is the opposite
+    of that shown above. If the option --disable-foo is listed, then
+    the default value for foo is enabled.
+
+Supported targets:
+EOF
+  show_targets ${all_platforms}
+  echo
+  exit 1
+}
+
+show_targets() {
+  while [ -n "$*" ]; do
+    if [ "${1%%-*}" = "${2%%-*}" ]; then
+      if [ "${2%%-*}" = "${3%%-*}" ]; then
+        printf "    %-24s %-24s %-24s\n" "$1" "$2" "$3"
+        shift; shift; shift
+      else
+        printf "    %-24s %-24s\n" "$1" "$2"
+        shift; shift
+      fi
+    else
+      printf "    %-24s\n" "$1"
+      shift
+    fi
+  done
+}
+
+show_help() {
+  show_help_pre
+  show_help_post
+}
+
+#
+# List Processing Functions
+#
+set_all(){
+  value=$1
+  shift
+  for var in $*; do
+    eval $var=$value
+  done
+}
+
+is_in(){
+  value=$1
+  shift
+  for var in $*; do
+    [ $var = $value ] && return 0
+  done
+  return 1
+}
+
+add_cflags() {
+  CFLAGS="${CFLAGS} $@"
+  CXXFLAGS="${CXXFLAGS} $@"
+}
+
+add_cflags_only() {
+  CFLAGS="${CFLAGS} $@"
+}
+
+add_cxxflags_only() {
+  CXXFLAGS="${CXXFLAGS} $@"
+}
+
+add_ldflags() {
+  LDFLAGS="${LDFLAGS} $@"
+}
+
+add_asflags() {
+  ASFLAGS="${ASFLAGS} $@"
+}
+
+add_extralibs() {
+  extralibs="${extralibs} $@"
+}
+
+#
+# Boolean Manipulation Functions
+#
+
+enable_feature(){
+  set_all yes $*
+}
+
+disable_feature(){
+  set_all no $*
+}
+
+enabled(){
+  eval test "x\$$1" = "xyes"
+}
+
+disabled(){
+  eval test "x\$$1" = "xno"
+}
+
+enable_codec(){
+  enabled "${1}" || echo "  enabling ${1}"
+  enable_feature "${1}"
+
+  is_in "${1}" vp8 vp9 && enable_feature "${1}_encoder" "${1}_decoder"
+}
+
+disable_codec(){
+  disabled "${1}" || echo "  disabling ${1}"
+  disable_feature "${1}"
+
+  is_in "${1}" vp8 vp9 && disable_feature "${1}_encoder" "${1}_decoder"
+}
+
+# Iterates through positional parameters, checks to confirm the parameter has
+# not been explicitly (force) disabled, and enables the setting controlled by
+# the parameter when the setting is not disabled.
+# Note: Does NOT alter RTCD generation options ($RTCD_OPTIONS).
+soft_enable() {
+  for var in $*; do
+    if ! disabled $var; then
+      enabled $var || log_echo "  enabling $var"
+      enable_feature $var
+    fi
+  done
+}
+
+# Iterates through positional parameters, checks to confirm the parameter has
+# not been explicitly (force) enabled, and disables the setting controlled by
+# the parameter when the setting is not enabled.
+# Note: Does NOT alter RTCD generation options ($RTCD_OPTIONS).
+soft_disable() {
+  for var in $*; do
+    if ! enabled $var; then
+      disabled $var || log_echo "  disabling $var"
+      disable_feature $var
+    fi
+  done
+}
+
+#
+# Text Processing Functions
+#
+toupper(){
+  echo "$@" | tr abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
+}
+
+tolower(){
+  echo "$@" | tr ABCDEFGHIJKLMNOPQRSTUVWXYZ abcdefghijklmnopqrstuvwxyz
+}
+
+#
+# Temporary File Functions
+#
+source_path=${0%/*}
+enable_feature source_path_used
+if [ -z "$source_path" ] || [ "$source_path" = "." ]; then
+  source_path="`pwd`"
+  disable_feature source_path_used
+fi
+
+if test ! -z "$TMPDIR" ; then
+  TMPDIRx="${TMPDIR}"
+elif test ! -z "$TEMPDIR" ; then
+  TMPDIRx="${TEMPDIR}"
+else
+  TMPDIRx="/tmp"
+fi
+RAND=$(awk 'BEGIN { srand(); printf "%d\n",(rand() * 32768)}')
+TMP_H="${TMPDIRx}/vpx-conf-$$-${RAND}.h"
+TMP_C="${TMPDIRx}/vpx-conf-$$-${RAND}.c"
+TMP_CC="${TMPDIRx}/vpx-conf-$$-${RAND}.cc"
+TMP_O="${TMPDIRx}/vpx-conf-$$-${RAND}.o"
+TMP_X="${TMPDIRx}/vpx-conf-$$-${RAND}.x"
+TMP_ASM="${TMPDIRx}/vpx-conf-$$-${RAND}.asm"
+
+clean_temp_files() {
+  rm -f ${TMP_C} ${TMP_CC} ${TMP_H} ${TMP_O} ${TMP_X} ${TMP_ASM}
+  enabled gcov && rm -f ${TMP_C%.c}.gcno ${TMP_CC%.cc}.gcno
+}
+
+#
+# Toolchain Check Functions
+#
+check_cmd() {
+  enabled external_build && return
+  log "$@"
+  "$@" >>${logfile} 2>&1
+}
+
+check_cc() {
+  log check_cc "$@"
+  cat >${TMP_C}
+  log_file ${TMP_C}
+  check_cmd ${CC} ${CFLAGS} "$@" -c -o ${TMP_O} ${TMP_C}
+}
+
+check_cxx() {
+  log check_cxx "$@"
+  cat >${TMP_CC}
+  log_file ${TMP_CC}
+  check_cmd ${CXX} ${CXXFLAGS} "$@" -c -o ${TMP_O} ${TMP_CC}
+}
+
+check_cpp() {
+  log check_cpp "$@"
+  cat > ${TMP_C}
+  log_file ${TMP_C}
+  check_cmd ${CC} ${CFLAGS} "$@" -E -o ${TMP_O} ${TMP_C}
+}
+
+check_ld() {
+  log check_ld "$@"
+  check_cc $@ \
+    && check_cmd ${LD} ${LDFLAGS} "$@" -o ${TMP_X} ${TMP_O} ${extralibs}
+}
+
+check_header(){
+  log check_header "$@"
+  header=$1
+  shift
+  var=`echo $header | sed 's/[^A-Za-z0-9_]/_/g'`
+  disable_feature $var
+  check_cpp "$@" <<EOF && enable_feature $var
+#include "$header"
+int x;
+EOF
+}
+
+check_cflags() {
+ log check_cflags "$@"
+ check_cc -Werror "$@" <<EOF
+int x;
+EOF
+}
+
+check_cxxflags() {
+  log check_cxxflags "$@"
+
+  # Catch CFLAGS that trigger CXX warnings
+  case "$CXX" in
+    *c++-analyzer|*clang++|*g++*)
+      check_cxx -Werror "$@" <<EOF
+int x;
+EOF
+      ;;
+    *)
+      check_cxx -Werror "$@" <<EOF
+int x;
+EOF
+      ;;
+    esac
+}
+
+check_add_cflags() {
+  check_cxxflags "$@" && add_cxxflags_only "$@"
+  check_cflags "$@" && add_cflags_only "$@"
+}
+
+check_add_cxxflags() {
+  check_cxxflags "$@" && add_cxxflags_only "$@"
+}
+
+check_add_asflags() {
+  log add_asflags "$@"
+  add_asflags "$@"
+}
+
+check_add_ldflags() {
+  log add_ldflags "$@"
+  add_ldflags "$@"
+}
+
+check_asm_align() {
+  log check_asm_align "$@"
+  cat >${TMP_ASM} <<EOF
+section .rodata
+align 16
+EOF
+  log_file ${TMP_ASM}
+  check_cmd ${AS} ${ASFLAGS} -o ${TMP_O} ${TMP_ASM}
+  readelf -WS ${TMP_O} >${TMP_X}
+  log_file ${TMP_X}
+  if ! grep -q '\.rodata .* 16$' ${TMP_X}; then
+    die "${AS} ${ASFLAGS} does not support section alignment (nasm <=2.08?)"
+  fi
+}
+
+# tests for -m$1 toggling the feature given in $2. If $2 is empty $1 is used.
+check_gcc_machine_option() {
+  opt="$1"
+  feature="$2"
+  [ -n "$feature" ] || feature="$opt"
+
+  if enabled gcc && ! disabled "$feature" && ! check_cflags "-m$opt"; then
+    RTCD_OPTIONS="${RTCD_OPTIONS}--disable-$feature "
+  else
+    soft_enable "$feature"
+  fi
+}
+
+write_common_config_banner() {
+  print_webm_license config.mk "##" ""
+  echo '# This file automatically generated by configure. Do not edit!' >> config.mk
+  echo "TOOLCHAIN := ${toolchain}" >> config.mk
+
+  case ${toolchain} in
+    *-linux-rvct)
+      echo "ALT_LIBC := ${alt_libc}" >> config.mk
+      ;;
+  esac
+}
+
+write_common_config_targets() {
+  for t in ${all_targets}; do
+    if enabled ${t}; then
+      if enabled child; then
+        fwrite config.mk "ALL_TARGETS += ${t}-${toolchain}"
+      else
+        fwrite config.mk "ALL_TARGETS += ${t}"
+      fi
+    fi
+    true;
+  done
+  true
+}
+
+write_common_target_config_mk() {
+  saved_CC="${CC}"
+  saved_CXX="${CXX}"
+  enabled ccache && CC="ccache ${CC}"
+  enabled ccache && CXX="ccache ${CXX}"
+  print_webm_license $1 "##" ""
+
+  cat >> $1 << EOF
+# This file automatically generated by configure. Do not edit!
+SRC_PATH="$source_path"
+SRC_PATH_BARE=$source_path
+BUILD_PFX=${BUILD_PFX}
+TOOLCHAIN=${toolchain}
+ASM_CONVERSION=${asm_conversion_cmd:-${source_path}/build/make/ads2gas.pl}
+GEN_VCPROJ=${gen_vcproj_cmd}
+MSVS_ARCH_DIR=${msvs_arch_dir}
+
+CC=${CC}
+CXX=${CXX}
+AR=${AR}
+LD=${LD}
+AS=${AS}
+STRIP=${STRIP}
+NM=${NM}
+
+CFLAGS  = ${CFLAGS}
+CXXFLAGS  = ${CXXFLAGS}
+ARFLAGS = -crs\$(if \$(quiet),,v)
+LDFLAGS = ${LDFLAGS}
+ASFLAGS = ${ASFLAGS}
+extralibs = ${extralibs}
+AS_SFX    = ${AS_SFX:-.asm}
+EXE_SFX   = ${EXE_SFX}
+VCPROJ_SFX = ${VCPROJ_SFX}
+RTCD_OPTIONS = ${RTCD_OPTIONS}
+EOF
+
+  if enabled rvct; then cat >> $1 << EOF
+fmt_deps = sed -e 's;^__image.axf;\${@:.d=.o} \$@;' #hide
+EOF
+  else cat >> $1 << EOF
+fmt_deps = sed -e 's;^\([a-zA-Z0-9_]*\)\.o;\${@:.d=.o} \$@;'
+EOF
+  fi
+
+  print_config_mk ARCH   "${1}" ${ARCH_LIST}
+  print_config_mk HAVE   "${1}" ${HAVE_LIST}
+  print_config_mk CONFIG "${1}" ${CONFIG_LIST}
+  print_config_mk HAVE   "${1}" gnu_strip
+
+  enabled msvs && echo "CONFIG_VS_VERSION=${vs_version}" >> "${1}"
+
+  CC="${saved_CC}"
+  CXX="${saved_CXX}"
+}
+
+write_common_target_config_h() {
+  print_webm_license ${TMP_H} "/*" " */"
+  cat >> ${TMP_H} << EOF
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    ${RESTRICT}
+#define INLINE      ${INLINE}
+EOF
+  print_config_h ARCH   "${TMP_H}" ${ARCH_LIST}
+  print_config_h HAVE   "${TMP_H}" ${HAVE_LIST}
+  print_config_h CONFIG "${TMP_H}" ${CONFIG_LIST}
+  print_config_vars_h   "${TMP_H}" ${VAR_LIST}
+  echo "#endif /* VPX_CONFIG_H */" >> ${TMP_H}
+  mkdir -p `dirname "$1"`
+  cmp "$1" ${TMP_H} >/dev/null 2>&1 || mv ${TMP_H} "$1"
+}
+
+process_common_cmdline() {
+  for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+      --child)
+        enable_feature child
+        ;;
+      --log*)
+        logging="$optval"
+        if ! disabled logging ; then
+          enabled logging || logfile="$logging"
+        else
+          logfile=/dev/null
+        fi
+        ;;
+      --target=*)
+        toolchain="${toolchain:-${optval}}"
+        ;;
+      --force-target=*)
+        toolchain="${toolchain:-${optval}}"
+        enable_feature force_toolchain
+        ;;
+      --cpu=*)
+        tune_cpu="$optval"
+        ;;
+      --extra-cflags=*)
+        extra_cflags="${optval}"
+        ;;
+      --extra-cxxflags=*)
+        extra_cxxflags="${optval}"
+        ;;
+      --enable-?*|--disable-?*)
+        eval `echo "$opt" | sed 's/--/action=/;s/-/ option=/;s/-/_/g'`
+        if is_in ${option} ${ARCH_EXT_LIST}; then
+          [ $action = "disable" ] && RTCD_OPTIONS="${RTCD_OPTIONS}--disable-${option} "
+        elif [ $action = "disable" ] && ! disabled $option ; then
+          is_in ${option} ${CMDLINE_SELECT} || die_unknown $opt
+          log_echo "  disabling $option"
+        elif [ $action = "enable" ] && ! enabled $option ; then
+          is_in ${option} ${CMDLINE_SELECT} || die_unknown $opt
+          log_echo "  enabling $option"
+        fi
+        ${action}_feature $option
+        ;;
+      --require-?*)
+        eval `echo "$opt" | sed 's/--/action=/;s/-/ option=/;s/-/_/g'`
+        if is_in ${option} ${ARCH_EXT_LIST}; then
+            RTCD_OPTIONS="${RTCD_OPTIONS}${opt} "
+        else
+            die_unknown $opt
+        fi
+        ;;
+      --force-enable-?*|--force-disable-?*)
+        eval `echo "$opt" | sed 's/--force-/action=/;s/-/ option=/;s/-/_/g'`
+        ${action}_feature $option
+        ;;
+      --libc=*)
+        [ -d "${optval}" ] || die "Not a directory: ${optval}"
+        disable_feature builtin_libc
+        alt_libc="${optval}"
+        ;;
+      --as=*)
+        [ "${optval}" = yasm ] || [ "${optval}" = nasm ] \
+          || [ "${optval}" = auto ] \
+          || die "Must be yasm, nasm or auto: ${optval}"
+        alt_as="${optval}"
+        ;;
+      --size-limit=*)
+        w="${optval%%x*}"
+        h="${optval##*x}"
+        VAR_LIST="DECODE_WIDTH_LIMIT ${w} DECODE_HEIGHT_LIMIT ${h}"
+        [ ${w} -gt 0 ] && [ ${h} -gt 0 ] || die "Invalid size-limit: too small."
+        [ ${w} -lt 65536 ] && [ ${h} -lt 65536 ] \
+            || die "Invalid size-limit: too big."
+        enable_feature size_limit
+        ;;
+      --prefix=*)
+        prefix="${optval}"
+        ;;
+      --libdir=*)
+        libdir="${optval}"
+        ;;
+      --sdk-path=*)
+        [ -d "${optval}" ] || die "Not a directory: ${optval}"
+        sdk_path="${optval}"
+        ;;
+      --libc|--as|--prefix|--libdir|--sdk-path)
+        die "Option ${opt} requires argument"
+        ;;
+      --help|-h)
+        show_help
+        ;;
+      *)
+        die_unknown $opt
+        ;;
+    esac
+  done
+}
+
+process_cmdline() {
+  for opt do
+    optval="${opt#*=}"
+    case "$opt" in
+      *)
+        process_common_cmdline $opt
+        ;;
+    esac
+  done
+}
+
+post_process_common_cmdline() {
+  prefix="${prefix:-/usr/local}"
+  prefix="${prefix%/}"
+  libdir="${libdir:-${prefix}/lib}"
+  libdir="${libdir%/}"
+  if [ "${libdir#${prefix}}" = "${libdir}" ]; then
+    die "Libdir ${libdir} must be a subdirectory of ${prefix}"
+  fi
+}
+
+post_process_cmdline() {
+  true;
+}
+
+setup_gnu_toolchain() {
+  CC=${CC:-${CROSS}gcc}
+  CXX=${CXX:-${CROSS}g++}
+  AR=${AR:-${CROSS}ar}
+  LD=${LD:-${CROSS}${link_with_cc:-ld}}
+  AS=${AS:-${CROSS}as}
+  STRIP=${STRIP:-${CROSS}strip}
+  NM=${NM:-${CROSS}nm}
+  AS_SFX=.s
+  EXE_SFX=
+}
+
+# Reliably find the newest available Darwin SDKs. (Older versions of
+# xcrun don't support --show-sdk-path.)
+show_darwin_sdk_path() {
+  xcrun --sdk $1 --show-sdk-path 2>/dev/null ||
+    xcodebuild -sdk $1 -version Path 2>/dev/null
+}
+
+# Print the major version number of the Darwin SDK specified by $1.
+show_darwin_sdk_major_version() {
+  xcrun --sdk $1 --show-sdk-version 2>/dev/null | cut -d. -f1
+}
+
+# Print the Xcode version.
+show_xcode_version() {
+  xcodebuild -version | head -n1 | cut -d' ' -f2
+}
+
+# Fails when Xcode version is less than 6.3.
+check_xcode_minimum_version() {
+  xcode_major=$(show_xcode_version | cut -f1 -d.)
+  xcode_minor=$(show_xcode_version | cut -f2 -d.)
+  xcode_min_major=6
+  xcode_min_minor=3
+  if [ ${xcode_major} -lt ${xcode_min_major} ]; then
+    return 1
+  fi
+  if [ ${xcode_major} -eq ${xcode_min_major} ] \
+    && [ ${xcode_minor} -lt ${xcode_min_minor} ]; then
+    return 1
+  fi
+}
+
+process_common_toolchain() {
+  if [ -z "$toolchain" ]; then
+    gcctarget="${CHOST:-$(gcc -dumpmachine 2> /dev/null)}"
+
+    # detect tgt_isa
+    case "$gcctarget" in
+      aarch64*)
+        tgt_isa=arm64
+        ;;
+      armv6*)
+        tgt_isa=armv6
+        ;;
+      armv7*-hardfloat* | armv7*-gnueabihf | arm-*-gnueabihf)
+        tgt_isa=armv7
+        float_abi=hard
+        ;;
+      armv7*)
+        tgt_isa=armv7
+        float_abi=softfp
+        ;;
+      *x86_64*|*amd64*)
+        tgt_isa=x86_64
+        ;;
+      *i[3456]86*)
+        tgt_isa=x86
+        ;;
+      *sparc*)
+        tgt_isa=sparc
+        ;;
+    esac
+
+    # detect tgt_os
+    case "$gcctarget" in
+      *darwin10*)
+        tgt_isa=x86_64
+        tgt_os=darwin10
+        ;;
+      *darwin11*)
+        tgt_isa=x86_64
+        tgt_os=darwin11
+        ;;
+      *darwin12*)
+        tgt_isa=x86_64
+        tgt_os=darwin12
+        ;;
+      *darwin13*)
+        tgt_isa=x86_64
+        tgt_os=darwin13
+        ;;
+      *darwin14*)
+        tgt_isa=x86_64
+        tgt_os=darwin14
+        ;;
+      *darwin15*)
+        tgt_isa=x86_64
+        tgt_os=darwin15
+        ;;
+      x86_64*mingw32*)
+        tgt_os=win64
+        ;;
+      *mingw32*|*cygwin*)
+        [ -z "$tgt_isa" ] && tgt_isa=x86
+        tgt_os=win32
+        ;;
+      *linux*|*bsd*)
+        tgt_os=linux
+        ;;
+      *solaris2.10)
+        tgt_os=solaris
+        ;;
+      *os2*)
+        tgt_os=os2
+        ;;
+    esac
+
+    if [ -n "$tgt_isa" ] && [ -n "$tgt_os" ]; then
+      toolchain=${tgt_isa}-${tgt_os}-gcc
+    fi
+  fi
+
+  toolchain=${toolchain:-generic-gnu}
+
+  is_in ${toolchain} ${all_platforms} || enabled force_toolchain \
+    || die "Unrecognized toolchain '${toolchain}'"
+
+  enabled child || log_echo "Configuring for target '${toolchain}'"
+
+  #
+  # Set up toolchain variables
+  #
+  tgt_isa=$(echo ${toolchain} | awk 'BEGIN{FS="-"}{print $1}')
+  tgt_os=$(echo ${toolchain} | awk 'BEGIN{FS="-"}{print $2}')
+  tgt_cc=$(echo ${toolchain} | awk 'BEGIN{FS="-"}{print $3}')
+
+  # Mark the specific ISA requested as enabled
+  soft_enable ${tgt_isa}
+  enable_feature ${tgt_os}
+  enable_feature ${tgt_cc}
+
+  # Enable the architecture family
+  case ${tgt_isa} in
+    arm*)
+      enable_feature arm
+      ;;
+    mips*)
+      enable_feature mips
+      ;;
+  esac
+
+  # PIC is probably what we want when building shared libs
+  enabled shared && soft_enable pic
+
+  # Minimum iOS version for all target platforms (darwin and iphonesimulator).
+  # Shared library framework builds are only possible on iOS 8 and later.
+  if enabled shared; then
+    IOS_VERSION_OPTIONS="--enable-shared"
+    IOS_VERSION_MIN="8.0"
+  else
+    IOS_VERSION_OPTIONS=""
+    IOS_VERSION_MIN="6.0"
+  fi
+
+  # Handle darwin variants. Newer SDKs allow targeting older
+  # platforms, so use the newest one available.
+  case ${toolchain} in
+    arm*-darwin*)
+      add_cflags "-miphoneos-version-min=${IOS_VERSION_MIN}"
+      iphoneos_sdk_dir="$(show_darwin_sdk_path iphoneos)"
+      if [ -d "${iphoneos_sdk_dir}" ]; then
+        add_cflags  "-isysroot ${iphoneos_sdk_dir}"
+        add_ldflags "-isysroot ${iphoneos_sdk_dir}"
+      fi
+      ;;
+    x86*-darwin*)
+      osx_sdk_dir="$(show_darwin_sdk_path macosx)"
+      if [ -d "${osx_sdk_dir}" ]; then
+        add_cflags  "-isysroot ${osx_sdk_dir}"
+        add_ldflags "-isysroot ${osx_sdk_dir}"
+      fi
+      ;;
+  esac
+
+  case ${toolchain} in
+    *-darwin8-*)
+      add_cflags  "-mmacosx-version-min=10.4"
+      add_ldflags "-mmacosx-version-min=10.4"
+      ;;
+    *-darwin9-*)
+      add_cflags  "-mmacosx-version-min=10.5"
+      add_ldflags "-mmacosx-version-min=10.5"
+      ;;
+    *-darwin10-*)
+      add_cflags  "-mmacosx-version-min=10.6"
+      add_ldflags "-mmacosx-version-min=10.6"
+      ;;
+    *-darwin11-*)
+      add_cflags  "-mmacosx-version-min=10.7"
+      add_ldflags "-mmacosx-version-min=10.7"
+      ;;
+    *-darwin12-*)
+      add_cflags  "-mmacosx-version-min=10.8"
+      add_ldflags "-mmacosx-version-min=10.8"
+      ;;
+    *-darwin13-*)
+      add_cflags  "-mmacosx-version-min=10.9"
+      add_ldflags "-mmacosx-version-min=10.9"
+      ;;
+    *-darwin14-*)
+      add_cflags  "-mmacosx-version-min=10.10"
+      add_ldflags "-mmacosx-version-min=10.10"
+      ;;
+    *-darwin15-*)
+      add_cflags  "-mmacosx-version-min=10.11"
+      add_ldflags "-mmacosx-version-min=10.11"
+      ;;
+    *-iphonesimulator-*)
+      add_cflags  "-miphoneos-version-min=${IOS_VERSION_MIN}"
+      add_ldflags "-miphoneos-version-min=${IOS_VERSION_MIN}"
+      iossim_sdk_dir="$(show_darwin_sdk_path iphonesimulator)"
+      if [ -d "${iossim_sdk_dir}" ]; then
+        add_cflags  "-isysroot ${iossim_sdk_dir}"
+        add_ldflags "-isysroot ${iossim_sdk_dir}"
+      fi
+      ;;
+  esac
+
+  # Handle Solaris variants. Solaris 10 needs -lposix4
+  case ${toolchain} in
+    sparc-solaris-*)
+      add_extralibs -lposix4
+      ;;
+    *-solaris-*)
+      add_extralibs -lposix4
+      ;;
+  esac
+
+  # Process ARM architecture variants
+  case ${toolchain} in
+    arm*)
+      # on arm, isa versions are supersets
+      case ${tgt_isa} in
+        arm64|armv8)
+          soft_enable neon
+          ;;
+        armv7|armv7s)
+          soft_enable neon
+          # Only enable neon_asm when neon is also enabled.
+          enabled neon && soft_enable neon_asm
+          # If someone tries to force it through, die.
+          if disabled neon && enabled neon_asm; then
+            die "Disabling neon while keeping neon-asm is not supported"
+          fi
+          case ${toolchain} in
+            # Apple iOS SDKs no longer support armv6 as of the version 9
+            # release (coincides with release of Xcode 7). Only enable media
+            # when using earlier SDK releases.
+            *-darwin*)
+              if [ "$(show_darwin_sdk_major_version iphoneos)" -lt 9 ]; then
+                soft_enable media
+              else
+                soft_disable media
+                RTCD_OPTIONS="${RTCD_OPTIONS}--disable-media "
+              fi
+              ;;
+            *)
+              soft_enable media
+              ;;
+          esac
+          ;;
+        armv6)
+          case ${toolchain} in
+            *-darwin*)
+              if [ "$(show_darwin_sdk_major_version iphoneos)" -lt 9 ]; then
+                soft_enable media
+              else
+                die "Your iOS SDK does not support armv6."
+              fi
+              ;;
+            *)
+              soft_enable media
+              ;;
+          esac
+          ;;
+      esac
+
+      asm_conversion_cmd="cat"
+
+      case ${tgt_cc} in
+        gcc)
+          link_with_cc=gcc
+          setup_gnu_toolchain
+          arch_int=${tgt_isa##armv}
+          arch_int=${arch_int%%te}
+          check_add_asflags --defsym ARCHITECTURE=${arch_int}
+          tune_cflags="-mtune="
+          if [ ${tgt_isa} = "armv7" ] || [ ${tgt_isa} = "armv7s" ]; then
+            if [ -z "${float_abi}" ]; then
+              check_cpp <<EOF && float_abi=hard || float_abi=softfp
+#ifndef __ARM_PCS_VFP
+#error "not hardfp"
+#endif
+EOF
+            fi
+            check_add_cflags  -march=armv7-a -mfloat-abi=${float_abi}
+            check_add_asflags -march=armv7-a -mfloat-abi=${float_abi}
+
+            if enabled neon || enabled neon_asm; then
+              check_add_cflags -mfpu=neon #-ftree-vectorize
+              check_add_asflags -mfpu=neon
+            fi
+          elif [ ${tgt_isa} = "arm64" ] || [ ${tgt_isa} = "armv8" ]; then
+            check_add_cflags -march=armv8-a
+            check_add_asflags -march=armv8-a
+          else
+            check_add_cflags -march=${tgt_isa}
+            check_add_asflags -march=${tgt_isa}
+          fi
+
+          enabled debug && add_asflags -g
+          asm_conversion_cmd="${source_path}/build/make/ads2gas.pl"
+          if enabled thumb; then
+            asm_conversion_cmd="$asm_conversion_cmd -thumb"
+            check_add_cflags -mthumb
+            check_add_asflags -mthumb -mimplicit-it=always
+          fi
+          ;;
+        vs*)
+          asm_conversion_cmd="${source_path}/build/make/ads2armasm_ms.pl"
+          AS_SFX=.s
+          msvs_arch_dir=arm-msvs
+          disable_feature multithread
+          disable_feature unit_tests
+          vs_version=${tgt_cc##vs}
+          if [ $vs_version -ge 12 ]; then
+            # MSVC 2013 doesn't allow doing plain .exe projects for ARM,
+            # only "AppContainerApplication" which requires an AppxManifest.
+            # Therefore disable the examples, just build the library.
+            disable_feature examples
+          fi
+          ;;
+        rvct)
+          CC=armcc
+          AR=armar
+          AS=armasm
+          LD="${source_path}/build/make/armlink_adapter.sh"
+          STRIP=arm-none-linux-gnueabi-strip
+          NM=arm-none-linux-gnueabi-nm
+          tune_cflags="--cpu="
+          tune_asflags="--cpu="
+          if [ -z "${tune_cpu}" ]; then
+            if [ ${tgt_isa} = "armv7" ]; then
+              if enabled neon || enabled neon_asm
+              then
+                check_add_cflags --fpu=softvfp+vfpv3
+                check_add_asflags --fpu=softvfp+vfpv3
+              fi
+              check_add_cflags --cpu=Cortex-A8
+              check_add_asflags --cpu=Cortex-A8
+            else
+              check_add_cflags --cpu=${tgt_isa##armv}
+              check_add_asflags --cpu=${tgt_isa##armv}
+            fi
+          fi
+          arch_int=${tgt_isa##armv}
+          arch_int=${arch_int%%te}
+          check_add_asflags --pd "\"ARCHITECTURE SETA ${arch_int}\""
+          enabled debug && add_asflags -g
+          add_cflags --gnu
+          add_cflags --enum_is_int
+          add_cflags --wchar32
+          ;;
+      esac
+
+      case ${tgt_os} in
+        none*)
+          disable_feature multithread
+          disable_feature os_support
+          ;;
+
+        android*)
+          if [ -z "${sdk_path}" ]; then
+            die "Must specify --sdk-path for Android builds."
+          fi
+
+          SDK_PATH=${sdk_path}
+          COMPILER_LOCATION=`find "${SDK_PATH}" \
+                             -name "arm-linux-androideabi-gcc*" -print -quit`
+          TOOLCHAIN_PATH=${COMPILER_LOCATION%/*}/arm-linux-androideabi-
+          CC=${TOOLCHAIN_PATH}gcc
+          CXX=${TOOLCHAIN_PATH}g++
+          AR=${TOOLCHAIN_PATH}ar
+          LD=${TOOLCHAIN_PATH}gcc
+          AS=${TOOLCHAIN_PATH}as
+          STRIP=${TOOLCHAIN_PATH}strip
+          NM=${TOOLCHAIN_PATH}nm
+
+          if [ -z "${alt_libc}" ]; then
+            alt_libc=`find "${SDK_PATH}" -name arch-arm -print | \
+              awk '{n = split($0,a,"/"); \
+                split(a[n-1],b,"-"); \
+                print $0 " " b[2]}' | \
+                sort -g -k 2 | \
+                awk '{ print $1 }' | tail -1`
+          fi
+
+          if [ -d "${alt_libc}" ]; then
+            add_cflags "--sysroot=${alt_libc}"
+            add_ldflags "--sysroot=${alt_libc}"
+          fi
+
+          # linker flag that routes around a CPU bug in some
+          # Cortex-A8 implementations (NDK Dev Guide)
+          add_ldflags "-Wl,--fix-cortex-a8"
+
+          enable_feature pic
+          soft_enable realtime_only
+          if [ ${tgt_isa} = "armv7" ]; then
+            soft_enable runtime_cpu_detect
+          fi
+          if enabled runtime_cpu_detect; then
+            add_cflags "-I${SDK_PATH}/sources/android/cpufeatures"
+          fi
+          ;;
+
+        darwin*)
+          XCRUN_FIND="xcrun --sdk iphoneos --find"
+          CXX="$(${XCRUN_FIND} clang++)"
+          CC="$(${XCRUN_FIND} clang)"
+          AR="$(${XCRUN_FIND} ar)"
+          AS="$(${XCRUN_FIND} as)"
+          STRIP="$(${XCRUN_FIND} strip)"
+          NM="$(${XCRUN_FIND} nm)"
+          RANLIB="$(${XCRUN_FIND} ranlib)"
+          AS_SFX=.s
+          LD="${CXX:-$(${XCRUN_FIND} ld)}"
+
+          # ASFLAGS is written here instead of using check_add_asflags
+          # because we need to overwrite all of ASFLAGS and purge the
+          # options that were put in above
+          ASFLAGS="-arch ${tgt_isa} -g"
+
+          add_cflags -arch ${tgt_isa}
+          add_ldflags -arch ${tgt_isa}
+
+          alt_libc="$(show_darwin_sdk_path iphoneos)"
+          if [ -d "${alt_libc}" ]; then
+            add_cflags -isysroot ${alt_libc}
+          fi
+
+          if [ "${LD}" = "${CXX}" ]; then
+            add_ldflags -miphoneos-version-min="${IOS_VERSION_MIN}"
+          else
+            add_ldflags -ios_version_min "${IOS_VERSION_MIN}"
+          fi
+
+          for d in lib usr/lib usr/lib/system; do
+            try_dir="${alt_libc}/${d}"
+            [ -d "${try_dir}" ] && add_ldflags -L"${try_dir}"
+          done
+
+          case ${tgt_isa} in
+            armv7|armv7s|armv8|arm64)
+              if enabled neon && ! check_xcode_minimum_version; then
+                soft_disable neon
+                log_echo "  neon disabled: upgrade Xcode (need v6.3+)."
+                if enabled neon_asm; then
+                  soft_disable neon_asm
+                  log_echo "  neon_asm disabled: upgrade Xcode (need v6.3+)."
+                fi
+              fi
+              ;;
+          esac
+
+          asm_conversion_cmd="${source_path}/build/make/ads2gas_apple.pl"
+
+          if [ "$(show_darwin_sdk_major_version iphoneos)" -gt 8 ]; then
+            check_add_cflags -fembed-bitcode
+            check_add_asflags -fembed-bitcode
+            check_add_ldflags -fembed-bitcode
+          fi
+          ;;
+
+        linux*)
+          enable_feature linux
+          if enabled rvct; then
+            # Check if we have CodeSourcery GCC in PATH. Needed for
+            # libraries
+            which arm-none-linux-gnueabi-gcc 2>&- || \
+              die "Couldn't find CodeSourcery GCC from PATH"
+
+            # Use armcc as a linker to enable translation of
+            # some gcc specific options such as -lm and -lpthread.
+            LD="armcc --translate_gcc"
+
+            # create configuration file (uses path to CodeSourcery GCC)
+            armcc --arm_linux_configure --arm_linux_config_file=arm_linux.cfg
+
+            add_cflags --arm_linux_paths --arm_linux_config_file=arm_linux.cfg
+            add_asflags --no_hide_all --apcs=/interwork
+            add_ldflags --arm_linux_paths --arm_linux_config_file=arm_linux.cfg
+            enabled pic && add_cflags --apcs=/fpic
+            enabled pic && add_asflags --apcs=/fpic
+            enabled shared && add_cflags --shared
+          fi
+          ;;
+      esac
+      ;;
+    mips*)
+      link_with_cc=gcc
+      setup_gnu_toolchain
+      tune_cflags="-mtune="
+      if enabled dspr2; then
+        check_add_cflags -mips32r2 -mdspr2
+      fi
+
+      if enabled runtime_cpu_detect; then
+        disable_feature runtime_cpu_detect
+      fi
+
+      if [ -n "${tune_cpu}" ]; then
+        case ${tune_cpu} in
+          p5600)
+            check_add_cflags -mips32r5 -funroll-loops -mload-store-pairs
+            check_add_cflags -msched-weight -mhard-float -mfp64
+            check_add_asflags -mips32r5 -mhard-float -mfp64
+            check_add_ldflags -mfp64
+            ;;
+          i6400)
+            check_add_cflags -mips64r6 -mabi=64 -funroll-loops -msched-weight
+            check_add_cflags  -mload-store-pairs -mhard-float -mfp64
+            check_add_asflags -mips64r6 -mabi=64 -mhard-float -mfp64
+            check_add_ldflags -mips64r6 -mabi=64 -mfp64
+            ;;
+        esac
+
+        if enabled msa; then
+          add_cflags -mmsa
+          add_asflags -mmsa
+          add_ldflags -mmsa
+        fi
+      fi
+
+      check_add_cflags -march=${tgt_isa}
+      check_add_asflags -march=${tgt_isa}
+      check_add_asflags -KPIC
+      ;;
+    x86*)
+      case  ${tgt_os} in
+        win*)
+          enabled gcc && add_cflags -fno-common
+          ;;
+        solaris*)
+          CC=${CC:-${CROSS}gcc}
+          CXX=${CXX:-${CROSS}g++}
+          LD=${LD:-${CROSS}gcc}
+          CROSS=${CROSS-g}
+          ;;
+        os2)
+          disable_feature pic
+          AS=${AS:-nasm}
+          add_ldflags -Zhigh-mem
+          ;;
+      esac
+
+      AS="${alt_as:-${AS:-auto}}"
+      case  ${tgt_cc} in
+        icc*)
+          CC=${CC:-icc}
+          LD=${LD:-icc}
+          setup_gnu_toolchain
+          add_cflags -use-msasm  # remove -use-msasm too?
+          # add -no-intel-extensions to suppress warning #10237
+          # refer to http://software.intel.com/en-us/forums/topic/280199
+          add_ldflags -i-static -no-intel-extensions
+          enabled x86_64 && add_cflags -ipo -static -O3 -no-prec-div
+          enabled x86_64 && AR=xiar
+          case ${tune_cpu} in
+            atom*)
+              tune_cflags="-x"
+              tune_cpu="SSE3_ATOM"
+              ;;
+            *)
+              tune_cflags="-march="
+              ;;
+          esac
+          ;;
+        gcc*)
+          link_with_cc=gcc
+          tune_cflags="-march="
+          setup_gnu_toolchain
+          #for 32 bit x86 builds, -O3 did not turn on this flag
+          enabled optimizations && disabled gprof && check_add_cflags -fomit-frame-pointer
+          ;;
+        vs*)
+          # When building with Microsoft Visual Studio the assembler is
+          # invoked directly. Checking at configure time is unnecessary.
+          # Skip the check by setting AS arbitrarily
+          AS=msvs
+          msvs_arch_dir=x86-msvs
+          vc_version=${tgt_cc##vs}
+          case $vc_version in
+            7|8|9|10)
+              echo "${tgt_cc} does not support avx/avx2, disabling....."
+              RTCD_OPTIONS="${RTCD_OPTIONS}--disable-avx --disable-avx2 "
+              soft_disable avx
+              soft_disable avx2
+              ;;
+          esac
+          case $vc_version in
+            7|8|9)
+              echo "${tgt_cc} omits stdint.h, disabling webm-io..."
+              soft_disable webm_io
+              ;;
+          esac
+          ;;
+      esac
+
+      bits=32
+      enabled x86_64 && bits=64
+      check_cpp <<EOF && bits=x32
+#if !defined(__ILP32__) || !defined(__x86_64__)
+#error "not x32"
+#endif
+EOF
+      case ${tgt_cc} in
+        gcc*)
+          add_cflags -m${bits}
+          add_ldflags -m${bits}
+          ;;
+      esac
+
+      soft_enable runtime_cpu_detect
+      # We can't use 'check_cflags' until the compiler is configured and CC is
+      # populated.
+      for ext in ${ARCH_EXT_LIST_X86}; do
+        # disable higher order extensions to simplify asm dependencies
+        if [ "$disable_exts" = "yes" ]; then
+          if ! disabled $ext; then
+            RTCD_OPTIONS="${RTCD_OPTIONS}--disable-${ext} "
+            disable_feature $ext
+          fi
+        elif disabled $ext; then
+          disable_exts="yes"
+        else
+          # use the shortened version for the flag: sse4_1 -> sse4
+          check_gcc_machine_option ${ext%_*} $ext
+        fi
+      done
+
+      if enabled external_build; then
+        log_echo "  skipping assembler detection"
+      else
+        case "${AS}" in
+          auto|"")
+            which nasm >/dev/null 2>&1 && AS=nasm
+            which yasm >/dev/null 2>&1 && AS=yasm
+            if [ "${AS}" = nasm ] ; then
+              # Apple ships version 0.98 of nasm through at least Xcode 6. Revisit
+              # this check if they start shipping a compatible version.
+              apple=`nasm -v | grep "Apple"`
+              [ -n "${apple}" ] \
+                && echo "Unsupported version of nasm: ${apple}" \
+                && AS=""
+            fi
+            [ "${AS}" = auto ] || [ -z "${AS}" ] \
+              && die "Neither yasm nor nasm have been found." \
+                     "See the prerequisites section in the README for more info."
+            ;;
+        esac
+        log_echo "  using $AS"
+      fi
+      [ "${AS##*/}" = nasm ] && add_asflags -Ox
+      AS_SFX=.asm
+      case  ${tgt_os} in
+        win32)
+          add_asflags -f win32
+          enabled debug && add_asflags -g cv8
+          EXE_SFX=.exe
+          ;;
+        win64)
+          add_asflags -f x64
+          enabled debug && add_asflags -g cv8
+          EXE_SFX=.exe
+          ;;
+        linux*|solaris*|android*)
+          add_asflags -f elf${bits}
+          enabled debug && [ "${AS}" = yasm ] && add_asflags -g dwarf2
+          enabled debug && [ "${AS}" = nasm ] && add_asflags -g
+          [ "${AS##*/}" = nasm ] && check_asm_align
+          ;;
+        darwin*)
+          add_asflags -f macho${bits}
+          enabled x86 && darwin_arch="-arch i386" || darwin_arch="-arch x86_64"
+          add_cflags  ${darwin_arch}
+          add_ldflags ${darwin_arch}
+          # -mdynamic-no-pic is still a bit of voodoo -- it was required at
+          # one time, but does not seem to be now, and it breaks some of the
+          # code that still relies on inline assembly.
+          # enabled icc && ! enabled pic && add_cflags -fno-pic -mdynamic-no-pic
+          enabled icc && ! enabled pic && add_cflags -fno-pic
+          ;;
+        iphonesimulator)
+          add_asflags -f macho${bits}
+          enabled x86 && sim_arch="-arch i386" || sim_arch="-arch x86_64"
+          add_cflags  ${sim_arch}
+          add_ldflags ${sim_arch}
+
+          if [ "$(show_darwin_sdk_major_version iphonesimulator)" -gt 8 ]; then
+            # yasm v1.3.0 doesn't know what -fembed-bitcode means, so turning it
+            # on is pointless (unless building a C-only lib). Warn the user, but
+            # do nothing here.
+            log "Warning: Bitcode embed disabled for simulator targets."
+          fi
+          ;;
+        os2)
+          add_asflags -f aout
+          enabled debug && add_asflags -g
+          EXE_SFX=.exe
+          ;;
+        *)
+          log "Warning: Unknown os $tgt_os while setting up $AS flags"
+          ;;
+      esac
+      ;;
+    *-gcc|generic-gnu)
+      link_with_cc=gcc
+      enable_feature gcc
+      setup_gnu_toolchain
+      ;;
+  esac
+
+  # Try to enable CPU specific tuning
+  if [ -n "${tune_cpu}" ]; then
+    if [ -n "${tune_cflags}" ]; then
+      check_add_cflags ${tune_cflags}${tune_cpu} || \
+        die "Requested CPU '${tune_cpu}' not supported by compiler"
+    fi
+    if [ -n "${tune_asflags}" ]; then
+      check_add_asflags ${tune_asflags}${tune_cpu} || \
+        die "Requested CPU '${tune_cpu}' not supported by assembler"
+    fi
+    if [ -z "${tune_cflags}${tune_asflags}" ]; then
+      log_echo "Warning: CPU tuning not supported by this toolchain"
+    fi
+  fi
+
+  if enabled debug; then
+    check_add_cflags -g && check_add_ldflags -g
+  else
+    check_add_cflags -DNDEBUG
+  fi
+
+  enabled gprof && check_add_cflags -pg && check_add_ldflags -pg
+  enabled gcov &&
+    check_add_cflags -fprofile-arcs -ftest-coverage &&
+    check_add_ldflags -fprofile-arcs -ftest-coverage
+
+  if enabled optimizations; then
+    if enabled rvct; then
+      enabled small && check_add_cflags -Ospace || check_add_cflags -Otime
+    else
+      enabled small && check_add_cflags -O2 ||  check_add_cflags -O3
+    fi
+  fi
+
+  if [ "${tgt_isa}" = "x86_64" ] || [ "${tgt_isa}" = "x86" ]; then
+    soft_enable use_x86inc
+  fi
+
+  # Position Independent Code (PIC) support, for building relocatable
+  # shared objects
+  enabled gcc && enabled pic && check_add_cflags -fPIC
+
+  # Work around longjmp interception on glibc >= 2.11, to improve binary
+  # compatibility. See http://code.google.com/p/webm/issues/detail?id=166
+  enabled linux && check_add_cflags -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=0
+
+  # Check for strip utility variant
+  ${STRIP} -V 2>/dev/null | grep GNU >/dev/null && enable_feature gnu_strip
+
+  # Try to determine target endianness
+  check_cc <<EOF
+unsigned int e = 'O'<<24 | '2'<<16 | 'B'<<8 | 'E';
+EOF
+    [ -f "${TMP_O}" ] && od -A n -t x1 "${TMP_O}" | tr -d '\n' |
+        grep '4f *32 *42 *45' >/dev/null 2>&1 && enable_feature big_endian
+
+    # Try to find which inline keywords are supported
+    check_cc <<EOF && INLINE="inline"
+static inline function() {}
+EOF
+
+  # Almost every platform uses pthreads.
+  if enabled multithread; then
+    case ${toolchain} in
+      *-win*-vs*)
+        ;;
+      *-android-gcc)
+        ;;
+      *)
+        check_header pthread.h && add_extralibs -lpthread
+        ;;
+    esac
+  fi
+
+  # only for MIPS platforms
+  case ${toolchain} in
+    mips*)
+      if enabled big_endian; then
+        if enabled dspr2; then
+          echo "dspr2 optimizations are available only for little endian platforms"
+          disable_feature dspr2
+        fi
+        if enabled msa; then
+          echo "msa optimizations are available only for little endian platforms"
+          disable_feature msa
+        fi
+      fi
+      ;;
+  esac
+
+  # glibc needs these
+  if enabled linux; then
+    add_cflags -D_LARGEFILE_SOURCE
+    add_cflags -D_FILE_OFFSET_BITS=64
+  fi
+}
+
+process_toolchain() {
+  process_common_toolchain
+}
+
+print_config_mk() {
+  saved_prefix="${prefix}"
+  prefix=$1
+  makefile=$2
+  shift 2
+  for cfg; do
+    if enabled $cfg; then
+      upname="`toupper $cfg`"
+      echo "${prefix}_${upname}=yes" >> $makefile
+    fi
+  done
+  prefix="${saved_prefix}"
+}
+
+print_config_h() {
+  saved_prefix="${prefix}"
+  prefix=$1
+  header=$2
+  shift 2
+  for cfg; do
+    upname="`toupper $cfg`"
+    if enabled $cfg; then
+      echo "#define ${prefix}_${upname} 1" >> $header
+    else
+      echo "#define ${prefix}_${upname} 0" >> $header
+    fi
+  done
+  prefix="${saved_prefix}"
+}
+
+print_config_vars_h() {
+  header=$1
+  shift
+  while [ $# -gt 0 ]; do
+    upname="`toupper $1`"
+    echo "#define ${upname} $2" >> $header
+    shift 2
+  done
+}
+
+print_webm_license() {
+  saved_prefix="${prefix}"
+  destination=$1
+  prefix="$2"
+  suffix="$3"
+  shift 3
+  cat <<EOF > ${destination}
+${prefix} Copyright (c) 2011 The WebM project authors. All Rights Reserved.${suffix}
+${prefix} ${suffix}
+${prefix} Use of this source code is governed by a BSD-style license${suffix}
+${prefix} that can be found in the LICENSE file in the root of the source${suffix}
+${prefix} tree. An additional intellectual property rights grant can be found${suffix}
+${prefix} in the file PATENTS.  All contributing project authors may${suffix}
+${prefix} be found in the AUTHORS file in the root of the source tree.${suffix}
+EOF
+  prefix="${saved_prefix}"
+}
+
+process_targets() {
+  true;
+}
+
+process_detect() {
+  true;
+}
+
+enable_feature logging
+logfile="config.log"
+self=$0
+process() {
+  cmdline_args="$@"
+  process_cmdline "$@"
+  if enabled child; then
+    echo "# ${self} $@" >> ${logfile}
+  else
+    echo "# ${self} $@" > ${logfile}
+  fi
+  post_process_common_cmdline
+  post_process_cmdline
+  process_toolchain
+  process_detect
+  process_targets
+
+  OOT_INSTALLS="${OOT_INSTALLS}"
+  if enabled source_path_used; then
+  # Prepare the PWD for building.
+  for f in ${OOT_INSTALLS}; do
+    install -D "${source_path}/$f" "$f"
+  done
+  fi
+  cp "${source_path}/build/make/Makefile" .
+
+  clean_temp_files
+  true
+}
diff --git a/libvpx/libvpx/build/make/gen_asm_deps.sh b/libvpx/libvpx/build/make/gen_asm_deps.sh
new file mode 100755
index 0000000..6a7bff9
--- /dev/null
+++ b/libvpx/libvpx/build/make/gen_asm_deps.sh
@@ -0,0 +1,64 @@
+#!/bin/sh
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+self=$0
+show_help() {
+    echo "usage: $self [options] <srcfile>"
+    echo
+    echo "Generate Makefile dependency information from assembly code source"
+    echo
+    exit 1
+}
+die_unknown(){
+    echo "Unknown option \"$1\"."
+    echo "See $0 --help for available options."
+    exit 1
+}
+for opt do
+    optval="${opt#*=}"
+    case "$opt" in
+    --build-pfx=*) pfx="${optval}"
+    ;;
+    --depfile=*) out="${optval}"
+    ;;
+    -I*) raw_inc_paths="${raw_inc_paths} ${opt}"
+         inc_path="${inc_path} ${opt#-I}"
+    ;;
+    -h|--help) show_help
+    ;;
+    *) [ -f "$opt" ] && srcfile="$opt"
+    ;;
+    esac
+done
+
+[ -n "$srcfile" ] || show_help
+sfx=${sfx:-asm}
+includes=$(LC_ALL=C egrep -i "include +\"?[a-z0-9_/]+\.${sfx}" $srcfile |
+           perl -p -e "s;.*?([a-z0-9_/]+.${sfx}).*;\1;")
+#" restore editor state
+for inc in ${includes}; do
+    found_inc_path=
+    for idir in ${inc_path}; do
+        [ -f "${idir}/${inc}" ] && found_inc_path="${idir}" && break
+    done
+    if [ -f `dirname $srcfile`/$inc ]; then
+        # Handle include files in the same directory as the source
+        $self --build-pfx=$pfx --depfile=$out ${raw_inc_paths} `dirname $srcfile`/$inc
+    elif [ -n "${found_inc_path}" ]; then
+        # Handle include files on the include path
+        $self --build-pfx=$pfx --depfile=$out ${raw_inc_paths} "${found_inc_path}/$inc"
+    else
+        # Handle generated includes in the build root (which may not exist yet)
+        echo ${out} ${out%d}o: "${pfx}${inc}"
+    fi
+done
+echo ${out} ${out%d}o: $srcfile
diff --git a/libvpx/libvpx/build/make/gen_msvs_def.sh b/libvpx/libvpx/build/make/gen_msvs_def.sh
new file mode 100755
index 0000000..4defcc2
--- /dev/null
+++ b/libvpx/libvpx/build/make/gen_msvs_def.sh
@@ -0,0 +1,83 @@
+#!/bin/bash
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+self=$0
+self_basename=${self##*/}
+EOL=$'\n'
+
+show_help() {
+    cat <<EOF
+Usage: ${self_basename} [options] file1 [file2 ...]
+
+This script generates a MSVC module definition file containing a list of symbols
+to export from a DLL. Source files are technically bash scripts (and thus may
+use #comment syntax) but in general, take the form of a list of symbols:
+
+  <kind> symbol1 [symbol2, symbol3, ...]
+
+where <kind> is either 'text' or 'data'
+
+
+Options:
+    --help                      Print this message
+    --out=filename              Write output to a file [stdout]
+    --name=project_name         Name of the library (required)
+EOF
+    exit 1
+}
+
+die() {
+    echo "${self_basename}: $@"
+    exit 1
+}
+
+die_unknown(){
+    echo "Unknown option \"$1\"."
+    echo "See ${self_basename} --help for available options."
+    exit 1
+}
+
+text() {
+    for sym in "$@"; do
+        echo "  $sym" >> ${outfile}
+    done
+}
+
+data() {
+    for sym in "$@"; do
+        printf "  %-40s DATA\n" "$sym" >> ${outfile}
+    done
+}
+
+# Process command line
+for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+    --help|-h) show_help
+    ;;
+    --out=*) outfile="$optval"
+    ;;
+    --name=*) name="${optval}"
+    ;;
+     -*) die_unknown $opt
+    ;;
+    *) file_list[${#file_list[@]}]="$opt"
+    esac
+done
+outfile=${outfile:-/dev/stdout}
+[ -n "$name" ] || die "Library name (--name) must be specified!"
+
+echo "LIBRARY ${name}" > ${outfile}
+echo "EXPORTS" >> ${outfile}
+for f in "${file_list[@]}"; do
+    . $f
+done
diff --git a/libvpx/libvpx/build/make/gen_msvs_proj.sh b/libvpx/libvpx/build/make/gen_msvs_proj.sh
new file mode 100755
index 0000000..2b91fbf
--- /dev/null
+++ b/libvpx/libvpx/build/make/gen_msvs_proj.sh
@@ -0,0 +1,490 @@
+#!/bin/bash
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+self=$0
+self_basename=${self##*/}
+self_dirname=$(dirname "$0")
+
+. "$self_dirname/msvs_common.sh"|| exit 127
+
+show_help() {
+    cat <<EOF
+Usage: ${self_basename} --name=projname [options] file1 [file2 ...]
+
+This script generates a Visual Studio project file from a list of source
+code files.
+
+Options:
+    --help                      Print this message
+    --exe                       Generate a project for building an Application
+    --lib                       Generate a project for creating a static library
+    --dll                       Generate a project for creating a dll
+    --static-crt                Use the static C runtime (/MT)
+    --target=isa-os-cc          Target specifier (required)
+    --out=filename              Write output to a file [stdout]
+    --name=project_name         Name of the project (required)
+    --proj-guid=GUID            GUID to use for the project
+    --module-def=filename       File containing export definitions (for DLLs)
+    --ver=version               Version (7,8,9) of visual studio to generate for
+    --src-path-bare=dir         Path to root of source tree
+    -Ipath/to/include           Additional include directories
+    -DFLAG[=value]              Preprocessor macros to define
+    -Lpath/to/lib               Additional library search paths
+    -llibname                   Library to link against
+EOF
+    exit 1
+}
+
+generate_filter() {
+    local var=$1
+    local name=$2
+    local pats=$3
+    local file_list_sz
+    local i
+    local f
+    local saveIFS="$IFS"
+    local pack
+    echo "generating filter '$name' from ${#file_list[@]} files" >&2
+    IFS=*
+
+    open_tag Filter \
+        Name=$name \
+        Filter=$pats \
+        UniqueIdentifier=`generate_uuid` \
+
+    file_list_sz=${#file_list[@]}
+    for i in ${!file_list[@]}; do
+        f=${file_list[i]}
+        for pat in ${pats//;/$IFS}; do
+            if [ "${f##*.}" == "$pat" ]; then
+                unset file_list[i]
+
+                objf=$(echo ${f%.*}.obj \
+                       | sed -e "s,$src_path_bare,," \
+                             -e 's/^[\./]\+//g' -e 's,[:/ ],_,g')
+                open_tag File RelativePath="$f"
+
+                if [ "$pat" == "asm" ] && $asm_use_custom_step; then
+                    # Avoid object file name collisions, i.e. vpx_config.c and
+                    # vpx_config.asm produce the same object file without
+                    # this additional suffix.
+                    objf=${objf%.obj}_asm.obj
+                    for plat in "${platforms[@]}"; do
+                        for cfg in Debug Release; do
+                            open_tag FileConfiguration \
+                                Name="${cfg}|${plat}" \
+
+                            tag Tool \
+                                Name="VCCustomBuildTool" \
+                                Description="Assembling \$(InputFileName)" \
+                                CommandLine="$(eval echo \$asm_${cfg}_cmdline) -o \$(IntDir)\\$objf" \
+                                Outputs="\$(IntDir)\\$objf" \
+
+                            close_tag FileConfiguration
+                        done
+                    done
+                fi
+                if [ "$pat" == "c" ] || \
+                   [ "$pat" == "cc" ] || [ "$pat" == "cpp" ]; then
+                    for plat in "${platforms[@]}"; do
+                        for cfg in Debug Release; do
+                            open_tag FileConfiguration \
+                                Name="${cfg}|${plat}" \
+
+                            tag Tool \
+                                Name="VCCLCompilerTool" \
+                                ObjectFile="\$(IntDir)\\$objf" \
+
+                            close_tag FileConfiguration
+                        done
+                    done
+                fi
+                close_tag File
+
+                break
+            fi
+        done
+    done
+
+    close_tag Filter
+    IFS="$saveIFS"
+}
+
+# Process command line
+unset target
+for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+        --help|-h) show_help
+        ;;
+        --target=*) target="${optval}"
+        ;;
+        --out=*) outfile="$optval"
+        ;;
+        --name=*) name="${optval}"
+        ;;
+        --proj-guid=*) guid="${optval}"
+        ;;
+        --module-def=*) link_opts="${link_opts} ModuleDefinitionFile=${optval}"
+        ;;
+        --exe) proj_kind="exe"
+        ;;
+        --dll) proj_kind="dll"
+        ;;
+        --lib) proj_kind="lib"
+        ;;
+        --src-path-bare=*)
+            src_path_bare=$(fix_path "$optval")
+            src_path_bare=${src_path_bare%/}
+        ;;
+        --static-crt) use_static_runtime=true
+        ;;
+        --ver=*)
+            vs_ver="$optval"
+            case "$optval" in
+                [789])
+                ;;
+                *) die Unrecognized Visual Studio Version in $opt
+                ;;
+            esac
+        ;;
+        -I*)
+            opt=${opt##-I}
+            opt=$(fix_path "$opt")
+            opt="${opt%/}"
+            incs="${incs}${incs:+;}&quot;${opt}&quot;"
+            yasmincs="${yasmincs} -I&quot;${opt}&quot;"
+        ;;
+        -D*) defines="${defines}${defines:+;}${opt##-D}"
+        ;;
+        -L*) # fudge . to $(OutDir)
+            if [ "${opt##-L}" == "." ]; then
+                libdirs="${libdirs}${libdirs:+;}&quot;\$(OutDir)&quot;"
+            else
+                 # Also try directories for this platform/configuration
+                 opt=${opt##-L}
+                 opt=$(fix_path "$opt")
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}&quot;"
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}/\$(PlatformName)/\$(ConfigurationName)&quot;"
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}/\$(PlatformName)&quot;"
+            fi
+        ;;
+        -l*) libs="${libs}${libs:+ }${opt##-l}.lib"
+        ;;
+        -*) die_unknown $opt
+        ;;
+        *)
+            # The paths in file_list are fixed outside of the loop.
+            file_list[${#file_list[@]}]="$opt"
+            case "$opt" in
+                 *.asm) uses_asm=true
+                 ;;
+            esac
+        ;;
+    esac
+done
+
+# Make one call to fix_path for file_list to improve performance.
+fix_file_list file_list
+
+outfile=${outfile:-/dev/stdout}
+guid=${guid:-`generate_uuid`}
+asm_use_custom_step=false
+uses_asm=${uses_asm:-false}
+case "${vs_ver:-8}" in
+    7) vs_ver_id="7.10"
+       asm_use_custom_step=$uses_asm
+       warn_64bit='Detect64BitPortabilityProblems=true'
+    ;;
+    8) vs_ver_id="8.00"
+       asm_use_custom_step=$uses_asm
+       warn_64bit='Detect64BitPortabilityProblems=true'
+    ;;
+    9) vs_ver_id="9.00"
+       asm_use_custom_step=$uses_asm
+       warn_64bit='Detect64BitPortabilityProblems=false'
+    ;;
+esac
+
+[ -n "$name" ] || die "Project name (--name) must be specified!"
+[ -n "$target" ] || die "Target (--target) must be specified!"
+
+if ${use_static_runtime:-false}; then
+    release_runtime=0
+    debug_runtime=1
+    lib_sfx=mt
+else
+    release_runtime=2
+    debug_runtime=3
+    lib_sfx=md
+fi
+
+# Calculate debug lib names: If a lib ends in ${lib_sfx}.lib, then rename
+# it to ${lib_sfx}d.lib. This precludes linking to release libs from a
+# debug exe, so this may need to be refactored later.
+for lib in ${libs}; do
+    if [ "$lib" != "${lib%${lib_sfx}.lib}" ]; then
+        lib=${lib%.lib}d.lib
+    fi
+    debug_libs="${debug_libs}${debug_libs:+ }${lib}"
+done
+
+
+# List Keyword for this target
+case "$target" in
+    x86*) keyword="ManagedCProj"
+    ;;
+    *) die "Unsupported target $target!"
+esac
+
+# List of all platforms supported for this target
+case "$target" in
+    x86_64*)
+        platforms[0]="x64"
+        asm_Debug_cmdline="yasm -Xvc -g cv8 -f win64 ${yasmincs} &quot;\$(InputPath)&quot;"
+        asm_Release_cmdline="yasm -Xvc -f win64 ${yasmincs} &quot;\$(InputPath)&quot;"
+    ;;
+    x86*)
+        platforms[0]="Win32"
+        asm_Debug_cmdline="yasm -Xvc -g cv8 -f win32 ${yasmincs} &quot;\$(InputPath)&quot;"
+        asm_Release_cmdline="yasm -Xvc -f win32 ${yasmincs} &quot;\$(InputPath)&quot;"
+    ;;
+    *) die "Unsupported target $target!"
+    ;;
+esac
+
+generate_vcproj() {
+    case "$proj_kind" in
+        exe) vs_ConfigurationType=1
+        ;;
+        dll) vs_ConfigurationType=2
+        ;;
+        *)   vs_ConfigurationType=4
+        ;;
+    esac
+
+    echo "<?xml version=\"1.0\" encoding=\"Windows-1252\"?>"
+    open_tag VisualStudioProject \
+        ProjectType="Visual C++" \
+        Version="${vs_ver_id}" \
+        Name="${name}" \
+        ProjectGUID="{${guid}}" \
+        RootNamespace="${name}" \
+        Keyword="${keyword}" \
+
+    open_tag Platforms
+    for plat in "${platforms[@]}"; do
+        tag Platform Name="$plat"
+    done
+    close_tag Platforms
+
+    open_tag Configurations
+    for plat in "${platforms[@]}"; do
+        plat_no_ws=`echo $plat | sed 's/[^A-Za-z0-9_]/_/g'`
+        open_tag Configuration \
+            Name="Debug|$plat" \
+            OutputDirectory="\$(SolutionDir)$plat_no_ws/\$(ConfigurationName)" \
+            IntermediateDirectory="$plat_no_ws/\$(ConfigurationName)/${name}" \
+            ConfigurationType="$vs_ConfigurationType" \
+            CharacterSet="1" \
+
+        case "$target" in
+            x86*)
+                case "$name" in
+                    vpx)
+                        tag Tool \
+                            Name="VCCLCompilerTool" \
+                            Optimization="0" \
+                            AdditionalIncludeDirectories="$incs" \
+                            PreprocessorDefinitions="WIN32;_DEBUG;_CRT_SECURE_NO_WARNINGS;_CRT_SECURE_NO_DEPRECATE;$defines" \
+                            RuntimeLibrary="$debug_runtime" \
+                            UsePrecompiledHeader="0" \
+                            WarningLevel="3" \
+                            DebugInformationFormat="2" \
+                            $warn_64bit \
+
+                        $uses_asm && tag Tool Name="YASM"  IncludePaths="$incs" Debug="true"
+                    ;;
+                    *)
+                        tag Tool \
+                            Name="VCCLCompilerTool" \
+                            Optimization="0" \
+                            AdditionalIncludeDirectories="$incs" \
+                            PreprocessorDefinitions="WIN32;_DEBUG;_CRT_SECURE_NO_WARNINGS;_CRT_SECURE_NO_DEPRECATE;$defines" \
+                            RuntimeLibrary="$debug_runtime" \
+                            UsePrecompiledHeader="0" \
+                            WarningLevel="3" \
+                            DebugInformationFormat="2" \
+                            $warn_64bit \
+
+                        $uses_asm && tag Tool Name="YASM"  IncludePaths="$incs" Debug="true"
+                    ;;
+                esac
+            ;;
+        esac
+
+        case "$proj_kind" in
+            exe)
+                case "$target" in
+                    x86*)
+                        case "$name" in
+                            *)
+                                tag Tool \
+                                    Name="VCLinkerTool" \
+                                    AdditionalDependencies="$debug_libs \$(NoInherit)" \
+                                    AdditionalLibraryDirectories="$libdirs" \
+                                    GenerateDebugInformation="true" \
+                                    ProgramDatabaseFile="\$(OutDir)/${name}.pdb" \
+                            ;;
+                        esac
+                    ;;
+                 esac
+            ;;
+            lib)
+                case "$target" in
+                    x86*)
+                        tag Tool \
+                            Name="VCLibrarianTool" \
+                            OutputFile="\$(OutDir)/${name}${lib_sfx}d.lib" \
+
+                    ;;
+                esac
+            ;;
+            dll)
+                tag Tool \
+                    Name="VCLinkerTool" \
+                    AdditionalDependencies="\$(NoInherit)" \
+                    LinkIncremental="2" \
+                    GenerateDebugInformation="true" \
+                    AssemblyDebug="1" \
+                    TargetMachine="1" \
+                    $link_opts \
+
+            ;;
+        esac
+
+        close_tag Configuration
+
+        open_tag Configuration \
+            Name="Release|$plat" \
+            OutputDirectory="\$(SolutionDir)$plat_no_ws/\$(ConfigurationName)" \
+            IntermediateDirectory="$plat_no_ws/\$(ConfigurationName)/${name}" \
+            ConfigurationType="$vs_ConfigurationType" \
+            CharacterSet="1" \
+            WholeProgramOptimization="0" \
+
+        case "$target" in
+            x86*)
+                case "$name" in
+                    vpx)
+                        tag Tool \
+                            Name="VCCLCompilerTool" \
+                            Optimization="2" \
+                            FavorSizeorSpeed="1" \
+                            AdditionalIncludeDirectories="$incs" \
+                            PreprocessorDefinitions="WIN32;NDEBUG;_CRT_SECURE_NO_WARNINGS;_CRT_SECURE_NO_DEPRECATE;$defines" \
+                            RuntimeLibrary="$release_runtime" \
+                            UsePrecompiledHeader="0" \
+                            WarningLevel="3" \
+                            DebugInformationFormat="0" \
+                            $warn_64bit \
+
+                        $uses_asm && tag Tool Name="YASM"  IncludePaths="$incs"
+                    ;;
+                    *)
+                        tag Tool \
+                            Name="VCCLCompilerTool" \
+                            AdditionalIncludeDirectories="$incs" \
+                            Optimization="2" \
+                            FavorSizeorSpeed="1" \
+                            PreprocessorDefinitions="WIN32;NDEBUG;_CRT_SECURE_NO_WARNINGS;_CRT_SECURE_NO_DEPRECATE;$defines" \
+                            RuntimeLibrary="$release_runtime" \
+                            UsePrecompiledHeader="0" \
+                            WarningLevel="3" \
+                            DebugInformationFormat="0" \
+                            $warn_64bit \
+
+                        $uses_asm && tag Tool Name="YASM"  IncludePaths="$incs"
+                    ;;
+                esac
+            ;;
+        esac
+
+        case "$proj_kind" in
+            exe)
+                case "$target" in
+                    x86*)
+                        case "$name" in
+                            *)
+                                tag Tool \
+                                    Name="VCLinkerTool" \
+                                    AdditionalDependencies="$libs \$(NoInherit)" \
+                                    AdditionalLibraryDirectories="$libdirs" \
+
+                            ;;
+                        esac
+                    ;;
+                 esac
+            ;;
+            lib)
+                case "$target" in
+                    x86*)
+                        tag Tool \
+                            Name="VCLibrarianTool" \
+                            OutputFile="\$(OutDir)/${name}${lib_sfx}.lib" \
+
+                    ;;
+                esac
+            ;;
+            dll) # note differences to debug version: LinkIncremental, AssemblyDebug
+                tag Tool \
+                    Name="VCLinkerTool" \
+                    AdditionalDependencies="\$(NoInherit)" \
+                    LinkIncremental="1" \
+                    GenerateDebugInformation="true" \
+                    TargetMachine="1" \
+                    $link_opts \
+
+            ;;
+        esac
+
+        close_tag Configuration
+    done
+    close_tag Configurations
+
+    open_tag Files
+    generate_filter srcs   "Source Files"   "c;cc;cpp;def;odl;idl;hpj;bat;asm;asmx"
+    generate_filter hdrs   "Header Files"   "h;hm;inl;inc;xsd"
+    generate_filter resrcs "Resource Files" "rc;ico;cur;bmp;dlg;rc2;rct;bin;rgs;gif;jpg;jpeg;jpe;resx;tiff;tif;png;wav"
+    generate_filter resrcs "Build Files"    "mk"
+    close_tag Files
+
+    tag       Globals
+    close_tag VisualStudioProject
+
+    # This must be done from within the {} subshell
+    echo "Ignored files list (${#file_list[@]} items) is:" >&2
+    for f in "${file_list[@]}"; do
+        echo "    $f" >&2
+    done
+}
+
+generate_vcproj |
+    sed  -e '/"/s;\([^ "]\)/;\1\\;g' > ${outfile}
+
+exit
+<!--
+TODO: Add any files not captured by filters.
+                <File
+                        RelativePath=".\ReadMe.txt"
+                        >
+                </File>
+-->
diff --git a/libvpx/libvpx/build/make/gen_msvs_sln.sh b/libvpx/libvpx/build/make/gen_msvs_sln.sh
new file mode 100755
index 0000000..664b404
--- /dev/null
+++ b/libvpx/libvpx/build/make/gen_msvs_sln.sh
@@ -0,0 +1,327 @@
+#!/bin/bash
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+self=$0
+self_basename=${self##*/}
+EOL=$'\n'
+EOLDOS=$'\r'
+
+show_help() {
+    cat <<EOF
+Usage: ${self_basename} [options] file1 [file2 ...]
+
+This script generates a Visual Studio solution file from a list of project
+files.
+
+Options:
+    --help                      Print this message
+    --out=outfile               Redirect output to a file
+    --ver=version               Version (7,8,9,10,11,12,14) of visual studio to generate for
+    --target=isa-os-cc          Target specifier
+EOF
+    exit 1
+}
+
+die() {
+    echo "${self_basename}: $@" >&2
+    [ -f "${outfile}" ] && rm -f ${outfile}{,.mk}
+    exit 1
+}
+
+die_unknown(){
+    echo "Unknown option \"$1\"." >&2
+    echo "See ${self_basename} --help for available options." >&2
+    [ -f "${outfile}" ] && rm -f ${outfile}{,.mk}
+    exit 1
+}
+
+indent1=$'\t'
+indent=""
+indent_push() {
+    indent="${indent}${indent1}"
+}
+indent_pop() {
+    indent="${indent%${indent1}}"
+}
+
+parse_project() {
+    local file=$1
+    if [ "$sfx" = "vcproj" ]; then
+        local name=`grep Name "$file" | awk 'BEGIN {FS="\""}{if (NR==1) print $2}'`
+        local guid=`grep ProjectGUID "$file" | awk 'BEGIN {FS="\""}{if (NR==1) print $2}'`
+    else
+        local name=`grep RootNamespace "$file" | sed 's,.*<.*>\(.*\)</.*>.*,\1,'`
+        local guid=`grep ProjectGuid "$file" | sed 's,.*<.*>\(.*\)</.*>.*,\1,'`
+    fi
+
+    # save the project GUID to a varaible, normalizing to the basename of the
+    # vcproj file without the extension
+    local var
+    var=${file##*/}
+    var=${var%%.${sfx}}
+    eval "${var}_file=\"$1\""
+    eval "${var}_name=$name"
+    eval "${var}_guid=$guid"
+
+    if [ "$sfx" = "vcproj" ]; then
+        cur_config_list=`grep -A1 '<Configuration' $file |
+            grep Name | cut -d\" -f2`
+    else
+        cur_config_list=`grep -B1 'Label="Configuration"' $file |
+            grep Condition | cut -d\' -f4`
+    fi
+    new_config_list=$(for i in $config_list $cur_config_list; do
+        echo $i
+    done | sort | uniq)
+    if [ "$config_list" != "" ] && [ "$config_list" != "$new_config_list" ]; then
+        mixed_platforms=1
+    fi
+    config_list="$new_config_list"
+    eval "${var}_config_list=\"$cur_config_list\""
+    proj_list="${proj_list} ${var}"
+}
+
+process_project() {
+    eval "local file=\${$1_file}"
+    eval "local name=\${$1_name}"
+    eval "local guid=\${$1_guid}"
+
+    # save the project GUID to a varaible, normalizing to the basename of the
+    # vcproj file without the extension
+    local var
+    var=${file##*/}
+    var=${var%%.${sfx}}
+    eval "${var}_guid=$guid"
+
+    echo "Project(\"{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}\") = \"$name\", \"$file\", \"$guid\""
+    indent_push
+
+    eval "local deps=\"\${${var}_deps}\""
+    if [ -n "$deps" ] && [ "$sfx" = "vcproj" ]; then
+        echo "${indent}ProjectSection(ProjectDependencies) = postProject"
+        indent_push
+
+        for dep in $deps; do
+            eval "local dep_guid=\${${dep}_guid}"
+            [ -z "${dep_guid}" ] && die "Unknown GUID for $dep (dependency of $var)"
+            echo "${indent}$dep_guid = $dep_guid"
+        done
+
+        indent_pop
+        echo "${indent}EndProjectSection"
+
+    fi
+
+    indent_pop
+    echo "EndProject"
+}
+
+process_global() {
+    echo "Global"
+    indent_push
+
+    #
+    # Solution Configuration Platforms
+    #
+    echo "${indent}GlobalSection(SolutionConfigurationPlatforms) = preSolution"
+    indent_push
+    IFS_bak=${IFS}
+    IFS=$'\r'$'\n'
+    if [ "$mixed_platforms" != "" ]; then
+        config_list="
+Release|Mixed Platforms
+Debug|Mixed Platforms"
+    fi
+    for config in ${config_list}; do
+        echo "${indent}$config = $config"
+    done
+    IFS=${IFS_bak}
+    indent_pop
+    echo "${indent}EndGlobalSection"
+
+    #
+    # Project Configuration Platforms
+    #
+    echo "${indent}GlobalSection(ProjectConfigurationPlatforms) = postSolution"
+    indent_push
+    for proj in ${proj_list}; do
+        eval "local proj_guid=\${${proj}_guid}"
+        eval "local proj_config_list=\${${proj}_config_list}"
+        IFS=$'\r'$'\n'
+        for config in ${proj_config_list}; do
+            if [ "$mixed_platforms" != "" ]; then
+                local c=${config%%|*}
+                echo "${indent}${proj_guid}.${c}|Mixed Platforms.ActiveCfg = ${config}"
+                echo "${indent}${proj_guid}.${c}|Mixed Platforms.Build.0 = ${config}"
+            else
+                echo "${indent}${proj_guid}.${config}.ActiveCfg = ${config}"
+                echo "${indent}${proj_guid}.${config}.Build.0 = ${config}"
+            fi
+
+        done
+        IFS=${IFS_bak}
+    done
+    indent_pop
+    echo "${indent}EndGlobalSection"
+
+    #
+    # Solution Properties
+    #
+    echo "${indent}GlobalSection(SolutionProperties) = preSolution"
+    indent_push
+    echo "${indent}HideSolutionNode = FALSE"
+    indent_pop
+    echo "${indent}EndGlobalSection"
+
+    indent_pop
+    echo "EndGlobal"
+}
+
+process_makefile() {
+    IFS_bak=${IFS}
+    IFS=$'\r'$'\n'
+    local TAB=$'\t'
+    cat <<EOF
+ifeq (\$(CONFIG_VS_VERSION),7)
+MSBUILD_TOOL := devenv.com
+else
+MSBUILD_TOOL := msbuild.exe
+endif
+found_devenv := \$(shell which \$(MSBUILD_TOOL) >/dev/null 2>&1 && echo yes)
+.nodevenv.once:
+${TAB}@echo "  * \$(MSBUILD_TOOL) not found in path."
+${TAB}@echo "  * "
+${TAB}@echo "  * You will have to build all configurations manually using the"
+${TAB}@echo "  * Visual Studio IDE. To allow make to build them automatically,"
+${TAB}@echo "  * add the Common7/IDE directory of your Visual Studio"
+${TAB}@echo "  * installation to your path, eg:"
+${TAB}@echo "  *   C:\Program Files\Microsoft Visual Studio 8\Common7\IDE"
+${TAB}@echo "  * "
+${TAB}@touch \$@
+CLEAN-OBJS += \$(if \$(found_devenv),,.nodevenv.once)
+
+EOF
+
+    for sln_config in ${config_list}; do
+        local config=${sln_config%%|*}
+        local platform=${sln_config##*|}
+        local nows_sln_config=`echo $sln_config | sed -e 's/[^a-zA-Z0-9]/_/g'`
+        cat <<EOF
+BUILD_TARGETS += \$(if \$(NO_LAUNCH_DEVENV),,$nows_sln_config)
+clean::
+${TAB}rm -rf "$platform"/"$config"
+.PHONY: $nows_sln_config
+ifneq (\$(found_devenv),)
+  ifeq (\$(CONFIG_VS_VERSION),7)
+$nows_sln_config: $outfile
+${TAB}\$(MSBUILD_TOOL) $outfile -build "$config"
+
+  else
+$nows_sln_config: $outfile
+${TAB}\$(MSBUILD_TOOL) $outfile -m -t:Build \\
+${TAB}${TAB}-p:Configuration="$config" -p:Platform="$platform"
+
+  endif
+else
+$nows_sln_config: $outfile .nodevenv.once
+${TAB}@echo "  * Skipping build of $sln_config (\$(MSBUILD_TOOL) not in path)."
+${TAB}@echo "  * "
+endif
+
+EOF
+    done
+    IFS=${IFS_bak}
+}
+
+# Process command line
+outfile=/dev/stdout
+for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+    --help|-h) show_help
+    ;;
+    --out=*) outfile="${optval}"; mkoutfile="${optval}".mk
+    ;;
+    --dep=*) eval "${optval%%:*}_deps=\"\${${optval%%:*}_deps} ${optval##*:}\""
+    ;;
+    --ver=*) vs_ver="$optval"
+             case $optval in
+             [789]|10|11|12|14)
+             ;;
+             *) die Unrecognized Visual Studio Version in $opt
+             ;;
+             esac
+    ;;
+    --ver=*) vs_ver="$optval"
+             case $optval in
+             7) sln_vers="8.00"
+                sln_vers_str="Visual Studio .NET 2003"
+             ;;
+             [89])
+             ;;
+             *) die "Unrecognized Visual Studio Version '$optval' in $opt"
+             ;;
+             esac
+    ;;
+    --target=*) target="${optval}"
+    ;;
+    -*) die_unknown $opt
+    ;;
+    *) file_list[${#file_list[@]}]="$opt"
+    esac
+done
+outfile=${outfile:-/dev/stdout}
+mkoutfile=${mkoutfile:-/dev/stdout}
+case "${vs_ver:-8}" in
+    7) sln_vers="8.00"
+       sln_vers_str="Visual Studio .NET 2003"
+    ;;
+    8) sln_vers="9.00"
+       sln_vers_str="Visual Studio 2005"
+    ;;
+    9) sln_vers="10.00"
+       sln_vers_str="Visual Studio 2008"
+    ;;
+    10) sln_vers="11.00"
+       sln_vers_str="Visual Studio 2010"
+    ;;
+    11) sln_vers="12.00"
+       sln_vers_str="Visual Studio 2012"
+    ;;
+    12) sln_vers="12.00"
+       sln_vers_str="Visual Studio 2013"
+    ;;
+    14) sln_vers="14.00"
+       sln_vers_str="Visual Studio 2015"
+    ;;
+esac
+case "${vs_ver:-8}" in
+    [789])
+    sfx=vcproj
+    ;;
+    10|11|12|14)
+    sfx=vcxproj
+    ;;
+esac
+
+for f in "${file_list[@]}"; do
+    parse_project $f
+done
+cat  >${outfile} <<EOF
+Microsoft Visual Studio Solution File, Format Version $sln_vers${EOLDOS}
+# $sln_vers_str${EOLDOS}
+EOF
+for proj in ${proj_list}; do
+    process_project $proj >>${outfile}
+done
+process_global >>${outfile}
+process_makefile >${mkoutfile}
diff --git a/libvpx/libvpx/build/make/gen_msvs_vcxproj.sh b/libvpx/libvpx/build/make/gen_msvs_vcxproj.sh
new file mode 100755
index 0000000..e98611d
--- /dev/null
+++ b/libvpx/libvpx/build/make/gen_msvs_vcxproj.sh
@@ -0,0 +1,490 @@
+#!/bin/bash
+##
+##  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+self=$0
+self_basename=${self##*/}
+self_dirname=$(dirname "$0")
+
+. "$self_dirname/msvs_common.sh"|| exit 127
+
+show_help() {
+    cat <<EOF
+Usage: ${self_basename} --name=projname [options] file1 [file2 ...]
+
+This script generates a Visual Studio project file from a list of source
+code files.
+
+Options:
+    --help                      Print this message
+    --exe                       Generate a project for building an Application
+    --lib                       Generate a project for creating a static library
+    --dll                       Generate a project for creating a dll
+    --static-crt                Use the static C runtime (/MT)
+    --enable-werror             Treat warnings as errors (/WX)
+    --target=isa-os-cc          Target specifier (required)
+    --out=filename              Write output to a file [stdout]
+    --name=project_name         Name of the project (required)
+    --proj-guid=GUID            GUID to use for the project
+    --module-def=filename       File containing export definitions (for DLLs)
+    --ver=version               Version (10,11,12,14) of visual studio to generate for
+    --src-path-bare=dir         Path to root of source tree
+    -Ipath/to/include           Additional include directories
+    -DFLAG[=value]              Preprocessor macros to define
+    -Lpath/to/lib               Additional library search paths
+    -llibname                   Library to link against
+EOF
+    exit 1
+}
+
+tag_content() {
+    local tag=$1
+    local content=$2
+    shift
+    shift
+    if [ $# -ne 0 ]; then
+        echo "${indent}<${tag}"
+        indent_push
+        tag_attributes "$@"
+        echo "${indent}>${content}</${tag}>"
+        indent_pop
+    else
+        echo "${indent}<${tag}>${content}</${tag}>"
+    fi
+}
+
+generate_filter() {
+    local name=$1
+    local pats=$2
+    local file_list_sz
+    local i
+    local f
+    local saveIFS="$IFS"
+    local pack
+    echo "generating filter '$name' from ${#file_list[@]} files" >&2
+    IFS=*
+
+    file_list_sz=${#file_list[@]}
+    for i in ${!file_list[@]}; do
+        f=${file_list[i]}
+        for pat in ${pats//;/$IFS}; do
+            if [ "${f##*.}" == "$pat" ]; then
+                unset file_list[i]
+
+                objf=$(echo ${f%.*}.obj \
+                       | sed -e "s,$src_path_bare,," \
+                             -e 's/^[\./]\+//g' -e 's,[:/ ],_,g')
+
+                if ([ "$pat" == "asm" ] || [ "$pat" == "s" ]) && $asm_use_custom_step; then
+                    # Avoid object file name collisions, i.e. vpx_config.c and
+                    # vpx_config.asm produce the same object file without
+                    # this additional suffix.
+                    objf=${objf%.obj}_asm.obj
+                    open_tag CustomBuild \
+                        Include="$f"
+                    for plat in "${platforms[@]}"; do
+                        for cfg in Debug Release; do
+                            tag_content Message "Assembling %(Filename)%(Extension)" \
+                                Condition="'\$(Configuration)|\$(Platform)'=='$cfg|$plat'"
+                            tag_content Command "$(eval echo \$asm_${cfg}_cmdline) -o \$(IntDir)$objf" \
+                                Condition="'\$(Configuration)|\$(Platform)'=='$cfg|$plat'"
+                            tag_content Outputs "\$(IntDir)$objf" \
+                                Condition="'\$(Configuration)|\$(Platform)'=='$cfg|$plat'"
+                        done
+                    done
+                    close_tag CustomBuild
+                elif [ "$pat" == "c" ] || \
+                     [ "$pat" == "cc" ] || [ "$pat" == "cpp" ]; then
+                    open_tag ClCompile \
+                        Include="$f"
+                    # Separate file names with Condition?
+                    tag_content ObjectFileName "\$(IntDir)$objf"
+                    # Check for AVX and turn it on to avoid warnings.
+                    if [[ $f =~ avx.?\.c$ ]]; then
+                        tag_content AdditionalOptions "/arch:AVX"
+                    fi
+                    close_tag ClCompile
+                elif [ "$pat" == "h" ] ; then
+                    tag ClInclude \
+                        Include="$f"
+                elif [ "$pat" == "vcxproj" ] ; then
+                    open_tag ProjectReference \
+                        Include="$f"
+                    depguid=`grep ProjectGuid "$f" | sed 's,.*<.*>\(.*\)</.*>.*,\1,'`
+                    tag_content Project "$depguid"
+                    tag_content ReferenceOutputAssembly false
+                    close_tag ProjectReference
+                else
+                    tag None \
+                        Include="$f"
+                fi
+
+                break
+            fi
+        done
+    done
+
+    IFS="$saveIFS"
+}
+
+# Process command line
+unset target
+for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+        --help|-h) show_help
+        ;;
+        --target=*) target="${optval}"
+        ;;
+        --out=*) outfile="$optval"
+        ;;
+        --name=*) name="${optval}"
+        ;;
+        --proj-guid=*) guid="${optval}"
+        ;;
+        --module-def=*) module_def="${optval}"
+        ;;
+        --exe) proj_kind="exe"
+        ;;
+        --dll) proj_kind="dll"
+        ;;
+        --lib) proj_kind="lib"
+        ;;
+        --src-path-bare=*)
+            src_path_bare=$(fix_path "$optval")
+            src_path_bare=${src_path_bare%/}
+        ;;
+        --static-crt) use_static_runtime=true
+        ;;
+        --enable-werror) werror=true
+        ;;
+        --ver=*)
+            vs_ver="$optval"
+            case "$optval" in
+                10|11|12|14)
+                ;;
+                *) die Unrecognized Visual Studio Version in $opt
+                ;;
+            esac
+        ;;
+        -I*)
+            opt=${opt##-I}
+            opt=$(fix_path "$opt")
+            opt="${opt%/}"
+            incs="${incs}${incs:+;}&quot;${opt}&quot;"
+            yasmincs="${yasmincs} -I&quot;${opt}&quot;"
+        ;;
+        -D*) defines="${defines}${defines:+;}${opt##-D}"
+        ;;
+        -L*) # fudge . to $(OutDir)
+            if [ "${opt##-L}" == "." ]; then
+                libdirs="${libdirs}${libdirs:+;}&quot;\$(OutDir)&quot;"
+            else
+                 # Also try directories for this platform/configuration
+                 opt=${opt##-L}
+                 opt=$(fix_path "$opt")
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}&quot;"
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}/\$(PlatformName)/\$(Configuration)&quot;"
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}/\$(PlatformName)&quot;"
+            fi
+        ;;
+        -l*) libs="${libs}${libs:+ }${opt##-l}.lib"
+        ;;
+        -*) die_unknown $opt
+        ;;
+        *)
+            # The paths in file_list are fixed outside of the loop.
+            file_list[${#file_list[@]}]="$opt"
+            case "$opt" in
+                 *.asm|*.s) uses_asm=true
+                 ;;
+            esac
+        ;;
+    esac
+done
+
+# Make one call to fix_path for file_list to improve performance.
+fix_file_list file_list
+
+outfile=${outfile:-/dev/stdout}
+guid=${guid:-`generate_uuid`}
+asm_use_custom_step=false
+uses_asm=${uses_asm:-false}
+case "${vs_ver:-11}" in
+    10|11|12|14)
+       asm_use_custom_step=$uses_asm
+    ;;
+esac
+
+[ -n "$name" ] || die "Project name (--name) must be specified!"
+[ -n "$target" ] || die "Target (--target) must be specified!"
+
+if ${use_static_runtime:-false}; then
+    release_runtime=MultiThreaded
+    debug_runtime=MultiThreadedDebug
+    lib_sfx=mt
+else
+    release_runtime=MultiThreadedDLL
+    debug_runtime=MultiThreadedDebugDLL
+    lib_sfx=md
+fi
+
+# Calculate debug lib names: If a lib ends in ${lib_sfx}.lib, then rename
+# it to ${lib_sfx}d.lib. This precludes linking to release libs from a
+# debug exe, so this may need to be refactored later.
+for lib in ${libs}; do
+    if [ "$lib" != "${lib%${lib_sfx}.lib}" ]; then
+        lib=${lib%.lib}d.lib
+    fi
+    debug_libs="${debug_libs}${debug_libs:+ }${lib}"
+done
+debug_libs=${debug_libs// /;}
+libs=${libs// /;}
+
+
+# List of all platforms supported for this target
+case "$target" in
+    x86_64*)
+        platforms[0]="x64"
+        asm_Debug_cmdline="yasm -Xvc -g cv8 -f win64 ${yasmincs} &quot;%(FullPath)&quot;"
+        asm_Release_cmdline="yasm -Xvc -f win64 ${yasmincs} &quot;%(FullPath)&quot;"
+    ;;
+    x86*)
+        platforms[0]="Win32"
+        asm_Debug_cmdline="yasm -Xvc -g cv8 -f win32 ${yasmincs} &quot;%(FullPath)&quot;"
+        asm_Release_cmdline="yasm -Xvc -f win32 ${yasmincs} &quot;%(FullPath)&quot;"
+    ;;
+    arm*)
+        platforms[0]="ARM"
+        asm_Debug_cmdline="armasm -nologo -oldit &quot;%(FullPath)&quot;"
+        asm_Release_cmdline="armasm -nologo -oldit &quot;%(FullPath)&quot;"
+    ;;
+    *) die "Unsupported target $target!"
+    ;;
+esac
+
+generate_vcxproj() {
+    echo "<?xml version=\"1.0\" encoding=\"utf-8\"?>"
+    open_tag Project \
+        DefaultTargets="Build" \
+        ToolsVersion="4.0" \
+        xmlns="http://schemas.microsoft.com/developer/msbuild/2003" \
+
+    open_tag ItemGroup \
+        Label="ProjectConfigurations"
+    for plat in "${platforms[@]}"; do
+        for config in Debug Release; do
+            open_tag ProjectConfiguration \
+                Include="$config|$plat"
+            tag_content Configuration $config
+            tag_content Platform $plat
+            close_tag ProjectConfiguration
+        done
+    done
+    close_tag ItemGroup
+
+    open_tag PropertyGroup \
+        Label="Globals"
+        tag_content ProjectGuid "{${guid}}"
+        tag_content RootNamespace ${name}
+        tag_content Keyword ManagedCProj
+        if [ $vs_ver -ge 12 ] && [ "${platforms[0]}" = "ARM" ]; then
+            tag_content AppContainerApplication true
+            # The application type can be one of "Windows Store",
+            # "Windows Phone" or "Windows Phone Silverlight". The
+            # actual value doesn't matter from the libvpx point of view,
+            # since a static library built for one works on the others.
+            # The PlatformToolset field needs to be set in sync with this;
+            # for Windows Store and Windows Phone Silverlight it should be
+            # v120 while it should be v120_wp81 if the type is Windows Phone.
+            tag_content ApplicationType "Windows Store"
+            tag_content ApplicationTypeRevision 8.1
+        fi
+    close_tag PropertyGroup
+
+    tag Import \
+        Project="\$(VCTargetsPath)\\Microsoft.Cpp.Default.props"
+
+    for plat in "${platforms[@]}"; do
+        for config in Release Debug; do
+            open_tag PropertyGroup \
+                Condition="'\$(Configuration)|\$(Platform)'=='$config|$plat'" \
+                Label="Configuration"
+            if [ "$proj_kind" = "exe" ]; then
+                tag_content ConfigurationType Application
+            elif [ "$proj_kind" = "dll" ]; then
+                tag_content ConfigurationType DynamicLibrary
+            else
+                tag_content ConfigurationType StaticLibrary
+            fi
+            if [ "$vs_ver" = "11" ]; then
+                if [ "$plat" = "ARM" ]; then
+                    # Setting the wp80 toolchain automatically sets the
+                    # WINAPI_FAMILY define, which is required for building
+                    # code for arm with the windows headers. Alternatively,
+                    # one could add AppContainerApplication=true in the Globals
+                    # section and add PrecompiledHeader=NotUsing and
+                    # CompileAsWinRT=false in ClCompile and SubSystem=Console
+                    # in Link.
+                    tag_content PlatformToolset v110_wp80
+                else
+                    tag_content PlatformToolset v110
+                fi
+            fi
+            if [ "$vs_ver" = "12" ]; then
+                # Setting a PlatformToolset indicating windows phone isn't
+                # enough to build code for arm with MSVC 2013, one strictly
+                # has to enable AppContainerApplication as well.
+                tag_content PlatformToolset v120
+            fi
+            if [ "$vs_ver" = "14" ]; then
+                tag_content PlatformToolset v140
+            fi
+            tag_content CharacterSet Unicode
+            if [ "$config" = "Release" ]; then
+                tag_content WholeProgramOptimization true
+            fi
+            close_tag PropertyGroup
+        done
+    done
+
+    tag Import \
+        Project="\$(VCTargetsPath)\\Microsoft.Cpp.props"
+
+    open_tag ImportGroup \
+        Label="PropertySheets"
+        tag Import \
+            Project="\$(UserRootDir)\\Microsoft.Cpp.\$(Platform).user.props" \
+            Condition="exists('\$(UserRootDir)\\Microsoft.Cpp.\$(Platform).user.props')" \
+            Label="LocalAppDataPlatform"
+    close_tag ImportGroup
+
+    tag PropertyGroup \
+        Label="UserMacros"
+
+    for plat in "${platforms[@]}"; do
+        plat_no_ws=`echo $plat | sed 's/[^A-Za-z0-9_]/_/g'`
+        for config in Debug Release; do
+            open_tag PropertyGroup \
+                Condition="'\$(Configuration)|\$(Platform)'=='$config|$plat'"
+            tag_content OutDir "\$(SolutionDir)$plat_no_ws\\\$(Configuration)\\"
+            tag_content IntDir "$plat_no_ws\\\$(Configuration)\\${name}\\"
+            if [ "$proj_kind" == "lib" ]; then
+              if [ "$config" == "Debug" ]; then
+                config_suffix=d
+              else
+                config_suffix=""
+              fi
+              tag_content TargetName "${name}${lib_sfx}${config_suffix}"
+            fi
+            close_tag PropertyGroup
+        done
+    done
+
+    for plat in "${platforms[@]}"; do
+        for config in Debug Release; do
+            open_tag ItemDefinitionGroup \
+                Condition="'\$(Configuration)|\$(Platform)'=='$config|$plat'"
+            if [ "$name" == "vpx" ]; then
+                hostplat=$plat
+                if [ "$hostplat" == "ARM" ]; then
+                    hostplat=Win32
+                fi
+            fi
+            open_tag ClCompile
+            if [ "$config" = "Debug" ]; then
+                opt=Disabled
+                runtime=$debug_runtime
+                curlibs=$debug_libs
+                debug=_DEBUG
+            else
+                opt=MaxSpeed
+                runtime=$release_runtime
+                curlibs=$libs
+                tag_content FavorSizeOrSpeed Speed
+                debug=NDEBUG
+            fi
+            extradefines=";$defines"
+            tag_content Optimization $opt
+            tag_content AdditionalIncludeDirectories "$incs;%(AdditionalIncludeDirectories)"
+            tag_content PreprocessorDefinitions "WIN32;$debug;_CRT_SECURE_NO_WARNINGS;_CRT_SECURE_NO_DEPRECATE$extradefines;%(PreprocessorDefinitions)"
+            tag_content RuntimeLibrary $runtime
+            tag_content WarningLevel Level3
+            if ${werror:-false}; then
+                tag_content TreatWarningAsError true
+            fi
+            if [ $vs_ver -ge 11 ]; then
+                # We need to override the defaults for these settings
+                # if AppContainerApplication is set.
+                tag_content CompileAsWinRT false
+                tag_content PrecompiledHeader NotUsing
+                tag_content SDLCheck false
+            fi
+            close_tag ClCompile
+            case "$proj_kind" in
+            exe)
+                open_tag Link
+                tag_content GenerateDebugInformation true
+                # Console is the default normally, but if
+                # AppContainerApplication is set, we need to override it.
+                tag_content SubSystem Console
+                close_tag Link
+                ;;
+            dll)
+                open_tag Link
+                tag_content GenerateDebugInformation true
+                tag_content ModuleDefinitionFile $module_def
+                close_tag Link
+                ;;
+            lib)
+                ;;
+            esac
+            close_tag ItemDefinitionGroup
+        done
+
+    done
+
+    open_tag ItemGroup
+    generate_filter "Source Files"   "c;cc;cpp;def;odl;idl;hpj;bat;asm;asmx;s"
+    close_tag ItemGroup
+    open_tag ItemGroup
+    generate_filter "Header Files"   "h;hm;inl;inc;xsd"
+    close_tag ItemGroup
+    open_tag ItemGroup
+    generate_filter "Build Files"    "mk"
+    close_tag ItemGroup
+    open_tag ItemGroup
+    generate_filter "References"     "vcxproj"
+    close_tag ItemGroup
+
+    tag Import \
+        Project="\$(VCTargetsPath)\\Microsoft.Cpp.targets"
+
+    open_tag ImportGroup \
+        Label="ExtensionTargets"
+    close_tag ImportGroup
+
+    close_tag Project
+
+    # This must be done from within the {} subshell
+    echo "Ignored files list (${#file_list[@]} items) is:" >&2
+    for f in "${file_list[@]}"; do
+        echo "    $f" >&2
+    done
+}
+
+# This regexp doesn't catch most of the strings in the vcxproj format,
+# since they're like <tag>path</tag> instead of <tag attr="path" />
+# as previously. It still seems to work ok despite this.
+generate_vcxproj |
+    sed  -e '/"/s;\([^ "]\)/;\1\\;g' |
+    sed  -e '/xmlns/s;\\;/;g' > ${outfile}
+
+exit
diff --git a/libvpx/libvpx/build/make/ios-Info.plist b/libvpx/libvpx/build/make/ios-Info.plist
new file mode 100644
index 0000000..d157b11
--- /dev/null
+++ b/libvpx/libvpx/build/make/ios-Info.plist
@@ -0,0 +1,37 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>CFBundleDevelopmentRegion</key>
+	<string>en</string>
+	<key>CFBundleExecutable</key>
+	<string>VPX</string>
+	<key>CFBundleIdentifier</key>
+	<string>org.webmproject.VPX</string>
+	<key>CFBundleInfoDictionaryVersion</key>
+	<string>6.0</string>
+	<key>CFBundleName</key>
+	<string>VPX</string>
+	<key>CFBundlePackageType</key>
+	<string>FMWK</string>
+	<key>CFBundleShortVersionString</key>
+	<string>${VERSION}</string>
+	<key>CFBundleSignature</key>
+	<string>????</string>
+	<key>CFBundleSupportedPlatforms</key>
+	<array>
+		<string>iPhoneOS</string>
+	</array>
+	<key>CFBundleVersion</key>
+	<string>${VERSION}</string>
+	<key>MinimumOSVersion</key>
+	<string>${IOS_VERSION_MIN}</string>
+	<key>UIDeviceFamily</key>
+	<array>
+		<integer>1</integer>
+		<integer>2</integer>
+	</array>
+	<key>VPXFullVersion</key>
+	<string>${FULLVERSION}</string>
+</dict>
+</plist>
diff --git a/libvpx/libvpx/build/make/iosbuild.sh b/libvpx/libvpx/build/make/iosbuild.sh
new file mode 100755
index 0000000..c703f22
--- /dev/null
+++ b/libvpx/libvpx/build/make/iosbuild.sh
@@ -0,0 +1,383 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##
+## This script generates 'VPX.framework'. An iOS app can encode and decode VPx
+## video by including 'VPX.framework'.
+##
+## Run iosbuild.sh to create 'VPX.framework' in the current directory.
+##
+set -e
+devnull='> /dev/null 2>&1'
+
+BUILD_ROOT="_iosbuild"
+CONFIGURE_ARGS="--disable-docs
+                --disable-examples
+                --disable-libyuv
+                --disable-unit-tests"
+DIST_DIR="_dist"
+FRAMEWORK_DIR="VPX.framework"
+FRAMEWORK_LIB="VPX.framework/VPX"
+HEADER_DIR="${FRAMEWORK_DIR}/Headers/vpx"
+SCRIPT_DIR=$(dirname "$0")
+LIBVPX_SOURCE_DIR=$(cd ${SCRIPT_DIR}/../..; pwd)
+LIPO=$(xcrun -sdk iphoneos${SDK} -find lipo)
+ORIG_PWD="$(pwd)"
+ARM_TARGETS="arm64-darwin-gcc
+             armv7-darwin-gcc
+             armv7s-darwin-gcc"
+SIM_TARGETS="x86-iphonesimulator-gcc
+             x86_64-iphonesimulator-gcc"
+OSX_TARGETS="x86-darwin15-gcc
+             x86_64-darwin15-gcc"
+TARGETS="${ARM_TARGETS} ${SIM_TARGETS}"
+
+# Configures for the target specified by $1, and invokes make with the dist
+# target using $DIST_DIR as the distribution output directory.
+build_target() {
+  local target="$1"
+  local old_pwd="$(pwd)"
+  local target_specific_flags=""
+
+  vlog "***Building target: ${target}***"
+
+  case "${target}" in
+    x86-*)
+      target_specific_flags="--enable-pic"
+      vlog "Enabled PIC for ${target}"
+      ;;
+  esac
+
+  mkdir "${target}"
+  cd "${target}"
+  eval "${LIBVPX_SOURCE_DIR}/configure" --target="${target}" \
+    ${CONFIGURE_ARGS} ${EXTRA_CONFIGURE_ARGS} ${target_specific_flags} \
+    ${devnull}
+  export DIST_DIR
+  eval make dist ${devnull}
+  cd "${old_pwd}"
+
+  vlog "***Done building target: ${target}***"
+}
+
+# Returns the preprocessor symbol for the target specified by $1.
+target_to_preproc_symbol() {
+  target="$1"
+  case "${target}" in
+    arm64-*)
+      echo "__aarch64__"
+      ;;
+    armv7-*)
+      echo "__ARM_ARCH_7A__"
+      ;;
+    armv7s-*)
+      echo "__ARM_ARCH_7S__"
+      ;;
+    x86-*)
+      echo "__i386__"
+      ;;
+    x86_64-*)
+      echo "__x86_64__"
+      ;;
+    *)
+      echo "#error ${target} unknown/unsupported"
+      return 1
+      ;;
+  esac
+}
+
+# Create a vpx_config.h shim that, based on preprocessor settings for the
+# current target CPU, includes the real vpx_config.h for the current target.
+# $1 is the list of targets.
+create_vpx_framework_config_shim() {
+  local targets="$1"
+  local config_file="${HEADER_DIR}/vpx_config.h"
+  local preproc_symbol=""
+  local target=""
+  local include_guard="VPX_FRAMEWORK_HEADERS_VPX_VPX_CONFIG_H_"
+
+  local file_header="/*
+ *  Copyright (c) $(date +%Y) The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+/* GENERATED FILE: DO NOT EDIT! */
+
+#ifndef ${include_guard}
+#define ${include_guard}
+
+#if defined"
+
+  printf "%s" "${file_header}" > "${config_file}"
+  for target in ${targets}; do
+    preproc_symbol=$(target_to_preproc_symbol "${target}")
+    printf " ${preproc_symbol}\n" >> "${config_file}"
+    printf "#define VPX_FRAMEWORK_TARGET \"${target}\"\n" >> "${config_file}"
+    printf "#include \"VPX/vpx/${target}/vpx_config.h\"\n" >> "${config_file}"
+    printf "#elif defined" >> "${config_file}"
+    mkdir "${HEADER_DIR}/${target}"
+    cp -p "${BUILD_ROOT}/${target}/vpx_config.h" "${HEADER_DIR}/${target}"
+  done
+
+  # Consume the last line of output from the loop: We don't want it.
+  sed -i '' -e '$d' "${config_file}"
+
+  printf "#endif\n\n" >> "${config_file}"
+  printf "#endif  // ${include_guard}" >> "${config_file}"
+}
+
+# Verifies that $FRAMEWORK_LIB fat library contains requested builds.
+verify_framework_targets() {
+  local requested_cpus=""
+  local cpu=""
+
+  # Extract CPU from full target name.
+  for target; do
+    cpu="${target%%-*}"
+    if [ "${cpu}" = "x86" ]; then
+      # lipo -info outputs i386 for libvpx x86 targets.
+      cpu="i386"
+    fi
+    requested_cpus="${requested_cpus}${cpu} "
+  done
+
+  # Get target CPUs present in framework library.
+  local targets_built=$(${LIPO} -info ${FRAMEWORK_LIB})
+
+  # $LIPO -info outputs a string like the following:
+  #   Architectures in the fat file: $FRAMEWORK_LIB <architectures>
+  # Capture only the architecture strings.
+  targets_built=${targets_built##*: }
+
+  # Sort CPU strings to make the next step a simple string compare.
+  local actual=$(echo ${targets_built} | tr " " "\n" | sort | tr "\n" " ")
+  local requested=$(echo ${requested_cpus} | tr " " "\n" | sort | tr "\n" " ")
+
+  vlog "Requested ${FRAMEWORK_LIB} CPUs: ${requested}"
+  vlog "Actual ${FRAMEWORK_LIB} CPUs: ${actual}"
+
+  if [ "${requested}" != "${actual}" ]; then
+    elog "Actual ${FRAMEWORK_LIB} targets do not match requested target list."
+    elog "  Requested target CPUs: ${requested}"
+    elog "  Actual target CPUs: ${actual}"
+    return 1
+  fi
+}
+
+# Configures and builds each target specified by $1, and then builds
+# VPX.framework.
+build_framework() {
+  local lib_list=""
+  local targets="$1"
+  local target=""
+  local target_dist_dir=""
+
+  # Clean up from previous build(s).
+  rm -rf "${BUILD_ROOT}" "${FRAMEWORK_DIR}"
+
+  # Create output dirs.
+  mkdir -p "${BUILD_ROOT}"
+  mkdir -p "${HEADER_DIR}"
+
+  cd "${BUILD_ROOT}"
+
+  for target in ${targets}; do
+    build_target "${target}"
+    target_dist_dir="${BUILD_ROOT}/${target}/${DIST_DIR}"
+    if [ "${ENABLE_SHARED}" = "yes" ]; then
+      local suffix="dylib"
+    else
+      local suffix="a"
+    fi
+    lib_list="${lib_list} ${target_dist_dir}/lib/libvpx.${suffix}"
+  done
+
+  cd "${ORIG_PWD}"
+
+  # The basic libvpx API includes are all the same; just grab the most recent
+  # set.
+  cp -p "${target_dist_dir}"/include/vpx/* "${HEADER_DIR}"
+
+  # Build the fat library.
+  ${LIPO} -create ${lib_list} -output ${FRAMEWORK_DIR}/VPX
+
+  # Create the vpx_config.h shim that allows usage of vpx_config.h from
+  # within VPX.framework.
+  create_vpx_framework_config_shim "${targets}"
+
+  # Copy in vpx_version.h.
+  cp -p "${BUILD_ROOT}/${target}/vpx_version.h" "${HEADER_DIR}"
+
+  if [ "${ENABLE_SHARED}" = "yes" ]; then
+    # Adjust the dylib's name so dynamic linking in apps works as expected.
+    install_name_tool -id '@rpath/VPX.framework/VPX' ${FRAMEWORK_DIR}/VPX
+
+    # Copy in Info.plist.
+    cat "${SCRIPT_DIR}/ios-Info.plist" \
+      | sed "s/\${FULLVERSION}/${FULLVERSION}/g" \
+      | sed "s/\${VERSION}/${VERSION}/g" \
+      | sed "s/\${IOS_VERSION_MIN}/${IOS_VERSION_MIN}/g" \
+      > "${FRAMEWORK_DIR}/Info.plist"
+  fi
+
+  # Confirm VPX.framework/VPX contains the targets requested.
+  verify_framework_targets ${targets}
+
+  vlog "Created fat library ${FRAMEWORK_LIB} containing:"
+  for lib in ${lib_list}; do
+    vlog "  $(echo ${lib} | awk -F / '{print $2, $NF}')"
+  done
+}
+
+# Trap function. Cleans up the subtree used to build all targets contained in
+# $TARGETS.
+cleanup() {
+  local readonly res=$?
+  cd "${ORIG_PWD}"
+
+  if [ $res -ne 0 ]; then
+    elog "build exited with error ($res)"
+  fi
+
+  if [ "${PRESERVE_BUILD_OUTPUT}" != "yes" ]; then
+    rm -rf "${BUILD_ROOT}"
+  fi
+}
+
+print_list() {
+  local indent="$1"
+  shift
+  local list="$@"
+  for entry in ${list}; do
+    echo "${indent}${entry}"
+  done
+}
+
+iosbuild_usage() {
+cat << EOF
+  Usage: ${0##*/} [arguments]
+    --help: Display this message and exit.
+    --enable-shared: Build a dynamic framework for use on iOS 8 or later.
+    --extra-configure-args <args>: Extra args to pass when configuring libvpx.
+    --macosx: Uses darwin15 targets instead of iphonesimulator targets for x86
+              and x86_64. Allows linking to framework when builds target MacOSX
+              instead of iOS.
+    --preserve-build-output: Do not delete the build directory.
+    --show-build-output: Show output from each library build.
+    --targets <targets>: Override default target list. Defaults:
+$(print_list "        " ${TARGETS})
+    --test-link: Confirms all targets can be linked. Functionally identical to
+                 passing --enable-examples via --extra-configure-args.
+    --verbose: Output information about the environment and each stage of the
+               build.
+EOF
+}
+
+elog() {
+  echo "${0##*/} failed because: $@" 1>&2
+}
+
+vlog() {
+  if [ "${VERBOSE}" = "yes" ]; then
+    echo "$@"
+  fi
+}
+
+trap cleanup EXIT
+
+# Parse the command line.
+while [ -n "$1" ]; do
+  case "$1" in
+    --extra-configure-args)
+      EXTRA_CONFIGURE_ARGS="$2"
+      shift
+      ;;
+    --help)
+      iosbuild_usage
+      exit
+      ;;
+    --enable-shared)
+      ENABLE_SHARED=yes
+      ;;
+    --preserve-build-output)
+      PRESERVE_BUILD_OUTPUT=yes
+      ;;
+    --show-build-output)
+      devnull=
+      ;;
+    --test-link)
+      EXTRA_CONFIGURE_ARGS="${EXTRA_CONFIGURE_ARGS} --enable-examples"
+      ;;
+    --targets)
+      TARGETS="$2"
+      shift
+      ;;
+    --macosx)
+      TARGETS="${ARM_TARGETS} ${OSX_TARGETS}"
+      ;;
+    --verbose)
+      VERBOSE=yes
+      ;;
+    *)
+      iosbuild_usage
+      exit 1
+      ;;
+  esac
+  shift
+done
+
+if [ "${ENABLE_SHARED}" = "yes" ]; then
+  CONFIGURE_ARGS="--enable-shared ${CONFIGURE_ARGS}"
+fi
+
+FULLVERSION=$("${SCRIPT_DIR}"/version.sh --bare "${LIBVPX_SOURCE_DIR}")
+VERSION=$(echo "${FULLVERSION}" | sed -E 's/^v([0-9]+\.[0-9]+\.[0-9]+).*$/\1/')
+
+if [ "$ENABLE_SHARED" = "yes" ]; then
+  IOS_VERSION_OPTIONS="--enable-shared"
+  IOS_VERSION_MIN="8.0"
+else
+  IOS_VERSION_OPTIONS=""
+  IOS_VERSION_MIN="6.0"
+fi
+
+if [ "${VERBOSE}" = "yes" ]; then
+cat << EOF
+  BUILD_ROOT=${BUILD_ROOT}
+  DIST_DIR=${DIST_DIR}
+  CONFIGURE_ARGS=${CONFIGURE_ARGS}
+  EXTRA_CONFIGURE_ARGS=${EXTRA_CONFIGURE_ARGS}
+  FRAMEWORK_DIR=${FRAMEWORK_DIR}
+  FRAMEWORK_LIB=${FRAMEWORK_LIB}
+  HEADER_DIR=${HEADER_DIR}
+  LIBVPX_SOURCE_DIR=${LIBVPX_SOURCE_DIR}
+  LIPO=${LIPO}
+  MAKEFLAGS=${MAKEFLAGS}
+  ORIG_PWD=${ORIG_PWD}
+  PRESERVE_BUILD_OUTPUT=${PRESERVE_BUILD_OUTPUT}
+  TARGETS="$(print_list "" ${TARGETS})"
+  ENABLE_SHARED=${ENABLE_SHARED}
+  OSX_TARGETS="${OSX_TARGETS}"
+  SIM_TARGETS="${SIM_TARGETS}"
+  SCRIPT_DIR="${SCRIPT_DIR}"
+  FULLVERSION="${FULLVERSION}"
+  VERSION="${VERSION}"
+  IOS_VERSION_MIN="${IOS_VERSION_MIN}"
+EOF
+fi
+
+build_framework "${TARGETS}"
+echo "Successfully built '${FRAMEWORK_DIR}' for:"
+print_list "" ${TARGETS}
diff --git a/libvpx/libvpx/build/make/msvs_common.sh b/libvpx/libvpx/build/make/msvs_common.sh
new file mode 100755
index 0000000..88f1cf9
--- /dev/null
+++ b/libvpx/libvpx/build/make/msvs_common.sh
@@ -0,0 +1,114 @@
+#!/bin/bash
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+if [ "$(uname -o 2>/dev/null)" = "Cygwin" ] \
+   && cygpath --help >/dev/null 2>&1; then
+    FIXPATH='cygpath -m'
+else
+    FIXPATH='echo_path'
+fi
+
+die() {
+    echo "${self_basename}: $@" >&2
+    exit 1
+}
+
+die_unknown(){
+    echo "Unknown option \"$1\"." >&2
+    echo "See ${self_basename} --help for available options." >&2
+    exit 1
+}
+
+echo_path() {
+    for path; do
+        echo "$path"
+    done
+}
+
+# Output one, possibly changed based on the system, path per line.
+fix_path() {
+    $FIXPATH "$@"
+}
+
+# Corrects the paths in file_list in one pass for efficiency.
+# $1 is the name of the array to be modified.
+fix_file_list() {
+    declare -n array_ref=$1
+    files=$(fix_path "${array_ref[@]}")
+    local IFS=$'\n'
+    array_ref=($files)
+}
+
+generate_uuid() {
+    local hex="0123456789ABCDEF"
+    local i
+    local uuid=""
+    local j
+    #93995380-89BD-4b04-88EB-625FBE52EBFB
+    for ((i=0; i<32; i++)); do
+        (( j = $RANDOM % 16 ))
+        uuid="${uuid}${hex:$j:1}"
+    done
+    echo "${uuid:0:8}-${uuid:8:4}-${uuid:12:4}-${uuid:16:4}-${uuid:20:12}"
+}
+
+indent1="    "
+indent=""
+indent_push() {
+    indent="${indent}${indent1}"
+}
+indent_pop() {
+    indent="${indent%${indent1}}"
+}
+
+tag_attributes() {
+    for opt in "$@"; do
+        optval="${opt#*=}"
+        [ -n "${optval}" ] ||
+            die "Missing attribute value in '$opt' while generating $tag tag"
+        echo "${indent}${opt%%=*}=\"${optval}\""
+    done
+}
+
+open_tag() {
+    local tag=$1
+    shift
+    if [ $# -ne 0 ]; then
+        echo "${indent}<${tag}"
+        indent_push
+        tag_attributes "$@"
+        echo "${indent}>"
+    else
+        echo "${indent}<${tag}>"
+        indent_push
+    fi
+}
+
+close_tag() {
+    local tag=$1
+    indent_pop
+    echo "${indent}</${tag}>"
+}
+
+tag() {
+    local tag=$1
+    shift
+    if [ $# -ne 0 ]; then
+        echo "${indent}<${tag}"
+        indent_push
+        tag_attributes "$@"
+        indent_pop
+        echo "${indent}/>"
+    else
+        echo "${indent}<${tag}/>"
+    fi
+}
+
diff --git a/libvpx/libvpx/build/make/rtcd.pl b/libvpx/libvpx/build/make/rtcd.pl
new file mode 100755
index 0000000..991b6ab
--- /dev/null
+++ b/libvpx/libvpx/build/make/rtcd.pl
@@ -0,0 +1,426 @@
+#!/usr/bin/env perl
+
+no strict 'refs';
+use warnings;
+use Getopt::Long;
+Getopt::Long::Configure("auto_help") if $Getopt::Long::VERSION > 2.32;
+
+my %ALL_FUNCS = ();
+my @ALL_ARCHS;
+my @ALL_FORWARD_DECLS;
+my @REQUIRES;
+
+my %opts = ();
+my %disabled = ();
+my %required = ();
+
+my @argv;
+foreach (@ARGV) {
+  $disabled{$1} = 1, next if /--disable-(.*)/;
+  $required{$1} = 1, next if /--require-(.*)/;
+  push @argv, $_;
+}
+
+# NB: use GetOptions() instead of GetOptionsFromArray() for compatibility.
+@ARGV = @argv;
+GetOptions(
+  \%opts,
+  'arch=s',
+  'sym=s',
+  'config=s',
+);
+
+foreach my $opt (qw/arch config/) {
+  if (!defined($opts{$opt})) {
+    warn "--$opt is required!\n";
+    Getopt::Long::HelpMessage('-exit' => 1);
+  }
+}
+
+foreach my $defs_file (@ARGV) {
+  if (!-f $defs_file) {
+    warn "$defs_file: $!\n";
+    Getopt::Long::HelpMessage('-exit' => 1);
+  }
+}
+
+open CONFIG_FILE, $opts{config} or
+  die "Error opening config file '$opts{config}': $!\n";
+
+my %config = ();
+while (<CONFIG_FILE>) {
+  next if !/^(?:CONFIG_|HAVE_)/;
+  chomp;
+  my @pair = split /=/;
+  $config{$pair[0]} = $pair[1];
+}
+close CONFIG_FILE;
+
+#
+# Routines for the RTCD DSL to call
+#
+sub vpx_config($) {
+  return (defined $config{$_[0]}) ? $config{$_[0]} : "";
+}
+
+sub specialize {
+  my $fn=$_[0];
+  shift;
+  foreach my $opt (@_) {
+    eval "\$${fn}_${opt}=${fn}_${opt}";
+  }
+}
+
+sub add_proto {
+  my $fn = splice(@_, -2, 1);
+  $ALL_FUNCS{$fn} = \@_;
+  specialize $fn, "c";
+}
+
+sub require {
+  foreach my $fn (keys %ALL_FUNCS) {
+    foreach my $opt (@_) {
+      my $ofn = eval "\$${fn}_${opt}";
+      next if !$ofn;
+
+      # if we already have a default, then we can disable it, as we know
+      # we can do better.
+      my $best = eval "\$${fn}_default";
+      if ($best) {
+        my $best_ofn = eval "\$${best}";
+        if ($best_ofn && "$best_ofn" ne "$ofn") {
+          eval "\$${best}_link = 'false'";
+        }
+      }
+      eval "\$${fn}_default=${fn}_${opt}";
+      eval "\$${fn}_${opt}_link='true'";
+    }
+  }
+}
+
+sub forward_decls {
+  push @ALL_FORWARD_DECLS, @_;
+}
+
+#
+# Include the user's directives
+#
+foreach my $f (@ARGV) {
+  open FILE, "<", $f or die "cannot open $f: $!\n";
+  my $contents = join('', <FILE>);
+  close FILE;
+  eval $contents or warn "eval failed: $@\n";
+}
+
+#
+# Process the directives according to the command line
+#
+sub process_forward_decls() {
+  foreach (@ALL_FORWARD_DECLS) {
+    $_->();
+  }
+}
+
+sub determine_indirection {
+  vpx_config("CONFIG_RUNTIME_CPU_DETECT") eq "yes" or &require(@ALL_ARCHS);
+  foreach my $fn (keys %ALL_FUNCS) {
+    my $n = "";
+    my @val = @{$ALL_FUNCS{$fn}};
+    my $args = pop @val;
+    my $rtyp = "@val";
+    my $dfn = eval "\$${fn}_default";
+    $dfn = eval "\$${dfn}";
+    foreach my $opt (@_) {
+      my $ofn = eval "\$${fn}_${opt}";
+      next if !$ofn;
+      my $link = eval "\$${fn}_${opt}_link";
+      next if $link && $link eq "false";
+      $n .= "x";
+    }
+    if ($n eq "x") {
+      eval "\$${fn}_indirect = 'false'";
+    } else {
+      eval "\$${fn}_indirect = 'true'";
+    }
+  }
+}
+
+sub declare_function_pointers {
+  foreach my $fn (sort keys %ALL_FUNCS) {
+    my @val = @{$ALL_FUNCS{$fn}};
+    my $args = pop @val;
+    my $rtyp = "@val";
+    my $dfn = eval "\$${fn}_default";
+    $dfn = eval "\$${dfn}";
+    foreach my $opt (@_) {
+      my $ofn = eval "\$${fn}_${opt}";
+      next if !$ofn;
+      print "$rtyp ${ofn}($args);\n";
+    }
+    if (eval "\$${fn}_indirect" eq "false") {
+      print "#define ${fn} ${dfn}\n";
+    } else {
+      print "RTCD_EXTERN $rtyp (*${fn})($args);\n";
+    }
+    print "\n";
+  }
+}
+
+sub set_function_pointers {
+  foreach my $fn (sort keys %ALL_FUNCS) {
+    my @val = @{$ALL_FUNCS{$fn}};
+    my $args = pop @val;
+    my $rtyp = "@val";
+    my $dfn = eval "\$${fn}_default";
+    $dfn = eval "\$${dfn}";
+    if (eval "\$${fn}_indirect" eq "true") {
+      print "    $fn = $dfn;\n";
+      foreach my $opt (@_) {
+        my $ofn = eval "\$${fn}_${opt}";
+        next if !$ofn;
+        next if "$ofn" eq "$dfn";
+        my $link = eval "\$${fn}_${opt}_link";
+        next if $link && $link eq "false";
+        my $cond = eval "\$have_${opt}";
+        print "    if (${cond}) $fn = $ofn;\n"
+      }
+    }
+  }
+}
+
+sub filter {
+  my @filtered;
+  foreach (@_) { push @filtered, $_ unless $disabled{$_}; }
+  return @filtered;
+}
+
+#
+# Helper functions for generating the arch specific RTCD files
+#
+sub common_top() {
+  my $include_guard = uc($opts{sym})."_H_";
+  print <<EOF;
+#ifndef ${include_guard}
+#define ${include_guard}
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+EOF
+
+process_forward_decls();
+print <<EOF;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+EOF
+declare_function_pointers("c", @ALL_ARCHS);
+
+print <<EOF;
+void $opts{sym}(void);
+
+EOF
+}
+
+sub common_bottom() {
+  print <<EOF;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
+EOF
+}
+
+sub x86() {
+  determine_indirection("c", @ALL_ARCHS);
+
+  # Assign the helper variable for each enabled extension
+  foreach my $opt (@ALL_ARCHS) {
+    my $opt_uc = uc $opt;
+    eval "\$have_${opt}=\"flags & HAS_${opt_uc}\"";
+  }
+
+  common_top;
+  print <<EOF;
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+EOF
+
+  set_function_pointers("c", @ALL_ARCHS);
+
+  print <<EOF;
+}
+#endif
+EOF
+  common_bottom;
+}
+
+sub arm() {
+  determine_indirection("c", @ALL_ARCHS);
+
+  # Assign the helper variable for each enabled extension
+  foreach my $opt (@ALL_ARCHS) {
+    my $opt_uc = uc $opt;
+    # Enable neon assembly based on HAVE_NEON logic instead of adding new
+    # HAVE_NEON_ASM logic
+    if ($opt eq 'neon_asm') { $opt_uc = 'NEON' }
+    eval "\$have_${opt}=\"flags & HAS_${opt_uc}\"";
+  }
+
+  common_top;
+  print <<EOF;
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+#include "vpx_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+EOF
+
+  set_function_pointers("c", @ALL_ARCHS);
+
+  print <<EOF;
+}
+#endif
+EOF
+  common_bottom;
+}
+
+sub mips() {
+  determine_indirection("c", @ALL_ARCHS);
+  common_top;
+
+  print <<EOF;
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+EOF
+
+  set_function_pointers("c", @ALL_ARCHS);
+
+  print <<EOF;
+#if HAVE_DSPR2
+void vpx_dsputil_static_init();
+#if CONFIG_VP8
+void dsputil_static_init();
+#endif
+
+vpx_dsputil_static_init();
+#if CONFIG_VP8
+dsputil_static_init();
+#endif
+#endif
+}
+#endif
+EOF
+  common_bottom;
+}
+
+sub unoptimized() {
+  determine_indirection "c";
+  common_top;
+  print <<EOF;
+#include "vpx_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+EOF
+
+  set_function_pointers "c";
+
+  print <<EOF;
+}
+#endif
+EOF
+  common_bottom;
+}
+
+#
+# Main Driver
+#
+
+&require("c");
+if ($opts{arch} eq 'x86') {
+  @ALL_ARCHS = filter(qw/mmx sse sse2 sse3 ssse3 sse4_1 avx avx2/);
+  x86;
+} elsif ($opts{arch} eq 'x86_64') {
+  @ALL_ARCHS = filter(qw/mmx sse sse2 sse3 ssse3 sse4_1 avx avx2/);
+  @REQUIRES = filter(keys %required ? keys %required : qw/mmx sse sse2/);
+  &require(@REQUIRES);
+  x86;
+} elsif ($opts{arch} eq 'mips32' || $opts{arch} eq 'mips64') {
+  @ALL_ARCHS = filter("$opts{arch}");
+  open CONFIG_FILE, $opts{config} or
+    die "Error opening config file '$opts{config}': $!\n";
+  while (<CONFIG_FILE>) {
+    if (/HAVE_DSPR2=yes/) {
+      @ALL_ARCHS = filter("$opts{arch}", qw/dspr2/);
+      last;
+    }
+    if (/HAVE_MSA=yes/) {
+      @ALL_ARCHS = filter("$opts{arch}", qw/msa/);
+      last;
+    }
+  }
+  close CONFIG_FILE;
+  mips;
+} elsif ($opts{arch} eq 'armv6') {
+  @ALL_ARCHS = filter(qw/media/);
+  arm;
+} elsif ($opts{arch} =~ /armv7\w?/) {
+  @ALL_ARCHS = filter(qw/media neon_asm neon/);
+  @REQUIRES = filter(keys %required ? keys %required : qw/media/);
+  &require(@REQUIRES);
+  arm;
+} elsif ($opts{arch} eq 'armv8' || $opts{arch} eq 'arm64' ) {
+  @ALL_ARCHS = filter(qw/neon/);
+  arm;
+} else {
+  unoptimized;
+}
+
+__END__
+
+=head1 NAME
+
+rtcd -
+
+=head1 SYNOPSIS
+
+Usage: rtcd.pl [options] FILE
+
+See 'perldoc rtcd.pl' for more details.
+
+=head1 DESCRIPTION
+
+Reads the Run Time CPU Detections definitions from FILE and generates a
+C header file on stdout.
+
+=head1 OPTIONS
+
+Options:
+  --arch=ARCH       Architecture to generate defs for (required)
+  --disable-EXT     Disable support for EXT extensions
+  --require-EXT     Require support for EXT extensions
+  --sym=SYMBOL      Unique symbol to use for RTCD initialization function
+  --config=FILE     File with CONFIG_FOO=yes lines to parse
diff --git a/libvpx/libvpx/build/make/thumb.pm b/libvpx/libvpx/build/make/thumb.pm
new file mode 100644
index 0000000..483c253
--- /dev/null
+++ b/libvpx/libvpx/build/make/thumb.pm
@@ -0,0 +1,70 @@
+#!/usr/bin/env perl
+##
+##  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+package thumb;
+
+sub FixThumbInstructions($$)
+{
+    my $short_branches = $_[1];
+    my $branch_shift_offset = $short_branches ? 1 : 0;
+
+    # Write additions with shifts, such as "add r10, r11, lsl #8",
+    # in three operand form, "add r10, r10, r11, lsl #8".
+    s/(add\s+)(r\d+),\s*(r\d+),\s*(lsl #\d+)/$1$2, $2, $3, $4/g;
+
+    # Convert additions with a non-constant shift into a sequence
+    # with left shift, addition and a right shift (to restore the
+    # register to the original value). Currently the right shift
+    # isn't necessary in the code base since the values in these
+    # registers aren't used, but doing the shift for consistency.
+    # This converts instructions such as "add r12, r12, r5, lsl r4"
+    # into the sequence "lsl r5, r4", "add r12, r12, r5", "lsr r5, r4".
+    s/^(\s*)(add)(\s+)(r\d+),\s*(r\d+),\s*(r\d+),\s*lsl (r\d+)/$1lsl$3$6, $7\n$1$2$3$4, $5, $6\n$1lsr$3$6, $7/g;
+
+    # Convert loads with right shifts in the indexing into a
+    # sequence of an add, load and sub. This converts
+    # "ldrb r4, [r9, lr, asr #1]" into "add r9, r9, lr, asr #1",
+    # "ldrb r9, [r9]", "sub r9, r9, lr, asr #1".
+    s/^(\s*)(ldrb)(\s+)(r\d+),\s*\[(\w+),\s*(\w+),\s*(asr #\d+)\]/$1add $3$5, $5, $6, $7\n$1$2$3$4, [$5]\n$1sub $3$5, $5, $6, $7/g;
+
+    # Convert register indexing with writeback into a separate add
+    # instruction. This converts "ldrb r12, [r1, r2]!" into
+    # "ldrb r12, [r1, r2]", "add r1, r1, r2".
+    s/^(\s*)(ldrb)(\s+)(r\d+),\s*\[(\w+),\s*(\w+)\]!/$1$2$3$4, [$5, $6]\n$1add $3$5, $6/g;
+
+    # Convert negative register indexing into separate sub/add instructions.
+    # This converts "ldrne r4, [src, -pstep, lsl #1]" into
+    # "subne src, src, pstep, lsl #1", "ldrne r4, [src]",
+    # "addne src, src, pstep, lsl #1". In a couple of cases where
+    # this is used, it's used for two subsequent load instructions,
+    # where a hand-written version of it could merge two subsequent
+    # add and sub instructions.
+    s/^(\s*)((ldr|str|pld)(ne)?)(\s+)(r\d+,\s*)?\[(\w+), -([^\]]+)\]/$1sub$4$5$7, $7, $8\n$1$2$5$6\[$7\]\n$1add$4$5$7, $7, $8/g;
+
+    # Convert register post indexing to a separate add instruction.
+    # This converts "ldrneb r9, [r0], r2" into "ldrneb r9, [r0]",
+    # "addne r0, r0, r2".
+    s/^(\s*)((ldr|str)(ne)?[bhd]?)(\s+)(\w+),(\s*\w+,)?\s*\[(\w+)\],\s*(\w+)/$1$2$5$6,$7 [$8]\n$1add$4$5$8, $8, $9/g;
+
+    # Convert a conditional addition to the pc register into a series of
+    # instructions. This converts "addlt pc, pc, r3, lsl #2" into
+    # "itttt lt", "movlt.n r12, pc", "addlt.w r12, #12",
+    # "addlt.w r12, r12, r3, lsl #2", "movlt.n pc, r12".
+    # This assumes that r12 is free at this point.
+    s/^(\s*)addlt(\s+)pc,\s*pc,\s*(\w+),\s*lsl\s*#(\d+)/$1itttt$2lt\n$1movlt.n$2r12, pc\n$1addlt.w$2r12, #12\n$1addlt.w$2r12, r12, $3, lsl #($4-$branch_shift_offset)\n$1movlt.n$2pc, r12/g;
+
+    # Convert "mov pc, lr" into "bx lr", since the former only works
+    # for switching from arm to thumb (and only in armv7), but not
+    # from thumb to arm.
+    s/mov(\s*)pc\s*,\s*lr/bx$1lr/g;
+}
+
+1;
diff --git a/libvpx/libvpx/build/make/version.sh b/libvpx/libvpx/build/make/version.sh
new file mode 100755
index 0000000..6967527
--- /dev/null
+++ b/libvpx/libvpx/build/make/version.sh
@@ -0,0 +1,77 @@
+#!/bin/sh
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+
+for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+    --bare) bare=true ;;
+    *) break ;;
+    esac
+    shift
+done
+source_path=${1:-.}
+out_file=${2}
+id=${3:-VERSION_STRING}
+
+git_version_id=""
+if [ -e "${source_path}/.git" ]; then
+    # Source Path is a git working copy. Check for local modifications.
+    # Note that git submodules may have a file as .git, not a directory.
+    export GIT_DIR="${source_path}/.git"
+    git_version_id=`git describe --match=v[0-9]* 2>/dev/null`
+fi
+
+changelog_version=""
+for p in "${source_path}" "${source_path}/.."; do
+    if [ -z "$git_version_id" -a -f "${p}/CHANGELOG" ]; then
+        changelog_version=`head -n1 "${p}/CHANGELOG" | awk '{print $2}'`
+        changelog_version="${changelog_version}"
+        break
+    fi
+done
+version_str="${changelog_version}${git_version_id}"
+bare_version=${version_str#v}
+major_version=${bare_version%%.*}
+bare_version=${bare_version#*.}
+minor_version=${bare_version%%.*}
+bare_version=${bare_version#*.}
+patch_version=${bare_version%%-*}
+bare_version=${bare_version#${patch_version}}
+extra_version=${bare_version##-}
+
+#since they'll be used as integers below make sure they are or force to 0
+for v in major_version minor_version patch_version; do
+    if eval echo \$$v |grep -E -q '[^[:digit:]]'; then
+        eval $v=0
+    fi
+done
+
+if [ ${bare} ]; then
+    echo "${changelog_version}${git_version_id}" > $$.tmp
+else
+    cat<<EOF>$$.tmp
+#define VERSION_MAJOR  $major_version
+#define VERSION_MINOR  $minor_version
+#define VERSION_PATCH  $patch_version
+#define VERSION_EXTRA  "$extra_version"
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define ${id}_NOSP "${version_str}"
+#define ${id}      " ${version_str}"
+EOF
+fi
+if [ -n "$out_file" ]; then
+diff $$.tmp ${out_file} >/dev/null 2>&1 || cat $$.tmp > ${out_file}
+else
+cat $$.tmp
+fi
+rm $$.tmp
diff --git a/libvpx/libvpx/codereview.settings b/libvpx/libvpx/codereview.settings
new file mode 100644
index 0000000..d7c8d39
--- /dev/null
+++ b/libvpx/libvpx/codereview.settings
@@ -0,0 +1,4 @@
+# This file is used by gcl to get repository specific information.
+GERRIT_HOST: chromium-review.googlesource.com
+GERRIT_PORT: 29418
+CODE_REVIEW_SERVER: chromium-review.googlesource.com
diff --git a/libvpx/libvpx/configure b/libvpx/libvpx/configure
new file mode 100755
index 0000000..f82ee04
--- /dev/null
+++ b/libvpx/libvpx/configure
@@ -0,0 +1,743 @@
+#!/bin/sh
+##
+##  configure
+##
+##  This script is the front-end to the build system. It provides a similar
+##  interface to standard configure scripts with some extra bits for dealing
+##  with toolchains that differ from the standard POSIX interface and
+##  for extracting subsets of the source tree. In theory, reusable parts
+##  of this script were intended to live in build/make/configure.sh,
+##  but in practice, the line is pretty blurry.
+##
+##  This build system is based in part on the FFmpeg configure script.
+##
+
+#source_path="`dirname \"$0\"`"
+source_path=${0%/*}
+. "${source_path}/build/make/configure.sh"
+
+show_help(){
+    show_help_pre
+    cat << EOF
+Advanced options:
+  ${toggle_libs}                  libraries
+  ${toggle_examples}              examples
+  ${toggle_docs}                  documentation
+  ${toggle_unit_tests}            unit tests
+  ${toggle_decode_perf_tests}     build decoder perf tests with unit tests
+  ${toggle_encode_perf_tests}     build encoder perf tests with unit tests
+  --cpu=CPU                       tune for the specified CPU (ARM: cortex-a8, X86: sse3)
+  --libc=PATH                     path to alternate libc
+  --size-limit=WxH                max size to allow in the decoder
+  --as={yasm|nasm|auto}           use specified assembler [auto, yasm preferred]
+  --sdk-path=PATH                 path to root of sdk (android builds only)
+  ${toggle_codec_srcs}            in/exclude codec library source code
+  ${toggle_debug_libs}            in/exclude debug version of libraries
+  ${toggle_static_msvcrt}         use static MSVCRT (VS builds only)
+  ${toggle_vp9_highbitdepth}      use VP9 high bit depth (10/12) profiles
+  ${toggle_better_hw_compatibility}
+                                  enable encoder to produce streams with better
+                                  hardware decoder compatibility
+  ${toggle_vp8}                   VP8 codec support
+  ${toggle_vp9}                   VP9 codec support
+  ${toggle_internal_stats}        output of encoder internal stats for debug, if supported (encoders)
+  ${toggle_postproc}              postprocessing
+  ${toggle_vp9_postproc}          vp9 specific postprocessing
+  ${toggle_multithread}           multithreaded encoding and decoding
+  ${toggle_spatial_resampling}    spatial sampling (scaling) support
+  ${toggle_realtime_only}         enable this option while building for real-time encoding
+  ${toggle_onthefly_bitpacking}   enable on-the-fly bitpacking in real-time encoding
+  ${toggle_error_concealment}     enable this option to get a decoder which is able to conceal losses
+  ${toggle_coefficient_range_checking}
+                                  enable decoder to check if intermediate
+                                  transform coefficients are in valid range
+  ${toggle_runtime_cpu_detect}    runtime cpu detection
+  ${toggle_shared}                shared library support
+  ${toggle_static}                static library support
+  ${toggle_small}                 favor smaller size over speed
+  ${toggle_postproc_visualizer}   macro block / block level visualizers
+  ${toggle_multi_res_encoding}    enable multiple-resolution encoding
+  ${toggle_temporal_denoising}    enable temporal denoising and disable the spatial denoiser
+  ${toggle_vp9_temporal_denoising}
+                                  enable vp9 temporal denoising
+  ${toggle_webm_io}               enable input from and output to WebM container
+  ${toggle_libyuv}                enable libyuv
+
+Codecs:
+  Codecs can be selectively enabled or disabled individually, or by family:
+      --disable-<codec>
+  is equivalent to:
+      --disable-<codec>-encoder
+      --disable-<codec>-decoder
+
+  Codecs available in this distribution:
+EOF
+#restore editor state '
+
+    family="";
+    last_family="";
+    c="";
+    str="";
+    for c in ${CODECS}; do
+        family=${c%_*}
+        if [ "${family}" != "${last_family}" ]; then
+            [ -z "${str}" ] || echo "${str}"
+            str="$(printf '    %10s:' ${family})"
+        fi
+        str="${str} $(printf '%10s' ${c#*_})"
+        last_family=${family}
+    done
+    echo "${str}"
+    show_help_post
+}
+
+##
+## BEGIN APPLICATION SPECIFIC CONFIGURATION
+##
+
+# all_platforms is a list of all supported target platforms. Maintain
+# alphabetically by architecture, generic-gnu last.
+all_platforms="${all_platforms} arm64-darwin-gcc"
+all_platforms="${all_platforms} arm64-linux-gcc"
+all_platforms="${all_platforms} armv6-linux-rvct"
+all_platforms="${all_platforms} armv6-linux-gcc"
+all_platforms="${all_platforms} armv6-none-rvct"
+all_platforms="${all_platforms} armv7-android-gcc"   #neon Cortex-A8
+all_platforms="${all_platforms} armv7-darwin-gcc"    #neon Cortex-A8
+all_platforms="${all_platforms} armv7-linux-rvct"    #neon Cortex-A8
+all_platforms="${all_platforms} armv7-linux-gcc"     #neon Cortex-A8
+all_platforms="${all_platforms} armv7-none-rvct"     #neon Cortex-A8
+all_platforms="${all_platforms} armv7-win32-vs11"
+all_platforms="${all_platforms} armv7-win32-vs12"
+all_platforms="${all_platforms} armv7-win32-vs14"
+all_platforms="${all_platforms} armv7s-darwin-gcc"
+all_platforms="${all_platforms} armv8-linux-gcc"
+all_platforms="${all_platforms} mips32-linux-gcc"
+all_platforms="${all_platforms} mips64-linux-gcc"
+all_platforms="${all_platforms} sparc-solaris-gcc"
+all_platforms="${all_platforms} x86-android-gcc"
+all_platforms="${all_platforms} x86-darwin8-gcc"
+all_platforms="${all_platforms} x86-darwin8-icc"
+all_platforms="${all_platforms} x86-darwin9-gcc"
+all_platforms="${all_platforms} x86-darwin9-icc"
+all_platforms="${all_platforms} x86-darwin10-gcc"
+all_platforms="${all_platforms} x86-darwin11-gcc"
+all_platforms="${all_platforms} x86-darwin12-gcc"
+all_platforms="${all_platforms} x86-darwin13-gcc"
+all_platforms="${all_platforms} x86-darwin14-gcc"
+all_platforms="${all_platforms} x86-darwin15-gcc"
+all_platforms="${all_platforms} x86-iphonesimulator-gcc"
+all_platforms="${all_platforms} x86-linux-gcc"
+all_platforms="${all_platforms} x86-linux-icc"
+all_platforms="${all_platforms} x86-os2-gcc"
+all_platforms="${all_platforms} x86-solaris-gcc"
+all_platforms="${all_platforms} x86-win32-gcc"
+all_platforms="${all_platforms} x86-win32-vs7"
+all_platforms="${all_platforms} x86-win32-vs8"
+all_platforms="${all_platforms} x86-win32-vs9"
+all_platforms="${all_platforms} x86-win32-vs10"
+all_platforms="${all_platforms} x86-win32-vs11"
+all_platforms="${all_platforms} x86-win32-vs12"
+all_platforms="${all_platforms} x86-win32-vs14"
+all_platforms="${all_platforms} x86_64-android-gcc"
+all_platforms="${all_platforms} x86_64-darwin9-gcc"
+all_platforms="${all_platforms} x86_64-darwin10-gcc"
+all_platforms="${all_platforms} x86_64-darwin11-gcc"
+all_platforms="${all_platforms} x86_64-darwin12-gcc"
+all_platforms="${all_platforms} x86_64-darwin13-gcc"
+all_platforms="${all_platforms} x86_64-darwin14-gcc"
+all_platforms="${all_platforms} x86_64-darwin15-gcc"
+all_platforms="${all_platforms} x86_64-iphonesimulator-gcc"
+all_platforms="${all_platforms} x86_64-linux-gcc"
+all_platforms="${all_platforms} x86_64-linux-icc"
+all_platforms="${all_platforms} x86_64-solaris-gcc"
+all_platforms="${all_platforms} x86_64-win64-gcc"
+all_platforms="${all_platforms} x86_64-win64-vs8"
+all_platforms="${all_platforms} x86_64-win64-vs9"
+all_platforms="${all_platforms} x86_64-win64-vs10"
+all_platforms="${all_platforms} x86_64-win64-vs11"
+all_platforms="${all_platforms} x86_64-win64-vs12"
+all_platforms="${all_platforms} x86_64-win64-vs14"
+all_platforms="${all_platforms} generic-gnu"
+
+# all_targets is a list of all targets that can be configured
+# note that these should be in dependency order for now.
+all_targets="libs examples docs"
+
+# all targets available are enabled, by default.
+for t in ${all_targets}; do
+    [ -f "${source_path}/${t}.mk" ] && enable_feature ${t}
+done
+
+if ! perl --version >/dev/null; then
+    die "Perl is required to build"
+fi
+
+
+if [ "`cd \"${source_path}\" && pwd`" != "`pwd`" ]; then
+  # test to see if source_path already configured
+  if [ -f "${source_path}/vpx_config.h" ]; then
+    die "source directory already configured; run 'make distclean' there first"
+  fi
+fi
+
+# check installed doxygen version
+doxy_version=$(doxygen --version 2>/dev/null)
+doxy_major=${doxy_version%%.*}
+if [ ${doxy_major:-0} -ge 1 ]; then
+    doxy_version=${doxy_version#*.}
+    doxy_minor=${doxy_version%%.*}
+    doxy_patch=${doxy_version##*.}
+
+    [ $doxy_major -gt 1 ] && enable_feature doxygen
+    [ $doxy_minor -gt 5 ] && enable_feature doxygen
+    [ $doxy_minor -eq 5 ] && [ $doxy_patch -ge 3 ] && enable_feature doxygen
+fi
+
+# disable codecs when their source directory does not exist
+[ -d "${source_path}/vp8" ] || disable_codec vp8
+[ -d "${source_path}/vp9" ] || disable_codec vp9
+
+# install everything except the sources, by default. sources will have
+# to be enabled when doing dist builds, since that's no longer a common
+# case.
+enabled doxygen && enable_feature install_docs
+enable_feature install_bins
+enable_feature install_libs
+
+enable_feature static
+enable_feature optimizations
+enable_feature dependency_tracking
+enable_feature spatial_resampling
+enable_feature multithread
+enable_feature os_support
+enable_feature temporal_denoising
+
+CODECS="
+    vp8_encoder
+    vp8_decoder
+    vp9_encoder
+    vp9_decoder
+"
+CODEC_FAMILIES="
+    vp8
+    vp9
+"
+
+ARCH_LIST="
+    arm
+    mips
+    x86
+    x86_64
+"
+ARCH_EXT_LIST_X86="
+    mmx
+    sse
+    sse2
+    sse3
+    ssse3
+    sse4_1
+    avx
+    avx2
+"
+ARCH_EXT_LIST="
+    edsp
+    media
+    neon
+    neon_asm
+
+    mips32
+    dspr2
+    msa
+    mips64
+
+    ${ARCH_EXT_LIST_X86}
+"
+HAVE_LIST="
+    ${ARCH_EXT_LIST}
+    vpx_ports
+    pthread_h
+    unistd_h
+"
+EXPERIMENT_LIST="
+    spatial_svc
+    fp_mb_stats
+    emulate_hardware
+    misc_fixes
+"
+CONFIG_LIST="
+    dependency_tracking
+    external_build
+    install_docs
+    install_bins
+    install_libs
+    install_srcs
+    use_x86inc
+    debug
+    gprof
+    gcov
+    rvct
+    gcc
+    msvs
+    pic
+    big_endian
+
+    codec_srcs
+    debug_libs
+
+    dequant_tokens
+    dc_recon
+    runtime_cpu_detect
+    postproc
+    vp9_postproc
+    multithread
+    internal_stats
+    ${CODECS}
+    ${CODEC_FAMILIES}
+    encoders
+    decoders
+    static_msvcrt
+    spatial_resampling
+    realtime_only
+    onthefly_bitpacking
+    error_concealment
+    shared
+    static
+    small
+    postproc_visualizer
+    os_support
+    unit_tests
+    webm_io
+    libyuv
+    decode_perf_tests
+    encode_perf_tests
+    multi_res_encoding
+    temporal_denoising
+    vp9_temporal_denoising
+    coefficient_range_checking
+    vp9_highbitdepth
+    better_hw_compatibility
+    experimental
+    size_limit
+    ${EXPERIMENT_LIST}
+"
+CMDLINE_SELECT="
+    dependency_tracking
+    external_build
+    extra_warnings
+    werror
+    install_docs
+    install_bins
+    install_libs
+    install_srcs
+    debug
+    gprof
+    gcov
+    pic
+    use_x86inc
+    optimizations
+    ccache
+    runtime_cpu_detect
+    thumb
+
+    libs
+    examples
+    docs
+    libc
+    as
+    size_limit
+    codec_srcs
+    debug_libs
+
+    dequant_tokens
+    dc_recon
+    postproc
+    vp9_postproc
+    multithread
+    internal_stats
+    ${CODECS}
+    ${CODEC_FAMILIES}
+    static_msvcrt
+    spatial_resampling
+    realtime_only
+    onthefly_bitpacking
+    error_concealment
+    shared
+    static
+    small
+    postproc_visualizer
+    unit_tests
+    webm_io
+    libyuv
+    decode_perf_tests
+    encode_perf_tests
+    multi_res_encoding
+    temporal_denoising
+    vp9_temporal_denoising
+    coefficient_range_checking
+    better_hw_compatibility
+    vp9_highbitdepth
+    experimental
+"
+
+process_cmdline() {
+    for opt do
+        optval="${opt#*=}"
+        case "$opt" in
+        --disable-codecs)
+          for c in ${CODEC_FAMILIES}; do disable_codec $c; done
+          ;;
+        --enable-?*|--disable-?*)
+        eval `echo "$opt" | sed 's/--/action=/;s/-/ option=/;s/-/_/g'`
+        if is_in ${option} ${EXPERIMENT_LIST}; then
+            if enabled experimental; then
+                ${action}_feature $option
+            else
+                log_echo "Ignoring $opt -- not in experimental mode."
+            fi
+        elif is_in ${option} "${CODECS} ${CODEC_FAMILIES}"; then
+            ${action}_codec ${option}
+        else
+            process_common_cmdline $opt
+        fi
+        ;;
+        *) process_common_cmdline "$opt"
+        ;;
+        esac
+    done
+}
+
+post_process_cmdline() {
+    c=""
+
+    # Enable all detected codecs, if they haven't been disabled
+    for c in ${CODECS}; do soft_enable $c; done
+
+    # Enable the codec family if any component of that family is enabled
+    for c in ${CODECS}; do
+        enabled $c && enable_feature ${c%_*}
+    done
+
+    # Set the {en,de}coders variable if any algorithm in that class is enabled
+    for c in ${CODECS}; do
+        enabled ${c} && enable_feature ${c##*_}s
+    done
+}
+
+
+process_targets() {
+    enabled child || write_common_config_banner
+    write_common_target_config_h ${BUILD_PFX}vpx_config.h
+    write_common_config_targets
+
+    # Calculate the default distribution name, based on the enabled features
+    cf=""
+    DIST_DIR=vpx
+    for cf in $CODEC_FAMILIES; do
+        if enabled ${cf}_encoder && enabled ${cf}_decoder; then
+            DIST_DIR="${DIST_DIR}-${cf}"
+        elif enabled ${cf}_encoder; then
+            DIST_DIR="${DIST_DIR}-${cf}cx"
+        elif enabled ${cf}_decoder; then
+            DIST_DIR="${DIST_DIR}-${cf}dx"
+        fi
+    done
+    enabled debug_libs && DIST_DIR="${DIST_DIR}-debug"
+    enabled codec_srcs && DIST_DIR="${DIST_DIR}-src"
+    ! enabled postproc && ! enabled vp9_postproc && DIST_DIR="${DIST_DIR}-nopost"
+    ! enabled multithread && DIST_DIR="${DIST_DIR}-nomt"
+    ! enabled install_docs && DIST_DIR="${DIST_DIR}-nodocs"
+    DIST_DIR="${DIST_DIR}-${tgt_isa}-${tgt_os}"
+    case "${tgt_os}" in
+    win*) enabled static_msvcrt && DIST_DIR="${DIST_DIR}mt" || DIST_DIR="${DIST_DIR}md"
+          DIST_DIR="${DIST_DIR}-${tgt_cc}"
+          ;;
+    esac
+    if [ -f "${source_path}/build/make/version.sh" ]; then
+        ver=`"$source_path/build/make/version.sh" --bare "$source_path"`
+        DIST_DIR="${DIST_DIR}-${ver}"
+        VERSION_STRING=${ver}
+        ver=${ver%%-*}
+        VERSION_PATCH=${ver##*.}
+        ver=${ver%.*}
+        VERSION_MINOR=${ver##*.}
+        ver=${ver#v}
+        VERSION_MAJOR=${ver%.*}
+    fi
+    enabled child || cat <<EOF >> config.mk
+
+PREFIX=${prefix}
+ifeq (\$(MAKECMDGOALS),dist)
+DIST_DIR?=${DIST_DIR}
+else
+DIST_DIR?=\$(DESTDIR)${prefix}
+endif
+LIBSUBDIR=${libdir##${prefix}/}
+
+VERSION_STRING=${VERSION_STRING}
+
+VERSION_MAJOR=${VERSION_MAJOR}
+VERSION_MINOR=${VERSION_MINOR}
+VERSION_PATCH=${VERSION_PATCH}
+
+CONFIGURE_ARGS=${CONFIGURE_ARGS}
+EOF
+    enabled child || echo "CONFIGURE_ARGS?=${CONFIGURE_ARGS}" >> config.mk
+
+    #
+    # Write makefiles for all enabled targets
+    #
+    for tgt in libs examples docs solution; do
+        tgt_fn="$tgt-$toolchain.mk"
+
+        if enabled $tgt; then
+            echo "Creating makefiles for ${toolchain} ${tgt}"
+            write_common_target_config_mk $tgt_fn ${BUILD_PFX}vpx_config.h
+            #write_${tgt}_config
+        fi
+    done
+
+}
+
+process_detect() {
+    if enabled shared; then
+        # Can only build shared libs on a subset of platforms. Doing this check
+        # here rather than at option parse time because the target auto-detect
+        # magic happens after the command line has been parsed.
+        case "${tgt_os}" in
+        linux|os2|darwin*|iphonesimulator*)
+            # Supported platforms
+            ;;
+        *)
+            if enabled gnu; then
+                echo "--enable-shared is only supported on ELF; assuming this is OK"
+            else
+                die "--enable-shared only supported on ELF, OS/2, and Darwin for now"
+            fi
+            ;;
+        esac
+    fi
+    if [ -z "$CC" ] || enabled external_build; then
+        echo "Bypassing toolchain for environment detection."
+        enable_feature external_build
+        check_header() {
+            log fake_check_header "$@"
+            header=$1
+            shift
+            var=`echo $header | sed 's/[^A-Za-z0-9_]/_/g'`
+            disable_feature $var
+            # Headers common to all environments
+            case $header in
+                stdio.h)
+                    true;
+                ;;
+                *)
+                    result=false
+                    for d in "$@"; do
+                        [ -f "${d##-I}/$header" ] && result=true && break
+                    done
+                    ${result:-true}
+            esac && enable_feature $var
+
+            # Specialize windows and POSIX environments.
+            case $toolchain in
+                *-win*-*)
+                    # Don't check for any headers in Windows builds.
+                    false
+                ;;
+                *)
+                    case $header in
+                        pthread.h) true;;
+                        unistd.h) true;;
+                        *) false;;
+                    esac && enable_feature $var
+            esac
+            enabled $var
+        }
+        check_ld() {
+            true
+        }
+    fi
+    check_header stdio.h || die "Unable to invoke compiler: ${CC} ${CFLAGS}"
+    check_ld <<EOF || die "Toolchain is unable to link executables"
+int main(void) {return 0;}
+EOF
+    # check system headers
+    check_header pthread.h
+    check_header unistd.h # for sysconf(3) and friends.
+
+    check_header vpx/vpx_integer.h -I${source_path} && enable_feature vpx_ports
+}
+
+process_toolchain() {
+    process_common_toolchain
+
+    # Enable some useful compiler flags
+    if enabled gcc; then
+        enabled werror && check_add_cflags -Werror
+        check_add_cflags -Wall
+        check_add_cflags -Wdeclaration-after-statement
+        check_add_cflags -Wdisabled-optimization
+        check_add_cflags -Wpointer-arith
+        check_add_cflags -Wtype-limits
+        check_add_cflags -Wcast-qual
+        check_add_cflags -Wvla
+        check_add_cflags -Wimplicit-function-declaration
+        check_add_cflags -Wuninitialized
+        check_add_cflags -Wunused-variable
+        case ${CC} in
+          *clang*)
+              # libvpx and/or clang have issues with aliasing:
+              # https://code.google.com/p/webm/issues/detail?id=603
+              # work around them until they are fixed
+              check_add_cflags -fno-strict-aliasing
+          ;;
+          *) check_add_cflags -Wunused-but-set-variable ;;
+        esac
+        if enabled mips || [ -z "${INLINE}" ]; then
+          enabled extra_warnings || check_add_cflags -Wno-unused-function
+        else
+          check_add_cflags -Wunused-function
+        fi
+    fi
+
+    if enabled icc; then
+        enabled werror && check_add_cflags -Werror
+        check_add_cflags -Wall
+        check_add_cflags -Wpointer-arith
+
+        # ICC has a number of floating point optimizations that we disable
+        # in favor of deterministic output WRT to other compilers
+        add_cflags -fp-model precise
+    fi
+
+    # Enable extra, harmless warnings. These might provide additional insight
+    # to what the compiler is doing and why, but in general, but they shouldn't
+    # be treated as fatal, even if we're treating warnings as errors.
+    GCC_EXTRA_WARNINGS="
+        -Wdisabled-optimization
+        -Winline
+    "
+    enabled gcc && EXTRA_WARNINGS="${GCC_EXTRA_WARNINGS}"
+    RVCT_EXTRA_WARNINGS="
+        --remarks
+    "
+    enabled rvct && EXTRA_WARNINGS="${RVCT_EXTRA_WARNINGS}"
+    if enabled extra_warnings; then
+        for w in ${EXTRA_WARNINGS}; do
+            check_add_cflags ${w}
+            enabled gcc && enabled werror && check_add_cflags -Wno-error=${w}
+        done
+    fi
+
+    # ccache only really works on gcc toolchains
+    enabled gcc || soft_disable ccache
+    if enabled mips; then
+        enable_feature dequant_tokens
+        enable_feature dc_recon
+    fi
+
+    if enabled internal_stats; then
+        enable_feature vp9_postproc
+    fi
+
+    # Enable the postbuild target if building for visual studio.
+    case "$tgt_cc" in
+        vs*) enable_feature msvs
+             enable_feature solution
+             vs_version=${tgt_cc##vs}
+             case $vs_version in
+             [789])
+                 VCPROJ_SFX=vcproj
+                 gen_vcproj_cmd=${source_path}/build/make/gen_msvs_proj.sh
+                 ;;
+             10|11|12|14)
+                 VCPROJ_SFX=vcxproj
+                 gen_vcproj_cmd=${source_path}/build/make/gen_msvs_vcxproj.sh
+                 enabled werror && gen_vcproj_cmd="${gen_vcproj_cmd} --enable-werror"
+                 ;;
+             esac
+             all_targets="${all_targets} solution"
+             INLINE="__forceinline"
+        ;;
+    esac
+
+    # Other toolchain specific defaults
+    case $toolchain in x86*) soft_enable postproc;; esac
+
+    if enabled postproc_visualizer; then
+        enabled postproc || die "postproc_visualizer requires postproc to be enabled"
+    fi
+
+    # Enable unit tests by default if we have a working C++ compiler.
+    case "$toolchain" in
+        *-vs*)
+            soft_enable unit_tests
+            soft_enable webm_io
+            soft_enable libyuv
+        ;;
+        *-android-*)
+            soft_enable webm_io
+            soft_enable libyuv
+            # GTestLog must be modified to use Android logging utilities.
+        ;;
+        *-darwin-*)
+            # iOS/ARM builds do not work with gtest. This does not match
+            # x86 targets.
+        ;;
+        *-iphonesimulator-*)
+            soft_enable webm_io
+            soft_enable libyuv
+        ;;
+        *-win*)
+            # Some mingw toolchains don't have pthread available by default.
+            # Treat these more like visual studio where threading in gtest
+            # would be disabled for the same reason.
+            check_cxx "$@" <<EOF && soft_enable unit_tests
+int z;
+EOF
+            check_cxx "$@" <<EOF && soft_enable webm_io
+int z;
+EOF
+            check_cxx "$@" <<EOF && soft_enable libyuv
+int z;
+EOF
+        ;;
+        *)
+            enabled pthread_h && check_cxx "$@" <<EOF && soft_enable unit_tests
+int z;
+EOF
+            check_cxx "$@" <<EOF && soft_enable webm_io
+int z;
+EOF
+            check_cxx "$@" <<EOF && soft_enable libyuv
+int z;
+EOF
+        ;;
+    esac
+    # libwebm needs to be linked with C++ standard library
+    enabled webm_io && LD=${CXX}
+
+    # append any user defined extra cflags
+    if [ -n "${extra_cflags}" ] ; then
+        check_add_cflags ${extra_cflags} || \
+        die "Requested extra CFLAGS '${extra_cflags}' not supported by compiler"
+    fi
+    if [ -n "${extra_cxxflags}" ]; then
+        check_add_cxxflags ${extra_cxxflags} || \
+        die "Requested extra CXXFLAGS '${extra_cxxflags}' not supported by compiler"
+    fi
+}
+
+
+##
+## END APPLICATION SPECIFIC CONFIGURATION
+##
+CONFIGURE_ARGS="$@"
+process "$@"
+print_webm_license ${BUILD_PFX}vpx_config.c "/*" " */"
+cat <<EOF >> ${BUILD_PFX}vpx_config.c
+#include "vpx/vpx_codec.h"
+static const char* const cfg = "$CONFIGURE_ARGS";
+const char *vpx_codec_build_config(void) {return cfg;}
+EOF
diff --git a/libvpx/libvpx/docs.mk b/libvpx/libvpx/docs.mk
new file mode 100644
index 0000000..889d182
--- /dev/null
+++ b/libvpx/libvpx/docs.mk
@@ -0,0 +1,48 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+INSTALL_MAPS += docs/%    docs/%
+INSTALL_MAPS += src/%     %
+INSTALL_MAPS += %         %
+
+# Static documentation authored in doxygen
+CODEC_DOX :=    mainpage.dox \
+		keywords.dox \
+		usage.dox \
+		usage_cx.dox \
+		usage_dx.dox \
+
+# Other doxy files sourced in Markdown
+TXT_DOX = $(call enabled,TXT_DOX)
+
+EXAMPLE_PATH += $(SRC_PATH_BARE) #for CHANGELOG, README, etc
+EXAMPLE_PATH += $(SRC_PATH_BARE)/examples
+
+doxyfile: $(if $(findstring examples, $(ALL_TARGETS)),examples.doxy)
+doxyfile: libs.doxy_template libs.doxy
+	@echo "    [CREATE] $@"
+	@cat $^ > $@
+	@echo "STRIP_FROM_PATH += $(SRC_PATH_BARE) $(BUILD_ROOT)" >> $@
+	@echo "INPUT += $(addprefix $(SRC_PATH_BARE)/,$(CODEC_DOX))" >> $@;
+	@echo "INPUT += $(TXT_DOX)" >> $@;
+	@echo "EXAMPLE_PATH += $(EXAMPLE_PATH)" >> $@
+
+CLEAN-OBJS += doxyfile $(wildcard docs/html/*)
+docs/html/index.html: doxyfile $(CODEC_DOX) $(TXT_DOX)
+	@echo "    [DOXYGEN] $<"
+	@doxygen $<
+DOCS-yes += docs/html/index.html
+
+DIST-DOCS-yes = $(wildcard docs/html/*)
+DIST-DOCS-$(CONFIG_CODEC_SRCS) += $(addprefix src/,$(CODEC_DOX))
+DIST-DOCS-$(CONFIG_CODEC_SRCS) += src/libs.doxy_template
+DIST-DOCS-yes                  += CHANGELOG
+DIST-DOCS-yes                  += README
diff --git a/libvpx/libvpx/examples.mk b/libvpx/libvpx/examples.mk
new file mode 100644
index 0000000..c891a54
--- /dev/null
+++ b/libvpx/libvpx/examples.mk
@@ -0,0 +1,393 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+LIBYUV_SRCS +=  third_party/libyuv/include/libyuv/basic_types.h  \
+                third_party/libyuv/include/libyuv/convert.h \
+                third_party/libyuv/include/libyuv/convert_argb.h \
+                third_party/libyuv/include/libyuv/convert_from.h \
+                third_party/libyuv/include/libyuv/cpu_id.h  \
+                third_party/libyuv/include/libyuv/planar_functions.h  \
+                third_party/libyuv/include/libyuv/rotate.h  \
+                third_party/libyuv/include/libyuv/row.h  \
+                third_party/libyuv/include/libyuv/scale.h  \
+                third_party/libyuv/include/libyuv/scale_row.h  \
+                third_party/libyuv/source/cpu_id.cc \
+                third_party/libyuv/source/planar_functions.cc \
+                third_party/libyuv/source/row_any.cc \
+                third_party/libyuv/source/row_common.cc \
+                third_party/libyuv/source/row_gcc.cc \
+                third_party/libyuv/source/row_mips.cc \
+                third_party/libyuv/source/row_neon.cc \
+                third_party/libyuv/source/row_neon64.cc \
+                third_party/libyuv/source/row_win.cc \
+                third_party/libyuv/source/scale.cc \
+                third_party/libyuv/source/scale_any.cc \
+                third_party/libyuv/source/scale_common.cc \
+                third_party/libyuv/source/scale_gcc.cc \
+                third_party/libyuv/source/scale_mips.cc \
+                third_party/libyuv/source/scale_neon.cc \
+                third_party/libyuv/source/scale_neon64.cc \
+                third_party/libyuv/source/scale_win.cc \
+
+LIBWEBM_COMMON_SRCS += third_party/libwebm/common/hdr_util.cc \
+                       third_party/libwebm/common/hdr_util.h \
+                       third_party/libwebm/common/webmids.h
+
+LIBWEBM_MUXER_SRCS += third_party/libwebm/mkvmuxer/mkvmuxer.cc \
+                      third_party/libwebm/mkvmuxer/mkvmuxerutil.cc \
+                      third_party/libwebm/mkvmuxer/mkvwriter.cc \
+                      third_party/libwebm/mkvmuxer/mkvmuxer.h \
+                      third_party/libwebm/mkvmuxer/mkvmuxertypes.h \
+                      third_party/libwebm/mkvmuxer/mkvmuxerutil.h \
+                      third_party/libwebm/mkvparser/mkvparser.h \
+                      third_party/libwebm/mkvmuxer/mkvwriter.h
+
+LIBWEBM_PARSER_SRCS = third_party/libwebm/mkvparser/mkvparser.cc \
+                      third_party/libwebm/mkvparser/mkvreader.cc \
+                      third_party/libwebm/mkvparser/mkvparser.h \
+                      third_party/libwebm/mkvparser/mkvreader.h
+
+# Add compile flags and include path for libwebm sources.
+ifeq ($(CONFIG_WEBM_IO),yes)
+  CXXFLAGS     += -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS
+  INC_PATH-yes += $(SRC_PATH_BARE)/third_party/libwebm
+endif
+
+
+# List of examples to build. UTILS are tools meant for distribution
+# while EXAMPLES demonstrate specific portions of the API.
+UTILS-$(CONFIG_DECODERS)    += vpxdec.c
+vpxdec.SRCS                 += md5_utils.c md5_utils.h
+vpxdec.SRCS                 += vpx_ports/mem_ops.h
+vpxdec.SRCS                 += vpx_ports/mem_ops_aligned.h
+vpxdec.SRCS                 += vpx_ports/msvc.h
+vpxdec.SRCS                 += vpx_ports/vpx_timer.h
+vpxdec.SRCS                 += vpx/vpx_integer.h
+vpxdec.SRCS                 += args.c args.h
+vpxdec.SRCS                 += ivfdec.c ivfdec.h
+vpxdec.SRCS                 += tools_common.c tools_common.h
+vpxdec.SRCS                 += y4menc.c y4menc.h
+ifeq ($(CONFIG_LIBYUV),yes)
+  vpxdec.SRCS                 += $(LIBYUV_SRCS)
+endif
+ifeq ($(CONFIG_WEBM_IO),yes)
+  vpxdec.SRCS                 += $(LIBWEBM_COMMON_SRCS)
+  vpxdec.SRCS                 += $(LIBWEBM_MUXER_SRCS)
+  vpxdec.SRCS                 += $(LIBWEBM_PARSER_SRCS)
+  vpxdec.SRCS                 += webmdec.cc webmdec.h
+endif
+vpxdec.GUID                  = BA5FE66F-38DD-E034-F542-B1578C5FB950
+vpxdec.DESCRIPTION           = Full featured decoder
+UTILS-$(CONFIG_ENCODERS)    += vpxenc.c
+vpxenc.SRCS                 += args.c args.h y4minput.c y4minput.h vpxenc.h
+vpxenc.SRCS                 += ivfdec.c ivfdec.h
+vpxenc.SRCS                 += ivfenc.c ivfenc.h
+vpxenc.SRCS                 += rate_hist.c rate_hist.h
+vpxenc.SRCS                 += tools_common.c tools_common.h
+vpxenc.SRCS                 += warnings.c warnings.h
+vpxenc.SRCS                 += vpx_ports/mem_ops.h
+vpxenc.SRCS                 += vpx_ports/mem_ops_aligned.h
+vpxenc.SRCS                 += vpx_ports/msvc.h
+vpxenc.SRCS                 += vpx_ports/vpx_timer.h
+vpxenc.SRCS                 += vpxstats.c vpxstats.h
+ifeq ($(CONFIG_LIBYUV),yes)
+  vpxenc.SRCS                 += $(LIBYUV_SRCS)
+endif
+ifeq ($(CONFIG_WEBM_IO),yes)
+  vpxenc.SRCS                 += $(LIBWEBM_COMMON_SRCS)
+  vpxenc.SRCS                 += $(LIBWEBM_MUXER_SRCS)
+  vpxenc.SRCS                 += $(LIBWEBM_PARSER_SRCS)
+  vpxenc.SRCS                 += webmenc.cc webmenc.h
+endif
+vpxenc.GUID                  = 548DEC74-7A15-4B2B-AFC3-AA102E7C25C1
+vpxenc.DESCRIPTION           = Full featured encoder
+ifeq ($(CONFIG_SPATIAL_SVC),yes)
+  EXAMPLES-$(CONFIG_VP9_ENCODER)      += vp9_spatial_svc_encoder.c
+  vp9_spatial_svc_encoder.SRCS        += args.c args.h
+  vp9_spatial_svc_encoder.SRCS        += ivfenc.c ivfenc.h
+  vp9_spatial_svc_encoder.SRCS        += tools_common.c tools_common.h
+  vp9_spatial_svc_encoder.SRCS        += video_common.h
+  vp9_spatial_svc_encoder.SRCS        += video_writer.h video_writer.c
+  vp9_spatial_svc_encoder.SRCS        += vpx_ports/msvc.h
+  vp9_spatial_svc_encoder.SRCS        += vpxstats.c vpxstats.h
+  vp9_spatial_svc_encoder.GUID        = 4A38598D-627D-4505-9C7B-D4020C84100D
+  vp9_spatial_svc_encoder.DESCRIPTION = VP9 Spatial SVC Encoder
+endif
+
+ifneq ($(CONFIG_SHARED),yes)
+EXAMPLES-$(CONFIG_VP9_ENCODER)    += resize_util.c
+endif
+
+EXAMPLES-$(CONFIG_ENCODERS)          += vpx_temporal_svc_encoder.c
+vpx_temporal_svc_encoder.SRCS        += ivfenc.c ivfenc.h
+vpx_temporal_svc_encoder.SRCS        += tools_common.c tools_common.h
+vpx_temporal_svc_encoder.SRCS        += video_common.h
+vpx_temporal_svc_encoder.SRCS        += video_writer.h video_writer.c
+vpx_temporal_svc_encoder.SRCS        += vpx_ports/msvc.h
+vpx_temporal_svc_encoder.GUID        = B18C08F2-A439-4502-A78E-849BE3D60947
+vpx_temporal_svc_encoder.DESCRIPTION = Temporal SVC Encoder
+EXAMPLES-$(CONFIG_DECODERS)        += simple_decoder.c
+simple_decoder.GUID                 = D3BBF1E9-2427-450D-BBFF-B2843C1D44CC
+simple_decoder.SRCS                += ivfdec.h ivfdec.c
+simple_decoder.SRCS                += tools_common.h tools_common.c
+simple_decoder.SRCS                += video_common.h
+simple_decoder.SRCS                += video_reader.h video_reader.c
+simple_decoder.SRCS                += vpx_ports/mem_ops.h
+simple_decoder.SRCS                += vpx_ports/mem_ops_aligned.h
+simple_decoder.SRCS                += vpx_ports/msvc.h
+simple_decoder.DESCRIPTION          = Simplified decoder loop
+EXAMPLES-$(CONFIG_DECODERS)        += postproc.c
+postproc.SRCS                      += ivfdec.h ivfdec.c
+postproc.SRCS                      += tools_common.h tools_common.c
+postproc.SRCS                      += video_common.h
+postproc.SRCS                      += video_reader.h video_reader.c
+postproc.SRCS                      += vpx_ports/mem_ops.h
+postproc.SRCS                      += vpx_ports/mem_ops_aligned.h
+postproc.SRCS                      += vpx_ports/msvc.h
+postproc.GUID                       = 65E33355-F35E-4088-884D-3FD4905881D7
+postproc.DESCRIPTION                = Decoder postprocessor control
+EXAMPLES-$(CONFIG_DECODERS)        += decode_to_md5.c
+decode_to_md5.SRCS                 += md5_utils.h md5_utils.c
+decode_to_md5.SRCS                 += ivfdec.h ivfdec.c
+decode_to_md5.SRCS                 += tools_common.h tools_common.c
+decode_to_md5.SRCS                 += video_common.h
+decode_to_md5.SRCS                 += video_reader.h video_reader.c
+decode_to_md5.SRCS                 += vpx_ports/mem_ops.h
+decode_to_md5.SRCS                 += vpx_ports/mem_ops_aligned.h
+decode_to_md5.SRCS                 += vpx_ports/msvc.h
+decode_to_md5.GUID                  = 59120B9B-2735-4BFE-B022-146CA340FE42
+decode_to_md5.DESCRIPTION           = Frame by frame MD5 checksum
+EXAMPLES-$(CONFIG_ENCODERS)     += simple_encoder.c
+simple_encoder.SRCS             += ivfenc.h ivfenc.c
+simple_encoder.SRCS             += tools_common.h tools_common.c
+simple_encoder.SRCS             += video_common.h
+simple_encoder.SRCS             += video_writer.h video_writer.c
+simple_encoder.SRCS             += vpx_ports/msvc.h
+simple_encoder.GUID              = 4607D299-8A71-4D2C-9B1D-071899B6FBFD
+simple_encoder.DESCRIPTION       = Simplified encoder loop
+EXAMPLES-$(CONFIG_VP9_ENCODER)  += vp9_lossless_encoder.c
+vp9_lossless_encoder.SRCS       += ivfenc.h ivfenc.c
+vp9_lossless_encoder.SRCS       += tools_common.h tools_common.c
+vp9_lossless_encoder.SRCS       += video_common.h
+vp9_lossless_encoder.SRCS       += video_writer.h video_writer.c
+vp9_lossless_encoder.SRCS       += vpx_ports/msvc.h
+vp9_lossless_encoder.GUID        = B63C7C88-5348-46DC-A5A6-CC151EF93366
+vp9_lossless_encoder.DESCRIPTION = Simplified lossless VP9 encoder
+EXAMPLES-$(CONFIG_ENCODERS)     += twopass_encoder.c
+twopass_encoder.SRCS            += ivfenc.h ivfenc.c
+twopass_encoder.SRCS            += tools_common.h tools_common.c
+twopass_encoder.SRCS            += video_common.h
+twopass_encoder.SRCS            += video_writer.h video_writer.c
+twopass_encoder.SRCS            += vpx_ports/msvc.h
+twopass_encoder.GUID             = 73494FA6-4AF9-4763-8FBB-265C92402FD8
+twopass_encoder.DESCRIPTION      = Two-pass encoder loop
+EXAMPLES-$(CONFIG_DECODERS)     += decode_with_drops.c
+decode_with_drops.SRCS          += ivfdec.h ivfdec.c
+decode_with_drops.SRCS          += tools_common.h tools_common.c
+decode_with_drops.SRCS          += video_common.h
+decode_with_drops.SRCS          += video_reader.h video_reader.c
+decode_with_drops.SRCS          += vpx_ports/mem_ops.h
+decode_with_drops.SRCS          += vpx_ports/mem_ops_aligned.h
+decode_with_drops.SRCS          += vpx_ports/msvc.h
+decode_with_drops.GUID           = CE5C53C4-8DDA-438A-86ED-0DDD3CDB8D26
+decode_with_drops.DESCRIPTION    = Drops frames while decoding
+EXAMPLES-$(CONFIG_ENCODERS)        += set_maps.c
+set_maps.SRCS                      += ivfenc.h ivfenc.c
+set_maps.SRCS                      += tools_common.h tools_common.c
+set_maps.SRCS                      += video_common.h
+set_maps.SRCS                      += video_writer.h video_writer.c
+set_maps.SRCS                      += vpx_ports/msvc.h
+set_maps.GUID                       = ECB2D24D-98B8-4015-A465-A4AF3DCC145F
+set_maps.DESCRIPTION                = Set active and ROI maps
+EXAMPLES-$(CONFIG_VP8_ENCODER)     += vp8cx_set_ref.c
+vp8cx_set_ref.SRCS                 += ivfenc.h ivfenc.c
+vp8cx_set_ref.SRCS                 += tools_common.h tools_common.c
+vp8cx_set_ref.SRCS                 += video_common.h
+vp8cx_set_ref.SRCS                 += video_writer.h video_writer.c
+vp8cx_set_ref.SRCS                 += vpx_ports/msvc.h
+vp8cx_set_ref.GUID                  = C5E31F7F-96F6-48BD-BD3E-10EBF6E8057A
+vp8cx_set_ref.DESCRIPTION           = VP8 set encoder reference frame
+
+
+ifeq ($(CONFIG_MULTI_RES_ENCODING),yes)
+ifeq ($(CONFIG_LIBYUV),yes)
+EXAMPLES-$(CONFIG_VP8_ENCODER)          += vp8_multi_resolution_encoder.c
+vp8_multi_resolution_encoder.SRCS       += ivfenc.h ivfenc.c
+vp8_multi_resolution_encoder.SRCS       += tools_common.h tools_common.c
+vp8_multi_resolution_encoder.SRCS       += video_writer.h video_writer.c
+vp8_multi_resolution_encoder.SRCS       += vpx_ports/msvc.h
+vp8_multi_resolution_encoder.SRCS       += $(LIBYUV_SRCS)
+vp8_multi_resolution_encoder.GUID        = 04f8738e-63c8-423b-90fa-7c2703a374de
+vp8_multi_resolution_encoder.DESCRIPTION = VP8 Multiple-resolution Encoding
+endif
+endif
+
+# Handle extra library flags depending on codec configuration
+
+# We should not link to math library (libm) on RVCT
+# when building for bare-metal targets
+ifeq ($(CONFIG_OS_SUPPORT), yes)
+CODEC_EXTRA_LIBS-$(CONFIG_VP8)         += m
+CODEC_EXTRA_LIBS-$(CONFIG_VP9)         += m
+else
+    ifeq ($(CONFIG_GCC), yes)
+    CODEC_EXTRA_LIBS-$(CONFIG_VP8)         += m
+    CODEC_EXTRA_LIBS-$(CONFIG_VP9)         += m
+    endif
+endif
+#
+# End of specified files. The rest of the build rules should happen
+# automagically from here.
+#
+
+
+# Examples need different flags based on whether we're building
+# from an installed tree or a version controlled tree. Determine
+# the proper paths.
+ifeq ($(HAVE_ALT_TREE_LAYOUT),yes)
+    LIB_PATH-yes := $(SRC_PATH_BARE)/../lib
+    INC_PATH-yes := $(SRC_PATH_BARE)/../include
+else
+    LIB_PATH-yes                     += $(if $(BUILD_PFX),$(BUILD_PFX),.)
+    INC_PATH-$(CONFIG_VP8_DECODER)   += $(SRC_PATH_BARE)/vp8
+    INC_PATH-$(CONFIG_VP8_ENCODER)   += $(SRC_PATH_BARE)/vp8
+    INC_PATH-$(CONFIG_VP9_DECODER)   += $(SRC_PATH_BARE)/vp9
+    INC_PATH-$(CONFIG_VP9_ENCODER)   += $(SRC_PATH_BARE)/vp9
+endif
+INC_PATH-$(CONFIG_LIBYUV) += $(SRC_PATH_BARE)/third_party/libyuv/include
+LIB_PATH := $(call enabled,LIB_PATH)
+INC_PATH := $(call enabled,INC_PATH)
+INTERNAL_CFLAGS = $(addprefix -I,$(INC_PATH))
+INTERNAL_LDFLAGS += $(addprefix -L,$(LIB_PATH))
+
+
+# Expand list of selected examples to build (as specified above)
+UTILS           = $(call enabled,UTILS)
+EXAMPLES        = $(addprefix examples/,$(call enabled,EXAMPLES))
+ALL_EXAMPLES    = $(UTILS) $(EXAMPLES)
+UTIL_SRCS       = $(foreach ex,$(UTILS),$($(ex:.c=).SRCS))
+ALL_SRCS        = $(foreach ex,$(ALL_EXAMPLES),$($(notdir $(ex:.c=)).SRCS))
+CODEC_EXTRA_LIBS=$(sort $(call enabled,CODEC_EXTRA_LIBS))
+
+
+# Expand all example sources into a variable containing all sources
+# for that example (not just them main one specified in UTILS/EXAMPLES)
+# and add this file to the list (for MSVS workspace generation)
+$(foreach ex,$(ALL_EXAMPLES),$(eval $(notdir $(ex:.c=)).SRCS += $(ex) examples.mk))
+
+
+# Create build/install dependencies for all examples. The common case
+# is handled here. The MSVS case is handled below.
+NOT_MSVS = $(if $(CONFIG_MSVS),,yes)
+DIST-BINS-$(NOT_MSVS)      += $(addprefix bin/,$(ALL_EXAMPLES:.c=$(EXE_SFX)))
+INSTALL-BINS-$(NOT_MSVS)   += $(addprefix bin/,$(UTILS:.c=$(EXE_SFX)))
+DIST-SRCS-yes              += $(ALL_SRCS)
+INSTALL-SRCS-yes           += $(UTIL_SRCS)
+OBJS-$(NOT_MSVS)           += $(call objs,$(ALL_SRCS))
+BINS-$(NOT_MSVS)           += $(addprefix $(BUILD_PFX),$(ALL_EXAMPLES:.c=$(EXE_SFX)))
+
+
+# Instantiate linker template for all examples.
+CODEC_LIB=$(if $(CONFIG_DEBUG_LIBS),vpx_g,vpx)
+ifneq ($(filter darwin%,$(TGT_OS)),)
+SHARED_LIB_SUF=.dylib
+else
+ifneq ($(filter os2%,$(TGT_OS)),)
+SHARED_LIB_SUF=_dll.a
+else
+SHARED_LIB_SUF=.so
+endif
+endif
+CODEC_LIB_SUF=$(if $(CONFIG_SHARED),$(SHARED_LIB_SUF),.a)
+$(foreach bin,$(BINS-yes),\
+    $(eval $(bin):$(LIB_PATH)/lib$(CODEC_LIB)$(CODEC_LIB_SUF))\
+    $(eval $(call linker_template,$(bin),\
+        $(call objs,$($(notdir $(bin:$(EXE_SFX)=)).SRCS)) \
+        -l$(CODEC_LIB) $(addprefix -l,$(CODEC_EXTRA_LIBS))\
+        )))
+
+# The following pairs define a mapping of locations in the distribution
+# tree to locations in the source/build trees.
+INSTALL_MAPS += src/%.c   %.c
+INSTALL_MAPS += src/%     $(SRC_PATH_BARE)/%
+INSTALL_MAPS += bin/%     %
+INSTALL_MAPS += %         %
+
+
+# Set up additional MSVS environment
+ifeq ($(CONFIG_MSVS),yes)
+CODEC_LIB=$(if $(CONFIG_SHARED),vpx,$(if $(CONFIG_STATIC_MSVCRT),vpxmt,vpxmd))
+# This variable uses deferred expansion intentionally, since the results of
+# $(wildcard) may change during the course of the Make.
+VS_PLATFORMS = $(foreach d,$(wildcard */Release/$(CODEC_LIB).lib),$(word 1,$(subst /, ,$(d))))
+INSTALL_MAPS += $(foreach p,$(VS_PLATFORMS),bin/$(p)/%  $(p)/Release/%)
+endif
+
+# Build Visual Studio Projects. We use a template here to instantiate
+# explicit rules rather than using an implicit rule because we want to
+# leverage make's VPATH searching rather than specifying the paths on
+# each file in ALL_EXAMPLES. This has the unfortunate side effect that
+# touching the source files trigger a rebuild of the project files
+# even though there is no real dependency there (the dependency is on
+# the makefiles). We may want to revisit this.
+define vcproj_template
+$(1): $($(1:.$(VCPROJ_SFX)=).SRCS) vpx.$(VCPROJ_SFX)
+	$(if $(quiet),@echo "    [vcproj] $$@")
+	$(qexec)$$(GEN_VCPROJ)\
+            --exe\
+            --target=$$(TOOLCHAIN)\
+            --name=$$(@:.$(VCPROJ_SFX)=)\
+            --ver=$$(CONFIG_VS_VERSION)\
+            --proj-guid=$$($$(@:.$(VCPROJ_SFX)=).GUID)\
+            --src-path-bare="$(SRC_PATH_BARE)" \
+            $$(if $$(CONFIG_STATIC_MSVCRT),--static-crt) \
+            --out=$$@ $$(INTERNAL_CFLAGS) $$(CFLAGS) \
+            $$(INTERNAL_LDFLAGS) $$(LDFLAGS) -l$$(CODEC_LIB) $$^
+endef
+ALL_EXAMPLES_BASENAME := $(notdir $(ALL_EXAMPLES))
+PROJECTS-$(CONFIG_MSVS) += $(ALL_EXAMPLES_BASENAME:.c=.$(VCPROJ_SFX))
+INSTALL-BINS-$(CONFIG_MSVS) += $(foreach p,$(VS_PLATFORMS),\
+                               $(addprefix bin/$(p)/,$(ALL_EXAMPLES_BASENAME:.c=.exe)))
+$(foreach proj,$(call enabled,PROJECTS),\
+    $(eval $(call vcproj_template,$(proj))))
+
+#
+# Documentation Rules
+#
+%.dox: %.c
+	@echo "    [DOXY] $@"
+	@mkdir -p $(dir $@)
+	@echo "/*!\page example_$(@F:.dox=) $(@F:.dox=)" > $@
+	@echo "   \includelineno $(<F)" >> $@
+	@echo "*/" >> $@
+
+samples.dox: examples.mk
+	@echo "    [DOXY] $@"
+	@echo "/*!\page samples Sample Code" > $@
+	@echo "    This SDK includes a number of sample applications."\
+	      "Each sample documents a feature of the SDK in both prose"\
+	      "and the associated C code."\
+	      "The following samples are included: ">>$@
+	@$(foreach ex,$(sort $(notdir $(EXAMPLES:.c=))),\
+	   echo "     - \subpage example_$(ex) $($(ex).DESCRIPTION)" >> $@;)
+	@echo >> $@
+	@echo "    In addition, the SDK contains a number of utilities."\
+              "Since these utilities are built upon the concepts described"\
+              "in the sample code listed above, they are not documented in"\
+              "pieces like the samples are. Their source is included here"\
+              "for reference. The following utilities are included:" >> $@
+	@$(foreach ex,$(sort $(UTILS:.c=)),\
+	   echo "     - \subpage example_$(ex) $($(ex).DESCRIPTION)" >> $@;)
+	@echo "*/" >> $@
+
+CLEAN-OBJS += examples.doxy samples.dox $(ALL_EXAMPLES:.c=.dox)
+DOCS-yes += examples.doxy samples.dox
+examples.doxy: samples.dox $(ALL_EXAMPLES:.c=.dox)
+	@echo "INPUT += $^" > $@
diff --git a/libvpx/libvpx/examples/decode_to_md5.c b/libvpx/libvpx/examples/decode_to_md5.c
new file mode 100644
index 0000000..1ae7a4b
--- /dev/null
+++ b/libvpx/libvpx/examples/decode_to_md5.c
@@ -0,0 +1,137 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Frame-by-frame MD5 Checksum
+// ===========================
+//
+// This example builds upon the simple decoder loop to show how checksums
+// of the decoded output can be generated. These are used for validating
+// decoder implementations against the reference implementation, for example.
+//
+// MD5 algorithm
+// -------------
+// The Message-Digest 5 (MD5) is a well known hash function. We have provided
+// an implementation derived from the RSA Data Security, Inc. MD5 Message-Digest
+// Algorithm for your use. Our implmentation only changes the interface of this
+// reference code. You must include the `md5_utils.h` header for access to these
+// functions.
+//
+// Processing The Decoded Data
+// ---------------------------
+// Each row of the image is passed to the MD5 accumulator. First the Y plane
+// is processed, then U, then V. It is important to honor the image's `stride`
+// values.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vp8dx.h"
+#include "vpx/vpx_decoder.h"
+
+#include "../md5_utils.h"
+#include "../tools_common.h"
+#include "../video_reader.h"
+#include "./vpx_config.h"
+
+static void get_image_md5(const vpx_image_t *img, unsigned char digest[16]) {
+  int plane, y;
+  MD5Context md5;
+
+  MD5Init(&md5);
+
+  for (plane = 0; plane < 3; ++plane) {
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int w = plane ? (img->d_w + 1) >> 1 : img->d_w;
+    const int h = plane ? (img->d_h + 1) >> 1 : img->d_h;
+
+    for (y = 0; y < h; ++y) {
+      MD5Update(&md5, buf, w);
+      buf += stride;
+    }
+  }
+
+  MD5Final(digest, &md5);
+}
+
+static void print_md5(FILE *stream, unsigned char digest[16]) {
+  int i;
+
+  for (i = 0; i < 16; ++i)
+    fprintf(stream, "%02x", digest[i]);
+}
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <infile> <outfile>\n", exec_name);
+  exit(EXIT_FAILURE);
+}
+
+int main(int argc, char **argv) {
+  int frame_cnt = 0;
+  FILE *outfile = NULL;
+  vpx_codec_ctx_t codec;
+  VpxVideoReader *reader = NULL;
+  const VpxVideoInfo *info = NULL;
+  const VpxInterface *decoder = NULL;
+
+  exec_name = argv[0];
+
+  if (argc != 3)
+    die("Invalid number of arguments.");
+
+  reader = vpx_video_reader_open(argv[1]);
+  if (!reader)
+    die("Failed to open %s for reading.", argv[1]);
+
+  if (!(outfile = fopen(argv[2], "wb")))
+    die("Failed to open %s for writing.", argv[2]);
+
+  info = vpx_video_reader_get_info(reader);
+
+  decoder = get_vpx_decoder_by_fourcc(info->codec_fourcc);
+  if (!decoder)
+    die("Unknown input codec.");
+
+  printf("Using %s\n", vpx_codec_iface_name(decoder->codec_interface()));
+
+  if (vpx_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0))
+    die_codec(&codec, "Failed to initialize decoder");
+
+  while (vpx_video_reader_read_frame(reader)) {
+    vpx_codec_iter_t iter = NULL;
+    vpx_image_t *img = NULL;
+    size_t frame_size = 0;
+    const unsigned char *frame = vpx_video_reader_get_frame(reader,
+                                                            &frame_size);
+    if (vpx_codec_decode(&codec, frame, (unsigned int)frame_size, NULL, 0))
+      die_codec(&codec, "Failed to decode frame");
+
+    while ((img = vpx_codec_get_frame(&codec, &iter)) != NULL) {
+      unsigned char digest[16];
+
+      get_image_md5(img, digest);
+      print_md5(outfile, digest);
+      fprintf(outfile, "  img-%dx%d-%04d.i420\n",
+              img->d_w, img->d_h, ++frame_cnt);
+    }
+  }
+
+  printf("Processed %d frames.\n", frame_cnt);
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec.");
+
+  vpx_video_reader_close(reader);
+
+  fclose(outfile);
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/decode_with_drops.c b/libvpx/libvpx/examples/decode_with_drops.c
new file mode 100644
index 0000000..2233e47
--- /dev/null
+++ b/libvpx/libvpx/examples/decode_with_drops.c
@@ -0,0 +1,152 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Decode With Drops Example
+// =========================
+//
+// This is an example utility which drops a series of frames, as specified
+// on the command line. This is useful for observing the error recovery
+// features of the codec.
+//
+// Usage
+// -----
+// This example adds a single argument to the `simple_decoder` example,
+// which specifies the range or pattern of frames to drop. The parameter is
+// parsed as follows:
+//
+// Dropping A Range Of Frames
+// --------------------------
+// To drop a range of frames, specify the starting frame and the ending
+// frame to drop, separated by a dash. The following command will drop
+// frames 5 through 10 (base 1).
+//
+//  $ ./decode_with_drops in.ivf out.i420 5-10
+//
+//
+// Dropping A Pattern Of Frames
+// ----------------------------
+// To drop a pattern of frames, specify the number of frames to drop and
+// the number of frames after which to repeat the pattern, separated by
+// a forward-slash. The following command will drop 3 of 7 frames.
+// Specifically, it will decode 4 frames, then drop 3 frames, and then
+// repeat.
+//
+//  $ ./decode_with_drops in.ivf out.i420 3/7
+//
+//
+// Extra Variables
+// ---------------
+// This example maintains the pattern passed on the command line in the
+// `n`, `m`, and `is_range` variables:
+//
+//
+// Making The Drop Decision
+// ------------------------
+// The example decides whether to drop the frame based on the current
+// frame number, immediately before decoding the frame.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vp8dx.h"
+#include "vpx/vpx_decoder.h"
+
+#include "../tools_common.h"
+#include "../video_reader.h"
+#include "./vpx_config.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <infile> <outfile> <N-M|N/M>\n", exec_name);
+  exit(EXIT_FAILURE);
+}
+
+int main(int argc, char **argv) {
+  int frame_cnt = 0;
+  FILE *outfile = NULL;
+  vpx_codec_ctx_t codec;
+  const VpxInterface *decoder = NULL;
+  VpxVideoReader *reader = NULL;
+  const VpxVideoInfo *info = NULL;
+  int n = 0;
+  int m = 0;
+  int is_range = 0;
+  char *nptr = NULL;
+
+  exec_name = argv[0];
+
+  if (argc != 4)
+    die("Invalid number of arguments.");
+
+  reader = vpx_video_reader_open(argv[1]);
+  if (!reader)
+    die("Failed to open %s for reading.", argv[1]);
+
+  if (!(outfile = fopen(argv[2], "wb")))
+    die("Failed to open %s for writing.", argv[2]);
+
+  n = strtol(argv[3], &nptr, 0);
+  m = strtol(nptr + 1, NULL, 0);
+  is_range = (*nptr == '-');
+  if (!n || !m || (*nptr != '-' && *nptr != '/'))
+    die("Couldn't parse pattern %s.\n", argv[3]);
+
+  info = vpx_video_reader_get_info(reader);
+
+  decoder = get_vpx_decoder_by_fourcc(info->codec_fourcc);
+  if (!decoder)
+    die("Unknown input codec.");
+
+  printf("Using %s\n", vpx_codec_iface_name(decoder->codec_interface()));
+
+  if (vpx_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0))
+    die_codec(&codec, "Failed to initialize decoder.");
+
+  while (vpx_video_reader_read_frame(reader)) {
+    vpx_codec_iter_t iter = NULL;
+    vpx_image_t *img = NULL;
+    size_t frame_size = 0;
+    int skip;
+    const unsigned char *frame = vpx_video_reader_get_frame(reader,
+                                                            &frame_size);
+    if (vpx_codec_decode(&codec, frame, (unsigned int)frame_size, NULL, 0))
+      die_codec(&codec, "Failed to decode frame.");
+
+    ++frame_cnt;
+
+    skip = (is_range && frame_cnt >= n && frame_cnt <= m) ||
+           (!is_range && m - (frame_cnt - 1) % m <= n);
+
+    if (!skip) {
+      putc('.', stdout);
+
+      while ((img = vpx_codec_get_frame(&codec, &iter)) != NULL)
+        vpx_img_write(img, outfile);
+    } else {
+      putc('X', stdout);
+    }
+
+    fflush(stdout);
+  }
+
+  printf("Processed %d frames.\n", frame_cnt);
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec.");
+
+  printf("Play: ffplay -f rawvideo -pix_fmt yuv420p -s %dx%d %s\n",
+         info->frame_width, info->frame_height, argv[2]);
+
+  vpx_video_reader_close(reader);
+  fclose(outfile);
+
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/postproc.c b/libvpx/libvpx/examples/postproc.c
new file mode 100644
index 0000000..a8ac208
--- /dev/null
+++ b/libvpx/libvpx/examples/postproc.c
@@ -0,0 +1,138 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Postprocessing Decoder
+// ======================
+//
+// This example adds postprocessing to the simple decoder loop.
+//
+// Initializing Postprocessing
+// ---------------------------
+// You must inform the codec that you might request postprocessing at
+// initialization time. This is done by passing the VPX_CODEC_USE_POSTPROC
+// flag to `vpx_codec_dec_init`. If the codec does not support
+// postprocessing, this call will return VPX_CODEC_INCAPABLE. For
+// demonstration purposes, we also fall back to default initialization if
+// the codec does not provide support.
+//
+// Using Adaptive Postprocessing
+// -----------------------------
+// VP6 provides "adaptive postprocessing." It will automatically select the
+// best postprocessing filter on a frame by frame basis based on the amount
+// of time remaining before the user's specified deadline expires. The
+// special value 0 indicates that the codec should take as long as
+// necessary to provide the best quality frame. This example gives the
+// codec 15ms (15000us) to return a frame. Remember that this is a soft
+// deadline, and the codec may exceed it doing its regular processing. In
+// these cases, no additional postprocessing will be done.
+//
+// Codec Specific Postprocessing Controls
+// --------------------------------------
+// Some codecs provide fine grained controls over their built-in
+// postprocessors. VP8 is one example. The following sample code toggles
+// postprocessing on and off every 15 frames.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vp8dx.h"
+#include "vpx/vpx_decoder.h"
+
+#include "../tools_common.h"
+#include "../video_reader.h"
+#include "./vpx_config.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <infile> <outfile>\n", exec_name);
+  exit(EXIT_FAILURE);
+}
+
+int main(int argc, char **argv) {
+  int frame_cnt = 0;
+  FILE *outfile = NULL;
+  vpx_codec_ctx_t codec;
+  vpx_codec_err_t res;
+  VpxVideoReader *reader = NULL;
+  const VpxInterface *decoder = NULL;
+  const VpxVideoInfo *info = NULL;
+
+  exec_name = argv[0];
+
+  if (argc != 3)
+    die("Invalid number of arguments.");
+
+  reader = vpx_video_reader_open(argv[1]);
+  if (!reader)
+    die("Failed to open %s for reading.", argv[1]);
+
+  if (!(outfile = fopen(argv[2], "wb")))
+    die("Failed to open %s for writing", argv[2]);
+
+  info = vpx_video_reader_get_info(reader);
+
+  decoder = get_vpx_decoder_by_fourcc(info->codec_fourcc);
+  if (!decoder)
+    die("Unknown input codec.");
+
+  printf("Using %s\n", vpx_codec_iface_name(decoder->codec_interface()));
+
+  res = vpx_codec_dec_init(&codec, decoder->codec_interface(), NULL,
+                           VPX_CODEC_USE_POSTPROC);
+  if (res == VPX_CODEC_INCAPABLE)
+    die_codec(&codec, "Postproc not supported by this decoder.");
+
+  if (res)
+    die_codec(&codec, "Failed to initialize decoder.");
+
+  while (vpx_video_reader_read_frame(reader)) {
+    vpx_codec_iter_t iter = NULL;
+    vpx_image_t *img = NULL;
+    size_t frame_size = 0;
+    const unsigned char *frame = vpx_video_reader_get_frame(reader,
+                                                            &frame_size);
+
+    ++frame_cnt;
+
+    if (frame_cnt % 30 == 1) {
+      vp8_postproc_cfg_t pp = {0, 0, 0};
+
+    if (vpx_codec_control(&codec, VP8_SET_POSTPROC, &pp))
+      die_codec(&codec, "Failed to turn off postproc.");
+    } else if (frame_cnt % 30 == 16) {
+      vp8_postproc_cfg_t pp = {VP8_DEBLOCK | VP8_DEMACROBLOCK | VP8_MFQE,
+                               4, 0};
+      if (vpx_codec_control(&codec, VP8_SET_POSTPROC, &pp))
+        die_codec(&codec, "Failed to turn on postproc.");
+    };
+
+    // Decode the frame with 15ms deadline
+    if (vpx_codec_decode(&codec, frame, (unsigned int)frame_size, NULL, 15000))
+      die_codec(&codec, "Failed to decode frame");
+
+    while ((img = vpx_codec_get_frame(&codec, &iter)) != NULL) {
+      vpx_img_write(img, outfile);
+    }
+  }
+
+  printf("Processed %d frames.\n", frame_cnt);
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec");
+
+  printf("Play: ffplay -f rawvideo -pix_fmt yuv420p -s %dx%d %s\n",
+         info->frame_width, info->frame_height, argv[2]);
+
+  vpx_video_reader_close(reader);
+
+  fclose(outfile);
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/resize_util.c b/libvpx/libvpx/examples/resize_util.c
new file mode 100644
index 0000000..e6fdd5b
--- /dev/null
+++ b/libvpx/libvpx/examples/resize_util.c
@@ -0,0 +1,130 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "../tools_common.h"
+#include "../vp9/encoder/vp9_resize.h"
+
+static const char *exec_name = NULL;
+
+static void usage() {
+  printf("Usage:\n");
+  printf("%s <input_yuv> <width>x<height> <target_width>x<target_height> ",
+         exec_name);
+  printf("<output_yuv> [<frames>]\n");
+}
+
+void usage_exit(void) {
+  usage();
+  exit(EXIT_FAILURE);
+}
+
+static int parse_dim(char *v, int *width, int *height) {
+  char *x = strchr(v, 'x');
+  if (x == NULL)
+    x = strchr(v, 'X');
+  if (x == NULL)
+    return 0;
+  *width = atoi(v);
+  *height = atoi(&x[1]);
+  if (*width <= 0 || *height <= 0)
+    return 0;
+  else
+    return 1;
+}
+
+int main(int argc, char *argv[]) {
+  char *fin, *fout;
+  FILE *fpin, *fpout;
+  uint8_t *inbuf, *outbuf;
+  uint8_t *inbuf_u, *outbuf_u;
+  uint8_t *inbuf_v, *outbuf_v;
+  int f, frames;
+  int width, height, target_width, target_height;
+
+  exec_name = argv[0];
+
+  if (argc < 5) {
+    printf("Incorrect parameters:\n");
+    usage();
+    return 1;
+  }
+
+  fin = argv[1];
+  fout = argv[4];
+  if (!parse_dim(argv[2], &width, &height)) {
+    printf("Incorrect parameters: %s\n", argv[2]);
+    usage();
+    return 1;
+  }
+  if (!parse_dim(argv[3], &target_width, &target_height)) {
+    printf("Incorrect parameters: %s\n", argv[3]);
+    usage();
+    return 1;
+  }
+
+  fpin = fopen(fin, "rb");
+  if (fpin == NULL) {
+    printf("Can't open file %s to read\n", fin);
+    usage();
+    return 1;
+  }
+  fpout = fopen(fout, "wb");
+  if (fpout == NULL) {
+    printf("Can't open file %s to write\n", fout);
+    usage();
+    return 1;
+  }
+  if (argc >= 6)
+    frames = atoi(argv[5]);
+  else
+    frames = INT_MAX;
+
+  printf("Input size:  %dx%d\n",
+         width, height);
+  printf("Target size: %dx%d, Frames: ",
+         target_width, target_height);
+  if (frames == INT_MAX)
+    printf("All\n");
+  else
+    printf("%d\n", frames);
+
+  inbuf = (uint8_t*)malloc(width * height * 3 / 2);
+  outbuf = (uint8_t*)malloc(target_width * target_height * 3 / 2);
+  inbuf_u = inbuf + width * height;
+  inbuf_v = inbuf_u + width * height / 4;
+  outbuf_u = outbuf + target_width * target_height;
+  outbuf_v = outbuf_u + target_width * target_height / 4;
+  f = 0;
+  while (f < frames) {
+    if (fread(inbuf, width * height * 3 / 2, 1, fpin) != 1)
+      break;
+    vp9_resize_frame420(inbuf, width, inbuf_u, inbuf_v, width / 2,
+                        height, width,
+                        outbuf, target_width, outbuf_u, outbuf_v,
+                        target_width / 2,
+                        target_height, target_width);
+    fwrite(outbuf, target_width * target_height * 3 / 2, 1, fpout);
+    f++;
+  }
+  printf("%d frames processed\n", f);
+  fclose(fpin);
+  fclose(fpout);
+
+  free(inbuf);
+  free(outbuf);
+  return 0;
+}
diff --git a/libvpx/libvpx/examples/set_maps.c b/libvpx/libvpx/examples/set_maps.c
new file mode 100644
index 0000000..1dc3ac0
--- /dev/null
+++ b/libvpx/libvpx/examples/set_maps.c
@@ -0,0 +1,255 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+// VP8 Set Active and ROI Maps
+// ===========================
+//
+// This is an example demonstrating how to control the VP8 encoder's
+// ROI and Active maps.
+//
+// ROI (Reigon of Interest) maps are a way for the application to assign
+// each macroblock in the image to a region, and then set quantizer and
+// filtering parameters on that image.
+//
+// Active maps are a way for the application to specify on a
+// macroblock-by-macroblock basis whether there is any activity in that
+// macroblock.
+//
+//
+// Configuration
+// -------------
+// An ROI map is set on frame 22. If the width of the image in macroblocks
+// is evenly divisble by 4, then the output will appear to have distinct
+// columns, where the quantizer, loopfilter, and static threshold differ
+// from column to column.
+//
+// An active map is set on frame 33. If the width of the image in macroblocks
+// is evenly divisble by 4, then the output will appear to have distinct
+// columns, where one column will have motion and the next will not.
+//
+// The active map is cleared on frame 44.
+//
+// Observing The Effects
+// ---------------------
+// Use the `simple_decoder` example to decode this sample, and observe
+// the change in the image at frames 22, 33, and 44.
+
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <codec> <width> <height> <infile> <outfile>\n",
+          exec_name);
+  exit(EXIT_FAILURE);
+}
+
+static void set_roi_map(const vpx_codec_enc_cfg_t *cfg,
+                        vpx_codec_ctx_t *codec) {
+  unsigned int i;
+  vpx_roi_map_t roi;
+  memset(&roi, 0, sizeof(roi));
+
+  roi.rows = (cfg->g_h + 15) / 16;
+  roi.cols = (cfg->g_w + 15) / 16;
+
+  roi.delta_q[0] = 0;
+  roi.delta_q[1] = -2;
+  roi.delta_q[2] = -4;
+  roi.delta_q[3] = -6;
+
+  roi.delta_lf[0] = 0;
+  roi.delta_lf[1] = 1;
+  roi.delta_lf[2] = 2;
+  roi.delta_lf[3] = 3;
+
+  roi.static_threshold[0] = 1500;
+  roi.static_threshold[1] = 1000;
+  roi.static_threshold[2] = 500;
+  roi.static_threshold[3] = 0;
+
+  roi.roi_map = (uint8_t *)malloc(roi.rows * roi.cols);
+  for (i = 0; i < roi.rows * roi.cols; ++i)
+    roi.roi_map[i] = i % 4;
+
+  if (vpx_codec_control(codec, VP8E_SET_ROI_MAP, &roi))
+    die_codec(codec, "Failed to set ROI map");
+
+  free(roi.roi_map);
+}
+
+static void set_active_map(const vpx_codec_enc_cfg_t *cfg,
+                           vpx_codec_ctx_t *codec) {
+  unsigned int i;
+  vpx_active_map_t map = {0, 0, 0};
+
+  map.rows = (cfg->g_h + 15) / 16;
+  map.cols = (cfg->g_w + 15) / 16;
+
+  map.active_map = (uint8_t *)malloc(map.rows * map.cols);
+  for (i = 0; i < map.rows * map.cols; ++i)
+    map.active_map[i] = i % 2;
+
+  if (vpx_codec_control(codec, VP8E_SET_ACTIVEMAP, &map))
+    die_codec(codec, "Failed to set active map");
+
+  free(map.active_map);
+}
+
+static void unset_active_map(const vpx_codec_enc_cfg_t *cfg,
+                             vpx_codec_ctx_t *codec) {
+  vpx_active_map_t map = {0, 0, 0};
+
+  map.rows = (cfg->g_h + 15) / 16;
+  map.cols = (cfg->g_w + 15) / 16;
+  map.active_map = NULL;
+
+  if (vpx_codec_control(codec, VP8E_SET_ACTIVEMAP, &map))
+    die_codec(codec, "Failed to set active map");
+}
+
+static int encode_frame(vpx_codec_ctx_t *codec,
+                        vpx_image_t *img,
+                        int frame_index,
+                        VpxVideoWriter *writer) {
+  int got_pkts = 0;
+  vpx_codec_iter_t iter = NULL;
+  const vpx_codec_cx_pkt_t *pkt = NULL;
+  const vpx_codec_err_t res = vpx_codec_encode(codec, img, frame_index, 1, 0,
+                                               VPX_DL_GOOD_QUALITY);
+  if (res != VPX_CODEC_OK)
+    die_codec(codec, "Failed to encode frame");
+
+  while ((pkt = vpx_codec_get_cx_data(codec, &iter)) != NULL) {
+    got_pkts = 1;
+
+    if (pkt->kind == VPX_CODEC_CX_FRAME_PKT) {
+      const int keyframe = (pkt->data.frame.flags & VPX_FRAME_IS_KEY) != 0;
+      if (!vpx_video_writer_write_frame(writer,
+                                        pkt->data.frame.buf,
+                                        pkt->data.frame.sz,
+                                        pkt->data.frame.pts)) {
+        die_codec(codec, "Failed to write compressed frame");
+      }
+
+      printf(keyframe ? "K" : ".");
+      fflush(stdout);
+    }
+  }
+
+  return got_pkts;
+}
+
+int main(int argc, char **argv) {
+  FILE *infile = NULL;
+  vpx_codec_ctx_t codec;
+  vpx_codec_enc_cfg_t cfg;
+  int frame_count = 0;
+  vpx_image_t raw;
+  vpx_codec_err_t res;
+  VpxVideoInfo info;
+  VpxVideoWriter *writer = NULL;
+  const VpxInterface *encoder = NULL;
+  const int fps = 2;        // TODO(dkovalev) add command line argument
+  const double bits_per_pixel_per_frame = 0.067;
+
+  exec_name = argv[0];
+  if (argc != 6)
+    die("Invalid number of arguments");
+
+  memset(&info, 0, sizeof(info));
+
+  encoder = get_vpx_encoder_by_name(argv[1]);
+  if (encoder == NULL) {
+    die("Unsupported codec.");
+  }
+  assert(encoder != NULL);
+  info.codec_fourcc = encoder->fourcc;
+  info.frame_width = strtol(argv[2], NULL, 0);
+  info.frame_height = strtol(argv[3], NULL, 0);
+  info.time_base.numerator = 1;
+  info.time_base.denominator = fps;
+
+  if (info.frame_width <= 0 ||
+      info.frame_height <= 0 ||
+      (info.frame_width % 2) != 0 ||
+      (info.frame_height % 2) != 0) {
+    die("Invalid frame size: %dx%d", info.frame_width, info.frame_height);
+  }
+
+  if (!vpx_img_alloc(&raw, VPX_IMG_FMT_I420, info.frame_width,
+                                             info.frame_height, 1)) {
+    die("Failed to allocate image.");
+  }
+
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
+
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res)
+    die_codec(&codec, "Failed to get default codec config.");
+
+  cfg.g_w = info.frame_width;
+  cfg.g_h = info.frame_height;
+  cfg.g_timebase.num = info.time_base.numerator;
+  cfg.g_timebase.den = info.time_base.denominator;
+  cfg.rc_target_bitrate = (unsigned int)(bits_per_pixel_per_frame * cfg.g_w *
+                                         cfg.g_h * fps / 1000);
+  cfg.g_lag_in_frames = 0;
+
+  writer = vpx_video_writer_open(argv[5], kContainerIVF, &info);
+  if (!writer)
+    die("Failed to open %s for writing.", argv[5]);
+
+  if (!(infile = fopen(argv[4], "rb")))
+    die("Failed to open %s for reading.", argv[4]);
+
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  // Encode frames.
+  while (vpx_img_read(&raw, infile)) {
+    ++frame_count;
+
+    if (frame_count == 22 && encoder->fourcc == VP8_FOURCC) {
+      set_roi_map(&cfg, &codec);
+    } else if (frame_count == 33) {
+      set_active_map(&cfg, &codec);
+    } else if (frame_count == 44) {
+      unset_active_map(&cfg, &codec);
+    }
+
+    encode_frame(&codec, &raw, frame_count, writer);
+  }
+
+  // Flush encoder.
+  while (encode_frame(&codec, NULL, -1, writer)) {}
+
+  printf("\n");
+  fclose(infile);
+  printf("Processed %d frames.\n", frame_count);
+
+  vpx_img_free(&raw);
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec.");
+
+  vpx_video_writer_close(writer);
+
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/simple_decoder.c b/libvpx/libvpx/examples/simple_decoder.c
new file mode 100644
index 0000000..8ccc810
--- /dev/null
+++ b/libvpx/libvpx/examples/simple_decoder.c
@@ -0,0 +1,154 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+// Simple Decoder
+// ==============
+//
+// This is an example of a simple decoder loop. It takes an input file
+// containing the compressed data (in IVF format), passes it through the
+// decoder, and writes the decompressed frames to disk. Other decoder
+// examples build upon this one.
+//
+// The details of the IVF format have been elided from this example for
+// simplicity of presentation, as IVF files will not generally be used by
+// your application. In general, an IVF file consists of a file header,
+// followed by a variable number of frames. Each frame consists of a frame
+// header followed by a variable length payload. The length of the payload
+// is specified in the first four bytes of the frame header. The payload is
+// the raw compressed data.
+//
+// Standard Includes
+// -----------------
+// For decoders, you only have to include `vpx_decoder.h` and then any
+// header files for the specific codecs you use. In this case, we're using
+// vp8.
+//
+// Initializing The Codec
+// ----------------------
+// The libvpx decoder is initialized by the call to vpx_codec_dec_init().
+// Determining the codec interface to use is handled by VpxVideoReader and the
+// functions prefixed with vpx_video_reader_. Discussion of those functions is
+// beyond the scope of this example, but the main gist is to open the input file
+// and parse just enough of it to determine if it's a VPx file and which VPx
+// codec is contained within the file.
+// Note the NULL pointer passed to vpx_codec_dec_init(). We do that in this
+// example because we want the algorithm to determine the stream configuration
+// (width/height) and allocate memory automatically.
+//
+// Decoding A Frame
+// ----------------
+// Once the frame has been read into memory, it is decoded using the
+// `vpx_codec_decode` function. The call takes a pointer to the data
+// (`frame`) and the length of the data (`frame_size`). No application data
+// is associated with the frame in this example, so the `user_priv`
+// parameter is NULL. The `deadline` parameter is left at zero for this
+// example. This parameter is generally only used when doing adaptive post
+// processing.
+//
+// Codecs may produce a variable number of output frames for every call to
+// `vpx_codec_decode`. These frames are retrieved by the
+// `vpx_codec_get_frame` iterator function. The iterator variable `iter` is
+// initialized to NULL each time `vpx_codec_decode` is called.
+// `vpx_codec_get_frame` is called in a loop, returning a pointer to a
+// decoded image or NULL to indicate the end of list.
+//
+// Processing The Decoded Data
+// ---------------------------
+// In this example, we simply write the encoded data to disk. It is
+// important to honor the image's `stride` values.
+//
+// Cleanup
+// -------
+// The `vpx_codec_destroy` call frees any memory allocated by the codec.
+//
+// Error Handling
+// --------------
+// This example does not special case any error return codes. If there was
+// an error, a descriptive message is printed and the program exits. With
+// few exceptions, vpx_codec functions return an enumerated error status,
+// with the value `0` indicating success.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vpx_decoder.h"
+
+#include "../tools_common.h"
+#include "../video_reader.h"
+#include "./vpx_config.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <infile> <outfile>\n", exec_name);
+  exit(EXIT_FAILURE);
+}
+
+int main(int argc, char **argv) {
+  int frame_cnt = 0;
+  FILE *outfile = NULL;
+  vpx_codec_ctx_t codec;
+  VpxVideoReader *reader = NULL;
+  const VpxInterface *decoder = NULL;
+  const VpxVideoInfo *info = NULL;
+
+  exec_name = argv[0];
+
+  if (argc != 3)
+    die("Invalid number of arguments.");
+
+  reader = vpx_video_reader_open(argv[1]);
+  if (!reader)
+    die("Failed to open %s for reading.", argv[1]);
+
+  if (!(outfile = fopen(argv[2], "wb")))
+    die("Failed to open %s for writing.", argv[2]);
+
+  info = vpx_video_reader_get_info(reader);
+
+  decoder = get_vpx_decoder_by_fourcc(info->codec_fourcc);
+  if (!decoder)
+    die("Unknown input codec.");
+
+  printf("Using %s\n", vpx_codec_iface_name(decoder->codec_interface()));
+
+  if (vpx_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0))
+    die_codec(&codec, "Failed to initialize decoder.");
+
+  while (vpx_video_reader_read_frame(reader)) {
+    vpx_codec_iter_t iter = NULL;
+    vpx_image_t *img = NULL;
+    size_t frame_size = 0;
+    const unsigned char *frame = vpx_video_reader_get_frame(reader,
+                                                            &frame_size);
+    if (vpx_codec_decode(&codec, frame, (unsigned int)frame_size, NULL, 0))
+      die_codec(&codec, "Failed to decode frame.");
+
+    while ((img = vpx_codec_get_frame(&codec, &iter)) != NULL) {
+      vpx_img_write(img, outfile);
+      ++frame_cnt;
+    }
+  }
+
+  printf("Processed %d frames.\n", frame_cnt);
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec");
+
+  printf("Play: ffplay -f rawvideo -pix_fmt yuv420p -s %dx%d %s\n",
+         info->frame_width, info->frame_height, argv[2]);
+
+  vpx_video_reader_close(reader);
+
+  fclose(outfile);
+
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/simple_encoder.c b/libvpx/libvpx/examples/simple_encoder.c
new file mode 100644
index 0000000..64f0a01
--- /dev/null
+++ b/libvpx/libvpx/examples/simple_encoder.c
@@ -0,0 +1,260 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Simple Encoder
+// ==============
+//
+// This is an example of a simple encoder loop. It takes an input file in
+// YV12 format, passes it through the encoder, and writes the compressed
+// frames to disk in IVF format. Other decoder examples build upon this
+// one.
+//
+// The details of the IVF format have been elided from this example for
+// simplicity of presentation, as IVF files will not generally be used by
+// your application. In general, an IVF file consists of a file header,
+// followed by a variable number of frames. Each frame consists of a frame
+// header followed by a variable length payload. The length of the payload
+// is specified in the first four bytes of the frame header. The payload is
+// the raw compressed data.
+//
+// Standard Includes
+// -----------------
+// For encoders, you only have to include `vpx_encoder.h` and then any
+// header files for the specific codecs you use. In this case, we're using
+// vp8.
+//
+// Getting The Default Configuration
+// ---------------------------------
+// Encoders have the notion of "usage profiles." For example, an encoder
+// may want to publish default configurations for both a video
+// conferencing application and a best quality offline encoder. These
+// obviously have very different default settings. Consult the
+// documentation for your codec to see if it provides any default
+// configurations. All codecs provide a default configuration, number 0,
+// which is valid for material in the vacinity of QCIF/QVGA.
+//
+// Updating The Configuration
+// ---------------------------------
+// Almost all applications will want to update the default configuration
+// with settings specific to their usage. Here we set the width and height
+// of the video file to that specified on the command line. We also scale
+// the default bitrate based on the ratio between the default resolution
+// and the resolution specified on the command line.
+//
+// Initializing The Codec
+// ----------------------
+// The encoder is initialized by the following code.
+//
+// Encoding A Frame
+// ----------------
+// The frame is read as a continuous block (size width * height * 3 / 2)
+// from the input file. If a frame was read (the input file has not hit
+// EOF) then the frame is passed to the encoder. Otherwise, a NULL
+// is passed, indicating the End-Of-Stream condition to the encoder. The
+// `frame_cnt` is reused as the presentation time stamp (PTS) and each
+// frame is shown for one frame-time in duration. The flags parameter is
+// unused in this example. The deadline is set to VPX_DL_REALTIME to
+// make the example run as quickly as possible.
+
+// Forced Keyframes
+// ----------------
+// Keyframes can be forced by setting the VPX_EFLAG_FORCE_KF bit of the
+// flags passed to `vpx_codec_control()`. In this example, we force a
+// keyframe every <keyframe-interval> frames. Note, the output stream can
+// contain additional keyframes beyond those that have been forced using the
+// VPX_EFLAG_FORCE_KF flag because of automatic keyframe placement by the
+// encoder.
+//
+// Processing The Encoded Data
+// ---------------------------
+// Each packet of type `VPX_CODEC_CX_FRAME_PKT` contains the encoded data
+// for this frame. We write a IVF frame header, followed by the raw data.
+//
+// Cleanup
+// -------
+// The `vpx_codec_destroy` call frees any memory allocated by the codec.
+//
+// Error Handling
+// --------------
+// This example does not special case any error return codes. If there was
+// an error, a descriptive message is printed and the program exits. With
+// few exeptions, vpx_codec functions return an enumerated error status,
+// with the value `0` indicating success.
+//
+// Error Resiliency Features
+// -------------------------
+// Error resiliency is controlled by the g_error_resilient member of the
+// configuration structure. Use the `decode_with_drops` example to decode with
+// frames 5-10 dropped. Compare the output for a file encoded with this example
+// versus one encoded with the `simple_encoder` example.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vpx_encoder.h"
+
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr,
+          "Usage: %s <codec> <width> <height> <infile> <outfile> "
+              "<keyframe-interval> <error-resilient> <frames to encode>\n"
+              "See comments in simple_encoder.c for more information.\n",
+          exec_name);
+  exit(EXIT_FAILURE);
+}
+
+static int encode_frame(vpx_codec_ctx_t *codec,
+                        vpx_image_t *img,
+                        int frame_index,
+                        int flags,
+                        VpxVideoWriter *writer) {
+  int got_pkts = 0;
+  vpx_codec_iter_t iter = NULL;
+  const vpx_codec_cx_pkt_t *pkt = NULL;
+  const vpx_codec_err_t res = vpx_codec_encode(codec, img, frame_index, 1,
+                                               flags, VPX_DL_GOOD_QUALITY);
+  if (res != VPX_CODEC_OK)
+    die_codec(codec, "Failed to encode frame");
+
+  while ((pkt = vpx_codec_get_cx_data(codec, &iter)) != NULL) {
+    got_pkts = 1;
+
+    if (pkt->kind == VPX_CODEC_CX_FRAME_PKT) {
+      const int keyframe = (pkt->data.frame.flags & VPX_FRAME_IS_KEY) != 0;
+      if (!vpx_video_writer_write_frame(writer,
+                                        pkt->data.frame.buf,
+                                        pkt->data.frame.sz,
+                                        pkt->data.frame.pts)) {
+        die_codec(codec, "Failed to write compressed frame");
+      }
+      printf(keyframe ? "K" : ".");
+      fflush(stdout);
+    }
+  }
+
+  return got_pkts;
+}
+
+// TODO(tomfinegan): Improve command line parsing and add args for bitrate/fps.
+int main(int argc, char **argv) {
+  FILE *infile = NULL;
+  vpx_codec_ctx_t codec;
+  vpx_codec_enc_cfg_t cfg;
+  int frame_count = 0;
+  vpx_image_t raw;
+  vpx_codec_err_t res;
+  VpxVideoInfo info = {0};
+  VpxVideoWriter *writer = NULL;
+  const VpxInterface *encoder = NULL;
+  const int fps = 30;
+  const int bitrate = 200;
+  int keyframe_interval = 0;
+  int max_frames = 0;
+  int frames_encoded = 0;
+  const char *codec_arg = NULL;
+  const char *width_arg = NULL;
+  const char *height_arg = NULL;
+  const char *infile_arg = NULL;
+  const char *outfile_arg = NULL;
+  const char *keyframe_interval_arg = NULL;
+
+  exec_name = argv[0];
+
+  if (argc != 9)
+    die("Invalid number of arguments");
+
+  codec_arg = argv[1];
+  width_arg = argv[2];
+  height_arg = argv[3];
+  infile_arg = argv[4];
+  outfile_arg = argv[5];
+  keyframe_interval_arg = argv[6];
+  max_frames = strtol(argv[8], NULL, 0);
+
+  encoder = get_vpx_encoder_by_name(codec_arg);
+  if (!encoder)
+     die("Unsupported codec.");
+
+  info.codec_fourcc = encoder->fourcc;
+  info.frame_width = strtol(width_arg, NULL, 0);
+  info.frame_height = strtol(height_arg, NULL, 0);
+  info.time_base.numerator = 1;
+  info.time_base.denominator = fps;
+
+  if (info.frame_width <= 0 ||
+      info.frame_height <= 0 ||
+      (info.frame_width % 2) != 0 ||
+      (info.frame_height % 2) != 0) {
+    die("Invalid frame size: %dx%d", info.frame_width, info.frame_height);
+  }
+
+  if (!vpx_img_alloc(&raw, VPX_IMG_FMT_I420, info.frame_width,
+                                             info.frame_height, 1)) {
+    die("Failed to allocate image.");
+  }
+
+  keyframe_interval = strtol(keyframe_interval_arg, NULL, 0);
+  if (keyframe_interval < 0)
+    die("Invalid keyframe interval value.");
+
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
+
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res)
+    die_codec(&codec, "Failed to get default codec config.");
+
+  cfg.g_w = info.frame_width;
+  cfg.g_h = info.frame_height;
+  cfg.g_timebase.num = info.time_base.numerator;
+  cfg.g_timebase.den = info.time_base.denominator;
+  cfg.rc_target_bitrate = bitrate;
+  cfg.g_error_resilient = strtol(argv[7], NULL, 0);
+
+  writer = vpx_video_writer_open(outfile_arg, kContainerIVF, &info);
+  if (!writer)
+    die("Failed to open %s for writing.", outfile_arg);
+
+  if (!(infile = fopen(infile_arg, "rb")))
+    die("Failed to open %s for reading.", infile_arg);
+
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  // Encode frames.
+  while (vpx_img_read(&raw, infile)) {
+    int flags = 0;
+    if (keyframe_interval > 0 && frame_count % keyframe_interval == 0)
+      flags |= VPX_EFLAG_FORCE_KF;
+    encode_frame(&codec, &raw, frame_count++, flags, writer);
+    frames_encoded++;
+    if (max_frames > 0 && frames_encoded >= max_frames)
+      break;
+  }
+
+  // Flush encoder.
+  while (encode_frame(&codec, NULL, -1, 0, writer)) {};
+
+  printf("\n");
+  fclose(infile);
+  printf("Processed %d frames.\n", frame_count);
+
+  vpx_img_free(&raw);
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec.");
+
+  vpx_video_writer_close(writer);
+
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/twopass_encoder.c b/libvpx/libvpx/examples/twopass_encoder.c
new file mode 100644
index 0000000..15a6617
--- /dev/null
+++ b/libvpx/libvpx/examples/twopass_encoder.c
@@ -0,0 +1,277 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Two Pass Encoder
+// ================
+//
+// This is an example of a two pass encoder loop. It takes an input file in
+// YV12 format, passes it through the encoder twice, and writes the compressed
+// frames to disk in IVF format. It builds upon the simple_encoder example.
+//
+// Twopass Variables
+// -----------------
+// Twopass mode needs to track the current pass number and the buffer of
+// statistics packets.
+//
+// Updating The Configuration
+// ---------------------------------
+// In two pass mode, the configuration has to be updated on each pass. The
+// statistics buffer is passed on the last pass.
+//
+// Encoding A Frame
+// ----------------
+// Encoding a frame in two pass mode is identical to the simple encoder
+// example. To increase the quality while sacrificing encoding speed,
+// VPX_DL_BEST_QUALITY can be used in place of VPX_DL_GOOD_QUALITY.
+//
+// Processing Statistics Packets
+// -----------------------------
+// Each packet of type `VPX_CODEC_CX_FRAME_PKT` contains the encoded data
+// for this frame. We write a IVF frame header, followed by the raw data.
+//
+//
+// Pass Progress Reporting
+// -----------------------------
+// It's sometimes helpful to see when each pass completes.
+//
+//
+// Clean-up
+// -----------------------------
+// Destruction of the encoder instance must be done on each pass. The
+// raw image should be destroyed at the end as usual.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vpx_encoder.h"
+
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr,
+          "Usage: %s <codec> <width> <height> <infile> <outfile> "
+              "<frame limit>\n",
+          exec_name);
+  exit(EXIT_FAILURE);
+}
+
+static int get_frame_stats(vpx_codec_ctx_t *ctx,
+                           const vpx_image_t *img,
+                           vpx_codec_pts_t pts,
+                           unsigned int duration,
+                           vpx_enc_frame_flags_t flags,
+                           unsigned int deadline,
+                           vpx_fixed_buf_t *stats) {
+  int got_pkts = 0;
+  vpx_codec_iter_t iter = NULL;
+  const vpx_codec_cx_pkt_t *pkt = NULL;
+  const vpx_codec_err_t res = vpx_codec_encode(ctx, img, pts, duration, flags,
+                                               deadline);
+  if (res != VPX_CODEC_OK)
+    die_codec(ctx, "Failed to get frame stats.");
+
+  while ((pkt = vpx_codec_get_cx_data(ctx, &iter)) != NULL) {
+    got_pkts = 1;
+
+    if (pkt->kind == VPX_CODEC_STATS_PKT) {
+      const uint8_t *const pkt_buf = pkt->data.twopass_stats.buf;
+      const size_t pkt_size = pkt->data.twopass_stats.sz;
+      stats->buf = realloc(stats->buf, stats->sz + pkt_size);
+      memcpy((uint8_t *)stats->buf + stats->sz, pkt_buf, pkt_size);
+      stats->sz += pkt_size;
+    }
+  }
+
+  return got_pkts;
+}
+
+static int encode_frame(vpx_codec_ctx_t *ctx,
+                        const vpx_image_t *img,
+                        vpx_codec_pts_t pts,
+                        unsigned int duration,
+                        vpx_enc_frame_flags_t flags,
+                        unsigned int deadline,
+                        VpxVideoWriter *writer) {
+  int got_pkts = 0;
+  vpx_codec_iter_t iter = NULL;
+  const vpx_codec_cx_pkt_t *pkt = NULL;
+  const vpx_codec_err_t res = vpx_codec_encode(ctx, img, pts, duration, flags,
+                                               deadline);
+  if (res != VPX_CODEC_OK)
+    die_codec(ctx, "Failed to encode frame.");
+
+  while ((pkt = vpx_codec_get_cx_data(ctx, &iter)) != NULL) {
+    got_pkts = 1;
+    if (pkt->kind == VPX_CODEC_CX_FRAME_PKT) {
+      const int keyframe = (pkt->data.frame.flags & VPX_FRAME_IS_KEY) != 0;
+
+      if (!vpx_video_writer_write_frame(writer, pkt->data.frame.buf,
+                                                pkt->data.frame.sz,
+                                                pkt->data.frame.pts))
+        die_codec(ctx, "Failed to write compressed frame.");
+      printf(keyframe ? "K" : ".");
+      fflush(stdout);
+    }
+  }
+
+  return got_pkts;
+}
+
+static vpx_fixed_buf_t pass0(vpx_image_t *raw,
+                             FILE *infile,
+                             const VpxInterface *encoder,
+                             const vpx_codec_enc_cfg_t *cfg,
+                             int max_frames) {
+  vpx_codec_ctx_t codec;
+  int frame_count = 0;
+  vpx_fixed_buf_t stats = {NULL, 0};
+
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  // Calculate frame statistics.
+  while (vpx_img_read(raw, infile)) {
+    ++frame_count;
+    get_frame_stats(&codec, raw, frame_count, 1, 0, VPX_DL_GOOD_QUALITY,
+                    &stats);
+    if (max_frames > 0 && frame_count >= max_frames)
+      break;
+  }
+
+  // Flush encoder.
+  while (get_frame_stats(&codec, NULL, frame_count, 1, 0,
+                         VPX_DL_GOOD_QUALITY, &stats)) {}
+
+  printf("Pass 0 complete. Processed %d frames.\n", frame_count);
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec.");
+
+  return stats;
+}
+
+static void pass1(vpx_image_t *raw,
+                  FILE *infile,
+                  const char *outfile_name,
+                  const VpxInterface *encoder,
+                  const vpx_codec_enc_cfg_t *cfg,
+                  int max_frames) {
+  VpxVideoInfo info = {
+    encoder->fourcc,
+    cfg->g_w,
+    cfg->g_h,
+    {cfg->g_timebase.num, cfg->g_timebase.den}
+  };
+  VpxVideoWriter *writer = NULL;
+  vpx_codec_ctx_t codec;
+  int frame_count = 0;
+
+  writer = vpx_video_writer_open(outfile_name, kContainerIVF, &info);
+  if (!writer)
+    die("Failed to open %s for writing", outfile_name);
+
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  // Encode frames.
+  while (vpx_img_read(raw, infile)) {
+    ++frame_count;
+    encode_frame(&codec, raw, frame_count, 1, 0, VPX_DL_GOOD_QUALITY, writer);
+
+    if (max_frames > 0 && frame_count >= max_frames)
+      break;
+  }
+
+  // Flush encoder.
+  while (encode_frame(&codec, NULL, -1, 1, 0, VPX_DL_GOOD_QUALITY, writer)) {}
+
+  printf("\n");
+
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec.");
+
+  vpx_video_writer_close(writer);
+
+  printf("Pass 1 complete. Processed %d frames.\n", frame_count);
+}
+
+int main(int argc, char **argv) {
+  FILE *infile = NULL;
+  int w, h;
+  vpx_codec_ctx_t codec;
+  vpx_codec_enc_cfg_t cfg;
+  vpx_image_t raw;
+  vpx_codec_err_t res;
+  vpx_fixed_buf_t stats;
+
+  const VpxInterface *encoder = NULL;
+  const int fps = 30;        // TODO(dkovalev) add command line argument
+  const int bitrate = 200;   // kbit/s TODO(dkovalev) add command line argument
+  const char *const codec_arg = argv[1];
+  const char *const width_arg = argv[2];
+  const char *const height_arg = argv[3];
+  const char *const infile_arg = argv[4];
+  const char *const outfile_arg = argv[5];
+  int max_frames = 0;
+  exec_name = argv[0];
+
+  if (argc != 7)
+    die("Invalid number of arguments.");
+
+  max_frames = strtol(argv[6], NULL, 0);
+
+  encoder = get_vpx_encoder_by_name(codec_arg);
+  if (!encoder)
+    die("Unsupported codec.");
+
+  w = strtol(width_arg, NULL, 0);
+  h = strtol(height_arg, NULL, 0);
+
+  if (w  <= 0 || h <= 0 || (w % 2) != 0 || (h  % 2) != 0)
+    die("Invalid frame size: %dx%d", w, h);
+
+  if (!vpx_img_alloc(&raw, VPX_IMG_FMT_I420, w, h, 1))
+    die("Failed to allocate image", w, h);
+
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
+
+  // Configuration
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res)
+    die_codec(&codec, "Failed to get default codec config.");
+
+  cfg.g_w = w;
+  cfg.g_h = h;
+  cfg.g_timebase.num = 1;
+  cfg.g_timebase.den = fps;
+  cfg.rc_target_bitrate = bitrate;
+
+  if (!(infile = fopen(infile_arg, "rb")))
+    die("Failed to open %s for reading", infile_arg);
+
+  // Pass 0
+  cfg.g_pass = VPX_RC_FIRST_PASS;
+  stats = pass0(&raw, infile, encoder, &cfg, max_frames);
+
+  // Pass 1
+  rewind(infile);
+  cfg.g_pass = VPX_RC_LAST_PASS;
+  cfg.rc_twopass_stats_in = stats;
+  pass1(&raw, infile, outfile_arg, encoder, &cfg, max_frames);
+  free(stats.buf);
+
+  vpx_img_free(&raw);
+  fclose(infile);
+
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/vp8_multi_resolution_encoder.c b/libvpx/libvpx/examples/vp8_multi_resolution_encoder.c
new file mode 100644
index 0000000..fc775ef
--- /dev/null
+++ b/libvpx/libvpx/examples/vp8_multi_resolution_encoder.c
@@ -0,0 +1,731 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+/*
+ * This is an example demonstrating multi-resolution encoding in VP8.
+ * High-resolution input video is down-sampled to lower-resolutions. The
+ * encoder then encodes the video and outputs multiple bitstreams with
+ * different resolutions.
+ *
+ * This test also allows for settings temporal layers for each spatial layer.
+ * Different number of temporal layers per spatial stream may be used.
+ * Currently up to 3 temporal layers per spatial stream (encoder) are supported
+ * in this test.
+ */
+
+#include "./vpx_config.h"
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdarg.h>
+#include <string.h>
+#include <math.h>
+#include <assert.h>
+#include <sys/time.h>
+#include "vpx_ports/vpx_timer.h"
+#include "vpx/vpx_encoder.h"
+#include "vpx/vp8cx.h"
+#include "vpx_ports/mem_ops.h"
+#include "../tools_common.h"
+#define interface (vpx_codec_vp8_cx())
+#define fourcc    0x30385056
+
+void usage_exit(void) {
+  exit(EXIT_FAILURE);
+}
+
+/*
+ * The input video frame is downsampled several times to generate a multi-level
+ * hierarchical structure. NUM_ENCODERS is defined as the number of encoding
+ * levels required. For example, if the size of input video is 1280x720,
+ * NUM_ENCODERS is 3, and down-sampling factor is 2, the encoder outputs 3
+ * bitstreams with resolution of 1280x720(level 0), 640x360(level 1), and
+ * 320x180(level 2) respectively.
+ */
+
+/* Number of encoders (spatial resolutions) used in this test. */
+#define NUM_ENCODERS 3
+
+/* Maximum number of temporal layers allowed for this test. */
+#define MAX_NUM_TEMPORAL_LAYERS 3
+
+/* This example uses the scaler function in libyuv. */
+#include "third_party/libyuv/include/libyuv/basic_types.h"
+#include "third_party/libyuv/include/libyuv/scale.h"
+#include "third_party/libyuv/include/libyuv/cpu_id.h"
+
+int (*read_frame_p)(FILE *f, vpx_image_t *img);
+
+static int read_frame(FILE *f, vpx_image_t *img) {
+    size_t nbytes, to_read;
+    int    res = 1;
+
+    to_read = img->w*img->h*3/2;
+    nbytes = fread(img->planes[0], 1, to_read, f);
+    if(nbytes != to_read) {
+        res = 0;
+        if(nbytes > 0)
+            printf("Warning: Read partial frame. Check your width & height!\n");
+    }
+    return res;
+}
+
+static int read_frame_by_row(FILE *f, vpx_image_t *img) {
+    size_t nbytes, to_read;
+    int    res = 1;
+    int plane;
+
+    for (plane = 0; plane < 3; plane++)
+    {
+        unsigned char *ptr;
+        int w = (plane ? (1 + img->d_w) / 2 : img->d_w);
+        int h = (plane ? (1 + img->d_h) / 2 : img->d_h);
+        int r;
+
+        /* Determine the correct plane based on the image format. The for-loop
+         * always counts in Y,U,V order, but this may not match the order of
+         * the data on disk.
+         */
+        switch (plane)
+        {
+        case 1:
+            ptr = img->planes[img->fmt==VPX_IMG_FMT_YV12? VPX_PLANE_V : VPX_PLANE_U];
+            break;
+        case 2:
+            ptr = img->planes[img->fmt==VPX_IMG_FMT_YV12?VPX_PLANE_U : VPX_PLANE_V];
+            break;
+        default:
+            ptr = img->planes[plane];
+        }
+
+        for (r = 0; r < h; r++)
+        {
+            to_read = w;
+
+            nbytes = fread(ptr, 1, to_read, f);
+            if(nbytes != to_read) {
+                res = 0;
+                if(nbytes > 0)
+                    printf("Warning: Read partial frame. Check your width & height!\n");
+                break;
+            }
+
+            ptr += img->stride[plane];
+        }
+        if (!res)
+            break;
+    }
+
+    return res;
+}
+
+static void write_ivf_file_header(FILE *outfile,
+                                  const vpx_codec_enc_cfg_t *cfg,
+                                  int frame_cnt) {
+    char header[32];
+
+    if(cfg->g_pass != VPX_RC_ONE_PASS && cfg->g_pass != VPX_RC_LAST_PASS)
+        return;
+    header[0] = 'D';
+    header[1] = 'K';
+    header[2] = 'I';
+    header[3] = 'F';
+    mem_put_le16(header+4,  0);                   /* version */
+    mem_put_le16(header+6,  32);                  /* headersize */
+    mem_put_le32(header+8,  fourcc);              /* headersize */
+    mem_put_le16(header+12, cfg->g_w);            /* width */
+    mem_put_le16(header+14, cfg->g_h);            /* height */
+    mem_put_le32(header+16, cfg->g_timebase.den); /* rate */
+    mem_put_le32(header+20, cfg->g_timebase.num); /* scale */
+    mem_put_le32(header+24, frame_cnt);           /* length */
+    mem_put_le32(header+28, 0);                   /* unused */
+
+    (void) fwrite(header, 1, 32, outfile);
+}
+
+static void write_ivf_frame_header(FILE *outfile,
+                                   const vpx_codec_cx_pkt_t *pkt)
+{
+    char             header[12];
+    vpx_codec_pts_t  pts;
+
+    if(pkt->kind != VPX_CODEC_CX_FRAME_PKT)
+        return;
+
+    pts = pkt->data.frame.pts;
+    mem_put_le32(header, pkt->data.frame.sz);
+    mem_put_le32(header+4, pts&0xFFFFFFFF);
+    mem_put_le32(header+8, pts >> 32);
+
+    (void) fwrite(header, 1, 12, outfile);
+}
+
+/* Temporal scaling parameters */
+/* This sets all the temporal layer parameters given |num_temporal_layers|,
+ * including the target bit allocation across temporal layers. Bit allocation
+ * parameters will be passed in as user parameters in another version.
+ */
+static void set_temporal_layer_pattern(int num_temporal_layers,
+                                       vpx_codec_enc_cfg_t *cfg,
+                                       int bitrate,
+                                       int *layer_flags)
+{
+    assert(num_temporal_layers <= MAX_NUM_TEMPORAL_LAYERS);
+    switch (num_temporal_layers)
+    {
+    case 1:
+    {
+        /* 1-layer */
+        cfg->ts_number_layers     = 1;
+        cfg->ts_periodicity       = 1;
+        cfg->ts_rate_decimator[0] = 1;
+        cfg->ts_layer_id[0] = 0;
+        cfg->ts_target_bitrate[0] = bitrate;
+
+        // Update L only.
+        layer_flags[0] = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+        break;
+    }
+
+    case 2:
+    {
+        /* 2-layers, with sync point at first frame of layer 1. */
+        cfg->ts_number_layers     = 2;
+        cfg->ts_periodicity       = 2;
+        cfg->ts_rate_decimator[0] = 2;
+        cfg->ts_rate_decimator[1] = 1;
+        cfg->ts_layer_id[0] = 0;
+        cfg->ts_layer_id[1] = 1;
+        // Use 60/40 bit allocation as example.
+        cfg->ts_target_bitrate[0] = 0.6f * bitrate;
+        cfg->ts_target_bitrate[1] = bitrate;
+
+        /* 0=L, 1=GF */
+        // ARF is used as predictor for all frames, and is only updated on
+        // key frame. Sync point every 8 frames.
+
+        // Layer 0: predict from L and ARF, update L and G.
+        layer_flags[0] = VP8_EFLAG_NO_REF_GF |
+                         VP8_EFLAG_NO_UPD_ARF;
+
+        // Layer 1: sync point: predict from L and ARF, and update G.
+        layer_flags[1] = VP8_EFLAG_NO_REF_GF |
+                         VP8_EFLAG_NO_UPD_LAST |
+                         VP8_EFLAG_NO_UPD_ARF;
+
+        // Layer 0, predict from L and ARF, update L.
+        layer_flags[2] = VP8_EFLAG_NO_REF_GF  |
+                         VP8_EFLAG_NO_UPD_GF  |
+                         VP8_EFLAG_NO_UPD_ARF;
+
+        // Layer 1: predict from L, G and ARF, and update G.
+        layer_flags[3] = VP8_EFLAG_NO_UPD_ARF |
+                         VP8_EFLAG_NO_UPD_LAST |
+                         VP8_EFLAG_NO_UPD_ENTROPY;
+
+        // Layer 0
+        layer_flags[4] = layer_flags[2];
+
+        // Layer 1
+        layer_flags[5] = layer_flags[3];
+
+        // Layer 0
+        layer_flags[6] = layer_flags[4];
+
+        // Layer 1
+        layer_flags[7] = layer_flags[5];
+        break;
+    }
+
+    case 3:
+    default:
+    {
+        // 3-layers structure where ARF is used as predictor for all frames,
+        // and is only updated on key frame.
+        // Sync points for layer 1 and 2 every 8 frames.
+        cfg->ts_number_layers     = 3;
+        cfg->ts_periodicity       = 4;
+        cfg->ts_rate_decimator[0] = 4;
+        cfg->ts_rate_decimator[1] = 2;
+        cfg->ts_rate_decimator[2] = 1;
+        cfg->ts_layer_id[0] = 0;
+        cfg->ts_layer_id[1] = 2;
+        cfg->ts_layer_id[2] = 1;
+        cfg->ts_layer_id[3] = 2;
+        // Use 40/20/40 bit allocation as example.
+        cfg->ts_target_bitrate[0] = 0.4f * bitrate;
+        cfg->ts_target_bitrate[1] = 0.6f * bitrate;
+        cfg->ts_target_bitrate[2] = bitrate;
+
+        /* 0=L, 1=GF, 2=ARF */
+
+        // Layer 0: predict from L and ARF; update L and G.
+        layer_flags[0] =  VP8_EFLAG_NO_UPD_ARF |
+                          VP8_EFLAG_NO_REF_GF;
+
+        // Layer 2: sync point: predict from L and ARF; update none.
+        layer_flags[1] = VP8_EFLAG_NO_REF_GF |
+                         VP8_EFLAG_NO_UPD_GF |
+                         VP8_EFLAG_NO_UPD_ARF |
+                         VP8_EFLAG_NO_UPD_LAST |
+                         VP8_EFLAG_NO_UPD_ENTROPY;
+
+        // Layer 1: sync point: predict from L and ARF; update G.
+        layer_flags[2] = VP8_EFLAG_NO_REF_GF |
+                         VP8_EFLAG_NO_UPD_ARF |
+                         VP8_EFLAG_NO_UPD_LAST;
+
+        // Layer 2: predict from L, G, ARF; update none.
+        layer_flags[3] = VP8_EFLAG_NO_UPD_GF |
+                         VP8_EFLAG_NO_UPD_ARF |
+                         VP8_EFLAG_NO_UPD_LAST |
+                         VP8_EFLAG_NO_UPD_ENTROPY;
+
+        // Layer 0: predict from L and ARF; update L.
+        layer_flags[4] = VP8_EFLAG_NO_UPD_GF |
+                         VP8_EFLAG_NO_UPD_ARF |
+                         VP8_EFLAG_NO_REF_GF;
+
+        // Layer 2: predict from L, G, ARF; update none.
+        layer_flags[5] = layer_flags[3];
+
+        // Layer 1: predict from L, G, ARF; update G.
+        layer_flags[6] = VP8_EFLAG_NO_UPD_ARF |
+                         VP8_EFLAG_NO_UPD_LAST;
+
+        // Layer 2: predict from L, G, ARF; update none.
+        layer_flags[7] = layer_flags[3];
+        break;
+    }
+    }
+}
+
+/* The periodicity of the pattern given the number of temporal layers. */
+static int periodicity_to_num_layers[MAX_NUM_TEMPORAL_LAYERS] = {1, 8, 8};
+
+int main(int argc, char **argv)
+{
+    FILE                 *infile, *outfile[NUM_ENCODERS];
+    FILE                 *downsampled_input[NUM_ENCODERS - 1];
+    char                 filename[50];
+    vpx_codec_ctx_t      codec[NUM_ENCODERS];
+    vpx_codec_enc_cfg_t  cfg[NUM_ENCODERS];
+    int                  frame_cnt = 0;
+    vpx_image_t          raw[NUM_ENCODERS];
+    vpx_codec_err_t      res[NUM_ENCODERS];
+
+    int                  i;
+    long                 width;
+    long                 height;
+    int                  length_frame;
+    int                  frame_avail;
+    int                  got_data;
+    int                  flags = 0;
+    int                  layer_id = 0;
+
+    int                  layer_flags[VPX_TS_MAX_PERIODICITY * NUM_ENCODERS]
+                                     = {0};
+    int                  flag_periodicity;
+
+    /*Currently, only realtime mode is supported in multi-resolution encoding.*/
+    int                  arg_deadline = VPX_DL_REALTIME;
+
+    /* Set show_psnr to 1/0 to show/not show PSNR. Choose show_psnr=0 if you
+       don't need to know PSNR, which will skip PSNR calculation and save
+       encoding time. */
+    int                  show_psnr = 0;
+    int                  key_frame_insert = 0;
+    uint64_t             psnr_sse_total[NUM_ENCODERS] = {0};
+    uint64_t             psnr_samples_total[NUM_ENCODERS] = {0};
+    double               psnr_totals[NUM_ENCODERS][4] = {{0,0}};
+    int                  psnr_count[NUM_ENCODERS] = {0};
+
+    int64_t              cx_time = 0;
+
+    /* Set the required target bitrates for each resolution level.
+     * If target bitrate for highest-resolution level is set to 0,
+     * (i.e. target_bitrate[0]=0), we skip encoding at that level.
+     */
+    unsigned int         target_bitrate[NUM_ENCODERS]={1000, 500, 100};
+
+    /* Enter the frame rate of the input video */
+    int                  framerate = 30;
+
+    /* Set down-sampling factor for each resolution level.
+       dsf[0] controls down sampling from level 0 to level 1;
+       dsf[1] controls down sampling from level 1 to level 2;
+       dsf[2] is not used. */
+    vpx_rational_t dsf[NUM_ENCODERS] = {{2, 1}, {2, 1}, {1, 1}};
+
+    /* Set the number of temporal layers for each encoder/resolution level,
+     * starting from highest resoln down to lowest resoln. */
+    unsigned int         num_temporal_layers[NUM_ENCODERS] = {3, 3, 3};
+
+    if(argc!= (7 + 3 * NUM_ENCODERS))
+        die("Usage: %s <width> <height> <frame_rate>  <infile> <outfile(s)> "
+            "<rate_encoder(s)> <temporal_layer(s)> <key_frame_insert> <output psnr?> \n",
+            argv[0]);
+
+    printf("Using %s\n",vpx_codec_iface_name(interface));
+
+    width = strtol(argv[1], NULL, 0);
+    height = strtol(argv[2], NULL, 0);
+    framerate = strtol(argv[3], NULL, 0);
+
+    if(width < 16 || width%2 || height <16 || height%2)
+        die("Invalid resolution: %ldx%ld", width, height);
+
+    /* Open input video file for encoding */
+    if(!(infile = fopen(argv[4], "rb")))
+        die("Failed to open %s for reading", argv[4]);
+
+    /* Open output file for each encoder to output bitstreams */
+    for (i=0; i< NUM_ENCODERS; i++)
+    {
+        if(!target_bitrate[i])
+        {
+            outfile[i] = NULL;
+            continue;
+        }
+
+        if(!(outfile[i] = fopen(argv[i+5], "wb")))
+            die("Failed to open %s for writing", argv[i+4]);
+    }
+
+    // Bitrates per spatial layer: overwrite default rates above.
+    for (i=0; i< NUM_ENCODERS; i++)
+    {
+        target_bitrate[i] = strtol(argv[NUM_ENCODERS + 5 + i], NULL, 0);
+    }
+
+    // Temporal layers per spatial layers: overwrite default settings above.
+    for (i=0; i< NUM_ENCODERS; i++)
+    {
+        num_temporal_layers[i] = strtol(argv[2 * NUM_ENCODERS + 5 + i], NULL, 0);
+        if (num_temporal_layers[i] < 1 || num_temporal_layers[i] > 3)
+          die("Invalid temporal layers: %d, Must be 1, 2, or 3. \n",
+              num_temporal_layers);
+    }
+
+    /* Open file to write out each spatially downsampled input stream. */
+    for (i=0; i< NUM_ENCODERS - 1; i++)
+    {
+       // Highest resoln is encoder 0.
+        if (sprintf(filename,"ds%d.yuv",NUM_ENCODERS - i) < 0)
+        {
+            return EXIT_FAILURE;
+        }
+        downsampled_input[i] = fopen(filename,"wb");
+    }
+
+    key_frame_insert = strtol(argv[3 * NUM_ENCODERS + 5], NULL, 0);
+
+    show_psnr = strtol(argv[3 * NUM_ENCODERS + 6], NULL, 0);
+
+
+    /* Populate default encoder configuration */
+    for (i=0; i< NUM_ENCODERS; i++)
+    {
+        res[i] = vpx_codec_enc_config_default(interface, &cfg[i], 0);
+        if(res[i]) {
+            printf("Failed to get config: %s\n", vpx_codec_err_to_string(res[i]));
+            return EXIT_FAILURE;
+        }
+    }
+
+    /*
+     * Update the default configuration according to needs of the application.
+     */
+    /* Highest-resolution encoder settings */
+    cfg[0].g_w = width;
+    cfg[0].g_h = height;
+    cfg[0].rc_dropframe_thresh = 0;
+    cfg[0].rc_end_usage = VPX_CBR;
+    cfg[0].rc_resize_allowed = 0;
+    cfg[0].rc_min_quantizer = 2;
+    cfg[0].rc_max_quantizer = 56;
+    cfg[0].rc_undershoot_pct = 100;
+    cfg[0].rc_overshoot_pct = 15;
+    cfg[0].rc_buf_initial_sz = 500;
+    cfg[0].rc_buf_optimal_sz = 600;
+    cfg[0].rc_buf_sz = 1000;
+    cfg[0].g_error_resilient = 1;              /* Enable error resilient mode */
+    cfg[0].g_lag_in_frames   = 0;
+
+    /* Disable automatic keyframe placement */
+    /* Note: These 3 settings are copied to all levels. But, except the lowest
+     * resolution level, all other levels are set to VPX_KF_DISABLED internally.
+     */
+    cfg[0].kf_mode           = VPX_KF_AUTO;
+    cfg[0].kf_min_dist = 3000;
+    cfg[0].kf_max_dist = 3000;
+
+    cfg[0].rc_target_bitrate = target_bitrate[0];       /* Set target bitrate */
+    cfg[0].g_timebase.num = 1;                          /* Set fps */
+    cfg[0].g_timebase.den = framerate;
+
+    /* Other-resolution encoder settings */
+    for (i=1; i< NUM_ENCODERS; i++)
+    {
+        memcpy(&cfg[i], &cfg[0], sizeof(vpx_codec_enc_cfg_t));
+
+        cfg[i].rc_target_bitrate = target_bitrate[i];
+
+        /* Note: Width & height of other-resolution encoders are calculated
+         * from the highest-resolution encoder's size and the corresponding
+         * down_sampling_factor.
+         */
+        {
+            unsigned int iw = cfg[i-1].g_w*dsf[i-1].den + dsf[i-1].num - 1;
+            unsigned int ih = cfg[i-1].g_h*dsf[i-1].den + dsf[i-1].num - 1;
+            cfg[i].g_w = iw/dsf[i-1].num;
+            cfg[i].g_h = ih/dsf[i-1].num;
+        }
+
+        /* Make width & height to be multiplier of 2. */
+        // Should support odd size ???
+        if((cfg[i].g_w)%2)cfg[i].g_w++;
+        if((cfg[i].g_h)%2)cfg[i].g_h++;
+    }
+
+
+    // Set the number of threads per encode/spatial layer.
+    // (1, 1, 1) means no encoder threading.
+    cfg[0].g_threads = 2;
+    cfg[1].g_threads = 1;
+    cfg[2].g_threads = 1;
+
+    /* Allocate image for each encoder */
+    for (i=0; i< NUM_ENCODERS; i++)
+        if(!vpx_img_alloc(&raw[i], VPX_IMG_FMT_I420, cfg[i].g_w, cfg[i].g_h, 32))
+            die("Failed to allocate image", cfg[i].g_w, cfg[i].g_h);
+
+    if (raw[0].stride[VPX_PLANE_Y] == raw[0].d_w)
+        read_frame_p = read_frame;
+    else
+        read_frame_p = read_frame_by_row;
+
+    for (i=0; i< NUM_ENCODERS; i++)
+        if(outfile[i])
+            write_ivf_file_header(outfile[i], &cfg[i], 0);
+
+    /* Temporal layers settings */
+    for ( i=0; i<NUM_ENCODERS; i++)
+    {
+        set_temporal_layer_pattern(num_temporal_layers[i],
+                                   &cfg[i],
+                                   cfg[i].rc_target_bitrate,
+                                   &layer_flags[i * VPX_TS_MAX_PERIODICITY]);
+    }
+
+    /* Initialize multi-encoder */
+    if(vpx_codec_enc_init_multi(&codec[0], interface, &cfg[0], NUM_ENCODERS,
+                                (show_psnr ? VPX_CODEC_USE_PSNR : 0), &dsf[0]))
+        die_codec(&codec[0], "Failed to initialize encoder");
+
+    /* The extra encoding configuration parameters can be set as follows. */
+    /* Set encoding speed */
+    for ( i=0; i<NUM_ENCODERS; i++)
+    {
+        int speed = -6;
+        /* Lower speed for the lowest resolution. */
+        if (i == NUM_ENCODERS - 1) speed = -4;
+        if(vpx_codec_control(&codec[i], VP8E_SET_CPUUSED, speed))
+            die_codec(&codec[i], "Failed to set cpu_used");
+    }
+
+    /* Set static threshold = 1 for all encoders */
+    for ( i=0; i<NUM_ENCODERS; i++)
+    {
+        if(vpx_codec_control(&codec[i], VP8E_SET_STATIC_THRESHOLD, 1))
+            die_codec(&codec[i], "Failed to set static threshold");
+    }
+
+    /* Set NOISE_SENSITIVITY to do TEMPORAL_DENOISING */
+    /* Enable denoising for the highest-resolution encoder. */
+    if(vpx_codec_control(&codec[0], VP8E_SET_NOISE_SENSITIVITY, 1))
+        die_codec(&codec[0], "Failed to set noise_sensitivity");
+    for ( i=1; i< NUM_ENCODERS; i++)
+    {
+        if(vpx_codec_control(&codec[i], VP8E_SET_NOISE_SENSITIVITY, 0))
+            die_codec(&codec[i], "Failed to set noise_sensitivity");
+    }
+
+    /* Set the number of token partitions */
+    for ( i=0; i<NUM_ENCODERS; i++)
+    {
+        if(vpx_codec_control(&codec[i], VP8E_SET_TOKEN_PARTITIONS, 1))
+            die_codec(&codec[i], "Failed to set static threshold");
+    }
+
+    /* Set the max intra target bitrate */
+    for ( i=0; i<NUM_ENCODERS; i++)
+    {
+        unsigned int max_intra_size_pct =
+            (int)(((double)cfg[0].rc_buf_optimal_sz * 0.5) * framerate / 10);
+        if(vpx_codec_control(&codec[i], VP8E_SET_MAX_INTRA_BITRATE_PCT,
+                             max_intra_size_pct))
+            die_codec(&codec[i], "Failed to set static threshold");
+       //printf("%d %d \n",i,max_intra_size_pct);
+    }
+
+    frame_avail = 1;
+    got_data = 0;
+
+    while(frame_avail || got_data)
+    {
+        struct vpx_usec_timer timer;
+        vpx_codec_iter_t iter[NUM_ENCODERS]={NULL};
+        const vpx_codec_cx_pkt_t *pkt[NUM_ENCODERS];
+
+        flags = 0;
+        frame_avail = read_frame_p(infile, &raw[0]);
+
+        if(frame_avail)
+        {
+            for ( i=1; i<NUM_ENCODERS; i++)
+            {
+                /*Scale the image down a number of times by downsampling factor*/
+                /* FilterMode 1 or 2 give better psnr than FilterMode 0. */
+                I420Scale(raw[i-1].planes[VPX_PLANE_Y], raw[i-1].stride[VPX_PLANE_Y],
+                          raw[i-1].planes[VPX_PLANE_U], raw[i-1].stride[VPX_PLANE_U],
+                          raw[i-1].planes[VPX_PLANE_V], raw[i-1].stride[VPX_PLANE_V],
+                          raw[i-1].d_w, raw[i-1].d_h,
+                          raw[i].planes[VPX_PLANE_Y], raw[i].stride[VPX_PLANE_Y],
+                          raw[i].planes[VPX_PLANE_U], raw[i].stride[VPX_PLANE_U],
+                          raw[i].planes[VPX_PLANE_V], raw[i].stride[VPX_PLANE_V],
+                          raw[i].d_w, raw[i].d_h, 1);
+                /* Write out down-sampled input. */
+                length_frame = cfg[i].g_w *  cfg[i].g_h *3/2;
+                if (fwrite(raw[i].planes[0], 1, length_frame,
+                           downsampled_input[NUM_ENCODERS - i - 1]) !=
+                               length_frame)
+                {
+                    return EXIT_FAILURE;
+                }
+            }
+        }
+
+        /* Set the flags (reference and update) for all the encoders.*/
+        for ( i=0; i<NUM_ENCODERS; i++)
+        {
+            layer_id = cfg[i].ts_layer_id[frame_cnt % cfg[i].ts_periodicity];
+            flags = 0;
+            flag_periodicity = periodicity_to_num_layers
+                [num_temporal_layers[i] - 1];
+            flags = layer_flags[i * VPX_TS_MAX_PERIODICITY +
+                                frame_cnt % flag_periodicity];
+            // Key frame flag for first frame.
+            if (frame_cnt == 0)
+            {
+                flags |= VPX_EFLAG_FORCE_KF;
+            }
+            if (frame_cnt > 0 && frame_cnt == key_frame_insert)
+            {
+                flags = VPX_EFLAG_FORCE_KF;
+            }
+
+            vpx_codec_control(&codec[i], VP8E_SET_FRAME_FLAGS, flags);
+            vpx_codec_control(&codec[i], VP8E_SET_TEMPORAL_LAYER_ID, layer_id);
+        }
+
+        /* Encode each frame at multi-levels */
+        /* Note the flags must be set to 0 in the encode call if they are set
+           for each frame with the vpx_codec_control(), as done above. */
+        vpx_usec_timer_start(&timer);
+        if(vpx_codec_encode(&codec[0], frame_avail? &raw[0] : NULL,
+            frame_cnt, 1, 0, arg_deadline))
+        {
+            die_codec(&codec[0], "Failed to encode frame");
+        }
+        vpx_usec_timer_mark(&timer);
+        cx_time += vpx_usec_timer_elapsed(&timer);
+
+        for (i=NUM_ENCODERS-1; i>=0 ; i--)
+        {
+            got_data = 0;
+            while( (pkt[i] = vpx_codec_get_cx_data(&codec[i], &iter[i])) )
+            {
+                got_data = 1;
+                switch(pkt[i]->kind) {
+                    case VPX_CODEC_CX_FRAME_PKT:
+                        write_ivf_frame_header(outfile[i], pkt[i]);
+                        (void) fwrite(pkt[i]->data.frame.buf, 1,
+                                      pkt[i]->data.frame.sz, outfile[i]);
+                    break;
+                    case VPX_CODEC_PSNR_PKT:
+                        if (show_psnr)
+                        {
+                            int j;
+
+                            psnr_sse_total[i] += pkt[i]->data.psnr.sse[0];
+                            psnr_samples_total[i] += pkt[i]->data.psnr.samples[0];
+                            for (j = 0; j < 4; j++)
+                            {
+                                psnr_totals[i][j] += pkt[i]->data.psnr.psnr[j];
+                            }
+                            psnr_count[i]++;
+                        }
+
+                        break;
+                    default:
+                        break;
+                }
+                printf(pkt[i]->kind == VPX_CODEC_CX_FRAME_PKT
+                       && (pkt[i]->data.frame.flags & VPX_FRAME_IS_KEY)? "K":"");
+                fflush(stdout);
+            }
+        }
+        frame_cnt++;
+    }
+    printf("\n");
+    printf("Frame cnt and encoding time/FPS stats for encoding: %d %f %f \n",
+            frame_cnt,
+            1000 * (float)cx_time / (double)(frame_cnt * 1000000),
+            1000000 * (double)frame_cnt / (double)cx_time);
+
+    fclose(infile);
+
+    printf("Processed %ld frames.\n",(long int)frame_cnt-1);
+    for (i=0; i< NUM_ENCODERS; i++)
+    {
+        /* Calculate PSNR and print it out */
+        if ( (show_psnr) && (psnr_count[i]>0) )
+        {
+            int j;
+            double ovpsnr = sse_to_psnr(psnr_samples_total[i], 255.0,
+                                        psnr_sse_total[i]);
+
+            fprintf(stderr, "\n ENC%d PSNR (Overall/Avg/Y/U/V)", i);
+
+            fprintf(stderr, " %.3lf", ovpsnr);
+            for (j = 0; j < 4; j++)
+            {
+                fprintf(stderr, " %.3lf", psnr_totals[i][j]/psnr_count[i]);
+            }
+        }
+
+        if(vpx_codec_destroy(&codec[i]))
+            die_codec(&codec[i], "Failed to destroy codec");
+
+        vpx_img_free(&raw[i]);
+
+        if(!outfile[i])
+            continue;
+
+        /* Try to rewrite the file header with the actual frame count */
+        if(!fseek(outfile[i], 0, SEEK_SET))
+            write_ivf_file_header(outfile[i], &cfg[i], frame_cnt-1);
+        fclose(outfile[i]);
+    }
+    printf("\n");
+
+    return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/vp8cx_set_ref.c b/libvpx/libvpx/examples/vp8cx_set_ref.c
new file mode 100644
index 0000000..8b4cc30
--- /dev/null
+++ b/libvpx/libvpx/examples/vp8cx_set_ref.c
@@ -0,0 +1,194 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+// VP8 Set Reference Frame
+// =======================
+//
+// This is an example demonstrating how to overwrite the VP8 encoder's
+// internal reference frame. In the sample we set the last frame to the
+// current frame. If this is done at a cut scene it will avoid a keyframe.
+// This technique could be used to bounce between two cameras.
+//
+// Note that the decoder would also have to set the reference frame to the
+// same value on the same frame, or the video will become corrupt.
+//
+// Usage
+// -----
+// This example adds a single argument to the `simple_encoder` example,
+// which specifies the frame number to update the reference frame on.
+// The parameter is parsed as follows:
+//
+//
+// Extra Variables
+// ---------------
+// This example maintains the frame number passed on the command line
+// in the `update_frame_num` variable.
+//
+//
+// Configuration
+// -------------
+//
+// The reference frame is updated on the frame specified on the command
+// line.
+//
+// Observing The Effects
+// ---------------------
+// Use the `simple_encoder` example to encode a sample with a cut scene.
+// Determine the frame number of the cut scene by looking for a generated
+// key-frame (indicated by a 'K'). Supply that frame number as an argument
+// to this example, and observe that no key-frame is generated.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <width> <height> <infile> <outfile> <frame>\n",
+          exec_name);
+  exit(EXIT_FAILURE);
+}
+
+static int encode_frame(vpx_codec_ctx_t *codec,
+                        vpx_image_t *img,
+                        int frame_index,
+                        VpxVideoWriter *writer) {
+  int got_pkts = 0;
+  vpx_codec_iter_t iter = NULL;
+  const vpx_codec_cx_pkt_t *pkt = NULL;
+  const vpx_codec_err_t res = vpx_codec_encode(codec, img, frame_index, 1, 0,
+                                               VPX_DL_GOOD_QUALITY);
+  if (res != VPX_CODEC_OK)
+    die_codec(codec, "Failed to encode frame");
+
+  while ((pkt = vpx_codec_get_cx_data(codec, &iter)) != NULL) {
+    got_pkts = 1;
+
+    if (pkt->kind == VPX_CODEC_CX_FRAME_PKT) {
+      const int keyframe = (pkt->data.frame.flags & VPX_FRAME_IS_KEY) != 0;
+      if (!vpx_video_writer_write_frame(writer,
+                                        pkt->data.frame.buf,
+                                        pkt->data.frame.sz,
+                                        pkt->data.frame.pts)) {
+        die_codec(codec, "Failed to write compressed frame");
+      }
+
+      printf(keyframe ? "K" : ".");
+      fflush(stdout);
+    }
+  }
+
+  return got_pkts;
+}
+
+int main(int argc, char **argv) {
+  FILE *infile = NULL;
+  vpx_codec_ctx_t codec = {0};
+  vpx_codec_enc_cfg_t cfg = {0};
+  int frame_count = 0;
+  vpx_image_t raw;
+  vpx_codec_err_t res;
+  VpxVideoInfo info = {0};
+  VpxVideoWriter *writer = NULL;
+  const VpxInterface *encoder = NULL;
+  int update_frame_num = 0;
+  const int fps = 30;        // TODO(dkovalev) add command line argument
+  const int bitrate = 200;   // kbit/s TODO(dkovalev) add command line argument
+
+  exec_name = argv[0];
+
+  if (argc != 6)
+    die("Invalid number of arguments");
+
+  // TODO(dkovalev): add vp9 support and rename the file accordingly
+  encoder = get_vpx_encoder_by_name("vp8");
+  if (!encoder)
+    die("Unsupported codec.");
+
+  update_frame_num = atoi(argv[5]);
+  if (!update_frame_num)
+    die("Couldn't parse frame number '%s'\n", argv[5]);
+
+  info.codec_fourcc = encoder->fourcc;
+  info.frame_width = strtol(argv[1], NULL, 0);
+  info.frame_height = strtol(argv[2], NULL, 0);
+  info.time_base.numerator = 1;
+  info.time_base.denominator = fps;
+
+  if (info.frame_width <= 0 ||
+      info.frame_height <= 0 ||
+      (info.frame_width % 2) != 0 ||
+      (info.frame_height % 2) != 0) {
+    die("Invalid frame size: %dx%d", info.frame_width, info.frame_height);
+  }
+
+  if (!vpx_img_alloc(&raw, VPX_IMG_FMT_I420, info.frame_width,
+                                             info.frame_height, 1)) {
+    die("Failed to allocate image.");
+  }
+
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
+
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res)
+    die_codec(&codec, "Failed to get default codec config.");
+
+  cfg.g_w = info.frame_width;
+  cfg.g_h = info.frame_height;
+  cfg.g_timebase.num = info.time_base.numerator;
+  cfg.g_timebase.den = info.time_base.denominator;
+  cfg.rc_target_bitrate = bitrate;
+
+  writer = vpx_video_writer_open(argv[4], kContainerIVF, &info);
+  if (!writer)
+    die("Failed to open %s for writing.", argv[4]);
+
+  if (!(infile = fopen(argv[3], "rb")))
+    die("Failed to open %s for reading.", argv[3]);
+
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  // Encode frames.
+  while (vpx_img_read(&raw, infile)) {
+    if (frame_count + 1 == update_frame_num) {
+      vpx_ref_frame_t ref;
+      ref.frame_type = VP8_LAST_FRAME;
+      ref.img = raw;
+      if (vpx_codec_control(&codec, VP8_SET_REFERENCE, &ref))
+        die_codec(&codec, "Failed to set reference frame");
+    }
+
+    encode_frame(&codec, &raw, frame_count++, writer);
+  }
+
+  // Flush encoder.
+  while (encode_frame(&codec, NULL, -1, writer)) {}
+
+  printf("\n");
+  fclose(infile);
+  printf("Processed %d frames.\n", frame_count);
+
+  vpx_img_free(&raw);
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec.");
+
+  vpx_video_writer_close(writer);
+
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/vp9_lossless_encoder.c b/libvpx/libvpx/examples/vp9_lossless_encoder.c
new file mode 100644
index 0000000..8272516
--- /dev/null
+++ b/libvpx/libvpx/examples/vp9_lossless_encoder.c
@@ -0,0 +1,144 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vpx_encoder.h"
+#include "vpx/vp8cx.h"
+
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "vp9_lossless_encoder: Example demonstrating VP9 lossless "
+                  "encoding feature. Supports raw input only.\n");
+  fprintf(stderr, "Usage: %s <width> <height> <infile> <outfile>\n", exec_name);
+  exit(EXIT_FAILURE);
+}
+
+static int encode_frame(vpx_codec_ctx_t *codec,
+                        vpx_image_t *img,
+                        int frame_index,
+                        int flags,
+                        VpxVideoWriter *writer) {
+  int got_pkts = 0;
+  vpx_codec_iter_t iter = NULL;
+  const vpx_codec_cx_pkt_t *pkt = NULL;
+  const vpx_codec_err_t res = vpx_codec_encode(codec, img, frame_index, 1,
+                                               flags, VPX_DL_GOOD_QUALITY);
+  if (res != VPX_CODEC_OK)
+    die_codec(codec, "Failed to encode frame");
+
+  while ((pkt = vpx_codec_get_cx_data(codec, &iter)) != NULL) {
+    got_pkts = 1;
+
+    if (pkt->kind == VPX_CODEC_CX_FRAME_PKT) {
+      const int keyframe = (pkt->data.frame.flags & VPX_FRAME_IS_KEY) != 0;
+      if (!vpx_video_writer_write_frame(writer,
+                                        pkt->data.frame.buf,
+                                        pkt->data.frame.sz,
+                                        pkt->data.frame.pts)) {
+        die_codec(codec, "Failed to write compressed frame");
+      }
+      printf(keyframe ? "K" : ".");
+      fflush(stdout);
+    }
+  }
+
+  return got_pkts;
+}
+
+int main(int argc, char **argv) {
+  FILE *infile = NULL;
+  vpx_codec_ctx_t codec;
+  vpx_codec_enc_cfg_t cfg;
+  int frame_count = 0;
+  vpx_image_t raw;
+  vpx_codec_err_t res;
+  VpxVideoInfo info = {0};
+  VpxVideoWriter *writer = NULL;
+  const VpxInterface *encoder = NULL;
+  const int fps = 30;
+
+  exec_name = argv[0];
+
+  if (argc < 5)
+    die("Invalid number of arguments");
+
+  encoder = get_vpx_encoder_by_name("vp9");
+  if (!encoder)
+     die("Unsupported codec.");
+
+  info.codec_fourcc = encoder->fourcc;
+  info.frame_width = strtol(argv[1], NULL, 0);
+  info.frame_height = strtol(argv[2], NULL, 0);
+  info.time_base.numerator = 1;
+  info.time_base.denominator = fps;
+
+  if (info.frame_width <= 0 ||
+      info.frame_height <= 0 ||
+      (info.frame_width % 2) != 0 ||
+      (info.frame_height % 2) != 0) {
+    die("Invalid frame size: %dx%d", info.frame_width, info.frame_height);
+  }
+
+  if (!vpx_img_alloc(&raw, VPX_IMG_FMT_I420, info.frame_width,
+                                             info.frame_height, 1)) {
+    die("Failed to allocate image.");
+  }
+
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
+
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res)
+    die_codec(&codec, "Failed to get default codec config.");
+
+  cfg.g_w = info.frame_width;
+  cfg.g_h = info.frame_height;
+  cfg.g_timebase.num = info.time_base.numerator;
+  cfg.g_timebase.den = info.time_base.denominator;
+
+  writer = vpx_video_writer_open(argv[4], kContainerIVF, &info);
+  if (!writer)
+    die("Failed to open %s for writing.", argv[4]);
+
+  if (!(infile = fopen(argv[3], "rb")))
+    die("Failed to open %s for reading.", argv[3]);
+
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  if (vpx_codec_control_(&codec, VP9E_SET_LOSSLESS, 1))
+    die_codec(&codec, "Failed to use lossless mode");
+
+  // Encode frames.
+  while (vpx_img_read(&raw, infile)) {
+    encode_frame(&codec, &raw, frame_count++, 0, writer);
+  }
+
+  // Flush encoder.
+  while (encode_frame(&codec, NULL, -1, 0, writer)) {}
+
+  printf("\n");
+  fclose(infile);
+  printf("Processed %d frames.\n", frame_count);
+
+  vpx_img_free(&raw);
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec.");
+
+  vpx_video_writer_close(writer);
+
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/vp9_spatial_svc_encoder.c b/libvpx/libvpx/examples/vp9_spatial_svc_encoder.c
new file mode 100644
index 0000000..271ab70
--- /dev/null
+++ b/libvpx/libvpx/examples/vp9_spatial_svc_encoder.c
@@ -0,0 +1,919 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+/*
+ * This is an example demonstrating how to implement a multi-layer
+ * VP9 encoding scheme based on spatial scalability for video applications
+ * that benefit from a scalable bitstream.
+ */
+
+#include <math.h>
+#include <stdarg.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+
+
+#include "../args.h"
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+#include "../vpx_ports/vpx_timer.h"
+#include "vpx/svc_context.h"
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+#include "../vpxstats.h"
+#include "vp9/encoder/vp9_encoder.h"
+#define OUTPUT_RC_STATS 1
+
+static const arg_def_t skip_frames_arg =
+    ARG_DEF("s", "skip-frames", 1, "input frames to skip");
+static const arg_def_t frames_arg =
+    ARG_DEF("f", "frames", 1, "number of frames to encode");
+static const arg_def_t threads_arg =
+    ARG_DEF("th", "threads", 1, "number of threads to use");
+#if OUTPUT_RC_STATS
+static const arg_def_t output_rc_stats_arg =
+    ARG_DEF("rcstat", "output_rc_stats", 1, "output rc stats");
+#endif
+static const arg_def_t width_arg = ARG_DEF("w", "width", 1, "source width");
+static const arg_def_t height_arg = ARG_DEF("h", "height", 1, "source height");
+static const arg_def_t timebase_arg =
+    ARG_DEF("t", "timebase", 1, "timebase (num/den)");
+static const arg_def_t bitrate_arg = ARG_DEF(
+    "b", "target-bitrate", 1, "encoding bitrate, in kilobits per second");
+static const arg_def_t spatial_layers_arg =
+    ARG_DEF("sl", "spatial-layers", 1, "number of spatial SVC layers");
+static const arg_def_t temporal_layers_arg =
+    ARG_DEF("tl", "temporal-layers", 1, "number of temporal SVC layers");
+static const arg_def_t temporal_layering_mode_arg =
+    ARG_DEF("tlm", "temporal-layering-mode", 1, "temporal layering scheme."
+        "VP9E_TEMPORAL_LAYERING_MODE");
+static const arg_def_t kf_dist_arg =
+    ARG_DEF("k", "kf-dist", 1, "number of frames between keyframes");
+static const arg_def_t scale_factors_arg =
+    ARG_DEF("r", "scale-factors", 1, "scale factors (lowest to highest layer)");
+static const arg_def_t passes_arg =
+    ARG_DEF("p", "passes", 1, "Number of passes (1/2)");
+static const arg_def_t pass_arg =
+    ARG_DEF(NULL, "pass", 1, "Pass to execute (1/2)");
+static const arg_def_t fpf_name_arg =
+    ARG_DEF(NULL, "fpf", 1, "First pass statistics file name");
+static const arg_def_t min_q_arg =
+    ARG_DEF(NULL, "min-q", 1, "Minimum quantizer");
+static const arg_def_t max_q_arg =
+    ARG_DEF(NULL, "max-q", 1, "Maximum quantizer");
+static const arg_def_t min_bitrate_arg =
+    ARG_DEF(NULL, "min-bitrate", 1, "Minimum bitrate");
+static const arg_def_t max_bitrate_arg =
+    ARG_DEF(NULL, "max-bitrate", 1, "Maximum bitrate");
+static const arg_def_t lag_in_frame_arg =
+    ARG_DEF(NULL, "lag-in-frames", 1, "Number of frame to input before "
+        "generating any outputs");
+static const arg_def_t rc_end_usage_arg =
+    ARG_DEF(NULL, "rc-end-usage", 1, "0 - 3: VBR, CBR, CQ, Q");
+static const arg_def_t speed_arg =
+    ARG_DEF("sp", "speed", 1, "speed configuration");
+static const arg_def_t aqmode_arg =
+    ARG_DEF("aq", "aqmode", 1, "aq-mode off/on");
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static const struct arg_enum_list bitdepth_enum[] = {
+  {"8",  VPX_BITS_8},
+  {"10", VPX_BITS_10},
+  {"12", VPX_BITS_12},
+  {NULL, 0}
+};
+
+static const arg_def_t bitdepth_arg =
+    ARG_DEF_ENUM("d", "bit-depth", 1, "Bit depth for codec 8, 10 or 12. ",
+                 bitdepth_enum);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+
+static const arg_def_t *svc_args[] = {
+  &frames_arg,        &width_arg,         &height_arg,
+  &timebase_arg,      &bitrate_arg,       &skip_frames_arg, &spatial_layers_arg,
+  &kf_dist_arg,       &scale_factors_arg, &passes_arg,      &pass_arg,
+  &fpf_name_arg,      &min_q_arg,         &max_q_arg,       &min_bitrate_arg,
+  &max_bitrate_arg,   &temporal_layers_arg, &temporal_layering_mode_arg,
+  &lag_in_frame_arg,  &threads_arg,       &aqmode_arg,
+#if OUTPUT_RC_STATS
+  &output_rc_stats_arg,
+#endif
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  &bitdepth_arg,
+#endif
+  &speed_arg,
+  &rc_end_usage_arg,  NULL
+};
+
+static const uint32_t default_frames_to_skip = 0;
+static const uint32_t default_frames_to_code = 60 * 60;
+static const uint32_t default_width = 1920;
+static const uint32_t default_height = 1080;
+static const uint32_t default_timebase_num = 1;
+static const uint32_t default_timebase_den = 60;
+static const uint32_t default_bitrate = 1000;
+static const uint32_t default_spatial_layers = 5;
+static const uint32_t default_temporal_layers = 1;
+static const uint32_t default_kf_dist = 100;
+static const uint32_t default_temporal_layering_mode = 0;
+static const uint32_t default_output_rc_stats = 0;
+static const int32_t default_speed = -1;  // -1 means use library default.
+static const uint32_t default_threads = 0;  // zero means use library default.
+
+typedef struct {
+  const char *input_filename;
+  const char *output_filename;
+  uint32_t frames_to_code;
+  uint32_t frames_to_skip;
+  struct VpxInputContext input_ctx;
+  stats_io_t rc_stats;
+  int passes;
+  int pass;
+} AppInput;
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <options> input_filename output_filename\n",
+          exec_name);
+  fprintf(stderr, "Options:\n");
+  arg_show_usage(stderr, svc_args);
+  exit(EXIT_FAILURE);
+}
+
+static void parse_command_line(int argc, const char **argv_,
+                               AppInput *app_input, SvcContext *svc_ctx,
+                               vpx_codec_enc_cfg_t *enc_cfg) {
+  struct arg arg = {0};
+  char **argv = NULL;
+  char **argi = NULL;
+  char **argj = NULL;
+  vpx_codec_err_t res;
+  int passes = 0;
+  int pass = 0;
+  const char *fpf_file_name = NULL;
+  unsigned int min_bitrate = 0;
+  unsigned int max_bitrate = 0;
+  char string_options[1024] = {0};
+
+  // initialize SvcContext with parameters that will be passed to vpx_svc_init
+  svc_ctx->log_level = SVC_LOG_DEBUG;
+  svc_ctx->spatial_layers = default_spatial_layers;
+  svc_ctx->temporal_layers = default_temporal_layers;
+  svc_ctx->temporal_layering_mode = default_temporal_layering_mode;
+#if OUTPUT_RC_STATS
+  svc_ctx->output_rc_stat = default_output_rc_stats;
+#endif
+  svc_ctx->speed = default_speed;
+  svc_ctx->threads = default_threads;
+
+  // start with default encoder configuration
+  res = vpx_codec_enc_config_default(vpx_codec_vp9_cx(), enc_cfg, 0);
+  if (res) {
+    die("Failed to get config: %s\n", vpx_codec_err_to_string(res));
+  }
+  // update enc_cfg with app default values
+  enc_cfg->g_w = default_width;
+  enc_cfg->g_h = default_height;
+  enc_cfg->g_timebase.num = default_timebase_num;
+  enc_cfg->g_timebase.den = default_timebase_den;
+  enc_cfg->rc_target_bitrate = default_bitrate;
+  enc_cfg->kf_min_dist = default_kf_dist;
+  enc_cfg->kf_max_dist = default_kf_dist;
+  enc_cfg->rc_end_usage = VPX_CQ;
+
+  // initialize AppInput with default values
+  app_input->frames_to_code = default_frames_to_code;
+  app_input->frames_to_skip = default_frames_to_skip;
+
+  // process command line options
+  argv = argv_dup(argc - 1, argv_ + 1);
+  for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
+    arg.argv_step = 1;
+
+    if (arg_match(&arg, &frames_arg, argi)) {
+      app_input->frames_to_code = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &width_arg, argi)) {
+      enc_cfg->g_w = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &height_arg, argi)) {
+      enc_cfg->g_h = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &timebase_arg, argi)) {
+      enc_cfg->g_timebase = arg_parse_rational(&arg);
+    } else if (arg_match(&arg, &bitrate_arg, argi)) {
+      enc_cfg->rc_target_bitrate = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &skip_frames_arg, argi)) {
+      app_input->frames_to_skip = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &spatial_layers_arg, argi)) {
+      svc_ctx->spatial_layers = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &temporal_layers_arg, argi)) {
+      svc_ctx->temporal_layers = arg_parse_uint(&arg);
+#if OUTPUT_RC_STATS
+    } else if (arg_match(&arg, &output_rc_stats_arg, argi)) {
+      svc_ctx->output_rc_stat = arg_parse_uint(&arg);
+#endif
+    } else if (arg_match(&arg, &speed_arg, argi)) {
+      svc_ctx->speed = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &aqmode_arg, argi)) {
+      svc_ctx->aqmode = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &threads_arg, argi)) {
+      svc_ctx->threads = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &temporal_layering_mode_arg, argi)) {
+      svc_ctx->temporal_layering_mode =
+          enc_cfg->temporal_layering_mode = arg_parse_int(&arg);
+      if (svc_ctx->temporal_layering_mode) {
+        enc_cfg->g_error_resilient = 1;
+      }
+    } else if (arg_match(&arg, &kf_dist_arg, argi)) {
+      enc_cfg->kf_min_dist = arg_parse_uint(&arg);
+      enc_cfg->kf_max_dist = enc_cfg->kf_min_dist;
+    } else if (arg_match(&arg, &scale_factors_arg, argi)) {
+      snprintf(string_options, sizeof(string_options), "%s scale-factors=%s",
+               string_options, arg.val);
+    } else if (arg_match(&arg, &passes_arg, argi)) {
+      passes = arg_parse_uint(&arg);
+      if (passes < 1 || passes > 2) {
+        die("Error: Invalid number of passes (%d)\n", passes);
+      }
+    } else if (arg_match(&arg, &pass_arg, argi)) {
+      pass = arg_parse_uint(&arg);
+      if (pass < 1 || pass > 2) {
+        die("Error: Invalid pass selected (%d)\n", pass);
+      }
+    } else if (arg_match(&arg, &fpf_name_arg, argi)) {
+      fpf_file_name = arg.val;
+    } else if (arg_match(&arg, &min_q_arg, argi)) {
+      snprintf(string_options, sizeof(string_options), "%s min-quantizers=%s",
+               string_options, arg.val);
+    } else if (arg_match(&arg, &max_q_arg, argi)) {
+      snprintf(string_options, sizeof(string_options), "%s max-quantizers=%s",
+               string_options, arg.val);
+    } else if (arg_match(&arg, &min_bitrate_arg, argi)) {
+      min_bitrate = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &max_bitrate_arg, argi)) {
+      max_bitrate = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &lag_in_frame_arg, argi)) {
+      enc_cfg->g_lag_in_frames = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &rc_end_usage_arg, argi)) {
+      enc_cfg->rc_end_usage = arg_parse_uint(&arg);
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else if (arg_match(&arg, &bitdepth_arg, argi)) {
+      enc_cfg->g_bit_depth = arg_parse_enum_or_int(&arg);
+      switch (enc_cfg->g_bit_depth) {
+        case VPX_BITS_8:
+          enc_cfg->g_input_bit_depth = 8;
+          enc_cfg->g_profile = 0;
+          break;
+        case VPX_BITS_10:
+          enc_cfg->g_input_bit_depth = 10;
+          enc_cfg->g_profile = 2;
+          break;
+         case VPX_BITS_12:
+          enc_cfg->g_input_bit_depth = 12;
+          enc_cfg->g_profile = 2;
+          break;
+        default:
+          die("Error: Invalid bit depth selected (%d)\n", enc_cfg->g_bit_depth);
+          break;
+      }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    } else {
+      ++argj;
+    }
+  }
+
+  // There will be a space in front of the string options
+  if (strlen(string_options) > 0)
+    vpx_svc_set_options(svc_ctx, string_options + 1);
+
+  if (passes == 0 || passes == 1) {
+    if (pass) {
+      fprintf(stderr, "pass is ignored since there's only one pass\n");
+    }
+    enc_cfg->g_pass = VPX_RC_ONE_PASS;
+  } else {
+    if (pass == 0) {
+      die("pass must be specified when passes is 2\n");
+    }
+
+    if (fpf_file_name == NULL) {
+      die("fpf must be specified when passes is 2\n");
+    }
+
+    if (pass == 1) {
+      enc_cfg->g_pass = VPX_RC_FIRST_PASS;
+      if (!stats_open_file(&app_input->rc_stats, fpf_file_name, 0)) {
+        fatal("Failed to open statistics store");
+      }
+    } else {
+      enc_cfg->g_pass = VPX_RC_LAST_PASS;
+      if (!stats_open_file(&app_input->rc_stats, fpf_file_name, 1)) {
+        fatal("Failed to open statistics store");
+      }
+      enc_cfg->rc_twopass_stats_in = stats_get(&app_input->rc_stats);
+    }
+    app_input->passes = passes;
+    app_input->pass = pass;
+  }
+
+  if (enc_cfg->rc_target_bitrate > 0) {
+    if (min_bitrate > 0) {
+      enc_cfg->rc_2pass_vbr_minsection_pct =
+          min_bitrate * 100 / enc_cfg->rc_target_bitrate;
+    }
+    if (max_bitrate > 0) {
+      enc_cfg->rc_2pass_vbr_maxsection_pct =
+          max_bitrate * 100 / enc_cfg->rc_target_bitrate;
+    }
+  }
+
+  // Check for unrecognized options
+  for (argi = argv; *argi; ++argi)
+    if (argi[0][0] == '-' && strlen(argi[0]) > 1)
+      die("Error: Unrecognized option %s\n", *argi);
+
+  if (argv[0] == NULL || argv[1] == 0) {
+    usage_exit();
+  }
+  app_input->input_filename = argv[0];
+  app_input->output_filename = argv[1];
+  free(argv);
+
+  if (enc_cfg->g_w < 16 || enc_cfg->g_w % 2 || enc_cfg->g_h < 16 ||
+      enc_cfg->g_h % 2)
+    die("Invalid resolution: %d x %d\n", enc_cfg->g_w, enc_cfg->g_h);
+
+  printf(
+      "Codec %s\nframes: %d, skip: %d\n"
+      "layers: %d\n"
+      "width %d, height: %d,\n"
+      "num: %d, den: %d, bitrate: %d,\n"
+      "gop size: %d\n",
+      vpx_codec_iface_name(vpx_codec_vp9_cx()), app_input->frames_to_code,
+      app_input->frames_to_skip,
+      svc_ctx->spatial_layers, enc_cfg->g_w, enc_cfg->g_h,
+      enc_cfg->g_timebase.num, enc_cfg->g_timebase.den,
+      enc_cfg->rc_target_bitrate, enc_cfg->kf_max_dist);
+}
+
+#if OUTPUT_RC_STATS
+// For rate control encoding stats.
+struct RateControlStats {
+  // Number of input frames per layer.
+  int layer_input_frames[VPX_MAX_LAYERS];
+  // Total (cumulative) number of encoded frames per layer.
+  int layer_tot_enc_frames[VPX_MAX_LAYERS];
+  // Number of encoded non-key frames per layer.
+  int layer_enc_frames[VPX_MAX_LAYERS];
+  // Framerate per layer (cumulative).
+  double layer_framerate[VPX_MAX_LAYERS];
+  // Target average frame size per layer (per-frame-bandwidth per layer).
+  double layer_pfb[VPX_MAX_LAYERS];
+  // Actual average frame size per layer.
+  double layer_avg_frame_size[VPX_MAX_LAYERS];
+  // Average rate mismatch per layer (|target - actual| / target).
+  double layer_avg_rate_mismatch[VPX_MAX_LAYERS];
+  // Actual encoding bitrate per layer (cumulative).
+  double layer_encoding_bitrate[VPX_MAX_LAYERS];
+  // Average of the short-time encoder actual bitrate.
+  // TODO(marpan): Should we add these short-time stats for each layer?
+  double avg_st_encoding_bitrate;
+  // Variance of the short-time encoder actual bitrate.
+  double variance_st_encoding_bitrate;
+  // Window (number of frames) for computing short-time encoding bitrate.
+  int window_size;
+  // Number of window measurements.
+  int window_count;
+};
+
+// Note: these rate control stats assume only 1 key frame in the
+// sequence (i.e., first frame only).
+static void set_rate_control_stats(struct RateControlStats *rc,
+                                     vpx_codec_enc_cfg_t *cfg) {
+  unsigned int sl, tl;
+  // Set the layer (cumulative) framerate and the target layer (non-cumulative)
+  // per-frame-bandwidth, for the rate control encoding stats below.
+  const double framerate = cfg->g_timebase.den / cfg->g_timebase.num;
+
+  for (sl = 0; sl < cfg->ss_number_layers; ++sl) {
+    for (tl = 0; tl < cfg->ts_number_layers; ++tl) {
+      const int layer = sl * cfg->ts_number_layers + tl;
+      const int tlayer0 = sl * cfg->ts_number_layers;
+      if (cfg->ts_number_layers == 1)
+        rc->layer_framerate[layer] = framerate;
+      else
+        rc->layer_framerate[layer] =
+          framerate / cfg->ts_rate_decimator[tl];
+      if (tl > 0) {
+        rc->layer_pfb[layer] = 1000.0 *
+            (cfg->layer_target_bitrate[layer] -
+                cfg->layer_target_bitrate[layer - 1]) /
+            (rc->layer_framerate[layer] -
+                rc->layer_framerate[layer - 1]);
+      } else {
+        rc->layer_pfb[tlayer0] = 1000.0 *
+            cfg->layer_target_bitrate[tlayer0] /
+            rc->layer_framerate[tlayer0];
+      }
+      rc->layer_input_frames[layer] = 0;
+      rc->layer_enc_frames[layer] = 0;
+      rc->layer_tot_enc_frames[layer] = 0;
+      rc->layer_encoding_bitrate[layer] = 0.0;
+      rc->layer_avg_frame_size[layer] = 0.0;
+      rc->layer_avg_rate_mismatch[layer] = 0.0;
+    }
+  }
+  rc->window_count = 0;
+  rc->window_size = 15;
+  rc->avg_st_encoding_bitrate = 0.0;
+  rc->variance_st_encoding_bitrate = 0.0;
+}
+
+static void printout_rate_control_summary(struct RateControlStats *rc,
+                                          vpx_codec_enc_cfg_t *cfg,
+                                          int frame_cnt) {
+  unsigned int sl, tl;
+  int tot_num_frames = 0;
+  double perc_fluctuation = 0.0;
+  printf("Total number of processed frames: %d\n\n", frame_cnt - 1);
+  printf("Rate control layer stats for sl%d tl%d layer(s):\n\n",
+      cfg->ss_number_layers, cfg->ts_number_layers);
+  for (sl = 0; sl < cfg->ss_number_layers; ++sl) {
+    for (tl = 0; tl < cfg->ts_number_layers; ++tl) {
+      const int layer = sl * cfg->ts_number_layers + tl;
+      const int num_dropped = (tl > 0) ?
+          (rc->layer_input_frames[layer] - rc->layer_enc_frames[layer]) :
+          (rc->layer_input_frames[layer] - rc->layer_enc_frames[layer] - 1);
+      if (!sl)
+        tot_num_frames += rc->layer_input_frames[layer];
+      rc->layer_encoding_bitrate[layer] = 0.001 * rc->layer_framerate[layer] *
+          rc->layer_encoding_bitrate[layer] / tot_num_frames;
+      rc->layer_avg_frame_size[layer] = rc->layer_avg_frame_size[layer] /
+          rc->layer_enc_frames[layer];
+      rc->layer_avg_rate_mismatch[layer] =
+          100.0 * rc->layer_avg_rate_mismatch[layer] /
+          rc->layer_enc_frames[layer];
+      printf("For layer#: sl%d tl%d \n", sl, tl);
+      printf("Bitrate (target vs actual): %d %f.0 kbps\n",
+             cfg->layer_target_bitrate[layer],
+             rc->layer_encoding_bitrate[layer]);
+      printf("Average frame size (target vs actual): %f %f bits\n",
+             rc->layer_pfb[layer], rc->layer_avg_frame_size[layer]);
+      printf("Average rate_mismatch: %f\n",
+             rc->layer_avg_rate_mismatch[layer]);
+      printf("Number of input frames, encoded (non-key) frames, "
+          "and percent dropped frames: %d %d %f.0 \n",
+          rc->layer_input_frames[layer], rc->layer_enc_frames[layer],
+          100.0 * num_dropped / rc->layer_input_frames[layer]);
+      printf("\n");
+    }
+  }
+  rc->avg_st_encoding_bitrate = rc->avg_st_encoding_bitrate / rc->window_count;
+  rc->variance_st_encoding_bitrate =
+      rc->variance_st_encoding_bitrate / rc->window_count -
+      (rc->avg_st_encoding_bitrate * rc->avg_st_encoding_bitrate);
+  perc_fluctuation = 100.0 * sqrt(rc->variance_st_encoding_bitrate) /
+      rc->avg_st_encoding_bitrate;
+  printf("Short-time stats, for window of %d frames: \n", rc->window_size);
+  printf("Average, rms-variance, and percent-fluct: %f %f %f \n",
+         rc->avg_st_encoding_bitrate,
+         sqrt(rc->variance_st_encoding_bitrate),
+         perc_fluctuation);
+  if (frame_cnt != tot_num_frames)
+    die("Error: Number of input frames not equal to output encoded frames != "
+        "%d tot_num_frames = %d\n", frame_cnt, tot_num_frames);
+}
+
+vpx_codec_err_t parse_superframe_index(const uint8_t *data,
+                                       size_t data_sz,
+                                       uint32_t sizes[8], int *count) {
+  // A chunk ending with a byte matching 0xc0 is an invalid chunk unless
+  // it is a super frame index. If the last byte of real video compression
+  // data is 0xc0 the encoder must add a 0 byte. If we have the marker but
+  // not the associated matching marker byte at the front of the index we have
+  // an invalid bitstream and need to return an error.
+
+  uint8_t marker;
+
+  marker = *(data + data_sz - 1);
+  *count = 0;
+
+
+  if ((marker & 0xe0) == 0xc0) {
+    const uint32_t frames = (marker & 0x7) + 1;
+    const uint32_t mag = ((marker >> 3) & 0x3) + 1;
+    const size_t index_sz = 2 + mag * frames;
+
+    // This chunk is marked as having a superframe index but doesn't have
+    // enough data for it, thus it's an invalid superframe index.
+    if (data_sz < index_sz)
+      return VPX_CODEC_CORRUPT_FRAME;
+
+    {
+      const uint8_t marker2 = *(data + data_sz - index_sz);
+
+      // This chunk is marked as having a superframe index but doesn't have
+      // the matching marker byte at the front of the index therefore it's an
+      // invalid chunk.
+      if (marker != marker2)
+        return VPX_CODEC_CORRUPT_FRAME;
+    }
+
+    {
+      // Found a valid superframe index.
+      uint32_t i, j;
+      const uint8_t *x = &data[data_sz - index_sz + 1];
+
+      for (i = 0; i < frames; ++i) {
+        uint32_t this_sz = 0;
+
+        for (j = 0; j < mag; ++j)
+          this_sz |= (*x++) << (j * 8);
+        sizes[i] = this_sz;
+      }
+      *count = frames;
+    }
+  }
+  return VPX_CODEC_OK;
+}
+#endif
+
+// Example pattern for spatial layers and 2 temporal layers used in the
+// bypass/flexible mode. The pattern corresponds to the pattern
+// VP9E_TEMPORAL_LAYERING_MODE_0101 (temporal_layering_mode == 2) used in
+// non-flexible mode.
+void set_frame_flags_bypass_mode(int sl, int tl, int num_spatial_layers,
+                                 int is_key_frame,
+                                 vpx_svc_ref_frame_config_t *ref_frame_config) {
+  for (sl = 0; sl < num_spatial_layers; ++sl) {
+    if (!tl) {
+      if (!sl) {
+        ref_frame_config->frame_flags[sl] = VP8_EFLAG_NO_REF_GF |
+                                            VP8_EFLAG_NO_REF_ARF |
+                                            VP8_EFLAG_NO_UPD_GF |
+                                            VP8_EFLAG_NO_UPD_ARF;
+      } else {
+        if (is_key_frame) {
+          ref_frame_config->frame_flags[sl] = VP8_EFLAG_NO_REF_LAST |
+                                              VP8_EFLAG_NO_REF_ARF |
+                                              VP8_EFLAG_NO_UPD_GF |
+                                              VP8_EFLAG_NO_UPD_ARF;
+        } else {
+        ref_frame_config->frame_flags[sl] = VP8_EFLAG_NO_REF_ARF |
+                                            VP8_EFLAG_NO_UPD_GF |
+                                            VP8_EFLAG_NO_UPD_ARF;
+        }
+      }
+    } else if (tl == 1) {
+      if (!sl) {
+        ref_frame_config->frame_flags[sl] = VP8_EFLAG_NO_REF_GF |
+                                            VP8_EFLAG_NO_REF_ARF |
+                                            VP8_EFLAG_NO_UPD_LAST |
+                                            VP8_EFLAG_NO_UPD_GF;
+      } else {
+        ref_frame_config->frame_flags[sl] = VP8_EFLAG_NO_REF_ARF |
+                                            VP8_EFLAG_NO_UPD_LAST |
+                                            VP8_EFLAG_NO_UPD_GF;
+      }
+    }
+    if (tl == 0) {
+      ref_frame_config->lst_fb_idx[sl] = sl;
+      if (sl)
+        ref_frame_config->gld_fb_idx[sl] = sl - 1;
+      else
+        ref_frame_config->gld_fb_idx[sl] = 0;
+      ref_frame_config->alt_fb_idx[sl] = 0;
+    } else if (tl == 1) {
+      ref_frame_config->lst_fb_idx[sl] = sl;
+      ref_frame_config->gld_fb_idx[sl] = num_spatial_layers + sl - 1;
+      ref_frame_config->alt_fb_idx[sl] = num_spatial_layers + sl;
+    }
+  }
+}
+
+int main(int argc, const char **argv) {
+  AppInput app_input = {0};
+  VpxVideoWriter *writer = NULL;
+  VpxVideoInfo info = {0};
+  vpx_codec_ctx_t codec;
+  vpx_codec_enc_cfg_t enc_cfg;
+  SvcContext svc_ctx;
+  uint32_t i;
+  uint32_t frame_cnt = 0;
+  vpx_image_t raw;
+  vpx_codec_err_t res;
+  int pts = 0;            /* PTS starts at 0 */
+  int frame_duration = 1; /* 1 timebase tick per frame */
+  FILE *infile = NULL;
+  int end_of_stream = 0;
+  int frames_received = 0;
+#if OUTPUT_RC_STATS
+  VpxVideoWriter *outfile[VPX_TS_MAX_LAYERS] = {NULL};
+  struct RateControlStats rc;
+  vpx_svc_layer_id_t layer_id;
+  vpx_svc_ref_frame_config_t ref_frame_config;
+  int sl, tl;
+  double sum_bitrate = 0.0;
+  double sum_bitrate2 = 0.0;
+  double framerate  = 30.0;
+#endif
+  struct vpx_usec_timer timer;
+  int64_t cx_time = 0;
+  memset(&svc_ctx, 0, sizeof(svc_ctx));
+  svc_ctx.log_print = 1;
+  exec_name = argv[0];
+  parse_command_line(argc, argv, &app_input, &svc_ctx, &enc_cfg);
+
+  // Allocate image buffer
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (!vpx_img_alloc(&raw, enc_cfg.g_input_bit_depth == 8 ?
+                         VPX_IMG_FMT_I420 : VPX_IMG_FMT_I42016,
+                     enc_cfg.g_w, enc_cfg.g_h, 32)) {
+    die("Failed to allocate image %dx%d\n", enc_cfg.g_w, enc_cfg.g_h);
+  }
+#else
+  if (!vpx_img_alloc(&raw, VPX_IMG_FMT_I420, enc_cfg.g_w, enc_cfg.g_h, 32)) {
+    die("Failed to allocate image %dx%d\n", enc_cfg.g_w, enc_cfg.g_h);
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  if (!(infile = fopen(app_input.input_filename, "rb")))
+    die("Failed to open %s for reading\n", app_input.input_filename);
+
+  // Initialize codec
+  if (vpx_svc_init(&svc_ctx, &codec, vpx_codec_vp9_cx(), &enc_cfg) !=
+      VPX_CODEC_OK)
+    die("Failed to initialize encoder\n");
+
+#if OUTPUT_RC_STATS
+  if (svc_ctx.output_rc_stat) {
+    set_rate_control_stats(&rc, &enc_cfg);
+    framerate = enc_cfg.g_timebase.den / enc_cfg.g_timebase.num;
+  }
+#endif
+
+  info.codec_fourcc = VP9_FOURCC;
+  info.time_base.numerator = enc_cfg.g_timebase.num;
+  info.time_base.denominator = enc_cfg.g_timebase.den;
+
+  if (!(app_input.passes == 2 && app_input.pass == 1)) {
+    // We don't save the bitstream for the 1st pass on two pass rate control
+    writer = vpx_video_writer_open(app_input.output_filename, kContainerIVF,
+                                   &info);
+    if (!writer)
+      die("Failed to open %s for writing\n", app_input.output_filename);
+  }
+#if OUTPUT_RC_STATS
+  // For now, just write temporal layer streams.
+  // TODO(wonkap): do spatial by re-writing superframe.
+  if (svc_ctx.output_rc_stat) {
+    for (tl = 0; tl < enc_cfg.ts_number_layers; ++tl) {
+      char file_name[PATH_MAX];
+
+      snprintf(file_name, sizeof(file_name), "%s_t%d.ivf",
+               app_input.output_filename, tl);
+      outfile[tl] = vpx_video_writer_open(file_name, kContainerIVF, &info);
+      if (!outfile[tl])
+        die("Failed to open %s for writing", file_name);
+    }
+  }
+#endif
+
+  // skip initial frames
+  for (i = 0; i < app_input.frames_to_skip; ++i)
+    vpx_img_read(&raw, infile);
+
+  if (svc_ctx.speed != -1)
+    vpx_codec_control(&codec, VP8E_SET_CPUUSED, svc_ctx.speed);
+  if (svc_ctx.threads)
+    vpx_codec_control(&codec, VP9E_SET_TILE_COLUMNS, (svc_ctx.threads >> 1));
+  if (svc_ctx.speed >= 5 && svc_ctx.aqmode == 1)
+    vpx_codec_control(&codec, VP9E_SET_AQ_MODE, 3);
+
+
+  // Encode frames
+  while (!end_of_stream) {
+    vpx_codec_iter_t iter = NULL;
+    const vpx_codec_cx_pkt_t *cx_pkt;
+    if (frame_cnt >= app_input.frames_to_code || !vpx_img_read(&raw, infile)) {
+      // We need one extra vpx_svc_encode call at end of stream to flush
+      // encoder and get remaining data
+      end_of_stream = 1;
+    }
+
+    // For BYPASS/FLEXIBLE mode, set the frame flags (reference and updates)
+    // and the buffer indices for each spatial layer of the current
+    // (super)frame to be encoded. The temporal layer_id for the current frame
+    // also needs to be set.
+    // TODO(marpan): Should rename the "VP9E_TEMPORAL_LAYERING_MODE_BYPASS"
+    // mode to "VP9E_LAYERING_MODE_BYPASS".
+    if (svc_ctx.temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_BYPASS) {
+      layer_id.spatial_layer_id = 0;
+      // Example for 2 temporal layers.
+      if (frame_cnt % 2 == 0)
+        layer_id.temporal_layer_id = 0;
+      else
+        layer_id.temporal_layer_id = 1;
+      // Note that we only set the temporal layer_id, since we are calling
+      // the encode for the whole superframe. The encoder will internally loop
+      // over all the spatial layers for the current superframe.
+      vpx_codec_control(&codec, VP9E_SET_SVC_LAYER_ID, &layer_id);
+      set_frame_flags_bypass_mode(sl, layer_id.temporal_layer_id,
+                                  svc_ctx.spatial_layers,
+                                  frame_cnt == 0,
+                                  &ref_frame_config);
+      vpx_codec_control(&codec, VP9E_SET_SVC_REF_FRAME_CONFIG,
+                        &ref_frame_config);
+      // Keep track of input frames, to account for frame drops in rate control
+      // stats/metrics.
+      for (sl = 0; sl < enc_cfg.ss_number_layers; ++sl) {
+        ++rc.layer_input_frames[sl * enc_cfg.ts_number_layers +
+                                layer_id.temporal_layer_id];
+      }
+    }
+
+    vpx_usec_timer_start(&timer);
+    res = vpx_svc_encode(&svc_ctx, &codec, (end_of_stream ? NULL : &raw),
+                         pts, frame_duration, svc_ctx.speed >= 5 ?
+                         VPX_DL_REALTIME : VPX_DL_GOOD_QUALITY);
+    vpx_usec_timer_mark(&timer);
+    cx_time += vpx_usec_timer_elapsed(&timer);
+
+    printf("%s", vpx_svc_get_message(&svc_ctx));
+    fflush(stdout);
+    if (res != VPX_CODEC_OK) {
+      die_codec(&codec, "Failed to encode frame");
+    }
+
+    while ((cx_pkt = vpx_codec_get_cx_data(&codec, &iter)) != NULL) {
+      switch (cx_pkt->kind) {
+        case VPX_CODEC_CX_FRAME_PKT: {
+          SvcInternal_t *const si = (SvcInternal_t *)svc_ctx.internal;
+          if (cx_pkt->data.frame.sz > 0) {
+#if OUTPUT_RC_STATS
+            uint32_t sizes[8];
+            int count = 0;
+#endif
+            vpx_video_writer_write_frame(writer,
+                                         cx_pkt->data.frame.buf,
+                                         cx_pkt->data.frame.sz,
+                                         cx_pkt->data.frame.pts);
+#if OUTPUT_RC_STATS
+            // TODO(marpan/wonkap): Put this (to line728) in separate function.
+            if (svc_ctx.output_rc_stat) {
+              vpx_codec_control(&codec, VP9E_GET_SVC_LAYER_ID, &layer_id);
+              parse_superframe_index(cx_pkt->data.frame.buf,
+                                     cx_pkt->data.frame.sz, sizes, &count);
+              // Note computing input_layer_frames here won't account for frame
+              // drops in rate control stats.
+              // TODO(marpan): Fix this for non-bypass mode so we can get stats
+              // for dropped frames.
+              if (svc_ctx.temporal_layering_mode !=
+                  VP9E_TEMPORAL_LAYERING_MODE_BYPASS) {
+                for (sl = 0; sl < enc_cfg.ss_number_layers; ++sl) {
+                  ++rc.layer_input_frames[sl * enc_cfg.ts_number_layers +
+                                         layer_id.temporal_layer_id];
+                }
+              }
+              for (tl = layer_id.temporal_layer_id;
+                  tl < enc_cfg.ts_number_layers; ++tl) {
+                vpx_video_writer_write_frame(outfile[tl],
+                                             cx_pkt->data.frame.buf,
+                                             cx_pkt->data.frame.sz,
+                                             cx_pkt->data.frame.pts);
+              }
+
+              for (sl = 0; sl < enc_cfg.ss_number_layers; ++sl) {
+                for (tl = layer_id.temporal_layer_id;
+                    tl < enc_cfg.ts_number_layers; ++tl) {
+                  const int layer = sl * enc_cfg.ts_number_layers + tl;
+                  ++rc.layer_tot_enc_frames[layer];
+                  rc.layer_encoding_bitrate[layer] += 8.0 * sizes[sl];
+                  // Keep count of rate control stats per layer, for non-key
+                  // frames.
+                  if (tl == layer_id.temporal_layer_id &&
+                      !(cx_pkt->data.frame.flags & VPX_FRAME_IS_KEY)) {
+                    rc.layer_avg_frame_size[layer] += 8.0 * sizes[sl];
+                    rc.layer_avg_rate_mismatch[layer] +=
+                        fabs(8.0 * sizes[sl] - rc.layer_pfb[layer]) /
+                        rc.layer_pfb[layer];
+                    ++rc.layer_enc_frames[layer];
+                  }
+                }
+              }
+
+              // Update for short-time encoding bitrate states, for moving
+              // window of size rc->window, shifted by rc->window / 2.
+              // Ignore first window segment, due to key frame.
+              if (frame_cnt > rc.window_size) {
+                tl = layer_id.temporal_layer_id;
+                for (sl = 0; sl < enc_cfg.ss_number_layers; ++sl) {
+                  sum_bitrate += 0.001 * 8.0 * sizes[sl] * framerate;
+                }
+                if (frame_cnt % rc.window_size == 0) {
+                  rc.window_count += 1;
+                  rc.avg_st_encoding_bitrate += sum_bitrate / rc.window_size;
+                  rc.variance_st_encoding_bitrate +=
+                      (sum_bitrate / rc.window_size) *
+                      (sum_bitrate / rc.window_size);
+                  sum_bitrate = 0.0;
+                }
+              }
+
+              // Second shifted window.
+              if (frame_cnt > rc.window_size + rc.window_size / 2) {
+               tl = layer_id.temporal_layer_id;
+               for (sl = 0; sl < enc_cfg.ss_number_layers; ++sl) {
+                 sum_bitrate2 += 0.001 * 8.0 * sizes[sl] * framerate;
+               }
+
+               if (frame_cnt > 2 * rc.window_size &&
+                  frame_cnt % rc.window_size == 0) {
+                 rc.window_count += 1;
+                 rc.avg_st_encoding_bitrate += sum_bitrate2 / rc.window_size;
+                 rc.variance_st_encoding_bitrate +=
+                    (sum_bitrate2 / rc.window_size) *
+                    (sum_bitrate2 / rc.window_size);
+                 sum_bitrate2 = 0.0;
+               }
+              }
+            }
+#endif
+          }
+
+          printf("SVC frame: %d, kf: %d, size: %d, pts: %d\n", frames_received,
+                 !!(cx_pkt->data.frame.flags & VPX_FRAME_IS_KEY),
+                 (int)cx_pkt->data.frame.sz, (int)cx_pkt->data.frame.pts);
+          if (enc_cfg.ss_number_layers == 1 && enc_cfg.ts_number_layers == 1)
+            si->bytes_sum[0] += (int)cx_pkt->data.frame.sz;
+          ++frames_received;
+          break;
+        }
+        case VPX_CODEC_STATS_PKT: {
+          stats_write(&app_input.rc_stats,
+                      cx_pkt->data.twopass_stats.buf,
+                      cx_pkt->data.twopass_stats.sz);
+          break;
+        }
+        default: {
+          break;
+        }
+      }
+    }
+
+    if (!end_of_stream) {
+      ++frame_cnt;
+      pts += frame_duration;
+    }
+  }
+
+  // Compensate for the extra frame count for the bypass mode.
+  if (svc_ctx.temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_BYPASS) {
+    for (sl = 0; sl < enc_cfg.ss_number_layers; ++sl) {
+      const int layer = sl * enc_cfg.ts_number_layers +
+          layer_id.temporal_layer_id;
+      --rc.layer_input_frames[layer];
+    }
+  }
+
+  printf("Processed %d frames\n", frame_cnt);
+  fclose(infile);
+#if OUTPUT_RC_STATS
+  if (svc_ctx.output_rc_stat) {
+    printout_rate_control_summary(&rc, &enc_cfg, frame_cnt);
+    printf("\n");
+  }
+#endif
+  if (vpx_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec");
+  if (app_input.passes == 2)
+    stats_close(&app_input.rc_stats, 1);
+  if (writer) {
+    vpx_video_writer_close(writer);
+  }
+#if OUTPUT_RC_STATS
+  if (svc_ctx.output_rc_stat) {
+    for (tl = 0; tl < enc_cfg.ts_number_layers; ++tl) {
+      vpx_video_writer_close(outfile[tl]);
+    }
+  }
+#endif
+  printf("Frame cnt and encoding time/FPS stats for encoding: %d %f %f \n",
+         frame_cnt,
+         1000 * (float)cx_time / (double)(frame_cnt * 1000000),
+         1000000 * (double)frame_cnt / (double)cx_time);
+  vpx_img_free(&raw);
+  // display average size, psnr
+  printf("%s", vpx_svc_dump_statistics(&svc_ctx));
+  vpx_svc_release(&svc_ctx);
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/examples/vpx_temporal_svc_encoder.c b/libvpx/libvpx/examples/vpx_temporal_svc_encoder.c
new file mode 100644
index 0000000..e6c09fb
--- /dev/null
+++ b/libvpx/libvpx/examples/vpx_temporal_svc_encoder.c
@@ -0,0 +1,852 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+//  This is an example demonstrating how to implement a multi-layer VPx
+//  encoding scheme based on temporal scalability for video applications
+//  that benefit from a scalable bitstream.
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "./vpx_config.h"
+#include "../vpx_ports/vpx_timer.h"
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  exit(EXIT_FAILURE);
+}
+
+// Denoiser states, for temporal denoising.
+enum denoiserState {
+  kDenoiserOff,
+  kDenoiserOnYOnly,
+  kDenoiserOnYUV,
+  kDenoiserOnYUVAggressive,
+  kDenoiserOnAdaptive
+};
+
+static int mode_to_num_layers[13] = {1, 2, 2, 3, 3, 3, 3, 5, 2, 3, 3, 3, 3};
+
+// For rate control encoding stats.
+struct RateControlMetrics {
+  // Number of input frames per layer.
+  int layer_input_frames[VPX_TS_MAX_LAYERS];
+  // Total (cumulative) number of encoded frames per layer.
+  int layer_tot_enc_frames[VPX_TS_MAX_LAYERS];
+  // Number of encoded non-key frames per layer.
+  int layer_enc_frames[VPX_TS_MAX_LAYERS];
+  // Framerate per layer layer (cumulative).
+  double layer_framerate[VPX_TS_MAX_LAYERS];
+  // Target average frame size per layer (per-frame-bandwidth per layer).
+  double layer_pfb[VPX_TS_MAX_LAYERS];
+  // Actual average frame size per layer.
+  double layer_avg_frame_size[VPX_TS_MAX_LAYERS];
+  // Average rate mismatch per layer (|target - actual| / target).
+  double layer_avg_rate_mismatch[VPX_TS_MAX_LAYERS];
+  // Actual encoding bitrate per layer (cumulative).
+  double layer_encoding_bitrate[VPX_TS_MAX_LAYERS];
+  // Average of the short-time encoder actual bitrate.
+  // TODO(marpan): Should we add these short-time stats for each layer?
+  double avg_st_encoding_bitrate;
+  // Variance of the short-time encoder actual bitrate.
+  double variance_st_encoding_bitrate;
+  // Window (number of frames) for computing short-timee encoding bitrate.
+  int window_size;
+  // Number of window measurements.
+  int window_count;
+  int layer_target_bitrate[VPX_MAX_LAYERS];
+};
+
+// Note: these rate control metrics assume only 1 key frame in the
+// sequence (i.e., first frame only). So for temporal pattern# 7
+// (which has key frame for every frame on base layer), the metrics
+// computation will be off/wrong.
+// TODO(marpan): Update these metrics to account for multiple key frames
+// in the stream.
+static void set_rate_control_metrics(struct RateControlMetrics *rc,
+                                     vpx_codec_enc_cfg_t *cfg) {
+  unsigned int i = 0;
+  // Set the layer (cumulative) framerate and the target layer (non-cumulative)
+  // per-frame-bandwidth, for the rate control encoding stats below.
+  const double framerate = cfg->g_timebase.den / cfg->g_timebase.num;
+  rc->layer_framerate[0] = framerate / cfg->ts_rate_decimator[0];
+  rc->layer_pfb[0] = 1000.0 * rc->layer_target_bitrate[0] /
+      rc->layer_framerate[0];
+  for (i = 0; i < cfg->ts_number_layers; ++i) {
+    if (i > 0) {
+      rc->layer_framerate[i] = framerate / cfg->ts_rate_decimator[i];
+      rc->layer_pfb[i] = 1000.0 *
+          (rc->layer_target_bitrate[i] - rc->layer_target_bitrate[i - 1]) /
+          (rc->layer_framerate[i] - rc->layer_framerate[i - 1]);
+    }
+    rc->layer_input_frames[i] = 0;
+    rc->layer_enc_frames[i] = 0;
+    rc->layer_tot_enc_frames[i] = 0;
+    rc->layer_encoding_bitrate[i] = 0.0;
+    rc->layer_avg_frame_size[i] = 0.0;
+    rc->layer_avg_rate_mismatch[i] = 0.0;
+  }
+  rc->window_count = 0;
+  rc->window_size = 15;
+  rc->avg_st_encoding_bitrate = 0.0;
+  rc->variance_st_encoding_bitrate = 0.0;
+}
+
+static void printout_rate_control_summary(struct RateControlMetrics *rc,
+                                          vpx_codec_enc_cfg_t *cfg,
+                                          int frame_cnt) {
+  unsigned int i = 0;
+  int tot_num_frames = 0;
+  double perc_fluctuation = 0.0;
+  printf("Total number of processed frames: %d\n\n", frame_cnt -1);
+  printf("Rate control layer stats for %d layer(s):\n\n",
+      cfg->ts_number_layers);
+  for (i = 0; i < cfg->ts_number_layers; ++i) {
+    const int num_dropped = (i > 0) ?
+        (rc->layer_input_frames[i] - rc->layer_enc_frames[i]) :
+        (rc->layer_input_frames[i] - rc->layer_enc_frames[i] - 1);
+    tot_num_frames += rc->layer_input_frames[i];
+    rc->layer_encoding_bitrate[i] = 0.001 * rc->layer_framerate[i] *
+        rc->layer_encoding_bitrate[i] / tot_num_frames;
+    rc->layer_avg_frame_size[i] = rc->layer_avg_frame_size[i] /
+        rc->layer_enc_frames[i];
+    rc->layer_avg_rate_mismatch[i] = 100.0 * rc->layer_avg_rate_mismatch[i] /
+        rc->layer_enc_frames[i];
+    printf("For layer#: %d \n", i);
+    printf("Bitrate (target vs actual): %d %f \n", rc->layer_target_bitrate[i],
+           rc->layer_encoding_bitrate[i]);
+    printf("Average frame size (target vs actual): %f %f \n", rc->layer_pfb[i],
+           rc->layer_avg_frame_size[i]);
+    printf("Average rate_mismatch: %f \n", rc->layer_avg_rate_mismatch[i]);
+    printf("Number of input frames, encoded (non-key) frames, "
+        "and perc dropped frames: %d %d %f \n", rc->layer_input_frames[i],
+        rc->layer_enc_frames[i],
+        100.0 * num_dropped / rc->layer_input_frames[i]);
+    printf("\n");
+  }
+  rc->avg_st_encoding_bitrate = rc->avg_st_encoding_bitrate / rc->window_count;
+  rc->variance_st_encoding_bitrate =
+      rc->variance_st_encoding_bitrate / rc->window_count -
+      (rc->avg_st_encoding_bitrate * rc->avg_st_encoding_bitrate);
+  perc_fluctuation = 100.0 * sqrt(rc->variance_st_encoding_bitrate) /
+      rc->avg_st_encoding_bitrate;
+  printf("Short-time stats, for window of %d frames: \n",rc->window_size);
+  printf("Average, rms-variance, and percent-fluct: %f %f %f \n",
+         rc->avg_st_encoding_bitrate,
+         sqrt(rc->variance_st_encoding_bitrate),
+         perc_fluctuation);
+  if ((frame_cnt - 1) != tot_num_frames)
+    die("Error: Number of input frames not equal to output! \n");
+}
+
+// Temporal scaling parameters:
+// NOTE: The 3 prediction frames cannot be used interchangeably due to
+// differences in the way they are handled throughout the code. The
+// frames should be allocated to layers in the order LAST, GF, ARF.
+// Other combinations work, but may produce slightly inferior results.
+static void set_temporal_layer_pattern(int layering_mode,
+                                       vpx_codec_enc_cfg_t *cfg,
+                                       int *layer_flags,
+                                       int *flag_periodicity) {
+  switch (layering_mode) {
+    case 0: {
+      // 1-layer.
+      int ids[1] = {0};
+      cfg->ts_periodicity = 1;
+      *flag_periodicity = 1;
+      cfg->ts_number_layers = 1;
+      cfg->ts_rate_decimator[0] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // Update L only.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF  | VP8_EFLAG_NO_UPD_GF |
+          VP8_EFLAG_NO_UPD_ARF;
+      break;
+    }
+    case 1: {
+      // 2-layers, 2-frame period.
+      int ids[2] = {0, 1};
+      cfg->ts_periodicity = 2;
+      *flag_periodicity = 2;
+      cfg->ts_number_layers = 2;
+      cfg->ts_rate_decimator[0] = 2;
+      cfg->ts_rate_decimator[1] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+#if 1
+      // 0=L, 1=GF, Intra-layer prediction enabled.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF  | VP8_EFLAG_NO_UPD_GF |
+          VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF;
+      layer_flags[1] = VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST |
+          VP8_EFLAG_NO_REF_ARF;
+#else
+       // 0=L, 1=GF, Intra-layer prediction disabled.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF  | VP8_EFLAG_NO_UPD_GF |
+          VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF;
+      layer_flags[1] = VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST |
+          VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_REF_LAST;
+#endif
+      break;
+    }
+    case 2: {
+      // 2-layers, 3-frame period.
+      int ids[3] = {0, 1, 1};
+      cfg->ts_periodicity = 3;
+      *flag_periodicity = 3;
+      cfg->ts_number_layers = 2;
+      cfg->ts_rate_decimator[0] = 3;
+      cfg->ts_rate_decimator[1] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // 0=L, 1=GF, Intra-layer prediction enabled.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF  | VP8_EFLAG_NO_REF_GF |
+          VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[1] =
+      layer_flags[2] = VP8_EFLAG_NO_REF_GF  | VP8_EFLAG_NO_REF_ARF |
+          VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST;
+      break;
+    }
+    case 3: {
+      // 3-layers, 6-frame period.
+      int ids[6] = {0, 2, 2, 1, 2, 2};
+      cfg->ts_periodicity = 6;
+      *flag_periodicity = 6;
+      cfg->ts_number_layers = 3;
+      cfg->ts_rate_decimator[0] = 6;
+      cfg->ts_rate_decimator[1] = 3;
+      cfg->ts_rate_decimator[2] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // 0=L, 1=GF, 2=ARF, Intra-layer prediction enabled.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF  | VP8_EFLAG_NO_REF_GF |
+          VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[3] = VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_ARF |
+          VP8_EFLAG_NO_UPD_LAST;
+      layer_flags[1] =
+      layer_flags[2] =
+      layer_flags[4] =
+      layer_flags[5] = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_LAST;
+      break;
+    }
+    case 4: {
+      // 3-layers, 4-frame period.
+      int ids[4] = {0, 2, 1, 2};
+      cfg->ts_periodicity = 4;
+      *flag_periodicity = 4;
+      cfg->ts_number_layers = 3;
+      cfg->ts_rate_decimator[0] = 4;
+      cfg->ts_rate_decimator[1] = 2;
+      cfg->ts_rate_decimator[2] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // 0=L, 1=GF, 2=ARF, Intra-layer prediction disabled.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF  | VP8_EFLAG_NO_REF_GF |
+          VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[2] = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF |
+          VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST;
+      layer_flags[1] =
+      layer_flags[3] = VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_LAST |
+          VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+      break;
+    }
+    case 5: {
+      // 3-layers, 4-frame period.
+      int ids[4] = {0, 2, 1, 2};
+      cfg->ts_periodicity = 4;
+      *flag_periodicity = 4;
+      cfg->ts_number_layers     = 3;
+      cfg->ts_rate_decimator[0] = 4;
+      cfg->ts_rate_decimator[1] = 2;
+      cfg->ts_rate_decimator[2] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // 0=L, 1=GF, 2=ARF, Intra-layer prediction enabled in layer 1, disabled
+      // in layer 2.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF  | VP8_EFLAG_NO_REF_GF |
+          VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[2] = VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_LAST |
+          VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[1] =
+      layer_flags[3] = VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_LAST |
+          VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+      break;
+    }
+    case 6: {
+      // 3-layers, 4-frame period.
+      int ids[4] = {0, 2, 1, 2};
+      cfg->ts_periodicity = 4;
+      *flag_periodicity = 4;
+      cfg->ts_number_layers = 3;
+      cfg->ts_rate_decimator[0] = 4;
+      cfg->ts_rate_decimator[1] = 2;
+      cfg->ts_rate_decimator[2] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // 0=L, 1=GF, 2=ARF, Intra-layer prediction enabled.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF  | VP8_EFLAG_NO_REF_GF |
+          VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[2] = VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_LAST |
+          VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[1] =
+      layer_flags[3] = VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF;
+      break;
+    }
+    case 7: {
+      // NOTE: Probably of academic interest only.
+      // 5-layers, 16-frame period.
+      int ids[16] = {0, 4, 3, 4, 2, 4, 3, 4, 1, 4, 3, 4, 2, 4, 3, 4};
+      cfg->ts_periodicity = 16;
+      *flag_periodicity = 16;
+      cfg->ts_number_layers = 5;
+      cfg->ts_rate_decimator[0] = 16;
+      cfg->ts_rate_decimator[1] = 8;
+      cfg->ts_rate_decimator[2] = 4;
+      cfg->ts_rate_decimator[3] = 2;
+      cfg->ts_rate_decimator[4] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      layer_flags[0]  = VPX_EFLAG_FORCE_KF;
+      layer_flags[1]  =
+      layer_flags[3]  =
+      layer_flags[5]  =
+      layer_flags[7]  =
+      layer_flags[9]  =
+      layer_flags[11] =
+      layer_flags[13] =
+      layer_flags[15] = VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF |
+          VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[2]  =
+      layer_flags[6]  =
+      layer_flags[10] =
+      layer_flags[14] = VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_GF;
+      layer_flags[4] =
+      layer_flags[12] = VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[8]  = VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF;
+      break;
+    }
+    case 8: {
+      // 2-layers, with sync point at first frame of layer 1.
+      int ids[2] = {0, 1};
+      cfg->ts_periodicity = 2;
+      *flag_periodicity = 8;
+      cfg->ts_number_layers = 2;
+      cfg->ts_rate_decimator[0] = 2;
+      cfg->ts_rate_decimator[1] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // 0=L, 1=GF.
+      // ARF is used as predictor for all frames, and is only updated on
+      // key frame. Sync point every 8 frames.
+
+      // Layer 0: predict from L and ARF, update L and G.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF  | VP8_EFLAG_NO_REF_GF |
+          VP8_EFLAG_NO_UPD_ARF;
+      // Layer 1: sync point: predict from L and ARF, and update G.
+      layer_flags[1] = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_UPD_LAST |
+          VP8_EFLAG_NO_UPD_ARF;
+      // Layer 0, predict from L and ARF, update L.
+      layer_flags[2] = VP8_EFLAG_NO_REF_GF  | VP8_EFLAG_NO_UPD_GF |
+          VP8_EFLAG_NO_UPD_ARF;
+      // Layer 1: predict from L, G and ARF, and update G.
+      layer_flags[3] = VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST |
+          VP8_EFLAG_NO_UPD_ENTROPY;
+      // Layer 0.
+      layer_flags[4] = layer_flags[2];
+      // Layer 1.
+      layer_flags[5] = layer_flags[3];
+      // Layer 0.
+      layer_flags[6] = layer_flags[4];
+      // Layer 1.
+      layer_flags[7] = layer_flags[5];
+     break;
+    }
+    case 9: {
+      // 3-layers: Sync points for layer 1 and 2 every 8 frames.
+      int ids[4] = {0, 2, 1, 2};
+      cfg->ts_periodicity = 4;
+      *flag_periodicity = 8;
+      cfg->ts_number_layers = 3;
+      cfg->ts_rate_decimator[0] = 4;
+      cfg->ts_rate_decimator[1] = 2;
+      cfg->ts_rate_decimator[2] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // 0=L, 1=GF, 2=ARF.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF  | VP8_EFLAG_NO_REF_GF |
+          VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[1] = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF |
+          VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF;
+      layer_flags[2] = VP8_EFLAG_NO_REF_GF   | VP8_EFLAG_NO_REF_ARF |
+          VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[3] =
+      layer_flags[5] = VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF;
+      layer_flags[4] = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF |
+          VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[6] = VP8_EFLAG_NO_REF_ARF | VP8_EFLAG_NO_UPD_LAST |
+          VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[7] = VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF |
+          VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_ENTROPY;
+      break;
+    }
+    case 10: {
+      // 3-layers structure where ARF is used as predictor for all frames,
+      // and is only updated on key frame.
+      // Sync points for layer 1 and 2 every 8 frames.
+
+      int ids[4] = {0, 2, 1, 2};
+      cfg->ts_periodicity = 4;
+      *flag_periodicity = 8;
+      cfg->ts_number_layers = 3;
+      cfg->ts_rate_decimator[0] = 4;
+      cfg->ts_rate_decimator[1] = 2;
+      cfg->ts_rate_decimator[2] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // 0=L, 1=GF, 2=ARF.
+      // Layer 0: predict from L and ARF; update L and G.
+      layer_flags[0] = VPX_EFLAG_FORCE_KF | VP8_EFLAG_NO_UPD_ARF |
+          VP8_EFLAG_NO_REF_GF;
+      // Layer 2: sync point: predict from L and ARF; update none.
+      layer_flags[1] = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_UPD_GF |
+          VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST |
+          VP8_EFLAG_NO_UPD_ENTROPY;
+      // Layer 1: sync point: predict from L and ARF; update G.
+      layer_flags[2] = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_UPD_ARF |
+          VP8_EFLAG_NO_UPD_LAST;
+      // Layer 2: predict from L, G, ARF; update none.
+      layer_flags[3] = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF |
+          VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_ENTROPY;
+      // Layer 0: predict from L and ARF; update L.
+      layer_flags[4] = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF |
+          VP8_EFLAG_NO_REF_GF;
+      // Layer 2: predict from L, G, ARF; update none.
+      layer_flags[5] = layer_flags[3];
+      // Layer 1: predict from L, G, ARF; update G.
+      layer_flags[6] = VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST;
+      // Layer 2: predict from L, G, ARF; update none.
+      layer_flags[7] = layer_flags[3];
+      break;
+    }
+    case 11: {
+      // 3-layers structure with one reference frame.
+      // This works same as temporal_layering_mode 3.
+      // This was added to compare with vp9_spatial_svc_encoder.
+
+      // 3-layers, 4-frame period.
+      int ids[4] = {0, 2, 1, 2};
+      cfg->ts_periodicity = 4;
+      *flag_periodicity = 4;
+      cfg->ts_number_layers = 3;
+      cfg->ts_rate_decimator[0] = 4;
+      cfg->ts_rate_decimator[1] = 2;
+      cfg->ts_rate_decimator[2] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // 0=L, 1=GF, 2=ARF, Intra-layer prediction disabled.
+      layer_flags[0] = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF |
+          VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
+      layer_flags[2] = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF |
+          VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST;
+      layer_flags[1] = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF |
+          VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF;
+      layer_flags[3] = VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_ARF |
+          VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF;
+      break;
+    }
+    case 12:
+    default: {
+      // 3-layers structure as in case 10, but no sync/refresh points for
+      // layer 1 and 2.
+      int ids[4] = {0, 2, 1, 2};
+      cfg->ts_periodicity = 4;
+      *flag_periodicity = 8;
+      cfg->ts_number_layers = 3;
+      cfg->ts_rate_decimator[0] = 4;
+      cfg->ts_rate_decimator[1] = 2;
+      cfg->ts_rate_decimator[2] = 1;
+      memcpy(cfg->ts_layer_id, ids, sizeof(ids));
+      // 0=L, 1=GF, 2=ARF.
+      // Layer 0: predict from L and ARF; update L.
+      layer_flags[0] = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF |
+          VP8_EFLAG_NO_REF_GF;
+      layer_flags[4] = layer_flags[0];
+      // Layer 1: predict from L, G, ARF; update G.
+      layer_flags[2] = VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST;
+      layer_flags[6] = layer_flags[2];
+      // Layer 2: predict from L, G, ARF; update none.
+      layer_flags[1] = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF |
+          VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_ENTROPY;
+      layer_flags[3] = layer_flags[1];
+      layer_flags[5] = layer_flags[1];
+      layer_flags[7] = layer_flags[1];
+      break;
+    }
+  }
+}
+
+int main(int argc, char **argv) {
+  VpxVideoWriter *outfile[VPX_TS_MAX_LAYERS] = {NULL};
+  vpx_codec_ctx_t codec;
+  vpx_codec_enc_cfg_t cfg;
+  int frame_cnt = 0;
+  vpx_image_t raw;
+  vpx_codec_err_t res;
+  unsigned int width;
+  unsigned int height;
+  int speed;
+  int frame_avail;
+  int got_data;
+  int flags = 0;
+  unsigned int i;
+  int pts = 0;  // PTS starts at 0.
+  int frame_duration = 1;  // 1 timebase tick per frame.
+  int layering_mode = 0;
+  int layer_flags[VPX_TS_MAX_PERIODICITY] = {0};
+  int flag_periodicity = 1;
+#if VPX_ENCODER_ABI_VERSION > (4 + VPX_CODEC_ABI_VERSION)
+  vpx_svc_layer_id_t layer_id = {0, 0};
+#else
+  vpx_svc_layer_id_t layer_id = {0};
+#endif
+  const VpxInterface *encoder = NULL;
+  FILE *infile = NULL;
+  struct RateControlMetrics rc;
+  int64_t cx_time = 0;
+  const int min_args_base = 11;
+#if CONFIG_VP9_HIGHBITDEPTH
+  vpx_bit_depth_t bit_depth = VPX_BITS_8;
+  int input_bit_depth = 8;
+  const int min_args = min_args_base + 1;
+#else
+  const int min_args = min_args_base;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  double sum_bitrate = 0.0;
+  double sum_bitrate2 = 0.0;
+  double framerate  = 30.0;
+
+  exec_name = argv[0];
+  // Check usage and arguments.
+  if (argc < min_args) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    die("Usage: %s <infile> <outfile> <codec_type(vp8/vp9)> <width> <height> "
+        "<rate_num> <rate_den> <speed> <frame_drop_threshold> <mode> "
+        "<Rate_0> ... <Rate_nlayers-1> <bit-depth> \n", argv[0]);
+#else
+    die("Usage: %s <infile> <outfile> <codec_type(vp8/vp9)> <width> <height> "
+        "<rate_num> <rate_den> <speed> <frame_drop_threshold> <mode> "
+        "<Rate_0> ... <Rate_nlayers-1> \n", argv[0]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+
+  encoder = get_vpx_encoder_by_name(argv[3]);
+  if (!encoder)
+    die("Unsupported codec.");
+
+  printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
+
+  width = strtol(argv[4], NULL, 0);
+  height = strtol(argv[5], NULL, 0);
+  if (width < 16 || width % 2 || height < 16 || height % 2) {
+    die("Invalid resolution: %d x %d", width, height);
+  }
+
+  layering_mode = strtol(argv[10], NULL, 0);
+  if (layering_mode < 0 || layering_mode > 13) {
+    die("Invalid layering mode (0..12) %s", argv[10]);
+  }
+
+  if (argc != min_args + mode_to_num_layers[layering_mode]) {
+    die("Invalid number of arguments");
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (strtol(argv[argc-1], NULL, 0)) {
+    case 8:
+      bit_depth = VPX_BITS_8;
+      input_bit_depth = 8;
+      break;
+    case 10:
+      bit_depth = VPX_BITS_10;
+      input_bit_depth = 10;
+      break;
+    case 12:
+      bit_depth = VPX_BITS_12;
+      input_bit_depth = 12;
+      break;
+    default:
+      die("Invalid bit depth (8, 10, 12) %s", argv[argc-1]);
+  }
+  if (!vpx_img_alloc(&raw,
+                     bit_depth == VPX_BITS_8 ? VPX_IMG_FMT_I420 :
+                                               VPX_IMG_FMT_I42016,
+                     width, height, 32)) {
+    die("Failed to allocate image", width, height);
+  }
+#else
+  if (!vpx_img_alloc(&raw, VPX_IMG_FMT_I420, width, height, 32)) {
+    die("Failed to allocate image", width, height);
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  // Populate encoder configuration.
+  res = vpx_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res) {
+    printf("Failed to get config: %s\n", vpx_codec_err_to_string(res));
+    return EXIT_FAILURE;
+  }
+
+  // Update the default configuration with our settings.
+  cfg.g_w = width;
+  cfg.g_h = height;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (bit_depth != VPX_BITS_8) {
+    cfg.g_bit_depth = bit_depth;
+    cfg.g_input_bit_depth = input_bit_depth;
+    cfg.g_profile = 2;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  // Timebase format e.g. 30fps: numerator=1, demoninator = 30.
+  cfg.g_timebase.num = strtol(argv[6], NULL, 0);
+  cfg.g_timebase.den = strtol(argv[7], NULL, 0);
+
+  speed = strtol(argv[8], NULL, 0);
+  if (speed < 0) {
+    die("Invalid speed setting: must be positive");
+  }
+
+  for (i = min_args_base;
+       (int)i < min_args_base + mode_to_num_layers[layering_mode];
+       ++i) {
+    rc.layer_target_bitrate[i - 11] = strtol(argv[i], NULL, 0);
+    if (strncmp(encoder->name, "vp8", 3) == 0)
+      cfg.ts_target_bitrate[i - 11] = rc.layer_target_bitrate[i - 11];
+    else if (strncmp(encoder->name, "vp9", 3) == 0)
+      cfg.layer_target_bitrate[i - 11] = rc.layer_target_bitrate[i - 11];
+  }
+
+  // Real time parameters.
+  cfg.rc_dropframe_thresh = strtol(argv[9], NULL, 0);
+  cfg.rc_end_usage = VPX_CBR;
+  cfg.rc_min_quantizer = 2;
+  cfg.rc_max_quantizer = 56;
+  if (strncmp(encoder->name, "vp9", 3) == 0)
+    cfg.rc_max_quantizer = 52;
+  cfg.rc_undershoot_pct = 50;
+  cfg.rc_overshoot_pct = 50;
+  cfg.rc_buf_initial_sz = 500;
+  cfg.rc_buf_optimal_sz = 600;
+  cfg.rc_buf_sz = 1000;
+
+  // Disable dynamic resizing by default.
+  cfg.rc_resize_allowed = 0;
+
+  // Use 1 thread as default.
+  cfg.g_threads = 1;
+
+  // Enable error resilient mode.
+  cfg.g_error_resilient = 1;
+  cfg.g_lag_in_frames   = 0;
+  cfg.kf_mode = VPX_KF_AUTO;
+
+  // Disable automatic keyframe placement.
+  cfg.kf_min_dist = cfg.kf_max_dist = 3000;
+
+  cfg.temporal_layering_mode = VP9E_TEMPORAL_LAYERING_MODE_BYPASS;
+
+  set_temporal_layer_pattern(layering_mode,
+                             &cfg,
+                             layer_flags,
+                             &flag_periodicity);
+
+  set_rate_control_metrics(&rc, &cfg);
+
+  // Target bandwidth for the whole stream.
+  // Set to layer_target_bitrate for highest layer (total bitrate).
+  cfg.rc_target_bitrate = rc.layer_target_bitrate[cfg.ts_number_layers - 1];
+
+  // Open input file.
+  if (!(infile = fopen(argv[1], "rb"))) {
+    die("Failed to open %s for reading", argv[1]);
+  }
+
+  framerate = cfg.g_timebase.den / cfg.g_timebase.num;
+  // Open an output file for each stream.
+  for (i = 0; i < cfg.ts_number_layers; ++i) {
+    char file_name[PATH_MAX];
+    VpxVideoInfo info;
+    info.codec_fourcc = encoder->fourcc;
+    info.frame_width = cfg.g_w;
+    info.frame_height = cfg.g_h;
+    info.time_base.numerator = cfg.g_timebase.num;
+    info.time_base.denominator = cfg.g_timebase.den;
+
+    snprintf(file_name, sizeof(file_name), "%s_%d.ivf", argv[2], i);
+    outfile[i] = vpx_video_writer_open(file_name, kContainerIVF, &info);
+    if (!outfile[i])
+      die("Failed to open %s for writing", file_name);
+
+    assert(outfile[i] != NULL);
+  }
+  // No spatial layers in this encoder.
+  cfg.ss_number_layers = 1;
+
+  // Initialize codec.
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (vpx_codec_enc_init(
+          &codec, encoder->codec_interface(), &cfg,
+          bit_depth == VPX_BITS_8 ? 0 : VPX_CODEC_USE_HIGHBITDEPTH))
+#else
+  if (vpx_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    die_codec(&codec, "Failed to initialize encoder");
+
+  if (strncmp(encoder->name, "vp8", 3) == 0) {
+    vpx_codec_control(&codec, VP8E_SET_CPUUSED, -speed);
+    vpx_codec_control(&codec, VP8E_SET_NOISE_SENSITIVITY, kDenoiserOff);
+    vpx_codec_control(&codec, VP8E_SET_STATIC_THRESHOLD, 1);
+  } else if (strncmp(encoder->name, "vp9", 3) == 0) {
+    vpx_svc_extra_cfg_t svc_params;
+    vpx_codec_control(&codec, VP8E_SET_CPUUSED, speed);
+    vpx_codec_control(&codec, VP9E_SET_AQ_MODE, 3);
+    vpx_codec_control(&codec, VP9E_SET_FRAME_PERIODIC_BOOST, 0);
+    vpx_codec_control(&codec, VP9E_SET_NOISE_SENSITIVITY, kDenoiserOff);
+    vpx_codec_control(&codec, VP8E_SET_STATIC_THRESHOLD, 1);
+    vpx_codec_control(&codec, VP9E_SET_TUNE_CONTENT, 0);
+    vpx_codec_control(&codec, VP9E_SET_TILE_COLUMNS, (cfg.g_threads >> 1));
+    if (vpx_codec_control(&codec, VP9E_SET_SVC, layering_mode > 0 ? 1: 0))
+      die_codec(&codec, "Failed to set SVC");
+    for (i = 0; i < cfg.ts_number_layers; ++i) {
+      svc_params.max_quantizers[i] = cfg.rc_max_quantizer;
+      svc_params.min_quantizers[i] = cfg.rc_min_quantizer;
+    }
+    svc_params.scaling_factor_num[0] = cfg.g_h;
+    svc_params.scaling_factor_den[0] = cfg.g_h;
+    vpx_codec_control(&codec, VP9E_SET_SVC_PARAMETERS, &svc_params);
+  }
+  if (strncmp(encoder->name, "vp8", 3) == 0) {
+    vpx_codec_control(&codec, VP8E_SET_SCREEN_CONTENT_MODE, 0);
+  }
+  vpx_codec_control(&codec, VP8E_SET_TOKEN_PARTITIONS, 1);
+  // This controls the maximum target size of the key frame.
+  // For generating smaller key frames, use a smaller max_intra_size_pct
+  // value, like 100 or 200.
+  {
+    const int max_intra_size_pct = 900;
+    vpx_codec_control(&codec, VP8E_SET_MAX_INTRA_BITRATE_PCT,
+                      max_intra_size_pct);
+  }
+
+  frame_avail = 1;
+  while (frame_avail || got_data) {
+    struct vpx_usec_timer timer;
+    vpx_codec_iter_t iter = NULL;
+    const vpx_codec_cx_pkt_t *pkt;
+#if VPX_ENCODER_ABI_VERSION > (4 + VPX_CODEC_ABI_VERSION)
+    // Update the temporal layer_id. No spatial layers in this test.
+    layer_id.spatial_layer_id = 0;
+#endif
+    layer_id.temporal_layer_id =
+        cfg.ts_layer_id[frame_cnt % cfg.ts_periodicity];
+    if (strncmp(encoder->name, "vp9", 3) == 0) {
+      vpx_codec_control(&codec, VP9E_SET_SVC_LAYER_ID, &layer_id);
+    } else if (strncmp(encoder->name, "vp8", 3) == 0) {
+      vpx_codec_control(&codec, VP8E_SET_TEMPORAL_LAYER_ID,
+                        layer_id.temporal_layer_id);
+    }
+    flags = layer_flags[frame_cnt % flag_periodicity];
+    if (layering_mode == 0)
+      flags = 0;
+    frame_avail = vpx_img_read(&raw, infile);
+    if (frame_avail)
+      ++rc.layer_input_frames[layer_id.temporal_layer_id];
+    vpx_usec_timer_start(&timer);
+    if (vpx_codec_encode(&codec, frame_avail? &raw : NULL, pts, 1, flags,
+        VPX_DL_REALTIME)) {
+      die_codec(&codec, "Failed to encode frame");
+    }
+    vpx_usec_timer_mark(&timer);
+    cx_time += vpx_usec_timer_elapsed(&timer);
+    // Reset KF flag.
+    if (layering_mode != 7) {
+      layer_flags[0] &= ~VPX_EFLAG_FORCE_KF;
+    }
+    got_data = 0;
+    while ( (pkt = vpx_codec_get_cx_data(&codec, &iter)) ) {
+      got_data = 1;
+      switch (pkt->kind) {
+        case VPX_CODEC_CX_FRAME_PKT:
+          for (i = cfg.ts_layer_id[frame_cnt % cfg.ts_periodicity];
+              i < cfg.ts_number_layers; ++i) {
+            vpx_video_writer_write_frame(outfile[i], pkt->data.frame.buf,
+                                         pkt->data.frame.sz, pts);
+            ++rc.layer_tot_enc_frames[i];
+            rc.layer_encoding_bitrate[i] += 8.0 * pkt->data.frame.sz;
+            // Keep count of rate control stats per layer (for non-key frames).
+            if (i == cfg.ts_layer_id[frame_cnt % cfg.ts_periodicity] &&
+                !(pkt->data.frame.flags & VPX_FRAME_IS_KEY)) {
+              rc.layer_avg_frame_size[i] += 8.0 * pkt->data.frame.sz;
+              rc.layer_avg_rate_mismatch[i] +=
+                  fabs(8.0 * pkt->data.frame.sz - rc.layer_pfb[i]) /
+                  rc.layer_pfb[i];
+              ++rc.layer_enc_frames[i];
+            }
+          }
+          // Update for short-time encoding bitrate states, for moving window
+          // of size rc->window, shifted by rc->window / 2.
+          // Ignore first window segment, due to key frame.
+          if (frame_cnt > rc.window_size) {
+            sum_bitrate += 0.001 * 8.0 * pkt->data.frame.sz * framerate;
+            if (frame_cnt % rc.window_size == 0) {
+              rc.window_count += 1;
+              rc.avg_st_encoding_bitrate += sum_bitrate / rc.window_size;
+              rc.variance_st_encoding_bitrate +=
+                  (sum_bitrate / rc.window_size) *
+                  (sum_bitrate / rc.window_size);
+              sum_bitrate = 0.0;
+            }
+          }
+          // Second shifted window.
+          if (frame_cnt > rc.window_size + rc.window_size / 2) {
+            sum_bitrate2 += 0.001 * 8.0 * pkt->data.frame.sz * framerate;
+            if (frame_cnt > 2 * rc.window_size &&
+                frame_cnt % rc.window_size == 0) {
+              rc.window_count += 1;
+              rc.avg_st_encoding_bitrate += sum_bitrate2 / rc.window_size;
+              rc.variance_st_encoding_bitrate +=
+                  (sum_bitrate2 / rc.window_size) *
+                  (sum_bitrate2 / rc.window_size);
+              sum_bitrate2 = 0.0;
+            }
+          }
+          break;
+          default:
+            break;
+      }
+    }
+    ++frame_cnt;
+    pts += frame_duration;
+  }
+  fclose(infile);
+  printout_rate_control_summary(&rc, &cfg, frame_cnt);
+  printf("\n");
+  printf("Frame cnt and encoding time/FPS stats for encoding: %d %f %f \n",
+          frame_cnt,
+          1000 * (float)cx_time / (double)(frame_cnt * 1000000),
+          1000000 * (double)frame_cnt / (double)cx_time);
+
+  if (vpx_codec_destroy(&codec))
+    die_codec(&codec, "Failed to destroy codec");
+
+  // Try to rewrite the output file headers with the actual frame count.
+  for (i = 0; i < cfg.ts_number_layers; ++i)
+    vpx_video_writer_close(outfile[i]);
+
+  vpx_img_free(&raw);
+  return EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/ivfdec.c b/libvpx/libvpx/ivfdec.c
new file mode 100644
index 0000000..7fc25a0
--- /dev/null
+++ b/libvpx/libvpx/ivfdec.c
@@ -0,0 +1,112 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx_ports/mem_ops.h"
+
+#include "./ivfdec.h"
+
+static const char *IVF_SIGNATURE = "DKIF";
+
+static void fix_framerate(int *num, int *den) {
+  // Some versions of vpxenc used 1/(2*fps) for the timebase, so
+  // we can guess the framerate using only the timebase in this
+  // case. Other files would require reading ahead to guess the
+  // timebase, like we do for webm.
+  if (*den > 0 && *den < 1000000000 && *num > 0 && *num < 1000) {
+    // Correct for the factor of 2 applied to the timebase in the encoder.
+    if (*num & 1)
+      *den *= 2;
+    else
+      *num /= 2;
+  } else {
+    // Don't know FPS for sure, and don't have readahead code
+    // (yet?), so just default to 30fps.
+    *num = 30;
+    *den = 1;
+  }
+}
+
+int file_is_ivf(struct VpxInputContext *input_ctx) {
+  char raw_hdr[32];
+  int is_ivf = 0;
+
+  if (fread(raw_hdr, 1, 32, input_ctx->file) == 32) {
+    if (memcmp(IVF_SIGNATURE, raw_hdr, 4) == 0) {
+      is_ivf = 1;
+
+      if (mem_get_le16(raw_hdr + 4) != 0) {
+        fprintf(stderr, "Error: Unrecognized IVF version! This file may not"
+                " decode properly.");
+      }
+
+      input_ctx->fourcc = mem_get_le32(raw_hdr + 8);
+      input_ctx->width = mem_get_le16(raw_hdr + 12);
+      input_ctx->height = mem_get_le16(raw_hdr + 14);
+      input_ctx->framerate.numerator = mem_get_le32(raw_hdr + 16);
+      input_ctx->framerate.denominator = mem_get_le32(raw_hdr + 20);
+      fix_framerate(&input_ctx->framerate.numerator,
+                    &input_ctx->framerate.denominator);
+    }
+  }
+
+  if (!is_ivf) {
+    rewind(input_ctx->file);
+    input_ctx->detect.buf_read = 0;
+  } else {
+    input_ctx->detect.position = 4;
+  }
+  return is_ivf;
+}
+
+int ivf_read_frame(FILE *infile, uint8_t **buffer,
+                   size_t *bytes_read, size_t *buffer_size) {
+  char raw_header[IVF_FRAME_HDR_SZ] = {0};
+  size_t frame_size = 0;
+
+  if (fread(raw_header, IVF_FRAME_HDR_SZ, 1, infile) != 1) {
+    if (!feof(infile))
+      warn("Failed to read frame size\n");
+  } else {
+    frame_size = mem_get_le32(raw_header);
+
+    if (frame_size > 256 * 1024 * 1024) {
+      warn("Read invalid frame size (%u)\n", (unsigned int)frame_size);
+      frame_size = 0;
+    }
+
+    if (frame_size > *buffer_size) {
+      uint8_t *new_buffer = realloc(*buffer, 2 * frame_size);
+
+      if (new_buffer) {
+        *buffer = new_buffer;
+        *buffer_size = 2 * frame_size;
+      } else {
+        warn("Failed to allocate compressed data buffer\n");
+        frame_size = 0;
+      }
+    }
+  }
+
+  if (!feof(infile)) {
+    if (fread(*buffer, 1, frame_size, infile) != frame_size) {
+      warn("Failed to read full frame\n");
+      return 1;
+    }
+
+    *bytes_read = frame_size;
+    return 0;
+  }
+
+  return 1;
+}
diff --git a/libvpx/libvpx/ivfdec.h b/libvpx/libvpx/ivfdec.h
new file mode 100644
index 0000000..dd29cc6
--- /dev/null
+++ b/libvpx/libvpx/ivfdec.h
@@ -0,0 +1,28 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef IVFDEC_H_
+#define IVFDEC_H_
+
+#include "./tools_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int file_is_ivf(struct VpxInputContext *input);
+
+int ivf_read_frame(FILE *infile, uint8_t **buffer,
+                   size_t *bytes_read, size_t *buffer_size);
+
+#ifdef __cplusplus
+}  /* extern "C" */
+#endif
+
+#endif  // IVFDEC_H_
diff --git a/libvpx/libvpx/ivfenc.c b/libvpx/libvpx/ivfenc.c
new file mode 100644
index 0000000..4a97c42
--- /dev/null
+++ b/libvpx/libvpx/ivfenc.c
@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./ivfenc.h"
+
+#include "vpx/vpx_encoder.h"
+#include "vpx_ports/mem_ops.h"
+
+void ivf_write_file_header(FILE *outfile,
+                           const struct vpx_codec_enc_cfg *cfg,
+                           unsigned int fourcc,
+                           int frame_cnt) {
+  char header[32];
+
+  header[0] = 'D';
+  header[1] = 'K';
+  header[2] = 'I';
+  header[3] = 'F';
+  mem_put_le16(header + 4, 0);                     // version
+  mem_put_le16(header + 6, 32);                    // header size
+  mem_put_le32(header + 8, fourcc);                // fourcc
+  mem_put_le16(header + 12, cfg->g_w);             // width
+  mem_put_le16(header + 14, cfg->g_h);             // height
+  mem_put_le32(header + 16, cfg->g_timebase.den);  // rate
+  mem_put_le32(header + 20, cfg->g_timebase.num);  // scale
+  mem_put_le32(header + 24, frame_cnt);            // length
+  mem_put_le32(header + 28, 0);                    // unused
+
+  fwrite(header, 1, 32, outfile);
+}
+
+void ivf_write_frame_header(FILE *outfile, int64_t pts, size_t frame_size) {
+  char header[12];
+
+  mem_put_le32(header, (int)frame_size);
+  mem_put_le32(header + 4, (int)(pts & 0xFFFFFFFF));
+  mem_put_le32(header + 8, (int)(pts >> 32));
+  fwrite(header, 1, 12, outfile);
+}
+
+void ivf_write_frame_size(FILE *outfile, size_t frame_size) {
+  char header[4];
+
+  mem_put_le32(header, (int)frame_size);
+  fwrite(header, 1, 4, outfile);
+}
diff --git a/libvpx/libvpx/ivfenc.h b/libvpx/libvpx/ivfenc.h
new file mode 100644
index 0000000..6623687
--- /dev/null
+++ b/libvpx/libvpx/ivfenc.h
@@ -0,0 +1,35 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef IVFENC_H_
+#define IVFENC_H_
+
+#include "./tools_common.h"
+
+struct vpx_codec_enc_cfg;
+struct vpx_codec_cx_pkt;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ivf_write_file_header(FILE *outfile,
+                           const struct vpx_codec_enc_cfg *cfg,
+                           uint32_t fourcc,
+                           int frame_cnt);
+
+void ivf_write_frame_header(FILE *outfile, int64_t pts, size_t frame_size);
+
+void ivf_write_frame_size(FILE *outfile, size_t frame_size);
+
+#ifdef __cplusplus
+}  /* extern "C" */
+#endif
+
+#endif  // IVFENC_H_
diff --git a/libvpx/libvpx/keywords.dox b/libvpx/libvpx/keywords.dox
new file mode 100644
index 0000000..56f5368
--- /dev/null
+++ b/libvpx/libvpx/keywords.dox
@@ -0,0 +1,51 @@
+/*!\page rfc2119 RFC2119 Keywords
+
+      The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL
+      NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED",  "MAY", and
+      "OPTIONAL" in this document are to be interpreted as described in
+      <a href="http://www.ietf.org/rfc/rfc2119.txt">RFC 2119.</a>
+
+Specifically, the following definitions are used:
+
+\section MUST
+\anchor REQUIRED
+\anchor SHALL
+   This word, or the terms "REQUIRED" or "SHALL", mean that the
+   definition is an absolute requirement of the specification.
+
+\section MUSTNOT MUST NOT
+\anchor SHALLNOT
+   This phrase, or the phrase "SHALL NOT", mean that the
+   definition is an absolute prohibition of the specification.
+
+\section SHOULD
+\anchor RECOMMENDED
+   This word, or the adjective "RECOMMENDED", mean that there
+   may exist valid reasons in particular circumstances to ignore a
+   particular item, but the full implications must be understood and
+   carefully weighed before choosing a different course.
+
+\section SHOULDNOT SHOULD NOT
+\anchor NOTRECOMMENDED
+   This phrase, or the phrase "NOT RECOMMENDED" mean that
+   there may exist valid reasons in particular circumstances when the
+   particular behavior is acceptable or even useful, but the full
+   implications should be understood and the case carefully weighed
+   before implementing any behavior described with this label.
+
+\section MAY
+\anchor OPTIONAL
+   This word, or the adjective "OPTIONAL", mean that an item is
+   truly optional.  One vendor may choose to include the item because a
+   particular marketplace requires it or because the vendor feels that
+   it enhances the product while another vendor may omit the same item.
+   An implementation which does not include a particular option \ref MUST be
+   prepared to interoperate with another implementation which does
+   include the option, though perhaps with reduced functionality. In the
+   same vein an implementation which does include a particular option
+   \ref MUST be prepared to interoperate with another implementation which
+   does not include the option (except, of course, for the feature the
+   option provides.)
+
+
+*/
diff --git a/libvpx/libvpx/libs.doxy_template b/libvpx/libvpx/libs.doxy_template
new file mode 100644
index 0000000..5a8f847
--- /dev/null
+++ b/libvpx/libvpx/libs.doxy_template
@@ -0,0 +1,1296 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+# Doxyfile 1.5.4
+
+# This file describes the settings to be used by the documentation system
+# doxygen (www.doxygen.org) for a project
+#
+# All text after a hash (#) is considered a comment and will be ignored
+# The format is:
+#       TAG = value [value, ...]
+# For lists items can also be appended using:
+#       TAG += value [value, ...]
+# Values that contain spaces should be placed between quotes (" ")
+
+#---------------------------------------------------------------------------
+# Project related configuration options
+#---------------------------------------------------------------------------
+
+# This tag specifies the encoding used for all characters in the config file that
+# follow. The default is UTF-8 which is also the encoding used for all text before
+# the first occurrence of this tag. Doxygen uses libiconv (or the iconv built into
+# libc) for the transcoding. See http://www.gnu.org/software/libiconv for the list of
+# possible encodings.
+
+DOXYFILE_ENCODING      = UTF-8
+
+# The PROJECT_NAME tag is a single word (or a sequence of words surrounded
+# by quotes) that should identify the project.
+
+PROJECT_NAME           = "WebM Codec SDK"
+
+# The OUTPUT_DIRECTORY tag is used to specify the (relative or absolute)
+# base path where the generated documentation will be put.
+# If a relative path is entered, it will be relative to the location
+# where doxygen was started. If left blank the current directory will be used.
+
+OUTPUT_DIRECTORY       = docs
+
+# If the CREATE_SUBDIRS tag is set to YES, then doxygen will create
+# 4096 sub-directories (in 2 levels) under the output directory of each output
+# format and will distribute the generated files over these directories.
+# Enabling this option can be useful when feeding doxygen a huge amount of
+# source files, where putting all generated files in the same directory would
+# otherwise cause performance problems for the file system.
+
+CREATE_SUBDIRS         = NO
+
+# The OUTPUT_LANGUAGE tag is used to specify the language in which all
+# documentation generated by doxygen is written. Doxygen will use this
+# information to generate all constant output in the proper language.
+# The default language is English, other supported languages are:
+# Afrikaans, Arabic, Brazilian, Catalan, Chinese, Chinese-Traditional,
+# Croatian, Czech, Danish, Dutch, Finnish, French, German, Greek, Hungarian,
+# Italian, Japanese, Japanese-en (Japanese with English messages), Korean,
+# Korean-en, Lithuanian, Norwegian, Polish, Portuguese, Romanian, Russian,
+# Serbian, Slovak, Slovene, Spanish, Swedish, and Ukrainian.
+
+OUTPUT_LANGUAGE        = English
+
+# If the BRIEF_MEMBER_DESC tag is set to YES (the default) Doxygen will
+# include brief member descriptions after the members that are listed in
+# the file and class documentation (similar to java_doc).
+# Set to NO to disable this.
+
+BRIEF_MEMBER_DESC      = YES
+
+# If the REPEAT_BRIEF tag is set to YES (the default) Doxygen will prepend
+# the brief description of a member or function before the detailed description.
+# Note: if both HIDE_UNDOC_MEMBERS and BRIEF_MEMBER_DESC are set to NO, the
+# brief descriptions will be completely suppressed.
+
+REPEAT_BRIEF           = YES
+
+# This tag implements a quasi-intelligent brief description abbreviator
+# that is used to form the text in various listings. Each string
+# in this list, if found as the leading text of the brief description, will be
+# stripped from the text and the result after processing the whole list, is
+# used as the annotated text. Otherwise, the brief description is used as-is.
+# If left blank, the following values are used ("$name" is automatically
+# replaced with the name of the entity): "The $name class" "The $name widget"
+# "The $name file" "is" "provides" "specifies" "contains"
+# "represents" "a" "an" "the"
+
+ABBREVIATE_BRIEF       =
+
+# If the ALWAYS_DETAILED_SEC and REPEAT_BRIEF tags are both set to YES then
+# Doxygen will generate a detailed section even if there is only a brief
+# description.
+
+ALWAYS_DETAILED_SEC    = NO
+
+# If the INLINE_INHERITED_MEMB tag is set to YES, doxygen will show all
+# inherited members of a class in the documentation of that class as if those
+# members were ordinary class members. Constructors, destructors and assignment
+# operators of the base classes will not be shown.
+
+INLINE_INHERITED_MEMB  = NO
+
+# If the FULL_PATH_NAMES tag is set to YES then Doxygen will prepend the full
+# path before files name in the file list and in the header files. If set
+# to NO the shortest path that makes the file name unique will be used.
+
+FULL_PATH_NAMES        = YES
+
+# If the FULL_PATH_NAMES tag is set to YES then the STRIP_FROM_PATH tag
+# can be used to strip a user-defined part of the path. Stripping is
+# only done if one of the specified strings matches the left-hand part of
+# the path. The tag can be used to show relative paths in the file list.
+# If left blank the directory from which doxygen is run is used as the
+# path to strip.
+
+STRIP_FROM_PATH        =
+
+# The STRIP_FROM_INC_PATH tag can be used to strip a user-defined part of
+# the path mentioned in the documentation of a class, which tells
+# the reader which header file to include in order to use a class.
+# If left blank only the name of the header file containing the class
+# definition is used. Otherwise one should specify the include paths that
+# are normally passed to the compiler using the -I flag.
+
+STRIP_FROM_INC_PATH    =
+
+# If the SHORT_NAMES tag is set to YES, doxygen will generate much shorter
+# (but less readable) file names. This can be useful is your file systems
+# doesn't support long names like on DOS, Mac, or CD-ROM.
+
+SHORT_NAMES            = NO
+
+# If the JAVADOC_AUTOBRIEF tag is set to YES then Doxygen
+# will interpret the first line (until the first dot) of a java_doc-style
+# comment as the brief description. If set to NO, the java_doc
+# comments will behave just like regular Qt-style comments
+# (thus requiring an explicit @brief command for a brief description.)
+
+JAVADOC_AUTOBRIEF      = NO
+
+# If the QT_AUTOBRIEF tag is set to YES then Doxygen will
+# interpret the first line (until the first dot) of a Qt-style
+# comment as the brief description. If set to NO, the comments
+# will behave just like regular Qt-style comments (thus requiring
+# an explicit \brief command for a brief description.)
+
+QT_AUTOBRIEF           = NO
+
+# The MULTILINE_CPP_IS_BRIEF tag can be set to YES to make Doxygen
+# treat a multi-line C++ special comment block (i.e. a block of //! or ///
+# comments) as a brief description. This used to be the default behaviour.
+# The new default is to treat a multi-line C++ comment block as a detailed
+# description. Set this tag to YES if you prefer the old behaviour instead.
+
+MULTILINE_CPP_IS_BRIEF = NO
+
+# If the INHERIT_DOCS tag is set to YES (the default) then an undocumented
+# member inherits the documentation from any documented member that it
+# re-implements.
+
+INHERIT_DOCS           = YES
+
+# If the SEPARATE_MEMBER_PAGES tag is set to YES, then doxygen will produce
+# a new page for each member. If set to NO, the documentation of a member will
+# be part of the file/class/namespace that contains it.
+
+SEPARATE_MEMBER_PAGES  = NO
+
+# The TAB_SIZE tag can be used to set the number of spaces in a tab.
+# Doxygen uses this value to replace tabs by spaces in code fragments.
+
+TAB_SIZE               = 4
+
+# This tag can be used to specify a number of aliases that acts
+# as commands in the documentation. An alias has the form "name=value".
+# For example adding "sideeffect=\par Side Effects:\n" will allow you to
+# put the command \sideeffect (or @sideeffect) in the documentation, which
+# will result in a user-defined paragraph with heading "Side Effects:".
+# You can put \n's in the value part of an alias to insert newlines.
+
+ALIASES                =
+
+# Set the OPTIMIZE_OUTPUT_FOR_C tag to YES if your project consists of C
+# sources only. Doxygen will then generate output that is more tailored for C.
+# For instance, some of the names that are used will be different. The list
+# of all members will be omitted, etc.
+
+OPTIMIZE_OUTPUT_FOR_C  = YES
+
+# Set the OPTIMIZE_OUTPUT_JAVA tag to YES if your project consists of Java
+# sources only. Doxygen will then generate output that is more tailored for Java.
+# For instance, namespaces will be presented as packages, qualified scopes
+# will look different, etc.
+
+OPTIMIZE_OUTPUT_JAVA   = NO
+
+# If you use STL classes (i.e. std::string, std::vector, etc.) but do not want to
+# include (a tag file for) the STL sources as input, then you should
+# set this tag to YES in order to let doxygen match functions declarations and
+# definitions whose arguments contain STL classes (e.g. func(std::string); v.s.
+# func(std::string) {}). This also make the inheritance and collaboration
+# diagrams that involve STL classes more complete and accurate.
+
+BUILTIN_STL_SUPPORT    = NO
+
+# If you use Microsoft's C++/CLI language, you should set this option to YES to
+# enable parsing support.
+
+CPP_CLI_SUPPORT        = NO
+
+# Set the SIP_SUPPORT tag to YES if your project consists of sip sources only.
+# Doxygen will parse them like normal C++ but will assume all classes use public
+# instead of private inheritance when no explicit protection keyword is present.
+
+SIP_SUPPORT            = NO
+
+# If member grouping is used in the documentation and the DISTRIBUTE_GROUP_DOC
+# tag is set to YES, then doxygen will reuse the documentation of the first
+# member in the group (if any) for the other members of the group. By default
+# all members of a group must be documented explicitly.
+
+DISTRIBUTE_GROUP_DOC   = NO
+
+# Set the SUBGROUPING tag to YES (the defqault) to allow class member groups of
+# the same type (for instance a group of public functions) to be put as a
+# subgroup of that type (e.g. under the Public Functions section). Set it to
+# NO to prevent subgrouping. Alternatively, this can be done per class using
+# the \nosubgrouping command.
+
+SUBGROUPING            = YES
+
+# When TYPEDEF_HIDES_STRUCT is enabled, a typedef of a struct (or union) is
+# documented as struct with the name of the typedef. So
+# typedef struct type_s {} type_t, will appear in the documentation as a struct
+# with name type_t. When disabled the typedef will appear as a member of a file,
+# namespace, or class. And the struct will be named type_s. This can typically
+# be useful for C code where the coding convention is that all structs are
+# typedef'ed and only the typedef is referenced never the struct's name.
+
+TYPEDEF_HIDES_STRUCT   = NO
+
+#---------------------------------------------------------------------------
+# Build related configuration options
+#---------------------------------------------------------------------------
+
+# If the EXTRACT_ALL tag is set to YES doxygen will assume all entities in
+# documentation are documented, even if no documentation was available.
+# Private class members and static file members will be hidden unless
+# the EXTRACT_PRIVATE and EXTRACT_STATIC tags are set to YES
+
+EXTRACT_ALL            = NO
+
+# If the EXTRACT_PRIVATE tag is set to YES all private members of a class
+# will be included in the documentation.
+
+EXTRACT_PRIVATE        = NO
+
+# If the EXTRACT_STATIC tag is set to YES all static members of a file
+# will be included in the documentation.
+
+EXTRACT_STATIC         = NO
+
+# If the EXTRACT_LOCAL_CLASSES tag is set to YES classes (and structs)
+# defined locally in source files will be included in the documentation.
+# If set to NO only classes defined in header files are included.
+
+EXTRACT_LOCAL_CLASSES  = YES
+
+# This flag is only useful for Objective-C code. When set to YES local
+# methods, which are defined in the implementation section but not in
+# the interface are included in the documentation.
+# If set to NO (the default) only methods in the interface are included.
+
+EXTRACT_LOCAL_METHODS  = NO
+
+# If this flag is set to YES, the members of anonymous namespaces will be extracted
+# and appear in the documentation as a namespace called 'anonymous_namespace{file}',
+# where file will be replaced with the base name of the file that contains the anonymous
+# namespace. By default anonymous namespace are hidden.
+
+EXTRACT_ANON_NSPACES   = NO
+
+# If the HIDE_UNDOC_MEMBERS tag is set to YES, Doxygen will hide all
+# undocumented members of documented classes, files or namespaces.
+# If set to NO (the default) these members will be included in the
+# various overviews, but no documentation section is generated.
+# This option has no effect if EXTRACT_ALL is enabled.
+
+HIDE_UNDOC_MEMBERS     = NO
+
+# If the HIDE_UNDOC_CLASSES tag is set to YES, Doxygen will hide all
+# undocumented classes that are normally visible in the class hierarchy.
+# If set to NO (the default) these classes will be included in the various
+# overviews. This option has no effect if EXTRACT_ALL is enabled.
+
+HIDE_UNDOC_CLASSES     = NO
+
+# If the HIDE_FRIEND_COMPOUNDS tag is set to YES, Doxygen will hide all
+# friend (class|struct|union) declarations.
+# If set to NO (the default) these declarations will be included in the
+# documentation.
+
+HIDE_FRIEND_COMPOUNDS  = NO
+
+# If the HIDE_IN_BODY_DOCS tag is set to YES, Doxygen will hide any
+# documentation blocks found inside the body of a function.
+# If set to NO (the default) these blocks will be appended to the
+# function's detailed documentation block.
+
+HIDE_IN_BODY_DOCS      = NO
+
+# The INTERNAL_DOCS tag determines if documentation
+# that is typed after a \internal command is included. If the tag is set
+# to NO (the default) then the documentation will be excluded.
+# Set it to YES to include the internal documentation.
+
+INTERNAL_DOCS          = NO
+
+# If the CASE_SENSE_NAMES tag is set to NO then Doxygen will only generate
+# file names in lower-case letters. If set to YES upper-case letters are also
+# allowed. This is useful if you have classes or files whose names only differ
+# in case and if your file system supports case sensitive file names. Windows
+# and Mac users are advised to set this option to NO.
+
+CASE_SENSE_NAMES       = YES
+
+# If the HIDE_SCOPE_NAMES tag is set to NO (the default) then Doxygen
+# will show members with their full class and namespace scopes in the
+# documentation. If set to YES the scope will be hidden.
+
+HIDE_SCOPE_NAMES       = NO
+
+# If the SHOW_INCLUDE_FILES tag is set to YES (the default) then Doxygen
+# will put a list of the files that are included by a file in the documentation
+# of that file.
+
+SHOW_INCLUDE_FILES     = YES
+
+# If the INLINE_INFO tag is set to YES (the default) then a tag [inline]
+# is inserted in the documentation for inline members.
+
+INLINE_INFO            = YES
+
+# If the SORT_MEMBER_DOCS tag is set to YES (the default) then doxygen
+# will sort the (detailed) documentation of file and class members
+# alphabetically by member name. If set to NO the members will appear in
+# declaration order.
+
+SORT_MEMBER_DOCS       = NO
+
+# If the SORT_BRIEF_DOCS tag is set to YES then doxygen will sort the
+# brief documentation of file, namespace and class members alphabetically
+# by member name. If set to NO (the default) the members will appear in
+# declaration order.
+
+SORT_BRIEF_DOCS        = NO
+
+# If the SORT_BY_SCOPE_NAME tag is set to YES, the class list will be
+# sorted by fully-qualified names, including namespaces. If set to
+# NO (the default), the class list will be sorted only by class name,
+# not including the namespace part.
+# Note: This option is not very useful if HIDE_SCOPE_NAMES is set to YES.
+# Note: This option applies only to the class list, not to the
+# alphabetical list.
+
+SORT_BY_SCOPE_NAME     = NO
+
+# The GENERATE_TODOLIST tag can be used to enable (YES) or
+# disable (NO) the todo list. This list is created by putting \todo
+# commands in the documentation.
+
+GENERATE_TODOLIST      = YES
+
+# The GENERATE_TESTLIST tag can be used to enable (YES) or
+# disable (NO) the test list. This list is created by putting \test
+# commands in the documentation.
+
+GENERATE_TESTLIST      = YES
+
+# The GENERATE_BUGLIST tag can be used to enable (YES) or
+# disable (NO) the bug list. This list is created by putting \bug
+# commands in the documentation.
+
+GENERATE_BUGLIST       = YES
+
+# The GENERATE_DEPRECATEDLIST tag can be used to enable (YES) or
+# disable (NO) the deprecated list. This list is created by putting
+# \deprecated commands in the documentation.
+
+GENERATE_DEPRECATEDLIST= YES
+
+# The ENABLED_SECTIONS tag can be used to enable conditional
+# documentation sections, marked by \if sectionname ... \endif.
+
+ENABLED_SECTIONS       =
+
+# The MAX_INITIALIZER_LINES tag determines the maximum number of lines
+# the initial value of a variable or define consists of for it to appear in
+# the documentation. If the initializer consists of more lines than specified
+# here it will be hidden. Use a value of 0 to hide initializers completely.
+# The appearance of the initializer of individual variables and defines in the
+# documentation can be controlled using \showinitializer or \hideinitializer
+# command in the documentation regardless of this setting.
+
+MAX_INITIALIZER_LINES  = 30
+
+# Set the SHOW_USED_FILES tag to NO to disable the list of files generated
+# at the bottom of the documentation of classes and structs. If set to YES the
+# list will mention the files that were used to generate the documentation.
+
+SHOW_USED_FILES        = YES
+
+# The FILE_VERSION_FILTER tag can be used to specify a program or script that
+# doxygen should invoke to get the current version for each file (typically from the
+# version control system). Doxygen will invoke the program by executing (via
+# popen()) the command <command> <input-file>, where <command> is the value of
+# the FILE_VERSION_FILTER tag, and <input-file> is the name of an input file
+# provided by doxygen. Whatever the program writes to standard output
+# is used as the file version. See the manual for examples.
+
+FILE_VERSION_FILTER    =
+
+#---------------------------------------------------------------------------
+# configuration options related to warning and progress messages
+#---------------------------------------------------------------------------
+
+# The QUIET tag can be used to turn on/off the messages that are generated
+# by doxygen. Possible values are YES and NO. If left blank NO is used.
+
+QUIET                  = YES
+
+# The WARNINGS tag can be used to turn on/off the warning messages that are
+# generated by doxygen. Possible values are YES and NO. If left blank
+# NO is used.
+
+WARNINGS               = YES
+
+# If WARN_IF_UNDOCUMENTED is set to YES, then doxygen will generate warnings
+# for undocumented members. If EXTRACT_ALL is set to YES then this flag will
+# automatically be disabled.
+
+WARN_IF_UNDOCUMENTED   = YES
+
+# If WARN_IF_DOC_ERROR is set to YES, doxygen will generate warnings for
+# potential errors in the documentation, such as not documenting some
+# parameters in a documented function, or documenting parameters that
+# don't exist or using markup commands wrongly.
+
+WARN_IF_DOC_ERROR      = YES
+
+# This WARN_NO_PARAMDOC option can be abled to get warnings for
+# functions that are documented, but have no documentation for their parameters
+# or return value. If set to NO (the default) doxygen will only warn about
+# wrong or incomplete parameter documentation, but not about the absence of
+# documentation.
+
+WARN_NO_PARAMDOC       = NO
+
+# The WARN_FORMAT tag determines the format of the warning messages that
+# doxygen can produce. The string should contain the $file, $line, and $text
+# tags, which will be replaced by the file and line number from which the
+# warning originated and the warning text. Optionally the format may contain
+# $version, which will be replaced by the version of the file (if it could
+# be obtained via FILE_VERSION_FILTER)
+
+WARN_FORMAT            = "$file:$line: $text"
+
+# The WARN_LOGFILE tag can be used to specify a file to which warning
+# and error messages should be written. If left blank the output is written
+# to stderr.
+
+WARN_LOGFILE           =
+
+#---------------------------------------------------------------------------
+# configuration options related to the input files
+#---------------------------------------------------------------------------
+
+# The INPUT tag can be used to specify the files and/or directories that contain
+# documented source files. You may enter file names like "myfile.cpp" or
+# directories like "/usr/src/myproject". Separate the files or directories
+# with spaces.
+
+INPUT =
+
+# This tag can be used to specify the character encoding of the source files that
+# doxygen parses. Internally doxygen uses the UTF-8 encoding, which is also the default
+# input encoding. Doxygen uses libiconv (or the iconv built into libc) for the transcoding.
+# See http://www.gnu.org/software/libiconv for the list of possible encodings.
+
+INPUT_ENCODING         = UTF-8
+
+# If the value of the INPUT tag contains directories, you can use the
+# FILE_PATTERNS tag to specify one or more wildcard pattern (like *.cpp
+# and *.h) to filter out the source-files in the directories. If left
+# blank the following patterns are tested:
+# *.c *.cc *.cxx *.cpp *.c++ *.java *.ii *.ixx *.ipp *.i++ *.inl *.h *.hh *.hxx
+# *.hpp *.h++ *.idl *.odl *.cs *.php *.php3 *.inc *.m *.mm *.py *.f90
+
+FILE_PATTERNS          =
+
+# The RECURSIVE tag can be used to turn specify whether or not subdirectories
+# should be searched for input files as well. Possible values are YES and NO.
+# If left blank NO is used.
+
+RECURSIVE              = NO
+
+# The EXCLUDE tag can be used to specify files and/or directories that should
+# excluded from the INPUT source files. This way you can easily exclude a
+# subdirectory from a directory tree whose root is specified with the INPUT tag.
+
+EXCLUDE                =
+
+# The EXCLUDE_SYMLINKS tag can be used select whether or not files or
+# directories that are symbolic links (a Unix filesystem feature) are excluded
+# from the input.
+
+EXCLUDE_SYMLINKS       = NO
+
+# If the value of the INPUT tag contains directories, you can use the
+# EXCLUDE_PATTERNS tag to specify one or more wildcard patterns to exclude
+# certain files from those directories. Note that the wildcards are matched
+# against the file with absolute path, so to exclude all test directories
+# for example use the pattern */test/*
+
+EXCLUDE_PATTERNS       =
+
+# The EXCLUDE_SYMBOLS tag can be used to specify one or more symbol names
+# (namespaces, classes, functions, etc.) that should be excluded from the output.
+# The symbol name can be a fully qualified name, a word, or if the wildcard * is used,
+# a substring. Examples: ANamespace, AClass, AClass::ANamespace, ANamespace::*Test
+
+EXCLUDE_SYMBOLS        =
+
+# The EXAMPLE_PATH tag can be used to specify one or more files or
+# directories that contain example code fragments that are included (see
+# the \include command).
+
+EXAMPLE_PATH           =
+
+# If the value of the EXAMPLE_PATH tag contains directories, you can use the
+# EXAMPLE_PATTERNS tag to specify one or more wildcard pattern (like *.cpp
+# and *.h) to filter out the source-files in the directories. If left
+# blank all files are included.
+
+EXAMPLE_PATTERNS       =
+
+# If the EXAMPLE_RECURSIVE tag is set to YES then subdirectories will be
+# searched for input files to be used with the \include or \dontinclude
+# commands irrespective of the value of the RECURSIVE tag.
+# Possible values are YES and NO. If left blank NO is used.
+
+EXAMPLE_RECURSIVE      = NO
+
+# The IMAGE_PATH tag can be used to specify one or more files or
+# directories that contain image that are included in the documentation (see
+# the \image command).
+
+IMAGE_PATH             =
+
+# The INPUT_FILTER tag can be used to specify a program that doxygen should
+# invoke to filter for each input file. Doxygen will invoke the filter program
+# by executing (via popen()) the command <filter> <input-file>, where <filter>
+# is the value of the INPUT_FILTER tag, and <input-file> is the name of an
+# input file. Doxygen will then use the output that the filter program writes
+# to standard output.  If FILTER_PATTERNS is specified, this tag will be
+# ignored.
+
+INPUT_FILTER           =
+
+# The FILTER_PATTERNS tag can be used to specify filters on a per file pattern
+# basis.  Doxygen will compare the file name with each pattern and apply the
+# filter if there is a match.  The filters are a list of the form:
+# pattern=filter (like *.cpp=my_cpp_filter). See INPUT_FILTER for further
+# info on how filters are used. If FILTER_PATTERNS is empty, INPUT_FILTER
+# is applied to all files.
+
+FILTER_PATTERNS        =
+
+# If the FILTER_SOURCE_FILES tag is set to YES, the input filter (if set using
+# INPUT_FILTER) will be used to filter the input files when producing source
+# files to browse (i.e. when SOURCE_BROWSER is set to YES).
+
+FILTER_SOURCE_FILES    = NO
+
+#---------------------------------------------------------------------------
+# configuration options related to source browsing
+#---------------------------------------------------------------------------
+
+# If the SOURCE_BROWSER tag is set to YES then a list of source files will
+# be generated. Documented entities will be cross-referenced with these sources.
+# Note: To get rid of all source code in the generated output, make sure also
+# VERBATIM_HEADERS is set to NO. If you have enabled CALL_GRAPH or CALLER_GRAPH
+# then you must also enable this option. If you don't then doxygen will produce
+# a warning and turn it on anyway
+
+SOURCE_BROWSER         = NO
+
+# Setting the INLINE_SOURCES tag to YES will include the body
+# of functions and classes directly in the documentation.
+
+INLINE_SOURCES         = NO
+
+# Setting the STRIP_CODE_COMMENTS tag to YES (the default) will instruct
+# doxygen to hide any special comment blocks from generated source code
+# fragments. Normal C and C++ comments will always remain visible.
+
+STRIP_CODE_COMMENTS    = YES
+
+# If the REFERENCED_BY_RELATION tag is set to YES (the default)
+# then for each documented function all documented
+# functions referencing it will be listed.
+
+REFERENCED_BY_RELATION = YES
+
+# If the REFERENCES_RELATION tag is set to YES (the default)
+# then for each documented function all documented entities
+# called/used by that function will be listed.
+
+REFERENCES_RELATION    = YES
+
+# If the REFERENCES_LINK_SOURCE tag is set to YES (the default)
+# and SOURCE_BROWSER tag is set to YES, then the hyperlinks from
+# functions in REFERENCES_RELATION and REFERENCED_BY_RELATION lists will
+# link to the source code.  Otherwise they will link to the documentstion.
+
+REFERENCES_LINK_SOURCE = YES
+
+# If the USE_HTAGS tag is set to YES then the references to source code
+# will point to the HTML generated by the htags(1) tool instead of doxygen
+# built-in source browser. The htags tool is part of GNU's global source
+# tagging system (see http://www.gnu.org/software/global/global.html). You
+# will need version 4.8.6 or higher.
+
+USE_HTAGS              = NO
+
+# If the VERBATIM_HEADERS tag is set to YES (the default) then Doxygen
+# will generate a verbatim copy of the header file for each class for
+# which an include is specified. Set to NO to disable this.
+
+VERBATIM_HEADERS       = YES
+
+#---------------------------------------------------------------------------
+# configuration options related to the alphabetical class index
+#---------------------------------------------------------------------------
+
+# If the ALPHABETICAL_INDEX tag is set to YES, an alphabetical index
+# of all compounds will be generated. Enable this if the project
+# contains a lot of classes, structs, unions or interfaces.
+
+ALPHABETICAL_INDEX     = NO
+
+# If the alphabetical index is enabled (see ALPHABETICAL_INDEX) then
+# the COLS_IN_ALPHA_INDEX tag can be used to specify the number of columns
+# in which this list will be split (can be a number in the range [1..20])
+
+COLS_IN_ALPHA_INDEX    = 5
+
+# In case all classes in a project start with a common prefix, all
+# classes will be put under the same header in the alphabetical index.
+# The IGNORE_PREFIX tag can be used to specify one or more prefixes that
+# should be ignored while generating the index headers.
+
+IGNORE_PREFIX          =
+
+#---------------------------------------------------------------------------
+# configuration options related to the HTML output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_HTML tag is set to YES (the default) Doxygen will
+# generate HTML output.
+
+GENERATE_HTML          = YES
+
+# The HTML_OUTPUT tag is used to specify where the HTML docs will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be
+# put in front of it. If left blank `html' will be used as the default path.
+
+HTML_OUTPUT            = html
+
+# The HTML_FILE_EXTENSION tag can be used to specify the file extension for
+# each generated HTML page (for example: .htm,.php,.asp). If it is left blank
+# doxygen will generate files with .html extension.
+
+HTML_FILE_EXTENSION    = .html
+
+# The HTML_HEADER tag can be used to specify a personal HTML header for
+# each generated HTML page. If it is left blank doxygen will generate a
+# standard header.
+
+HTML_HEADER            =
+
+# The HTML_FOOTER tag can be used to specify a personal HTML footer for
+# each generated HTML page. If it is left blank doxygen will generate a
+# standard footer.
+
+HTML_FOOTER            =
+
+# The HTML_STYLESHEET tag can be used to specify a user-defined cascading
+# style sheet that is used by each HTML page. It can be used to
+# fine-tune the look of the HTML output. If the tag is left blank doxygen
+# will generate a default style sheet. Note that doxygen will try to copy
+# the style sheet file to the HTML output directory, so don't put your own
+# stylesheet in the HTML output directory as well, or it will be erased!
+
+HTML_STYLESHEET        =
+
+# If the GENERATE_HTMLHELP tag is set to YES, additional index files
+# will be generated that can be used as input for tools like the
+# Microsoft HTML help workshop to generate a compressed HTML help file (.chm)
+# of the generated HTML documentation.
+
+GENERATE_HTMLHELP      = NO
+
+# If the HTML_DYNAMIC_SECTIONS tag is set to YES then the generated HTML
+# documentation will contain sections that can be hidden and shown after the
+# page has loaded. For this to work a browser that supports
+# java_script and DHTML is required (for instance Mozilla 1.0+, Firefox
+# Netscape 6.0+, Internet explorer 5.0+, Konqueror, or Safari).
+
+HTML_DYNAMIC_SECTIONS  = NO
+
+# If the GENERATE_HTMLHELP tag is set to YES, the CHM_FILE tag can
+# be used to specify the file name of the resulting .chm file. You
+# can add a path in front of the file if the result should not be
+# written to the html output directory.
+
+CHM_FILE               =
+
+# If the GENERATE_HTMLHELP tag is set to YES, the HHC_LOCATION tag can
+# be used to specify the location (absolute path including file name) of
+# the HTML help compiler (hhc.exe). If non-empty doxygen will try to run
+# the HTML help compiler on the generated index.hhp.
+
+HHC_LOCATION           =
+
+# If the GENERATE_HTMLHELP tag is set to YES, the GENERATE_CHI flag
+# controls if a separate .chi index file is generated (YES) or that
+# it should be included in the master .chm file (NO).
+
+GENERATE_CHI           = NO
+
+# If the GENERATE_HTMLHELP tag is set to YES, the BINARY_TOC flag
+# controls whether a binary table of contents is generated (YES) or a
+# normal table of contents (NO) in the .chm file.
+
+BINARY_TOC             = NO
+
+# The TOC_EXPAND flag can be set to YES to add extra items for group members
+# to the contents of the HTML help documentation and to the tree view.
+
+TOC_EXPAND             = NO
+
+# The DISABLE_INDEX tag can be used to turn on/off the condensed index at
+# top of each HTML page. The value NO (the default) enables the index and
+# the value YES disables it.
+
+DISABLE_INDEX          = NO
+
+# This tag can be used to set the number of enum values (range [1..20])
+# that doxygen will group on one line in the generated HTML documentation.
+
+ENUM_VALUES_PER_LINE   = 4
+
+# If the GENERATE_TREEVIEW tag is set to YES, a side panel will be
+# generated containing a tree-like index structure (just like the one that
+# is generated for HTML Help). For this to work a browser that supports
+# java_script, DHTML, CSS and frames is required (for instance Mozilla 1.0+,
+# Netscape 6.0+, Internet explorer 5.0+, or Konqueror). Windows users are
+# probably better off using the HTML help feature.
+
+GENERATE_TREEVIEW      = NO
+
+# If the treeview is enabled (see GENERATE_TREEVIEW) then this tag can be
+# used to set the initial width (in pixels) of the frame in which the tree
+# is shown.
+
+TREEVIEW_WIDTH         = 250
+
+#---------------------------------------------------------------------------
+# configuration options related to the la_te_x output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_LATEX tag is set to YES (the default) Doxygen will
+# generate Latex output.
+
+GENERATE_LATEX         = YES
+
+# The LATEX_OUTPUT tag is used to specify where the la_te_x docs will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be
+# put in front of it. If left blank `latex' will be used as the default path.
+
+LATEX_OUTPUT           = latex
+
+# The LATEX_CMD_NAME tag can be used to specify the la_te_x command name to be
+# invoked. If left blank `latex' will be used as the default command name.
+
+LATEX_CMD_NAME         = latex
+
+# The MAKEINDEX_CMD_NAME tag can be used to specify the command name to
+# generate index for la_te_x. If left blank `makeindex' will be used as the
+# default command name.
+
+MAKEINDEX_CMD_NAME     = makeindex
+
+# If the COMPACT_LATEX tag is set to YES Doxygen generates more compact
+# la_te_x documents. This may be useful for small projects and may help to
+# save some trees in general.
+
+COMPACT_LATEX          = YES
+
+# The PAPER_TYPE tag can be used to set the paper type that is used
+# by the printer. Possible values are: a4, a4wide, letter, legal and
+# executive. If left blank a4wide will be used.
+
+PAPER_TYPE             = letter
+
+# The EXTRA_PACKAGES tag can be to specify one or more names of la_te_x
+# packages that should be included in the la_te_x output.
+
+EXTRA_PACKAGES         =
+
+# The LATEX_HEADER tag can be used to specify a personal la_te_x header for
+# the generated latex document. The header should contain everything until
+# the first chapter. If it is left blank doxygen will generate a
+# standard header. Notice: only use this tag if you know what you are doing!
+
+LATEX_HEADER           =
+
+# If the PDF_HYPERLINKS tag is set to YES, the la_te_x that is generated
+# is prepared for conversion to pdf (using ps2pdf). The pdf file will
+# contain links (just like the HTML output) instead of page references
+# This makes the output suitable for online browsing using a pdf viewer.
+
+PDF_HYPERLINKS         = YES
+
+# If the USE_PDFLATEX tag is set to YES, pdflatex will be used instead of
+# plain latex in the generated Makefile. Set this option to YES to get a
+# higher quality PDF documentation.
+
+USE_PDFLATEX           = YES
+
+# If the LATEX_BATCHMODE tag is set to YES, doxygen will add the \\batchmode.
+# command to the generated la_te_x files. This will instruct la_te_x to keep
+# running if errors occur, instead of asking the user for help.
+# This option is also used when generating formulas in HTML.
+
+LATEX_BATCHMODE        = NO
+
+# If LATEX_HIDE_INDICES is set to YES then doxygen will not
+# include the index chapters (such as File Index, Compound Index, etc.)
+# in the output.
+
+LATEX_HIDE_INDICES     = NO
+
+#---------------------------------------------------------------------------
+# configuration options related to the RTF output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_RTF tag is set to YES Doxygen will generate RTF output
+# The RTF output is optimized for Word 97 and may not look very pretty with
+# other RTF readers or editors.
+
+GENERATE_RTF           = NO
+
+# The RTF_OUTPUT tag is used to specify where the RTF docs will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be
+# put in front of it. If left blank `rtf' will be used as the default path.
+
+RTF_OUTPUT             = rtf
+
+# If the COMPACT_RTF tag is set to YES Doxygen generates more compact
+# RTF documents. This may be useful for small projects and may help to
+# save some trees in general.
+
+COMPACT_RTF            = NO
+
+# If the RTF_HYPERLINKS tag is set to YES, the RTF that is generated
+# will contain hyperlink fields. The RTF file will
+# contain links (just like the HTML output) instead of page references.
+# This makes the output suitable for online browsing using WORD or other
+# programs which support those fields.
+# Note: wordpad (write) and others do not support links.
+
+RTF_HYPERLINKS         = NO
+
+# Load stylesheet definitions from file. Syntax is similar to doxygen's
+# config file, i.e. a series of assignments. You only have to provide
+# replacements, missing definitions are set to their default value.
+
+RTF_STYLESHEET_FILE    =
+
+# Set optional variables used in the generation of an rtf document.
+# Syntax is similar to doxygen's config file.
+
+RTF_EXTENSIONS_FILE    =
+
+#---------------------------------------------------------------------------
+# configuration options related to the man page output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_MAN tag is set to YES (the default) Doxygen will
+# generate man pages
+
+GENERATE_MAN           = NO
+
+# The MAN_OUTPUT tag is used to specify where the man pages will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be
+# put in front of it. If left blank `man' will be used as the default path.
+
+MAN_OUTPUT             = man
+
+# The MAN_EXTENSION tag determines the extension that is added to
+# the generated man pages (default is the subroutine's section .3)
+
+MAN_EXTENSION          = .3
+
+# If the MAN_LINKS tag is set to YES and Doxygen generates man output,
+# then it will generate one additional man file for each entity
+# documented in the real man page(s). These additional files
+# only source the real man page, but without them the man command
+# would be unable to find the correct page. The default is NO.
+
+MAN_LINKS              = YES
+
+#---------------------------------------------------------------------------
+# configuration options related to the XML output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_XML tag is set to YES Doxygen will
+# generate an XML file that captures the structure of
+# the code including all documentation.
+
+GENERATE_XML           = NO
+
+# The XML_OUTPUT tag is used to specify where the XML pages will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be
+# put in front of it. If left blank `xml' will be used as the default path.
+
+XML_OUTPUT             = xml
+
+# The XML_SCHEMA tag can be used to specify an XML schema,
+# which can be used by a validating XML parser to check the
+# syntax of the XML files.
+
+XML_SCHEMA             =
+
+# The XML_DTD tag can be used to specify an XML DTD,
+# which can be used by a validating XML parser to check the
+# syntax of the XML files.
+
+XML_DTD                =
+
+# If the XML_PROGRAMLISTING tag is set to YES Doxygen will
+# dump the program listings (including syntax highlighting
+# and cross-referencing information) to the XML output. Note that
+# enabling this will significantly increase the size of the XML output.
+
+XML_PROGRAMLISTING     = YES
+
+#---------------------------------------------------------------------------
+# configuration options for the auto_gen Definitions output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_AUTOGEN_DEF tag is set to YES Doxygen will
+# generate an auto_gen Definitions (see autogen.sf.net) file
+# that captures the structure of the code including all
+# documentation. Note that this feature is still experimental
+# and incomplete at the moment.
+
+GENERATE_AUTOGEN_DEF   = NO
+
+#---------------------------------------------------------------------------
+# configuration options related to the Perl module output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_PERLMOD tag is set to YES Doxygen will
+# generate a Perl module file that captures the structure of
+# the code including all documentation. Note that this
+# feature is still experimental and incomplete at the
+# moment.
+
+GENERATE_PERLMOD       = NO
+
+# If the PERLMOD_LATEX tag is set to YES Doxygen will generate
+# the necessary Makefile rules, Perl scripts and la_te_x code to be able
+# to generate PDF and DVI output from the Perl module output.
+
+PERLMOD_LATEX          = NO
+
+# If the PERLMOD_PRETTY tag is set to YES the Perl module output will be
+# nicely formatted so it can be parsed by a human reader.  This is useful
+# if you want to understand what is going on.  On the other hand, if this
+# tag is set to NO the size of the Perl module output will be much smaller
+# and Perl will parse it just the same.
+
+PERLMOD_PRETTY         = YES
+
+# The names of the make variables in the generated doxyrules.make file
+# are prefixed with the string contained in PERLMOD_MAKEVAR_PREFIX.
+# This is useful so different doxyrules.make files included by the same
+# Makefile don't overwrite each other's variables.
+
+PERLMOD_MAKEVAR_PREFIX =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the preprocessor
+#---------------------------------------------------------------------------
+
+# If the ENABLE_PREPROCESSING tag is set to YES (the default) Doxygen will
+# evaluate all C-preprocessor directives found in the sources and include
+# files.
+
+ENABLE_PREPROCESSING   = YES
+
+# If the MACRO_EXPANSION tag is set to YES Doxygen will expand all macro
+# names in the source code. If set to NO (the default) only conditional
+# compilation will be performed. Macro expansion can be done in a controlled
+# way by setting EXPAND_ONLY_PREDEF to YES.
+
+MACRO_EXPANSION        = YES
+
+# If the EXPAND_ONLY_PREDEF and MACRO_EXPANSION tags are both set to YES
+# then the macro expansion is limited to the macros specified with the
+# PREDEFINED and EXPAND_AS_DEFINED tags.
+
+EXPAND_ONLY_PREDEF     = NO
+
+# If the SEARCH_INCLUDES tag is set to YES (the default) the includes files
+# in the INCLUDE_PATH (see below) will be search if a #include is found.
+
+SEARCH_INCLUDES        = YES
+
+# The INCLUDE_PATH tag can be used to specify one or more directories that
+# contain include files that are not input files but should be processed by
+# the preprocessor.
+
+INCLUDE_PATH           =
+
+# You can use the INCLUDE_FILE_PATTERNS tag to specify one or more wildcard
+# patterns (like *.h and *.hpp) to filter out the header-files in the
+# directories. If left blank, the patterns specified with FILE_PATTERNS will
+# be used.
+
+INCLUDE_FILE_PATTERNS  = *.h
+
+# The PREDEFINED tag can be used to specify one or more macro names that
+# are defined before the preprocessor is started (similar to the -D option of
+# gcc). The argument of the tag is a list of macros of the form: name
+# or name=definition (no spaces). If the definition and the = are
+# omitted =1 is assumed. To prevent a macro definition from being
+# undefined via #undef or recursively expanded use the := operator
+# instead of the = operator.
+
+PREDEFINED             =
+
+# If the MACRO_EXPANSION and EXPAND_ONLY_PREDEF tags are set to YES then
+# this tag can be used to specify a list of macro names that should be expanded.
+# The macro definition that is found in the sources will be used.
+# Use the PREDEFINED tag if you want to use a different macro definition.
+
+EXPAND_AS_DEFINED      =
+
+# If the SKIP_FUNCTION_MACROS tag is set to YES (the default) then
+# doxygen's preprocessor will remove all function-like macros that are alone
+# on a line, have an all uppercase name, and do not end with a semicolon. Such
+# function macros are typically used for boiler-plate code, and will confuse
+# the parser if not removed.
+
+SKIP_FUNCTION_MACROS   = YES
+
+#---------------------------------------------------------------------------
+# Configuration::additions related to external references
+#---------------------------------------------------------------------------
+
+# The TAGFILES option can be used to specify one or more tagfiles.
+# Optionally an initial location of the external documentation
+# can be added for each tagfile. The format of a tag file without
+# this location is as follows:
+#   TAGFILES = file1 file2 ...
+# Adding location for the tag files is done as follows:
+#   TAGFILES = file1=loc1 "file2 = loc2" ...
+# where "loc1" and "loc2" can be relative or absolute paths or
+# URLs. If a location is present for each tag, the installdox tool
+# does not have to be run to correct the links.
+# Note that each tag file must have a unique name
+# (where the name does NOT include the path)
+# If a tag file is not located in the directory in which doxygen
+# is run, you must also specify the path to the tagfile here.
+
+TAGFILES               =
+
+# When a file name is specified after GENERATE_TAGFILE, doxygen will create
+# a tag file that is based on the input files it reads.
+
+GENERATE_TAGFILE       =
+
+# If the ALLEXTERNALS tag is set to YES all external classes will be listed
+# in the class index. If set to NO only the inherited external classes
+# will be listed.
+
+ALLEXTERNALS           = NO
+
+# If the EXTERNAL_GROUPS tag is set to YES all external groups will be listed
+# in the modules index. If set to NO, only the current project's groups will
+# be listed.
+
+EXTERNAL_GROUPS        = YES
+
+# The PERL_PATH should be the absolute path and name of the perl script
+# interpreter (i.e. the result of `which perl').
+
+PERL_PATH              = /usr/bin/perl
+
+#---------------------------------------------------------------------------
+# Configuration options related to the dot tool
+#---------------------------------------------------------------------------
+
+# If the CLASS_DIAGRAMS tag is set to YES (the default) Doxygen will
+# generate a inheritance diagram (in HTML, RTF and la_te_x) for classes with base
+# or super classes. Setting the tag to NO turns the diagrams off. Note that
+# this option is superseded by the HAVE_DOT option below. This is only a
+# fallback. It is recommended to install and use dot, since it yields more
+# powerful graphs.
+
+CLASS_DIAGRAMS         = YES
+
+# You can define message sequence charts within doxygen comments using the \msc
+# command. Doxygen will then run the mscgen tool (see http://www.mcternan.me.uk/mscgen/) to
+# produce the chart and insert it in the documentation. The MSCGEN_PATH tag allows you to
+# specify the directory where the mscgen tool resides. If left empty the tool is assumed to
+# be found in the default search path.
+
+MSCGEN_PATH            =
+
+# If set to YES, the inheritance and collaboration graphs will hide
+# inheritance and usage relations if the target is undocumented
+# or is not a class.
+
+HIDE_UNDOC_RELATIONS   = YES
+
+# If you set the HAVE_DOT tag to YES then doxygen will assume the dot tool is
+# available from the path. This tool is part of Graphviz, a graph visualization
+# toolkit from AT&T and Lucent Bell Labs. The other options in this section
+# have no effect if this option is set to NO (the default)
+
+HAVE_DOT               = NO
+
+# If the CLASS_GRAPH and HAVE_DOT tags are set to YES then doxygen
+# will generate a graph for each documented class showing the direct and
+# indirect inheritance relations. Setting this tag to YES will force the
+# the CLASS_DIAGRAMS tag to NO.
+
+CLASS_GRAPH            = YES
+
+# If the COLLABORATION_GRAPH and HAVE_DOT tags are set to YES then doxygen
+# will generate a graph for each documented class showing the direct and
+# indirect implementation dependencies (inheritance, containment, and
+# class references variables) of the class with other documented classes.
+
+COLLABORATION_GRAPH    = YES
+
+# If the GROUP_GRAPHS and HAVE_DOT tags are set to YES then doxygen
+# will generate a graph for groups, showing the direct groups dependencies
+
+GROUP_GRAPHS           = YES
+
+# If the UML_LOOK tag is set to YES doxygen will generate inheritance and
+# collaboration diagrams in a style similar to the OMG's Unified Modeling
+# Language.
+
+UML_LOOK               = NO
+
+# If set to YES, the inheritance and collaboration graphs will show the
+# relations between templates and their instances.
+
+TEMPLATE_RELATIONS     = NO
+
+# If the ENABLE_PREPROCESSING, SEARCH_INCLUDES, INCLUDE_GRAPH, and HAVE_DOT
+# tags are set to YES then doxygen will generate a graph for each documented
+# file showing the direct and indirect include dependencies of the file with
+# other documented files.
+
+INCLUDE_GRAPH          = YES
+
+# If the ENABLE_PREPROCESSING, SEARCH_INCLUDES, INCLUDED_BY_GRAPH, and
+# HAVE_DOT tags are set to YES then doxygen will generate a graph for each
+# documented header file showing the documented files that directly or
+# indirectly include this file.
+
+INCLUDED_BY_GRAPH      = YES
+
+# If the CALL_GRAPH, SOURCE_BROWSER and HAVE_DOT tags are set to YES then doxygen will
+# generate a call dependency graph for every global function or class method.
+# Note that enabling this option will significantly increase the time of a run.
+# So in most cases it will be better to enable call graphs for selected
+# functions only using the \callgraph command.
+
+CALL_GRAPH             = NO
+
+# If the CALLER_GRAPH, SOURCE_BROWSER and HAVE_DOT tags are set to YES then doxygen will
+# generate a caller dependency graph for every global function or class method.
+# Note that enabling this option will significantly increase the time of a run.
+# So in most cases it will be better to enable caller graphs for selected
+# functions only using the \callergraph command.
+
+CALLER_GRAPH           = NO
+
+# If the GRAPHICAL_HIERARCHY and HAVE_DOT tags are set to YES then doxygen
+# will graphical hierarchy of all classes instead of a textual one.
+
+GRAPHICAL_HIERARCHY    = YES
+
+# If the DIRECTORY_GRAPH, SHOW_DIRECTORIES and HAVE_DOT tags are set to YES
+# then doxygen will show the dependencies a directory has on other directories
+# in a graphical way. The dependency relations are determined by the #include
+# relations between the files in the directories.
+
+DIRECTORY_GRAPH        = YES
+
+# The DOT_IMAGE_FORMAT tag can be used to set the image format of the images
+# generated by dot. Possible values are png, jpg, or gif
+# If left blank png will be used.
+
+DOT_IMAGE_FORMAT       = png
+
+# The tag DOT_PATH can be used to specify the path where the dot tool can be
+# found. If left blank, it is assumed the dot tool can be found in the path.
+
+DOT_PATH               =
+
+# The DOTFILE_DIRS tag can be used to specify one or more directories that
+# contain dot files that are included in the documentation (see the
+# \dotfile command).
+
+DOTFILE_DIRS           =
+
+# The MAX_DOT_GRAPH_MAX_NODES tag can be used to set the maximum number of
+# nodes that will be shown in the graph. If the number of nodes in a graph
+# becomes larger than this value, doxygen will truncate the graph, which is
+# visualized by representing a node as a red box. Note that doxygen if the number
+# of direct children of the root node in a graph is already larger than
+# MAX_DOT_GRAPH_NOTES then the graph will not be shown at all. Also note
+# that the size of a graph can be further restricted by MAX_DOT_GRAPH_DEPTH.
+
+DOT_GRAPH_MAX_NODES    = 50
+
+# The MAX_DOT_GRAPH_DEPTH tag can be used to set the maximum depth of the
+# graphs generated by dot. A depth value of 3 means that only nodes reachable
+# from the root by following a path via at most 3 edges will be shown. Nodes
+# that lay further from the root node will be omitted. Note that setting this
+# option to 1 or 2 may greatly reduce the computation time needed for large
+# code bases. Also note that the size of a graph can be further restricted by
+# DOT_GRAPH_MAX_NODES. Using a depth of 0 means no depth restriction.
+
+MAX_DOT_GRAPH_DEPTH    = 0
+
+# Set the DOT_TRANSPARENT tag to YES to generate images with a transparent
+# background. This is disabled by default, which results in a white background.
+# Warning: Depending on the platform used, enabling this option may lead to
+# badly anti-aliased labels on the edges of a graph (i.e. they become hard to
+# read).
+
+DOT_TRANSPARENT        = YES
+
+# Set the DOT_MULTI_TARGETS tag to YES allow dot to generate multiple output
+# files in one run (i.e. multiple -o and -T options on the command line). This
+# makes dot run faster, but since only newer versions of dot (>1.8.10)
+# support this, this feature is disabled by default.
+
+DOT_MULTI_TARGETS      = NO
+
+# If the GENERATE_LEGEND tag is set to YES (the default) Doxygen will
+# generate a legend page explaining the meaning of the various boxes and
+# arrows in the dot generated graphs.
+
+GENERATE_LEGEND        = YES
+
+# If the DOT_CLEANUP tag is set to YES (the default) Doxygen will
+# remove the intermediate dot files that are used to generate
+# the various graphs.
+
+DOT_CLEANUP            = YES
+
+#---------------------------------------------------------------------------
+# Configuration::additions related to the search engine
+#---------------------------------------------------------------------------
+
+# The SEARCHENGINE tag specifies whether or not a search engine should be
+# used. If set to NO the values of all tags below this one will be ignored.
+
+SEARCHENGINE           = NO
diff --git a/libvpx/libvpx/libs.mk b/libvpx/libvpx/libs.mk
new file mode 100644
index 0000000..9a6092a
--- /dev/null
+++ b/libvpx/libvpx/libs.mk
@@ -0,0 +1,615 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+# ARM assembly files are written in RVCT-style. We use some make magic to
+# filter those files to allow GCC compilation
+ifeq ($(ARCH_ARM),yes)
+  ASM:=$(if $(filter yes,$(CONFIG_GCC)$(CONFIG_MSVS)),.asm.s,.asm)
+else
+  ASM:=.asm
+endif
+
+#
+# Rule to generate runtime cpu detection files
+#
+define rtcd_h_template
+$$(BUILD_PFX)$(1).h: $$(SRC_PATH_BARE)/$(2)
+	@echo "    [CREATE] $$@"
+	$$(qexec)$$(SRC_PATH_BARE)/build/make/rtcd.pl --arch=$$(TGT_ISA) \
+          --sym=$(1) \
+          --config=$$(CONFIG_DIR)$$(target)-$$(TOOLCHAIN).mk \
+          $$(RTCD_OPTIONS) $$^ > $$@
+CLEAN-OBJS += $$(BUILD_PFX)$(1).h
+RTCD += $$(BUILD_PFX)$(1).h
+endef
+
+CODEC_SRCS-yes += CHANGELOG
+CODEC_SRCS-yes += libs.mk
+
+include $(SRC_PATH_BARE)/vpx/vpx_codec.mk
+CODEC_SRCS-yes += $(addprefix vpx/,$(call enabled,API_SRCS))
+CODEC_DOC_SRCS += $(addprefix vpx/,$(call enabled,API_DOC_SRCS))
+
+include $(SRC_PATH_BARE)/vpx_mem/vpx_mem.mk
+CODEC_SRCS-yes += $(addprefix vpx_mem/,$(call enabled,MEM_SRCS))
+
+include $(SRC_PATH_BARE)/vpx_scale/vpx_scale.mk
+CODEC_SRCS-yes += $(addprefix vpx_scale/,$(call enabled,SCALE_SRCS))
+
+include $(SRC_PATH_BARE)/vpx_ports/vpx_ports.mk
+CODEC_SRCS-yes += $(addprefix vpx_ports/,$(call enabled,PORTS_SRCS))
+
+include $(SRC_PATH_BARE)/vpx_dsp/vpx_dsp.mk
+CODEC_SRCS-yes += $(addprefix vpx_dsp/,$(call enabled,DSP_SRCS))
+
+include $(SRC_PATH_BARE)/vpx_util/vpx_util.mk
+CODEC_SRCS-yes += $(addprefix vpx_util/,$(call enabled,UTIL_SRCS))
+
+ifeq ($(CONFIG_VP8),yes)
+  VP8_PREFIX=vp8/
+  include $(SRC_PATH_BARE)/$(VP8_PREFIX)vp8_common.mk
+endif
+
+ifeq ($(CONFIG_VP8_ENCODER),yes)
+  include $(SRC_PATH_BARE)/$(VP8_PREFIX)vp8cx.mk
+  CODEC_SRCS-yes += $(addprefix $(VP8_PREFIX),$(call enabled,VP8_CX_SRCS))
+  CODEC_EXPORTS-yes += $(addprefix $(VP8_PREFIX),$(VP8_CX_EXPORTS))
+  INSTALL-LIBS-yes += include/vpx/vp8.h include/vpx/vp8cx.h
+  INSTALL_MAPS += include/vpx/% $(SRC_PATH_BARE)/$(VP8_PREFIX)/%
+  CODEC_DOC_SECTIONS += vp8 vp8_encoder
+endif
+
+ifeq ($(CONFIG_VP8_DECODER),yes)
+  include $(SRC_PATH_BARE)/$(VP8_PREFIX)vp8dx.mk
+  CODEC_SRCS-yes += $(addprefix $(VP8_PREFIX),$(call enabled,VP8_DX_SRCS))
+  CODEC_EXPORTS-yes += $(addprefix $(VP8_PREFIX),$(VP8_DX_EXPORTS))
+  INSTALL-LIBS-yes += include/vpx/vp8.h include/vpx/vp8dx.h
+  INSTALL_MAPS += include/vpx/% $(SRC_PATH_BARE)/$(VP8_PREFIX)/%
+  CODEC_DOC_SECTIONS += vp8 vp8_decoder
+endif
+
+ifeq ($(CONFIG_VP9),yes)
+  VP9_PREFIX=vp9/
+  include $(SRC_PATH_BARE)/$(VP9_PREFIX)vp9_common.mk
+endif
+
+ifeq ($(CONFIG_VP9_ENCODER),yes)
+  VP9_PREFIX=vp9/
+  include $(SRC_PATH_BARE)/$(VP9_PREFIX)vp9cx.mk
+  CODEC_SRCS-yes += $(addprefix $(VP9_PREFIX),$(call enabled,VP9_CX_SRCS))
+  CODEC_EXPORTS-yes += $(addprefix $(VP9_PREFIX),$(VP9_CX_EXPORTS))
+  CODEC_SRCS-yes += $(VP9_PREFIX)vp9cx.mk vpx/vp8.h vpx/vp8cx.h
+  INSTALL-LIBS-yes += include/vpx/vp8.h include/vpx/vp8cx.h
+  INSTALL-LIBS-$(CONFIG_SPATIAL_SVC) += include/vpx/svc_context.h
+  INSTALL_MAPS += include/vpx/% $(SRC_PATH_BARE)/$(VP9_PREFIX)/%
+  CODEC_DOC_SRCS += vpx/vp8.h vpx/vp8cx.h
+  CODEC_DOC_SECTIONS += vp9 vp9_encoder
+endif
+
+ifeq ($(CONFIG_VP9_DECODER),yes)
+  VP9_PREFIX=vp9/
+  include $(SRC_PATH_BARE)/$(VP9_PREFIX)vp9dx.mk
+  CODEC_SRCS-yes += $(addprefix $(VP9_PREFIX),$(call enabled,VP9_DX_SRCS))
+  CODEC_EXPORTS-yes += $(addprefix $(VP9_PREFIX),$(VP9_DX_EXPORTS))
+  CODEC_SRCS-yes += $(VP9_PREFIX)vp9dx.mk vpx/vp8.h vpx/vp8dx.h
+  INSTALL-LIBS-yes += include/vpx/vp8.h include/vpx/vp8dx.h
+  INSTALL_MAPS += include/vpx/% $(SRC_PATH_BARE)/$(VP9_PREFIX)/%
+  CODEC_DOC_SRCS += vpx/vp8.h vpx/vp8dx.h
+  CODEC_DOC_SECTIONS += vp9 vp9_decoder
+endif
+
+VP9_PREFIX=vp9/
+$(BUILD_PFX)$(VP9_PREFIX)%.c.o: CFLAGS += -Wextra
+
+ifeq ($(CONFIG_ENCODERS),yes)
+  CODEC_DOC_SECTIONS += encoder
+endif
+ifeq ($(CONFIG_DECODERS),yes)
+  CODEC_DOC_SECTIONS += decoder
+endif
+
+
+ifeq ($(CONFIG_MSVS),yes)
+CODEC_LIB=$(if $(CONFIG_STATIC_MSVCRT),vpxmt,vpxmd)
+GTEST_LIB=$(if $(CONFIG_STATIC_MSVCRT),gtestmt,gtestmd)
+# This variable uses deferred expansion intentionally, since the results of
+# $(wildcard) may change during the course of the Make.
+VS_PLATFORMS = $(foreach d,$(wildcard */Release/$(CODEC_LIB).lib),$(word 1,$(subst /, ,$(d))))
+endif
+
+# The following pairs define a mapping of locations in the distribution
+# tree to locations in the source/build trees.
+INSTALL_MAPS += include/vpx/% $(SRC_PATH_BARE)/vpx/%
+INSTALL_MAPS += include/vpx/% $(SRC_PATH_BARE)/vpx_ports/%
+INSTALL_MAPS += $(LIBSUBDIR)/%     %
+INSTALL_MAPS += src/%     $(SRC_PATH_BARE)/%
+ifeq ($(CONFIG_MSVS),yes)
+INSTALL_MAPS += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/%  $(p)/Release/%)
+INSTALL_MAPS += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/%  $(p)/Debug/%)
+endif
+
+CODEC_SRCS-yes += build/make/version.sh
+CODEC_SRCS-yes += build/make/rtcd.pl
+CODEC_SRCS-yes += vpx_ports/emmintrin_compat.h
+CODEC_SRCS-yes += vpx_ports/mem_ops.h
+CODEC_SRCS-yes += vpx_ports/mem_ops_aligned.h
+CODEC_SRCS-yes += vpx_ports/vpx_once.h
+CODEC_SRCS-yes += $(BUILD_PFX)vpx_config.c
+INSTALL-SRCS-no += $(BUILD_PFX)vpx_config.c
+ifeq ($(ARCH_X86)$(ARCH_X86_64),yes)
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += third_party/x86inc/x86inc.asm
+endif
+CODEC_EXPORTS-yes += vpx/exports_com
+CODEC_EXPORTS-$(CONFIG_ENCODERS) += vpx/exports_enc
+ifeq ($(CONFIG_SPATIAL_SVC),yes)
+CODEC_EXPORTS-$(CONFIG_ENCODERS) += vpx/exports_spatial_svc
+endif
+CODEC_EXPORTS-$(CONFIG_DECODERS) += vpx/exports_dec
+
+INSTALL-LIBS-yes += include/vpx/vpx_codec.h
+INSTALL-LIBS-yes += include/vpx/vpx_frame_buffer.h
+INSTALL-LIBS-yes += include/vpx/vpx_image.h
+INSTALL-LIBS-yes += include/vpx/vpx_integer.h
+INSTALL-LIBS-$(CONFIG_DECODERS) += include/vpx/vpx_decoder.h
+INSTALL-LIBS-$(CONFIG_ENCODERS) += include/vpx/vpx_encoder.h
+ifeq ($(CONFIG_EXTERNAL_BUILD),yes)
+ifeq ($(CONFIG_MSVS),yes)
+INSTALL-LIBS-yes                  += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/$(CODEC_LIB).lib)
+INSTALL-LIBS-$(CONFIG_DEBUG_LIBS) += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/$(CODEC_LIB)d.lib)
+INSTALL-LIBS-$(CONFIG_SHARED) += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/vpx.dll)
+INSTALL-LIBS-$(CONFIG_SHARED) += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/vpx.exp)
+endif
+else
+INSTALL-LIBS-$(CONFIG_STATIC) += $(LIBSUBDIR)/libvpx.a
+INSTALL-LIBS-$(CONFIG_DEBUG_LIBS) += $(LIBSUBDIR)/libvpx_g.a
+endif
+
+CODEC_SRCS=$(call enabled,CODEC_SRCS)
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += $(CODEC_SRCS)
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += $(call enabled,CODEC_EXPORTS)
+
+
+# Generate a list of all enabled sources, in particular for exporting to gyp
+# based build systems.
+libvpx_srcs.txt:
+	@echo "    [CREATE] $@"
+	@echo $(CODEC_SRCS) | xargs -n1 echo | LC_ALL=C sort -u > $@
+CLEAN-OBJS += libvpx_srcs.txt
+
+
+ifeq ($(CONFIG_EXTERNAL_BUILD),yes)
+ifeq ($(CONFIG_MSVS),yes)
+
+vpx.def: $(call enabled,CODEC_EXPORTS)
+	@echo "    [CREATE] $@"
+	$(qexec)$(SRC_PATH_BARE)/build/make/gen_msvs_def.sh\
+            --name=vpx\
+            --out=$@ $^
+CLEAN-OBJS += vpx.def
+
+# Assembly files that are included, but don't define symbols themselves.
+# Filtered out to avoid Visual Studio build warnings.
+ASM_INCLUDES := \
+    third_party/x86inc/x86inc.asm \
+    vpx_config.asm \
+    vpx_ports/x86_abi_support.asm \
+
+vpx.$(VCPROJ_SFX): $(CODEC_SRCS) vpx.def
+	@echo "    [CREATE] $@"
+	$(qexec)$(GEN_VCPROJ) \
+            $(if $(CONFIG_SHARED),--dll,--lib) \
+            --target=$(TOOLCHAIN) \
+            $(if $(CONFIG_STATIC_MSVCRT),--static-crt) \
+            --name=vpx \
+            --proj-guid=DCE19DAF-69AC-46DB-B14A-39F0FAA5DB74 \
+            --module-def=vpx.def \
+            --ver=$(CONFIG_VS_VERSION) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
+            --out=$@ $(CFLAGS) \
+            $(filter-out $(addprefix %, $(ASM_INCLUDES)), $^) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
+
+PROJECTS-yes += vpx.$(VCPROJ_SFX)
+
+vpx.$(VCPROJ_SFX): vpx_config.asm
+vpx.$(VCPROJ_SFX): $(RTCD)
+
+endif
+else
+LIBVPX_OBJS=$(call objs,$(CODEC_SRCS))
+OBJS-yes += $(LIBVPX_OBJS)
+LIBS-$(if yes,$(CONFIG_STATIC)) += $(BUILD_PFX)libvpx.a $(BUILD_PFX)libvpx_g.a
+$(BUILD_PFX)libvpx_g.a: $(LIBVPX_OBJS)
+
+SO_VERSION_MAJOR := 4
+SO_VERSION_MINOR := 0
+SO_VERSION_PATCH := 0
+ifeq ($(filter darwin%,$(TGT_OS)),$(TGT_OS))
+LIBVPX_SO               := libvpx.$(SO_VERSION_MAJOR).dylib
+SHARED_LIB_SUF          := .dylib
+EXPORT_FILE             := libvpx.syms
+LIBVPX_SO_SYMLINKS      := $(addprefix $(LIBSUBDIR)/, \
+                             libvpx.dylib  )
+else
+ifeq ($(filter iphonesimulator%,$(TGT_OS)),$(TGT_OS))
+LIBVPX_SO               := libvpx.$(SO_VERSION_MAJOR).dylib
+SHARED_LIB_SUF          := .dylib
+EXPORT_FILE             := libvpx.syms
+LIBVPX_SO_SYMLINKS      := $(addprefix $(LIBSUBDIR)/, libvpx.dylib)
+else
+ifeq ($(filter os2%,$(TGT_OS)),$(TGT_OS))
+LIBVPX_SO               := libvpx$(SO_VERSION_MAJOR).dll
+SHARED_LIB_SUF          := _dll.a
+EXPORT_FILE             := libvpx.def
+LIBVPX_SO_SYMLINKS      :=
+LIBVPX_SO_IMPLIB        := libvpx_dll.a
+else
+LIBVPX_SO               := libvpx.so.$(SO_VERSION_MAJOR).$(SO_VERSION_MINOR).$(SO_VERSION_PATCH)
+SHARED_LIB_SUF          := .so
+EXPORT_FILE             := libvpx.ver
+LIBVPX_SO_SYMLINKS      := $(addprefix $(LIBSUBDIR)/, \
+                             libvpx.so libvpx.so.$(SO_VERSION_MAJOR) \
+                             libvpx.so.$(SO_VERSION_MAJOR).$(SO_VERSION_MINOR))
+endif
+endif
+endif
+
+LIBS-$(CONFIG_SHARED) += $(BUILD_PFX)$(LIBVPX_SO)\
+                           $(notdir $(LIBVPX_SO_SYMLINKS)) \
+                           $(if $(LIBVPX_SO_IMPLIB), $(BUILD_PFX)$(LIBVPX_SO_IMPLIB))
+$(BUILD_PFX)$(LIBVPX_SO): $(LIBVPX_OBJS) $(EXPORT_FILE)
+$(BUILD_PFX)$(LIBVPX_SO): extralibs += -lm
+$(BUILD_PFX)$(LIBVPX_SO): SONAME = libvpx.so.$(SO_VERSION_MAJOR)
+$(BUILD_PFX)$(LIBVPX_SO): EXPORTS_FILE = $(EXPORT_FILE)
+
+libvpx.ver: $(call enabled,CODEC_EXPORTS)
+	@echo "    [CREATE] $@"
+	$(qexec)echo "{ global:" > $@
+	$(qexec)for f in $?; do awk '{print $$2";"}' < $$f >>$@; done
+	$(qexec)echo "local: *; };" >> $@
+CLEAN-OBJS += libvpx.ver
+
+libvpx.syms: $(call enabled,CODEC_EXPORTS)
+	@echo "    [CREATE] $@"
+	$(qexec)awk '{print "_"$$2}' $^ >$@
+CLEAN-OBJS += libvpx.syms
+
+libvpx.def: $(call enabled,CODEC_EXPORTS)
+	@echo "    [CREATE] $@"
+	$(qexec)echo LIBRARY $(LIBVPX_SO:.dll=) INITINSTANCE TERMINSTANCE > $@
+	$(qexec)echo "DATA MULTIPLE NONSHARED" >> $@
+	$(qexec)echo "EXPORTS" >> $@
+	$(qexec)awk '!/vpx_svc_*/ {print "_"$$2}' $^ >>$@
+CLEAN-OBJS += libvpx.def
+
+libvpx_dll.a: $(LIBVPX_SO)
+	@echo "    [IMPLIB] $@"
+	$(qexec)emximp -o $@ $<
+CLEAN-OBJS += libvpx_dll.a
+
+define libvpx_symlink_template
+$(1): $(2)
+	@echo "    [LN]     $(2) $$@"
+	$(qexec)mkdir -p $$(dir $$@)
+	$(qexec)ln -sf $(2) $$@
+endef
+
+$(eval $(call libvpx_symlink_template,\
+    $(addprefix $(BUILD_PFX),$(notdir $(LIBVPX_SO_SYMLINKS))),\
+    $(BUILD_PFX)$(LIBVPX_SO)))
+$(eval $(call libvpx_symlink_template,\
+    $(addprefix $(DIST_DIR)/,$(LIBVPX_SO_SYMLINKS)),\
+    $(LIBVPX_SO)))
+
+
+INSTALL-LIBS-$(CONFIG_SHARED) += $(LIBVPX_SO_SYMLINKS)
+INSTALL-LIBS-$(CONFIG_SHARED) += $(LIBSUBDIR)/$(LIBVPX_SO)
+INSTALL-LIBS-$(CONFIG_SHARED) += $(if $(LIBVPX_SO_IMPLIB),$(LIBSUBDIR)/$(LIBVPX_SO_IMPLIB))
+
+
+LIBS-yes += vpx.pc
+vpx.pc: config.mk libs.mk
+	@echo "    [CREATE] $@"
+	$(qexec)echo '# pkg-config file from libvpx $(VERSION_STRING)' > $@
+	$(qexec)echo 'prefix=$(PREFIX)' >> $@
+	$(qexec)echo 'exec_prefix=$${prefix}' >> $@
+	$(qexec)echo 'libdir=$${prefix}/$(LIBSUBDIR)' >> $@
+	$(qexec)echo 'includedir=$${prefix}/include' >> $@
+	$(qexec)echo '' >> $@
+	$(qexec)echo 'Name: vpx' >> $@
+	$(qexec)echo 'Description: WebM Project VPx codec implementation' >> $@
+	$(qexec)echo 'Version: $(VERSION_MAJOR).$(VERSION_MINOR).$(VERSION_PATCH)' >> $@
+	$(qexec)echo 'Requires:' >> $@
+	$(qexec)echo 'Conflicts:' >> $@
+	$(qexec)echo 'Libs: -L$${libdir} -lvpx -lm' >> $@
+ifeq ($(HAVE_PTHREAD_H),yes)
+	$(qexec)echo 'Libs.private: -lm -lpthread' >> $@
+else
+	$(qexec)echo 'Libs.private: -lm' >> $@
+endif
+	$(qexec)echo 'Cflags: -I$${includedir}' >> $@
+INSTALL-LIBS-yes += $(LIBSUBDIR)/pkgconfig/vpx.pc
+INSTALL_MAPS += $(LIBSUBDIR)/pkgconfig/%.pc %.pc
+CLEAN-OBJS += vpx.pc
+endif
+
+#
+# Rule to make assembler configuration file from C configuration file
+#
+ifeq ($(ARCH_X86)$(ARCH_X86_64),yes)
+# YASM
+$(BUILD_PFX)vpx_config.asm: $(BUILD_PFX)vpx_config.h
+	@echo "    [CREATE] $@"
+	@egrep "#define [A-Z0-9_]+ [01]" $< \
+	    | awk '{print $$2 " equ " $$3}' > $@
+else
+ADS2GAS=$(if $(filter yes,$(CONFIG_GCC)),| $(ASM_CONVERSION))
+$(BUILD_PFX)vpx_config.asm: $(BUILD_PFX)vpx_config.h
+	@echo "    [CREATE] $@"
+	@egrep "#define [A-Z0-9_]+ [01]" $< \
+	    | awk '{print $$2 " EQU " $$3}' $(ADS2GAS) > $@
+	@echo "        END" $(ADS2GAS) >> $@
+CLEAN-OBJS += $(BUILD_PFX)vpx_config.asm
+endif
+
+#
+# Add assembler dependencies for configuration.
+#
+$(filter %.s.o,$(OBJS-yes)):     $(BUILD_PFX)vpx_config.asm
+$(filter %$(ASM).o,$(OBJS-yes)): $(BUILD_PFX)vpx_config.asm
+
+
+$(shell $(SRC_PATH_BARE)/build/make/version.sh "$(SRC_PATH_BARE)" $(BUILD_PFX)vpx_version.h)
+CLEAN-OBJS += $(BUILD_PFX)vpx_version.h
+
+#
+# Add include path for libwebm sources.
+#
+ifeq ($(CONFIG_WEBM_IO),yes)
+  CXXFLAGS += -I$(SRC_PATH_BARE)/third_party/libwebm
+endif
+
+##
+## libvpx test directives
+##
+ifeq ($(CONFIG_UNIT_TESTS),yes)
+LIBVPX_TEST_DATA_PATH ?= .
+
+include $(SRC_PATH_BARE)/test/test.mk
+LIBVPX_TEST_SRCS=$(addprefix test/,$(call enabled,LIBVPX_TEST_SRCS))
+LIBVPX_TEST_BIN=./test_libvpx$(EXE_SFX)
+LIBVPX_TEST_DATA=$(addprefix $(LIBVPX_TEST_DATA_PATH)/,\
+                     $(call enabled,LIBVPX_TEST_DATA))
+libvpx_test_data_url=http://downloads.webmproject.org/test_data/libvpx/$(1)
+
+TEST_INTRA_PRED_SPEED_BIN=./test_intra_pred_speed$(EXE_SFX)
+TEST_INTRA_PRED_SPEED_SRCS=$(addprefix test/,$(call enabled,TEST_INTRA_PRED_SPEED_SRCS))
+TEST_INTRA_PRED_SPEED_OBJS := $(sort $(call objs,$(TEST_INTRA_PRED_SPEED_SRCS)))
+
+libvpx_test_srcs.txt:
+	@echo "    [CREATE] $@"
+	@echo $(LIBVPX_TEST_SRCS) | xargs -n1 echo | LC_ALL=C sort -u > $@
+CLEAN-OBJS += libvpx_test_srcs.txt
+
+$(LIBVPX_TEST_DATA): $(SRC_PATH_BARE)/test/test-data.sha1
+	@echo "    [DOWNLOAD] $@"
+	$(qexec)trap 'rm -f $@' INT TERM &&\
+            curl -L -o $@ $(call libvpx_test_data_url,$(@F))
+
+testdata:: $(LIBVPX_TEST_DATA)
+	$(qexec)[ -x "$$(which sha1sum)" ] && sha1sum=sha1sum;\
+          [ -x "$$(which shasum)" ] && sha1sum=shasum;\
+          [ -x "$$(which sha1)" ] && sha1sum=sha1;\
+          if [ -n "$${sha1sum}" ]; then\
+            set -e;\
+            echo "Checking test data:";\
+            for f in $(call enabled,LIBVPX_TEST_DATA); do\
+                grep $$f $(SRC_PATH_BARE)/test/test-data.sha1 |\
+                    (cd $(LIBVPX_TEST_DATA_PATH); $${sha1sum} -c);\
+            done; \
+        else\
+            echo "Skipping test data integrity check, sha1sum not found.";\
+        fi
+
+ifeq ($(CONFIG_EXTERNAL_BUILD),yes)
+ifeq ($(CONFIG_MSVS),yes)
+
+gtest.$(VCPROJ_SFX): $(SRC_PATH_BARE)/third_party/googletest/src/src/gtest-all.cc
+	@echo "    [CREATE] $@"
+	$(qexec)$(GEN_VCPROJ) \
+            --lib \
+            --target=$(TOOLCHAIN) \
+            $(if $(CONFIG_STATIC_MSVCRT),--static-crt) \
+            --name=gtest \
+            --proj-guid=EC00E1EC-AF68-4D92-A255-181690D1C9B1 \
+            --ver=$(CONFIG_VS_VERSION) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
+            -D_VARIADIC_MAX=10 \
+            --out=gtest.$(VCPROJ_SFX) $(SRC_PATH_BARE)/third_party/googletest/src/src/gtest-all.cc \
+            -I. -I"$(SRC_PATH_BARE)/third_party/googletest/src/include" -I"$(SRC_PATH_BARE)/third_party/googletest/src"
+
+PROJECTS-$(CONFIG_MSVS) += gtest.$(VCPROJ_SFX)
+
+test_libvpx.$(VCPROJ_SFX): $(LIBVPX_TEST_SRCS) vpx.$(VCPROJ_SFX) gtest.$(VCPROJ_SFX)
+	@echo "    [CREATE] $@"
+	$(qexec)$(GEN_VCPROJ) \
+            --exe \
+            --target=$(TOOLCHAIN) \
+            --name=test_libvpx \
+            -D_VARIADIC_MAX=10 \
+            --proj-guid=CD837F5F-52D8-4314-A370-895D614166A7 \
+            --ver=$(CONFIG_VS_VERSION) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
+            $(if $(CONFIG_STATIC_MSVCRT),--static-crt) \
+            --out=$@ $(INTERNAL_CFLAGS) $(CFLAGS) \
+            -I. -I"$(SRC_PATH_BARE)/third_party/googletest/src/include" \
+            $(if $(CONFIG_WEBM_IO),-I"$(SRC_PATH_BARE)/third_party/libwebm") \
+            -L. -l$(CODEC_LIB) -l$(GTEST_LIB) $^
+
+PROJECTS-$(CONFIG_MSVS) += test_libvpx.$(VCPROJ_SFX)
+
+LIBVPX_TEST_BIN := $(addprefix $(TGT_OS:win64=x64)/Release/,$(notdir $(LIBVPX_TEST_BIN)))
+
+ifneq ($(strip $(TEST_INTRA_PRED_SPEED_OBJS)),)
+PROJECTS-$(CONFIG_MSVS) += test_intra_pred_speed.$(VCPROJ_SFX)
+test_intra_pred_speed.$(VCPROJ_SFX): $(TEST_INTRA_PRED_SPEED_SRCS) vpx.$(VCPROJ_SFX) gtest.$(VCPROJ_SFX)
+	@echo "    [CREATE] $@"
+	$(qexec)$(GEN_VCPROJ) \
+            --exe \
+            --target=$(TOOLCHAIN) \
+            --name=test_intra_pred_speed \
+            -D_VARIADIC_MAX=10 \
+            --proj-guid=CD837F5F-52D8-4314-A370-895D614166A7 \
+            --ver=$(CONFIG_VS_VERSION) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
+            $(if $(CONFIG_STATIC_MSVCRT),--static-crt) \
+            --out=$@ $(INTERNAL_CFLAGS) $(CFLAGS) \
+            -I. -I"$(SRC_PATH_BARE)/third_party/googletest/src/include" \
+            -L. -l$(CODEC_LIB) -l$(GTEST_LIB) $^
+endif  # TEST_INTRA_PRED_SPEED
+endif
+else
+
+include $(SRC_PATH_BARE)/third_party/googletest/gtest.mk
+GTEST_SRCS := $(addprefix third_party/googletest/src/,$(call enabled,GTEST_SRCS))
+GTEST_OBJS=$(call objs,$(GTEST_SRCS))
+ifeq ($(filter win%,$(TGT_OS)),$(TGT_OS))
+# Disabling pthreads globally will cause issues on darwin and possibly elsewhere
+$(GTEST_OBJS) $(GTEST_OBJS:.o=.d): CXXFLAGS += -DGTEST_HAS_PTHREAD=0
+endif
+GTEST_INCLUDES := -I$(SRC_PATH_BARE)/third_party/googletest/src
+GTEST_INCLUDES += -I$(SRC_PATH_BARE)/third_party/googletest/src/include
+$(GTEST_OBJS) $(GTEST_OBJS:.o=.d): CXXFLAGS += $(GTEST_INCLUDES)
+OBJS-yes += $(GTEST_OBJS)
+LIBS-yes += $(BUILD_PFX)libgtest.a $(BUILD_PFX)libgtest_g.a
+$(BUILD_PFX)libgtest_g.a: $(GTEST_OBJS)
+
+LIBVPX_TEST_OBJS=$(sort $(call objs,$(LIBVPX_TEST_SRCS)))
+$(LIBVPX_TEST_OBJS) $(LIBVPX_TEST_OBJS:.o=.d): CXXFLAGS += $(GTEST_INCLUDES)
+OBJS-yes += $(LIBVPX_TEST_OBJS)
+BINS-yes += $(LIBVPX_TEST_BIN)
+
+CODEC_LIB=$(if $(CONFIG_DEBUG_LIBS),vpx_g,vpx)
+CODEC_LIB_SUF=$(if $(CONFIG_SHARED),$(SHARED_LIB_SUF),.a)
+TEST_LIBS := lib$(CODEC_LIB)$(CODEC_LIB_SUF) libgtest.a
+$(LIBVPX_TEST_BIN): $(TEST_LIBS)
+$(eval $(call linkerxx_template,$(LIBVPX_TEST_BIN), \
+              $(LIBVPX_TEST_OBJS) \
+              -L. -lvpx -lgtest $(extralibs) -lm))
+
+ifneq ($(strip $(TEST_INTRA_PRED_SPEED_OBJS)),)
+$(TEST_INTRA_PRED_SPEED_OBJS) $(TEST_INTRA_PRED_SPEED_OBJS:.o=.d): CXXFLAGS += $(GTEST_INCLUDES)
+OBJS-yes += $(TEST_INTRA_PRED_SPEED_OBJS)
+BINS-yes += $(TEST_INTRA_PRED_SPEED_BIN)
+
+$(TEST_INTRA_PRED_SPEED_BIN): $(TEST_LIBS)
+$(eval $(call linkerxx_template,$(TEST_INTRA_PRED_SPEED_BIN), \
+              $(TEST_INTRA_PRED_SPEED_OBJS) \
+              -L. -lvpx -lgtest $(extralibs) -lm))
+endif  # TEST_INTRA_PRED_SPEED
+
+endif  # CONFIG_UNIT_TESTS
+
+# Install test sources only if codec source is included
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += $(patsubst $(SRC_PATH_BARE)/%,%,\
+    $(shell find $(SRC_PATH_BARE)/third_party/googletest -type f))
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += $(LIBVPX_TEST_SRCS)
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += $(TEST_INTRA_PRED_SPEED_SRCS)
+
+define test_shard_template
+test:: test_shard.$(1)
+test-no-data-check:: test_shard_ndc.$(1)
+test_shard.$(1) test_shard_ndc.$(1): $(LIBVPX_TEST_BIN)
+	@set -e; \
+	 export GTEST_SHARD_INDEX=$(1); \
+	 export GTEST_TOTAL_SHARDS=$(2); \
+	 $(LIBVPX_TEST_BIN)
+test_shard.$(1): testdata
+.PHONY: test_shard.$(1)
+endef
+
+NUM_SHARDS := 10
+SHARDS := 0 1 2 3 4 5 6 7 8 9
+$(foreach s,$(SHARDS),$(eval $(call test_shard_template,$(s),$(NUM_SHARDS))))
+
+endif
+
+##
+## documentation directives
+##
+CLEAN-OBJS += libs.doxy
+DOCS-yes += libs.doxy
+libs.doxy: $(CODEC_DOC_SRCS)
+	@echo "    [CREATE] $@"
+	@rm -f $@
+	@echo "INPUT += $^" >> $@
+	@echo "INCLUDE_PATH += ." >> $@;
+	@echo "ENABLED_SECTIONS += $(sort $(CODEC_DOC_SECTIONS))" >> $@
+
+## Generate rtcd.h for all objects
+ifeq ($(CONFIG_DEPENDENCY_TRACKING),yes)
+$(OBJS-yes:.o=.d): $(RTCD)
+else
+$(OBJS-yes): $(RTCD)
+endif
+
+## Update the global src list
+SRCS += $(CODEC_SRCS) $(LIBVPX_TEST_SRCS) $(GTEST_SRCS)
+
+##
+## vpxdec/vpxenc tests.
+##
+ifeq ($(CONFIG_UNIT_TESTS),yes)
+TEST_BIN_PATH = .
+ifeq ($(CONFIG_MSVS),yes)
+# MSVC will build both Debug and Release configurations of tools in a
+# sub directory named for the current target. Assume the user wants to
+# run the Release tools, and assign TEST_BIN_PATH accordingly.
+# TODO(tomfinegan): Is this adequate for ARM?
+# TODO(tomfinegan): Support running the debug versions of tools?
+TEST_BIN_PATH := $(addsuffix /$(TGT_OS:win64=x64)/Release, $(TEST_BIN_PATH))
+endif
+utiltest utiltest-no-data-check:
+	$(qexec)$(SRC_PATH_BARE)/test/vpxdec.sh \
+		--test-data-path $(LIBVPX_TEST_DATA_PATH) \
+		--bin-path $(TEST_BIN_PATH)
+	$(qexec)$(SRC_PATH_BARE)/test/vpxenc.sh \
+		--test-data-path $(LIBVPX_TEST_DATA_PATH) \
+		--bin-path $(TEST_BIN_PATH)
+utiltest: testdata
+else
+utiltest utiltest-no-data-check:
+	@echo Unit tests must be enabled to make the utiltest target.
+endif
+
+##
+## Example tests.
+##
+ifeq ($(CONFIG_UNIT_TESTS),yes)
+# All non-MSVC targets output example targets in a sub dir named examples.
+EXAMPLES_BIN_PATH = examples
+ifeq ($(CONFIG_MSVS),yes)
+# MSVC will build both Debug and Release configurations of the examples in a
+# sub directory named for the current target. Assume the user wants to
+# run the Release tools, and assign EXAMPLES_BIN_PATH accordingly.
+# TODO(tomfinegan): Is this adequate for ARM?
+# TODO(tomfinegan): Support running the debug versions of tools?
+EXAMPLES_BIN_PATH := $(TGT_OS:win64=x64)/Release
+endif
+exampletest exampletest-no-data-check: examples
+	$(qexec)$(SRC_PATH_BARE)/test/examples.sh \
+		--test-data-path $(LIBVPX_TEST_DATA_PATH) \
+		--bin-path $(EXAMPLES_BIN_PATH)
+exampletest: testdata
+else
+exampletest exampletest-no-data-check:
+	@echo Unit tests must be enabled to make the exampletest target.
+endif
diff --git a/libvpx/libvpx/mainpage.dox b/libvpx/libvpx/mainpage.dox
new file mode 100644
index 0000000..ec202fa
--- /dev/null
+++ b/libvpx/libvpx/mainpage.dox
@@ -0,0 +1,53 @@
+/*!\mainpage WebM Codec SDK
+
+  \section main_contents Page Contents
+  - \ref main_intro
+  - \ref main_startpoints
+  - \ref main_support
+
+  \section main_intro Introduction
+  Welcome to the WebM Codec SDK. This SDK allows you to integrate your
+  applications with the VP8 and VP9 video codecs, high quality, royalty free,
+  open source codecs deployed on billions of computers and devices worldwide.
+
+  This distribution of the WebM Codec SDK includes the following support:
+
+  \if vp8_encoder
+  - \ref vp8_encoder
+  \endif
+  \if vp8_decoder
+  - \ref vp8_decoder
+  \endif
+
+
+  \section main_startpoints Starting Points
+  - Consult the \ref changelog for a complete list of improvements in this
+    release.
+  - The \ref readme contains instructions on recompiling the sample applications.
+  - Read the \ref usage "usage" for a narrative on codec usage.
+  - Read the \ref samples "sample code" for examples of how to interact with the
+    codec.
+  - \ref codec reference
+  \if encoder
+  - \ref encoder reference
+  \endif
+  \if decoder
+  - \ref decoder reference
+  \endif
+
+  \section main_support Support Options & FAQ
+  The WebM project is an open source project supported by its community. For
+  questions about this SDK, please mail the apps-devel@webmproject.org list.
+  To contribute, see http://www.webmproject.org/code/contribute and mail
+  codec-devel@webmproject.org.
+*/
+
+/*!\page changelog CHANGELOG
+   \verbinclude CHANGELOG
+*/
+
+/*!\page readme README
+   \verbinclude README
+*/
+
+/*!\defgroup codecs Supported Codecs */
diff --git a/libvpx/libvpx/md5_utils.c b/libvpx/libvpx/md5_utils.c
new file mode 100644
index 0000000..a9b979a
--- /dev/null
+++ b/libvpx/libvpx/md5_utils.c
@@ -0,0 +1,254 @@
+/*
+ * This code implements the MD5 message-digest algorithm.
+ * The algorithm is due to Ron Rivest.  This code was
+ * written by Colin Plumb in 1993, no copyright is claimed.
+ * This code is in the public domain; do with it what you wish.
+ *
+ * Equivalent code is available from RSA Data Security, Inc.
+ * This code has been tested against that, and is equivalent,
+ * except that you don't need to include two pages of legalese
+ * with every copy.
+ *
+ * To compute the message digest of a chunk of bytes, declare an
+ * MD5Context structure, pass it to MD5Init, call MD5Update as
+ * needed on buffers full of bytes, and then call MD5Final, which
+ * will fill a supplied 16-byte array with the digest.
+ *
+ * Changed so as no longer to depend on Colin Plumb's `usual.h' header
+ * definitions
+ *  - Ian Jackson <ian@chiark.greenend.org.uk>.
+ * Still in the public domain.
+ */
+
+#include <string.h>   /* for memcpy() */
+
+#include "md5_utils.h"
+
+static void
+byteSwap(UWORD32 *buf, unsigned words) {
+  md5byte *p;
+
+  /* Only swap bytes for big endian machines */
+  int i = 1;
+
+  if (*(char *)&i == 1)
+    return;
+
+  p = (md5byte *)buf;
+
+  do {
+    *buf++ = (UWORD32)((unsigned)p[3] << 8 | p[2]) << 16 |
+             ((unsigned)p[1] << 8 | p[0]);
+    p += 4;
+  } while (--words);
+}
+
+/*
+ * Start MD5 accumulation.  Set bit count to 0 and buffer to mysterious
+ * initialization constants.
+ */
+void
+MD5Init(struct MD5Context *ctx) {
+  ctx->buf[0] = 0x67452301;
+  ctx->buf[1] = 0xefcdab89;
+  ctx->buf[2] = 0x98badcfe;
+  ctx->buf[3] = 0x10325476;
+
+  ctx->bytes[0] = 0;
+  ctx->bytes[1] = 0;
+}
+
+/*
+ * Update context to reflect the concatenation of another buffer full
+ * of bytes.
+ */
+void
+MD5Update(struct MD5Context *ctx, md5byte const *buf, unsigned len) {
+  UWORD32 t;
+
+  /* Update byte count */
+
+  t = ctx->bytes[0];
+
+  if ((ctx->bytes[0] = t + len) < t)
+    ctx->bytes[1]++;  /* Carry from low to high */
+
+  t = 64 - (t & 0x3f);  /* Space available in ctx->in (at least 1) */
+
+  if (t > len) {
+    memcpy((md5byte *)ctx->in + 64 - t, buf, len);
+    return;
+  }
+
+  /* First chunk is an odd size */
+  memcpy((md5byte *)ctx->in + 64 - t, buf, t);
+  byteSwap(ctx->in, 16);
+  MD5Transform(ctx->buf, ctx->in);
+  buf += t;
+  len -= t;
+
+  /* Process data in 64-byte chunks */
+  while (len >= 64) {
+    memcpy(ctx->in, buf, 64);
+    byteSwap(ctx->in, 16);
+    MD5Transform(ctx->buf, ctx->in);
+    buf += 64;
+    len -= 64;
+  }
+
+  /* Handle any remaining bytes of data. */
+  memcpy(ctx->in, buf, len);
+}
+
+/*
+ * Final wrapup - pad to 64-byte boundary with the bit pattern
+ * 1 0* (64-bit count of bits processed, MSB-first)
+ */
+void
+MD5Final(md5byte digest[16], struct MD5Context *ctx) {
+  int count = ctx->bytes[0] & 0x3f; /* Number of bytes in ctx->in */
+  md5byte *p = (md5byte *)ctx->in + count;
+
+  /* Set the first char of padding to 0x80.  There is always room. */
+  *p++ = 0x80;
+
+  /* Bytes of padding needed to make 56 bytes (-8..55) */
+  count = 56 - 1 - count;
+
+  if (count < 0) {  /* Padding forces an extra block */
+    memset(p, 0, count + 8);
+    byteSwap(ctx->in, 16);
+    MD5Transform(ctx->buf, ctx->in);
+    p = (md5byte *)ctx->in;
+    count = 56;
+  }
+
+  memset(p, 0, count);
+  byteSwap(ctx->in, 14);
+
+  /* Append length in bits and transform */
+  ctx->in[14] = ctx->bytes[0] << 3;
+  ctx->in[15] = ctx->bytes[1] << 3 | ctx->bytes[0] >> 29;
+  MD5Transform(ctx->buf, ctx->in);
+
+  byteSwap(ctx->buf, 4);
+  memcpy(digest, ctx->buf, 16);
+  memset(ctx, 0, sizeof(*ctx)); /* In case it's sensitive */
+}
+
+#ifndef ASM_MD5
+
+/* The four core functions - F1 is optimized somewhat */
+
+/* #define F1(x, y, z) (x & y | ~x & z) */
+#define F1(x, y, z) (z ^ (x & (y ^ z)))
+#define F2(x, y, z) F1(z, x, y)
+#define F3(x, y, z) (x ^ y ^ z)
+#define F4(x, y, z) (y ^ (x | ~z))
+
+/* This is the central step in the MD5 algorithm. */
+#define MD5STEP(f,w,x,y,z,in,s) \
+  (w += f(x,y,z) + in, w = (w<<s | w>>(32-s)) + x)
+
+#if defined(__clang__) && defined(__has_attribute)
+#if __has_attribute(no_sanitize)
+#define VPX_NO_UNSIGNED_OVERFLOW_CHECK \
+  __attribute__((no_sanitize("unsigned-integer-overflow")))
+#endif
+#endif
+
+#ifndef VPX_NO_UNSIGNED_OVERFLOW_CHECK
+#define VPX_NO_UNSIGNED_OVERFLOW_CHECK
+#endif
+
+/*
+ * The core of the MD5 algorithm, this alters an existing MD5 hash to
+ * reflect the addition of 16 longwords of new data.  MD5Update blocks
+ * the data and converts bytes into longwords for this routine.
+ */
+VPX_NO_UNSIGNED_OVERFLOW_CHECK void
+MD5Transform(UWORD32 buf[4], UWORD32 const in[16]) {
+  register UWORD32 a, b, c, d;
+
+  a = buf[0];
+  b = buf[1];
+  c = buf[2];
+  d = buf[3];
+
+  MD5STEP(F1, a, b, c, d, in[0] + 0xd76aa478, 7);
+  MD5STEP(F1, d, a, b, c, in[1] + 0xe8c7b756, 12);
+  MD5STEP(F1, c, d, a, b, in[2] + 0x242070db, 17);
+  MD5STEP(F1, b, c, d, a, in[3] + 0xc1bdceee, 22);
+  MD5STEP(F1, a, b, c, d, in[4] + 0xf57c0faf, 7);
+  MD5STEP(F1, d, a, b, c, in[5] + 0x4787c62a, 12);
+  MD5STEP(F1, c, d, a, b, in[6] + 0xa8304613, 17);
+  MD5STEP(F1, b, c, d, a, in[7] + 0xfd469501, 22);
+  MD5STEP(F1, a, b, c, d, in[8] + 0x698098d8, 7);
+  MD5STEP(F1, d, a, b, c, in[9] + 0x8b44f7af, 12);
+  MD5STEP(F1, c, d, a, b, in[10] + 0xffff5bb1, 17);
+  MD5STEP(F1, b, c, d, a, in[11] + 0x895cd7be, 22);
+  MD5STEP(F1, a, b, c, d, in[12] + 0x6b901122, 7);
+  MD5STEP(F1, d, a, b, c, in[13] + 0xfd987193, 12);
+  MD5STEP(F1, c, d, a, b, in[14] + 0xa679438e, 17);
+  MD5STEP(F1, b, c, d, a, in[15] + 0x49b40821, 22);
+
+  MD5STEP(F2, a, b, c, d, in[1] + 0xf61e2562, 5);
+  MD5STEP(F2, d, a, b, c, in[6] + 0xc040b340, 9);
+  MD5STEP(F2, c, d, a, b, in[11] + 0x265e5a51, 14);
+  MD5STEP(F2, b, c, d, a, in[0] + 0xe9b6c7aa, 20);
+  MD5STEP(F2, a, b, c, d, in[5] + 0xd62f105d, 5);
+  MD5STEP(F2, d, a, b, c, in[10] + 0x02441453, 9);
+  MD5STEP(F2, c, d, a, b, in[15] + 0xd8a1e681, 14);
+  MD5STEP(F2, b, c, d, a, in[4] + 0xe7d3fbc8, 20);
+  MD5STEP(F2, a, b, c, d, in[9] + 0x21e1cde6, 5);
+  MD5STEP(F2, d, a, b, c, in[14] + 0xc33707d6, 9);
+  MD5STEP(F2, c, d, a, b, in[3] + 0xf4d50d87, 14);
+  MD5STEP(F2, b, c, d, a, in[8] + 0x455a14ed, 20);
+  MD5STEP(F2, a, b, c, d, in[13] + 0xa9e3e905, 5);
+  MD5STEP(F2, d, a, b, c, in[2] + 0xfcefa3f8, 9);
+  MD5STEP(F2, c, d, a, b, in[7] + 0x676f02d9, 14);
+  MD5STEP(F2, b, c, d, a, in[12] + 0x8d2a4c8a, 20);
+
+  MD5STEP(F3, a, b, c, d, in[5] + 0xfffa3942, 4);
+  MD5STEP(F3, d, a, b, c, in[8] + 0x8771f681, 11);
+  MD5STEP(F3, c, d, a, b, in[11] + 0x6d9d6122, 16);
+  MD5STEP(F3, b, c, d, a, in[14] + 0xfde5380c, 23);
+  MD5STEP(F3, a, b, c, d, in[1] + 0xa4beea44, 4);
+  MD5STEP(F3, d, a, b, c, in[4] + 0x4bdecfa9, 11);
+  MD5STEP(F3, c, d, a, b, in[7] + 0xf6bb4b60, 16);
+  MD5STEP(F3, b, c, d, a, in[10] + 0xbebfbc70, 23);
+  MD5STEP(F3, a, b, c, d, in[13] + 0x289b7ec6, 4);
+  MD5STEP(F3, d, a, b, c, in[0] + 0xeaa127fa, 11);
+  MD5STEP(F3, c, d, a, b, in[3] + 0xd4ef3085, 16);
+  MD5STEP(F3, b, c, d, a, in[6] + 0x04881d05, 23);
+  MD5STEP(F3, a, b, c, d, in[9] + 0xd9d4d039, 4);
+  MD5STEP(F3, d, a, b, c, in[12] + 0xe6db99e5, 11);
+  MD5STEP(F3, c, d, a, b, in[15] + 0x1fa27cf8, 16);
+  MD5STEP(F3, b, c, d, a, in[2] + 0xc4ac5665, 23);
+
+  MD5STEP(F4, a, b, c, d, in[0] + 0xf4292244, 6);
+  MD5STEP(F4, d, a, b, c, in[7] + 0x432aff97, 10);
+  MD5STEP(F4, c, d, a, b, in[14] + 0xab9423a7, 15);
+  MD5STEP(F4, b, c, d, a, in[5] + 0xfc93a039, 21);
+  MD5STEP(F4, a, b, c, d, in[12] + 0x655b59c3, 6);
+  MD5STEP(F4, d, a, b, c, in[3] + 0x8f0ccc92, 10);
+  MD5STEP(F4, c, d, a, b, in[10] + 0xffeff47d, 15);
+  MD5STEP(F4, b, c, d, a, in[1] + 0x85845dd1, 21);
+  MD5STEP(F4, a, b, c, d, in[8] + 0x6fa87e4f, 6);
+  MD5STEP(F4, d, a, b, c, in[15] + 0xfe2ce6e0, 10);
+  MD5STEP(F4, c, d, a, b, in[6] + 0xa3014314, 15);
+  MD5STEP(F4, b, c, d, a, in[13] + 0x4e0811a1, 21);
+  MD5STEP(F4, a, b, c, d, in[4] + 0xf7537e82, 6);
+  MD5STEP(F4, d, a, b, c, in[11] + 0xbd3af235, 10);
+  MD5STEP(F4, c, d, a, b, in[2] + 0x2ad7d2bb, 15);
+  MD5STEP(F4, b, c, d, a, in[9] + 0xeb86d391, 21);
+
+  buf[0] += a;
+  buf[1] += b;
+  buf[2] += c;
+  buf[3] += d;
+}
+
+#undef VPX_NO_UNSIGNED_OVERFLOW_CHECK
+
+#endif
diff --git a/libvpx/libvpx/md5_utils.h b/libvpx/libvpx/md5_utils.h
new file mode 100644
index 0000000..bd4991b
--- /dev/null
+++ b/libvpx/libvpx/md5_utils.h
@@ -0,0 +1,49 @@
+/*
+ * This is the header file for the MD5 message-digest algorithm.
+ * The algorithm is due to Ron Rivest.  This code was
+ * written by Colin Plumb in 1993, no copyright is claimed.
+ * This code is in the public domain; do with it what you wish.
+ *
+ * Equivalent code is available from RSA Data Security, Inc.
+ * This code has been tested against that, and is equivalent,
+ * except that you don't need to include two pages of legalese
+ * with every copy.
+ *
+ * To compute the message digest of a chunk of bytes, declare an
+ * MD5Context structure, pass it to MD5Init, call MD5Update as
+ * needed on buffers full of bytes, and then call MD5Final, which
+ * will fill a supplied 16-byte array with the digest.
+ *
+ * Changed so as no longer to depend on Colin Plumb's `usual.h'
+ * header definitions
+ *  - Ian Jackson <ian@chiark.greenend.org.uk>.
+ * Still in the public domain.
+ */
+
+#ifndef MD5_UTILS_H_
+#define MD5_UTILS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define md5byte unsigned char
+#define UWORD32 unsigned int
+
+typedef struct MD5Context MD5Context;
+struct MD5Context {
+  UWORD32 buf[4];
+  UWORD32 bytes[2];
+  UWORD32 in[16];
+};
+
+void MD5Init(struct MD5Context *context);
+void MD5Update(struct MD5Context *context, md5byte const *buf, unsigned len);
+void MD5Final(unsigned char digest[16], struct MD5Context *context);
+void MD5Transform(UWORD32 buf[4], UWORD32 const in[16]);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // MD5_UTILS_H_
diff --git a/libvpx/libvpx/rate_hist.c b/libvpx/libvpx/rate_hist.c
new file mode 100644
index 0000000..a77222b
--- /dev/null
+++ b/libvpx/libvpx/rate_hist.c
@@ -0,0 +1,285 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdlib.h>
+#include <limits.h>
+#include <stdio.h>
+#include <math.h>
+
+#include "./rate_hist.h"
+
+#define RATE_BINS 100
+#define HIST_BAR_MAX 40
+
+struct hist_bucket {
+  int low;
+  int high;
+  int count;
+};
+
+struct rate_hist {
+  int64_t *pts;
+  int *sz;
+  int samples;
+  int frames;
+  struct hist_bucket bucket[RATE_BINS];
+  int total;
+};
+
+struct rate_hist *init_rate_histogram(const vpx_codec_enc_cfg_t *cfg,
+                                      const vpx_rational_t *fps) {
+  int i;
+  struct rate_hist *hist = malloc(sizeof(*hist));
+
+  // Determine the number of samples in the buffer. Use the file's framerate
+  // to determine the number of frames in rc_buf_sz milliseconds, with an
+  // adjustment (5/4) to account for alt-refs
+  hist->samples = cfg->rc_buf_sz * 5 / 4 * fps->num / fps->den / 1000;
+
+  // prevent division by zero
+  if (hist->samples == 0)
+    hist->samples = 1;
+
+  hist->frames = 0;
+  hist->total = 0;
+
+  hist->pts = calloc(hist->samples, sizeof(*hist->pts));
+  hist->sz = calloc(hist->samples, sizeof(*hist->sz));
+  for (i = 0; i < RATE_BINS; i++) {
+    hist->bucket[i].low = INT_MAX;
+    hist->bucket[i].high = 0;
+    hist->bucket[i].count = 0;
+  }
+
+  return hist;
+}
+
+void destroy_rate_histogram(struct rate_hist *hist) {
+  if (hist) {
+    free(hist->pts);
+    free(hist->sz);
+    free(hist);
+  }
+}
+
+void update_rate_histogram(struct rate_hist *hist,
+                           const vpx_codec_enc_cfg_t *cfg,
+                           const vpx_codec_cx_pkt_t *pkt) {
+  int i;
+  int64_t then = 0;
+  int64_t avg_bitrate = 0;
+  int64_t sum_sz = 0;
+  const int64_t now = pkt->data.frame.pts * 1000 *
+                          (uint64_t)cfg->g_timebase.num /
+                              (uint64_t)cfg->g_timebase.den;
+
+  int idx = hist->frames++ % hist->samples;
+  hist->pts[idx] = now;
+  hist->sz[idx] = (int)pkt->data.frame.sz;
+
+  if (now < cfg->rc_buf_initial_sz)
+    return;
+
+  if (!cfg->rc_target_bitrate)
+    return;
+
+  then = now;
+
+  /* Sum the size over the past rc_buf_sz ms */
+  for (i = hist->frames; i > 0 && hist->frames - i < hist->samples; i--) {
+    const int i_idx = (i - 1) % hist->samples;
+
+    then = hist->pts[i_idx];
+    if (now - then > cfg->rc_buf_sz)
+      break;
+    sum_sz += hist->sz[i_idx];
+  }
+
+  if (now == then)
+    return;
+
+  avg_bitrate = sum_sz * 8 * 1000 / (now - then);
+  idx = (int)(avg_bitrate * (RATE_BINS / 2) / (cfg->rc_target_bitrate * 1000));
+  if (idx < 0)
+    idx = 0;
+  if (idx > RATE_BINS - 1)
+    idx = RATE_BINS - 1;
+  if (hist->bucket[idx].low > avg_bitrate)
+    hist->bucket[idx].low = (int)avg_bitrate;
+  if (hist->bucket[idx].high < avg_bitrate)
+    hist->bucket[idx].high = (int)avg_bitrate;
+  hist->bucket[idx].count++;
+  hist->total++;
+}
+
+static int merge_hist_buckets(struct hist_bucket *bucket,
+                              int max_buckets, int *num_buckets) {
+  int small_bucket = 0, merge_bucket = INT_MAX, big_bucket = 0;
+  int buckets = *num_buckets;
+  int i;
+
+  /* Find the extrema for this list of buckets */
+  big_bucket = small_bucket = 0;
+  for (i = 0; i < buckets; i++) {
+    if (bucket[i].count < bucket[small_bucket].count)
+      small_bucket = i;
+    if (bucket[i].count > bucket[big_bucket].count)
+      big_bucket = i;
+  }
+
+  /* If we have too many buckets, merge the smallest with an adjacent
+   * bucket.
+   */
+  while (buckets > max_buckets) {
+    int last_bucket = buckets - 1;
+
+    /* merge the small bucket with an adjacent one. */
+    if (small_bucket == 0)
+      merge_bucket = 1;
+    else if (small_bucket == last_bucket)
+      merge_bucket = last_bucket - 1;
+    else if (bucket[small_bucket - 1].count < bucket[small_bucket + 1].count)
+      merge_bucket = small_bucket - 1;
+    else
+      merge_bucket = small_bucket + 1;
+
+    assert(abs(merge_bucket - small_bucket) <= 1);
+    assert(small_bucket < buckets);
+    assert(big_bucket < buckets);
+    assert(merge_bucket < buckets);
+
+    if (merge_bucket < small_bucket) {
+      bucket[merge_bucket].high = bucket[small_bucket].high;
+      bucket[merge_bucket].count += bucket[small_bucket].count;
+    } else {
+      bucket[small_bucket].high = bucket[merge_bucket].high;
+      bucket[small_bucket].count += bucket[merge_bucket].count;
+      merge_bucket = small_bucket;
+    }
+
+    assert(bucket[merge_bucket].low != bucket[merge_bucket].high);
+
+    buckets--;
+
+    /* Remove the merge_bucket from the list, and find the new small
+     * and big buckets while we're at it
+     */
+    big_bucket = small_bucket = 0;
+    for (i = 0; i < buckets; i++) {
+      if (i > merge_bucket)
+        bucket[i] = bucket[i + 1];
+
+      if (bucket[i].count < bucket[small_bucket].count)
+        small_bucket = i;
+      if (bucket[i].count > bucket[big_bucket].count)
+        big_bucket = i;
+    }
+  }
+
+  *num_buckets = buckets;
+  return bucket[big_bucket].count;
+}
+
+static void show_histogram(const struct hist_bucket *bucket,
+                           int buckets, int total, int scale) {
+  const char *pat1, *pat2;
+  int i;
+
+  switch ((int)(log(bucket[buckets - 1].high) / log(10)) + 1) {
+    case 1:
+    case 2:
+      pat1 = "%4d %2s: ";
+      pat2 = "%4d-%2d: ";
+      break;
+    case 3:
+      pat1 = "%5d %3s: ";
+      pat2 = "%5d-%3d: ";
+      break;
+    case 4:
+      pat1 = "%6d %4s: ";
+      pat2 = "%6d-%4d: ";
+      break;
+    case 5:
+      pat1 = "%7d %5s: ";
+      pat2 = "%7d-%5d: ";
+      break;
+    case 6:
+      pat1 = "%8d %6s: ";
+      pat2 = "%8d-%6d: ";
+      break;
+    case 7:
+      pat1 = "%9d %7s: ";
+      pat2 = "%9d-%7d: ";
+      break;
+    default:
+      pat1 = "%12d %10s: ";
+      pat2 = "%12d-%10d: ";
+      break;
+  }
+
+  for (i = 0; i < buckets; i++) {
+    int len;
+    int j;
+    float pct;
+
+    pct = (float)(100.0 * bucket[i].count / total);
+    len = HIST_BAR_MAX * bucket[i].count / scale;
+    if (len < 1)
+      len = 1;
+    assert(len <= HIST_BAR_MAX);
+
+    if (bucket[i].low == bucket[i].high)
+      fprintf(stderr, pat1, bucket[i].low, "");
+    else
+      fprintf(stderr, pat2, bucket[i].low, bucket[i].high);
+
+    for (j = 0; j < HIST_BAR_MAX; j++)
+      fprintf(stderr, j < len ? "=" : " ");
+    fprintf(stderr, "\t%5d (%6.2f%%)\n", bucket[i].count, pct);
+  }
+}
+
+void show_q_histogram(const int counts[64], int max_buckets) {
+  struct hist_bucket bucket[64];
+  int buckets = 0;
+  int total = 0;
+  int scale;
+  int i;
+
+  for (i = 0; i < 64; i++) {
+    if (counts[i]) {
+      bucket[buckets].low = bucket[buckets].high = i;
+      bucket[buckets].count = counts[i];
+      buckets++;
+      total += counts[i];
+    }
+  }
+
+  fprintf(stderr, "\nQuantizer Selection:\n");
+  scale = merge_hist_buckets(bucket, max_buckets, &buckets);
+  show_histogram(bucket, buckets, total, scale);
+}
+
+void show_rate_histogram(struct rate_hist *hist,
+                         const vpx_codec_enc_cfg_t *cfg, int max_buckets) {
+  int i, scale;
+  int buckets = 0;
+
+  for (i = 0; i < RATE_BINS; i++) {
+    if (hist->bucket[i].low == INT_MAX)
+      continue;
+    hist->bucket[buckets++] = hist->bucket[i];
+  }
+
+  fprintf(stderr, "\nRate (over %dms window):\n", cfg->rc_buf_sz);
+  scale = merge_hist_buckets(hist->bucket, max_buckets, &buckets);
+  show_histogram(hist->bucket, buckets, hist->total, scale);
+}
diff --git a/libvpx/libvpx/rate_hist.h b/libvpx/libvpx/rate_hist.h
new file mode 100644
index 0000000..00a1676
--- /dev/null
+++ b/libvpx/libvpx/rate_hist.h
@@ -0,0 +1,40 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef RATE_HIST_H_
+#define RATE_HIST_H_
+
+#include "vpx/vpx_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct rate_hist;
+
+struct rate_hist *init_rate_histogram(const vpx_codec_enc_cfg_t *cfg,
+                                      const vpx_rational_t *fps);
+
+void destroy_rate_histogram(struct rate_hist *hist);
+
+void update_rate_histogram(struct rate_hist *hist,
+                           const vpx_codec_enc_cfg_t *cfg,
+                           const vpx_codec_cx_pkt_t *pkt);
+
+void show_q_histogram(const int counts[64], int max_buckets);
+
+void show_rate_histogram(struct rate_hist *hist, const vpx_codec_enc_cfg_t *cfg,
+                         int max_buckets);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // RATE_HIST_H_
diff --git a/libvpx/libvpx/solution.mk b/libvpx/libvpx/solution.mk
new file mode 100644
index 0000000..145adc0
--- /dev/null
+++ b/libvpx/libvpx/solution.mk
@@ -0,0 +1,31 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+# libvpx reverse dependencies (targets that depend on libvpx)
+VPX_NONDEPS=$(addsuffix .$(VCPROJ_SFX),vpx gtest)
+VPX_RDEPS=$(foreach vcp,\
+              $(filter-out $(VPX_NONDEPS),$^), --dep=$(vcp:.$(VCPROJ_SFX)=):vpx)
+
+vpx.sln: $(wildcard *.$(VCPROJ_SFX))
+	@echo "    [CREATE] $@"
+	$(SRC_PATH_BARE)/build/make/gen_msvs_sln.sh \
+            $(if $(filter vpx.$(VCPROJ_SFX),$^),$(VPX_RDEPS)) \
+            --dep=test_libvpx:gtest \
+            --ver=$(CONFIG_VS_VERSION)\
+            --out=$@ $^
+vpx.sln.mk: vpx.sln
+	@true
+
+PROJECTS-yes += vpx.sln vpx.sln.mk
+-include vpx.sln.mk
+
+# Always install this file, as it is an unconditional post-build rule.
+INSTALL_MAPS += src/%     $(SRC_PATH_BARE)/%
+INSTALL-SRCS-yes            += $(target).mk
diff --git a/libvpx/libvpx/test/acm_random.h b/libvpx/libvpx/test/acm_random.h
new file mode 100644
index 0000000..b94b6e1
--- /dev/null
+++ b/libvpx/libvpx/test/acm_random.h
@@ -0,0 +1,73 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef TEST_ACM_RANDOM_H_
+#define TEST_ACM_RANDOM_H_
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "vpx/vpx_integer.h"
+
+namespace libvpx_test {
+
+class ACMRandom {
+ public:
+  ACMRandom() : random_(DeterministicSeed()) {}
+
+  explicit ACMRandom(int seed) : random_(seed) {}
+
+  void Reset(int seed) {
+    random_.Reseed(seed);
+  }
+  uint16_t Rand16(void) {
+    const uint32_t value =
+        random_.Generate(testing::internal::Random::kMaxRange);
+    return (value >> 15) & 0xffff;
+  }
+
+  int16_t Rand9Signed(void) {
+    // Use 9 bits: values between 255 (0x0FF) and -256 (0x100).
+    const uint32_t value = random_.Generate(512);
+    return static_cast<int16_t>(value) - 256;
+  }
+
+  uint8_t Rand8(void) {
+    const uint32_t value =
+        random_.Generate(testing::internal::Random::kMaxRange);
+    // There's a bit more entropy in the upper bits of this implementation.
+    return (value >> 23) & 0xff;
+  }
+
+  uint8_t Rand8Extremes(void) {
+    // Returns a random value near 0 or near 255, to better exercise
+    // saturation behavior.
+    const uint8_t r = Rand8();
+    return r < 128 ? r << 4 : r >> 4;
+  }
+
+  int PseudoUniform(int range) {
+    return random_.Generate(range);
+  }
+
+  int operator()(int n) {
+    return PseudoUniform(n);
+  }
+
+  static int DeterministicSeed(void) {
+    return 0xbaba;
+  }
+
+ private:
+  testing::internal::Random random_;
+};
+
+}  // namespace libvpx_test
+
+#endif  // TEST_ACM_RANDOM_H_
diff --git a/libvpx/libvpx/test/active_map_refresh_test.cc b/libvpx/libvpx/test/active_map_refresh_test.cc
new file mode 100644
index 0000000..c945661
--- /dev/null
+++ b/libvpx/libvpx/test/active_map_refresh_test.cc
@@ -0,0 +1,127 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <algorithm>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+
+namespace {
+
+// Check if any pixel in a 16x16 macroblock varies between frames.
+int CheckMb(const vpx_image_t &current, const vpx_image_t &previous,
+            int mb_r, int mb_c) {
+  for (int plane = 0; plane < 3; plane++) {
+    int r = 16 * mb_r;
+    int c0 = 16 * mb_c;
+    int r_top = std::min(r + 16, static_cast<int>(current.d_h));
+    int c_top = std::min(c0 + 16, static_cast<int>(current.d_w));
+    r = std::max(r, 0);
+    c0 = std::max(c0, 0);
+    if (plane > 0 && current.x_chroma_shift) {
+      c_top = (c_top + 1) >> 1;
+      c0 >>= 1;
+    }
+    if (plane > 0 && current.y_chroma_shift) {
+      r_top = (r_top + 1) >> 1;
+      r >>= 1;
+    }
+    for (; r < r_top; ++r) {
+      for (int c = c0; c < c_top; ++c) {
+        if (current.planes[plane][current.stride[plane] * r + c] !=
+            previous.planes[plane][previous.stride[plane] * r + c])
+          return 1;
+      }
+    }
+  }
+  return 0;
+}
+
+void GenerateMap(int mb_rows, int mb_cols, const vpx_image_t &current,
+                 const vpx_image_t &previous, uint8_t *map) {
+  for (int mb_r = 0; mb_r < mb_rows; ++mb_r) {
+    for (int mb_c = 0; mb_c < mb_cols; ++mb_c) {
+      map[mb_r * mb_cols + mb_c] = CheckMb(current, previous, mb_r, mb_c);
+    }
+  }
+}
+
+const int kAqModeCyclicRefresh = 3;
+
+class ActiveMapRefreshTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
+ protected:
+  ActiveMapRefreshTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~ActiveMapRefreshTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    cpu_used_ = GET_PARAM(2);
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    ::libvpx_test::Y4mVideoSource *y4m_video =
+        static_cast<libvpx_test::Y4mVideoSource *>(video);
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_CPUUSED, cpu_used_);
+      encoder->Control(VP9E_SET_AQ_MODE, kAqModeCyclicRefresh);
+    } else if (video->frame() >= 2 && video->img()) {
+      vpx_image_t *current = video->img();
+      vpx_image_t *previous = y4m_holder_->img();
+      ASSERT_TRUE(previous != NULL);
+      vpx_active_map_t map = vpx_active_map_t();
+      const int width = static_cast<int>(current->d_w);
+      const int height = static_cast<int>(current->d_h);
+      const int mb_width = (width + 15) / 16;
+      const int mb_height = (height + 15) / 16;
+      uint8_t *active_map = new uint8_t[mb_width * mb_height];
+      GenerateMap(mb_height, mb_width, *current, *previous, active_map);
+      map.cols = mb_width;
+      map.rows = mb_height;
+      map.active_map = active_map;
+      encoder->Control(VP8E_SET_ACTIVEMAP, &map);
+      delete[] active_map;
+    }
+    if (video->img()) {
+      y4m_video->SwapBuffers(y4m_holder_);
+    }
+  }
+
+  int cpu_used_;
+  ::libvpx_test::Y4mVideoSource *y4m_holder_;
+};
+
+TEST_P(ActiveMapRefreshTest, Test) {
+  cfg_.g_lag_in_frames = 0;
+  cfg_.g_profile = 1;
+  cfg_.rc_target_bitrate = 600;
+  cfg_.rc_resize_allowed = 0;
+  cfg_.rc_min_quantizer = 8;
+  cfg_.rc_max_quantizer = 30;
+  cfg_.g_pass = VPX_RC_ONE_PASS;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.kf_max_dist = 90000;
+
+  ::libvpx_test::Y4mVideoSource video("desktop_credits.y4m", 0, 30);
+  ::libvpx_test::Y4mVideoSource video_holder("desktop_credits.y4m", 0, 30);
+  video_holder.Begin();
+  y4m_holder_ = &video_holder;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+VP9_INSTANTIATE_TEST_CASE(ActiveMapRefreshTest,
+                          ::testing::Values(::libvpx_test::kRealTime),
+                          ::testing::Range(5, 6));
+}  // namespace
diff --git a/libvpx/libvpx/test/active_map_test.cc b/libvpx/libvpx/test/active_map_test.cc
new file mode 100644
index 0000000..dc3de72
--- /dev/null
+++ b/libvpx/libvpx/test/active_map_test.cc
@@ -0,0 +1,89 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <climits>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+class ActiveMapTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
+ protected:
+  static const int kWidth = 208;
+  static const int kHeight = 144;
+
+  ActiveMapTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~ActiveMapTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    cpu_used_ = GET_PARAM(2);
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_CPUUSED, cpu_used_);
+    } else if (video->frame() == 3) {
+      vpx_active_map_t map = vpx_active_map_t();
+      uint8_t active_map[9 * 13] = {
+        1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0,
+        1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0,
+        1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0,
+        1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0,
+        0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 1,
+        0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 1, 0, 1,
+        0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 1, 0, 1,
+        0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0, 1, 1,
+        1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 0,
+      };
+      map.cols = (kWidth + 15) / 16;
+      map.rows = (kHeight + 15) / 16;
+      ASSERT_EQ(map.cols, 13u);
+      ASSERT_EQ(map.rows, 9u);
+      map.active_map = active_map;
+      encoder->Control(VP8E_SET_ACTIVEMAP, &map);
+    } else if (video->frame() == 15) {
+      vpx_active_map_t map = vpx_active_map_t();
+      map.cols = (kWidth + 15) / 16;
+      map.rows = (kHeight + 15) / 16;
+      map.active_map = NULL;
+      encoder->Control(VP8E_SET_ACTIVEMAP, &map);
+    }
+  }
+
+  int cpu_used_;
+};
+
+TEST_P(ActiveMapTest, Test) {
+  // Validate that this non multiple of 64 wide clip encodes
+  cfg_.g_lag_in_frames = 0;
+  cfg_.rc_target_bitrate = 400;
+  cfg_.rc_resize_allowed = 0;
+  cfg_.g_pass = VPX_RC_ONE_PASS;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.kf_max_dist = 90000;
+
+  ::libvpx_test::I420VideoSource video("hantro_odd.yuv", kWidth, kHeight, 30,
+                                       1, 0, 20);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+VP9_INSTANTIATE_TEST_CASE(ActiveMapTest,
+                          ::testing::Values(::libvpx_test::kRealTime),
+                          ::testing::Range(0, 9));
+}  // namespace
diff --git a/libvpx/libvpx/test/add_noise_test.cc b/libvpx/libvpx/test/add_noise_test.cc
new file mode 100644
index 0000000..e9945c4
--- /dev/null
+++ b/libvpx/libvpx/test/add_noise_test.cc
@@ -0,0 +1,197 @@
+/*
+ *  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <math.h>
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+
+// TODO(jimbankoski): make width and height integers not unsigned.
+typedef void (*AddNoiseFunc)(unsigned char *start, char *noise,
+                             char blackclamp[16], char whiteclamp[16],
+                             char bothclamp[16], unsigned int width,
+                             unsigned int height, int pitch);
+
+class AddNoiseTest
+    : public ::testing::TestWithParam<AddNoiseFunc> {
+ public:
+  virtual void TearDown() {
+    libvpx_test::ClearSystemState();
+  }
+  virtual ~AddNoiseTest() {}
+};
+
+double stddev6(char a, char b, char c, char d, char e, char f) {
+  const double n = (a + b + c + d + e + f) / 6.0;
+  const double v = ((a - n) * (a - n) + (b - n) * (b - n) + (c - n) * (c - n) +
+                    (d - n) * (d - n) + (e - n) * (e - n) + (f - n) * (f - n)) /
+                   6.0;
+  return sqrt(v);
+}
+
+// TODO(jimbankoski): The following 2 functions are duplicated in each codec.
+// For now the vp9 one has been copied into the test as is. We should normalize
+// these in vpx_dsp and not have 3 copies of these unless there is different
+// noise we add for each codec.
+
+double gaussian(double sigma, double mu, double x) {
+  return 1 / (sigma * sqrt(2.0 * 3.14159265)) *
+         (exp(-(x - mu) * (x - mu) / (2 * sigma * sigma)));
+}
+
+int setup_noise(int size_noise, char *noise) {
+  char char_dist[300];
+  const int ai = 4;
+  const int qi = 24;
+  const double sigma = ai + .5 + .6 * (63 - qi) / 63.0;
+
+  /* set up a lookup table of 256 entries that matches
+   * a gaussian distribution with sigma determined by q.
+   */
+  int next = 0;
+
+  for (int i = -32; i < 32; i++) {
+    int a_i = (int) (0.5 + 256 * gaussian(sigma, 0, i));
+
+    if (a_i) {
+      for (int j = 0; j < a_i; j++) {
+        char_dist[next + j] = (char)(i);
+      }
+
+      next = next + a_i;
+    }
+  }
+
+  for (; next < 256; next++)
+    char_dist[next] = 0;
+
+  for (int i = 0; i < size_noise; i++) {
+    noise[i] = char_dist[rand() & 0xff];  // NOLINT
+  }
+
+  // Returns the most negative value in distribution.
+  return char_dist[0];
+}
+
+TEST_P(AddNoiseTest, CheckNoiseAdded) {
+  DECLARE_ALIGNED(16, char, blackclamp[16]);
+  DECLARE_ALIGNED(16, char, whiteclamp[16]);
+  DECLARE_ALIGNED(16, char, bothclamp[16]);
+  const int width  = 64;
+  const int height = 64;
+  const int image_size = width * height;
+  char noise[3072];
+
+  const int clamp = setup_noise(3072, noise);
+  for (int i = 0; i < 16; i++) {
+    blackclamp[i] = -clamp;
+    whiteclamp[i] = -clamp;
+    bothclamp[i] = -2 * clamp;
+  }
+
+  uint8_t *const s = reinterpret_cast<uint8_t *>(vpx_calloc(image_size, 1));
+  memset(s, 99, image_size);
+
+  ASM_REGISTER_STATE_CHECK(GetParam()(s, noise, blackclamp, whiteclamp,
+                                      bothclamp, width, height, width));
+
+  // Check to make sure we don't end up having either the same or no added
+  // noise either vertically or horizontally.
+  for (int i = 0; i < image_size - 6 * width - 6; ++i) {
+    const double hd = stddev6(s[i] - 99, s[i + 1] - 99, s[i + 2] - 99,
+                              s[i + 3] - 99, s[i + 4] - 99, s[i + 5] - 99);
+    const double vd = stddev6(s[i] - 99, s[i + width] - 99,
+                              s[i + 2 * width] - 99, s[i + 3 * width] - 99,
+                              s[i + 4 * width] - 99, s[i + 5 * width] - 99);
+
+    EXPECT_NE(hd, 0);
+    EXPECT_NE(vd, 0);
+  }
+
+  // Initialize pixels in the image to 255 and check for roll over.
+  memset(s, 255, image_size);
+
+  ASM_REGISTER_STATE_CHECK(GetParam()(s, noise, blackclamp, whiteclamp,
+                                      bothclamp, width, height, width));
+
+  // Check to make sure don't roll over.
+  for (int i = 0; i < image_size; ++i) {
+    EXPECT_GT((int)s[i], 10) << "i = " << i;
+  }
+
+  // Initialize pixels in the image to 0 and check for roll under.
+  memset(s, 0, image_size);
+
+  ASM_REGISTER_STATE_CHECK(GetParam()(s, noise, blackclamp, whiteclamp,
+                                      bothclamp, width, height, width));
+
+  // Check to make sure don't roll under.
+  for (int i = 0; i < image_size; ++i) {
+    EXPECT_LT((int)s[i], 245) << "i = " << i;
+  }
+
+  vpx_free(s);
+}
+
+TEST_P(AddNoiseTest, CheckCvsAssembly) {
+  DECLARE_ALIGNED(16, char, blackclamp[16]);
+  DECLARE_ALIGNED(16, char, whiteclamp[16]);
+  DECLARE_ALIGNED(16, char, bothclamp[16]);
+  const int width  = 64;
+  const int height = 64;
+  const int image_size = width * height;
+  char noise[3072];
+
+  const int clamp = setup_noise(3072, noise);
+  for (int i = 0; i < 16; i++) {
+    blackclamp[i] = -clamp;
+    whiteclamp[i] = -clamp;
+    bothclamp[i] = -2 * clamp;
+  }
+
+  uint8_t *const s = reinterpret_cast<uint8_t *>(vpx_calloc(image_size, 1));
+  uint8_t *const d = reinterpret_cast<uint8_t *>(vpx_calloc(image_size, 1));
+
+  memset(s, 99, image_size);
+  memset(d, 99, image_size);
+
+  srand(0);
+  ASM_REGISTER_STATE_CHECK(GetParam()(s, noise, blackclamp, whiteclamp,
+                                      bothclamp, width, height, width));
+  srand(0);
+  ASM_REGISTER_STATE_CHECK(vpx_plane_add_noise_c(d, noise, blackclamp,
+                                                 whiteclamp, bothclamp,
+                                                 width, height, width));
+
+  for (int i = 0; i < image_size; ++i) {
+    EXPECT_EQ((int)s[i], (int)d[i]) << "i = " << i;
+  }
+
+  vpx_free(d);
+  vpx_free(s);
+}
+
+INSTANTIATE_TEST_CASE_P(C, AddNoiseTest,
+                        ::testing::Values(vpx_plane_add_noise_c));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, AddNoiseTest,
+                        ::testing::Values(vpx_plane_add_noise_sse2));
+#endif
+
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(MSA, AddNoiseTest,
+                        ::testing::Values(vpx_plane_add_noise_msa));
+#endif
+}  // namespace
diff --git a/libvpx/libvpx/test/altref_test.cc b/libvpx/libvpx/test/altref_test.cc
new file mode 100644
index 0000000..d9f83d8
--- /dev/null
+++ b/libvpx/libvpx/test/altref_test.cc
@@ -0,0 +1,159 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+namespace {
+
+#if CONFIG_VP8_ENCODER
+
+// lookahead range: [kLookAheadMin, kLookAheadMax).
+const int kLookAheadMin = 5;
+const int kLookAheadMax = 26;
+
+class AltRefTest : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWithParam<int> {
+ protected:
+  AltRefTest() : EncoderTest(GET_PARAM(0)), altref_count_(0) {}
+  virtual ~AltRefTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(libvpx_test::kTwoPassGood);
+  }
+
+  virtual void BeginPassHook(unsigned int pass) {
+    altref_count_ = 0;
+  }
+
+  virtual void PreEncodeFrameHook(libvpx_test::VideoSource *video,
+                                  libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+      encoder->Control(VP8E_SET_CPUUSED, 3);
+    }
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->data.frame.flags & VPX_FRAME_IS_INVISIBLE) ++altref_count_;
+  }
+
+  int altref_count() const { return altref_count_; }
+
+ private:
+  int altref_count_;
+};
+
+TEST_P(AltRefTest, MonotonicTimestamps) {
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 1000;
+  cfg_.g_lag_in_frames = GET_PARAM(1);
+
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 30);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_GE(altref_count(), 1);
+}
+
+VP8_INSTANTIATE_TEST_CASE(AltRefTest,
+                          ::testing::Range(kLookAheadMin, kLookAheadMax));
+
+#endif  // CONFIG_VP8_ENCODER
+
+class AltRefForcedKeyTestLarge
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
+ protected:
+  AltRefForcedKeyTestLarge()
+      : EncoderTest(GET_PARAM(0)),
+        encoding_mode_(GET_PARAM(1)),
+        cpu_used_(GET_PARAM(2)),
+        forced_kf_frame_num_(1),
+        frame_num_(0) {}
+  virtual ~AltRefForcedKeyTestLarge() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+    cfg_.rc_end_usage = VPX_VBR;
+    cfg_.g_threads = 0;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      encoder->Control(VP8E_SET_CPUUSED, cpu_used_);
+      encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+      // override test default for tile columns if necessary.
+#if CONFIG_VP9_ENCODER
+      if (GET_PARAM(0) == &libvpx_test::kVP9) {
+        encoder->Control(VP9E_SET_TILE_COLUMNS, 6);
+      }
+#endif
+    }
+    frame_flags_ =
+        (video->frame() == forced_kf_frame_num_) ? VPX_EFLAG_FORCE_KF : 0;
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (frame_num_ == forced_kf_frame_num_) {
+      ASSERT_TRUE(!!(pkt->data.frame.flags & VPX_FRAME_IS_KEY))
+          << "Frame #" << frame_num_ << " isn't a keyframe!";
+    }
+    ++frame_num_;
+  }
+
+  ::libvpx_test::TestMode encoding_mode_;
+  int cpu_used_;
+  unsigned int forced_kf_frame_num_;
+  unsigned int frame_num_;
+};
+
+TEST_P(AltRefForcedKeyTestLarge, Frame1IsKey) {
+  const vpx_rational timebase = { 1, 30 };
+  const int lag_values[] = { 3, 15, 25, -1 };
+
+  forced_kf_frame_num_ = 1;
+  for (int i = 0; lag_values[i] != -1; ++i) {
+    frame_num_ = 0;
+    cfg_.g_lag_in_frames = lag_values[i];
+    libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       timebase.den, timebase.num, 0, 30);
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  }
+}
+
+TEST_P(AltRefForcedKeyTestLarge, ForcedFrameIsKey) {
+  const vpx_rational timebase = { 1, 30 };
+  const int lag_values[] = { 3, 15, 25, -1 };
+
+  for (int i = 0; lag_values[i] != -1; ++i) {
+    frame_num_ = 0;
+    forced_kf_frame_num_ = lag_values[i] - 1;
+    cfg_.g_lag_in_frames = lag_values[i];
+    libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       timebase.den, timebase.num, 0, 30);
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  }
+}
+
+VP8_INSTANTIATE_TEST_CASE(
+    AltRefForcedKeyTestLarge,
+    ::testing::Values(::libvpx_test::kOnePassGood),
+    ::testing::Range(0, 9));
+
+VP9_INSTANTIATE_TEST_CASE(
+    AltRefForcedKeyTestLarge,
+    ::testing::Values(::libvpx_test::kOnePassGood),
+    ::testing::Range(0, 9));
+}  // namespace
diff --git a/libvpx/libvpx/test/android/Android.mk b/libvpx/libvpx/test/android/Android.mk
new file mode 100644
index 0000000..48872a2
--- /dev/null
+++ b/libvpx/libvpx/test/android/Android.mk
@@ -0,0 +1,56 @@
+# Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS.  All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+#
+# This make file builds vpx_test app for android.
+# The test app itself runs on the command line through adb shell
+# The paths are really messed up as the libvpx make file
+# expects to be made from a parent directory.
+CUR_WD := $(call my-dir)
+BINDINGS_DIR := $(CUR_WD)/../../..
+LOCAL_PATH := $(CUR_WD)/../../..
+
+#libwebm
+include $(CLEAR_VARS)
+include $(BINDINGS_DIR)/libvpx/third_party/libwebm/Android.mk
+LOCAL_PATH := $(CUR_WD)/../../..
+
+#libvpx
+include $(CLEAR_VARS)
+LOCAL_STATIC_LIBRARIES := libwebm
+include $(BINDINGS_DIR)/libvpx/build/make/Android.mk
+LOCAL_PATH := $(CUR_WD)/../..
+
+#libgtest
+include $(CLEAR_VARS)
+LOCAL_ARM_MODE := arm
+LOCAL_CPP_EXTENSION := .cc
+LOCAL_MODULE := gtest
+LOCAL_C_INCLUDES := $(LOCAL_PATH)/third_party/googletest/src/
+LOCAL_C_INCLUDES += $(LOCAL_PATH)/third_party/googletest/src/include/
+LOCAL_SRC_FILES := ./third_party/googletest/src/src/gtest-all.cc
+include $(BUILD_STATIC_LIBRARY)
+
+#libvpx_test
+include $(CLEAR_VARS)
+LOCAL_ARM_MODE := arm
+LOCAL_MODULE := libvpx_test
+LOCAL_STATIC_LIBRARIES := gtest libwebm
+
+ifeq ($(ENABLE_SHARED),1)
+  LOCAL_SHARED_LIBRARIES := vpx
+else
+  LOCAL_STATIC_LIBRARIES += vpx
+endif
+
+include $(LOCAL_PATH)/test/test.mk
+LOCAL_C_INCLUDES := $(BINDINGS_DIR)
+FILTERED_SRC := $(sort $(filter %.cc %.c, $(LIBVPX_TEST_SRCS-yes)))
+LOCAL_SRC_FILES := $(addprefix ./test/, $(FILTERED_SRC))
+# some test files depend on *_rtcd.h, ensure they're generated first.
+$(eval $(call rtcd_dep_template))
+include $(BUILD_EXECUTABLE)
diff --git a/libvpx/libvpx/test/android/README b/libvpx/libvpx/test/android/README
new file mode 100644
index 0000000..4a1adcf
--- /dev/null
+++ b/libvpx/libvpx/test/android/README
@@ -0,0 +1,32 @@
+Android.mk will build vpx unittests on android.
+1) Configure libvpx from the parent directory:
+./libvpx/configure --target=armv7-android-gcc --enable-external-build \
+  --enable-postproc --disable-install-srcs --enable-multi-res-encoding \
+  --enable-temporal-denoising --disable-unit-tests --disable-install-docs \
+  --disable-examples --disable-runtime-cpu-detect --sdk-path=$NDK
+
+2) From the parent directory, invoke ndk-build:
+NDK_PROJECT_PATH=. ndk-build APP_BUILD_SCRIPT=./libvpx/test/android/Android.mk \
+  APP_ABI=armeabi-v7a APP_PLATFORM=android-18 APP_OPTIM=release \
+  APP_STL=gnustl_static
+
+Note: Both adb and ndk-build are available prebuilt at:
+  https://chromium.googlesource.com/android_tools
+
+3) Run get_files.py to download the test files:
+python get_files.py -i /path/to/test-data.sha1 -o /path/to/put/files \
+  -u http://downloads.webmproject.org/test_data/libvpx
+
+4) Transfer files to device using adb. Ensure you have proper permissions for
+the target
+
+adb push /path/to/test_files /data/local/tmp
+adb push /path/to/built_libs /data/local/tmp
+
+NOTE: Built_libs defaults to parent_dir/libs/armeabi-v7a
+
+5) Run tests:
+adb shell
+(on device)
+cd /data/local/tmp
+LD_LIBRARY_PATH=. ./vpx_test
diff --git a/libvpx/libvpx/test/android/get_files.py b/libvpx/libvpx/test/android/get_files.py
new file mode 100644
index 0000000..1c69740
--- /dev/null
+++ b/libvpx/libvpx/test/android/get_files.py
@@ -0,0 +1,118 @@
+# Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS.  All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+#
+# This simple script pulls test files from the webm homepage
+# It is intelligent enough to only pull files if
+#   1) File / test_data folder does not exist
+#   2) SHA mismatch
+
+import pycurl
+import csv
+import hashlib
+import re
+import os.path
+import time
+import itertools
+import sys
+import getopt
+
+#globals
+url = ''
+file_list_path = ''
+local_resource_path = ''
+
+# Helper functions:
+# A simple function which returns the sha hash of a file in hex
+def get_file_sha(filename):
+  try:
+    sha_hash = hashlib.sha1()
+    with open(filename, 'rb') as file:
+      buf = file.read(HASH_CHUNK)
+      while len(buf) > 0:
+        sha_hash.update(buf)
+        buf = file.read(HASH_CHUNK)
+      return sha_hash.hexdigest()
+  except IOError:
+    print "Error reading " + filename
+
+# Downloads a file from a url, and then checks the sha against the passed
+# in sha
+def download_and_check_sha(url, filename, sha):
+  path = os.path.join(local_resource_path, filename)
+  fp = open(path, "wb")
+  curl = pycurl.Curl()
+  curl.setopt(pycurl.URL, url + "/" + filename)
+  curl.setopt(pycurl.WRITEDATA, fp)
+  curl.perform()
+  curl.close()
+  fp.close()
+  return get_file_sha(path) == sha
+
+#constants
+ftp_retries = 3
+
+SHA_COL = 0
+NAME_COL = 1
+EXPECTED_COL = 2
+HASH_CHUNK = 65536
+
+# Main script
+try:
+  opts, args = \
+      getopt.getopt(sys.argv[1:], \
+                    "u:i:o:", ["url=", "input_csv=", "output_dir="])
+except:
+  print 'get_files.py -u <url> -i <input_csv> -o <output_dir>'
+  sys.exit(2)
+
+for opt, arg in opts:
+  if opt == '-u':
+    url = arg
+  elif opt in ("-i", "--input_csv"):
+    file_list_path = os.path.join(arg)
+  elif opt in ("-o", "--output_dir"):
+    local_resource_path = os.path.join(arg)
+
+if len(sys.argv) != 7:
+  print "Expects two paths and a url!"
+  exit(1)
+
+if not os.path.isdir(local_resource_path):
+  os.makedirs(local_resource_path)
+
+file_list_csv = open(file_list_path, "rb")
+
+# Our 'csv' file uses multiple spaces as a delimiter, python's
+# csv class only uses single character delimiters, so we convert them below
+file_list_reader = csv.reader((re.sub(' +', ' ', line) \
+    for line in file_list_csv), delimiter = ' ')
+
+file_shas = []
+file_names = []
+
+for row in file_list_reader:
+  if len(row) != EXPECTED_COL:
+      continue
+  file_shas.append(row[SHA_COL])
+  file_names.append(row[NAME_COL])
+
+file_list_csv.close()
+
+# Download files, only if they don't already exist and have correct shas
+for filename, sha in itertools.izip(file_names, file_shas):
+  path = os.path.join(local_resource_path, filename)
+  if os.path.isfile(path) \
+      and get_file_sha(path) == sha:
+    print path + ' exists, skipping'
+    continue
+  for retry in range(0, ftp_retries):
+    print "Downloading " + path
+    if not download_and_check_sha(url, filename, sha):
+      print "Sha does not match, retrying..."
+    else:
+      break
diff --git a/libvpx/libvpx/test/android/scrape_gtest_log.py b/libvpx/libvpx/test/android/scrape_gtest_log.py
new file mode 100644
index 0000000..487845c
--- /dev/null
+++ b/libvpx/libvpx/test/android/scrape_gtest_log.py
@@ -0,0 +1,57 @@
+# Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+#
+# Use of this source code is governed by a BSD-style license
+# that can be found in the LICENSE file in the root of the source
+# tree. An additional intellectual property rights grant can be found
+# in the file PATENTS.  All contributing project authors may
+# be found in the AUTHORS file in the root of the source tree.
+
+"""Standalone script which parses a gtest log for json.
+
+Json is returned returns as an array.  This script is used by the libvpx
+waterfall to gather json results mixed in with gtest logs.  This is
+dubious software engineering.
+"""
+
+import getopt
+import json
+import os
+import re
+import sys
+
+
+def main():
+  if len(sys.argv) != 3:
+    print "Expects a file to write json to!"
+    exit(1)
+
+  try:
+    opts, _ = \
+        getopt.getopt(sys.argv[1:], \
+                      'o:', ['output-json='])
+  except getopt.GetOptError:
+    print 'scrape_gtest_log.py -o <output_json>'
+    sys.exit(2)
+
+  output_json = ''
+  for opt, arg in opts:
+    if opt in ('-o', '--output-json'):
+      output_json = os.path.join(arg)
+
+  blob = sys.stdin.read()
+  json_string = '[' + ','.join('{' + x + '}' for x in
+                               re.findall(r'{([^}]*.?)}', blob)) + ']'
+  print blob
+
+  output = json.dumps(json.loads(json_string), indent=4, sort_keys=True)
+  print output
+
+  path = os.path.dirname(output_json)
+  if path and not os.path.exists(path):
+    os.makedirs(path)
+
+  outfile = open(output_json, 'w')
+  outfile.write(output)
+
+if __name__ == '__main__':
+  sys.exit(main())
diff --git a/libvpx/libvpx/test/aq_segment_test.cc b/libvpx/libvpx/test/aq_segment_test.cc
new file mode 100644
index 0000000..1b9c943
--- /dev/null
+++ b/libvpx/libvpx/test/aq_segment_test.cc
@@ -0,0 +1,109 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+class AqSegmentTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
+ protected:
+  AqSegmentTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~AqSegmentTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    set_cpu_used_ = GET_PARAM(2);
+    aq_mode_ = 0;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_CPUUSED, set_cpu_used_);
+      encoder->Control(VP9E_SET_AQ_MODE, aq_mode_);
+      encoder->Control(VP8E_SET_MAX_INTRA_BITRATE_PCT, 100);
+    }
+  }
+
+  int set_cpu_used_;
+  int aq_mode_;
+};
+
+// Validate that this AQ segmentation mode (AQ=1, variance_ap)
+// encodes and decodes without a mismatch.
+TEST_P(AqSegmentTest, TestNoMisMatchAQ1) {
+  cfg_.rc_min_quantizer = 8;
+  cfg_.rc_max_quantizer = 56;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_target_bitrate = 300;
+
+  aq_mode_ = 1;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                        30, 1, 0, 100);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+// Validate that this AQ segmentation mode (AQ=2, complexity_aq)
+// encodes and decodes without a mismatch.
+TEST_P(AqSegmentTest, TestNoMisMatchAQ2) {
+  cfg_.rc_min_quantizer = 8;
+  cfg_.rc_max_quantizer = 56;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_target_bitrate = 300;
+
+  aq_mode_ = 2;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                        30, 1, 0, 100);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+// Validate that this AQ segmentation mode (AQ=3, cyclic_refresh_aq)
+// encodes and decodes without a mismatch.
+TEST_P(AqSegmentTest, TestNoMisMatchAQ3) {
+  cfg_.rc_min_quantizer = 8;
+  cfg_.rc_max_quantizer = 56;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_target_bitrate = 300;
+
+  aq_mode_ = 3;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                        30, 1, 0, 100);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+VP9_INSTANTIATE_TEST_CASE(AqSegmentTest,
+                          ::testing::Values(::libvpx_test::kRealTime,
+                                            ::libvpx_test::kOnePassGood),
+                          ::testing::Range(3, 9));
+}  // namespace
diff --git a/libvpx/libvpx/test/avg_test.cc b/libvpx/libvpx/test/avg_test.cc
new file mode 100644
index 0000000..44d8dd7
--- /dev/null
+++ b/libvpx/libvpx/test/avg_test.cc
@@ -0,0 +1,411 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vpx_mem/vpx_mem.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+class AverageTestBase : public ::testing::Test {
+ public:
+  AverageTestBase(int width, int height) : width_(width), height_(height) {}
+
+  static void SetUpTestCase() {
+    source_data_ = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kDataBlockSize));
+  }
+
+  static void TearDownTestCase() {
+    vpx_free(source_data_);
+    source_data_ = NULL;
+  }
+
+  virtual void TearDown() {
+    libvpx_test::ClearSystemState();
+  }
+
+ protected:
+  // Handle blocks up to 4 blocks 64x64 with stride up to 128
+  static const int kDataAlignment = 16;
+  static const int kDataBlockSize = 64 * 128;
+
+  virtual void SetUp() {
+    source_stride_ = (width_ + 31) & ~31;
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+  }
+
+  // Sum Pixels
+  unsigned int ReferenceAverage8x8(const uint8_t* source, int pitch) {
+    unsigned int average = 0;
+    for (int h = 0; h < 8; ++h)
+      for (int w = 0; w < 8; ++w)
+        average += source[h * pitch + w];
+    return ((average + 32) >> 6);
+  }
+
+  unsigned int ReferenceAverage4x4(const uint8_t* source, int pitch) {
+    unsigned int average = 0;
+    for (int h = 0; h < 4; ++h)
+      for (int w = 0; w < 4; ++w)
+        average += source[h * pitch + w];
+    return ((average + 8) >> 4);
+  }
+
+  void FillConstant(uint8_t fill_constant) {
+    for (int i = 0; i < width_ * height_; ++i) {
+        source_data_[i] = fill_constant;
+    }
+  }
+
+  void FillRandom() {
+    for (int i = 0; i < width_ * height_; ++i) {
+        source_data_[i] = rnd_.Rand8();
+    }
+  }
+
+  int width_, height_;
+  static uint8_t* source_data_;
+  int source_stride_;
+
+  ACMRandom rnd_;
+};
+typedef unsigned int (*AverageFunction)(const uint8_t* s, int pitch);
+
+typedef std::tr1::tuple<int, int, int, int, AverageFunction> AvgFunc;
+
+class AverageTest
+    : public AverageTestBase,
+      public ::testing::WithParamInterface<AvgFunc>{
+ public:
+  AverageTest() : AverageTestBase(GET_PARAM(0), GET_PARAM(1)) {}
+
+ protected:
+  void CheckAverages() {
+    unsigned int expected = 0;
+    if (GET_PARAM(3) == 8) {
+      expected = ReferenceAverage8x8(source_data_+ GET_PARAM(2),
+                                     source_stride_);
+    } else  if (GET_PARAM(3) == 4) {
+      expected = ReferenceAverage4x4(source_data_+ GET_PARAM(2),
+                                     source_stride_);
+    }
+
+    ASM_REGISTER_STATE_CHECK(GET_PARAM(4)(source_data_+ GET_PARAM(2),
+                                          source_stride_));
+    unsigned int actual = GET_PARAM(4)(source_data_+ GET_PARAM(2),
+                                       source_stride_);
+
+    EXPECT_EQ(expected, actual);
+  }
+};
+
+typedef void (*IntProRowFunc)(int16_t hbuf[16], uint8_t const *ref,
+                              const int ref_stride, const int height);
+
+typedef std::tr1::tuple<int, IntProRowFunc, IntProRowFunc> IntProRowParam;
+
+class IntProRowTest
+    : public AverageTestBase,
+      public ::testing::WithParamInterface<IntProRowParam> {
+ public:
+  IntProRowTest()
+    : AverageTestBase(16, GET_PARAM(0)),
+      hbuf_asm_(NULL),
+      hbuf_c_(NULL) {
+    asm_func_ = GET_PARAM(1);
+    c_func_ = GET_PARAM(2);
+  }
+
+ protected:
+  virtual void SetUp() {
+    hbuf_asm_ = reinterpret_cast<int16_t*>(
+        vpx_memalign(kDataAlignment, sizeof(*hbuf_asm_) * 16));
+    hbuf_c_ = reinterpret_cast<int16_t*>(
+        vpx_memalign(kDataAlignment, sizeof(*hbuf_c_) * 16));
+  }
+
+  virtual void TearDown() {
+    vpx_free(hbuf_c_);
+    hbuf_c_ = NULL;
+    vpx_free(hbuf_asm_);
+    hbuf_asm_ = NULL;
+  }
+
+  void RunComparison() {
+    ASM_REGISTER_STATE_CHECK(c_func_(hbuf_c_, source_data_, 0, height_));
+    ASM_REGISTER_STATE_CHECK(asm_func_(hbuf_asm_, source_data_, 0, height_));
+    EXPECT_EQ(0, memcmp(hbuf_c_, hbuf_asm_, sizeof(*hbuf_c_) * 16))
+        << "Output mismatch";
+  }
+
+ private:
+  IntProRowFunc asm_func_;
+  IntProRowFunc c_func_;
+  int16_t *hbuf_asm_;
+  int16_t *hbuf_c_;
+};
+
+typedef int16_t (*IntProColFunc)(uint8_t const *ref, const int width);
+
+typedef std::tr1::tuple<int, IntProColFunc, IntProColFunc> IntProColParam;
+
+class IntProColTest
+    : public AverageTestBase,
+      public ::testing::WithParamInterface<IntProColParam> {
+ public:
+  IntProColTest() : AverageTestBase(GET_PARAM(0), 1), sum_asm_(0), sum_c_(0) {
+    asm_func_ = GET_PARAM(1);
+    c_func_ = GET_PARAM(2);
+  }
+
+ protected:
+  void RunComparison() {
+    ASM_REGISTER_STATE_CHECK(sum_c_ = c_func_(source_data_, width_));
+    ASM_REGISTER_STATE_CHECK(sum_asm_ = asm_func_(source_data_, width_));
+    EXPECT_EQ(sum_c_, sum_asm_) << "Output mismatch";
+  }
+
+ private:
+  IntProColFunc asm_func_;
+  IntProColFunc c_func_;
+  int16_t sum_asm_;
+  int16_t sum_c_;
+};
+
+typedef int (*SatdFunc)(const int16_t *coeffs, int length);
+typedef std::tr1::tuple<int, SatdFunc> SatdTestParam;
+
+class SatdTest
+    : public ::testing::Test,
+      public ::testing::WithParamInterface<SatdTestParam> {
+ protected:
+  virtual void SetUp() {
+    satd_size_ = GET_PARAM(0);
+    satd_func_ = GET_PARAM(1);
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+    src_ = reinterpret_cast<int16_t*>(
+        vpx_memalign(16, sizeof(*src_) * satd_size_));
+    ASSERT_TRUE(src_ != NULL);
+  }
+
+  virtual void TearDown() {
+    libvpx_test::ClearSystemState();
+    vpx_free(src_);
+  }
+
+  void FillConstant(const int16_t val) {
+    for (int i = 0; i < satd_size_; ++i) src_[i] = val;
+  }
+
+  void FillRandom() {
+    for (int i = 0; i < satd_size_; ++i) src_[i] = rnd_.Rand16();
+  }
+
+  void Check(const int expected) {
+    int total;
+    ASM_REGISTER_STATE_CHECK(total = satd_func_(src_, satd_size_));
+    EXPECT_EQ(expected, total);
+  }
+
+  int satd_size_;
+
+ private:
+  int16_t *src_;
+  SatdFunc satd_func_;
+  ACMRandom rnd_;
+};
+
+uint8_t* AverageTestBase::source_data_ = NULL;
+
+TEST_P(AverageTest, MinValue) {
+  FillConstant(0);
+  CheckAverages();
+}
+
+TEST_P(AverageTest, MaxValue) {
+  FillConstant(255);
+  CheckAverages();
+}
+
+TEST_P(AverageTest, Random) {
+  // The reference frame, but not the source frame, may be unaligned for
+  // certain types of searches.
+  for (int i = 0; i < 1000; i++) {
+    FillRandom();
+    CheckAverages();
+  }
+}
+
+TEST_P(IntProRowTest, MinValue) {
+  FillConstant(0);
+  RunComparison();
+}
+
+TEST_P(IntProRowTest, MaxValue) {
+  FillConstant(255);
+  RunComparison();
+}
+
+TEST_P(IntProRowTest, Random) {
+  FillRandom();
+  RunComparison();
+}
+
+TEST_P(IntProColTest, MinValue) {
+  FillConstant(0);
+  RunComparison();
+}
+
+TEST_P(IntProColTest, MaxValue) {
+  FillConstant(255);
+  RunComparison();
+}
+
+TEST_P(IntProColTest, Random) {
+  FillRandom();
+  RunComparison();
+}
+
+
+TEST_P(SatdTest, MinValue) {
+  const int kMin = -32640;
+  const int expected = -kMin * satd_size_;
+  FillConstant(kMin);
+  Check(expected);
+}
+
+TEST_P(SatdTest, MaxValue) {
+  const int kMax = 32640;
+  const int expected = kMax * satd_size_;
+  FillConstant(kMax);
+  Check(expected);
+}
+
+TEST_P(SatdTest, Random) {
+  int expected;
+  switch (satd_size_) {
+    case 16: expected = 205298; break;
+    case 64: expected = 1113950; break;
+    case 256: expected = 4268415; break;
+    case 1024: expected = 16954082; break;
+    default:
+      FAIL() << "Invalid satd size (" << satd_size_
+             << ") valid: 16/64/256/1024";
+  }
+  FillRandom();
+  Check(expected);
+}
+
+using std::tr1::make_tuple;
+
+INSTANTIATE_TEST_CASE_P(
+    C, AverageTest,
+    ::testing::Values(
+        make_tuple(16, 16, 1, 8, &vpx_avg_8x8_c),
+        make_tuple(16, 16, 1, 4, &vpx_avg_4x4_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    C, SatdTest,
+    ::testing::Values(
+        make_tuple(16, &vpx_satd_c),
+        make_tuple(64, &vpx_satd_c),
+        make_tuple(256, &vpx_satd_c),
+        make_tuple(1024, &vpx_satd_c)));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, AverageTest,
+    ::testing::Values(
+        make_tuple(16, 16, 0, 8, &vpx_avg_8x8_sse2),
+        make_tuple(16, 16, 5, 8, &vpx_avg_8x8_sse2),
+        make_tuple(32, 32, 15, 8, &vpx_avg_8x8_sse2),
+        make_tuple(16, 16, 0, 4, &vpx_avg_4x4_sse2),
+        make_tuple(16, 16, 5, 4, &vpx_avg_4x4_sse2),
+        make_tuple(32, 32, 15, 4, &vpx_avg_4x4_sse2)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, IntProRowTest, ::testing::Values(
+        make_tuple(16, &vpx_int_pro_row_sse2, &vpx_int_pro_row_c),
+        make_tuple(32, &vpx_int_pro_row_sse2, &vpx_int_pro_row_c),
+        make_tuple(64, &vpx_int_pro_row_sse2, &vpx_int_pro_row_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, IntProColTest, ::testing::Values(
+        make_tuple(16, &vpx_int_pro_col_sse2, &vpx_int_pro_col_c),
+        make_tuple(32, &vpx_int_pro_col_sse2, &vpx_int_pro_col_c),
+        make_tuple(64, &vpx_int_pro_col_sse2, &vpx_int_pro_col_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, SatdTest,
+    ::testing::Values(
+        make_tuple(16, &vpx_satd_sse2),
+        make_tuple(64, &vpx_satd_sse2),
+        make_tuple(256, &vpx_satd_sse2),
+        make_tuple(1024, &vpx_satd_sse2)));
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+    NEON, AverageTest,
+    ::testing::Values(
+        make_tuple(16, 16, 0, 8, &vpx_avg_8x8_neon),
+        make_tuple(16, 16, 5, 8, &vpx_avg_8x8_neon),
+        make_tuple(32, 32, 15, 8, &vpx_avg_8x8_neon),
+        make_tuple(16, 16, 0, 4, &vpx_avg_4x4_neon),
+        make_tuple(16, 16, 5, 4, &vpx_avg_4x4_neon),
+        make_tuple(32, 32, 15, 4, &vpx_avg_4x4_neon)));
+
+INSTANTIATE_TEST_CASE_P(
+    NEON, IntProRowTest, ::testing::Values(
+        make_tuple(16, &vpx_int_pro_row_neon, &vpx_int_pro_row_c),
+        make_tuple(32, &vpx_int_pro_row_neon, &vpx_int_pro_row_c),
+        make_tuple(64, &vpx_int_pro_row_neon, &vpx_int_pro_row_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    NEON, IntProColTest, ::testing::Values(
+        make_tuple(16, &vpx_int_pro_col_neon, &vpx_int_pro_col_c),
+        make_tuple(32, &vpx_int_pro_col_neon, &vpx_int_pro_col_c),
+        make_tuple(64, &vpx_int_pro_col_neon, &vpx_int_pro_col_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    NEON, SatdTest,
+    ::testing::Values(
+        make_tuple(16, &vpx_satd_neon),
+        make_tuple(64, &vpx_satd_neon),
+        make_tuple(256, &vpx_satd_neon),
+        make_tuple(1024, &vpx_satd_neon)));
+#endif
+
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(
+    MSA, AverageTest,
+    ::testing::Values(
+        make_tuple(16, 16, 0, 8, &vpx_avg_8x8_msa),
+        make_tuple(16, 16, 5, 8, &vpx_avg_8x8_msa),
+        make_tuple(32, 32, 15, 8, &vpx_avg_8x8_msa),
+        make_tuple(16, 16, 0, 4, &vpx_avg_4x4_msa),
+        make_tuple(16, 16, 5, 4, &vpx_avg_4x4_msa),
+        make_tuple(32, 32, 15, 4, &vpx_avg_4x4_msa)));
+#endif
+
+}  // namespace
diff --git a/libvpx/libvpx/test/blockiness_test.cc b/libvpx/libvpx/test/blockiness_test.cc
new file mode 100644
index 0000000..0c60baa
--- /dev/null
+++ b/libvpx/libvpx/test/blockiness_test.cc
@@ -0,0 +1,229 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#if CONFIG_VP9_ENCODER
+#include "./vp9_rtcd.h"
+#endif
+
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+#include "vpx_mem/vpx_mem.h"
+
+
+extern "C"
+double vp9_get_blockiness(const unsigned char *img1, int img1_pitch,
+                          const unsigned char *img2, int img2_pitch,
+                          int width, int height);
+
+using libvpx_test::ACMRandom;
+
+namespace {
+class BlockinessTestBase : public ::testing::Test {
+ public:
+  BlockinessTestBase(int width, int height) : width_(width), height_(height) {}
+
+  static void SetUpTestCase() {
+    source_data_ = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kDataBufferSize));
+    reference_data_ = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kDataBufferSize));
+  }
+
+  static void TearDownTestCase() {
+    vpx_free(source_data_);
+    source_data_ = NULL;
+    vpx_free(reference_data_);
+    reference_data_ = NULL;
+  }
+
+  virtual void TearDown() {
+    libvpx_test::ClearSystemState();
+  }
+
+ protected:
+  // Handle frames up to 640x480
+  static const int kDataAlignment = 16;
+  static const int kDataBufferSize = 640*480;
+
+  virtual void SetUp() {
+    source_stride_ = (width_ + 31) & ~31;
+    reference_stride_ = width_ * 2;
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+  }
+
+  void FillConstant(uint8_t *data, int stride, uint8_t fill_constant,
+                    int width, int height) {
+    for (int h = 0; h < height; ++h) {
+      for (int w = 0; w < width; ++w) {
+        data[h * stride + w] = fill_constant;
+      }
+    }
+  }
+
+  void FillConstant(uint8_t *data, int stride, uint8_t fill_constant) {
+    FillConstant(data, stride, fill_constant, width_, height_);
+  }
+
+  void FillRandom(uint8_t *data, int stride, int width, int height) {
+    for (int h = 0; h < height; ++h) {
+      for (int w = 0; w < width; ++w) {
+        data[h * stride + w] = rnd_.Rand8();
+      }
+    }
+  }
+
+  void FillRandom(uint8_t *data, int stride) {
+    FillRandom(data, stride, width_, height_);
+  }
+
+  void FillRandomBlocky(uint8_t *data, int stride) {
+    for (int h = 0; h < height_; h += 4) {
+      for (int w = 0; w < width_; w += 4) {
+        FillRandom(data + h * stride + w, stride, 4, 4);
+      }
+    }
+  }
+
+  void FillCheckerboard(uint8_t *data, int stride) {
+    for (int h = 0; h < height_; h += 4) {
+      for (int w = 0; w < width_; w += 4) {
+        if (((h/4) ^ (w/4)) & 1)
+          FillConstant(data + h * stride + w, stride, 255, 4, 4);
+        else
+          FillConstant(data + h * stride + w, stride, 0, 4, 4);
+      }
+    }
+  }
+
+  void Blur(uint8_t *data, int stride, int taps) {
+    int sum = 0;
+    int half_taps = taps / 2;
+    for (int h = 0; h < height_; ++h) {
+      for (int w = 0; w < taps; ++w) {
+        sum += data[w + h * stride];
+      }
+      for (int w = taps; w < width_; ++w) {
+        sum += data[w + h * stride] - data[w - taps + h * stride];
+        data[w - half_taps + h * stride] = (sum + half_taps) / taps;
+      }
+    }
+    for (int w = 0; w < width_; ++w) {
+      for (int h = 0; h < taps; ++h) {
+        sum += data[h + w * stride];
+      }
+      for (int h = taps; h < height_; ++h) {
+        sum += data[w + h * stride] - data[(h - taps) * stride + w];
+        data[(h - half_taps) * stride + w] = (sum + half_taps) / taps;
+      }
+    }
+  }
+  int width_, height_;
+  static uint8_t* source_data_;
+  int source_stride_;
+  static uint8_t* reference_data_;
+  int reference_stride_;
+
+  ACMRandom rnd_;
+};
+
+#if CONFIG_VP9_ENCODER
+typedef std::tr1::tuple<int, int> BlockinessParam;
+class BlockinessVP9Test
+    : public BlockinessTestBase,
+      public ::testing::WithParamInterface<BlockinessParam> {
+ public:
+  BlockinessVP9Test() : BlockinessTestBase(GET_PARAM(0), GET_PARAM(1)) {}
+
+ protected:
+  int CheckBlockiness() {
+    return vp9_get_blockiness(source_data_, source_stride_,
+                              reference_data_, reference_stride_,
+                              width_, height_);
+  }
+};
+#endif  // CONFIG_VP9_ENCODER
+
+uint8_t* BlockinessTestBase::source_data_ = NULL;
+uint8_t* BlockinessTestBase::reference_data_ = NULL;
+
+#if CONFIG_VP9_ENCODER
+TEST_P(BlockinessVP9Test, SourceBlockierThanReference) {
+  // Source is blockier than reference.
+  FillRandomBlocky(source_data_, source_stride_);
+  FillConstant(reference_data_, reference_stride_, 128);
+  int super_blocky = CheckBlockiness();
+
+  EXPECT_EQ(0, super_blocky) << "Blocky source should produce 0 blockiness.";
+}
+
+TEST_P(BlockinessVP9Test, ReferenceBlockierThanSource) {
+  // Source is blockier than reference.
+  FillConstant(source_data_, source_stride_, 128);
+  FillRandomBlocky(reference_data_, reference_stride_);
+  int super_blocky = CheckBlockiness();
+
+  EXPECT_GT(super_blocky, 0.0)
+      << "Blocky reference should score high for blockiness.";
+}
+
+TEST_P(BlockinessVP9Test, BlurringDecreasesBlockiness) {
+  // Source is blockier than reference.
+  FillConstant(source_data_, source_stride_, 128);
+  FillRandomBlocky(reference_data_, reference_stride_);
+  int super_blocky = CheckBlockiness();
+
+  Blur(reference_data_, reference_stride_, 4);
+  int less_blocky = CheckBlockiness();
+
+  EXPECT_GT(super_blocky, less_blocky)
+      << "A straight blur should decrease blockiness.";
+}
+
+TEST_P(BlockinessVP9Test, WorstCaseBlockiness) {
+  // Source is blockier than reference.
+  FillConstant(source_data_, source_stride_, 128);
+  FillCheckerboard(reference_data_, reference_stride_);
+
+  int super_blocky = CheckBlockiness();
+
+  Blur(reference_data_, reference_stride_, 4);
+  int less_blocky = CheckBlockiness();
+
+  EXPECT_GT(super_blocky, less_blocky)
+      << "A straight blur should decrease blockiness.";
+}
+#endif  // CONFIG_VP9_ENCODER
+
+
+using std::tr1::make_tuple;
+
+//------------------------------------------------------------------------------
+// C functions
+
+#if CONFIG_VP9_ENCODER
+const BlockinessParam c_vp9_tests[] = {
+  make_tuple(320, 240),
+  make_tuple(318, 242),
+  make_tuple(318, 238),
+};
+INSTANTIATE_TEST_CASE_P(C, BlockinessVP9Test, ::testing::ValuesIn(c_vp9_tests));
+#endif
+
+}  // namespace
diff --git a/libvpx/libvpx/test/borders_test.cc b/libvpx/libvpx/test/borders_test.cc
new file mode 100644
index 0000000..bd3ac39
--- /dev/null
+++ b/libvpx/libvpx/test/borders_test.cc
@@ -0,0 +1,83 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <climits>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+class BordersTest : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ protected:
+  BordersTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~BordersTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_CPUUSED, 1);
+      encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+      encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+      encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+      encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+    }
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->data.frame.flags & VPX_FRAME_IS_KEY) {
+    }
+  }
+};
+
+TEST_P(BordersTest, TestEncodeHighBitrate) {
+  // Validate that this non multiple of 64 wide clip encodes and decodes
+  // without a mismatch when passing in a very low max q.  This pushes
+  // the encoder to producing lots of big partitions which will likely
+  // extend into the border and test the border condition.
+  cfg_.g_lag_in_frames = 25;
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.rc_max_quantizer = 10;
+
+  ::libvpx_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       40);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+TEST_P(BordersTest, TestLowBitrate) {
+  // Validate that this clip encodes and decodes without a mismatch
+  // when passing in a very high min q.  This pushes the encoder to producing
+  // lots of small partitions which might will test the other condition.
+
+  cfg_.g_lag_in_frames = 25;
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 200;
+  cfg_.rc_min_quantizer = 40;
+
+  ::libvpx_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       40);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+VP9_INSTANTIATE_TEST_CASE(BordersTest, ::testing::Values(
+    ::libvpx_test::kTwoPassGood));
+}  // namespace
diff --git a/libvpx/libvpx/test/byte_alignment_test.cc b/libvpx/libvpx/test/byte_alignment_test.cc
new file mode 100644
index 0000000..3a808b0
--- /dev/null
+++ b/libvpx/libvpx/test/byte_alignment_test.cc
@@ -0,0 +1,189 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+
+namespace {
+
+#if CONFIG_WEBM_IO
+
+const int kLegacyByteAlignment = 0;
+const int kLegacyYPlaneByteAlignment = 32;
+const int kNumPlanesToCheck = 3;
+const char kVP9TestFile[] = "vp90-2-02-size-lf-1920x1080.webm";
+const char kVP9Md5File[] = "vp90-2-02-size-lf-1920x1080.webm.md5";
+
+struct ByteAlignmentTestParam {
+  int byte_alignment;
+  vpx_codec_err_t expected_value;
+  bool decode_remaining;
+};
+
+const ByteAlignmentTestParam kBaTestParams[] = {
+  {kLegacyByteAlignment, VPX_CODEC_OK, true},
+  {32, VPX_CODEC_OK, true},
+  {64, VPX_CODEC_OK, true},
+  {128, VPX_CODEC_OK, true},
+  {256, VPX_CODEC_OK, true},
+  {512, VPX_CODEC_OK, true},
+  {1024, VPX_CODEC_OK, true},
+  {1, VPX_CODEC_INVALID_PARAM, false},
+  {-2, VPX_CODEC_INVALID_PARAM, false},
+  {4, VPX_CODEC_INVALID_PARAM, false},
+  {16, VPX_CODEC_INVALID_PARAM, false},
+  {255, VPX_CODEC_INVALID_PARAM, false},
+  {2048, VPX_CODEC_INVALID_PARAM, false},
+};
+
+// Class for testing byte alignment of reference buffers.
+class ByteAlignmentTest
+    : public ::testing::TestWithParam<ByteAlignmentTestParam> {
+ protected:
+  ByteAlignmentTest()
+      : video_(NULL),
+        decoder_(NULL),
+        md5_file_(NULL) {}
+
+  virtual void SetUp() {
+    video_ = new libvpx_test::WebMVideoSource(kVP9TestFile);
+    ASSERT_TRUE(video_ != NULL);
+    video_->Init();
+    video_->Begin();
+
+    const vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+    decoder_ = new libvpx_test::VP9Decoder(cfg, 0);
+    ASSERT_TRUE(decoder_ != NULL);
+
+    OpenMd5File(kVP9Md5File);
+  }
+
+  virtual void TearDown() {
+    if (md5_file_ != NULL)
+      fclose(md5_file_);
+
+    delete decoder_;
+    delete video_;
+  }
+
+  void SetByteAlignment(int byte_alignment, vpx_codec_err_t expected_value) {
+    decoder_->Control(VP9_SET_BYTE_ALIGNMENT, byte_alignment, expected_value);
+  }
+
+  vpx_codec_err_t DecodeOneFrame(int byte_alignment_to_check) {
+    const vpx_codec_err_t res =
+        decoder_->DecodeFrame(video_->cxdata(), video_->frame_size());
+    CheckDecodedFrames(byte_alignment_to_check);
+    if (res == VPX_CODEC_OK)
+      video_->Next();
+    return res;
+  }
+
+  vpx_codec_err_t DecodeRemainingFrames(int byte_alignment_to_check) {
+    for (; video_->cxdata() != NULL; video_->Next()) {
+      const vpx_codec_err_t res =
+          decoder_->DecodeFrame(video_->cxdata(), video_->frame_size());
+      if (res != VPX_CODEC_OK)
+        return res;
+      CheckDecodedFrames(byte_alignment_to_check);
+    }
+    return VPX_CODEC_OK;
+  }
+
+ private:
+  // Check if |data| is aligned to |byte_alignment_to_check|.
+  // |byte_alignment_to_check| must be a power of 2.
+  void CheckByteAlignment(const uint8_t *data, int byte_alignment_to_check) {
+    ASSERT_EQ(0u, reinterpret_cast<size_t>(data) % byte_alignment_to_check);
+  }
+
+  // Iterate through the planes of the decoded frames and check for
+  // alignment based off |byte_alignment_to_check|.
+  void CheckDecodedFrames(int byte_alignment_to_check) {
+    libvpx_test::DxDataIterator dec_iter = decoder_->GetDxData();
+    const vpx_image_t *img;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next()) != NULL) {
+      if (byte_alignment_to_check == kLegacyByteAlignment) {
+        CheckByteAlignment(img->planes[0], kLegacyYPlaneByteAlignment);
+      } else {
+        for (int i = 0; i < kNumPlanesToCheck; ++i) {
+          CheckByteAlignment(img->planes[i], byte_alignment_to_check);
+        }
+      }
+      CheckMd5(*img);
+    }
+  }
+
+  // TODO(fgalligan): Move the MD5 testing code into another class.
+  void OpenMd5File(const std::string &md5_file_name_) {
+    md5_file_ = libvpx_test::OpenTestDataFile(md5_file_name_);
+    ASSERT_TRUE(md5_file_ != NULL) << "MD5 file open failed. Filename: "
+        << md5_file_name_;
+  }
+
+  void CheckMd5(const vpx_image_t &img) {
+    ASSERT_TRUE(md5_file_ != NULL);
+    char expected_md5[33];
+    char junk[128];
+
+    // Read correct md5 checksums.
+    const int res = fscanf(md5_file_, "%s  %s", expected_md5, junk);
+    ASSERT_NE(EOF, res) << "Read md5 data failed";
+    expected_md5[32] = '\0';
+
+    ::libvpx_test::MD5 md5_res;
+    md5_res.Add(&img);
+    const char *const actual_md5 = md5_res.Get();
+
+    // Check md5 match.
+    ASSERT_STREQ(expected_md5, actual_md5) << "MD5 checksums don't match";
+  }
+
+  libvpx_test::WebMVideoSource *video_;
+  libvpx_test::VP9Decoder *decoder_;
+  FILE *md5_file_;
+};
+
+TEST_F(ByteAlignmentTest, SwitchByteAlignment) {
+  const int num_elements = 14;
+  const int byte_alignments[] = { 0, 32, 64, 128, 256, 512, 1024,
+                                  0, 1024, 32, 512, 64, 256, 128 };
+
+  for (int i = 0; i < num_elements; ++i) {
+    SetByteAlignment(byte_alignments[i], VPX_CODEC_OK);
+    ASSERT_EQ(VPX_CODEC_OK, DecodeOneFrame(byte_alignments[i]));
+  }
+  SetByteAlignment(byte_alignments[0], VPX_CODEC_OK);
+  ASSERT_EQ(VPX_CODEC_OK, DecodeRemainingFrames(byte_alignments[0]));
+}
+
+TEST_P(ByteAlignmentTest, TestAlignment) {
+  const ByteAlignmentTestParam t = GetParam();
+  SetByteAlignment(t.byte_alignment, t.expected_value);
+  if (t.decode_remaining)
+    ASSERT_EQ(VPX_CODEC_OK, DecodeRemainingFrames(t.byte_alignment));
+}
+
+INSTANTIATE_TEST_CASE_P(Alignments, ByteAlignmentTest,
+                        ::testing::ValuesIn(kBaTestParams));
+
+#endif  // CONFIG_WEBM_IO
+
+}  // namespace
diff --git a/libvpx/libvpx/test/clear_system_state.h b/libvpx/libvpx/test/clear_system_state.h
new file mode 100644
index 0000000..5e76797
--- /dev/null
+++ b/libvpx/libvpx/test/clear_system_state.h
@@ -0,0 +1,29 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef TEST_CLEAR_SYSTEM_STATE_H_
+#define TEST_CLEAR_SYSTEM_STATE_H_
+
+#include "./vpx_config.h"
+#if ARCH_X86 || ARCH_X86_64
+# include "vpx_ports/x86.h"
+#endif
+
+namespace libvpx_test {
+
+// Reset system to a known state. This function should be used for all non-API
+// test cases.
+inline void ClearSystemState() {
+#if ARCH_X86 || ARCH_X86_64
+  vpx_reset_mmx_state();
+#endif
+}
+
+}  // namespace libvpx_test
+#endif  // TEST_CLEAR_SYSTEM_STATE_H_
diff --git a/libvpx/libvpx/test/codec_factory.h b/libvpx/libvpx/test/codec_factory.h
new file mode 100644
index 0000000..429d40d
--- /dev/null
+++ b/libvpx/libvpx/test/codec_factory.h
@@ -0,0 +1,255 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef TEST_CODEC_FACTORY_H_
+#define TEST_CODEC_FACTORY_H_
+
+#include "./vpx_config.h"
+#include "vpx/vpx_decoder.h"
+#include "vpx/vpx_encoder.h"
+#if CONFIG_VP8_ENCODER || CONFIG_VP9_ENCODER
+#include "vpx/vp8cx.h"
+#endif
+#if CONFIG_VP8_DECODER || CONFIG_VP9_DECODER
+#include "vpx/vp8dx.h"
+#endif
+
+#include "test/decode_test_driver.h"
+#include "test/encode_test_driver.h"
+namespace libvpx_test {
+
+const int kCodecFactoryParam = 0;
+
+class CodecFactory {
+ public:
+  CodecFactory() {}
+
+  virtual ~CodecFactory() {}
+
+  virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
+                                 unsigned long deadline) const = 0;
+
+  virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
+                                 const vpx_codec_flags_t flags,
+                                 unsigned long deadline)  // NOLINT(runtime/int)
+                                 const = 0;
+
+  virtual Encoder* CreateEncoder(vpx_codec_enc_cfg_t cfg,
+                                 unsigned long deadline,
+                                 const unsigned long init_flags,
+                                 TwopassStatsStore *stats) const = 0;
+
+  virtual vpx_codec_err_t DefaultEncoderConfig(vpx_codec_enc_cfg_t *cfg,
+                                               int usage) const = 0;
+};
+
+/* Provide CodecTestWith<n>Params classes for a variable number of parameters
+ * to avoid having to include a pointer to the CodecFactory in every test
+ * definition.
+ */
+template<class T1>
+class CodecTestWithParam : public ::testing::TestWithParam<
+    std::tr1::tuple< const libvpx_test::CodecFactory*, T1 > > {
+};
+
+template<class T1, class T2>
+class CodecTestWith2Params : public ::testing::TestWithParam<
+    std::tr1::tuple< const libvpx_test::CodecFactory*, T1, T2 > > {
+};
+
+template<class T1, class T2, class T3>
+class CodecTestWith3Params : public ::testing::TestWithParam<
+    std::tr1::tuple< const libvpx_test::CodecFactory*, T1, T2, T3 > > {
+};
+
+/*
+ * VP8 Codec Definitions
+ */
+#if CONFIG_VP8
+class VP8Decoder : public Decoder {
+ public:
+  VP8Decoder(vpx_codec_dec_cfg_t cfg, unsigned long deadline)
+      : Decoder(cfg, deadline) {}
+
+  VP8Decoder(vpx_codec_dec_cfg_t cfg, const vpx_codec_flags_t flag,
+             unsigned long deadline)  // NOLINT
+      : Decoder(cfg, flag, deadline) {}
+
+ protected:
+  virtual vpx_codec_iface_t* CodecInterface() const {
+#if CONFIG_VP8_DECODER
+    return &vpx_codec_vp8_dx_algo;
+#else
+    return NULL;
+#endif
+  }
+};
+
+class VP8Encoder : public Encoder {
+ public:
+  VP8Encoder(vpx_codec_enc_cfg_t cfg, unsigned long deadline,
+             const unsigned long init_flags, TwopassStatsStore *stats)
+      : Encoder(cfg, deadline, init_flags, stats) {}
+
+ protected:
+  virtual vpx_codec_iface_t* CodecInterface() const {
+#if CONFIG_VP8_ENCODER
+    return &vpx_codec_vp8_cx_algo;
+#else
+    return NULL;
+#endif
+  }
+};
+
+class VP8CodecFactory : public CodecFactory {
+ public:
+  VP8CodecFactory() : CodecFactory() {}
+
+  virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
+                                 unsigned long deadline) const {
+    return CreateDecoder(cfg, 0, deadline);
+  }
+
+  virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
+                                 const vpx_codec_flags_t flags,
+                                 unsigned long deadline) const {  // NOLINT
+#if CONFIG_VP8_DECODER
+    return new VP8Decoder(cfg, flags, deadline);
+#else
+    return NULL;
+#endif
+  }
+
+  virtual Encoder* CreateEncoder(vpx_codec_enc_cfg_t cfg,
+                                 unsigned long deadline,
+                                 const unsigned long init_flags,
+                                 TwopassStatsStore *stats) const {
+#if CONFIG_VP8_ENCODER
+    return new VP8Encoder(cfg, deadline, init_flags, stats);
+#else
+    return NULL;
+#endif
+  }
+
+  virtual vpx_codec_err_t DefaultEncoderConfig(vpx_codec_enc_cfg_t *cfg,
+                                               int usage) const {
+#if CONFIG_VP8_ENCODER
+    return vpx_codec_enc_config_default(&vpx_codec_vp8_cx_algo, cfg, usage);
+#else
+    return VPX_CODEC_INCAPABLE;
+#endif
+  }
+};
+
+const libvpx_test::VP8CodecFactory kVP8;
+
+#define VP8_INSTANTIATE_TEST_CASE(test, ...)\
+  INSTANTIATE_TEST_CASE_P(VP8, test, \
+      ::testing::Combine( \
+          ::testing::Values(static_cast<const libvpx_test::CodecFactory*>( \
+              &libvpx_test::kVP8)), \
+          __VA_ARGS__))
+#else
+#define VP8_INSTANTIATE_TEST_CASE(test, ...)
+#endif  // CONFIG_VP8
+
+
+/*
+ * VP9 Codec Definitions
+ */
+#if CONFIG_VP9
+class VP9Decoder : public Decoder {
+ public:
+  VP9Decoder(vpx_codec_dec_cfg_t cfg, unsigned long deadline)
+      : Decoder(cfg, deadline) {}
+
+  VP9Decoder(vpx_codec_dec_cfg_t cfg, const vpx_codec_flags_t flag,
+             unsigned long deadline)  // NOLINT
+      : Decoder(cfg, flag, deadline) {}
+
+ protected:
+  virtual vpx_codec_iface_t* CodecInterface() const {
+#if CONFIG_VP9_DECODER
+    return &vpx_codec_vp9_dx_algo;
+#else
+    return NULL;
+#endif
+  }
+};
+
+class VP9Encoder : public Encoder {
+ public:
+  VP9Encoder(vpx_codec_enc_cfg_t cfg, unsigned long deadline,
+             const unsigned long init_flags, TwopassStatsStore *stats)
+      : Encoder(cfg, deadline, init_flags, stats) {}
+
+ protected:
+  virtual vpx_codec_iface_t* CodecInterface() const {
+#if CONFIG_VP9_ENCODER
+    return &vpx_codec_vp9_cx_algo;
+#else
+    return NULL;
+#endif
+  }
+};
+
+class VP9CodecFactory : public CodecFactory {
+ public:
+  VP9CodecFactory() : CodecFactory() {}
+
+  virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
+                                 unsigned long deadline) const {
+    return CreateDecoder(cfg, 0, deadline);
+  }
+
+  virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
+                                 const vpx_codec_flags_t flags,
+                                 unsigned long deadline) const {  // NOLINT
+#if CONFIG_VP9_DECODER
+    return new VP9Decoder(cfg, flags, deadline);
+#else
+    return NULL;
+#endif
+  }
+
+  virtual Encoder* CreateEncoder(vpx_codec_enc_cfg_t cfg,
+                                 unsigned long deadline,
+                                 const unsigned long init_flags,
+                                 TwopassStatsStore *stats) const {
+#if CONFIG_VP9_ENCODER
+    return new VP9Encoder(cfg, deadline, init_flags, stats);
+#else
+    return NULL;
+#endif
+  }
+
+  virtual vpx_codec_err_t DefaultEncoderConfig(vpx_codec_enc_cfg_t *cfg,
+                                               int usage) const {
+#if CONFIG_VP9_ENCODER
+    return vpx_codec_enc_config_default(&vpx_codec_vp9_cx_algo, cfg, usage);
+#else
+    return VPX_CODEC_INCAPABLE;
+#endif
+  }
+};
+
+const libvpx_test::VP9CodecFactory kVP9;
+
+#define VP9_INSTANTIATE_TEST_CASE(test, ...)\
+  INSTANTIATE_TEST_CASE_P(VP9, test, \
+      ::testing::Combine( \
+          ::testing::Values(static_cast<const libvpx_test::CodecFactory*>( \
+               &libvpx_test::kVP9)), \
+          __VA_ARGS__))
+#else
+#define VP9_INSTANTIATE_TEST_CASE(test, ...)
+#endif  // CONFIG_VP9
+
+}  // namespace libvpx_test
+#endif  // TEST_CODEC_FACTORY_H_
diff --git a/libvpx/libvpx/test/config_test.cc b/libvpx/libvpx/test/config_test.cc
new file mode 100644
index 0000000..0493110
--- /dev/null
+++ b/libvpx/libvpx/test/config_test.cc
@@ -0,0 +1,60 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/video_source.h"
+
+namespace {
+
+class ConfigTest : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ protected:
+  ConfigTest() : EncoderTest(GET_PARAM(0)),
+                 frame_count_in_(0), frame_count_out_(0), frame_count_max_(0) {}
+  virtual ~ConfigTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    frame_count_in_ = 0;
+    frame_count_out_ = 0;
+  }
+
+  virtual void PreEncodeFrameHook(libvpx_test::VideoSource* /*video*/) {
+    ++frame_count_in_;
+    abort_ |= (frame_count_in_ >= frame_count_max_);
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t* /*pkt*/) {
+    ++frame_count_out_;
+  }
+
+  unsigned int frame_count_in_;
+  unsigned int frame_count_out_;
+  unsigned int frame_count_max_;
+};
+
+TEST_P(ConfigTest, LagIsDisabled) {
+  frame_count_max_ = 2;
+  cfg_.g_lag_in_frames = 15;
+
+  libvpx_test::DummyVideoSource video;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  EXPECT_EQ(frame_count_in_, frame_count_out_);
+}
+
+VP8_INSTANTIATE_TEST_CASE(ConfigTest, ONE_PASS_TEST_MODES);
+}  // namespace
diff --git a/libvpx/libvpx/test/consistency_test.cc b/libvpx/libvpx/test/consistency_test.cc
new file mode 100644
index 0000000..9c2fd55
--- /dev/null
+++ b/libvpx/libvpx/test/consistency_test.cc
@@ -0,0 +1,224 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#if CONFIG_VP9_ENCODER
+#include "./vp9_rtcd.h"
+#endif
+
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vpx_dsp/ssim.h"
+#include "vpx_mem/vpx_mem.h"
+
+extern "C"
+double vpx_get_ssim_metrics(uint8_t *img1, int img1_pitch,
+                            uint8_t *img2, int img2_pitch,
+                            int width, int height,
+                            Ssimv *sv2, Metrics *m,
+                            int do_inconsistency);
+
+using libvpx_test::ACMRandom;
+
+namespace {
+class ConsistencyTestBase : public ::testing::Test {
+ public:
+  ConsistencyTestBase(int width, int height) : width_(width), height_(height) {}
+
+  static void SetUpTestCase() {
+    source_data_[0] = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kDataBufferSize));
+    reference_data_[0] = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kDataBufferSize));
+    source_data_[1] = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kDataBufferSize));
+    reference_data_[1] = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kDataBufferSize));
+    ssim_array_ = new Ssimv[kDataBufferSize / 16];
+  }
+
+  static void ClearSsim() {
+    memset(ssim_array_, 0, kDataBufferSize / 16);
+  }
+  static void TearDownTestCase() {
+    vpx_free(source_data_[0]);
+    source_data_[0] = NULL;
+    vpx_free(reference_data_[0]);
+    reference_data_[0] = NULL;
+    vpx_free(source_data_[1]);
+    source_data_[1] = NULL;
+    vpx_free(reference_data_[1]);
+    reference_data_[1] = NULL;
+
+    delete[] ssim_array_;
+  }
+
+  virtual void TearDown() {
+    libvpx_test::ClearSystemState();
+  }
+
+ protected:
+  // Handle frames up to 640x480
+  static const int kDataAlignment = 16;
+  static const int kDataBufferSize = 640*480;
+
+  virtual void SetUp() {
+    source_stride_ = (width_ + 31) & ~31;
+    reference_stride_ = width_ * 2;
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+  }
+
+  void FillRandom(uint8_t *data, int stride, int width, int height) {
+    for (int h = 0; h < height; ++h) {
+      for (int w = 0; w < width; ++w) {
+        data[h * stride + w] = rnd_.Rand8();
+      }
+    }
+  }
+
+  void FillRandom(uint8_t *data, int stride) {
+    FillRandom(data, stride, width_, height_);
+  }
+
+  void Copy(uint8_t *reference, uint8_t *source) {
+    memcpy(reference, source, kDataBufferSize);
+  }
+
+  void Blur(uint8_t *data, int stride, int taps) {
+    int sum = 0;
+    int half_taps = taps / 2;
+    for (int h = 0; h < height_; ++h) {
+      for (int w = 0; w < taps; ++w) {
+        sum += data[w + h * stride];
+      }
+      for (int w = taps; w < width_; ++w) {
+        sum += data[w + h * stride] - data[w - taps + h * stride];
+        data[w - half_taps + h * stride] = (sum + half_taps) / taps;
+      }
+    }
+    for (int w = 0; w < width_; ++w) {
+      for (int h = 0; h < taps; ++h) {
+        sum += data[h + w * stride];
+      }
+      for (int h = taps; h < height_; ++h) {
+        sum += data[w + h * stride] - data[(h - taps) * stride + w];
+        data[(h - half_taps) * stride + w] = (sum + half_taps) / taps;
+      }
+    }
+  }
+  int width_, height_;
+  static uint8_t* source_data_[2];
+  int source_stride_;
+  static uint8_t* reference_data_[2];
+  int reference_stride_;
+  static Ssimv *ssim_array_;
+  Metrics metrics_;
+
+  ACMRandom rnd_;
+};
+
+#if CONFIG_VP9_ENCODER
+typedef std::tr1::tuple<int, int> ConsistencyParam;
+class ConsistencyVP9Test
+    : public ConsistencyTestBase,
+      public ::testing::WithParamInterface<ConsistencyParam> {
+ public:
+  ConsistencyVP9Test() : ConsistencyTestBase(GET_PARAM(0), GET_PARAM(1)) {}
+
+ protected:
+  double CheckConsistency(int frame) {
+    EXPECT_LT(frame, 2)<< "Frame to check has to be less than 2.";
+    return
+        vpx_get_ssim_metrics(source_data_[frame], source_stride_,
+                             reference_data_[frame], reference_stride_,
+                             width_, height_, ssim_array_, &metrics_, 1);
+  }
+};
+#endif  // CONFIG_VP9_ENCODER
+
+uint8_t* ConsistencyTestBase::source_data_[2] = {NULL, NULL};
+uint8_t* ConsistencyTestBase::reference_data_[2] = {NULL, NULL};
+Ssimv* ConsistencyTestBase::ssim_array_ = NULL;
+
+#if CONFIG_VP9_ENCODER
+TEST_P(ConsistencyVP9Test, ConsistencyIsZero) {
+  FillRandom(source_data_[0], source_stride_);
+  Copy(source_data_[1], source_data_[0]);
+  Copy(reference_data_[0], source_data_[0]);
+  Blur(reference_data_[0], reference_stride_, 3);
+  Copy(reference_data_[1], source_data_[0]);
+  Blur(reference_data_[1], reference_stride_, 3);
+
+  double inconsistency = CheckConsistency(1);
+  inconsistency = CheckConsistency(0);
+  EXPECT_EQ(inconsistency, 0.0)
+      << "Should have 0 inconsistency if they are exactly the same.";
+
+  // If sources are not consistent reference frames inconsistency should
+  // be less than if the source is consistent.
+  FillRandom(source_data_[0], source_stride_);
+  FillRandom(source_data_[1], source_stride_);
+  FillRandom(reference_data_[0], reference_stride_);
+  FillRandom(reference_data_[1], reference_stride_);
+  CheckConsistency(0);
+  inconsistency = CheckConsistency(1);
+
+  Copy(source_data_[1], source_data_[0]);
+  CheckConsistency(0);
+  double inconsistency2 = CheckConsistency(1);
+  EXPECT_LT(inconsistency, inconsistency2)
+      << "Should have less inconsistency if source itself is inconsistent.";
+
+  // Less of a blur should be less inconsistent than more blur coming off a
+  // a frame with no blur.
+  ClearSsim();
+  FillRandom(source_data_[0], source_stride_);
+  Copy(source_data_[1], source_data_[0]);
+  Copy(reference_data_[0], source_data_[0]);
+  Copy(reference_data_[1], source_data_[0]);
+  Blur(reference_data_[1], reference_stride_, 4);
+  CheckConsistency(0);
+  inconsistency = CheckConsistency(1);
+  ClearSsim();
+  Copy(reference_data_[1], source_data_[0]);
+  Blur(reference_data_[1], reference_stride_, 8);
+  CheckConsistency(0);
+  inconsistency2 = CheckConsistency(1);
+
+  EXPECT_LT(inconsistency, inconsistency2)
+      << "Stronger Blur should produce more inconsistency.";
+}
+#endif  // CONFIG_VP9_ENCODER
+
+
+using std::tr1::make_tuple;
+
+//------------------------------------------------------------------------------
+// C functions
+
+#if CONFIG_VP9_ENCODER
+const ConsistencyParam c_vp9_tests[] = {
+  make_tuple(320, 240),
+  make_tuple(318, 242),
+  make_tuple(318, 238),
+};
+INSTANTIATE_TEST_CASE_P(C, ConsistencyVP9Test,
+                        ::testing::ValuesIn(c_vp9_tests));
+#endif
+
+}  // namespace
diff --git a/libvpx/libvpx/test/convolve_test.cc b/libvpx/libvpx/test/convolve_test.cc
new file mode 100644
index 0000000..73b0edb
--- /dev/null
+++ b/libvpx/libvpx/test/convolve_test.cc
@@ -0,0 +1,1238 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_filter.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/vpx_filter.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+namespace {
+
+static const unsigned int kMaxDimension = 64;
+
+typedef void (*ConvolveFunc)(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int filter_x_stride,
+                             const int16_t *filter_y, int filter_y_stride,
+                             int w, int h);
+
+struct ConvolveFunctions {
+  ConvolveFunctions(ConvolveFunc copy, ConvolveFunc avg,
+                    ConvolveFunc h8, ConvolveFunc h8_avg,
+                    ConvolveFunc v8, ConvolveFunc v8_avg,
+                    ConvolveFunc hv8, ConvolveFunc hv8_avg,
+                    ConvolveFunc sh8, ConvolveFunc sh8_avg,
+                    ConvolveFunc sv8, ConvolveFunc sv8_avg,
+                    ConvolveFunc shv8, ConvolveFunc shv8_avg,
+                    int bd)
+      : copy_(copy), avg_(avg), h8_(h8), v8_(v8), hv8_(hv8), h8_avg_(h8_avg),
+        v8_avg_(v8_avg), hv8_avg_(hv8_avg), sh8_(sh8), sv8_(sv8), shv8_(shv8),
+        sh8_avg_(sh8_avg), sv8_avg_(sv8_avg), shv8_avg_(shv8_avg),
+        use_highbd_(bd) {}
+
+  ConvolveFunc copy_;
+  ConvolveFunc avg_;
+  ConvolveFunc h8_;
+  ConvolveFunc v8_;
+  ConvolveFunc hv8_;
+  ConvolveFunc h8_avg_;
+  ConvolveFunc v8_avg_;
+  ConvolveFunc hv8_avg_;
+  ConvolveFunc sh8_;        // scaled horiz
+  ConvolveFunc sv8_;        // scaled vert
+  ConvolveFunc shv8_;       // scaled horiz/vert
+  ConvolveFunc sh8_avg_;    // scaled avg horiz
+  ConvolveFunc sv8_avg_;    // scaled avg vert
+  ConvolveFunc shv8_avg_;   // scaled avg horiz/vert
+  int use_highbd_;  // 0 if high bitdepth not used, else the actual bit depth.
+};
+
+typedef std::tr1::tuple<int, int, const ConvolveFunctions *> ConvolveParam;
+
+#define ALL_SIZES(convolve_fn) \
+    make_tuple(4, 4, &convolve_fn),     \
+    make_tuple(8, 4, &convolve_fn),     \
+    make_tuple(4, 8, &convolve_fn),     \
+    make_tuple(8, 8, &convolve_fn),     \
+    make_tuple(16, 8, &convolve_fn),    \
+    make_tuple(8, 16, &convolve_fn),    \
+    make_tuple(16, 16, &convolve_fn),   \
+    make_tuple(32, 16, &convolve_fn),   \
+    make_tuple(16, 32, &convolve_fn),   \
+    make_tuple(32, 32, &convolve_fn),   \
+    make_tuple(64, 32, &convolve_fn),   \
+    make_tuple(32, 64, &convolve_fn),   \
+    make_tuple(64, 64, &convolve_fn)
+
+// Reference 8-tap subpixel filter, slightly modified to fit into this test.
+#define VP9_FILTER_WEIGHT 128
+#define VP9_FILTER_SHIFT 7
+uint8_t clip_pixel(int x) {
+  return x < 0 ? 0 :
+         x > 255 ? 255 :
+         x;
+}
+
+void filter_block2d_8_c(const uint8_t *src_ptr,
+                        const unsigned int src_stride,
+                        const int16_t *HFilter,
+                        const int16_t *VFilter,
+                        uint8_t *dst_ptr,
+                        unsigned int dst_stride,
+                        unsigned int output_width,
+                        unsigned int output_height) {
+  // Between passes, we use an intermediate buffer whose height is extended to
+  // have enough horizontally filtered values as input for the vertical pass.
+  // This buffer is allocated to be big enough for the largest block type we
+  // support.
+  const int kInterp_Extend = 4;
+  const unsigned int intermediate_height =
+      (kInterp_Extend - 1) + output_height + kInterp_Extend;
+  unsigned int i, j;
+
+  // Size of intermediate_buffer is max_intermediate_height * filter_max_width,
+  // where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height
+  //                                 + kInterp_Extend
+  //                               = 3 + 16 + 4
+  //                               = 23
+  // and filter_max_width          = 16
+  //
+  uint8_t intermediate_buffer[71 * kMaxDimension];
+  const int intermediate_next_stride =
+      1 - static_cast<int>(intermediate_height * output_width);
+
+  // Horizontal pass (src -> transposed intermediate).
+  uint8_t *output_ptr = intermediate_buffer;
+  const int src_next_row_stride = src_stride - output_width;
+  src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1);
+  for (i = 0; i < intermediate_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      // Apply filter...
+      const int temp = (src_ptr[0] * HFilter[0]) +
+          (src_ptr[1] * HFilter[1]) +
+          (src_ptr[2] * HFilter[2]) +
+          (src_ptr[3] * HFilter[3]) +
+          (src_ptr[4] * HFilter[4]) +
+          (src_ptr[5] * HFilter[5]) +
+          (src_ptr[6] * HFilter[6]) +
+          (src_ptr[7] * HFilter[7]) +
+          (VP9_FILTER_WEIGHT >> 1);  // Rounding
+
+      // Normalize back to 0-255...
+      *output_ptr = clip_pixel(temp >> VP9_FILTER_SHIFT);
+      ++src_ptr;
+      output_ptr += intermediate_height;
+    }
+    src_ptr += src_next_row_stride;
+    output_ptr += intermediate_next_stride;
+  }
+
+  // Vertical pass (transposed intermediate -> dst).
+  src_ptr = intermediate_buffer;
+  const int dst_next_row_stride = dst_stride - output_width;
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      // Apply filter...
+      const int temp = (src_ptr[0] * VFilter[0]) +
+          (src_ptr[1] * VFilter[1]) +
+          (src_ptr[2] * VFilter[2]) +
+          (src_ptr[3] * VFilter[3]) +
+          (src_ptr[4] * VFilter[4]) +
+          (src_ptr[5] * VFilter[5]) +
+          (src_ptr[6] * VFilter[6]) +
+          (src_ptr[7] * VFilter[7]) +
+          (VP9_FILTER_WEIGHT >> 1);  // Rounding
+
+      // Normalize back to 0-255...
+      *dst_ptr++ = clip_pixel(temp >> VP9_FILTER_SHIFT);
+      src_ptr += intermediate_height;
+    }
+    src_ptr += intermediate_next_stride;
+    dst_ptr += dst_next_row_stride;
+  }
+}
+
+void block2d_average_c(uint8_t *src,
+                       unsigned int src_stride,
+                       uint8_t *output_ptr,
+                       unsigned int output_stride,
+                       unsigned int output_width,
+                       unsigned int output_height) {
+  unsigned int i, j;
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      output_ptr[j] = (output_ptr[j] + src[i * src_stride + j] + 1) >> 1;
+    }
+    output_ptr += output_stride;
+  }
+}
+
+void filter_average_block2d_8_c(const uint8_t *src_ptr,
+                                const unsigned int src_stride,
+                                const int16_t *HFilter,
+                                const int16_t *VFilter,
+                                uint8_t *dst_ptr,
+                                unsigned int dst_stride,
+                                unsigned int output_width,
+                                unsigned int output_height) {
+  uint8_t tmp[kMaxDimension * kMaxDimension];
+
+  assert(output_width <= kMaxDimension);
+  assert(output_height <= kMaxDimension);
+  filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, 64,
+                     output_width, output_height);
+  block2d_average_c(tmp, 64, dst_ptr, dst_stride,
+                    output_width, output_height);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void highbd_filter_block2d_8_c(const uint16_t *src_ptr,
+                               const unsigned int src_stride,
+                               const int16_t *HFilter,
+                               const int16_t *VFilter,
+                               uint16_t *dst_ptr,
+                               unsigned int dst_stride,
+                               unsigned int output_width,
+                               unsigned int output_height,
+                               int bd) {
+  // Between passes, we use an intermediate buffer whose height is extended to
+  // have enough horizontally filtered values as input for the vertical pass.
+  // This buffer is allocated to be big enough for the largest block type we
+  // support.
+  const int kInterp_Extend = 4;
+  const unsigned int intermediate_height =
+      (kInterp_Extend - 1) + output_height + kInterp_Extend;
+
+  /* Size of intermediate_buffer is max_intermediate_height * filter_max_width,
+   * where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height
+   *                                 + kInterp_Extend
+   *                               = 3 + 16 + 4
+   *                               = 23
+   * and filter_max_width = 16
+   */
+  uint16_t intermediate_buffer[71 * kMaxDimension];
+  const int intermediate_next_stride =
+      1 - static_cast<int>(intermediate_height * output_width);
+
+  // Horizontal pass (src -> transposed intermediate).
+  {
+    uint16_t *output_ptr = intermediate_buffer;
+    const int src_next_row_stride = src_stride - output_width;
+    unsigned int i, j;
+    src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1);
+    for (i = 0; i < intermediate_height; ++i) {
+      for (j = 0; j < output_width; ++j) {
+        // Apply filter...
+        const int temp = (src_ptr[0] * HFilter[0]) +
+                         (src_ptr[1] * HFilter[1]) +
+                         (src_ptr[2] * HFilter[2]) +
+                         (src_ptr[3] * HFilter[3]) +
+                         (src_ptr[4] * HFilter[4]) +
+                         (src_ptr[5] * HFilter[5]) +
+                         (src_ptr[6] * HFilter[6]) +
+                         (src_ptr[7] * HFilter[7]) +
+                         (VP9_FILTER_WEIGHT >> 1);  // Rounding
+
+        // Normalize back to 0-255...
+        *output_ptr = clip_pixel_highbd(temp >> VP9_FILTER_SHIFT, bd);
+        ++src_ptr;
+        output_ptr += intermediate_height;
+      }
+      src_ptr += src_next_row_stride;
+      output_ptr += intermediate_next_stride;
+    }
+  }
+
+  // Vertical pass (transposed intermediate -> dst).
+  {
+    uint16_t *src_ptr = intermediate_buffer;
+    const int dst_next_row_stride = dst_stride - output_width;
+    unsigned int i, j;
+    for (i = 0; i < output_height; ++i) {
+      for (j = 0; j < output_width; ++j) {
+        // Apply filter...
+        const int temp = (src_ptr[0] * VFilter[0]) +
+                         (src_ptr[1] * VFilter[1]) +
+                         (src_ptr[2] * VFilter[2]) +
+                         (src_ptr[3] * VFilter[3]) +
+                         (src_ptr[4] * VFilter[4]) +
+                         (src_ptr[5] * VFilter[5]) +
+                         (src_ptr[6] * VFilter[6]) +
+                         (src_ptr[7] * VFilter[7]) +
+                         (VP9_FILTER_WEIGHT >> 1);  // Rounding
+
+        // Normalize back to 0-255...
+        *dst_ptr++ = clip_pixel_highbd(temp >> VP9_FILTER_SHIFT, bd);
+        src_ptr += intermediate_height;
+      }
+      src_ptr += intermediate_next_stride;
+      dst_ptr += dst_next_row_stride;
+    }
+  }
+}
+
+void highbd_block2d_average_c(uint16_t *src,
+                              unsigned int src_stride,
+                              uint16_t *output_ptr,
+                              unsigned int output_stride,
+                              unsigned int output_width,
+                              unsigned int output_height) {
+  unsigned int i, j;
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      output_ptr[j] = (output_ptr[j] + src[i * src_stride + j] + 1) >> 1;
+    }
+    output_ptr += output_stride;
+  }
+}
+
+void highbd_filter_average_block2d_8_c(const uint16_t *src_ptr,
+                                       const unsigned int src_stride,
+                                       const int16_t *HFilter,
+                                       const int16_t *VFilter,
+                                       uint16_t *dst_ptr,
+                                       unsigned int dst_stride,
+                                       unsigned int output_width,
+                                       unsigned int output_height,
+                                       int bd) {
+  uint16_t tmp[kMaxDimension * kMaxDimension];
+
+  assert(output_width <= kMaxDimension);
+  assert(output_height <= kMaxDimension);
+  highbd_filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, 64,
+                            output_width, output_height, bd);
+  highbd_block2d_average_c(tmp, 64, dst_ptr, dst_stride,
+                           output_width, output_height);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
+ public:
+  static void SetUpTestCase() {
+    // Force input_ to be unaligned, output to be 16 byte aligned.
+    input_ = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kInputBufferSize + 1)) + 1;
+    output_ = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kOutputBufferSize));
+    output_ref_ = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kOutputBufferSize));
+#if CONFIG_VP9_HIGHBITDEPTH
+    input16_ = reinterpret_cast<uint16_t*>(
+        vpx_memalign(kDataAlignment,
+                     (kInputBufferSize + 1) * sizeof(uint16_t))) + 1;
+    output16_ = reinterpret_cast<uint16_t*>(
+        vpx_memalign(kDataAlignment, (kOutputBufferSize) * sizeof(uint16_t)));
+    output16_ref_ = reinterpret_cast<uint16_t*>(
+        vpx_memalign(kDataAlignment, (kOutputBufferSize) * sizeof(uint16_t)));
+#endif
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+  static void TearDownTestCase() {
+    vpx_free(input_ - 1);
+    input_ = NULL;
+    vpx_free(output_);
+    output_ = NULL;
+    vpx_free(output_ref_);
+    output_ref_ = NULL;
+#if CONFIG_VP9_HIGHBITDEPTH
+    vpx_free(input16_ - 1);
+    input16_ = NULL;
+    vpx_free(output16_);
+    output16_ = NULL;
+    vpx_free(output16_ref_);
+    output16_ref_ = NULL;
+#endif
+  }
+
+ protected:
+  static const int kDataAlignment = 16;
+  static const int kOuterBlockSize = 256;
+  static const int kInputStride = kOuterBlockSize;
+  static const int kOutputStride = kOuterBlockSize;
+  static const int kInputBufferSize = kOuterBlockSize * kOuterBlockSize;
+  static const int kOutputBufferSize = kOuterBlockSize * kOuterBlockSize;
+
+  int Width() const { return GET_PARAM(0); }
+  int Height() const { return GET_PARAM(1); }
+  int BorderLeft() const {
+    const int center = (kOuterBlockSize - Width()) / 2;
+    return (center + (kDataAlignment - 1)) & ~(kDataAlignment - 1);
+  }
+  int BorderTop() const { return (kOuterBlockSize - Height()) / 2; }
+
+  bool IsIndexInBorder(int i) {
+    return (i < BorderTop() * kOuterBlockSize ||
+            i >= (BorderTop() + Height()) * kOuterBlockSize ||
+            i % kOuterBlockSize < BorderLeft() ||
+            i % kOuterBlockSize >= (BorderLeft() + Width()));
+  }
+
+  virtual void SetUp() {
+    UUT_ = GET_PARAM(2);
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (UUT_->use_highbd_ != 0)
+      mask_ = (1 << UUT_->use_highbd_) - 1;
+    else
+      mask_ = 255;
+#endif
+    /* Set up guard blocks for an inner block centered in the outer block */
+    for (int i = 0; i < kOutputBufferSize; ++i) {
+      if (IsIndexInBorder(i))
+        output_[i] = 255;
+      else
+        output_[i] = 0;
+    }
+
+    ::libvpx_test::ACMRandom prng;
+    for (int i = 0; i < kInputBufferSize; ++i) {
+      if (i & 1) {
+        input_[i] = 255;
+#if CONFIG_VP9_HIGHBITDEPTH
+        input16_[i] = mask_;
+#endif
+      } else {
+        input_[i] = prng.Rand8Extremes();
+#if CONFIG_VP9_HIGHBITDEPTH
+        input16_[i] = prng.Rand16() & mask_;
+#endif
+      }
+    }
+  }
+
+  void SetConstantInput(int value) {
+    memset(input_, value, kInputBufferSize);
+#if CONFIG_VP9_HIGHBITDEPTH
+    vpx_memset16(input16_, value, kInputBufferSize);
+#endif
+  }
+
+  void CopyOutputToRef() {
+    memcpy(output_ref_, output_, kOutputBufferSize);
+#if CONFIG_VP9_HIGHBITDEPTH
+    memcpy(output16_ref_, output16_, kOutputBufferSize);
+#endif
+  }
+
+  void CheckGuardBlocks() {
+    for (int i = 0; i < kOutputBufferSize; ++i) {
+      if (IsIndexInBorder(i))
+        EXPECT_EQ(255, output_[i]);
+    }
+  }
+
+  uint8_t *input() const {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      return input_ + BorderTop() * kOuterBlockSize + BorderLeft();
+    } else {
+      return CONVERT_TO_BYTEPTR(input16_ + BorderTop() * kOuterBlockSize +
+                                BorderLeft());
+    }
+#else
+    return input_ + BorderTop() * kOuterBlockSize + BorderLeft();
+#endif
+  }
+
+  uint8_t *output() const {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      return output_ + BorderTop() * kOuterBlockSize + BorderLeft();
+    } else {
+      return CONVERT_TO_BYTEPTR(output16_ + BorderTop() * kOuterBlockSize +
+                                BorderLeft());
+    }
+#else
+    return output_ + BorderTop() * kOuterBlockSize + BorderLeft();
+#endif
+  }
+
+  uint8_t *output_ref() const {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      return output_ref_ + BorderTop() * kOuterBlockSize + BorderLeft();
+    } else {
+      return CONVERT_TO_BYTEPTR(output16_ref_ + BorderTop() * kOuterBlockSize +
+                                BorderLeft());
+    }
+#else
+    return output_ref_ + BorderTop() * kOuterBlockSize + BorderLeft();
+#endif
+  }
+
+  uint16_t lookup(uint8_t *list, int index) const {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      return list[index];
+    } else {
+      return CONVERT_TO_SHORTPTR(list)[index];
+    }
+#else
+    return list[index];
+#endif
+  }
+
+  void assign_val(uint8_t *list, int index, uint16_t val) const {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      list[index] = (uint8_t) val;
+    } else {
+      CONVERT_TO_SHORTPTR(list)[index] = val;
+    }
+#else
+    list[index] = (uint8_t) val;
+#endif
+  }
+
+  void wrapper_filter_average_block2d_8_c(const uint8_t *src_ptr,
+                                          const unsigned int src_stride,
+                                          const int16_t *HFilter,
+                                          const int16_t *VFilter,
+                                          uint8_t *dst_ptr,
+                                          unsigned int dst_stride,
+                                          unsigned int output_width,
+                                          unsigned int output_height) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      filter_average_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
+                                 dst_ptr, dst_stride, output_width,
+                                 output_height);
+    } else {
+      highbd_filter_average_block2d_8_c(CONVERT_TO_SHORTPTR(src_ptr),
+                                        src_stride, HFilter, VFilter,
+                                        CONVERT_TO_SHORTPTR(dst_ptr),
+                                        dst_stride, output_width, output_height,
+                                        UUT_->use_highbd_);
+    }
+#else
+    filter_average_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
+                               dst_ptr, dst_stride, output_width,
+                               output_height);
+#endif
+  }
+
+  void wrapper_filter_block2d_8_c(const uint8_t *src_ptr,
+                                  const unsigned int src_stride,
+                                  const int16_t *HFilter,
+                                  const int16_t *VFilter,
+                                  uint8_t *dst_ptr,
+                                  unsigned int dst_stride,
+                                  unsigned int output_width,
+                                  unsigned int output_height) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
+                         dst_ptr, dst_stride, output_width, output_height);
+    } else {
+      highbd_filter_block2d_8_c(CONVERT_TO_SHORTPTR(src_ptr), src_stride,
+                                HFilter, VFilter,
+                                CONVERT_TO_SHORTPTR(dst_ptr), dst_stride,
+                                output_width, output_height, UUT_->use_highbd_);
+    }
+#else
+    filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
+                       dst_ptr, dst_stride, output_width, output_height);
+#endif
+  }
+
+  const ConvolveFunctions* UUT_;
+  static uint8_t* input_;
+  static uint8_t* output_;
+  static uint8_t* output_ref_;
+#if CONFIG_VP9_HIGHBITDEPTH
+  static uint16_t* input16_;
+  static uint16_t* output16_;
+  static uint16_t* output16_ref_;
+  int mask_;
+#endif
+};
+
+uint8_t* ConvolveTest::input_ = NULL;
+uint8_t* ConvolveTest::output_ = NULL;
+uint8_t* ConvolveTest::output_ref_ = NULL;
+#if CONFIG_VP9_HIGHBITDEPTH
+uint16_t* ConvolveTest::input16_ = NULL;
+uint16_t* ConvolveTest::output16_ = NULL;
+uint16_t* ConvolveTest::output16_ref_ = NULL;
+#endif
+
+TEST_P(ConvolveTest, GuardBlocks) {
+  CheckGuardBlocks();
+}
+
+TEST_P(ConvolveTest, Copy) {
+  uint8_t* const in = input();
+  uint8_t* const out = output();
+
+  ASM_REGISTER_STATE_CHECK(
+      UUT_->copy_(in, kInputStride, out, kOutputStride, NULL, 0, NULL, 0,
+                  Width(), Height()));
+
+  CheckGuardBlocks();
+
+  for (int y = 0; y < Height(); ++y)
+    for (int x = 0; x < Width(); ++x)
+      ASSERT_EQ(lookup(out, y * kOutputStride + x),
+                lookup(in, y * kInputStride + x))
+          << "(" << x << "," << y << ")";
+}
+
+TEST_P(ConvolveTest, Avg) {
+  uint8_t* const in = input();
+  uint8_t* const out = output();
+  uint8_t* const out_ref = output_ref();
+  CopyOutputToRef();
+
+  ASM_REGISTER_STATE_CHECK(
+      UUT_->avg_(in, kInputStride, out, kOutputStride, NULL, 0, NULL, 0,
+                Width(), Height()));
+
+  CheckGuardBlocks();
+
+  for (int y = 0; y < Height(); ++y)
+    for (int x = 0; x < Width(); ++x)
+      ASSERT_EQ(lookup(out, y * kOutputStride + x),
+                ROUND_POWER_OF_TWO(lookup(in, y * kInputStride + x) +
+                                   lookup(out_ref, y * kOutputStride + x), 1))
+          << "(" << x << "," << y << ")";
+}
+
+TEST_P(ConvolveTest, CopyHoriz) {
+  uint8_t* const in = input();
+  uint8_t* const out = output();
+  DECLARE_ALIGNED(256, const int16_t, filter8[8]) = {0, 0, 0, 128, 0, 0, 0, 0};
+
+  ASM_REGISTER_STATE_CHECK(
+      UUT_->sh8_(in, kInputStride, out, kOutputStride, filter8, 16, filter8, 16,
+                 Width(), Height()));
+
+  CheckGuardBlocks();
+
+  for (int y = 0; y < Height(); ++y)
+    for (int x = 0; x < Width(); ++x)
+      ASSERT_EQ(lookup(out, y * kOutputStride + x),
+                lookup(in, y * kInputStride + x))
+          << "(" << x << "," << y << ")";
+}
+
+TEST_P(ConvolveTest, CopyVert) {
+  uint8_t* const in = input();
+  uint8_t* const out = output();
+  DECLARE_ALIGNED(256, const int16_t, filter8[8]) = {0, 0, 0, 128, 0, 0, 0, 0};
+
+  ASM_REGISTER_STATE_CHECK(
+      UUT_->sv8_(in, kInputStride, out, kOutputStride, filter8, 16, filter8, 16,
+                 Width(), Height()));
+
+  CheckGuardBlocks();
+
+  for (int y = 0; y < Height(); ++y)
+    for (int x = 0; x < Width(); ++x)
+      ASSERT_EQ(lookup(out, y * kOutputStride + x),
+                lookup(in, y * kInputStride + x))
+          << "(" << x << "," << y << ")";
+}
+
+TEST_P(ConvolveTest, Copy2D) {
+  uint8_t* const in = input();
+  uint8_t* const out = output();
+  DECLARE_ALIGNED(256, const int16_t, filter8[8]) = {0, 0, 0, 128, 0, 0, 0, 0};
+
+  ASM_REGISTER_STATE_CHECK(
+      UUT_->shv8_(in, kInputStride, out, kOutputStride, filter8, 16, filter8,
+                  16, Width(), Height()));
+
+  CheckGuardBlocks();
+
+  for (int y = 0; y < Height(); ++y)
+    for (int x = 0; x < Width(); ++x)
+      ASSERT_EQ(lookup(out, y * kOutputStride + x),
+                lookup(in, y * kInputStride + x))
+          << "(" << x << "," << y << ")";
+}
+
+const int kNumFilterBanks = 4;
+const int kNumFilters = 16;
+
+TEST(ConvolveTest, FiltersWontSaturateWhenAddedPairwise) {
+  for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
+    const InterpKernel *filters =
+        vp9_filter_kernels[static_cast<INTERP_FILTER>(filter_bank)];
+    for (int i = 0; i < kNumFilters; i++) {
+      const int p0 = filters[i][0] + filters[i][1];
+      const int p1 = filters[i][2] + filters[i][3];
+      const int p2 = filters[i][4] + filters[i][5];
+      const int p3 = filters[i][6] + filters[i][7];
+      EXPECT_LE(p0, 128);
+      EXPECT_LE(p1, 128);
+      EXPECT_LE(p2, 128);
+      EXPECT_LE(p3, 128);
+      EXPECT_LE(p0 + p3, 128);
+      EXPECT_LE(p0 + p3 + p1, 128);
+      EXPECT_LE(p0 + p3 + p1 + p2, 128);
+      EXPECT_EQ(p0 + p1 + p2 + p3, 128);
+    }
+  }
+}
+
+const int16_t kInvalidFilter[8] = { 0 };
+
+TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) {
+  uint8_t* const in = input();
+  uint8_t* const out = output();
+#if CONFIG_VP9_HIGHBITDEPTH
+  uint8_t ref8[kOutputStride * kMaxDimension];
+  uint16_t ref16[kOutputStride * kMaxDimension];
+  uint8_t* ref;
+  if (UUT_->use_highbd_ == 0) {
+    ref = ref8;
+  } else {
+    ref = CONVERT_TO_BYTEPTR(ref16);
+  }
+#else
+  uint8_t ref[kOutputStride * kMaxDimension];
+#endif
+
+  for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
+    const InterpKernel *filters =
+        vp9_filter_kernels[static_cast<INTERP_FILTER>(filter_bank)];
+
+    for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
+      for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
+        wrapper_filter_block2d_8_c(in, kInputStride,
+                                   filters[filter_x], filters[filter_y],
+                                   ref, kOutputStride,
+                                   Width(), Height());
+
+        if (filter_x && filter_y)
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->hv8_(in, kInputStride, out, kOutputStride,
+                         filters[filter_x], 16, filters[filter_y], 16,
+                         Width(), Height()));
+        else if (filter_y)
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->v8_(in, kInputStride, out, kOutputStride,
+                        kInvalidFilter, 16, filters[filter_y], 16,
+                        Width(), Height()));
+        else if (filter_x)
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->h8_(in, kInputStride, out, kOutputStride,
+                        filters[filter_x], 16, kInvalidFilter, 16,
+                        Width(), Height()));
+        else
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->copy_(in, kInputStride, out, kOutputStride,
+                          kInvalidFilter, 0, kInvalidFilter, 0,
+                          Width(), Height()));
+
+        CheckGuardBlocks();
+
+        for (int y = 0; y < Height(); ++y)
+          for (int x = 0; x < Width(); ++x)
+            ASSERT_EQ(lookup(ref, y * kOutputStride + x),
+                      lookup(out, y * kOutputStride + x))
+                << "mismatch at (" << x << "," << y << "), "
+                << "filters (" << filter_bank << ","
+                << filter_x << "," << filter_y << ")";
+      }
+    }
+  }
+}
+
+TEST_P(ConvolveTest, MatchesReferenceAveragingSubpixelFilter) {
+  uint8_t* const in = input();
+  uint8_t* const out = output();
+#if CONFIG_VP9_HIGHBITDEPTH
+  uint8_t ref8[kOutputStride * kMaxDimension];
+  uint16_t ref16[kOutputStride * kMaxDimension];
+  uint8_t* ref;
+  if (UUT_->use_highbd_ == 0) {
+    ref = ref8;
+  } else {
+    ref = CONVERT_TO_BYTEPTR(ref16);
+  }
+#else
+  uint8_t ref[kOutputStride * kMaxDimension];
+#endif
+
+  // Populate ref and out with some random data
+  ::libvpx_test::ACMRandom prng;
+  for (int y = 0; y < Height(); ++y) {
+    for (int x = 0; x < Width(); ++x) {
+      uint16_t r;
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (UUT_->use_highbd_ == 0 || UUT_->use_highbd_ == 8) {
+        r = prng.Rand8Extremes();
+      } else {
+        r = prng.Rand16() & mask_;
+      }
+#else
+      r = prng.Rand8Extremes();
+#endif
+
+      assign_val(out, y * kOutputStride + x, r);
+      assign_val(ref, y * kOutputStride + x, r);
+    }
+  }
+
+  for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
+    const InterpKernel *filters =
+        vp9_filter_kernels[static_cast<INTERP_FILTER>(filter_bank)];
+
+    for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
+      for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
+        wrapper_filter_average_block2d_8_c(in, kInputStride,
+                                           filters[filter_x], filters[filter_y],
+                                           ref, kOutputStride,
+                                           Width(), Height());
+
+        if (filter_x && filter_y)
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->hv8_avg_(in, kInputStride, out, kOutputStride,
+                             filters[filter_x], 16, filters[filter_y], 16,
+                             Width(), Height()));
+        else if (filter_y)
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->v8_avg_(in, kInputStride, out, kOutputStride,
+                            kInvalidFilter, 16, filters[filter_y], 16,
+                            Width(), Height()));
+        else if (filter_x)
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->h8_avg_(in, kInputStride, out, kOutputStride,
+                            filters[filter_x], 16, kInvalidFilter, 16,
+                            Width(), Height()));
+        else
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->avg_(in, kInputStride, out, kOutputStride,
+                          kInvalidFilter, 0, kInvalidFilter, 0,
+                          Width(), Height()));
+
+        CheckGuardBlocks();
+
+        for (int y = 0; y < Height(); ++y)
+          for (int x = 0; x < Width(); ++x)
+            ASSERT_EQ(lookup(ref, y * kOutputStride + x),
+                      lookup(out, y * kOutputStride + x))
+                << "mismatch at (" << x << "," << y << "), "
+                << "filters (" << filter_bank << ","
+                << filter_x << "," << filter_y << ")";
+      }
+    }
+  }
+}
+
+TEST_P(ConvolveTest, FilterExtremes) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+#if CONFIG_VP9_HIGHBITDEPTH
+  uint8_t ref8[kOutputStride * kMaxDimension];
+  uint16_t ref16[kOutputStride * kMaxDimension];
+  uint8_t *ref;
+  if (UUT_->use_highbd_ == 0) {
+    ref = ref8;
+  } else {
+    ref = CONVERT_TO_BYTEPTR(ref16);
+  }
+#else
+  uint8_t ref[kOutputStride * kMaxDimension];
+#endif
+
+  // Populate ref and out with some random data
+  ::libvpx_test::ACMRandom prng;
+  for (int y = 0; y < Height(); ++y) {
+    for (int x = 0; x < Width(); ++x) {
+      uint16_t r;
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (UUT_->use_highbd_ == 0 || UUT_->use_highbd_ == 8) {
+        r = prng.Rand8Extremes();
+      } else {
+        r = prng.Rand16() & mask_;
+      }
+#else
+      r = prng.Rand8Extremes();
+#endif
+      assign_val(out, y * kOutputStride + x, r);
+      assign_val(ref, y * kOutputStride + x, r);
+    }
+  }
+
+  for (int axis = 0; axis < 2; axis++) {
+    int seed_val = 0;
+    while (seed_val < 256) {
+      for (int y = 0; y < 8; ++y) {
+        for (int x = 0; x < 8; ++x) {
+#if CONFIG_VP9_HIGHBITDEPTH
+            assign_val(in, y * kOutputStride + x - SUBPEL_TAPS / 2 + 1,
+                       ((seed_val >> (axis ? y : x)) & 1) * mask_);
+#else
+            assign_val(in, y * kOutputStride + x - SUBPEL_TAPS / 2 + 1,
+                       ((seed_val >> (axis ? y : x)) & 1) * 255);
+#endif
+          if (axis) seed_val++;
+        }
+        if (axis)
+          seed_val-= 8;
+        else
+          seed_val++;
+      }
+      if (axis) seed_val += 8;
+
+      for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
+        const InterpKernel *filters =
+            vp9_filter_kernels[static_cast<INTERP_FILTER>(filter_bank)];
+        for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
+          for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
+            wrapper_filter_block2d_8_c(in, kInputStride,
+                                       filters[filter_x], filters[filter_y],
+                                       ref, kOutputStride,
+                                       Width(), Height());
+            if (filter_x && filter_y)
+              ASM_REGISTER_STATE_CHECK(
+                  UUT_->hv8_(in, kInputStride, out, kOutputStride,
+                             filters[filter_x], 16, filters[filter_y], 16,
+                             Width(), Height()));
+            else if (filter_y)
+              ASM_REGISTER_STATE_CHECK(
+                  UUT_->v8_(in, kInputStride, out, kOutputStride,
+                            kInvalidFilter, 16, filters[filter_y], 16,
+                            Width(), Height()));
+            else if (filter_x)
+              ASM_REGISTER_STATE_CHECK(
+                  UUT_->h8_(in, kInputStride, out, kOutputStride,
+                            filters[filter_x], 16, kInvalidFilter, 16,
+                            Width(), Height()));
+            else
+              ASM_REGISTER_STATE_CHECK(
+                  UUT_->copy_(in, kInputStride, out, kOutputStride,
+                              kInvalidFilter, 0, kInvalidFilter, 0,
+                              Width(), Height()));
+
+            for (int y = 0; y < Height(); ++y)
+              for (int x = 0; x < Width(); ++x)
+                ASSERT_EQ(lookup(ref, y * kOutputStride + x),
+                          lookup(out, y * kOutputStride + x))
+                    << "mismatch at (" << x << "," << y << "), "
+                    << "filters (" << filter_bank << ","
+                    << filter_x << "," << filter_y << ")";
+          }
+        }
+      }
+    }
+  }
+}
+
+/* This test exercises that enough rows and columns are filtered with every
+   possible initial fractional positions and scaling steps. */
+TEST_P(ConvolveTest, CheckScalingFiltering) {
+  uint8_t* const in = input();
+  uint8_t* const out = output();
+  const InterpKernel *const eighttap = vp9_filter_kernels[EIGHTTAP];
+
+  SetConstantInput(127);
+
+  for (int frac = 0; frac < 16; ++frac) {
+    for (int step = 1; step <= 32; ++step) {
+      /* Test the horizontal and vertical filters in combination. */
+      ASM_REGISTER_STATE_CHECK(UUT_->shv8_(in, kInputStride, out, kOutputStride,
+                                           eighttap[frac], step,
+                                           eighttap[frac], step,
+                                           Width(), Height()));
+
+      CheckGuardBlocks();
+
+      for (int y = 0; y < Height(); ++y) {
+        for (int x = 0; x < Width(); ++x) {
+          ASSERT_EQ(lookup(in, y * kInputStride + x),
+                    lookup(out, y * kOutputStride + x))
+              << "x == " << x << ", y == " << y
+              << ", frac == " << frac << ", step == " << step;
+        }
+      }
+    }
+  }
+}
+
+using std::tr1::make_tuple;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+#define WRAP(func, bd) \
+void wrap_ ## func ## _ ## bd(const uint8_t *src, ptrdiff_t src_stride, \
+                              uint8_t *dst, ptrdiff_t dst_stride, \
+                              const int16_t *filter_x, \
+                              int filter_x_stride, \
+                              const int16_t *filter_y, \
+                              int filter_y_stride, \
+                              int w, int h) { \
+  vpx_highbd_ ## func(src, src_stride, dst, dst_stride, filter_x, \
+                      filter_x_stride, filter_y, filter_y_stride, \
+                      w, h, bd); \
+}
+#if HAVE_SSE2 && ARCH_X86_64
+#if CONFIG_USE_X86INC
+WRAP(convolve_copy_sse2, 8)
+WRAP(convolve_avg_sse2, 8)
+WRAP(convolve_copy_sse2, 10)
+WRAP(convolve_avg_sse2, 10)
+WRAP(convolve_copy_sse2, 12)
+WRAP(convolve_avg_sse2, 12)
+#endif  // CONFIG_USE_X86INC
+WRAP(convolve8_horiz_sse2, 8)
+WRAP(convolve8_avg_horiz_sse2, 8)
+WRAP(convolve8_vert_sse2, 8)
+WRAP(convolve8_avg_vert_sse2, 8)
+WRAP(convolve8_sse2, 8)
+WRAP(convolve8_avg_sse2, 8)
+WRAP(convolve8_horiz_sse2, 10)
+WRAP(convolve8_avg_horiz_sse2, 10)
+WRAP(convolve8_vert_sse2, 10)
+WRAP(convolve8_avg_vert_sse2, 10)
+WRAP(convolve8_sse2, 10)
+WRAP(convolve8_avg_sse2, 10)
+WRAP(convolve8_horiz_sse2, 12)
+WRAP(convolve8_avg_horiz_sse2, 12)
+WRAP(convolve8_vert_sse2, 12)
+WRAP(convolve8_avg_vert_sse2, 12)
+WRAP(convolve8_sse2, 12)
+WRAP(convolve8_avg_sse2, 12)
+#endif  // HAVE_SSE2 && ARCH_X86_64
+
+WRAP(convolve_copy_c, 8)
+WRAP(convolve_avg_c, 8)
+WRAP(convolve8_horiz_c, 8)
+WRAP(convolve8_avg_horiz_c, 8)
+WRAP(convolve8_vert_c, 8)
+WRAP(convolve8_avg_vert_c, 8)
+WRAP(convolve8_c, 8)
+WRAP(convolve8_avg_c, 8)
+WRAP(convolve_copy_c, 10)
+WRAP(convolve_avg_c, 10)
+WRAP(convolve8_horiz_c, 10)
+WRAP(convolve8_avg_horiz_c, 10)
+WRAP(convolve8_vert_c, 10)
+WRAP(convolve8_avg_vert_c, 10)
+WRAP(convolve8_c, 10)
+WRAP(convolve8_avg_c, 10)
+WRAP(convolve_copy_c, 12)
+WRAP(convolve_avg_c, 12)
+WRAP(convolve8_horiz_c, 12)
+WRAP(convolve8_avg_horiz_c, 12)
+WRAP(convolve8_vert_c, 12)
+WRAP(convolve8_avg_vert_c, 12)
+WRAP(convolve8_c, 12)
+WRAP(convolve8_avg_c, 12)
+#undef WRAP
+
+const ConvolveFunctions convolve8_c(
+    wrap_convolve_copy_c_8, wrap_convolve_avg_c_8,
+    wrap_convolve8_horiz_c_8, wrap_convolve8_avg_horiz_c_8,
+    wrap_convolve8_vert_c_8, wrap_convolve8_avg_vert_c_8,
+    wrap_convolve8_c_8, wrap_convolve8_avg_c_8,
+    wrap_convolve8_horiz_c_8, wrap_convolve8_avg_horiz_c_8,
+    wrap_convolve8_vert_c_8, wrap_convolve8_avg_vert_c_8,
+    wrap_convolve8_c_8, wrap_convolve8_avg_c_8, 8);
+const ConvolveFunctions convolve10_c(
+    wrap_convolve_copy_c_10, wrap_convolve_avg_c_10,
+    wrap_convolve8_horiz_c_10, wrap_convolve8_avg_horiz_c_10,
+    wrap_convolve8_vert_c_10, wrap_convolve8_avg_vert_c_10,
+    wrap_convolve8_c_10, wrap_convolve8_avg_c_10,
+    wrap_convolve8_horiz_c_10, wrap_convolve8_avg_horiz_c_10,
+    wrap_convolve8_vert_c_10, wrap_convolve8_avg_vert_c_10,
+    wrap_convolve8_c_10, wrap_convolve8_avg_c_10, 10);
+const ConvolveFunctions convolve12_c(
+    wrap_convolve_copy_c_12, wrap_convolve_avg_c_12,
+    wrap_convolve8_horiz_c_12, wrap_convolve8_avg_horiz_c_12,
+    wrap_convolve8_vert_c_12, wrap_convolve8_avg_vert_c_12,
+    wrap_convolve8_c_12, wrap_convolve8_avg_c_12,
+    wrap_convolve8_horiz_c_12, wrap_convolve8_avg_horiz_c_12,
+    wrap_convolve8_vert_c_12, wrap_convolve8_avg_vert_c_12,
+    wrap_convolve8_c_12, wrap_convolve8_avg_c_12, 12);
+const ConvolveParam kArrayConvolve_c[] = {
+    ALL_SIZES(convolve8_c),
+    ALL_SIZES(convolve10_c),
+    ALL_SIZES(convolve12_c)
+};
+
+#else
+const ConvolveFunctions convolve8_c(
+    vpx_convolve_copy_c, vpx_convolve_avg_c,
+    vpx_convolve8_horiz_c, vpx_convolve8_avg_horiz_c,
+    vpx_convolve8_vert_c, vpx_convolve8_avg_vert_c,
+    vpx_convolve8_c, vpx_convolve8_avg_c,
+    vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c,
+    vpx_scaled_vert_c, vpx_scaled_avg_vert_c,
+    vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
+const ConvolveParam kArrayConvolve_c[] = { ALL_SIZES(convolve8_c) };
+#endif
+INSTANTIATE_TEST_CASE_P(C, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve_c));
+
+#if HAVE_SSE2 && ARCH_X86_64
+#if CONFIG_VP9_HIGHBITDEPTH
+const ConvolveFunctions convolve8_sse2(
+#if CONFIG_USE_X86INC
+    wrap_convolve_copy_sse2_8, wrap_convolve_avg_sse2_8,
+#else
+    wrap_convolve_copy_c_8, wrap_convolve_avg_c_8,
+#endif  // CONFIG_USE_X86INC
+    wrap_convolve8_horiz_sse2_8, wrap_convolve8_avg_horiz_sse2_8,
+    wrap_convolve8_vert_sse2_8, wrap_convolve8_avg_vert_sse2_8,
+    wrap_convolve8_sse2_8, wrap_convolve8_avg_sse2_8,
+    wrap_convolve8_horiz_sse2_8, wrap_convolve8_avg_horiz_sse2_8,
+    wrap_convolve8_vert_sse2_8, wrap_convolve8_avg_vert_sse2_8,
+    wrap_convolve8_sse2_8, wrap_convolve8_avg_sse2_8, 8);
+const ConvolveFunctions convolve10_sse2(
+#if CONFIG_USE_X86INC
+    wrap_convolve_copy_sse2_10, wrap_convolve_avg_sse2_10,
+#else
+    wrap_convolve_copy_c_10, wrap_convolve_avg_c_10,
+#endif  // CONFIG_USE_X86INC
+    wrap_convolve8_horiz_sse2_10, wrap_convolve8_avg_horiz_sse2_10,
+    wrap_convolve8_vert_sse2_10, wrap_convolve8_avg_vert_sse2_10,
+    wrap_convolve8_sse2_10, wrap_convolve8_avg_sse2_10,
+    wrap_convolve8_horiz_sse2_10, wrap_convolve8_avg_horiz_sse2_10,
+    wrap_convolve8_vert_sse2_10, wrap_convolve8_avg_vert_sse2_10,
+    wrap_convolve8_sse2_10, wrap_convolve8_avg_sse2_10, 10);
+const ConvolveFunctions convolve12_sse2(
+#if CONFIG_USE_X86INC
+    wrap_convolve_copy_sse2_12, wrap_convolve_avg_sse2_12,
+#else
+    wrap_convolve_copy_c_12, wrap_convolve_avg_c_12,
+#endif  // CONFIG_USE_X86INC
+    wrap_convolve8_horiz_sse2_12, wrap_convolve8_avg_horiz_sse2_12,
+    wrap_convolve8_vert_sse2_12, wrap_convolve8_avg_vert_sse2_12,
+    wrap_convolve8_sse2_12, wrap_convolve8_avg_sse2_12,
+    wrap_convolve8_horiz_sse2_12, wrap_convolve8_avg_horiz_sse2_12,
+    wrap_convolve8_vert_sse2_12, wrap_convolve8_avg_vert_sse2_12,
+    wrap_convolve8_sse2_12, wrap_convolve8_avg_sse2_12, 12);
+const ConvolveParam kArrayConvolve_sse2[] = {
+    ALL_SIZES(convolve8_sse2),
+    ALL_SIZES(convolve10_sse2),
+    ALL_SIZES(convolve12_sse2)
+};
+#else
+const ConvolveFunctions convolve8_sse2(
+#if CONFIG_USE_X86INC
+    vpx_convolve_copy_sse2, vpx_convolve_avg_sse2,
+#else
+    vpx_convolve_copy_c, vpx_convolve_avg_c,
+#endif  // CONFIG_USE_X86INC
+    vpx_convolve8_horiz_sse2, vpx_convolve8_avg_horiz_sse2,
+    vpx_convolve8_vert_sse2, vpx_convolve8_avg_vert_sse2,
+    vpx_convolve8_sse2, vpx_convolve8_avg_sse2,
+    vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c,
+    vpx_scaled_vert_c, vpx_scaled_avg_vert_c,
+    vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
+
+const ConvolveParam kArrayConvolve_sse2[] = { ALL_SIZES(convolve8_sse2) };
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(SSE2, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve_sse2));
+#endif
+
+#if HAVE_SSSE3
+const ConvolveFunctions convolve8_ssse3(
+    vpx_convolve_copy_c, vpx_convolve_avg_c,
+    vpx_convolve8_horiz_ssse3, vpx_convolve8_avg_horiz_ssse3,
+    vpx_convolve8_vert_ssse3, vpx_convolve8_avg_vert_ssse3,
+    vpx_convolve8_ssse3, vpx_convolve8_avg_ssse3,
+    vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c,
+    vpx_scaled_vert_c, vpx_scaled_avg_vert_c,
+    vpx_scaled_2d_ssse3, vpx_scaled_avg_2d_c, 0);
+
+const ConvolveParam kArrayConvolve8_ssse3[] = { ALL_SIZES(convolve8_ssse3) };
+INSTANTIATE_TEST_CASE_P(SSSE3, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve8_ssse3));
+#endif
+
+#if HAVE_AVX2 && HAVE_SSSE3
+const ConvolveFunctions convolve8_avx2(
+    vpx_convolve_copy_c, vpx_convolve_avg_c,
+    vpx_convolve8_horiz_avx2, vpx_convolve8_avg_horiz_ssse3,
+    vpx_convolve8_vert_avx2, vpx_convolve8_avg_vert_ssse3,
+    vpx_convolve8_avx2, vpx_convolve8_avg_ssse3,
+    vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c,
+    vpx_scaled_vert_c, vpx_scaled_avg_vert_c,
+    vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
+
+const ConvolveParam kArrayConvolve8_avx2[] = { ALL_SIZES(convolve8_avx2) };
+INSTANTIATE_TEST_CASE_P(AVX2, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve8_avx2));
+#endif  // HAVE_AVX2 && HAVE_SSSE3
+
+#if HAVE_NEON
+#if HAVE_NEON_ASM
+const ConvolveFunctions convolve8_neon(
+    vpx_convolve_copy_neon, vpx_convolve_avg_neon,
+    vpx_convolve8_horiz_neon, vpx_convolve8_avg_horiz_neon,
+    vpx_convolve8_vert_neon, vpx_convolve8_avg_vert_neon,
+    vpx_convolve8_neon, vpx_convolve8_avg_neon,
+    vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c,
+    vpx_scaled_vert_c, vpx_scaled_avg_vert_c,
+    vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
+#else  // HAVE_NEON
+const ConvolveFunctions convolve8_neon(
+    vpx_convolve_copy_neon, vpx_convolve_avg_neon,
+    vpx_convolve8_horiz_neon, vpx_convolve8_avg_horiz_neon,
+    vpx_convolve8_vert_neon, vpx_convolve8_avg_vert_neon,
+    vpx_convolve8_neon, vpx_convolve8_avg_neon,
+    vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c,
+    vpx_scaled_vert_c, vpx_scaled_avg_vert_c,
+    vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
+#endif  // HAVE_NEON_ASM
+
+const ConvolveParam kArrayConvolve8_neon[] = { ALL_SIZES(convolve8_neon) };
+INSTANTIATE_TEST_CASE_P(NEON, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve8_neon));
+#endif  // HAVE_NEON
+
+#if HAVE_DSPR2
+const ConvolveFunctions convolve8_dspr2(
+    vpx_convolve_copy_dspr2, vpx_convolve_avg_dspr2,
+    vpx_convolve8_horiz_dspr2, vpx_convolve8_avg_horiz_dspr2,
+    vpx_convolve8_vert_dspr2, vpx_convolve8_avg_vert_dspr2,
+    vpx_convolve8_dspr2, vpx_convolve8_avg_dspr2,
+    vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c,
+    vpx_scaled_vert_c, vpx_scaled_avg_vert_c,
+    vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
+
+const ConvolveParam kArrayConvolve8_dspr2[] = { ALL_SIZES(convolve8_dspr2) };
+INSTANTIATE_TEST_CASE_P(DSPR2, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve8_dspr2));
+#endif  // HAVE_DSPR2
+
+#if HAVE_MSA
+const ConvolveFunctions convolve8_msa(
+    vpx_convolve_copy_msa, vpx_convolve_avg_msa,
+    vpx_convolve8_horiz_msa, vpx_convolve8_avg_horiz_msa,
+    vpx_convolve8_vert_msa, vpx_convolve8_avg_vert_msa,
+    vpx_convolve8_msa, vpx_convolve8_avg_msa,
+    vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c,
+    vpx_scaled_vert_c, vpx_scaled_avg_vert_c,
+    vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
+
+const ConvolveParam kArrayConvolve8_msa[] = { ALL_SIZES(convolve8_msa) };
+INSTANTIATE_TEST_CASE_P(MSA, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve8_msa));
+#endif  // HAVE_MSA
+}  // namespace
diff --git a/libvpx/libvpx/test/cpu_speed_test.cc b/libvpx/libvpx/test/cpu_speed_test.cc
new file mode 100644
index 0000000..2cad30f
--- /dev/null
+++ b/libvpx/libvpx/test/cpu_speed_test.cc
@@ -0,0 +1,161 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+
+namespace {
+
+const int kMaxPSNR = 100;
+
+class CpuSpeedTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
+ protected:
+  CpuSpeedTest()
+      : EncoderTest(GET_PARAM(0)),
+        encoding_mode_(GET_PARAM(1)),
+        set_cpu_used_(GET_PARAM(2)),
+        min_psnr_(kMaxPSNR),
+        tune_content_(VP9E_CONTENT_DEFAULT) {}
+  virtual ~CpuSpeedTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+    if (encoding_mode_ != ::libvpx_test::kRealTime) {
+      cfg_.g_lag_in_frames = 25;
+      cfg_.rc_end_usage = VPX_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = VPX_CBR;
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    min_psnr_ = kMaxPSNR;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_CPUUSED, set_cpu_used_);
+      encoder->Control(VP9E_SET_TUNE_CONTENT, tune_content_);
+      if (encoding_mode_ != ::libvpx_test::kRealTime) {
+        encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+        encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+      }
+    }
+  }
+
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->data.psnr.psnr[0] < min_psnr_)
+      min_psnr_ = pkt->data.psnr.psnr[0];
+  }
+
+  ::libvpx_test::TestMode encoding_mode_;
+  int set_cpu_used_;
+  double min_psnr_;
+  int tune_content_;
+};
+
+TEST_P(CpuSpeedTest, TestQ0) {
+  // Validate that this non multiple of 64 wide clip encodes and decodes
+  // without a mismatch when passing in a very low max q.  This pushes
+  // the encoder to producing lots of big partitions which will likely
+  // extend into the border and test the border condition.
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 400;
+  cfg_.rc_max_quantizer = 0;
+  cfg_.rc_min_quantizer = 0;
+
+  ::libvpx_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       20);
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_GE(min_psnr_, kMaxPSNR);
+}
+
+TEST_P(CpuSpeedTest, TestScreencastQ0) {
+  ::libvpx_test::Y4mVideoSource video("screendata.y4m", 0, 25);
+  cfg_.g_timebase = video.timebase();
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 400;
+  cfg_.rc_max_quantizer = 0;
+  cfg_.rc_min_quantizer = 0;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_GE(min_psnr_, kMaxPSNR);
+}
+
+TEST_P(CpuSpeedTest, TestTuneScreen) {
+  ::libvpx_test::Y4mVideoSource video("screendata.y4m", 0, 25);
+  cfg_.g_timebase = video.timebase();
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_minsection_pct = 2000;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_min_quantizer = 0;
+  tune_content_ = VP9E_CONTENT_SCREEN;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+TEST_P(CpuSpeedTest, TestEncodeHighBitrate) {
+  // Validate that this non multiple of 64 wide clip encodes and decodes
+  // without a mismatch when passing in a very low max q.  This pushes
+  // the encoder to producing lots of big partitions which will likely
+  // extend into the border and test the border condition.
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 12000;
+  cfg_.rc_max_quantizer = 10;
+  cfg_.rc_min_quantizer = 0;
+
+  ::libvpx_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       20);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+TEST_P(CpuSpeedTest, TestLowBitrate) {
+  // Validate that this clip encodes and decodes without a mismatch
+  // when passing in a very high min q.  This pushes the encoder to producing
+  // lots of small partitions which might will test the other condition.
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 200;
+  cfg_.rc_min_quantizer = 40;
+
+  ::libvpx_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       20);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+VP9_INSTANTIATE_TEST_CASE(
+    CpuSpeedTest,
+    ::testing::Values(::libvpx_test::kTwoPassGood, ::libvpx_test::kOnePassGood,
+                      ::libvpx_test::kRealTime),
+    ::testing::Range(0, 9));
+}  // namespace
diff --git a/libvpx/libvpx/test/cq_test.cc b/libvpx/libvpx/test/cq_test.cc
new file mode 100644
index 0000000..4e8019a
--- /dev/null
+++ b/libvpx/libvpx/test/cq_test.cc
@@ -0,0 +1,134 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <cmath>
+#include <map>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+// CQ level range: [kCQLevelMin, kCQLevelMax).
+const int kCQLevelMin = 4;
+const int kCQLevelMax = 63;
+const int kCQLevelStep = 8;
+const unsigned int kCQTargetBitrate = 2000;
+
+class CQTest : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWithParam<int> {
+ public:
+  // maps the cqlevel to the bitrate produced.
+  typedef std::map<int, uint32_t> BitrateMap;
+
+  static void SetUpTestCase() {
+    bitrates_.clear();
+  }
+
+  static void TearDownTestCase() {
+    ASSERT_TRUE(!HasFailure())
+        << "skipping bitrate validation due to earlier failure.";
+    uint32_t prev_actual_bitrate = kCQTargetBitrate;
+    for (BitrateMap::const_iterator iter = bitrates_.begin();
+         iter != bitrates_.end(); ++iter) {
+      const uint32_t cq_actual_bitrate = iter->second;
+      EXPECT_LE(cq_actual_bitrate, prev_actual_bitrate)
+          << "cq_level: " << iter->first
+          << ", bitrate should decrease with increase in CQ level.";
+      prev_actual_bitrate = cq_actual_bitrate;
+    }
+  }
+
+ protected:
+  CQTest() : EncoderTest(GET_PARAM(0)), cq_level_(GET_PARAM(1)) {
+    init_flags_ = VPX_CODEC_USE_PSNR;
+  }
+
+  virtual ~CQTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(libvpx_test::kTwoPassGood);
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    file_size_ = 0;
+    psnr_ = 0.0;
+    n_frames_ = 0;
+  }
+
+  virtual void PreEncodeFrameHook(libvpx_test::VideoSource *video,
+                                  libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      if (cfg_.rc_end_usage == VPX_CQ) {
+        encoder->Control(VP8E_SET_CQ_LEVEL, cq_level_);
+      }
+      encoder->Control(VP8E_SET_CPUUSED, 3);
+    }
+  }
+
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    psnr_ += pow(10.0, pkt->data.psnr.psnr[0] / 10.0);
+    n_frames_++;
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    file_size_ += pkt->data.frame.sz;
+  }
+
+  double GetLinearPSNROverBitrate() const {
+    double avg_psnr = log10(psnr_ / n_frames_) * 10.0;
+    return pow(10.0, avg_psnr / 10.0) / file_size_;
+  }
+
+  int cq_level() const { return cq_level_; }
+  size_t file_size() const { return file_size_; }
+  int n_frames() const { return n_frames_; }
+
+  static BitrateMap bitrates_;
+
+ private:
+  int cq_level_;
+  size_t file_size_;
+  double psnr_;
+  int n_frames_;
+};
+
+CQTest::BitrateMap CQTest::bitrates_;
+
+TEST_P(CQTest, LinearPSNRIsHigherForCQLevel) {
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = kCQTargetBitrate;
+  cfg_.g_lag_in_frames = 25;
+
+  cfg_.rc_end_usage = VPX_CQ;
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 30);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double cq_psnr_lin = GetLinearPSNROverBitrate();
+  const unsigned int cq_actual_bitrate =
+      static_cast<unsigned int>(file_size()) * 8 * 30 / (n_frames() * 1000);
+  EXPECT_LE(cq_actual_bitrate, kCQTargetBitrate);
+  bitrates_[cq_level()] = cq_actual_bitrate;
+
+  // try targeting the approximate same bitrate with VBR mode
+  cfg_.rc_end_usage = VPX_VBR;
+  cfg_.rc_target_bitrate = cq_actual_bitrate;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double vbr_psnr_lin = GetLinearPSNROverBitrate();
+  EXPECT_GE(cq_psnr_lin, vbr_psnr_lin);
+}
+
+VP8_INSTANTIATE_TEST_CASE(CQTest,
+                          ::testing::Range(kCQLevelMin, kCQLevelMax,
+                                           kCQLevelStep));
+}  // namespace
diff --git a/libvpx/libvpx/test/datarate_test.cc b/libvpx/libvpx/test/datarate_test.cc
new file mode 100644
index 0000000..220cbf3
--- /dev/null
+++ b/libvpx/libvpx/test/datarate_test.cc
@@ -0,0 +1,1156 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_config.h"
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "vpx/vpx_codec.h"
+
+namespace {
+
+class DatarateTestLarge : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ public:
+  DatarateTestLarge() : EncoderTest(GET_PARAM(0)) {}
+
+  virtual ~DatarateTestLarge() {}
+
+ protected:
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    ResetModel();
+  }
+
+  virtual void ResetModel() {
+    last_pts_ = 0;
+    bits_in_buffer_model_ = cfg_.rc_target_bitrate * cfg_.rc_buf_initial_sz;
+    frame_number_ = 0;
+    first_drop_ = 0;
+    bits_total_ = 0;
+    duration_ = 0.0;
+    denoiser_offon_test_ = 0;
+    denoiser_offon_period_ = -1;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 0)
+      encoder->Control(VP8E_SET_NOISE_SENSITIVITY, denoiser_on_);
+
+    if (denoiser_offon_test_) {
+      ASSERT_GT(denoiser_offon_period_, 0)
+          << "denoiser_offon_period_ is not positive.";
+      if ((video->frame() + 1) % denoiser_offon_period_ == 0) {
+        // Flip denoiser_on_ periodically
+        denoiser_on_ ^= 1;
+      }
+      encoder->Control(VP8E_SET_NOISE_SENSITIVITY, denoiser_on_);
+    }
+
+    const vpx_rational_t tb = video->timebase();
+    timebase_ = static_cast<double>(tb.num) / tb.den;
+    duration_ = 0;
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    // Time since last timestamp = duration.
+    vpx_codec_pts_t duration = pkt->data.frame.pts - last_pts_;
+
+    // TODO(jimbankoski): Remove these lines when the issue:
+    // http://code.google.com/p/webm/issues/detail?id=496 is fixed.
+    // For now the codec assumes buffer starts at starting buffer rate
+    // plus one frame's time.
+    if (last_pts_ == 0)
+      duration = 1;
+
+    // Add to the buffer the bits we'd expect from a constant bitrate server.
+    bits_in_buffer_model_ += static_cast<int64_t>(
+        duration * timebase_ * cfg_.rc_target_bitrate * 1000);
+
+    /* Test the buffer model here before subtracting the frame. Do so because
+     * the way the leaky bucket model works in libvpx is to allow the buffer to
+     * empty - and then stop showing frames until we've got enough bits to
+     * show one. As noted in comment below (issue 495), this does not currently
+     * apply to key frames. For now exclude key frames in condition below. */
+    const bool key_frame = (pkt->data.frame.flags & VPX_FRAME_IS_KEY)
+                         ? true: false;
+    if (!key_frame) {
+      ASSERT_GE(bits_in_buffer_model_, 0) << "Buffer Underrun at frame "
+          << pkt->data.frame.pts;
+    }
+
+    const int64_t frame_size_in_bits = pkt->data.frame.sz * 8;
+
+    // Subtract from the buffer the bits associated with a played back frame.
+    bits_in_buffer_model_ -= frame_size_in_bits;
+
+    // Update the running total of bits for end of test datarate checks.
+    bits_total_ += frame_size_in_bits;
+
+    // If first drop not set and we have a drop set it to this time.
+    if (!first_drop_ && duration > 1)
+      first_drop_ = last_pts_ + 1;
+
+    // Update the most recent pts.
+    last_pts_ = pkt->data.frame.pts;
+
+    // We update this so that we can calculate the datarate minus the last
+    // frame encoded in the file.
+    bits_in_last_frame_ = frame_size_in_bits;
+
+    ++frame_number_;
+  }
+
+  virtual void EndPassHook(void) {
+    if (bits_total_) {
+      const double file_size_in_kb = bits_total_ / 1000.;  // bits per kilobit
+
+      duration_ = (last_pts_ + 1) * timebase_;
+
+      // Effective file datarate includes the time spent prebuffering.
+      effective_datarate_ = (bits_total_ - bits_in_last_frame_) / 1000.0
+          / (cfg_.rc_buf_initial_sz / 1000.0 + duration_);
+
+      file_datarate_ = file_size_in_kb / duration_;
+    }
+  }
+
+  vpx_codec_pts_t last_pts_;
+  int64_t bits_in_buffer_model_;
+  double timebase_;
+  int frame_number_;
+  vpx_codec_pts_t first_drop_;
+  int64_t bits_total_;
+  double duration_;
+  double file_datarate_;
+  double effective_datarate_;
+  int64_t bits_in_last_frame_;
+  int denoiser_on_;
+  int denoiser_offon_test_;
+  int denoiser_offon_period_;
+};
+
+#if CONFIG_TEMPORAL_DENOISING
+// Check basic datarate targeting, for a single bitrate, but loop over the
+// various denoiser settings.
+TEST_P(DatarateTestLarge, DenoiserLevels) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_max_quantizer = 56;
+  cfg_.rc_end_usage = VPX_CBR;
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 140);
+  for (int j = 1; j < 5; ++j) {
+    // Run over the denoiser levels.
+    // For the temporal denoiser (#if CONFIG_TEMPORAL_DENOISING) the level j
+    // refers to the 4 denoiser modes: denoiserYonly, denoiserOnYUV,
+    // denoiserOnAggressive, and denoiserOnAdaptive.
+    // For the spatial denoiser (if !CONFIG_TEMPORAL_DENOISING), the level j
+    // refers to the blur thresholds: 20, 40, 60 80.
+    // The j = 0 case (denoiser off) is covered in the tests below.
+    denoiser_on_ = j;
+    cfg_.rc_target_bitrate = 300;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(cfg_.rc_target_bitrate, effective_datarate_ * 0.95)
+        << " The datarate for the file exceeds the target!";
+
+    ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.3)
+        << " The datarate for the file missed the target!";
+  }
+}
+
+// Check basic datarate targeting, for a single bitrate, when denoiser is off
+// and on.
+TEST_P(DatarateTestLarge, DenoiserOffOn) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_max_quantizer = 56;
+  cfg_.rc_end_usage = VPX_CBR;
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 299);
+  cfg_.rc_target_bitrate = 300;
+  ResetModel();
+  // The denoiser is off by default.
+  denoiser_on_ = 0;
+  // Set the offon test flag.
+  denoiser_offon_test_ = 1;
+  denoiser_offon_period_ = 100;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  ASSERT_GE(cfg_.rc_target_bitrate, effective_datarate_ * 0.95)
+      << " The datarate for the file exceeds the target!";
+  ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.3)
+      << " The datarate for the file missed the target!";
+}
+#endif  // CONFIG_TEMPORAL_DENOISING
+
+TEST_P(DatarateTestLarge, BasicBufferModel) {
+  denoiser_on_ = 0;
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_max_quantizer = 56;
+  cfg_.rc_end_usage = VPX_CBR;
+  // 2 pass cbr datarate control has a bug hidden by the small # of
+  // frames selected in this encode. The problem is that even if the buffer is
+  // negative we produce a keyframe on a cutscene. Ignoring datarate
+  // constraints
+  // TODO(jimbankoski): ( Fix when issue
+  // http://code.google.com/p/webm/issues/detail?id=495 is addressed. )
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 140);
+
+  // There is an issue for low bitrates in real-time mode, where the
+  // effective_datarate slightly overshoots the target bitrate.
+  // This is same the issue as noted about (#495).
+  // TODO(jimbankoski/marpan): Update test to run for lower bitrates (< 100),
+  // when the issue is resolved.
+  for (int i = 100; i < 800; i += 200) {
+    cfg_.rc_target_bitrate = i;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(cfg_.rc_target_bitrate, effective_datarate_ * 0.95)
+        << " The datarate for the file exceeds the target!";
+
+    ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.3)
+        << " The datarate for the file missed the target!";
+  }
+}
+
+TEST_P(DatarateTestLarge, ChangingDropFrameThresh) {
+  denoiser_on_ = 0;
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_max_quantizer = 36;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.rc_target_bitrate = 200;
+  cfg_.kf_mode = VPX_KF_DISABLED;
+
+  const int frame_count = 40;
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, frame_count);
+
+  // Here we check that the first dropped frame gets earlier and earlier
+  // as the drop frame threshold is increased.
+
+  const int kDropFrameThreshTestStep = 30;
+  vpx_codec_pts_t last_drop = frame_count;
+  for (int i = 1; i < 91; i += kDropFrameThreshTestStep) {
+    cfg_.rc_dropframe_thresh = i;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_LE(first_drop_, last_drop)
+        << " The first dropped frame for drop_thresh " << i
+        << " > first dropped frame for drop_thresh "
+        << i - kDropFrameThreshTestStep;
+    last_drop = first_drop_;
+  }
+}
+
+class DatarateTestVP9Large : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
+ public:
+  DatarateTestVP9Large() : EncoderTest(GET_PARAM(0)) {}
+
+ protected:
+  virtual ~DatarateTestVP9Large() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    set_cpu_used_ = GET_PARAM(2);
+    ResetModel();
+  }
+
+  virtual void ResetModel() {
+    last_pts_ = 0;
+    bits_in_buffer_model_ = cfg_.rc_target_bitrate * cfg_.rc_buf_initial_sz;
+    frame_number_ = 0;
+    tot_frame_number_ = 0;
+    first_drop_ = 0;
+    num_drops_ = 0;
+    // Denoiser is off by default.
+    denoiser_on_ = 0;
+    // For testing up to 3 layers.
+    for (int i = 0; i < 3; ++i) {
+      bits_total_[i] = 0;
+    }
+    denoiser_offon_test_ = 0;
+    denoiser_offon_period_ = -1;
+  }
+
+  //
+  // Frame flags and layer id for temporal layers.
+  //
+
+  // For two layers, test pattern is:
+  //   1     3
+  // 0    2     .....
+  // For three layers, test pattern is:
+  //   1      3    5      7
+  //      2           6
+  // 0          4            ....
+  // LAST is always update on base/layer 0, GOLDEN is updated on layer 1.
+  // For this 3 layer example, the 2nd enhancement layer (layer 2) does not
+  // update any reference frames.
+  int SetFrameFlags(int frame_num, int num_temp_layers) {
+    int frame_flags = 0;
+    if (num_temp_layers == 2) {
+      if (frame_num % 2 == 0) {
+        // Layer 0: predict from L and ARF, update L.
+        frame_flags = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_UPD_GF |
+                      VP8_EFLAG_NO_UPD_ARF;
+      } else {
+        // Layer 1: predict from L, G and ARF, and update G.
+        frame_flags = VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST |
+                      VP8_EFLAG_NO_UPD_ENTROPY;
+      }
+    } else if (num_temp_layers == 3) {
+      if (frame_num % 4 == 0) {
+        // Layer 0: predict from L and ARF; update L.
+        frame_flags = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF |
+                      VP8_EFLAG_NO_REF_GF;
+      } else if ((frame_num - 2) % 4 == 0) {
+        // Layer 1: predict from L, G, ARF; update G.
+        frame_flags = VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST;
+      }  else if ((frame_num - 1) % 2 == 0) {
+        // Layer 2: predict from L, G, ARF; update none.
+        frame_flags = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF |
+                      VP8_EFLAG_NO_UPD_LAST;
+      }
+    }
+    return frame_flags;
+  }
+
+  int SetLayerId(int frame_num, int num_temp_layers) {
+    int layer_id = 0;
+    if (num_temp_layers == 2) {
+      if (frame_num % 2 == 0) {
+        layer_id = 0;
+      } else {
+        layer_id = 1;
+      }
+    } else if (num_temp_layers == 3) {
+      if (frame_num % 4 == 0) {
+        layer_id = 0;
+      } else if ((frame_num - 2) % 4 == 0) {
+        layer_id = 1;
+      } else if ((frame_num - 1) % 2 == 0) {
+        layer_id = 2;
+      }
+    }
+    return layer_id;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 0)
+      encoder->Control(VP8E_SET_CPUUSED, set_cpu_used_);
+
+    if (denoiser_offon_test_) {
+      ASSERT_GT(denoiser_offon_period_, 0)
+          << "denoiser_offon_period_ is not positive.";
+      if ((video->frame() + 1) % denoiser_offon_period_ == 0) {
+        // Flip denoiser_on_ periodically
+        denoiser_on_ ^= 1;
+      }
+    }
+
+    encoder->Control(VP9E_SET_NOISE_SENSITIVITY, denoiser_on_);
+
+    if (cfg_.ts_number_layers > 1) {
+      if (video->frame() == 0) {
+        encoder->Control(VP9E_SET_SVC, 1);
+      }
+      vpx_svc_layer_id_t layer_id;
+      layer_id.spatial_layer_id = 0;
+      frame_flags_ = SetFrameFlags(video->frame(), cfg_.ts_number_layers);
+      layer_id.temporal_layer_id = SetLayerId(video->frame(),
+                                              cfg_.ts_number_layers);
+      encoder->Control(VP9E_SET_SVC_LAYER_ID, &layer_id);
+    }
+    const vpx_rational_t tb = video->timebase();
+    timebase_ = static_cast<double>(tb.num) / tb.den;
+    duration_ = 0;
+  }
+
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    // Time since last timestamp = duration.
+    vpx_codec_pts_t duration = pkt->data.frame.pts - last_pts_;
+
+    if (duration > 1) {
+      // If first drop not set and we have a drop set it to this time.
+      if (!first_drop_)
+        first_drop_ = last_pts_ + 1;
+      // Update the number of frame drops.
+      num_drops_ += static_cast<int>(duration - 1);
+      // Update counter for total number of frames (#frames input to encoder).
+      // Needed for setting the proper layer_id below.
+      tot_frame_number_ += static_cast<int>(duration - 1);
+    }
+
+    int layer = SetLayerId(tot_frame_number_, cfg_.ts_number_layers);
+
+    // Add to the buffer the bits we'd expect from a constant bitrate server.
+    bits_in_buffer_model_ += static_cast<int64_t>(
+        duration * timebase_ * cfg_.rc_target_bitrate * 1000);
+
+    // Buffer should not go negative.
+    ASSERT_GE(bits_in_buffer_model_, 0) << "Buffer Underrun at frame "
+        << pkt->data.frame.pts;
+
+    const size_t frame_size_in_bits = pkt->data.frame.sz * 8;
+
+    // Update the total encoded bits. For temporal layers, update the cumulative
+    // encoded bits per layer.
+    for (int i = layer; i < static_cast<int>(cfg_.ts_number_layers); ++i) {
+      bits_total_[i] += frame_size_in_bits;
+    }
+
+    // Update the most recent pts.
+    last_pts_ = pkt->data.frame.pts;
+    ++frame_number_;
+    ++tot_frame_number_;
+  }
+
+  virtual void EndPassHook(void) {
+    for (int layer = 0; layer < static_cast<int>(cfg_.ts_number_layers);
+        ++layer) {
+      duration_ = (last_pts_ + 1) * timebase_;
+      if (bits_total_[layer]) {
+        // Effective file datarate:
+        effective_datarate_[layer] = (bits_total_[layer] / 1000.0) / duration_;
+      }
+    }
+  }
+
+  vpx_codec_pts_t last_pts_;
+  double timebase_;
+  int frame_number_;      // Counter for number of non-dropped/encoded frames.
+  int tot_frame_number_;  // Counter for total number of input frames.
+  int64_t bits_total_[3];
+  double duration_;
+  double effective_datarate_[3];
+  int set_cpu_used_;
+  int64_t bits_in_buffer_model_;
+  vpx_codec_pts_t first_drop_;
+  int num_drops_;
+  int denoiser_on_;
+  int denoiser_offon_test_;
+  int denoiser_offon_period_;
+};
+
+// Check basic rate targeting for VBR mode.
+TEST_P(DatarateTestVP9Large, BasicRateTargetingVBR) {
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.g_error_resilient = 0;
+  cfg_.rc_end_usage = VPX_VBR;
+  cfg_.g_lag_in_frames = 0;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 300);
+  for (int i = 400; i <= 800; i += 400) {
+    cfg_.rc_target_bitrate = i;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(effective_datarate_[0], cfg_.rc_target_bitrate * 0.75)
+        << " The datarate for the file is lower than target by too much!";
+    ASSERT_LE(effective_datarate_[0], cfg_.rc_target_bitrate * 1.25)
+        << " The datarate for the file is greater than target by too much!";
+  }
+}
+
+// Check basic rate targeting for CBR,
+TEST_P(DatarateTestVP9Large, BasicRateTargeting) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 140);
+  for (int i = 150; i < 800; i += 200) {
+    cfg_.rc_target_bitrate = i;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(effective_datarate_[0], cfg_.rc_target_bitrate * 0.85)
+        << " The datarate for the file is lower than target by too much!";
+    ASSERT_LE(effective_datarate_[0], cfg_.rc_target_bitrate * 1.15)
+        << " The datarate for the file is greater than target by too much!";
+  }
+}
+
+// Check basic rate targeting for CBR.
+TEST_P(DatarateTestVP9Large, BasicRateTargeting444) {
+  ::libvpx_test::Y4mVideoSource video("rush_hour_444.y4m", 0, 140);
+
+  cfg_.g_profile = 1;
+  cfg_.g_timebase = video.timebase();
+
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = VPX_CBR;
+
+  for (int i = 250; i < 900; i += 200) {
+    cfg_.rc_target_bitrate = i;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(static_cast<double>(cfg_.rc_target_bitrate),
+              effective_datarate_[0] * 0.85)
+        << " The datarate for the file exceeds the target by too much!";
+    ASSERT_LE(static_cast<double>(cfg_.rc_target_bitrate),
+              effective_datarate_[0] * 1.15)
+        << " The datarate for the file missed the target!"
+        << cfg_.rc_target_bitrate << " "<< effective_datarate_;
+  }
+}
+
+// Check that (1) the first dropped frame gets earlier and earlier
+// as the drop frame threshold is increased, and (2) that the total number of
+// frame drops does not decrease as we increase frame drop threshold.
+// Use a lower qp-max to force some frame drops.
+TEST_P(DatarateTestVP9Large, ChangingDropFrameThresh) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_undershoot_pct = 20;
+  cfg_.rc_undershoot_pct = 20;
+  cfg_.rc_dropframe_thresh = 10;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 50;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.rc_target_bitrate = 200;
+  cfg_.g_lag_in_frames = 0;
+  // TODO(marpan): Investigate datarate target failures with a smaller keyframe
+  // interval (128).
+  cfg_.kf_max_dist = 9999;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 140);
+
+  const int kDropFrameThreshTestStep = 30;
+  vpx_codec_pts_t last_drop = 140;
+  int last_num_drops = 0;
+  for (int i = 10; i < 100; i += kDropFrameThreshTestStep) {
+    cfg_.rc_dropframe_thresh = i;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(effective_datarate_[0], cfg_.rc_target_bitrate * 0.85)
+        << " The datarate for the file is lower than target by too much!";
+    ASSERT_LE(effective_datarate_[0], cfg_.rc_target_bitrate * 1.15)
+        << " The datarate for the file is greater than target by too much!";
+    ASSERT_LE(first_drop_, last_drop)
+        << " The first dropped frame for drop_thresh " << i
+        << " > first dropped frame for drop_thresh "
+        << i - kDropFrameThreshTestStep;
+    ASSERT_GE(num_drops_, last_num_drops * 0.85)
+        << " The number of dropped frames for drop_thresh " << i
+        << " < number of dropped frames for drop_thresh "
+        << i - kDropFrameThreshTestStep;
+    last_drop = first_drop_;
+    last_num_drops = num_drops_;
+  }
+}
+
+// Check basic rate targeting for 2 temporal layers.
+TEST_P(DatarateTestVP9Large, BasicRateTargeting2TemporalLayers) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+
+  // 2 Temporal layers, no spatial layers: Framerate decimation (2, 1).
+  cfg_.ss_number_layers = 1;
+  cfg_.ts_number_layers = 2;
+  cfg_.ts_rate_decimator[0] = 2;
+  cfg_.ts_rate_decimator[1] = 1;
+
+  cfg_.temporal_layering_mode = VP9E_TEMPORAL_LAYERING_MODE_BYPASS;
+
+  if (deadline_ == VPX_DL_REALTIME)
+    cfg_.g_error_resilient = 1;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 200);
+  for (int i = 200; i <= 800; i += 200) {
+    cfg_.rc_target_bitrate = i;
+    ResetModel();
+    // 60-40 bitrate allocation for 2 temporal layers.
+    cfg_.layer_target_bitrate[0] = 60 * cfg_.rc_target_bitrate / 100;
+    cfg_.layer_target_bitrate[1] = cfg_.rc_target_bitrate;
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    for (int j = 0; j < static_cast<int>(cfg_.ts_number_layers); ++j) {
+      ASSERT_GE(effective_datarate_[j], cfg_.layer_target_bitrate[j] * 0.85)
+          << " The datarate for the file is lower than target by too much, "
+              "for layer: " << j;
+      ASSERT_LE(effective_datarate_[j], cfg_.layer_target_bitrate[j] * 1.15)
+          << " The datarate for the file is greater than target by too much, "
+              "for layer: " << j;
+    }
+  }
+}
+
+// Check basic rate targeting for 3 temporal layers.
+TEST_P(DatarateTestVP9Large, BasicRateTargeting3TemporalLayers) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+
+  // 3 Temporal layers, no spatial layers: Framerate decimation (4, 2, 1).
+  cfg_.ss_number_layers = 1;
+  cfg_.ts_number_layers = 3;
+  cfg_.ts_rate_decimator[0] = 4;
+  cfg_.ts_rate_decimator[1] = 2;
+  cfg_.ts_rate_decimator[2] = 1;
+
+  cfg_.temporal_layering_mode = VP9E_TEMPORAL_LAYERING_MODE_BYPASS;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 200);
+  for (int i = 200; i <= 800; i += 200) {
+    cfg_.rc_target_bitrate = i;
+    ResetModel();
+    // 40-20-40 bitrate allocation for 3 temporal layers.
+    cfg_.layer_target_bitrate[0] = 40 * cfg_.rc_target_bitrate / 100;
+    cfg_.layer_target_bitrate[1] = 60 * cfg_.rc_target_bitrate / 100;
+    cfg_.layer_target_bitrate[2] = cfg_.rc_target_bitrate;
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    for (int j = 0; j < static_cast<int>(cfg_.ts_number_layers); ++j) {
+      // TODO(yaowu): Work out more stable rc control strategy and
+      //              Adjust the thresholds to be tighter than .75.
+      ASSERT_GE(effective_datarate_[j], cfg_.layer_target_bitrate[j] * 0.75)
+          << " The datarate for the file is lower than target by too much, "
+              "for layer: " << j;
+      // TODO(yaowu): Work out more stable rc control strategy and
+      //              Adjust the thresholds to be tighter than 1.25.
+      ASSERT_LE(effective_datarate_[j], cfg_.layer_target_bitrate[j] * 1.25)
+          << " The datarate for the file is greater than target by too much, "
+              "for layer: " << j;
+    }
+  }
+}
+
+// Check basic rate targeting for 3 temporal layers, with frame dropping.
+// Only for one (low) bitrate with lower max_quantizer, and somewhat higher
+// frame drop threshold, to force frame dropping.
+TEST_P(DatarateTestVP9Large, BasicRateTargeting3TemporalLayersFrameDropping) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  // Set frame drop threshold and rc_max_quantizer to force some frame drops.
+  cfg_.rc_dropframe_thresh = 20;
+  cfg_.rc_max_quantizer = 45;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+
+  // 3 Temporal layers, no spatial layers: Framerate decimation (4, 2, 1).
+  cfg_.ss_number_layers = 1;
+  cfg_.ts_number_layers = 3;
+  cfg_.ts_rate_decimator[0] = 4;
+  cfg_.ts_rate_decimator[1] = 2;
+  cfg_.ts_rate_decimator[2] = 1;
+
+  cfg_.temporal_layering_mode = VP9E_TEMPORAL_LAYERING_MODE_BYPASS;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 200);
+  cfg_.rc_target_bitrate = 200;
+  ResetModel();
+  // 40-20-40 bitrate allocation for 3 temporal layers.
+  cfg_.layer_target_bitrate[0] = 40 * cfg_.rc_target_bitrate / 100;
+  cfg_.layer_target_bitrate[1] = 60 * cfg_.rc_target_bitrate / 100;
+  cfg_.layer_target_bitrate[2] = cfg_.rc_target_bitrate;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  for (int j = 0; j < static_cast<int>(cfg_.ts_number_layers); ++j) {
+    ASSERT_GE(effective_datarate_[j], cfg_.layer_target_bitrate[j] * 0.85)
+        << " The datarate for the file is lower than target by too much, "
+            "for layer: " << j;
+    ASSERT_LE(effective_datarate_[j], cfg_.layer_target_bitrate[j] * 1.15)
+        << " The datarate for the file is greater than target by too much, "
+            "for layer: " << j;
+    // Expect some frame drops in this test: for this 200 frames test,
+    // expect at least 10% and not more than 60% drops.
+    ASSERT_GE(num_drops_, 20);
+    ASSERT_LE(num_drops_, 130);
+  }
+}
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+// Check basic datarate targeting, for a single bitrate, when denoiser is on.
+TEST_P(DatarateTestVP9Large, DenoiserLevels) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_min_quantizer = 2;
+  cfg_.rc_max_quantizer = 56;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 140);
+
+  // For the temporal denoiser (#if CONFIG_VP9_TEMPORAL_DENOISING),
+  // there is only one denoiser mode: denoiserYonly(which is 1),
+  // but may add more modes in the future.
+  cfg_.rc_target_bitrate = 300;
+  ResetModel();
+  // Turn on the denoiser.
+  denoiser_on_ = 1;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  ASSERT_GE(effective_datarate_[0], cfg_.rc_target_bitrate * 0.85)
+      << " The datarate for the file is lower than target by too much!";
+  ASSERT_LE(effective_datarate_[0], cfg_.rc_target_bitrate * 1.15)
+      << " The datarate for the file is greater than target by too much!";
+}
+
+// Check basic datarate targeting, for a single bitrate, when denoiser is off
+// and on.
+TEST_P(DatarateTestVP9Large, DenoiserOffOn) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_min_quantizer = 2;
+  cfg_.rc_max_quantizer = 56;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 299);
+
+  // For the temporal denoiser (#if CONFIG_VP9_TEMPORAL_DENOISING),
+  // there is only one denoiser mode: denoiserYonly(which is 1),
+  // but may add more modes in the future.
+  cfg_.rc_target_bitrate = 300;
+  ResetModel();
+  // The denoiser is off by default.
+  denoiser_on_ = 0;
+  // Set the offon test flag.
+  denoiser_offon_test_ = 1;
+  denoiser_offon_period_ = 100;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  ASSERT_GE(effective_datarate_[0], cfg_.rc_target_bitrate * 0.85)
+      << " The datarate for the file is lower than target by too much!";
+  ASSERT_LE(effective_datarate_[0], cfg_.rc_target_bitrate * 1.15)
+      << " The datarate for the file is greater than target by too much!";
+}
+#endif  // CONFIG_VP9_TEMPORAL_DENOISING
+
+class DatarateOnePassCbrSvc : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
+ public:
+  DatarateOnePassCbrSvc() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~DatarateOnePassCbrSvc() {}
+ protected:
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    speed_setting_ = GET_PARAM(2);
+    ResetModel();
+  }
+  virtual void ResetModel() {
+    last_pts_ = 0;
+    bits_in_buffer_model_ = cfg_.rc_target_bitrate * cfg_.rc_buf_initial_sz;
+    frame_number_ = 0;
+    first_drop_ = 0;
+    bits_total_ = 0;
+    duration_ = 0.0;
+    mismatch_psnr_ = 0.0;
+    mismatch_nframes_ = 0;
+  }
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+  }
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      int i;
+      for (i = 0; i < VPX_MAX_LAYERS; ++i) {
+        svc_params_.max_quantizers[i] = 63;
+        svc_params_.min_quantizers[i] = 0;
+      }
+      encoder->Control(VP9E_SET_SVC, 1);
+      encoder->Control(VP9E_SET_SVC_PARAMETERS, &svc_params_);
+      encoder->Control(VP8E_SET_CPUUSED, speed_setting_);
+      encoder->Control(VP9E_SET_TILE_COLUMNS, 0);
+      encoder->Control(VP8E_SET_MAX_INTRA_BITRATE_PCT, 300);
+      encoder->Control(VP9E_SET_TILE_COLUMNS, (cfg_.g_threads >> 1));
+    }
+    const vpx_rational_t tb = video->timebase();
+    timebase_ = static_cast<double>(tb.num) / tb.den;
+    duration_ = 0;
+  }
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    vpx_codec_pts_t duration = pkt->data.frame.pts - last_pts_;
+    if (last_pts_ == 0)
+      duration = 1;
+    bits_in_buffer_model_ += static_cast<int64_t>(
+        duration * timebase_ * cfg_.rc_target_bitrate * 1000);
+    const bool key_frame = (pkt->data.frame.flags & VPX_FRAME_IS_KEY)
+                         ? true: false;
+    if (!key_frame) {
+      ASSERT_GE(bits_in_buffer_model_, 0) << "Buffer Underrun at frame "
+          << pkt->data.frame.pts;
+    }
+    const size_t frame_size_in_bits = pkt->data.frame.sz * 8;
+    bits_in_buffer_model_ -= frame_size_in_bits;
+    bits_total_ += frame_size_in_bits;
+    if (!first_drop_ && duration > 1)
+      first_drop_ = last_pts_ + 1;
+    last_pts_ = pkt->data.frame.pts;
+    bits_in_last_frame_ = frame_size_in_bits;
+    ++frame_number_;
+  }
+  virtual void EndPassHook(void) {
+    if (bits_total_) {
+      const double file_size_in_kb = bits_total_ / 1000.;  // bits per kilobit
+      duration_ = (last_pts_ + 1) * timebase_;
+      file_datarate_ = file_size_in_kb / duration_;
+    }
+  }
+
+  virtual void MismatchHook(const vpx_image_t *img1,
+                            const vpx_image_t *img2) {
+    double mismatch_psnr = compute_psnr(img1, img2);
+    mismatch_psnr_ += mismatch_psnr;
+    ++mismatch_nframes_;
+  }
+
+  unsigned int GetMismatchFrames() {
+    return mismatch_nframes_;
+  }
+
+  vpx_codec_pts_t last_pts_;
+  int64_t bits_in_buffer_model_;
+  double timebase_;
+  int frame_number_;
+  vpx_codec_pts_t first_drop_;
+  int64_t bits_total_;
+  double duration_;
+  double file_datarate_;
+  size_t bits_in_last_frame_;
+  vpx_svc_extra_cfg_t svc_params_;
+  int speed_setting_;
+  double mismatch_psnr_;
+  int mismatch_nframes_;
+};
+static void assign_layer_bitrates(vpx_codec_enc_cfg_t *const enc_cfg,
+    const vpx_svc_extra_cfg_t *svc_params,
+    int spatial_layers,
+    int temporal_layers,
+    int temporal_layering_mode) {
+  int sl, spatial_layer_target;
+  float total = 0;
+  float alloc_ratio[VPX_MAX_LAYERS] = {0};
+  for (sl = 0; sl < spatial_layers; ++sl) {
+    if (svc_params->scaling_factor_den[sl] > 0) {
+      alloc_ratio[sl] = (float)(svc_params->scaling_factor_num[sl] *
+          1.0 / svc_params->scaling_factor_den[sl]);
+      total += alloc_ratio[sl];
+    }
+  }
+  for (sl = 0; sl < spatial_layers; ++sl) {
+    enc_cfg->ss_target_bitrate[sl] = spatial_layer_target =
+        (unsigned int)(enc_cfg->rc_target_bitrate *
+            alloc_ratio[sl] / total);
+    const int index = sl * temporal_layers;
+    if (temporal_layering_mode == 3) {
+      enc_cfg->layer_target_bitrate[index] =
+          spatial_layer_target >> 1;
+      enc_cfg->layer_target_bitrate[index + 1] =
+          (spatial_layer_target >> 1) + (spatial_layer_target >> 2);
+      enc_cfg->layer_target_bitrate[index + 2] =
+          spatial_layer_target;
+    } else if (temporal_layering_mode == 2) {
+      enc_cfg->layer_target_bitrate[index] =
+          spatial_layer_target * 2 / 3;
+      enc_cfg->layer_target_bitrate[index + 1] =
+          spatial_layer_target;
+    }
+  }
+}
+
+// Check basic rate targeting for 1 pass CBR SVC: 2 spatial layers and
+// 3 temporal layers. Run CIF clip with 1 thread.
+TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc2SpatialLayers) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.ss_number_layers = 2;
+  cfg_.ts_number_layers = 3;
+  cfg_.ts_rate_decimator[0] = 4;
+  cfg_.ts_rate_decimator[1] = 2;
+  cfg_.ts_rate_decimator[2] = 1;
+  cfg_.g_error_resilient = 1;
+  cfg_.g_threads = 1;
+  cfg_.temporal_layering_mode = 3;
+  svc_params_.scaling_factor_num[0] = 144;
+  svc_params_.scaling_factor_den[0] = 288;
+  svc_params_.scaling_factor_num[1] = 288;
+  svc_params_.scaling_factor_den[1] = 288;
+  cfg_.rc_dropframe_thresh = 10;
+  cfg_.kf_max_dist = 9999;
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 200);
+  // TODO(wonkap/marpan): Check that effective_datarate for each layer hits the
+  // layer target_bitrate.
+  for (int i = 200; i <= 800; i += 200) {
+    cfg_.rc_target_bitrate = i;
+    ResetModel();
+    assign_layer_bitrates(&cfg_, &svc_params_, cfg_.ss_number_layers,
+        cfg_.ts_number_layers, cfg_.temporal_layering_mode);
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(cfg_.rc_target_bitrate, file_datarate_ * 0.85)
+            << " The datarate for the file exceeds the target by too much!";
+    ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.15)
+        << " The datarate for the file is lower than the target by too much!";
+    EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+  }
+}
+
+// Check basic rate targeting for 1 pass CBR SVC: 2 spatial layers and 3
+// temporal layers. Run CIF clip with 1 thread, and few short key frame periods.
+TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc2SpatialLayersSmallKf) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.ss_number_layers = 2;
+  cfg_.ts_number_layers = 3;
+  cfg_.ts_rate_decimator[0] = 4;
+  cfg_.ts_rate_decimator[1] = 2;
+  cfg_.ts_rate_decimator[2] = 1;
+  cfg_.g_error_resilient = 1;
+  cfg_.g_threads = 1;
+  cfg_.temporal_layering_mode = 3;
+  svc_params_.scaling_factor_num[0] = 144;
+  svc_params_.scaling_factor_den[0] = 288;
+  svc_params_.scaling_factor_num[1] = 288;
+  svc_params_.scaling_factor_den[1] = 288;
+  cfg_.rc_dropframe_thresh = 10;
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 200);
+  cfg_.rc_target_bitrate = 400;
+  // For this 3 temporal layer case, pattern repeats every 4 frames, so choose
+  // 4 key neighboring key frame periods (so key frame will land on 0-2-1-2).
+  for (int j = 64; j <= 67; j++) {
+    cfg_.kf_max_dist = j;
+    ResetModel();
+    assign_layer_bitrates(&cfg_, &svc_params_, cfg_.ss_number_layers,
+        cfg_.ts_number_layers, cfg_.temporal_layering_mode);
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(cfg_.rc_target_bitrate, file_datarate_ * 0.85)
+            << " The datarate for the file exceeds the target by too much!";
+    ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.15)
+        << " The datarate for the file is lower than the target by too much!";
+    EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+  }
+}
+
+// Check basic rate targeting for 1 pass CBR SVC: 2 spatial layers and
+// 3 temporal layers. Run HD clip with 4 threads.
+TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc2SpatialLayers4threads) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.ss_number_layers = 2;
+  cfg_.ts_number_layers = 3;
+  cfg_.ts_rate_decimator[0] = 4;
+  cfg_.ts_rate_decimator[1] = 2;
+  cfg_.ts_rate_decimator[2] = 1;
+  cfg_.g_error_resilient = 1;
+  cfg_.g_threads = 4;
+  cfg_.temporal_layering_mode = 3;
+  svc_params_.scaling_factor_num[0] = 144;
+  svc_params_.scaling_factor_den[0] = 288;
+  svc_params_.scaling_factor_num[1] = 288;
+  svc_params_.scaling_factor_den[1] = 288;
+  cfg_.rc_dropframe_thresh = 10;
+  cfg_.kf_max_dist = 9999;
+  ::libvpx_test::I420VideoSource video("niklas_1280_720_30.y4m", 1280, 720,
+                                       30, 1, 0, 300);
+  cfg_.rc_target_bitrate = 800;
+  ResetModel();
+  assign_layer_bitrates(&cfg_, &svc_params_, cfg_.ss_number_layers,
+      cfg_.ts_number_layers, cfg_.temporal_layering_mode);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  ASSERT_GE(cfg_.rc_target_bitrate, file_datarate_ * 0.85)
+          << " The datarate for the file exceeds the target by too much!";
+  ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.15)
+      << " The datarate for the file is lower than the target by too much!";
+  EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+}
+
+// Check basic rate targeting for 1 pass CBR SVC: 3 spatial layers and
+// 3 temporal layers. Run CIF clip with 1 thread.
+TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc3SpatialLayers) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.ss_number_layers = 3;
+  cfg_.ts_number_layers = 3;
+  cfg_.ts_rate_decimator[0] = 4;
+  cfg_.ts_rate_decimator[1] = 2;
+  cfg_.ts_rate_decimator[2] = 1;
+  cfg_.g_error_resilient = 1;
+  cfg_.g_threads = 1;
+  cfg_.temporal_layering_mode = 3;
+  svc_params_.scaling_factor_num[0] = 72;
+  svc_params_.scaling_factor_den[0] = 288;
+  svc_params_.scaling_factor_num[1] = 144;
+  svc_params_.scaling_factor_den[1] = 288;
+  svc_params_.scaling_factor_num[2] = 288;
+  svc_params_.scaling_factor_den[2] = 288;
+  cfg_.rc_dropframe_thresh = 10;
+  cfg_.kf_max_dist = 9999;
+  ::libvpx_test::I420VideoSource video("niklas_1280_720_30.y4m", 1280, 720,
+                                       30, 1, 0, 300);
+  cfg_.rc_target_bitrate = 800;
+  ResetModel();
+  assign_layer_bitrates(&cfg_, &svc_params_, cfg_.ss_number_layers,
+     cfg_.ts_number_layers, cfg_.temporal_layering_mode);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  ASSERT_GE(cfg_.rc_target_bitrate, file_datarate_ * 0.85)
+          << " The datarate for the file exceeds the target by too much!";
+  ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.22)
+      << " The datarate for the file is lower than the target by too much!";
+  EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+}
+
+// Check basic rate targeting for 1 pass CBR SVC: 3 spatial layers and 3
+// temporal layers. Run CIF clip with 1 thread, and few short key frame periods.
+TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc3SpatialLayersSmallKf) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.ss_number_layers = 3;
+  cfg_.ts_number_layers = 3;
+  cfg_.ts_rate_decimator[0] = 4;
+  cfg_.ts_rate_decimator[1] = 2;
+  cfg_.ts_rate_decimator[2] = 1;
+  cfg_.g_error_resilient = 1;
+  cfg_.g_threads = 1;
+  cfg_.temporal_layering_mode = 3;
+  svc_params_.scaling_factor_num[0] = 72;
+  svc_params_.scaling_factor_den[0] = 288;
+  svc_params_.scaling_factor_num[1] = 144;
+  svc_params_.scaling_factor_den[1] = 288;
+  svc_params_.scaling_factor_num[2] = 288;
+  svc_params_.scaling_factor_den[2] = 288;
+  cfg_.rc_dropframe_thresh = 10;
+  ::libvpx_test::I420VideoSource video("niklas_1280_720_30.y4m", 1280, 720,
+                                       30, 1, 0, 300);
+  cfg_.rc_target_bitrate = 800;
+  // For this 3 temporal layer case, pattern repeats every 4 frames, so choose
+  // 4 key neighboring key frame periods (so key frame will land on 0-2-1-2).
+  for (int j = 32; j <= 35; j++) {
+    cfg_.kf_max_dist = j;
+    ResetModel();
+    assign_layer_bitrates(&cfg_, &svc_params_, cfg_.ss_number_layers,
+       cfg_.ts_number_layers, cfg_.temporal_layering_mode);
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(cfg_.rc_target_bitrate, file_datarate_ * 0.85)
+            << " The datarate for the file exceeds the target by too much!";
+    ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.30)
+        << " The datarate for the file is lower than the target by too much!";
+    EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+  }
+}
+
+// Check basic rate targeting for 1 pass CBR SVC: 3 spatial layers and
+// 3 temporal layers. Run HD clip with 4 threads.
+TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc3SpatialLayers4threads) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.ss_number_layers = 3;
+  cfg_.ts_number_layers = 3;
+  cfg_.ts_rate_decimator[0] = 4;
+  cfg_.ts_rate_decimator[1] = 2;
+  cfg_.ts_rate_decimator[2] = 1;
+  cfg_.g_error_resilient = 1;
+  cfg_.g_threads = 4;
+  cfg_.temporal_layering_mode = 3;
+  svc_params_.scaling_factor_num[0] = 72;
+  svc_params_.scaling_factor_den[0] = 288;
+  svc_params_.scaling_factor_num[1] = 144;
+  svc_params_.scaling_factor_den[1] = 288;
+  svc_params_.scaling_factor_num[2] = 288;
+  svc_params_.scaling_factor_den[2] = 288;
+  cfg_.rc_dropframe_thresh = 10;
+  cfg_.kf_max_dist = 9999;
+  ::libvpx_test::I420VideoSource video("niklas_1280_720_30.y4m", 1280, 720,
+                                       30, 1, 0, 300);
+  cfg_.rc_target_bitrate = 800;
+  ResetModel();
+  assign_layer_bitrates(&cfg_, &svc_params_, cfg_.ss_number_layers,
+      cfg_.ts_number_layers, cfg_.temporal_layering_mode);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  ASSERT_GE(cfg_.rc_target_bitrate, file_datarate_ * 0.85)
+          << " The datarate for the file exceeds the target by too much!";
+  ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.22)
+      << " The datarate for the file is lower than the target by too much!";
+  EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+}
+
+VP8_INSTANTIATE_TEST_CASE(DatarateTestLarge, ALL_TEST_MODES);
+VP9_INSTANTIATE_TEST_CASE(DatarateTestVP9Large,
+                          ::testing::Values(::libvpx_test::kOnePassGood,
+                                            ::libvpx_test::kRealTime),
+                          ::testing::Range(2, 9));
+VP9_INSTANTIATE_TEST_CASE(DatarateOnePassCbrSvc,
+                          ::testing::Values(::libvpx_test::kRealTime),
+                          ::testing::Range(5, 9));
+}  // namespace
diff --git a/libvpx/libvpx/test/dct16x16_test.cc b/libvpx/libvpx/test/dct16x16_test.cc
new file mode 100644
index 0000000..ddaf939
--- /dev/null
+++ b/libvpx/libvpx/test/dct16x16_test.cc
@@ -0,0 +1,994 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_scan.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+
+#ifdef _MSC_VER
+static int round(double x) {
+  if (x < 0)
+    return static_cast<int>(ceil(x - 0.5));
+  else
+    return static_cast<int>(floor(x + 0.5));
+}
+#endif
+
+const int kNumCoeffs = 256;
+const double C1 = 0.995184726672197;
+const double C2 = 0.98078528040323;
+const double C3 = 0.956940335732209;
+const double C4 = 0.923879532511287;
+const double C5 = 0.881921264348355;
+const double C6 = 0.831469612302545;
+const double C7 = 0.773010453362737;
+const double C8 = 0.707106781186548;
+const double C9 = 0.634393284163646;
+const double C10 = 0.555570233019602;
+const double C11 = 0.471396736825998;
+const double C12 = 0.38268343236509;
+const double C13 = 0.290284677254462;
+const double C14 = 0.195090322016128;
+const double C15 = 0.098017140329561;
+
+void butterfly_16x16_dct_1d(double input[16], double output[16]) {
+  double step[16];
+  double intermediate[16];
+  double temp1, temp2;
+
+  // step 1
+  step[ 0] = input[0] + input[15];
+  step[ 1] = input[1] + input[14];
+  step[ 2] = input[2] + input[13];
+  step[ 3] = input[3] + input[12];
+  step[ 4] = input[4] + input[11];
+  step[ 5] = input[5] + input[10];
+  step[ 6] = input[6] + input[ 9];
+  step[ 7] = input[7] + input[ 8];
+  step[ 8] = input[7] - input[ 8];
+  step[ 9] = input[6] - input[ 9];
+  step[10] = input[5] - input[10];
+  step[11] = input[4] - input[11];
+  step[12] = input[3] - input[12];
+  step[13] = input[2] - input[13];
+  step[14] = input[1] - input[14];
+  step[15] = input[0] - input[15];
+
+  // step 2
+  output[0] = step[0] + step[7];
+  output[1] = step[1] + step[6];
+  output[2] = step[2] + step[5];
+  output[3] = step[3] + step[4];
+  output[4] = step[3] - step[4];
+  output[5] = step[2] - step[5];
+  output[6] = step[1] - step[6];
+  output[7] = step[0] - step[7];
+
+  temp1 = step[ 8] * C7;
+  temp2 = step[15] * C9;
+  output[ 8] = temp1 + temp2;
+
+  temp1 = step[ 9] * C11;
+  temp2 = step[14] * C5;
+  output[ 9] = temp1 - temp2;
+
+  temp1 = step[10] * C3;
+  temp2 = step[13] * C13;
+  output[10] = temp1 + temp2;
+
+  temp1 = step[11] * C15;
+  temp2 = step[12] * C1;
+  output[11] = temp1 - temp2;
+
+  temp1 = step[11] * C1;
+  temp2 = step[12] * C15;
+  output[12] = temp2 + temp1;
+
+  temp1 = step[10] * C13;
+  temp2 = step[13] * C3;
+  output[13] = temp2 - temp1;
+
+  temp1 = step[ 9] * C5;
+  temp2 = step[14] * C11;
+  output[14] = temp2 + temp1;
+
+  temp1 = step[ 8] * C9;
+  temp2 = step[15] * C7;
+  output[15] = temp2 - temp1;
+
+  // step 3
+  step[ 0] = output[0] + output[3];
+  step[ 1] = output[1] + output[2];
+  step[ 2] = output[1] - output[2];
+  step[ 3] = output[0] - output[3];
+
+  temp1 = output[4] * C14;
+  temp2 = output[7] * C2;
+  step[ 4] = temp1 + temp2;
+
+  temp1 = output[5] * C10;
+  temp2 = output[6] * C6;
+  step[ 5] = temp1 + temp2;
+
+  temp1 = output[5] * C6;
+  temp2 = output[6] * C10;
+  step[ 6] = temp2 - temp1;
+
+  temp1 = output[4] * C2;
+  temp2 = output[7] * C14;
+  step[ 7] = temp2 - temp1;
+
+  step[ 8] = output[ 8] + output[11];
+  step[ 9] = output[ 9] + output[10];
+  step[10] = output[ 9] - output[10];
+  step[11] = output[ 8] - output[11];
+
+  step[12] = output[12] + output[15];
+  step[13] = output[13] + output[14];
+  step[14] = output[13] - output[14];
+  step[15] = output[12] - output[15];
+
+  // step 4
+  output[ 0] = (step[ 0] + step[ 1]);
+  output[ 8] = (step[ 0] - step[ 1]);
+
+  temp1 = step[2] * C12;
+  temp2 = step[3] * C4;
+  temp1 = temp1 + temp2;
+  output[ 4] = 2*(temp1 * C8);
+
+  temp1 = step[2] * C4;
+  temp2 = step[3] * C12;
+  temp1 = temp2 - temp1;
+  output[12] = 2 * (temp1 * C8);
+
+  output[ 2] = 2 * ((step[4] + step[ 5]) * C8);
+  output[14] = 2 * ((step[7] - step[ 6]) * C8);
+
+  temp1 = step[4] - step[5];
+  temp2 = step[6] + step[7];
+  output[ 6] = (temp1 + temp2);
+  output[10] = (temp1 - temp2);
+
+  intermediate[8] = step[8] + step[14];
+  intermediate[9] = step[9] + step[15];
+
+  temp1 = intermediate[8] * C12;
+  temp2 = intermediate[9] * C4;
+  temp1 = temp1 - temp2;
+  output[3] = 2 * (temp1 * C8);
+
+  temp1 = intermediate[8] * C4;
+  temp2 = intermediate[9] * C12;
+  temp1 = temp2 + temp1;
+  output[13] = 2 * (temp1 * C8);
+
+  output[ 9] = 2 * ((step[10] + step[11]) * C8);
+
+  intermediate[11] = step[10] - step[11];
+  intermediate[12] = step[12] + step[13];
+  intermediate[13] = step[12] - step[13];
+  intermediate[14] = step[ 8] - step[14];
+  intermediate[15] = step[ 9] - step[15];
+
+  output[15] = (intermediate[11] + intermediate[12]);
+  output[ 1] = -(intermediate[11] - intermediate[12]);
+
+  output[ 7] = 2 * (intermediate[13] * C8);
+
+  temp1 = intermediate[14] * C12;
+  temp2 = intermediate[15] * C4;
+  temp1 = temp1 - temp2;
+  output[11] = -2 * (temp1 * C8);
+
+  temp1 = intermediate[14] * C4;
+  temp2 = intermediate[15] * C12;
+  temp1 = temp2 + temp1;
+  output[ 5] = 2 * (temp1 * C8);
+}
+
+void reference_16x16_dct_2d(int16_t input[256], double output[256]) {
+  // First transform columns
+  for (int i = 0; i < 16; ++i) {
+    double temp_in[16], temp_out[16];
+    for (int j = 0; j < 16; ++j)
+      temp_in[j] = input[j * 16 + i];
+    butterfly_16x16_dct_1d(temp_in, temp_out);
+    for (int j = 0; j < 16; ++j)
+      output[j * 16 + i] = temp_out[j];
+  }
+  // Then transform rows
+  for (int i = 0; i < 16; ++i) {
+    double temp_in[16], temp_out[16];
+    for (int j = 0; j < 16; ++j)
+      temp_in[j] = output[j + i * 16];
+    butterfly_16x16_dct_1d(temp_in, temp_out);
+    // Scale by some magic number
+    for (int j = 0; j < 16; ++j)
+      output[j + i * 16] = temp_out[j]/2;
+  }
+}
+
+typedef void (*FdctFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*IdctFunc)(const tran_low_t *in, uint8_t *out, int stride);
+typedef void (*FhtFunc)(const int16_t *in, tran_low_t *out, int stride,
+                        int tx_type);
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+
+typedef std::tr1::tuple<FdctFunc, IdctFunc, int, vpx_bit_depth_t> Dct16x16Param;
+typedef std::tr1::tuple<FhtFunc, IhtFunc, int, vpx_bit_depth_t> Ht16x16Param;
+typedef std::tr1::tuple<IdctFunc, IdctFunc, int, vpx_bit_depth_t>
+    Idct16x16Param;
+
+void fdct16x16_ref(const int16_t *in, tran_low_t *out, int stride,
+                   int /*tx_type*/) {
+  vpx_fdct16x16_c(in, out, stride);
+}
+
+void idct16x16_ref(const tran_low_t *in, uint8_t *dest, int stride,
+                   int /*tx_type*/) {
+  vpx_idct16x16_256_add_c(in, dest, stride);
+}
+
+void fht16x16_ref(const int16_t *in, tran_low_t *out, int stride,
+                  int tx_type) {
+  vp9_fht16x16_c(in, out, stride, tx_type);
+}
+
+void iht16x16_ref(const tran_low_t *in, uint8_t *dest, int stride,
+                  int tx_type) {
+  vp9_iht16x16_256_add_c(in, dest, stride, tx_type);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void idct16x16_10(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct16x16_256_add_c(in, out, stride, 10);
+}
+
+void idct16x16_12(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct16x16_256_add_c(in, out, stride, 12);
+}
+
+void idct16x16_10_ref(const tran_low_t *in, uint8_t *out, int stride,
+                      int /*tx_type*/) {
+  idct16x16_10(in, out, stride);
+}
+
+void idct16x16_12_ref(const tran_low_t *in, uint8_t *out, int stride,
+                      int /*tx_type*/) {
+  idct16x16_12(in, out, stride);
+}
+
+void iht16x16_10(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  vp9_highbd_iht16x16_256_add_c(in, out, stride, tx_type, 10);
+}
+
+void iht16x16_12(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  vp9_highbd_iht16x16_256_add_c(in, out, stride, tx_type, 12);
+}
+
+#if HAVE_SSE2
+void idct16x16_10_add_10_c(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct16x16_10_add_c(in, out, stride, 10);
+}
+
+void idct16x16_10_add_12_c(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct16x16_10_add_c(in, out, stride, 12);
+}
+
+void idct16x16_256_add_10_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct16x16_256_add_sse2(in, out, stride, 10);
+}
+
+void idct16x16_256_add_12_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct16x16_256_add_sse2(in, out, stride, 12);
+}
+
+void idct16x16_10_add_10_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct16x16_10_add_sse2(in, out, stride, 10);
+}
+
+void idct16x16_10_add_12_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct16x16_10_add_sse2(in, out, stride, 12);
+}
+#endif  // HAVE_SSE2
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+class Trans16x16TestBase {
+ public:
+  virtual ~Trans16x16TestBase() {}
+
+ protected:
+  virtual void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) = 0;
+
+  virtual void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) = 0;
+
+  void RunAccuracyCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    uint32_t max_error = 0;
+    int64_t total_error = 0;
+    const int count_test_block = 10000;
+    for (int i = 0; i < count_test_block; ++i) {
+      DECLARE_ALIGNED(16, int16_t, test_input_block[kNumCoeffs]);
+      DECLARE_ALIGNED(16, tran_low_t, test_temp_block[kNumCoeffs]);
+      DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+      DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_VP9_HIGHBITDEPTH
+      DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+      DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+#endif
+
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (bit_depth_ == VPX_BITS_8) {
+          src[j] = rnd.Rand8();
+          dst[j] = rnd.Rand8();
+          test_input_block[j] = src[j] - dst[j];
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand16() & mask_;
+          dst16[j] = rnd.Rand16() & mask_;
+          test_input_block[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(test_input_block,
+                                          test_temp_block, pitch_));
+      if (bit_depth_ == VPX_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(test_temp_block, dst, pitch_));
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(test_temp_block, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        const int32_t diff =
+            bit_depth_ == VPX_BITS_8 ?  dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int32_t diff = dst[j] - src[j];
+#endif
+        const uint32_t error = diff * diff;
+        if (max_error < error)
+          max_error = error;
+        total_error += error;
+      }
+    }
+
+    EXPECT_GE(1u  << 2 * (bit_depth_ - 8), max_error)
+        << "Error: 16x16 FHT/IHT has an individual round trip error > 1";
+
+    EXPECT_GE(count_test_block << 2 * (bit_depth_ - 8), total_error)
+        << "Error: 16x16 FHT/IHT has average round trip error > 1 per block";
+  }
+
+  void RunCoeffCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, input_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]);
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j)
+        input_block[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+
+      fwd_txfm_ref(input_block, output_ref_block, pitch_, tx_type_);
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(input_block, output_block, pitch_));
+
+      // The minimum quant value is 4.
+      for (int j = 0; j < kNumCoeffs; ++j)
+        EXPECT_EQ(output_block[j], output_ref_block[j]);
+    }
+  }
+
+  void RunMemCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, input_extreme_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]);
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        input_extreme_block[j] = rnd.Rand8() % 2 ? mask_ : -mask_;
+      }
+      if (i == 0) {
+        for (int j = 0; j < kNumCoeffs; ++j)
+          input_extreme_block[j] = mask_;
+      } else if (i == 1) {
+        for (int j = 0; j < kNumCoeffs; ++j)
+          input_extreme_block[j] = -mask_;
+      }
+
+      fwd_txfm_ref(input_extreme_block, output_ref_block, pitch_, tx_type_);
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(input_extreme_block,
+                                          output_block, pitch_));
+
+      // The minimum quant value is 4.
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        EXPECT_EQ(output_block[j], output_ref_block[j]);
+        EXPECT_GE(4 * DCT_MAX_VALUE << (bit_depth_ - 8), abs(output_block[j]))
+            << "Error: 16x16 FDCT has coefficient larger than 4*DCT_MAX_VALUE";
+      }
+    }
+  }
+
+  void RunQuantCheck(int dc_thred, int ac_thred) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 100000;
+    DECLARE_ALIGNED(16, int16_t, input_extreme_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, ref[kNumCoeffs]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, ref16[kNumCoeffs]);
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        input_extreme_block[j] = rnd.Rand8() % 2 ? mask_ : -mask_;
+      }
+      if (i == 0)
+        for (int j = 0; j < kNumCoeffs; ++j)
+          input_extreme_block[j] = mask_;
+      if (i == 1)
+        for (int j = 0; j < kNumCoeffs; ++j)
+          input_extreme_block[j] = -mask_;
+
+      fwd_txfm_ref(input_extreme_block, output_ref_block, pitch_, tx_type_);
+
+      // clear reconstructed pixel buffers
+      memset(dst, 0, kNumCoeffs * sizeof(uint8_t));
+      memset(ref, 0, kNumCoeffs * sizeof(uint8_t));
+#if CONFIG_VP9_HIGHBITDEPTH
+      memset(dst16, 0, kNumCoeffs * sizeof(uint16_t));
+      memset(ref16, 0, kNumCoeffs * sizeof(uint16_t));
+#endif
+
+      // quantization with maximum allowed step sizes
+      output_ref_block[0] = (output_ref_block[0] / dc_thred) * dc_thred;
+      for (int j = 1; j < kNumCoeffs; ++j)
+        output_ref_block[j] = (output_ref_block[j] / ac_thred) * ac_thred;
+      if (bit_depth_ == VPX_BITS_8) {
+        inv_txfm_ref(output_ref_block, ref, pitch_, tx_type_);
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(output_ref_block, dst, pitch_));
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        inv_txfm_ref(output_ref_block, CONVERT_TO_BYTEPTR(ref16), pitch_,
+                     tx_type_);
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(output_ref_block,
+                                            CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+      if (bit_depth_ == VPX_BITS_8) {
+        for (int j = 0; j < kNumCoeffs; ++j)
+          EXPECT_EQ(ref[j], dst[j]);
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        for (int j = 0; j < kNumCoeffs; ++j)
+          EXPECT_EQ(ref16[j], dst16[j]);
+#endif
+      }
+    }
+  }
+
+  void RunInvAccuracyCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, in[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    for (int i = 0; i < count_test_block; ++i) {
+      double out_r[kNumCoeffs];
+
+      // Initialize a test block with input range [-255, 255].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (bit_depth_ == VPX_BITS_8) {
+          src[j] = rnd.Rand8();
+          dst[j] = rnd.Rand8();
+          in[j] = src[j] - dst[j];
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand16() & mask_;
+          dst16[j] = rnd.Rand16() & mask_;
+          in[j] = src16[j] - dst16[j];
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        }
+      }
+
+      reference_16x16_dct_2d(in, out_r);
+      for (int j = 0; j < kNumCoeffs; ++j)
+        coeff[j] = static_cast<tran_low_t>(round(out_r[j]));
+
+      if (bit_depth_ == VPX_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, 16));
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, CONVERT_TO_BYTEPTR(dst16),
+                                            16));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        const uint32_t diff =
+            bit_depth_ == VPX_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const uint32_t diff = dst[j] - src[j];
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        const uint32_t error = diff * diff;
+        EXPECT_GE(1u, error)
+            << "Error: 16x16 IDCT has error " << error
+            << " at index " << j;
+      }
+    }
+  }
+
+  void CompareInvReference(IdctFunc ref_txfm, int thresh) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 10000;
+    const int eob = 10;
+    const int16_t *scan = vp9_default_scan_orders[TX_16X16].scan;
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, ref[kNumCoeffs]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, ref16[kNumCoeffs]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    for (int i = 0; i < count_test_block; ++i) {
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (j < eob) {
+          // Random values less than the threshold, either positive or negative
+          coeff[scan[j]] = rnd(thresh) * (1 - 2 * (i % 2));
+        } else {
+          coeff[scan[j]] = 0;
+        }
+        if (bit_depth_ == VPX_BITS_8) {
+          dst[j] = 0;
+          ref[j] = 0;
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          dst16[j] = 0;
+          ref16[j] = 0;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        }
+      }
+      if (bit_depth_ == VPX_BITS_8) {
+        ref_txfm(coeff, ref, pitch_);
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, pitch_));
+      } else {
+#if CONFIG_VP9_HIGHBITDEPTH
+        ref_txfm(coeff, CONVERT_TO_BYTEPTR(ref16), pitch_);
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, CONVERT_TO_BYTEPTR(dst16),
+                                 pitch_));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        const uint32_t diff =
+            bit_depth_ == VPX_BITS_8 ? dst[j] - ref[j] : dst16[j] - ref16[j];
+#else
+        const uint32_t diff = dst[j] - ref[j];
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        const uint32_t error = diff * diff;
+        EXPECT_EQ(0u, error)
+            << "Error: 16x16 IDCT Comparison has error " << error
+            << " at index " << j;
+      }
+    }
+  }
+
+  int pitch_;
+  int tx_type_;
+  vpx_bit_depth_t bit_depth_;
+  int mask_;
+  FhtFunc fwd_txfm_ref;
+  IhtFunc inv_txfm_ref;
+};
+
+class Trans16x16DCT
+    : public Trans16x16TestBase,
+      public ::testing::TestWithParam<Dct16x16Param> {
+ public:
+  virtual ~Trans16x16DCT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_  = GET_PARAM(2);
+    bit_depth_ = GET_PARAM(3);
+    pitch_    = 16;
+    fwd_txfm_ref = fdct16x16_ref;
+    inv_txfm_ref = idct16x16_ref;
+    mask_ = (1 << bit_depth_) - 1;
+#if CONFIG_VP9_HIGHBITDEPTH
+    switch (bit_depth_) {
+      case VPX_BITS_10:
+        inv_txfm_ref = idct16x16_10_ref;
+        break;
+      case VPX_BITS_12:
+        inv_txfm_ref = idct16x16_12_ref;
+        break;
+      default:
+        inv_txfm_ref = idct16x16_ref;
+        break;
+    }
+#else
+    inv_txfm_ref = idct16x16_ref;
+#endif
+  }
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride);
+  }
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
+};
+
+TEST_P(Trans16x16DCT, AccuracyCheck) {
+  RunAccuracyCheck();
+}
+
+TEST_P(Trans16x16DCT, CoeffCheck) {
+  RunCoeffCheck();
+}
+
+TEST_P(Trans16x16DCT, MemCheck) {
+  RunMemCheck();
+}
+
+TEST_P(Trans16x16DCT, QuantCheck) {
+  // Use maximally allowed quantization step sizes for DC and AC
+  // coefficients respectively.
+  RunQuantCheck(1336, 1828);
+}
+
+TEST_P(Trans16x16DCT, InvAccuracyCheck) {
+  RunInvAccuracyCheck();
+}
+
+class Trans16x16HT
+    : public Trans16x16TestBase,
+      public ::testing::TestWithParam<Ht16x16Param> {
+ public:
+  virtual ~Trans16x16HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_  = GET_PARAM(2);
+    bit_depth_ = GET_PARAM(3);
+    pitch_    = 16;
+    fwd_txfm_ref = fht16x16_ref;
+    inv_txfm_ref = iht16x16_ref;
+    mask_ = (1 << bit_depth_) - 1;
+#if CONFIG_VP9_HIGHBITDEPTH
+    switch (bit_depth_) {
+      case VPX_BITS_10:
+        inv_txfm_ref = iht16x16_10;
+        break;
+      case VPX_BITS_12:
+        inv_txfm_ref = iht16x16_12;
+        break;
+      default:
+        inv_txfm_ref = iht16x16_ref;
+        break;
+    }
+#else
+    inv_txfm_ref = iht16x16_ref;
+#endif
+  }
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(Trans16x16HT, AccuracyCheck) {
+  RunAccuracyCheck();
+}
+
+TEST_P(Trans16x16HT, CoeffCheck) {
+  RunCoeffCheck();
+}
+
+TEST_P(Trans16x16HT, MemCheck) {
+  RunMemCheck();
+}
+
+TEST_P(Trans16x16HT, QuantCheck) {
+  // The encoder skips any non-DC intra prediction modes,
+  // when the quantization step size goes beyond 988.
+  RunQuantCheck(429, 729);
+}
+
+class InvTrans16x16DCT
+    : public Trans16x16TestBase,
+      public ::testing::TestWithParam<Idct16x16Param> {
+ public:
+  virtual ~InvTrans16x16DCT() {}
+
+  virtual void SetUp() {
+    ref_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    thresh_ = GET_PARAM(2);
+    bit_depth_ = GET_PARAM(3);
+    pitch_ = 16;
+    mask_ = (1 << bit_depth_) - 1;
+}
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(int16_t * /*in*/, tran_low_t * /*out*/, int /*stride*/) {}
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  IdctFunc ref_txfm_;
+  IdctFunc inv_txfm_;
+  int thresh_;
+};
+
+TEST_P(InvTrans16x16DCT, CompareReference) {
+  CompareInvReference(ref_txfm_, thresh_);
+}
+
+class PartialTrans16x16Test
+    : public ::testing::TestWithParam<
+          std::tr1::tuple<FdctFunc, vpx_bit_depth_t> > {
+ public:
+  virtual ~PartialTrans16x16Test() {}
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    bit_depth_ = GET_PARAM(1);
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  vpx_bit_depth_t bit_depth_;
+  FdctFunc fwd_txfm_;
+};
+
+TEST_P(PartialTrans16x16Test, Extremes) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int16_t maxval =
+      static_cast<int16_t>(clip_pixel_highbd(1 << 30, bit_depth_));
+#else
+  const int16_t maxval = 255;
+#endif
+  const int minval = -maxval;
+  DECLARE_ALIGNED(16, int16_t, input[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output[kNumCoeffs]);
+
+  for (int i = 0; i < kNumCoeffs; ++i) input[i] = maxval;
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 16));
+  EXPECT_EQ((maxval * kNumCoeffs) >> 1, output[0]);
+
+  for (int i = 0; i < kNumCoeffs; ++i) input[i] = minval;
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 16));
+  EXPECT_EQ((minval * kNumCoeffs) >> 1, output[0]);
+}
+
+TEST_P(PartialTrans16x16Test, Random) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int16_t maxval =
+      static_cast<int16_t>(clip_pixel_highbd(1 << 30, bit_depth_));
+#else
+  const int16_t maxval = 255;
+#endif
+  DECLARE_ALIGNED(16, int16_t, input[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output[kNumCoeffs]);
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+  int sum = 0;
+  for (int i = 0; i < kNumCoeffs; ++i) {
+    const int val = (i & 1) ? -rnd(maxval + 1) : rnd(maxval + 1);
+    input[i] = val;
+    sum += val;
+  }
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 16));
+  EXPECT_EQ(sum >> 1, output[0]);
+}
+
+using std::tr1::make_tuple;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans16x16DCT,
+    ::testing::Values(
+        make_tuple(&vpx_highbd_fdct16x16_c, &idct16x16_10, 0, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct16x16_c, &idct16x16_12, 0, VPX_BITS_12),
+        make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c, 0, VPX_BITS_8)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, Trans16x16DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct16x16_c, &vpx_idct16x16_256_add_c, 0, VPX_BITS_8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&vp9_highbd_fht16x16_c, &iht16x16_10, 0, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht16x16_c, &iht16x16_10, 1, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht16x16_c, &iht16x16_10, 2, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht16x16_c, &iht16x16_10, 3, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht16x16_c, &iht16x16_12, 0, VPX_BITS_12),
+        make_tuple(&vp9_highbd_fht16x16_c, &iht16x16_12, 1, VPX_BITS_12),
+        make_tuple(&vp9_highbd_fht16x16_c, &iht16x16_12, 2, VPX_BITS_12),
+        make_tuple(&vp9_highbd_fht16x16_c, &iht16x16_12, 3, VPX_BITS_12),
+        make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 3, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    C, PartialTrans16x16Test,
+    ::testing::Values(make_tuple(&vpx_highbd_fdct16x16_1_c, VPX_BITS_8),
+                      make_tuple(&vpx_highbd_fdct16x16_1_c, VPX_BITS_10),
+                      make_tuple(&vpx_highbd_fdct16x16_1_c, VPX_BITS_12)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_c, &vp9_iht16x16_256_add_c, 3, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(C, PartialTrans16x16Test,
+                        ::testing::Values(make_tuple(&vpx_fdct16x16_1_c,
+                                                     VPX_BITS_8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if HAVE_NEON_ASM && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    NEON, Trans16x16DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct16x16_c,
+                   &vpx_idct16x16_256_add_neon, 0, VPX_BITS_8)));
+#endif
+
+#if HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans16x16DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct16x16_sse2,
+                   &vpx_idct16x16_256_add_sse2, 0, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht16x16_sse2, &vp9_iht16x16_256_add_sse2, 0,
+                   VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_sse2, &vp9_iht16x16_256_add_sse2, 1,
+                   VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_sse2, &vp9_iht16x16_256_add_sse2, 2,
+                   VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_sse2, &vp9_iht16x16_256_add_sse2, 3,
+                   VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(SSE2, PartialTrans16x16Test,
+                        ::testing::Values(make_tuple(&vpx_fdct16x16_1_sse2,
+                                                     VPX_BITS_8)));
+#endif  // HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans16x16DCT,
+    ::testing::Values(
+        make_tuple(&vpx_highbd_fdct16x16_sse2,
+                   &idct16x16_10, 0, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct16x16_c,
+                   &idct16x16_256_add_10_sse2, 0, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct16x16_sse2,
+                   &idct16x16_12, 0, VPX_BITS_12),
+        make_tuple(&vpx_highbd_fdct16x16_c,
+                   &idct16x16_256_add_12_sse2, 0, VPX_BITS_12),
+        make_tuple(&vpx_fdct16x16_sse2,
+                   &vpx_idct16x16_256_add_c, 0, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht16x16_sse2, &vp9_iht16x16_256_add_c, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_sse2, &vp9_iht16x16_256_add_c, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_sse2, &vp9_iht16x16_256_add_c, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_sse2, &vp9_iht16x16_256_add_c, 3,
+                   VPX_BITS_8)));
+// Optimizations take effect at a threshold of 3155, so we use a value close to
+// that to test both branches.
+INSTANTIATE_TEST_CASE_P(
+    SSE2, InvTrans16x16DCT,
+    ::testing::Values(
+        make_tuple(&idct16x16_10_add_10_c,
+                   &idct16x16_10_add_10_sse2, 3167, VPX_BITS_10),
+        make_tuple(&idct16x16_10,
+                   &idct16x16_256_add_10_sse2, 3167, VPX_BITS_10),
+        make_tuple(&idct16x16_10_add_12_c,
+                   &idct16x16_10_add_12_sse2, 3167, VPX_BITS_12),
+        make_tuple(&idct16x16_12,
+                   &idct16x16_256_add_12_sse2, 3167, VPX_BITS_12)));
+INSTANTIATE_TEST_CASE_P(SSE2, PartialTrans16x16Test,
+                        ::testing::Values(make_tuple(&vpx_fdct16x16_1_sse2,
+                                                     VPX_BITS_8)));
+#endif  // HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    MSA, Trans16x16DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct16x16_msa,
+                   &vpx_idct16x16_256_add_msa, 0, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    MSA, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht16x16_msa, &vp9_iht16x16_256_add_msa, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_msa, &vp9_iht16x16_256_add_msa, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_msa, &vp9_iht16x16_256_add_msa, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht16x16_msa, &vp9_iht16x16_256_add_msa, 3,
+                   VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(MSA, PartialTrans16x16Test,
+                        ::testing::Values(make_tuple(&vpx_fdct16x16_1_msa,
+                                                     VPX_BITS_8)));
+#endif  // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+}  // namespace
diff --git a/libvpx/libvpx/test/dct32x32_test.cc b/libvpx/libvpx/test/dct32x32_test.cc
new file mode 100644
index 0000000..16d8825
--- /dev/null
+++ b/libvpx/libvpx/test/dct32x32_test.cc
@@ -0,0 +1,469 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+#ifdef _MSC_VER
+static int round(double x) {
+  if (x < 0)
+    return static_cast<int>(ceil(x - 0.5));
+  else
+    return static_cast<int>(floor(x + 0.5));
+}
+#endif
+
+const int kNumCoeffs = 1024;
+const double kPi = 3.141592653589793238462643383279502884;
+void reference_32x32_dct_1d(const double in[32], double out[32]) {
+  const double kInvSqrt2 = 0.707106781186547524400844362104;
+  for (int k = 0; k < 32; k++) {
+    out[k] = 0.0;
+    for (int n = 0; n < 32; n++)
+      out[k] += in[n] * cos(kPi * (2 * n + 1) * k / 64.0);
+    if (k == 0)
+      out[k] = out[k] * kInvSqrt2;
+  }
+}
+
+void reference_32x32_dct_2d(const int16_t input[kNumCoeffs],
+                            double output[kNumCoeffs]) {
+  // First transform columns
+  for (int i = 0; i < 32; ++i) {
+    double temp_in[32], temp_out[32];
+    for (int j = 0; j < 32; ++j)
+      temp_in[j] = input[j*32 + i];
+    reference_32x32_dct_1d(temp_in, temp_out);
+    for (int j = 0; j < 32; ++j)
+      output[j * 32 + i] = temp_out[j];
+  }
+  // Then transform rows
+  for (int i = 0; i < 32; ++i) {
+    double temp_in[32], temp_out[32];
+    for (int j = 0; j < 32; ++j)
+      temp_in[j] = output[j + i*32];
+    reference_32x32_dct_1d(temp_in, temp_out);
+    // Scale by some magic number
+    for (int j = 0; j < 32; ++j)
+      output[j + i * 32] = temp_out[j] / 4;
+  }
+}
+
+typedef void (*FwdTxfmFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*InvTxfmFunc)(const tran_low_t *in, uint8_t *out, int stride);
+
+typedef std::tr1::tuple<FwdTxfmFunc, InvTxfmFunc, int, vpx_bit_depth_t>
+    Trans32x32Param;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void idct32x32_10(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct32x32_1024_add_c(in, out, stride, 10);
+}
+
+void idct32x32_12(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct32x32_1024_add_c(in, out, stride, 12);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+class Trans32x32Test : public ::testing::TestWithParam<Trans32x32Param> {
+ public:
+  virtual ~Trans32x32Test() {}
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    version_  = GET_PARAM(2);  // 0: high precision forward transform
+                               // 1: low precision version for rd loop
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  int version_;
+  vpx_bit_depth_t bit_depth_;
+  int mask_;
+  FwdTxfmFunc fwd_txfm_;
+  InvTxfmFunc inv_txfm_;
+};
+
+TEST_P(Trans32x32Test, AccuracyCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  uint32_t max_error = 0;
+  int64_t total_error = 0;
+  const int count_test_block = 10000;
+  DECLARE_ALIGNED(16, int16_t, test_input_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, test_temp_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_VP9_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+#endif
+
+  for (int i = 0; i < count_test_block; ++i) {
+    // Initialize a test block with input range [-mask_, mask_].
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      if (bit_depth_ == VPX_BITS_8) {
+        src[j] = rnd.Rand8();
+        dst[j] = rnd.Rand8();
+        test_input_block[j] = src[j] - dst[j];
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        src16[j] = rnd.Rand16() & mask_;
+        dst16[j] = rnd.Rand16() & mask_;
+        test_input_block[j] = src16[j] - dst16[j];
+#endif
+      }
+    }
+
+    ASM_REGISTER_STATE_CHECK(fwd_txfm_(test_input_block, test_temp_block, 32));
+    if (bit_depth_ == VPX_BITS_8) {
+      ASM_REGISTER_STATE_CHECK(inv_txfm_(test_temp_block, dst, 32));
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else {
+      ASM_REGISTER_STATE_CHECK(inv_txfm_(test_temp_block,
+                                         CONVERT_TO_BYTEPTR(dst16), 32));
+#endif
+    }
+
+    for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+      const int32_t diff =
+          bit_depth_ == VPX_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+      const int32_t diff = dst[j] - src[j];
+#endif
+      const uint32_t error = diff * diff;
+      if (max_error < error)
+        max_error = error;
+      total_error += error;
+    }
+  }
+
+  if (version_ == 1) {
+    max_error /= 2;
+    total_error /= 45;
+  }
+
+  EXPECT_GE(1u << 2 * (bit_depth_ - 8), max_error)
+      << "Error: 32x32 FDCT/IDCT has an individual round-trip error > 1";
+
+  EXPECT_GE(count_test_block << 2 * (bit_depth_ - 8), total_error)
+      << "Error: 32x32 FDCT/IDCT has average round-trip error > 1 per block";
+}
+
+TEST_P(Trans32x32Test, CoeffCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 1000;
+
+  DECLARE_ALIGNED(16, int16_t, input_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]);
+
+  for (int i = 0; i < count_test_block; ++i) {
+    for (int j = 0; j < kNumCoeffs; ++j)
+      input_block[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+
+    const int stride = 32;
+    vpx_fdct32x32_c(input_block, output_ref_block, stride);
+    ASM_REGISTER_STATE_CHECK(fwd_txfm_(input_block, output_block, stride));
+
+    if (version_ == 0) {
+      for (int j = 0; j < kNumCoeffs; ++j)
+        EXPECT_EQ(output_block[j], output_ref_block[j])
+            << "Error: 32x32 FDCT versions have mismatched coefficients";
+    } else {
+      for (int j = 0; j < kNumCoeffs; ++j)
+        EXPECT_GE(6, abs(output_block[j] - output_ref_block[j]))
+            << "Error: 32x32 FDCT rd has mismatched coefficients";
+    }
+  }
+}
+
+TEST_P(Trans32x32Test, MemCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 2000;
+
+  DECLARE_ALIGNED(16, int16_t, input_extreme_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]);
+
+  for (int i = 0; i < count_test_block; ++i) {
+    // Initialize a test block with input range [-mask_, mask_].
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      input_extreme_block[j] = rnd.Rand8() & 1 ? mask_ : -mask_;
+    }
+    if (i == 0) {
+      for (int j = 0; j < kNumCoeffs; ++j)
+        input_extreme_block[j] = mask_;
+    } else if (i == 1) {
+      for (int j = 0; j < kNumCoeffs; ++j)
+        input_extreme_block[j] = -mask_;
+    }
+
+    const int stride = 32;
+    vpx_fdct32x32_c(input_extreme_block, output_ref_block, stride);
+    ASM_REGISTER_STATE_CHECK(
+        fwd_txfm_(input_extreme_block, output_block, stride));
+
+    // The minimum quant value is 4.
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      if (version_ == 0) {
+        EXPECT_EQ(output_block[j], output_ref_block[j])
+            << "Error: 32x32 FDCT versions have mismatched coefficients";
+      } else {
+        EXPECT_GE(6, abs(output_block[j] - output_ref_block[j]))
+            << "Error: 32x32 FDCT rd has mismatched coefficients";
+      }
+      EXPECT_GE(4 * DCT_MAX_VALUE << (bit_depth_ - 8), abs(output_ref_block[j]))
+          << "Error: 32x32 FDCT C has coefficient larger than 4*DCT_MAX_VALUE";
+      EXPECT_GE(4 * DCT_MAX_VALUE << (bit_depth_ - 8), abs(output_block[j]))
+          << "Error: 32x32 FDCT has coefficient larger than "
+          << "4*DCT_MAX_VALUE";
+    }
+  }
+}
+
+TEST_P(Trans32x32Test, InverseAccuracy) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 1000;
+  DECLARE_ALIGNED(16, int16_t, in[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_VP9_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+#endif
+
+  for (int i = 0; i < count_test_block; ++i) {
+    double out_r[kNumCoeffs];
+
+    // Initialize a test block with input range [-255, 255]
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      if (bit_depth_ == VPX_BITS_8) {
+        src[j] = rnd.Rand8();
+        dst[j] = rnd.Rand8();
+        in[j] = src[j] - dst[j];
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        src16[j] = rnd.Rand16() & mask_;
+        dst16[j] = rnd.Rand16() & mask_;
+        in[j] = src16[j] - dst16[j];
+#endif
+      }
+    }
+
+    reference_32x32_dct_2d(in, out_r);
+    for (int j = 0; j < kNumCoeffs; ++j)
+      coeff[j] = static_cast<tran_low_t>(round(out_r[j]));
+    if (bit_depth_ == VPX_BITS_8) {
+      ASM_REGISTER_STATE_CHECK(inv_txfm_(coeff, dst, 32));
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else {
+      ASM_REGISTER_STATE_CHECK(inv_txfm_(coeff, CONVERT_TO_BYTEPTR(dst16), 32));
+#endif
+    }
+    for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+      const int diff =
+          bit_depth_ == VPX_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+      const int diff = dst[j] - src[j];
+#endif
+      const int error = diff * diff;
+      EXPECT_GE(1, error)
+          << "Error: 32x32 IDCT has error " << error
+          << " at index " << j;
+    }
+  }
+}
+
+class PartialTrans32x32Test
+    : public ::testing::TestWithParam<
+          std::tr1::tuple<FwdTxfmFunc, vpx_bit_depth_t> > {
+ public:
+  virtual ~PartialTrans32x32Test() {}
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    bit_depth_ = GET_PARAM(1);
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  vpx_bit_depth_t bit_depth_;
+  FwdTxfmFunc fwd_txfm_;
+};
+
+TEST_P(PartialTrans32x32Test, Extremes) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int16_t maxval =
+      static_cast<int16_t>(clip_pixel_highbd(1 << 30, bit_depth_));
+#else
+  const int16_t maxval = 255;
+#endif
+  const int minval = -maxval;
+  DECLARE_ALIGNED(16, int16_t, input[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output[kNumCoeffs]);
+
+  for (int i = 0; i < kNumCoeffs; ++i) input[i] = maxval;
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 32));
+  EXPECT_EQ((maxval * kNumCoeffs) >> 3, output[0]);
+
+  for (int i = 0; i < kNumCoeffs; ++i) input[i] = minval;
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 32));
+  EXPECT_EQ((minval * kNumCoeffs) >> 3, output[0]);
+}
+
+TEST_P(PartialTrans32x32Test, Random) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int16_t maxval =
+      static_cast<int16_t>(clip_pixel_highbd(1 << 30, bit_depth_));
+#else
+  const int16_t maxval = 255;
+#endif
+  DECLARE_ALIGNED(16, int16_t, input[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output[kNumCoeffs]);
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+  int sum = 0;
+  for (int i = 0; i < kNumCoeffs; ++i) {
+    const int val = (i & 1) ? -rnd(maxval + 1) : rnd(maxval + 1);
+    input[i] = val;
+    sum += val;
+  }
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 32));
+  EXPECT_EQ(sum >> 3, output[0]);
+}
+
+using std::tr1::make_tuple;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans32x32Test,
+    ::testing::Values(
+        make_tuple(&vpx_highbd_fdct32x32_c,
+                   &idct32x32_10, 0, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct32x32_rd_c,
+                   &idct32x32_10, 1, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct32x32_c,
+                   &idct32x32_12, 0, VPX_BITS_12),
+        make_tuple(&vpx_highbd_fdct32x32_rd_c,
+                   &idct32x32_12, 1, VPX_BITS_12),
+        make_tuple(&vpx_fdct32x32_c,
+                   &vpx_idct32x32_1024_add_c, 0, VPX_BITS_8),
+        make_tuple(&vpx_fdct32x32_rd_c,
+                   &vpx_idct32x32_1024_add_c, 1, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    C, PartialTrans32x32Test,
+    ::testing::Values(make_tuple(&vpx_highbd_fdct32x32_1_c, VPX_BITS_8),
+                      make_tuple(&vpx_highbd_fdct32x32_1_c, VPX_BITS_10),
+                      make_tuple(&vpx_highbd_fdct32x32_1_c, VPX_BITS_12)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, Trans32x32Test,
+    ::testing::Values(
+        make_tuple(&vpx_fdct32x32_c,
+                   &vpx_idct32x32_1024_add_c, 0, VPX_BITS_8),
+        make_tuple(&vpx_fdct32x32_rd_c,
+                   &vpx_idct32x32_1024_add_c, 1, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(C, PartialTrans32x32Test,
+                        ::testing::Values(make_tuple(&vpx_fdct32x32_1_c,
+                                                     VPX_BITS_8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    NEON, Trans32x32Test,
+    ::testing::Values(
+        make_tuple(&vpx_fdct32x32_c,
+                   &vpx_idct32x32_1024_add_neon, 0, VPX_BITS_8),
+        make_tuple(&vpx_fdct32x32_rd_c,
+                   &vpx_idct32x32_1024_add_neon, 1, VPX_BITS_8)));
+#endif  // HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans32x32Test,
+    ::testing::Values(
+        make_tuple(&vpx_fdct32x32_sse2,
+                   &vpx_idct32x32_1024_add_sse2, 0, VPX_BITS_8),
+        make_tuple(&vpx_fdct32x32_rd_sse2,
+                   &vpx_idct32x32_1024_add_sse2, 1, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(SSE2, PartialTrans32x32Test,
+                        ::testing::Values(make_tuple(&vpx_fdct32x32_1_sse2,
+                                                     VPX_BITS_8)));
+#endif  // HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans32x32Test,
+    ::testing::Values(
+        make_tuple(&vpx_highbd_fdct32x32_sse2, &idct32x32_10, 0, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct32x32_rd_sse2, &idct32x32_10, 1,
+                   VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct32x32_sse2, &idct32x32_12, 0, VPX_BITS_12),
+        make_tuple(&vpx_highbd_fdct32x32_rd_sse2, &idct32x32_12, 1,
+                   VPX_BITS_12),
+        make_tuple(&vpx_fdct32x32_sse2, &vpx_idct32x32_1024_add_c, 0,
+                   VPX_BITS_8),
+        make_tuple(&vpx_fdct32x32_rd_sse2, &vpx_idct32x32_1024_add_c, 1,
+                   VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(SSE2, PartialTrans32x32Test,
+                        ::testing::Values(make_tuple(&vpx_fdct32x32_1_sse2,
+                                                     VPX_BITS_8)));
+#endif  // HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_AVX2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    AVX2, Trans32x32Test,
+    ::testing::Values(
+        make_tuple(&vpx_fdct32x32_avx2,
+                   &vpx_idct32x32_1024_add_sse2, 0, VPX_BITS_8),
+        make_tuple(&vpx_fdct32x32_rd_avx2,
+                   &vpx_idct32x32_1024_add_sse2, 1, VPX_BITS_8)));
+#endif  // HAVE_AVX2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    MSA, Trans32x32Test,
+    ::testing::Values(
+        make_tuple(&vpx_fdct32x32_msa,
+                   &vpx_idct32x32_1024_add_msa, 0, VPX_BITS_8),
+        make_tuple(&vpx_fdct32x32_rd_msa,
+                   &vpx_idct32x32_1024_add_msa, 1, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(MSA, PartialTrans32x32Test,
+                        ::testing::Values(make_tuple(&vpx_fdct32x32_1_msa,
+                                                     VPX_BITS_8)));
+#endif  // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+}  // namespace
diff --git a/libvpx/libvpx/test/decode_api_test.cc b/libvpx/libvpx/test/decode_api_test.cc
new file mode 100644
index 0000000..99b4db1
--- /dev/null
+++ b/libvpx/libvpx/test/decode_api_test.cc
@@ -0,0 +1,182 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "test/ivf_video_source.h"
+#include "vpx/vp8dx.h"
+#include "vpx/vpx_decoder.h"
+
+namespace {
+
+#define NELEMENTS(x) static_cast<int>(sizeof(x) / sizeof(x[0]))
+
+TEST(DecodeAPI, InvalidParams) {
+  static const vpx_codec_iface_t *kCodecs[] = {
+#if CONFIG_VP8_DECODER
+    &vpx_codec_vp8_dx_algo,
+#endif
+#if CONFIG_VP9_DECODER
+    &vpx_codec_vp9_dx_algo,
+#endif
+  };
+  uint8_t buf[1] = {0};
+  vpx_codec_ctx_t dec;
+
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_dec_init(NULL, NULL, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_dec_init(&dec, NULL, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_decode(NULL, NULL, 0, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_decode(NULL, buf, 0, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+            vpx_codec_decode(NULL, buf, NELEMENTS(buf), NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+            vpx_codec_decode(NULL, NULL, NELEMENTS(buf), NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_destroy(NULL));
+  EXPECT_TRUE(vpx_codec_error(NULL) != NULL);
+
+  for (int i = 0; i < NELEMENTS(kCodecs); ++i) {
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_dec_init(NULL, kCodecs[i], NULL, 0));
+
+    EXPECT_EQ(VPX_CODEC_OK, vpx_codec_dec_init(&dec, kCodecs[i], NULL, 0));
+    EXPECT_EQ(VPX_CODEC_UNSUP_BITSTREAM,
+              vpx_codec_decode(&dec, buf, NELEMENTS(buf), NULL, 0));
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_decode(&dec, NULL, NELEMENTS(buf), NULL, 0));
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_decode(&dec, buf, 0, NULL, 0));
+
+    EXPECT_EQ(VPX_CODEC_OK, vpx_codec_destroy(&dec));
+  }
+}
+
+#if CONFIG_VP8_DECODER
+TEST(DecodeAPI, OptionalParams) {
+  vpx_codec_ctx_t dec;
+
+#if CONFIG_ERROR_CONCEALMENT
+  EXPECT_EQ(VPX_CODEC_OK, vpx_codec_dec_init(&dec, &vpx_codec_vp8_dx_algo, NULL,
+                                             VPX_CODEC_USE_ERROR_CONCEALMENT));
+#else
+  EXPECT_EQ(VPX_CODEC_INCAPABLE,
+            vpx_codec_dec_init(&dec, &vpx_codec_vp8_dx_algo, NULL,
+                               VPX_CODEC_USE_ERROR_CONCEALMENT));
+#endif  // CONFIG_ERROR_CONCEALMENT
+}
+#endif  // CONFIG_VP8_DECODER
+
+#if CONFIG_VP9_DECODER
+// Test VP9 codec controls after a decode error to ensure the code doesn't
+// misbehave.
+void TestVp9Controls(vpx_codec_ctx_t *dec) {
+  static const int kControls[] = {
+    VP8D_GET_LAST_REF_UPDATES,
+    VP8D_GET_FRAME_CORRUPTED,
+    VP9D_GET_DISPLAY_SIZE,
+    VP9D_GET_FRAME_SIZE
+  };
+  int val[2];
+
+  for (int i = 0; i < NELEMENTS(kControls); ++i) {
+    const vpx_codec_err_t res = vpx_codec_control_(dec, kControls[i], val);
+    switch (kControls[i]) {
+      case VP8D_GET_FRAME_CORRUPTED:
+        EXPECT_EQ(VPX_CODEC_ERROR, res) << kControls[i];
+        break;
+      default:
+        EXPECT_EQ(VPX_CODEC_OK, res) << kControls[i];
+        break;
+    }
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_control_(dec, kControls[i], NULL));
+  }
+
+  vp9_ref_frame_t ref;
+  ref.idx = 0;
+  EXPECT_EQ(VPX_CODEC_ERROR, vpx_codec_control(dec, VP9_GET_REFERENCE, &ref));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+            vpx_codec_control(dec, VP9_GET_REFERENCE, NULL));
+
+  vpx_ref_frame_t ref_copy;
+  const int width = 352;
+  const int height = 288;
+  ASSERT_TRUE(
+      vpx_img_alloc(&ref_copy.img, VPX_IMG_FMT_I420, width, height, 1) != NULL);
+  ref_copy.frame_type = VP8_LAST_FRAME;
+  EXPECT_EQ(VPX_CODEC_ERROR,
+            vpx_codec_control(dec, VP8_COPY_REFERENCE, &ref_copy));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+            vpx_codec_control(dec, VP8_COPY_REFERENCE, NULL));
+  vpx_img_free(&ref_copy.img);
+}
+
+TEST(DecodeAPI, Vp9InvalidDecode) {
+  const vpx_codec_iface_t *const codec = &vpx_codec_vp9_dx_algo;
+  const char filename[] =
+      "invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf";
+  libvpx_test::IVFVideoSource video(filename);
+  video.Init();
+  video.Begin();
+  ASSERT_TRUE(!HasFailure());
+
+  vpx_codec_ctx_t dec;
+  EXPECT_EQ(VPX_CODEC_OK, vpx_codec_dec_init(&dec, codec, NULL, 0));
+  const uint32_t frame_size = static_cast<uint32_t>(video.frame_size());
+#if CONFIG_VP9_HIGHBITDEPTH
+  EXPECT_EQ(VPX_CODEC_MEM_ERROR,
+            vpx_codec_decode(&dec, video.cxdata(), frame_size, NULL, 0));
+#else
+  EXPECT_EQ(VPX_CODEC_UNSUP_BITSTREAM,
+            vpx_codec_decode(&dec, video.cxdata(), frame_size, NULL, 0));
+#endif
+  vpx_codec_iter_t iter = NULL;
+  EXPECT_EQ(NULL, vpx_codec_get_frame(&dec, &iter));
+
+  TestVp9Controls(&dec);
+  EXPECT_EQ(VPX_CODEC_OK, vpx_codec_destroy(&dec));
+}
+
+TEST(DecodeAPI, Vp9PeekSI) {
+  const vpx_codec_iface_t *const codec = &vpx_codec_vp9_dx_algo;
+  // The first 9 bytes are valid and the rest of the bytes are made up. Until
+  // size 10, this should return VPX_CODEC_UNSUP_BITSTREAM and after that it
+  // should return VPX_CODEC_CORRUPT_FRAME.
+  const uint8_t data[32] = {
+    0x85, 0xa4, 0xc1, 0xa1, 0x38, 0x81, 0xa3, 0x49,
+    0x83, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+  };
+
+  for (uint32_t data_sz = 1; data_sz <= 32; ++data_sz) {
+    // Verify behavior of vpx_codec_decode. vpx_codec_decode doesn't even get
+    // to decoder_peek_si_internal on frames of size < 8.
+    if (data_sz >= 8) {
+      vpx_codec_ctx_t dec;
+      EXPECT_EQ(VPX_CODEC_OK, vpx_codec_dec_init(&dec, codec, NULL, 0));
+      EXPECT_EQ((data_sz < 10) ?
+                    VPX_CODEC_UNSUP_BITSTREAM : VPX_CODEC_CORRUPT_FRAME,
+                vpx_codec_decode(&dec, data, data_sz, NULL, 0));
+      vpx_codec_iter_t iter = NULL;
+      EXPECT_EQ(NULL, vpx_codec_get_frame(&dec, &iter));
+      EXPECT_EQ(VPX_CODEC_OK, vpx_codec_destroy(&dec));
+    }
+
+    // Verify behavior of vpx_codec_peek_stream_info.
+    vpx_codec_stream_info_t si;
+    si.sz = sizeof(si);
+    EXPECT_EQ((data_sz < 10) ? VPX_CODEC_UNSUP_BITSTREAM : VPX_CODEC_OK,
+              vpx_codec_peek_stream_info(codec, data, data_sz, &si));
+  }
+}
+#endif  // CONFIG_VP9_DECODER
+
+}  // namespace
diff --git a/libvpx/libvpx/test/decode_perf_test.cc b/libvpx/libvpx/test/decode_perf_test.cc
new file mode 100644
index 0000000..c24d517
--- /dev/null
+++ b/libvpx/libvpx/test/decode_perf_test.cc
@@ -0,0 +1,273 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/ivf_video_source.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#include "test/webm_video_source.h"
+#include "vpx_ports/vpx_timer.h"
+#include "./ivfenc.h"
+#include "./vpx_version.h"
+
+using std::tr1::make_tuple;
+
+namespace {
+
+#define VIDEO_NAME 0
+#define THREADS 1
+
+const int kMaxPsnr = 100;
+const double kUsecsInSec = 1000000.0;
+const char kNewEncodeOutputFile[] = "new_encode.ivf";
+
+/*
+ DecodePerfTest takes a tuple of filename + number of threads to decode with
+ */
+typedef std::tr1::tuple<const char *, unsigned> DecodePerfParam;
+
+const DecodePerfParam kVP9DecodePerfVectors[] = {
+  make_tuple("vp90-2-bbb_426x240_tile_1x1_180kbps.webm", 1),
+  make_tuple("vp90-2-bbb_640x360_tile_1x2_337kbps.webm", 2),
+  make_tuple("vp90-2-bbb_854x480_tile_1x2_651kbps.webm", 2),
+  make_tuple("vp90-2-bbb_1280x720_tile_1x4_1310kbps.webm", 4),
+  make_tuple("vp90-2-bbb_1920x1080_tile_1x1_2581kbps.webm", 1),
+  make_tuple("vp90-2-bbb_1920x1080_tile_1x4_2586kbps.webm", 4),
+  make_tuple("vp90-2-bbb_1920x1080_tile_1x4_fpm_2304kbps.webm", 4),
+  make_tuple("vp90-2-sintel_426x182_tile_1x1_171kbps.webm", 1),
+  make_tuple("vp90-2-sintel_640x272_tile_1x2_318kbps.webm", 2),
+  make_tuple("vp90-2-sintel_854x364_tile_1x2_621kbps.webm", 2),
+  make_tuple("vp90-2-sintel_1280x546_tile_1x4_1257kbps.webm", 4),
+  make_tuple("vp90-2-sintel_1920x818_tile_1x4_fpm_2279kbps.webm", 4),
+  make_tuple("vp90-2-tos_426x178_tile_1x1_181kbps.webm", 1),
+  make_tuple("vp90-2-tos_640x266_tile_1x2_336kbps.webm", 2),
+  make_tuple("vp90-2-tos_854x356_tile_1x2_656kbps.webm", 2),
+  make_tuple("vp90-2-tos_854x356_tile_1x2_fpm_546kbps.webm", 2),
+  make_tuple("vp90-2-tos_1280x534_tile_1x4_1306kbps.webm", 4),
+  make_tuple("vp90-2-tos_1280x534_tile_1x4_fpm_952kbps.webm", 4),
+  make_tuple("vp90-2-tos_1920x800_tile_1x4_fpm_2335kbps.webm", 4),
+};
+
+/*
+ In order to reflect real world performance as much as possible, Perf tests
+ *DO NOT* do any correctness checks. Please run them alongside correctness
+ tests to ensure proper codec integrity. Furthermore, in this test we
+ deliberately limit the amount of system calls we make to avoid OS
+ preemption.
+
+ TODO(joshualitt) create a more detailed perf measurement test to collect
+   power/temp/min max frame decode times/etc
+ */
+
+class DecodePerfTest : public ::testing::TestWithParam<DecodePerfParam> {
+};
+
+TEST_P(DecodePerfTest, PerfTest) {
+  const char *const video_name = GET_PARAM(VIDEO_NAME);
+  const unsigned threads = GET_PARAM(THREADS);
+
+  libvpx_test::WebMVideoSource video(video_name);
+  video.Init();
+
+  vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+  cfg.threads = threads;
+  libvpx_test::VP9Decoder decoder(cfg, 0);
+
+  vpx_usec_timer t;
+  vpx_usec_timer_start(&t);
+
+  for (video.Begin(); video.cxdata() != NULL; video.Next()) {
+    decoder.DecodeFrame(video.cxdata(), video.frame_size());
+  }
+
+  vpx_usec_timer_mark(&t);
+  const double elapsed_secs = double(vpx_usec_timer_elapsed(&t))
+                              / kUsecsInSec;
+  const unsigned frames = video.frame_number();
+  const double fps = double(frames) / elapsed_secs;
+
+  printf("{\n");
+  printf("\t\"type\" : \"decode_perf_test\",\n");
+  printf("\t\"version\" : \"%s\",\n", VERSION_STRING_NOSP);
+  printf("\t\"videoName\" : \"%s\",\n", video_name);
+  printf("\t\"threadCount\" : %u,\n", threads);
+  printf("\t\"decodeTimeSecs\" : %f,\n", elapsed_secs);
+  printf("\t\"totalFrames\" : %u,\n", frames);
+  printf("\t\"framesPerSecond\" : %f\n", fps);
+  printf("}\n");
+}
+
+INSTANTIATE_TEST_CASE_P(VP9, DecodePerfTest,
+                        ::testing::ValuesIn(kVP9DecodePerfVectors));
+
+class VP9NewEncodeDecodePerfTest :
+    public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ protected:
+  VP9NewEncodeDecodePerfTest()
+      : EncoderTest(GET_PARAM(0)),
+        encoding_mode_(GET_PARAM(1)),
+        speed_(0),
+        outfile_(0),
+        out_frames_(0) {
+  }
+
+  virtual ~VP9NewEncodeDecodePerfTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+
+    cfg_.g_lag_in_frames = 25;
+    cfg_.rc_min_quantizer = 2;
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_dropframe_thresh = 0;
+    cfg_.rc_undershoot_pct = 50;
+    cfg_.rc_overshoot_pct = 50;
+    cfg_.rc_buf_sz = 1000;
+    cfg_.rc_buf_initial_sz = 500;
+    cfg_.rc_buf_optimal_sz = 600;
+    cfg_.rc_resize_allowed = 0;
+    cfg_.rc_end_usage = VPX_VBR;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_CPUUSED, speed_);
+      encoder->Control(VP9E_SET_FRAME_PARALLEL_DECODING, 1);
+      encoder->Control(VP9E_SET_TILE_COLUMNS, 2);
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    const std::string data_path = getenv("LIBVPX_TEST_DATA_PATH");
+    const std::string path_to_source = data_path + "/" + kNewEncodeOutputFile;
+    outfile_ = fopen(path_to_source.c_str(), "wb");
+    ASSERT_TRUE(outfile_ != NULL);
+  }
+
+  virtual void EndPassHook() {
+    if (outfile_ != NULL) {
+      if (!fseek(outfile_, 0, SEEK_SET))
+        ivf_write_file_header(outfile_, &cfg_, VP9_FOURCC, out_frames_);
+      fclose(outfile_);
+      outfile_ = NULL;
+    }
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    ++out_frames_;
+
+    // Write initial file header if first frame.
+    if (pkt->data.frame.pts == 0)
+      ivf_write_file_header(outfile_, &cfg_, VP9_FOURCC, out_frames_);
+
+    // Write frame header and data.
+    ivf_write_frame_header(outfile_, out_frames_, pkt->data.frame.sz);
+    ASSERT_EQ(fwrite(pkt->data.frame.buf, 1, pkt->data.frame.sz, outfile_),
+              pkt->data.frame.sz);
+  }
+
+  virtual bool DoDecode() { return false; }
+
+  void set_speed(unsigned int speed) {
+    speed_ = speed;
+  }
+
+ private:
+  libvpx_test::TestMode encoding_mode_;
+  uint32_t speed_;
+  FILE *outfile_;
+  uint32_t out_frames_;
+};
+
+struct EncodePerfTestVideo {
+  EncodePerfTestVideo(const char *name_, uint32_t width_, uint32_t height_,
+                      uint32_t bitrate_, int frames_)
+      : name(name_),
+        width(width_),
+        height(height_),
+        bitrate(bitrate_),
+        frames(frames_) {}
+  const char *name;
+  uint32_t width;
+  uint32_t height;
+  uint32_t bitrate;
+  int frames;
+};
+
+const EncodePerfTestVideo kVP9EncodePerfTestVectors[] = {
+  EncodePerfTestVideo("niklas_1280_720_30.yuv", 1280, 720, 600, 470),
+};
+
+TEST_P(VP9NewEncodeDecodePerfTest, PerfTest) {
+  SetUp();
+
+  // TODO(JBB): Make this work by going through the set of given files.
+  const int i = 0;
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = kVP9EncodePerfTestVectors[i].bitrate;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  const char *video_name = kVP9EncodePerfTestVectors[i].name;
+  libvpx_test::I420VideoSource video(
+      video_name,
+      kVP9EncodePerfTestVectors[i].width,
+      kVP9EncodePerfTestVectors[i].height,
+      timebase.den, timebase.num, 0,
+      kVP9EncodePerfTestVectors[i].frames);
+  set_speed(2);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  const uint32_t threads = 4;
+
+  libvpx_test::IVFVideoSource decode_video(kNewEncodeOutputFile);
+  decode_video.Init();
+
+  vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+  cfg.threads = threads;
+  libvpx_test::VP9Decoder decoder(cfg, 0);
+
+  vpx_usec_timer t;
+  vpx_usec_timer_start(&t);
+
+  for (decode_video.Begin(); decode_video.cxdata() != NULL;
+       decode_video.Next()) {
+    decoder.DecodeFrame(decode_video.cxdata(), decode_video.frame_size());
+  }
+
+  vpx_usec_timer_mark(&t);
+  const double elapsed_secs =
+      static_cast<double>(vpx_usec_timer_elapsed(&t)) / kUsecsInSec;
+  const unsigned decode_frames = decode_video.frame_number();
+  const double fps = static_cast<double>(decode_frames) / elapsed_secs;
+
+  printf("{\n");
+  printf("\t\"type\" : \"decode_perf_test\",\n");
+  printf("\t\"version\" : \"%s\",\n", VERSION_STRING_NOSP);
+  printf("\t\"videoName\" : \"%s\",\n", kNewEncodeOutputFile);
+  printf("\t\"threadCount\" : %u,\n", threads);
+  printf("\t\"decodeTimeSecs\" : %f,\n", elapsed_secs);
+  printf("\t\"totalFrames\" : %u,\n", decode_frames);
+  printf("\t\"framesPerSecond\" : %f\n", fps);
+  printf("}\n");
+}
+
+VP9_INSTANTIATE_TEST_CASE(
+  VP9NewEncodeDecodePerfTest, ::testing::Values(::libvpx_test::kTwoPassGood));
+}  // namespace
diff --git a/libvpx/libvpx/test/decode_test_driver.cc b/libvpx/libvpx/test/decode_test_driver.cc
new file mode 100644
index 0000000..ad861c3
--- /dev/null
+++ b/libvpx/libvpx/test/decode_test_driver.cc
@@ -0,0 +1,123 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/register_state_check.h"
+#include "test/video_source.h"
+
+namespace libvpx_test {
+
+const char kVP8Name[] = "WebM Project VP8";
+
+vpx_codec_err_t Decoder::PeekStream(const uint8_t *cxdata, size_t size,
+                                    vpx_codec_stream_info_t *stream_info) {
+  return vpx_codec_peek_stream_info(CodecInterface(),
+                                    cxdata, static_cast<unsigned int>(size),
+                                    stream_info);
+}
+
+vpx_codec_err_t Decoder::DecodeFrame(const uint8_t *cxdata, size_t size) {
+  return DecodeFrame(cxdata, size, NULL);
+}
+
+vpx_codec_err_t Decoder::DecodeFrame(const uint8_t *cxdata, size_t size,
+                                     void *user_priv) {
+  vpx_codec_err_t res_dec;
+  InitOnce();
+  API_REGISTER_STATE_CHECK(
+      res_dec = vpx_codec_decode(&decoder_,
+                                 cxdata, static_cast<unsigned int>(size),
+                                 user_priv, 0));
+  return res_dec;
+}
+
+bool Decoder::IsVP8() const {
+  const char *codec_name = GetDecoderName();
+  return strncmp(kVP8Name, codec_name, sizeof(kVP8Name) - 1) == 0;
+}
+
+void DecoderTest::HandlePeekResult(Decoder *const decoder,
+                                   CompressedVideoSource *video,
+                                   const vpx_codec_err_t res_peek) {
+  const bool is_vp8 = decoder->IsVP8();
+  if (is_vp8) {
+    /* Vp8's implementation of PeekStream returns an error if the frame you
+     * pass it is not a keyframe, so we only expect VPX_CODEC_OK on the first
+     * frame, which must be a keyframe. */
+    if (video->frame_number() == 0)
+      ASSERT_EQ(VPX_CODEC_OK, res_peek) << "Peek return failed: "
+                                        << vpx_codec_err_to_string(res_peek);
+  } else {
+    /* The Vp9 implementation of PeekStream returns an error only if the
+     * data passed to it isn't a valid Vp9 chunk. */
+    ASSERT_EQ(VPX_CODEC_OK, res_peek) << "Peek return failed: "
+                                      << vpx_codec_err_to_string(res_peek);
+  }
+}
+
+void DecoderTest::RunLoop(CompressedVideoSource *video,
+                          const vpx_codec_dec_cfg_t &dec_cfg) {
+  Decoder* const decoder = codec_->CreateDecoder(dec_cfg, flags_, 0);
+  ASSERT_TRUE(decoder != NULL);
+  bool end_of_file = false;
+
+  // Decode frames.
+  for (video->Begin(); !::testing::Test::HasFailure() && !end_of_file;
+       video->Next()) {
+    PreDecodeFrameHook(*video, decoder);
+
+    vpx_codec_stream_info_t stream_info;
+    stream_info.sz = sizeof(stream_info);
+
+    if (video->cxdata() != NULL) {
+      const vpx_codec_err_t res_peek = decoder->PeekStream(video->cxdata(),
+                                                           video->frame_size(),
+                                                           &stream_info);
+      HandlePeekResult(decoder, video, res_peek);
+      ASSERT_FALSE(::testing::Test::HasFailure());
+
+      vpx_codec_err_t res_dec = decoder->DecodeFrame(video->cxdata(),
+                                                     video->frame_size());
+      if (!HandleDecodeResult(res_dec, *video, decoder))
+        break;
+    } else {
+      // Signal end of the file to the decoder.
+      const vpx_codec_err_t res_dec = decoder->DecodeFrame(NULL, 0);
+      ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+      end_of_file = true;
+    }
+
+    DxDataIterator dec_iter = decoder->GetDxData();
+    const vpx_image_t *img = NULL;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next()))
+      DecompressedFrameHook(*img, video->frame_number());
+  }
+  delete decoder;
+}
+
+void DecoderTest::RunLoop(CompressedVideoSource *video) {
+  vpx_codec_dec_cfg_t dec_cfg = vpx_codec_dec_cfg_t();
+  RunLoop(video, dec_cfg);
+}
+
+void DecoderTest::set_cfg(const vpx_codec_dec_cfg_t &dec_cfg) {
+  memcpy(&cfg_, &dec_cfg, sizeof(cfg_));
+}
+
+void DecoderTest::set_flags(const vpx_codec_flags_t flags) {
+  flags_ = flags;
+}
+
+}  // namespace libvpx_test
diff --git a/libvpx/libvpx/test/decode_test_driver.h b/libvpx/libvpx/test/decode_test_driver.h
new file mode 100644
index 0000000..f566c53
--- /dev/null
+++ b/libvpx/libvpx/test/decode_test_driver.h
@@ -0,0 +1,181 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef TEST_DECODE_TEST_DRIVER_H_
+#define TEST_DECODE_TEST_DRIVER_H_
+#include <cstring>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "vpx/vpx_decoder.h"
+
+namespace libvpx_test {
+
+class CodecFactory;
+class CompressedVideoSource;
+
+// Provides an object to handle decoding output
+class DxDataIterator {
+ public:
+  explicit DxDataIterator(vpx_codec_ctx_t *decoder)
+      : decoder_(decoder), iter_(NULL) {}
+
+  const vpx_image_t *Next() {
+    return vpx_codec_get_frame(decoder_, &iter_);
+  }
+
+ private:
+  vpx_codec_ctx_t  *decoder_;
+  vpx_codec_iter_t  iter_;
+};
+
+// Provides a simplified interface to manage one video decoding.
+// Similar to Encoder class, the exact services should be added
+// as more tests are added.
+class Decoder {
+ public:
+  Decoder(vpx_codec_dec_cfg_t cfg, unsigned long deadline)
+      : cfg_(cfg), flags_(0), deadline_(deadline), init_done_(false) {
+    memset(&decoder_, 0, sizeof(decoder_));
+  }
+
+  Decoder(vpx_codec_dec_cfg_t cfg, const vpx_codec_flags_t flag,
+          unsigned long deadline)  // NOLINT
+      : cfg_(cfg), flags_(flag), deadline_(deadline), init_done_(false) {
+    memset(&decoder_, 0, sizeof(decoder_));
+  }
+
+  virtual ~Decoder() {
+    vpx_codec_destroy(&decoder_);
+  }
+
+  vpx_codec_err_t PeekStream(const uint8_t *cxdata, size_t size,
+                             vpx_codec_stream_info_t *stream_info);
+
+  vpx_codec_err_t DecodeFrame(const uint8_t *cxdata, size_t size);
+
+  vpx_codec_err_t DecodeFrame(const uint8_t *cxdata, size_t size,
+                              void *user_priv);
+
+  DxDataIterator GetDxData() {
+    return DxDataIterator(&decoder_);
+  }
+
+  void set_deadline(unsigned long deadline) {
+    deadline_ = deadline;
+  }
+
+  void Control(int ctrl_id, int arg) {
+    Control(ctrl_id, arg, VPX_CODEC_OK);
+  }
+
+  void Control(int ctrl_id, const void *arg) {
+    InitOnce();
+    const vpx_codec_err_t res = vpx_codec_control_(&decoder_, ctrl_id, arg);
+    ASSERT_EQ(VPX_CODEC_OK, res) << DecodeError();
+  }
+
+  void Control(int ctrl_id, int arg, vpx_codec_err_t expected_value) {
+    InitOnce();
+    const vpx_codec_err_t res = vpx_codec_control_(&decoder_, ctrl_id, arg);
+    ASSERT_EQ(expected_value, res) << DecodeError();
+  }
+
+  const char* DecodeError() {
+    const char *detail = vpx_codec_error_detail(&decoder_);
+    return detail ? detail : vpx_codec_error(&decoder_);
+  }
+
+  // Passes the external frame buffer information to libvpx.
+  vpx_codec_err_t SetFrameBufferFunctions(
+      vpx_get_frame_buffer_cb_fn_t cb_get,
+      vpx_release_frame_buffer_cb_fn_t cb_release, void *user_priv) {
+    InitOnce();
+    return vpx_codec_set_frame_buffer_functions(
+        &decoder_, cb_get, cb_release, user_priv);
+  }
+
+  const char* GetDecoderName() const {
+    return vpx_codec_iface_name(CodecInterface());
+  }
+
+  bool IsVP8() const;
+
+  vpx_codec_ctx_t * GetDecoder() {
+    return &decoder_;
+  }
+
+ protected:
+  virtual vpx_codec_iface_t* CodecInterface() const = 0;
+
+  void InitOnce() {
+    if (!init_done_) {
+      const vpx_codec_err_t res = vpx_codec_dec_init(&decoder_,
+                                                     CodecInterface(),
+                                                     &cfg_, flags_);
+      ASSERT_EQ(VPX_CODEC_OK, res) << DecodeError();
+      init_done_ = true;
+    }
+  }
+
+  vpx_codec_ctx_t     decoder_;
+  vpx_codec_dec_cfg_t cfg_;
+  vpx_codec_flags_t   flags_;
+  unsigned int        deadline_;
+  bool                init_done_;
+};
+
+// Common test functionality for all Decoder tests.
+class DecoderTest {
+ public:
+  // Main decoding loop
+  virtual void RunLoop(CompressedVideoSource *video);
+  virtual void RunLoop(CompressedVideoSource *video,
+                       const vpx_codec_dec_cfg_t &dec_cfg);
+
+  virtual void set_cfg(const vpx_codec_dec_cfg_t &dec_cfg);
+  virtual void set_flags(const vpx_codec_flags_t flags);
+
+  // Hook to be called before decompressing every frame.
+  virtual void PreDecodeFrameHook(const CompressedVideoSource& /*video*/,
+                                  Decoder* /*decoder*/) {}
+
+  // Hook to be called to handle decode result. Return true to continue.
+  virtual bool HandleDecodeResult(const vpx_codec_err_t res_dec,
+                                  const CompressedVideoSource& /*video*/,
+                                  Decoder *decoder) {
+    EXPECT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+    return VPX_CODEC_OK == res_dec;
+  }
+
+  // Hook to be called on every decompressed frame.
+  virtual void DecompressedFrameHook(const vpx_image_t& /*img*/,
+                                     const unsigned int /*frame_number*/) {}
+
+  // Hook to be called on peek result
+  virtual void HandlePeekResult(Decoder* const decoder,
+                                CompressedVideoSource *video,
+                                const vpx_codec_err_t res_peek);
+
+ protected:
+  explicit DecoderTest(const CodecFactory *codec)
+      : codec_(codec),
+        cfg_(),
+        flags_(0) {}
+
+  virtual ~DecoderTest() {}
+
+  const CodecFactory *codec_;
+  vpx_codec_dec_cfg_t cfg_;
+  vpx_codec_flags_t   flags_;
+};
+
+}  // namespace libvpx_test
+
+#endif  // TEST_DECODE_TEST_DRIVER_H_
diff --git a/libvpx/libvpx/test/decode_to_md5.sh b/libvpx/libvpx/test/decode_to_md5.sh
new file mode 100755
index 0000000..854b74f
--- /dev/null
+++ b/libvpx/libvpx/test/decode_to_md5.sh
@@ -0,0 +1,73 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx decode_to_md5 example. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to decode_to_md5_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+#   $VP8_IVF_FILE and $VP9_IVF_FILE are required.
+decode_to_md5_verify_environment() {
+  if [ ! -e "${VP8_IVF_FILE}" ] || [ ! -e "${VP9_IVF_FILE}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs decode_to_md5 on $1 and captures the md5 sum for the final frame. $2 is
+# interpreted as codec name and used solely to name the output file. $3 is the
+# expected md5 sum: It must match that of the final frame.
+decode_to_md5() {
+  local decoder="${LIBVPX_BIN_PATH}/decode_to_md5${VPX_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local expected_md5="$3"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/decode_to_md5_${codec}"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+
+  local md5_last_frame="$(tail -n1 "${output_file}" | awk '{print $1}')"
+  local actual_md5="$(echo "${md5_last_frame}" | awk '{print $1}')"
+  [ "${actual_md5}" = "${expected_md5}" ] || return 1
+}
+
+decode_to_md5_vp8() {
+  # expected MD5 sum for the last frame.
+  local expected_md5="56794d911b02190212bca92f88ad60c6"
+
+  if [ "$(vp8_decode_available)" = "yes" ]; then
+    decode_to_md5 "${VP8_IVF_FILE}" "vp8" "${expected_md5}"
+  fi
+}
+
+decode_to_md5_vp9() {
+  # expected MD5 sum for the last frame.
+  local expected_md5="2952c0eae93f3dadd1aa84c50d3fd6d2"
+
+  if [ "$(vp9_decode_available)" = "yes" ]; then
+    decode_to_md5 "${VP9_IVF_FILE}" "vp9" "${expected_md5}"
+  fi
+}
+
+decode_to_md5_tests="decode_to_md5_vp8
+                     decode_to_md5_vp9"
+
+run_tests decode_to_md5_verify_environment "${decode_to_md5_tests}"
diff --git a/libvpx/libvpx/test/decode_with_drops.sh b/libvpx/libvpx/test/decode_with_drops.sh
new file mode 100755
index 0000000..9b2edb6
--- /dev/null
+++ b/libvpx/libvpx/test/decode_with_drops.sh
@@ -0,0 +1,79 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx decode_with_drops example. To add new tests to
+##  this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to decode_with_drops_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+#   $VP8_IVF_FILE and $VP9_IVF_FILE are required.
+decode_with_drops_verify_environment() {
+  if [ ! -e "${VP8_IVF_FILE}" ] || [ ! -e "${VP9_IVF_FILE}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs decode_with_drops on $1, $2 is interpreted as codec name and used solely
+# to name the output file. $3 is the drop mode, and is passed directly to
+# decode_with_drops.
+decode_with_drops() {
+  local decoder="${LIBVPX_BIN_PATH}/decode_with_drops${VPX_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/decode_with_drops_${codec}"
+  local drop_mode="$3"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      "${drop_mode}" ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+# Decodes $VP8_IVF_FILE while dropping frames, twice: once in sequence mode,
+# and once in pattern mode.
+# Note: This test assumes that $VP8_IVF_FILE has exactly 29 frames, and could
+# break if the file is modified.
+decode_with_drops_vp8() {
+  if [ "$(vp8_decode_available)" = "yes" ]; then
+    # Test sequence mode: Drop frames 2-28.
+    decode_with_drops "${VP8_IVF_FILE}" "vp8" "2-28"
+
+    # Test pattern mode: Drop 3 of every 4 frames.
+    decode_with_drops "${VP8_IVF_FILE}" "vp8" "3/4"
+  fi
+}
+
+# Decodes $VP9_IVF_FILE while dropping frames, twice: once in sequence mode,
+# and once in pattern mode.
+# Note: This test assumes that $VP9_IVF_FILE has exactly 20 frames, and could
+# break if the file is modified.
+decode_with_drops_vp9() {
+  if [ "$(vp9_decode_available)" = "yes" ]; then
+    # Test sequence mode: Drop frames 2-28.
+    decode_with_drops "${VP9_IVF_FILE}" "vp9" "2-19"
+
+    # Test pattern mode: Drop 3 of every 4 frames.
+    decode_with_drops "${VP9_IVF_FILE}" "vp9" "3/4"
+  fi
+}
+
+decode_with_drops_tests="decode_with_drops_vp8
+                         decode_with_drops_vp9"
+
+run_tests decode_with_drops_verify_environment "${decode_with_drops_tests}"
diff --git a/libvpx/libvpx/test/encode_api_test.cc b/libvpx/libvpx/test/encode_api_test.cc
new file mode 100644
index 0000000..94afdde
--- /dev/null
+++ b/libvpx/libvpx/test/encode_api_test.cc
@@ -0,0 +1,65 @@
+/*
+ *  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+
+namespace {
+
+#define NELEMENTS(x) static_cast<int>(sizeof(x) / sizeof(x[0]))
+
+TEST(EncodeAPI, InvalidParams) {
+  static const vpx_codec_iface_t *kCodecs[] = {
+#if CONFIG_VP8_ENCODER
+    &vpx_codec_vp8_cx_algo,
+#endif
+#if CONFIG_VP9_ENCODER
+    &vpx_codec_vp9_cx_algo,
+#endif
+  };
+  uint8_t buf[1] = {0};
+  vpx_image_t img;
+  vpx_codec_ctx_t enc;
+  vpx_codec_enc_cfg_t cfg;
+
+  EXPECT_EQ(&img, vpx_img_wrap(&img, VPX_IMG_FMT_I420, 1, 1, 1, buf));
+
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_enc_init(NULL, NULL, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_enc_init(&enc, NULL, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_encode(NULL, NULL, 0, 0, 0, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_encode(NULL, &img, 0, 0, 0, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, vpx_codec_destroy(NULL));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+            vpx_codec_enc_config_default(NULL, NULL, 0));
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+            vpx_codec_enc_config_default(NULL, &cfg, 0));
+  EXPECT_TRUE(vpx_codec_error(NULL) != NULL);
+
+  for (int i = 0; i < NELEMENTS(kCodecs); ++i) {
+    SCOPED_TRACE(vpx_codec_iface_name(kCodecs[i]));
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_enc_init(NULL, kCodecs[i], NULL, 0));
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_enc_init(&enc, kCodecs[i], NULL, 0));
+    EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+              vpx_codec_enc_config_default(kCodecs[i], &cfg, 1));
+
+    EXPECT_EQ(VPX_CODEC_OK, vpx_codec_enc_config_default(kCodecs[i], &cfg, 0));
+    EXPECT_EQ(VPX_CODEC_OK, vpx_codec_enc_init(&enc, kCodecs[i], &cfg, 0));
+    EXPECT_EQ(VPX_CODEC_OK, vpx_codec_encode(&enc, NULL, 0, 0, 0, 0));
+
+    EXPECT_EQ(VPX_CODEC_OK, vpx_codec_destroy(&enc));
+  }
+}
+
+}  // namespace
diff --git a/libvpx/libvpx/test/encode_perf_test.cc b/libvpx/libvpx/test/encode_perf_test.cc
new file mode 100644
index 0000000..7e9f0d6
--- /dev/null
+++ b/libvpx/libvpx/test/encode_perf_test.cc
@@ -0,0 +1,202 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <string>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "./vpx_version.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "vpx_ports/vpx_timer.h"
+
+namespace {
+
+const int kMaxPsnr = 100;
+const double kUsecsInSec = 1000000.0;
+
+struct EncodePerfTestVideo {
+  EncodePerfTestVideo(const char *name_, uint32_t width_, uint32_t height_,
+                      uint32_t bitrate_, int frames_)
+      : name(name_),
+        width(width_),
+        height(height_),
+        bitrate(bitrate_),
+        frames(frames_) {}
+  const char *name;
+  uint32_t width;
+  uint32_t height;
+  uint32_t bitrate;
+  int frames;
+};
+
+const EncodePerfTestVideo kVP9EncodePerfTestVectors[] = {
+  EncodePerfTestVideo("desktop_640_360_30.yuv", 640, 360, 200, 2484),
+  EncodePerfTestVideo("kirland_640_480_30.yuv", 640, 480, 200, 300),
+  EncodePerfTestVideo("macmarcomoving_640_480_30.yuv", 640, 480, 200, 987),
+  EncodePerfTestVideo("macmarcostationary_640_480_30.yuv", 640, 480, 200, 718),
+  EncodePerfTestVideo("niklas_640_480_30.yuv", 640, 480, 200, 471),
+  EncodePerfTestVideo("tacomanarrows_640_480_30.yuv", 640, 480, 200, 300),
+  EncodePerfTestVideo("tacomasmallcameramovement_640_480_30.yuv",
+                      640, 480, 200, 300),
+  EncodePerfTestVideo("thaloundeskmtg_640_480_30.yuv", 640, 480, 200, 300),
+  EncodePerfTestVideo("niklas_1280_720_30.yuv", 1280, 720, 600, 470),
+};
+
+const int kEncodePerfTestSpeeds[] = { 5, 6, 7, 8 };
+const int kEncodePerfTestThreads[] = { 1, 2, 4 };
+
+#define NELEMENTS(x) (sizeof((x)) / sizeof((x)[0]))
+
+class VP9EncodePerfTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ protected:
+  VP9EncodePerfTest()
+      : EncoderTest(GET_PARAM(0)),
+        min_psnr_(kMaxPsnr),
+        nframes_(0),
+        encoding_mode_(GET_PARAM(1)),
+        speed_(0),
+        threads_(1) {}
+
+  virtual ~VP9EncodePerfTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+
+    cfg_.g_lag_in_frames = 0;
+    cfg_.rc_min_quantizer = 2;
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_dropframe_thresh = 0;
+    cfg_.rc_undershoot_pct = 50;
+    cfg_.rc_overshoot_pct = 50;
+    cfg_.rc_buf_sz = 1000;
+    cfg_.rc_buf_initial_sz = 500;
+    cfg_.rc_buf_optimal_sz = 600;
+    cfg_.rc_resize_allowed = 0;
+    cfg_.rc_end_usage = VPX_CBR;
+    cfg_.g_error_resilient = 1;
+    cfg_.g_threads = threads_;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      const int log2_tile_columns = 3;
+      encoder->Control(VP8E_SET_CPUUSED, speed_);
+      encoder->Control(VP9E_SET_TILE_COLUMNS, log2_tile_columns);
+      encoder->Control(VP9E_SET_FRAME_PARALLEL_DECODING, 1);
+      encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 0);
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    min_psnr_ = kMaxPsnr;
+    nframes_ = 0;
+  }
+
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->data.psnr.psnr[0] < min_psnr_) {
+      min_psnr_= pkt->data.psnr.psnr[0];
+    }
+  }
+
+  // for performance reasons don't decode
+  virtual bool DoDecode() { return 0; }
+
+  double min_psnr() const {
+    return min_psnr_;
+  }
+
+  void set_speed(unsigned int speed) {
+    speed_ = speed;
+  }
+
+  void set_threads(unsigned int threads) {
+    threads_ = threads;
+  }
+
+ private:
+  double min_psnr_;
+  unsigned int nframes_;
+  libvpx_test::TestMode encoding_mode_;
+  unsigned speed_;
+  unsigned int threads_;
+};
+
+TEST_P(VP9EncodePerfTest, PerfTest) {
+  for (size_t i = 0; i < NELEMENTS(kVP9EncodePerfTestVectors); ++i) {
+    for (size_t j = 0; j < NELEMENTS(kEncodePerfTestSpeeds); ++j) {
+      for (size_t k = 0; k < NELEMENTS(kEncodePerfTestThreads); ++k) {
+        if (kVP9EncodePerfTestVectors[i].width < 512 &&
+            kEncodePerfTestThreads[k] > 1)
+          continue;
+        else if (kVP9EncodePerfTestVectors[i].width < 1024 &&
+                 kEncodePerfTestThreads[k] > 2)
+          continue;
+
+        set_threads(kEncodePerfTestThreads[k]);
+        SetUp();
+
+        const vpx_rational timebase = { 33333333, 1000000000 };
+        cfg_.g_timebase = timebase;
+        cfg_.rc_target_bitrate = kVP9EncodePerfTestVectors[i].bitrate;
+
+        init_flags_ = VPX_CODEC_USE_PSNR;
+
+        const unsigned frames = kVP9EncodePerfTestVectors[i].frames;
+        const char *video_name = kVP9EncodePerfTestVectors[i].name;
+        libvpx_test::I420VideoSource video(
+            video_name,
+            kVP9EncodePerfTestVectors[i].width,
+            kVP9EncodePerfTestVectors[i].height,
+            timebase.den, timebase.num, 0,
+            kVP9EncodePerfTestVectors[i].frames);
+        set_speed(kEncodePerfTestSpeeds[j]);
+
+        vpx_usec_timer t;
+        vpx_usec_timer_start(&t);
+
+        ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+        vpx_usec_timer_mark(&t);
+        const double elapsed_secs = vpx_usec_timer_elapsed(&t) / kUsecsInSec;
+        const double fps = frames / elapsed_secs;
+        const double minimum_psnr = min_psnr();
+        std::string display_name(video_name);
+        if (kEncodePerfTestThreads[k] > 1) {
+          char thread_count[32];
+          snprintf(thread_count, sizeof(thread_count), "_t-%d",
+                   kEncodePerfTestThreads[k]);
+          display_name += thread_count;
+        }
+
+        printf("{\n");
+        printf("\t\"type\" : \"encode_perf_test\",\n");
+        printf("\t\"version\" : \"%s\",\n", VERSION_STRING_NOSP);
+        printf("\t\"videoName\" : \"%s\",\n", display_name.c_str());
+        printf("\t\"encodeTimeSecs\" : %f,\n", elapsed_secs);
+        printf("\t\"totalFrames\" : %u,\n", frames);
+        printf("\t\"framesPerSecond\" : %f,\n", fps);
+        printf("\t\"minPsnr\" : %f,\n", minimum_psnr);
+        printf("\t\"speed\" : %d,\n", kEncodePerfTestSpeeds[j]);
+        printf("\t\"threads\" : %d\n", kEncodePerfTestThreads[k]);
+        printf("}\n");
+      }
+    }
+  }
+}
+
+VP9_INSTANTIATE_TEST_CASE(
+    VP9EncodePerfTest, ::testing::Values(::libvpx_test::kRealTime));
+}  // namespace
diff --git a/libvpx/libvpx/test/encode_test_driver.cc b/libvpx/libvpx/test/encode_test_driver.cc
new file mode 100644
index 0000000..b8c7371
--- /dev/null
+++ b/libvpx/libvpx/test/encode_test_driver.cc
@@ -0,0 +1,273 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/encode_test_driver.h"
+#include "test/register_state_check.h"
+#include "test/video_source.h"
+
+namespace libvpx_test {
+void Encoder::InitEncoder(VideoSource *video) {
+  vpx_codec_err_t res;
+  const vpx_image_t *img = video->img();
+
+  if (video->img() && !encoder_.priv) {
+    cfg_.g_w = img->d_w;
+    cfg_.g_h = img->d_h;
+    cfg_.g_timebase = video->timebase();
+    cfg_.rc_twopass_stats_in = stats_->buf();
+
+    res = vpx_codec_enc_init(&encoder_, CodecInterface(), &cfg_,
+                             init_flags_);
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+
+#if CONFIG_VP9_ENCODER
+    if (CodecInterface() == &vpx_codec_vp9_cx_algo) {
+      // Default to 1 tile column for VP9.
+      const int log2_tile_columns = 0;
+      res = vpx_codec_control_(&encoder_, VP9E_SET_TILE_COLUMNS,
+                               log2_tile_columns);
+      ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+    } else
+#endif
+    {
+#if CONFIG_VP8_ENCODER
+      ASSERT_EQ(&vpx_codec_vp8_cx_algo, CodecInterface())
+          << "Unknown Codec Interface";
+#endif
+    }
+  }
+}
+
+void Encoder::EncodeFrame(VideoSource *video, const unsigned long frame_flags) {
+  if (video->img())
+    EncodeFrameInternal(*video, frame_flags);
+  else
+    Flush();
+
+  // Handle twopass stats
+  CxDataIterator iter = GetCxData();
+
+  while (const vpx_codec_cx_pkt_t *pkt = iter.Next()) {
+    if (pkt->kind != VPX_CODEC_STATS_PKT)
+      continue;
+
+    stats_->Append(*pkt);
+  }
+}
+
+void Encoder::EncodeFrameInternal(const VideoSource &video,
+                                  const unsigned long frame_flags) {
+  vpx_codec_err_t res;
+  const vpx_image_t *img = video.img();
+
+  // Handle frame resizing
+  if (cfg_.g_w != img->d_w || cfg_.g_h != img->d_h) {
+    cfg_.g_w = img->d_w;
+    cfg_.g_h = img->d_h;
+    res = vpx_codec_enc_config_set(&encoder_, &cfg_);
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+  }
+
+  // Encode the frame
+  API_REGISTER_STATE_CHECK(
+      res = vpx_codec_encode(&encoder_, img, video.pts(), video.duration(),
+                             frame_flags, deadline_));
+  ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+}
+
+void Encoder::Flush() {
+  const vpx_codec_err_t res = vpx_codec_encode(&encoder_, NULL, 0, 0, 0,
+                                               deadline_);
+  if (!encoder_.priv)
+    ASSERT_EQ(VPX_CODEC_ERROR, res) << EncoderError();
+  else
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+}
+
+void EncoderTest::InitializeConfig() {
+  const vpx_codec_err_t res = codec_->DefaultEncoderConfig(&cfg_, 0);
+  dec_cfg_ = vpx_codec_dec_cfg_t();
+  ASSERT_EQ(VPX_CODEC_OK, res);
+}
+
+void EncoderTest::SetMode(TestMode mode) {
+  switch (mode) {
+    case kRealTime:
+      deadline_ = VPX_DL_REALTIME;
+      break;
+
+    case kOnePassGood:
+    case kTwoPassGood:
+      deadline_ = VPX_DL_GOOD_QUALITY;
+      break;
+
+    case kOnePassBest:
+    case kTwoPassBest:
+      deadline_ = VPX_DL_BEST_QUALITY;
+      break;
+
+    default:
+      ASSERT_TRUE(false) << "Unexpected mode " << mode;
+  }
+
+  if (mode == kTwoPassGood || mode == kTwoPassBest)
+    passes_ = 2;
+  else
+    passes_ = 1;
+}
+// The function should return "true" most of the time, therefore no early
+// break-out is implemented within the match checking process.
+static bool compare_img(const vpx_image_t *img1,
+                        const vpx_image_t *img2) {
+  bool match = (img1->fmt == img2->fmt) &&
+               (img1->cs == img2->cs) &&
+               (img1->d_w == img2->d_w) &&
+               (img1->d_h == img2->d_h);
+
+  const unsigned int width_y  = img1->d_w;
+  const unsigned int height_y = img1->d_h;
+  unsigned int i;
+  for (i = 0; i < height_y; ++i)
+    match = (memcmp(img1->planes[VPX_PLANE_Y] + i * img1->stride[VPX_PLANE_Y],
+                    img2->planes[VPX_PLANE_Y] + i * img2->stride[VPX_PLANE_Y],
+                    width_y) == 0) && match;
+  const unsigned int width_uv  = (img1->d_w + 1) >> 1;
+  const unsigned int height_uv = (img1->d_h + 1) >> 1;
+  for (i = 0; i <  height_uv; ++i)
+    match = (memcmp(img1->planes[VPX_PLANE_U] + i * img1->stride[VPX_PLANE_U],
+                    img2->planes[VPX_PLANE_U] + i * img2->stride[VPX_PLANE_U],
+                    width_uv) == 0) && match;
+  for (i = 0; i < height_uv; ++i)
+    match = (memcmp(img1->planes[VPX_PLANE_V] + i * img1->stride[VPX_PLANE_V],
+                    img2->planes[VPX_PLANE_V] + i * img2->stride[VPX_PLANE_V],
+                    width_uv) == 0) && match;
+  return match;
+}
+
+void EncoderTest::MismatchHook(const vpx_image_t* /*img1*/,
+                               const vpx_image_t* /*img2*/) {
+  ASSERT_TRUE(0) << "Encode/Decode mismatch found";
+}
+
+void EncoderTest::RunLoop(VideoSource *video) {
+  vpx_codec_dec_cfg_t dec_cfg = vpx_codec_dec_cfg_t();
+
+  stats_.Reset();
+
+  ASSERT_TRUE(passes_ == 1 || passes_ == 2);
+  for (unsigned int pass = 0; pass < passes_; pass++) {
+    last_pts_ = 0;
+
+    if (passes_ == 1)
+      cfg_.g_pass = VPX_RC_ONE_PASS;
+    else if (pass == 0)
+      cfg_.g_pass = VPX_RC_FIRST_PASS;
+    else
+      cfg_.g_pass = VPX_RC_LAST_PASS;
+
+    BeginPassHook(pass);
+    Encoder* const encoder = codec_->CreateEncoder(cfg_, deadline_, init_flags_,
+                                                   &stats_);
+    ASSERT_TRUE(encoder != NULL);
+
+    video->Begin();
+    encoder->InitEncoder(video);
+    ASSERT_FALSE(::testing::Test::HasFatalFailure());
+
+    unsigned long dec_init_flags = 0;  // NOLINT
+    // Use fragment decoder if encoder outputs partitions.
+    // NOTE: fragment decoder and partition encoder are only supported by VP8.
+    if (init_flags_ & VPX_CODEC_USE_OUTPUT_PARTITION)
+      dec_init_flags |= VPX_CODEC_USE_INPUT_FRAGMENTS;
+    Decoder* const decoder = codec_->CreateDecoder(dec_cfg, dec_init_flags, 0);
+    bool again;
+    for (again = true; again; video->Next()) {
+      again = (video->img() != NULL);
+
+      PreEncodeFrameHook(video);
+      PreEncodeFrameHook(video, encoder);
+      encoder->EncodeFrame(video, frame_flags_);
+
+      CxDataIterator iter = encoder->GetCxData();
+
+      bool has_cxdata = false;
+      bool has_dxdata = false;
+      while (const vpx_codec_cx_pkt_t *pkt = iter.Next()) {
+        pkt = MutateEncoderOutputHook(pkt);
+        again = true;
+        switch (pkt->kind) {
+          case VPX_CODEC_CX_FRAME_PKT:
+            has_cxdata = true;
+            if (decoder && DoDecode()) {
+              vpx_codec_err_t res_dec = decoder->DecodeFrame(
+                  (const uint8_t*)pkt->data.frame.buf, pkt->data.frame.sz);
+
+              if (!HandleDecodeResult(res_dec, *video, decoder))
+                break;
+
+              has_dxdata = true;
+            }
+            ASSERT_GE(pkt->data.frame.pts, last_pts_);
+            last_pts_ = pkt->data.frame.pts;
+            FramePktHook(pkt);
+            break;
+
+          case VPX_CODEC_PSNR_PKT:
+            PSNRPktHook(pkt);
+            break;
+
+          default:
+            break;
+        }
+      }
+
+      // Flush the decoder when there are no more fragments.
+      if ((init_flags_ & VPX_CODEC_USE_OUTPUT_PARTITION) && has_dxdata) {
+        const vpx_codec_err_t res_dec = decoder->DecodeFrame(NULL, 0);
+        if (!HandleDecodeResult(res_dec, *video, decoder))
+          break;
+      }
+
+      if (has_dxdata && has_cxdata) {
+        const vpx_image_t *img_enc = encoder->GetPreviewFrame();
+        DxDataIterator dec_iter = decoder->GetDxData();
+        const vpx_image_t *img_dec = dec_iter.Next();
+        if (img_enc && img_dec) {
+          const bool res = compare_img(img_enc, img_dec);
+          if (!res) {  // Mismatch
+            MismatchHook(img_enc, img_dec);
+          }
+        }
+        if (img_dec)
+          DecompressedFrameHook(*img_dec, video->pts());
+      }
+      if (!Continue())
+        break;
+    }
+
+    EndPassHook();
+
+    if (decoder)
+      delete decoder;
+    delete encoder;
+
+    if (!Continue())
+      break;
+  }
+}
+
+}  // namespace libvpx_test
diff --git a/libvpx/libvpx/test/encode_test_driver.h b/libvpx/libvpx/test/encode_test_driver.h
new file mode 100644
index 0000000..d14ddc7
--- /dev/null
+++ b/libvpx/libvpx/test/encode_test_driver.h
@@ -0,0 +1,278 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef TEST_ENCODE_TEST_DRIVER_H_
+#define TEST_ENCODE_TEST_DRIVER_H_
+
+#include <string>
+#include <vector>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#if CONFIG_VP8_ENCODER || CONFIG_VP9_ENCODER
+#include "vpx/vp8cx.h"
+#endif
+#include "vpx/vpx_encoder.h"
+
+namespace libvpx_test {
+
+class CodecFactory;
+class VideoSource;
+
+enum TestMode {
+  kRealTime,
+  kOnePassGood,
+  kOnePassBest,
+  kTwoPassGood,
+  kTwoPassBest
+};
+#define ALL_TEST_MODES ::testing::Values(::libvpx_test::kRealTime, \
+                                         ::libvpx_test::kOnePassGood, \
+                                         ::libvpx_test::kOnePassBest, \
+                                         ::libvpx_test::kTwoPassGood, \
+                                         ::libvpx_test::kTwoPassBest)
+
+#define ONE_PASS_TEST_MODES ::testing::Values(::libvpx_test::kRealTime, \
+                                              ::libvpx_test::kOnePassGood, \
+                                              ::libvpx_test::kOnePassBest)
+
+#define TWO_PASS_TEST_MODES ::testing::Values(::libvpx_test::kTwoPassGood, \
+                                              ::libvpx_test::kTwoPassBest)
+
+
+// Provides an object to handle the libvpx get_cx_data() iteration pattern
+class CxDataIterator {
+ public:
+  explicit CxDataIterator(vpx_codec_ctx_t *encoder)
+      : encoder_(encoder), iter_(NULL) {}
+
+  const vpx_codec_cx_pkt_t *Next() {
+    return vpx_codec_get_cx_data(encoder_, &iter_);
+  }
+
+ private:
+  vpx_codec_ctx_t  *encoder_;
+  vpx_codec_iter_t  iter_;
+};
+
+// Implements an in-memory store for libvpx twopass statistics
+class TwopassStatsStore {
+ public:
+  void Append(const vpx_codec_cx_pkt_t &pkt) {
+    buffer_.append(reinterpret_cast<char *>(pkt.data.twopass_stats.buf),
+                   pkt.data.twopass_stats.sz);
+  }
+
+  vpx_fixed_buf_t buf() {
+    const vpx_fixed_buf_t buf = { &buffer_[0], buffer_.size() };
+    return buf;
+  }
+
+  void Reset() {
+    buffer_.clear();
+  }
+
+ protected:
+  std::string  buffer_;
+};
+
+
+// Provides a simplified interface to manage one video encoding pass, given
+// a configuration and video source.
+//
+// TODO(jkoleszar): The exact services it provides and the appropriate
+// level of abstraction will be fleshed out as more tests are written.
+class Encoder {
+ public:
+  Encoder(vpx_codec_enc_cfg_t cfg, unsigned long deadline,
+          const unsigned long init_flags, TwopassStatsStore *stats)
+      : cfg_(cfg), deadline_(deadline), init_flags_(init_flags), stats_(stats) {
+    memset(&encoder_, 0, sizeof(encoder_));
+  }
+
+  virtual ~Encoder() {
+    vpx_codec_destroy(&encoder_);
+  }
+
+  CxDataIterator GetCxData() {
+    return CxDataIterator(&encoder_);
+  }
+
+  void InitEncoder(VideoSource *video);
+
+  const vpx_image_t *GetPreviewFrame() {
+    return vpx_codec_get_preview_frame(&encoder_);
+  }
+  // This is a thin wrapper around vpx_codec_encode(), so refer to
+  // vpx_encoder.h for its semantics.
+  void EncodeFrame(VideoSource *video, const unsigned long frame_flags);
+
+  // Convenience wrapper for EncodeFrame()
+  void EncodeFrame(VideoSource *video) {
+    EncodeFrame(video, 0);
+  }
+
+  void Control(int ctrl_id, int arg) {
+    const vpx_codec_err_t res = vpx_codec_control_(&encoder_, ctrl_id, arg);
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+  }
+
+  void Control(int ctrl_id, int *arg) {
+    const vpx_codec_err_t res = vpx_codec_control_(&encoder_, ctrl_id, arg);
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+  }
+
+  void Control(int ctrl_id, struct vpx_scaling_mode *arg) {
+    const vpx_codec_err_t res = vpx_codec_control_(&encoder_, ctrl_id, arg);
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+  }
+
+  void Control(int ctrl_id, struct vpx_svc_layer_id *arg) {
+    const vpx_codec_err_t res = vpx_codec_control_(&encoder_, ctrl_id, arg);
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+  }
+
+  void Control(int ctrl_id, struct vpx_svc_parameters *arg) {
+    const vpx_codec_err_t res = vpx_codec_control_(&encoder_, ctrl_id, arg);
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+  }
+#if CONFIG_VP8_ENCODER || CONFIG_VP9_ENCODER
+  void Control(int ctrl_id, vpx_active_map_t *arg) {
+    const vpx_codec_err_t res = vpx_codec_control_(&encoder_, ctrl_id, arg);
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+  }
+#endif
+
+  void Config(const vpx_codec_enc_cfg_t *cfg) {
+    const vpx_codec_err_t res = vpx_codec_enc_config_set(&encoder_, cfg);
+    ASSERT_EQ(VPX_CODEC_OK, res) << EncoderError();
+    cfg_ = *cfg;
+  }
+
+  void set_deadline(unsigned long deadline) {
+    deadline_ = deadline;
+  }
+
+ protected:
+  virtual vpx_codec_iface_t* CodecInterface() const = 0;
+
+  const char *EncoderError() {
+    const char *detail = vpx_codec_error_detail(&encoder_);
+    return detail ? detail : vpx_codec_error(&encoder_);
+  }
+
+  // Encode an image
+  void EncodeFrameInternal(const VideoSource &video,
+                           const unsigned long frame_flags);
+
+  // Flush the encoder on EOS
+  void Flush();
+
+  vpx_codec_ctx_t      encoder_;
+  vpx_codec_enc_cfg_t  cfg_;
+  unsigned long        deadline_;
+  unsigned long        init_flags_;
+  TwopassStatsStore   *stats_;
+};
+
+// Common test functionality for all Encoder tests.
+//
+// This class is a mixin which provides the main loop common to all
+// encoder tests. It provides hooks which can be overridden by subclasses
+// to implement each test's specific behavior, while centralizing the bulk
+// of the boilerplate. Note that it doesn't inherit the gtest testing
+// classes directly, so that tests can be parameterized differently.
+class EncoderTest {
+ protected:
+  explicit EncoderTest(const CodecFactory *codec)
+      : codec_(codec), abort_(false), init_flags_(0), frame_flags_(0),
+        last_pts_(0) {
+    // Default to 1 thread.
+    cfg_.g_threads = 1;
+  }
+
+  virtual ~EncoderTest() {}
+
+  // Initialize the cfg_ member with the default configuration.
+  void InitializeConfig();
+
+  // Map the TestMode enum to the deadline_ and passes_ variables.
+  void SetMode(TestMode mode);
+
+  // Set encoder flag.
+  void set_init_flags(unsigned long flag) {  // NOLINT(runtime/int)
+    init_flags_ = flag;
+  }
+
+  // Main loop
+  virtual void RunLoop(VideoSource *video);
+
+  // Hook to be called at the beginning of a pass.
+  virtual void BeginPassHook(unsigned int /*pass*/) {}
+
+  // Hook to be called at the end of a pass.
+  virtual void EndPassHook() {}
+
+  // Hook to be called before encoding a frame.
+  virtual void PreEncodeFrameHook(VideoSource* /*video*/) {}
+  virtual void PreEncodeFrameHook(VideoSource* /*video*/,
+                                  Encoder* /*encoder*/) {}
+
+  // Hook to be called on every compressed data packet.
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t* /*pkt*/) {}
+
+  // Hook to be called on every PSNR packet.
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t* /*pkt*/) {}
+
+  // Hook to determine whether the encode loop should continue.
+  virtual bool Continue() const {
+    return !(::testing::Test::HasFatalFailure() || abort_);
+  }
+
+  const CodecFactory   *codec_;
+  // Hook to determine whether to decode frame after encoding
+  virtual bool DoDecode() const { return 1; }
+
+  // Hook to handle encode/decode mismatch
+  virtual void MismatchHook(const vpx_image_t *img1,
+                            const vpx_image_t *img2);
+
+  // Hook to be called on every decompressed frame.
+  virtual void DecompressedFrameHook(const vpx_image_t& /*img*/,
+                                     vpx_codec_pts_t /*pts*/) {}
+
+  // Hook to be called to handle decode result. Return true to continue.
+  virtual bool HandleDecodeResult(const vpx_codec_err_t res_dec,
+                                  const VideoSource& /*video*/,
+                                  Decoder *decoder) {
+    EXPECT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+    return VPX_CODEC_OK == res_dec;
+  }
+
+  // Hook that can modify the encoder's output data
+  virtual const vpx_codec_cx_pkt_t *MutateEncoderOutputHook(
+      const vpx_codec_cx_pkt_t *pkt) {
+    return pkt;
+  }
+
+  bool                 abort_;
+  vpx_codec_enc_cfg_t  cfg_;
+  vpx_codec_dec_cfg_t  dec_cfg_;
+  unsigned int         passes_;
+  unsigned long        deadline_;
+  TwopassStatsStore    stats_;
+  unsigned long        init_flags_;
+  unsigned long        frame_flags_;
+  vpx_codec_pts_t      last_pts_;
+};
+
+}  // namespace libvpx_test
+
+#endif  // TEST_ENCODE_TEST_DRIVER_H_
diff --git a/libvpx/libvpx/test/error_resilience_test.cc b/libvpx/libvpx/test/error_resilience_test.cc
new file mode 100644
index 0000000..00a095c
--- /dev/null
+++ b/libvpx/libvpx/test/error_resilience_test.cc
@@ -0,0 +1,599 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+const int kMaxErrorFrames = 12;
+const int kMaxDroppableFrames = 12;
+
+class ErrorResilienceTestLarge : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, bool> {
+ protected:
+  ErrorResilienceTestLarge()
+      : EncoderTest(GET_PARAM(0)),
+        svc_support_(GET_PARAM(2)),
+        psnr_(0.0),
+        nframes_(0),
+        mismatch_psnr_(0.0),
+        mismatch_nframes_(0),
+        encoding_mode_(GET_PARAM(1)) {
+    Reset();
+  }
+
+  virtual ~ErrorResilienceTestLarge() {}
+
+  void Reset() {
+    error_nframes_ = 0;
+    droppable_nframes_ = 0;
+    pattern_switch_ = 0;
+  }
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    psnr_ = 0.0;
+    nframes_ = 0;
+    mismatch_psnr_ = 0.0;
+    mismatch_nframes_ = 0;
+  }
+
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    psnr_ += pkt->data.psnr.psnr[0];
+    nframes_++;
+  }
+
+  //
+  // Frame flags and layer id for temporal layers.
+  // For two layers, test pattern is:
+  //   1     3
+  // 0    2     .....
+  // LAST is updated on base/layer 0, GOLDEN  updated on layer 1.
+  // Non-zero pattern_switch parameter means pattern will switch to
+  // not using LAST for frame_num >= pattern_switch.
+  int SetFrameFlags(int frame_num,
+                    int num_temp_layers,
+                    int pattern_switch) {
+    int frame_flags = 0;
+    if (num_temp_layers == 2) {
+        if (frame_num % 2 == 0) {
+          if (frame_num < pattern_switch || pattern_switch == 0) {
+            // Layer 0: predict from LAST and ARF, update LAST.
+            frame_flags = VP8_EFLAG_NO_REF_GF |
+                          VP8_EFLAG_NO_UPD_GF |
+                          VP8_EFLAG_NO_UPD_ARF;
+          } else {
+            // Layer 0: predict from GF and ARF, update GF.
+            frame_flags = VP8_EFLAG_NO_REF_LAST |
+                          VP8_EFLAG_NO_UPD_LAST |
+                          VP8_EFLAG_NO_UPD_ARF;
+          }
+        } else {
+          if (frame_num < pattern_switch || pattern_switch == 0) {
+            // Layer 1: predict from L, GF, and ARF, update GF.
+            frame_flags = VP8_EFLAG_NO_UPD_ARF |
+                          VP8_EFLAG_NO_UPD_LAST;
+          } else {
+            // Layer 1: predict from GF and ARF, update GF.
+            frame_flags = VP8_EFLAG_NO_REF_LAST |
+                          VP8_EFLAG_NO_UPD_LAST |
+                          VP8_EFLAG_NO_UPD_ARF;
+          }
+        }
+    }
+    return frame_flags;
+  }
+
+  virtual void PreEncodeFrameHook(libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder * /*encoder*/) {
+    frame_flags_ &= ~(VP8_EFLAG_NO_UPD_LAST |
+                      VP8_EFLAG_NO_UPD_GF |
+                      VP8_EFLAG_NO_UPD_ARF);
+    // For temporal layer case.
+    if (cfg_.ts_number_layers > 1) {
+      frame_flags_ = SetFrameFlags(video->frame(),
+                                   cfg_.ts_number_layers,
+                                   pattern_switch_);
+      for (unsigned int i = 0; i < droppable_nframes_; ++i) {
+        if (droppable_frames_[i] == video->frame()) {
+          std::cout << "Encoding droppable frame: "
+                    << droppable_frames_[i] << "\n";
+        }
+      }
+    } else {
+       if (droppable_nframes_ > 0 &&
+         (cfg_.g_pass == VPX_RC_LAST_PASS || cfg_.g_pass == VPX_RC_ONE_PASS)) {
+         for (unsigned int i = 0; i < droppable_nframes_; ++i) {
+           if (droppable_frames_[i] == video->frame()) {
+             std::cout << "Encoding droppable frame: "
+                       << droppable_frames_[i] << "\n";
+             frame_flags_ |= (VP8_EFLAG_NO_UPD_LAST |
+                              VP8_EFLAG_NO_UPD_GF |
+                              VP8_EFLAG_NO_UPD_ARF);
+             return;
+           }
+         }
+       }
+    }
+  }
+
+  double GetAveragePsnr() const {
+    if (nframes_)
+      return psnr_ / nframes_;
+    return 0.0;
+  }
+
+  double GetAverageMismatchPsnr() const {
+    if (mismatch_nframes_)
+      return mismatch_psnr_ / mismatch_nframes_;
+    return 0.0;
+  }
+
+  virtual bool DoDecode() const {
+    if (error_nframes_ > 0 &&
+        (cfg_.g_pass == VPX_RC_LAST_PASS || cfg_.g_pass == VPX_RC_ONE_PASS)) {
+      for (unsigned int i = 0; i < error_nframes_; ++i) {
+        if (error_frames_[i] == nframes_ - 1) {
+          std::cout << "             Skipping decoding frame: "
+                    << error_frames_[i] << "\n";
+          return 0;
+        }
+      }
+    }
+    return 1;
+  }
+
+  virtual void MismatchHook(const vpx_image_t *img1,
+                            const vpx_image_t *img2) {
+    double mismatch_psnr = compute_psnr(img1, img2);
+    mismatch_psnr_ += mismatch_psnr;
+    ++mismatch_nframes_;
+    // std::cout << "Mismatch frame psnr: " << mismatch_psnr << "\n";
+  }
+
+  void SetErrorFrames(int num, unsigned int *list) {
+    if (num > kMaxErrorFrames)
+      num = kMaxErrorFrames;
+    else if (num < 0)
+      num = 0;
+    error_nframes_ = num;
+    for (unsigned int i = 0; i < error_nframes_; ++i)
+      error_frames_[i] = list[i];
+  }
+
+  void SetDroppableFrames(int num, unsigned int *list) {
+    if (num > kMaxDroppableFrames)
+      num = kMaxDroppableFrames;
+    else if (num < 0)
+      num = 0;
+    droppable_nframes_ = num;
+    for (unsigned int i = 0; i < droppable_nframes_; ++i)
+      droppable_frames_[i] = list[i];
+  }
+
+  unsigned int GetMismatchFrames() {
+    return mismatch_nframes_;
+  }
+
+  void SetPatternSwitch(int frame_switch) {
+     pattern_switch_ = frame_switch;
+   }
+
+  bool svc_support_;
+
+ private:
+  double psnr_;
+  unsigned int nframes_;
+  unsigned int error_nframes_;
+  unsigned int droppable_nframes_;
+  unsigned int pattern_switch_;
+  double mismatch_psnr_;
+  unsigned int mismatch_nframes_;
+  unsigned int error_frames_[kMaxErrorFrames];
+  unsigned int droppable_frames_[kMaxDroppableFrames];
+  libvpx_test::TestMode encoding_mode_;
+};
+
+TEST_P(ErrorResilienceTestLarge, OnVersusOff) {
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.g_lag_in_frames = 10;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 30);
+
+  // Error resilient mode OFF.
+  cfg_.g_error_resilient = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_resilience_off = GetAveragePsnr();
+  EXPECT_GT(psnr_resilience_off, 25.0);
+
+  // Error resilient mode ON.
+  cfg_.g_error_resilient = 1;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_resilience_on = GetAveragePsnr();
+  EXPECT_GT(psnr_resilience_on, 25.0);
+
+  // Test that turning on error resilient mode hurts by 10% at most.
+  if (psnr_resilience_off > 0.0) {
+    const double psnr_ratio = psnr_resilience_on / psnr_resilience_off;
+    EXPECT_GE(psnr_ratio, 0.9);
+    EXPECT_LE(psnr_ratio, 1.1);
+  }
+}
+
+// Check for successful decoding and no encoder/decoder mismatch
+// if we lose (i.e., drop before decoding) a set of droppable
+// frames (i.e., frames that don't update any reference buffers).
+// Check both isolated and consecutive loss.
+TEST_P(ErrorResilienceTestLarge, DropFramesWithoutRecovery) {
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 500;
+  // FIXME(debargha): Fix this to work for any lag.
+  // Currently this test only works for lag = 0
+  cfg_.g_lag_in_frames = 0;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 40);
+
+  // Error resilient mode ON.
+  cfg_.g_error_resilient = 1;
+  cfg_.kf_mode = VPX_KF_DISABLED;
+
+  // Set an arbitrary set of error frames same as droppable frames.
+  // In addition to isolated loss/drop, add a long consecutive series
+  // (of size 9) of dropped frames.
+  unsigned int num_droppable_frames = 11;
+  unsigned int droppable_frame_list[] = {5, 16, 22, 23, 24, 25, 26, 27, 28,
+                                         29, 30};
+  SetDroppableFrames(num_droppable_frames, droppable_frame_list);
+  SetErrorFrames(num_droppable_frames, droppable_frame_list);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  // Test that no mismatches have been found
+  std::cout << "             Mismatch frames: "
+            << GetMismatchFrames() << "\n";
+  EXPECT_EQ(GetMismatchFrames(), (unsigned int) 0);
+
+  // Reset previously set of error/droppable frames.
+  Reset();
+
+#if 0
+  // TODO(jkoleszar): This test is disabled for the time being as too
+  // sensitive. It's not clear how to set a reasonable threshold for
+  // this behavior.
+
+  // Now set an arbitrary set of error frames that are non-droppable
+  unsigned int num_error_frames = 3;
+  unsigned int error_frame_list[] = {3, 10, 20};
+  SetErrorFrames(num_error_frames, error_frame_list);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  // Test that dropping an arbitrary set of inter frames does not hurt too much
+  // Note the Average Mismatch PSNR is the average of the PSNR between
+  // decoded frame and encoder's version of the same frame for all frames
+  // with mismatch.
+  const double psnr_resilience_mismatch = GetAverageMismatchPsnr();
+  std::cout << "             Mismatch PSNR: "
+            << psnr_resilience_mismatch << "\n";
+  EXPECT_GT(psnr_resilience_mismatch, 20.0);
+#endif
+}
+
+// Check for successful decoding and no encoder/decoder mismatch
+// if we lose (i.e., drop before decoding) the enhancement layer frames for a
+// two layer temporal pattern. The base layer does not predict from the top
+// layer, so successful decoding is expected.
+TEST_P(ErrorResilienceTestLarge, 2LayersDropEnhancement) {
+  // This test doesn't run if SVC is not supported.
+  if (!svc_support_)
+    return;
+
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 500;
+  cfg_.g_lag_in_frames = 0;
+
+  cfg_.rc_end_usage = VPX_CBR;
+  // 2 Temporal layers, no spatial layers, CBR mode.
+  cfg_.ss_number_layers = 1;
+  cfg_.ts_number_layers = 2;
+  cfg_.ts_rate_decimator[0] = 2;
+  cfg_.ts_rate_decimator[1] = 1;
+  cfg_.ts_periodicity = 2;
+  cfg_.ts_target_bitrate[0] = 60 * cfg_.rc_target_bitrate / 100;
+  cfg_.ts_target_bitrate[1] = cfg_.rc_target_bitrate;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 40);
+
+  // Error resilient mode ON.
+  cfg_.g_error_resilient = 1;
+  cfg_.kf_mode = VPX_KF_DISABLED;
+  SetPatternSwitch(0);
+
+  // The odd frames are the enhancement layer for 2 layer pattern, so set
+  // those frames as droppable. Drop the last 7 frames.
+  unsigned int num_droppable_frames = 7;
+  unsigned int droppable_frame_list[] = {27, 29, 31, 33, 35, 37, 39};
+  SetDroppableFrames(num_droppable_frames, droppable_frame_list);
+  SetErrorFrames(num_droppable_frames, droppable_frame_list);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  // Test that no mismatches have been found
+  std::cout << "             Mismatch frames: "
+            << GetMismatchFrames() << "\n";
+  EXPECT_EQ(GetMismatchFrames(), (unsigned int) 0);
+
+  // Reset previously set of error/droppable frames.
+  Reset();
+}
+
+// Check for successful decoding and no encoder/decoder mismatch
+// for a two layer temporal pattern, where at some point in the
+// sequence, the LAST ref is not used anymore.
+TEST_P(ErrorResilienceTestLarge, 2LayersNoRefLast) {
+  // This test doesn't run if SVC is not supported.
+  if (!svc_support_)
+    return;
+
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 500;
+  cfg_.g_lag_in_frames = 0;
+
+  cfg_.rc_end_usage = VPX_CBR;
+  // 2 Temporal layers, no spatial layers, CBR mode.
+  cfg_.ss_number_layers = 1;
+  cfg_.ts_number_layers = 2;
+  cfg_.ts_rate_decimator[0] = 2;
+  cfg_.ts_rate_decimator[1] = 1;
+  cfg_.ts_periodicity = 2;
+  cfg_.ts_target_bitrate[0] = 60 * cfg_.rc_target_bitrate / 100;
+  cfg_.ts_target_bitrate[1] = cfg_.rc_target_bitrate;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 100);
+
+  // Error resilient mode ON.
+  cfg_.g_error_resilient = 1;
+  cfg_.kf_mode = VPX_KF_DISABLED;
+  SetPatternSwitch(60);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  // Test that no mismatches have been found
+  std::cout << "             Mismatch frames: "
+            << GetMismatchFrames() << "\n";
+  EXPECT_EQ(GetMismatchFrames(), (unsigned int) 0);
+
+  // Reset previously set of error/droppable frames.
+  Reset();
+}
+
+class ErrorResilienceTestLargeCodecControls : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ protected:
+  ErrorResilienceTestLargeCodecControls()
+      : EncoderTest(GET_PARAM(0)),
+        encoding_mode_(GET_PARAM(1)) {
+    Reset();
+  }
+
+  virtual ~ErrorResilienceTestLargeCodecControls() {}
+
+  void Reset() {
+    last_pts_ = 0;
+    tot_frame_number_ = 0;
+    // For testing up to 3 layers.
+    for (int i = 0; i < 3; ++i) {
+      bits_total_[i] = 0;
+    }
+    duration_ = 0.0;
+  }
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+  }
+
+  //
+  // Frame flags and layer id for temporal layers.
+  //
+
+  // For two layers, test pattern is:
+  //   1     3
+  // 0    2     .....
+  // For three layers, test pattern is:
+  //   1      3    5      7
+  //      2           6
+  // 0          4            ....
+  // LAST is always update on base/layer 0, GOLDEN is updated on layer 1,
+  // and ALTREF is updated on top layer for 3 layer pattern.
+  int SetFrameFlags(int frame_num, int num_temp_layers) {
+    int frame_flags = 0;
+    if (num_temp_layers == 2) {
+      if (frame_num % 2 == 0) {
+        // Layer 0: predict from L and ARF, update L.
+        frame_flags = VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_UPD_GF |
+                      VP8_EFLAG_NO_UPD_ARF;
+      } else {
+        // Layer 1: predict from L, G and ARF, and update G.
+        frame_flags = VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST |
+                      VP8_EFLAG_NO_UPD_ENTROPY;
+      }
+    } else if (num_temp_layers == 3) {
+      if (frame_num % 4 == 0) {
+        // Layer 0: predict from L, update L.
+        frame_flags = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF |
+                      VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF;
+      } else if ((frame_num - 2) % 4 == 0) {
+        // Layer 1: predict from L, G,  update G.
+        frame_flags = VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_NO_UPD_LAST |
+                      VP8_EFLAG_NO_REF_ARF;
+      }  else if ((frame_num - 1) % 2 == 0) {
+        // Layer 2: predict from L, G, ARF; update ARG.
+        frame_flags = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_LAST;
+      }
+    }
+    return frame_flags;
+  }
+
+  int SetLayerId(int frame_num, int num_temp_layers) {
+    int layer_id = 0;
+    if (num_temp_layers == 2) {
+      if (frame_num % 2 == 0) {
+        layer_id = 0;
+      } else {
+         layer_id = 1;
+      }
+    } else if (num_temp_layers == 3) {
+      if (frame_num % 4 == 0) {
+        layer_id = 0;
+      } else if ((frame_num - 2) % 4 == 0) {
+        layer_id = 1;
+      } else if ((frame_num - 1) % 2 == 0) {
+        layer_id = 2;
+      }
+    }
+    return layer_id;
+  }
+
+  virtual void PreEncodeFrameHook(libvpx_test::VideoSource *video,
+                                  libvpx_test::Encoder *encoder) {
+    if (cfg_.ts_number_layers > 1) {
+        int layer_id = SetLayerId(video->frame(), cfg_.ts_number_layers);
+        int frame_flags = SetFrameFlags(video->frame(), cfg_.ts_number_layers);
+        if (video->frame() > 0) {
+          encoder->Control(VP8E_SET_TEMPORAL_LAYER_ID, layer_id);
+          encoder->Control(VP8E_SET_FRAME_FLAGS, frame_flags);
+        }
+       const vpx_rational_t tb = video->timebase();
+       timebase_ = static_cast<double>(tb.num) / tb.den;
+       duration_ = 0;
+       return;
+    }
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    // Time since last timestamp = duration.
+    vpx_codec_pts_t duration = pkt->data.frame.pts - last_pts_;
+    if (duration > 1) {
+      // Update counter for total number of frames (#frames input to encoder).
+      // Needed for setting the proper layer_id below.
+      tot_frame_number_ += static_cast<int>(duration - 1);
+    }
+    int layer = SetLayerId(tot_frame_number_, cfg_.ts_number_layers);
+    const size_t frame_size_in_bits = pkt->data.frame.sz * 8;
+    // Update the total encoded bits. For temporal layers, update the cumulative
+    // encoded bits per layer.
+    for (int i = layer; i < static_cast<int>(cfg_.ts_number_layers); ++i) {
+      bits_total_[i] += frame_size_in_bits;
+    }
+    // Update the most recent pts.
+    last_pts_ = pkt->data.frame.pts;
+    ++tot_frame_number_;
+  }
+
+  virtual void EndPassHook(void) {
+    duration_ = (last_pts_ + 1) * timebase_;
+    if (cfg_.ts_number_layers  > 1) {
+      for (int layer = 0; layer < static_cast<int>(cfg_.ts_number_layers);
+          ++layer) {
+        if (bits_total_[layer]) {
+          // Effective file datarate:
+          effective_datarate_[layer] = (bits_total_[layer] / 1000.0) / duration_;
+        }
+      }
+    }
+  }
+
+  double effective_datarate_[3];
+   private:
+    libvpx_test::TestMode encoding_mode_;
+    vpx_codec_pts_t last_pts_;
+    double timebase_;
+    int64_t bits_total_[3];
+    double duration_;
+    int tot_frame_number_;
+  };
+
+// Check two codec controls used for:
+// (1) for setting temporal layer id, and (2) for settings encoder flags.
+// This test invokes those controls for each frame, and verifies encoder/decoder
+// mismatch and basic rate control response.
+// TODO(marpan): Maybe move this test to datarate_test.cc.
+TEST_P(ErrorResilienceTestLargeCodecControls, CodecControl3TemporalLayers) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_min_quantizer = 2;
+  cfg_.rc_max_quantizer = 56;
+  cfg_.rc_end_usage = VPX_CBR;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.kf_mode = VPX_KF_DISABLED;
+  cfg_.g_error_resilient = 1;
+
+  // 3 Temporal layers. Framerate decimation (4, 2, 1).
+  cfg_.ts_number_layers = 3;
+  cfg_.ts_rate_decimator[0] = 4;
+  cfg_.ts_rate_decimator[1] = 2;
+  cfg_.ts_rate_decimator[2] = 1;
+  cfg_.ts_periodicity = 4;
+  cfg_.ts_layer_id[0] = 0;
+  cfg_.ts_layer_id[1] = 2;
+  cfg_.ts_layer_id[2] = 1;
+  cfg_.ts_layer_id[3] = 2;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 200);
+  for (int i = 200; i <= 800; i += 200) {
+    cfg_.rc_target_bitrate = i;
+    Reset();
+    // 40-20-40 bitrate allocation for 3 temporal layers.
+    cfg_.ts_target_bitrate[0] = 40 * cfg_.rc_target_bitrate / 100;
+    cfg_.ts_target_bitrate[1] = 60 * cfg_.rc_target_bitrate / 100;
+    cfg_.ts_target_bitrate[2] = cfg_.rc_target_bitrate;
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    for (int j = 0; j < static_cast<int>(cfg_.ts_number_layers); ++j) {
+      ASSERT_GE(effective_datarate_[j], cfg_.ts_target_bitrate[j] * 0.75)
+          << " The datarate for the file is lower than target by too much, "
+              "for layer: " << j;
+      ASSERT_LE(effective_datarate_[j], cfg_.ts_target_bitrate[j] * 1.25)
+          << " The datarate for the file is greater than target by too much, "
+              "for layer: " << j;
+    }
+  }
+}
+
+VP8_INSTANTIATE_TEST_CASE(ErrorResilienceTestLarge, ONE_PASS_TEST_MODES,
+                          ::testing::Values(true));
+VP8_INSTANTIATE_TEST_CASE(ErrorResilienceTestLargeCodecControls,
+                          ONE_PASS_TEST_MODES);
+VP9_INSTANTIATE_TEST_CASE(ErrorResilienceTestLarge, ONE_PASS_TEST_MODES,
+                          ::testing::Values(true));
+}  // namespace
diff --git a/libvpx/libvpx/test/examples.sh b/libvpx/libvpx/test/examples.sh
new file mode 100755
index 0000000..39f7e39
--- /dev/null
+++ b/libvpx/libvpx/test/examples.sh
@@ -0,0 +1,29 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file runs all of the tests for the libvpx examples.
+##
+. $(dirname $0)/tools_common.sh
+
+example_tests=$(ls $(dirname $0)/*.sh)
+
+# List of script names to exclude.
+exclude_list="examples tools_common"
+
+# Filter out the scripts in $exclude_list.
+for word in ${exclude_list}; do
+  example_tests=$(filter_strings "${example_tests}" "${word}" exclude)
+done
+
+for test in ${example_tests}; do
+  # Source each test script so that exporting variables can be avoided.
+  VPX_TEST_NAME="$(basename ${test%.*})"
+  . "${test}"
+done
diff --git a/libvpx/libvpx/test/external_frame_buffer_test.cc b/libvpx/libvpx/test/external_frame_buffer_test.cc
new file mode 100644
index 0000000..2570f44
--- /dev/null
+++ b/libvpx/libvpx/test/external_frame_buffer_test.cc
@@ -0,0 +1,493 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/md5_helper.h"
+#include "test/test_vectors.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+
+namespace {
+
+const int kVideoNameParam = 1;
+
+struct ExternalFrameBuffer {
+  uint8_t *data;
+  size_t size;
+  int in_use;
+};
+
+// Class to manipulate a list of external frame buffers.
+class ExternalFrameBufferList {
+ public:
+  ExternalFrameBufferList()
+      : num_buffers_(0),
+        ext_fb_list_(NULL) {}
+
+  virtual ~ExternalFrameBufferList() {
+    for (int i = 0; i < num_buffers_; ++i) {
+      delete [] ext_fb_list_[i].data;
+    }
+    delete [] ext_fb_list_;
+  }
+
+  // Creates the list to hold the external buffers. Returns true on success.
+  bool CreateBufferList(int num_buffers) {
+    if (num_buffers < 0)
+      return false;
+
+    num_buffers_ = num_buffers;
+    ext_fb_list_ = new ExternalFrameBuffer[num_buffers_];
+    EXPECT_TRUE(ext_fb_list_ != NULL);
+    memset(ext_fb_list_, 0, sizeof(ext_fb_list_[0]) * num_buffers_);
+    return true;
+  }
+
+  // Searches the frame buffer list for a free frame buffer. Makes sure
+  // that the frame buffer is at least |min_size| in bytes. Marks that the
+  // frame buffer is in use by libvpx. Finally sets |fb| to point to the
+  // external frame buffer. Returns < 0 on an error.
+  int GetFreeFrameBuffer(size_t min_size, vpx_codec_frame_buffer_t *fb) {
+    EXPECT_TRUE(fb != NULL);
+    const int idx = FindFreeBufferIndex();
+    if (idx == num_buffers_)
+      return -1;
+
+    if (ext_fb_list_[idx].size < min_size) {
+      delete [] ext_fb_list_[idx].data;
+      ext_fb_list_[idx].data = new uint8_t[min_size];
+      memset(ext_fb_list_[idx].data, 0, min_size);
+      ext_fb_list_[idx].size = min_size;
+    }
+
+    SetFrameBuffer(idx, fb);
+    return 0;
+  }
+
+  // Test function that will not allocate any data for the frame buffer.
+  // Returns < 0 on an error.
+  int GetZeroFrameBuffer(size_t min_size, vpx_codec_frame_buffer_t *fb) {
+    EXPECT_TRUE(fb != NULL);
+    const int idx = FindFreeBufferIndex();
+    if (idx == num_buffers_)
+      return -1;
+
+    if (ext_fb_list_[idx].size < min_size) {
+      delete [] ext_fb_list_[idx].data;
+      ext_fb_list_[idx].data = NULL;
+      ext_fb_list_[idx].size = min_size;
+    }
+
+    SetFrameBuffer(idx, fb);
+    return 0;
+  }
+
+  // Marks the external frame buffer that |fb| is pointing to as free.
+  // Returns < 0 on an error.
+  int ReturnFrameBuffer(vpx_codec_frame_buffer_t *fb) {
+    if (fb == NULL) {
+      EXPECT_TRUE(fb != NULL);
+      return -1;
+    }
+    ExternalFrameBuffer *const ext_fb =
+        reinterpret_cast<ExternalFrameBuffer*>(fb->priv);
+    if (ext_fb == NULL) {
+      EXPECT_TRUE(ext_fb != NULL);
+      return -1;
+    }
+    EXPECT_EQ(1, ext_fb->in_use);
+    ext_fb->in_use = 0;
+    return 0;
+  }
+
+  // Checks that the ximage data is contained within the external frame buffer
+  // private data passed back in the ximage.
+  void CheckXImageFrameBuffer(const vpx_image_t *img) {
+    if (img->fb_priv != NULL) {
+      const struct ExternalFrameBuffer *const ext_fb =
+          reinterpret_cast<ExternalFrameBuffer*>(img->fb_priv);
+
+      ASSERT_TRUE(img->planes[0] >= ext_fb->data &&
+                  img->planes[0] < (ext_fb->data + ext_fb->size));
+    }
+  }
+
+ private:
+  // Returns the index of the first free frame buffer. Returns |num_buffers_|
+  // if there are no free frame buffers.
+  int FindFreeBufferIndex() {
+    int i;
+    // Find a free frame buffer.
+    for (i = 0; i < num_buffers_; ++i) {
+      if (!ext_fb_list_[i].in_use)
+        break;
+    }
+    return i;
+  }
+
+  // Sets |fb| to an external frame buffer. idx is the index into the frame
+  // buffer list.
+  void SetFrameBuffer(int idx, vpx_codec_frame_buffer_t *fb) {
+    ASSERT_TRUE(fb != NULL);
+    fb->data = ext_fb_list_[idx].data;
+    fb->size = ext_fb_list_[idx].size;
+    ASSERT_EQ(0, ext_fb_list_[idx].in_use);
+    ext_fb_list_[idx].in_use = 1;
+    fb->priv = &ext_fb_list_[idx];
+  }
+
+  int num_buffers_;
+  ExternalFrameBuffer *ext_fb_list_;
+};
+
+#if CONFIG_WEBM_IO
+
+// Callback used by libvpx to request the application to return a frame
+// buffer of at least |min_size| in bytes.
+int get_vp9_frame_buffer(void *user_priv, size_t min_size,
+                         vpx_codec_frame_buffer_t *fb) {
+  ExternalFrameBufferList *const fb_list =
+      reinterpret_cast<ExternalFrameBufferList*>(user_priv);
+  return fb_list->GetFreeFrameBuffer(min_size, fb);
+}
+
+// Callback used by libvpx to tell the application that |fb| is not needed
+// anymore.
+int release_vp9_frame_buffer(void *user_priv,
+                             vpx_codec_frame_buffer_t *fb) {
+  ExternalFrameBufferList *const fb_list =
+      reinterpret_cast<ExternalFrameBufferList*>(user_priv);
+  return fb_list->ReturnFrameBuffer(fb);
+}
+
+// Callback will not allocate data for frame buffer.
+int get_vp9_zero_frame_buffer(void *user_priv, size_t min_size,
+                              vpx_codec_frame_buffer_t *fb) {
+  ExternalFrameBufferList *const fb_list =
+      reinterpret_cast<ExternalFrameBufferList*>(user_priv);
+  return fb_list->GetZeroFrameBuffer(min_size, fb);
+}
+
+// Callback will allocate one less byte than |min_size|.
+int get_vp9_one_less_byte_frame_buffer(void *user_priv, size_t min_size,
+                                       vpx_codec_frame_buffer_t *fb) {
+  ExternalFrameBufferList *const fb_list =
+      reinterpret_cast<ExternalFrameBufferList*>(user_priv);
+  return fb_list->GetFreeFrameBuffer(min_size - 1, fb);
+}
+
+// Callback will not release the external frame buffer.
+int do_not_release_vp9_frame_buffer(void *user_priv,
+                                    vpx_codec_frame_buffer_t *fb) {
+  (void)user_priv;
+  (void)fb;
+  return 0;
+}
+
+#endif  // CONFIG_WEBM_IO
+
+// Class for testing passing in external frame buffers to libvpx.
+class ExternalFrameBufferMD5Test
+    : public ::libvpx_test::DecoderTest,
+      public ::libvpx_test::CodecTestWithParam<const char*> {
+ protected:
+  ExternalFrameBufferMD5Test()
+      : DecoderTest(GET_PARAM(::libvpx_test::kCodecFactoryParam)),
+        md5_file_(NULL),
+        num_buffers_(0) {}
+
+  virtual ~ExternalFrameBufferMD5Test() {
+    if (md5_file_ != NULL)
+      fclose(md5_file_);
+  }
+
+  virtual void PreDecodeFrameHook(
+      const libvpx_test::CompressedVideoSource &video,
+      libvpx_test::Decoder *decoder) {
+    if (num_buffers_ > 0 && video.frame_number() == 0) {
+      // Have libvpx use frame buffers we create.
+      ASSERT_TRUE(fb_list_.CreateBufferList(num_buffers_));
+      ASSERT_EQ(VPX_CODEC_OK,
+                decoder->SetFrameBufferFunctions(
+                    GetVP9FrameBuffer, ReleaseVP9FrameBuffer, this));
+    }
+  }
+
+  void OpenMD5File(const std::string &md5_file_name_) {
+    md5_file_ = libvpx_test::OpenTestDataFile(md5_file_name_);
+    ASSERT_TRUE(md5_file_ != NULL) << "Md5 file open failed. Filename: "
+        << md5_file_name_;
+  }
+
+  virtual void DecompressedFrameHook(const vpx_image_t &img,
+                                     const unsigned int frame_number) {
+    ASSERT_TRUE(md5_file_ != NULL);
+    char expected_md5[33];
+    char junk[128];
+
+    // Read correct md5 checksums.
+    const int res = fscanf(md5_file_, "%s  %s", expected_md5, junk);
+    ASSERT_NE(EOF, res) << "Read md5 data failed";
+    expected_md5[32] = '\0';
+
+    ::libvpx_test::MD5 md5_res;
+    md5_res.Add(&img);
+    const char *const actual_md5 = md5_res.Get();
+
+    // Check md5 match.
+    ASSERT_STREQ(expected_md5, actual_md5)
+        << "Md5 checksums don't match: frame number = " << frame_number;
+  }
+
+  // Callback to get a free external frame buffer. Return value < 0 is an
+  // error.
+  static int GetVP9FrameBuffer(void *user_priv, size_t min_size,
+                               vpx_codec_frame_buffer_t *fb) {
+    ExternalFrameBufferMD5Test *const md5Test =
+        reinterpret_cast<ExternalFrameBufferMD5Test*>(user_priv);
+    return md5Test->fb_list_.GetFreeFrameBuffer(min_size, fb);
+  }
+
+  // Callback to release an external frame buffer. Return value < 0 is an
+  // error.
+  static int ReleaseVP9FrameBuffer(void *user_priv,
+                                   vpx_codec_frame_buffer_t *fb) {
+    ExternalFrameBufferMD5Test *const md5Test =
+        reinterpret_cast<ExternalFrameBufferMD5Test*>(user_priv);
+    return md5Test->fb_list_.ReturnFrameBuffer(fb);
+  }
+
+  void set_num_buffers(int num_buffers) { num_buffers_ = num_buffers; }
+  int num_buffers() const { return num_buffers_; }
+
+ private:
+  FILE *md5_file_;
+  int num_buffers_;
+  ExternalFrameBufferList fb_list_;
+};
+
+#if CONFIG_WEBM_IO
+const char kVP9TestFile[] = "vp90-2-02-size-lf-1920x1080.webm";
+
+// Class for testing passing in external frame buffers to libvpx.
+class ExternalFrameBufferTest : public ::testing::Test {
+ protected:
+  ExternalFrameBufferTest()
+      : video_(NULL),
+        decoder_(NULL),
+        num_buffers_(0) {}
+
+  virtual void SetUp() {
+    video_ = new libvpx_test::WebMVideoSource(kVP9TestFile);
+    ASSERT_TRUE(video_ != NULL);
+    video_->Init();
+    video_->Begin();
+
+    vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+    decoder_ = new libvpx_test::VP9Decoder(cfg, 0);
+    ASSERT_TRUE(decoder_ != NULL);
+  }
+
+  virtual void TearDown() {
+    delete decoder_;
+    delete video_;
+  }
+
+  // Passes the external frame buffer information to libvpx.
+  vpx_codec_err_t SetFrameBufferFunctions(
+      int num_buffers,
+      vpx_get_frame_buffer_cb_fn_t cb_get,
+      vpx_release_frame_buffer_cb_fn_t cb_release) {
+    if (num_buffers > 0) {
+      num_buffers_ = num_buffers;
+      EXPECT_TRUE(fb_list_.CreateBufferList(num_buffers_));
+    }
+
+    return decoder_->SetFrameBufferFunctions(cb_get, cb_release, &fb_list_);
+  }
+
+  vpx_codec_err_t DecodeOneFrame() {
+    const vpx_codec_err_t res =
+        decoder_->DecodeFrame(video_->cxdata(), video_->frame_size());
+    CheckDecodedFrames();
+    if (res == VPX_CODEC_OK)
+      video_->Next();
+    return res;
+  }
+
+  vpx_codec_err_t DecodeRemainingFrames() {
+    for (; video_->cxdata() != NULL; video_->Next()) {
+      const vpx_codec_err_t res =
+          decoder_->DecodeFrame(video_->cxdata(), video_->frame_size());
+      if (res != VPX_CODEC_OK)
+        return res;
+      CheckDecodedFrames();
+    }
+    return VPX_CODEC_OK;
+  }
+
+ private:
+  void CheckDecodedFrames() {
+    libvpx_test::DxDataIterator dec_iter = decoder_->GetDxData();
+    const vpx_image_t *img = NULL;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next()) != NULL) {
+      fb_list_.CheckXImageFrameBuffer(img);
+    }
+  }
+
+  libvpx_test::WebMVideoSource *video_;
+  libvpx_test::VP9Decoder *decoder_;
+  int num_buffers_;
+  ExternalFrameBufferList fb_list_;
+};
+#endif  // CONFIG_WEBM_IO
+
+// This test runs through the set of test vectors, and decodes them.
+// Libvpx will call into the application to allocate a frame buffer when
+// needed. The md5 checksums are computed for each frame in the video file.
+// If md5 checksums match the correct md5 data, then the test is passed.
+// Otherwise, the test failed.
+TEST_P(ExternalFrameBufferMD5Test, ExtFBMD5Match) {
+  const std::string filename = GET_PARAM(kVideoNameParam);
+  libvpx_test::CompressedVideoSource *video = NULL;
+
+  // Number of buffers equals #VP9_MAXIMUM_REF_BUFFERS +
+  // #VPX_MAXIMUM_WORK_BUFFERS + four jitter buffers.
+  const int jitter_buffers = 4;
+  const int num_buffers =
+      VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS + jitter_buffers;
+  set_num_buffers(num_buffers);
+
+#if CONFIG_VP8_DECODER
+  // Tell compiler we are not using kVP8TestVectors.
+  (void)libvpx_test::kVP8TestVectors;
+#endif
+
+  // Open compressed video file.
+  if (filename.substr(filename.length() - 3, 3) == "ivf") {
+    video = new libvpx_test::IVFVideoSource(filename);
+  } else {
+#if CONFIG_WEBM_IO
+    video = new libvpx_test::WebMVideoSource(filename);
+#else
+    fprintf(stderr, "WebM IO is disabled, skipping test vector %s\n",
+            filename.c_str());
+    return;
+#endif
+  }
+  ASSERT_TRUE(video != NULL);
+  video->Init();
+
+  // Construct md5 file name.
+  const std::string md5_filename = filename + ".md5";
+  OpenMD5File(md5_filename);
+
+  // Decode frame, and check the md5 matching.
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video));
+  delete video;
+}
+
+#if CONFIG_WEBM_IO
+TEST_F(ExternalFrameBufferTest, MinFrameBuffers) {
+  // Minimum number of external frame buffers for VP9 is
+  // #VP9_MAXIMUM_REF_BUFFERS + #VPX_MAXIMUM_WORK_BUFFERS.
+  const int num_buffers = VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS;
+  ASSERT_EQ(VPX_CODEC_OK,
+            SetFrameBufferFunctions(
+                num_buffers, get_vp9_frame_buffer, release_vp9_frame_buffer));
+  ASSERT_EQ(VPX_CODEC_OK, DecodeRemainingFrames());
+}
+
+TEST_F(ExternalFrameBufferTest, EightJitterBuffers) {
+  // Number of buffers equals #VP9_MAXIMUM_REF_BUFFERS +
+  // #VPX_MAXIMUM_WORK_BUFFERS + eight jitter buffers.
+  const int jitter_buffers = 8;
+  const int num_buffers =
+      VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS + jitter_buffers;
+  ASSERT_EQ(VPX_CODEC_OK,
+            SetFrameBufferFunctions(
+                num_buffers, get_vp9_frame_buffer, release_vp9_frame_buffer));
+  ASSERT_EQ(VPX_CODEC_OK, DecodeRemainingFrames());
+}
+
+TEST_F(ExternalFrameBufferTest, NotEnoughBuffers) {
+  // Minimum number of external frame buffers for VP9 is
+  // #VP9_MAXIMUM_REF_BUFFERS + #VPX_MAXIMUM_WORK_BUFFERS. Most files will
+  // only use 5 frame buffers at one time.
+  const int num_buffers = 2;
+  ASSERT_EQ(VPX_CODEC_OK,
+            SetFrameBufferFunctions(
+                num_buffers, get_vp9_frame_buffer, release_vp9_frame_buffer));
+  ASSERT_EQ(VPX_CODEC_OK, DecodeOneFrame());
+  ASSERT_EQ(VPX_CODEC_MEM_ERROR, DecodeRemainingFrames());
+}
+
+TEST_F(ExternalFrameBufferTest, NoRelease) {
+  const int num_buffers = VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS;
+  ASSERT_EQ(VPX_CODEC_OK,
+            SetFrameBufferFunctions(num_buffers, get_vp9_frame_buffer,
+                                    do_not_release_vp9_frame_buffer));
+  ASSERT_EQ(VPX_CODEC_OK, DecodeOneFrame());
+  ASSERT_EQ(VPX_CODEC_MEM_ERROR, DecodeRemainingFrames());
+}
+
+TEST_F(ExternalFrameBufferTest, NullRealloc) {
+  const int num_buffers = VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS;
+  ASSERT_EQ(VPX_CODEC_OK,
+            SetFrameBufferFunctions(num_buffers, get_vp9_zero_frame_buffer,
+                                    release_vp9_frame_buffer));
+  ASSERT_EQ(VPX_CODEC_MEM_ERROR, DecodeOneFrame());
+}
+
+TEST_F(ExternalFrameBufferTest, ReallocOneLessByte) {
+  const int num_buffers = VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS;
+  ASSERT_EQ(VPX_CODEC_OK,
+            SetFrameBufferFunctions(
+                num_buffers, get_vp9_one_less_byte_frame_buffer,
+                release_vp9_frame_buffer));
+  ASSERT_EQ(VPX_CODEC_MEM_ERROR, DecodeOneFrame());
+}
+
+TEST_F(ExternalFrameBufferTest, NullGetFunction) {
+  const int num_buffers = VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS;
+  ASSERT_EQ(VPX_CODEC_INVALID_PARAM,
+            SetFrameBufferFunctions(num_buffers, NULL,
+                                    release_vp9_frame_buffer));
+}
+
+TEST_F(ExternalFrameBufferTest, NullReleaseFunction) {
+  const int num_buffers = VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS;
+  ASSERT_EQ(VPX_CODEC_INVALID_PARAM,
+            SetFrameBufferFunctions(num_buffers, get_vp9_frame_buffer, NULL));
+}
+
+TEST_F(ExternalFrameBufferTest, SetAfterDecode) {
+  const int num_buffers = VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS;
+  ASSERT_EQ(VPX_CODEC_OK, DecodeOneFrame());
+  ASSERT_EQ(VPX_CODEC_ERROR,
+            SetFrameBufferFunctions(
+                num_buffers, get_vp9_frame_buffer, release_vp9_frame_buffer));
+}
+#endif  // CONFIG_WEBM_IO
+
+VP9_INSTANTIATE_TEST_CASE(ExternalFrameBufferMD5Test,
+                          ::testing::ValuesIn(libvpx_test::kVP9TestVectors,
+                                              libvpx_test::kVP9TestVectors +
+                                              libvpx_test::kNumVP9TestVectors));
+}  // namespace
diff --git a/libvpx/libvpx/test/fdct4x4_test.cc b/libvpx/libvpx/test/fdct4x4_test.cc
new file mode 100644
index 0000000..735cccf
--- /dev/null
+++ b/libvpx/libvpx/test/fdct4x4_test.cc
@@ -0,0 +1,546 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+const int kNumCoeffs = 16;
+typedef void (*FdctFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*IdctFunc)(const tran_low_t *in, uint8_t *out, int stride);
+typedef void (*FhtFunc)(const int16_t *in, tran_low_t *out, int stride,
+                        int tx_type);
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+
+typedef std::tr1::tuple<FdctFunc, IdctFunc, int, vpx_bit_depth_t> Dct4x4Param;
+typedef std::tr1::tuple<FhtFunc, IhtFunc, int, vpx_bit_depth_t> Ht4x4Param;
+
+void fdct4x4_ref(const int16_t *in, tran_low_t *out, int stride,
+                 int /*tx_type*/) {
+  vpx_fdct4x4_c(in, out, stride);
+}
+
+void fht4x4_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  vp9_fht4x4_c(in, out, stride, tx_type);
+}
+
+void fwht4x4_ref(const int16_t *in, tran_low_t *out, int stride,
+                 int /*tx_type*/) {
+  vp9_fwht4x4_c(in, out, stride);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void idct4x4_10(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct4x4_16_add_c(in, out, stride, 10);
+}
+
+void idct4x4_12(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct4x4_16_add_c(in, out, stride, 12);
+}
+
+void iht4x4_10(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  vp9_highbd_iht4x4_16_add_c(in, out, stride, tx_type, 10);
+}
+
+void iht4x4_12(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  vp9_highbd_iht4x4_16_add_c(in, out, stride, tx_type, 12);
+}
+
+void iwht4x4_10(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_iwht4x4_16_add_c(in, out, stride, 10);
+}
+
+void iwht4x4_12(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_iwht4x4_16_add_c(in, out, stride, 12);
+}
+
+#if HAVE_SSE2
+void idct4x4_10_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct4x4_16_add_sse2(in, out, stride, 10);
+}
+
+void idct4x4_12_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct4x4_16_add_sse2(in, out, stride, 12);
+}
+#endif  // HAVE_SSE2
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+class Trans4x4TestBase {
+ public:
+  virtual ~Trans4x4TestBase() {}
+
+ protected:
+  virtual void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) = 0;
+
+  virtual void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) = 0;
+
+  void RunAccuracyCheck(int limit) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    uint32_t max_error = 0;
+    int64_t total_error = 0;
+    const int count_test_block = 10000;
+    for (int i = 0; i < count_test_block; ++i) {
+      DECLARE_ALIGNED(16, int16_t, test_input_block[kNumCoeffs]);
+      DECLARE_ALIGNED(16, tran_low_t, test_temp_block[kNumCoeffs]);
+      DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+      DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_VP9_HIGHBITDEPTH
+      DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+      DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+#endif
+
+      // Initialize a test block with input range [-255, 255].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (bit_depth_ == VPX_BITS_8) {
+          src[j] = rnd.Rand8();
+          dst[j] = rnd.Rand8();
+          test_input_block[j] = src[j] - dst[j];
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand16() & mask_;
+          dst16[j] = rnd.Rand16() & mask_;
+          test_input_block[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(test_input_block,
+                                          test_temp_block, pitch_));
+      if (bit_depth_ == VPX_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(test_temp_block, dst, pitch_));
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(test_temp_block,
+                                            CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == VPX_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        ASSERT_EQ(VPX_BITS_8, bit_depth_);
+        const int diff = dst[j] - src[j];
+#endif
+        const uint32_t error = diff * diff;
+        if (max_error < error)
+          max_error = error;
+        total_error += error;
+      }
+    }
+
+    EXPECT_GE(static_cast<uint32_t>(limit), max_error)
+        << "Error: 4x4 FHT/IHT has an individual round trip error > "
+        << limit;
+
+    EXPECT_GE(count_test_block * limit, total_error)
+        << "Error: 4x4 FHT/IHT has average round trip error > " << limit
+        << " per block";
+  }
+
+  void RunCoeffCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 5000;
+    DECLARE_ALIGNED(16, int16_t, input_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]);
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j)
+        input_block[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+
+      fwd_txfm_ref(input_block, output_ref_block, pitch_, tx_type_);
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(input_block, output_block, pitch_));
+
+      // The minimum quant value is 4.
+      for (int j = 0; j < kNumCoeffs; ++j)
+        EXPECT_EQ(output_block[j], output_ref_block[j]);
+    }
+  }
+
+  void RunMemCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 5000;
+    DECLARE_ALIGNED(16, int16_t, input_extreme_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]);
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        input_extreme_block[j] = rnd.Rand8() % 2 ? mask_ : -mask_;
+      }
+      if (i == 0) {
+        for (int j = 0; j < kNumCoeffs; ++j)
+          input_extreme_block[j] = mask_;
+      } else if (i == 1) {
+        for (int j = 0; j < kNumCoeffs; ++j)
+          input_extreme_block[j] = -mask_;
+      }
+
+      fwd_txfm_ref(input_extreme_block, output_ref_block, pitch_, tx_type_);
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(input_extreme_block,
+                                          output_block, pitch_));
+
+      // The minimum quant value is 4.
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        EXPECT_EQ(output_block[j], output_ref_block[j]);
+        EXPECT_GE(4 * DCT_MAX_VALUE << (bit_depth_ - 8), abs(output_block[j]))
+            << "Error: 4x4 FDCT has coefficient larger than 4*DCT_MAX_VALUE";
+      }
+    }
+  }
+
+  void RunInvAccuracyCheck(int limit) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, in[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (bit_depth_ == VPX_BITS_8) {
+          src[j] = rnd.Rand8();
+          dst[j] = rnd.Rand8();
+          in[j] = src[j] - dst[j];
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand16() & mask_;
+          dst16[j] = rnd.Rand16() & mask_;
+          in[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      fwd_txfm_ref(in, coeff, pitch_, tx_type_);
+
+      if (bit_depth_ == VPX_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, pitch_));
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, CONVERT_TO_BYTEPTR(dst16),
+                                            pitch_));
+#endif
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == VPX_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int diff = dst[j] - src[j];
+#endif
+        const uint32_t error = diff * diff;
+        EXPECT_GE(static_cast<uint32_t>(limit), error)
+            << "Error: 4x4 IDCT has error " << error
+            << " at index " << j;
+      }
+    }
+  }
+
+  int pitch_;
+  int tx_type_;
+  FhtFunc fwd_txfm_ref;
+  vpx_bit_depth_t bit_depth_;
+  int mask_;
+};
+
+class Trans4x4DCT
+    : public Trans4x4TestBase,
+      public ::testing::TestWithParam<Dct4x4Param> {
+ public:
+  virtual ~Trans4x4DCT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_  = GET_PARAM(2);
+    pitch_    = 4;
+    fwd_txfm_ref = fdct4x4_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride);
+  }
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
+};
+
+TEST_P(Trans4x4DCT, AccuracyCheck) {
+  RunAccuracyCheck(1);
+}
+
+TEST_P(Trans4x4DCT, CoeffCheck) {
+  RunCoeffCheck();
+}
+
+TEST_P(Trans4x4DCT, MemCheck) {
+  RunMemCheck();
+}
+
+TEST_P(Trans4x4DCT, InvAccuracyCheck) {
+  RunInvAccuracyCheck(1);
+}
+
+class Trans4x4HT
+    : public Trans4x4TestBase,
+      public ::testing::TestWithParam<Ht4x4Param> {
+ public:
+  virtual ~Trans4x4HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_  = GET_PARAM(2);
+    pitch_    = 4;
+    fwd_txfm_ref = fht4x4_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(Trans4x4HT, AccuracyCheck) {
+  RunAccuracyCheck(1);
+}
+
+TEST_P(Trans4x4HT, CoeffCheck) {
+  RunCoeffCheck();
+}
+
+TEST_P(Trans4x4HT, MemCheck) {
+  RunMemCheck();
+}
+
+TEST_P(Trans4x4HT, InvAccuracyCheck) {
+  RunInvAccuracyCheck(1);
+}
+
+class Trans4x4WHT
+    : public Trans4x4TestBase,
+      public ::testing::TestWithParam<Dct4x4Param> {
+ public:
+  virtual ~Trans4x4WHT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_  = GET_PARAM(2);
+    pitch_    = 4;
+    fwd_txfm_ref = fwht4x4_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride);
+  }
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
+};
+
+TEST_P(Trans4x4WHT, AccuracyCheck) {
+  RunAccuracyCheck(0);
+}
+
+TEST_P(Trans4x4WHT, CoeffCheck) {
+  RunCoeffCheck();
+}
+
+TEST_P(Trans4x4WHT, MemCheck) {
+  RunMemCheck();
+}
+
+TEST_P(Trans4x4WHT, InvAccuracyCheck) {
+  RunInvAccuracyCheck(0);
+}
+using std::tr1::make_tuple;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4DCT,
+    ::testing::Values(
+        make_tuple(&vpx_highbd_fdct4x4_c, &idct4x4_10, 0, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct4x4_c, &idct4x4_12, 0, VPX_BITS_12),
+        make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c, 0, VPX_BITS_8)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct4x4_c, &vpx_idct4x4_16_add_c, 0, VPX_BITS_8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&vp9_highbd_fht4x4_c, &iht4x4_10, 0, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht4x4_c, &iht4x4_10, 1, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht4x4_c, &iht4x4_10, 2, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht4x4_c, &iht4x4_10, 3, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht4x4_c, &iht4x4_12, 0, VPX_BITS_12),
+        make_tuple(&vp9_highbd_fht4x4_c, &iht4x4_12, 1, VPX_BITS_12),
+        make_tuple(&vp9_highbd_fht4x4_c, &iht4x4_12, 2, VPX_BITS_12),
+        make_tuple(&vp9_highbd_fht4x4_c, &iht4x4_12, 3, VPX_BITS_12),
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 3, VPX_BITS_8)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_c, 3, VPX_BITS_8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4WHT,
+    ::testing::Values(
+        make_tuple(&vp9_highbd_fwht4x4_c, &iwht4x4_10, 0, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fwht4x4_c, &iwht4x4_12, 0, VPX_BITS_12),
+        make_tuple(&vp9_fwht4x4_c, &vpx_iwht4x4_16_add_c, 0, VPX_BITS_8)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4WHT,
+    ::testing::Values(
+        make_tuple(&vp9_fwht4x4_c, &vpx_iwht4x4_16_add_c, 0, VPX_BITS_8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if HAVE_NEON_ASM && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    NEON, Trans4x4DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct4x4_c,
+                   &vpx_idct4x4_16_add_neon, 0, VPX_BITS_8)));
+#endif  // HAVE_NEON_ASM && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    NEON, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_neon, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_neon, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_neon, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_c, &vp9_iht4x4_16_add_neon, 3, VPX_BITS_8)));
+#endif  // HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if CONFIG_USE_X86INC && HAVE_SSE2 && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans4x4WHT,
+    ::testing::Values(
+        make_tuple(&vp9_fwht4x4_sse2, &vpx_iwht4x4_16_add_c, 0, VPX_BITS_8),
+        make_tuple(&vp9_fwht4x4_c, &vpx_iwht4x4_16_add_sse2, 0, VPX_BITS_8)));
+#endif
+
+#if HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans4x4DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct4x4_sse2,
+                   &vpx_idct4x4_16_add_sse2, 0, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht4x4_sse2, &vp9_iht4x4_16_add_sse2, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_sse2, &vp9_iht4x4_16_add_sse2, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_sse2, &vp9_iht4x4_16_add_sse2, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_sse2, &vp9_iht4x4_16_add_sse2, 3, VPX_BITS_8)));
+#endif  // HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans4x4DCT,
+    ::testing::Values(
+        make_tuple(&vpx_highbd_fdct4x4_c,    &idct4x4_10_sse2, 0, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct4x4_sse2, &idct4x4_10_sse2, 0, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct4x4_c,    &idct4x4_12_sse2, 0, VPX_BITS_12),
+        make_tuple(&vpx_highbd_fdct4x4_sse2, &idct4x4_12_sse2, 0, VPX_BITS_12),
+        make_tuple(&vpx_fdct4x4_sse2,      &vpx_idct4x4_16_add_c, 0,
+                   VPX_BITS_8)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht4x4_sse2, &vp9_iht4x4_16_add_c, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_sse2, &vp9_iht4x4_16_add_c, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_sse2, &vp9_iht4x4_16_add_c, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_sse2, &vp9_iht4x4_16_add_c, 3, VPX_BITS_8)));
+#endif  // HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    MSA, Trans4x4DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct4x4_msa, &vpx_idct4x4_16_add_msa, 0, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    MSA, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht4x4_msa, &vp9_iht4x4_16_add_msa, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_msa, &vp9_iht4x4_16_add_msa, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_msa, &vp9_iht4x4_16_add_msa, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht4x4_msa, &vp9_iht4x4_16_add_msa, 3, VPX_BITS_8)));
+#endif  // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+}  // namespace
diff --git a/libvpx/libvpx/test/fdct8x8_test.cc b/libvpx/libvpx/test/fdct8x8_test.cc
new file mode 100644
index 0000000..29f2158
--- /dev/null
+++ b/libvpx/libvpx/test/fdct8x8_test.cc
@@ -0,0 +1,791 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_scan.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+
+const int kNumCoeffs = 64;
+const double kPi = 3.141592653589793238462643383279502884;
+
+const int kSignBiasMaxDiff255 = 1500;
+const int kSignBiasMaxDiff15 = 10000;
+
+typedef void (*FdctFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*IdctFunc)(const tran_low_t *in, uint8_t *out, int stride);
+typedef void (*FhtFunc)(const int16_t *in, tran_low_t *out, int stride,
+                        int tx_type);
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+
+typedef std::tr1::tuple<FdctFunc, IdctFunc, int, vpx_bit_depth_t> Dct8x8Param;
+typedef std::tr1::tuple<FhtFunc, IhtFunc, int, vpx_bit_depth_t> Ht8x8Param;
+typedef std::tr1::tuple<IdctFunc, IdctFunc, int, vpx_bit_depth_t> Idct8x8Param;
+
+void reference_8x8_dct_1d(const double in[8], double out[8]) {
+  const double kInvSqrt2 = 0.707106781186547524400844362104;
+  for (int k = 0; k < 8; k++) {
+    out[k] = 0.0;
+    for (int n = 0; n < 8; n++)
+      out[k] += in[n] * cos(kPi * (2 * n + 1) * k / 16.0);
+    if (k == 0)
+      out[k] = out[k] * kInvSqrt2;
+  }
+}
+
+void reference_8x8_dct_2d(const int16_t input[kNumCoeffs],
+                          double output[kNumCoeffs]) {
+  // First transform columns
+  for (int i = 0; i < 8; ++i) {
+    double temp_in[8], temp_out[8];
+    for (int j = 0; j < 8; ++j)
+      temp_in[j] = input[j*8 + i];
+    reference_8x8_dct_1d(temp_in, temp_out);
+    for (int j = 0; j < 8; ++j)
+      output[j * 8 + i] = temp_out[j];
+  }
+  // Then transform rows
+  for (int i = 0; i < 8; ++i) {
+    double temp_in[8], temp_out[8];
+    for (int j = 0; j < 8; ++j)
+      temp_in[j] = output[j + i*8];
+    reference_8x8_dct_1d(temp_in, temp_out);
+    // Scale by some magic number
+    for (int j = 0; j < 8; ++j)
+      output[j + i * 8] = temp_out[j] * 2;
+  }
+}
+
+
+void fdct8x8_ref(const int16_t *in, tran_low_t *out, int stride,
+                 int /*tx_type*/) {
+  vpx_fdct8x8_c(in, out, stride);
+}
+
+void fht8x8_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  vp9_fht8x8_c(in, out, stride, tx_type);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void idct8x8_10(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct8x8_64_add_c(in, out, stride, 10);
+}
+
+void idct8x8_12(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct8x8_64_add_c(in, out, stride, 12);
+}
+
+void iht8x8_10(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  vp9_highbd_iht8x8_64_add_c(in, out, stride, tx_type, 10);
+}
+
+void iht8x8_12(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  vp9_highbd_iht8x8_64_add_c(in, out, stride, tx_type, 12);
+}
+
+#if HAVE_SSE2
+
+void idct8x8_10_add_10_c(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct8x8_10_add_c(in, out, stride, 10);
+}
+
+void idct8x8_10_add_12_c(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct8x8_10_add_c(in, out, stride, 12);
+}
+
+void idct8x8_10_add_10_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct8x8_10_add_sse2(in, out, stride, 10);
+}
+
+void idct8x8_10_add_12_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct8x8_10_add_sse2(in, out, stride, 12);
+}
+
+void idct8x8_64_add_10_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct8x8_64_add_sse2(in, out, stride, 10);
+}
+
+void idct8x8_64_add_12_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  vpx_highbd_idct8x8_64_add_sse2(in, out, stride, 12);
+}
+#endif  // HAVE_SSE2
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+class FwdTrans8x8TestBase {
+ public:
+  virtual ~FwdTrans8x8TestBase() {}
+
+ protected:
+  virtual void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) = 0;
+  virtual void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) = 0;
+
+  void RunSignBiasCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    DECLARE_ALIGNED(16, int16_t, test_input_block[64]);
+    DECLARE_ALIGNED(16, tran_low_t, test_output_block[64]);
+    int count_sign_block[64][2];
+    const int count_test_block = 100000;
+
+    memset(count_sign_block, 0, sizeof(count_sign_block));
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-255, 255].
+      for (int j = 0; j < 64; ++j)
+        test_input_block[j] = ((rnd.Rand16() >> (16 - bit_depth_)) & mask_) -
+                              ((rnd.Rand16() >> (16 - bit_depth_)) & mask_);
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(test_input_block, test_output_block, pitch_));
+
+      for (int j = 0; j < 64; ++j) {
+        if (test_output_block[j] < 0)
+          ++count_sign_block[j][0];
+        else if (test_output_block[j] > 0)
+          ++count_sign_block[j][1];
+      }
+    }
+
+    for (int j = 0; j < 64; ++j) {
+      const int diff = abs(count_sign_block[j][0] - count_sign_block[j][1]);
+      const int max_diff = kSignBiasMaxDiff255;
+      EXPECT_LT(diff, max_diff << (bit_depth_ - 8))
+          << "Error: 8x8 FDCT/FHT has a sign bias > "
+          << 1. * max_diff / count_test_block * 100 << "%"
+          << " for input range [-255, 255] at index " << j
+          << " count0: " << count_sign_block[j][0]
+          << " count1: " << count_sign_block[j][1]
+          << " diff: " << diff;
+    }
+
+    memset(count_sign_block, 0, sizeof(count_sign_block));
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_ / 16, mask_ / 16].
+      for (int j = 0; j < 64; ++j)
+        test_input_block[j] = ((rnd.Rand16() & mask_) >> 4) -
+                              ((rnd.Rand16() & mask_) >> 4);
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(test_input_block, test_output_block, pitch_));
+
+      for (int j = 0; j < 64; ++j) {
+        if (test_output_block[j] < 0)
+          ++count_sign_block[j][0];
+        else if (test_output_block[j] > 0)
+          ++count_sign_block[j][1];
+      }
+    }
+
+    for (int j = 0; j < 64; ++j) {
+      const int diff = abs(count_sign_block[j][0] - count_sign_block[j][1]);
+      const int max_diff = kSignBiasMaxDiff15;
+      EXPECT_LT(diff, max_diff << (bit_depth_ - 8))
+          << "Error: 8x8 FDCT/FHT has a sign bias > "
+          << 1. * max_diff / count_test_block * 100 << "%"
+          << " for input range [-15, 15] at index " << j
+          << " count0: " << count_sign_block[j][0]
+          << " count1: " << count_sign_block[j][1]
+          << " diff: " << diff;
+    }
+  }
+
+  void RunRoundTripErrorCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    int max_error = 0;
+    int total_error = 0;
+    const int count_test_block = 100000;
+    DECLARE_ALIGNED(16, int16_t, test_input_block[64]);
+    DECLARE_ALIGNED(16, tran_low_t, test_temp_block[64]);
+    DECLARE_ALIGNED(16, uint8_t, dst[64]);
+    DECLARE_ALIGNED(16, uint8_t, src[64]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[64]);
+    DECLARE_ALIGNED(16, uint16_t, src16[64]);
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < 64; ++j) {
+        if (bit_depth_ == VPX_BITS_8) {
+          src[j] = rnd.Rand8();
+          dst[j] = rnd.Rand8();
+          test_input_block[j] = src[j] - dst[j];
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand16() & mask_;
+          dst16[j] = rnd.Rand16() & mask_;
+          test_input_block[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(test_input_block, test_temp_block, pitch_));
+      for (int j = 0; j < 64; ++j) {
+          if (test_temp_block[j] > 0) {
+            test_temp_block[j] += 2;
+            test_temp_block[j] /= 4;
+            test_temp_block[j] *= 4;
+          } else {
+            test_temp_block[j] -= 2;
+            test_temp_block[j] /= 4;
+            test_temp_block[j] *= 4;
+          }
+      }
+      if (bit_depth_ == VPX_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(test_temp_block, dst, pitch_));
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(test_temp_block, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < 64; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == VPX_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int diff = dst[j] - src[j];
+#endif
+        const int error = diff * diff;
+        if (max_error < error)
+          max_error = error;
+        total_error += error;
+      }
+    }
+
+    EXPECT_GE(1 << 2 * (bit_depth_ - 8), max_error)
+      << "Error: 8x8 FDCT/IDCT or FHT/IHT has an individual"
+      << " roundtrip error > 1";
+
+    EXPECT_GE((count_test_block << 2 * (bit_depth_ - 8))/5, total_error)
+      << "Error: 8x8 FDCT/IDCT or FHT/IHT has average roundtrip "
+      << "error > 1/5 per block";
+  }
+
+  void RunExtremalCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    int max_error = 0;
+    int total_error = 0;
+    int total_coeff_error = 0;
+    const int count_test_block = 100000;
+    DECLARE_ALIGNED(16, int16_t, test_input_block[64]);
+    DECLARE_ALIGNED(16, tran_low_t, test_temp_block[64]);
+    DECLARE_ALIGNED(16, tran_low_t, ref_temp_block[64]);
+    DECLARE_ALIGNED(16, uint8_t, dst[64]);
+    DECLARE_ALIGNED(16, uint8_t, src[64]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[64]);
+    DECLARE_ALIGNED(16, uint16_t, src16[64]);
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < 64; ++j) {
+        if (bit_depth_ == VPX_BITS_8) {
+          if (i == 0) {
+            src[j] = 255;
+            dst[j] = 0;
+          } else if (i == 1) {
+            src[j] = 0;
+            dst[j] = 255;
+          } else {
+            src[j] = rnd.Rand8() % 2 ? 255 : 0;
+            dst[j] = rnd.Rand8() % 2 ? 255 : 0;
+          }
+          test_input_block[j] = src[j] - dst[j];
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          if (i == 0) {
+            src16[j] = mask_;
+            dst16[j] = 0;
+          } else if (i == 1) {
+            src16[j] = 0;
+            dst16[j] = mask_;
+          } else {
+            src16[j] = rnd.Rand8() % 2 ? mask_ : 0;
+            dst16[j] = rnd.Rand8() % 2 ? mask_ : 0;
+          }
+          test_input_block[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(test_input_block, test_temp_block, pitch_));
+      ASM_REGISTER_STATE_CHECK(
+          fwd_txfm_ref(test_input_block, ref_temp_block, pitch_, tx_type_));
+      if (bit_depth_ == VPX_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(test_temp_block, dst, pitch_));
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(test_temp_block, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < 64; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == VPX_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int diff = dst[j] - src[j];
+#endif
+        const int error = diff * diff;
+        if (max_error < error)
+          max_error = error;
+        total_error += error;
+
+        const int coeff_diff = test_temp_block[j] - ref_temp_block[j];
+        total_coeff_error += abs(coeff_diff);
+      }
+
+      EXPECT_GE(1 << 2 * (bit_depth_ - 8), max_error)
+          << "Error: Extremal 8x8 FDCT/IDCT or FHT/IHT has"
+          << "an individual roundtrip error > 1";
+
+      EXPECT_GE((count_test_block << 2 * (bit_depth_ - 8))/5, total_error)
+          << "Error: Extremal 8x8 FDCT/IDCT or FHT/IHT has average"
+          << " roundtrip error > 1/5 per block";
+
+      EXPECT_EQ(0, total_coeff_error)
+          << "Error: Extremal 8x8 FDCT/FHT has"
+          << "overflow issues in the intermediate steps > 1";
+    }
+  }
+
+  void RunInvAccuracyCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, in[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      double out_r[kNumCoeffs];
+
+      // Initialize a test block with input range [-255, 255].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (bit_depth_ == VPX_BITS_8) {
+          src[j] = rnd.Rand8() % 2 ? 255 : 0;
+          dst[j] = src[j] > 0 ? 0 : 255;
+          in[j] = src[j] - dst[j];
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand8() % 2 ? mask_ : 0;
+          dst16[j] = src16[j] > 0 ? 0 : mask_;
+          in[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      reference_8x8_dct_2d(in, out_r);
+      for (int j = 0; j < kNumCoeffs; ++j)
+        coeff[j] = static_cast<tran_low_t>(round(out_r[j]));
+
+      if (bit_depth_ == VPX_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, pitch_));
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, CONVERT_TO_BYTEPTR(dst16),
+                                            pitch_));
+#endif
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == VPX_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int diff = dst[j] - src[j];
+#endif
+        const uint32_t error = diff * diff;
+        EXPECT_GE(1u << 2 * (bit_depth_ - 8), error)
+            << "Error: 8x8 IDCT has error " << error
+            << " at index " << j;
+      }
+    }
+  }
+
+  void RunFwdAccuracyCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, in[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, coeff_r[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+
+    for (int i = 0; i < count_test_block; ++i) {
+      double out_r[kNumCoeffs];
+
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j)
+        in[j] = rnd.Rand8() % 2 == 0 ? mask_ : -mask_;
+
+      RunFwdTxfm(in, coeff, pitch_);
+      reference_8x8_dct_2d(in, out_r);
+      for (int j = 0; j < kNumCoeffs; ++j)
+        coeff_r[j] = static_cast<tran_low_t>(round(out_r[j]));
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        const int32_t diff = coeff[j] - coeff_r[j];
+        const uint32_t error = diff * diff;
+        EXPECT_GE(9u << 2 * (bit_depth_ - 8), error)
+            << "Error: 8x8 DCT has error " << error
+            << " at index " << j;
+      }
+    }
+  }
+
+void CompareInvReference(IdctFunc ref_txfm, int thresh) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 10000;
+    const int eob = 12;
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, ref[kNumCoeffs]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, ref16[kNumCoeffs]);
+#endif
+    const int16_t *scan = vp9_default_scan_orders[TX_8X8].scan;
+
+    for (int i = 0; i < count_test_block; ++i) {
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (j < eob) {
+          // Random values less than the threshold, either positive or negative
+          coeff[scan[j]] = rnd(thresh) * (1-2*(i%2));
+        } else {
+          coeff[scan[j]] = 0;
+        }
+        if (bit_depth_ == VPX_BITS_8) {
+          dst[j] = 0;
+          ref[j] = 0;
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          dst16[j] = 0;
+          ref16[j] = 0;
+#endif
+        }
+      }
+      if (bit_depth_ == VPX_BITS_8) {
+        ref_txfm(coeff, ref, pitch_);
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, pitch_));
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        ref_txfm(coeff, CONVERT_TO_BYTEPTR(ref16), pitch_);
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, CONVERT_TO_BYTEPTR(dst16),
+                                            pitch_));
+#endif
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == VPX_BITS_8 ? dst[j] - ref[j] : dst16[j] - ref16[j];
+#else
+        const int diff = dst[j] - ref[j];
+#endif
+        const uint32_t error = diff * diff;
+        EXPECT_EQ(0u, error)
+            << "Error: 8x8 IDCT has error " << error
+            << " at index " << j;
+      }
+    }
+  }
+  int pitch_;
+  int tx_type_;
+  FhtFunc fwd_txfm_ref;
+  vpx_bit_depth_t bit_depth_;
+  int mask_;
+};
+
+class FwdTrans8x8DCT
+    : public FwdTrans8x8TestBase,
+      public ::testing::TestWithParam<Dct8x8Param> {
+ public:
+  virtual ~FwdTrans8x8DCT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_  = GET_PARAM(2);
+    pitch_    = 8;
+    fwd_txfm_ref = fdct8x8_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride);
+  }
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
+};
+
+TEST_P(FwdTrans8x8DCT, SignBiasCheck) {
+  RunSignBiasCheck();
+}
+
+TEST_P(FwdTrans8x8DCT, RoundTripErrorCheck) {
+  RunRoundTripErrorCheck();
+}
+
+TEST_P(FwdTrans8x8DCT, ExtremalCheck) {
+  RunExtremalCheck();
+}
+
+TEST_P(FwdTrans8x8DCT, FwdAccuracyCheck) {
+  RunFwdAccuracyCheck();
+}
+
+TEST_P(FwdTrans8x8DCT, InvAccuracyCheck) {
+  RunInvAccuracyCheck();
+}
+
+class FwdTrans8x8HT
+    : public FwdTrans8x8TestBase,
+      public ::testing::TestWithParam<Ht8x8Param> {
+ public:
+  virtual ~FwdTrans8x8HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_  = GET_PARAM(2);
+    pitch_    = 8;
+    fwd_txfm_ref = fht8x8_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(FwdTrans8x8HT, SignBiasCheck) {
+  RunSignBiasCheck();
+}
+
+TEST_P(FwdTrans8x8HT, RoundTripErrorCheck) {
+  RunRoundTripErrorCheck();
+}
+
+TEST_P(FwdTrans8x8HT, ExtremalCheck) {
+  RunExtremalCheck();
+}
+
+class InvTrans8x8DCT
+    : public FwdTrans8x8TestBase,
+      public ::testing::TestWithParam<Idct8x8Param> {
+ public:
+  virtual ~InvTrans8x8DCT() {}
+
+  virtual void SetUp() {
+    ref_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    thresh_ = GET_PARAM(2);
+    pitch_ = 8;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+  void RunFwdTxfm(int16_t * /*out*/, tran_low_t * /*dst*/, int /*stride*/) {}
+
+  IdctFunc ref_txfm_;
+  IdctFunc inv_txfm_;
+  int thresh_;
+};
+
+TEST_P(InvTrans8x8DCT, CompareReference) {
+  CompareInvReference(ref_txfm_, thresh_);
+}
+
+using std::tr1::make_tuple;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, FwdTrans8x8DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c, 0, VPX_BITS_8),
+        make_tuple(&vpx_highbd_fdct8x8_c, &idct8x8_10, 0, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct8x8_c, &idct8x8_12, 0, VPX_BITS_12)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, FwdTrans8x8DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct8x8_c, &vpx_idct8x8_64_add_c, 0, VPX_BITS_8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_c, 0, VPX_BITS_8),
+        make_tuple(&vp9_highbd_fht8x8_c, &iht8x8_10, 0, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht8x8_c, &iht8x8_10, 1, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht8x8_c, &iht8x8_10, 2, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht8x8_c, &iht8x8_10, 3, VPX_BITS_10),
+        make_tuple(&vp9_highbd_fht8x8_c, &iht8x8_12, 0, VPX_BITS_12),
+        make_tuple(&vp9_highbd_fht8x8_c, &iht8x8_12, 1, VPX_BITS_12),
+        make_tuple(&vp9_highbd_fht8x8_c, &iht8x8_12, 2, VPX_BITS_12),
+        make_tuple(&vp9_highbd_fht8x8_c, &iht8x8_12, 3, VPX_BITS_12),
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_c, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_c, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_c, 3, VPX_BITS_8)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_c, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_c, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_c, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_c, 3, VPX_BITS_8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if HAVE_NEON_ASM && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    NEON, FwdTrans8x8DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct8x8_neon, &vpx_idct8x8_64_add_neon, 0,
+                   VPX_BITS_8)));
+#endif  // HAVE_NEON_ASM && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    NEON, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_neon, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_neon, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_neon, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_c, &vp9_iht8x8_64_add_neon, 3, VPX_BITS_8)));
+#endif  // HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSE2, FwdTrans8x8DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct8x8_sse2, &vpx_idct8x8_64_add_sse2, 0,
+                   VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    SSE2, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht8x8_sse2, &vp9_iht8x8_64_add_sse2, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_sse2, &vp9_iht8x8_64_add_sse2, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_sse2, &vp9_iht8x8_64_add_sse2, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_sse2, &vp9_iht8x8_64_add_sse2, 3, VPX_BITS_8)));
+#endif  // HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSE2, FwdTrans8x8DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct8x8_sse2, &vpx_idct8x8_64_add_c, 0, VPX_BITS_8),
+        make_tuple(&vpx_highbd_fdct8x8_c,
+                   &idct8x8_64_add_10_sse2, 12, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct8x8_sse2,
+                   &idct8x8_64_add_10_sse2, 12, VPX_BITS_10),
+        make_tuple(&vpx_highbd_fdct8x8_c,
+                   &idct8x8_64_add_12_sse2, 12, VPX_BITS_12),
+        make_tuple(&vpx_highbd_fdct8x8_sse2,
+                   &idct8x8_64_add_12_sse2, 12, VPX_BITS_12)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht8x8_sse2, &vp9_iht8x8_64_add_c, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_sse2, &vp9_iht8x8_64_add_c, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_sse2, &vp9_iht8x8_64_add_c, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_sse2, &vp9_iht8x8_64_add_c, 3, VPX_BITS_8)));
+
+// Optimizations take effect at a threshold of 6201, so we use a value close to
+// that to test both branches.
+INSTANTIATE_TEST_CASE_P(
+    SSE2, InvTrans8x8DCT,
+    ::testing::Values(
+        make_tuple(&idct8x8_10_add_10_c,
+                   &idct8x8_10_add_10_sse2, 6225, VPX_BITS_10),
+        make_tuple(&idct8x8_10,
+                   &idct8x8_64_add_10_sse2, 6225, VPX_BITS_10),
+        make_tuple(&idct8x8_10_add_12_c,
+                   &idct8x8_10_add_12_sse2, 6225, VPX_BITS_12),
+        make_tuple(&idct8x8_12,
+                   &idct8x8_64_add_12_sse2, 6225, VPX_BITS_12)));
+#endif  // HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_SSSE3 && CONFIG_USE_X86INC && ARCH_X86_64 && \
+    !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, FwdTrans8x8DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct8x8_ssse3, &vpx_idct8x8_64_add_ssse3, 0,
+                   VPX_BITS_8)));
+#endif
+
+#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    MSA, FwdTrans8x8DCT,
+    ::testing::Values(
+        make_tuple(&vpx_fdct8x8_msa, &vpx_idct8x8_64_add_msa, 0, VPX_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    MSA, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&vp9_fht8x8_msa, &vp9_iht8x8_64_add_msa, 0, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_msa, &vp9_iht8x8_64_add_msa, 1, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_msa, &vp9_iht8x8_64_add_msa, 2, VPX_BITS_8),
+        make_tuple(&vp9_fht8x8_msa, &vp9_iht8x8_64_add_msa, 3, VPX_BITS_8)));
+#endif  // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+}  // namespace
diff --git a/libvpx/libvpx/test/frame_size_tests.cc b/libvpx/libvpx/test/frame_size_tests.cc
new file mode 100644
index 0000000..d39c8f6
--- /dev/null
+++ b/libvpx/libvpx/test/frame_size_tests.cc
@@ -0,0 +1,96 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/video_source.h"
+
+namespace {
+
+class VP9FrameSizeTestsLarge
+    : public ::libvpx_test::EncoderTest,
+      public ::testing::Test {
+ protected:
+  VP9FrameSizeTestsLarge() : EncoderTest(&::libvpx_test::kVP9),
+                             expected_res_(VPX_CODEC_OK) {}
+  virtual ~VP9FrameSizeTestsLarge() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(::libvpx_test::kRealTime);
+  }
+
+  virtual bool HandleDecodeResult(const vpx_codec_err_t res_dec,
+                                  const libvpx_test::VideoSource& /*video*/,
+                                  libvpx_test::Decoder *decoder) {
+    EXPECT_EQ(expected_res_, res_dec) << decoder->DecodeError();
+    return !::testing::Test::HasFailure();
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_CPUUSED, 7);
+      encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+      encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+      encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+      encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+    }
+  }
+
+  int expected_res_;
+};
+
+TEST_F(VP9FrameSizeTestsLarge, TestInvalidSizes) {
+  ::libvpx_test::RandomVideoSource video;
+
+#if CONFIG_SIZE_LIMIT
+  video.SetSize(DECODE_WIDTH_LIMIT + 16, DECODE_HEIGHT_LIMIT + 16);
+  video.set_limit(2);
+  expected_res_ = VPX_CODEC_CORRUPT_FRAME;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+#endif
+}
+
+TEST_F(VP9FrameSizeTestsLarge, ValidSizes) {
+  ::libvpx_test::RandomVideoSource video;
+
+#if CONFIG_SIZE_LIMIT
+  video.SetSize(DECODE_WIDTH_LIMIT, DECODE_HEIGHT_LIMIT);
+  video.set_limit(2);
+  expected_res_ = VPX_CODEC_OK;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+#else
+  // This test produces a pretty large single frame allocation,  (roughly
+  // 25 megabits). The encoder allocates a good number of these frames
+  // one for each lag in frames (for 2 pass), and then one for each possible
+  // reference buffer (8) - we can end up with up to 30 buffers of roughly this
+  // size or almost 1 gig of memory.
+  // In total the allocations will exceed 2GiB which may cause a failure with
+  // mingw + wine, use a smaller size in that case.
+#if defined(_WIN32) && !defined(_WIN64) || defined(__OS2__)
+  video.SetSize(4096, 3072);
+#else
+  video.SetSize(4096, 4096);
+#endif
+  video.set_limit(2);
+  expected_res_ = VPX_CODEC_OK;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+#endif
+}
+
+TEST_F(VP9FrameSizeTestsLarge, OneByOneVideo) {
+  ::libvpx_test::RandomVideoSource video;
+
+  video.SetSize(1, 1);
+  video.set_limit(2);
+  expected_res_ = VPX_CODEC_OK;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+}  // namespace
diff --git a/libvpx/libvpx/test/hadamard_test.cc b/libvpx/libvpx/test/hadamard_test.cc
new file mode 100644
index 0000000..7a5bd5b
--- /dev/null
+++ b/libvpx/libvpx/test/hadamard_test.cc
@@ -0,0 +1,220 @@
+/*
+ *  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <algorithm>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_dsp_rtcd.h"
+
+#include "test/acm_random.h"
+#include "test/register_state_check.h"
+
+namespace {
+
+using ::libvpx_test::ACMRandom;
+
+typedef void (*HadamardFunc)(const int16_t *a, int a_stride, int16_t *b);
+
+void hadamard_loop(const int16_t *a, int a_stride, int16_t *out) {
+  int16_t b[8];
+  for (int i = 0; i < 8; i += 2) {
+    b[i + 0] = a[i * a_stride] + a[(i + 1) * a_stride];
+    b[i + 1] = a[i * a_stride] - a[(i + 1) * a_stride];
+  }
+  int16_t c[8];
+  for (int i = 0; i < 8; i += 4) {
+    c[i + 0] = b[i + 0] + b[i + 2];
+    c[i + 1] = b[i + 1] + b[i + 3];
+    c[i + 2] = b[i + 0] - b[i + 2];
+    c[i + 3] = b[i + 1] - b[i + 3];
+  }
+  out[0] = c[0] + c[4];
+  out[7] = c[1] + c[5];
+  out[3] = c[2] + c[6];
+  out[4] = c[3] + c[7];
+  out[2] = c[0] - c[4];
+  out[6] = c[1] - c[5];
+  out[1] = c[2] - c[6];
+  out[5] = c[3] - c[7];
+}
+
+void reference_hadamard8x8(const int16_t *a, int a_stride, int16_t *b) {
+  int16_t buf[64];
+  for (int i = 0; i < 8; ++i) {
+    hadamard_loop(a + i, a_stride, buf + i * 8);
+  }
+
+  for (int i = 0; i < 8; ++i) {
+    hadamard_loop(buf + i, 8, b + i * 8);
+  }
+}
+
+void reference_hadamard16x16(const int16_t *a, int a_stride, int16_t *b) {
+  /* The source is a 16x16 block. The destination is rearranged to 8x32.
+   * Input is 9 bit. */
+  reference_hadamard8x8(a + 0 + 0 * a_stride, a_stride, b + 0);
+  reference_hadamard8x8(a + 8 + 0 * a_stride, a_stride, b + 64);
+  reference_hadamard8x8(a + 0 + 8 * a_stride, a_stride, b + 128);
+  reference_hadamard8x8(a + 8 + 8 * a_stride, a_stride, b + 192);
+
+  /* Overlay the 8x8 blocks and combine. */
+  for (int i = 0; i < 64; ++i) {
+    /* 8x8 steps the range up to 15 bits. */
+    const int16_t a0 = b[0];
+    const int16_t a1 = b[64];
+    const int16_t a2 = b[128];
+    const int16_t a3 = b[192];
+
+    /* Prevent the result from escaping int16_t. */
+    const int16_t b0 = (a0 + a1) >> 1;
+    const int16_t b1 = (a0 - a1) >> 1;
+    const int16_t b2 = (a2 + a3) >> 1;
+    const int16_t b3 = (a2 - a3) >> 1;
+
+    /* Store a 16 bit value. */
+    b[  0] = b0 + b2;
+    b[ 64] = b1 + b3;
+    b[128] = b0 - b2;
+    b[192] = b1 - b3;
+
+    ++b;
+  }
+}
+
+class HadamardTestBase : public ::testing::TestWithParam<HadamardFunc> {
+ public:
+  virtual void SetUp() {
+    h_func_ = GetParam();
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+  }
+
+ protected:
+  HadamardFunc h_func_;
+  ACMRandom rnd_;
+};
+
+class Hadamard8x8Test : public HadamardTestBase {};
+
+TEST_P(Hadamard8x8Test, CompareReferenceRandom) {
+  DECLARE_ALIGNED(16, int16_t, a[64]);
+  DECLARE_ALIGNED(16, int16_t, b[64]);
+  int16_t b_ref[64];
+  for (int i = 0; i < 64; ++i) {
+    a[i] = rnd_.Rand9Signed();
+  }
+  memset(b, 0, sizeof(b));
+  memset(b_ref, 0, sizeof(b_ref));
+
+  reference_hadamard8x8(a, 8, b_ref);
+  ASM_REGISTER_STATE_CHECK(h_func_(a, 8, b));
+
+  // The order of the output is not important. Sort before checking.
+  std::sort(b, b + 64);
+  std::sort(b_ref, b_ref + 64);
+  EXPECT_EQ(0, memcmp(b, b_ref, sizeof(b)));
+}
+
+TEST_P(Hadamard8x8Test, VaryStride) {
+  DECLARE_ALIGNED(16, int16_t, a[64 * 8]);
+  DECLARE_ALIGNED(16, int16_t, b[64]);
+  int16_t b_ref[64];
+  for (int i = 0; i < 64 * 8; ++i) {
+    a[i] = rnd_.Rand9Signed();
+  }
+
+  for (int i = 8; i < 64; i += 8) {
+    memset(b, 0, sizeof(b));
+    memset(b_ref, 0, sizeof(b_ref));
+
+    reference_hadamard8x8(a, i, b_ref);
+    ASM_REGISTER_STATE_CHECK(h_func_(a, i, b));
+
+    // The order of the output is not important. Sort before checking.
+    std::sort(b, b + 64);
+    std::sort(b_ref, b_ref + 64);
+    EXPECT_EQ(0, memcmp(b, b_ref, sizeof(b)));
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(C, Hadamard8x8Test,
+                        ::testing::Values(&vpx_hadamard_8x8_c));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, Hadamard8x8Test,
+                        ::testing::Values(&vpx_hadamard_8x8_sse2));
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3 && CONFIG_USE_X86INC && ARCH_X86_64
+INSTANTIATE_TEST_CASE_P(SSSE3, Hadamard8x8Test,
+                        ::testing::Values(&vpx_hadamard_8x8_ssse3));
+#endif  // HAVE_SSSE3 && CONFIG_USE_X86INC && ARCH_X86_64
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, Hadamard8x8Test,
+                        ::testing::Values(&vpx_hadamard_8x8_neon));
+#endif  // HAVE_NEON
+
+class Hadamard16x16Test : public HadamardTestBase {};
+
+TEST_P(Hadamard16x16Test, CompareReferenceRandom) {
+  DECLARE_ALIGNED(16, int16_t, a[16 * 16]);
+  DECLARE_ALIGNED(16, int16_t, b[16 * 16]);
+  int16_t b_ref[16 * 16];
+  for (int i = 0; i < 16 * 16; ++i) {
+    a[i] = rnd_.Rand9Signed();
+  }
+  memset(b, 0, sizeof(b));
+  memset(b_ref, 0, sizeof(b_ref));
+
+  reference_hadamard16x16(a, 16, b_ref);
+  ASM_REGISTER_STATE_CHECK(h_func_(a, 16, b));
+
+  // The order of the output is not important. Sort before checking.
+  std::sort(b, b + 16 * 16);
+  std::sort(b_ref, b_ref + 16 * 16);
+  EXPECT_EQ(0, memcmp(b, b_ref, sizeof(b)));
+}
+
+TEST_P(Hadamard16x16Test, VaryStride) {
+  DECLARE_ALIGNED(16, int16_t, a[16 * 16 * 8]);
+  DECLARE_ALIGNED(16, int16_t, b[16 * 16]);
+  int16_t b_ref[16 * 16];
+  for (int i = 0; i < 16 * 16 * 8; ++i) {
+    a[i] = rnd_.Rand9Signed();
+  }
+
+  for (int i = 8; i < 64; i += 8) {
+    memset(b, 0, sizeof(b));
+    memset(b_ref, 0, sizeof(b_ref));
+
+    reference_hadamard16x16(a, i, b_ref);
+    ASM_REGISTER_STATE_CHECK(h_func_(a, i, b));
+
+    // The order of the output is not important. Sort before checking.
+    std::sort(b, b + 16 * 16);
+    std::sort(b_ref, b_ref + 16 * 16);
+    EXPECT_EQ(0, memcmp(b, b_ref, sizeof(b)));
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(C, Hadamard16x16Test,
+                        ::testing::Values(&vpx_hadamard_16x16_c));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, Hadamard16x16Test,
+                        ::testing::Values(&vpx_hadamard_16x16_sse2));
+#endif  // HAVE_SSE2
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, Hadamard16x16Test,
+                        ::testing::Values(&vpx_hadamard_16x16_neon));
+#endif  // HAVE_NEON
+}  // namespace
diff --git a/libvpx/libvpx/test/i420_video_source.h b/libvpx/libvpx/test/i420_video_source.h
new file mode 100644
index 0000000..0a18480
--- /dev/null
+++ b/libvpx/libvpx/test/i420_video_source.h
@@ -0,0 +1,36 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef TEST_I420_VIDEO_SOURCE_H_
+#define TEST_I420_VIDEO_SOURCE_H_
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+
+#include "test/yuv_video_source.h"
+
+namespace libvpx_test {
+
+// This class extends VideoSource to allow parsing of raw yv12
+// so that we can do actual file encodes.
+class I420VideoSource : public YUVVideoSource {
+ public:
+  I420VideoSource(const std::string &file_name,
+                  unsigned int width, unsigned int height,
+                  int rate_numerator, int rate_denominator,
+                  unsigned int start, int limit)
+      : YUVVideoSource(file_name, VPX_IMG_FMT_I420,
+                       width, height,
+                       rate_numerator, rate_denominator,
+                       start, limit) {}
+};
+
+}  // namespace libvpx_test
+
+#endif  // TEST_I420_VIDEO_SOURCE_H_
diff --git a/libvpx/libvpx/test/idct8x8_test.cc b/libvpx/libvpx/test/idct8x8_test.cc
new file mode 100644
index 0000000..7f9d751
--- /dev/null
+++ b/libvpx/libvpx/test/idct8x8_test.cc
@@ -0,0 +1,101 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "vpx/vpx_integer.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+
+#ifdef _MSC_VER
+static int round(double x) {
+  if (x < 0)
+    return static_cast<int>(ceil(x - 0.5));
+  else
+    return static_cast<int>(floor(x + 0.5));
+}
+#endif
+
+void reference_dct_1d(double input[8], double output[8]) {
+  const double kPi = 3.141592653589793238462643383279502884;
+  const double kInvSqrt2 = 0.707106781186547524400844362104;
+  for (int k = 0; k < 8; k++) {
+    output[k] = 0.0;
+    for (int n = 0; n < 8; n++)
+      output[k] += input[n]*cos(kPi*(2*n+1)*k/16.0);
+    if (k == 0)
+      output[k] = output[k]*kInvSqrt2;
+  }
+}
+
+void reference_dct_2d(int16_t input[64], double output[64]) {
+  // First transform columns
+  for (int i = 0; i < 8; ++i) {
+    double temp_in[8], temp_out[8];
+    for (int j = 0; j < 8; ++j)
+      temp_in[j] = input[j*8 + i];
+    reference_dct_1d(temp_in, temp_out);
+    for (int j = 0; j < 8; ++j)
+      output[j*8 + i] = temp_out[j];
+  }
+  // Then transform rows
+  for (int i = 0; i < 8; ++i) {
+    double temp_in[8], temp_out[8];
+    for (int j = 0; j < 8; ++j)
+      temp_in[j] = output[j + i*8];
+    reference_dct_1d(temp_in, temp_out);
+    for (int j = 0; j < 8; ++j)
+      output[j + i*8] = temp_out[j];
+  }
+  // Scale by some magic number
+  for (int i = 0; i < 64; ++i)
+    output[i] *= 2;
+}
+
+TEST(VP9Idct8x8Test, AccuracyCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 10000;
+  for (int i = 0; i < count_test_block; ++i) {
+    int16_t input[64];
+    tran_low_t coeff[64];
+    double output_r[64];
+    uint8_t dst[64], src[64];
+
+    for (int j = 0; j < 64; ++j) {
+      src[j] = rnd.Rand8();
+      dst[j] = rnd.Rand8();
+    }
+    // Initialize a test block with input range [-255, 255].
+    for (int j = 0; j < 64; ++j)
+      input[j] = src[j] - dst[j];
+
+    reference_dct_2d(input, output_r);
+    for (int j = 0; j < 64; ++j)
+      coeff[j] = round(output_r[j]);
+    vpx_idct8x8_64_add_c(coeff, dst, 8);
+    for (int j = 0; j < 64; ++j) {
+      const int diff = dst[j] - src[j];
+      const int error = diff * diff;
+      EXPECT_GE(1, error)
+          << "Error: 8x8 FDCT/IDCT has error " << error
+          << " at index " << j;
+    }
+  }
+}
+
+}  // namespace
diff --git a/libvpx/libvpx/test/idct_test.cc b/libvpx/libvpx/test/idct_test.cc
new file mode 100644
index 0000000..39db3e4
--- /dev/null
+++ b/libvpx/libvpx/test/idct_test.cc
@@ -0,0 +1,121 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vp8_rtcd.h"
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "vpx/vpx_integer.h"
+
+typedef void (*IdctFunc)(int16_t *input, unsigned char *pred_ptr,
+                         int pred_stride, unsigned char *dst_ptr,
+                         int dst_stride);
+namespace {
+class IDCTTest : public ::testing::TestWithParam<IdctFunc> {
+ protected:
+  virtual void SetUp() {
+    int i;
+
+    UUT = GetParam();
+    memset(input, 0, sizeof(input));
+    /* Set up guard blocks */
+    for (i = 0; i < 256; i++) output[i] = ((i & 0xF) < 4 && (i < 64)) ? 0 : -1;
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+  IdctFunc UUT;
+  int16_t input[16];
+  unsigned char output[256];
+  unsigned char predict[256];
+};
+
+TEST_P(IDCTTest, TestGuardBlocks) {
+  int i;
+
+  for (i = 0; i < 256; i++)
+    if ((i & 0xF) < 4 && i < 64)
+      EXPECT_EQ(0, output[i]) << i;
+    else
+      EXPECT_EQ(255, output[i]);
+}
+
+TEST_P(IDCTTest, TestAllZeros) {
+  int i;
+
+  ASM_REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
+
+  for (i = 0; i < 256; i++)
+    if ((i & 0xF) < 4 && i < 64)
+      EXPECT_EQ(0, output[i]) << "i==" << i;
+    else
+      EXPECT_EQ(255, output[i]) << "i==" << i;
+}
+
+TEST_P(IDCTTest, TestAllOnes) {
+  int i;
+
+  input[0] = 4;
+  ASM_REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
+
+  for (i = 0; i < 256; i++)
+    if ((i & 0xF) < 4 && i < 64)
+      EXPECT_EQ(1, output[i]) << "i==" << i;
+    else
+      EXPECT_EQ(255, output[i]) << "i==" << i;
+}
+
+TEST_P(IDCTTest, TestAddOne) {
+  int i;
+
+  for (i = 0; i < 256; i++) predict[i] = i;
+  input[0] = 4;
+  ASM_REGISTER_STATE_CHECK(UUT(input, predict, 16, output, 16));
+
+  for (i = 0; i < 256; i++)
+    if ((i & 0xF) < 4 && i < 64)
+      EXPECT_EQ(i + 1, output[i]) << "i==" << i;
+    else
+      EXPECT_EQ(255, output[i]) << "i==" << i;
+}
+
+TEST_P(IDCTTest, TestWithData) {
+  int i;
+
+  for (i = 0; i < 16; i++) input[i] = i;
+
+  ASM_REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
+
+  for (i = 0; i < 256; i++)
+    if ((i & 0xF) > 3 || i > 63)
+      EXPECT_EQ(255, output[i]) << "i==" << i;
+    else if (i == 0)
+      EXPECT_EQ(11, output[i]) << "i==" << i;
+    else if (i == 34)
+      EXPECT_EQ(1, output[i]) << "i==" << i;
+    else if (i == 2 || i == 17 || i == 32)
+      EXPECT_EQ(3, output[i]) << "i==" << i;
+    else
+      EXPECT_EQ(0, output[i]) << "i==" << i;
+}
+
+INSTANTIATE_TEST_CASE_P(C, IDCTTest, ::testing::Values(vp8_short_idct4x4llm_c));
+#if HAVE_MMX
+INSTANTIATE_TEST_CASE_P(MMX, IDCTTest,
+                        ::testing::Values(vp8_short_idct4x4llm_mmx));
+#endif
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(MSA, IDCTTest,
+                        ::testing::Values(vp8_short_idct4x4llm_msa));
+#endif
+}
diff --git a/libvpx/libvpx/test/invalid_file_test.cc b/libvpx/libvpx/test/invalid_file_test.cc
new file mode 100644
index 0000000..f4241eb
--- /dev/null
+++ b/libvpx/libvpx/test/invalid_file_test.cc
@@ -0,0 +1,182 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+
+struct DecodeParam {
+  int threads;
+  const char *filename;
+};
+
+std::ostream &operator<<(std::ostream &os, const DecodeParam &dp) {
+  return os << "threads: " << dp.threads << " file: " << dp.filename;
+}
+
+class InvalidFileTest
+    : public ::libvpx_test::DecoderTest,
+      public ::libvpx_test::CodecTestWithParam<DecodeParam> {
+ protected:
+  InvalidFileTest() : DecoderTest(GET_PARAM(0)), res_file_(NULL) {}
+
+  virtual ~InvalidFileTest() {
+    if (res_file_ != NULL)
+      fclose(res_file_);
+  }
+
+  void OpenResFile(const std::string &res_file_name_) {
+    res_file_ = libvpx_test::OpenTestDataFile(res_file_name_);
+    ASSERT_TRUE(res_file_ != NULL) << "Result file open failed. Filename: "
+        << res_file_name_;
+  }
+
+  virtual bool HandleDecodeResult(
+      const vpx_codec_err_t res_dec,
+      const libvpx_test::CompressedVideoSource &video,
+      libvpx_test::Decoder *decoder) {
+    EXPECT_TRUE(res_file_ != NULL);
+    int expected_res_dec;
+
+    // Read integer result.
+    const int res = fscanf(res_file_, "%d", &expected_res_dec);
+    EXPECT_NE(res, EOF) << "Read result data failed";
+
+    // Check results match.
+    const DecodeParam input = GET_PARAM(1);
+    if (input.threads > 1) {
+      // The serial decode check is too strict for tile-threaded decoding as
+      // there is no guarantee on the decode order nor which specific error
+      // will take precedence. Currently a tile-level error is not forwarded so
+      // the frame will simply be marked corrupt.
+      EXPECT_TRUE(res_dec == expected_res_dec ||
+                  res_dec == VPX_CODEC_CORRUPT_FRAME)
+          << "Results don't match: frame number = " << video.frame_number()
+          << ". (" << decoder->DecodeError() << "). Expected: "
+          << expected_res_dec << " or " << VPX_CODEC_CORRUPT_FRAME;
+    } else {
+      EXPECT_EQ(expected_res_dec, res_dec)
+          << "Results don't match: frame number = " << video.frame_number()
+          << ". (" << decoder->DecodeError() << ")";
+    }
+
+    return !HasFailure();
+  }
+
+  void RunTest() {
+    const DecodeParam input = GET_PARAM(1);
+    libvpx_test::CompressedVideoSource *video = NULL;
+    vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+    cfg.threads = input.threads;
+    const std::string filename = input.filename;
+
+    // Open compressed video file.
+    if (filename.substr(filename.length() - 3, 3) == "ivf") {
+      video = new libvpx_test::IVFVideoSource(filename);
+    } else if (filename.substr(filename.length() - 4, 4) == "webm") {
+#if CONFIG_WEBM_IO
+      video = new libvpx_test::WebMVideoSource(filename);
+#else
+      fprintf(stderr, "WebM IO is disabled, skipping test vector %s\n",
+              filename.c_str());
+      return;
+#endif
+    }
+    video->Init();
+
+    // Construct result file name. The file holds a list of expected integer
+    // results, one for each decoded frame.  Any result that doesn't match
+    // the files list will cause a test failure.
+    const std::string res_filename = filename + ".res";
+    OpenResFile(res_filename);
+
+    // Decode frame, and check the md5 matching.
+    ASSERT_NO_FATAL_FAILURE(RunLoop(video, cfg));
+    delete video;
+  }
+
+ private:
+  FILE *res_file_;
+};
+
+TEST_P(InvalidFileTest, ReturnCode) {
+  RunTest();
+}
+
+const DecodeParam kVP9InvalidFileTests[] = {
+  {1, "invalid-vp90-02-v2.webm"},
+#if CONFIG_VP9_HIGHBITDEPTH
+  {1, "invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf"},
+#endif
+  {1, "invalid-vp90-03-v3.webm"},
+  {1, "invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf"},
+  {1, "invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf"},
+  {1, "invalid-vp90-2-12-droppable_1.ivf.s3676_r01-05_b6-.ivf"},
+  {1, "invalid-vp90-2-05-resize.ivf.s59293_r01-05_b6-.ivf"},
+  {1, "invalid-vp90-2-09-subpixel-00.ivf.s20492_r01-05_b6-.v2.ivf"},
+  {1, "invalid-vp91-2-mixedrefcsp-444to420.ivf"},
+  {1, "invalid-vp90-2-12-droppable_1.ivf.s73804_r01-05_b6-.ivf"},
+  {1, "invalid-vp90-2-03-size-224x196.webm.ivf.s44156_r01-05_b6-.ivf"},
+  {1, "invalid-vp90-2-03-size-202x210.webm.ivf.s113306_r01-05_b6-.ivf"},
+};
+
+VP9_INSTANTIATE_TEST_CASE(InvalidFileTest,
+                          ::testing::ValuesIn(kVP9InvalidFileTests));
+
+// This class will include test vectors that are expected to fail
+// peek. However they are still expected to have no fatal failures.
+class InvalidFileInvalidPeekTest : public InvalidFileTest {
+ protected:
+  InvalidFileInvalidPeekTest() : InvalidFileTest() {}
+  virtual void HandlePeekResult(libvpx_test::Decoder *const /*decoder*/,
+                                libvpx_test::CompressedVideoSource* /*video*/,
+                                const vpx_codec_err_t /*res_peek*/) {}
+};
+
+TEST_P(InvalidFileInvalidPeekTest, ReturnCode) {
+  RunTest();
+}
+
+const DecodeParam kVP9InvalidFileInvalidPeekTests[] = {
+  {1, "invalid-vp90-01-v3.webm"},
+};
+
+VP9_INSTANTIATE_TEST_CASE(InvalidFileInvalidPeekTest,
+                          ::testing::ValuesIn(kVP9InvalidFileInvalidPeekTests));
+
+const DecodeParam kMultiThreadedVP9InvalidFileTests[] = {
+  {4, "invalid-vp90-2-08-tile_1x4_frame_parallel_all_key.webm"},
+  {4, "invalid-"
+      "vp90-2-08-tile_1x2_frame_parallel.webm.ivf.s47039_r01-05_b6-.ivf"},
+  {4, "invalid-vp90-2-08-tile_1x8_frame_parallel.webm.ivf.s288_r01-05_b6-.ivf"},
+  {2, "invalid-vp90-2-09-aq2.webm.ivf.s3984_r01-05_b6-.v2.ivf"},
+  {4, "invalid-vp90-2-09-subpixel-00.ivf.s19552_r01-05_b6-.v2.ivf"},
+};
+
+INSTANTIATE_TEST_CASE_P(
+    VP9MultiThreaded, InvalidFileTest,
+    ::testing::Combine(
+        ::testing::Values(
+            static_cast<const libvpx_test::CodecFactory*>(&libvpx_test::kVP9)),
+        ::testing::ValuesIn(kMultiThreadedVP9InvalidFileTests)));
+}  // namespace
diff --git a/libvpx/libvpx/test/ivf_video_source.h b/libvpx/libvpx/test/ivf_video_source.h
new file mode 100644
index 0000000..824a39d
--- /dev/null
+++ b/libvpx/libvpx/test/ivf_video_source.h
@@ -0,0 +1,111 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef TEST_IVF_VIDEO_SOURCE_H_
+#define TEST_IVF_VIDEO_SOURCE_H_
+#include <cstdio>
+#include <cstdlib>
+#include <new>
+#include <string>
+#include "test/video_source.h"
+
+namespace libvpx_test {
+const unsigned int kCodeBufferSize = 256 * 1024;
+const unsigned int kIvfFileHdrSize = 32;
+const unsigned int kIvfFrameHdrSize = 12;
+
+static unsigned int MemGetLe32(const uint8_t *mem) {
+  return (mem[3] << 24) | (mem[2] << 16) | (mem[1] << 8) | (mem[0]);
+}
+
+// This class extends VideoSource to allow parsing of ivf files,
+// so that we can do actual file decodes.
+class IVFVideoSource : public CompressedVideoSource {
+ public:
+  explicit IVFVideoSource(const std::string &file_name)
+      : file_name_(file_name),
+        input_file_(NULL),
+        compressed_frame_buf_(NULL),
+        frame_sz_(0),
+        frame_(0),
+        end_of_file_(false) {
+  }
+
+  virtual ~IVFVideoSource() {
+    delete[] compressed_frame_buf_;
+
+    if (input_file_)
+      fclose(input_file_);
+  }
+
+  virtual void Init() {
+    // Allocate a buffer for read in the compressed video frame.
+    compressed_frame_buf_ = new uint8_t[libvpx_test::kCodeBufferSize];
+    ASSERT_TRUE(compressed_frame_buf_ != NULL)
+        << "Allocate frame buffer failed";
+  }
+
+  virtual void Begin() {
+    input_file_ = OpenTestDataFile(file_name_);
+    ASSERT_TRUE(input_file_ != NULL) << "Input file open failed. Filename: "
+        << file_name_;
+
+    // Read file header
+    uint8_t file_hdr[kIvfFileHdrSize];
+    ASSERT_EQ(kIvfFileHdrSize, fread(file_hdr, 1, kIvfFileHdrSize, input_file_))
+        << "File header read failed.";
+    // Check file header
+    ASSERT_TRUE(file_hdr[0] == 'D' && file_hdr[1] == 'K' && file_hdr[2] == 'I'
+                && file_hdr[3] == 'F') << "Input is not an IVF file.";
+
+    FillFrame();
+  }
+
+  virtual void Next() {
+    ++frame_;
+    FillFrame();
+  }
+
+  void FillFrame() {
+    ASSERT_TRUE(input_file_ != NULL);
+    uint8_t frame_hdr[kIvfFrameHdrSize];
+    // Check frame header and read a frame from input_file.
+    if (fread(frame_hdr, 1, kIvfFrameHdrSize, input_file_)
+        != kIvfFrameHdrSize) {
+      end_of_file_ = true;
+    } else {
+      end_of_file_ = false;
+
+      frame_sz_ = MemGetLe32(frame_hdr);
+      ASSERT_LE(frame_sz_, kCodeBufferSize)
+          << "Frame is too big for allocated code buffer";
+      ASSERT_EQ(frame_sz_,
+                fread(compressed_frame_buf_, 1, frame_sz_, input_file_))
+          << "Failed to read complete frame";
+    }
+  }
+
+  virtual const uint8_t *cxdata() const {
+    return end_of_file_ ? NULL : compressed_frame_buf_;
+  }
+  virtual size_t frame_size() const { return frame_sz_; }
+  virtual unsigned int frame_number() const { return frame_; }
+
+ protected:
+  std::string file_name_;
+  FILE *input_file_;
+  uint8_t *compressed_frame_buf_;
+  size_t frame_sz_;
+  unsigned int frame_;
+  bool end_of_file_;
+};
+
+}  // namespace libvpx_test
+
+#endif  // TEST_IVF_VIDEO_SOURCE_H_
diff --git a/libvpx/libvpx/test/keyframe_test.cc b/libvpx/libvpx/test/keyframe_test.cc
new file mode 100644
index 0000000..d8b21a1
--- /dev/null
+++ b/libvpx/libvpx/test/keyframe_test.cc
@@ -0,0 +1,145 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <climits>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+class KeyframeTest : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ protected:
+  KeyframeTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~KeyframeTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    kf_count_ = 0;
+    kf_count_max_ = INT_MAX;
+    kf_do_force_kf_ = false;
+    set_cpu_used_ = 0;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (kf_do_force_kf_)
+      frame_flags_ = (video->frame() % 3) ? 0 : VPX_EFLAG_FORCE_KF;
+    if (set_cpu_used_ && video->frame() == 1)
+      encoder->Control(VP8E_SET_CPUUSED, set_cpu_used_);
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->data.frame.flags & VPX_FRAME_IS_KEY) {
+      kf_pts_list_.push_back(pkt->data.frame.pts);
+      kf_count_++;
+      abort_ |= kf_count_ > kf_count_max_;
+    }
+  }
+
+  bool kf_do_force_kf_;
+  int kf_count_;
+  int kf_count_max_;
+  std::vector<vpx_codec_pts_t> kf_pts_list_;
+  int set_cpu_used_;
+};
+
+TEST_P(KeyframeTest, TestRandomVideoSource) {
+  // Validate that encoding the RandomVideoSource produces multiple keyframes.
+  // This validates the results of the TestDisableKeyframes test.
+  kf_count_max_ = 2;  // early exit successful tests.
+
+  ::libvpx_test::RandomVideoSource video;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  // In realtime mode - auto placed keyframes are exceedingly rare,  don't
+  // bother with this check   if(GetParam() > 0)
+  if (GET_PARAM(1) > 0)
+    EXPECT_GT(kf_count_, 1);
+}
+
+TEST_P(KeyframeTest, TestDisableKeyframes) {
+  cfg_.kf_mode = VPX_KF_DISABLED;
+  kf_count_max_ = 1;  // early exit failed tests.
+
+  ::libvpx_test::RandomVideoSource video;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  EXPECT_EQ(1, kf_count_);
+}
+
+TEST_P(KeyframeTest, TestForceKeyframe) {
+  cfg_.kf_mode = VPX_KF_DISABLED;
+  kf_do_force_kf_ = true;
+
+  ::libvpx_test::DummyVideoSource video;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  // verify that every third frame is a keyframe.
+  for (std::vector<vpx_codec_pts_t>::const_iterator iter = kf_pts_list_.begin();
+       iter != kf_pts_list_.end(); ++iter) {
+    ASSERT_EQ(0, *iter % 3) << "Unexpected keyframe at frame " << *iter;
+  }
+}
+
+TEST_P(KeyframeTest, TestKeyframeMaxDistance) {
+  cfg_.kf_max_dist = 25;
+
+  ::libvpx_test::DummyVideoSource video;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  // verify that keyframe interval matches kf_max_dist
+  for (std::vector<vpx_codec_pts_t>::const_iterator iter = kf_pts_list_.begin();
+       iter != kf_pts_list_.end(); ++iter) {
+    ASSERT_EQ(0, *iter % 25) << "Unexpected keyframe at frame " << *iter;
+  }
+}
+
+TEST_P(KeyframeTest, TestAutoKeyframe) {
+  cfg_.kf_mode = VPX_KF_AUTO;
+  kf_do_force_kf_ = false;
+
+  // Force a deterministic speed step in Real Time mode, as the faster modes
+  // may not produce a keyframe like we expect. This is necessary when running
+  // on very slow environments (like Valgrind). The step -11 was determined
+  // experimentally as the fastest mode that still throws the keyframe.
+  if (deadline_ == VPX_DL_REALTIME)
+    set_cpu_used_ = -11;
+
+  // This clip has a cut scene every 30 frames -> Frame 0, 30, 60, 90, 120.
+  // I check only the first 40 frames to make sure there's a keyframe at frame
+  // 0 and 30.
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 40);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  // In realtime mode - auto placed keyframes are exceedingly rare,  don't
+  // bother with this check
+  if (GET_PARAM(1) > 0)
+    EXPECT_EQ(2u, kf_pts_list_.size()) << " Not the right number of keyframes ";
+
+  // Verify that keyframes match the file keyframes in the file.
+  for (std::vector<vpx_codec_pts_t>::const_iterator iter = kf_pts_list_.begin();
+       iter != kf_pts_list_.end(); ++iter) {
+    if (deadline_ == VPX_DL_REALTIME && *iter > 0)
+      EXPECT_EQ(0, (*iter - 1) % 30) << "Unexpected keyframe at frame "
+        << *iter;
+    else
+      EXPECT_EQ(0, *iter % 30) << "Unexpected keyframe at frame " << *iter;
+  }
+}
+
+VP8_INSTANTIATE_TEST_CASE(KeyframeTest, ALL_TEST_MODES);
+}  // namespace
diff --git a/libvpx/libvpx/test/level_test.cc b/libvpx/libvpx/test/level_test.cc
new file mode 100644
index 0000000..62d0247
--- /dev/null
+++ b/libvpx/libvpx/test/level_test.cc
@@ -0,0 +1,119 @@
+/*
+ *  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+class LevelTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
+ protected:
+  LevelTest()
+     : EncoderTest(GET_PARAM(0)),
+       encoding_mode_(GET_PARAM(1)),
+       cpu_used_(GET_PARAM(2)),
+       min_gf_internal_(24),
+       target_level_(0),
+       level_(0) {}
+  virtual ~LevelTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+    if (encoding_mode_ != ::libvpx_test::kRealTime) {
+      cfg_.g_lag_in_frames = 25;
+      cfg_.rc_end_usage = VPX_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = VPX_CBR;
+    }
+    cfg_.rc_2pass_vbr_minsection_pct = 5;
+    cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+    cfg_.rc_target_bitrate = 400;
+    cfg_.rc_max_quantizer = 63;
+    cfg_.rc_min_quantizer = 0;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      encoder->Control(VP8E_SET_CPUUSED, cpu_used_);
+      encoder->Control(VP9E_SET_TARGET_LEVEL, target_level_);
+      encoder->Control(VP9E_SET_MIN_GF_INTERVAL, min_gf_internal_);
+      if (encoding_mode_ != ::libvpx_test::kRealTime) {
+        encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+        encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+      }
+    }
+    encoder->Control(VP9E_GET_LEVEL, &level_);
+    ASSERT_LE(level_, 51);
+    ASSERT_GE(level_, 0);
+  }
+
+  ::libvpx_test::TestMode encoding_mode_;
+  int cpu_used_;
+  int min_gf_internal_;
+  int target_level_;
+  int level_;
+};
+
+// Test for keeping level stats only
+TEST_P(LevelTest, TestTargetLevel0) {
+  ::libvpx_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       40);
+  target_level_ = 0;
+  min_gf_internal_ = 4;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  ASSERT_EQ(11, level_);
+
+  cfg_.rc_target_bitrate = 1600;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  ASSERT_EQ(20, level_);
+}
+
+// Test for level control being turned off
+TEST_P(LevelTest, TestTargetLevel255) {
+  ::libvpx_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       30);
+  target_level_ = 255;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+TEST_P(LevelTest, TestTargetLevelApi) {
+  ::libvpx_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0, 1);
+  static const vpx_codec_iface_t *codec = &vpx_codec_vp9_cx_algo;
+  vpx_codec_ctx_t enc;
+  vpx_codec_enc_cfg_t cfg;
+  EXPECT_EQ(VPX_CODEC_OK, vpx_codec_enc_config_default(codec, &cfg, 0));
+  EXPECT_EQ(VPX_CODEC_OK, vpx_codec_enc_init(&enc, codec, &cfg, 0));
+  for (int level = 0; level <= 256; ++level) {
+    if (level == 10 || level == 11 || level == 20 || level == 21 ||
+        level == 30 || level == 31 || level == 40 || level == 41 ||
+        level == 50 || level == 51 || level == 52 || level == 60 ||
+        level == 61 || level == 62 || level == 0 || level == 255)
+      EXPECT_EQ(VPX_CODEC_OK,
+                vpx_codec_control(&enc, VP9E_SET_TARGET_LEVEL, level));
+    else
+      EXPECT_EQ(VPX_CODEC_INVALID_PARAM,
+                vpx_codec_control(&enc, VP9E_SET_TARGET_LEVEL, level));
+  }
+  EXPECT_EQ(VPX_CODEC_OK, vpx_codec_destroy(&enc));
+}
+
+VP9_INSTANTIATE_TEST_CASE(LevelTest,
+                          ::testing::Values(::libvpx_test::kTwoPassGood,
+                                            ::libvpx_test::kOnePassGood),
+                          ::testing::Range(0, 9));
+}  // namespace
diff --git a/libvpx/libvpx/test/lpf_8_test.cc b/libvpx/libvpx/test/lpf_8_test.cc
new file mode 100644
index 0000000..94646e4
--- /dev/null
+++ b/libvpx/libvpx/test/lpf_8_test.cc
@@ -0,0 +1,672 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cmath>
+#include <cstdlib>
+#include <string>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vpx/vpx_integer.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+// Horizontally and Vertically need 32x32: 8  Coeffs preceeding filtered section
+//                                         16 Coefs within filtered section
+//                                         8  Coeffs following filtered section
+const int kNumCoeffs = 1024;
+
+const int number_of_iterations = 10000;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*loop_op_t)(uint16_t *s, int p, const uint8_t *blimit,
+                          const uint8_t *limit, const uint8_t *thresh,
+                          int bd);
+typedef void (*dual_loop_op_t)(uint16_t *s, int p, const uint8_t *blimit0,
+                               const uint8_t *limit0, const uint8_t *thresh0,
+                               const uint8_t *blimit1, const uint8_t *limit1,
+                               const uint8_t *thresh1, int bd);
+#else
+typedef void (*loop_op_t)(uint8_t *s, int p, const uint8_t *blimit,
+                          const uint8_t *limit, const uint8_t *thresh);
+typedef void (*dual_loop_op_t)(uint8_t *s, int p, const uint8_t *blimit0,
+                               const uint8_t *limit0, const uint8_t *thresh0,
+                               const uint8_t *blimit1, const uint8_t *limit1,
+                               const uint8_t *thresh1);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+typedef std::tr1::tuple<loop_op_t, loop_op_t, int> loop8_param_t;
+typedef std::tr1::tuple<dual_loop_op_t, dual_loop_op_t, int> dualloop8_param_t;
+
+class Loop8Test6Param : public ::testing::TestWithParam<loop8_param_t> {
+ public:
+  virtual ~Loop8Test6Param() {}
+  virtual void SetUp() {
+    loopfilter_op_ = GET_PARAM(0);
+    ref_loopfilter_op_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  int bit_depth_;
+  int mask_;
+  loop_op_t loopfilter_op_;
+  loop_op_t ref_loopfilter_op_;
+};
+
+class Loop8Test9Param : public ::testing::TestWithParam<dualloop8_param_t> {
+ public:
+  virtual ~Loop8Test9Param() {}
+  virtual void SetUp() {
+    loopfilter_op_ = GET_PARAM(0);
+    ref_loopfilter_op_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  int bit_depth_;
+  int mask_;
+  dual_loop_op_t loopfilter_op_;
+  dual_loop_op_t ref_loopfilter_op_;
+};
+
+TEST_P(Loop8Test6Param, OperationCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = number_of_iterations;
+#if CONFIG_VP9_HIGHBITDEPTH
+  int32_t bd = bit_depth_;
+  DECLARE_ALIGNED(16, uint16_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, ref_s[kNumCoeffs]);
+#else
+  DECLARE_ALIGNED(8, uint8_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(8, uint8_t, ref_s[kNumCoeffs]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  int err_count_total = 0;
+  int first_failure = -1;
+  for (int i = 0; i < count_test_block; ++i) {
+    int err_count = 0;
+    uint8_t tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t, blimit[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t, limit[16])  = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t, thresh[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    int32_t p = kNumCoeffs/32;
+
+    uint16_t tmp_s[kNumCoeffs];
+    int j = 0;
+    while (j < kNumCoeffs) {
+      uint8_t val = rnd.Rand8();
+      if (val & 0x80) {  // 50% chance to choose a new value.
+        tmp_s[j] = rnd.Rand16();
+        j++;
+      } else {  // 50% chance to repeat previous value in row X times
+        int k = 0;
+        while (k++ < ((val & 0x1f) + 1) && j < kNumCoeffs) {
+          if (j < 1) {
+            tmp_s[j] = rnd.Rand16();
+          } else if (val & 0x20) {  // Increment by an value within the limit
+            tmp_s[j] = (tmp_s[j - 1] + (*limit - 1));
+          } else {  // Decrement by an value within the limit
+            tmp_s[j] = (tmp_s[j - 1] - (*limit - 1));
+          }
+          j++;
+        }
+      }
+    }
+    for (j = 0; j < kNumCoeffs; j++) {
+      if (i % 2) {
+        s[j] = tmp_s[j] & mask_;
+      } else {
+        s[j] = tmp_s[p * (j % p) + j / p] & mask_;
+      }
+      ref_s[j] = s[j];
+    }
+#if CONFIG_VP9_HIGHBITDEPTH
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit, limit, thresh, bd);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh, bd));
+#else
+    ref_loopfilter_op_(ref_s+8+p*8, p, blimit, limit, thresh);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      err_count += ref_s[j] != s[j];
+    }
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Loop8Test6Param, C output doesn't match SSE2 "
+         "loopfilter output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(Loop8Test6Param, ValueCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = number_of_iterations;
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int32_t bd = bit_depth_;
+  DECLARE_ALIGNED(16, uint16_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, ref_s[kNumCoeffs]);
+#else
+  DECLARE_ALIGNED(8, uint8_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(8, uint8_t, ref_s[kNumCoeffs]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  int err_count_total = 0;
+  int first_failure = -1;
+
+  // NOTE: The code in vp9_loopfilter.c:update_sharpness computes mblim as a
+  // function of sharpness_lvl and the loopfilter lvl as:
+  // block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));
+  // ...
+  // memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
+  //        SIMD_WIDTH);
+  // This means that the largest value for mblim will occur when sharpness_lvl
+  // is equal to 0, and lvl is equal to its greatest value (MAX_LOOP_FILTER).
+  // In this case block_inside_limit will be equal to MAX_LOOP_FILTER and
+  // therefore mblim will be equal to (2 * (lvl + 2) + block_inside_limit) =
+  // 2 * (MAX_LOOP_FILTER + 2) + MAX_LOOP_FILTER = 3 * MAX_LOOP_FILTER + 4
+
+  for (int i = 0; i < count_test_block; ++i) {
+    int err_count = 0;
+    uint8_t tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t, blimit[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t, limit[16])  = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t, thresh[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    int32_t p = kNumCoeffs / 32;
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      s[j] = rnd.Rand16() & mask_;
+      ref_s[j] = s[j];
+    }
+#if CONFIG_VP9_HIGHBITDEPTH
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit, limit, thresh, bd);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh, bd));
+#else
+    ref_loopfilter_op_(ref_s+8+p*8, p, blimit, limit, thresh);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      err_count += ref_s[j] != s[j];
+    }
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Loop8Test6Param, C output doesn't match SSE2 "
+         "loopfilter output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(Loop8Test9Param, OperationCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = number_of_iterations;
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int32_t bd = bit_depth_;
+  DECLARE_ALIGNED(16, uint16_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, ref_s[kNumCoeffs]);
+#else
+  DECLARE_ALIGNED(8,  uint8_t,  s[kNumCoeffs]);
+  DECLARE_ALIGNED(8,  uint8_t,  ref_s[kNumCoeffs]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  int err_count_total = 0;
+  int first_failure = -1;
+  for (int i = 0; i < count_test_block; ++i) {
+    int err_count = 0;
+    uint8_t tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t, blimit0[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t, limit0[16])  = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t, thresh0[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t, blimit1[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t, limit1[16])  = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t, thresh1[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    int32_t p = kNumCoeffs / 32;
+    uint16_t tmp_s[kNumCoeffs];
+    int j = 0;
+    const uint8_t limit = *limit0 < *limit1 ? *limit0 : *limit1;
+    while (j < kNumCoeffs) {
+      uint8_t val = rnd.Rand8();
+      if (val & 0x80) {  // 50% chance to choose a new value.
+        tmp_s[j] = rnd.Rand16();
+        j++;
+      } else {  // 50% chance to repeat previous value in row X times.
+        int k = 0;
+        while (k++ < ((val & 0x1f) + 1) && j < kNumCoeffs) {
+          if (j < 1) {
+            tmp_s[j] = rnd.Rand16();
+          } else if (val & 0x20) {  // Increment by a value within the limit.
+            tmp_s[j] = (tmp_s[j - 1] + (limit - 1));
+          } else {  // Decrement by an value within the limit.
+            tmp_s[j] = (tmp_s[j - 1] - (limit - 1));
+          }
+          j++;
+        }
+      }
+    }
+    for (j = 0; j < kNumCoeffs; j++) {
+      if (i % 2) {
+        s[j] = tmp_s[j] & mask_;
+      } else {
+        s[j] = tmp_s[p * (j % p) + j / p] & mask_;
+      }
+      ref_s[j] = s[j];
+    }
+#if CONFIG_VP9_HIGHBITDEPTH
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0,
+                       blimit1, limit1, thresh1, bd);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0, thresh0,
+                       blimit1, limit1, thresh1, bd));
+#else
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0,
+                       blimit1, limit1, thresh1);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0, thresh0,
+                       blimit1, limit1, thresh1));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      err_count += ref_s[j] != s[j];
+    }
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Loop8Test9Param, C output doesn't match SSE2 "
+         "loopfilter output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(Loop8Test9Param, ValueCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = number_of_iterations;
+#if CONFIG_VP9_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, ref_s[kNumCoeffs]);
+#else
+  DECLARE_ALIGNED(8,  uint8_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(8,  uint8_t, ref_s[kNumCoeffs]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  int err_count_total = 0;
+  int first_failure = -1;
+  for (int i = 0; i < count_test_block; ++i) {
+    int err_count = 0;
+    uint8_t tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t, blimit0[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t, limit0[16])  = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t, thresh0[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t, blimit1[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t, limit1[16])  = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t, thresh1[16]) = {
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+        tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp
+    };
+    int32_t p = kNumCoeffs / 32;  // TODO(pdlf) can we have non-square here?
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      s[j] = rnd.Rand16() & mask_;
+      ref_s[j] = s[j];
+    }
+#if CONFIG_VP9_HIGHBITDEPTH
+    const int32_t bd = bit_depth_;
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0,
+                       blimit1, limit1, thresh1, bd);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0,
+                       thresh0, blimit1, limit1, thresh1, bd));
+#else
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0,
+                       blimit1, limit1, thresh1);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0, thresh0,
+                       blimit1, limit1, thresh1));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      err_count += ref_s[j] != s[j];
+    }
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Loop8Test9Param, C output doesn't match SSE2"
+         "loopfilter output. "
+      << "First failed at test case " << first_failure;
+}
+
+using std::tr1::make_tuple;
+
+#if HAVE_SSE2
+#if CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Loop8Test6Param,
+    ::testing::Values(
+        make_tuple(&vpx_highbd_lpf_horizontal_4_sse2,
+                   &vpx_highbd_lpf_horizontal_4_c, 8),
+        make_tuple(&vpx_highbd_lpf_vertical_4_sse2,
+                   &vpx_highbd_lpf_vertical_4_c, 8),
+        make_tuple(&vpx_highbd_lpf_horizontal_8_sse2,
+                   &vpx_highbd_lpf_horizontal_8_c, 8),
+        make_tuple(&vpx_highbd_lpf_horizontal_edge_8_sse2,
+                   &vpx_highbd_lpf_horizontal_edge_8_c, 8),
+        make_tuple(&vpx_highbd_lpf_horizontal_edge_16_sse2,
+                   &vpx_highbd_lpf_horizontal_edge_16_c, 8),
+        make_tuple(&vpx_highbd_lpf_vertical_8_sse2,
+                   &vpx_highbd_lpf_vertical_8_c, 8),
+        make_tuple(&vpx_highbd_lpf_vertical_16_sse2,
+                   &vpx_highbd_lpf_vertical_16_c, 8),
+        make_tuple(&vpx_highbd_lpf_horizontal_4_sse2,
+                   &vpx_highbd_lpf_horizontal_4_c, 10),
+        make_tuple(&vpx_highbd_lpf_vertical_4_sse2,
+                   &vpx_highbd_lpf_vertical_4_c, 10),
+        make_tuple(&vpx_highbd_lpf_horizontal_8_sse2,
+                   &vpx_highbd_lpf_horizontal_8_c, 10),
+        make_tuple(&vpx_highbd_lpf_horizontal_edge_8_sse2,
+                   &vpx_highbd_lpf_horizontal_edge_8_c, 10),
+        make_tuple(&vpx_highbd_lpf_horizontal_edge_16_sse2,
+                   &vpx_highbd_lpf_horizontal_edge_16_c, 10),
+        make_tuple(&vpx_highbd_lpf_vertical_8_sse2,
+                   &vpx_highbd_lpf_vertical_8_c, 10),
+        make_tuple(&vpx_highbd_lpf_vertical_16_sse2,
+                   &vpx_highbd_lpf_vertical_16_c, 10),
+        make_tuple(&vpx_highbd_lpf_horizontal_4_sse2,
+                   &vpx_highbd_lpf_horizontal_4_c, 12),
+        make_tuple(&vpx_highbd_lpf_vertical_4_sse2,
+                   &vpx_highbd_lpf_vertical_4_c, 12),
+        make_tuple(&vpx_highbd_lpf_horizontal_8_sse2,
+                   &vpx_highbd_lpf_horizontal_8_c, 12),
+        make_tuple(&vpx_highbd_lpf_horizontal_edge_8_sse2,
+                   &vpx_highbd_lpf_horizontal_edge_8_c, 12),
+        make_tuple(&vpx_highbd_lpf_horizontal_edge_16_sse2,
+                   &vpx_highbd_lpf_horizontal_edge_16_c, 12),
+        make_tuple(&vpx_highbd_lpf_vertical_8_sse2,
+                   &vpx_highbd_lpf_vertical_8_c, 12),
+        make_tuple(&vpx_highbd_lpf_vertical_16_sse2,
+                   &vpx_highbd_lpf_vertical_16_c, 12),
+        make_tuple(&vpx_highbd_lpf_vertical_16_dual_sse2,
+                   &vpx_highbd_lpf_vertical_16_dual_c, 8),
+        make_tuple(&vpx_highbd_lpf_vertical_16_dual_sse2,
+                   &vpx_highbd_lpf_vertical_16_dual_c, 10),
+        make_tuple(&vpx_highbd_lpf_vertical_16_dual_sse2,
+                   &vpx_highbd_lpf_vertical_16_dual_c, 12)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Loop8Test6Param,
+    ::testing::Values(
+        make_tuple(&vpx_lpf_horizontal_4_sse2,
+                   &vpx_lpf_horizontal_4_c, 8),
+        make_tuple(&vpx_lpf_horizontal_8_sse2,
+                   &vpx_lpf_horizontal_8_c, 8),
+        make_tuple(&vpx_lpf_horizontal_edge_8_sse2,
+                   &vpx_lpf_horizontal_edge_8_c, 8),
+        make_tuple(&vpx_lpf_horizontal_edge_16_sse2,
+                   &vpx_lpf_horizontal_edge_16_c, 8),
+        make_tuple(&vpx_lpf_vertical_4_sse2,
+                   &vpx_lpf_vertical_4_c, 8),
+        make_tuple(&vpx_lpf_vertical_8_sse2,
+                   &vpx_lpf_vertical_8_c, 8),
+        make_tuple(&vpx_lpf_vertical_16_sse2,
+                   &vpx_lpf_vertical_16_c, 8),
+        make_tuple(&vpx_lpf_vertical_16_dual_sse2,
+                   &vpx_lpf_vertical_16_dual_c, 8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#endif
+
+#if HAVE_AVX2 && (!CONFIG_VP9_HIGHBITDEPTH)
+INSTANTIATE_TEST_CASE_P(
+    AVX2, Loop8Test6Param,
+    ::testing::Values(
+        make_tuple(&vpx_lpf_horizontal_edge_8_avx2,
+                   &vpx_lpf_horizontal_edge_8_c, 8),
+        make_tuple(&vpx_lpf_horizontal_edge_16_avx2,
+                   &vpx_lpf_horizontal_edge_16_c, 8)));
+#endif
+
+#if HAVE_SSE2
+#if CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Loop8Test9Param,
+    ::testing::Values(
+        make_tuple(&vpx_highbd_lpf_horizontal_4_dual_sse2,
+                   &vpx_highbd_lpf_horizontal_4_dual_c, 8),
+        make_tuple(&vpx_highbd_lpf_horizontal_8_dual_sse2,
+                   &vpx_highbd_lpf_horizontal_8_dual_c, 8),
+        make_tuple(&vpx_highbd_lpf_vertical_4_dual_sse2,
+                   &vpx_highbd_lpf_vertical_4_dual_c, 8),
+        make_tuple(&vpx_highbd_lpf_vertical_8_dual_sse2,
+                   &vpx_highbd_lpf_vertical_8_dual_c, 8),
+        make_tuple(&vpx_highbd_lpf_horizontal_4_dual_sse2,
+                   &vpx_highbd_lpf_horizontal_4_dual_c, 10),
+        make_tuple(&vpx_highbd_lpf_horizontal_8_dual_sse2,
+                   &vpx_highbd_lpf_horizontal_8_dual_c, 10),
+        make_tuple(&vpx_highbd_lpf_vertical_4_dual_sse2,
+                   &vpx_highbd_lpf_vertical_4_dual_c, 10),
+        make_tuple(&vpx_highbd_lpf_vertical_8_dual_sse2,
+                   &vpx_highbd_lpf_vertical_8_dual_c, 10),
+        make_tuple(&vpx_highbd_lpf_horizontal_4_dual_sse2,
+                   &vpx_highbd_lpf_horizontal_4_dual_c, 12),
+        make_tuple(&vpx_highbd_lpf_horizontal_8_dual_sse2,
+                   &vpx_highbd_lpf_horizontal_8_dual_c, 12),
+        make_tuple(&vpx_highbd_lpf_vertical_4_dual_sse2,
+                   &vpx_highbd_lpf_vertical_4_dual_c, 12),
+        make_tuple(&vpx_highbd_lpf_vertical_8_dual_sse2,
+                   &vpx_highbd_lpf_vertical_8_dual_c, 12)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Loop8Test9Param,
+    ::testing::Values(
+        make_tuple(&vpx_lpf_horizontal_4_dual_sse2,
+                   &vpx_lpf_horizontal_4_dual_c, 8),
+        make_tuple(&vpx_lpf_horizontal_8_dual_sse2,
+                   &vpx_lpf_horizontal_8_dual_c, 8),
+        make_tuple(&vpx_lpf_vertical_4_dual_sse2,
+                   &vpx_lpf_vertical_4_dual_c, 8),
+        make_tuple(&vpx_lpf_vertical_8_dual_sse2,
+                   &vpx_lpf_vertical_8_dual_c, 8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#endif
+
+#if HAVE_NEON
+#if CONFIG_VP9_HIGHBITDEPTH
+// No neon high bitdepth functions.
+#else
+INSTANTIATE_TEST_CASE_P(
+    NEON, Loop8Test6Param,
+    ::testing::Values(
+#if HAVE_NEON_ASM
+// Using #if inside the macro is unsupported on MSVS but the tests are not
+// currently built for MSVS with ARM and NEON.
+        make_tuple(&vpx_lpf_horizontal_edge_8_neon,
+                   &vpx_lpf_horizontal_edge_8_c, 8),
+        make_tuple(&vpx_lpf_horizontal_edge_16_neon,
+                   &vpx_lpf_horizontal_edge_16_c, 8),
+        make_tuple(&vpx_lpf_vertical_16_neon,
+                   &vpx_lpf_vertical_16_c, 8),
+        make_tuple(&vpx_lpf_vertical_16_dual_neon,
+                   &vpx_lpf_vertical_16_dual_c, 8),
+#endif  // HAVE_NEON_ASM
+        make_tuple(&vpx_lpf_horizontal_8_neon,
+                   &vpx_lpf_horizontal_8_c, 8),
+        make_tuple(&vpx_lpf_vertical_8_neon,
+                   &vpx_lpf_vertical_8_c, 8),
+        make_tuple(&vpx_lpf_horizontal_4_neon,
+                   &vpx_lpf_horizontal_4_c, 8),
+        make_tuple(&vpx_lpf_vertical_4_neon,
+                   &vpx_lpf_vertical_4_c, 8)));
+INSTANTIATE_TEST_CASE_P(
+    NEON, Loop8Test9Param,
+    ::testing::Values(
+#if HAVE_NEON_ASM
+        make_tuple(&vpx_lpf_horizontal_8_dual_neon,
+                   &vpx_lpf_horizontal_8_dual_c, 8),
+        make_tuple(&vpx_lpf_vertical_8_dual_neon,
+                   &vpx_lpf_vertical_8_dual_c, 8),
+#endif  // HAVE_NEON_ASM
+        make_tuple(&vpx_lpf_horizontal_4_dual_neon,
+                   &vpx_lpf_horizontal_4_dual_c, 8),
+        make_tuple(&vpx_lpf_vertical_4_dual_neon,
+                   &vpx_lpf_vertical_4_dual_c, 8)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#endif  // HAVE_NEON
+
+#if HAVE_DSPR2 && !CONFIG_VP9_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    DSPR2, Loop8Test6Param,
+    ::testing::Values(
+        make_tuple(&vpx_lpf_horizontal_4_dspr2,
+                   &vpx_lpf_horizontal_4_c, 8),
+        make_tuple(&vpx_lpf_horizontal_8_dspr2,
+                   &vpx_lpf_horizontal_8_c, 8),
+        make_tuple(&vpx_lpf_horizontal_edge_8,
+                   &vpx_lpf_horizontal_edge_8, 8),
+        make_tuple(&vpx_lpf_horizontal_edge_16,
+                   &vpx_lpf_horizontal_edge_16, 8),
+        make_tuple(&vpx_lpf_vertical_4_dspr2,
+                   &vpx_lpf_vertical_4_c, 8),
+        make_tuple(&vpx_lpf_vertical_8_dspr2,
+                   &vpx_lpf_vertical_8_c, 8),
+        make_tuple(&vpx_lpf_vertical_16_dspr2,
+                   &vpx_lpf_vertical_16_c, 8),
+        make_tuple(&vpx_lpf_vertical_16_dual_dspr2,
+                   &vpx_lpf_vertical_16_dual_c, 8)));
+
+INSTANTIATE_TEST_CASE_P(
+    DSPR2, Loop8Test9Param,
+    ::testing::Values(
+        make_tuple(&vpx_lpf_horizontal_4_dual_dspr2,
+                   &vpx_lpf_horizontal_4_dual_c, 8),
+        make_tuple(&vpx_lpf_horizontal_8_dual_dspr2,
+                   &vpx_lpf_horizontal_8_dual_c, 8),
+        make_tuple(&vpx_lpf_vertical_4_dual_dspr2,
+                   &vpx_lpf_vertical_4_dual_c, 8),
+        make_tuple(&vpx_lpf_vertical_8_dual_dspr2,
+                   &vpx_lpf_vertical_8_dual_c, 8)));
+#endif  // HAVE_DSPR2 && !CONFIG_VP9_HIGHBITDEPTH
+
+#if HAVE_MSA && (!CONFIG_VP9_HIGHBITDEPTH)
+INSTANTIATE_TEST_CASE_P(
+    MSA, Loop8Test6Param,
+    ::testing::Values(
+        make_tuple(&vpx_lpf_horizontal_4_msa,
+                   &vpx_lpf_horizontal_4_c, 8),
+        make_tuple(&vpx_lpf_horizontal_8_msa,
+                   &vpx_lpf_horizontal_8_c, 8),
+        make_tuple(&vpx_lpf_horizontal_edge_8_msa,
+                   &vpx_lpf_horizontal_edge_8_c, 8),
+        make_tuple(&vpx_lpf_horizontal_edge_16_msa,
+                   &vpx_lpf_horizontal_edge_16_c, 8),
+        make_tuple(&vpx_lpf_vertical_4_msa,
+                   &vpx_lpf_vertical_4_c, 8),
+        make_tuple(&vpx_lpf_vertical_8_msa,
+                   &vpx_lpf_vertical_8_c, 8),
+        make_tuple(&vpx_lpf_vertical_16_msa,
+                   &vpx_lpf_vertical_16_c, 8)));
+
+INSTANTIATE_TEST_CASE_P(
+    MSA, Loop8Test9Param,
+    ::testing::Values(
+        make_tuple(&vpx_lpf_horizontal_4_dual_msa,
+                   &vpx_lpf_horizontal_4_dual_c, 8),
+        make_tuple(&vpx_lpf_horizontal_8_dual_msa,
+                   &vpx_lpf_horizontal_8_dual_c, 8),
+        make_tuple(&vpx_lpf_vertical_4_dual_msa,
+                   &vpx_lpf_vertical_4_dual_c, 8),
+        make_tuple(&vpx_lpf_vertical_8_dual_msa,
+                   &vpx_lpf_vertical_8_dual_c, 8)));
+#endif  // HAVE_MSA && (!CONFIG_VP9_HIGHBITDEPTH)
+
+}  // namespace
diff --git a/libvpx/libvpx/test/md5_helper.h b/libvpx/libvpx/test/md5_helper.h
new file mode 100644
index 0000000..742cf0b
--- /dev/null
+++ b/libvpx/libvpx/test/md5_helper.h
@@ -0,0 +1,74 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef TEST_MD5_HELPER_H_
+#define TEST_MD5_HELPER_H_
+
+#include "./md5_utils.h"
+#include "vpx/vpx_decoder.h"
+
+namespace libvpx_test {
+class MD5 {
+ public:
+  MD5() {
+    MD5Init(&md5_);
+  }
+
+  void Add(const vpx_image_t *img) {
+    for (int plane = 0; plane < 3; ++plane) {
+      const uint8_t *buf = img->planes[plane];
+      // Calculate the width and height to do the md5 check. For the chroma
+      // plane, we never want to round down and thus skip a pixel so if
+      // we are shifting by 1 (chroma_shift) we add 1 before doing the shift.
+      // This works only for chroma_shift of 0 and 1.
+      const int bytes_per_sample =
+          (img->fmt & VPX_IMG_FMT_HIGHBITDEPTH) ? 2 : 1;
+      const int h = plane ? (img->d_h + img->y_chroma_shift) >>
+                    img->y_chroma_shift : img->d_h;
+      const int w = (plane ? (img->d_w + img->x_chroma_shift) >>
+                     img->x_chroma_shift : img->d_w) * bytes_per_sample;
+
+      for (int y = 0; y < h; ++y) {
+        MD5Update(&md5_, buf, w);
+        buf += img->stride[plane];
+      }
+    }
+  }
+
+  void Add(const uint8_t *data, size_t size) {
+    MD5Update(&md5_, data, static_cast<uint32_t>(size));
+  }
+
+  const char *Get(void) {
+    static const char hex[16] = {
+      '0', '1', '2', '3', '4', '5', '6', '7',
+      '8', '9', 'a', 'b', 'c', 'd', 'e', 'f',
+    };
+    uint8_t tmp[16];
+    MD5Context ctx_tmp = md5_;
+
+    MD5Final(tmp, &ctx_tmp);
+    for (int i = 0; i < 16; i++) {
+      res_[i * 2 + 0]  = hex[tmp[i] >> 4];
+      res_[i * 2 + 1]  = hex[tmp[i] & 0xf];
+    }
+    res_[32] = 0;
+
+    return res_;
+  }
+
+ protected:
+  char res_[33];
+  MD5Context md5_;
+};
+
+}  // namespace libvpx_test
+
+#endif  // TEST_MD5_HELPER_H_
diff --git a/libvpx/libvpx/test/minmax_test.cc b/libvpx/libvpx/test/minmax_test.cc
new file mode 100644
index 0000000..dbe4342
--- /dev/null
+++ b/libvpx/libvpx/test/minmax_test.cc
@@ -0,0 +1,132 @@
+/*
+ *  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+#include "test/acm_random.h"
+#include "test/register_state_check.h"
+
+namespace {
+
+using ::libvpx_test::ACMRandom;
+
+typedef void (*MinMaxFunc)(const uint8_t *a, int a_stride,
+                           const uint8_t *b, int b_stride,
+                           int *min, int *max);
+
+class MinMaxTest : public ::testing::TestWithParam<MinMaxFunc> {
+ public:
+  virtual void SetUp() {
+    mm_func_ = GetParam();
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+  }
+
+ protected:
+  MinMaxFunc mm_func_;
+  ACMRandom rnd_;
+};
+
+void reference_minmax(const uint8_t *a, int a_stride,
+                      const uint8_t *b, int b_stride,
+                      int *min_ret, int *max_ret) {
+  int min = 255;
+  int max = 0;
+  for (int i = 0; i < 8; i++) {
+    for (int j = 0; j < 8; j++) {
+      const int diff = abs(a[i * a_stride + j] - b[i * b_stride + j]);
+      if (min > diff) min = diff;
+      if (max < diff) max = diff;
+    }
+  }
+
+  *min_ret = min;
+  *max_ret = max;
+}
+
+TEST_P(MinMaxTest, MinValue) {
+  for (int i = 0; i < 64; i++) {
+    uint8_t a[64], b[64];
+    memset(a, 0, sizeof(a));
+    memset(b, 255, sizeof(b));
+    b[i] = i;  // Set a minimum difference of i.
+
+    int min, max;
+    ASM_REGISTER_STATE_CHECK(mm_func_(a, 8, b, 8, &min, &max));
+    EXPECT_EQ(255, max);
+    EXPECT_EQ(i, min);
+  }
+}
+
+TEST_P(MinMaxTest, MaxValue) {
+  for (int i = 0; i < 64; i++) {
+    uint8_t a[64], b[64];
+    memset(a, 0, sizeof(a));
+    memset(b, 0, sizeof(b));
+    b[i] = i;  // Set a maximum difference of i.
+
+    int min, max;
+    ASM_REGISTER_STATE_CHECK(mm_func_(a, 8, b, 8, &min, &max));
+    EXPECT_EQ(i, max);
+    EXPECT_EQ(0, min);
+  }
+}
+
+TEST_P(MinMaxTest, CompareReference) {
+  uint8_t a[64], b[64];
+  for (int j = 0; j < 64; j++) {
+    a[j] = rnd_.Rand8();
+    b[j] = rnd_.Rand8();
+  }
+
+  int min_ref, max_ref, min, max;
+  reference_minmax(a, 8, b, 8, &min_ref, &max_ref);
+  ASM_REGISTER_STATE_CHECK(mm_func_(a, 8, b, 8, &min, &max));
+  EXPECT_EQ(max_ref, max);
+  EXPECT_EQ(min_ref, min);
+}
+
+TEST_P(MinMaxTest, CompareReferenceAndVaryStride) {
+  uint8_t a[8 * 64], b[8 * 64];
+  for (int i = 0; i < 8 * 64; i++) {
+    a[i] = rnd_.Rand8();
+    b[i] = rnd_.Rand8();
+  }
+  for (int a_stride = 8; a_stride <= 64; a_stride += 8) {
+    for (int b_stride = 8; b_stride <= 64; b_stride += 8) {
+      int min_ref, max_ref, min, max;
+      reference_minmax(a, a_stride, b, b_stride, &min_ref, &max_ref);
+      ASM_REGISTER_STATE_CHECK(mm_func_(a, a_stride, b, b_stride, &min, &max));
+      EXPECT_EQ(max_ref, max) << "when a_stride = " << a_stride
+                              << " and b_stride = " << b_stride;;
+      EXPECT_EQ(min_ref, min) << "when a_stride = " << a_stride
+                              << " and b_stride = " << b_stride;;
+    }
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(C, MinMaxTest, ::testing::Values(&vpx_minmax_8x8_c));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, MinMaxTest,
+                        ::testing::Values(&vpx_minmax_8x8_sse2));
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, MinMaxTest,
+                        ::testing::Values(&vpx_minmax_8x8_neon));
+#endif
+
+}  // namespace
diff --git a/libvpx/libvpx/test/partial_idct_test.cc b/libvpx/libvpx/test/partial_idct_test.cc
new file mode 100644
index 0000000..6c82412
--- /dev/null
+++ b/libvpx/libvpx/test/partial_idct_test.cc
@@ -0,0 +1,343 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_scan.h"
+#include "vpx/vpx_integer.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+typedef void (*FwdTxfmFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*InvTxfmFunc)(const tran_low_t *in, uint8_t *out, int stride);
+typedef std::tr1::tuple<FwdTxfmFunc,
+                        InvTxfmFunc,
+                        InvTxfmFunc,
+                        TX_SIZE, int> PartialInvTxfmParam;
+const int kMaxNumCoeffs = 1024;
+class PartialIDctTest : public ::testing::TestWithParam<PartialInvTxfmParam> {
+ public:
+  virtual ~PartialIDctTest() {}
+  virtual void SetUp() {
+    ftxfm_ = GET_PARAM(0);
+    full_itxfm_ = GET_PARAM(1);
+    partial_itxfm_ = GET_PARAM(2);
+    tx_size_  = GET_PARAM(3);
+    last_nonzero_ = GET_PARAM(4);
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  int last_nonzero_;
+  TX_SIZE tx_size_;
+  FwdTxfmFunc ftxfm_;
+  InvTxfmFunc full_itxfm_;
+  InvTxfmFunc partial_itxfm_;
+};
+
+TEST_P(PartialIDctTest, RunQuantCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int size;
+  switch (tx_size_) {
+    case TX_4X4:
+      size = 4;
+      break;
+    case TX_8X8:
+      size = 8;
+      break;
+    case TX_16X16:
+      size = 16;
+      break;
+    case TX_32X32:
+      size = 32;
+      break;
+    default:
+      FAIL() << "Wrong Size!";
+      break;
+  }
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block2[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst2[kMaxNumCoeffs]);
+
+  const int count_test_block = 1000;
+  const int block_size = size * size;
+
+  DECLARE_ALIGNED(16, int16_t, input_extreme_block[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kMaxNumCoeffs]);
+
+  int max_error = 0;
+  for (int i = 0; i < count_test_block; ++i) {
+    // clear out destination buffer
+    memset(dst1, 0, sizeof(*dst1) * block_size);
+    memset(dst2, 0, sizeof(*dst2) * block_size);
+    memset(test_coef_block1, 0, sizeof(*test_coef_block1) * block_size);
+    memset(test_coef_block2, 0, sizeof(*test_coef_block2) * block_size);
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-255, 255].
+      if (i == 0) {
+        for (int j = 0; j < block_size; ++j)
+          input_extreme_block[j] = 255;
+      } else if (i == 1) {
+        for (int j = 0; j < block_size; ++j)
+          input_extreme_block[j] = -255;
+      } else {
+        for (int j = 0; j < block_size; ++j) {
+          input_extreme_block[j] = rnd.Rand8() % 2 ? 255 : -255;
+        }
+      }
+
+      ftxfm_(input_extreme_block, output_ref_block, size);
+
+      // quantization with maximum allowed step sizes
+      test_coef_block1[0] = (output_ref_block[0] / 1336) * 1336;
+      for (int j = 1; j < last_nonzero_; ++j)
+        test_coef_block1[vp9_default_scan_orders[tx_size_].scan[j]]
+                         = (output_ref_block[j] / 1828) * 1828;
+    }
+
+    ASM_REGISTER_STATE_CHECK(full_itxfm_(test_coef_block1, dst1, size));
+    ASM_REGISTER_STATE_CHECK(partial_itxfm_(test_coef_block1, dst2, size));
+
+    for (int j = 0; j < block_size; ++j) {
+      const int diff = dst1[j] - dst2[j];
+      const int error = diff * diff;
+      if (max_error < error)
+        max_error = error;
+    }
+  }
+
+  EXPECT_EQ(0, max_error)
+      << "Error: partial inverse transform produces different results";
+}
+
+TEST_P(PartialIDctTest, ResultsMatch) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int size;
+  switch (tx_size_) {
+    case TX_4X4:
+      size = 4;
+      break;
+    case TX_8X8:
+      size = 8;
+      break;
+    case TX_16X16:
+      size = 16;
+      break;
+    case TX_32X32:
+      size = 32;
+      break;
+    default:
+      FAIL() << "Wrong Size!";
+      break;
+  }
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block2[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst2[kMaxNumCoeffs]);
+  const int count_test_block = 1000;
+  const int max_coeff = 32766 / 4;
+  const int block_size = size * size;
+  int max_error = 0;
+  for (int i = 0; i < count_test_block; ++i) {
+    // clear out destination buffer
+    memset(dst1, 0, sizeof(*dst1) * block_size);
+    memset(dst2, 0, sizeof(*dst2) * block_size);
+    memset(test_coef_block1, 0, sizeof(*test_coef_block1) * block_size);
+    memset(test_coef_block2, 0, sizeof(*test_coef_block2) * block_size);
+    int max_energy_leftover = max_coeff * max_coeff;
+    for (int j = 0; j < last_nonzero_; ++j) {
+      int16_t coef = static_cast<int16_t>(sqrt(1.0 * max_energy_leftover) *
+                                          (rnd.Rand16() - 32768) / 65536);
+      max_energy_leftover -= coef * coef;
+      if (max_energy_leftover < 0) {
+        max_energy_leftover = 0;
+        coef = 0;
+      }
+      test_coef_block1[vp9_default_scan_orders[tx_size_].scan[j]] = coef;
+    }
+
+    memcpy(test_coef_block2, test_coef_block1,
+           sizeof(*test_coef_block2) * block_size);
+
+    ASM_REGISTER_STATE_CHECK(full_itxfm_(test_coef_block1, dst1, size));
+    ASM_REGISTER_STATE_CHECK(partial_itxfm_(test_coef_block2, dst2, size));
+
+    for (int j = 0; j < block_size; ++j) {
+      const int diff = dst1[j] - dst2[j];
+      const int error = diff * diff;
+      if (max_error < error)
+        max_error = error;
+    }
+  }
+
+  EXPECT_EQ(0, max_error)
+      << "Error: partial inverse transform produces different results";
+}
+using std::tr1::make_tuple;
+
+INSTANTIATE_TEST_CASE_P(
+    C, PartialIDctTest,
+    ::testing::Values(
+        make_tuple(&vpx_fdct32x32_c,
+                   &vpx_idct32x32_1024_add_c,
+                   &vpx_idct32x32_34_add_c,
+                   TX_32X32, 34),
+        make_tuple(&vpx_fdct32x32_c,
+                   &vpx_idct32x32_1024_add_c,
+                   &vpx_idct32x32_1_add_c,
+                   TX_32X32, 1),
+        make_tuple(&vpx_fdct16x16_c,
+                   &vpx_idct16x16_256_add_c,
+                   &vpx_idct16x16_10_add_c,
+                   TX_16X16, 10),
+        make_tuple(&vpx_fdct16x16_c,
+                   &vpx_idct16x16_256_add_c,
+                   &vpx_idct16x16_1_add_c,
+                   TX_16X16, 1),
+        make_tuple(&vpx_fdct8x8_c,
+                   &vpx_idct8x8_64_add_c,
+                   &vpx_idct8x8_12_add_c,
+                   TX_8X8, 12),
+        make_tuple(&vpx_fdct8x8_c,
+                   &vpx_idct8x8_64_add_c,
+                   &vpx_idct8x8_1_add_c,
+                   TX_8X8, 1),
+        make_tuple(&vpx_fdct4x4_c,
+                   &vpx_idct4x4_16_add_c,
+                   &vpx_idct4x4_1_add_c,
+                   TX_4X4, 1)));
+
+#if HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    NEON, PartialIDctTest,
+    ::testing::Values(
+        make_tuple(&vpx_fdct32x32_c,
+                   &vpx_idct32x32_1024_add_c,
+                   &vpx_idct32x32_1_add_neon,
+                   TX_32X32, 1),
+        make_tuple(&vpx_fdct16x16_c,
+                   &vpx_idct16x16_256_add_c,
+                   &vpx_idct16x16_10_add_neon,
+                   TX_16X16, 10),
+        make_tuple(&vpx_fdct16x16_c,
+                   &vpx_idct16x16_256_add_c,
+                   &vpx_idct16x16_1_add_neon,
+                   TX_16X16, 1),
+        make_tuple(&vpx_fdct8x8_c,
+                   &vpx_idct8x8_64_add_c,
+                   &vpx_idct8x8_12_add_neon,
+                   TX_8X8, 12),
+        make_tuple(&vpx_fdct8x8_c,
+                   &vpx_idct8x8_64_add_c,
+                   &vpx_idct8x8_1_add_neon,
+                   TX_8X8, 1),
+        make_tuple(&vpx_fdct4x4_c,
+                   &vpx_idct4x4_16_add_c,
+                   &vpx_idct4x4_1_add_neon,
+                   TX_4X4, 1)));
+#endif  // HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+#if HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSE2, PartialIDctTest,
+    ::testing::Values(
+        make_tuple(&vpx_fdct32x32_c,
+                   &vpx_idct32x32_1024_add_c,
+                   &vpx_idct32x32_34_add_sse2,
+                   TX_32X32, 34),
+        make_tuple(&vpx_fdct32x32_c,
+                   &vpx_idct32x32_1024_add_c,
+                   &vpx_idct32x32_1_add_sse2,
+                   TX_32X32, 1),
+        make_tuple(&vpx_fdct16x16_c,
+                   &vpx_idct16x16_256_add_c,
+                   &vpx_idct16x16_10_add_sse2,
+                   TX_16X16, 10),
+        make_tuple(&vpx_fdct16x16_c,
+                   &vpx_idct16x16_256_add_c,
+                   &vpx_idct16x16_1_add_sse2,
+                   TX_16X16, 1),
+        make_tuple(&vpx_fdct8x8_c,
+                   &vpx_idct8x8_64_add_c,
+                   &vpx_idct8x8_12_add_sse2,
+                   TX_8X8, 12),
+        make_tuple(&vpx_fdct8x8_c,
+                   &vpx_idct8x8_64_add_c,
+                   &vpx_idct8x8_1_add_sse2,
+                   TX_8X8, 1),
+        make_tuple(&vpx_fdct4x4_c,
+                   &vpx_idct4x4_16_add_c,
+                   &vpx_idct4x4_1_add_sse2,
+                   TX_4X4, 1)));
+#endif
+
+#if HAVE_SSSE3 && CONFIG_USE_X86INC && ARCH_X86_64 && \
+    !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    SSSE3_64, PartialIDctTest,
+    ::testing::Values(
+        make_tuple(&vpx_fdct8x8_c,
+                   &vpx_idct8x8_64_add_c,
+                   &vpx_idct8x8_12_add_ssse3,
+                   TX_8X8, 12)));
+#endif
+
+#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+INSTANTIATE_TEST_CASE_P(
+    MSA, PartialIDctTest,
+    ::testing::Values(
+        make_tuple(&vpx_fdct32x32_c,
+                   &vpx_idct32x32_1024_add_c,
+                   &vpx_idct32x32_34_add_msa,
+                   TX_32X32, 34),
+        make_tuple(&vpx_fdct32x32_c,
+                   &vpx_idct32x32_1024_add_c,
+                   &vpx_idct32x32_1_add_msa,
+                   TX_32X32, 1),
+        make_tuple(&vpx_fdct16x16_c,
+                   &vpx_idct16x16_256_add_c,
+                   &vpx_idct16x16_10_add_msa,
+                   TX_16X16, 10),
+        make_tuple(&vpx_fdct16x16_c,
+                   &vpx_idct16x16_256_add_c,
+                   &vpx_idct16x16_1_add_msa,
+                   TX_16X16, 1),
+        make_tuple(&vpx_fdct8x8_c,
+                   &vpx_idct8x8_64_add_c,
+                   &vpx_idct8x8_12_add_msa,
+                   TX_8X8, 10),
+        make_tuple(&vpx_fdct8x8_c,
+                   &vpx_idct8x8_64_add_c,
+                   &vpx_idct8x8_1_add_msa,
+                   TX_8X8, 1),
+        make_tuple(&vpx_fdct4x4_c,
+                   &vpx_idct4x4_16_add_c,
+                   &vpx_idct4x4_1_add_msa,
+                   TX_4X4, 1)));
+#endif  // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE
+
+}  // namespace
diff --git a/libvpx/libvpx/test/postproc.sh b/libvpx/libvpx/test/postproc.sh
new file mode 100755
index 0000000..939a3e7
--- /dev/null
+++ b/libvpx/libvpx/test/postproc.sh
@@ -0,0 +1,63 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx postproc example code. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to postproc_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+#   $VP8_IVF_FILE and $VP9_IVF_FILE are required.
+postproc_verify_environment() {
+  if [ ! -e "${VP8_IVF_FILE}" ] || [ ! -e "${VP9_IVF_FILE}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs postproc using $1 as input file. $2 is the codec name, and is used
+# solely to name the output file.
+postproc() {
+  local decoder="${LIBVPX_BIN_PATH}/postproc${VPX_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/postproc_${codec}.raw"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+postproc_vp8() {
+  if [ "$(vp8_decode_available)" = "yes" ]; then
+    postproc "${VP8_IVF_FILE}" vp8 || return 1
+  fi
+}
+
+postproc_vp9() {
+  if [ "$(vpx_config_option_enabled CONFIG_VP9_POSTPROC)" = "yes" ]; then
+    if [ "$(vp9_decode_available)" = "yes" ]; then
+      postproc "${VP9_IVF_FILE}" vp9 || return 1
+    fi
+  fi
+}
+
+postproc_tests="postproc_vp8
+                postproc_vp9"
+
+run_tests postproc_verify_environment "${postproc_tests}"
diff --git a/libvpx/libvpx/test/pp_filter_test.cc b/libvpx/libvpx/test/pp_filter_test.cc
new file mode 100644
index 0000000..e4688dd
--- /dev/null
+++ b/libvpx/libvpx/test/pp_filter_test.cc
@@ -0,0 +1,118 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "./vp8_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+
+typedef void (*PostProcFunc)(unsigned char *src_ptr,
+                             unsigned char *dst_ptr,
+                             int src_pixels_per_line,
+                             int dst_pixels_per_line,
+                             int cols,
+                             unsigned char *flimit,
+                             int size);
+
+namespace {
+
+class VP8PostProcessingFilterTest
+    : public ::testing::TestWithParam<PostProcFunc> {
+ public:
+  virtual void TearDown() {
+    libvpx_test::ClearSystemState();
+  }
+};
+
+// Test routine for the VP8 post-processing function
+// vp8_post_proc_down_and_across_mb_row_c.
+
+TEST_P(VP8PostProcessingFilterTest, FilterOutputCheck) {
+  // Size of the underlying data block that will be filtered.
+  const int block_width  = 16;
+  const int block_height = 16;
+
+  // 5-tap filter needs 2 padding rows above and below the block in the input.
+  const int input_width = block_width;
+  const int input_height = block_height + 4;
+  const int input_stride = input_width;
+  const int input_size = input_width * input_height;
+
+  // Filter extends output block by 8 samples at left and right edges.
+  const int output_width = block_width + 16;
+  const int output_height = block_height;
+  const int output_stride = output_width;
+  const int output_size = output_width * output_height;
+
+  uint8_t *const src_image =
+      reinterpret_cast<uint8_t*>(vpx_calloc(input_size, 1));
+  uint8_t *const dst_image =
+      reinterpret_cast<uint8_t*>(vpx_calloc(output_size, 1));
+
+  // Pointers to top-left pixel of block in the input and output images.
+  uint8_t *const src_image_ptr = src_image + (input_stride << 1);
+  uint8_t *const dst_image_ptr = dst_image + 8;
+  uint8_t *const flimits =
+      reinterpret_cast<uint8_t *>(vpx_memalign(16, block_width));
+  (void)memset(flimits, 255, block_width);
+
+  // Initialize pixels in the input:
+  //   block pixels to value 1,
+  //   border pixels to value 10.
+  (void)memset(src_image, 10, input_size);
+  uint8_t *pixel_ptr = src_image_ptr;
+  for (int i = 0; i < block_height; ++i) {
+    for (int j = 0; j < block_width; ++j) {
+      pixel_ptr[j] = 1;
+    }
+    pixel_ptr += input_stride;
+  }
+
+  // Initialize pixels in the output to 99.
+  (void)memset(dst_image, 99, output_size);
+
+  ASM_REGISTER_STATE_CHECK(
+      GetParam()(src_image_ptr, dst_image_ptr, input_stride,
+                 output_stride, block_width, flimits, 16));
+
+  static const uint8_t expected_data[block_height] = {
+    4, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 3, 4
+  };
+
+  pixel_ptr = dst_image_ptr;
+  for (int i = 0; i < block_height; ++i) {
+    for (int j = 0; j < block_width; ++j) {
+      EXPECT_EQ(expected_data[i], pixel_ptr[j])
+          << "VP8PostProcessingFilterTest failed with invalid filter output";
+    }
+    pixel_ptr += output_stride;
+  }
+
+  vpx_free(src_image);
+  vpx_free(dst_image);
+  vpx_free(flimits);
+};
+
+INSTANTIATE_TEST_CASE_P(C, VP8PostProcessingFilterTest,
+    ::testing::Values(vp8_post_proc_down_and_across_mb_row_c));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, VP8PostProcessingFilterTest,
+    ::testing::Values(vp8_post_proc_down_and_across_mb_row_sse2));
+#endif
+
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(MSA, VP8PostProcessingFilterTest,
+    ::testing::Values(vp8_post_proc_down_and_across_mb_row_msa));
+#endif
+
+}  // namespace
diff --git a/libvpx/libvpx/test/quantize_test.cc b/libvpx/libvpx/test/quantize_test.cc
new file mode 100644
index 0000000..69da899
--- /dev/null
+++ b/libvpx/libvpx/test/quantize_test.cc
@@ -0,0 +1,203 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vp8_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp8/common/blockd.h"
+#include "vp8/common/onyx.h"
+#include "vp8/encoder/block.h"
+#include "vp8/encoder/onyx_int.h"
+#include "vp8/encoder/quantize.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+
+const int kNumBlocks = 25;
+const int kNumBlockEntries = 16;
+
+typedef void (*VP8Quantize)(BLOCK *b, BLOCKD *d);
+
+typedef std::tr1::tuple<VP8Quantize, VP8Quantize> VP8QuantizeParam;
+
+using libvpx_test::ACMRandom;
+using std::tr1::make_tuple;
+
+// Create and populate a VP8_COMP instance which has a complete set of
+// quantization inputs as well as a second MACROBLOCKD for output.
+class QuantizeTestBase {
+ public:
+  virtual ~QuantizeTestBase() {
+    vp8_remove_compressor(&vp8_comp_);
+    vp8_comp_ = NULL;
+    vpx_free(macroblockd_dst_);
+    macroblockd_dst_ = NULL;
+    libvpx_test::ClearSystemState();
+  }
+
+ protected:
+  void SetupCompressor() {
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+
+    // The full configuration is necessary to generate the quantization tables.
+    VP8_CONFIG vp8_config;
+    memset(&vp8_config, 0, sizeof(vp8_config));
+
+    vp8_comp_ = vp8_create_compressor(&vp8_config);
+
+    // Set the tables based on a quantizer of 0.
+    vp8_set_quantizer(vp8_comp_, 0);
+
+    // Set up all the block/blockd pointers for the mb in vp8_comp_.
+    vp8cx_frame_init_quantizer(vp8_comp_);
+
+    // Copy macroblockd from the reference to get pre-set-up dequant values.
+    macroblockd_dst_ = reinterpret_cast<MACROBLOCKD *>(
+        vpx_memalign(32, sizeof(*macroblockd_dst_)));
+    memcpy(macroblockd_dst_, &vp8_comp_->mb.e_mbd, sizeof(*macroblockd_dst_));
+    // Fix block pointers - currently they point to the blocks in the reference
+    // structure.
+    vp8_setup_block_dptrs(macroblockd_dst_);
+  }
+
+  void UpdateQuantizer(int q) {
+    vp8_set_quantizer(vp8_comp_, q);
+
+    memcpy(macroblockd_dst_, &vp8_comp_->mb.e_mbd, sizeof(*macroblockd_dst_));
+    vp8_setup_block_dptrs(macroblockd_dst_);
+  }
+
+  void FillCoeffConstant(int16_t c) {
+    for (int i = 0; i < kNumBlocks * kNumBlockEntries; ++i) {
+      vp8_comp_->mb.coeff[i] = c;
+    }
+  }
+
+  void FillCoeffRandom() {
+    for (int i = 0; i < kNumBlocks * kNumBlockEntries; ++i) {
+      vp8_comp_->mb.coeff[i] = rnd_.Rand8();
+    }
+  }
+
+  void CheckOutput() {
+    EXPECT_EQ(0, memcmp(vp8_comp_->mb.e_mbd.qcoeff, macroblockd_dst_->qcoeff,
+                        sizeof(*macroblockd_dst_->qcoeff) * kNumBlocks *
+                            kNumBlockEntries))
+        << "qcoeff mismatch";
+    EXPECT_EQ(0, memcmp(vp8_comp_->mb.e_mbd.dqcoeff, macroblockd_dst_->dqcoeff,
+                        sizeof(*macroblockd_dst_->dqcoeff) * kNumBlocks *
+                            kNumBlockEntries))
+        << "dqcoeff mismatch";
+    EXPECT_EQ(0, memcmp(vp8_comp_->mb.e_mbd.eobs, macroblockd_dst_->eobs,
+                        sizeof(*macroblockd_dst_->eobs) * kNumBlocks))
+        << "eobs mismatch";
+  }
+
+  VP8_COMP *vp8_comp_;
+  MACROBLOCKD *macroblockd_dst_;
+
+ private:
+  ACMRandom rnd_;
+};
+
+class QuantizeTest : public QuantizeTestBase,
+                     public ::testing::TestWithParam<VP8QuantizeParam> {
+ protected:
+  virtual void SetUp() {
+    SetupCompressor();
+    asm_quant_ = GET_PARAM(0);
+    c_quant_ = GET_PARAM(1);
+  }
+
+  void RunComparison() {
+    for (int i = 0; i < kNumBlocks; ++i) {
+      ASM_REGISTER_STATE_CHECK(
+          c_quant_(&vp8_comp_->mb.block[i], &vp8_comp_->mb.e_mbd.block[i]));
+      ASM_REGISTER_STATE_CHECK(
+          asm_quant_(&vp8_comp_->mb.block[i], &macroblockd_dst_->block[i]));
+    }
+
+    CheckOutput();
+  }
+
+ private:
+  VP8Quantize asm_quant_;
+  VP8Quantize c_quant_;
+};
+
+TEST_P(QuantizeTest, TestZeroInput) {
+  FillCoeffConstant(0);
+  RunComparison();
+}
+
+TEST_P(QuantizeTest, TestLargeNegativeInput) {
+  FillCoeffConstant(0);
+  // Generate a qcoeff which contains 512/-512 (0x0100/0xFE00) to catch issues
+  // like BUG=883 where the constant being compared was incorrectly initialized.
+  vp8_comp_->mb.coeff[0] = -8191;
+  RunComparison();
+}
+
+TEST_P(QuantizeTest, TestRandomInput) {
+  FillCoeffRandom();
+  RunComparison();
+}
+
+TEST_P(QuantizeTest, TestMultipleQ) {
+  for (int q = 0; q < QINDEX_RANGE; ++q) {
+    UpdateQuantizer(q);
+    FillCoeffRandom();
+    RunComparison();
+  }
+}
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, QuantizeTest,
+    ::testing::Values(
+        make_tuple(&vp8_fast_quantize_b_sse2, &vp8_fast_quantize_b_c),
+        make_tuple(&vp8_regular_quantize_b_sse2, &vp8_regular_quantize_b_c)));
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(SSSE3, QuantizeTest,
+                        ::testing::Values(make_tuple(&vp8_fast_quantize_b_ssse3,
+                                                     &vp8_fast_quantize_b_c)));
+#endif  // HAVE_SSSE3
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, QuantizeTest,
+    ::testing::Values(make_tuple(&vp8_regular_quantize_b_sse4_1,
+                                 &vp8_regular_quantize_b_c)));
+#endif  // HAVE_SSE4_1
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, QuantizeTest,
+                        ::testing::Values(make_tuple(&vp8_fast_quantize_b_neon,
+                                                     &vp8_fast_quantize_b_c)));
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(
+    MSA, QuantizeTest,
+    ::testing::Values(
+        make_tuple(&vp8_fast_quantize_b_msa, &vp8_fast_quantize_b_c),
+        make_tuple(&vp8_regular_quantize_b_msa, &vp8_regular_quantize_b_c)));
+#endif  // HAVE_MSA
+}  // namespace
diff --git a/libvpx/libvpx/test/realtime_test.cc b/libvpx/libvpx/test/realtime_test.cc
new file mode 100644
index 0000000..24749e4
--- /dev/null
+++ b/libvpx/libvpx/test/realtime_test.cc
@@ -0,0 +1,64 @@
+/*
+ *  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/video_source.h"
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+namespace {
+
+const int kVideoSourceWidth = 320;
+const int kVideoSourceHeight = 240;
+const int kFramesToEncode = 2;
+
+class RealtimeTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ protected:
+  RealtimeTest()
+      : EncoderTest(GET_PARAM(0)), frame_packets_(0) {}
+  virtual ~RealtimeTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    cfg_.g_lag_in_frames = 0;
+    SetMode(::libvpx_test::kRealTime);
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    // TODO(tomfinegan): We're changing the pass value here to make sure
+    // we get frames when real time mode is combined with |g_pass| set to
+    // VPX_RC_FIRST_PASS. This is necessary because EncoderTest::RunLoop() sets
+    // the pass value based on the mode passed into EncoderTest::SetMode(),
+    // which overrides the one specified in SetUp() above.
+    cfg_.g_pass = VPX_RC_FIRST_PASS;
+  }
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t * /*pkt*/) {
+    frame_packets_++;
+  }
+
+  int frame_packets_;
+};
+
+TEST_P(RealtimeTest, RealtimeFirstPassProducesFrames) {
+  ::libvpx_test::RandomVideoSource video;
+  video.SetSize(kVideoSourceWidth, kVideoSourceHeight);
+  video.set_limit(kFramesToEncode);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_EQ(kFramesToEncode, frame_packets_);
+}
+
+VP8_INSTANTIATE_TEST_CASE(RealtimeTest,
+                          ::testing::Values(::libvpx_test::kRealTime));
+VP9_INSTANTIATE_TEST_CASE(RealtimeTest,
+                          ::testing::Values(::libvpx_test::kRealTime));
+
+}  // namespace
diff --git a/libvpx/libvpx/test/register_state_check.h b/libvpx/libvpx/test/register_state_check.h
new file mode 100644
index 0000000..5336f2f
--- /dev/null
+++ b/libvpx/libvpx/test/register_state_check.h
@@ -0,0 +1,192 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef TEST_REGISTER_STATE_CHECK_H_
+#define TEST_REGISTER_STATE_CHECK_H_
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+// ASM_REGISTER_STATE_CHECK(asm_function)
+//   Minimally validates the environment pre & post function execution. This
+//   variant should be used with assembly functions which are not expected to
+//   fully restore the system state. See platform implementations of
+//   RegisterStateCheck for details.
+//
+// API_REGISTER_STATE_CHECK(api_function)
+//   Performs all the checks done by ASM_REGISTER_STATE_CHECK() and any
+//   additional checks to ensure the environment is in a consistent state pre &
+//   post function execution. This variant should be used with API functions.
+//   See platform implementations of RegisterStateCheckXXX for details.
+//
+
+#if defined(_WIN64)
+
+#undef NOMINMAX
+#define NOMINMAX
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#include <winnt.h>
+
+inline bool operator==(const M128A& lhs, const M128A& rhs) {
+  return (lhs.Low == rhs.Low && lhs.High == rhs.High);
+}
+
+namespace libvpx_test {
+
+// Compares the state of xmm[6-15] at construction with their state at
+// destruction. These registers should be preserved by the callee on
+// Windows x64.
+class RegisterStateCheck {
+ public:
+  RegisterStateCheck() { initialized_ = StoreRegisters(&pre_context_); }
+  ~RegisterStateCheck() { EXPECT_TRUE(Check()); }
+
+ private:
+  static bool StoreRegisters(CONTEXT* const context) {
+    const HANDLE this_thread = GetCurrentThread();
+    EXPECT_TRUE(this_thread != NULL);
+    context->ContextFlags = CONTEXT_FLOATING_POINT;
+    const bool context_saved = GetThreadContext(this_thread, context) == TRUE;
+    EXPECT_TRUE(context_saved) << "GetLastError: " << GetLastError();
+    return context_saved;
+  }
+
+  // Compares the register state. Returns true if the states match.
+  bool Check() const {
+    if (!initialized_) return false;
+    CONTEXT post_context;
+    if (!StoreRegisters(&post_context)) return false;
+
+    const M128A* xmm_pre = &pre_context_.Xmm6;
+    const M128A* xmm_post = &post_context.Xmm6;
+    for (int i = 6; i <= 15; ++i) {
+      EXPECT_EQ(*xmm_pre, *xmm_post) << "xmm" << i << " has been modified!";
+      ++xmm_pre;
+      ++xmm_post;
+    }
+    return !testing::Test::HasNonfatalFailure();
+  }
+
+  bool initialized_;
+  CONTEXT pre_context_;
+};
+
+#define ASM_REGISTER_STATE_CHECK(statement) do {  \
+  libvpx_test::RegisterStateCheck reg_check;      \
+  statement;                                      \
+} while (false)
+
+}  // namespace libvpx_test
+
+#elif defined(CONFIG_SHARED) && defined(HAVE_NEON_ASM) && defined(CONFIG_VP9) \
+      && !CONFIG_SHARED && HAVE_NEON_ASM && CONFIG_VP9
+
+extern "C" {
+// Save the d8-d15 registers into store.
+void vpx_push_neon(int64_t *store);
+}
+
+namespace libvpx_test {
+
+// Compares the state of d8-d15 at construction with their state at
+// destruction. These registers should be preserved by the callee on
+// arm platform.
+class RegisterStateCheck {
+ public:
+  RegisterStateCheck() { initialized_ = StoreRegisters(pre_store_); }
+  ~RegisterStateCheck() { EXPECT_TRUE(Check()); }
+
+ private:
+  static bool StoreRegisters(int64_t store[8]) {
+    vpx_push_neon(store);
+    return true;
+  }
+
+  // Compares the register state. Returns true if the states match.
+  bool Check() const {
+    if (!initialized_) return false;
+    int64_t post_store[8];
+    vpx_push_neon(post_store);
+    for (int i = 0; i < 8; ++i) {
+      EXPECT_EQ(pre_store_[i], post_store[i]) << "d"
+          << i + 8 << " has been modified";
+    }
+    return !testing::Test::HasNonfatalFailure();
+  }
+
+  bool initialized_;
+  int64_t pre_store_[8];
+};
+
+#define ASM_REGISTER_STATE_CHECK(statement) do {  \
+  libvpx_test::RegisterStateCheck reg_check;      \
+  statement;                                      \
+} while (false)
+
+}  // namespace libvpx_test
+
+#else
+
+namespace libvpx_test {
+
+class RegisterStateCheck {};
+#define ASM_REGISTER_STATE_CHECK(statement) statement
+
+}  // namespace libvpx_test
+
+#endif  // _WIN64
+
+#if ARCH_X86 || ARCH_X86_64
+#if defined(__GNUC__)
+
+namespace libvpx_test {
+
+// Checks the FPU tag word pre/post execution to ensure emms has been called.
+class RegisterStateCheckMMX {
+ public:
+  RegisterStateCheckMMX() {
+    __asm__ volatile("fstenv %0" : "=rm"(pre_fpu_env_));
+  }
+  ~RegisterStateCheckMMX() { EXPECT_TRUE(Check()); }
+
+ private:
+  // Checks the FPU tag word pre/post execution, returning false if not cleared
+  // to 0xffff.
+  bool Check() const {
+    EXPECT_EQ(0xffff, pre_fpu_env_[4])
+        << "FPU was in an inconsistent state prior to call";
+
+    uint16_t post_fpu_env[14];
+    __asm__ volatile("fstenv %0" : "=rm"(post_fpu_env));
+    EXPECT_EQ(0xffff, post_fpu_env[4])
+        << "FPU was left in an inconsistent state after call";
+    return !testing::Test::HasNonfatalFailure();
+  }
+
+  uint16_t pre_fpu_env_[14];
+};
+
+#define API_REGISTER_STATE_CHECK(statement) do {  \
+  libvpx_test::RegisterStateCheckMMX reg_check;   \
+  ASM_REGISTER_STATE_CHECK(statement);            \
+} while (false)
+
+}  // namespace libvpx_test
+
+#endif  // __GNUC__
+#endif  // ARCH_X86 || ARCH_X86_64
+
+#ifndef API_REGISTER_STATE_CHECK
+#define API_REGISTER_STATE_CHECK ASM_REGISTER_STATE_CHECK
+#endif
+
+#endif  // TEST_REGISTER_STATE_CHECK_H_
diff --git a/libvpx/libvpx/test/resize_test.cc b/libvpx/libvpx/test/resize_test.cc
new file mode 100644
index 0000000..90f5452
--- /dev/null
+++ b/libvpx/libvpx/test/resize_test.cc
@@ -0,0 +1,735 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <stdio.h>
+
+#include <climits>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/video_source.h"
+#include "test/util.h"
+
+// Enable(1) or Disable(0) writing of the compressed bitstream.
+#define WRITE_COMPRESSED_STREAM 0
+
+namespace {
+
+#if WRITE_COMPRESSED_STREAM
+static void mem_put_le16(char *const mem, const unsigned int val) {
+  mem[0] = val;
+  mem[1] = val >> 8;
+}
+
+static void mem_put_le32(char *const mem, const unsigned int val) {
+  mem[0] = val;
+  mem[1] = val >> 8;
+  mem[2] = val >> 16;
+  mem[3] = val >> 24;
+}
+
+static void write_ivf_file_header(const vpx_codec_enc_cfg_t *const cfg,
+                                  int frame_cnt, FILE *const outfile) {
+  char header[32];
+
+  header[0] = 'D';
+  header[1] = 'K';
+  header[2] = 'I';
+  header[3] = 'F';
+  mem_put_le16(header + 4,  0);                   /* version */
+  mem_put_le16(header + 6,  32);                  /* headersize */
+  mem_put_le32(header + 8,  0x30395056);          /* fourcc (vp9) */
+  mem_put_le16(header + 12, cfg->g_w);            /* width */
+  mem_put_le16(header + 14, cfg->g_h);            /* height */
+  mem_put_le32(header + 16, cfg->g_timebase.den); /* rate */
+  mem_put_le32(header + 20, cfg->g_timebase.num); /* scale */
+  mem_put_le32(header + 24, frame_cnt);           /* length */
+  mem_put_le32(header + 28, 0);                   /* unused */
+
+  (void)fwrite(header, 1, 32, outfile);
+}
+
+static void write_ivf_frame_size(FILE *const outfile, const size_t size) {
+  char header[4];
+  mem_put_le32(header, static_cast<unsigned int>(size));
+  (void)fwrite(header, 1, 4, outfile);
+}
+
+static void write_ivf_frame_header(const vpx_codec_cx_pkt_t *const pkt,
+                                   FILE *const outfile) {
+  char header[12];
+  vpx_codec_pts_t pts;
+
+  if (pkt->kind != VPX_CODEC_CX_FRAME_PKT)
+    return;
+
+  pts = pkt->data.frame.pts;
+  mem_put_le32(header, static_cast<unsigned int>(pkt->data.frame.sz));
+  mem_put_le32(header + 4, pts & 0xFFFFFFFF);
+  mem_put_le32(header + 8, pts >> 32);
+
+  (void)fwrite(header, 1, 12, outfile);
+}
+#endif  // WRITE_COMPRESSED_STREAM
+
+const unsigned int kInitialWidth = 320;
+const unsigned int kInitialHeight = 240;
+
+struct FrameInfo {
+  FrameInfo(vpx_codec_pts_t _pts, unsigned int _w, unsigned int _h)
+      : pts(_pts), w(_w), h(_h) {}
+
+  vpx_codec_pts_t pts;
+  unsigned int w;
+  unsigned int h;
+};
+
+void ScaleForFrameNumber(unsigned int frame,
+                         unsigned int initial_w,
+                         unsigned int initial_h,
+                         unsigned int *w,
+                         unsigned int *h,
+                         int flag_codec) {
+  if (frame < 10) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 20) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 30) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 40) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 50) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 60) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 70) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 80) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 90) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 100) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 110) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 120) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 130) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 140) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 150) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 160) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 170) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 180) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 190) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 200) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 210) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 220) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 230) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 240) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 250) {
+    *w = initial_w  / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 260) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  // Go down very low.
+  if (frame < 270) {
+    *w = initial_w / 4;
+    *h = initial_h / 4;
+    return;
+  }
+  if (flag_codec == 1) {
+    // Cases that only works for VP9.
+    // For VP9: Swap width and height of original.
+    if (frame < 320) {
+      *w = initial_h;
+      *h = initial_w;
+      return;
+    }
+  }
+  *w = initial_w;
+  *h = initial_h;
+}
+
+class ResizingVideoSource : public ::libvpx_test::DummyVideoSource {
+ public:
+  ResizingVideoSource() {
+    SetSize(kInitialWidth, kInitialHeight);
+    limit_ = 350;
+  }
+  int flag_codec_;
+  virtual ~ResizingVideoSource() {}
+
+ protected:
+  virtual void Next() {
+    ++frame_;
+    unsigned int width;
+    unsigned int height;
+    ScaleForFrameNumber(frame_, kInitialWidth, kInitialHeight, &width, &height,
+                        flag_codec_);
+    SetSize(width, height);
+    FillFrame();
+  }
+};
+
+class ResizeTest : public ::libvpx_test::EncoderTest,
+  public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ protected:
+  ResizeTest() : EncoderTest(GET_PARAM(0)) {}
+
+  virtual ~ResizeTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+  }
+
+  virtual void DecompressedFrameHook(const vpx_image_t &img,
+                                     vpx_codec_pts_t pts) {
+    frame_info_list_.push_back(FrameInfo(pts, img.d_w, img.d_h));
+  }
+
+  std::vector< FrameInfo > frame_info_list_;
+};
+
+TEST_P(ResizeTest, TestExternalResizeWorks) {
+  ResizingVideoSource video;
+  video.flag_codec_ = 0;
+  cfg_.g_lag_in_frames = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  for (std::vector<FrameInfo>::const_iterator info = frame_info_list_.begin();
+       info != frame_info_list_.end(); ++info) {
+    const unsigned int frame = static_cast<unsigned>(info->pts);
+    unsigned int expected_w;
+    unsigned int expected_h;
+    ScaleForFrameNumber(frame, kInitialWidth, kInitialHeight,
+                        &expected_w, &expected_h, 0);
+    EXPECT_EQ(expected_w, info->w)
+        << "Frame " << frame << " had unexpected width";
+    EXPECT_EQ(expected_h, info->h)
+        << "Frame " << frame << " had unexpected height";
+  }
+}
+
+const unsigned int kStepDownFrame = 3;
+const unsigned int kStepUpFrame = 6;
+
+class ResizeInternalTest : public ResizeTest {
+ protected:
+#if WRITE_COMPRESSED_STREAM
+  ResizeInternalTest()
+      : ResizeTest(),
+        frame0_psnr_(0.0),
+        outfile_(NULL),
+        out_frames_(0) {}
+#else
+  ResizeInternalTest() : ResizeTest(), frame0_psnr_(0.0) {}
+#endif
+
+  virtual ~ResizeInternalTest() {}
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+#if WRITE_COMPRESSED_STREAM
+    outfile_ = fopen("vp90-2-05-resize.ivf", "wb");
+#endif
+  }
+
+  virtual void EndPassHook() {
+#if WRITE_COMPRESSED_STREAM
+    if (outfile_) {
+      if (!fseek(outfile_, 0, SEEK_SET))
+        write_ivf_file_header(&cfg_, out_frames_, outfile_);
+      fclose(outfile_);
+      outfile_ = NULL;
+    }
+#endif
+  }
+
+  virtual void PreEncodeFrameHook(libvpx_test::VideoSource *video,
+                                  libvpx_test::Encoder *encoder) {
+    if (change_config_) {
+      int new_q = 60;
+      if (video->frame() == 0) {
+        struct vpx_scaling_mode mode = {VP8E_ONETWO, VP8E_ONETWO};
+        encoder->Control(VP8E_SET_SCALEMODE, &mode);
+      }
+      if (video->frame() == 1) {
+        struct vpx_scaling_mode mode = {VP8E_NORMAL, VP8E_NORMAL};
+        encoder->Control(VP8E_SET_SCALEMODE, &mode);
+        cfg_.rc_min_quantizer = cfg_.rc_max_quantizer = new_q;
+        encoder->Config(&cfg_);
+      }
+    } else {
+      if (video->frame() == kStepDownFrame) {
+        struct vpx_scaling_mode mode = {VP8E_FOURFIVE, VP8E_THREEFIVE};
+        encoder->Control(VP8E_SET_SCALEMODE, &mode);
+      }
+      if (video->frame() == kStepUpFrame) {
+        struct vpx_scaling_mode mode = {VP8E_NORMAL, VP8E_NORMAL};
+        encoder->Control(VP8E_SET_SCALEMODE, &mode);
+      }
+    }
+  }
+
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (!frame0_psnr_)
+      frame0_psnr_ = pkt->data.psnr.psnr[0];
+    EXPECT_NEAR(pkt->data.psnr.psnr[0], frame0_psnr_, 2.0);
+  }
+
+#if WRITE_COMPRESSED_STREAM
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    ++out_frames_;
+
+    // Write initial file header if first frame.
+    if (pkt->data.frame.pts == 0)
+      write_ivf_file_header(&cfg_, 0, outfile_);
+
+    // Write frame header and data.
+    write_ivf_frame_header(pkt, outfile_);
+    (void)fwrite(pkt->data.frame.buf, 1, pkt->data.frame.sz, outfile_);
+  }
+#endif
+
+  double frame0_psnr_;
+  bool change_config_;
+#if WRITE_COMPRESSED_STREAM
+  FILE *outfile_;
+  unsigned int out_frames_;
+#endif
+};
+
+TEST_P(ResizeInternalTest, TestInternalResizeWorks) {
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 10);
+  init_flags_ = VPX_CODEC_USE_PSNR;
+  change_config_ = false;
+
+  // q picked such that initial keyframe on this clip is ~30dB PSNR
+  cfg_.rc_min_quantizer = cfg_.rc_max_quantizer = 48;
+
+  // If the number of frames being encoded is smaller than g_lag_in_frames
+  // the encoded frame is unavailable using the current API. Comparing
+  // frames to detect mismatch would then not be possible. Set
+  // g_lag_in_frames = 0 to get around this.
+  cfg_.g_lag_in_frames = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  for (std::vector<FrameInfo>::const_iterator info = frame_info_list_.begin();
+       info != frame_info_list_.end(); ++info) {
+    const vpx_codec_pts_t pts = info->pts;
+    if (pts >= kStepDownFrame && pts < kStepUpFrame) {
+      ASSERT_EQ(282U, info->w) << "Frame " << pts << " had unexpected width";
+      ASSERT_EQ(173U, info->h) << "Frame " << pts << " had unexpected height";
+    } else {
+      EXPECT_EQ(352U, info->w) << "Frame " << pts << " had unexpected width";
+      EXPECT_EQ(288U, info->h) << "Frame " << pts << " had unexpected height";
+    }
+  }
+}
+
+TEST_P(ResizeInternalTest, TestInternalResizeChangeConfig) {
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 10);
+  cfg_.g_w = 352;
+  cfg_.g_h = 288;
+  change_config_ = true;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+class ResizeRealtimeTest : public ::libvpx_test::EncoderTest,
+  public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
+ protected:
+  ResizeRealtimeTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~ResizeRealtimeTest() {}
+
+  virtual void PreEncodeFrameHook(libvpx_test::VideoSource *video,
+                                  libvpx_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      encoder->Control(VP9E_SET_AQ_MODE, 3);
+      encoder->Control(VP8E_SET_CPUUSED, set_cpu_used_);
+    }
+
+    if (change_bitrate_ && video->frame() == 120) {
+      change_bitrate_ = false;
+      cfg_.rc_target_bitrate = 500;
+      encoder->Config(&cfg_);
+    }
+  }
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    set_cpu_used_ = GET_PARAM(2);
+  }
+
+  virtual void DecompressedFrameHook(const vpx_image_t &img,
+                                     vpx_codec_pts_t pts) {
+    frame_info_list_.push_back(FrameInfo(pts, img.d_w, img.d_h));
+  }
+
+  virtual void MismatchHook(const vpx_image_t *img1,
+                             const vpx_image_t *img2) {
+    double mismatch_psnr = compute_psnr(img1, img2);
+    mismatch_psnr_ += mismatch_psnr;
+    ++mismatch_nframes_;
+  }
+
+  unsigned int GetMismatchFrames() {
+      return mismatch_nframes_;
+  }
+
+  void DefaultConfig() {
+    cfg_.rc_buf_initial_sz = 500;
+    cfg_.rc_buf_optimal_sz = 600;
+    cfg_.rc_buf_sz = 1000;
+    cfg_.rc_min_quantizer = 2;
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_undershoot_pct = 50;
+    cfg_.rc_overshoot_pct = 50;
+    cfg_.rc_end_usage = VPX_CBR;
+    cfg_.kf_mode = VPX_KF_AUTO;
+    cfg_.g_lag_in_frames = 0;
+    cfg_.kf_min_dist = cfg_.kf_max_dist = 3000;
+    // Enable dropped frames.
+    cfg_.rc_dropframe_thresh = 1;
+    // Enable error_resilience mode.
+    cfg_.g_error_resilient  = 1;
+    // Enable dynamic resizing.
+    cfg_.rc_resize_allowed = 1;
+    // Run at low bitrate.
+    cfg_.rc_target_bitrate = 200;
+  }
+
+  std::vector< FrameInfo > frame_info_list_;
+  int set_cpu_used_;
+  bool change_bitrate_;
+  double mismatch_psnr_;
+  int mismatch_nframes_;
+};
+
+TEST_P(ResizeRealtimeTest, TestExternalResizeWorks) {
+  ResizingVideoSource video;
+  video.flag_codec_ = 1;
+  DefaultConfig();
+  // Disable internal resize for this test.
+  cfg_.rc_resize_allowed = 0;
+  change_bitrate_ = false;
+  mismatch_psnr_ = 0.0;
+  mismatch_nframes_ = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  for (std::vector<FrameInfo>::const_iterator info = frame_info_list_.begin();
+       info != frame_info_list_.end(); ++info) {
+    const unsigned int frame = static_cast<unsigned>(info->pts);
+    unsigned int expected_w;
+    unsigned int expected_h;
+    ScaleForFrameNumber(frame, kInitialWidth, kInitialHeight,
+                        &expected_w, &expected_h, 1);
+    EXPECT_EQ(expected_w, info->w)
+        << "Frame " << frame << " had unexpected width";
+    EXPECT_EQ(expected_h, info->h)
+        << "Frame " << frame << " had unexpected height";
+    EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+  }
+}
+
+// Verify the dynamic resizer behavior for real time, 1 pass CBR mode.
+// Run at low bitrate, with resize_allowed = 1, and verify that we get
+// one resize down event.
+TEST_P(ResizeRealtimeTest, TestInternalResizeDown) {
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 299);
+  DefaultConfig();
+  cfg_.g_w = 352;
+  cfg_.g_h = 288;
+  change_bitrate_ = false;
+  mismatch_psnr_ = 0.0;
+  mismatch_nframes_ = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  unsigned int last_w = cfg_.g_w;
+  unsigned int last_h = cfg_.g_h;
+  int resize_count = 0;
+  for (std::vector<FrameInfo>::const_iterator info = frame_info_list_.begin();
+       info != frame_info_list_.end(); ++info) {
+    if (info->w != last_w || info->h != last_h) {
+      // Verify that resize down occurs.
+      ASSERT_LT(info->w, last_w);
+      ASSERT_LT(info->h, last_h);
+      last_w = info->w;
+      last_h = info->h;
+      resize_count++;
+    }
+  }
+
+#if CONFIG_VP9_DECODER
+  // Verify that we get 1 resize down event in this test.
+  ASSERT_EQ(1, resize_count) << "Resizing should occur.";
+  EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+#else
+  printf("Warning: VP9 decoder unavailable, unable to check resize count!\n");
+#endif
+}
+
+// Verify the dynamic resizer behavior for real time, 1 pass CBR mode.
+// Start at low target bitrate, raise the bitrate in the middle of the clip,
+// scaling-up should occur after bitrate changed.
+TEST_P(ResizeRealtimeTest, TestInternalResizeDownUpChangeBitRate) {
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 359);
+  DefaultConfig();
+  cfg_.g_w = 352;
+  cfg_.g_h = 288;
+  change_bitrate_ = true;
+  mismatch_psnr_ = 0.0;
+  mismatch_nframes_ = 0;
+  // Disable dropped frames.
+  cfg_.rc_dropframe_thresh = 0;
+  // Starting bitrate low.
+  cfg_.rc_target_bitrate = 80;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  unsigned int last_w = cfg_.g_w;
+  unsigned int last_h = cfg_.g_h;
+  int resize_count = 0;
+  for (std::vector<FrameInfo>::const_iterator info = frame_info_list_.begin();
+       info != frame_info_list_.end(); ++info) {
+    if (info->w != last_w || info->h != last_h) {
+      resize_count++;
+      if (resize_count == 1) {
+        // Verify that resize down occurs.
+        ASSERT_LT(info->w, last_w);
+        ASSERT_LT(info->h, last_h);
+      } else if (resize_count == 2) {
+        // Verify that resize up occurs.
+        ASSERT_GT(info->w, last_w);
+        ASSERT_GT(info->h, last_h);
+      }
+      last_w = info->w;
+      last_h = info->h;
+    }
+  }
+
+#if CONFIG_VP9_DECODER
+  // Verify that we get 2 resize events in this test.
+  ASSERT_EQ(resize_count, 2) << "Resizing should occur twice.";
+  EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+#else
+  printf("Warning: VP9 decoder unavailable, unable to check resize count!\n");
+#endif
+}
+
+vpx_img_fmt_t CspForFrameNumber(int frame) {
+  if (frame < 10)
+    return VPX_IMG_FMT_I420;
+  if (frame < 20)
+    return VPX_IMG_FMT_I444;
+  return VPX_IMG_FMT_I420;
+}
+
+class ResizeCspTest : public ResizeTest {
+ protected:
+#if WRITE_COMPRESSED_STREAM
+  ResizeCspTest()
+      : ResizeTest(),
+        frame0_psnr_(0.0),
+        outfile_(NULL),
+        out_frames_(0) {}
+#else
+  ResizeCspTest() : ResizeTest(), frame0_psnr_(0.0) {}
+#endif
+
+  virtual ~ResizeCspTest() {}
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+#if WRITE_COMPRESSED_STREAM
+    outfile_ = fopen("vp91-2-05-cspchape.ivf", "wb");
+#endif
+  }
+
+  virtual void EndPassHook() {
+#if WRITE_COMPRESSED_STREAM
+    if (outfile_) {
+      if (!fseek(outfile_, 0, SEEK_SET))
+        write_ivf_file_header(&cfg_, out_frames_, outfile_);
+      fclose(outfile_);
+      outfile_ = NULL;
+    }
+#endif
+  }
+
+  virtual void PreEncodeFrameHook(libvpx_test::VideoSource *video,
+                                  libvpx_test::Encoder *encoder) {
+    if (CspForFrameNumber(video->frame()) != VPX_IMG_FMT_I420 &&
+        cfg_.g_profile != 1) {
+      cfg_.g_profile = 1;
+      encoder->Config(&cfg_);
+    }
+    if (CspForFrameNumber(video->frame()) == VPX_IMG_FMT_I420 &&
+        cfg_.g_profile != 0) {
+      cfg_.g_profile = 0;
+      encoder->Config(&cfg_);
+    }
+  }
+
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (!frame0_psnr_)
+      frame0_psnr_ = pkt->data.psnr.psnr[0];
+    EXPECT_NEAR(pkt->data.psnr.psnr[0], frame0_psnr_, 2.0);
+  }
+
+#if WRITE_COMPRESSED_STREAM
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    ++out_frames_;
+
+    // Write initial file header if first frame.
+    if (pkt->data.frame.pts == 0)
+      write_ivf_file_header(&cfg_, 0, outfile_);
+
+    // Write frame header and data.
+    write_ivf_frame_header(pkt, outfile_);
+    (void)fwrite(pkt->data.frame.buf, 1, pkt->data.frame.sz, outfile_);
+  }
+#endif
+
+  double frame0_psnr_;
+#if WRITE_COMPRESSED_STREAM
+  FILE *outfile_;
+  unsigned int out_frames_;
+#endif
+};
+
+class ResizingCspVideoSource : public ::libvpx_test::DummyVideoSource {
+ public:
+  ResizingCspVideoSource() {
+    SetSize(kInitialWidth, kInitialHeight);
+    limit_ = 30;
+  }
+
+  virtual ~ResizingCspVideoSource() {}
+
+ protected:
+  virtual void Next() {
+    ++frame_;
+    SetImageFormat(CspForFrameNumber(frame_));
+    FillFrame();
+  }
+};
+
+TEST_P(ResizeCspTest, TestResizeCspWorks) {
+  ResizingCspVideoSource video;
+  init_flags_ = VPX_CODEC_USE_PSNR;
+  cfg_.rc_min_quantizer = cfg_.rc_max_quantizer = 48;
+  cfg_.g_lag_in_frames = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+VP8_INSTANTIATE_TEST_CASE(ResizeTest, ONE_PASS_TEST_MODES);
+VP9_INSTANTIATE_TEST_CASE(ResizeTest,
+                          ::testing::Values(::libvpx_test::kRealTime));
+VP9_INSTANTIATE_TEST_CASE(ResizeInternalTest,
+                          ::testing::Values(::libvpx_test::kOnePassBest));
+VP9_INSTANTIATE_TEST_CASE(ResizeRealtimeTest,
+                          ::testing::Values(::libvpx_test::kRealTime),
+                          ::testing::Range(5, 9));
+VP9_INSTANTIATE_TEST_CASE(ResizeCspTest,
+                          ::testing::Values(::libvpx_test::kRealTime));
+}  // namespace
diff --git a/libvpx/libvpx/test/resize_util.sh b/libvpx/libvpx/test/resize_util.sh
new file mode 100755
index 0000000..5e47271
--- /dev/null
+++ b/libvpx/libvpx/test/resize_util.sh
@@ -0,0 +1,69 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx resize_util example code. To add new tests to
+##  this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to resize_util_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+resize_util_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Resizes $YUV_RAW_INPUT using the resize_util example. $1 is the output
+# dimensions that will be passed to resize_util.
+resize_util() {
+  local resizer="${LIBVPX_BIN_PATH}/resize_util${VPX_TEST_EXE_SUFFIX}"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/resize_util.raw"
+  local frames_to_resize="10"
+  local target_dimensions="$1"
+
+  # resize_util is available only when CONFIG_SHARED is disabled.
+  if [ -z "$(vpx_config_option_enabled CONFIG_SHARED)" ]; then
+    if [ ! -x "${resizer}" ]; then
+      elog "${resizer} does not exist or is not executable."
+      return 1
+    fi
+
+    eval "${VPX_TEST_PREFIX}" "${resizer}" "${YUV_RAW_INPUT}" \
+        "${YUV_RAW_INPUT_WIDTH}x${YUV_RAW_INPUT_HEIGHT}" \
+        "${target_dimensions}" "${output_file}" ${frames_to_resize} \
+        ${devnull}
+
+    [ -e "${output_file}" ] || return 1
+  fi
+}
+
+# Halves each dimension of $YUV_RAW_INPUT using resize_util().
+resize_down() {
+  local target_width=$((${YUV_RAW_INPUT_WIDTH} / 2))
+  local target_height=$((${YUV_RAW_INPUT_HEIGHT} / 2))
+
+  resize_util "${target_width}x${target_height}"
+}
+
+# Doubles each dimension of $YUV_RAW_INPUT using resize_util().
+resize_up() {
+  local target_width=$((${YUV_RAW_INPUT_WIDTH} * 2))
+  local target_height=$((${YUV_RAW_INPUT_HEIGHT} * 2))
+
+  resize_util "${target_width}x${target_height}"
+}
+
+resize_util_tests="resize_down
+                   resize_up"
+
+run_tests resize_util_verify_environment "${resize_util_tests}"
diff --git a/libvpx/libvpx/test/sad_test.cc b/libvpx/libvpx/test/sad_test.cc
new file mode 100644
index 0000000..e6bd0d7
--- /dev/null
+++ b/libvpx/libvpx/test/sad_test.cc
@@ -0,0 +1,951 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <string.h>
+#include <limits.h>
+#include <stdio.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vpx/vpx_codec.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+typedef unsigned int (*SadMxNFunc)(const uint8_t *src_ptr,
+                                   int src_stride,
+                                   const uint8_t *ref_ptr,
+                                   int ref_stride);
+typedef std::tr1::tuple<int, int, SadMxNFunc, int> SadMxNParam;
+
+typedef uint32_t (*SadMxNAvgFunc)(const uint8_t *src_ptr,
+                                  int src_stride,
+                                  const uint8_t *ref_ptr,
+                                  int ref_stride,
+                                  const uint8_t *second_pred);
+typedef std::tr1::tuple<int, int, SadMxNAvgFunc, int> SadMxNAvgParam;
+
+typedef void (*SadMxNx4Func)(const uint8_t *src_ptr,
+                             int src_stride,
+                             const uint8_t *const ref_ptr[],
+                             int ref_stride,
+                             uint32_t *sad_array);
+typedef std::tr1::tuple<int, int, SadMxNx4Func, int> SadMxNx4Param;
+
+using libvpx_test::ACMRandom;
+
+namespace {
+class SADTestBase : public ::testing::Test {
+ public:
+  SADTestBase(int width, int height, int bit_depth) :
+      width_(width), height_(height), bd_(bit_depth) {}
+
+  static void SetUpTestCase() {
+    source_data8_ = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kDataBlockSize));
+    reference_data8_ = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, kDataBufferSize));
+    second_pred8_ = reinterpret_cast<uint8_t*>(
+        vpx_memalign(kDataAlignment, 64*64));
+    source_data16_ = reinterpret_cast<uint16_t*>(
+        vpx_memalign(kDataAlignment, kDataBlockSize*sizeof(uint16_t)));
+    reference_data16_ = reinterpret_cast<uint16_t*>(
+        vpx_memalign(kDataAlignment, kDataBufferSize*sizeof(uint16_t)));
+    second_pred16_ = reinterpret_cast<uint16_t*>(
+        vpx_memalign(kDataAlignment, 64*64*sizeof(uint16_t)));
+  }
+
+  static void TearDownTestCase() {
+    vpx_free(source_data8_);
+    source_data8_ = NULL;
+    vpx_free(reference_data8_);
+    reference_data8_ = NULL;
+    vpx_free(second_pred8_);
+    second_pred8_ = NULL;
+    vpx_free(source_data16_);
+    source_data16_ = NULL;
+    vpx_free(reference_data16_);
+    reference_data16_ = NULL;
+    vpx_free(second_pred16_);
+    second_pred16_ = NULL;
+  }
+
+  virtual void TearDown() {
+    libvpx_test::ClearSystemState();
+  }
+
+ protected:
+  // Handle blocks up to 4 blocks 64x64 with stride up to 128
+  static const int kDataAlignment = 16;
+  static const int kDataBlockSize = 64 * 128;
+  static const int kDataBufferSize = 4 * kDataBlockSize;
+
+  virtual void SetUp() {
+    if (bd_ == -1) {
+      use_high_bit_depth_ = false;
+      bit_depth_ = VPX_BITS_8;
+      source_data_ = source_data8_;
+      reference_data_ = reference_data8_;
+      second_pred_ = second_pred8_;
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else {
+      use_high_bit_depth_ = true;
+      bit_depth_ = static_cast<vpx_bit_depth_t>(bd_);
+      source_data_ = CONVERT_TO_BYTEPTR(source_data16_);
+      reference_data_ = CONVERT_TO_BYTEPTR(reference_data16_);
+      second_pred_ = CONVERT_TO_BYTEPTR(second_pred16_);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    }
+    mask_ = (1 << bit_depth_) - 1;
+    source_stride_ = (width_ + 31) & ~31;
+    reference_stride_ = width_ * 2;
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+  }
+
+  virtual uint8_t *GetReference(int block_idx) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (use_high_bit_depth_)
+      return CONVERT_TO_BYTEPTR(CONVERT_TO_SHORTPTR(reference_data_) +
+                                block_idx * kDataBlockSize);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    return reference_data_ + block_idx * kDataBlockSize;
+  }
+
+  // Sum of Absolute Differences. Given two blocks, calculate the absolute
+  // difference between two pixels in the same relative location; accumulate.
+  unsigned int ReferenceSAD(int block_idx) {
+    unsigned int sad = 0;
+      const uint8_t *const reference8 = GetReference(block_idx);
+      const uint8_t *const source8 = source_data_;
+#if CONFIG_VP9_HIGHBITDEPTH
+      const uint16_t *const reference16 =
+          CONVERT_TO_SHORTPTR(GetReference(block_idx));
+      const uint16_t *const source16 = CONVERT_TO_SHORTPTR(source_data_);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    for (int h = 0; h < height_; ++h) {
+      for (int w = 0; w < width_; ++w) {
+        if (!use_high_bit_depth_) {
+          sad += abs(source8[h * source_stride_ + w] -
+                     reference8[h * reference_stride_ + w]);
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          sad += abs(source16[h * source_stride_ + w] -
+                     reference16[h * reference_stride_ + w]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        }
+      }
+    }
+    return sad;
+  }
+
+  // Sum of Absolute Differences Average. Given two blocks, and a prediction
+  // calculate the absolute difference between one pixel and average of the
+  // corresponding and predicted pixels; accumulate.
+  unsigned int ReferenceSADavg(int block_idx) {
+    unsigned int sad = 0;
+    const uint8_t *const reference8 = GetReference(block_idx);
+    const uint8_t *const source8 = source_data_;
+    const uint8_t *const second_pred8 = second_pred_;
+#if CONFIG_VP9_HIGHBITDEPTH
+    const uint16_t *const reference16 =
+        CONVERT_TO_SHORTPTR(GetReference(block_idx));
+    const uint16_t *const source16 = CONVERT_TO_SHORTPTR(source_data_);
+    const uint16_t *const second_pred16 = CONVERT_TO_SHORTPTR(second_pred_);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    for (int h = 0; h < height_; ++h) {
+      for (int w = 0; w < width_; ++w) {
+        if (!use_high_bit_depth_) {
+          const int tmp = second_pred8[h * width_ + w] +
+              reference8[h * reference_stride_ + w];
+          const uint8_t comp_pred = ROUND_POWER_OF_TWO(tmp, 1);
+          sad += abs(source8[h * source_stride_ + w] - comp_pred);
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          const int tmp = second_pred16[h * width_ + w] +
+              reference16[h * reference_stride_ + w];
+          const uint16_t comp_pred = ROUND_POWER_OF_TWO(tmp, 1);
+          sad += abs(source16[h * source_stride_ + w] - comp_pred);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        }
+      }
+    }
+    return sad;
+  }
+
+  void FillConstant(uint8_t *data, int stride, uint16_t fill_constant) {
+    uint8_t *data8 = data;
+#if CONFIG_VP9_HIGHBITDEPTH
+    uint16_t *data16 = CONVERT_TO_SHORTPTR(data);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    for (int h = 0; h < height_; ++h) {
+      for (int w = 0; w < width_; ++w) {
+        if (!use_high_bit_depth_) {
+          data8[h * stride + w] = static_cast<uint8_t>(fill_constant);
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          data16[h * stride + w] = fill_constant;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        }
+      }
+    }
+  }
+
+  void FillRandom(uint8_t *data, int stride) {
+    uint8_t *data8 = data;
+#if CONFIG_VP9_HIGHBITDEPTH
+    uint16_t *data16 = CONVERT_TO_SHORTPTR(data);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    for (int h = 0; h < height_; ++h) {
+      for (int w = 0; w < width_; ++w) {
+        if (!use_high_bit_depth_) {
+          data8[h * stride + w] = rnd_.Rand8();
+#if CONFIG_VP9_HIGHBITDEPTH
+        } else {
+          data16[h * stride + w] = rnd_.Rand16() & mask_;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        }
+      }
+    }
+  }
+
+  int width_, height_, mask_, bd_;
+  vpx_bit_depth_t bit_depth_;
+  static uint8_t *source_data_;
+  static uint8_t *reference_data_;
+  static uint8_t *second_pred_;
+  int source_stride_;
+  bool use_high_bit_depth_;
+  static uint8_t *source_data8_;
+  static uint8_t *reference_data8_;
+  static uint8_t *second_pred8_;
+  static uint16_t *source_data16_;
+  static uint16_t *reference_data16_;
+  static uint16_t *second_pred16_;
+  int reference_stride_;
+
+  ACMRandom rnd_;
+};
+
+class SADx4Test
+    : public SADTestBase,
+      public ::testing::WithParamInterface<SadMxNx4Param> {
+ public:
+  SADx4Test() : SADTestBase(GET_PARAM(0), GET_PARAM(1), GET_PARAM(3)) {}
+
+ protected:
+  void SADs(unsigned int *results) {
+    const uint8_t *references[] = {GetReference(0), GetReference(1),
+                                   GetReference(2), GetReference(3)};
+
+    ASM_REGISTER_STATE_CHECK(GET_PARAM(2)(source_data_, source_stride_,
+                                          references, reference_stride_,
+                                          results));
+  }
+
+  void CheckSADs() {
+    unsigned int reference_sad, exp_sad[4];
+
+    SADs(exp_sad);
+    for (int block = 0; block < 4; ++block) {
+      reference_sad = ReferenceSAD(block);
+
+      EXPECT_EQ(reference_sad, exp_sad[block]) << "block " << block;
+    }
+  }
+};
+
+class SADTest
+    : public SADTestBase,
+      public ::testing::WithParamInterface<SadMxNParam> {
+ public:
+  SADTest() : SADTestBase(GET_PARAM(0), GET_PARAM(1), GET_PARAM(3)) {}
+
+ protected:
+  unsigned int SAD(int block_idx) {
+    unsigned int ret;
+    const uint8_t *const reference = GetReference(block_idx);
+
+    ASM_REGISTER_STATE_CHECK(ret = GET_PARAM(2)(source_data_, source_stride_,
+                                                reference, reference_stride_));
+    return ret;
+  }
+
+  void CheckSAD() {
+    const unsigned int reference_sad = ReferenceSAD(0);
+    const unsigned int exp_sad = SAD(0);
+
+    ASSERT_EQ(reference_sad, exp_sad);
+  }
+};
+
+class SADavgTest
+    : public SADTestBase,
+      public ::testing::WithParamInterface<SadMxNAvgParam> {
+ public:
+  SADavgTest() : SADTestBase(GET_PARAM(0), GET_PARAM(1), GET_PARAM(3)) {}
+
+ protected:
+  unsigned int SAD_avg(int block_idx) {
+    unsigned int ret;
+    const uint8_t *const reference = GetReference(block_idx);
+
+    ASM_REGISTER_STATE_CHECK(ret = GET_PARAM(2)(source_data_, source_stride_,
+                                                reference, reference_stride_,
+                                                second_pred_));
+    return ret;
+  }
+
+  void CheckSAD() {
+    const unsigned int reference_sad = ReferenceSADavg(0);
+    const unsigned int exp_sad = SAD_avg(0);
+
+    ASSERT_EQ(reference_sad, exp_sad);
+  }
+};
+
+uint8_t *SADTestBase::source_data_ = NULL;
+uint8_t *SADTestBase::reference_data_ = NULL;
+uint8_t *SADTestBase::second_pred_ = NULL;
+uint8_t *SADTestBase::source_data8_ = NULL;
+uint8_t *SADTestBase::reference_data8_ = NULL;
+uint8_t *SADTestBase::second_pred8_ = NULL;
+uint16_t *SADTestBase::source_data16_ = NULL;
+uint16_t *SADTestBase::reference_data16_ = NULL;
+uint16_t *SADTestBase::second_pred16_ = NULL;
+
+TEST_P(SADTest, MaxRef) {
+  FillConstant(source_data_, source_stride_, 0);
+  FillConstant(reference_data_, reference_stride_, mask_);
+  CheckSAD();
+}
+
+TEST_P(SADTest, MaxSrc) {
+  FillConstant(source_data_, source_stride_, mask_);
+  FillConstant(reference_data_, reference_stride_, 0);
+  CheckSAD();
+}
+
+TEST_P(SADTest, ShortRef) {
+  const int tmp_stride = reference_stride_;
+  reference_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADTest, UnalignedRef) {
+  // The reference frame, but not the source frame, may be unaligned for
+  // certain types of searches.
+  const int tmp_stride = reference_stride_;
+  reference_stride_ -= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADTest, ShortSrc) {
+  const int tmp_stride = source_stride_;
+  source_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD();
+  source_stride_ = tmp_stride;
+}
+
+TEST_P(SADavgTest, MaxRef) {
+  FillConstant(source_data_, source_stride_, 0);
+  FillConstant(reference_data_, reference_stride_, mask_);
+  FillConstant(second_pred_, width_, 0);
+  CheckSAD();
+}
+TEST_P(SADavgTest, MaxSrc) {
+  FillConstant(source_data_, source_stride_, mask_);
+  FillConstant(reference_data_, reference_stride_, 0);
+  FillConstant(second_pred_, width_, 0);
+  CheckSAD();
+}
+
+TEST_P(SADavgTest, ShortRef) {
+  const int tmp_stride = reference_stride_;
+  reference_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  FillRandom(second_pred_, width_);
+  CheckSAD();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADavgTest, UnalignedRef) {
+  // The reference frame, but not the source frame, may be unaligned for
+  // certain types of searches.
+  const int tmp_stride = reference_stride_;
+  reference_stride_ -= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  FillRandom(second_pred_, width_);
+  CheckSAD();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADavgTest, ShortSrc) {
+  const int tmp_stride = source_stride_;
+  source_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  FillRandom(second_pred_, width_);
+  CheckSAD();
+  source_stride_ = tmp_stride;
+}
+
+TEST_P(SADx4Test, MaxRef) {
+  FillConstant(source_data_, source_stride_, 0);
+  FillConstant(GetReference(0), reference_stride_, mask_);
+  FillConstant(GetReference(1), reference_stride_, mask_);
+  FillConstant(GetReference(2), reference_stride_, mask_);
+  FillConstant(GetReference(3), reference_stride_, mask_);
+  CheckSADs();
+}
+
+TEST_P(SADx4Test, MaxSrc) {
+  FillConstant(source_data_, source_stride_, mask_);
+  FillConstant(GetReference(0), reference_stride_, 0);
+  FillConstant(GetReference(1), reference_stride_, 0);
+  FillConstant(GetReference(2), reference_stride_, 0);
+  FillConstant(GetReference(3), reference_stride_, 0);
+  CheckSADs();
+}
+
+TEST_P(SADx4Test, ShortRef) {
+  int tmp_stride = reference_stride_;
+  reference_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(GetReference(0), reference_stride_);
+  FillRandom(GetReference(1), reference_stride_);
+  FillRandom(GetReference(2), reference_stride_);
+  FillRandom(GetReference(3), reference_stride_);
+  CheckSADs();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADx4Test, UnalignedRef) {
+  // The reference frame, but not the source frame, may be unaligned for
+  // certain types of searches.
+  int tmp_stride = reference_stride_;
+  reference_stride_ -= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(GetReference(0), reference_stride_);
+  FillRandom(GetReference(1), reference_stride_);
+  FillRandom(GetReference(2), reference_stride_);
+  FillRandom(GetReference(3), reference_stride_);
+  CheckSADs();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADx4Test, ShortSrc) {
+  int tmp_stride = source_stride_;
+  source_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(GetReference(0), reference_stride_);
+  FillRandom(GetReference(1), reference_stride_);
+  FillRandom(GetReference(2), reference_stride_);
+  FillRandom(GetReference(3), reference_stride_);
+  CheckSADs();
+  source_stride_ = tmp_stride;
+}
+
+TEST_P(SADx4Test, SrcAlignedByWidth) {
+  uint8_t * tmp_source_data = source_data_;
+  source_data_ += width_;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(GetReference(0), reference_stride_);
+  FillRandom(GetReference(1), reference_stride_);
+  FillRandom(GetReference(2), reference_stride_);
+  FillRandom(GetReference(3), reference_stride_);
+  CheckSADs();
+  source_data_ = tmp_source_data;
+}
+
+using std::tr1::make_tuple;
+
+//------------------------------------------------------------------------------
+// C functions
+const SadMxNParam c_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64_c, -1),
+  make_tuple(64, 32, &vpx_sad64x32_c, -1),
+  make_tuple(32, 64, &vpx_sad32x64_c, -1),
+  make_tuple(32, 32, &vpx_sad32x32_c, -1),
+  make_tuple(32, 16, &vpx_sad32x16_c, -1),
+  make_tuple(16, 32, &vpx_sad16x32_c, -1),
+  make_tuple(16, 16, &vpx_sad16x16_c, -1),
+  make_tuple(16, 8, &vpx_sad16x8_c, -1),
+  make_tuple(8, 16, &vpx_sad8x16_c, -1),
+  make_tuple(8, 8, &vpx_sad8x8_c, -1),
+  make_tuple(8, 4, &vpx_sad8x4_c, -1),
+  make_tuple(4, 8, &vpx_sad4x8_c, -1),
+  make_tuple(4, 4, &vpx_sad4x4_c, -1),
+#if CONFIG_VP9_HIGHBITDEPTH
+  make_tuple(64, 64, &vpx_highbd_sad64x64_c, 8),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_c, 8),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_c, 8),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_c, 8),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_c, 8),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_c, 8),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_c, 8),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_c, 8),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_c, 8),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_c, 8),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_c, 8),
+  make_tuple(4, 8, &vpx_highbd_sad4x8_c, 8),
+  make_tuple(4, 4, &vpx_highbd_sad4x4_c, 8),
+  make_tuple(64, 64, &vpx_highbd_sad64x64_c, 10),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_c, 10),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_c, 10),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_c, 10),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_c, 10),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_c, 10),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_c, 10),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_c, 10),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_c, 10),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_c, 10),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_c, 10),
+  make_tuple(4, 8, &vpx_highbd_sad4x8_c, 10),
+  make_tuple(4, 4, &vpx_highbd_sad4x4_c, 10),
+  make_tuple(64, 64, &vpx_highbd_sad64x64_c, 12),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_c, 12),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_c, 12),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_c, 12),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_c, 12),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_c, 12),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_c, 12),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_c, 12),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_c, 12),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_c, 12),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_c, 12),
+  make_tuple(4, 8, &vpx_highbd_sad4x8_c, 12),
+  make_tuple(4, 4, &vpx_highbd_sad4x4_c, 12),
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(C, SADTest, ::testing::ValuesIn(c_tests));
+
+const SadMxNAvgParam avg_c_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64_avg_c, -1),
+  make_tuple(64, 32, &vpx_sad64x32_avg_c, -1),
+  make_tuple(32, 64, &vpx_sad32x64_avg_c, -1),
+  make_tuple(32, 32, &vpx_sad32x32_avg_c, -1),
+  make_tuple(32, 16, &vpx_sad32x16_avg_c, -1),
+  make_tuple(16, 32, &vpx_sad16x32_avg_c, -1),
+  make_tuple(16, 16, &vpx_sad16x16_avg_c, -1),
+  make_tuple(16, 8, &vpx_sad16x8_avg_c, -1),
+  make_tuple(8, 16, &vpx_sad8x16_avg_c, -1),
+  make_tuple(8, 8, &vpx_sad8x8_avg_c, -1),
+  make_tuple(8, 4, &vpx_sad8x4_avg_c, -1),
+  make_tuple(4, 8, &vpx_sad4x8_avg_c, -1),
+  make_tuple(4, 4, &vpx_sad4x4_avg_c, -1),
+#if CONFIG_VP9_HIGHBITDEPTH
+  make_tuple(64, 64, &vpx_highbd_sad64x64_avg_c, 8),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_avg_c, 8),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_avg_c, 8),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_avg_c, 8),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_avg_c, 8),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_avg_c, 8),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_avg_c, 8),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_avg_c, 8),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_avg_c, 8),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_avg_c, 8),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_avg_c, 8),
+  make_tuple(4, 8, &vpx_highbd_sad4x8_avg_c, 8),
+  make_tuple(4, 4, &vpx_highbd_sad4x4_avg_c, 8),
+  make_tuple(64, 64, &vpx_highbd_sad64x64_avg_c, 10),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_avg_c, 10),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_avg_c, 10),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_avg_c, 10),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_avg_c, 10),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_avg_c, 10),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_avg_c, 10),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_avg_c, 10),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_avg_c, 10),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_avg_c, 10),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_avg_c, 10),
+  make_tuple(4, 8, &vpx_highbd_sad4x8_avg_c, 10),
+  make_tuple(4, 4, &vpx_highbd_sad4x4_avg_c, 10),
+  make_tuple(64, 64, &vpx_highbd_sad64x64_avg_c, 12),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_avg_c, 12),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_avg_c, 12),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_avg_c, 12),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_avg_c, 12),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_avg_c, 12),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_avg_c, 12),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_avg_c, 12),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_avg_c, 12),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_avg_c, 12),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_avg_c, 12),
+  make_tuple(4, 8, &vpx_highbd_sad4x8_avg_c, 12),
+  make_tuple(4, 4, &vpx_highbd_sad4x4_avg_c, 12),
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(C, SADavgTest, ::testing::ValuesIn(avg_c_tests));
+
+const SadMxNx4Param x4d_c_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64x4d_c, -1),
+  make_tuple(64, 32, &vpx_sad64x32x4d_c, -1),
+  make_tuple(32, 64, &vpx_sad32x64x4d_c, -1),
+  make_tuple(32, 32, &vpx_sad32x32x4d_c, -1),
+  make_tuple(32, 16, &vpx_sad32x16x4d_c, -1),
+  make_tuple(16, 32, &vpx_sad16x32x4d_c, -1),
+  make_tuple(16, 16, &vpx_sad16x16x4d_c, -1),
+  make_tuple(16, 8, &vpx_sad16x8x4d_c, -1),
+  make_tuple(8, 16, &vpx_sad8x16x4d_c, -1),
+  make_tuple(8, 8, &vpx_sad8x8x4d_c, -1),
+  make_tuple(8, 4, &vpx_sad8x4x4d_c, -1),
+  make_tuple(4, 8, &vpx_sad4x8x4d_c, -1),
+  make_tuple(4, 4, &vpx_sad4x4x4d_c, -1),
+#if CONFIG_VP9_HIGHBITDEPTH
+  make_tuple(64, 64, &vpx_highbd_sad64x64x4d_c, 8),
+  make_tuple(64, 32, &vpx_highbd_sad64x32x4d_c, 8),
+  make_tuple(32, 64, &vpx_highbd_sad32x64x4d_c, 8),
+  make_tuple(32, 32, &vpx_highbd_sad32x32x4d_c, 8),
+  make_tuple(32, 16, &vpx_highbd_sad32x16x4d_c, 8),
+  make_tuple(16, 32, &vpx_highbd_sad16x32x4d_c, 8),
+  make_tuple(16, 16, &vpx_highbd_sad16x16x4d_c, 8),
+  make_tuple(16, 8, &vpx_highbd_sad16x8x4d_c, 8),
+  make_tuple(8, 16, &vpx_highbd_sad8x16x4d_c, 8),
+  make_tuple(8, 8, &vpx_highbd_sad8x8x4d_c, 8),
+  make_tuple(8, 4, &vpx_highbd_sad8x4x4d_c, 8),
+  make_tuple(4, 8, &vpx_highbd_sad4x8x4d_c, 8),
+  make_tuple(4, 4, &vpx_highbd_sad4x4x4d_c, 8),
+  make_tuple(64, 64, &vpx_highbd_sad64x64x4d_c, 10),
+  make_tuple(64, 32, &vpx_highbd_sad64x32x4d_c, 10),
+  make_tuple(32, 64, &vpx_highbd_sad32x64x4d_c, 10),
+  make_tuple(32, 32, &vpx_highbd_sad32x32x4d_c, 10),
+  make_tuple(32, 16, &vpx_highbd_sad32x16x4d_c, 10),
+  make_tuple(16, 32, &vpx_highbd_sad16x32x4d_c, 10),
+  make_tuple(16, 16, &vpx_highbd_sad16x16x4d_c, 10),
+  make_tuple(16, 8, &vpx_highbd_sad16x8x4d_c, 10),
+  make_tuple(8, 16, &vpx_highbd_sad8x16x4d_c, 10),
+  make_tuple(8, 8, &vpx_highbd_sad8x8x4d_c, 10),
+  make_tuple(8, 4, &vpx_highbd_sad8x4x4d_c, 10),
+  make_tuple(4, 8, &vpx_highbd_sad4x8x4d_c, 10),
+  make_tuple(4, 4, &vpx_highbd_sad4x4x4d_c, 10),
+  make_tuple(64, 64, &vpx_highbd_sad64x64x4d_c, 12),
+  make_tuple(64, 32, &vpx_highbd_sad64x32x4d_c, 12),
+  make_tuple(32, 64, &vpx_highbd_sad32x64x4d_c, 12),
+  make_tuple(32, 32, &vpx_highbd_sad32x32x4d_c, 12),
+  make_tuple(32, 16, &vpx_highbd_sad32x16x4d_c, 12),
+  make_tuple(16, 32, &vpx_highbd_sad16x32x4d_c, 12),
+  make_tuple(16, 16, &vpx_highbd_sad16x16x4d_c, 12),
+  make_tuple(16, 8, &vpx_highbd_sad16x8x4d_c, 12),
+  make_tuple(8, 16, &vpx_highbd_sad8x16x4d_c, 12),
+  make_tuple(8, 8, &vpx_highbd_sad8x8x4d_c, 12),
+  make_tuple(8, 4, &vpx_highbd_sad8x4x4d_c, 12),
+  make_tuple(4, 8, &vpx_highbd_sad4x8x4d_c, 12),
+  make_tuple(4, 4, &vpx_highbd_sad4x4x4d_c, 12),
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(C, SADx4Test, ::testing::ValuesIn(x4d_c_tests));
+
+//------------------------------------------------------------------------------
+// ARM functions
+#if HAVE_MEDIA
+const SadMxNParam media_tests[] = {
+  make_tuple(16, 16, &vpx_sad16x16_media, -1),
+};
+INSTANTIATE_TEST_CASE_P(MEDIA, SADTest, ::testing::ValuesIn(media_tests));
+#endif  // HAVE_MEDIA
+
+#if HAVE_NEON
+const SadMxNParam neon_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64_neon, -1),
+  make_tuple(32, 32, &vpx_sad32x32_neon, -1),
+  make_tuple(16, 16, &vpx_sad16x16_neon, -1),
+  make_tuple(16, 8, &vpx_sad16x8_neon, -1),
+  make_tuple(8, 16, &vpx_sad8x16_neon, -1),
+  make_tuple(8, 8, &vpx_sad8x8_neon, -1),
+  make_tuple(4, 4, &vpx_sad4x4_neon, -1),
+};
+INSTANTIATE_TEST_CASE_P(NEON, SADTest, ::testing::ValuesIn(neon_tests));
+
+const SadMxNx4Param x4d_neon_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64x4d_neon, -1),
+  make_tuple(32, 32, &vpx_sad32x32x4d_neon, -1),
+  make_tuple(16, 16, &vpx_sad16x16x4d_neon, -1),
+};
+INSTANTIATE_TEST_CASE_P(NEON, SADx4Test, ::testing::ValuesIn(x4d_neon_tests));
+#endif  // HAVE_NEON
+
+//------------------------------------------------------------------------------
+// x86 functions
+#if HAVE_SSE2
+#if CONFIG_USE_X86INC
+const SadMxNParam sse2_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64_sse2, -1),
+  make_tuple(64, 32, &vpx_sad64x32_sse2, -1),
+  make_tuple(32, 64, &vpx_sad32x64_sse2, -1),
+  make_tuple(32, 32, &vpx_sad32x32_sse2, -1),
+  make_tuple(32, 16, &vpx_sad32x16_sse2, -1),
+  make_tuple(16, 32, &vpx_sad16x32_sse2, -1),
+  make_tuple(16, 16, &vpx_sad16x16_sse2, -1),
+  make_tuple(16, 8, &vpx_sad16x8_sse2, -1),
+  make_tuple(8, 16, &vpx_sad8x16_sse2, -1),
+  make_tuple(8, 8, &vpx_sad8x8_sse2, -1),
+  make_tuple(8, 4, &vpx_sad8x4_sse2, -1),
+  make_tuple(4, 8, &vpx_sad4x8_sse2, -1),
+  make_tuple(4, 4, &vpx_sad4x4_sse2, -1),
+#if CONFIG_VP9_HIGHBITDEPTH
+  make_tuple(64, 64, &vpx_highbd_sad64x64_sse2, 8),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_sse2, 8),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_sse2, 8),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_sse2, 8),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_sse2, 8),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_sse2, 8),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_sse2, 8),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_sse2, 8),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_sse2, 8),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_sse2, 8),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_sse2, 8),
+  make_tuple(64, 64, &vpx_highbd_sad64x64_sse2, 10),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_sse2, 10),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_sse2, 10),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_sse2, 10),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_sse2, 10),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_sse2, 10),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_sse2, 10),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_sse2, 10),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_sse2, 10),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_sse2, 10),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_sse2, 10),
+  make_tuple(64, 64, &vpx_highbd_sad64x64_sse2, 12),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_sse2, 12),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_sse2, 12),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_sse2, 12),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_sse2, 12),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_sse2, 12),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_sse2, 12),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_sse2, 12),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_sse2, 12),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_sse2, 12),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_sse2, 12),
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(SSE2, SADTest, ::testing::ValuesIn(sse2_tests));
+
+const SadMxNAvgParam avg_sse2_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64_avg_sse2, -1),
+  make_tuple(64, 32, &vpx_sad64x32_avg_sse2, -1),
+  make_tuple(32, 64, &vpx_sad32x64_avg_sse2, -1),
+  make_tuple(32, 32, &vpx_sad32x32_avg_sse2, -1),
+  make_tuple(32, 16, &vpx_sad32x16_avg_sse2, -1),
+  make_tuple(16, 32, &vpx_sad16x32_avg_sse2, -1),
+  make_tuple(16, 16, &vpx_sad16x16_avg_sse2, -1),
+  make_tuple(16, 8, &vpx_sad16x8_avg_sse2, -1),
+  make_tuple(8, 16, &vpx_sad8x16_avg_sse2, -1),
+  make_tuple(8, 8, &vpx_sad8x8_avg_sse2, -1),
+  make_tuple(8, 4, &vpx_sad8x4_avg_sse2, -1),
+  make_tuple(4, 8, &vpx_sad4x8_avg_sse2, -1),
+  make_tuple(4, 4, &vpx_sad4x4_avg_sse2, -1),
+#if CONFIG_VP9_HIGHBITDEPTH
+  make_tuple(64, 64, &vpx_highbd_sad64x64_avg_sse2, 8),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_avg_sse2, 8),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_avg_sse2, 8),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_avg_sse2, 8),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_avg_sse2, 8),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_avg_sse2, 8),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_avg_sse2, 8),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_avg_sse2, 8),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_avg_sse2, 8),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_avg_sse2, 8),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_avg_sse2, 8),
+  make_tuple(64, 64, &vpx_highbd_sad64x64_avg_sse2, 10),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_avg_sse2, 10),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_avg_sse2, 10),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_avg_sse2, 10),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_avg_sse2, 10),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_avg_sse2, 10),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_avg_sse2, 10),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_avg_sse2, 10),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_avg_sse2, 10),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_avg_sse2, 10),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_avg_sse2, 10),
+  make_tuple(64, 64, &vpx_highbd_sad64x64_avg_sse2, 12),
+  make_tuple(64, 32, &vpx_highbd_sad64x32_avg_sse2, 12),
+  make_tuple(32, 64, &vpx_highbd_sad32x64_avg_sse2, 12),
+  make_tuple(32, 32, &vpx_highbd_sad32x32_avg_sse2, 12),
+  make_tuple(32, 16, &vpx_highbd_sad32x16_avg_sse2, 12),
+  make_tuple(16, 32, &vpx_highbd_sad16x32_avg_sse2, 12),
+  make_tuple(16, 16, &vpx_highbd_sad16x16_avg_sse2, 12),
+  make_tuple(16, 8, &vpx_highbd_sad16x8_avg_sse2, 12),
+  make_tuple(8, 16, &vpx_highbd_sad8x16_avg_sse2, 12),
+  make_tuple(8, 8, &vpx_highbd_sad8x8_avg_sse2, 12),
+  make_tuple(8, 4, &vpx_highbd_sad8x4_avg_sse2, 12),
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(SSE2, SADavgTest, ::testing::ValuesIn(avg_sse2_tests));
+
+const SadMxNx4Param x4d_sse2_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64x4d_sse2, -1),
+  make_tuple(64, 32, &vpx_sad64x32x4d_sse2, -1),
+  make_tuple(32, 64, &vpx_sad32x64x4d_sse2, -1),
+  make_tuple(32, 32, &vpx_sad32x32x4d_sse2, -1),
+  make_tuple(32, 16, &vpx_sad32x16x4d_sse2, -1),
+  make_tuple(16, 32, &vpx_sad16x32x4d_sse2, -1),
+  make_tuple(16, 16, &vpx_sad16x16x4d_sse2, -1),
+  make_tuple(16, 8, &vpx_sad16x8x4d_sse2, -1),
+  make_tuple(8, 16, &vpx_sad8x16x4d_sse2, -1),
+  make_tuple(8, 8, &vpx_sad8x8x4d_sse2, -1),
+  make_tuple(8, 4, &vpx_sad8x4x4d_sse2, -1),
+  make_tuple(4, 8, &vpx_sad4x8x4d_sse2, -1),
+  make_tuple(4, 4, &vpx_sad4x4x4d_sse2, -1),
+#if CONFIG_VP9_HIGHBITDEPTH
+  make_tuple(64, 64, &vpx_highbd_sad64x64x4d_sse2, 8),
+  make_tuple(64, 32, &vpx_highbd_sad64x32x4d_sse2, 8),
+  make_tuple(32, 64, &vpx_highbd_sad32x64x4d_sse2, 8),
+  make_tuple(32, 32, &vpx_highbd_sad32x32x4d_sse2, 8),
+  make_tuple(32, 16, &vpx_highbd_sad32x16x4d_sse2, 8),
+  make_tuple(16, 32, &vpx_highbd_sad16x32x4d_sse2, 8),
+  make_tuple(16, 16, &vpx_highbd_sad16x16x4d_sse2, 8),
+  make_tuple(16, 8, &vpx_highbd_sad16x8x4d_sse2, 8),
+  make_tuple(8, 16, &vpx_highbd_sad8x16x4d_sse2, 8),
+  make_tuple(8, 8, &vpx_highbd_sad8x8x4d_sse2, 8),
+  make_tuple(8, 4, &vpx_highbd_sad8x4x4d_sse2, 8),
+  make_tuple(4, 8, &vpx_highbd_sad4x8x4d_sse2, 8),
+  make_tuple(4, 4, &vpx_highbd_sad4x4x4d_sse2, 8),
+  make_tuple(64, 64, &vpx_highbd_sad64x64x4d_sse2, 10),
+  make_tuple(64, 32, &vpx_highbd_sad64x32x4d_sse2, 10),
+  make_tuple(32, 64, &vpx_highbd_sad32x64x4d_sse2, 10),
+  make_tuple(32, 32, &vpx_highbd_sad32x32x4d_sse2, 10),
+  make_tuple(32, 16, &vpx_highbd_sad32x16x4d_sse2, 10),
+  make_tuple(16, 32, &vpx_highbd_sad16x32x4d_sse2, 10),
+  make_tuple(16, 16, &vpx_highbd_sad16x16x4d_sse2, 10),
+  make_tuple(16, 8, &vpx_highbd_sad16x8x4d_sse2, 10),
+  make_tuple(8, 16, &vpx_highbd_sad8x16x4d_sse2, 10),
+  make_tuple(8, 8, &vpx_highbd_sad8x8x4d_sse2, 10),
+  make_tuple(8, 4, &vpx_highbd_sad8x4x4d_sse2, 10),
+  make_tuple(4, 8, &vpx_highbd_sad4x8x4d_sse2, 10),
+  make_tuple(4, 4, &vpx_highbd_sad4x4x4d_sse2, 10),
+  make_tuple(64, 64, &vpx_highbd_sad64x64x4d_sse2, 12),
+  make_tuple(64, 32, &vpx_highbd_sad64x32x4d_sse2, 12),
+  make_tuple(32, 64, &vpx_highbd_sad32x64x4d_sse2, 12),
+  make_tuple(32, 32, &vpx_highbd_sad32x32x4d_sse2, 12),
+  make_tuple(32, 16, &vpx_highbd_sad32x16x4d_sse2, 12),
+  make_tuple(16, 32, &vpx_highbd_sad16x32x4d_sse2, 12),
+  make_tuple(16, 16, &vpx_highbd_sad16x16x4d_sse2, 12),
+  make_tuple(16, 8, &vpx_highbd_sad16x8x4d_sse2, 12),
+  make_tuple(8, 16, &vpx_highbd_sad8x16x4d_sse2, 12),
+  make_tuple(8, 8, &vpx_highbd_sad8x8x4d_sse2, 12),
+  make_tuple(8, 4, &vpx_highbd_sad8x4x4d_sse2, 12),
+  make_tuple(4, 8, &vpx_highbd_sad4x8x4d_sse2, 12),
+  make_tuple(4, 4, &vpx_highbd_sad4x4x4d_sse2, 12),
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(SSE2, SADx4Test, ::testing::ValuesIn(x4d_sse2_tests));
+#endif  // CONFIG_USE_X86INC
+#endif  // HAVE_SSE2
+
+#if HAVE_SSE3
+// Only functions are x3, which do not have tests.
+#endif  // HAVE_SSE3
+
+#if HAVE_SSSE3
+// Only functions are x3, which do not have tests.
+#endif  // HAVE_SSSE3
+
+#if HAVE_SSE4_1
+// Only functions are x8, which do not have tests.
+#endif  // HAVE_SSE4_1
+
+#if HAVE_AVX2
+const SadMxNParam avx2_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64_avx2, -1),
+  make_tuple(64, 32, &vpx_sad64x32_avx2, -1),
+  make_tuple(32, 64, &vpx_sad32x64_avx2, -1),
+  make_tuple(32, 32, &vpx_sad32x32_avx2, -1),
+  make_tuple(32, 16, &vpx_sad32x16_avx2, -1),
+};
+INSTANTIATE_TEST_CASE_P(AVX2, SADTest, ::testing::ValuesIn(avx2_tests));
+
+const SadMxNAvgParam avg_avx2_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64_avg_avx2, -1),
+  make_tuple(64, 32, &vpx_sad64x32_avg_avx2, -1),
+  make_tuple(32, 64, &vpx_sad32x64_avg_avx2, -1),
+  make_tuple(32, 32, &vpx_sad32x32_avg_avx2, -1),
+  make_tuple(32, 16, &vpx_sad32x16_avg_avx2, -1),
+};
+INSTANTIATE_TEST_CASE_P(AVX2, SADavgTest, ::testing::ValuesIn(avg_avx2_tests));
+
+const SadMxNx4Param x4d_avx2_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64x4d_avx2, -1),
+  make_tuple(32, 32, &vpx_sad32x32x4d_avx2, -1),
+};
+INSTANTIATE_TEST_CASE_P(AVX2, SADx4Test, ::testing::ValuesIn(x4d_avx2_tests));
+#endif  // HAVE_AVX2
+
+//------------------------------------------------------------------------------
+// MIPS functions
+#if HAVE_MSA
+const SadMxNParam msa_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64_msa, -1),
+  make_tuple(64, 32, &vpx_sad64x32_msa, -1),
+  make_tuple(32, 64, &vpx_sad32x64_msa, -1),
+  make_tuple(32, 32, &vpx_sad32x32_msa, -1),
+  make_tuple(32, 16, &vpx_sad32x16_msa, -1),
+  make_tuple(16, 32, &vpx_sad16x32_msa, -1),
+  make_tuple(16, 16, &vpx_sad16x16_msa, -1),
+  make_tuple(16, 8, &vpx_sad16x8_msa, -1),
+  make_tuple(8, 16, &vpx_sad8x16_msa, -1),
+  make_tuple(8, 8, &vpx_sad8x8_msa, -1),
+  make_tuple(8, 4, &vpx_sad8x4_msa, -1),
+  make_tuple(4, 8, &vpx_sad4x8_msa, -1),
+  make_tuple(4, 4, &vpx_sad4x4_msa, -1),
+};
+INSTANTIATE_TEST_CASE_P(MSA, SADTest, ::testing::ValuesIn(msa_tests));
+
+const SadMxNAvgParam avg_msa_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64_avg_msa, -1),
+  make_tuple(64, 32, &vpx_sad64x32_avg_msa, -1),
+  make_tuple(32, 64, &vpx_sad32x64_avg_msa, -1),
+  make_tuple(32, 32, &vpx_sad32x32_avg_msa, -1),
+  make_tuple(32, 16, &vpx_sad32x16_avg_msa, -1),
+  make_tuple(16, 32, &vpx_sad16x32_avg_msa, -1),
+  make_tuple(16, 16, &vpx_sad16x16_avg_msa, -1),
+  make_tuple(16, 8, &vpx_sad16x8_avg_msa, -1),
+  make_tuple(8, 16, &vpx_sad8x16_avg_msa, -1),
+  make_tuple(8, 8, &vpx_sad8x8_avg_msa, -1),
+  make_tuple(8, 4, &vpx_sad8x4_avg_msa, -1),
+  make_tuple(4, 8, &vpx_sad4x8_avg_msa, -1),
+  make_tuple(4, 4, &vpx_sad4x4_avg_msa, -1),
+};
+INSTANTIATE_TEST_CASE_P(MSA, SADavgTest, ::testing::ValuesIn(avg_msa_tests));
+
+const SadMxNx4Param x4d_msa_tests[] = {
+  make_tuple(64, 64, &vpx_sad64x64x4d_msa, -1),
+  make_tuple(64, 32, &vpx_sad64x32x4d_msa, -1),
+  make_tuple(32, 64, &vpx_sad32x64x4d_msa, -1),
+  make_tuple(32, 32, &vpx_sad32x32x4d_msa, -1),
+  make_tuple(32, 16, &vpx_sad32x16x4d_msa, -1),
+  make_tuple(16, 32, &vpx_sad16x32x4d_msa, -1),
+  make_tuple(16, 16, &vpx_sad16x16x4d_msa, -1),
+  make_tuple(16, 8, &vpx_sad16x8x4d_msa, -1),
+  make_tuple(8, 16, &vpx_sad8x16x4d_msa, -1),
+  make_tuple(8, 8, &vpx_sad8x8x4d_msa, -1),
+  make_tuple(8, 4, &vpx_sad8x4x4d_msa, -1),
+  make_tuple(4, 8, &vpx_sad4x8x4d_msa, -1),
+  make_tuple(4, 4, &vpx_sad4x4x4d_msa, -1),
+};
+INSTANTIATE_TEST_CASE_P(MSA, SADx4Test, ::testing::ValuesIn(x4d_msa_tests));
+#endif  // HAVE_MSA
+
+}  // namespace
diff --git a/libvpx/libvpx/test/set_maps.sh b/libvpx/libvpx/test/set_maps.sh
new file mode 100755
index 0000000..e7c8d43
--- /dev/null
+++ b/libvpx/libvpx/test/set_maps.sh
@@ -0,0 +1,59 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx set_maps example. To add new tests to this file,
+##  do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to set_maps_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required, and set_maps must exist in
+# $LIBVPX_BIN_PATH.
+set_maps_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+  if [ -z "$(vpx_tool_path set_maps)" ]; then
+    elog "set_maps not found. It must exist in LIBVPX_BIN_PATH or its parent."
+    return 1
+  fi
+}
+
+# Runs set_maps using the codec specified by $1.
+set_maps() {
+  local encoder="$(vpx_tool_path set_maps)"
+  local codec="$1"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/set_maps_${codec}.ivf"
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+set_maps_vp8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    set_maps vp8 || return 1
+  fi
+}
+
+set_maps_vp9() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    set_maps vp9 || return 1
+  fi
+}
+
+set_maps_tests="set_maps_vp8
+                set_maps_vp9"
+
+run_tests set_maps_verify_environment "${set_maps_tests}"
diff --git a/libvpx/libvpx/test/set_roi.cc b/libvpx/libvpx/test/set_roi.cc
new file mode 100644
index 0000000..fea8cca
--- /dev/null
+++ b/libvpx/libvpx/test/set_roi.cc
@@ -0,0 +1,184 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <math.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/types.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/acm_random.h"
+#include "vp8/encoder/onyx_int.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+
+TEST(VP8RoiMapTest, ParameterCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int delta_q[MAX_MB_SEGMENTS] = { -2, -25, 0, 31 };
+  int delta_lf[MAX_MB_SEGMENTS] = { -2, -25, 0, 31 };
+  unsigned int threshold[MAX_MB_SEGMENTS] = { 0, 100, 200, 300 };
+
+  const int internalq_trans[] = {
+    0,   1,  2,  3,  4,  5,  7,  8,
+    9,  10, 12, 13, 15, 17, 18, 19,
+    20,  21, 23, 24, 25, 26, 27, 28,
+    29,  30, 31, 33, 35, 37, 39, 41,
+    43,  45, 47, 49, 51, 53, 55, 57,
+    59,  61, 64, 67, 70, 73, 76, 79,
+    82,  85, 88, 91, 94, 97, 100, 103,
+    106, 109, 112, 115, 118, 121, 124, 127,
+  };
+
+  // Initialize elements of cpi with valid defaults.
+  VP8_COMP cpi;
+  cpi.mb.e_mbd.mb_segement_abs_delta = SEGMENT_DELTADATA;
+  cpi.cyclic_refresh_mode_enabled = 0;
+  cpi.mb.e_mbd.segmentation_enabled = 0;
+  cpi.mb.e_mbd.update_mb_segmentation_map = 0;
+  cpi.mb.e_mbd.update_mb_segmentation_data = 0;
+  cpi.common.mb_rows = 240 >> 4;
+  cpi.common.mb_cols = 320 >> 4;
+  const int mbs = (cpi.common.mb_rows * cpi.common.mb_cols);
+  memset(cpi.segment_feature_data, 0, sizeof(cpi.segment_feature_data));
+
+  // Segment map
+  cpi.segmentation_map = reinterpret_cast<unsigned char *>(vpx_calloc(mbs, 1));
+
+  // Allocate memory for the source memory map.
+  unsigned char *roi_map =
+    reinterpret_cast<unsigned char *>(vpx_calloc(mbs, 1));
+  memset(&roi_map[mbs >> 2], 1, (mbs >> 2));
+  memset(&roi_map[mbs >> 1], 2, (mbs >> 2));
+  memset(&roi_map[mbs -(mbs >> 2)], 3, (mbs >> 2));
+
+  // Do a test call with valid parameters.
+  int roi_retval = vp8_set_roimap(&cpi, roi_map, cpi.common.mb_rows,
+                                  cpi.common.mb_cols, delta_q, delta_lf,
+                                  threshold);
+  EXPECT_EQ(0, roi_retval)
+        << "vp8_set_roimap roi failed with default test parameters";
+
+  // Check that the values in the cpi structure get set as expected.
+  if (roi_retval == 0) {
+    // Check that the segment map got set.
+    const int mapcompare = memcmp(roi_map, cpi.segmentation_map, mbs);
+    EXPECT_EQ(0, mapcompare) << "segment map error";
+
+    // Check the q deltas (note the need to translate into
+    // the interanl range of 0-127.
+    for (int i = 0; i < MAX_MB_SEGMENTS; ++i) {
+      const int transq = internalq_trans[abs(delta_q[i])];
+      if (abs(cpi.segment_feature_data[MB_LVL_ALT_Q][i]) != transq) {
+          EXPECT_EQ(transq, cpi.segment_feature_data[MB_LVL_ALT_Q][i])
+                    << "segment delta_q  error";
+          break;
+      }
+    }
+
+    // Check the loop filter deltas
+    for (int i = 0; i < MAX_MB_SEGMENTS; ++i) {
+      if (cpi.segment_feature_data[MB_LVL_ALT_LF][i] != delta_lf[i]) {
+        EXPECT_EQ(delta_lf[i], cpi.segment_feature_data[MB_LVL_ALT_LF][i])
+                  << "segment delta_lf error";
+        break;
+      }
+    }
+
+    // Check the breakout thresholds
+    for (int i = 0; i < MAX_MB_SEGMENTS; ++i) {
+      unsigned int breakout =
+        static_cast<unsigned int>(cpi.segment_encode_breakout[i]);
+
+      if (threshold[i] != breakout) {
+        EXPECT_EQ(threshold[i], breakout)
+                  << "breakout threshold error";
+        break;
+      }
+    }
+
+    // Segmentation, and segmentation update flages should be set.
+    EXPECT_EQ(1, cpi.mb.e_mbd.segmentation_enabled)
+              << "segmentation_enabled error";
+    EXPECT_EQ(1, cpi.mb.e_mbd.update_mb_segmentation_map)
+              << "update_mb_segmentation_map error";
+    EXPECT_EQ(1, cpi.mb.e_mbd.update_mb_segmentation_data)
+              << "update_mb_segmentation_data error";
+
+
+    // Try a range of delta q and lf parameters (some legal, some not)
+    for (int i = 0; i < 1000; ++i) {
+      int rand_deltas[4];
+      int deltas_valid;
+      rand_deltas[0] = rnd(160) - 80;
+      rand_deltas[1] = rnd(160) - 80;
+      rand_deltas[2] = rnd(160) - 80;
+      rand_deltas[3] = rnd(160) - 80;
+
+      deltas_valid = ((abs(rand_deltas[0]) <= 63) &&
+                      (abs(rand_deltas[1]) <= 63) &&
+                      (abs(rand_deltas[2]) <= 63) &&
+                      (abs(rand_deltas[3]) <= 63)) ? 0 : -1;
+
+      // Test with random delta q values.
+      roi_retval = vp8_set_roimap(&cpi, roi_map, cpi.common.mb_rows,
+                                  cpi.common.mb_cols, rand_deltas,
+                                  delta_lf, threshold);
+      EXPECT_EQ(deltas_valid, roi_retval) << "dq range check error";
+
+      // One delta_q error shown at a time
+      if (deltas_valid != roi_retval)
+        break;
+
+      // Test with random loop filter values.
+      roi_retval = vp8_set_roimap(&cpi, roi_map, cpi.common.mb_rows,
+                                  cpi.common.mb_cols, delta_q,
+                                  rand_deltas, threshold);
+      EXPECT_EQ(deltas_valid, roi_retval) << "dlf range check error";
+
+      // One delta loop filter error shown at a time
+      if (deltas_valid != roi_retval)
+        break;
+    }
+
+    // Test that we report and error if cyclic refresh is enabled.
+    cpi.cyclic_refresh_mode_enabled = 1;
+    roi_retval = vp8_set_roimap(&cpi, roi_map, cpi.common.mb_rows,
+                                cpi.common.mb_cols, delta_q,
+                                delta_lf, threshold);
+    EXPECT_EQ(-1, roi_retval) << "cyclic refresh check error";
+    cpi.cyclic_refresh_mode_enabled = 0;
+
+    // Test invalid number of rows or colums.
+    roi_retval = vp8_set_roimap(&cpi, roi_map, cpi.common.mb_rows + 1,
+                                cpi.common.mb_cols, delta_q,
+                                delta_lf, threshold);
+    EXPECT_EQ(-1, roi_retval) << "MB rows bounds check error";
+
+    roi_retval = vp8_set_roimap(&cpi, roi_map, cpi.common.mb_rows,
+                                cpi.common.mb_cols - 1, delta_q,
+                                delta_lf, threshold);
+    EXPECT_EQ(-1, roi_retval) << "MB cols bounds check error";
+  }
+
+  // Free allocated memory
+  if (cpi.segmentation_map)
+    vpx_free(cpi.segmentation_map);
+  if (roi_map)
+    vpx_free(roi_map);
+};
+
+}  // namespace
diff --git a/libvpx/libvpx/test/simple_decoder.sh b/libvpx/libvpx/test/simple_decoder.sh
new file mode 100755
index 0000000..7eeaf71
--- /dev/null
+++ b/libvpx/libvpx/test/simple_decoder.sh
@@ -0,0 +1,61 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx simple_decoder example code. To add new tests to
+##  this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to simple_decoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+#   $VP8_IVF_FILE and $VP9_IVF_FILE are required.
+simple_decoder_verify_environment() {
+  if [ ! -e "${VP8_IVF_FILE}" ] || [ ! -e "${VP9_IVF_FILE}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs simple_decoder using $1 as input file. $2 is the codec name, and is used
+# solely to name the output file.
+simple_decoder() {
+  local decoder="${LIBVPX_BIN_PATH}/simple_decoder${VPX_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/simple_decoder_${codec}.raw"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+simple_decoder_vp8() {
+  if [ "$(vp8_decode_available)" = "yes" ]; then
+    simple_decoder "${VP8_IVF_FILE}" vp8 || return 1
+  fi
+}
+
+simple_decoder_vp9() {
+  if [ "$(vp9_decode_available)" = "yes" ]; then
+    simple_decoder "${VP9_IVF_FILE}" vp9 || return 1
+  fi
+}
+
+simple_decoder_tests="simple_decoder_vp8
+                      simple_decoder_vp9"
+
+run_tests simple_decoder_verify_environment "${simple_decoder_tests}"
diff --git a/libvpx/libvpx/test/simple_encoder.sh b/libvpx/libvpx/test/simple_encoder.sh
new file mode 100755
index 0000000..ee633ae
--- /dev/null
+++ b/libvpx/libvpx/test/simple_encoder.sh
@@ -0,0 +1,59 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx simple_encoder example. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to simple_encoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+simple_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs simple_encoder using the codec specified by $1 with a frame limit of 100.
+simple_encoder() {
+  local encoder="${LIBVPX_BIN_PATH}/simple_encoder${VPX_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/simple_encoder_${codec}.ivf"
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" 9999 0 100 \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+simple_encoder_vp8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    simple_encoder vp8 || return 1
+  fi
+}
+
+simple_encoder_vp9() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    simple_encoder vp9 || return 1
+  fi
+}
+
+simple_encoder_tests="simple_encoder_vp8
+                      simple_encoder_vp9"
+
+run_tests simple_encoder_verify_environment "${simple_encoder_tests}"
diff --git a/libvpx/libvpx/test/sixtap_predict_test.cc b/libvpx/libvpx/test/sixtap_predict_test.cc
new file mode 100644
index 0000000..304a148
--- /dev/null
+++ b/libvpx/libvpx/test/sixtap_predict_test.cc
@@ -0,0 +1,233 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vp8_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+
+typedef void (*SixtapPredictFunc)(uint8_t *src_ptr,
+                                  int src_pixels_per_line,
+                                  int xoffset,
+                                  int yoffset,
+                                  uint8_t *dst_ptr,
+                                  int dst_pitch);
+
+typedef std::tr1::tuple<int, int, SixtapPredictFunc> SixtapPredictParam;
+
+class SixtapPredictTest
+    : public ::testing::TestWithParam<SixtapPredictParam> {
+ public:
+  static void SetUpTestCase() {
+    src_ = reinterpret_cast<uint8_t*>(vpx_memalign(kDataAlignment, kSrcSize));
+    dst_ = reinterpret_cast<uint8_t*>(vpx_memalign(kDataAlignment, kDstSize));
+    dst_c_ = reinterpret_cast<uint8_t*>(vpx_memalign(kDataAlignment, kDstSize));
+  }
+
+  static void TearDownTestCase() {
+    vpx_free(src_);
+    src_ = NULL;
+    vpx_free(dst_);
+    dst_ = NULL;
+    vpx_free(dst_c_);
+    dst_c_ = NULL;
+  }
+
+  virtual void TearDown() {
+    libvpx_test::ClearSystemState();
+  }
+
+ protected:
+  // Make test arrays big enough for 16x16 functions. Six-tap filters
+  // need 5 extra pixels outside of the macroblock.
+  static const int kSrcStride = 21;
+  static const int kDstStride = 16;
+  static const int kDataAlignment = 16;
+  static const int kSrcSize = kSrcStride * kSrcStride + 1;
+  static const int kDstSize = kDstStride * kDstStride;
+
+  virtual void SetUp() {
+    width_ = GET_PARAM(0);
+    height_ = GET_PARAM(1);
+    sixtap_predict_ = GET_PARAM(2);
+    memset(src_, 0, kSrcSize);
+    memset(dst_, 0, kDstSize);
+    memset(dst_c_, 0, kDstSize);
+  }
+
+  int width_;
+  int height_;
+  SixtapPredictFunc sixtap_predict_;
+  // The src stores the macroblock we will filter on, and makes it 1 byte larger
+  // in order to test unaligned access. The result is stored in dst and dst_c(c
+  // reference code result).
+  static uint8_t* src_;
+  static uint8_t* dst_;
+  static uint8_t* dst_c_;
+};
+
+uint8_t* SixtapPredictTest::src_ = NULL;
+uint8_t* SixtapPredictTest::dst_ = NULL;
+uint8_t* SixtapPredictTest::dst_c_ = NULL;
+
+TEST_P(SixtapPredictTest, TestWithPresetData) {
+  // Test input
+  static const uint8_t test_data[kSrcSize] = {
+    216, 184, 4, 191, 82, 92, 41, 0, 1, 226, 236, 172, 20, 182, 42, 226, 177,
+    79, 94, 77, 179, 203, 206, 198, 22, 192, 19, 75, 17, 192, 44, 233, 120,
+    48, 168, 203, 141, 210, 203, 143, 180, 184, 59, 201, 110, 102, 171, 32,
+    182, 10, 109, 105, 213, 60, 47, 236, 253, 67, 55, 14, 3, 99, 247, 124,
+    148, 159, 71, 34, 114, 19, 177, 38, 203, 237, 239, 58, 83, 155, 91, 10,
+    166, 201, 115, 124, 5, 163, 104, 2, 231, 160, 16, 234, 4, 8, 103, 153,
+    167, 174, 187, 26, 193, 109, 64, 141, 90, 48, 200, 174, 204, 36, 184,
+    114, 237, 43, 238, 242, 207, 86, 245, 182, 247, 6, 161, 251, 14, 8, 148,
+    182, 182, 79, 208, 120, 188, 17, 6, 23, 65, 206, 197, 13, 242, 126, 128,
+    224, 170, 110, 211, 121, 197, 200, 47, 188, 207, 208, 184, 221, 216, 76,
+    148, 143, 156, 100, 8, 89, 117, 14, 112, 183, 221, 54, 197, 208, 180, 69,
+    176, 94, 180, 131, 215, 121, 76, 7, 54, 28, 216, 238, 249, 176, 58, 142,
+    64, 215, 242, 72, 49, 104, 87, 161, 32, 52, 216, 230, 4, 141, 44, 181,
+    235, 224, 57, 195, 89, 134, 203, 144, 162, 163, 126, 156, 84, 185, 42,
+    148, 145, 29, 221, 194, 134, 52, 100, 166, 105, 60, 140, 110, 201, 184,
+    35, 181, 153, 93, 121, 243, 227, 68, 131, 134, 232, 2, 35, 60, 187, 77,
+    209, 76, 106, 174, 15, 241, 227, 115, 151, 77, 175, 36, 187, 121, 221,
+    223, 47, 118, 61, 168, 105, 32, 237, 236, 167, 213, 238, 202, 17, 170,
+    24, 226, 247, 131, 145, 6, 116, 117, 121, 11, 194, 41, 48, 126, 162, 13,
+    93, 209, 131, 154, 122, 237, 187, 103, 217, 99, 60, 200, 45, 78, 115, 69,
+    49, 106, 200, 194, 112, 60, 56, 234, 72, 251, 19, 120, 121, 182, 134, 215,
+    135, 10, 114, 2, 247, 46, 105, 209, 145, 165, 153, 191, 243, 12, 5, 36,
+    119, 206, 231, 231, 11, 32, 209, 83, 27, 229, 204, 149, 155, 83, 109, 35,
+    93, 223, 37, 84, 14, 142, 37, 160, 52, 191, 96, 40, 204, 101, 77, 67, 52,
+    53, 43, 63, 85, 253, 147, 113, 226, 96, 6, 125, 179, 115, 161, 17, 83,
+    198, 101, 98, 85, 139, 3, 137, 75, 99, 178, 23, 201, 255, 91, 253, 52,
+    134, 60, 138, 131, 208, 251, 101, 48, 2, 227, 228, 118, 132, 245, 202,
+    75, 91, 44, 160, 231, 47, 41, 50, 147, 220, 74, 92, 219, 165, 89, 16
+  };
+
+  // Expected result
+  static const uint8_t expected_dst[kDstSize] = {
+    117, 102, 74, 135, 42, 98, 175, 206, 70, 73, 222, 197, 50, 24, 39, 49, 38,
+    105, 90, 47, 169, 40, 171, 215, 200, 73, 109, 141, 53, 85, 177, 164, 79,
+    208, 124, 89, 212, 18, 81, 145, 151, 164, 217, 153, 91, 154, 102, 102,
+    159, 75, 164, 152, 136, 51, 213, 219, 186, 116, 193, 224, 186, 36, 231,
+    208, 84, 211, 155, 167, 35, 59, 42, 76, 216, 149, 73, 201, 78, 149, 184,
+    100, 96, 196, 189, 198, 188, 235, 195, 117, 129, 120, 129, 49, 25, 133,
+    113, 69, 221, 114, 70, 143, 99, 157, 108, 189, 140, 78, 6, 55, 65, 240,
+    255, 245, 184, 72, 90, 100, 116, 131, 39, 60, 234, 167, 33, 160, 88, 185,
+    200, 157, 159, 176, 127, 151, 138, 102, 168, 106, 170, 86, 82, 219, 189,
+    76, 33, 115, 197, 106, 96, 198, 136, 97, 141, 237, 151, 98, 137, 191,
+    185, 2, 57, 95, 142, 91, 255, 185, 97, 137, 76, 162, 94, 173, 131, 193,
+    161, 81, 106, 72, 135, 222, 234, 137, 66, 137, 106, 243, 210, 147, 95,
+    15, 137, 110, 85, 66, 16, 96, 167, 147, 150, 173, 203, 140, 118, 196,
+    84, 147, 160, 19, 95, 101, 123, 74, 132, 202, 82, 166, 12, 131, 166,
+    189, 170, 159, 85, 79, 66, 57, 152, 132, 203, 194, 0, 1, 56, 146, 180,
+    224, 156, 28, 83, 181, 79, 76, 80, 46, 160, 175, 59, 106, 43, 87, 75,
+    136, 85, 189, 46, 71, 200, 90
+  };
+
+  uint8_t *src = const_cast<uint8_t*>(test_data);
+
+  ASM_REGISTER_STATE_CHECK(
+      sixtap_predict_(&src[kSrcStride * 2 + 2 + 1], kSrcStride,
+                      2, 2, dst_, kDstStride));
+
+  for (int i = 0; i < height_; ++i)
+    for (int j = 0; j < width_; ++j)
+      ASSERT_EQ(expected_dst[i * kDstStride + j], dst_[i * kDstStride + j])
+          << "i==" << (i * width_ + j);
+}
+
+using libvpx_test::ACMRandom;
+
+TEST_P(SixtapPredictTest, TestWithRandomData) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  for (int i = 0; i < kSrcSize; ++i)
+    src_[i] = rnd.Rand8();
+
+  // Run tests for all possible offsets.
+  for (int xoffset = 0; xoffset < 8; ++xoffset) {
+    for (int yoffset = 0; yoffset < 8; ++yoffset) {
+      // Call c reference function.
+      // Move start point to next pixel to test if the function reads
+      // unaligned data correctly.
+      vp8_sixtap_predict16x16_c(&src_[kSrcStride * 2 + 2 + 1], kSrcStride,
+                                xoffset, yoffset, dst_c_, kDstStride);
+
+      // Run test.
+      ASM_REGISTER_STATE_CHECK(
+          sixtap_predict_(&src_[kSrcStride * 2 + 2 + 1], kSrcStride,
+                          xoffset, yoffset, dst_, kDstStride));
+
+      for (int i = 0; i < height_; ++i)
+        for (int j = 0; j < width_; ++j)
+          ASSERT_EQ(dst_c_[i * kDstStride + j], dst_[i * kDstStride + j])
+              << "i==" << (i * width_ + j);
+    }
+  }
+}
+
+using std::tr1::make_tuple;
+
+INSTANTIATE_TEST_CASE_P(
+    C, SixtapPredictTest, ::testing::Values(
+        make_tuple(16, 16, &vp8_sixtap_predict16x16_c),
+        make_tuple(8, 8, &vp8_sixtap_predict8x8_c),
+        make_tuple(8, 4, &vp8_sixtap_predict8x4_c),
+        make_tuple(4, 4, &vp8_sixtap_predict4x4_c)));
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+    NEON, SixtapPredictTest, ::testing::Values(
+        make_tuple(16, 16, &vp8_sixtap_predict16x16_neon),
+        make_tuple(8, 8, &vp8_sixtap_predict8x8_neon),
+        make_tuple(8, 4, &vp8_sixtap_predict8x4_neon)));
+#endif
+#if HAVE_MMX
+INSTANTIATE_TEST_CASE_P(
+    MMX, SixtapPredictTest, ::testing::Values(
+        make_tuple(16, 16, &vp8_sixtap_predict16x16_mmx),
+        make_tuple(8, 8, &vp8_sixtap_predict8x8_mmx),
+        make_tuple(8, 4, &vp8_sixtap_predict8x4_mmx),
+        make_tuple(4, 4, &vp8_sixtap_predict4x4_mmx)));
+#endif
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, SixtapPredictTest, ::testing::Values(
+        make_tuple(16, 16, &vp8_sixtap_predict16x16_sse2),
+        make_tuple(8, 8, &vp8_sixtap_predict8x8_sse2),
+        make_tuple(8, 4, &vp8_sixtap_predict8x4_sse2)));
+#endif
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, SixtapPredictTest, ::testing::Values(
+        make_tuple(16, 16, &vp8_sixtap_predict16x16_ssse3),
+        make_tuple(8, 8, &vp8_sixtap_predict8x8_ssse3),
+        make_tuple(8, 4, &vp8_sixtap_predict8x4_ssse3),
+        make_tuple(4, 4, &vp8_sixtap_predict4x4_ssse3)));
+#endif
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(
+    MSA, SixtapPredictTest, ::testing::Values(
+        make_tuple(16, 16, &vp8_sixtap_predict16x16_msa),
+        make_tuple(8, 8, &vp8_sixtap_predict8x8_msa),
+        make_tuple(8, 4, &vp8_sixtap_predict8x4_msa),
+        make_tuple(4, 4, &vp8_sixtap_predict4x4_msa)));
+#endif
+}  // namespace
diff --git a/libvpx/libvpx/test/superframe_test.cc b/libvpx/libvpx/test/superframe_test.cc
new file mode 100644
index 0000000..b07bcb2
--- /dev/null
+++ b/libvpx/libvpx/test/superframe_test.cc
@@ -0,0 +1,104 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <climits>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+const int kTestMode = 0;
+
+typedef std::tr1::tuple<libvpx_test::TestMode,int> SuperframeTestParam;
+
+class SuperframeTest : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWithParam<SuperframeTestParam> {
+ protected:
+  SuperframeTest() : EncoderTest(GET_PARAM(0)), modified_buf_(NULL),
+      last_sf_pts_(0) {}
+  virtual ~SuperframeTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    const SuperframeTestParam input = GET_PARAM(1);
+    const libvpx_test::TestMode mode = std::tr1::get<kTestMode>(input);
+    SetMode(mode);
+    sf_count_ = 0;
+    sf_count_max_ = INT_MAX;
+  }
+
+  virtual void TearDown() {
+    delete[] modified_buf_;
+  }
+
+  virtual void PreEncodeFrameHook(libvpx_test::VideoSource *video,
+                                  libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+    }
+  }
+
+  virtual const vpx_codec_cx_pkt_t * MutateEncoderOutputHook(
+      const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->kind != VPX_CODEC_CX_FRAME_PKT)
+      return pkt;
+
+    const uint8_t *buffer = reinterpret_cast<uint8_t*>(pkt->data.frame.buf);
+    const uint8_t marker = buffer[pkt->data.frame.sz - 1];
+    const int frames = (marker & 0x7) + 1;
+    const int mag = ((marker >> 3) & 3) + 1;
+    const unsigned int index_sz = 2 + mag * frames;
+    if ((marker & 0xe0) == 0xc0 &&
+        pkt->data.frame.sz >= index_sz &&
+        buffer[pkt->data.frame.sz - index_sz] == marker) {
+      // frame is a superframe. strip off the index.
+      if (modified_buf_)
+        delete[] modified_buf_;
+      modified_buf_ = new uint8_t[pkt->data.frame.sz - index_sz];
+      memcpy(modified_buf_, pkt->data.frame.buf,
+             pkt->data.frame.sz - index_sz);
+      modified_pkt_ = *pkt;
+      modified_pkt_.data.frame.buf = modified_buf_;
+      modified_pkt_.data.frame.sz -= index_sz;
+
+      sf_count_++;
+      last_sf_pts_ = pkt->data.frame.pts;
+      return &modified_pkt_;
+    }
+
+    // Make sure we do a few frames after the last SF
+    abort_ |= sf_count_ > sf_count_max_ &&
+              pkt->data.frame.pts - last_sf_pts_ >= 5;
+    return pkt;
+  }
+
+  int sf_count_;
+  int sf_count_max_;
+  vpx_codec_cx_pkt_t modified_pkt_;
+  uint8_t *modified_buf_;
+  vpx_codec_pts_t last_sf_pts_;
+};
+
+TEST_P(SuperframeTest, TestSuperframeIndexIsOptional) {
+  sf_count_max_ = 0;  // early exit on successful test.
+  cfg_.g_lag_in_frames = 25;
+
+  ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 40);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_EQ(sf_count_, 1);
+}
+
+VP9_INSTANTIATE_TEST_CASE(SuperframeTest, ::testing::Combine(
+    ::testing::Values(::libvpx_test::kTwoPassGood),
+    ::testing::Values(0)));
+}  // namespace
diff --git a/libvpx/libvpx/test/svc_test.cc b/libvpx/libvpx/test/svc_test.cc
new file mode 100644
index 0000000..b955cee
--- /dev/null
+++ b/libvpx/libvpx/test/svc_test.cc
@@ -0,0 +1,797 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/i420_video_source.h"
+
+#include "vp9/decoder/vp9_decoder.h"
+
+#include "vpx/svc_context.h"
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+
+namespace {
+
+using libvpx_test::CodecFactory;
+using libvpx_test::Decoder;
+using libvpx_test::DxDataIterator;
+using libvpx_test::VP9CodecFactory;
+
+class SvcTest : public ::testing::Test {
+ protected:
+  static const uint32_t kWidth = 352;
+  static const uint32_t kHeight = 288;
+
+  SvcTest()
+      : codec_iface_(0),
+        test_file_name_("hantro_collage_w352h288.yuv"),
+        codec_initialized_(false),
+        decoder_(0) {
+    memset(&svc_, 0, sizeof(svc_));
+    memset(&codec_, 0, sizeof(codec_));
+    memset(&codec_enc_, 0, sizeof(codec_enc_));
+  }
+
+  virtual ~SvcTest() {}
+
+  virtual void SetUp() {
+    svc_.log_level = SVC_LOG_DEBUG;
+    svc_.log_print = 0;
+
+    codec_iface_ = vpx_codec_vp9_cx();
+    const vpx_codec_err_t res =
+        vpx_codec_enc_config_default(codec_iface_, &codec_enc_, 0);
+    EXPECT_EQ(VPX_CODEC_OK, res);
+
+    codec_enc_.g_w = kWidth;
+    codec_enc_.g_h = kHeight;
+    codec_enc_.g_timebase.num = 1;
+    codec_enc_.g_timebase.den = 60;
+    codec_enc_.kf_min_dist = 100;
+    codec_enc_.kf_max_dist = 100;
+
+    vpx_codec_dec_cfg_t dec_cfg = vpx_codec_dec_cfg_t();
+    VP9CodecFactory codec_factory;
+    decoder_ = codec_factory.CreateDecoder(dec_cfg, 0);
+
+    tile_columns_ = 0;
+    tile_rows_ = 0;
+  }
+
+  virtual void TearDown() {
+    ReleaseEncoder();
+    delete(decoder_);
+  }
+
+  void InitializeEncoder() {
+    const vpx_codec_err_t res =
+        vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+    EXPECT_EQ(VPX_CODEC_OK, res);
+    vpx_codec_control(&codec_, VP8E_SET_CPUUSED, 4);  // Make the test faster
+    vpx_codec_control(&codec_, VP9E_SET_TILE_COLUMNS, tile_columns_);
+    vpx_codec_control(&codec_, VP9E_SET_TILE_ROWS, tile_rows_);
+    codec_initialized_ = true;
+  }
+
+  void ReleaseEncoder() {
+    vpx_svc_release(&svc_);
+    if (codec_initialized_) vpx_codec_destroy(&codec_);
+    codec_initialized_ = false;
+  }
+
+  void GetStatsData(std::string *const stats_buf) {
+    vpx_codec_iter_t iter = NULL;
+    const vpx_codec_cx_pkt_t *cx_pkt;
+
+    while ((cx_pkt = vpx_codec_get_cx_data(&codec_, &iter)) != NULL) {
+      if (cx_pkt->kind == VPX_CODEC_STATS_PKT) {
+        EXPECT_GT(cx_pkt->data.twopass_stats.sz, 0U);
+        ASSERT_TRUE(cx_pkt->data.twopass_stats.buf != NULL);
+        stats_buf->append(static_cast<char*>(cx_pkt->data.twopass_stats.buf),
+                          cx_pkt->data.twopass_stats.sz);
+      }
+    }
+  }
+
+  void Pass1EncodeNFrames(const int n, const int layers,
+                          std::string *const stats_buf) {
+    vpx_codec_err_t res;
+
+    ASSERT_GT(n, 0);
+    ASSERT_GT(layers, 0);
+    svc_.spatial_layers = layers;
+    codec_enc_.g_pass = VPX_RC_FIRST_PASS;
+    InitializeEncoder();
+
+    libvpx_test::I420VideoSource video(test_file_name_,
+                                       codec_enc_.g_w, codec_enc_.g_h,
+                                       codec_enc_.g_timebase.den,
+                                       codec_enc_.g_timebase.num, 0, 30);
+    video.Begin();
+
+    for (int i = 0; i < n; ++i) {
+      res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
+                           video.duration(), VPX_DL_GOOD_QUALITY);
+      ASSERT_EQ(VPX_CODEC_OK, res);
+      GetStatsData(stats_buf);
+      video.Next();
+    }
+
+    // Flush encoder and test EOS packet.
+    res = vpx_svc_encode(&svc_, &codec_, NULL, video.pts(),
+                         video.duration(), VPX_DL_GOOD_QUALITY);
+    ASSERT_EQ(VPX_CODEC_OK, res);
+    GetStatsData(stats_buf);
+
+    ReleaseEncoder();
+  }
+
+  void StoreFrames(const size_t max_frame_received,
+                   struct vpx_fixed_buf *const outputs,
+                   size_t *const frame_received) {
+    vpx_codec_iter_t iter = NULL;
+    const vpx_codec_cx_pkt_t *cx_pkt;
+
+    while ((cx_pkt = vpx_codec_get_cx_data(&codec_, &iter)) != NULL) {
+      if (cx_pkt->kind == VPX_CODEC_CX_FRAME_PKT) {
+        const size_t frame_size = cx_pkt->data.frame.sz;
+
+        EXPECT_GT(frame_size, 0U);
+        ASSERT_TRUE(cx_pkt->data.frame.buf != NULL);
+        ASSERT_LT(*frame_received, max_frame_received);
+
+        if (*frame_received == 0)
+          EXPECT_EQ(1, !!(cx_pkt->data.frame.flags & VPX_FRAME_IS_KEY));
+
+        outputs[*frame_received].buf = malloc(frame_size + 16);
+        ASSERT_TRUE(outputs[*frame_received].buf != NULL);
+        memcpy(outputs[*frame_received].buf, cx_pkt->data.frame.buf,
+               frame_size);
+        outputs[*frame_received].sz = frame_size;
+        ++(*frame_received);
+      }
+    }
+  }
+
+  void Pass2EncodeNFrames(std::string *const stats_buf,
+                          const int n, const int layers,
+                          struct vpx_fixed_buf *const outputs) {
+    vpx_codec_err_t res;
+    size_t frame_received = 0;
+
+    ASSERT_TRUE(outputs != NULL);
+    ASSERT_GT(n, 0);
+    ASSERT_GT(layers, 0);
+    svc_.spatial_layers = layers;
+    codec_enc_.rc_target_bitrate = 500;
+    if (codec_enc_.g_pass == VPX_RC_LAST_PASS) {
+      ASSERT_TRUE(stats_buf != NULL);
+      ASSERT_GT(stats_buf->size(), 0U);
+      codec_enc_.rc_twopass_stats_in.buf = &(*stats_buf)[0];
+      codec_enc_.rc_twopass_stats_in.sz = stats_buf->size();
+    }
+    InitializeEncoder();
+
+    libvpx_test::I420VideoSource video(test_file_name_,
+                                       codec_enc_.g_w, codec_enc_.g_h,
+                                       codec_enc_.g_timebase.den,
+                                       codec_enc_.g_timebase.num, 0, 30);
+    video.Begin();
+
+    for (int i = 0; i < n; ++i) {
+      res = vpx_svc_encode(&svc_, &codec_, video.img(), video.pts(),
+                           video.duration(), VPX_DL_GOOD_QUALITY);
+      ASSERT_EQ(VPX_CODEC_OK, res);
+      StoreFrames(n, outputs, &frame_received);
+      video.Next();
+    }
+
+    // Flush encoder.
+    res = vpx_svc_encode(&svc_, &codec_, NULL, 0,
+                         video.duration(), VPX_DL_GOOD_QUALITY);
+    EXPECT_EQ(VPX_CODEC_OK, res);
+    StoreFrames(n, outputs, &frame_received);
+
+    EXPECT_EQ(frame_received, static_cast<size_t>(n));
+
+    ReleaseEncoder();
+  }
+
+  void DecodeNFrames(const struct vpx_fixed_buf *const inputs, const int n) {
+    int decoded_frames = 0;
+    int received_frames = 0;
+
+    ASSERT_TRUE(inputs != NULL);
+    ASSERT_GT(n, 0);
+
+    for (int i = 0; i < n; ++i) {
+      ASSERT_TRUE(inputs[i].buf != NULL);
+      ASSERT_GT(inputs[i].sz, 0U);
+      const vpx_codec_err_t res_dec =
+          decoder_->DecodeFrame(static_cast<const uint8_t *>(inputs[i].buf),
+                                inputs[i].sz);
+      ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder_->DecodeError();
+      ++decoded_frames;
+
+      DxDataIterator dec_iter = decoder_->GetDxData();
+      while (dec_iter.Next() != NULL) {
+        ++received_frames;
+      }
+    }
+    EXPECT_EQ(decoded_frames, n);
+    EXPECT_EQ(received_frames, n);
+  }
+
+  void DropEnhancementLayers(struct vpx_fixed_buf *const inputs,
+                             const int num_super_frames,
+                             const int remained_spatial_layers) {
+    ASSERT_TRUE(inputs != NULL);
+    ASSERT_GT(num_super_frames, 0);
+    ASSERT_GT(remained_spatial_layers, 0);
+
+    for (int i = 0; i < num_super_frames; ++i) {
+      uint32_t frame_sizes[8] = {0};
+      int frame_count = 0;
+      int frames_found = 0;
+      int frame;
+      ASSERT_TRUE(inputs[i].buf != NULL);
+      ASSERT_GT(inputs[i].sz, 0U);
+
+      vpx_codec_err_t res =
+          vp9_parse_superframe_index(static_cast<const uint8_t*>(inputs[i].buf),
+                                     inputs[i].sz, frame_sizes, &frame_count,
+                                     NULL, NULL);
+      ASSERT_EQ(VPX_CODEC_OK, res);
+
+      if (frame_count == 0) {
+        // There's no super frame but only a single frame.
+        ASSERT_EQ(1, remained_spatial_layers);
+      } else {
+        // Found a super frame.
+        uint8_t *frame_data = static_cast<uint8_t*>(inputs[i].buf);
+        uint8_t *frame_start = frame_data;
+        for (frame = 0; frame < frame_count; ++frame) {
+          // Looking for a visible frame.
+          if (frame_data[0] & 0x02) {
+            ++frames_found;
+            if (frames_found == remained_spatial_layers)
+              break;
+          }
+          frame_data += frame_sizes[frame];
+        }
+        ASSERT_LT(frame, frame_count) << "Couldn't find a visible frame. "
+            << "remained_spatial_layers: " << remained_spatial_layers
+            << "    super_frame: " << i;
+        if (frame == frame_count - 1)
+          continue;
+
+        frame_data += frame_sizes[frame];
+
+        // We need to add one more frame for multiple frame contexts.
+        uint8_t marker =
+            static_cast<const uint8_t*>(inputs[i].buf)[inputs[i].sz - 1];
+        const uint32_t mag = ((marker >> 3) & 0x3) + 1;
+        const size_t index_sz = 2 + mag * frame_count;
+        const size_t new_index_sz = 2 + mag * (frame + 1);
+        marker &= 0x0f8;
+        marker |= frame;
+
+        // Copy existing frame sizes.
+        memmove(frame_data + 1, frame_start + inputs[i].sz - index_sz + 1,
+                new_index_sz - 2);
+        // New marker.
+        frame_data[0] = marker;
+        frame_data += (mag * (frame + 1) + 1);
+
+        *frame_data++ = marker;
+        inputs[i].sz = frame_data - frame_start;
+      }
+    }
+  }
+
+  void FreeBitstreamBuffers(struct vpx_fixed_buf *const inputs, const int n) {
+    ASSERT_TRUE(inputs != NULL);
+    ASSERT_GT(n, 0);
+
+    for (int i = 0; i < n; ++i) {
+      free(inputs[i].buf);
+      inputs[i].buf = NULL;
+      inputs[i].sz = 0;
+    }
+  }
+
+  SvcContext svc_;
+  vpx_codec_ctx_t codec_;
+  struct vpx_codec_enc_cfg codec_enc_;
+  vpx_codec_iface_t *codec_iface_;
+  std::string test_file_name_;
+  bool codec_initialized_;
+  Decoder *decoder_;
+  int tile_columns_;
+  int tile_rows_;
+};
+
+TEST_F(SvcTest, SvcInit) {
+  // test missing parameters
+  vpx_codec_err_t res = vpx_svc_init(NULL, &codec_, codec_iface_, &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+  res = vpx_svc_init(&svc_, NULL, codec_iface_, &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+  res = vpx_svc_init(&svc_, &codec_, NULL, &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_init(&svc_, &codec_, codec_iface_, NULL);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  svc_.spatial_layers = 6;  // too many layers
+  res = vpx_svc_init(&svc_, &codec_, codec_iface_, &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  svc_.spatial_layers = 0;  // use default layers
+  InitializeEncoder();
+  EXPECT_EQ(VPX_SS_DEFAULT_LAYERS, svc_.spatial_layers);
+}
+
+TEST_F(SvcTest, InitTwoLayers) {
+  svc_.spatial_layers = 2;
+  InitializeEncoder();
+}
+
+TEST_F(SvcTest, InvalidOptions) {
+  vpx_codec_err_t res = vpx_svc_set_options(&svc_, NULL);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_set_options(&svc_, "not-an-option=1");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+}
+
+TEST_F(SvcTest, SetLayersOption) {
+  vpx_codec_err_t res = vpx_svc_set_options(&svc_, "spatial-layers=3");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  InitializeEncoder();
+  EXPECT_EQ(3, svc_.spatial_layers);
+}
+
+TEST_F(SvcTest, SetMultipleOptions) {
+  vpx_codec_err_t res =
+      vpx_svc_set_options(&svc_, "spatial-layers=2 scale-factors=1/3,2/3");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  InitializeEncoder();
+  EXPECT_EQ(2, svc_.spatial_layers);
+}
+
+TEST_F(SvcTest, SetScaleFactorsOption) {
+  svc_.spatial_layers = 2;
+  vpx_codec_err_t res =
+      vpx_svc_set_options(&svc_, "scale-factors=not-scale-factors");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_set_options(&svc_, "scale-factors=1/3, 3*3");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_set_options(&svc_, "scale-factors=1/3");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_set_options(&svc_, "scale-factors=1/3,2/3");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  InitializeEncoder();
+}
+
+TEST_F(SvcTest, SetQuantizersOption) {
+  svc_.spatial_layers = 2;
+  vpx_codec_err_t res = vpx_svc_set_options(&svc_, "max-quantizers=nothing");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_set_options(&svc_, "min-quantizers=nothing");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_set_options(&svc_, "max-quantizers=40");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_set_options(&svc_, "min-quantizers=40");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_set_options(&svc_, "max-quantizers=30,30 min-quantizers=40,40");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_set_options(&svc_, "max-quantizers=40,40 min-quantizers=30,30");
+  InitializeEncoder();
+}
+
+TEST_F(SvcTest, SetAutoAltRefOption) {
+  svc_.spatial_layers = 5;
+  vpx_codec_err_t res = vpx_svc_set_options(&svc_, "auto-alt-refs=none");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  res = vpx_svc_set_options(&svc_, "auto-alt-refs=1,1,1,1,0");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  vpx_svc_set_options(&svc_, "auto-alt-refs=0,1,1,1,0");
+  InitializeEncoder();
+}
+
+// Test that decoder can handle an SVC frame as the first frame in a sequence.
+TEST_F(SvcTest, OnePassEncodeOneFrame) {
+  codec_enc_.g_pass = VPX_RC_ONE_PASS;
+  vpx_fixed_buf output = {0};
+  Pass2EncodeNFrames(NULL, 1, 2, &output);
+  DecodeNFrames(&output, 1);
+  FreeBitstreamBuffers(&output, 1);
+}
+
+TEST_F(SvcTest, OnePassEncodeThreeFrames) {
+  codec_enc_.g_pass = VPX_RC_ONE_PASS;
+  codec_enc_.g_lag_in_frames = 0;
+  vpx_fixed_buf outputs[3];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(NULL, 3, 2, &outputs[0]);
+  DecodeNFrames(&outputs[0], 3);
+  FreeBitstreamBuffers(&outputs[0], 3);
+}
+
+TEST_F(SvcTest, TwoPassEncode10Frames) {
+  // First pass encode
+  std::string stats_buf;
+  Pass1EncodeNFrames(10, 2, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 2, &outputs[0]);
+  DecodeNFrames(&outputs[0], 10);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest, TwoPassEncode20FramesWithAltRef) {
+  // First pass encode
+  std::string stats_buf;
+  Pass1EncodeNFrames(20, 2, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1,1");
+  vpx_fixed_buf outputs[20];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 20, 2, &outputs[0]);
+  DecodeNFrames(&outputs[0], 20);
+  FreeBitstreamBuffers(&outputs[0], 20);
+}
+
+TEST_F(SvcTest, TwoPassEncode2SpatialLayersDecodeBaseLayerOnly) {
+  // First pass encode
+  std::string stats_buf;
+  Pass1EncodeNFrames(10, 2, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1,1");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 2, &outputs[0]);
+  DropEnhancementLayers(&outputs[0], 10, 1);
+  DecodeNFrames(&outputs[0], 10);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest, TwoPassEncode5SpatialLayersDecode54321Layers) {
+  // First pass encode
+  std::string stats_buf;
+  Pass1EncodeNFrames(10, 5, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=0,1,1,1,0");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 5, &outputs[0]);
+
+  DecodeNFrames(&outputs[0], 10);
+  DropEnhancementLayers(&outputs[0], 10, 4);
+  DecodeNFrames(&outputs[0], 10);
+  DropEnhancementLayers(&outputs[0], 10, 3);
+  DecodeNFrames(&outputs[0], 10);
+  DropEnhancementLayers(&outputs[0], 10, 2);
+  DecodeNFrames(&outputs[0], 10);
+  DropEnhancementLayers(&outputs[0], 10, 1);
+  DecodeNFrames(&outputs[0], 10);
+
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest, TwoPassEncode2SNRLayers) {
+  // First pass encode
+  std::string stats_buf;
+  vpx_svc_set_options(&svc_, "scale-factors=1/1,1/1");
+  Pass1EncodeNFrames(20, 2, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  vpx_svc_set_options(&svc_,
+                      "auto-alt-refs=1,1 scale-factors=1/1,1/1");
+  vpx_fixed_buf outputs[20];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 20, 2, &outputs[0]);
+  DecodeNFrames(&outputs[0], 20);
+  FreeBitstreamBuffers(&outputs[0], 20);
+}
+
+TEST_F(SvcTest, TwoPassEncode3SNRLayersDecode321Layers) {
+  // First pass encode
+  std::string stats_buf;
+  vpx_svc_set_options(&svc_, "scale-factors=1/1,1/1,1/1");
+  Pass1EncodeNFrames(20, 3, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  vpx_svc_set_options(&svc_,
+                      "auto-alt-refs=1,1,1 scale-factors=1/1,1/1,1/1");
+  vpx_fixed_buf outputs[20];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 20, 3, &outputs[0]);
+  DecodeNFrames(&outputs[0], 20);
+  DropEnhancementLayers(&outputs[0], 20, 2);
+  DecodeNFrames(&outputs[0], 20);
+  DropEnhancementLayers(&outputs[0], 20, 1);
+  DecodeNFrames(&outputs[0], 20);
+
+  FreeBitstreamBuffers(&outputs[0], 20);
+}
+
+TEST_F(SvcTest, SetMultipleFrameContextsOption) {
+  svc_.spatial_layers = 5;
+  vpx_codec_err_t res =
+      vpx_svc_set_options(&svc_, "multi-frame-contexts=1");
+  EXPECT_EQ(VPX_CODEC_OK, res);
+  res = vpx_svc_init(&svc_, &codec_, vpx_codec_vp9_cx(), &codec_enc_);
+  EXPECT_EQ(VPX_CODEC_INVALID_PARAM, res);
+
+  svc_.spatial_layers = 2;
+  res = vpx_svc_set_options(&svc_, "multi-frame-contexts=1");
+  InitializeEncoder();
+}
+
+TEST_F(SvcTest, TwoPassEncode2SpatialLayersWithMultipleFrameContexts) {
+  // First pass encode
+  std::string stats_buf;
+  Pass1EncodeNFrames(10, 2, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  codec_enc_.g_error_resilient = 0;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1,1 multi-frame-contexts=1");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 2, &outputs[0]);
+  DecodeNFrames(&outputs[0], 10);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest,
+       TwoPassEncode2SpatialLayersWithMultipleFrameContextsDecodeBaselayer) {
+  // First pass encode
+  std::string stats_buf;
+  Pass1EncodeNFrames(10, 2, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  codec_enc_.g_error_resilient = 0;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1,1 multi-frame-contexts=1");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 2, &outputs[0]);
+  DropEnhancementLayers(&outputs[0], 10, 1);
+  DecodeNFrames(&outputs[0], 10);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest, TwoPassEncode2SNRLayersWithMultipleFrameContexts) {
+  // First pass encode
+  std::string stats_buf;
+  vpx_svc_set_options(&svc_, "scale-factors=1/1,1/1");
+  Pass1EncodeNFrames(10, 2, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  codec_enc_.g_error_resilient = 0;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1,1 scale-factors=1/1,1/1 "
+                      "multi-frame-contexts=1");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 2, &outputs[0]);
+  DecodeNFrames(&outputs[0], 10);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest,
+       TwoPassEncode3SNRLayersWithMultipleFrameContextsDecode321Layer) {
+  // First pass encode
+  std::string stats_buf;
+  vpx_svc_set_options(&svc_, "scale-factors=1/1,1/1,1/1");
+  Pass1EncodeNFrames(10, 3, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  codec_enc_.g_error_resilient = 0;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1,1,1 scale-factors=1/1,1/1,1/1 "
+                      "multi-frame-contexts=1");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 3, &outputs[0]);
+
+  DecodeNFrames(&outputs[0], 10);
+  DropEnhancementLayers(&outputs[0], 10, 2);
+  DecodeNFrames(&outputs[0], 10);
+  DropEnhancementLayers(&outputs[0], 10, 1);
+  DecodeNFrames(&outputs[0], 10);
+
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest, TwoPassEncode2TemporalLayers) {
+  // First pass encode
+  std::string stats_buf;
+  vpx_svc_set_options(&svc_, "scale-factors=1/1");
+  svc_.temporal_layers = 2;
+  Pass1EncodeNFrames(10, 1, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  svc_.temporal_layers = 2;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1 scale-factors=1/1");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 1, &outputs[0]);
+  DecodeNFrames(&outputs[0], 10);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest, TwoPassEncode2TemporalLayersWithMultipleFrameContexts) {
+  // First pass encode
+  std::string stats_buf;
+  vpx_svc_set_options(&svc_, "scale-factors=1/1");
+  svc_.temporal_layers = 2;
+  Pass1EncodeNFrames(10, 1, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  svc_.temporal_layers = 2;
+  codec_enc_.g_error_resilient = 0;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1 scale-factors=1/1 "
+                      "multi-frame-contexts=1");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 1, &outputs[0]);
+  DecodeNFrames(&outputs[0], 10);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest, TwoPassEncode2TemporalLayersDecodeBaseLayer) {
+  // First pass encode
+  std::string stats_buf;
+  vpx_svc_set_options(&svc_, "scale-factors=1/1");
+  svc_.temporal_layers = 2;
+  Pass1EncodeNFrames(10, 1, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  svc_.temporal_layers = 2;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1 scale-factors=1/1");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 1, &outputs[0]);
+
+  vpx_fixed_buf base_layer[5];
+  for (int i = 0; i < 5; ++i)
+    base_layer[i] = outputs[i * 2];
+
+  DecodeNFrames(&base_layer[0], 5);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest,
+       TwoPassEncode2TemporalLayersWithMultipleFrameContextsDecodeBaseLayer) {
+  // First pass encode
+  std::string stats_buf;
+  vpx_svc_set_options(&svc_, "scale-factors=1/1");
+  svc_.temporal_layers = 2;
+  Pass1EncodeNFrames(10, 1, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  svc_.temporal_layers = 2;
+  codec_enc_.g_error_resilient = 0;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1 scale-factors=1/1 "
+                      "multi-frame-contexts=1");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 1, &outputs[0]);
+
+  vpx_fixed_buf base_layer[5];
+  for (int i = 0; i < 5; ++i)
+    base_layer[i] = outputs[i * 2];
+
+  DecodeNFrames(&base_layer[0], 5);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest, TwoPassEncode2TemporalLayersWithTiles) {
+  // First pass encode
+  std::string stats_buf;
+  vpx_svc_set_options(&svc_, "scale-factors=1/1");
+  svc_.temporal_layers = 2;
+  Pass1EncodeNFrames(10, 1, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  svc_.temporal_layers = 2;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1 scale-factors=1/1");
+  codec_enc_.g_w = 704;
+  codec_enc_.g_h = 144;
+  tile_columns_ = 1;
+  tile_rows_ = 1;
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 1, &outputs[0]);
+  DecodeNFrames(&outputs[0], 10);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+TEST_F(SvcTest,
+       TwoPassEncode2TemporalLayersWithMultipleFrameContextsAndTiles) {
+  // First pass encode
+  std::string stats_buf;
+  vpx_svc_set_options(&svc_, "scale-factors=1/1");
+  svc_.temporal_layers = 2;
+  Pass1EncodeNFrames(10, 1, &stats_buf);
+
+  // Second pass encode
+  codec_enc_.g_pass = VPX_RC_LAST_PASS;
+  svc_.temporal_layers = 2;
+  codec_enc_.g_error_resilient = 0;
+  codec_enc_.g_w = 704;
+  codec_enc_.g_h = 144;
+  tile_columns_ = 1;
+  tile_rows_ = 1;
+  vpx_svc_set_options(&svc_, "auto-alt-refs=1 scale-factors=1/1 "
+                      "multi-frame-contexts=1");
+  vpx_fixed_buf outputs[10];
+  memset(&outputs[0], 0, sizeof(outputs));
+  Pass2EncodeNFrames(&stats_buf, 10, 1, &outputs[0]);
+  DecodeNFrames(&outputs[0], 10);
+  FreeBitstreamBuffers(&outputs[0], 10);
+}
+
+}  // namespace
diff --git a/libvpx/libvpx/test/test-data.mk b/libvpx/libvpx/test/test-data.mk
new file mode 100644
index 0000000..05a0885
--- /dev/null
+++ b/libvpx/libvpx/test/test-data.mk
@@ -0,0 +1,863 @@
+LIBVPX_TEST_SRCS-yes += test-data.mk
+
+# Encoder test source
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += hantro_collage_w352h288.yuv
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += hantro_odd.yuv
+
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_420.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_422.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_444.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_440.yuv
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_420.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_422.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_444.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_440.yuv
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_420_a10-1.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_420.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_422.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_444.y4m
+LIBVPX_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_440.yuv
+
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += desktop_credits.y4m
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += niklas_1280_720_30.y4m
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += rush_hour_444.y4m
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += screendata.y4m
+
+# Test vectors
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-001.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-001.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-002.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-002.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-003.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-003.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-004.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-004.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-005.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-005.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-006.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-006.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-007.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-007.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-008.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-008.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-009.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-009.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-010.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-010.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-011.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-011.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-012.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-012.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-013.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-013.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-014.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-014.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-015.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-015.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-016.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-016.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-017.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-017.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-018.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-00-comprehensive-018.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-01-intra-1400.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-01-intra-1400.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-01-intra-1411.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-01-intra-1411.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-01-intra-1416.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-01-intra-1416.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-01-intra-1417.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-01-intra-1417.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-02-inter-1402.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-02-inter-1402.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-02-inter-1412.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-02-inter-1412.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-02-inter-1418.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-02-inter-1418.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-02-inter-1424.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-02-inter-1424.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-01.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-01.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-02.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-02.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-03.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-03.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-04.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-04.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1401.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1401.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1403.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1403.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1407.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1407.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1408.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1408.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1409.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1409.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1410.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1410.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1413.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1413.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1414.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1414.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1415.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1415.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1425.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1425.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1426.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1426.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1427.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1427.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1432.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1432.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1435.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1435.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1436.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1436.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1437.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1437.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1441.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1441.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1442.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-03-segmentation-1442.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-04-partitions-1404.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-04-partitions-1404.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-04-partitions-1405.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-04-partitions-1405.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-04-partitions-1406.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-04-partitions-1406.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1428.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1428.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1429.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1429.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1430.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1430.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1431.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1431.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1433.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1433.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1434.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1434.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1438.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1438.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1439.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1439.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1440.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1440.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1443.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-05-sharpness-1443.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-06-smallsize.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP8_DECODER) += vp80-06-smallsize.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-00.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-00.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-01.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-01.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-02.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-02.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-03.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-03.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-04.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-04.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-05.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-05.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-06.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-06.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-07.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-07.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-08.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-08.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-09.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-09.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-10.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-10.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-11.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-11.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-12.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-12.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-13.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-13.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-14.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-14.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-15.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-15.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-17.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-17.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-18.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-18.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-19.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-19.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-20.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-20.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-21.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-21.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-22.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-22.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-23.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-23.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-24.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-24.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-25.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-25.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-26.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-26.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-27.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-27.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-28.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-28.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-29.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-29.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-30.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-30.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-31.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-31.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-32.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-32.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-33.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-33.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-34.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-34.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-35.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-35.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-36.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-36.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-37.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-37.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-38.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-38.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-39.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-39.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-40.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-40.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-41.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-41.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-42.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-42.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-43.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-43.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-44.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-44.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-45.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-45.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-46.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-46.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-47.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-47.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-48.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-48.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-49.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-49.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-50.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-50.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-51.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-51.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-52.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-52.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-53.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-53.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-54.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-54.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-55.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-55.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-56.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-56.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-57.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-57.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-58.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-58.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-59.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-59.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-60.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-60.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-61.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-61.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-62.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-62.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-63.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-00-quantizer-63.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-3.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-3.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-5.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-5.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-6.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-6.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-7.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-01-sharpness-7.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x08.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x08.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x10.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x10.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x18.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x18.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x32.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x32.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x34.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x34.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x64.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x64.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x66.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-08x66.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x08.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x08.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x10.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x10.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x18.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x18.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x32.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x32.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x34.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x34.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x64.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x64.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x66.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-10x66.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x08.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x08.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x10.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x10.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x18.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x18.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x32.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x32.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x34.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x34.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x64.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x64.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x66.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-16x66.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x08.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x08.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x10.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x10.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x18.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x18.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x32.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x32.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x34.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x34.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x64.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x64.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x66.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-18x66.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x08.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x08.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x10.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x10.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x18.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x18.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x32.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x32.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x34.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x34.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x64.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x64.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x66.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-32x66.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x08.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x08.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x10.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x10.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x18.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x18.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x32.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x32.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x34.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x34.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x64.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x64.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x66.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-34x66.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x08.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x08.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x10.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x10.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x18.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x18.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x32.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x32.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x34.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x34.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x64.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x64.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x66.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-64x66.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x08.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x08.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x10.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x10.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x18.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x18.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x32.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x32.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x34.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x34.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x64.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x64.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x66.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-66x66.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-130x132.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-130x132.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-132x130.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-132x130.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-132x132.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-132x132.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-178x180.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-178x180.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-180x178.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-180x178.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-180x180.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-180x180.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-lf-1920x1080.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-02-size-lf-1920x1080.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-deltaq.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-deltaq.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x196.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x196.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x198.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x198.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x200.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x200.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x202.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x202.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x208.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x208.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x210.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x210.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x224.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x224.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x226.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-196x226.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x196.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x196.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x198.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x198.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x200.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x200.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x202.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x202.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x208.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x208.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x210.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x210.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x224.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x224.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x226.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-198x226.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x196.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x196.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x198.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x198.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x200.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x200.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x202.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x202.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x208.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x208.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x210.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x210.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x224.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x224.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x226.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-200x226.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x196.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x196.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x198.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x198.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x200.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x200.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x202.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x202.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x208.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x208.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x210.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x210.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x224.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x224.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x226.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-202x226.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x196.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x196.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x198.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x198.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x200.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x200.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x202.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x202.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x208.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x208.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x210.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x210.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x224.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x224.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x226.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-208x226.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x196.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x196.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x198.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x198.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x200.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x200.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x202.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x202.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x208.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x208.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x210.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x210.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x224.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x224.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x226.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-210x226.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x196.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x196.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x198.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x198.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x200.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x200.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x202.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x202.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x208.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x208.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x210.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x210.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x224.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x224.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x226.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-224x226.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x196.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x196.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x198.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x198.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x200.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x200.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x202.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x202.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x208.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x208.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x210.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x210.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x224.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x224.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x226.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-226x226.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-352x288.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-03-size-352x288.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-05-resize.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-05-resize.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-06-bilinear.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-06-bilinear.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-07-frame_parallel.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-07-frame_parallel.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-07-frame_parallel-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-07-frame_parallel-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile-4x1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile-4x1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile-4x4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile-4x4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x2_frame_parallel.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x2_frame_parallel.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x4_frame_parallel.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x4_frame_parallel.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x8.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x8.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x8_frame_parallel.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile_1x8_frame_parallel.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-09-aq2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-09-aq2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-09-lf_deltas.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-09-lf_deltas.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-09-subpixel-00.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-09-subpixel-00.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-10-show-existing-frame2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-351x287.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-351x287.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-351x288.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-351x288.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-352x287.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-11-size-352x287.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-12-droppable_1.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-12-droppable_1.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-12-droppable_2.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-12-droppable_2.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-12-droppable_3.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-12-droppable_3.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-13-largescaling.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-13-largescaling.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-8.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-1-8.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-8-4-2-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-8-4-2-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-8.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-16-8.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-2-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-2-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-2-16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-2-16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-2-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-2-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-2-8.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-2-8.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-4-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-4-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-4-16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-4-16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-4-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-4-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-4-8.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-4-8.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-8-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-8-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-8-16.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-8-16.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-8-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-8-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-8-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-fp-tiles-8-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-1-2-4-8.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-1-2-4-8.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-1-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-1-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-1-8.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-1-8.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-2-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-2-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-2-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-2-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-2-8.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-2-8.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-4-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-4-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-4-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-4-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-4-8.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-4-8.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-8-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-8-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-8-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-8-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-8-4-2-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-8-4-2-1.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-8-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-14-resize-10frames-fp-tiles-8-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey_adpq.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey_adpq.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-16-intra-only.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-16-intra-only.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-17-show-existing-frame.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-17-show-existing-frame.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-18-resize.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-18-resize.ivf.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-19-skip.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-19-skip.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-19-skip-01.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-19-skip-01.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-19-skip-02.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-19-skip-02.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv422.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv422.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv440.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv440.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv444.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv444.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-20-big_superframe-01.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-20-big_superframe-01.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-20-big_superframe-02.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-20-big_superframe-02.webm.md5
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp92-2-20-10bit-yuv420.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp92-2-20-10bit-yuv420.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp92-2-20-12bit-yuv420.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp92-2-20-12bit-yuv420.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-10bit-yuv422.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-10bit-yuv422.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-12bit-yuv422.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-12bit-yuv422.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-10bit-yuv440.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-10bit-yuv440.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-12bit-yuv440.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-12bit-yuv440.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-10bit-yuv444.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-10bit-yuv444.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-12bit-yuv444.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp93-2-20-12bit-yuv444.webm.md5
+endif  # CONFIG_VP9_HIGHBITDEPTH
+
+# Invalid files for testing libvpx error checking.
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-01-v3.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-01-v3.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-02-v2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-02-v2.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-03-v3.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-03-v3.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-03-size-202x210.webm.ivf.s113306_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-03-size-202x210.webm.ivf.s113306_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-03-size-224x196.webm.ivf.s44156_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-03-size-224x196.webm.ivf.s44156_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-05-resize.ivf.s59293_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-05-resize.ivf.s59293_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-08-tile_1x2_frame_parallel.webm.ivf.s47039_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-08-tile_1x2_frame_parallel.webm.ivf.s47039_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-08-tile_1x8_frame_parallel.webm.ivf.s288_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-08-tile_1x8_frame_parallel.webm.ivf.s288_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-08-tile_1x4_frame_parallel_all_key.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-08-tile_1x4_frame_parallel_all_key.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-09-aq2.webm.ivf.s3984_r01-05_b6-.v2.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-09-aq2.webm.ivf.s3984_r01-05_b6-.v2.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-09-subpixel-00.ivf.s19552_r01-05_b6-.v2.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-09-subpixel-00.ivf.s19552_r01-05_b6-.v2.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-09-subpixel-00.ivf.s20492_r01-05_b6-.v2.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-09-subpixel-00.ivf.s20492_r01-05_b6-.v2.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-12-droppable_1.ivf.s3676_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-12-droppable_1.ivf.s3676_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-12-droppable_1.ivf.s73804_r01-05_b6-.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-12-droppable_1.ivf.s73804_r01-05_b6-.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp91-2-mixedrefcsp-444to420.ivf
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp91-2-mixedrefcsp-444to420.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-07-frame_parallel-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-07-frame_parallel-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-07-frame_parallel-3.webm
+
+ifeq ($(CONFIG_DECODE_PERF_TESTS),yes)
+# Encode / Decode test
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += niklas_1280_720_30.yuv
+# BBB VP9 streams
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-bbb_426x240_tile_1x1_180kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-bbb_640x360_tile_1x2_337kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-bbb_854x480_tile_1x2_651kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-bbb_1280x720_tile_1x4_1310kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-bbb_1920x1080_tile_1x1_2581kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-bbb_1920x1080_tile_1x4_2586kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-bbb_1920x1080_tile_1x4_fpm_2304kbps.webm
+# Sintel VP9 streams
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-sintel_426x182_tile_1x1_171kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-sintel_640x272_tile_1x2_318kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-sintel_854x364_tile_1x2_621kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-sintel_1280x546_tile_1x4_1257kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-sintel_1920x818_tile_1x4_fpm_2279kbps.webm
+# TOS VP9 streams
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-tos_426x178_tile_1x1_181kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-tos_640x266_tile_1x2_336kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-tos_854x356_tile_1x2_656kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-tos_854x356_tile_1x2_fpm_546kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-tos_1280x534_tile_1x4_1306kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-tos_1280x534_tile_1x4_fpm_952kbps.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-tos_1920x800_tile_1x4_fpm_2335kbps.webm
+endif  # CONFIG_DECODE_PERF_TESTS
+
+ifeq ($(CONFIG_ENCODE_PERF_TESTS),yes)
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += desktop_640_360_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += kirland_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += macmarcomoving_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += macmarcostationary_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += niklas_1280_720_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += niklas_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += tacomanarrows_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += tacomasmallcameramovement_640_480_30.yuv
+LIBVPX_TEST_DATA-$(CONFIG_VP9_ENCODER) += thaloundeskmtg_640_480_30.yuv
+endif  # CONFIG_ENCODE_PERF_TESTS
+
+# sort and remove duplicates
+LIBVPX_TEST_DATA-yes := $(sort $(LIBVPX_TEST_DATA-yes))
+
+# VP9 dynamic resizing test (decoder)
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x180_5_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x180_5_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x180_5_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x180_5_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x180_7_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x180_7_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x180_7_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x180_7_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x240_5_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x240_5_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x240_5_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x240_5_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x240_7_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x240_7_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x240_7_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_320x240_7_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x360_5_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x360_5_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x360_5_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x360_5_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x360_7_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x360_7_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x360_7_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x360_7_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x480_5_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x480_5_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x480_5_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x480_5_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x480_7_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x480_7_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x480_7_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_640x480_7_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1280x720_5_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1280x720_5_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1280x720_5_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1280x720_5_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1280x720_7_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1280x720_7_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1280x720_7_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1280x720_7_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1920x1080_5_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1920x1080_5_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1920x1080_5_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1920x1080_5_3-4.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1920x1080_7_1-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1920x1080_7_1-2.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1920x1080_7_3-4.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-21-resize_inter_1920x1080_7_3-4.webm.md5
diff --git a/libvpx/libvpx/test/test-data.sha1 b/libvpx/libvpx/test/test-data.sha1
new file mode 100644
index 0000000..a4ed174
--- /dev/null
+++ b/libvpx/libvpx/test/test-data.sha1
@@ -0,0 +1,836 @@
+d5dfb0151c9051f8c85999255645d7a23916d3c0 *hantro_collage_w352h288.yuv
+b87815bf86020c592ccc7a846ba2e28ec8043902 *hantro_odd.yuv
+76024eb753cdac6a5e5703aaea189d35c3c30ac7 *invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf
+7448d8798a4380162d4b56f9b452e2f6f9e24e7a *invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.v2.ivf.res
+83f50908c8dc0ef8760595447a2ff7727489542e *invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf
+456d1493e52d32a5c30edf44a27debc1fa6b253a *invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf.res
+c123d1f9f02fb4143abb5e271916e3a3080de8f6 *invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf
+456d1493e52d32a5c30edf44a27debc1fa6b253a *invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf.res
+fe346136b9b8c1e6f6084cc106485706915795e4 *invalid-vp90-01-v3.webm
+5d9474c0309b7ca09a182d888f73b37a8fe1362c *invalid-vp90-01-v3.webm.res
+d78e2fceba5ac942246503ec8366f879c4775ca5 *invalid-vp90-02-v2.webm
+8e2eff4af87d2b561cce2365713269e301457ef3 *invalid-vp90-02-v2.webm.res
+df1a1453feb3c00d7d89746c7003b4163523bff3 *invalid-vp90-03-v3.webm
+4935c62becc68c13642a03db1e6d3e2331c1c612 *invalid-vp90-03-v3.webm.res
+d637297561dd904eb2c97a9015deeb31c4a1e8d2 *invalid-vp90-2-08-tile_1x4_frame_parallel_all_key.webm
+3a204bdbeaa3c6458b77bcebb8366d107267f55d *invalid-vp90-2-08-tile_1x4_frame_parallel_all_key.webm.res
+a432f96ff0a787268e2f94a8092ab161a18d1b06 *park_joy_90p_10_420.y4m
+0b194cc312c3a2e84d156a221b0a5eb615dfddc5 *park_joy_90p_10_422.y4m
+ff0e0a21dc2adc95b8c1b37902713700655ced17 *park_joy_90p_10_444.y4m
+c934da6fb8cc54ee2a8c17c54cf6076dac37ead0 *park_joy_90p_10_440.yuv
+614c32ae1eca391e867c70d19974f0d62664dd99 *park_joy_90p_12_420.y4m
+c92825f1ea25c5c37855083a69faac6ac4641a9e *park_joy_90p_12_422.y4m
+b592189b885b6cc85db55cc98512a197d73d3b34 *park_joy_90p_12_444.y4m
+82c1bfcca368c2f22bad7d693d690d5499ecdd11 *park_joy_90p_12_440.yuv
+b9e1e90aece2be6e2c90d89e6ab2372d5f8c792d *park_joy_90p_8_420_a10-1.y4m
+4e0eb61e76f0684188d9bc9f3ce61f6b6b77bb2c *park_joy_90p_8_420.y4m
+7a193ff7dfeb96ba5f82b2afd7afa9e1fe83d947 *park_joy_90p_8_422.y4m
+bdb7856e6bc93599bdda05c2e773a9f22b6c6d03 *park_joy_90p_8_444.y4m
+81e1f3843748438b8f2e71db484eb22daf72e939 *park_joy_90p_8_440.yuv
+b1f1c3ec79114b9a0651af24ce634afb44a9a419 *rush_hour_444.y4m
+5184c46ddca8b1fadd16742e8500115bc8f749da *vp80-00-comprehensive-001.ivf
+65bf1bbbced81b97bd030f376d1b7f61a224793f *vp80-00-comprehensive-002.ivf
+906b4c1e99eb734504c504b3f1ad8052137ce672 *vp80-00-comprehensive-003.ivf
+ec144b1af53af895db78355785650b96dd3f0ade *vp80-00-comprehensive-004.ivf
+afc7091785c62f1c121c4554a2830c30704587d9 *vp80-00-comprehensive-005.ivf
+42ea9d55c818145d06a9b633b8e85c6a6164fd3e *vp80-00-comprehensive-006.ivf
+e5b3a73ab79fe024c14309d653d6bed92902ee3b *vp80-00-comprehensive-007.ivf
+f3c50a58875930adfb84525c0ef59d7e4c08540c *vp80-00-comprehensive-008.ivf
+4b2841fdb83db51ae322096ae468bbb9dc2c8362 *vp80-00-comprehensive-009.ivf
+efbff736e3a91ab6a98c5bc2dce65d645944c7b1 *vp80-00-comprehensive-010.ivf
+6b315102cae008d22a3d2c231be92cb704a222f8 *vp80-00-comprehensive-011.ivf
+f3214a4fea14c2d5ec689936c1613f274c859ee8 *vp80-00-comprehensive-012.ivf
+e4094e96d308c8a35b74c480a43d853c5294cd34 *vp80-00-comprehensive-013.ivf
+5b0adfaf60a69e0aaf3ec021a39d0a68fc0e1b5a *vp80-00-comprehensive-014.ivf
+e8467688ddf26b5000664f904faf0d70506aa653 *vp80-00-comprehensive-015.ivf
+aab55582337dfd2a39ff54fb2576a91910d49337 *vp80-00-comprehensive-016.ivf
+1ba24724f80203c9bae4f1d0f99d534721980016 *vp80-00-comprehensive-017.ivf
+143a15512b46f436280ddb4d0e6411eb4af434f2 *vp80-00-comprehensive-018.ivf
+c5baeaf5714fdfb3a8bc960a8e33ac438e83b16b *vp80-01-intra-1400.ivf
+f383955229afe3408453e316d11553d923ca60d5 *vp80-01-intra-1411.ivf
+84e1f4343f174c9f3c83f834bac3196fb325bf2c *vp80-01-intra-1416.ivf
+fb6e712a47dd57a28a3727d2ae2c97a8b7c7ca51 *vp80-01-intra-1417.ivf
+71ea772d3e9d315b8cbecf41207b8a237c34853b *vp80-02-inter-1402.ivf
+d85dbc4271525dcd128c503f936fe69091d1f8d0 *vp80-02-inter-1412.ivf
+d4e5d3ad56511867d025f93724d090f92ba6ec3d *vp80-02-inter-1418.ivf
+91791cbcc37c60f35dbd8090bacb54e5ec6dd4fa *vp80-02-inter-1424.ivf
+17fbfe2fea70f6e2f3fa6ca4efaae6c0b03b5f02 *vp80-03-segmentation-01.ivf
+3c3600dbbcde08e20d54c66fe3b7eadd4f09bdbb *vp80-03-segmentation-02.ivf
+c156778d5340967d4b369c490848076e92f1f875 *vp80-03-segmentation-03.ivf
+d25dcff6c60e87a1af70945b8911b6b4998533b0 *vp80-03-segmentation-04.ivf
+362baba2ce454c9db21218f35e81c27a5ed0b730 *vp80-03-segmentation-1401.ivf
+d223ae7ee748ce07e74c4679bfd219e84aa9f4b0 *vp80-03-segmentation-1403.ivf
+033adf7f3a13836a3f1cffcb87c1972900f2b5c6 *vp80-03-segmentation-1407.ivf
+4d51dfbf9f3e2c590ec99d1d6f59dd731d04375f *vp80-03-segmentation-1408.ivf
+f37a62b197c2600d75e0ccfbb31b60efdedac251 *vp80-03-segmentation-1409.ivf
+eb25bd7bfba5b2f6935018a930f42d123b1e7fcd *vp80-03-segmentation-1410.ivf
+b9d5c436663a30c27cfff84b53a002e501258843 *vp80-03-segmentation-1413.ivf
+6da92b9d1a180cc3a8afe348ab12258f5a37be1a *vp80-03-segmentation-1414.ivf
+a4f5842602886bd669f115f93d8a35c035cb0948 *vp80-03-segmentation-1415.ivf
+f295dceb8ef278b77251b3f9df8aee22e161d547 *vp80-03-segmentation-1425.ivf
+198dbf9f36f733200e432664cc8c5752d59779de *vp80-03-segmentation-1426.ivf
+7704804e32f5de976803929934a7fafe101ac7b0 *vp80-03-segmentation-1427.ivf
+831ccd862ea95ca025d2f3bd8b88678752f5416d *vp80-03-segmentation-1432.ivf
+b3c11978529289f9109f2766fcaba3ebc40e11ef *vp80-03-segmentation-1435.ivf
+a835a731f5520ebfc1002c40121264d0020559ac *vp80-03-segmentation-1436.ivf
+1d1732942f773bb2a5775fcb9689b1579ce28eab *vp80-03-segmentation-1437.ivf
+db04799adfe089dfdf74dbd43cc05ede7161f99e *vp80-03-segmentation-1441.ivf
+7caf39b3f20cfd52b998210878062e52a5edf1e6 *vp80-03-segmentation-1442.ivf
+3607f6bb4ee106c38fa1ea370dc4ff8b8cde2261 *vp80-04-partitions-1404.ivf
+93cc323b6b6867f1b12dd48773424549c6960a6b *vp80-04-partitions-1405.ivf
+047eedb14b865bdac8a3538e63801054e0295e9c *vp80-04-partitions-1406.ivf
+0f1233bd2bc33f56ce5e495dbd455d122339f384 *vp80-05-sharpness-1428.ivf
+51767fc136488a9535c2a4c38067c542ee2048df *vp80-05-sharpness-1429.ivf
+9805aa107672de25d6fb8c35e20d06deca5efe18 *vp80-05-sharpness-1430.ivf
+61db6b965f9c27aebe71b85bf2d5877e58e4bbdf *vp80-05-sharpness-1431.ivf
+10420d266290d2923555f84af38eeb96edbd3ae8 *vp80-05-sharpness-1433.ivf
+3ed24f9a80cddfdf75824ba95cdb4ff9286cb443 *vp80-05-sharpness-1434.ivf
+c87599cbecd72d4cd4f7ace3313b7a6bc6eb8163 *vp80-05-sharpness-1438.ivf
+aff51d865c2621b60510459244ea83e958e4baed *vp80-05-sharpness-1439.ivf
+da386e72b19b5485a6af199c5eb60ef25e510dd1 *vp80-05-sharpness-1440.ivf
+6759a095203d96ccd267ce09b1b050b8cc4c2f1f *vp80-05-sharpness-1443.ivf
+b95d3cc1d0df991e63e150a801710a72f20d9ba0 *vp80-06-smallsize.ivf
+db55ec7fd02c864ba996ff060b25b1e08611330b *vp80-00-comprehensive-001.ivf.md5
+29db0ad011cba1e45f856d5623cd38dac3e3bf19 *vp80-00-comprehensive-002.ivf.md5
+e84f258f69e173e7d68f8f8c037a0a3766902182 *vp80-00-comprehensive-003.ivf.md5
+eb7912eaf69559a16fd82bc3f5fb1524cf4a4466 *vp80-00-comprehensive-004.ivf.md5
+4206f71c94894bd5b5b376f6c09b3817dbc65206 *vp80-00-comprehensive-005.ivf.md5
+4f89b356f6f2fecb928f330a10f804f00f5325f5 *vp80-00-comprehensive-006.ivf.md5
+2813236a32964dd8007e17648bcf035a20fcda6c *vp80-00-comprehensive-007.ivf.md5
+10746c72098f872803c900e17c5680e451f5f498 *vp80-00-comprehensive-008.ivf.md5
+39a23d0692ce64421a7bb7cdf6ccec5928d37fff *vp80-00-comprehensive-009.ivf.md5
+f6e3de8931a0cc659bda8fbc14050346955e72d4 *vp80-00-comprehensive-010.ivf.md5
+101683ec195b6e944f7cd1e468fc8921439363e6 *vp80-00-comprehensive-011.ivf.md5
+1f592751ce46d8688998fa0fa4fbdcda0fd4058c *vp80-00-comprehensive-012.ivf.md5
+6066176f90ca790251e795fca1a5797d59999841 *vp80-00-comprehensive-013.ivf.md5
+2656da94ba93691f23edc4d60b3a09e2be46c217 *vp80-00-comprehensive-014.ivf.md5
+c6e0d5f5d61460c8ac8edfa4e701f10312c03133 *vp80-00-comprehensive-015.ivf.md5
+ee60fee501d8493e34e8d6a1fe315b51ed09b24a *vp80-00-comprehensive-016.ivf.md5
+9f1914ceffcad4546c0a29de3ef591d8bea304dc *vp80-00-comprehensive-017.ivf.md5
+e0305178fe288a9fd8082b39e2d03181edb19054 *vp80-00-comprehensive-018.ivf.md5
+612494da2fa799cc9d76dcdd835ae6c7cb2e5c05 *vp80-01-intra-1400.ivf.md5
+48ea06097ac8269c5e8c2131d3d0639f431fcf0e *vp80-01-intra-1411.ivf.md5
+6e2ab4e7677ad0ba868083ca6bc387ee922b400c *vp80-01-intra-1416.ivf.md5
+eca0a90348959ce3854142f8d8641b13050e8349 *vp80-01-intra-1417.ivf.md5
+920feea203145d5c2258a91c4e6991934a79a99e *vp80-02-inter-1402.ivf.md5
+f71d97909fe2b3dd65be7e1f56c72237f0cef200 *vp80-02-inter-1412.ivf.md5
+e911254569a30bbb2a237ff8b79f69ed9da0672d *vp80-02-inter-1418.ivf.md5
+58c789c50c9bb9cc90580bed291164a0939d28ba *vp80-02-inter-1424.ivf.md5
+ff3e2f441327b9c20a0b37c524e0f5a48a36de7b *vp80-03-segmentation-01.ivf.md5
+0791f417f076a542ae66fbc3426ab4d94cbd6c75 *vp80-03-segmentation-02.ivf.md5
+722e50f1a6a91c34302d68681faffc1c26d1cc57 *vp80-03-segmentation-03.ivf.md5
+c701f1885bcfb27fb8e70cc65606b289172ef889 *vp80-03-segmentation-04.ivf.md5
+f79bc9ec189a2b4807632a3d0c5bf04a178b5300 *vp80-03-segmentation-1401.ivf.md5
+b9aa4c74c0219b639811c44760d0b24cd8bb436a *vp80-03-segmentation-1403.ivf.md5
+70d5a2207ca1891bcaebd5cf6dd88ce8d57b4334 *vp80-03-segmentation-1407.ivf.md5
+265f962ee781531f9a93b9309461316fd32b2a1d *vp80-03-segmentation-1408.ivf.md5
+0c4ecbbd6dc042d30e626d951b65f460dd6cd563 *vp80-03-segmentation-1409.ivf.md5
+cf779af36a937f06570a0fca9db64ba133451dee *vp80-03-segmentation-1410.ivf.md5
+0e6c5036d51ab078842f133934926c598a9cff02 *vp80-03-segmentation-1413.ivf.md5
+eb3930aaf229116c80d507516c34759c3f6cdf69 *vp80-03-segmentation-1414.ivf.md5
+123d6c0f72ee87911c4ae7538e87b7d163b22d6c *vp80-03-segmentation-1415.ivf.md5
+e70551d1a38920e097a5d8782390b79ecaeb7505 *vp80-03-segmentation-1425.ivf.md5
+44e8f4117e46dbb302b2cfd81171cc1a1846e431 *vp80-03-segmentation-1426.ivf.md5
+52636e54aee5f95bbace37021bd67de5db767e9a *vp80-03-segmentation-1427.ivf.md5
+b1ad3eff20215c28e295b15ef3636ed926d59cba *vp80-03-segmentation-1432.ivf.md5
+24c22a552fa28a90e5978f67f57181cc2d7546d7 *vp80-03-segmentation-1435.ivf.md5
+96c49c390abfced18a7a8c9b9ea10af778e10edb *vp80-03-segmentation-1436.ivf.md5
+f95eb6214571434f1f73ab7833b9ccdf47588020 *vp80-03-segmentation-1437.ivf.md5
+1c0700ca27c9b0090a7747a4b0b4dc21d1843181 *vp80-03-segmentation-1441.ivf.md5
+81d4f23ca32667ee958bae579c8f5e97ba72eb97 *vp80-03-segmentation-1442.ivf.md5
+272efcef07a3a30fbca51bfd566063d8258ec0be *vp80-04-partitions-1404.ivf.md5
+66ed219ab812ac801b256d35cf495d193d4cf478 *vp80-04-partitions-1405.ivf.md5
+36083f37f56f502bd60ec5e07502ee9e6b8699b0 *vp80-04-partitions-1406.ivf.md5
+6ca909bf168a64c09415626294665dc1be3d1973 *vp80-05-sharpness-1428.ivf.md5
+1667d2ee2334e5fdea8a8a866f4ccf3cf76f033a *vp80-05-sharpness-1429.ivf.md5
+71bcbe5357d36a19df5b07fbe3e27bffa8893f0a *vp80-05-sharpness-1430.ivf.md5
+89a09b1dffce2d55770a89e58d9925c70ef79bf8 *vp80-05-sharpness-1431.ivf.md5
+08444a18b4e6ba3450c0796dd728d48c399a2dc9 *vp80-05-sharpness-1433.ivf.md5
+6d6223719a90c13e848aa2a8a6642098cdb5977a *vp80-05-sharpness-1434.ivf.md5
+41d70bb5fa45bc88da1604a0af466930b8dd77b5 *vp80-05-sharpness-1438.ivf.md5
+086c56378df81b6cee264d7540a7b8f2b405c7a4 *vp80-05-sharpness-1439.ivf.md5
+d32dc2c4165eb266ea4c23c14a45459b363def32 *vp80-05-sharpness-1440.ivf.md5
+8c69dc3d8e563f56ffab5ad1e400d9e689dd23df *vp80-05-sharpness-1443.ivf.md5
+d6f246df012c241b5fa6c1345019a3703d85c419 *vp80-06-smallsize.ivf.md5
+ce881e567fe1d0fbcb2d3e9e6281a1a8d74d82e0 *vp90-2-00-quantizer-00.webm
+ac5eda33407d0521c7afca43a63fd305c0cd9d13 *vp90-2-00-quantizer-00.webm.md5
+2ca0463f2cfb93d25d7dded174db70b7cb87cb48 *vp90-2-00-quantizer-01.webm
+10d98884fc6d9a5f47a2057922b8e25dd48d7786 *vp90-2-00-quantizer-01.webm.md5
+d80a2920a5e0819d69dcba8fe260c01f820f8982 *vp90-2-00-quantizer-02.webm
+c964c8e5e04165fabbf1c6ee8ee5121d35921965 *vp90-2-00-quantizer-02.webm.md5
+fdef046777b5b75c962b715d809dbe2ea331afb9 *vp90-2-00-quantizer-03.webm
+f270bee0b0c7aa2bf4c5afe098556b4f3f890faf *vp90-2-00-quantizer-03.webm.md5
+66d98609e809394a6ac730787e6724e3badc075a *vp90-2-00-quantizer-04.webm
+427433bfe121c4aea1095ec3124fdc174d200e3a *vp90-2-00-quantizer-04.webm.md5
+e6e42626d8cadf0b5be16313f69212981b96fee5 *vp90-2-00-quantizer-05.webm
+c98f6a9a1af4cfd71416792827304266aad4bd46 *vp90-2-00-quantizer-05.webm.md5
+413ef09b721f5dcec1a96e937a97e5873c2e6db6 *vp90-2-00-quantizer-06.webm
+5080e940a23805c82e578e21b57fc2c511e76376 *vp90-2-00-quantizer-06.webm.md5
+4a50a5f4ac717c30dfaae8bb46702e3542e867de *vp90-2-00-quantizer-07.webm
+76c429a02b56762e10ee4db88729d8834b3a70f4 *vp90-2-00-quantizer-07.webm.md5
+d2f4e464780bf8b7e647efa18ac777a930e62bc0 *vp90-2-00-quantizer-08.webm
+ab94aabf9316111b52d7c531962ed4123313b6ba *vp90-2-00-quantizer-08.webm.md5
+174bc58433936dd79550398d744f1072ce7f5693 *vp90-2-00-quantizer-09.webm
+e1f7690cd83ccc56d045e17cce552544a5f03810 *vp90-2-00-quantizer-09.webm.md5
+52bc1dfd3a97b24d922eb8a31d07527891561f2a *vp90-2-00-quantizer-10.webm
+9b37bed893b5f6a4e12f2aa40f02dd40f944d0f8 *vp90-2-00-quantizer-10.webm.md5
+10031eecafde1e1d8e6323fe2b2a1d7e77a66869 *vp90-2-00-quantizer-11.webm
+fe4620a4bb0e4f5cb9bbfedc4039a22b81b0f5c0 *vp90-2-00-quantizer-11.webm.md5
+78e9f7bb77e8e348155bbdfa12790789d1d50c34 *vp90-2-00-quantizer-12.webm
+0961d060cc8dd469c6dac8d7d75f927c0bb971b8 *vp90-2-00-quantizer-12.webm.md5
+133b77a3bbcef652552d74ffc46afbfe3b8a1cba *vp90-2-00-quantizer-13.webm
+df29e5e0f95772af482f540d776f6b9dea4bfa29 *vp90-2-00-quantizer-13.webm.md5
+27323afdaf8987e025c27129c74c86502315a206 *vp90-2-00-quantizer-14.webm
+ce96a2cc312942f0427a463f15a392870dd69764 *vp90-2-00-quantizer-14.webm.md5
+ab58d0b41037829f6bc993910999f4af0212aafd *vp90-2-00-quantizer-15.webm
+40f700db606501aa7cb49049624cbdde6409b122 *vp90-2-00-quantizer-15.webm.md5
+cd948e66448aafb65998815ce37241f95d7c9ee7 *vp90-2-00-quantizer-16.webm
+039b742d149c945ed79c7b9a6384352852a1c116 *vp90-2-00-quantizer-16.webm.md5
+62f56e663e13c576764e491cf08f19bd46a71999 *vp90-2-00-quantizer-17.webm
+90c5a39bf76e6b3e0a1c0d3e9b68a9fd78be963e *vp90-2-00-quantizer-17.webm.md5
+f26ecad7263cd66a614e53ba5d7c00df181affeb *vp90-2-00-quantizer-18.webm
+cda0a1c0fca2ec2976ae55124a8a67305508bae6 *vp90-2-00-quantizer-18.webm.md5
+94bfc4c04fcfe139a63b98c569e8c14ba98c401f *vp90-2-00-quantizer-19.webm
+5b8ec169ccf67d8a0a8e46a62eb173f5a1dbaf4f *vp90-2-00-quantizer-19.webm.md5
+0ee88e9318985e1e245de78c2c4a665885ab76a7 *vp90-2-00-quantizer-20.webm
+4b26f7edb4fcd3a1b4cce9ba3cb8650e3ee6e063 *vp90-2-00-quantizer-20.webm.md5
+6a995cb2b1db33da8087321df1e646f95c3e32d1 *vp90-2-00-quantizer-21.webm
+e216b4a1eceac03efcc433759be54ab8ea87b24b *vp90-2-00-quantizer-21.webm.md5
+aa7722fc427e7180115f3c9cd96bb6b2768e7296 *vp90-2-00-quantizer-22.webm
+1aa813bd45ae831bf5e79ace4d73dfd25989a07d *vp90-2-00-quantizer-22.webm.md5
+7677e5b929ed6d142041f19b8a9cd5822ee1504a *vp90-2-00-quantizer-23.webm
+0de0af34abd843d5b37e58baf3ed96a6104b64c3 *vp90-2-00-quantizer-23.webm.md5
+b2995cbe1128b2d4926f1b28d01c501ecb6be8c8 *vp90-2-00-quantizer-24.webm
+db6033af2ba2f2bca62468fb4b8808e474f93923 *vp90-2-00-quantizer-24.webm.md5
+8135ba35587fd92cd4667be7896323d9b634401c *vp90-2-00-quantizer-25.webm
+3499e00c2cc15876f61f07e3d3cfca54ebcd98fd *vp90-2-00-quantizer-25.webm.md5
+af0fa2907746db82d345f6d831fcc1b2862a29fb *vp90-2-00-quantizer-26.webm
+cd6fe3d14dab48886ebf65be00e6ed9616ebe5a7 *vp90-2-00-quantizer-26.webm.md5
+bd0002e91323776beb5ff11e06edcf19fc08e9b9 *vp90-2-00-quantizer-27.webm
+fe72154ef196067d6c272521012dd79706496cac *vp90-2-00-quantizer-27.webm.md5
+fc15eb606f81455ff03df16bf3432296b002c43c *vp90-2-00-quantizer-28.webm
+40b2e24b542206a6bfd746ef199e49ccea07678a *vp90-2-00-quantizer-28.webm.md5
+3090bbf913cad0b2eddca7228f5ed51a58378b8d *vp90-2-00-quantizer-29.webm
+eb59745e0912d8ed6c928268bcf265237c9ba93f *vp90-2-00-quantizer-29.webm.md5
+c615abdca9c25e1cb110d908edbedfb3b7c92b91 *vp90-2-00-quantizer-30.webm
+ad0f4fe6733e4e7cdfe8ef8722bb341dcc7538c0 *vp90-2-00-quantizer-30.webm.md5
+037d9f242086cfb085518f6416259defa82d5fc2 *vp90-2-00-quantizer-31.webm
+4654b40792572f0a790874c6347ef9196d86c1a7 *vp90-2-00-quantizer-31.webm.md5
+505899f3f3515044c5c8b3213d9b9d16f614619d *vp90-2-00-quantizer-32.webm
+659a2e6dd02df323f62600626859006640b445df *vp90-2-00-quantizer-32.webm.md5
+8b32ec9c3b7e5ca8ddc6b8aea1c1cb7ca996bccc *vp90-2-00-quantizer-33.webm
+5b175ef1120ddeba4feae1247bf381bbc4e816ce *vp90-2-00-quantizer-33.webm.md5
+4d283755d17e287b1d099a80604398f60d7fb6ea *vp90-2-00-quantizer-34.webm
+22a739de95acfeb27524e3700b8f678a9ad744d8 *vp90-2-00-quantizer-34.webm.md5
+4296f56a892a412d3d4f64824718dd566c4e6459 *vp90-2-00-quantizer-35.webm
+c532c9c8dc7b3506fc6a51e5c20c17ef0ac039e7 *vp90-2-00-quantizer-35.webm.md5
+6f54e11da461e4410dd9075b015e2d9bc1d07dfb *vp90-2-00-quantizer-36.webm
+0b3573f5addea4e3eb11a0b85f068299d5bdad78 *vp90-2-00-quantizer-36.webm.md5
+210581682a26c2c4375efc785c36e07539888bc2 *vp90-2-00-quantizer-37.webm
+2b4fb6f8ba975237858e61cc8f560bcfc87cb38e *vp90-2-00-quantizer-37.webm.md5
+a15ef31283dfc4860f837fe200eb32a445f59629 *vp90-2-00-quantizer-38.webm
+fb76771f3a795054b9936f70da7505c3ac585284 *vp90-2-00-quantizer-38.webm.md5
+1df8433a441412831daae6726df89fa70d21b14d *vp90-2-00-quantizer-39.webm
+39e162c09a20e7e684868097766347014371fee6 *vp90-2-00-quantizer-39.webm.md5
+5330e4788ab9129dbb25a7a7d5411104521248b6 *vp90-2-00-quantizer-40.webm
+872cc0f2cc9dbf000f89eadb4d8f9940e48e00b1 *vp90-2-00-quantizer-40.webm.md5
+d88d03b982889e399a78d7a06eeb1cf30e6c2da2 *vp90-2-00-quantizer-41.webm
+5b4f7217e57fa2a221011d0b32f8d0409496b7b6 *vp90-2-00-quantizer-41.webm.md5
+9e16406e3e26955a6e17d455ef1ef64bbfa26e53 *vp90-2-00-quantizer-42.webm
+0219d090cf37daabe19256ba8e932ba4874b92e4 *vp90-2-00-quantizer-42.webm.md5
+a9b15843486fb05f8cd15437ef279782a42b75db *vp90-2-00-quantizer-43.webm
+3c9b0b4c607f9579a31726bfcf56729334ddc686 *vp90-2-00-quantizer-43.webm.md5
+1dbc931ac446c91eabe7213efff55b596cccf07c *vp90-2-00-quantizer-44.webm
+73bc8f675103abaef3d9f73a2742b3bffd726d23 *vp90-2-00-quantizer-44.webm.md5
+7c6c1be15beb9d6201204b018966c8c4f9777efc *vp90-2-00-quantizer-45.webm
+c907b29da821f790c6748de61f592689312e4e36 *vp90-2-00-quantizer-45.webm.md5
+07b434da1a467580f73b32177ee11b3e00f65a0d *vp90-2-00-quantizer-46.webm
+7b2b7ce60c50bc970bc0ada46d7a7ce440148da3 *vp90-2-00-quantizer-46.webm.md5
+233d0465fb1a6fa36e9f89bd2193ac79bd4d2809 *vp90-2-00-quantizer-47.webm
+527e0a9fb932efe915027ffe077f9e8d3a4fb139 *vp90-2-00-quantizer-47.webm.md5
+719613df7307e205c3fdb6acfb373849c5ab23c7 *vp90-2-00-quantizer-48.webm
+65ab6c9d1b682c183b201c7ff42b90343ce3e304 *vp90-2-00-quantizer-48.webm.md5
+3bf04a598325ed0eabae1598ec7f718f715ec672 *vp90-2-00-quantizer-49.webm
+ac68c4387ce11fcc998d8ba455ab9b2bb361d240 *vp90-2-00-quantizer-49.webm.md5
+d59238fb3a654931c9b65a11e7321b40d1f702e9 *vp90-2-00-quantizer-50.webm
+d0576bfede46fd55659f028f2fd28554ceb3e6cc *vp90-2-00-quantizer-50.webm.md5
+3f579785101d4209360dd96f8c2ffe9beddf3bee *vp90-2-00-quantizer-51.webm
+89fcfe04f4457a7f02ab4a2f94aacbb88aee5789 *vp90-2-00-quantizer-51.webm.md5
+28be5836e2fedefe4babf12fc9b79e460ab0a0f4 *vp90-2-00-quantizer-52.webm
+f3dd52b70c18345fee740220f35da9c4def2017a *vp90-2-00-quantizer-52.webm.md5
+488ad4058c17170665b6acd1021fade9a02771e4 *vp90-2-00-quantizer-53.webm
+1cdcb1d4f3a37cf83ad235eb27ec62ed2a01afc7 *vp90-2-00-quantizer-53.webm.md5
+682978289cb28cc8c9d39bc797300e45d6039de7 *vp90-2-00-quantizer-54.webm
+36c35353f2c03cb099bd710d9994de7d9ed88834 *vp90-2-00-quantizer-54.webm.md5
+c398ce49af762a48f10cc4da9fae0769aae5f226 *vp90-2-00-quantizer-55.webm
+2cf3570542d984f167ab087f59493c7fb47e0ed2 *vp90-2-00-quantizer-55.webm.md5
+3071f18b2fce261aa82d61f81a7ae4ca9a75d0e3 *vp90-2-00-quantizer-56.webm
+d3f93f8272b6de31cffb011a26f11abb514efb12 *vp90-2-00-quantizer-56.webm.md5
+f4e8e14b1f278801a7eb6f11734780a01b1668e9 *vp90-2-00-quantizer-57.webm
+6478fdf1d7faf6db5f19dffc5e1363af358699ee *vp90-2-00-quantizer-57.webm.md5
+307dc264f57cc618fff211fa44d7f52767ed9660 *vp90-2-00-quantizer-58.webm
+cf231d4a52d492fa692ea4194ec5eb7511fec54e *vp90-2-00-quantizer-58.webm.md5
+1fd7cd596170afce2de0b1441b7674bda5723440 *vp90-2-00-quantizer-59.webm
+4681f7ef96f63e085c41bb1a964b0df7e67e0b38 *vp90-2-00-quantizer-59.webm.md5
+34cdcc81c0ba7085aefbb22d7b4aa9bca3dd7c62 *vp90-2-00-quantizer-60.webm
+58691ef53b6b623810e2c57ded374c77535df935 *vp90-2-00-quantizer-60.webm.md5
+e6e812406aab81021bb16e772c1db03f75906cb6 *vp90-2-00-quantizer-61.webm
+76436eace62f08ff92b61a0845e66667a027db1b *vp90-2-00-quantizer-61.webm.md5
+84d811bceed70c950a6a08e572a6e274866e72b1 *vp90-2-00-quantizer-62.webm
+2d937cc011eeddd95222b960982da5cd18db580f *vp90-2-00-quantizer-62.webm.md5
+0912b295ba0ea09359315315ffd67d22d046f883 *vp90-2-00-quantizer-63.webm
+5a829031055d70565f57dbcd47a6ac33619952b3 *vp90-2-00-quantizer-63.webm.md5
+0cf9e5ebe0112bdb47b5887ee5d58eb9d4727c00 *vp90-2-01-sharpness-1.webm
+5a0476be4448bae8f8ca17ea236c98793a755948 *vp90-2-01-sharpness-1.webm.md5
+51e02d7911810cdf5be8b68ac40aedab479a3179 *vp90-2-01-sharpness-2.webm
+a0ca5bc87a5ed7c7051f59078daa0d03be1b45b6 *vp90-2-01-sharpness-2.webm.md5
+0603f8ad239c07a531d948187f4dafcaf51eda8d *vp90-2-01-sharpness-3.webm
+3af8000a69c72fe77881e3176f026c2affb78cc7 *vp90-2-01-sharpness-3.webm.md5
+4ca4839f48146252fb261ed88838d80211804841 *vp90-2-01-sharpness-4.webm
+08832a1494f84fa9edd40e080bcf2c0e80100c76 *vp90-2-01-sharpness-4.webm.md5
+95099dc8f9cbaf9b9a7dd65311923e441ff70731 *vp90-2-01-sharpness-5.webm
+93ceee30c140f0b406726c0d896b9db6031c4c7f *vp90-2-01-sharpness-5.webm.md5
+ceb4116fb7b078d266d153233b6d62a255a34e4c *vp90-2-01-sharpness-6.webm
+da83efe59e537ce538e8b03a6eac63cf25849c9a *vp90-2-01-sharpness-6.webm.md5
+b5f7cd19aece3880f9d616a778e5cc24c6b9b505 *vp90-2-01-sharpness-7.webm
+2957408d20deac8633941a2169f801bae6f086e1 *vp90-2-01-sharpness-7.webm.md5
+ffc096c2ce1050450ad462b5fabd2a5220846319 *vp90-2-02-size-08x08.webm
+e36d2ed6fa2746347710b750586aafa6a01ff3ae *vp90-2-02-size-08x08.webm.md5
+895b986f9fd55cd879472b31c6a06b82094418c8 *vp90-2-02-size-08x10.webm
+079157a19137ccaebba606f2871f45a397347150 *vp90-2-02-size-08x10.webm.md5
+1c5992203e62a2b83040ccbecd748b604e19f4c0 *vp90-2-02-size-08x16.webm
+9aa45ffdf2078f883bbed01450031b691819c144 *vp90-2-02-size-08x16.webm.md5
+d0a8953da1f85f484487408fee5da9e2a8391901 *vp90-2-02-size-08x18.webm
+59a5cc17d354c6a23e5e959d666b1456a5d49c56 *vp90-2-02-size-08x18.webm.md5
+1b13461a9fc65cb041bacfe4ea6f02d363397d61 *vp90-2-02-size-08x32.webm
+2bdddd6878f05d37d84cde056a3f5e7f926ba3d6 *vp90-2-02-size-08x32.webm.md5
+2861f0a0daadb62295b0504a1fbe5b50c79a8f59 *vp90-2-02-size-08x34.webm
+6b5812cfb8a82d378ea2913bf009e93668020147 *vp90-2-02-size-08x34.webm.md5
+02f948216d4246579dc53c47fe55d8fb264ba251 *vp90-2-02-size-08x64.webm
+84b55fdee6d9aa820c7a8c62822446184b191767 *vp90-2-02-size-08x64.webm.md5
+4b011242cbf42516efd2b197baebb61dd34562c9 *vp90-2-02-size-08x66.webm
+6b1fa0a885947b3cc0fe58f75f838e662bd9bb8b *vp90-2-02-size-08x66.webm.md5
+4057796be9dd12df48ab607f502ae6aa70eeeab6 *vp90-2-02-size-10x08.webm
+71c752c51aec9f48de286b93f4c20e9c11cad7d0 *vp90-2-02-size-10x08.webm.md5
+6583c853fa43fc53d51743eac5f3a43a359d45d0 *vp90-2-02-size-10x10.webm
+1da524d24af1944b671d4d3f2b398d6e336584c3 *vp90-2-02-size-10x10.webm.md5
+ba442fc03ccd3a705c64c83b36f5ada67d198874 *vp90-2-02-size-10x16.webm
+7cfd960f232c34c641a4a2a9411b6fd0efb2fc50 *vp90-2-02-size-10x16.webm.md5
+cc92ed40eef14f52e4d080cb2c57939dd8326374 *vp90-2-02-size-10x18.webm
+db5626275cc55ce970b91c995e74f6838d943aca *vp90-2-02-size-10x18.webm.md5
+3a93d501d22325e9fd4c9d8b82e2a432de33c351 *vp90-2-02-size-10x32.webm
+5cae51b0c71cfc131651f345f87583eb2903afaf *vp90-2-02-size-10x32.webm.md5
+50d2f2b15a9a5178153db44a9e03aaf32b227f67 *vp90-2-02-size-10x34.webm
+bb0efe058122641e7f73e94497dda2b9e6c21efd *vp90-2-02-size-10x34.webm.md5
+01624ec173e533e0b33fd9bdb91eb7360c7c9175 *vp90-2-02-size-10x64.webm
+b9c0e3b054463546356acf5157f9be92fd34732f *vp90-2-02-size-10x64.webm.md5
+2942879baf1c09e96b14d0fc84806abfe129c706 *vp90-2-02-size-10x66.webm
+bab5f539c2f91952e187456b4beafbb4c01e25ee *vp90-2-02-size-10x66.webm.md5
+88d2b63ca5e9ee163d8f20e8886f3df3ff301a66 *vp90-2-02-size-16x08.webm
+7f48a0fcf8c25963f3057d7f6669c5f2415834b8 *vp90-2-02-size-16x08.webm.md5
+59261eb34c15ea9b5ddd2d416215c1a8b9e6dc1f *vp90-2-02-size-16x10.webm
+73a7c209a46dd051c9f7339b6e02ccd5b3b9fc81 *vp90-2-02-size-16x10.webm.md5
+066834fef9cf5b9a72932cf4dea5f253e14a976d *vp90-2-02-size-16x16.webm
+faec542f52f37601cb9c480d887ae9355be99372 *vp90-2-02-size-16x16.webm.md5
+195307b4eb3192271ee4a935b0e48deef0c54cc2 *vp90-2-02-size-16x18.webm
+5a92e19e624c0376321d4d0e22c0c91995bc23e1 *vp90-2-02-size-16x18.webm.md5
+14f3f884216d7ae16ec521f024a2f2d31bbf9c1a *vp90-2-02-size-16x32.webm
+ea622d1c817dd174556f7ee7ccfe4942b34d4845 *vp90-2-02-size-16x32.webm.md5
+2e0501100578a5da9dd47e4beea160f945bdd1ba *vp90-2-02-size-16x34.webm
+1b8645ef64239334921c5f56b24ce815e6070b05 *vp90-2-02-size-16x34.webm.md5
+89a6797fbebebe93215f367229a9152277f5dcfe *vp90-2-02-size-16x64.webm
+a03d8c1179ca626a8856fb416d635dbf377979cd *vp90-2-02-size-16x64.webm.md5
+0f3a182e0750fcbae0b9eae80c7a53aabafdd18d *vp90-2-02-size-16x66.webm
+8cb6736dc2d897c1283919a32068af377d66c59c *vp90-2-02-size-16x66.webm.md5
+68fe70dc7914cc1d8d6dcd97388b79196ba3e7f1 *vp90-2-02-size-18x08.webm
+874c7fb505be9db3160c57cb405c4dbd5b990dc2 *vp90-2-02-size-18x08.webm.md5
+0546352dd78496d4dd86c3727ac2ff36c9e72032 *vp90-2-02-size-18x10.webm
+1d80eb36557ea5f25a386495a36f93da0f25316b *vp90-2-02-size-18x10.webm.md5
+60fe99e5f5cc99706efa3e0b894e45cbcf0d6330 *vp90-2-02-size-18x16.webm
+1ab6cdd89a53662995d103546e6611c84f9292ab *vp90-2-02-size-18x16.webm.md5
+f9a8f5fb749d69fd555db6ca093b7f77800c7b4f *vp90-2-02-size-18x18.webm
+ace8a66328f7802b15f9989c2720c029c6abd279 *vp90-2-02-size-18x18.webm.md5
+a197123a527ec25913a9bf52dc8c347749e00045 *vp90-2-02-size-18x32.webm
+34fbd7036752232d1663e70d7f7cdc93f7129202 *vp90-2-02-size-18x32.webm.md5
+f219655a639a774a2c9c0a9f45c28dc0b5e75e24 *vp90-2-02-size-18x34.webm
+2c4d622a9ea548791c1a07903d3702e9774388bb *vp90-2-02-size-18x34.webm.md5
+5308578da48c677d477a5404e19391d1303033c9 *vp90-2-02-size-18x64.webm
+e7fd4462527bac38559518ba80e41847db880f15 *vp90-2-02-size-18x64.webm.md5
+e109a7e013bd179f97e378542e1e81689ed06802 *vp90-2-02-size-18x66.webm
+45c04e422fb383c1f3be04beefaa4490e83bdb1a *vp90-2-02-size-18x66.webm.md5
+38844cae5d99caf445f7de33c3ae78494ce36c01 *vp90-2-02-size-32x08.webm
+ad018be39e493ca2405225034b1a5b7a42af6f3a *vp90-2-02-size-32x08.webm.md5
+7b57eaad55906f9de9903c8657a3fcb2aaf792ea *vp90-2-02-size-32x10.webm
+2294425d4e55d275af5e25a0beac9738a1b4ee73 *vp90-2-02-size-32x10.webm.md5
+f47ca2ced0d47f761bb0a5fdcd911d3f450fdcc1 *vp90-2-02-size-32x16.webm
+ae10981d93913f0ab1f28c1146255e01769aa8c0 *vp90-2-02-size-32x16.webm.md5
+08b23ad838b6cf1fbfe3ad7e7775d95573e815fc *vp90-2-02-size-32x18.webm
+1ba76f4c4a4ac7aabfa3ce195c1b473535eb7cc8 *vp90-2-02-size-32x18.webm.md5
+d5b88ae6c8c25c53dee74d9f1e6ca64244349a57 *vp90-2-02-size-32x32.webm
+e39c067a8ee2da52a51641eb1cb7f8eba935eb6b *vp90-2-02-size-32x32.webm.md5
+529429920dc36bd899059fa75a767f02c8c60874 *vp90-2-02-size-32x34.webm
+56888e7834f52b106e8911e3a7fc0f473b609995 *vp90-2-02-size-32x34.webm.md5
+38e848e160391c2b1a55040aadde613b9f4bf15e *vp90-2-02-size-32x64.webm
+8950485fb3f68b0e8be234db860e4ec5f5490fd0 *vp90-2-02-size-32x64.webm.md5
+5e8670f0b8ec9cefa8795b8959ffbe1a8e1aea94 *vp90-2-02-size-32x66.webm
+225df9d7d72ec711b0b60f4aeb65311c97db054a *vp90-2-02-size-32x66.webm.md5
+695f929e2ce6fb11a1f180322d46c5cb1c97fa61 *vp90-2-02-size-34x08.webm
+5bb4262030018dd01883965c6aa6070185924ef6 *vp90-2-02-size-34x08.webm.md5
+5adf74ec906d2ad3f7526e06bd29f5ad7d966a90 *vp90-2-02-size-34x10.webm
+71c100b437d3e8701632ae8d65c3555339b1c68f *vp90-2-02-size-34x10.webm.md5
+d0918923c987fba2d00193d83797b21289fe54aa *vp90-2-02-size-34x16.webm
+5d5a52f3535b4d2698dd3d87f4a13fdc9b57163d *vp90-2-02-size-34x16.webm.md5
+553ab0042cf87f5e668ec31b2e4b2a4b6ec196fd *vp90-2-02-size-34x18.webm
+a164c7f3c424987df2340496e6a8cf76e973f0f1 *vp90-2-02-size-34x18.webm.md5
+baf3e233634f150de81c18ba5d8848068e1c3c54 *vp90-2-02-size-34x32.webm
+22a79d3bd1c9b85dfe8c70bb2e19f08a92a8be03 *vp90-2-02-size-34x32.webm.md5
+6d50a533774a7167350e4a7ef43c94a5622179a2 *vp90-2-02-size-34x34.webm
+0c099638e79c273546523e06704553e42eb00b00 *vp90-2-02-size-34x34.webm.md5
+698cdd0a5e895cc202c488675e682a8c537ede4f *vp90-2-02-size-34x64.webm
+9317b63987cddab8389510a27b86f9f3d46e3fa5 *vp90-2-02-size-34x64.webm.md5
+4b5335ca06f082b6b69f584eb8e7886bdcafefd3 *vp90-2-02-size-34x66.webm
+e18d68b35428f46a84a947c646804a51ef1d7cec *vp90-2-02-size-34x66.webm.md5
+a54ae7b494906ec928a876e8290e5574f2f9f6a2 *vp90-2-02-size-64x08.webm
+87f9f7087b6489d45e9e4b38ede2c5aef4a4928f *vp90-2-02-size-64x08.webm.md5
+24522c70804a3c23d937df2d829ae63965b23f38 *vp90-2-02-size-64x10.webm
+447ce03938ab53bffcb4a841ee0bfaa90462dcb9 *vp90-2-02-size-64x10.webm.md5
+2a5035d035d214ae614af8051930690ef623989b *vp90-2-02-size-64x16.webm
+84e355761dd2e0361b904c84c52a0dd0384d89cf *vp90-2-02-size-64x16.webm.md5
+3a293ef4e270a19438e59b817fbe5f43eed4d36b *vp90-2-02-size-64x18.webm
+666824e5ba746779eb46079e0631853dcc86d48b *vp90-2-02-size-64x18.webm.md5
+ed32fae837095c9e8fc95d223ec68101812932c2 *vp90-2-02-size-64x32.webm
+97086eadedce1d0d9c072b585ba7b49aec69b1e7 *vp90-2-02-size-64x32.webm.md5
+696c7a7250bdfff594f4dfd88af34239092ecd00 *vp90-2-02-size-64x34.webm
+253a1d38d452e7826b086846c6f872f829c276bb *vp90-2-02-size-64x34.webm.md5
+fc508e0e3c2e6872c60919a60b812c5232e9c2b0 *vp90-2-02-size-64x64.webm
+2cd6ebeca0f82e9f505616825c07950371b905ab *vp90-2-02-size-64x64.webm.md5
+0f8a4fc1d6521187660425c283f08dff8c66e476 *vp90-2-02-size-64x66.webm
+5806be11a1d346be235f88d3683e69f73746166c *vp90-2-02-size-64x66.webm.md5
+273b0c36e3658685cde250408a478116d7ae92f1 *vp90-2-02-size-66x08.webm
+23c3cd0dca20a2f71f036e77ea92025ff4e7a298 *vp90-2-02-size-66x08.webm.md5
+4844c59c3306d1e671bb0568f00e344bf797e66e *vp90-2-02-size-66x10.webm
+e041eaf6841d775f8fde8bbb4949d2733fdaab7f *vp90-2-02-size-66x10.webm.md5
+bdf3f1582b234fcd2805ffec59f9d716a2345302 *vp90-2-02-size-66x16.webm
+2ec85ee18119e6798968571ea6e1b93ca386e3af *vp90-2-02-size-66x16.webm.md5
+0acce9af12b13b025d5274013da7ef6f568f075f *vp90-2-02-size-66x18.webm
+77c4d53e2a5c96b70af9d575fe6811e0f5ee627b *vp90-2-02-size-66x18.webm.md5
+682b36a25774bbdedcd603f504d18eb63f0167d4 *vp90-2-02-size-66x32.webm
+53728fae2a428f16d376a29f341a64ddca97996a *vp90-2-02-size-66x32.webm.md5
+e71b70e901e29eaa6672a6aa4f37f6f5faa02bd6 *vp90-2-02-size-66x34.webm
+f69a6a555e3f614b0a35f9bfc313d8ebb35bc725 *vp90-2-02-size-66x34.webm.md5
+4151b8c29452d5c2266397a7b9bf688899a2937b *vp90-2-02-size-66x64.webm
+69486e7fd9e380b6c97a03d3e167affc79f73840 *vp90-2-02-size-66x64.webm.md5
+68784a1ecac776fe2a3f230345af32f06f123536 *vp90-2-02-size-66x66.webm
+7f008c7f48d55e652fbd6bac405b51e0015c94f2 *vp90-2-02-size-66x66.webm.md5
+7e1bc449231ac1c5c2a11c9a6333b3e828763798 *vp90-2-03-size-196x196.webm
+6788a561466dace32d500194bf042e19cccc35e1 *vp90-2-03-size-196x196.webm.md5
+a170c9a88ec1dd854c7a471ff55fb2a97ac31870 *vp90-2-03-size-196x198.webm
+6bf9d6a8e2bdc5bf4f8a78071a3fed5ca02ad6f2 *vp90-2-03-size-196x198.webm.md5
+68f861d21c4c8b03d572c3d3fcd9f4fbf1f4503f *vp90-2-03-size-196x200.webm
+bbfc260b2bfd872cc6054272bb6b7f959a9e1c6e *vp90-2-03-size-196x200.webm.md5
+fc34889feeca2b7e5b27b4f1ce22d2e2b8e3e4b1 *vp90-2-03-size-196x202.webm
+158ee72af578f39aad0c3b8f4cbed2fc78b57e0f *vp90-2-03-size-196x202.webm.md5
+dd28fb7247af534bdf5e6795a3ac429610489a0b *vp90-2-03-size-196x208.webm
+7546be847efce2d1c0a23f807bfb03f91b764e1e *vp90-2-03-size-196x208.webm.md5
+41d5cf5ed65b722a1b6dc035e67f978ea8ffecf8 *vp90-2-03-size-196x210.webm
+9444fdf632d6a1b6143f4cb10fed8f63c1d67ec1 *vp90-2-03-size-196x210.webm.md5
+5007bc618143437c009d6dde5fc2e86f72d37dc2 *vp90-2-03-size-196x224.webm
+858361d8f79b44df5545feabbc9754ec9ede632f *vp90-2-03-size-196x224.webm.md5
+0bcbe357fbc776c3fa68e7117179574ed7564a44 *vp90-2-03-size-196x226.webm
+72006a5f42031a43d70a2cd9fc1958962a86628f *vp90-2-03-size-196x226.webm.md5
+000239f048cceaac055558e97ef07078ebf65502 *vp90-2-03-size-198x196.webm
+2d6841901b72000c5340f30be602853438c1b787 *vp90-2-03-size-198x196.webm.md5
+ae75b766306a6404c3b3b35a6b6d53633c14fbdb *vp90-2-03-size-198x198.webm
+3f2544b4f3b4b643a98f2c3b15ea5826fc702fa1 *vp90-2-03-size-198x198.webm.md5
+95ffd573fa84ccef1cd59e1583e6054f56a5c83d *vp90-2-03-size-198x200.webm
+5d537e3c9b9c54418c79677543454c4cda3de1af *vp90-2-03-size-198x200.webm.md5
+ecc845bf574375f469bc91bf5c75c79dc00073d6 *vp90-2-03-size-198x202.webm
+1b59f5e111265615a7a459eeda8cc9045178d228 *vp90-2-03-size-198x202.webm.md5
+432fb27144fe421b9f51cf44d2750a26133ed585 *vp90-2-03-size-198x208.webm
+a58a67f4fb357c73ca078aeecbc0f782975630b1 *vp90-2-03-size-198x208.webm.md5
+ff5058e7e6a47435046612afc8536f2040989e6f *vp90-2-03-size-198x210.webm
+18d3be7935e52217e2e9400b6f2c681a9e45dc89 *vp90-2-03-size-198x210.webm.md5
+a0d55263c1ed2c03817454dd4ec4090d36dbc864 *vp90-2-03-size-198x224.webm
+efa366a299817e2da51c00623b165aab9fbb8d91 *vp90-2-03-size-198x224.webm.md5
+ccd142fa2920fc85bb753f049160c1c353ad1574 *vp90-2-03-size-198x226.webm
+534524a0b2dbff852e0b92ef09939db072f83243 *vp90-2-03-size-198x226.webm.md5
+0d483b94ed40abc8ab6e49f960432ee54ad9c7f1 *vp90-2-03-size-200x196.webm
+41795f548181717906e7a504ba551f06c32102ae *vp90-2-03-size-200x196.webm.md5
+f6c2dc54e0989d50f01333fe40c91661fcbf849a *vp90-2-03-size-200x198.webm
+43df5d8c46a40089441392e6d096c588c1079a68 *vp90-2-03-size-200x198.webm.md5
+2f6e9df82e44fc145f0d9212dcccbed3de605e23 *vp90-2-03-size-200x200.webm
+757b2ef96b82093255725bab9690bbafe27f3caf *vp90-2-03-size-200x200.webm.md5
+40c5ea60415642a4a2e75c0d127b06309baadfab *vp90-2-03-size-200x202.webm
+3022c4a1c625b5dc04fdb1052d17d45b4171cfba *vp90-2-03-size-200x202.webm.md5
+6942ed5b27476bb8506d10e600d6ff60887780ca *vp90-2-03-size-200x208.webm
+c4ab8c66f3cf2dc8e8dd7abae9ac21f4d32cd6be *vp90-2-03-size-200x208.webm.md5
+71dbc99b83c49d1da45589b91eabb98e2f4a7b1e *vp90-2-03-size-200x210.webm
+3f0b40da7eef7974b9bc326562f251feb67d9c7c *vp90-2-03-size-200x210.webm.md5
+6b6b8489081cfefb377cc5f18eb754ec2383f655 *vp90-2-03-size-200x224.webm
+a259df2ac0e294492e3f9d4315baa34cab044f04 *vp90-2-03-size-200x224.webm.md5
+c9adc1c9bb07559349a0b054df4af56f7a6edbb9 *vp90-2-03-size-200x226.webm
+714cec61e3575581e4f1a0e3921f4dfdbbd316c5 *vp90-2-03-size-200x226.webm.md5
+f9bdc936bdf53f8be9ce78fecd41a21d31ff3943 *vp90-2-03-size-202x196.webm
+5b8e2e50fcea2c43b12fc067b8a9cc117af77bda *vp90-2-03-size-202x196.webm.md5
+c7b66ea3da87613deb47ff24a111247d3c384fec *vp90-2-03-size-202x198.webm
+517e91204b25586da943556f4adc5951c9be8bee *vp90-2-03-size-202x198.webm.md5
+935ef56b01cfdb4265a7e24696645209ccb20970 *vp90-2-03-size-202x200.webm
+55b8ec4a2513183144a8e27564596c06c7576fce *vp90-2-03-size-202x200.webm.md5
+849acf75e4f1d8d90046704e1103a18c64f30e35 *vp90-2-03-size-202x202.webm
+c79afc6660df2824e7df314e5bfd71f0d8acf76b *vp90-2-03-size-202x202.webm.md5
+17b3a4d55576b770626ccb856b9f1a6c8f6ae476 *vp90-2-03-size-202x208.webm
+0b887ff30409c58f2ccdc3bfacd6be7c69f8997a *vp90-2-03-size-202x208.webm.md5
+032d0ade4230fb2eef6d19915a7a1c9aa4a52617 *vp90-2-03-size-202x210.webm
+f78f8e79533c0c88dd2bfdcec9b1c07848568ece *vp90-2-03-size-202x210.webm.md5
+915a38c31fe425d5b93c837121cfa8082f5ea5bc *vp90-2-03-size-202x224.webm
+bf52a104074d0c5942aa7a5b31e11db47e43d48e *vp90-2-03-size-202x224.webm.md5
+be5cfde35666fa435e47d544d9258215beb1cf29 *vp90-2-03-size-202x226.webm
+2fa2f87502fda756b319389c8975204e130a2e3f *vp90-2-03-size-202x226.webm.md5
+15d908e97862b5b4bf295610df011fb9aa09909b *vp90-2-03-size-208x196.webm
+50c60792305d6a99be376dd596a6ff979325e6cc *vp90-2-03-size-208x196.webm.md5
+a367c7bc9fde56d6f4848cc573c7d4c1ce75e348 *vp90-2-03-size-208x198.webm
+be85fb2c8d435a75484231356f07d06ebddd13cd *vp90-2-03-size-208x198.webm.md5
+05fd46deb7288e7253742091f56e54a9a441a187 *vp90-2-03-size-208x200.webm
+74f8ec3b3a2fe81767ed1ab36a47bc0062d6223c *vp90-2-03-size-208x200.webm.md5
+d8985c4b386513a7385a4b3639bf91e469f1378b *vp90-2-03-size-208x202.webm
+0614a1e8d92048852adcf605a51333f5fabc7f03 *vp90-2-03-size-208x202.webm.md5
+28b002242238479165ba4fb87ee6b442c64b32e4 *vp90-2-03-size-208x208.webm
+37de5aca59bb900228400b0e115d3229edb9dcc0 *vp90-2-03-size-208x208.webm.md5
+c545be0050c2fad7c68427dbf86c62a739e94ab3 *vp90-2-03-size-208x210.webm
+d646eccb3cd578f94b54777e32b88898bef6e17a *vp90-2-03-size-208x210.webm.md5
+63a0cfe295b661026dd7b1bebb67acace1db766f *vp90-2-03-size-208x224.webm
+85c0361d93bf85a335248fef2767ff43eeef23db *vp90-2-03-size-208x224.webm.md5
+f911cc718d66e4fe8a865226088939c9eb1b7825 *vp90-2-03-size-208x226.webm
+a6d583a57876e7b7ec48625b2b2cdbcf70cab837 *vp90-2-03-size-208x226.webm.md5
+5bbb0f36da9a4683cf04e724124d8696332911bf *vp90-2-03-size-210x196.webm
+a3580fc7816d7fbcfb54fdba501cabbd06ba2f1d *vp90-2-03-size-210x196.webm.md5
+8db64d6f9ce36dd382013b42ae4e292deba697bc *vp90-2-03-size-210x198.webm
+eda20f8268c7f4147bead4059e9c4897e09140a9 *vp90-2-03-size-210x198.webm.md5
+ce391505eeaf1d12406563101cd6b2dbbbb44bfc *vp90-2-03-size-210x200.webm
+79d73b7f623082d2a00aa33e95c79d11c7d9c3a8 *vp90-2-03-size-210x200.webm.md5
+852db6fdc206e72391fc69b807f1954934679949 *vp90-2-03-size-210x202.webm
+f69414c5677ed2f2b8b37ae76429e509a92276a5 *vp90-2-03-size-210x202.webm.md5
+c424cc3edd2308da7d33f27acb36b54db5bf2595 *vp90-2-03-size-210x208.webm
+27b18562faa1b3184256f4eae8114b539b3e9d3e *vp90-2-03-size-210x208.webm.md5
+dd029eba719d50a2851592fa8b9b2efe88904930 *vp90-2-03-size-210x210.webm
+c853a1670465eaa04ca31b3511995f1b6ed4f58f *vp90-2-03-size-210x210.webm.md5
+d962e8ae676c54d0c3ea04ec7c04b37ae6a786e3 *vp90-2-03-size-210x224.webm
+93b793e79d987065b39ad8e2e71244368435fc25 *vp90-2-03-size-210x224.webm.md5
+3d0825fe83bcc125be1f78145ff43ca6d7588784 *vp90-2-03-size-210x226.webm
+5230f31a57ca3b5311698a12035d2644533b3ec4 *vp90-2-03-size-210x226.webm.md5
+6622f8bd9279e1ce45509a58a31a990052d45e14 *vp90-2-03-size-224x196.webm
+65411da07f60113f2be05c807879072b161d561e *vp90-2-03-size-224x196.webm.md5
+6744ff2ee2c41eb08c62ff30880833b6d77b585b *vp90-2-03-size-224x198.webm
+46ea3641d41acd4bff347b224646c060d5620385 *vp90-2-03-size-224x198.webm.md5
+8eb91f3416a1404705f370caecd74b2b458351b1 *vp90-2-03-size-224x200.webm
+196aefb854c8b95b9330263d6690b7ee15693ecf *vp90-2-03-size-224x200.webm.md5
+256a5a23ef4e6d5ef2871af5afb8cd13d28cec00 *vp90-2-03-size-224x202.webm
+840ad8455dcf2be378c14b007e66fa642fc8196d *vp90-2-03-size-224x202.webm.md5
+db4606480ab48b96c9a6ff5e639f1f1aea2a12e4 *vp90-2-03-size-224x208.webm
+40b9801d5620467499ac70fa6b7c40aaa5e1c331 *vp90-2-03-size-224x208.webm.md5
+e37159e687fe1cb24cffddfae059301adbaf4212 *vp90-2-03-size-224x210.webm
+1e4acd4b6334ae260c3eed08652d0ba8122073f2 *vp90-2-03-size-224x210.webm.md5
+0de1eb4bb6285ae621e4f2b613d2aa4a8c95a130 *vp90-2-03-size-224x224.webm
+37db449ad86fb286c2c02d94aa8fe0379c05044a *vp90-2-03-size-224x224.webm.md5
+32ebbf903a7d7881bcfe59639f1d472371f3bf27 *vp90-2-03-size-224x226.webm
+5cc3ac5dc9f6912491aa2ddac863f8187f34c569 *vp90-2-03-size-224x226.webm.md5
+9480ff5c2c32b1870ac760c87514912616e6cf01 *vp90-2-03-size-226x196.webm
+fe83655c0f1888f0af7b047785f01ba7ca9f1324 *vp90-2-03-size-226x196.webm.md5
+09cad4221996315cdddad4e502dbfabf53ca1d6a *vp90-2-03-size-226x198.webm
+e3ddfdc650acb95adb45abd9b634e1f09ea8ac96 *vp90-2-03-size-226x198.webm.md5
+c34f49d55fe39e3f0b607e3cc95e30244225cecb *vp90-2-03-size-226x200.webm
+abb83edc868a3523ccd4e5523fac2efbe7c3df1f *vp90-2-03-size-226x200.webm.md5
+d17bc08eedfc60c4c23d576a6c964a21bf854d1f *vp90-2-03-size-226x202.webm
+1d22d2d0f375251c2d5a1acb4714bc35d963865b *vp90-2-03-size-226x202.webm.md5
+9bd537c4f92a25596ccd29fedfe181feac948b92 *vp90-2-03-size-226x208.webm
+6feb0e7325386275719f3511ada9e248a2ae7df4 *vp90-2-03-size-226x208.webm.md5
+4487067f6cedd495b93696b44b37fe0a3e7eda14 *vp90-2-03-size-226x210.webm
+49a8fa87945f47208168d541c068e78d878075d5 *vp90-2-03-size-226x210.webm.md5
+559fea2f8da42b33c1aa1dbc34d1d6781009847a *vp90-2-03-size-226x224.webm
+83c6d8f2969b759e10e5c6542baca1265c874c29 *vp90-2-03-size-226x224.webm.md5
+fe0af2ee47b1e5f6a66db369e2d7e9d870b38dce *vp90-2-03-size-226x226.webm
+94ad19b8b699cea105e2ff18f0df2afd7242bcf7 *vp90-2-03-size-226x226.webm.md5
+52bc1dfd3a97b24d922eb8a31d07527891561f2a *vp90-2-03-size-352x288.webm
+3084d6d0a1eec22e85a394422fbc8faae58930a5 *vp90-2-03-size-352x288.webm.md5
+b6524e4084d15b5d0caaa3d3d1368db30cbee69c *vp90-2-03-deltaq.webm
+65f45ec9a55537aac76104818278e0978f94a678 *vp90-2-03-deltaq.webm.md5
+4dbb87494c7f565ffc266c98d17d0d8c7a5c5aba *vp90-2-05-resize.ivf
+7f6d8879336239a43dbb6c9f13178cb11cf7ed09 *vp90-2-05-resize.ivf.md5
+bf61ddc1f716eba58d4c9837d4e91031d9ce4ffe *vp90-2-06-bilinear.webm
+f6235f937552e11d8eb331ec55da6b3aa596b9ac *vp90-2-06-bilinear.webm.md5
+0c83a1e414fde3bccd6dc451bbaee68e59974c76 *vp90-2-07-frame_parallel.webm
+e5c2c9fb383e5bf3b563480adaeba5b7e3475ecd *vp90-2-07-frame_parallel.webm.md5
+086c7edcffd699ae7d99d710fd7e53b18910ca5b *vp90-2-08-tile_1x2_frame_parallel.webm
+e981ecaabb29a80e0cbc1f4002384965ce8e95bb *vp90-2-08-tile_1x2_frame_parallel.webm.md5
+ed79be026a6f28646c5825da1c12d1fbc70f96a4 *vp90-2-08-tile_1x2.webm
+45b404e025841c9750895fc1a9f6bd384fe6a315 *vp90-2-08-tile_1x2.webm.md5
+cf8ea970c776797aae71dac8317ea926d9431cab *vp90-2-08-tile_1x4_frame_parallel.webm
+a481fbea465010b57af5a19ebf6d4a5cfe5b9278 *vp90-2-08-tile_1x4_frame_parallel.webm.md5
+0203ec456277a01aec401e7fb6c72c9a7e5e3f9d *vp90-2-08-tile_1x4.webm
+c9b237dfcc01c1b414fbcaa481d014a906ef7998 *vp90-2-08-tile_1x4.webm.md5
+20c75157e91ab41f82f70ffa73d5d01df8469287 *vp90-2-08-tile-4x4.webm
+ae7451810247fd13975cc257aa0301ff17102255 *vp90-2-08-tile-4x4.webm.md5
+2ec6e15422ac7a61af072dc5f27fcaf1942ce116 *vp90-2-08-tile-4x1.webm
+0094f5ee5e46345017c30e0aa4835b550212d853 *vp90-2-08-tile-4x1.webm.md5
+edea45dac4a3c2e5372339f8851d24c9bef803d6 *vp90-2-09-subpixel-00.ivf
+5428efc4bf92191faedf4a727fcd1d94966a7abc *vp90-2-09-subpixel-00.ivf.md5
+8cdd435d89029987ee196896e21520e5f879f04d *vp90-2-bbb_1280x720_tile_1x4_1310kbps.webm
+091b373aa2ecb59aa5c647affd5bcafcc7547364 *vp90-2-bbb_1920x1080_tile_1x1_2581kbps.webm
+87ee28032b0963a44b73a850fcc816a6dc83efbb *vp90-2-bbb_1920x1080_tile_1x4_2586kbps.webm
+c6ce25c4bfd4bdfc2932b70428e3dfe11210ec4f *vp90-2-bbb_1920x1080_tile_1x4_fpm_2304kbps.webm
+2064bdb22aa71c2691e0469fb62e8087a43f08f8 *vp90-2-bbb_426x240_tile_1x1_180kbps.webm
+8080eda22694910162f0996e8a962612f381a57f *vp90-2-bbb_640x360_tile_1x2_337kbps.webm
+a484b335c27ea189c0f0d77babea4a510ce12d50 *vp90-2-bbb_854x480_tile_1x2_651kbps.webm
+3eacf1f006250be4cc5c92a7ef146e385ee62653 *vp90-2-sintel_1280x546_tile_1x4_1257kbps.webm
+217f089a16447490823127b36ce0d945522accfd *vp90-2-sintel_1920x818_tile_1x4_fpm_2279kbps.webm
+eedb3c641e60dacbe082491a16df529a5c9187df *vp90-2-sintel_426x182_tile_1x1_171kbps.webm
+cb7e4955af183dff33bcba0c837f0922ab066400 *vp90-2-sintel_640x272_tile_1x2_318kbps.webm
+48613f9380e2580002f8a09d6e412ea4e89a52b9 *vp90-2-sintel_854x364_tile_1x2_621kbps.webm
+990a91f24dd284562d21d714ae773dff5452cad8 *vp90-2-tos_1280x534_tile_1x4_1306kbps.webm
+aa402217577a659cfc670157735b4b8e9aa670fe *vp90-2-tos_1280x534_tile_1x4_fpm_952kbps.webm
+b6dd558c90bca466b4bcbd03b3371648186465a7 *vp90-2-tos_1920x800_tile_1x4_fpm_2335kbps.webm
+1a9c2914ba932a38f0a143efc1ad0e318e78888b *vp90-2-tos_426x178_tile_1x1_181kbps.webm
+a3d2b09f24debad4747a1b3066f572be4273bced *vp90-2-tos_640x266_tile_1x2_336kbps.webm
+c64b03b5c090e6888cb39685c31f00a6b79fa45c *vp90-2-tos_854x356_tile_1x2_656kbps.webm
+94b533dbcf94292001e27cc51fec87f9e8c90c0b *vp90-2-tos_854x356_tile_1x2_fpm_546kbps.webm
+0e7cd4135b231c9cea8d76c19f9e84b6fd77acec *vp90-2-08-tile_1x8_frame_parallel.webm
+c9b6850af28579b031791066457f4cb40df6e1c7 *vp90-2-08-tile_1x8_frame_parallel.webm.md5
+e448b6e83490bca0f8d58b4f4b1126a17baf4b0c *vp90-2-08-tile_1x8.webm
+5e524165f0397e6141d914f4f0a66267d7658376 *vp90-2-08-tile_1x8.webm.md5
+a34e14923d6d17b1144254d8187d7f85b700a63c *vp90-2-02-size-lf-1920x1080.webm
+e3b28ddcfaeb37fb4d132b93f92642a9ad17c22d *vp90-2-02-size-lf-1920x1080.webm.md5
+d48c5db1b0f8e60521a7c749696b8067886033a3 *vp90-2-09-aq2.webm
+84c1599298aac78f2fc05ae2274575d10569dfa0 *vp90-2-09-aq2.webm.md5
+55fc55ed73d578ed60fad05692579873f8bad758 *vp90-2-09-lf_deltas.webm
+54638c38009198c38c8f3b25c182b709b6c1fd2e *vp90-2-09-lf_deltas.webm.md5
+510d95f3beb3b51c572611fdaeeece12277dac30 *vp90-2-10-show-existing-frame.webm
+14d631096f4bfa2d71f7f739aec1448fb3c33bad *vp90-2-10-show-existing-frame.webm.md5
+d2feea7728e8d2c615981d0f47427a4a5a45d881 *vp90-2-10-show-existing-frame2.webm
+5f7c7811baa3e4f03be1dd78c33971b727846821 *vp90-2-10-show-existing-frame2.webm.md5
+b4318e75f73a6a08992c7326de2fb589c2a794c7 *vp90-2-11-size-351x287.webm
+b3c48382cf7d0454e83a02497c229d27720f9e20 *vp90-2-11-size-351x287.webm.md5
+8e0096475ea2535bac71d3e2fc09e0c451c444df *vp90-2-11-size-351x288.webm
+19e003804ec1dfc5464813b32339a15d5ba7b42f *vp90-2-11-size-351x288.webm.md5
+40cd1d6a188d7a88b21ebac1e573d3f270ab261e *vp90-2-11-size-352x287.webm
+68f515abe3858fc1eded46c8e6b2f727d43b5331 *vp90-2-11-size-352x287.webm.md5
+9a510769ff23db410880ec3029d433e87d17f7fc *vp90-2-12-droppable_1.ivf
+952eaac6eefa6f62179ed1db3e922fd42fecc624 *vp90-2-12-droppable_1.ivf.md5
+9a510769ff23db410880ec3029d433e87d17f7fc *vp90-2-12-droppable_2.ivf
+92a756469fa438220524e7fa6ac1d38c89514d17 *vp90-2-12-droppable_2.ivf.md5
+c21e97e4ba486520118d78b01a5cb6e6dc33e190 *vp90-2-12-droppable_3.ivf
+601abc9e4176c70f82ac0381365e9b151fdd24cd *vp90-2-12-droppable_3.ivf.md5
+61c640dad23cd4f7ad811b867e7b7e3521f4e3ba *vp90-2-13-largescaling.webm
+bca1b02eebdb088fa3f389fe0e7571e75a71f523 *vp90-2-13-largescaling.webm.md5
+c740708fa390806eebaf669909c1285ab464f886 *vp90-2-14-resize-fp-tiles-1-2.webm
+c7b85ffd8e11500f73f52e7dc5a47f57c393d47f *vp90-2-14-resize-fp-tiles-1-2.webm.md5
+ec8faa352a08f7033c60f29f80d505e2d7daa103 *vp90-2-14-resize-fp-tiles-1-4.webm
+6852c783fb421bda5ded3d4c5a3ffc46de03fbc1 *vp90-2-14-resize-fp-tiles-1-4.webm.md5
+8af61853ac0d07c4cb5bf7c2016661ba350b3497 *vp90-2-14-resize-fp-tiles-1-8.webm
+571353bac89fea60b5706073409aa3c0d42aefe9 *vp90-2-14-resize-fp-tiles-1-8.webm.md5
+b1c187ed69931496b82ec194017a79831bafceef *vp90-2-14-resize-fp-tiles-1-16.webm
+1c199a41afe42ce303944d70089eaaa2263b4a09 *vp90-2-14-resize-fp-tiles-1-16.webm.md5
+8eaae5a6f2dff934610b0c7a917d7f583ba74aa5 *vp90-2-14-resize-fp-tiles-2-1.webm
+db18fcf915f7ffaea6c39feab8bda6c1688af011 *vp90-2-14-resize-fp-tiles-2-1.webm.md5
+bc3046d138941e2a20e9ceec0ff6d25c25d12af3 *vp90-2-14-resize-fp-tiles-4-1.webm
+393211b808030d09a79927b17a4374b2f68a60ae *vp90-2-14-resize-fp-tiles-4-1.webm.md5
+6e8f8e31721a0f7f68a2964e36e0e698c2e276b1 *vp90-2-14-resize-fp-tiles-8-1.webm
+491fd3cd78fb0577bfe905bb64bbf64bd7d29140 *vp90-2-14-resize-fp-tiles-8-1.webm.md5
+cc5958da2a7edf739cd2cfeb18bd05e77903087e *vp90-2-14-resize-fp-tiles-16-1.webm
+0b58daf55aaf9063bf5b4fb33393d18b417dc428 *vp90-2-14-resize-fp-tiles-16-1.webm.md5
+821eeecc9d8c6a316134dd42d1ff057787d8047b *vp90-2-14-resize-fp-tiles-2-4.webm
+374c549f2839a3d0b732c4e3650700144037e76c *vp90-2-14-resize-fp-tiles-2-4.webm.md5
+dff8c8e49aacea9f4c7f22cb882da984e2a1b405 *vp90-2-14-resize-fp-tiles-2-8.webm
+e5b8820a7c823b21297d6e889e57ec401882c210 *vp90-2-14-resize-fp-tiles-2-8.webm.md5
+77629e4b23e32896aadf6e994c78bd4ffa1c7797 *vp90-2-14-resize-fp-tiles-2-16.webm
+1937f5df032664ac345d4613ad4417b4967b1230 *vp90-2-14-resize-fp-tiles-2-16.webm.md5
+380ba5702bb1ec7947697314ab0300b5c56a1665 *vp90-2-14-resize-fp-tiles-4-2.webm
+fde7b30d2aa64c1e851a4852f655d79fc542cf66 *vp90-2-14-resize-fp-tiles-4-2.webm.md5
+dc784b258ffa2abc2ae693d11792acf0bb9cb74f *vp90-2-14-resize-fp-tiles-8-2.webm
+edf26f0130aeee8342d49c2c8f0793ad008782d9 *vp90-2-14-resize-fp-tiles-8-2.webm.md5
+8e575789fd63ebf69e8eff1b9a4351a249a73bee *vp90-2-14-resize-fp-tiles-16-2.webm
+b6415318c1c589a1f64b9d569ce3cabbec2e0d52 *vp90-2-14-resize-fp-tiles-16-2.webm.md5
+e3adc944a11c4c5517e63664c84ebb0847b64d81 *vp90-2-14-resize-fp-tiles-4-8.webm
+03cba0532bc90a05b1990db830bf5701e24e7982 *vp90-2-14-resize-fp-tiles-4-8.webm.md5
+3b27a991eb6d78dce38efab35b7db682e8cbbee3 *vp90-2-14-resize-fp-tiles-4-16.webm
+5d16b7f82bf59f802724ddfd97abb487150b1c9d *vp90-2-14-resize-fp-tiles-4-16.webm.md5
+d5fed8c28c1d4c7e232ebbd25cf758757313ed96 *vp90-2-14-resize-fp-tiles-8-4.webm
+5a8ff8a52cbbde7bfab569beb6d971c5f8b904f7 *vp90-2-14-resize-fp-tiles-8-4.webm.md5
+17a5faa023d77ee9dad423a4e0d3145796bbc500 *vp90-2-14-resize-fp-tiles-16-4.webm
+2ef8daa3c3e750fd745130d0a76a39fe86f0448f *vp90-2-14-resize-fp-tiles-16-4.webm.md5
+9361e031f5cc990d8740863e310abb5167ae351e *vp90-2-14-resize-fp-tiles-8-16.webm
+57f13a2197486584f4e1a4f82ad969f3abc5a1a2 *vp90-2-14-resize-fp-tiles-8-16.webm.md5
+5803fc6fcbfb47b7661f3fcc6499158a32b56675 *vp90-2-14-resize-fp-tiles-16-8.webm
+be0fe64a1a4933696ff92d93f9bdecdbd886dc13 *vp90-2-14-resize-fp-tiles-16-8.webm.md5
+0ac0f6d20a0afed77f742a3b9acb59fd7b9cb093 *vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm
+1765315acccfe6cd12230e731369fcb15325ebfa *vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm.md5
+4a2b7a683576fe8e330c7d1c4f098ff4e70a43a8 *vp90-2-14-resize-fp-tiles-16-8-4-2-1.webm
+1ef480392112b3509cb190afbb96f9a38dd9fbac *vp90-2-14-resize-fp-tiles-16-8-4-2-1.webm.md5
+e615575ded499ea1d992f3b38e3baa434509cdcd *vp90-2-15-segkey.webm
+e3ab35d4316c5e81325c50f5236ceca4bc0d35df *vp90-2-15-segkey.webm.md5
+9b7ca2cac09d34c4a5d296c1900f93b1e2f69d0d *vp90-2-15-segkey_adpq.webm
+8f46ba5f785d0c2170591a153e0d0d146a7c8090 *vp90-2-15-segkey_adpq.webm.md5
+698a6910a97486b833073ef0c0b18d75dce57ee8 *vp90-2-16-intra-only.webm
+5661b0168752969f055eec37b05fa9fa947dc7eb *vp90-2-16-intra-only.webm.md5
+c01bb7938f9a9f25e0c37afdec2f2fb73b6cc7fa *vp90-2-17-show-existing-frame.webm
+cc75f351818b9a619818f5cc77b9bc013d0c1e11 *vp90-2-17-show-existing-frame.webm.md5
+013708bd043f0821a3e56fb8404d82e7a0c7af6c *vp91-2-04-yuv422.webm
+1e58a7d23adad830a672f1733c9d2ae17890d59c *vp91-2-04-yuv422.webm.md5
+25d78f28948789d159a9453ebc13048b818251b1 *vp91-2-04-yuv440.webm
+81b3870b27a7f695ef6a43e87ab04bbdb5aee2f5 *vp91-2-04-yuv440.webm.md5
+0321d507ce62dedc8a51b4e9011f7a19aed9c3dc *vp91-2-04-yuv444.webm
+367e423dd41fdb49aa028574a2cfec5c2f325c5c *vp91-2-04-yuv444.webm.md5
+f77673b566f686853adefe0c578ad251b7241281 *vp92-2-20-10bit-yuv420.webm
+abdedfaddacbbe1a15ac7a54e86360f03629fb7a *vp92-2-20-10bit-yuv420.webm.md5
+0c2c355a1b17b28537c5a3b19997c8783b69f1af *vp92-2-20-12bit-yuv420.webm
+afb2c2798703e039189b0a15c8ac5685aa51d33f *vp92-2-20-12bit-yuv420.webm.md5
+0d661bc6e83da33238981481efd1b1802d323d88 *vp93-2-20-10bit-yuv422.webm
+10318907063db22eb02fad332556edbbecd443cc *vp93-2-20-10bit-yuv422.webm.md5
+ebc6be2f7511a0bdeac0b18c67f84ba7168839c7 *vp93-2-20-12bit-yuv422.webm
+235232267c6a1dc8a11e45d600f1c99d2f8b42d4 *vp93-2-20-12bit-yuv422.webm.md5
+f76b11b26d4beaceac7a7e7729dd5054d095164f *vp93-2-20-10bit-yuv440.webm
+757b33b5ac969c5999999488a731a3d1e6d9fb88 *vp93-2-20-10bit-yuv440.webm.md5
+df8807dbd29bec795c2db9c3c18e511fbb988101 *vp93-2-20-12bit-yuv440.webm
+ea4100930c3f59a1c23fbb33ab0ea01151cae159 *vp93-2-20-12bit-yuv440.webm.md5
+189c1b5f404ff41a50a7fc96341085ad541314a9 *vp93-2-20-10bit-yuv444.webm
+2dd0177c2f9d970b6e698892634c653630f91f40 *vp93-2-20-10bit-yuv444.webm.md5
+bd44cf6e1c27343e3639df9ac21346aedd5d6973 *vp93-2-20-12bit-yuv444.webm
+f36e5bdf5ec3213f32c0ddc82f95d82c5133bf27 *vp93-2-20-12bit-yuv444.webm.md5
+eb438c6540eb429f74404eedfa3228d409c57874 *desktop_640_360_30.yuv
+89e70ebd22c27d275fe14dc2f1a41841a6d8b9ab *kirland_640_480_30.yuv
+33c533192759e5bb4f07abfbac389dc259db4686 *macmarcomoving_640_480_30.yuv
+8bfaab121080821b8f03b23467911e59ec59b8fe *macmarcostationary_640_480_30.yuv
+70894878d916a599842d9ad0dcd24e10c13e5467 *niklas_640_480_30.yuv
+8784b6df2d8cc946195a90ac00540500d2e522e4 *tacomanarrows_640_480_30.yuv
+edd86a1f5e62fd9da9a9d46078247759c2638009 *tacomasmallcameramovement_640_480_30.yuv
+9a70e8b7d14fba9234d0e51dce876635413ce444 *thaloundeskmtg_640_480_30.yuv
+e7d315dbf4f3928779e0dc624311196d44491d32 *niklas_1280_720_30.yuv
+c77e4a26616add298a05dd5d12397be22c0e40c5 *vp90-2-18-resize.ivf
+c12918cf0a716417fba2de35c3fc5ab90e52dfce *vp90-2-18-resize.ivf.md5
+717da707afcaa1f692ff1946f291054eb75a4f06 *screendata.y4m
+b7c1296630cdf1a7ef493d15ff4f9eb2999202f6 *invalid-vp90-2-08-tile_1x2_frame_parallel.webm.ivf.s47039_r01-05_b6-.ivf
+0a3884edb3fd8f9d9b500223e650f7de257b67d8 *invalid-vp90-2-08-tile_1x2_frame_parallel.webm.ivf.s47039_r01-05_b6-.ivf.res
+359e138dfb66863828397b77000ea7a83c844d02 *invalid-vp90-2-08-tile_1x8_frame_parallel.webm.ivf.s288_r01-05_b6-.ivf
+bbd33de01c17b165b4ce00308e8a19a942023ab8 *invalid-vp90-2-08-tile_1x8_frame_parallel.webm.ivf.s288_r01-05_b6-.ivf.res
+fac89b5735be8a86b0dc05159f996a5c3208ae32 *invalid-vp90-2-09-aq2.webm.ivf.s3984_r01-05_b6-.v2.ivf
+0a3884edb3fd8f9d9b500223e650f7de257b67d8 *invalid-vp90-2-09-aq2.webm.ivf.s3984_r01-05_b6-.v2.ivf.res
+4506dfdcdf8ee4250924b075a0dcf1f070f72e5a *invalid-vp90-2-09-subpixel-00.ivf.s19552_r01-05_b6-.v2.ivf
+bcdedaf168ac225575468fda77502d2dc9fd5baa *invalid-vp90-2-09-subpixel-00.ivf.s19552_r01-05_b6-.v2.ivf.res
+65e93f9653bcf65b022f7d225268d1a90a76e7bb *vp90-2-19-skip.webm
+368dccdde5288c13c25695d2eacdc7402cadf613 *vp90-2-19-skip.webm.md5
+ffe460282df2b0e7d4603c2158653ad96f574b02 *vp90-2-19-skip-01.webm
+bd21bc9eda4a4a36b221d71ede3a139fc3c7bd85 *vp90-2-19-skip-01.webm.md5
+178f5bd239e38cc1cc2657a7a5e1a9f52ad2d3fe *vp90-2-19-skip-02.webm
+9020d5e260bd7df08e2b3d4b86f8623cee3daea2 *vp90-2-19-skip-02.webm.md5
+b03c408cf23158638da18dbc3323b99a1635c68a *invalid-vp90-2-12-droppable_1.ivf.s3676_r01-05_b6-.ivf
+0a3884edb3fd8f9d9b500223e650f7de257b67d8 *invalid-vp90-2-12-droppable_1.ivf.s3676_r01-05_b6-.ivf.res
+5e67e24e7f53fd189e565513cef8519b1bd6c712 *invalid-vp90-2-05-resize.ivf.s59293_r01-05_b6-.ivf
+741158f67c0d9d23726624d06bdc482ad368afc9 *invalid-vp90-2-05-resize.ivf.s59293_r01-05_b6-.ivf.res
+8b1f7bf7e86c0976d277f60e8fcd9539e75a079a *invalid-vp90-2-09-subpixel-00.ivf.s20492_r01-05_b6-.v2.ivf
+9c6bdf048fb2e66f07d4b4db5b32e6f303bd6109 *invalid-vp90-2-09-subpixel-00.ivf.s20492_r01-05_b6-.v2.ivf.res
+552e372e9b78127389fb06b34545df2cec15ba6d *invalid-vp91-2-mixedrefcsp-444to420.ivf
+a61774cf03fc584bd9f0904fc145253bb8ea6c4c *invalid-vp91-2-mixedrefcsp-444to420.ivf.res
+812d05a64a0d83c1b504d0519927ddc5a2cdb273 *invalid-vp90-2-12-droppable_1.ivf.s73804_r01-05_b6-.ivf
+1e472baaf5f6113459f0399a38a5a5e68d17799d *invalid-vp90-2-12-droppable_1.ivf.s73804_r01-05_b6-.ivf.res
+f97088c7359fc8d3d5aa5eafe57bc7308b3ee124 *vp90-2-20-big_superframe-01.webm
+47d7d409785afa33b123376de0c907336e6c7bd7 *vp90-2-20-big_superframe-01.webm.md5
+65ade6d2786209582c50d34cfe22b3cdb033abaf *vp90-2-20-big_superframe-02.webm
+7c0ed8d04c4d06c5411dd2e5de2411d37f092db5 *vp90-2-20-big_superframe-02.webm.md5
+667ec8718c982aef6be07eb94f083c2efb9d2d16 *vp90-2-07-frame_parallel-1.webm
+bfc82bf848e9c05020d61e3ffc1e62f25df81d19 *vp90-2-07-frame_parallel-1.webm.md5
+efd5a51d175cfdacd169ed23477729dc558030dc *invalid-vp90-2-07-frame_parallel-1.webm
+9f912712ec418be69adb910e2ca886a63c4cec08 *invalid-vp90-2-07-frame_parallel-2.webm
+445f5a53ca9555341852997ccdd480a51540bd14 *invalid-vp90-2-07-frame_parallel-3.webm
+d18c90709a0d03c82beadf10898b27d88fff719c *invalid-vp90-2-03-size-224x196.webm.ivf.s44156_r01-05_b6-.ivf
+d06285d109ecbaef63b0cbcc44d70a129186f51c *invalid-vp90-2-03-size-224x196.webm.ivf.s44156_r01-05_b6-.ivf.res
+e60d859b0ef2b331b21740cf6cb83fabe469b079 *invalid-vp90-2-03-size-202x210.webm.ivf.s113306_r01-05_b6-.ivf
+0ae808dca4d3c1152a9576e14830b6faa39f1b4a *invalid-vp90-2-03-size-202x210.webm.ivf.s113306_r01-05_b6-.ivf.res
+9cfc855459e7549fd015c79e8eca512b2f2cb7e3 *niklas_1280_720_30.y4m
+5b5763b388b1b52a81bb82b39f7ec25c4bd3d0e1 *desktop_credits.y4m
+85771f6ab44e4a0226e206c0cde8351dd5918953 *vp90-2-02-size-130x132.webm
+512dad5eabbed37b4bbbc64ce153f1a5484427b8 *vp90-2-02-size-130x132.webm.md5
+01f7127d40360289db63b27f61cb9afcda350e95 *vp90-2-02-size-132x130.webm
+4a94275328ae076cf60f966c097a8721010fbf5a *vp90-2-02-size-132x130.webm.md5
+f41c0400b5716b4b70552c40dd03d44be131e1cc *vp90-2-02-size-132x132.webm
+1a69e989f697e424bfe3e3e8a77bb0c0992c8e47 *vp90-2-02-size-132x132.webm.md5
+94a5cbfacacba100e0c5f7861c72a1b417feca0f *vp90-2-02-size-178x180.webm
+dedfecf1d784bcf70629592fa5e6f01d5441ccc9 *vp90-2-02-size-178x180.webm.md5
+4828b62478c04014bba3095a83106911a71cf387 *vp90-2-02-size-180x178.webm
+423da2b861050c969d78ed8e8f8f14045d1d8199 *vp90-2-02-size-180x178.webm.md5
+338f7c9282f43e29940f5391118aadd17e4f9234 *vp90-2-02-size-180x180.webm
+6c2ef013392310778dca5dd5351160eca66b0a60 *vp90-2-02-size-180x180.webm.md5
+679fa7d6807e936ff937d7b282e7dbd8ac76447e *vp90-2-14-resize-10frames-fp-tiles-1-2-4-8.webm
+fc7267ab8fc2bf5d6c234e34ee6c078a967b4888 *vp90-2-14-resize-10frames-fp-tiles-1-2-4-8.webm.md5
+9d33a137c819792209c5ce4e4e1ee5da73d574fe *vp90-2-14-resize-10frames-fp-tiles-1-2.webm
+0c78a154956a8605d050bdd75e0dcc4d39c040a6 *vp90-2-14-resize-10frames-fp-tiles-1-2.webm.md5
+d6a8d8c57f66a91d23e8e7df480f9ae841e56c37 *vp90-2-14-resize-10frames-fp-tiles-1-4.webm
+e9b4e8c7b33b5fda745d340c3f47e6623ae40cf2 *vp90-2-14-resize-10frames-fp-tiles-1-4.webm.md5
+aa6fe043a0c4a42b49c87ebbe812d4afd9945bec *vp90-2-14-resize-10frames-fp-tiles-1-8.webm
+028520578994c2d013d4c0129033d4f2ff31bbe0 *vp90-2-14-resize-10frames-fp-tiles-1-8.webm.md5
+d1d5463c9ea7b5cc5f609ddedccddf656f348d1a *vp90-2-14-resize-10frames-fp-tiles-2-1.webm
+92d5872f5bdffbed721703b7e959b4f885e3d77a *vp90-2-14-resize-10frames-fp-tiles-2-1.webm.md5
+677cb29de1215d97346015af5807a9b1faad54cf *vp90-2-14-resize-10frames-fp-tiles-2-4.webm
+a5db19f977094ec3fd60b4f7671b3e6740225e12 *vp90-2-14-resize-10frames-fp-tiles-2-4.webm.md5
+cdd3c52ba21067efdbb2de917fe2a965bf27332e *vp90-2-14-resize-10frames-fp-tiles-2-8.webm
+db17ec5d894ea8b8d0b7f32206d0dd3d46dcfa6d *vp90-2-14-resize-10frames-fp-tiles-2-8.webm.md5
+0f6093c472125d05b764d7d1965c1d56771c0ea2 *vp90-2-14-resize-10frames-fp-tiles-4-1.webm
+bc7c79e1bee07926dd970462ce6f64fc30eec3e1 *vp90-2-14-resize-10frames-fp-tiles-4-1.webm.md5
+c5142e2bff4091338196c8ea8bc9266e64f548bc *vp90-2-14-resize-10frames-fp-tiles-4-2.webm
+22aa3dd430b69fd3d92f6561bac86deeed90486d *vp90-2-14-resize-10frames-fp-tiles-4-2.webm.md5
+ede8b1466d2f26e1b1bd9602addb9cd1017e1d8c *vp90-2-14-resize-10frames-fp-tiles-4-8.webm
+508d5ebb9c0eac2a4100281a3ee052ec2fc19217 *vp90-2-14-resize-10frames-fp-tiles-4-8.webm.md5
+2b292e3392854cd1d76ae597a6f53656cf741cfa *vp90-2-14-resize-10frames-fp-tiles-8-1.webm
+1c24e54fa19e94e1722f24676404444e941c3d31 *vp90-2-14-resize-10frames-fp-tiles-8-1.webm.md5
+61beda21064e09634564caa6697ab90bd53c9af7 *vp90-2-14-resize-10frames-fp-tiles-8-2.webm
+9c0657b4d9e1d0e4c9d28a90e5a8630a65519124 *vp90-2-14-resize-10frames-fp-tiles-8-2.webm.md5
+1758c50a11a7c92522749b4a251664705f1f0d4b *vp90-2-14-resize-10frames-fp-tiles-8-4-2-1.webm
+4f454a06750614314ae15a44087b79016fe2db97 *vp90-2-14-resize-10frames-fp-tiles-8-4-2-1.webm.md5
+3920c95ba94f1f048a731d9d9b416043b44aa4bd *vp90-2-14-resize-10frames-fp-tiles-8-4.webm
+4eb347a0456d2c49a1e1d8de5aa1c51acc39887e *vp90-2-14-resize-10frames-fp-tiles-8-4.webm.md5
+4b95a74c032a473b6683d7ad5754db1b0ec378e9 *vp90-2-21-resize_inter_1280x720_5_1-2.webm
+a7826dd386bedfe69d02736969bfb47fb6a40a5e *vp90-2-21-resize_inter_1280x720_5_1-2.webm.md5
+5cfff79e82c4d69964ccb8e75b4f0c53b9295167 *vp90-2-21-resize_inter_1280x720_5_3-4.webm
+a18f57db4a25e1f543a99f2ceb182e00db0ee22f *vp90-2-21-resize_inter_1280x720_5_3-4.webm.md5
+d26db0811bf30eb4131d928669713e2485f8e833 *vp90-2-21-resize_inter_1280x720_7_1-2.webm
+fd6f9f332cd5bea4c0f0d57be4297bea493cc5a1 *vp90-2-21-resize_inter_1280x720_7_1-2.webm.md5
+5c7d73d4d268e2ba9593b31cb091fd339505c7fd *vp90-2-21-resize_inter_1280x720_7_3-4.webm
+7bbb949cabc1e70dadcc74582739f63b833034e0 *vp90-2-21-resize_inter_1280x720_7_3-4.webm.md5
+f2d2a41a60eb894aff0c5854afca15931f1445a8 *vp90-2-21-resize_inter_1920x1080_5_1-2.webm
+66d7789992613ac9d678ff905ff1059daa1b89e4 *vp90-2-21-resize_inter_1920x1080_5_1-2.webm.md5
+764edb75fe7dd64e73a1b4f3b4b2b1bf237a4dea *vp90-2-21-resize_inter_1920x1080_5_3-4.webm
+f78bea1075983fd990e7f25d4f31438f9b5efa34 *vp90-2-21-resize_inter_1920x1080_5_3-4.webm.md5
+96496f2ade764a5de9f0c27917c7df1f120fb2ef *vp90-2-21-resize_inter_1920x1080_7_1-2.webm
+2632b635135ed5ecd67fd22dec7990d29c4f4cb5 *vp90-2-21-resize_inter_1920x1080_7_1-2.webm.md5
+74889ea42001bf41428cb742ca74e65129c886dc *vp90-2-21-resize_inter_1920x1080_7_3-4.webm
+d2cf3b25956415bb579d368e7098097e482dd73a *vp90-2-21-resize_inter_1920x1080_7_3-4.webm.md5
+4658986a8ce36ebfcc80a1903e446eaab3985336 *vp90-2-21-resize_inter_320x180_5_1-2.webm
+8a3d8cf325109ffa913cc9426c32eea8c202a09a *vp90-2-21-resize_inter_320x180_5_1-2.webm.md5
+16303aa45176520ee42c2c425247aadc1506b881 *vp90-2-21-resize_inter_320x180_5_3-4.webm
+41cab1ddf7715b680a4dbce42faa9bcd72af4e5c *vp90-2-21-resize_inter_320x180_5_3-4.webm.md5
+56648adcee66dd0e5cb6ac947f5ee1b9cc8ba129 *vp90-2-21-resize_inter_320x180_7_1-2.webm
+70047377787003cc03dda7b2394e6d7eaa666d9e *vp90-2-21-resize_inter_320x180_7_1-2.webm.md5
+d2ff99165488499cc55f75929f1ce5ca9c9e359b *vp90-2-21-resize_inter_320x180_7_3-4.webm
+e69019e378114a4643db283b66d1a7e304761a56 *vp90-2-21-resize_inter_320x180_7_3-4.webm.md5
+4834d129bed0f4289d3a88f2ae3a1736f77621b0 *vp90-2-21-resize_inter_320x240_5_1-2.webm
+a75653c53d22b623c1927fc0088da21dafef21f4 *vp90-2-21-resize_inter_320x240_5_1-2.webm.md5
+19818e1b7fd1c1e63d8873c31b0babe29dd33ba6 *vp90-2-21-resize_inter_320x240_5_3-4.webm
+8d89814ff469a186312111651b16601dfbce4336 *vp90-2-21-resize_inter_320x240_5_3-4.webm.md5
+ac8057bae52498f324ce92a074d5f8207cc4a4a7 *vp90-2-21-resize_inter_320x240_7_1-2.webm
+2643440898c83c08cc47bc744245af696b877c24 *vp90-2-21-resize_inter_320x240_7_1-2.webm.md5
+cf4a4cd38ac8b18c42d8c25a3daafdb39132256b *vp90-2-21-resize_inter_320x240_7_3-4.webm
+70ba8ec9120b26e9b0ffa2c79b432f16cbcb50ec *vp90-2-21-resize_inter_320x240_7_3-4.webm.md5
+669f10409fe1c4a054010162ca47773ea1fdbead *vp90-2-21-resize_inter_640x360_5_1-2.webm
+6355a04249004a35fb386dd1024214234f044383 *vp90-2-21-resize_inter_640x360_5_1-2.webm.md5
+c23763b950b8247c1775d1f8158d93716197676c *vp90-2-21-resize_inter_640x360_5_3-4.webm
+59e6fc381e3ec3b7bdaac586334e0bc944d18fb6 *vp90-2-21-resize_inter_640x360_5_3-4.webm.md5
+71b45cbfdd068baa1f679a69e5e6f421d256a85f *vp90-2-21-resize_inter_640x360_7_1-2.webm
+1416fc761b690c54a955c4cf017fa078520e8c18 *vp90-2-21-resize_inter_640x360_7_1-2.webm.md5
+6c409903279448a697e4db63bab1061784bcd8d2 *vp90-2-21-resize_inter_640x360_7_3-4.webm
+60de1299793433a630b71130cf76c9f5965758e2 *vp90-2-21-resize_inter_640x360_7_3-4.webm.md5
+852b597b8af096d90c80bf0ed6ed3b336b851f19 *vp90-2-21-resize_inter_640x480_5_1-2.webm
+f6856f19236ee46ed462bd0a2e7e72b9c3b9cea6 *vp90-2-21-resize_inter_640x480_5_1-2.webm.md5
+792a16c6f60043bd8dceb515f0b95b8891647858 *vp90-2-21-resize_inter_640x480_5_3-4.webm
+68ffe59877e9a7863805e1c0a3ce18ce037d7c9d *vp90-2-21-resize_inter_640x480_5_3-4.webm.md5
+61e044c4759972a35ea3db8c1478a988910a4ef4 *vp90-2-21-resize_inter_640x480_7_1-2.webm
+7739bfca167b1b43fea72f807f01e097b7cb98d8 *vp90-2-21-resize_inter_640x480_7_1-2.webm.md5
+7291af354b4418917eee00e3a7e366086a0b7a10 *vp90-2-21-resize_inter_640x480_7_3-4.webm
+4a18b09ccb36564193f0215f599d745d95bb558c *vp90-2-21-resize_inter_640x480_7_3-4.webm.md5
diff --git a/libvpx/libvpx/test/test.mk b/libvpx/libvpx/test/test.mk
new file mode 100644
index 0000000..2d50ce8
--- /dev/null
+++ b/libvpx/libvpx/test/test.mk
@@ -0,0 +1,179 @@
+LIBVPX_TEST_SRCS-yes += acm_random.h
+LIBVPX_TEST_SRCS-yes += clear_system_state.h
+LIBVPX_TEST_SRCS-yes += codec_factory.h
+LIBVPX_TEST_SRCS-yes += md5_helper.h
+LIBVPX_TEST_SRCS-yes += register_state_check.h
+LIBVPX_TEST_SRCS-yes += test.mk
+LIBVPX_TEST_SRCS-yes += test_libvpx.cc
+LIBVPX_TEST_SRCS-yes += test_vectors.cc
+LIBVPX_TEST_SRCS-yes += test_vectors.h
+LIBVPX_TEST_SRCS-yes += util.h
+LIBVPX_TEST_SRCS-yes += video_source.h
+
+##
+## BLACK BOX TESTS
+##
+## Black box tests only use the public API.
+##
+LIBVPX_TEST_SRCS-yes                   += ../md5_utils.h ../md5_utils.c
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ivf_video_source.h
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += ../y4minput.h ../y4minput.c
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += altref_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += aq_segment_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += datarate_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += encode_api_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += error_resilience_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += i420_video_source.h
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += realtime_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += resize_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += y4m_video_source.h
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += yuv_video_source.h
+
+LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += config_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += cq_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += keyframe_test.cc
+
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += byte_alignment_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += external_frame_buffer_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += invalid_file_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += user_priv_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += vp9_frame_parallel_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += active_map_refresh_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += active_map_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += borders_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += cpu_speed_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += frame_size_tests.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_lossless_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_end_to_end_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_ethread_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += level_test.cc
+
+LIBVPX_TEST_SRCS-yes                   += decode_test_driver.cc
+LIBVPX_TEST_SRCS-yes                   += decode_test_driver.h
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += encode_test_driver.cc
+LIBVPX_TEST_SRCS-yes                   += encode_test_driver.h
+
+## IVF writing.
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += ../ivfenc.c ../ivfenc.h
+
+## Y4m parsing.
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += y4m_test.cc ../y4menc.c ../y4menc.h
+
+## WebM Parsing
+ifeq ($(CONFIG_WEBM_IO), yes)
+LIBWEBM_PARSER_SRCS += ../third_party/libwebm/mkvparser/mkvparser.cc
+LIBWEBM_PARSER_SRCS += ../third_party/libwebm/mkvparser/mkvreader.cc
+LIBWEBM_PARSER_SRCS += ../third_party/libwebm/mkvparser/mkvparser.h
+LIBWEBM_PARSER_SRCS += ../third_party/libwebm/mkvparser/mkvreader.h
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += $(LIBWEBM_PARSER_SRCS)
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ../tools_common.h
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ../webmdec.cc
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += ../webmdec.h
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += webm_video_source.h
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += vp9_skip_loopfilter_test.cc
+endif
+
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += decode_api_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_DECODERS)    += test_vector_test.cc
+
+# Currently we only support decoder perf tests for vp9. Also they read from WebM
+# files, so WebM IO is required.
+ifeq ($(CONFIG_DECODE_PERF_TESTS)$(CONFIG_VP9_DECODER)$(CONFIG_WEBM_IO), \
+      yesyesyes)
+LIBVPX_TEST_SRCS-yes                   += decode_perf_test.cc
+endif
+
+# encode perf tests are vp9 only
+ifeq ($(CONFIG_ENCODE_PERF_TESTS)$(CONFIG_VP9_ENCODER), yesyes)
+LIBVPX_TEST_SRCS-yes += encode_perf_test.cc
+endif
+
+##
+## WHITE BOX TESTS
+##
+## Whitebox tests invoke functions not exposed via the public API. Certain
+## shared library builds don't make these functions accessible.
+##
+ifeq ($(CONFIG_SHARED),)
+
+## VP8
+ifeq ($(CONFIG_VP8),yes)
+
+# These tests require both the encoder and decoder to be built.
+ifeq ($(CONFIG_VP8_ENCODER)$(CONFIG_VP8_DECODER),yesyes)
+LIBVPX_TEST_SRCS-yes                   += vp8_boolcoder_test.cc
+LIBVPX_TEST_SRCS-yes                   += vp8_fragments_test.cc
+endif
+
+LIBVPX_TEST_SRCS-$(CONFIG_POSTPROC)    += add_noise_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_POSTPROC)    += pp_filter_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP8_DECODER) += vp8_decrypt_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += quantize_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += set_roi.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += variance_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += vp8_fdct4x4_test.cc
+
+LIBVPX_TEST_SRCS-yes                   += idct_test.cc
+LIBVPX_TEST_SRCS-yes                   += sixtap_predict_test.cc
+LIBVPX_TEST_SRCS-yes                   += vpx_scale_test.cc
+
+ifeq ($(CONFIG_VP8_ENCODER)$(CONFIG_TEMPORAL_DENOISING),yesyes)
+LIBVPX_TEST_SRCS-$(HAVE_SSE2) += vp8_denoiser_sse2_test.cc
+endif
+
+endif # VP8
+
+## VP9
+ifeq ($(CONFIG_VP9),yes)
+
+# These tests require both the encoder and decoder to be built.
+ifeq ($(CONFIG_VP9_ENCODER)$(CONFIG_VP9_DECODER),yesyes)
+# IDCT test currently depends on FDCT function
+LIBVPX_TEST_SRCS-yes                   += idct8x8_test.cc
+LIBVPX_TEST_SRCS-yes                   += partial_idct_test.cc
+LIBVPX_TEST_SRCS-yes                   += superframe_test.cc
+LIBVPX_TEST_SRCS-yes                   += tile_independence_test.cc
+LIBVPX_TEST_SRCS-yes                   += vp9_boolcoder_test.cc
+LIBVPX_TEST_SRCS-yes                   += vp9_encoder_parms_get_to_decoder.cc
+endif
+
+LIBVPX_TEST_SRCS-yes                   += convolve_test.cc
+LIBVPX_TEST_SRCS-yes                   += lpf_8_test.cc
+LIBVPX_TEST_SRCS-yes                   += vp9_intrapred_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += vp9_decrypt_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += vp9_thread_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += avg_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += dct16x16_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += dct32x32_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += fdct4x4_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += fdct8x8_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += hadamard_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += minmax_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += variance_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_error_block_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_quantize_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_subtract_test.cc
+
+ifeq ($(CONFIG_VP9_ENCODER),yes)
+LIBVPX_TEST_SRCS-$(CONFIG_SPATIAL_SVC) += svc_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_INTERNAL_STATS) += blockiness_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_INTERNAL_STATS) += consistency_test.cc
+endif
+
+ifeq ($(CONFIG_VP9_ENCODER)$(CONFIG_VP9_TEMPORAL_DENOISING),yesyes)
+LIBVPX_TEST_SRCS-$(HAVE_SSE2) += vp9_denoiser_sse2_test.cc
+endif
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_arf_freq_test.cc
+
+endif # VP9
+
+## Multi-codec / unconditional whitebox tests.
+
+LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS) += sad_test.cc
+
+TEST_INTRA_PRED_SPEED_SRCS-yes := test_intra_pred_speed.cc
+TEST_INTRA_PRED_SPEED_SRCS-yes += ../md5_utils.h ../md5_utils.c
+
+endif # CONFIG_SHARED
+
+include $(SRC_PATH_BARE)/test/test-data.mk
diff --git a/libvpx/libvpx/test/test_intra_pred_speed.cc b/libvpx/libvpx/test/test_intra_pred_speed.cc
new file mode 100644
index 0000000..2acf744
--- /dev/null
+++ b/libvpx/libvpx/test/test_intra_pred_speed.cc
@@ -0,0 +1,373 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+//  Test and time VPX intra-predictor functions
+
+#include <stdio.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/md5_helper.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/vpx_timer.h"
+
+// -----------------------------------------------------------------------------
+
+namespace {
+
+typedef void (*VpxPredFunc)(uint8_t *dst, ptrdiff_t y_stride,
+                            const uint8_t *above, const uint8_t *left);
+
+const int kNumVp9IntraPredFuncs = 13;
+const char *kVp9IntraPredNames[kNumVp9IntraPredFuncs] = {
+  "DC_PRED", "DC_LEFT_PRED", "DC_TOP_PRED", "DC_128_PRED", "V_PRED", "H_PRED",
+  "D45_PRED", "D135_PRED", "D117_PRED", "D153_PRED", "D207_PRED", "D63_PRED",
+  "TM_PRED"
+};
+
+void TestIntraPred(const char name[], VpxPredFunc const *pred_funcs,
+                   const char *const pred_func_names[], int num_funcs,
+                   const char *const signatures[], int block_size,
+                   int num_pixels_per_test) {
+  libvpx_test::ACMRandom rnd(libvpx_test::ACMRandom::DeterministicSeed());
+  const int kBPS = 32;
+  const int kTotalPixels = 32 * kBPS;
+  DECLARE_ALIGNED(16, uint8_t, src[kTotalPixels]);
+  DECLARE_ALIGNED(16, uint8_t, ref_src[kTotalPixels]);
+  DECLARE_ALIGNED(16, uint8_t, left[kBPS]);
+  DECLARE_ALIGNED(16, uint8_t, above_mem[2 * kBPS + 16]);
+  uint8_t *const above = above_mem + 16;
+  for (int i = 0; i < kTotalPixels; ++i) ref_src[i] = rnd.Rand8();
+  for (int i = 0; i < kBPS; ++i) left[i] = rnd.Rand8();
+  for (int i = -1; i < kBPS; ++i) above[i] = rnd.Rand8();
+  const int kNumTests = static_cast<int>(2.e10 / num_pixels_per_test);
+
+  // some code assumes the top row has been extended:
+  // d45/d63 C-code, for instance, but not the assembly.
+  // TODO(jzern): this style of extension isn't strictly necessary.
+  ASSERT_LE(block_size, kBPS);
+  memset(above + block_size, above[block_size - 1], 2 * kBPS - block_size);
+
+  for (int k = 0; k < num_funcs; ++k) {
+    if (pred_funcs[k] == NULL) continue;
+    memcpy(src, ref_src, sizeof(src));
+    vpx_usec_timer timer;
+    vpx_usec_timer_start(&timer);
+    for (int num_tests = 0; num_tests < kNumTests; ++num_tests) {
+      pred_funcs[k](src, kBPS, above, left);
+    }
+    libvpx_test::ClearSystemState();
+    vpx_usec_timer_mark(&timer);
+    const int elapsed_time =
+        static_cast<int>(vpx_usec_timer_elapsed(&timer) / 1000);
+    libvpx_test::MD5 md5;
+    md5.Add(src, sizeof(src));
+    printf("Mode %s[%12s]: %5d ms     MD5: %s\n", name, pred_func_names[k],
+           elapsed_time, md5.Get());
+    EXPECT_STREQ(signatures[k], md5.Get());
+  }
+}
+
+void TestIntraPred4(VpxPredFunc const *pred_funcs) {
+  static const int kNumVp9IntraFuncs = 13;
+  static const char *const kSignatures[kNumVp9IntraFuncs] = {
+    "4334156168b34ab599d9b5b30f522fe9",
+    "bc4649d5ba47c7ff178d92e475960fb0",
+    "8d316e5933326dcac24e1064794b5d12",
+    "a27270fed024eafd762c95de85f4da51",
+    "c33dff000d4256c2b8f3bf9e9bab14d2",
+    "44d8cddc2ad8f79b8ed3306051722b4f",
+    "eb54839b2bad6699d8946f01ec041cd0",
+    "ecb0d56ae5f677ea45127ce9d5c058e4",
+    "0b7936841f6813da818275944895b574",
+    "9117972ef64f91a58ff73e1731c81db2",
+    "c56d5e8c729e46825f46dd5d3b5d508a",
+    "c0889e2039bcf7bcb5d2f33cdca69adc",
+    "309a618577b27c648f9c5ee45252bc8f",
+  };
+  TestIntraPred("Intra4", pred_funcs, kVp9IntraPredNames, kNumVp9IntraFuncs,
+                kSignatures, 4, 4 * 4 * kNumVp9IntraFuncs);
+}
+
+void TestIntraPred8(VpxPredFunc const *pred_funcs) {
+  static const int kNumVp9IntraFuncs = 13;
+  static const char *const kSignatures[kNumVp9IntraFuncs] = {
+    "7694ddeeefed887faf9d339d18850928",
+    "7d726b1213591b99f736be6dec65065b",
+    "19c5711281357a485591aaf9c96c0a67",
+    "ba6b66877a089e71cd938e3b8c40caac",
+    "802440c93317e0f8ba93fab02ef74265",
+    "9e09a47a15deb0b9d8372824f9805080",
+    "b7c2d8c662268c0c427da412d7b0311d",
+    "78339c1c60bb1d67d248ab8c4da08b7f",
+    "5c97d70f7d47de1882a6cd86c165c8a9",
+    "8182bf60688b42205acd95e59e967157",
+    "08323400005a297f16d7e57e7fe1eaac",
+    "95f7bfc262329a5849eda66d8f7c68ce",
+    "815b75c8e0d91cc1ae766dc5d3e445a3",
+  };
+  TestIntraPred("Intra8", pred_funcs, kVp9IntraPredNames, kNumVp9IntraFuncs,
+                kSignatures, 8, 8 * 8 * kNumVp9IntraFuncs);
+}
+
+void TestIntraPred16(VpxPredFunc const *pred_funcs) {
+  static const int kNumVp9IntraFuncs = 13;
+  static const char *const kSignatures[kNumVp9IntraFuncs] = {
+    "b40dbb555d5d16a043dc361e6694fe53",
+    "fb08118cee3b6405d64c1fd68be878c6",
+    "6c190f341475c837cc38c2e566b64875",
+    "db5c34ccbe2c7f595d9b08b0dc2c698c",
+    "a62cbfd153a1f0b9fed13e62b8408a7a",
+    "143df5b4c89335e281103f610f5052e4",
+    "d87feb124107cdf2cfb147655aa0bb3c",
+    "7841fae7d4d47b519322e6a03eeed9dc",
+    "f6ebed3f71cbcf8d6d0516ce87e11093",
+    "3cc480297dbfeed01a1c2d78dd03d0c5",
+    "b9f69fa6532b372c545397dcb78ef311",
+    "a8fe1c70432f09d0c20c67bdb6432c4d",
+    "b8a41aa968ec108af447af4217cba91b",
+  };
+  TestIntraPred("Intra16", pred_funcs, kVp9IntraPredNames, kNumVp9IntraFuncs,
+                kSignatures, 16, 16 * 16 * kNumVp9IntraFuncs);
+}
+
+void TestIntraPred32(VpxPredFunc const *pred_funcs) {
+  static const int kNumVp9IntraFuncs = 13;
+  static const char *const kSignatures[kNumVp9IntraFuncs] = {
+    "558541656d84f9ae7896db655826febe",
+    "b3587a1f9a01495fa38c8cd3c8e2a1bf",
+    "4c6501e64f25aacc55a2a16c7e8f0255",
+    "b3b01379ba08916ef6b1b35f7d9ad51c",
+    "0f1eb38b6cbddb3d496199ef9f329071",
+    "911c06efb9ed1c3b4c104b232b55812f",
+    "9225beb0ddfa7a1d24eaa1be430a6654",
+    "0a6d584a44f8db9aa7ade2e2fdb9fc9e",
+    "b01c9076525216925f3456f034fb6eee",
+    "d267e20ad9e5cd2915d1a47254d3d149",
+    "ed012a4a5da71f36c2393023184a0e59",
+    "f162b51ed618d28b936974cff4391da5",
+    "9e1370c6d42e08d357d9612c93a71cfc",
+  };
+  TestIntraPred("Intra32", pred_funcs, kVp9IntraPredNames, kNumVp9IntraFuncs,
+                kSignatures, 32, 32 * 32 * kNumVp9IntraFuncs);
+}
+
+}  // namespace
+
+// Defines a test case for |arch| (e.g., C, SSE2, ...) passing the predictors
+// to |test_func|. The test name is 'arch.test_func', e.g., C.TestIntraPred4.
+#define INTRA_PRED_TEST(arch, test_func, dc, dc_left, dc_top, dc_128, v, h, \
+                        d45, d135, d117, d153, d207, d63, tm)               \
+  TEST(arch, test_func) {                                                   \
+    static const VpxPredFunc vpx_intra_pred[] = {                           \
+        dc,   dc_left, dc_top, dc_128, v,   h, d45,                         \
+        d135, d117,    d153,   d207,   d63, tm};                            \
+    test_func(vpx_intra_pred);                                              \
+  }
+
+// -----------------------------------------------------------------------------
+// 4x4
+
+INTRA_PRED_TEST(C, TestIntraPred4, vpx_dc_predictor_4x4_c,
+                vpx_dc_left_predictor_4x4_c, vpx_dc_top_predictor_4x4_c,
+                vpx_dc_128_predictor_4x4_c, vpx_v_predictor_4x4_c,
+                vpx_h_predictor_4x4_c, vpx_d45_predictor_4x4_c,
+                vpx_d135_predictor_4x4_c, vpx_d117_predictor_4x4_c,
+                vpx_d153_predictor_4x4_c, vpx_d207_predictor_4x4_c,
+                vpx_d63_predictor_4x4_c, vpx_tm_predictor_4x4_c)
+
+#if HAVE_SSE2 && CONFIG_USE_X86INC
+INTRA_PRED_TEST(SSE2, TestIntraPred4, vpx_dc_predictor_4x4_sse2,
+                vpx_dc_left_predictor_4x4_sse2, vpx_dc_top_predictor_4x4_sse2,
+                vpx_dc_128_predictor_4x4_sse2, vpx_v_predictor_4x4_sse2,
+                vpx_h_predictor_4x4_sse2, vpx_d45_predictor_4x4_sse2, NULL,
+                NULL, NULL, vpx_d207_predictor_4x4_sse2, NULL,
+                vpx_tm_predictor_4x4_sse2)
+#endif  // HAVE_SSE2 && CONFIG_USE_X86INC
+
+#if HAVE_SSSE3 && CONFIG_USE_X86INC
+INTRA_PRED_TEST(SSSE3, TestIntraPred4, NULL, NULL, NULL, NULL, NULL,
+                NULL, NULL, NULL, NULL,
+                vpx_d153_predictor_4x4_ssse3, NULL,
+                vpx_d63_predictor_4x4_ssse3, NULL)
+#endif  // HAVE_SSSE3 && CONFIG_USE_X86INC
+
+#if HAVE_DSPR2
+INTRA_PRED_TEST(DSPR2, TestIntraPred4, vpx_dc_predictor_4x4_dspr2, NULL, NULL,
+                NULL, NULL, vpx_h_predictor_4x4_dspr2, NULL, NULL, NULL, NULL,
+                NULL, NULL, vpx_tm_predictor_4x4_dspr2)
+#endif  // HAVE_DSPR2
+
+#if HAVE_NEON
+INTRA_PRED_TEST(NEON, TestIntraPred4, vpx_dc_predictor_4x4_neon,
+                vpx_dc_left_predictor_4x4_neon, vpx_dc_top_predictor_4x4_neon,
+                vpx_dc_128_predictor_4x4_neon, vpx_v_predictor_4x4_neon,
+                vpx_h_predictor_4x4_neon, vpx_d45_predictor_4x4_neon,
+                vpx_d135_predictor_4x4_neon, NULL, NULL, NULL, NULL,
+                vpx_tm_predictor_4x4_neon)
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+INTRA_PRED_TEST(MSA, TestIntraPred4, vpx_dc_predictor_4x4_msa,
+                vpx_dc_left_predictor_4x4_msa, vpx_dc_top_predictor_4x4_msa,
+                vpx_dc_128_predictor_4x4_msa, vpx_v_predictor_4x4_msa,
+                vpx_h_predictor_4x4_msa, NULL, NULL, NULL, NULL, NULL,
+                NULL, vpx_tm_predictor_4x4_msa)
+#endif  // HAVE_MSA
+
+// -----------------------------------------------------------------------------
+// 8x8
+
+INTRA_PRED_TEST(C, TestIntraPred8, vpx_dc_predictor_8x8_c,
+                vpx_dc_left_predictor_8x8_c, vpx_dc_top_predictor_8x8_c,
+                vpx_dc_128_predictor_8x8_c, vpx_v_predictor_8x8_c,
+                vpx_h_predictor_8x8_c, vpx_d45_predictor_8x8_c,
+                vpx_d135_predictor_8x8_c, vpx_d117_predictor_8x8_c,
+                vpx_d153_predictor_8x8_c, vpx_d207_predictor_8x8_c,
+                vpx_d63_predictor_8x8_c, vpx_tm_predictor_8x8_c)
+
+#if HAVE_SSE2 && CONFIG_USE_X86INC
+INTRA_PRED_TEST(SSE2, TestIntraPred8, vpx_dc_predictor_8x8_sse2,
+                vpx_dc_left_predictor_8x8_sse2, vpx_dc_top_predictor_8x8_sse2,
+                vpx_dc_128_predictor_8x8_sse2, vpx_v_predictor_8x8_sse2,
+                vpx_h_predictor_8x8_sse2, vpx_d45_predictor_8x8_sse2, NULL,
+                NULL, NULL, NULL, NULL, vpx_tm_predictor_8x8_sse2)
+#endif  // HAVE_SSE2 && CONFIG_USE_X86INC
+
+#if HAVE_SSSE3 && CONFIG_USE_X86INC
+INTRA_PRED_TEST(SSSE3, TestIntraPred8, NULL, NULL, NULL, NULL, NULL,
+                NULL, NULL, NULL, NULL,
+                vpx_d153_predictor_8x8_ssse3, vpx_d207_predictor_8x8_ssse3,
+                vpx_d63_predictor_8x8_ssse3, NULL)
+#endif  // HAVE_SSSE3 && CONFIG_USE_X86INC
+
+#if HAVE_DSPR2
+INTRA_PRED_TEST(DSPR2, TestIntraPred8, vpx_dc_predictor_8x8_dspr2, NULL, NULL,
+                NULL, NULL, vpx_h_predictor_8x8_dspr2, NULL, NULL, NULL, NULL,
+                NULL, NULL, vpx_tm_predictor_8x8_c)
+#endif  // HAVE_DSPR2
+
+#if HAVE_NEON
+INTRA_PRED_TEST(NEON, TestIntraPred8, vpx_dc_predictor_8x8_neon,
+                vpx_dc_left_predictor_8x8_neon, vpx_dc_top_predictor_8x8_neon,
+                vpx_dc_128_predictor_8x8_neon, vpx_v_predictor_8x8_neon,
+                vpx_h_predictor_8x8_neon, vpx_d45_predictor_8x8_neon, NULL,
+                NULL, NULL, NULL, NULL, vpx_tm_predictor_8x8_neon)
+
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+INTRA_PRED_TEST(MSA, TestIntraPred8, vpx_dc_predictor_8x8_msa,
+                vpx_dc_left_predictor_8x8_msa, vpx_dc_top_predictor_8x8_msa,
+                vpx_dc_128_predictor_8x8_msa, vpx_v_predictor_8x8_msa,
+                vpx_h_predictor_8x8_msa, NULL, NULL, NULL, NULL, NULL,
+                NULL, vpx_tm_predictor_8x8_msa)
+#endif  // HAVE_MSA
+
+// -----------------------------------------------------------------------------
+// 16x16
+
+INTRA_PRED_TEST(C, TestIntraPred16, vpx_dc_predictor_16x16_c,
+                vpx_dc_left_predictor_16x16_c, vpx_dc_top_predictor_16x16_c,
+                vpx_dc_128_predictor_16x16_c, vpx_v_predictor_16x16_c,
+                vpx_h_predictor_16x16_c, vpx_d45_predictor_16x16_c,
+                vpx_d135_predictor_16x16_c, vpx_d117_predictor_16x16_c,
+                vpx_d153_predictor_16x16_c, vpx_d207_predictor_16x16_c,
+                vpx_d63_predictor_16x16_c, vpx_tm_predictor_16x16_c)
+
+#if HAVE_SSE2 && CONFIG_USE_X86INC
+INTRA_PRED_TEST(SSE2, TestIntraPred16, vpx_dc_predictor_16x16_sse2,
+                vpx_dc_left_predictor_16x16_sse2,
+                vpx_dc_top_predictor_16x16_sse2,
+                vpx_dc_128_predictor_16x16_sse2, vpx_v_predictor_16x16_sse2,
+                vpx_h_predictor_16x16_sse2, NULL, NULL, NULL, NULL, NULL, NULL,
+                vpx_tm_predictor_16x16_sse2)
+#endif  // HAVE_SSE2 && CONFIG_USE_X86INC
+
+#if HAVE_SSSE3 && CONFIG_USE_X86INC
+INTRA_PRED_TEST(SSSE3, TestIntraPred16, NULL, NULL, NULL, NULL, NULL,
+                NULL, vpx_d45_predictor_16x16_ssse3,
+                NULL, NULL, vpx_d153_predictor_16x16_ssse3,
+                vpx_d207_predictor_16x16_ssse3, vpx_d63_predictor_16x16_ssse3,
+                NULL)
+#endif  // HAVE_SSSE3 && CONFIG_USE_X86INC
+
+#if HAVE_DSPR2
+INTRA_PRED_TEST(DSPR2, TestIntraPred16, vpx_dc_predictor_16x16_dspr2, NULL,
+                NULL, NULL, NULL, vpx_h_predictor_16x16_dspr2, NULL, NULL, NULL,
+                NULL, NULL, NULL, NULL)
+#endif  // HAVE_DSPR2
+
+#if HAVE_NEON
+INTRA_PRED_TEST(NEON, TestIntraPred16, vpx_dc_predictor_16x16_neon,
+                vpx_dc_left_predictor_16x16_neon,
+                vpx_dc_top_predictor_16x16_neon,
+                vpx_dc_128_predictor_16x16_neon, vpx_v_predictor_16x16_neon,
+                vpx_h_predictor_16x16_neon, vpx_d45_predictor_16x16_neon, NULL,
+                NULL, NULL, NULL, NULL, vpx_tm_predictor_16x16_neon)
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+INTRA_PRED_TEST(MSA, TestIntraPred16, vpx_dc_predictor_16x16_msa,
+                vpx_dc_left_predictor_16x16_msa, vpx_dc_top_predictor_16x16_msa,
+                vpx_dc_128_predictor_16x16_msa, vpx_v_predictor_16x16_msa,
+                vpx_h_predictor_16x16_msa, NULL, NULL, NULL, NULL, NULL,
+                NULL, vpx_tm_predictor_16x16_msa)
+#endif  // HAVE_MSA
+
+// -----------------------------------------------------------------------------
+// 32x32
+
+INTRA_PRED_TEST(C, TestIntraPred32, vpx_dc_predictor_32x32_c,
+                vpx_dc_left_predictor_32x32_c, vpx_dc_top_predictor_32x32_c,
+                vpx_dc_128_predictor_32x32_c, vpx_v_predictor_32x32_c,
+                vpx_h_predictor_32x32_c, vpx_d45_predictor_32x32_c,
+                vpx_d135_predictor_32x32_c, vpx_d117_predictor_32x32_c,
+                vpx_d153_predictor_32x32_c, vpx_d207_predictor_32x32_c,
+                vpx_d63_predictor_32x32_c, vpx_tm_predictor_32x32_c)
+
+#if HAVE_SSE2 && CONFIG_USE_X86INC
+INTRA_PRED_TEST(SSE2, TestIntraPred32, vpx_dc_predictor_32x32_sse2,
+                vpx_dc_left_predictor_32x32_sse2,
+                vpx_dc_top_predictor_32x32_sse2,
+                vpx_dc_128_predictor_32x32_sse2, vpx_v_predictor_32x32_sse2,
+                vpx_h_predictor_32x32_sse2, NULL, NULL, NULL, NULL, NULL,
+                NULL, vpx_tm_predictor_32x32_sse2)
+#endif  // HAVE_SSE2 && CONFIG_USE_X86INC
+
+#if HAVE_SSSE3 && CONFIG_USE_X86INC
+INTRA_PRED_TEST(SSSE3, TestIntraPred32, NULL, NULL, NULL, NULL, NULL,
+                NULL, vpx_d45_predictor_32x32_ssse3, NULL, NULL,
+                vpx_d153_predictor_32x32_ssse3, vpx_d207_predictor_32x32_ssse3,
+                vpx_d63_predictor_32x32_ssse3, NULL)
+#endif  // HAVE_SSSE3 && CONFIG_USE_X86INC
+
+#if HAVE_NEON
+INTRA_PRED_TEST(NEON, TestIntraPred32, vpx_dc_predictor_32x32_neon,
+                vpx_dc_left_predictor_32x32_neon,
+                vpx_dc_top_predictor_32x32_neon,
+                vpx_dc_128_predictor_32x32_neon, vpx_v_predictor_32x32_neon,
+                vpx_h_predictor_32x32_neon, NULL, NULL, NULL, NULL, NULL, NULL,
+                vpx_tm_predictor_32x32_neon)
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+INTRA_PRED_TEST(MSA, TestIntraPred32, vpx_dc_predictor_32x32_msa,
+                vpx_dc_left_predictor_32x32_msa, vpx_dc_top_predictor_32x32_msa,
+                vpx_dc_128_predictor_32x32_msa, vpx_v_predictor_32x32_msa,
+                vpx_h_predictor_32x32_msa, NULL, NULL, NULL, NULL, NULL,
+                NULL, vpx_tm_predictor_32x32_msa)
+#endif  // HAVE_MSA
+
+#include "test/test_libvpx.cc"
diff --git a/libvpx/libvpx/test/test_libvpx.cc b/libvpx/libvpx/test/test_libvpx.cc
new file mode 100644
index 0000000..005ea8d
--- /dev/null
+++ b/libvpx/libvpx/test/test_libvpx.cc
@@ -0,0 +1,77 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <string>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#if ARCH_X86 || ARCH_X86_64
+#include "vpx_ports/x86.h"
+#endif
+extern "C" {
+#if CONFIG_VP8
+extern void vp8_rtcd();
+#endif  // CONFIG_VP8
+#if CONFIG_VP9
+extern void vp9_rtcd();
+#endif  // CONFIG_VP9
+extern void vpx_dsp_rtcd();
+extern void vpx_scale_rtcd();
+}
+
+#if ARCH_X86 || ARCH_X86_64
+static void append_negative_gtest_filter(const char *str) {
+  std::string filter = ::testing::FLAGS_gtest_filter;
+  // Negative patterns begin with one '-' followed by a ':' separated list.
+  if (filter.find('-') == std::string::npos) filter += '-';
+  filter += str;
+  ::testing::FLAGS_gtest_filter = filter;
+}
+#endif  // ARCH_X86 || ARCH_X86_64
+
+int main(int argc, char **argv) {
+  ::testing::InitGoogleTest(&argc, argv);
+
+#if ARCH_X86 || ARCH_X86_64
+  const int simd_caps = x86_simd_caps();
+  if (!(simd_caps & HAS_MMX))
+    append_negative_gtest_filter(":MMX.*:MMX/*");
+  if (!(simd_caps & HAS_SSE))
+    append_negative_gtest_filter(":SSE.*:SSE/*");
+  if (!(simd_caps & HAS_SSE2))
+    append_negative_gtest_filter(":SSE2.*:SSE2/*");
+  if (!(simd_caps & HAS_SSE3))
+    append_negative_gtest_filter(":SSE3.*:SSE3/*");
+  if (!(simd_caps & HAS_SSSE3))
+    append_negative_gtest_filter(":SSSE3.*:SSSE3/*");
+  if (!(simd_caps & HAS_SSE4_1))
+    append_negative_gtest_filter(":SSE4_1.*:SSE4_1/*");
+  if (!(simd_caps & HAS_AVX))
+    append_negative_gtest_filter(":AVX.*:AVX/*");
+  if (!(simd_caps & HAS_AVX2))
+    append_negative_gtest_filter(":AVX2.*:AVX2/*");
+#endif  // ARCH_X86 || ARCH_X86_64
+
+#if !CONFIG_SHARED
+// Shared library builds don't support whitebox tests
+// that exercise internal symbols.
+
+#if CONFIG_VP8
+  vp8_rtcd();
+#endif  // CONFIG_VP8
+#if CONFIG_VP9
+  vp9_rtcd();
+#endif  // CONFIG_VP9
+  vpx_dsp_rtcd();
+  vpx_scale_rtcd();
+#endif  // !CONFIG_SHARED
+
+  return RUN_ALL_TESTS();
+}
diff --git a/libvpx/libvpx/test/test_vector_test.cc b/libvpx/libvpx/test/test_vector_test.cc
new file mode 100644
index 0000000..f1aa4d7
--- /dev/null
+++ b/libvpx/libvpx/test/test_vector_test.cc
@@ -0,0 +1,192 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <set>
+#include <string>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "../tools_common.h"
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/md5_helper.h"
+#include "test/test_vectors.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+
+enum DecodeMode {
+  kSerialMode,
+  kFrameParallelMode
+};
+
+const int kDecodeMode = 0;
+const int kThreads = 1;
+const int kFileName = 2;
+
+typedef std::tr1::tuple<int, int, const char*> DecodeParam;
+
+class TestVectorTest : public ::libvpx_test::DecoderTest,
+    public ::libvpx_test::CodecTestWithParam<DecodeParam> {
+ protected:
+  TestVectorTest()
+      : DecoderTest(GET_PARAM(0)),
+        md5_file_(NULL) {
+#if CONFIG_VP9_DECODER
+    resize_clips_.insert(
+      ::libvpx_test::kVP9TestVectorsResize,
+      ::libvpx_test::kVP9TestVectorsResize +
+          ::libvpx_test::kNumVP9TestVectorsResize);
+#endif
+  }
+
+  virtual ~TestVectorTest() {
+    if (md5_file_)
+      fclose(md5_file_);
+  }
+
+  void OpenMD5File(const std::string& md5_file_name_) {
+    md5_file_ = libvpx_test::OpenTestDataFile(md5_file_name_);
+    ASSERT_TRUE(md5_file_ != NULL) << "Md5 file open failed. Filename: "
+        << md5_file_name_;
+  }
+
+  virtual void DecompressedFrameHook(const vpx_image_t& img,
+                                     const unsigned int frame_number) {
+    ASSERT_TRUE(md5_file_ != NULL);
+    char expected_md5[33];
+    char junk[128];
+
+    // Read correct md5 checksums.
+    const int res = fscanf(md5_file_, "%s  %s", expected_md5, junk);
+    ASSERT_NE(res, EOF) << "Read md5 data failed";
+    expected_md5[32] = '\0';
+
+    ::libvpx_test::MD5 md5_res;
+    md5_res.Add(&img);
+    const char *actual_md5 = md5_res.Get();
+
+    // Check md5 match.
+    ASSERT_STREQ(expected_md5, actual_md5)
+        << "Md5 checksums don't match: frame number = " << frame_number;
+  }
+
+#if CONFIG_VP9_DECODER
+  std::set<std::string> resize_clips_;
+#endif
+
+ private:
+  FILE *md5_file_;
+};
+
+// This test runs through the whole set of test vectors, and decodes them.
+// The md5 checksums are computed for each frame in the video file. If md5
+// checksums match the correct md5 data, then the test is passed. Otherwise,
+// the test failed.
+TEST_P(TestVectorTest, MD5Match) {
+  const DecodeParam input = GET_PARAM(1);
+  const std::string filename = std::tr1::get<kFileName>(input);
+  const int threads = std::tr1::get<kThreads>(input);
+  const int mode = std::tr1::get<kDecodeMode>(input);
+  libvpx_test::CompressedVideoSource *video = NULL;
+  vpx_codec_flags_t flags = 0;
+  vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+  char str[256];
+
+  if (mode == kFrameParallelMode) {
+    flags |= VPX_CODEC_USE_FRAME_THREADING;
+#if CONFIG_VP9_DECODER
+    // TODO(hkuang): Fix frame parallel decode bug. See issue 1086.
+    if (resize_clips_.find(filename) != resize_clips_.end()) {
+      printf("Skipping the test file: %s, due to frame parallel decode bug.\n",
+             filename.c_str());
+      return;
+    }
+#endif
+  }
+
+  cfg.threads = threads;
+
+  snprintf(str, sizeof(str) / sizeof(str[0]) - 1,
+           "file: %s  mode: %s threads: %d",
+           filename.c_str(), mode == 0 ? "Serial" : "Parallel", threads);
+  SCOPED_TRACE(str);
+
+  // Open compressed video file.
+  if (filename.substr(filename.length() - 3, 3) == "ivf") {
+    video = new libvpx_test::IVFVideoSource(filename);
+  } else if (filename.substr(filename.length() - 4, 4) == "webm") {
+#if CONFIG_WEBM_IO
+    video = new libvpx_test::WebMVideoSource(filename);
+#else
+    fprintf(stderr, "WebM IO is disabled, skipping test vector %s\n",
+            filename.c_str());
+    return;
+#endif
+  }
+  video->Init();
+
+  // Construct md5 file name.
+  const std::string md5_filename = filename + ".md5";
+  OpenMD5File(md5_filename);
+
+  // Set decode config and flags.
+  set_cfg(cfg);
+  set_flags(flags);
+
+  // Decode frame, and check the md5 matching.
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video, cfg));
+  delete video;
+}
+
+// Test VP8 decode in serial mode with single thread.
+// NOTE: VP8 only support serial mode.
+#if CONFIG_VP8_DECODER
+VP8_INSTANTIATE_TEST_CASE(
+    TestVectorTest,
+    ::testing::Combine(
+        ::testing::Values(0),  // Serial Mode.
+        ::testing::Values(1),  // Single thread.
+        ::testing::ValuesIn(libvpx_test::kVP8TestVectors,
+                            libvpx_test::kVP8TestVectors +
+                                libvpx_test::kNumVP8TestVectors)));
+#endif  // CONFIG_VP8_DECODER
+
+// Test VP9 decode in serial mode with single thread.
+#if CONFIG_VP9_DECODER
+VP9_INSTANTIATE_TEST_CASE(
+    TestVectorTest,
+    ::testing::Combine(
+        ::testing::Values(0),  // Serial Mode.
+        ::testing::Values(1),  // Single thread.
+        ::testing::ValuesIn(libvpx_test::kVP9TestVectors,
+                            libvpx_test::kVP9TestVectors +
+                                libvpx_test::kNumVP9TestVectors)));
+
+// Test VP9 decode in frame parallel mode with different number of threads.
+INSTANTIATE_TEST_CASE_P(
+    VP9MultiThreadedFrameParallel, TestVectorTest,
+    ::testing::Combine(
+        ::testing::Values(
+            static_cast<const libvpx_test::CodecFactory *>(&libvpx_test::kVP9)),
+        ::testing::Combine(
+            ::testing::Values(1),        // Frame Parallel mode.
+            ::testing::Range(2, 9),      // With 2 ~ 8 threads.
+            ::testing::ValuesIn(libvpx_test::kVP9TestVectors,
+                                libvpx_test::kVP9TestVectors +
+                                    libvpx_test::kNumVP9TestVectors))));
+#endif
+}  // namespace
diff --git a/libvpx/libvpx/test/test_vectors.cc b/libvpx/libvpx/test/test_vectors.cc
new file mode 100644
index 0000000..c822479
--- /dev/null
+++ b/libvpx/libvpx/test/test_vectors.cc
@@ -0,0 +1,251 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "test/test_vectors.h"
+
+namespace libvpx_test {
+
+#define NELEMENTS(x) static_cast<int>(sizeof(x) / sizeof(x[0]))
+
+#if CONFIG_VP8_DECODER
+const char *const kVP8TestVectors[] = {
+  "vp80-00-comprehensive-001.ivf",
+  "vp80-00-comprehensive-002.ivf", "vp80-00-comprehensive-003.ivf",
+  "vp80-00-comprehensive-004.ivf", "vp80-00-comprehensive-005.ivf",
+  "vp80-00-comprehensive-006.ivf", "vp80-00-comprehensive-007.ivf",
+  "vp80-00-comprehensive-008.ivf", "vp80-00-comprehensive-009.ivf",
+  "vp80-00-comprehensive-010.ivf", "vp80-00-comprehensive-011.ivf",
+  "vp80-00-comprehensive-012.ivf", "vp80-00-comprehensive-013.ivf",
+  "vp80-00-comprehensive-014.ivf", "vp80-00-comprehensive-015.ivf",
+  "vp80-00-comprehensive-016.ivf", "vp80-00-comprehensive-017.ivf",
+  "vp80-00-comprehensive-018.ivf", "vp80-01-intra-1400.ivf",
+  "vp80-01-intra-1411.ivf", "vp80-01-intra-1416.ivf",
+  "vp80-01-intra-1417.ivf", "vp80-02-inter-1402.ivf",
+  "vp80-02-inter-1412.ivf", "vp80-02-inter-1418.ivf",
+  "vp80-02-inter-1424.ivf", "vp80-03-segmentation-01.ivf",
+  "vp80-03-segmentation-02.ivf", "vp80-03-segmentation-03.ivf",
+  "vp80-03-segmentation-04.ivf", "vp80-03-segmentation-1401.ivf",
+  "vp80-03-segmentation-1403.ivf", "vp80-03-segmentation-1407.ivf",
+  "vp80-03-segmentation-1408.ivf", "vp80-03-segmentation-1409.ivf",
+  "vp80-03-segmentation-1410.ivf", "vp80-03-segmentation-1413.ivf",
+  "vp80-03-segmentation-1414.ivf", "vp80-03-segmentation-1415.ivf",
+  "vp80-03-segmentation-1425.ivf", "vp80-03-segmentation-1426.ivf",
+  "vp80-03-segmentation-1427.ivf", "vp80-03-segmentation-1432.ivf",
+  "vp80-03-segmentation-1435.ivf", "vp80-03-segmentation-1436.ivf",
+  "vp80-03-segmentation-1437.ivf", "vp80-03-segmentation-1441.ivf",
+  "vp80-03-segmentation-1442.ivf", "vp80-04-partitions-1404.ivf",
+  "vp80-04-partitions-1405.ivf", "vp80-04-partitions-1406.ivf",
+  "vp80-05-sharpness-1428.ivf", "vp80-05-sharpness-1429.ivf",
+  "vp80-05-sharpness-1430.ivf", "vp80-05-sharpness-1431.ivf",
+  "vp80-05-sharpness-1433.ivf", "vp80-05-sharpness-1434.ivf",
+  "vp80-05-sharpness-1438.ivf", "vp80-05-sharpness-1439.ivf",
+  "vp80-05-sharpness-1440.ivf", "vp80-05-sharpness-1443.ivf",
+  "vp80-06-smallsize.ivf"
+};
+const int kNumVP8TestVectors = NELEMENTS(kVP8TestVectors);
+#endif  // CONFIG_VP8_DECODER
+#if CONFIG_VP9_DECODER
+#define RESIZE_TEST_VECTORS "vp90-2-21-resize_inter_320x180_5_1-2.webm", \
+  "vp90-2-21-resize_inter_320x180_5_3-4.webm", \
+  "vp90-2-21-resize_inter_320x180_7_1-2.webm", \
+  "vp90-2-21-resize_inter_320x180_7_3-4.webm", \
+  "vp90-2-21-resize_inter_320x240_5_1-2.webm", \
+  "vp90-2-21-resize_inter_320x240_5_3-4.webm", \
+  "vp90-2-21-resize_inter_320x240_7_1-2.webm", \
+  "vp90-2-21-resize_inter_320x240_7_3-4.webm", \
+  "vp90-2-21-resize_inter_640x360_5_1-2.webm", \
+  "vp90-2-21-resize_inter_640x360_5_3-4.webm", \
+  "vp90-2-21-resize_inter_640x360_7_1-2.webm", \
+  "vp90-2-21-resize_inter_640x360_7_3-4.webm", \
+  "vp90-2-21-resize_inter_640x480_5_1-2.webm", \
+  "vp90-2-21-resize_inter_640x480_5_3-4.webm", \
+  "vp90-2-21-resize_inter_640x480_7_1-2.webm", \
+  "vp90-2-21-resize_inter_640x480_7_3-4.webm", \
+  "vp90-2-21-resize_inter_1280x720_5_1-2.webm", \
+  "vp90-2-21-resize_inter_1280x720_5_3-4.webm", \
+  "vp90-2-21-resize_inter_1280x720_7_1-2.webm", \
+  "vp90-2-21-resize_inter_1280x720_7_3-4.webm", \
+  "vp90-2-21-resize_inter_1920x1080_5_1-2.webm", \
+  "vp90-2-21-resize_inter_1920x1080_5_3-4.webm", \
+  "vp90-2-21-resize_inter_1920x1080_7_1-2.webm", \
+  "vp90-2-21-resize_inter_1920x1080_7_3-4.webm",
+
+const char *const kVP9TestVectors[] = {
+  "vp90-2-00-quantizer-00.webm", "vp90-2-00-quantizer-01.webm",
+  "vp90-2-00-quantizer-02.webm", "vp90-2-00-quantizer-03.webm",
+  "vp90-2-00-quantizer-04.webm", "vp90-2-00-quantizer-05.webm",
+  "vp90-2-00-quantizer-06.webm", "vp90-2-00-quantizer-07.webm",
+  "vp90-2-00-quantizer-08.webm", "vp90-2-00-quantizer-09.webm",
+  "vp90-2-00-quantizer-10.webm", "vp90-2-00-quantizer-11.webm",
+  "vp90-2-00-quantizer-12.webm", "vp90-2-00-quantizer-13.webm",
+  "vp90-2-00-quantizer-14.webm", "vp90-2-00-quantizer-15.webm",
+  "vp90-2-00-quantizer-16.webm", "vp90-2-00-quantizer-17.webm",
+  "vp90-2-00-quantizer-18.webm", "vp90-2-00-quantizer-19.webm",
+  "vp90-2-00-quantizer-20.webm", "vp90-2-00-quantizer-21.webm",
+  "vp90-2-00-quantizer-22.webm", "vp90-2-00-quantizer-23.webm",
+  "vp90-2-00-quantizer-24.webm", "vp90-2-00-quantizer-25.webm",
+  "vp90-2-00-quantizer-26.webm", "vp90-2-00-quantizer-27.webm",
+  "vp90-2-00-quantizer-28.webm", "vp90-2-00-quantizer-29.webm",
+  "vp90-2-00-quantizer-30.webm", "vp90-2-00-quantizer-31.webm",
+  "vp90-2-00-quantizer-32.webm", "vp90-2-00-quantizer-33.webm",
+  "vp90-2-00-quantizer-34.webm", "vp90-2-00-quantizer-35.webm",
+  "vp90-2-00-quantizer-36.webm", "vp90-2-00-quantizer-37.webm",
+  "vp90-2-00-quantizer-38.webm", "vp90-2-00-quantizer-39.webm",
+  "vp90-2-00-quantizer-40.webm", "vp90-2-00-quantizer-41.webm",
+  "vp90-2-00-quantizer-42.webm", "vp90-2-00-quantizer-43.webm",
+  "vp90-2-00-quantizer-44.webm", "vp90-2-00-quantizer-45.webm",
+  "vp90-2-00-quantizer-46.webm", "vp90-2-00-quantizer-47.webm",
+  "vp90-2-00-quantizer-48.webm", "vp90-2-00-quantizer-49.webm",
+  "vp90-2-00-quantizer-50.webm", "vp90-2-00-quantizer-51.webm",
+  "vp90-2-00-quantizer-52.webm", "vp90-2-00-quantizer-53.webm",
+  "vp90-2-00-quantizer-54.webm", "vp90-2-00-quantizer-55.webm",
+  "vp90-2-00-quantizer-56.webm", "vp90-2-00-quantizer-57.webm",
+  "vp90-2-00-quantizer-58.webm", "vp90-2-00-quantizer-59.webm",
+  "vp90-2-00-quantizer-60.webm", "vp90-2-00-quantizer-61.webm",
+  "vp90-2-00-quantizer-62.webm", "vp90-2-00-quantizer-63.webm",
+  "vp90-2-01-sharpness-1.webm", "vp90-2-01-sharpness-2.webm",
+  "vp90-2-01-sharpness-3.webm", "vp90-2-01-sharpness-4.webm",
+  "vp90-2-01-sharpness-5.webm", "vp90-2-01-sharpness-6.webm",
+  "vp90-2-01-sharpness-7.webm", "vp90-2-02-size-08x08.webm",
+  "vp90-2-02-size-08x10.webm", "vp90-2-02-size-08x16.webm",
+  "vp90-2-02-size-08x18.webm", "vp90-2-02-size-08x32.webm",
+  "vp90-2-02-size-08x34.webm", "vp90-2-02-size-08x64.webm",
+  "vp90-2-02-size-08x66.webm", "vp90-2-02-size-10x08.webm",
+  "vp90-2-02-size-10x10.webm", "vp90-2-02-size-10x16.webm",
+  "vp90-2-02-size-10x18.webm", "vp90-2-02-size-10x32.webm",
+  "vp90-2-02-size-10x34.webm", "vp90-2-02-size-10x64.webm",
+  "vp90-2-02-size-10x66.webm", "vp90-2-02-size-16x08.webm",
+  "vp90-2-02-size-16x10.webm", "vp90-2-02-size-16x16.webm",
+  "vp90-2-02-size-16x18.webm", "vp90-2-02-size-16x32.webm",
+  "vp90-2-02-size-16x34.webm", "vp90-2-02-size-16x64.webm",
+  "vp90-2-02-size-16x66.webm", "vp90-2-02-size-18x08.webm",
+  "vp90-2-02-size-18x10.webm", "vp90-2-02-size-18x16.webm",
+  "vp90-2-02-size-18x18.webm", "vp90-2-02-size-18x32.webm",
+  "vp90-2-02-size-18x34.webm", "vp90-2-02-size-18x64.webm",
+  "vp90-2-02-size-18x66.webm", "vp90-2-02-size-32x08.webm",
+  "vp90-2-02-size-32x10.webm", "vp90-2-02-size-32x16.webm",
+  "vp90-2-02-size-32x18.webm", "vp90-2-02-size-32x32.webm",
+  "vp90-2-02-size-32x34.webm", "vp90-2-02-size-32x64.webm",
+  "vp90-2-02-size-32x66.webm", "vp90-2-02-size-34x08.webm",
+  "vp90-2-02-size-34x10.webm", "vp90-2-02-size-34x16.webm",
+  "vp90-2-02-size-34x18.webm", "vp90-2-02-size-34x32.webm",
+  "vp90-2-02-size-34x34.webm", "vp90-2-02-size-34x64.webm",
+  "vp90-2-02-size-34x66.webm", "vp90-2-02-size-64x08.webm",
+  "vp90-2-02-size-64x10.webm", "vp90-2-02-size-64x16.webm",
+  "vp90-2-02-size-64x18.webm", "vp90-2-02-size-64x32.webm",
+  "vp90-2-02-size-64x34.webm", "vp90-2-02-size-64x64.webm",
+  "vp90-2-02-size-64x66.webm", "vp90-2-02-size-66x08.webm",
+  "vp90-2-02-size-66x10.webm", "vp90-2-02-size-66x16.webm",
+  "vp90-2-02-size-66x18.webm", "vp90-2-02-size-66x32.webm",
+  "vp90-2-02-size-66x34.webm", "vp90-2-02-size-66x64.webm",
+  "vp90-2-02-size-66x66.webm", "vp90-2-02-size-130x132.webm",
+  "vp90-2-02-size-132x130.webm", "vp90-2-02-size-132x132.webm",
+  "vp90-2-02-size-178x180.webm", "vp90-2-02-size-180x178.webm",
+  "vp90-2-02-size-180x180.webm", "vp90-2-03-size-196x196.webm",
+  "vp90-2-03-size-196x198.webm", "vp90-2-03-size-196x200.webm",
+  "vp90-2-03-size-196x202.webm", "vp90-2-03-size-196x208.webm",
+  "vp90-2-03-size-196x210.webm", "vp90-2-03-size-196x224.webm",
+  "vp90-2-03-size-196x226.webm", "vp90-2-03-size-198x196.webm",
+  "vp90-2-03-size-198x198.webm", "vp90-2-03-size-198x200.webm",
+  "vp90-2-03-size-198x202.webm", "vp90-2-03-size-198x208.webm",
+  "vp90-2-03-size-198x210.webm", "vp90-2-03-size-198x224.webm",
+  "vp90-2-03-size-198x226.webm", "vp90-2-03-size-200x196.webm",
+  "vp90-2-03-size-200x198.webm", "vp90-2-03-size-200x200.webm",
+  "vp90-2-03-size-200x202.webm", "vp90-2-03-size-200x208.webm",
+  "vp90-2-03-size-200x210.webm", "vp90-2-03-size-200x224.webm",
+  "vp90-2-03-size-200x226.webm", "vp90-2-03-size-202x196.webm",
+  "vp90-2-03-size-202x198.webm", "vp90-2-03-size-202x200.webm",
+  "vp90-2-03-size-202x202.webm", "vp90-2-03-size-202x208.webm",
+  "vp90-2-03-size-202x210.webm", "vp90-2-03-size-202x224.webm",
+  "vp90-2-03-size-202x226.webm", "vp90-2-03-size-208x196.webm",
+  "vp90-2-03-size-208x198.webm", "vp90-2-03-size-208x200.webm",
+  "vp90-2-03-size-208x202.webm", "vp90-2-03-size-208x208.webm",
+  "vp90-2-03-size-208x210.webm", "vp90-2-03-size-208x224.webm",
+  "vp90-2-03-size-208x226.webm", "vp90-2-03-size-210x196.webm",
+  "vp90-2-03-size-210x198.webm", "vp90-2-03-size-210x200.webm",
+  "vp90-2-03-size-210x202.webm", "vp90-2-03-size-210x208.webm",
+  "vp90-2-03-size-210x210.webm", "vp90-2-03-size-210x224.webm",
+  "vp90-2-03-size-210x226.webm", "vp90-2-03-size-224x196.webm",
+  "vp90-2-03-size-224x198.webm", "vp90-2-03-size-224x200.webm",
+  "vp90-2-03-size-224x202.webm", "vp90-2-03-size-224x208.webm",
+  "vp90-2-03-size-224x210.webm", "vp90-2-03-size-224x224.webm",
+  "vp90-2-03-size-224x226.webm", "vp90-2-03-size-226x196.webm",
+  "vp90-2-03-size-226x198.webm", "vp90-2-03-size-226x200.webm",
+  "vp90-2-03-size-226x202.webm", "vp90-2-03-size-226x208.webm",
+  "vp90-2-03-size-226x210.webm", "vp90-2-03-size-226x224.webm",
+  "vp90-2-03-size-226x226.webm", "vp90-2-03-size-352x288.webm",
+  "vp90-2-03-deltaq.webm",
+  "vp90-2-05-resize.ivf", "vp90-2-06-bilinear.webm",
+  "vp90-2-07-frame_parallel.webm", "vp90-2-08-tile_1x2_frame_parallel.webm",
+  "vp90-2-08-tile_1x2.webm", "vp90-2-08-tile_1x4_frame_parallel.webm",
+  "vp90-2-08-tile_1x4.webm", "vp90-2-08-tile_1x8_frame_parallel.webm",
+  "vp90-2-08-tile_1x8.webm", "vp90-2-08-tile-4x4.webm",
+  "vp90-2-08-tile-4x1.webm", "vp90-2-09-subpixel-00.ivf",
+  "vp90-2-02-size-lf-1920x1080.webm", "vp90-2-09-aq2.webm",
+  "vp90-2-09-lf_deltas.webm", "vp90-2-10-show-existing-frame.webm",
+  "vp90-2-10-show-existing-frame2.webm",
+  "vp90-2-11-size-351x287.webm", "vp90-2-11-size-351x288.webm",
+  "vp90-2-11-size-352x287.webm", "vp90-2-12-droppable_1.ivf",
+  "vp90-2-12-droppable_2.ivf", "vp90-2-12-droppable_3.ivf",
+#if !CONFIG_SIZE_LIMIT || \
+    (DECODE_WIDTH_LIMIT >= 20400 && DECODE_HEIGHT_LIMIT >= 120)
+  "vp90-2-13-largescaling.webm",
+#endif
+  "vp90-2-14-resize-fp-tiles-1-16.webm",
+  "vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm",
+  "vp90-2-14-resize-fp-tiles-1-2.webm", "vp90-2-14-resize-fp-tiles-1-4.webm",
+  "vp90-2-14-resize-fp-tiles-16-1.webm", "vp90-2-14-resize-fp-tiles-16-2.webm",
+  "vp90-2-14-resize-fp-tiles-16-4.webm",
+  "vp90-2-14-resize-fp-tiles-16-8-4-2-1.webm",
+  "vp90-2-14-resize-fp-tiles-16-8.webm", "vp90-2-14-resize-fp-tiles-1-8.webm",
+  "vp90-2-14-resize-fp-tiles-2-16.webm", "vp90-2-14-resize-fp-tiles-2-1.webm",
+  "vp90-2-14-resize-fp-tiles-2-4.webm", "vp90-2-14-resize-fp-tiles-2-8.webm",
+  "vp90-2-14-resize-fp-tiles-4-16.webm", "vp90-2-14-resize-fp-tiles-4-1.webm",
+  "vp90-2-14-resize-fp-tiles-4-2.webm", "vp90-2-14-resize-fp-tiles-4-8.webm",
+  "vp90-2-14-resize-fp-tiles-8-16.webm", "vp90-2-14-resize-fp-tiles-8-1.webm",
+  "vp90-2-14-resize-fp-tiles-8-2.webm", "vp90-2-14-resize-fp-tiles-8-4.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-1-2-4-8.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-1-2.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-1-4.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-1-8.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-2-1.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-2-4.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-2-8.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-4-1.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-4-2.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-4-8.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-8-1.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-8-2.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-8-4-2-1.webm",
+  "vp90-2-14-resize-10frames-fp-tiles-8-4.webm",
+  "vp90-2-15-segkey.webm", "vp90-2-15-segkey_adpq.webm",
+  "vp90-2-16-intra-only.webm", "vp90-2-17-show-existing-frame.webm",
+  "vp90-2-18-resize.ivf", "vp90-2-19-skip.webm",
+  "vp90-2-19-skip-01.webm", "vp90-2-19-skip-02.webm",
+  "vp91-2-04-yuv444.webm",
+  "vp91-2-04-yuv422.webm", "vp91-2-04-yuv440.webm",
+#if CONFIG_VP9_HIGHBITDEPTH
+  "vp92-2-20-10bit-yuv420.webm", "vp92-2-20-12bit-yuv420.webm",
+  "vp93-2-20-10bit-yuv422.webm", "vp93-2-20-12bit-yuv422.webm",
+  "vp93-2-20-10bit-yuv440.webm", "vp93-2-20-12bit-yuv440.webm",
+  "vp93-2-20-10bit-yuv444.webm", "vp93-2-20-12bit-yuv444.webm",
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  "vp90-2-20-big_superframe-01.webm", "vp90-2-20-big_superframe-02.webm",
+  RESIZE_TEST_VECTORS
+};
+const int kNumVP9TestVectors = NELEMENTS(kVP9TestVectors);
+const char *const kVP9TestVectorsResize[] = {
+  RESIZE_TEST_VECTORS
+};
+const int kNumVP9TestVectorsResize = NELEMENTS(kVP9TestVectorsResize);
+#undef RESIZE_TEST_VECTORS
+#endif  // CONFIG_VP9_DECODER
+
+}  // namespace libvpx_test
diff --git a/libvpx/libvpx/test/test_vectors.h b/libvpx/libvpx/test/test_vectors.h
new file mode 100644
index 0000000..2c6918a
--- /dev/null
+++ b/libvpx/libvpx/test/test_vectors.h
@@ -0,0 +1,32 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef TEST_TEST_VECTORS_H_
+#define TEST_TEST_VECTORS_H_
+
+#include "./vpx_config.h"
+
+namespace libvpx_test {
+
+#if CONFIG_VP8_DECODER
+extern const int kNumVP8TestVectors;
+extern const char *const kVP8TestVectors[];
+#endif
+
+#if CONFIG_VP9_DECODER
+extern const int kNumVP9TestVectors;
+extern const char *const kVP9TestVectors[];
+extern const int kNumVP9TestVectorsResize;
+extern const char *const kVP9TestVectorsResize[];
+#endif  // CONFIG_VP9_DECODER
+
+}  // namespace libvpx_test
+
+#endif  // TEST_TEST_VECTORS_H_
diff --git a/libvpx/libvpx/test/tile_independence_test.cc b/libvpx/libvpx/test/tile_independence_test.cc
new file mode 100644
index 0000000..f15d94a
--- /dev/null
+++ b/libvpx/libvpx/test/tile_independence_test.cc
@@ -0,0 +1,106 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/md5_helper.h"
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+class TileIndependenceTest : public ::libvpx_test::EncoderTest,
+                             public ::libvpx_test::CodecTestWithParam<int> {
+ protected:
+  TileIndependenceTest()
+      : EncoderTest(GET_PARAM(0)),
+        md5_fw_order_(),
+        md5_inv_order_(),
+        n_tiles_(GET_PARAM(1)) {
+    init_flags_ = VPX_CODEC_USE_PSNR;
+    vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+    cfg.w = 704;
+    cfg.h = 144;
+    cfg.threads = 1;
+    fw_dec_ = codec_->CreateDecoder(cfg, 0);
+    inv_dec_ = codec_->CreateDecoder(cfg, 0);
+    inv_dec_->Control(VP9_INVERT_TILE_DECODE_ORDER, 1);
+  }
+
+  virtual ~TileIndependenceTest() {
+    delete fw_dec_;
+    delete inv_dec_;
+  }
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(libvpx_test::kTwoPassGood);
+  }
+
+  virtual void PreEncodeFrameHook(libvpx_test::VideoSource *video,
+                                  libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP9E_SET_TILE_COLUMNS, n_tiles_);
+    }
+  }
+
+  void UpdateMD5(::libvpx_test::Decoder *dec, const vpx_codec_cx_pkt_t *pkt,
+                 ::libvpx_test::MD5 *md5) {
+    const vpx_codec_err_t res = dec->DecodeFrame(
+        reinterpret_cast<uint8_t*>(pkt->data.frame.buf), pkt->data.frame.sz);
+    if (res != VPX_CODEC_OK) {
+      abort_ = true;
+      ASSERT_EQ(VPX_CODEC_OK, res);
+    }
+    const vpx_image_t *img = dec->GetDxData().Next();
+    md5->Add(img);
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    UpdateMD5(fw_dec_, pkt, &md5_fw_order_);
+    UpdateMD5(inv_dec_, pkt, &md5_inv_order_);
+  }
+
+  ::libvpx_test::MD5 md5_fw_order_, md5_inv_order_;
+  ::libvpx_test::Decoder *fw_dec_, *inv_dec_;
+
+ private:
+  int n_tiles_;
+};
+
+// run an encode with 2 or 4 tiles, and do the decode both in normal and
+// inverted tile ordering. Ensure that the MD5 of the output in both cases
+// is identical. If so, tiles are considered independent and the test passes.
+TEST_P(TileIndependenceTest, MD5Match) {
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 500;
+  cfg_.g_lag_in_frames = 25;
+  cfg_.rc_end_usage = VPX_VBR;
+
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 704, 144,
+                                     timebase.den, timebase.num, 0, 30);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  const char *md5_fw_str = md5_fw_order_.Get();
+  const char *md5_inv_str = md5_inv_order_.Get();
+
+  // could use ASSERT_EQ(!memcmp(.., .., 16) here, but this gives nicer
+  // output if it fails. Not sure if it's helpful since it's really just
+  // a MD5...
+  ASSERT_STREQ(md5_fw_str, md5_inv_str);
+}
+
+VP9_INSTANTIATE_TEST_CASE(TileIndependenceTest, ::testing::Range(0, 2, 1));
+}  // namespace
diff --git a/libvpx/libvpx/test/tools_common.sh b/libvpx/libvpx/test/tools_common.sh
new file mode 100755
index 0000000..0bdcc08
--- /dev/null
+++ b/libvpx/libvpx/test/tools_common.sh
@@ -0,0 +1,438 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file contains shell code shared by test scripts for libvpx tools.
+
+# Use $VPX_TEST_TOOLS_COMMON_SH as a pseudo include guard.
+if [ -z "${VPX_TEST_TOOLS_COMMON_SH}" ]; then
+VPX_TEST_TOOLS_COMMON_SH=included
+
+set -e
+devnull='> /dev/null 2>&1'
+VPX_TEST_PREFIX=""
+
+elog() {
+  echo "$@" 1>&2
+}
+
+vlog() {
+  if [ "${VPX_TEST_VERBOSE_OUTPUT}" = "yes" ]; then
+    echo "$@"
+  fi
+}
+
+# Sets $VPX_TOOL_TEST to the name specified by positional parameter one.
+test_begin() {
+  VPX_TOOL_TEST="${1}"
+}
+
+# Clears the VPX_TOOL_TEST variable after confirming that $VPX_TOOL_TEST matches
+# positional parameter one.
+test_end() {
+  if [ "$1" != "${VPX_TOOL_TEST}" ]; then
+    echo "FAIL completed test mismatch!."
+    echo "  completed test: ${1}"
+    echo "  active test: ${VPX_TOOL_TEST}."
+    return 1
+  fi
+  VPX_TOOL_TEST='<unset>'
+}
+
+# Echoes the target configuration being tested.
+test_configuration_target() {
+  vpx_config_mk="${LIBVPX_CONFIG_PATH}/config.mk"
+  # Find the TOOLCHAIN line, split it using ':=' as the field separator, and
+  # print the last field to get the value. Then pipe the value to tr to consume
+  # any leading/trailing spaces while allowing tr to echo the output to stdout.
+  awk -F ':=' '/TOOLCHAIN/ { print $NF }' "${vpx_config_mk}" | tr -d ' '
+}
+
+# Trap function used for failure reports and tool output directory removal.
+# When the contents of $VPX_TOOL_TEST do not match the string '<unset>', reports
+# failure of test stored in $VPX_TOOL_TEST.
+cleanup() {
+  if [ -n "${VPX_TOOL_TEST}" ] && [ "${VPX_TOOL_TEST}" != '<unset>' ]; then
+    echo "FAIL: $VPX_TOOL_TEST"
+  fi
+  if [ -n "${VPX_TEST_OUTPUT_DIR}" ] && [ -d "${VPX_TEST_OUTPUT_DIR}" ]; then
+    rm -rf "${VPX_TEST_OUTPUT_DIR}"
+  fi
+}
+
+# Echoes the git hash portion of the VERSION_STRING variable defined in
+# $LIBVPX_CONFIG_PATH/config.mk to stdout, or the version number string when
+# no git hash is contained in VERSION_STRING.
+config_hash() {
+  vpx_config_mk="${LIBVPX_CONFIG_PATH}/config.mk"
+  # Find VERSION_STRING line, split it with "-g" and print the last field to
+  # output the git hash to stdout.
+  vpx_version=$(awk -F -g '/VERSION_STRING/ {print $NF}' "${vpx_config_mk}")
+  # Handle two situations here:
+  # 1. The default case: $vpx_version is a git hash, so echo it unchanged.
+  # 2. When being run a non-dev tree, the -g portion is not present in the
+  #    version string: It's only the version number.
+  #    In this case $vpx_version is something like 'VERSION_STRING=v1.3.0', so
+  #    we echo only what is after the '='.
+  echo "${vpx_version##*=}"
+}
+
+# Echoes the short form of the current git hash.
+current_hash() {
+  if git --version > /dev/null 2>&1; then
+    (cd "$(dirname "${0}")"
+    git rev-parse --short HEAD)
+  else
+    # Return the config hash if git is unavailable: Fail silently, git hashes
+    # are used only for warnings.
+    config_hash
+  fi
+}
+
+# Echoes warnings to stdout when git hash in vpx_config.h does not match the
+# current git hash.
+check_git_hashes() {
+  hash_at_configure_time=$(config_hash)
+  hash_now=$(current_hash)
+
+  if [ "${hash_at_configure_time}" != "${hash_now}" ]; then
+    echo "Warning: git hash has changed since last configure."
+  fi
+}
+
+# $1 is the name of an environment variable containing a directory name to
+# test.
+test_env_var_dir() {
+  local dir=$(eval echo "\${$1}")
+  if [ ! -d "${dir}" ]; then
+    elog "'${dir}': No such directory"
+    elog "The $1 environment variable must be set to a valid directory."
+    return 1
+  fi
+}
+
+# This script requires that the LIBVPX_BIN_PATH, LIBVPX_CONFIG_PATH, and
+# LIBVPX_TEST_DATA_PATH variables are in the environment: Confirm that
+# the variables are set and that they all evaluate to directory paths.
+verify_vpx_test_environment() {
+  test_env_var_dir "LIBVPX_BIN_PATH" \
+    && test_env_var_dir "LIBVPX_CONFIG_PATH" \
+    && test_env_var_dir "LIBVPX_TEST_DATA_PATH"
+}
+
+# Greps vpx_config.h in LIBVPX_CONFIG_PATH for positional parameter one, which
+# should be a LIBVPX preprocessor flag. Echoes yes to stdout when the feature
+# is available.
+vpx_config_option_enabled() {
+  vpx_config_option="${1}"
+  vpx_config_file="${LIBVPX_CONFIG_PATH}/vpx_config.h"
+  config_line=$(grep "${vpx_config_option}" "${vpx_config_file}")
+  if echo "${config_line}" | egrep -q '1$'; then
+    echo yes
+  fi
+}
+
+# Echoes yes when output of test_configuration_target() contains win32 or win64.
+is_windows_target() {
+  if test_configuration_target \
+     | grep -q -e win32 -e win64 > /dev/null 2>&1; then
+    echo yes
+  fi
+}
+
+# Echoes path to $1 when it's executable and exists in ${LIBVPX_BIN_PATH}, or an
+# empty string. Caller is responsible for testing the string once the function
+# returns.
+vpx_tool_path() {
+  local readonly tool_name="$1"
+  local tool_path="${LIBVPX_BIN_PATH}/${tool_name}${VPX_TEST_EXE_SUFFIX}"
+  if [ ! -x "${tool_path}" ]; then
+    # Try one directory up: when running via examples.sh the tool could be in
+    # the parent directory of $LIBVPX_BIN_PATH.
+    tool_path="${LIBVPX_BIN_PATH}/../${tool_name}${VPX_TEST_EXE_SUFFIX}"
+  fi
+
+  if [ ! -x "${tool_path}" ]; then
+    tool_path=""
+  fi
+  echo "${tool_path}"
+}
+
+# Echoes yes to stdout when the file named by positional parameter one exists
+# in LIBVPX_BIN_PATH, and is executable.
+vpx_tool_available() {
+  local tool_name="$1"
+  local tool="${LIBVPX_BIN_PATH}/${tool_name}${VPX_TEST_EXE_SUFFIX}"
+  [ -x "${tool}" ] && echo yes
+}
+
+# Echoes yes to stdout when vpx_config_option_enabled() reports yes for
+# CONFIG_VP8_DECODER.
+vp8_decode_available() {
+  [ "$(vpx_config_option_enabled CONFIG_VP8_DECODER)" = "yes" ] && echo yes
+}
+
+# Echoes yes to stdout when vpx_config_option_enabled() reports yes for
+# CONFIG_VP8_ENCODER.
+vp8_encode_available() {
+  [ "$(vpx_config_option_enabled CONFIG_VP8_ENCODER)" = "yes" ] && echo yes
+}
+
+# Echoes yes to stdout when vpx_config_option_enabled() reports yes for
+# CONFIG_VP9_DECODER.
+vp9_decode_available() {
+  [ "$(vpx_config_option_enabled CONFIG_VP9_DECODER)" = "yes" ] && echo yes
+}
+
+# Echoes yes to stdout when vpx_config_option_enabled() reports yes for
+# CONFIG_VP9_ENCODER.
+vp9_encode_available() {
+  [ "$(vpx_config_option_enabled CONFIG_VP9_ENCODER)" = "yes" ] && echo yes
+}
+
+# Echoes yes to stdout when vpx_config_option_enabled() reports yes for
+# CONFIG_WEBM_IO.
+webm_io_available() {
+  [ "$(vpx_config_option_enabled CONFIG_WEBM_IO)" = "yes" ] && echo yes
+}
+
+# Filters strings from $1 using the filter specified by $2. Filter behavior
+# depends on the presence of $3. When $3 is present, strings that match the
+# filter are excluded. When $3 is omitted, strings matching the filter are
+# included.
+# The filtered result is echoed to stdout.
+filter_strings() {
+  strings=${1}
+  filter=${2}
+  exclude=${3}
+
+  if [ -n "${exclude}" ]; then
+    # When positional parameter three exists the caller wants to remove strings.
+    # Tell grep to invert matches using the -v argument.
+    exclude='-v'
+  else
+    unset exclude
+  fi
+
+  if [ -n "${filter}" ]; then
+    for s in ${strings}; do
+      if echo "${s}" | egrep -q ${exclude} "${filter}" > /dev/null 2>&1; then
+        filtered_strings="${filtered_strings} ${s}"
+      fi
+    done
+  else
+    filtered_strings="${strings}"
+  fi
+  echo "${filtered_strings}"
+}
+
+# Runs user test functions passed via positional parameters one and two.
+# Functions in positional parameter one are treated as environment verification
+# functions and are run unconditionally. Functions in positional parameter two
+# are run according to the rules specified in vpx_test_usage().
+run_tests() {
+  local env_tests="verify_vpx_test_environment $1"
+  local tests_to_filter="$2"
+  local test_name="${VPX_TEST_NAME}"
+
+  if [ -z "${test_name}" ]; then
+    test_name="$(basename "${0%.*}")"
+  fi
+
+  if [ "${VPX_TEST_RUN_DISABLED_TESTS}" != "yes" ]; then
+    # Filter out DISABLED tests.
+    tests_to_filter=$(filter_strings "${tests_to_filter}" ^DISABLED exclude)
+  fi
+
+  if [ -n "${VPX_TEST_FILTER}" ]; then
+    # Remove tests not matching the user's filter.
+    tests_to_filter=$(filter_strings "${tests_to_filter}" ${VPX_TEST_FILTER})
+  fi
+
+  # User requested test listing: Dump test names and return.
+  if [ "${VPX_TEST_LIST_TESTS}" = "yes" ]; then
+    for test_name in $tests_to_filter; do
+      echo ${test_name}
+    done
+    return
+  fi
+
+  # Don't bother with the environment tests if everything else was disabled.
+  [ -z "${tests_to_filter}" ] && return
+
+  # Combine environment and actual tests.
+  local tests_to_run="${env_tests} ${tests_to_filter}"
+
+  check_git_hashes
+
+  # Run tests.
+  for test in ${tests_to_run}; do
+    test_begin "${test}"
+    vlog "  RUN  ${test}"
+    "${test}"
+    vlog "  PASS ${test}"
+    test_end "${test}"
+  done
+
+  local tested_config="$(test_configuration_target) @ $(current_hash)"
+  echo "${test_name}: Done, all tests pass for ${tested_config}."
+}
+
+vpx_test_usage() {
+cat << EOF
+  Usage: ${0##*/} [arguments]
+    --bin-path <path to libvpx binaries directory>
+    --config-path <path to libvpx config directory>
+    --filter <filter>: User test filter. Only tests matching filter are run.
+    --run-disabled-tests: Run disabled tests.
+    --help: Display this message and exit.
+    --test-data-path <path to libvpx test data directory>
+    --show-program-output: Shows output from all programs being tested.
+    --prefix: Allows for a user specified prefix to be inserted before all test
+              programs. Grants the ability, for example, to run test programs
+              within valgrind.
+    --list-tests: List all test names and exit without actually running tests.
+    --verbose: Verbose output.
+
+    When the --bin-path option is not specified the script attempts to use
+    \$LIBVPX_BIN_PATH and then the current directory.
+
+    When the --config-path option is not specified the script attempts to use
+    \$LIBVPX_CONFIG_PATH and then the current directory.
+
+    When the -test-data-path option is not specified the script attempts to use
+    \$LIBVPX_TEST_DATA_PATH and then the current directory.
+EOF
+}
+
+# Returns non-zero (failure) when required environment variables are empty
+# strings.
+vpx_test_check_environment() {
+  if [ -z "${LIBVPX_BIN_PATH}" ] || \
+     [ -z "${LIBVPX_CONFIG_PATH}" ] || \
+     [ -z "${LIBVPX_TEST_DATA_PATH}" ]; then
+    return 1
+  fi
+}
+
+# Parse the command line.
+while [ -n "$1" ]; do
+  case "$1" in
+    --bin-path)
+      LIBVPX_BIN_PATH="$2"
+      shift
+      ;;
+    --config-path)
+      LIBVPX_CONFIG_PATH="$2"
+      shift
+      ;;
+    --filter)
+      VPX_TEST_FILTER="$2"
+      shift
+      ;;
+    --run-disabled-tests)
+      VPX_TEST_RUN_DISABLED_TESTS=yes
+      ;;
+    --help)
+      vpx_test_usage
+      exit
+      ;;
+    --test-data-path)
+      LIBVPX_TEST_DATA_PATH="$2"
+      shift
+      ;;
+    --prefix)
+      VPX_TEST_PREFIX="$2"
+      shift
+      ;;
+    --verbose)
+      VPX_TEST_VERBOSE_OUTPUT=yes
+      ;;
+    --show-program-output)
+      devnull=
+      ;;
+    --list-tests)
+      VPX_TEST_LIST_TESTS=yes
+      ;;
+    *)
+      vpx_test_usage
+      exit 1
+      ;;
+  esac
+  shift
+done
+
+# Handle running the tests from a build directory without arguments when running
+# the tests on *nix/macosx.
+LIBVPX_BIN_PATH="${LIBVPX_BIN_PATH:-.}"
+LIBVPX_CONFIG_PATH="${LIBVPX_CONFIG_PATH:-.}"
+LIBVPX_TEST_DATA_PATH="${LIBVPX_TEST_DATA_PATH:-.}"
+
+# Create a temporary directory for output files, and a trap to clean it up.
+if [ -n "${TMPDIR}" ]; then
+  VPX_TEST_TEMP_ROOT="${TMPDIR}"
+elif [ -n "${TEMPDIR}" ]; then
+  VPX_TEST_TEMP_ROOT="${TEMPDIR}"
+else
+  VPX_TEST_TEMP_ROOT=/tmp
+fi
+
+VPX_TEST_OUTPUT_DIR="${VPX_TEST_TEMP_ROOT}/vpx_test_$$"
+
+if ! mkdir -p "${VPX_TEST_OUTPUT_DIR}" || \
+   [ ! -d "${VPX_TEST_OUTPUT_DIR}" ]; then
+  echo "${0##*/}: Cannot create output directory, giving up."
+  echo "${0##*/}:   VPX_TEST_OUTPUT_DIR=${VPX_TEST_OUTPUT_DIR}"
+  exit 1
+fi
+
+if [ "$(is_windows_target)" = "yes" ]; then
+  VPX_TEST_EXE_SUFFIX=".exe"
+fi
+
+# Variables shared by tests.
+VP8_IVF_FILE="${LIBVPX_TEST_DATA_PATH}/vp80-00-comprehensive-001.ivf"
+VP9_IVF_FILE="${LIBVPX_TEST_DATA_PATH}/vp90-2-09-subpixel-00.ivf"
+
+VP9_WEBM_FILE="${LIBVPX_TEST_DATA_PATH}/vp90-2-00-quantizer-00.webm"
+VP9_FPM_WEBM_FILE="${LIBVPX_TEST_DATA_PATH}/vp90-2-07-frame_parallel-1.webm"
+VP9_LT_50_FRAMES_WEBM_FILE="${LIBVPX_TEST_DATA_PATH}/vp90-2-02-size-32x08.webm"
+
+YUV_RAW_INPUT="${LIBVPX_TEST_DATA_PATH}/hantro_collage_w352h288.yuv"
+YUV_RAW_INPUT_WIDTH=352
+YUV_RAW_INPUT_HEIGHT=288
+
+Y4M_NOSQ_PAR_INPUT="${LIBVPX_TEST_DATA_PATH}/park_joy_90p_8_420_a10-1.y4m"
+Y4M_720P_INPUT="${LIBVPX_TEST_DATA_PATH}/niklas_1280_720_30.y4m"
+
+# Setup a trap function to clean up after tests complete.
+trap cleanup EXIT
+
+vlog "$(basename "${0%.*}") test configuration:
+  LIBVPX_BIN_PATH=${LIBVPX_BIN_PATH}
+  LIBVPX_CONFIG_PATH=${LIBVPX_CONFIG_PATH}
+  LIBVPX_TEST_DATA_PATH=${LIBVPX_TEST_DATA_PATH}
+  VP8_IVF_FILE=${VP8_IVF_FILE}
+  VP9_IVF_FILE=${VP9_IVF_FILE}
+  VP9_WEBM_FILE=${VP9_WEBM_FILE}
+  VPX_TEST_EXE_SUFFIX=${VPX_TEST_EXE_SUFFIX}
+  VPX_TEST_FILTER=${VPX_TEST_FILTER}
+  VPX_TEST_LIST_TESTS=${VPX_TEST_LIST_TESTS}
+  VPX_TEST_OUTPUT_DIR=${VPX_TEST_OUTPUT_DIR}
+  VPX_TEST_PREFIX=${VPX_TEST_PREFIX}
+  VPX_TEST_RUN_DISABLED_TESTS=${VPX_TEST_RUN_DISABLED_TESTS}
+  VPX_TEST_SHOW_PROGRAM_OUTPUT=${VPX_TEST_SHOW_PROGRAM_OUTPUT}
+  VPX_TEST_TEMP_ROOT=${VPX_TEST_TEMP_ROOT}
+  VPX_TEST_VERBOSE_OUTPUT=${VPX_TEST_VERBOSE_OUTPUT}
+  YUV_RAW_INPUT=${YUV_RAW_INPUT}
+  YUV_RAW_INPUT_WIDTH=${YUV_RAW_INPUT_WIDTH}
+  YUV_RAW_INPUT_HEIGHT=${YUV_RAW_INPUT_HEIGHT}
+  Y4M_NOSQ_PAR_INPUT=${Y4M_NOSQ_PAR_INPUT}"
+
+fi  # End $VPX_TEST_TOOLS_COMMON_SH pseudo include guard.
diff --git a/libvpx/libvpx/test/twopass_encoder.sh b/libvpx/libvpx/test/twopass_encoder.sh
new file mode 100755
index 0000000..7a223f2
--- /dev/null
+++ b/libvpx/libvpx/test/twopass_encoder.sh
@@ -0,0 +1,60 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx twopass_encoder example. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to twopass_encoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+twopass_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs twopass_encoder using the codec specified by $1 with a frame limit of
+# 100.
+twopass_encoder() {
+  local encoder="${LIBVPX_BIN_PATH}/twopass_encoder${VPX_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/twopass_encoder_${codec}.ivf"
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" 100 \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+twopass_encoder_vp8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    twopass_encoder vp8 || return 1
+  fi
+}
+
+twopass_encoder_vp9() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    twopass_encoder vp9 || return 1
+  fi
+}
+
+twopass_encoder_tests="twopass_encoder_vp8
+                       twopass_encoder_vp9"
+
+run_tests twopass_encoder_verify_environment "${twopass_encoder_tests}"
diff --git a/libvpx/libvpx/test/user_priv_test.cc b/libvpx/libvpx/test/user_priv_test.cc
new file mode 100644
index 0000000..8512d88
--- /dev/null
+++ b/libvpx/libvpx/test/user_priv_test.cc
@@ -0,0 +1,100 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "test/acm_random.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+#include "vpx/vp8.h"
+
+namespace {
+
+using std::string;
+using libvpx_test::ACMRandom;
+
+#if CONFIG_WEBM_IO
+
+void CheckUserPrivateData(void *user_priv, int *target) {
+  // actual pointer value should be the same as expected.
+  EXPECT_EQ(reinterpret_cast<void *>(target), user_priv) <<
+      "user_priv pointer value does not match.";
+}
+
+// Decodes |filename|. Passes in user_priv data when calling DecodeFrame and
+// compares the user_priv from return img with the original user_priv to see if
+// they match. Both the pointer values and the values inside the addresses
+// should match.
+string DecodeFile(const string &filename) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  libvpx_test::WebMVideoSource video(filename);
+  video.Init();
+
+  vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+  libvpx_test::VP9Decoder decoder(cfg, 0);
+
+  libvpx_test::MD5 md5;
+  int frame_num = 0;
+  for (video.Begin(); !::testing::Test::HasFailure() && video.cxdata();
+       video.Next()) {
+    void *user_priv = reinterpret_cast<void *>(&frame_num);
+    const vpx_codec_err_t res =
+        decoder.DecodeFrame(video.cxdata(), video.frame_size(),
+                            (frame_num == 0) ? NULL : user_priv);
+    if (res != VPX_CODEC_OK) {
+      EXPECT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+      break;
+    }
+    libvpx_test::DxDataIterator dec_iter = decoder.GetDxData();
+    const vpx_image_t *img = NULL;
+
+    // Get decompressed data.
+    while ((img = dec_iter.Next())) {
+      if (frame_num == 0) {
+        CheckUserPrivateData(img->user_priv, NULL);
+      } else {
+        CheckUserPrivateData(img->user_priv, &frame_num);
+
+        // Also test ctrl_get_reference api.
+        struct vp9_ref_frame ref;
+        // Randomly fetch a reference frame.
+        ref.idx = rnd.Rand8() % 3;
+        decoder.Control(VP9_GET_REFERENCE, &ref);
+
+        CheckUserPrivateData(ref.img.user_priv, NULL);
+      }
+      md5.Add(img);
+    }
+
+    frame_num++;
+  }
+  return string(md5.Get());
+}
+
+TEST(UserPrivTest, VideoDecode) {
+  // no tiles or frame parallel; this exercises the decoding to test the
+  // user_priv.
+  EXPECT_STREQ("b35a1b707b28e82be025d960aba039bc",
+               DecodeFile("vp90-2-03-size-226x226.webm").c_str());
+}
+
+#endif  // CONFIG_WEBM_IO
+
+}  // namespace
diff --git a/libvpx/libvpx/test/util.h b/libvpx/libvpx/test/util.h
new file mode 100644
index 0000000..b27bffa
--- /dev/null
+++ b/libvpx/libvpx/test/util.h
@@ -0,0 +1,46 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef TEST_UTIL_H_
+#define TEST_UTIL_H_
+
+#include <stdio.h>
+#include <math.h>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "vpx/vpx_image.h"
+
+// Macros
+#define GET_PARAM(k) std::tr1::get< k >(GetParam())
+
+inline double compute_psnr(const vpx_image_t *img1, const vpx_image_t *img2) {
+  assert((img1->fmt == img2->fmt) &&
+         (img1->d_w == img2->d_w) &&
+         (img1->d_h == img2->d_h));
+
+  const unsigned int width_y  = img1->d_w;
+  const unsigned int height_y = img1->d_h;
+  unsigned int i, j;
+
+  int64_t sqrerr = 0;
+  for (i = 0; i < height_y; ++i)
+    for (j = 0; j < width_y; ++j) {
+      int64_t d = img1->planes[VPX_PLANE_Y][i * img1->stride[VPX_PLANE_Y] + j] -
+                  img2->planes[VPX_PLANE_Y][i * img2->stride[VPX_PLANE_Y] + j];
+      sqrerr += d * d;
+    }
+  double mse = static_cast<double>(sqrerr) / (width_y * height_y);
+  double psnr = 100.0;
+  if (mse > 0.0) {
+    psnr = 10 * log10(255.0 * 255.0 / mse);
+  }
+  return psnr;
+}
+
+#endif  // TEST_UTIL_H_
diff --git a/libvpx/libvpx/test/variance_test.cc b/libvpx/libvpx/test/variance_test.cc
new file mode 100644
index 0000000..cb63390
--- /dev/null
+++ b/libvpx/libvpx/test/variance_test.cc
@@ -0,0 +1,1337 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdlib>
+#include <new>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+namespace {
+
+typedef unsigned int (*VarianceMxNFunc)(const uint8_t *a, int a_stride,
+                                        const uint8_t *b, int b_stride,
+                                        unsigned int *sse);
+typedef unsigned int (*SubpixVarMxNFunc)(const uint8_t *a, int a_stride,
+                                         int xoffset, int yoffset,
+                                         const uint8_t *b, int b_stride,
+                                         unsigned int *sse);
+typedef unsigned int (*SubpixAvgVarMxNFunc)(const uint8_t *a, int a_stride,
+                                            int xoffset, int yoffset,
+                                            const uint8_t *b, int b_stride,
+                                            uint32_t *sse,
+                                            const uint8_t *second_pred);
+typedef unsigned int (*Get4x4SseFunc)(const uint8_t *a, int a_stride,
+                                      const uint8_t *b, int b_stride);
+typedef unsigned int (*SumOfSquaresFunction)(const int16_t *src);
+
+
+using ::std::tr1::get;
+using ::std::tr1::make_tuple;
+using ::std::tr1::tuple;
+using libvpx_test::ACMRandom;
+
+// Truncate high bit depth results by downshifting (with rounding) by:
+// 2 * (bit_depth - 8) for sse
+// (bit_depth - 8) for se
+static void RoundHighBitDepth(int bit_depth, int64_t *se, uint64_t *sse) {
+  switch (bit_depth) {
+    case VPX_BITS_12:
+      *sse = (*sse + 128) >> 8;
+      *se = (*se + 8) >> 4;
+      break;
+    case VPX_BITS_10:
+      *sse = (*sse + 8) >> 4;
+      *se = (*se + 2) >> 2;
+      break;
+    case VPX_BITS_8:
+    default:
+      break;
+  }
+}
+
+static unsigned int mb_ss_ref(const int16_t *src) {
+  unsigned int res = 0;
+  for (int i = 0; i < 256; ++i) {
+    res += src[i] * src[i];
+  }
+  return res;
+}
+
+/* Note:
+ *  Our codebase calculates the "diff" value in the variance algorithm by
+ *  (src - ref).
+ */
+static uint32_t variance_ref(const uint8_t *src, const uint8_t *ref,
+                             int l2w, int l2h, int src_stride_coeff,
+                             int ref_stride_coeff, uint32_t *sse_ptr,
+                             bool use_high_bit_depth_,
+                             vpx_bit_depth_t bit_depth) {
+  int64_t se = 0;
+  uint64_t sse = 0;
+  const int w = 1 << l2w;
+  const int h = 1 << l2h;
+  for (int y = 0; y < h; y++) {
+    for (int x = 0; x < w; x++) {
+      int diff;
+      if (!use_high_bit_depth_) {
+        diff = src[w * y * src_stride_coeff + x] -
+               ref[w * y * ref_stride_coeff + x];
+        se += diff;
+        sse += diff * diff;
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        diff = CONVERT_TO_SHORTPTR(src)[w * y * src_stride_coeff + x] -
+               CONVERT_TO_SHORTPTR(ref)[w * y * ref_stride_coeff + x];
+        se += diff;
+        sse += diff * diff;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+    }
+  }
+  RoundHighBitDepth(bit_depth, &se, &sse);
+  *sse_ptr = static_cast<uint32_t>(sse);
+  return static_cast<uint32_t>(sse -
+                               ((static_cast<int64_t>(se) * se) >>
+                                (l2w + l2h)));
+}
+
+/* The subpel reference functions differ from the codec version in one aspect:
+ * they calculate the bilinear factors directly instead of using a lookup table
+ * and therefore upshift xoff and yoff by 1. Only every other calculated value
+ * is used so the codec version shrinks the table to save space and maintain
+ * compatibility with vp8.
+ */
+static uint32_t subpel_variance_ref(const uint8_t *ref, const uint8_t *src,
+                                    int l2w, int l2h, int xoff, int yoff,
+                                    uint32_t *sse_ptr,
+                                    bool use_high_bit_depth_,
+                                    vpx_bit_depth_t bit_depth) {
+  int64_t se = 0;
+  uint64_t sse = 0;
+  const int w = 1 << l2w;
+  const int h = 1 << l2h;
+
+  xoff <<= 1;
+  yoff <<= 1;
+
+  for (int y = 0; y < h; y++) {
+    for (int x = 0; x < w; x++) {
+      // Bilinear interpolation at a 16th pel step.
+      if (!use_high_bit_depth_) {
+        const int a1 = ref[(w + 1) * (y + 0) + x + 0];
+        const int a2 = ref[(w + 1) * (y + 0) + x + 1];
+        const int b1 = ref[(w + 1) * (y + 1) + x + 0];
+        const int b2 = ref[(w + 1) * (y + 1) + x + 1];
+        const int a = a1 + (((a2 - a1) * xoff + 8) >> 4);
+        const int b = b1 + (((b2 - b1) * xoff + 8) >> 4);
+        const int r = a + (((b - a) * yoff + 8) >> 4);
+        const int diff = r - src[w * y + x];
+        se += diff;
+        sse += diff * diff;
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        uint16_t *ref16 = CONVERT_TO_SHORTPTR(ref);
+        uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+        const int a1 = ref16[(w + 1) * (y + 0) + x + 0];
+        const int a2 = ref16[(w + 1) * (y + 0) + x + 1];
+        const int b1 = ref16[(w + 1) * (y + 1) + x + 0];
+        const int b2 = ref16[(w + 1) * (y + 1) + x + 1];
+        const int a = a1 + (((a2 - a1) * xoff + 8) >> 4);
+        const int b = b1 + (((b2 - b1) * xoff + 8) >> 4);
+        const int r = a + (((b - a) * yoff + 8) >> 4);
+        const int diff = r - src16[w * y + x];
+        se += diff;
+        sse += diff * diff;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+    }
+  }
+  RoundHighBitDepth(bit_depth, &se, &sse);
+  *sse_ptr = static_cast<uint32_t>(sse);
+  return static_cast<uint32_t>(sse -
+                               ((static_cast<int64_t>(se) * se) >>
+                                (l2w + l2h)));
+}
+
+class SumOfSquaresTest : public ::testing::TestWithParam<SumOfSquaresFunction> {
+ public:
+  SumOfSquaresTest() : func_(GetParam()) {}
+
+  virtual ~SumOfSquaresTest() {
+    libvpx_test::ClearSystemState();
+  }
+
+ protected:
+  void ConstTest();
+  void RefTest();
+
+  SumOfSquaresFunction func_;
+  ACMRandom rnd_;
+};
+
+void SumOfSquaresTest::ConstTest() {
+  int16_t mem[256];
+  unsigned int res;
+  for (int v = 0; v < 256; ++v) {
+    for (int i = 0; i < 256; ++i) {
+      mem[i] = v;
+    }
+    ASM_REGISTER_STATE_CHECK(res = func_(mem));
+    EXPECT_EQ(256u * (v * v), res);
+  }
+}
+
+void SumOfSquaresTest::RefTest() {
+  int16_t mem[256];
+  for (int i = 0; i < 100; ++i) {
+    for (int j = 0; j < 256; ++j) {
+      mem[j] = rnd_.Rand8() - rnd_.Rand8();
+    }
+
+    const unsigned int expected = mb_ss_ref(mem);
+    unsigned int res;
+    ASM_REGISTER_STATE_CHECK(res = func_(mem));
+    EXPECT_EQ(expected, res);
+  }
+}
+
+template<typename VarianceFunctionType>
+class VarianceTest
+    : public ::testing::TestWithParam<tuple<int, int,
+                                            VarianceFunctionType, int> > {
+ public:
+  virtual void SetUp() {
+    const tuple<int, int, VarianceFunctionType, int>& params = this->GetParam();
+    log2width_  = get<0>(params);
+    width_ = 1 << log2width_;
+    log2height_ = get<1>(params);
+    height_ = 1 << log2height_;
+    variance_ = get<2>(params);
+    if (get<3>(params)) {
+      bit_depth_ = static_cast<vpx_bit_depth_t>(get<3>(params));
+      use_high_bit_depth_ = true;
+    } else {
+      bit_depth_ = VPX_BITS_8;
+      use_high_bit_depth_ = false;
+    }
+    mask_ = (1 << bit_depth_) - 1;
+
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+    block_size_ = width_ * height_;
+    if (!use_high_bit_depth_) {
+      src_ = reinterpret_cast<uint8_t *>(vpx_memalign(16, block_size_ * 2));
+      ref_ = new uint8_t[block_size_ * 2];
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else {
+      src_ = CONVERT_TO_BYTEPTR(reinterpret_cast<uint16_t *>(
+          vpx_memalign(16, block_size_ * 2 * sizeof(uint16_t))));
+      ref_ = CONVERT_TO_BYTEPTR(new uint16_t[block_size_ * 2]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    }
+    ASSERT_TRUE(src_ != NULL);
+    ASSERT_TRUE(ref_ != NULL);
+  }
+
+  virtual void TearDown() {
+    if (!use_high_bit_depth_) {
+      vpx_free(src_);
+      delete[] ref_;
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else {
+      vpx_free(CONVERT_TO_SHORTPTR(src_));
+      delete[] CONVERT_TO_SHORTPTR(ref_);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    }
+    libvpx_test::ClearSystemState();
+  }
+
+ protected:
+  void ZeroTest();
+  void RefTest();
+  void RefStrideTest();
+  void OneQuarterTest();
+
+  ACMRandom rnd_;
+  uint8_t *src_;
+  uint8_t *ref_;
+  int width_, log2width_;
+  int height_, log2height_;
+  vpx_bit_depth_t bit_depth_;
+  int mask_;
+  bool use_high_bit_depth_;
+  int block_size_;
+  VarianceFunctionType variance_;
+};
+
+template<typename VarianceFunctionType>
+void VarianceTest<VarianceFunctionType>::ZeroTest() {
+  for (int i = 0; i <= 255; ++i) {
+    if (!use_high_bit_depth_) {
+      memset(src_, i, block_size_);
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else {
+      vpx_memset16(CONVERT_TO_SHORTPTR(src_), i << (bit_depth_ - 8),
+                   block_size_);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    }
+    for (int j = 0; j <= 255; ++j) {
+      if (!use_high_bit_depth_) {
+        memset(ref_, j, block_size_);
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        vpx_memset16(CONVERT_TO_SHORTPTR(ref_), j  << (bit_depth_ - 8),
+                     block_size_);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+      unsigned int sse;
+      unsigned int var;
+      ASM_REGISTER_STATE_CHECK(
+          var = variance_(src_, width_, ref_, width_, &sse));
+      EXPECT_EQ(0u, var) << "src values: " << i << " ref values: " << j;
+    }
+  }
+}
+
+template<typename VarianceFunctionType>
+void VarianceTest<VarianceFunctionType>::RefTest() {
+  for (int i = 0; i < 10; ++i) {
+    for (int j = 0; j < block_size_; j++) {
+      if (!use_high_bit_depth_) {
+        src_[j] = rnd_.Rand8();
+        ref_[j] = rnd_.Rand8();
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        CONVERT_TO_SHORTPTR(src_)[j] = rnd_.Rand16() & mask_;
+        CONVERT_TO_SHORTPTR(ref_)[j] = rnd_.Rand16() & mask_;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+    }
+    unsigned int sse1, sse2;
+    unsigned int var1;
+    const int stride_coeff = 1;
+    ASM_REGISTER_STATE_CHECK(
+        var1 = variance_(src_, width_, ref_, width_, &sse1));
+    const unsigned int var2 = variance_ref(src_, ref_, log2width_,
+                                           log2height_, stride_coeff,
+                                           stride_coeff, &sse2,
+                                           use_high_bit_depth_, bit_depth_);
+    EXPECT_EQ(sse1, sse2)
+        << "Error at test index: " << i;
+    EXPECT_EQ(var1, var2)
+        << "Error at test index: " << i;
+  }
+}
+
+template<typename VarianceFunctionType>
+void VarianceTest<VarianceFunctionType>::RefStrideTest() {
+  for (int i = 0; i < 10; ++i) {
+    int ref_stride_coeff = i % 2;
+    int src_stride_coeff = (i >> 1) % 2;
+    for (int j = 0; j < block_size_; j++) {
+      int ref_ind = (j / width_) * ref_stride_coeff * width_ + j % width_;
+      int src_ind = (j / width_) * src_stride_coeff * width_ + j % width_;
+      if (!use_high_bit_depth_) {
+        src_[src_ind] = rnd_.Rand8();
+        ref_[ref_ind] = rnd_.Rand8();
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        CONVERT_TO_SHORTPTR(src_)[src_ind] = rnd_.Rand16() & mask_;
+        CONVERT_TO_SHORTPTR(ref_)[ref_ind] = rnd_.Rand16() & mask_;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+    }
+    unsigned int sse1, sse2;
+    unsigned int var1;
+
+    ASM_REGISTER_STATE_CHECK(
+        var1 = variance_(src_, width_ * src_stride_coeff,
+                         ref_, width_ * ref_stride_coeff, &sse1));
+    const unsigned int var2 = variance_ref(src_, ref_, log2width_,
+                                           log2height_, src_stride_coeff,
+                                           ref_stride_coeff, &sse2,
+                                           use_high_bit_depth_, bit_depth_);
+    EXPECT_EQ(sse1, sse2)
+        << "Error at test index: " << i;
+    EXPECT_EQ(var1, var2)
+        << "Error at test index: " << i;
+  }
+}
+
+template<typename VarianceFunctionType>
+void VarianceTest<VarianceFunctionType>::OneQuarterTest() {
+  const int half = block_size_ / 2;
+  if (!use_high_bit_depth_) {
+    memset(src_, 255, block_size_);
+    memset(ref_, 255, half);
+    memset(ref_ + half, 0, half);
+#if CONFIG_VP9_HIGHBITDEPTH
+  } else {
+    vpx_memset16(CONVERT_TO_SHORTPTR(src_), 255 << (bit_depth_ - 8),
+                 block_size_);
+    vpx_memset16(CONVERT_TO_SHORTPTR(ref_), 255 << (bit_depth_ - 8), half);
+    vpx_memset16(CONVERT_TO_SHORTPTR(ref_) + half, 0, half);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+  unsigned int sse;
+  unsigned int var;
+  ASM_REGISTER_STATE_CHECK(var = variance_(src_, width_, ref_, width_, &sse));
+  const unsigned int expected = block_size_ * 255 * 255 / 4;
+  EXPECT_EQ(expected, var);
+}
+
+template<typename MseFunctionType>
+class MseTest
+    : public ::testing::TestWithParam<tuple<int, int, MseFunctionType> > {
+ public:
+  virtual void SetUp() {
+    const tuple<int, int, MseFunctionType>& params = this->GetParam();
+    log2width_  = get<0>(params);
+    width_ = 1 << log2width_;
+    log2height_ = get<1>(params);
+    height_ = 1 << log2height_;
+    mse_ = get<2>(params);
+
+    rnd(ACMRandom::DeterministicSeed());
+    block_size_ = width_ * height_;
+    src_ = reinterpret_cast<uint8_t *>(vpx_memalign(16, block_size_));
+    ref_ = new uint8_t[block_size_];
+    ASSERT_TRUE(src_ != NULL);
+    ASSERT_TRUE(ref_ != NULL);
+  }
+
+  virtual void TearDown() {
+    vpx_free(src_);
+    delete[] ref_;
+    libvpx_test::ClearSystemState();
+  }
+
+ protected:
+  void RefTest_mse();
+  void RefTest_sse();
+  void MaxTest_mse();
+  void MaxTest_sse();
+
+  ACMRandom rnd;
+  uint8_t* src_;
+  uint8_t* ref_;
+  int width_, log2width_;
+  int height_, log2height_;
+  int block_size_;
+  MseFunctionType mse_;
+};
+
+template<typename MseFunctionType>
+void MseTest<MseFunctionType>::RefTest_mse() {
+  for (int i = 0; i < 10; ++i) {
+    for (int j = 0; j < block_size_; j++) {
+      src_[j] = rnd.Rand8();
+      ref_[j] = rnd.Rand8();
+    }
+    unsigned int sse1, sse2;
+    const int stride_coeff = 1;
+    ASM_REGISTER_STATE_CHECK(mse_(src_, width_, ref_, width_, &sse1));
+    variance_ref(src_, ref_, log2width_, log2height_, stride_coeff,
+                 stride_coeff, &sse2, false, VPX_BITS_8);
+    EXPECT_EQ(sse1, sse2);
+  }
+}
+
+template<typename MseFunctionType>
+void MseTest<MseFunctionType>::RefTest_sse() {
+  for (int i = 0; i < 10; ++i) {
+    for (int j = 0; j < block_size_; j++) {
+      src_[j] = rnd.Rand8();
+      ref_[j] = rnd.Rand8();
+    }
+    unsigned int sse2;
+    unsigned int var1;
+    const int stride_coeff = 1;
+    ASM_REGISTER_STATE_CHECK(var1 = mse_(src_, width_, ref_, width_));
+    variance_ref(src_, ref_, log2width_, log2height_, stride_coeff,
+                 stride_coeff, &sse2, false, VPX_BITS_8);
+    EXPECT_EQ(var1, sse2);
+  }
+}
+
+template<typename MseFunctionType>
+void MseTest<MseFunctionType>::MaxTest_mse() {
+  memset(src_, 255, block_size_);
+  memset(ref_, 0, block_size_);
+  unsigned int sse;
+  ASM_REGISTER_STATE_CHECK(mse_(src_, width_, ref_, width_, &sse));
+  const unsigned int expected = block_size_ * 255 * 255;
+  EXPECT_EQ(expected, sse);
+}
+
+template<typename MseFunctionType>
+void MseTest<MseFunctionType>::MaxTest_sse() {
+  memset(src_, 255, block_size_);
+  memset(ref_, 0, block_size_);
+  unsigned int var;
+  ASM_REGISTER_STATE_CHECK(var = mse_(src_, width_, ref_, width_));
+  const unsigned int expected = block_size_ * 255 * 255;
+  EXPECT_EQ(expected, var);
+}
+
+static uint32_t subpel_avg_variance_ref(const uint8_t *ref,
+                                        const uint8_t *src,
+                                        const uint8_t *second_pred,
+                                        int l2w, int l2h,
+                                        int xoff, int yoff,
+                                        uint32_t *sse_ptr,
+                                        bool use_high_bit_depth,
+                                        vpx_bit_depth_t bit_depth) {
+  int64_t se = 0;
+  uint64_t sse = 0;
+  const int w = 1 << l2w;
+  const int h = 1 << l2h;
+
+  xoff <<= 1;
+  yoff <<= 1;
+
+  for (int y = 0; y < h; y++) {
+    for (int x = 0; x < w; x++) {
+      // bilinear interpolation at a 16th pel step
+      if (!use_high_bit_depth) {
+        const int a1 = ref[(w + 1) * (y + 0) + x + 0];
+        const int a2 = ref[(w + 1) * (y + 0) + x + 1];
+        const int b1 = ref[(w + 1) * (y + 1) + x + 0];
+        const int b2 = ref[(w + 1) * (y + 1) + x + 1];
+        const int a = a1 + (((a2 - a1) * xoff + 8) >> 4);
+        const int b = b1 + (((b2 - b1) * xoff + 8) >> 4);
+        const int r = a + (((b - a) * yoff + 8) >> 4);
+        const int diff = ((r + second_pred[w * y + x] + 1) >> 1) - src[w * y + x];
+        se += diff;
+        sse += diff * diff;
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        uint16_t *ref16 = CONVERT_TO_SHORTPTR(ref);
+        uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+        uint16_t *sec16   = CONVERT_TO_SHORTPTR(second_pred);
+        const int a1 = ref16[(w + 1) * (y + 0) + x + 0];
+        const int a2 = ref16[(w + 1) * (y + 0) + x + 1];
+        const int b1 = ref16[(w + 1) * (y + 1) + x + 0];
+        const int b2 = ref16[(w + 1) * (y + 1) + x + 1];
+        const int a = a1 + (((a2 - a1) * xoff + 8) >> 4);
+        const int b = b1 + (((b2 - b1) * xoff + 8) >> 4);
+        const int r = a + (((b - a) * yoff + 8) >> 4);
+        const int diff = ((r + sec16[w * y + x] + 1) >> 1) - src16[w * y + x];
+        se += diff;
+        sse += diff * diff;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+    }
+  }
+  RoundHighBitDepth(bit_depth, &se, &sse);
+  *sse_ptr = static_cast<uint32_t>(sse);
+  return static_cast<uint32_t>(sse -
+                               ((static_cast<int64_t>(se) * se) >>
+                                (l2w + l2h)));
+}
+
+template<typename SubpelVarianceFunctionType>
+class SubpelVarianceTest
+    : public ::testing::TestWithParam<tuple<int, int,
+                                            SubpelVarianceFunctionType, int> > {
+ public:
+  virtual void SetUp() {
+    const tuple<int, int, SubpelVarianceFunctionType, int>& params =
+        this->GetParam();
+    log2width_  = get<0>(params);
+    width_ = 1 << log2width_;
+    log2height_ = get<1>(params);
+    height_ = 1 << log2height_;
+    subpel_variance_ = get<2>(params);
+    if (get<3>(params)) {
+      bit_depth_ = (vpx_bit_depth_t) get<3>(params);
+      use_high_bit_depth_ = true;
+    } else {
+      bit_depth_ = VPX_BITS_8;
+      use_high_bit_depth_ = false;
+    }
+    mask_ = (1 << bit_depth_)-1;
+
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+    block_size_ = width_ * height_;
+    if (!use_high_bit_depth_) {
+      src_ = reinterpret_cast<uint8_t *>(vpx_memalign(16, block_size_));
+      sec_ = reinterpret_cast<uint8_t *>(vpx_memalign(16, block_size_));
+      ref_ = new uint8_t[block_size_ + width_ + height_ + 1];
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else {
+      src_ = CONVERT_TO_BYTEPTR(
+          reinterpret_cast<uint16_t *>(
+              vpx_memalign(16, block_size_*sizeof(uint16_t))));
+      sec_ = CONVERT_TO_BYTEPTR(
+          reinterpret_cast<uint16_t *>(
+              vpx_memalign(16, block_size_*sizeof(uint16_t))));
+      ref_ = CONVERT_TO_BYTEPTR(
+          new uint16_t[block_size_ + width_ + height_ + 1]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    }
+    ASSERT_TRUE(src_ != NULL);
+    ASSERT_TRUE(sec_ != NULL);
+    ASSERT_TRUE(ref_ != NULL);
+  }
+
+  virtual void TearDown() {
+    if (!use_high_bit_depth_) {
+      vpx_free(src_);
+      delete[] ref_;
+      vpx_free(sec_);
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else {
+      vpx_free(CONVERT_TO_SHORTPTR(src_));
+      delete[] CONVERT_TO_SHORTPTR(ref_);
+      vpx_free(CONVERT_TO_SHORTPTR(sec_));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    }
+    libvpx_test::ClearSystemState();
+  }
+
+ protected:
+  void RefTest();
+  void ExtremeRefTest();
+
+  ACMRandom rnd_;
+  uint8_t *src_;
+  uint8_t *ref_;
+  uint8_t *sec_;
+  bool use_high_bit_depth_;
+  vpx_bit_depth_t bit_depth_;
+  int width_, log2width_;
+  int height_, log2height_;
+  int block_size_,  mask_;
+  SubpelVarianceFunctionType subpel_variance_;
+};
+
+template<typename SubpelVarianceFunctionType>
+void SubpelVarianceTest<SubpelVarianceFunctionType>::RefTest() {
+  for (int x = 0; x < 8; ++x) {
+    for (int y = 0; y < 8; ++y) {
+      if (!use_high_bit_depth_) {
+        for (int j = 0; j < block_size_; j++) {
+          src_[j] = rnd_.Rand8();
+        }
+        for (int j = 0; j < block_size_ + width_ + height_ + 1; j++) {
+          ref_[j] = rnd_.Rand8();
+        }
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        for (int j = 0; j < block_size_; j++) {
+          CONVERT_TO_SHORTPTR(src_)[j] = rnd_.Rand16() & mask_;
+        }
+        for (int j = 0; j < block_size_ + width_ + height_ + 1; j++) {
+          CONVERT_TO_SHORTPTR(ref_)[j] = rnd_.Rand16() & mask_;
+        }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+      unsigned int sse1, sse2;
+      unsigned int var1;
+      ASM_REGISTER_STATE_CHECK(var1 = subpel_variance_(ref_, width_ + 1, x, y,
+                                                       src_, width_, &sse1));
+      const unsigned int var2 = subpel_variance_ref(ref_, src_,
+                                                    log2width_, log2height_,
+                                                    x, y, &sse2,
+                                                    use_high_bit_depth_,
+                                                    bit_depth_);
+      EXPECT_EQ(sse1, sse2) << "at position " << x << ", " << y;
+      EXPECT_EQ(var1, var2) << "at position " << x << ", " << y;
+    }
+  }
+}
+
+template<typename SubpelVarianceFunctionType>
+void SubpelVarianceTest<SubpelVarianceFunctionType>::ExtremeRefTest() {
+  // Compare against reference.
+  // Src: Set the first half of values to 0, the second half to the maximum.
+  // Ref: Set the first half of values to the maximum, the second half to 0.
+  for (int x = 0; x < 8; ++x) {
+    for (int y = 0; y < 8; ++y) {
+      const int half = block_size_ / 2;
+      if (!use_high_bit_depth_) {
+        memset(src_, 0, half);
+        memset(src_ + half, 255, half);
+        memset(ref_, 255, half);
+        memset(ref_ + half, 0, half + width_ + height_ + 1);
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        vpx_memset16(CONVERT_TO_SHORTPTR(src_), mask_, half);
+        vpx_memset16(CONVERT_TO_SHORTPTR(src_) + half, 0, half);
+        vpx_memset16(CONVERT_TO_SHORTPTR(ref_), 0, half);
+        vpx_memset16(CONVERT_TO_SHORTPTR(ref_) + half, mask_,
+                     half + width_ + height_ + 1);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+      unsigned int sse1, sse2;
+      unsigned int var1;
+      ASM_REGISTER_STATE_CHECK(
+          var1 = subpel_variance_(ref_, width_ + 1, x, y, src_, width_, &sse1));
+      const unsigned int var2 =
+          subpel_variance_ref(ref_, src_, log2width_, log2height_,
+                              x, y, &sse2, use_high_bit_depth_, bit_depth_);
+      EXPECT_EQ(sse1, sse2) << "for xoffset " << x << " and yoffset " << y;
+      EXPECT_EQ(var1, var2) << "for xoffset " << x << " and yoffset " << y;
+    }
+  }
+}
+
+template<>
+void SubpelVarianceTest<SubpixAvgVarMxNFunc>::RefTest() {
+  for (int x = 0; x < 8; ++x) {
+    for (int y = 0; y < 8; ++y) {
+      if (!use_high_bit_depth_) {
+        for (int j = 0; j < block_size_; j++) {
+          src_[j] = rnd_.Rand8();
+          sec_[j] = rnd_.Rand8();
+        }
+        for (int j = 0; j < block_size_ + width_ + height_ + 1; j++) {
+          ref_[j] = rnd_.Rand8();
+        }
+#if CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        for (int j = 0; j < block_size_; j++) {
+          CONVERT_TO_SHORTPTR(src_)[j] = rnd_.Rand16() & mask_;
+          CONVERT_TO_SHORTPTR(sec_)[j] = rnd_.Rand16() & mask_;
+        }
+        for (int j = 0; j < block_size_ + width_ + height_ + 1; j++) {
+          CONVERT_TO_SHORTPTR(ref_)[j] = rnd_.Rand16() & mask_;
+        }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      }
+      unsigned int sse1, sse2;
+      unsigned int var1;
+      ASM_REGISTER_STATE_CHECK(
+          var1 = subpel_variance_(ref_, width_ + 1, x, y,
+                                  src_, width_, &sse1, sec_));
+      const unsigned int var2 = subpel_avg_variance_ref(ref_, src_, sec_,
+                                                        log2width_, log2height_,
+                                                        x, y, &sse2,
+                                                        use_high_bit_depth_,
+                                                        bit_depth_);
+      EXPECT_EQ(sse1, sse2) << "at position " << x << ", " << y;
+      EXPECT_EQ(var1, var2) << "at position " << x << ", " << y;
+    }
+  }
+}
+
+typedef MseTest<Get4x4SseFunc> VpxSseTest;
+typedef MseTest<VarianceMxNFunc> VpxMseTest;
+typedef VarianceTest<VarianceMxNFunc> VpxVarianceTest;
+typedef SubpelVarianceTest<SubpixVarMxNFunc> VpxSubpelVarianceTest;
+typedef SubpelVarianceTest<SubpixAvgVarMxNFunc> VpxSubpelAvgVarianceTest;
+
+TEST_P(VpxSseTest, Ref_sse) { RefTest_sse(); }
+TEST_P(VpxSseTest, Max_sse) { MaxTest_sse(); }
+TEST_P(VpxMseTest, Ref_mse) { RefTest_mse(); }
+TEST_P(VpxMseTest, Max_mse) { MaxTest_mse(); }
+TEST_P(VpxVarianceTest, Zero) { ZeroTest(); }
+TEST_P(VpxVarianceTest, Ref) { RefTest(); }
+TEST_P(VpxVarianceTest, RefStride) { RefStrideTest(); }
+TEST_P(VpxVarianceTest, OneQuarter) { OneQuarterTest(); }
+TEST_P(SumOfSquaresTest, Const) { ConstTest(); }
+TEST_P(SumOfSquaresTest, Ref) { RefTest(); }
+TEST_P(VpxSubpelVarianceTest, Ref) { RefTest(); }
+TEST_P(VpxSubpelVarianceTest, ExtremeRef) { ExtremeRefTest(); }
+TEST_P(VpxSubpelAvgVarianceTest, Ref) { RefTest(); }
+
+INSTANTIATE_TEST_CASE_P(C, SumOfSquaresTest,
+                        ::testing::Values(vpx_get_mb_ss_c));
+
+INSTANTIATE_TEST_CASE_P(C, VpxSseTest,
+                        ::testing::Values(make_tuple(2, 2,
+                                                     &vpx_get4x4sse_cs_c)));
+
+INSTANTIATE_TEST_CASE_P(C, VpxMseTest,
+                        ::testing::Values(make_tuple(4, 4, &vpx_mse16x16_c),
+                                          make_tuple(4, 3, &vpx_mse16x8_c),
+                                          make_tuple(3, 4, &vpx_mse8x16_c),
+                                          make_tuple(3, 3, &vpx_mse8x8_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    C, VpxVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_variance64x64_c, 0),
+                      make_tuple(6, 5, &vpx_variance64x32_c, 0),
+                      make_tuple(5, 6, &vpx_variance32x64_c, 0),
+                      make_tuple(5, 5, &vpx_variance32x32_c, 0),
+                      make_tuple(5, 4, &vpx_variance32x16_c, 0),
+                      make_tuple(4, 5, &vpx_variance16x32_c, 0),
+                      make_tuple(4, 4, &vpx_variance16x16_c, 0),
+                      make_tuple(4, 3, &vpx_variance16x8_c, 0),
+                      make_tuple(3, 4, &vpx_variance8x16_c, 0),
+                      make_tuple(3, 3, &vpx_variance8x8_c, 0),
+                      make_tuple(3, 2, &vpx_variance8x4_c, 0),
+                      make_tuple(2, 3, &vpx_variance4x8_c, 0),
+                      make_tuple(2, 2, &vpx_variance4x4_c, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    C, VpxSubpelVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_sub_pixel_variance64x64_c, 0),
+                      make_tuple(6, 5, &vpx_sub_pixel_variance64x32_c, 0),
+                      make_tuple(5, 6, &vpx_sub_pixel_variance32x64_c, 0),
+                      make_tuple(5, 5, &vpx_sub_pixel_variance32x32_c, 0),
+                      make_tuple(5, 4, &vpx_sub_pixel_variance32x16_c, 0),
+                      make_tuple(4, 5, &vpx_sub_pixel_variance16x32_c, 0),
+                      make_tuple(4, 4, &vpx_sub_pixel_variance16x16_c, 0),
+                      make_tuple(4, 3, &vpx_sub_pixel_variance16x8_c, 0),
+                      make_tuple(3, 4, &vpx_sub_pixel_variance8x16_c, 0),
+                      make_tuple(3, 3, &vpx_sub_pixel_variance8x8_c, 0),
+                      make_tuple(3, 2, &vpx_sub_pixel_variance8x4_c, 0),
+                      make_tuple(2, 3, &vpx_sub_pixel_variance4x8_c, 0),
+                      make_tuple(2, 2, &vpx_sub_pixel_variance4x4_c, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    C, VpxSubpelAvgVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_sub_pixel_avg_variance64x64_c, 0),
+                      make_tuple(6, 5, &vpx_sub_pixel_avg_variance64x32_c, 0),
+                      make_tuple(5, 6, &vpx_sub_pixel_avg_variance32x64_c, 0),
+                      make_tuple(5, 5, &vpx_sub_pixel_avg_variance32x32_c, 0),
+                      make_tuple(5, 4, &vpx_sub_pixel_avg_variance32x16_c, 0),
+                      make_tuple(4, 5, &vpx_sub_pixel_avg_variance16x32_c, 0),
+                      make_tuple(4, 4, &vpx_sub_pixel_avg_variance16x16_c, 0),
+                      make_tuple(4, 3, &vpx_sub_pixel_avg_variance16x8_c, 0),
+                      make_tuple(3, 4, &vpx_sub_pixel_avg_variance8x16_c, 0),
+                      make_tuple(3, 3, &vpx_sub_pixel_avg_variance8x8_c, 0),
+                      make_tuple(3, 2, &vpx_sub_pixel_avg_variance8x4_c, 0),
+                      make_tuple(2, 3, &vpx_sub_pixel_avg_variance4x8_c, 0),
+                      make_tuple(2, 2, &vpx_sub_pixel_avg_variance4x4_c, 0)));
+
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef MseTest<VarianceMxNFunc> VpxHBDMseTest;
+typedef VarianceTest<VarianceMxNFunc> VpxHBDVarianceTest;
+typedef SubpelVarianceTest<SubpixVarMxNFunc> VpxHBDSubpelVarianceTest;
+typedef SubpelVarianceTest<SubpixAvgVarMxNFunc>
+    VpxHBDSubpelAvgVarianceTest;
+
+TEST_P(VpxHBDMseTest, Ref_mse) { RefTest_mse(); }
+TEST_P(VpxHBDMseTest, Max_mse) { MaxTest_mse(); }
+TEST_P(VpxHBDVarianceTest, Zero) { ZeroTest(); }
+TEST_P(VpxHBDVarianceTest, Ref) { RefTest(); }
+TEST_P(VpxHBDVarianceTest, RefStride) { RefStrideTest(); }
+TEST_P(VpxHBDVarianceTest, OneQuarter) { OneQuarterTest(); }
+TEST_P(VpxHBDSubpelVarianceTest, Ref) { RefTest(); }
+TEST_P(VpxHBDSubpelVarianceTest, ExtremeRef) { ExtremeRefTest(); }
+TEST_P(VpxHBDSubpelAvgVarianceTest, Ref) { RefTest(); }
+
+/* TODO(debargha): This test does not support the highbd version
+INSTANTIATE_TEST_CASE_P(
+    C, VpxHBDMseTest,
+    ::testing::Values(make_tuple(4, 4, &vpx_highbd_12_mse16x16_c),
+                      make_tuple(4, 4, &vpx_highbd_12_mse16x8_c),
+                      make_tuple(4, 4, &vpx_highbd_12_mse8x16_c),
+                      make_tuple(4, 4, &vpx_highbd_12_mse8x8_c),
+                      make_tuple(4, 4, &vpx_highbd_10_mse16x16_c),
+                      make_tuple(4, 4, &vpx_highbd_10_mse16x8_c),
+                      make_tuple(4, 4, &vpx_highbd_10_mse8x16_c),
+                      make_tuple(4, 4, &vpx_highbd_10_mse8x8_c),
+                      make_tuple(4, 4, &vpx_highbd_8_mse16x16_c),
+                      make_tuple(4, 4, &vpx_highbd_8_mse16x8_c),
+                      make_tuple(4, 4, &vpx_highbd_8_mse8x16_c),
+                      make_tuple(4, 4, &vpx_highbd_8_mse8x8_c)));
+*/
+
+INSTANTIATE_TEST_CASE_P(
+    C, VpxHBDVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_highbd_12_variance64x64_c, 12),
+                      make_tuple(6, 5, &vpx_highbd_12_variance64x32_c, 12),
+                      make_tuple(5, 6, &vpx_highbd_12_variance32x64_c, 12),
+                      make_tuple(5, 5, &vpx_highbd_12_variance32x32_c, 12),
+                      make_tuple(5, 4, &vpx_highbd_12_variance32x16_c, 12),
+                      make_tuple(4, 5, &vpx_highbd_12_variance16x32_c, 12),
+                      make_tuple(4, 4, &vpx_highbd_12_variance16x16_c, 12),
+                      make_tuple(4, 3, &vpx_highbd_12_variance16x8_c, 12),
+                      make_tuple(3, 4, &vpx_highbd_12_variance8x16_c, 12),
+                      make_tuple(3, 3, &vpx_highbd_12_variance8x8_c, 12),
+                      make_tuple(3, 2, &vpx_highbd_12_variance8x4_c, 12),
+                      make_tuple(2, 3, &vpx_highbd_12_variance4x8_c, 12),
+                      make_tuple(2, 2, &vpx_highbd_12_variance4x4_c, 12),
+                      make_tuple(6, 6, &vpx_highbd_10_variance64x64_c, 10),
+                      make_tuple(6, 5, &vpx_highbd_10_variance64x32_c, 10),
+                      make_tuple(5, 6, &vpx_highbd_10_variance32x64_c, 10),
+                      make_tuple(5, 5, &vpx_highbd_10_variance32x32_c, 10),
+                      make_tuple(5, 4, &vpx_highbd_10_variance32x16_c, 10),
+                      make_tuple(4, 5, &vpx_highbd_10_variance16x32_c, 10),
+                      make_tuple(4, 4, &vpx_highbd_10_variance16x16_c, 10),
+                      make_tuple(4, 3, &vpx_highbd_10_variance16x8_c, 10),
+                      make_tuple(3, 4, &vpx_highbd_10_variance8x16_c, 10),
+                      make_tuple(3, 3, &vpx_highbd_10_variance8x8_c, 10),
+                      make_tuple(3, 2, &vpx_highbd_10_variance8x4_c, 10),
+                      make_tuple(2, 3, &vpx_highbd_10_variance4x8_c, 10),
+                      make_tuple(2, 2, &vpx_highbd_10_variance4x4_c, 10),
+                      make_tuple(6, 6, &vpx_highbd_8_variance64x64_c, 8),
+                      make_tuple(6, 5, &vpx_highbd_8_variance64x32_c, 8),
+                      make_tuple(5, 6, &vpx_highbd_8_variance32x64_c, 8),
+                      make_tuple(5, 5, &vpx_highbd_8_variance32x32_c, 8),
+                      make_tuple(5, 4, &vpx_highbd_8_variance32x16_c, 8),
+                      make_tuple(4, 5, &vpx_highbd_8_variance16x32_c, 8),
+                      make_tuple(4, 4, &vpx_highbd_8_variance16x16_c, 8),
+                      make_tuple(4, 3, &vpx_highbd_8_variance16x8_c, 8),
+                      make_tuple(3, 4, &vpx_highbd_8_variance8x16_c, 8),
+                      make_tuple(3, 3, &vpx_highbd_8_variance8x8_c, 8),
+                      make_tuple(3, 2, &vpx_highbd_8_variance8x4_c, 8),
+                      make_tuple(2, 3, &vpx_highbd_8_variance4x8_c, 8),
+                      make_tuple(2, 2, &vpx_highbd_8_variance4x4_c, 8)));
+
+INSTANTIATE_TEST_CASE_P(
+    C, VpxHBDSubpelVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &vpx_highbd_8_sub_pixel_variance64x64_c, 8),
+        make_tuple(6, 5, &vpx_highbd_8_sub_pixel_variance64x32_c, 8),
+        make_tuple(5, 6, &vpx_highbd_8_sub_pixel_variance32x64_c, 8),
+        make_tuple(5, 5, &vpx_highbd_8_sub_pixel_variance32x32_c, 8),
+        make_tuple(5, 4, &vpx_highbd_8_sub_pixel_variance32x16_c, 8),
+        make_tuple(4, 5, &vpx_highbd_8_sub_pixel_variance16x32_c, 8),
+        make_tuple(4, 4, &vpx_highbd_8_sub_pixel_variance16x16_c, 8),
+        make_tuple(4, 3, &vpx_highbd_8_sub_pixel_variance16x8_c, 8),
+        make_tuple(3, 4, &vpx_highbd_8_sub_pixel_variance8x16_c, 8),
+        make_tuple(3, 3, &vpx_highbd_8_sub_pixel_variance8x8_c, 8),
+        make_tuple(3, 2, &vpx_highbd_8_sub_pixel_variance8x4_c, 8),
+        make_tuple(2, 3, &vpx_highbd_8_sub_pixel_variance4x8_c, 8),
+        make_tuple(2, 2, &vpx_highbd_8_sub_pixel_variance4x4_c, 8),
+        make_tuple(6, 6, &vpx_highbd_10_sub_pixel_variance64x64_c, 10),
+        make_tuple(6, 5, &vpx_highbd_10_sub_pixel_variance64x32_c, 10),
+        make_tuple(5, 6, &vpx_highbd_10_sub_pixel_variance32x64_c, 10),
+        make_tuple(5, 5, &vpx_highbd_10_sub_pixel_variance32x32_c, 10),
+        make_tuple(5, 4, &vpx_highbd_10_sub_pixel_variance32x16_c, 10),
+        make_tuple(4, 5, &vpx_highbd_10_sub_pixel_variance16x32_c, 10),
+        make_tuple(4, 4, &vpx_highbd_10_sub_pixel_variance16x16_c, 10),
+        make_tuple(4, 3, &vpx_highbd_10_sub_pixel_variance16x8_c, 10),
+        make_tuple(3, 4, &vpx_highbd_10_sub_pixel_variance8x16_c, 10),
+        make_tuple(3, 3, &vpx_highbd_10_sub_pixel_variance8x8_c, 10),
+        make_tuple(3, 2, &vpx_highbd_10_sub_pixel_variance8x4_c, 10),
+        make_tuple(2, 3, &vpx_highbd_10_sub_pixel_variance4x8_c, 10),
+        make_tuple(2, 2, &vpx_highbd_10_sub_pixel_variance4x4_c, 10),
+        make_tuple(6, 6, &vpx_highbd_12_sub_pixel_variance64x64_c, 12),
+        make_tuple(6, 5, &vpx_highbd_12_sub_pixel_variance64x32_c, 12),
+        make_tuple(5, 6, &vpx_highbd_12_sub_pixel_variance32x64_c, 12),
+        make_tuple(5, 5, &vpx_highbd_12_sub_pixel_variance32x32_c, 12),
+        make_tuple(5, 4, &vpx_highbd_12_sub_pixel_variance32x16_c, 12),
+        make_tuple(4, 5, &vpx_highbd_12_sub_pixel_variance16x32_c, 12),
+        make_tuple(4, 4, &vpx_highbd_12_sub_pixel_variance16x16_c, 12),
+        make_tuple(4, 3, &vpx_highbd_12_sub_pixel_variance16x8_c, 12),
+        make_tuple(3, 4, &vpx_highbd_12_sub_pixel_variance8x16_c, 12),
+        make_tuple(3, 3, &vpx_highbd_12_sub_pixel_variance8x8_c, 12),
+        make_tuple(3, 2, &vpx_highbd_12_sub_pixel_variance8x4_c, 12),
+        make_tuple(2, 3, &vpx_highbd_12_sub_pixel_variance4x8_c, 12),
+        make_tuple(2, 2, &vpx_highbd_12_sub_pixel_variance4x4_c, 12)));
+
+INSTANTIATE_TEST_CASE_P(
+    C, VpxHBDSubpelAvgVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &vpx_highbd_8_sub_pixel_avg_variance64x64_c, 8),
+        make_tuple(6, 5, &vpx_highbd_8_sub_pixel_avg_variance64x32_c, 8),
+        make_tuple(5, 6, &vpx_highbd_8_sub_pixel_avg_variance32x64_c, 8),
+        make_tuple(5, 5, &vpx_highbd_8_sub_pixel_avg_variance32x32_c, 8),
+        make_tuple(5, 4, &vpx_highbd_8_sub_pixel_avg_variance32x16_c, 8),
+        make_tuple(4, 5, &vpx_highbd_8_sub_pixel_avg_variance16x32_c, 8),
+        make_tuple(4, 4, &vpx_highbd_8_sub_pixel_avg_variance16x16_c, 8),
+        make_tuple(4, 3, &vpx_highbd_8_sub_pixel_avg_variance16x8_c, 8),
+        make_tuple(3, 4, &vpx_highbd_8_sub_pixel_avg_variance8x16_c, 8),
+        make_tuple(3, 3, &vpx_highbd_8_sub_pixel_avg_variance8x8_c, 8),
+        make_tuple(3, 2, &vpx_highbd_8_sub_pixel_avg_variance8x4_c, 8),
+        make_tuple(2, 3, &vpx_highbd_8_sub_pixel_avg_variance4x8_c, 8),
+        make_tuple(2, 2, &vpx_highbd_8_sub_pixel_avg_variance4x4_c, 8),
+        make_tuple(6, 6, &vpx_highbd_10_sub_pixel_avg_variance64x64_c, 10),
+        make_tuple(6, 5, &vpx_highbd_10_sub_pixel_avg_variance64x32_c, 10),
+        make_tuple(5, 6, &vpx_highbd_10_sub_pixel_avg_variance32x64_c, 10),
+        make_tuple(5, 5, &vpx_highbd_10_sub_pixel_avg_variance32x32_c, 10),
+        make_tuple(5, 4, &vpx_highbd_10_sub_pixel_avg_variance32x16_c, 10),
+        make_tuple(4, 5, &vpx_highbd_10_sub_pixel_avg_variance16x32_c, 10),
+        make_tuple(4, 4, &vpx_highbd_10_sub_pixel_avg_variance16x16_c, 10),
+        make_tuple(4, 3, &vpx_highbd_10_sub_pixel_avg_variance16x8_c, 10),
+        make_tuple(3, 4, &vpx_highbd_10_sub_pixel_avg_variance8x16_c, 10),
+        make_tuple(3, 3, &vpx_highbd_10_sub_pixel_avg_variance8x8_c, 10),
+        make_tuple(3, 2, &vpx_highbd_10_sub_pixel_avg_variance8x4_c, 10),
+        make_tuple(2, 3, &vpx_highbd_10_sub_pixel_avg_variance4x8_c, 10),
+        make_tuple(2, 2, &vpx_highbd_10_sub_pixel_avg_variance4x4_c, 10),
+        make_tuple(6, 6, &vpx_highbd_12_sub_pixel_avg_variance64x64_c, 12),
+        make_tuple(6, 5, &vpx_highbd_12_sub_pixel_avg_variance64x32_c, 12),
+        make_tuple(5, 6, &vpx_highbd_12_sub_pixel_avg_variance32x64_c, 12),
+        make_tuple(5, 5, &vpx_highbd_12_sub_pixel_avg_variance32x32_c, 12),
+        make_tuple(5, 4, &vpx_highbd_12_sub_pixel_avg_variance32x16_c, 12),
+        make_tuple(4, 5, &vpx_highbd_12_sub_pixel_avg_variance16x32_c, 12),
+        make_tuple(4, 4, &vpx_highbd_12_sub_pixel_avg_variance16x16_c, 12),
+        make_tuple(4, 3, &vpx_highbd_12_sub_pixel_avg_variance16x8_c, 12),
+        make_tuple(3, 4, &vpx_highbd_12_sub_pixel_avg_variance8x16_c, 12),
+        make_tuple(3, 3, &vpx_highbd_12_sub_pixel_avg_variance8x8_c, 12),
+        make_tuple(3, 2, &vpx_highbd_12_sub_pixel_avg_variance8x4_c, 12),
+        make_tuple(2, 3, &vpx_highbd_12_sub_pixel_avg_variance4x8_c, 12),
+        make_tuple(2, 2, &vpx_highbd_12_sub_pixel_avg_variance4x4_c, 12)));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, SumOfSquaresTest,
+                        ::testing::Values(vpx_get_mb_ss_sse2));
+
+INSTANTIATE_TEST_CASE_P(SSE2, VpxMseTest,
+                        ::testing::Values(make_tuple(4, 4, &vpx_mse16x16_sse2),
+                                          make_tuple(4, 3, &vpx_mse16x8_sse2),
+                                          make_tuple(3, 4, &vpx_mse8x16_sse2),
+                                          make_tuple(3, 3, &vpx_mse8x8_sse2)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VpxVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_variance64x64_sse2, 0),
+                      make_tuple(6, 5, &vpx_variance64x32_sse2, 0),
+                      make_tuple(5, 6, &vpx_variance32x64_sse2, 0),
+                      make_tuple(5, 5, &vpx_variance32x32_sse2, 0),
+                      make_tuple(5, 4, &vpx_variance32x16_sse2, 0),
+                      make_tuple(4, 5, &vpx_variance16x32_sse2, 0),
+                      make_tuple(4, 4, &vpx_variance16x16_sse2, 0),
+                      make_tuple(4, 3, &vpx_variance16x8_sse2, 0),
+                      make_tuple(3, 4, &vpx_variance8x16_sse2, 0),
+                      make_tuple(3, 3, &vpx_variance8x8_sse2, 0),
+                      make_tuple(3, 2, &vpx_variance8x4_sse2, 0),
+                      make_tuple(2, 3, &vpx_variance4x8_sse2, 0),
+                      make_tuple(2, 2, &vpx_variance4x4_sse2, 0)));
+
+#if CONFIG_USE_X86INC
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VpxSubpelVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_sub_pixel_variance64x64_sse2, 0),
+                      make_tuple(6, 5, &vpx_sub_pixel_variance64x32_sse2, 0),
+                      make_tuple(5, 6, &vpx_sub_pixel_variance32x64_sse2, 0),
+                      make_tuple(5, 5, &vpx_sub_pixel_variance32x32_sse2, 0),
+                      make_tuple(5, 4, &vpx_sub_pixel_variance32x16_sse2, 0),
+                      make_tuple(4, 5, &vpx_sub_pixel_variance16x32_sse2, 0),
+                      make_tuple(4, 4, &vpx_sub_pixel_variance16x16_sse2, 0),
+                      make_tuple(4, 3, &vpx_sub_pixel_variance16x8_sse2, 0),
+                      make_tuple(3, 4, &vpx_sub_pixel_variance8x16_sse2, 0),
+                      make_tuple(3, 3, &vpx_sub_pixel_variance8x8_sse2, 0),
+                      make_tuple(3, 2, &vpx_sub_pixel_variance8x4_sse2, 0),
+                      make_tuple(2, 3, &vpx_sub_pixel_variance4x8_sse2, 0),
+                      make_tuple(2, 2, &vpx_sub_pixel_variance4x4_sse2, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VpxSubpelAvgVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &vpx_sub_pixel_avg_variance64x64_sse2, 0),
+        make_tuple(6, 5, &vpx_sub_pixel_avg_variance64x32_sse2, 0),
+        make_tuple(5, 6, &vpx_sub_pixel_avg_variance32x64_sse2, 0),
+        make_tuple(5, 5, &vpx_sub_pixel_avg_variance32x32_sse2, 0),
+        make_tuple(5, 4, &vpx_sub_pixel_avg_variance32x16_sse2, 0),
+        make_tuple(4, 5, &vpx_sub_pixel_avg_variance16x32_sse2, 0),
+        make_tuple(4, 4, &vpx_sub_pixel_avg_variance16x16_sse2, 0),
+        make_tuple(4, 3, &vpx_sub_pixel_avg_variance16x8_sse2, 0),
+        make_tuple(3, 4, &vpx_sub_pixel_avg_variance8x16_sse2, 0),
+        make_tuple(3, 3, &vpx_sub_pixel_avg_variance8x8_sse2, 0),
+        make_tuple(3, 2, &vpx_sub_pixel_avg_variance8x4_sse2, 0),
+        make_tuple(2, 3, &vpx_sub_pixel_avg_variance4x8_sse2, 0),
+        make_tuple(2, 2, &vpx_sub_pixel_avg_variance4x4_sse2, 0)));
+#endif  // CONFIG_USE_X86INC
+
+#if CONFIG_VP9_HIGHBITDEPTH
+/* TODO(debargha): This test does not support the highbd version
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VpxHBDMseTest,
+    ::testing::Values(make_tuple(4, 4, &vpx_highbd_12_mse16x16_sse2),
+                      make_tuple(4, 3, &vpx_highbd_12_mse16x8_sse2),
+                      make_tuple(3, 4, &vpx_highbd_12_mse8x16_sse2),
+                      make_tuple(3, 3, &vpx_highbd_12_mse8x8_sse2),
+                      make_tuple(4, 4, &vpx_highbd_10_mse16x16_sse2),
+                      make_tuple(4, 3, &vpx_highbd_10_mse16x8_sse2),
+                      make_tuple(3, 4, &vpx_highbd_10_mse8x16_sse2),
+                      make_tuple(3, 3, &vpx_highbd_10_mse8x8_sse2),
+                      make_tuple(4, 4, &vpx_highbd_8_mse16x16_sse2),
+                      make_tuple(4, 3, &vpx_highbd_8_mse16x8_sse2),
+                      make_tuple(3, 4, &vpx_highbd_8_mse8x16_sse2),
+                      make_tuple(3, 3, &vpx_highbd_8_mse8x8_sse2)));
+*/
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VpxHBDVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_highbd_12_variance64x64_sse2, 12),
+                      make_tuple(6, 5, &vpx_highbd_12_variance64x32_sse2, 12),
+                      make_tuple(5, 6, &vpx_highbd_12_variance32x64_sse2, 12),
+                      make_tuple(5, 5, &vpx_highbd_12_variance32x32_sse2, 12),
+                      make_tuple(5, 4, &vpx_highbd_12_variance32x16_sse2, 12),
+                      make_tuple(4, 5, &vpx_highbd_12_variance16x32_sse2, 12),
+                      make_tuple(4, 4, &vpx_highbd_12_variance16x16_sse2, 12),
+                      make_tuple(4, 3, &vpx_highbd_12_variance16x8_sse2, 12),
+                      make_tuple(3, 4, &vpx_highbd_12_variance8x16_sse2, 12),
+                      make_tuple(3, 3, &vpx_highbd_12_variance8x8_sse2, 12),
+                      make_tuple(6, 6, &vpx_highbd_10_variance64x64_sse2, 10),
+                      make_tuple(6, 5, &vpx_highbd_10_variance64x32_sse2, 10),
+                      make_tuple(5, 6, &vpx_highbd_10_variance32x64_sse2, 10),
+                      make_tuple(5, 5, &vpx_highbd_10_variance32x32_sse2, 10),
+                      make_tuple(5, 4, &vpx_highbd_10_variance32x16_sse2, 10),
+                      make_tuple(4, 5, &vpx_highbd_10_variance16x32_sse2, 10),
+                      make_tuple(4, 4, &vpx_highbd_10_variance16x16_sse2, 10),
+                      make_tuple(4, 3, &vpx_highbd_10_variance16x8_sse2, 10),
+                      make_tuple(3, 4, &vpx_highbd_10_variance8x16_sse2, 10),
+                      make_tuple(3, 3, &vpx_highbd_10_variance8x8_sse2, 10),
+                      make_tuple(6, 6, &vpx_highbd_8_variance64x64_sse2, 8),
+                      make_tuple(6, 5, &vpx_highbd_8_variance64x32_sse2, 8),
+                      make_tuple(5, 6, &vpx_highbd_8_variance32x64_sse2, 8),
+                      make_tuple(5, 5, &vpx_highbd_8_variance32x32_sse2, 8),
+                      make_tuple(5, 4, &vpx_highbd_8_variance32x16_sse2, 8),
+                      make_tuple(4, 5, &vpx_highbd_8_variance16x32_sse2, 8),
+                      make_tuple(4, 4, &vpx_highbd_8_variance16x16_sse2, 8),
+                      make_tuple(4, 3, &vpx_highbd_8_variance16x8_sse2, 8),
+                      make_tuple(3, 4, &vpx_highbd_8_variance8x16_sse2, 8),
+                      make_tuple(3, 3, &vpx_highbd_8_variance8x8_sse2, 8)));
+
+#if CONFIG_USE_X86INC
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VpxHBDSubpelVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &vpx_highbd_12_sub_pixel_variance64x64_sse2, 12),
+        make_tuple(6, 5, &vpx_highbd_12_sub_pixel_variance64x32_sse2, 12),
+        make_tuple(5, 6, &vpx_highbd_12_sub_pixel_variance32x64_sse2, 12),
+        make_tuple(5, 5, &vpx_highbd_12_sub_pixel_variance32x32_sse2, 12),
+        make_tuple(5, 4, &vpx_highbd_12_sub_pixel_variance32x16_sse2, 12),
+        make_tuple(4, 5, &vpx_highbd_12_sub_pixel_variance16x32_sse2, 12),
+        make_tuple(4, 4, &vpx_highbd_12_sub_pixel_variance16x16_sse2, 12),
+        make_tuple(4, 3, &vpx_highbd_12_sub_pixel_variance16x8_sse2, 12),
+        make_tuple(3, 4, &vpx_highbd_12_sub_pixel_variance8x16_sse2, 12),
+        make_tuple(3, 3, &vpx_highbd_12_sub_pixel_variance8x8_sse2, 12),
+        make_tuple(3, 2, &vpx_highbd_12_sub_pixel_variance8x4_sse2, 12),
+        make_tuple(6, 6, &vpx_highbd_10_sub_pixel_variance64x64_sse2, 10),
+        make_tuple(6, 5, &vpx_highbd_10_sub_pixel_variance64x32_sse2, 10),
+        make_tuple(5, 6, &vpx_highbd_10_sub_pixel_variance32x64_sse2, 10),
+        make_tuple(5, 5, &vpx_highbd_10_sub_pixel_variance32x32_sse2, 10),
+        make_tuple(5, 4, &vpx_highbd_10_sub_pixel_variance32x16_sse2, 10),
+        make_tuple(4, 5, &vpx_highbd_10_sub_pixel_variance16x32_sse2, 10),
+        make_tuple(4, 4, &vpx_highbd_10_sub_pixel_variance16x16_sse2, 10),
+        make_tuple(4, 3, &vpx_highbd_10_sub_pixel_variance16x8_sse2, 10),
+        make_tuple(3, 4, &vpx_highbd_10_sub_pixel_variance8x16_sse2, 10),
+        make_tuple(3, 3, &vpx_highbd_10_sub_pixel_variance8x8_sse2, 10),
+        make_tuple(3, 2, &vpx_highbd_10_sub_pixel_variance8x4_sse2, 10),
+        make_tuple(6, 6, &vpx_highbd_8_sub_pixel_variance64x64_sse2, 8),
+        make_tuple(6, 5, &vpx_highbd_8_sub_pixel_variance64x32_sse2, 8),
+        make_tuple(5, 6, &vpx_highbd_8_sub_pixel_variance32x64_sse2, 8),
+        make_tuple(5, 5, &vpx_highbd_8_sub_pixel_variance32x32_sse2, 8),
+        make_tuple(5, 4, &vpx_highbd_8_sub_pixel_variance32x16_sse2, 8),
+        make_tuple(4, 5, &vpx_highbd_8_sub_pixel_variance16x32_sse2, 8),
+        make_tuple(4, 4, &vpx_highbd_8_sub_pixel_variance16x16_sse2, 8),
+        make_tuple(4, 3, &vpx_highbd_8_sub_pixel_variance16x8_sse2, 8),
+        make_tuple(3, 4, &vpx_highbd_8_sub_pixel_variance8x16_sse2, 8),
+        make_tuple(3, 3, &vpx_highbd_8_sub_pixel_variance8x8_sse2, 8),
+        make_tuple(3, 2, &vpx_highbd_8_sub_pixel_variance8x4_sse2, 8)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VpxHBDSubpelAvgVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &vpx_highbd_12_sub_pixel_avg_variance64x64_sse2, 12),
+        make_tuple(6, 5, &vpx_highbd_12_sub_pixel_avg_variance64x32_sse2, 12),
+        make_tuple(5, 6, &vpx_highbd_12_sub_pixel_avg_variance32x64_sse2, 12),
+        make_tuple(5, 5, &vpx_highbd_12_sub_pixel_avg_variance32x32_sse2, 12),
+        make_tuple(5, 4, &vpx_highbd_12_sub_pixel_avg_variance32x16_sse2, 12),
+        make_tuple(4, 5, &vpx_highbd_12_sub_pixel_avg_variance16x32_sse2, 12),
+        make_tuple(4, 4, &vpx_highbd_12_sub_pixel_avg_variance16x16_sse2, 12),
+        make_tuple(4, 3, &vpx_highbd_12_sub_pixel_avg_variance16x8_sse2, 12),
+        make_tuple(3, 4, &vpx_highbd_12_sub_pixel_avg_variance8x16_sse2, 12),
+        make_tuple(3, 3, &vpx_highbd_12_sub_pixel_avg_variance8x8_sse2, 12),
+        make_tuple(3, 2, &vpx_highbd_12_sub_pixel_avg_variance8x4_sse2, 12),
+        make_tuple(6, 6, &vpx_highbd_10_sub_pixel_avg_variance64x64_sse2, 10),
+        make_tuple(6, 5, &vpx_highbd_10_sub_pixel_avg_variance64x32_sse2, 10),
+        make_tuple(5, 6, &vpx_highbd_10_sub_pixel_avg_variance32x64_sse2, 10),
+        make_tuple(5, 5, &vpx_highbd_10_sub_pixel_avg_variance32x32_sse2, 10),
+        make_tuple(5, 4, &vpx_highbd_10_sub_pixel_avg_variance32x16_sse2, 10),
+        make_tuple(4, 5, &vpx_highbd_10_sub_pixel_avg_variance16x32_sse2, 10),
+        make_tuple(4, 4, &vpx_highbd_10_sub_pixel_avg_variance16x16_sse2, 10),
+        make_tuple(4, 3, &vpx_highbd_10_sub_pixel_avg_variance16x8_sse2, 10),
+        make_tuple(3, 4, &vpx_highbd_10_sub_pixel_avg_variance8x16_sse2, 10),
+        make_tuple(3, 3, &vpx_highbd_10_sub_pixel_avg_variance8x8_sse2, 10),
+        make_tuple(3, 2, &vpx_highbd_10_sub_pixel_avg_variance8x4_sse2, 10),
+        make_tuple(6, 6, &vpx_highbd_8_sub_pixel_avg_variance64x64_sse2, 8),
+        make_tuple(6, 5, &vpx_highbd_8_sub_pixel_avg_variance64x32_sse2, 8),
+        make_tuple(5, 6, &vpx_highbd_8_sub_pixel_avg_variance32x64_sse2, 8),
+        make_tuple(5, 5, &vpx_highbd_8_sub_pixel_avg_variance32x32_sse2, 8),
+        make_tuple(5, 4, &vpx_highbd_8_sub_pixel_avg_variance32x16_sse2, 8),
+        make_tuple(4, 5, &vpx_highbd_8_sub_pixel_avg_variance16x32_sse2, 8),
+        make_tuple(4, 4, &vpx_highbd_8_sub_pixel_avg_variance16x16_sse2, 8),
+        make_tuple(4, 3, &vpx_highbd_8_sub_pixel_avg_variance16x8_sse2, 8),
+        make_tuple(3, 4, &vpx_highbd_8_sub_pixel_avg_variance8x16_sse2, 8),
+        make_tuple(3, 3, &vpx_highbd_8_sub_pixel_avg_variance8x8_sse2, 8),
+        make_tuple(3, 2, &vpx_highbd_8_sub_pixel_avg_variance8x4_sse2, 8)));
+#endif  // CONFIG_USE_X86INC
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3
+#if CONFIG_USE_X86INC
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, VpxSubpelVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_sub_pixel_variance64x64_ssse3, 0),
+                      make_tuple(6, 5, &vpx_sub_pixel_variance64x32_ssse3, 0),
+                      make_tuple(5, 6, &vpx_sub_pixel_variance32x64_ssse3, 0),
+                      make_tuple(5, 5, &vpx_sub_pixel_variance32x32_ssse3, 0),
+                      make_tuple(5, 4, &vpx_sub_pixel_variance32x16_ssse3, 0),
+                      make_tuple(4, 5, &vpx_sub_pixel_variance16x32_ssse3, 0),
+                      make_tuple(4, 4, &vpx_sub_pixel_variance16x16_ssse3, 0),
+                      make_tuple(4, 3, &vpx_sub_pixel_variance16x8_ssse3, 0),
+                      make_tuple(3, 4, &vpx_sub_pixel_variance8x16_ssse3, 0),
+                      make_tuple(3, 3, &vpx_sub_pixel_variance8x8_ssse3, 0),
+                      make_tuple(3, 2, &vpx_sub_pixel_variance8x4_ssse3, 0),
+                      make_tuple(2, 3, &vpx_sub_pixel_variance4x8_ssse3, 0),
+                      make_tuple(2, 2, &vpx_sub_pixel_variance4x4_ssse3, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, VpxSubpelAvgVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &vpx_sub_pixel_avg_variance64x64_ssse3, 0),
+        make_tuple(6, 5, &vpx_sub_pixel_avg_variance64x32_ssse3, 0),
+        make_tuple(5, 6, &vpx_sub_pixel_avg_variance32x64_ssse3, 0),
+        make_tuple(5, 5, &vpx_sub_pixel_avg_variance32x32_ssse3, 0),
+        make_tuple(5, 4, &vpx_sub_pixel_avg_variance32x16_ssse3, 0),
+        make_tuple(4, 5, &vpx_sub_pixel_avg_variance16x32_ssse3, 0),
+        make_tuple(4, 4, &vpx_sub_pixel_avg_variance16x16_ssse3, 0),
+        make_tuple(4, 3, &vpx_sub_pixel_avg_variance16x8_ssse3, 0),
+        make_tuple(3, 4, &vpx_sub_pixel_avg_variance8x16_ssse3, 0),
+        make_tuple(3, 3, &vpx_sub_pixel_avg_variance8x8_ssse3, 0),
+        make_tuple(3, 2, &vpx_sub_pixel_avg_variance8x4_ssse3, 0),
+        make_tuple(2, 3, &vpx_sub_pixel_avg_variance4x8_ssse3, 0),
+        make_tuple(2, 2, &vpx_sub_pixel_avg_variance4x4_ssse3, 0)));
+#endif  // CONFIG_USE_X86INC
+#endif  // HAVE_SSSE3
+
+#if HAVE_AVX2
+INSTANTIATE_TEST_CASE_P(AVX2, VpxMseTest,
+                        ::testing::Values(make_tuple(4, 4,
+                                                     &vpx_mse16x16_avx2)));
+
+INSTANTIATE_TEST_CASE_P(
+    AVX2, VpxVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_variance64x64_avx2, 0),
+                      make_tuple(6, 5, &vpx_variance64x32_avx2, 0),
+                      make_tuple(5, 5, &vpx_variance32x32_avx2, 0),
+                      make_tuple(5, 4, &vpx_variance32x16_avx2, 0),
+                      make_tuple(4, 4, &vpx_variance16x16_avx2, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    AVX2, VpxSubpelVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_sub_pixel_variance64x64_avx2, 0),
+                      make_tuple(5, 5, &vpx_sub_pixel_variance32x32_avx2, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    AVX2, VpxSubpelAvgVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &vpx_sub_pixel_avg_variance64x64_avx2, 0),
+        make_tuple(5, 5, &vpx_sub_pixel_avg_variance32x32_avx2, 0)));
+#endif  // HAVE_AVX2
+
+#if HAVE_MEDIA
+INSTANTIATE_TEST_CASE_P(MEDIA, VpxMseTest,
+                        ::testing::Values(make_tuple(4, 4,
+                                                     &vpx_mse16x16_media)));
+
+INSTANTIATE_TEST_CASE_P(
+    MEDIA, VpxVarianceTest,
+    ::testing::Values(make_tuple(4, 4, &vpx_variance16x16_media, 0),
+                      make_tuple(3, 3, &vpx_variance8x8_media, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    MEDIA, VpxSubpelVarianceTest,
+    ::testing::Values(make_tuple(4, 4, &vpx_sub_pixel_variance16x16_media, 0),
+                      make_tuple(3, 3, &vpx_sub_pixel_variance8x8_media, 0)));
+#endif  // HAVE_MEDIA
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, VpxSseTest,
+                        ::testing::Values(make_tuple(2, 2,
+                                                     &vpx_get4x4sse_cs_neon)));
+
+INSTANTIATE_TEST_CASE_P(NEON, VpxMseTest,
+                        ::testing::Values(make_tuple(4, 4,
+                                                     &vpx_mse16x16_neon)));
+
+INSTANTIATE_TEST_CASE_P(
+    NEON, VpxVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_variance64x64_neon, 0),
+                      make_tuple(6, 5, &vpx_variance64x32_neon, 0),
+                      make_tuple(5, 6, &vpx_variance32x64_neon, 0),
+                      make_tuple(5, 5, &vpx_variance32x32_neon, 0),
+                      make_tuple(4, 4, &vpx_variance16x16_neon, 0),
+                      make_tuple(4, 3, &vpx_variance16x8_neon, 0),
+                      make_tuple(3, 4, &vpx_variance8x16_neon, 0),
+                      make_tuple(3, 3, &vpx_variance8x8_neon, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    NEON, VpxSubpelVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_sub_pixel_variance64x64_neon, 0),
+                      make_tuple(5, 5, &vpx_sub_pixel_variance32x32_neon, 0),
+                      make_tuple(4, 4, &vpx_sub_pixel_variance16x16_neon, 0),
+                      make_tuple(3, 3, &vpx_sub_pixel_variance8x8_neon, 0)));
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(MSA, SumOfSquaresTest,
+                        ::testing::Values(vpx_get_mb_ss_msa));
+
+INSTANTIATE_TEST_CASE_P(MSA, VpxSseTest,
+                        ::testing::Values(make_tuple(2, 2,
+                                                     &vpx_get4x4sse_cs_msa)));
+
+INSTANTIATE_TEST_CASE_P(MSA, VpxMseTest,
+                        ::testing::Values(make_tuple(4, 4, &vpx_mse16x16_msa),
+                                          make_tuple(4, 3, &vpx_mse16x8_msa),
+                                          make_tuple(3, 4, &vpx_mse8x16_msa),
+                                          make_tuple(3, 3, &vpx_mse8x8_msa)));
+
+INSTANTIATE_TEST_CASE_P(
+    MSA, VpxVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_variance64x64_msa, 0),
+                      make_tuple(6, 5, &vpx_variance64x32_msa, 0),
+                      make_tuple(5, 6, &vpx_variance32x64_msa, 0),
+                      make_tuple(5, 5, &vpx_variance32x32_msa, 0),
+                      make_tuple(5, 4, &vpx_variance32x16_msa, 0),
+                      make_tuple(4, 5, &vpx_variance16x32_msa, 0),
+                      make_tuple(4, 4, &vpx_variance16x16_msa, 0),
+                      make_tuple(4, 3, &vpx_variance16x8_msa, 0),
+                      make_tuple(3, 4, &vpx_variance8x16_msa, 0),
+                      make_tuple(3, 3, &vpx_variance8x8_msa, 0),
+                      make_tuple(3, 2, &vpx_variance8x4_msa, 0),
+                      make_tuple(2, 3, &vpx_variance4x8_msa, 0),
+                      make_tuple(2, 2, &vpx_variance4x4_msa, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    MSA, VpxSubpelVarianceTest,
+    ::testing::Values(make_tuple(2, 2, &vpx_sub_pixel_variance4x4_msa, 0),
+                      make_tuple(2, 3, &vpx_sub_pixel_variance4x8_msa, 0),
+                      make_tuple(3, 2, &vpx_sub_pixel_variance8x4_msa, 0),
+                      make_tuple(3, 3, &vpx_sub_pixel_variance8x8_msa, 0),
+                      make_tuple(3, 4, &vpx_sub_pixel_variance8x16_msa, 0),
+                      make_tuple(4, 3, &vpx_sub_pixel_variance16x8_msa, 0),
+                      make_tuple(4, 4, &vpx_sub_pixel_variance16x16_msa, 0),
+                      make_tuple(4, 5, &vpx_sub_pixel_variance16x32_msa, 0),
+                      make_tuple(5, 4, &vpx_sub_pixel_variance32x16_msa, 0),
+                      make_tuple(5, 5, &vpx_sub_pixel_variance32x32_msa, 0),
+                      make_tuple(5, 6, &vpx_sub_pixel_variance32x64_msa, 0),
+                      make_tuple(6, 5, &vpx_sub_pixel_variance64x32_msa, 0),
+                      make_tuple(6, 6, &vpx_sub_pixel_variance64x64_msa, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    MSA, VpxSubpelAvgVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &vpx_sub_pixel_avg_variance64x64_msa, 0),
+                      make_tuple(6, 5, &vpx_sub_pixel_avg_variance64x32_msa, 0),
+                      make_tuple(5, 6, &vpx_sub_pixel_avg_variance32x64_msa, 0),
+                      make_tuple(5, 5, &vpx_sub_pixel_avg_variance32x32_msa, 0),
+                      make_tuple(5, 4, &vpx_sub_pixel_avg_variance32x16_msa, 0),
+                      make_tuple(4, 5, &vpx_sub_pixel_avg_variance16x32_msa, 0),
+                      make_tuple(4, 4, &vpx_sub_pixel_avg_variance16x16_msa, 0),
+                      make_tuple(4, 3, &vpx_sub_pixel_avg_variance16x8_msa, 0),
+                      make_tuple(3, 4, &vpx_sub_pixel_avg_variance8x16_msa, 0),
+                      make_tuple(3, 3, &vpx_sub_pixel_avg_variance8x8_msa, 0),
+                      make_tuple(3, 2, &vpx_sub_pixel_avg_variance8x4_msa, 0),
+                      make_tuple(2, 3, &vpx_sub_pixel_avg_variance4x8_msa, 0),
+                      make_tuple(2, 2, &vpx_sub_pixel_avg_variance4x4_msa, 0)));
+#endif  // HAVE_MSA
+}  // namespace
diff --git a/libvpx/libvpx/test/video_source.h b/libvpx/libvpx/test/video_source.h
new file mode 100644
index 0000000..ade323e
--- /dev/null
+++ b/libvpx/libvpx/test/video_source.h
@@ -0,0 +1,270 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef TEST_VIDEO_SOURCE_H_
+#define TEST_VIDEO_SOURCE_H_
+
+#if defined(_WIN32)
+#undef NOMINMAX
+#define NOMINMAX
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#endif
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include "test/acm_random.h"
+#include "vpx/vpx_encoder.h"
+
+namespace libvpx_test {
+
+// Helper macros to ensure LIBVPX_TEST_DATA_PATH is a quoted string.
+// These are undefined right below GetDataPath
+// NOTE: LIBVPX_TEST_DATA_PATH MUST NOT be a quoted string before
+// Stringification or the GetDataPath will fail at runtime
+#define TO_STRING(S) #S
+#define STRINGIFY(S) TO_STRING(S)
+
+// A simple function to encapsulate cross platform retrieval of test data path
+static std::string GetDataPath() {
+  const char *const data_path = getenv("LIBVPX_TEST_DATA_PATH");
+  if (data_path == NULL) {
+#ifdef LIBVPX_TEST_DATA_PATH
+    // In some environments, we cannot set environment variables
+    // Instead, we set the data path by using a preprocessor symbol
+    // which can be set from make files
+    return STRINGIFY(LIBVPX_TEST_DATA_PATH);
+#else
+    return ".";
+#endif
+  }
+  return data_path;
+}
+
+// Undefining stringification macros because they are not used elsewhere
+#undef TO_STRING
+#undef STRINGIFY
+
+inline FILE *OpenTestDataFile(const std::string& file_name) {
+  const std::string path_to_source = GetDataPath() + "/" + file_name;
+  return fopen(path_to_source.c_str(), "rb");
+}
+
+static FILE *GetTempOutFile(std::string *file_name) {
+  file_name->clear();
+#if defined(_WIN32)
+  char fname[MAX_PATH];
+  char tmppath[MAX_PATH];
+  if (GetTempPathA(MAX_PATH, tmppath)) {
+    // Assume for now that the filename generated is unique per process
+    if (GetTempFileNameA(tmppath, "lvx", 0, fname)) {
+      file_name->assign(fname);
+      return fopen(fname, "wb+");
+    }
+  }
+  return NULL;
+#else
+  return tmpfile();
+#endif
+}
+
+class TempOutFile {
+ public:
+  TempOutFile() {
+    file_ = GetTempOutFile(&file_name_);
+  }
+  ~TempOutFile() {
+    CloseFile();
+    if (!file_name_.empty()) {
+      EXPECT_EQ(0, remove(file_name_.c_str()));
+    }
+  }
+  FILE *file() {
+    return file_;
+  }
+  const std::string& file_name() {
+    return file_name_;
+  }
+
+ protected:
+  void CloseFile() {
+    if (file_) {
+      fclose(file_);
+      file_ = NULL;
+    }
+  }
+  FILE *file_;
+  std::string file_name_;
+};
+
+// Abstract base class for test video sources, which provide a stream of
+// vpx_image_t images with associated timestamps and duration.
+class VideoSource {
+ public:
+  virtual ~VideoSource() {}
+
+  // Prepare the stream for reading, rewind/open as necessary.
+  virtual void Begin() = 0;
+
+  // Advance the cursor to the next frame
+  virtual void Next() = 0;
+
+  // Get the current video frame, or NULL on End-Of-Stream.
+  virtual vpx_image_t *img() const = 0;
+
+  // Get the presentation timestamp of the current frame.
+  virtual vpx_codec_pts_t pts() const = 0;
+
+  // Get the current frame's duration
+  virtual unsigned long duration() const = 0;
+
+  // Get the timebase for the stream
+  virtual vpx_rational_t timebase() const = 0;
+
+  // Get the current frame counter, starting at 0.
+  virtual unsigned int frame() const = 0;
+
+  // Get the current file limit.
+  virtual unsigned int limit() const = 0;
+};
+
+
+class DummyVideoSource : public VideoSource {
+ public:
+  DummyVideoSource()
+      : img_(NULL),
+        limit_(100),
+        width_(80),
+        height_(64),
+        format_(VPX_IMG_FMT_I420) {
+    ReallocImage();
+  }
+
+  virtual ~DummyVideoSource() { vpx_img_free(img_); }
+
+  virtual void Begin() {
+    frame_ = 0;
+    FillFrame();
+  }
+
+  virtual void Next() {
+    ++frame_;
+    FillFrame();
+  }
+
+  virtual vpx_image_t *img() const {
+    return (frame_ < limit_) ? img_ : NULL;
+  }
+
+  // Models a stream where Timebase = 1/FPS, so pts == frame.
+  virtual vpx_codec_pts_t pts() const { return frame_; }
+
+  virtual unsigned long duration() const { return 1; }
+
+  virtual vpx_rational_t timebase() const {
+    const vpx_rational_t t = {1, 30};
+    return t;
+  }
+
+  virtual unsigned int frame() const { return frame_; }
+
+  virtual unsigned int limit() const { return limit_; }
+
+  void set_limit(unsigned int limit) {
+    limit_ = limit;
+  }
+
+  void SetSize(unsigned int width, unsigned int height) {
+    if (width != width_ || height != height_) {
+      width_ = width;
+      height_ = height;
+      ReallocImage();
+    }
+  }
+
+  void SetImageFormat(vpx_img_fmt_t format) {
+    if (format_ != format) {
+      format_ = format;
+      ReallocImage();
+    }
+  }
+
+ protected:
+  virtual void FillFrame() { if (img_) memset(img_->img_data, 0, raw_sz_); }
+
+  void ReallocImage() {
+    vpx_img_free(img_);
+    img_ = vpx_img_alloc(NULL, format_, width_, height_, 32);
+    raw_sz_ = ((img_->w + 31) & ~31) * img_->h * img_->bps / 8;
+  }
+
+  vpx_image_t *img_;
+  size_t       raw_sz_;
+  unsigned int limit_;
+  unsigned int frame_;
+  unsigned int width_;
+  unsigned int height_;
+  vpx_img_fmt_t format_;
+};
+
+
+class RandomVideoSource : public DummyVideoSource {
+ public:
+  RandomVideoSource(int seed = ACMRandom::DeterministicSeed())
+      : rnd_(seed),
+        seed_(seed) { }
+
+ protected:
+  // Reset the RNG to get a matching stream for the second pass
+  virtual void Begin() {
+    frame_ = 0;
+    rnd_.Reset(seed_);
+    FillFrame();
+  }
+
+  // 15 frames of noise, followed by 15 static frames. Reset to 0 rather
+  // than holding previous frames to encourage keyframes to be thrown.
+  virtual void FillFrame() {
+    if (img_) {
+      if (frame_ % 30 < 15)
+        for (size_t i = 0; i < raw_sz_; ++i)
+          img_->img_data[i] = rnd_.Rand8();
+      else
+        memset(img_->img_data, 0, raw_sz_);
+    }
+  }
+
+  ACMRandom rnd_;
+  int seed_;
+};
+
+// Abstract base class for test video sources, which provide a stream of
+// decompressed images to the decoder.
+class CompressedVideoSource {
+ public:
+  virtual ~CompressedVideoSource() {}
+
+  virtual void Init() = 0;
+
+  // Prepare the stream for reading, rewind/open as necessary.
+  virtual void Begin() = 0;
+
+  // Advance the cursor to the next frame
+  virtual void Next() = 0;
+
+  virtual const uint8_t *cxdata() const = 0;
+
+  virtual size_t frame_size() const = 0;
+
+  virtual unsigned int frame_number() const = 0;
+};
+
+}  // namespace libvpx_test
+
+#endif  // TEST_VIDEO_SOURCE_H_
diff --git a/libvpx/libvpx/test/vp8_boolcoder_test.cc b/libvpx/libvpx/test/vp8_boolcoder_test.cc
new file mode 100644
index 0000000..02d7162
--- /dev/null
+++ b/libvpx/libvpx/test/vp8_boolcoder_test.cc
@@ -0,0 +1,116 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <math.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/types.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "test/acm_random.h"
+#include "vp8/decoder/dboolhuff.h"
+#include "vp8/encoder/boolhuff.h"
+#include "vpx/vpx_integer.h"
+
+namespace {
+const int num_tests = 10;
+
+// In a real use the 'decrypt_state' parameter will be a pointer to a struct
+// with whatever internal state the decryptor uses. For testing we'll just
+// xor with a constant key, and decrypt_state will point to the start of
+// the original buffer.
+const uint8_t secret_key[16] = {
+  0x01, 0x12, 0x23, 0x34, 0x45, 0x56, 0x67, 0x78,
+  0x89, 0x9a, 0xab, 0xbc, 0xcd, 0xde, 0xef, 0xf0
+};
+
+void encrypt_buffer(uint8_t *buffer, size_t size) {
+  for (size_t i = 0; i < size; ++i) {
+    buffer[i] ^= secret_key[i & 15];
+  }
+}
+
+void test_decrypt_cb(void *decrypt_state, const uint8_t *input,
+                     uint8_t *output, int count) {
+  const size_t offset = input - reinterpret_cast<uint8_t*>(decrypt_state);
+  for (int i = 0; i < count; i++) {
+    output[i] = input[i] ^ secret_key[(offset + i) & 15];
+  }
+}
+
+}  // namespace
+
+using libvpx_test::ACMRandom;
+
+TEST(VP8, TestBitIO) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  for (int n = 0; n < num_tests; ++n) {
+    for (int method = 0; method <= 7; ++method) {   // we generate various proba
+      const int kBitsToTest = 1000;
+      uint8_t probas[kBitsToTest];
+
+      for (int i = 0; i < kBitsToTest; ++i) {
+        const int parity = i & 1;
+        probas[i] =
+            (method == 0) ? 0 : (method == 1) ? 255 :
+            (method == 2) ? 128 :
+            (method == 3) ? rnd.Rand8() :
+            (method == 4) ? (parity ? 0 : 255) :
+            // alternate between low and high proba:
+            (method == 5) ? (parity ? rnd(128) : 255 - rnd(128)) :
+            (method == 6) ?
+                (parity ? rnd(64) : 255 - rnd(64)) :
+                (parity ? rnd(32) : 255 - rnd(32));
+      }
+      for (int bit_method = 0; bit_method <= 3; ++bit_method) {
+        const int random_seed = 6432;
+        const int kBufferSize = 10000;
+        ACMRandom bit_rnd(random_seed);
+        BOOL_CODER bw;
+        uint8_t bw_buffer[kBufferSize];
+        vp8_start_encode(&bw, bw_buffer, bw_buffer + kBufferSize);
+
+        int bit = (bit_method == 0) ? 0 : (bit_method == 1) ? 1 : 0;
+        for (int i = 0; i < kBitsToTest; ++i) {
+          if (bit_method == 2) {
+            bit = (i & 1);
+          } else if (bit_method == 3) {
+            bit = bit_rnd(2);
+          }
+          vp8_encode_bool(&bw, bit, static_cast<int>(probas[i]));
+        }
+
+        vp8_stop_encode(&bw);
+
+        BOOL_DECODER br;
+        encrypt_buffer(bw_buffer, kBufferSize);
+        vp8dx_start_decode(&br, bw_buffer, kBufferSize,
+                           test_decrypt_cb,
+                           reinterpret_cast<void *>(bw_buffer));
+        bit_rnd.Reset(random_seed);
+        for (int i = 0; i < kBitsToTest; ++i) {
+          if (bit_method == 2) {
+            bit = (i & 1);
+          } else if (bit_method == 3) {
+            bit = bit_rnd(2);
+          }
+          GTEST_ASSERT_EQ(vp8dx_decode_bool(&br, probas[i]), bit)
+              << "pos: "<< i << " / " << kBitsToTest
+              << " bit_method: " << bit_method
+              << " method: " << method;
+        }
+      }
+    }
+  }
+}
diff --git a/libvpx/libvpx/test/vp8_decrypt_test.cc b/libvpx/libvpx/test/vp8_decrypt_test.cc
new file mode 100644
index 0000000..972a1d9
--- /dev/null
+++ b/libvpx/libvpx/test/vp8_decrypt_test.cc
@@ -0,0 +1,71 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/ivf_video_source.h"
+
+namespace {
+// In a real use the 'decrypt_state' parameter will be a pointer to a struct
+// with whatever internal state the decryptor uses. For testing we'll just
+// xor with a constant key, and decrypt_state will point to the start of
+// the original buffer.
+const uint8_t test_key[16] = {
+  0x01, 0x12, 0x23, 0x34, 0x45, 0x56, 0x67, 0x78,
+  0x89, 0x9a, 0xab, 0xbc, 0xcd, 0xde, 0xef, 0xf0
+};
+
+void encrypt_buffer(const uint8_t *src, uint8_t *dst, size_t size,
+                    ptrdiff_t offset) {
+  for (size_t i = 0; i < size; ++i) {
+    dst[i] = src[i] ^ test_key[(offset + i) & 15];
+  }
+}
+
+void test_decrypt_cb(void *decrypt_state, const uint8_t *input,
+                     uint8_t *output, int count) {
+  encrypt_buffer(input, output, count,
+                 input - reinterpret_cast<uint8_t *>(decrypt_state));
+}
+
+}  // namespace
+
+namespace libvpx_test {
+
+TEST(TestDecrypt, DecryptWorksVp8) {
+  libvpx_test::IVFVideoSource video("vp80-00-comprehensive-001.ivf");
+  video.Init();
+
+  vpx_codec_dec_cfg_t dec_cfg = vpx_codec_dec_cfg_t();
+  VP8Decoder decoder(dec_cfg, 0);
+
+  video.Begin();
+
+  // no decryption
+  vpx_codec_err_t res = decoder.DecodeFrame(video.cxdata(), video.frame_size());
+  ASSERT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+
+  // decrypt frame
+  video.Next();
+
+  std::vector<uint8_t> encrypted(video.frame_size());
+  encrypt_buffer(video.cxdata(), &encrypted[0], video.frame_size(), 0);
+  vpx_decrypt_init di = { test_decrypt_cb, &encrypted[0] };
+  decoder.Control(VPXD_SET_DECRYPTOR, &di);
+
+  res = decoder.DecodeFrame(&encrypted[0], encrypted.size());
+  ASSERT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+}
+
+}  // namespace libvpx_test
diff --git a/libvpx/libvpx/test/vp8_denoiser_sse2_test.cc b/libvpx/libvpx/test/vp8_denoiser_sse2_test.cc
new file mode 100644
index 0000000..e8ca8d3
--- /dev/null
+++ b/libvpx/libvpx/test/vp8_denoiser_sse2_test.cc
@@ -0,0 +1,116 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+#include "vp8/encoder/denoising.h"
+#include "vp8/common/reconinter.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+
+const int kNumPixels = 16 * 16;
+class VP8DenoiserTest : public ::testing::TestWithParam<int> {
+ public:
+  virtual ~VP8DenoiserTest() {}
+
+  virtual void SetUp() {
+    increase_denoising_ = GetParam();
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  int increase_denoising_;
+};
+
+TEST_P(VP8DenoiserTest, BitexactCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 4000;
+  const int stride = 16;
+
+  // Allocate the space for input and output,
+  // where sig_block_c/_sse2 is the block to be denoised,
+  // mc_avg_block is the denoised reference block,
+  // avg_block_c is the denoised result from C code,
+  // avg_block_sse2 is the denoised result from SSE2 code.
+  DECLARE_ALIGNED(16, uint8_t, sig_block_c[kNumPixels]);
+  // Since in VP8 denoiser, the source signal will be changed,
+  // we need another copy of the source signal as the input of sse2 code.
+  DECLARE_ALIGNED(16, uint8_t, sig_block_sse2[kNumPixels]);
+  DECLARE_ALIGNED(16, uint8_t, mc_avg_block[kNumPixels]);
+  DECLARE_ALIGNED(16, uint8_t, avg_block_c[kNumPixels]);
+  DECLARE_ALIGNED(16, uint8_t, avg_block_sse2[kNumPixels]);
+
+  for (int i = 0; i < count_test_block; ++i) {
+    // Generate random motion magnitude, 20% of which exceed the threshold.
+    const int motion_magnitude_ran =
+        rnd.Rand8() % static_cast<int>(MOTION_MAGNITUDE_THRESHOLD * 1.2);
+
+    // Initialize a test block with random number in range [0, 255].
+    for (int j = 0; j < kNumPixels; ++j) {
+      int temp = 0;
+      sig_block_sse2[j] = sig_block_c[j] = rnd.Rand8();
+      // The pixels in mc_avg_block are generated by adding a random
+      // number in range [-19, 19] to corresponding pixels in sig_block.
+      temp = sig_block_c[j] + (rnd.Rand8() % 2 == 0 ? -1 : 1) *
+             (rnd.Rand8() % 20);
+      // Clip.
+      mc_avg_block[j] = (temp < 0) ? 0 : ((temp > 255) ? 255 : temp);
+    }
+
+    // Test denosiser on Y component.
+    ASM_REGISTER_STATE_CHECK(vp8_denoiser_filter_c(
+        mc_avg_block, stride, avg_block_c, stride, sig_block_c, stride,
+        motion_magnitude_ran, increase_denoising_));
+
+    ASM_REGISTER_STATE_CHECK(vp8_denoiser_filter_sse2(
+        mc_avg_block, stride, avg_block_sse2, stride, sig_block_sse2, stride,
+        motion_magnitude_ran, increase_denoising_));
+
+    // Check bitexactness.
+    for (int h = 0; h < 16; ++h) {
+      for (int w = 0; w < 16; ++w) {
+        EXPECT_EQ(avg_block_c[h * stride + w], avg_block_sse2[h * stride + w]);
+      }
+    }
+
+    // Test denoiser on UV component.
+    ASM_REGISTER_STATE_CHECK(vp8_denoiser_filter_uv_c(
+        mc_avg_block, stride, avg_block_c, stride, sig_block_c, stride,
+        motion_magnitude_ran, increase_denoising_));
+
+    ASM_REGISTER_STATE_CHECK(vp8_denoiser_filter_uv_sse2(
+        mc_avg_block, stride, avg_block_sse2, stride, sig_block_sse2, stride,
+        motion_magnitude_ran, increase_denoising_));
+
+    // Check bitexactness.
+    for (int h = 0; h < 16; ++h) {
+      for (int w = 0; w < 16; ++w) {
+        EXPECT_EQ(avg_block_c[h * stride + w], avg_block_sse2[h * stride + w]);
+      }
+    }
+  }
+}
+
+// Test for all block size.
+INSTANTIATE_TEST_CASE_P(SSE2, VP8DenoiserTest, ::testing::Values(0, 1));
+}  // namespace
diff --git a/libvpx/libvpx/test/vp8_fdct4x4_test.cc b/libvpx/libvpx/test/vp8_fdct4x4_test.cc
new file mode 100644
index 0000000..11a653d
--- /dev/null
+++ b/libvpx/libvpx/test/vp8_fdct4x4_test.cc
@@ -0,0 +1,164 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/types.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vp8_rtcd.h"
+#include "test/acm_random.h"
+#include "vpx/vpx_integer.h"
+
+namespace {
+
+const int cospi8sqrt2minus1 = 20091;
+const int sinpi8sqrt2 = 35468;
+
+void reference_idct4x4(const int16_t *input, int16_t *output) {
+  const int16_t *ip = input;
+  int16_t *op = output;
+
+  for (int i = 0; i < 4; ++i) {
+    const int a1 = ip[0] + ip[8];
+    const int b1 = ip[0] - ip[8];
+    const int temp1 = (ip[4] * sinpi8sqrt2) >> 16;
+    const int temp2 = ip[12] + ((ip[12] * cospi8sqrt2minus1) >> 16);
+    const int c1 = temp1 - temp2;
+    const int temp3 = ip[4] + ((ip[4] * cospi8sqrt2minus1) >> 16);
+    const int temp4 = (ip[12] * sinpi8sqrt2) >> 16;
+    const int d1 = temp3 + temp4;
+    op[0] = a1 + d1;
+    op[12] = a1 - d1;
+    op[4] = b1 + c1;
+    op[8] = b1 - c1;
+    ++ip;
+    ++op;
+  }
+  ip = output;
+  op = output;
+  for (int i = 0; i < 4; ++i) {
+    const int a1 = ip[0] + ip[2];
+    const int b1 = ip[0] - ip[2];
+    const int temp1 = (ip[1] * sinpi8sqrt2) >> 16;
+    const int temp2 = ip[3] + ((ip[3] * cospi8sqrt2minus1) >> 16);
+    const int c1 = temp1 - temp2;
+    const int temp3 = ip[1] + ((ip[1] * cospi8sqrt2minus1) >> 16);
+    const int temp4 = (ip[3] * sinpi8sqrt2) >> 16;
+    const int d1 = temp3 + temp4;
+    op[0] = (a1 + d1 + 4) >> 3;
+    op[3] = (a1 - d1 + 4) >> 3;
+    op[1] = (b1 + c1 + 4) >> 3;
+    op[2] = (b1 - c1 + 4) >> 3;
+    ip += 4;
+    op += 4;
+  }
+}
+
+using libvpx_test::ACMRandom;
+
+TEST(VP8FdctTest, SignBiasCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int16_t test_input_block[16];
+  int16_t test_output_block[16];
+  const int pitch = 8;
+  int count_sign_block[16][2];
+  const int count_test_block = 1000000;
+
+  memset(count_sign_block, 0, sizeof(count_sign_block));
+
+  for (int i = 0; i < count_test_block; ++i) {
+    // Initialize a test block with input range [-255, 255].
+    for (int j = 0; j < 16; ++j)
+      test_input_block[j] = rnd.Rand8() - rnd.Rand8();
+
+    vp8_short_fdct4x4_c(test_input_block, test_output_block, pitch);
+
+    for (int j = 0; j < 16; ++j) {
+      if (test_output_block[j] < 0)
+        ++count_sign_block[j][0];
+      else if (test_output_block[j] > 0)
+        ++count_sign_block[j][1];
+    }
+  }
+
+  bool bias_acceptable = true;
+  for (int j = 0; j < 16; ++j)
+    bias_acceptable = bias_acceptable &&
+    (abs(count_sign_block[j][0] - count_sign_block[j][1]) < 10000);
+
+  EXPECT_EQ(true, bias_acceptable)
+    << "Error: 4x4 FDCT has a sign bias > 1% for input range [-255, 255]";
+
+  memset(count_sign_block, 0, sizeof(count_sign_block));
+
+  for (int i = 0; i < count_test_block; ++i) {
+    // Initialize a test block with input range [-15, 15].
+    for (int j = 0; j < 16; ++j)
+      test_input_block[j] = (rnd.Rand8() >> 4) - (rnd.Rand8() >> 4);
+
+    vp8_short_fdct4x4_c(test_input_block, test_output_block, pitch);
+
+    for (int j = 0; j < 16; ++j) {
+      if (test_output_block[j] < 0)
+        ++count_sign_block[j][0];
+      else if (test_output_block[j] > 0)
+        ++count_sign_block[j][1];
+    }
+  }
+
+  bias_acceptable = true;
+  for (int j = 0; j < 16; ++j)
+    bias_acceptable = bias_acceptable &&
+    (abs(count_sign_block[j][0] - count_sign_block[j][1]) < 100000);
+
+  EXPECT_EQ(true, bias_acceptable)
+    << "Error: 4x4 FDCT has a sign bias > 10% for input range [-15, 15]";
+};
+
+TEST(VP8FdctTest, RoundTripErrorCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int max_error = 0;
+  double total_error = 0;
+  const int count_test_block = 1000000;
+  for (int i = 0; i < count_test_block; ++i) {
+    int16_t test_input_block[16];
+    int16_t test_temp_block[16];
+    int16_t test_output_block[16];
+
+    // Initialize a test block with input range [-255, 255].
+    for (int j = 0; j < 16; ++j)
+      test_input_block[j] = rnd.Rand8() - rnd.Rand8();
+
+    const int pitch = 8;
+    vp8_short_fdct4x4_c(test_input_block, test_temp_block, pitch);
+    reference_idct4x4(test_temp_block, test_output_block);
+
+    for (int j = 0; j < 16; ++j) {
+      const int diff = test_input_block[j] - test_output_block[j];
+      const int error = diff * diff;
+      if (max_error < error)
+        max_error = error;
+      total_error += error;
+    }
+  }
+
+  EXPECT_GE(1, max_error )
+    << "Error: FDCT/IDCT has an individual roundtrip error > 1";
+
+  EXPECT_GE(count_test_block, total_error)
+    << "Error: FDCT/IDCT has average roundtrip error > 1 per block";
+};
+
+}  // namespace
diff --git a/libvpx/libvpx/test/vp8_fragments_test.cc b/libvpx/libvpx/test/vp8_fragments_test.cc
new file mode 100644
index 0000000..cb0d1a1
--- /dev/null
+++ b/libvpx/libvpx/test/vp8_fragments_test.cc
@@ -0,0 +1,37 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/video_source.h"
+
+namespace {
+
+class VP8FramgmentsTest
+    : public ::libvpx_test::EncoderTest,
+      public ::testing::Test {
+ protected:
+  VP8FramgmentsTest() : EncoderTest(&::libvpx_test::kVP8) {}
+  virtual ~VP8FramgmentsTest() {}
+
+  virtual void SetUp() {
+    const unsigned long init_flags =  // NOLINT(runtime/int)
+        VPX_CODEC_USE_OUTPUT_PARTITION;
+    InitializeConfig();
+    SetMode(::libvpx_test::kRealTime);
+    set_init_flags(init_flags);
+  }
+};
+
+TEST_F(VP8FramgmentsTest, TestFragmentsEncodeDecode) {
+  ::libvpx_test::RandomVideoSource video;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+}  // namespace
diff --git a/libvpx/libvpx/test/vp8_multi_resolution_encoder.sh b/libvpx/libvpx/test/vp8_multi_resolution_encoder.sh
new file mode 100755
index 0000000..a8b7fe7
--- /dev/null
+++ b/libvpx/libvpx/test/vp8_multi_resolution_encoder.sh
@@ -0,0 +1,75 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx vp8_multi_resolution_encoder example. To add new
+##  tests to this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vp8_mre_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+vp8_multi_resolution_encoder_verify_environment() {
+  if [ "$(vpx_config_option_enabled CONFIG_MULTI_RES_ENCODING)" = "yes" ]; then
+    if [ ! -e "${YUV_RAW_INPUT}" ]; then
+      elog "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+      return 1
+    fi
+    local readonly app="vp8_multi_resolution_encoder"
+    if [ -z "$(vpx_tool_path "${app}")" ]; then
+      elog "${app} not found. It must exist in LIBVPX_BIN_PATH or its parent."
+      return 1
+    fi
+  fi
+}
+
+# Runs vp8_multi_resolution_encoder. Simply forwards all arguments to
+# vp8_multi_resolution_encoder after building path to the executable.
+vp8_mre() {
+  local readonly encoder="$(vpx_tool_path vp8_multi_resolution_encoder)"
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "$@" ${devnull}
+}
+
+vp8_multi_resolution_encoder_three_formats() {
+  local readonly output_files="${VPX_TEST_OUTPUT_DIR}/vp8_mre_0.ivf
+                               ${VPX_TEST_OUTPUT_DIR}/vp8_mre_1.ivf
+                               ${VPX_TEST_OUTPUT_DIR}/vp8_mre_2.ivf"
+
+  if [ "$(vpx_config_option_enabled CONFIG_MULTI_RES_ENCODING)" = "yes" ]; then
+    if [ "$(vp8_encode_available)" = "yes" ]; then
+      # Param order:
+      #  Input width
+      #  Input height
+      #  Input file path
+      #  Output file names
+      #  Output PSNR
+      vp8_mre "${YUV_RAW_INPUT_WIDTH}" \
+        "${YUV_RAW_INPUT_HEIGHT}" \
+        "${YUV_RAW_INPUT}" \
+        ${output_files} \
+        0
+
+      for output_file in ${output_files}; do
+        if [ ! -e "${output_file}" ]; then
+          elog "Missing output file: ${output_file}"
+          return 1
+        fi
+      done
+    fi
+  fi
+}
+
+vp8_mre_tests="vp8_multi_resolution_encoder_three_formats"
+run_tests vp8_multi_resolution_encoder_verify_environment "${vp8_mre_tests}"
diff --git a/libvpx/libvpx/test/vp8cx_set_ref.sh b/libvpx/libvpx/test/vp8cx_set_ref.sh
new file mode 100755
index 0000000..5d760bc
--- /dev/null
+++ b/libvpx/libvpx/test/vp8cx_set_ref.sh
@@ -0,0 +1,57 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx vp8cx_set_ref example. To add new tests to this
+##  file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vp8cx_set_ref_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+vp8cx_set_ref_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs vp8cx_set_ref and updates the reference frame before encoding frame 90.
+# $1 is the codec name, which vp8cx_set_ref does not support at present: It's
+# currently used only to name the output file.
+# TODO(tomfinegan): Pass the codec param once the example is updated to support
+# VP9.
+vpx_set_ref() {
+  local encoder="${LIBVPX_BIN_PATH}/vp8cx_set_ref${VPX_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/vp8cx_set_ref_${codec}.ivf"
+  local ref_frame_num=90
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" \
+      "${ref_frame_num}" ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+vp8cx_set_ref_vp8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_set_ref vp8 || return 1
+  fi
+}
+
+vp8cx_set_ref_tests="vp8cx_set_ref_vp8"
+
+run_tests vp8cx_set_ref_verify_environment "${vp8cx_set_ref_tests}"
diff --git a/libvpx/libvpx/test/vp9_arf_freq_test.cc b/libvpx/libvpx/test/vp9_arf_freq_test.cc
new file mode 100644
index 0000000..aa3e34d
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_arf_freq_test.cc
@@ -0,0 +1,232 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "test/yuv_video_source.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+
+namespace {
+
+const unsigned int kFrames = 100;
+const int kBitrate = 500;
+
+#define ARF_NOT_SEEN               1000001
+#define ARF_SEEN_ONCE              1000000
+
+typedef struct {
+  const char *filename;
+  unsigned int width;
+  unsigned int height;
+  unsigned int framerate_num;
+  unsigned int framerate_den;
+  unsigned int input_bit_depth;
+  vpx_img_fmt fmt;
+  vpx_bit_depth_t bit_depth;
+  unsigned int profile;
+} TestVideoParam;
+
+typedef struct {
+  libvpx_test::TestMode mode;
+  int cpu_used;
+} TestEncodeParam;
+
+const TestVideoParam kTestVectors[] = {
+  // artificially increase framerate to trigger default check
+  {"hantro_collage_w352h288.yuv", 352, 288, 5000, 1,
+    8, VPX_IMG_FMT_I420, VPX_BITS_8, 0},
+  {"hantro_collage_w352h288.yuv", 352, 288, 30, 1,
+    8, VPX_IMG_FMT_I420, VPX_BITS_8, 0},
+  {"rush_hour_444.y4m", 352, 288, 30, 1,
+    8, VPX_IMG_FMT_I444, VPX_BITS_8, 1},
+#if CONFIG_VP9_HIGHBITDEPTH
+  // Add list of profile 2/3 test videos here ...
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+};
+
+const TestEncodeParam kEncodeVectors[] = {
+  {::libvpx_test::kOnePassGood, 2},
+  {::libvpx_test::kOnePassGood, 5},
+  {::libvpx_test::kTwoPassGood, 1},
+  {::libvpx_test::kTwoPassGood, 2},
+  {::libvpx_test::kTwoPassGood, 5},
+  {::libvpx_test::kRealTime, 5},
+};
+
+const int kMinArfVectors[] = {
+  // NOTE: 0 refers to the default built-in logic in:
+  //       vp9_rc_get_default_min_gf_interval(...)
+  0, 4, 8, 12, 15
+};
+
+int is_extension_y4m(const char *filename) {
+  const char *dot = strrchr(filename, '.');
+  if (!dot || dot == filename)
+    return 0;
+  else
+    return !strcmp(dot, ".y4m");
+}
+
+class ArfFreqTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith3Params<TestVideoParam, \
+                                                 TestEncodeParam, int> {
+ protected:
+  ArfFreqTest()
+      : EncoderTest(GET_PARAM(0)),
+        test_video_param_(GET_PARAM(1)),
+        test_encode_param_(GET_PARAM(2)),
+        min_arf_requested_(GET_PARAM(3)) {
+  }
+
+  virtual ~ArfFreqTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(test_encode_param_.mode);
+    if (test_encode_param_.mode != ::libvpx_test::kRealTime) {
+      cfg_.g_lag_in_frames = 25;
+      cfg_.rc_end_usage = VPX_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = VPX_CBR;
+      cfg_.rc_buf_sz = 1000;
+      cfg_.rc_buf_initial_sz = 500;
+      cfg_.rc_buf_optimal_sz = 600;
+    }
+    dec_cfg_.threads = 4;
+  }
+
+  virtual void BeginPassHook(unsigned int) {
+    min_run_ = ARF_NOT_SEEN;
+    run_of_visible_frames_ = 0;
+  }
+
+  int GetNumFramesInPkt(const vpx_codec_cx_pkt_t *pkt) {
+    const uint8_t *buffer = reinterpret_cast<uint8_t*>(pkt->data.frame.buf);
+    const uint8_t marker = buffer[pkt->data.frame.sz - 1];
+    const int mag = ((marker >> 3) & 3) + 1;
+    int frames = (marker & 0x7) + 1;
+    const unsigned int index_sz = 2 + mag  * frames;
+    // Check for superframe or not.
+    // Assume superframe has only one visible frame, the rest being
+    // invisible. If superframe index is not found, then there is only
+    // one frame.
+    if (!((marker & 0xe0) == 0xc0 &&
+          pkt->data.frame.sz >= index_sz &&
+          buffer[pkt->data.frame.sz - index_sz] == marker)) {
+      frames = 1;
+    }
+    return frames;
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->kind != VPX_CODEC_CX_FRAME_PKT)
+      return;
+    const int frames = GetNumFramesInPkt(pkt);
+    if (frames == 1) {
+      run_of_visible_frames_++;
+    } else if (frames == 2) {
+      if (min_run_ == ARF_NOT_SEEN) {
+        min_run_ = ARF_SEEN_ONCE;
+      } else if (min_run_ == ARF_SEEN_ONCE ||
+                 run_of_visible_frames_ < min_run_) {
+        min_run_ = run_of_visible_frames_;
+      }
+      run_of_visible_frames_ = 1;
+    } else {
+      min_run_ = 0;
+      run_of_visible_frames_ = 1;
+    }
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      encoder->Control(VP9E_SET_FRAME_PARALLEL_DECODING, 1);
+      encoder->Control(VP9E_SET_TILE_COLUMNS, 4);
+      encoder->Control(VP8E_SET_CPUUSED, test_encode_param_.cpu_used);
+      encoder->Control(VP9E_SET_MIN_GF_INTERVAL, min_arf_requested_);
+      if (test_encode_param_.mode != ::libvpx_test::kRealTime) {
+        encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+        encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+      }
+    }
+  }
+
+  int GetMinVisibleRun() const {
+    return min_run_;
+  }
+
+  int GetMinArfDistanceRequested() const {
+    if (min_arf_requested_)
+      return min_arf_requested_;
+    else
+      return vp9_rc_get_default_min_gf_interval(
+          test_video_param_.width, test_video_param_.height,
+          (double)test_video_param_.framerate_num /
+          test_video_param_.framerate_den);
+  }
+
+  TestVideoParam test_video_param_;
+  TestEncodeParam test_encode_param_;
+
+ private:
+  int min_arf_requested_;
+  int min_run_;
+  int run_of_visible_frames_;
+};
+
+TEST_P(ArfFreqTest, MinArfFreqTest) {
+  cfg_.rc_target_bitrate = kBitrate;
+  cfg_.g_error_resilient = 0;
+  cfg_.g_profile = test_video_param_.profile;
+  cfg_.g_input_bit_depth = test_video_param_.input_bit_depth;
+  cfg_.g_bit_depth = test_video_param_.bit_depth;
+  init_flags_ = VPX_CODEC_USE_PSNR;
+  if (cfg_.g_bit_depth > 8)
+    init_flags_ |= VPX_CODEC_USE_HIGHBITDEPTH;
+
+  libvpx_test::VideoSource *video;
+  if (is_extension_y4m(test_video_param_.filename)) {
+    video = new libvpx_test::Y4mVideoSource(test_video_param_.filename,
+                                            0, kFrames);
+  } else {
+    video = new libvpx_test::YUVVideoSource(test_video_param_.filename,
+                                            test_video_param_.fmt,
+                                            test_video_param_.width,
+                                            test_video_param_.height,
+                                            test_video_param_.framerate_num,
+                                            test_video_param_.framerate_den,
+                                            0, kFrames);
+  }
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video));
+  const int min_run = GetMinVisibleRun();
+  const int min_arf_dist_requested = GetMinArfDistanceRequested();
+  if (min_run != ARF_NOT_SEEN && min_run != ARF_SEEN_ONCE) {
+    const int min_arf_dist = min_run + 1;
+    EXPECT_GE(min_arf_dist, min_arf_dist_requested);
+  }
+  delete(video);
+}
+
+VP9_INSTANTIATE_TEST_CASE(
+    ArfFreqTest,
+    ::testing::ValuesIn(kTestVectors),
+    ::testing::ValuesIn(kEncodeVectors),
+    ::testing::ValuesIn(kMinArfVectors));
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_boolcoder_test.cc b/libvpx/libvpx/test/vp9_boolcoder_test.cc
new file mode 100644
index 0000000..c61bb4a
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_boolcoder_test.cc
@@ -0,0 +1,88 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "test/acm_random.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/bitreader.h"
+#include "vpx_dsp/bitwriter.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+const int num_tests = 10;
+}  // namespace
+
+TEST(VP9, TestBitIO) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  for (int n = 0; n < num_tests; ++n) {
+    for (int method = 0; method <= 7; ++method) {   // we generate various proba
+      const int kBitsToTest = 1000;
+      uint8_t probas[kBitsToTest];
+
+      for (int i = 0; i < kBitsToTest; ++i) {
+        const int parity = i & 1;
+        probas[i] =
+          (method == 0) ? 0 : (method == 1) ? 255 :
+          (method == 2) ? 128 :
+          (method == 3) ? rnd.Rand8() :
+          (method == 4) ? (parity ? 0 : 255) :
+            // alternate between low and high proba:
+            (method == 5) ? (parity ? rnd(128) : 255 - rnd(128)) :
+            (method == 6) ?
+            (parity ? rnd(64) : 255 - rnd(64)) :
+            (parity ? rnd(32) : 255 - rnd(32));
+      }
+      for (int bit_method = 0; bit_method <= 3; ++bit_method) {
+        const int random_seed = 6432;
+        const int kBufferSize = 10000;
+        ACMRandom bit_rnd(random_seed);
+        vpx_writer bw;
+        uint8_t bw_buffer[kBufferSize];
+        vpx_start_encode(&bw, bw_buffer);
+
+        int bit = (bit_method == 0) ? 0 : (bit_method == 1) ? 1 : 0;
+        for (int i = 0; i < kBitsToTest; ++i) {
+          if (bit_method == 2) {
+            bit = (i & 1);
+          } else if (bit_method == 3) {
+            bit = bit_rnd(2);
+          }
+          vpx_write(&bw, bit, static_cast<int>(probas[i]));
+        }
+
+        vpx_stop_encode(&bw);
+
+        // First bit should be zero
+        GTEST_ASSERT_EQ(bw_buffer[0] & 0x80, 0);
+
+        vpx_reader br;
+        vpx_reader_init(&br, bw_buffer, kBufferSize, NULL, NULL);
+        bit_rnd.Reset(random_seed);
+        for (int i = 0; i < kBitsToTest; ++i) {
+          if (bit_method == 2) {
+            bit = (i & 1);
+          } else if (bit_method == 3) {
+            bit = bit_rnd(2);
+          }
+          GTEST_ASSERT_EQ(vpx_read(&br, probas[i]), bit)
+              << "pos: " << i << " / " << kBitsToTest
+              << " bit_method: " << bit_method
+              << " method: " << method;
+        }
+      }
+    }
+  }
+}
diff --git a/libvpx/libvpx/test/vp9_decrypt_test.cc b/libvpx/libvpx/test/vp9_decrypt_test.cc
new file mode 100644
index 0000000..d988612
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_decrypt_test.cc
@@ -0,0 +1,71 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/ivf_video_source.h"
+
+namespace {
+// In a real use the 'decrypt_state' parameter will be a pointer to a struct
+// with whatever internal state the decryptor uses. For testing we'll just
+// xor with a constant key, and decrypt_state will point to the start of
+// the original buffer.
+const uint8_t test_key[16] = {
+  0x01, 0x12, 0x23, 0x34, 0x45, 0x56, 0x67, 0x78,
+  0x89, 0x9a, 0xab, 0xbc, 0xcd, 0xde, 0xef, 0xf0
+};
+
+void encrypt_buffer(const uint8_t *src, uint8_t *dst, size_t size,
+                    ptrdiff_t offset) {
+  for (size_t i = 0; i < size; ++i) {
+    dst[i] = src[i] ^ test_key[(offset + i) & 15];
+  }
+}
+
+void test_decrypt_cb(void *decrypt_state, const uint8_t *input,
+                     uint8_t *output, int count) {
+  encrypt_buffer(input, output, count,
+                 input - reinterpret_cast<uint8_t *>(decrypt_state));
+}
+
+}  // namespace
+
+namespace libvpx_test {
+
+TEST(TestDecrypt, DecryptWorksVp9) {
+  libvpx_test::IVFVideoSource video("vp90-2-05-resize.ivf");
+  video.Init();
+
+  vpx_codec_dec_cfg_t dec_cfg = vpx_codec_dec_cfg_t();
+  VP9Decoder decoder(dec_cfg, 0);
+
+  video.Begin();
+
+  // no decryption
+  vpx_codec_err_t res = decoder.DecodeFrame(video.cxdata(), video.frame_size());
+  ASSERT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+
+  // decrypt frame
+  video.Next();
+
+  std::vector<uint8_t> encrypted(video.frame_size());
+  encrypt_buffer(video.cxdata(), &encrypted[0], video.frame_size(), 0);
+  vpx_decrypt_init di = { test_decrypt_cb, &encrypted[0] };
+  decoder.Control(VPXD_SET_DECRYPTOR, &di);
+
+  res = decoder.DecodeFrame(&encrypted[0], encrypted.size());
+  ASSERT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+}
+
+}  // namespace libvpx_test
diff --git a/libvpx/libvpx/test/vp9_denoiser_sse2_test.cc b/libvpx/libvpx/test/vp9_denoiser_sse2_test.cc
new file mode 100644
index 0000000..c84d7ff
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_denoiser_sse2_test.cc
@@ -0,0 +1,100 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_denoiser.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+
+const int kNumPixels = 64 * 64;
+class VP9DenoiserTest : public ::testing::TestWithParam<BLOCK_SIZE> {
+ public:
+  virtual ~VP9DenoiserTest() {}
+
+  virtual void SetUp() {
+    bs_ = GetParam();
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  BLOCK_SIZE bs_;
+};
+
+TEST_P(VP9DenoiserTest, BitexactCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 4000;
+
+  // Allocate the space for input and output,
+  // where sig_block is the block to be denoised,
+  // mc_avg_block is the denoised reference block,
+  // avg_block_c is the denoised result from C code,
+  // avg_block_sse2 is the denoised result from SSE2 code.
+  DECLARE_ALIGNED(16, uint8_t, sig_block[kNumPixels]);
+  DECLARE_ALIGNED(16, uint8_t, mc_avg_block[kNumPixels]);
+  DECLARE_ALIGNED(16, uint8_t, avg_block_c[kNumPixels]);
+  DECLARE_ALIGNED(16, uint8_t, avg_block_sse2[kNumPixels]);
+
+  for (int i = 0; i < count_test_block; ++i) {
+    // Generate random motion magnitude, 20% of which exceed the threshold.
+    const int motion_magnitude_random =
+        rnd.Rand8() % static_cast<int>(MOTION_MAGNITUDE_THRESHOLD * 1.2);
+
+    // Initialize a test block with random number in range [0, 255].
+    for (int j = 0; j < kNumPixels; ++j) {
+      int temp = 0;
+      sig_block[j] = rnd.Rand8();
+      // The pixels in mc_avg_block are generated by adding a random
+      // number in range [-19, 19] to corresponding pixels in sig_block.
+      temp = sig_block[j] + ((rnd.Rand8() % 2 == 0) ? -1 : 1) *
+             (rnd.Rand8() % 20);
+      // Clip.
+      mc_avg_block[j] = (temp < 0) ? 0 : ((temp > 255) ? 255 : temp);
+    }
+
+    ASM_REGISTER_STATE_CHECK(vp9_denoiser_filter_c(
+        sig_block, 64, mc_avg_block, 64, avg_block_c,
+        64, 0, bs_, motion_magnitude_random));
+
+    ASM_REGISTER_STATE_CHECK(vp9_denoiser_filter_sse2(
+        sig_block, 64, mc_avg_block, 64, avg_block_sse2,
+        64, 0, bs_, motion_magnitude_random));
+
+    // Test bitexactness.
+    for (int h = 0; h < (4 << b_height_log2_lookup[bs_]); ++h) {
+      for (int w = 0; w < (4 << b_width_log2_lookup[bs_]); ++w) {
+        EXPECT_EQ(avg_block_c[h * 64 + w], avg_block_sse2[h * 64 + w]);
+      }
+    }
+  }
+}
+
+// Test for all block size.
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VP9DenoiserTest,
+    ::testing::Values(BLOCK_8X8, BLOCK_8X16, BLOCK_16X8, BLOCK_16X16,
+                      BLOCK_16X32, BLOCK_32X16, BLOCK_32X32, BLOCK_32X64,
+                      BLOCK_64X32, BLOCK_64X64));
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_encoder_parms_get_to_decoder.cc b/libvpx/libvpx/test/vp9_encoder_parms_get_to_decoder.cc
new file mode 100644
index 0000000..bd84098
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_encoder_parms_get_to_decoder.cc
@@ -0,0 +1,153 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "vp9/vp9_dx_iface.h"
+
+namespace {
+
+const int kCpuUsed = 2;
+
+struct EncodePerfTestVideo {
+  const char *name;
+  uint32_t width;
+  uint32_t height;
+  uint32_t bitrate;
+  int frames;
+};
+
+const EncodePerfTestVideo kVP9EncodePerfTestVectors[] = {
+  {"niklas_1280_720_30.y4m", 1280, 720, 600, 10},
+};
+
+struct EncodeParameters {
+  int32_t tile_rows;
+  int32_t tile_cols;
+  int32_t lossless;
+  int32_t error_resilient;
+  int32_t frame_parallel;
+  vpx_color_range_t color_range;
+  vpx_color_space_t cs;
+  int render_size[2];
+  // TODO(JBB): quantizers / bitrate
+};
+
+const EncodeParameters kVP9EncodeParameterSet[] = {
+  {0, 0, 0, 1, 0, VPX_CR_STUDIO_RANGE, VPX_CS_BT_601, { 0, 0 }},
+  {0, 0, 0, 0, 0, VPX_CR_FULL_RANGE, VPX_CS_BT_709, { 0, 0 }},
+  {0, 0, 1, 0, 0, VPX_CR_FULL_RANGE, VPX_CS_BT_2020, { 0, 0 }},
+  {0, 2, 0, 0, 1, VPX_CR_STUDIO_RANGE, VPX_CS_UNKNOWN, { 640, 480 }},
+  // TODO(JBB): Test profiles (requires more work).
+};
+
+class VpxEncoderParmsGetToDecoder
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith2Params<EncodeParameters,
+                                                 EncodePerfTestVideo> {
+ protected:
+  VpxEncoderParmsGetToDecoder()
+      : EncoderTest(GET_PARAM(0)), encode_parms(GET_PARAM(1)) {}
+
+  virtual ~VpxEncoderParmsGetToDecoder() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(::libvpx_test::kTwoPassGood);
+    cfg_.g_lag_in_frames = 25;
+    cfg_.g_error_resilient = encode_parms.error_resilient;
+    dec_cfg_.threads = 4;
+    test_video_ = GET_PARAM(2);
+    cfg_.rc_target_bitrate = test_video_.bitrate;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP9E_SET_COLOR_SPACE, encode_parms.cs);
+      encoder->Control(VP9E_SET_COLOR_RANGE, encode_parms.color_range);
+      encoder->Control(VP9E_SET_LOSSLESS, encode_parms.lossless);
+      encoder->Control(VP9E_SET_FRAME_PARALLEL_DECODING,
+                       encode_parms.frame_parallel);
+      encoder->Control(VP9E_SET_TILE_ROWS, encode_parms.tile_rows);
+      encoder->Control(VP9E_SET_TILE_COLUMNS, encode_parms.tile_cols);
+      encoder->Control(VP8E_SET_CPUUSED, kCpuUsed);
+      encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+      encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+      encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+      encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+      if (encode_parms.render_size[0] > 0 && encode_parms.render_size[1] > 0)
+        encoder->Control(VP9E_SET_RENDER_SIZE, encode_parms.render_size);
+    }
+  }
+
+  virtual bool HandleDecodeResult(const vpx_codec_err_t res_dec,
+                                  const libvpx_test::VideoSource & /*video*/,
+                                  libvpx_test::Decoder *decoder) {
+    vpx_codec_ctx_t *const vp9_decoder = decoder->GetDecoder();
+    vpx_codec_alg_priv_t *const priv =
+        reinterpret_cast<vpx_codec_alg_priv_t *>(vp9_decoder->priv);
+    FrameWorkerData *const worker_data =
+        reinterpret_cast<FrameWorkerData *>(priv->frame_workers[0].data1);
+    VP9_COMMON *const common = &worker_data->pbi->common;
+
+    if (encode_parms.lossless) {
+      EXPECT_EQ(0, common->base_qindex);
+      EXPECT_EQ(0, common->y_dc_delta_q);
+      EXPECT_EQ(0, common->uv_dc_delta_q);
+      EXPECT_EQ(0, common->uv_ac_delta_q);
+      EXPECT_EQ(ONLY_4X4, common->tx_mode);
+    }
+    EXPECT_EQ(encode_parms.error_resilient, common->error_resilient_mode);
+    if (encode_parms.error_resilient) {
+      EXPECT_EQ(1, common->frame_parallel_decoding_mode);
+      EXPECT_EQ(0, common->use_prev_frame_mvs);
+    } else {
+      EXPECT_EQ(encode_parms.frame_parallel,
+                common->frame_parallel_decoding_mode);
+    }
+    EXPECT_EQ(encode_parms.color_range, common->color_range);
+    EXPECT_EQ(encode_parms.cs, common->color_space);
+    if (encode_parms.render_size[0] > 0 && encode_parms.render_size[1] > 0) {
+      EXPECT_EQ(encode_parms.render_size[0], common->render_width);
+      EXPECT_EQ(encode_parms.render_size[1], common->render_height);
+    }
+    EXPECT_EQ(encode_parms.tile_cols, common->log2_tile_cols);
+    EXPECT_EQ(encode_parms.tile_rows, common->log2_tile_rows);
+
+    EXPECT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+    return VPX_CODEC_OK == res_dec;
+  }
+
+  EncodePerfTestVideo test_video_;
+
+ private:
+  EncodeParameters encode_parms;
+};
+
+TEST_P(VpxEncoderParmsGetToDecoder, BitstreamParms) {
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  libvpx_test::VideoSource *const video =
+      new libvpx_test::Y4mVideoSource(test_video_.name, 0, test_video_.frames);
+  ASSERT_TRUE(video != NULL);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video));
+  delete video;
+}
+
+VP9_INSTANTIATE_TEST_CASE(VpxEncoderParmsGetToDecoder,
+                          ::testing::ValuesIn(kVP9EncodeParameterSet),
+                          ::testing::ValuesIn(kVP9EncodePerfTestVectors));
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_end_to_end_test.cc b/libvpx/libvpx/test/vp9_end_to_end_test.cc
new file mode 100644
index 0000000..666919f
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_end_to_end_test.cc
@@ -0,0 +1,189 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "test/yuv_video_source.h"
+
+namespace {
+
+const unsigned int kWidth  = 160;
+const unsigned int kHeight = 90;
+const unsigned int kFramerate = 50;
+const unsigned int kFrames = 10;
+const int kBitrate = 500;
+// List of psnr thresholds for speed settings 0-7 and 5 encoding modes
+const double kPsnrThreshold[][5] = {
+  { 36.0, 37.0, 37.0, 37.0, 37.0 },
+  { 35.0, 36.0, 36.0, 36.0, 36.0 },
+  { 34.0, 35.0, 35.0, 35.0, 35.0 },
+  { 33.0, 34.0, 34.0, 34.0, 34.0 },
+  { 32.0, 33.0, 33.0, 33.0, 33.0 },
+  { 31.0, 32.0, 32.0, 32.0, 32.0 },
+  { 30.0, 31.0, 31.0, 31.0, 31.0 },
+  { 29.0, 30.0, 30.0, 30.0, 30.0 },
+};
+
+typedef struct {
+  const char *filename;
+  unsigned int input_bit_depth;
+  vpx_img_fmt fmt;
+  vpx_bit_depth_t bit_depth;
+  unsigned int profile;
+} TestVideoParam;
+
+const TestVideoParam kTestVectors[] = {
+  {"park_joy_90p_8_420.y4m", 8, VPX_IMG_FMT_I420, VPX_BITS_8, 0},
+  {"park_joy_90p_8_422.y4m", 8, VPX_IMG_FMT_I422, VPX_BITS_8, 1},
+  {"park_joy_90p_8_444.y4m", 8, VPX_IMG_FMT_I444, VPX_BITS_8, 1},
+  {"park_joy_90p_8_440.yuv", 8, VPX_IMG_FMT_I440, VPX_BITS_8, 1},
+#if CONFIG_VP9_HIGHBITDEPTH
+  {"park_joy_90p_10_420.y4m", 10, VPX_IMG_FMT_I42016, VPX_BITS_10, 2},
+  {"park_joy_90p_10_422.y4m", 10, VPX_IMG_FMT_I42216, VPX_BITS_10, 3},
+  {"park_joy_90p_10_444.y4m", 10, VPX_IMG_FMT_I44416, VPX_BITS_10, 3},
+  {"park_joy_90p_10_440.yuv", 10, VPX_IMG_FMT_I44016, VPX_BITS_10, 3},
+  {"park_joy_90p_12_420.y4m", 12, VPX_IMG_FMT_I42016, VPX_BITS_12, 2},
+  {"park_joy_90p_12_422.y4m", 12, VPX_IMG_FMT_I42216, VPX_BITS_12, 3},
+  {"park_joy_90p_12_444.y4m", 12, VPX_IMG_FMT_I44416, VPX_BITS_12, 3},
+  {"park_joy_90p_12_440.yuv", 12, VPX_IMG_FMT_I44016, VPX_BITS_12, 3},
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+};
+
+// Encoding modes tested
+const libvpx_test::TestMode kEncodingModeVectors[] = {
+  ::libvpx_test::kTwoPassGood,
+  ::libvpx_test::kOnePassGood,
+  ::libvpx_test::kRealTime,
+};
+
+// Speed settings tested
+const int kCpuUsedVectors[] = {1, 2, 3, 5, 6};
+
+int is_extension_y4m(const char *filename) {
+  const char *dot = strrchr(filename, '.');
+  if (!dot || dot == filename)
+    return 0;
+  else
+    return !strcmp(dot, ".y4m");
+}
+
+class EndToEndTestLarge
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith3Params<libvpx_test::TestMode, \
+                                                 TestVideoParam, int> {
+ protected:
+  EndToEndTestLarge()
+      : EncoderTest(GET_PARAM(0)),
+        test_video_param_(GET_PARAM(2)),
+        cpu_used_(GET_PARAM(3)),
+        psnr_(0.0),
+        nframes_(0),
+        encoding_mode_(GET_PARAM(1)) {
+  }
+
+  virtual ~EndToEndTestLarge() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+    if (encoding_mode_ != ::libvpx_test::kRealTime) {
+      cfg_.g_lag_in_frames = 5;
+      cfg_.rc_end_usage = VPX_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = VPX_CBR;
+      cfg_.rc_buf_sz = 1000;
+      cfg_.rc_buf_initial_sz = 500;
+      cfg_.rc_buf_optimal_sz = 600;
+    }
+    dec_cfg_.threads = 4;
+  }
+
+  virtual void BeginPassHook(unsigned int) {
+    psnr_ = 0.0;
+    nframes_ = 0;
+  }
+
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    psnr_ += pkt->data.psnr.psnr[0];
+    nframes_++;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(VP9E_SET_FRAME_PARALLEL_DECODING, 1);
+      encoder->Control(VP9E_SET_TILE_COLUMNS, 4);
+      encoder->Control(VP8E_SET_CPUUSED, cpu_used_);
+      if (encoding_mode_ != ::libvpx_test::kRealTime) {
+        encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+        encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+      }
+    }
+  }
+
+  double GetAveragePsnr() const {
+    if (nframes_)
+      return psnr_ / nframes_;
+    return 0.0;
+  }
+
+  double GetPsnrThreshold() {
+    return kPsnrThreshold[cpu_used_][encoding_mode_];
+  }
+
+  TestVideoParam test_video_param_;
+  int cpu_used_;
+
+ private:
+  double psnr_;
+  unsigned int nframes_;
+  libvpx_test::TestMode encoding_mode_;
+};
+
+TEST_P(EndToEndTestLarge, EndtoEndPSNRTest) {
+  cfg_.rc_target_bitrate = kBitrate;
+  cfg_.g_error_resilient = 0;
+  cfg_.g_profile = test_video_param_.profile;
+  cfg_.g_input_bit_depth = test_video_param_.input_bit_depth;
+  cfg_.g_bit_depth = test_video_param_.bit_depth;
+  init_flags_ = VPX_CODEC_USE_PSNR;
+  if (cfg_.g_bit_depth > 8)
+    init_flags_ |= VPX_CODEC_USE_HIGHBITDEPTH;
+
+  libvpx_test::VideoSource *video;
+  if (is_extension_y4m(test_video_param_.filename)) {
+    video = new libvpx_test::Y4mVideoSource(test_video_param_.filename,
+                                            0, kFrames);
+  } else {
+    video = new libvpx_test::YUVVideoSource(test_video_param_.filename,
+                                            test_video_param_.fmt,
+                                            kWidth, kHeight,
+                                            kFramerate, 1, 0, kFrames);
+  }
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video));
+  const double psnr = GetAveragePsnr();
+  EXPECT_GT(psnr, GetPsnrThreshold());
+  delete(video);
+}
+
+VP9_INSTANTIATE_TEST_CASE(
+    EndToEndTestLarge,
+    ::testing::ValuesIn(kEncodingModeVectors),
+    ::testing::ValuesIn(kTestVectors),
+    ::testing::ValuesIn(kCpuUsedVectors));
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_error_block_test.cc b/libvpx/libvpx/test/vp9_error_block_test.cc
new file mode 100644
index 0000000..23a249e
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_error_block_test.cc
@@ -0,0 +1,211 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cmath>
+#include <cstdlib>
+#include <string>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_integer.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+#if CONFIG_VP9_HIGHBITDEPTH
+const int kNumIterations = 1000;
+
+typedef int64_t (*ErrorBlockFunc)(const tran_low_t *coeff,
+                                  const tran_low_t *dqcoeff,
+                                  intptr_t block_size,
+                                  int64_t *ssz, int bps);
+
+typedef std::tr1::tuple<ErrorBlockFunc, ErrorBlockFunc, vpx_bit_depth_t>
+                        ErrorBlockParam;
+
+class ErrorBlockTest
+  : public ::testing::TestWithParam<ErrorBlockParam> {
+ public:
+  virtual ~ErrorBlockTest() {}
+  virtual void SetUp() {
+    error_block_op_     = GET_PARAM(0);
+    ref_error_block_op_ = GET_PARAM(1);
+    bit_depth_  = GET_PARAM(2);
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  vpx_bit_depth_t bit_depth_;
+  ErrorBlockFunc error_block_op_;
+  ErrorBlockFunc ref_error_block_op_;
+};
+
+TEST_P(ErrorBlockTest, OperationCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, tran_low_t, coeff[4096]);
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[4096]);
+  int err_count_total = 0;
+  int first_failure = -1;
+  intptr_t block_size;
+  int64_t ssz;
+  int64_t ret;
+  int64_t ref_ssz;
+  int64_t ref_ret;
+  const int msb = bit_depth_ + 8 - 1;
+  for (int i = 0; i < kNumIterations; ++i) {
+    int err_count = 0;
+    block_size = 16 << (i % 9);  // All block sizes from 4x4, 8x4 ..64x64
+    for (int j = 0; j < block_size; j++) {
+      // coeff and dqcoeff will always have at least the same sign, and this
+      // can be used for optimization, so generate test input precisely.
+      if (rnd(2)) {
+        // Positive number
+        coeff[j]   = rnd(1 << msb);
+        dqcoeff[j] = rnd(1 << msb);
+      } else {
+        // Negative number
+        coeff[j]   = -rnd(1 << msb);
+        dqcoeff[j] = -rnd(1 << msb);
+      }
+    }
+    ref_ret = ref_error_block_op_(coeff, dqcoeff, block_size, &ref_ssz,
+                                  bit_depth_);
+    ASM_REGISTER_STATE_CHECK(ret = error_block_op_(coeff, dqcoeff, block_size,
+                                                   &ssz, bit_depth_));
+    err_count += (ref_ret != ret) | (ref_ssz != ssz);
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Error Block Test, C output doesn't match optimized output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(ErrorBlockTest, ExtremeValues) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, tran_low_t, coeff[4096]);
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[4096]);
+  int err_count_total = 0;
+  int first_failure = -1;
+  intptr_t block_size;
+  int64_t ssz;
+  int64_t ret;
+  int64_t ref_ssz;
+  int64_t ref_ret;
+  const int msb = bit_depth_ + 8 - 1;
+  int max_val = ((1 << msb) - 1);
+  for (int i = 0; i < kNumIterations; ++i) {
+    int err_count = 0;
+    int k = (i / 9) % 9;
+
+    // Change the maximum coeff value, to test different bit boundaries
+    if ( k == 8 && (i % 9) == 0 ) {
+      max_val >>= 1;
+    }
+    block_size = 16 << (i % 9);  // All block sizes from 4x4, 8x4 ..64x64
+    for (int j = 0; j < block_size; j++) {
+      if (k < 4) {
+        // Test at positive maximum values
+        coeff[j]   = k % 2 ? max_val : 0;
+        dqcoeff[j] = (k >> 1) % 2 ? max_val : 0;
+      } else if (k < 8) {
+        // Test at negative maximum values
+        coeff[j]   = k % 2 ? -max_val : 0;
+        dqcoeff[j] = (k >> 1) % 2 ? -max_val : 0;
+      } else {
+        if (rnd(2)) {
+          // Positive number
+          coeff[j]   = rnd(1 << 14);
+          dqcoeff[j] = rnd(1 << 14);
+        } else {
+          // Negative number
+          coeff[j]   = -rnd(1 << 14);
+          dqcoeff[j] = -rnd(1 << 14);
+        }
+      }
+    }
+    ref_ret = ref_error_block_op_(coeff, dqcoeff, block_size, &ref_ssz,
+                                  bit_depth_);
+    ASM_REGISTER_STATE_CHECK(ret = error_block_op_(coeff, dqcoeff, block_size,
+                                                   &ssz, bit_depth_));
+    err_count += (ref_ret != ret) | (ref_ssz != ssz);
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Error Block Test, C output doesn't match optimized output. "
+      << "First failed at test case " << first_failure;
+}
+
+using std::tr1::make_tuple;
+
+#if CONFIG_USE_X86INC
+int64_t wrap_vp9_highbd_block_error_8bit_c(const tran_low_t *coeff,
+                                           const tran_low_t *dqcoeff,
+                                           intptr_t block_size,
+                                           int64_t *ssz, int bps) {
+  EXPECT_EQ(8, bps);
+  return vp9_highbd_block_error_8bit_c(coeff, dqcoeff, block_size, ssz);
+}
+
+#if HAVE_SSE2
+int64_t wrap_vp9_highbd_block_error_8bit_sse2(const tran_low_t *coeff,
+                                              const tran_low_t *dqcoeff,
+                                              intptr_t block_size,
+                                              int64_t *ssz, int bps) {
+  EXPECT_EQ(8, bps);
+  return vp9_highbd_block_error_8bit_sse2(coeff, dqcoeff, block_size, ssz);
+}
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, ErrorBlockTest,
+    ::testing::Values(
+        make_tuple(&vp9_highbd_block_error_sse2,
+                   &vp9_highbd_block_error_c, VPX_BITS_10),
+        make_tuple(&vp9_highbd_block_error_sse2,
+                   &vp9_highbd_block_error_c, VPX_BITS_12),
+        make_tuple(&vp9_highbd_block_error_sse2,
+                   &vp9_highbd_block_error_c, VPX_BITS_8),
+        make_tuple(&wrap_vp9_highbd_block_error_8bit_sse2,
+                   &wrap_vp9_highbd_block_error_8bit_c, VPX_BITS_8)));
+#endif  // HAVE_SSE2
+
+#if HAVE_AVX
+int64_t wrap_vp9_highbd_block_error_8bit_avx(const tran_low_t *coeff,
+                                              const tran_low_t *dqcoeff,
+                                              intptr_t block_size,
+                                              int64_t *ssz, int bps) {
+  EXPECT_EQ(8, bps);
+  return vp9_highbd_block_error_8bit_avx(coeff, dqcoeff, block_size, ssz);
+}
+
+INSTANTIATE_TEST_CASE_P(
+    AVX, ErrorBlockTest,
+    ::testing::Values(
+        make_tuple(&wrap_vp9_highbd_block_error_8bit_avx,
+                   &wrap_vp9_highbd_block_error_8bit_c, VPX_BITS_8)));
+#endif  // HAVE_AVX
+
+#endif  // CONFIG_USE_X86INC
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_ethread_test.cc b/libvpx/libvpx/test/vp9_ethread_test.cc
new file mode 100644
index 0000000..62b9109
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_ethread_test.cc
@@ -0,0 +1,131 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+#include <vector>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+
+namespace {
+class VPxEncoderThreadTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith2Params<libvpx_test::TestMode, int> {
+ protected:
+  VPxEncoderThreadTest()
+      : EncoderTest(GET_PARAM(0)),
+        encoder_initialized_(false),
+        tiles_(2),
+        encoding_mode_(GET_PARAM(1)),
+        set_cpu_used_(GET_PARAM(2)) {
+    init_flags_ = VPX_CODEC_USE_PSNR;
+    md5_.clear();
+  }
+  virtual ~VPxEncoderThreadTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+
+    if (encoding_mode_ != ::libvpx_test::kRealTime) {
+      cfg_.g_lag_in_frames = 3;
+      cfg_.rc_end_usage = VPX_VBR;
+      cfg_.rc_2pass_vbr_minsection_pct = 5;
+      cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = VPX_CBR;
+      cfg_.g_error_resilient = 1;
+    }
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_min_quantizer = 0;
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    encoder_initialized_ = false;
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource * /*video*/,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (!encoder_initialized_) {
+      // Encode 4 column tiles.
+      encoder->Control(VP9E_SET_TILE_COLUMNS, tiles_);
+      encoder->Control(VP8E_SET_CPUUSED, set_cpu_used_);
+      if (encoding_mode_ != ::libvpx_test::kRealTime) {
+        encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+        encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+      } else {
+        encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 0);
+        encoder->Control(VP9E_SET_AQ_MODE, 3);
+      }
+      encoder_initialized_ = true;
+    }
+  }
+
+  virtual void DecompressedFrameHook(const vpx_image_t &img,
+                                     vpx_codec_pts_t /*pts*/) {
+    ::libvpx_test::MD5 md5_res;
+    md5_res.Add(&img);
+    md5_.push_back(md5_res.Get());
+  }
+
+  virtual bool HandleDecodeResult(const vpx_codec_err_t res,
+                                  const libvpx_test::VideoSource& /*video*/,
+                                  libvpx_test::Decoder * /*decoder*/) {
+    if (res != VPX_CODEC_OK) {
+      EXPECT_EQ(VPX_CODEC_OK, res);
+      return false;
+    }
+
+    return true;
+  }
+
+  bool encoder_initialized_;
+  int tiles_;
+  ::libvpx_test::TestMode encoding_mode_;
+  int set_cpu_used_;
+  std::vector<std::string> md5_;
+};
+
+TEST_P(VPxEncoderThreadTest, EncoderResultTest) {
+  std::vector<std::string> single_thr_md5, multi_thr_md5;
+
+  ::libvpx_test::Y4mVideoSource video("niklas_1280_720_30.y4m", 15, 20);
+
+  cfg_.rc_target_bitrate = 1000;
+
+  // Encode using single thread.
+  cfg_.g_threads = 1;
+  init_flags_ = VPX_CODEC_USE_PSNR;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  single_thr_md5 = md5_;
+  md5_.clear();
+
+  // Encode using multiple threads.
+  cfg_.g_threads = 4;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  multi_thr_md5 = md5_;
+  md5_.clear();
+
+  // Compare to check if two vectors are equal.
+  ASSERT_EQ(single_thr_md5, multi_thr_md5);
+}
+
+VP9_INSTANTIATE_TEST_CASE(
+    VPxEncoderThreadTest,
+    ::testing::Values(::libvpx_test::kTwoPassGood, ::libvpx_test::kOnePassGood,
+                      ::libvpx_test::kRealTime),
+    ::testing::Range(1, 9));
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_frame_parallel_test.cc b/libvpx/libvpx/test/vp9_frame_parallel_test.cc
new file mode 100644
index 0000000..f0df88a
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_frame_parallel_test.cc
@@ -0,0 +1,220 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+
+using std::string;
+
+#if CONFIG_WEBM_IO
+
+struct PauseFileList {
+  const char *name;
+  // md5 sum for decoded frames which does not include skipped frames.
+  const char *expected_md5;
+  const int pause_frame_num;
+};
+
+// Decodes |filename| with |num_threads|. Pause at the specified frame_num,
+// seek to next key frame and then continue decoding until the end. Return
+// the md5 of the decoded frames which does not include skipped frames.
+string DecodeFileWithPause(const string &filename, int num_threads,
+                           int pause_num) {
+  libvpx_test::WebMVideoSource video(filename);
+  video.Init();
+  int in_frames = 0;
+  int out_frames = 0;
+
+  vpx_codec_dec_cfg_t cfg = {0};
+  cfg.threads = num_threads;
+  vpx_codec_flags_t flags = 0;
+  flags |= VPX_CODEC_USE_FRAME_THREADING;
+  libvpx_test::VP9Decoder decoder(cfg, flags, 0);
+
+  libvpx_test::MD5 md5;
+  video.Begin();
+
+  do {
+    ++in_frames;
+    const vpx_codec_err_t res =
+        decoder.DecodeFrame(video.cxdata(), video.frame_size());
+    if (res != VPX_CODEC_OK) {
+      EXPECT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+      break;
+    }
+
+    // Pause at specified frame number.
+    if (in_frames == pause_num) {
+      // Flush the decoder and then seek to next key frame.
+      decoder.DecodeFrame(NULL, 0);
+      video.SeekToNextKeyFrame();
+    } else {
+      video.Next();
+    }
+
+    // Flush the decoder at the end of the video.
+    if (!video.cxdata())
+      decoder.DecodeFrame(NULL, 0);
+
+    libvpx_test::DxDataIterator dec_iter = decoder.GetDxData();
+    const vpx_image_t *img;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next())) {
+      ++out_frames;
+      md5.Add(img);
+    }
+  } while (video.cxdata() != NULL);
+
+  EXPECT_EQ(in_frames, out_frames) <<
+      "Input frame count does not match output frame count";
+
+  return string(md5.Get());
+}
+
+void DecodeFilesWithPause(const PauseFileList files[]) {
+  for (const PauseFileList *iter = files; iter->name != NULL; ++iter) {
+    SCOPED_TRACE(iter->name);
+    for (int t = 2; t <= 8; ++t) {
+      EXPECT_EQ(iter->expected_md5,
+                DecodeFileWithPause(iter->name, t, iter->pause_frame_num))
+          << "threads = " << t;
+    }
+  }
+}
+
+TEST(VP9MultiThreadedFrameParallel, PauseSeekResume) {
+  // vp90-2-07-frame_parallel-1.webm is a 40 frame video file with
+  // one key frame for every ten frames.
+  static const PauseFileList files[] = {
+    { "vp90-2-07-frame_parallel-1.webm",
+      "6ea7c3875d67252e7caf2bc6e75b36b1", 6 },
+    { "vp90-2-07-frame_parallel-1.webm",
+      "4bb634160c7356a8d7d4299b6dc83a45", 12 },
+    { "vp90-2-07-frame_parallel-1.webm",
+      "89772591e6ef461f9fa754f916c78ed8", 26 },
+    { NULL, NULL, 0 },
+  };
+  DecodeFilesWithPause(files);
+}
+
+struct FileList {
+  const char *name;
+  // md5 sum for decoded frames which does not include corrupted frames.
+  const char *expected_md5;
+  // Expected number of decoded frames which does not include corrupted frames.
+  const int expected_frame_count;
+};
+
+// Decodes |filename| with |num_threads|. Return the md5 of the decoded
+// frames which does not include corrupted frames.
+string DecodeFile(const string &filename, int num_threads,
+                  int expected_frame_count) {
+  libvpx_test::WebMVideoSource video(filename);
+  video.Init();
+
+  vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+  cfg.threads = num_threads;
+  const vpx_codec_flags_t flags = VPX_CODEC_USE_FRAME_THREADING;
+  libvpx_test::VP9Decoder decoder(cfg, flags, 0);
+
+  libvpx_test::MD5 md5;
+  video.Begin();
+
+  int out_frames = 0;
+  do {
+    const vpx_codec_err_t res =
+        decoder.DecodeFrame(video.cxdata(), video.frame_size());
+    // TODO(hkuang): frame parallel mode should return an error on corruption.
+    if (res != VPX_CODEC_OK) {
+      EXPECT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+      break;
+    }
+
+    video.Next();
+
+    // Flush the decoder at the end of the video.
+    if (!video.cxdata())
+      decoder.DecodeFrame(NULL, 0);
+
+    libvpx_test::DxDataIterator dec_iter = decoder.GetDxData();
+    const vpx_image_t *img;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next())) {
+      ++out_frames;
+      md5.Add(img);
+    }
+  } while (video.cxdata() != NULL);
+
+  EXPECT_EQ(expected_frame_count, out_frames) <<
+      "Input frame count does not match expected output frame count";
+
+  return string(md5.Get());
+}
+
+void DecodeFiles(const FileList files[]) {
+  for (const FileList *iter = files; iter->name != NULL; ++iter) {
+    SCOPED_TRACE(iter->name);
+    for (int t = 2; t <= 8; ++t) {
+      EXPECT_EQ(iter->expected_md5,
+                DecodeFile(iter->name, t, iter->expected_frame_count))
+          << "threads = " << t;
+    }
+  }
+}
+
+TEST(VP9MultiThreadedFrameParallel, InvalidFileTest) {
+  static const FileList files[] = {
+    // invalid-vp90-2-07-frame_parallel-1.webm is a 40 frame video file with
+    // one key frame for every ten frames. The 11th frame has corrupted data.
+    { "invalid-vp90-2-07-frame_parallel-1.webm",
+      "0549d0f45f60deaef8eb708e6c0eb6cb", 30 },
+    // invalid-vp90-2-07-frame_parallel-2.webm is a 40 frame video file with
+    // one key frame for every ten frames. The 1st and 31st frames have
+    // corrupted data.
+    { "invalid-vp90-2-07-frame_parallel-2.webm",
+      "6a1f3cf6f9e7a364212fadb9580d525e", 20 },
+    // invalid-vp90-2-07-frame_parallel-3.webm is a 40 frame video file with
+    // one key frame for every ten frames. The 5th and 13th frames have
+    // corrupted data.
+    { "invalid-vp90-2-07-frame_parallel-3.webm",
+      "8256544308de926b0681e04685b98677", 27 },
+    { NULL, NULL, 0 },
+  };
+  DecodeFiles(files);
+}
+
+TEST(VP9MultiThreadedFrameParallel, ValidFileTest) {
+  static const FileList files[] = {
+#if CONFIG_VP9_HIGHBITDEPTH
+    { "vp92-2-20-10bit-yuv420.webm",
+      "a16b99df180c584e8db2ffeda987d293", 10 },
+#endif
+    { NULL, NULL, 0 },
+  };
+  DecodeFiles(files);
+}
+#endif  // CONFIG_WEBM_IO
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_intrapred_test.cc b/libvpx/libvpx/test/vp9_intrapred_test.cc
new file mode 100644
index 0000000..416f3c3
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_intrapred_test.cc
@@ -0,0 +1,231 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+
+using libvpx_test::ACMRandom;
+
+const int count_test_block = 100000;
+
+// Base class for VP9 intra prediction tests.
+class VP9IntraPredBase {
+ public:
+  virtual ~VP9IntraPredBase() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  virtual void Predict() = 0;
+
+  void CheckPrediction(int test_case_number, int *error_count) const {
+    // For each pixel ensure that the calculated value is the same as reference.
+    for (int y = 0; y < block_size_; y++) {
+      for (int x = 0; x < block_size_; x++) {
+        *error_count += ref_dst_[x + y * stride_] != dst_[x + y * stride_];
+        if (*error_count == 1) {
+          ASSERT_EQ(ref_dst_[x + y * stride_], dst_[x + y * stride_])
+              << " Failed on Test Case Number "<< test_case_number;
+        }
+      }
+    }
+  }
+
+  void RunTest(uint16_t* left_col, uint16_t* above_data,
+               uint16_t* dst, uint16_t* ref_dst) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    left_col_ = left_col;
+    dst_ = dst;
+    ref_dst_ = ref_dst;
+    above_row_ = above_data + 16;
+    int error_count = 0;
+    for (int i = 0; i < count_test_block; ++i) {
+      // Fill edges with random data, try first with saturated values.
+      for (int x = -1; x <= block_size_*2; x++) {
+        if (i == 0) {
+          above_row_[x] = mask_;
+        } else {
+          above_row_[x] = rnd.Rand16() & mask_;
+        }
+      }
+      for (int y = 0; y < block_size_; y++) {
+        if (i == 0) {
+          left_col_[y] = mask_;
+        } else {
+          left_col_[y] = rnd.Rand16() & mask_;
+        }
+      }
+      Predict();
+      CheckPrediction(i, &error_count);
+    }
+    ASSERT_EQ(0, error_count);
+  }
+
+  int block_size_;
+  uint16_t *above_row_;
+  uint16_t *left_col_;
+  uint16_t *dst_;
+  uint16_t *ref_dst_;
+  ptrdiff_t stride_;
+  int mask_;
+};
+
+typedef void (*intra_pred_fn_t)(
+      uint16_t *dst, ptrdiff_t stride, const uint16_t *above,
+      const uint16_t *left, int bps);
+typedef std::tr1::tuple<intra_pred_fn_t,
+                        intra_pred_fn_t, int, int> intra_pred_params_t;
+class VP9IntraPredTest
+    : public VP9IntraPredBase,
+      public ::testing::TestWithParam<intra_pred_params_t> {
+
+  virtual void SetUp() {
+    pred_fn_    = GET_PARAM(0);
+    ref_fn_     = GET_PARAM(1);
+    block_size_ = GET_PARAM(2);
+    bit_depth_  = GET_PARAM(3);
+    stride_     = block_size_ * 3;
+    mask_       = (1 << bit_depth_) - 1;
+  }
+
+  virtual void Predict() {
+    const uint16_t *const_above_row = above_row_;
+    const uint16_t *const_left_col = left_col_;
+    ref_fn_(ref_dst_, stride_, const_above_row, const_left_col, bit_depth_);
+    ASM_REGISTER_STATE_CHECK(pred_fn_(dst_, stride_, const_above_row,
+                                      const_left_col, bit_depth_));
+  }
+  intra_pred_fn_t pred_fn_;
+  intra_pred_fn_t ref_fn_;
+  int bit_depth_;
+};
+
+TEST_P(VP9IntraPredTest, IntraPredTests) {
+  // max block size is 32
+  DECLARE_ALIGNED(16, uint16_t, left_col[2*32]);
+  DECLARE_ALIGNED(16, uint16_t, above_data[2*32+32]);
+  DECLARE_ALIGNED(16, uint16_t, dst[3 * 32 * 32]);
+  DECLARE_ALIGNED(16, uint16_t, ref_dst[3 * 32 * 32]);
+  RunTest(left_col, above_data, dst, ref_dst);
+}
+
+using std::tr1::make_tuple;
+
+#if HAVE_SSE2
+#if CONFIG_VP9_HIGHBITDEPTH
+#if CONFIG_USE_X86INC
+INSTANTIATE_TEST_CASE_P(SSE2_TO_C_8, VP9IntraPredTest,
+                        ::testing::Values(
+                            make_tuple(&vpx_highbd_dc_predictor_32x32_sse2,
+                                       &vpx_highbd_dc_predictor_32x32_c, 32, 8),
+                            make_tuple(&vpx_highbd_tm_predictor_16x16_sse2,
+                                       &vpx_highbd_tm_predictor_16x16_c, 16, 8),
+                            make_tuple(&vpx_highbd_tm_predictor_32x32_sse2,
+                                       &vpx_highbd_tm_predictor_32x32_c, 32, 8),
+                            make_tuple(&vpx_highbd_dc_predictor_4x4_sse2,
+                                       &vpx_highbd_dc_predictor_4x4_c, 4, 8),
+                            make_tuple(&vpx_highbd_dc_predictor_8x8_sse2,
+                                       &vpx_highbd_dc_predictor_8x8_c, 8, 8),
+                            make_tuple(&vpx_highbd_dc_predictor_16x16_sse2,
+                                       &vpx_highbd_dc_predictor_16x16_c, 16, 8),
+                            make_tuple(&vpx_highbd_v_predictor_4x4_sse2,
+                                       &vpx_highbd_v_predictor_4x4_c, 4, 8),
+                            make_tuple(&vpx_highbd_v_predictor_8x8_sse2,
+                                       &vpx_highbd_v_predictor_8x8_c, 8, 8),
+                            make_tuple(&vpx_highbd_v_predictor_16x16_sse2,
+                                       &vpx_highbd_v_predictor_16x16_c, 16, 8),
+                            make_tuple(&vpx_highbd_v_predictor_32x32_sse2,
+                                       &vpx_highbd_v_predictor_32x32_c, 32, 8),
+                            make_tuple(&vpx_highbd_tm_predictor_4x4_sse2,
+                                       &vpx_highbd_tm_predictor_4x4_c, 4, 8),
+                            make_tuple(&vpx_highbd_tm_predictor_8x8_sse2,
+                                       &vpx_highbd_tm_predictor_8x8_c, 8, 8)));
+
+INSTANTIATE_TEST_CASE_P(SSE2_TO_C_10, VP9IntraPredTest,
+                        ::testing::Values(
+                            make_tuple(&vpx_highbd_dc_predictor_32x32_sse2,
+                                       &vpx_highbd_dc_predictor_32x32_c, 32,
+                                       10),
+                            make_tuple(&vpx_highbd_tm_predictor_16x16_sse2,
+                                       &vpx_highbd_tm_predictor_16x16_c, 16,
+                                       10),
+                            make_tuple(&vpx_highbd_tm_predictor_32x32_sse2,
+                                       &vpx_highbd_tm_predictor_32x32_c, 32,
+                                       10),
+                            make_tuple(&vpx_highbd_dc_predictor_4x4_sse2,
+                                       &vpx_highbd_dc_predictor_4x4_c, 4, 10),
+                            make_tuple(&vpx_highbd_dc_predictor_8x8_sse2,
+                                       &vpx_highbd_dc_predictor_8x8_c, 8, 10),
+                            make_tuple(&vpx_highbd_dc_predictor_16x16_sse2,
+                                       &vpx_highbd_dc_predictor_16x16_c, 16,
+                                       10),
+                            make_tuple(&vpx_highbd_v_predictor_4x4_sse2,
+                                       &vpx_highbd_v_predictor_4x4_c, 4, 10),
+                            make_tuple(&vpx_highbd_v_predictor_8x8_sse2,
+                                       &vpx_highbd_v_predictor_8x8_c, 8, 10),
+                            make_tuple(&vpx_highbd_v_predictor_16x16_sse2,
+                                       &vpx_highbd_v_predictor_16x16_c, 16,
+                                       10),
+                            make_tuple(&vpx_highbd_v_predictor_32x32_sse2,
+                                       &vpx_highbd_v_predictor_32x32_c, 32,
+                                       10),
+                            make_tuple(&vpx_highbd_tm_predictor_4x4_sse2,
+                                       &vpx_highbd_tm_predictor_4x4_c, 4, 10),
+                            make_tuple(&vpx_highbd_tm_predictor_8x8_sse2,
+                                       &vpx_highbd_tm_predictor_8x8_c, 8, 10)));
+
+INSTANTIATE_TEST_CASE_P(SSE2_TO_C_12, VP9IntraPredTest,
+                        ::testing::Values(
+                            make_tuple(&vpx_highbd_dc_predictor_32x32_sse2,
+                                       &vpx_highbd_dc_predictor_32x32_c, 32,
+                                       12),
+                            make_tuple(&vpx_highbd_tm_predictor_16x16_sse2,
+                                       &vpx_highbd_tm_predictor_16x16_c, 16,
+                                       12),
+                            make_tuple(&vpx_highbd_tm_predictor_32x32_sse2,
+                                       &vpx_highbd_tm_predictor_32x32_c, 32,
+                                       12),
+                            make_tuple(&vpx_highbd_dc_predictor_4x4_sse2,
+                                       &vpx_highbd_dc_predictor_4x4_c, 4, 12),
+                            make_tuple(&vpx_highbd_dc_predictor_8x8_sse2,
+                                       &vpx_highbd_dc_predictor_8x8_c, 8, 12),
+                            make_tuple(&vpx_highbd_dc_predictor_16x16_sse2,
+                                       &vpx_highbd_dc_predictor_16x16_c, 16,
+                                       12),
+                            make_tuple(&vpx_highbd_v_predictor_4x4_sse2,
+                                       &vpx_highbd_v_predictor_4x4_c, 4, 12),
+                            make_tuple(&vpx_highbd_v_predictor_8x8_sse2,
+                                       &vpx_highbd_v_predictor_8x8_c, 8, 12),
+                            make_tuple(&vpx_highbd_v_predictor_16x16_sse2,
+                                       &vpx_highbd_v_predictor_16x16_c, 16,
+                                       12),
+                            make_tuple(&vpx_highbd_v_predictor_32x32_sse2,
+                                       &vpx_highbd_v_predictor_32x32_c, 32,
+                                       12),
+                            make_tuple(&vpx_highbd_tm_predictor_4x4_sse2,
+                                       &vpx_highbd_tm_predictor_4x4_c, 4, 12),
+                            make_tuple(&vpx_highbd_tm_predictor_8x8_sse2,
+                                       &vpx_highbd_tm_predictor_8x8_c, 8, 12)));
+
+#endif  // CONFIG_USE_X86INC
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#endif  // HAVE_SSE2
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_lossless_test.cc b/libvpx/libvpx/test/vp9_lossless_test.cc
new file mode 100644
index 0000000..4177393
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_lossless_test.cc
@@ -0,0 +1,130 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+
+namespace {
+
+const int kMaxPsnr = 100;
+
+class LosslessTest : public ::libvpx_test::EncoderTest,
+    public ::libvpx_test::CodecTestWithParam<libvpx_test::TestMode> {
+ protected:
+  LosslessTest()
+      : EncoderTest(GET_PARAM(0)),
+        psnr_(kMaxPsnr),
+        nframes_(0),
+        encoding_mode_(GET_PARAM(1)) {
+  }
+
+  virtual ~LosslessTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      // Only call Control if quantizer > 0 to verify that using quantizer
+      // alone will activate lossless
+      if (cfg_.rc_max_quantizer > 0 || cfg_.rc_min_quantizer > 0) {
+        encoder->Control(VP9E_SET_LOSSLESS, 1);
+      }
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    psnr_ = kMaxPsnr;
+    nframes_ = 0;
+  }
+
+  virtual void PSNRPktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->data.psnr.psnr[0] < psnr_)
+      psnr_= pkt->data.psnr.psnr[0];
+  }
+
+  double GetMinPsnr() const {
+      return psnr_;
+  }
+
+ private:
+  double psnr_;
+  unsigned int nframes_;
+  libvpx_test::TestMode encoding_mode_;
+};
+
+TEST_P(LosslessTest, TestLossLessEncoding) {
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.g_lag_in_frames = 25;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 0;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  // intentionally changed the dimension for better testing coverage
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 10);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_lossless = GetMinPsnr();
+  EXPECT_GE(psnr_lossless, kMaxPsnr);
+}
+
+TEST_P(LosslessTest, TestLossLessEncoding444) {
+  libvpx_test::Y4mVideoSource video("rush_hour_444.y4m", 0, 10);
+
+  cfg_.g_profile = 1;
+  cfg_.g_timebase = video.timebase();
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.g_lag_in_frames = 25;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 0;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_lossless = GetMinPsnr();
+  EXPECT_GE(psnr_lossless, kMaxPsnr);
+}
+
+TEST_P(LosslessTest, TestLossLessEncodingCtrl) {
+  const vpx_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.g_lag_in_frames = 25;
+  // Intentionally set Q > 0, to make sure control can be used to activate
+  // lossless
+  cfg_.rc_min_quantizer = 10;
+  cfg_.rc_max_quantizer = 20;
+
+  init_flags_ = VPX_CODEC_USE_PSNR;
+
+  libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 10);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_lossless = GetMinPsnr();
+  EXPECT_GE(psnr_lossless, kMaxPsnr);
+}
+
+VP9_INSTANTIATE_TEST_CASE(LosslessTest,
+                          ::testing::Values(::libvpx_test::kRealTime,
+                                            ::libvpx_test::kOnePassGood,
+                                            ::libvpx_test::kTwoPassGood));
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_quantize_test.cc b/libvpx/libvpx/test/vp9_quantize_test.cc
new file mode 100644
index 0000000..81d31fd
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_quantize_test.cc
@@ -0,0 +1,351 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_scan.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_integer.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+#if CONFIG_VP9_HIGHBITDEPTH
+const int number_of_iterations = 100;
+
+typedef void (*QuantizeFunc)(const tran_low_t *coeff, intptr_t count,
+                             int skip_block, const int16_t *zbin,
+                             const int16_t *round, const int16_t *quant,
+                             const int16_t *quant_shift,
+                             tran_low_t *qcoeff, tran_low_t *dqcoeff,
+                             const int16_t *dequant,
+                             uint16_t *eob, const int16_t *scan,
+                             const int16_t *iscan);
+typedef std::tr1::tuple<QuantizeFunc, QuantizeFunc, vpx_bit_depth_t>
+    QuantizeParam;
+
+class VP9QuantizeTest : public ::testing::TestWithParam<QuantizeParam> {
+ public:
+  virtual ~VP9QuantizeTest() {}
+  virtual void SetUp() {
+    quantize_op_   = GET_PARAM(0);
+    ref_quantize_op_ = GET_PARAM(1);
+    bit_depth_  = GET_PARAM(2);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  vpx_bit_depth_t bit_depth_;
+  int mask_;
+  QuantizeFunc quantize_op_;
+  QuantizeFunc ref_quantize_op_;
+};
+
+class VP9Quantize32Test : public ::testing::TestWithParam<QuantizeParam> {
+ public:
+  virtual ~VP9Quantize32Test() {}
+  virtual void SetUp() {
+    quantize_op_   = GET_PARAM(0);
+    ref_quantize_op_ = GET_PARAM(1);
+    bit_depth_  = GET_PARAM(2);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  vpx_bit_depth_t bit_depth_;
+  int mask_;
+  QuantizeFunc quantize_op_;
+  QuantizeFunc ref_quantize_op_;
+};
+
+TEST_P(VP9QuantizeTest, OperationCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, tran_low_t, coeff_ptr[256]);
+  DECLARE_ALIGNED(16, int16_t, zbin_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, round_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, quant_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, quant_shift_ptr[2]);
+  DECLARE_ALIGNED(16, tran_low_t, qcoeff_ptr[256]);
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff_ptr[256]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_qcoeff_ptr[256]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_dqcoeff_ptr[256]);
+  DECLARE_ALIGNED(16, int16_t, dequant_ptr[2]);
+  DECLARE_ALIGNED(16, uint16_t, eob_ptr[1]);
+  DECLARE_ALIGNED(16, uint16_t, ref_eob_ptr[1]);
+  int err_count_total = 0;
+  int first_failure = -1;
+  for (int i = 0; i < number_of_iterations; ++i) {
+    const int skip_block = i == 0;
+    const TX_SIZE sz = (TX_SIZE)(i % 3);  // TX_4X4, TX_8X8 TX_16X16
+    const TX_TYPE tx_type = (TX_TYPE)((i >> 2) % 3);
+    const scan_order *scan_order = &vp9_scan_orders[sz][tx_type];
+    const int count = (4 << sz) * (4 << sz);  // 16, 64, 256
+    int err_count = 0;
+    *eob_ptr = rnd.Rand16();
+    *ref_eob_ptr = *eob_ptr;
+    for (int j = 0; j < count; j++) {
+      coeff_ptr[j] = rnd.Rand16()&mask_;
+    }
+    for (int j = 0; j < 2; j++) {
+      zbin_ptr[j] = rnd.Rand16()&mask_;
+      round_ptr[j] = rnd.Rand16();
+      quant_ptr[j] = rnd.Rand16();
+      quant_shift_ptr[j] = rnd.Rand16();
+      dequant_ptr[j] = rnd.Rand16();
+    }
+    ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr,
+                     quant_ptr, quant_shift_ptr, ref_qcoeff_ptr,
+                     ref_dqcoeff_ptr, dequant_ptr,
+                     ref_eob_ptr, scan_order->scan, scan_order->iscan);
+    ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block,
+                                          zbin_ptr, round_ptr, quant_ptr,
+                                          quant_shift_ptr, qcoeff_ptr,
+                                          dqcoeff_ptr, dequant_ptr, eob_ptr,
+                                          scan_order->scan, scan_order->iscan));
+    for (int j = 0; j < sz; ++j) {
+      err_count += (ref_qcoeff_ptr[j]  != qcoeff_ptr[j]) |
+          (ref_dqcoeff_ptr[j] != dqcoeff_ptr[j]);
+    }
+    err_count += (*ref_eob_ptr != *eob_ptr);
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Quantization Test, C output doesn't match SSE2 output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(VP9Quantize32Test, OperationCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, tran_low_t, coeff_ptr[1024]);
+  DECLARE_ALIGNED(16, int16_t, zbin_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, round_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, quant_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, quant_shift_ptr[2]);
+  DECLARE_ALIGNED(16, tran_low_t, qcoeff_ptr[1024]);
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff_ptr[1024]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_qcoeff_ptr[1024]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_dqcoeff_ptr[1024]);
+  DECLARE_ALIGNED(16, int16_t, dequant_ptr[2]);
+  DECLARE_ALIGNED(16, uint16_t, eob_ptr[1]);
+  DECLARE_ALIGNED(16, uint16_t, ref_eob_ptr[1]);
+  int err_count_total = 0;
+  int first_failure = -1;
+  for (int i = 0; i < number_of_iterations; ++i) {
+    const int skip_block = i == 0;
+    const TX_SIZE sz = TX_32X32;
+    const TX_TYPE tx_type = (TX_TYPE)(i % 4);
+    const scan_order *scan_order = &vp9_scan_orders[sz][tx_type];
+    const int count = (4 << sz) * (4 << sz);  // 1024
+    int err_count = 0;
+    *eob_ptr = rnd.Rand16();
+    *ref_eob_ptr = *eob_ptr;
+    for (int j = 0; j < count; j++) {
+      coeff_ptr[j] = rnd.Rand16()&mask_;
+    }
+    for (int j = 0; j < 2; j++) {
+      zbin_ptr[j] = rnd.Rand16()&mask_;
+      round_ptr[j] = rnd.Rand16();
+      quant_ptr[j] = rnd.Rand16();
+      quant_shift_ptr[j] = rnd.Rand16();
+      dequant_ptr[j] = rnd.Rand16();
+    }
+    ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr,
+                     quant_ptr, quant_shift_ptr, ref_qcoeff_ptr,
+                     ref_dqcoeff_ptr, dequant_ptr,
+                     ref_eob_ptr, scan_order->scan, scan_order->iscan);
+    ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block,
+                                          zbin_ptr, round_ptr, quant_ptr,
+                                          quant_shift_ptr, qcoeff_ptr,
+                                          dqcoeff_ptr, dequant_ptr, eob_ptr,
+                                          scan_order->scan, scan_order->iscan));
+    for (int j = 0; j < sz; ++j) {
+      err_count += (ref_qcoeff_ptr[j]  != qcoeff_ptr[j]) |
+          (ref_dqcoeff_ptr[j] != dqcoeff_ptr[j]);
+    }
+    err_count += (*ref_eob_ptr != *eob_ptr);
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Quantization Test, C output doesn't match SSE2 output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(VP9QuantizeTest, EOBCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, tran_low_t, coeff_ptr[256]);
+  DECLARE_ALIGNED(16, int16_t, zbin_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, round_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, quant_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, quant_shift_ptr[2]);
+  DECLARE_ALIGNED(16, tran_low_t, qcoeff_ptr[256]);
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff_ptr[256]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_qcoeff_ptr[256]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_dqcoeff_ptr[256]);
+  DECLARE_ALIGNED(16, int16_t, dequant_ptr[2]);
+  DECLARE_ALIGNED(16, uint16_t, eob_ptr[1]);
+  DECLARE_ALIGNED(16, uint16_t, ref_eob_ptr[1]);
+  int err_count_total = 0;
+  int first_failure = -1;
+  for (int i = 0; i < number_of_iterations; ++i) {
+    int skip_block = i == 0;
+    TX_SIZE sz = (TX_SIZE)(i % 3);  // TX_4X4, TX_8X8 TX_16X16
+    TX_TYPE tx_type = (TX_TYPE)((i >> 2) % 3);
+    const scan_order *scan_order = &vp9_scan_orders[sz][tx_type];
+    int count = (4 << sz) * (4 << sz);  // 16, 64, 256
+    int err_count = 0;
+    *eob_ptr = rnd.Rand16();
+    *ref_eob_ptr = *eob_ptr;
+    // Two random entries
+    for (int j = 0; j < count; j++) {
+      coeff_ptr[j] = 0;
+    }
+    coeff_ptr[rnd(count)] = rnd.Rand16()&mask_;
+    coeff_ptr[rnd(count)] = rnd.Rand16()&mask_;
+    for (int j = 0; j < 2; j++) {
+      zbin_ptr[j] = rnd.Rand16()&mask_;
+      round_ptr[j] = rnd.Rand16();
+      quant_ptr[j] = rnd.Rand16();
+      quant_shift_ptr[j] = rnd.Rand16();
+      dequant_ptr[j] = rnd.Rand16();
+    }
+
+    ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr,
+                     quant_ptr, quant_shift_ptr, ref_qcoeff_ptr,
+                     ref_dqcoeff_ptr, dequant_ptr,
+                     ref_eob_ptr, scan_order->scan, scan_order->iscan);
+    ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block,
+                                          zbin_ptr, round_ptr, quant_ptr,
+                                          quant_shift_ptr, qcoeff_ptr,
+                                          dqcoeff_ptr, dequant_ptr, eob_ptr,
+                                          scan_order->scan, scan_order->iscan));
+
+    for (int j = 0; j < sz; ++j) {
+      err_count += (ref_qcoeff_ptr[j]  != qcoeff_ptr[j]) |
+          (ref_dqcoeff_ptr[j] != dqcoeff_ptr[j]);
+    }
+    err_count += (*ref_eob_ptr != *eob_ptr);
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Quantization Test, C output doesn't match SSE2 output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(VP9Quantize32Test, EOBCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, tran_low_t, coeff_ptr[1024]);
+  DECLARE_ALIGNED(16, int16_t, zbin_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, round_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, quant_ptr[2]);
+  DECLARE_ALIGNED(16, int16_t, quant_shift_ptr[2]);
+  DECLARE_ALIGNED(16, tran_low_t, qcoeff_ptr[1024]);
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff_ptr[1024]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_qcoeff_ptr[1024]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_dqcoeff_ptr[1024]);
+  DECLARE_ALIGNED(16, int16_t, dequant_ptr[2]);
+  DECLARE_ALIGNED(16, uint16_t, eob_ptr[1]);
+  DECLARE_ALIGNED(16, uint16_t, ref_eob_ptr[1]);
+  int err_count_total = 0;
+  int first_failure = -1;
+  for (int i = 0; i < number_of_iterations; ++i) {
+    int skip_block = i == 0;
+    TX_SIZE sz = TX_32X32;
+    TX_TYPE tx_type = (TX_TYPE)(i % 4);
+    const scan_order *scan_order = &vp9_scan_orders[sz][tx_type];
+    int count = (4 << sz) * (4 << sz);  // 1024
+    int err_count = 0;
+    *eob_ptr = rnd.Rand16();
+    *ref_eob_ptr = *eob_ptr;
+    for (int j = 0; j < count; j++) {
+      coeff_ptr[j] = 0;
+    }
+    // Two random entries
+    coeff_ptr[rnd(count)] = rnd.Rand16()&mask_;
+    coeff_ptr[rnd(count)] = rnd.Rand16()&mask_;
+    for (int j = 0; j < 2; j++) {
+      zbin_ptr[j] = rnd.Rand16()&mask_;
+      round_ptr[j] = rnd.Rand16();
+      quant_ptr[j] = rnd.Rand16();
+      quant_shift_ptr[j] = rnd.Rand16();
+      dequant_ptr[j] = rnd.Rand16();
+    }
+
+    ref_quantize_op_(coeff_ptr, count, skip_block, zbin_ptr, round_ptr,
+                     quant_ptr, quant_shift_ptr, ref_qcoeff_ptr,
+                     ref_dqcoeff_ptr, dequant_ptr,
+                     ref_eob_ptr, scan_order->scan, scan_order->iscan);
+    ASM_REGISTER_STATE_CHECK(quantize_op_(coeff_ptr, count, skip_block,
+                                          zbin_ptr, round_ptr, quant_ptr,
+                                          quant_shift_ptr, qcoeff_ptr,
+                                          dqcoeff_ptr, dequant_ptr, eob_ptr,
+                                          scan_order->scan, scan_order->iscan));
+
+    for (int j = 0; j < sz; ++j) {
+      err_count += (ref_qcoeff_ptr[j]  != qcoeff_ptr[j]) |
+          (ref_dqcoeff_ptr[j] != dqcoeff_ptr[j]);
+    }
+    err_count += (*ref_eob_ptr != *eob_ptr);
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Quantization Test, C output doesn't match SSE2 output. "
+      << "First failed at test case " << first_failure;
+}
+using std::tr1::make_tuple;
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VP9QuantizeTest,
+    ::testing::Values(
+        make_tuple(&vpx_highbd_quantize_b_sse2,
+                   &vpx_highbd_quantize_b_c, VPX_BITS_8),
+        make_tuple(&vpx_highbd_quantize_b_sse2,
+                   &vpx_highbd_quantize_b_c, VPX_BITS_10),
+        make_tuple(&vpx_highbd_quantize_b_sse2,
+                   &vpx_highbd_quantize_b_c, VPX_BITS_12)));
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VP9Quantize32Test,
+    ::testing::Values(
+        make_tuple(&vpx_highbd_quantize_b_32x32_sse2,
+                   &vpx_highbd_quantize_b_32x32_c, VPX_BITS_8),
+        make_tuple(&vpx_highbd_quantize_b_32x32_sse2,
+                   &vpx_highbd_quantize_b_32x32_c, VPX_BITS_10),
+        make_tuple(&vpx_highbd_quantize_b_32x32_sse2,
+                   &vpx_highbd_quantize_b_32x32_c, VPX_BITS_12)));
+#endif  // HAVE_SSE2
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_skip_loopfilter_test.cc b/libvpx/libvpx/test/vp9_skip_loopfilter_test.cc
new file mode 100644
index 0000000..b0cc7ba
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_skip_loopfilter_test.cc
@@ -0,0 +1,180 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#include "test/webm_video_source.h"
+
+namespace {
+
+const char kVp9TestFile[] = "vp90-2-08-tile_1x8_frame_parallel.webm";
+const char kVp9Md5File[] = "vp90-2-08-tile_1x8_frame_parallel.webm.md5";
+
+// Class for testing shutting off the loop filter.
+class SkipLoopFilterTest {
+ public:
+  SkipLoopFilterTest()
+      : video_(NULL),
+        decoder_(NULL),
+        md5_file_(NULL) {}
+
+  ~SkipLoopFilterTest() {
+    if (md5_file_ != NULL)
+      fclose(md5_file_);
+    delete decoder_;
+    delete video_;
+  }
+
+  // If |threads| > 0 then set the decoder with that number of threads.
+  void Init(int num_threads) {
+    expected_md5_[0] = '\0';
+    junk_[0] = '\0';
+    video_ = new libvpx_test::WebMVideoSource(kVp9TestFile);
+    ASSERT_TRUE(video_ != NULL);
+    video_->Init();
+    video_->Begin();
+
+    vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+    if (num_threads > 0)
+      cfg.threads = num_threads;
+    decoder_ = new libvpx_test::VP9Decoder(cfg, 0);
+    ASSERT_TRUE(decoder_ != NULL);
+
+    OpenMd5File(kVp9Md5File);
+  }
+
+  // Set the VP9 skipLoopFilter control value.
+  void SetSkipLoopFilter(int value, vpx_codec_err_t expected_value) {
+    decoder_->Control(VP9_SET_SKIP_LOOP_FILTER, value, expected_value);
+  }
+
+  vpx_codec_err_t DecodeOneFrame() {
+    const vpx_codec_err_t res =
+        decoder_->DecodeFrame(video_->cxdata(), video_->frame_size());
+    if (res == VPX_CODEC_OK) {
+      ReadMd5();
+      video_->Next();
+    }
+    return res;
+  }
+
+  vpx_codec_err_t DecodeRemainingFrames() {
+    for (; video_->cxdata() != NULL; video_->Next()) {
+      const vpx_codec_err_t res =
+          decoder_->DecodeFrame(video_->cxdata(), video_->frame_size());
+      if (res != VPX_CODEC_OK)
+        return res;
+      ReadMd5();
+    }
+    return VPX_CODEC_OK;
+  }
+
+  // Checks if MD5 matches or doesn't.
+  void CheckMd5(bool matches) {
+    libvpx_test::DxDataIterator dec_iter = decoder_->GetDxData();
+    const vpx_image_t *img = dec_iter.Next();
+    CheckMd5Vpx(*img, matches);
+  }
+
+ private:
+  // TODO(fgalligan): Move the MD5 testing code into another class.
+  void OpenMd5File(const std::string &md5_file_name) {
+    md5_file_ = libvpx_test::OpenTestDataFile(md5_file_name);
+    ASSERT_TRUE(md5_file_ != NULL) << "MD5 file open failed. Filename: "
+        << md5_file_name;
+  }
+
+  // Reads the next line of the MD5 file.
+  void ReadMd5() {
+    ASSERT_TRUE(md5_file_ != NULL);
+    const int res = fscanf(md5_file_, "%s  %s", expected_md5_, junk_);
+    ASSERT_NE(EOF, res) << "Read md5 data failed";
+    expected_md5_[32] = '\0';
+  }
+
+  // Checks if the last read MD5 matches |img| or doesn't.
+  void CheckMd5Vpx(const vpx_image_t &img, bool matches) {
+    ::libvpx_test::MD5 md5_res;
+    md5_res.Add(&img);
+    const char *const actual_md5 = md5_res.Get();
+
+    // Check MD5.
+    if (matches)
+      ASSERT_STREQ(expected_md5_, actual_md5) << "MD5 checksums don't match";
+    else
+      ASSERT_STRNE(expected_md5_, actual_md5) << "MD5 checksums match";
+  }
+
+  libvpx_test::WebMVideoSource *video_;
+  libvpx_test::VP9Decoder *decoder_;
+  FILE *md5_file_;
+  char expected_md5_[33];
+  char junk_[128];
+};
+
+TEST(SkipLoopFilterTest, ShutOffLoopFilter) {
+  const int non_zero_value = 1;
+  const int num_threads = 0;
+  SkipLoopFilterTest skip_loop_filter;
+  skip_loop_filter.Init(num_threads);
+  skip_loop_filter.SetSkipLoopFilter(non_zero_value, VPX_CODEC_OK);
+  ASSERT_EQ(VPX_CODEC_OK, skip_loop_filter.DecodeRemainingFrames());
+  skip_loop_filter.CheckMd5(false);
+}
+
+TEST(SkipLoopFilterTest, ShutOffLoopFilterSingleThread) {
+  const int non_zero_value = 1;
+  const int num_threads = 1;
+  SkipLoopFilterTest skip_loop_filter;
+  skip_loop_filter.Init(num_threads);
+  skip_loop_filter.SetSkipLoopFilter(non_zero_value, VPX_CODEC_OK);
+  ASSERT_EQ(VPX_CODEC_OK, skip_loop_filter.DecodeRemainingFrames());
+  skip_loop_filter.CheckMd5(false);
+}
+
+TEST(SkipLoopFilterTest, ShutOffLoopFilter8Threads) {
+  const int non_zero_value = 1;
+  const int num_threads = 8;
+  SkipLoopFilterTest skip_loop_filter;
+  skip_loop_filter.Init(num_threads);
+  skip_loop_filter.SetSkipLoopFilter(non_zero_value, VPX_CODEC_OK);
+  ASSERT_EQ(VPX_CODEC_OK, skip_loop_filter.DecodeRemainingFrames());
+  skip_loop_filter.CheckMd5(false);
+}
+
+TEST(SkipLoopFilterTest, WithLoopFilter) {
+  const int non_zero_value = 1;
+  const int num_threads = 0;
+  SkipLoopFilterTest skip_loop_filter;
+  skip_loop_filter.Init(num_threads);
+  skip_loop_filter.SetSkipLoopFilter(non_zero_value, VPX_CODEC_OK);
+  skip_loop_filter.SetSkipLoopFilter(0, VPX_CODEC_OK);
+  ASSERT_EQ(VPX_CODEC_OK, skip_loop_filter.DecodeRemainingFrames());
+  skip_loop_filter.CheckMd5(true);
+}
+
+TEST(SkipLoopFilterTest, ToggleLoopFilter) {
+  const int num_threads = 0;
+  SkipLoopFilterTest skip_loop_filter;
+  skip_loop_filter.Init(num_threads);
+
+  for (int i = 0; i < 10; ++i) {
+    skip_loop_filter.SetSkipLoopFilter(i % 2, VPX_CODEC_OK);
+    ASSERT_EQ(VPX_CODEC_OK, skip_loop_filter.DecodeOneFrame());
+  }
+  ASSERT_EQ(VPX_CODEC_OK, skip_loop_filter.DecodeRemainingFrames());
+  skip_loop_filter.CheckMd5(false);
+}
+
+}  // namespace
diff --git a/libvpx/libvpx/test/vp9_spatial_svc_encoder.sh b/libvpx/libvpx/test/vp9_spatial_svc_encoder.sh
new file mode 100755
index 0000000..6503107
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_spatial_svc_encoder.sh
@@ -0,0 +1,72 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx vp9_spatial_svc_encoder example. To add new
+##  tests to to this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vp9_spatial_svc_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+vp9_spatial_svc_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs vp9_spatial_svc_encoder. $1 is the test name.
+vp9_spatial_svc_encoder() {
+  local readonly \
+    encoder="${LIBVPX_BIN_PATH}/vp9_spatial_svc_encoder${VPX_TEST_EXE_SUFFIX}"
+  local readonly test_name="$1"
+  local readonly \
+    output_file="${VPX_TEST_OUTPUT_DIR}/vp9_ssvc_encoder${test_name}.ivf"
+  local readonly frames_to_encode=10
+  local readonly max_kf=9999
+
+  shift
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" -w "${YUV_RAW_INPUT_WIDTH}" \
+    -h "${YUV_RAW_INPUT_HEIGHT}" -k "${max_kf}" -f "${frames_to_encode}" \
+    "$@" "${YUV_RAW_INPUT}" "${output_file}" ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+# Each test is run with layer count 1-$vp9_ssvc_test_layers.
+vp9_ssvc_test_layers=5
+
+vp9_spatial_svc() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    local readonly test_name="vp9_spatial_svc"
+    for layers in $(seq 1 ${vp9_ssvc_test_layers}); do
+      vp9_spatial_svc_encoder "${test_name}" -sl ${layers}
+    done
+  fi
+}
+
+readonly vp9_spatial_svc_tests="DISABLED_vp9_spatial_svc_mode_i
+                                DISABLED_vp9_spatial_svc_mode_altip
+                                DISABLED_vp9_spatial_svc_mode_ip
+                                DISABLED_vp9_spatial_svc_mode_gf
+                                vp9_spatial_svc"
+
+if [ "$(vpx_config_option_enabled CONFIG_SPATIAL_SVC)" = "yes" ]; then
+  run_tests \
+    vp9_spatial_svc_encoder_verify_environment \
+    "${vp9_spatial_svc_tests}"
+fi
diff --git a/libvpx/libvpx/test/vp9_subtract_test.cc b/libvpx/libvpx/test/vp9_subtract_test.cc
new file mode 100644
index 0000000..3cad4d7
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_subtract_test.cc
@@ -0,0 +1,109 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vpx_mem/vpx_mem.h"
+
+typedef void (*SubtractFunc)(int rows, int cols,
+                             int16_t *diff_ptr, ptrdiff_t diff_stride,
+                             const uint8_t *src_ptr, ptrdiff_t src_stride,
+                             const uint8_t *pred_ptr, ptrdiff_t pred_stride);
+
+namespace vp9 {
+
+class VP9SubtractBlockTest : public ::testing::TestWithParam<SubtractFunc> {
+ public:
+  virtual void TearDown() {
+    libvpx_test::ClearSystemState();
+  }
+};
+
+using libvpx_test::ACMRandom;
+
+TEST_P(VP9SubtractBlockTest, SimpleSubtract) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+  // FIXME(rbultje) split in its own file
+  for (BLOCK_SIZE bsize = BLOCK_4X4; bsize < BLOCK_SIZES;
+       bsize = static_cast<BLOCK_SIZE>(static_cast<int>(bsize) + 1)) {
+    const int block_width = 4 * num_4x4_blocks_wide_lookup[bsize];
+    const int block_height = 4 * num_4x4_blocks_high_lookup[bsize];
+    int16_t *diff = reinterpret_cast<int16_t *>(
+        vpx_memalign(16, sizeof(*diff) * block_width * block_height * 2));
+    uint8_t *pred = reinterpret_cast<uint8_t *>(
+        vpx_memalign(16, block_width * block_height * 2));
+    uint8_t *src  = reinterpret_cast<uint8_t *>(
+        vpx_memalign(16, block_width * block_height * 2));
+
+    for (int n = 0; n < 100; n++) {
+      for (int r = 0; r < block_height; ++r) {
+        for (int c = 0; c < block_width * 2; ++c) {
+          src[r * block_width * 2 + c] = rnd.Rand8();
+          pred[r * block_width * 2 + c] = rnd.Rand8();
+        }
+      }
+
+      GetParam()(block_height, block_width, diff, block_width,
+                 src, block_width, pred, block_width);
+
+      for (int r = 0; r < block_height; ++r) {
+        for (int c = 0; c < block_width; ++c) {
+          EXPECT_EQ(diff[r * block_width + c],
+                    (src[r * block_width + c] -
+                     pred[r * block_width + c])) << "r = " << r
+                                                 << ", c = " << c
+                                                 << ", bs = " << bsize;
+        }
+      }
+
+      GetParam()(block_height, block_width, diff, block_width * 2,
+                 src, block_width * 2, pred, block_width * 2);
+
+      for (int r = 0; r < block_height; ++r) {
+        for (int c = 0; c < block_width; ++c) {
+          EXPECT_EQ(diff[r * block_width * 2 + c],
+                    (src[r * block_width * 2 + c] -
+                     pred[r * block_width * 2 + c])) << "r = " << r
+                                                     << ", c = " << c
+                                                     << ", bs = " << bsize;
+        }
+      }
+    }
+    vpx_free(diff);
+    vpx_free(pred);
+    vpx_free(src);
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(C, VP9SubtractBlockTest,
+                        ::testing::Values(vpx_subtract_block_c));
+
+#if HAVE_SSE2 && CONFIG_USE_X86INC
+INSTANTIATE_TEST_CASE_P(SSE2, VP9SubtractBlockTest,
+                        ::testing::Values(vpx_subtract_block_sse2));
+#endif
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, VP9SubtractBlockTest,
+                        ::testing::Values(vpx_subtract_block_neon));
+#endif
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(MSA, VP9SubtractBlockTest,
+                        ::testing::Values(vpx_subtract_block_msa));
+#endif
+
+}  // namespace vp9
diff --git a/libvpx/libvpx/test/vp9_thread_test.cc b/libvpx/libvpx/test/vp9_thread_test.cc
new file mode 100644
index 0000000..92e4b96
--- /dev/null
+++ b/libvpx/libvpx/test/vp9_thread_test.cc
@@ -0,0 +1,326 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/md5_helper.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "vpx_util/vpx_thread.h"
+
+namespace {
+
+using std::string;
+
+class VPxWorkerThreadTest : public ::testing::TestWithParam<bool> {
+ protected:
+  virtual ~VPxWorkerThreadTest() {}
+  virtual void SetUp() {
+    vpx_get_worker_interface()->init(&worker_);
+  }
+
+  virtual void TearDown() {
+    vpx_get_worker_interface()->end(&worker_);
+  }
+
+  void Run(VPxWorker* worker) {
+    const bool synchronous = GetParam();
+    if (synchronous) {
+      vpx_get_worker_interface()->execute(worker);
+    } else {
+      vpx_get_worker_interface()->launch(worker);
+    }
+  }
+
+  VPxWorker worker_;
+};
+
+int ThreadHook(void* data, void* return_value) {
+  int* const hook_data = reinterpret_cast<int*>(data);
+  *hook_data = 5;
+  return *reinterpret_cast<int*>(return_value);
+}
+
+TEST_P(VPxWorkerThreadTest, HookSuccess) {
+  // should be a no-op.
+  EXPECT_NE(vpx_get_worker_interface()->sync(&worker_), 0);
+
+  for (int i = 0; i < 2; ++i) {
+    EXPECT_NE(vpx_get_worker_interface()->reset(&worker_), 0);
+
+    int hook_data = 0;
+    int return_value = 1;  // return successfully from the hook
+    worker_.hook = ThreadHook;
+    worker_.data1 = &hook_data;
+    worker_.data2 = &return_value;
+
+    Run(&worker_);
+    EXPECT_NE(vpx_get_worker_interface()->sync(&worker_), 0);
+    EXPECT_FALSE(worker_.had_error);
+    EXPECT_EQ(5, hook_data);
+
+    // should be a no-op.
+    EXPECT_NE(vpx_get_worker_interface()->sync(&worker_), 0);
+  }
+}
+
+TEST_P(VPxWorkerThreadTest, HookFailure) {
+  EXPECT_NE(vpx_get_worker_interface()->reset(&worker_), 0);
+
+  int hook_data = 0;
+  int return_value = 0;  // return failure from the hook
+  worker_.hook = ThreadHook;
+  worker_.data1 = &hook_data;
+  worker_.data2 = &return_value;
+
+  Run(&worker_);
+  EXPECT_FALSE(vpx_get_worker_interface()->sync(&worker_));
+  EXPECT_EQ(1, worker_.had_error);
+
+  // Ensure _reset() clears the error and _launch() can be called again.
+  return_value = 1;
+  EXPECT_NE(vpx_get_worker_interface()->reset(&worker_), 0);
+  EXPECT_FALSE(worker_.had_error);
+  vpx_get_worker_interface()->launch(&worker_);
+  EXPECT_NE(vpx_get_worker_interface()->sync(&worker_), 0);
+  EXPECT_FALSE(worker_.had_error);
+}
+
+TEST_P(VPxWorkerThreadTest, EndWithoutSync) {
+  // Create a large number of threads to increase the chances of detecting a
+  // race. Doing more work in the hook is no guarantee as any race would occur
+  // post hook execution in the main thread loop driver.
+  static const int kNumWorkers = 64;
+  VPxWorker workers[kNumWorkers];
+  int hook_data[kNumWorkers];
+  int return_value[kNumWorkers];
+
+  for (int n = 0; n < kNumWorkers; ++n) {
+    vpx_get_worker_interface()->init(&workers[n]);
+    return_value[n] = 1;  // return successfully from the hook
+    workers[n].hook = ThreadHook;
+    workers[n].data1 = &hook_data[n];
+    workers[n].data2 = &return_value[n];
+  }
+
+  for (int i = 0; i < 2; ++i) {
+    for (int n = 0; n < kNumWorkers; ++n) {
+      EXPECT_NE(vpx_get_worker_interface()->reset(&workers[n]), 0);
+      hook_data[n] = 0;
+    }
+
+    for (int n = 0; n < kNumWorkers; ++n) {
+      Run(&workers[n]);
+    }
+
+    for (int n = kNumWorkers - 1; n >= 0; --n) {
+      vpx_get_worker_interface()->end(&workers[n]);
+    }
+  }
+}
+
+TEST(VPxWorkerThreadTest, TestInterfaceAPI) {
+  EXPECT_EQ(0, vpx_set_worker_interface(NULL));
+  EXPECT_TRUE(vpx_get_worker_interface() != NULL);
+  for (int i = 0; i < 6; ++i) {
+    VPxWorkerInterface winterface = *vpx_get_worker_interface();
+    switch (i) {
+      default:
+      case 0: winterface.init = NULL; break;
+      case 1: winterface.reset = NULL; break;
+      case 2: winterface.sync = NULL; break;
+      case 3: winterface.launch = NULL; break;
+      case 4: winterface.execute = NULL; break;
+      case 5: winterface.end = NULL; break;
+    }
+    EXPECT_EQ(0, vpx_set_worker_interface(&winterface));
+  }
+}
+
+// -----------------------------------------------------------------------------
+// Multi-threaded decode tests
+
+#if CONFIG_WEBM_IO
+struct FileList {
+  const char *name;
+  const char *expected_md5;
+};
+
+// Decodes |filename| with |num_threads|. Returns the md5 of the decoded frames.
+string DecodeFile(const string& filename, int num_threads) {
+  libvpx_test::WebMVideoSource video(filename);
+  video.Init();
+
+  vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+  cfg.threads = num_threads;
+  libvpx_test::VP9Decoder decoder(cfg, 0);
+
+  libvpx_test::MD5 md5;
+  for (video.Begin(); video.cxdata(); video.Next()) {
+    const vpx_codec_err_t res =
+        decoder.DecodeFrame(video.cxdata(), video.frame_size());
+    if (res != VPX_CODEC_OK) {
+      EXPECT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+      break;
+    }
+
+    libvpx_test::DxDataIterator dec_iter = decoder.GetDxData();
+    const vpx_image_t *img = NULL;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next())) {
+      md5.Add(img);
+    }
+  }
+  return string(md5.Get());
+}
+
+void DecodeFiles(const FileList files[]) {
+  for (const FileList *iter = files; iter->name != NULL; ++iter) {
+    SCOPED_TRACE(iter->name);
+    for (int t = 1; t <= 8; ++t) {
+      EXPECT_EQ(iter->expected_md5, DecodeFile(iter->name, t))
+          << "threads = " << t;
+    }
+  }
+}
+
+// Trivial serialized thread worker interface implementation.
+// Note any worker that requires synchronization between other workers will
+// hang.
+namespace impl {
+
+void Init(VPxWorker *const worker) { memset(worker, 0, sizeof(*worker)); }
+int Reset(VPxWorker *const /*worker*/) { return 1; }
+int Sync(VPxWorker *const worker) { return !worker->had_error; }
+
+void Execute(VPxWorker *const worker) {
+  worker->had_error |= !worker->hook(worker->data1, worker->data2);
+}
+
+void Launch(VPxWorker *const worker) { Execute(worker); }
+void End(VPxWorker *const /*worker*/) {}
+
+}  // namespace impl
+
+TEST(VPxWorkerThreadTest, TestSerialInterface) {
+  static const VPxWorkerInterface serial_interface = {
+    impl::Init, impl::Reset, impl::Sync, impl::Launch, impl::Execute, impl::End
+  };
+  // TODO(jzern): Avoid using a file that will use the row-based thread
+  // loopfilter, with the simple serialized implementation it will hang. This is
+  // due to its expectation that rows will be run in parallel as they wait on
+  // progress in the row above before proceeding.
+  static const char expected_md5[] = "b35a1b707b28e82be025d960aba039bc";
+  static const char filename[] = "vp90-2-03-size-226x226.webm";
+  VPxWorkerInterface default_interface = *vpx_get_worker_interface();
+
+  EXPECT_NE(vpx_set_worker_interface(&serial_interface), 0);
+  EXPECT_EQ(expected_md5, DecodeFile(filename, 2));
+
+  // Reset the interface.
+  EXPECT_NE(vpx_set_worker_interface(&default_interface), 0);
+  EXPECT_EQ(expected_md5, DecodeFile(filename, 2));
+}
+
+TEST(VP9DecodeMultiThreadedTest, NoTilesNonFrameParallel) {
+  // no tiles or frame parallel; this exercises loop filter threading.
+  EXPECT_EQ("b35a1b707b28e82be025d960aba039bc",
+            DecodeFile("vp90-2-03-size-226x226.webm", 2));
+}
+
+TEST(VP9DecodeMultiThreadedTest, FrameParallel) {
+  static const FileList files[] = {
+    { "vp90-2-08-tile_1x2_frame_parallel.webm",
+      "68ede6abd66bae0a2edf2eb9232241b6" },
+    { "vp90-2-08-tile_1x4_frame_parallel.webm",
+      "368ebc6ebf3a5e478d85b2c3149b2848" },
+    { "vp90-2-08-tile_1x8_frame_parallel.webm",
+      "17e439da2388aff3a0f69cb22579c6c1" },
+    { NULL, NULL }
+  };
+
+  DecodeFiles(files);
+}
+
+TEST(VP9DecodeMultiThreadedTest, FrameParallelResize) {
+  static const FileList files[] = {
+    { "vp90-2-14-resize-fp-tiles-1-16.webm",
+      "0cd5e632c326297e975f38949c31ea94" },
+    { "vp90-2-14-resize-fp-tiles-1-2-4-8-16.webm",
+      "5c78a96a42e7f4a4f6b2edcdb791e44c" },
+    { "vp90-2-14-resize-fp-tiles-1-2.webm",
+      "e030450ae85c3277be2a418769df98e2" },
+    { "vp90-2-14-resize-fp-tiles-1-4.webm",
+      "312eed4e2b64eb7a4e7f18916606a430" },
+    { "vp90-2-14-resize-fp-tiles-16-1.webm",
+      "1755c16d8af16a9cb3fe7338d90abe52" },
+    { "vp90-2-14-resize-fp-tiles-16-2.webm",
+      "500300592d3fcb6f12fab25e48aaf4df" },
+    { "vp90-2-14-resize-fp-tiles-16-4.webm",
+      "47c48379fa6331215d91c67648e1af6e" },
+    { "vp90-2-14-resize-fp-tiles-16-8-4-2-1.webm",
+      "eecf17290739bc708506fa4827665989" },
+    { "vp90-2-14-resize-fp-tiles-16-8.webm",
+      "29b6bb54e4c26b5ca85d5de5fed94e76" },
+    { "vp90-2-14-resize-fp-tiles-1-8.webm",
+      "1b6f175e08cd82cf84bb800ac6d1caa3" },
+    { "vp90-2-14-resize-fp-tiles-2-16.webm",
+      "ca3b03e4197995d8d5444ede7a6c0804" },
+    { "vp90-2-14-resize-fp-tiles-2-1.webm",
+      "99aec065369d70bbb78ccdff65afed3f" },
+    { "vp90-2-14-resize-fp-tiles-2-4.webm",
+      "22d0ebdb49b87d2920a85aea32e1afd5" },
+    { "vp90-2-14-resize-fp-tiles-2-8.webm",
+      "c2115cf051c62e0f7db1d4a783831541" },
+    { "vp90-2-14-resize-fp-tiles-4-16.webm",
+      "c690d7e1719b31367564cac0af0939cb" },
+    { "vp90-2-14-resize-fp-tiles-4-1.webm",
+      "a926020b2cc3e15ad4cc271853a0ff26" },
+    { "vp90-2-14-resize-fp-tiles-4-2.webm",
+      "42699063d9e581f1993d0cf890c2be78" },
+    { "vp90-2-14-resize-fp-tiles-4-8.webm",
+      "7f76d96036382f45121e3d5aa6f8ec52" },
+    { "vp90-2-14-resize-fp-tiles-8-16.webm",
+      "76a43fcdd7e658542913ea43216ec55d" },
+    { "vp90-2-14-resize-fp-tiles-8-1.webm",
+      "8e3fbe89486ca60a59299dea9da91378" },
+    { "vp90-2-14-resize-fp-tiles-8-2.webm",
+      "ae96f21f21b6370cc0125621b441fc52" },
+    { "vp90-2-14-resize-fp-tiles-8-4.webm",
+      "3eb4f24f10640d42218f7fd7b9fd30d4" },
+    { NULL, NULL }
+  };
+
+  DecodeFiles(files);
+}
+
+TEST(VP9DecodeMultiThreadedTest, NonFrameParallel) {
+  static const FileList files[] = {
+    { "vp90-2-08-tile_1x2.webm", "570b4a5d5a70d58b5359671668328a16" },
+    { "vp90-2-08-tile_1x4.webm", "988d86049e884c66909d2d163a09841a" },
+    { "vp90-2-08-tile_1x8.webm", "0941902a52e9092cb010905eab16364c" },
+    { "vp90-2-08-tile-4x1.webm", "06505aade6647c583c8e00a2f582266f" },
+    { "vp90-2-08-tile-4x4.webm", "85c2299892460d76e2c600502d52bfe2" },
+    { NULL, NULL }
+  };
+
+  DecodeFiles(files);
+}
+#endif  // CONFIG_WEBM_IO
+
+INSTANTIATE_TEST_CASE_P(Synchronous, VPxWorkerThreadTest, ::testing::Bool());
+
+}  // namespace
diff --git a/libvpx/libvpx/test/vpx_scale_test.cc b/libvpx/libvpx/test/vpx_scale_test.cc
new file mode 100644
index 0000000..ef716fc
--- /dev/null
+++ b/libvpx/libvpx/test/vpx_scale_test.cc
@@ -0,0 +1,255 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./vpx_scale_rtcd.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_scale/yv12config.h"
+
+namespace {
+
+typedef void (*ExtendFrameBorderFunc)(YV12_BUFFER_CONFIG *ybf);
+typedef void (*CopyFrameFunc)(const YV12_BUFFER_CONFIG *src_ybf,
+                              YV12_BUFFER_CONFIG *dst_ybf);
+
+class VpxScaleBase {
+ public:
+  virtual ~VpxScaleBase() {
+    libvpx_test::ClearSystemState();
+  }
+
+  void ResetImage(int width, int height) {
+    width_ = width;
+    height_ = height;
+    memset(&img_, 0, sizeof(img_));
+    ASSERT_EQ(0, vp8_yv12_alloc_frame_buffer(&img_, width_, height_,
+                                             VP8BORDERINPIXELS));
+    memset(img_.buffer_alloc, kBufFiller, img_.frame_size);
+    FillPlane(img_.y_buffer, img_.y_crop_width, img_.y_crop_height,
+              img_.y_stride);
+    FillPlane(img_.u_buffer, img_.uv_crop_width, img_.uv_crop_height,
+              img_.uv_stride);
+    FillPlane(img_.v_buffer, img_.uv_crop_width, img_.uv_crop_height,
+              img_.uv_stride);
+
+    memset(&ref_img_, 0, sizeof(ref_img_));
+    ASSERT_EQ(0, vp8_yv12_alloc_frame_buffer(&ref_img_, width_, height_,
+                                             VP8BORDERINPIXELS));
+    memset(ref_img_.buffer_alloc, kBufFiller, ref_img_.frame_size);
+
+    memset(&cpy_img_, 0, sizeof(cpy_img_));
+    ASSERT_EQ(0, vp8_yv12_alloc_frame_buffer(&cpy_img_, width_, height_,
+                                             VP8BORDERINPIXELS));
+    memset(cpy_img_.buffer_alloc, kBufFiller, cpy_img_.frame_size);
+    ReferenceCopyFrame();
+  }
+
+  void DeallocImage() {
+    vp8_yv12_de_alloc_frame_buffer(&img_);
+    vp8_yv12_de_alloc_frame_buffer(&ref_img_);
+    vp8_yv12_de_alloc_frame_buffer(&cpy_img_);
+  }
+
+ protected:
+  static const int kBufFiller = 123;
+  static const int kBufMax = kBufFiller - 1;
+
+  static void FillPlane(uint8_t *buf, int width, int height, int stride) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        buf[x + (y * stride)] = (x + (width * y)) % kBufMax;
+      }
+    }
+  }
+
+  static void ExtendPlane(uint8_t *buf, int crop_width, int crop_height,
+                          int width, int height, int stride, int padding) {
+    // Copy the outermost visible pixel to a distance of at least 'padding.'
+    // The buffers are allocated such that there may be excess space outside the
+    // padding. As long as the minimum amount of padding is achieved it is not
+    // necessary to fill this space as well.
+    uint8_t *left = buf - padding;
+    uint8_t *right = buf + crop_width;
+    const int right_extend = padding + (width - crop_width);
+    const int bottom_extend = padding + (height - crop_height);
+
+    // Fill the border pixels from the nearest image pixel.
+    for (int y = 0; y < crop_height; ++y) {
+      memset(left, left[padding], padding);
+      memset(right, right[-1], right_extend);
+      left += stride;
+      right += stride;
+    }
+
+    left = buf - padding;
+    uint8_t *top = left - (stride * padding);
+    // The buffer does not always extend as far as the stride.
+    // Equivalent to padding + width + padding.
+    const int extend_width = padding + crop_width + right_extend;
+
+    // The first row was already extended to the left and right. Copy it up.
+    for (int y = 0; y < padding; ++y) {
+      memcpy(top, left, extend_width);
+      top += stride;
+    }
+
+    uint8_t *bottom = left + (crop_height * stride);
+    for (int y = 0; y <  bottom_extend; ++y) {
+      memcpy(bottom, left + (crop_height - 1) * stride, extend_width);
+      bottom += stride;
+    }
+  }
+
+  void ReferenceExtendBorder() {
+    ExtendPlane(ref_img_.y_buffer,
+                ref_img_.y_crop_width, ref_img_.y_crop_height,
+                ref_img_.y_width, ref_img_.y_height,
+                ref_img_.y_stride,
+                ref_img_.border);
+    ExtendPlane(ref_img_.u_buffer,
+                ref_img_.uv_crop_width, ref_img_.uv_crop_height,
+                ref_img_.uv_width, ref_img_.uv_height,
+                ref_img_.uv_stride,
+                ref_img_.border / 2);
+    ExtendPlane(ref_img_.v_buffer,
+                ref_img_.uv_crop_width, ref_img_.uv_crop_height,
+                ref_img_.uv_width, ref_img_.uv_height,
+                ref_img_.uv_stride,
+                ref_img_.border / 2);
+  }
+
+  void ReferenceCopyFrame() {
+    // Copy img_ to ref_img_ and extend frame borders. This will be used for
+    // verifying extend_fn_ as well as copy_frame_fn_.
+    EXPECT_EQ(ref_img_.frame_size, img_.frame_size);
+    for (int y = 0; y < img_.y_crop_height; ++y) {
+      for (int x = 0; x < img_.y_crop_width; ++x) {
+        ref_img_.y_buffer[x + y * ref_img_.y_stride] =
+            img_.y_buffer[x + y * img_.y_stride];
+      }
+    }
+
+    for (int y = 0; y < img_.uv_crop_height; ++y) {
+      for (int x = 0; x < img_.uv_crop_width; ++x) {
+        ref_img_.u_buffer[x + y * ref_img_.uv_stride] =
+            img_.u_buffer[x + y * img_.uv_stride];
+        ref_img_.v_buffer[x + y * ref_img_.uv_stride] =
+            img_.v_buffer[x + y * img_.uv_stride];
+      }
+    }
+
+    ReferenceExtendBorder();
+  }
+
+  void CompareImages(const YV12_BUFFER_CONFIG actual) {
+    EXPECT_EQ(ref_img_.frame_size, actual.frame_size);
+    EXPECT_EQ(0, memcmp(ref_img_.buffer_alloc, actual.buffer_alloc,
+                        ref_img_.frame_size));
+  }
+
+  YV12_BUFFER_CONFIG img_;
+  YV12_BUFFER_CONFIG ref_img_;
+  YV12_BUFFER_CONFIG cpy_img_;
+  int width_;
+  int height_;
+};
+
+class ExtendBorderTest
+    : public VpxScaleBase,
+      public ::testing::TestWithParam<ExtendFrameBorderFunc> {
+ public:
+  virtual ~ExtendBorderTest() {}
+
+ protected:
+  virtual void SetUp() {
+    extend_fn_ = GetParam();
+  }
+
+  void ExtendBorder() {
+    ASM_REGISTER_STATE_CHECK(extend_fn_(&img_));
+  }
+
+  void RunTest() {
+#if ARCH_ARM
+    // Some arm devices OOM when trying to allocate the largest buffers.
+    static const int kNumSizesToTest = 6;
+#else
+    static const int kNumSizesToTest = 7;
+#endif
+    static const int kSizesToTest[] = {1, 15, 33, 145, 512, 1025, 16383};
+    for (int h = 0; h < kNumSizesToTest; ++h) {
+      for (int w = 0; w < kNumSizesToTest; ++w) {
+        ResetImage(kSizesToTest[w], kSizesToTest[h]);
+        ExtendBorder();
+        ReferenceExtendBorder();
+        CompareImages(img_);
+        DeallocImage();
+      }
+    }
+  }
+
+  ExtendFrameBorderFunc extend_fn_;
+};
+
+TEST_P(ExtendBorderTest, ExtendBorder) {
+  ASSERT_NO_FATAL_FAILURE(RunTest());
+}
+
+INSTANTIATE_TEST_CASE_P(C, ExtendBorderTest,
+                        ::testing::Values(vp8_yv12_extend_frame_borders_c));
+
+class CopyFrameTest
+    : public VpxScaleBase,
+      public ::testing::TestWithParam<CopyFrameFunc> {
+ public:
+  virtual ~CopyFrameTest() {}
+
+ protected:
+  virtual void SetUp() {
+    copy_frame_fn_ = GetParam();
+  }
+
+  void CopyFrame() {
+    ASM_REGISTER_STATE_CHECK(copy_frame_fn_(&img_, &cpy_img_));
+  }
+
+  void RunTest() {
+#if ARCH_ARM
+    // Some arm devices OOM when trying to allocate the largest buffers.
+    static const int kNumSizesToTest = 6;
+#else
+    static const int kNumSizesToTest = 7;
+#endif
+    static const int kSizesToTest[] = {1, 15, 33, 145, 512, 1025, 16383};
+    for (int h = 0; h < kNumSizesToTest; ++h) {
+      for (int w = 0; w < kNumSizesToTest; ++w) {
+        ResetImage(kSizesToTest[w], kSizesToTest[h]);
+        ReferenceCopyFrame();
+        CopyFrame();
+        CompareImages(cpy_img_);
+        DeallocImage();
+      }
+    }
+  }
+
+  CopyFrameFunc copy_frame_fn_;
+};
+
+TEST_P(CopyFrameTest, CopyFrame) {
+  ASSERT_NO_FATAL_FAILURE(RunTest());
+}
+
+INSTANTIATE_TEST_CASE_P(C, CopyFrameTest,
+                        ::testing::Values(vp8_yv12_copy_frame_c));
+}  // namespace
diff --git a/libvpx/libvpx/test/vpx_temporal_svc_encoder.sh b/libvpx/libvpx/test/vpx_temporal_svc_encoder.sh
new file mode 100755
index 0000000..fcc8cb4
--- /dev/null
+++ b/libvpx/libvpx/test/vpx_temporal_svc_encoder.sh
@@ -0,0 +1,290 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests the libvpx vpx_temporal_svc_encoder example. To add new
+##  tests to this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vpx_tsvc_encoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+vpx_tsvc_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+  if [ "$(vpx_config_option_enabled CONFIG_TEMPORAL_DENOISING)" != "yes" ]; then
+    elog "Warning: Temporal denoising is disabled! Spatial denoising will be " \
+      "used instead, which is probably not what you want for this test."
+  fi
+}
+
+# Runs vpx_temporal_svc_encoder using the codec specified by $1 and output file
+# name by $2. Additional positional parameters are passed directly to
+# vpx_temporal_svc_encoder.
+vpx_tsvc_encoder() {
+  local encoder="${LIBVPX_BIN_PATH}/vpx_temporal_svc_encoder"
+  encoder="${encoder}${VPX_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file_base="$2"
+  local output_file="${VPX_TEST_OUTPUT_DIR}/${output_file_base}"
+  local timebase_num="1"
+  local timebase_den="1000"
+  local speed="6"
+  local frame_drop_thresh="30"
+
+  shift 2
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "${YUV_RAW_INPUT}" "${output_file}" \
+      "${codec}" "${YUV_RAW_INPUT_WIDTH}" "${YUV_RAW_INPUT_HEIGHT}" \
+      "${timebase_num}" "${timebase_den}" "${speed}" "${frame_drop_thresh}" \
+      "$@" \
+      ${devnull}
+}
+
+# Confirms that all expected output files exist given the output file name
+# passed to vpx_temporal_svc_encoder.
+# The file name passed to vpx_temporal_svc_encoder is joined with the stream
+# number and the extension .ivf to produce per stream output files.  Here $1 is
+# file name, and $2 is expected number of files.
+files_exist() {
+  local file_name="${VPX_TEST_OUTPUT_DIR}/$1"
+  local num_files="$(($2 - 1))"
+  for stream_num in $(seq 0 ${num_files}); do
+    [ -e "${file_name}_${stream_num}.ivf" ] || return 1
+  done
+}
+
+# Run vpx_temporal_svc_encoder in all supported modes for vp8 and vp9.
+
+vpx_tsvc_encoder_vp8_mode_0() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 0 200 || return 1
+    # Mode 0 produces 1 stream
+    files_exist "${FUNCNAME}" 1 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_1() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 1 200 400 || return 1
+    # Mode 1 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_2() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 2 200 400 || return 1
+    # Mode 2 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_3() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 3 200 400 600 || return 1
+    # Mode 3 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_4() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 4 200 400 600 || return 1
+    # Mode 4 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_5() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 5 200 400 600 || return 1
+    # Mode 5 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_6() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 6 200 400 600 || return 1
+    # Mode 6 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_7() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 7 200 400 600 800 1000 || return 1
+    # Mode 7 produces 5 streams
+    files_exist "${FUNCNAME}" 5 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 8 200 400 || return 1
+    # Mode 8 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_9() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 9 200 400 600 || return 1
+    # Mode 9 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_10() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 10 200 400 600 || return 1
+    # Mode 10 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp8_mode_11() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp8 "${FUNCNAME}" 11 200 400 600 || return 1
+    # Mode 11 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_0() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 0 200 || return 1
+    # Mode 0 produces 1 stream
+    files_exist "${FUNCNAME}" 1 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_1() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 1 200 400 || return 1
+    # Mode 1 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_2() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 2 200 400 || return 1
+    # Mode 2 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_3() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 3 200 400 600 || return 1
+    # Mode 3 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_4() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 4 200 400 600 || return 1
+    # Mode 4 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_5() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 5 200 400 600 || return 1
+    # Mode 5 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_6() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 6 200 400 600 || return 1
+    # Mode 6 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_7() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 7 200 400 600 800 1000 || return 1
+    # Mode 7 produces 5 streams
+    files_exist "${FUNCNAME}" 5 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_8() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 8 200 400 || return 1
+    # Mode 8 produces 2 streams
+    files_exist "${FUNCNAME}" 2 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_9() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 9 200 400 600 || return 1
+    # Mode 9 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_10() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 10 200 400 600 || return 1
+    # Mode 10 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_vp9_mode_11() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    vpx_tsvc_encoder vp9 "${FUNCNAME}" 11 200 400 600 || return 1
+    # Mode 11 produces 3 streams
+    files_exist "${FUNCNAME}" 3 || return 1
+  fi
+}
+
+vpx_tsvc_encoder_tests="vpx_tsvc_encoder_vp8_mode_0
+                        vpx_tsvc_encoder_vp8_mode_1
+                        vpx_tsvc_encoder_vp8_mode_2
+                        vpx_tsvc_encoder_vp8_mode_3
+                        vpx_tsvc_encoder_vp8_mode_4
+                        vpx_tsvc_encoder_vp8_mode_5
+                        vpx_tsvc_encoder_vp8_mode_6
+                        vpx_tsvc_encoder_vp8_mode_7
+                        vpx_tsvc_encoder_vp8_mode_8
+                        vpx_tsvc_encoder_vp8_mode_9
+                        vpx_tsvc_encoder_vp8_mode_10
+                        vpx_tsvc_encoder_vp8_mode_11
+                        vpx_tsvc_encoder_vp9_mode_0
+                        vpx_tsvc_encoder_vp9_mode_1
+                        vpx_tsvc_encoder_vp9_mode_2
+                        vpx_tsvc_encoder_vp9_mode_3
+                        vpx_tsvc_encoder_vp9_mode_4
+                        vpx_tsvc_encoder_vp9_mode_5
+                        vpx_tsvc_encoder_vp9_mode_6
+                        vpx_tsvc_encoder_vp9_mode_7
+                        vpx_tsvc_encoder_vp9_mode_8
+                        vpx_tsvc_encoder_vp9_mode_9
+                        vpx_tsvc_encoder_vp9_mode_10
+                        vpx_tsvc_encoder_vp9_mode_11"
+
+run_tests vpx_tsvc_encoder_verify_environment "${vpx_tsvc_encoder_tests}"
diff --git a/libvpx/libvpx/test/vpxdec.sh b/libvpx/libvpx/test/vpxdec.sh
new file mode 100755
index 0000000..de51c80
--- /dev/null
+++ b/libvpx/libvpx/test/vpxdec.sh
@@ -0,0 +1,116 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests vpxdec. To add new tests to this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vpxdec_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available.
+vpxdec_verify_environment() {
+  if [ ! -e "${VP8_IVF_FILE}" ] || [ ! -e "${VP9_WEBM_FILE}" ] || \
+    [ ! -e "${VP9_FPM_WEBM_FILE}" ] || \
+    [ ! -e "${VP9_LT_50_FRAMES_WEBM_FILE}" ] ; then
+    elog "Libvpx test data must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+  if [ -z "$(vpx_tool_path vpxdec)" ]; then
+    elog "vpxdec not found. It must exist in LIBVPX_BIN_PATH or its parent."
+    return 1
+  fi
+}
+
+# Wrapper function for running vpxdec with pipe input. Requires that
+# LIBVPX_BIN_PATH points to the directory containing vpxdec. $1 is used as the
+# input file path and shifted away. All remaining parameters are passed through
+# to vpxdec.
+vpxdec_pipe() {
+  local readonly decoder="$(vpx_tool_path vpxdec)"
+  local readonly input="$1"
+  shift
+  cat "${input}" | eval "${VPX_TEST_PREFIX}" "${decoder}" - "$@" ${devnull}
+}
+
+# Wrapper function for running vpxdec. Requires that LIBVPX_BIN_PATH points to
+# the directory containing vpxdec. $1 one is used as the input file path and
+# shifted away. All remaining parameters are passed through to vpxdec.
+vpxdec() {
+  local readonly decoder="$(vpx_tool_path vpxdec)"
+  local readonly input="$1"
+  shift
+  eval "${VPX_TEST_PREFIX}" "${decoder}" "$input" "$@" ${devnull}
+}
+
+vpxdec_can_decode_vp8() {
+  if [ "$(vp8_decode_available)" = "yes" ]; then
+    echo yes
+  fi
+}
+
+vpxdec_can_decode_vp9() {
+  if [ "$(vp9_decode_available)" = "yes" ]; then
+    echo yes
+  fi
+}
+
+vpxdec_vp8_ivf() {
+  if [ "$(vpxdec_can_decode_vp8)" = "yes" ]; then
+    vpxdec "${VP8_IVF_FILE}" --summary --noblit
+  fi
+}
+
+vpxdec_vp8_ivf_pipe_input() {
+  if [ "$(vpxdec_can_decode_vp8)" = "yes" ]; then
+    vpxdec_pipe "${VP8_IVF_FILE}" --summary --noblit
+  fi
+}
+
+vpxdec_vp9_webm() {
+  if [ "$(vpxdec_can_decode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    vpxdec "${VP9_WEBM_FILE}" --summary --noblit
+  fi
+}
+
+vpxdec_vp9_webm_frame_parallel() {
+  if [ "$(vpxdec_can_decode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    for threads in 2 3 4 5 6 7 8; do
+      vpxdec "${VP9_FPM_WEBM_FILE}" --summary --noblit --threads=$threads \
+        --frame-parallel
+    done
+  fi
+}
+
+vpxdec_vp9_webm_less_than_50_frames() {
+  # ensure that reaching eof in webm_guess_framerate doesn't result in invalid
+  # frames in actual webm_read_frame calls.
+  if [ "$(vpxdec_can_decode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly decoder="$(vpx_tool_path vpxdec)"
+    local readonly expected=10
+    local readonly num_frames=$(${VPX_TEST_PREFIX} "${decoder}" \
+      "${VP9_LT_50_FRAMES_WEBM_FILE}" --summary --noblit 2>&1 \
+      | awk '/^[0-9]+ decoded frames/ { print $1 }')
+    if [ "$num_frames" -ne "$expected" ]; then
+      elog "Output frames ($num_frames) != expected ($expected)"
+      return 1
+    fi
+  fi
+}
+
+vpxdec_tests="vpxdec_vp8_ivf
+              vpxdec_vp8_ivf_pipe_input
+              vpxdec_vp9_webm
+              vpxdec_vp9_webm_frame_parallel
+              vpxdec_vp9_webm_less_than_50_frames"
+
+run_tests vpxdec_verify_environment "${vpxdec_tests}"
diff --git a/libvpx/libvpx/test/vpxenc.sh b/libvpx/libvpx/test/vpxenc.sh
new file mode 100755
index 0000000..e899499
--- /dev/null
+++ b/libvpx/libvpx/test/vpxenc.sh
@@ -0,0 +1,429 @@
+#!/bin/sh
+##
+##  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+##  This file tests vpxenc using hantro_collage_w352h288.yuv as input. To add
+##  new tests to this file, do the following:
+##    1. Write a shell function (this is your test).
+##    2. Add the function to vpxenc_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+readonly TEST_FRAMES=10
+
+# Environment check: Make sure input is available.
+vpxenc_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    elog "The file ${YUV_RAW_INPUT##*/} must exist in LIBVPX_TEST_DATA_PATH."
+    return 1
+  fi
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ]; then
+    if [ ! -e "${Y4M_NOSQ_PAR_INPUT}" ]; then
+      elog "The file ${Y4M_NOSQ_PAR_INPUT##*/} must exist in"
+      elog "LIBVPX_TEST_DATA_PATH."
+      return 1
+    fi
+  fi
+  if [ -z "$(vpx_tool_path vpxenc)" ]; then
+    elog "vpxenc not found. It must exist in LIBVPX_BIN_PATH or its parent."
+    return 1
+  fi
+}
+
+vpxenc_can_encode_vp8() {
+  if [ "$(vp8_encode_available)" = "yes" ]; then
+    echo yes
+  fi
+}
+
+vpxenc_can_encode_vp9() {
+  if [ "$(vp9_encode_available)" = "yes" ]; then
+    echo yes
+  fi
+}
+
+# Echo vpxenc command line parameters allowing use of
+# hantro_collage_w352h288.yuv as input.
+yuv_input_hantro_collage() {
+  echo ""${YUV_RAW_INPUT}"
+       --width="${YUV_RAW_INPUT_WIDTH}"
+       --height="${YUV_RAW_INPUT_HEIGHT}""
+}
+
+y4m_input_non_square_par() {
+  echo ""${Y4M_NOSQ_PAR_INPUT}""
+}
+
+y4m_input_720p() {
+  echo ""${Y4M_720P_INPUT}""
+}
+
+# Echo default vpxenc real time encoding params. $1 is the codec, which defaults
+# to vp8 if unspecified.
+vpxenc_rt_params() {
+  local readonly codec="${1:-vp8}"
+  echo "--codec=${codec}
+    --buf-initial-sz=500
+    --buf-optimal-sz=600
+    --buf-sz=1000
+    --cpu-used=-6
+    --end-usage=cbr
+    --error-resilient=1
+    --kf-max-dist=90000
+    --lag-in-frames=0
+    --max-intra-rate=300
+    --max-q=56
+    --min-q=2
+    --noise-sensitivity=0
+    --overshoot-pct=50
+    --passes=1
+    --profile=0
+    --resize-allowed=0
+    --rt
+    --static-thresh=0
+    --undershoot-pct=50"
+}
+
+# Wrapper function for running vpxenc with pipe input. Requires that
+# LIBVPX_BIN_PATH points to the directory containing vpxenc. $1 is used as the
+# input file path and shifted away. All remaining parameters are passed through
+# to vpxenc.
+vpxenc_pipe() {
+  local readonly encoder="$(vpx_tool_path vpxenc)"
+  local readonly input="$1"
+  shift
+  cat "${input}" | eval "${VPX_TEST_PREFIX}" "${encoder}" - \
+    --test-decode=fatal \
+    "$@" ${devnull}
+}
+
+# Wrapper function for running vpxenc. Requires that LIBVPX_BIN_PATH points to
+# the directory containing vpxenc. $1 one is used as the input file path and
+# shifted away. All remaining parameters are passed through to vpxenc.
+vpxenc() {
+  local readonly encoder="$(vpx_tool_path vpxenc)"
+  local readonly input="$1"
+  shift
+  eval "${VPX_TEST_PREFIX}" "${encoder}" "${input}" \
+    --test-decode=fatal \
+    "$@" ${devnull}
+}
+
+vpxenc_vp8_ivf() {
+  if [ "$(vpxenc_can_encode_vp8)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp8.ivf"
+    vpxenc $(yuv_input_hantro_collage) \
+      --codec=vp8 \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp8_webm() {
+  if [ "$(vpxenc_can_encode_vp8)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp8.webm"
+    vpxenc $(yuv_input_hantro_collage) \
+      --codec=vp8 \
+      --limit="${TEST_FRAMES}" \
+      --output="${output}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp8_webm_rt() {
+  if [ "$(vpxenc_can_encode_vp8)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp8_rt.webm"
+    vpxenc $(yuv_input_hantro_collage) \
+      $(vpxenc_rt_params vp8) \
+      --output="${output}"
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp8_webm_2pass() {
+  if [ "$(vpxenc_can_encode_vp8)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp8.webm"
+    vpxenc $(yuv_input_hantro_collage) \
+      --codec=vp8 \
+      --limit="${TEST_FRAMES}" \
+      --output="${output}" \
+      --passes=2
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp8_webm_lag10_frames20() {
+  if [ "$(vpxenc_can_encode_vp8)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly lag_total_frames=20
+    local readonly lag_frames=10
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp8_lag10_frames20.webm"
+    vpxenc $(yuv_input_hantro_collage) \
+      --codec=vp8 \
+      --limit="${lag_total_frames}" \
+      --lag-in-frames="${lag_frames}" \
+      --output="${output}" \
+      --auto-alt-ref=1 \
+      --passes=2
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp8_ivf_piped_input() {
+  if [ "$(vpxenc_can_encode_vp8)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp8_piped_input.ivf"
+    vpxenc_pipe $(yuv_input_hantro_collage) \
+      --codec=vp8 \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp9_ivf() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9.ivf"
+    vpxenc $(yuv_input_hantro_collage) \
+      --codec=vp9 \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp9_webm() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9.webm"
+    vpxenc $(yuv_input_hantro_collage) \
+      --codec=vp9 \
+      --limit="${TEST_FRAMES}" \
+      --output="${output}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp9_webm_rt() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9_rt.webm"
+    vpxenc $(yuv_input_hantro_collage) \
+      $(vpxenc_rt_params vp9) \
+      --output="${output}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp9_webm_rt_multithread_tiled() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9_rt_multithread_tiled.webm"
+    local readonly tilethread_min=2
+    local readonly tilethread_max=4
+    local readonly num_threads="$(seq ${tilethread_min} ${tilethread_max})"
+    local readonly num_tile_cols="$(seq ${tilethread_min} ${tilethread_max})"
+
+    for threads in ${num_threads}; do
+      for tile_cols in ${num_tile_cols}; do
+        vpxenc $(y4m_input_720p) \
+          $(vpxenc_rt_params vp9) \
+          --threads=${threads} \
+          --tile-columns=${tile_cols} \
+          --output="${output}"
+      done
+    done
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+
+    rm "${output}"
+  fi
+}
+
+vpxenc_vp9_webm_rt_multithread_tiled_frameparallel() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9_rt_mt_t_fp.webm"
+    local readonly tilethread_min=2
+    local readonly tilethread_max=4
+    local readonly num_threads="$(seq ${tilethread_min} ${tilethread_max})"
+    local readonly num_tile_cols="$(seq ${tilethread_min} ${tilethread_max})"
+
+    for threads in ${num_threads}; do
+      for tile_cols in ${num_tile_cols}; do
+        vpxenc $(y4m_input_720p) \
+          $(vpxenc_rt_params vp9) \
+          --threads=${threads} \
+          --tile-columns=${tile_cols} \
+          --frame-parallel=1 \
+          --output="${output}"
+      done
+    done
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+
+    rm "${output}"
+  fi
+}
+
+vpxenc_vp9_webm_2pass() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9.webm"
+    vpxenc $(yuv_input_hantro_collage) \
+      --codec=vp9 \
+      --limit="${TEST_FRAMES}" \
+      --output="${output}" \
+      --passes=2
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp9_ivf_lossless() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9_lossless.ivf"
+    vpxenc $(yuv_input_hantro_collage) \
+      --codec=vp9 \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}" \
+      --lossless=1
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp9_ivf_minq0_maxq0() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9_lossless_minq0_maxq0.ivf"
+    vpxenc $(yuv_input_hantro_collage) \
+      --codec=vp9 \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}" \
+      --min-q=0 \
+      --max-q=0
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_vp9_webm_lag10_frames20() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly lag_total_frames=20
+    local readonly lag_frames=10
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9_lag10_frames20.webm"
+    vpxenc $(yuv_input_hantro_collage) \
+      --codec=vp9 \
+      --limit="${lag_total_frames}" \
+      --lag-in-frames="${lag_frames}" \
+      --output="${output}" \
+      --passes=2 \
+      --auto-alt-ref=1
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+# TODO(fgalligan): Test that DisplayWidth is different than video width.
+vpxenc_vp9_webm_non_square_par() {
+  if [ "$(vpxenc_can_encode_vp9)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${VPX_TEST_OUTPUT_DIR}/vp9_non_square_par.webm"
+    vpxenc $(y4m_input_non_square_par) \
+      --codec=vp9 \
+      --limit="${TEST_FRAMES}" \
+      --output="${output}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+vpxenc_tests="vpxenc_vp8_ivf
+              vpxenc_vp8_webm
+              vpxenc_vp8_webm_rt
+              vpxenc_vp8_webm_2pass
+              vpxenc_vp8_webm_lag10_frames20
+              vpxenc_vp8_ivf_piped_input
+              vpxenc_vp9_ivf
+              vpxenc_vp9_webm
+              vpxenc_vp9_webm_rt
+              vpxenc_vp9_webm_rt_multithread_tiled
+              vpxenc_vp9_webm_rt_multithread_tiled_frameparallel
+              vpxenc_vp9_webm_2pass
+              vpxenc_vp9_ivf_lossless
+              vpxenc_vp9_ivf_minq0_maxq0
+              vpxenc_vp9_webm_lag10_frames20
+              vpxenc_vp9_webm_non_square_par"
+
+run_tests vpxenc_verify_environment "${vpxenc_tests}"
diff --git a/libvpx/libvpx/test/webm_video_source.h b/libvpx/libvpx/test/webm_video_source.h
new file mode 100644
index 0000000..8258756
--- /dev/null
+++ b/libvpx/libvpx/test/webm_video_source.h
@@ -0,0 +1,102 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef TEST_WEBM_VIDEO_SOURCE_H_
+#define TEST_WEBM_VIDEO_SOURCE_H_
+#include <cstdarg>
+#include <cstdio>
+#include <cstdlib>
+#include <new>
+#include <string>
+#include "../tools_common.h"
+#include "../webmdec.h"
+#include "test/video_source.h"
+
+namespace libvpx_test {
+
+// This class extends VideoSource to allow parsing of WebM files,
+// so that we can do actual file decodes.
+class WebMVideoSource : public CompressedVideoSource {
+ public:
+  explicit WebMVideoSource(const std::string &file_name)
+      : file_name_(file_name),
+        vpx_ctx_(new VpxInputContext()),
+        webm_ctx_(new WebmInputContext()),
+        buf_(NULL),
+        buf_sz_(0),
+        frame_(0),
+        end_of_file_(false) {
+  }
+
+  virtual ~WebMVideoSource() {
+    if (vpx_ctx_->file != NULL)
+      fclose(vpx_ctx_->file);
+    webm_free(webm_ctx_);
+    delete vpx_ctx_;
+    delete webm_ctx_;
+  }
+
+  virtual void Init() {
+  }
+
+  virtual void Begin() {
+    vpx_ctx_->file = OpenTestDataFile(file_name_);
+    ASSERT_TRUE(vpx_ctx_->file != NULL) << "Input file open failed. Filename: "
+        << file_name_;
+
+    ASSERT_EQ(file_is_webm(webm_ctx_, vpx_ctx_), 1) << "file is not WebM";
+
+    FillFrame();
+  }
+
+  virtual void Next() {
+    ++frame_;
+    FillFrame();
+  }
+
+  void FillFrame() {
+    ASSERT_TRUE(vpx_ctx_->file != NULL);
+    const int status = webm_read_frame(webm_ctx_, &buf_, &buf_sz_);
+    ASSERT_GE(status, 0) << "webm_read_frame failed";
+    if (status == 1) {
+      end_of_file_ = true;
+    }
+  }
+
+  void SeekToNextKeyFrame() {
+    ASSERT_TRUE(vpx_ctx_->file != NULL);
+    do {
+      const int status = webm_read_frame(webm_ctx_, &buf_, &buf_sz_);
+      ASSERT_GE(status, 0) << "webm_read_frame failed";
+      ++frame_;
+      if (status == 1) {
+        end_of_file_ = true;
+      }
+    } while (!webm_ctx_->is_key_frame && !end_of_file_);
+  }
+
+  virtual const uint8_t *cxdata() const {
+    return end_of_file_ ? NULL : buf_;
+  }
+  virtual size_t frame_size() const { return buf_sz_; }
+  virtual unsigned int frame_number() const { return frame_; }
+
+ protected:
+  std::string file_name_;
+  VpxInputContext *vpx_ctx_;
+  WebmInputContext *webm_ctx_;
+  uint8_t *buf_;
+  size_t buf_sz_;
+  unsigned int frame_;
+  bool end_of_file_;
+};
+
+}  // namespace libvpx_test
+
+#endif  // TEST_WEBM_VIDEO_SOURCE_H_
diff --git a/libvpx/libvpx/test/y4m_test.cc b/libvpx/libvpx/test/y4m_test.cc
new file mode 100644
index 0000000..a555329
--- /dev/null
+++ b/libvpx/libvpx/test/y4m_test.cc
@@ -0,0 +1,195 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./vpx_config.h"
+#include "./y4menc.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+
+namespace {
+
+using std::string;
+
+static const unsigned int kWidth  = 160;
+static const unsigned int kHeight = 90;
+static const unsigned int kFrames = 10;
+
+struct Y4mTestParam {
+  const char *filename;
+  unsigned int bit_depth;
+  vpx_img_fmt format;
+  const char *md5raw;
+};
+
+const Y4mTestParam kY4mTestVectors[] = {
+  {"park_joy_90p_8_420.y4m", 8, VPX_IMG_FMT_I420,
+    "e5406275b9fc6bb3436c31d4a05c1cab"},
+  {"park_joy_90p_8_422.y4m", 8, VPX_IMG_FMT_I422,
+    "284a47a47133b12884ec3a14e959a0b6"},
+  {"park_joy_90p_8_444.y4m", 8, VPX_IMG_FMT_I444,
+    "90517ff33843d85de712fd4fe60dbed0"},
+  {"park_joy_90p_10_420.y4m", 10, VPX_IMG_FMT_I42016,
+    "63f21f9f717d8b8631bd2288ee87137b"},
+  {"park_joy_90p_10_422.y4m", 10, VPX_IMG_FMT_I42216,
+    "48ab51fb540aed07f7ff5af130c9b605"},
+  {"park_joy_90p_10_444.y4m", 10, VPX_IMG_FMT_I44416,
+    "067bfd75aa85ff9bae91fa3e0edd1e3e"},
+  {"park_joy_90p_12_420.y4m", 12, VPX_IMG_FMT_I42016,
+    "9e6d8f6508c6e55625f6b697bc461cef"},
+  {"park_joy_90p_12_422.y4m", 12, VPX_IMG_FMT_I42216,
+    "b239c6b301c0b835485be349ca83a7e3"},
+  {"park_joy_90p_12_444.y4m", 12, VPX_IMG_FMT_I44416,
+    "5a6481a550821dab6d0192f5c63845e9"},
+};
+
+static void write_image_file(const vpx_image_t *img, FILE *file) {
+  int plane, y;
+  for (plane = 0; plane < 3; ++plane) {
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int bytes_per_sample = (img->fmt & VPX_IMG_FMT_HIGHBITDEPTH) ? 2 : 1;
+    const int h = (plane ? (img->d_h + img->y_chroma_shift) >>
+                   img->y_chroma_shift : img->d_h);
+    const int w = (plane ? (img->d_w + img->x_chroma_shift) >>
+                   img->x_chroma_shift : img->d_w);
+    for (y = 0; y < h; ++y) {
+      fwrite(buf, bytes_per_sample, w, file);
+      buf += stride;
+    }
+  }
+}
+
+class Y4mVideoSourceTest
+    : public ::testing::TestWithParam<Y4mTestParam>,
+      public ::libvpx_test::Y4mVideoSource {
+ protected:
+  Y4mVideoSourceTest() : Y4mVideoSource("", 0, 0) {}
+
+  virtual ~Y4mVideoSourceTest() {
+    CloseSource();
+  }
+
+  virtual void Init(const std::string &file_name, int limit) {
+    file_name_ = file_name;
+    start_ = 0;
+    limit_ = limit;
+    frame_ = 0;
+    Begin();
+  }
+
+  // Checks y4m header information
+  void HeaderChecks(unsigned int bit_depth, vpx_img_fmt_t fmt) {
+    ASSERT_TRUE(input_file_ != NULL);
+    ASSERT_EQ(y4m_.pic_w, (int)kWidth);
+    ASSERT_EQ(y4m_.pic_h, (int)kHeight);
+    ASSERT_EQ(img()->d_w, kWidth);
+    ASSERT_EQ(img()->d_h, kHeight);
+    ASSERT_EQ(y4m_.bit_depth, bit_depth);
+    ASSERT_EQ(y4m_.vpx_fmt, fmt);
+    if (fmt == VPX_IMG_FMT_I420 || fmt == VPX_IMG_FMT_I42016) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 3 / 2);
+      ASSERT_EQ(img()->x_chroma_shift, 1U);
+      ASSERT_EQ(img()->y_chroma_shift, 1U);
+    }
+    if (fmt == VPX_IMG_FMT_I422 || fmt == VPX_IMG_FMT_I42216) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 2);
+      ASSERT_EQ(img()->x_chroma_shift, 1U);
+      ASSERT_EQ(img()->y_chroma_shift, 0U);
+    }
+    if (fmt == VPX_IMG_FMT_I444 || fmt == VPX_IMG_FMT_I44416) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 3);
+      ASSERT_EQ(img()->x_chroma_shift, 0U);
+      ASSERT_EQ(img()->y_chroma_shift, 0U);
+    }
+  }
+
+  // Checks MD5 of the raw frame data
+  void Md5Check(const string &expected_md5) {
+    ASSERT_TRUE(input_file_ != NULL);
+    libvpx_test::MD5 md5;
+    for (unsigned int i = start_; i < limit_; i++) {
+      md5.Add(img());
+      Next();
+    }
+    ASSERT_EQ(string(md5.Get()), expected_md5);
+  }
+};
+
+TEST_P(Y4mVideoSourceTest, SourceTest) {
+  const Y4mTestParam t = GetParam();
+  Init(t.filename, kFrames);
+  HeaderChecks(t.bit_depth, t.format);
+  Md5Check(t.md5raw);
+}
+
+INSTANTIATE_TEST_CASE_P(C, Y4mVideoSourceTest,
+                        ::testing::ValuesIn(kY4mTestVectors));
+
+class Y4mVideoWriteTest
+    : public Y4mVideoSourceTest {
+ protected:
+  Y4mVideoWriteTest() {}
+
+  virtual ~Y4mVideoWriteTest() {
+    delete tmpfile_;
+    input_file_ = NULL;
+  }
+
+  void ReplaceInputFile(FILE *input_file) {
+    CloseSource();
+    frame_ = 0;
+    input_file_ = input_file;
+    rewind(input_file_);
+    ReadSourceToStart();
+  }
+
+  // Writes out a y4m file and then reads it back
+  void WriteY4mAndReadBack() {
+    ASSERT_TRUE(input_file_ != NULL);
+    char buf[Y4M_BUFFER_SIZE] = {0};
+    const struct VpxRational framerate = {y4m_.fps_n, y4m_.fps_d};
+    tmpfile_ = new libvpx_test::TempOutFile;
+    ASSERT_TRUE(tmpfile_->file() != NULL);
+    y4m_write_file_header(buf, sizeof(buf),
+                          kWidth, kHeight,
+                          &framerate, y4m_.vpx_fmt,
+                          y4m_.bit_depth);
+    fputs(buf, tmpfile_->file());
+    for (unsigned int i = start_; i < limit_; i++) {
+      y4m_write_frame_header(buf, sizeof(buf));
+      fputs(buf, tmpfile_->file());
+      write_image_file(img(), tmpfile_->file());
+      Next();
+    }
+    ReplaceInputFile(tmpfile_->file());
+  }
+
+  virtual void Init(const std::string &file_name, int limit) {
+    Y4mVideoSourceTest::Init(file_name, limit);
+    WriteY4mAndReadBack();
+  }
+  libvpx_test::TempOutFile *tmpfile_;
+};
+
+TEST_P(Y4mVideoWriteTest, WriteTest) {
+  const Y4mTestParam t = GetParam();
+  Init(t.filename, kFrames);
+  HeaderChecks(t.bit_depth, t.format);
+  Md5Check(t.md5raw);
+}
+
+INSTANTIATE_TEST_CASE_P(C, Y4mVideoWriteTest,
+                        ::testing::ValuesIn(kY4mTestVectors));
+}  // namespace
diff --git a/libvpx/libvpx/test/y4m_video_source.h b/libvpx/libvpx/test/y4m_video_source.h
new file mode 100644
index 0000000..03d9388
--- /dev/null
+++ b/libvpx/libvpx/test/y4m_video_source.h
@@ -0,0 +1,130 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef TEST_Y4M_VIDEO_SOURCE_H_
+#define TEST_Y4M_VIDEO_SOURCE_H_
+#include <algorithm>
+#include <string>
+
+#include "test/video_source.h"
+#include "./y4minput.h"
+
+namespace libvpx_test {
+
+// This class extends VideoSource to allow parsing of raw yv12
+// so that we can do actual file encodes.
+class Y4mVideoSource : public VideoSource {
+ public:
+  Y4mVideoSource(const std::string &file_name,
+                  unsigned int start, int limit)
+      : file_name_(file_name),
+        input_file_(NULL),
+        img_(new vpx_image_t()),
+        start_(start),
+        limit_(limit),
+        frame_(0),
+        framerate_numerator_(0),
+        framerate_denominator_(0),
+        y4m_() {
+  }
+
+  virtual ~Y4mVideoSource() {
+    vpx_img_free(img_.get());
+    CloseSource();
+  }
+
+  virtual void OpenSource() {
+    CloseSource();
+    input_file_ = OpenTestDataFile(file_name_);
+    ASSERT_TRUE(input_file_ != NULL) << "Input file open failed. Filename: "
+                                     << file_name_;
+  }
+
+  virtual void ReadSourceToStart() {
+    ASSERT_TRUE(input_file_ != NULL);
+    ASSERT_FALSE(y4m_input_open(&y4m_, input_file_, NULL, 0, 0));
+    framerate_numerator_ = y4m_.fps_n;
+    framerate_denominator_ = y4m_.fps_d;
+    frame_ = 0;
+    for (unsigned int i = 0; i < start_; i++) {
+      Next();
+    }
+    FillFrame();
+  }
+
+  virtual void Begin() {
+    OpenSource();
+    ReadSourceToStart();
+  }
+
+  virtual void Next() {
+    ++frame_;
+    FillFrame();
+  }
+
+  virtual vpx_image_t *img() const {
+    return (frame_ < limit_) ? img_.get() : NULL;
+  }
+
+  // Models a stream where Timebase = 1/FPS, so pts == frame.
+  virtual vpx_codec_pts_t pts() const { return frame_; }
+
+  virtual unsigned long duration() const { return 1; }
+
+  virtual vpx_rational_t timebase() const {
+    const vpx_rational_t t = { framerate_denominator_, framerate_numerator_ };
+    return t;
+  }
+
+  virtual unsigned int frame() const { return frame_; }
+
+  virtual unsigned int limit() const { return limit_; }
+
+  virtual void FillFrame() {
+    ASSERT_TRUE(input_file_ != NULL);
+    // Read a frame from input_file.
+    y4m_input_fetch_frame(&y4m_, input_file_, img_.get());
+  }
+
+  // Swap buffers with another y4m source. This allows reading a new frame
+  // while keeping the old frame around. A whole Y4mSource is required and
+  // not just a vpx_image_t because of how the y4m reader manipulates
+  // vpx_image_t internals,
+  void SwapBuffers(Y4mVideoSource *other) {
+    std::swap(other->y4m_.dst_buf, y4m_.dst_buf);
+    vpx_image_t *tmp;
+    tmp = other->img_.release();
+    other->img_.reset(img_.release());
+    img_.reset(tmp);
+  }
+
+ protected:
+  void CloseSource() {
+    y4m_input_close(&y4m_);
+    y4m_ = y4m_input();
+    if (input_file_ != NULL) {
+      fclose(input_file_);
+      input_file_ = NULL;
+    }
+  }
+
+  std::string file_name_;
+  FILE *input_file_;
+  testing::internal::scoped_ptr<vpx_image_t> img_;
+  unsigned int start_;
+  unsigned int limit_;
+  unsigned int frame_;
+  int framerate_numerator_;
+  int framerate_denominator_;
+  y4m_input y4m_;
+};
+
+}  // namespace libvpx_test
+
+#endif  // TEST_Y4M_VIDEO_SOURCE_H_
diff --git a/libvpx/libvpx/test/yuv_video_source.h b/libvpx/libvpx/test/yuv_video_source.h
new file mode 100644
index 0000000..3c852b2
--- /dev/null
+++ b/libvpx/libvpx/test/yuv_video_source.h
@@ -0,0 +1,151 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef TEST_YUV_VIDEO_SOURCE_H_
+#define TEST_YUV_VIDEO_SOURCE_H_
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+
+#include "test/video_source.h"
+#include "vpx/vpx_image.h"
+
+namespace libvpx_test {
+
+// This class extends VideoSource to allow parsing of raw YUV
+// formats of various color sampling and bit-depths so that we can
+// do actual file encodes.
+class YUVVideoSource : public VideoSource {
+ public:
+  YUVVideoSource(const std::string &file_name, vpx_img_fmt format,
+                 unsigned int width, unsigned int height,
+                 int rate_numerator, int rate_denominator,
+                 unsigned int start, int limit)
+      : file_name_(file_name),
+        input_file_(NULL),
+        img_(NULL),
+        start_(start),
+        limit_(limit),
+        frame_(0),
+        width_(0),
+        height_(0),
+        format_(VPX_IMG_FMT_NONE),
+        framerate_numerator_(rate_numerator),
+        framerate_denominator_(rate_denominator) {
+    // This initializes format_, raw_size_, width_, height_ and allocates img.
+    SetSize(width, height, format);
+  }
+
+  virtual ~YUVVideoSource() {
+    vpx_img_free(img_);
+    if (input_file_)
+      fclose(input_file_);
+  }
+
+  virtual void Begin() {
+    if (input_file_)
+      fclose(input_file_);
+    input_file_ = OpenTestDataFile(file_name_);
+    ASSERT_TRUE(input_file_ != NULL) << "Input file open failed. Filename: "
+                                     << file_name_;
+    if (start_)
+      fseek(input_file_, static_cast<unsigned>(raw_size_) * start_, SEEK_SET);
+
+    frame_ = start_;
+    FillFrame();
+  }
+
+  virtual void Next() {
+    ++frame_;
+    FillFrame();
+  }
+
+  virtual vpx_image_t *img() const { return (frame_ < limit_) ? img_ : NULL;  }
+
+  // Models a stream where Timebase = 1/FPS, so pts == frame.
+  virtual vpx_codec_pts_t pts() const { return frame_; }
+
+  virtual unsigned long duration() const { return 1; }
+
+  virtual vpx_rational_t timebase() const {
+    const vpx_rational_t t = { framerate_denominator_, framerate_numerator_ };
+    return t;
+  }
+
+  virtual unsigned int frame() const { return frame_; }
+
+  virtual unsigned int limit() const { return limit_; }
+
+  virtual void SetSize(unsigned int width, unsigned int height,
+                       vpx_img_fmt format) {
+    if (width != width_ || height != height_ || format != format_) {
+      vpx_img_free(img_);
+      img_ = vpx_img_alloc(NULL, format, width, height, 1);
+      ASSERT_TRUE(img_ != NULL);
+      width_ = width;
+      height_ = height;
+      format_ = format;
+      switch (format) {
+        case VPX_IMG_FMT_I420:
+          raw_size_ = width * height * 3 / 2;
+          break;
+        case VPX_IMG_FMT_I422:
+          raw_size_ = width * height * 2;
+          break;
+        case VPX_IMG_FMT_I440:
+          raw_size_ = width * height * 2;
+          break;
+        case VPX_IMG_FMT_I444:
+          raw_size_ = width * height * 3;
+          break;
+        case VPX_IMG_FMT_I42016:
+          raw_size_ = width * height * 3;
+          break;
+        case VPX_IMG_FMT_I42216:
+          raw_size_ = width * height * 4;
+          break;
+        case VPX_IMG_FMT_I44016:
+          raw_size_ = width * height * 4;
+          break;
+        case VPX_IMG_FMT_I44416:
+          raw_size_ = width * height * 6;
+          break;
+        default:
+          ASSERT_TRUE(0);
+      }
+    }
+  }
+
+  virtual void FillFrame() {
+    ASSERT_TRUE(input_file_ != NULL);
+    // Read a frame from input_file.
+    if (fread(img_->img_data, raw_size_, 1, input_file_) == 0) {
+      limit_ = frame_;
+    }
+  }
+
+ protected:
+  std::string file_name_;
+  FILE *input_file_;
+  vpx_image_t *img_;
+  size_t raw_size_;
+  unsigned int start_;
+  unsigned int limit_;
+  unsigned int frame_;
+  unsigned int width_;
+  unsigned int height_;
+  vpx_img_fmt format_;
+  int framerate_numerator_;
+  int framerate_denominator_;
+};
+
+}  // namespace libvpx_test
+
+#endif  // TEST_YUV_VIDEO_SOURCE_H_
diff --git a/libvpx/libvpx/third_party/googletest/README.libvpx b/libvpx/libvpx/third_party/googletest/README.libvpx
new file mode 100644
index 0000000..1eca78d
--- /dev/null
+++ b/libvpx/libvpx/third_party/googletest/README.libvpx
@@ -0,0 +1,19 @@
+URL: http://code.google.com/p/googletest/
+Version: 1.7.0
+License: BSD
+License File: COPYING
+
+Description:
+Google's framework for writing C++ tests on a variety of platforms
+(Linux, Mac OS X, Windows, Windows CE, Symbian, etc).  Based on the
+xUnit architecture.  Supports automatic test discovery, a rich set of
+assertions, user-defined assertions, death tests, fatal and non-fatal
+failures, various options for running the tests, and XML test report
+generation.
+
+Local Modifications:
+- Removed unused declarations of kPathSeparatorString to have warning
+  free build.
+- Added GTEST_ATTRIBUTE_UNUSED_ to test registering dummies in TEST_P
+  and INSTANTIATE_TEST_CASE_P to remove warnings about unused variables
+  under GCC 5.
\ No newline at end of file
diff --git a/libvpx/libvpx/third_party/googletest/gtest.mk b/libvpx/libvpx/third_party/googletest/gtest.mk
new file mode 100644
index 0000000..0de3113
--- /dev/null
+++ b/libvpx/libvpx/third_party/googletest/gtest.mk
@@ -0,0 +1 @@
+GTEST_SRCS-yes += src/gtest-all.cc 
diff --git a/libvpx/libvpx/third_party/googletest/src/CHANGES b/libvpx/libvpx/third_party/googletest/src/CHANGES
new file mode 100644
index 0000000..0552132
--- /dev/null
+++ b/libvpx/libvpx/third_party/googletest/src/CHANGES
@@ -0,0 +1,157 @@
+Changes for 1.7.0:
+
+* New feature: death tests are supported on OpenBSD and in iOS
+  simulator now.
+* New feature: Google Test now implements a protocol to allow
+  a test runner to detect that a test program has exited
+  prematurely and report it as a failure (before it would be
+  falsely reported as a success if the exit code is 0).
+* New feature: Test::RecordProperty() can now be used outside of the
+  lifespan of a test method, in which case it will be attributed to
+  the current test case or the test program in the XML report.
+* New feature (potentially breaking): --gtest_list_tests now prints
+  the type parameters and value parameters for each test.
+* Improvement: char pointers and char arrays are now escaped properly
+  in failure messages.
+* Improvement: failure summary in XML reports now includes file and
+  line information.
+* Improvement: the <testsuites> XML element now has a timestamp attribute.
+* Improvement: When --gtest_filter is specified, XML report now doesn't
+  contain information about tests that are filtered out.
+* Fixed the bug where long --gtest_filter flag values are truncated in
+  death tests.
+* Potentially breaking change: RUN_ALL_TESTS() is now implemented as a
+  function instead of a macro in order to work better with Clang.
+* Compatibility fixes with C++ 11 and various platforms.
+* Bug/warning fixes.
+
+Changes for 1.6.0:
+
+* New feature: ADD_FAILURE_AT() for reporting a test failure at the
+  given source location -- useful for writing testing utilities.
+* New feature: the universal value printer is moved from Google Mock
+  to Google Test.
+* New feature: type parameters and value parameters are reported in
+  the XML report now.
+* A gtest_disable_pthreads CMake option.
+* Colored output works in GNU Screen sessions now.
+* Parameters of value-parameterized tests are now printed in the
+  textual output.
+* Failures from ad hoc test assertions run before RUN_ALL_TESTS() are
+  now correctly reported.
+* Arguments of ASSERT_XY and EXPECT_XY no longer need to support << to
+  ostream.
+* More complete handling of exceptions.
+* GTEST_ASSERT_XY can be used instead of ASSERT_XY in case the latter
+  name is already used by another library.
+* --gtest_catch_exceptions is now true by default, allowing a test
+  program to continue after an exception is thrown.
+* Value-parameterized test fixtures can now derive from Test and
+  WithParamInterface<T> separately, easing conversion of legacy tests.
+* Death test messages are clearly marked to make them more
+  distinguishable from other messages.
+* Compatibility fixes for Android, Google Native Client, MinGW, HP UX,
+  PowerPC, Lucid autotools, libCStd, Sun C++, Borland C++ Builder (Code Gear),
+  IBM XL C++ (Visual Age C++), and C++0x.
+* Bug fixes and implementation clean-ups.
+* Potentially incompatible changes: disables the harmful 'make install'
+  command in autotools.
+
+Changes for 1.5.0:
+
+ * New feature: assertions can be safely called in multiple threads
+   where the pthreads library is available.
+ * New feature: predicates used inside EXPECT_TRUE() and friends
+   can now generate custom failure messages.
+ * New feature: Google Test can now be compiled as a DLL.
+ * New feature: fused source files are included.
+ * New feature: prints help when encountering unrecognized Google Test flags.
+ * Experimental feature: CMake build script (requires CMake 2.6.4+).
+ * Experimental feature: the Pump script for meta programming.
+ * double values streamed to an assertion are printed with enough precision
+   to differentiate any two different values.
+ * Google Test now works on Solaris and AIX.
+ * Build and test script improvements.
+ * Bug fixes and implementation clean-ups.
+
+ Potentially breaking changes:
+
+ * Stopped supporting VC++ 7.1 with exceptions disabled.
+ * Dropped support for 'make install'.
+
+Changes for 1.4.0:
+
+ * New feature: the event listener API
+ * New feature: test shuffling
+ * New feature: the XML report format is closer to junitreport and can
+   be parsed by Hudson now.
+ * New feature: when a test runs under Visual Studio, its failures are
+   integrated in the IDE.
+ * New feature: /MD(d) versions of VC++ projects.
+ * New feature: elapsed time for the tests is printed by default.
+ * New feature: comes with a TR1 tuple implementation such that Boost
+   is no longer needed for Combine().
+ * New feature: EXPECT_DEATH_IF_SUPPORTED macro and friends.
+ * New feature: the Xcode project can now produce static gtest
+   libraries in addition to a framework.
+ * Compatibility fixes for Solaris, Cygwin, minGW, Windows Mobile,
+   Symbian, gcc, and C++Builder.
+ * Bug fixes and implementation clean-ups.
+
+Changes for 1.3.0:
+
+ * New feature: death tests on Windows, Cygwin, and Mac.
+ * New feature: ability to use Google Test assertions in other testing
+   frameworks.
+ * New feature: ability to run disabled test via
+   --gtest_also_run_disabled_tests.
+ * New feature: the --help flag for printing the usage.
+ * New feature: access to Google Test flag values in user code.
+ * New feature: a script that packs Google Test into one .h and one
+   .cc file for easy deployment.
+ * New feature: support for distributing test functions to multiple
+   machines (requires support from the test runner).
+ * Bug fixes and implementation clean-ups.
+
+Changes for 1.2.1:
+
+ * Compatibility fixes for Linux IA-64 and IBM z/OS.
+ * Added support for using Boost and other TR1 implementations.
+ * Changes to the build scripts to support upcoming release of Google C++
+   Mocking Framework.
+ * Added Makefile to the distribution package.
+ * Improved build instructions in README.
+
+Changes for 1.2.0:
+
+ * New feature: value-parameterized tests.
+ * New feature: the ASSERT/EXPECT_(NON)FATAL_FAILURE(_ON_ALL_THREADS)
+   macros.
+ * Changed the XML report format to match JUnit/Ant's.
+ * Added tests to the Xcode project.
+ * Added scons/SConscript for building with SCons.
+ * Added src/gtest-all.cc for building Google Test from a single file.
+ * Fixed compatibility with Solaris and z/OS.
+ * Enabled running Python tests on systems with python 2.3 installed,
+   e.g. Mac OS X 10.4.
+ * Bug fixes.
+
+Changes for 1.1.0:
+
+ * New feature: type-parameterized tests.
+ * New feature: exception assertions.
+ * New feature: printing elapsed time of tests.
+ * Improved the robustness of death tests.
+ * Added an Xcode project and samples.
+ * Adjusted the output format on Windows to be understandable by Visual Studio.
+ * Minor bug fixes.
+
+Changes for 1.0.1:
+
+ * Added project files for Visual Studio 7.1.
+ * Fixed issues with compiling on Mac OS X.
+ * Fixed issues with compiling on Cygwin.
+
+Changes for 1.0.0:
+
+ * Initial Open Source release of Google Test
diff --git a/libvpx/libvpx/third_party/googletest/src/CONTRIBUTORS b/libvpx/libvpx/third_party/googletest/src/CONTRIBUTORS
new file mode 100644
index 0000000..feae2fc
--- /dev/null
+++ b/libvpx/libvpx/third_party/googletest/src/CONTRIBUTORS
@@ -0,0 +1,37 @@
+# This file contains a list of people who've made non-trivial
+# contribution to the Google C++ Testing Framework project.  People
+# who commit code to the project are encouraged to add their names
+# here.  Please keep the list sorted by first names.
+
+Ajay Joshi <jaj@google.com>
+Balázs Dán <balazs.dan@gmail.com>
+Bharat Mediratta <bharat@menalto.com>
+Chandler Carruth <chandlerc@google.com>
+Chris Prince <cprince@google.com>
+Chris Taylor <taylorc@google.com>
+Dan Egnor <egnor@google.com>
+Eric Roman <eroman@chromium.org>
+Hady Zalek <hady.zalek@gmail.com>
+Jeffrey Yasskin <jyasskin@google.com>
+Jói Sigurðsson <joi@google.com>
+Keir Mierle <mierle@gmail.com>
+Keith Ray <keith.ray@gmail.com>
+Kenton Varda <kenton@google.com>
+Manuel Klimek <klimek@google.com>
+Markus Heule <markus.heule@gmail.com>
+Mika Raento <mikie@iki.fi>
+Miklós Fazekas <mfazekas@szemafor.com>
+Pasi Valminen <pasi.valminen@gmail.com>
+Patrick Hanna <phanna@google.com>
+Patrick Riley <pfr@google.com>
+Peter Kaminski <piotrk@google.com>
+Preston Jackson <preston.a.jackson@gmail.com>
+Rainer Klaffenboeck <rainer.klaffenboeck@dynatrace.com>
+Russ Cox <rsc@google.com>
+Russ Rufer <russ@pentad.com>
+Sean Mcafee <eefacm@gmail.com>
+Sigurður Ásgeirsson <siggi@google.com>
+Tracy Bialik <tracy@pentad.com>
+Vadim Berman <vadimb@google.com>
+Vlad Losev <vladl@google.com>
+Zhanyong Wan <wan@google.com>
diff --git a/libvpx/libvpx/third_party/googletest/src/LICENSE b/libvpx/libvpx/third_party/googletest/src/LICENSE
new file mode 100644
index 0000000..1941a11
--- /dev/null
+++ b/libvpx/libvpx/third_party/googletest/src/LICENSE
@@ -0,0 +1,28 @@
+Copyright 2008, Google Inc.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+    * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+    * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/libvpx/libvpx/third_party/googletest/src/README b/libvpx/libvpx/third_party/googletest/src/README
new file mode 100644
index 0000000..26f35a8
--- /dev/null
+++ b/libvpx/libvpx/third_party/googletest/src/README
@@ -0,0 +1,435 @@
+Google C++ Testing Framework
+============================
+
+http://code.google.com/p/googletest/
+
+Overview
+--------
+
+Google's framework for writing C++ tests on a variety of platforms
+(Linux, Mac OS X, Windows, Windows CE, Symbian, etc).  Based on the
+xUnit architecture.  Supports automatic test discovery, a rich set of
+assertions, user-defined assertions, death tests, fatal and non-fatal
+failures, various options for running the tests, and XML test report
+generation.
+
+Please see the project page above for more information as well as the
+mailing list for questions, discussions, and development.  There is
+also an IRC channel on OFTC (irc.oftc.net) #gtest available.  Please
+join us!
+
+Requirements for End Users
+--------------------------
+
+Google Test is designed to have fairly minimal requirements to build
+and use with your projects, but there are some.  Currently, we support
+Linux, Windows, Mac OS X, and Cygwin.  We will also make our best
+effort to support other platforms (e.g. Solaris, AIX, and z/OS).
+However, since core members of the Google Test project have no access
+to these platforms, Google Test may have outstanding issues there.  If
+you notice any problems on your platform, please notify
+googletestframework@googlegroups.com.  Patches for fixing them are
+even more welcome!
+
+### Linux Requirements ###
+
+These are the base requirements to build and use Google Test from a source
+package (as described below):
+  * GNU-compatible Make or gmake
+  * POSIX-standard shell
+  * POSIX(-2) Regular Expressions (regex.h)
+  * A C++98-standard-compliant compiler
+
+### Windows Requirements ###
+
+  * Microsoft Visual C++ 7.1 or newer
+
+### Cygwin Requirements ###
+
+  * Cygwin 1.5.25-14 or newer
+
+### Mac OS X Requirements ###
+
+  * Mac OS X 10.4 Tiger or newer
+  * Developer Tools Installed
+
+Also, you'll need CMake 2.6.4 or higher if you want to build the
+samples using the provided CMake script, regardless of the platform.
+
+Requirements for Contributors
+-----------------------------
+
+We welcome patches.  If you plan to contribute a patch, you need to
+build Google Test and its own tests from an SVN checkout (described
+below), which has further requirements:
+
+  * Python version 2.3 or newer (for running some of the tests and
+    re-generating certain source files from templates)
+  * CMake 2.6.4 or newer
+
+Getting the Source
+------------------
+
+There are two primary ways of getting Google Test's source code: you
+can download a stable source release in your preferred archive format,
+or directly check out the source from our Subversion (SVN) repositary.
+The SVN checkout requires a few extra steps and some extra software
+packages on your system, but lets you track the latest development and
+make patches much more easily, so we highly encourage it.
+
+### Source Package ###
+
+Google Test is released in versioned source packages which can be
+downloaded from the download page [1].  Several different archive
+formats are provided, but the only difference is the tools used to
+manipulate them, and the size of the resulting file.  Download
+whichever you are most comfortable with.
+
+  [1] http://code.google.com/p/googletest/downloads/list
+
+Once the package is downloaded, expand it using whichever tools you
+prefer for that type.  This will result in a new directory with the
+name "gtest-X.Y.Z" which contains all of the source code.  Here are
+some examples on Linux:
+
+  tar -xvzf gtest-X.Y.Z.tar.gz
+  tar -xvjf gtest-X.Y.Z.tar.bz2
+  unzip gtest-X.Y.Z.zip
+
+### SVN Checkout ###
+
+To check out the main branch (also known as the "trunk") of Google
+Test, run the following Subversion command:
+
+  svn checkout http://googletest.googlecode.com/svn/trunk/ gtest-svn
+
+Setting up the Build
+--------------------
+
+To build Google Test and your tests that use it, you need to tell your
+build system where to find its headers and source files.  The exact
+way to do it depends on which build system you use, and is usually
+straightforward.
+
+### Generic Build Instructions ###
+
+Suppose you put Google Test in directory ${GTEST_DIR}.  To build it,
+create a library build target (or a project as called by Visual Studio
+and Xcode) to compile
+
+  ${GTEST_DIR}/src/gtest-all.cc
+
+with ${GTEST_DIR}/include in the system header search path and ${GTEST_DIR}
+in the normal header search path.  Assuming a Linux-like system and gcc,
+something like the following will do:
+
+  g++ -isystem ${GTEST_DIR}/include -I${GTEST_DIR} \
+      -pthread -c ${GTEST_DIR}/src/gtest-all.cc
+  ar -rv libgtest.a gtest-all.o
+
+(We need -pthread as Google Test uses threads.)
+
+Next, you should compile your test source file with
+${GTEST_DIR}/include in the system header search path, and link it
+with gtest and any other necessary libraries:
+
+  g++ -isystem ${GTEST_DIR}/include -pthread path/to/your_test.cc libgtest.a \
+      -o your_test
+
+As an example, the make/ directory contains a Makefile that you can
+use to build Google Test on systems where GNU make is available
+(e.g. Linux, Mac OS X, and Cygwin).  It doesn't try to build Google
+Test's own tests.  Instead, it just builds the Google Test library and
+a sample test.  You can use it as a starting point for your own build
+script.
+
+If the default settings are correct for your environment, the
+following commands should succeed:
+
+  cd ${GTEST_DIR}/make
+  make
+  ./sample1_unittest
+
+If you see errors, try to tweak the contents of make/Makefile to make
+them go away.  There are instructions in make/Makefile on how to do
+it.
+
+### Using CMake ###
+
+Google Test comes with a CMake build script (CMakeLists.txt) that can
+be used on a wide range of platforms ("C" stands for cross-platofrm.).
+If you don't have CMake installed already, you can download it for
+free from http://www.cmake.org/.
+
+CMake works by generating native makefiles or build projects that can
+be used in the compiler environment of your choice.  The typical
+workflow starts with:
+
+  mkdir mybuild       # Create a directory to hold the build output.
+  cd mybuild
+  cmake ${GTEST_DIR}  # Generate native build scripts.
+
+If you want to build Google Test's samples, you should replace the
+last command with
+
+  cmake -Dgtest_build_samples=ON ${GTEST_DIR}
+
+If you are on a *nix system, you should now see a Makefile in the
+current directory.  Just type 'make' to build gtest.
+
+If you use Windows and have Vistual Studio installed, a gtest.sln file
+and several .vcproj files will be created.  You can then build them
+using Visual Studio.
+
+On Mac OS X with Xcode installed, a .xcodeproj file will be generated.
+
+### Legacy Build Scripts ###
+
+Before settling on CMake, we have been providing hand-maintained build
+projects/scripts for Visual Studio, Xcode, and Autotools.  While we
+continue to provide them for convenience, they are not actively
+maintained any more.  We highly recommend that you follow the
+instructions in the previous two sections to integrate Google Test
+with your existing build system.
+
+If you still need to use the legacy build scripts, here's how:
+
+The msvc\ folder contains two solutions with Visual C++ projects.
+Open the gtest.sln or gtest-md.sln file using Visual Studio, and you
+are ready to build Google Test the same way you build any Visual
+Studio project.  Files that have names ending with -md use DLL
+versions of Microsoft runtime libraries (the /MD or the /MDd compiler
+option).  Files without that suffix use static versions of the runtime
+libraries (the /MT or the /MTd option).  Please note that one must use
+the same option to compile both gtest and the test code.  If you use
+Visual Studio 2005 or above, we recommend the -md version as /MD is
+the default for new projects in these versions of Visual Studio.
+
+On Mac OS X, open the gtest.xcodeproj in the xcode/ folder using
+Xcode.  Build the "gtest" target.  The universal binary framework will
+end up in your selected build directory (selected in the Xcode
+"Preferences..." -> "Building" pane and defaults to xcode/build).
+Alternatively, at the command line, enter:
+
+  xcodebuild
+
+This will build the "Release" configuration of gtest.framework in your
+default build location.  See the "xcodebuild" man page for more
+information about building different configurations and building in
+different locations.
+
+If you wish to use the Google Test Xcode project with Xcode 4.x and
+above, you need to either:
+ * update the SDK configuration options in xcode/Config/General.xconfig.
+   Comment options SDKROOT, MACOS_DEPLOYMENT_TARGET, and GCC_VERSION. If
+   you choose this route you lose the ability to target earlier versions
+   of MacOS X.
+ * Install an SDK for an earlier version. This doesn't appear to be
+   supported by Apple, but has been reported to work
+   (http://stackoverflow.com/questions/5378518).
+
+Tweaking Google Test
+--------------------
+
+Google Test can be used in diverse environments.  The default
+configuration may not work (or may not work well) out of the box in
+some environments.  However, you can easily tweak Google Test by
+defining control macros on the compiler command line.  Generally,
+these macros are named like GTEST_XYZ and you define them to either 1
+or 0 to enable or disable a certain feature.
+
+We list the most frequently used macros below.  For a complete list,
+see file include/gtest/internal/gtest-port.h.
+
+### Choosing a TR1 Tuple Library ###
+
+Some Google Test features require the C++ Technical Report 1 (TR1)
+tuple library, which is not yet available with all compilers.  The
+good news is that Google Test implements a subset of TR1 tuple that's
+enough for its own need, and will automatically use this when the
+compiler doesn't provide TR1 tuple.
+
+Usually you don't need to care about which tuple library Google Test
+uses.  However, if your project already uses TR1 tuple, you need to
+tell Google Test to use the same TR1 tuple library the rest of your
+project uses, or the two tuple implementations will clash.  To do
+that, add
+
+  -DGTEST_USE_OWN_TR1_TUPLE=0
+
+to the compiler flags while compiling Google Test and your tests.  If
+you want to force Google Test to use its own tuple library, just add
+
+  -DGTEST_USE_OWN_TR1_TUPLE=1
+
+to the compiler flags instead.
+
+If you don't want Google Test to use tuple at all, add
+
+  -DGTEST_HAS_TR1_TUPLE=0
+
+and all features using tuple will be disabled.
+
+### Multi-threaded Tests ###
+
+Google Test is thread-safe where the pthread library is available.
+After #include "gtest/gtest.h", you can check the GTEST_IS_THREADSAFE
+macro to see whether this is the case (yes if the macro is #defined to
+1, no if it's undefined.).
+
+If Google Test doesn't correctly detect whether pthread is available
+in your environment, you can force it with
+
+  -DGTEST_HAS_PTHREAD=1
+
+or
+
+  -DGTEST_HAS_PTHREAD=0
+
+When Google Test uses pthread, you may need to add flags to your
+compiler and/or linker to select the pthread library, or you'll get
+link errors.  If you use the CMake script or the deprecated Autotools
+script, this is taken care of for you.  If you use your own build
+script, you'll need to read your compiler and linker's manual to
+figure out what flags to add.
+
+### As a Shared Library (DLL) ###
+
+Google Test is compact, so most users can build and link it as a
+static library for the simplicity.  You can choose to use Google Test
+as a shared library (known as a DLL on Windows) if you prefer.
+
+To compile *gtest* as a shared library, add
+
+  -DGTEST_CREATE_SHARED_LIBRARY=1
+
+to the compiler flags.  You'll also need to tell the linker to produce
+a shared library instead - consult your linker's manual for how to do
+it.
+
+To compile your *tests* that use the gtest shared library, add
+
+  -DGTEST_LINKED_AS_SHARED_LIBRARY=1
+
+to the compiler flags.
+
+Note: while the above steps aren't technically necessary today when
+using some compilers (e.g. GCC), they may become necessary in the
+future, if we decide to improve the speed of loading the library (see
+http://gcc.gnu.org/wiki/Visibility for details).  Therefore you are
+recommended to always add the above flags when using Google Test as a
+shared library.  Otherwise a future release of Google Test may break
+your build script.
+
+### Avoiding Macro Name Clashes ###
+
+In C++, macros don't obey namespaces.  Therefore two libraries that
+both define a macro of the same name will clash if you #include both
+definitions.  In case a Google Test macro clashes with another
+library, you can force Google Test to rename its macro to avoid the
+conflict.
+
+Specifically, if both Google Test and some other code define macro
+FOO, you can add
+
+  -DGTEST_DONT_DEFINE_FOO=1
+
+to the compiler flags to tell Google Test to change the macro's name
+from FOO to GTEST_FOO.  Currently FOO can be FAIL, SUCCEED, or TEST.
+For example, with -DGTEST_DONT_DEFINE_TEST=1, you'll need to write
+
+  GTEST_TEST(SomeTest, DoesThis) { ... }
+
+instead of
+
+  TEST(SomeTest, DoesThis) { ... }
+
+in order to define a test.
+
+Upgrating from an Earlier Version
+---------------------------------
+
+We strive to keep Google Test releases backward compatible.
+Sometimes, though, we have to make some breaking changes for the
+users' long-term benefits.  This section describes what you'll need to
+do if you are upgrading from an earlier version of Google Test.
+
+### Upgrading from 1.3.0 or Earlier ###
+
+You may need to explicitly enable or disable Google Test's own TR1
+tuple library.  See the instructions in section "Choosing a TR1 Tuple
+Library".
+
+### Upgrading from 1.4.0 or Earlier ###
+
+The Autotools build script (configure + make) is no longer officially
+supportted.  You are encouraged to migrate to your own build system or
+use CMake.  If you still need to use Autotools, you can find
+instructions in the README file from Google Test 1.4.0.
+
+On platforms where the pthread library is available, Google Test uses
+it in order to be thread-safe.  See the "Multi-threaded Tests" section
+for what this means to your build script.
+
+If you use Microsoft Visual C++ 7.1 with exceptions disabled, Google
+Test will no longer compile.  This should affect very few people, as a
+large portion of STL (including <string>) doesn't compile in this mode
+anyway.  We decided to stop supporting it in order to greatly simplify
+Google Test's implementation.
+
+Developing Google Test
+----------------------
+
+This section discusses how to make your own changes to Google Test.
+
+### Testing Google Test Itself ###
+
+To make sure your changes work as intended and don't break existing
+functionality, you'll want to compile and run Google Test's own tests.
+For that you can use CMake:
+
+  mkdir mybuild
+  cd mybuild
+  cmake -Dgtest_build_tests=ON ${GTEST_DIR}
+
+Make sure you have Python installed, as some of Google Test's tests
+are written in Python.  If the cmake command complains about not being
+able to find Python ("Could NOT find PythonInterp (missing:
+PYTHON_EXECUTABLE)"), try telling it explicitly where your Python
+executable can be found:
+
+  cmake -DPYTHON_EXECUTABLE=path/to/python -Dgtest_build_tests=ON ${GTEST_DIR}
+
+Next, you can build Google Test and all of its own tests.  On *nix,
+this is usually done by 'make'.  To run the tests, do
+
+  make test
+
+All tests should pass.
+
+### Regenerating Source Files ###
+
+Some of Google Test's source files are generated from templates (not
+in the C++ sense) using a script.  A template file is named FOO.pump,
+where FOO is the name of the file it will generate.  For example, the
+file include/gtest/internal/gtest-type-util.h.pump is used to generate
+gtest-type-util.h in the same directory.
+
+Normally you don't need to worry about regenerating the source files,
+unless you need to modify them.  In that case, you should modify the
+corresponding .pump files instead and run the pump.py Python script to
+regenerate them.  You can find pump.py in the scripts/ directory.
+Read the Pump manual [2] for how to use it.
+
+  [2] http://code.google.com/p/googletest/wiki/PumpManual
+
+### Contributing a Patch ###
+
+We welcome patches.  Please read the Google Test developer's guide [3]
+for how you can contribute.  In particular, make sure you have signed
+the Contributor License Agreement, or we won't be able to accept the
+patch.
+
+  [3] http://code.google.com/p/googletest/wiki/GoogleTestDevGuide
+
+Happy testing!
diff --git a/libvpx/libvpx/third_party/googletest/src/include/gtest/gtest.h b/libvpx/libvpx/third_party/googletest/src/include/gtest/gtest.h
new file mode 100644
index 0000000..581a44e
--- /dev/null
+++ b/libvpx/libvpx/third_party/googletest/src/include/gtest/gtest.h
@@ -0,0 +1,20061 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines the public API for Google Test.  It should be
+// included by any test program that uses Google Test.
+//
+// IMPORTANT NOTE: Due to limitation of the C++ language, we have to
+// leave some internal implementation details in this header file.
+// They are clearly marked by comments like this:
+//
+//   // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+//
+// Such code is NOT meant to be used by a user directly, and is subject
+// to CHANGE WITHOUT NOTICE.  Therefore DO NOT DEPEND ON IT in a user
+// program!
+//
+// Acknowledgment: Google Test borrowed the idea of automatic test
+// registration from Barthelemy Dagenais' (barthelemy@prologique.com)
+// easyUnit framework.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_H_
+
+#include <limits>
+#include <ostream>
+#include <vector>
+
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan), eefacm@gmail.com (Sean Mcafee)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file declares functions and macros used internally by
+// Google Test.  They are subject to change without notice.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_INTERNAL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_INTERNAL_H_
+
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan)
+//
+// Low-level types and utilities for porting Google Test to various
+// platforms.  They are subject to change without notice.  DO NOT USE
+// THEM IN USER CODE.
+//
+// This file is fundamental to Google Test.  All other Google Test source
+// files are expected to #include this.  Therefore, it cannot #include
+// any other Google Test header.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_H_
+
+// The user can define the following macros in the build script to
+// control Google Test's behavior.  If the user doesn't define a macro
+// in this list, Google Test will define it.
+//
+//   GTEST_HAS_CLONE          - Define it to 1/0 to indicate that clone(2)
+//                              is/isn't available.
+//   GTEST_HAS_EXCEPTIONS     - Define it to 1/0 to indicate that exceptions
+//                              are enabled.
+//   GTEST_HAS_GLOBAL_STRING  - Define it to 1/0 to indicate that ::string
+//                              is/isn't available (some systems define
+//                              ::string, which is different to std::string).
+//   GTEST_HAS_GLOBAL_WSTRING - Define it to 1/0 to indicate that ::string
+//                              is/isn't available (some systems define
+//                              ::wstring, which is different to std::wstring).
+//   GTEST_HAS_POSIX_RE       - Define it to 1/0 to indicate that POSIX regular
+//                              expressions are/aren't available.
+//   GTEST_HAS_PTHREAD        - Define it to 1/0 to indicate that <pthread.h>
+//                              is/isn't available.
+//   GTEST_HAS_RTTI           - Define it to 1/0 to indicate that RTTI is/isn't
+//                              enabled.
+//   GTEST_HAS_STD_WSTRING    - Define it to 1/0 to indicate that
+//                              std::wstring does/doesn't work (Google Test can
+//                              be used where std::wstring is unavailable).
+//   GTEST_HAS_TR1_TUPLE      - Define it to 1/0 to indicate tr1::tuple
+//                              is/isn't available.
+//   GTEST_HAS_SEH            - Define it to 1/0 to indicate whether the
+//                              compiler supports Microsoft's "Structured
+//                              Exception Handling".
+//   GTEST_HAS_STREAM_REDIRECTION
+//                            - Define it to 1/0 to indicate whether the
+//                              platform supports I/O stream redirection using
+//                              dup() and dup2().
+//   GTEST_USE_OWN_TR1_TUPLE  - Define it to 1/0 to indicate whether Google
+//                              Test's own tr1 tuple implementation should be
+//                              used.  Unused when the user sets
+//                              GTEST_HAS_TR1_TUPLE to 0.
+//   GTEST_LANG_CXX11         - Define it to 1/0 to indicate that Google Test
+//                              is building in C++11/C++98 mode.
+//   GTEST_LINKED_AS_SHARED_LIBRARY
+//                            - Define to 1 when compiling tests that use
+//                              Google Test as a shared library (known as
+//                              DLL on Windows).
+//   GTEST_CREATE_SHARED_LIBRARY
+//                            - Define to 1 when compiling Google Test itself
+//                              as a shared library.
+
+// This header defines the following utilities:
+//
+// Macros indicating the current platform (defined to 1 if compiled on
+// the given platform; otherwise undefined):
+//   GTEST_OS_AIX      - IBM AIX
+//   GTEST_OS_CYGWIN   - Cygwin
+//   GTEST_OS_HPUX     - HP-UX
+//   GTEST_OS_LINUX    - Linux
+//     GTEST_OS_LINUX_ANDROID - Google Android
+//   GTEST_OS_MAC      - Mac OS X
+//     GTEST_OS_IOS    - iOS
+//       GTEST_OS_IOS_SIMULATOR - iOS simulator
+//   GTEST_OS_NACL     - Google Native Client (NaCl)
+//   GTEST_OS_OPENBSD  - OpenBSD
+//   GTEST_OS_QNX      - QNX
+//   GTEST_OS_SOLARIS  - Sun Solaris
+//   GTEST_OS_SYMBIAN  - Symbian
+//   GTEST_OS_WINDOWS  - Windows (Desktop, MinGW, or Mobile)
+//     GTEST_OS_WINDOWS_DESKTOP  - Windows Desktop
+//     GTEST_OS_WINDOWS_MINGW    - MinGW
+//     GTEST_OS_WINDOWS_MOBILE   - Windows Mobile
+//   GTEST_OS_ZOS      - z/OS
+//
+// Among the platforms, Cygwin, Linux, Max OS X, and Windows have the
+// most stable support.  Since core members of the Google Test project
+// don't have access to other platforms, support for them may be less
+// stable.  If you notice any problems on your platform, please notify
+// googletestframework@googlegroups.com (patches for fixing them are
+// even more welcome!).
+//
+// Note that it is possible that none of the GTEST_OS_* macros are defined.
+//
+// Macros indicating available Google Test features (defined to 1 if
+// the corresponding feature is supported; otherwise undefined):
+//   GTEST_HAS_COMBINE      - the Combine() function (for value-parameterized
+//                            tests)
+//   GTEST_HAS_DEATH_TEST   - death tests
+//   GTEST_HAS_PARAM_TEST   - value-parameterized tests
+//   GTEST_HAS_TYPED_TEST   - typed tests
+//   GTEST_HAS_TYPED_TEST_P - type-parameterized tests
+//   GTEST_USES_POSIX_RE    - enhanced POSIX regex is used. Do not confuse with
+//                            GTEST_HAS_POSIX_RE (see above) which users can
+//                            define themselves.
+//   GTEST_USES_SIMPLE_RE   - our own simple regex is used;
+//                            the above two are mutually exclusive.
+//   GTEST_CAN_COMPARE_NULL - accepts untyped NULL in EXPECT_EQ().
+//
+// Macros for basic C++ coding:
+//   GTEST_AMBIGUOUS_ELSE_BLOCKER_ - for disabling a gcc warning.
+//   GTEST_ATTRIBUTE_UNUSED_  - declares that a class' instances or a
+//                              variable don't have to be used.
+//   GTEST_DISALLOW_ASSIGN_   - disables operator=.
+//   GTEST_DISALLOW_COPY_AND_ASSIGN_ - disables copy ctor and operator=.
+//   GTEST_MUST_USE_RESULT_   - declares that a function's result must be used.
+//
+// Synchronization:
+//   Mutex, MutexLock, ThreadLocal, GetThreadCount()
+//                  - synchronization primitives.
+//   GTEST_IS_THREADSAFE - defined to 1 to indicate that the above
+//                         synchronization primitives have real implementations
+//                         and Google Test is thread-safe; or 0 otherwise.
+//
+// Template meta programming:
+//   is_pointer     - as in TR1; needed on Symbian and IBM XL C/C++ only.
+//   IteratorTraits - partial implementation of std::iterator_traits, which
+//                    is not available in libCstd when compiled with Sun C++.
+//
+// Smart pointers:
+//   scoped_ptr     - as in TR2.
+//
+// Regular expressions:
+//   RE             - a simple regular expression class using the POSIX
+//                    Extended Regular Expression syntax on UNIX-like
+//                    platforms, or a reduced regular exception syntax on
+//                    other platforms, including Windows.
+//
+// Logging:
+//   GTEST_LOG_()   - logs messages at the specified severity level.
+//   LogToStderr()  - directs all log messages to stderr.
+//   FlushInfoLog() - flushes informational log messages.
+//
+// Stdout and stderr capturing:
+//   CaptureStdout()     - starts capturing stdout.
+//   GetCapturedStdout() - stops capturing stdout and returns the captured
+//                         string.
+//   CaptureStderr()     - starts capturing stderr.
+//   GetCapturedStderr() - stops capturing stderr and returns the captured
+//                         string.
+//
+// Integer types:
+//   TypeWithSize   - maps an integer to a int type.
+//   Int32, UInt32, Int64, UInt64, TimeInMillis
+//                  - integers of known sizes.
+//   BiggestInt     - the biggest signed integer type.
+//
+// Command-line utilities:
+//   GTEST_FLAG()       - references a flag.
+//   GTEST_DECLARE_*()  - declares a flag.
+//   GTEST_DEFINE_*()   - defines a flag.
+//   GetInjectableArgvs() - returns the command line as a vector of strings.
+//
+// Environment variable utilities:
+//   GetEnv()             - gets the value of an environment variable.
+//   BoolFromGTestEnv()   - parses a bool environment variable.
+//   Int32FromGTestEnv()  - parses an Int32 environment variable.
+//   StringFromGTestEnv() - parses a string environment variable.
+
+#include <ctype.h>   // for isspace, etc
+#include <stddef.h>  // for ptrdiff_t
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#ifndef _WIN32_WCE
+# include <sys/types.h>
+# include <sys/stat.h>
+#endif  // !_WIN32_WCE
+
+#if defined __APPLE__
+# include <AvailabilityMacros.h>
+# include <TargetConditionals.h>
+#endif
+
+#include <iostream>  // NOLINT
+#include <sstream>  // NOLINT
+#include <string>  // NOLINT
+
+#define GTEST_DEV_EMAIL_ "googletestframework@@googlegroups.com"
+#define GTEST_FLAG_PREFIX_ "gtest_"
+#define GTEST_FLAG_PREFIX_DASH_ "gtest-"
+#define GTEST_FLAG_PREFIX_UPPER_ "GTEST_"
+#define GTEST_NAME_ "Google Test"
+#define GTEST_PROJECT_URL_ "http://code.google.com/p/googletest/"
+
+// Determines the version of gcc that is used to compile this.
+#ifdef __GNUC__
+// 40302 means version 4.3.2.
+# define GTEST_GCC_VER_ \
+    (__GNUC__*10000 + __GNUC_MINOR__*100 + __GNUC_PATCHLEVEL__)
+#endif  // __GNUC__
+
+// Determines the platform on which Google Test is compiled.
+#ifdef __CYGWIN__
+# define GTEST_OS_CYGWIN 1
+#elif defined __SYMBIAN32__
+# define GTEST_OS_SYMBIAN 1
+#elif defined _WIN32
+# define GTEST_OS_WINDOWS 1
+# ifdef _WIN32_WCE
+#  define GTEST_OS_WINDOWS_MOBILE 1
+# elif defined(__MINGW__) || defined(__MINGW32__)
+#  define GTEST_OS_WINDOWS_MINGW 1
+# else
+#  define GTEST_OS_WINDOWS_DESKTOP 1
+# endif  // _WIN32_WCE
+#elif defined __APPLE__
+# define GTEST_OS_MAC 1
+# if TARGET_OS_IPHONE
+#  define GTEST_OS_IOS 1
+#  if TARGET_IPHONE_SIMULATOR
+#   define GTEST_OS_IOS_SIMULATOR 1
+#  endif
+# endif
+#elif defined __linux__
+# define GTEST_OS_LINUX 1
+# if defined __ANDROID__
+#  define GTEST_OS_LINUX_ANDROID 1
+# endif
+#elif defined __MVS__
+# define GTEST_OS_ZOS 1
+#elif defined(__sun) && defined(__SVR4)
+# define GTEST_OS_SOLARIS 1
+#elif defined(_AIX)
+# define GTEST_OS_AIX 1
+#elif defined(__hpux)
+# define GTEST_OS_HPUX 1
+#elif defined __native_client__
+# define GTEST_OS_NACL 1
+#elif defined __OpenBSD__
+# define GTEST_OS_OPENBSD 1
+#elif defined __QNX__
+# define GTEST_OS_QNX 1
+#endif  // __CYGWIN__
+
+#ifndef GTEST_LANG_CXX11
+// gcc and clang define __GXX_EXPERIMENTAL_CXX0X__ when
+// -std={c,gnu}++{0x,11} is passed.  The C++11 standard specifies a
+// value for __cplusplus, and recent versions of clang, gcc, and
+// probably other compilers set that too in C++11 mode.
+# if __GXX_EXPERIMENTAL_CXX0X__ || __cplusplus >= 201103L
+// Compiling in at least C++11 mode.
+#  define GTEST_LANG_CXX11 1
+# else
+#  define GTEST_LANG_CXX11 0
+# endif
+#endif
+
+// Brings in definitions for functions used in the testing::internal::posix
+// namespace (read, write, close, chdir, isatty, stat). We do not currently
+// use them on Windows Mobile.
+#if !GTEST_OS_WINDOWS
+// This assumes that non-Windows OSes provide unistd.h. For OSes where this
+// is not the case, we need to include headers that provide the functions
+// mentioned above.
+# include <unistd.h>
+# include <strings.h>
+#elif !GTEST_OS_WINDOWS_MOBILE
+# include <direct.h>
+# include <io.h>
+#endif
+
+#if GTEST_OS_LINUX_ANDROID
+// Used to define __ANDROID_API__ matching the target NDK API level.
+#  include <android/api-level.h>  // NOLINT
+#endif
+
+// Defines this to true iff Google Test can use POSIX regular expressions.
+#ifndef GTEST_HAS_POSIX_RE
+# if GTEST_OS_LINUX_ANDROID
+// On Android, <regex.h> is only available starting with Gingerbread.
+#  define GTEST_HAS_POSIX_RE (__ANDROID_API__ >= 9)
+# else
+#  define GTEST_HAS_POSIX_RE (!GTEST_OS_WINDOWS)
+# endif
+#endif
+
+#if GTEST_HAS_POSIX_RE
+
+// On some platforms, <regex.h> needs someone to define size_t, and
+// won't compile otherwise.  We can #include it here as we already
+// included <stdlib.h>, which is guaranteed to define size_t through
+// <stddef.h>.
+# include <regex.h>  // NOLINT
+
+# define GTEST_USES_POSIX_RE 1
+
+#elif GTEST_OS_WINDOWS
+
+// <regex.h> is not available on Windows.  Use our own simple regex
+// implementation instead.
+# define GTEST_USES_SIMPLE_RE 1
+
+#else
+
+// <regex.h> may not be available on this platform.  Use our own
+// simple regex implementation instead.
+# define GTEST_USES_SIMPLE_RE 1
+
+#endif  // GTEST_HAS_POSIX_RE
+
+#ifndef GTEST_HAS_EXCEPTIONS
+// The user didn't tell us whether exceptions are enabled, so we need
+// to figure it out.
+# if defined(_MSC_VER) || defined(__BORLANDC__)
+// MSVC's and C++Builder's implementations of the STL use the _HAS_EXCEPTIONS
+// macro to enable exceptions, so we'll do the same.
+// Assumes that exceptions are enabled by default.
+#  ifndef _HAS_EXCEPTIONS
+#   define _HAS_EXCEPTIONS 1
+#  endif  // _HAS_EXCEPTIONS
+#  define GTEST_HAS_EXCEPTIONS _HAS_EXCEPTIONS
+# elif defined(__GNUC__) && __EXCEPTIONS
+// gcc defines __EXCEPTIONS to 1 iff exceptions are enabled.
+#  define GTEST_HAS_EXCEPTIONS 1
+# elif defined(__SUNPRO_CC)
+// Sun Pro CC supports exceptions.  However, there is no compile-time way of
+// detecting whether they are enabled or not.  Therefore, we assume that
+// they are enabled unless the user tells us otherwise.
+#  define GTEST_HAS_EXCEPTIONS 1
+# elif defined(__IBMCPP__) && __EXCEPTIONS
+// xlC defines __EXCEPTIONS to 1 iff exceptions are enabled.
+#  define GTEST_HAS_EXCEPTIONS 1
+# elif defined(__HP_aCC)
+// Exception handling is in effect by default in HP aCC compiler. It has to
+// be turned of by +noeh compiler option if desired.
+#  define GTEST_HAS_EXCEPTIONS 1
+# else
+// For other compilers, we assume exceptions are disabled to be
+// conservative.
+#  define GTEST_HAS_EXCEPTIONS 0
+# endif  // defined(_MSC_VER) || defined(__BORLANDC__)
+#endif  // GTEST_HAS_EXCEPTIONS
+
+#if !defined(GTEST_HAS_STD_STRING)
+// Even though we don't use this macro any longer, we keep it in case
+// some clients still depend on it.
+# define GTEST_HAS_STD_STRING 1
+#elif !GTEST_HAS_STD_STRING
+// The user told us that ::std::string isn't available.
+# error "Google Test cannot be used where ::std::string isn't available."
+#endif  // !defined(GTEST_HAS_STD_STRING)
+
+#ifndef GTEST_HAS_GLOBAL_STRING
+// The user didn't tell us whether ::string is available, so we need
+// to figure it out.
+
+# define GTEST_HAS_GLOBAL_STRING 0
+
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+#ifndef GTEST_HAS_STD_WSTRING
+// The user didn't tell us whether ::std::wstring is available, so we need
+// to figure it out.
+// TODO(wan@google.com): uses autoconf to detect whether ::std::wstring
+//   is available.
+
+// Cygwin 1.7 and below doesn't support ::std::wstring.
+// Solaris' libc++ doesn't support it either.  Android has
+// no support for it at least as recent as Froyo (2.2).
+# define GTEST_HAS_STD_WSTRING \
+    (!(GTEST_OS_LINUX_ANDROID || GTEST_OS_CYGWIN || GTEST_OS_SOLARIS))
+
+#endif  // GTEST_HAS_STD_WSTRING
+
+#ifndef GTEST_HAS_GLOBAL_WSTRING
+// The user didn't tell us whether ::wstring is available, so we need
+// to figure it out.
+# define GTEST_HAS_GLOBAL_WSTRING \
+    (GTEST_HAS_STD_WSTRING && GTEST_HAS_GLOBAL_STRING)
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+// Determines whether RTTI is available.
+#ifndef GTEST_HAS_RTTI
+// The user didn't tell us whether RTTI is enabled, so we need to
+// figure it out.
+
+# ifdef _MSC_VER
+
+#  ifdef _CPPRTTI  // MSVC defines this macro iff RTTI is enabled.
+#   define GTEST_HAS_RTTI 1
+#  else
+#   define GTEST_HAS_RTTI 0
+#  endif
+
+// Starting with version 4.3.2, gcc defines __GXX_RTTI iff RTTI is enabled.
+# elif defined(__GNUC__) && (GTEST_GCC_VER_ >= 40302)
+
+#  ifdef __GXX_RTTI
+// When building against STLport with the Android NDK and with
+// -frtti -fno-exceptions, the build fails at link time with undefined
+// references to __cxa_bad_typeid. Note sure if STL or toolchain bug,
+// so disable RTTI when detected.
+#   if GTEST_OS_LINUX_ANDROID && defined(_STLPORT_MAJOR) && \
+       !defined(__EXCEPTIONS)
+#    define GTEST_HAS_RTTI 0
+#   else
+#    define GTEST_HAS_RTTI 1
+#   endif  // GTEST_OS_LINUX_ANDROID && __STLPORT_MAJOR && !__EXCEPTIONS
+#  else
+#   define GTEST_HAS_RTTI 0
+#  endif  // __GXX_RTTI
+
+// Clang defines __GXX_RTTI starting with version 3.0, but its manual recommends
+// using has_feature instead. has_feature(cxx_rtti) is supported since 2.7, the
+// first version with C++ support.
+# elif defined(__clang__)
+
+#  define GTEST_HAS_RTTI __has_feature(cxx_rtti)
+
+// Starting with version 9.0 IBM Visual Age defines __RTTI_ALL__ to 1 if
+// both the typeid and dynamic_cast features are present.
+# elif defined(__IBMCPP__) && (__IBMCPP__ >= 900)
+
+#  ifdef __RTTI_ALL__
+#   define GTEST_HAS_RTTI 1
+#  else
+#   define GTEST_HAS_RTTI 0
+#  endif
+
+# else
+
+// For all other compilers, we assume RTTI is enabled.
+#  define GTEST_HAS_RTTI 1
+
+# endif  // _MSC_VER
+
+#endif  // GTEST_HAS_RTTI
+
+// It's this header's responsibility to #include <typeinfo> when RTTI
+// is enabled.
+#if GTEST_HAS_RTTI
+# include <typeinfo>
+#endif
+
+// Determines whether Google Test can use the pthreads library.
+#ifndef GTEST_HAS_PTHREAD
+// The user didn't tell us explicitly, so we assume pthreads support is
+// available on Linux and Mac.
+//
+// To disable threading support in Google Test, add -DGTEST_HAS_PTHREAD=0
+// to your compiler flags.
+# define GTEST_HAS_PTHREAD (GTEST_OS_LINUX || GTEST_OS_MAC || GTEST_OS_HPUX \
+    || GTEST_OS_QNX)
+#endif  // GTEST_HAS_PTHREAD
+
+#if GTEST_HAS_PTHREAD
+// gtest-port.h guarantees to #include <pthread.h> when GTEST_HAS_PTHREAD is
+// true.
+# include <pthread.h>  // NOLINT
+
+// For timespec and nanosleep, used below.
+# include <time.h>  // NOLINT
+#endif
+
+// Determines whether Google Test can use tr1/tuple.  You can define
+// this macro to 0 to prevent Google Test from using tuple (any
+// feature depending on tuple with be disabled in this mode).
+#ifndef GTEST_HAS_TR1_TUPLE
+# if GTEST_OS_LINUX_ANDROID && defined(_STLPORT_MAJOR)
+// STLport, provided with the Android NDK, has neither <tr1/tuple> or <tuple>.
+#  define GTEST_HAS_TR1_TUPLE 0
+# else
+// The user didn't tell us not to do it, so we assume it's OK.
+#  define GTEST_HAS_TR1_TUPLE 1
+# endif
+#endif  // GTEST_HAS_TR1_TUPLE
+
+// Determines whether Google Test's own tr1 tuple implementation
+// should be used.
+#ifndef GTEST_USE_OWN_TR1_TUPLE
+// The user didn't tell us, so we need to figure it out.
+
+// We use our own TR1 tuple if we aren't sure the user has an
+// implementation of it already.  At this time, libstdc++ 4.0.0+ and
+// MSVC 2010 are the only mainstream standard libraries that come
+// with a TR1 tuple implementation.  NVIDIA's CUDA NVCC compiler
+// pretends to be GCC by defining __GNUC__ and friends, but cannot
+// compile GCC's tuple implementation.  MSVC 2008 (9.0) provides TR1
+// tuple in a 323 MB Feature Pack download, which we cannot assume the
+// user has.  QNX's QCC compiler is a modified GCC but it doesn't
+// support TR1 tuple.  libc++ only provides std::tuple, in C++11 mode,
+// and it can be used with some compilers that define __GNUC__.
+# if (defined(__GNUC__) && !defined(__CUDACC__) && (GTEST_GCC_VER_ >= 40000) \
+      && !GTEST_OS_QNX && !defined(_LIBCPP_VERSION)) || _MSC_VER >= 1600
+#  define GTEST_ENV_HAS_TR1_TUPLE_ 1
+# endif
+
+// C++11 specifies that <tuple> provides std::tuple. Use that if gtest is used
+// in C++11 mode and libstdc++ isn't very old (binaries targeting OS X 10.6
+// can build with clang but need to use gcc4.2's libstdc++).
+# if GTEST_LANG_CXX11 && (!defined(__GLIBCXX__) || __GLIBCXX__ > 20110325)
+#  define GTEST_ENV_HAS_STD_TUPLE_ 1
+# endif
+
+# if GTEST_ENV_HAS_TR1_TUPLE_ || GTEST_ENV_HAS_STD_TUPLE_
+#  define GTEST_USE_OWN_TR1_TUPLE 0
+# else
+#  define GTEST_USE_OWN_TR1_TUPLE 1
+# endif
+
+#endif  // GTEST_USE_OWN_TR1_TUPLE
+
+// To avoid conditional compilation everywhere, we make it
+// gtest-port.h's responsibility to #include the header implementing
+// tr1/tuple.
+#if GTEST_HAS_TR1_TUPLE
+
+# if GTEST_USE_OWN_TR1_TUPLE
+// This file was GENERATED by command:
+//     pump.py gtest-tuple.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2009 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Implements a subset of TR1 tuple needed by Google Test and Google Mock.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
+
+#include <utility>  // For ::std::pair.
+
+// The compiler used in Symbian has a bug that prevents us from declaring the
+// tuple template as a friend (it complains that tuple is redefined).  This
+// hack bypasses the bug by declaring the members that should otherwise be
+// private as public.
+// Sun Studio versions < 12 also have the above bug.
+#if defined(__SYMBIAN32__) || (defined(__SUNPRO_CC) && __SUNPRO_CC < 0x590)
+# define GTEST_DECLARE_TUPLE_AS_FRIEND_ public:
+#else
+# define GTEST_DECLARE_TUPLE_AS_FRIEND_ \
+    template <GTEST_10_TYPENAMES_(U)> friend class tuple; \
+   private:
+#endif
+
+// GTEST_n_TUPLE_(T) is the type of an n-tuple.
+#define GTEST_0_TUPLE_(T) tuple<>
+#define GTEST_1_TUPLE_(T) tuple<T##0, void, void, void, void, void, void, \
+    void, void, void>
+#define GTEST_2_TUPLE_(T) tuple<T##0, T##1, void, void, void, void, void, \
+    void, void, void>
+#define GTEST_3_TUPLE_(T) tuple<T##0, T##1, T##2, void, void, void, void, \
+    void, void, void>
+#define GTEST_4_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, void, void, void, \
+    void, void, void>
+#define GTEST_5_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, void, void, \
+    void, void, void>
+#define GTEST_6_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, T##5, void, \
+    void, void, void>
+#define GTEST_7_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, T##5, T##6, \
+    void, void, void>
+#define GTEST_8_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, T##5, T##6, \
+    T##7, void, void>
+#define GTEST_9_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, T##5, T##6, \
+    T##7, T##8, void>
+#define GTEST_10_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, T##5, T##6, \
+    T##7, T##8, T##9>
+
+// GTEST_n_TYPENAMES_(T) declares a list of n typenames.
+#define GTEST_0_TYPENAMES_(T)
+#define GTEST_1_TYPENAMES_(T) typename T##0
+#define GTEST_2_TYPENAMES_(T) typename T##0, typename T##1
+#define GTEST_3_TYPENAMES_(T) typename T##0, typename T##1, typename T##2
+#define GTEST_4_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3
+#define GTEST_5_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4
+#define GTEST_6_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5
+#define GTEST_7_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6
+#define GTEST_8_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6, typename T##7
+#define GTEST_9_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6, \
+    typename T##7, typename T##8
+#define GTEST_10_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6, \
+    typename T##7, typename T##8, typename T##9
+
+// In theory, defining stuff in the ::std namespace is undefined
+// behavior.  We can do this as we are playing the role of a standard
+// library vendor.
+namespace std {
+namespace tr1 {
+
+template <typename T0 = void, typename T1 = void, typename T2 = void,
+    typename T3 = void, typename T4 = void, typename T5 = void,
+    typename T6 = void, typename T7 = void, typename T8 = void,
+    typename T9 = void>
+class tuple;
+
+// Anything in namespace gtest_internal is Google Test's INTERNAL
+// IMPLEMENTATION DETAIL and MUST NOT BE USED DIRECTLY in user code.
+namespace gtest_internal {
+
+// ByRef<T>::type is T if T is a reference; otherwise it's const T&.
+template <typename T>
+struct ByRef { typedef const T& type; };  // NOLINT
+template <typename T>
+struct ByRef<T&> { typedef T& type; };  // NOLINT
+
+// A handy wrapper for ByRef.
+#define GTEST_BY_REF_(T) typename ::std::tr1::gtest_internal::ByRef<T>::type
+
+// AddRef<T>::type is T if T is a reference; otherwise it's T&.  This
+// is the same as tr1::add_reference<T>::type.
+template <typename T>
+struct AddRef { typedef T& type; };  // NOLINT
+template <typename T>
+struct AddRef<T&> { typedef T& type; };  // NOLINT
+
+// A handy wrapper for AddRef.
+#define GTEST_ADD_REF_(T) typename ::std::tr1::gtest_internal::AddRef<T>::type
+
+// A helper for implementing get<k>().
+template <int k> class Get;
+
+// A helper for implementing tuple_element<k, T>.  kIndexValid is true
+// iff k < the number of fields in tuple type T.
+template <bool kIndexValid, int kIndex, class Tuple>
+struct TupleElement;
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 0, GTEST_10_TUPLE_(T) > {
+  typedef T0 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 1, GTEST_10_TUPLE_(T) > {
+  typedef T1 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 2, GTEST_10_TUPLE_(T) > {
+  typedef T2 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 3, GTEST_10_TUPLE_(T) > {
+  typedef T3 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 4, GTEST_10_TUPLE_(T) > {
+  typedef T4 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 5, GTEST_10_TUPLE_(T) > {
+  typedef T5 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 6, GTEST_10_TUPLE_(T) > {
+  typedef T6 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 7, GTEST_10_TUPLE_(T) > {
+  typedef T7 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 8, GTEST_10_TUPLE_(T) > {
+  typedef T8 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 9, GTEST_10_TUPLE_(T) > {
+  typedef T9 type;
+};
+
+}  // namespace gtest_internal
+
+template <>
+class tuple<> {
+ public:
+  tuple() {}
+  tuple(const tuple& /* t */)  {}
+  tuple& operator=(const tuple& /* t */) { return *this; }
+};
+
+template <GTEST_1_TYPENAMES_(T)>
+class GTEST_1_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0) : f0_(f0) {}
+
+  tuple(const tuple& t) : f0_(t.f0_) {}
+
+  template <GTEST_1_TYPENAMES_(U)>
+  tuple(const GTEST_1_TUPLE_(U)& t) : f0_(t.f0_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_1_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_1_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_1_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_1_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    return *this;
+  }
+
+  T0 f0_;
+};
+
+template <GTEST_2_TYPENAMES_(T)>
+class GTEST_2_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1) : f0_(f0),
+      f1_(f1) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_) {}
+
+  template <GTEST_2_TYPENAMES_(U)>
+  tuple(const GTEST_2_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_) {}
+  template <typename U0, typename U1>
+  tuple(const ::std::pair<U0, U1>& p) : f0_(p.first), f1_(p.second) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_2_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_2_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+  template <typename U0, typename U1>
+  tuple& operator=(const ::std::pair<U0, U1>& p) {
+    f0_ = p.first;
+    f1_ = p.second;
+    return *this;
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_2_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_2_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+};
+
+template <GTEST_3_TYPENAMES_(T)>
+class GTEST_3_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2) : f0_(f0), f1_(f1), f2_(f2) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_) {}
+
+  template <GTEST_3_TYPENAMES_(U)>
+  tuple(const GTEST_3_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_3_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_3_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_3_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_3_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+};
+
+template <GTEST_4_TYPENAMES_(T)>
+class GTEST_4_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3) : f0_(f0), f1_(f1), f2_(f2),
+      f3_(f3) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_) {}
+
+  template <GTEST_4_TYPENAMES_(U)>
+  tuple(const GTEST_4_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_4_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_4_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_4_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_4_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+};
+
+template <GTEST_5_TYPENAMES_(T)>
+class GTEST_5_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3,
+      GTEST_BY_REF_(T4) f4) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_) {}
+
+  template <GTEST_5_TYPENAMES_(U)>
+  tuple(const GTEST_5_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_5_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_5_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_5_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_5_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+};
+
+template <GTEST_6_TYPENAMES_(T)>
+class GTEST_6_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4),
+      f5_(f5) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_) {}
+
+  template <GTEST_6_TYPENAMES_(U)>
+  tuple(const GTEST_6_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_6_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_6_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_6_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_6_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+};
+
+template <GTEST_7_TYPENAMES_(T)>
+class GTEST_7_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6) : f0_(f0), f1_(f1), f2_(f2),
+      f3_(f3), f4_(f4), f5_(f5), f6_(f6) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_) {}
+
+  template <GTEST_7_TYPENAMES_(U)>
+  tuple(const GTEST_7_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_7_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_7_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_7_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_7_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+};
+
+template <GTEST_8_TYPENAMES_(T)>
+class GTEST_8_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_(), f7_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6,
+      GTEST_BY_REF_(T7) f7) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4),
+      f5_(f5), f6_(f6), f7_(f7) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_) {}
+
+  template <GTEST_8_TYPENAMES_(U)>
+  tuple(const GTEST_8_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_8_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_8_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_8_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_8_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    f7_ = t.f7_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+  T7 f7_;
+};
+
+template <GTEST_9_TYPENAMES_(T)>
+class GTEST_9_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_(), f7_(), f8_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6, GTEST_BY_REF_(T7) f7,
+      GTEST_BY_REF_(T8) f8) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4),
+      f5_(f5), f6_(f6), f7_(f7), f8_(f8) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_) {}
+
+  template <GTEST_9_TYPENAMES_(U)>
+  tuple(const GTEST_9_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_9_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_9_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_9_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_9_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    f7_ = t.f7_;
+    f8_ = t.f8_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+  T7 f7_;
+  T8 f8_;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+class tuple {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_(), f7_(), f8_(),
+      f9_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6, GTEST_BY_REF_(T7) f7,
+      GTEST_BY_REF_(T8) f8, GTEST_BY_REF_(T9) f9) : f0_(f0), f1_(f1), f2_(f2),
+      f3_(f3), f4_(f4), f5_(f5), f6_(f6), f7_(f7), f8_(f8), f9_(f9) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_), f9_(t.f9_) {}
+
+  template <GTEST_10_TYPENAMES_(U)>
+  tuple(const GTEST_10_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_),
+      f9_(t.f9_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_10_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_10_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_10_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_10_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    f7_ = t.f7_;
+    f8_ = t.f8_;
+    f9_ = t.f9_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+  T7 f7_;
+  T8 f8_;
+  T9 f9_;
+};
+
+// 6.1.3.2 Tuple creation functions.
+
+// Known limitations: we don't support passing an
+// std::tr1::reference_wrapper<T> to make_tuple().  And we don't
+// implement tie().
+
+inline tuple<> make_tuple() { return tuple<>(); }
+
+template <GTEST_1_TYPENAMES_(T)>
+inline GTEST_1_TUPLE_(T) make_tuple(const T0& f0) {
+  return GTEST_1_TUPLE_(T)(f0);
+}
+
+template <GTEST_2_TYPENAMES_(T)>
+inline GTEST_2_TUPLE_(T) make_tuple(const T0& f0, const T1& f1) {
+  return GTEST_2_TUPLE_(T)(f0, f1);
+}
+
+template <GTEST_3_TYPENAMES_(T)>
+inline GTEST_3_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2) {
+  return GTEST_3_TUPLE_(T)(f0, f1, f2);
+}
+
+template <GTEST_4_TYPENAMES_(T)>
+inline GTEST_4_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3) {
+  return GTEST_4_TUPLE_(T)(f0, f1, f2, f3);
+}
+
+template <GTEST_5_TYPENAMES_(T)>
+inline GTEST_5_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4) {
+  return GTEST_5_TUPLE_(T)(f0, f1, f2, f3, f4);
+}
+
+template <GTEST_6_TYPENAMES_(T)>
+inline GTEST_6_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5) {
+  return GTEST_6_TUPLE_(T)(f0, f1, f2, f3, f4, f5);
+}
+
+template <GTEST_7_TYPENAMES_(T)>
+inline GTEST_7_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6) {
+  return GTEST_7_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6);
+}
+
+template <GTEST_8_TYPENAMES_(T)>
+inline GTEST_8_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6, const T7& f7) {
+  return GTEST_8_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6, f7);
+}
+
+template <GTEST_9_TYPENAMES_(T)>
+inline GTEST_9_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6, const T7& f7,
+    const T8& f8) {
+  return GTEST_9_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6, f7, f8);
+}
+
+template <GTEST_10_TYPENAMES_(T)>
+inline GTEST_10_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6, const T7& f7,
+    const T8& f8, const T9& f9) {
+  return GTEST_10_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6, f7, f8, f9);
+}
+
+// 6.1.3.3 Tuple helper classes.
+
+template <typename Tuple> struct tuple_size;
+
+template <GTEST_0_TYPENAMES_(T)>
+struct tuple_size<GTEST_0_TUPLE_(T) > {
+  static const int value = 0;
+};
+
+template <GTEST_1_TYPENAMES_(T)>
+struct tuple_size<GTEST_1_TUPLE_(T) > {
+  static const int value = 1;
+};
+
+template <GTEST_2_TYPENAMES_(T)>
+struct tuple_size<GTEST_2_TUPLE_(T) > {
+  static const int value = 2;
+};
+
+template <GTEST_3_TYPENAMES_(T)>
+struct tuple_size<GTEST_3_TUPLE_(T) > {
+  static const int value = 3;
+};
+
+template <GTEST_4_TYPENAMES_(T)>
+struct tuple_size<GTEST_4_TUPLE_(T) > {
+  static const int value = 4;
+};
+
+template <GTEST_5_TYPENAMES_(T)>
+struct tuple_size<GTEST_5_TUPLE_(T) > {
+  static const int value = 5;
+};
+
+template <GTEST_6_TYPENAMES_(T)>
+struct tuple_size<GTEST_6_TUPLE_(T) > {
+  static const int value = 6;
+};
+
+template <GTEST_7_TYPENAMES_(T)>
+struct tuple_size<GTEST_7_TUPLE_(T) > {
+  static const int value = 7;
+};
+
+template <GTEST_8_TYPENAMES_(T)>
+struct tuple_size<GTEST_8_TUPLE_(T) > {
+  static const int value = 8;
+};
+
+template <GTEST_9_TYPENAMES_(T)>
+struct tuple_size<GTEST_9_TUPLE_(T) > {
+  static const int value = 9;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct tuple_size<GTEST_10_TUPLE_(T) > {
+  static const int value = 10;
+};
+
+template <int k, class Tuple>
+struct tuple_element {
+  typedef typename gtest_internal::TupleElement<
+      k < (tuple_size<Tuple>::value), k, Tuple>::type type;
+};
+
+#define GTEST_TUPLE_ELEMENT_(k, Tuple) typename tuple_element<k, Tuple >::type
+
+// 6.1.3.4 Element access.
+
+namespace gtest_internal {
+
+template <>
+class Get<0> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(0, Tuple))
+  Field(Tuple& t) { return t.f0_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(0, Tuple))
+  ConstField(const Tuple& t) { return t.f0_; }
+};
+
+template <>
+class Get<1> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(1, Tuple))
+  Field(Tuple& t) { return t.f1_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(1, Tuple))
+  ConstField(const Tuple& t) { return t.f1_; }
+};
+
+template <>
+class Get<2> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(2, Tuple))
+  Field(Tuple& t) { return t.f2_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(2, Tuple))
+  ConstField(const Tuple& t) { return t.f2_; }
+};
+
+template <>
+class Get<3> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(3, Tuple))
+  Field(Tuple& t) { return t.f3_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(3, Tuple))
+  ConstField(const Tuple& t) { return t.f3_; }
+};
+
+template <>
+class Get<4> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(4, Tuple))
+  Field(Tuple& t) { return t.f4_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(4, Tuple))
+  ConstField(const Tuple& t) { return t.f4_; }
+};
+
+template <>
+class Get<5> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(5, Tuple))
+  Field(Tuple& t) { return t.f5_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(5, Tuple))
+  ConstField(const Tuple& t) { return t.f5_; }
+};
+
+template <>
+class Get<6> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(6, Tuple))
+  Field(Tuple& t) { return t.f6_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(6, Tuple))
+  ConstField(const Tuple& t) { return t.f6_; }
+};
+
+template <>
+class Get<7> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(7, Tuple))
+  Field(Tuple& t) { return t.f7_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(7, Tuple))
+  ConstField(const Tuple& t) { return t.f7_; }
+};
+
+template <>
+class Get<8> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(8, Tuple))
+  Field(Tuple& t) { return t.f8_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(8, Tuple))
+  ConstField(const Tuple& t) { return t.f8_; }
+};
+
+template <>
+class Get<9> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(9, Tuple))
+  Field(Tuple& t) { return t.f9_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(9, Tuple))
+  ConstField(const Tuple& t) { return t.f9_; }
+};
+
+}  // namespace gtest_internal
+
+template <int k, GTEST_10_TYPENAMES_(T)>
+GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(k, GTEST_10_TUPLE_(T)))
+get(GTEST_10_TUPLE_(T)& t) {
+  return gtest_internal::Get<k>::Field(t);
+}
+
+template <int k, GTEST_10_TYPENAMES_(T)>
+GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(k,  GTEST_10_TUPLE_(T)))
+get(const GTEST_10_TUPLE_(T)& t) {
+  return gtest_internal::Get<k>::ConstField(t);
+}
+
+// 6.1.3.5 Relational operators
+
+// We only implement == and !=, as we don't have a need for the rest yet.
+
+namespace gtest_internal {
+
+// SameSizeTuplePrefixComparator<k, k>::Eq(t1, t2) returns true if the
+// first k fields of t1 equals the first k fields of t2.
+// SameSizeTuplePrefixComparator(k1, k2) would be a compiler error if
+// k1 != k2.
+template <int kSize1, int kSize2>
+struct SameSizeTuplePrefixComparator;
+
+template <>
+struct SameSizeTuplePrefixComparator<0, 0> {
+  template <class Tuple1, class Tuple2>
+  static bool Eq(const Tuple1& /* t1 */, const Tuple2& /* t2 */) {
+    return true;
+  }
+};
+
+template <int k>
+struct SameSizeTuplePrefixComparator<k, k> {
+  template <class Tuple1, class Tuple2>
+  static bool Eq(const Tuple1& t1, const Tuple2& t2) {
+    return SameSizeTuplePrefixComparator<k - 1, k - 1>::Eq(t1, t2) &&
+        ::std::tr1::get<k - 1>(t1) == ::std::tr1::get<k - 1>(t2);
+  }
+};
+
+}  // namespace gtest_internal
+
+template <GTEST_10_TYPENAMES_(T), GTEST_10_TYPENAMES_(U)>
+inline bool operator==(const GTEST_10_TUPLE_(T)& t,
+                       const GTEST_10_TUPLE_(U)& u) {
+  return gtest_internal::SameSizeTuplePrefixComparator<
+      tuple_size<GTEST_10_TUPLE_(T) >::value,
+      tuple_size<GTEST_10_TUPLE_(U) >::value>::Eq(t, u);
+}
+
+template <GTEST_10_TYPENAMES_(T), GTEST_10_TYPENAMES_(U)>
+inline bool operator!=(const GTEST_10_TUPLE_(T)& t,
+                       const GTEST_10_TUPLE_(U)& u) { return !(t == u); }
+
+// 6.1.4 Pairs.
+// Unimplemented.
+
+}  // namespace tr1
+}  // namespace std
+
+#undef GTEST_0_TUPLE_
+#undef GTEST_1_TUPLE_
+#undef GTEST_2_TUPLE_
+#undef GTEST_3_TUPLE_
+#undef GTEST_4_TUPLE_
+#undef GTEST_5_TUPLE_
+#undef GTEST_6_TUPLE_
+#undef GTEST_7_TUPLE_
+#undef GTEST_8_TUPLE_
+#undef GTEST_9_TUPLE_
+#undef GTEST_10_TUPLE_
+
+#undef GTEST_0_TYPENAMES_
+#undef GTEST_1_TYPENAMES_
+#undef GTEST_2_TYPENAMES_
+#undef GTEST_3_TYPENAMES_
+#undef GTEST_4_TYPENAMES_
+#undef GTEST_5_TYPENAMES_
+#undef GTEST_6_TYPENAMES_
+#undef GTEST_7_TYPENAMES_
+#undef GTEST_8_TYPENAMES_
+#undef GTEST_9_TYPENAMES_
+#undef GTEST_10_TYPENAMES_
+
+#undef GTEST_DECLARE_TUPLE_AS_FRIEND_
+#undef GTEST_BY_REF_
+#undef GTEST_ADD_REF_
+#undef GTEST_TUPLE_ELEMENT_
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
+# elif GTEST_ENV_HAS_STD_TUPLE_
+#  include <tuple>
+// C++11 puts its tuple into the ::std namespace rather than
+// ::std::tr1.  gtest expects tuple to live in ::std::tr1, so put it there.
+// This causes undefined behavior, but supported compilers react in
+// the way we intend.
+namespace std {
+namespace tr1 {
+using ::std::get;
+using ::std::make_tuple;
+using ::std::tuple;
+using ::std::tuple_element;
+using ::std::tuple_size;
+}
+}
+
+# elif GTEST_OS_SYMBIAN
+
+// On Symbian, BOOST_HAS_TR1_TUPLE causes Boost's TR1 tuple library to
+// use STLport's tuple implementation, which unfortunately doesn't
+// work as the copy of STLport distributed with Symbian is incomplete.
+// By making sure BOOST_HAS_TR1_TUPLE is undefined, we force Boost to
+// use its own tuple implementation.
+#  ifdef BOOST_HAS_TR1_TUPLE
+#   undef BOOST_HAS_TR1_TUPLE
+#  endif  // BOOST_HAS_TR1_TUPLE
+
+// This prevents <boost/tr1/detail/config.hpp>, which defines
+// BOOST_HAS_TR1_TUPLE, from being #included by Boost's <tuple>.
+#  define BOOST_TR1_DETAIL_CONFIG_HPP_INCLUDED
+#  include <tuple>
+
+# elif defined(__GNUC__) && (GTEST_GCC_VER_ >= 40000)
+// GCC 4.0+ implements tr1/tuple in the <tr1/tuple> header.  This does
+// not conform to the TR1 spec, which requires the header to be <tuple>.
+
+#  if !GTEST_HAS_RTTI && GTEST_GCC_VER_ < 40302
+// Until version 4.3.2, gcc has a bug that causes <tr1/functional>,
+// which is #included by <tr1/tuple>, to not compile when RTTI is
+// disabled.  _TR1_FUNCTIONAL is the header guard for
+// <tr1/functional>.  Hence the following #define is a hack to prevent
+// <tr1/functional> from being included.
+#   define _TR1_FUNCTIONAL 1
+#   include <tr1/tuple>
+#   undef _TR1_FUNCTIONAL  // Allows the user to #include
+                        // <tr1/functional> if he chooses to.
+#  else
+#   include <tr1/tuple>  // NOLINT
+#  endif  // !GTEST_HAS_RTTI && GTEST_GCC_VER_ < 40302
+
+# else
+// If the compiler is not GCC 4.0+, we assume the user is using a
+// spec-conforming TR1 implementation.
+#  include <tuple>  // NOLINT
+# endif  // GTEST_USE_OWN_TR1_TUPLE
+
+#endif  // GTEST_HAS_TR1_TUPLE
+
+// Determines whether clone(2) is supported.
+// Usually it will only be available on Linux, excluding
+// Linux on the Itanium architecture.
+// Also see http://linux.die.net/man/2/clone.
+#ifndef GTEST_HAS_CLONE
+// The user didn't tell us, so we need to figure it out.
+
+# if GTEST_OS_LINUX && !defined(__ia64__)
+#  if GTEST_OS_LINUX_ANDROID
+// On Android, clone() is only available on ARM starting with Gingerbread.
+#    if defined(__arm__) && __ANDROID_API__ >= 9
+#     define GTEST_HAS_CLONE 1
+#    else
+#     define GTEST_HAS_CLONE 0
+#    endif
+#  else
+#   define GTEST_HAS_CLONE 1
+#  endif
+# else
+#  define GTEST_HAS_CLONE 0
+# endif  // GTEST_OS_LINUX && !defined(__ia64__)
+
+#endif  // GTEST_HAS_CLONE
+
+// Determines whether to support stream redirection. This is used to test
+// output correctness and to implement death tests.
+#ifndef GTEST_HAS_STREAM_REDIRECTION
+// By default, we assume that stream redirection is supported on all
+// platforms except known mobile ones.
+# if GTEST_OS_WINDOWS_MOBILE || GTEST_OS_SYMBIAN
+#  define GTEST_HAS_STREAM_REDIRECTION 0
+# else
+#  define GTEST_HAS_STREAM_REDIRECTION 1
+# endif  // !GTEST_OS_WINDOWS_MOBILE && !GTEST_OS_SYMBIAN
+#endif  // GTEST_HAS_STREAM_REDIRECTION
+
+// Determines whether to support death tests.
+// Google Test does not support death tests for VC 7.1 and earlier as
+// abort() in a VC 7.1 application compiled as GUI in debug config
+// pops up a dialog window that cannot be suppressed programmatically.
+#if (GTEST_OS_LINUX || GTEST_OS_CYGWIN || GTEST_OS_SOLARIS || \
+     (GTEST_OS_MAC && !GTEST_OS_IOS) || GTEST_OS_IOS_SIMULATOR || \
+     (GTEST_OS_WINDOWS_DESKTOP && _MSC_VER >= 1400) || \
+     GTEST_OS_WINDOWS_MINGW || GTEST_OS_AIX || GTEST_OS_HPUX || \
+     GTEST_OS_OPENBSD || GTEST_OS_QNX)
+# define GTEST_HAS_DEATH_TEST 1
+# include <vector>  // NOLINT
+#endif
+
+// We don't support MSVC 7.1 with exceptions disabled now.  Therefore
+// all the compilers we care about are adequate for supporting
+// value-parameterized tests.
+#define GTEST_HAS_PARAM_TEST 1
+
+// Determines whether to support type-driven tests.
+
+// Typed tests need <typeinfo> and variadic macros, which GCC, VC++ 8.0,
+// Sun Pro CC, IBM Visual Age, and HP aCC support.
+#if defined(__GNUC__) || (_MSC_VER >= 1400) || defined(__SUNPRO_CC) || \
+    defined(__IBMCPP__) || defined(__HP_aCC)
+# define GTEST_HAS_TYPED_TEST 1
+# define GTEST_HAS_TYPED_TEST_P 1
+#endif
+
+// Determines whether to support Combine(). This only makes sense when
+// value-parameterized tests are enabled.  The implementation doesn't
+// work on Sun Studio since it doesn't understand templated conversion
+// operators.
+#if GTEST_HAS_PARAM_TEST && GTEST_HAS_TR1_TUPLE && !defined(__SUNPRO_CC)
+# define GTEST_HAS_COMBINE 1
+#endif
+
+// Determines whether the system compiler uses UTF-16 for encoding wide strings.
+#define GTEST_WIDE_STRING_USES_UTF16_ \
+    (GTEST_OS_WINDOWS || GTEST_OS_CYGWIN || GTEST_OS_SYMBIAN || GTEST_OS_AIX)
+
+// Determines whether test results can be streamed to a socket.
+#if GTEST_OS_LINUX
+# define GTEST_CAN_STREAM_RESULTS_ 1
+#endif
+
+// Defines some utility macros.
+
+// The GNU compiler emits a warning if nested "if" statements are followed by
+// an "else" statement and braces are not used to explicitly disambiguate the
+// "else" binding.  This leads to problems with code like:
+//
+//   if (gate)
+//     ASSERT_*(condition) << "Some message";
+//
+// The "switch (0) case 0:" idiom is used to suppress this.
+#ifdef __INTEL_COMPILER
+# define GTEST_AMBIGUOUS_ELSE_BLOCKER_
+#else
+# define GTEST_AMBIGUOUS_ELSE_BLOCKER_ switch (0) case 0: default:  // NOLINT
+#endif
+
+// Use this annotation at the end of a struct/class definition to
+// prevent the compiler from optimizing away instances that are never
+// used.  This is useful when all interesting logic happens inside the
+// c'tor and / or d'tor.  Example:
+//
+//   struct Foo {
+//     Foo() { ... }
+//   } GTEST_ATTRIBUTE_UNUSED_;
+//
+// Also use it after a variable or parameter declaration to tell the
+// compiler the variable/parameter does not have to be used.
+#if defined(__GNUC__) && !defined(COMPILER_ICC)
+# define GTEST_ATTRIBUTE_UNUSED_ __attribute__ ((unused))
+#else
+# define GTEST_ATTRIBUTE_UNUSED_
+#endif
+
+// A macro to disallow operator=
+// This should be used in the private: declarations for a class.
+#define GTEST_DISALLOW_ASSIGN_(type)\
+  void operator=(type const &)
+
+// A macro to disallow copy constructor and operator=
+// This should be used in the private: declarations for a class.
+#define GTEST_DISALLOW_COPY_AND_ASSIGN_(type)\
+  type(type const &);\
+  GTEST_DISALLOW_ASSIGN_(type)
+
+// Tell the compiler to warn about unused return values for functions declared
+// with this macro.  The macro should be used on function declarations
+// following the argument list:
+//
+//   Sprocket* AllocateSprocket() GTEST_MUST_USE_RESULT_;
+#if defined(__GNUC__) && (GTEST_GCC_VER_ >= 30400) && !defined(COMPILER_ICC)
+# define GTEST_MUST_USE_RESULT_ __attribute__ ((warn_unused_result))
+#else
+# define GTEST_MUST_USE_RESULT_
+#endif  // __GNUC__ && (GTEST_GCC_VER_ >= 30400) && !COMPILER_ICC
+
+// Determine whether the compiler supports Microsoft's Structured Exception
+// Handling.  This is supported by several Windows compilers but generally
+// does not exist on any other system.
+#ifndef GTEST_HAS_SEH
+// The user didn't tell us, so we need to figure it out.
+
+# if defined(_MSC_VER) || defined(__BORLANDC__)
+// These two compilers are known to support SEH.
+#  define GTEST_HAS_SEH 1
+# else
+// Assume no SEH.
+#  define GTEST_HAS_SEH 0
+# endif
+
+#endif  // GTEST_HAS_SEH
+
+#ifdef _MSC_VER
+
+# if GTEST_LINKED_AS_SHARED_LIBRARY
+#  define GTEST_API_ __declspec(dllimport)
+# elif GTEST_CREATE_SHARED_LIBRARY
+#  define GTEST_API_ __declspec(dllexport)
+# endif
+
+#endif  // _MSC_VER
+
+#ifndef GTEST_API_
+# define GTEST_API_
+#endif
+
+#ifdef __GNUC__
+// Ask the compiler to never inline a given function.
+# define GTEST_NO_INLINE_ __attribute__((noinline))
+#else
+# define GTEST_NO_INLINE_
+#endif
+
+// _LIBCPP_VERSION is defined by the libc++ library from the LLVM project.
+#if defined(__GLIBCXX__) || defined(_LIBCPP_VERSION)
+# define GTEST_HAS_CXXABI_H_ 1
+#else
+# define GTEST_HAS_CXXABI_H_ 0
+#endif
+
+namespace testing {
+
+class Message;
+
+namespace internal {
+
+// A secret type that Google Test users don't know about.  It has no
+// definition on purpose.  Therefore it's impossible to create a
+// Secret object, which is what we want.
+class Secret;
+
+// The GTEST_COMPILE_ASSERT_ macro can be used to verify that a compile time
+// expression is true. For example, you could use it to verify the
+// size of a static array:
+//
+//   GTEST_COMPILE_ASSERT_(ARRAYSIZE(content_type_names) == CONTENT_NUM_TYPES,
+//                         content_type_names_incorrect_size);
+//
+// or to make sure a struct is smaller than a certain size:
+//
+//   GTEST_COMPILE_ASSERT_(sizeof(foo) < 128, foo_too_large);
+//
+// The second argument to the macro is the name of the variable. If
+// the expression is false, most compilers will issue a warning/error
+// containing the name of the variable.
+
+template <bool>
+struct CompileAssert {
+};
+
+#define GTEST_COMPILE_ASSERT_(expr, msg) \
+  typedef ::testing::internal::CompileAssert<(static_cast<bool>(expr))> \
+      msg[static_cast<bool>(expr) ? 1 : -1] GTEST_ATTRIBUTE_UNUSED_
+
+// Implementation details of GTEST_COMPILE_ASSERT_:
+//
+// - GTEST_COMPILE_ASSERT_ works by defining an array type that has -1
+//   elements (and thus is invalid) when the expression is false.
+//
+// - The simpler definition
+//
+//    #define GTEST_COMPILE_ASSERT_(expr, msg) typedef char msg[(expr) ? 1 : -1]
+//
+//   does not work, as gcc supports variable-length arrays whose sizes
+//   are determined at run-time (this is gcc's extension and not part
+//   of the C++ standard).  As a result, gcc fails to reject the
+//   following code with the simple definition:
+//
+//     int foo;
+//     GTEST_COMPILE_ASSERT_(foo, msg); // not supposed to compile as foo is
+//                                      // not a compile-time constant.
+//
+// - By using the type CompileAssert<(bool(expr))>, we ensures that
+//   expr is a compile-time constant.  (Template arguments must be
+//   determined at compile-time.)
+//
+// - The outter parentheses in CompileAssert<(bool(expr))> are necessary
+//   to work around a bug in gcc 3.4.4 and 4.0.1.  If we had written
+//
+//     CompileAssert<bool(expr)>
+//
+//   instead, these compilers will refuse to compile
+//
+//     GTEST_COMPILE_ASSERT_(5 > 0, some_message);
+//
+//   (They seem to think the ">" in "5 > 0" marks the end of the
+//   template argument list.)
+//
+// - The array size is (bool(expr) ? 1 : -1), instead of simply
+//
+//     ((expr) ? 1 : -1).
+//
+//   This is to avoid running into a bug in MS VC 7.1, which
+//   causes ((0.0) ? 1 : -1) to incorrectly evaluate to 1.
+
+// StaticAssertTypeEqHelper is used by StaticAssertTypeEq defined in gtest.h.
+//
+// This template is declared, but intentionally undefined.
+template <typename T1, typename T2>
+struct StaticAssertTypeEqHelper;
+
+template <typename T>
+struct StaticAssertTypeEqHelper<T, T> {};
+
+#if GTEST_HAS_GLOBAL_STRING
+typedef ::string string;
+#else
+typedef ::std::string string;
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+#if GTEST_HAS_GLOBAL_WSTRING
+typedef ::wstring wstring;
+#elif GTEST_HAS_STD_WSTRING
+typedef ::std::wstring wstring;
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+// A helper for suppressing warnings on constant condition.  It just
+// returns 'condition'.
+GTEST_API_ bool IsTrue(bool condition);
+
+// Defines scoped_ptr.
+
+// This implementation of scoped_ptr is PARTIAL - it only contains
+// enough stuff to satisfy Google Test's need.
+template <typename T>
+class scoped_ptr {
+ public:
+  typedef T element_type;
+
+  explicit scoped_ptr(T* p = NULL) : ptr_(p) {}
+  ~scoped_ptr() { reset(); }
+
+  T& operator*() const { return *ptr_; }
+  T* operator->() const { return ptr_; }
+  T* get() const { return ptr_; }
+
+  T* release() {
+    T* const ptr = ptr_;
+    ptr_ = NULL;
+    return ptr;
+  }
+
+  void reset(T* p = NULL) {
+    if (p != ptr_) {
+      if (IsTrue(sizeof(T) > 0)) {  // Makes sure T is a complete type.
+        delete ptr_;
+      }
+      ptr_ = p;
+    }
+  }
+
+ private:
+  T* ptr_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(scoped_ptr);
+};
+
+// Defines RE.
+
+// A simple C++ wrapper for <regex.h>.  It uses the POSIX Extended
+// Regular Expression syntax.
+class GTEST_API_ RE {
+ public:
+  // A copy constructor is required by the Standard to initialize object
+  // references from r-values.
+  RE(const RE& other) { Init(other.pattern()); }
+
+  // Constructs an RE from a string.
+  RE(const ::std::string& regex) { Init(regex.c_str()); }  // NOLINT
+
+#if GTEST_HAS_GLOBAL_STRING
+
+  RE(const ::string& regex) { Init(regex.c_str()); }  // NOLINT
+
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+  RE(const char* regex) { Init(regex); }  // NOLINT
+  ~RE();
+
+  // Returns the string representation of the regex.
+  const char* pattern() const { return pattern_; }
+
+  // FullMatch(str, re) returns true iff regular expression re matches
+  // the entire str.
+  // PartialMatch(str, re) returns true iff regular expression re
+  // matches a substring of str (including str itself).
+  //
+  // TODO(wan@google.com): make FullMatch() and PartialMatch() work
+  // when str contains NUL characters.
+  static bool FullMatch(const ::std::string& str, const RE& re) {
+    return FullMatch(str.c_str(), re);
+  }
+  static bool PartialMatch(const ::std::string& str, const RE& re) {
+    return PartialMatch(str.c_str(), re);
+  }
+
+#if GTEST_HAS_GLOBAL_STRING
+
+  static bool FullMatch(const ::string& str, const RE& re) {
+    return FullMatch(str.c_str(), re);
+  }
+  static bool PartialMatch(const ::string& str, const RE& re) {
+    return PartialMatch(str.c_str(), re);
+  }
+
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+  static bool FullMatch(const char* str, const RE& re);
+  static bool PartialMatch(const char* str, const RE& re);
+
+ private:
+  void Init(const char* regex);
+
+  // We use a const char* instead of an std::string, as Google Test used to be
+  // used where std::string is not available.  TODO(wan@google.com): change to
+  // std::string.
+  const char* pattern_;
+  bool is_valid_;
+
+#if GTEST_USES_POSIX_RE
+
+  regex_t full_regex_;     // For FullMatch().
+  regex_t partial_regex_;  // For PartialMatch().
+
+#else  // GTEST_USES_SIMPLE_RE
+
+  const char* full_pattern_;  // For FullMatch();
+
+#endif
+
+  GTEST_DISALLOW_ASSIGN_(RE);
+};
+
+// Formats a source file path and a line number as they would appear
+// in an error message from the compiler used to compile this code.
+GTEST_API_ ::std::string FormatFileLocation(const char* file, int line);
+
+// Formats a file location for compiler-independent XML output.
+// Although this function is not platform dependent, we put it next to
+// FormatFileLocation in order to contrast the two functions.
+GTEST_API_ ::std::string FormatCompilerIndependentFileLocation(const char* file,
+                                                               int line);
+
+// Defines logging utilities:
+//   GTEST_LOG_(severity) - logs messages at the specified severity level. The
+//                          message itself is streamed into the macro.
+//   LogToStderr()  - directs all log messages to stderr.
+//   FlushInfoLog() - flushes informational log messages.
+
+enum GTestLogSeverity {
+  GTEST_INFO,
+  GTEST_WARNING,
+  GTEST_ERROR,
+  GTEST_FATAL
+};
+
+// Formats log entry severity, provides a stream object for streaming the
+// log message, and terminates the message with a newline when going out of
+// scope.
+class GTEST_API_ GTestLog {
+ public:
+  GTestLog(GTestLogSeverity severity, const char* file, int line);
+
+  // Flushes the buffers and, if severity is GTEST_FATAL, aborts the program.
+  ~GTestLog();
+
+  ::std::ostream& GetStream() { return ::std::cerr; }
+
+ private:
+  const GTestLogSeverity severity_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(GTestLog);
+};
+
+#define GTEST_LOG_(severity) \
+    ::testing::internal::GTestLog(::testing::internal::GTEST_##severity, \
+                                  __FILE__, __LINE__).GetStream()
+
+inline void LogToStderr() {}
+inline void FlushInfoLog() { fflush(NULL); }
+
+// INTERNAL IMPLEMENTATION - DO NOT USE.
+//
+// GTEST_CHECK_ is an all-mode assert. It aborts the program if the condition
+// is not satisfied.
+//  Synopsys:
+//    GTEST_CHECK_(boolean_condition);
+//     or
+//    GTEST_CHECK_(boolean_condition) << "Additional message";
+//
+//    This checks the condition and if the condition is not satisfied
+//    it prints message about the condition violation, including the
+//    condition itself, plus additional message streamed into it, if any,
+//    and then it aborts the program. It aborts the program irrespective of
+//    whether it is built in the debug mode or not.
+#define GTEST_CHECK_(condition) \
+    GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+    if (::testing::internal::IsTrue(condition)) \
+      ; \
+    else \
+      GTEST_LOG_(FATAL) << "Condition " #condition " failed. "
+
+// An all-mode assert to verify that the given POSIX-style function
+// call returns 0 (indicating success).  Known limitation: this
+// doesn't expand to a balanced 'if' statement, so enclose the macro
+// in {} if you need to use it as the only statement in an 'if'
+// branch.
+#define GTEST_CHECK_POSIX_SUCCESS_(posix_call) \
+  if (const int gtest_error = (posix_call)) \
+    GTEST_LOG_(FATAL) << #posix_call << "failed with error " \
+                      << gtest_error
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Use ImplicitCast_ as a safe version of static_cast for upcasting in
+// the type hierarchy (e.g. casting a Foo* to a SuperclassOfFoo* or a
+// const Foo*).  When you use ImplicitCast_, the compiler checks that
+// the cast is safe.  Such explicit ImplicitCast_s are necessary in
+// surprisingly many situations where C++ demands an exact type match
+// instead of an argument type convertable to a target type.
+//
+// The syntax for using ImplicitCast_ is the same as for static_cast:
+//
+//   ImplicitCast_<ToType>(expr)
+//
+// ImplicitCast_ would have been part of the C++ standard library,
+// but the proposal was submitted too late.  It will probably make
+// its way into the language in the future.
+//
+// This relatively ugly name is intentional. It prevents clashes with
+// similar functions users may have (e.g., implicit_cast). The internal
+// namespace alone is not enough because the function can be found by ADL.
+template<typename To>
+inline To ImplicitCast_(To x) { return x; }
+
+// When you upcast (that is, cast a pointer from type Foo to type
+// SuperclassOfFoo), it's fine to use ImplicitCast_<>, since upcasts
+// always succeed.  When you downcast (that is, cast a pointer from
+// type Foo to type SubclassOfFoo), static_cast<> isn't safe, because
+// how do you know the pointer is really of type SubclassOfFoo?  It
+// could be a bare Foo, or of type DifferentSubclassOfFoo.  Thus,
+// when you downcast, you should use this macro.  In debug mode, we
+// use dynamic_cast<> to double-check the downcast is legal (we die
+// if it's not).  In normal mode, we do the efficient static_cast<>
+// instead.  Thus, it's important to test in debug mode to make sure
+// the cast is legal!
+//    This is the only place in the code we should use dynamic_cast<>.
+// In particular, you SHOULDN'T be using dynamic_cast<> in order to
+// do RTTI (eg code like this:
+//    if (dynamic_cast<Subclass1>(foo)) HandleASubclass1Object(foo);
+//    if (dynamic_cast<Subclass2>(foo)) HandleASubclass2Object(foo);
+// You should design the code some other way not to need this.
+//
+// This relatively ugly name is intentional. It prevents clashes with
+// similar functions users may have (e.g., down_cast). The internal
+// namespace alone is not enough because the function can be found by ADL.
+template<typename To, typename From>  // use like this: DownCast_<T*>(foo);
+inline To DownCast_(From* f) {  // so we only accept pointers
+  // Ensures that To is a sub-type of From *.  This test is here only
+  // for compile-time type checking, and has no overhead in an
+  // optimized build at run-time, as it will be optimized away
+  // completely.
+  if (false) {
+    const To to = NULL;
+    ::testing::internal::ImplicitCast_<From*>(to);
+  }
+
+#if GTEST_HAS_RTTI
+  // RTTI: debug mode only!
+  GTEST_CHECK_(f == NULL || dynamic_cast<To>(f) != NULL);
+#endif
+  return static_cast<To>(f);
+}
+
+// Downcasts the pointer of type Base to Derived.
+// Derived must be a subclass of Base. The parameter MUST
+// point to a class of type Derived, not any subclass of it.
+// When RTTI is available, the function performs a runtime
+// check to enforce this.
+template <class Derived, class Base>
+Derived* CheckedDowncastToActualType(Base* base) {
+#if GTEST_HAS_RTTI
+  GTEST_CHECK_(typeid(*base) == typeid(Derived));
+  return dynamic_cast<Derived*>(base);  // NOLINT
+#else
+  return static_cast<Derived*>(base);  // Poor man's downcast.
+#endif
+}
+
+#if GTEST_HAS_STREAM_REDIRECTION
+
+// Defines the stderr capturer:
+//   CaptureStdout     - starts capturing stdout.
+//   GetCapturedStdout - stops capturing stdout and returns the captured string.
+//   CaptureStderr     - starts capturing stderr.
+//   GetCapturedStderr - stops capturing stderr and returns the captured string.
+//
+GTEST_API_ void CaptureStdout();
+GTEST_API_ std::string GetCapturedStdout();
+GTEST_API_ void CaptureStderr();
+GTEST_API_ std::string GetCapturedStderr();
+
+#endif  // GTEST_HAS_STREAM_REDIRECTION
+
+
+#if GTEST_HAS_DEATH_TEST
+
+const ::std::vector<testing::internal::string>& GetInjectableArgvs();
+void SetInjectableArgvs(const ::std::vector<testing::internal::string>*
+                             new_argvs);
+
+// A copy of all command line arguments.  Set by InitGoogleTest().
+extern ::std::vector<testing::internal::string> g_argvs;
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+// Defines synchronization primitives.
+
+#if GTEST_HAS_PTHREAD
+
+// Sleeps for (roughly) n milli-seconds.  This function is only for
+// testing Google Test's own constructs.  Don't use it in user tests,
+// either directly or indirectly.
+inline void SleepMilliseconds(int n) {
+  const timespec time = {
+    0,                  // 0 seconds.
+    n * 1000L * 1000L,  // And n ms.
+  };
+  nanosleep(&time, NULL);
+}
+
+// Allows a controller thread to pause execution of newly created
+// threads until notified.  Instances of this class must be created
+// and destroyed in the controller thread.
+//
+// This class is only for testing Google Test's own constructs. Do not
+// use it in user tests, either directly or indirectly.
+class Notification {
+ public:
+  Notification() : notified_(false) {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_init(&mutex_, NULL));
+  }
+  ~Notification() {
+    pthread_mutex_destroy(&mutex_);
+  }
+
+  // Notifies all threads created with this notification to start. Must
+  // be called from the controller thread.
+  void Notify() {
+    pthread_mutex_lock(&mutex_);
+    notified_ = true;
+    pthread_mutex_unlock(&mutex_);
+  }
+
+  // Blocks until the controller thread notifies. Must be called from a test
+  // thread.
+  void WaitForNotification() {
+    for (;;) {
+      pthread_mutex_lock(&mutex_);
+      const bool notified = notified_;
+      pthread_mutex_unlock(&mutex_);
+      if (notified)
+        break;
+      SleepMilliseconds(10);
+    }
+  }
+
+ private:
+  pthread_mutex_t mutex_;
+  bool notified_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Notification);
+};
+
+// As a C-function, ThreadFuncWithCLinkage cannot be templated itself.
+// Consequently, it cannot select a correct instantiation of ThreadWithParam
+// in order to call its Run(). Introducing ThreadWithParamBase as a
+// non-templated base class for ThreadWithParam allows us to bypass this
+// problem.
+class ThreadWithParamBase {
+ public:
+  virtual ~ThreadWithParamBase() {}
+  virtual void Run() = 0;
+};
+
+// pthread_create() accepts a pointer to a function type with the C linkage.
+// According to the Standard (7.5/1), function types with different linkages
+// are different even if they are otherwise identical.  Some compilers (for
+// example, SunStudio) treat them as different types.  Since class methods
+// cannot be defined with C-linkage we need to define a free C-function to
+// pass into pthread_create().
+extern "C" inline void* ThreadFuncWithCLinkage(void* thread) {
+  static_cast<ThreadWithParamBase*>(thread)->Run();
+  return NULL;
+}
+
+// Helper class for testing Google Test's multi-threading constructs.
+// To use it, write:
+//
+//   void ThreadFunc(int param) { /* Do things with param */ }
+//   Notification thread_can_start;
+//   ...
+//   // The thread_can_start parameter is optional; you can supply NULL.
+//   ThreadWithParam<int> thread(&ThreadFunc, 5, &thread_can_start);
+//   thread_can_start.Notify();
+//
+// These classes are only for testing Google Test's own constructs. Do
+// not use them in user tests, either directly or indirectly.
+template <typename T>
+class ThreadWithParam : public ThreadWithParamBase {
+ public:
+  typedef void (*UserThreadFunc)(T);
+
+  ThreadWithParam(
+      UserThreadFunc func, T param, Notification* thread_can_start)
+      : func_(func),
+        param_(param),
+        thread_can_start_(thread_can_start),
+        finished_(false) {
+    ThreadWithParamBase* const base = this;
+    // The thread can be created only after all fields except thread_
+    // have been initialized.
+    GTEST_CHECK_POSIX_SUCCESS_(
+        pthread_create(&thread_, 0, &ThreadFuncWithCLinkage, base));
+  }
+  ~ThreadWithParam() { Join(); }
+
+  void Join() {
+    if (!finished_) {
+      GTEST_CHECK_POSIX_SUCCESS_(pthread_join(thread_, 0));
+      finished_ = true;
+    }
+  }
+
+  virtual void Run() {
+    if (thread_can_start_ != NULL)
+      thread_can_start_->WaitForNotification();
+    func_(param_);
+  }
+
+ private:
+  const UserThreadFunc func_;  // User-supplied thread function.
+  const T param_;  // User-supplied parameter to the thread function.
+  // When non-NULL, used to block execution until the controller thread
+  // notifies.
+  Notification* const thread_can_start_;
+  bool finished_;  // true iff we know that the thread function has finished.
+  pthread_t thread_;  // The native thread object.
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ThreadWithParam);
+};
+
+// MutexBase and Mutex implement mutex on pthreads-based platforms. They
+// are used in conjunction with class MutexLock:
+//
+//   Mutex mutex;
+//   ...
+//   MutexLock lock(&mutex);  // Acquires the mutex and releases it at the end
+//                            // of the current scope.
+//
+// MutexBase implements behavior for both statically and dynamically
+// allocated mutexes.  Do not use MutexBase directly.  Instead, write
+// the following to define a static mutex:
+//
+//   GTEST_DEFINE_STATIC_MUTEX_(g_some_mutex);
+//
+// You can forward declare a static mutex like this:
+//
+//   GTEST_DECLARE_STATIC_MUTEX_(g_some_mutex);
+//
+// To create a dynamic mutex, just define an object of type Mutex.
+class MutexBase {
+ public:
+  // Acquires this mutex.
+  void Lock() {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_lock(&mutex_));
+    owner_ = pthread_self();
+    has_owner_ = true;
+  }
+
+  // Releases this mutex.
+  void Unlock() {
+    // Since the lock is being released the owner_ field should no longer be
+    // considered valid. We don't protect writing to has_owner_ here, as it's
+    // the caller's responsibility to ensure that the current thread holds the
+    // mutex when this is called.
+    has_owner_ = false;
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_unlock(&mutex_));
+  }
+
+  // Does nothing if the current thread holds the mutex. Otherwise, crashes
+  // with high probability.
+  void AssertHeld() const {
+    GTEST_CHECK_(has_owner_ && pthread_equal(owner_, pthread_self()))
+        << "The current thread is not holding the mutex @" << this;
+  }
+
+  // A static mutex may be used before main() is entered.  It may even
+  // be used before the dynamic initialization stage.  Therefore we
+  // must be able to initialize a static mutex object at link time.
+  // This means MutexBase has to be a POD and its member variables
+  // have to be public.
+ public:
+  pthread_mutex_t mutex_;  // The underlying pthread mutex.
+  // has_owner_ indicates whether the owner_ field below contains a valid thread
+  // ID and is therefore safe to inspect (e.g., to use in pthread_equal()). All
+  // accesses to the owner_ field should be protected by a check of this field.
+  // An alternative might be to memset() owner_ to all zeros, but there's no
+  // guarantee that a zero'd pthread_t is necessarily invalid or even different
+  // from pthread_self().
+  bool has_owner_;
+  pthread_t owner_;  // The thread holding the mutex.
+};
+
+// Forward-declares a static mutex.
+# define GTEST_DECLARE_STATIC_MUTEX_(mutex) \
+    extern ::testing::internal::MutexBase mutex
+
+// Defines and statically (i.e. at link time) initializes a static mutex.
+// The initialization list here does not explicitly initialize each field,
+// instead relying on default initialization for the unspecified fields. In
+// particular, the owner_ field (a pthread_t) is not explicitly initialized.
+// This allows initialization to work whether pthread_t is a scalar or struct.
+// The flag -Wmissing-field-initializers must not be specified for this to work.
+# define GTEST_DEFINE_STATIC_MUTEX_(mutex) \
+    ::testing::internal::MutexBase mutex = { PTHREAD_MUTEX_INITIALIZER, false }
+
+// The Mutex class can only be used for mutexes created at runtime. It
+// shares its API with MutexBase otherwise.
+class Mutex : public MutexBase {
+ public:
+  Mutex() {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_init(&mutex_, NULL));
+    has_owner_ = false;
+  }
+  ~Mutex() {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_destroy(&mutex_));
+  }
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Mutex);
+};
+
+// We cannot name this class MutexLock as the ctor declaration would
+// conflict with a macro named MutexLock, which is defined on some
+// platforms.  Hence the typedef trick below.
+class GTestMutexLock {
+ public:
+  explicit GTestMutexLock(MutexBase* mutex)
+      : mutex_(mutex) { mutex_->Lock(); }
+
+  ~GTestMutexLock() { mutex_->Unlock(); }
+
+ private:
+  MutexBase* const mutex_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(GTestMutexLock);
+};
+
+typedef GTestMutexLock MutexLock;
+
+// Helpers for ThreadLocal.
+
+// pthread_key_create() requires DeleteThreadLocalValue() to have
+// C-linkage.  Therefore it cannot be templatized to access
+// ThreadLocal<T>.  Hence the need for class
+// ThreadLocalValueHolderBase.
+class ThreadLocalValueHolderBase {
+ public:
+  virtual ~ThreadLocalValueHolderBase() {}
+};
+
+// Called by pthread to delete thread-local data stored by
+// pthread_setspecific().
+extern "C" inline void DeleteThreadLocalValue(void* value_holder) {
+  delete static_cast<ThreadLocalValueHolderBase*>(value_holder);
+}
+
+// Implements thread-local storage on pthreads-based systems.
+//
+//   // Thread 1
+//   ThreadLocal<int> tl(100);  // 100 is the default value for each thread.
+//
+//   // Thread 2
+//   tl.set(150);  // Changes the value for thread 2 only.
+//   EXPECT_EQ(150, tl.get());
+//
+//   // Thread 1
+//   EXPECT_EQ(100, tl.get());  // In thread 1, tl has the original value.
+//   tl.set(200);
+//   EXPECT_EQ(200, tl.get());
+//
+// The template type argument T must have a public copy constructor.
+// In addition, the default ThreadLocal constructor requires T to have
+// a public default constructor.
+//
+// An object managed for a thread by a ThreadLocal instance is deleted
+// when the thread exits.  Or, if the ThreadLocal instance dies in
+// that thread, when the ThreadLocal dies.  It's the user's
+// responsibility to ensure that all other threads using a ThreadLocal
+// have exited when it dies, or the per-thread objects for those
+// threads will not be deleted.
+//
+// Google Test only uses global ThreadLocal objects.  That means they
+// will die after main() has returned.  Therefore, no per-thread
+// object managed by Google Test will be leaked as long as all threads
+// using Google Test have exited when main() returns.
+template <typename T>
+class ThreadLocal {
+ public:
+  ThreadLocal() : key_(CreateKey()),
+                  default_() {}
+  explicit ThreadLocal(const T& value) : key_(CreateKey()),
+                                         default_(value) {}
+
+  ~ThreadLocal() {
+    // Destroys the managed object for the current thread, if any.
+    DeleteThreadLocalValue(pthread_getspecific(key_));
+
+    // Releases resources associated with the key.  This will *not*
+    // delete managed objects for other threads.
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_key_delete(key_));
+  }
+
+  T* pointer() { return GetOrCreateValue(); }
+  const T* pointer() const { return GetOrCreateValue(); }
+  const T& get() const { return *pointer(); }
+  void set(const T& value) { *pointer() = value; }
+
+ private:
+  // Holds a value of type T.
+  class ValueHolder : public ThreadLocalValueHolderBase {
+   public:
+    explicit ValueHolder(const T& value) : value_(value) {}
+
+    T* pointer() { return &value_; }
+
+   private:
+    T value_;
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(ValueHolder);
+  };
+
+  static pthread_key_t CreateKey() {
+    pthread_key_t key;
+    // When a thread exits, DeleteThreadLocalValue() will be called on
+    // the object managed for that thread.
+    GTEST_CHECK_POSIX_SUCCESS_(
+        pthread_key_create(&key, &DeleteThreadLocalValue));
+    return key;
+  }
+
+  T* GetOrCreateValue() const {
+    ThreadLocalValueHolderBase* const holder =
+        static_cast<ThreadLocalValueHolderBase*>(pthread_getspecific(key_));
+    if (holder != NULL) {
+      return CheckedDowncastToActualType<ValueHolder>(holder)->pointer();
+    }
+
+    ValueHolder* const new_holder = new ValueHolder(default_);
+    ThreadLocalValueHolderBase* const holder_base = new_holder;
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_setspecific(key_, holder_base));
+    return new_holder->pointer();
+  }
+
+  // A key pthreads uses for looking up per-thread values.
+  const pthread_key_t key_;
+  const T default_;  // The default value for each thread.
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ThreadLocal);
+};
+
+# define GTEST_IS_THREADSAFE 1
+
+#else  // GTEST_HAS_PTHREAD
+
+// A dummy implementation of synchronization primitives (mutex, lock,
+// and thread-local variable).  Necessary for compiling Google Test where
+// mutex is not supported - using Google Test in multiple threads is not
+// supported on such platforms.
+
+class Mutex {
+ public:
+  Mutex() {}
+  void Lock() {}
+  void Unlock() {}
+  void AssertHeld() const {}
+};
+
+# define GTEST_DECLARE_STATIC_MUTEX_(mutex) \
+  extern ::testing::internal::Mutex mutex
+
+# define GTEST_DEFINE_STATIC_MUTEX_(mutex) ::testing::internal::Mutex mutex
+
+class GTestMutexLock {
+ public:
+  explicit GTestMutexLock(Mutex*) {}  // NOLINT
+};
+
+typedef GTestMutexLock MutexLock;
+
+template <typename T>
+class ThreadLocal {
+ public:
+  ThreadLocal() : value_() {}
+  explicit ThreadLocal(const T& value) : value_(value) {}
+  T* pointer() { return &value_; }
+  const T* pointer() const { return &value_; }
+  const T& get() const { return value_; }
+  void set(const T& value) { value_ = value; }
+ private:
+  T value_;
+};
+
+// The above synchronization primitives have dummy implementations.
+// Therefore Google Test is not thread-safe.
+# define GTEST_IS_THREADSAFE 0
+
+#endif  // GTEST_HAS_PTHREAD
+
+// Returns the number of threads running in the process, or 0 to indicate that
+// we cannot detect it.
+GTEST_API_ size_t GetThreadCount();
+
+// Passing non-POD classes through ellipsis (...) crashes the ARM
+// compiler and generates a warning in Sun Studio.  The Nokia Symbian
+// and the IBM XL C/C++ compiler try to instantiate a copy constructor
+// for objects passed through ellipsis (...), failing for uncopyable
+// objects.  We define this to ensure that only POD is passed through
+// ellipsis on these systems.
+#if defined(__SYMBIAN32__) || defined(__IBMCPP__) || defined(__SUNPRO_CC)
+// We lose support for NULL detection where the compiler doesn't like
+// passing non-POD classes through ellipsis (...).
+# define GTEST_ELLIPSIS_NEEDS_POD_ 1
+#else
+# define GTEST_CAN_COMPARE_NULL 1
+#endif
+
+// The Nokia Symbian and IBM XL C/C++ compilers cannot decide between
+// const T& and const T* in a function template.  These compilers
+// _can_ decide between class template specializations for T and T*,
+// so a tr1::type_traits-like is_pointer works.
+#if defined(__SYMBIAN32__) || defined(__IBMCPP__)
+# define GTEST_NEEDS_IS_POINTER_ 1
+#endif
+
+template <bool bool_value>
+struct bool_constant {
+  typedef bool_constant<bool_value> type;
+  static const bool value = bool_value;
+};
+template <bool bool_value> const bool bool_constant<bool_value>::value;
+
+typedef bool_constant<false> false_type;
+typedef bool_constant<true> true_type;
+
+template <typename T>
+struct is_pointer : public false_type {};
+
+template <typename T>
+struct is_pointer<T*> : public true_type {};
+
+template <typename Iterator>
+struct IteratorTraits {
+  typedef typename Iterator::value_type value_type;
+};
+
+template <typename T>
+struct IteratorTraits<T*> {
+  typedef T value_type;
+};
+
+template <typename T>
+struct IteratorTraits<const T*> {
+  typedef T value_type;
+};
+
+#if GTEST_OS_WINDOWS
+# define GTEST_PATH_SEP_ "\\"
+# define GTEST_HAS_ALT_PATH_SEP_ 1
+// The biggest signed integer type the compiler supports.
+typedef __int64 BiggestInt;
+#else
+# define GTEST_PATH_SEP_ "/"
+# define GTEST_HAS_ALT_PATH_SEP_ 0
+typedef long long BiggestInt;  // NOLINT
+#endif  // GTEST_OS_WINDOWS
+
+// Utilities for char.
+
+// isspace(int ch) and friends accept an unsigned char or EOF.  char
+// may be signed, depending on the compiler (or compiler flags).
+// Therefore we need to cast a char to unsigned char before calling
+// isspace(), etc.
+
+inline bool IsAlpha(char ch) {
+  return isalpha(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsAlNum(char ch) {
+  return isalnum(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsDigit(char ch) {
+  return isdigit(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsLower(char ch) {
+  return islower(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsSpace(char ch) {
+  return isspace(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsUpper(char ch) {
+  return isupper(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsXDigit(char ch) {
+  return isxdigit(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsXDigit(wchar_t ch) {
+  const unsigned char low_byte = static_cast<unsigned char>(ch);
+  return ch == low_byte && isxdigit(low_byte) != 0;
+}
+
+inline char ToLower(char ch) {
+  return static_cast<char>(tolower(static_cast<unsigned char>(ch)));
+}
+inline char ToUpper(char ch) {
+  return static_cast<char>(toupper(static_cast<unsigned char>(ch)));
+}
+
+// The testing::internal::posix namespace holds wrappers for common
+// POSIX functions.  These wrappers hide the differences between
+// Windows/MSVC and POSIX systems.  Since some compilers define these
+// standard functions as macros, the wrapper cannot have the same name
+// as the wrapped function.
+
+namespace posix {
+
+// Functions with a different name on Windows.
+
+#if GTEST_OS_WINDOWS
+
+typedef struct _stat StatStruct;
+
+# ifdef __BORLANDC__
+inline int IsATTY(int fd) { return isatty(fd); }
+inline int StrCaseCmp(const char* s1, const char* s2) {
+  return stricmp(s1, s2);
+}
+inline char* StrDup(const char* src) { return strdup(src); }
+# else  // !__BORLANDC__
+#  if GTEST_OS_WINDOWS_MOBILE
+inline int IsATTY(int /* fd */) { return 0; }
+#  else
+inline int IsATTY(int fd) { return _isatty(fd); }
+#  endif  // GTEST_OS_WINDOWS_MOBILE
+inline int StrCaseCmp(const char* s1, const char* s2) {
+  return _stricmp(s1, s2);
+}
+inline char* StrDup(const char* src) { return _strdup(src); }
+# endif  // __BORLANDC__
+
+# if GTEST_OS_WINDOWS_MOBILE
+inline int FileNo(FILE* file) { return reinterpret_cast<int>(_fileno(file)); }
+// Stat(), RmDir(), and IsDir() are not needed on Windows CE at this
+// time and thus not defined there.
+# else
+inline int FileNo(FILE* file) { return _fileno(file); }
+inline int Stat(const char* path, StatStruct* buf) { return _stat(path, buf); }
+inline int RmDir(const char* dir) { return _rmdir(dir); }
+inline bool IsDir(const StatStruct& st) {
+  return (_S_IFDIR & st.st_mode) != 0;
+}
+# endif  // GTEST_OS_WINDOWS_MOBILE
+
+#else
+
+typedef struct stat StatStruct;
+
+inline int FileNo(FILE* file) { return fileno(file); }
+inline int IsATTY(int fd) { return isatty(fd); }
+inline int Stat(const char* path, StatStruct* buf) { return stat(path, buf); }
+inline int StrCaseCmp(const char* s1, const char* s2) {
+  return strcasecmp(s1, s2);
+}
+inline char* StrDup(const char* src) { return strdup(src); }
+inline int RmDir(const char* dir) { return rmdir(dir); }
+inline bool IsDir(const StatStruct& st) { return S_ISDIR(st.st_mode); }
+
+#endif  // GTEST_OS_WINDOWS
+
+// Functions deprecated by MSVC 8.0.
+
+#ifdef _MSC_VER
+// Temporarily disable warning 4996 (deprecated function).
+# pragma warning(push)
+# pragma warning(disable:4996)
+#endif
+
+inline const char* StrNCpy(char* dest, const char* src, size_t n) {
+  return strncpy(dest, src, n);
+}
+
+// ChDir(), FReopen(), FDOpen(), Read(), Write(), Close(), and
+// StrError() aren't needed on Windows CE at this time and thus not
+// defined there.
+
+#if !GTEST_OS_WINDOWS_MOBILE
+inline int ChDir(const char* dir) { return chdir(dir); }
+#endif
+inline FILE* FOpen(const char* path, const char* mode) {
+  return fopen(path, mode);
+}
+#if !GTEST_OS_WINDOWS_MOBILE
+inline FILE *FReopen(const char* path, const char* mode, FILE* stream) {
+  return freopen(path, mode, stream);
+}
+inline FILE* FDOpen(int fd, const char* mode) { return fdopen(fd, mode); }
+#endif
+inline int FClose(FILE* fp) { return fclose(fp); }
+#if !GTEST_OS_WINDOWS_MOBILE
+inline int Read(int fd, void* buf, unsigned int count) {
+  return static_cast<int>(read(fd, buf, count));
+}
+inline int Write(int fd, const void* buf, unsigned int count) {
+  return static_cast<int>(write(fd, buf, count));
+}
+inline int Close(int fd) { return close(fd); }
+inline const char* StrError(int errnum) { return strerror(errnum); }
+#endif
+inline const char* GetEnv(const char* name) {
+#if GTEST_OS_WINDOWS_MOBILE
+  // We are on Windows CE, which has no environment variables.
+  return NULL;
+#elif defined(__BORLANDC__) || defined(__SunOS_5_8) || defined(__SunOS_5_9)
+  // Environment variables which we programmatically clear will be set to the
+  // empty string rather than unset (NULL).  Handle that case.
+  const char* const env = getenv(name);
+  return (env != NULL && env[0] != '\0') ? env : NULL;
+#else
+  return getenv(name);
+#endif
+}
+
+#ifdef _MSC_VER
+# pragma warning(pop)  // Restores the warning state.
+#endif
+
+#if GTEST_OS_WINDOWS_MOBILE
+// Windows CE has no C library. The abort() function is used in
+// several places in Google Test. This implementation provides a reasonable
+// imitation of standard behaviour.
+void Abort();
+#else
+inline void Abort() { abort(); }
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+}  // namespace posix
+
+// MSVC "deprecates" snprintf and issues warnings wherever it is used.  In
+// order to avoid these warnings, we need to use _snprintf or _snprintf_s on
+// MSVC-based platforms.  We map the GTEST_SNPRINTF_ macro to the appropriate
+// function in order to achieve that.  We use macro definition here because
+// snprintf is a variadic function.
+#if _MSC_VER >= 1400 && !GTEST_OS_WINDOWS_MOBILE
+// MSVC 2005 and above support variadic macros.
+# define GTEST_SNPRINTF_(buffer, size, format, ...) \
+     _snprintf_s(buffer, size, size, format, __VA_ARGS__)
+#elif defined(_MSC_VER)
+// Windows CE does not define _snprintf_s and MSVC prior to 2005 doesn't
+// complain about _snprintf.
+# define GTEST_SNPRINTF_ _snprintf
+#else
+# define GTEST_SNPRINTF_ snprintf
+#endif
+
+// The maximum number a BiggestInt can represent.  This definition
+// works no matter BiggestInt is represented in one's complement or
+// two's complement.
+//
+// We cannot rely on numeric_limits in STL, as __int64 and long long
+// are not part of standard C++ and numeric_limits doesn't need to be
+// defined for them.
+const BiggestInt kMaxBiggestInt =
+    ~(static_cast<BiggestInt>(1) << (8*sizeof(BiggestInt) - 1));
+
+// This template class serves as a compile-time function from size to
+// type.  It maps a size in bytes to a primitive type with that
+// size. e.g.
+//
+//   TypeWithSize<4>::UInt
+//
+// is typedef-ed to be unsigned int (unsigned integer made up of 4
+// bytes).
+//
+// Such functionality should belong to STL, but I cannot find it
+// there.
+//
+// Google Test uses this class in the implementation of floating-point
+// comparison.
+//
+// For now it only handles UInt (unsigned int) as that's all Google Test
+// needs.  Other types can be easily added in the future if need
+// arises.
+template <size_t size>
+class TypeWithSize {
+ public:
+  // This prevents the user from using TypeWithSize<N> with incorrect
+  // values of N.
+  typedef void UInt;
+};
+
+// The specialization for size 4.
+template <>
+class TypeWithSize<4> {
+ public:
+  // unsigned int has size 4 in both gcc and MSVC.
+  //
+  // As base/basictypes.h doesn't compile on Windows, we cannot use
+  // uint32, uint64, and etc here.
+  typedef int Int;
+  typedef unsigned int UInt;
+};
+
+// The specialization for size 8.
+template <>
+class TypeWithSize<8> {
+ public:
+#if GTEST_OS_WINDOWS
+  typedef __int64 Int;
+  typedef unsigned __int64 UInt;
+#else
+  typedef long long Int;  // NOLINT
+  typedef unsigned long long UInt;  // NOLINT
+#endif  // GTEST_OS_WINDOWS
+};
+
+// Integer types of known sizes.
+typedef TypeWithSize<4>::Int Int32;
+typedef TypeWithSize<4>::UInt UInt32;
+typedef TypeWithSize<8>::Int Int64;
+typedef TypeWithSize<8>::UInt UInt64;
+typedef TypeWithSize<8>::Int TimeInMillis;  // Represents time in milliseconds.
+
+// Utilities for command line flags and environment variables.
+
+// Macro for referencing flags.
+#define GTEST_FLAG(name) FLAGS_gtest_##name
+
+// Macros for declaring flags.
+#define GTEST_DECLARE_bool_(name) GTEST_API_ extern bool GTEST_FLAG(name)
+#define GTEST_DECLARE_int32_(name) \
+    GTEST_API_ extern ::testing::internal::Int32 GTEST_FLAG(name)
+#define GTEST_DECLARE_string_(name) \
+    GTEST_API_ extern ::std::string GTEST_FLAG(name)
+
+// Macros for defining flags.
+#define GTEST_DEFINE_bool_(name, default_val, doc) \
+    GTEST_API_ bool GTEST_FLAG(name) = (default_val)
+#define GTEST_DEFINE_int32_(name, default_val, doc) \
+    GTEST_API_ ::testing::internal::Int32 GTEST_FLAG(name) = (default_val)
+#define GTEST_DEFINE_string_(name, default_val, doc) \
+    GTEST_API_ ::std::string GTEST_FLAG(name) = (default_val)
+
+// Thread annotations
+#define GTEST_EXCLUSIVE_LOCK_REQUIRED_(locks)
+#define GTEST_LOCK_EXCLUDED_(locks)
+
+// Parses 'str' for a 32-bit signed integer.  If successful, writes the result
+// to *value and returns true; otherwise leaves *value unchanged and returns
+// false.
+// TODO(chandlerc): Find a better way to refactor flag and environment parsing
+// out of both gtest-port.cc and gtest.cc to avoid exporting this utility
+// function.
+bool ParseInt32(const Message& src_text, const char* str, Int32* value);
+
+// Parses a bool/Int32/string from the environment variable
+// corresponding to the given Google Test flag.
+bool BoolFromGTestEnv(const char* flag, bool default_val);
+GTEST_API_ Int32 Int32FromGTestEnv(const char* flag, Int32 default_val);
+const char* StringFromGTestEnv(const char* flag, const char* default_val);
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_H_
+
+#if GTEST_OS_LINUX
+# include <stdlib.h>
+# include <sys/types.h>
+# include <sys/wait.h>
+# include <unistd.h>
+#endif  // GTEST_OS_LINUX
+
+#if GTEST_HAS_EXCEPTIONS
+# include <stdexcept>
+#endif
+
+#include <ctype.h>
+#include <float.h>
+#include <string.h>
+#include <iomanip>
+#include <limits>
+#include <set>
+
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines the Message class.
+//
+// IMPORTANT NOTE: Due to limitation of the C++ language, we have to
+// leave some internal implementation details in this header file.
+// They are clearly marked by comments like this:
+//
+//   // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+//
+// Such code is NOT meant to be used by a user directly, and is subject
+// to CHANGE WITHOUT NOTICE.  Therefore DO NOT DEPEND ON IT in a user
+// program!
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_MESSAGE_H_
+#define GTEST_INCLUDE_GTEST_GTEST_MESSAGE_H_
+
+#include <limits>
+
+
+// Ensures that there is at least one operator<< in the global namespace.
+// See Message& operator<<(...) below for why.
+void operator<<(const testing::internal::Secret&, int);
+
+namespace testing {
+
+// The Message class works like an ostream repeater.
+//
+// Typical usage:
+//
+//   1. You stream a bunch of values to a Message object.
+//      It will remember the text in a stringstream.
+//   2. Then you stream the Message object to an ostream.
+//      This causes the text in the Message to be streamed
+//      to the ostream.
+//
+// For example;
+//
+//   testing::Message foo;
+//   foo << 1 << " != " << 2;
+//   std::cout << foo;
+//
+// will print "1 != 2".
+//
+// Message is not intended to be inherited from.  In particular, its
+// destructor is not virtual.
+//
+// Note that stringstream behaves differently in gcc and in MSVC.  You
+// can stream a NULL char pointer to it in the former, but not in the
+// latter (it causes an access violation if you do).  The Message
+// class hides this difference by treating a NULL char pointer as
+// "(null)".
+class GTEST_API_ Message {
+ private:
+  // The type of basic IO manipulators (endl, ends, and flush) for
+  // narrow streams.
+  typedef std::ostream& (*BasicNarrowIoManip)(std::ostream&);
+
+ public:
+  // Constructs an empty Message.
+  Message();
+
+  // Copy constructor.
+  Message(const Message& msg) : ss_(new ::std::stringstream) {  // NOLINT
+    *ss_ << msg.GetString();
+  }
+
+  // Constructs a Message from a C-string.
+  explicit Message(const char* str) : ss_(new ::std::stringstream) {
+    *ss_ << str;
+  }
+
+#if GTEST_OS_SYMBIAN
+  // Streams a value (either a pointer or not) to this object.
+  template <typename T>
+  inline Message& operator <<(const T& value) {
+    StreamHelper(typename internal::is_pointer<T>::type(), value);
+    return *this;
+  }
+#else
+  // Streams a non-pointer value to this object.
+  template <typename T>
+  inline Message& operator <<(const T& val) {
+    // Some libraries overload << for STL containers.  These
+    // overloads are defined in the global namespace instead of ::std.
+    //
+    // C++'s symbol lookup rule (i.e. Koenig lookup) says that these
+    // overloads are visible in either the std namespace or the global
+    // namespace, but not other namespaces, including the testing
+    // namespace which Google Test's Message class is in.
+    //
+    // To allow STL containers (and other types that has a << operator
+    // defined in the global namespace) to be used in Google Test
+    // assertions, testing::Message must access the custom << operator
+    // from the global namespace.  With this using declaration,
+    // overloads of << defined in the global namespace and those
+    // visible via Koenig lookup are both exposed in this function.
+    using ::operator <<;
+    *ss_ << val;
+    return *this;
+  }
+
+  // Streams a pointer value to this object.
+  //
+  // This function is an overload of the previous one.  When you
+  // stream a pointer to a Message, this definition will be used as it
+  // is more specialized.  (The C++ Standard, section
+  // [temp.func.order].)  If you stream a non-pointer, then the
+  // previous definition will be used.
+  //
+  // The reason for this overload is that streaming a NULL pointer to
+  // ostream is undefined behavior.  Depending on the compiler, you
+  // may get "0", "(nil)", "(null)", or an access violation.  To
+  // ensure consistent result across compilers, we always treat NULL
+  // as "(null)".
+  template <typename T>
+  inline Message& operator <<(T* const& pointer) {  // NOLINT
+    if (pointer == NULL) {
+      *ss_ << "(null)";
+    } else {
+      *ss_ << pointer;
+    }
+    return *this;
+  }
+#endif  // GTEST_OS_SYMBIAN
+
+  // Since the basic IO manipulators are overloaded for both narrow
+  // and wide streams, we have to provide this specialized definition
+  // of operator <<, even though its body is the same as the
+  // templatized version above.  Without this definition, streaming
+  // endl or other basic IO manipulators to Message will confuse the
+  // compiler.
+  Message& operator <<(BasicNarrowIoManip val) {
+    *ss_ << val;
+    return *this;
+  }
+
+  // Instead of 1/0, we want to see true/false for bool values.
+  Message& operator <<(bool b) {
+    return *this << (b ? "true" : "false");
+  }
+
+  // These two overloads allow streaming a wide C string to a Message
+  // using the UTF-8 encoding.
+  Message& operator <<(const wchar_t* wide_c_str);
+  Message& operator <<(wchar_t* wide_c_str);
+
+#if GTEST_HAS_STD_WSTRING
+  // Converts the given wide string to a narrow string using the UTF-8
+  // encoding, and streams the result to this Message object.
+  Message& operator <<(const ::std::wstring& wstr);
+#endif  // GTEST_HAS_STD_WSTRING
+
+#if GTEST_HAS_GLOBAL_WSTRING
+  // Converts the given wide string to a narrow string using the UTF-8
+  // encoding, and streams the result to this Message object.
+  Message& operator <<(const ::wstring& wstr);
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+  // Gets the text streamed to this object so far as an std::string.
+  // Each '\0' character in the buffer is replaced with "\\0".
+  //
+  // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+  std::string GetString() const;
+
+ private:
+
+#if GTEST_OS_SYMBIAN
+  // These are needed as the Nokia Symbian Compiler cannot decide between
+  // const T& and const T* in a function template. The Nokia compiler _can_
+  // decide between class template specializations for T and T*, so a
+  // tr1::type_traits-like is_pointer works, and we can overload on that.
+  template <typename T>
+  inline void StreamHelper(internal::true_type /*is_pointer*/, T* pointer) {
+    if (pointer == NULL) {
+      *ss_ << "(null)";
+    } else {
+      *ss_ << pointer;
+    }
+  }
+  template <typename T>
+  inline void StreamHelper(internal::false_type /*is_pointer*/,
+                           const T& value) {
+    // See the comments in Message& operator <<(const T&) above for why
+    // we need this using statement.
+    using ::operator <<;
+    *ss_ << value;
+  }
+#endif  // GTEST_OS_SYMBIAN
+
+  // We'll hold the text streamed to this object here.
+  const internal::scoped_ptr< ::std::stringstream> ss_;
+
+  // We declare (but don't implement) this to prevent the compiler
+  // from implementing the assignment operator.
+  void operator=(const Message&);
+};
+
+// Streams a Message to an ostream.
+inline std::ostream& operator <<(std::ostream& os, const Message& sb) {
+  return os << sb.GetString();
+}
+
+namespace internal {
+
+// Converts a streamable value to an std::string.  A NULL pointer is
+// converted to "(null)".  When the input value is a ::string,
+// ::std::string, ::wstring, or ::std::wstring object, each NUL
+// character in it is replaced with "\\0".
+template <typename T>
+std::string StreamableToString(const T& streamable) {
+  return (Message() << streamable).GetString();
+}
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_MESSAGE_H_
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan), eefacm@gmail.com (Sean Mcafee)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file declares the String class and functions used internally by
+// Google Test.  They are subject to change without notice. They should not used
+// by code external to Google Test.
+//
+// This header file is #included by <gtest/internal/gtest-internal.h>.
+// It should not be #included by other files.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_STRING_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_STRING_H_
+
+#ifdef __BORLANDC__
+// string.h is not guaranteed to provide strcpy on C++ Builder.
+# include <mem.h>
+#endif
+
+#include <string.h>
+#include <string>
+
+
+namespace testing {
+namespace internal {
+
+// String - an abstract class holding static string utilities.
+class GTEST_API_ String {
+ public:
+  // Static utility methods
+
+  // Clones a 0-terminated C string, allocating memory using new.  The
+  // caller is responsible for deleting the return value using
+  // delete[].  Returns the cloned string, or NULL if the input is
+  // NULL.
+  //
+  // This is different from strdup() in string.h, which allocates
+  // memory using malloc().
+  static const char* CloneCString(const char* c_str);
+
+#if GTEST_OS_WINDOWS_MOBILE
+  // Windows CE does not have the 'ANSI' versions of Win32 APIs. To be
+  // able to pass strings to Win32 APIs on CE we need to convert them
+  // to 'Unicode', UTF-16.
+
+  // Creates a UTF-16 wide string from the given ANSI string, allocating
+  // memory using new. The caller is responsible for deleting the return
+  // value using delete[]. Returns the wide string, or NULL if the
+  // input is NULL.
+  //
+  // The wide string is created using the ANSI codepage (CP_ACP) to
+  // match the behaviour of the ANSI versions of Win32 calls and the
+  // C runtime.
+  static LPCWSTR AnsiToUtf16(const char* c_str);
+
+  // Creates an ANSI string from the given wide string, allocating
+  // memory using new. The caller is responsible for deleting the return
+  // value using delete[]. Returns the ANSI string, or NULL if the
+  // input is NULL.
+  //
+  // The returned string is created using the ANSI codepage (CP_ACP) to
+  // match the behaviour of the ANSI versions of Win32 calls and the
+  // C runtime.
+  static const char* Utf16ToAnsi(LPCWSTR utf16_str);
+#endif
+
+  // Compares two C strings.  Returns true iff they have the same content.
+  //
+  // Unlike strcmp(), this function can handle NULL argument(s).  A
+  // NULL C string is considered different to any non-NULL C string,
+  // including the empty string.
+  static bool CStringEquals(const char* lhs, const char* rhs);
+
+  // Converts a wide C string to a String using the UTF-8 encoding.
+  // NULL will be converted to "(null)".  If an error occurred during
+  // the conversion, "(failed to convert from wide string)" is
+  // returned.
+  static std::string ShowWideCString(const wchar_t* wide_c_str);
+
+  // Compares two wide C strings.  Returns true iff they have the same
+  // content.
+  //
+  // Unlike wcscmp(), this function can handle NULL argument(s).  A
+  // NULL C string is considered different to any non-NULL C string,
+  // including the empty string.
+  static bool WideCStringEquals(const wchar_t* lhs, const wchar_t* rhs);
+
+  // Compares two C strings, ignoring case.  Returns true iff they
+  // have the same content.
+  //
+  // Unlike strcasecmp(), this function can handle NULL argument(s).
+  // A NULL C string is considered different to any non-NULL C string,
+  // including the empty string.
+  static bool CaseInsensitiveCStringEquals(const char* lhs,
+                                           const char* rhs);
+
+  // Compares two wide C strings, ignoring case.  Returns true iff they
+  // have the same content.
+  //
+  // Unlike wcscasecmp(), this function can handle NULL argument(s).
+  // A NULL C string is considered different to any non-NULL wide C string,
+  // including the empty string.
+  // NB: The implementations on different platforms slightly differ.
+  // On windows, this method uses _wcsicmp which compares according to LC_CTYPE
+  // environment variable. On GNU platform this method uses wcscasecmp
+  // which compares according to LC_CTYPE category of the current locale.
+  // On MacOS X, it uses towlower, which also uses LC_CTYPE category of the
+  // current locale.
+  static bool CaseInsensitiveWideCStringEquals(const wchar_t* lhs,
+                                               const wchar_t* rhs);
+
+  // Returns true iff the given string ends with the given suffix, ignoring
+  // case. Any string is considered to end with an empty suffix.
+  static bool EndsWithCaseInsensitive(
+      const std::string& str, const std::string& suffix);
+
+  // Formats an int value as "%02d".
+  static std::string FormatIntWidth2(int value);  // "%02d" for width == 2
+
+  // Formats an int value as "%X".
+  static std::string FormatHexInt(int value);
+
+  // Formats a byte as "%02X".
+  static std::string FormatByte(unsigned char value);
+
+ private:
+  String();  // Not meant to be instantiated.
+};  // class String
+
+// Gets the content of the stringstream's buffer as an std::string.  Each '\0'
+// character in the buffer is replaced with "\\0".
+GTEST_API_ std::string StringStreamToString(::std::stringstream* stream);
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_STRING_H_
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: keith.ray@gmail.com (Keith Ray)
+//
+// Google Test filepath utilities
+//
+// This header file declares classes and functions used internally by
+// Google Test.  They are subject to change without notice.
+//
+// This file is #included in <gtest/internal/gtest-internal.h>.
+// Do not include this header file separately!
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_FILEPATH_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_FILEPATH_H_
+
+
+namespace testing {
+namespace internal {
+
+// FilePath - a class for file and directory pathname manipulation which
+// handles platform-specific conventions (like the pathname separator).
+// Used for helper functions for naming files in a directory for xml output.
+// Except for Set methods, all methods are const or static, which provides an
+// "immutable value object" -- useful for peace of mind.
+// A FilePath with a value ending in a path separator ("like/this/") represents
+// a directory, otherwise it is assumed to represent a file. In either case,
+// it may or may not represent an actual file or directory in the file system.
+// Names are NOT checked for syntax correctness -- no checking for illegal
+// characters, malformed paths, etc.
+
+class GTEST_API_ FilePath {
+ public:
+  FilePath() : pathname_("") { }
+  FilePath(const FilePath& rhs) : pathname_(rhs.pathname_) { }
+
+  explicit FilePath(const std::string& pathname) : pathname_(pathname) {
+    Normalize();
+  }
+
+  FilePath& operator=(const FilePath& rhs) {
+    Set(rhs);
+    return *this;
+  }
+
+  void Set(const FilePath& rhs) {
+    pathname_ = rhs.pathname_;
+  }
+
+  const std::string& string() const { return pathname_; }
+  const char* c_str() const { return pathname_.c_str(); }
+
+  // Returns the current working directory, or "" if unsuccessful.
+  static FilePath GetCurrentDir();
+
+  // Given directory = "dir", base_name = "test", number = 0,
+  // extension = "xml", returns "dir/test.xml". If number is greater
+  // than zero (e.g., 12), returns "dir/test_12.xml".
+  // On Windows platform, uses \ as the separator rather than /.
+  static FilePath MakeFileName(const FilePath& directory,
+                               const FilePath& base_name,
+                               int number,
+                               const char* extension);
+
+  // Given directory = "dir", relative_path = "test.xml",
+  // returns "dir/test.xml".
+  // On Windows, uses \ as the separator rather than /.
+  static FilePath ConcatPaths(const FilePath& directory,
+                              const FilePath& relative_path);
+
+  // Returns a pathname for a file that does not currently exist. The pathname
+  // will be directory/base_name.extension or
+  // directory/base_name_<number>.extension if directory/base_name.extension
+  // already exists. The number will be incremented until a pathname is found
+  // that does not already exist.
+  // Examples: 'dir/foo_test.xml' or 'dir/foo_test_1.xml'.
+  // There could be a race condition if two or more processes are calling this
+  // function at the same time -- they could both pick the same filename.
+  static FilePath GenerateUniqueFileName(const FilePath& directory,
+                                         const FilePath& base_name,
+                                         const char* extension);
+
+  // Returns true iff the path is "".
+  bool IsEmpty() const { return pathname_.empty(); }
+
+  // If input name has a trailing separator character, removes it and returns
+  // the name, otherwise return the name string unmodified.
+  // On Windows platform, uses \ as the separator, other platforms use /.
+  FilePath RemoveTrailingPathSeparator() const;
+
+  // Returns a copy of the FilePath with the directory part removed.
+  // Example: FilePath("path/to/file").RemoveDirectoryName() returns
+  // FilePath("file"). If there is no directory part ("just_a_file"), it returns
+  // the FilePath unmodified. If there is no file part ("just_a_dir/") it
+  // returns an empty FilePath ("").
+  // On Windows platform, '\' is the path separator, otherwise it is '/'.
+  FilePath RemoveDirectoryName() const;
+
+  // RemoveFileName returns the directory path with the filename removed.
+  // Example: FilePath("path/to/file").RemoveFileName() returns "path/to/".
+  // If the FilePath is "a_file" or "/a_file", RemoveFileName returns
+  // FilePath("./") or, on Windows, FilePath(".\\"). If the filepath does
+  // not have a file, like "just/a/dir/", it returns the FilePath unmodified.
+  // On Windows platform, '\' is the path separator, otherwise it is '/'.
+  FilePath RemoveFileName() const;
+
+  // Returns a copy of the FilePath with the case-insensitive extension removed.
+  // Example: FilePath("dir/file.exe").RemoveExtension("EXE") returns
+  // FilePath("dir/file"). If a case-insensitive extension is not
+  // found, returns a copy of the original FilePath.
+  FilePath RemoveExtension(const char* extension) const;
+
+  // Creates directories so that path exists. Returns true if successful or if
+  // the directories already exist; returns false if unable to create
+  // directories for any reason. Will also return false if the FilePath does
+  // not represent a directory (that is, it doesn't end with a path separator).
+  bool CreateDirectoriesRecursively() const;
+
+  // Create the directory so that path exists. Returns true if successful or
+  // if the directory already exists; returns false if unable to create the
+  // directory for any reason, including if the parent directory does not
+  // exist. Not named "CreateDirectory" because that's a macro on Windows.
+  bool CreateFolder() const;
+
+  // Returns true if FilePath describes something in the file-system,
+  // either a file, directory, or whatever, and that something exists.
+  bool FileOrDirectoryExists() const;
+
+  // Returns true if pathname describes a directory in the file-system
+  // that exists.
+  bool DirectoryExists() const;
+
+  // Returns true if FilePath ends with a path separator, which indicates that
+  // it is intended to represent a directory. Returns false otherwise.
+  // This does NOT check that a directory (or file) actually exists.
+  bool IsDirectory() const;
+
+  // Returns true if pathname describes a root directory. (Windows has one
+  // root directory per disk drive.)
+  bool IsRootDirectory() const;
+
+  // Returns true if pathname describes an absolute path.
+  bool IsAbsolutePath() const;
+
+ private:
+  // Replaces multiple consecutive separators with a single separator.
+  // For example, "bar///foo" becomes "bar/foo". Does not eliminate other
+  // redundancies that might be in a pathname involving "." or "..".
+  //
+  // A pathname with multiple consecutive separators may occur either through
+  // user error or as a result of some scripts or APIs that generate a pathname
+  // with a trailing separator. On other platforms the same API or script
+  // may NOT generate a pathname with a trailing "/". Then elsewhere that
+  // pathname may have another "/" and pathname components added to it,
+  // without checking for the separator already being there.
+  // The script language and operating system may allow paths like "foo//bar"
+  // but some of the functions in FilePath will not handle that correctly. In
+  // particular, RemoveTrailingPathSeparator() only removes one separator, and
+  // it is called in CreateDirectoriesRecursively() assuming that it will change
+  // a pathname from directory syntax (trailing separator) to filename syntax.
+  //
+  // On Windows this method also replaces the alternate path separator '/' with
+  // the primary path separator '\\', so that for example "bar\\/\\foo" becomes
+  // "bar\\foo".
+
+  void Normalize();
+
+  // Returns a pointer to the last occurence of a valid path separator in
+  // the FilePath. On Windows, for example, both '/' and '\' are valid path
+  // separators. Returns NULL if no path separator was found.
+  const char* FindLastPathSeparator() const;
+
+  std::string pathname_;
+};  // class FilePath
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_FILEPATH_H_
+// This file was GENERATED by command:
+//     pump.py gtest-type-util.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Type utilities needed for implementing typed and type-parameterized
+// tests.  This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+// Currently we support at most 50 types in a list, and at most 50
+// type-parameterized tests in one type-parameterized test case.
+// Please contact googletestframework@googlegroups.com if you need
+// more.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
+
+
+// #ifdef __GNUC__ is too general here.  It is possible to use gcc without using
+// libstdc++ (which is where cxxabi.h comes from).
+# if GTEST_HAS_CXXABI_H_
+#  include <cxxabi.h>
+# elif defined(__HP_aCC)
+#  include <acxx_demangle.h>
+# endif  // GTEST_HASH_CXXABI_H_
+
+namespace testing {
+namespace internal {
+
+// GetTypeName<T>() returns a human-readable name of type T.
+// NB: This function is also used in Google Mock, so don't move it inside of
+// the typed-test-only section below.
+template <typename T>
+std::string GetTypeName() {
+# if GTEST_HAS_RTTI
+
+  const char* const name = typeid(T).name();
+#  if GTEST_HAS_CXXABI_H_ || defined(__HP_aCC)
+  int status = 0;
+  // gcc's implementation of typeid(T).name() mangles the type name,
+  // so we have to demangle it.
+#   if GTEST_HAS_CXXABI_H_
+  using abi::__cxa_demangle;
+#   endif  // GTEST_HAS_CXXABI_H_
+  char* const readable_name = __cxa_demangle(name, 0, 0, &status);
+  const std::string name_str(status == 0 ? readable_name : name);
+  free(readable_name);
+  return name_str;
+#  else
+  return name;
+#  endif  // GTEST_HAS_CXXABI_H_ || __HP_aCC
+
+# else
+
+  return "<type>";
+
+# endif  // GTEST_HAS_RTTI
+}
+
+#if GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+// AssertyTypeEq<T1, T2>::type is defined iff T1 and T2 are the same
+// type.  This can be used as a compile-time assertion to ensure that
+// two types are equal.
+
+template <typename T1, typename T2>
+struct AssertTypeEq;
+
+template <typename T>
+struct AssertTypeEq<T, T> {
+  typedef bool type;
+};
+
+// A unique type used as the default value for the arguments of class
+// template Types.  This allows us to simulate variadic templates
+// (e.g. Types<int>, Type<int, double>, and etc), which C++ doesn't
+// support directly.
+struct None {};
+
+// The following family of struct and struct templates are used to
+// represent type lists.  In particular, TypesN<T1, T2, ..., TN>
+// represents a type list with N types (T1, T2, ..., and TN) in it.
+// Except for Types0, every struct in the family has two member types:
+// Head for the first type in the list, and Tail for the rest of the
+// list.
+
+// The empty type list.
+struct Types0 {};
+
+// Type lists of length 1, 2, 3, and so on.
+
+template <typename T1>
+struct Types1 {
+  typedef T1 Head;
+  typedef Types0 Tail;
+};
+template <typename T1, typename T2>
+struct Types2 {
+  typedef T1 Head;
+  typedef Types1<T2> Tail;
+};
+
+template <typename T1, typename T2, typename T3>
+struct Types3 {
+  typedef T1 Head;
+  typedef Types2<T2, T3> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4>
+struct Types4 {
+  typedef T1 Head;
+  typedef Types3<T2, T3, T4> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+struct Types5 {
+  typedef T1 Head;
+  typedef Types4<T2, T3, T4, T5> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6>
+struct Types6 {
+  typedef T1 Head;
+  typedef Types5<T2, T3, T4, T5, T6> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7>
+struct Types7 {
+  typedef T1 Head;
+  typedef Types6<T2, T3, T4, T5, T6, T7> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8>
+struct Types8 {
+  typedef T1 Head;
+  typedef Types7<T2, T3, T4, T5, T6, T7, T8> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9>
+struct Types9 {
+  typedef T1 Head;
+  typedef Types8<T2, T3, T4, T5, T6, T7, T8, T9> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10>
+struct Types10 {
+  typedef T1 Head;
+  typedef Types9<T2, T3, T4, T5, T6, T7, T8, T9, T10> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11>
+struct Types11 {
+  typedef T1 Head;
+  typedef Types10<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12>
+struct Types12 {
+  typedef T1 Head;
+  typedef Types11<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13>
+struct Types13 {
+  typedef T1 Head;
+  typedef Types12<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14>
+struct Types14 {
+  typedef T1 Head;
+  typedef Types13<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15>
+struct Types15 {
+  typedef T1 Head;
+  typedef Types14<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16>
+struct Types16 {
+  typedef T1 Head;
+  typedef Types15<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17>
+struct Types17 {
+  typedef T1 Head;
+  typedef Types16<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18>
+struct Types18 {
+  typedef T1 Head;
+  typedef Types17<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19>
+struct Types19 {
+  typedef T1 Head;
+  typedef Types18<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20>
+struct Types20 {
+  typedef T1 Head;
+  typedef Types19<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21>
+struct Types21 {
+  typedef T1 Head;
+  typedef Types20<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22>
+struct Types22 {
+  typedef T1 Head;
+  typedef Types21<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23>
+struct Types23 {
+  typedef T1 Head;
+  typedef Types22<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24>
+struct Types24 {
+  typedef T1 Head;
+  typedef Types23<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25>
+struct Types25 {
+  typedef T1 Head;
+  typedef Types24<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26>
+struct Types26 {
+  typedef T1 Head;
+  typedef Types25<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27>
+struct Types27 {
+  typedef T1 Head;
+  typedef Types26<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28>
+struct Types28 {
+  typedef T1 Head;
+  typedef Types27<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29>
+struct Types29 {
+  typedef T1 Head;
+  typedef Types28<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30>
+struct Types30 {
+  typedef T1 Head;
+  typedef Types29<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31>
+struct Types31 {
+  typedef T1 Head;
+  typedef Types30<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32>
+struct Types32 {
+  typedef T1 Head;
+  typedef Types31<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33>
+struct Types33 {
+  typedef T1 Head;
+  typedef Types32<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34>
+struct Types34 {
+  typedef T1 Head;
+  typedef Types33<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35>
+struct Types35 {
+  typedef T1 Head;
+  typedef Types34<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36>
+struct Types36 {
+  typedef T1 Head;
+  typedef Types35<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37>
+struct Types37 {
+  typedef T1 Head;
+  typedef Types36<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38>
+struct Types38 {
+  typedef T1 Head;
+  typedef Types37<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39>
+struct Types39 {
+  typedef T1 Head;
+  typedef Types38<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40>
+struct Types40 {
+  typedef T1 Head;
+  typedef Types39<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41>
+struct Types41 {
+  typedef T1 Head;
+  typedef Types40<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42>
+struct Types42 {
+  typedef T1 Head;
+  typedef Types41<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43>
+struct Types43 {
+  typedef T1 Head;
+  typedef Types42<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44>
+struct Types44 {
+  typedef T1 Head;
+  typedef Types43<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45>
+struct Types45 {
+  typedef T1 Head;
+  typedef Types44<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46>
+struct Types46 {
+  typedef T1 Head;
+  typedef Types45<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45, T46> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47>
+struct Types47 {
+  typedef T1 Head;
+  typedef Types46<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45, T46, T47> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48>
+struct Types48 {
+  typedef T1 Head;
+  typedef Types47<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45, T46, T47, T48> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49>
+struct Types49 {
+  typedef T1 Head;
+  typedef Types48<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45, T46, T47, T48, T49> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49, typename T50>
+struct Types50 {
+  typedef T1 Head;
+  typedef Types49<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45, T46, T47, T48, T49, T50> Tail;
+};
+
+
+}  // namespace internal
+
+// We don't want to require the users to write TypesN<...> directly,
+// as that would require them to count the length.  Types<...> is much
+// easier to write, but generates horrible messages when there is a
+// compiler error, as gcc insists on printing out each template
+// argument, even if it has the default value (this means Types<int>
+// will appear as Types<int, None, None, ..., None> in the compiler
+// errors).
+//
+// Our solution is to combine the best part of the two approaches: a
+// user would write Types<T1, ..., TN>, and Google Test will translate
+// that to TypesN<T1, ..., TN> internally to make error messages
+// readable.  The translation is done by the 'type' member of the
+// Types template.
+template <typename T1 = internal::None, typename T2 = internal::None,
+    typename T3 = internal::None, typename T4 = internal::None,
+    typename T5 = internal::None, typename T6 = internal::None,
+    typename T7 = internal::None, typename T8 = internal::None,
+    typename T9 = internal::None, typename T10 = internal::None,
+    typename T11 = internal::None, typename T12 = internal::None,
+    typename T13 = internal::None, typename T14 = internal::None,
+    typename T15 = internal::None, typename T16 = internal::None,
+    typename T17 = internal::None, typename T18 = internal::None,
+    typename T19 = internal::None, typename T20 = internal::None,
+    typename T21 = internal::None, typename T22 = internal::None,
+    typename T23 = internal::None, typename T24 = internal::None,
+    typename T25 = internal::None, typename T26 = internal::None,
+    typename T27 = internal::None, typename T28 = internal::None,
+    typename T29 = internal::None, typename T30 = internal::None,
+    typename T31 = internal::None, typename T32 = internal::None,
+    typename T33 = internal::None, typename T34 = internal::None,
+    typename T35 = internal::None, typename T36 = internal::None,
+    typename T37 = internal::None, typename T38 = internal::None,
+    typename T39 = internal::None, typename T40 = internal::None,
+    typename T41 = internal::None, typename T42 = internal::None,
+    typename T43 = internal::None, typename T44 = internal::None,
+    typename T45 = internal::None, typename T46 = internal::None,
+    typename T47 = internal::None, typename T48 = internal::None,
+    typename T49 = internal::None, typename T50 = internal::None>
+struct Types {
+  typedef internal::Types50<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46, T47, T48, T49, T50> type;
+};
+
+template <>
+struct Types<internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types0 type;
+};
+template <typename T1>
+struct Types<T1, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types1<T1> type;
+};
+template <typename T1, typename T2>
+struct Types<T1, T2, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types2<T1, T2> type;
+};
+template <typename T1, typename T2, typename T3>
+struct Types<T1, T2, T3, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types3<T1, T2, T3> type;
+};
+template <typename T1, typename T2, typename T3, typename T4>
+struct Types<T1, T2, T3, T4, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types4<T1, T2, T3, T4> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+struct Types<T1, T2, T3, T4, T5, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types5<T1, T2, T3, T4, T5> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6>
+struct Types<T1, T2, T3, T4, T5, T6, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types6<T1, T2, T3, T4, T5, T6> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7>
+struct Types<T1, T2, T3, T4, T5, T6, T7, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types7<T1, T2, T3, T4, T5, T6, T7> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types8<T1, T2, T3, T4, T5, T6, T7, T8> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types9<T1, T2, T3, T4, T5, T6, T7, T8, T9> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types10<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types11<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types12<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types13<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types14<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types15<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types16<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types17<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types18<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types19<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types20<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types21<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types22<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types23<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types24<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types25<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types26<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types27<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types28<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types29<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types30<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types31<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types32<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types33<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types34<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types35<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types36<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types37<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types38<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types39<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types40<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types41<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types42<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types43<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types44<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44, T45,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types45<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44, T45,
+    T46, internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types46<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44, T45,
+    T46, T47, internal::None, internal::None, internal::None> {
+  typedef internal::Types47<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46, T47> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44, T45,
+    T46, T47, T48, internal::None, internal::None> {
+  typedef internal::Types48<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46, T47, T48> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44, T45,
+    T46, T47, T48, T49, internal::None> {
+  typedef internal::Types49<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46, T47, T48, T49> type;
+};
+
+namespace internal {
+
+# define GTEST_TEMPLATE_ template <typename T> class
+
+// The template "selector" struct TemplateSel<Tmpl> is used to
+// represent Tmpl, which must be a class template with one type
+// parameter, as a type.  TemplateSel<Tmpl>::Bind<T>::type is defined
+// as the type Tmpl<T>.  This allows us to actually instantiate the
+// template "selected" by TemplateSel<Tmpl>.
+//
+// This trick is necessary for simulating typedef for class templates,
+// which C++ doesn't support directly.
+template <GTEST_TEMPLATE_ Tmpl>
+struct TemplateSel {
+  template <typename T>
+  struct Bind {
+    typedef Tmpl<T> type;
+  };
+};
+
+# define GTEST_BIND_(TmplSel, T) \
+  TmplSel::template Bind<T>::type
+
+// A unique struct template used as the default value for the
+// arguments of class template Templates.  This allows us to simulate
+// variadic templates (e.g. Templates<int>, Templates<int, double>,
+// and etc), which C++ doesn't support directly.
+template <typename T>
+struct NoneT {};
+
+// The following family of struct and struct templates are used to
+// represent template lists.  In particular, TemplatesN<T1, T2, ...,
+// TN> represents a list of N templates (T1, T2, ..., and TN).  Except
+// for Templates0, every struct in the family has two member types:
+// Head for the selector of the first template in the list, and Tail
+// for the rest of the list.
+
+// The empty template list.
+struct Templates0 {};
+
+// Template lists of length 1, 2, 3, and so on.
+
+template <GTEST_TEMPLATE_ T1>
+struct Templates1 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates0 Tail;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2>
+struct Templates2 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates1<T2> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3>
+struct Templates3 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates2<T2, T3> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4>
+struct Templates4 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates3<T2, T3, T4> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5>
+struct Templates5 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates4<T2, T3, T4, T5> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6>
+struct Templates6 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates5<T2, T3, T4, T5, T6> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7>
+struct Templates7 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates6<T2, T3, T4, T5, T6, T7> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8>
+struct Templates8 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates7<T2, T3, T4, T5, T6, T7, T8> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9>
+struct Templates9 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates8<T2, T3, T4, T5, T6, T7, T8, T9> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10>
+struct Templates10 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates9<T2, T3, T4, T5, T6, T7, T8, T9, T10> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11>
+struct Templates11 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates10<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12>
+struct Templates12 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates11<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13>
+struct Templates13 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates12<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14>
+struct Templates14 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates13<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15>
+struct Templates15 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates14<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16>
+struct Templates16 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates15<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17>
+struct Templates17 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates16<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18>
+struct Templates18 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates17<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19>
+struct Templates19 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates18<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20>
+struct Templates20 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates19<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21>
+struct Templates21 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates20<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22>
+struct Templates22 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates21<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23>
+struct Templates23 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates22<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24>
+struct Templates24 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates23<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25>
+struct Templates25 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates24<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26>
+struct Templates26 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates25<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27>
+struct Templates27 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates26<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28>
+struct Templates28 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates27<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29>
+struct Templates29 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates28<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30>
+struct Templates30 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates29<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31>
+struct Templates31 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates30<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32>
+struct Templates32 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates31<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33>
+struct Templates33 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates32<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34>
+struct Templates34 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates33<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35>
+struct Templates35 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates34<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36>
+struct Templates36 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates35<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37>
+struct Templates37 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates36<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38>
+struct Templates38 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates37<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39>
+struct Templates39 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates38<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40>
+struct Templates40 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates39<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41>
+struct Templates41 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates40<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42>
+struct Templates42 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates41<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43>
+struct Templates43 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates42<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44>
+struct Templates44 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates43<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45>
+struct Templates45 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates44<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46>
+struct Templates46 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates45<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45, T46> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47>
+struct Templates47 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates46<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45, T46, T47> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47, GTEST_TEMPLATE_ T48>
+struct Templates48 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates47<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45, T46, T47, T48> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47, GTEST_TEMPLATE_ T48,
+    GTEST_TEMPLATE_ T49>
+struct Templates49 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates48<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45, T46, T47, T48, T49> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47, GTEST_TEMPLATE_ T48,
+    GTEST_TEMPLATE_ T49, GTEST_TEMPLATE_ T50>
+struct Templates50 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates49<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45, T46, T47, T48, T49, T50> Tail;
+};
+
+
+// We don't want to require the users to write TemplatesN<...> directly,
+// as that would require them to count the length.  Templates<...> is much
+// easier to write, but generates horrible messages when there is a
+// compiler error, as gcc insists on printing out each template
+// argument, even if it has the default value (this means Templates<list>
+// will appear as Templates<list, NoneT, NoneT, ..., NoneT> in the compiler
+// errors).
+//
+// Our solution is to combine the best part of the two approaches: a
+// user would write Templates<T1, ..., TN>, and Google Test will translate
+// that to TemplatesN<T1, ..., TN> internally to make error messages
+// readable.  The translation is done by the 'type' member of the
+// Templates template.
+template <GTEST_TEMPLATE_ T1 = NoneT, GTEST_TEMPLATE_ T2 = NoneT,
+    GTEST_TEMPLATE_ T3 = NoneT, GTEST_TEMPLATE_ T4 = NoneT,
+    GTEST_TEMPLATE_ T5 = NoneT, GTEST_TEMPLATE_ T6 = NoneT,
+    GTEST_TEMPLATE_ T7 = NoneT, GTEST_TEMPLATE_ T8 = NoneT,
+    GTEST_TEMPLATE_ T9 = NoneT, GTEST_TEMPLATE_ T10 = NoneT,
+    GTEST_TEMPLATE_ T11 = NoneT, GTEST_TEMPLATE_ T12 = NoneT,
+    GTEST_TEMPLATE_ T13 = NoneT, GTEST_TEMPLATE_ T14 = NoneT,
+    GTEST_TEMPLATE_ T15 = NoneT, GTEST_TEMPLATE_ T16 = NoneT,
+    GTEST_TEMPLATE_ T17 = NoneT, GTEST_TEMPLATE_ T18 = NoneT,
+    GTEST_TEMPLATE_ T19 = NoneT, GTEST_TEMPLATE_ T20 = NoneT,
+    GTEST_TEMPLATE_ T21 = NoneT, GTEST_TEMPLATE_ T22 = NoneT,
+    GTEST_TEMPLATE_ T23 = NoneT, GTEST_TEMPLATE_ T24 = NoneT,
+    GTEST_TEMPLATE_ T25 = NoneT, GTEST_TEMPLATE_ T26 = NoneT,
+    GTEST_TEMPLATE_ T27 = NoneT, GTEST_TEMPLATE_ T28 = NoneT,
+    GTEST_TEMPLATE_ T29 = NoneT, GTEST_TEMPLATE_ T30 = NoneT,
+    GTEST_TEMPLATE_ T31 = NoneT, GTEST_TEMPLATE_ T32 = NoneT,
+    GTEST_TEMPLATE_ T33 = NoneT, GTEST_TEMPLATE_ T34 = NoneT,
+    GTEST_TEMPLATE_ T35 = NoneT, GTEST_TEMPLATE_ T36 = NoneT,
+    GTEST_TEMPLATE_ T37 = NoneT, GTEST_TEMPLATE_ T38 = NoneT,
+    GTEST_TEMPLATE_ T39 = NoneT, GTEST_TEMPLATE_ T40 = NoneT,
+    GTEST_TEMPLATE_ T41 = NoneT, GTEST_TEMPLATE_ T42 = NoneT,
+    GTEST_TEMPLATE_ T43 = NoneT, GTEST_TEMPLATE_ T44 = NoneT,
+    GTEST_TEMPLATE_ T45 = NoneT, GTEST_TEMPLATE_ T46 = NoneT,
+    GTEST_TEMPLATE_ T47 = NoneT, GTEST_TEMPLATE_ T48 = NoneT,
+    GTEST_TEMPLATE_ T49 = NoneT, GTEST_TEMPLATE_ T50 = NoneT>
+struct Templates {
+  typedef Templates50<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45, T46, T47, T48, T49, T50> type;
+};
+
+template <>
+struct Templates<NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT> {
+  typedef Templates0 type;
+};
+template <GTEST_TEMPLATE_ T1>
+struct Templates<T1, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT> {
+  typedef Templates1<T1> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2>
+struct Templates<T1, T2, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT> {
+  typedef Templates2<T1, T2> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3>
+struct Templates<T1, T2, T3, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates3<T1, T2, T3> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4>
+struct Templates<T1, T2, T3, T4, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates4<T1, T2, T3, T4> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5>
+struct Templates<T1, T2, T3, T4, T5, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates5<T1, T2, T3, T4, T5> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6>
+struct Templates<T1, T2, T3, T4, T5, T6, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates6<T1, T2, T3, T4, T5, T6> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates7<T1, T2, T3, T4, T5, T6, T7> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates8<T1, T2, T3, T4, T5, T6, T7, T8> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates9<T1, T2, T3, T4, T5, T6, T7, T8, T9> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates10<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates11<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates12<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates13<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates14<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates15<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates16<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates17<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates18<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates19<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates20<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates21<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT> {
+  typedef Templates22<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT> {
+  typedef Templates23<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT> {
+  typedef Templates24<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT> {
+  typedef Templates25<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT> {
+  typedef Templates26<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT> {
+  typedef Templates27<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT> {
+  typedef Templates28<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT> {
+  typedef Templates29<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates30<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates31<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates32<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates33<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates34<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates35<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates36<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates37<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates38<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates39<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates40<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates41<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates42<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates43<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates44<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    T45, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates45<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    T45, T46, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates46<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45, T46> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    T45, T46, T47, NoneT, NoneT, NoneT> {
+  typedef Templates47<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45, T46, T47> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47, GTEST_TEMPLATE_ T48>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    T45, T46, T47, T48, NoneT, NoneT> {
+  typedef Templates48<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45, T46, T47, T48> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47, GTEST_TEMPLATE_ T48,
+    GTEST_TEMPLATE_ T49>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    T45, T46, T47, T48, T49, NoneT> {
+  typedef Templates49<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45, T46, T47, T48, T49> type;
+};
+
+// The TypeList template makes it possible to use either a single type
+// or a Types<...> list in TYPED_TEST_CASE() and
+// INSTANTIATE_TYPED_TEST_CASE_P().
+
+template <typename T>
+struct TypeList {
+  typedef Types1<T> type;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49, typename T50>
+struct TypeList<Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46, T47, T48, T49, T50> > {
+  typedef typename Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46, T47, T48, T49, T50>::type type;
+};
+
+#endif  // GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
+
+// Due to C++ preprocessor weirdness, we need double indirection to
+// concatenate two tokens when one of them is __LINE__.  Writing
+//
+//   foo ## __LINE__
+//
+// will result in the token foo__LINE__, instead of foo followed by
+// the current line number.  For more details, see
+// http://www.parashift.com/c++-faq-lite/misc-technical-issues.html#faq-39.6
+#define GTEST_CONCAT_TOKEN_(foo, bar) GTEST_CONCAT_TOKEN_IMPL_(foo, bar)
+#define GTEST_CONCAT_TOKEN_IMPL_(foo, bar) foo ## bar
+
+class ProtocolMessage;
+namespace proto2 { class Message; }
+
+namespace testing {
+
+// Forward declarations.
+
+class AssertionResult;                 // Result of an assertion.
+class Message;                         // Represents a failure message.
+class Test;                            // Represents a test.
+class TestInfo;                        // Information about a test.
+class TestPartResult;                  // Result of a test part.
+class UnitTest;                        // A collection of test cases.
+
+template <typename T>
+::std::string PrintToString(const T& value);
+
+namespace internal {
+
+struct TraceInfo;                      // Information about a trace point.
+class ScopedTrace;                     // Implements scoped trace.
+class TestInfoImpl;                    // Opaque implementation of TestInfo
+class UnitTestImpl;                    // Opaque implementation of UnitTest
+
+// How many times InitGoogleTest() has been called.
+GTEST_API_ extern int g_init_gtest_count;
+
+// The text used in failure messages to indicate the start of the
+// stack trace.
+GTEST_API_ extern const char kStackTraceMarker[];
+
+// Two overloaded helpers for checking at compile time whether an
+// expression is a null pointer literal (i.e. NULL or any 0-valued
+// compile-time integral constant).  Their return values have
+// different sizes, so we can use sizeof() to test which version is
+// picked by the compiler.  These helpers have no implementations, as
+// we only need their signatures.
+//
+// Given IsNullLiteralHelper(x), the compiler will pick the first
+// version if x can be implicitly converted to Secret*, and pick the
+// second version otherwise.  Since Secret is a secret and incomplete
+// type, the only expression a user can write that has type Secret* is
+// a null pointer literal.  Therefore, we know that x is a null
+// pointer literal if and only if the first version is picked by the
+// compiler.
+char IsNullLiteralHelper(Secret* p);
+char (&IsNullLiteralHelper(...))[2];  // NOLINT
+
+// A compile-time bool constant that is true if and only if x is a
+// null pointer literal (i.e. NULL or any 0-valued compile-time
+// integral constant).
+#ifdef GTEST_ELLIPSIS_NEEDS_POD_
+// We lose support for NULL detection where the compiler doesn't like
+// passing non-POD classes through ellipsis (...).
+# define GTEST_IS_NULL_LITERAL_(x) false
+#else
+# define GTEST_IS_NULL_LITERAL_(x) \
+    (sizeof(::testing::internal::IsNullLiteralHelper(x)) == 1)
+#endif  // GTEST_ELLIPSIS_NEEDS_POD_
+
+// Appends the user-supplied message to the Google-Test-generated message.
+GTEST_API_ std::string AppendUserMessage(
+    const std::string& gtest_msg, const Message& user_msg);
+
+#if GTEST_HAS_EXCEPTIONS
+
+// This exception is thrown by (and only by) a failed Google Test
+// assertion when GTEST_FLAG(throw_on_failure) is true (if exceptions
+// are enabled).  We derive it from std::runtime_error, which is for
+// errors presumably detectable only at run time.  Since
+// std::runtime_error inherits from std::exception, many testing
+// frameworks know how to extract and print the message inside it.
+class GTEST_API_ GoogleTestFailureException : public ::std::runtime_error {
+ public:
+  explicit GoogleTestFailureException(const TestPartResult& failure);
+};
+
+#endif  // GTEST_HAS_EXCEPTIONS
+
+// A helper class for creating scoped traces in user programs.
+class GTEST_API_ ScopedTrace {
+ public:
+  // The c'tor pushes the given source file location and message onto
+  // a trace stack maintained by Google Test.
+  ScopedTrace(const char* file, int line, const Message& message);
+
+  // The d'tor pops the info pushed by the c'tor.
+  //
+  // Note that the d'tor is not virtual in order to be efficient.
+  // Don't inherit from ScopedTrace!
+  ~ScopedTrace();
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ScopedTrace);
+} GTEST_ATTRIBUTE_UNUSED_;  // A ScopedTrace object does its job in its
+                            // c'tor and d'tor.  Therefore it doesn't
+                            // need to be used otherwise.
+
+// Constructs and returns the message for an equality assertion
+// (e.g. ASSERT_EQ, EXPECT_STREQ, etc) failure.
+//
+// The first four parameters are the expressions used in the assertion
+// and their values, as strings.  For example, for ASSERT_EQ(foo, bar)
+// where foo is 5 and bar is 6, we have:
+//
+//   expected_expression: "foo"
+//   actual_expression:   "bar"
+//   expected_value:      "5"
+//   actual_value:        "6"
+//
+// The ignoring_case parameter is true iff the assertion is a
+// *_STRCASEEQ*.  When it's true, the string " (ignoring case)" will
+// be inserted into the message.
+GTEST_API_ AssertionResult EqFailure(const char* expected_expression,
+                                     const char* actual_expression,
+                                     const std::string& expected_value,
+                                     const std::string& actual_value,
+                                     bool ignoring_case);
+
+// Constructs a failure message for Boolean assertions such as EXPECT_TRUE.
+GTEST_API_ std::string GetBoolAssertionFailureMessage(
+    const AssertionResult& assertion_result,
+    const char* expression_text,
+    const char* actual_predicate_value,
+    const char* expected_predicate_value);
+
+// This template class represents an IEEE floating-point number
+// (either single-precision or double-precision, depending on the
+// template parameters).
+//
+// The purpose of this class is to do more sophisticated number
+// comparison.  (Due to round-off error, etc, it's very unlikely that
+// two floating-points will be equal exactly.  Hence a naive
+// comparison by the == operation often doesn't work.)
+//
+// Format of IEEE floating-point:
+//
+//   The most-significant bit being the leftmost, an IEEE
+//   floating-point looks like
+//
+//     sign_bit exponent_bits fraction_bits
+//
+//   Here, sign_bit is a single bit that designates the sign of the
+//   number.
+//
+//   For float, there are 8 exponent bits and 23 fraction bits.
+//
+//   For double, there are 11 exponent bits and 52 fraction bits.
+//
+//   More details can be found at
+//   http://en.wikipedia.org/wiki/IEEE_floating-point_standard.
+//
+// Template parameter:
+//
+//   RawType: the raw floating-point type (either float or double)
+template <typename RawType>
+class FloatingPoint {
+ public:
+  // Defines the unsigned integer type that has the same size as the
+  // floating point number.
+  typedef typename TypeWithSize<sizeof(RawType)>::UInt Bits;
+
+  // Constants.
+
+  // # of bits in a number.
+  static const size_t kBitCount = 8*sizeof(RawType);
+
+  // # of fraction bits in a number.
+  static const size_t kFractionBitCount =
+    std::numeric_limits<RawType>::digits - 1;
+
+  // # of exponent bits in a number.
+  static const size_t kExponentBitCount = kBitCount - 1 - kFractionBitCount;
+
+  // The mask for the sign bit.
+  static const Bits kSignBitMask = static_cast<Bits>(1) << (kBitCount - 1);
+
+  // The mask for the fraction bits.
+  static const Bits kFractionBitMask =
+    ~static_cast<Bits>(0) >> (kExponentBitCount + 1);
+
+  // The mask for the exponent bits.
+  static const Bits kExponentBitMask = ~(kSignBitMask | kFractionBitMask);
+
+  // How many ULP's (Units in the Last Place) we want to tolerate when
+  // comparing two numbers.  The larger the value, the more error we
+  // allow.  A 0 value means that two numbers must be exactly the same
+  // to be considered equal.
+  //
+  // The maximum error of a single floating-point operation is 0.5
+  // units in the last place.  On Intel CPU's, all floating-point
+  // calculations are done with 80-bit precision, while double has 64
+  // bits.  Therefore, 4 should be enough for ordinary use.
+  //
+  // See the following article for more details on ULP:
+  // http://randomascii.wordpress.com/2012/02/25/comparing-floating-point-numbers-2012-edition/
+  static const size_t kMaxUlps = 4;
+
+  // Constructs a FloatingPoint from a raw floating-point number.
+  //
+  // On an Intel CPU, passing a non-normalized NAN (Not a Number)
+  // around may change its bits, although the new value is guaranteed
+  // to be also a NAN.  Therefore, don't expect this constructor to
+  // preserve the bits in x when x is a NAN.
+  explicit FloatingPoint(const RawType& x) { u_.value_ = x; }
+
+  // Static methods
+
+  // Reinterprets a bit pattern as a floating-point number.
+  //
+  // This function is needed to test the AlmostEquals() method.
+  static RawType ReinterpretBits(const Bits bits) {
+    FloatingPoint fp(0);
+    fp.u_.bits_ = bits;
+    return fp.u_.value_;
+  }
+
+  // Returns the floating-point number that represent positive infinity.
+  static RawType Infinity() {
+    return ReinterpretBits(kExponentBitMask);
+  }
+
+  // Returns the maximum representable finite floating-point number.
+  static RawType Max();
+
+  // Non-static methods
+
+  // Returns the bits that represents this number.
+  const Bits &bits() const { return u_.bits_; }
+
+  // Returns the exponent bits of this number.
+  Bits exponent_bits() const { return kExponentBitMask & u_.bits_; }
+
+  // Returns the fraction bits of this number.
+  Bits fraction_bits() const { return kFractionBitMask & u_.bits_; }
+
+  // Returns the sign bit of this number.
+  Bits sign_bit() const { return kSignBitMask & u_.bits_; }
+
+  // Returns true iff this is NAN (not a number).
+  bool is_nan() const {
+    // It's a NAN if the exponent bits are all ones and the fraction
+    // bits are not entirely zeros.
+    return (exponent_bits() == kExponentBitMask) && (fraction_bits() != 0);
+  }
+
+  // Returns true iff this number is at most kMaxUlps ULP's away from
+  // rhs.  In particular, this function:
+  //
+  //   - returns false if either number is (or both are) NAN.
+  //   - treats really large numbers as almost equal to infinity.
+  //   - thinks +0.0 and -0.0 are 0 DLP's apart.
+  bool AlmostEquals(const FloatingPoint& rhs) const {
+    // The IEEE standard says that any comparison operation involving
+    // a NAN must return false.
+    if (is_nan() || rhs.is_nan()) return false;
+
+    return DistanceBetweenSignAndMagnitudeNumbers(u_.bits_, rhs.u_.bits_)
+        <= kMaxUlps;
+  }
+
+ private:
+  // The data type used to store the actual floating-point number.
+  union FloatingPointUnion {
+    RawType value_;  // The raw floating-point number.
+    Bits bits_;      // The bits that represent the number.
+  };
+
+  // Converts an integer from the sign-and-magnitude representation to
+  // the biased representation.  More precisely, let N be 2 to the
+  // power of (kBitCount - 1), an integer x is represented by the
+  // unsigned number x + N.
+  //
+  // For instance,
+  //
+  //   -N + 1 (the most negative number representable using
+  //          sign-and-magnitude) is represented by 1;
+  //   0      is represented by N; and
+  //   N - 1  (the biggest number representable using
+  //          sign-and-magnitude) is represented by 2N - 1.
+  //
+  // Read http://en.wikipedia.org/wiki/Signed_number_representations
+  // for more details on signed number representations.
+  static Bits SignAndMagnitudeToBiased(const Bits &sam) {
+    if (kSignBitMask & sam) {
+      // sam represents a negative number.
+      return ~sam + 1;
+    } else {
+      // sam represents a positive number.
+      return kSignBitMask | sam;
+    }
+  }
+
+  // Given two numbers in the sign-and-magnitude representation,
+  // returns the distance between them as an unsigned number.
+  static Bits DistanceBetweenSignAndMagnitudeNumbers(const Bits &sam1,
+                                                     const Bits &sam2) {
+    const Bits biased1 = SignAndMagnitudeToBiased(sam1);
+    const Bits biased2 = SignAndMagnitudeToBiased(sam2);
+    return (biased1 >= biased2) ? (biased1 - biased2) : (biased2 - biased1);
+  }
+
+  FloatingPointUnion u_;
+};
+
+// We cannot use std::numeric_limits<T>::max() as it clashes with the max()
+// macro defined by <windows.h>.
+template <>
+inline float FloatingPoint<float>::Max() { return FLT_MAX; }
+template <>
+inline double FloatingPoint<double>::Max() { return DBL_MAX; }
+
+// Typedefs the instances of the FloatingPoint template class that we
+// care to use.
+typedef FloatingPoint<float> Float;
+typedef FloatingPoint<double> Double;
+
+// In order to catch the mistake of putting tests that use different
+// test fixture classes in the same test case, we need to assign
+// unique IDs to fixture classes and compare them.  The TypeId type is
+// used to hold such IDs.  The user should treat TypeId as an opaque
+// type: the only operation allowed on TypeId values is to compare
+// them for equality using the == operator.
+typedef const void* TypeId;
+
+template <typename T>
+class TypeIdHelper {
+ public:
+  // dummy_ must not have a const type.  Otherwise an overly eager
+  // compiler (e.g. MSVC 7.1 & 8.0) may try to merge
+  // TypeIdHelper<T>::dummy_ for different Ts as an "optimization".
+  static bool dummy_;
+};
+
+template <typename T>
+bool TypeIdHelper<T>::dummy_ = false;
+
+// GetTypeId<T>() returns the ID of type T.  Different values will be
+// returned for different types.  Calling the function twice with the
+// same type argument is guaranteed to return the same ID.
+template <typename T>
+TypeId GetTypeId() {
+  // The compiler is required to allocate a different
+  // TypeIdHelper<T>::dummy_ variable for each T used to instantiate
+  // the template.  Therefore, the address of dummy_ is guaranteed to
+  // be unique.
+  return &(TypeIdHelper<T>::dummy_);
+}
+
+// Returns the type ID of ::testing::Test.  Always call this instead
+// of GetTypeId< ::testing::Test>() to get the type ID of
+// ::testing::Test, as the latter may give the wrong result due to a
+// suspected linker bug when compiling Google Test as a Mac OS X
+// framework.
+GTEST_API_ TypeId GetTestTypeId();
+
+// Defines the abstract factory interface that creates instances
+// of a Test object.
+class TestFactoryBase {
+ public:
+  virtual ~TestFactoryBase() {}
+
+  // Creates a test instance to run. The instance is both created and destroyed
+  // within TestInfoImpl::Run()
+  virtual Test* CreateTest() = 0;
+
+ protected:
+  TestFactoryBase() {}
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestFactoryBase);
+};
+
+// This class provides implementation of TeastFactoryBase interface.
+// It is used in TEST and TEST_F macros.
+template <class TestClass>
+class TestFactoryImpl : public TestFactoryBase {
+ public:
+  virtual Test* CreateTest() { return new TestClass; }
+};
+
+#if GTEST_OS_WINDOWS
+
+// Predicate-formatters for implementing the HRESULT checking macros
+// {ASSERT|EXPECT}_HRESULT_{SUCCEEDED|FAILED}
+// We pass a long instead of HRESULT to avoid causing an
+// include dependency for the HRESULT type.
+GTEST_API_ AssertionResult IsHRESULTSuccess(const char* expr,
+                                            long hr);  // NOLINT
+GTEST_API_ AssertionResult IsHRESULTFailure(const char* expr,
+                                            long hr);  // NOLINT
+
+#endif  // GTEST_OS_WINDOWS
+
+// Types of SetUpTestCase() and TearDownTestCase() functions.
+typedef void (*SetUpTestCaseFunc)();
+typedef void (*TearDownTestCaseFunc)();
+
+// Creates a new TestInfo object and registers it with Google Test;
+// returns the created object.
+//
+// Arguments:
+//
+//   test_case_name:   name of the test case
+//   name:             name of the test
+//   type_param        the name of the test's type parameter, or NULL if
+//                     this is not a typed or a type-parameterized test.
+//   value_param       text representation of the test's value parameter,
+//                     or NULL if this is not a type-parameterized test.
+//   fixture_class_id: ID of the test fixture class
+//   set_up_tc:        pointer to the function that sets up the test case
+//   tear_down_tc:     pointer to the function that tears down the test case
+//   factory:          pointer to the factory that creates a test object.
+//                     The newly created TestInfo instance will assume
+//                     ownership of the factory object.
+GTEST_API_ TestInfo* MakeAndRegisterTestInfo(
+    const char* test_case_name,
+    const char* name,
+    const char* type_param,
+    const char* value_param,
+    TypeId fixture_class_id,
+    SetUpTestCaseFunc set_up_tc,
+    TearDownTestCaseFunc tear_down_tc,
+    TestFactoryBase* factory);
+
+// If *pstr starts with the given prefix, modifies *pstr to be right
+// past the prefix and returns true; otherwise leaves *pstr unchanged
+// and returns false.  None of pstr, *pstr, and prefix can be NULL.
+GTEST_API_ bool SkipPrefix(const char* prefix, const char** pstr);
+
+#if GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+// State of the definition of a type-parameterized test case.
+class GTEST_API_ TypedTestCasePState {
+ public:
+  TypedTestCasePState() : registered_(false) {}
+
+  // Adds the given test name to defined_test_names_ and return true
+  // if the test case hasn't been registered; otherwise aborts the
+  // program.
+  bool AddTestName(const char* file, int line, const char* case_name,
+                   const char* test_name) {
+    if (registered_) {
+      fprintf(stderr, "%s Test %s must be defined before "
+              "REGISTER_TYPED_TEST_CASE_P(%s, ...).\n",
+              FormatFileLocation(file, line).c_str(), test_name, case_name);
+      fflush(stderr);
+      posix::Abort();
+    }
+    defined_test_names_.insert(test_name);
+    return true;
+  }
+
+  // Verifies that registered_tests match the test names in
+  // defined_test_names_; returns registered_tests if successful, or
+  // aborts the program otherwise.
+  const char* VerifyRegisteredTestNames(
+      const char* file, int line, const char* registered_tests);
+
+ private:
+  bool registered_;
+  ::std::set<const char*> defined_test_names_;
+};
+
+// Skips to the first non-space char after the first comma in 'str';
+// returns NULL if no comma is found in 'str'.
+inline const char* SkipComma(const char* str) {
+  const char* comma = strchr(str, ',');
+  if (comma == NULL) {
+    return NULL;
+  }
+  while (IsSpace(*(++comma))) {}
+  return comma;
+}
+
+// Returns the prefix of 'str' before the first comma in it; returns
+// the entire string if it contains no comma.
+inline std::string GetPrefixUntilComma(const char* str) {
+  const char* comma = strchr(str, ',');
+  return comma == NULL ? str : std::string(str, comma);
+}
+
+// TypeParameterizedTest<Fixture, TestSel, Types>::Register()
+// registers a list of type-parameterized tests with Google Test.  The
+// return value is insignificant - we just need to return something
+// such that we can call this function in a namespace scope.
+//
+// Implementation note: The GTEST_TEMPLATE_ macro declares a template
+// template parameter.  It's defined in gtest-type-util.h.
+template <GTEST_TEMPLATE_ Fixture, class TestSel, typename Types>
+class TypeParameterizedTest {
+ public:
+  // 'index' is the index of the test in the type list 'Types'
+  // specified in INSTANTIATE_TYPED_TEST_CASE_P(Prefix, TestCase,
+  // Types).  Valid values for 'index' are [0, N - 1] where N is the
+  // length of Types.
+  static bool Register(const char* prefix, const char* case_name,
+                       const char* test_names, int index) {
+    typedef typename Types::Head Type;
+    typedef Fixture<Type> FixtureClass;
+    typedef typename GTEST_BIND_(TestSel, Type) TestClass;
+
+    // First, registers the first type-parameterized test in the type
+    // list.
+    MakeAndRegisterTestInfo(
+        (std::string(prefix) + (prefix[0] == '\0' ? "" : "/") + case_name + "/"
+         + StreamableToString(index)).c_str(),
+        GetPrefixUntilComma(test_names).c_str(),
+        GetTypeName<Type>().c_str(),
+        NULL,  // No value parameter.
+        GetTypeId<FixtureClass>(),
+        TestClass::SetUpTestCase,
+        TestClass::TearDownTestCase,
+        new TestFactoryImpl<TestClass>);
+
+    // Next, recurses (at compile time) with the tail of the type list.
+    return TypeParameterizedTest<Fixture, TestSel, typename Types::Tail>
+        ::Register(prefix, case_name, test_names, index + 1);
+  }
+};
+
+// The base case for the compile time recursion.
+template <GTEST_TEMPLATE_ Fixture, class TestSel>
+class TypeParameterizedTest<Fixture, TestSel, Types0> {
+ public:
+  static bool Register(const char* /*prefix*/, const char* /*case_name*/,
+                       const char* /*test_names*/, int /*index*/) {
+    return true;
+  }
+};
+
+// TypeParameterizedTestCase<Fixture, Tests, Types>::Register()
+// registers *all combinations* of 'Tests' and 'Types' with Google
+// Test.  The return value is insignificant - we just need to return
+// something such that we can call this function in a namespace scope.
+template <GTEST_TEMPLATE_ Fixture, typename Tests, typename Types>
+class TypeParameterizedTestCase {
+ public:
+  static bool Register(const char* prefix, const char* case_name,
+                       const char* test_names) {
+    typedef typename Tests::Head Head;
+
+    // First, register the first test in 'Test' for each type in 'Types'.
+    TypeParameterizedTest<Fixture, Head, Types>::Register(
+        prefix, case_name, test_names, 0);
+
+    // Next, recurses (at compile time) with the tail of the test list.
+    return TypeParameterizedTestCase<Fixture, typename Tests::Tail, Types>
+        ::Register(prefix, case_name, SkipComma(test_names));
+  }
+};
+
+// The base case for the compile time recursion.
+template <GTEST_TEMPLATE_ Fixture, typename Types>
+class TypeParameterizedTestCase<Fixture, Templates0, Types> {
+ public:
+  static bool Register(const char* /*prefix*/, const char* /*case_name*/,
+                       const char* /*test_names*/) {
+    return true;
+  }
+};
+
+#endif  // GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+// Returns the current OS stack trace as an std::string.
+//
+// The maximum number of stack frames to be included is specified by
+// the gtest_stack_trace_depth flag.  The skip_count parameter
+// specifies the number of top frames to be skipped, which doesn't
+// count against the number of frames to be included.
+//
+// For example, if Foo() calls Bar(), which in turn calls
+// GetCurrentOsStackTraceExceptTop(..., 1), Foo() will be included in
+// the trace but Bar() and GetCurrentOsStackTraceExceptTop() won't.
+GTEST_API_ std::string GetCurrentOsStackTraceExceptTop(
+    UnitTest* unit_test, int skip_count);
+
+// Helpers for suppressing warnings on unreachable code or constant
+// condition.
+
+// Always returns true.
+GTEST_API_ bool AlwaysTrue();
+
+// Always returns false.
+inline bool AlwaysFalse() { return !AlwaysTrue(); }
+
+// Helper for suppressing false warning from Clang on a const char*
+// variable declared in a conditional expression always being NULL in
+// the else branch.
+struct GTEST_API_ ConstCharPtr {
+  ConstCharPtr(const char* str) : value(str) {}
+  operator bool() const { return true; }
+  const char* value;
+};
+
+// A simple Linear Congruential Generator for generating random
+// numbers with a uniform distribution.  Unlike rand() and srand(), it
+// doesn't use global state (and therefore can't interfere with user
+// code).  Unlike rand_r(), it's portable.  An LCG isn't very random,
+// but it's good enough for our purposes.
+class GTEST_API_ Random {
+ public:
+  static const UInt32 kMaxRange = 1u << 31;
+
+  explicit Random(UInt32 seed) : state_(seed) {}
+
+  void Reseed(UInt32 seed) { state_ = seed; }
+
+  // Generates a random number from [0, range).  Crashes if 'range' is
+  // 0 or greater than kMaxRange.
+  UInt32 Generate(UInt32 range);
+
+ private:
+  UInt32 state_;
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Random);
+};
+
+// Defining a variable of type CompileAssertTypesEqual<T1, T2> will cause a
+// compiler error iff T1 and T2 are different types.
+template <typename T1, typename T2>
+struct CompileAssertTypesEqual;
+
+template <typename T>
+struct CompileAssertTypesEqual<T, T> {
+};
+
+// Removes the reference from a type if it is a reference type,
+// otherwise leaves it unchanged.  This is the same as
+// tr1::remove_reference, which is not widely available yet.
+template <typename T>
+struct RemoveReference { typedef T type; };  // NOLINT
+template <typename T>
+struct RemoveReference<T&> { typedef T type; };  // NOLINT
+
+// A handy wrapper around RemoveReference that works when the argument
+// T depends on template parameters.
+#define GTEST_REMOVE_REFERENCE_(T) \
+    typename ::testing::internal::RemoveReference<T>::type
+
+// Removes const from a type if it is a const type, otherwise leaves
+// it unchanged.  This is the same as tr1::remove_const, which is not
+// widely available yet.
+template <typename T>
+struct RemoveConst { typedef T type; };  // NOLINT
+template <typename T>
+struct RemoveConst<const T> { typedef T type; };  // NOLINT
+
+// MSVC 8.0, Sun C++, and IBM XL C++ have a bug which causes the above
+// definition to fail to remove the const in 'const int[3]' and 'const
+// char[3][4]'.  The following specialization works around the bug.
+template <typename T, size_t N>
+struct RemoveConst<const T[N]> {
+  typedef typename RemoveConst<T>::type type[N];
+};
+
+#if defined(_MSC_VER) && _MSC_VER < 1400
+// This is the only specialization that allows VC++ 7.1 to remove const in
+// 'const int[3] and 'const int[3][4]'.  However, it causes trouble with GCC
+// and thus needs to be conditionally compiled.
+template <typename T, size_t N>
+struct RemoveConst<T[N]> {
+  typedef typename RemoveConst<T>::type type[N];
+};
+#endif
+
+// A handy wrapper around RemoveConst that works when the argument
+// T depends on template parameters.
+#define GTEST_REMOVE_CONST_(T) \
+    typename ::testing::internal::RemoveConst<T>::type
+
+// Turns const U&, U&, const U, and U all into U.
+#define GTEST_REMOVE_REFERENCE_AND_CONST_(T) \
+    GTEST_REMOVE_CONST_(GTEST_REMOVE_REFERENCE_(T))
+
+// Adds reference to a type if it is not a reference type,
+// otherwise leaves it unchanged.  This is the same as
+// tr1::add_reference, which is not widely available yet.
+template <typename T>
+struct AddReference { typedef T& type; };  // NOLINT
+template <typename T>
+struct AddReference<T&> { typedef T& type; };  // NOLINT
+
+// A handy wrapper around AddReference that works when the argument T
+// depends on template parameters.
+#define GTEST_ADD_REFERENCE_(T) \
+    typename ::testing::internal::AddReference<T>::type
+
+// Adds a reference to const on top of T as necessary.  For example,
+// it transforms
+//
+//   char         ==> const char&
+//   const char   ==> const char&
+//   char&        ==> const char&
+//   const char&  ==> const char&
+//
+// The argument T must depend on some template parameters.
+#define GTEST_REFERENCE_TO_CONST_(T) \
+    GTEST_ADD_REFERENCE_(const GTEST_REMOVE_REFERENCE_(T))
+
+// ImplicitlyConvertible<From, To>::value is a compile-time bool
+// constant that's true iff type From can be implicitly converted to
+// type To.
+template <typename From, typename To>
+class ImplicitlyConvertible {
+ private:
+  // We need the following helper functions only for their types.
+  // They have no implementations.
+
+  // MakeFrom() is an expression whose type is From.  We cannot simply
+  // use From(), as the type From may not have a public default
+  // constructor.
+  static From MakeFrom();
+
+  // These two functions are overloaded.  Given an expression
+  // Helper(x), the compiler will pick the first version if x can be
+  // implicitly converted to type To; otherwise it will pick the
+  // second version.
+  //
+  // The first version returns a value of size 1, and the second
+  // version returns a value of size 2.  Therefore, by checking the
+  // size of Helper(x), which can be done at compile time, we can tell
+  // which version of Helper() is used, and hence whether x can be
+  // implicitly converted to type To.
+  static char Helper(To);
+  static char (&Helper(...))[2];  // NOLINT
+
+  // We have to put the 'public' section after the 'private' section,
+  // or MSVC refuses to compile the code.
+ public:
+  // MSVC warns about implicitly converting from double to int for
+  // possible loss of data, so we need to temporarily disable the
+  // warning.
+#ifdef _MSC_VER
+# pragma warning(push)          // Saves the current warning state.
+# pragma warning(disable:4244)  // Temporarily disables warning 4244.
+
+  static const bool value =
+      sizeof(Helper(ImplicitlyConvertible::MakeFrom())) == 1;
+# pragma warning(pop)           // Restores the warning state.
+#elif defined(__BORLANDC__)
+  // C++Builder cannot use member overload resolution during template
+  // instantiation.  The simplest workaround is to use its C++0x type traits
+  // functions (C++Builder 2009 and above only).
+  static const bool value = __is_convertible(From, To);
+#else
+  static const bool value =
+      sizeof(Helper(ImplicitlyConvertible::MakeFrom())) == 1;
+#endif  // _MSV_VER
+};
+template <typename From, typename To>
+const bool ImplicitlyConvertible<From, To>::value;
+
+// IsAProtocolMessage<T>::value is a compile-time bool constant that's
+// true iff T is type ProtocolMessage, proto2::Message, or a subclass
+// of those.
+template <typename T>
+struct IsAProtocolMessage
+    : public bool_constant<
+  ImplicitlyConvertible<const T*, const ::ProtocolMessage*>::value ||
+  ImplicitlyConvertible<const T*, const ::proto2::Message*>::value> {
+};
+
+// When the compiler sees expression IsContainerTest<C>(0), if C is an
+// STL-style container class, the first overload of IsContainerTest
+// will be viable (since both C::iterator* and C::const_iterator* are
+// valid types and NULL can be implicitly converted to them).  It will
+// be picked over the second overload as 'int' is a perfect match for
+// the type of argument 0.  If C::iterator or C::const_iterator is not
+// a valid type, the first overload is not viable, and the second
+// overload will be picked.  Therefore, we can determine whether C is
+// a container class by checking the type of IsContainerTest<C>(0).
+// The value of the expression is insignificant.
+//
+// Note that we look for both C::iterator and C::const_iterator.  The
+// reason is that C++ injects the name of a class as a member of the
+// class itself (e.g. you can refer to class iterator as either
+// 'iterator' or 'iterator::iterator').  If we look for C::iterator
+// only, for example, we would mistakenly think that a class named
+// iterator is an STL container.
+//
+// Also note that the simpler approach of overloading
+// IsContainerTest(typename C::const_iterator*) and
+// IsContainerTest(...) doesn't work with Visual Age C++ and Sun C++.
+typedef int IsContainer;
+template <class C>
+IsContainer IsContainerTest(int /* dummy */,
+                            typename C::iterator* /* it */ = NULL,
+                            typename C::const_iterator* /* const_it */ = NULL) {
+  return 0;
+}
+
+typedef char IsNotContainer;
+template <class C>
+IsNotContainer IsContainerTest(long /* dummy */) { return '\0'; }
+
+// EnableIf<condition>::type is void when 'Cond' is true, and
+// undefined when 'Cond' is false.  To use SFINAE to make a function
+// overload only apply when a particular expression is true, add
+// "typename EnableIf<expression>::type* = 0" as the last parameter.
+template<bool> struct EnableIf;
+template<> struct EnableIf<true> { typedef void type; };  // NOLINT
+
+// Utilities for native arrays.
+
+// ArrayEq() compares two k-dimensional native arrays using the
+// elements' operator==, where k can be any integer >= 0.  When k is
+// 0, ArrayEq() degenerates into comparing a single pair of values.
+
+template <typename T, typename U>
+bool ArrayEq(const T* lhs, size_t size, const U* rhs);
+
+// This generic version is used when k is 0.
+template <typename T, typename U>
+inline bool ArrayEq(const T& lhs, const U& rhs) { return lhs == rhs; }
+
+// This overload is used when k >= 1.
+template <typename T, typename U, size_t N>
+inline bool ArrayEq(const T(&lhs)[N], const U(&rhs)[N]) {
+  return internal::ArrayEq(lhs, N, rhs);
+}
+
+// This helper reduces code bloat.  If we instead put its logic inside
+// the previous ArrayEq() function, arrays with different sizes would
+// lead to different copies of the template code.
+template <typename T, typename U>
+bool ArrayEq(const T* lhs, size_t size, const U* rhs) {
+  for (size_t i = 0; i != size; i++) {
+    if (!internal::ArrayEq(lhs[i], rhs[i]))
+      return false;
+  }
+  return true;
+}
+
+// Finds the first element in the iterator range [begin, end) that
+// equals elem.  Element may be a native array type itself.
+template <typename Iter, typename Element>
+Iter ArrayAwareFind(Iter begin, Iter end, const Element& elem) {
+  for (Iter it = begin; it != end; ++it) {
+    if (internal::ArrayEq(*it, elem))
+      return it;
+  }
+  return end;
+}
+
+// CopyArray() copies a k-dimensional native array using the elements'
+// operator=, where k can be any integer >= 0.  When k is 0,
+// CopyArray() degenerates into copying a single value.
+
+template <typename T, typename U>
+void CopyArray(const T* from, size_t size, U* to);
+
+// This generic version is used when k is 0.
+template <typename T, typename U>
+inline void CopyArray(const T& from, U* to) { *to = from; }
+
+// This overload is used when k >= 1.
+template <typename T, typename U, size_t N>
+inline void CopyArray(const T(&from)[N], U(*to)[N]) {
+  internal::CopyArray(from, N, *to);
+}
+
+// This helper reduces code bloat.  If we instead put its logic inside
+// the previous CopyArray() function, arrays with different sizes
+// would lead to different copies of the template code.
+template <typename T, typename U>
+void CopyArray(const T* from, size_t size, U* to) {
+  for (size_t i = 0; i != size; i++) {
+    internal::CopyArray(from[i], to + i);
+  }
+}
+
+// The relation between an NativeArray object (see below) and the
+// native array it represents.
+enum RelationToSource {
+  kReference,  // The NativeArray references the native array.
+  kCopy        // The NativeArray makes a copy of the native array and
+               // owns the copy.
+};
+
+// Adapts a native array to a read-only STL-style container.  Instead
+// of the complete STL container concept, this adaptor only implements
+// members useful for Google Mock's container matchers.  New members
+// should be added as needed.  To simplify the implementation, we only
+// support Element being a raw type (i.e. having no top-level const or
+// reference modifier).  It's the client's responsibility to satisfy
+// this requirement.  Element can be an array type itself (hence
+// multi-dimensional arrays are supported).
+template <typename Element>
+class NativeArray {
+ public:
+  // STL-style container typedefs.
+  typedef Element value_type;
+  typedef Element* iterator;
+  typedef const Element* const_iterator;
+
+  // Constructs from a native array.
+  NativeArray(const Element* array, size_t count, RelationToSource relation) {
+    Init(array, count, relation);
+  }
+
+  // Copy constructor.
+  NativeArray(const NativeArray& rhs) {
+    Init(rhs.array_, rhs.size_, rhs.relation_to_source_);
+  }
+
+  ~NativeArray() {
+    // Ensures that the user doesn't instantiate NativeArray with a
+    // const or reference type.
+    static_cast<void>(StaticAssertTypeEqHelper<Element,
+        GTEST_REMOVE_REFERENCE_AND_CONST_(Element)>());
+    if (relation_to_source_ == kCopy)
+      delete[] array_;
+  }
+
+  // STL-style container methods.
+  size_t size() const { return size_; }
+  const_iterator begin() const { return array_; }
+  const_iterator end() const { return array_ + size_; }
+  bool operator==(const NativeArray& rhs) const {
+    return size() == rhs.size() &&
+        ArrayEq(begin(), size(), rhs.begin());
+  }
+
+ private:
+  // Initializes this object; makes a copy of the input array if
+  // 'relation' is kCopy.
+  void Init(const Element* array, size_t a_size, RelationToSource relation) {
+    if (relation == kReference) {
+      array_ = array;
+    } else {
+      Element* const copy = new Element[a_size];
+      CopyArray(array, a_size, copy);
+      array_ = copy;
+    }
+    size_ = a_size;
+    relation_to_source_ = relation;
+  }
+
+  const Element* array_;
+  size_t size_;
+  RelationToSource relation_to_source_;
+
+  GTEST_DISALLOW_ASSIGN_(NativeArray);
+};
+
+}  // namespace internal
+}  // namespace testing
+
+#define GTEST_MESSAGE_AT_(file, line, message, result_type) \
+  ::testing::internal::AssertHelper(result_type, file, line, message) \
+    = ::testing::Message()
+
+#define GTEST_MESSAGE_(message, result_type) \
+  GTEST_MESSAGE_AT_(__FILE__, __LINE__, message, result_type)
+
+#define GTEST_FATAL_FAILURE_(message) \
+  return GTEST_MESSAGE_(message, ::testing::TestPartResult::kFatalFailure)
+
+#define GTEST_NONFATAL_FAILURE_(message) \
+  GTEST_MESSAGE_(message, ::testing::TestPartResult::kNonFatalFailure)
+
+#define GTEST_SUCCESS_(message) \
+  GTEST_MESSAGE_(message, ::testing::TestPartResult::kSuccess)
+
+// Suppresses MSVC warnings 4072 (unreachable code) for the code following
+// statement if it returns or throws (or doesn't return or throw in some
+// situations).
+#define GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement) \
+  if (::testing::internal::AlwaysTrue()) { statement; }
+
+#define GTEST_TEST_THROW_(statement, expected_exception, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::ConstCharPtr gtest_msg = "") { \
+    bool gtest_caught_expected = false; \
+    try { \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    } \
+    catch (expected_exception const&) { \
+      gtest_caught_expected = true; \
+    } \
+    catch (...) { \
+      gtest_msg.value = \
+          "Expected: " #statement " throws an exception of type " \
+          #expected_exception ".\n  Actual: it throws a different type."; \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testthrow_, __LINE__); \
+    } \
+    if (!gtest_caught_expected) { \
+      gtest_msg.value = \
+          "Expected: " #statement " throws an exception of type " \
+          #expected_exception ".\n  Actual: it throws nothing."; \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testthrow_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testthrow_, __LINE__): \
+      fail(gtest_msg.value)
+
+#define GTEST_TEST_NO_THROW_(statement, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    try { \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    } \
+    catch (...) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testnothrow_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testnothrow_, __LINE__): \
+      fail("Expected: " #statement " doesn't throw an exception.\n" \
+           "  Actual: it throws.")
+
+#define GTEST_TEST_ANY_THROW_(statement, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    bool gtest_caught_any = false; \
+    try { \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    } \
+    catch (...) { \
+      gtest_caught_any = true; \
+    } \
+    if (!gtest_caught_any) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testanythrow_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testanythrow_, __LINE__): \
+      fail("Expected: " #statement " throws an exception.\n" \
+           "  Actual: it doesn't.")
+
+
+// Implements Boolean test assertions such as EXPECT_TRUE. expression can be
+// either a boolean expression or an AssertionResult. text is a textual
+// represenation of expression as it was passed into the EXPECT_TRUE.
+#define GTEST_TEST_BOOLEAN_(expression, text, actual, expected, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (const ::testing::AssertionResult gtest_ar_ = \
+      ::testing::AssertionResult(expression)) \
+    ; \
+  else \
+    fail(::testing::internal::GetBoolAssertionFailureMessage(\
+        gtest_ar_, text, #actual, #expected).c_str())
+
+#define GTEST_TEST_NO_FATAL_FAILURE_(statement, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    ::testing::internal::HasNewFatalFailureHelper gtest_fatal_failure_checker; \
+    GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    if (gtest_fatal_failure_checker.has_new_fatal_failure()) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testnofatal_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testnofatal_, __LINE__): \
+      fail("Expected: " #statement " doesn't generate new fatal " \
+           "failures in the current thread.\n" \
+           "  Actual: it does.")
+
+// Expands to the name of the class that implements the given test.
+#define GTEST_TEST_CLASS_NAME_(test_case_name, test_name) \
+  test_case_name##_##test_name##_Test
+
+// Helper macro for defining tests.
+#define GTEST_TEST_(test_case_name, test_name, parent_class, parent_id)\
+class GTEST_TEST_CLASS_NAME_(test_case_name, test_name) : public parent_class {\
+ public:\
+  GTEST_TEST_CLASS_NAME_(test_case_name, test_name)() {}\
+ private:\
+  virtual void TestBody();\
+  static ::testing::TestInfo* const test_info_ GTEST_ATTRIBUTE_UNUSED_;\
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(\
+      GTEST_TEST_CLASS_NAME_(test_case_name, test_name));\
+};\
+\
+::testing::TestInfo* const GTEST_TEST_CLASS_NAME_(test_case_name, test_name)\
+  ::test_info_ =\
+    ::testing::internal::MakeAndRegisterTestInfo(\
+        #test_case_name, #test_name, NULL, NULL, \
+        (parent_id), \
+        parent_class::SetUpTestCase, \
+        parent_class::TearDownTestCase, \
+        new ::testing::internal::TestFactoryImpl<\
+            GTEST_TEST_CLASS_NAME_(test_case_name, test_name)>);\
+void GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::TestBody()
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_INTERNAL_H_
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines the public API for death tests.  It is
+// #included by gtest.h so a user doesn't need to include this
+// directly.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_DEATH_TEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_DEATH_TEST_H_
+
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan), eefacm@gmail.com (Sean Mcafee)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines internal utilities needed for implementing
+// death tests.  They are subject to change without notice.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_DEATH_TEST_INTERNAL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_DEATH_TEST_INTERNAL_H_
+
+
+#include <stdio.h>
+
+namespace testing {
+namespace internal {
+
+GTEST_DECLARE_string_(internal_run_death_test);
+
+// Names of the flags (needed for parsing Google Test flags).
+const char kDeathTestStyleFlag[] = "death_test_style";
+const char kDeathTestUseFork[] = "death_test_use_fork";
+const char kInternalRunDeathTestFlag[] = "internal_run_death_test";
+
+#if GTEST_HAS_DEATH_TEST
+
+// DeathTest is a class that hides much of the complexity of the
+// GTEST_DEATH_TEST_ macro.  It is abstract; its static Create method
+// returns a concrete class that depends on the prevailing death test
+// style, as defined by the --gtest_death_test_style and/or
+// --gtest_internal_run_death_test flags.
+
+// In describing the results of death tests, these terms are used with
+// the corresponding definitions:
+//
+// exit status:  The integer exit information in the format specified
+//               by wait(2)
+// exit code:    The integer code passed to exit(3), _exit(2), or
+//               returned from main()
+class GTEST_API_ DeathTest {
+ public:
+  // Create returns false if there was an error determining the
+  // appropriate action to take for the current death test; for example,
+  // if the gtest_death_test_style flag is set to an invalid value.
+  // The LastMessage method will return a more detailed message in that
+  // case.  Otherwise, the DeathTest pointer pointed to by the "test"
+  // argument is set.  If the death test should be skipped, the pointer
+  // is set to NULL; otherwise, it is set to the address of a new concrete
+  // DeathTest object that controls the execution of the current test.
+  static bool Create(const char* statement, const RE* regex,
+                     const char* file, int line, DeathTest** test);
+  DeathTest();
+  virtual ~DeathTest() { }
+
+  // A helper class that aborts a death test when it's deleted.
+  class ReturnSentinel {
+   public:
+    explicit ReturnSentinel(DeathTest* test) : test_(test) { }
+    ~ReturnSentinel() { test_->Abort(TEST_ENCOUNTERED_RETURN_STATEMENT); }
+   private:
+    DeathTest* const test_;
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(ReturnSentinel);
+  } GTEST_ATTRIBUTE_UNUSED_;
+
+  // An enumeration of possible roles that may be taken when a death
+  // test is encountered.  EXECUTE means that the death test logic should
+  // be executed immediately.  OVERSEE means that the program should prepare
+  // the appropriate environment for a child process to execute the death
+  // test, then wait for it to complete.
+  enum TestRole { OVERSEE_TEST, EXECUTE_TEST };
+
+  // An enumeration of the three reasons that a test might be aborted.
+  enum AbortReason {
+    TEST_ENCOUNTERED_RETURN_STATEMENT,
+    TEST_THREW_EXCEPTION,
+    TEST_DID_NOT_DIE
+  };
+
+  // Assumes one of the above roles.
+  virtual TestRole AssumeRole() = 0;
+
+  // Waits for the death test to finish and returns its status.
+  virtual int Wait() = 0;
+
+  // Returns true if the death test passed; that is, the test process
+  // exited during the test, its exit status matches a user-supplied
+  // predicate, and its stderr output matches a user-supplied regular
+  // expression.
+  // The user-supplied predicate may be a macro expression rather
+  // than a function pointer or functor, or else Wait and Passed could
+  // be combined.
+  virtual bool Passed(bool exit_status_ok) = 0;
+
+  // Signals that the death test did not die as expected.
+  virtual void Abort(AbortReason reason) = 0;
+
+  // Returns a human-readable outcome message regarding the outcome of
+  // the last death test.
+  static const char* LastMessage();
+
+  static void set_last_death_test_message(const std::string& message);
+
+ private:
+  // A string containing a description of the outcome of the last death test.
+  static std::string last_death_test_message_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(DeathTest);
+};
+
+// Factory interface for death tests.  May be mocked out for testing.
+class DeathTestFactory {
+ public:
+  virtual ~DeathTestFactory() { }
+  virtual bool Create(const char* statement, const RE* regex,
+                      const char* file, int line, DeathTest** test) = 0;
+};
+
+// A concrete DeathTestFactory implementation for normal use.
+class DefaultDeathTestFactory : public DeathTestFactory {
+ public:
+  virtual bool Create(const char* statement, const RE* regex,
+                      const char* file, int line, DeathTest** test);
+};
+
+// Returns true if exit_status describes a process that was terminated
+// by a signal, or exited normally with a nonzero exit code.
+GTEST_API_ bool ExitedUnsuccessfully(int exit_status);
+
+// Traps C++ exceptions escaping statement and reports them as test
+// failures. Note that trapping SEH exceptions is not implemented here.
+# if GTEST_HAS_EXCEPTIONS
+#  define GTEST_EXECUTE_DEATH_TEST_STATEMENT_(statement, death_test) \
+  try { \
+    GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+  } catch (const ::std::exception& gtest_exception) { \
+    fprintf(\
+        stderr, \
+        "\n%s: Caught std::exception-derived exception escaping the " \
+        "death test statement. Exception message: %s\n", \
+        ::testing::internal::FormatFileLocation(__FILE__, __LINE__).c_str(), \
+        gtest_exception.what()); \
+    fflush(stderr); \
+    death_test->Abort(::testing::internal::DeathTest::TEST_THREW_EXCEPTION); \
+  } catch (...) { \
+    death_test->Abort(::testing::internal::DeathTest::TEST_THREW_EXCEPTION); \
+  }
+
+# else
+#  define GTEST_EXECUTE_DEATH_TEST_STATEMENT_(statement, death_test) \
+  GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement)
+
+# endif
+
+// This macro is for implementing ASSERT_DEATH*, EXPECT_DEATH*,
+// ASSERT_EXIT*, and EXPECT_EXIT*.
+# define GTEST_DEATH_TEST_(statement, predicate, regex, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    const ::testing::internal::RE& gtest_regex = (regex); \
+    ::testing::internal::DeathTest* gtest_dt; \
+    if (!::testing::internal::DeathTest::Create(#statement, &gtest_regex, \
+        __FILE__, __LINE__, &gtest_dt)) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_, __LINE__); \
+    } \
+    if (gtest_dt != NULL) { \
+      ::testing::internal::scoped_ptr< ::testing::internal::DeathTest> \
+          gtest_dt_ptr(gtest_dt); \
+      switch (gtest_dt->AssumeRole()) { \
+        case ::testing::internal::DeathTest::OVERSEE_TEST: \
+          if (!gtest_dt->Passed(predicate(gtest_dt->Wait()))) { \
+            goto GTEST_CONCAT_TOKEN_(gtest_label_, __LINE__); \
+          } \
+          break; \
+        case ::testing::internal::DeathTest::EXECUTE_TEST: { \
+          ::testing::internal::DeathTest::ReturnSentinel \
+              gtest_sentinel(gtest_dt); \
+          GTEST_EXECUTE_DEATH_TEST_STATEMENT_(statement, gtest_dt); \
+          gtest_dt->Abort(::testing::internal::DeathTest::TEST_DID_NOT_DIE); \
+          break; \
+        } \
+        default: \
+          break; \
+      } \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_, __LINE__): \
+      fail(::testing::internal::DeathTest::LastMessage())
+// The symbol "fail" here expands to something into which a message
+// can be streamed.
+
+// This macro is for implementing ASSERT/EXPECT_DEBUG_DEATH when compiled in
+// NDEBUG mode. In this case we need the statements to be executed, the regex is
+// ignored, and the macro must accept a streamed message even though the message
+// is never printed.
+# define GTEST_EXECUTE_STATEMENT_(statement, regex) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+     GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+  } else \
+    ::testing::Message()
+
+// A class representing the parsed contents of the
+// --gtest_internal_run_death_test flag, as it existed when
+// RUN_ALL_TESTS was called.
+class InternalRunDeathTestFlag {
+ public:
+  InternalRunDeathTestFlag(const std::string& a_file,
+                           int a_line,
+                           int an_index,
+                           int a_write_fd)
+      : file_(a_file), line_(a_line), index_(an_index),
+        write_fd_(a_write_fd) {}
+
+  ~InternalRunDeathTestFlag() {
+    if (write_fd_ >= 0)
+      posix::Close(write_fd_);
+  }
+
+  const std::string& file() const { return file_; }
+  int line() const { return line_; }
+  int index() const { return index_; }
+  int write_fd() const { return write_fd_; }
+
+ private:
+  std::string file_;
+  int line_;
+  int index_;
+  int write_fd_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(InternalRunDeathTestFlag);
+};
+
+// Returns a newly created InternalRunDeathTestFlag object with fields
+// initialized from the GTEST_FLAG(internal_run_death_test) flag if
+// the flag is specified; otherwise returns NULL.
+InternalRunDeathTestFlag* ParseInternalRunDeathTestFlag();
+
+#else  // GTEST_HAS_DEATH_TEST
+
+// This macro is used for implementing macros such as
+// EXPECT_DEATH_IF_SUPPORTED and ASSERT_DEATH_IF_SUPPORTED on systems where
+// death tests are not supported. Those macros must compile on such systems
+// iff EXPECT_DEATH and ASSERT_DEATH compile with the same parameters on
+// systems that support death tests. This allows one to write such a macro
+// on a system that does not support death tests and be sure that it will
+// compile on a death-test supporting system.
+//
+// Parameters:
+//   statement -  A statement that a macro such as EXPECT_DEATH would test
+//                for program termination. This macro has to make sure this
+//                statement is compiled but not executed, to ensure that
+//                EXPECT_DEATH_IF_SUPPORTED compiles with a certain
+//                parameter iff EXPECT_DEATH compiles with it.
+//   regex     -  A regex that a macro such as EXPECT_DEATH would use to test
+//                the output of statement.  This parameter has to be
+//                compiled but not evaluated by this macro, to ensure that
+//                this macro only accepts expressions that a macro such as
+//                EXPECT_DEATH would accept.
+//   terminator - Must be an empty statement for EXPECT_DEATH_IF_SUPPORTED
+//                and a return statement for ASSERT_DEATH_IF_SUPPORTED.
+//                This ensures that ASSERT_DEATH_IF_SUPPORTED will not
+//                compile inside functions where ASSERT_DEATH doesn't
+//                compile.
+//
+//  The branch that has an always false condition is used to ensure that
+//  statement and regex are compiled (and thus syntactically correct) but
+//  never executed. The unreachable code macro protects the terminator
+//  statement from generating an 'unreachable code' warning in case
+//  statement unconditionally returns or throws. The Message constructor at
+//  the end allows the syntax of streaming additional messages into the
+//  macro, for compilational compatibility with EXPECT_DEATH/ASSERT_DEATH.
+# define GTEST_UNSUPPORTED_DEATH_TEST_(statement, regex, terminator) \
+    GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+    if (::testing::internal::AlwaysTrue()) { \
+      GTEST_LOG_(WARNING) \
+          << "Death tests are not supported on this platform.\n" \
+          << "Statement '" #statement "' cannot be verified."; \
+    } else if (::testing::internal::AlwaysFalse()) { \
+      ::testing::internal::RE::PartialMatch(".*", (regex)); \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+      terminator; \
+    } else \
+      ::testing::Message()
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_DEATH_TEST_INTERNAL_H_
+
+namespace testing {
+
+// This flag controls the style of death tests.  Valid values are "threadsafe",
+// meaning that the death test child process will re-execute the test binary
+// from the start, running only a single death test, or "fast",
+// meaning that the child process will execute the test logic immediately
+// after forking.
+GTEST_DECLARE_string_(death_test_style);
+
+#if GTEST_HAS_DEATH_TEST
+
+namespace internal {
+
+// Returns a Boolean value indicating whether the caller is currently
+// executing in the context of the death test child process.  Tools such as
+// Valgrind heap checkers may need this to modify their behavior in death
+// tests.  IMPORTANT: This is an internal utility.  Using it may break the
+// implementation of death tests.  User code MUST NOT use it.
+GTEST_API_ bool InDeathTestChild();
+
+}  // namespace internal
+
+// The following macros are useful for writing death tests.
+
+// Here's what happens when an ASSERT_DEATH* or EXPECT_DEATH* is
+// executed:
+//
+//   1. It generates a warning if there is more than one active
+//   thread.  This is because it's safe to fork() or clone() only
+//   when there is a single thread.
+//
+//   2. The parent process clone()s a sub-process and runs the death
+//   test in it; the sub-process exits with code 0 at the end of the
+//   death test, if it hasn't exited already.
+//
+//   3. The parent process waits for the sub-process to terminate.
+//
+//   4. The parent process checks the exit code and error message of
+//   the sub-process.
+//
+// Examples:
+//
+//   ASSERT_DEATH(server.SendMessage(56, "Hello"), "Invalid port number");
+//   for (int i = 0; i < 5; i++) {
+//     EXPECT_DEATH(server.ProcessRequest(i),
+//                  "Invalid request .* in ProcessRequest()")
+//                  << "Failed to die on request " << i;
+//   }
+//
+//   ASSERT_EXIT(server.ExitNow(), ::testing::ExitedWithCode(0), "Exiting");
+//
+//   bool KilledBySIGHUP(int exit_code) {
+//     return WIFSIGNALED(exit_code) && WTERMSIG(exit_code) == SIGHUP;
+//   }
+//
+//   ASSERT_EXIT(client.HangUpServer(), KilledBySIGHUP, "Hanging up!");
+//
+// On the regular expressions used in death tests:
+//
+//   On POSIX-compliant systems (*nix), we use the <regex.h> library,
+//   which uses the POSIX extended regex syntax.
+//
+//   On other platforms (e.g. Windows), we only support a simple regex
+//   syntax implemented as part of Google Test.  This limited
+//   implementation should be enough most of the time when writing
+//   death tests; though it lacks many features you can find in PCRE
+//   or POSIX extended regex syntax.  For example, we don't support
+//   union ("x|y"), grouping ("(xy)"), brackets ("[xy]"), and
+//   repetition count ("x{5,7}"), among others.
+//
+//   Below is the syntax that we do support.  We chose it to be a
+//   subset of both PCRE and POSIX extended regex, so it's easy to
+//   learn wherever you come from.  In the following: 'A' denotes a
+//   literal character, period (.), or a single \\ escape sequence;
+//   'x' and 'y' denote regular expressions; 'm' and 'n' are for
+//   natural numbers.
+//
+//     c     matches any literal character c
+//     \\d   matches any decimal digit
+//     \\D   matches any character that's not a decimal digit
+//     \\f   matches \f
+//     \\n   matches \n
+//     \\r   matches \r
+//     \\s   matches any ASCII whitespace, including \n
+//     \\S   matches any character that's not a whitespace
+//     \\t   matches \t
+//     \\v   matches \v
+//     \\w   matches any letter, _, or decimal digit
+//     \\W   matches any character that \\w doesn't match
+//     \\c   matches any literal character c, which must be a punctuation
+//     .     matches any single character except \n
+//     A?    matches 0 or 1 occurrences of A
+//     A*    matches 0 or many occurrences of A
+//     A+    matches 1 or many occurrences of A
+//     ^     matches the beginning of a string (not that of each line)
+//     $     matches the end of a string (not that of each line)
+//     xy    matches x followed by y
+//
+//   If you accidentally use PCRE or POSIX extended regex features
+//   not implemented by us, you will get a run-time failure.  In that
+//   case, please try to rewrite your regular expression within the
+//   above syntax.
+//
+//   This implementation is *not* meant to be as highly tuned or robust
+//   as a compiled regex library, but should perform well enough for a
+//   death test, which already incurs significant overhead by launching
+//   a child process.
+//
+// Known caveats:
+//
+//   A "threadsafe" style death test obtains the path to the test
+//   program from argv[0] and re-executes it in the sub-process.  For
+//   simplicity, the current implementation doesn't search the PATH
+//   when launching the sub-process.  This means that the user must
+//   invoke the test program via a path that contains at least one
+//   path separator (e.g. path/to/foo_test and
+//   /absolute/path/to/bar_test are fine, but foo_test is not).  This
+//   is rarely a problem as people usually don't put the test binary
+//   directory in PATH.
+//
+// TODO(wan@google.com): make thread-safe death tests search the PATH.
+
+// Asserts that a given statement causes the program to exit, with an
+// integer exit status that satisfies predicate, and emitting error output
+// that matches regex.
+# define ASSERT_EXIT(statement, predicate, regex) \
+    GTEST_DEATH_TEST_(statement, predicate, regex, GTEST_FATAL_FAILURE_)
+
+// Like ASSERT_EXIT, but continues on to successive tests in the
+// test case, if any:
+# define EXPECT_EXIT(statement, predicate, regex) \
+    GTEST_DEATH_TEST_(statement, predicate, regex, GTEST_NONFATAL_FAILURE_)
+
+// Asserts that a given statement causes the program to exit, either by
+// explicitly exiting with a nonzero exit code or being killed by a
+// signal, and emitting error output that matches regex.
+# define ASSERT_DEATH(statement, regex) \
+    ASSERT_EXIT(statement, ::testing::internal::ExitedUnsuccessfully, regex)
+
+// Like ASSERT_DEATH, but continues on to successive tests in the
+// test case, if any:
+# define EXPECT_DEATH(statement, regex) \
+    EXPECT_EXIT(statement, ::testing::internal::ExitedUnsuccessfully, regex)
+
+// Two predicate classes that can be used in {ASSERT,EXPECT}_EXIT*:
+
+// Tests that an exit code describes a normal exit with a given exit code.
+class GTEST_API_ ExitedWithCode {
+ public:
+  explicit ExitedWithCode(int exit_code);
+  bool operator()(int exit_status) const;
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ExitedWithCode& other);
+
+  const int exit_code_;
+};
+
+# if !GTEST_OS_WINDOWS
+// Tests that an exit code describes an exit due to termination by a
+// given signal.
+class GTEST_API_ KilledBySignal {
+ public:
+  explicit KilledBySignal(int signum);
+  bool operator()(int exit_status) const;
+ private:
+  const int signum_;
+};
+# endif  // !GTEST_OS_WINDOWS
+
+// EXPECT_DEBUG_DEATH asserts that the given statements die in debug mode.
+// The death testing framework causes this to have interesting semantics,
+// since the sideeffects of the call are only visible in opt mode, and not
+// in debug mode.
+//
+// In practice, this can be used to test functions that utilize the
+// LOG(DFATAL) macro using the following style:
+//
+// int DieInDebugOr12(int* sideeffect) {
+//   if (sideeffect) {
+//     *sideeffect = 12;
+//   }
+//   LOG(DFATAL) << "death";
+//   return 12;
+// }
+//
+// TEST(TestCase, TestDieOr12WorksInDgbAndOpt) {
+//   int sideeffect = 0;
+//   // Only asserts in dbg.
+//   EXPECT_DEBUG_DEATH(DieInDebugOr12(&sideeffect), "death");
+//
+// #ifdef NDEBUG
+//   // opt-mode has sideeffect visible.
+//   EXPECT_EQ(12, sideeffect);
+// #else
+//   // dbg-mode no visible sideeffect.
+//   EXPECT_EQ(0, sideeffect);
+// #endif
+// }
+//
+// This will assert that DieInDebugReturn12InOpt() crashes in debug
+// mode, usually due to a DCHECK or LOG(DFATAL), but returns the
+// appropriate fallback value (12 in this case) in opt mode. If you
+// need to test that a function has appropriate side-effects in opt
+// mode, include assertions against the side-effects.  A general
+// pattern for this is:
+//
+// EXPECT_DEBUG_DEATH({
+//   // Side-effects here will have an effect after this statement in
+//   // opt mode, but none in debug mode.
+//   EXPECT_EQ(12, DieInDebugOr12(&sideeffect));
+// }, "death");
+//
+# ifdef NDEBUG
+
+#  define EXPECT_DEBUG_DEATH(statement, regex) \
+  GTEST_EXECUTE_STATEMENT_(statement, regex)
+
+#  define ASSERT_DEBUG_DEATH(statement, regex) \
+  GTEST_EXECUTE_STATEMENT_(statement, regex)
+
+# else
+
+#  define EXPECT_DEBUG_DEATH(statement, regex) \
+  EXPECT_DEATH(statement, regex)
+
+#  define ASSERT_DEBUG_DEATH(statement, regex) \
+  ASSERT_DEATH(statement, regex)
+
+# endif  // NDEBUG for EXPECT_DEBUG_DEATH
+#endif  // GTEST_HAS_DEATH_TEST
+
+// EXPECT_DEATH_IF_SUPPORTED(statement, regex) and
+// ASSERT_DEATH_IF_SUPPORTED(statement, regex) expand to real death tests if
+// death tests are supported; otherwise they just issue a warning.  This is
+// useful when you are combining death test assertions with normal test
+// assertions in one test.
+#if GTEST_HAS_DEATH_TEST
+# define EXPECT_DEATH_IF_SUPPORTED(statement, regex) \
+    EXPECT_DEATH(statement, regex)
+# define ASSERT_DEATH_IF_SUPPORTED(statement, regex) \
+    ASSERT_DEATH(statement, regex)
+#else
+# define EXPECT_DEATH_IF_SUPPORTED(statement, regex) \
+    GTEST_UNSUPPORTED_DEATH_TEST_(statement, regex, )
+# define ASSERT_DEATH_IF_SUPPORTED(statement, regex) \
+    GTEST_UNSUPPORTED_DEATH_TEST_(statement, regex, return)
+#endif
+
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_DEATH_TEST_H_
+// This file was GENERATED by command:
+//     pump.py gtest-param-test.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: vladl@google.com (Vlad Losev)
+//
+// Macros and functions for implementing parameterized tests
+// in Google C++ Testing Framework (Google Test)
+//
+// This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
+
+
+// Value-parameterized tests allow you to test your code with different
+// parameters without writing multiple copies of the same test.
+//
+// Here is how you use value-parameterized tests:
+
+#if 0
+
+// To write value-parameterized tests, first you should define a fixture
+// class. It is usually derived from testing::TestWithParam<T> (see below for
+// another inheritance scheme that's sometimes useful in more complicated
+// class hierarchies), where the type of your parameter values.
+// TestWithParam<T> is itself derived from testing::Test. T can be any
+// copyable type. If it's a raw pointer, you are responsible for managing the
+// lifespan of the pointed values.
+
+class FooTest : public ::testing::TestWithParam<const char*> {
+  // You can implement all the usual class fixture members here.
+};
+
+// Then, use the TEST_P macro to define as many parameterized tests
+// for this fixture as you want. The _P suffix is for "parameterized"
+// or "pattern", whichever you prefer to think.
+
+TEST_P(FooTest, DoesBlah) {
+  // Inside a test, access the test parameter with the GetParam() method
+  // of the TestWithParam<T> class:
+  EXPECT_TRUE(foo.Blah(GetParam()));
+  ...
+}
+
+TEST_P(FooTest, HasBlahBlah) {
+  ...
+}
+
+// Finally, you can use INSTANTIATE_TEST_CASE_P to instantiate the test
+// case with any set of parameters you want. Google Test defines a number
+// of functions for generating test parameters. They return what we call
+// (surprise!) parameter generators. Here is a  summary of them, which
+// are all in the testing namespace:
+//
+//
+//  Range(begin, end [, step]) - Yields values {begin, begin+step,
+//                               begin+step+step, ...}. The values do not
+//                               include end. step defaults to 1.
+//  Values(v1, v2, ..., vN)    - Yields values {v1, v2, ..., vN}.
+//  ValuesIn(container)        - Yields values from a C-style array, an STL
+//  ValuesIn(begin,end)          container, or an iterator range [begin, end).
+//  Bool()                     - Yields sequence {false, true}.
+//  Combine(g1, g2, ..., gN)   - Yields all combinations (the Cartesian product
+//                               for the math savvy) of the values generated
+//                               by the N generators.
+//
+// For more details, see comments at the definitions of these functions below
+// in this file.
+//
+// The following statement will instantiate tests from the FooTest test case
+// each with parameter values "meeny", "miny", and "moe".
+
+INSTANTIATE_TEST_CASE_P(InstantiationName,
+                        FooTest,
+                        Values("meeny", "miny", "moe"));
+
+// To distinguish different instances of the pattern, (yes, you
+// can instantiate it more then once) the first argument to the
+// INSTANTIATE_TEST_CASE_P macro is a prefix that will be added to the
+// actual test case name. Remember to pick unique prefixes for different
+// instantiations. The tests from the instantiation above will have
+// these names:
+//
+//    * InstantiationName/FooTest.DoesBlah/0 for "meeny"
+//    * InstantiationName/FooTest.DoesBlah/1 for "miny"
+//    * InstantiationName/FooTest.DoesBlah/2 for "moe"
+//    * InstantiationName/FooTest.HasBlahBlah/0 for "meeny"
+//    * InstantiationName/FooTest.HasBlahBlah/1 for "miny"
+//    * InstantiationName/FooTest.HasBlahBlah/2 for "moe"
+//
+// You can use these names in --gtest_filter.
+//
+// This statement will instantiate all tests from FooTest again, each
+// with parameter values "cat" and "dog":
+
+const char* pets[] = {"cat", "dog"};
+INSTANTIATE_TEST_CASE_P(AnotherInstantiationName, FooTest, ValuesIn(pets));
+
+// The tests from the instantiation above will have these names:
+//
+//    * AnotherInstantiationName/FooTest.DoesBlah/0 for "cat"
+//    * AnotherInstantiationName/FooTest.DoesBlah/1 for "dog"
+//    * AnotherInstantiationName/FooTest.HasBlahBlah/0 for "cat"
+//    * AnotherInstantiationName/FooTest.HasBlahBlah/1 for "dog"
+//
+// Please note that INSTANTIATE_TEST_CASE_P will instantiate all tests
+// in the given test case, whether their definitions come before or
+// AFTER the INSTANTIATE_TEST_CASE_P statement.
+//
+// Please also note that generator expressions (including parameters to the
+// generators) are evaluated in InitGoogleTest(), after main() has started.
+// This allows the user on one hand, to adjust generator parameters in order
+// to dynamically determine a set of tests to run and on the other hand,
+// give the user a chance to inspect the generated tests with Google Test
+// reflection API before RUN_ALL_TESTS() is executed.
+//
+// You can see samples/sample7_unittest.cc and samples/sample8_unittest.cc
+// for more examples.
+//
+// In the future, we plan to publish the API for defining new parameter
+// generators. But for now this interface remains part of the internal
+// implementation and is subject to change.
+//
+//
+// A parameterized test fixture must be derived from testing::Test and from
+// testing::WithParamInterface<T>, where T is the type of the parameter
+// values. Inheriting from TestWithParam<T> satisfies that requirement because
+// TestWithParam<T> inherits from both Test and WithParamInterface. In more
+// complicated hierarchies, however, it is occasionally useful to inherit
+// separately from Test and WithParamInterface. For example:
+
+class BaseTest : public ::testing::Test {
+  // You can inherit all the usual members for a non-parameterized test
+  // fixture here.
+};
+
+class DerivedTest : public BaseTest, public ::testing::WithParamInterface<int> {
+  // The usual test fixture members go here too.
+};
+
+TEST_F(BaseTest, HasFoo) {
+  // This is an ordinary non-parameterized test.
+}
+
+TEST_P(DerivedTest, DoesBlah) {
+  // GetParam works just the same here as if you inherit from TestWithParam.
+  EXPECT_TRUE(foo.Blah(GetParam()));
+}
+
+#endif  // 0
+
+
+#if !GTEST_OS_SYMBIAN
+# include <utility>
+#endif
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: vladl@google.com (Vlad Losev)
+
+// Type and function utilities for implementing parameterized tests.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_H_
+
+#include <iterator>
+#include <utility>
+#include <vector>
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+// Copyright 2003 Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: Dan Egnor (egnor@google.com)
+//
+// A "smart" pointer type with reference tracking.  Every pointer to a
+// particular object is kept on a circular linked list.  When the last pointer
+// to an object is destroyed or reassigned, the object is deleted.
+//
+// Used properly, this deletes the object when the last reference goes away.
+// There are several caveats:
+// - Like all reference counting schemes, cycles lead to leaks.
+// - Each smart pointer is actually two pointers (8 bytes instead of 4).
+// - Every time a pointer is assigned, the entire list of pointers to that
+//   object is traversed.  This class is therefore NOT SUITABLE when there
+//   will often be more than two or three pointers to a particular object.
+// - References are only tracked as long as linked_ptr<> objects are copied.
+//   If a linked_ptr<> is converted to a raw pointer and back, BAD THINGS
+//   will happen (double deletion).
+//
+// A good use of this class is storing object references in STL containers.
+// You can safely put linked_ptr<> in a vector<>.
+// Other uses may not be as good.
+//
+// Note: If you use an incomplete type with linked_ptr<>, the class
+// *containing* linked_ptr<> must have a constructor and destructor (even
+// if they do nothing!).
+//
+// Bill Gibbons suggested we use something like this.
+//
+// Thread Safety:
+//   Unlike other linked_ptr implementations, in this implementation
+//   a linked_ptr object is thread-safe in the sense that:
+//     - it's safe to copy linked_ptr objects concurrently,
+//     - it's safe to copy *from* a linked_ptr and read its underlying
+//       raw pointer (e.g. via get()) concurrently, and
+//     - it's safe to write to two linked_ptrs that point to the same
+//       shared object concurrently.
+// TODO(wan@google.com): rename this to safe_linked_ptr to avoid
+// confusion with normal linked_ptr.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_LINKED_PTR_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_LINKED_PTR_H_
+
+#include <stdlib.h>
+#include <assert.h>
+
+
+namespace testing {
+namespace internal {
+
+// Protects copying of all linked_ptr objects.
+GTEST_API_ GTEST_DECLARE_STATIC_MUTEX_(g_linked_ptr_mutex);
+
+// This is used internally by all instances of linked_ptr<>.  It needs to be
+// a non-template class because different types of linked_ptr<> can refer to
+// the same object (linked_ptr<Superclass>(obj) vs linked_ptr<Subclass>(obj)).
+// So, it needs to be possible for different types of linked_ptr to participate
+// in the same circular linked list, so we need a single class type here.
+//
+// DO NOT USE THIS CLASS DIRECTLY YOURSELF.  Use linked_ptr<T>.
+class linked_ptr_internal {
+ public:
+  // Create a new circle that includes only this instance.
+  void join_new() {
+    next_ = this;
+  }
+
+  // Many linked_ptr operations may change p.link_ for some linked_ptr
+  // variable p in the same circle as this object.  Therefore we need
+  // to prevent two such operations from occurring concurrently.
+  //
+  // Note that different types of linked_ptr objects can coexist in a
+  // circle (e.g. linked_ptr<Base>, linked_ptr<Derived1>, and
+  // linked_ptr<Derived2>).  Therefore we must use a single mutex to
+  // protect all linked_ptr objects.  This can create serious
+  // contention in production code, but is acceptable in a testing
+  // framework.
+
+  // Join an existing circle.
+  void join(linked_ptr_internal const* ptr)
+      GTEST_LOCK_EXCLUDED_(g_linked_ptr_mutex) {
+    MutexLock lock(&g_linked_ptr_mutex);
+
+    linked_ptr_internal const* p = ptr;
+    while (p->next_ != ptr) p = p->next_;
+    p->next_ = this;
+    next_ = ptr;
+  }
+
+  // Leave whatever circle we're part of.  Returns true if we were the
+  // last member of the circle.  Once this is done, you can join() another.
+  bool depart()
+      GTEST_LOCK_EXCLUDED_(g_linked_ptr_mutex) {
+    MutexLock lock(&g_linked_ptr_mutex);
+
+    if (next_ == this) return true;
+    linked_ptr_internal const* p = next_;
+    while (p->next_ != this) p = p->next_;
+    p->next_ = next_;
+    return false;
+  }
+
+ private:
+  mutable linked_ptr_internal const* next_;
+};
+
+template <typename T>
+class linked_ptr {
+ public:
+  typedef T element_type;
+
+  // Take over ownership of a raw pointer.  This should happen as soon as
+  // possible after the object is created.
+  explicit linked_ptr(T* ptr = NULL) { capture(ptr); }
+  ~linked_ptr() { depart(); }
+
+  // Copy an existing linked_ptr<>, adding ourselves to the list of references.
+  template <typename U> linked_ptr(linked_ptr<U> const& ptr) { copy(&ptr); }
+  linked_ptr(linked_ptr const& ptr) {  // NOLINT
+    assert(&ptr != this);
+    copy(&ptr);
+  }
+
+  // Assignment releases the old value and acquires the new.
+  template <typename U> linked_ptr& operator=(linked_ptr<U> const& ptr) {
+    depart();
+    copy(&ptr);
+    return *this;
+  }
+
+  linked_ptr& operator=(linked_ptr const& ptr) {
+    if (&ptr != this) {
+      depart();
+      copy(&ptr);
+    }
+    return *this;
+  }
+
+  // Smart pointer members.
+  void reset(T* ptr = NULL) {
+    depart();
+    capture(ptr);
+  }
+  T* get() const { return value_; }
+  T* operator->() const { return value_; }
+  T& operator*() const { return *value_; }
+
+  bool operator==(T* p) const { return value_ == p; }
+  bool operator!=(T* p) const { return value_ != p; }
+  template <typename U>
+  bool operator==(linked_ptr<U> const& ptr) const {
+    return value_ == ptr.get();
+  }
+  template <typename U>
+  bool operator!=(linked_ptr<U> const& ptr) const {
+    return value_ != ptr.get();
+  }
+
+ private:
+  template <typename U>
+  friend class linked_ptr;
+
+  T* value_;
+  linked_ptr_internal link_;
+
+  void depart() {
+    if (link_.depart()) delete value_;
+  }
+
+  void capture(T* ptr) {
+    value_ = ptr;
+    link_.join_new();
+  }
+
+  template <typename U> void copy(linked_ptr<U> const* ptr) {
+    value_ = ptr->get();
+    if (value_)
+      link_.join(&ptr->link_);
+    else
+      link_.join_new();
+  }
+};
+
+template<typename T> inline
+bool operator==(T* ptr, const linked_ptr<T>& x) {
+  return ptr == x.get();
+}
+
+template<typename T> inline
+bool operator!=(T* ptr, const linked_ptr<T>& x) {
+  return ptr != x.get();
+}
+
+// A function to convert T* into linked_ptr<T>
+// Doing e.g. make_linked_ptr(new FooBarBaz<type>(arg)) is a shorter notation
+// for linked_ptr<FooBarBaz<type> >(new FooBarBaz<type>(arg))
+template <typename T>
+linked_ptr<T> make_linked_ptr(T* ptr) {
+  return linked_ptr<T>(ptr);
+}
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_LINKED_PTR_H_
+// Copyright 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Google Test - The Google C++ Testing Framework
+//
+// This file implements a universal value printer that can print a
+// value of any type T:
+//
+//   void ::testing::internal::UniversalPrinter<T>::Print(value, ostream_ptr);
+//
+// A user can teach this function how to print a class type T by
+// defining either operator<<() or PrintTo() in the namespace that
+// defines T.  More specifically, the FIRST defined function in the
+// following list will be used (assuming T is defined in namespace
+// foo):
+//
+//   1. foo::PrintTo(const T&, ostream*)
+//   2. operator<<(ostream&, const T&) defined in either foo or the
+//      global namespace.
+//
+// If none of the above is defined, it will print the debug string of
+// the value if it is a protocol buffer, or print the raw bytes in the
+// value otherwise.
+//
+// To aid debugging: when T is a reference type, the address of the
+// value is also printed; when T is a (const) char pointer, both the
+// pointer value and the NUL-terminated string it points to are
+// printed.
+//
+// We also provide some convenient wrappers:
+//
+//   // Prints a value to a string.  For a (const or not) char
+//   // pointer, the NUL-terminated string (but not the pointer) is
+//   // printed.
+//   std::string ::testing::PrintToString(const T& value);
+//
+//   // Prints a value tersely: for a reference type, the referenced
+//   // value (but not the address) is printed; for a (const or not) char
+//   // pointer, the NUL-terminated string (but not the pointer) is
+//   // printed.
+//   void ::testing::internal::UniversalTersePrint(const T& value, ostream*);
+//
+//   // Prints value using the type inferred by the compiler.  The difference
+//   // from UniversalTersePrint() is that this function prints both the
+//   // pointer and the NUL-terminated string for a (const or not) char pointer.
+//   void ::testing::internal::UniversalPrint(const T& value, ostream*);
+//
+//   // Prints the fields of a tuple tersely to a string vector, one
+//   // element for each field. Tuple support must be enabled in
+//   // gtest-port.h.
+//   std::vector<string> UniversalTersePrintTupleFieldsToStrings(
+//       const Tuple& value);
+//
+// Known limitation:
+//
+// The print primitives print the elements of an STL-style container
+// using the compiler-inferred type of *iter where iter is a
+// const_iterator of the container.  When const_iterator is an input
+// iterator but not a forward iterator, this inferred type may not
+// match value_type, and the print output may be incorrect.  In
+// practice, this is rarely a problem as for most containers
+// const_iterator is a forward iterator.  We'll fix this if there's an
+// actual need for it.  Note that this fix cannot rely on value_type
+// being defined as many user-defined container types don't have
+// value_type.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PRINTERS_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PRINTERS_H_
+
+#include <ostream>  // NOLINT
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+
+namespace testing {
+
+// Definitions in the 'internal' and 'internal2' name spaces are
+// subject to change without notice.  DO NOT USE THEM IN USER CODE!
+namespace internal2 {
+
+// Prints the given number of bytes in the given object to the given
+// ostream.
+GTEST_API_ void PrintBytesInObjectTo(const unsigned char* obj_bytes,
+                                     size_t count,
+                                     ::std::ostream* os);
+
+// For selecting which printer to use when a given type has neither <<
+// nor PrintTo().
+enum TypeKind {
+  kProtobuf,              // a protobuf type
+  kConvertibleToInteger,  // a type implicitly convertible to BiggestInt
+                          // (e.g. a named or unnamed enum type)
+  kOtherType              // anything else
+};
+
+// TypeWithoutFormatter<T, kTypeKind>::PrintValue(value, os) is called
+// by the universal printer to print a value of type T when neither
+// operator<< nor PrintTo() is defined for T, where kTypeKind is the
+// "kind" of T as defined by enum TypeKind.
+template <typename T, TypeKind kTypeKind>
+class TypeWithoutFormatter {
+ public:
+  // This default version is called when kTypeKind is kOtherType.
+  static void PrintValue(const T& value, ::std::ostream* os) {
+    PrintBytesInObjectTo(reinterpret_cast<const unsigned char*>(&value),
+                         sizeof(value), os);
+  }
+};
+
+// We print a protobuf using its ShortDebugString() when the string
+// doesn't exceed this many characters; otherwise we print it using
+// DebugString() for better readability.
+const size_t kProtobufOneLinerMaxLength = 50;
+
+template <typename T>
+class TypeWithoutFormatter<T, kProtobuf> {
+ public:
+  static void PrintValue(const T& value, ::std::ostream* os) {
+    const ::testing::internal::string short_str = value.ShortDebugString();
+    const ::testing::internal::string pretty_str =
+        short_str.length() <= kProtobufOneLinerMaxLength ?
+        short_str : ("\n" + value.DebugString());
+    *os << ("<" + pretty_str + ">");
+  }
+};
+
+template <typename T>
+class TypeWithoutFormatter<T, kConvertibleToInteger> {
+ public:
+  // Since T has no << operator or PrintTo() but can be implicitly
+  // converted to BiggestInt, we print it as a BiggestInt.
+  //
+  // Most likely T is an enum type (either named or unnamed), in which
+  // case printing it as an integer is the desired behavior.  In case
+  // T is not an enum, printing it as an integer is the best we can do
+  // given that it has no user-defined printer.
+  static void PrintValue(const T& value, ::std::ostream* os) {
+    const internal::BiggestInt kBigInt = value;
+    *os << kBigInt;
+  }
+};
+
+// Prints the given value to the given ostream.  If the value is a
+// protocol message, its debug string is printed; if it's an enum or
+// of a type implicitly convertible to BiggestInt, it's printed as an
+// integer; otherwise the bytes in the value are printed.  This is
+// what UniversalPrinter<T>::Print() does when it knows nothing about
+// type T and T has neither << operator nor PrintTo().
+//
+// A user can override this behavior for a class type Foo by defining
+// a << operator in the namespace where Foo is defined.
+//
+// We put this operator in namespace 'internal2' instead of 'internal'
+// to simplify the implementation, as much code in 'internal' needs to
+// use << in STL, which would conflict with our own << were it defined
+// in 'internal'.
+//
+// Note that this operator<< takes a generic std::basic_ostream<Char,
+// CharTraits> type instead of the more restricted std::ostream.  If
+// we define it to take an std::ostream instead, we'll get an
+// "ambiguous overloads" compiler error when trying to print a type
+// Foo that supports streaming to std::basic_ostream<Char,
+// CharTraits>, as the compiler cannot tell whether
+// operator<<(std::ostream&, const T&) or
+// operator<<(std::basic_stream<Char, CharTraits>, const Foo&) is more
+// specific.
+template <typename Char, typename CharTraits, typename T>
+::std::basic_ostream<Char, CharTraits>& operator<<(
+    ::std::basic_ostream<Char, CharTraits>& os, const T& x) {
+  TypeWithoutFormatter<T,
+      (internal::IsAProtocolMessage<T>::value ? kProtobuf :
+       internal::ImplicitlyConvertible<const T&, internal::BiggestInt>::value ?
+       kConvertibleToInteger : kOtherType)>::PrintValue(x, &os);
+  return os;
+}
+
+}  // namespace internal2
+}  // namespace testing
+
+// This namespace MUST NOT BE NESTED IN ::testing, or the name look-up
+// magic needed for implementing UniversalPrinter won't work.
+namespace testing_internal {
+
+// Used to print a value that is not an STL-style container when the
+// user doesn't define PrintTo() for it.
+template <typename T>
+void DefaultPrintNonContainerTo(const T& value, ::std::ostream* os) {
+  // With the following statement, during unqualified name lookup,
+  // testing::internal2::operator<< appears as if it was declared in
+  // the nearest enclosing namespace that contains both
+  // ::testing_internal and ::testing::internal2, i.e. the global
+  // namespace.  For more details, refer to the C++ Standard section
+  // 7.3.4-1 [namespace.udir].  This allows us to fall back onto
+  // testing::internal2::operator<< in case T doesn't come with a <<
+  // operator.
+  //
+  // We cannot write 'using ::testing::internal2::operator<<;', which
+  // gcc 3.3 fails to compile due to a compiler bug.
+  using namespace ::testing::internal2;  // NOLINT
+
+  // Assuming T is defined in namespace foo, in the next statement,
+  // the compiler will consider all of:
+  //
+  //   1. foo::operator<< (thanks to Koenig look-up),
+  //   2. ::operator<< (as the current namespace is enclosed in ::),
+  //   3. testing::internal2::operator<< (thanks to the using statement above).
+  //
+  // The operator<< whose type matches T best will be picked.
+  //
+  // We deliberately allow #2 to be a candidate, as sometimes it's
+  // impossible to define #1 (e.g. when foo is ::std, defining
+  // anything in it is undefined behavior unless you are a compiler
+  // vendor.).
+  *os << value;
+}
+
+}  // namespace testing_internal
+
+namespace testing {
+namespace internal {
+
+// UniversalPrinter<T>::Print(value, ostream_ptr) prints the given
+// value to the given ostream.  The caller must ensure that
+// 'ostream_ptr' is not NULL, or the behavior is undefined.
+//
+// We define UniversalPrinter as a class template (as opposed to a
+// function template), as we need to partially specialize it for
+// reference types, which cannot be done with function templates.
+template <typename T>
+class UniversalPrinter;
+
+template <typename T>
+void UniversalPrint(const T& value, ::std::ostream* os);
+
+// Used to print an STL-style container when the user doesn't define
+// a PrintTo() for it.
+template <typename C>
+void DefaultPrintTo(IsContainer /* dummy */,
+                    false_type /* is not a pointer */,
+                    const C& container, ::std::ostream* os) {
+  const size_t kMaxCount = 32;  // The maximum number of elements to print.
+  *os << '{';
+  size_t count = 0;
+  for (typename C::const_iterator it = container.begin();
+       it != container.end(); ++it, ++count) {
+    if (count > 0) {
+      *os << ',';
+      if (count == kMaxCount) {  // Enough has been printed.
+        *os << " ...";
+        break;
+      }
+    }
+    *os << ' ';
+    // We cannot call PrintTo(*it, os) here as PrintTo() doesn't
+    // handle *it being a native array.
+    internal::UniversalPrint(*it, os);
+  }
+
+  if (count > 0) {
+    *os << ' ';
+  }
+  *os << '}';
+}
+
+// Used to print a pointer that is neither a char pointer nor a member
+// pointer, when the user doesn't define PrintTo() for it.  (A member
+// variable pointer or member function pointer doesn't really point to
+// a location in the address space.  Their representation is
+// implementation-defined.  Therefore they will be printed as raw
+// bytes.)
+template <typename T>
+void DefaultPrintTo(IsNotContainer /* dummy */,
+                    true_type /* is a pointer */,
+                    T* p, ::std::ostream* os) {
+  if (p == NULL) {
+    *os << "NULL";
+  } else {
+    // C++ doesn't allow casting from a function pointer to any object
+    // pointer.
+    //
+    // IsTrue() silences warnings: "Condition is always true",
+    // "unreachable code".
+    if (IsTrue(ImplicitlyConvertible<T*, const void*>::value)) {
+      // T is not a function type.  We just call << to print p,
+      // relying on ADL to pick up user-defined << for their pointer
+      // types, if any.
+      *os << p;
+    } else {
+      // T is a function type, so '*os << p' doesn't do what we want
+      // (it just prints p as bool).  We want to print p as a const
+      // void*.  However, we cannot cast it to const void* directly,
+      // even using reinterpret_cast, as earlier versions of gcc
+      // (e.g. 3.4.5) cannot compile the cast when p is a function
+      // pointer.  Casting to UInt64 first solves the problem.
+      *os << reinterpret_cast<const void*>(
+          reinterpret_cast<internal::UInt64>(p));
+    }
+  }
+}
+
+// Used to print a non-container, non-pointer value when the user
+// doesn't define PrintTo() for it.
+template <typename T>
+void DefaultPrintTo(IsNotContainer /* dummy */,
+                    false_type /* is not a pointer */,
+                    const T& value, ::std::ostream* os) {
+  ::testing_internal::DefaultPrintNonContainerTo(value, os);
+}
+
+// Prints the given value using the << operator if it has one;
+// otherwise prints the bytes in it.  This is what
+// UniversalPrinter<T>::Print() does when PrintTo() is not specialized
+// or overloaded for type T.
+//
+// A user can override this behavior for a class type Foo by defining
+// an overload of PrintTo() in the namespace where Foo is defined.  We
+// give the user this option as sometimes defining a << operator for
+// Foo is not desirable (e.g. the coding style may prevent doing it,
+// or there is already a << operator but it doesn't do what the user
+// wants).
+template <typename T>
+void PrintTo(const T& value, ::std::ostream* os) {
+  // DefaultPrintTo() is overloaded.  The type of its first two
+  // arguments determine which version will be picked.  If T is an
+  // STL-style container, the version for container will be called; if
+  // T is a pointer, the pointer version will be called; otherwise the
+  // generic version will be called.
+  //
+  // Note that we check for container types here, prior to we check
+  // for protocol message types in our operator<<.  The rationale is:
+  //
+  // For protocol messages, we want to give people a chance to
+  // override Google Mock's format by defining a PrintTo() or
+  // operator<<.  For STL containers, other formats can be
+  // incompatible with Google Mock's format for the container
+  // elements; therefore we check for container types here to ensure
+  // that our format is used.
+  //
+  // The second argument of DefaultPrintTo() is needed to bypass a bug
+  // in Symbian's C++ compiler that prevents it from picking the right
+  // overload between:
+  //
+  //   PrintTo(const T& x, ...);
+  //   PrintTo(T* x, ...);
+  DefaultPrintTo(IsContainerTest<T>(0), is_pointer<T>(), value, os);
+}
+
+// The following list of PrintTo() overloads tells
+// UniversalPrinter<T>::Print() how to print standard types (built-in
+// types, strings, plain arrays, and pointers).
+
+// Overloads for various char types.
+GTEST_API_ void PrintTo(unsigned char c, ::std::ostream* os);
+GTEST_API_ void PrintTo(signed char c, ::std::ostream* os);
+inline void PrintTo(char c, ::std::ostream* os) {
+  // When printing a plain char, we always treat it as unsigned.  This
+  // way, the output won't be affected by whether the compiler thinks
+  // char is signed or not.
+  PrintTo(static_cast<unsigned char>(c), os);
+}
+
+// Overloads for other simple built-in types.
+inline void PrintTo(bool x, ::std::ostream* os) {
+  *os << (x ? "true" : "false");
+}
+
+// Overload for wchar_t type.
+// Prints a wchar_t as a symbol if it is printable or as its internal
+// code otherwise and also as its decimal code (except for L'\0').
+// The L'\0' char is printed as "L'\\0'". The decimal code is printed
+// as signed integer when wchar_t is implemented by the compiler
+// as a signed type and is printed as an unsigned integer when wchar_t
+// is implemented as an unsigned type.
+GTEST_API_ void PrintTo(wchar_t wc, ::std::ostream* os);
+
+// Overloads for C strings.
+GTEST_API_ void PrintTo(const char* s, ::std::ostream* os);
+inline void PrintTo(char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const char*>(s), os);
+}
+
+// signed/unsigned char is often used for representing binary data, so
+// we print pointers to it as void* to be safe.
+inline void PrintTo(const signed char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const void*>(s), os);
+}
+inline void PrintTo(signed char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const void*>(s), os);
+}
+inline void PrintTo(const unsigned char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const void*>(s), os);
+}
+inline void PrintTo(unsigned char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const void*>(s), os);
+}
+
+// MSVC can be configured to define wchar_t as a typedef of unsigned
+// short.  It defines _NATIVE_WCHAR_T_DEFINED when wchar_t is a native
+// type.  When wchar_t is a typedef, defining an overload for const
+// wchar_t* would cause unsigned short* be printed as a wide string,
+// possibly causing invalid memory accesses.
+#if !defined(_MSC_VER) || defined(_NATIVE_WCHAR_T_DEFINED)
+// Overloads for wide C strings
+GTEST_API_ void PrintTo(const wchar_t* s, ::std::ostream* os);
+inline void PrintTo(wchar_t* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const wchar_t*>(s), os);
+}
+#endif
+
+// Overload for C arrays.  Multi-dimensional arrays are printed
+// properly.
+
+// Prints the given number of elements in an array, without printing
+// the curly braces.
+template <typename T>
+void PrintRawArrayTo(const T a[], size_t count, ::std::ostream* os) {
+  UniversalPrint(a[0], os);
+  for (size_t i = 1; i != count; i++) {
+    *os << ", ";
+    UniversalPrint(a[i], os);
+  }
+}
+
+// Overloads for ::string and ::std::string.
+#if GTEST_HAS_GLOBAL_STRING
+GTEST_API_ void PrintStringTo(const ::string&s, ::std::ostream* os);
+inline void PrintTo(const ::string& s, ::std::ostream* os) {
+  PrintStringTo(s, os);
+}
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+GTEST_API_ void PrintStringTo(const ::std::string&s, ::std::ostream* os);
+inline void PrintTo(const ::std::string& s, ::std::ostream* os) {
+  PrintStringTo(s, os);
+}
+
+// Overloads for ::wstring and ::std::wstring.
+#if GTEST_HAS_GLOBAL_WSTRING
+GTEST_API_ void PrintWideStringTo(const ::wstring&s, ::std::ostream* os);
+inline void PrintTo(const ::wstring& s, ::std::ostream* os) {
+  PrintWideStringTo(s, os);
+}
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+#if GTEST_HAS_STD_WSTRING
+GTEST_API_ void PrintWideStringTo(const ::std::wstring&s, ::std::ostream* os);
+inline void PrintTo(const ::std::wstring& s, ::std::ostream* os) {
+  PrintWideStringTo(s, os);
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+#if GTEST_HAS_TR1_TUPLE
+// Overload for ::std::tr1::tuple.  Needed for printing function arguments,
+// which are packed as tuples.
+
+// Helper function for printing a tuple.  T must be instantiated with
+// a tuple type.
+template <typename T>
+void PrintTupleTo(const T& t, ::std::ostream* os);
+
+// Overloaded PrintTo() for tuples of various arities.  We support
+// tuples of up-to 10 fields.  The following implementation works
+// regardless of whether tr1::tuple is implemented using the
+// non-standard variadic template feature or not.
+
+inline void PrintTo(const ::std::tr1::tuple<>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1>
+void PrintTo(const ::std::tr1::tuple<T1>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2>
+void PrintTo(const ::std::tr1::tuple<T1, T2>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4, T5>& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+          typename T6>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4, T5, T6>& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+          typename T6, typename T7>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7>& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+          typename T6, typename T7, typename T8>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8>& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+          typename T6, typename T7, typename T8, typename T9>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8, T9>& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+          typename T6, typename T7, typename T8, typename T9, typename T10>
+void PrintTo(
+    const ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10>& t,
+    ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+#endif  // GTEST_HAS_TR1_TUPLE
+
+// Overload for std::pair.
+template <typename T1, typename T2>
+void PrintTo(const ::std::pair<T1, T2>& value, ::std::ostream* os) {
+  *os << '(';
+  // We cannot use UniversalPrint(value.first, os) here, as T1 may be
+  // a reference type.  The same for printing value.second.
+  UniversalPrinter<T1>::Print(value.first, os);
+  *os << ", ";
+  UniversalPrinter<T2>::Print(value.second, os);
+  *os << ')';
+}
+
+// Implements printing a non-reference type T by letting the compiler
+// pick the right overload of PrintTo() for T.
+template <typename T>
+class UniversalPrinter {
+ public:
+  // MSVC warns about adding const to a function type, so we want to
+  // disable the warning.
+#ifdef _MSC_VER
+# pragma warning(push)          // Saves the current warning state.
+# pragma warning(disable:4180)  // Temporarily disables warning 4180.
+#endif  // _MSC_VER
+
+  // Note: we deliberately don't call this PrintTo(), as that name
+  // conflicts with ::testing::internal::PrintTo in the body of the
+  // function.
+  static void Print(const T& value, ::std::ostream* os) {
+    // By default, ::testing::internal::PrintTo() is used for printing
+    // the value.
+    //
+    // Thanks to Koenig look-up, if T is a class and has its own
+    // PrintTo() function defined in its namespace, that function will
+    // be visible here.  Since it is more specific than the generic ones
+    // in ::testing::internal, it will be picked by the compiler in the
+    // following statement - exactly what we want.
+    PrintTo(value, os);
+  }
+
+#ifdef _MSC_VER
+# pragma warning(pop)           // Restores the warning state.
+#endif  // _MSC_VER
+};
+
+// UniversalPrintArray(begin, len, os) prints an array of 'len'
+// elements, starting at address 'begin'.
+template <typename T>
+void UniversalPrintArray(const T* begin, size_t len, ::std::ostream* os) {
+  if (len == 0) {
+    *os << "{}";
+  } else {
+    *os << "{ ";
+    const size_t kThreshold = 18;
+    const size_t kChunkSize = 8;
+    // If the array has more than kThreshold elements, we'll have to
+    // omit some details by printing only the first and the last
+    // kChunkSize elements.
+    // TODO(wan@google.com): let the user control the threshold using a flag.
+    if (len <= kThreshold) {
+      PrintRawArrayTo(begin, len, os);
+    } else {
+      PrintRawArrayTo(begin, kChunkSize, os);
+      *os << ", ..., ";
+      PrintRawArrayTo(begin + len - kChunkSize, kChunkSize, os);
+    }
+    *os << " }";
+  }
+}
+// This overload prints a (const) char array compactly.
+GTEST_API_ void UniversalPrintArray(
+    const char* begin, size_t len, ::std::ostream* os);
+
+// This overload prints a (const) wchar_t array compactly.
+GTEST_API_ void UniversalPrintArray(
+    const wchar_t* begin, size_t len, ::std::ostream* os);
+
+// Implements printing an array type T[N].
+template <typename T, size_t N>
+class UniversalPrinter<T[N]> {
+ public:
+  // Prints the given array, omitting some elements when there are too
+  // many.
+  static void Print(const T (&a)[N], ::std::ostream* os) {
+    UniversalPrintArray(a, N, os);
+  }
+};
+
+// Implements printing a reference type T&.
+template <typename T>
+class UniversalPrinter<T&> {
+ public:
+  // MSVC warns about adding const to a function type, so we want to
+  // disable the warning.
+#ifdef _MSC_VER
+# pragma warning(push)          // Saves the current warning state.
+# pragma warning(disable:4180)  // Temporarily disables warning 4180.
+#endif  // _MSC_VER
+
+  static void Print(const T& value, ::std::ostream* os) {
+    // Prints the address of the value.  We use reinterpret_cast here
+    // as static_cast doesn't compile when T is a function type.
+    *os << "@" << reinterpret_cast<const void*>(&value) << " ";
+
+    // Then prints the value itself.
+    UniversalPrint(value, os);
+  }
+
+#ifdef _MSC_VER
+# pragma warning(pop)           // Restores the warning state.
+#endif  // _MSC_VER
+};
+
+// Prints a value tersely: for a reference type, the referenced value
+// (but not the address) is printed; for a (const) char pointer, the
+// NUL-terminated string (but not the pointer) is printed.
+
+template <typename T>
+class UniversalTersePrinter {
+ public:
+  static void Print(const T& value, ::std::ostream* os) {
+    UniversalPrint(value, os);
+  }
+};
+template <typename T>
+class UniversalTersePrinter<T&> {
+ public:
+  static void Print(const T& value, ::std::ostream* os) {
+    UniversalPrint(value, os);
+  }
+};
+template <typename T, size_t N>
+class UniversalTersePrinter<T[N]> {
+ public:
+  static void Print(const T (&value)[N], ::std::ostream* os) {
+    UniversalPrinter<T[N]>::Print(value, os);
+  }
+};
+template <>
+class UniversalTersePrinter<const char*> {
+ public:
+  static void Print(const char* str, ::std::ostream* os) {
+    if (str == NULL) {
+      *os << "NULL";
+    } else {
+      UniversalPrint(string(str), os);
+    }
+  }
+};
+template <>
+class UniversalTersePrinter<char*> {
+ public:
+  static void Print(char* str, ::std::ostream* os) {
+    UniversalTersePrinter<const char*>::Print(str, os);
+  }
+};
+
+#if GTEST_HAS_STD_WSTRING
+template <>
+class UniversalTersePrinter<const wchar_t*> {
+ public:
+  static void Print(const wchar_t* str, ::std::ostream* os) {
+    if (str == NULL) {
+      *os << "NULL";
+    } else {
+      UniversalPrint(::std::wstring(str), os);
+    }
+  }
+};
+#endif
+
+template <>
+class UniversalTersePrinter<wchar_t*> {
+ public:
+  static void Print(wchar_t* str, ::std::ostream* os) {
+    UniversalTersePrinter<const wchar_t*>::Print(str, os);
+  }
+};
+
+template <typename T>
+void UniversalTersePrint(const T& value, ::std::ostream* os) {
+  UniversalTersePrinter<T>::Print(value, os);
+}
+
+// Prints a value using the type inferred by the compiler.  The
+// difference between this and UniversalTersePrint() is that for a
+// (const) char pointer, this prints both the pointer and the
+// NUL-terminated string.
+template <typename T>
+void UniversalPrint(const T& value, ::std::ostream* os) {
+  // A workarond for the bug in VC++ 7.1 that prevents us from instantiating
+  // UniversalPrinter with T directly.
+  typedef T T1;
+  UniversalPrinter<T1>::Print(value, os);
+}
+
+#if GTEST_HAS_TR1_TUPLE
+typedef ::std::vector<string> Strings;
+
+// This helper template allows PrintTo() for tuples and
+// UniversalTersePrintTupleFieldsToStrings() to be defined by
+// induction on the number of tuple fields.  The idea is that
+// TuplePrefixPrinter<N>::PrintPrefixTo(t, os) prints the first N
+// fields in tuple t, and can be defined in terms of
+// TuplePrefixPrinter<N - 1>.
+
+// The inductive case.
+template <size_t N>
+struct TuplePrefixPrinter {
+  // Prints the first N fields of a tuple.
+  template <typename Tuple>
+  static void PrintPrefixTo(const Tuple& t, ::std::ostream* os) {
+    TuplePrefixPrinter<N - 1>::PrintPrefixTo(t, os);
+    *os << ", ";
+    UniversalPrinter<typename ::std::tr1::tuple_element<N - 1, Tuple>::type>
+        ::Print(::std::tr1::get<N - 1>(t), os);
+  }
+
+  // Tersely prints the first N fields of a tuple to a string vector,
+  // one element for each field.
+  template <typename Tuple>
+  static void TersePrintPrefixToStrings(const Tuple& t, Strings* strings) {
+    TuplePrefixPrinter<N - 1>::TersePrintPrefixToStrings(t, strings);
+    ::std::stringstream ss;
+    UniversalTersePrint(::std::tr1::get<N - 1>(t), &ss);
+    strings->push_back(ss.str());
+  }
+};
+
+// Base cases.
+template <>
+struct TuplePrefixPrinter<0> {
+  template <typename Tuple>
+  static void PrintPrefixTo(const Tuple&, ::std::ostream*) {}
+
+  template <typename Tuple>
+  static void TersePrintPrefixToStrings(const Tuple&, Strings*) {}
+};
+// We have to specialize the entire TuplePrefixPrinter<> class
+// template here, even though the definition of
+// TersePrintPrefixToStrings() is the same as the generic version, as
+// Embarcadero (formerly CodeGear, formerly Borland) C++ doesn't
+// support specializing a method template of a class template.
+template <>
+struct TuplePrefixPrinter<1> {
+  template <typename Tuple>
+  static void PrintPrefixTo(const Tuple& t, ::std::ostream* os) {
+    UniversalPrinter<typename ::std::tr1::tuple_element<0, Tuple>::type>::
+        Print(::std::tr1::get<0>(t), os);
+  }
+
+  template <typename Tuple>
+  static void TersePrintPrefixToStrings(const Tuple& t, Strings* strings) {
+    ::std::stringstream ss;
+    UniversalTersePrint(::std::tr1::get<0>(t), &ss);
+    strings->push_back(ss.str());
+  }
+};
+
+// Helper function for printing a tuple.  T must be instantiated with
+// a tuple type.
+template <typename T>
+void PrintTupleTo(const T& t, ::std::ostream* os) {
+  *os << "(";
+  TuplePrefixPrinter< ::std::tr1::tuple_size<T>::value>::
+      PrintPrefixTo(t, os);
+  *os << ")";
+}
+
+// Prints the fields of a tuple tersely to a string vector, one
+// element for each field.  See the comment before
+// UniversalTersePrint() for how we define "tersely".
+template <typename Tuple>
+Strings UniversalTersePrintTupleFieldsToStrings(const Tuple& value) {
+  Strings result;
+  TuplePrefixPrinter< ::std::tr1::tuple_size<Tuple>::value>::
+      TersePrintPrefixToStrings(value, &result);
+  return result;
+}
+#endif  // GTEST_HAS_TR1_TUPLE
+
+}  // namespace internal
+
+template <typename T>
+::std::string PrintToString(const T& value) {
+  ::std::stringstream ss;
+  internal::UniversalTersePrinter<T>::Print(value, &ss);
+  return ss.str();
+}
+
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PRINTERS_H_
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+namespace internal {
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Outputs a message explaining invalid registration of different
+// fixture class for the same test case. This may happen when
+// TEST_P macro is used to define two tests with the same name
+// but in different namespaces.
+GTEST_API_ void ReportInvalidTestCaseType(const char* test_case_name,
+                                          const char* file, int line);
+
+template <typename> class ParamGeneratorInterface;
+template <typename> class ParamGenerator;
+
+// Interface for iterating over elements provided by an implementation
+// of ParamGeneratorInterface<T>.
+template <typename T>
+class ParamIteratorInterface {
+ public:
+  virtual ~ParamIteratorInterface() {}
+  // A pointer to the base generator instance.
+  // Used only for the purposes of iterator comparison
+  // to make sure that two iterators belong to the same generator.
+  virtual const ParamGeneratorInterface<T>* BaseGenerator() const = 0;
+  // Advances iterator to point to the next element
+  // provided by the generator. The caller is responsible
+  // for not calling Advance() on an iterator equal to
+  // BaseGenerator()->End().
+  virtual void Advance() = 0;
+  // Clones the iterator object. Used for implementing copy semantics
+  // of ParamIterator<T>.
+  virtual ParamIteratorInterface* Clone() const = 0;
+  // Dereferences the current iterator and provides (read-only) access
+  // to the pointed value. It is the caller's responsibility not to call
+  // Current() on an iterator equal to BaseGenerator()->End().
+  // Used for implementing ParamGenerator<T>::operator*().
+  virtual const T* Current() const = 0;
+  // Determines whether the given iterator and other point to the same
+  // element in the sequence generated by the generator.
+  // Used for implementing ParamGenerator<T>::operator==().
+  virtual bool Equals(const ParamIteratorInterface& other) const = 0;
+};
+
+// Class iterating over elements provided by an implementation of
+// ParamGeneratorInterface<T>. It wraps ParamIteratorInterface<T>
+// and implements the const forward iterator concept.
+template <typename T>
+class ParamIterator {
+ public:
+  typedef T value_type;
+  typedef const T& reference;
+  typedef ptrdiff_t difference_type;
+
+  // ParamIterator assumes ownership of the impl_ pointer.
+  ParamIterator(const ParamIterator& other) : impl_(other.impl_->Clone()) {}
+  ParamIterator& operator=(const ParamIterator& other) {
+    if (this != &other)
+      impl_.reset(other.impl_->Clone());
+    return *this;
+  }
+
+  const T& operator*() const { return *impl_->Current(); }
+  const T* operator->() const { return impl_->Current(); }
+  // Prefix version of operator++.
+  ParamIterator& operator++() {
+    impl_->Advance();
+    return *this;
+  }
+  // Postfix version of operator++.
+  ParamIterator operator++(int /*unused*/) {
+    ParamIteratorInterface<T>* clone = impl_->Clone();
+    impl_->Advance();
+    return ParamIterator(clone);
+  }
+  bool operator==(const ParamIterator& other) const {
+    return impl_.get() == other.impl_.get() || impl_->Equals(*other.impl_);
+  }
+  bool operator!=(const ParamIterator& other) const {
+    return !(*this == other);
+  }
+
+ private:
+  friend class ParamGenerator<T>;
+  explicit ParamIterator(ParamIteratorInterface<T>* impl) : impl_(impl) {}
+  scoped_ptr<ParamIteratorInterface<T> > impl_;
+};
+
+// ParamGeneratorInterface<T> is the binary interface to access generators
+// defined in other translation units.
+template <typename T>
+class ParamGeneratorInterface {
+ public:
+  typedef T ParamType;
+
+  virtual ~ParamGeneratorInterface() {}
+
+  // Generator interface definition
+  virtual ParamIteratorInterface<T>* Begin() const = 0;
+  virtual ParamIteratorInterface<T>* End() const = 0;
+};
+
+// Wraps ParamGeneratorInterface<T> and provides general generator syntax
+// compatible with the STL Container concept.
+// This class implements copy initialization semantics and the contained
+// ParamGeneratorInterface<T> instance is shared among all copies
+// of the original object. This is possible because that instance is immutable.
+template<typename T>
+class ParamGenerator {
+ public:
+  typedef ParamIterator<T> iterator;
+
+  explicit ParamGenerator(ParamGeneratorInterface<T>* impl) : impl_(impl) {}
+  ParamGenerator(const ParamGenerator& other) : impl_(other.impl_) {}
+
+  ParamGenerator& operator=(const ParamGenerator& other) {
+    impl_ = other.impl_;
+    return *this;
+  }
+
+  iterator begin() const { return iterator(impl_->Begin()); }
+  iterator end() const { return iterator(impl_->End()); }
+
+ private:
+  linked_ptr<const ParamGeneratorInterface<T> > impl_;
+};
+
+// Generates values from a range of two comparable values. Can be used to
+// generate sequences of user-defined types that implement operator+() and
+// operator<().
+// This class is used in the Range() function.
+template <typename T, typename IncrementT>
+class RangeGenerator : public ParamGeneratorInterface<T> {
+ public:
+  RangeGenerator(T begin, T end, IncrementT step)
+      : begin_(begin), end_(end),
+        step_(step), end_index_(CalculateEndIndex(begin, end, step)) {}
+  virtual ~RangeGenerator() {}
+
+  virtual ParamIteratorInterface<T>* Begin() const {
+    return new Iterator(this, begin_, 0, step_);
+  }
+  virtual ParamIteratorInterface<T>* End() const {
+    return new Iterator(this, end_, end_index_, step_);
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<T> {
+   public:
+    Iterator(const ParamGeneratorInterface<T>* base, T value, int index,
+             IncrementT step)
+        : base_(base), value_(value), index_(index), step_(step) {}
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<T>* BaseGenerator() const {
+      return base_;
+    }
+    virtual void Advance() {
+      value_ = value_ + step_;
+      index_++;
+    }
+    virtual ParamIteratorInterface<T>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const T* Current() const { return &value_; }
+    virtual bool Equals(const ParamIteratorInterface<T>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const int other_index =
+          CheckedDowncastToActualType<const Iterator>(&other)->index_;
+      return index_ == other_index;
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : ParamIteratorInterface<T>(),
+          base_(other.base_), value_(other.value_), index_(other.index_),
+          step_(other.step_) {}
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<T>* const base_;
+    T value_;
+    int index_;
+    const IncrementT step_;
+  };  // class RangeGenerator::Iterator
+
+  static int CalculateEndIndex(const T& begin,
+                               const T& end,
+                               const IncrementT& step) {
+    int end_index = 0;
+    for (T i = begin; i < end; i = i + step)
+      end_index++;
+    return end_index;
+  }
+
+  // No implementation - assignment is unsupported.
+  void operator=(const RangeGenerator& other);
+
+  const T begin_;
+  const T end_;
+  const IncrementT step_;
+  // The index for the end() iterator. All the elements in the generated
+  // sequence are indexed (0-based) to aid iterator comparison.
+  const int end_index_;
+};  // class RangeGenerator
+
+
+// Generates values from a pair of STL-style iterators. Used in the
+// ValuesIn() function. The elements are copied from the source range
+// since the source can be located on the stack, and the generator
+// is likely to persist beyond that stack frame.
+template <typename T>
+class ValuesInIteratorRangeGenerator : public ParamGeneratorInterface<T> {
+ public:
+  template <typename ForwardIterator>
+  ValuesInIteratorRangeGenerator(ForwardIterator begin, ForwardIterator end)
+      : container_(begin, end) {}
+  virtual ~ValuesInIteratorRangeGenerator() {}
+
+  virtual ParamIteratorInterface<T>* Begin() const {
+    return new Iterator(this, container_.begin());
+  }
+  virtual ParamIteratorInterface<T>* End() const {
+    return new Iterator(this, container_.end());
+  }
+
+ private:
+  typedef typename ::std::vector<T> ContainerType;
+
+  class Iterator : public ParamIteratorInterface<T> {
+   public:
+    Iterator(const ParamGeneratorInterface<T>* base,
+             typename ContainerType::const_iterator iterator)
+        : base_(base), iterator_(iterator) {}
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<T>* BaseGenerator() const {
+      return base_;
+    }
+    virtual void Advance() {
+      ++iterator_;
+      value_.reset();
+    }
+    virtual ParamIteratorInterface<T>* Clone() const {
+      return new Iterator(*this);
+    }
+    // We need to use cached value referenced by iterator_ because *iterator_
+    // can return a temporary object (and of type other then T), so just
+    // having "return &*iterator_;" doesn't work.
+    // value_ is updated here and not in Advance() because Advance()
+    // can advance iterator_ beyond the end of the range, and we cannot
+    // detect that fact. The client code, on the other hand, is
+    // responsible for not calling Current() on an out-of-range iterator.
+    virtual const T* Current() const {
+      if (value_.get() == NULL)
+        value_.reset(new T(*iterator_));
+      return value_.get();
+    }
+    virtual bool Equals(const ParamIteratorInterface<T>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      return iterator_ ==
+          CheckedDowncastToActualType<const Iterator>(&other)->iterator_;
+    }
+
+   private:
+    Iterator(const Iterator& other)
+          // The explicit constructor call suppresses a false warning
+          // emitted by gcc when supplied with the -Wextra option.
+        : ParamIteratorInterface<T>(),
+          base_(other.base_),
+          iterator_(other.iterator_) {}
+
+    const ParamGeneratorInterface<T>* const base_;
+    typename ContainerType::const_iterator iterator_;
+    // A cached value of *iterator_. We keep it here to allow access by
+    // pointer in the wrapping iterator's operator->().
+    // value_ needs to be mutable to be accessed in Current().
+    // Use of scoped_ptr helps manage cached value's lifetime,
+    // which is bound by the lifespan of the iterator itself.
+    mutable scoped_ptr<const T> value_;
+  };  // class ValuesInIteratorRangeGenerator::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const ValuesInIteratorRangeGenerator& other);
+
+  const ContainerType container_;
+};  // class ValuesInIteratorRangeGenerator
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Stores a parameter value and later creates tests parameterized with that
+// value.
+template <class TestClass>
+class ParameterizedTestFactory : public TestFactoryBase {
+ public:
+  typedef typename TestClass::ParamType ParamType;
+  explicit ParameterizedTestFactory(ParamType parameter) :
+      parameter_(parameter) {}
+  virtual Test* CreateTest() {
+    TestClass::SetParam(&parameter_);
+    return new TestClass();
+  }
+
+ private:
+  const ParamType parameter_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestFactory);
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// TestMetaFactoryBase is a base class for meta-factories that create
+// test factories for passing into MakeAndRegisterTestInfo function.
+template <class ParamType>
+class TestMetaFactoryBase {
+ public:
+  virtual ~TestMetaFactoryBase() {}
+
+  virtual TestFactoryBase* CreateTestFactory(ParamType parameter) = 0;
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// TestMetaFactory creates test factories for passing into
+// MakeAndRegisterTestInfo function. Since MakeAndRegisterTestInfo receives
+// ownership of test factory pointer, same factory object cannot be passed
+// into that method twice. But ParameterizedTestCaseInfo is going to call
+// it for each Test/Parameter value combination. Thus it needs meta factory
+// creator class.
+template <class TestCase>
+class TestMetaFactory
+    : public TestMetaFactoryBase<typename TestCase::ParamType> {
+ public:
+  typedef typename TestCase::ParamType ParamType;
+
+  TestMetaFactory() {}
+
+  virtual TestFactoryBase* CreateTestFactory(ParamType parameter) {
+    return new ParameterizedTestFactory<TestCase>(parameter);
+  }
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestMetaFactory);
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// ParameterizedTestCaseInfoBase is a generic interface
+// to ParameterizedTestCaseInfo classes. ParameterizedTestCaseInfoBase
+// accumulates test information provided by TEST_P macro invocations
+// and generators provided by INSTANTIATE_TEST_CASE_P macro invocations
+// and uses that information to register all resulting test instances
+// in RegisterTests method. The ParameterizeTestCaseRegistry class holds
+// a collection of pointers to the ParameterizedTestCaseInfo objects
+// and calls RegisterTests() on each of them when asked.
+class ParameterizedTestCaseInfoBase {
+ public:
+  virtual ~ParameterizedTestCaseInfoBase() {}
+
+  // Base part of test case name for display purposes.
+  virtual const string& GetTestCaseName() const = 0;
+  // Test case id to verify identity.
+  virtual TypeId GetTestCaseTypeId() const = 0;
+  // UnitTest class invokes this method to register tests in this
+  // test case right before running them in RUN_ALL_TESTS macro.
+  // This method should not be called more then once on any single
+  // instance of a ParameterizedTestCaseInfoBase derived class.
+  virtual void RegisterTests() = 0;
+
+ protected:
+  ParameterizedTestCaseInfoBase() {}
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestCaseInfoBase);
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// ParameterizedTestCaseInfo accumulates tests obtained from TEST_P
+// macro invocations for a particular test case and generators
+// obtained from INSTANTIATE_TEST_CASE_P macro invocations for that
+// test case. It registers tests with all values generated by all
+// generators when asked.
+template <class TestCase>
+class ParameterizedTestCaseInfo : public ParameterizedTestCaseInfoBase {
+ public:
+  // ParamType and GeneratorCreationFunc are private types but are required
+  // for declarations of public methods AddTestPattern() and
+  // AddTestCaseInstantiation().
+  typedef typename TestCase::ParamType ParamType;
+  // A function that returns an instance of appropriate generator type.
+  typedef ParamGenerator<ParamType>(GeneratorCreationFunc)();
+
+  explicit ParameterizedTestCaseInfo(const char* name)
+      : test_case_name_(name) {}
+
+  // Test case base name for display purposes.
+  virtual const string& GetTestCaseName() const { return test_case_name_; }
+  // Test case id to verify identity.
+  virtual TypeId GetTestCaseTypeId() const { return GetTypeId<TestCase>(); }
+  // TEST_P macro uses AddTestPattern() to record information
+  // about a single test in a LocalTestInfo structure.
+  // test_case_name is the base name of the test case (without invocation
+  // prefix). test_base_name is the name of an individual test without
+  // parameter index. For the test SequenceA/FooTest.DoBar/1 FooTest is
+  // test case base name and DoBar is test base name.
+  void AddTestPattern(const char* test_case_name,
+                      const char* test_base_name,
+                      TestMetaFactoryBase<ParamType>* meta_factory) {
+    tests_.push_back(linked_ptr<TestInfo>(new TestInfo(test_case_name,
+                                                       test_base_name,
+                                                       meta_factory)));
+  }
+  // INSTANTIATE_TEST_CASE_P macro uses AddGenerator() to record information
+  // about a generator.
+  int AddTestCaseInstantiation(const string& instantiation_name,
+                               GeneratorCreationFunc* func,
+                               const char* /* file */,
+                               int /* line */) {
+    instantiations_.push_back(::std::make_pair(instantiation_name, func));
+    return 0;  // Return value used only to run this method in namespace scope.
+  }
+  // UnitTest class invokes this method to register tests in this test case
+  // test cases right before running tests in RUN_ALL_TESTS macro.
+  // This method should not be called more then once on any single
+  // instance of a ParameterizedTestCaseInfoBase derived class.
+  // UnitTest has a guard to prevent from calling this method more then once.
+  virtual void RegisterTests() {
+    for (typename TestInfoContainer::iterator test_it = tests_.begin();
+         test_it != tests_.end(); ++test_it) {
+      linked_ptr<TestInfo> test_info = *test_it;
+      for (typename InstantiationContainer::iterator gen_it =
+               instantiations_.begin(); gen_it != instantiations_.end();
+               ++gen_it) {
+        const string& instantiation_name = gen_it->first;
+        ParamGenerator<ParamType> generator((*gen_it->second)());
+
+        string test_case_name;
+        if ( !instantiation_name.empty() )
+          test_case_name = instantiation_name + "/";
+        test_case_name += test_info->test_case_base_name;
+
+        int i = 0;
+        for (typename ParamGenerator<ParamType>::iterator param_it =
+                 generator.begin();
+             param_it != generator.end(); ++param_it, ++i) {
+          Message test_name_stream;
+          test_name_stream << test_info->test_base_name << "/" << i;
+          MakeAndRegisterTestInfo(
+              test_case_name.c_str(),
+              test_name_stream.GetString().c_str(),
+              NULL,  // No type parameter.
+              PrintToString(*param_it).c_str(),
+              GetTestCaseTypeId(),
+              TestCase::SetUpTestCase,
+              TestCase::TearDownTestCase,
+              test_info->test_meta_factory->CreateTestFactory(*param_it));
+        }  // for param_it
+      }  // for gen_it
+    }  // for test_it
+  }  // RegisterTests
+
+ private:
+  // LocalTestInfo structure keeps information about a single test registered
+  // with TEST_P macro.
+  struct TestInfo {
+    TestInfo(const char* a_test_case_base_name,
+             const char* a_test_base_name,
+             TestMetaFactoryBase<ParamType>* a_test_meta_factory) :
+        test_case_base_name(a_test_case_base_name),
+        test_base_name(a_test_base_name),
+        test_meta_factory(a_test_meta_factory) {}
+
+    const string test_case_base_name;
+    const string test_base_name;
+    const scoped_ptr<TestMetaFactoryBase<ParamType> > test_meta_factory;
+  };
+  typedef ::std::vector<linked_ptr<TestInfo> > TestInfoContainer;
+  // Keeps pairs of <Instantiation name, Sequence generator creation function>
+  // received from INSTANTIATE_TEST_CASE_P macros.
+  typedef ::std::vector<std::pair<string, GeneratorCreationFunc*> >
+      InstantiationContainer;
+
+  const string test_case_name_;
+  TestInfoContainer tests_;
+  InstantiationContainer instantiations_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestCaseInfo);
+};  // class ParameterizedTestCaseInfo
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// ParameterizedTestCaseRegistry contains a map of ParameterizedTestCaseInfoBase
+// classes accessed by test case names. TEST_P and INSTANTIATE_TEST_CASE_P
+// macros use it to locate their corresponding ParameterizedTestCaseInfo
+// descriptors.
+class ParameterizedTestCaseRegistry {
+ public:
+  ParameterizedTestCaseRegistry() {}
+  ~ParameterizedTestCaseRegistry() {
+    for (TestCaseInfoContainer::iterator it = test_case_infos_.begin();
+         it != test_case_infos_.end(); ++it) {
+      delete *it;
+    }
+  }
+
+  // Looks up or creates and returns a structure containing information about
+  // tests and instantiations of a particular test case.
+  template <class TestCase>
+  ParameterizedTestCaseInfo<TestCase>* GetTestCasePatternHolder(
+      const char* test_case_name,
+      const char* file,
+      int line) {
+    ParameterizedTestCaseInfo<TestCase>* typed_test_info = NULL;
+    for (TestCaseInfoContainer::iterator it = test_case_infos_.begin();
+         it != test_case_infos_.end(); ++it) {
+      if ((*it)->GetTestCaseName() == test_case_name) {
+        if ((*it)->GetTestCaseTypeId() != GetTypeId<TestCase>()) {
+          // Complain about incorrect usage of Google Test facilities
+          // and terminate the program since we cannot guaranty correct
+          // test case setup and tear-down in this case.
+          ReportInvalidTestCaseType(test_case_name,  file, line);
+          posix::Abort();
+        } else {
+          // At this point we are sure that the object we found is of the same
+          // type we are looking for, so we downcast it to that type
+          // without further checks.
+          typed_test_info = CheckedDowncastToActualType<
+              ParameterizedTestCaseInfo<TestCase> >(*it);
+        }
+        break;
+      }
+    }
+    if (typed_test_info == NULL) {
+      typed_test_info = new ParameterizedTestCaseInfo<TestCase>(test_case_name);
+      test_case_infos_.push_back(typed_test_info);
+    }
+    return typed_test_info;
+  }
+  void RegisterTests() {
+    for (TestCaseInfoContainer::iterator it = test_case_infos_.begin();
+         it != test_case_infos_.end(); ++it) {
+      (*it)->RegisterTests();
+    }
+  }
+
+ private:
+  typedef ::std::vector<ParameterizedTestCaseInfoBase*> TestCaseInfoContainer;
+
+  TestCaseInfoContainer test_case_infos_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestCaseRegistry);
+};
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  //  GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_H_
+// This file was GENERATED by command:
+//     pump.py gtest-param-util-generated.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: vladl@google.com (Vlad Losev)
+
+// Type and function utilities for implementing parameterized tests.
+// This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+// Currently Google Test supports at most 50 arguments in Values,
+// and at most 10 arguments in Combine. Please contact
+// googletestframework@googlegroups.com if you need more.
+// Please note that the number of arguments to Combine is limited
+// by the maximum arity of the implementation of tr1::tuple which is
+// currently set at 10.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+
+// Forward declarations of ValuesIn(), which is implemented in
+// include/gtest/gtest-param-test.h.
+template <typename ForwardIterator>
+internal::ParamGenerator<
+  typename ::testing::internal::IteratorTraits<ForwardIterator>::value_type>
+ValuesIn(ForwardIterator begin, ForwardIterator end);
+
+template <typename T, size_t N>
+internal::ParamGenerator<T> ValuesIn(const T (&array)[N]);
+
+template <class Container>
+internal::ParamGenerator<typename Container::value_type> ValuesIn(
+    const Container& container);
+
+namespace internal {
+
+// Used in the Values() function to provide polymorphic capabilities.
+template <typename T1>
+class ValueArray1 {
+ public:
+  explicit ValueArray1(T1 v1) : v1_(v1) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const { return ValuesIn(&v1_, &v1_ + 1); }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray1& other);
+
+  const T1 v1_;
+};
+
+template <typename T1, typename T2>
+class ValueArray2 {
+ public:
+  ValueArray2(T1 v1, T2 v2) : v1_(v1), v2_(v2) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray2& other);
+
+  const T1 v1_;
+  const T2 v2_;
+};
+
+template <typename T1, typename T2, typename T3>
+class ValueArray3 {
+ public:
+  ValueArray3(T1 v1, T2 v2, T3 v3) : v1_(v1), v2_(v2), v3_(v3) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray3& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4>
+class ValueArray4 {
+ public:
+  ValueArray4(T1 v1, T2 v2, T3 v3, T4 v4) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray4& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+class ValueArray5 {
+ public:
+  ValueArray5(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray5& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6>
+class ValueArray6 {
+ public:
+  ValueArray6(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray6& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7>
+class ValueArray7 {
+ public:
+  ValueArray7(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray7& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8>
+class ValueArray8 {
+ public:
+  ValueArray8(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+      T8 v8) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray8& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9>
+class ValueArray9 {
+ public:
+  ValueArray9(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+      T9 v9) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray9& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10>
+class ValueArray10 {
+ public:
+  ValueArray10(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray10& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11>
+class ValueArray11 {
+ public:
+  ValueArray11(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray11& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12>
+class ValueArray12 {
+ public:
+  ValueArray12(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray12& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13>
+class ValueArray13 {
+ public:
+  ValueArray13(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray13& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14>
+class ValueArray14 {
+ public:
+  ValueArray14(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray14& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15>
+class ValueArray15 {
+ public:
+  ValueArray15(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray15& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16>
+class ValueArray16 {
+ public:
+  ValueArray16(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray16& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17>
+class ValueArray17 {
+ public:
+  ValueArray17(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16,
+      T17 v17) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray17& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18>
+class ValueArray18 {
+ public:
+  ValueArray18(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray18& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19>
+class ValueArray19 {
+ public:
+  ValueArray19(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray19& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20>
+class ValueArray20 {
+ public:
+  ValueArray20(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray20& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21>
+class ValueArray21 {
+ public:
+  ValueArray21(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray21& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22>
+class ValueArray22 {
+ public:
+  ValueArray22(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray22& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23>
+class ValueArray23 {
+ public:
+  ValueArray23(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray23& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24>
+class ValueArray24 {
+ public:
+  ValueArray24(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray24& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25>
+class ValueArray25 {
+ public:
+  ValueArray25(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24,
+      T25 v25) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray25& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26>
+class ValueArray26 {
+ public:
+  ValueArray26(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray26& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27>
+class ValueArray27 {
+ public:
+  ValueArray27(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19),
+      v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25),
+      v26_(v26), v27_(v27) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray27& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28>
+class ValueArray28 {
+ public:
+  ValueArray28(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24),
+      v25_(v25), v26_(v26), v27_(v27), v28_(v28) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray28& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29>
+class ValueArray29 {
+ public:
+  ValueArray29(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23),
+      v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray29& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30>
+class ValueArray30 {
+ public:
+  ValueArray30(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray30& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31>
+class ValueArray31 {
+ public:
+  ValueArray31(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray31& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32>
+class ValueArray32 {
+ public:
+  ValueArray32(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27),
+      v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray32& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33>
+class ValueArray33 {
+ public:
+  ValueArray33(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32,
+      T33 v33) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray33& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34>
+class ValueArray34 {
+ public:
+  ValueArray34(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray34& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35>
+class ValueArray35 {
+ public:
+  ValueArray35(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19),
+      v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25),
+      v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31),
+      v32_(v32), v33_(v33), v34_(v34), v35_(v35) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray35& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36>
+class ValueArray36 {
+ public:
+  ValueArray36(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24),
+      v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30),
+      v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35), v36_(v36) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray36& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37>
+class ValueArray37 {
+ public:
+  ValueArray37(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23),
+      v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29),
+      v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35),
+      v36_(v36), v37_(v37) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray37& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38>
+class ValueArray38 {
+ public:
+  ValueArray38(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray38& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39>
+class ValueArray39 {
+ public:
+  ValueArray39(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray39& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40>
+class ValueArray40 {
+ public:
+  ValueArray40(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27),
+      v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33),
+      v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39),
+      v40_(v40) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray40& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41>
+class ValueArray41 {
+ public:
+  ValueArray41(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40,
+      T41 v41) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray41& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42>
+class ValueArray42 {
+ public:
+  ValueArray42(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41), v42_(v42) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray42& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43>
+class ValueArray43 {
+ public:
+  ValueArray43(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19),
+      v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25),
+      v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31),
+      v32_(v32), v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37),
+      v38_(v38), v39_(v39), v40_(v40), v41_(v41), v42_(v42), v43_(v43) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray43& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44>
+class ValueArray44 {
+ public:
+  ValueArray44(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24),
+      v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30),
+      v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35), v36_(v36),
+      v37_(v37), v38_(v38), v39_(v39), v40_(v40), v41_(v41), v42_(v42),
+      v43_(v43), v44_(v44) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray44& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45>
+class ValueArray45 {
+ public:
+  ValueArray45(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23),
+      v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29),
+      v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35),
+      v36_(v36), v37_(v37), v38_(v38), v39_(v39), v40_(v40), v41_(v41),
+      v42_(v42), v43_(v43), v44_(v44), v45_(v45) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray45& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46>
+class ValueArray46 {
+ public:
+  ValueArray46(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39), v40_(v40),
+      v41_(v41), v42_(v42), v43_(v43), v44_(v44), v45_(v45), v46_(v46) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_), static_cast<T>(v46_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray46& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47>
+class ValueArray47 {
+ public:
+  ValueArray47(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39), v40_(v40),
+      v41_(v41), v42_(v42), v43_(v43), v44_(v44), v45_(v45), v46_(v46),
+      v47_(v47) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_), static_cast<T>(v46_), static_cast<T>(v47_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray47& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48>
+class ValueArray48 {
+ public:
+  ValueArray48(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47, T48 v48) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27),
+      v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33),
+      v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39),
+      v40_(v40), v41_(v41), v42_(v42), v43_(v43), v44_(v44), v45_(v45),
+      v46_(v46), v47_(v47), v48_(v48) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_), static_cast<T>(v46_), static_cast<T>(v47_),
+        static_cast<T>(v48_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray48& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+  const T48 v48_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49>
+class ValueArray49 {
+ public:
+  ValueArray49(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47, T48 v48,
+      T49 v49) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41), v42_(v42), v43_(v43), v44_(v44),
+      v45_(v45), v46_(v46), v47_(v47), v48_(v48), v49_(v49) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_), static_cast<T>(v46_), static_cast<T>(v47_),
+        static_cast<T>(v48_), static_cast<T>(v49_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray49& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+  const T48 v48_;
+  const T49 v49_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49, typename T50>
+class ValueArray50 {
+ public:
+  ValueArray50(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47, T48 v48, T49 v49,
+      T50 v50) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41), v42_(v42), v43_(v43), v44_(v44),
+      v45_(v45), v46_(v46), v47_(v47), v48_(v48), v49_(v49), v50_(v50) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_), static_cast<T>(v46_), static_cast<T>(v47_),
+        static_cast<T>(v48_), static_cast<T>(v49_), static_cast<T>(v50_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray50& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+  const T48 v48_;
+  const T49 v49_;
+  const T50 v50_;
+};
+
+# if GTEST_HAS_COMBINE
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Generates values from the Cartesian product of values produced
+// by the argument generators.
+//
+template <typename T1, typename T2>
+class CartesianProductGenerator2
+    : public ParamGeneratorInterface< ::std::tr1::tuple<T1, T2> > {
+ public:
+  typedef ::std::tr1::tuple<T1, T2> ParamType;
+
+  CartesianProductGenerator2(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2)
+      : g1_(g1), g2_(g2) {}
+  virtual ~CartesianProductGenerator2() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current2_;
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator2::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator2& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+};  // class CartesianProductGenerator2
+
+
+template <typename T1, typename T2, typename T3>
+class CartesianProductGenerator3
+    : public ParamGeneratorInterface< ::std::tr1::tuple<T1, T2, T3> > {
+ public:
+  typedef ::std::tr1::tuple<T1, T2, T3> ParamType;
+
+  CartesianProductGenerator3(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3)
+      : g1_(g1), g2_(g2), g3_(g3) {}
+  virtual ~CartesianProductGenerator3() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current3_;
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator3::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator3& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+};  // class CartesianProductGenerator3
+
+
+template <typename T1, typename T2, typename T3, typename T4>
+class CartesianProductGenerator4
+    : public ParamGeneratorInterface< ::std::tr1::tuple<T1, T2, T3, T4> > {
+ public:
+  typedef ::std::tr1::tuple<T1, T2, T3, T4> ParamType;
+
+  CartesianProductGenerator4(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4) {}
+  virtual ~CartesianProductGenerator4() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current4_;
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator4::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator4& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+};  // class CartesianProductGenerator4
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+class CartesianProductGenerator5
+    : public ParamGeneratorInterface< ::std::tr1::tuple<T1, T2, T3, T4, T5> > {
+ public:
+  typedef ::std::tr1::tuple<T1, T2, T3, T4, T5> ParamType;
+
+  CartesianProductGenerator5(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5) {}
+  virtual ~CartesianProductGenerator5() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current5_;
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator5::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator5& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+};  // class CartesianProductGenerator5
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6>
+class CartesianProductGenerator6
+    : public ParamGeneratorInterface< ::std::tr1::tuple<T1, T2, T3, T4, T5,
+        T6> > {
+ public:
+  typedef ::std::tr1::tuple<T1, T2, T3, T4, T5, T6> ParamType;
+
+  CartesianProductGenerator6(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5,
+      const ParamGenerator<T6>& g6)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6) {}
+  virtual ~CartesianProductGenerator6() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5,
+      const ParamGenerator<T6>& g6,
+      const typename ParamGenerator<T6>::iterator& current6)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current6_;
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    const typename ParamGenerator<T6>::iterator begin6_;
+    const typename ParamGenerator<T6>::iterator end6_;
+    typename ParamGenerator<T6>::iterator current6_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator6::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator6& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+  const ParamGenerator<T6> g6_;
+};  // class CartesianProductGenerator6
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7>
+class CartesianProductGenerator7
+    : public ParamGeneratorInterface< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6,
+        T7> > {
+ public:
+  typedef ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7> ParamType;
+
+  CartesianProductGenerator7(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5,
+      const ParamGenerator<T6>& g6, const ParamGenerator<T7>& g7)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7) {}
+  virtual ~CartesianProductGenerator7() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5,
+      const ParamGenerator<T6>& g6,
+      const typename ParamGenerator<T6>::iterator& current6,
+      const ParamGenerator<T7>& g7,
+      const typename ParamGenerator<T7>::iterator& current7)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current7_;
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    const typename ParamGenerator<T6>::iterator begin6_;
+    const typename ParamGenerator<T6>::iterator end6_;
+    typename ParamGenerator<T6>::iterator current6_;
+    const typename ParamGenerator<T7>::iterator begin7_;
+    const typename ParamGenerator<T7>::iterator end7_;
+    typename ParamGenerator<T7>::iterator current7_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator7::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator7& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+  const ParamGenerator<T6> g6_;
+  const ParamGenerator<T7> g7_;
+};  // class CartesianProductGenerator7
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8>
+class CartesianProductGenerator8
+    : public ParamGeneratorInterface< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6,
+        T7, T8> > {
+ public:
+  typedef ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8> ParamType;
+
+  CartesianProductGenerator8(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5,
+      const ParamGenerator<T6>& g6, const ParamGenerator<T7>& g7,
+      const ParamGenerator<T8>& g8)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7),
+          g8_(g8) {}
+  virtual ~CartesianProductGenerator8() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin(), g8_, g8_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end(), g8_,
+        g8_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5,
+      const ParamGenerator<T6>& g6,
+      const typename ParamGenerator<T6>::iterator& current6,
+      const ParamGenerator<T7>& g7,
+      const typename ParamGenerator<T7>::iterator& current7,
+      const ParamGenerator<T8>& g8,
+      const typename ParamGenerator<T8>::iterator& current8)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7),
+          begin8_(g8.begin()), end8_(g8.end()), current8_(current8)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current8_;
+      if (current8_ == end8_) {
+        current8_ = begin8_;
+        ++current7_;
+      }
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_ &&
+          current8_ == typed_other->current8_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_),
+        begin8_(other.begin8_),
+        end8_(other.end8_),
+        current8_(other.current8_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_, *current8_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_ ||
+          current8_ == end8_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    const typename ParamGenerator<T6>::iterator begin6_;
+    const typename ParamGenerator<T6>::iterator end6_;
+    typename ParamGenerator<T6>::iterator current6_;
+    const typename ParamGenerator<T7>::iterator begin7_;
+    const typename ParamGenerator<T7>::iterator end7_;
+    typename ParamGenerator<T7>::iterator current7_;
+    const typename ParamGenerator<T8>::iterator begin8_;
+    const typename ParamGenerator<T8>::iterator end8_;
+    typename ParamGenerator<T8>::iterator current8_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator8::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator8& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+  const ParamGenerator<T6> g6_;
+  const ParamGenerator<T7> g7_;
+  const ParamGenerator<T8> g8_;
+};  // class CartesianProductGenerator8
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9>
+class CartesianProductGenerator9
+    : public ParamGeneratorInterface< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6,
+        T7, T8, T9> > {
+ public:
+  typedef ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8, T9> ParamType;
+
+  CartesianProductGenerator9(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5,
+      const ParamGenerator<T6>& g6, const ParamGenerator<T7>& g7,
+      const ParamGenerator<T8>& g8, const ParamGenerator<T9>& g9)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9) {}
+  virtual ~CartesianProductGenerator9() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin(), g8_, g8_.begin(), g9_, g9_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end(), g8_,
+        g8_.end(), g9_, g9_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5,
+      const ParamGenerator<T6>& g6,
+      const typename ParamGenerator<T6>::iterator& current6,
+      const ParamGenerator<T7>& g7,
+      const typename ParamGenerator<T7>::iterator& current7,
+      const ParamGenerator<T8>& g8,
+      const typename ParamGenerator<T8>::iterator& current8,
+      const ParamGenerator<T9>& g9,
+      const typename ParamGenerator<T9>::iterator& current9)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7),
+          begin8_(g8.begin()), end8_(g8.end()), current8_(current8),
+          begin9_(g9.begin()), end9_(g9.end()), current9_(current9)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current9_;
+      if (current9_ == end9_) {
+        current9_ = begin9_;
+        ++current8_;
+      }
+      if (current8_ == end8_) {
+        current8_ = begin8_;
+        ++current7_;
+      }
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_ &&
+          current8_ == typed_other->current8_ &&
+          current9_ == typed_other->current9_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_),
+        begin8_(other.begin8_),
+        end8_(other.end8_),
+        current8_(other.current8_),
+        begin9_(other.begin9_),
+        end9_(other.end9_),
+        current9_(other.current9_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_, *current8_,
+            *current9_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_ ||
+          current8_ == end8_ ||
+          current9_ == end9_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    const typename ParamGenerator<T6>::iterator begin6_;
+    const typename ParamGenerator<T6>::iterator end6_;
+    typename ParamGenerator<T6>::iterator current6_;
+    const typename ParamGenerator<T7>::iterator begin7_;
+    const typename ParamGenerator<T7>::iterator end7_;
+    typename ParamGenerator<T7>::iterator current7_;
+    const typename ParamGenerator<T8>::iterator begin8_;
+    const typename ParamGenerator<T8>::iterator end8_;
+    typename ParamGenerator<T8>::iterator current8_;
+    const typename ParamGenerator<T9>::iterator begin9_;
+    const typename ParamGenerator<T9>::iterator end9_;
+    typename ParamGenerator<T9>::iterator current9_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator9::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator9& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+  const ParamGenerator<T6> g6_;
+  const ParamGenerator<T7> g7_;
+  const ParamGenerator<T8> g8_;
+  const ParamGenerator<T9> g9_;
+};  // class CartesianProductGenerator9
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10>
+class CartesianProductGenerator10
+    : public ParamGeneratorInterface< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6,
+        T7, T8, T9, T10> > {
+ public:
+  typedef ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10> ParamType;
+
+  CartesianProductGenerator10(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5,
+      const ParamGenerator<T6>& g6, const ParamGenerator<T7>& g7,
+      const ParamGenerator<T8>& g8, const ParamGenerator<T9>& g9,
+      const ParamGenerator<T10>& g10)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9), g10_(g10) {}
+  virtual ~CartesianProductGenerator10() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin(), g8_, g8_.begin(), g9_, g9_.begin(), g10_, g10_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end(), g8_,
+        g8_.end(), g9_, g9_.end(), g10_, g10_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5,
+      const ParamGenerator<T6>& g6,
+      const typename ParamGenerator<T6>::iterator& current6,
+      const ParamGenerator<T7>& g7,
+      const typename ParamGenerator<T7>::iterator& current7,
+      const ParamGenerator<T8>& g8,
+      const typename ParamGenerator<T8>::iterator& current8,
+      const ParamGenerator<T9>& g9,
+      const typename ParamGenerator<T9>::iterator& current9,
+      const ParamGenerator<T10>& g10,
+      const typename ParamGenerator<T10>::iterator& current10)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7),
+          begin8_(g8.begin()), end8_(g8.end()), current8_(current8),
+          begin9_(g9.begin()), end9_(g9.end()), current9_(current9),
+          begin10_(g10.begin()), end10_(g10.end()), current10_(current10)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current10_;
+      if (current10_ == end10_) {
+        current10_ = begin10_;
+        ++current9_;
+      }
+      if (current9_ == end9_) {
+        current9_ = begin9_;
+        ++current8_;
+      }
+      if (current8_ == end8_) {
+        current8_ = begin8_;
+        ++current7_;
+      }
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_ &&
+          current8_ == typed_other->current8_ &&
+          current9_ == typed_other->current9_ &&
+          current10_ == typed_other->current10_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_),
+        begin8_(other.begin8_),
+        end8_(other.end8_),
+        current8_(other.current8_),
+        begin9_(other.begin9_),
+        end9_(other.end9_),
+        current9_(other.current9_),
+        begin10_(other.begin10_),
+        end10_(other.end10_),
+        current10_(other.current10_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_, *current8_,
+            *current9_, *current10_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_ ||
+          current8_ == end8_ ||
+          current9_ == end9_ ||
+          current10_ == end10_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    const typename ParamGenerator<T6>::iterator begin6_;
+    const typename ParamGenerator<T6>::iterator end6_;
+    typename ParamGenerator<T6>::iterator current6_;
+    const typename ParamGenerator<T7>::iterator begin7_;
+    const typename ParamGenerator<T7>::iterator end7_;
+    typename ParamGenerator<T7>::iterator current7_;
+    const typename ParamGenerator<T8>::iterator begin8_;
+    const typename ParamGenerator<T8>::iterator end8_;
+    typename ParamGenerator<T8>::iterator current8_;
+    const typename ParamGenerator<T9>::iterator begin9_;
+    const typename ParamGenerator<T9>::iterator end9_;
+    typename ParamGenerator<T9>::iterator current9_;
+    const typename ParamGenerator<T10>::iterator begin10_;
+    const typename ParamGenerator<T10>::iterator end10_;
+    typename ParamGenerator<T10>::iterator current10_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator10::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator10& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+  const ParamGenerator<T6> g6_;
+  const ParamGenerator<T7> g7_;
+  const ParamGenerator<T8> g8_;
+  const ParamGenerator<T9> g9_;
+  const ParamGenerator<T10> g10_;
+};  // class CartesianProductGenerator10
+
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Helper classes providing Combine() with polymorphic features. They allow
+// casting CartesianProductGeneratorN<T> to ParamGenerator<U> if T is
+// convertible to U.
+//
+template <class Generator1, class Generator2>
+class CartesianProductHolder2 {
+ public:
+CartesianProductHolder2(const Generator1& g1, const Generator2& g2)
+      : g1_(g1), g2_(g2) {}
+  template <typename T1, typename T2>
+  operator ParamGenerator< ::std::tr1::tuple<T1, T2> >() const {
+    return ParamGenerator< ::std::tr1::tuple<T1, T2> >(
+        new CartesianProductGenerator2<T1, T2>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder2& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+};  // class CartesianProductHolder2
+
+template <class Generator1, class Generator2, class Generator3>
+class CartesianProductHolder3 {
+ public:
+CartesianProductHolder3(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3)
+      : g1_(g1), g2_(g2), g3_(g3) {}
+  template <typename T1, typename T2, typename T3>
+  operator ParamGenerator< ::std::tr1::tuple<T1, T2, T3> >() const {
+    return ParamGenerator< ::std::tr1::tuple<T1, T2, T3> >(
+        new CartesianProductGenerator3<T1, T2, T3>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder3& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+};  // class CartesianProductHolder3
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4>
+class CartesianProductHolder4 {
+ public:
+CartesianProductHolder4(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4) {}
+  template <typename T1, typename T2, typename T3, typename T4>
+  operator ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4> >() const {
+    return ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4> >(
+        new CartesianProductGenerator4<T1, T2, T3, T4>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder4& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+};  // class CartesianProductHolder4
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5>
+class CartesianProductHolder5 {
+ public:
+CartesianProductHolder5(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5>
+  operator ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5> >() const {
+    return ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5> >(
+        new CartesianProductGenerator5<T1, T2, T3, T4, T5>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder5& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+};  // class CartesianProductHolder5
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5, class Generator6>
+class CartesianProductHolder6 {
+ public:
+CartesianProductHolder6(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5,
+      typename T6>
+  operator ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6> >() const {
+    return ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6> >(
+        new CartesianProductGenerator6<T1, T2, T3, T4, T5, T6>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_),
+        static_cast<ParamGenerator<T6> >(g6_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder6& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+};  // class CartesianProductHolder6
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5, class Generator6, class Generator7>
+class CartesianProductHolder7 {
+ public:
+CartesianProductHolder7(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5,
+      typename T6, typename T7>
+  operator ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6,
+      T7> >() const {
+    return ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7> >(
+        new CartesianProductGenerator7<T1, T2, T3, T4, T5, T6, T7>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_),
+        static_cast<ParamGenerator<T6> >(g6_),
+        static_cast<ParamGenerator<T7> >(g7_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder7& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+};  // class CartesianProductHolder7
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5, class Generator6, class Generator7,
+    class Generator8>
+class CartesianProductHolder8 {
+ public:
+CartesianProductHolder8(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7, const Generator8& g8)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7),
+          g8_(g8) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5,
+      typename T6, typename T7, typename T8>
+  operator ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7,
+      T8> >() const {
+    return ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8> >(
+        new CartesianProductGenerator8<T1, T2, T3, T4, T5, T6, T7, T8>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_),
+        static_cast<ParamGenerator<T6> >(g6_),
+        static_cast<ParamGenerator<T7> >(g7_),
+        static_cast<ParamGenerator<T8> >(g8_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder8& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+  const Generator8 g8_;
+};  // class CartesianProductHolder8
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5, class Generator6, class Generator7,
+    class Generator8, class Generator9>
+class CartesianProductHolder9 {
+ public:
+CartesianProductHolder9(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7, const Generator8& g8,
+    const Generator9& g9)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5,
+      typename T6, typename T7, typename T8, typename T9>
+  operator ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8,
+      T9> >() const {
+    return ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8,
+        T9> >(
+        new CartesianProductGenerator9<T1, T2, T3, T4, T5, T6, T7, T8, T9>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_),
+        static_cast<ParamGenerator<T6> >(g6_),
+        static_cast<ParamGenerator<T7> >(g7_),
+        static_cast<ParamGenerator<T8> >(g8_),
+        static_cast<ParamGenerator<T9> >(g9_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder9& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+  const Generator8 g8_;
+  const Generator9 g9_;
+};  // class CartesianProductHolder9
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5, class Generator6, class Generator7,
+    class Generator8, class Generator9, class Generator10>
+class CartesianProductHolder10 {
+ public:
+CartesianProductHolder10(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7, const Generator8& g8,
+    const Generator9& g9, const Generator10& g10)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9), g10_(g10) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5,
+      typename T6, typename T7, typename T8, typename T9, typename T10>
+  operator ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8,
+      T9, T10> >() const {
+    return ParamGenerator< ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8,
+        T9, T10> >(
+        new CartesianProductGenerator10<T1, T2, T3, T4, T5, T6, T7, T8, T9,
+            T10>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_),
+        static_cast<ParamGenerator<T6> >(g6_),
+        static_cast<ParamGenerator<T7> >(g7_),
+        static_cast<ParamGenerator<T8> >(g8_),
+        static_cast<ParamGenerator<T9> >(g9_),
+        static_cast<ParamGenerator<T10> >(g10_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder10& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+  const Generator8 g8_;
+  const Generator9 g9_;
+  const Generator10 g10_;
+};  // class CartesianProductHolder10
+
+# endif  // GTEST_HAS_COMBINE
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  //  GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+
+// Functions producing parameter generators.
+//
+// Google Test uses these generators to produce parameters for value-
+// parameterized tests. When a parameterized test case is instantiated
+// with a particular generator, Google Test creates and runs tests
+// for each element in the sequence produced by the generator.
+//
+// In the following sample, tests from test case FooTest are instantiated
+// each three times with parameter values 3, 5, and 8:
+//
+// class FooTest : public TestWithParam<int> { ... };
+//
+// TEST_P(FooTest, TestThis) {
+// }
+// TEST_P(FooTest, TestThat) {
+// }
+// INSTANTIATE_TEST_CASE_P(TestSequence, FooTest, Values(3, 5, 8));
+//
+
+// Range() returns generators providing sequences of values in a range.
+//
+// Synopsis:
+// Range(start, end)
+//   - returns a generator producing a sequence of values {start, start+1,
+//     start+2, ..., }.
+// Range(start, end, step)
+//   - returns a generator producing a sequence of values {start, start+step,
+//     start+step+step, ..., }.
+// Notes:
+//   * The generated sequences never include end. For example, Range(1, 5)
+//     returns a generator producing a sequence {1, 2, 3, 4}. Range(1, 9, 2)
+//     returns a generator producing {1, 3, 5, 7}.
+//   * start and end must have the same type. That type may be any integral or
+//     floating-point type or a user defined type satisfying these conditions:
+//     * It must be assignable (have operator=() defined).
+//     * It must have operator+() (operator+(int-compatible type) for
+//       two-operand version).
+//     * It must have operator<() defined.
+//     Elements in the resulting sequences will also have that type.
+//   * Condition start < end must be satisfied in order for resulting sequences
+//     to contain any elements.
+//
+template <typename T, typename IncrementT>
+internal::ParamGenerator<T> Range(T start, T end, IncrementT step) {
+  return internal::ParamGenerator<T>(
+      new internal::RangeGenerator<T, IncrementT>(start, end, step));
+}
+
+template <typename T>
+internal::ParamGenerator<T> Range(T start, T end) {
+  return Range(start, end, 1);
+}
+
+// ValuesIn() function allows generation of tests with parameters coming from
+// a container.
+//
+// Synopsis:
+// ValuesIn(const T (&array)[N])
+//   - returns a generator producing sequences with elements from
+//     a C-style array.
+// ValuesIn(const Container& container)
+//   - returns a generator producing sequences with elements from
+//     an STL-style container.
+// ValuesIn(Iterator begin, Iterator end)
+//   - returns a generator producing sequences with elements from
+//     a range [begin, end) defined by a pair of STL-style iterators. These
+//     iterators can also be plain C pointers.
+//
+// Please note that ValuesIn copies the values from the containers
+// passed in and keeps them to generate tests in RUN_ALL_TESTS().
+//
+// Examples:
+//
+// This instantiates tests from test case StringTest
+// each with C-string values of "foo", "bar", and "baz":
+//
+// const char* strings[] = {"foo", "bar", "baz"};
+// INSTANTIATE_TEST_CASE_P(StringSequence, SrtingTest, ValuesIn(strings));
+//
+// This instantiates tests from test case StlStringTest
+// each with STL strings with values "a" and "b":
+//
+// ::std::vector< ::std::string> GetParameterStrings() {
+//   ::std::vector< ::std::string> v;
+//   v.push_back("a");
+//   v.push_back("b");
+//   return v;
+// }
+//
+// INSTANTIATE_TEST_CASE_P(CharSequence,
+//                         StlStringTest,
+//                         ValuesIn(GetParameterStrings()));
+//
+//
+// This will also instantiate tests from CharTest
+// each with parameter values 'a' and 'b':
+//
+// ::std::list<char> GetParameterChars() {
+//   ::std::list<char> list;
+//   list.push_back('a');
+//   list.push_back('b');
+//   return list;
+// }
+// ::std::list<char> l = GetParameterChars();
+// INSTANTIATE_TEST_CASE_P(CharSequence2,
+//                         CharTest,
+//                         ValuesIn(l.begin(), l.end()));
+//
+template <typename ForwardIterator>
+internal::ParamGenerator<
+  typename ::testing::internal::IteratorTraits<ForwardIterator>::value_type>
+ValuesIn(ForwardIterator begin, ForwardIterator end) {
+  typedef typename ::testing::internal::IteratorTraits<ForwardIterator>
+      ::value_type ParamType;
+  return internal::ParamGenerator<ParamType>(
+      new internal::ValuesInIteratorRangeGenerator<ParamType>(begin, end));
+}
+
+template <typename T, size_t N>
+internal::ParamGenerator<T> ValuesIn(const T (&array)[N]) {
+  return ValuesIn(array, array + N);
+}
+
+template <class Container>
+internal::ParamGenerator<typename Container::value_type> ValuesIn(
+    const Container& container) {
+  return ValuesIn(container.begin(), container.end());
+}
+
+// Values() allows generating tests from explicitly specified list of
+// parameters.
+//
+// Synopsis:
+// Values(T v1, T v2, ..., T vN)
+//   - returns a generator producing sequences with elements v1, v2, ..., vN.
+//
+// For example, this instantiates tests from test case BarTest each
+// with values "one", "two", and "three":
+//
+// INSTANTIATE_TEST_CASE_P(NumSequence, BarTest, Values("one", "two", "three"));
+//
+// This instantiates tests from test case BazTest each with values 1, 2, 3.5.
+// The exact type of values will depend on the type of parameter in BazTest.
+//
+// INSTANTIATE_TEST_CASE_P(FloatingNumbers, BazTest, Values(1, 2, 3.5));
+//
+// Currently, Values() supports from 1 to 50 parameters.
+//
+template <typename T1>
+internal::ValueArray1<T1> Values(T1 v1) {
+  return internal::ValueArray1<T1>(v1);
+}
+
+template <typename T1, typename T2>
+internal::ValueArray2<T1, T2> Values(T1 v1, T2 v2) {
+  return internal::ValueArray2<T1, T2>(v1, v2);
+}
+
+template <typename T1, typename T2, typename T3>
+internal::ValueArray3<T1, T2, T3> Values(T1 v1, T2 v2, T3 v3) {
+  return internal::ValueArray3<T1, T2, T3>(v1, v2, v3);
+}
+
+template <typename T1, typename T2, typename T3, typename T4>
+internal::ValueArray4<T1, T2, T3, T4> Values(T1 v1, T2 v2, T3 v3, T4 v4) {
+  return internal::ValueArray4<T1, T2, T3, T4>(v1, v2, v3, v4);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+internal::ValueArray5<T1, T2, T3, T4, T5> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5) {
+  return internal::ValueArray5<T1, T2, T3, T4, T5>(v1, v2, v3, v4, v5);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6>
+internal::ValueArray6<T1, T2, T3, T4, T5, T6> Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6) {
+  return internal::ValueArray6<T1, T2, T3, T4, T5, T6>(v1, v2, v3, v4, v5, v6);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7>
+internal::ValueArray7<T1, T2, T3, T4, T5, T6, T7> Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6, T7 v7) {
+  return internal::ValueArray7<T1, T2, T3, T4, T5, T6, T7>(v1, v2, v3, v4, v5,
+      v6, v7);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8>
+internal::ValueArray8<T1, T2, T3, T4, T5, T6, T7, T8> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8) {
+  return internal::ValueArray8<T1, T2, T3, T4, T5, T6, T7, T8>(v1, v2, v3, v4,
+      v5, v6, v7, v8);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9>
+internal::ValueArray9<T1, T2, T3, T4, T5, T6, T7, T8, T9> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9) {
+  return internal::ValueArray9<T1, T2, T3, T4, T5, T6, T7, T8, T9>(v1, v2, v3,
+      v4, v5, v6, v7, v8, v9);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10>
+internal::ValueArray10<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10> Values(T1 v1,
+    T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10) {
+  return internal::ValueArray10<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10>(v1,
+      v2, v3, v4, v5, v6, v7, v8, v9, v10);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11>
+internal::ValueArray11<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10,
+    T11> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11) {
+  return internal::ValueArray11<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10,
+      T11>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12>
+internal::ValueArray12<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+    T12> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12) {
+  return internal::ValueArray12<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13>
+internal::ValueArray13<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+    T13> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13) {
+  return internal::ValueArray13<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14>
+internal::ValueArray14<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14) {
+  return internal::ValueArray14<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+      v14);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15>
+internal::ValueArray15<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+    T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15) {
+  return internal::ValueArray15<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+      v13, v14, v15);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16>
+internal::ValueArray16<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16) {
+  return internal::ValueArray16<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+      v12, v13, v14, v15, v16);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17>
+internal::ValueArray17<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17) {
+  return internal::ValueArray17<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10,
+      v11, v12, v13, v14, v15, v16, v17);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18>
+internal::ValueArray18<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6,
+    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18) {
+  return internal::ValueArray18<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18>(v1, v2, v3, v4, v5, v6, v7, v8, v9,
+      v10, v11, v12, v13, v14, v15, v16, v17, v18);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19>
+internal::ValueArray19<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5,
+    T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14,
+    T15 v15, T16 v16, T17 v17, T18 v18, T19 v19) {
+  return internal::ValueArray19<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19>(v1, v2, v3, v4, v5, v6, v7, v8,
+      v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20>
+internal::ValueArray20<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20) {
+  return internal::ValueArray20<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20>(v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21>
+internal::ValueArray21<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21) {
+  return internal::ValueArray21<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21>(v1, v2, v3, v4, v5, v6,
+      v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22>
+internal::ValueArray22<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22> Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22) {
+  return internal::ValueArray22<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22>(v1, v2, v3, v4,
+      v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23>
+internal::ValueArray23<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23) {
+  return internal::ValueArray23<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23>(v1, v2, v3,
+      v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24>
+internal::ValueArray24<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24) {
+  return internal::ValueArray24<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24>(v1, v2,
+      v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18,
+      v19, v20, v21, v22, v23, v24);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25>
+internal::ValueArray25<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25> Values(T1 v1,
+    T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11,
+    T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19,
+    T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25) {
+  return internal::ValueArray25<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25>(v1,
+      v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17,
+      v18, v19, v20, v21, v22, v23, v24, v25);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26>
+internal::ValueArray26<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+    T26> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26) {
+  return internal::ValueArray26<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15,
+      v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27>
+internal::ValueArray27<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+    T27> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27) {
+  return internal::ValueArray27<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14,
+      v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28>
+internal::ValueArray28<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+    T28> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28) {
+  return internal::ValueArray28<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+      v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27,
+      v28);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29>
+internal::ValueArray29<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29) {
+  return internal::ValueArray29<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+      v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26,
+      v27, v28, v29);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30>
+internal::ValueArray30<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+    T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16,
+    T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24,
+    T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30) {
+  return internal::ValueArray30<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+      v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25,
+      v26, v27, v28, v29, v30);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31>
+internal::ValueArray31<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31) {
+  return internal::ValueArray31<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10,
+      v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24,
+      v25, v26, v27, v28, v29, v30, v31);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32>
+internal::ValueArray32<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32) {
+  return internal::ValueArray32<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32>(v1, v2, v3, v4, v5, v6, v7, v8, v9,
+      v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33>
+internal::ValueArray33<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6,
+    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33) {
+  return internal::ValueArray33<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33>(v1, v2, v3, v4, v5, v6, v7, v8,
+      v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32, v33);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34>
+internal::ValueArray34<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5,
+    T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14,
+    T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22,
+    T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30,
+    T31 v31, T32 v32, T33 v33, T34 v34) {
+  return internal::ValueArray34<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34>(v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22,
+      v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35>
+internal::ValueArray35<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21,
+    T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29,
+    T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35) {
+  return internal::ValueArray35<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35>(v1, v2, v3, v4, v5, v6,
+      v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21,
+      v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36>
+internal::ValueArray36<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21,
+    T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29,
+    T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36) {
+  return internal::ValueArray36<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36>(v1, v2, v3, v4,
+      v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33,
+      v34, v35, v36);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37>
+internal::ValueArray37<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37> Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28,
+    T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36,
+    T37 v37) {
+  return internal::ValueArray37<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37>(v1, v2, v3,
+      v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33,
+      v34, v35, v36, v37);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38>
+internal::ValueArray38<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28,
+    T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36,
+    T37 v37, T38 v38) {
+  return internal::ValueArray38<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38>(v1, v2,
+      v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18,
+      v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32,
+      v33, v34, v35, v36, v37, v38);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39>
+internal::ValueArray39<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28,
+    T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36,
+    T37 v37, T38 v38, T39 v39) {
+  return internal::ValueArray39<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39>(v1,
+      v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17,
+      v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31,
+      v32, v33, v34, v35, v36, v37, v38, v39);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40>
+internal::ValueArray40<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40> Values(T1 v1,
+    T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11,
+    T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19,
+    T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27,
+    T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35,
+    T36 v36, T37 v37, T38 v38, T39 v39, T40 v40) {
+  return internal::ValueArray40<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15,
+      v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29,
+      v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41>
+internal::ValueArray41<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+    T41> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41) {
+  return internal::ValueArray41<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14,
+      v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28,
+      v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42>
+internal::ValueArray42<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+    T42> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+    T42 v42) {
+  return internal::ValueArray42<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+      v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27,
+      v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41,
+      v42);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43>
+internal::ValueArray43<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+    T43> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+    T42 v42, T43 v43) {
+  return internal::ValueArray43<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+      v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26,
+      v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40,
+      v41, v42, v43);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44>
+internal::ValueArray44<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+    T42 v42, T43 v43, T44 v44) {
+  return internal::ValueArray44<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+      v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25,
+      v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39,
+      v40, v41, v42, v43, v44);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45>
+internal::ValueArray45<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+    T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16,
+    T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24,
+    T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32,
+    T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40,
+    T41 v41, T42 v42, T43 v43, T44 v44, T45 v45) {
+  return internal::ValueArray45<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10,
+      v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24,
+      v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38,
+      v39, v40, v41, v42, v43, v44, v45);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46>
+internal::ValueArray46<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39,
+    T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46) {
+  return internal::ValueArray46<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45, T46>(v1, v2, v3, v4, v5, v6, v7, v8, v9,
+      v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37,
+      v38, v39, v40, v41, v42, v43, v44, v45, v46);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47>
+internal::ValueArray47<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46, T47> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39,
+    T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47) {
+  return internal::ValueArray47<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45, T46, T47>(v1, v2, v3, v4, v5, v6, v7, v8,
+      v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37,
+      v38, v39, v40, v41, v42, v43, v44, v45, v46, v47);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48>
+internal::ValueArray48<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46, T47, T48> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6,
+    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39,
+    T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47,
+    T48 v48) {
+  return internal::ValueArray48<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45, T46, T47, T48>(v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22,
+      v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36,
+      v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49>
+internal::ValueArray49<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46, T47, T48, T49> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5,
+    T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14,
+    T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22,
+    T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30,
+    T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38,
+    T39 v39, T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46,
+    T47 v47, T48 v48, T49 v49) {
+  return internal::ValueArray49<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45, T46, T47, T48, T49>(v1, v2, v3, v4, v5, v6,
+      v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21,
+      v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35,
+      v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49, typename T50>
+internal::ValueArray50<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46, T47, T48, T49, T50> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21,
+    T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29,
+    T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37,
+    T38 v38, T39 v39, T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45,
+    T46 v46, T47 v47, T48 v48, T49 v49, T50 v50) {
+  return internal::ValueArray50<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45, T46, T47, T48, T49, T50>(v1, v2, v3, v4,
+      v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33,
+      v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47,
+      v48, v49, v50);
+}
+
+// Bool() allows generating tests with parameters in a set of (false, true).
+//
+// Synopsis:
+// Bool()
+//   - returns a generator producing sequences with elements {false, true}.
+//
+// It is useful when testing code that depends on Boolean flags. Combinations
+// of multiple flags can be tested when several Bool()'s are combined using
+// Combine() function.
+//
+// In the following example all tests in the test case FlagDependentTest
+// will be instantiated twice with parameters false and true.
+//
+// class FlagDependentTest : public testing::TestWithParam<bool> {
+//   virtual void SetUp() {
+//     external_flag = GetParam();
+//   }
+// }
+// INSTANTIATE_TEST_CASE_P(BoolSequence, FlagDependentTest, Bool());
+//
+inline internal::ParamGenerator<bool> Bool() {
+  return Values(false, true);
+}
+
+# if GTEST_HAS_COMBINE
+// Combine() allows the user to combine two or more sequences to produce
+// values of a Cartesian product of those sequences' elements.
+//
+// Synopsis:
+// Combine(gen1, gen2, ..., genN)
+//   - returns a generator producing sequences with elements coming from
+//     the Cartesian product of elements from the sequences generated by
+//     gen1, gen2, ..., genN. The sequence elements will have a type of
+//     tuple<T1, T2, ..., TN> where T1, T2, ..., TN are the types
+//     of elements from sequences produces by gen1, gen2, ..., genN.
+//
+// Combine can have up to 10 arguments. This number is currently limited
+// by the maximum number of elements in the tuple implementation used by Google
+// Test.
+//
+// Example:
+//
+// This will instantiate tests in test case AnimalTest each one with
+// the parameter values tuple("cat", BLACK), tuple("cat", WHITE),
+// tuple("dog", BLACK), and tuple("dog", WHITE):
+//
+// enum Color { BLACK, GRAY, WHITE };
+// class AnimalTest
+//     : public testing::TestWithParam<tuple<const char*, Color> > {...};
+//
+// TEST_P(AnimalTest, AnimalLooksNice) {...}
+//
+// INSTANTIATE_TEST_CASE_P(AnimalVariations, AnimalTest,
+//                         Combine(Values("cat", "dog"),
+//                                 Values(BLACK, WHITE)));
+//
+// This will instantiate tests in FlagDependentTest with all variations of two
+// Boolean flags:
+//
+// class FlagDependentTest
+//     : public testing::TestWithParam<tuple<bool, bool> > {
+//   virtual void SetUp() {
+//     // Assigns external_flag_1 and external_flag_2 values from the tuple.
+//     tie(external_flag_1, external_flag_2) = GetParam();
+//   }
+// };
+//
+// TEST_P(FlagDependentTest, TestFeature1) {
+//   // Test your code using external_flag_1 and external_flag_2 here.
+// }
+// INSTANTIATE_TEST_CASE_P(TwoBoolSequence, FlagDependentTest,
+//                         Combine(Bool(), Bool()));
+//
+template <typename Generator1, typename Generator2>
+internal::CartesianProductHolder2<Generator1, Generator2> Combine(
+    const Generator1& g1, const Generator2& g2) {
+  return internal::CartesianProductHolder2<Generator1, Generator2>(
+      g1, g2);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3>
+internal::CartesianProductHolder3<Generator1, Generator2, Generator3> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3) {
+  return internal::CartesianProductHolder3<Generator1, Generator2, Generator3>(
+      g1, g2, g3);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4>
+internal::CartesianProductHolder4<Generator1, Generator2, Generator3,
+    Generator4> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4) {
+  return internal::CartesianProductHolder4<Generator1, Generator2, Generator3,
+      Generator4>(
+      g1, g2, g3, g4);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5>
+internal::CartesianProductHolder5<Generator1, Generator2, Generator3,
+    Generator4, Generator5> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5) {
+  return internal::CartesianProductHolder5<Generator1, Generator2, Generator3,
+      Generator4, Generator5>(
+      g1, g2, g3, g4, g5);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5, typename Generator6>
+internal::CartesianProductHolder6<Generator1, Generator2, Generator3,
+    Generator4, Generator5, Generator6> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6) {
+  return internal::CartesianProductHolder6<Generator1, Generator2, Generator3,
+      Generator4, Generator5, Generator6>(
+      g1, g2, g3, g4, g5, g6);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5, typename Generator6,
+    typename Generator7>
+internal::CartesianProductHolder7<Generator1, Generator2, Generator3,
+    Generator4, Generator5, Generator6, Generator7> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7) {
+  return internal::CartesianProductHolder7<Generator1, Generator2, Generator3,
+      Generator4, Generator5, Generator6, Generator7>(
+      g1, g2, g3, g4, g5, g6, g7);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5, typename Generator6,
+    typename Generator7, typename Generator8>
+internal::CartesianProductHolder8<Generator1, Generator2, Generator3,
+    Generator4, Generator5, Generator6, Generator7, Generator8> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7, const Generator8& g8) {
+  return internal::CartesianProductHolder8<Generator1, Generator2, Generator3,
+      Generator4, Generator5, Generator6, Generator7, Generator8>(
+      g1, g2, g3, g4, g5, g6, g7, g8);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5, typename Generator6,
+    typename Generator7, typename Generator8, typename Generator9>
+internal::CartesianProductHolder9<Generator1, Generator2, Generator3,
+    Generator4, Generator5, Generator6, Generator7, Generator8,
+    Generator9> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7, const Generator8& g8, const Generator9& g9) {
+  return internal::CartesianProductHolder9<Generator1, Generator2, Generator3,
+      Generator4, Generator5, Generator6, Generator7, Generator8, Generator9>(
+      g1, g2, g3, g4, g5, g6, g7, g8, g9);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5, typename Generator6,
+    typename Generator7, typename Generator8, typename Generator9,
+    typename Generator10>
+internal::CartesianProductHolder10<Generator1, Generator2, Generator3,
+    Generator4, Generator5, Generator6, Generator7, Generator8, Generator9,
+    Generator10> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7, const Generator8& g8, const Generator9& g9,
+        const Generator10& g10) {
+  return internal::CartesianProductHolder10<Generator1, Generator2, Generator3,
+      Generator4, Generator5, Generator6, Generator7, Generator8, Generator9,
+      Generator10>(
+      g1, g2, g3, g4, g5, g6, g7, g8, g9, g10);
+}
+# endif  // GTEST_HAS_COMBINE
+
+
+
+# define TEST_P(test_case_name, test_name) \
+  class GTEST_TEST_CLASS_NAME_(test_case_name, test_name) \
+      : public test_case_name { \
+   public: \
+    GTEST_TEST_CLASS_NAME_(test_case_name, test_name)() {} \
+    virtual void TestBody(); \
+   private: \
+    static int AddToRegistry() { \
+      ::testing::UnitTest::GetInstance()->parameterized_test_registry(). \
+          GetTestCasePatternHolder<test_case_name>(\
+              #test_case_name, __FILE__, __LINE__)->AddTestPattern(\
+                  #test_case_name, \
+                  #test_name, \
+                  new ::testing::internal::TestMetaFactory< \
+                      GTEST_TEST_CLASS_NAME_(test_case_name, test_name)>()); \
+      return 0; \
+    } \
+    static int gtest_registering_dummy_ GTEST_ATTRIBUTE_UNUSED_; \
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(\
+        GTEST_TEST_CLASS_NAME_(test_case_name, test_name)); \
+  }; \
+  int GTEST_TEST_CLASS_NAME_(test_case_name, \
+                             test_name)::gtest_registering_dummy_ = \
+      GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::AddToRegistry(); \
+  void GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::TestBody()
+
+# define INSTANTIATE_TEST_CASE_P(prefix, test_case_name, generator) \
+  ::testing::internal::ParamGenerator<test_case_name::ParamType> \
+      gtest_##prefix##test_case_name##_EvalGenerator_() { return generator; } \
+  int gtest_##prefix##test_case_name##_dummy_ GTEST_ATTRIBUTE_UNUSED_ = \
+      ::testing::UnitTest::GetInstance()->parameterized_test_registry(). \
+          GetTestCasePatternHolder<test_case_name>(\
+              #test_case_name, __FILE__, __LINE__)->AddTestCaseInstantiation(\
+                  #prefix, \
+                  &gtest_##prefix##test_case_name##_EvalGenerator_, \
+                  __FILE__, __LINE__)
+
+}  // namespace testing
+
+#endif  // GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
+// Copyright 2006, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// Google C++ Testing Framework definitions useful in production code.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PROD_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PROD_H_
+
+// When you need to test the private or protected members of a class,
+// use the FRIEND_TEST macro to declare your tests as friends of the
+// class.  For example:
+//
+// class MyClass {
+//  private:
+//   void MyMethod();
+//   FRIEND_TEST(MyClassTest, MyMethod);
+// };
+//
+// class MyClassTest : public testing::Test {
+//   // ...
+// };
+//
+// TEST_F(MyClassTest, MyMethod) {
+//   // Can call MyClass::MyMethod() here.
+// }
+
+#define FRIEND_TEST(test_case_name, test_name)\
+friend class test_case_name##_##test_name##_Test
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PROD_H_
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: mheule@google.com (Markus Heule)
+//
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_TEST_PART_H_
+#define GTEST_INCLUDE_GTEST_GTEST_TEST_PART_H_
+
+#include <iosfwd>
+#include <vector>
+
+namespace testing {
+
+// A copyable object representing the result of a test part (i.e. an
+// assertion or an explicit FAIL(), ADD_FAILURE(), or SUCCESS()).
+//
+// Don't inherit from TestPartResult as its destructor is not virtual.
+class GTEST_API_ TestPartResult {
+ public:
+  // The possible outcomes of a test part (i.e. an assertion or an
+  // explicit SUCCEED(), FAIL(), or ADD_FAILURE()).
+  enum Type {
+    kSuccess,          // Succeeded.
+    kNonFatalFailure,  // Failed but the test can continue.
+    kFatalFailure      // Failed and the test should be terminated.
+  };
+
+  // C'tor.  TestPartResult does NOT have a default constructor.
+  // Always use this constructor (with parameters) to create a
+  // TestPartResult object.
+  TestPartResult(Type a_type,
+                 const char* a_file_name,
+                 int a_line_number,
+                 const char* a_message)
+      : type_(a_type),
+        file_name_(a_file_name == NULL ? "" : a_file_name),
+        line_number_(a_line_number),
+        summary_(ExtractSummary(a_message)),
+        message_(a_message) {
+  }
+
+  // Gets the outcome of the test part.
+  Type type() const { return type_; }
+
+  // Gets the name of the source file where the test part took place, or
+  // NULL if it's unknown.
+  const char* file_name() const {
+    return file_name_.empty() ? NULL : file_name_.c_str();
+  }
+
+  // Gets the line in the source file where the test part took place,
+  // or -1 if it's unknown.
+  int line_number() const { return line_number_; }
+
+  // Gets the summary of the failure message.
+  const char* summary() const { return summary_.c_str(); }
+
+  // Gets the message associated with the test part.
+  const char* message() const { return message_.c_str(); }
+
+  // Returns true iff the test part passed.
+  bool passed() const { return type_ == kSuccess; }
+
+  // Returns true iff the test part failed.
+  bool failed() const { return type_ != kSuccess; }
+
+  // Returns true iff the test part non-fatally failed.
+  bool nonfatally_failed() const { return type_ == kNonFatalFailure; }
+
+  // Returns true iff the test part fatally failed.
+  bool fatally_failed() const { return type_ == kFatalFailure; }
+
+ private:
+  Type type_;
+
+  // Gets the summary of the failure message by omitting the stack
+  // trace in it.
+  static std::string ExtractSummary(const char* message);
+
+  // The name of the source file where the test part took place, or
+  // "" if the source file is unknown.
+  std::string file_name_;
+  // The line in the source file where the test part took place, or -1
+  // if the line number is unknown.
+  int line_number_;
+  std::string summary_;  // The test failure summary.
+  std::string message_;  // The test failure message.
+};
+
+// Prints a TestPartResult object.
+std::ostream& operator<<(std::ostream& os, const TestPartResult& result);
+
+// An array of TestPartResult objects.
+//
+// Don't inherit from TestPartResultArray as its destructor is not
+// virtual.
+class GTEST_API_ TestPartResultArray {
+ public:
+  TestPartResultArray() {}
+
+  // Appends the given TestPartResult to the array.
+  void Append(const TestPartResult& result);
+
+  // Returns the TestPartResult at the given index (0-based).
+  const TestPartResult& GetTestPartResult(int index) const;
+
+  // Returns the number of TestPartResult objects in the array.
+  int size() const;
+
+ private:
+  std::vector<TestPartResult> array_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestPartResultArray);
+};
+
+// This interface knows how to report a test part result.
+class TestPartResultReporterInterface {
+ public:
+  virtual ~TestPartResultReporterInterface() {}
+
+  virtual void ReportTestPartResult(const TestPartResult& result) = 0;
+};
+
+namespace internal {
+
+// This helper class is used by {ASSERT|EXPECT}_NO_FATAL_FAILURE to check if a
+// statement generates new fatal failures. To do so it registers itself as the
+// current test part result reporter. Besides checking if fatal failures were
+// reported, it only delegates the reporting to the former result reporter.
+// The original result reporter is restored in the destructor.
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+class GTEST_API_ HasNewFatalFailureHelper
+    : public TestPartResultReporterInterface {
+ public:
+  HasNewFatalFailureHelper();
+  virtual ~HasNewFatalFailureHelper();
+  virtual void ReportTestPartResult(const TestPartResult& result);
+  bool has_new_fatal_failure() const { return has_new_fatal_failure_; }
+ private:
+  bool has_new_fatal_failure_;
+  TestPartResultReporterInterface* original_reporter_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(HasNewFatalFailureHelper);
+};
+
+}  // namespace internal
+
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_TEST_PART_H_
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_TYPED_TEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_TYPED_TEST_H_
+
+// This header implements typed tests and type-parameterized tests.
+
+// Typed (aka type-driven) tests repeat the same test for types in a
+// list.  You must know which types you want to test with when writing
+// typed tests. Here's how you do it:
+
+#if 0
+
+// First, define a fixture class template.  It should be parameterized
+// by a type.  Remember to derive it from testing::Test.
+template <typename T>
+class FooTest : public testing::Test {
+ public:
+  ...
+  typedef std::list<T> List;
+  static T shared_;
+  T value_;
+};
+
+// Next, associate a list of types with the test case, which will be
+// repeated for each type in the list.  The typedef is necessary for
+// the macro to parse correctly.
+typedef testing::Types<char, int, unsigned int> MyTypes;
+TYPED_TEST_CASE(FooTest, MyTypes);
+
+// If the type list contains only one type, you can write that type
+// directly without Types<...>:
+//   TYPED_TEST_CASE(FooTest, int);
+
+// Then, use TYPED_TEST() instead of TEST_F() to define as many typed
+// tests for this test case as you want.
+TYPED_TEST(FooTest, DoesBlah) {
+  // Inside a test, refer to TypeParam to get the type parameter.
+  // Since we are inside a derived class template, C++ requires use to
+  // visit the members of FooTest via 'this'.
+  TypeParam n = this->value_;
+
+  // To visit static members of the fixture, add the TestFixture::
+  // prefix.
+  n += TestFixture::shared_;
+
+  // To refer to typedefs in the fixture, add the "typename
+  // TestFixture::" prefix.
+  typename TestFixture::List values;
+  values.push_back(n);
+  ...
+}
+
+TYPED_TEST(FooTest, HasPropertyA) { ... }
+
+#endif  // 0
+
+// Type-parameterized tests are abstract test patterns parameterized
+// by a type.  Compared with typed tests, type-parameterized tests
+// allow you to define the test pattern without knowing what the type
+// parameters are.  The defined pattern can be instantiated with
+// different types any number of times, in any number of translation
+// units.
+//
+// If you are designing an interface or concept, you can define a
+// suite of type-parameterized tests to verify properties that any
+// valid implementation of the interface/concept should have.  Then,
+// each implementation can easily instantiate the test suite to verify
+// that it conforms to the requirements, without having to write
+// similar tests repeatedly.  Here's an example:
+
+#if 0
+
+// First, define a fixture class template.  It should be parameterized
+// by a type.  Remember to derive it from testing::Test.
+template <typename T>
+class FooTest : public testing::Test {
+  ...
+};
+
+// Next, declare that you will define a type-parameterized test case
+// (the _P suffix is for "parameterized" or "pattern", whichever you
+// prefer):
+TYPED_TEST_CASE_P(FooTest);
+
+// Then, use TYPED_TEST_P() to define as many type-parameterized tests
+// for this type-parameterized test case as you want.
+TYPED_TEST_P(FooTest, DoesBlah) {
+  // Inside a test, refer to TypeParam to get the type parameter.
+  TypeParam n = 0;
+  ...
+}
+
+TYPED_TEST_P(FooTest, HasPropertyA) { ... }
+
+// Now the tricky part: you need to register all test patterns before
+// you can instantiate them.  The first argument of the macro is the
+// test case name; the rest are the names of the tests in this test
+// case.
+REGISTER_TYPED_TEST_CASE_P(FooTest,
+                           DoesBlah, HasPropertyA);
+
+// Finally, you are free to instantiate the pattern with the types you
+// want.  If you put the above code in a header file, you can #include
+// it in multiple C++ source files and instantiate it multiple times.
+//
+// To distinguish different instances of the pattern, the first
+// argument to the INSTANTIATE_* macro is a prefix that will be added
+// to the actual test case name.  Remember to pick unique prefixes for
+// different instances.
+typedef testing::Types<char, int, unsigned int> MyTypes;
+INSTANTIATE_TYPED_TEST_CASE_P(My, FooTest, MyTypes);
+
+// If the type list contains only one type, you can write that type
+// directly without Types<...>:
+//   INSTANTIATE_TYPED_TEST_CASE_P(My, FooTest, int);
+
+#endif  // 0
+
+
+// Implements typed tests.
+
+#if GTEST_HAS_TYPED_TEST
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Expands to the name of the typedef for the type parameters of the
+// given test case.
+# define GTEST_TYPE_PARAMS_(TestCaseName) gtest_type_params_##TestCaseName##_
+
+// The 'Types' template argument below must have spaces around it
+// since some compilers may choke on '>>' when passing a template
+// instance (e.g. Types<int>)
+# define TYPED_TEST_CASE(CaseName, Types) \
+  typedef ::testing::internal::TypeList< Types >::type \
+      GTEST_TYPE_PARAMS_(CaseName)
+
+# define TYPED_TEST(CaseName, TestName) \
+  template <typename gtest_TypeParam_> \
+  class GTEST_TEST_CLASS_NAME_(CaseName, TestName) \
+      : public CaseName<gtest_TypeParam_> { \
+   private: \
+    typedef CaseName<gtest_TypeParam_> TestFixture; \
+    typedef gtest_TypeParam_ TypeParam; \
+    virtual void TestBody(); \
+  }; \
+  bool gtest_##CaseName##_##TestName##_registered_ GTEST_ATTRIBUTE_UNUSED_ = \
+      ::testing::internal::TypeParameterizedTest< \
+          CaseName, \
+          ::testing::internal::TemplateSel< \
+              GTEST_TEST_CLASS_NAME_(CaseName, TestName)>, \
+          GTEST_TYPE_PARAMS_(CaseName)>::Register(\
+              "", #CaseName, #TestName, 0); \
+  template <typename gtest_TypeParam_> \
+  void GTEST_TEST_CLASS_NAME_(CaseName, TestName)<gtest_TypeParam_>::TestBody()
+
+#endif  // GTEST_HAS_TYPED_TEST
+
+// Implements type-parameterized tests.
+
+#if GTEST_HAS_TYPED_TEST_P
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Expands to the namespace name that the type-parameterized tests for
+// the given type-parameterized test case are defined in.  The exact
+// name of the namespace is subject to change without notice.
+# define GTEST_CASE_NAMESPACE_(TestCaseName) \
+  gtest_case_##TestCaseName##_
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Expands to the name of the variable used to remember the names of
+// the defined tests in the given test case.
+# define GTEST_TYPED_TEST_CASE_P_STATE_(TestCaseName) \
+  gtest_typed_test_case_p_state_##TestCaseName##_
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE DIRECTLY.
+//
+// Expands to the name of the variable used to remember the names of
+// the registered tests in the given test case.
+# define GTEST_REGISTERED_TEST_NAMES_(TestCaseName) \
+  gtest_registered_test_names_##TestCaseName##_
+
+// The variables defined in the type-parameterized test macros are
+// static as typically these macros are used in a .h file that can be
+// #included in multiple translation units linked together.
+# define TYPED_TEST_CASE_P(CaseName) \
+  static ::testing::internal::TypedTestCasePState \
+      GTEST_TYPED_TEST_CASE_P_STATE_(CaseName)
+
+# define TYPED_TEST_P(CaseName, TestName) \
+  namespace GTEST_CASE_NAMESPACE_(CaseName) { \
+  template <typename gtest_TypeParam_> \
+  class TestName : public CaseName<gtest_TypeParam_> { \
+   private: \
+    typedef CaseName<gtest_TypeParam_> TestFixture; \
+    typedef gtest_TypeParam_ TypeParam; \
+    virtual void TestBody(); \
+  }; \
+  static bool gtest_##TestName##_defined_ GTEST_ATTRIBUTE_UNUSED_ = \
+      GTEST_TYPED_TEST_CASE_P_STATE_(CaseName).AddTestName(\
+          __FILE__, __LINE__, #CaseName, #TestName); \
+  } \
+  template <typename gtest_TypeParam_> \
+  void GTEST_CASE_NAMESPACE_(CaseName)::TestName<gtest_TypeParam_>::TestBody()
+
+# define REGISTER_TYPED_TEST_CASE_P(CaseName, ...) \
+  namespace GTEST_CASE_NAMESPACE_(CaseName) { \
+  typedef ::testing::internal::Templates<__VA_ARGS__>::type gtest_AllTests_; \
+  } \
+  static const char* const GTEST_REGISTERED_TEST_NAMES_(CaseName) = \
+      GTEST_TYPED_TEST_CASE_P_STATE_(CaseName).VerifyRegisteredTestNames(\
+          __FILE__, __LINE__, #__VA_ARGS__)
+
+// The 'Types' template argument below must have spaces around it
+// since some compilers may choke on '>>' when passing a template
+// instance (e.g. Types<int>)
+# define INSTANTIATE_TYPED_TEST_CASE_P(Prefix, CaseName, Types) \
+  bool gtest_##Prefix##_##CaseName GTEST_ATTRIBUTE_UNUSED_ = \
+      ::testing::internal::TypeParameterizedTestCase<CaseName, \
+          GTEST_CASE_NAMESPACE_(CaseName)::gtest_AllTests_, \
+          ::testing::internal::TypeList< Types >::type>::Register(\
+              #Prefix, #CaseName, GTEST_REGISTERED_TEST_NAMES_(CaseName))
+
+#endif  // GTEST_HAS_TYPED_TEST_P
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_TYPED_TEST_H_
+
+// Depending on the platform, different string classes are available.
+// On Linux, in addition to ::std::string, Google also makes use of
+// class ::string, which has the same interface as ::std::string, but
+// has a different implementation.
+//
+// The user can define GTEST_HAS_GLOBAL_STRING to 1 to indicate that
+// ::string is available AND is a distinct type to ::std::string, or
+// define it to 0 to indicate otherwise.
+//
+// If the user's ::std::string and ::string are the same class due to
+// aliasing, he should define GTEST_HAS_GLOBAL_STRING to 0.
+//
+// If the user doesn't define GTEST_HAS_GLOBAL_STRING, it is defined
+// heuristically.
+
+namespace testing {
+
+// Declares the flags.
+
+// This flag temporary enables the disabled tests.
+GTEST_DECLARE_bool_(also_run_disabled_tests);
+
+// This flag brings the debugger on an assertion failure.
+GTEST_DECLARE_bool_(break_on_failure);
+
+// This flag controls whether Google Test catches all test-thrown exceptions
+// and logs them as failures.
+GTEST_DECLARE_bool_(catch_exceptions);
+
+// This flag enables using colors in terminal output. Available values are
+// "yes" to enable colors, "no" (disable colors), or "auto" (the default)
+// to let Google Test decide.
+GTEST_DECLARE_string_(color);
+
+// This flag sets up the filter to select by name using a glob pattern
+// the tests to run. If the filter is not given all tests are executed.
+GTEST_DECLARE_string_(filter);
+
+// This flag causes the Google Test to list tests. None of the tests listed
+// are actually run if the flag is provided.
+GTEST_DECLARE_bool_(list_tests);
+
+// This flag controls whether Google Test emits a detailed XML report to a file
+// in addition to its normal textual output.
+GTEST_DECLARE_string_(output);
+
+// This flags control whether Google Test prints the elapsed time for each
+// test.
+GTEST_DECLARE_bool_(print_time);
+
+// This flag specifies the random number seed.
+GTEST_DECLARE_int32_(random_seed);
+
+// This flag sets how many times the tests are repeated. The default value
+// is 1. If the value is -1 the tests are repeating forever.
+GTEST_DECLARE_int32_(repeat);
+
+// This flag controls whether Google Test includes Google Test internal
+// stack frames in failure stack traces.
+GTEST_DECLARE_bool_(show_internal_stack_frames);
+
+// When this flag is specified, tests' order is randomized on every iteration.
+GTEST_DECLARE_bool_(shuffle);
+
+// This flag specifies the maximum number of stack frames to be
+// printed in a failure message.
+GTEST_DECLARE_int32_(stack_trace_depth);
+
+// When this flag is specified, a failed assertion will throw an
+// exception if exceptions are enabled, or exit the program with a
+// non-zero code otherwise.
+GTEST_DECLARE_bool_(throw_on_failure);
+
+// When this flag is set with a "host:port" string, on supported
+// platforms test results are streamed to the specified port on
+// the specified host machine.
+GTEST_DECLARE_string_(stream_result_to);
+
+// The upper limit for valid stack trace depths.
+const int kMaxStackTraceDepth = 100;
+
+namespace internal {
+
+class AssertHelper;
+class DefaultGlobalTestPartResultReporter;
+class ExecDeathTest;
+class NoExecDeathTest;
+class FinalSuccessChecker;
+class GTestFlagSaver;
+class StreamingListenerTest;
+class TestResultAccessor;
+class TestEventListenersAccessor;
+class TestEventRepeater;
+class UnitTestRecordPropertyTestHelper;
+class WindowsDeathTest;
+class UnitTestImpl* GetUnitTestImpl();
+void ReportFailureInUnknownLocation(TestPartResult::Type result_type,
+                                    const std::string& message);
+
+}  // namespace internal
+
+// The friend relationship of some of these classes is cyclic.
+// If we don't forward declare them the compiler might confuse the classes
+// in friendship clauses with same named classes on the scope.
+class Test;
+class TestCase;
+class TestInfo;
+class UnitTest;
+
+// A class for indicating whether an assertion was successful.  When
+// the assertion wasn't successful, the AssertionResult object
+// remembers a non-empty message that describes how it failed.
+//
+// To create an instance of this class, use one of the factory functions
+// (AssertionSuccess() and AssertionFailure()).
+//
+// This class is useful for two purposes:
+//   1. Defining predicate functions to be used with Boolean test assertions
+//      EXPECT_TRUE/EXPECT_FALSE and their ASSERT_ counterparts
+//   2. Defining predicate-format functions to be
+//      used with predicate assertions (ASSERT_PRED_FORMAT*, etc).
+//
+// For example, if you define IsEven predicate:
+//
+//   testing::AssertionResult IsEven(int n) {
+//     if ((n % 2) == 0)
+//       return testing::AssertionSuccess();
+//     else
+//       return testing::AssertionFailure() << n << " is odd";
+//   }
+//
+// Then the failed expectation EXPECT_TRUE(IsEven(Fib(5)))
+// will print the message
+//
+//   Value of: IsEven(Fib(5))
+//     Actual: false (5 is odd)
+//   Expected: true
+//
+// instead of a more opaque
+//
+//   Value of: IsEven(Fib(5))
+//     Actual: false
+//   Expected: true
+//
+// in case IsEven is a simple Boolean predicate.
+//
+// If you expect your predicate to be reused and want to support informative
+// messages in EXPECT_FALSE and ASSERT_FALSE (negative assertions show up
+// about half as often as positive ones in our tests), supply messages for
+// both success and failure cases:
+//
+//   testing::AssertionResult IsEven(int n) {
+//     if ((n % 2) == 0)
+//       return testing::AssertionSuccess() << n << " is even";
+//     else
+//       return testing::AssertionFailure() << n << " is odd";
+//   }
+//
+// Then a statement EXPECT_FALSE(IsEven(Fib(6))) will print
+//
+//   Value of: IsEven(Fib(6))
+//     Actual: true (8 is even)
+//   Expected: false
+//
+// NB: Predicates that support negative Boolean assertions have reduced
+// performance in positive ones so be careful not to use them in tests
+// that have lots (tens of thousands) of positive Boolean assertions.
+//
+// To use this class with EXPECT_PRED_FORMAT assertions such as:
+//
+//   // Verifies that Foo() returns an even number.
+//   EXPECT_PRED_FORMAT1(IsEven, Foo());
+//
+// you need to define:
+//
+//   testing::AssertionResult IsEven(const char* expr, int n) {
+//     if ((n % 2) == 0)
+//       return testing::AssertionSuccess();
+//     else
+//       return testing::AssertionFailure()
+//         << "Expected: " << expr << " is even\n  Actual: it's " << n;
+//   }
+//
+// If Foo() returns 5, you will see the following message:
+//
+//   Expected: Foo() is even
+//     Actual: it's 5
+//
+class GTEST_API_ AssertionResult {
+ public:
+  // Copy constructor.
+  // Used in EXPECT_TRUE/FALSE(assertion_result).
+  AssertionResult(const AssertionResult& other);
+  // Used in the EXPECT_TRUE/FALSE(bool_expression).
+  explicit AssertionResult(bool success) : success_(success) {}
+
+  // Returns true iff the assertion succeeded.
+  operator bool() const { return success_; }  // NOLINT
+
+  // Returns the assertion's negation. Used with EXPECT/ASSERT_FALSE.
+  AssertionResult operator!() const;
+
+  // Returns the text streamed into this AssertionResult. Test assertions
+  // use it when they fail (i.e., the predicate's outcome doesn't match the
+  // assertion's expectation). When nothing has been streamed into the
+  // object, returns an empty string.
+  const char* message() const {
+    return message_.get() != NULL ?  message_->c_str() : "";
+  }
+  // TODO(vladl@google.com): Remove this after making sure no clients use it.
+  // Deprecated; please use message() instead.
+  const char* failure_message() const { return message(); }
+
+  // Streams a custom failure message into this object.
+  template <typename T> AssertionResult& operator<<(const T& value) {
+    AppendMessage(Message() << value);
+    return *this;
+  }
+
+  // Allows streaming basic output manipulators such as endl or flush into
+  // this object.
+  AssertionResult& operator<<(
+      ::std::ostream& (*basic_manipulator)(::std::ostream& stream)) {
+    AppendMessage(Message() << basic_manipulator);
+    return *this;
+  }
+
+ private:
+  // Appends the contents of message to message_.
+  void AppendMessage(const Message& a_message) {
+    if (message_.get() == NULL)
+      message_.reset(new ::std::string);
+    message_->append(a_message.GetString().c_str());
+  }
+
+  // Stores result of the assertion predicate.
+  bool success_;
+  // Stores the message describing the condition in case the expectation
+  // construct is not satisfied with the predicate's outcome.
+  // Referenced via a pointer to avoid taking too much stack frame space
+  // with test assertions.
+  internal::scoped_ptr< ::std::string> message_;
+
+  GTEST_DISALLOW_ASSIGN_(AssertionResult);
+};
+
+// Makes a successful assertion result.
+GTEST_API_ AssertionResult AssertionSuccess();
+
+// Makes a failed assertion result.
+GTEST_API_ AssertionResult AssertionFailure();
+
+// Makes a failed assertion result with the given failure message.
+// Deprecated; use AssertionFailure() << msg.
+GTEST_API_ AssertionResult AssertionFailure(const Message& msg);
+
+// The abstract class that all tests inherit from.
+//
+// In Google Test, a unit test program contains one or many TestCases, and
+// each TestCase contains one or many Tests.
+//
+// When you define a test using the TEST macro, you don't need to
+// explicitly derive from Test - the TEST macro automatically does
+// this for you.
+//
+// The only time you derive from Test is when defining a test fixture
+// to be used a TEST_F.  For example:
+//
+//   class FooTest : public testing::Test {
+//    protected:
+//     virtual void SetUp() { ... }
+//     virtual void TearDown() { ... }
+//     ...
+//   };
+//
+//   TEST_F(FooTest, Bar) { ... }
+//   TEST_F(FooTest, Baz) { ... }
+//
+// Test is not copyable.
+class GTEST_API_ Test {
+ public:
+  friend class TestInfo;
+
+  // Defines types for pointers to functions that set up and tear down
+  // a test case.
+  typedef internal::SetUpTestCaseFunc SetUpTestCaseFunc;
+  typedef internal::TearDownTestCaseFunc TearDownTestCaseFunc;
+
+  // The d'tor is virtual as we intend to inherit from Test.
+  virtual ~Test();
+
+  // Sets up the stuff shared by all tests in this test case.
+  //
+  // Google Test will call Foo::SetUpTestCase() before running the first
+  // test in test case Foo.  Hence a sub-class can define its own
+  // SetUpTestCase() method to shadow the one defined in the super
+  // class.
+  static void SetUpTestCase() {}
+
+  // Tears down the stuff shared by all tests in this test case.
+  //
+  // Google Test will call Foo::TearDownTestCase() after running the last
+  // test in test case Foo.  Hence a sub-class can define its own
+  // TearDownTestCase() method to shadow the one defined in the super
+  // class.
+  static void TearDownTestCase() {}
+
+  // Returns true iff the current test has a fatal failure.
+  static bool HasFatalFailure();
+
+  // Returns true iff the current test has a non-fatal failure.
+  static bool HasNonfatalFailure();
+
+  // Returns true iff the current test has a (either fatal or
+  // non-fatal) failure.
+  static bool HasFailure() { return HasFatalFailure() || HasNonfatalFailure(); }
+
+  // Logs a property for the current test, test case, or for the entire
+  // invocation of the test program when used outside of the context of a
+  // test case.  Only the last value for a given key is remembered.  These
+  // are public static so they can be called from utility functions that are
+  // not members of the test fixture.  Calls to RecordProperty made during
+  // lifespan of the test (from the moment its constructor starts to the
+  // moment its destructor finishes) will be output in XML as attributes of
+  // the <testcase> element.  Properties recorded from fixture's
+  // SetUpTestCase or TearDownTestCase are logged as attributes of the
+  // corresponding <testsuite> element.  Calls to RecordProperty made in the
+  // global context (before or after invocation of RUN_ALL_TESTS and from
+  // SetUp/TearDown method of Environment objects registered with Google
+  // Test) will be output as attributes of the <testsuites> element.
+  static void RecordProperty(const std::string& key, const std::string& value);
+  static void RecordProperty(const std::string& key, int value);
+
+ protected:
+  // Creates a Test object.
+  Test();
+
+  // Sets up the test fixture.
+  virtual void SetUp();
+
+  // Tears down the test fixture.
+  virtual void TearDown();
+
+ private:
+  // Returns true iff the current test has the same fixture class as
+  // the first test in the current test case.
+  static bool HasSameFixtureClass();
+
+  // Runs the test after the test fixture has been set up.
+  //
+  // A sub-class must implement this to define the test logic.
+  //
+  // DO NOT OVERRIDE THIS FUNCTION DIRECTLY IN A USER PROGRAM.
+  // Instead, use the TEST or TEST_F macro.
+  virtual void TestBody() = 0;
+
+  // Sets up, executes, and tears down the test.
+  void Run();
+
+  // Deletes self.  We deliberately pick an unusual name for this
+  // internal method to avoid clashing with names used in user TESTs.
+  void DeleteSelf_() { delete this; }
+
+  // Uses a GTestFlagSaver to save and restore all Google Test flags.
+  const internal::GTestFlagSaver* const gtest_flag_saver_;
+
+  // Often a user mis-spells SetUp() as Setup() and spends a long time
+  // wondering why it is never called by Google Test.  The declaration of
+  // the following method is solely for catching such an error at
+  // compile time:
+  //
+  //   - The return type is deliberately chosen to be not void, so it
+  //   will be a conflict if a user declares void Setup() in his test
+  //   fixture.
+  //
+  //   - This method is private, so it will be another compiler error
+  //   if a user calls it from his test fixture.
+  //
+  // DO NOT OVERRIDE THIS FUNCTION.
+  //
+  // If you see an error about overriding the following function or
+  // about it being private, you have mis-spelled SetUp() as Setup().
+  struct Setup_should_be_spelled_SetUp {};
+  virtual Setup_should_be_spelled_SetUp* Setup() { return NULL; }
+
+  // We disallow copying Tests.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Test);
+};
+
+typedef internal::TimeInMillis TimeInMillis;
+
+// A copyable object representing a user specified test property which can be
+// output as a key/value string pair.
+//
+// Don't inherit from TestProperty as its destructor is not virtual.
+class TestProperty {
+ public:
+  // C'tor.  TestProperty does NOT have a default constructor.
+  // Always use this constructor (with parameters) to create a
+  // TestProperty object.
+  TestProperty(const std::string& a_key, const std::string& a_value) :
+    key_(a_key), value_(a_value) {
+  }
+
+  // Gets the user supplied key.
+  const char* key() const {
+    return key_.c_str();
+  }
+
+  // Gets the user supplied value.
+  const char* value() const {
+    return value_.c_str();
+  }
+
+  // Sets a new value, overriding the one supplied in the constructor.
+  void SetValue(const std::string& new_value) {
+    value_ = new_value;
+  }
+
+ private:
+  // The key supplied by the user.
+  std::string key_;
+  // The value supplied by the user.
+  std::string value_;
+};
+
+// The result of a single Test.  This includes a list of
+// TestPartResults, a list of TestProperties, a count of how many
+// death tests there are in the Test, and how much time it took to run
+// the Test.
+//
+// TestResult is not copyable.
+class GTEST_API_ TestResult {
+ public:
+  // Creates an empty TestResult.
+  TestResult();
+
+  // D'tor.  Do not inherit from TestResult.
+  ~TestResult();
+
+  // Gets the number of all test parts.  This is the sum of the number
+  // of successful test parts and the number of failed test parts.
+  int total_part_count() const;
+
+  // Returns the number of the test properties.
+  int test_property_count() const;
+
+  // Returns true iff the test passed (i.e. no test part failed).
+  bool Passed() const { return !Failed(); }
+
+  // Returns true iff the test failed.
+  bool Failed() const;
+
+  // Returns true iff the test fatally failed.
+  bool HasFatalFailure() const;
+
+  // Returns true iff the test has a non-fatal failure.
+  bool HasNonfatalFailure() const;
+
+  // Returns the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const { return elapsed_time_; }
+
+  // Returns the i-th test part result among all the results. i can range
+  // from 0 to test_property_count() - 1. If i is not in that range, aborts
+  // the program.
+  const TestPartResult& GetTestPartResult(int i) const;
+
+  // Returns the i-th test property. i can range from 0 to
+  // test_property_count() - 1. If i is not in that range, aborts the
+  // program.
+  const TestProperty& GetTestProperty(int i) const;
+
+ private:
+  friend class TestInfo;
+  friend class TestCase;
+  friend class UnitTest;
+  friend class internal::DefaultGlobalTestPartResultReporter;
+  friend class internal::ExecDeathTest;
+  friend class internal::TestResultAccessor;
+  friend class internal::UnitTestImpl;
+  friend class internal::WindowsDeathTest;
+
+  // Gets the vector of TestPartResults.
+  const std::vector<TestPartResult>& test_part_results() const {
+    return test_part_results_;
+  }
+
+  // Gets the vector of TestProperties.
+  const std::vector<TestProperty>& test_properties() const {
+    return test_properties_;
+  }
+
+  // Sets the elapsed time.
+  void set_elapsed_time(TimeInMillis elapsed) { elapsed_time_ = elapsed; }
+
+  // Adds a test property to the list. The property is validated and may add
+  // a non-fatal failure if invalid (e.g., if it conflicts with reserved
+  // key names). If a property is already recorded for the same key, the
+  // value will be updated, rather than storing multiple values for the same
+  // key.  xml_element specifies the element for which the property is being
+  // recorded and is used for validation.
+  void RecordProperty(const std::string& xml_element,
+                      const TestProperty& test_property);
+
+  // Adds a failure if the key is a reserved attribute of Google Test
+  // testcase tags.  Returns true if the property is valid.
+  // TODO(russr): Validate attribute names are legal and human readable.
+  static bool ValidateTestProperty(const std::string& xml_element,
+                                   const TestProperty& test_property);
+
+  // Adds a test part result to the list.
+  void AddTestPartResult(const TestPartResult& test_part_result);
+
+  // Returns the death test count.
+  int death_test_count() const { return death_test_count_; }
+
+  // Increments the death test count, returning the new count.
+  int increment_death_test_count() { return ++death_test_count_; }
+
+  // Clears the test part results.
+  void ClearTestPartResults();
+
+  // Clears the object.
+  void Clear();
+
+  // Protects mutable state of the property vector and of owned
+  // properties, whose values may be updated.
+  internal::Mutex test_properites_mutex_;
+
+  // The vector of TestPartResults
+  std::vector<TestPartResult> test_part_results_;
+  // The vector of TestProperties
+  std::vector<TestProperty> test_properties_;
+  // Running count of death tests.
+  int death_test_count_;
+  // The elapsed time, in milliseconds.
+  TimeInMillis elapsed_time_;
+
+  // We disallow copying TestResult.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestResult);
+};  // class TestResult
+
+// A TestInfo object stores the following information about a test:
+//
+//   Test case name
+//   Test name
+//   Whether the test should be run
+//   A function pointer that creates the test object when invoked
+//   Test result
+//
+// The constructor of TestInfo registers itself with the UnitTest
+// singleton such that the RUN_ALL_TESTS() macro knows which tests to
+// run.
+class GTEST_API_ TestInfo {
+ public:
+  // Destructs a TestInfo object.  This function is not virtual, so
+  // don't inherit from TestInfo.
+  ~TestInfo();
+
+  // Returns the test case name.
+  const char* test_case_name() const { return test_case_name_.c_str(); }
+
+  // Returns the test name.
+  const char* name() const { return name_.c_str(); }
+
+  // Returns the name of the parameter type, or NULL if this is not a typed
+  // or a type-parameterized test.
+  const char* type_param() const {
+    if (type_param_.get() != NULL)
+      return type_param_->c_str();
+    return NULL;
+  }
+
+  // Returns the text representation of the value parameter, or NULL if this
+  // is not a value-parameterized test.
+  const char* value_param() const {
+    if (value_param_.get() != NULL)
+      return value_param_->c_str();
+    return NULL;
+  }
+
+  // Returns true if this test should run, that is if the test is not
+  // disabled (or it is disabled but the also_run_disabled_tests flag has
+  // been specified) and its full name matches the user-specified filter.
+  //
+  // Google Test allows the user to filter the tests by their full names.
+  // The full name of a test Bar in test case Foo is defined as
+  // "Foo.Bar".  Only the tests that match the filter will run.
+  //
+  // A filter is a colon-separated list of glob (not regex) patterns,
+  // optionally followed by a '-' and a colon-separated list of
+  // negative patterns (tests to exclude).  A test is run if it
+  // matches one of the positive patterns and does not match any of
+  // the negative patterns.
+  //
+  // For example, *A*:Foo.* is a filter that matches any string that
+  // contains the character 'A' or starts with "Foo.".
+  bool should_run() const { return should_run_; }
+
+  // Returns true iff this test will appear in the XML report.
+  bool is_reportable() const {
+    // For now, the XML report includes all tests matching the filter.
+    // In the future, we may trim tests that are excluded because of
+    // sharding.
+    return matches_filter_;
+  }
+
+  // Returns the result of the test.
+  const TestResult* result() const { return &result_; }
+
+ private:
+#if GTEST_HAS_DEATH_TEST
+  friend class internal::DefaultDeathTestFactory;
+#endif  // GTEST_HAS_DEATH_TEST
+  friend class Test;
+  friend class TestCase;
+  friend class internal::UnitTestImpl;
+  friend class internal::StreamingListenerTest;
+  friend TestInfo* internal::MakeAndRegisterTestInfo(
+      const char* test_case_name,
+      const char* name,
+      const char* type_param,
+      const char* value_param,
+      internal::TypeId fixture_class_id,
+      Test::SetUpTestCaseFunc set_up_tc,
+      Test::TearDownTestCaseFunc tear_down_tc,
+      internal::TestFactoryBase* factory);
+
+  // Constructs a TestInfo object. The newly constructed instance assumes
+  // ownership of the factory object.
+  TestInfo(const std::string& test_case_name,
+           const std::string& name,
+           const char* a_type_param,   // NULL if not a type-parameterized test
+           const char* a_value_param,  // NULL if not a value-parameterized test
+           internal::TypeId fixture_class_id,
+           internal::TestFactoryBase* factory);
+
+  // Increments the number of death tests encountered in this test so
+  // far.
+  int increment_death_test_count() {
+    return result_.increment_death_test_count();
+  }
+
+  // Creates the test object, runs it, records its result, and then
+  // deletes it.
+  void Run();
+
+  static void ClearTestResult(TestInfo* test_info) {
+    test_info->result_.Clear();
+  }
+
+  // These fields are immutable properties of the test.
+  const std::string test_case_name_;     // Test case name
+  const std::string name_;               // Test name
+  // Name of the parameter type, or NULL if this is not a typed or a
+  // type-parameterized test.
+  const internal::scoped_ptr<const ::std::string> type_param_;
+  // Text representation of the value parameter, or NULL if this is not a
+  // value-parameterized test.
+  const internal::scoped_ptr<const ::std::string> value_param_;
+  const internal::TypeId fixture_class_id_;   // ID of the test fixture class
+  bool should_run_;                 // True iff this test should run
+  bool is_disabled_;                // True iff this test is disabled
+  bool matches_filter_;             // True if this test matches the
+                                    // user-specified filter.
+  internal::TestFactoryBase* const factory_;  // The factory that creates
+                                              // the test object
+
+  // This field is mutable and needs to be reset before running the
+  // test for the second time.
+  TestResult result_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestInfo);
+};
+
+// A test case, which consists of a vector of TestInfos.
+//
+// TestCase is not copyable.
+class GTEST_API_ TestCase {
+ public:
+  // Creates a TestCase with the given name.
+  //
+  // TestCase does NOT have a default constructor.  Always use this
+  // constructor to create a TestCase object.
+  //
+  // Arguments:
+  //
+  //   name:         name of the test case
+  //   a_type_param: the name of the test's type parameter, or NULL if
+  //                 this is not a type-parameterized test.
+  //   set_up_tc:    pointer to the function that sets up the test case
+  //   tear_down_tc: pointer to the function that tears down the test case
+  TestCase(const char* name, const char* a_type_param,
+           Test::SetUpTestCaseFunc set_up_tc,
+           Test::TearDownTestCaseFunc tear_down_tc);
+
+  // Destructor of TestCase.
+  virtual ~TestCase();
+
+  // Gets the name of the TestCase.
+  const char* name() const { return name_.c_str(); }
+
+  // Returns the name of the parameter type, or NULL if this is not a
+  // type-parameterized test case.
+  const char* type_param() const {
+    if (type_param_.get() != NULL)
+      return type_param_->c_str();
+    return NULL;
+  }
+
+  // Returns true if any test in this test case should run.
+  bool should_run() const { return should_run_; }
+
+  // Gets the number of successful tests in this test case.
+  int successful_test_count() const;
+
+  // Gets the number of failed tests in this test case.
+  int failed_test_count() const;
+
+  // Gets the number of disabled tests that will be reported in the XML report.
+  int reportable_disabled_test_count() const;
+
+  // Gets the number of disabled tests in this test case.
+  int disabled_test_count() const;
+
+  // Gets the number of tests to be printed in the XML report.
+  int reportable_test_count() const;
+
+  // Get the number of tests in this test case that should run.
+  int test_to_run_count() const;
+
+  // Gets the number of all tests in this test case.
+  int total_test_count() const;
+
+  // Returns true iff the test case passed.
+  bool Passed() const { return !Failed(); }
+
+  // Returns true iff the test case failed.
+  bool Failed() const { return failed_test_count() > 0; }
+
+  // Returns the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const { return elapsed_time_; }
+
+  // Returns the i-th test among all the tests. i can range from 0 to
+  // total_test_count() - 1. If i is not in that range, returns NULL.
+  const TestInfo* GetTestInfo(int i) const;
+
+  // Returns the TestResult that holds test properties recorded during
+  // execution of SetUpTestCase and TearDownTestCase.
+  const TestResult& ad_hoc_test_result() const { return ad_hoc_test_result_; }
+
+ private:
+  friend class Test;
+  friend class internal::UnitTestImpl;
+
+  // Gets the (mutable) vector of TestInfos in this TestCase.
+  std::vector<TestInfo*>& test_info_list() { return test_info_list_; }
+
+  // Gets the (immutable) vector of TestInfos in this TestCase.
+  const std::vector<TestInfo*>& test_info_list() const {
+    return test_info_list_;
+  }
+
+  // Returns the i-th test among all the tests. i can range from 0 to
+  // total_test_count() - 1. If i is not in that range, returns NULL.
+  TestInfo* GetMutableTestInfo(int i);
+
+  // Sets the should_run member.
+  void set_should_run(bool should) { should_run_ = should; }
+
+  // Adds a TestInfo to this test case.  Will delete the TestInfo upon
+  // destruction of the TestCase object.
+  void AddTestInfo(TestInfo * test_info);
+
+  // Clears the results of all tests in this test case.
+  void ClearResult();
+
+  // Clears the results of all tests in the given test case.
+  static void ClearTestCaseResult(TestCase* test_case) {
+    test_case->ClearResult();
+  }
+
+  // Runs every test in this TestCase.
+  void Run();
+
+  // Runs SetUpTestCase() for this TestCase.  This wrapper is needed
+  // for catching exceptions thrown from SetUpTestCase().
+  void RunSetUpTestCase() { (*set_up_tc_)(); }
+
+  // Runs TearDownTestCase() for this TestCase.  This wrapper is
+  // needed for catching exceptions thrown from TearDownTestCase().
+  void RunTearDownTestCase() { (*tear_down_tc_)(); }
+
+  // Returns true iff test passed.
+  static bool TestPassed(const TestInfo* test_info) {
+    return test_info->should_run() && test_info->result()->Passed();
+  }
+
+  // Returns true iff test failed.
+  static bool TestFailed(const TestInfo* test_info) {
+    return test_info->should_run() && test_info->result()->Failed();
+  }
+
+  // Returns true iff the test is disabled and will be reported in the XML
+  // report.
+  static bool TestReportableDisabled(const TestInfo* test_info) {
+    return test_info->is_reportable() && test_info->is_disabled_;
+  }
+
+  // Returns true iff test is disabled.
+  static bool TestDisabled(const TestInfo* test_info) {
+    return test_info->is_disabled_;
+  }
+
+  // Returns true iff this test will appear in the XML report.
+  static bool TestReportable(const TestInfo* test_info) {
+    return test_info->is_reportable();
+  }
+
+  // Returns true if the given test should run.
+  static bool ShouldRunTest(const TestInfo* test_info) {
+    return test_info->should_run();
+  }
+
+  // Shuffles the tests in this test case.
+  void ShuffleTests(internal::Random* random);
+
+  // Restores the test order to before the first shuffle.
+  void UnshuffleTests();
+
+  // Name of the test case.
+  std::string name_;
+  // Name of the parameter type, or NULL if this is not a typed or a
+  // type-parameterized test.
+  const internal::scoped_ptr<const ::std::string> type_param_;
+  // The vector of TestInfos in their original order.  It owns the
+  // elements in the vector.
+  std::vector<TestInfo*> test_info_list_;
+  // Provides a level of indirection for the test list to allow easy
+  // shuffling and restoring the test order.  The i-th element in this
+  // vector is the index of the i-th test in the shuffled test list.
+  std::vector<int> test_indices_;
+  // Pointer to the function that sets up the test case.
+  Test::SetUpTestCaseFunc set_up_tc_;
+  // Pointer to the function that tears down the test case.
+  Test::TearDownTestCaseFunc tear_down_tc_;
+  // True iff any test in this test case should run.
+  bool should_run_;
+  // Elapsed time, in milliseconds.
+  TimeInMillis elapsed_time_;
+  // Holds test properties recorded during execution of SetUpTestCase and
+  // TearDownTestCase.
+  TestResult ad_hoc_test_result_;
+
+  // We disallow copying TestCases.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestCase);
+};
+
+// An Environment object is capable of setting up and tearing down an
+// environment.  The user should subclass this to define his own
+// environment(s).
+//
+// An Environment object does the set-up and tear-down in virtual
+// methods SetUp() and TearDown() instead of the constructor and the
+// destructor, as:
+//
+//   1. You cannot safely throw from a destructor.  This is a problem
+//      as in some cases Google Test is used where exceptions are enabled, and
+//      we may want to implement ASSERT_* using exceptions where they are
+//      available.
+//   2. You cannot use ASSERT_* directly in a constructor or
+//      destructor.
+class Environment {
+ public:
+  // The d'tor is virtual as we need to subclass Environment.
+  virtual ~Environment() {}
+
+  // Override this to define how to set up the environment.
+  virtual void SetUp() {}
+
+  // Override this to define how to tear down the environment.
+  virtual void TearDown() {}
+ private:
+  // If you see an error about overriding the following function or
+  // about it being private, you have mis-spelled SetUp() as Setup().
+  struct Setup_should_be_spelled_SetUp {};
+  virtual Setup_should_be_spelled_SetUp* Setup() { return NULL; }
+};
+
+// The interface for tracing execution of tests. The methods are organized in
+// the order the corresponding events are fired.
+class TestEventListener {
+ public:
+  virtual ~TestEventListener() {}
+
+  // Fired before any test activity starts.
+  virtual void OnTestProgramStart(const UnitTest& unit_test) = 0;
+
+  // Fired before each iteration of tests starts.  There may be more than
+  // one iteration if GTEST_FLAG(repeat) is set. iteration is the iteration
+  // index, starting from 0.
+  virtual void OnTestIterationStart(const UnitTest& unit_test,
+                                    int iteration) = 0;
+
+  // Fired before environment set-up for each iteration of tests starts.
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& unit_test) = 0;
+
+  // Fired after environment set-up for each iteration of tests ends.
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& unit_test) = 0;
+
+  // Fired before the test case starts.
+  virtual void OnTestCaseStart(const TestCase& test_case) = 0;
+
+  // Fired before the test starts.
+  virtual void OnTestStart(const TestInfo& test_info) = 0;
+
+  // Fired after a failed assertion or a SUCCEED() invocation.
+  virtual void OnTestPartResult(const TestPartResult& test_part_result) = 0;
+
+  // Fired after the test ends.
+  virtual void OnTestEnd(const TestInfo& test_info) = 0;
+
+  // Fired after the test case ends.
+  virtual void OnTestCaseEnd(const TestCase& test_case) = 0;
+
+  // Fired before environment tear-down for each iteration of tests starts.
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& unit_test) = 0;
+
+  // Fired after environment tear-down for each iteration of tests ends.
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& unit_test) = 0;
+
+  // Fired after each iteration of tests finishes.
+  virtual void OnTestIterationEnd(const UnitTest& unit_test,
+                                  int iteration) = 0;
+
+  // Fired after all test activities have ended.
+  virtual void OnTestProgramEnd(const UnitTest& unit_test) = 0;
+};
+
+// The convenience class for users who need to override just one or two
+// methods and are not concerned that a possible change to a signature of
+// the methods they override will not be caught during the build.  For
+// comments about each method please see the definition of TestEventListener
+// above.
+class EmptyTestEventListener : public TestEventListener {
+ public:
+  virtual void OnTestProgramStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationStart(const UnitTest& /*unit_test*/,
+                                    int /*iteration*/) {}
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestCaseStart(const TestCase& /*test_case*/) {}
+  virtual void OnTestStart(const TestInfo& /*test_info*/) {}
+  virtual void OnTestPartResult(const TestPartResult& /*test_part_result*/) {}
+  virtual void OnTestEnd(const TestInfo& /*test_info*/) {}
+  virtual void OnTestCaseEnd(const TestCase& /*test_case*/) {}
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationEnd(const UnitTest& /*unit_test*/,
+                                  int /*iteration*/) {}
+  virtual void OnTestProgramEnd(const UnitTest& /*unit_test*/) {}
+};
+
+// TestEventListeners lets users add listeners to track events in Google Test.
+class GTEST_API_ TestEventListeners {
+ public:
+  TestEventListeners();
+  ~TestEventListeners();
+
+  // Appends an event listener to the end of the list. Google Test assumes
+  // the ownership of the listener (i.e. it will delete the listener when
+  // the test program finishes).
+  void Append(TestEventListener* listener);
+
+  // Removes the given event listener from the list and returns it.  It then
+  // becomes the caller's responsibility to delete the listener. Returns
+  // NULL if the listener is not found in the list.
+  TestEventListener* Release(TestEventListener* listener);
+
+  // Returns the standard listener responsible for the default console
+  // output.  Can be removed from the listeners list to shut down default
+  // console output.  Note that removing this object from the listener list
+  // with Release transfers its ownership to the caller and makes this
+  // function return NULL the next time.
+  TestEventListener* default_result_printer() const {
+    return default_result_printer_;
+  }
+
+  // Returns the standard listener responsible for the default XML output
+  // controlled by the --gtest_output=xml flag.  Can be removed from the
+  // listeners list by users who want to shut down the default XML output
+  // controlled by this flag and substitute it with custom one.  Note that
+  // removing this object from the listener list with Release transfers its
+  // ownership to the caller and makes this function return NULL the next
+  // time.
+  TestEventListener* default_xml_generator() const {
+    return default_xml_generator_;
+  }
+
+ private:
+  friend class TestCase;
+  friend class TestInfo;
+  friend class internal::DefaultGlobalTestPartResultReporter;
+  friend class internal::NoExecDeathTest;
+  friend class internal::TestEventListenersAccessor;
+  friend class internal::UnitTestImpl;
+
+  // Returns repeater that broadcasts the TestEventListener events to all
+  // subscribers.
+  TestEventListener* repeater();
+
+  // Sets the default_result_printer attribute to the provided listener.
+  // The listener is also added to the listener list and previous
+  // default_result_printer is removed from it and deleted. The listener can
+  // also be NULL in which case it will not be added to the list. Does
+  // nothing if the previous and the current listener objects are the same.
+  void SetDefaultResultPrinter(TestEventListener* listener);
+
+  // Sets the default_xml_generator attribute to the provided listener.  The
+  // listener is also added to the listener list and previous
+  // default_xml_generator is removed from it and deleted. The listener can
+  // also be NULL in which case it will not be added to the list. Does
+  // nothing if the previous and the current listener objects are the same.
+  void SetDefaultXmlGenerator(TestEventListener* listener);
+
+  // Controls whether events will be forwarded by the repeater to the
+  // listeners in the list.
+  bool EventForwardingEnabled() const;
+  void SuppressEventForwarding();
+
+  // The actual list of listeners.
+  internal::TestEventRepeater* repeater_;
+  // Listener responsible for the standard result output.
+  TestEventListener* default_result_printer_;
+  // Listener responsible for the creation of the XML output file.
+  TestEventListener* default_xml_generator_;
+
+  // We disallow copying TestEventListeners.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestEventListeners);
+};
+
+// A UnitTest consists of a vector of TestCases.
+//
+// This is a singleton class.  The only instance of UnitTest is
+// created when UnitTest::GetInstance() is first called.  This
+// instance is never deleted.
+//
+// UnitTest is not copyable.
+//
+// This class is thread-safe as long as the methods are called
+// according to their specification.
+class GTEST_API_ UnitTest {
+ public:
+  // Gets the singleton UnitTest object.  The first time this method
+  // is called, a UnitTest object is constructed and returned.
+  // Consecutive calls will return the same object.
+  static UnitTest* GetInstance();
+
+  // Runs all tests in this UnitTest object and prints the result.
+  // Returns 0 if successful, or 1 otherwise.
+  //
+  // This method can only be called from the main thread.
+  //
+  // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+  int Run() GTEST_MUST_USE_RESULT_;
+
+  // Returns the working directory when the first TEST() or TEST_F()
+  // was executed.  The UnitTest object owns the string.
+  const char* original_working_dir() const;
+
+  // Returns the TestCase object for the test that's currently running,
+  // or NULL if no test is running.
+  const TestCase* current_test_case() const
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Returns the TestInfo object for the test that's currently running,
+  // or NULL if no test is running.
+  const TestInfo* current_test_info() const
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Returns the random seed used at the start of the current test run.
+  int random_seed() const;
+
+#if GTEST_HAS_PARAM_TEST
+  // Returns the ParameterizedTestCaseRegistry object used to keep track of
+  // value-parameterized tests and instantiate and register them.
+  //
+  // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+  internal::ParameterizedTestCaseRegistry& parameterized_test_registry()
+      GTEST_LOCK_EXCLUDED_(mutex_);
+#endif  // GTEST_HAS_PARAM_TEST
+
+  // Gets the number of successful test cases.
+  int successful_test_case_count() const;
+
+  // Gets the number of failed test cases.
+  int failed_test_case_count() const;
+
+  // Gets the number of all test cases.
+  int total_test_case_count() const;
+
+  // Gets the number of all test cases that contain at least one test
+  // that should run.
+  int test_case_to_run_count() const;
+
+  // Gets the number of successful tests.
+  int successful_test_count() const;
+
+  // Gets the number of failed tests.
+  int failed_test_count() const;
+
+  // Gets the number of disabled tests that will be reported in the XML report.
+  int reportable_disabled_test_count() const;
+
+  // Gets the number of disabled tests.
+  int disabled_test_count() const;
+
+  // Gets the number of tests to be printed in the XML report.
+  int reportable_test_count() const;
+
+  // Gets the number of all tests.
+  int total_test_count() const;
+
+  // Gets the number of tests that should run.
+  int test_to_run_count() const;
+
+  // Gets the time of the test program start, in ms from the start of the
+  // UNIX epoch.
+  TimeInMillis start_timestamp() const;
+
+  // Gets the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const;
+
+  // Returns true iff the unit test passed (i.e. all test cases passed).
+  bool Passed() const;
+
+  // Returns true iff the unit test failed (i.e. some test case failed
+  // or something outside of all tests failed).
+  bool Failed() const;
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  const TestCase* GetTestCase(int i) const;
+
+  // Returns the TestResult containing information on test failures and
+  // properties logged outside of individual test cases.
+  const TestResult& ad_hoc_test_result() const;
+
+  // Returns the list of event listeners that can be used to track events
+  // inside Google Test.
+  TestEventListeners& listeners();
+
+ private:
+  // Registers and returns a global test environment.  When a test
+  // program is run, all global test environments will be set-up in
+  // the order they were registered.  After all tests in the program
+  // have finished, all global test environments will be torn-down in
+  // the *reverse* order they were registered.
+  //
+  // The UnitTest object takes ownership of the given environment.
+  //
+  // This method can only be called from the main thread.
+  Environment* AddEnvironment(Environment* env);
+
+  // Adds a TestPartResult to the current TestResult object.  All
+  // Google Test assertion macros (e.g. ASSERT_TRUE, EXPECT_EQ, etc)
+  // eventually call this to report their results.  The user code
+  // should use the assertion macros instead of calling this directly.
+  void AddTestPartResult(TestPartResult::Type result_type,
+                         const char* file_name,
+                         int line_number,
+                         const std::string& message,
+                         const std::string& os_stack_trace)
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Adds a TestProperty to the current TestResult object when invoked from
+  // inside a test, to current TestCase's ad_hoc_test_result_ when invoked
+  // from SetUpTestCase or TearDownTestCase, or to the global property set
+  // when invoked elsewhere.  If the result already contains a property with
+  // the same key, the value will be updated.
+  void RecordProperty(const std::string& key, const std::string& value);
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  TestCase* GetMutableTestCase(int i);
+
+  // Accessors for the implementation object.
+  internal::UnitTestImpl* impl() { return impl_; }
+  const internal::UnitTestImpl* impl() const { return impl_; }
+
+  // These classes and funcions are friends as they need to access private
+  // members of UnitTest.
+  friend class Test;
+  friend class internal::AssertHelper;
+  friend class internal::ScopedTrace;
+  friend class internal::StreamingListenerTest;
+  friend class internal::UnitTestRecordPropertyTestHelper;
+  friend Environment* AddGlobalTestEnvironment(Environment* env);
+  friend internal::UnitTestImpl* internal::GetUnitTestImpl();
+  friend void internal::ReportFailureInUnknownLocation(
+      TestPartResult::Type result_type,
+      const std::string& message);
+
+  // Creates an empty UnitTest.
+  UnitTest();
+
+  // D'tor
+  virtual ~UnitTest();
+
+  // Pushes a trace defined by SCOPED_TRACE() on to the per-thread
+  // Google Test trace stack.
+  void PushGTestTrace(const internal::TraceInfo& trace)
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Pops a trace from the per-thread Google Test trace stack.
+  void PopGTestTrace()
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Protects mutable state in *impl_.  This is mutable as some const
+  // methods need to lock it too.
+  mutable internal::Mutex mutex_;
+
+  // Opaque implementation object.  This field is never changed once
+  // the object is constructed.  We don't mark it as const here, as
+  // doing so will cause a warning in the constructor of UnitTest.
+  // Mutable state in *impl_ is protected by mutex_.
+  internal::UnitTestImpl* impl_;
+
+  // We disallow copying UnitTest.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(UnitTest);
+};
+
+// A convenient wrapper for adding an environment for the test
+// program.
+//
+// You should call this before RUN_ALL_TESTS() is called, probably in
+// main().  If you use gtest_main, you need to call this before main()
+// starts for it to take effect.  For example, you can define a global
+// variable like this:
+//
+//   testing::Environment* const foo_env =
+//       testing::AddGlobalTestEnvironment(new FooEnvironment);
+//
+// However, we strongly recommend you to write your own main() and
+// call AddGlobalTestEnvironment() there, as relying on initialization
+// of global variables makes the code harder to read and may cause
+// problems when you register multiple environments from different
+// translation units and the environments have dependencies among them
+// (remember that the compiler doesn't guarantee the order in which
+// global variables from different translation units are initialized).
+inline Environment* AddGlobalTestEnvironment(Environment* env) {
+  return UnitTest::GetInstance()->AddEnvironment(env);
+}
+
+// Initializes Google Test.  This must be called before calling
+// RUN_ALL_TESTS().  In particular, it parses a command line for the
+// flags that Google Test recognizes.  Whenever a Google Test flag is
+// seen, it is removed from argv, and *argc is decremented.
+//
+// No value is returned.  Instead, the Google Test flag variables are
+// updated.
+//
+// Calling the function for the second time has no user-visible effect.
+GTEST_API_ void InitGoogleTest(int* argc, char** argv);
+
+// This overloaded version can be used in Windows programs compiled in
+// UNICODE mode.
+GTEST_API_ void InitGoogleTest(int* argc, wchar_t** argv);
+
+namespace internal {
+
+// FormatForComparison<ToPrint, OtherOperand>::Format(value) formats a
+// value of type ToPrint that is an operand of a comparison assertion
+// (e.g. ASSERT_EQ).  OtherOperand is the type of the other operand in
+// the comparison, and is used to help determine the best way to
+// format the value.  In particular, when the value is a C string
+// (char pointer) and the other operand is an STL string object, we
+// want to format the C string as a string, since we know it is
+// compared by value with the string object.  If the value is a char
+// pointer but the other operand is not an STL string object, we don't
+// know whether the pointer is supposed to point to a NUL-terminated
+// string, and thus want to print it as a pointer to be safe.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+
+// The default case.
+template <typename ToPrint, typename OtherOperand>
+class FormatForComparison {
+ public:
+  static ::std::string Format(const ToPrint& value) {
+    return ::testing::PrintToString(value);
+  }
+};
+
+// Array.
+template <typename ToPrint, size_t N, typename OtherOperand>
+class FormatForComparison<ToPrint[N], OtherOperand> {
+ public:
+  static ::std::string Format(const ToPrint* value) {
+    return FormatForComparison<const ToPrint*, OtherOperand>::Format(value);
+  }
+};
+
+// By default, print C string as pointers to be safe, as we don't know
+// whether they actually point to a NUL-terminated string.
+
+#define GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(CharType)                \
+  template <typename OtherOperand>                                      \
+  class FormatForComparison<CharType*, OtherOperand> {                  \
+   public:                                                              \
+    static ::std::string Format(CharType* value) {                      \
+      return ::testing::PrintToString(static_cast<const void*>(value)); \
+    }                                                                   \
+  }
+
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(char);
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(const char);
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(wchar_t);
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(const wchar_t);
+
+#undef GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_
+
+// If a C string is compared with an STL string object, we know it's meant
+// to point to a NUL-terminated string, and thus can print it as a string.
+
+#define GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(CharType, OtherStringType) \
+  template <>                                                           \
+  class FormatForComparison<CharType*, OtherStringType> {               \
+   public:                                                              \
+    static ::std::string Format(CharType* value) {                      \
+      return ::testing::PrintToString(value);                           \
+    }                                                                   \
+  }
+
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(char, ::std::string);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const char, ::std::string);
+
+#if GTEST_HAS_GLOBAL_STRING
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(char, ::string);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const char, ::string);
+#endif
+
+#if GTEST_HAS_GLOBAL_WSTRING
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(wchar_t, ::wstring);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const wchar_t, ::wstring);
+#endif
+
+#if GTEST_HAS_STD_WSTRING
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(wchar_t, ::std::wstring);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const wchar_t, ::std::wstring);
+#endif
+
+#undef GTEST_IMPL_FORMAT_C_STRING_AS_STRING_
+
+// Formats a comparison assertion (e.g. ASSERT_EQ, EXPECT_LT, and etc)
+// operand to be used in a failure message.  The type (but not value)
+// of the other operand may affect the format.  This allows us to
+// print a char* as a raw pointer when it is compared against another
+// char* or void*, and print it as a C string when it is compared
+// against an std::string object, for example.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+template <typename T1, typename T2>
+std::string FormatForComparisonFailureMessage(
+    const T1& value, const T2& /* other_operand */) {
+  return FormatForComparison<T1, T2>::Format(value);
+}
+
+// The helper function for {ASSERT|EXPECT}_EQ.
+template <typename T1, typename T2>
+AssertionResult CmpHelperEQ(const char* expected_expression,
+                            const char* actual_expression,
+                            const T1& expected,
+                            const T2& actual) {
+#ifdef _MSC_VER
+# pragma warning(push)          // Saves the current warning state.
+# pragma warning(disable:4389)  // Temporarily disables warning on
+                                // signed/unsigned mismatch.
+#endif
+
+  if (expected == actual) {
+    return AssertionSuccess();
+  }
+
+#ifdef _MSC_VER
+# pragma warning(pop)          // Restores the warning state.
+#endif
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   FormatForComparisonFailureMessage(expected, actual),
+                   FormatForComparisonFailureMessage(actual, expected),
+                   false);
+}
+
+// With this overloaded version, we allow anonymous enums to be used
+// in {ASSERT|EXPECT}_EQ when compiled with gcc 4, as anonymous enums
+// can be implicitly cast to BiggestInt.
+GTEST_API_ AssertionResult CmpHelperEQ(const char* expected_expression,
+                                       const char* actual_expression,
+                                       BiggestInt expected,
+                                       BiggestInt actual);
+
+// The helper class for {ASSERT|EXPECT}_EQ.  The template argument
+// lhs_is_null_literal is true iff the first argument to ASSERT_EQ()
+// is a null pointer literal.  The following default implementation is
+// for lhs_is_null_literal being false.
+template <bool lhs_is_null_literal>
+class EqHelper {
+ public:
+  // This templatized version is for the general case.
+  template <typename T1, typename T2>
+  static AssertionResult Compare(const char* expected_expression,
+                                 const char* actual_expression,
+                                 const T1& expected,
+                                 const T2& actual) {
+    return CmpHelperEQ(expected_expression, actual_expression, expected,
+                       actual);
+  }
+
+  // With this overloaded version, we allow anonymous enums to be used
+  // in {ASSERT|EXPECT}_EQ when compiled with gcc 4, as anonymous
+  // enums can be implicitly cast to BiggestInt.
+  //
+  // Even though its body looks the same as the above version, we
+  // cannot merge the two, as it will make anonymous enums unhappy.
+  static AssertionResult Compare(const char* expected_expression,
+                                 const char* actual_expression,
+                                 BiggestInt expected,
+                                 BiggestInt actual) {
+    return CmpHelperEQ(expected_expression, actual_expression, expected,
+                       actual);
+  }
+};
+
+// This specialization is used when the first argument to ASSERT_EQ()
+// is a null pointer literal, like NULL, false, or 0.
+template <>
+class EqHelper<true> {
+ public:
+  // We define two overloaded versions of Compare().  The first
+  // version will be picked when the second argument to ASSERT_EQ() is
+  // NOT a pointer, e.g. ASSERT_EQ(0, AnIntFunction()) or
+  // EXPECT_EQ(false, a_bool).
+  template <typename T1, typename T2>
+  static AssertionResult Compare(
+      const char* expected_expression,
+      const char* actual_expression,
+      const T1& expected,
+      const T2& actual,
+      // The following line prevents this overload from being considered if T2
+      // is not a pointer type.  We need this because ASSERT_EQ(NULL, my_ptr)
+      // expands to Compare("", "", NULL, my_ptr), which requires a conversion
+      // to match the Secret* in the other overload, which would otherwise make
+      // this template match better.
+      typename EnableIf<!is_pointer<T2>::value>::type* = 0) {
+    return CmpHelperEQ(expected_expression, actual_expression, expected,
+                       actual);
+  }
+
+  // This version will be picked when the second argument to ASSERT_EQ() is a
+  // pointer, e.g. ASSERT_EQ(NULL, a_pointer).
+  template <typename T>
+  static AssertionResult Compare(
+      const char* expected_expression,
+      const char* actual_expression,
+      // We used to have a second template parameter instead of Secret*.  That
+      // template parameter would deduce to 'long', making this a better match
+      // than the first overload even without the first overload's EnableIf.
+      // Unfortunately, gcc with -Wconversion-null warns when "passing NULL to
+      // non-pointer argument" (even a deduced integral argument), so the old
+      // implementation caused warnings in user code.
+      Secret* /* expected (NULL) */,
+      T* actual) {
+    // We already know that 'expected' is a null pointer.
+    return CmpHelperEQ(expected_expression, actual_expression,
+                       static_cast<T*>(NULL), actual);
+  }
+};
+
+// A macro for implementing the helper functions needed to implement
+// ASSERT_?? and EXPECT_??.  It is here just to avoid copy-and-paste
+// of similar code.
+//
+// For each templatized helper function, we also define an overloaded
+// version for BiggestInt in order to reduce code bloat and allow
+// anonymous enums to be used with {ASSERT|EXPECT}_?? when compiled
+// with gcc 4.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+#define GTEST_IMPL_CMP_HELPER_(op_name, op)\
+template <typename T1, typename T2>\
+AssertionResult CmpHelper##op_name(const char* expr1, const char* expr2, \
+                                   const T1& val1, const T2& val2) {\
+  if (val1 op val2) {\
+    return AssertionSuccess();\
+  } else {\
+    return AssertionFailure() \
+        << "Expected: (" << expr1 << ") " #op " (" << expr2\
+        << "), actual: " << FormatForComparisonFailureMessage(val1, val2)\
+        << " vs " << FormatForComparisonFailureMessage(val2, val1);\
+  }\
+}\
+GTEST_API_ AssertionResult CmpHelper##op_name(\
+    const char* expr1, const char* expr2, BiggestInt val1, BiggestInt val2)
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+
+// Implements the helper function for {ASSERT|EXPECT}_NE
+GTEST_IMPL_CMP_HELPER_(NE, !=);
+// Implements the helper function for {ASSERT|EXPECT}_LE
+GTEST_IMPL_CMP_HELPER_(LE, <=);
+// Implements the helper function for {ASSERT|EXPECT}_LT
+GTEST_IMPL_CMP_HELPER_(LT, <);
+// Implements the helper function for {ASSERT|EXPECT}_GE
+GTEST_IMPL_CMP_HELPER_(GE, >=);
+// Implements the helper function for {ASSERT|EXPECT}_GT
+GTEST_IMPL_CMP_HELPER_(GT, >);
+
+#undef GTEST_IMPL_CMP_HELPER_
+
+// The helper function for {ASSERT|EXPECT}_STREQ.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTREQ(const char* expected_expression,
+                                          const char* actual_expression,
+                                          const char* expected,
+                                          const char* actual);
+
+// The helper function for {ASSERT|EXPECT}_STRCASEEQ.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRCASEEQ(const char* expected_expression,
+                                              const char* actual_expression,
+                                              const char* expected,
+                                              const char* actual);
+
+// The helper function for {ASSERT|EXPECT}_STRNE.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                                          const char* s2_expression,
+                                          const char* s1,
+                                          const char* s2);
+
+// The helper function for {ASSERT|EXPECT}_STRCASENE.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRCASENE(const char* s1_expression,
+                                              const char* s2_expression,
+                                              const char* s1,
+                                              const char* s2);
+
+
+// Helper function for *_STREQ on wide strings.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTREQ(const char* expected_expression,
+                                          const char* actual_expression,
+                                          const wchar_t* expected,
+                                          const wchar_t* actual);
+
+// Helper function for *_STRNE on wide strings.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                                          const char* s2_expression,
+                                          const wchar_t* s1,
+                                          const wchar_t* s2);
+
+}  // namespace internal
+
+// IsSubstring() and IsNotSubstring() are intended to be used as the
+// first argument to {EXPECT,ASSERT}_PRED_FORMAT2(), not by
+// themselves.  They check whether needle is a substring of haystack
+// (NULL is considered a substring of itself only), and return an
+// appropriate error message when they fail.
+//
+// The {needle,haystack}_expr arguments are the stringified
+// expressions that generated the two real arguments.
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack);
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack);
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack);
+
+#if GTEST_HAS_STD_WSTRING
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack);
+#endif  // GTEST_HAS_STD_WSTRING
+
+namespace internal {
+
+// Helper template function for comparing floating-points.
+//
+// Template parameter:
+//
+//   RawType: the raw floating-point type (either float or double)
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+template <typename RawType>
+AssertionResult CmpHelperFloatingPointEQ(const char* expected_expression,
+                                         const char* actual_expression,
+                                         RawType expected,
+                                         RawType actual) {
+  const FloatingPoint<RawType> lhs(expected), rhs(actual);
+
+  if (lhs.AlmostEquals(rhs)) {
+    return AssertionSuccess();
+  }
+
+  ::std::stringstream expected_ss;
+  expected_ss << std::setprecision(std::numeric_limits<RawType>::digits10 + 2)
+              << expected;
+
+  ::std::stringstream actual_ss;
+  actual_ss << std::setprecision(std::numeric_limits<RawType>::digits10 + 2)
+            << actual;
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   StringStreamToString(&expected_ss),
+                   StringStreamToString(&actual_ss),
+                   false);
+}
+
+// Helper function for implementing ASSERT_NEAR.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult DoubleNearPredFormat(const char* expr1,
+                                                const char* expr2,
+                                                const char* abs_error_expr,
+                                                double val1,
+                                                double val2,
+                                                double abs_error);
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+// A class that enables one to stream messages to assertion macros
+class GTEST_API_ AssertHelper {
+ public:
+  // Constructor.
+  AssertHelper(TestPartResult::Type type,
+               const char* file,
+               int line,
+               const char* message);
+  ~AssertHelper();
+
+  // Message assignment is a semantic trick to enable assertion
+  // streaming; see the GTEST_MESSAGE_ macro below.
+  void operator=(const Message& message) const;
+
+ private:
+  // We put our data in a struct so that the size of the AssertHelper class can
+  // be as small as possible.  This is important because gcc is incapable of
+  // re-using stack space even for temporary variables, so every EXPECT_EQ
+  // reserves stack space for another AssertHelper.
+  struct AssertHelperData {
+    AssertHelperData(TestPartResult::Type t,
+                     const char* srcfile,
+                     int line_num,
+                     const char* msg)
+        : type(t), file(srcfile), line(line_num), message(msg) { }
+
+    TestPartResult::Type const type;
+    const char* const file;
+    int const line;
+    std::string const message;
+
+   private:
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(AssertHelperData);
+  };
+
+  AssertHelperData* const data_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(AssertHelper);
+};
+
+}  // namespace internal
+
+#if GTEST_HAS_PARAM_TEST
+// The pure interface class that all value-parameterized tests inherit from.
+// A value-parameterized class must inherit from both ::testing::Test and
+// ::testing::WithParamInterface. In most cases that just means inheriting
+// from ::testing::TestWithParam, but more complicated test hierarchies
+// may need to inherit from Test and WithParamInterface at different levels.
+//
+// This interface has support for accessing the test parameter value via
+// the GetParam() method.
+//
+// Use it with one of the parameter generator defining functions, like Range(),
+// Values(), ValuesIn(), Bool(), and Combine().
+//
+// class FooTest : public ::testing::TestWithParam<int> {
+//  protected:
+//   FooTest() {
+//     // Can use GetParam() here.
+//   }
+//   virtual ~FooTest() {
+//     // Can use GetParam() here.
+//   }
+//   virtual void SetUp() {
+//     // Can use GetParam() here.
+//   }
+//   virtual void TearDown {
+//     // Can use GetParam() here.
+//   }
+// };
+// TEST_P(FooTest, DoesBar) {
+//   // Can use GetParam() method here.
+//   Foo foo;
+//   ASSERT_TRUE(foo.DoesBar(GetParam()));
+// }
+// INSTANTIATE_TEST_CASE_P(OneToTenRange, FooTest, ::testing::Range(1, 10));
+
+template <typename T>
+class WithParamInterface {
+ public:
+  typedef T ParamType;
+  virtual ~WithParamInterface() {}
+
+  // The current parameter value. Is also available in the test fixture's
+  // constructor. This member function is non-static, even though it only
+  // references static data, to reduce the opportunity for incorrect uses
+  // like writing 'WithParamInterface<bool>::GetParam()' for a test that
+  // uses a fixture whose parameter type is int.
+  const ParamType& GetParam() const {
+    GTEST_CHECK_(parameter_ != NULL)
+        << "GetParam() can only be called inside a value-parameterized test "
+        << "-- did you intend to write TEST_P instead of TEST_F?";
+    return *parameter_;
+  }
+
+ private:
+  // Sets parameter value. The caller is responsible for making sure the value
+  // remains alive and unchanged throughout the current test.
+  static void SetParam(const ParamType* parameter) {
+    parameter_ = parameter;
+  }
+
+  // Static value used for accessing parameter during a test lifetime.
+  static const ParamType* parameter_;
+
+  // TestClass must be a subclass of WithParamInterface<T> and Test.
+  template <class TestClass> friend class internal::ParameterizedTestFactory;
+};
+
+template <typename T>
+const T* WithParamInterface<T>::parameter_ = NULL;
+
+// Most value-parameterized classes can ignore the existence of
+// WithParamInterface, and can just inherit from ::testing::TestWithParam.
+
+template <typename T>
+class TestWithParam : public Test, public WithParamInterface<T> {
+};
+
+#endif  // GTEST_HAS_PARAM_TEST
+
+// Macros for indicating success/failure in test code.
+
+// ADD_FAILURE unconditionally adds a failure to the current test.
+// SUCCEED generates a success - it doesn't automatically make the
+// current test successful, as a test is only successful when it has
+// no failure.
+//
+// EXPECT_* verifies that a certain condition is satisfied.  If not,
+// it behaves like ADD_FAILURE.  In particular:
+//
+//   EXPECT_TRUE  verifies that a Boolean condition is true.
+//   EXPECT_FALSE verifies that a Boolean condition is false.
+//
+// FAIL and ASSERT_* are similar to ADD_FAILURE and EXPECT_*, except
+// that they will also abort the current function on failure.  People
+// usually want the fail-fast behavior of FAIL and ASSERT_*, but those
+// writing data-driven tests often find themselves using ADD_FAILURE
+// and EXPECT_* more.
+
+// Generates a nonfatal failure with a generic message.
+#define ADD_FAILURE() GTEST_NONFATAL_FAILURE_("Failed")
+
+// Generates a nonfatal failure at the given source file location with
+// a generic message.
+#define ADD_FAILURE_AT(file, line) \
+  GTEST_MESSAGE_AT_(file, line, "Failed", \
+                    ::testing::TestPartResult::kNonFatalFailure)
+
+// Generates a fatal failure with a generic message.
+#define GTEST_FAIL() GTEST_FATAL_FAILURE_("Failed")
+
+// Define this macro to 1 to omit the definition of FAIL(), which is a
+// generic name and clashes with some other libraries.
+#if !GTEST_DONT_DEFINE_FAIL
+# define FAIL() GTEST_FAIL()
+#endif
+
+// Generates a success with a generic message.
+#define GTEST_SUCCEED() GTEST_SUCCESS_("Succeeded")
+
+// Define this macro to 1 to omit the definition of SUCCEED(), which
+// is a generic name and clashes with some other libraries.
+#if !GTEST_DONT_DEFINE_SUCCEED
+# define SUCCEED() GTEST_SUCCEED()
+#endif
+
+// Macros for testing exceptions.
+//
+//    * {ASSERT|EXPECT}_THROW(statement, expected_exception):
+//         Tests that the statement throws the expected exception.
+//    * {ASSERT|EXPECT}_NO_THROW(statement):
+//         Tests that the statement doesn't throw any exception.
+//    * {ASSERT|EXPECT}_ANY_THROW(statement):
+//         Tests that the statement throws an exception.
+
+#define EXPECT_THROW(statement, expected_exception) \
+  GTEST_TEST_THROW_(statement, expected_exception, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_NO_THROW(statement) \
+  GTEST_TEST_NO_THROW_(statement, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_ANY_THROW(statement) \
+  GTEST_TEST_ANY_THROW_(statement, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_THROW(statement, expected_exception) \
+  GTEST_TEST_THROW_(statement, expected_exception, GTEST_FATAL_FAILURE_)
+#define ASSERT_NO_THROW(statement) \
+  GTEST_TEST_NO_THROW_(statement, GTEST_FATAL_FAILURE_)
+#define ASSERT_ANY_THROW(statement) \
+  GTEST_TEST_ANY_THROW_(statement, GTEST_FATAL_FAILURE_)
+
+// Boolean assertions. Condition can be either a Boolean expression or an
+// AssertionResult. For more information on how to use AssertionResult with
+// these macros see comments on that class.
+#define EXPECT_TRUE(condition) \
+  GTEST_TEST_BOOLEAN_(condition, #condition, false, true, \
+                      GTEST_NONFATAL_FAILURE_)
+#define EXPECT_FALSE(condition) \
+  GTEST_TEST_BOOLEAN_(!(condition), #condition, true, false, \
+                      GTEST_NONFATAL_FAILURE_)
+#define ASSERT_TRUE(condition) \
+  GTEST_TEST_BOOLEAN_(condition, #condition, false, true, \
+                      GTEST_FATAL_FAILURE_)
+#define ASSERT_FALSE(condition) \
+  GTEST_TEST_BOOLEAN_(!(condition), #condition, true, false, \
+                      GTEST_FATAL_FAILURE_)
+
+// Includes the auto-generated header that implements a family of
+// generic predicate assertion macros.
+// Copyright 2006, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// This file is AUTOMATICALLY GENERATED on 10/31/2011 by command
+// 'gen_gtest_pred_impl.py 5'.  DO NOT EDIT BY HAND!
+//
+// Implements a family of generic predicate assertion macros.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PRED_IMPL_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PRED_IMPL_H_
+
+// Makes sure this header is not included before gtest.h.
+#ifndef GTEST_INCLUDE_GTEST_GTEST_H_
+# error Do not include gtest_pred_impl.h directly.  Include gtest.h instead.
+#endif  // GTEST_INCLUDE_GTEST_GTEST_H_
+
+// This header implements a family of generic predicate assertion
+// macros:
+//
+//   ASSERT_PRED_FORMAT1(pred_format, v1)
+//   ASSERT_PRED_FORMAT2(pred_format, v1, v2)
+//   ...
+//
+// where pred_format is a function or functor that takes n (in the
+// case of ASSERT_PRED_FORMATn) values and their source expression
+// text, and returns a testing::AssertionResult.  See the definition
+// of ASSERT_EQ in gtest.h for an example.
+//
+// If you don't care about formatting, you can use the more
+// restrictive version:
+//
+//   ASSERT_PRED1(pred, v1)
+//   ASSERT_PRED2(pred, v1, v2)
+//   ...
+//
+// where pred is an n-ary function or functor that returns bool,
+// and the values v1, v2, ..., must support the << operator for
+// streaming to std::ostream.
+//
+// We also define the EXPECT_* variations.
+//
+// For now we only support predicates whose arity is at most 5.
+// Please email googletestframework@googlegroups.com if you need
+// support for higher arities.
+
+// GTEST_ASSERT_ is the basic statement to which all of the assertions
+// in this file reduce.  Don't use this in your code.
+
+#define GTEST_ASSERT_(expression, on_failure) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (const ::testing::AssertionResult gtest_ar = (expression)) \
+    ; \
+  else \
+    on_failure(gtest_ar.failure_message())
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED1.  Don't use
+// this in your code.
+template <typename Pred,
+          typename T1>
+AssertionResult AssertPred1Helper(const char* pred_text,
+                                  const char* e1,
+                                  Pred pred,
+                                  const T1& v1) {
+  if (pred(v1)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT1.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT1_(pred_format, v1, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, v1), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED1.  Don't use
+// this in your code.
+#define GTEST_PRED1_(pred, v1, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred1Helper(#pred, \
+                                             #v1, \
+                                             pred, \
+                                             v1), on_failure)
+
+// Unary predicate assertion macros.
+#define EXPECT_PRED_FORMAT1(pred_format, v1) \
+  GTEST_PRED_FORMAT1_(pred_format, v1, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED1(pred, v1) \
+  GTEST_PRED1_(pred, v1, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT1(pred_format, v1) \
+  GTEST_PRED_FORMAT1_(pred_format, v1, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED1(pred, v1) \
+  GTEST_PRED1_(pred, v1, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED2.  Don't use
+// this in your code.
+template <typename Pred,
+          typename T1,
+          typename T2>
+AssertionResult AssertPred2Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2) {
+  if (pred(v1, v2)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT2.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT2_(pred_format, v1, v2, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, v1, v2), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED2.  Don't use
+// this in your code.
+#define GTEST_PRED2_(pred, v1, v2, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred2Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             pred, \
+                                             v1, \
+                                             v2), on_failure)
+
+// Binary predicate assertion macros.
+#define EXPECT_PRED_FORMAT2(pred_format, v1, v2) \
+  GTEST_PRED_FORMAT2_(pred_format, v1, v2, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED2(pred, v1, v2) \
+  GTEST_PRED2_(pred, v1, v2, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT2(pred_format, v1, v2) \
+  GTEST_PRED_FORMAT2_(pred_format, v1, v2, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED2(pred, v1, v2) \
+  GTEST_PRED2_(pred, v1, v2, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED3.  Don't use
+// this in your code.
+template <typename Pred,
+          typename T1,
+          typename T2,
+          typename T3>
+AssertionResult AssertPred3Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  const char* e3,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2,
+                                  const T3& v3) {
+  if (pred(v1, v2, v3)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ", "
+                            << e3 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2
+                            << "\n" << e3 << " evaluates to " << v3;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT3.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT3_(pred_format, v1, v2, v3, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, #v3, v1, v2, v3), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED3.  Don't use
+// this in your code.
+#define GTEST_PRED3_(pred, v1, v2, v3, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred3Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             #v3, \
+                                             pred, \
+                                             v1, \
+                                             v2, \
+                                             v3), on_failure)
+
+// Ternary predicate assertion macros.
+#define EXPECT_PRED_FORMAT3(pred_format, v1, v2, v3) \
+  GTEST_PRED_FORMAT3_(pred_format, v1, v2, v3, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED3(pred, v1, v2, v3) \
+  GTEST_PRED3_(pred, v1, v2, v3, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT3(pred_format, v1, v2, v3) \
+  GTEST_PRED_FORMAT3_(pred_format, v1, v2, v3, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED3(pred, v1, v2, v3) \
+  GTEST_PRED3_(pred, v1, v2, v3, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED4.  Don't use
+// this in your code.
+template <typename Pred,
+          typename T1,
+          typename T2,
+          typename T3,
+          typename T4>
+AssertionResult AssertPred4Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  const char* e3,
+                                  const char* e4,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2,
+                                  const T3& v3,
+                                  const T4& v4) {
+  if (pred(v1, v2, v3, v4)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ", "
+                            << e3 << ", "
+                            << e4 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2
+                            << "\n" << e3 << " evaluates to " << v3
+                            << "\n" << e4 << " evaluates to " << v4;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT4.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT4_(pred_format, v1, v2, v3, v4, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, #v3, #v4, v1, v2, v3, v4), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED4.  Don't use
+// this in your code.
+#define GTEST_PRED4_(pred, v1, v2, v3, v4, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred4Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             #v3, \
+                                             #v4, \
+                                             pred, \
+                                             v1, \
+                                             v2, \
+                                             v3, \
+                                             v4), on_failure)
+
+// 4-ary predicate assertion macros.
+#define EXPECT_PRED_FORMAT4(pred_format, v1, v2, v3, v4) \
+  GTEST_PRED_FORMAT4_(pred_format, v1, v2, v3, v4, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED4(pred, v1, v2, v3, v4) \
+  GTEST_PRED4_(pred, v1, v2, v3, v4, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT4(pred_format, v1, v2, v3, v4) \
+  GTEST_PRED_FORMAT4_(pred_format, v1, v2, v3, v4, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED4(pred, v1, v2, v3, v4) \
+  GTEST_PRED4_(pred, v1, v2, v3, v4, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED5.  Don't use
+// this in your code.
+template <typename Pred,
+          typename T1,
+          typename T2,
+          typename T3,
+          typename T4,
+          typename T5>
+AssertionResult AssertPred5Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  const char* e3,
+                                  const char* e4,
+                                  const char* e5,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2,
+                                  const T3& v3,
+                                  const T4& v4,
+                                  const T5& v5) {
+  if (pred(v1, v2, v3, v4, v5)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ", "
+                            << e3 << ", "
+                            << e4 << ", "
+                            << e5 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2
+                            << "\n" << e3 << " evaluates to " << v3
+                            << "\n" << e4 << " evaluates to " << v4
+                            << "\n" << e5 << " evaluates to " << v5;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT5.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT5_(pred_format, v1, v2, v3, v4, v5, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, #v3, #v4, #v5, v1, v2, v3, v4, v5), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED5.  Don't use
+// this in your code.
+#define GTEST_PRED5_(pred, v1, v2, v3, v4, v5, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred5Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             #v3, \
+                                             #v4, \
+                                             #v5, \
+                                             pred, \
+                                             v1, \
+                                             v2, \
+                                             v3, \
+                                             v4, \
+                                             v5), on_failure)
+
+// 5-ary predicate assertion macros.
+#define EXPECT_PRED_FORMAT5(pred_format, v1, v2, v3, v4, v5) \
+  GTEST_PRED_FORMAT5_(pred_format, v1, v2, v3, v4, v5, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED5(pred, v1, v2, v3, v4, v5) \
+  GTEST_PRED5_(pred, v1, v2, v3, v4, v5, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT5(pred_format, v1, v2, v3, v4, v5) \
+  GTEST_PRED_FORMAT5_(pred_format, v1, v2, v3, v4, v5, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED5(pred, v1, v2, v3, v4, v5) \
+  GTEST_PRED5_(pred, v1, v2, v3, v4, v5, GTEST_FATAL_FAILURE_)
+
+
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PRED_IMPL_H_
+
+// Macros for testing equalities and inequalities.
+//
+//    * {ASSERT|EXPECT}_EQ(expected, actual): Tests that expected == actual
+//    * {ASSERT|EXPECT}_NE(v1, v2):           Tests that v1 != v2
+//    * {ASSERT|EXPECT}_LT(v1, v2):           Tests that v1 < v2
+//    * {ASSERT|EXPECT}_LE(v1, v2):           Tests that v1 <= v2
+//    * {ASSERT|EXPECT}_GT(v1, v2):           Tests that v1 > v2
+//    * {ASSERT|EXPECT}_GE(v1, v2):           Tests that v1 >= v2
+//
+// When they are not, Google Test prints both the tested expressions and
+// their actual values.  The values must be compatible built-in types,
+// or you will get a compiler error.  By "compatible" we mean that the
+// values can be compared by the respective operator.
+//
+// Note:
+//
+//   1. It is possible to make a user-defined type work with
+//   {ASSERT|EXPECT}_??(), but that requires overloading the
+//   comparison operators and is thus discouraged by the Google C++
+//   Usage Guide.  Therefore, you are advised to use the
+//   {ASSERT|EXPECT}_TRUE() macro to assert that two objects are
+//   equal.
+//
+//   2. The {ASSERT|EXPECT}_??() macros do pointer comparisons on
+//   pointers (in particular, C strings).  Therefore, if you use it
+//   with two C strings, you are testing how their locations in memory
+//   are related, not how their content is related.  To compare two C
+//   strings by content, use {ASSERT|EXPECT}_STR*().
+//
+//   3. {ASSERT|EXPECT}_EQ(expected, actual) is preferred to
+//   {ASSERT|EXPECT}_TRUE(expected == actual), as the former tells you
+//   what the actual value is when it fails, and similarly for the
+//   other comparisons.
+//
+//   4. Do not depend on the order in which {ASSERT|EXPECT}_??()
+//   evaluate their arguments, which is undefined.
+//
+//   5. These macros evaluate their arguments exactly once.
+//
+// Examples:
+//
+//   EXPECT_NE(5, Foo());
+//   EXPECT_EQ(NULL, a_pointer);
+//   ASSERT_LT(i, array_size);
+//   ASSERT_GT(records.size(), 0) << "There is no record left.";
+
+#define EXPECT_EQ(expected, actual) \
+  EXPECT_PRED_FORMAT2(::testing::internal:: \
+                      EqHelper<GTEST_IS_NULL_LITERAL_(expected)>::Compare, \
+                      expected, actual)
+#define EXPECT_NE(expected, actual) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperNE, expected, actual)
+#define EXPECT_LE(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperLE, val1, val2)
+#define EXPECT_LT(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperLT, val1, val2)
+#define EXPECT_GE(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperGE, val1, val2)
+#define EXPECT_GT(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperGT, val1, val2)
+
+#define GTEST_ASSERT_EQ(expected, actual) \
+  ASSERT_PRED_FORMAT2(::testing::internal:: \
+                      EqHelper<GTEST_IS_NULL_LITERAL_(expected)>::Compare, \
+                      expected, actual)
+#define GTEST_ASSERT_NE(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperNE, val1, val2)
+#define GTEST_ASSERT_LE(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperLE, val1, val2)
+#define GTEST_ASSERT_LT(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperLT, val1, val2)
+#define GTEST_ASSERT_GE(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperGE, val1, val2)
+#define GTEST_ASSERT_GT(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperGT, val1, val2)
+
+// Define macro GTEST_DONT_DEFINE_ASSERT_XY to 1 to omit the definition of
+// ASSERT_XY(), which clashes with some users' own code.
+
+#if !GTEST_DONT_DEFINE_ASSERT_EQ
+# define ASSERT_EQ(val1, val2) GTEST_ASSERT_EQ(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_NE
+# define ASSERT_NE(val1, val2) GTEST_ASSERT_NE(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_LE
+# define ASSERT_LE(val1, val2) GTEST_ASSERT_LE(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_LT
+# define ASSERT_LT(val1, val2) GTEST_ASSERT_LT(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_GE
+# define ASSERT_GE(val1, val2) GTEST_ASSERT_GE(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_GT
+# define ASSERT_GT(val1, val2) GTEST_ASSERT_GT(val1, val2)
+#endif
+
+// C-string Comparisons.  All tests treat NULL and any non-NULL string
+// as different.  Two NULLs are equal.
+//
+//    * {ASSERT|EXPECT}_STREQ(s1, s2):     Tests that s1 == s2
+//    * {ASSERT|EXPECT}_STRNE(s1, s2):     Tests that s1 != s2
+//    * {ASSERT|EXPECT}_STRCASEEQ(s1, s2): Tests that s1 == s2, ignoring case
+//    * {ASSERT|EXPECT}_STRCASENE(s1, s2): Tests that s1 != s2, ignoring case
+//
+// For wide or narrow string objects, you can use the
+// {ASSERT|EXPECT}_??() macros.
+//
+// Don't depend on the order in which the arguments are evaluated,
+// which is undefined.
+//
+// These macros evaluate their arguments exactly once.
+
+#define EXPECT_STREQ(expected, actual) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTREQ, expected, actual)
+#define EXPECT_STRNE(s1, s2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTRNE, s1, s2)
+#define EXPECT_STRCASEEQ(expected, actual) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASEEQ, expected, actual)
+#define EXPECT_STRCASENE(s1, s2)\
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASENE, s1, s2)
+
+#define ASSERT_STREQ(expected, actual) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTREQ, expected, actual)
+#define ASSERT_STRNE(s1, s2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTRNE, s1, s2)
+#define ASSERT_STRCASEEQ(expected, actual) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASEEQ, expected, actual)
+#define ASSERT_STRCASENE(s1, s2)\
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASENE, s1, s2)
+
+// Macros for comparing floating-point numbers.
+//
+//    * {ASSERT|EXPECT}_FLOAT_EQ(expected, actual):
+//         Tests that two float values are almost equal.
+//    * {ASSERT|EXPECT}_DOUBLE_EQ(expected, actual):
+//         Tests that two double values are almost equal.
+//    * {ASSERT|EXPECT}_NEAR(v1, v2, abs_error):
+//         Tests that v1 and v2 are within the given distance to each other.
+//
+// Google Test uses ULP-based comparison to automatically pick a default
+// error bound that is appropriate for the operands.  See the
+// FloatingPoint template class in gtest-internal.h if you are
+// interested in the implementation details.
+
+#define EXPECT_FLOAT_EQ(expected, actual)\
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ<float>, \
+                      expected, actual)
+
+#define EXPECT_DOUBLE_EQ(expected, actual)\
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ<double>, \
+                      expected, actual)
+
+#define ASSERT_FLOAT_EQ(expected, actual)\
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ<float>, \
+                      expected, actual)
+
+#define ASSERT_DOUBLE_EQ(expected, actual)\
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ<double>, \
+                      expected, actual)
+
+#define EXPECT_NEAR(val1, val2, abs_error)\
+  EXPECT_PRED_FORMAT3(::testing::internal::DoubleNearPredFormat, \
+                      val1, val2, abs_error)
+
+#define ASSERT_NEAR(val1, val2, abs_error)\
+  ASSERT_PRED_FORMAT3(::testing::internal::DoubleNearPredFormat, \
+                      val1, val2, abs_error)
+
+// These predicate format functions work on floating-point values, and
+// can be used in {ASSERT|EXPECT}_PRED_FORMAT2*(), e.g.
+//
+//   EXPECT_PRED_FORMAT2(testing::DoubleLE, Foo(), 5.0);
+
+// Asserts that val1 is less than, or almost equal to, val2.  Fails
+// otherwise.  In particular, it fails if either val1 or val2 is NaN.
+GTEST_API_ AssertionResult FloatLE(const char* expr1, const char* expr2,
+                                   float val1, float val2);
+GTEST_API_ AssertionResult DoubleLE(const char* expr1, const char* expr2,
+                                    double val1, double val2);
+
+
+#if GTEST_OS_WINDOWS
+
+// Macros that test for HRESULT failure and success, these are only useful
+// on Windows, and rely on Windows SDK macros and APIs to compile.
+//
+//    * {ASSERT|EXPECT}_HRESULT_{SUCCEEDED|FAILED}(expr)
+//
+// When expr unexpectedly fails or succeeds, Google Test prints the
+// expected result and the actual result with both a human-readable
+// string representation of the error, if available, as well as the
+// hex result code.
+# define EXPECT_HRESULT_SUCCEEDED(expr) \
+    EXPECT_PRED_FORMAT1(::testing::internal::IsHRESULTSuccess, (expr))
+
+# define ASSERT_HRESULT_SUCCEEDED(expr) \
+    ASSERT_PRED_FORMAT1(::testing::internal::IsHRESULTSuccess, (expr))
+
+# define EXPECT_HRESULT_FAILED(expr) \
+    EXPECT_PRED_FORMAT1(::testing::internal::IsHRESULTFailure, (expr))
+
+# define ASSERT_HRESULT_FAILED(expr) \
+    ASSERT_PRED_FORMAT1(::testing::internal::IsHRESULTFailure, (expr))
+
+#endif  // GTEST_OS_WINDOWS
+
+// Macros that execute statement and check that it doesn't generate new fatal
+// failures in the current thread.
+//
+//   * {ASSERT|EXPECT}_NO_FATAL_FAILURE(statement);
+//
+// Examples:
+//
+//   EXPECT_NO_FATAL_FAILURE(Process());
+//   ASSERT_NO_FATAL_FAILURE(Process()) << "Process() failed";
+//
+#define ASSERT_NO_FATAL_FAILURE(statement) \
+    GTEST_TEST_NO_FATAL_FAILURE_(statement, GTEST_FATAL_FAILURE_)
+#define EXPECT_NO_FATAL_FAILURE(statement) \
+    GTEST_TEST_NO_FATAL_FAILURE_(statement, GTEST_NONFATAL_FAILURE_)
+
+// Causes a trace (including the source file path, the current line
+// number, and the given message) to be included in every test failure
+// message generated by code in the current scope.  The effect is
+// undone when the control leaves the current scope.
+//
+// The message argument can be anything streamable to std::ostream.
+//
+// In the implementation, we include the current line number as part
+// of the dummy variable name, thus allowing multiple SCOPED_TRACE()s
+// to appear in the same block - as long as they are on different
+// lines.
+#define SCOPED_TRACE(message) \
+  ::testing::internal::ScopedTrace GTEST_CONCAT_TOKEN_(gtest_trace_, __LINE__)(\
+    __FILE__, __LINE__, ::testing::Message() << (message))
+
+// Compile-time assertion for type equality.
+// StaticAssertTypeEq<type1, type2>() compiles iff type1 and type2 are
+// the same type.  The value it returns is not interesting.
+//
+// Instead of making StaticAssertTypeEq a class template, we make it a
+// function template that invokes a helper class template.  This
+// prevents a user from misusing StaticAssertTypeEq<T1, T2> by
+// defining objects of that type.
+//
+// CAVEAT:
+//
+// When used inside a method of a class template,
+// StaticAssertTypeEq<T1, T2>() is effective ONLY IF the method is
+// instantiated.  For example, given:
+//
+//   template <typename T> class Foo {
+//    public:
+//     void Bar() { testing::StaticAssertTypeEq<int, T>(); }
+//   };
+//
+// the code:
+//
+//   void Test1() { Foo<bool> foo; }
+//
+// will NOT generate a compiler error, as Foo<bool>::Bar() is never
+// actually instantiated.  Instead, you need:
+//
+//   void Test2() { Foo<bool> foo; foo.Bar(); }
+//
+// to cause a compiler error.
+template <typename T1, typename T2>
+bool StaticAssertTypeEq() {
+  (void)internal::StaticAssertTypeEqHelper<T1, T2>();
+  return true;
+}
+
+// Defines a test.
+//
+// The first parameter is the name of the test case, and the second
+// parameter is the name of the test within the test case.
+//
+// The convention is to end the test case name with "Test".  For
+// example, a test case for the Foo class can be named FooTest.
+//
+// The user should put his test code between braces after using this
+// macro.  Example:
+//
+//   TEST(FooTest, InitializesCorrectly) {
+//     Foo foo;
+//     EXPECT_TRUE(foo.StatusIsOK());
+//   }
+
+// Note that we call GetTestTypeId() instead of GetTypeId<
+// ::testing::Test>() here to get the type ID of testing::Test.  This
+// is to work around a suspected linker bug when using Google Test as
+// a framework on Mac OS X.  The bug causes GetTypeId<
+// ::testing::Test>() to return different values depending on whether
+// the call is from the Google Test framework itself or from user test
+// code.  GetTestTypeId() is guaranteed to always return the same
+// value, as it always calls GetTypeId<>() from the Google Test
+// framework.
+#define GTEST_TEST(test_case_name, test_name)\
+  GTEST_TEST_(test_case_name, test_name, \
+              ::testing::Test, ::testing::internal::GetTestTypeId())
+
+// Define this macro to 1 to omit the definition of TEST(), which
+// is a generic name and clashes with some other libraries.
+#if !GTEST_DONT_DEFINE_TEST
+# define TEST(test_case_name, test_name) GTEST_TEST(test_case_name, test_name)
+#endif
+
+// Defines a test that uses a test fixture.
+//
+// The first parameter is the name of the test fixture class, which
+// also doubles as the test case name.  The second parameter is the
+// name of the test within the test case.
+//
+// A test fixture class must be declared earlier.  The user should put
+// his test code between braces after using this macro.  Example:
+//
+//   class FooTest : public testing::Test {
+//    protected:
+//     virtual void SetUp() { b_.AddElement(3); }
+//
+//     Foo a_;
+//     Foo b_;
+//   };
+//
+//   TEST_F(FooTest, InitializesCorrectly) {
+//     EXPECT_TRUE(a_.StatusIsOK());
+//   }
+//
+//   TEST_F(FooTest, ReturnsElementCountCorrectly) {
+//     EXPECT_EQ(0, a_.size());
+//     EXPECT_EQ(1, b_.size());
+//   }
+
+#define TEST_F(test_fixture, test_name)\
+  GTEST_TEST_(test_fixture, test_name, test_fixture, \
+              ::testing::internal::GetTypeId<test_fixture>())
+
+}  // namespace testing
+
+// Use this function in main() to run all tests.  It returns 0 if all
+// tests are successful, or 1 otherwise.
+//
+// RUN_ALL_TESTS() should be invoked after the command line has been
+// parsed by InitGoogleTest().
+//
+// This function was formerly a macro; thus, it is in the global
+// namespace and has an all-caps name.
+int RUN_ALL_TESTS() GTEST_MUST_USE_RESULT_;
+
+inline int RUN_ALL_TESTS() {
+  return ::testing::UnitTest::GetInstance()->Run();
+}
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_H_
diff --git a/libvpx/libvpx/third_party/googletest/src/src/gtest-all.cc b/libvpx/libvpx/third_party/googletest/src/src/gtest-all.cc
new file mode 100644
index 0000000..8d90627
--- /dev/null
+++ b/libvpx/libvpx/third_party/googletest/src/src/gtest-all.cc
@@ -0,0 +1,9590 @@
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: mheule@google.com (Markus Heule)
+//
+// Google C++ Testing Framework (Google Test)
+//
+// Sometimes it's desirable to build Google Test by compiling a single file.
+// This file serves this purpose.
+
+// This line ensures that gtest.h can be compiled on its own, even
+// when it's fused.
+#include "gtest/gtest.h"
+
+// The following lines pull in the real gtest *.cc files.
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+
+// Copyright 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// Utilities for testing Google Test itself and code that uses Google Test
+// (e.g. frameworks built on top of Google Test).
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_SPI_H_
+#define GTEST_INCLUDE_GTEST_GTEST_SPI_H_
+
+
+namespace testing {
+
+// This helper class can be used to mock out Google Test failure reporting
+// so that we can test Google Test or code that builds on Google Test.
+//
+// An object of this class appends a TestPartResult object to the
+// TestPartResultArray object given in the constructor whenever a Google Test
+// failure is reported. It can either intercept only failures that are
+// generated in the same thread that created this object or it can intercept
+// all generated failures. The scope of this mock object can be controlled with
+// the second argument to the two arguments constructor.
+class GTEST_API_ ScopedFakeTestPartResultReporter
+    : public TestPartResultReporterInterface {
+ public:
+  // The two possible mocking modes of this object.
+  enum InterceptMode {
+    INTERCEPT_ONLY_CURRENT_THREAD,  // Intercepts only thread local failures.
+    INTERCEPT_ALL_THREADS           // Intercepts all failures.
+  };
+
+  // The c'tor sets this object as the test part result reporter used
+  // by Google Test.  The 'result' parameter specifies where to report the
+  // results. This reporter will only catch failures generated in the current
+  // thread. DEPRECATED
+  explicit ScopedFakeTestPartResultReporter(TestPartResultArray* result);
+
+  // Same as above, but you can choose the interception scope of this object.
+  ScopedFakeTestPartResultReporter(InterceptMode intercept_mode,
+                                   TestPartResultArray* result);
+
+  // The d'tor restores the previous test part result reporter.
+  virtual ~ScopedFakeTestPartResultReporter();
+
+  // Appends the TestPartResult object to the TestPartResultArray
+  // received in the constructor.
+  //
+  // This method is from the TestPartResultReporterInterface
+  // interface.
+  virtual void ReportTestPartResult(const TestPartResult& result);
+ private:
+  void Init();
+
+  const InterceptMode intercept_mode_;
+  TestPartResultReporterInterface* old_reporter_;
+  TestPartResultArray* const result_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ScopedFakeTestPartResultReporter);
+};
+
+namespace internal {
+
+// A helper class for implementing EXPECT_FATAL_FAILURE() and
+// EXPECT_NONFATAL_FAILURE().  Its destructor verifies that the given
+// TestPartResultArray contains exactly one failure that has the given
+// type and contains the given substring.  If that's not the case, a
+// non-fatal failure will be generated.
+class GTEST_API_ SingleFailureChecker {
+ public:
+  // The constructor remembers the arguments.
+  SingleFailureChecker(const TestPartResultArray* results,
+                       TestPartResult::Type type,
+                       const string& substr);
+  ~SingleFailureChecker();
+ private:
+  const TestPartResultArray* const results_;
+  const TestPartResult::Type type_;
+  const string substr_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(SingleFailureChecker);
+};
+
+}  // namespace internal
+
+}  // namespace testing
+
+// A set of macros for testing Google Test assertions or code that's expected
+// to generate Google Test fatal failures.  It verifies that the given
+// statement will cause exactly one fatal Google Test failure with 'substr'
+// being part of the failure message.
+//
+// There are two different versions of this macro. EXPECT_FATAL_FAILURE only
+// affects and considers failures generated in the current thread and
+// EXPECT_FATAL_FAILURE_ON_ALL_THREADS does the same but for all threads.
+//
+// The verification of the assertion is done correctly even when the statement
+// throws an exception or aborts the current function.
+//
+// Known restrictions:
+//   - 'statement' cannot reference local non-static variables or
+//     non-static members of the current object.
+//   - 'statement' cannot return a value.
+//   - You cannot stream a failure message to this macro.
+//
+// Note that even though the implementations of the following two
+// macros are much alike, we cannot refactor them to use a common
+// helper macro, due to some peculiarity in how the preprocessor
+// works.  The AcceptsMacroThatExpandsToUnprotectedComma test in
+// gtest_unittest.cc will fail to compile if we do that.
+#define EXPECT_FATAL_FAILURE(statement, substr) \
+  do { \
+    class GTestExpectFatalFailureHelper {\
+     public:\
+      static void Execute() { statement; }\
+    };\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        &gtest_failures, ::testing::TestPartResult::kFatalFailure, (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter:: \
+          INTERCEPT_ONLY_CURRENT_THREAD, &gtest_failures);\
+      GTestExpectFatalFailureHelper::Execute();\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+#define EXPECT_FATAL_FAILURE_ON_ALL_THREADS(statement, substr) \
+  do { \
+    class GTestExpectFatalFailureHelper {\
+     public:\
+      static void Execute() { statement; }\
+    };\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        &gtest_failures, ::testing::TestPartResult::kFatalFailure, (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter:: \
+          INTERCEPT_ALL_THREADS, &gtest_failures);\
+      GTestExpectFatalFailureHelper::Execute();\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+// A macro for testing Google Test assertions or code that's expected to
+// generate Google Test non-fatal failures.  It asserts that the given
+// statement will cause exactly one non-fatal Google Test failure with 'substr'
+// being part of the failure message.
+//
+// There are two different versions of this macro. EXPECT_NONFATAL_FAILURE only
+// affects and considers failures generated in the current thread and
+// EXPECT_NONFATAL_FAILURE_ON_ALL_THREADS does the same but for all threads.
+//
+// 'statement' is allowed to reference local variables and members of
+// the current object.
+//
+// The verification of the assertion is done correctly even when the statement
+// throws an exception or aborts the current function.
+//
+// Known restrictions:
+//   - You cannot stream a failure message to this macro.
+//
+// Note that even though the implementations of the following two
+// macros are much alike, we cannot refactor them to use a common
+// helper macro, due to some peculiarity in how the preprocessor
+// works.  If we do that, the code won't compile when the user gives
+// EXPECT_NONFATAL_FAILURE() a statement that contains a macro that
+// expands to code containing an unprotected comma.  The
+// AcceptsMacroThatExpandsToUnprotectedComma test in gtest_unittest.cc
+// catches that.
+//
+// For the same reason, we have to write
+//   if (::testing::internal::AlwaysTrue()) { statement; }
+// instead of
+//   GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement)
+// to avoid an MSVC warning on unreachable code.
+#define EXPECT_NONFATAL_FAILURE(statement, substr) \
+  do {\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        &gtest_failures, ::testing::TestPartResult::kNonFatalFailure, \
+        (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter:: \
+          INTERCEPT_ONLY_CURRENT_THREAD, &gtest_failures);\
+      if (::testing::internal::AlwaysTrue()) { statement; }\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+#define EXPECT_NONFATAL_FAILURE_ON_ALL_THREADS(statement, substr) \
+  do {\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        &gtest_failures, ::testing::TestPartResult::kNonFatalFailure, \
+        (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter::INTERCEPT_ALL_THREADS, \
+          &gtest_failures);\
+      if (::testing::internal::AlwaysTrue()) { statement; }\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_SPI_H_
+
+#include <ctype.h>
+#include <math.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <time.h>
+#include <wchar.h>
+#include <wctype.h>
+
+#include <algorithm>
+#include <iomanip>
+#include <limits>
+#include <ostream>  // NOLINT
+#include <sstream>
+#include <vector>
+
+#if GTEST_OS_LINUX
+
+// TODO(kenton@google.com): Use autoconf to detect availability of
+// gettimeofday().
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+
+# include <fcntl.h>  // NOLINT
+# include <limits.h>  // NOLINT
+# include <sched.h>  // NOLINT
+// Declares vsnprintf().  This header is not available on Windows.
+# include <strings.h>  // NOLINT
+# include <sys/mman.h>  // NOLINT
+# include <sys/time.h>  // NOLINT
+# include <unistd.h>  // NOLINT
+# include <string>
+
+#elif GTEST_OS_SYMBIAN
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+# include <sys/time.h>  // NOLINT
+
+#elif GTEST_OS_ZOS
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+# include <sys/time.h>  // NOLINT
+
+// On z/OS we additionally need strings.h for strcasecmp.
+# include <strings.h>  // NOLINT
+
+#elif GTEST_OS_WINDOWS_MOBILE  // We are on Windows CE.
+
+# include <windows.h>  // NOLINT
+
+#elif GTEST_OS_WINDOWS  // We are on Windows proper.
+
+# include <io.h>  // NOLINT
+# include <sys/timeb.h>  // NOLINT
+# include <sys/types.h>  // NOLINT
+# include <sys/stat.h>  // NOLINT
+
+# if GTEST_OS_WINDOWS_MINGW
+// MinGW has gettimeofday() but not _ftime64().
+// TODO(kenton@google.com): Use autoconf to detect availability of
+//   gettimeofday().
+// TODO(kenton@google.com): There are other ways to get the time on
+//   Windows, like GetTickCount() or GetSystemTimeAsFileTime().  MinGW
+//   supports these.  consider using them instead.
+#  define GTEST_HAS_GETTIMEOFDAY_ 1
+#  include <sys/time.h>  // NOLINT
+# endif  // GTEST_OS_WINDOWS_MINGW
+
+// cpplint thinks that the header is already included, so we want to
+// silence it.
+# include <windows.h>  // NOLINT
+
+#else
+
+// Assume other platforms have gettimeofday().
+// TODO(kenton@google.com): Use autoconf to detect availability of
+//   gettimeofday().
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+
+// cpplint thinks that the header is already included, so we want to
+// silence it.
+# include <sys/time.h>  // NOLINT
+# include <unistd.h>  // NOLINT
+
+#endif  // GTEST_OS_LINUX
+
+#if GTEST_HAS_EXCEPTIONS
+# include <stdexcept>
+#endif
+
+#if GTEST_CAN_STREAM_RESULTS_
+# include <arpa/inet.h>  // NOLINT
+# include <netdb.h>  // NOLINT
+#endif
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick is to
+// prevent a user from accidentally including gtest-internal-inl.h in
+// his code.
+#define GTEST_IMPLEMENTATION_ 1
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Utility functions and classes used by the Google C++ testing framework.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// This file contains purely Google Test's internal implementation.  Please
+// DO NOT #INCLUDE IT IN A USER PROGRAM.
+
+#ifndef GTEST_SRC_GTEST_INTERNAL_INL_H_
+#define GTEST_SRC_GTEST_INTERNAL_INL_H_
+
+// GTEST_IMPLEMENTATION_ is defined to 1 iff the current translation unit is
+// part of Google Test's implementation; otherwise it's undefined.
+#if !GTEST_IMPLEMENTATION_
+// A user is trying to include this from his code - just say no.
+# error "gtest-internal-inl.h is part of Google Test's internal implementation."
+# error "It must not be included except by Google Test itself."
+#endif  // GTEST_IMPLEMENTATION_
+
+#ifndef _WIN32_WCE
+# include <errno.h>
+#endif  // !_WIN32_WCE
+#include <stddef.h>
+#include <stdlib.h>  // For strtoll/_strtoul64/malloc/free.
+#include <string.h>  // For memmove.
+
+#include <algorithm>
+#include <string>
+#include <vector>
+
+
+#if GTEST_CAN_STREAM_RESULTS_
+# include <arpa/inet.h>  // NOLINT
+# include <netdb.h>  // NOLINT
+#endif
+
+#if GTEST_OS_WINDOWS
+# include <windows.h>  // NOLINT
+#endif  // GTEST_OS_WINDOWS
+
+
+namespace testing {
+
+// Declares the flags.
+//
+// We don't want the users to modify this flag in the code, but want
+// Google Test's own unit tests to be able to access it. Therefore we
+// declare it here as opposed to in gtest.h.
+GTEST_DECLARE_bool_(death_test_use_fork);
+
+namespace internal {
+
+// The value of GetTestTypeId() as seen from within the Google Test
+// library.  This is solely for testing GetTestTypeId().
+GTEST_API_ extern const TypeId kTestTypeIdInGoogleTest;
+
+// Names of the flags (needed for parsing Google Test flags).
+const char kAlsoRunDisabledTestsFlag[] = "also_run_disabled_tests";
+const char kBreakOnFailureFlag[] = "break_on_failure";
+const char kCatchExceptionsFlag[] = "catch_exceptions";
+const char kColorFlag[] = "color";
+const char kFilterFlag[] = "filter";
+const char kListTestsFlag[] = "list_tests";
+const char kOutputFlag[] = "output";
+const char kPrintTimeFlag[] = "print_time";
+const char kRandomSeedFlag[] = "random_seed";
+const char kRepeatFlag[] = "repeat";
+const char kShuffleFlag[] = "shuffle";
+const char kStackTraceDepthFlag[] = "stack_trace_depth";
+const char kStreamResultToFlag[] = "stream_result_to";
+const char kThrowOnFailureFlag[] = "throw_on_failure";
+
+// A valid random seed must be in [1, kMaxRandomSeed].
+const int kMaxRandomSeed = 99999;
+
+// g_help_flag is true iff the --help flag or an equivalent form is
+// specified on the command line.
+GTEST_API_ extern bool g_help_flag;
+
+// Returns the current time in milliseconds.
+GTEST_API_ TimeInMillis GetTimeInMillis();
+
+// Returns true iff Google Test should use colors in the output.
+GTEST_API_ bool ShouldUseColor(bool stdout_is_tty);
+
+// Formats the given time in milliseconds as seconds.
+GTEST_API_ std::string FormatTimeInMillisAsSeconds(TimeInMillis ms);
+
+// Converts the given time in milliseconds to a date string in the ISO 8601
+// format, without the timezone information.  N.B.: due to the use the
+// non-reentrant localtime() function, this function is not thread safe.  Do
+// not use it in any code that can be called from multiple threads.
+GTEST_API_ std::string FormatEpochTimeInMillisAsIso8601(TimeInMillis ms);
+
+// Parses a string for an Int32 flag, in the form of "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+GTEST_API_ bool ParseInt32Flag(
+    const char* str, const char* flag, Int32* value);
+
+// Returns a random seed in range [1, kMaxRandomSeed] based on the
+// given --gtest_random_seed flag value.
+inline int GetRandomSeedFromFlag(Int32 random_seed_flag) {
+  const unsigned int raw_seed = (random_seed_flag == 0) ?
+      static_cast<unsigned int>(GetTimeInMillis()) :
+      static_cast<unsigned int>(random_seed_flag);
+
+  // Normalizes the actual seed to range [1, kMaxRandomSeed] such that
+  // it's easy to type.
+  const int normalized_seed =
+      static_cast<int>((raw_seed - 1U) %
+                       static_cast<unsigned int>(kMaxRandomSeed)) + 1;
+  return normalized_seed;
+}
+
+// Returns the first valid random seed after 'seed'.  The behavior is
+// undefined if 'seed' is invalid.  The seed after kMaxRandomSeed is
+// considered to be 1.
+inline int GetNextRandomSeed(int seed) {
+  GTEST_CHECK_(1 <= seed && seed <= kMaxRandomSeed)
+      << "Invalid random seed " << seed << " - must be in [1, "
+      << kMaxRandomSeed << "].";
+  const int next_seed = seed + 1;
+  return (next_seed > kMaxRandomSeed) ? 1 : next_seed;
+}
+
+// This class saves the values of all Google Test flags in its c'tor, and
+// restores them in its d'tor.
+class GTestFlagSaver {
+ public:
+  // The c'tor.
+  GTestFlagSaver() {
+    also_run_disabled_tests_ = GTEST_FLAG(also_run_disabled_tests);
+    break_on_failure_ = GTEST_FLAG(break_on_failure);
+    catch_exceptions_ = GTEST_FLAG(catch_exceptions);
+    color_ = GTEST_FLAG(color);
+    death_test_style_ = GTEST_FLAG(death_test_style);
+    death_test_use_fork_ = GTEST_FLAG(death_test_use_fork);
+    filter_ = GTEST_FLAG(filter);
+    internal_run_death_test_ = GTEST_FLAG(internal_run_death_test);
+    list_tests_ = GTEST_FLAG(list_tests);
+    output_ = GTEST_FLAG(output);
+    print_time_ = GTEST_FLAG(print_time);
+    random_seed_ = GTEST_FLAG(random_seed);
+    repeat_ = GTEST_FLAG(repeat);
+    shuffle_ = GTEST_FLAG(shuffle);
+    stack_trace_depth_ = GTEST_FLAG(stack_trace_depth);
+    stream_result_to_ = GTEST_FLAG(stream_result_to);
+    throw_on_failure_ = GTEST_FLAG(throw_on_failure);
+  }
+
+  // The d'tor is not virtual.  DO NOT INHERIT FROM THIS CLASS.
+  ~GTestFlagSaver() {
+    GTEST_FLAG(also_run_disabled_tests) = also_run_disabled_tests_;
+    GTEST_FLAG(break_on_failure) = break_on_failure_;
+    GTEST_FLAG(catch_exceptions) = catch_exceptions_;
+    GTEST_FLAG(color) = color_;
+    GTEST_FLAG(death_test_style) = death_test_style_;
+    GTEST_FLAG(death_test_use_fork) = death_test_use_fork_;
+    GTEST_FLAG(filter) = filter_;
+    GTEST_FLAG(internal_run_death_test) = internal_run_death_test_;
+    GTEST_FLAG(list_tests) = list_tests_;
+    GTEST_FLAG(output) = output_;
+    GTEST_FLAG(print_time) = print_time_;
+    GTEST_FLAG(random_seed) = random_seed_;
+    GTEST_FLAG(repeat) = repeat_;
+    GTEST_FLAG(shuffle) = shuffle_;
+    GTEST_FLAG(stack_trace_depth) = stack_trace_depth_;
+    GTEST_FLAG(stream_result_to) = stream_result_to_;
+    GTEST_FLAG(throw_on_failure) = throw_on_failure_;
+  }
+
+ private:
+  // Fields for saving the original values of flags.
+  bool also_run_disabled_tests_;
+  bool break_on_failure_;
+  bool catch_exceptions_;
+  std::string color_;
+  std::string death_test_style_;
+  bool death_test_use_fork_;
+  std::string filter_;
+  std::string internal_run_death_test_;
+  bool list_tests_;
+  std::string output_;
+  bool print_time_;
+  internal::Int32 random_seed_;
+  internal::Int32 repeat_;
+  bool shuffle_;
+  internal::Int32 stack_trace_depth_;
+  std::string stream_result_to_;
+  bool throw_on_failure_;
+} GTEST_ATTRIBUTE_UNUSED_;
+
+// Converts a Unicode code point to a narrow string in UTF-8 encoding.
+// code_point parameter is of type UInt32 because wchar_t may not be
+// wide enough to contain a code point.
+// If the code_point is not a valid Unicode code point
+// (i.e. outside of Unicode range U+0 to U+10FFFF) it will be converted
+// to "(Invalid Unicode 0xXXXXXXXX)".
+GTEST_API_ std::string CodePointToUtf8(UInt32 code_point);
+
+// Converts a wide string to a narrow string in UTF-8 encoding.
+// The wide string is assumed to have the following encoding:
+//   UTF-16 if sizeof(wchar_t) == 2 (on Windows, Cygwin, Symbian OS)
+//   UTF-32 if sizeof(wchar_t) == 4 (on Linux)
+// Parameter str points to a null-terminated wide string.
+// Parameter num_chars may additionally limit the number
+// of wchar_t characters processed. -1 is used when the entire string
+// should be processed.
+// If the string contains code points that are not valid Unicode code points
+// (i.e. outside of Unicode range U+0 to U+10FFFF) they will be output
+// as '(Invalid Unicode 0xXXXXXXXX)'. If the string is in UTF16 encoding
+// and contains invalid UTF-16 surrogate pairs, values in those pairs
+// will be encoded as individual Unicode characters from Basic Normal Plane.
+GTEST_API_ std::string WideStringToUtf8(const wchar_t* str, int num_chars);
+
+// Reads the GTEST_SHARD_STATUS_FILE environment variable, and creates the file
+// if the variable is present. If a file already exists at this location, this
+// function will write over it. If the variable is present, but the file cannot
+// be created, prints an error and exits.
+void WriteToShardStatusFileIfNeeded();
+
+// Checks whether sharding is enabled by examining the relevant
+// environment variable values. If the variables are present,
+// but inconsistent (e.g., shard_index >= total_shards), prints
+// an error and exits. If in_subprocess_for_death_test, sharding is
+// disabled because it must only be applied to the original test
+// process. Otherwise, we could filter out death tests we intended to execute.
+GTEST_API_ bool ShouldShard(const char* total_shards_str,
+                            const char* shard_index_str,
+                            bool in_subprocess_for_death_test);
+
+// Parses the environment variable var as an Int32. If it is unset,
+// returns default_val. If it is not an Int32, prints an error and
+// and aborts.
+GTEST_API_ Int32 Int32FromEnvOrDie(const char* env_var, Int32 default_val);
+
+// Given the total number of shards, the shard index, and the test id,
+// returns true iff the test should be run on this shard. The test id is
+// some arbitrary but unique non-negative integer assigned to each test
+// method. Assumes that 0 <= shard_index < total_shards.
+GTEST_API_ bool ShouldRunTestOnShard(
+    int total_shards, int shard_index, int test_id);
+
+// STL container utilities.
+
+// Returns the number of elements in the given container that satisfy
+// the given predicate.
+template <class Container, typename Predicate>
+inline int CountIf(const Container& c, Predicate predicate) {
+  // Implemented as an explicit loop since std::count_if() in libCstd on
+  // Solaris has a non-standard signature.
+  int count = 0;
+  for (typename Container::const_iterator it = c.begin(); it != c.end(); ++it) {
+    if (predicate(*it))
+      ++count;
+  }
+  return count;
+}
+
+// Applies a function/functor to each element in the container.
+template <class Container, typename Functor>
+void ForEach(const Container& c, Functor functor) {
+  std::for_each(c.begin(), c.end(), functor);
+}
+
+// Returns the i-th element of the vector, or default_value if i is not
+// in range [0, v.size()).
+template <typename E>
+inline E GetElementOr(const std::vector<E>& v, int i, E default_value) {
+  return (i < 0 || i >= static_cast<int>(v.size())) ? default_value : v[i];
+}
+
+// Performs an in-place shuffle of a range of the vector's elements.
+// 'begin' and 'end' are element indices as an STL-style range;
+// i.e. [begin, end) are shuffled, where 'end' == size() means to
+// shuffle to the end of the vector.
+template <typename E>
+void ShuffleRange(internal::Random* random, int begin, int end,
+                  std::vector<E>* v) {
+  const int size = static_cast<int>(v->size());
+  GTEST_CHECK_(0 <= begin && begin <= size)
+      << "Invalid shuffle range start " << begin << ": must be in range [0, "
+      << size << "].";
+  GTEST_CHECK_(begin <= end && end <= size)
+      << "Invalid shuffle range finish " << end << ": must be in range ["
+      << begin << ", " << size << "].";
+
+  // Fisher-Yates shuffle, from
+  // http://en.wikipedia.org/wiki/Fisher-Yates_shuffle
+  for (int range_width = end - begin; range_width >= 2; range_width--) {
+    const int last_in_range = begin + range_width - 1;
+    const int selected = begin + random->Generate(range_width);
+    std::swap((*v)[selected], (*v)[last_in_range]);
+  }
+}
+
+// Performs an in-place shuffle of the vector's elements.
+template <typename E>
+inline void Shuffle(internal::Random* random, std::vector<E>* v) {
+  ShuffleRange(random, 0, static_cast<int>(v->size()), v);
+}
+
+// A function for deleting an object.  Handy for being used as a
+// functor.
+template <typename T>
+static void Delete(T* x) {
+  delete x;
+}
+
+// A predicate that checks the key of a TestProperty against a known key.
+//
+// TestPropertyKeyIs is copyable.
+class TestPropertyKeyIs {
+ public:
+  // Constructor.
+  //
+  // TestPropertyKeyIs has NO default constructor.
+  explicit TestPropertyKeyIs(const std::string& key) : key_(key) {}
+
+  // Returns true iff the test name of test property matches on key_.
+  bool operator()(const TestProperty& test_property) const {
+    return test_property.key() == key_;
+  }
+
+ private:
+  std::string key_;
+};
+
+// Class UnitTestOptions.
+//
+// This class contains functions for processing options the user
+// specifies when running the tests.  It has only static members.
+//
+// In most cases, the user can specify an option using either an
+// environment variable or a command line flag.  E.g. you can set the
+// test filter using either GTEST_FILTER or --gtest_filter.  If both
+// the variable and the flag are present, the latter overrides the
+// former.
+class GTEST_API_ UnitTestOptions {
+ public:
+  // Functions for processing the gtest_output flag.
+
+  // Returns the output format, or "" for normal printed output.
+  static std::string GetOutputFormat();
+
+  // Returns the absolute path of the requested output file, or the
+  // default (test_detail.xml in the original working directory) if
+  // none was explicitly specified.
+  static std::string GetAbsolutePathToOutputFile();
+
+  // Functions for processing the gtest_filter flag.
+
+  // Returns true iff the wildcard pattern matches the string.  The
+  // first ':' or '\0' character in pattern marks the end of it.
+  //
+  // This recursive algorithm isn't very efficient, but is clear and
+  // works well enough for matching test names, which are short.
+  static bool PatternMatchesString(const char *pattern, const char *str);
+
+  // Returns true iff the user-specified filter matches the test case
+  // name and the test name.
+  static bool FilterMatchesTest(const std::string &test_case_name,
+                                const std::string &test_name);
+
+#if GTEST_OS_WINDOWS
+  // Function for supporting the gtest_catch_exception flag.
+
+  // Returns EXCEPTION_EXECUTE_HANDLER if Google Test should handle the
+  // given SEH exception, or EXCEPTION_CONTINUE_SEARCH otherwise.
+  // This function is useful as an __except condition.
+  static int GTestShouldProcessSEH(DWORD exception_code);
+#endif  // GTEST_OS_WINDOWS
+
+  // Returns true if "name" matches the ':' separated list of glob-style
+  // filters in "filter".
+  static bool MatchesFilter(const std::string& name, const char* filter);
+};
+
+// Returns the current application's name, removing directory path if that
+// is present.  Used by UnitTestOptions::GetOutputFile.
+GTEST_API_ FilePath GetCurrentExecutableName();
+
+// The role interface for getting the OS stack trace as a string.
+class OsStackTraceGetterInterface {
+ public:
+  OsStackTraceGetterInterface() {}
+  virtual ~OsStackTraceGetterInterface() {}
+
+  // Returns the current OS stack trace as an std::string.  Parameters:
+  //
+  //   max_depth  - the maximum number of stack frames to be included
+  //                in the trace.
+  //   skip_count - the number of top frames to be skipped; doesn't count
+  //                against max_depth.
+  virtual string CurrentStackTrace(int max_depth, int skip_count) = 0;
+
+  // UponLeavingGTest() should be called immediately before Google Test calls
+  // user code. It saves some information about the current stack that
+  // CurrentStackTrace() will use to find and hide Google Test stack frames.
+  virtual void UponLeavingGTest() = 0;
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(OsStackTraceGetterInterface);
+};
+
+// A working implementation of the OsStackTraceGetterInterface interface.
+class OsStackTraceGetter : public OsStackTraceGetterInterface {
+ public:
+  OsStackTraceGetter() : caller_frame_(NULL) {}
+
+  virtual string CurrentStackTrace(int max_depth, int skip_count)
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  virtual void UponLeavingGTest() GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // This string is inserted in place of stack frames that are part of
+  // Google Test's implementation.
+  static const char* const kElidedFramesMarker;
+
+ private:
+  Mutex mutex_;  // protects all internal state
+
+  // We save the stack frame below the frame that calls user code.
+  // We do this because the address of the frame immediately below
+  // the user code changes between the call to UponLeavingGTest()
+  // and any calls to CurrentStackTrace() from within the user code.
+  void* caller_frame_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(OsStackTraceGetter);
+};
+
+// Information about a Google Test trace point.
+struct TraceInfo {
+  const char* file;
+  int line;
+  std::string message;
+};
+
+// This is the default global test part result reporter used in UnitTestImpl.
+// This class should only be used by UnitTestImpl.
+class DefaultGlobalTestPartResultReporter
+  : public TestPartResultReporterInterface {
+ public:
+  explicit DefaultGlobalTestPartResultReporter(UnitTestImpl* unit_test);
+  // Implements the TestPartResultReporterInterface. Reports the test part
+  // result in the current test.
+  virtual void ReportTestPartResult(const TestPartResult& result);
+
+ private:
+  UnitTestImpl* const unit_test_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(DefaultGlobalTestPartResultReporter);
+};
+
+// This is the default per thread test part result reporter used in
+// UnitTestImpl. This class should only be used by UnitTestImpl.
+class DefaultPerThreadTestPartResultReporter
+    : public TestPartResultReporterInterface {
+ public:
+  explicit DefaultPerThreadTestPartResultReporter(UnitTestImpl* unit_test);
+  // Implements the TestPartResultReporterInterface. The implementation just
+  // delegates to the current global test part result reporter of *unit_test_.
+  virtual void ReportTestPartResult(const TestPartResult& result);
+
+ private:
+  UnitTestImpl* const unit_test_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(DefaultPerThreadTestPartResultReporter);
+};
+
+// The private implementation of the UnitTest class.  We don't protect
+// the methods under a mutex, as this class is not accessible by a
+// user and the UnitTest class that delegates work to this class does
+// proper locking.
+class GTEST_API_ UnitTestImpl {
+ public:
+  explicit UnitTestImpl(UnitTest* parent);
+  virtual ~UnitTestImpl();
+
+  // There are two different ways to register your own TestPartResultReporter.
+  // You can register your own repoter to listen either only for test results
+  // from the current thread or for results from all threads.
+  // By default, each per-thread test result repoter just passes a new
+  // TestPartResult to the global test result reporter, which registers the
+  // test part result for the currently running test.
+
+  // Returns the global test part result reporter.
+  TestPartResultReporterInterface* GetGlobalTestPartResultReporter();
+
+  // Sets the global test part result reporter.
+  void SetGlobalTestPartResultReporter(
+      TestPartResultReporterInterface* reporter);
+
+  // Returns the test part result reporter for the current thread.
+  TestPartResultReporterInterface* GetTestPartResultReporterForCurrentThread();
+
+  // Sets the test part result reporter for the current thread.
+  void SetTestPartResultReporterForCurrentThread(
+      TestPartResultReporterInterface* reporter);
+
+  // Gets the number of successful test cases.
+  int successful_test_case_count() const;
+
+  // Gets the number of failed test cases.
+  int failed_test_case_count() const;
+
+  // Gets the number of all test cases.
+  int total_test_case_count() const;
+
+  // Gets the number of all test cases that contain at least one test
+  // that should run.
+  int test_case_to_run_count() const;
+
+  // Gets the number of successful tests.
+  int successful_test_count() const;
+
+  // Gets the number of failed tests.
+  int failed_test_count() const;
+
+  // Gets the number of disabled tests that will be reported in the XML report.
+  int reportable_disabled_test_count() const;
+
+  // Gets the number of disabled tests.
+  int disabled_test_count() const;
+
+  // Gets the number of tests to be printed in the XML report.
+  int reportable_test_count() const;
+
+  // Gets the number of all tests.
+  int total_test_count() const;
+
+  // Gets the number of tests that should run.
+  int test_to_run_count() const;
+
+  // Gets the time of the test program start, in ms from the start of the
+  // UNIX epoch.
+  TimeInMillis start_timestamp() const { return start_timestamp_; }
+
+  // Gets the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const { return elapsed_time_; }
+
+  // Returns true iff the unit test passed (i.e. all test cases passed).
+  bool Passed() const { return !Failed(); }
+
+  // Returns true iff the unit test failed (i.e. some test case failed
+  // or something outside of all tests failed).
+  bool Failed() const {
+    return failed_test_case_count() > 0 || ad_hoc_test_result()->Failed();
+  }
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  const TestCase* GetTestCase(int i) const {
+    const int index = GetElementOr(test_case_indices_, i, -1);
+    return index < 0 ? NULL : test_cases_[i];
+  }
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  TestCase* GetMutableTestCase(int i) {
+    const int index = GetElementOr(test_case_indices_, i, -1);
+    return index < 0 ? NULL : test_cases_[index];
+  }
+
+  // Provides access to the event listener list.
+  TestEventListeners* listeners() { return &listeners_; }
+
+  // Returns the TestResult for the test that's currently running, or
+  // the TestResult for the ad hoc test if no test is running.
+  TestResult* current_test_result();
+
+  // Returns the TestResult for the ad hoc test.
+  const TestResult* ad_hoc_test_result() const { return &ad_hoc_test_result_; }
+
+  // Sets the OS stack trace getter.
+  //
+  // Does nothing if the input and the current OS stack trace getter
+  // are the same; otherwise, deletes the old getter and makes the
+  // input the current getter.
+  void set_os_stack_trace_getter(OsStackTraceGetterInterface* getter);
+
+  // Returns the current OS stack trace getter if it is not NULL;
+  // otherwise, creates an OsStackTraceGetter, makes it the current
+  // getter, and returns it.
+  OsStackTraceGetterInterface* os_stack_trace_getter();
+
+  // Returns the current OS stack trace as an std::string.
+  //
+  // The maximum number of stack frames to be included is specified by
+  // the gtest_stack_trace_depth flag.  The skip_count parameter
+  // specifies the number of top frames to be skipped, which doesn't
+  // count against the number of frames to be included.
+  //
+  // For example, if Foo() calls Bar(), which in turn calls
+  // CurrentOsStackTraceExceptTop(1), Foo() will be included in the
+  // trace but Bar() and CurrentOsStackTraceExceptTop() won't.
+  std::string CurrentOsStackTraceExceptTop(int skip_count) GTEST_NO_INLINE_;
+
+  // Finds and returns a TestCase with the given name.  If one doesn't
+  // exist, creates one and returns it.
+  //
+  // Arguments:
+  //
+  //   test_case_name: name of the test case
+  //   type_param:     the name of the test's type parameter, or NULL if
+  //                   this is not a typed or a type-parameterized test.
+  //   set_up_tc:      pointer to the function that sets up the test case
+  //   tear_down_tc:   pointer to the function that tears down the test case
+  TestCase* GetTestCase(const char* test_case_name,
+                        const char* type_param,
+                        Test::SetUpTestCaseFunc set_up_tc,
+                        Test::TearDownTestCaseFunc tear_down_tc);
+
+  // Adds a TestInfo to the unit test.
+  //
+  // Arguments:
+  //
+  //   set_up_tc:    pointer to the function that sets up the test case
+  //   tear_down_tc: pointer to the function that tears down the test case
+  //   test_info:    the TestInfo object
+  void AddTestInfo(Test::SetUpTestCaseFunc set_up_tc,
+                   Test::TearDownTestCaseFunc tear_down_tc,
+                   TestInfo* test_info) {
+    // In order to support thread-safe death tests, we need to
+    // remember the original working directory when the test program
+    // was first invoked.  We cannot do this in RUN_ALL_TESTS(), as
+    // the user may have changed the current directory before calling
+    // RUN_ALL_TESTS().  Therefore we capture the current directory in
+    // AddTestInfo(), which is called to register a TEST or TEST_F
+    // before main() is reached.
+    if (original_working_dir_.IsEmpty()) {
+      original_working_dir_.Set(FilePath::GetCurrentDir());
+      GTEST_CHECK_(!original_working_dir_.IsEmpty())
+          << "Failed to get the current working directory.";
+    }
+
+    GetTestCase(test_info->test_case_name(),
+                test_info->type_param(),
+                set_up_tc,
+                tear_down_tc)->AddTestInfo(test_info);
+  }
+
+#if GTEST_HAS_PARAM_TEST
+  // Returns ParameterizedTestCaseRegistry object used to keep track of
+  // value-parameterized tests and instantiate and register them.
+  internal::ParameterizedTestCaseRegistry& parameterized_test_registry() {
+    return parameterized_test_registry_;
+  }
+#endif  // GTEST_HAS_PARAM_TEST
+
+  // Sets the TestCase object for the test that's currently running.
+  void set_current_test_case(TestCase* a_current_test_case) {
+    current_test_case_ = a_current_test_case;
+  }
+
+  // Sets the TestInfo object for the test that's currently running.  If
+  // current_test_info is NULL, the assertion results will be stored in
+  // ad_hoc_test_result_.
+  void set_current_test_info(TestInfo* a_current_test_info) {
+    current_test_info_ = a_current_test_info;
+  }
+
+  // Registers all parameterized tests defined using TEST_P and
+  // INSTANTIATE_TEST_CASE_P, creating regular tests for each test/parameter
+  // combination. This method can be called more then once; it has guards
+  // protecting from registering the tests more then once.  If
+  // value-parameterized tests are disabled, RegisterParameterizedTests is
+  // present but does nothing.
+  void RegisterParameterizedTests();
+
+  // Runs all tests in this UnitTest object, prints the result, and
+  // returns true if all tests are successful.  If any exception is
+  // thrown during a test, this test is considered to be failed, but
+  // the rest of the tests will still be run.
+  bool RunAllTests();
+
+  // Clears the results of all tests, except the ad hoc tests.
+  void ClearNonAdHocTestResult() {
+    ForEach(test_cases_, TestCase::ClearTestCaseResult);
+  }
+
+  // Clears the results of ad-hoc test assertions.
+  void ClearAdHocTestResult() {
+    ad_hoc_test_result_.Clear();
+  }
+
+  // Adds a TestProperty to the current TestResult object when invoked in a
+  // context of a test or a test case, or to the global property set. If the
+  // result already contains a property with the same key, the value will be
+  // updated.
+  void RecordProperty(const TestProperty& test_property);
+
+  enum ReactionToSharding {
+    HONOR_SHARDING_PROTOCOL,
+    IGNORE_SHARDING_PROTOCOL
+  };
+
+  // Matches the full name of each test against the user-specified
+  // filter to decide whether the test should run, then records the
+  // result in each TestCase and TestInfo object.
+  // If shard_tests == HONOR_SHARDING_PROTOCOL, further filters tests
+  // based on sharding variables in the environment.
+  // Returns the number of tests that should run.
+  int FilterTests(ReactionToSharding shard_tests);
+
+  // Prints the names of the tests matching the user-specified filter flag.
+  void ListTestsMatchingFilter();
+
+  const TestCase* current_test_case() const { return current_test_case_; }
+  TestInfo* current_test_info() { return current_test_info_; }
+  const TestInfo* current_test_info() const { return current_test_info_; }
+
+  // Returns the vector of environments that need to be set-up/torn-down
+  // before/after the tests are run.
+  std::vector<Environment*>& environments() { return environments_; }
+
+  // Getters for the per-thread Google Test trace stack.
+  std::vector<TraceInfo>& gtest_trace_stack() {
+    return *(gtest_trace_stack_.pointer());
+  }
+  const std::vector<TraceInfo>& gtest_trace_stack() const {
+    return gtest_trace_stack_.get();
+  }
+
+#if GTEST_HAS_DEATH_TEST
+  void InitDeathTestSubprocessControlInfo() {
+    internal_run_death_test_flag_.reset(ParseInternalRunDeathTestFlag());
+  }
+  // Returns a pointer to the parsed --gtest_internal_run_death_test
+  // flag, or NULL if that flag was not specified.
+  // This information is useful only in a death test child process.
+  // Must not be called before a call to InitGoogleTest.
+  const InternalRunDeathTestFlag* internal_run_death_test_flag() const {
+    return internal_run_death_test_flag_.get();
+  }
+
+  // Returns a pointer to the current death test factory.
+  internal::DeathTestFactory* death_test_factory() {
+    return death_test_factory_.get();
+  }
+
+  void SuppressTestEventsIfInSubprocess();
+
+  friend class ReplaceDeathTestFactory;
+#endif  // GTEST_HAS_DEATH_TEST
+
+  // Initializes the event listener performing XML output as specified by
+  // UnitTestOptions. Must not be called before InitGoogleTest.
+  void ConfigureXmlOutput();
+
+#if GTEST_CAN_STREAM_RESULTS_
+  // Initializes the event listener for streaming test results to a socket.
+  // Must not be called before InitGoogleTest.
+  void ConfigureStreamingOutput();
+#endif
+
+  // Performs initialization dependent upon flag values obtained in
+  // ParseGoogleTestFlagsOnly.  Is called from InitGoogleTest after the call to
+  // ParseGoogleTestFlagsOnly.  In case a user neglects to call InitGoogleTest
+  // this function is also called from RunAllTests.  Since this function can be
+  // called more than once, it has to be idempotent.
+  void PostFlagParsingInit();
+
+  // Gets the random seed used at the start of the current test iteration.
+  int random_seed() const { return random_seed_; }
+
+  // Gets the random number generator.
+  internal::Random* random() { return &random_; }
+
+  // Shuffles all test cases, and the tests within each test case,
+  // making sure that death tests are still run first.
+  void ShuffleTests();
+
+  // Restores the test cases and tests to their order before the first shuffle.
+  void UnshuffleTests();
+
+  // Returns the value of GTEST_FLAG(catch_exceptions) at the moment
+  // UnitTest::Run() starts.
+  bool catch_exceptions() const { return catch_exceptions_; }
+
+ private:
+  friend class ::testing::UnitTest;
+
+  // Used by UnitTest::Run() to capture the state of
+  // GTEST_FLAG(catch_exceptions) at the moment it starts.
+  void set_catch_exceptions(bool value) { catch_exceptions_ = value; }
+
+  // The UnitTest object that owns this implementation object.
+  UnitTest* const parent_;
+
+  // The working directory when the first TEST() or TEST_F() was
+  // executed.
+  internal::FilePath original_working_dir_;
+
+  // The default test part result reporters.
+  DefaultGlobalTestPartResultReporter default_global_test_part_result_reporter_;
+  DefaultPerThreadTestPartResultReporter
+      default_per_thread_test_part_result_reporter_;
+
+  // Points to (but doesn't own) the global test part result reporter.
+  TestPartResultReporterInterface* global_test_part_result_repoter_;
+
+  // Protects read and write access to global_test_part_result_reporter_.
+  internal::Mutex global_test_part_result_reporter_mutex_;
+
+  // Points to (but doesn't own) the per-thread test part result reporter.
+  internal::ThreadLocal<TestPartResultReporterInterface*>
+      per_thread_test_part_result_reporter_;
+
+  // The vector of environments that need to be set-up/torn-down
+  // before/after the tests are run.
+  std::vector<Environment*> environments_;
+
+  // The vector of TestCases in their original order.  It owns the
+  // elements in the vector.
+  std::vector<TestCase*> test_cases_;
+
+  // Provides a level of indirection for the test case list to allow
+  // easy shuffling and restoring the test case order.  The i-th
+  // element of this vector is the index of the i-th test case in the
+  // shuffled order.
+  std::vector<int> test_case_indices_;
+
+#if GTEST_HAS_PARAM_TEST
+  // ParameterizedTestRegistry object used to register value-parameterized
+  // tests.
+  internal::ParameterizedTestCaseRegistry parameterized_test_registry_;
+
+  // Indicates whether RegisterParameterizedTests() has been called already.
+  bool parameterized_tests_registered_;
+#endif  // GTEST_HAS_PARAM_TEST
+
+  // Index of the last death test case registered.  Initially -1.
+  int last_death_test_case_;
+
+  // This points to the TestCase for the currently running test.  It
+  // changes as Google Test goes through one test case after another.
+  // When no test is running, this is set to NULL and Google Test
+  // stores assertion results in ad_hoc_test_result_.  Initially NULL.
+  TestCase* current_test_case_;
+
+  // This points to the TestInfo for the currently running test.  It
+  // changes as Google Test goes through one test after another.  When
+  // no test is running, this is set to NULL and Google Test stores
+  // assertion results in ad_hoc_test_result_.  Initially NULL.
+  TestInfo* current_test_info_;
+
+  // Normally, a user only writes assertions inside a TEST or TEST_F,
+  // or inside a function called by a TEST or TEST_F.  Since Google
+  // Test keeps track of which test is current running, it can
+  // associate such an assertion with the test it belongs to.
+  //
+  // If an assertion is encountered when no TEST or TEST_F is running,
+  // Google Test attributes the assertion result to an imaginary "ad hoc"
+  // test, and records the result in ad_hoc_test_result_.
+  TestResult ad_hoc_test_result_;
+
+  // The list of event listeners that can be used to track events inside
+  // Google Test.
+  TestEventListeners listeners_;
+
+  // The OS stack trace getter.  Will be deleted when the UnitTest
+  // object is destructed.  By default, an OsStackTraceGetter is used,
+  // but the user can set this field to use a custom getter if that is
+  // desired.
+  OsStackTraceGetterInterface* os_stack_trace_getter_;
+
+  // True iff PostFlagParsingInit() has been called.
+  bool post_flag_parse_init_performed_;
+
+  // The random number seed used at the beginning of the test run.
+  int random_seed_;
+
+  // Our random number generator.
+  internal::Random random_;
+
+  // The time of the test program start, in ms from the start of the
+  // UNIX epoch.
+  TimeInMillis start_timestamp_;
+
+  // How long the test took to run, in milliseconds.
+  TimeInMillis elapsed_time_;
+
+#if GTEST_HAS_DEATH_TEST
+  // The decomposed components of the gtest_internal_run_death_test flag,
+  // parsed when RUN_ALL_TESTS is called.
+  internal::scoped_ptr<InternalRunDeathTestFlag> internal_run_death_test_flag_;
+  internal::scoped_ptr<internal::DeathTestFactory> death_test_factory_;
+#endif  // GTEST_HAS_DEATH_TEST
+
+  // A per-thread stack of traces created by the SCOPED_TRACE() macro.
+  internal::ThreadLocal<std::vector<TraceInfo> > gtest_trace_stack_;
+
+  // The value of GTEST_FLAG(catch_exceptions) at the moment RunAllTests()
+  // starts.
+  bool catch_exceptions_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(UnitTestImpl);
+};  // class UnitTestImpl
+
+// Convenience function for accessing the global UnitTest
+// implementation object.
+inline UnitTestImpl* GetUnitTestImpl() {
+  return UnitTest::GetInstance()->impl();
+}
+
+#if GTEST_USES_SIMPLE_RE
+
+// Internal helper functions for implementing the simple regular
+// expression matcher.
+GTEST_API_ bool IsInSet(char ch, const char* str);
+GTEST_API_ bool IsAsciiDigit(char ch);
+GTEST_API_ bool IsAsciiPunct(char ch);
+GTEST_API_ bool IsRepeat(char ch);
+GTEST_API_ bool IsAsciiWhiteSpace(char ch);
+GTEST_API_ bool IsAsciiWordChar(char ch);
+GTEST_API_ bool IsValidEscape(char ch);
+GTEST_API_ bool AtomMatchesChar(bool escaped, char pattern, char ch);
+GTEST_API_ bool ValidateRegex(const char* regex);
+GTEST_API_ bool MatchRegexAtHead(const char* regex, const char* str);
+GTEST_API_ bool MatchRepetitionAndRegexAtHead(
+    bool escaped, char ch, char repeat, const char* regex, const char* str);
+GTEST_API_ bool MatchRegexAnywhere(const char* regex, const char* str);
+
+#endif  // GTEST_USES_SIMPLE_RE
+
+// Parses the command line for Google Test flags, without initializing
+// other parts of Google Test.
+GTEST_API_ void ParseGoogleTestFlagsOnly(int* argc, char** argv);
+GTEST_API_ void ParseGoogleTestFlagsOnly(int* argc, wchar_t** argv);
+
+#if GTEST_HAS_DEATH_TEST
+
+// Returns the message describing the last system error, regardless of the
+// platform.
+GTEST_API_ std::string GetLastErrnoDescription();
+
+# if GTEST_OS_WINDOWS
+// Provides leak-safe Windows kernel handle ownership.
+class AutoHandle {
+ public:
+  AutoHandle() : handle_(INVALID_HANDLE_VALUE) {}
+  explicit AutoHandle(HANDLE handle) : handle_(handle) {}
+
+  ~AutoHandle() { Reset(); }
+
+  HANDLE Get() const { return handle_; }
+  void Reset() { Reset(INVALID_HANDLE_VALUE); }
+  void Reset(HANDLE handle) {
+    if (handle != handle_) {
+      if (handle_ != INVALID_HANDLE_VALUE)
+        ::CloseHandle(handle_);
+      handle_ = handle;
+    }
+  }
+
+ private:
+  HANDLE handle_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(AutoHandle);
+};
+# endif  // GTEST_OS_WINDOWS
+
+// Attempts to parse a string into a positive integer pointed to by the
+// number parameter.  Returns true if that is possible.
+// GTEST_HAS_DEATH_TEST implies that we have ::std::string, so we can use
+// it here.
+template <typename Integer>
+bool ParseNaturalNumber(const ::std::string& str, Integer* number) {
+  // Fail fast if the given string does not begin with a digit;
+  // this bypasses strtoXXX's "optional leading whitespace and plus
+  // or minus sign" semantics, which are undesirable here.
+  if (str.empty() || !IsDigit(str[0])) {
+    return false;
+  }
+  errno = 0;
+
+  char* end;
+  // BiggestConvertible is the largest integer type that system-provided
+  // string-to-number conversion routines can return.
+
+# if GTEST_OS_WINDOWS && !defined(__GNUC__)
+
+  // MSVC and C++ Builder define __int64 instead of the standard long long.
+  typedef unsigned __int64 BiggestConvertible;
+  const BiggestConvertible parsed = _strtoui64(str.c_str(), &end, 10);
+
+# else
+
+  typedef unsigned long long BiggestConvertible;  // NOLINT
+  const BiggestConvertible parsed = strtoull(str.c_str(), &end, 10);
+
+# endif  // GTEST_OS_WINDOWS && !defined(__GNUC__)
+
+  const bool parse_success = *end == '\0' && errno == 0;
+
+  // TODO(vladl@google.com): Convert this to compile time assertion when it is
+  // available.
+  GTEST_CHECK_(sizeof(Integer) <= sizeof(parsed));
+
+  const Integer result = static_cast<Integer>(parsed);
+  if (parse_success && static_cast<BiggestConvertible>(result) == parsed) {
+    *number = result;
+    return true;
+  }
+  return false;
+}
+#endif  // GTEST_HAS_DEATH_TEST
+
+// TestResult contains some private methods that should be hidden from
+// Google Test user but are required for testing. This class allow our tests
+// to access them.
+//
+// This class is supplied only for the purpose of testing Google Test's own
+// constructs. Do not use it in user tests, either directly or indirectly.
+class TestResultAccessor {
+ public:
+  static void RecordProperty(TestResult* test_result,
+                             const std::string& xml_element,
+                             const TestProperty& property) {
+    test_result->RecordProperty(xml_element, property);
+  }
+
+  static void ClearTestPartResults(TestResult* test_result) {
+    test_result->ClearTestPartResults();
+  }
+
+  static const std::vector<testing::TestPartResult>& test_part_results(
+      const TestResult& test_result) {
+    return test_result.test_part_results();
+  }
+};
+
+#if GTEST_CAN_STREAM_RESULTS_
+
+// Streams test results to the given port on the given host machine.
+class StreamingListener : public EmptyTestEventListener {
+ public:
+  // Abstract base class for writing strings to a socket.
+  class AbstractSocketWriter {
+   public:
+    virtual ~AbstractSocketWriter() {}
+
+    // Sends a string to the socket.
+    virtual void Send(const string& message) = 0;
+
+    // Closes the socket.
+    virtual void CloseConnection() {}
+
+    // Sends a string and a newline to the socket.
+    void SendLn(const string& message) {
+      Send(message + "\n");
+    }
+  };
+
+  // Concrete class for actually writing strings to a socket.
+  class SocketWriter : public AbstractSocketWriter {
+   public:
+    SocketWriter(const string& host, const string& port)
+        : sockfd_(-1), host_name_(host), port_num_(port) {
+      MakeConnection();
+    }
+
+    virtual ~SocketWriter() {
+      if (sockfd_ != -1)
+        CloseConnection();
+    }
+
+    // Sends a string to the socket.
+    virtual void Send(const string& message) {
+      GTEST_CHECK_(sockfd_ != -1)
+          << "Send() can be called only when there is a connection.";
+
+      const int len = static_cast<int>(message.length());
+      if (write(sockfd_, message.c_str(), len) != len) {
+        GTEST_LOG_(WARNING)
+            << "stream_result_to: failed to stream to "
+            << host_name_ << ":" << port_num_;
+      }
+    }
+
+   private:
+    // Creates a client socket and connects to the server.
+    void MakeConnection();
+
+    // Closes the socket.
+    void CloseConnection() {
+      GTEST_CHECK_(sockfd_ != -1)
+          << "CloseConnection() can be called only when there is a connection.";
+
+      close(sockfd_);
+      sockfd_ = -1;
+    }
+
+    int sockfd_;  // socket file descriptor
+    const string host_name_;
+    const string port_num_;
+
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(SocketWriter);
+  };  // class SocketWriter
+
+  // Escapes '=', '&', '%', and '\n' characters in str as "%xx".
+  static string UrlEncode(const char* str);
+
+  StreamingListener(const string& host, const string& port)
+      : socket_writer_(new SocketWriter(host, port)) { Start(); }
+
+  explicit StreamingListener(AbstractSocketWriter* socket_writer)
+      : socket_writer_(socket_writer) { Start(); }
+
+  void OnTestProgramStart(const UnitTest& /* unit_test */) {
+    SendLn("event=TestProgramStart");
+  }
+
+  void OnTestProgramEnd(const UnitTest& unit_test) {
+    // Note that Google Test current only report elapsed time for each
+    // test iteration, not for the entire test program.
+    SendLn("event=TestProgramEnd&passed=" + FormatBool(unit_test.Passed()));
+
+    // Notify the streaming server to stop.
+    socket_writer_->CloseConnection();
+  }
+
+  void OnTestIterationStart(const UnitTest& /* unit_test */, int iteration) {
+    SendLn("event=TestIterationStart&iteration=" +
+           StreamableToString(iteration));
+  }
+
+  void OnTestIterationEnd(const UnitTest& unit_test, int /* iteration */) {
+    SendLn("event=TestIterationEnd&passed=" +
+           FormatBool(unit_test.Passed()) + "&elapsed_time=" +
+           StreamableToString(unit_test.elapsed_time()) + "ms");
+  }
+
+  void OnTestCaseStart(const TestCase& test_case) {
+    SendLn(std::string("event=TestCaseStart&name=") + test_case.name());
+  }
+
+  void OnTestCaseEnd(const TestCase& test_case) {
+    SendLn("event=TestCaseEnd&passed=" + FormatBool(test_case.Passed())
+           + "&elapsed_time=" + StreamableToString(test_case.elapsed_time())
+           + "ms");
+  }
+
+  void OnTestStart(const TestInfo& test_info) {
+    SendLn(std::string("event=TestStart&name=") + test_info.name());
+  }
+
+  void OnTestEnd(const TestInfo& test_info) {
+    SendLn("event=TestEnd&passed=" +
+           FormatBool((test_info.result())->Passed()) +
+           "&elapsed_time=" +
+           StreamableToString((test_info.result())->elapsed_time()) + "ms");
+  }
+
+  void OnTestPartResult(const TestPartResult& test_part_result) {
+    const char* file_name = test_part_result.file_name();
+    if (file_name == NULL)
+      file_name = "";
+    SendLn("event=TestPartResult&file=" + UrlEncode(file_name) +
+           "&line=" + StreamableToString(test_part_result.line_number()) +
+           "&message=" + UrlEncode(test_part_result.message()));
+  }
+
+ private:
+  // Sends the given message and a newline to the socket.
+  void SendLn(const string& message) { socket_writer_->SendLn(message); }
+
+  // Called at the start of streaming to notify the receiver what
+  // protocol we are using.
+  void Start() { SendLn("gtest_streaming_protocol_version=1.0"); }
+
+  string FormatBool(bool value) { return value ? "1" : "0"; }
+
+  const scoped_ptr<AbstractSocketWriter> socket_writer_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(StreamingListener);
+};  // class StreamingListener
+
+#endif  // GTEST_CAN_STREAM_RESULTS_
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_SRC_GTEST_INTERNAL_INL_H_
+#undef GTEST_IMPLEMENTATION_
+
+#if GTEST_OS_WINDOWS
+# define vsnprintf _vsnprintf
+#endif  // GTEST_OS_WINDOWS
+
+namespace testing {
+
+using internal::CountIf;
+using internal::ForEach;
+using internal::GetElementOr;
+using internal::Shuffle;
+
+// Constants.
+
+// A test whose test case name or test name matches this filter is
+// disabled and not run.
+static const char kDisableTestFilter[] = "DISABLED_*:*/DISABLED_*";
+
+// A test case whose name matches this filter is considered a death
+// test case and will be run before test cases whose name doesn't
+// match this filter.
+static const char kDeathTestCaseFilter[] = "*DeathTest:*DeathTest/*";
+
+// A test filter that matches everything.
+static const char kUniversalFilter[] = "*";
+
+// The default output file for XML output.
+static const char kDefaultOutputFile[] = "test_detail.xml";
+
+// The environment variable name for the test shard index.
+static const char kTestShardIndex[] = "GTEST_SHARD_INDEX";
+// The environment variable name for the total number of test shards.
+static const char kTestTotalShards[] = "GTEST_TOTAL_SHARDS";
+// The environment variable name for the test shard status file.
+static const char kTestShardStatusFile[] = "GTEST_SHARD_STATUS_FILE";
+
+namespace internal {
+
+// The text used in failure messages to indicate the start of the
+// stack trace.
+const char kStackTraceMarker[] = "\nStack trace:\n";
+
+// g_help_flag is true iff the --help flag or an equivalent form is
+// specified on the command line.
+bool g_help_flag = false;
+
+}  // namespace internal
+
+static const char* GetDefaultFilter() {
+  return kUniversalFilter;
+}
+
+GTEST_DEFINE_bool_(
+    also_run_disabled_tests,
+    internal::BoolFromGTestEnv("also_run_disabled_tests", false),
+    "Run disabled tests too, in addition to the tests normally being run.");
+
+GTEST_DEFINE_bool_(
+    break_on_failure,
+    internal::BoolFromGTestEnv("break_on_failure", false),
+    "True iff a failed assertion should be a debugger break-point.");
+
+GTEST_DEFINE_bool_(
+    catch_exceptions,
+    internal::BoolFromGTestEnv("catch_exceptions", true),
+    "True iff " GTEST_NAME_
+    " should catch exceptions and treat them as test failures.");
+
+GTEST_DEFINE_string_(
+    color,
+    internal::StringFromGTestEnv("color", "auto"),
+    "Whether to use colors in the output.  Valid values: yes, no, "
+    "and auto.  'auto' means to use colors if the output is "
+    "being sent to a terminal and the TERM environment variable "
+    "is set to a terminal type that supports colors.");
+
+GTEST_DEFINE_string_(
+    filter,
+    internal::StringFromGTestEnv("filter", GetDefaultFilter()),
+    "A colon-separated list of glob (not regex) patterns "
+    "for filtering the tests to run, optionally followed by a "
+    "'-' and a : separated list of negative patterns (tests to "
+    "exclude).  A test is run if it matches one of the positive "
+    "patterns and does not match any of the negative patterns.");
+
+GTEST_DEFINE_bool_(list_tests, false,
+                   "List all tests without running them.");
+
+GTEST_DEFINE_string_(
+    output,
+    internal::StringFromGTestEnv("output", ""),
+    "A format (currently must be \"xml\"), optionally followed "
+    "by a colon and an output file name or directory. A directory "
+    "is indicated by a trailing pathname separator. "
+    "Examples: \"xml:filename.xml\", \"xml::directoryname/\". "
+    "If a directory is specified, output files will be created "
+    "within that directory, with file-names based on the test "
+    "executable's name and, if necessary, made unique by adding "
+    "digits.");
+
+GTEST_DEFINE_bool_(
+    print_time,
+    internal::BoolFromGTestEnv("print_time", true),
+    "True iff " GTEST_NAME_
+    " should display elapsed time in text output.");
+
+GTEST_DEFINE_int32_(
+    random_seed,
+    internal::Int32FromGTestEnv("random_seed", 0),
+    "Random number seed to use when shuffling test orders.  Must be in range "
+    "[1, 99999], or 0 to use a seed based on the current time.");
+
+GTEST_DEFINE_int32_(
+    repeat,
+    internal::Int32FromGTestEnv("repeat", 1),
+    "How many times to repeat each test.  Specify a negative number "
+    "for repeating forever.  Useful for shaking out flaky tests.");
+
+GTEST_DEFINE_bool_(
+    show_internal_stack_frames, false,
+    "True iff " GTEST_NAME_ " should include internal stack frames when "
+    "printing test failure stack traces.");
+
+GTEST_DEFINE_bool_(
+    shuffle,
+    internal::BoolFromGTestEnv("shuffle", false),
+    "True iff " GTEST_NAME_
+    " should randomize tests' order on every run.");
+
+GTEST_DEFINE_int32_(
+    stack_trace_depth,
+    internal::Int32FromGTestEnv("stack_trace_depth", kMaxStackTraceDepth),
+    "The maximum number of stack frames to print when an "
+    "assertion fails.  The valid range is 0 through 100, inclusive.");
+
+GTEST_DEFINE_string_(
+    stream_result_to,
+    internal::StringFromGTestEnv("stream_result_to", ""),
+    "This flag specifies the host name and the port number on which to stream "
+    "test results. Example: \"localhost:555\". The flag is effective only on "
+    "Linux.");
+
+GTEST_DEFINE_bool_(
+    throw_on_failure,
+    internal::BoolFromGTestEnv("throw_on_failure", false),
+    "When this flag is specified, a failed assertion will throw an exception "
+    "if exceptions are enabled or exit the program with a non-zero code "
+    "otherwise.");
+
+namespace internal {
+
+// Generates a random number from [0, range), using a Linear
+// Congruential Generator (LCG).  Crashes if 'range' is 0 or greater
+// than kMaxRange.
+UInt32 Random::Generate(UInt32 range) {
+  // These constants are the same as are used in glibc's rand(3).
+  state_ = (1103515245U*state_ + 12345U) % kMaxRange;
+
+  GTEST_CHECK_(range > 0)
+      << "Cannot generate a number in the range [0, 0).";
+  GTEST_CHECK_(range <= kMaxRange)
+      << "Generation of a number in [0, " << range << ") was requested, "
+      << "but this can only generate numbers in [0, " << kMaxRange << ").";
+
+  // Converting via modulus introduces a bit of downward bias, but
+  // it's simple, and a linear congruential generator isn't too good
+  // to begin with.
+  return state_ % range;
+}
+
+// GTestIsInitialized() returns true iff the user has initialized
+// Google Test.  Useful for catching the user mistake of not initializing
+// Google Test before calling RUN_ALL_TESTS().
+//
+// A user must call testing::InitGoogleTest() to initialize Google
+// Test.  g_init_gtest_count is set to the number of times
+// InitGoogleTest() has been called.  We don't protect this variable
+// under a mutex as it is only accessed in the main thread.
+GTEST_API_ int g_init_gtest_count = 0;
+static bool GTestIsInitialized() { return g_init_gtest_count != 0; }
+
+// Iterates over a vector of TestCases, keeping a running sum of the
+// results of calling a given int-returning method on each.
+// Returns the sum.
+static int SumOverTestCaseList(const std::vector<TestCase*>& case_list,
+                               int (TestCase::*method)() const) {
+  int sum = 0;
+  for (size_t i = 0; i < case_list.size(); i++) {
+    sum += (case_list[i]->*method)();
+  }
+  return sum;
+}
+
+// Returns true iff the test case passed.
+static bool TestCasePassed(const TestCase* test_case) {
+  return test_case->should_run() && test_case->Passed();
+}
+
+// Returns true iff the test case failed.
+static bool TestCaseFailed(const TestCase* test_case) {
+  return test_case->should_run() && test_case->Failed();
+}
+
+// Returns true iff test_case contains at least one test that should
+// run.
+static bool ShouldRunTestCase(const TestCase* test_case) {
+  return test_case->should_run();
+}
+
+// AssertHelper constructor.
+AssertHelper::AssertHelper(TestPartResult::Type type,
+                           const char* file,
+                           int line,
+                           const char* message)
+    : data_(new AssertHelperData(type, file, line, message)) {
+}
+
+AssertHelper::~AssertHelper() {
+  delete data_;
+}
+
+// Message assignment, for assertion streaming support.
+void AssertHelper::operator=(const Message& message) const {
+  UnitTest::GetInstance()->
+    AddTestPartResult(data_->type, data_->file, data_->line,
+                      AppendUserMessage(data_->message, message),
+                      UnitTest::GetInstance()->impl()
+                      ->CurrentOsStackTraceExceptTop(1)
+                      // Skips the stack frame for this function itself.
+                      );  // NOLINT
+}
+
+// Mutex for linked pointers.
+GTEST_API_ GTEST_DEFINE_STATIC_MUTEX_(g_linked_ptr_mutex);
+
+// Application pathname gotten in InitGoogleTest.
+std::string g_executable_path;
+
+// Returns the current application's name, removing directory path if that
+// is present.
+FilePath GetCurrentExecutableName() {
+  FilePath result;
+
+#if GTEST_OS_WINDOWS
+  result.Set(FilePath(g_executable_path).RemoveExtension("exe"));
+#else
+  result.Set(FilePath(g_executable_path));
+#endif  // GTEST_OS_WINDOWS
+
+  return result.RemoveDirectoryName();
+}
+
+// Functions for processing the gtest_output flag.
+
+// Returns the output format, or "" for normal printed output.
+std::string UnitTestOptions::GetOutputFormat() {
+  const char* const gtest_output_flag = GTEST_FLAG(output).c_str();
+  if (gtest_output_flag == NULL) return std::string("");
+
+  const char* const colon = strchr(gtest_output_flag, ':');
+  return (colon == NULL) ?
+      std::string(gtest_output_flag) :
+      std::string(gtest_output_flag, colon - gtest_output_flag);
+}
+
+// Returns the name of the requested output file, or the default if none
+// was explicitly specified.
+std::string UnitTestOptions::GetAbsolutePathToOutputFile() {
+  const char* const gtest_output_flag = GTEST_FLAG(output).c_str();
+  if (gtest_output_flag == NULL)
+    return "";
+
+  const char* const colon = strchr(gtest_output_flag, ':');
+  if (colon == NULL)
+    return internal::FilePath::ConcatPaths(
+        internal::FilePath(
+            UnitTest::GetInstance()->original_working_dir()),
+        internal::FilePath(kDefaultOutputFile)).string();
+
+  internal::FilePath output_name(colon + 1);
+  if (!output_name.IsAbsolutePath())
+    // TODO(wan@google.com): on Windows \some\path is not an absolute
+    // path (as its meaning depends on the current drive), yet the
+    // following logic for turning it into an absolute path is wrong.
+    // Fix it.
+    output_name = internal::FilePath::ConcatPaths(
+        internal::FilePath(UnitTest::GetInstance()->original_working_dir()),
+        internal::FilePath(colon + 1));
+
+  if (!output_name.IsDirectory())
+    return output_name.string();
+
+  internal::FilePath result(internal::FilePath::GenerateUniqueFileName(
+      output_name, internal::GetCurrentExecutableName(),
+      GetOutputFormat().c_str()));
+  return result.string();
+}
+
+// Returns true iff the wildcard pattern matches the string.  The
+// first ':' or '\0' character in pattern marks the end of it.
+//
+// This recursive algorithm isn't very efficient, but is clear and
+// works well enough for matching test names, which are short.
+bool UnitTestOptions::PatternMatchesString(const char *pattern,
+                                           const char *str) {
+  switch (*pattern) {
+    case '\0':
+    case ':':  // Either ':' or '\0' marks the end of the pattern.
+      return *str == '\0';
+    case '?':  // Matches any single character.
+      return *str != '\0' && PatternMatchesString(pattern + 1, str + 1);
+    case '*':  // Matches any string (possibly empty) of characters.
+      return (*str != '\0' && PatternMatchesString(pattern, str + 1)) ||
+          PatternMatchesString(pattern + 1, str);
+    default:  // Non-special character.  Matches itself.
+      return *pattern == *str &&
+          PatternMatchesString(pattern + 1, str + 1);
+  }
+}
+
+bool UnitTestOptions::MatchesFilter(
+    const std::string& name, const char* filter) {
+  const char *cur_pattern = filter;
+  for (;;) {
+    if (PatternMatchesString(cur_pattern, name.c_str())) {
+      return true;
+    }
+
+    // Finds the next pattern in the filter.
+    cur_pattern = strchr(cur_pattern, ':');
+
+    // Returns if no more pattern can be found.
+    if (cur_pattern == NULL) {
+      return false;
+    }
+
+    // Skips the pattern separater (the ':' character).
+    cur_pattern++;
+  }
+}
+
+// Returns true iff the user-specified filter matches the test case
+// name and the test name.
+bool UnitTestOptions::FilterMatchesTest(const std::string &test_case_name,
+                                        const std::string &test_name) {
+  const std::string& full_name = test_case_name + "." + test_name.c_str();
+
+  // Split --gtest_filter at '-', if there is one, to separate into
+  // positive filter and negative filter portions
+  const char* const p = GTEST_FLAG(filter).c_str();
+  const char* const dash = strchr(p, '-');
+  std::string positive;
+  std::string negative;
+  if (dash == NULL) {
+    positive = GTEST_FLAG(filter).c_str();  // Whole string is a positive filter
+    negative = "";
+  } else {
+    positive = std::string(p, dash);   // Everything up to the dash
+    negative = std::string(dash + 1);  // Everything after the dash
+    if (positive.empty()) {
+      // Treat '-test1' as the same as '*-test1'
+      positive = kUniversalFilter;
+    }
+  }
+
+  // A filter is a colon-separated list of patterns.  It matches a
+  // test if any pattern in it matches the test.
+  return (MatchesFilter(full_name, positive.c_str()) &&
+          !MatchesFilter(full_name, negative.c_str()));
+}
+
+#if GTEST_HAS_SEH
+// Returns EXCEPTION_EXECUTE_HANDLER if Google Test should handle the
+// given SEH exception, or EXCEPTION_CONTINUE_SEARCH otherwise.
+// This function is useful as an __except condition.
+int UnitTestOptions::GTestShouldProcessSEH(DWORD exception_code) {
+  // Google Test should handle a SEH exception if:
+  //   1. the user wants it to, AND
+  //   2. this is not a breakpoint exception, AND
+  //   3. this is not a C++ exception (VC++ implements them via SEH,
+  //      apparently).
+  //
+  // SEH exception code for C++ exceptions.
+  // (see http://support.microsoft.com/kb/185294 for more information).
+  const DWORD kCxxExceptionCode = 0xe06d7363;
+
+  bool should_handle = true;
+
+  if (!GTEST_FLAG(catch_exceptions))
+    should_handle = false;
+  else if (exception_code == EXCEPTION_BREAKPOINT)
+    should_handle = false;
+  else if (exception_code == kCxxExceptionCode)
+    should_handle = false;
+
+  return should_handle ? EXCEPTION_EXECUTE_HANDLER : EXCEPTION_CONTINUE_SEARCH;
+}
+#endif  // GTEST_HAS_SEH
+
+}  // namespace internal
+
+// The c'tor sets this object as the test part result reporter used by
+// Google Test.  The 'result' parameter specifies where to report the
+// results. Intercepts only failures from the current thread.
+ScopedFakeTestPartResultReporter::ScopedFakeTestPartResultReporter(
+    TestPartResultArray* result)
+    : intercept_mode_(INTERCEPT_ONLY_CURRENT_THREAD),
+      result_(result) {
+  Init();
+}
+
+// The c'tor sets this object as the test part result reporter used by
+// Google Test.  The 'result' parameter specifies where to report the
+// results.
+ScopedFakeTestPartResultReporter::ScopedFakeTestPartResultReporter(
+    InterceptMode intercept_mode, TestPartResultArray* result)
+    : intercept_mode_(intercept_mode),
+      result_(result) {
+  Init();
+}
+
+void ScopedFakeTestPartResultReporter::Init() {
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  if (intercept_mode_ == INTERCEPT_ALL_THREADS) {
+    old_reporter_ = impl->GetGlobalTestPartResultReporter();
+    impl->SetGlobalTestPartResultReporter(this);
+  } else {
+    old_reporter_ = impl->GetTestPartResultReporterForCurrentThread();
+    impl->SetTestPartResultReporterForCurrentThread(this);
+  }
+}
+
+// The d'tor restores the test part result reporter used by Google Test
+// before.
+ScopedFakeTestPartResultReporter::~ScopedFakeTestPartResultReporter() {
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  if (intercept_mode_ == INTERCEPT_ALL_THREADS) {
+    impl->SetGlobalTestPartResultReporter(old_reporter_);
+  } else {
+    impl->SetTestPartResultReporterForCurrentThread(old_reporter_);
+  }
+}
+
+// Increments the test part result count and remembers the result.
+// This method is from the TestPartResultReporterInterface interface.
+void ScopedFakeTestPartResultReporter::ReportTestPartResult(
+    const TestPartResult& result) {
+  result_->Append(result);
+}
+
+namespace internal {
+
+// Returns the type ID of ::testing::Test.  We should always call this
+// instead of GetTypeId< ::testing::Test>() to get the type ID of
+// testing::Test.  This is to work around a suspected linker bug when
+// using Google Test as a framework on Mac OS X.  The bug causes
+// GetTypeId< ::testing::Test>() to return different values depending
+// on whether the call is from the Google Test framework itself or
+// from user test code.  GetTestTypeId() is guaranteed to always
+// return the same value, as it always calls GetTypeId<>() from the
+// gtest.cc, which is within the Google Test framework.
+TypeId GetTestTypeId() {
+  return GetTypeId<Test>();
+}
+
+// The value of GetTestTypeId() as seen from within the Google Test
+// library.  This is solely for testing GetTestTypeId().
+extern const TypeId kTestTypeIdInGoogleTest = GetTestTypeId();
+
+// This predicate-formatter checks that 'results' contains a test part
+// failure of the given type and that the failure message contains the
+// given substring.
+AssertionResult HasOneFailure(const char* /* results_expr */,
+                              const char* /* type_expr */,
+                              const char* /* substr_expr */,
+                              const TestPartResultArray& results,
+                              TestPartResult::Type type,
+                              const string& substr) {
+  const std::string expected(type == TestPartResult::kFatalFailure ?
+                        "1 fatal failure" :
+                        "1 non-fatal failure");
+  Message msg;
+  if (results.size() != 1) {
+    msg << "Expected: " << expected << "\n"
+        << "  Actual: " << results.size() << " failures";
+    for (int i = 0; i < results.size(); i++) {
+      msg << "\n" << results.GetTestPartResult(i);
+    }
+    return AssertionFailure() << msg;
+  }
+
+  const TestPartResult& r = results.GetTestPartResult(0);
+  if (r.type() != type) {
+    return AssertionFailure() << "Expected: " << expected << "\n"
+                              << "  Actual:\n"
+                              << r;
+  }
+
+  if (strstr(r.message(), substr.c_str()) == NULL) {
+    return AssertionFailure() << "Expected: " << expected << " containing \""
+                              << substr << "\"\n"
+                              << "  Actual:\n"
+                              << r;
+  }
+
+  return AssertionSuccess();
+}
+
+// The constructor of SingleFailureChecker remembers where to look up
+// test part results, what type of failure we expect, and what
+// substring the failure message should contain.
+SingleFailureChecker:: SingleFailureChecker(
+    const TestPartResultArray* results,
+    TestPartResult::Type type,
+    const string& substr)
+    : results_(results),
+      type_(type),
+      substr_(substr) {}
+
+// The destructor of SingleFailureChecker verifies that the given
+// TestPartResultArray contains exactly one failure that has the given
+// type and contains the given substring.  If that's not the case, a
+// non-fatal failure will be generated.
+SingleFailureChecker::~SingleFailureChecker() {
+  EXPECT_PRED_FORMAT3(HasOneFailure, *results_, type_, substr_);
+}
+
+DefaultGlobalTestPartResultReporter::DefaultGlobalTestPartResultReporter(
+    UnitTestImpl* unit_test) : unit_test_(unit_test) {}
+
+void DefaultGlobalTestPartResultReporter::ReportTestPartResult(
+    const TestPartResult& result) {
+  unit_test_->current_test_result()->AddTestPartResult(result);
+  unit_test_->listeners()->repeater()->OnTestPartResult(result);
+}
+
+DefaultPerThreadTestPartResultReporter::DefaultPerThreadTestPartResultReporter(
+    UnitTestImpl* unit_test) : unit_test_(unit_test) {}
+
+void DefaultPerThreadTestPartResultReporter::ReportTestPartResult(
+    const TestPartResult& result) {
+  unit_test_->GetGlobalTestPartResultReporter()->ReportTestPartResult(result);
+}
+
+// Returns the global test part result reporter.
+TestPartResultReporterInterface*
+UnitTestImpl::GetGlobalTestPartResultReporter() {
+  internal::MutexLock lock(&global_test_part_result_reporter_mutex_);
+  return global_test_part_result_repoter_;
+}
+
+// Sets the global test part result reporter.
+void UnitTestImpl::SetGlobalTestPartResultReporter(
+    TestPartResultReporterInterface* reporter) {
+  internal::MutexLock lock(&global_test_part_result_reporter_mutex_);
+  global_test_part_result_repoter_ = reporter;
+}
+
+// Returns the test part result reporter for the current thread.
+TestPartResultReporterInterface*
+UnitTestImpl::GetTestPartResultReporterForCurrentThread() {
+  return per_thread_test_part_result_reporter_.get();
+}
+
+// Sets the test part result reporter for the current thread.
+void UnitTestImpl::SetTestPartResultReporterForCurrentThread(
+    TestPartResultReporterInterface* reporter) {
+  per_thread_test_part_result_reporter_.set(reporter);
+}
+
+// Gets the number of successful test cases.
+int UnitTestImpl::successful_test_case_count() const {
+  return CountIf(test_cases_, TestCasePassed);
+}
+
+// Gets the number of failed test cases.
+int UnitTestImpl::failed_test_case_count() const {
+  return CountIf(test_cases_, TestCaseFailed);
+}
+
+// Gets the number of all test cases.
+int UnitTestImpl::total_test_case_count() const {
+  return static_cast<int>(test_cases_.size());
+}
+
+// Gets the number of all test cases that contain at least one test
+// that should run.
+int UnitTestImpl::test_case_to_run_count() const {
+  return CountIf(test_cases_, ShouldRunTestCase);
+}
+
+// Gets the number of successful tests.
+int UnitTestImpl::successful_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::successful_test_count);
+}
+
+// Gets the number of failed tests.
+int UnitTestImpl::failed_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::failed_test_count);
+}
+
+// Gets the number of disabled tests that will be reported in the XML report.
+int UnitTestImpl::reportable_disabled_test_count() const {
+  return SumOverTestCaseList(test_cases_,
+                             &TestCase::reportable_disabled_test_count);
+}
+
+// Gets the number of disabled tests.
+int UnitTestImpl::disabled_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::disabled_test_count);
+}
+
+// Gets the number of tests to be printed in the XML report.
+int UnitTestImpl::reportable_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::reportable_test_count);
+}
+
+// Gets the number of all tests.
+int UnitTestImpl::total_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::total_test_count);
+}
+
+// Gets the number of tests that should run.
+int UnitTestImpl::test_to_run_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::test_to_run_count);
+}
+
+// Returns the current OS stack trace as an std::string.
+//
+// The maximum number of stack frames to be included is specified by
+// the gtest_stack_trace_depth flag.  The skip_count parameter
+// specifies the number of top frames to be skipped, which doesn't
+// count against the number of frames to be included.
+//
+// For example, if Foo() calls Bar(), which in turn calls
+// CurrentOsStackTraceExceptTop(1), Foo() will be included in the
+// trace but Bar() and CurrentOsStackTraceExceptTop() won't.
+std::string UnitTestImpl::CurrentOsStackTraceExceptTop(int skip_count) {
+  (void)skip_count;
+  return "";
+}
+
+// Returns the current time in milliseconds.
+TimeInMillis GetTimeInMillis() {
+#if GTEST_OS_WINDOWS_MOBILE || defined(__BORLANDC__)
+  // Difference between 1970-01-01 and 1601-01-01 in milliseconds.
+  // http://analogous.blogspot.com/2005/04/epoch.html
+  const TimeInMillis kJavaEpochToWinFileTimeDelta =
+    static_cast<TimeInMillis>(116444736UL) * 100000UL;
+  const DWORD kTenthMicrosInMilliSecond = 10000;
+
+  SYSTEMTIME now_systime;
+  FILETIME now_filetime;
+  ULARGE_INTEGER now_int64;
+  // TODO(kenton@google.com): Shouldn't this just use
+  //   GetSystemTimeAsFileTime()?
+  GetSystemTime(&now_systime);
+  if (SystemTimeToFileTime(&now_systime, &now_filetime)) {
+    now_int64.LowPart = now_filetime.dwLowDateTime;
+    now_int64.HighPart = now_filetime.dwHighDateTime;
+    now_int64.QuadPart = (now_int64.QuadPart / kTenthMicrosInMilliSecond) -
+      kJavaEpochToWinFileTimeDelta;
+    return now_int64.QuadPart;
+  }
+  return 0;
+#elif GTEST_OS_WINDOWS && !GTEST_HAS_GETTIMEOFDAY_
+  __timeb64 now;
+
+# ifdef _MSC_VER
+
+  // MSVC 8 deprecates _ftime64(), so we want to suppress warning 4996
+  // (deprecated function) there.
+  // TODO(kenton@google.com): Use GetTickCount()?  Or use
+  //   SystemTimeToFileTime()
+#  pragma warning(push)          // Saves the current warning state.
+#  pragma warning(disable:4996)  // Temporarily disables warning 4996.
+  _ftime64(&now);
+#  pragma warning(pop)           // Restores the warning state.
+# else
+
+  _ftime64(&now);
+
+# endif  // _MSC_VER
+
+  return static_cast<TimeInMillis>(now.time) * 1000 + now.millitm;
+#elif GTEST_HAS_GETTIMEOFDAY_
+  struct timeval now;
+  gettimeofday(&now, NULL);
+  return static_cast<TimeInMillis>(now.tv_sec) * 1000 + now.tv_usec / 1000;
+#else
+# error "Don't know how to get the current time on your system."
+#endif
+}
+
+// Utilities
+
+// class String.
+
+#if GTEST_OS_WINDOWS_MOBILE
+// Creates a UTF-16 wide string from the given ANSI string, allocating
+// memory using new. The caller is responsible for deleting the return
+// value using delete[]. Returns the wide string, or NULL if the
+// input is NULL.
+LPCWSTR String::AnsiToUtf16(const char* ansi) {
+  if (!ansi) return NULL;
+  const int length = strlen(ansi);
+  const int unicode_length =
+      MultiByteToWideChar(CP_ACP, 0, ansi, length,
+                          NULL, 0);
+  WCHAR* unicode = new WCHAR[unicode_length + 1];
+  MultiByteToWideChar(CP_ACP, 0, ansi, length,
+                      unicode, unicode_length);
+  unicode[unicode_length] = 0;
+  return unicode;
+}
+
+// Creates an ANSI string from the given wide string, allocating
+// memory using new. The caller is responsible for deleting the return
+// value using delete[]. Returns the ANSI string, or NULL if the
+// input is NULL.
+const char* String::Utf16ToAnsi(LPCWSTR utf16_str)  {
+  if (!utf16_str) return NULL;
+  const int ansi_length =
+      WideCharToMultiByte(CP_ACP, 0, utf16_str, -1,
+                          NULL, 0, NULL, NULL);
+  char* ansi = new char[ansi_length + 1];
+  WideCharToMultiByte(CP_ACP, 0, utf16_str, -1,
+                      ansi, ansi_length, NULL, NULL);
+  ansi[ansi_length] = 0;
+  return ansi;
+}
+
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+// Compares two C strings.  Returns true iff they have the same content.
+//
+// Unlike strcmp(), this function can handle NULL argument(s).  A NULL
+// C string is considered different to any non-NULL C string,
+// including the empty string.
+bool String::CStringEquals(const char * lhs, const char * rhs) {
+  if ( lhs == NULL ) return rhs == NULL;
+
+  if ( rhs == NULL ) return false;
+
+  return strcmp(lhs, rhs) == 0;
+}
+
+#if GTEST_HAS_STD_WSTRING || GTEST_HAS_GLOBAL_WSTRING
+
+// Converts an array of wide chars to a narrow string using the UTF-8
+// encoding, and streams the result to the given Message object.
+static void StreamWideCharsToMessage(const wchar_t* wstr, size_t length,
+                                     Message* msg) {
+  for (size_t i = 0; i != length; ) {  // NOLINT
+    if (wstr[i] != L'\0') {
+      *msg << WideStringToUtf8(wstr + i, static_cast<int>(length - i));
+      while (i != length && wstr[i] != L'\0')
+        i++;
+    } else {
+      *msg << '\0';
+      i++;
+    }
+  }
+}
+
+#endif  // GTEST_HAS_STD_WSTRING || GTEST_HAS_GLOBAL_WSTRING
+
+}  // namespace internal
+
+// Constructs an empty Message.
+// We allocate the stringstream separately because otherwise each use of
+// ASSERT/EXPECT in a procedure adds over 200 bytes to the procedure's
+// stack frame leading to huge stack frames in some cases; gcc does not reuse
+// the stack space.
+Message::Message() : ss_(new ::std::stringstream) {
+  // By default, we want there to be enough precision when printing
+  // a double to a Message.
+  *ss_ << std::setprecision(std::numeric_limits<double>::digits10 + 2);
+}
+
+// These two overloads allow streaming a wide C string to a Message
+// using the UTF-8 encoding.
+Message& Message::operator <<(const wchar_t* wide_c_str) {
+  return *this << internal::String::ShowWideCString(wide_c_str);
+}
+Message& Message::operator <<(wchar_t* wide_c_str) {
+  return *this << internal::String::ShowWideCString(wide_c_str);
+}
+
+#if GTEST_HAS_STD_WSTRING
+// Converts the given wide string to a narrow string using the UTF-8
+// encoding, and streams the result to this Message object.
+Message& Message::operator <<(const ::std::wstring& wstr) {
+  internal::StreamWideCharsToMessage(wstr.c_str(), wstr.length(), this);
+  return *this;
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+#if GTEST_HAS_GLOBAL_WSTRING
+// Converts the given wide string to a narrow string using the UTF-8
+// encoding, and streams the result to this Message object.
+Message& Message::operator <<(const ::wstring& wstr) {
+  internal::StreamWideCharsToMessage(wstr.c_str(), wstr.length(), this);
+  return *this;
+}
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+// Gets the text streamed to this object so far as an std::string.
+// Each '\0' character in the buffer is replaced with "\\0".
+std::string Message::GetString() const {
+  return internal::StringStreamToString(ss_.get());
+}
+
+// AssertionResult constructors.
+// Used in EXPECT_TRUE/FALSE(assertion_result).
+AssertionResult::AssertionResult(const AssertionResult& other)
+    : success_(other.success_),
+      message_(other.message_.get() != NULL ?
+               new ::std::string(*other.message_) :
+               static_cast< ::std::string*>(NULL)) {
+}
+
+// Returns the assertion's negation. Used with EXPECT/ASSERT_FALSE.
+AssertionResult AssertionResult::operator!() const {
+  AssertionResult negation(!success_);
+  if (message_.get() != NULL)
+    negation << *message_;
+  return negation;
+}
+
+// Makes a successful assertion result.
+AssertionResult AssertionSuccess() {
+  return AssertionResult(true);
+}
+
+// Makes a failed assertion result.
+AssertionResult AssertionFailure() {
+  return AssertionResult(false);
+}
+
+// Makes a failed assertion result with the given failure message.
+// Deprecated; use AssertionFailure() << message.
+AssertionResult AssertionFailure(const Message& message) {
+  return AssertionFailure() << message;
+}
+
+namespace internal {
+
+// Constructs and returns the message for an equality assertion
+// (e.g. ASSERT_EQ, EXPECT_STREQ, etc) failure.
+//
+// The first four parameters are the expressions used in the assertion
+// and their values, as strings.  For example, for ASSERT_EQ(foo, bar)
+// where foo is 5 and bar is 6, we have:
+//
+//   expected_expression: "foo"
+//   actual_expression:   "bar"
+//   expected_value:      "5"
+//   actual_value:        "6"
+//
+// The ignoring_case parameter is true iff the assertion is a
+// *_STRCASEEQ*.  When it's true, the string " (ignoring case)" will
+// be inserted into the message.
+AssertionResult EqFailure(const char* expected_expression,
+                          const char* actual_expression,
+                          const std::string& expected_value,
+                          const std::string& actual_value,
+                          bool ignoring_case) {
+  Message msg;
+  msg << "Value of: " << actual_expression;
+  if (actual_value != actual_expression) {
+    msg << "\n  Actual: " << actual_value;
+  }
+
+  msg << "\nExpected: " << expected_expression;
+  if (ignoring_case) {
+    msg << " (ignoring case)";
+  }
+  if (expected_value != expected_expression) {
+    msg << "\nWhich is: " << expected_value;
+  }
+
+  return AssertionFailure() << msg;
+}
+
+// Constructs a failure message for Boolean assertions such as EXPECT_TRUE.
+std::string GetBoolAssertionFailureMessage(
+    const AssertionResult& assertion_result,
+    const char* expression_text,
+    const char* actual_predicate_value,
+    const char* expected_predicate_value) {
+  const char* actual_message = assertion_result.message();
+  Message msg;
+  msg << "Value of: " << expression_text
+      << "\n  Actual: " << actual_predicate_value;
+  if (actual_message[0] != '\0')
+    msg << " (" << actual_message << ")";
+  msg << "\nExpected: " << expected_predicate_value;
+  return msg.GetString();
+}
+
+// Helper function for implementing ASSERT_NEAR.
+AssertionResult DoubleNearPredFormat(const char* expr1,
+                                     const char* expr2,
+                                     const char* abs_error_expr,
+                                     double val1,
+                                     double val2,
+                                     double abs_error) {
+  const double diff = fabs(val1 - val2);
+  if (diff <= abs_error) return AssertionSuccess();
+
+  // TODO(wan): do not print the value of an expression if it's
+  // already a literal.
+  return AssertionFailure()
+      << "The difference between " << expr1 << " and " << expr2
+      << " is " << diff << ", which exceeds " << abs_error_expr << ", where\n"
+      << expr1 << " evaluates to " << val1 << ",\n"
+      << expr2 << " evaluates to " << val2 << ", and\n"
+      << abs_error_expr << " evaluates to " << abs_error << ".";
+}
+
+
+// Helper template for implementing FloatLE() and DoubleLE().
+template <typename RawType>
+AssertionResult FloatingPointLE(const char* expr1,
+                                const char* expr2,
+                                RawType val1,
+                                RawType val2) {
+  // Returns success if val1 is less than val2,
+  if (val1 < val2) {
+    return AssertionSuccess();
+  }
+
+  // or if val1 is almost equal to val2.
+  const FloatingPoint<RawType> lhs(val1), rhs(val2);
+  if (lhs.AlmostEquals(rhs)) {
+    return AssertionSuccess();
+  }
+
+  // Note that the above two checks will both fail if either val1 or
+  // val2 is NaN, as the IEEE floating-point standard requires that
+  // any predicate involving a NaN must return false.
+
+  ::std::stringstream val1_ss;
+  val1_ss << std::setprecision(std::numeric_limits<RawType>::digits10 + 2)
+          << val1;
+
+  ::std::stringstream val2_ss;
+  val2_ss << std::setprecision(std::numeric_limits<RawType>::digits10 + 2)
+          << val2;
+
+  return AssertionFailure()
+      << "Expected: (" << expr1 << ") <= (" << expr2 << ")\n"
+      << "  Actual: " << StringStreamToString(&val1_ss) << " vs "
+      << StringStreamToString(&val2_ss);
+}
+
+}  // namespace internal
+
+// Asserts that val1 is less than, or almost equal to, val2.  Fails
+// otherwise.  In particular, it fails if either val1 or val2 is NaN.
+AssertionResult FloatLE(const char* expr1, const char* expr2,
+                        float val1, float val2) {
+  return internal::FloatingPointLE<float>(expr1, expr2, val1, val2);
+}
+
+// Asserts that val1 is less than, or almost equal to, val2.  Fails
+// otherwise.  In particular, it fails if either val1 or val2 is NaN.
+AssertionResult DoubleLE(const char* expr1, const char* expr2,
+                         double val1, double val2) {
+  return internal::FloatingPointLE<double>(expr1, expr2, val1, val2);
+}
+
+namespace internal {
+
+// The helper function for {ASSERT|EXPECT}_EQ with int or enum
+// arguments.
+AssertionResult CmpHelperEQ(const char* expected_expression,
+                            const char* actual_expression,
+                            BiggestInt expected,
+                            BiggestInt actual) {
+  if (expected == actual) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   FormatForComparisonFailureMessage(expected, actual),
+                   FormatForComparisonFailureMessage(actual, expected),
+                   false);
+}
+
+// A macro for implementing the helper functions needed to implement
+// ASSERT_?? and EXPECT_?? with integer or enum arguments.  It is here
+// just to avoid copy-and-paste of similar code.
+#define GTEST_IMPL_CMP_HELPER_(op_name, op)\
+AssertionResult CmpHelper##op_name(const char* expr1, const char* expr2, \
+                                   BiggestInt val1, BiggestInt val2) {\
+  if (val1 op val2) {\
+    return AssertionSuccess();\
+  } else {\
+    return AssertionFailure() \
+        << "Expected: (" << expr1 << ") " #op " (" << expr2\
+        << "), actual: " << FormatForComparisonFailureMessage(val1, val2)\
+        << " vs " << FormatForComparisonFailureMessage(val2, val1);\
+  }\
+}
+
+// Implements the helper function for {ASSERT|EXPECT}_NE with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(NE, !=)
+// Implements the helper function for {ASSERT|EXPECT}_LE with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(LE, <=)
+// Implements the helper function for {ASSERT|EXPECT}_LT with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(LT, < )
+// Implements the helper function for {ASSERT|EXPECT}_GE with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(GE, >=)
+// Implements the helper function for {ASSERT|EXPECT}_GT with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(GT, > )
+
+#undef GTEST_IMPL_CMP_HELPER_
+
+// The helper function for {ASSERT|EXPECT}_STREQ.
+AssertionResult CmpHelperSTREQ(const char* expected_expression,
+                               const char* actual_expression,
+                               const char* expected,
+                               const char* actual) {
+  if (String::CStringEquals(expected, actual)) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   PrintToString(expected),
+                   PrintToString(actual),
+                   false);
+}
+
+// The helper function for {ASSERT|EXPECT}_STRCASEEQ.
+AssertionResult CmpHelperSTRCASEEQ(const char* expected_expression,
+                                   const char* actual_expression,
+                                   const char* expected,
+                                   const char* actual) {
+  if (String::CaseInsensitiveCStringEquals(expected, actual)) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   PrintToString(expected),
+                   PrintToString(actual),
+                   true);
+}
+
+// The helper function for {ASSERT|EXPECT}_STRNE.
+AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                               const char* s2_expression,
+                               const char* s1,
+                               const char* s2) {
+  if (!String::CStringEquals(s1, s2)) {
+    return AssertionSuccess();
+  } else {
+    return AssertionFailure() << "Expected: (" << s1_expression << ") != ("
+                              << s2_expression << "), actual: \""
+                              << s1 << "\" vs \"" << s2 << "\"";
+  }
+}
+
+// The helper function for {ASSERT|EXPECT}_STRCASENE.
+AssertionResult CmpHelperSTRCASENE(const char* s1_expression,
+                                   const char* s2_expression,
+                                   const char* s1,
+                                   const char* s2) {
+  if (!String::CaseInsensitiveCStringEquals(s1, s2)) {
+    return AssertionSuccess();
+  } else {
+    return AssertionFailure()
+        << "Expected: (" << s1_expression << ") != ("
+        << s2_expression << ") (ignoring case), actual: \""
+        << s1 << "\" vs \"" << s2 << "\"";
+  }
+}
+
+}  // namespace internal
+
+namespace {
+
+// Helper functions for implementing IsSubString() and IsNotSubstring().
+
+// This group of overloaded functions return true iff needle is a
+// substring of haystack.  NULL is considered a substring of itself
+// only.
+
+bool IsSubstringPred(const char* needle, const char* haystack) {
+  if (needle == NULL || haystack == NULL)
+    return needle == haystack;
+
+  return strstr(haystack, needle) != NULL;
+}
+
+bool IsSubstringPred(const wchar_t* needle, const wchar_t* haystack) {
+  if (needle == NULL || haystack == NULL)
+    return needle == haystack;
+
+  return wcsstr(haystack, needle) != NULL;
+}
+
+// StringType here can be either ::std::string or ::std::wstring.
+template <typename StringType>
+bool IsSubstringPred(const StringType& needle,
+                     const StringType& haystack) {
+  return haystack.find(needle) != StringType::npos;
+}
+
+// This function implements either IsSubstring() or IsNotSubstring(),
+// depending on the value of the expected_to_be_substring parameter.
+// StringType here can be const char*, const wchar_t*, ::std::string,
+// or ::std::wstring.
+template <typename StringType>
+AssertionResult IsSubstringImpl(
+    bool expected_to_be_substring,
+    const char* needle_expr, const char* haystack_expr,
+    const StringType& needle, const StringType& haystack) {
+  if (IsSubstringPred(needle, haystack) == expected_to_be_substring)
+    return AssertionSuccess();
+
+  const bool is_wide_string = sizeof(needle[0]) > 1;
+  const char* const begin_string_quote = is_wide_string ? "L\"" : "\"";
+  return AssertionFailure()
+      << "Value of: " << needle_expr << "\n"
+      << "  Actual: " << begin_string_quote << needle << "\"\n"
+      << "Expected: " << (expected_to_be_substring ? "" : "not ")
+      << "a substring of " << haystack_expr << "\n"
+      << "Which is: " << begin_string_quote << haystack << "\"";
+}
+
+}  // namespace
+
+// IsSubstring() and IsNotSubstring() check whether needle is a
+// substring of haystack (NULL is considered a substring of itself
+// only), and return an appropriate error message when they fail.
+
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+
+#if GTEST_HAS_STD_WSTRING
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+namespace internal {
+
+#if GTEST_OS_WINDOWS
+
+namespace {
+
+// Helper function for IsHRESULT{SuccessFailure} predicates
+AssertionResult HRESULTFailureHelper(const char* expr,
+                                     const char* expected,
+                                     long hr) {  // NOLINT
+# if GTEST_OS_WINDOWS_MOBILE
+
+  // Windows CE doesn't support FormatMessage.
+  const char error_text[] = "";
+
+# else
+
+  // Looks up the human-readable system message for the HRESULT code
+  // and since we're not passing any params to FormatMessage, we don't
+  // want inserts expanded.
+  const DWORD kFlags = FORMAT_MESSAGE_FROM_SYSTEM |
+                       FORMAT_MESSAGE_IGNORE_INSERTS;
+  const DWORD kBufSize = 4096;
+  // Gets the system's human readable message string for this HRESULT.
+  char error_text[kBufSize] = { '\0' };
+  DWORD message_length = ::FormatMessageA(kFlags,
+                                          0,  // no source, we're asking system
+                                          hr,  // the error
+                                          0,  // no line width restrictions
+                                          error_text,  // output buffer
+                                          kBufSize,  // buf size
+                                          NULL);  // no arguments for inserts
+  // Trims tailing white space (FormatMessage leaves a trailing CR-LF)
+  for (; message_length && IsSpace(error_text[message_length - 1]);
+          --message_length) {
+    error_text[message_length - 1] = '\0';
+  }
+
+# endif  // GTEST_OS_WINDOWS_MOBILE
+
+  const std::string error_hex("0x" + String::FormatHexInt(hr));
+  return ::testing::AssertionFailure()
+      << "Expected: " << expr << " " << expected << ".\n"
+      << "  Actual: " << error_hex << " " << error_text << "\n";
+}
+
+}  // namespace
+
+AssertionResult IsHRESULTSuccess(const char* expr, long hr) {  // NOLINT
+  if (SUCCEEDED(hr)) {
+    return AssertionSuccess();
+  }
+  return HRESULTFailureHelper(expr, "succeeds", hr);
+}
+
+AssertionResult IsHRESULTFailure(const char* expr, long hr) {  // NOLINT
+  if (FAILED(hr)) {
+    return AssertionSuccess();
+  }
+  return HRESULTFailureHelper(expr, "fails", hr);
+}
+
+#endif  // GTEST_OS_WINDOWS
+
+// Utility functions for encoding Unicode text (wide strings) in
+// UTF-8.
+
+// A Unicode code-point can have upto 21 bits, and is encoded in UTF-8
+// like this:
+//
+// Code-point length   Encoding
+//   0 -  7 bits       0xxxxxxx
+//   8 - 11 bits       110xxxxx 10xxxxxx
+//  12 - 16 bits       1110xxxx 10xxxxxx 10xxxxxx
+//  17 - 21 bits       11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+
+// The maximum code-point a one-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint1 = (static_cast<UInt32>(1) <<  7) - 1;
+
+// The maximum code-point a two-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint2 = (static_cast<UInt32>(1) << (5 + 6)) - 1;
+
+// The maximum code-point a three-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint3 = (static_cast<UInt32>(1) << (4 + 2*6)) - 1;
+
+// The maximum code-point a four-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint4 = (static_cast<UInt32>(1) << (3 + 3*6)) - 1;
+
+// Chops off the n lowest bits from a bit pattern.  Returns the n
+// lowest bits.  As a side effect, the original bit pattern will be
+// shifted to the right by n bits.
+inline UInt32 ChopLowBits(UInt32* bits, int n) {
+  const UInt32 low_bits = *bits & ((static_cast<UInt32>(1) << n) - 1);
+  *bits >>= n;
+  return low_bits;
+}
+
+// Converts a Unicode code point to a narrow string in UTF-8 encoding.
+// code_point parameter is of type UInt32 because wchar_t may not be
+// wide enough to contain a code point.
+// If the code_point is not a valid Unicode code point
+// (i.e. outside of Unicode range U+0 to U+10FFFF) it will be converted
+// to "(Invalid Unicode 0xXXXXXXXX)".
+std::string CodePointToUtf8(UInt32 code_point) {
+  if (code_point > kMaxCodePoint4) {
+    return "(Invalid Unicode 0x" + String::FormatHexInt(code_point) + ")";
+  }
+
+  char str[5];  // Big enough for the largest valid code point.
+  if (code_point <= kMaxCodePoint1) {
+    str[1] = '\0';
+    str[0] = static_cast<char>(code_point);                          // 0xxxxxxx
+  } else if (code_point <= kMaxCodePoint2) {
+    str[2] = '\0';
+    str[1] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[0] = static_cast<char>(0xC0 | code_point);                   // 110xxxxx
+  } else if (code_point <= kMaxCodePoint3) {
+    str[3] = '\0';
+    str[2] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[1] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[0] = static_cast<char>(0xE0 | code_point);                   // 1110xxxx
+  } else {  // code_point <= kMaxCodePoint4
+    str[4] = '\0';
+    str[3] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[2] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[1] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[0] = static_cast<char>(0xF0 | code_point);                   // 11110xxx
+  }
+  return str;
+}
+
+// The following two functions only make sense if the the system
+// uses UTF-16 for wide string encoding. All supported systems
+// with 16 bit wchar_t (Windows, Cygwin, Symbian OS) do use UTF-16.
+
+// Determines if the arguments constitute UTF-16 surrogate pair
+// and thus should be combined into a single Unicode code point
+// using CreateCodePointFromUtf16SurrogatePair.
+inline bool IsUtf16SurrogatePair(wchar_t first, wchar_t second) {
+  return sizeof(wchar_t) == 2 &&
+      (first & 0xFC00) == 0xD800 && (second & 0xFC00) == 0xDC00;
+}
+
+// Creates a Unicode code point from UTF16 surrogate pair.
+inline UInt32 CreateCodePointFromUtf16SurrogatePair(wchar_t first,
+                                                    wchar_t second) {
+  const UInt32 mask = (1 << 10) - 1;
+  return (sizeof(wchar_t) == 2) ?
+      (((first & mask) << 10) | (second & mask)) + 0x10000 :
+      // This function should not be called when the condition is
+      // false, but we provide a sensible default in case it is.
+      static_cast<UInt32>(first);
+}
+
+// Converts a wide string to a narrow string in UTF-8 encoding.
+// The wide string is assumed to have the following encoding:
+//   UTF-16 if sizeof(wchar_t) == 2 (on Windows, Cygwin, Symbian OS)
+//   UTF-32 if sizeof(wchar_t) == 4 (on Linux)
+// Parameter str points to a null-terminated wide string.
+// Parameter num_chars may additionally limit the number
+// of wchar_t characters processed. -1 is used when the entire string
+// should be processed.
+// If the string contains code points that are not valid Unicode code points
+// (i.e. outside of Unicode range U+0 to U+10FFFF) they will be output
+// as '(Invalid Unicode 0xXXXXXXXX)'. If the string is in UTF16 encoding
+// and contains invalid UTF-16 surrogate pairs, values in those pairs
+// will be encoded as individual Unicode characters from Basic Normal Plane.
+std::string WideStringToUtf8(const wchar_t* str, int num_chars) {
+  if (num_chars == -1)
+    num_chars = static_cast<int>(wcslen(str));
+
+  ::std::stringstream stream;
+  for (int i = 0; i < num_chars; ++i) {
+    UInt32 unicode_code_point;
+
+    if (str[i] == L'\0') {
+      break;
+    } else if (i + 1 < num_chars && IsUtf16SurrogatePair(str[i], str[i + 1])) {
+      unicode_code_point = CreateCodePointFromUtf16SurrogatePair(str[i],
+                                                                 str[i + 1]);
+      i++;
+    } else {
+      unicode_code_point = static_cast<UInt32>(str[i]);
+    }
+
+    stream << CodePointToUtf8(unicode_code_point);
+  }
+  return StringStreamToString(&stream);
+}
+
+// Converts a wide C string to an std::string using the UTF-8 encoding.
+// NULL will be converted to "(null)".
+std::string String::ShowWideCString(const wchar_t * wide_c_str) {
+  if (wide_c_str == NULL)  return "(null)";
+
+  return internal::WideStringToUtf8(wide_c_str, -1);
+}
+
+// Compares two wide C strings.  Returns true iff they have the same
+// content.
+//
+// Unlike wcscmp(), this function can handle NULL argument(s).  A NULL
+// C string is considered different to any non-NULL C string,
+// including the empty string.
+bool String::WideCStringEquals(const wchar_t * lhs, const wchar_t * rhs) {
+  if (lhs == NULL) return rhs == NULL;
+
+  if (rhs == NULL) return false;
+
+  return wcscmp(lhs, rhs) == 0;
+}
+
+// Helper function for *_STREQ on wide strings.
+AssertionResult CmpHelperSTREQ(const char* expected_expression,
+                               const char* actual_expression,
+                               const wchar_t* expected,
+                               const wchar_t* actual) {
+  if (String::WideCStringEquals(expected, actual)) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(expected_expression,
+                   actual_expression,
+                   PrintToString(expected),
+                   PrintToString(actual),
+                   false);
+}
+
+// Helper function for *_STRNE on wide strings.
+AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                               const char* s2_expression,
+                               const wchar_t* s1,
+                               const wchar_t* s2) {
+  if (!String::WideCStringEquals(s1, s2)) {
+    return AssertionSuccess();
+  }
+
+  return AssertionFailure() << "Expected: (" << s1_expression << ") != ("
+                            << s2_expression << "), actual: "
+                            << PrintToString(s1)
+                            << " vs " << PrintToString(s2);
+}
+
+// Compares two C strings, ignoring case.  Returns true iff they have
+// the same content.
+//
+// Unlike strcasecmp(), this function can handle NULL argument(s).  A
+// NULL C string is considered different to any non-NULL C string,
+// including the empty string.
+bool String::CaseInsensitiveCStringEquals(const char * lhs, const char * rhs) {
+  if (lhs == NULL)
+    return rhs == NULL;
+  if (rhs == NULL)
+    return false;
+  return posix::StrCaseCmp(lhs, rhs) == 0;
+}
+
+  // Compares two wide C strings, ignoring case.  Returns true iff they
+  // have the same content.
+  //
+  // Unlike wcscasecmp(), this function can handle NULL argument(s).
+  // A NULL C string is considered different to any non-NULL wide C string,
+  // including the empty string.
+  // NB: The implementations on different platforms slightly differ.
+  // On windows, this method uses _wcsicmp which compares according to LC_CTYPE
+  // environment variable. On GNU platform this method uses wcscasecmp
+  // which compares according to LC_CTYPE category of the current locale.
+  // On MacOS X, it uses towlower, which also uses LC_CTYPE category of the
+  // current locale.
+bool String::CaseInsensitiveWideCStringEquals(const wchar_t* lhs,
+                                              const wchar_t* rhs) {
+  if (lhs == NULL) return rhs == NULL;
+
+  if (rhs == NULL) return false;
+
+#if GTEST_OS_WINDOWS
+  return _wcsicmp(lhs, rhs) == 0;
+#elif GTEST_OS_LINUX && !GTEST_OS_LINUX_ANDROID
+  return wcscasecmp(lhs, rhs) == 0;
+#else
+  // Android, Mac OS X and Cygwin don't define wcscasecmp.
+  // Other unknown OSes may not define it either.
+  wint_t left, right;
+  do {
+    left = towlower(*lhs++);
+    right = towlower(*rhs++);
+  } while (left && left == right);
+  return left == right;
+#endif  // OS selector
+}
+
+// Returns true iff str ends with the given suffix, ignoring case.
+// Any string is considered to end with an empty suffix.
+bool String::EndsWithCaseInsensitive(
+    const std::string& str, const std::string& suffix) {
+  const size_t str_len = str.length();
+  const size_t suffix_len = suffix.length();
+  return (str_len >= suffix_len) &&
+         CaseInsensitiveCStringEquals(str.c_str() + str_len - suffix_len,
+                                      suffix.c_str());
+}
+
+// Formats an int value as "%02d".
+std::string String::FormatIntWidth2(int value) {
+  std::stringstream ss;
+  ss << std::setfill('0') << std::setw(2) << value;
+  return ss.str();
+}
+
+// Formats an int value as "%X".
+std::string String::FormatHexInt(int value) {
+  std::stringstream ss;
+  ss << std::hex << std::uppercase << value;
+  return ss.str();
+}
+
+// Formats a byte as "%02X".
+std::string String::FormatByte(unsigned char value) {
+  std::stringstream ss;
+  ss << std::setfill('0') << std::setw(2) << std::hex << std::uppercase
+     << static_cast<unsigned int>(value);
+  return ss.str();
+}
+
+// Converts the buffer in a stringstream to an std::string, converting NUL
+// bytes to "\\0" along the way.
+std::string StringStreamToString(::std::stringstream* ss) {
+  const ::std::string& str = ss->str();
+  const char* const start = str.c_str();
+  const char* const end = start + str.length();
+
+  std::string result;
+  result.reserve(2 * (end - start));
+  for (const char* ch = start; ch != end; ++ch) {
+    if (*ch == '\0') {
+      result += "\\0";  // Replaces NUL with "\\0";
+    } else {
+      result += *ch;
+    }
+  }
+
+  return result;
+}
+
+// Appends the user-supplied message to the Google-Test-generated message.
+std::string AppendUserMessage(const std::string& gtest_msg,
+                              const Message& user_msg) {
+  // Appends the user message if it's non-empty.
+  const std::string user_msg_string = user_msg.GetString();
+  if (user_msg_string.empty()) {
+    return gtest_msg;
+  }
+
+  return gtest_msg + "\n" + user_msg_string;
+}
+
+}  // namespace internal
+
+// class TestResult
+
+// Creates an empty TestResult.
+TestResult::TestResult()
+    : death_test_count_(0),
+      elapsed_time_(0) {
+}
+
+// D'tor.
+TestResult::~TestResult() {
+}
+
+// Returns the i-th test part result among all the results. i can
+// range from 0 to total_part_count() - 1. If i is not in that range,
+// aborts the program.
+const TestPartResult& TestResult::GetTestPartResult(int i) const {
+  if (i < 0 || i >= total_part_count())
+    internal::posix::Abort();
+  return test_part_results_.at(i);
+}
+
+// Returns the i-th test property. i can range from 0 to
+// test_property_count() - 1. If i is not in that range, aborts the
+// program.
+const TestProperty& TestResult::GetTestProperty(int i) const {
+  if (i < 0 || i >= test_property_count())
+    internal::posix::Abort();
+  return test_properties_.at(i);
+}
+
+// Clears the test part results.
+void TestResult::ClearTestPartResults() {
+  test_part_results_.clear();
+}
+
+// Adds a test part result to the list.
+void TestResult::AddTestPartResult(const TestPartResult& test_part_result) {
+  test_part_results_.push_back(test_part_result);
+}
+
+// Adds a test property to the list. If a property with the same key as the
+// supplied property is already represented, the value of this test_property
+// replaces the old value for that key.
+void TestResult::RecordProperty(const std::string& xml_element,
+                                const TestProperty& test_property) {
+  if (!ValidateTestProperty(xml_element, test_property)) {
+    return;
+  }
+  internal::MutexLock lock(&test_properites_mutex_);
+  const std::vector<TestProperty>::iterator property_with_matching_key =
+      std::find_if(test_properties_.begin(), test_properties_.end(),
+                   internal::TestPropertyKeyIs(test_property.key()));
+  if (property_with_matching_key == test_properties_.end()) {
+    test_properties_.push_back(test_property);
+    return;
+  }
+  property_with_matching_key->SetValue(test_property.value());
+}
+
+// The list of reserved attributes used in the <testsuites> element of XML
+// output.
+static const char* const kReservedTestSuitesAttributes[] = {
+  "disabled",
+  "errors",
+  "failures",
+  "name",
+  "random_seed",
+  "tests",
+  "time",
+  "timestamp"
+};
+
+// The list of reserved attributes used in the <testsuite> element of XML
+// output.
+static const char* const kReservedTestSuiteAttributes[] = {
+  "disabled",
+  "errors",
+  "failures",
+  "name",
+  "tests",
+  "time"
+};
+
+// The list of reserved attributes used in the <testcase> element of XML output.
+static const char* const kReservedTestCaseAttributes[] = {
+  "classname",
+  "name",
+  "status",
+  "time",
+  "type_param",
+  "value_param"
+};
+
+template <int kSize>
+std::vector<std::string> ArrayAsVector(const char* const (&array)[kSize]) {
+  return std::vector<std::string>(array, array + kSize);
+}
+
+static std::vector<std::string> GetReservedAttributesForElement(
+    const std::string& xml_element) {
+  if (xml_element == "testsuites") {
+    return ArrayAsVector(kReservedTestSuitesAttributes);
+  } else if (xml_element == "testsuite") {
+    return ArrayAsVector(kReservedTestSuiteAttributes);
+  } else if (xml_element == "testcase") {
+    return ArrayAsVector(kReservedTestCaseAttributes);
+  } else {
+    GTEST_CHECK_(false) << "Unrecognized xml_element provided: " << xml_element;
+  }
+  // This code is unreachable but some compilers may not realizes that.
+  return std::vector<std::string>();
+}
+
+static std::string FormatWordList(const std::vector<std::string>& words) {
+  Message word_list;
+  for (size_t i = 0; i < words.size(); ++i) {
+    if (i > 0 && words.size() > 2) {
+      word_list << ", ";
+    }
+    if (i == words.size() - 1) {
+      word_list << "and ";
+    }
+    word_list << "'" << words[i] << "'";
+  }
+  return word_list.GetString();
+}
+
+bool ValidateTestPropertyName(const std::string& property_name,
+                              const std::vector<std::string>& reserved_names) {
+  if (std::find(reserved_names.begin(), reserved_names.end(), property_name) !=
+          reserved_names.end()) {
+    ADD_FAILURE() << "Reserved key used in RecordProperty(): " << property_name
+                  << " (" << FormatWordList(reserved_names)
+                  << " are reserved by " << GTEST_NAME_ << ")";
+    return false;
+  }
+  return true;
+}
+
+// Adds a failure if the key is a reserved attribute of the element named
+// xml_element.  Returns true if the property is valid.
+bool TestResult::ValidateTestProperty(const std::string& xml_element,
+                                      const TestProperty& test_property) {
+  return ValidateTestPropertyName(test_property.key(),
+                                  GetReservedAttributesForElement(xml_element));
+}
+
+// Clears the object.
+void TestResult::Clear() {
+  test_part_results_.clear();
+  test_properties_.clear();
+  death_test_count_ = 0;
+  elapsed_time_ = 0;
+}
+
+// Returns true iff the test failed.
+bool TestResult::Failed() const {
+  for (int i = 0; i < total_part_count(); ++i) {
+    if (GetTestPartResult(i).failed())
+      return true;
+  }
+  return false;
+}
+
+// Returns true iff the test part fatally failed.
+static bool TestPartFatallyFailed(const TestPartResult& result) {
+  return result.fatally_failed();
+}
+
+// Returns true iff the test fatally failed.
+bool TestResult::HasFatalFailure() const {
+  return CountIf(test_part_results_, TestPartFatallyFailed) > 0;
+}
+
+// Returns true iff the test part non-fatally failed.
+static bool TestPartNonfatallyFailed(const TestPartResult& result) {
+  return result.nonfatally_failed();
+}
+
+// Returns true iff the test has a non-fatal failure.
+bool TestResult::HasNonfatalFailure() const {
+  return CountIf(test_part_results_, TestPartNonfatallyFailed) > 0;
+}
+
+// Gets the number of all test parts.  This is the sum of the number
+// of successful test parts and the number of failed test parts.
+int TestResult::total_part_count() const {
+  return static_cast<int>(test_part_results_.size());
+}
+
+// Returns the number of the test properties.
+int TestResult::test_property_count() const {
+  return static_cast<int>(test_properties_.size());
+}
+
+// class Test
+
+// Creates a Test object.
+
+// The c'tor saves the values of all Google Test flags.
+Test::Test()
+    : gtest_flag_saver_(new internal::GTestFlagSaver) {
+}
+
+// The d'tor restores the values of all Google Test flags.
+Test::~Test() {
+  delete gtest_flag_saver_;
+}
+
+// Sets up the test fixture.
+//
+// A sub-class may override this.
+void Test::SetUp() {
+}
+
+// Tears down the test fixture.
+//
+// A sub-class may override this.
+void Test::TearDown() {
+}
+
+// Allows user supplied key value pairs to be recorded for later output.
+void Test::RecordProperty(const std::string& key, const std::string& value) {
+  UnitTest::GetInstance()->RecordProperty(key, value);
+}
+
+// Allows user supplied key value pairs to be recorded for later output.
+void Test::RecordProperty(const std::string& key, int value) {
+  Message value_message;
+  value_message << value;
+  RecordProperty(key, value_message.GetString().c_str());
+}
+
+namespace internal {
+
+void ReportFailureInUnknownLocation(TestPartResult::Type result_type,
+                                    const std::string& message) {
+  // This function is a friend of UnitTest and as such has access to
+  // AddTestPartResult.
+  UnitTest::GetInstance()->AddTestPartResult(
+      result_type,
+      NULL,  // No info about the source file where the exception occurred.
+      -1,    // We have no info on which line caused the exception.
+      message,
+      "");   // No stack trace, either.
+}
+
+}  // namespace internal
+
+// Google Test requires all tests in the same test case to use the same test
+// fixture class.  This function checks if the current test has the
+// same fixture class as the first test in the current test case.  If
+// yes, it returns true; otherwise it generates a Google Test failure and
+// returns false.
+bool Test::HasSameFixtureClass() {
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  const TestCase* const test_case = impl->current_test_case();
+
+  // Info about the first test in the current test case.
+  const TestInfo* const first_test_info = test_case->test_info_list()[0];
+  const internal::TypeId first_fixture_id = first_test_info->fixture_class_id_;
+  const char* const first_test_name = first_test_info->name();
+
+  // Info about the current test.
+  const TestInfo* const this_test_info = impl->current_test_info();
+  const internal::TypeId this_fixture_id = this_test_info->fixture_class_id_;
+  const char* const this_test_name = this_test_info->name();
+
+  if (this_fixture_id != first_fixture_id) {
+    // Is the first test defined using TEST?
+    const bool first_is_TEST = first_fixture_id == internal::GetTestTypeId();
+    // Is this test defined using TEST?
+    const bool this_is_TEST = this_fixture_id == internal::GetTestTypeId();
+
+    if (first_is_TEST || this_is_TEST) {
+      // The user mixed TEST and TEST_F in this test case - we'll tell
+      // him/her how to fix it.
+
+      // Gets the name of the TEST and the name of the TEST_F.  Note
+      // that first_is_TEST and this_is_TEST cannot both be true, as
+      // the fixture IDs are different for the two tests.
+      const char* const TEST_name =
+          first_is_TEST ? first_test_name : this_test_name;
+      const char* const TEST_F_name =
+          first_is_TEST ? this_test_name : first_test_name;
+
+      ADD_FAILURE()
+          << "All tests in the same test case must use the same test fixture\n"
+          << "class, so mixing TEST_F and TEST in the same test case is\n"
+          << "illegal.  In test case " << this_test_info->test_case_name()
+          << ",\n"
+          << "test " << TEST_F_name << " is defined using TEST_F but\n"
+          << "test " << TEST_name << " is defined using TEST.  You probably\n"
+          << "want to change the TEST to TEST_F or move it to another test\n"
+          << "case.";
+    } else {
+      // The user defined two fixture classes with the same name in
+      // two namespaces - we'll tell him/her how to fix it.
+      ADD_FAILURE()
+          << "All tests in the same test case must use the same test fixture\n"
+          << "class.  However, in test case "
+          << this_test_info->test_case_name() << ",\n"
+          << "you defined test " << first_test_name
+          << " and test " << this_test_name << "\n"
+          << "using two different test fixture classes.  This can happen if\n"
+          << "the two classes are from different namespaces or translation\n"
+          << "units and have the same name.  You should probably rename one\n"
+          << "of the classes to put the tests into different test cases.";
+    }
+    return false;
+  }
+
+  return true;
+}
+
+#if GTEST_HAS_SEH
+
+// Adds an "exception thrown" fatal failure to the current test.  This
+// function returns its result via an output parameter pointer because VC++
+// prohibits creation of objects with destructors on stack in functions
+// using __try (see error C2712).
+static std::string* FormatSehExceptionMessage(DWORD exception_code,
+                                              const char* location) {
+  Message message;
+  message << "SEH exception with code 0x" << std::setbase(16) <<
+    exception_code << std::setbase(10) << " thrown in " << location << ".";
+
+  return new std::string(message.GetString());
+}
+
+#endif  // GTEST_HAS_SEH
+
+namespace internal {
+
+#if GTEST_HAS_EXCEPTIONS
+
+// Adds an "exception thrown" fatal failure to the current test.
+static std::string FormatCxxExceptionMessage(const char* description,
+                                             const char* location) {
+  Message message;
+  if (description != NULL) {
+    message << "C++ exception with description \"" << description << "\"";
+  } else {
+    message << "Unknown C++ exception";
+  }
+  message << " thrown in " << location << ".";
+
+  return message.GetString();
+}
+
+static std::string PrintTestPartResultToString(
+    const TestPartResult& test_part_result);
+
+GoogleTestFailureException::GoogleTestFailureException(
+    const TestPartResult& failure)
+    : ::std::runtime_error(PrintTestPartResultToString(failure).c_str()) {}
+
+#endif  // GTEST_HAS_EXCEPTIONS
+
+// We put these helper functions in the internal namespace as IBM's xlC
+// compiler rejects the code if they were declared static.
+
+// Runs the given method and handles SEH exceptions it throws, when
+// SEH is supported; returns the 0-value for type Result in case of an
+// SEH exception.  (Microsoft compilers cannot handle SEH and C++
+// exceptions in the same function.  Therefore, we provide a separate
+// wrapper function for handling SEH exceptions.)
+template <class T, typename Result>
+Result HandleSehExceptionsInMethodIfSupported(
+    T* object, Result (T::*method)(), const char* location) {
+#if GTEST_HAS_SEH
+  __try {
+    return (object->*method)();
+  } __except (internal::UnitTestOptions::GTestShouldProcessSEH(  // NOLINT
+      GetExceptionCode())) {
+    // We create the exception message on the heap because VC++ prohibits
+    // creation of objects with destructors on stack in functions using __try
+    // (see error C2712).
+    std::string* exception_message = FormatSehExceptionMessage(
+        GetExceptionCode(), location);
+    internal::ReportFailureInUnknownLocation(TestPartResult::kFatalFailure,
+                                             *exception_message);
+    delete exception_message;
+    return static_cast<Result>(0);
+  }
+#else
+  (void)location;
+  return (object->*method)();
+#endif  // GTEST_HAS_SEH
+}
+
+// Runs the given method and catches and reports C++ and/or SEH-style
+// exceptions, if they are supported; returns the 0-value for type
+// Result in case of an SEH exception.
+template <class T, typename Result>
+Result HandleExceptionsInMethodIfSupported(
+    T* object, Result (T::*method)(), const char* location) {
+  // NOTE: The user code can affect the way in which Google Test handles
+  // exceptions by setting GTEST_FLAG(catch_exceptions), but only before
+  // RUN_ALL_TESTS() starts. It is technically possible to check the flag
+  // after the exception is caught and either report or re-throw the
+  // exception based on the flag's value:
+  //
+  // try {
+  //   // Perform the test method.
+  // } catch (...) {
+  //   if (GTEST_FLAG(catch_exceptions))
+  //     // Report the exception as failure.
+  //   else
+  //     throw;  // Re-throws the original exception.
+  // }
+  //
+  // However, the purpose of this flag is to allow the program to drop into
+  // the debugger when the exception is thrown. On most platforms, once the
+  // control enters the catch block, the exception origin information is
+  // lost and the debugger will stop the program at the point of the
+  // re-throw in this function -- instead of at the point of the original
+  // throw statement in the code under test.  For this reason, we perform
+  // the check early, sacrificing the ability to affect Google Test's
+  // exception handling in the method where the exception is thrown.
+  if (internal::GetUnitTestImpl()->catch_exceptions()) {
+#if GTEST_HAS_EXCEPTIONS
+    try {
+      return HandleSehExceptionsInMethodIfSupported(object, method, location);
+    } catch (const internal::GoogleTestFailureException&) {  // NOLINT
+      // This exception type can only be thrown by a failed Google
+      // Test assertion with the intention of letting another testing
+      // framework catch it.  Therefore we just re-throw it.
+      throw;
+    } catch (const std::exception& e) {  // NOLINT
+      internal::ReportFailureInUnknownLocation(
+          TestPartResult::kFatalFailure,
+          FormatCxxExceptionMessage(e.what(), location));
+    } catch (...) {  // NOLINT
+      internal::ReportFailureInUnknownLocation(
+          TestPartResult::kFatalFailure,
+          FormatCxxExceptionMessage(NULL, location));
+    }
+    return static_cast<Result>(0);
+#else
+    return HandleSehExceptionsInMethodIfSupported(object, method, location);
+#endif  // GTEST_HAS_EXCEPTIONS
+  } else {
+    return (object->*method)();
+  }
+}
+
+}  // namespace internal
+
+// Runs the test and updates the test result.
+void Test::Run() {
+  if (!HasSameFixtureClass()) return;
+
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(this, &Test::SetUp, "SetUp()");
+  // We will run the test only if SetUp() was successful.
+  if (!HasFatalFailure()) {
+    impl->os_stack_trace_getter()->UponLeavingGTest();
+    internal::HandleExceptionsInMethodIfSupported(
+        this, &Test::TestBody, "the test body");
+  }
+
+  // However, we want to clean up as much as possible.  Hence we will
+  // always call TearDown(), even if SetUp() or the test body has
+  // failed.
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      this, &Test::TearDown, "TearDown()");
+}
+
+// Returns true iff the current test has a fatal failure.
+bool Test::HasFatalFailure() {
+  return internal::GetUnitTestImpl()->current_test_result()->HasFatalFailure();
+}
+
+// Returns true iff the current test has a non-fatal failure.
+bool Test::HasNonfatalFailure() {
+  return internal::GetUnitTestImpl()->current_test_result()->
+      HasNonfatalFailure();
+}
+
+// class TestInfo
+
+// Constructs a TestInfo object. It assumes ownership of the test factory
+// object.
+TestInfo::TestInfo(const std::string& a_test_case_name,
+                   const std::string& a_name,
+                   const char* a_type_param,
+                   const char* a_value_param,
+                   internal::TypeId fixture_class_id,
+                   internal::TestFactoryBase* factory)
+    : test_case_name_(a_test_case_name),
+      name_(a_name),
+      type_param_(a_type_param ? new std::string(a_type_param) : NULL),
+      value_param_(a_value_param ? new std::string(a_value_param) : NULL),
+      fixture_class_id_(fixture_class_id),
+      should_run_(false),
+      is_disabled_(false),
+      matches_filter_(false),
+      factory_(factory),
+      result_() {}
+
+// Destructs a TestInfo object.
+TestInfo::~TestInfo() { delete factory_; }
+
+namespace internal {
+
+// Creates a new TestInfo object and registers it with Google Test;
+// returns the created object.
+//
+// Arguments:
+//
+//   test_case_name:   name of the test case
+//   name:             name of the test
+//   type_param:       the name of the test's type parameter, or NULL if
+//                     this is not a typed or a type-parameterized test.
+//   value_param:      text representation of the test's value parameter,
+//                     or NULL if this is not a value-parameterized test.
+//   fixture_class_id: ID of the test fixture class
+//   set_up_tc:        pointer to the function that sets up the test case
+//   tear_down_tc:     pointer to the function that tears down the test case
+//   factory:          pointer to the factory that creates a test object.
+//                     The newly created TestInfo instance will assume
+//                     ownership of the factory object.
+TestInfo* MakeAndRegisterTestInfo(
+    const char* test_case_name,
+    const char* name,
+    const char* type_param,
+    const char* value_param,
+    TypeId fixture_class_id,
+    SetUpTestCaseFunc set_up_tc,
+    TearDownTestCaseFunc tear_down_tc,
+    TestFactoryBase* factory) {
+  TestInfo* const test_info =
+      new TestInfo(test_case_name, name, type_param, value_param,
+                   fixture_class_id, factory);
+  GetUnitTestImpl()->AddTestInfo(set_up_tc, tear_down_tc, test_info);
+  return test_info;
+}
+
+#if GTEST_HAS_PARAM_TEST
+void ReportInvalidTestCaseType(const char* test_case_name,
+                               const char* file, int line) {
+  Message errors;
+  errors
+      << "Attempted redefinition of test case " << test_case_name << ".\n"
+      << "All tests in the same test case must use the same test fixture\n"
+      << "class.  However, in test case " << test_case_name << ", you tried\n"
+      << "to define a test using a fixture class different from the one\n"
+      << "used earlier. This can happen if the two fixture classes are\n"
+      << "from different namespaces and have the same name. You should\n"
+      << "probably rename one of the classes to put the tests into different\n"
+      << "test cases.";
+
+  fprintf(stderr, "%s %s", FormatFileLocation(file, line).c_str(),
+          errors.GetString().c_str());
+}
+#endif  // GTEST_HAS_PARAM_TEST
+
+}  // namespace internal
+
+namespace {
+
+// A predicate that checks the test name of a TestInfo against a known
+// value.
+//
+// This is used for implementation of the TestCase class only.  We put
+// it in the anonymous namespace to prevent polluting the outer
+// namespace.
+//
+// TestNameIs is copyable.
+class TestNameIs {
+ public:
+  // Constructor.
+  //
+  // TestNameIs has NO default constructor.
+  explicit TestNameIs(const char* name)
+      : name_(name) {}
+
+  // Returns true iff the test name of test_info matches name_.
+  bool operator()(const TestInfo * test_info) const {
+    return test_info && test_info->name() == name_;
+  }
+
+ private:
+  std::string name_;
+};
+
+}  // namespace
+
+namespace internal {
+
+// This method expands all parameterized tests registered with macros TEST_P
+// and INSTANTIATE_TEST_CASE_P into regular tests and registers those.
+// This will be done just once during the program runtime.
+void UnitTestImpl::RegisterParameterizedTests() {
+#if GTEST_HAS_PARAM_TEST
+  if (!parameterized_tests_registered_) {
+    parameterized_test_registry_.RegisterTests();
+    parameterized_tests_registered_ = true;
+  }
+#endif
+}
+
+}  // namespace internal
+
+// Creates the test object, runs it, records its result, and then
+// deletes it.
+void TestInfo::Run() {
+  if (!should_run_) return;
+
+  // Tells UnitTest where to store test result.
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  impl->set_current_test_info(this);
+
+  TestEventListener* repeater = UnitTest::GetInstance()->listeners().repeater();
+
+  // Notifies the unit test event listeners that a test is about to start.
+  repeater->OnTestStart(*this);
+
+  const TimeInMillis start = internal::GetTimeInMillis();
+
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+
+  // Creates the test object.
+  Test* const test = internal::HandleExceptionsInMethodIfSupported(
+      factory_, &internal::TestFactoryBase::CreateTest,
+      "the test fixture's constructor");
+
+  // Runs the test only if the test object was created and its
+  // constructor didn't generate a fatal failure.
+  if ((test != NULL) && !Test::HasFatalFailure()) {
+    // This doesn't throw as all user code that can throw are wrapped into
+    // exception handling code.
+    test->Run();
+  }
+
+  // Deletes the test object.
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      test, &Test::DeleteSelf_, "the test fixture's destructor");
+
+  result_.set_elapsed_time(internal::GetTimeInMillis() - start);
+
+  // Notifies the unit test event listener that a test has just finished.
+  repeater->OnTestEnd(*this);
+
+  // Tells UnitTest to stop associating assertion results to this
+  // test.
+  impl->set_current_test_info(NULL);
+}
+
+// class TestCase
+
+// Gets the number of successful tests in this test case.
+int TestCase::successful_test_count() const {
+  return CountIf(test_info_list_, TestPassed);
+}
+
+// Gets the number of failed tests in this test case.
+int TestCase::failed_test_count() const {
+  return CountIf(test_info_list_, TestFailed);
+}
+
+// Gets the number of disabled tests that will be reported in the XML report.
+int TestCase::reportable_disabled_test_count() const {
+  return CountIf(test_info_list_, TestReportableDisabled);
+}
+
+// Gets the number of disabled tests in this test case.
+int TestCase::disabled_test_count() const {
+  return CountIf(test_info_list_, TestDisabled);
+}
+
+// Gets the number of tests to be printed in the XML report.
+int TestCase::reportable_test_count() const {
+  return CountIf(test_info_list_, TestReportable);
+}
+
+// Get the number of tests in this test case that should run.
+int TestCase::test_to_run_count() const {
+  return CountIf(test_info_list_, ShouldRunTest);
+}
+
+// Gets the number of all tests.
+int TestCase::total_test_count() const {
+  return static_cast<int>(test_info_list_.size());
+}
+
+// Creates a TestCase with the given name.
+//
+// Arguments:
+//
+//   name:         name of the test case
+//   a_type_param: the name of the test case's type parameter, or NULL if
+//                 this is not a typed or a type-parameterized test case.
+//   set_up_tc:    pointer to the function that sets up the test case
+//   tear_down_tc: pointer to the function that tears down the test case
+TestCase::TestCase(const char* a_name, const char* a_type_param,
+                   Test::SetUpTestCaseFunc set_up_tc,
+                   Test::TearDownTestCaseFunc tear_down_tc)
+    : name_(a_name),
+      type_param_(a_type_param ? new std::string(a_type_param) : NULL),
+      set_up_tc_(set_up_tc),
+      tear_down_tc_(tear_down_tc),
+      should_run_(false),
+      elapsed_time_(0) {
+}
+
+// Destructor of TestCase.
+TestCase::~TestCase() {
+  // Deletes every Test in the collection.
+  ForEach(test_info_list_, internal::Delete<TestInfo>);
+}
+
+// Returns the i-th test among all the tests. i can range from 0 to
+// total_test_count() - 1. If i is not in that range, returns NULL.
+const TestInfo* TestCase::GetTestInfo(int i) const {
+  const int index = GetElementOr(test_indices_, i, -1);
+  return index < 0 ? NULL : test_info_list_[index];
+}
+
+// Returns the i-th test among all the tests. i can range from 0 to
+// total_test_count() - 1. If i is not in that range, returns NULL.
+TestInfo* TestCase::GetMutableTestInfo(int i) {
+  const int index = GetElementOr(test_indices_, i, -1);
+  return index < 0 ? NULL : test_info_list_[index];
+}
+
+// Adds a test to this test case.  Will delete the test upon
+// destruction of the TestCase object.
+void TestCase::AddTestInfo(TestInfo * test_info) {
+  test_info_list_.push_back(test_info);
+  test_indices_.push_back(static_cast<int>(test_indices_.size()));
+}
+
+// Runs every test in this TestCase.
+void TestCase::Run() {
+  if (!should_run_) return;
+
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  impl->set_current_test_case(this);
+
+  TestEventListener* repeater = UnitTest::GetInstance()->listeners().repeater();
+
+  repeater->OnTestCaseStart(*this);
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      this, &TestCase::RunSetUpTestCase, "SetUpTestCase()");
+
+  const internal::TimeInMillis start = internal::GetTimeInMillis();
+  for (int i = 0; i < total_test_count(); i++) {
+    GetMutableTestInfo(i)->Run();
+  }
+  elapsed_time_ = internal::GetTimeInMillis() - start;
+
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      this, &TestCase::RunTearDownTestCase, "TearDownTestCase()");
+
+  repeater->OnTestCaseEnd(*this);
+  impl->set_current_test_case(NULL);
+}
+
+// Clears the results of all tests in this test case.
+void TestCase::ClearResult() {
+  ad_hoc_test_result_.Clear();
+  ForEach(test_info_list_, TestInfo::ClearTestResult);
+}
+
+// Shuffles the tests in this test case.
+void TestCase::ShuffleTests(internal::Random* random) {
+  Shuffle(random, &test_indices_);
+}
+
+// Restores the test order to before the first shuffle.
+void TestCase::UnshuffleTests() {
+  for (size_t i = 0; i < test_indices_.size(); i++) {
+    test_indices_[i] = static_cast<int>(i);
+  }
+}
+
+// Formats a countable noun.  Depending on its quantity, either the
+// singular form or the plural form is used. e.g.
+//
+// FormatCountableNoun(1, "formula", "formuli") returns "1 formula".
+// FormatCountableNoun(5, "book", "books") returns "5 books".
+static std::string FormatCountableNoun(int count,
+                                       const char * singular_form,
+                                       const char * plural_form) {
+  return internal::StreamableToString(count) + " " +
+      (count == 1 ? singular_form : plural_form);
+}
+
+// Formats the count of tests.
+static std::string FormatTestCount(int test_count) {
+  return FormatCountableNoun(test_count, "test", "tests");
+}
+
+// Formats the count of test cases.
+static std::string FormatTestCaseCount(int test_case_count) {
+  return FormatCountableNoun(test_case_count, "test case", "test cases");
+}
+
+// Converts a TestPartResult::Type enum to human-friendly string
+// representation.  Both kNonFatalFailure and kFatalFailure are translated
+// to "Failure", as the user usually doesn't care about the difference
+// between the two when viewing the test result.
+static const char * TestPartResultTypeToString(TestPartResult::Type type) {
+  switch (type) {
+    case TestPartResult::kSuccess:
+      return "Success";
+
+    case TestPartResult::kNonFatalFailure:
+    case TestPartResult::kFatalFailure:
+#ifdef _MSC_VER
+      return "error: ";
+#else
+      return "Failure\n";
+#endif
+    default:
+      return "Unknown result type";
+  }
+}
+
+namespace internal {
+
+// Prints a TestPartResult to an std::string.
+static std::string PrintTestPartResultToString(
+    const TestPartResult& test_part_result) {
+  return (Message()
+          << internal::FormatFileLocation(test_part_result.file_name(),
+                                          test_part_result.line_number())
+          << " " << TestPartResultTypeToString(test_part_result.type())
+          << test_part_result.message()).GetString();
+}
+
+// Prints a TestPartResult.
+static void PrintTestPartResult(const TestPartResult& test_part_result) {
+  const std::string& result =
+      PrintTestPartResultToString(test_part_result);
+  printf("%s\n", result.c_str());
+  fflush(stdout);
+  // If the test program runs in Visual Studio or a debugger, the
+  // following statements add the test part result message to the Output
+  // window such that the user can double-click on it to jump to the
+  // corresponding source code location; otherwise they do nothing.
+#if GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+  // We don't call OutputDebugString*() on Windows Mobile, as printing
+  // to stdout is done by OutputDebugString() there already - we don't
+  // want the same message printed twice.
+  ::OutputDebugStringA(result.c_str());
+  ::OutputDebugStringA("\n");
+#endif
+}
+
+// class PrettyUnitTestResultPrinter
+
+enum GTestColor {
+  COLOR_DEFAULT,
+  COLOR_RED,
+  COLOR_GREEN,
+  COLOR_YELLOW
+};
+
+#if GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+
+// Returns the character attribute for the given color.
+WORD GetColorAttribute(GTestColor color) {
+  switch (color) {
+    case COLOR_RED:    return FOREGROUND_RED;
+    case COLOR_GREEN:  return FOREGROUND_GREEN;
+    case COLOR_YELLOW: return FOREGROUND_RED | FOREGROUND_GREEN;
+    default:           return 0;
+  }
+}
+
+#else
+
+// Returns the ANSI color code for the given color.  COLOR_DEFAULT is
+// an invalid input.
+const char* GetAnsiColorCode(GTestColor color) {
+  switch (color) {
+    case COLOR_RED:     return "1";
+    case COLOR_GREEN:   return "2";
+    case COLOR_YELLOW:  return "3";
+    default:            return NULL;
+  };
+}
+
+#endif  // GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+
+// Returns true iff Google Test should use colors in the output.
+bool ShouldUseColor(bool stdout_is_tty) {
+  const char* const gtest_color = GTEST_FLAG(color).c_str();
+
+  if (String::CaseInsensitiveCStringEquals(gtest_color, "auto")) {
+#if GTEST_OS_WINDOWS
+    // On Windows the TERM variable is usually not set, but the
+    // console there does support colors.
+    return stdout_is_tty;
+#else
+    // On non-Windows platforms, we rely on the TERM variable.
+    const char* const term = posix::GetEnv("TERM");
+    const bool term_supports_color =
+        String::CStringEquals(term, "xterm") ||
+        String::CStringEquals(term, "xterm-color") ||
+        String::CStringEquals(term, "xterm-256color") ||
+        String::CStringEquals(term, "screen") ||
+        String::CStringEquals(term, "screen-256color") ||
+        String::CStringEquals(term, "linux") ||
+        String::CStringEquals(term, "cygwin");
+    return stdout_is_tty && term_supports_color;
+#endif  // GTEST_OS_WINDOWS
+  }
+
+  return String::CaseInsensitiveCStringEquals(gtest_color, "yes") ||
+      String::CaseInsensitiveCStringEquals(gtest_color, "true") ||
+      String::CaseInsensitiveCStringEquals(gtest_color, "t") ||
+      String::CStringEquals(gtest_color, "1");
+  // We take "yes", "true", "t", and "1" as meaning "yes".  If the
+  // value is neither one of these nor "auto", we treat it as "no" to
+  // be conservative.
+}
+
+// Helpers for printing colored strings to stdout. Note that on Windows, we
+// cannot simply emit special characters and have the terminal change colors.
+// This routine must actually emit the characters rather than return a string
+// that would be colored when printed, as can be done on Linux.
+void ColoredPrintf(GTestColor color, const char* fmt, ...) {
+  va_list args;
+  va_start(args, fmt);
+
+#if GTEST_OS_WINDOWS_MOBILE || GTEST_OS_SYMBIAN || GTEST_OS_ZOS || GTEST_OS_IOS
+  const bool use_color = false;
+#else
+  static const bool in_color_mode =
+      ShouldUseColor(posix::IsATTY(posix::FileNo(stdout)) != 0);
+  const bool use_color = in_color_mode && (color != COLOR_DEFAULT);
+#endif  // GTEST_OS_WINDOWS_MOBILE || GTEST_OS_SYMBIAN || GTEST_OS_ZOS
+  // The '!= 0' comparison is necessary to satisfy MSVC 7.1.
+
+  if (!use_color) {
+    vprintf(fmt, args);
+    va_end(args);
+    return;
+  }
+
+#if GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+  const HANDLE stdout_handle = GetStdHandle(STD_OUTPUT_HANDLE);
+
+  // Gets the current text color.
+  CONSOLE_SCREEN_BUFFER_INFO buffer_info;
+  GetConsoleScreenBufferInfo(stdout_handle, &buffer_info);
+  const WORD old_color_attrs = buffer_info.wAttributes;
+
+  // We need to flush the stream buffers into the console before each
+  // SetConsoleTextAttribute call lest it affect the text that is already
+  // printed but has not yet reached the console.
+  fflush(stdout);
+  SetConsoleTextAttribute(stdout_handle,
+                          GetColorAttribute(color) | FOREGROUND_INTENSITY);
+  vprintf(fmt, args);
+
+  fflush(stdout);
+  // Restores the text color.
+  SetConsoleTextAttribute(stdout_handle, old_color_attrs);
+#else
+  printf("\033[0;3%sm", GetAnsiColorCode(color));
+  vprintf(fmt, args);
+  printf("\033[m");  // Resets the terminal to default.
+#endif  // GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+  va_end(args);
+}
+
+// Text printed in Google Test's text output and --gunit_list_tests
+// output to label the type parameter and value parameter for a test.
+static const char kTypeParamLabel[] = "TypeParam";
+static const char kValueParamLabel[] = "GetParam()";
+
+void PrintFullTestCommentIfPresent(const TestInfo& test_info) {
+  const char* const type_param = test_info.type_param();
+  const char* const value_param = test_info.value_param();
+
+  if (type_param != NULL || value_param != NULL) {
+    printf(", where ");
+    if (type_param != NULL) {
+      printf("%s = %s", kTypeParamLabel, type_param);
+      if (value_param != NULL)
+        printf(" and ");
+    }
+    if (value_param != NULL) {
+      printf("%s = %s", kValueParamLabel, value_param);
+    }
+  }
+}
+
+// This class implements the TestEventListener interface.
+//
+// Class PrettyUnitTestResultPrinter is copyable.
+class PrettyUnitTestResultPrinter : public TestEventListener {
+ public:
+  PrettyUnitTestResultPrinter() {}
+  static void PrintTestName(const char * test_case, const char * test) {
+    printf("%s.%s", test_case, test);
+  }
+
+  // The following methods override what's in the TestEventListener class.
+  virtual void OnTestProgramStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationStart(const UnitTest& unit_test, int iteration);
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestCaseStart(const TestCase& test_case);
+  virtual void OnTestStart(const TestInfo& test_info);
+  virtual void OnTestPartResult(const TestPartResult& result);
+  virtual void OnTestEnd(const TestInfo& test_info);
+  virtual void OnTestCaseEnd(const TestCase& test_case);
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationEnd(const UnitTest& unit_test, int iteration);
+  virtual void OnTestProgramEnd(const UnitTest& /*unit_test*/) {}
+
+ private:
+  static void PrintFailedTests(const UnitTest& unit_test);
+};
+
+  // Fired before each iteration of tests starts.
+void PrettyUnitTestResultPrinter::OnTestIterationStart(
+    const UnitTest& unit_test, int iteration) {
+  if (GTEST_FLAG(repeat) != 1)
+    printf("\nRepeating all tests (iteration %d) . . .\n\n", iteration + 1);
+
+  const char* const filter = GTEST_FLAG(filter).c_str();
+
+  // Prints the filter if it's not *.  This reminds the user that some
+  // tests may be skipped.
+  if (!String::CStringEquals(filter, kUniversalFilter)) {
+    ColoredPrintf(COLOR_YELLOW,
+                  "Note: %s filter = %s\n", GTEST_NAME_, filter);
+  }
+
+  if (internal::ShouldShard(kTestTotalShards, kTestShardIndex, false)) {
+    const Int32 shard_index = Int32FromEnvOrDie(kTestShardIndex, -1);
+    ColoredPrintf(COLOR_YELLOW,
+                  "Note: This is test shard %d of %s.\n",
+                  static_cast<int>(shard_index) + 1,
+                  internal::posix::GetEnv(kTestTotalShards));
+  }
+
+  if (GTEST_FLAG(shuffle)) {
+    ColoredPrintf(COLOR_YELLOW,
+                  "Note: Randomizing tests' orders with a seed of %d .\n",
+                  unit_test.random_seed());
+  }
+
+  ColoredPrintf(COLOR_GREEN,  "[==========] ");
+  printf("Running %s from %s.\n",
+         FormatTestCount(unit_test.test_to_run_count()).c_str(),
+         FormatTestCaseCount(unit_test.test_case_to_run_count()).c_str());
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnEnvironmentsSetUpStart(
+    const UnitTest& /*unit_test*/) {
+  ColoredPrintf(COLOR_GREEN,  "[----------] ");
+  printf("Global test environment set-up.\n");
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestCaseStart(const TestCase& test_case) {
+  const std::string counts =
+      FormatCountableNoun(test_case.test_to_run_count(), "test", "tests");
+  ColoredPrintf(COLOR_GREEN, "[----------] ");
+  printf("%s from %s", counts.c_str(), test_case.name());
+  if (test_case.type_param() == NULL) {
+    printf("\n");
+  } else {
+    printf(", where %s = %s\n", kTypeParamLabel, test_case.type_param());
+  }
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestStart(const TestInfo& test_info) {
+  ColoredPrintf(COLOR_GREEN,  "[ RUN      ] ");
+  PrintTestName(test_info.test_case_name(), test_info.name());
+  printf("\n");
+  fflush(stdout);
+}
+
+// Called after an assertion failure.
+void PrettyUnitTestResultPrinter::OnTestPartResult(
+    const TestPartResult& result) {
+  // If the test part succeeded, we don't need to do anything.
+  if (result.type() == TestPartResult::kSuccess)
+    return;
+
+  // Print failure message from the assertion (e.g. expected this and got that).
+  PrintTestPartResult(result);
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestEnd(const TestInfo& test_info) {
+  if (test_info.result()->Passed()) {
+    ColoredPrintf(COLOR_GREEN, "[       OK ] ");
+  } else {
+    ColoredPrintf(COLOR_RED, "[  FAILED  ] ");
+  }
+  PrintTestName(test_info.test_case_name(), test_info.name());
+  if (test_info.result()->Failed())
+    PrintFullTestCommentIfPresent(test_info);
+
+  if (GTEST_FLAG(print_time)) {
+    printf(" (%s ms)\n", internal::StreamableToString(
+           test_info.result()->elapsed_time()).c_str());
+  } else {
+    printf("\n");
+  }
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestCaseEnd(const TestCase& test_case) {
+  if (!GTEST_FLAG(print_time)) return;
+
+  const std::string counts =
+      FormatCountableNoun(test_case.test_to_run_count(), "test", "tests");
+  ColoredPrintf(COLOR_GREEN, "[----------] ");
+  printf("%s from %s (%s ms total)\n\n",
+         counts.c_str(), test_case.name(),
+         internal::StreamableToString(test_case.elapsed_time()).c_str());
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnEnvironmentsTearDownStart(
+    const UnitTest& /*unit_test*/) {
+  ColoredPrintf(COLOR_GREEN,  "[----------] ");
+  printf("Global test environment tear-down\n");
+  fflush(stdout);
+}
+
+// Internal helper for printing the list of failed tests.
+void PrettyUnitTestResultPrinter::PrintFailedTests(const UnitTest& unit_test) {
+  const int failed_test_count = unit_test.failed_test_count();
+  if (failed_test_count == 0) {
+    return;
+  }
+
+  for (int i = 0; i < unit_test.total_test_case_count(); ++i) {
+    const TestCase& test_case = *unit_test.GetTestCase(i);
+    if (!test_case.should_run() || (test_case.failed_test_count() == 0)) {
+      continue;
+    }
+    for (int j = 0; j < test_case.total_test_count(); ++j) {
+      const TestInfo& test_info = *test_case.GetTestInfo(j);
+      if (!test_info.should_run() || test_info.result()->Passed()) {
+        continue;
+      }
+      ColoredPrintf(COLOR_RED, "[  FAILED  ] ");
+      printf("%s.%s", test_case.name(), test_info.name());
+      PrintFullTestCommentIfPresent(test_info);
+      printf("\n");
+    }
+  }
+}
+
+void PrettyUnitTestResultPrinter::OnTestIterationEnd(const UnitTest& unit_test,
+                                                     int /*iteration*/) {
+  ColoredPrintf(COLOR_GREEN,  "[==========] ");
+  printf("%s from %s ran.",
+         FormatTestCount(unit_test.test_to_run_count()).c_str(),
+         FormatTestCaseCount(unit_test.test_case_to_run_count()).c_str());
+  if (GTEST_FLAG(print_time)) {
+    printf(" (%s ms total)",
+           internal::StreamableToString(unit_test.elapsed_time()).c_str());
+  }
+  printf("\n");
+  ColoredPrintf(COLOR_GREEN,  "[  PASSED  ] ");
+  printf("%s.\n", FormatTestCount(unit_test.successful_test_count()).c_str());
+
+  int num_failures = unit_test.failed_test_count();
+  if (!unit_test.Passed()) {
+    const int failed_test_count = unit_test.failed_test_count();
+    ColoredPrintf(COLOR_RED,  "[  FAILED  ] ");
+    printf("%s, listed below:\n", FormatTestCount(failed_test_count).c_str());
+    PrintFailedTests(unit_test);
+    printf("\n%2d FAILED %s\n", num_failures,
+                        num_failures == 1 ? "TEST" : "TESTS");
+  }
+
+  int num_disabled = unit_test.reportable_disabled_test_count();
+  if (num_disabled && !GTEST_FLAG(also_run_disabled_tests)) {
+    if (!num_failures) {
+      printf("\n");  // Add a spacer if no FAILURE banner is displayed.
+    }
+    ColoredPrintf(COLOR_YELLOW,
+                  "  YOU HAVE %d DISABLED %s\n\n",
+                  num_disabled,
+                  num_disabled == 1 ? "TEST" : "TESTS");
+  }
+  // Ensure that Google Test output is printed before, e.g., heapchecker output.
+  fflush(stdout);
+}
+
+// End PrettyUnitTestResultPrinter
+
+// class TestEventRepeater
+//
+// This class forwards events to other event listeners.
+class TestEventRepeater : public TestEventListener {
+ public:
+  TestEventRepeater() : forwarding_enabled_(true) {}
+  virtual ~TestEventRepeater();
+  void Append(TestEventListener *listener);
+  TestEventListener* Release(TestEventListener* listener);
+
+  // Controls whether events will be forwarded to listeners_. Set to false
+  // in death test child processes.
+  bool forwarding_enabled() const { return forwarding_enabled_; }
+  void set_forwarding_enabled(bool enable) { forwarding_enabled_ = enable; }
+
+  virtual void OnTestProgramStart(const UnitTest& unit_test);
+  virtual void OnTestIterationStart(const UnitTest& unit_test, int iteration);
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& unit_test);
+  virtual void OnTestCaseStart(const TestCase& test_case);
+  virtual void OnTestStart(const TestInfo& test_info);
+  virtual void OnTestPartResult(const TestPartResult& result);
+  virtual void OnTestEnd(const TestInfo& test_info);
+  virtual void OnTestCaseEnd(const TestCase& test_case);
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& unit_test);
+  virtual void OnTestIterationEnd(const UnitTest& unit_test, int iteration);
+  virtual void OnTestProgramEnd(const UnitTest& unit_test);
+
+ private:
+  // Controls whether events will be forwarded to listeners_. Set to false
+  // in death test child processes.
+  bool forwarding_enabled_;
+  // The list of listeners that receive events.
+  std::vector<TestEventListener*> listeners_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestEventRepeater);
+};
+
+TestEventRepeater::~TestEventRepeater() {
+  ForEach(listeners_, Delete<TestEventListener>);
+}
+
+void TestEventRepeater::Append(TestEventListener *listener) {
+  listeners_.push_back(listener);
+}
+
+// TODO(vladl@google.com): Factor the search functionality into Vector::Find.
+TestEventListener* TestEventRepeater::Release(TestEventListener *listener) {
+  for (size_t i = 0; i < listeners_.size(); ++i) {
+    if (listeners_[i] == listener) {
+      listeners_.erase(listeners_.begin() + i);
+      return listener;
+    }
+  }
+
+  return NULL;
+}
+
+// Since most methods are very similar, use macros to reduce boilerplate.
+// This defines a member that forwards the call to all listeners.
+#define GTEST_REPEATER_METHOD_(Name, Type) \
+void TestEventRepeater::Name(const Type& parameter) { \
+  if (forwarding_enabled_) { \
+    for (size_t i = 0; i < listeners_.size(); i++) { \
+      listeners_[i]->Name(parameter); \
+    } \
+  } \
+}
+// This defines a member that forwards the call to all listeners in reverse
+// order.
+#define GTEST_REVERSE_REPEATER_METHOD_(Name, Type) \
+void TestEventRepeater::Name(const Type& parameter) { \
+  if (forwarding_enabled_) { \
+    for (int i = static_cast<int>(listeners_.size()) - 1; i >= 0; i--) { \
+      listeners_[i]->Name(parameter); \
+    } \
+  } \
+}
+
+GTEST_REPEATER_METHOD_(OnTestProgramStart, UnitTest)
+GTEST_REPEATER_METHOD_(OnEnvironmentsSetUpStart, UnitTest)
+GTEST_REPEATER_METHOD_(OnTestCaseStart, TestCase)
+GTEST_REPEATER_METHOD_(OnTestStart, TestInfo)
+GTEST_REPEATER_METHOD_(OnTestPartResult, TestPartResult)
+GTEST_REPEATER_METHOD_(OnEnvironmentsTearDownStart, UnitTest)
+GTEST_REVERSE_REPEATER_METHOD_(OnEnvironmentsSetUpEnd, UnitTest)
+GTEST_REVERSE_REPEATER_METHOD_(OnEnvironmentsTearDownEnd, UnitTest)
+GTEST_REVERSE_REPEATER_METHOD_(OnTestEnd, TestInfo)
+GTEST_REVERSE_REPEATER_METHOD_(OnTestCaseEnd, TestCase)
+GTEST_REVERSE_REPEATER_METHOD_(OnTestProgramEnd, UnitTest)
+
+#undef GTEST_REPEATER_METHOD_
+#undef GTEST_REVERSE_REPEATER_METHOD_
+
+void TestEventRepeater::OnTestIterationStart(const UnitTest& unit_test,
+                                             int iteration) {
+  if (forwarding_enabled_) {
+    for (size_t i = 0; i < listeners_.size(); i++) {
+      listeners_[i]->OnTestIterationStart(unit_test, iteration);
+    }
+  }
+}
+
+void TestEventRepeater::OnTestIterationEnd(const UnitTest& unit_test,
+                                           int iteration) {
+  if (forwarding_enabled_) {
+    for (int i = static_cast<int>(listeners_.size()) - 1; i >= 0; i--) {
+      listeners_[i]->OnTestIterationEnd(unit_test, iteration);
+    }
+  }
+}
+
+// End TestEventRepeater
+
+// This class generates an XML output file.
+class XmlUnitTestResultPrinter : public EmptyTestEventListener {
+ public:
+  explicit XmlUnitTestResultPrinter(const char* output_file);
+
+  virtual void OnTestIterationEnd(const UnitTest& unit_test, int iteration);
+
+ private:
+  // Is c a whitespace character that is normalized to a space character
+  // when it appears in an XML attribute value?
+  static bool IsNormalizableWhitespace(char c) {
+    return c == 0x9 || c == 0xA || c == 0xD;
+  }
+
+  // May c appear in a well-formed XML document?
+  static bool IsValidXmlCharacter(char c) {
+    return IsNormalizableWhitespace(c) || c >= 0x20;
+  }
+
+  // Returns an XML-escaped copy of the input string str.  If
+  // is_attribute is true, the text is meant to appear as an attribute
+  // value, and normalizable whitespace is preserved by replacing it
+  // with character references.
+  static std::string EscapeXml(const std::string& str, bool is_attribute);
+
+  // Returns the given string with all characters invalid in XML removed.
+  static std::string RemoveInvalidXmlCharacters(const std::string& str);
+
+  // Convenience wrapper around EscapeXml when str is an attribute value.
+  static std::string EscapeXmlAttribute(const std::string& str) {
+    return EscapeXml(str, true);
+  }
+
+  // Convenience wrapper around EscapeXml when str is not an attribute value.
+  static std::string EscapeXmlText(const char* str) {
+    return EscapeXml(str, false);
+  }
+
+  // Verifies that the given attribute belongs to the given element and
+  // streams the attribute as XML.
+  static void OutputXmlAttribute(std::ostream* stream,
+                                 const std::string& element_name,
+                                 const std::string& name,
+                                 const std::string& value);
+
+  // Streams an XML CDATA section, escaping invalid CDATA sequences as needed.
+  static void OutputXmlCDataSection(::std::ostream* stream, const char* data);
+
+  // Streams an XML representation of a TestInfo object.
+  static void OutputXmlTestInfo(::std::ostream* stream,
+                                const char* test_case_name,
+                                const TestInfo& test_info);
+
+  // Prints an XML representation of a TestCase object
+  static void PrintXmlTestCase(::std::ostream* stream,
+                               const TestCase& test_case);
+
+  // Prints an XML summary of unit_test to output stream out.
+  static void PrintXmlUnitTest(::std::ostream* stream,
+                               const UnitTest& unit_test);
+
+  // Produces a string representing the test properties in a result as space
+  // delimited XML attributes based on the property key="value" pairs.
+  // When the std::string is not empty, it includes a space at the beginning,
+  // to delimit this attribute from prior attributes.
+  static std::string TestPropertiesAsXmlAttributes(const TestResult& result);
+
+  // The output file.
+  const std::string output_file_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(XmlUnitTestResultPrinter);
+};
+
+// Creates a new XmlUnitTestResultPrinter.
+XmlUnitTestResultPrinter::XmlUnitTestResultPrinter(const char* output_file)
+    : output_file_(output_file) {
+  if (output_file_.c_str() == NULL || output_file_.empty()) {
+    fprintf(stderr, "XML output file may not be null\n");
+    fflush(stderr);
+    exit(EXIT_FAILURE);
+  }
+}
+
+// Called after the unit test ends.
+void XmlUnitTestResultPrinter::OnTestIterationEnd(const UnitTest& unit_test,
+                                                  int /*iteration*/) {
+  FILE* xmlout = NULL;
+  FilePath output_file(output_file_);
+  FilePath output_dir(output_file.RemoveFileName());
+
+  if (output_dir.CreateDirectoriesRecursively()) {
+    xmlout = posix::FOpen(output_file_.c_str(), "w");
+  }
+  if (xmlout == NULL) {
+    // TODO(wan): report the reason of the failure.
+    //
+    // We don't do it for now as:
+    //
+    //   1. There is no urgent need for it.
+    //   2. It's a bit involved to make the errno variable thread-safe on
+    //      all three operating systems (Linux, Windows, and Mac OS).
+    //   3. To interpret the meaning of errno in a thread-safe way,
+    //      we need the strerror_r() function, which is not available on
+    //      Windows.
+    fprintf(stderr,
+            "Unable to open file \"%s\"\n",
+            output_file_.c_str());
+    fflush(stderr);
+    exit(EXIT_FAILURE);
+  }
+  std::stringstream stream;
+  PrintXmlUnitTest(&stream, unit_test);
+  fprintf(xmlout, "%s", StringStreamToString(&stream).c_str());
+  fclose(xmlout);
+}
+
+// Returns an XML-escaped copy of the input string str.  If is_attribute
+// is true, the text is meant to appear as an attribute value, and
+// normalizable whitespace is preserved by replacing it with character
+// references.
+//
+// Invalid XML characters in str, if any, are stripped from the output.
+// It is expected that most, if not all, of the text processed by this
+// module will consist of ordinary English text.
+// If this module is ever modified to produce version 1.1 XML output,
+// most invalid characters can be retained using character references.
+// TODO(wan): It might be nice to have a minimally invasive, human-readable
+// escaping scheme for invalid characters, rather than dropping them.
+std::string XmlUnitTestResultPrinter::EscapeXml(
+    const std::string& str, bool is_attribute) {
+  Message m;
+
+  for (size_t i = 0; i < str.size(); ++i) {
+    const char ch = str[i];
+    switch (ch) {
+      case '<':
+        m << "&lt;";
+        break;
+      case '>':
+        m << "&gt;";
+        break;
+      case '&':
+        m << "&amp;";
+        break;
+      case '\'':
+        if (is_attribute)
+          m << "&apos;";
+        else
+          m << '\'';
+        break;
+      case '"':
+        if (is_attribute)
+          m << "&quot;";
+        else
+          m << '"';
+        break;
+      default:
+        if (IsValidXmlCharacter(ch)) {
+          if (is_attribute && IsNormalizableWhitespace(ch))
+            m << "&#x" << String::FormatByte(static_cast<unsigned char>(ch))
+              << ";";
+          else
+            m << ch;
+        }
+        break;
+    }
+  }
+
+  return m.GetString();
+}
+
+// Returns the given string with all characters invalid in XML removed.
+// Currently invalid characters are dropped from the string. An
+// alternative is to replace them with certain characters such as . or ?.
+std::string XmlUnitTestResultPrinter::RemoveInvalidXmlCharacters(
+    const std::string& str) {
+  std::string output;
+  output.reserve(str.size());
+  for (std::string::const_iterator it = str.begin(); it != str.end(); ++it)
+    if (IsValidXmlCharacter(*it))
+      output.push_back(*it);
+
+  return output;
+}
+
+// The following routines generate an XML representation of a UnitTest
+// object.
+//
+// This is how Google Test concepts map to the DTD:
+//
+// <testsuites name="AllTests">        <-- corresponds to a UnitTest object
+//   <testsuite name="testcase-name">  <-- corresponds to a TestCase object
+//     <testcase name="test-name">     <-- corresponds to a TestInfo object
+//       <failure message="...">...</failure>
+//       <failure message="...">...</failure>
+//       <failure message="...">...</failure>
+//                                     <-- individual assertion failures
+//     </testcase>
+//   </testsuite>
+// </testsuites>
+
+// Formats the given time in milliseconds as seconds.
+std::string FormatTimeInMillisAsSeconds(TimeInMillis ms) {
+  ::std::stringstream ss;
+  ss << ms/1000.0;
+  return ss.str();
+}
+
+// Converts the given epoch time in milliseconds to a date string in the ISO
+// 8601 format, without the timezone information.
+std::string FormatEpochTimeInMillisAsIso8601(TimeInMillis ms) {
+  // Using non-reentrant version as localtime_r is not portable.
+  time_t seconds = static_cast<time_t>(ms / 1000);
+#ifdef _MSC_VER
+# pragma warning(push)          // Saves the current warning state.
+# pragma warning(disable:4996)  // Temporarily disables warning 4996
+                                // (function or variable may be unsafe).
+  const struct tm* const time_struct = localtime(&seconds);  // NOLINT
+# pragma warning(pop)           // Restores the warning state again.
+#else
+  const struct tm* const time_struct = localtime(&seconds);  // NOLINT
+#endif
+  if (time_struct == NULL)
+    return "";  // Invalid ms value
+
+  // YYYY-MM-DDThh:mm:ss
+  return StreamableToString(time_struct->tm_year + 1900) + "-" +
+      String::FormatIntWidth2(time_struct->tm_mon + 1) + "-" +
+      String::FormatIntWidth2(time_struct->tm_mday) + "T" +
+      String::FormatIntWidth2(time_struct->tm_hour) + ":" +
+      String::FormatIntWidth2(time_struct->tm_min) + ":" +
+      String::FormatIntWidth2(time_struct->tm_sec);
+}
+
+// Streams an XML CDATA section, escaping invalid CDATA sequences as needed.
+void XmlUnitTestResultPrinter::OutputXmlCDataSection(::std::ostream* stream,
+                                                     const char* data) {
+  const char* segment = data;
+  *stream << "<![CDATA[";
+  for (;;) {
+    const char* const next_segment = strstr(segment, "]]>");
+    if (next_segment != NULL) {
+      stream->write(
+          segment, static_cast<std::streamsize>(next_segment - segment));
+      *stream << "]]>]]&gt;<![CDATA[";
+      segment = next_segment + strlen("]]>");
+    } else {
+      *stream << segment;
+      break;
+    }
+  }
+  *stream << "]]>";
+}
+
+void XmlUnitTestResultPrinter::OutputXmlAttribute(
+    std::ostream* stream,
+    const std::string& element_name,
+    const std::string& name,
+    const std::string& value) {
+  const std::vector<std::string>& allowed_names =
+      GetReservedAttributesForElement(element_name);
+
+  GTEST_CHECK_(std::find(allowed_names.begin(), allowed_names.end(), name) !=
+                   allowed_names.end())
+      << "Attribute " << name << " is not allowed for element <" << element_name
+      << ">.";
+
+  *stream << " " << name << "=\"" << EscapeXmlAttribute(value) << "\"";
+}
+
+// Prints an XML representation of a TestInfo object.
+// TODO(wan): There is also value in printing properties with the plain printer.
+void XmlUnitTestResultPrinter::OutputXmlTestInfo(::std::ostream* stream,
+                                                 const char* test_case_name,
+                                                 const TestInfo& test_info) {
+  const TestResult& result = *test_info.result();
+  const std::string kTestcase = "testcase";
+
+  *stream << "    <testcase";
+  OutputXmlAttribute(stream, kTestcase, "name", test_info.name());
+
+  if (test_info.value_param() != NULL) {
+    OutputXmlAttribute(stream, kTestcase, "value_param",
+                       test_info.value_param());
+  }
+  if (test_info.type_param() != NULL) {
+    OutputXmlAttribute(stream, kTestcase, "type_param", test_info.type_param());
+  }
+
+  OutputXmlAttribute(stream, kTestcase, "status",
+                     test_info.should_run() ? "run" : "notrun");
+  OutputXmlAttribute(stream, kTestcase, "time",
+                     FormatTimeInMillisAsSeconds(result.elapsed_time()));
+  OutputXmlAttribute(stream, kTestcase, "classname", test_case_name);
+  *stream << TestPropertiesAsXmlAttributes(result);
+
+  int failures = 0;
+  for (int i = 0; i < result.total_part_count(); ++i) {
+    const TestPartResult& part = result.GetTestPartResult(i);
+    if (part.failed()) {
+      if (++failures == 1) {
+        *stream << ">\n";
+      }
+      const string location = internal::FormatCompilerIndependentFileLocation(
+          part.file_name(), part.line_number());
+      const string summary = location + "\n" + part.summary();
+      *stream << "      <failure message=\""
+              << EscapeXmlAttribute(summary.c_str())
+              << "\" type=\"\">";
+      const string detail = location + "\n" + part.message();
+      OutputXmlCDataSection(stream, RemoveInvalidXmlCharacters(detail).c_str());
+      *stream << "</failure>\n";
+    }
+  }
+
+  if (failures == 0)
+    *stream << " />\n";
+  else
+    *stream << "    </testcase>\n";
+}
+
+// Prints an XML representation of a TestCase object
+void XmlUnitTestResultPrinter::PrintXmlTestCase(std::ostream* stream,
+                                                const TestCase& test_case) {
+  const std::string kTestsuite = "testsuite";
+  *stream << "  <" << kTestsuite;
+  OutputXmlAttribute(stream, kTestsuite, "name", test_case.name());
+  OutputXmlAttribute(stream, kTestsuite, "tests",
+                     StreamableToString(test_case.reportable_test_count()));
+  OutputXmlAttribute(stream, kTestsuite, "failures",
+                     StreamableToString(test_case.failed_test_count()));
+  OutputXmlAttribute(
+      stream, kTestsuite, "disabled",
+      StreamableToString(test_case.reportable_disabled_test_count()));
+  OutputXmlAttribute(stream, kTestsuite, "errors", "0");
+  OutputXmlAttribute(stream, kTestsuite, "time",
+                     FormatTimeInMillisAsSeconds(test_case.elapsed_time()));
+  *stream << TestPropertiesAsXmlAttributes(test_case.ad_hoc_test_result())
+          << ">\n";
+
+  for (int i = 0; i < test_case.total_test_count(); ++i) {
+    if (test_case.GetTestInfo(i)->is_reportable())
+      OutputXmlTestInfo(stream, test_case.name(), *test_case.GetTestInfo(i));
+  }
+  *stream << "  </" << kTestsuite << ">\n";
+}
+
+// Prints an XML summary of unit_test to output stream out.
+void XmlUnitTestResultPrinter::PrintXmlUnitTest(std::ostream* stream,
+                                                const UnitTest& unit_test) {
+  const std::string kTestsuites = "testsuites";
+
+  *stream << "<?xml version=\"1.0\" encoding=\"UTF-8\"?>\n";
+  *stream << "<" << kTestsuites;
+
+  OutputXmlAttribute(stream, kTestsuites, "tests",
+                     StreamableToString(unit_test.reportable_test_count()));
+  OutputXmlAttribute(stream, kTestsuites, "failures",
+                     StreamableToString(unit_test.failed_test_count()));
+  OutputXmlAttribute(
+      stream, kTestsuites, "disabled",
+      StreamableToString(unit_test.reportable_disabled_test_count()));
+  OutputXmlAttribute(stream, kTestsuites, "errors", "0");
+  OutputXmlAttribute(
+      stream, kTestsuites, "timestamp",
+      FormatEpochTimeInMillisAsIso8601(unit_test.start_timestamp()));
+  OutputXmlAttribute(stream, kTestsuites, "time",
+                     FormatTimeInMillisAsSeconds(unit_test.elapsed_time()));
+
+  if (GTEST_FLAG(shuffle)) {
+    OutputXmlAttribute(stream, kTestsuites, "random_seed",
+                       StreamableToString(unit_test.random_seed()));
+  }
+
+  *stream << TestPropertiesAsXmlAttributes(unit_test.ad_hoc_test_result());
+
+  OutputXmlAttribute(stream, kTestsuites, "name", "AllTests");
+  *stream << ">\n";
+
+  for (int i = 0; i < unit_test.total_test_case_count(); ++i) {
+    if (unit_test.GetTestCase(i)->reportable_test_count() > 0)
+      PrintXmlTestCase(stream, *unit_test.GetTestCase(i));
+  }
+  *stream << "</" << kTestsuites << ">\n";
+}
+
+// Produces a string representing the test properties in a result as space
+// delimited XML attributes based on the property key="value" pairs.
+std::string XmlUnitTestResultPrinter::TestPropertiesAsXmlAttributes(
+    const TestResult& result) {
+  Message attributes;
+  for (int i = 0; i < result.test_property_count(); ++i) {
+    const TestProperty& property = result.GetTestProperty(i);
+    attributes << " " << property.key() << "="
+        << "\"" << EscapeXmlAttribute(property.value()) << "\"";
+  }
+  return attributes.GetString();
+}
+
+// End XmlUnitTestResultPrinter
+
+#if GTEST_CAN_STREAM_RESULTS_
+
+// Checks if str contains '=', '&', '%' or '\n' characters. If yes,
+// replaces them by "%xx" where xx is their hexadecimal value. For
+// example, replaces "=" with "%3D".  This algorithm is O(strlen(str))
+// in both time and space -- important as the input str may contain an
+// arbitrarily long test failure message and stack trace.
+string StreamingListener::UrlEncode(const char* str) {
+  string result;
+  result.reserve(strlen(str) + 1);
+  for (char ch = *str; ch != '\0'; ch = *++str) {
+    switch (ch) {
+      case '%':
+      case '=':
+      case '&':
+      case '\n':
+        result.append("%" + String::FormatByte(static_cast<unsigned char>(ch)));
+        break;
+      default:
+        result.push_back(ch);
+        break;
+    }
+  }
+  return result;
+}
+
+void StreamingListener::SocketWriter::MakeConnection() {
+  GTEST_CHECK_(sockfd_ == -1)
+      << "MakeConnection() can't be called when there is already a connection.";
+
+  addrinfo hints;
+  memset(&hints, 0, sizeof(hints));
+  hints.ai_family = AF_UNSPEC;    // To allow both IPv4 and IPv6 addresses.
+  hints.ai_socktype = SOCK_STREAM;
+  addrinfo* servinfo = NULL;
+
+  // Use the getaddrinfo() to get a linked list of IP addresses for
+  // the given host name.
+  const int error_num = getaddrinfo(
+      host_name_.c_str(), port_num_.c_str(), &hints, &servinfo);
+  if (error_num != 0) {
+    GTEST_LOG_(WARNING) << "stream_result_to: getaddrinfo() failed: "
+                        << gai_strerror(error_num);
+  }
+
+  // Loop through all the results and connect to the first we can.
+  for (addrinfo* cur_addr = servinfo; sockfd_ == -1 && cur_addr != NULL;
+       cur_addr = cur_addr->ai_next) {
+    sockfd_ = socket(
+        cur_addr->ai_family, cur_addr->ai_socktype, cur_addr->ai_protocol);
+    if (sockfd_ != -1) {
+      // Connect the client socket to the server socket.
+      if (connect(sockfd_, cur_addr->ai_addr, cur_addr->ai_addrlen) == -1) {
+        close(sockfd_);
+        sockfd_ = -1;
+      }
+    }
+  }
+
+  freeaddrinfo(servinfo);  // all done with this structure
+
+  if (sockfd_ == -1) {
+    GTEST_LOG_(WARNING) << "stream_result_to: failed to connect to "
+                        << host_name_ << ":" << port_num_;
+  }
+}
+
+// End of class Streaming Listener
+#endif  // GTEST_CAN_STREAM_RESULTS__
+
+// Class ScopedTrace
+
+// Pushes the given source file location and message onto a per-thread
+// trace stack maintained by Google Test.
+ScopedTrace::ScopedTrace(const char* file, int line, const Message& message)
+    GTEST_LOCK_EXCLUDED_(&UnitTest::mutex_) {
+  TraceInfo trace;
+  trace.file = file;
+  trace.line = line;
+  trace.message = message.GetString();
+
+  UnitTest::GetInstance()->PushGTestTrace(trace);
+}
+
+// Pops the info pushed by the c'tor.
+ScopedTrace::~ScopedTrace()
+    GTEST_LOCK_EXCLUDED_(&UnitTest::mutex_) {
+  UnitTest::GetInstance()->PopGTestTrace();
+}
+
+
+// class OsStackTraceGetter
+
+// Returns the current OS stack trace as an std::string.  Parameters:
+//
+//   max_depth  - the maximum number of stack frames to be included
+//                in the trace.
+//   skip_count - the number of top frames to be skipped; doesn't count
+//                against max_depth.
+//
+string OsStackTraceGetter::CurrentStackTrace(int /* max_depth */,
+                                             int /* skip_count */)
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  return "";
+}
+
+void OsStackTraceGetter::UponLeavingGTest()
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+}
+
+const char* const
+OsStackTraceGetter::kElidedFramesMarker =
+    "... " GTEST_NAME_ " internal frames ...";
+
+// A helper class that creates the premature-exit file in its
+// constructor and deletes the file in its destructor.
+class ScopedPrematureExitFile {
+ public:
+  explicit ScopedPrematureExitFile(const char* premature_exit_filepath)
+      : premature_exit_filepath_(premature_exit_filepath) {
+    // If a path to the premature-exit file is specified...
+    if (premature_exit_filepath != NULL && *premature_exit_filepath != '\0') {
+      // create the file with a single "0" character in it.  I/O
+      // errors are ignored as there's nothing better we can do and we
+      // don't want to fail the test because of this.
+      FILE* pfile = posix::FOpen(premature_exit_filepath, "w");
+      fwrite("0", 1, 1, pfile);
+      fclose(pfile);
+    }
+  }
+
+  ~ScopedPrematureExitFile() {
+    if (premature_exit_filepath_ != NULL && *premature_exit_filepath_ != '\0') {
+      remove(premature_exit_filepath_);
+    }
+  }
+
+ private:
+  const char* const premature_exit_filepath_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ScopedPrematureExitFile);
+};
+
+}  // namespace internal
+
+// class TestEventListeners
+
+TestEventListeners::TestEventListeners()
+    : repeater_(new internal::TestEventRepeater()),
+      default_result_printer_(NULL),
+      default_xml_generator_(NULL) {
+}
+
+TestEventListeners::~TestEventListeners() { delete repeater_; }
+
+// Returns the standard listener responsible for the default console
+// output.  Can be removed from the listeners list to shut down default
+// console output.  Note that removing this object from the listener list
+// with Release transfers its ownership to the user.
+void TestEventListeners::Append(TestEventListener* listener) {
+  repeater_->Append(listener);
+}
+
+// Removes the given event listener from the list and returns it.  It then
+// becomes the caller's responsibility to delete the listener. Returns
+// NULL if the listener is not found in the list.
+TestEventListener* TestEventListeners::Release(TestEventListener* listener) {
+  if (listener == default_result_printer_)
+    default_result_printer_ = NULL;
+  else if (listener == default_xml_generator_)
+    default_xml_generator_ = NULL;
+  return repeater_->Release(listener);
+}
+
+// Returns repeater that broadcasts the TestEventListener events to all
+// subscribers.
+TestEventListener* TestEventListeners::repeater() { return repeater_; }
+
+// Sets the default_result_printer attribute to the provided listener.
+// The listener is also added to the listener list and previous
+// default_result_printer is removed from it and deleted. The listener can
+// also be NULL in which case it will not be added to the list. Does
+// nothing if the previous and the current listener objects are the same.
+void TestEventListeners::SetDefaultResultPrinter(TestEventListener* listener) {
+  if (default_result_printer_ != listener) {
+    // It is an error to pass this method a listener that is already in the
+    // list.
+    delete Release(default_result_printer_);
+    default_result_printer_ = listener;
+    if (listener != NULL)
+      Append(listener);
+  }
+}
+
+// Sets the default_xml_generator attribute to the provided listener.  The
+// listener is also added to the listener list and previous
+// default_xml_generator is removed from it and deleted. The listener can
+// also be NULL in which case it will not be added to the list. Does
+// nothing if the previous and the current listener objects are the same.
+void TestEventListeners::SetDefaultXmlGenerator(TestEventListener* listener) {
+  if (default_xml_generator_ != listener) {
+    // It is an error to pass this method a listener that is already in the
+    // list.
+    delete Release(default_xml_generator_);
+    default_xml_generator_ = listener;
+    if (listener != NULL)
+      Append(listener);
+  }
+}
+
+// Controls whether events will be forwarded by the repeater to the
+// listeners in the list.
+bool TestEventListeners::EventForwardingEnabled() const {
+  return repeater_->forwarding_enabled();
+}
+
+void TestEventListeners::SuppressEventForwarding() {
+  repeater_->set_forwarding_enabled(false);
+}
+
+// class UnitTest
+
+// Gets the singleton UnitTest object.  The first time this method is
+// called, a UnitTest object is constructed and returned.  Consecutive
+// calls will return the same object.
+//
+// We don't protect this under mutex_ as a user is not supposed to
+// call this before main() starts, from which point on the return
+// value will never change.
+UnitTest* UnitTest::GetInstance() {
+  // When compiled with MSVC 7.1 in optimized mode, destroying the
+  // UnitTest object upon exiting the program messes up the exit code,
+  // causing successful tests to appear failed.  We have to use a
+  // different implementation in this case to bypass the compiler bug.
+  // This implementation makes the compiler happy, at the cost of
+  // leaking the UnitTest object.
+
+  // CodeGear C++Builder insists on a public destructor for the
+  // default implementation.  Use this implementation to keep good OO
+  // design with private destructor.
+
+#if (_MSC_VER == 1310 && !defined(_DEBUG)) || defined(__BORLANDC__)
+  static UnitTest* const instance = new UnitTest;
+  return instance;
+#else
+  static UnitTest instance;
+  return &instance;
+#endif  // (_MSC_VER == 1310 && !defined(_DEBUG)) || defined(__BORLANDC__)
+}
+
+// Gets the number of successful test cases.
+int UnitTest::successful_test_case_count() const {
+  return impl()->successful_test_case_count();
+}
+
+// Gets the number of failed test cases.
+int UnitTest::failed_test_case_count() const {
+  return impl()->failed_test_case_count();
+}
+
+// Gets the number of all test cases.
+int UnitTest::total_test_case_count() const {
+  return impl()->total_test_case_count();
+}
+
+// Gets the number of all test cases that contain at least one test
+// that should run.
+int UnitTest::test_case_to_run_count() const {
+  return impl()->test_case_to_run_count();
+}
+
+// Gets the number of successful tests.
+int UnitTest::successful_test_count() const {
+  return impl()->successful_test_count();
+}
+
+// Gets the number of failed tests.
+int UnitTest::failed_test_count() const { return impl()->failed_test_count(); }
+
+// Gets the number of disabled tests that will be reported in the XML report.
+int UnitTest::reportable_disabled_test_count() const {
+  return impl()->reportable_disabled_test_count();
+}
+
+// Gets the number of disabled tests.
+int UnitTest::disabled_test_count() const {
+  return impl()->disabled_test_count();
+}
+
+// Gets the number of tests to be printed in the XML report.
+int UnitTest::reportable_test_count() const {
+  return impl()->reportable_test_count();
+}
+
+// Gets the number of all tests.
+int UnitTest::total_test_count() const { return impl()->total_test_count(); }
+
+// Gets the number of tests that should run.
+int UnitTest::test_to_run_count() const { return impl()->test_to_run_count(); }
+
+// Gets the time of the test program start, in ms from the start of the
+// UNIX epoch.
+internal::TimeInMillis UnitTest::start_timestamp() const {
+    return impl()->start_timestamp();
+}
+
+// Gets the elapsed time, in milliseconds.
+internal::TimeInMillis UnitTest::elapsed_time() const {
+  return impl()->elapsed_time();
+}
+
+// Returns true iff the unit test passed (i.e. all test cases passed).
+bool UnitTest::Passed() const { return impl()->Passed(); }
+
+// Returns true iff the unit test failed (i.e. some test case failed
+// or something outside of all tests failed).
+bool UnitTest::Failed() const { return impl()->Failed(); }
+
+// Gets the i-th test case among all the test cases. i can range from 0 to
+// total_test_case_count() - 1. If i is not in that range, returns NULL.
+const TestCase* UnitTest::GetTestCase(int i) const {
+  return impl()->GetTestCase(i);
+}
+
+// Returns the TestResult containing information on test failures and
+// properties logged outside of individual test cases.
+const TestResult& UnitTest::ad_hoc_test_result() const {
+  return *impl()->ad_hoc_test_result();
+}
+
+// Gets the i-th test case among all the test cases. i can range from 0 to
+// total_test_case_count() - 1. If i is not in that range, returns NULL.
+TestCase* UnitTest::GetMutableTestCase(int i) {
+  return impl()->GetMutableTestCase(i);
+}
+
+// Returns the list of event listeners that can be used to track events
+// inside Google Test.
+TestEventListeners& UnitTest::listeners() {
+  return *impl()->listeners();
+}
+
+// Registers and returns a global test environment.  When a test
+// program is run, all global test environments will be set-up in the
+// order they were registered.  After all tests in the program have
+// finished, all global test environments will be torn-down in the
+// *reverse* order they were registered.
+//
+// The UnitTest object takes ownership of the given environment.
+//
+// We don't protect this under mutex_, as we only support calling it
+// from the main thread.
+Environment* UnitTest::AddEnvironment(Environment* env) {
+  if (env == NULL) {
+    return NULL;
+  }
+
+  impl_->environments().push_back(env);
+  return env;
+}
+
+// Adds a TestPartResult to the current TestResult object.  All Google Test
+// assertion macros (e.g. ASSERT_TRUE, EXPECT_EQ, etc) eventually call
+// this to report their results.  The user code should use the
+// assertion macros instead of calling this directly.
+void UnitTest::AddTestPartResult(
+    TestPartResult::Type result_type,
+    const char* file_name,
+    int line_number,
+    const std::string& message,
+    const std::string& os_stack_trace) GTEST_LOCK_EXCLUDED_(mutex_) {
+  Message msg;
+  msg << message;
+
+  internal::MutexLock lock(&mutex_);
+  if (impl_->gtest_trace_stack().size() > 0) {
+    msg << "\n" << GTEST_NAME_ << " trace:";
+
+    for (int i = static_cast<int>(impl_->gtest_trace_stack().size());
+         i > 0; --i) {
+      const internal::TraceInfo& trace = impl_->gtest_trace_stack()[i - 1];
+      msg << "\n" << internal::FormatFileLocation(trace.file, trace.line)
+          << " " << trace.message;
+    }
+  }
+
+  if (os_stack_trace.c_str() != NULL && !os_stack_trace.empty()) {
+    msg << internal::kStackTraceMarker << os_stack_trace;
+  }
+
+  const TestPartResult result =
+    TestPartResult(result_type, file_name, line_number,
+                   msg.GetString().c_str());
+  impl_->GetTestPartResultReporterForCurrentThread()->
+      ReportTestPartResult(result);
+
+  if (result_type != TestPartResult::kSuccess) {
+    // gtest_break_on_failure takes precedence over
+    // gtest_throw_on_failure.  This allows a user to set the latter
+    // in the code (perhaps in order to use Google Test assertions
+    // with another testing framework) and specify the former on the
+    // command line for debugging.
+    if (GTEST_FLAG(break_on_failure)) {
+#if GTEST_OS_WINDOWS
+      // Using DebugBreak on Windows allows gtest to still break into a debugger
+      // when a failure happens and both the --gtest_break_on_failure and
+      // the --gtest_catch_exceptions flags are specified.
+      DebugBreak();
+#else
+      // Dereference NULL through a volatile pointer to prevent the compiler
+      // from removing. We use this rather than abort() or __builtin_trap() for
+      // portability: Symbian doesn't implement abort() well, and some debuggers
+      // don't correctly trap abort().
+      *static_cast<volatile int*>(NULL) = 1;
+#endif  // GTEST_OS_WINDOWS
+    } else if (GTEST_FLAG(throw_on_failure)) {
+#if GTEST_HAS_EXCEPTIONS
+      throw internal::GoogleTestFailureException(result);
+#else
+      // We cannot call abort() as it generates a pop-up in debug mode
+      // that cannot be suppressed in VC 7.1 or below.
+      exit(1);
+#endif
+    }
+  }
+}
+
+// Adds a TestProperty to the current TestResult object when invoked from
+// inside a test, to current TestCase's ad_hoc_test_result_ when invoked
+// from SetUpTestCase or TearDownTestCase, or to the global property set
+// when invoked elsewhere.  If the result already contains a property with
+// the same key, the value will be updated.
+void UnitTest::RecordProperty(const std::string& key,
+                              const std::string& value) {
+  impl_->RecordProperty(TestProperty(key, value));
+}
+
+// Runs all tests in this UnitTest object and prints the result.
+// Returns 0 if successful, or 1 otherwise.
+//
+// We don't protect this under mutex_, as we only support calling it
+// from the main thread.
+int UnitTest::Run() {
+  const bool in_death_test_child_process =
+      internal::GTEST_FLAG(internal_run_death_test).length() > 0;
+
+  // Google Test implements this protocol for catching that a test
+  // program exits before returning control to Google Test:
+  //
+  //   1. Upon start, Google Test creates a file whose absolute path
+  //      is specified by the environment variable
+  //      TEST_PREMATURE_EXIT_FILE.
+  //   2. When Google Test has finished its work, it deletes the file.
+  //
+  // This allows a test runner to set TEST_PREMATURE_EXIT_FILE before
+  // running a Google-Test-based test program and check the existence
+  // of the file at the end of the test execution to see if it has
+  // exited prematurely.
+
+  // If we are in the child process of a death test, don't
+  // create/delete the premature exit file, as doing so is unnecessary
+  // and will confuse the parent process.  Otherwise, create/delete
+  // the file upon entering/leaving this function.  If the program
+  // somehow exits before this function has a chance to return, the
+  // premature-exit file will be left undeleted, causing a test runner
+  // that understands the premature-exit-file protocol to report the
+  // test as having failed.
+  const internal::ScopedPrematureExitFile premature_exit_file(
+      in_death_test_child_process ?
+      NULL : internal::posix::GetEnv("TEST_PREMATURE_EXIT_FILE"));
+
+  // Captures the value of GTEST_FLAG(catch_exceptions).  This value will be
+  // used for the duration of the program.
+  impl()->set_catch_exceptions(GTEST_FLAG(catch_exceptions));
+
+#if GTEST_HAS_SEH
+  // Either the user wants Google Test to catch exceptions thrown by the
+  // tests or this is executing in the context of death test child
+  // process. In either case the user does not want to see pop-up dialogs
+  // about crashes - they are expected.
+  if (impl()->catch_exceptions() || in_death_test_child_process) {
+# if !GTEST_OS_WINDOWS_MOBILE
+    // SetErrorMode doesn't exist on CE.
+    SetErrorMode(SEM_FAILCRITICALERRORS | SEM_NOALIGNMENTFAULTEXCEPT |
+                 SEM_NOGPFAULTERRORBOX | SEM_NOOPENFILEERRORBOX);
+# endif  // !GTEST_OS_WINDOWS_MOBILE
+
+# if (defined(_MSC_VER) || GTEST_OS_WINDOWS_MINGW) && !GTEST_OS_WINDOWS_MOBILE
+    // Death test children can be terminated with _abort().  On Windows,
+    // _abort() can show a dialog with a warning message.  This forces the
+    // abort message to go to stderr instead.
+    _set_error_mode(_OUT_TO_STDERR);
+# endif
+
+# if _MSC_VER >= 1400 && !GTEST_OS_WINDOWS_MOBILE
+    // In the debug version, Visual Studio pops up a separate dialog
+    // offering a choice to debug the aborted program. We need to suppress
+    // this dialog or it will pop up for every EXPECT/ASSERT_DEATH statement
+    // executed. Google Test will notify the user of any unexpected
+    // failure via stderr.
+    //
+    // VC++ doesn't define _set_abort_behavior() prior to the version 8.0.
+    // Users of prior VC versions shall suffer the agony and pain of
+    // clicking through the countless debug dialogs.
+    // TODO(vladl@google.com): find a way to suppress the abort dialog() in the
+    // debug mode when compiled with VC 7.1 or lower.
+    if (!GTEST_FLAG(break_on_failure))
+      _set_abort_behavior(
+          0x0,                                    // Clear the following flags:
+          _WRITE_ABORT_MSG | _CALL_REPORTFAULT);  // pop-up window, core dump.
+# endif
+  }
+#endif  // GTEST_HAS_SEH
+
+  return internal::HandleExceptionsInMethodIfSupported(
+      impl(),
+      &internal::UnitTestImpl::RunAllTests,
+      "auxiliary test code (environments or event listeners)") ? 0 : 1;
+}
+
+// Returns the working directory when the first TEST() or TEST_F() was
+// executed.
+const char* UnitTest::original_working_dir() const {
+  return impl_->original_working_dir_.c_str();
+}
+
+// Returns the TestCase object for the test that's currently running,
+// or NULL if no test is running.
+const TestCase* UnitTest::current_test_case() const
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  return impl_->current_test_case();
+}
+
+// Returns the TestInfo object for the test that's currently running,
+// or NULL if no test is running.
+const TestInfo* UnitTest::current_test_info() const
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  return impl_->current_test_info();
+}
+
+// Returns the random seed used at the start of the current test run.
+int UnitTest::random_seed() const { return impl_->random_seed(); }
+
+#if GTEST_HAS_PARAM_TEST
+// Returns ParameterizedTestCaseRegistry object used to keep track of
+// value-parameterized tests and instantiate and register them.
+internal::ParameterizedTestCaseRegistry&
+    UnitTest::parameterized_test_registry()
+        GTEST_LOCK_EXCLUDED_(mutex_) {
+  return impl_->parameterized_test_registry();
+}
+#endif  // GTEST_HAS_PARAM_TEST
+
+// Creates an empty UnitTest.
+UnitTest::UnitTest() {
+  impl_ = new internal::UnitTestImpl(this);
+}
+
+// Destructor of UnitTest.
+UnitTest::~UnitTest() {
+  delete impl_;
+}
+
+// Pushes a trace defined by SCOPED_TRACE() on to the per-thread
+// Google Test trace stack.
+void UnitTest::PushGTestTrace(const internal::TraceInfo& trace)
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  impl_->gtest_trace_stack().push_back(trace);
+}
+
+// Pops a trace from the per-thread Google Test trace stack.
+void UnitTest::PopGTestTrace()
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  impl_->gtest_trace_stack().pop_back();
+}
+
+namespace internal {
+
+UnitTestImpl::UnitTestImpl(UnitTest* parent)
+    : parent_(parent),
+#ifdef _MSC_VER
+# pragma warning(push)                    // Saves the current warning state.
+# pragma warning(disable:4355)            // Temporarily disables warning 4355
+                                         // (using this in initializer).
+      default_global_test_part_result_reporter_(this),
+      default_per_thread_test_part_result_reporter_(this),
+# pragma warning(pop)                     // Restores the warning state again.
+#else
+      default_global_test_part_result_reporter_(this),
+      default_per_thread_test_part_result_reporter_(this),
+#endif  // _MSC_VER
+      global_test_part_result_repoter_(
+          &default_global_test_part_result_reporter_),
+      per_thread_test_part_result_reporter_(
+          &default_per_thread_test_part_result_reporter_),
+#if GTEST_HAS_PARAM_TEST
+      parameterized_test_registry_(),
+      parameterized_tests_registered_(false),
+#endif  // GTEST_HAS_PARAM_TEST
+      last_death_test_case_(-1),
+      current_test_case_(NULL),
+      current_test_info_(NULL),
+      ad_hoc_test_result_(),
+      os_stack_trace_getter_(NULL),
+      post_flag_parse_init_performed_(false),
+      random_seed_(0),  // Will be overridden by the flag before first use.
+      random_(0),  // Will be reseeded before first use.
+      start_timestamp_(0),
+      elapsed_time_(0),
+#if GTEST_HAS_DEATH_TEST
+      death_test_factory_(new DefaultDeathTestFactory),
+#endif
+      // Will be overridden by the flag before first use.
+      catch_exceptions_(false) {
+  listeners()->SetDefaultResultPrinter(new PrettyUnitTestResultPrinter);
+}
+
+UnitTestImpl::~UnitTestImpl() {
+  // Deletes every TestCase.
+  ForEach(test_cases_, internal::Delete<TestCase>);
+
+  // Deletes every Environment.
+  ForEach(environments_, internal::Delete<Environment>);
+
+  delete os_stack_trace_getter_;
+}
+
+// Adds a TestProperty to the current TestResult object when invoked in a
+// context of a test, to current test case's ad_hoc_test_result when invoke
+// from SetUpTestCase/TearDownTestCase, or to the global property set
+// otherwise.  If the result already contains a property with the same key,
+// the value will be updated.
+void UnitTestImpl::RecordProperty(const TestProperty& test_property) {
+  std::string xml_element;
+  TestResult* test_result;  // TestResult appropriate for property recording.
+
+  if (current_test_info_ != NULL) {
+    xml_element = "testcase";
+    test_result = &(current_test_info_->result_);
+  } else if (current_test_case_ != NULL) {
+    xml_element = "testsuite";
+    test_result = &(current_test_case_->ad_hoc_test_result_);
+  } else {
+    xml_element = "testsuites";
+    test_result = &ad_hoc_test_result_;
+  }
+  test_result->RecordProperty(xml_element, test_property);
+}
+
+#if GTEST_HAS_DEATH_TEST
+// Disables event forwarding if the control is currently in a death test
+// subprocess. Must not be called before InitGoogleTest.
+void UnitTestImpl::SuppressTestEventsIfInSubprocess() {
+  if (internal_run_death_test_flag_.get() != NULL)
+    listeners()->SuppressEventForwarding();
+}
+#endif  // GTEST_HAS_DEATH_TEST
+
+// Initializes event listeners performing XML output as specified by
+// UnitTestOptions. Must not be called before InitGoogleTest.
+void UnitTestImpl::ConfigureXmlOutput() {
+  const std::string& output_format = UnitTestOptions::GetOutputFormat();
+  if (output_format == "xml") {
+    listeners()->SetDefaultXmlGenerator(new XmlUnitTestResultPrinter(
+        UnitTestOptions::GetAbsolutePathToOutputFile().c_str()));
+  } else if (output_format != "") {
+    printf("WARNING: unrecognized output format \"%s\" ignored.\n",
+           output_format.c_str());
+    fflush(stdout);
+  }
+}
+
+#if GTEST_CAN_STREAM_RESULTS_
+// Initializes event listeners for streaming test results in string form.
+// Must not be called before InitGoogleTest.
+void UnitTestImpl::ConfigureStreamingOutput() {
+  const std::string& target = GTEST_FLAG(stream_result_to);
+  if (!target.empty()) {
+    const size_t pos = target.find(':');
+    if (pos != std::string::npos) {
+      listeners()->Append(new StreamingListener(target.substr(0, pos),
+                                                target.substr(pos+1)));
+    } else {
+      printf("WARNING: unrecognized streaming target \"%s\" ignored.\n",
+             target.c_str());
+      fflush(stdout);
+    }
+  }
+}
+#endif  // GTEST_CAN_STREAM_RESULTS_
+
+// Performs initialization dependent upon flag values obtained in
+// ParseGoogleTestFlagsOnly.  Is called from InitGoogleTest after the call to
+// ParseGoogleTestFlagsOnly.  In case a user neglects to call InitGoogleTest
+// this function is also called from RunAllTests.  Since this function can be
+// called more than once, it has to be idempotent.
+void UnitTestImpl::PostFlagParsingInit() {
+  // Ensures that this function does not execute more than once.
+  if (!post_flag_parse_init_performed_) {
+    post_flag_parse_init_performed_ = true;
+
+#if GTEST_HAS_DEATH_TEST
+    InitDeathTestSubprocessControlInfo();
+    SuppressTestEventsIfInSubprocess();
+#endif  // GTEST_HAS_DEATH_TEST
+
+    // Registers parameterized tests. This makes parameterized tests
+    // available to the UnitTest reflection API without running
+    // RUN_ALL_TESTS.
+    RegisterParameterizedTests();
+
+    // Configures listeners for XML output. This makes it possible for users
+    // to shut down the default XML output before invoking RUN_ALL_TESTS.
+    ConfigureXmlOutput();
+
+#if GTEST_CAN_STREAM_RESULTS_
+    // Configures listeners for streaming test results to the specified server.
+    ConfigureStreamingOutput();
+#endif  // GTEST_CAN_STREAM_RESULTS_
+  }
+}
+
+// A predicate that checks the name of a TestCase against a known
+// value.
+//
+// This is used for implementation of the UnitTest class only.  We put
+// it in the anonymous namespace to prevent polluting the outer
+// namespace.
+//
+// TestCaseNameIs is copyable.
+class TestCaseNameIs {
+ public:
+  // Constructor.
+  explicit TestCaseNameIs(const std::string& name)
+      : name_(name) {}
+
+  // Returns true iff the name of test_case matches name_.
+  bool operator()(const TestCase* test_case) const {
+    return test_case != NULL && strcmp(test_case->name(), name_.c_str()) == 0;
+  }
+
+ private:
+  std::string name_;
+};
+
+// Finds and returns a TestCase with the given name.  If one doesn't
+// exist, creates one and returns it.  It's the CALLER'S
+// RESPONSIBILITY to ensure that this function is only called WHEN THE
+// TESTS ARE NOT SHUFFLED.
+//
+// Arguments:
+//
+//   test_case_name: name of the test case
+//   type_param:     the name of the test case's type parameter, or NULL if
+//                   this is not a typed or a type-parameterized test case.
+//   set_up_tc:      pointer to the function that sets up the test case
+//   tear_down_tc:   pointer to the function that tears down the test case
+TestCase* UnitTestImpl::GetTestCase(const char* test_case_name,
+                                    const char* type_param,
+                                    Test::SetUpTestCaseFunc set_up_tc,
+                                    Test::TearDownTestCaseFunc tear_down_tc) {
+  // Can we find a TestCase with the given name?
+  const std::vector<TestCase*>::const_iterator test_case =
+      std::find_if(test_cases_.begin(), test_cases_.end(),
+                   TestCaseNameIs(test_case_name));
+
+  if (test_case != test_cases_.end())
+    return *test_case;
+
+  // No.  Let's create one.
+  TestCase* const new_test_case =
+      new TestCase(test_case_name, type_param, set_up_tc, tear_down_tc);
+
+  // Is this a death test case?
+  if (internal::UnitTestOptions::MatchesFilter(test_case_name,
+                                               kDeathTestCaseFilter)) {
+    // Yes.  Inserts the test case after the last death test case
+    // defined so far.  This only works when the test cases haven't
+    // been shuffled.  Otherwise we may end up running a death test
+    // after a non-death test.
+    ++last_death_test_case_;
+    test_cases_.insert(test_cases_.begin() + last_death_test_case_,
+                       new_test_case);
+  } else {
+    // No.  Appends to the end of the list.
+    test_cases_.push_back(new_test_case);
+  }
+
+  test_case_indices_.push_back(static_cast<int>(test_case_indices_.size()));
+  return new_test_case;
+}
+
+// Helpers for setting up / tearing down the given environment.  They
+// are for use in the ForEach() function.
+static void SetUpEnvironment(Environment* env) { env->SetUp(); }
+static void TearDownEnvironment(Environment* env) { env->TearDown(); }
+
+// Runs all tests in this UnitTest object, prints the result, and
+// returns true if all tests are successful.  If any exception is
+// thrown during a test, the test is considered to be failed, but the
+// rest of the tests will still be run.
+//
+// When parameterized tests are enabled, it expands and registers
+// parameterized tests first in RegisterParameterizedTests().
+// All other functions called from RunAllTests() may safely assume that
+// parameterized tests are ready to be counted and run.
+bool UnitTestImpl::RunAllTests() {
+  // Makes sure InitGoogleTest() was called.
+  if (!GTestIsInitialized()) {
+    printf("%s",
+           "\nThis test program did NOT call ::testing::InitGoogleTest "
+           "before calling RUN_ALL_TESTS().  Please fix it.\n");
+    return false;
+  }
+
+  // Do not run any test if the --help flag was specified.
+  if (g_help_flag)
+    return true;
+
+  // Repeats the call to the post-flag parsing initialization in case the
+  // user didn't call InitGoogleTest.
+  PostFlagParsingInit();
+
+  // Even if sharding is not on, test runners may want to use the
+  // GTEST_SHARD_STATUS_FILE to query whether the test supports the sharding
+  // protocol.
+  internal::WriteToShardStatusFileIfNeeded();
+
+  // True iff we are in a subprocess for running a thread-safe-style
+  // death test.
+  bool in_subprocess_for_death_test = false;
+
+#if GTEST_HAS_DEATH_TEST
+  in_subprocess_for_death_test = (internal_run_death_test_flag_.get() != NULL);
+#endif  // GTEST_HAS_DEATH_TEST
+
+  const bool should_shard = ShouldShard(kTestTotalShards, kTestShardIndex,
+                                        in_subprocess_for_death_test);
+
+  // Compares the full test names with the filter to decide which
+  // tests to run.
+  const bool has_tests_to_run = FilterTests(should_shard
+                                              ? HONOR_SHARDING_PROTOCOL
+                                              : IGNORE_SHARDING_PROTOCOL) > 0;
+
+  // Lists the tests and exits if the --gtest_list_tests flag was specified.
+  if (GTEST_FLAG(list_tests)) {
+    // This must be called *after* FilterTests() has been called.
+    ListTestsMatchingFilter();
+    return true;
+  }
+
+  random_seed_ = GTEST_FLAG(shuffle) ?
+      GetRandomSeedFromFlag(GTEST_FLAG(random_seed)) : 0;
+
+  // True iff at least one test has failed.
+  bool failed = false;
+
+  TestEventListener* repeater = listeners()->repeater();
+
+  start_timestamp_ = GetTimeInMillis();
+  repeater->OnTestProgramStart(*parent_);
+
+  // How many times to repeat the tests?  We don't want to repeat them
+  // when we are inside the subprocess of a death test.
+  const int repeat = in_subprocess_for_death_test ? 1 : GTEST_FLAG(repeat);
+  // Repeats forever if the repeat count is negative.
+  const bool forever = repeat < 0;
+  for (int i = 0; forever || i != repeat; i++) {
+    // We want to preserve failures generated by ad-hoc test
+    // assertions executed before RUN_ALL_TESTS().
+    ClearNonAdHocTestResult();
+
+    const TimeInMillis start = GetTimeInMillis();
+
+    // Shuffles test cases and tests if requested.
+    if (has_tests_to_run && GTEST_FLAG(shuffle)) {
+      random()->Reseed(random_seed_);
+      // This should be done before calling OnTestIterationStart(),
+      // such that a test event listener can see the actual test order
+      // in the event.
+      ShuffleTests();
+    }
+
+    // Tells the unit test event listeners that the tests are about to start.
+    repeater->OnTestIterationStart(*parent_, i);
+
+    // Runs each test case if there is at least one test to run.
+    if (has_tests_to_run) {
+      // Sets up all environments beforehand.
+      repeater->OnEnvironmentsSetUpStart(*parent_);
+      ForEach(environments_, SetUpEnvironment);
+      repeater->OnEnvironmentsSetUpEnd(*parent_);
+
+      // Runs the tests only if there was no fatal failure during global
+      // set-up.
+      if (!Test::HasFatalFailure()) {
+        for (int test_index = 0; test_index < total_test_case_count();
+             test_index++) {
+          GetMutableTestCase(test_index)->Run();
+        }
+      }
+
+      // Tears down all environments in reverse order afterwards.
+      repeater->OnEnvironmentsTearDownStart(*parent_);
+      std::for_each(environments_.rbegin(), environments_.rend(),
+                    TearDownEnvironment);
+      repeater->OnEnvironmentsTearDownEnd(*parent_);
+    }
+
+    elapsed_time_ = GetTimeInMillis() - start;
+
+    // Tells the unit test event listener that the tests have just finished.
+    repeater->OnTestIterationEnd(*parent_, i);
+
+    // Gets the result and clears it.
+    if (!Passed()) {
+      failed = true;
+    }
+
+    // Restores the original test order after the iteration.  This
+    // allows the user to quickly repro a failure that happens in the
+    // N-th iteration without repeating the first (N - 1) iterations.
+    // This is not enclosed in "if (GTEST_FLAG(shuffle)) { ... }", in
+    // case the user somehow changes the value of the flag somewhere
+    // (it's always safe to unshuffle the tests).
+    UnshuffleTests();
+
+    if (GTEST_FLAG(shuffle)) {
+      // Picks a new random seed for each iteration.
+      random_seed_ = GetNextRandomSeed(random_seed_);
+    }
+  }
+
+  repeater->OnTestProgramEnd(*parent_);
+
+  return !failed;
+}
+
+// Reads the GTEST_SHARD_STATUS_FILE environment variable, and creates the file
+// if the variable is present. If a file already exists at this location, this
+// function will write over it. If the variable is present, but the file cannot
+// be created, prints an error and exits.
+void WriteToShardStatusFileIfNeeded() {
+  const char* const test_shard_file = posix::GetEnv(kTestShardStatusFile);
+  if (test_shard_file != NULL) {
+    FILE* const file = posix::FOpen(test_shard_file, "w");
+    if (file == NULL) {
+      ColoredPrintf(COLOR_RED,
+                    "Could not write to the test shard status file \"%s\" "
+                    "specified by the %s environment variable.\n",
+                    test_shard_file, kTestShardStatusFile);
+      fflush(stdout);
+      exit(EXIT_FAILURE);
+    }
+    fclose(file);
+  }
+}
+
+// Checks whether sharding is enabled by examining the relevant
+// environment variable values. If the variables are present,
+// but inconsistent (i.e., shard_index >= total_shards), prints
+// an error and exits. If in_subprocess_for_death_test, sharding is
+// disabled because it must only be applied to the original test
+// process. Otherwise, we could filter out death tests we intended to execute.
+bool ShouldShard(const char* total_shards_env,
+                 const char* shard_index_env,
+                 bool in_subprocess_for_death_test) {
+  if (in_subprocess_for_death_test) {
+    return false;
+  }
+
+  const Int32 total_shards = Int32FromEnvOrDie(total_shards_env, -1);
+  const Int32 shard_index = Int32FromEnvOrDie(shard_index_env, -1);
+
+  if (total_shards == -1 && shard_index == -1) {
+    return false;
+  } else if (total_shards == -1 && shard_index != -1) {
+    const Message msg = Message()
+      << "Invalid environment variables: you have "
+      << kTestShardIndex << " = " << shard_index
+      << ", but have left " << kTestTotalShards << " unset.\n";
+    ColoredPrintf(COLOR_RED, msg.GetString().c_str());
+    fflush(stdout);
+    exit(EXIT_FAILURE);
+  } else if (total_shards != -1 && shard_index == -1) {
+    const Message msg = Message()
+      << "Invalid environment variables: you have "
+      << kTestTotalShards << " = " << total_shards
+      << ", but have left " << kTestShardIndex << " unset.\n";
+    ColoredPrintf(COLOR_RED, msg.GetString().c_str());
+    fflush(stdout);
+    exit(EXIT_FAILURE);
+  } else if (shard_index < 0 || shard_index >= total_shards) {
+    const Message msg = Message()
+      << "Invalid environment variables: we require 0 <= "
+      << kTestShardIndex << " < " << kTestTotalShards
+      << ", but you have " << kTestShardIndex << "=" << shard_index
+      << ", " << kTestTotalShards << "=" << total_shards << ".\n";
+    ColoredPrintf(COLOR_RED, msg.GetString().c_str());
+    fflush(stdout);
+    exit(EXIT_FAILURE);
+  }
+
+  return total_shards > 1;
+}
+
+// Parses the environment variable var as an Int32. If it is unset,
+// returns default_val. If it is not an Int32, prints an error
+// and aborts.
+Int32 Int32FromEnvOrDie(const char* var, Int32 default_val) {
+  const char* str_val = posix::GetEnv(var);
+  if (str_val == NULL) {
+    return default_val;
+  }
+
+  Int32 result;
+  if (!ParseInt32(Message() << "The value of environment variable " << var,
+                  str_val, &result)) {
+    exit(EXIT_FAILURE);
+  }
+  return result;
+}
+
+// Given the total number of shards, the shard index, and the test id,
+// returns true iff the test should be run on this shard. The test id is
+// some arbitrary but unique non-negative integer assigned to each test
+// method. Assumes that 0 <= shard_index < total_shards.
+bool ShouldRunTestOnShard(int total_shards, int shard_index, int test_id) {
+  return (test_id % total_shards) == shard_index;
+}
+
+// Compares the name of each test with the user-specified filter to
+// decide whether the test should be run, then records the result in
+// each TestCase and TestInfo object.
+// If shard_tests == true, further filters tests based on sharding
+// variables in the environment - see
+// http://code.google.com/p/googletest/wiki/GoogleTestAdvancedGuide.
+// Returns the number of tests that should run.
+int UnitTestImpl::FilterTests(ReactionToSharding shard_tests) {
+  const Int32 total_shards = shard_tests == HONOR_SHARDING_PROTOCOL ?
+      Int32FromEnvOrDie(kTestTotalShards, -1) : -1;
+  const Int32 shard_index = shard_tests == HONOR_SHARDING_PROTOCOL ?
+      Int32FromEnvOrDie(kTestShardIndex, -1) : -1;
+
+  // num_runnable_tests are the number of tests that will
+  // run across all shards (i.e., match filter and are not disabled).
+  // num_selected_tests are the number of tests to be run on
+  // this shard.
+  int num_runnable_tests = 0;
+  int num_selected_tests = 0;
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    TestCase* const test_case = test_cases_[i];
+    const std::string &test_case_name = test_case->name();
+    test_case->set_should_run(false);
+
+    for (size_t j = 0; j < test_case->test_info_list().size(); j++) {
+      TestInfo* const test_info = test_case->test_info_list()[j];
+      const std::string test_name(test_info->name());
+      // A test is disabled if test case name or test name matches
+      // kDisableTestFilter.
+      const bool is_disabled =
+          internal::UnitTestOptions::MatchesFilter(test_case_name,
+                                                   kDisableTestFilter) ||
+          internal::UnitTestOptions::MatchesFilter(test_name,
+                                                   kDisableTestFilter);
+      test_info->is_disabled_ = is_disabled;
+
+      const bool matches_filter =
+          internal::UnitTestOptions::FilterMatchesTest(test_case_name,
+                                                       test_name);
+      test_info->matches_filter_ = matches_filter;
+
+      const bool is_runnable =
+          (GTEST_FLAG(also_run_disabled_tests) || !is_disabled) &&
+          matches_filter;
+
+      const bool is_selected = is_runnable &&
+          (shard_tests == IGNORE_SHARDING_PROTOCOL ||
+           ShouldRunTestOnShard(total_shards, shard_index,
+                                num_runnable_tests));
+
+      num_runnable_tests += is_runnable;
+      num_selected_tests += is_selected;
+
+      test_info->should_run_ = is_selected;
+      test_case->set_should_run(test_case->should_run() || is_selected);
+    }
+  }
+  return num_selected_tests;
+}
+
+// Prints the given C-string on a single line by replacing all '\n'
+// characters with string "\\n".  If the output takes more than
+// max_length characters, only prints the first max_length characters
+// and "...".
+static void PrintOnOneLine(const char* str, int max_length) {
+  if (str != NULL) {
+    for (int i = 0; *str != '\0'; ++str) {
+      if (i >= max_length) {
+        printf("...");
+        break;
+      }
+      if (*str == '\n') {
+        printf("\\n");
+        i += 2;
+      } else {
+        printf("%c", *str);
+        ++i;
+      }
+    }
+  }
+}
+
+// Prints the names of the tests matching the user-specified filter flag.
+void UnitTestImpl::ListTestsMatchingFilter() {
+  // Print at most this many characters for each type/value parameter.
+  const int kMaxParamLength = 250;
+
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    const TestCase* const test_case = test_cases_[i];
+    bool printed_test_case_name = false;
+
+    for (size_t j = 0; j < test_case->test_info_list().size(); j++) {
+      const TestInfo* const test_info =
+          test_case->test_info_list()[j];
+      if (test_info->matches_filter_) {
+        if (!printed_test_case_name) {
+          printed_test_case_name = true;
+          printf("%s.", test_case->name());
+          if (test_case->type_param() != NULL) {
+            printf("  # %s = ", kTypeParamLabel);
+            // We print the type parameter on a single line to make
+            // the output easy to parse by a program.
+            PrintOnOneLine(test_case->type_param(), kMaxParamLength);
+          }
+          printf("\n");
+        }
+        printf("  %s", test_info->name());
+        if (test_info->value_param() != NULL) {
+          printf("  # %s = ", kValueParamLabel);
+          // We print the value parameter on a single line to make the
+          // output easy to parse by a program.
+          PrintOnOneLine(test_info->value_param(), kMaxParamLength);
+        }
+        printf("\n");
+      }
+    }
+  }
+  fflush(stdout);
+}
+
+// Sets the OS stack trace getter.
+//
+// Does nothing if the input and the current OS stack trace getter are
+// the same; otherwise, deletes the old getter and makes the input the
+// current getter.
+void UnitTestImpl::set_os_stack_trace_getter(
+    OsStackTraceGetterInterface* getter) {
+  if (os_stack_trace_getter_ != getter) {
+    delete os_stack_trace_getter_;
+    os_stack_trace_getter_ = getter;
+  }
+}
+
+// Returns the current OS stack trace getter if it is not NULL;
+// otherwise, creates an OsStackTraceGetter, makes it the current
+// getter, and returns it.
+OsStackTraceGetterInterface* UnitTestImpl::os_stack_trace_getter() {
+  if (os_stack_trace_getter_ == NULL) {
+    os_stack_trace_getter_ = new OsStackTraceGetter;
+  }
+
+  return os_stack_trace_getter_;
+}
+
+// Returns the TestResult for the test that's currently running, or
+// the TestResult for the ad hoc test if no test is running.
+TestResult* UnitTestImpl::current_test_result() {
+  return current_test_info_ ?
+      &(current_test_info_->result_) : &ad_hoc_test_result_;
+}
+
+// Shuffles all test cases, and the tests within each test case,
+// making sure that death tests are still run first.
+void UnitTestImpl::ShuffleTests() {
+  // Shuffles the death test cases.
+  ShuffleRange(random(), 0, last_death_test_case_ + 1, &test_case_indices_);
+
+  // Shuffles the non-death test cases.
+  ShuffleRange(random(), last_death_test_case_ + 1,
+               static_cast<int>(test_cases_.size()), &test_case_indices_);
+
+  // Shuffles the tests inside each test case.
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    test_cases_[i]->ShuffleTests(random());
+  }
+}
+
+// Restores the test cases and tests to their order before the first shuffle.
+void UnitTestImpl::UnshuffleTests() {
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    // Unshuffles the tests in each test case.
+    test_cases_[i]->UnshuffleTests();
+    // Resets the index of each test case.
+    test_case_indices_[i] = static_cast<int>(i);
+  }
+}
+
+// Returns the current OS stack trace as an std::string.
+//
+// The maximum number of stack frames to be included is specified by
+// the gtest_stack_trace_depth flag.  The skip_count parameter
+// specifies the number of top frames to be skipped, which doesn't
+// count against the number of frames to be included.
+//
+// For example, if Foo() calls Bar(), which in turn calls
+// GetCurrentOsStackTraceExceptTop(..., 1), Foo() will be included in
+// the trace but Bar() and GetCurrentOsStackTraceExceptTop() won't.
+std::string GetCurrentOsStackTraceExceptTop(UnitTest* /*unit_test*/,
+                                            int skip_count) {
+  // We pass skip_count + 1 to skip this wrapper function in addition
+  // to what the user really wants to skip.
+  return GetUnitTestImpl()->CurrentOsStackTraceExceptTop(skip_count + 1);
+}
+
+// Used by the GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_ macro to
+// suppress unreachable code warnings.
+namespace {
+class ClassUniqueToAlwaysTrue {};
+}
+
+bool IsTrue(bool condition) { return condition; }
+
+bool AlwaysTrue() {
+#if GTEST_HAS_EXCEPTIONS
+  // This condition is always false so AlwaysTrue() never actually throws,
+  // but it makes the compiler think that it may throw.
+  if (IsTrue(false))
+    throw ClassUniqueToAlwaysTrue();
+#endif  // GTEST_HAS_EXCEPTIONS
+  return true;
+}
+
+// If *pstr starts with the given prefix, modifies *pstr to be right
+// past the prefix and returns true; otherwise leaves *pstr unchanged
+// and returns false.  None of pstr, *pstr, and prefix can be NULL.
+bool SkipPrefix(const char* prefix, const char** pstr) {
+  const size_t prefix_len = strlen(prefix);
+  if (strncmp(*pstr, prefix, prefix_len) == 0) {
+    *pstr += prefix_len;
+    return true;
+  }
+  return false;
+}
+
+// Parses a string as a command line flag.  The string should have
+// the format "--flag=value".  When def_optional is true, the "=value"
+// part can be omitted.
+//
+// Returns the value of the flag, or NULL if the parsing failed.
+const char* ParseFlagValue(const char* str,
+                           const char* flag,
+                           bool def_optional) {
+  // str and flag must not be NULL.
+  if (str == NULL || flag == NULL) return NULL;
+
+  // The flag must start with "--" followed by GTEST_FLAG_PREFIX_.
+  const std::string flag_str = std::string("--") + GTEST_FLAG_PREFIX_ + flag;
+  const size_t flag_len = flag_str.length();
+  if (strncmp(str, flag_str.c_str(), flag_len) != 0) return NULL;
+
+  // Skips the flag name.
+  const char* flag_end = str + flag_len;
+
+  // When def_optional is true, it's OK to not have a "=value" part.
+  if (def_optional && (flag_end[0] == '\0')) {
+    return flag_end;
+  }
+
+  // If def_optional is true and there are more characters after the
+  // flag name, or if def_optional is false, there must be a '=' after
+  // the flag name.
+  if (flag_end[0] != '=') return NULL;
+
+  // Returns the string after "=".
+  return flag_end + 1;
+}
+
+// Parses a string for a bool flag, in the form of either
+// "--flag=value" or "--flag".
+//
+// In the former case, the value is taken as true as long as it does
+// not start with '0', 'f', or 'F'.
+//
+// In the latter case, the value is taken as true.
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseBoolFlag(const char* str, const char* flag, bool* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, true);
+
+  // Aborts if the parsing failed.
+  if (value_str == NULL) return false;
+
+  // Converts the string value to a bool.
+  *value = !(*value_str == '0' || *value_str == 'f' || *value_str == 'F');
+  return true;
+}
+
+// Parses a string for an Int32 flag, in the form of
+// "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseInt32Flag(const char* str, const char* flag, Int32* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, false);
+
+  // Aborts if the parsing failed.
+  if (value_str == NULL) return false;
+
+  // Sets *value to the value of the flag.
+  return ParseInt32(Message() << "The value of flag --" << flag,
+                    value_str, value);
+}
+
+// Parses a string for a string flag, in the form of
+// "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseStringFlag(const char* str, const char* flag, std::string* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, false);
+
+  // Aborts if the parsing failed.
+  if (value_str == NULL) return false;
+
+  // Sets *value to the value of the flag.
+  *value = value_str;
+  return true;
+}
+
+// Determines whether a string has a prefix that Google Test uses for its
+// flags, i.e., starts with GTEST_FLAG_PREFIX_ or GTEST_FLAG_PREFIX_DASH_.
+// If Google Test detects that a command line flag has its prefix but is not
+// recognized, it will print its help message. Flags starting with
+// GTEST_INTERNAL_PREFIX_ followed by "internal_" are considered Google Test
+// internal flags and do not trigger the help message.
+static bool HasGoogleTestFlagPrefix(const char* str) {
+  return (SkipPrefix("--", &str) ||
+          SkipPrefix("-", &str) ||
+          SkipPrefix("/", &str)) &&
+         !SkipPrefix(GTEST_FLAG_PREFIX_ "internal_", &str) &&
+         (SkipPrefix(GTEST_FLAG_PREFIX_, &str) ||
+          SkipPrefix(GTEST_FLAG_PREFIX_DASH_, &str));
+}
+
+// Prints a string containing code-encoded text.  The following escape
+// sequences can be used in the string to control the text color:
+//
+//   @@    prints a single '@' character.
+//   @R    changes the color to red.
+//   @G    changes the color to green.
+//   @Y    changes the color to yellow.
+//   @D    changes to the default terminal text color.
+//
+// TODO(wan@google.com): Write tests for this once we add stdout
+// capturing to Google Test.
+static void PrintColorEncoded(const char* str) {
+  GTestColor color = COLOR_DEFAULT;  // The current color.
+
+  // Conceptually, we split the string into segments divided by escape
+  // sequences.  Then we print one segment at a time.  At the end of
+  // each iteration, the str pointer advances to the beginning of the
+  // next segment.
+  for (;;) {
+    const char* p = strchr(str, '@');
+    if (p == NULL) {
+      ColoredPrintf(color, "%s", str);
+      return;
+    }
+
+    ColoredPrintf(color, "%s", std::string(str, p).c_str());
+
+    const char ch = p[1];
+    str = p + 2;
+    if (ch == '@') {
+      ColoredPrintf(color, "@");
+    } else if (ch == 'D') {
+      color = COLOR_DEFAULT;
+    } else if (ch == 'R') {
+      color = COLOR_RED;
+    } else if (ch == 'G') {
+      color = COLOR_GREEN;
+    } else if (ch == 'Y') {
+      color = COLOR_YELLOW;
+    } else {
+      --str;
+    }
+  }
+}
+
+static const char kColorEncodedHelpMessage[] =
+"This program contains tests written using " GTEST_NAME_ ". You can use the\n"
+"following command line flags to control its behavior:\n"
+"\n"
+"Test Selection:\n"
+"  @G--" GTEST_FLAG_PREFIX_ "list_tests@D\n"
+"      List the names of all tests instead of running them. The name of\n"
+"      TEST(Foo, Bar) is \"Foo.Bar\".\n"
+"  @G--" GTEST_FLAG_PREFIX_ "filter=@YPOSTIVE_PATTERNS"
+    "[@G-@YNEGATIVE_PATTERNS]@D\n"
+"      Run only the tests whose name matches one of the positive patterns but\n"
+"      none of the negative patterns. '?' matches any single character; '*'\n"
+"      matches any substring; ':' separates two patterns.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "also_run_disabled_tests@D\n"
+"      Run all disabled tests too.\n"
+"\n"
+"Test Execution:\n"
+"  @G--" GTEST_FLAG_PREFIX_ "repeat=@Y[COUNT]@D\n"
+"      Run the tests repeatedly; use a negative count to repeat forever.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "shuffle@D\n"
+"      Randomize tests' orders on every iteration.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "random_seed=@Y[NUMBER]@D\n"
+"      Random number seed to use for shuffling test orders (between 1 and\n"
+"      99999, or 0 to use a seed based on the current time).\n"
+"\n"
+"Test Output:\n"
+"  @G--" GTEST_FLAG_PREFIX_ "color=@Y(@Gyes@Y|@Gno@Y|@Gauto@Y)@D\n"
+"      Enable/disable colored output. The default is @Gauto@D.\n"
+"  -@G-" GTEST_FLAG_PREFIX_ "print_time=0@D\n"
+"      Don't print the elapsed time of each test.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "output=xml@Y[@G:@YDIRECTORY_PATH@G"
+    GTEST_PATH_SEP_ "@Y|@G:@YFILE_PATH]@D\n"
+"      Generate an XML report in the given directory or with the given file\n"
+"      name. @YFILE_PATH@D defaults to @Gtest_details.xml@D.\n"
+#if GTEST_CAN_STREAM_RESULTS_
+"  @G--" GTEST_FLAG_PREFIX_ "stream_result_to=@YHOST@G:@YPORT@D\n"
+"      Stream test results to the given server.\n"
+#endif  // GTEST_CAN_STREAM_RESULTS_
+"\n"
+"Assertion Behavior:\n"
+#if GTEST_HAS_DEATH_TEST && !GTEST_OS_WINDOWS
+"  @G--" GTEST_FLAG_PREFIX_ "death_test_style=@Y(@Gfast@Y|@Gthreadsafe@Y)@D\n"
+"      Set the default death test style.\n"
+#endif  // GTEST_HAS_DEATH_TEST && !GTEST_OS_WINDOWS
+"  @G--" GTEST_FLAG_PREFIX_ "break_on_failure@D\n"
+"      Turn assertion failures into debugger break-points.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "throw_on_failure@D\n"
+"      Turn assertion failures into C++ exceptions.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "catch_exceptions=0@D\n"
+"      Do not report exceptions as test failures. Instead, allow them\n"
+"      to crash the program or throw a pop-up (on Windows).\n"
+"\n"
+"Except for @G--" GTEST_FLAG_PREFIX_ "list_tests@D, you can alternatively set "
+    "the corresponding\n"
+"environment variable of a flag (all letters in upper-case). For example, to\n"
+"disable colored text output, you can either specify @G--" GTEST_FLAG_PREFIX_
+    "color=no@D or set\n"
+"the @G" GTEST_FLAG_PREFIX_UPPER_ "COLOR@D environment variable to @Gno@D.\n"
+"\n"
+"For more information, please read the " GTEST_NAME_ " documentation at\n"
+"@G" GTEST_PROJECT_URL_ "@D. If you find a bug in " GTEST_NAME_ "\n"
+"(not one in your own code or tests), please report it to\n"
+"@G<" GTEST_DEV_EMAIL_ ">@D.\n";
+
+// Parses the command line for Google Test flags, without initializing
+// other parts of Google Test.  The type parameter CharType can be
+// instantiated to either char or wchar_t.
+template <typename CharType>
+void ParseGoogleTestFlagsOnlyImpl(int* argc, CharType** argv) {
+  for (int i = 1; i < *argc; i++) {
+    const std::string arg_string = StreamableToString(argv[i]);
+    const char* const arg = arg_string.c_str();
+
+    using internal::ParseBoolFlag;
+    using internal::ParseInt32Flag;
+    using internal::ParseStringFlag;
+
+    // Do we see a Google Test flag?
+    if (ParseBoolFlag(arg, kAlsoRunDisabledTestsFlag,
+                      &GTEST_FLAG(also_run_disabled_tests)) ||
+        ParseBoolFlag(arg, kBreakOnFailureFlag,
+                      &GTEST_FLAG(break_on_failure)) ||
+        ParseBoolFlag(arg, kCatchExceptionsFlag,
+                      &GTEST_FLAG(catch_exceptions)) ||
+        ParseStringFlag(arg, kColorFlag, &GTEST_FLAG(color)) ||
+        ParseStringFlag(arg, kDeathTestStyleFlag,
+                        &GTEST_FLAG(death_test_style)) ||
+        ParseBoolFlag(arg, kDeathTestUseFork,
+                      &GTEST_FLAG(death_test_use_fork)) ||
+        ParseStringFlag(arg, kFilterFlag, &GTEST_FLAG(filter)) ||
+        ParseStringFlag(arg, kInternalRunDeathTestFlag,
+                        &GTEST_FLAG(internal_run_death_test)) ||
+        ParseBoolFlag(arg, kListTestsFlag, &GTEST_FLAG(list_tests)) ||
+        ParseStringFlag(arg, kOutputFlag, &GTEST_FLAG(output)) ||
+        ParseBoolFlag(arg, kPrintTimeFlag, &GTEST_FLAG(print_time)) ||
+        ParseInt32Flag(arg, kRandomSeedFlag, &GTEST_FLAG(random_seed)) ||
+        ParseInt32Flag(arg, kRepeatFlag, &GTEST_FLAG(repeat)) ||
+        ParseBoolFlag(arg, kShuffleFlag, &GTEST_FLAG(shuffle)) ||
+        ParseInt32Flag(arg, kStackTraceDepthFlag,
+                       &GTEST_FLAG(stack_trace_depth)) ||
+        ParseStringFlag(arg, kStreamResultToFlag,
+                        &GTEST_FLAG(stream_result_to)) ||
+        ParseBoolFlag(arg, kThrowOnFailureFlag,
+                      &GTEST_FLAG(throw_on_failure))
+        ) {
+      // Yes.  Shift the remainder of the argv list left by one.  Note
+      // that argv has (*argc + 1) elements, the last one always being
+      // NULL.  The following loop moves the trailing NULL element as
+      // well.
+      for (int j = i; j != *argc; j++) {
+        argv[j] = argv[j + 1];
+      }
+
+      // Decrements the argument count.
+      (*argc)--;
+
+      // We also need to decrement the iterator as we just removed
+      // an element.
+      i--;
+    } else if (arg_string == "--help" || arg_string == "-h" ||
+               arg_string == "-?" || arg_string == "/?" ||
+               HasGoogleTestFlagPrefix(arg)) {
+      // Both help flag and unrecognized Google Test flags (excluding
+      // internal ones) trigger help display.
+      g_help_flag = true;
+    }
+  }
+
+  if (g_help_flag) {
+    // We print the help here instead of in RUN_ALL_TESTS(), as the
+    // latter may not be called at all if the user is using Google
+    // Test with another testing framework.
+    PrintColorEncoded(kColorEncodedHelpMessage);
+  }
+}
+
+// Parses the command line for Google Test flags, without initializing
+// other parts of Google Test.
+void ParseGoogleTestFlagsOnly(int* argc, char** argv) {
+  ParseGoogleTestFlagsOnlyImpl(argc, argv);
+}
+void ParseGoogleTestFlagsOnly(int* argc, wchar_t** argv) {
+  ParseGoogleTestFlagsOnlyImpl(argc, argv);
+}
+
+// The internal implementation of InitGoogleTest().
+//
+// The type parameter CharType can be instantiated to either char or
+// wchar_t.
+template <typename CharType>
+void InitGoogleTestImpl(int* argc, CharType** argv) {
+  g_init_gtest_count++;
+
+  // We don't want to run the initialization code twice.
+  if (g_init_gtest_count != 1) return;
+
+  if (*argc <= 0) return;
+
+  internal::g_executable_path = internal::StreamableToString(argv[0]);
+
+#if GTEST_HAS_DEATH_TEST
+
+  g_argvs.clear();
+  for (int i = 0; i != *argc; i++) {
+    g_argvs.push_back(StreamableToString(argv[i]));
+  }
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+  ParseGoogleTestFlagsOnly(argc, argv);
+  GetUnitTestImpl()->PostFlagParsingInit();
+}
+
+}  // namespace internal
+
+// Initializes Google Test.  This must be called before calling
+// RUN_ALL_TESTS().  In particular, it parses a command line for the
+// flags that Google Test recognizes.  Whenever a Google Test flag is
+// seen, it is removed from argv, and *argc is decremented.
+//
+// No value is returned.  Instead, the Google Test flag variables are
+// updated.
+//
+// Calling the function for the second time has no user-visible effect.
+void InitGoogleTest(int* argc, char** argv) {
+  internal::InitGoogleTestImpl(argc, argv);
+}
+
+// This overloaded version can be used in Windows programs compiled in
+// UNICODE mode.
+void InitGoogleTest(int* argc, wchar_t** argv) {
+  internal::InitGoogleTestImpl(argc, argv);
+}
+
+}  // namespace testing
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan), vladl@google.com (Vlad Losev)
+//
+// This file implements death tests.
+
+
+#if GTEST_HAS_DEATH_TEST
+
+# if GTEST_OS_MAC
+#  include <crt_externs.h>
+# endif  // GTEST_OS_MAC
+
+# include <errno.h>
+# include <fcntl.h>
+# include <limits.h>
+
+# if GTEST_OS_LINUX
+#  include <signal.h>
+# endif  // GTEST_OS_LINUX
+
+# include <stdarg.h>
+
+# if GTEST_OS_WINDOWS
+#  include <windows.h>
+# else
+#  include <sys/mman.h>
+#  include <sys/wait.h>
+# endif  // GTEST_OS_WINDOWS
+
+# if GTEST_OS_QNX
+#  include <spawn.h>
+# endif  // GTEST_OS_QNX
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick is to
+// prevent a user from accidentally including gtest-internal-inl.h in
+// his code.
+#define GTEST_IMPLEMENTATION_ 1
+#undef GTEST_IMPLEMENTATION_
+
+namespace testing {
+
+// Constants.
+
+// The default death test style.
+static const char kDefaultDeathTestStyle[] = "fast";
+
+GTEST_DEFINE_string_(
+    death_test_style,
+    internal::StringFromGTestEnv("death_test_style", kDefaultDeathTestStyle),
+    "Indicates how to run a death test in a forked child process: "
+    "\"threadsafe\" (child process re-executes the test binary "
+    "from the beginning, running only the specific death test) or "
+    "\"fast\" (child process runs the death test immediately "
+    "after forking).");
+
+GTEST_DEFINE_bool_(
+    death_test_use_fork,
+    internal::BoolFromGTestEnv("death_test_use_fork", false),
+    "Instructs to use fork()/_exit() instead of clone() in death tests. "
+    "Ignored and always uses fork() on POSIX systems where clone() is not "
+    "implemented. Useful when running under valgrind or similar tools if "
+    "those do not support clone(). Valgrind 3.3.1 will just fail if "
+    "it sees an unsupported combination of clone() flags. "
+    "It is not recommended to use this flag w/o valgrind though it will "
+    "work in 99% of the cases. Once valgrind is fixed, this flag will "
+    "most likely be removed.");
+
+namespace internal {
+GTEST_DEFINE_string_(
+    internal_run_death_test, "",
+    "Indicates the file, line number, temporal index of "
+    "the single death test to run, and a file descriptor to "
+    "which a success code may be sent, all separated by "
+    "the '|' characters.  This flag is specified if and only if the current "
+    "process is a sub-process launched for running a thread-safe "
+    "death test.  FOR INTERNAL USE ONLY.");
+}  // namespace internal
+
+#if GTEST_HAS_DEATH_TEST
+
+namespace internal {
+
+// Valid only for fast death tests. Indicates the code is running in the
+// child process of a fast style death test.
+static bool g_in_fast_death_test_child = false;
+
+// Returns a Boolean value indicating whether the caller is currently
+// executing in the context of the death test child process.  Tools such as
+// Valgrind heap checkers may need this to modify their behavior in death
+// tests.  IMPORTANT: This is an internal utility.  Using it may break the
+// implementation of death tests.  User code MUST NOT use it.
+bool InDeathTestChild() {
+# if GTEST_OS_WINDOWS
+
+  // On Windows, death tests are thread-safe regardless of the value of the
+  // death_test_style flag.
+  return !GTEST_FLAG(internal_run_death_test).empty();
+
+# else
+
+  if (GTEST_FLAG(death_test_style) == "threadsafe")
+    return !GTEST_FLAG(internal_run_death_test).empty();
+  else
+    return g_in_fast_death_test_child;
+#endif
+}
+
+}  // namespace internal
+
+// ExitedWithCode constructor.
+ExitedWithCode::ExitedWithCode(int exit_code) : exit_code_(exit_code) {
+}
+
+// ExitedWithCode function-call operator.
+bool ExitedWithCode::operator()(int exit_status) const {
+# if GTEST_OS_WINDOWS
+
+  return exit_status == exit_code_;
+
+# else
+
+  return WIFEXITED(exit_status) && WEXITSTATUS(exit_status) == exit_code_;
+
+# endif  // GTEST_OS_WINDOWS
+}
+
+# if !GTEST_OS_WINDOWS
+// KilledBySignal constructor.
+KilledBySignal::KilledBySignal(int signum) : signum_(signum) {
+}
+
+// KilledBySignal function-call operator.
+bool KilledBySignal::operator()(int exit_status) const {
+  return WIFSIGNALED(exit_status) && WTERMSIG(exit_status) == signum_;
+}
+# endif  // !GTEST_OS_WINDOWS
+
+namespace internal {
+
+// Utilities needed for death tests.
+
+// Generates a textual description of a given exit code, in the format
+// specified by wait(2).
+static std::string ExitSummary(int exit_code) {
+  Message m;
+
+# if GTEST_OS_WINDOWS
+
+  m << "Exited with exit status " << exit_code;
+
+# else
+
+  if (WIFEXITED(exit_code)) {
+    m << "Exited with exit status " << WEXITSTATUS(exit_code);
+  } else if (WIFSIGNALED(exit_code)) {
+    m << "Terminated by signal " << WTERMSIG(exit_code);
+  }
+#  ifdef WCOREDUMP
+  if (WCOREDUMP(exit_code)) {
+    m << " (core dumped)";
+  }
+#  endif
+# endif  // GTEST_OS_WINDOWS
+
+  return m.GetString();
+}
+
+// Returns true if exit_status describes a process that was terminated
+// by a signal, or exited normally with a nonzero exit code.
+bool ExitedUnsuccessfully(int exit_status) {
+  return !ExitedWithCode(0)(exit_status);
+}
+
+# if !GTEST_OS_WINDOWS
+// Generates a textual failure message when a death test finds more than
+// one thread running, or cannot determine the number of threads, prior
+// to executing the given statement.  It is the responsibility of the
+// caller not to pass a thread_count of 1.
+static std::string DeathTestThreadWarning(size_t thread_count) {
+  Message msg;
+  msg << "Death tests use fork(), which is unsafe particularly"
+      << " in a threaded context. For this test, " << GTEST_NAME_ << " ";
+  if (thread_count == 0)
+    msg << "couldn't detect the number of threads.";
+  else
+    msg << "detected " << thread_count << " threads.";
+  return msg.GetString();
+}
+# endif  // !GTEST_OS_WINDOWS
+
+// Flag characters for reporting a death test that did not die.
+static const char kDeathTestLived = 'L';
+static const char kDeathTestReturned = 'R';
+static const char kDeathTestThrew = 'T';
+static const char kDeathTestInternalError = 'I';
+
+// An enumeration describing all of the possible ways that a death test can
+// conclude.  DIED means that the process died while executing the test
+// code; LIVED means that process lived beyond the end of the test code;
+// RETURNED means that the test statement attempted to execute a return
+// statement, which is not allowed; THREW means that the test statement
+// returned control by throwing an exception.  IN_PROGRESS means the test
+// has not yet concluded.
+// TODO(vladl@google.com): Unify names and possibly values for
+// AbortReason, DeathTestOutcome, and flag characters above.
+enum DeathTestOutcome { IN_PROGRESS, DIED, LIVED, RETURNED, THREW };
+
+// Routine for aborting the program which is safe to call from an
+// exec-style death test child process, in which case the error
+// message is propagated back to the parent process.  Otherwise, the
+// message is simply printed to stderr.  In either case, the program
+// then exits with status 1.
+void DeathTestAbort(const std::string& message) {
+  // On a POSIX system, this function may be called from a threadsafe-style
+  // death test child process, which operates on a very small stack.  Use
+  // the heap for any additional non-minuscule memory requirements.
+  const InternalRunDeathTestFlag* const flag =
+      GetUnitTestImpl()->internal_run_death_test_flag();
+  if (flag != NULL) {
+    FILE* parent = posix::FDOpen(flag->write_fd(), "w");
+    fputc(kDeathTestInternalError, parent);
+    fprintf(parent, "%s", message.c_str());
+    fflush(parent);
+    _exit(1);
+  } else {
+    fprintf(stderr, "%s", message.c_str());
+    fflush(stderr);
+    posix::Abort();
+  }
+}
+
+// A replacement for CHECK that calls DeathTestAbort if the assertion
+// fails.
+# define GTEST_DEATH_TEST_CHECK_(expression) \
+  do { \
+    if (!::testing::internal::IsTrue(expression)) { \
+      DeathTestAbort( \
+          ::std::string("CHECK failed: File ") + __FILE__ +  ", line " \
+          + ::testing::internal::StreamableToString(__LINE__) + ": " \
+          + #expression); \
+    } \
+  } while (::testing::internal::AlwaysFalse())
+
+// This macro is similar to GTEST_DEATH_TEST_CHECK_, but it is meant for
+// evaluating any system call that fulfills two conditions: it must return
+// -1 on failure, and set errno to EINTR when it is interrupted and
+// should be tried again.  The macro expands to a loop that repeatedly
+// evaluates the expression as long as it evaluates to -1 and sets
+// errno to EINTR.  If the expression evaluates to -1 but errno is
+// something other than EINTR, DeathTestAbort is called.
+# define GTEST_DEATH_TEST_CHECK_SYSCALL_(expression) \
+  do { \
+    int gtest_retval; \
+    do { \
+      gtest_retval = (expression); \
+    } while (gtest_retval == -1 && errno == EINTR); \
+    if (gtest_retval == -1) { \
+      DeathTestAbort( \
+          ::std::string("CHECK failed: File ") + __FILE__ + ", line " \
+          + ::testing::internal::StreamableToString(__LINE__) + ": " \
+          + #expression + " != -1"); \
+    } \
+  } while (::testing::internal::AlwaysFalse())
+
+// Returns the message describing the last system error in errno.
+std::string GetLastErrnoDescription() {
+    return errno == 0 ? "" : posix::StrError(errno);
+}
+
+// This is called from a death test parent process to read a failure
+// message from the death test child process and log it with the FATAL
+// severity. On Windows, the message is read from a pipe handle. On other
+// platforms, it is read from a file descriptor.
+static void FailFromInternalError(int fd) {
+  Message error;
+  char buffer[256];
+  int num_read;
+
+  do {
+    while ((num_read = posix::Read(fd, buffer, 255)) > 0) {
+      buffer[num_read] = '\0';
+      error << buffer;
+    }
+  } while (num_read == -1 && errno == EINTR);
+
+  if (num_read == 0) {
+    GTEST_LOG_(FATAL) << error.GetString();
+  } else {
+    const int last_error = errno;
+    GTEST_LOG_(FATAL) << "Error while reading death test internal: "
+                      << GetLastErrnoDescription() << " [" << last_error << "]";
+  }
+}
+
+// Death test constructor.  Increments the running death test count
+// for the current test.
+DeathTest::DeathTest() {
+  TestInfo* const info = GetUnitTestImpl()->current_test_info();
+  if (info == NULL) {
+    DeathTestAbort("Cannot run a death test outside of a TEST or "
+                   "TEST_F construct");
+  }
+}
+
+// Creates and returns a death test by dispatching to the current
+// death test factory.
+bool DeathTest::Create(const char* statement, const RE* regex,
+                       const char* file, int line, DeathTest** test) {
+  return GetUnitTestImpl()->death_test_factory()->Create(
+      statement, regex, file, line, test);
+}
+
+const char* DeathTest::LastMessage() {
+  return last_death_test_message_.c_str();
+}
+
+void DeathTest::set_last_death_test_message(const std::string& message) {
+  last_death_test_message_ = message;
+}
+
+std::string DeathTest::last_death_test_message_;
+
+// Provides cross platform implementation for some death functionality.
+class DeathTestImpl : public DeathTest {
+ protected:
+  DeathTestImpl(const char* a_statement, const RE* a_regex)
+      : statement_(a_statement),
+        regex_(a_regex),
+        spawned_(false),
+        status_(-1),
+        outcome_(IN_PROGRESS),
+        read_fd_(-1),
+        write_fd_(-1) {}
+
+  // read_fd_ is expected to be closed and cleared by a derived class.
+  ~DeathTestImpl() { GTEST_DEATH_TEST_CHECK_(read_fd_ == -1); }
+
+  void Abort(AbortReason reason);
+  virtual bool Passed(bool status_ok);
+
+  const char* statement() const { return statement_; }
+  const RE* regex() const { return regex_; }
+  bool spawned() const { return spawned_; }
+  void set_spawned(bool is_spawned) { spawned_ = is_spawned; }
+  int status() const { return status_; }
+  void set_status(int a_status) { status_ = a_status; }
+  DeathTestOutcome outcome() const { return outcome_; }
+  void set_outcome(DeathTestOutcome an_outcome) { outcome_ = an_outcome; }
+  int read_fd() const { return read_fd_; }
+  void set_read_fd(int fd) { read_fd_ = fd; }
+  int write_fd() const { return write_fd_; }
+  void set_write_fd(int fd) { write_fd_ = fd; }
+
+  // Called in the parent process only. Reads the result code of the death
+  // test child process via a pipe, interprets it to set the outcome_
+  // member, and closes read_fd_.  Outputs diagnostics and terminates in
+  // case of unexpected codes.
+  void ReadAndInterpretStatusByte();
+
+ private:
+  // The textual content of the code this object is testing.  This class
+  // doesn't own this string and should not attempt to delete it.
+  const char* const statement_;
+  // The regular expression which test output must match.  DeathTestImpl
+  // doesn't own this object and should not attempt to delete it.
+  const RE* const regex_;
+  // True if the death test child process has been successfully spawned.
+  bool spawned_;
+  // The exit status of the child process.
+  int status_;
+  // How the death test concluded.
+  DeathTestOutcome outcome_;
+  // Descriptor to the read end of the pipe to the child process.  It is
+  // always -1 in the child process.  The child keeps its write end of the
+  // pipe in write_fd_.
+  int read_fd_;
+  // Descriptor to the child's write end of the pipe to the parent process.
+  // It is always -1 in the parent process.  The parent keeps its end of the
+  // pipe in read_fd_.
+  int write_fd_;
+};
+
+// Called in the parent process only. Reads the result code of the death
+// test child process via a pipe, interprets it to set the outcome_
+// member, and closes read_fd_.  Outputs diagnostics and terminates in
+// case of unexpected codes.
+void DeathTestImpl::ReadAndInterpretStatusByte() {
+  char flag;
+  int bytes_read;
+
+  // The read() here blocks until data is available (signifying the
+  // failure of the death test) or until the pipe is closed (signifying
+  // its success), so it's okay to call this in the parent before
+  // the child process has exited.
+  do {
+    bytes_read = posix::Read(read_fd(), &flag, 1);
+  } while (bytes_read == -1 && errno == EINTR);
+
+  if (bytes_read == 0) {
+    set_outcome(DIED);
+  } else if (bytes_read == 1) {
+    switch (flag) {
+      case kDeathTestReturned:
+        set_outcome(RETURNED);
+        break;
+      case kDeathTestThrew:
+        set_outcome(THREW);
+        break;
+      case kDeathTestLived:
+        set_outcome(LIVED);
+        break;
+      case kDeathTestInternalError:
+        FailFromInternalError(read_fd());  // Does not return.
+        break;
+      default:
+        GTEST_LOG_(FATAL) << "Death test child process reported "
+                          << "unexpected status byte ("
+                          << static_cast<unsigned int>(flag) << ")";
+    }
+  } else {
+    GTEST_LOG_(FATAL) << "Read from death test child process failed: "
+                      << GetLastErrnoDescription();
+  }
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(posix::Close(read_fd()));
+  set_read_fd(-1);
+}
+
+// Signals that the death test code which should have exited, didn't.
+// Should be called only in a death test child process.
+// Writes a status byte to the child's status file descriptor, then
+// calls _exit(1).
+void DeathTestImpl::Abort(AbortReason reason) {
+  // The parent process considers the death test to be a failure if
+  // it finds any data in our pipe.  So, here we write a single flag byte
+  // to the pipe, then exit.
+  const char status_ch =
+      reason == TEST_DID_NOT_DIE ? kDeathTestLived :
+      reason == TEST_THREW_EXCEPTION ? kDeathTestThrew : kDeathTestReturned;
+
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(posix::Write(write_fd(), &status_ch, 1));
+  // We are leaking the descriptor here because on some platforms (i.e.,
+  // when built as Windows DLL), destructors of global objects will still
+  // run after calling _exit(). On such systems, write_fd_ will be
+  // indirectly closed from the destructor of UnitTestImpl, causing double
+  // close if it is also closed here. On debug configurations, double close
+  // may assert. As there are no in-process buffers to flush here, we are
+  // relying on the OS to close the descriptor after the process terminates
+  // when the destructors are not run.
+  _exit(1);  // Exits w/o any normal exit hooks (we were supposed to crash)
+}
+
+// Returns an indented copy of stderr output for a death test.
+// This makes distinguishing death test output lines from regular log lines
+// much easier.
+static ::std::string FormatDeathTestOutput(const ::std::string& output) {
+  ::std::string ret;
+  for (size_t at = 0; ; ) {
+    const size_t line_end = output.find('\n', at);
+    ret += "[  DEATH   ] ";
+    if (line_end == ::std::string::npos) {
+      ret += output.substr(at);
+      break;
+    }
+    ret += output.substr(at, line_end + 1 - at);
+    at = line_end + 1;
+  }
+  return ret;
+}
+
+// Assesses the success or failure of a death test, using both private
+// members which have previously been set, and one argument:
+//
+// Private data members:
+//   outcome:  An enumeration describing how the death test
+//             concluded: DIED, LIVED, THREW, or RETURNED.  The death test
+//             fails in the latter three cases.
+//   status:   The exit status of the child process. On *nix, it is in the
+//             in the format specified by wait(2). On Windows, this is the
+//             value supplied to the ExitProcess() API or a numeric code
+//             of the exception that terminated the program.
+//   regex:    A regular expression object to be applied to
+//             the test's captured standard error output; the death test
+//             fails if it does not match.
+//
+// Argument:
+//   status_ok: true if exit_status is acceptable in the context of
+//              this particular death test, which fails if it is false
+//
+// Returns true iff all of the above conditions are met.  Otherwise, the
+// first failing condition, in the order given above, is the one that is
+// reported. Also sets the last death test message string.
+bool DeathTestImpl::Passed(bool status_ok) {
+  if (!spawned())
+    return false;
+
+  const std::string error_message = GetCapturedStderr();
+
+  bool success = false;
+  Message buffer;
+
+  buffer << "Death test: " << statement() << "\n";
+  switch (outcome()) {
+    case LIVED:
+      buffer << "    Result: failed to die.\n"
+             << " Error msg:\n" << FormatDeathTestOutput(error_message);
+      break;
+    case THREW:
+      buffer << "    Result: threw an exception.\n"
+             << " Error msg:\n" << FormatDeathTestOutput(error_message);
+      break;
+    case RETURNED:
+      buffer << "    Result: illegal return in test statement.\n"
+             << " Error msg:\n" << FormatDeathTestOutput(error_message);
+      break;
+    case DIED:
+      if (status_ok) {
+        const bool matched = RE::PartialMatch(error_message.c_str(), *regex());
+        if (matched) {
+          success = true;
+        } else {
+          buffer << "    Result: died but not with expected error.\n"
+                 << "  Expected: " << regex()->pattern() << "\n"
+                 << "Actual msg:\n" << FormatDeathTestOutput(error_message);
+        }
+      } else {
+        buffer << "    Result: died but not with expected exit code:\n"
+               << "            " << ExitSummary(status()) << "\n"
+               << "Actual msg:\n" << FormatDeathTestOutput(error_message);
+      }
+      break;
+    case IN_PROGRESS:
+    default:
+      GTEST_LOG_(FATAL)
+          << "DeathTest::Passed somehow called before conclusion of test";
+  }
+
+  DeathTest::set_last_death_test_message(buffer.GetString());
+  return success;
+}
+
+# if GTEST_OS_WINDOWS
+// WindowsDeathTest implements death tests on Windows. Due to the
+// specifics of starting new processes on Windows, death tests there are
+// always threadsafe, and Google Test considers the
+// --gtest_death_test_style=fast setting to be equivalent to
+// --gtest_death_test_style=threadsafe there.
+//
+// A few implementation notes:  Like the Linux version, the Windows
+// implementation uses pipes for child-to-parent communication. But due to
+// the specifics of pipes on Windows, some extra steps are required:
+//
+// 1. The parent creates a communication pipe and stores handles to both
+//    ends of it.
+// 2. The parent starts the child and provides it with the information
+//    necessary to acquire the handle to the write end of the pipe.
+// 3. The child acquires the write end of the pipe and signals the parent
+//    using a Windows event.
+// 4. Now the parent can release the write end of the pipe on its side. If
+//    this is done before step 3, the object's reference count goes down to
+//    0 and it is destroyed, preventing the child from acquiring it. The
+//    parent now has to release it, or read operations on the read end of
+//    the pipe will not return when the child terminates.
+// 5. The parent reads child's output through the pipe (outcome code and
+//    any possible error messages) from the pipe, and its stderr and then
+//    determines whether to fail the test.
+//
+// Note: to distinguish Win32 API calls from the local method and function
+// calls, the former are explicitly resolved in the global namespace.
+//
+class WindowsDeathTest : public DeathTestImpl {
+ public:
+  WindowsDeathTest(const char* a_statement,
+                   const RE* a_regex,
+                   const char* file,
+                   int line)
+      : DeathTestImpl(a_statement, a_regex), file_(file), line_(line) {}
+
+  // All of these virtual functions are inherited from DeathTest.
+  virtual int Wait();
+  virtual TestRole AssumeRole();
+
+ private:
+  // The name of the file in which the death test is located.
+  const char* const file_;
+  // The line number on which the death test is located.
+  const int line_;
+  // Handle to the write end of the pipe to the child process.
+  AutoHandle write_handle_;
+  // Child process handle.
+  AutoHandle child_handle_;
+  // Event the child process uses to signal the parent that it has
+  // acquired the handle to the write end of the pipe. After seeing this
+  // event the parent can release its own handles to make sure its
+  // ReadFile() calls return when the child terminates.
+  AutoHandle event_handle_;
+};
+
+// Waits for the child in a death test to exit, returning its exit
+// status, or 0 if no child process exists.  As a side effect, sets the
+// outcome data member.
+int WindowsDeathTest::Wait() {
+  if (!spawned())
+    return 0;
+
+  // Wait until the child either signals that it has acquired the write end
+  // of the pipe or it dies.
+  const HANDLE wait_handles[2] = { child_handle_.Get(), event_handle_.Get() };
+  switch (::WaitForMultipleObjects(2,
+                                   wait_handles,
+                                   FALSE,  // Waits for any of the handles.
+                                   INFINITE)) {
+    case WAIT_OBJECT_0:
+    case WAIT_OBJECT_0 + 1:
+      break;
+    default:
+      GTEST_DEATH_TEST_CHECK_(false);  // Should not get here.
+  }
+
+  // The child has acquired the write end of the pipe or exited.
+  // We release the handle on our side and continue.
+  write_handle_.Reset();
+  event_handle_.Reset();
+
+  ReadAndInterpretStatusByte();
+
+  // Waits for the child process to exit if it haven't already. This
+  // returns immediately if the child has already exited, regardless of
+  // whether previous calls to WaitForMultipleObjects synchronized on this
+  // handle or not.
+  GTEST_DEATH_TEST_CHECK_(
+      WAIT_OBJECT_0 == ::WaitForSingleObject(child_handle_.Get(),
+                                             INFINITE));
+  DWORD status_code;
+  GTEST_DEATH_TEST_CHECK_(
+      ::GetExitCodeProcess(child_handle_.Get(), &status_code) != FALSE);
+  child_handle_.Reset();
+  set_status(static_cast<int>(status_code));
+  return status();
+}
+
+// The AssumeRole process for a Windows death test.  It creates a child
+// process with the same executable as the current process to run the
+// death test.  The child process is given the --gtest_filter and
+// --gtest_internal_run_death_test flags such that it knows to run the
+// current death test only.
+DeathTest::TestRole WindowsDeathTest::AssumeRole() {
+  const UnitTestImpl* const impl = GetUnitTestImpl();
+  const InternalRunDeathTestFlag* const flag =
+      impl->internal_run_death_test_flag();
+  const TestInfo* const info = impl->current_test_info();
+  const int death_test_index = info->result()->death_test_count();
+
+  if (flag != NULL) {
+    // ParseInternalRunDeathTestFlag() has performed all the necessary
+    // processing.
+    set_write_fd(flag->write_fd());
+    return EXECUTE_TEST;
+  }
+
+  // WindowsDeathTest uses an anonymous pipe to communicate results of
+  // a death test.
+  SECURITY_ATTRIBUTES handles_are_inheritable = {
+    sizeof(SECURITY_ATTRIBUTES), NULL, TRUE };
+  HANDLE read_handle, write_handle;
+  GTEST_DEATH_TEST_CHECK_(
+      ::CreatePipe(&read_handle, &write_handle, &handles_are_inheritable,
+                   0)  // Default buffer size.
+      != FALSE);
+  set_read_fd(::_open_osfhandle(reinterpret_cast<intptr_t>(read_handle),
+                                O_RDONLY));
+  write_handle_.Reset(write_handle);
+  event_handle_.Reset(::CreateEvent(
+      &handles_are_inheritable,
+      TRUE,    // The event will automatically reset to non-signaled state.
+      FALSE,   // The initial state is non-signalled.
+      NULL));  // The even is unnamed.
+  GTEST_DEATH_TEST_CHECK_(event_handle_.Get() != NULL);
+  const std::string filter_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kFilterFlag + "=" +
+      info->test_case_name() + "." + info->name();
+  const std::string internal_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kInternalRunDeathTestFlag +
+      "=" + file_ + "|" + StreamableToString(line_) + "|" +
+      StreamableToString(death_test_index) + "|" +
+      StreamableToString(static_cast<unsigned int>(::GetCurrentProcessId())) +
+      // size_t has the same width as pointers on both 32-bit and 64-bit
+      // Windows platforms.
+      // See http://msdn.microsoft.com/en-us/library/tcxf1dw6.aspx.
+      "|" + StreamableToString(reinterpret_cast<size_t>(write_handle)) +
+      "|" + StreamableToString(reinterpret_cast<size_t>(event_handle_.Get()));
+
+  char executable_path[_MAX_PATH + 1];  // NOLINT
+  GTEST_DEATH_TEST_CHECK_(
+      _MAX_PATH + 1 != ::GetModuleFileNameA(NULL,
+                                            executable_path,
+                                            _MAX_PATH));
+
+  std::string command_line =
+      std::string(::GetCommandLineA()) + " " + filter_flag + " \"" +
+      internal_flag + "\"";
+
+  DeathTest::set_last_death_test_message("");
+
+  CaptureStderr();
+  // Flush the log buffers since the log streams are shared with the child.
+  FlushInfoLog();
+
+  // The child process will share the standard handles with the parent.
+  STARTUPINFOA startup_info;
+  memset(&startup_info, 0, sizeof(STARTUPINFO));
+  startup_info.dwFlags = STARTF_USESTDHANDLES;
+  startup_info.hStdInput = ::GetStdHandle(STD_INPUT_HANDLE);
+  startup_info.hStdOutput = ::GetStdHandle(STD_OUTPUT_HANDLE);
+  startup_info.hStdError = ::GetStdHandle(STD_ERROR_HANDLE);
+
+  PROCESS_INFORMATION process_info;
+  GTEST_DEATH_TEST_CHECK_(::CreateProcessA(
+      executable_path,
+      const_cast<char*>(command_line.c_str()),
+      NULL,   // Retuned process handle is not inheritable.
+      NULL,   // Retuned thread handle is not inheritable.
+      TRUE,   // Child inherits all inheritable handles (for write_handle_).
+      0x0,    // Default creation flags.
+      NULL,   // Inherit the parent's environment.
+      UnitTest::GetInstance()->original_working_dir(),
+      &startup_info,
+      &process_info) != FALSE);
+  child_handle_.Reset(process_info.hProcess);
+  ::CloseHandle(process_info.hThread);
+  set_spawned(true);
+  return OVERSEE_TEST;
+}
+# else  // We are not on Windows.
+
+// ForkingDeathTest provides implementations for most of the abstract
+// methods of the DeathTest interface.  Only the AssumeRole method is
+// left undefined.
+class ForkingDeathTest : public DeathTestImpl {
+ public:
+  ForkingDeathTest(const char* statement, const RE* regex);
+
+  // All of these virtual functions are inherited from DeathTest.
+  virtual int Wait();
+
+ protected:
+  void set_child_pid(pid_t child_pid) { child_pid_ = child_pid; }
+
+ private:
+  // PID of child process during death test; 0 in the child process itself.
+  pid_t child_pid_;
+};
+
+// Constructs a ForkingDeathTest.
+ForkingDeathTest::ForkingDeathTest(const char* a_statement, const RE* a_regex)
+    : DeathTestImpl(a_statement, a_regex),
+      child_pid_(-1) {}
+
+// Waits for the child in a death test to exit, returning its exit
+// status, or 0 if no child process exists.  As a side effect, sets the
+// outcome data member.
+int ForkingDeathTest::Wait() {
+  if (!spawned())
+    return 0;
+
+  ReadAndInterpretStatusByte();
+
+  int status_value;
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(waitpid(child_pid_, &status_value, 0));
+  set_status(status_value);
+  return status_value;
+}
+
+// A concrete death test class that forks, then immediately runs the test
+// in the child process.
+class NoExecDeathTest : public ForkingDeathTest {
+ public:
+  NoExecDeathTest(const char* a_statement, const RE* a_regex) :
+      ForkingDeathTest(a_statement, a_regex) { }
+  virtual TestRole AssumeRole();
+};
+
+// The AssumeRole process for a fork-and-run death test.  It implements a
+// straightforward fork, with a simple pipe to transmit the status byte.
+DeathTest::TestRole NoExecDeathTest::AssumeRole() {
+  const size_t thread_count = GetThreadCount();
+  if (thread_count != 1) {
+    GTEST_LOG_(WARNING) << DeathTestThreadWarning(thread_count);
+  }
+
+  int pipe_fd[2];
+  GTEST_DEATH_TEST_CHECK_(pipe(pipe_fd) != -1);
+
+  DeathTest::set_last_death_test_message("");
+  CaptureStderr();
+  // When we fork the process below, the log file buffers are copied, but the
+  // file descriptors are shared.  We flush all log files here so that closing
+  // the file descriptors in the child process doesn't throw off the
+  // synchronization between descriptors and buffers in the parent process.
+  // This is as close to the fork as possible to avoid a race condition in case
+  // there are multiple threads running before the death test, and another
+  // thread writes to the log file.
+  FlushInfoLog();
+
+  const pid_t child_pid = fork();
+  GTEST_DEATH_TEST_CHECK_(child_pid != -1);
+  set_child_pid(child_pid);
+  if (child_pid == 0) {
+    GTEST_DEATH_TEST_CHECK_SYSCALL_(close(pipe_fd[0]));
+    set_write_fd(pipe_fd[1]);
+    // Redirects all logging to stderr in the child process to prevent
+    // concurrent writes to the log files.  We capture stderr in the parent
+    // process and append the child process' output to a log.
+    LogToStderr();
+    // Event forwarding to the listeners of event listener API mush be shut
+    // down in death test subprocesses.
+    GetUnitTestImpl()->listeners()->SuppressEventForwarding();
+    g_in_fast_death_test_child = true;
+    return EXECUTE_TEST;
+  } else {
+    GTEST_DEATH_TEST_CHECK_SYSCALL_(close(pipe_fd[1]));
+    set_read_fd(pipe_fd[0]);
+    set_spawned(true);
+    return OVERSEE_TEST;
+  }
+}
+
+// A concrete death test class that forks and re-executes the main
+// program from the beginning, with command-line flags set that cause
+// only this specific death test to be run.
+class ExecDeathTest : public ForkingDeathTest {
+ public:
+  ExecDeathTest(const char* a_statement, const RE* a_regex,
+                const char* file, int line) :
+      ForkingDeathTest(a_statement, a_regex), file_(file), line_(line) { }
+  virtual TestRole AssumeRole();
+ private:
+  static ::std::vector<testing::internal::string>
+  GetArgvsForDeathTestChildProcess() {
+    ::std::vector<testing::internal::string> args = GetInjectableArgvs();
+    return args;
+  }
+  // The name of the file in which the death test is located.
+  const char* const file_;
+  // The line number on which the death test is located.
+  const int line_;
+};
+
+// Utility class for accumulating command-line arguments.
+class Arguments {
+ public:
+  Arguments() {
+    args_.push_back(NULL);
+  }
+
+  ~Arguments() {
+    for (std::vector<char*>::iterator i = args_.begin(); i != args_.end();
+         ++i) {
+      free(*i);
+    }
+  }
+  void AddArgument(const char* argument) {
+    args_.insert(args_.end() - 1, posix::StrDup(argument));
+  }
+
+  template <typename Str>
+  void AddArguments(const ::std::vector<Str>& arguments) {
+    for (typename ::std::vector<Str>::const_iterator i = arguments.begin();
+         i != arguments.end();
+         ++i) {
+      args_.insert(args_.end() - 1, posix::StrDup(i->c_str()));
+    }
+  }
+  char* const* Argv() {
+    return &args_[0];
+  }
+
+ private:
+  std::vector<char*> args_;
+};
+
+// A struct that encompasses the arguments to the child process of a
+// threadsafe-style death test process.
+struct ExecDeathTestArgs {
+  char* const* argv;  // Command-line arguments for the child's call to exec
+  int close_fd;       // File descriptor to close; the read end of a pipe
+};
+
+#  if GTEST_OS_MAC
+inline char** GetEnviron() {
+  // When Google Test is built as a framework on MacOS X, the environ variable
+  // is unavailable. Apple's documentation (man environ) recommends using
+  // _NSGetEnviron() instead.
+  return *_NSGetEnviron();
+}
+#  else
+// Some POSIX platforms expect you to declare environ. extern "C" makes
+// it reside in the global namespace.
+extern "C" char** environ;
+inline char** GetEnviron() { return environ; }
+#  endif  // GTEST_OS_MAC
+
+#  if !GTEST_OS_QNX
+// The main function for a threadsafe-style death test child process.
+// This function is called in a clone()-ed process and thus must avoid
+// any potentially unsafe operations like malloc or libc functions.
+static int ExecDeathTestChildMain(void* child_arg) {
+  ExecDeathTestArgs* const args = static_cast<ExecDeathTestArgs*>(child_arg);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(close(args->close_fd));
+
+  // We need to execute the test program in the same environment where
+  // it was originally invoked.  Therefore we change to the original
+  // working directory first.
+  const char* const original_dir =
+      UnitTest::GetInstance()->original_working_dir();
+  // We can safely call chdir() as it's a direct system call.
+  if (chdir(original_dir) != 0) {
+    DeathTestAbort(std::string("chdir(\"") + original_dir + "\") failed: " +
+                   GetLastErrnoDescription());
+    return EXIT_FAILURE;
+  }
+
+  // We can safely call execve() as it's a direct system call.  We
+  // cannot use execvp() as it's a libc function and thus potentially
+  // unsafe.  Since execve() doesn't search the PATH, the user must
+  // invoke the test program via a valid path that contains at least
+  // one path separator.
+  execve(args->argv[0], args->argv, GetEnviron());
+  DeathTestAbort(std::string("execve(") + args->argv[0] + ", ...) in " +
+                 original_dir + " failed: " +
+                 GetLastErrnoDescription());
+  return EXIT_FAILURE;
+}
+#  endif  // !GTEST_OS_QNX
+
+// Two utility routines that together determine the direction the stack
+// grows.
+// This could be accomplished more elegantly by a single recursive
+// function, but we want to guard against the unlikely possibility of
+// a smart compiler optimizing the recursion away.
+//
+// GTEST_NO_INLINE_ is required to prevent GCC 4.6 from inlining
+// StackLowerThanAddress into StackGrowsDown, which then doesn't give
+// correct answer.
+void StackLowerThanAddress(const void* ptr, bool* result) GTEST_NO_INLINE_;
+void StackLowerThanAddress(const void* ptr, bool* result) {
+  int dummy;
+  *result = (&dummy < ptr);
+}
+
+bool StackGrowsDown() {
+  int dummy;
+  bool result;
+  StackLowerThanAddress(&dummy, &result);
+  return result;
+}
+
+// Spawns a child process with the same executable as the current process in
+// a thread-safe manner and instructs it to run the death test.  The
+// implementation uses fork(2) + exec.  On systems where clone(2) is
+// available, it is used instead, being slightly more thread-safe.  On QNX,
+// fork supports only single-threaded environments, so this function uses
+// spawn(2) there instead.  The function dies with an error message if
+// anything goes wrong.
+static pid_t ExecDeathTestSpawnChild(char* const* argv, int close_fd) {
+  ExecDeathTestArgs args = { argv, close_fd };
+  pid_t child_pid = -1;
+
+#  if GTEST_OS_QNX
+  // Obtains the current directory and sets it to be closed in the child
+  // process.
+  const int cwd_fd = open(".", O_RDONLY);
+  GTEST_DEATH_TEST_CHECK_(cwd_fd != -1);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(fcntl(cwd_fd, F_SETFD, FD_CLOEXEC));
+  // We need to execute the test program in the same environment where
+  // it was originally invoked.  Therefore we change to the original
+  // working directory first.
+  const char* const original_dir =
+      UnitTest::GetInstance()->original_working_dir();
+  // We can safely call chdir() as it's a direct system call.
+  if (chdir(original_dir) != 0) {
+    DeathTestAbort(std::string("chdir(\"") + original_dir + "\") failed: " +
+                   GetLastErrnoDescription());
+    return EXIT_FAILURE;
+  }
+
+  int fd_flags;
+  // Set close_fd to be closed after spawn.
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(fd_flags = fcntl(close_fd, F_GETFD));
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(fcntl(close_fd, F_SETFD,
+                                        fd_flags | FD_CLOEXEC));
+  struct inheritance inherit = {0};
+  // spawn is a system call.
+  child_pid = spawn(args.argv[0], 0, NULL, &inherit, args.argv, GetEnviron());
+  // Restores the current working directory.
+  GTEST_DEATH_TEST_CHECK_(fchdir(cwd_fd) != -1);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(close(cwd_fd));
+
+#  else   // GTEST_OS_QNX
+#   if GTEST_OS_LINUX
+  // When a SIGPROF signal is received while fork() or clone() are executing,
+  // the process may hang. To avoid this, we ignore SIGPROF here and re-enable
+  // it after the call to fork()/clone() is complete.
+  struct sigaction saved_sigprof_action;
+  struct sigaction ignore_sigprof_action;
+  memset(&ignore_sigprof_action, 0, sizeof(ignore_sigprof_action));
+  sigemptyset(&ignore_sigprof_action.sa_mask);
+  ignore_sigprof_action.sa_handler = SIG_IGN;
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(sigaction(
+      SIGPROF, &ignore_sigprof_action, &saved_sigprof_action));
+#   endif  // GTEST_OS_LINUX
+
+#   if GTEST_HAS_CLONE
+  const bool use_fork = GTEST_FLAG(death_test_use_fork);
+
+  if (!use_fork) {
+    static const bool stack_grows_down = StackGrowsDown();
+    const size_t stack_size = getpagesize();
+    // MMAP_ANONYMOUS is not defined on Mac, so we use MAP_ANON instead.
+    void* const stack = mmap(NULL, stack_size, PROT_READ | PROT_WRITE,
+                             MAP_ANON | MAP_PRIVATE, -1, 0);
+    GTEST_DEATH_TEST_CHECK_(stack != MAP_FAILED);
+
+    // Maximum stack alignment in bytes:  For a downward-growing stack, this
+    // amount is subtracted from size of the stack space to get an address
+    // that is within the stack space and is aligned on all systems we care
+    // about.  As far as I know there is no ABI with stack alignment greater
+    // than 64.  We assume stack and stack_size already have alignment of
+    // kMaxStackAlignment.
+    const size_t kMaxStackAlignment = 64;
+    void* const stack_top =
+        static_cast<char*>(stack) +
+            (stack_grows_down ? stack_size - kMaxStackAlignment : 0);
+    GTEST_DEATH_TEST_CHECK_(stack_size > kMaxStackAlignment &&
+        reinterpret_cast<intptr_t>(stack_top) % kMaxStackAlignment == 0);
+
+    child_pid = clone(&ExecDeathTestChildMain, stack_top, SIGCHLD, &args);
+
+    GTEST_DEATH_TEST_CHECK_(munmap(stack, stack_size) != -1);
+  }
+#   else
+  const bool use_fork = true;
+#   endif  // GTEST_HAS_CLONE
+
+  if (use_fork && (child_pid = fork()) == 0) {
+      ExecDeathTestChildMain(&args);
+      _exit(0);
+  }
+#  endif  // GTEST_OS_QNX
+#  if GTEST_OS_LINUX
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(
+      sigaction(SIGPROF, &saved_sigprof_action, NULL));
+#  endif  // GTEST_OS_LINUX
+
+  GTEST_DEATH_TEST_CHECK_(child_pid != -1);
+  return child_pid;
+}
+
+// The AssumeRole process for a fork-and-exec death test.  It re-executes the
+// main program from the beginning, setting the --gtest_filter
+// and --gtest_internal_run_death_test flags to cause only the current
+// death test to be re-run.
+DeathTest::TestRole ExecDeathTest::AssumeRole() {
+  const UnitTestImpl* const impl = GetUnitTestImpl();
+  const InternalRunDeathTestFlag* const flag =
+      impl->internal_run_death_test_flag();
+  const TestInfo* const info = impl->current_test_info();
+  const int death_test_index = info->result()->death_test_count();
+
+  if (flag != NULL) {
+    set_write_fd(flag->write_fd());
+    return EXECUTE_TEST;
+  }
+
+  int pipe_fd[2];
+  GTEST_DEATH_TEST_CHECK_(pipe(pipe_fd) != -1);
+  // Clear the close-on-exec flag on the write end of the pipe, lest
+  // it be closed when the child process does an exec:
+  GTEST_DEATH_TEST_CHECK_(fcntl(pipe_fd[1], F_SETFD, 0) != -1);
+
+  const std::string filter_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kFilterFlag + "="
+      + info->test_case_name() + "." + info->name();
+  const std::string internal_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kInternalRunDeathTestFlag + "="
+      + file_ + "|" + StreamableToString(line_) + "|"
+      + StreamableToString(death_test_index) + "|"
+      + StreamableToString(pipe_fd[1]);
+  Arguments args;
+  args.AddArguments(GetArgvsForDeathTestChildProcess());
+  args.AddArgument(filter_flag.c_str());
+  args.AddArgument(internal_flag.c_str());
+
+  DeathTest::set_last_death_test_message("");
+
+  CaptureStderr();
+  // See the comment in NoExecDeathTest::AssumeRole for why the next line
+  // is necessary.
+  FlushInfoLog();
+
+  const pid_t child_pid = ExecDeathTestSpawnChild(args.Argv(), pipe_fd[0]);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(close(pipe_fd[1]));
+  set_child_pid(child_pid);
+  set_read_fd(pipe_fd[0]);
+  set_spawned(true);
+  return OVERSEE_TEST;
+}
+
+# endif  // !GTEST_OS_WINDOWS
+
+// Creates a concrete DeathTest-derived class that depends on the
+// --gtest_death_test_style flag, and sets the pointer pointed to
+// by the "test" argument to its address.  If the test should be
+// skipped, sets that pointer to NULL.  Returns true, unless the
+// flag is set to an invalid value.
+bool DefaultDeathTestFactory::Create(const char* statement, const RE* regex,
+                                     const char* file, int line,
+                                     DeathTest** test) {
+  UnitTestImpl* const impl = GetUnitTestImpl();
+  const InternalRunDeathTestFlag* const flag =
+      impl->internal_run_death_test_flag();
+  const int death_test_index = impl->current_test_info()
+      ->increment_death_test_count();
+
+  if (flag != NULL) {
+    if (death_test_index > flag->index()) {
+      DeathTest::set_last_death_test_message(
+          "Death test count (" + StreamableToString(death_test_index)
+          + ") somehow exceeded expected maximum ("
+          + StreamableToString(flag->index()) + ")");
+      return false;
+    }
+
+    if (!(flag->file() == file && flag->line() == line &&
+          flag->index() == death_test_index)) {
+      *test = NULL;
+      return true;
+    }
+  }
+
+# if GTEST_OS_WINDOWS
+
+  if (GTEST_FLAG(death_test_style) == "threadsafe" ||
+      GTEST_FLAG(death_test_style) == "fast") {
+    *test = new WindowsDeathTest(statement, regex, file, line);
+  }
+
+# else
+
+  if (GTEST_FLAG(death_test_style) == "threadsafe") {
+    *test = new ExecDeathTest(statement, regex, file, line);
+  } else if (GTEST_FLAG(death_test_style) == "fast") {
+    *test = new NoExecDeathTest(statement, regex);
+  }
+
+# endif  // GTEST_OS_WINDOWS
+
+  else {  // NOLINT - this is more readable than unbalanced brackets inside #if.
+    DeathTest::set_last_death_test_message(
+        "Unknown death test style \"" + GTEST_FLAG(death_test_style)
+        + "\" encountered");
+    return false;
+  }
+
+  return true;
+}
+
+// Splits a given string on a given delimiter, populating a given
+// vector with the fields.  GTEST_HAS_DEATH_TEST implies that we have
+// ::std::string, so we can use it here.
+static void SplitString(const ::std::string& str, char delimiter,
+                        ::std::vector< ::std::string>* dest) {
+  ::std::vector< ::std::string> parsed;
+  ::std::string::size_type pos = 0;
+  while (::testing::internal::AlwaysTrue()) {
+    const ::std::string::size_type colon = str.find(delimiter, pos);
+    if (colon == ::std::string::npos) {
+      parsed.push_back(str.substr(pos));
+      break;
+    } else {
+      parsed.push_back(str.substr(pos, colon - pos));
+      pos = colon + 1;
+    }
+  }
+  dest->swap(parsed);
+}
+
+# if GTEST_OS_WINDOWS
+// Recreates the pipe and event handles from the provided parameters,
+// signals the event, and returns a file descriptor wrapped around the pipe
+// handle. This function is called in the child process only.
+int GetStatusFileDescriptor(unsigned int parent_process_id,
+                            size_t write_handle_as_size_t,
+                            size_t event_handle_as_size_t) {
+  AutoHandle parent_process_handle(::OpenProcess(PROCESS_DUP_HANDLE,
+                                                   FALSE,  // Non-inheritable.
+                                                   parent_process_id));
+  if (parent_process_handle.Get() == INVALID_HANDLE_VALUE) {
+    DeathTestAbort("Unable to open parent process " +
+                   StreamableToString(parent_process_id));
+  }
+
+  // TODO(vladl@google.com): Replace the following check with a
+  // compile-time assertion when available.
+  GTEST_CHECK_(sizeof(HANDLE) <= sizeof(size_t));
+
+  const HANDLE write_handle =
+      reinterpret_cast<HANDLE>(write_handle_as_size_t);
+  HANDLE dup_write_handle;
+
+  // The newly initialized handle is accessible only in in the parent
+  // process. To obtain one accessible within the child, we need to use
+  // DuplicateHandle.
+  if (!::DuplicateHandle(parent_process_handle.Get(), write_handle,
+                         ::GetCurrentProcess(), &dup_write_handle,
+                         0x0,    // Requested privileges ignored since
+                                 // DUPLICATE_SAME_ACCESS is used.
+                         FALSE,  // Request non-inheritable handler.
+                         DUPLICATE_SAME_ACCESS)) {
+    DeathTestAbort("Unable to duplicate the pipe handle " +
+                   StreamableToString(write_handle_as_size_t) +
+                   " from the parent process " +
+                   StreamableToString(parent_process_id));
+  }
+
+  const HANDLE event_handle = reinterpret_cast<HANDLE>(event_handle_as_size_t);
+  HANDLE dup_event_handle;
+
+  if (!::DuplicateHandle(parent_process_handle.Get(), event_handle,
+                         ::GetCurrentProcess(), &dup_event_handle,
+                         0x0,
+                         FALSE,
+                         DUPLICATE_SAME_ACCESS)) {
+    DeathTestAbort("Unable to duplicate the event handle " +
+                   StreamableToString(event_handle_as_size_t) +
+                   " from the parent process " +
+                   StreamableToString(parent_process_id));
+  }
+
+  const int write_fd =
+      ::_open_osfhandle(reinterpret_cast<intptr_t>(dup_write_handle), O_APPEND);
+  if (write_fd == -1) {
+    DeathTestAbort("Unable to convert pipe handle " +
+                   StreamableToString(write_handle_as_size_t) +
+                   " to a file descriptor");
+  }
+
+  // Signals the parent that the write end of the pipe has been acquired
+  // so the parent can release its own write end.
+  ::SetEvent(dup_event_handle);
+
+  return write_fd;
+}
+# endif  // GTEST_OS_WINDOWS
+
+// Returns a newly created InternalRunDeathTestFlag object with fields
+// initialized from the GTEST_FLAG(internal_run_death_test) flag if
+// the flag is specified; otherwise returns NULL.
+InternalRunDeathTestFlag* ParseInternalRunDeathTestFlag() {
+  if (GTEST_FLAG(internal_run_death_test) == "") return NULL;
+
+  // GTEST_HAS_DEATH_TEST implies that we have ::std::string, so we
+  // can use it here.
+  int line = -1;
+  int index = -1;
+  ::std::vector< ::std::string> fields;
+  SplitString(GTEST_FLAG(internal_run_death_test).c_str(), '|', &fields);
+  int write_fd = -1;
+
+# if GTEST_OS_WINDOWS
+
+  unsigned int parent_process_id = 0;
+  size_t write_handle_as_size_t = 0;
+  size_t event_handle_as_size_t = 0;
+
+  if (fields.size() != 6
+      || !ParseNaturalNumber(fields[1], &line)
+      || !ParseNaturalNumber(fields[2], &index)
+      || !ParseNaturalNumber(fields[3], &parent_process_id)
+      || !ParseNaturalNumber(fields[4], &write_handle_as_size_t)
+      || !ParseNaturalNumber(fields[5], &event_handle_as_size_t)) {
+    DeathTestAbort("Bad --gtest_internal_run_death_test flag: " +
+                   GTEST_FLAG(internal_run_death_test));
+  }
+  write_fd = GetStatusFileDescriptor(parent_process_id,
+                                     write_handle_as_size_t,
+                                     event_handle_as_size_t);
+# else
+
+  if (fields.size() != 4
+      || !ParseNaturalNumber(fields[1], &line)
+      || !ParseNaturalNumber(fields[2], &index)
+      || !ParseNaturalNumber(fields[3], &write_fd)) {
+    DeathTestAbort("Bad --gtest_internal_run_death_test flag: "
+        + GTEST_FLAG(internal_run_death_test));
+  }
+
+# endif  // GTEST_OS_WINDOWS
+
+  return new InternalRunDeathTestFlag(fields[0], line, index, write_fd);
+}
+
+}  // namespace internal
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+}  // namespace testing
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: keith.ray@gmail.com (Keith Ray)
+
+
+#include <stdlib.h>
+
+#if GTEST_OS_WINDOWS_MOBILE
+# include <windows.h>
+#elif GTEST_OS_WINDOWS
+# include <direct.h>
+# include <io.h>
+#elif GTEST_OS_SYMBIAN
+// Symbian OpenC has PATH_MAX in sys/syslimits.h
+# include <sys/syslimits.h>
+#else
+# include <limits.h>
+# include <climits>  // Some Linux distributions define PATH_MAX here.
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+#if GTEST_OS_WINDOWS
+# define GTEST_PATH_MAX_ _MAX_PATH
+#elif defined(PATH_MAX)
+# define GTEST_PATH_MAX_ PATH_MAX
+#elif defined(_XOPEN_PATH_MAX)
+# define GTEST_PATH_MAX_ _XOPEN_PATH_MAX
+#else
+# define GTEST_PATH_MAX_ _POSIX_PATH_MAX
+#endif  // GTEST_OS_WINDOWS
+
+
+namespace testing {
+namespace internal {
+
+#if GTEST_OS_WINDOWS
+// On Windows, '\\' is the standard path separator, but many tools and the
+// Windows API also accept '/' as an alternate path separator. Unless otherwise
+// noted, a file path can contain either kind of path separators, or a mixture
+// of them.
+const char kPathSeparator = '\\';
+const char kAlternatePathSeparator = '/';
+const char kAlternatePathSeparatorString[] = "/";
+# if GTEST_OS_WINDOWS_MOBILE
+// Windows CE doesn't have a current directory. You should not use
+// the current directory in tests on Windows CE, but this at least
+// provides a reasonable fallback.
+const char kCurrentDirectoryString[] = "\\";
+// Windows CE doesn't define INVALID_FILE_ATTRIBUTES
+const DWORD kInvalidFileAttributes = 0xffffffff;
+# else
+const char kCurrentDirectoryString[] = ".\\";
+# endif  // GTEST_OS_WINDOWS_MOBILE
+#else
+const char kPathSeparator = '/';
+const char kCurrentDirectoryString[] = "./";
+#endif  // GTEST_OS_WINDOWS
+
+// Returns whether the given character is a valid path separator.
+static bool IsPathSeparator(char c) {
+#if GTEST_HAS_ALT_PATH_SEP_
+  return (c == kPathSeparator) || (c == kAlternatePathSeparator);
+#else
+  return c == kPathSeparator;
+#endif
+}
+
+// Returns the current working directory, or "" if unsuccessful.
+FilePath FilePath::GetCurrentDir() {
+#if GTEST_OS_WINDOWS_MOBILE
+  // Windows CE doesn't have a current directory, so we just return
+  // something reasonable.
+  return FilePath(kCurrentDirectoryString);
+#elif GTEST_OS_WINDOWS
+  char cwd[GTEST_PATH_MAX_ + 1] = { '\0' };
+  return FilePath(_getcwd(cwd, sizeof(cwd)) == NULL ? "" : cwd);
+#else
+  char cwd[GTEST_PATH_MAX_ + 1] = { '\0' };
+  return FilePath(getcwd(cwd, sizeof(cwd)) == NULL ? "" : cwd);
+#endif  // GTEST_OS_WINDOWS_MOBILE
+}
+
+// Returns a copy of the FilePath with the case-insensitive extension removed.
+// Example: FilePath("dir/file.exe").RemoveExtension("EXE") returns
+// FilePath("dir/file"). If a case-insensitive extension is not
+// found, returns a copy of the original FilePath.
+FilePath FilePath::RemoveExtension(const char* extension) const {
+  const std::string dot_extension = std::string(".") + extension;
+  if (String::EndsWithCaseInsensitive(pathname_, dot_extension)) {
+    return FilePath(pathname_.substr(
+        0, pathname_.length() - dot_extension.length()));
+  }
+  return *this;
+}
+
+// Returns a pointer to the last occurence of a valid path separator in
+// the FilePath. On Windows, for example, both '/' and '\' are valid path
+// separators. Returns NULL if no path separator was found.
+const char* FilePath::FindLastPathSeparator() const {
+  const char* const last_sep = strrchr(c_str(), kPathSeparator);
+#if GTEST_HAS_ALT_PATH_SEP_
+  const char* const last_alt_sep = strrchr(c_str(), kAlternatePathSeparator);
+  // Comparing two pointers of which only one is NULL is undefined.
+  if (last_alt_sep != NULL &&
+      (last_sep == NULL || last_alt_sep > last_sep)) {
+    return last_alt_sep;
+  }
+#endif
+  return last_sep;
+}
+
+// Returns a copy of the FilePath with the directory part removed.
+// Example: FilePath("path/to/file").RemoveDirectoryName() returns
+// FilePath("file"). If there is no directory part ("just_a_file"), it returns
+// the FilePath unmodified. If there is no file part ("just_a_dir/") it
+// returns an empty FilePath ("").
+// On Windows platform, '\' is the path separator, otherwise it is '/'.
+FilePath FilePath::RemoveDirectoryName() const {
+  const char* const last_sep = FindLastPathSeparator();
+  return last_sep ? FilePath(last_sep + 1) : *this;
+}
+
+// RemoveFileName returns the directory path with the filename removed.
+// Example: FilePath("path/to/file").RemoveFileName() returns "path/to/".
+// If the FilePath is "a_file" or "/a_file", RemoveFileName returns
+// FilePath("./") or, on Windows, FilePath(".\\"). If the filepath does
+// not have a file, like "just/a/dir/", it returns the FilePath unmodified.
+// On Windows platform, '\' is the path separator, otherwise it is '/'.
+FilePath FilePath::RemoveFileName() const {
+  const char* const last_sep = FindLastPathSeparator();
+  std::string dir;
+  if (last_sep) {
+    dir = std::string(c_str(), last_sep + 1 - c_str());
+  } else {
+    dir = kCurrentDirectoryString;
+  }
+  return FilePath(dir);
+}
+
+// Helper functions for naming files in a directory for xml output.
+
+// Given directory = "dir", base_name = "test", number = 0,
+// extension = "xml", returns "dir/test.xml". If number is greater
+// than zero (e.g., 12), returns "dir/test_12.xml".
+// On Windows platform, uses \ as the separator rather than /.
+FilePath FilePath::MakeFileName(const FilePath& directory,
+                                const FilePath& base_name,
+                                int number,
+                                const char* extension) {
+  std::string file;
+  if (number == 0) {
+    file = base_name.string() + "." + extension;
+  } else {
+    file = base_name.string() + "_" + StreamableToString(number)
+        + "." + extension;
+  }
+  return ConcatPaths(directory, FilePath(file));
+}
+
+// Given directory = "dir", relative_path = "test.xml", returns "dir/test.xml".
+// On Windows, uses \ as the separator rather than /.
+FilePath FilePath::ConcatPaths(const FilePath& directory,
+                               const FilePath& relative_path) {
+  if (directory.IsEmpty())
+    return relative_path;
+  const FilePath dir(directory.RemoveTrailingPathSeparator());
+  return FilePath(dir.string() + kPathSeparator + relative_path.string());
+}
+
+// Returns true if pathname describes something findable in the file-system,
+// either a file, directory, or whatever.
+bool FilePath::FileOrDirectoryExists() const {
+#if GTEST_OS_WINDOWS_MOBILE
+  LPCWSTR unicode = String::AnsiToUtf16(pathname_.c_str());
+  const DWORD attributes = GetFileAttributes(unicode);
+  delete [] unicode;
+  return attributes != kInvalidFileAttributes;
+#else
+  posix::StatStruct file_stat;
+  return posix::Stat(pathname_.c_str(), &file_stat) == 0;
+#endif  // GTEST_OS_WINDOWS_MOBILE
+}
+
+// Returns true if pathname describes a directory in the file-system
+// that exists.
+bool FilePath::DirectoryExists() const {
+  bool result = false;
+#if GTEST_OS_WINDOWS
+  // Don't strip off trailing separator if path is a root directory on
+  // Windows (like "C:\\").
+  const FilePath& path(IsRootDirectory() ? *this :
+                                           RemoveTrailingPathSeparator());
+#else
+  const FilePath& path(*this);
+#endif
+
+#if GTEST_OS_WINDOWS_MOBILE
+  LPCWSTR unicode = String::AnsiToUtf16(path.c_str());
+  const DWORD attributes = GetFileAttributes(unicode);
+  delete [] unicode;
+  if ((attributes != kInvalidFileAttributes) &&
+      (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+    result = true;
+  }
+#else
+  posix::StatStruct file_stat;
+  result = posix::Stat(path.c_str(), &file_stat) == 0 &&
+      posix::IsDir(file_stat);
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+  return result;
+}
+
+// Returns true if pathname describes a root directory. (Windows has one
+// root directory per disk drive.)
+bool FilePath::IsRootDirectory() const {
+#if GTEST_OS_WINDOWS
+  // TODO(wan@google.com): on Windows a network share like
+  // \\server\share can be a root directory, although it cannot be the
+  // current directory.  Handle this properly.
+  return pathname_.length() == 3 && IsAbsolutePath();
+#else
+  return pathname_.length() == 1 && IsPathSeparator(pathname_.c_str()[0]);
+#endif
+}
+
+// Returns true if pathname describes an absolute path.
+bool FilePath::IsAbsolutePath() const {
+  const char* const name = pathname_.c_str();
+#if GTEST_OS_WINDOWS
+  return pathname_.length() >= 3 &&
+     ((name[0] >= 'a' && name[0] <= 'z') ||
+      (name[0] >= 'A' && name[0] <= 'Z')) &&
+     name[1] == ':' &&
+     IsPathSeparator(name[2]);
+#else
+  return IsPathSeparator(name[0]);
+#endif
+}
+
+// Returns a pathname for a file that does not currently exist. The pathname
+// will be directory/base_name.extension or
+// directory/base_name_<number>.extension if directory/base_name.extension
+// already exists. The number will be incremented until a pathname is found
+// that does not already exist.
+// Examples: 'dir/foo_test.xml' or 'dir/foo_test_1.xml'.
+// There could be a race condition if two or more processes are calling this
+// function at the same time -- they could both pick the same filename.
+FilePath FilePath::GenerateUniqueFileName(const FilePath& directory,
+                                          const FilePath& base_name,
+                                          const char* extension) {
+  FilePath full_pathname;
+  int number = 0;
+  do {
+    full_pathname.Set(MakeFileName(directory, base_name, number++, extension));
+  } while (full_pathname.FileOrDirectoryExists());
+  return full_pathname;
+}
+
+// Returns true if FilePath ends with a path separator, which indicates that
+// it is intended to represent a directory. Returns false otherwise.
+// This does NOT check that a directory (or file) actually exists.
+bool FilePath::IsDirectory() const {
+  return !pathname_.empty() &&
+         IsPathSeparator(pathname_.c_str()[pathname_.length() - 1]);
+}
+
+// Create directories so that path exists. Returns true if successful or if
+// the directories already exist; returns false if unable to create directories
+// for any reason.
+bool FilePath::CreateDirectoriesRecursively() const {
+  if (!this->IsDirectory()) {
+    return false;
+  }
+
+  if (pathname_.length() == 0 || this->DirectoryExists()) {
+    return true;
+  }
+
+  const FilePath parent(this->RemoveTrailingPathSeparator().RemoveFileName());
+  return parent.CreateDirectoriesRecursively() && this->CreateFolder();
+}
+
+// Create the directory so that path exists. Returns true if successful or
+// if the directory already exists; returns false if unable to create the
+// directory for any reason, including if the parent directory does not
+// exist. Not named "CreateDirectory" because that's a macro on Windows.
+bool FilePath::CreateFolder() const {
+#if GTEST_OS_WINDOWS_MOBILE
+  FilePath removed_sep(this->RemoveTrailingPathSeparator());
+  LPCWSTR unicode = String::AnsiToUtf16(removed_sep.c_str());
+  int result = CreateDirectory(unicode, NULL) ? 0 : -1;
+  delete [] unicode;
+#elif GTEST_OS_WINDOWS
+  int result = _mkdir(pathname_.c_str());
+#else
+  int result = mkdir(pathname_.c_str(), 0777);
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+  if (result == -1) {
+    return this->DirectoryExists();  // An error is OK if the directory exists.
+  }
+  return true;  // No error.
+}
+
+// If input name has a trailing separator character, remove it and return the
+// name, otherwise return the name string unmodified.
+// On Windows platform, uses \ as the separator, other platforms use /.
+FilePath FilePath::RemoveTrailingPathSeparator() const {
+  return IsDirectory()
+      ? FilePath(pathname_.substr(0, pathname_.length() - 1))
+      : *this;
+}
+
+// Removes any redundant separators that might be in the pathname.
+// For example, "bar///foo" becomes "bar/foo". Does not eliminate other
+// redundancies that might be in a pathname involving "." or "..".
+// TODO(wan@google.com): handle Windows network shares (e.g. \\server\share).
+void FilePath::Normalize() {
+  if (pathname_.c_str() == NULL) {
+    pathname_ = "";
+    return;
+  }
+  const char* src = pathname_.c_str();
+  char* const dest = new char[pathname_.length() + 1];
+  char* dest_ptr = dest;
+  memset(dest_ptr, 0, pathname_.length() + 1);
+
+  while (*src != '\0') {
+    *dest_ptr = *src;
+    if (!IsPathSeparator(*src)) {
+      src++;
+    } else {
+#if GTEST_HAS_ALT_PATH_SEP_
+      if (*dest_ptr == kAlternatePathSeparator) {
+        *dest_ptr = kPathSeparator;
+      }
+#endif
+      while (IsPathSeparator(*src))
+        src++;
+    }
+    dest_ptr++;
+  }
+  *dest_ptr = '\0';
+  pathname_ = dest;
+  delete[] dest;
+}
+
+}  // namespace internal
+}  // namespace testing
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+
+#include <limits.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+
+#if GTEST_OS_WINDOWS_MOBILE
+# include <windows.h>  // For TerminateProcess()
+#elif GTEST_OS_WINDOWS
+# include <io.h>
+# include <sys/stat.h>
+#else
+# include <unistd.h>
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+#if GTEST_OS_MAC
+# include <mach/mach_init.h>
+# include <mach/task.h>
+# include <mach/vm_map.h>
+#endif  // GTEST_OS_MAC
+
+#if GTEST_OS_QNX
+# include <devctl.h>
+# include <sys/procfs.h>
+#endif  // GTEST_OS_QNX
+
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick is to
+// prevent a user from accidentally including gtest-internal-inl.h in
+// his code.
+#define GTEST_IMPLEMENTATION_ 1
+#undef GTEST_IMPLEMENTATION_
+
+namespace testing {
+namespace internal {
+
+#if defined(_MSC_VER) || defined(__BORLANDC__)
+// MSVC and C++Builder do not provide a definition of STDERR_FILENO.
+const int kStdOutFileno = 1;
+const int kStdErrFileno = 2;
+#else
+const int kStdOutFileno = STDOUT_FILENO;
+const int kStdErrFileno = STDERR_FILENO;
+#endif  // _MSC_VER
+
+#if GTEST_OS_MAC
+
+// Returns the number of threads running in the process, or 0 to indicate that
+// we cannot detect it.
+size_t GetThreadCount() {
+  const task_t task = mach_task_self();
+  mach_msg_type_number_t thread_count;
+  thread_act_array_t thread_list;
+  const kern_return_t status = task_threads(task, &thread_list, &thread_count);
+  if (status == KERN_SUCCESS) {
+    // task_threads allocates resources in thread_list and we need to free them
+    // to avoid leaks.
+    vm_deallocate(task,
+                  reinterpret_cast<vm_address_t>(thread_list),
+                  sizeof(thread_t) * thread_count);
+    return static_cast<size_t>(thread_count);
+  } else {
+    return 0;
+  }
+}
+
+#elif GTEST_OS_QNX
+
+// Returns the number of threads running in the process, or 0 to indicate that
+// we cannot detect it.
+size_t GetThreadCount() {
+  const int fd = open("/proc/self/as", O_RDONLY);
+  if (fd < 0) {
+    return 0;
+  }
+  procfs_info process_info;
+  const int status =
+      devctl(fd, DCMD_PROC_INFO, &process_info, sizeof(process_info), NULL);
+  close(fd);
+  if (status == EOK) {
+    return static_cast<size_t>(process_info.num_threads);
+  } else {
+    return 0;
+  }
+}
+
+#else
+
+size_t GetThreadCount() {
+  // There's no portable way to detect the number of threads, so we just
+  // return 0 to indicate that we cannot detect it.
+  return 0;
+}
+
+#endif  // GTEST_OS_MAC
+
+#if GTEST_USES_POSIX_RE
+
+// Implements RE.  Currently only needed for death tests.
+
+RE::~RE() {
+  if (is_valid_) {
+    // regfree'ing an invalid regex might crash because the content
+    // of the regex is undefined. Since the regex's are essentially
+    // the same, one cannot be valid (or invalid) without the other
+    // being so too.
+    regfree(&partial_regex_);
+    regfree(&full_regex_);
+  }
+  free(const_cast<char*>(pattern_));
+}
+
+// Returns true iff regular expression re matches the entire str.
+bool RE::FullMatch(const char* str, const RE& re) {
+  if (!re.is_valid_) return false;
+
+  regmatch_t match;
+  return regexec(&re.full_regex_, str, 1, &match, 0) == 0;
+}
+
+// Returns true iff regular expression re matches a substring of str
+// (including str itself).
+bool RE::PartialMatch(const char* str, const RE& re) {
+  if (!re.is_valid_) return false;
+
+  regmatch_t match;
+  return regexec(&re.partial_regex_, str, 1, &match, 0) == 0;
+}
+
+// Initializes an RE from its string representation.
+void RE::Init(const char* regex) {
+  pattern_ = posix::StrDup(regex);
+
+  // Reserves enough bytes to hold the regular expression used for a
+  // full match.
+  const size_t full_regex_len = strlen(regex) + 10;
+  char* const full_pattern = new char[full_regex_len];
+
+  snprintf(full_pattern, full_regex_len, "^(%s)$", regex);
+  is_valid_ = regcomp(&full_regex_, full_pattern, REG_EXTENDED) == 0;
+  // We want to call regcomp(&partial_regex_, ...) even if the
+  // previous expression returns false.  Otherwise partial_regex_ may
+  // not be properly initialized can may cause trouble when it's
+  // freed.
+  //
+  // Some implementation of POSIX regex (e.g. on at least some
+  // versions of Cygwin) doesn't accept the empty string as a valid
+  // regex.  We change it to an equivalent form "()" to be safe.
+  if (is_valid_) {
+    const char* const partial_regex = (*regex == '\0') ? "()" : regex;
+    is_valid_ = regcomp(&partial_regex_, partial_regex, REG_EXTENDED) == 0;
+  }
+  EXPECT_TRUE(is_valid_)
+      << "Regular expression \"" << regex
+      << "\" is not a valid POSIX Extended regular expression.";
+
+  delete[] full_pattern;
+}
+
+#elif GTEST_USES_SIMPLE_RE
+
+// Returns true iff ch appears anywhere in str (excluding the
+// terminating '\0' character).
+bool IsInSet(char ch, const char* str) {
+  return ch != '\0' && strchr(str, ch) != NULL;
+}
+
+// Returns true iff ch belongs to the given classification.  Unlike
+// similar functions in <ctype.h>, these aren't affected by the
+// current locale.
+bool IsAsciiDigit(char ch) { return '0' <= ch && ch <= '9'; }
+bool IsAsciiPunct(char ch) {
+  return IsInSet(ch, "^-!\"#$%&'()*+,./:;<=>?@[\\]_`{|}~");
+}
+bool IsRepeat(char ch) { return IsInSet(ch, "?*+"); }
+bool IsAsciiWhiteSpace(char ch) { return IsInSet(ch, " \f\n\r\t\v"); }
+bool IsAsciiWordChar(char ch) {
+  return ('a' <= ch && ch <= 'z') || ('A' <= ch && ch <= 'Z') ||
+      ('0' <= ch && ch <= '9') || ch == '_';
+}
+
+// Returns true iff "\\c" is a supported escape sequence.
+bool IsValidEscape(char c) {
+  return (IsAsciiPunct(c) || IsInSet(c, "dDfnrsStvwW"));
+}
+
+// Returns true iff the given atom (specified by escaped and pattern)
+// matches ch.  The result is undefined if the atom is invalid.
+bool AtomMatchesChar(bool escaped, char pattern_char, char ch) {
+  if (escaped) {  // "\\p" where p is pattern_char.
+    switch (pattern_char) {
+      case 'd': return IsAsciiDigit(ch);
+      case 'D': return !IsAsciiDigit(ch);
+      case 'f': return ch == '\f';
+      case 'n': return ch == '\n';
+      case 'r': return ch == '\r';
+      case 's': return IsAsciiWhiteSpace(ch);
+      case 'S': return !IsAsciiWhiteSpace(ch);
+      case 't': return ch == '\t';
+      case 'v': return ch == '\v';
+      case 'w': return IsAsciiWordChar(ch);
+      case 'W': return !IsAsciiWordChar(ch);
+    }
+    return IsAsciiPunct(pattern_char) && pattern_char == ch;
+  }
+
+  return (pattern_char == '.' && ch != '\n') || pattern_char == ch;
+}
+
+// Helper function used by ValidateRegex() to format error messages.
+std::string FormatRegexSyntaxError(const char* regex, int index) {
+  return (Message() << "Syntax error at index " << index
+          << " in simple regular expression \"" << regex << "\": ").GetString();
+}
+
+// Generates non-fatal failures and returns false if regex is invalid;
+// otherwise returns true.
+bool ValidateRegex(const char* regex) {
+  if (regex == NULL) {
+    // TODO(wan@google.com): fix the source file location in the
+    // assertion failures to match where the regex is used in user
+    // code.
+    ADD_FAILURE() << "NULL is not a valid simple regular expression.";
+    return false;
+  }
+
+  bool is_valid = true;
+
+  // True iff ?, *, or + can follow the previous atom.
+  bool prev_repeatable = false;
+  for (int i = 0; regex[i]; i++) {
+    if (regex[i] == '\\') {  // An escape sequence
+      i++;
+      if (regex[i] == '\0') {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i - 1)
+                      << "'\\' cannot appear at the end.";
+        return false;
+      }
+
+      if (!IsValidEscape(regex[i])) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i - 1)
+                      << "invalid escape sequence \"\\" << regex[i] << "\".";
+        is_valid = false;
+      }
+      prev_repeatable = true;
+    } else {  // Not an escape sequence.
+      const char ch = regex[i];
+
+      if (ch == '^' && i > 0) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'^' can only appear at the beginning.";
+        is_valid = false;
+      } else if (ch == '$' && regex[i + 1] != '\0') {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'$' can only appear at the end.";
+        is_valid = false;
+      } else if (IsInSet(ch, "()[]{}|")) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'" << ch << "' is unsupported.";
+        is_valid = false;
+      } else if (IsRepeat(ch) && !prev_repeatable) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'" << ch << "' can only follow a repeatable token.";
+        is_valid = false;
+      }
+
+      prev_repeatable = !IsInSet(ch, "^$?*+");
+    }
+  }
+
+  return is_valid;
+}
+
+// Matches a repeated regex atom followed by a valid simple regular
+// expression.  The regex atom is defined as c if escaped is false,
+// or \c otherwise.  repeat is the repetition meta character (?, *,
+// or +).  The behavior is undefined if str contains too many
+// characters to be indexable by size_t, in which case the test will
+// probably time out anyway.  We are fine with this limitation as
+// std::string has it too.
+bool MatchRepetitionAndRegexAtHead(
+    bool escaped, char c, char repeat, const char* regex,
+    const char* str) {
+  const size_t min_count = (repeat == '+') ? 1 : 0;
+  const size_t max_count = (repeat == '?') ? 1 :
+      static_cast<size_t>(-1) - 1;
+  // We cannot call numeric_limits::max() as it conflicts with the
+  // max() macro on Windows.
+
+  for (size_t i = 0; i <= max_count; ++i) {
+    // We know that the atom matches each of the first i characters in str.
+    if (i >= min_count && MatchRegexAtHead(regex, str + i)) {
+      // We have enough matches at the head, and the tail matches too.
+      // Since we only care about *whether* the pattern matches str
+      // (as opposed to *how* it matches), there is no need to find a
+      // greedy match.
+      return true;
+    }
+    if (str[i] == '\0' || !AtomMatchesChar(escaped, c, str[i]))
+      return false;
+  }
+  return false;
+}
+
+// Returns true iff regex matches a prefix of str.  regex must be a
+// valid simple regular expression and not start with "^", or the
+// result is undefined.
+bool MatchRegexAtHead(const char* regex, const char* str) {
+  if (*regex == '\0')  // An empty regex matches a prefix of anything.
+    return true;
+
+  // "$" only matches the end of a string.  Note that regex being
+  // valid guarantees that there's nothing after "$" in it.
+  if (*regex == '$')
+    return *str == '\0';
+
+  // Is the first thing in regex an escape sequence?
+  const bool escaped = *regex == '\\';
+  if (escaped)
+    ++regex;
+  if (IsRepeat(regex[1])) {
+    // MatchRepetitionAndRegexAtHead() calls MatchRegexAtHead(), so
+    // here's an indirect recursion.  It terminates as the regex gets
+    // shorter in each recursion.
+    return MatchRepetitionAndRegexAtHead(
+        escaped, regex[0], regex[1], regex + 2, str);
+  } else {
+    // regex isn't empty, isn't "$", and doesn't start with a
+    // repetition.  We match the first atom of regex with the first
+    // character of str and recurse.
+    return (*str != '\0') && AtomMatchesChar(escaped, *regex, *str) &&
+        MatchRegexAtHead(regex + 1, str + 1);
+  }
+}
+
+// Returns true iff regex matches any substring of str.  regex must be
+// a valid simple regular expression, or the result is undefined.
+//
+// The algorithm is recursive, but the recursion depth doesn't exceed
+// the regex length, so we won't need to worry about running out of
+// stack space normally.  In rare cases the time complexity can be
+// exponential with respect to the regex length + the string length,
+// but usually it's must faster (often close to linear).
+bool MatchRegexAnywhere(const char* regex, const char* str) {
+  if (regex == NULL || str == NULL)
+    return false;
+
+  if (*regex == '^')
+    return MatchRegexAtHead(regex + 1, str);
+
+  // A successful match can be anywhere in str.
+  do {
+    if (MatchRegexAtHead(regex, str))
+      return true;
+  } while (*str++ != '\0');
+  return false;
+}
+
+// Implements the RE class.
+
+RE::~RE() {
+  free(const_cast<char*>(pattern_));
+  free(const_cast<char*>(full_pattern_));
+}
+
+// Returns true iff regular expression re matches the entire str.
+bool RE::FullMatch(const char* str, const RE& re) {
+  return re.is_valid_ && MatchRegexAnywhere(re.full_pattern_, str);
+}
+
+// Returns true iff regular expression re matches a substring of str
+// (including str itself).
+bool RE::PartialMatch(const char* str, const RE& re) {
+  return re.is_valid_ && MatchRegexAnywhere(re.pattern_, str);
+}
+
+// Initializes an RE from its string representation.
+void RE::Init(const char* regex) {
+  pattern_ = full_pattern_ = NULL;
+  if (regex != NULL) {
+    pattern_ = posix::StrDup(regex);
+  }
+
+  is_valid_ = ValidateRegex(regex);
+  if (!is_valid_) {
+    // No need to calculate the full pattern when the regex is invalid.
+    return;
+  }
+
+  const size_t len = strlen(regex);
+  // Reserves enough bytes to hold the regular expression used for a
+  // full match: we need space to prepend a '^', append a '$', and
+  // terminate the string with '\0'.
+  char* buffer = static_cast<char*>(malloc(len + 3));
+  full_pattern_ = buffer;
+
+  if (*regex != '^')
+    *buffer++ = '^';  // Makes sure full_pattern_ starts with '^'.
+
+  // We don't use snprintf or strncpy, as they trigger a warning when
+  // compiled with VC++ 8.0.
+  memcpy(buffer, regex, len);
+  buffer += len;
+
+  if (len == 0 || regex[len - 1] != '$')
+    *buffer++ = '$';  // Makes sure full_pattern_ ends with '$'.
+
+  *buffer = '\0';
+}
+
+#endif  // GTEST_USES_POSIX_RE
+
+const char kUnknownFile[] = "unknown file";
+
+// Formats a source file path and a line number as they would appear
+// in an error message from the compiler used to compile this code.
+GTEST_API_ ::std::string FormatFileLocation(const char* file, int line) {
+  const std::string file_name(file == NULL ? kUnknownFile : file);
+
+  if (line < 0) {
+    return file_name + ":";
+  }
+#ifdef _MSC_VER
+  return file_name + "(" + StreamableToString(line) + "):";
+#else
+  return file_name + ":" + StreamableToString(line) + ":";
+#endif  // _MSC_VER
+}
+
+// Formats a file location for compiler-independent XML output.
+// Although this function is not platform dependent, we put it next to
+// FormatFileLocation in order to contrast the two functions.
+// Note that FormatCompilerIndependentFileLocation() does NOT append colon
+// to the file location it produces, unlike FormatFileLocation().
+GTEST_API_ ::std::string FormatCompilerIndependentFileLocation(
+    const char* file, int line) {
+  const std::string file_name(file == NULL ? kUnknownFile : file);
+
+  if (line < 0)
+    return file_name;
+  else
+    return file_name + ":" + StreamableToString(line);
+}
+
+
+GTestLog::GTestLog(GTestLogSeverity severity, const char* file, int line)
+    : severity_(severity) {
+  const char* const marker =
+      severity == GTEST_INFO ?    "[  INFO ]" :
+      severity == GTEST_WARNING ? "[WARNING]" :
+      severity == GTEST_ERROR ?   "[ ERROR ]" : "[ FATAL ]";
+  GetStream() << ::std::endl << marker << " "
+              << FormatFileLocation(file, line).c_str() << ": ";
+}
+
+// Flushes the buffers and, if severity is GTEST_FATAL, aborts the program.
+GTestLog::~GTestLog() {
+  GetStream() << ::std::endl;
+  if (severity_ == GTEST_FATAL) {
+    fflush(stderr);
+    posix::Abort();
+  }
+}
+// Disable Microsoft deprecation warnings for POSIX functions called from
+// this class (creat, dup, dup2, and close)
+#ifdef _MSC_VER
+# pragma warning(push)
+# pragma warning(disable: 4996)
+#endif  // _MSC_VER
+
+#if GTEST_HAS_STREAM_REDIRECTION
+
+// Object that captures an output stream (stdout/stderr).
+class CapturedStream {
+ public:
+  // The ctor redirects the stream to a temporary file.
+  explicit CapturedStream(int fd) : fd_(fd), uncaptured_fd_(dup(fd)) {
+# if GTEST_OS_WINDOWS
+    char temp_dir_path[MAX_PATH + 1] = { '\0' };  // NOLINT
+    char temp_file_path[MAX_PATH + 1] = { '\0' };  // NOLINT
+
+    ::GetTempPathA(sizeof(temp_dir_path), temp_dir_path);
+    const UINT success = ::GetTempFileNameA(temp_dir_path,
+                                            "gtest_redir",
+                                            0,  // Generate unique file name.
+                                            temp_file_path);
+    GTEST_CHECK_(success != 0)
+        << "Unable to create a temporary file in " << temp_dir_path;
+    const int captured_fd = creat(temp_file_path, _S_IREAD | _S_IWRITE);
+    GTEST_CHECK_(captured_fd != -1) << "Unable to open temporary file "
+                                    << temp_file_path;
+    filename_ = temp_file_path;
+# else
+    // There's no guarantee that a test has write access to the current
+    // directory, so we create the temporary file in the /tmp directory
+    // instead. We use /tmp on most systems, and /sdcard on Android.
+    // That's because Android doesn't have /tmp.
+#  if GTEST_OS_LINUX_ANDROID
+    // Note: Android applications are expected to call the framework's
+    // Context.getExternalStorageDirectory() method through JNI to get
+    // the location of the world-writable SD Card directory. However,
+    // this requires a Context handle, which cannot be retrieved
+    // globally from native code. Doing so also precludes running the
+    // code as part of a regular standalone executable, which doesn't
+    // run in a Dalvik process (e.g. when running it through 'adb shell').
+    //
+    // The location /sdcard is directly accessible from native code
+    // and is the only location (unofficially) supported by the Android
+    // team. It's generally a symlink to the real SD Card mount point
+    // which can be /mnt/sdcard, /mnt/sdcard0, /system/media/sdcard, or
+    // other OEM-customized locations. Never rely on these, and always
+    // use /sdcard.
+    char name_template[] = "/sdcard/gtest_captured_stream.XXXXXX";
+#  else
+    char name_template[] = "/tmp/captured_stream.XXXXXX";
+#  endif  // GTEST_OS_LINUX_ANDROID
+    const int captured_fd = mkstemp(name_template);
+    filename_ = name_template;
+# endif  // GTEST_OS_WINDOWS
+    fflush(NULL);
+    dup2(captured_fd, fd_);
+    close(captured_fd);
+  }
+
+  ~CapturedStream() {
+    remove(filename_.c_str());
+  }
+
+  std::string GetCapturedString() {
+    if (uncaptured_fd_ != -1) {
+      // Restores the original stream.
+      fflush(NULL);
+      dup2(uncaptured_fd_, fd_);
+      close(uncaptured_fd_);
+      uncaptured_fd_ = -1;
+    }
+
+    FILE* const file = posix::FOpen(filename_.c_str(), "r");
+    const std::string content = ReadEntireFile(file);
+    posix::FClose(file);
+    return content;
+  }
+
+ private:
+  // Reads the entire content of a file as an std::string.
+  static std::string ReadEntireFile(FILE* file);
+
+  // Returns the size (in bytes) of a file.
+  static size_t GetFileSize(FILE* file);
+
+  const int fd_;  // A stream to capture.
+  int uncaptured_fd_;
+  // Name of the temporary file holding the stderr output.
+  ::std::string filename_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(CapturedStream);
+};
+
+// Returns the size (in bytes) of a file.
+size_t CapturedStream::GetFileSize(FILE* file) {
+  fseek(file, 0, SEEK_END);
+  return static_cast<size_t>(ftell(file));
+}
+
+// Reads the entire content of a file as a string.
+std::string CapturedStream::ReadEntireFile(FILE* file) {
+  const size_t file_size = GetFileSize(file);
+  char* const buffer = new char[file_size];
+
+  size_t bytes_last_read = 0;  // # of bytes read in the last fread()
+  size_t bytes_read = 0;       // # of bytes read so far
+
+  fseek(file, 0, SEEK_SET);
+
+  // Keeps reading the file until we cannot read further or the
+  // pre-determined file size is reached.
+  do {
+    bytes_last_read = fread(buffer+bytes_read, 1, file_size-bytes_read, file);
+    bytes_read += bytes_last_read;
+  } while (bytes_last_read > 0 && bytes_read < file_size);
+
+  const std::string content(buffer, bytes_read);
+  delete[] buffer;
+
+  return content;
+}
+
+# ifdef _MSC_VER
+#  pragma warning(pop)
+# endif  // _MSC_VER
+
+static CapturedStream* g_captured_stderr = NULL;
+static CapturedStream* g_captured_stdout = NULL;
+
+// Starts capturing an output stream (stdout/stderr).
+void CaptureStream(int fd, const char* stream_name, CapturedStream** stream) {
+  if (*stream != NULL) {
+    GTEST_LOG_(FATAL) << "Only one " << stream_name
+                      << " capturer can exist at a time.";
+  }
+  *stream = new CapturedStream(fd);
+}
+
+// Stops capturing the output stream and returns the captured string.
+std::string GetCapturedStream(CapturedStream** captured_stream) {
+  const std::string content = (*captured_stream)->GetCapturedString();
+
+  delete *captured_stream;
+  *captured_stream = NULL;
+
+  return content;
+}
+
+// Starts capturing stdout.
+void CaptureStdout() {
+  CaptureStream(kStdOutFileno, "stdout", &g_captured_stdout);
+}
+
+// Starts capturing stderr.
+void CaptureStderr() {
+  CaptureStream(kStdErrFileno, "stderr", &g_captured_stderr);
+}
+
+// Stops capturing stdout and returns the captured string.
+std::string GetCapturedStdout() {
+  return GetCapturedStream(&g_captured_stdout);
+}
+
+// Stops capturing stderr and returns the captured string.
+std::string GetCapturedStderr() {
+  return GetCapturedStream(&g_captured_stderr);
+}
+
+#endif  // GTEST_HAS_STREAM_REDIRECTION
+
+#if GTEST_HAS_DEATH_TEST
+
+// A copy of all command line arguments.  Set by InitGoogleTest().
+::std::vector<testing::internal::string> g_argvs;
+
+static const ::std::vector<testing::internal::string>* g_injected_test_argvs =
+                                        NULL;  // Owned.
+
+void SetInjectableArgvs(const ::std::vector<testing::internal::string>* argvs) {
+  if (g_injected_test_argvs != argvs)
+    delete g_injected_test_argvs;
+  g_injected_test_argvs = argvs;
+}
+
+const ::std::vector<testing::internal::string>& GetInjectableArgvs() {
+  if (g_injected_test_argvs != NULL) {
+    return *g_injected_test_argvs;
+  }
+  return g_argvs;
+}
+#endif  // GTEST_HAS_DEATH_TEST
+
+#if GTEST_OS_WINDOWS_MOBILE
+namespace posix {
+void Abort() {
+  DebugBreak();
+  TerminateProcess(GetCurrentProcess(), 1);
+}
+}  // namespace posix
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+// Returns the name of the environment variable corresponding to the
+// given flag.  For example, FlagToEnvVar("foo") will return
+// "GTEST_FOO" in the open-source version.
+static std::string FlagToEnvVar(const char* flag) {
+  const std::string full_flag =
+      (Message() << GTEST_FLAG_PREFIX_ << flag).GetString();
+
+  Message env_var;
+  for (size_t i = 0; i != full_flag.length(); i++) {
+    env_var << ToUpper(full_flag.c_str()[i]);
+  }
+
+  return env_var.GetString();
+}
+
+// Parses 'str' for a 32-bit signed integer.  If successful, writes
+// the result to *value and returns true; otherwise leaves *value
+// unchanged and returns false.
+bool ParseInt32(const Message& src_text, const char* str, Int32* value) {
+  // Parses the environment variable as a decimal integer.
+  char* end = NULL;
+  const long long_value = strtol(str, &end, 10);  // NOLINT
+
+  // Has strtol() consumed all characters in the string?
+  if (*end != '\0') {
+    // No - an invalid character was encountered.
+    Message msg;
+    msg << "WARNING: " << src_text
+        << " is expected to be a 32-bit integer, but actually"
+        << " has value \"" << str << "\".\n";
+    printf("%s", msg.GetString().c_str());
+    fflush(stdout);
+    return false;
+  }
+
+  // Is the parsed value in the range of an Int32?
+  const Int32 result = static_cast<Int32>(long_value);
+  if (long_value == LONG_MAX || long_value == LONG_MIN ||
+      // The parsed value overflows as a long.  (strtol() returns
+      // LONG_MAX or LONG_MIN when the input overflows.)
+      result != long_value
+      // The parsed value overflows as an Int32.
+      ) {
+    Message msg;
+    msg << "WARNING: " << src_text
+        << " is expected to be a 32-bit integer, but actually"
+        << " has value " << str << ", which overflows.\n";
+    printf("%s", msg.GetString().c_str());
+    fflush(stdout);
+    return false;
+  }
+
+  *value = result;
+  return true;
+}
+
+// Reads and returns the Boolean environment variable corresponding to
+// the given flag; if it's not set, returns default_value.
+//
+// The value is considered true iff it's not "0".
+bool BoolFromGTestEnv(const char* flag, bool default_value) {
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const string_value = posix::GetEnv(env_var.c_str());
+  return string_value == NULL ?
+      default_value : strcmp(string_value, "0") != 0;
+}
+
+// Reads and returns a 32-bit integer stored in the environment
+// variable corresponding to the given flag; if it isn't set or
+// doesn't represent a valid 32-bit integer, returns default_value.
+Int32 Int32FromGTestEnv(const char* flag, Int32 default_value) {
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const string_value = posix::GetEnv(env_var.c_str());
+  if (string_value == NULL) {
+    // The environment variable is not set.
+    return default_value;
+  }
+
+  Int32 result = default_value;
+  if (!ParseInt32(Message() << "Environment variable " << env_var,
+                  string_value, &result)) {
+    printf("The default value %s is used.\n",
+           (Message() << default_value).GetString().c_str());
+    fflush(stdout);
+    return default_value;
+  }
+
+  return result;
+}
+
+// Reads and returns the string environment variable corresponding to
+// the given flag; if it's not set, returns default_value.
+const char* StringFromGTestEnv(const char* flag, const char* default_value) {
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const value = posix::GetEnv(env_var.c_str());
+  return value == NULL ? default_value : value;
+}
+
+}  // namespace internal
+}  // namespace testing
+// Copyright 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Google Test - The Google C++ Testing Framework
+//
+// This file implements a universal value printer that can print a
+// value of any type T:
+//
+//   void ::testing::internal::UniversalPrinter<T>::Print(value, ostream_ptr);
+//
+// It uses the << operator when possible, and prints the bytes in the
+// object otherwise.  A user can override its behavior for a class
+// type Foo by defining either operator<<(::std::ostream&, const Foo&)
+// or void PrintTo(const Foo&, ::std::ostream*) in the namespace that
+// defines Foo.
+
+#include <ctype.h>
+#include <stdio.h>
+#include <ostream>  // NOLINT
+#include <string>
+
+namespace testing {
+
+namespace {
+
+using ::std::ostream;
+
+// Prints a segment of bytes in the given object.
+void PrintByteSegmentInObjectTo(const unsigned char* obj_bytes, size_t start,
+                                size_t count, ostream* os) {
+  char text[5] = "";
+  for (size_t i = 0; i != count; i++) {
+    const size_t j = start + i;
+    if (i != 0) {
+      // Organizes the bytes into groups of 2 for easy parsing by
+      // human.
+      if ((j % 2) == 0)
+        *os << ' ';
+      else
+        *os << '-';
+    }
+    GTEST_SNPRINTF_(text, sizeof(text), "%02X", obj_bytes[j]);
+    *os << text;
+  }
+}
+
+// Prints the bytes in the given value to the given ostream.
+void PrintBytesInObjectToImpl(const unsigned char* obj_bytes, size_t count,
+                              ostream* os) {
+  // Tells the user how big the object is.
+  *os << count << "-byte object <";
+
+  const size_t kThreshold = 132;
+  const size_t kChunkSize = 64;
+  // If the object size is bigger than kThreshold, we'll have to omit
+  // some details by printing only the first and the last kChunkSize
+  // bytes.
+  // TODO(wan): let the user control the threshold using a flag.
+  if (count < kThreshold) {
+    PrintByteSegmentInObjectTo(obj_bytes, 0, count, os);
+  } else {
+    PrintByteSegmentInObjectTo(obj_bytes, 0, kChunkSize, os);
+    *os << " ... ";
+    // Rounds up to 2-byte boundary.
+    const size_t resume_pos = (count - kChunkSize + 1)/2*2;
+    PrintByteSegmentInObjectTo(obj_bytes, resume_pos, count - resume_pos, os);
+  }
+  *os << ">";
+}
+
+}  // namespace
+
+namespace internal2 {
+
+// Delegates to PrintBytesInObjectToImpl() to print the bytes in the
+// given object.  The delegation simplifies the implementation, which
+// uses the << operator and thus is easier done outside of the
+// ::testing::internal namespace, which contains a << operator that
+// sometimes conflicts with the one in STL.
+void PrintBytesInObjectTo(const unsigned char* obj_bytes, size_t count,
+                          ostream* os) {
+  PrintBytesInObjectToImpl(obj_bytes, count, os);
+}
+
+}  // namespace internal2
+
+namespace internal {
+
+// Depending on the value of a char (or wchar_t), we print it in one
+// of three formats:
+//   - as is if it's a printable ASCII (e.g. 'a', '2', ' '),
+//   - as a hexidecimal escape sequence (e.g. '\x7F'), or
+//   - as a special escape sequence (e.g. '\r', '\n').
+enum CharFormat {
+  kAsIs,
+  kHexEscape,
+  kSpecialEscape
+};
+
+// Returns true if c is a printable ASCII character.  We test the
+// value of c directly instead of calling isprint(), which is buggy on
+// Windows Mobile.
+inline bool IsPrintableAscii(wchar_t c) {
+  return 0x20 <= c && c <= 0x7E;
+}
+
+// Prints a wide or narrow char c as a character literal without the
+// quotes, escaping it when necessary; returns how c was formatted.
+// The template argument UnsignedChar is the unsigned version of Char,
+// which is the type of c.
+template <typename UnsignedChar, typename Char>
+static CharFormat PrintAsCharLiteralTo(Char c, ostream* os) {
+  switch (static_cast<wchar_t>(c)) {
+    case L'\0':
+      *os << "\\0";
+      break;
+    case L'\'':
+      *os << "\\'";
+      break;
+    case L'\\':
+      *os << "\\\\";
+      break;
+    case L'\a':
+      *os << "\\a";
+      break;
+    case L'\b':
+      *os << "\\b";
+      break;
+    case L'\f':
+      *os << "\\f";
+      break;
+    case L'\n':
+      *os << "\\n";
+      break;
+    case L'\r':
+      *os << "\\r";
+      break;
+    case L'\t':
+      *os << "\\t";
+      break;
+    case L'\v':
+      *os << "\\v";
+      break;
+    default:
+      if (IsPrintableAscii(c)) {
+        *os << static_cast<char>(c);
+        return kAsIs;
+      } else {
+        *os << "\\x" + String::FormatHexInt(static_cast<UnsignedChar>(c));
+        return kHexEscape;
+      }
+  }
+  return kSpecialEscape;
+}
+
+// Prints a wchar_t c as if it's part of a string literal, escaping it when
+// necessary; returns how c was formatted.
+static CharFormat PrintAsStringLiteralTo(wchar_t c, ostream* os) {
+  switch (c) {
+    case L'\'':
+      *os << "'";
+      return kAsIs;
+    case L'"':
+      *os << "\\\"";
+      return kSpecialEscape;
+    default:
+      return PrintAsCharLiteralTo<wchar_t>(c, os);
+  }
+}
+
+// Prints a char c as if it's part of a string literal, escaping it when
+// necessary; returns how c was formatted.
+static CharFormat PrintAsStringLiteralTo(char c, ostream* os) {
+  return PrintAsStringLiteralTo(
+      static_cast<wchar_t>(static_cast<unsigned char>(c)), os);
+}
+
+// Prints a wide or narrow character c and its code.  '\0' is printed
+// as "'\\0'", other unprintable characters are also properly escaped
+// using the standard C++ escape sequence.  The template argument
+// UnsignedChar is the unsigned version of Char, which is the type of c.
+template <typename UnsignedChar, typename Char>
+void PrintCharAndCodeTo(Char c, ostream* os) {
+  // First, print c as a literal in the most readable form we can find.
+  *os << ((sizeof(c) > 1) ? "L'" : "'");
+  const CharFormat format = PrintAsCharLiteralTo<UnsignedChar>(c, os);
+  *os << "'";
+
+  // To aid user debugging, we also print c's code in decimal, unless
+  // it's 0 (in which case c was printed as '\\0', making the code
+  // obvious).
+  if (c == 0)
+    return;
+  *os << " (" << static_cast<int>(c);
+
+  // For more convenience, we print c's code again in hexidecimal,
+  // unless c was already printed in the form '\x##' or the code is in
+  // [1, 9].
+  if (format == kHexEscape || (1 <= c && c <= 9)) {
+    // Do nothing.
+  } else {
+    *os << ", 0x" << String::FormatHexInt(static_cast<UnsignedChar>(c));
+  }
+  *os << ")";
+}
+
+void PrintTo(unsigned char c, ::std::ostream* os) {
+  PrintCharAndCodeTo<unsigned char>(c, os);
+}
+void PrintTo(signed char c, ::std::ostream* os) {
+  PrintCharAndCodeTo<unsigned char>(c, os);
+}
+
+// Prints a wchar_t as a symbol if it is printable or as its internal
+// code otherwise and also as its code.  L'\0' is printed as "L'\\0'".
+void PrintTo(wchar_t wc, ostream* os) {
+  PrintCharAndCodeTo<wchar_t>(wc, os);
+}
+
+// Prints the given array of characters to the ostream.  CharType must be either
+// char or wchar_t.
+// The array starts at begin, the length is len, it may include '\0' characters
+// and may not be NUL-terminated.
+template <typename CharType>
+static void PrintCharsAsStringTo(
+    const CharType* begin, size_t len, ostream* os) {
+  const char* const kQuoteBegin = sizeof(CharType) == 1 ? "\"" : "L\"";
+  *os << kQuoteBegin;
+  bool is_previous_hex = false;
+  for (size_t index = 0; index < len; ++index) {
+    const CharType cur = begin[index];
+    if (is_previous_hex && IsXDigit(cur)) {
+      // Previous character is of '\x..' form and this character can be
+      // interpreted as another hexadecimal digit in its number. Break string to
+      // disambiguate.
+      *os << "\" " << kQuoteBegin;
+    }
+    is_previous_hex = PrintAsStringLiteralTo(cur, os) == kHexEscape;
+  }
+  *os << "\"";
+}
+
+// Prints a (const) char/wchar_t array of 'len' elements, starting at address
+// 'begin'.  CharType must be either char or wchar_t.
+template <typename CharType>
+static void UniversalPrintCharArray(
+    const CharType* begin, size_t len, ostream* os) {
+  // The code
+  //   const char kFoo[] = "foo";
+  // generates an array of 4, not 3, elements, with the last one being '\0'.
+  //
+  // Therefore when printing a char array, we don't print the last element if
+  // it's '\0', such that the output matches the string literal as it's
+  // written in the source code.
+  if (len > 0 && begin[len - 1] == '\0') {
+    PrintCharsAsStringTo(begin, len - 1, os);
+    return;
+  }
+
+  // If, however, the last element in the array is not '\0', e.g.
+  //    const char kFoo[] = { 'f', 'o', 'o' };
+  // we must print the entire array.  We also print a message to indicate
+  // that the array is not NUL-terminated.
+  PrintCharsAsStringTo(begin, len, os);
+  *os << " (no terminating NUL)";
+}
+
+// Prints a (const) char array of 'len' elements, starting at address 'begin'.
+void UniversalPrintArray(const char* begin, size_t len, ostream* os) {
+  UniversalPrintCharArray(begin, len, os);
+}
+
+// Prints a (const) wchar_t array of 'len' elements, starting at address
+// 'begin'.
+void UniversalPrintArray(const wchar_t* begin, size_t len, ostream* os) {
+  UniversalPrintCharArray(begin, len, os);
+}
+
+// Prints the given C string to the ostream.
+void PrintTo(const char* s, ostream* os) {
+  if (s == NULL) {
+    *os << "NULL";
+  } else {
+    *os << ImplicitCast_<const void*>(s) << " pointing to ";
+    PrintCharsAsStringTo(s, strlen(s), os);
+  }
+}
+
+// MSVC compiler can be configured to define whar_t as a typedef
+// of unsigned short. Defining an overload for const wchar_t* in that case
+// would cause pointers to unsigned shorts be printed as wide strings,
+// possibly accessing more memory than intended and causing invalid
+// memory accesses. MSVC defines _NATIVE_WCHAR_T_DEFINED symbol when
+// wchar_t is implemented as a native type.
+#if !defined(_MSC_VER) || defined(_NATIVE_WCHAR_T_DEFINED)
+// Prints the given wide C string to the ostream.
+void PrintTo(const wchar_t* s, ostream* os) {
+  if (s == NULL) {
+    *os << "NULL";
+  } else {
+    *os << ImplicitCast_<const void*>(s) << " pointing to ";
+    PrintCharsAsStringTo(s, wcslen(s), os);
+  }
+}
+#endif  // wchar_t is native
+
+// Prints a ::string object.
+#if GTEST_HAS_GLOBAL_STRING
+void PrintStringTo(const ::string& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+void PrintStringTo(const ::std::string& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+
+// Prints a ::wstring object.
+#if GTEST_HAS_GLOBAL_WSTRING
+void PrintWideStringTo(const ::wstring& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+#if GTEST_HAS_STD_WSTRING
+void PrintWideStringTo(const ::std::wstring& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+}  // namespace internal
+
+}  // namespace testing
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: mheule@google.com (Markus Heule)
+//
+// The Google C++ Testing Framework (Google Test)
+
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick is to
+// prevent a user from accidentally including gtest-internal-inl.h in
+// his code.
+#define GTEST_IMPLEMENTATION_ 1
+#undef GTEST_IMPLEMENTATION_
+
+namespace testing {
+
+using internal::GetUnitTestImpl;
+
+// Gets the summary of the failure message by omitting the stack trace
+// in it.
+std::string TestPartResult::ExtractSummary(const char* message) {
+  const char* const stack_trace = strstr(message, internal::kStackTraceMarker);
+  return stack_trace == NULL ? message :
+      std::string(message, stack_trace);
+}
+
+// Prints a TestPartResult object.
+std::ostream& operator<<(std::ostream& os, const TestPartResult& result) {
+  return os
+      << result.file_name() << ":" << result.line_number() << ": "
+      << (result.type() == TestPartResult::kSuccess ? "Success" :
+          result.type() == TestPartResult::kFatalFailure ? "Fatal failure" :
+          "Non-fatal failure") << ":\n"
+      << result.message() << std::endl;
+}
+
+// Appends a TestPartResult to the array.
+void TestPartResultArray::Append(const TestPartResult& result) {
+  array_.push_back(result);
+}
+
+// Returns the TestPartResult at the given index (0-based).
+const TestPartResult& TestPartResultArray::GetTestPartResult(int index) const {
+  if (index < 0 || index >= size()) {
+    printf("\nInvalid index (%d) into TestPartResultArray.\n", index);
+    internal::posix::Abort();
+  }
+
+  return array_[index];
+}
+
+// Returns the number of TestPartResult objects in the array.
+int TestPartResultArray::size() const {
+  return static_cast<int>(array_.size());
+}
+
+namespace internal {
+
+HasNewFatalFailureHelper::HasNewFatalFailureHelper()
+    : has_new_fatal_failure_(false),
+      original_reporter_(GetUnitTestImpl()->
+                         GetTestPartResultReporterForCurrentThread()) {
+  GetUnitTestImpl()->SetTestPartResultReporterForCurrentThread(this);
+}
+
+HasNewFatalFailureHelper::~HasNewFatalFailureHelper() {
+  GetUnitTestImpl()->SetTestPartResultReporterForCurrentThread(
+      original_reporter_);
+}
+
+void HasNewFatalFailureHelper::ReportTestPartResult(
+    const TestPartResult& result) {
+  if (result.fatally_failed())
+    has_new_fatal_failure_ = true;
+  original_reporter_->ReportTestPartResult(result);
+}
+
+}  // namespace internal
+
+}  // namespace testing
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+
+namespace testing {
+namespace internal {
+
+#if GTEST_HAS_TYPED_TEST_P
+
+// Skips to the first non-space char in str. Returns an empty string if str
+// contains only whitespace characters.
+static const char* SkipSpaces(const char* str) {
+  while (IsSpace(*str))
+    str++;
+  return str;
+}
+
+// Verifies that registered_tests match the test names in
+// defined_test_names_; returns registered_tests if successful, or
+// aborts the program otherwise.
+const char* TypedTestCasePState::VerifyRegisteredTestNames(
+    const char* file, int line, const char* registered_tests) {
+  typedef ::std::set<const char*>::const_iterator DefinedTestIter;
+  registered_ = true;
+
+  // Skip initial whitespace in registered_tests since some
+  // preprocessors prefix stringizied literals with whitespace.
+  registered_tests = SkipSpaces(registered_tests);
+
+  Message errors;
+  ::std::set<std::string> tests;
+  for (const char* names = registered_tests; names != NULL;
+       names = SkipComma(names)) {
+    const std::string name = GetPrefixUntilComma(names);
+    if (tests.count(name) != 0) {
+      errors << "Test " << name << " is listed more than once.\n";
+      continue;
+    }
+
+    bool found = false;
+    for (DefinedTestIter it = defined_test_names_.begin();
+         it != defined_test_names_.end();
+         ++it) {
+      if (name == *it) {
+        found = true;
+        break;
+      }
+    }
+
+    if (found) {
+      tests.insert(name);
+    } else {
+      errors << "No test named " << name
+             << " can be found in this test case.\n";
+    }
+  }
+
+  for (DefinedTestIter it = defined_test_names_.begin();
+       it != defined_test_names_.end();
+       ++it) {
+    if (tests.count(*it) == 0) {
+      errors << "You forgot to list test " << *it << ".\n";
+    }
+  }
+
+  const std::string& errors_str = errors.GetString();
+  if (errors_str != "") {
+    fprintf(stderr, "%s %s", FormatFileLocation(file, line).c_str(),
+            errors_str.c_str());
+    fflush(stderr);
+    posix::Abort();
+  }
+
+  return registered_tests;
+}
+
+#endif  // GTEST_HAS_TYPED_TEST_P
+
+}  // namespace internal
+}  // namespace testing
diff --git a/libvpx/libvpx/third_party/googletest/src/src/gtest_main.cc b/libvpx/libvpx/third_party/googletest/src/src/gtest_main.cc
new file mode 100644
index 0000000..f302822
--- /dev/null
+++ b/libvpx/libvpx/third_party/googletest/src/src/gtest_main.cc
@@ -0,0 +1,38 @@
+// Copyright 2006, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+#include <stdio.h>
+
+#include "gtest/gtest.h"
+
+GTEST_API_ int main(int argc, char **argv) {
+  printf("Running main() from gtest_main.cc\n");
+  testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/libvpx/libvpx/third_party/libwebm/AUTHORS.TXT b/libvpx/libvpx/third_party/libwebm/AUTHORS.TXT
new file mode 100644
index 0000000..8ab6f79
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/AUTHORS.TXT
@@ -0,0 +1,4 @@
+# Names should be added to this file like so:
+# Name or Organization <email address>
+
+Google Inc.
diff --git a/libvpx/libvpx/third_party/libwebm/Android.mk b/libvpx/libvpx/third_party/libwebm/Android.mk
new file mode 100644
index 0000000..8149a08
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/Android.mk
@@ -0,0 +1,17 @@
+LOCAL_PATH:= $(call my-dir)
+
+include $(CLEAR_VARS)
+LOCAL_MODULE:= libwebm
+LOCAL_CPPFLAGS:=-D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS
+LOCAL_CPPFLAGS+=-D__STDC_LIMIT_MACROS -Wno-extern-c-compat
+LOCAL_C_INCLUDES:= $(LOCAL_PATH)
+LOCAL_EXPORT_C_INCLUDES:= $(LOCAL_PATH)
+
+LOCAL_SRC_FILES:= common/file_util.cc \
+                  common/hdr_util.cc \
+                  mkvparser/mkvparser.cc \
+                  mkvparser/mkvreader.cc \
+                  mkvmuxer/mkvmuxer.cc \
+                  mkvmuxer/mkvmuxerutil.cc \
+                  mkvmuxer/mkvwriter.cc
+include $(BUILD_STATIC_LIBRARY)
diff --git a/libvpx/libvpx/third_party/libwebm/LICENSE.TXT b/libvpx/libvpx/third_party/libwebm/LICENSE.TXT
new file mode 100644
index 0000000..7a6f995
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/LICENSE.TXT
@@ -0,0 +1,30 @@
+Copyright (c) 2010, Google Inc. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+  * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+
+  * Redistributions in binary form must reproduce the above copyright
+    notice, this list of conditions and the following disclaimer in
+    the documentation and/or other materials provided with the
+    distribution.
+
+  * Neither the name of Google nor the names of its contributors may
+    be used to endorse or promote products derived from this software
+    without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
diff --git a/libvpx/libvpx/third_party/libwebm/PATENTS.TXT b/libvpx/libvpx/third_party/libwebm/PATENTS.TXT
new file mode 100644
index 0000000..caedf60
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/PATENTS.TXT
@@ -0,0 +1,23 @@
+Additional IP Rights Grant (Patents)
+------------------------------------
+
+"These implementations" means the copyrightable works that implement the WebM
+codecs distributed by Google as part of the WebM Project.
+
+Google hereby grants to you a perpetual, worldwide, non-exclusive, no-charge,
+royalty-free, irrevocable (except as stated in this section) patent license to
+make, have made, use, offer to sell, sell, import, transfer, and otherwise
+run, modify and propagate the contents of these implementations of WebM, where
+such license applies only to those patent claims, both currently owned by
+Google and acquired in the future, licensable by Google that are necessarily
+infringed by these implementations of WebM. This grant does not include claims
+that would be infringed only as a consequence of further modification of these
+implementations. If you or your agent or exclusive licensee institute or order
+or agree to the institution of patent litigation or any other patent
+enforcement activity against any entity (including a cross-claim or
+counterclaim in a lawsuit) alleging that any of these implementations of WebM
+or any code incorporated within any of these implementations of WebM
+constitute direct or contributory patent infringement, or inducement of
+patent infringement, then any patent rights granted to you under this License
+for these implementations of WebM shall terminate as of the date such
+litigation is filed.
diff --git a/libvpx/libvpx/third_party/libwebm/README.libvpx b/libvpx/libvpx/third_party/libwebm/README.libvpx
new file mode 100644
index 0000000..73f8303
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/README.libvpx
@@ -0,0 +1,10 @@
+URL: https://chromium.googlesource.com/webm/libwebm
+Version: 32d5ac49414a8914ec1e1f285f3f927c6e8ec29d
+License: BSD
+License File: LICENSE.txt
+
+Description:
+libwebm is used to handle WebM container I/O.
+
+Local Changes:
+* <none>
diff --git a/libvpx/libvpx/third_party/libwebm/common/file_util.cc b/libvpx/libvpx/third_party/libwebm/common/file_util.cc
new file mode 100644
index 0000000..4f91318
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/common/file_util.cc
@@ -0,0 +1,67 @@
+// Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#include "common/file_util.h"
+
+#include <sys/stat.h>
+#ifndef _MSC_VER
+#include <unistd.h>  // close()
+#endif
+
+#include <cstdio>
+#include <cstdlib>
+#include <fstream>
+#include <ios>
+
+namespace libwebm {
+
+std::string GetTempFileName() {
+#if !defined _MSC_VER && !defined __MINGW32__
+  char temp_file_name_template[] = "libwebm_temp.XXXXXX";
+  int fd = mkstemp(temp_file_name_template);
+  if (fd != -1) {
+    close(fd);
+    return std::string(temp_file_name_template);
+  }
+  return std::string();
+#else
+  char tmp_file_name[_MAX_PATH];
+  errno_t err = tmpnam_s(tmp_file_name);
+  if (err == 0) {
+    return std::string(tmp_file_name);
+  }
+  return std::string();
+#endif
+}
+
+uint64_t GetFileSize(const std::string& file_name) {
+  uint64_t file_size = 0;
+#ifndef _MSC_VER
+  struct stat st;
+  st.st_size = 0;
+  if (stat(file_name.c_str(), &st) == 0) {
+#else
+  struct _stat st;
+  st.st_size = 0;
+  if (_stat(file_name.c_str(), &st) == 0) {
+#endif
+    file_size = st.st_size;
+  }
+  return file_size;
+}
+
+TempFileDeleter::TempFileDeleter() { file_name_ = GetTempFileName(); }
+
+TempFileDeleter::~TempFileDeleter() {
+  std::ifstream file(file_name_.c_str());
+  if (file.good()) {
+    file.close();
+    std::remove(file_name_.c_str());
+  }
+}
+
+}  // namespace libwebm
diff --git a/libvpx/libvpx/third_party/libwebm/common/file_util.h b/libvpx/libvpx/third_party/libwebm/common/file_util.h
new file mode 100644
index 0000000..0e71eac
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/common/file_util.h
@@ -0,0 +1,41 @@
+// Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef LIBWEBM_COMMON_FILE_UTIL_H_
+#define LIBWEBM_COMMON_FILE_UTIL_H_
+
+#include <stdint.h>
+
+#include <string>
+
+#include "mkvmuxer/mkvmuxertypes.h"  // LIBWEBM_DISALLOW_COPY_AND_ASSIGN()
+
+namespace libwebm {
+
+// Returns a temporary file name.
+std::string GetTempFileName();
+
+// Returns size of file specified by |file_name|, or 0 upon failure.
+uint64_t GetFileSize(const std::string& file_name);
+
+// Manages life of temporary file specified at time of construction. Deletes
+// file upon destruction.
+class TempFileDeleter {
+ public:
+  TempFileDeleter();
+  explicit TempFileDeleter(std::string file_name) : file_name_(file_name) {}
+  ~TempFileDeleter();
+  const std::string& name() const { return file_name_; }
+
+ private:
+  std::string file_name_;
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(TempFileDeleter);
+};
+
+}  // namespace libwebm
+
+#endif  // LIBWEBM_COMMON_FILE_UTIL_H_
\ No newline at end of file
diff --git a/libvpx/libvpx/third_party/libwebm/common/hdr_util.cc b/libvpx/libvpx/third_party/libwebm/common/hdr_util.cc
new file mode 100644
index 0000000..e1a9842
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/common/hdr_util.cc
@@ -0,0 +1,182 @@
+// Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#include "hdr_util.h"
+
+#include <cstddef>
+#include <new>
+
+#include "mkvparser/mkvparser.h"
+
+namespace libwebm {
+bool CopyPrimaryChromaticity(const mkvparser::PrimaryChromaticity& parser_pc,
+                             PrimaryChromaticityPtr* muxer_pc) {
+  muxer_pc->reset(new (std::nothrow)
+                      mkvmuxer::PrimaryChromaticity(parser_pc.x, parser_pc.y));
+  if (!muxer_pc->get())
+    return false;
+  return true;
+}
+
+bool MasteringMetadataValuePresent(double value) {
+  return value != mkvparser::MasteringMetadata::kValueNotPresent;
+}
+
+bool CopyMasteringMetadata(const mkvparser::MasteringMetadata& parser_mm,
+                           mkvmuxer::MasteringMetadata* muxer_mm) {
+  if (MasteringMetadataValuePresent(parser_mm.luminance_max))
+    muxer_mm->luminance_max = parser_mm.luminance_max;
+  if (MasteringMetadataValuePresent(parser_mm.luminance_min))
+    muxer_mm->luminance_min = parser_mm.luminance_min;
+
+  PrimaryChromaticityPtr r_ptr(NULL);
+  PrimaryChromaticityPtr g_ptr(NULL);
+  PrimaryChromaticityPtr b_ptr(NULL);
+  PrimaryChromaticityPtr wp_ptr(NULL);
+
+  if (parser_mm.r) {
+    if (!CopyPrimaryChromaticity(*parser_mm.r, &r_ptr))
+      return false;
+  }
+  if (parser_mm.g) {
+    if (!CopyPrimaryChromaticity(*parser_mm.g, &g_ptr))
+      return false;
+  }
+  if (parser_mm.b) {
+    if (!CopyPrimaryChromaticity(*parser_mm.b, &b_ptr))
+      return false;
+  }
+  if (parser_mm.white_point) {
+    if (!CopyPrimaryChromaticity(*parser_mm.white_point, &wp_ptr))
+      return false;
+  }
+
+  if (!muxer_mm->SetChromaticity(r_ptr.get(), g_ptr.get(), b_ptr.get(),
+                                 wp_ptr.get())) {
+    return false;
+  }
+
+  return true;
+}
+
+bool ColourValuePresent(long long value) {
+  return value != mkvparser::Colour::kValueNotPresent;
+}
+
+bool CopyColour(const mkvparser::Colour& parser_colour,
+                mkvmuxer::Colour* muxer_colour) {
+  if (!muxer_colour)
+    return false;
+
+  if (ColourValuePresent(parser_colour.matrix_coefficients))
+    muxer_colour->matrix_coefficients = parser_colour.matrix_coefficients;
+  if (ColourValuePresent(parser_colour.bits_per_channel))
+    muxer_colour->bits_per_channel = parser_colour.bits_per_channel;
+  if (ColourValuePresent(parser_colour.chroma_subsampling_horz))
+    muxer_colour->chroma_subsampling_horz =
+        parser_colour.chroma_subsampling_horz;
+  if (ColourValuePresent(parser_colour.chroma_subsampling_vert))
+    muxer_colour->chroma_subsampling_vert =
+        parser_colour.chroma_subsampling_vert;
+  if (ColourValuePresent(parser_colour.cb_subsampling_horz))
+    muxer_colour->cb_subsampling_horz = parser_colour.cb_subsampling_horz;
+  if (ColourValuePresent(parser_colour.cb_subsampling_vert))
+    muxer_colour->cb_subsampling_vert = parser_colour.cb_subsampling_vert;
+  if (ColourValuePresent(parser_colour.chroma_siting_horz))
+    muxer_colour->chroma_siting_horz = parser_colour.chroma_siting_horz;
+  if (ColourValuePresent(parser_colour.chroma_siting_vert))
+    muxer_colour->chroma_siting_vert = parser_colour.chroma_siting_vert;
+  if (ColourValuePresent(parser_colour.range))
+    muxer_colour->range = parser_colour.range;
+  if (ColourValuePresent(parser_colour.transfer_characteristics))
+    muxer_colour->transfer_characteristics =
+        parser_colour.transfer_characteristics;
+  if (ColourValuePresent(parser_colour.primaries))
+    muxer_colour->primaries = parser_colour.primaries;
+  if (ColourValuePresent(parser_colour.max_cll))
+    muxer_colour->max_cll = parser_colour.max_cll;
+  if (ColourValuePresent(parser_colour.max_fall))
+    muxer_colour->max_fall = parser_colour.max_fall;
+
+  if (parser_colour.mastering_metadata) {
+    mkvmuxer::MasteringMetadata muxer_mm;
+    if (!CopyMasteringMetadata(*parser_colour.mastering_metadata, &muxer_mm))
+      return false;
+    if (!muxer_colour->SetMasteringMetadata(muxer_mm))
+      return false;
+  }
+  return true;
+}
+
+// Format of VPx private data:
+//
+//   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+//  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+//  |    ID Byte    |             Length            |               |
+//  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+               |
+//  |                                                               |
+//  :               Bytes 1..Length of Codec Feature                :
+//  |                                                               |
+//  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+//
+// ID Byte Format
+// ID byte is an unsigned byte.
+//   0 1 2 3 4 5 6 7
+//  +-+-+-+-+-+-+-+-+
+//  |X|    ID       |
+//  +-+-+-+-+-+-+-+-+
+//
+// The X bit is reserved.
+//
+// Currently only profile level is supported. ID byte must be set to 1, and
+// length must be 1. Supported values are:
+//
+//   10: Level 1
+//   11: Level 1.1
+//   20: Level 2
+//   21: Level 2.1
+//   30: Level 3
+//   31: Level 3.1
+//   40: Level 4
+//   41: Level 4.1
+//   50: Level 5
+//   51: Level 5.1
+//   52: Level 5.2
+//   60: Level 6
+//   61: Level 6.1
+//   62: Level 6.2
+//
+// See the following link for more information:
+// http://www.webmproject.org/vp9/profiles/
+int ParseVpxCodecPrivate(const uint8_t* private_data, int32_t length) {
+  const int kVpxCodecPrivateLength = 3;
+  if (!private_data || length != kVpxCodecPrivateLength)
+    return 0;
+
+  const uint8_t id_byte = *private_data;
+  if (id_byte != 1)
+    return 0;
+
+  const int kVpxProfileLength = 1;
+  const uint8_t length_byte = private_data[1];
+  if (length_byte != kVpxProfileLength)
+    return 0;
+
+  const int level = static_cast<int>(private_data[2]);
+
+  const int kNumLevels = 14;
+  const int levels[kNumLevels] = {10, 11, 20, 21, 30, 31, 40,
+                                  41, 50, 51, 52, 60, 61, 62};
+
+  for (int i = 0; i < kNumLevels; ++i) {
+    if (level == levels[i])
+      return level;
+  }
+
+  return 0;
+}
+}  // namespace libwebm
diff --git a/libvpx/libvpx/third_party/libwebm/common/hdr_util.h b/libvpx/libvpx/third_party/libwebm/common/hdr_util.h
new file mode 100644
index 0000000..d30c2b9
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/common/hdr_util.h
@@ -0,0 +1,51 @@
+// Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef LIBWEBM_COMMON_HDR_UTIL_H_
+#define LIBWEBM_COMMON_HDR_UTIL_H_
+
+#include <stdint.h>
+
+#include <memory>
+
+#include "mkvmuxer/mkvmuxer.h"
+
+namespace mkvparser {
+struct Colour;
+struct MasteringMetadata;
+struct PrimaryChromaticity;
+}  // namespace mkvparser
+
+namespace libwebm {
+// Utility types and functions for working with the Colour element and its
+// children. Copiers return true upon success. Presence functions return true
+// when the specified element is present.
+
+// TODO(tomfinegan): These should be moved to libwebm_utils once c++11 is
+// required by libwebm.
+
+typedef std::auto_ptr<mkvmuxer::PrimaryChromaticity> PrimaryChromaticityPtr;
+
+bool CopyPrimaryChromaticity(const mkvparser::PrimaryChromaticity& parser_pc,
+                             PrimaryChromaticityPtr* muxer_pc);
+
+bool MasteringMetadataValuePresent(double value);
+
+bool CopyMasteringMetadata(const mkvparser::MasteringMetadata& parser_mm,
+                           mkvmuxer::MasteringMetadata* muxer_mm);
+
+bool ColourValuePresent(long long value);
+
+bool CopyColour(const mkvparser::Colour& parser_colour,
+                mkvmuxer::Colour* muxer_colour);
+
+// Returns VP9 profile upon success or 0 upon failure.
+int ParseVpxCodecPrivate(const uint8_t* private_data, int32_t length);
+
+}  // namespace libwebm
+
+#endif  // LIBWEBM_COMMON_HDR_UTIL_H_
diff --git a/libvpx/libvpx/third_party/libwebm/common/webmids.h b/libvpx/libvpx/third_party/libwebm/common/webmids.h
new file mode 100644
index 0000000..32a0c5f
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/common/webmids.h
@@ -0,0 +1,184 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#ifndef COMMON_WEBMIDS_H_
+#define COMMON_WEBMIDS_H_
+
+namespace libwebm {
+
+enum MkvId {
+  kMkvEBML = 0x1A45DFA3,
+  kMkvEBMLVersion = 0x4286,
+  kMkvEBMLReadVersion = 0x42F7,
+  kMkvEBMLMaxIDLength = 0x42F2,
+  kMkvEBMLMaxSizeLength = 0x42F3,
+  kMkvDocType = 0x4282,
+  kMkvDocTypeVersion = 0x4287,
+  kMkvDocTypeReadVersion = 0x4285,
+  kMkvVoid = 0xEC,
+  kMkvSignatureSlot = 0x1B538667,
+  kMkvSignatureAlgo = 0x7E8A,
+  kMkvSignatureHash = 0x7E9A,
+  kMkvSignaturePublicKey = 0x7EA5,
+  kMkvSignature = 0x7EB5,
+  kMkvSignatureElements = 0x7E5B,
+  kMkvSignatureElementList = 0x7E7B,
+  kMkvSignedElement = 0x6532,
+  // segment
+  kMkvSegment = 0x18538067,
+  // Meta Seek Information
+  kMkvSeekHead = 0x114D9B74,
+  kMkvSeek = 0x4DBB,
+  kMkvSeekID = 0x53AB,
+  kMkvSeekPosition = 0x53AC,
+  // Segment Information
+  kMkvInfo = 0x1549A966,
+  kMkvTimecodeScale = 0x2AD7B1,
+  kMkvDuration = 0x4489,
+  kMkvDateUTC = 0x4461,
+  kMkvTitle = 0x7BA9,
+  kMkvMuxingApp = 0x4D80,
+  kMkvWritingApp = 0x5741,
+  // Cluster
+  kMkvCluster = 0x1F43B675,
+  kMkvTimecode = 0xE7,
+  kMkvPrevSize = 0xAB,
+  kMkvBlockGroup = 0xA0,
+  kMkvBlock = 0xA1,
+  kMkvBlockDuration = 0x9B,
+  kMkvReferenceBlock = 0xFB,
+  kMkvLaceNumber = 0xCC,
+  kMkvSimpleBlock = 0xA3,
+  kMkvBlockAdditions = 0x75A1,
+  kMkvBlockMore = 0xA6,
+  kMkvBlockAddID = 0xEE,
+  kMkvBlockAdditional = 0xA5,
+  kMkvDiscardPadding = 0x75A2,
+  // Track
+  kMkvTracks = 0x1654AE6B,
+  kMkvTrackEntry = 0xAE,
+  kMkvTrackNumber = 0xD7,
+  kMkvTrackUID = 0x73C5,
+  kMkvTrackType = 0x83,
+  kMkvFlagEnabled = 0xB9,
+  kMkvFlagDefault = 0x88,
+  kMkvFlagForced = 0x55AA,
+  kMkvFlagLacing = 0x9C,
+  kMkvDefaultDuration = 0x23E383,
+  kMkvMaxBlockAdditionID = 0x55EE,
+  kMkvName = 0x536E,
+  kMkvLanguage = 0x22B59C,
+  kMkvCodecID = 0x86,
+  kMkvCodecPrivate = 0x63A2,
+  kMkvCodecName = 0x258688,
+  kMkvCodecDelay = 0x56AA,
+  kMkvSeekPreRoll = 0x56BB,
+  // video
+  kMkvVideo = 0xE0,
+  kMkvFlagInterlaced = 0x9A,
+  kMkvStereoMode = 0x53B8,
+  kMkvAlphaMode = 0x53C0,
+  kMkvPixelWidth = 0xB0,
+  kMkvPixelHeight = 0xBA,
+  kMkvPixelCropBottom = 0x54AA,
+  kMkvPixelCropTop = 0x54BB,
+  kMkvPixelCropLeft = 0x54CC,
+  kMkvPixelCropRight = 0x54DD,
+  kMkvDisplayWidth = 0x54B0,
+  kMkvDisplayHeight = 0x54BA,
+  kMkvDisplayUnit = 0x54B2,
+  kMkvAspectRatioType = 0x54B3,
+  kMkvFrameRate = 0x2383E3,
+  // end video
+  // colour
+  kMkvColour = 0x55B0,
+  kMkvMatrixCoefficients = 0x55B1,
+  kMkvBitsPerChannel = 0x55B2,
+  kMkvChromaSubsamplingHorz = 0x55B3,
+  kMkvChromaSubsamplingVert = 0x55B4,
+  kMkvCbSubsamplingHorz = 0x55B5,
+  kMkvCbSubsamplingVert = 0x55B6,
+  kMkvChromaSitingHorz = 0x55B7,
+  kMkvChromaSitingVert = 0x55B8,
+  kMkvRange = 0x55B9,
+  kMkvTransferCharacteristics = 0x55BA,
+  kMkvPrimaries = 0x55BB,
+  kMkvMaxCLL = 0x55BC,
+  kMkvMaxFALL = 0x55BD,
+  // mastering metadata
+  kMkvMasteringMetadata = 0x55D0,
+  kMkvPrimaryRChromaticityX = 0x55D1,
+  kMkvPrimaryRChromaticityY = 0x55D2,
+  kMkvPrimaryGChromaticityX = 0x55D3,
+  kMkvPrimaryGChromaticityY = 0x55D4,
+  kMkvPrimaryBChromaticityX = 0x55D5,
+  kMkvPrimaryBChromaticityY = 0x55D6,
+  kMkvWhitePointChromaticityX = 0x55D7,
+  kMkvWhitePointChromaticityY = 0x55D8,
+  kMkvLuminanceMax = 0x55D9,
+  kMkvLuminanceMin = 0x55DA,
+  // end mastering metadata
+  // end colour
+  // audio
+  kMkvAudio = 0xE1,
+  kMkvSamplingFrequency = 0xB5,
+  kMkvOutputSamplingFrequency = 0x78B5,
+  kMkvChannels = 0x9F,
+  kMkvBitDepth = 0x6264,
+  // end audio
+  // ContentEncodings
+  kMkvContentEncodings = 0x6D80,
+  kMkvContentEncoding = 0x6240,
+  kMkvContentEncodingOrder = 0x5031,
+  kMkvContentEncodingScope = 0x5032,
+  kMkvContentEncodingType = 0x5033,
+  kMkvContentCompression = 0x5034,
+  kMkvContentCompAlgo = 0x4254,
+  kMkvContentCompSettings = 0x4255,
+  kMkvContentEncryption = 0x5035,
+  kMkvContentEncAlgo = 0x47E1,
+  kMkvContentEncKeyID = 0x47E2,
+  kMkvContentSignature = 0x47E3,
+  kMkvContentSigKeyID = 0x47E4,
+  kMkvContentSigAlgo = 0x47E5,
+  kMkvContentSigHashAlgo = 0x47E6,
+  kMkvContentEncAESSettings = 0x47E7,
+  kMkvAESSettingsCipherMode = 0x47E8,
+  kMkvAESSettingsCipherInitData = 0x47E9,
+  // end ContentEncodings
+  // Cueing Data
+  kMkvCues = 0x1C53BB6B,
+  kMkvCuePoint = 0xBB,
+  kMkvCueTime = 0xB3,
+  kMkvCueTrackPositions = 0xB7,
+  kMkvCueTrack = 0xF7,
+  kMkvCueClusterPosition = 0xF1,
+  kMkvCueBlockNumber = 0x5378,
+  // Chapters
+  kMkvChapters = 0x1043A770,
+  kMkvEditionEntry = 0x45B9,
+  kMkvChapterAtom = 0xB6,
+  kMkvChapterUID = 0x73C4,
+  kMkvChapterStringUID = 0x5654,
+  kMkvChapterTimeStart = 0x91,
+  kMkvChapterTimeEnd = 0x92,
+  kMkvChapterDisplay = 0x80,
+  kMkvChapString = 0x85,
+  kMkvChapLanguage = 0x437C,
+  kMkvChapCountry = 0x437E,
+  // Tags
+  kMkvTags = 0x1254C367,
+  kMkvTag = 0x7373,
+  kMkvSimpleTag = 0x67C8,
+  kMkvTagName = 0x45A3,
+  kMkvTagString = 0x4487
+};
+
+}  // namespace libwebm
+
+#endif  // COMMON_WEBMIDS_H_
diff --git a/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxer.cc b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxer.cc
new file mode 100644
index 0000000..c79ce24
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxer.cc
@@ -0,0 +1,3769 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#include "mkvmuxer/mkvmuxer.h"
+
+#include <cfloat>
+#include <climits>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <ctime>
+#include <memory>
+#include <new>
+#include <vector>
+
+#include "common/webmids.h"
+#include "mkvmuxer/mkvmuxerutil.h"
+#include "mkvmuxer/mkvwriter.h"
+#include "mkvparser/mkvparser.h"
+
+namespace mkvmuxer {
+
+const float MasteringMetadata::kValueNotPresent = FLT_MAX;
+const uint64_t Colour::kValueNotPresent = UINT64_MAX;
+
+namespace {
+// Deallocate the string designated by |dst|, and then copy the |src|
+// string to |dst|.  The caller owns both the |src| string and the
+// |dst| copy (hence the caller is responsible for eventually
+// deallocating the strings, either directly, or indirectly via
+// StrCpy).  Returns true if the source string was successfully copied
+// to the destination.
+bool StrCpy(const char* src, char** dst_ptr) {
+  if (dst_ptr == NULL)
+    return false;
+
+  char*& dst = *dst_ptr;
+
+  delete[] dst;
+  dst = NULL;
+
+  if (src == NULL)
+    return true;
+
+  const size_t size = strlen(src) + 1;
+
+  dst = new (std::nothrow) char[size];  // NOLINT
+  if (dst == NULL)
+    return false;
+
+  strcpy(dst, src);  // NOLINT
+  return true;
+}
+
+typedef std::auto_ptr<PrimaryChromaticity> PrimaryChromaticityPtr;
+bool CopyChromaticity(const PrimaryChromaticity* src,
+                      PrimaryChromaticityPtr* dst) {
+  if (!dst)
+    return false;
+
+  dst->reset(new (std::nothrow) PrimaryChromaticity(src->x, src->y));
+  if (!dst->get())
+    return false;
+
+  return true;
+}
+
+}  // namespace
+
+///////////////////////////////////////////////////////////////
+//
+// IMkvWriter Class
+
+IMkvWriter::IMkvWriter() {}
+
+IMkvWriter::~IMkvWriter() {}
+
+bool WriteEbmlHeader(IMkvWriter* writer, uint64_t doc_type_version) {
+  // Level 0
+  uint64_t size = EbmlElementSize(libwebm::kMkvEBMLVersion, UINT64_C(1));
+  size += EbmlElementSize(libwebm::kMkvEBMLReadVersion, UINT64_C(1));
+  size += EbmlElementSize(libwebm::kMkvEBMLMaxIDLength, UINT64_C(4));
+  size += EbmlElementSize(libwebm::kMkvEBMLMaxSizeLength, UINT64_C(8));
+  size += EbmlElementSize(libwebm::kMkvDocType, "webm");
+  size += EbmlElementSize(libwebm::kMkvDocTypeVersion, doc_type_version);
+  size += EbmlElementSize(libwebm::kMkvDocTypeReadVersion, UINT64_C(2));
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvEBML, size))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvEBMLVersion, UINT64_C(1)))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvEBMLReadVersion, UINT64_C(1)))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvEBMLMaxIDLength, UINT64_C(4)))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvEBMLMaxSizeLength, UINT64_C(8)))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvDocType, "webm"))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvDocTypeVersion, doc_type_version))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvDocTypeReadVersion, UINT64_C(2)))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlHeader(IMkvWriter* writer) {
+  return WriteEbmlHeader(writer, mkvmuxer::Segment::kDefaultDocTypeVersion);
+}
+
+bool ChunkedCopy(mkvparser::IMkvReader* source, mkvmuxer::IMkvWriter* dst,
+                 int64_t start, int64_t size) {
+  // TODO(vigneshv): Check if this is a reasonable value.
+  const uint32_t kBufSize = 2048;
+  uint8_t* buf = new uint8_t[kBufSize];
+  int64_t offset = start;
+  while (size > 0) {
+    const int64_t read_len = (size > kBufSize) ? kBufSize : size;
+    if (source->Read(offset, static_cast<long>(read_len), buf))
+      return false;
+    dst->Write(buf, static_cast<uint32_t>(read_len));
+    offset += read_len;
+    size -= read_len;
+  }
+  delete[] buf;
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Frame Class
+
+Frame::Frame()
+    : add_id_(0),
+      additional_(NULL),
+      additional_length_(0),
+      duration_(0),
+      duration_set_(false),
+      frame_(NULL),
+      is_key_(false),
+      length_(0),
+      track_number_(0),
+      timestamp_(0),
+      discard_padding_(0),
+      reference_block_timestamp_(0),
+      reference_block_timestamp_set_(false) {}
+
+Frame::~Frame() {
+  delete[] frame_;
+  delete[] additional_;
+}
+
+bool Frame::CopyFrom(const Frame& frame) {
+  delete[] frame_;
+  frame_ = NULL;
+  length_ = 0;
+  if (frame.length() > 0 && frame.frame() != NULL &&
+      !Init(frame.frame(), frame.length())) {
+    return false;
+  }
+  add_id_ = 0;
+  delete[] additional_;
+  additional_ = NULL;
+  additional_length_ = 0;
+  if (frame.additional_length() > 0 && frame.additional() != NULL &&
+      !AddAdditionalData(frame.additional(), frame.additional_length(),
+                         frame.add_id())) {
+    return false;
+  }
+  duration_ = frame.duration();
+  duration_set_ = frame.duration_set();
+  is_key_ = frame.is_key();
+  track_number_ = frame.track_number();
+  timestamp_ = frame.timestamp();
+  discard_padding_ = frame.discard_padding();
+  reference_block_timestamp_ = frame.reference_block_timestamp();
+  reference_block_timestamp_set_ = frame.reference_block_timestamp_set();
+  return true;
+}
+
+bool Frame::Init(const uint8_t* frame, uint64_t length) {
+  uint8_t* const data =
+      new (std::nothrow) uint8_t[static_cast<size_t>(length)];  // NOLINT
+  if (!data)
+    return false;
+
+  delete[] frame_;
+  frame_ = data;
+  length_ = length;
+
+  memcpy(frame_, frame, static_cast<size_t>(length_));
+  return true;
+}
+
+bool Frame::AddAdditionalData(const uint8_t* additional, uint64_t length,
+                              uint64_t add_id) {
+  uint8_t* const data =
+      new (std::nothrow) uint8_t[static_cast<size_t>(length)];  // NOLINT
+  if (!data)
+    return false;
+
+  delete[] additional_;
+  additional_ = data;
+  additional_length_ = length;
+  add_id_ = add_id;
+
+  memcpy(additional_, additional, static_cast<size_t>(additional_length_));
+  return true;
+}
+
+bool Frame::IsValid() const {
+  if (length_ == 0 || !frame_) {
+    return false;
+  }
+  if ((additional_length_ != 0 && !additional_) ||
+      (additional_ != NULL && additional_length_ == 0)) {
+    return false;
+  }
+  if (track_number_ == 0 || track_number_ > kMaxTrackNumber) {
+    return false;
+  }
+  if (!CanBeSimpleBlock() && !is_key_ && !reference_block_timestamp_set_) {
+    return false;
+  }
+  return true;
+}
+
+bool Frame::CanBeSimpleBlock() const {
+  return additional_ == NULL && discard_padding_ == 0 && duration_ == 0;
+}
+
+void Frame::set_duration(uint64_t duration) {
+  duration_ = duration;
+  duration_set_ = true;
+}
+
+void Frame::set_reference_block_timestamp(int64_t reference_block_timestamp) {
+  reference_block_timestamp_ = reference_block_timestamp;
+  reference_block_timestamp_set_ = true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// CuePoint Class
+
+CuePoint::CuePoint()
+    : time_(0),
+      track_(0),
+      cluster_pos_(0),
+      block_number_(1),
+      output_block_number_(true) {}
+
+CuePoint::~CuePoint() {}
+
+bool CuePoint::Write(IMkvWriter* writer) const {
+  if (!writer || track_ < 1 || cluster_pos_ < 1)
+    return false;
+
+  uint64_t size =
+      EbmlElementSize(libwebm::kMkvCueClusterPosition, cluster_pos_);
+  size += EbmlElementSize(libwebm::kMkvCueTrack, track_);
+  if (output_block_number_ && block_number_ > 1)
+    size += EbmlElementSize(libwebm::kMkvCueBlockNumber, block_number_);
+  const uint64_t track_pos_size =
+      EbmlMasterElementSize(libwebm::kMkvCueTrackPositions, size) + size;
+  const uint64_t payload_size =
+      EbmlElementSize(libwebm::kMkvCueTime, time_) + track_pos_size;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvCuePoint, payload_size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvCueTime, time_))
+    return false;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvCueTrackPositions, size))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvCueTrack, track_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvCueClusterPosition, cluster_pos_))
+    return false;
+  if (output_block_number_ && block_number_ > 1)
+    if (!WriteEbmlElement(writer, libwebm::kMkvCueBlockNumber, block_number_))
+      return false;
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0)
+    return false;
+
+  if (stop_position - payload_position != static_cast<int64_t>(payload_size))
+    return false;
+
+  return true;
+}
+
+uint64_t CuePoint::PayloadSize() const {
+  uint64_t size =
+      EbmlElementSize(libwebm::kMkvCueClusterPosition, cluster_pos_);
+  size += EbmlElementSize(libwebm::kMkvCueTrack, track_);
+  if (output_block_number_ && block_number_ > 1)
+    size += EbmlElementSize(libwebm::kMkvCueBlockNumber, block_number_);
+  const uint64_t track_pos_size =
+      EbmlMasterElementSize(libwebm::kMkvCueTrackPositions, size) + size;
+  const uint64_t payload_size =
+      EbmlElementSize(libwebm::kMkvCueTime, time_) + track_pos_size;
+
+  return payload_size;
+}
+
+uint64_t CuePoint::Size() const {
+  const uint64_t payload_size = PayloadSize();
+  return EbmlMasterElementSize(libwebm::kMkvCuePoint, payload_size) +
+         payload_size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Cues Class
+
+Cues::Cues()
+    : cue_entries_capacity_(0),
+      cue_entries_size_(0),
+      cue_entries_(NULL),
+      output_block_number_(true) {}
+
+Cues::~Cues() {
+  if (cue_entries_) {
+    for (int32_t i = 0; i < cue_entries_size_; ++i) {
+      CuePoint* const cue = cue_entries_[i];
+      delete cue;
+    }
+    delete[] cue_entries_;
+  }
+}
+
+bool Cues::AddCue(CuePoint* cue) {
+  if (!cue)
+    return false;
+
+  if ((cue_entries_size_ + 1) > cue_entries_capacity_) {
+    // Add more CuePoints.
+    const int32_t new_capacity =
+        (!cue_entries_capacity_) ? 2 : cue_entries_capacity_ * 2;
+
+    if (new_capacity < 1)
+      return false;
+
+    CuePoint** const cues =
+        new (std::nothrow) CuePoint*[new_capacity];  // NOLINT
+    if (!cues)
+      return false;
+
+    for (int32_t i = 0; i < cue_entries_size_; ++i) {
+      cues[i] = cue_entries_[i];
+    }
+
+    delete[] cue_entries_;
+
+    cue_entries_ = cues;
+    cue_entries_capacity_ = new_capacity;
+  }
+
+  cue->set_output_block_number(output_block_number_);
+  cue_entries_[cue_entries_size_++] = cue;
+  return true;
+}
+
+CuePoint* Cues::GetCueByIndex(int32_t index) const {
+  if (cue_entries_ == NULL)
+    return NULL;
+
+  if (index >= cue_entries_size_)
+    return NULL;
+
+  return cue_entries_[index];
+}
+
+uint64_t Cues::Size() {
+  uint64_t size = 0;
+  for (int32_t i = 0; i < cue_entries_size_; ++i)
+    size += GetCueByIndex(i)->Size();
+  size += EbmlMasterElementSize(libwebm::kMkvCues, size);
+  return size;
+}
+
+bool Cues::Write(IMkvWriter* writer) const {
+  if (!writer)
+    return false;
+
+  uint64_t size = 0;
+  for (int32_t i = 0; i < cue_entries_size_; ++i) {
+    const CuePoint* const cue = GetCueByIndex(i);
+
+    if (!cue)
+      return false;
+
+    size += cue->Size();
+  }
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvCues, size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  for (int32_t i = 0; i < cue_entries_size_; ++i) {
+    const CuePoint* const cue = GetCueByIndex(i);
+
+    if (!cue->Write(writer))
+      return false;
+  }
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0)
+    return false;
+
+  if (stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// ContentEncAESSettings Class
+
+ContentEncAESSettings::ContentEncAESSettings() : cipher_mode_(kCTR) {}
+
+uint64_t ContentEncAESSettings::Size() const {
+  const uint64_t payload = PayloadSize();
+  const uint64_t size =
+      EbmlMasterElementSize(libwebm::kMkvContentEncAESSettings, payload) +
+      payload;
+  return size;
+}
+
+bool ContentEncAESSettings::Write(IMkvWriter* writer) const {
+  const uint64_t payload = PayloadSize();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvContentEncAESSettings,
+                              payload))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvAESSettingsCipherMode,
+                        cipher_mode_))
+    return false;
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(payload))
+    return false;
+
+  return true;
+}
+
+uint64_t ContentEncAESSettings::PayloadSize() const {
+  uint64_t size =
+      EbmlElementSize(libwebm::kMkvAESSettingsCipherMode, cipher_mode_);
+  return size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// ContentEncoding Class
+
+ContentEncoding::ContentEncoding()
+    : enc_algo_(5),
+      enc_key_id_(NULL),
+      encoding_order_(0),
+      encoding_scope_(1),
+      encoding_type_(1),
+      enc_key_id_length_(0) {}
+
+ContentEncoding::~ContentEncoding() { delete[] enc_key_id_; }
+
+bool ContentEncoding::SetEncryptionID(const uint8_t* id, uint64_t length) {
+  if (!id || length < 1)
+    return false;
+
+  delete[] enc_key_id_;
+
+  enc_key_id_ =
+      new (std::nothrow) uint8_t[static_cast<size_t>(length)];  // NOLINT
+  if (!enc_key_id_)
+    return false;
+
+  memcpy(enc_key_id_, id, static_cast<size_t>(length));
+  enc_key_id_length_ = length;
+
+  return true;
+}
+
+uint64_t ContentEncoding::Size() const {
+  const uint64_t encryption_size = EncryptionSize();
+  const uint64_t encoding_size = EncodingSize(0, encryption_size);
+  const uint64_t encodings_size =
+      EbmlMasterElementSize(libwebm::kMkvContentEncoding, encoding_size) +
+      encoding_size;
+
+  return encodings_size;
+}
+
+bool ContentEncoding::Write(IMkvWriter* writer) const {
+  const uint64_t encryption_size = EncryptionSize();
+  const uint64_t encoding_size = EncodingSize(0, encryption_size);
+  const uint64_t size =
+      EbmlMasterElementSize(libwebm::kMkvContentEncoding, encoding_size) +
+      encoding_size;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvContentEncoding,
+                              encoding_size))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvContentEncodingOrder,
+                        encoding_order_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvContentEncodingScope,
+                        encoding_scope_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvContentEncodingType,
+                        encoding_type_))
+    return false;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvContentEncryption,
+                              encryption_size))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvContentEncAlgo, enc_algo_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvContentEncKeyID, enc_key_id_,
+                        enc_key_id_length_))
+    return false;
+
+  if (!enc_aes_settings_.Write(writer))
+    return false;
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  return true;
+}
+
+uint64_t ContentEncoding::EncodingSize(uint64_t compresion_size,
+                                       uint64_t encryption_size) const {
+  // TODO(fgalligan): Add support for compression settings.
+  if (compresion_size != 0)
+    return 0;
+
+  uint64_t encoding_size = 0;
+
+  if (encryption_size > 0) {
+    encoding_size +=
+        EbmlMasterElementSize(libwebm::kMkvContentEncryption, encryption_size) +
+        encryption_size;
+  }
+  encoding_size +=
+      EbmlElementSize(libwebm::kMkvContentEncodingType, encoding_type_);
+  encoding_size +=
+      EbmlElementSize(libwebm::kMkvContentEncodingScope, encoding_scope_);
+  encoding_size +=
+      EbmlElementSize(libwebm::kMkvContentEncodingOrder, encoding_order_);
+
+  return encoding_size;
+}
+
+uint64_t ContentEncoding::EncryptionSize() const {
+  const uint64_t aes_size = enc_aes_settings_.Size();
+
+  uint64_t encryption_size = EbmlElementSize(libwebm::kMkvContentEncKeyID,
+                                             enc_key_id_, enc_key_id_length_);
+  encryption_size += EbmlElementSize(libwebm::kMkvContentEncAlgo, enc_algo_);
+
+  return encryption_size + aes_size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Track Class
+
+Track::Track(unsigned int* seed)
+    : codec_id_(NULL),
+      codec_private_(NULL),
+      language_(NULL),
+      max_block_additional_id_(0),
+      name_(NULL),
+      number_(0),
+      type_(0),
+      uid_(MakeUID(seed)),
+      codec_delay_(0),
+      seek_pre_roll_(0),
+      default_duration_(0),
+      codec_private_length_(0),
+      content_encoding_entries_(NULL),
+      content_encoding_entries_size_(0) {}
+
+Track::~Track() {
+  delete[] codec_id_;
+  delete[] codec_private_;
+  delete[] language_;
+  delete[] name_;
+
+  if (content_encoding_entries_) {
+    for (uint32_t i = 0; i < content_encoding_entries_size_; ++i) {
+      ContentEncoding* const encoding = content_encoding_entries_[i];
+      delete encoding;
+    }
+    delete[] content_encoding_entries_;
+  }
+}
+
+bool Track::AddContentEncoding() {
+  const uint32_t count = content_encoding_entries_size_ + 1;
+
+  ContentEncoding** const content_encoding_entries =
+      new (std::nothrow) ContentEncoding*[count];  // NOLINT
+  if (!content_encoding_entries)
+    return false;
+
+  ContentEncoding* const content_encoding =
+      new (std::nothrow) ContentEncoding();  // NOLINT
+  if (!content_encoding) {
+    delete[] content_encoding_entries;
+    return false;
+  }
+
+  for (uint32_t i = 0; i < content_encoding_entries_size_; ++i) {
+    content_encoding_entries[i] = content_encoding_entries_[i];
+  }
+
+  delete[] content_encoding_entries_;
+
+  content_encoding_entries_ = content_encoding_entries;
+  content_encoding_entries_[content_encoding_entries_size_] = content_encoding;
+  content_encoding_entries_size_ = count;
+  return true;
+}
+
+ContentEncoding* Track::GetContentEncodingByIndex(uint32_t index) const {
+  if (content_encoding_entries_ == NULL)
+    return NULL;
+
+  if (index >= content_encoding_entries_size_)
+    return NULL;
+
+  return content_encoding_entries_[index];
+}
+
+uint64_t Track::PayloadSize() const {
+  uint64_t size = EbmlElementSize(libwebm::kMkvTrackNumber, number_);
+  size += EbmlElementSize(libwebm::kMkvTrackUID, uid_);
+  size += EbmlElementSize(libwebm::kMkvTrackType, type_);
+  if (codec_id_)
+    size += EbmlElementSize(libwebm::kMkvCodecID, codec_id_);
+  if (codec_private_)
+    size += EbmlElementSize(libwebm::kMkvCodecPrivate, codec_private_,
+                            codec_private_length_);
+  if (language_)
+    size += EbmlElementSize(libwebm::kMkvLanguage, language_);
+  if (name_)
+    size += EbmlElementSize(libwebm::kMkvName, name_);
+  if (max_block_additional_id_)
+    size += EbmlElementSize(libwebm::kMkvMaxBlockAdditionID,
+                            max_block_additional_id_);
+  if (codec_delay_)
+    size += EbmlElementSize(libwebm::kMkvCodecDelay, codec_delay_);
+  if (seek_pre_roll_)
+    size += EbmlElementSize(libwebm::kMkvSeekPreRoll, seek_pre_roll_);
+  if (default_duration_)
+    size += EbmlElementSize(libwebm::kMkvDefaultDuration, default_duration_);
+
+  if (content_encoding_entries_size_ > 0) {
+    uint64_t content_encodings_size = 0;
+    for (uint32_t i = 0; i < content_encoding_entries_size_; ++i) {
+      ContentEncoding* const encoding = content_encoding_entries_[i];
+      content_encodings_size += encoding->Size();
+    }
+
+    size += EbmlMasterElementSize(libwebm::kMkvContentEncodings,
+                                  content_encodings_size) +
+            content_encodings_size;
+  }
+
+  return size;
+}
+
+uint64_t Track::Size() const {
+  uint64_t size = PayloadSize();
+  size += EbmlMasterElementSize(libwebm::kMkvTrackEntry, size);
+  return size;
+}
+
+bool Track::Write(IMkvWriter* writer) const {
+  if (!writer)
+    return false;
+
+  // mandatory elements without a default value.
+  if (!type_ || !codec_id_)
+    return false;
+
+  // |size| may be bigger than what is written out in this function because
+  // derived classes may write out more data in the Track element.
+  const uint64_t payload_size = PayloadSize();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvTrackEntry, payload_size))
+    return false;
+
+  uint64_t size = EbmlElementSize(libwebm::kMkvTrackNumber, number_);
+  size += EbmlElementSize(libwebm::kMkvTrackUID, uid_);
+  size += EbmlElementSize(libwebm::kMkvTrackType, type_);
+  if (codec_id_)
+    size += EbmlElementSize(libwebm::kMkvCodecID, codec_id_);
+  if (codec_private_)
+    size += EbmlElementSize(libwebm::kMkvCodecPrivate, codec_private_,
+                            codec_private_length_);
+  if (language_)
+    size += EbmlElementSize(libwebm::kMkvLanguage, language_);
+  if (name_)
+    size += EbmlElementSize(libwebm::kMkvName, name_);
+  if (max_block_additional_id_)
+    size += EbmlElementSize(libwebm::kMkvMaxBlockAdditionID,
+                            max_block_additional_id_);
+  if (codec_delay_)
+    size += EbmlElementSize(libwebm::kMkvCodecDelay, codec_delay_);
+  if (seek_pre_roll_)
+    size += EbmlElementSize(libwebm::kMkvSeekPreRoll, seek_pre_roll_);
+  if (default_duration_)
+    size += EbmlElementSize(libwebm::kMkvDefaultDuration, default_duration_);
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvTrackNumber, number_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvTrackUID, uid_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvTrackType, type_))
+    return false;
+  if (max_block_additional_id_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvMaxBlockAdditionID,
+                          max_block_additional_id_)) {
+      return false;
+    }
+  }
+  if (codec_delay_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvCodecDelay, codec_delay_))
+      return false;
+  }
+  if (seek_pre_roll_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvSeekPreRoll, seek_pre_roll_))
+      return false;
+  }
+  if (default_duration_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvDefaultDuration,
+                          default_duration_))
+      return false;
+  }
+  if (codec_id_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvCodecID, codec_id_))
+      return false;
+  }
+  if (codec_private_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvCodecPrivate, codec_private_,
+                          codec_private_length_))
+      return false;
+  }
+  if (language_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvLanguage, language_))
+      return false;
+  }
+  if (name_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvName, name_))
+      return false;
+  }
+
+  int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  if (content_encoding_entries_size_ > 0) {
+    uint64_t content_encodings_size = 0;
+    for (uint32_t i = 0; i < content_encoding_entries_size_; ++i) {
+      ContentEncoding* const encoding = content_encoding_entries_[i];
+      content_encodings_size += encoding->Size();
+    }
+
+    if (!WriteEbmlMasterElement(writer, libwebm::kMkvContentEncodings,
+                                content_encodings_size))
+      return false;
+
+    for (uint32_t i = 0; i < content_encoding_entries_size_; ++i) {
+      ContentEncoding* const encoding = content_encoding_entries_[i];
+      if (!encoding->Write(writer))
+        return false;
+    }
+  }
+
+  stop_position = writer->Position();
+  if (stop_position < 0)
+    return false;
+  return true;
+}
+
+bool Track::SetCodecPrivate(const uint8_t* codec_private, uint64_t length) {
+  if (!codec_private || length < 1)
+    return false;
+
+  delete[] codec_private_;
+
+  codec_private_ =
+      new (std::nothrow) uint8_t[static_cast<size_t>(length)];  // NOLINT
+  if (!codec_private_)
+    return false;
+
+  memcpy(codec_private_, codec_private, static_cast<size_t>(length));
+  codec_private_length_ = length;
+
+  return true;
+}
+
+void Track::set_codec_id(const char* codec_id) {
+  if (codec_id) {
+    delete[] codec_id_;
+
+    const size_t length = strlen(codec_id) + 1;
+    codec_id_ = new (std::nothrow) char[length];  // NOLINT
+    if (codec_id_) {
+#ifdef _MSC_VER
+      strcpy_s(codec_id_, length, codec_id);
+#else
+      strcpy(codec_id_, codec_id);
+#endif
+    }
+  }
+}
+
+// TODO(fgalligan): Vet the language parameter.
+void Track::set_language(const char* language) {
+  if (language) {
+    delete[] language_;
+
+    const size_t length = strlen(language) + 1;
+    language_ = new (std::nothrow) char[length];  // NOLINT
+    if (language_) {
+#ifdef _MSC_VER
+      strcpy_s(language_, length, language);
+#else
+      strcpy(language_, language);
+#endif
+    }
+  }
+}
+
+void Track::set_name(const char* name) {
+  if (name) {
+    delete[] name_;
+
+    const size_t length = strlen(name) + 1;
+    name_ = new (std::nothrow) char[length];  // NOLINT
+    if (name_) {
+#ifdef _MSC_VER
+      strcpy_s(name_, length, name);
+#else
+      strcpy(name_, name);
+#endif
+    }
+  }
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Colour and its child elements
+
+uint64_t PrimaryChromaticity::PrimaryChromaticityPayloadSize(
+    libwebm::MkvId x_id, libwebm::MkvId y_id) const {
+  return EbmlElementSize(x_id, x) + EbmlElementSize(y_id, y);
+}
+
+bool PrimaryChromaticity::Write(IMkvWriter* writer, libwebm::MkvId x_id,
+                                libwebm::MkvId y_id) const {
+  return WriteEbmlElement(writer, x_id, x) && WriteEbmlElement(writer, y_id, y);
+}
+
+uint64_t MasteringMetadata::MasteringMetadataSize() const {
+  uint64_t size = PayloadSize();
+
+  if (size > 0)
+    size += EbmlMasterElementSize(libwebm::kMkvMasteringMetadata, size);
+
+  return size;
+}
+
+bool MasteringMetadata::Write(IMkvWriter* writer) const {
+  const uint64_t size = PayloadSize();
+
+  // Don't write an empty element.
+  if (size == 0)
+    return true;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvMasteringMetadata, size))
+    return false;
+  if (luminance_max != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvLuminanceMax, luminance_max)) {
+    return false;
+  }
+  if (luminance_min != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvLuminanceMin, luminance_min)) {
+    return false;
+  }
+  if (r_ &&
+      !r_->Write(writer, libwebm::kMkvPrimaryRChromaticityX,
+                 libwebm::kMkvPrimaryRChromaticityY)) {
+    return false;
+  }
+  if (g_ &&
+      !g_->Write(writer, libwebm::kMkvPrimaryGChromaticityX,
+                 libwebm::kMkvPrimaryGChromaticityY)) {
+    return false;
+  }
+  if (b_ &&
+      !b_->Write(writer, libwebm::kMkvPrimaryBChromaticityX,
+                 libwebm::kMkvPrimaryBChromaticityY)) {
+    return false;
+  }
+  if (white_point_ &&
+      !white_point_->Write(writer, libwebm::kMkvWhitePointChromaticityX,
+                           libwebm::kMkvWhitePointChromaticityY)) {
+    return false;
+  }
+
+  return true;
+}
+
+bool MasteringMetadata::SetChromaticity(
+    const PrimaryChromaticity* r, const PrimaryChromaticity* g,
+    const PrimaryChromaticity* b, const PrimaryChromaticity* white_point) {
+  PrimaryChromaticityPtr r_ptr(NULL);
+  if (r) {
+    if (!CopyChromaticity(r, &r_ptr))
+      return false;
+  }
+  PrimaryChromaticityPtr g_ptr(NULL);
+  if (g) {
+    if (!CopyChromaticity(g, &g_ptr))
+      return false;
+  }
+  PrimaryChromaticityPtr b_ptr(NULL);
+  if (b) {
+    if (!CopyChromaticity(b, &b_ptr))
+      return false;
+  }
+  PrimaryChromaticityPtr wp_ptr(NULL);
+  if (white_point) {
+    if (!CopyChromaticity(white_point, &wp_ptr))
+      return false;
+  }
+
+  r_ = r_ptr.release();
+  g_ = g_ptr.release();
+  b_ = b_ptr.release();
+  white_point_ = wp_ptr.release();
+  return true;
+}
+
+uint64_t MasteringMetadata::PayloadSize() const {
+  uint64_t size = 0;
+
+  if (luminance_max != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvLuminanceMax, luminance_max);
+  if (luminance_min != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvLuminanceMin, luminance_min);
+
+  if (r_) {
+    size += r_->PrimaryChromaticityPayloadSize(
+        libwebm::kMkvPrimaryRChromaticityX, libwebm::kMkvPrimaryRChromaticityY);
+  }
+  if (g_) {
+    size += g_->PrimaryChromaticityPayloadSize(
+        libwebm::kMkvPrimaryGChromaticityX, libwebm::kMkvPrimaryGChromaticityY);
+  }
+  if (b_) {
+    size += b_->PrimaryChromaticityPayloadSize(
+        libwebm::kMkvPrimaryBChromaticityX, libwebm::kMkvPrimaryBChromaticityY);
+  }
+  if (white_point_) {
+    size += white_point_->PrimaryChromaticityPayloadSize(
+        libwebm::kMkvWhitePointChromaticityX,
+        libwebm::kMkvWhitePointChromaticityY);
+  }
+
+  return size;
+}
+
+uint64_t Colour::ColourSize() const {
+  uint64_t size = PayloadSize();
+
+  if (size > 0)
+    size += EbmlMasterElementSize(libwebm::kMkvColour, size);
+
+  return size;
+}
+
+bool Colour::Write(IMkvWriter* writer) const {
+  const uint64_t size = PayloadSize();
+
+  // Don't write an empty element.
+  if (size == 0)
+    return true;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvColour, size))
+    return false;
+
+  if (matrix_coefficients != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvMatrixCoefficients,
+                        matrix_coefficients)) {
+    return false;
+  }
+  if (bits_per_channel != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvBitsPerChannel,
+                        bits_per_channel)) {
+    return false;
+  }
+  if (chroma_subsampling_horz != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvChromaSubsamplingHorz,
+                        chroma_subsampling_horz)) {
+    return false;
+  }
+  if (chroma_subsampling_vert != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvChromaSubsamplingVert,
+                        chroma_subsampling_vert)) {
+    return false;
+  }
+
+  if (cb_subsampling_horz != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvCbSubsamplingHorz,
+                        cb_subsampling_horz)) {
+    return false;
+  }
+  if (cb_subsampling_vert != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvCbSubsamplingVert,
+                        cb_subsampling_vert)) {
+    return false;
+  }
+  if (chroma_siting_horz != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvChromaSitingHorz,
+                        chroma_siting_horz)) {
+    return false;
+  }
+  if (chroma_siting_vert != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvChromaSitingVert,
+                        chroma_siting_vert)) {
+    return false;
+  }
+  if (range != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvRange, range)) {
+    return false;
+  }
+  if (transfer_characteristics != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvTransferCharacteristics,
+                        transfer_characteristics)) {
+    return false;
+  }
+  if (primaries != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvPrimaries, primaries)) {
+    return false;
+  }
+  if (max_cll != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvMaxCLL, max_cll)) {
+    return false;
+  }
+  if (max_fall != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvMaxFALL, max_fall)) {
+    return false;
+  }
+
+  if (mastering_metadata_ && !mastering_metadata_->Write(writer))
+    return false;
+
+  return true;
+}
+
+bool Colour::SetMasteringMetadata(const MasteringMetadata& mastering_metadata) {
+  std::auto_ptr<MasteringMetadata> mm_ptr(new MasteringMetadata());
+  if (!mm_ptr.get())
+    return false;
+
+  mm_ptr->luminance_max = mastering_metadata.luminance_max;
+  mm_ptr->luminance_min = mastering_metadata.luminance_min;
+
+  if (!mm_ptr->SetChromaticity(mastering_metadata.r(), mastering_metadata.g(),
+                               mastering_metadata.b(),
+                               mastering_metadata.white_point())) {
+    return false;
+  }
+
+  delete mastering_metadata_;
+  mastering_metadata_ = mm_ptr.release();
+  return true;
+}
+
+uint64_t Colour::PayloadSize() const {
+  uint64_t size = 0;
+
+  if (matrix_coefficients != kValueNotPresent)
+    size +=
+        EbmlElementSize(libwebm::kMkvMatrixCoefficients, matrix_coefficients);
+  if (bits_per_channel != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvBitsPerChannel, bits_per_channel);
+  if (chroma_subsampling_horz != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvChromaSubsamplingHorz,
+                            chroma_subsampling_horz);
+  if (chroma_subsampling_vert != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvChromaSubsamplingVert,
+                            chroma_subsampling_vert);
+  if (cb_subsampling_horz != kValueNotPresent)
+    size +=
+        EbmlElementSize(libwebm::kMkvCbSubsamplingHorz, cb_subsampling_horz);
+  if (cb_subsampling_vert != kValueNotPresent)
+    size +=
+        EbmlElementSize(libwebm::kMkvCbSubsamplingVert, cb_subsampling_vert);
+  if (chroma_siting_horz != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvChromaSitingHorz, chroma_siting_horz);
+  if (chroma_siting_vert != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvChromaSitingVert, chroma_siting_vert);
+  if (range != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvRange, range);
+  if (transfer_characteristics != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvTransferCharacteristics,
+                            transfer_characteristics);
+  if (primaries != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvPrimaries, primaries);
+  if (max_cll != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvMaxCLL, max_cll);
+  if (max_fall != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvMaxFALL, max_fall);
+
+  if (mastering_metadata_)
+    size += mastering_metadata_->MasteringMetadataSize();
+
+  return size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// VideoTrack Class
+
+VideoTrack::VideoTrack(unsigned int* seed)
+    : Track(seed),
+      display_height_(0),
+      display_width_(0),
+      crop_left_(0),
+      crop_right_(0),
+      crop_top_(0),
+      crop_bottom_(0),
+      frame_rate_(0.0),
+      height_(0),
+      stereo_mode_(0),
+      alpha_mode_(0),
+      width_(0),
+      colour_(NULL) {}
+
+VideoTrack::~VideoTrack() { delete colour_; }
+
+bool VideoTrack::SetStereoMode(uint64_t stereo_mode) {
+  if (stereo_mode != kMono && stereo_mode != kSideBySideLeftIsFirst &&
+      stereo_mode != kTopBottomRightIsFirst &&
+      stereo_mode != kTopBottomLeftIsFirst &&
+      stereo_mode != kSideBySideRightIsFirst)
+    return false;
+
+  stereo_mode_ = stereo_mode;
+  return true;
+}
+
+bool VideoTrack::SetAlphaMode(uint64_t alpha_mode) {
+  if (alpha_mode != kNoAlpha && alpha_mode != kAlpha)
+    return false;
+
+  alpha_mode_ = alpha_mode;
+  return true;
+}
+
+uint64_t VideoTrack::PayloadSize() const {
+  const uint64_t parent_size = Track::PayloadSize();
+
+  uint64_t size = VideoPayloadSize();
+  size += EbmlMasterElementSize(libwebm::kMkvVideo, size);
+
+  return parent_size + size;
+}
+
+bool VideoTrack::Write(IMkvWriter* writer) const {
+  if (!Track::Write(writer))
+    return false;
+
+  const uint64_t size = VideoPayloadSize();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvVideo, size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvPixelWidth, width_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvPixelHeight, height_))
+    return false;
+  if (display_width_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvDisplayWidth, display_width_))
+      return false;
+  }
+  if (display_height_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvDisplayHeight, display_height_))
+      return false;
+  }
+  if (crop_left_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvPixelCropLeft, crop_left_))
+      return false;
+  }
+  if (crop_right_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvPixelCropRight, crop_right_))
+      return false;
+  }
+  if (crop_top_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvPixelCropTop, crop_top_))
+      return false;
+  }
+  if (crop_bottom_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvPixelCropBottom, crop_bottom_))
+      return false;
+  }
+  if (stereo_mode_ > kMono) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvStereoMode, stereo_mode_))
+      return false;
+  }
+  if (alpha_mode_ > kNoAlpha) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvAlphaMode, alpha_mode_))
+      return false;
+  }
+  if (frame_rate_ > 0.0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvFrameRate,
+                          static_cast<float>(frame_rate_))) {
+      return false;
+    }
+  }
+  if (colour_) {
+    if (!colour_->Write(writer))
+      return false;
+  }
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size)) {
+    return false;
+  }
+
+  return true;
+}
+
+bool VideoTrack::SetColour(const Colour& colour) {
+  std::auto_ptr<Colour> colour_ptr(new Colour());
+  if (!colour_ptr.get())
+    return false;
+
+  if (colour.mastering_metadata()) {
+    if (!colour_ptr->SetMasteringMetadata(*colour.mastering_metadata()))
+      return false;
+  }
+
+  colour_ptr->matrix_coefficients = colour.matrix_coefficients;
+  colour_ptr->bits_per_channel = colour.bits_per_channel;
+  colour_ptr->chroma_subsampling_horz = colour.chroma_subsampling_horz;
+  colour_ptr->chroma_subsampling_vert = colour.chroma_subsampling_vert;
+  colour_ptr->cb_subsampling_horz = colour.cb_subsampling_horz;
+  colour_ptr->cb_subsampling_vert = colour.cb_subsampling_vert;
+  colour_ptr->chroma_siting_horz = colour.chroma_siting_horz;
+  colour_ptr->chroma_siting_vert = colour.chroma_siting_vert;
+  colour_ptr->range = colour.range;
+  colour_ptr->transfer_characteristics = colour.transfer_characteristics;
+  colour_ptr->primaries = colour.primaries;
+  colour_ptr->max_cll = colour.max_cll;
+  colour_ptr->max_fall = colour.max_fall;
+  colour_ = colour_ptr.release();
+  return true;
+}
+
+uint64_t VideoTrack::VideoPayloadSize() const {
+  uint64_t size = EbmlElementSize(libwebm::kMkvPixelWidth, width_);
+  size += EbmlElementSize(libwebm::kMkvPixelHeight, height_);
+  if (display_width_ > 0)
+    size += EbmlElementSize(libwebm::kMkvDisplayWidth, display_width_);
+  if (display_height_ > 0)
+    size += EbmlElementSize(libwebm::kMkvDisplayHeight, display_height_);
+  if (crop_left_ > 0)
+    size += EbmlElementSize(libwebm::kMkvPixelCropLeft, crop_left_);
+  if (crop_right_ > 0)
+    size += EbmlElementSize(libwebm::kMkvPixelCropRight, crop_right_);
+  if (crop_top_ > 0)
+    size += EbmlElementSize(libwebm::kMkvPixelCropTop, crop_top_);
+  if (crop_bottom_ > 0)
+    size += EbmlElementSize(libwebm::kMkvPixelCropBottom, crop_bottom_);
+  if (stereo_mode_ > kMono)
+    size += EbmlElementSize(libwebm::kMkvStereoMode, stereo_mode_);
+  if (alpha_mode_ > kNoAlpha)
+    size += EbmlElementSize(libwebm::kMkvAlphaMode, alpha_mode_);
+  if (frame_rate_ > 0.0)
+    size += EbmlElementSize(libwebm::kMkvFrameRate,
+                            static_cast<float>(frame_rate_));
+  if (colour_)
+    size += colour_->ColourSize();
+
+  return size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// AudioTrack Class
+
+AudioTrack::AudioTrack(unsigned int* seed)
+    : Track(seed), bit_depth_(0), channels_(1), sample_rate_(0.0) {}
+
+AudioTrack::~AudioTrack() {}
+
+uint64_t AudioTrack::PayloadSize() const {
+  const uint64_t parent_size = Track::PayloadSize();
+
+  uint64_t size = EbmlElementSize(libwebm::kMkvSamplingFrequency,
+                                  static_cast<float>(sample_rate_));
+  size += EbmlElementSize(libwebm::kMkvChannels, channels_);
+  if (bit_depth_ > 0)
+    size += EbmlElementSize(libwebm::kMkvBitDepth, bit_depth_);
+  size += EbmlMasterElementSize(libwebm::kMkvAudio, size);
+
+  return parent_size + size;
+}
+
+bool AudioTrack::Write(IMkvWriter* writer) const {
+  if (!Track::Write(writer))
+    return false;
+
+  // Calculate AudioSettings size.
+  uint64_t size = EbmlElementSize(libwebm::kMkvSamplingFrequency,
+                                  static_cast<float>(sample_rate_));
+  size += EbmlElementSize(libwebm::kMkvChannels, channels_);
+  if (bit_depth_ > 0)
+    size += EbmlElementSize(libwebm::kMkvBitDepth, bit_depth_);
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvAudio, size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvSamplingFrequency,
+                        static_cast<float>(sample_rate_)))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvChannels, channels_))
+    return false;
+  if (bit_depth_ > 0)
+    if (!WriteEbmlElement(writer, libwebm::kMkvBitDepth, bit_depth_))
+      return false;
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Tracks Class
+
+const char Tracks::kOpusCodecId[] = "A_OPUS";
+const char Tracks::kVorbisCodecId[] = "A_VORBIS";
+const char Tracks::kVp8CodecId[] = "V_VP8";
+const char Tracks::kVp9CodecId[] = "V_VP9";
+const char Tracks::kVp10CodecId[] = "V_VP10";
+
+Tracks::Tracks()
+    : track_entries_(NULL), track_entries_size_(0), wrote_tracks_(false) {}
+
+Tracks::~Tracks() {
+  if (track_entries_) {
+    for (uint32_t i = 0; i < track_entries_size_; ++i) {
+      Track* const track = track_entries_[i];
+      delete track;
+    }
+    delete[] track_entries_;
+  }
+}
+
+bool Tracks::AddTrack(Track* track, int32_t number) {
+  if (number < 0 || wrote_tracks_)
+    return false;
+
+  // This muxer only supports track numbers in the range [1, 126], in
+  // order to be able (to use Matroska integer representation) to
+  // serialize the block header (of which the track number is a part)
+  // for a frame using exactly 4 bytes.
+
+  if (number > 0x7E)
+    return false;
+
+  uint32_t track_num = number;
+
+  if (track_num > 0) {
+    // Check to make sure a track does not already have |track_num|.
+    for (uint32_t i = 0; i < track_entries_size_; ++i) {
+      if (track_entries_[i]->number() == track_num)
+        return false;
+    }
+  }
+
+  const uint32_t count = track_entries_size_ + 1;
+
+  Track** const track_entries = new (std::nothrow) Track*[count];  // NOLINT
+  if (!track_entries)
+    return false;
+
+  for (uint32_t i = 0; i < track_entries_size_; ++i) {
+    track_entries[i] = track_entries_[i];
+  }
+
+  delete[] track_entries_;
+
+  // Find the lowest availible track number > 0.
+  if (track_num == 0) {
+    track_num = count;
+
+    // Check to make sure a track does not already have |track_num|.
+    bool exit = false;
+    do {
+      exit = true;
+      for (uint32_t i = 0; i < track_entries_size_; ++i) {
+        if (track_entries[i]->number() == track_num) {
+          track_num++;
+          exit = false;
+          break;
+        }
+      }
+    } while (!exit);
+  }
+  track->set_number(track_num);
+
+  track_entries_ = track_entries;
+  track_entries_[track_entries_size_] = track;
+  track_entries_size_ = count;
+  return true;
+}
+
+const Track* Tracks::GetTrackByIndex(uint32_t index) const {
+  if (track_entries_ == NULL)
+    return NULL;
+
+  if (index >= track_entries_size_)
+    return NULL;
+
+  return track_entries_[index];
+}
+
+Track* Tracks::GetTrackByNumber(uint64_t track_number) const {
+  const int32_t count = track_entries_size();
+  for (int32_t i = 0; i < count; ++i) {
+    if (track_entries_[i]->number() == track_number)
+      return track_entries_[i];
+  }
+
+  return NULL;
+}
+
+bool Tracks::TrackIsAudio(uint64_t track_number) const {
+  const Track* const track = GetTrackByNumber(track_number);
+
+  if (track->type() == kAudio)
+    return true;
+
+  return false;
+}
+
+bool Tracks::TrackIsVideo(uint64_t track_number) const {
+  const Track* const track = GetTrackByNumber(track_number);
+
+  if (track->type() == kVideo)
+    return true;
+
+  return false;
+}
+
+bool Tracks::Write(IMkvWriter* writer) const {
+  uint64_t size = 0;
+  const int32_t count = track_entries_size();
+  for (int32_t i = 0; i < count; ++i) {
+    const Track* const track = GetTrackByIndex(i);
+
+    if (!track)
+      return false;
+
+    size += track->Size();
+  }
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvTracks, size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  for (int32_t i = 0; i < count; ++i) {
+    const Track* const track = GetTrackByIndex(i);
+    if (!track->Write(writer))
+      return false;
+  }
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  wrote_tracks_ = true;
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Chapter Class
+
+bool Chapter::set_id(const char* id) { return StrCpy(id, &id_); }
+
+void Chapter::set_time(const Segment& segment, uint64_t start_ns,
+                       uint64_t end_ns) {
+  const SegmentInfo* const info = segment.GetSegmentInfo();
+  const uint64_t timecode_scale = info->timecode_scale();
+  start_timecode_ = start_ns / timecode_scale;
+  end_timecode_ = end_ns / timecode_scale;
+}
+
+bool Chapter::add_string(const char* title, const char* language,
+                         const char* country) {
+  if (!ExpandDisplaysArray())
+    return false;
+
+  Display& d = displays_[displays_count_++];
+  d.Init();
+
+  if (!d.set_title(title))
+    return false;
+
+  if (!d.set_language(language))
+    return false;
+
+  if (!d.set_country(country))
+    return false;
+
+  return true;
+}
+
+Chapter::Chapter() {
+  // This ctor only constructs the object.  Proper initialization is
+  // done in Init() (called in Chapters::AddChapter()).  The only
+  // reason we bother implementing this ctor is because we had to
+  // declare it as private (along with the dtor), in order to prevent
+  // clients from creating Chapter instances (a privelege we grant
+  // only to the Chapters class).  Doing no initialization here also
+  // means that creating arrays of chapter objects is more efficient,
+  // because we only initialize each new chapter object as it becomes
+  // active on the array.
+}
+
+Chapter::~Chapter() {}
+
+void Chapter::Init(unsigned int* seed) {
+  id_ = NULL;
+  start_timecode_ = 0;
+  end_timecode_ = 0;
+  displays_ = NULL;
+  displays_size_ = 0;
+  displays_count_ = 0;
+  uid_ = MakeUID(seed);
+}
+
+void Chapter::ShallowCopy(Chapter* dst) const {
+  dst->id_ = id_;
+  dst->start_timecode_ = start_timecode_;
+  dst->end_timecode_ = end_timecode_;
+  dst->uid_ = uid_;
+  dst->displays_ = displays_;
+  dst->displays_size_ = displays_size_;
+  dst->displays_count_ = displays_count_;
+}
+
+void Chapter::Clear() {
+  StrCpy(NULL, &id_);
+
+  while (displays_count_ > 0) {
+    Display& d = displays_[--displays_count_];
+    d.Clear();
+  }
+
+  delete[] displays_;
+  displays_ = NULL;
+
+  displays_size_ = 0;
+}
+
+bool Chapter::ExpandDisplaysArray() {
+  if (displays_size_ > displays_count_)
+    return true;  // nothing to do yet
+
+  const int size = (displays_size_ == 0) ? 1 : 2 * displays_size_;
+
+  Display* const displays = new (std::nothrow) Display[size];  // NOLINT
+  if (displays == NULL)
+    return false;
+
+  for (int idx = 0; idx < displays_count_; ++idx) {
+    displays[idx] = displays_[idx];  // shallow copy
+  }
+
+  delete[] displays_;
+
+  displays_ = displays;
+  displays_size_ = size;
+
+  return true;
+}
+
+uint64_t Chapter::WriteAtom(IMkvWriter* writer) const {
+  uint64_t payload_size =
+      EbmlElementSize(libwebm::kMkvChapterStringUID, id_) +
+      EbmlElementSize(libwebm::kMkvChapterUID, uid_) +
+      EbmlElementSize(libwebm::kMkvChapterTimeStart, start_timecode_) +
+      EbmlElementSize(libwebm::kMkvChapterTimeEnd, end_timecode_);
+
+  for (int idx = 0; idx < displays_count_; ++idx) {
+    const Display& d = displays_[idx];
+    payload_size += d.WriteDisplay(NULL);
+  }
+
+  const uint64_t atom_size =
+      EbmlMasterElementSize(libwebm::kMkvChapterAtom, payload_size) +
+      payload_size;
+
+  if (writer == NULL)
+    return atom_size;
+
+  const int64_t start = writer->Position();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvChapterAtom, payload_size))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvChapterStringUID, id_))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvChapterUID, uid_))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvChapterTimeStart, start_timecode_))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvChapterTimeEnd, end_timecode_))
+    return 0;
+
+  for (int idx = 0; idx < displays_count_; ++idx) {
+    const Display& d = displays_[idx];
+
+    if (!d.WriteDisplay(writer))
+      return 0;
+  }
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != atom_size)
+    return 0;
+
+  return atom_size;
+}
+
+void Chapter::Display::Init() {
+  title_ = NULL;
+  language_ = NULL;
+  country_ = NULL;
+}
+
+void Chapter::Display::Clear() {
+  StrCpy(NULL, &title_);
+  StrCpy(NULL, &language_);
+  StrCpy(NULL, &country_);
+}
+
+bool Chapter::Display::set_title(const char* title) {
+  return StrCpy(title, &title_);
+}
+
+bool Chapter::Display::set_language(const char* language) {
+  return StrCpy(language, &language_);
+}
+
+bool Chapter::Display::set_country(const char* country) {
+  return StrCpy(country, &country_);
+}
+
+uint64_t Chapter::Display::WriteDisplay(IMkvWriter* writer) const {
+  uint64_t payload_size = EbmlElementSize(libwebm::kMkvChapString, title_);
+
+  if (language_)
+    payload_size += EbmlElementSize(libwebm::kMkvChapLanguage, language_);
+
+  if (country_)
+    payload_size += EbmlElementSize(libwebm::kMkvChapCountry, country_);
+
+  const uint64_t display_size =
+      EbmlMasterElementSize(libwebm::kMkvChapterDisplay, payload_size) +
+      payload_size;
+
+  if (writer == NULL)
+    return display_size;
+
+  const int64_t start = writer->Position();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvChapterDisplay,
+                              payload_size))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvChapString, title_))
+    return 0;
+
+  if (language_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvChapLanguage, language_))
+      return 0;
+  }
+
+  if (country_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvChapCountry, country_))
+      return 0;
+  }
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != display_size)
+    return 0;
+
+  return display_size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Chapters Class
+
+Chapters::Chapters() : chapters_size_(0), chapters_count_(0), chapters_(NULL) {}
+
+Chapters::~Chapters() {
+  while (chapters_count_ > 0) {
+    Chapter& chapter = chapters_[--chapters_count_];
+    chapter.Clear();
+  }
+
+  delete[] chapters_;
+  chapters_ = NULL;
+}
+
+int Chapters::Count() const { return chapters_count_; }
+
+Chapter* Chapters::AddChapter(unsigned int* seed) {
+  if (!ExpandChaptersArray())
+    return NULL;
+
+  Chapter& chapter = chapters_[chapters_count_++];
+  chapter.Init(seed);
+
+  return &chapter;
+}
+
+bool Chapters::Write(IMkvWriter* writer) const {
+  if (writer == NULL)
+    return false;
+
+  const uint64_t payload_size = WriteEdition(NULL);  // return size only
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvChapters, payload_size))
+    return false;
+
+  const int64_t start = writer->Position();
+
+  if (WriteEdition(writer) == 0)  // error
+    return false;
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != payload_size)
+    return false;
+
+  return true;
+}
+
+bool Chapters::ExpandChaptersArray() {
+  if (chapters_size_ > chapters_count_)
+    return true;  // nothing to do yet
+
+  const int size = (chapters_size_ == 0) ? 1 : 2 * chapters_size_;
+
+  Chapter* const chapters = new (std::nothrow) Chapter[size];  // NOLINT
+  if (chapters == NULL)
+    return false;
+
+  for (int idx = 0; idx < chapters_count_; ++idx) {
+    const Chapter& src = chapters_[idx];
+    Chapter* const dst = chapters + idx;
+    src.ShallowCopy(dst);
+  }
+
+  delete[] chapters_;
+
+  chapters_ = chapters;
+  chapters_size_ = size;
+
+  return true;
+}
+
+uint64_t Chapters::WriteEdition(IMkvWriter* writer) const {
+  uint64_t payload_size = 0;
+
+  for (int idx = 0; idx < chapters_count_; ++idx) {
+    const Chapter& chapter = chapters_[idx];
+    payload_size += chapter.WriteAtom(NULL);
+  }
+
+  const uint64_t edition_size =
+      EbmlMasterElementSize(libwebm::kMkvEditionEntry, payload_size) +
+      payload_size;
+
+  if (writer == NULL)  // return size only
+    return edition_size;
+
+  const int64_t start = writer->Position();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvEditionEntry, payload_size))
+    return 0;  // error
+
+  for (int idx = 0; idx < chapters_count_; ++idx) {
+    const Chapter& chapter = chapters_[idx];
+
+    const uint64_t chapter_size = chapter.WriteAtom(writer);
+    if (chapter_size == 0)  // error
+      return 0;
+  }
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != edition_size)
+    return 0;
+
+  return edition_size;
+}
+
+// Tag Class
+
+bool Tag::add_simple_tag(const char* tag_name, const char* tag_string) {
+  if (!ExpandSimpleTagsArray())
+    return false;
+
+  SimpleTag& st = simple_tags_[simple_tags_count_++];
+  st.Init();
+
+  if (!st.set_tag_name(tag_name))
+    return false;
+
+  if (!st.set_tag_string(tag_string))
+    return false;
+
+  return true;
+}
+
+Tag::Tag() {
+  simple_tags_ = NULL;
+  simple_tags_size_ = 0;
+  simple_tags_count_ = 0;
+}
+
+Tag::~Tag() {}
+
+void Tag::ShallowCopy(Tag* dst) const {
+  dst->simple_tags_ = simple_tags_;
+  dst->simple_tags_size_ = simple_tags_size_;
+  dst->simple_tags_count_ = simple_tags_count_;
+}
+
+void Tag::Clear() {
+  while (simple_tags_count_ > 0) {
+    SimpleTag& st = simple_tags_[--simple_tags_count_];
+    st.Clear();
+  }
+
+  delete[] simple_tags_;
+  simple_tags_ = NULL;
+
+  simple_tags_size_ = 0;
+}
+
+bool Tag::ExpandSimpleTagsArray() {
+  if (simple_tags_size_ > simple_tags_count_)
+    return true;  // nothing to do yet
+
+  const int size = (simple_tags_size_ == 0) ? 1 : 2 * simple_tags_size_;
+
+  SimpleTag* const simple_tags = new (std::nothrow) SimpleTag[size];  // NOLINT
+  if (simple_tags == NULL)
+    return false;
+
+  for (int idx = 0; idx < simple_tags_count_; ++idx) {
+    simple_tags[idx] = simple_tags_[idx];  // shallow copy
+  }
+
+  delete[] simple_tags_;
+
+  simple_tags_ = simple_tags;
+  simple_tags_size_ = size;
+
+  return true;
+}
+
+uint64_t Tag::Write(IMkvWriter* writer) const {
+  uint64_t payload_size = 0;
+
+  for (int idx = 0; idx < simple_tags_count_; ++idx) {
+    const SimpleTag& st = simple_tags_[idx];
+    payload_size += st.Write(NULL);
+  }
+
+  const uint64_t tag_size =
+      EbmlMasterElementSize(libwebm::kMkvTag, payload_size) + payload_size;
+
+  if (writer == NULL)
+    return tag_size;
+
+  const int64_t start = writer->Position();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvTag, payload_size))
+    return 0;
+
+  for (int idx = 0; idx < simple_tags_count_; ++idx) {
+    const SimpleTag& st = simple_tags_[idx];
+
+    if (!st.Write(writer))
+      return 0;
+  }
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != tag_size)
+    return 0;
+
+  return tag_size;
+}
+
+// Tag::SimpleTag
+
+void Tag::SimpleTag::Init() {
+  tag_name_ = NULL;
+  tag_string_ = NULL;
+}
+
+void Tag::SimpleTag::Clear() {
+  StrCpy(NULL, &tag_name_);
+  StrCpy(NULL, &tag_string_);
+}
+
+bool Tag::SimpleTag::set_tag_name(const char* tag_name) {
+  return StrCpy(tag_name, &tag_name_);
+}
+
+bool Tag::SimpleTag::set_tag_string(const char* tag_string) {
+  return StrCpy(tag_string, &tag_string_);
+}
+
+uint64_t Tag::SimpleTag::Write(IMkvWriter* writer) const {
+  uint64_t payload_size = EbmlElementSize(libwebm::kMkvTagName, tag_name_);
+
+  payload_size += EbmlElementSize(libwebm::kMkvTagString, tag_string_);
+
+  const uint64_t simple_tag_size =
+      EbmlMasterElementSize(libwebm::kMkvSimpleTag, payload_size) +
+      payload_size;
+
+  if (writer == NULL)
+    return simple_tag_size;
+
+  const int64_t start = writer->Position();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvSimpleTag, payload_size))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvTagName, tag_name_))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvTagString, tag_string_))
+    return 0;
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != simple_tag_size)
+    return 0;
+
+  return simple_tag_size;
+}
+
+// Tags Class
+
+Tags::Tags() : tags_size_(0), tags_count_(0), tags_(NULL) {}
+
+Tags::~Tags() {
+  while (tags_count_ > 0) {
+    Tag& tag = tags_[--tags_count_];
+    tag.Clear();
+  }
+
+  delete[] tags_;
+  tags_ = NULL;
+}
+
+int Tags::Count() const { return tags_count_; }
+
+Tag* Tags::AddTag() {
+  if (!ExpandTagsArray())
+    return NULL;
+
+  Tag& tag = tags_[tags_count_++];
+
+  return &tag;
+}
+
+bool Tags::Write(IMkvWriter* writer) const {
+  if (writer == NULL)
+    return false;
+
+  uint64_t payload_size = 0;
+
+  for (int idx = 0; idx < tags_count_; ++idx) {
+    const Tag& tag = tags_[idx];
+    payload_size += tag.Write(NULL);
+  }
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvTags, payload_size))
+    return false;
+
+  const int64_t start = writer->Position();
+
+  for (int idx = 0; idx < tags_count_; ++idx) {
+    const Tag& tag = tags_[idx];
+
+    const uint64_t tag_size = tag.Write(writer);
+    if (tag_size == 0)  // error
+      return 0;
+  }
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != payload_size)
+    return false;
+
+  return true;
+}
+
+bool Tags::ExpandTagsArray() {
+  if (tags_size_ > tags_count_)
+    return true;  // nothing to do yet
+
+  const int size = (tags_size_ == 0) ? 1 : 2 * tags_size_;
+
+  Tag* const tags = new (std::nothrow) Tag[size];  // NOLINT
+  if (tags == NULL)
+    return false;
+
+  for (int idx = 0; idx < tags_count_; ++idx) {
+    const Tag& src = tags_[idx];
+    Tag* const dst = tags + idx;
+    src.ShallowCopy(dst);
+  }
+
+  delete[] tags_;
+
+  tags_ = tags;
+  tags_size_ = size;
+
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Cluster class
+
+Cluster::Cluster(uint64_t timecode, int64_t cues_pos, uint64_t timecode_scale,
+                 bool write_last_frame_with_duration, bool fixed_size_timecode)
+    : blocks_added_(0),
+      finalized_(false),
+      fixed_size_timecode_(fixed_size_timecode),
+      header_written_(false),
+      payload_size_(0),
+      position_for_cues_(cues_pos),
+      size_position_(-1),
+      timecode_(timecode),
+      timecode_scale_(timecode_scale),
+      write_last_frame_with_duration_(write_last_frame_with_duration),
+      writer_(NULL) {}
+
+Cluster::~Cluster() {}
+
+bool Cluster::Init(IMkvWriter* ptr_writer) {
+  if (!ptr_writer) {
+    return false;
+  }
+  writer_ = ptr_writer;
+  return true;
+}
+
+bool Cluster::AddFrame(const Frame* const frame) {
+  return QueueOrWriteFrame(frame);
+}
+
+bool Cluster::AddFrame(const uint8_t* data, uint64_t length,
+                       uint64_t track_number, uint64_t abs_timecode,
+                       bool is_key) {
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_track_number(track_number);
+  frame.set_timestamp(abs_timecode);
+  frame.set_is_key(is_key);
+  return QueueOrWriteFrame(&frame);
+}
+
+bool Cluster::AddFrameWithAdditional(const uint8_t* data, uint64_t length,
+                                     const uint8_t* additional,
+                                     uint64_t additional_length,
+                                     uint64_t add_id, uint64_t track_number,
+                                     uint64_t abs_timecode, bool is_key) {
+  if (!additional || additional_length == 0) {
+    return false;
+  }
+  Frame frame;
+  if (!frame.Init(data, length) ||
+      !frame.AddAdditionalData(additional, additional_length, add_id)) {
+    return false;
+  }
+  frame.set_track_number(track_number);
+  frame.set_timestamp(abs_timecode);
+  frame.set_is_key(is_key);
+  return QueueOrWriteFrame(&frame);
+}
+
+bool Cluster::AddFrameWithDiscardPadding(const uint8_t* data, uint64_t length,
+                                         int64_t discard_padding,
+                                         uint64_t track_number,
+                                         uint64_t abs_timecode, bool is_key) {
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_discard_padding(discard_padding);
+  frame.set_track_number(track_number);
+  frame.set_timestamp(abs_timecode);
+  frame.set_is_key(is_key);
+  return QueueOrWriteFrame(&frame);
+}
+
+bool Cluster::AddMetadata(const uint8_t* data, uint64_t length,
+                          uint64_t track_number, uint64_t abs_timecode,
+                          uint64_t duration_timecode) {
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_track_number(track_number);
+  frame.set_timestamp(abs_timecode);
+  frame.set_duration(duration_timecode);
+  frame.set_is_key(true);  // All metadata blocks are keyframes.
+  return QueueOrWriteFrame(&frame);
+}
+
+void Cluster::AddPayloadSize(uint64_t size) { payload_size_ += size; }
+
+bool Cluster::Finalize() {
+  return !write_last_frame_with_duration_ && Finalize(false, 0);
+}
+
+bool Cluster::Finalize(bool set_last_frame_duration, uint64_t duration) {
+  if (!writer_ || finalized_)
+    return false;
+
+  if (write_last_frame_with_duration_) {
+    // Write out held back Frames. This essentially performs a k-way merge
+    // across all tracks in the increasing order of timestamps.
+    while (!stored_frames_.empty()) {
+      Frame* frame = stored_frames_.begin()->second.front();
+
+      // Get the next frame to write (frame with least timestamp across all
+      // tracks).
+      for (FrameMapIterator frames_iterator = ++stored_frames_.begin();
+           frames_iterator != stored_frames_.end(); ++frames_iterator) {
+        if (frames_iterator->second.front()->timestamp() < frame->timestamp()) {
+          frame = frames_iterator->second.front();
+        }
+      }
+
+      // Set the duration if it's the last frame for the track.
+      if (set_last_frame_duration &&
+          stored_frames_[frame->track_number()].size() == 1 &&
+          !frame->duration_set()) {
+        frame->set_duration(duration - frame->timestamp());
+        if (!frame->is_key() && !frame->reference_block_timestamp_set()) {
+          frame->set_reference_block_timestamp(
+              last_block_timestamp_[frame->track_number()]);
+        }
+      }
+
+      // Write the frame and remove it from |stored_frames_|.
+      const bool wrote_frame = DoWriteFrame(frame);
+      stored_frames_[frame->track_number()].pop_front();
+      if (stored_frames_[frame->track_number()].empty()) {
+        stored_frames_.erase(frame->track_number());
+      }
+      delete frame;
+      if (!wrote_frame)
+        return false;
+    }
+  }
+
+  if (size_position_ == -1)
+    return false;
+
+  if (writer_->Seekable()) {
+    const int64_t pos = writer_->Position();
+
+    if (writer_->Position(size_position_))
+      return false;
+
+    if (WriteUIntSize(writer_, payload_size(), 8))
+      return false;
+
+    if (writer_->Position(pos))
+      return false;
+  }
+
+  finalized_ = true;
+
+  return true;
+}
+
+uint64_t Cluster::Size() const {
+  const uint64_t element_size =
+      EbmlMasterElementSize(libwebm::kMkvCluster, 0xFFFFFFFFFFFFFFFFULL) +
+      payload_size_;
+  return element_size;
+}
+
+bool Cluster::PreWriteBlock() {
+  if (finalized_)
+    return false;
+
+  if (!header_written_) {
+    if (!WriteClusterHeader())
+      return false;
+  }
+
+  return true;
+}
+
+void Cluster::PostWriteBlock(uint64_t element_size) {
+  AddPayloadSize(element_size);
+  ++blocks_added_;
+}
+
+int64_t Cluster::GetRelativeTimecode(int64_t abs_timecode) const {
+  const int64_t cluster_timecode = this->Cluster::timecode();
+  const int64_t rel_timecode =
+      static_cast<int64_t>(abs_timecode) - cluster_timecode;
+
+  if (rel_timecode < 0 || rel_timecode > kMaxBlockTimecode)
+    return -1;
+
+  return rel_timecode;
+}
+
+bool Cluster::DoWriteFrame(const Frame* const frame) {
+  if (!frame || !frame->IsValid())
+    return false;
+
+  if (!PreWriteBlock())
+    return false;
+
+  const uint64_t element_size = WriteFrame(writer_, frame, this);
+  if (element_size == 0)
+    return false;
+
+  PostWriteBlock(element_size);
+  last_block_timestamp_[frame->track_number()] = frame->timestamp();
+  return true;
+}
+
+bool Cluster::QueueOrWriteFrame(const Frame* const frame) {
+  if (!frame || !frame->IsValid())
+    return false;
+
+  // If |write_last_frame_with_duration_| is not set, then write the frame right
+  // away.
+  if (!write_last_frame_with_duration_) {
+    return DoWriteFrame(frame);
+  }
+
+  // Queue the current frame.
+  uint64_t track_number = frame->track_number();
+  Frame* const frame_to_store = new Frame();
+  frame_to_store->CopyFrom(*frame);
+  stored_frames_[track_number].push_back(frame_to_store);
+
+  // Iterate through all queued frames in the current track except the last one
+  // and write it if it is okay to do so (i.e.) no other track has an held back
+  // frame with timestamp <= the timestamp of the frame in question.
+  std::vector<std::list<Frame*>::iterator> frames_to_erase;
+  for (std::list<Frame *>::iterator
+           current_track_iterator = stored_frames_[track_number].begin(),
+           end = --stored_frames_[track_number].end();
+       current_track_iterator != end; ++current_track_iterator) {
+    const Frame* const frame_to_write = *current_track_iterator;
+    bool okay_to_write = true;
+    for (FrameMapIterator track_iterator = stored_frames_.begin();
+         track_iterator != stored_frames_.end(); ++track_iterator) {
+      if (track_iterator->first == track_number) {
+        continue;
+      }
+      if (track_iterator->second.front()->timestamp() <
+          frame_to_write->timestamp()) {
+        okay_to_write = false;
+        break;
+      }
+    }
+    if (okay_to_write) {
+      const bool wrote_frame = DoWriteFrame(frame_to_write);
+      delete frame_to_write;
+      if (!wrote_frame)
+        return false;
+      frames_to_erase.push_back(current_track_iterator);
+    } else {
+      break;
+    }
+  }
+  for (std::vector<std::list<Frame*>::iterator>::iterator iterator =
+           frames_to_erase.begin();
+       iterator != frames_to_erase.end(); ++iterator) {
+    stored_frames_[track_number].erase(*iterator);
+  }
+  return true;
+}
+
+bool Cluster::WriteClusterHeader() {
+  if (finalized_)
+    return false;
+
+  if (WriteID(writer_, libwebm::kMkvCluster))
+    return false;
+
+  // Save for later.
+  size_position_ = writer_->Position();
+
+  // Write "unknown" (EBML coded -1) as cluster size value. We need to write 8
+  // bytes because we do not know how big our cluster will be.
+  if (SerializeInt(writer_, kEbmlUnknownValue, 8))
+    return false;
+
+  if (!WriteEbmlElement(writer_, libwebm::kMkvTimecode, timecode(),
+                        fixed_size_timecode_ ? 8 : 0)) {
+    return false;
+  }
+  AddPayloadSize(EbmlElementSize(libwebm::kMkvTimecode, timecode(),
+                                 fixed_size_timecode_ ? 8 : 0));
+  header_written_ = true;
+
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// SeekHead Class
+
+SeekHead::SeekHead() : start_pos_(0ULL) {
+  for (int32_t i = 0; i < kSeekEntryCount; ++i) {
+    seek_entry_id_[i] = 0;
+    seek_entry_pos_[i] = 0;
+  }
+}
+
+SeekHead::~SeekHead() {}
+
+bool SeekHead::Finalize(IMkvWriter* writer) const {
+  if (writer->Seekable()) {
+    if (start_pos_ == -1)
+      return false;
+
+    uint64_t payload_size = 0;
+    uint64_t entry_size[kSeekEntryCount];
+
+    for (int32_t i = 0; i < kSeekEntryCount; ++i) {
+      if (seek_entry_id_[i] != 0) {
+        entry_size[i] = EbmlElementSize(
+            libwebm::kMkvSeekID, static_cast<uint64_t>(seek_entry_id_[i]));
+        entry_size[i] +=
+            EbmlElementSize(libwebm::kMkvSeekPosition, seek_entry_pos_[i]);
+
+        payload_size +=
+            EbmlMasterElementSize(libwebm::kMkvSeek, entry_size[i]) +
+            entry_size[i];
+      }
+    }
+
+    // No SeekHead elements
+    if (payload_size == 0)
+      return true;
+
+    const int64_t pos = writer->Position();
+    if (writer->Position(start_pos_))
+      return false;
+
+    if (!WriteEbmlMasterElement(writer, libwebm::kMkvSeekHead, payload_size))
+      return false;
+
+    for (int32_t i = 0; i < kSeekEntryCount; ++i) {
+      if (seek_entry_id_[i] != 0) {
+        if (!WriteEbmlMasterElement(writer, libwebm::kMkvSeek, entry_size[i]))
+          return false;
+
+        if (!WriteEbmlElement(writer, libwebm::kMkvSeekID,
+                              static_cast<uint64_t>(seek_entry_id_[i])))
+          return false;
+
+        if (!WriteEbmlElement(writer, libwebm::kMkvSeekPosition,
+                              seek_entry_pos_[i]))
+          return false;
+      }
+    }
+
+    const uint64_t total_entry_size = kSeekEntryCount * MaxEntrySize();
+    const uint64_t total_size =
+        EbmlMasterElementSize(libwebm::kMkvSeekHead, total_entry_size) +
+        total_entry_size;
+    const int64_t size_left = total_size - (writer->Position() - start_pos_);
+
+    const uint64_t bytes_written = WriteVoidElement(writer, size_left);
+    if (!bytes_written)
+      return false;
+
+    if (writer->Position(pos))
+      return false;
+  }
+
+  return true;
+}
+
+bool SeekHead::Write(IMkvWriter* writer) {
+  const uint64_t entry_size = kSeekEntryCount * MaxEntrySize();
+  const uint64_t size =
+      EbmlMasterElementSize(libwebm::kMkvSeekHead, entry_size);
+
+  start_pos_ = writer->Position();
+
+  const uint64_t bytes_written = WriteVoidElement(writer, size + entry_size);
+  if (!bytes_written)
+    return false;
+
+  return true;
+}
+
+bool SeekHead::AddSeekEntry(uint32_t id, uint64_t pos) {
+  for (int32_t i = 0; i < kSeekEntryCount; ++i) {
+    if (seek_entry_id_[i] == 0) {
+      seek_entry_id_[i] = id;
+      seek_entry_pos_[i] = pos;
+      return true;
+    }
+  }
+  return false;
+}
+
+uint32_t SeekHead::GetId(int index) const {
+  if (index < 0 || index >= kSeekEntryCount)
+    return UINT_MAX;
+  return seek_entry_id_[index];
+}
+
+uint64_t SeekHead::GetPosition(int index) const {
+  if (index < 0 || index >= kSeekEntryCount)
+    return ULLONG_MAX;
+  return seek_entry_pos_[index];
+}
+
+bool SeekHead::SetSeekEntry(int index, uint32_t id, uint64_t position) {
+  if (index < 0 || index >= kSeekEntryCount)
+    return false;
+  seek_entry_id_[index] = id;
+  seek_entry_pos_[index] = position;
+  return true;
+}
+
+uint64_t SeekHead::MaxEntrySize() const {
+  const uint64_t max_entry_payload_size =
+      EbmlElementSize(libwebm::kMkvSeekID, UINT64_C(0xffffffff)) +
+      EbmlElementSize(libwebm::kMkvSeekPosition, UINT64_C(0xffffffffffffffff));
+  const uint64_t max_entry_size =
+      EbmlMasterElementSize(libwebm::kMkvSeek, max_entry_payload_size) +
+      max_entry_payload_size;
+
+  return max_entry_size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// SegmentInfo Class
+
+SegmentInfo::SegmentInfo()
+    : duration_(-1.0),
+      muxing_app_(NULL),
+      timecode_scale_(1000000ULL),
+      writing_app_(NULL),
+      date_utc_(LLONG_MIN),
+      duration_pos_(-1) {}
+
+SegmentInfo::~SegmentInfo() {
+  delete[] muxing_app_;
+  delete[] writing_app_;
+}
+
+bool SegmentInfo::Init() {
+  int32_t major;
+  int32_t minor;
+  int32_t build;
+  int32_t revision;
+  GetVersion(&major, &minor, &build, &revision);
+  char temp[256];
+#ifdef _MSC_VER
+  sprintf_s(temp, sizeof(temp) / sizeof(temp[0]), "libwebm-%d.%d.%d.%d", major,
+            minor, build, revision);
+#else
+  snprintf(temp, sizeof(temp) / sizeof(temp[0]), "libwebm-%d.%d.%d.%d", major,
+           minor, build, revision);
+#endif
+
+  const size_t app_len = strlen(temp) + 1;
+
+  delete[] muxing_app_;
+
+  muxing_app_ = new (std::nothrow) char[app_len];  // NOLINT
+  if (!muxing_app_)
+    return false;
+
+#ifdef _MSC_VER
+  strcpy_s(muxing_app_, app_len, temp);
+#else
+  strcpy(muxing_app_, temp);
+#endif
+
+  set_writing_app(temp);
+  if (!writing_app_)
+    return false;
+  return true;
+}
+
+bool SegmentInfo::Finalize(IMkvWriter* writer) const {
+  if (!writer)
+    return false;
+
+  if (duration_ > 0.0) {
+    if (writer->Seekable()) {
+      if (duration_pos_ == -1)
+        return false;
+
+      const int64_t pos = writer->Position();
+
+      if (writer->Position(duration_pos_))
+        return false;
+
+      if (!WriteEbmlElement(writer, libwebm::kMkvDuration,
+                            static_cast<float>(duration_)))
+        return false;
+
+      if (writer->Position(pos))
+        return false;
+    }
+  }
+
+  return true;
+}
+
+bool SegmentInfo::Write(IMkvWriter* writer) {
+  if (!writer || !muxing_app_ || !writing_app_)
+    return false;
+
+  uint64_t size = EbmlElementSize(libwebm::kMkvTimecodeScale, timecode_scale_);
+  if (duration_ > 0.0)
+    size +=
+        EbmlElementSize(libwebm::kMkvDuration, static_cast<float>(duration_));
+  if (date_utc_ != LLONG_MIN)
+    size += EbmlDateElementSize(libwebm::kMkvDateUTC);
+  size += EbmlElementSize(libwebm::kMkvMuxingApp, muxing_app_);
+  size += EbmlElementSize(libwebm::kMkvWritingApp, writing_app_);
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvInfo, size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvTimecodeScale, timecode_scale_))
+    return false;
+
+  if (duration_ > 0.0) {
+    // Save for later
+    duration_pos_ = writer->Position();
+
+    if (!WriteEbmlElement(writer, libwebm::kMkvDuration,
+                          static_cast<float>(duration_)))
+      return false;
+  }
+
+  if (date_utc_ != LLONG_MIN)
+    WriteEbmlDateElement(writer, libwebm::kMkvDateUTC, date_utc_);
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvMuxingApp, muxing_app_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvWritingApp, writing_app_))
+    return false;
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  return true;
+}
+
+void SegmentInfo::set_muxing_app(const char* app) {
+  if (app) {
+    const size_t length = strlen(app) + 1;
+    char* temp_str = new (std::nothrow) char[length];  // NOLINT
+    if (!temp_str)
+      return;
+
+#ifdef _MSC_VER
+    strcpy_s(temp_str, length, app);
+#else
+    strcpy(temp_str, app);
+#endif
+
+    delete[] muxing_app_;
+    muxing_app_ = temp_str;
+  }
+}
+
+void SegmentInfo::set_writing_app(const char* app) {
+  if (app) {
+    const size_t length = strlen(app) + 1;
+    char* temp_str = new (std::nothrow) char[length];  // NOLINT
+    if (!temp_str)
+      return;
+
+#ifdef _MSC_VER
+    strcpy_s(temp_str, length, app);
+#else
+    strcpy(temp_str, app);
+#endif
+
+    delete[] writing_app_;
+    writing_app_ = temp_str;
+  }
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Segment Class
+
+Segment::Segment()
+    : chunk_count_(0),
+      chunk_name_(NULL),
+      chunk_writer_cluster_(NULL),
+      chunk_writer_cues_(NULL),
+      chunk_writer_header_(NULL),
+      chunking_(false),
+      chunking_base_name_(NULL),
+      cluster_list_(NULL),
+      cluster_list_capacity_(0),
+      cluster_list_size_(0),
+      cues_position_(kAfterClusters),
+      cues_track_(0),
+      force_new_cluster_(false),
+      frames_(NULL),
+      frames_capacity_(0),
+      frames_size_(0),
+      has_video_(false),
+      header_written_(false),
+      last_block_duration_(0),
+      last_timestamp_(0),
+      max_cluster_duration_(kDefaultMaxClusterDuration),
+      max_cluster_size_(0),
+      mode_(kFile),
+      new_cuepoint_(false),
+      output_cues_(true),
+      accurate_cluster_duration_(false),
+      fixed_size_cluster_timecode_(false),
+      payload_pos_(0),
+      size_position_(0),
+      doc_type_version_(kDefaultDocTypeVersion),
+      doc_type_version_written_(0),
+      writer_cluster_(NULL),
+      writer_cues_(NULL),
+      writer_header_(NULL) {
+  const time_t curr_time = time(NULL);
+  seed_ = static_cast<unsigned int>(curr_time);
+#ifdef _WIN32
+  srand(seed_);
+#endif
+}
+
+Segment::~Segment() {
+  if (cluster_list_) {
+    for (int32_t i = 0; i < cluster_list_size_; ++i) {
+      Cluster* const cluster = cluster_list_[i];
+      delete cluster;
+    }
+    delete[] cluster_list_;
+  }
+
+  if (frames_) {
+    for (int32_t i = 0; i < frames_size_; ++i) {
+      Frame* const frame = frames_[i];
+      delete frame;
+    }
+    delete[] frames_;
+  }
+
+  delete[] chunk_name_;
+  delete[] chunking_base_name_;
+
+  if (chunk_writer_cluster_) {
+    chunk_writer_cluster_->Close();
+    delete chunk_writer_cluster_;
+  }
+  if (chunk_writer_cues_) {
+    chunk_writer_cues_->Close();
+    delete chunk_writer_cues_;
+  }
+  if (chunk_writer_header_) {
+    chunk_writer_header_->Close();
+    delete chunk_writer_header_;
+  }
+}
+
+void Segment::MoveCuesBeforeClustersHelper(uint64_t diff, int32_t index,
+                                           uint64_t* cues_size) {
+  CuePoint* const cue_point = cues_.GetCueByIndex(index);
+  if (cue_point == NULL)
+    return;
+  const uint64_t old_cue_point_size = cue_point->Size();
+  const uint64_t cluster_pos = cue_point->cluster_pos() + diff;
+  cue_point->set_cluster_pos(cluster_pos);  // update the new cluster position
+  // New size of the cue is computed as follows
+  //    Let a = current sum of size of all CuePoints
+  //    Let b = Increase in Cue Point's size due to this iteration
+  //    Let c = Increase in size of Cues Element's length due to this iteration
+  //            (This is computed as CodedSize(a + b) - CodedSize(a))
+  //    Let d = b + c. Now d is the |diff| passed to the next recursive call.
+  //    Let e = a + b. Now e is the |cues_size| passed to the next recursive
+  //                   call.
+  const uint64_t cue_point_size_diff = cue_point->Size() - old_cue_point_size;
+  const uint64_t cue_size_diff =
+      GetCodedUIntSize(*cues_size + cue_point_size_diff) -
+      GetCodedUIntSize(*cues_size);
+  *cues_size += cue_point_size_diff;
+  diff = cue_size_diff + cue_point_size_diff;
+  if (diff > 0) {
+    for (int32_t i = 0; i < cues_.cue_entries_size(); ++i) {
+      MoveCuesBeforeClustersHelper(diff, i, cues_size);
+    }
+  }
+}
+
+void Segment::MoveCuesBeforeClusters() {
+  const uint64_t current_cue_size = cues_.Size();
+  uint64_t cue_size = 0;
+  for (int32_t i = 0; i < cues_.cue_entries_size(); ++i)
+    cue_size += cues_.GetCueByIndex(i)->Size();
+  for (int32_t i = 0; i < cues_.cue_entries_size(); ++i)
+    MoveCuesBeforeClustersHelper(current_cue_size, i, &cue_size);
+
+  // Adjust the Seek Entry to reflect the change in position
+  // of Cluster and Cues
+  int32_t cluster_index = 0;
+  int32_t cues_index = 0;
+  for (int32_t i = 0; i < SeekHead::kSeekEntryCount; ++i) {
+    if (seek_head_.GetId(i) == libwebm::kMkvCluster)
+      cluster_index = i;
+    if (seek_head_.GetId(i) == libwebm::kMkvCues)
+      cues_index = i;
+  }
+  seek_head_.SetSeekEntry(cues_index, libwebm::kMkvCues,
+                          seek_head_.GetPosition(cluster_index));
+  seek_head_.SetSeekEntry(cluster_index, libwebm::kMkvCluster,
+                          cues_.Size() + seek_head_.GetPosition(cues_index));
+}
+
+bool Segment::Init(IMkvWriter* ptr_writer) {
+  if (!ptr_writer) {
+    return false;
+  }
+  writer_cluster_ = ptr_writer;
+  writer_cues_ = ptr_writer;
+  writer_header_ = ptr_writer;
+  return segment_info_.Init();
+}
+
+bool Segment::CopyAndMoveCuesBeforeClusters(mkvparser::IMkvReader* reader,
+                                            IMkvWriter* writer) {
+  if (!writer->Seekable() || chunking_)
+    return false;
+  const int64_t cluster_offset =
+      cluster_list_[0]->size_position() - GetUIntSize(libwebm::kMkvCluster);
+
+  // Copy the headers.
+  if (!ChunkedCopy(reader, writer, 0, cluster_offset))
+    return false;
+
+  // Recompute cue positions and seek entries.
+  MoveCuesBeforeClusters();
+
+  // Write cues and seek entries.
+  // TODO(vigneshv): As of now, it's safe to call seek_head_.Finalize() for the
+  // second time with a different writer object. But the name Finalize() doesn't
+  // indicate something we want to call more than once. So consider renaming it
+  // to write() or some such.
+  if (!cues_.Write(writer) || !seek_head_.Finalize(writer))
+    return false;
+
+  // Copy the Clusters.
+  if (!ChunkedCopy(reader, writer, cluster_offset,
+                   cluster_end_offset_ - cluster_offset))
+    return false;
+
+  // Update the Segment size in case the Cues size has changed.
+  const int64_t pos = writer->Position();
+  const int64_t segment_size = writer->Position() - payload_pos_;
+  if (writer->Position(size_position_) ||
+      WriteUIntSize(writer, segment_size, 8) || writer->Position(pos))
+    return false;
+  return true;
+}
+
+bool Segment::Finalize() {
+  if (WriteFramesAll() < 0)
+    return false;
+
+  if (cluster_list_size_ > 0) {
+    // Update last cluster's size
+    Cluster* const old_cluster = cluster_list_[cluster_list_size_ - 1];
+
+    // For the last frame of the last Cluster, we don't write it as a BlockGroup
+    // with Duration unless the frame itself has duration set explicitly.
+    if (!old_cluster || !old_cluster->Finalize(false, 0))
+      return false;
+  }
+
+  if (mode_ == kFile) {
+    if (chunking_ && chunk_writer_cluster_) {
+      chunk_writer_cluster_->Close();
+      chunk_count_++;
+    }
+
+    const double duration =
+        (static_cast<double>(last_timestamp_) + last_block_duration_) /
+        segment_info_.timecode_scale();
+    segment_info_.set_duration(duration);
+    if (!segment_info_.Finalize(writer_header_))
+      return false;
+
+    if (output_cues_)
+      if (!seek_head_.AddSeekEntry(libwebm::kMkvCues, MaxOffset()))
+        return false;
+
+    if (chunking_) {
+      if (!chunk_writer_cues_)
+        return false;
+
+      char* name = NULL;
+      if (!UpdateChunkName("cues", &name))
+        return false;
+
+      const bool cues_open = chunk_writer_cues_->Open(name);
+      delete[] name;
+      if (!cues_open)
+        return false;
+    }
+
+    cluster_end_offset_ = writer_cluster_->Position();
+
+    // Write the seek headers and cues
+    if (output_cues_)
+      if (!cues_.Write(writer_cues_))
+        return false;
+
+    if (!seek_head_.Finalize(writer_header_))
+      return false;
+
+    if (writer_header_->Seekable()) {
+      if (size_position_ == -1)
+        return false;
+
+      const int64_t segment_size = MaxOffset();
+      if (segment_size < 1)
+        return false;
+
+      const int64_t pos = writer_header_->Position();
+      UpdateDocTypeVersion();
+      if (doc_type_version_ != doc_type_version_written_) {
+        if (writer_header_->Position(0))
+          return false;
+
+        if (!WriteEbmlHeader(writer_header_, doc_type_version_))
+          return false;
+        if (writer_header_->Position() != ebml_header_size_)
+          return false;
+
+        doc_type_version_written_ = doc_type_version_;
+      }
+
+      if (writer_header_->Position(size_position_))
+        return false;
+
+      if (WriteUIntSize(writer_header_, segment_size, 8))
+        return false;
+
+      if (writer_header_->Position(pos))
+        return false;
+    }
+
+    if (chunking_) {
+      // Do not close any writers until the segment size has been written,
+      // otherwise the size may be off.
+      if (!chunk_writer_cues_ || !chunk_writer_header_)
+        return false;
+
+      chunk_writer_cues_->Close();
+      chunk_writer_header_->Close();
+    }
+  }
+
+  return true;
+}
+
+Track* Segment::AddTrack(int32_t number) {
+  Track* const track = new (std::nothrow) Track(&seed_);  // NOLINT
+
+  if (!track)
+    return NULL;
+
+  if (!tracks_.AddTrack(track, number)) {
+    delete track;
+    return NULL;
+  }
+
+  return track;
+}
+
+Chapter* Segment::AddChapter() { return chapters_.AddChapter(&seed_); }
+
+Tag* Segment::AddTag() { return tags_.AddTag(); }
+
+uint64_t Segment::AddVideoTrack(int32_t width, int32_t height, int32_t number) {
+  VideoTrack* const track = new (std::nothrow) VideoTrack(&seed_);  // NOLINT
+  if (!track)
+    return 0;
+
+  track->set_type(Tracks::kVideo);
+  track->set_codec_id(Tracks::kVp8CodecId);
+  track->set_width(width);
+  track->set_height(height);
+
+  tracks_.AddTrack(track, number);
+  has_video_ = true;
+
+  return track->number();
+}
+
+bool Segment::AddCuePoint(uint64_t timestamp, uint64_t track) {
+  if (cluster_list_size_ < 1)
+    return false;
+
+  const Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
+  if (!cluster)
+    return false;
+
+  CuePoint* const cue = new (std::nothrow) CuePoint();  // NOLINT
+  if (!cue)
+    return false;
+
+  cue->set_time(timestamp / segment_info_.timecode_scale());
+  cue->set_block_number(cluster->blocks_added());
+  cue->set_cluster_pos(cluster->position_for_cues());
+  cue->set_track(track);
+  if (!cues_.AddCue(cue))
+    return false;
+
+  new_cuepoint_ = false;
+  return true;
+}
+
+uint64_t Segment::AddAudioTrack(int32_t sample_rate, int32_t channels,
+                                int32_t number) {
+  AudioTrack* const track = new (std::nothrow) AudioTrack(&seed_);  // NOLINT
+  if (!track)
+    return 0;
+
+  track->set_type(Tracks::kAudio);
+  track->set_codec_id(Tracks::kVorbisCodecId);
+  track->set_sample_rate(sample_rate);
+  track->set_channels(channels);
+
+  tracks_.AddTrack(track, number);
+
+  return track->number();
+}
+
+bool Segment::AddFrame(const uint8_t* data, uint64_t length,
+                       uint64_t track_number, uint64_t timestamp, bool is_key) {
+  if (!data)
+    return false;
+
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_track_number(track_number);
+  frame.set_timestamp(timestamp);
+  frame.set_is_key(is_key);
+  return AddGenericFrame(&frame);
+}
+
+bool Segment::AddFrameWithAdditional(const uint8_t* data, uint64_t length,
+                                     const uint8_t* additional,
+                                     uint64_t additional_length,
+                                     uint64_t add_id, uint64_t track_number,
+                                     uint64_t timestamp, bool is_key) {
+  if (!data || !additional)
+    return false;
+
+  Frame frame;
+  if (!frame.Init(data, length) ||
+      !frame.AddAdditionalData(additional, additional_length, add_id)) {
+    return false;
+  }
+  frame.set_track_number(track_number);
+  frame.set_timestamp(timestamp);
+  frame.set_is_key(is_key);
+  return AddGenericFrame(&frame);
+}
+
+bool Segment::AddFrameWithDiscardPadding(const uint8_t* data, uint64_t length,
+                                         int64_t discard_padding,
+                                         uint64_t track_number,
+                                         uint64_t timestamp, bool is_key) {
+  if (!data)
+    return false;
+
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_discard_padding(discard_padding);
+  frame.set_track_number(track_number);
+  frame.set_timestamp(timestamp);
+  frame.set_is_key(is_key);
+  return AddGenericFrame(&frame);
+}
+
+bool Segment::AddMetadata(const uint8_t* data, uint64_t length,
+                          uint64_t track_number, uint64_t timestamp_ns,
+                          uint64_t duration_ns) {
+  if (!data)
+    return false;
+
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_track_number(track_number);
+  frame.set_timestamp(timestamp_ns);
+  frame.set_duration(duration_ns);
+  frame.set_is_key(true);  // All metadata blocks are keyframes.
+  return AddGenericFrame(&frame);
+}
+
+bool Segment::AddGenericFrame(const Frame* frame) {
+  if (!frame)
+    return false;
+
+  if (!CheckHeaderInfo())
+    return false;
+
+  // Check for non-monotonically increasing timestamps.
+  if (frame->timestamp() < last_timestamp_)
+    return false;
+
+  // Check if the track number is valid.
+  if (!tracks_.GetTrackByNumber(frame->track_number()))
+    return false;
+
+  if (frame->discard_padding() != 0)
+    doc_type_version_ = 4;
+
+  // If the segment has a video track hold onto audio frames to make sure the
+  // audio that is associated with the start time of a video key-frame is
+  // muxed into the same cluster.
+  if (has_video_ && tracks_.TrackIsAudio(frame->track_number()) &&
+      !force_new_cluster_) {
+    Frame* const new_frame = new (std::nothrow) Frame();
+    if (!new_frame || !new_frame->CopyFrom(*frame))
+      return false;
+    return QueueFrame(new_frame);
+  }
+
+  if (!DoNewClusterProcessing(frame->track_number(), frame->timestamp(),
+                              frame->is_key())) {
+    return false;
+  }
+
+  if (cluster_list_size_ < 1)
+    return false;
+
+  Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
+  if (!cluster)
+    return false;
+
+  // If the Frame is not a SimpleBlock, then set the reference_block_timestamp
+  // if it is not set already.
+  bool frame_created = false;
+  if (!frame->CanBeSimpleBlock() && !frame->is_key() &&
+      !frame->reference_block_timestamp_set()) {
+    Frame* const new_frame = new (std::nothrow) Frame();
+    if (!new_frame->CopyFrom(*frame))
+      return false;
+    new_frame->set_reference_block_timestamp(
+        last_track_timestamp_[frame->track_number() - 1]);
+    frame = new_frame;
+    frame_created = true;
+  }
+
+  if (!cluster->AddFrame(frame))
+    return false;
+
+  if (new_cuepoint_ && cues_track_ == frame->track_number()) {
+    if (!AddCuePoint(frame->timestamp(), cues_track_))
+      return false;
+  }
+
+  last_timestamp_ = frame->timestamp();
+  last_track_timestamp_[frame->track_number() - 1] = frame->timestamp();
+  last_block_duration_ = frame->duration();
+
+  if (frame_created)
+    delete frame;
+
+  return true;
+}
+
+void Segment::OutputCues(bool output_cues) { output_cues_ = output_cues; }
+
+void Segment::AccurateClusterDuration(bool accurate_cluster_duration) {
+  accurate_cluster_duration_ = accurate_cluster_duration;
+}
+
+void Segment::UseFixedSizeClusterTimecode(bool fixed_size_cluster_timecode) {
+  fixed_size_cluster_timecode_ = fixed_size_cluster_timecode;
+}
+
+bool Segment::SetChunking(bool chunking, const char* filename) {
+  if (chunk_count_ > 0)
+    return false;
+
+  if (chunking) {
+    if (!filename)
+      return false;
+
+    // Check if we are being set to what is already set.
+    if (chunking_ && !strcmp(filename, chunking_base_name_))
+      return true;
+
+    const size_t name_length = strlen(filename) + 1;
+    char* const temp = new (std::nothrow) char[name_length];  // NOLINT
+    if (!temp)
+      return false;
+
+#ifdef _MSC_VER
+    strcpy_s(temp, name_length, filename);
+#else
+    strcpy(temp, filename);
+#endif
+
+    delete[] chunking_base_name_;
+    chunking_base_name_ = temp;
+
+    if (!UpdateChunkName("chk", &chunk_name_))
+      return false;
+
+    if (!chunk_writer_cluster_) {
+      chunk_writer_cluster_ = new (std::nothrow) MkvWriter();  // NOLINT
+      if (!chunk_writer_cluster_)
+        return false;
+    }
+
+    if (!chunk_writer_cues_) {
+      chunk_writer_cues_ = new (std::nothrow) MkvWriter();  // NOLINT
+      if (!chunk_writer_cues_)
+        return false;
+    }
+
+    if (!chunk_writer_header_) {
+      chunk_writer_header_ = new (std::nothrow) MkvWriter();  // NOLINT
+      if (!chunk_writer_header_)
+        return false;
+    }
+
+    if (!chunk_writer_cluster_->Open(chunk_name_))
+      return false;
+
+    const size_t header_length = strlen(filename) + strlen(".hdr") + 1;
+    char* const header = new (std::nothrow) char[header_length];  // NOLINT
+    if (!header)
+      return false;
+
+#ifdef _MSC_VER
+    strcpy_s(header, header_length - strlen(".hdr"), chunking_base_name_);
+    strcat_s(header, header_length, ".hdr");
+#else
+    strcpy(header, chunking_base_name_);
+    strcat(header, ".hdr");
+#endif
+    if (!chunk_writer_header_->Open(header)) {
+      delete[] header;
+      return false;
+    }
+
+    writer_cluster_ = chunk_writer_cluster_;
+    writer_cues_ = chunk_writer_cues_;
+    writer_header_ = chunk_writer_header_;
+
+    delete[] header;
+  }
+
+  chunking_ = chunking;
+
+  return true;
+}
+
+bool Segment::CuesTrack(uint64_t track_number) {
+  const Track* const track = GetTrackByNumber(track_number);
+  if (!track)
+    return false;
+
+  cues_track_ = track_number;
+  return true;
+}
+
+void Segment::ForceNewClusterOnNextFrame() { force_new_cluster_ = true; }
+
+Track* Segment::GetTrackByNumber(uint64_t track_number) const {
+  return tracks_.GetTrackByNumber(track_number);
+}
+
+bool Segment::WriteSegmentHeader() {
+  UpdateDocTypeVersion();
+
+  // TODO(fgalligan): Support more than one segment.
+  if (!WriteEbmlHeader(writer_header_, doc_type_version_))
+    return false;
+  doc_type_version_written_ = doc_type_version_;
+  ebml_header_size_ = static_cast<int32_t>(writer_header_->Position());
+
+  // Write "unknown" (-1) as segment size value. If mode is kFile, Segment
+  // will write over duration when the file is finalized.
+  if (WriteID(writer_header_, libwebm::kMkvSegment))
+    return false;
+
+  // Save for later.
+  size_position_ = writer_header_->Position();
+
+  // Write "unknown" (EBML coded -1) as segment size value. We need to write 8
+  // bytes because if we are going to overwrite the segment size later we do
+  // not know how big our segment will be.
+  if (SerializeInt(writer_header_, kEbmlUnknownValue, 8))
+    return false;
+
+  payload_pos_ = writer_header_->Position();
+
+  if (mode_ == kFile && writer_header_->Seekable()) {
+    // Set the duration > 0.0 so SegmentInfo will write out the duration. When
+    // the muxer is done writing we will set the correct duration and have
+    // SegmentInfo upadte it.
+    segment_info_.set_duration(1.0);
+
+    if (!seek_head_.Write(writer_header_))
+      return false;
+  }
+
+  if (!seek_head_.AddSeekEntry(libwebm::kMkvInfo, MaxOffset()))
+    return false;
+  if (!segment_info_.Write(writer_header_))
+    return false;
+
+  if (!seek_head_.AddSeekEntry(libwebm::kMkvTracks, MaxOffset()))
+    return false;
+  if (!tracks_.Write(writer_header_))
+    return false;
+
+  if (chapters_.Count() > 0) {
+    if (!seek_head_.AddSeekEntry(libwebm::kMkvChapters, MaxOffset()))
+      return false;
+    if (!chapters_.Write(writer_header_))
+      return false;
+  }
+
+  if (tags_.Count() > 0) {
+    if (!seek_head_.AddSeekEntry(libwebm::kMkvTags, MaxOffset()))
+      return false;
+    if (!tags_.Write(writer_header_))
+      return false;
+  }
+
+  if (chunking_ && (mode_ == kLive || !writer_header_->Seekable())) {
+    if (!chunk_writer_header_)
+      return false;
+
+    chunk_writer_header_->Close();
+  }
+
+  header_written_ = true;
+
+  return true;
+}
+
+// Here we are testing whether to create a new cluster, given a frame
+// having time frame_timestamp_ns.
+//
+int Segment::TestFrame(uint64_t track_number, uint64_t frame_timestamp_ns,
+                       bool is_key) const {
+  if (force_new_cluster_)
+    return 1;
+
+  // If no clusters have been created yet, then create a new cluster
+  // and write this frame immediately, in the new cluster.  This path
+  // should only be followed once, the first time we attempt to write
+  // a frame.
+
+  if (cluster_list_size_ <= 0)
+    return 1;
+
+  // There exists at least one cluster. We must compare the frame to
+  // the last cluster, in order to determine whether the frame is
+  // written to the existing cluster, or that a new cluster should be
+  // created.
+
+  const uint64_t timecode_scale = segment_info_.timecode_scale();
+  const uint64_t frame_timecode = frame_timestamp_ns / timecode_scale;
+
+  const Cluster* const last_cluster = cluster_list_[cluster_list_size_ - 1];
+  const uint64_t last_cluster_timecode = last_cluster->timecode();
+
+  // For completeness we test for the case when the frame's timecode
+  // is less than the cluster's timecode.  Although in principle that
+  // is allowed, this muxer doesn't actually write clusters like that,
+  // so this indicates a bug somewhere in our algorithm.
+
+  if (frame_timecode < last_cluster_timecode)  // should never happen
+    return -1;
+
+  // If the frame has a timestamp significantly larger than the last
+  // cluster (in Matroska, cluster-relative timestamps are serialized
+  // using a 16-bit signed integer), then we cannot write this frame
+  // to that cluster, and so we must create a new cluster.
+
+  const int64_t delta_timecode = frame_timecode - last_cluster_timecode;
+
+  if (delta_timecode > kMaxBlockTimecode)
+    return 2;
+
+  // We decide to create a new cluster when we have a video keyframe.
+  // This will flush queued (audio) frames, and write the keyframe
+  // immediately, in the newly-created cluster.
+
+  if (is_key && tracks_.TrackIsVideo(track_number))
+    return 1;
+
+  // Create a new cluster if we have accumulated too many frames
+  // already, where "too many" is defined as "the total time of frames
+  // in the cluster exceeds a threshold".
+
+  const uint64_t delta_ns = delta_timecode * timecode_scale;
+
+  if (max_cluster_duration_ > 0 && delta_ns >= max_cluster_duration_)
+    return 1;
+
+  // This is similar to the case above, with the difference that a new
+  // cluster is created when the size of the current cluster exceeds a
+  // threshold.
+
+  const uint64_t cluster_size = last_cluster->payload_size();
+
+  if (max_cluster_size_ > 0 && cluster_size >= max_cluster_size_)
+    return 1;
+
+  // There's no need to create a new cluster, so emit this frame now.
+
+  return 0;
+}
+
+bool Segment::MakeNewCluster(uint64_t frame_timestamp_ns) {
+  const int32_t new_size = cluster_list_size_ + 1;
+
+  if (new_size > cluster_list_capacity_) {
+    // Add more clusters.
+    const int32_t new_capacity =
+        (cluster_list_capacity_ <= 0) ? 1 : cluster_list_capacity_ * 2;
+    Cluster** const clusters =
+        new (std::nothrow) Cluster*[new_capacity];  // NOLINT
+    if (!clusters)
+      return false;
+
+    for (int32_t i = 0; i < cluster_list_size_; ++i) {
+      clusters[i] = cluster_list_[i];
+    }
+
+    delete[] cluster_list_;
+
+    cluster_list_ = clusters;
+    cluster_list_capacity_ = new_capacity;
+  }
+
+  if (!WriteFramesLessThan(frame_timestamp_ns))
+    return false;
+
+  if (cluster_list_size_ > 0) {
+    // Update old cluster's size
+    Cluster* const old_cluster = cluster_list_[cluster_list_size_ - 1];
+
+    if (!old_cluster || !old_cluster->Finalize(true, frame_timestamp_ns))
+      return false;
+  }
+
+  if (output_cues_)
+    new_cuepoint_ = true;
+
+  if (chunking_ && cluster_list_size_ > 0) {
+    chunk_writer_cluster_->Close();
+    chunk_count_++;
+
+    if (!UpdateChunkName("chk", &chunk_name_))
+      return false;
+    if (!chunk_writer_cluster_->Open(chunk_name_))
+      return false;
+  }
+
+  const uint64_t timecode_scale = segment_info_.timecode_scale();
+  const uint64_t frame_timecode = frame_timestamp_ns / timecode_scale;
+
+  uint64_t cluster_timecode = frame_timecode;
+
+  if (frames_size_ > 0) {
+    const Frame* const f = frames_[0];  // earliest queued frame
+    const uint64_t ns = f->timestamp();
+    const uint64_t tc = ns / timecode_scale;
+
+    if (tc < cluster_timecode)
+      cluster_timecode = tc;
+  }
+
+  Cluster*& cluster = cluster_list_[cluster_list_size_];
+  const int64_t offset = MaxOffset();
+  cluster = new (std::nothrow)
+      Cluster(cluster_timecode, offset, segment_info_.timecode_scale(),
+              accurate_cluster_duration_, fixed_size_cluster_timecode_);
+  if (!cluster)
+    return false;
+
+  if (!cluster->Init(writer_cluster_))
+    return false;
+
+  cluster_list_size_ = new_size;
+  return true;
+}
+
+bool Segment::DoNewClusterProcessing(uint64_t track_number,
+                                     uint64_t frame_timestamp_ns, bool is_key) {
+  for (;;) {
+    // Based on the characteristics of the current frame and current
+    // cluster, decide whether to create a new cluster.
+    const int result = TestFrame(track_number, frame_timestamp_ns, is_key);
+    if (result < 0)  // error
+      return false;
+
+    // Always set force_new_cluster_ to false after TestFrame.
+    force_new_cluster_ = false;
+
+    // A non-zero result means create a new cluster.
+    if (result > 0 && !MakeNewCluster(frame_timestamp_ns))
+      return false;
+
+    // Write queued (audio) frames.
+    const int frame_count = WriteFramesAll();
+    if (frame_count < 0)  // error
+      return false;
+
+    // Write the current frame to the current cluster (if TestFrame
+    // returns 0) or to a newly created cluster (TestFrame returns 1).
+    if (result <= 1)
+      return true;
+
+    // TestFrame returned 2, which means there was a large time
+    // difference between the cluster and the frame itself.  Do the
+    // test again, comparing the frame to the new cluster.
+  }
+}
+
+bool Segment::CheckHeaderInfo() {
+  if (!header_written_) {
+    if (!WriteSegmentHeader())
+      return false;
+
+    if (!seek_head_.AddSeekEntry(libwebm::kMkvCluster, MaxOffset()))
+      return false;
+
+    if (output_cues_ && cues_track_ == 0) {
+      // Check for a video track
+      for (uint32_t i = 0; i < tracks_.track_entries_size(); ++i) {
+        const Track* const track = tracks_.GetTrackByIndex(i);
+        if (!track)
+          return false;
+
+        if (tracks_.TrackIsVideo(track->number())) {
+          cues_track_ = track->number();
+          break;
+        }
+      }
+
+      // Set first track found
+      if (cues_track_ == 0) {
+        const Track* const track = tracks_.GetTrackByIndex(0);
+        if (!track)
+          return false;
+
+        cues_track_ = track->number();
+      }
+    }
+  }
+  return true;
+}
+
+void Segment::UpdateDocTypeVersion() {
+  for (uint32_t index = 0; index < tracks_.track_entries_size(); ++index) {
+    const Track* track = tracks_.GetTrackByIndex(index);
+    if (track == NULL)
+      break;
+    if ((track->codec_delay() || track->seek_pre_roll()) &&
+        doc_type_version_ < 4) {
+      doc_type_version_ = 4;
+      break;
+    }
+  }
+}
+
+bool Segment::UpdateChunkName(const char* ext, char** name) const {
+  if (!name || !ext)
+    return false;
+
+  char ext_chk[64];
+#ifdef _MSC_VER
+  sprintf_s(ext_chk, sizeof(ext_chk), "_%06d.%s", chunk_count_, ext);
+#else
+  snprintf(ext_chk, sizeof(ext_chk), "_%06d.%s", chunk_count_, ext);
+#endif
+
+  const size_t length = strlen(chunking_base_name_) + strlen(ext_chk) + 1;
+  char* const str = new (std::nothrow) char[length];  // NOLINT
+  if (!str)
+    return false;
+
+#ifdef _MSC_VER
+  strcpy_s(str, length - strlen(ext_chk), chunking_base_name_);
+  strcat_s(str, length, ext_chk);
+#else
+  strcpy(str, chunking_base_name_);
+  strcat(str, ext_chk);
+#endif
+
+  delete[] * name;
+  *name = str;
+
+  return true;
+}
+
+int64_t Segment::MaxOffset() {
+  if (!writer_header_)
+    return -1;
+
+  int64_t offset = writer_header_->Position() - payload_pos_;
+
+  if (chunking_) {
+    for (int32_t i = 0; i < cluster_list_size_; ++i) {
+      Cluster* const cluster = cluster_list_[i];
+      offset += cluster->Size();
+    }
+
+    if (writer_cues_)
+      offset += writer_cues_->Position();
+  }
+
+  return offset;
+}
+
+bool Segment::QueueFrame(Frame* frame) {
+  const int32_t new_size = frames_size_ + 1;
+
+  if (new_size > frames_capacity_) {
+    // Add more frames.
+    const int32_t new_capacity = (!frames_capacity_) ? 2 : frames_capacity_ * 2;
+
+    if (new_capacity < 1)
+      return false;
+
+    Frame** const frames = new (std::nothrow) Frame*[new_capacity];  // NOLINT
+    if (!frames)
+      return false;
+
+    for (int32_t i = 0; i < frames_size_; ++i) {
+      frames[i] = frames_[i];
+    }
+
+    delete[] frames_;
+    frames_ = frames;
+    frames_capacity_ = new_capacity;
+  }
+
+  frames_[frames_size_++] = frame;
+
+  return true;
+}
+
+int Segment::WriteFramesAll() {
+  if (frames_ == NULL)
+    return 0;
+
+  if (cluster_list_size_ < 1)
+    return -1;
+
+  Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
+
+  if (!cluster)
+    return -1;
+
+  for (int32_t i = 0; i < frames_size_; ++i) {
+    Frame*& frame = frames_[i];
+    // TODO(jzern/vigneshv): using Segment::AddGenericFrame here would limit the
+    // places where |doc_type_version_| needs to be updated.
+    if (frame->discard_padding() != 0)
+      doc_type_version_ = 4;
+    if (!cluster->AddFrame(frame))
+      return -1;
+
+    if (new_cuepoint_ && cues_track_ == frame->track_number()) {
+      if (!AddCuePoint(frame->timestamp(), cues_track_))
+        return -1;
+    }
+
+    if (frame->timestamp() > last_timestamp_) {
+      last_timestamp_ = frame->timestamp();
+      last_track_timestamp_[frame->track_number() - 1] = frame->timestamp();
+    }
+
+    delete frame;
+    frame = NULL;
+  }
+
+  const int result = frames_size_;
+  frames_size_ = 0;
+
+  return result;
+}
+
+bool Segment::WriteFramesLessThan(uint64_t timestamp) {
+  // Check |cluster_list_size_| to see if this is the first cluster. If it is
+  // the first cluster the audio frames that are less than the first video
+  // timesatmp will be written in a later step.
+  if (frames_size_ > 0 && cluster_list_size_ > 0) {
+    if (!frames_)
+      return false;
+
+    Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
+    if (!cluster)
+      return false;
+
+    int32_t shift_left = 0;
+
+    // TODO(fgalligan): Change this to use the durations of frames instead of
+    // the next frame's start time if the duration is accurate.
+    for (int32_t i = 1; i < frames_size_; ++i) {
+      const Frame* const frame_curr = frames_[i];
+
+      if (frame_curr->timestamp() > timestamp)
+        break;
+
+      const Frame* const frame_prev = frames_[i - 1];
+      if (frame_prev->discard_padding() != 0)
+        doc_type_version_ = 4;
+      if (!cluster->AddFrame(frame_prev))
+        return false;
+
+      if (new_cuepoint_ && cues_track_ == frame_prev->track_number()) {
+        if (!AddCuePoint(frame_prev->timestamp(), cues_track_))
+          return false;
+      }
+
+      ++shift_left;
+      if (frame_prev->timestamp() > last_timestamp_) {
+        last_timestamp_ = frame_prev->timestamp();
+        last_track_timestamp_[frame_prev->track_number() - 1] =
+            frame_prev->timestamp();
+      }
+
+      delete frame_prev;
+    }
+
+    if (shift_left > 0) {
+      if (shift_left >= frames_size_)
+        return false;
+
+      const int32_t new_frames_size = frames_size_ - shift_left;
+      for (int32_t i = 0; i < new_frames_size; ++i) {
+        frames_[i] = frames_[i + shift_left];
+      }
+
+      frames_size_ = new_frames_size;
+    }
+  }
+
+  return true;
+}
+
+}  // namespace mkvmuxer
diff --git a/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxer.h b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxer.h
new file mode 100644
index 0000000..55ba071
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxer.h
@@ -0,0 +1,1695 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#ifndef MKVMUXER_MKVMUXER_H_
+#define MKVMUXER_MKVMUXER_H_
+
+#include <stdint.h>
+
+#include <cstddef>
+#include <list>
+#include <map>
+
+#include "common/webmids.h"
+#include "mkvmuxer/mkvmuxertypes.h"
+
+// For a description of the WebM elements see
+// http://www.webmproject.org/code/specs/container/.
+
+namespace mkvparser {
+class IMkvReader;
+}  // namespace mkvparser
+
+namespace mkvmuxer {
+
+class MkvWriter;
+class Segment;
+
+const uint64_t kMaxTrackNumber = 126;
+
+///////////////////////////////////////////////////////////////
+// Interface used by the mkvmuxer to write out the Mkv data.
+class IMkvWriter {
+ public:
+  // Writes out |len| bytes of |buf|. Returns 0 on success.
+  virtual int32 Write(const void* buf, uint32 len) = 0;
+
+  // Returns the offset of the output position from the beginning of the
+  // output.
+  virtual int64 Position() const = 0;
+
+  // Set the current File position. Returns 0 on success.
+  virtual int32 Position(int64 position) = 0;
+
+  // Returns true if the writer is seekable.
+  virtual bool Seekable() const = 0;
+
+  // Element start notification. Called whenever an element identifier is about
+  // to be written to the stream. |element_id| is the element identifier, and
+  // |position| is the location in the WebM stream where the first octet of the
+  // element identifier will be written.
+  // Note: the |MkvId| enumeration in webmids.hpp defines element values.
+  virtual void ElementStartNotify(uint64 element_id, int64 position) = 0;
+
+ protected:
+  IMkvWriter();
+  virtual ~IMkvWriter();
+
+ private:
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(IMkvWriter);
+};
+
+// Writes out the EBML header for a WebM file. This function must be called
+// before any other libwebm writing functions are called.
+bool WriteEbmlHeader(IMkvWriter* writer, uint64_t doc_type_version);
+
+// Deprecated. Writes out EBML header with doc_type_version as
+// kDefaultDocTypeVersion. Exists for backward compatibility.
+bool WriteEbmlHeader(IMkvWriter* writer);
+
+// Copies in Chunk from source to destination between the given byte positions
+bool ChunkedCopy(mkvparser::IMkvReader* source, IMkvWriter* dst, int64_t start,
+                 int64_t size);
+
+///////////////////////////////////////////////////////////////
+// Class to hold data the will be written to a block.
+class Frame {
+ public:
+  Frame();
+  ~Frame();
+
+  // Sets this frame's contents based on |frame|. Returns true on success. On
+  // failure, this frame's existing contents may be lost.
+  bool CopyFrom(const Frame& frame);
+
+  // Copies |frame| data into |frame_|. Returns true on success.
+  bool Init(const uint8_t* frame, uint64_t length);
+
+  // Copies |additional| data into |additional_|. Returns true on success.
+  bool AddAdditionalData(const uint8_t* additional, uint64_t length,
+                         uint64_t add_id);
+
+  // Returns true if the frame has valid parameters.
+  bool IsValid() const;
+
+  // Returns true if the frame can be written as a SimpleBlock based on current
+  // parameters.
+  bool CanBeSimpleBlock() const;
+
+  uint64_t add_id() const { return add_id_; }
+  const uint8_t* additional() const { return additional_; }
+  uint64_t additional_length() const { return additional_length_; }
+  void set_duration(uint64_t duration);
+  uint64_t duration() const { return duration_; }
+  bool duration_set() const { return duration_set_; }
+  const uint8_t* frame() const { return frame_; }
+  void set_is_key(bool key) { is_key_ = key; }
+  bool is_key() const { return is_key_; }
+  uint64_t length() const { return length_; }
+  void set_track_number(uint64_t track_number) { track_number_ = track_number; }
+  uint64_t track_number() const { return track_number_; }
+  void set_timestamp(uint64_t timestamp) { timestamp_ = timestamp; }
+  uint64_t timestamp() const { return timestamp_; }
+  void set_discard_padding(int64_t discard_padding) {
+    discard_padding_ = discard_padding;
+  }
+  int64_t discard_padding() const { return discard_padding_; }
+  void set_reference_block_timestamp(int64_t reference_block_timestamp);
+  int64_t reference_block_timestamp() const {
+    return reference_block_timestamp_;
+  }
+  bool reference_block_timestamp_set() const {
+    return reference_block_timestamp_set_;
+  }
+
+ private:
+  // Id of the Additional data.
+  uint64_t add_id_;
+
+  // Pointer to additional data. Owned by this class.
+  uint8_t* additional_;
+
+  // Length of the additional data.
+  uint64_t additional_length_;
+
+  // Duration of the frame in nanoseconds.
+  uint64_t duration_;
+
+  // Flag indicating that |duration_| has been set. Setting duration causes the
+  // frame to be written out as a Block with BlockDuration instead of as a
+  // SimpleBlock.
+  bool duration_set_;
+
+  // Pointer to the data. Owned by this class.
+  uint8_t* frame_;
+
+  // Flag telling if the data should set the key flag of a block.
+  bool is_key_;
+
+  // Length of the data.
+  uint64_t length_;
+
+  // Mkv track number the data is associated with.
+  uint64_t track_number_;
+
+  // Timestamp of the data in nanoseconds.
+  uint64_t timestamp_;
+
+  // Discard padding for the frame.
+  int64_t discard_padding_;
+
+  // Reference block timestamp.
+  int64_t reference_block_timestamp_;
+
+  // Flag indicating if |reference_block_timestamp_| has been set.
+  bool reference_block_timestamp_set_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Frame);
+};
+
+///////////////////////////////////////////////////////////////
+// Class to hold one cue point in a Cues element.
+class CuePoint {
+ public:
+  CuePoint();
+  ~CuePoint();
+
+  // Returns the size in bytes for the entire CuePoint element.
+  uint64_t Size() const;
+
+  // Output the CuePoint element to the writer. Returns true on success.
+  bool Write(IMkvWriter* writer) const;
+
+  void set_time(uint64_t time) { time_ = time; }
+  uint64_t time() const { return time_; }
+  void set_track(uint64_t track) { track_ = track; }
+  uint64_t track() const { return track_; }
+  void set_cluster_pos(uint64_t cluster_pos) { cluster_pos_ = cluster_pos; }
+  uint64_t cluster_pos() const { return cluster_pos_; }
+  void set_block_number(uint64_t block_number) { block_number_ = block_number; }
+  uint64_t block_number() const { return block_number_; }
+  void set_output_block_number(bool output_block_number) {
+    output_block_number_ = output_block_number;
+  }
+  bool output_block_number() const { return output_block_number_; }
+
+ private:
+  // Returns the size in bytes for the payload of the CuePoint element.
+  uint64_t PayloadSize() const;
+
+  // Absolute timecode according to the segment time base.
+  uint64_t time_;
+
+  // The Track element associated with the CuePoint.
+  uint64_t track_;
+
+  // The position of the Cluster containing the Block.
+  uint64_t cluster_pos_;
+
+  // Number of the Block within the Cluster, starting from 1.
+  uint64_t block_number_;
+
+  // If true the muxer will write out the block number for the cue if the
+  // block number is different than the default of 1. Default is set to true.
+  bool output_block_number_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(CuePoint);
+};
+
+///////////////////////////////////////////////////////////////
+// Cues element.
+class Cues {
+ public:
+  Cues();
+  ~Cues();
+
+  // Adds a cue point to the Cues element. Returns true on success.
+  bool AddCue(CuePoint* cue);
+
+  // Returns the cue point by index. Returns NULL if there is no cue point
+  // match.
+  CuePoint* GetCueByIndex(int32_t index) const;
+
+  // Returns the total size of the Cues element
+  uint64_t Size();
+
+  // Output the Cues element to the writer. Returns true on success.
+  bool Write(IMkvWriter* writer) const;
+
+  int32_t cue_entries_size() const { return cue_entries_size_; }
+  void set_output_block_number(bool output_block_number) {
+    output_block_number_ = output_block_number;
+  }
+  bool output_block_number() const { return output_block_number_; }
+
+ private:
+  // Number of allocated elements in |cue_entries_|.
+  int32_t cue_entries_capacity_;
+
+  // Number of CuePoints in |cue_entries_|.
+  int32_t cue_entries_size_;
+
+  // CuePoint list.
+  CuePoint** cue_entries_;
+
+  // If true the muxer will write out the block number for the cue if the
+  // block number is different than the default of 1. Default is set to true.
+  bool output_block_number_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Cues);
+};
+
+///////////////////////////////////////////////////////////////
+// ContentEncAESSettings element
+class ContentEncAESSettings {
+ public:
+  enum { kCTR = 1 };
+
+  ContentEncAESSettings();
+  ~ContentEncAESSettings() {}
+
+  // Returns the size in bytes for the ContentEncAESSettings element.
+  uint64_t Size() const;
+
+  // Writes out the ContentEncAESSettings element to |writer|. Returns true on
+  // success.
+  bool Write(IMkvWriter* writer) const;
+
+  uint64_t cipher_mode() const { return cipher_mode_; }
+
+ private:
+  // Returns the size in bytes for the payload of the ContentEncAESSettings
+  // element.
+  uint64_t PayloadSize() const;
+
+  // Sub elements
+  uint64_t cipher_mode_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(ContentEncAESSettings);
+};
+
+///////////////////////////////////////////////////////////////
+// ContentEncoding element
+// Elements used to describe if the track data has been encrypted or
+// compressed with zlib or header stripping.
+// Currently only whole frames can be encrypted with AES. This dictates that
+// ContentEncodingOrder will be 0, ContentEncodingScope will be 1,
+// ContentEncodingType will be 1, and ContentEncAlgo will be 5.
+class ContentEncoding {
+ public:
+  ContentEncoding();
+  ~ContentEncoding();
+
+  // Sets the content encryption id. Copies |length| bytes from |id| to
+  // |enc_key_id_|. Returns true on success.
+  bool SetEncryptionID(const uint8_t* id, uint64_t length);
+
+  // Returns the size in bytes for the ContentEncoding element.
+  uint64_t Size() const;
+
+  // Writes out the ContentEncoding element to |writer|. Returns true on
+  // success.
+  bool Write(IMkvWriter* writer) const;
+
+  uint64_t enc_algo() const { return enc_algo_; }
+  uint64_t encoding_order() const { return encoding_order_; }
+  uint64_t encoding_scope() const { return encoding_scope_; }
+  uint64_t encoding_type() const { return encoding_type_; }
+  ContentEncAESSettings* enc_aes_settings() { return &enc_aes_settings_; }
+
+ private:
+  // Returns the size in bytes for the encoding elements.
+  uint64_t EncodingSize(uint64_t compresion_size,
+                        uint64_t encryption_size) const;
+
+  // Returns the size in bytes for the encryption elements.
+  uint64_t EncryptionSize() const;
+
+  // Track element names
+  uint64_t enc_algo_;
+  uint8_t* enc_key_id_;
+  uint64_t encoding_order_;
+  uint64_t encoding_scope_;
+  uint64_t encoding_type_;
+
+  // ContentEncAESSettings element.
+  ContentEncAESSettings enc_aes_settings_;
+
+  // Size of the ContentEncKeyID data in bytes.
+  uint64_t enc_key_id_length_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(ContentEncoding);
+};
+
+///////////////////////////////////////////////////////////////
+// Colour element.
+struct PrimaryChromaticity {
+  PrimaryChromaticity(float x_val, float y_val) : x(x_val), y(y_val) {}
+  PrimaryChromaticity() : x(0), y(0) {}
+  ~PrimaryChromaticity() {}
+  uint64_t PrimaryChromaticityPayloadSize(libwebm::MkvId x_id,
+                                          libwebm::MkvId y_id) const;
+  bool Write(IMkvWriter* writer, libwebm::MkvId x_id,
+             libwebm::MkvId y_id) const;
+
+  float x;
+  float y;
+};
+
+class MasteringMetadata {
+ public:
+  static const float kValueNotPresent;
+
+  MasteringMetadata()
+      : luminance_max(kValueNotPresent),
+        luminance_min(kValueNotPresent),
+        r_(NULL),
+        g_(NULL),
+        b_(NULL),
+        white_point_(NULL) {}
+  ~MasteringMetadata() {
+    delete r_;
+    delete g_;
+    delete b_;
+    delete white_point_;
+  }
+
+  // Returns total size of the MasteringMetadata element.
+  uint64_t MasteringMetadataSize() const;
+  bool Write(IMkvWriter* writer) const;
+
+  // Copies non-null chromaticity.
+  bool SetChromaticity(const PrimaryChromaticity* r,
+                       const PrimaryChromaticity* g,
+                       const PrimaryChromaticity* b,
+                       const PrimaryChromaticity* white_point);
+  const PrimaryChromaticity* r() const { return r_; }
+  const PrimaryChromaticity* g() const { return g_; }
+  const PrimaryChromaticity* b() const { return b_; }
+  const PrimaryChromaticity* white_point() const { return white_point_; }
+
+  float luminance_max;
+  float luminance_min;
+
+ private:
+  // Returns size of MasteringMetadata child elements.
+  uint64_t PayloadSize() const;
+
+  PrimaryChromaticity* r_;
+  PrimaryChromaticity* g_;
+  PrimaryChromaticity* b_;
+  PrimaryChromaticity* white_point_;
+};
+
+class Colour {
+ public:
+  static const uint64_t kValueNotPresent;
+  Colour()
+      : matrix_coefficients(kValueNotPresent),
+        bits_per_channel(kValueNotPresent),
+        chroma_subsampling_horz(kValueNotPresent),
+        chroma_subsampling_vert(kValueNotPresent),
+        cb_subsampling_horz(kValueNotPresent),
+        cb_subsampling_vert(kValueNotPresent),
+        chroma_siting_horz(kValueNotPresent),
+        chroma_siting_vert(kValueNotPresent),
+        range(kValueNotPresent),
+        transfer_characteristics(kValueNotPresent),
+        primaries(kValueNotPresent),
+        max_cll(kValueNotPresent),
+        max_fall(kValueNotPresent),
+        mastering_metadata_(NULL) {}
+  ~Colour() { delete mastering_metadata_; }
+
+  // Returns total size of the Colour element.
+  uint64_t ColourSize() const;
+  bool Write(IMkvWriter* writer) const;
+
+  // Deep copies |mastering_metadata|.
+  bool SetMasteringMetadata(const MasteringMetadata& mastering_metadata);
+
+  const MasteringMetadata* mastering_metadata() const {
+    return mastering_metadata_;
+  }
+
+  uint64_t matrix_coefficients;
+  uint64_t bits_per_channel;
+  uint64_t chroma_subsampling_horz;
+  uint64_t chroma_subsampling_vert;
+  uint64_t cb_subsampling_horz;
+  uint64_t cb_subsampling_vert;
+  uint64_t chroma_siting_horz;
+  uint64_t chroma_siting_vert;
+  uint64_t range;
+  uint64_t transfer_characteristics;
+  uint64_t primaries;
+  uint64_t max_cll;
+  uint64_t max_fall;
+
+ private:
+  // Returns size of Colour child elements.
+  uint64_t PayloadSize() const;
+
+  MasteringMetadata* mastering_metadata_;
+};
+
+///////////////////////////////////////////////////////////////
+// Track element.
+class Track {
+ public:
+  // The |seed| parameter is used to synthesize a UID for the track.
+  explicit Track(unsigned int* seed);
+  virtual ~Track();
+
+  // Adds a ContentEncoding element to the Track. Returns true on success.
+  virtual bool AddContentEncoding();
+
+  // Returns the ContentEncoding by index. Returns NULL if there is no
+  // ContentEncoding match.
+  ContentEncoding* GetContentEncodingByIndex(uint32_t index) const;
+
+  // Returns the size in bytes for the payload of the Track element.
+  virtual uint64_t PayloadSize() const;
+
+  // Returns the size in bytes of the Track element.
+  virtual uint64_t Size() const;
+
+  // Output the Track element to the writer. Returns true on success.
+  virtual bool Write(IMkvWriter* writer) const;
+
+  // Sets the CodecPrivate element of the Track element. Copies |length|
+  // bytes from |codec_private| to |codec_private_|. Returns true on success.
+  bool SetCodecPrivate(const uint8_t* codec_private, uint64_t length);
+
+  void set_codec_id(const char* codec_id);
+  const char* codec_id() const { return codec_id_; }
+  const uint8_t* codec_private() const { return codec_private_; }
+  void set_language(const char* language);
+  const char* language() const { return language_; }
+  void set_max_block_additional_id(uint64_t max_block_additional_id) {
+    max_block_additional_id_ = max_block_additional_id;
+  }
+  uint64_t max_block_additional_id() const { return max_block_additional_id_; }
+  void set_name(const char* name);
+  const char* name() const { return name_; }
+  void set_number(uint64_t number) { number_ = number; }
+  uint64_t number() const { return number_; }
+  void set_type(uint64_t type) { type_ = type; }
+  uint64_t type() const { return type_; }
+  void set_uid(uint64_t uid) { uid_ = uid; }
+  uint64_t uid() const { return uid_; }
+  void set_codec_delay(uint64_t codec_delay) { codec_delay_ = codec_delay; }
+  uint64_t codec_delay() const { return codec_delay_; }
+  void set_seek_pre_roll(uint64_t seek_pre_roll) {
+    seek_pre_roll_ = seek_pre_roll;
+  }
+  uint64_t seek_pre_roll() const { return seek_pre_roll_; }
+  void set_default_duration(uint64_t default_duration) {
+    default_duration_ = default_duration;
+  }
+  uint64_t default_duration() const { return default_duration_; }
+
+  uint64_t codec_private_length() const { return codec_private_length_; }
+  uint32_t content_encoding_entries_size() const {
+    return content_encoding_entries_size_;
+  }
+
+ private:
+  // Track element names.
+  char* codec_id_;
+  uint8_t* codec_private_;
+  char* language_;
+  uint64_t max_block_additional_id_;
+  char* name_;
+  uint64_t number_;
+  uint64_t type_;
+  uint64_t uid_;
+  uint64_t codec_delay_;
+  uint64_t seek_pre_roll_;
+  uint64_t default_duration_;
+
+  // Size of the CodecPrivate data in bytes.
+  uint64_t codec_private_length_;
+
+  // ContentEncoding element list.
+  ContentEncoding** content_encoding_entries_;
+
+  // Number of ContentEncoding elements added.
+  uint32_t content_encoding_entries_size_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Track);
+};
+
+///////////////////////////////////////////////////////////////
+// Track that has video specific elements.
+class VideoTrack : public Track {
+ public:
+  // Supported modes for stereo 3D.
+  enum StereoMode {
+    kMono = 0,
+    kSideBySideLeftIsFirst = 1,
+    kTopBottomRightIsFirst = 2,
+    kTopBottomLeftIsFirst = 3,
+    kSideBySideRightIsFirst = 11
+  };
+
+  enum AlphaMode { kNoAlpha = 0, kAlpha = 1 };
+
+  // The |seed| parameter is used to synthesize a UID for the track.
+  explicit VideoTrack(unsigned int* seed);
+  virtual ~VideoTrack();
+
+  // Returns the size in bytes for the payload of the Track element plus the
+  // video specific elements.
+  virtual uint64_t PayloadSize() const;
+
+  // Output the VideoTrack element to the writer. Returns true on success.
+  virtual bool Write(IMkvWriter* writer) const;
+
+  // Sets the video's stereo mode. Returns true on success.
+  bool SetStereoMode(uint64_t stereo_mode);
+
+  // Sets the video's alpha mode. Returns true on success.
+  bool SetAlphaMode(uint64_t alpha_mode);
+
+  void set_display_height(uint64_t height) { display_height_ = height; }
+  uint64_t display_height() const { return display_height_; }
+  void set_display_width(uint64_t width) { display_width_ = width; }
+  uint64_t display_width() const { return display_width_; }
+
+  void set_crop_left(uint64_t crop_left) { crop_left_ = crop_left; }
+  uint64_t crop_left() const { return crop_left_; }
+  void set_crop_right(uint64_t crop_right) { crop_right_ = crop_right; }
+  uint64_t crop_right() const { return crop_right_; }
+  void set_crop_top(uint64_t crop_top) { crop_top_ = crop_top; }
+  uint64_t crop_top() const { return crop_top_; }
+  void set_crop_bottom(uint64_t crop_bottom) { crop_bottom_ = crop_bottom; }
+  uint64_t crop_bottom() const { return crop_bottom_; }
+
+  void set_frame_rate(double frame_rate) { frame_rate_ = frame_rate; }
+  double frame_rate() const { return frame_rate_; }
+  void set_height(uint64_t height) { height_ = height; }
+  uint64_t height() const { return height_; }
+  uint64_t stereo_mode() { return stereo_mode_; }
+  uint64_t alpha_mode() { return alpha_mode_; }
+  void set_width(uint64_t width) { width_ = width; }
+  uint64_t width() const { return width_; }
+
+  Colour* colour() { return colour_; }
+
+  // Deep copies |colour|.
+  bool SetColour(const Colour& colour);
+
+ private:
+  // Returns the size in bytes of the Video element.
+  uint64_t VideoPayloadSize() const;
+
+  // Video track element names.
+  uint64_t display_height_;
+  uint64_t display_width_;
+  uint64_t crop_left_;
+  uint64_t crop_right_;
+  uint64_t crop_top_;
+  uint64_t crop_bottom_;
+  double frame_rate_;
+  uint64_t height_;
+  uint64_t stereo_mode_;
+  uint64_t alpha_mode_;
+  uint64_t width_;
+
+  Colour* colour_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(VideoTrack);
+};
+
+///////////////////////////////////////////////////////////////
+// Track that has audio specific elements.
+class AudioTrack : public Track {
+ public:
+  // The |seed| parameter is used to synthesize a UID for the track.
+  explicit AudioTrack(unsigned int* seed);
+  virtual ~AudioTrack();
+
+  // Returns the size in bytes for the payload of the Track element plus the
+  // audio specific elements.
+  virtual uint64_t PayloadSize() const;
+
+  // Output the AudioTrack element to the writer. Returns true on success.
+  virtual bool Write(IMkvWriter* writer) const;
+
+  void set_bit_depth(uint64_t bit_depth) { bit_depth_ = bit_depth; }
+  uint64_t bit_depth() const { return bit_depth_; }
+  void set_channels(uint64_t channels) { channels_ = channels; }
+  uint64_t channels() const { return channels_; }
+  void set_sample_rate(double sample_rate) { sample_rate_ = sample_rate; }
+  double sample_rate() const { return sample_rate_; }
+
+ private:
+  // Audio track element names.
+  uint64_t bit_depth_;
+  uint64_t channels_;
+  double sample_rate_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(AudioTrack);
+};
+
+///////////////////////////////////////////////////////////////
+// Tracks element
+class Tracks {
+ public:
+  // Audio and video type defined by the Matroska specs.
+  enum { kVideo = 0x1, kAudio = 0x2 };
+
+  static const char kOpusCodecId[];
+  static const char kVorbisCodecId[];
+  static const char kVp8CodecId[];
+  static const char kVp9CodecId[];
+  static const char kVp10CodecId[];
+
+  Tracks();
+  ~Tracks();
+
+  // Adds a Track element to the Tracks object. |track| will be owned and
+  // deleted by the Tracks object. Returns true on success. |number| is the
+  // number to use for the track. |number| must be >= 0. If |number| == 0
+  // then the muxer will decide on the track number.
+  bool AddTrack(Track* track, int32_t number);
+
+  // Returns the track by index. Returns NULL if there is no track match.
+  const Track* GetTrackByIndex(uint32_t idx) const;
+
+  // Search the Tracks and return the track that matches |tn|. Returns NULL
+  // if there is no track match.
+  Track* GetTrackByNumber(uint64_t track_number) const;
+
+  // Returns true if the track number is an audio track.
+  bool TrackIsAudio(uint64_t track_number) const;
+
+  // Returns true if the track number is a video track.
+  bool TrackIsVideo(uint64_t track_number) const;
+
+  // Output the Tracks element to the writer. Returns true on success.
+  bool Write(IMkvWriter* writer) const;
+
+  uint32_t track_entries_size() const { return track_entries_size_; }
+
+ private:
+  // Track element list.
+  Track** track_entries_;
+
+  // Number of Track elements added.
+  uint32_t track_entries_size_;
+
+  // Whether or not Tracks element has already been written via IMkvWriter.
+  mutable bool wrote_tracks_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Tracks);
+};
+
+///////////////////////////////////////////////////////////////
+// Chapter element
+//
+class Chapter {
+ public:
+  // Set the identifier for this chapter.  (This corresponds to the
+  // Cue Identifier line in WebVTT.)
+  // TODO(matthewjheaney): the actual serialization of this item in
+  // MKV is pending.
+  bool set_id(const char* id);
+
+  // Converts the nanosecond start and stop times of this chapter to
+  // their corresponding timecode values, and stores them that way.
+  void set_time(const Segment& segment, uint64_t start_time_ns,
+                uint64_t end_time_ns);
+
+  // Sets the uid for this chapter. Primarily used to enable
+  // deterministic output from the muxer.
+  void set_uid(const uint64_t uid) { uid_ = uid; }
+
+  // Add a title string to this chapter, per the semantics described
+  // here:
+  //  http://www.matroska.org/technical/specs/index.html
+  //
+  // The title ("chapter string") is a UTF-8 string.
+  //
+  // The language has ISO 639-2 representation, described here:
+  //  http://www.loc.gov/standards/iso639-2/englangn.html
+  //  http://www.loc.gov/standards/iso639-2/php/English_list.php
+  // If you specify NULL as the language value, this implies
+  // English ("eng").
+  //
+  // The country value corresponds to the codes listed here:
+  //  http://www.iana.org/domains/root/db/
+  //
+  // The function returns false if the string could not be allocated.
+  bool add_string(const char* title, const char* language, const char* country);
+
+ private:
+  friend class Chapters;
+
+  // For storage of chapter titles that differ by language.
+  class Display {
+   public:
+    // Establish representation invariant for new Display object.
+    void Init();
+
+    // Reclaim resources, in anticipation of destruction.
+    void Clear();
+
+    // Copies the title to the |title_| member.  Returns false on
+    // error.
+    bool set_title(const char* title);
+
+    // Copies the language to the |language_| member.  Returns false
+    // on error.
+    bool set_language(const char* language);
+
+    // Copies the country to the |country_| member.  Returns false on
+    // error.
+    bool set_country(const char* country);
+
+    // If |writer| is non-NULL, serialize the Display sub-element of
+    // the Atom into the stream.  Returns the Display element size on
+    // success, 0 if error.
+    uint64_t WriteDisplay(IMkvWriter* writer) const;
+
+   private:
+    char* title_;
+    char* language_;
+    char* country_;
+  };
+
+  Chapter();
+  ~Chapter();
+
+  // Establish the representation invariant for a newly-created
+  // Chapter object.  The |seed| parameter is used to create the UID
+  // for this chapter atom.
+  void Init(unsigned int* seed);
+
+  // Copies this Chapter object to a different one.  This is used when
+  // expanding a plain array of Chapter objects (see Chapters).
+  void ShallowCopy(Chapter* dst) const;
+
+  // Reclaim resources used by this Chapter object, pending its
+  // destruction.
+  void Clear();
+
+  // If there is no storage remaining on the |displays_| array for a
+  // new display object, creates a new, longer array and copies the
+  // existing Display objects to the new array.  Returns false if the
+  // array cannot be expanded.
+  bool ExpandDisplaysArray();
+
+  // If |writer| is non-NULL, serialize the Atom sub-element into the
+  // stream.  Returns the total size of the element on success, 0 if
+  // error.
+  uint64_t WriteAtom(IMkvWriter* writer) const;
+
+  // The string identifier for this chapter (corresponds to WebVTT cue
+  // identifier).
+  char* id_;
+
+  // Start timecode of the chapter.
+  uint64_t start_timecode_;
+
+  // Stop timecode of the chapter.
+  uint64_t end_timecode_;
+
+  // The binary identifier for this chapter.
+  uint64_t uid_;
+
+  // The Atom element can contain multiple Display sub-elements, as
+  // the same logical title can be rendered in different languages.
+  Display* displays_;
+
+  // The physical length (total size) of the |displays_| array.
+  int displays_size_;
+
+  // The logical length (number of active elements) on the |displays_|
+  // array.
+  int displays_count_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Chapter);
+};
+
+///////////////////////////////////////////////////////////////
+// Chapters element
+//
+class Chapters {
+ public:
+  Chapters();
+  ~Chapters();
+
+  Chapter* AddChapter(unsigned int* seed);
+
+  // Returns the number of chapters that have been added.
+  int Count() const;
+
+  // Output the Chapters element to the writer. Returns true on success.
+  bool Write(IMkvWriter* writer) const;
+
+ private:
+  // Expands the chapters_ array if there is not enough space to contain
+  // another chapter object.  Returns true on success.
+  bool ExpandChaptersArray();
+
+  // If |writer| is non-NULL, serialize the Edition sub-element of the
+  // Chapters element into the stream.  Returns the Edition element
+  // size on success, 0 if error.
+  uint64_t WriteEdition(IMkvWriter* writer) const;
+
+  // Total length of the chapters_ array.
+  int chapters_size_;
+
+  // Number of active chapters on the chapters_ array.
+  int chapters_count_;
+
+  // Array for storage of chapter objects.
+  Chapter* chapters_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Chapters);
+};
+
+///////////////////////////////////////////////////////////////
+// Tag element
+//
+class Tag {
+ public:
+  bool add_simple_tag(const char* tag_name, const char* tag_string);
+
+ private:
+  // Tags calls Clear and the destructor of Tag
+  friend class Tags;
+
+  // For storage of simple tags
+  class SimpleTag {
+   public:
+    // Establish representation invariant for new SimpleTag object.
+    void Init();
+
+    // Reclaim resources, in anticipation of destruction.
+    void Clear();
+
+    // Copies the title to the |tag_name_| member.  Returns false on
+    // error.
+    bool set_tag_name(const char* tag_name);
+
+    // Copies the language to the |tag_string_| member.  Returns false
+    // on error.
+    bool set_tag_string(const char* tag_string);
+
+    // If |writer| is non-NULL, serialize the SimpleTag sub-element of
+    // the Atom into the stream.  Returns the SimpleTag element size on
+    // success, 0 if error.
+    uint64_t Write(IMkvWriter* writer) const;
+
+   private:
+    char* tag_name_;
+    char* tag_string_;
+  };
+
+  Tag();
+  ~Tag();
+
+  // Copies this Tag object to a different one.  This is used when
+  // expanding a plain array of Tag objects (see Tags).
+  void ShallowCopy(Tag* dst) const;
+
+  // Reclaim resources used by this Tag object, pending its
+  // destruction.
+  void Clear();
+
+  // If there is no storage remaining on the |simple_tags_| array for a
+  // new display object, creates a new, longer array and copies the
+  // existing SimpleTag objects to the new array.  Returns false if the
+  // array cannot be expanded.
+  bool ExpandSimpleTagsArray();
+
+  // If |writer| is non-NULL, serialize the Tag sub-element into the
+  // stream.  Returns the total size of the element on success, 0 if
+  // error.
+  uint64_t Write(IMkvWriter* writer) const;
+
+  // The Atom element can contain multiple SimpleTag sub-elements
+  SimpleTag* simple_tags_;
+
+  // The physical length (total size) of the |simple_tags_| array.
+  int simple_tags_size_;
+
+  // The logical length (number of active elements) on the |simple_tags_|
+  // array.
+  int simple_tags_count_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Tag);
+};
+
+///////////////////////////////////////////////////////////////
+// Tags element
+//
+class Tags {
+ public:
+  Tags();
+  ~Tags();
+
+  Tag* AddTag();
+
+  // Returns the number of tags that have been added.
+  int Count() const;
+
+  // Output the Tags element to the writer. Returns true on success.
+  bool Write(IMkvWriter* writer) const;
+
+ private:
+  // Expands the tags_ array if there is not enough space to contain
+  // another tag object.  Returns true on success.
+  bool ExpandTagsArray();
+
+  // Total length of the tags_ array.
+  int tags_size_;
+
+  // Number of active tags on the tags_ array.
+  int tags_count_;
+
+  // Array for storage of tag objects.
+  Tag* tags_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Tags);
+};
+
+///////////////////////////////////////////////////////////////
+// Cluster element
+//
+// Notes:
+//  |Init| must be called before any other method in this class.
+class Cluster {
+ public:
+  // |timecode| is the absolute timecode of the cluster. |cues_pos| is the
+  // position for the cluster within the segment that should be written in
+  // the cues element. |timecode_scale| is the timecode scale of the segment.
+  Cluster(uint64_t timecode, int64_t cues_pos, uint64_t timecode_scale,
+          bool write_last_frame_with_duration = false,
+          bool fixed_size_timecode = false);
+  ~Cluster();
+
+  bool Init(IMkvWriter* ptr_writer);
+
+  // Adds a frame to be output in the file. The frame is written out through
+  // |writer_| if successful. Returns true on success.
+  bool AddFrame(const Frame* frame);
+
+  // Adds a frame to be output in the file. The frame is written out through
+  // |writer_| if successful. Returns true on success.
+  // Inputs:
+  //   data: Pointer to the data
+  //   length: Length of the data
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.  The range of allowed values is [1, 126].
+  //   timecode:     Absolute (not relative to cluster) timestamp of the
+  //                 frame, expressed in timecode units.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrame(const uint8_t* data, uint64_t length, uint64_t track_number,
+                uint64_t timecode,  // timecode units (absolute)
+                bool is_key);
+
+  // Adds a frame to be output in the file. The frame is written out through
+  // |writer_| if successful. Returns true on success.
+  // Inputs:
+  //   data: Pointer to the data
+  //   length: Length of the data
+  //   additional: Pointer to the additional data
+  //   additional_length: Length of the additional data
+  //   add_id: Value of BlockAddID element
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.  The range of allowed values is [1, 126].
+  //   abs_timecode: Absolute (not relative to cluster) timestamp of the
+  //                 frame, expressed in timecode units.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrameWithAdditional(const uint8_t* data, uint64_t length,
+                              const uint8_t* additional,
+                              uint64_t additional_length, uint64_t add_id,
+                              uint64_t track_number, uint64_t abs_timecode,
+                              bool is_key);
+
+  // Adds a frame to be output in the file. The frame is written out through
+  // |writer_| if successful. Returns true on success.
+  // Inputs:
+  //   data: Pointer to the data.
+  //   length: Length of the data.
+  //   discard_padding: DiscardPadding element value.
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.  The range of allowed values is [1, 126].
+  //   abs_timecode: Absolute (not relative to cluster) timestamp of the
+  //                 frame, expressed in timecode units.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrameWithDiscardPadding(const uint8_t* data, uint64_t length,
+                                  int64_t discard_padding,
+                                  uint64_t track_number, uint64_t abs_timecode,
+                                  bool is_key);
+
+  // Writes a frame of metadata to the output medium; returns true on
+  // success.
+  // Inputs:
+  //   data: Pointer to the data
+  //   length: Length of the data
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.  The range of allowed values is [1, 126].
+  //   timecode:     Absolute (not relative to cluster) timestamp of the
+  //                 metadata frame, expressed in timecode units.
+  //   duration:     Duration of metadata frame, in timecode units.
+  //
+  // The metadata frame is written as a block group, with a duration
+  // sub-element but no reference time sub-elements (indicating that
+  // it is considered a keyframe, per Matroska semantics).
+  bool AddMetadata(const uint8_t* data, uint64_t length, uint64_t track_number,
+                   uint64_t timecode, uint64_t duration);
+
+  // Increments the size of the cluster's data in bytes.
+  void AddPayloadSize(uint64_t size);
+
+  // Closes the cluster so no more data can be written to it. Will update the
+  // cluster's size if |writer_| is seekable. Returns true on success. This
+  // variant of Finalize() fails when |write_last_frame_with_duration_| is set
+  // to true.
+  bool Finalize();
+
+  // Closes the cluster so no more data can be written to it. Will update the
+  // cluster's size if |writer_| is seekable. Returns true on success.
+  // Inputs:
+  //   set_last_frame_duration: Boolean indicating whether or not the duration
+  //                            of the last frame should be set. If set to
+  //                            false, the |duration| value is ignored and
+  //                            |write_last_frame_with_duration_| will not be
+  //                            honored.
+  //   duration: Duration of the Cluster in timecode scale.
+  bool Finalize(bool set_last_frame_duration, uint64_t duration);
+
+  // Returns the size in bytes for the entire Cluster element.
+  uint64_t Size() const;
+
+  // Given |abs_timecode|, calculates timecode relative to most recent timecode.
+  // Returns -1 on failure, or a relative timecode.
+  int64_t GetRelativeTimecode(int64_t abs_timecode) const;
+
+  int64_t size_position() const { return size_position_; }
+  int32_t blocks_added() const { return blocks_added_; }
+  uint64_t payload_size() const { return payload_size_; }
+  int64_t position_for_cues() const { return position_for_cues_; }
+  uint64_t timecode() const { return timecode_; }
+  uint64_t timecode_scale() const { return timecode_scale_; }
+  void set_write_last_frame_with_duration(bool write_last_frame_with_duration) {
+    write_last_frame_with_duration_ = write_last_frame_with_duration;
+  }
+  bool write_last_frame_with_duration() const {
+    return write_last_frame_with_duration_;
+  }
+
+ private:
+  // Iterator type for the |stored_frames_| map.
+  typedef std::map<uint64_t, std::list<Frame*> >::iterator FrameMapIterator;
+
+  // Utility method that confirms that blocks can still be added, and that the
+  // cluster header has been written. Used by |DoWriteFrame*|. Returns true
+  // when successful.
+  bool PreWriteBlock();
+
+  // Utility method used by the |DoWriteFrame*| methods that handles the book
+  // keeping required after each block is written.
+  void PostWriteBlock(uint64_t element_size);
+
+  // Does some verification and calls WriteFrame.
+  bool DoWriteFrame(const Frame* const frame);
+
+  // Either holds back the given frame, or writes it out depending on whether or
+  // not |write_last_frame_with_duration_| is set.
+  bool QueueOrWriteFrame(const Frame* const frame);
+
+  // Outputs the Cluster header to |writer_|. Returns true on success.
+  bool WriteClusterHeader();
+
+  // Number of blocks added to the cluster.
+  int32_t blocks_added_;
+
+  // Flag telling if the cluster has been closed.
+  bool finalized_;
+
+  // Flag indicating whether the cluster's timecode will always be written out
+  // using 8 bytes.
+  bool fixed_size_timecode_;
+
+  // Flag telling if the cluster's header has been written.
+  bool header_written_;
+
+  // The size of the cluster elements in bytes.
+  uint64_t payload_size_;
+
+  // The file position used for cue points.
+  const int64_t position_for_cues_;
+
+  // The file position of the cluster's size element.
+  int64_t size_position_;
+
+  // The absolute timecode of the cluster.
+  const uint64_t timecode_;
+
+  // The timecode scale of the Segment containing the cluster.
+  const uint64_t timecode_scale_;
+
+  // Flag indicating whether the last frame of the cluster should be written as
+  // a Block with Duration. If set to true, then it will result in holding back
+  // of frames and the parameterized version of Finalize() must be called to
+  // finish writing the Cluster.
+  bool write_last_frame_with_duration_;
+
+  // Map used to hold back frames, if required. Track number is the key.
+  std::map<uint64_t, std::list<Frame*> > stored_frames_;
+
+  // Map from track number to the timestamp of the last block written for that
+  // track.
+  std::map<uint64_t, uint64_t> last_block_timestamp_;
+
+  // Pointer to the writer object. Not owned by this class.
+  IMkvWriter* writer_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Cluster);
+};
+
+///////////////////////////////////////////////////////////////
+// SeekHead element
+class SeekHead {
+ public:
+  SeekHead();
+  ~SeekHead();
+
+  // TODO(fgalligan): Change this to reserve a certain size. Then check how
+  // big the seek entry to be added is as not every seek entry will be the
+  // maximum size it could be.
+  // Adds a seek entry to be written out when the element is finalized. |id|
+  // must be the coded mkv element id. |pos| is the file position of the
+  // element. Returns true on success.
+  bool AddSeekEntry(uint32_t id, uint64_t pos);
+
+  // Writes out SeekHead and SeekEntry elements. Returns true on success.
+  bool Finalize(IMkvWriter* writer) const;
+
+  // Returns the id of the Seek Entry at the given index. Returns -1 if index is
+  // out of range.
+  uint32_t GetId(int index) const;
+
+  // Returns the position of the Seek Entry at the given index. Returns -1 if
+  // index is out of range.
+  uint64_t GetPosition(int index) const;
+
+  // Sets the Seek Entry id and position at given index.
+  // Returns true on success.
+  bool SetSeekEntry(int index, uint32_t id, uint64_t position);
+
+  // Reserves space by writing out a Void element which will be updated with
+  // a SeekHead element later. Returns true on success.
+  bool Write(IMkvWriter* writer);
+
+  // We are going to put a cap on the number of Seek Entries.
+  const static int32_t kSeekEntryCount = 5;
+
+ private:
+  // Returns the maximum size in bytes of one seek entry.
+  uint64_t MaxEntrySize() const;
+
+  // Seek entry id element list.
+  uint32_t seek_entry_id_[kSeekEntryCount];
+
+  // Seek entry pos element list.
+  uint64_t seek_entry_pos_[kSeekEntryCount];
+
+  // The file position of SeekHead element.
+  int64_t start_pos_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(SeekHead);
+};
+
+///////////////////////////////////////////////////////////////
+// Segment Information element
+class SegmentInfo {
+ public:
+  SegmentInfo();
+  ~SegmentInfo();
+
+  // Will update the duration if |duration_| is > 0.0. Returns true on success.
+  bool Finalize(IMkvWriter* writer) const;
+
+  // Sets |muxing_app_| and |writing_app_|.
+  bool Init();
+
+  // Output the Segment Information element to the writer. Returns true on
+  // success.
+  bool Write(IMkvWriter* writer);
+
+  void set_duration(double duration) { duration_ = duration; }
+  double duration() const { return duration_; }
+  void set_muxing_app(const char* app);
+  const char* muxing_app() const { return muxing_app_; }
+  void set_timecode_scale(uint64_t scale) { timecode_scale_ = scale; }
+  uint64_t timecode_scale() const { return timecode_scale_; }
+  void set_writing_app(const char* app);
+  const char* writing_app() const { return writing_app_; }
+  void set_date_utc(int64_t date_utc) { date_utc_ = date_utc; }
+  int64_t date_utc() const { return date_utc_; }
+
+ private:
+  // Segment Information element names.
+  // Initially set to -1 to signify that a duration has not been set and should
+  // not be written out.
+  double duration_;
+  // Set to libwebm-%d.%d.%d.%d, major, minor, build, revision.
+  char* muxing_app_;
+  uint64_t timecode_scale_;
+  // Initially set to libwebm-%d.%d.%d.%d, major, minor, build, revision.
+  char* writing_app_;
+  // LLONG_MIN when DateUTC is not set.
+  int64_t date_utc_;
+
+  // The file position of the duration element.
+  int64_t duration_pos_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(SegmentInfo);
+};
+
+///////////////////////////////////////////////////////////////
+// This class represents the main segment in a WebM file. Currently only
+// supports one Segment element.
+//
+// Notes:
+//  |Init| must be called before any other method in this class.
+class Segment {
+ public:
+  enum Mode { kLive = 0x1, kFile = 0x2 };
+
+  enum CuesPosition {
+    kAfterClusters = 0x0,  // Position Cues after Clusters - Default
+    kBeforeClusters = 0x1  // Position Cues before Clusters
+  };
+
+  const static uint32_t kDefaultDocTypeVersion = 2;
+  const static uint64_t kDefaultMaxClusterDuration = 30000000000ULL;
+
+  Segment();
+  ~Segment();
+
+  // Initializes |SegmentInfo| and returns result. Always returns false when
+  // |ptr_writer| is NULL.
+  bool Init(IMkvWriter* ptr_writer);
+
+  // Adds a generic track to the segment.  Returns the newly-allocated
+  // track object (which is owned by the segment) on success, NULL on
+  // error. |number| is the number to use for the track.  |number|
+  // must be >= 0. If |number| == 0 then the muxer will decide on the
+  // track number.
+  Track* AddTrack(int32_t number);
+
+  // Adds a Vorbis audio track to the segment. Returns the number of the track
+  // on success, 0 on error. |number| is the number to use for the audio track.
+  // |number| must be >= 0. If |number| == 0 then the muxer will decide on
+  // the track number.
+  uint64_t AddAudioTrack(int32_t sample_rate, int32_t channels, int32_t number);
+
+  // Adds an empty chapter to the chapters of this segment.  Returns
+  // non-NULL on success.  After adding the chapter, the caller should
+  // populate its fields via the Chapter member functions.
+  Chapter* AddChapter();
+
+  // Adds an empty tag to the tags of this segment.  Returns
+  // non-NULL on success.  After adding the tag, the caller should
+  // populate its fields via the Tag member functions.
+  Tag* AddTag();
+
+  // Adds a cue point to the Cues element. |timestamp| is the time in
+  // nanoseconds of the cue's time. |track| is the Track of the Cue. This
+  // function must be called after AddFrame to calculate the correct
+  // BlockNumber for the CuePoint. Returns true on success.
+  bool AddCuePoint(uint64_t timestamp, uint64_t track);
+
+  // Adds a frame to be output in the file. Returns true on success.
+  // Inputs:
+  //   data: Pointer to the data
+  //   length: Length of the data
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.
+  //   timestamp:    Timestamp of the frame in nanoseconds from 0.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrame(const uint8_t* data, uint64_t length, uint64_t track_number,
+                uint64_t timestamp_ns, bool is_key);
+
+  // Writes a frame of metadata to the output medium; returns true on
+  // success.
+  // Inputs:
+  //   data: Pointer to the data
+  //   length: Length of the data
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.
+  //   timecode:     Absolute timestamp of the metadata frame, expressed
+  //                 in nanosecond units.
+  //   duration:     Duration of metadata frame, in nanosecond units.
+  //
+  // The metadata frame is written as a block group, with a duration
+  // sub-element but no reference time sub-elements (indicating that
+  // it is considered a keyframe, per Matroska semantics).
+  bool AddMetadata(const uint8_t* data, uint64_t length, uint64_t track_number,
+                   uint64_t timestamp_ns, uint64_t duration_ns);
+
+  // Writes a frame with additional data to the output medium; returns true on
+  // success.
+  // Inputs:
+  //   data: Pointer to the data.
+  //   length: Length of the data.
+  //   additional: Pointer to additional data.
+  //   additional_length: Length of additional data.
+  //   add_id: Additional ID which identifies the type of additional data.
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.
+  //   timestamp:    Absolute timestamp of the frame, expressed in nanosecond
+  //                 units.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrameWithAdditional(const uint8_t* data, uint64_t length,
+                              const uint8_t* additional,
+                              uint64_t additional_length, uint64_t add_id,
+                              uint64_t track_number, uint64_t timestamp,
+                              bool is_key);
+
+  // Writes a frame with DiscardPadding to the output medium; returns true on
+  // success.
+  // Inputs:
+  //   data: Pointer to the data.
+  //   length: Length of the data.
+  //   discard_padding: DiscardPadding element value.
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.
+  //   timestamp:    Absolute timestamp of the frame, expressed in nanosecond
+  //                 units.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrameWithDiscardPadding(const uint8_t* data, uint64_t length,
+                                  int64_t discard_padding,
+                                  uint64_t track_number, uint64_t timestamp,
+                                  bool is_key);
+
+  // Writes a Frame to the output medium. Chooses the correct way of writing
+  // the frame (Block vs SimpleBlock) based on the parameters passed.
+  // Inputs:
+  //   frame: frame object
+  bool AddGenericFrame(const Frame* frame);
+
+  // Adds a VP8 video track to the segment. Returns the number of the track on
+  // success, 0 on error. |number| is the number to use for the video track.
+  // |number| must be >= 0. If |number| == 0 then the muxer will decide on
+  // the track number.
+  uint64_t AddVideoTrack(int32_t width, int32_t height, int32_t number);
+
+  // This function must be called after Finalize() if you need a copy of the
+  // output with Cues written before the Clusters. It will return false if the
+  // writer is not seekable of if chunking is set to true.
+  // Input parameters:
+  // reader - an IMkvReader object created with the same underlying file of the
+  //          current writer object. Make sure to close the existing writer
+  //          object before creating this so that all the data is properly
+  //          flushed and available for reading.
+  // writer - an IMkvWriter object pointing to a *different* file than the one
+  //          pointed by the current writer object. This file will contain the
+  //          Cues element before the Clusters.
+  bool CopyAndMoveCuesBeforeClusters(mkvparser::IMkvReader* reader,
+                                     IMkvWriter* writer);
+
+  // Sets which track to use for the Cues element. Must have added the track
+  // before calling this function. Returns true on success. |track_number| is
+  // returned by the Add track functions.
+  bool CuesTrack(uint64_t track_number);
+
+  // This will force the muxer to create a new Cluster when the next frame is
+  // added.
+  void ForceNewClusterOnNextFrame();
+
+  // Writes out any frames that have not been written out. Finalizes the last
+  // cluster. May update the size and duration of the segment. May output the
+  // Cues element. May finalize the SeekHead element. Returns true on success.
+  bool Finalize();
+
+  // Returns the Cues object.
+  Cues* GetCues() { return &cues_; }
+
+  // Returns the Segment Information object.
+  const SegmentInfo* GetSegmentInfo() const { return &segment_info_; }
+  SegmentInfo* GetSegmentInfo() { return &segment_info_; }
+
+  // Search the Tracks and return the track that matches |track_number|.
+  // Returns NULL if there is no track match.
+  Track* GetTrackByNumber(uint64_t track_number) const;
+
+  // Toggles whether to output a cues element.
+  void OutputCues(bool output_cues);
+
+  // Toggles whether to write the last frame in each Cluster with Duration.
+  void AccurateClusterDuration(bool accurate_cluster_duration);
+
+  // Toggles whether to write the Cluster Timecode using exactly 8 bytes.
+  void UseFixedSizeClusterTimecode(bool fixed_size_cluster_timecode);
+
+  // Sets if the muxer will output files in chunks or not. |chunking| is a
+  // flag telling whether or not to turn on chunking. |filename| is the base
+  // filename for the chunk files. The header chunk file will be named
+  // |filename|.hdr and the data chunks will be named
+  // |filename|_XXXXXX.chk. Chunking implies that the muxer will be writing
+  // to files so the muxer will use the default MkvWriter class to control
+  // what data is written to what files. Returns true on success.
+  // TODO: Should we change the IMkvWriter Interface to add Open and Close?
+  // That will force the interface to be dependent on files.
+  bool SetChunking(bool chunking, const char* filename);
+
+  bool chunking() const { return chunking_; }
+  uint64_t cues_track() const { return cues_track_; }
+  void set_max_cluster_duration(uint64_t max_cluster_duration) {
+    max_cluster_duration_ = max_cluster_duration;
+  }
+  uint64_t max_cluster_duration() const { return max_cluster_duration_; }
+  void set_max_cluster_size(uint64_t max_cluster_size) {
+    max_cluster_size_ = max_cluster_size;
+  }
+  uint64_t max_cluster_size() const { return max_cluster_size_; }
+  void set_mode(Mode mode) { mode_ = mode; }
+  Mode mode() const { return mode_; }
+  CuesPosition cues_position() const { return cues_position_; }
+  bool output_cues() const { return output_cues_; }
+  const SegmentInfo* segment_info() const { return &segment_info_; }
+
+ private:
+  // Checks if header information has been output and initialized. If not it
+  // will output the Segment element and initialize the SeekHead elment and
+  // Cues elements.
+  bool CheckHeaderInfo();
+
+  // Sets |doc_type_version_| based on the current element requirements.
+  void UpdateDocTypeVersion();
+
+  // Sets |name| according to how many chunks have been written. |ext| is the
+  // file extension. |name| must be deleted by the calling app. Returns true
+  // on success.
+  bool UpdateChunkName(const char* ext, char** name) const;
+
+  // Returns the maximum offset within the segment's payload. When chunking
+  // this function is needed to determine offsets of elements within the
+  // chunked files. Returns -1 on error.
+  int64_t MaxOffset();
+
+  // Adds the frame to our frame array.
+  bool QueueFrame(Frame* frame);
+
+  // Output all frames that are queued. Returns -1 on error, otherwise
+  // it returns the number of frames written.
+  int WriteFramesAll();
+
+  // Output all frames that are queued that have an end time that is less
+  // then |timestamp|. Returns true on success and if there are no frames
+  // queued.
+  bool WriteFramesLessThan(uint64_t timestamp);
+
+  // Outputs the segment header, Segment Information element, SeekHead element,
+  // and Tracks element to |writer_|.
+  bool WriteSegmentHeader();
+
+  // Given a frame with the specified timestamp (nanosecond units) and
+  // keyframe status, determine whether a new cluster should be
+  // created, before writing enqueued frames and the frame itself. The
+  // function returns one of the following values:
+  //  -1 = error: an out-of-order frame was detected
+  //  0 = do not create a new cluster, and write frame to the existing cluster
+  //  1 = create a new cluster, and write frame to that new cluster
+  //  2 = create a new cluster, and re-run test
+  int TestFrame(uint64_t track_num, uint64_t timestamp_ns, bool key) const;
+
+  // Create a new cluster, using the earlier of the first enqueued
+  // frame, or the indicated time. Returns true on success.
+  bool MakeNewCluster(uint64_t timestamp_ns);
+
+  // Checks whether a new cluster needs to be created, and if so
+  // creates a new cluster. Returns false if creation of a new cluster
+  // was necessary but creation was not successful.
+  bool DoNewClusterProcessing(uint64_t track_num, uint64_t timestamp_ns,
+                              bool key);
+
+  // Adjusts Cue Point values (to place Cues before Clusters) so that they
+  // reflect the correct offsets.
+  void MoveCuesBeforeClusters();
+
+  // This function recursively computes the correct cluster offsets (this is
+  // done to move the Cues before Clusters). It recursively updates the change
+  // in size (which indicates a change in cluster offset) until no sizes change.
+  // Parameters:
+  // diff - indicates the difference in size of the Cues element that needs to
+  //        accounted for.
+  // index - index in the list of Cues which is currently being adjusted.
+  // cue_size - sum of size of all the CuePoint elements.
+  void MoveCuesBeforeClustersHelper(uint64_t diff, int index,
+                                    uint64_t* cue_size);
+
+  // Seeds the random number generator used to make UIDs.
+  unsigned int seed_;
+
+  // WebM elements
+  Cues cues_;
+  SeekHead seek_head_;
+  SegmentInfo segment_info_;
+  Tracks tracks_;
+  Chapters chapters_;
+  Tags tags_;
+
+  // Number of chunks written.
+  int chunk_count_;
+
+  // Current chunk filename.
+  char* chunk_name_;
+
+  // Default MkvWriter object created by this class used for writing clusters
+  // out in separate files.
+  MkvWriter* chunk_writer_cluster_;
+
+  // Default MkvWriter object created by this class used for writing Cues
+  // element out to a file.
+  MkvWriter* chunk_writer_cues_;
+
+  // Default MkvWriter object created by this class used for writing the
+  // Matroska header out to a file.
+  MkvWriter* chunk_writer_header_;
+
+  // Flag telling whether or not the muxer is chunking output to multiple
+  // files.
+  bool chunking_;
+
+  // Base filename for the chunked files.
+  char* chunking_base_name_;
+
+  // File position offset where the Clusters end.
+  int64_t cluster_end_offset_;
+
+  // List of clusters.
+  Cluster** cluster_list_;
+
+  // Number of cluster pointers allocated in the cluster list.
+  int32_t cluster_list_capacity_;
+
+  // Number of clusters in the cluster list.
+  int32_t cluster_list_size_;
+
+  // Indicates whether Cues should be written before or after Clusters
+  CuesPosition cues_position_;
+
+  // Track number that is associated with the cues element for this segment.
+  uint64_t cues_track_;
+
+  // Tells the muxer to force a new cluster on the next Block.
+  bool force_new_cluster_;
+
+  // List of stored audio frames. These variables are used to store frames so
+  // the muxer can follow the guideline "Audio blocks that contain the video
+  // key frame's timecode should be in the same cluster as the video key frame
+  // block."
+  Frame** frames_;
+
+  // Number of frame pointers allocated in the frame list.
+  int32_t frames_capacity_;
+
+  // Number of frames in the frame list.
+  int32_t frames_size_;
+
+  // Flag telling if a video track has been added to the segment.
+  bool has_video_;
+
+  // Flag telling if the segment's header has been written.
+  bool header_written_;
+
+  // Duration of the last block in nanoseconds.
+  uint64_t last_block_duration_;
+
+  // Last timestamp in nanoseconds added to a cluster.
+  uint64_t last_timestamp_;
+
+  // Last timestamp in nanoseconds by track number added to a cluster.
+  uint64_t last_track_timestamp_[kMaxTrackNumber];
+
+  // Maximum time in nanoseconds for a cluster duration. This variable is a
+  // guideline and some clusters may have a longer duration. Default is 30
+  // seconds.
+  uint64_t max_cluster_duration_;
+
+  // Maximum size in bytes for a cluster. This variable is a guideline and
+  // some clusters may have a larger size. Default is 0 which signifies that
+  // the muxer will decide the size.
+  uint64_t max_cluster_size_;
+
+  // The mode that segment is in. If set to |kLive| the writer must not
+  // seek backwards.
+  Mode mode_;
+
+  // Flag telling the muxer that a new cue point should be added.
+  bool new_cuepoint_;
+
+  // TODO(fgalligan): Should we add support for more than one Cues element?
+  // Flag whether or not the muxer should output a Cues element.
+  bool output_cues_;
+
+  // Flag whether or not the last frame in each Cluster will have a Duration
+  // element in it.
+  bool accurate_cluster_duration_;
+
+  // Flag whether or not to write the Cluster Timecode using exactly 8 bytes.
+  bool fixed_size_cluster_timecode_;
+
+  // The size of the EBML header, used to validate the header if
+  // WriteEbmlHeader() is called more than once.
+  int32_t ebml_header_size_;
+
+  // The file position of the segment's payload.
+  int64_t payload_pos_;
+
+  // The file position of the element's size.
+  int64_t size_position_;
+
+  // Current DocTypeVersion (|doc_type_version_|) and that written in
+  // WriteSegmentHeader().
+  // WriteEbmlHeader() will be called from Finalize() if |doc_type_version_|
+  // differs from |doc_type_version_written_|.
+  uint32_t doc_type_version_;
+  uint32_t doc_type_version_written_;
+
+  // Pointer to the writer objects. Not owned by this class.
+  IMkvWriter* writer_cluster_;
+  IMkvWriter* writer_cues_;
+  IMkvWriter* writer_header_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Segment);
+};
+
+}  // namespace mkvmuxer
+
+#endif  // MKVMUXER_MKVMUXER_H_
diff --git a/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxertypes.h b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxertypes.h
new file mode 100644
index 0000000..e5db121
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxertypes.h
@@ -0,0 +1,28 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#ifndef MKVMUXER_MKVMUXERTYPES_H_
+#define MKVMUXER_MKVMUXERTYPES_H_
+
+namespace mkvmuxer {
+typedef unsigned char uint8;
+typedef short int16;
+typedef int int32;
+typedef unsigned int uint32;
+typedef long long int64;
+typedef unsigned long long uint64;
+}  // namespace mkvmuxer
+
+// Copied from Chromium basictypes.h
+// A macro to disallow the copy constructor and operator= functions
+// This should be used in the private: declarations for a class
+#define LIBWEBM_DISALLOW_COPY_AND_ASSIGN(TypeName) \
+  TypeName(const TypeName&);                       \
+  void operator=(const TypeName&)
+
+#endif  // MKVMUXER_MKVMUXERTYPES_HPP_
diff --git a/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxerutil.cc b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxerutil.cc
new file mode 100644
index 0000000..3562b8a
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxerutil.cc
@@ -0,0 +1,650 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#include "mkvmuxer/mkvmuxerutil.h"
+
+#ifdef __ANDROID__
+#include <fcntl.h>
+#endif
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <ctime>
+#include <new>
+
+#include "common/webmids.h"
+#include "mkvmuxer/mkvmuxer.h"
+#include "mkvmuxer/mkvwriter.h"
+
+namespace mkvmuxer {
+
+namespace {
+
+// Date elements are always 8 octets in size.
+const int kDateElementSize = 8;
+
+uint64_t WriteBlock(IMkvWriter* writer, const Frame* const frame,
+                    int64_t timecode, uint64_t timecode_scale) {
+  uint64_t block_additional_elem_size = 0;
+  uint64_t block_addid_elem_size = 0;
+  uint64_t block_more_payload_size = 0;
+  uint64_t block_more_elem_size = 0;
+  uint64_t block_additions_payload_size = 0;
+  uint64_t block_additions_elem_size = 0;
+  if (frame->additional()) {
+    block_additional_elem_size =
+        EbmlElementSize(libwebm::kMkvBlockAdditional, frame->additional(),
+                        frame->additional_length());
+    block_addid_elem_size =
+        EbmlElementSize(libwebm::kMkvBlockAddID, frame->add_id());
+
+    block_more_payload_size =
+        block_addid_elem_size + block_additional_elem_size;
+    block_more_elem_size =
+        EbmlMasterElementSize(libwebm::kMkvBlockMore, block_more_payload_size) +
+        block_more_payload_size;
+    block_additions_payload_size = block_more_elem_size;
+    block_additions_elem_size =
+        EbmlMasterElementSize(libwebm::kMkvBlockAdditions,
+                              block_additions_payload_size) +
+        block_additions_payload_size;
+  }
+
+  uint64_t discard_padding_elem_size = 0;
+  if (frame->discard_padding() != 0) {
+    discard_padding_elem_size =
+        EbmlElementSize(libwebm::kMkvDiscardPadding, frame->discard_padding());
+  }
+
+  const uint64_t reference_block_timestamp =
+      frame->reference_block_timestamp() / timecode_scale;
+  uint64_t reference_block_elem_size = 0;
+  if (!frame->is_key()) {
+    reference_block_elem_size =
+        EbmlElementSize(libwebm::kMkvReferenceBlock, reference_block_timestamp);
+  }
+
+  const uint64_t duration = frame->duration() / timecode_scale;
+  uint64_t block_duration_elem_size = 0;
+  if (duration > 0)
+    block_duration_elem_size =
+        EbmlElementSize(libwebm::kMkvBlockDuration, duration);
+
+  const uint64_t block_payload_size = 4 + frame->length();
+  const uint64_t block_elem_size =
+      EbmlMasterElementSize(libwebm::kMkvBlock, block_payload_size) +
+      block_payload_size;
+
+  const uint64_t block_group_payload_size =
+      block_elem_size + block_additions_elem_size + block_duration_elem_size +
+      discard_padding_elem_size + reference_block_elem_size;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvBlockGroup,
+                              block_group_payload_size)) {
+    return 0;
+  }
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvBlock, block_payload_size))
+    return 0;
+
+  if (WriteUInt(writer, frame->track_number()))
+    return 0;
+
+  if (SerializeInt(writer, timecode, 2))
+    return 0;
+
+  // For a Block, flags is always 0.
+  if (SerializeInt(writer, 0, 1))
+    return 0;
+
+  if (writer->Write(frame->frame(), static_cast<uint32_t>(frame->length())))
+    return 0;
+
+  if (frame->additional()) {
+    if (!WriteEbmlMasterElement(writer, libwebm::kMkvBlockAdditions,
+                                block_additions_payload_size)) {
+      return 0;
+    }
+
+    if (!WriteEbmlMasterElement(writer, libwebm::kMkvBlockMore,
+                                block_more_payload_size))
+      return 0;
+
+    if (!WriteEbmlElement(writer, libwebm::kMkvBlockAddID, frame->add_id()))
+      return 0;
+
+    if (!WriteEbmlElement(writer, libwebm::kMkvBlockAdditional,
+                          frame->additional(), frame->additional_length())) {
+      return 0;
+    }
+  }
+
+  if (frame->discard_padding() != 0 &&
+      !WriteEbmlElement(writer, libwebm::kMkvDiscardPadding,
+                        frame->discard_padding())) {
+    return false;
+  }
+
+  if (!frame->is_key() &&
+      !WriteEbmlElement(writer, libwebm::kMkvReferenceBlock,
+                        reference_block_timestamp)) {
+    return false;
+  }
+
+  if (duration > 0 &&
+      !WriteEbmlElement(writer, libwebm::kMkvBlockDuration, duration)) {
+    return false;
+  }
+  return EbmlMasterElementSize(libwebm::kMkvBlockGroup,
+                               block_group_payload_size) +
+         block_group_payload_size;
+}
+
+uint64_t WriteSimpleBlock(IMkvWriter* writer, const Frame* const frame,
+                          int64_t timecode) {
+  if (WriteID(writer, libwebm::kMkvSimpleBlock))
+    return 0;
+
+  const int32_t size = static_cast<int32_t>(frame->length()) + 4;
+  if (WriteUInt(writer, size))
+    return 0;
+
+  if (WriteUInt(writer, static_cast<uint64_t>(frame->track_number())))
+    return 0;
+
+  if (SerializeInt(writer, timecode, 2))
+    return 0;
+
+  uint64_t flags = 0;
+  if (frame->is_key())
+    flags |= 0x80;
+
+  if (SerializeInt(writer, flags, 1))
+    return 0;
+
+  if (writer->Write(frame->frame(), static_cast<uint32_t>(frame->length())))
+    return 0;
+
+  return static_cast<uint64_t>(GetUIntSize(libwebm::kMkvSimpleBlock) +
+                               GetCodedUIntSize(size) + 4 + frame->length());
+}
+
+}  // namespace
+
+int32_t GetCodedUIntSize(uint64_t value) {
+  if (value < 0x000000000000007FULL)
+    return 1;
+  else if (value < 0x0000000000003FFFULL)
+    return 2;
+  else if (value < 0x00000000001FFFFFULL)
+    return 3;
+  else if (value < 0x000000000FFFFFFFULL)
+    return 4;
+  else if (value < 0x00000007FFFFFFFFULL)
+    return 5;
+  else if (value < 0x000003FFFFFFFFFFULL)
+    return 6;
+  else if (value < 0x0001FFFFFFFFFFFFULL)
+    return 7;
+  return 8;
+}
+
+int32_t GetUIntSize(uint64_t value) {
+  if (value < 0x0000000000000100ULL)
+    return 1;
+  else if (value < 0x0000000000010000ULL)
+    return 2;
+  else if (value < 0x0000000001000000ULL)
+    return 3;
+  else if (value < 0x0000000100000000ULL)
+    return 4;
+  else if (value < 0x0000010000000000ULL)
+    return 5;
+  else if (value < 0x0001000000000000ULL)
+    return 6;
+  else if (value < 0x0100000000000000ULL)
+    return 7;
+  return 8;
+}
+
+int32_t GetIntSize(int64_t value) {
+  // Doubling the requested value ensures positive values with their high bit
+  // set are written with 0-padding to avoid flipping the signedness.
+  const uint64_t v = (value < 0) ? value ^ -1LL : value;
+  return GetUIntSize(2 * v);
+}
+
+uint64_t EbmlMasterElementSize(uint64_t type, uint64_t value) {
+  // Size of EBML ID
+  int32_t ebml_size = GetUIntSize(type);
+
+  // Datasize
+  ebml_size += GetCodedUIntSize(value);
+
+  return static_cast<uint64_t>(ebml_size);
+}
+
+uint64_t EbmlElementSize(uint64_t type, int64_t value) {
+  // Size of EBML ID
+  int32_t ebml_size = GetUIntSize(type);
+
+  // Datasize
+  ebml_size += GetIntSize(value);
+
+  // Size of Datasize
+  ebml_size++;
+
+  return static_cast<uint64_t>(ebml_size);
+}
+
+uint64_t EbmlElementSize(uint64_t type, uint64_t value) {
+  return EbmlElementSize(type, value, 0);
+}
+
+uint64_t EbmlElementSize(uint64_t type, uint64_t value, uint64_t fixed_size) {
+  // Size of EBML ID
+  uint64_t ebml_size = static_cast<uint64_t>(GetUIntSize(type));
+
+  // Datasize
+  ebml_size +=
+      (fixed_size > 0) ? fixed_size : static_cast<uint64_t>(GetUIntSize(value));
+
+  // Size of Datasize
+  ebml_size++;
+
+  return ebml_size;
+}
+
+uint64_t EbmlElementSize(uint64_t type, float /* value */) {
+  // Size of EBML ID
+  uint64_t ebml_size = static_cast<uint64_t>(GetUIntSize(type));
+
+  // Datasize
+  ebml_size += sizeof(float);
+
+  // Size of Datasize
+  ebml_size++;
+
+  return ebml_size;
+}
+
+uint64_t EbmlElementSize(uint64_t type, const char* value) {
+  if (!value)
+    return 0;
+
+  // Size of EBML ID
+  uint64_t ebml_size = static_cast<uint64_t>(GetUIntSize(type));
+
+  // Datasize
+  ebml_size += strlen(value);
+
+  // Size of Datasize
+  ebml_size++;
+
+  return ebml_size;
+}
+
+uint64_t EbmlElementSize(uint64_t type, const uint8_t* value, uint64_t size) {
+  if (!value)
+    return 0;
+
+  // Size of EBML ID
+  uint64_t ebml_size = static_cast<uint64_t>(GetUIntSize(type));
+
+  // Datasize
+  ebml_size += size;
+
+  // Size of Datasize
+  ebml_size += GetCodedUIntSize(size);
+
+  return ebml_size;
+}
+
+uint64_t EbmlDateElementSize(uint64_t type) {
+  // Size of EBML ID
+  uint64_t ebml_size = static_cast<uint64_t>(GetUIntSize(type));
+
+  // Datasize
+  ebml_size += kDateElementSize;
+
+  // Size of Datasize
+  ebml_size++;
+
+  return ebml_size;
+}
+
+int32_t SerializeInt(IMkvWriter* writer, int64_t value, int32_t size) {
+  if (!writer || size < 1 || size > 8)
+    return -1;
+
+  for (int32_t i = 1; i <= size; ++i) {
+    const int32_t byte_count = size - i;
+    const int32_t bit_count = byte_count * 8;
+
+    const int64_t bb = value >> bit_count;
+    const uint8_t b = static_cast<uint8_t>(bb);
+
+    const int32_t status = writer->Write(&b, 1);
+
+    if (status < 0)
+      return status;
+  }
+
+  return 0;
+}
+
+int32_t SerializeFloat(IMkvWriter* writer, float f) {
+  if (!writer)
+    return -1;
+
+  assert(sizeof(uint32_t) == sizeof(float));
+  // This union is merely used to avoid a reinterpret_cast from float& to
+  // uint32& which will result in violation of strict aliasing.
+  union U32 {
+    uint32_t u32;
+    float f;
+  } value;
+  value.f = f;
+
+  for (int32_t i = 1; i <= 4; ++i) {
+    const int32_t byte_count = 4 - i;
+    const int32_t bit_count = byte_count * 8;
+
+    const uint8_t byte = static_cast<uint8_t>(value.u32 >> bit_count);
+
+    const int32_t status = writer->Write(&byte, 1);
+
+    if (status < 0)
+      return status;
+  }
+
+  return 0;
+}
+
+int32_t WriteUInt(IMkvWriter* writer, uint64_t value) {
+  if (!writer)
+    return -1;
+
+  int32_t size = GetCodedUIntSize(value);
+
+  return WriteUIntSize(writer, value, size);
+}
+
+int32_t WriteUIntSize(IMkvWriter* writer, uint64_t value, int32_t size) {
+  if (!writer || size < 0 || size > 8)
+    return -1;
+
+  if (size > 0) {
+    const uint64_t bit = 1LL << (size * 7);
+
+    if (value > (bit - 2))
+      return -1;
+
+    value |= bit;
+  } else {
+    size = 1;
+    int64_t bit;
+
+    for (;;) {
+      bit = 1LL << (size * 7);
+      const uint64_t max = bit - 2;
+
+      if (value <= max)
+        break;
+
+      ++size;
+    }
+
+    if (size > 8)
+      return false;
+
+    value |= bit;
+  }
+
+  return SerializeInt(writer, value, size);
+}
+
+int32_t WriteID(IMkvWriter* writer, uint64_t type) {
+  if (!writer)
+    return -1;
+
+  writer->ElementStartNotify(type, writer->Position());
+
+  const int32_t size = GetUIntSize(type);
+
+  return SerializeInt(writer, type, size);
+}
+
+bool WriteEbmlMasterElement(IMkvWriter* writer, uint64_t type, uint64_t size) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  if (WriteUInt(writer, size))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, uint64_t value) {
+  return WriteEbmlElement(writer, type, value, 0);
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, uint64_t value,
+                      uint64_t fixed_size) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  uint64_t size = static_cast<uint64_t>(GetUIntSize(value));
+  if (fixed_size > 0) {
+    if (size > fixed_size)
+      return false;
+    size = fixed_size;
+  }
+  if (WriteUInt(writer, size))
+    return false;
+
+  if (SerializeInt(writer, value, static_cast<int32_t>(size)))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, int64_t value) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return 0;
+
+  const uint64_t size = GetIntSize(value);
+  if (WriteUInt(writer, size))
+    return false;
+
+  if (SerializeInt(writer, value, static_cast<int32_t>(size)))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, float value) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  if (WriteUInt(writer, 4))
+    return false;
+
+  if (SerializeFloat(writer, value))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, const char* value) {
+  if (!writer || !value)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  const uint64_t length = strlen(value);
+  if (WriteUInt(writer, length))
+    return false;
+
+  if (writer->Write(value, static_cast<const uint32_t>(length)))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, const uint8_t* value,
+                      uint64_t size) {
+  if (!writer || !value || size < 1)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  if (WriteUInt(writer, size))
+    return false;
+
+  if (writer->Write(value, static_cast<uint32_t>(size)))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlDateElement(IMkvWriter* writer, uint64_t type, int64_t value) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  if (WriteUInt(writer, kDateElementSize))
+    return false;
+
+  if (SerializeInt(writer, value, kDateElementSize))
+    return false;
+
+  return true;
+}
+
+uint64_t WriteFrame(IMkvWriter* writer, const Frame* const frame,
+                    Cluster* cluster) {
+  if (!writer || !frame || !frame->IsValid() || !cluster ||
+      !cluster->timecode_scale())
+    return 0;
+
+  //  Technically the timecode for a block can be less than the
+  //  timecode for the cluster itself (remember that block timecode
+  //  is a signed, 16-bit integer).  However, as a simplification we
+  //  only permit non-negative cluster-relative timecodes for blocks.
+  const int64_t relative_timecode = cluster->GetRelativeTimecode(
+      frame->timestamp() / cluster->timecode_scale());
+  if (relative_timecode < 0 || relative_timecode > kMaxBlockTimecode)
+    return 0;
+
+  return frame->CanBeSimpleBlock() ?
+             WriteSimpleBlock(writer, frame, relative_timecode) :
+             WriteBlock(writer, frame, relative_timecode,
+                        cluster->timecode_scale());
+}
+
+uint64_t WriteVoidElement(IMkvWriter* writer, uint64_t size) {
+  if (!writer)
+    return false;
+
+  // Subtract one for the void ID and the coded size.
+  uint64_t void_entry_size = size - 1 - GetCodedUIntSize(size - 1);
+  uint64_t void_size =
+      EbmlMasterElementSize(libwebm::kMkvVoid, void_entry_size) +
+      void_entry_size;
+
+  if (void_size != size)
+    return 0;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return 0;
+
+  if (WriteID(writer, libwebm::kMkvVoid))
+    return 0;
+
+  if (WriteUInt(writer, void_entry_size))
+    return 0;
+
+  const uint8_t value = 0;
+  for (int32_t i = 0; i < static_cast<int32_t>(void_entry_size); ++i) {
+    if (writer->Write(&value, 1))
+      return 0;
+  }
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(void_size))
+    return 0;
+
+  return void_size;
+}
+
+void GetVersion(int32_t* major, int32_t* minor, int32_t* build,
+                int32_t* revision) {
+  *major = 0;
+  *minor = 2;
+  *build = 1;
+  *revision = 0;
+}
+
+uint64_t MakeUID(unsigned int* seed) {
+  uint64_t uid = 0;
+
+#ifdef __MINGW32__
+  srand(*seed);
+#endif
+
+  for (int i = 0; i < 7; ++i) {  // avoid problems with 8-byte values
+    uid <<= 8;
+
+// TODO(fgalligan): Move random number generation to platform specific code.
+#ifdef _MSC_VER
+    (void)seed;
+    const int32_t nn = rand();
+#elif __ANDROID__
+    int32_t temp_num = 1;
+    int fd = open("/dev/urandom", O_RDONLY);
+    if (fd != -1) {
+      read(fd, &temp_num, sizeof(temp_num));
+      close(fd);
+    }
+    const int32_t nn = temp_num;
+#elif defined __MINGW32__
+    const int32_t nn = rand();
+#else
+    const int32_t nn = rand_r(seed);
+#endif
+    const int32_t n = 0xFF & (nn >> 4);  // throw away low-order bits
+
+    uid |= n;
+  }
+
+  return uid;
+}
+
+}  // namespace mkvmuxer
diff --git a/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxerutil.h b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxerutil.h
new file mode 100644
index 0000000..0e21a2d
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvmuxerutil.h
@@ -0,0 +1,95 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef MKVMUXER_MKVMUXERUTIL_H_
+#define MKVMUXER_MKVMUXERUTIL_H_
+
+#include <stdint.h>
+
+namespace mkvmuxer {
+class Cluster;
+class Frame;
+class IMkvWriter;
+
+const uint64_t kEbmlUnknownValue = 0x01FFFFFFFFFFFFFFULL;
+const int64_t kMaxBlockTimecode = 0x07FFFLL;
+
+// Writes out |value| in Big Endian order. Returns 0 on success.
+int32_t SerializeInt(IMkvWriter* writer, int64_t value, int32_t size);
+
+// Returns the size in bytes of the element.
+int32_t GetUIntSize(uint64_t value);
+int32_t GetIntSize(int64_t value);
+int32_t GetCodedUIntSize(uint64_t value);
+uint64_t EbmlMasterElementSize(uint64_t type, uint64_t value);
+uint64_t EbmlElementSize(uint64_t type, int64_t value);
+uint64_t EbmlElementSize(uint64_t type, uint64_t value);
+uint64_t EbmlElementSize(uint64_t type, float value);
+uint64_t EbmlElementSize(uint64_t type, const char* value);
+uint64_t EbmlElementSize(uint64_t type, const uint8_t* value, uint64_t size);
+uint64_t EbmlDateElementSize(uint64_t type);
+
+// Returns the size in bytes of the element assuming that the element was
+// written using |fixed_size| bytes. If |fixed_size| is set to zero, then it
+// computes the necessary number of bytes based on |value|.
+uint64_t EbmlElementSize(uint64_t type, uint64_t value, uint64_t fixed_size);
+
+// Creates an EBML coded number from |value| and writes it out. The size of
+// the coded number is determined by the value of |value|. |value| must not
+// be in a coded form. Returns 0 on success.
+int32_t WriteUInt(IMkvWriter* writer, uint64_t value);
+
+// Creates an EBML coded number from |value| and writes it out. The size of
+// the coded number is determined by the value of |size|. |value| must not
+// be in a coded form. Returns 0 on success.
+int32_t WriteUIntSize(IMkvWriter* writer, uint64_t value, int32_t size);
+
+// Output an Mkv master element. Returns true if the element was written.
+bool WriteEbmlMasterElement(IMkvWriter* writer, uint64_t value, uint64_t size);
+
+// Outputs an Mkv ID, calls |IMkvWriter::ElementStartNotify|, and passes the
+// ID to |SerializeInt|. Returns 0 on success.
+int32_t WriteID(IMkvWriter* writer, uint64_t type);
+
+// Output an Mkv non-master element. Returns true if the element was written.
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, uint64_t value);
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, int64_t value);
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, float value);
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, const char* value);
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, const uint8_t* value,
+                      uint64_t size);
+bool WriteEbmlDateElement(IMkvWriter* writer, uint64_t type, int64_t value);
+
+// Output an Mkv non-master element using fixed size. The element will be
+// written out using exactly |fixed_size| bytes. If |fixed_size| is set to zero
+// then it computes the necessary number of bytes based on |value|. Returns true
+// if the element was written.
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, uint64_t value,
+                      uint64_t fixed_size);
+
+// Output a Mkv Frame. It decides the correct element to write (Block vs
+// SimpleBlock) based on the parameters of the Frame.
+uint64_t WriteFrame(IMkvWriter* writer, const Frame* const frame,
+                    Cluster* cluster);
+
+// Output a void element. |size| must be the entire size in bytes that will be
+// void. The function will calculate the size of the void header and subtract
+// it from |size|.
+uint64_t WriteVoidElement(IMkvWriter* writer, uint64_t size);
+
+// Returns the version number of the muxer in |major|, |minor|, |build|,
+// and |revision|.
+void GetVersion(int32_t* major, int32_t* minor, int32_t* build,
+                int32_t* revision);
+
+// Returns a random number to be used for UID, using |seed| to seed
+// the random-number generator (see POSIX rand_r() for semantics).
+uint64_t MakeUID(unsigned int* seed);
+
+}  // namespace mkvmuxer
+
+#endif  // MKVMUXER_MKVMUXERUTIL_H_
diff --git a/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvwriter.cc b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvwriter.cc
new file mode 100644
index 0000000..ca48e14
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvwriter.cc
@@ -0,0 +1,88 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#include "mkvmuxer/mkvwriter.h"
+
+#ifdef _MSC_VER
+#include <share.h>  // for _SH_DENYWR
+#endif
+
+namespace mkvmuxer {
+
+MkvWriter::MkvWriter() : file_(NULL), writer_owns_file_(true) {}
+
+MkvWriter::MkvWriter(FILE* fp) : file_(fp), writer_owns_file_(false) {}
+
+MkvWriter::~MkvWriter() { Close(); }
+
+int32 MkvWriter::Write(const void* buffer, uint32 length) {
+  if (!file_)
+    return -1;
+
+  if (length == 0)
+    return 0;
+
+  if (buffer == NULL)
+    return -1;
+
+  const size_t bytes_written = fwrite(buffer, 1, length, file_);
+
+  return (bytes_written == length) ? 0 : -1;
+}
+
+bool MkvWriter::Open(const char* filename) {
+  if (filename == NULL)
+    return false;
+
+  if (file_)
+    return false;
+
+#ifdef _MSC_VER
+  file_ = _fsopen(filename, "wb", _SH_DENYWR);
+#else
+  file_ = fopen(filename, "wb");
+#endif
+  if (file_ == NULL)
+    return false;
+  return true;
+}
+
+void MkvWriter::Close() {
+  if (file_ && writer_owns_file_) {
+    fclose(file_);
+  }
+  file_ = NULL;
+}
+
+int64 MkvWriter::Position() const {
+  if (!file_)
+    return 0;
+
+#ifdef _MSC_VER
+  return _ftelli64(file_);
+#else
+  return ftell(file_);
+#endif
+}
+
+int32 MkvWriter::Position(int64 position) {
+  if (!file_)
+    return -1;
+
+#ifdef _MSC_VER
+  return _fseeki64(file_, position, SEEK_SET);
+#else
+  return fseek(file_, position, SEEK_SET);
+#endif
+}
+
+bool MkvWriter::Seekable() const { return true; }
+
+void MkvWriter::ElementStartNotify(uint64, int64) {}
+
+}  // namespace mkvmuxer
diff --git a/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvwriter.h b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvwriter.h
new file mode 100644
index 0000000..4227c63
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvmuxer/mkvwriter.h
@@ -0,0 +1,51 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#ifndef MKVMUXER_MKVWRITER_H_
+#define MKVMUXER_MKVWRITER_H_
+
+#include <stdio.h>
+
+#include "mkvmuxer/mkvmuxer.h"
+#include "mkvmuxer/mkvmuxertypes.h"
+
+namespace mkvmuxer {
+
+// Default implementation of the IMkvWriter interface on Windows.
+class MkvWriter : public IMkvWriter {
+ public:
+  MkvWriter();
+  explicit MkvWriter(FILE* fp);
+  virtual ~MkvWriter();
+
+  // IMkvWriter interface
+  virtual int64 Position() const;
+  virtual int32 Position(int64 position);
+  virtual bool Seekable() const;
+  virtual int32 Write(const void* buffer, uint32 length);
+  virtual void ElementStartNotify(uint64 element_id, int64 position);
+
+  // Creates and opens a file for writing. |filename| is the name of the file
+  // to open. This function will overwrite the contents of |filename|. Returns
+  // true on success.
+  bool Open(const char* filename);
+
+  // Closes an opened file.
+  void Close();
+
+ private:
+  // File handle to output file.
+  FILE* file_;
+  bool writer_owns_file_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(MkvWriter);
+};
+
+}  // namespace mkvmuxer
+
+#endif  // MKVMUXER_MKVWRITER_H_
diff --git a/libvpx/libvpx/third_party/libwebm/mkvparser/mkvparser.cc b/libvpx/libvpx/third_party/libwebm/mkvparser/mkvparser.cc
new file mode 100644
index 0000000..2180115
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvparser/mkvparser.cc
@@ -0,0 +1,7940 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#include "mkvparser/mkvparser.h"
+
+#if defined(_MSC_VER) && _MSC_VER < 1800
+#include <float.h>  // _isnan() / _finite()
+#define MSC_COMPAT
+#endif
+
+#include <cassert>
+#include <cfloat>
+#include <climits>
+#include <cmath>
+#include <cstring>
+#include <memory>
+#include <new>
+
+#include "common/webmids.h"
+
+namespace mkvparser {
+const float MasteringMetadata::kValueNotPresent = FLT_MAX;
+const long long Colour::kValueNotPresent = LLONG_MAX;
+
+#ifdef MSC_COMPAT
+inline bool isnan(double val) { return !!_isnan(val); }
+inline bool isinf(double val) { return !_finite(val); }
+#else
+inline bool isnan(double val) { return std::isnan(val); }
+inline bool isinf(double val) { return std::isinf(val); }
+#endif  // MSC_COMPAT
+
+IMkvReader::~IMkvReader() {}
+
+template <typename Type>
+Type* SafeArrayAlloc(unsigned long long num_elements,
+                     unsigned long long element_size) {
+  if (num_elements == 0 || element_size == 0)
+    return NULL;
+
+  const size_t kMaxAllocSize = 0x80000000;  // 2GiB
+  const unsigned long long num_bytes = num_elements * element_size;
+  if (element_size > (kMaxAllocSize / num_elements))
+    return NULL;
+  if (num_bytes != static_cast<size_t>(num_bytes))
+    return NULL;
+
+  return new (std::nothrow) Type[static_cast<size_t>(num_bytes)];
+}
+
+void GetVersion(int& major, int& minor, int& build, int& revision) {
+  major = 1;
+  minor = 0;
+  build = 0;
+  revision = 30;
+}
+
+long long ReadUInt(IMkvReader* pReader, long long pos, long& len) {
+  if (!pReader || pos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  len = 1;
+  unsigned char b;
+  int status = pReader->Read(pos, 1, &b);
+
+  if (status < 0)  // error or underflow
+    return status;
+
+  if (status > 0)  // interpreted as "underflow"
+    return E_BUFFER_NOT_FULL;
+
+  if (b == 0)  // we can't handle u-int values larger than 8 bytes
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char m = 0x80;
+
+  while (!(b & m)) {
+    m >>= 1;
+    ++len;
+  }
+
+  long long result = b & (~m);
+  ++pos;
+
+  for (int i = 1; i < len; ++i) {
+    status = pReader->Read(pos, 1, &b);
+
+    if (status < 0) {
+      len = 1;
+      return status;
+    }
+
+    if (status > 0) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result <<= 8;
+    result |= b;
+
+    ++pos;
+  }
+
+  return result;
+}
+
+// Reads an EBML ID and returns it.
+// An ID must at least 1 byte long, cannot exceed 4, and its value must be
+// greater than 0.
+// See known EBML values and EBMLMaxIDLength:
+// http://www.matroska.org/technical/specs/index.html
+// Returns the ID, or a value less than 0 to report an error while reading the
+// ID.
+long long ReadID(IMkvReader* pReader, long long pos, long& len) {
+  if (pReader == NULL || pos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  // Read the first byte. The length in bytes of the ID is determined by
+  // finding the first set bit in the first byte of the ID.
+  unsigned char temp_byte = 0;
+  int read_status = pReader->Read(pos, 1, &temp_byte);
+
+  if (read_status < 0)
+    return E_FILE_FORMAT_INVALID;
+  else if (read_status > 0)  // No data to read.
+    return E_BUFFER_NOT_FULL;
+
+  if (temp_byte == 0)  // ID length > 8 bytes; invalid file.
+    return E_FILE_FORMAT_INVALID;
+
+  int bit_pos = 0;
+  const int kMaxIdLengthInBytes = 4;
+  const int kCheckByte = 0x80;
+
+  // Find the first bit that's set.
+  bool found_bit = false;
+  for (; bit_pos < kMaxIdLengthInBytes; ++bit_pos) {
+    if ((kCheckByte >> bit_pos) & temp_byte) {
+      found_bit = true;
+      break;
+    }
+  }
+
+  if (!found_bit) {
+    // The value is too large to be a valid ID.
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  // Read the remaining bytes of the ID (if any).
+  const int id_length = bit_pos + 1;
+  long long ebml_id = temp_byte;
+  for (int i = 1; i < id_length; ++i) {
+    ebml_id <<= 8;
+    read_status = pReader->Read(pos + i, 1, &temp_byte);
+
+    if (read_status < 0)
+      return E_FILE_FORMAT_INVALID;
+    else if (read_status > 0)
+      return E_BUFFER_NOT_FULL;
+
+    ebml_id |= temp_byte;
+  }
+
+  len = id_length;
+  return ebml_id;
+}
+
+long long GetUIntLength(IMkvReader* pReader, long long pos, long& len) {
+  if (!pReader || pos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  long long total, available;
+
+  int status = pReader->Length(&total, &available);
+  if (status < 0 || (total >= 0 && available > total))
+    return E_FILE_FORMAT_INVALID;
+
+  len = 1;
+
+  if (pos >= available)
+    return pos;  // too few bytes available
+
+  unsigned char b;
+
+  status = pReader->Read(pos, 1, &b);
+
+  if (status != 0)
+    return status;
+
+  if (b == 0)  // we can't handle u-int values larger than 8 bytes
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char m = 0x80;
+
+  while (!(b & m)) {
+    m >>= 1;
+    ++len;
+  }
+
+  return 0;  // success
+}
+
+// TODO(vigneshv): This function assumes that unsigned values never have their
+// high bit set.
+long long UnserializeUInt(IMkvReader* pReader, long long pos, long long size) {
+  if (!pReader || pos < 0 || (size <= 0) || (size > 8))
+    return E_FILE_FORMAT_INVALID;
+
+  long long result = 0;
+
+  for (long long i = 0; i < size; ++i) {
+    unsigned char b;
+
+    const long status = pReader->Read(pos, 1, &b);
+
+    if (status < 0)
+      return status;
+
+    result <<= 8;
+    result |= b;
+
+    ++pos;
+  }
+
+  return result;
+}
+
+long UnserializeFloat(IMkvReader* pReader, long long pos, long long size_,
+                      double& result) {
+  if (!pReader || pos < 0 || ((size_ != 4) && (size_ != 8)))
+    return E_FILE_FORMAT_INVALID;
+
+  const long size = static_cast<long>(size_);
+
+  unsigned char buf[8];
+
+  const int status = pReader->Read(pos, size, buf);
+
+  if (status < 0)  // error
+    return status;
+
+  if (size == 4) {
+    union {
+      float f;
+      unsigned long ff;
+    };
+
+    ff = 0;
+
+    for (int i = 0;;) {
+      ff |= buf[i];
+
+      if (++i >= 4)
+        break;
+
+      ff <<= 8;
+    }
+
+    result = f;
+  } else {
+    union {
+      double d;
+      unsigned long long dd;
+    };
+
+    dd = 0;
+
+    for (int i = 0;;) {
+      dd |= buf[i];
+
+      if (++i >= 8)
+        break;
+
+      dd <<= 8;
+    }
+
+    result = d;
+  }
+
+  if (mkvparser::isinf(result) || mkvparser::isnan(result))
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+long UnserializeInt(IMkvReader* pReader, long long pos, long long size,
+                    long long& result_ref) {
+  if (!pReader || pos < 0 || size < 1 || size > 8)
+    return E_FILE_FORMAT_INVALID;
+
+  signed char first_byte = 0;
+  const long status = pReader->Read(pos, 1, (unsigned char*)&first_byte);
+
+  if (status < 0)
+    return status;
+
+  unsigned long long result = first_byte;
+  ++pos;
+
+  for (long i = 1; i < size; ++i) {
+    unsigned char b;
+
+    const long status = pReader->Read(pos, 1, &b);
+
+    if (status < 0)
+      return status;
+
+    result <<= 8;
+    result |= b;
+
+    ++pos;
+  }
+
+  result_ref = static_cast<long long>(result);
+  return 0;
+}
+
+long UnserializeString(IMkvReader* pReader, long long pos, long long size,
+                       char*& str) {
+  delete[] str;
+  str = NULL;
+
+  if (size >= LONG_MAX || size < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  // +1 for '\0' terminator
+  const long required_size = static_cast<long>(size) + 1;
+
+  str = SafeArrayAlloc<char>(1, required_size);
+  if (str == NULL)
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char* const buf = reinterpret_cast<unsigned char*>(str);
+
+  const long status = pReader->Read(pos, static_cast<long>(size), buf);
+
+  if (status) {
+    delete[] str;
+    str = NULL;
+
+    return status;
+  }
+
+  str[required_size - 1] = '\0';
+  return 0;
+}
+
+long ParseElementHeader(IMkvReader* pReader, long long& pos, long long stop,
+                        long long& id, long long& size) {
+  if (stop >= 0 && pos >= stop)
+    return E_FILE_FORMAT_INVALID;
+
+  long len;
+
+  id = ReadID(pReader, pos, len);
+
+  if (id < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume id
+
+  if (stop >= 0 && pos >= stop)
+    return E_FILE_FORMAT_INVALID;
+
+  size = ReadUInt(pReader, pos, len);
+
+  if (size < 0 || len < 1 || len > 8) {
+    // Invalid: Negative payload size, negative or 0 length integer, or integer
+    // larger than 64 bits (libwebm cannot handle them).
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  // Avoid rolling over pos when very close to LLONG_MAX.
+  const unsigned long long rollover_check =
+      static_cast<unsigned long long>(pos) + len;
+  if (rollover_check > LLONG_MAX)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume length of size
+
+  // pos now designates payload
+
+  if (stop >= 0 && pos > stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;  // success
+}
+
+bool Match(IMkvReader* pReader, long long& pos, unsigned long expected_id,
+           long long& val) {
+  if (!pReader || pos < 0)
+    return false;
+
+  long long total = 0;
+  long long available = 0;
+
+  const long status = pReader->Length(&total, &available);
+  if (status < 0 || (total >= 0 && available > total))
+    return false;
+
+  long len = 0;
+
+  const long long id = ReadID(pReader, pos, len);
+  if (id < 0 || (available - pos) > len)
+    return false;
+
+  if (static_cast<unsigned long>(id) != expected_id)
+    return false;
+
+  pos += len;  // consume id
+
+  const long long size = ReadUInt(pReader, pos, len);
+  if (size < 0 || size > 8 || len < 1 || len > 8 || (available - pos) > len)
+    return false;
+
+  pos += len;  // consume length of size of payload
+
+  val = UnserializeUInt(pReader, pos, size);
+  if (val < 0)
+    return false;
+
+  pos += size;  // consume size of payload
+
+  return true;
+}
+
+bool Match(IMkvReader* pReader, long long& pos, unsigned long expected_id,
+           unsigned char*& buf, size_t& buflen) {
+  if (!pReader || pos < 0)
+    return false;
+
+  long long total = 0;
+  long long available = 0;
+
+  long status = pReader->Length(&total, &available);
+  if (status < 0 || (total >= 0 && available > total))
+    return false;
+
+  long len = 0;
+  const long long id = ReadID(pReader, pos, len);
+  if (id < 0 || (available - pos) > len)
+    return false;
+
+  if (static_cast<unsigned long>(id) != expected_id)
+    return false;
+
+  pos += len;  // consume id
+
+  const long long size = ReadUInt(pReader, pos, len);
+  if (size < 0 || len <= 0 || len > 8 || (available - pos) > len)
+    return false;
+
+  unsigned long long rollover_check =
+      static_cast<unsigned long long>(pos) + len;
+  if (rollover_check > LLONG_MAX)
+    return false;
+
+  pos += len;  // consume length of size of payload
+
+  rollover_check = static_cast<unsigned long long>(pos) + size;
+  if (rollover_check > LLONG_MAX)
+    return false;
+
+  if ((pos + size) > available)
+    return false;
+
+  if (size >= LONG_MAX)
+    return false;
+
+  const long buflen_ = static_cast<long>(size);
+
+  buf = SafeArrayAlloc<unsigned char>(1, buflen_);
+  if (!buf)
+    return false;
+
+  status = pReader->Read(pos, buflen_, buf);
+  if (status != 0)
+    return false;
+
+  buflen = buflen_;
+
+  pos += size;  // consume size of payload
+  return true;
+}
+
+EBMLHeader::EBMLHeader() : m_docType(NULL) { Init(); }
+
+EBMLHeader::~EBMLHeader() { delete[] m_docType; }
+
+void EBMLHeader::Init() {
+  m_version = 1;
+  m_readVersion = 1;
+  m_maxIdLength = 4;
+  m_maxSizeLength = 8;
+
+  if (m_docType) {
+    delete[] m_docType;
+    m_docType = NULL;
+  }
+
+  m_docTypeVersion = 1;
+  m_docTypeReadVersion = 1;
+}
+
+long long EBMLHeader::Parse(IMkvReader* pReader, long long& pos) {
+  if (!pReader)
+    return E_FILE_FORMAT_INVALID;
+
+  long long total, available;
+
+  long status = pReader->Length(&total, &available);
+
+  if (status < 0)  // error
+    return status;
+
+  pos = 0;
+
+  // Scan until we find what looks like the first byte of the EBML header.
+  const long long kMaxScanBytes = (available >= 1024) ? 1024 : available;
+  const unsigned char kEbmlByte0 = 0x1A;
+  unsigned char scan_byte = 0;
+
+  while (pos < kMaxScanBytes) {
+    status = pReader->Read(pos, 1, &scan_byte);
+
+    if (status < 0)  // error
+      return status;
+    else if (status > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if (scan_byte == kEbmlByte0)
+      break;
+
+    ++pos;
+  }
+
+  long len = 0;
+  const long long ebml_id = ReadID(pReader, pos, len);
+
+  if (ebml_id == E_BUFFER_NOT_FULL)
+    return E_BUFFER_NOT_FULL;
+
+  if (len != 4 || ebml_id != libwebm::kMkvEBML)
+    return E_FILE_FORMAT_INVALID;
+
+  // Move read pos forward to the EBML header size field.
+  pos += 4;
+
+  // Read length of size field.
+  long long result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error
+    return E_FILE_FORMAT_INVALID;
+  else if (result > 0)  // need more data
+    return E_BUFFER_NOT_FULL;
+
+  if (len < 1 || len > 8)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((total >= 0) && ((total - pos) < len))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((available - pos) < len)
+    return pos + len;  // try again later
+
+  // Read the EBML header size.
+  result = ReadUInt(pReader, pos, len);
+
+  if (result < 0)  // error
+    return result;
+
+  pos += len;  // consume size field
+
+  // pos now designates start of payload
+
+  if ((total >= 0) && ((total - pos) < result))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((available - pos) < result)
+    return pos + result;
+
+  const long long end = pos + result;
+
+  Init();
+
+  while (pos < end) {
+    long long id, size;
+
+    status = ParseElementHeader(pReader, pos, end, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvEBMLVersion) {
+      m_version = UnserializeUInt(pReader, pos, size);
+
+      if (m_version <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvEBMLReadVersion) {
+      m_readVersion = UnserializeUInt(pReader, pos, size);
+
+      if (m_readVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvEBMLMaxIDLength) {
+      m_maxIdLength = UnserializeUInt(pReader, pos, size);
+
+      if (m_maxIdLength <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvEBMLMaxSizeLength) {
+      m_maxSizeLength = UnserializeUInt(pReader, pos, size);
+
+      if (m_maxSizeLength <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDocType) {
+      if (m_docType)
+        return E_FILE_FORMAT_INVALID;
+
+      status = UnserializeString(pReader, pos, size, m_docType);
+
+      if (status)  // error
+        return status;
+    } else if (id == libwebm::kMkvDocTypeVersion) {
+      m_docTypeVersion = UnserializeUInt(pReader, pos, size);
+
+      if (m_docTypeVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDocTypeReadVersion) {
+      m_docTypeReadVersion = UnserializeUInt(pReader, pos, size);
+
+      if (m_docTypeReadVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;
+  }
+
+  if (pos != end)
+    return E_FILE_FORMAT_INVALID;
+
+  // Make sure DocType, DocTypeReadVersion, and DocTypeVersion are valid.
+  if (m_docType == NULL || m_docTypeReadVersion <= 0 || m_docTypeVersion <= 0)
+    return E_FILE_FORMAT_INVALID;
+
+  // Make sure EBMLMaxIDLength and EBMLMaxSizeLength are valid.
+  if (m_maxIdLength <= 0 || m_maxIdLength > 4 || m_maxSizeLength <= 0 ||
+      m_maxSizeLength > 8)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+Segment::Segment(IMkvReader* pReader, long long elem_start,
+                 // long long elem_size,
+                 long long start, long long size)
+    : m_pReader(pReader),
+      m_element_start(elem_start),
+      // m_element_size(elem_size),
+      m_start(start),
+      m_size(size),
+      m_pos(start),
+      m_pUnknownSize(0),
+      m_pSeekHead(NULL),
+      m_pInfo(NULL),
+      m_pTracks(NULL),
+      m_pCues(NULL),
+      m_pChapters(NULL),
+      m_pTags(NULL),
+      m_clusters(NULL),
+      m_clusterCount(0),
+      m_clusterPreloadCount(0),
+      m_clusterSize(0) {}
+
+Segment::~Segment() {
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  Cluster** i = m_clusters;
+  Cluster** j = m_clusters + count;
+
+  while (i != j) {
+    Cluster* const p = *i++;
+    delete p;
+  }
+
+  delete[] m_clusters;
+
+  delete m_pTracks;
+  delete m_pInfo;
+  delete m_pCues;
+  delete m_pChapters;
+  delete m_pTags;
+  delete m_pSeekHead;
+}
+
+long long Segment::CreateInstance(IMkvReader* pReader, long long pos,
+                                  Segment*& pSegment) {
+  if (pReader == NULL || pos < 0)
+    return E_PARSE_FAILED;
+
+  pSegment = NULL;
+
+  long long total, available;
+
+  const long status = pReader->Length(&total, &available);
+
+  if (status < 0)  // error
+    return status;
+
+  if (available < 0)
+    return -1;
+
+  if ((total >= 0) && (available > total))
+    return -1;
+
+  // I would assume that in practice this loop would execute
+  // exactly once, but we allow for other elements (e.g. Void)
+  // to immediately follow the EBML header.  This is fine for
+  // the source filter case (since the entire file is available),
+  // but in the splitter case over a network we should probably
+  // just give up early.  We could for example decide only to
+  // execute this loop a maximum of, say, 10 times.
+  // TODO:
+  // There is an implied "give up early" by only parsing up
+  // to the available limit.  We do do that, but only if the
+  // total file size is unknown.  We could decide to always
+  // use what's available as our limit (irrespective of whether
+  // we happen to know the total file length).  This would have
+  // as its sense "parse this much of the file before giving up",
+  // which a slightly different sense from "try to parse up to
+  // 10 EMBL elements before giving up".
+
+  for (;;) {
+    if ((total >= 0) && (pos >= total))
+      return E_FILE_FORMAT_INVALID;
+
+    // Read ID
+    long len;
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result)  // error, or too few available bytes
+      return result;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    const long long idpos = pos;
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result)  // error, or too few available bytes
+      return result;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return size;
+
+    pos += len;  // consume length of size of element
+
+    // Pos now points to start of payload
+
+    // Handle "unknown size" for live streaming of webm files.
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (id == libwebm::kMkvSegment) {
+      if (size == unknown_size)
+        size = -1;
+
+      else if (total < 0)
+        size = -1;
+
+      else if ((pos + size) > total)
+        size = -1;
+
+      pSegment = new (std::nothrow) Segment(pReader, idpos, pos, size);
+      if (pSegment == NULL)
+        return E_PARSE_FAILED;
+
+      return 0;  // success
+    }
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && ((pos + size) > total))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + size) > available)
+      return pos + size;
+
+    pos += size;  // consume payload
+  }
+}
+
+long long Segment::ParseHeaders() {
+  // Outermost (level 0) segment object has been constructed,
+  // and pos designates start of payload.  We need to find the
+  // inner (level 1) elements.
+  long long total, available;
+
+  const int status = m_pReader->Length(&total, &available);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total > 0 && available > total)
+    return E_FILE_FORMAT_INVALID;
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  if ((segment_stop >= 0 && total >= 0 && segment_stop > total) ||
+      (segment_stop >= 0 && m_pos > segment_stop)) {
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  for (;;) {
+    if ((total >= 0) && (m_pos >= total))
+      break;
+
+    if ((segment_stop >= 0) && (m_pos >= segment_stop))
+      break;
+
+    long long pos = m_pos;
+    const long long element_start = pos;
+
+    // Avoid rolling over pos when very close to LLONG_MAX.
+    unsigned long long rollover_check = pos + 1ULL;
+    if (rollover_check > LLONG_MAX)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + 1) > available)
+      return (pos + 1);
+
+    long len;
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return result;
+
+    if (result > 0) {
+      // MkvReader doesn't have enough data to satisfy this read attempt.
+      return (pos + 1);
+    }
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    const long long idpos = pos;
+    const long long id = ReadID(m_pReader, idpos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvCluster)
+      break;
+
+    pos += len;  // consume ID
+
+    if ((pos + 1) > available)
+      return (pos + 1);
+
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return result;
+
+    if (result > 0) {
+      // MkvReader doesn't have enough data to satisfy this read attempt.
+      return (pos + 1);
+    }
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0 || len < 1 || len > 8) {
+      // TODO(tomfinegan): ReadUInt should return an error when len is < 1 or
+      // len > 8 is true instead of checking this _everywhere_.
+      return size;
+    }
+
+    pos += len;  // consume length of size of element
+
+    // Avoid rolling over pos when very close to LLONG_MAX.
+    rollover_check = static_cast<unsigned long long>(pos) + size;
+    if (rollover_check > LLONG_MAX)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long element_size = size + pos - element_start;
+
+    // Pos now points to start of payload
+
+    if ((segment_stop >= 0) && ((pos + size) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // We read EBML elements either in total or nothing at all.
+
+    if ((pos + size) > available)
+      return pos + size;
+
+    if (id == libwebm::kMkvInfo) {
+      if (m_pInfo)
+        return E_FILE_FORMAT_INVALID;
+
+      m_pInfo = new (std::nothrow)
+          SegmentInfo(this, pos, size, element_start, element_size);
+
+      if (m_pInfo == NULL)
+        return -1;
+
+      const long status = m_pInfo->Parse();
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvTracks) {
+      if (m_pTracks)
+        return E_FILE_FORMAT_INVALID;
+
+      m_pTracks = new (std::nothrow)
+          Tracks(this, pos, size, element_start, element_size);
+
+      if (m_pTracks == NULL)
+        return -1;
+
+      const long status = m_pTracks->Parse();
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvCues) {
+      if (m_pCues == NULL) {
+        m_pCues = new (std::nothrow)
+            Cues(this, pos, size, element_start, element_size);
+
+        if (m_pCues == NULL)
+          return -1;
+      }
+    } else if (id == libwebm::kMkvSeekHead) {
+      if (m_pSeekHead == NULL) {
+        m_pSeekHead = new (std::nothrow)
+            SeekHead(this, pos, size, element_start, element_size);
+
+        if (m_pSeekHead == NULL)
+          return -1;
+
+        const long status = m_pSeekHead->Parse();
+
+        if (status)
+          return status;
+      }
+    } else if (id == libwebm::kMkvChapters) {
+      if (m_pChapters == NULL) {
+        m_pChapters = new (std::nothrow)
+            Chapters(this, pos, size, element_start, element_size);
+
+        if (m_pChapters == NULL)
+          return -1;
+
+        const long status = m_pChapters->Parse();
+
+        if (status)
+          return status;
+      }
+    } else if (id == libwebm::kMkvTags) {
+      if (m_pTags == NULL) {
+        m_pTags = new (std::nothrow)
+            Tags(this, pos, size, element_start, element_size);
+
+        if (m_pTags == NULL)
+          return -1;
+
+        const long status = m_pTags->Parse();
+
+        if (status)
+          return status;
+      }
+    }
+
+    m_pos = pos + size;  // consume payload
+  }
+
+  if (segment_stop >= 0 && m_pos > segment_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (m_pInfo == NULL)  // TODO: liberalize this behavior
+    return E_FILE_FORMAT_INVALID;
+
+  if (m_pTracks == NULL)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;  // success
+}
+
+long Segment::LoadCluster(long long& pos, long& len) {
+  for (;;) {
+    const long result = DoLoadCluster(pos, len);
+
+    if (result <= 1)
+      return result;
+  }
+}
+
+long Segment::DoLoadCluster(long long& pos, long& len) {
+  if (m_pos < 0)
+    return DoLoadClusterUnknownSize(pos, len);
+
+  long long total, avail;
+
+  long status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total >= 0 && avail > total)
+    return E_FILE_FORMAT_INVALID;
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  long long cluster_off = -1;  // offset relative to start of segment
+  long long cluster_size = -1;  // size of cluster payload
+
+  for (;;) {
+    if ((total >= 0) && (m_pos >= total))
+      return 1;  // no more clusters
+
+    if ((segment_stop >= 0) && (m_pos >= segment_stop))
+      return 1;  // no more clusters
+
+    pos = m_pos;
+
+    // Read ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long idpos = pos;
+    const long long id = ReadID(m_pReader, idpos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume length of size of element
+
+    // pos now points to start of payload
+
+    if (size == 0) {
+      // Missing element payload: move on.
+      m_pos = pos;
+      continue;
+    }
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if ((segment_stop >= 0) && (size != unknown_size) &&
+        ((pos + size) > segment_stop)) {
+      return E_FILE_FORMAT_INVALID;
+    }
+
+    if (id == libwebm::kMkvCues) {
+      if (size == unknown_size) {
+        // Cues element of unknown size: Not supported.
+        return E_FILE_FORMAT_INVALID;
+      }
+
+      if (m_pCues == NULL) {
+        const long long element_size = (pos - idpos) + size;
+
+        m_pCues = new (std::nothrow) Cues(this, pos, size, idpos, element_size);
+        if (m_pCues == NULL)
+          return -1;
+      }
+
+      m_pos = pos + size;  // consume payload
+      continue;
+    }
+
+    if (id != libwebm::kMkvCluster) {
+      // Besides the Segment, Libwebm allows only cluster elements of unknown
+      // size. Fail the parse upon encountering a non-cluster element reporting
+      // unknown size.
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
+
+      m_pos = pos + size;  // consume payload
+      continue;
+    }
+
+    // We have a cluster.
+
+    cluster_off = idpos - m_start;  // relative pos
+
+    if (size != unknown_size)
+      cluster_size = size;
+
+    break;
+  }
+
+  if (cluster_off < 0) {
+    // No cluster, die.
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  long long pos_;
+  long len_;
+
+  status = Cluster::HasBlockEntries(this, cluster_off, pos_, len_);
+
+  if (status < 0) {  // error, or underflow
+    pos = pos_;
+    len = len_;
+
+    return status;
+  }
+
+  // status == 0 means "no block entries found"
+  // status > 0 means "found at least one block entry"
+
+  // TODO:
+  // The issue here is that the segment increments its own
+  // pos ptr past the most recent cluster parsed, and then
+  // starts from there to parse the next cluster.  If we
+  // don't know the size of the current cluster, then we
+  // must either parse its payload (as we do below), looking
+  // for the cluster (or cues) ID to terminate the parse.
+  // This isn't really what we want: rather, we really need
+  // a way to create the curr cluster object immediately.
+  // The pity is that cluster::parse can determine its own
+  // boundary, and we largely duplicate that same logic here.
+  //
+  // Maybe we need to get rid of our look-ahead preloading
+  // in source::parse???
+  //
+  // As we're parsing the blocks in the curr cluster
+  //(in cluster::parse), we should have some way to signal
+  // to the segment that we have determined the boundary,
+  // so it can adjust its own segment::m_pos member.
+  //
+  // The problem is that we're asserting in asyncreadinit,
+  // because we adjust the pos down to the curr seek pos,
+  // and the resulting adjusted len is > 2GB.  I'm suspicious
+  // that this is even correct, but even if it is, we can't
+  // be loading that much data in the cache anyway.
+
+  const long idx = m_clusterCount;
+
+  if (m_clusterPreloadCount > 0) {
+    if (idx >= m_clusterSize)
+      return E_FILE_FORMAT_INVALID;
+
+    Cluster* const pCluster = m_clusters[idx];
+    if (pCluster == NULL || pCluster->m_index >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long off = pCluster->GetPosition();
+    if (off < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (off == cluster_off) {  // preloaded already
+      if (status == 0)  // no entries found
+        return E_FILE_FORMAT_INVALID;
+
+      if (cluster_size >= 0)
+        pos += cluster_size;
+      else {
+        const long long element_size = pCluster->GetElementSize();
+
+        if (element_size <= 0)
+          return E_FILE_FORMAT_INVALID;  // TODO: handle this case
+
+        pos = pCluster->m_element_start + element_size;
+      }
+
+      pCluster->m_index = idx;  // move from preloaded to loaded
+      ++m_clusterCount;
+      --m_clusterPreloadCount;
+
+      m_pos = pos;  // consume payload
+      if (segment_stop >= 0 && m_pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      return 0;  // success
+    }
+  }
+
+  if (status == 0) {  // no entries found
+    if (cluster_size >= 0)
+      pos += cluster_size;
+
+    if ((total >= 0) && (pos >= total)) {
+      m_pos = total;
+      return 1;  // no more clusters
+    }
+
+    if ((segment_stop >= 0) && (pos >= segment_stop)) {
+      m_pos = segment_stop;
+      return 1;  // no more clusters
+    }
+
+    m_pos = pos;
+    return 2;  // try again
+  }
+
+  // status > 0 means we have an entry
+
+  Cluster* const pCluster = Cluster::Create(this, idx, cluster_off);
+  if (pCluster == NULL)
+    return -1;
+
+  if (!AppendCluster(pCluster)) {
+    delete pCluster;
+    return -1;
+  }
+
+  if (cluster_size >= 0) {
+    pos += cluster_size;
+
+    m_pos = pos;
+
+    if (segment_stop > 0 && m_pos > segment_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    return 0;
+  }
+
+  m_pUnknownSize = pCluster;
+  m_pos = -pos;
+
+  return 0;  // partial success, since we have a new cluster
+
+  // status == 0 means "no block entries found"
+  // pos designates start of payload
+  // m_pos has NOT been adjusted yet (in case we need to come back here)
+}
+
+long Segment::DoLoadClusterUnknownSize(long long& pos, long& len) {
+  if (m_pos >= 0 || m_pUnknownSize == NULL)
+    return E_PARSE_FAILED;
+
+  const long status = m_pUnknownSize->Parse(pos, len);
+
+  if (status < 0)  // error or underflow
+    return status;
+
+  if (status == 0)  // parsed a block
+    return 2;  // continue parsing
+
+  const long long start = m_pUnknownSize->m_element_start;
+  const long long size = m_pUnknownSize->GetElementSize();
+
+  if (size < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos = start + size;
+  m_pos = pos;
+
+  m_pUnknownSize = 0;
+
+  return 2;  // continue parsing
+}
+
+bool Segment::AppendCluster(Cluster* pCluster) {
+  if (pCluster == NULL || pCluster->m_index < 0)
+    return false;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  long& size = m_clusterSize;
+  const long idx = pCluster->m_index;
+
+  if (size < count || idx != m_clusterCount)
+    return false;
+
+  if (count >= size) {
+    const long n = (size <= 0) ? 2048 : 2 * size;
+
+    Cluster** const qq = new (std::nothrow) Cluster*[n];
+    if (qq == NULL)
+      return false;
+
+    Cluster** q = qq;
+    Cluster** p = m_clusters;
+    Cluster** const pp = p + count;
+
+    while (p != pp)
+      *q++ = *p++;
+
+    delete[] m_clusters;
+
+    m_clusters = qq;
+    size = n;
+  }
+
+  if (m_clusterPreloadCount > 0) {
+    Cluster** const p = m_clusters + m_clusterCount;
+    if (*p == NULL || (*p)->m_index >= 0)
+      return false;
+
+    Cluster** q = p + m_clusterPreloadCount;
+    if (q >= (m_clusters + size))
+      return false;
+
+    for (;;) {
+      Cluster** const qq = q - 1;
+      if ((*qq)->m_index >= 0)
+        return false;
+
+      *q = *qq;
+      q = qq;
+
+      if (q == p)
+        break;
+    }
+  }
+
+  m_clusters[idx] = pCluster;
+  ++m_clusterCount;
+  return true;
+}
+
+bool Segment::PreloadCluster(Cluster* pCluster, ptrdiff_t idx) {
+  if (pCluster == NULL || pCluster->m_index >= 0 || idx < m_clusterCount)
+    return false;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  long& size = m_clusterSize;
+  if (size < count)
+    return false;
+
+  if (count >= size) {
+    const long n = (size <= 0) ? 2048 : 2 * size;
+
+    Cluster** const qq = new (std::nothrow) Cluster*[n];
+    if (qq == NULL)
+      return false;
+    Cluster** q = qq;
+
+    Cluster** p = m_clusters;
+    Cluster** const pp = p + count;
+
+    while (p != pp)
+      *q++ = *p++;
+
+    delete[] m_clusters;
+
+    m_clusters = qq;
+    size = n;
+  }
+
+  if (m_clusters == NULL)
+    return false;
+
+  Cluster** const p = m_clusters + idx;
+
+  Cluster** q = m_clusters + count;
+  if (q < p || q >= (m_clusters + size))
+    return false;
+
+  while (q > p) {
+    Cluster** const qq = q - 1;
+
+    if ((*qq)->m_index >= 0)
+      return false;
+
+    *q = *qq;
+    q = qq;
+  }
+
+  m_clusters[idx] = pCluster;
+  ++m_clusterPreloadCount;
+  return true;
+}
+
+long Segment::Load() {
+  if (m_clusters != NULL || m_clusterSize != 0 || m_clusterCount != 0)
+    return E_PARSE_FAILED;
+
+  // Outermost (level 0) segment object has been constructed,
+  // and pos designates start of payload.  We need to find the
+  // inner (level 1) elements.
+
+  const long long header_status = ParseHeaders();
+
+  if (header_status < 0)  // error
+    return static_cast<long>(header_status);
+
+  if (header_status > 0)  // underflow
+    return E_BUFFER_NOT_FULL;
+
+  if (m_pInfo == NULL || m_pTracks == NULL)
+    return E_FILE_FORMAT_INVALID;
+
+  for (;;) {
+    const long status = LoadCluster();
+
+    if (status < 0)  // error
+      return status;
+
+    if (status >= 1)  // no more clusters
+      return 0;
+  }
+}
+
+SeekHead::SeekHead(Segment* pSegment, long long start, long long size_,
+                   long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_entries(0),
+      m_entry_count(0),
+      m_void_elements(0),
+      m_void_element_count(0) {}
+
+SeekHead::~SeekHead() {
+  delete[] m_entries;
+  delete[] m_void_elements;
+}
+
+long SeekHead::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
+
+  // first count the seek head entries
+
+  int entry_count = 0;
+  int void_element_count = 0;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvSeek)
+      ++entry_count;
+    else if (id == libwebm::kMkvVoid)
+      ++void_element_count;
+
+    pos += size;  // consume payload
+
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  m_entries = new (std::nothrow) Entry[entry_count];
+
+  if (m_entries == NULL)
+    return -1;
+
+  m_void_elements = new (std::nothrow) VoidElement[void_element_count];
+
+  if (m_void_elements == NULL)
+    return -1;
+
+  // now parse the entries and void elements
+
+  Entry* pEntry = m_entries;
+  VoidElement* pVoidElement = m_void_elements;
+
+  pos = m_start;
+
+  while (pos < stop) {
+    const long long idpos = pos;
+
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvSeek) {
+      if (ParseEntry(pReader, pos, size, pEntry)) {
+        Entry& e = *pEntry++;
+
+        e.element_start = idpos;
+        e.element_size = (pos + size) - idpos;
+      }
+    } else if (id == libwebm::kMkvVoid) {
+      VoidElement& e = *pVoidElement++;
+
+      e.element_start = idpos;
+      e.element_size = (pos + size) - idpos;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  ptrdiff_t count_ = ptrdiff_t(pEntry - m_entries);
+  assert(count_ >= 0);
+  assert(count_ <= entry_count);
+
+  m_entry_count = static_cast<int>(count_);
+
+  count_ = ptrdiff_t(pVoidElement - m_void_elements);
+  assert(count_ >= 0);
+  assert(count_ <= void_element_count);
+
+  m_void_element_count = static_cast<int>(count_);
+
+  return 0;
+}
+
+int SeekHead::GetCount() const { return m_entry_count; }
+
+const SeekHead::Entry* SeekHead::GetEntry(int idx) const {
+  if (idx < 0)
+    return 0;
+
+  if (idx >= m_entry_count)
+    return 0;
+
+  return m_entries + idx;
+}
+
+int SeekHead::GetVoidElementCount() const { return m_void_element_count; }
+
+const SeekHead::VoidElement* SeekHead::GetVoidElement(int idx) const {
+  if (idx < 0)
+    return 0;
+
+  if (idx >= m_void_element_count)
+    return 0;
+
+  return m_void_elements + idx;
+}
+
+long Segment::ParseCues(long long off, long long& pos, long& len) {
+  if (m_pCues)
+    return 0;  // success
+
+  if (off < 0)
+    return -1;
+
+  long long total, avail;
+
+  const int status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  pos = m_start + off;
+
+  if ((total < 0) || (pos >= total))
+    return 1;  // don't bother parsing cues
+
+  const long long element_start = pos;
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long result = GetUIntLength(m_pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)  // underflow (weird)
+  {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long idpos = pos;
+
+  const long long id = ReadID(m_pReader, idpos, len);
+
+  if (id != libwebm::kMkvCues)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume ID
+  assert((segment_stop < 0) || (pos <= segment_stop));
+
+  // Read Size
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  result = GetUIntLength(m_pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)  // underflow (weird)
+  {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long size = ReadUInt(m_pReader, pos, len);
+
+  if (size < 0)  // error
+    return static_cast<long>(size);
+
+  if (size == 0)  // weird, although technically not illegal
+    return 1;  // done
+
+  pos += len;  // consume length of size of element
+  assert((segment_stop < 0) || (pos <= segment_stop));
+
+  // Pos now points to start of payload
+
+  const long long element_stop = pos + size;
+
+  if ((segment_stop >= 0) && (element_stop > segment_stop))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((total >= 0) && (element_stop > total))
+    return 1;  // don't bother parsing anymore
+
+  len = static_cast<long>(size);
+
+  if (element_stop > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long element_size = element_stop - element_start;
+
+  m_pCues =
+      new (std::nothrow) Cues(this, pos, size, element_start, element_size);
+  if (m_pCues == NULL)
+    return -1;
+
+  return 0;  // success
+}
+
+bool SeekHead::ParseEntry(IMkvReader* pReader, long long start, long long size_,
+                          Entry* pEntry) {
+  if (size_ <= 0)
+    return false;
+
+  long long pos = start;
+  const long long stop = start + size_;
+
+  long len;
+
+  // parse the container for the level-1 element ID
+
+  const long long seekIdId = ReadID(pReader, pos, len);
+  if (seekIdId < 0)
+    return false;
+
+  if (seekIdId != libwebm::kMkvSeekID)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume SeekID id
+
+  const long long seekIdSize = ReadUInt(pReader, pos, len);
+
+  if (seekIdSize <= 0)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume size of field
+
+  if ((pos + seekIdSize) > stop)
+    return false;
+
+  // Note that the SeekId payload really is serialized
+  // as a "Matroska integer", not as a plain binary value.
+  // In fact, Matroska requires that ID values in the
+  // stream exactly match the binary representation as listed
+  // in the Matroska specification.
+  //
+  // This parser is more liberal, and permits IDs to have
+  // any width.  (This could make the representation in the stream
+  // different from what's in the spec, but it doesn't matter here,
+  // since we always normalize "Matroska integer" values.)
+
+  pEntry->id = ReadUInt(pReader, pos, len);  // payload
+
+  if (pEntry->id <= 0)
+    return false;
+
+  if (len != seekIdSize)
+    return false;
+
+  pos += seekIdSize;  // consume SeekID payload
+
+  const long long seekPosId = ReadID(pReader, pos, len);
+
+  if (seekPosId != libwebm::kMkvSeekPosition)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume id
+
+  const long long seekPosSize = ReadUInt(pReader, pos, len);
+
+  if (seekPosSize <= 0)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume size
+
+  if ((pos + seekPosSize) > stop)
+    return false;
+
+  pEntry->pos = UnserializeUInt(pReader, pos, seekPosSize);
+
+  if (pEntry->pos < 0)
+    return false;
+
+  pos += seekPosSize;  // consume payload
+
+  if (pos != stop)
+    return false;
+
+  return true;
+}
+
+Cues::Cues(Segment* pSegment, long long start_, long long size_,
+           long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start_),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_cue_points(NULL),
+      m_count(0),
+      m_preload_count(0),
+      m_pos(start_) {}
+
+Cues::~Cues() {
+  const long n = m_count + m_preload_count;
+
+  CuePoint** p = m_cue_points;
+  CuePoint** const q = p + n;
+
+  while (p != q) {
+    CuePoint* const pCP = *p++;
+    assert(pCP);
+
+    delete pCP;
+  }
+
+  delete[] m_cue_points;
+}
+
+long Cues::GetCount() const {
+  if (m_cue_points == NULL)
+    return -1;
+
+  return m_count;  // TODO: really ignore preload count?
+}
+
+bool Cues::DoneParsing() const {
+  const long long stop = m_start + m_size;
+  return (m_pos >= stop);
+}
+
+bool Cues::Init() const {
+  if (m_cue_points)
+    return true;
+
+  if (m_count != 0 || m_preload_count != 0)
+    return false;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  const long long stop = m_start + m_size;
+  long long pos = m_start;
+
+  long cue_points_size = 0;
+
+  while (pos < stop) {
+    const long long idpos = pos;
+
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if (id < 0 || (pos + len) > stop) {
+      return false;
+    }
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    if (size < 0 || (pos + len > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume Size field
+    if (pos + size > stop) {
+      return false;
+    }
+
+    if (id == libwebm::kMkvCuePoint) {
+      if (!PreloadCuePoint(cue_points_size, idpos))
+        return false;
+    }
+
+    pos += size;  // skip payload
+  }
+  return true;
+}
+
+bool Cues::PreloadCuePoint(long& cue_points_size, long long pos) const {
+  if (m_count != 0)
+    return false;
+
+  if (m_preload_count >= cue_points_size) {
+    const long n = (cue_points_size <= 0) ? 2048 : 2 * cue_points_size;
+
+    CuePoint** const qq = new (std::nothrow) CuePoint*[n];
+    if (qq == NULL)
+      return false;
+
+    CuePoint** q = qq;  // beginning of target
+
+    CuePoint** p = m_cue_points;  // beginning of source
+    CuePoint** const pp = p + m_preload_count;  // end of source
+
+    while (p != pp)
+      *q++ = *p++;
+
+    delete[] m_cue_points;
+
+    m_cue_points = qq;
+    cue_points_size = n;
+  }
+
+  CuePoint* const pCP = new (std::nothrow) CuePoint(m_preload_count, pos);
+  if (pCP == NULL)
+    return false;
+
+  m_cue_points[m_preload_count++] = pCP;
+  return true;
+}
+
+bool Cues::LoadCuePoint() const {
+  const long long stop = m_start + m_size;
+
+  if (m_pos >= stop)
+    return false;  // nothing else to do
+
+  if (!Init()) {
+    m_pos = stop;
+    return false;
+  }
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  while (m_pos < stop) {
+    const long long idpos = m_pos;
+
+    long len;
+
+    const long long id = ReadID(pReader, m_pos, len);
+    if (id < 0 || (m_pos + len) > stop)
+      return false;
+
+    m_pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, m_pos, len);
+    if (size < 0 || (m_pos + len) > stop)
+      return false;
+
+    m_pos += len;  // consume Size field
+    if ((m_pos + size) > stop)
+      return false;
+
+    if (id != libwebm::kMkvCuePoint) {
+      m_pos += size;  // consume payload
+      if (m_pos > stop)
+        return false;
+
+      continue;
+    }
+
+    if (m_preload_count < 1)
+      return false;
+
+    CuePoint* const pCP = m_cue_points[m_count];
+    if (!pCP || (pCP->GetTimeCode() < 0 && (-pCP->GetTimeCode() != idpos)))
+      return false;
+
+    if (!pCP->Load(pReader)) {
+      m_pos = stop;
+      return false;
+    }
+    ++m_count;
+    --m_preload_count;
+
+    m_pos += size;  // consume payload
+    if (m_pos > stop)
+      return false;
+
+    return true;  // yes, we loaded a cue point
+  }
+
+  return false;  // no, we did not load a cue point
+}
+
+bool Cues::Find(long long time_ns, const Track* pTrack, const CuePoint*& pCP,
+                const CuePoint::TrackPosition*& pTP) const {
+  if (time_ns < 0 || pTrack == NULL || m_cue_points == NULL || m_count == 0)
+    return false;
+
+  CuePoint** const ii = m_cue_points;
+  CuePoint** i = ii;
+
+  CuePoint** const jj = ii + m_count;
+  CuePoint** j = jj;
+
+  pCP = *i;
+  if (pCP == NULL)
+    return false;
+
+  if (time_ns <= pCP->GetTime(m_pSegment)) {
+    pTP = pCP->Find(pTrack);
+    return (pTP != NULL);
+  }
+
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) <= time_ns
+    //[i, j)  ?
+    //[j, jj) > time_ns
+
+    CuePoint** const k = i + (j - i) / 2;
+    if (k >= jj)
+      return false;
+
+    CuePoint* const pCP = *k;
+    if (pCP == NULL)
+      return false;
+
+    const long long t = pCP->GetTime(m_pSegment);
+
+    if (t <= time_ns)
+      i = k + 1;
+    else
+      j = k;
+
+    if (i > j)
+      return false;
+  }
+
+  if (i != j || i > jj || i <= ii)
+    return false;
+
+  pCP = *--i;
+
+  if (pCP == NULL || pCP->GetTime(m_pSegment) > time_ns)
+    return false;
+
+  // TODO: here and elsewhere, it's probably not correct to search
+  // for the cue point with this time, and then search for a matching
+  // track.  In principle, the matching track could be on some earlier
+  // cue point, and with our current algorithm, we'd miss it.  To make
+  // this bullet-proof, we'd need to create a secondary structure,
+  // with a list of cue points that apply to a track, and then search
+  // that track-based structure for a matching cue point.
+
+  pTP = pCP->Find(pTrack);
+  return (pTP != NULL);
+}
+
+const CuePoint* Cues::GetFirst() const {
+  if (m_cue_points == NULL || m_count == 0)
+    return NULL;
+
+  CuePoint* const* const pp = m_cue_points;
+  if (pp == NULL)
+    return NULL;
+
+  CuePoint* const pCP = pp[0];
+  if (pCP == NULL || pCP->GetTimeCode() < 0)
+    return NULL;
+
+  return pCP;
+}
+
+const CuePoint* Cues::GetLast() const {
+  if (m_cue_points == NULL || m_count <= 0)
+    return NULL;
+
+  const long index = m_count - 1;
+
+  CuePoint* const* const pp = m_cue_points;
+  if (pp == NULL)
+    return NULL;
+
+  CuePoint* const pCP = pp[index];
+  if (pCP == NULL || pCP->GetTimeCode() < 0)
+    return NULL;
+
+  return pCP;
+}
+
+const CuePoint* Cues::GetNext(const CuePoint* pCurr) const {
+  if (pCurr == NULL || pCurr->GetTimeCode() < 0 || m_cue_points == NULL ||
+      m_count < 1) {
+    return NULL;
+  }
+
+  long index = pCurr->m_index;
+  if (index >= m_count)
+    return NULL;
+
+  CuePoint* const* const pp = m_cue_points;
+  if (pp == NULL || pp[index] != pCurr)
+    return NULL;
+
+  ++index;
+
+  if (index >= m_count)
+    return NULL;
+
+  CuePoint* const pNext = pp[index];
+
+  if (pNext == NULL || pNext->GetTimeCode() < 0)
+    return NULL;
+
+  return pNext;
+}
+
+const BlockEntry* Cues::GetBlock(const CuePoint* pCP,
+                                 const CuePoint::TrackPosition* pTP) const {
+  if (pCP == NULL || pTP == NULL)
+    return NULL;
+
+  return m_pSegment->GetBlock(*pCP, *pTP);
+}
+
+const BlockEntry* Segment::GetBlock(const CuePoint& cp,
+                                    const CuePoint::TrackPosition& tp) {
+  Cluster** const ii = m_clusters;
+  Cluster** i = ii;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  Cluster** const jj = ii + count;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) < pTP->m_pos
+    //[i, j) ?
+    //[j, jj)  > pTP->m_pos
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    Cluster* const pCluster = *k;
+    assert(pCluster);
+
+    // const long long pos_ = pCluster->m_pos;
+    // assert(pos_);
+    // const long long pos = pos_ * ((pos_ < 0) ? -1 : 1);
+
+    const long long pos = pCluster->GetPosition();
+    assert(pos >= 0);
+
+    if (pos < tp.m_pos)
+      i = k + 1;
+    else if (pos > tp.m_pos)
+      j = k;
+    else
+      return pCluster->GetEntry(cp, tp);
+  }
+
+  assert(i == j);
+  // assert(Cluster::HasBlockEntries(this, tp.m_pos));
+
+  Cluster* const pCluster = Cluster::Create(this, -1, tp.m_pos);  //, -1);
+  if (pCluster == NULL)
+    return NULL;
+
+  const ptrdiff_t idx = i - m_clusters;
+
+  if (!PreloadCluster(pCluster, idx)) {
+    delete pCluster;
+    return NULL;
+  }
+  assert(m_clusters);
+  assert(m_clusterPreloadCount > 0);
+  assert(m_clusters[idx] == pCluster);
+
+  return pCluster->GetEntry(cp, tp);
+}
+
+const Cluster* Segment::FindOrPreloadCluster(long long requested_pos) {
+  if (requested_pos < 0)
+    return 0;
+
+  Cluster** const ii = m_clusters;
+  Cluster** i = ii;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  Cluster** const jj = ii + count;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) < pTP->m_pos
+    //[i, j) ?
+    //[j, jj)  > pTP->m_pos
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    Cluster* const pCluster = *k;
+    assert(pCluster);
+
+    // const long long pos_ = pCluster->m_pos;
+    // assert(pos_);
+    // const long long pos = pos_ * ((pos_ < 0) ? -1 : 1);
+
+    const long long pos = pCluster->GetPosition();
+    assert(pos >= 0);
+
+    if (pos < requested_pos)
+      i = k + 1;
+    else if (pos > requested_pos)
+      j = k;
+    else
+      return pCluster;
+  }
+
+  assert(i == j);
+  // assert(Cluster::HasBlockEntries(this, tp.m_pos));
+
+  Cluster* const pCluster = Cluster::Create(this, -1, requested_pos);
+  if (pCluster == NULL)
+    return NULL;
+
+  const ptrdiff_t idx = i - m_clusters;
+
+  if (!PreloadCluster(pCluster, idx)) {
+    delete pCluster;
+    return NULL;
+  }
+  assert(m_clusters);
+  assert(m_clusterPreloadCount > 0);
+  assert(m_clusters[idx] == pCluster);
+
+  return pCluster;
+}
+
+CuePoint::CuePoint(long idx, long long pos)
+    : m_element_start(0),
+      m_element_size(0),
+      m_index(idx),
+      m_timecode(-1 * pos),
+      m_track_positions(NULL),
+      m_track_positions_count(0) {
+  assert(pos > 0);
+}
+
+CuePoint::~CuePoint() { delete[] m_track_positions; }
+
+bool CuePoint::Load(IMkvReader* pReader) {
+  // odbgstream os;
+  // os << "CuePoint::Load(begin): timecode=" << m_timecode << endl;
+
+  if (m_timecode >= 0)  // already loaded
+    return true;
+
+  assert(m_track_positions == NULL);
+  assert(m_track_positions_count == 0);
+
+  long long pos_ = -m_timecode;
+  const long long element_start = pos_;
+
+  long long stop;
+
+  {
+    long len;
+
+    const long long id = ReadID(pReader, pos_, len);
+    if (id != libwebm::kMkvCuePoint)
+      return false;
+
+    pos_ += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos_, len);
+    assert(size >= 0);
+
+    pos_ += len;  // consume Size field
+    // pos_ now points to start of payload
+
+    stop = pos_ + size;
+  }
+
+  const long long element_size = stop - element_start;
+
+  long long pos = pos_;
+
+  // First count number of track positions
+
+  while (pos < stop) {
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if ((id < 0) || (pos + len > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    if ((size < 0) || (pos + len > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume Size field
+    if ((pos + size) > stop) {
+      return false;
+    }
+
+    if (id == libwebm::kMkvCueTime)
+      m_timecode = UnserializeUInt(pReader, pos, size);
+
+    else if (id == libwebm::kMkvCueTrackPositions)
+      ++m_track_positions_count;
+
+    pos += size;  // consume payload
+  }
+
+  if (m_timecode < 0 || m_track_positions_count <= 0) {
+    return false;
+  }
+
+  // os << "CuePoint::Load(cont'd): idpos=" << idpos
+  //   << " timecode=" << m_timecode
+  //   << endl;
+
+  m_track_positions = new (std::nothrow) TrackPosition[m_track_positions_count];
+  if (m_track_positions == NULL)
+    return false;
+
+  // Now parse track positions
+
+  TrackPosition* p = m_track_positions;
+  pos = pos_;
+
+  while (pos < stop) {
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if (id < 0 || (pos + len) > stop)
+      return false;
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);
+    assert((pos + len) <= stop);
+
+    pos += len;  // consume Size field
+    assert((pos + size) <= stop);
+
+    if (id == libwebm::kMkvCueTrackPositions) {
+      TrackPosition& tp = *p++;
+      if (!tp.Parse(pReader, pos, size)) {
+        return false;
+      }
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return false;
+  }
+
+  assert(size_t(p - m_track_positions) == m_track_positions_count);
+
+  m_element_start = element_start;
+  m_element_size = element_size;
+
+  return true;
+}
+
+bool CuePoint::TrackPosition::Parse(IMkvReader* pReader, long long start_,
+                                    long long size_) {
+  const long long stop = start_ + size_;
+  long long pos = start_;
+
+  m_track = -1;
+  m_pos = -1;
+  m_block = 1;  // default
+
+  while (pos < stop) {
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if ((id < 0) || ((pos + len) > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    if ((size < 0) || ((pos + len) > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume Size field
+    if ((pos + size) > stop) {
+      return false;
+    }
+
+    if (id == libwebm::kMkvCueTrack)
+      m_track = UnserializeUInt(pReader, pos, size);
+    else if (id == libwebm::kMkvCueClusterPosition)
+      m_pos = UnserializeUInt(pReader, pos, size);
+    else if (id == libwebm::kMkvCueBlockNumber)
+      m_block = UnserializeUInt(pReader, pos, size);
+
+    pos += size;  // consume payload
+  }
+
+  if ((m_pos < 0) || (m_track <= 0)) {
+    return false;
+  }
+
+  return true;
+}
+
+const CuePoint::TrackPosition* CuePoint::Find(const Track* pTrack) const {
+  assert(pTrack);
+
+  const long long n = pTrack->GetNumber();
+
+  const TrackPosition* i = m_track_positions;
+  const TrackPosition* const j = i + m_track_positions_count;
+
+  while (i != j) {
+    const TrackPosition& p = *i++;
+
+    if (p.m_track == n)
+      return &p;
+  }
+
+  return NULL;  // no matching track number found
+}
+
+long long CuePoint::GetTimeCode() const { return m_timecode; }
+
+long long CuePoint::GetTime(const Segment* pSegment) const {
+  assert(pSegment);
+  assert(m_timecode >= 0);
+
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+  assert(pInfo);
+
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
+
+  const long long time = scale * m_timecode;
+
+  return time;
+}
+
+bool Segment::DoneParsing() const {
+  if (m_size < 0) {
+    long long total, avail;
+
+    const int status = m_pReader->Length(&total, &avail);
+
+    if (status < 0)  // error
+      return true;  // must assume done
+
+    if (total < 0)
+      return false;  // assume live stream
+
+    return (m_pos >= total);
+  }
+
+  const long long stop = m_start + m_size;
+
+  return (m_pos >= stop);
+}
+
+const Cluster* Segment::GetFirst() const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
+
+  Cluster* const pCluster = m_clusters[0];
+  assert(pCluster);
+
+  return pCluster;
+}
+
+const Cluster* Segment::GetLast() const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
+
+  const long idx = m_clusterCount - 1;
+
+  Cluster* const pCluster = m_clusters[idx];
+  assert(pCluster);
+
+  return pCluster;
+}
+
+unsigned long Segment::GetCount() const { return m_clusterCount; }
+
+const Cluster* Segment::GetNext(const Cluster* pCurr) {
+  assert(pCurr);
+  assert(pCurr != &m_eos);
+  assert(m_clusters);
+
+  long idx = pCurr->m_index;
+
+  if (idx >= 0) {
+    assert(m_clusterCount > 0);
+    assert(idx < m_clusterCount);
+    assert(pCurr == m_clusters[idx]);
+
+    ++idx;
+
+    if (idx >= m_clusterCount)
+      return &m_eos;  // caller will LoadCluster as desired
+
+    Cluster* const pNext = m_clusters[idx];
+    assert(pNext);
+    assert(pNext->m_index >= 0);
+    assert(pNext->m_index == idx);
+
+    return pNext;
+  }
+
+  assert(m_clusterPreloadCount > 0);
+
+  long long pos = pCurr->m_element_start;
+
+  assert(m_size >= 0);  // TODO
+  const long long stop = m_start + m_size;  // end of segment
+
+  {
+    long len;
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);
+    assert((pos + len) <= stop);  // TODO
+    if (result != 0)
+      return NULL;
+
+    const long long id = ReadID(m_pReader, pos, len);
+    if (id != libwebm::kMkvCluster)
+      return NULL;
+
+    pos += len;  // consume ID
+
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);  // TODO
+    assert((pos + len) <= stop);  // TODO
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+    assert(size > 0);  // TODO
+    // assert((pCurr->m_size <= 0) || (pCurr->m_size == size));
+
+    pos += len;  // consume length of size of element
+    assert((pos + size) <= stop);  // TODO
+
+    // Pos now points to start of payload
+
+    pos += size;  // consume payload
+  }
+
+  long long off_next = 0;
+
+  while (pos < stop) {
+    long len;
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);
+    assert((pos + len) <= stop);  // TODO
+    if (result != 0)
+      return NULL;
+
+    const long long idpos = pos;  // pos of next (potential) cluster
+
+    const long long id = ReadID(m_pReader, idpos, len);
+    if (id < 0)
+      return NULL;
+
+    pos += len;  // consume ID
+
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);  // TODO
+    assert((pos + len) <= stop);  // TODO
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+    assert(size >= 0);  // TODO
+
+    pos += len;  // consume length of size of element
+    assert((pos + size) <= stop);  // TODO
+
+    // Pos now points to start of payload
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvCluster) {
+      const long long off_next_ = idpos - m_start;
+
+      long long pos_;
+      long len_;
+
+      const long status = Cluster::HasBlockEntries(this, off_next_, pos_, len_);
+
+      assert(status >= 0);
+
+      if (status > 0) {
+        off_next = off_next_;
+        break;
+      }
+    }
+
+    pos += size;  // consume payload
+  }
+
+  if (off_next <= 0)
+    return 0;
+
+  Cluster** const ii = m_clusters + m_clusterCount;
+  Cluster** i = ii;
+
+  Cluster** const jj = ii + m_clusterPreloadCount;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) < pos_next
+    //[i, j) ?
+    //[j, jj)  > pos_next
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    Cluster* const pNext = *k;
+    assert(pNext);
+    assert(pNext->m_index < 0);
+
+    // const long long pos_ = pNext->m_pos;
+    // assert(pos_);
+    // pos = pos_ * ((pos_ < 0) ? -1 : 1);
+
+    pos = pNext->GetPosition();
+
+    if (pos < off_next)
+      i = k + 1;
+    else if (pos > off_next)
+      j = k;
+    else
+      return pNext;
+  }
+
+  assert(i == j);
+
+  Cluster* const pNext = Cluster::Create(this, -1, off_next);
+  if (pNext == NULL)
+    return NULL;
+
+  const ptrdiff_t idx_next = i - m_clusters;  // insertion position
+
+  if (!PreloadCluster(pNext, idx_next)) {
+    delete pNext;
+    return NULL;
+  }
+  assert(m_clusters);
+  assert(idx_next < m_clusterSize);
+  assert(m_clusters[idx_next] == pNext);
+
+  return pNext;
+}
+
+long Segment::ParseNext(const Cluster* pCurr, const Cluster*& pResult,
+                        long long& pos, long& len) {
+  assert(pCurr);
+  assert(!pCurr->EOS());
+  assert(m_clusters);
+
+  pResult = 0;
+
+  if (pCurr->m_index >= 0) {  // loaded (not merely preloaded)
+    assert(m_clusters[pCurr->m_index] == pCurr);
+
+    const long next_idx = pCurr->m_index + 1;
+
+    if (next_idx < m_clusterCount) {
+      pResult = m_clusters[next_idx];
+      return 0;  // success
+    }
+
+    // curr cluster is last among loaded
+
+    const long result = LoadCluster(pos, len);
+
+    if (result < 0)  // error or underflow
+      return result;
+
+    if (result > 0)  // no more clusters
+    {
+      // pResult = &m_eos;
+      return 1;
+    }
+
+    pResult = GetLast();
+    return 0;  // success
+  }
+
+  assert(m_pos > 0);
+
+  long long total, avail;
+
+  long status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  // interrogate curr cluster
+
+  pos = pCurr->m_element_start;
+
+  if (pCurr->m_element_size >= 0)
+    pos += pCurr->m_element_size;
+  else {
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadUInt(m_pReader, pos, len);
+
+    if (id != libwebm::kMkvCluster)
+      return -1;
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume size field
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)  // TODO: should never happen
+      return E_FILE_FORMAT_INVALID;  // TODO: resolve this
+
+    // assert((pCurr->m_size <= 0) || (pCurr->m_size == size));
+
+    if ((segment_stop >= 0) && ((pos + size) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // Pos now points to start of payload
+
+    pos += size;  // consume payload (that is, the current cluster)
+    if (segment_stop >= 0 && pos > segment_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    // By consuming the payload, we are assuming that the curr
+    // cluster isn't interesting.  That is, we don't bother checking
+    // whether the payload of the curr cluster is less than what
+    // happens to be available (obtained via IMkvReader::Length).
+    // Presumably the caller has already dispensed with the current
+    // cluster, and really does want the next cluster.
+  }
+
+  // pos now points to just beyond the last fully-loaded cluster
+
+  for (;;) {
+    const long status = DoParseNext(pResult, pos, len);
+
+    if (status <= 1)
+      return status;
+  }
+}
+
+long Segment::DoParseNext(const Cluster*& pResult, long long& pos, long& len) {
+  long long total, avail;
+
+  long status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  // Parse next cluster.  This is strictly a parsing activity.
+  // Creation of a new cluster object happens later, after the
+  // parsing is done.
+
+  long long off_next = 0;
+  long long cluster_size = -1;
+
+  for (;;) {
+    if ((total >= 0) && (pos >= total))
+      return 1;  // EOF
+
+    if ((segment_stop >= 0) && (pos >= segment_stop))
+      return 1;  // EOF
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long idpos = pos;  // absolute
+    const long long idoff = pos - m_start;  // relative
+
+    const long long id = ReadID(m_pReader, idpos, len);  // absolute
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id == 0)  // weird
+      return -1;  // generic error
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume length of size of element
+
+    // Pos now points to start of payload
+
+    if (size == 0)  // weird
+      continue;
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if ((segment_stop >= 0) && (size != unknown_size) &&
+        ((pos + size) > segment_stop)) {
+      return E_FILE_FORMAT_INVALID;
+    }
+
+    if (id == libwebm::kMkvCues) {
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
+
+      const long long element_stop = pos + size;
+
+      if ((segment_stop >= 0) && (element_stop > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      const long long element_start = idpos;
+      const long long element_size = element_stop - element_start;
+
+      if (m_pCues == NULL) {
+        m_pCues = new (std::nothrow)
+            Cues(this, pos, size, element_start, element_size);
+        if (m_pCues == NULL)
+          return false;
+      }
+
+      pos += size;  // consume payload
+      if (segment_stop >= 0 && pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      continue;
+    }
+
+    if (id != libwebm::kMkvCluster) {  // not a Cluster ID
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
+
+      pos += size;  // consume payload
+      if (segment_stop >= 0 && pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      continue;
+    }
+
+    // We have a cluster.
+    off_next = idoff;
+
+    if (size != unknown_size)
+      cluster_size = size;
+
+    break;
+  }
+
+  assert(off_next > 0);  // have cluster
+
+  // We have parsed the next cluster.
+  // We have not created a cluster object yet.  What we need
+  // to do now is determine whether it has already be preloaded
+  //(in which case, an object for this cluster has already been
+  // created), and if not, create a new cluster object.
+
+  Cluster** const ii = m_clusters + m_clusterCount;
+  Cluster** i = ii;
+
+  Cluster** const jj = ii + m_clusterPreloadCount;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) < pos_next
+    //[i, j) ?
+    //[j, jj)  > pos_next
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    const Cluster* const pNext = *k;
+    assert(pNext);
+    assert(pNext->m_index < 0);
+
+    pos = pNext->GetPosition();
+    assert(pos >= 0);
+
+    if (pos < off_next)
+      i = k + 1;
+    else if (pos > off_next)
+      j = k;
+    else {
+      pResult = pNext;
+      return 0;  // success
+    }
+  }
+
+  assert(i == j);
+
+  long long pos_;
+  long len_;
+
+  status = Cluster::HasBlockEntries(this, off_next, pos_, len_);
+
+  if (status < 0) {  // error or underflow
+    pos = pos_;
+    len = len_;
+
+    return status;
+  }
+
+  if (status > 0) {  // means "found at least one block entry"
+    Cluster* const pNext = Cluster::Create(this,
+                                           -1,  // preloaded
+                                           off_next);
+    if (pNext == NULL)
+      return -1;
+
+    const ptrdiff_t idx_next = i - m_clusters;  // insertion position
+
+    if (!PreloadCluster(pNext, idx_next)) {
+      delete pNext;
+      return -1;
+    }
+    assert(m_clusters);
+    assert(idx_next < m_clusterSize);
+    assert(m_clusters[idx_next] == pNext);
+
+    pResult = pNext;
+    return 0;  // success
+  }
+
+  // status == 0 means "no block entries found"
+
+  if (cluster_size < 0) {  // unknown size
+    const long long payload_pos = pos;  // absolute pos of cluster payload
+
+    for (;;) {  // determine cluster size
+      if ((total >= 0) && (pos >= total))
+        break;
+
+      if ((segment_stop >= 0) && (pos >= segment_stop))
+        break;  // no more clusters
+
+      // Read ID
+
+      if ((pos + 1) > avail) {
+        len = 1;
+        return E_BUFFER_NOT_FULL;
+      }
+
+      long long result = GetUIntLength(m_pReader, pos, len);
+
+      if (result < 0)  // error
+        return static_cast<long>(result);
+
+      if (result > 0)  // weird
+        return E_BUFFER_NOT_FULL;
+
+      if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      if ((pos + len) > avail)
+        return E_BUFFER_NOT_FULL;
+
+      const long long idpos = pos;
+      const long long id = ReadID(m_pReader, idpos, len);
+
+      if (id < 0)  // error (or underflow)
+        return static_cast<long>(id);
+
+      // This is the distinguished set of ID's we use to determine
+      // that we have exhausted the sub-element's inside the cluster
+      // whose ID we parsed earlier.
+
+      if (id == libwebm::kMkvCluster || id == libwebm::kMkvCues)
+        break;
+
+      pos += len;  // consume ID (of sub-element)
+
+      // Read Size
+
+      if ((pos + 1) > avail) {
+        len = 1;
+        return E_BUFFER_NOT_FULL;
+      }
+
+      result = GetUIntLength(m_pReader, pos, len);
+
+      if (result < 0)  // error
+        return static_cast<long>(result);
+
+      if (result > 0)  // weird
+        return E_BUFFER_NOT_FULL;
+
+      if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      if ((pos + len) > avail)
+        return E_BUFFER_NOT_FULL;
+
+      const long long size = ReadUInt(m_pReader, pos, len);
+
+      if (size < 0)  // error
+        return static_cast<long>(size);
+
+      pos += len;  // consume size field of element
+
+      // pos now points to start of sub-element's payload
+
+      if (size == 0)  // weird
+        continue;
+
+      const long long unknown_size = (1LL << (7 * len)) - 1;
+
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;  // not allowed for sub-elements
+
+      if ((segment_stop >= 0) && ((pos + size) > segment_stop))  // weird
+        return E_FILE_FORMAT_INVALID;
+
+      pos += size;  // consume payload of sub-element
+      if (segment_stop >= 0 && pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+    }  // determine cluster size
+
+    cluster_size = pos - payload_pos;
+    assert(cluster_size >= 0);  // TODO: handle cluster_size = 0
+
+    pos = payload_pos;  // reset and re-parse original cluster
+  }
+
+  pos += cluster_size;  // consume payload
+  if (segment_stop >= 0 && pos > segment_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 2;  // try to find a cluster that follows next
+}
+
+const Cluster* Segment::FindCluster(long long time_ns) const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
+
+  {
+    Cluster* const pCluster = m_clusters[0];
+    assert(pCluster);
+    assert(pCluster->m_index == 0);
+
+    if (time_ns <= pCluster->GetTime())
+      return pCluster;
+  }
+
+  // Binary search of cluster array
+
+  long i = 0;
+  long j = m_clusterCount;
+
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) <= time_ns
+    //[i, j) ?
+    //[j, m_clusterCount)  > time_ns
+
+    const long k = i + (j - i) / 2;
+    assert(k < m_clusterCount);
+
+    Cluster* const pCluster = m_clusters[k];
+    assert(pCluster);
+    assert(pCluster->m_index == k);
+
+    const long long t = pCluster->GetTime();
+
+    if (t <= time_ns)
+      i = k + 1;
+    else
+      j = k;
+
+    assert(i <= j);
+  }
+
+  assert(i == j);
+  assert(i > 0);
+  assert(i <= m_clusterCount);
+
+  const long k = i - 1;
+
+  Cluster* const pCluster = m_clusters[k];
+  assert(pCluster);
+  assert(pCluster->m_index == k);
+  assert(pCluster->GetTime() <= time_ns);
+
+  return pCluster;
+}
+
+const Tracks* Segment::GetTracks() const { return m_pTracks; }
+const SegmentInfo* Segment::GetInfo() const { return m_pInfo; }
+const Cues* Segment::GetCues() const { return m_pCues; }
+const Chapters* Segment::GetChapters() const { return m_pChapters; }
+const Tags* Segment::GetTags() const { return m_pTags; }
+const SeekHead* Segment::GetSeekHead() const { return m_pSeekHead; }
+
+long long Segment::GetDuration() const {
+  assert(m_pInfo);
+  return m_pInfo->GetDuration();
+}
+
+Chapters::Chapters(Segment* pSegment, long long payload_start,
+                   long long payload_size, long long element_start,
+                   long long element_size)
+    : m_pSegment(pSegment),
+      m_start(payload_start),
+      m_size(payload_size),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_editions(NULL),
+      m_editions_size(0),
+      m_editions_count(0) {}
+
+Chapters::~Chapters() {
+  while (m_editions_count > 0) {
+    Edition& e = m_editions[--m_editions_count];
+    e.Clear();
+  }
+  delete[] m_editions;
+}
+
+long Chapters::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;  // payload start
+  const long long stop = pos + m_size;  // payload stop
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvEditionEntry) {
+      status = ParseEdition(pos, size);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+int Chapters::GetEditionCount() const { return m_editions_count; }
+
+const Chapters::Edition* Chapters::GetEdition(int idx) const {
+  if (idx < 0)
+    return NULL;
+
+  if (idx >= m_editions_count)
+    return NULL;
+
+  return m_editions + idx;
+}
+
+bool Chapters::ExpandEditionsArray() {
+  if (m_editions_size > m_editions_count)
+    return true;  // nothing else to do
+
+  const int size = (m_editions_size == 0) ? 1 : 2 * m_editions_size;
+
+  Edition* const editions = new (std::nothrow) Edition[size];
+
+  if (editions == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_editions_count; ++idx) {
+    m_editions[idx].ShallowCopy(editions[idx]);
+  }
+
+  delete[] m_editions;
+  m_editions = editions;
+
+  m_editions_size = size;
+  return true;
+}
+
+long Chapters::ParseEdition(long long pos, long long size) {
+  if (!ExpandEditionsArray())
+    return -1;
+
+  Edition& e = m_editions[m_editions_count++];
+  e.Init();
+
+  return e.Parse(m_pSegment->m_pReader, pos, size);
+}
+
+Chapters::Edition::Edition() {}
+
+Chapters::Edition::~Edition() {}
+
+int Chapters::Edition::GetAtomCount() const { return m_atoms_count; }
+
+const Chapters::Atom* Chapters::Edition::GetAtom(int index) const {
+  if (index < 0)
+    return NULL;
+
+  if (index >= m_atoms_count)
+    return NULL;
+
+  return m_atoms + index;
+}
+
+void Chapters::Edition::Init() {
+  m_atoms = NULL;
+  m_atoms_size = 0;
+  m_atoms_count = 0;
+}
+
+void Chapters::Edition::ShallowCopy(Edition& rhs) const {
+  rhs.m_atoms = m_atoms;
+  rhs.m_atoms_size = m_atoms_size;
+  rhs.m_atoms_count = m_atoms_count;
+}
+
+void Chapters::Edition::Clear() {
+  while (m_atoms_count > 0) {
+    Atom& a = m_atoms[--m_atoms_count];
+    a.Clear();
+  }
+
+  delete[] m_atoms;
+  m_atoms = NULL;
+
+  m_atoms_size = 0;
+}
+
+long Chapters::Edition::Parse(IMkvReader* pReader, long long pos,
+                              long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)
+      continue;
+
+    if (id == libwebm::kMkvChapterAtom) {
+      status = ParseAtom(pReader, pos, size);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long Chapters::Edition::ParseAtom(IMkvReader* pReader, long long pos,
+                                  long long size) {
+  if (!ExpandAtomsArray())
+    return -1;
+
+  Atom& a = m_atoms[m_atoms_count++];
+  a.Init();
+
+  return a.Parse(pReader, pos, size);
+}
+
+bool Chapters::Edition::ExpandAtomsArray() {
+  if (m_atoms_size > m_atoms_count)
+    return true;  // nothing else to do
+
+  const int size = (m_atoms_size == 0) ? 1 : 2 * m_atoms_size;
+
+  Atom* const atoms = new (std::nothrow) Atom[size];
+
+  if (atoms == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_atoms_count; ++idx) {
+    m_atoms[idx].ShallowCopy(atoms[idx]);
+  }
+
+  delete[] m_atoms;
+  m_atoms = atoms;
+
+  m_atoms_size = size;
+  return true;
+}
+
+Chapters::Atom::Atom() {}
+
+Chapters::Atom::~Atom() {}
+
+unsigned long long Chapters::Atom::GetUID() const { return m_uid; }
+
+const char* Chapters::Atom::GetStringUID() const { return m_string_uid; }
+
+long long Chapters::Atom::GetStartTimecode() const { return m_start_timecode; }
+
+long long Chapters::Atom::GetStopTimecode() const { return m_stop_timecode; }
+
+long long Chapters::Atom::GetStartTime(const Chapters* pChapters) const {
+  return GetTime(pChapters, m_start_timecode);
+}
+
+long long Chapters::Atom::GetStopTime(const Chapters* pChapters) const {
+  return GetTime(pChapters, m_stop_timecode);
+}
+
+int Chapters::Atom::GetDisplayCount() const { return m_displays_count; }
+
+const Chapters::Display* Chapters::Atom::GetDisplay(int index) const {
+  if (index < 0)
+    return NULL;
+
+  if (index >= m_displays_count)
+    return NULL;
+
+  return m_displays + index;
+}
+
+void Chapters::Atom::Init() {
+  m_string_uid = NULL;
+  m_uid = 0;
+  m_start_timecode = -1;
+  m_stop_timecode = -1;
+
+  m_displays = NULL;
+  m_displays_size = 0;
+  m_displays_count = 0;
+}
+
+void Chapters::Atom::ShallowCopy(Atom& rhs) const {
+  rhs.m_string_uid = m_string_uid;
+  rhs.m_uid = m_uid;
+  rhs.m_start_timecode = m_start_timecode;
+  rhs.m_stop_timecode = m_stop_timecode;
+
+  rhs.m_displays = m_displays;
+  rhs.m_displays_size = m_displays_size;
+  rhs.m_displays_count = m_displays_count;
+}
+
+void Chapters::Atom::Clear() {
+  delete[] m_string_uid;
+  m_string_uid = NULL;
+
+  while (m_displays_count > 0) {
+    Display& d = m_displays[--m_displays_count];
+    d.Clear();
+  }
+
+  delete[] m_displays;
+  m_displays = NULL;
+
+  m_displays_size = 0;
+}
+
+long Chapters::Atom::Parse(IMkvReader* pReader, long long pos, long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // 0 length payload, skip.
+      continue;
+
+    if (id == libwebm::kMkvChapterDisplay) {
+      status = ParseDisplay(pReader, pos, size);
+
+      if (status < 0)  // error
+        return status;
+    } else if (id == libwebm::kMkvChapterStringUID) {
+      status = UnserializeString(pReader, pos, size, m_string_uid);
+
+      if (status < 0)  // error
+        return status;
+    } else if (id == libwebm::kMkvChapterUID) {
+      long long val;
+      status = UnserializeInt(pReader, pos, size, val);
+
+      if (status < 0)  // error
+        return status;
+
+      m_uid = static_cast<unsigned long long>(val);
+    } else if (id == libwebm::kMkvChapterTimeStart) {
+      const long long val = UnserializeUInt(pReader, pos, size);
+
+      if (val < 0)  // error
+        return static_cast<long>(val);
+
+      m_start_timecode = val;
+    } else if (id == libwebm::kMkvChapterTimeEnd) {
+      const long long val = UnserializeUInt(pReader, pos, size);
+
+      if (val < 0)  // error
+        return static_cast<long>(val);
+
+      m_stop_timecode = val;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long long Chapters::Atom::GetTime(const Chapters* pChapters,
+                                  long long timecode) {
+  if (pChapters == NULL)
+    return -1;
+
+  Segment* const pSegment = pChapters->m_pSegment;
+
+  if (pSegment == NULL)  // weird
+    return -1;
+
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+
+  if (pInfo == NULL)
+    return -1;
+
+  const long long timecode_scale = pInfo->GetTimeCodeScale();
+
+  if (timecode_scale < 1)  // weird
+    return -1;
+
+  if (timecode < 0)
+    return -1;
+
+  const long long result = timecode_scale * timecode;
+
+  return result;
+}
+
+long Chapters::Atom::ParseDisplay(IMkvReader* pReader, long long pos,
+                                  long long size) {
+  if (!ExpandDisplaysArray())
+    return -1;
+
+  Display& d = m_displays[m_displays_count++];
+  d.Init();
+
+  return d.Parse(pReader, pos, size);
+}
+
+bool Chapters::Atom::ExpandDisplaysArray() {
+  if (m_displays_size > m_displays_count)
+    return true;  // nothing else to do
+
+  const int size = (m_displays_size == 0) ? 1 : 2 * m_displays_size;
+
+  Display* const displays = new (std::nothrow) Display[size];
+
+  if (displays == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_displays_count; ++idx) {
+    m_displays[idx].ShallowCopy(displays[idx]);
+  }
+
+  delete[] m_displays;
+  m_displays = displays;
+
+  m_displays_size = size;
+  return true;
+}
+
+Chapters::Display::Display() {}
+
+Chapters::Display::~Display() {}
+
+const char* Chapters::Display::GetString() const { return m_string; }
+
+const char* Chapters::Display::GetLanguage() const { return m_language; }
+
+const char* Chapters::Display::GetCountry() const { return m_country; }
+
+void Chapters::Display::Init() {
+  m_string = NULL;
+  m_language = NULL;
+  m_country = NULL;
+}
+
+void Chapters::Display::ShallowCopy(Display& rhs) const {
+  rhs.m_string = m_string;
+  rhs.m_language = m_language;
+  rhs.m_country = m_country;
+}
+
+void Chapters::Display::Clear() {
+  delete[] m_string;
+  m_string = NULL;
+
+  delete[] m_language;
+  m_language = NULL;
+
+  delete[] m_country;
+  m_country = NULL;
+}
+
+long Chapters::Display::Parse(IMkvReader* pReader, long long pos,
+                              long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // No payload.
+      continue;
+
+    if (id == libwebm::kMkvChapString) {
+      status = UnserializeString(pReader, pos, size, m_string);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvChapLanguage) {
+      status = UnserializeString(pReader, pos, size, m_language);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvChapCountry) {
+      status = UnserializeString(pReader, pos, size, m_country);
+
+      if (status)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+Tags::Tags(Segment* pSegment, long long payload_start, long long payload_size,
+           long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(payload_start),
+      m_size(payload_size),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_tags(NULL),
+      m_tags_size(0),
+      m_tags_count(0) {}
+
+Tags::~Tags() {
+  while (m_tags_count > 0) {
+    Tag& t = m_tags[--m_tags_count];
+    t.Clear();
+  }
+  delete[] m_tags;
+}
+
+long Tags::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;  // payload start
+  const long long stop = pos + m_size;  // payload stop
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)
+      return status;
+
+    if (size == 0)  // 0 length tag, read another
+      continue;
+
+    if (id == libwebm::kMkvTag) {
+      status = ParseTag(pos, size);
+
+      if (status < 0)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+int Tags::GetTagCount() const { return m_tags_count; }
+
+const Tags::Tag* Tags::GetTag(int idx) const {
+  if (idx < 0)
+    return NULL;
+
+  if (idx >= m_tags_count)
+    return NULL;
+
+  return m_tags + idx;
+}
+
+bool Tags::ExpandTagsArray() {
+  if (m_tags_size > m_tags_count)
+    return true;  // nothing else to do
+
+  const int size = (m_tags_size == 0) ? 1 : 2 * m_tags_size;
+
+  Tag* const tags = new (std::nothrow) Tag[size];
+
+  if (tags == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_tags_count; ++idx) {
+    m_tags[idx].ShallowCopy(tags[idx]);
+  }
+
+  delete[] m_tags;
+  m_tags = tags;
+
+  m_tags_size = size;
+  return true;
+}
+
+long Tags::ParseTag(long long pos, long long size) {
+  if (!ExpandTagsArray())
+    return -1;
+
+  Tag& t = m_tags[m_tags_count++];
+  t.Init();
+
+  return t.Parse(m_pSegment->m_pReader, pos, size);
+}
+
+Tags::Tag::Tag() {}
+
+Tags::Tag::~Tag() {}
+
+int Tags::Tag::GetSimpleTagCount() const { return m_simple_tags_count; }
+
+const Tags::SimpleTag* Tags::Tag::GetSimpleTag(int index) const {
+  if (index < 0)
+    return NULL;
+
+  if (index >= m_simple_tags_count)
+    return NULL;
+
+  return m_simple_tags + index;
+}
+
+void Tags::Tag::Init() {
+  m_simple_tags = NULL;
+  m_simple_tags_size = 0;
+  m_simple_tags_count = 0;
+}
+
+void Tags::Tag::ShallowCopy(Tag& rhs) const {
+  rhs.m_simple_tags = m_simple_tags;
+  rhs.m_simple_tags_size = m_simple_tags_size;
+  rhs.m_simple_tags_count = m_simple_tags_count;
+}
+
+void Tags::Tag::Clear() {
+  while (m_simple_tags_count > 0) {
+    SimpleTag& d = m_simple_tags[--m_simple_tags_count];
+    d.Clear();
+  }
+
+  delete[] m_simple_tags;
+  m_simple_tags = NULL;
+
+  m_simple_tags_size = 0;
+}
+
+long Tags::Tag::Parse(IMkvReader* pReader, long long pos, long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)
+      return status;
+
+    if (size == 0)  // 0 length tag, read another
+      continue;
+
+    if (id == libwebm::kMkvSimpleTag) {
+      status = ParseSimpleTag(pReader, pos, size);
+
+      if (status < 0)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long Tags::Tag::ParseSimpleTag(IMkvReader* pReader, long long pos,
+                               long long size) {
+  if (!ExpandSimpleTagsArray())
+    return -1;
+
+  SimpleTag& st = m_simple_tags[m_simple_tags_count++];
+  st.Init();
+
+  return st.Parse(pReader, pos, size);
+}
+
+bool Tags::Tag::ExpandSimpleTagsArray() {
+  if (m_simple_tags_size > m_simple_tags_count)
+    return true;  // nothing else to do
+
+  const int size = (m_simple_tags_size == 0) ? 1 : 2 * m_simple_tags_size;
+
+  SimpleTag* const displays = new (std::nothrow) SimpleTag[size];
+
+  if (displays == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_simple_tags_count; ++idx) {
+    m_simple_tags[idx].ShallowCopy(displays[idx]);
+  }
+
+  delete[] m_simple_tags;
+  m_simple_tags = displays;
+
+  m_simple_tags_size = size;
+  return true;
+}
+
+Tags::SimpleTag::SimpleTag() {}
+
+Tags::SimpleTag::~SimpleTag() {}
+
+const char* Tags::SimpleTag::GetTagName() const { return m_tag_name; }
+
+const char* Tags::SimpleTag::GetTagString() const { return m_tag_string; }
+
+void Tags::SimpleTag::Init() {
+  m_tag_name = NULL;
+  m_tag_string = NULL;
+}
+
+void Tags::SimpleTag::ShallowCopy(SimpleTag& rhs) const {
+  rhs.m_tag_name = m_tag_name;
+  rhs.m_tag_string = m_tag_string;
+}
+
+void Tags::SimpleTag::Clear() {
+  delete[] m_tag_name;
+  m_tag_name = NULL;
+
+  delete[] m_tag_string;
+  m_tag_string = NULL;
+}
+
+long Tags::SimpleTag::Parse(IMkvReader* pReader, long long pos,
+                            long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvTagName) {
+      status = UnserializeString(pReader, pos, size, m_tag_name);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvTagString) {
+      status = UnserializeString(pReader, pos, size, m_tag_string);
+
+      if (status)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+SegmentInfo::SegmentInfo(Segment* pSegment, long long start, long long size_,
+                         long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_pMuxingAppAsUTF8(NULL),
+      m_pWritingAppAsUTF8(NULL),
+      m_pTitleAsUTF8(NULL) {}
+
+SegmentInfo::~SegmentInfo() {
+  delete[] m_pMuxingAppAsUTF8;
+  m_pMuxingAppAsUTF8 = NULL;
+
+  delete[] m_pWritingAppAsUTF8;
+  m_pWritingAppAsUTF8 = NULL;
+
+  delete[] m_pTitleAsUTF8;
+  m_pTitleAsUTF8 = NULL;
+}
+
+long SegmentInfo::Parse() {
+  assert(m_pMuxingAppAsUTF8 == NULL);
+  assert(m_pWritingAppAsUTF8 == NULL);
+  assert(m_pTitleAsUTF8 == NULL);
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
+
+  m_timecodeScale = 1000000;
+  m_duration = -1;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvTimecodeScale) {
+      m_timecodeScale = UnserializeUInt(pReader, pos, size);
+
+      if (m_timecodeScale <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDuration) {
+      const long status = UnserializeFloat(pReader, pos, size, m_duration);
+
+      if (status < 0)
+        return status;
+
+      if (m_duration < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvMuxingApp) {
+      const long status =
+          UnserializeString(pReader, pos, size, m_pMuxingAppAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvWritingApp) {
+      const long status =
+          UnserializeString(pReader, pos, size, m_pWritingAppAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvTitle) {
+      const long status = UnserializeString(pReader, pos, size, m_pTitleAsUTF8);
+
+      if (status)
+        return status;
+    }
+
+    pos += size;
+
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  const double rollover_check = m_duration * m_timecodeScale;
+  if (rollover_check > LLONG_MAX)
+    return E_FILE_FORMAT_INVALID;
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+long long SegmentInfo::GetTimeCodeScale() const { return m_timecodeScale; }
+
+long long SegmentInfo::GetDuration() const {
+  if (m_duration < 0)
+    return -1;
+
+  assert(m_timecodeScale >= 1);
+
+  const double dd = double(m_duration) * double(m_timecodeScale);
+  const long long d = static_cast<long long>(dd);
+
+  return d;
+}
+
+const char* SegmentInfo::GetMuxingAppAsUTF8() const {
+  return m_pMuxingAppAsUTF8;
+}
+
+const char* SegmentInfo::GetWritingAppAsUTF8() const {
+  return m_pWritingAppAsUTF8;
+}
+
+const char* SegmentInfo::GetTitleAsUTF8() const { return m_pTitleAsUTF8; }
+
+///////////////////////////////////////////////////////////////
+// ContentEncoding element
+ContentEncoding::ContentCompression::ContentCompression()
+    : algo(0), settings(NULL), settings_len(0) {}
+
+ContentEncoding::ContentCompression::~ContentCompression() {
+  delete[] settings;
+}
+
+ContentEncoding::ContentEncryption::ContentEncryption()
+    : algo(0),
+      key_id(NULL),
+      key_id_len(0),
+      signature(NULL),
+      signature_len(0),
+      sig_key_id(NULL),
+      sig_key_id_len(0),
+      sig_algo(0),
+      sig_hash_algo(0) {}
+
+ContentEncoding::ContentEncryption::~ContentEncryption() {
+  delete[] key_id;
+  delete[] signature;
+  delete[] sig_key_id;
+}
+
+ContentEncoding::ContentEncoding()
+    : compression_entries_(NULL),
+      compression_entries_end_(NULL),
+      encryption_entries_(NULL),
+      encryption_entries_end_(NULL),
+      encoding_order_(0),
+      encoding_scope_(1),
+      encoding_type_(0) {}
+
+ContentEncoding::~ContentEncoding() {
+  ContentCompression** comp_i = compression_entries_;
+  ContentCompression** const comp_j = compression_entries_end_;
+
+  while (comp_i != comp_j) {
+    ContentCompression* const comp = *comp_i++;
+    delete comp;
+  }
+
+  delete[] compression_entries_;
+
+  ContentEncryption** enc_i = encryption_entries_;
+  ContentEncryption** const enc_j = encryption_entries_end_;
+
+  while (enc_i != enc_j) {
+    ContentEncryption* const enc = *enc_i++;
+    delete enc;
+  }
+
+  delete[] encryption_entries_;
+}
+
+const ContentEncoding::ContentCompression*
+    ContentEncoding::GetCompressionByIndex(unsigned long idx) const {
+  const ptrdiff_t count = compression_entries_end_ - compression_entries_;
+  assert(count >= 0);
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return compression_entries_[idx];
+}
+
+unsigned long ContentEncoding::GetCompressionCount() const {
+  const ptrdiff_t count = compression_entries_end_ - compression_entries_;
+  assert(count >= 0);
+
+  return static_cast<unsigned long>(count);
+}
+
+const ContentEncoding::ContentEncryption* ContentEncoding::GetEncryptionByIndex(
+    unsigned long idx) const {
+  const ptrdiff_t count = encryption_entries_end_ - encryption_entries_;
+  assert(count >= 0);
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return encryption_entries_[idx];
+}
+
+unsigned long ContentEncoding::GetEncryptionCount() const {
+  const ptrdiff_t count = encryption_entries_end_ - encryption_entries_;
+  assert(count >= 0);
+
+  return static_cast<unsigned long>(count);
+}
+
+long ContentEncoding::ParseContentEncAESSettingsEntry(
+    long long start, long long size, IMkvReader* pReader,
+    ContentEncAESSettings* aes) {
+  assert(pReader);
+  assert(aes);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvAESSettingsCipherMode) {
+      aes->cipher_mode = UnserializeUInt(pReader, pos, size);
+      if (aes->cipher_mode != 1)
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  return 0;
+}
+
+long ContentEncoding::ParseContentEncodingEntry(long long start, long long size,
+                                                IMkvReader* pReader) {
+  assert(pReader);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  // Count ContentCompression and ContentEncryption elements.
+  int compression_count = 0;
+  int encryption_count = 0;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentCompression)
+      ++compression_count;
+
+    if (id == libwebm::kMkvContentEncryption)
+      ++encryption_count;
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (compression_count <= 0 && encryption_count <= 0)
+    return -1;
+
+  if (compression_count > 0) {
+    compression_entries_ =
+        new (std::nothrow) ContentCompression*[compression_count];
+    if (!compression_entries_)
+      return -1;
+    compression_entries_end_ = compression_entries_;
+  }
+
+  if (encryption_count > 0) {
+    encryption_entries_ =
+        new (std::nothrow) ContentEncryption*[encryption_count];
+    if (!encryption_entries_) {
+      delete[] compression_entries_;
+      return -1;
+    }
+    encryption_entries_end_ = encryption_entries_;
+  }
+
+  pos = start;
+  while (pos < stop) {
+    long long id, size;
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentEncodingOrder) {
+      encoding_order_ = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentEncodingScope) {
+      encoding_scope_ = UnserializeUInt(pReader, pos, size);
+      if (encoding_scope_ < 1)
+        return -1;
+    } else if (id == libwebm::kMkvContentEncodingType) {
+      encoding_type_ = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentCompression) {
+      ContentCompression* const compression =
+          new (std::nothrow) ContentCompression();
+      if (!compression)
+        return -1;
+
+      status = ParseCompressionEntry(pos, size, pReader, compression);
+      if (status) {
+        delete compression;
+        return status;
+      }
+      *compression_entries_end_++ = compression;
+    } else if (id == libwebm::kMkvContentEncryption) {
+      ContentEncryption* const encryption =
+          new (std::nothrow) ContentEncryption();
+      if (!encryption)
+        return -1;
+
+      status = ParseEncryptionEntry(pos, size, pReader, encryption);
+      if (status) {
+        delete encryption;
+        return status;
+      }
+      *encryption_entries_end_++ = encryption;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long ContentEncoding::ParseCompressionEntry(long long start, long long size,
+                                            IMkvReader* pReader,
+                                            ContentCompression* compression) {
+  assert(pReader);
+  assert(compression);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  bool valid = false;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentCompAlgo) {
+      long long algo = UnserializeUInt(pReader, pos, size);
+      if (algo < 0)
+        return E_FILE_FORMAT_INVALID;
+      compression->algo = algo;
+      valid = true;
+    } else if (id == libwebm::kMkvContentCompSettings) {
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      compression->settings = buf;
+      compression->settings_len = buflen;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  // ContentCompAlgo is mandatory
+  if (!valid)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+long ContentEncoding::ParseEncryptionEntry(long long start, long long size,
+                                           IMkvReader* pReader,
+                                           ContentEncryption* encryption) {
+  assert(pReader);
+  assert(encryption);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentEncAlgo) {
+      encryption->algo = UnserializeUInt(pReader, pos, size);
+      if (encryption->algo != 5)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvContentEncKeyID) {
+      delete[] encryption->key_id;
+      encryption->key_id = NULL;
+      encryption->key_id_len = 0;
+
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      encryption->key_id = buf;
+      encryption->key_id_len = buflen;
+    } else if (id == libwebm::kMkvContentSignature) {
+      delete[] encryption->signature;
+      encryption->signature = NULL;
+      encryption->signature_len = 0;
+
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      encryption->signature = buf;
+      encryption->signature_len = buflen;
+    } else if (id == libwebm::kMkvContentSigKeyID) {
+      delete[] encryption->sig_key_id;
+      encryption->sig_key_id = NULL;
+      encryption->sig_key_id_len = 0;
+
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      encryption->sig_key_id = buf;
+      encryption->sig_key_id_len = buflen;
+    } else if (id == libwebm::kMkvContentSigAlgo) {
+      encryption->sig_algo = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentSigHashAlgo) {
+      encryption->sig_hash_algo = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentEncAESSettings) {
+      const long status = ParseContentEncAESSettingsEntry(
+          pos, size, pReader, &encryption->aes_settings);
+      if (status)
+        return status;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  return 0;
+}
+
+Track::Track(Segment* pSegment, long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      content_encoding_entries_(NULL),
+      content_encoding_entries_end_(NULL) {}
+
+Track::~Track() {
+  Info& info = const_cast<Info&>(m_info);
+  info.Clear();
+
+  ContentEncoding** i = content_encoding_entries_;
+  ContentEncoding** const j = content_encoding_entries_end_;
+
+  while (i != j) {
+    ContentEncoding* const encoding = *i++;
+    delete encoding;
+  }
+
+  delete[] content_encoding_entries_;
+}
+
+long Track::Create(Segment* pSegment, const Info& info, long long element_start,
+                   long long element_size, Track*& pResult) {
+  if (pResult)
+    return -1;
+
+  Track* const pTrack =
+      new (std::nothrow) Track(pSegment, element_start, element_size);
+
+  if (pTrack == NULL)
+    return -1;  // generic error
+
+  const int status = info.Copy(pTrack->m_info);
+
+  if (status) {  // error
+    delete pTrack;
+    return status;
+  }
+
+  pResult = pTrack;
+  return 0;  // success
+}
+
+Track::Info::Info()
+    : uid(0),
+      defaultDuration(0),
+      codecDelay(0),
+      seekPreRoll(0),
+      nameAsUTF8(NULL),
+      language(NULL),
+      codecId(NULL),
+      codecNameAsUTF8(NULL),
+      codecPrivate(NULL),
+      codecPrivateSize(0),
+      lacing(false) {}
+
+Track::Info::~Info() { Clear(); }
+
+void Track::Info::Clear() {
+  delete[] nameAsUTF8;
+  nameAsUTF8 = NULL;
+
+  delete[] language;
+  language = NULL;
+
+  delete[] codecId;
+  codecId = NULL;
+
+  delete[] codecPrivate;
+  codecPrivate = NULL;
+  codecPrivateSize = 0;
+
+  delete[] codecNameAsUTF8;
+  codecNameAsUTF8 = NULL;
+}
+
+int Track::Info::CopyStr(char* Info::*str, Info& dst_) const {
+  if (str == static_cast<char * Info::*>(NULL))
+    return -1;
+
+  char*& dst = dst_.*str;
+
+  if (dst)  // should be NULL already
+    return -1;
+
+  const char* const src = this->*str;
+
+  if (src == NULL)
+    return 0;
+
+  const size_t len = strlen(src);
+
+  dst = SafeArrayAlloc<char>(1, len + 1);
+
+  if (dst == NULL)
+    return -1;
+
+  strcpy(dst, src);
+
+  return 0;
+}
+
+int Track::Info::Copy(Info& dst) const {
+  if (&dst == this)
+    return 0;
+
+  dst.type = type;
+  dst.number = number;
+  dst.defaultDuration = defaultDuration;
+  dst.codecDelay = codecDelay;
+  dst.seekPreRoll = seekPreRoll;
+  dst.uid = uid;
+  dst.lacing = lacing;
+  dst.settings = settings;
+
+  // We now copy the string member variables from src to dst.
+  // This involves memory allocation so in principle the operation
+  // can fail (indeed, that's why we have Info::Copy), so we must
+  // report this to the caller.  An error return from this function
+  // therefore implies that the copy was only partially successful.
+
+  if (int status = CopyStr(&Info::nameAsUTF8, dst))
+    return status;
+
+  if (int status = CopyStr(&Info::language, dst))
+    return status;
+
+  if (int status = CopyStr(&Info::codecId, dst))
+    return status;
+
+  if (int status = CopyStr(&Info::codecNameAsUTF8, dst))
+    return status;
+
+  if (codecPrivateSize > 0) {
+    if (codecPrivate == NULL)
+      return -1;
+
+    if (dst.codecPrivate)
+      return -1;
+
+    if (dst.codecPrivateSize != 0)
+      return -1;
+
+    dst.codecPrivate = SafeArrayAlloc<unsigned char>(1, codecPrivateSize);
+
+    if (dst.codecPrivate == NULL)
+      return -1;
+
+    memcpy(dst.codecPrivate, codecPrivate, codecPrivateSize);
+    dst.codecPrivateSize = codecPrivateSize;
+  }
+
+  return 0;
+}
+
+const BlockEntry* Track::GetEOS() const { return &m_eos; }
+
+long Track::GetType() const { return m_info.type; }
+
+long Track::GetNumber() const { return m_info.number; }
+
+unsigned long long Track::GetUid() const { return m_info.uid; }
+
+const char* Track::GetNameAsUTF8() const { return m_info.nameAsUTF8; }
+
+const char* Track::GetLanguage() const { return m_info.language; }
+
+const char* Track::GetCodecNameAsUTF8() const { return m_info.codecNameAsUTF8; }
+
+const char* Track::GetCodecId() const { return m_info.codecId; }
+
+const unsigned char* Track::GetCodecPrivate(size_t& size) const {
+  size = m_info.codecPrivateSize;
+  return m_info.codecPrivate;
+}
+
+bool Track::GetLacing() const { return m_info.lacing; }
+
+unsigned long long Track::GetDefaultDuration() const {
+  return m_info.defaultDuration;
+}
+
+unsigned long long Track::GetCodecDelay() const { return m_info.codecDelay; }
+
+unsigned long long Track::GetSeekPreRoll() const { return m_info.seekPreRoll; }
+
+long Track::GetFirst(const BlockEntry*& pBlockEntry) const {
+  const Cluster* pCluster = m_pSegment->GetFirst();
+
+  for (int i = 0;;) {
+    if (pCluster == NULL) {
+      pBlockEntry = GetEOS();
+      return 1;
+    }
+
+    if (pCluster->EOS()) {
+      if (m_pSegment->DoneParsing()) {
+        pBlockEntry = GetEOS();
+        return 1;
+      }
+
+      pBlockEntry = 0;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long status = pCluster->GetFirst(pBlockEntry);
+
+    if (status < 0)  // error
+      return status;
+
+    if (pBlockEntry == 0) {  // empty cluster
+      pCluster = m_pSegment->GetNext(pCluster);
+      continue;
+    }
+
+    for (;;) {
+      const Block* const pBlock = pBlockEntry->GetBlock();
+      assert(pBlock);
+
+      const long long tn = pBlock->GetTrackNumber();
+
+      if ((tn == m_info.number) && VetEntry(pBlockEntry))
+        return 0;
+
+      const BlockEntry* pNextEntry;
+
+      status = pCluster->GetNext(pBlockEntry, pNextEntry);
+
+      if (status < 0)  // error
+        return status;
+
+      if (pNextEntry == 0)
+        break;
+
+      pBlockEntry = pNextEntry;
+    }
+
+    ++i;
+
+    if (i >= 100)
+      break;
+
+    pCluster = m_pSegment->GetNext(pCluster);
+  }
+
+  // NOTE: if we get here, it means that we didn't find a block with
+  // a matching track number.  We interpret that as an error (which
+  // might be too conservative).
+
+  pBlockEntry = GetEOS();  // so we can return a non-NULL value
+  return 1;
+}
+
+long Track::GetNext(const BlockEntry* pCurrEntry,
+                    const BlockEntry*& pNextEntry) const {
+  assert(pCurrEntry);
+  assert(!pCurrEntry->EOS());  //?
+
+  const Block* const pCurrBlock = pCurrEntry->GetBlock();
+  assert(pCurrBlock && pCurrBlock->GetTrackNumber() == m_info.number);
+  if (!pCurrBlock || pCurrBlock->GetTrackNumber() != m_info.number)
+    return -1;
+
+  const Cluster* pCluster = pCurrEntry->GetCluster();
+  assert(pCluster);
+  assert(!pCluster->EOS());
+
+  long status = pCluster->GetNext(pCurrEntry, pNextEntry);
+
+  if (status < 0)  // error
+    return status;
+
+  for (int i = 0;;) {
+    while (pNextEntry) {
+      const Block* const pNextBlock = pNextEntry->GetBlock();
+      assert(pNextBlock);
+
+      if (pNextBlock->GetTrackNumber() == m_info.number)
+        return 0;
+
+      pCurrEntry = pNextEntry;
+
+      status = pCluster->GetNext(pCurrEntry, pNextEntry);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    pCluster = m_pSegment->GetNext(pCluster);
+
+    if (pCluster == NULL) {
+      pNextEntry = GetEOS();
+      return 1;
+    }
+
+    if (pCluster->EOS()) {
+      if (m_pSegment->DoneParsing()) {
+        pNextEntry = GetEOS();
+        return 1;
+      }
+
+      // TODO: there is a potential O(n^2) problem here: we tell the
+      // caller to (pre)load another cluster, which he does, but then he
+      // calls GetNext again, which repeats the same search.  This is
+      // a pathological case, since the only way it can happen is if
+      // there exists a long sequence of clusters none of which contain a
+      // block from this track.  One way around this problem is for the
+      // caller to be smarter when he loads another cluster: don't call
+      // us back until you have a cluster that contains a block from this
+      // track. (Of course, that's not cheap either, since our caller
+      // would have to scan the each cluster as it's loaded, so that
+      // would just push back the problem.)
+
+      pNextEntry = NULL;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    status = pCluster->GetFirst(pNextEntry);
+
+    if (status < 0)  // error
+      return status;
+
+    if (pNextEntry == NULL)  // empty cluster
+      continue;
+
+    ++i;
+
+    if (i >= 100)
+      break;
+  }
+
+  // NOTE: if we get here, it means that we didn't find a block with
+  // a matching track number after lots of searching, so we give
+  // up trying.
+
+  pNextEntry = GetEOS();  // so we can return a non-NULL value
+  return 1;
+}
+
+bool Track::VetEntry(const BlockEntry* pBlockEntry) const {
+  assert(pBlockEntry);
+  const Block* const pBlock = pBlockEntry->GetBlock();
+  assert(pBlock);
+  assert(pBlock->GetTrackNumber() == m_info.number);
+  if (!pBlock || pBlock->GetTrackNumber() != m_info.number)
+    return false;
+
+  // This function is used during a seek to determine whether the
+  // frame is a valid seek target.  This default function simply
+  // returns true, which means all frames are valid seek targets.
+  // It gets overridden by the VideoTrack class, because only video
+  // keyframes can be used as seek target.
+
+  return true;
+}
+
+long Track::Seek(long long time_ns, const BlockEntry*& pResult) const {
+  const long status = GetFirst(pResult);
+
+  if (status < 0)  // buffer underflow, etc
+    return status;
+
+  assert(pResult);
+
+  if (pResult->EOS())
+    return 0;
+
+  const Cluster* pCluster = pResult->GetCluster();
+  assert(pCluster);
+  assert(pCluster->GetIndex() >= 0);
+
+  if (time_ns <= pResult->GetBlock()->GetTime(pCluster))
+    return 0;
+
+  Cluster** const clusters = m_pSegment->m_clusters;
+  assert(clusters);
+
+  const long count = m_pSegment->GetCount();  // loaded only, not preloaded
+  assert(count > 0);
+
+  Cluster** const i = clusters + pCluster->GetIndex();
+  assert(i);
+  assert(*i == pCluster);
+  assert(pCluster->GetTime() <= time_ns);
+
+  Cluster** const j = clusters + count;
+
+  Cluster** lo = i;
+  Cluster** hi = j;
+
+  while (lo < hi) {
+    // INVARIANT:
+    //[i, lo) <= time_ns
+    //[lo, hi) ?
+    //[hi, j)  > time_ns
+
+    Cluster** const mid = lo + (hi - lo) / 2;
+    assert(mid < hi);
+
+    pCluster = *mid;
+    assert(pCluster);
+    assert(pCluster->GetIndex() >= 0);
+    assert(pCluster->GetIndex() == long(mid - m_pSegment->m_clusters));
+
+    const long long t = pCluster->GetTime();
+
+    if (t <= time_ns)
+      lo = mid + 1;
+    else
+      hi = mid;
+
+    assert(lo <= hi);
+  }
+
+  assert(lo == hi);
+  assert(lo > i);
+  assert(lo <= j);
+
+  while (lo > i) {
+    pCluster = *--lo;
+    assert(pCluster);
+    assert(pCluster->GetTime() <= time_ns);
+
+    pResult = pCluster->GetEntry(this);
+
+    if ((pResult != 0) && !pResult->EOS())
+      return 0;
+
+    // landed on empty cluster (no entries)
+  }
+
+  pResult = GetEOS();  // weird
+  return 0;
+}
+
+const ContentEncoding* Track::GetContentEncodingByIndex(
+    unsigned long idx) const {
+  const ptrdiff_t count =
+      content_encoding_entries_end_ - content_encoding_entries_;
+  assert(count >= 0);
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return content_encoding_entries_[idx];
+}
+
+unsigned long Track::GetContentEncodingCount() const {
+  const ptrdiff_t count =
+      content_encoding_entries_end_ - content_encoding_entries_;
+  assert(count >= 0);
+
+  return static_cast<unsigned long>(count);
+}
+
+long Track::ParseContentEncodingsEntry(long long start, long long size) {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+  assert(pReader);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  // Count ContentEncoding elements.
+  int count = 0;
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    // pos now designates start of element
+    if (id == libwebm::kMkvContentEncoding)
+      ++count;
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (count <= 0)
+    return -1;
+
+  content_encoding_entries_ = new (std::nothrow) ContentEncoding*[count];
+  if (!content_encoding_entries_)
+    return -1;
+
+  content_encoding_entries_end_ = content_encoding_entries_;
+
+  pos = start;
+  while (pos < stop) {
+    long long id, size;
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    // pos now designates start of element
+    if (id == libwebm::kMkvContentEncoding) {
+      ContentEncoding* const content_encoding =
+          new (std::nothrow) ContentEncoding();
+      if (!content_encoding)
+        return -1;
+
+      status = content_encoding->ParseContentEncodingEntry(pos, size, pReader);
+      if (status) {
+        delete content_encoding;
+        return status;
+      }
+
+      *content_encoding_entries_end_++ = content_encoding;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+Track::EOSBlock::EOSBlock() : BlockEntry(NULL, LONG_MIN) {}
+
+BlockEntry::Kind Track::EOSBlock::GetKind() const { return kBlockEOS; }
+
+const Block* Track::EOSBlock::GetBlock() const { return NULL; }
+
+bool PrimaryChromaticity::Parse(IMkvReader* reader, long long read_pos,
+                                long long value_size, bool is_x,
+                                PrimaryChromaticity** chromaticity) {
+  if (!reader)
+    return false;
+
+  std::auto_ptr<PrimaryChromaticity> chromaticity_ptr;
+
+  if (!*chromaticity) {
+    chromaticity_ptr.reset(new PrimaryChromaticity());
+  } else {
+    chromaticity_ptr.reset(*chromaticity);
+  }
+
+  if (!chromaticity_ptr.get())
+    return false;
+
+  float* value = is_x ? &chromaticity_ptr->x : &chromaticity_ptr->y;
+
+  double parser_value = 0;
+  const long long value_parse_status =
+      UnserializeFloat(reader, read_pos, value_size, parser_value);
+
+  *value = static_cast<float>(parser_value);
+
+  if (value_parse_status < 0 || *value < 0.0 || *value > 1.0)
+    return false;
+
+  *chromaticity = chromaticity_ptr.release();
+  return true;
+}
+
+bool MasteringMetadata::Parse(IMkvReader* reader, long long mm_start,
+                              long long mm_size, MasteringMetadata** mm) {
+  if (!reader || *mm)
+    return false;
+
+  std::auto_ptr<MasteringMetadata> mm_ptr(new MasteringMetadata());
+  if (!mm_ptr.get())
+    return false;
+
+  const long long mm_end = mm_start + mm_size;
+  long long read_pos = mm_start;
+
+  while (read_pos < mm_end) {
+    long long child_id = 0;
+    long long child_size = 0;
+
+    const long long status =
+        ParseElementHeader(reader, read_pos, mm_end, child_id, child_size);
+    if (status < 0)
+      return false;
+
+    if (child_id == libwebm::kMkvLuminanceMax) {
+      double value = 0;
+      const long long value_parse_status =
+          UnserializeFloat(reader, read_pos, child_size, value);
+      mm_ptr->luminance_max = static_cast<float>(value);
+      if (value_parse_status < 0 || mm_ptr->luminance_max < 0.0 ||
+          mm_ptr->luminance_max > 9999.99) {
+        return false;
+      }
+    } else if (child_id == libwebm::kMkvLuminanceMin) {
+      double value = 0;
+      const long long value_parse_status =
+          UnserializeFloat(reader, read_pos, child_size, value);
+      mm_ptr->luminance_min = static_cast<float>(value);
+      if (value_parse_status < 0 || mm_ptr->luminance_min < 0.0 ||
+          mm_ptr->luminance_min > 999.9999) {
+        return false;
+      }
+    } else {
+      bool is_x = false;
+      PrimaryChromaticity** chromaticity;
+      switch (child_id) {
+        case libwebm::kMkvPrimaryRChromaticityX:
+        case libwebm::kMkvPrimaryRChromaticityY:
+          is_x = child_id == libwebm::kMkvPrimaryRChromaticityX;
+          chromaticity = &mm_ptr->r;
+          break;
+        case libwebm::kMkvPrimaryGChromaticityX:
+        case libwebm::kMkvPrimaryGChromaticityY:
+          is_x = child_id == libwebm::kMkvPrimaryGChromaticityX;
+          chromaticity = &mm_ptr->g;
+          break;
+        case libwebm::kMkvPrimaryBChromaticityX:
+        case libwebm::kMkvPrimaryBChromaticityY:
+          is_x = child_id == libwebm::kMkvPrimaryBChromaticityX;
+          chromaticity = &mm_ptr->b;
+          break;
+        case libwebm::kMkvWhitePointChromaticityX:
+        case libwebm::kMkvWhitePointChromaticityY:
+          is_x = child_id == libwebm::kMkvWhitePointChromaticityX;
+          chromaticity = &mm_ptr->white_point;
+          break;
+        default:
+          return false;
+      }
+      const bool value_parse_status = PrimaryChromaticity::Parse(
+          reader, read_pos, child_size, is_x, chromaticity);
+      if (!value_parse_status)
+        return false;
+    }
+
+    read_pos += child_size;
+    if (read_pos > mm_end)
+      return false;
+  }
+
+  *mm = mm_ptr.release();
+  return true;
+}
+
+bool Colour::Parse(IMkvReader* reader, long long colour_start,
+                   long long colour_size, Colour** colour) {
+  if (!reader || *colour)
+    return false;
+
+  std::auto_ptr<Colour> colour_ptr(new Colour());
+  if (!colour_ptr.get())
+    return false;
+
+  const long long colour_end = colour_start + colour_size;
+  long long read_pos = colour_start;
+
+  while (read_pos < colour_end) {
+    long long child_id = 0;
+    long long child_size = 0;
+
+    const long status =
+        ParseElementHeader(reader, read_pos, colour_end, child_id, child_size);
+    if (status < 0)
+      return false;
+
+    if (child_id == libwebm::kMkvMatrixCoefficients) {
+      colour_ptr->matrix_coefficients =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->matrix_coefficients < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvBitsPerChannel) {
+      colour_ptr->bits_per_channel =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->bits_per_channel < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSubsamplingHorz) {
+      colour_ptr->chroma_subsampling_horz =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_subsampling_horz < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSubsamplingVert) {
+      colour_ptr->chroma_subsampling_vert =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_subsampling_vert < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvCbSubsamplingHorz) {
+      colour_ptr->cb_subsampling_horz =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->cb_subsampling_horz < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvCbSubsamplingVert) {
+      colour_ptr->cb_subsampling_vert =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->cb_subsampling_vert < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSitingHorz) {
+      colour_ptr->chroma_siting_horz =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_siting_horz < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSitingVert) {
+      colour_ptr->chroma_siting_vert =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_siting_vert < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvRange) {
+      colour_ptr->range = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->range < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvTransferCharacteristics) {
+      colour_ptr->transfer_characteristics =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->transfer_characteristics < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvPrimaries) {
+      colour_ptr->primaries = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->primaries < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvMaxCLL) {
+      colour_ptr->max_cll = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->max_cll < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvMaxFALL) {
+      colour_ptr->max_fall = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->max_fall < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvMasteringMetadata) {
+      if (!MasteringMetadata::Parse(reader, read_pos, child_size,
+                                    &colour_ptr->mastering_metadata))
+        return false;
+    } else {
+      return false;
+    }
+
+    read_pos += child_size;
+    if (read_pos > colour_end)
+      return false;
+  }
+  *colour = colour_ptr.release();
+  return true;
+}
+
+VideoTrack::VideoTrack(Segment* pSegment, long long element_start,
+                       long long element_size)
+    : Track(pSegment, element_start, element_size), m_colour(NULL) {}
+
+VideoTrack::~VideoTrack() { delete m_colour; }
+
+long VideoTrack::Parse(Segment* pSegment, const Info& info,
+                       long long element_start, long long element_size,
+                       VideoTrack*& pResult) {
+  if (pResult)
+    return -1;
+
+  if (info.type != Track::kVideo)
+    return -1;
+
+  long long width = 0;
+  long long height = 0;
+  long long display_width = 0;
+  long long display_height = 0;
+  long long display_unit = 0;
+  long long stereo_mode = 0;
+
+  double rate = 0.0;
+
+  IMkvReader* const pReader = pSegment->m_pReader;
+
+  const Settings& s = info.settings;
+  assert(s.start >= 0);
+  assert(s.size >= 0);
+
+  long long pos = s.start;
+  assert(pos >= 0);
+
+  const long long stop = pos + s.size;
+
+  Colour* colour = NULL;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvPixelWidth) {
+      width = UnserializeUInt(pReader, pos, size);
+
+      if (width <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvPixelHeight) {
+      height = UnserializeUInt(pReader, pos, size);
+
+      if (height <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDisplayWidth) {
+      display_width = UnserializeUInt(pReader, pos, size);
+
+      if (display_width <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDisplayHeight) {
+      display_height = UnserializeUInt(pReader, pos, size);
+
+      if (display_height <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDisplayUnit) {
+      display_unit = UnserializeUInt(pReader, pos, size);
+
+      if (display_unit < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvStereoMode) {
+      stereo_mode = UnserializeUInt(pReader, pos, size);
+
+      if (stereo_mode < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvFrameRate) {
+      const long status = UnserializeFloat(pReader, pos, size, rate);
+
+      if (status < 0)
+        return status;
+
+      if (rate <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvColour) {
+      if (!Colour::Parse(pReader, pos, size, &colour))
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  VideoTrack* const pTrack =
+      new (std::nothrow) VideoTrack(pSegment, element_start, element_size);
+
+  if (pTrack == NULL)
+    return -1;  // generic error
+
+  const int status = info.Copy(pTrack->m_info);
+
+  if (status) {  // error
+    delete pTrack;
+    return status;
+  }
+
+  pTrack->m_width = width;
+  pTrack->m_height = height;
+  pTrack->m_display_width = display_width;
+  pTrack->m_display_height = display_height;
+  pTrack->m_display_unit = display_unit;
+  pTrack->m_stereo_mode = stereo_mode;
+  pTrack->m_rate = rate;
+  pTrack->m_colour = colour;
+
+  pResult = pTrack;
+  return 0;  // success
+}
+
+bool VideoTrack::VetEntry(const BlockEntry* pBlockEntry) const {
+  return Track::VetEntry(pBlockEntry) && pBlockEntry->GetBlock()->IsKey();
+}
+
+long VideoTrack::Seek(long long time_ns, const BlockEntry*& pResult) const {
+  const long status = GetFirst(pResult);
+
+  if (status < 0)  // buffer underflow, etc
+    return status;
+
+  assert(pResult);
+
+  if (pResult->EOS())
+    return 0;
+
+  const Cluster* pCluster = pResult->GetCluster();
+  assert(pCluster);
+  assert(pCluster->GetIndex() >= 0);
+
+  if (time_ns <= pResult->GetBlock()->GetTime(pCluster))
+    return 0;
+
+  Cluster** const clusters = m_pSegment->m_clusters;
+  assert(clusters);
+
+  const long count = m_pSegment->GetCount();  // loaded only, not pre-loaded
+  assert(count > 0);
+
+  Cluster** const i = clusters + pCluster->GetIndex();
+  assert(i);
+  assert(*i == pCluster);
+  assert(pCluster->GetTime() <= time_ns);
+
+  Cluster** const j = clusters + count;
+
+  Cluster** lo = i;
+  Cluster** hi = j;
+
+  while (lo < hi) {
+    // INVARIANT:
+    //[i, lo) <= time_ns
+    //[lo, hi) ?
+    //[hi, j)  > time_ns
+
+    Cluster** const mid = lo + (hi - lo) / 2;
+    assert(mid < hi);
+
+    pCluster = *mid;
+    assert(pCluster);
+    assert(pCluster->GetIndex() >= 0);
+    assert(pCluster->GetIndex() == long(mid - m_pSegment->m_clusters));
+
+    const long long t = pCluster->GetTime();
+
+    if (t <= time_ns)
+      lo = mid + 1;
+    else
+      hi = mid;
+
+    assert(lo <= hi);
+  }
+
+  assert(lo == hi);
+  assert(lo > i);
+  assert(lo <= j);
+
+  pCluster = *--lo;
+  assert(pCluster);
+  assert(pCluster->GetTime() <= time_ns);
+
+  pResult = pCluster->GetEntry(this, time_ns);
+
+  if ((pResult != 0) && !pResult->EOS())  // found a keyframe
+    return 0;
+
+  while (lo != i) {
+    pCluster = *--lo;
+    assert(pCluster);
+    assert(pCluster->GetTime() <= time_ns);
+
+    pResult = pCluster->GetEntry(this, time_ns);
+
+    if ((pResult != 0) && !pResult->EOS())
+      return 0;
+  }
+
+  // weird: we're on the first cluster, but no keyframe found
+  // should never happen but we must return something anyway
+
+  pResult = GetEOS();
+  return 0;
+}
+
+Colour* VideoTrack::GetColour() const { return m_colour; }
+
+long long VideoTrack::GetWidth() const { return m_width; }
+
+long long VideoTrack::GetHeight() const { return m_height; }
+
+long long VideoTrack::GetDisplayWidth() const {
+  return m_display_width > 0 ? m_display_width : GetWidth();
+}
+
+long long VideoTrack::GetDisplayHeight() const {
+  return m_display_height > 0 ? m_display_height : GetHeight();
+}
+
+long long VideoTrack::GetDisplayUnit() const { return m_display_unit; }
+
+long long VideoTrack::GetStereoMode() const { return m_stereo_mode; }
+
+double VideoTrack::GetFrameRate() const { return m_rate; }
+
+AudioTrack::AudioTrack(Segment* pSegment, long long element_start,
+                       long long element_size)
+    : Track(pSegment, element_start, element_size) {}
+
+long AudioTrack::Parse(Segment* pSegment, const Info& info,
+                       long long element_start, long long element_size,
+                       AudioTrack*& pResult) {
+  if (pResult)
+    return -1;
+
+  if (info.type != Track::kAudio)
+    return -1;
+
+  IMkvReader* const pReader = pSegment->m_pReader;
+
+  const Settings& s = info.settings;
+  assert(s.start >= 0);
+  assert(s.size >= 0);
+
+  long long pos = s.start;
+  assert(pos >= 0);
+
+  const long long stop = pos + s.size;
+
+  double rate = 8000.0;  // MKV default
+  long long channels = 1;
+  long long bit_depth = 0;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvSamplingFrequency) {
+      status = UnserializeFloat(pReader, pos, size, rate);
+
+      if (status < 0)
+        return status;
+
+      if (rate <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvChannels) {
+      channels = UnserializeUInt(pReader, pos, size);
+
+      if (channels <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvBitDepth) {
+      bit_depth = UnserializeUInt(pReader, pos, size);
+
+      if (bit_depth <= 0)
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  AudioTrack* const pTrack =
+      new (std::nothrow) AudioTrack(pSegment, element_start, element_size);
+
+  if (pTrack == NULL)
+    return -1;  // generic error
+
+  const int status = info.Copy(pTrack->m_info);
+
+  if (status) {
+    delete pTrack;
+    return status;
+  }
+
+  pTrack->m_rate = rate;
+  pTrack->m_channels = channels;
+  pTrack->m_bitDepth = bit_depth;
+
+  pResult = pTrack;
+  return 0;  // success
+}
+
+double AudioTrack::GetSamplingRate() const { return m_rate; }
+
+long long AudioTrack::GetChannels() const { return m_channels; }
+
+long long AudioTrack::GetBitDepth() const { return m_bitDepth; }
+
+Tracks::Tracks(Segment* pSegment, long long start, long long size_,
+               long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_trackEntries(NULL),
+      m_trackEntriesEnd(NULL) {}
+
+long Tracks::Parse() {
+  assert(m_trackEntries == NULL);
+  assert(m_trackEntriesEnd == NULL);
+
+  const long long stop = m_start + m_size;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  int count = 0;
+  long long pos = m_start;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvTrackEntry)
+      ++count;
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (count <= 0)
+    return 0;  // success
+
+  m_trackEntries = new (std::nothrow) Track*[count];
+
+  if (m_trackEntries == NULL)
+    return -1;
+
+  m_trackEntriesEnd = m_trackEntries;
+
+  pos = m_start;
+
+  while (pos < stop) {
+    const long long element_start = pos;
+
+    long long id, payload_size;
+
+    const long status =
+        ParseElementHeader(pReader, pos, stop, id, payload_size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (payload_size == 0)  // weird
+      continue;
+
+    const long long payload_stop = pos + payload_size;
+    assert(payload_stop <= stop);  // checked in ParseElement
+
+    const long long element_size = payload_stop - element_start;
+
+    if (id == libwebm::kMkvTrackEntry) {
+      Track*& pTrack = *m_trackEntriesEnd;
+      pTrack = NULL;
+
+      const long status = ParseTrackEntry(pos, payload_size, element_start,
+                                          element_size, pTrack);
+      if (status)
+        return status;
+
+      if (pTrack)
+        ++m_trackEntriesEnd;
+    }
+
+    pos = payload_stop;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;  // success
+}
+
+unsigned long Tracks::GetTracksCount() const {
+  const ptrdiff_t result = m_trackEntriesEnd - m_trackEntries;
+  assert(result >= 0);
+
+  return static_cast<unsigned long>(result);
+}
+
+long Tracks::ParseTrackEntry(long long track_start, long long track_size,
+                             long long element_start, long long element_size,
+                             Track*& pResult) const {
+  if (pResult)
+    return -1;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = track_start;
+  const long long track_stop = track_start + track_size;
+
+  Track::Info info;
+
+  info.type = 0;
+  info.number = 0;
+  info.uid = 0;
+  info.defaultDuration = 0;
+
+  Track::Settings v;
+  v.start = -1;
+  v.size = -1;
+
+  Track::Settings a;
+  a.start = -1;
+  a.size = -1;
+
+  Track::Settings e;  // content_encodings_settings;
+  e.start = -1;
+  e.size = -1;
+
+  long long lacing = 1;  // default is true
+
+  while (pos < track_stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, track_stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long start = pos;
+
+    if (id == libwebm::kMkvVideo) {
+      v.start = start;
+      v.size = size;
+    } else if (id == libwebm::kMkvAudio) {
+      a.start = start;
+      a.size = size;
+    } else if (id == libwebm::kMkvContentEncodings) {
+      e.start = start;
+      e.size = size;
+    } else if (id == libwebm::kMkvTrackUID) {
+      if (size > 8)
+        return E_FILE_FORMAT_INVALID;
+
+      info.uid = 0;
+
+      long long pos_ = start;
+      const long long pos_end = start + size;
+
+      while (pos_ != pos_end) {
+        unsigned char b;
+
+        const int status = pReader->Read(pos_, 1, &b);
+
+        if (status)
+          return status;
+
+        info.uid <<= 8;
+        info.uid |= b;
+
+        ++pos_;
+      }
+    } else if (id == libwebm::kMkvTrackNumber) {
+      const long long num = UnserializeUInt(pReader, pos, size);
+
+      if ((num <= 0) || (num > 127))
+        return E_FILE_FORMAT_INVALID;
+
+      info.number = static_cast<long>(num);
+    } else if (id == libwebm::kMkvTrackType) {
+      const long long type = UnserializeUInt(pReader, pos, size);
+
+      if ((type <= 0) || (type > 254))
+        return E_FILE_FORMAT_INVALID;
+
+      info.type = static_cast<long>(type);
+    } else if (id == libwebm::kMkvName) {
+      const long status =
+          UnserializeString(pReader, pos, size, info.nameAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvLanguage) {
+      const long status = UnserializeString(pReader, pos, size, info.language);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvDefaultDuration) {
+      const long long duration = UnserializeUInt(pReader, pos, size);
+
+      if (duration < 0)
+        return E_FILE_FORMAT_INVALID;
+
+      info.defaultDuration = static_cast<unsigned long long>(duration);
+    } else if (id == libwebm::kMkvCodecID) {
+      const long status = UnserializeString(pReader, pos, size, info.codecId);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvFlagLacing) {
+      lacing = UnserializeUInt(pReader, pos, size);
+
+      if ((lacing < 0) || (lacing > 1))
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvCodecPrivate) {
+      delete[] info.codecPrivate;
+      info.codecPrivate = NULL;
+      info.codecPrivateSize = 0;
+
+      const size_t buflen = static_cast<size_t>(size);
+
+      if (buflen) {
+        unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+
+        if (buf == NULL)
+          return -1;
+
+        const int status = pReader->Read(pos, static_cast<long>(buflen), buf);
+
+        if (status) {
+          delete[] buf;
+          return status;
+        }
+
+        info.codecPrivate = buf;
+        info.codecPrivateSize = buflen;
+      }
+    } else if (id == libwebm::kMkvCodecName) {
+      const long status =
+          UnserializeString(pReader, pos, size, info.codecNameAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvCodecDelay) {
+      info.codecDelay = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvSeekPreRoll) {
+      info.seekPreRoll = UnserializeUInt(pReader, pos, size);
+    }
+
+    pos += size;  // consume payload
+    if (pos > track_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != track_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (info.number <= 0)  // not specified
+    return E_FILE_FORMAT_INVALID;
+
+  if (GetTrackByNumber(info.number))
+    return E_FILE_FORMAT_INVALID;
+
+  if (info.type <= 0)  // not specified
+    return E_FILE_FORMAT_INVALID;
+
+  info.lacing = (lacing > 0) ? true : false;
+
+  if (info.type == Track::kVideo) {
+    if (v.start < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (a.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    info.settings = v;
+
+    VideoTrack* pTrack = NULL;
+
+    const long status = VideoTrack::Parse(m_pSegment, info, element_start,
+                                          element_size, pTrack);
+
+    if (status)
+      return status;
+
+    pResult = pTrack;
+    assert(pResult);
+
+    if (e.start >= 0)
+      pResult->ParseContentEncodingsEntry(e.start, e.size);
+  } else if (info.type == Track::kAudio) {
+    if (a.start < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (v.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    info.settings = a;
+
+    AudioTrack* pTrack = NULL;
+
+    const long status = AudioTrack::Parse(m_pSegment, info, element_start,
+                                          element_size, pTrack);
+
+    if (status)
+      return status;
+
+    pResult = pTrack;
+    assert(pResult);
+
+    if (e.start >= 0)
+      pResult->ParseContentEncodingsEntry(e.start, e.size);
+  } else {
+    // neither video nor audio - probably metadata or subtitles
+
+    if (a.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (v.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (info.type == Track::kMetadata && e.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    info.settings.start = -1;
+    info.settings.size = 0;
+
+    Track* pTrack = NULL;
+
+    const long status =
+        Track::Create(m_pSegment, info, element_start, element_size, pTrack);
+
+    if (status)
+      return status;
+
+    pResult = pTrack;
+    assert(pResult);
+  }
+
+  return 0;  // success
+}
+
+Tracks::~Tracks() {
+  Track** i = m_trackEntries;
+  Track** const j = m_trackEntriesEnd;
+
+  while (i != j) {
+    Track* const pTrack = *i++;
+    delete pTrack;
+  }
+
+  delete[] m_trackEntries;
+}
+
+const Track* Tracks::GetTrackByNumber(long tn) const {
+  if (tn < 0)
+    return NULL;
+
+  Track** i = m_trackEntries;
+  Track** const j = m_trackEntriesEnd;
+
+  while (i != j) {
+    Track* const pTrack = *i++;
+
+    if (pTrack == NULL)
+      continue;
+
+    if (tn == pTrack->GetNumber())
+      return pTrack;
+  }
+
+  return NULL;  // not found
+}
+
+const Track* Tracks::GetTrackByIndex(unsigned long idx) const {
+  const ptrdiff_t count = m_trackEntriesEnd - m_trackEntries;
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return m_trackEntries[idx];
+}
+
+long Cluster::Load(long long& pos, long& len) const {
+  if (m_pSegment == NULL)
+    return E_PARSE_FAILED;
+
+  if (m_timecode >= 0)  // at least partially loaded
+    return 0;
+
+  if (m_pos != m_element_start || m_element_size >= 0)
+    return E_PARSE_FAILED;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+  long long total, avail;
+  const int status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total >= 0 && (avail > total || m_pos > total))
+    return E_FILE_FORMAT_INVALID;
+
+  pos = m_pos;
+
+  long long cluster_size = -1;
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error or underflow
+    return static_cast<long>(result);
+
+  if (result > 0)
+    return E_BUFFER_NOT_FULL;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long id_ = ReadID(pReader, pos, len);
+
+  if (id_ < 0)  // error
+    return static_cast<long>(id_);
+
+  if (id_ != libwebm::kMkvCluster)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume id
+
+  // read cluster size
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)
+    return E_BUFFER_NOT_FULL;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long size = ReadUInt(pReader, pos, len);
+
+  if (size < 0)  // error
+    return static_cast<long>(cluster_size);
+
+  if (size == 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume length of size of element
+
+  const long long unknown_size = (1LL << (7 * len)) - 1;
+
+  if (size != unknown_size)
+    cluster_size = size;
+
+  // pos points to start of payload
+  long long timecode = -1;
+  long long new_pos = -1;
+  bool bBlock = false;
+
+  long long cluster_stop = (cluster_size < 0) ? -1 : pos + cluster_size;
+
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      break;
+
+    // Parse ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id == 0)
+      return E_FILE_FORMAT_INVALID;
+
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
+
+    if (id == libwebm::kMkvCluster)
+      break;
+
+    if (id == libwebm::kMkvCues)
+      break;
+
+    pos += len;  // consume ID field
+
+    // Parse Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume size field
+
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // pos now points to start of payload
+
+    if (size == 0)
+      continue;
+
+    if ((cluster_stop >= 0) && ((pos + size) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvTimecode) {
+      len = static_cast<long>(size);
+
+      if ((pos + size) > avail)
+        return E_BUFFER_NOT_FULL;
+
+      timecode = UnserializeUInt(pReader, pos, size);
+
+      if (timecode < 0)  // error (or underflow)
+        return static_cast<long>(timecode);
+
+      new_pos = pos + size;
+
+      if (bBlock)
+        break;
+    } else if (id == libwebm::kMkvBlockGroup) {
+      bBlock = true;
+      break;
+    } else if (id == libwebm::kMkvSimpleBlock) {
+      bBlock = true;
+      break;
+    }
+
+    pos += size;  // consume payload
+    if (cluster_stop >= 0 && pos > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (cluster_stop >= 0 && pos > cluster_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (timecode < 0)  // no timecode found
+    return E_FILE_FORMAT_INVALID;
+
+  if (!bBlock)
+    return E_FILE_FORMAT_INVALID;
+
+  m_pos = new_pos;  // designates position just beyond timecode payload
+  m_timecode = timecode;  // m_timecode >= 0 means we're partially loaded
+
+  if (cluster_size >= 0)
+    m_element_size = cluster_stop - m_element_start;
+
+  return 0;
+}
+
+long Cluster::Parse(long long& pos, long& len) const {
+  long status = Load(pos, len);
+
+  if (status < 0)
+    return status;
+
+  if (m_pos < m_element_start || m_timecode < 0)
+    return E_PARSE_FAILED;
+
+  const long long cluster_stop =
+      (m_element_size < 0) ? -1 : m_element_start + m_element_size;
+
+  if ((cluster_stop >= 0) && (m_pos >= cluster_stop))
+    return 1;  // nothing else to do
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long total, avail;
+
+  status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total >= 0 && avail > total)
+    return E_FILE_FORMAT_INVALID;
+
+  pos = m_pos;
+
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      break;
+
+    if ((total >= 0) && (pos >= total)) {
+      if (m_element_size < 0)
+        m_element_size = pos - m_element_start;
+
+      break;
+    }
+
+    // Parse ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
+
+    if ((id == libwebm::kMkvCluster) || (id == libwebm::kMkvCues)) {
+      if (m_element_size < 0)
+        m_element_size = pos - m_element_start;
+
+      break;
+    }
+
+    pos += len;  // consume ID field
+
+    // Parse Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume size field
+
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // pos now points to start of payload
+
+    if (size == 0)
+      continue;
+
+    // const long long block_start = pos;
+    const long long block_stop = pos + size;
+
+    if (cluster_stop >= 0) {
+      if (block_stop > cluster_stop) {
+        if (id == libwebm::kMkvBlockGroup || id == libwebm::kMkvSimpleBlock) {
+          return E_FILE_FORMAT_INVALID;
+        }
+
+        pos = cluster_stop;
+        break;
+      }
+    } else if ((total >= 0) && (block_stop > total)) {
+      m_element_size = total - m_element_start;
+      pos = total;
+      break;
+    } else if (block_stop > avail) {
+      len = static_cast<long>(size);
+      return E_BUFFER_NOT_FULL;
+    }
+
+    Cluster* const this_ = const_cast<Cluster*>(this);
+
+    if (id == libwebm::kMkvBlockGroup)
+      return this_->ParseBlockGroup(size, pos, len);
+
+    if (id == libwebm::kMkvSimpleBlock)
+      return this_->ParseSimpleBlock(size, pos, len);
+
+    pos += size;  // consume payload
+    if (cluster_stop >= 0 && pos > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (m_element_size < 1)
+    return E_FILE_FORMAT_INVALID;
+
+  m_pos = pos;
+  if (cluster_stop >= 0 && m_pos > cluster_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (m_entries_count > 0) {
+    const long idx = m_entries_count - 1;
+
+    const BlockEntry* const pLast = m_entries[idx];
+    if (pLast == NULL)
+      return E_PARSE_FAILED;
+
+    const Block* const pBlock = pLast->GetBlock();
+    if (pBlock == NULL)
+      return E_PARSE_FAILED;
+
+    const long long start = pBlock->m_start;
+
+    if ((total >= 0) && (start > total))
+      return E_PARSE_FAILED;  // defend against trucated stream
+
+    const long long size = pBlock->m_size;
+
+    const long long stop = start + size;
+    if (cluster_stop >= 0 && stop > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && (stop > total))
+      return E_PARSE_FAILED;  // defend against trucated stream
+  }
+
+  return 1;  // no more entries
+}
+
+long Cluster::ParseSimpleBlock(long long block_size, long long& pos,
+                               long& len) {
+  const long long block_start = pos;
+  const long long block_stop = pos + block_size;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long total, avail;
+
+  long status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  // parse track number
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)  // weird
+    return E_BUFFER_NOT_FULL;
+
+  if ((pos + len) > block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long track = ReadUInt(pReader, pos, len);
+
+  if (track < 0)  // error
+    return static_cast<long>(track);
+
+  if (track == 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume track number
+
+  if ((pos + 2) > block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + 2) > avail) {
+    len = 2;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  pos += 2;  // consume timecode
+
+  if ((pos + 1) > block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  unsigned char flags;
+
+  status = pReader->Read(pos, 1, &flags);
+
+  if (status < 0) {  // error or underflow
+    len = 1;
+    return status;
+  }
+
+  ++pos;  // consume flags byte
+  assert(pos <= avail);
+
+  if (pos >= block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  const int lacing = int(flags & 0x06) >> 1;
+
+  if ((lacing != 0) && (block_stop > avail)) {
+    len = static_cast<long>(block_stop - pos);
+    return E_BUFFER_NOT_FULL;
+  }
+
+  status = CreateBlock(libwebm::kMkvSimpleBlock, block_start, block_size,
+                       0);  // DiscardPadding
+
+  if (status != 0)
+    return status;
+
+  m_pos = block_stop;
+
+  return 0;  // success
+}
+
+long Cluster::ParseBlockGroup(long long payload_size, long long& pos,
+                              long& len) {
+  const long long payload_start = pos;
+  const long long payload_stop = pos + payload_size;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long total, avail;
+
+  long status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  if ((total >= 0) && (payload_stop > total))
+    return E_FILE_FORMAT_INVALID;
+
+  if (payload_stop > avail) {
+    len = static_cast<long>(payload_size);
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long discard_padding = 0;
+
+  while (pos < payload_stop) {
+    // parse sub-block element ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((pos + len) > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id == 0)  // not a valid ID
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID field
+
+    // Parse Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((pos + len) > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume size field
+
+    // pos now points to start of sub-block group payload
+
+    if (pos > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if (size == 0)  // weird
+      continue;
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvDiscardPadding) {
+      status = UnserializeInt(pReader, pos, size, discard_padding);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    if (id != libwebm::kMkvBlock) {
+      pos += size;  // consume sub-part of block group
+
+      if (pos > payload_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      continue;
+    }
+
+    const long long block_stop = pos + size;
+
+    if (block_stop > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    // parse track number
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((pos + len) > block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long track = ReadUInt(pReader, pos, len);
+
+    if (track < 0)  // error
+      return static_cast<long>(track);
+
+    if (track == 0)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume track number
+
+    if ((pos + 2) > block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + 2) > avail) {
+      len = 2;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    pos += 2;  // consume timecode
+
+    if ((pos + 1) > block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    unsigned char flags;
+
+    status = pReader->Read(pos, 1, &flags);
+
+    if (status < 0) {  // error or underflow
+      len = 1;
+      return status;
+    }
+
+    ++pos;  // consume flags byte
+    assert(pos <= avail);
+
+    if (pos >= block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    const int lacing = int(flags & 0x06) >> 1;
+
+    if ((lacing != 0) && (block_stop > avail)) {
+      len = static_cast<long>(block_stop - pos);
+      return E_BUFFER_NOT_FULL;
+    }
+
+    pos = block_stop;  // consume block-part of block group
+    if (pos > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != payload_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  status = CreateBlock(libwebm::kMkvBlockGroup, payload_start, payload_size,
+                       discard_padding);
+  if (status != 0)
+    return status;
+
+  m_pos = payload_stop;
+
+  return 0;  // success
+}
+
+long Cluster::GetEntry(long index, const mkvparser::BlockEntry*& pEntry) const {
+  assert(m_pos >= m_element_start);
+
+  pEntry = NULL;
+
+  if (index < 0)
+    return -1;  // generic error
+
+  if (m_entries_count < 0)
+    return E_BUFFER_NOT_FULL;
+
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count <= m_entries_size);
+
+  if (index < m_entries_count) {
+    pEntry = m_entries[index];
+    assert(pEntry);
+
+    return 1;  // found entry
+  }
+
+  if (m_element_size < 0)  // we don't know cluster end yet
+    return E_BUFFER_NOT_FULL;  // underflow
+
+  const long long element_stop = m_element_start + m_element_size;
+
+  if (m_pos >= element_stop)
+    return 0;  // nothing left to parse
+
+  return E_BUFFER_NOT_FULL;  // underflow, since more remains to be parsed
+}
+
+Cluster* Cluster::Create(Segment* pSegment, long idx, long long off) {
+  if (!pSegment || off < 0)
+    return NULL;
+
+  const long long element_start = pSegment->m_start + off;
+
+  Cluster* const pCluster =
+      new (std::nothrow) Cluster(pSegment, idx, element_start);
+
+  return pCluster;
+}
+
+Cluster::Cluster()
+    : m_pSegment(NULL),
+      m_element_start(0),
+      m_index(0),
+      m_pos(0),
+      m_element_size(0),
+      m_timecode(0),
+      m_entries(NULL),
+      m_entries_size(0),
+      m_entries_count(0)  // means "no entries"
+{}
+
+Cluster::Cluster(Segment* pSegment, long idx, long long element_start
+                 /* long long element_size */)
+    : m_pSegment(pSegment),
+      m_element_start(element_start),
+      m_index(idx),
+      m_pos(element_start),
+      m_element_size(-1 /* element_size */),
+      m_timecode(-1),
+      m_entries(NULL),
+      m_entries_size(0),
+      m_entries_count(-1)  // means "has not been parsed yet"
+{}
+
+Cluster::~Cluster() {
+  if (m_entries_count <= 0)
+    return;
+
+  BlockEntry** i = m_entries;
+  BlockEntry** const j = m_entries + m_entries_count;
+
+  while (i != j) {
+    BlockEntry* p = *i++;
+    assert(p);
+
+    delete p;
+  }
+
+  delete[] m_entries;
+}
+
+bool Cluster::EOS() const { return (m_pSegment == NULL); }
+
+long Cluster::GetIndex() const { return m_index; }
+
+long long Cluster::GetPosition() const {
+  const long long pos = m_element_start - m_pSegment->m_start;
+  assert(pos >= 0);
+
+  return pos;
+}
+
+long long Cluster::GetElementSize() const { return m_element_size; }
+
+long Cluster::HasBlockEntries(
+    const Segment* pSegment,
+    long long off,  // relative to start of segment payload
+    long long& pos, long& len) {
+  assert(pSegment);
+  assert(off >= 0);  // relative to segment
+
+  IMkvReader* const pReader = pSegment->m_pReader;
+
+  long long total, avail;
+
+  long status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  pos = pSegment->m_start + off;  // absolute
+
+  if ((total >= 0) && (pos >= total))
+    return 0;  // we don't even have a complete cluster
+
+  const long long segment_stop =
+      (pSegment->m_size < 0) ? -1 : pSegment->m_start + pSegment->m_size;
+
+  long long cluster_stop = -1;  // interpreted later to mean "unknown size"
+
+  {
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // need more data
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return 0;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id != libwebm::kMkvCluster)
+      return E_PARSE_FAILED;
+
+    pos += len;  // consume Cluster ID field
+
+    // read size field
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return 0;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    if (size == 0)
+      return 0;  // cluster does not have entries
+
+    pos += len;  // consume size field
+
+    // pos now points to start of payload
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size != unknown_size) {
+      cluster_stop = pos + size;
+      assert(cluster_stop >= 0);
+
+      if ((segment_stop >= 0) && (cluster_stop > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      if ((total >= 0) && (cluster_stop > total))
+        // return E_FILE_FORMAT_INVALID;  //too conservative
+        return 0;  // cluster does not have any entries
+    }
+  }
+
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      return 0;  // no entries detected
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // need more data
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
+
+    if (id == libwebm::kMkvCluster)
+      return 0;  // no entries found
+
+    if (id == libwebm::kMkvCues)
+      return 0;  // no entries found
+
+    pos += len;  // consume id field
+
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // read size field
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // underflow
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume size field
+
+    // pos now points to start of payload
+
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if (size == 0)  // weird
+      continue;
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;  // not supported inside cluster
+
+    if ((cluster_stop >= 0) && ((pos + size) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvBlockGroup)
+      return 1;  // have at least one entry
+
+    if (id == libwebm::kMkvSimpleBlock)
+      return 1;  // have at least one entry
+
+    pos += size;  // consume payload
+    if (cluster_stop >= 0 && pos > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+}
+
+long long Cluster::GetTimeCode() const {
+  long long pos;
+  long len;
+
+  const long status = Load(pos, len);
+
+  if (status < 0)  // error
+    return status;
+
+  return m_timecode;
+}
+
+long long Cluster::GetTime() const {
+  const long long tc = GetTimeCode();
+
+  if (tc < 0)
+    return tc;
+
+  const SegmentInfo* const pInfo = m_pSegment->GetInfo();
+  assert(pInfo);
+
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
+
+  const long long t = m_timecode * scale;
+
+  return t;
+}
+
+long long Cluster::GetFirstTime() const {
+  const BlockEntry* pEntry;
+
+  const long status = GetFirst(pEntry);
+
+  if (status < 0)  // error
+    return status;
+
+  if (pEntry == NULL)  // empty cluster
+    return GetTime();
+
+  const Block* const pBlock = pEntry->GetBlock();
+  assert(pBlock);
+
+  return pBlock->GetTime(this);
+}
+
+long long Cluster::GetLastTime() const {
+  const BlockEntry* pEntry;
+
+  const long status = GetLast(pEntry);
+
+  if (status < 0)  // error
+    return status;
+
+  if (pEntry == NULL)  // empty cluster
+    return GetTime();
+
+  const Block* const pBlock = pEntry->GetBlock();
+  assert(pBlock);
+
+  return pBlock->GetTime(this);
+}
+
+long Cluster::CreateBlock(long long id,
+                          long long pos,  // absolute pos of payload
+                          long long size, long long discard_padding) {
+  if (id != libwebm::kMkvBlockGroup && id != libwebm::kMkvSimpleBlock)
+    return E_PARSE_FAILED;
+
+  if (m_entries_count < 0) {  // haven't parsed anything yet
+    assert(m_entries == NULL);
+    assert(m_entries_size == 0);
+
+    m_entries_size = 1024;
+    m_entries = new (std::nothrow) BlockEntry*[m_entries_size];
+    if (m_entries == NULL)
+      return -1;
+
+    m_entries_count = 0;
+  } else {
+    assert(m_entries);
+    assert(m_entries_size > 0);
+    assert(m_entries_count <= m_entries_size);
+
+    if (m_entries_count >= m_entries_size) {
+      const long entries_size = 2 * m_entries_size;
+
+      BlockEntry** const entries = new (std::nothrow) BlockEntry*[entries_size];
+      if (entries == NULL)
+        return -1;
+
+      BlockEntry** src = m_entries;
+      BlockEntry** const src_end = src + m_entries_count;
+
+      BlockEntry** dst = entries;
+
+      while (src != src_end)
+        *dst++ = *src++;
+
+      delete[] m_entries;
+
+      m_entries = entries;
+      m_entries_size = entries_size;
+    }
+  }
+
+  if (id == libwebm::kMkvBlockGroup)
+    return CreateBlockGroup(pos, size, discard_padding);
+  else
+    return CreateSimpleBlock(pos, size);
+}
+
+long Cluster::CreateBlockGroup(long long start_offset, long long size,
+                               long long discard_padding) {
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count >= 0);
+  assert(m_entries_count < m_entries_size);
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = start_offset;
+  const long long stop = start_offset + size;
+
+  // For WebM files, there is a bias towards previous reference times
+  //(in order to support alt-ref frames, which refer back to the previous
+  // keyframe).  Normally a 0 value is not possible, but here we tenatively
+  // allow 0 as the value of a reference frame, with the interpretation
+  // that this is a "previous" reference time.
+
+  long long prev = 1;  // nonce
+  long long next = 0;  // nonce
+  long long duration = -1;  // really, this is unsigned
+
+  long long bpos = -1;
+  long long bsize = -1;
+
+  while (pos < stop) {
+    long len;
+    const long long id = ReadID(pReader, pos, len);
+    if (id < 0 || (pos + len) > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);  // TODO
+    assert((pos + len) <= stop);
+
+    pos += len;  // consume size
+
+    if (id == libwebm::kMkvBlock) {
+      if (bpos < 0) {  // Block ID
+        bpos = pos;
+        bsize = size;
+      }
+    } else if (id == libwebm::kMkvBlockDuration) {
+      if (size > 8)
+        return E_FILE_FORMAT_INVALID;
+
+      duration = UnserializeUInt(pReader, pos, size);
+
+      if (duration < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvReferenceBlock) {
+      if (size > 8 || size <= 0)
+        return E_FILE_FORMAT_INVALID;
+      const long size_ = static_cast<long>(size);
+
+      long long time;
+
+      long status = UnserializeInt(pReader, pos, size_, time);
+      assert(status == 0);
+      if (status != 0)
+        return -1;
+
+      if (time <= 0)  // see note above
+        prev = time;
+      else
+        next = time;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+  if (bpos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  assert(bsize >= 0);
+
+  const long idx = m_entries_count;
+
+  BlockEntry** const ppEntry = m_entries + idx;
+  BlockEntry*& pEntry = *ppEntry;
+
+  pEntry = new (std::nothrow)
+      BlockGroup(this, idx, bpos, bsize, prev, next, duration, discard_padding);
+
+  if (pEntry == NULL)
+    return -1;  // generic error
+
+  BlockGroup* const p = static_cast<BlockGroup*>(pEntry);
+
+  const long status = p->Parse();
+
+  if (status == 0) {  // success
+    ++m_entries_count;
+    return 0;
+  }
+
+  delete pEntry;
+  pEntry = 0;
+
+  return status;
+}
+
+long Cluster::CreateSimpleBlock(long long st, long long sz) {
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count >= 0);
+  assert(m_entries_count < m_entries_size);
+
+  const long idx = m_entries_count;
+
+  BlockEntry** const ppEntry = m_entries + idx;
+  BlockEntry*& pEntry = *ppEntry;
+
+  pEntry = new (std::nothrow) SimpleBlock(this, idx, st, sz);
+
+  if (pEntry == NULL)
+    return -1;  // generic error
+
+  SimpleBlock* const p = static_cast<SimpleBlock*>(pEntry);
+
+  const long status = p->Parse();
+
+  if (status == 0) {
+    ++m_entries_count;
+    return 0;
+  }
+
+  delete pEntry;
+  pEntry = 0;
+
+  return status;
+}
+
+long Cluster::GetFirst(const BlockEntry*& pFirst) const {
+  if (m_entries_count <= 0) {
+    long long pos;
+    long len;
+
+    const long status = Parse(pos, len);
+
+    if (status < 0) {  // error
+      pFirst = NULL;
+      return status;
+    }
+
+    if (m_entries_count <= 0) {  // empty cluster
+      pFirst = NULL;
+      return 0;
+    }
+  }
+
+  assert(m_entries);
+
+  pFirst = m_entries[0];
+  assert(pFirst);
+
+  return 0;  // success
+}
+
+long Cluster::GetLast(const BlockEntry*& pLast) const {
+  for (;;) {
+    long long pos;
+    long len;
+
+    const long status = Parse(pos, len);
+
+    if (status < 0) {  // error
+      pLast = NULL;
+      return status;
+    }
+
+    if (status > 0)  // no new block
+      break;
+  }
+
+  if (m_entries_count <= 0) {
+    pLast = NULL;
+    return 0;
+  }
+
+  assert(m_entries);
+
+  const long idx = m_entries_count - 1;
+
+  pLast = m_entries[idx];
+  assert(pLast);
+
+  return 0;
+}
+
+long Cluster::GetNext(const BlockEntry* pCurr, const BlockEntry*& pNext) const {
+  assert(pCurr);
+  assert(m_entries);
+  assert(m_entries_count > 0);
+
+  size_t idx = pCurr->GetIndex();
+  assert(idx < size_t(m_entries_count));
+  assert(m_entries[idx] == pCurr);
+
+  ++idx;
+
+  if (idx >= size_t(m_entries_count)) {
+    long long pos;
+    long len;
+
+    const long status = Parse(pos, len);
+
+    if (status < 0) {  // error
+      pNext = NULL;
+      return status;
+    }
+
+    if (status > 0) {
+      pNext = NULL;
+      return 0;
+    }
+
+    assert(m_entries);
+    assert(m_entries_count > 0);
+    assert(idx < size_t(m_entries_count));
+  }
+
+  pNext = m_entries[idx];
+  assert(pNext);
+
+  return 0;
+}
+
+long Cluster::GetEntryCount() const { return m_entries_count; }
+
+const BlockEntry* Cluster::GetEntry(const Track* pTrack,
+                                    long long time_ns) const {
+  assert(pTrack);
+
+  if (m_pSegment == NULL)  // this is the special EOS cluster
+    return pTrack->GetEOS();
+
+  const BlockEntry* pResult = pTrack->GetEOS();
+
+  long index = 0;
+
+  for (;;) {
+    if (index >= m_entries_count) {
+      long long pos;
+      long len;
+
+      const long status = Parse(pos, len);
+      assert(status >= 0);
+
+      if (status > 0)  // completely parsed, and no more entries
+        return pResult;
+
+      if (status < 0)  // should never happen
+        return 0;
+
+      assert(m_entries);
+      assert(index < m_entries_count);
+    }
+
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
+
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
+
+    if (pBlock->GetTrackNumber() != pTrack->GetNumber()) {
+      ++index;
+      continue;
+    }
+
+    if (pTrack->VetEntry(pEntry)) {
+      if (time_ns < 0)  // just want first candidate block
+        return pEntry;
+
+      const long long ns = pBlock->GetTime(this);
+
+      if (ns > time_ns)
+        return pResult;
+
+      pResult = pEntry;  // have a candidate
+    } else if (time_ns >= 0) {
+      const long long ns = pBlock->GetTime(this);
+
+      if (ns > time_ns)
+        return pResult;
+    }
+
+    ++index;
+  }
+}
+
+const BlockEntry* Cluster::GetEntry(const CuePoint& cp,
+                                    const CuePoint::TrackPosition& tp) const {
+  assert(m_pSegment);
+  const long long tc = cp.GetTimeCode();
+
+  if (tp.m_block > 0) {
+    const long block = static_cast<long>(tp.m_block);
+    const long index = block - 1;
+
+    while (index >= m_entries_count) {
+      long long pos;
+      long len;
+
+      const long status = Parse(pos, len);
+
+      if (status < 0)  // TODO: can this happen?
+        return NULL;
+
+      if (status > 0)  // nothing remains to be parsed
+        return NULL;
+    }
+
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
+
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
+
+    if ((pBlock->GetTrackNumber() == tp.m_track) &&
+        (pBlock->GetTimeCode(this) == tc)) {
+      return pEntry;
+    }
+  }
+
+  long index = 0;
+
+  for (;;) {
+    if (index >= m_entries_count) {
+      long long pos;
+      long len;
+
+      const long status = Parse(pos, len);
+
+      if (status < 0)  // TODO: can this happen?
+        return NULL;
+
+      if (status > 0)  // nothing remains to be parsed
+        return NULL;
+
+      assert(m_entries);
+      assert(index < m_entries_count);
+    }
+
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
+
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
+
+    if (pBlock->GetTrackNumber() != tp.m_track) {
+      ++index;
+      continue;
+    }
+
+    const long long tc_ = pBlock->GetTimeCode(this);
+
+    if (tc_ < tc) {
+      ++index;
+      continue;
+    }
+
+    if (tc_ > tc)
+      return NULL;
+
+    const Tracks* const pTracks = m_pSegment->GetTracks();
+    assert(pTracks);
+
+    const long tn = static_cast<long>(tp.m_track);
+    const Track* const pTrack = pTracks->GetTrackByNumber(tn);
+
+    if (pTrack == NULL)
+      return NULL;
+
+    const long long type = pTrack->GetType();
+
+    if (type == 2)  // audio
+      return pEntry;
+
+    if (type != 1)  // not video
+      return NULL;
+
+    if (!pBlock->IsKey())
+      return NULL;
+
+    return pEntry;
+  }
+}
+
+BlockEntry::BlockEntry(Cluster* p, long idx) : m_pCluster(p), m_index(idx) {}
+BlockEntry::~BlockEntry() {}
+const Cluster* BlockEntry::GetCluster() const { return m_pCluster; }
+long BlockEntry::GetIndex() const { return m_index; }
+
+SimpleBlock::SimpleBlock(Cluster* pCluster, long idx, long long start,
+                         long long size)
+    : BlockEntry(pCluster, idx), m_block(start, size, 0) {}
+
+long SimpleBlock::Parse() { return m_block.Parse(m_pCluster); }
+BlockEntry::Kind SimpleBlock::GetKind() const { return kBlockSimple; }
+const Block* SimpleBlock::GetBlock() const { return &m_block; }
+
+BlockGroup::BlockGroup(Cluster* pCluster, long idx, long long block_start,
+                       long long block_size, long long prev, long long next,
+                       long long duration, long long discard_padding)
+    : BlockEntry(pCluster, idx),
+      m_block(block_start, block_size, discard_padding),
+      m_prev(prev),
+      m_next(next),
+      m_duration(duration) {}
+
+long BlockGroup::Parse() {
+  const long status = m_block.Parse(m_pCluster);
+
+  if (status)
+    return status;
+
+  m_block.SetKey((m_prev > 0) && (m_next <= 0));
+
+  return 0;
+}
+
+BlockEntry::Kind BlockGroup::GetKind() const { return kBlockGroup; }
+const Block* BlockGroup::GetBlock() const { return &m_block; }
+long long BlockGroup::GetPrevTimeCode() const { return m_prev; }
+long long BlockGroup::GetNextTimeCode() const { return m_next; }
+long long BlockGroup::GetDurationTimeCode() const { return m_duration; }
+
+Block::Block(long long start, long long size_, long long discard_padding)
+    : m_start(start),
+      m_size(size_),
+      m_track(0),
+      m_timecode(-1),
+      m_flags(0),
+      m_frames(NULL),
+      m_frame_count(-1),
+      m_discard_padding(discard_padding) {}
+
+Block::~Block() { delete[] m_frames; }
+
+long Block::Parse(const Cluster* pCluster) {
+  if (pCluster == NULL)
+    return -1;
+
+  if (pCluster->m_pSegment == NULL)
+    return -1;
+
+  assert(m_start >= 0);
+  assert(m_size >= 0);
+  assert(m_track <= 0);
+  assert(m_frames == NULL);
+  assert(m_frame_count <= 0);
+
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
+
+  long len;
+
+  IMkvReader* const pReader = pCluster->m_pSegment->m_pReader;
+
+  m_track = ReadUInt(pReader, pos, len);
+
+  if (m_track <= 0)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > stop)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume track number
+
+  if ((stop - pos) < 2)
+    return E_FILE_FORMAT_INVALID;
+
+  long status;
+  long long value;
+
+  status = UnserializeInt(pReader, pos, 2, value);
+
+  if (status)
+    return E_FILE_FORMAT_INVALID;
+
+  if (value < SHRT_MIN)
+    return E_FILE_FORMAT_INVALID;
+
+  if (value > SHRT_MAX)
+    return E_FILE_FORMAT_INVALID;
+
+  m_timecode = static_cast<short>(value);
+
+  pos += 2;
+
+  if ((stop - pos) <= 0)
+    return E_FILE_FORMAT_INVALID;
+
+  status = pReader->Read(pos, 1, &m_flags);
+
+  if (status)
+    return E_FILE_FORMAT_INVALID;
+
+  const int lacing = int(m_flags & 0x06) >> 1;
+
+  ++pos;  // consume flags byte
+
+  if (lacing == 0) {  // no lacing
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    m_frame_count = 1;
+    m_frames = new (std::nothrow) Frame[m_frame_count];
+    if (m_frames == NULL)
+      return -1;
+
+    Frame& f = m_frames[0];
+    f.pos = pos;
+
+    const long long frame_size = stop - pos;
+
+    if (frame_size > LONG_MAX || frame_size <= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    f.len = static_cast<long>(frame_size);
+
+    return 0;  // success
+  }
+
+  if (pos >= stop)
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char biased_count;
+
+  status = pReader->Read(pos, 1, &biased_count);
+
+  if (status)
+    return E_FILE_FORMAT_INVALID;
+
+  ++pos;  // consume frame count
+  if (pos > stop)
+    return E_FILE_FORMAT_INVALID;
+
+  m_frame_count = int(biased_count) + 1;
+
+  m_frames = new (std::nothrow) Frame[m_frame_count];
+  if (m_frames == NULL)
+    return -1;
+
+  if (!m_frames)
+    return E_FILE_FORMAT_INVALID;
+
+  if (lacing == 1) {  // Xiph
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
+
+    long long size = 0;
+    int frame_count = m_frame_count;
+
+    while (frame_count > 1) {
+      long frame_size = 0;
+
+      for (;;) {
+        unsigned char val;
+
+        if (pos >= stop)
+          return E_FILE_FORMAT_INVALID;
+
+        status = pReader->Read(pos, 1, &val);
+
+        if (status)
+          return E_FILE_FORMAT_INVALID;
+
+        ++pos;  // consume xiph size byte
+
+        frame_size += val;
+
+        if (val < 255)
+          break;
+      }
+
+      Frame& f = *pf++;
+      assert(pf < pf_end);
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = 0;  // patch later
+
+      if (frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      f.len = frame_size;
+      size += frame_size;  // contribution of this frame
+
+      --frame_count;
+    }
+
+    if (pf >= pf_end || pos > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    {
+      Frame& f = *pf++;
+
+      if (pf != pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = 0;  // patch later
+
+      const long long total_size = stop - pos;
+
+      if (total_size < size)
+        return E_FILE_FORMAT_INVALID;
+
+      const long long frame_size = total_size - size;
+
+      if (frame_size > LONG_MAX || frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      f.len = static_cast<long>(frame_size);
+    }
+
+    pf = m_frames;
+    while (pf != pf_end) {
+      Frame& f = *pf++;
+      assert((pos + f.len) <= stop);
+
+      if ((pos + f.len) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = pos;
+      pos += f.len;
+    }
+
+    assert(pos == stop);
+    if (pos != stop)
+      return E_FILE_FORMAT_INVALID;
+
+  } else if (lacing == 2) {  // fixed-size lacing
+    if (pos >= stop)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long total_size = stop - pos;
+
+    if ((total_size % m_frame_count) != 0)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long frame_size = total_size / m_frame_count;
+
+    if (frame_size > LONG_MAX || frame_size <= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
+
+    while (pf != pf_end) {
+      assert((pos + frame_size) <= stop);
+      if ((pos + frame_size) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      Frame& f = *pf++;
+
+      f.pos = pos;
+      f.len = static_cast<long>(frame_size);
+
+      pos += frame_size;
+    }
+
+    assert(pos == stop);
+    if (pos != stop)
+      return E_FILE_FORMAT_INVALID;
+
+  } else {
+    assert(lacing == 3);  // EBML lacing
+
+    if (pos >= stop)
+      return E_FILE_FORMAT_INVALID;
+
+    long long size = 0;
+    int frame_count = m_frame_count;
+
+    long long frame_size = ReadUInt(pReader, pos, len);
+
+    if (frame_size <= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (frame_size > LONG_MAX)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume length of size of first frame
+
+    if ((pos + frame_size) > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
+
+    {
+      Frame& curr = *pf;
+
+      curr.pos = 0;  // patch later
+
+      curr.len = static_cast<long>(frame_size);
+      size += curr.len;  // contribution of this frame
+    }
+
+    --frame_count;
+
+    while (frame_count > 1) {
+      if (pos >= stop)
+        return E_FILE_FORMAT_INVALID;
+
+      assert(pf < pf_end);
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      const Frame& prev = *pf++;
+      assert(prev.len == frame_size);
+      if (prev.len != frame_size)
+        return E_FILE_FORMAT_INVALID;
+
+      assert(pf < pf_end);
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      Frame& curr = *pf;
+
+      curr.pos = 0;  // patch later
+
+      const long long delta_size_ = ReadUInt(pReader, pos, len);
+
+      if (delta_size_ < 0)
+        return E_FILE_FORMAT_INVALID;
+
+      if ((pos + len) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      pos += len;  // consume length of (delta) size
+      if (pos > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      const long exp = 7 * len - 1;
+      const long long bias = (1LL << exp) - 1LL;
+      const long long delta_size = delta_size_ - bias;
+
+      frame_size += delta_size;
+
+      if (frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      if (frame_size > LONG_MAX)
+        return E_FILE_FORMAT_INVALID;
+
+      curr.len = static_cast<long>(frame_size);
+      size += curr.len;  // contribution of this frame
+
+      --frame_count;
+    }
+
+    // parse last frame
+    if (frame_count > 0) {
+      if (pos > stop || pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      const Frame& prev = *pf++;
+      assert(prev.len == frame_size);
+      if (prev.len != frame_size)
+        return E_FILE_FORMAT_INVALID;
+
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      Frame& curr = *pf++;
+      if (pf != pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      curr.pos = 0;  // patch later
+
+      const long long total_size = stop - pos;
+
+      if (total_size < size)
+        return E_FILE_FORMAT_INVALID;
+
+      frame_size = total_size - size;
+
+      if (frame_size > LONG_MAX || frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      curr.len = static_cast<long>(frame_size);
+    }
+
+    pf = m_frames;
+    while (pf != pf_end) {
+      Frame& f = *pf++;
+      assert((pos + f.len) <= stop);
+      if ((pos + f.len) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = pos;
+      pos += f.len;
+    }
+
+    if (pos != stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  return 0;  // success
+}
+
+long long Block::GetTimeCode(const Cluster* pCluster) const {
+  if (pCluster == 0)
+    return m_timecode;
+
+  const long long tc0 = pCluster->GetTimeCode();
+  assert(tc0 >= 0);
+
+  const long long tc = tc0 + m_timecode;
+
+  return tc;  // unscaled timecode units
+}
+
+long long Block::GetTime(const Cluster* pCluster) const {
+  assert(pCluster);
+
+  const long long tc = GetTimeCode(pCluster);
+
+  const Segment* const pSegment = pCluster->m_pSegment;
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+  assert(pInfo);
+
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
+
+  const long long ns = tc * scale;
+
+  return ns;
+}
+
+long long Block::GetTrackNumber() const { return m_track; }
+
+bool Block::IsKey() const {
+  return ((m_flags & static_cast<unsigned char>(1 << 7)) != 0);
+}
+
+void Block::SetKey(bool bKey) {
+  if (bKey)
+    m_flags |= static_cast<unsigned char>(1 << 7);
+  else
+    m_flags &= 0x7F;
+}
+
+bool Block::IsInvisible() const { return bool(int(m_flags & 0x08) != 0); }
+
+Block::Lacing Block::GetLacing() const {
+  const int value = int(m_flags & 0x06) >> 1;
+  return static_cast<Lacing>(value);
+}
+
+int Block::GetFrameCount() const { return m_frame_count; }
+
+const Block::Frame& Block::GetFrame(int idx) const {
+  assert(idx >= 0);
+  assert(idx < m_frame_count);
+
+  const Frame& f = m_frames[idx];
+  assert(f.pos > 0);
+  assert(f.len > 0);
+
+  return f;
+}
+
+long Block::Frame::Read(IMkvReader* pReader, unsigned char* buf) const {
+  assert(pReader);
+  assert(buf);
+
+  const long status = pReader->Read(pos, len, buf);
+  return status;
+}
+
+long long Block::GetDiscardPadding() const { return m_discard_padding; }
+
+}  // namespace mkvparser
diff --git a/libvpx/libvpx/third_party/libwebm/mkvparser/mkvparser.h b/libvpx/libvpx/third_party/libwebm/mkvparser/mkvparser.h
new file mode 100644
index 0000000..42e6e88
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvparser/mkvparser.h
@@ -0,0 +1,1112 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef MKVPARSER_MKVPARSER_H_
+#define MKVPARSER_MKVPARSER_H_
+
+#include <cstddef>
+
+namespace mkvparser {
+
+const int E_PARSE_FAILED = -1;
+const int E_FILE_FORMAT_INVALID = -2;
+const int E_BUFFER_NOT_FULL = -3;
+
+class IMkvReader {
+ public:
+  virtual int Read(long long pos, long len, unsigned char* buf) = 0;
+  virtual int Length(long long* total, long long* available) = 0;
+
+ protected:
+  virtual ~IMkvReader();
+};
+
+template <typename Type>
+Type* SafeArrayAlloc(unsigned long long num_elements,
+                     unsigned long long element_size);
+long long GetUIntLength(IMkvReader*, long long, long&);
+long long ReadUInt(IMkvReader*, long long, long&);
+long long ReadID(IMkvReader* pReader, long long pos, long& len);
+long long UnserializeUInt(IMkvReader*, long long pos, long long size);
+
+long UnserializeFloat(IMkvReader*, long long pos, long long size, double&);
+long UnserializeInt(IMkvReader*, long long pos, long long size,
+                    long long& result);
+
+long UnserializeString(IMkvReader*, long long pos, long long size, char*& str);
+
+long ParseElementHeader(IMkvReader* pReader,
+                        long long& pos,  // consume id and size fields
+                        long long stop,  // if you know size of element's parent
+                        long long& id, long long& size);
+
+bool Match(IMkvReader*, long long&, unsigned long, long long&);
+bool Match(IMkvReader*, long long&, unsigned long, unsigned char*&, size_t&);
+
+void GetVersion(int& major, int& minor, int& build, int& revision);
+
+struct EBMLHeader {
+  EBMLHeader();
+  ~EBMLHeader();
+  long long m_version;
+  long long m_readVersion;
+  long long m_maxIdLength;
+  long long m_maxSizeLength;
+  char* m_docType;
+  long long m_docTypeVersion;
+  long long m_docTypeReadVersion;
+
+  long long Parse(IMkvReader*, long long&);
+  void Init();
+};
+
+class Segment;
+class Track;
+class Cluster;
+
+class Block {
+  Block(const Block&);
+  Block& operator=(const Block&);
+
+ public:
+  const long long m_start;
+  const long long m_size;
+
+  Block(long long start, long long size, long long discard_padding);
+  ~Block();
+
+  long Parse(const Cluster*);
+
+  long long GetTrackNumber() const;
+  long long GetTimeCode(const Cluster*) const;  // absolute, but not scaled
+  long long GetTime(const Cluster*) const;  // absolute, and scaled (ns)
+  bool IsKey() const;
+  void SetKey(bool);
+  bool IsInvisible() const;
+
+  enum Lacing { kLacingNone, kLacingXiph, kLacingFixed, kLacingEbml };
+  Lacing GetLacing() const;
+
+  int GetFrameCount() const;  // to index frames: [0, count)
+
+  struct Frame {
+    long long pos;  // absolute offset
+    long len;
+
+    long Read(IMkvReader*, unsigned char*) const;
+  };
+
+  const Frame& GetFrame(int frame_index) const;
+
+  long long GetDiscardPadding() const;
+
+ private:
+  long long m_track;  // Track::Number()
+  short m_timecode;  // relative to cluster
+  unsigned char m_flags;
+
+  Frame* m_frames;
+  int m_frame_count;
+
+ protected:
+  const long long m_discard_padding;
+};
+
+class BlockEntry {
+  BlockEntry(const BlockEntry&);
+  BlockEntry& operator=(const BlockEntry&);
+
+ protected:
+  BlockEntry(Cluster*, long index);
+
+ public:
+  virtual ~BlockEntry();
+
+  bool EOS() const { return (GetKind() == kBlockEOS); }
+  const Cluster* GetCluster() const;
+  long GetIndex() const;
+  virtual const Block* GetBlock() const = 0;
+
+  enum Kind { kBlockEOS, kBlockSimple, kBlockGroup };
+  virtual Kind GetKind() const = 0;
+
+ protected:
+  Cluster* const m_pCluster;
+  const long m_index;
+};
+
+class SimpleBlock : public BlockEntry {
+  SimpleBlock(const SimpleBlock&);
+  SimpleBlock& operator=(const SimpleBlock&);
+
+ public:
+  SimpleBlock(Cluster*, long index, long long start, long long size);
+  long Parse();
+
+  Kind GetKind() const;
+  const Block* GetBlock() const;
+
+ protected:
+  Block m_block;
+};
+
+class BlockGroup : public BlockEntry {
+  BlockGroup(const BlockGroup&);
+  BlockGroup& operator=(const BlockGroup&);
+
+ public:
+  BlockGroup(Cluster*, long index,
+             long long block_start,  // absolute pos of block's payload
+             long long block_size,  // size of block's payload
+             long long prev, long long next, long long duration,
+             long long discard_padding);
+
+  long Parse();
+
+  Kind GetKind() const;
+  const Block* GetBlock() const;
+
+  long long GetPrevTimeCode() const;  // relative to block's time
+  long long GetNextTimeCode() const;  // as above
+  long long GetDurationTimeCode() const;
+
+ private:
+  Block m_block;
+  const long long m_prev;
+  const long long m_next;
+  const long long m_duration;
+};
+
+///////////////////////////////////////////////////////////////
+// ContentEncoding element
+// Elements used to describe if the track data has been encrypted or
+// compressed with zlib or header stripping.
+class ContentEncoding {
+ public:
+  enum { kCTR = 1 };
+
+  ContentEncoding();
+  ~ContentEncoding();
+
+  // ContentCompression element names
+  struct ContentCompression {
+    ContentCompression();
+    ~ContentCompression();
+
+    unsigned long long algo;
+    unsigned char* settings;
+    long long settings_len;
+  };
+
+  // ContentEncAESSettings element names
+  struct ContentEncAESSettings {
+    ContentEncAESSettings() : cipher_mode(kCTR) {}
+    ~ContentEncAESSettings() {}
+
+    unsigned long long cipher_mode;
+  };
+
+  // ContentEncryption element names
+  struct ContentEncryption {
+    ContentEncryption();
+    ~ContentEncryption();
+
+    unsigned long long algo;
+    unsigned char* key_id;
+    long long key_id_len;
+    unsigned char* signature;
+    long long signature_len;
+    unsigned char* sig_key_id;
+    long long sig_key_id_len;
+    unsigned long long sig_algo;
+    unsigned long long sig_hash_algo;
+
+    ContentEncAESSettings aes_settings;
+  };
+
+  // Returns ContentCompression represented by |idx|. Returns NULL if |idx|
+  // is out of bounds.
+  const ContentCompression* GetCompressionByIndex(unsigned long idx) const;
+
+  // Returns number of ContentCompression elements in this ContentEncoding
+  // element.
+  unsigned long GetCompressionCount() const;
+
+  // Parses the ContentCompression element from |pReader|. |start| is the
+  // starting offset of the ContentCompression payload. |size| is the size in
+  // bytes of the ContentCompression payload. |compression| is where the parsed
+  // values will be stored.
+  long ParseCompressionEntry(long long start, long long size,
+                             IMkvReader* pReader,
+                             ContentCompression* compression);
+
+  // Returns ContentEncryption represented by |idx|. Returns NULL if |idx|
+  // is out of bounds.
+  const ContentEncryption* GetEncryptionByIndex(unsigned long idx) const;
+
+  // Returns number of ContentEncryption elements in this ContentEncoding
+  // element.
+  unsigned long GetEncryptionCount() const;
+
+  // Parses the ContentEncAESSettings element from |pReader|. |start| is the
+  // starting offset of the ContentEncAESSettings payload. |size| is the
+  // size in bytes of the ContentEncAESSettings payload. |encryption| is
+  // where the parsed values will be stored.
+  long ParseContentEncAESSettingsEntry(long long start, long long size,
+                                       IMkvReader* pReader,
+                                       ContentEncAESSettings* aes);
+
+  // Parses the ContentEncoding element from |pReader|. |start| is the
+  // starting offset of the ContentEncoding payload. |size| is the size in
+  // bytes of the ContentEncoding payload. Returns true on success.
+  long ParseContentEncodingEntry(long long start, long long size,
+                                 IMkvReader* pReader);
+
+  // Parses the ContentEncryption element from |pReader|. |start| is the
+  // starting offset of the ContentEncryption payload. |size| is the size in
+  // bytes of the ContentEncryption payload. |encryption| is where the parsed
+  // values will be stored.
+  long ParseEncryptionEntry(long long start, long long size,
+                            IMkvReader* pReader, ContentEncryption* encryption);
+
+  unsigned long long encoding_order() const { return encoding_order_; }
+  unsigned long long encoding_scope() const { return encoding_scope_; }
+  unsigned long long encoding_type() const { return encoding_type_; }
+
+ private:
+  // Member variables for list of ContentCompression elements.
+  ContentCompression** compression_entries_;
+  ContentCompression** compression_entries_end_;
+
+  // Member variables for list of ContentEncryption elements.
+  ContentEncryption** encryption_entries_;
+  ContentEncryption** encryption_entries_end_;
+
+  // ContentEncoding element names
+  unsigned long long encoding_order_;
+  unsigned long long encoding_scope_;
+  unsigned long long encoding_type_;
+
+  // LIBWEBM_DISALLOW_COPY_AND_ASSIGN(ContentEncoding);
+  ContentEncoding(const ContentEncoding&);
+  ContentEncoding& operator=(const ContentEncoding&);
+};
+
+class Track {
+  Track(const Track&);
+  Track& operator=(const Track&);
+
+ public:
+  class Info;
+  static long Create(Segment*, const Info&, long long element_start,
+                     long long element_size, Track*&);
+
+  enum Type { kVideo = 1, kAudio = 2, kSubtitle = 0x11, kMetadata = 0x21 };
+
+  Segment* const m_pSegment;
+  const long long m_element_start;
+  const long long m_element_size;
+  virtual ~Track();
+
+  long GetType() const;
+  long GetNumber() const;
+  unsigned long long GetUid() const;
+  const char* GetNameAsUTF8() const;
+  const char* GetLanguage() const;
+  const char* GetCodecNameAsUTF8() const;
+  const char* GetCodecId() const;
+  const unsigned char* GetCodecPrivate(size_t&) const;
+  bool GetLacing() const;
+  unsigned long long GetDefaultDuration() const;
+  unsigned long long GetCodecDelay() const;
+  unsigned long long GetSeekPreRoll() const;
+
+  const BlockEntry* GetEOS() const;
+
+  struct Settings {
+    long long start;
+    long long size;
+  };
+
+  class Info {
+   public:
+    Info();
+    ~Info();
+    int Copy(Info&) const;
+    void Clear();
+    long type;
+    long number;
+    unsigned long long uid;
+    unsigned long long defaultDuration;
+    unsigned long long codecDelay;
+    unsigned long long seekPreRoll;
+    char* nameAsUTF8;
+    char* language;
+    char* codecId;
+    char* codecNameAsUTF8;
+    unsigned char* codecPrivate;
+    size_t codecPrivateSize;
+    bool lacing;
+    Settings settings;
+
+   private:
+    Info(const Info&);
+    Info& operator=(const Info&);
+    int CopyStr(char* Info::*str, Info&) const;
+  };
+
+  long GetFirst(const BlockEntry*&) const;
+  long GetNext(const BlockEntry* pCurr, const BlockEntry*& pNext) const;
+  virtual bool VetEntry(const BlockEntry*) const;
+  virtual long Seek(long long time_ns, const BlockEntry*&) const;
+
+  const ContentEncoding* GetContentEncodingByIndex(unsigned long idx) const;
+  unsigned long GetContentEncodingCount() const;
+
+  long ParseContentEncodingsEntry(long long start, long long size);
+
+ protected:
+  Track(Segment*, long long element_start, long long element_size);
+
+  Info m_info;
+
+  class EOSBlock : public BlockEntry {
+   public:
+    EOSBlock();
+
+    Kind GetKind() const;
+    const Block* GetBlock() const;
+  };
+
+  EOSBlock m_eos;
+
+ private:
+  ContentEncoding** content_encoding_entries_;
+  ContentEncoding** content_encoding_entries_end_;
+};
+
+struct PrimaryChromaticity {
+  PrimaryChromaticity() : x(0), y(0) {}
+  ~PrimaryChromaticity() {}
+  static bool Parse(IMkvReader* reader, long long read_pos,
+                    long long value_size, bool is_x,
+                    PrimaryChromaticity** chromaticity);
+  float x;
+  float y;
+};
+
+struct MasteringMetadata {
+  static const float kValueNotPresent;
+
+  MasteringMetadata()
+      : r(NULL),
+        g(NULL),
+        b(NULL),
+        white_point(NULL),
+        luminance_max(kValueNotPresent),
+        luminance_min(kValueNotPresent) {}
+  ~MasteringMetadata() {
+    delete r;
+    delete g;
+    delete b;
+    delete white_point;
+  }
+
+  static bool Parse(IMkvReader* reader, long long element_start,
+                    long long element_size,
+                    MasteringMetadata** mastering_metadata);
+
+  PrimaryChromaticity* r;
+  PrimaryChromaticity* g;
+  PrimaryChromaticity* b;
+  PrimaryChromaticity* white_point;
+  float luminance_max;
+  float luminance_min;
+};
+
+struct Colour {
+  static const long long kValueNotPresent;
+
+  // Unless otherwise noted all values assigned upon construction are the
+  // equivalent of unspecified/default.
+  Colour()
+      : matrix_coefficients(kValueNotPresent),
+        bits_per_channel(kValueNotPresent),
+        chroma_subsampling_horz(kValueNotPresent),
+        chroma_subsampling_vert(kValueNotPresent),
+        cb_subsampling_horz(kValueNotPresent),
+        cb_subsampling_vert(kValueNotPresent),
+        chroma_siting_horz(kValueNotPresent),
+        chroma_siting_vert(kValueNotPresent),
+        range(kValueNotPresent),
+        transfer_characteristics(kValueNotPresent),
+        primaries(kValueNotPresent),
+        max_cll(kValueNotPresent),
+        max_fall(kValueNotPresent),
+        mastering_metadata(NULL) {}
+  ~Colour() {
+    delete mastering_metadata;
+    mastering_metadata = NULL;
+  }
+
+  static bool Parse(IMkvReader* reader, long long element_start,
+                    long long element_size, Colour** colour);
+
+  long long matrix_coefficients;
+  long long bits_per_channel;
+  long long chroma_subsampling_horz;
+  long long chroma_subsampling_vert;
+  long long cb_subsampling_horz;
+  long long cb_subsampling_vert;
+  long long chroma_siting_horz;
+  long long chroma_siting_vert;
+  long long range;
+  long long transfer_characteristics;
+  long long primaries;
+  long long max_cll;
+  long long max_fall;
+
+  MasteringMetadata* mastering_metadata;
+};
+
+class VideoTrack : public Track {
+  VideoTrack(const VideoTrack&);
+  VideoTrack& operator=(const VideoTrack&);
+
+  VideoTrack(Segment*, long long element_start, long long element_size);
+
+ public:
+  virtual ~VideoTrack();
+  static long Parse(Segment*, const Info&, long long element_start,
+                    long long element_size, VideoTrack*&);
+
+  long long GetWidth() const;
+  long long GetHeight() const;
+  long long GetDisplayWidth() const;
+  long long GetDisplayHeight() const;
+  long long GetDisplayUnit() const;
+  long long GetStereoMode() const;
+  double GetFrameRate() const;
+
+  bool VetEntry(const BlockEntry*) const;
+  long Seek(long long time_ns, const BlockEntry*&) const;
+
+  Colour* GetColour() const;
+
+ private:
+  long long m_width;
+  long long m_height;
+  long long m_display_width;
+  long long m_display_height;
+  long long m_display_unit;
+  long long m_stereo_mode;
+
+  double m_rate;
+
+  Colour* m_colour;
+};
+
+class AudioTrack : public Track {
+  AudioTrack(const AudioTrack&);
+  AudioTrack& operator=(const AudioTrack&);
+
+  AudioTrack(Segment*, long long element_start, long long element_size);
+
+ public:
+  static long Parse(Segment*, const Info&, long long element_start,
+                    long long element_size, AudioTrack*&);
+
+  double GetSamplingRate() const;
+  long long GetChannels() const;
+  long long GetBitDepth() const;
+
+ private:
+  double m_rate;
+  long long m_channels;
+  long long m_bitDepth;
+};
+
+class Tracks {
+  Tracks(const Tracks&);
+  Tracks& operator=(const Tracks&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  Tracks(Segment*, long long start, long long size, long long element_start,
+         long long element_size);
+
+  ~Tracks();
+
+  long Parse();
+
+  unsigned long GetTracksCount() const;
+
+  const Track* GetTrackByNumber(long tn) const;
+  const Track* GetTrackByIndex(unsigned long idx) const;
+
+ private:
+  Track** m_trackEntries;
+  Track** m_trackEntriesEnd;
+
+  long ParseTrackEntry(long long payload_start, long long payload_size,
+                       long long element_start, long long element_size,
+                       Track*&) const;
+};
+
+class Chapters {
+  Chapters(const Chapters&);
+  Chapters& operator=(const Chapters&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  Chapters(Segment*, long long payload_start, long long payload_size,
+           long long element_start, long long element_size);
+
+  ~Chapters();
+
+  long Parse();
+
+  class Atom;
+  class Edition;
+
+  class Display {
+    friend class Atom;
+    Display();
+    Display(const Display&);
+    ~Display();
+    Display& operator=(const Display&);
+
+   public:
+    const char* GetString() const;
+    const char* GetLanguage() const;
+    const char* GetCountry() const;
+
+   private:
+    void Init();
+    void ShallowCopy(Display&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    char* m_string;
+    char* m_language;
+    char* m_country;
+  };
+
+  class Atom {
+    friend class Edition;
+    Atom();
+    Atom(const Atom&);
+    ~Atom();
+    Atom& operator=(const Atom&);
+
+   public:
+    unsigned long long GetUID() const;
+    const char* GetStringUID() const;
+
+    long long GetStartTimecode() const;
+    long long GetStopTimecode() const;
+
+    long long GetStartTime(const Chapters*) const;
+    long long GetStopTime(const Chapters*) const;
+
+    int GetDisplayCount() const;
+    const Display* GetDisplay(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Atom&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+    static long long GetTime(const Chapters*, long long timecode);
+
+    long ParseDisplay(IMkvReader*, long long pos, long long size);
+    bool ExpandDisplaysArray();
+
+    char* m_string_uid;
+    unsigned long long m_uid;
+    long long m_start_timecode;
+    long long m_stop_timecode;
+
+    Display* m_displays;
+    int m_displays_size;
+    int m_displays_count;
+  };
+
+  class Edition {
+    friend class Chapters;
+    Edition();
+    Edition(const Edition&);
+    ~Edition();
+    Edition& operator=(const Edition&);
+
+   public:
+    int GetAtomCount() const;
+    const Atom* GetAtom(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Edition&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    long ParseAtom(IMkvReader*, long long pos, long long size);
+    bool ExpandAtomsArray();
+
+    Atom* m_atoms;
+    int m_atoms_size;
+    int m_atoms_count;
+  };
+
+  int GetEditionCount() const;
+  const Edition* GetEdition(int index) const;
+
+ private:
+  long ParseEdition(long long pos, long long size);
+  bool ExpandEditionsArray();
+
+  Edition* m_editions;
+  int m_editions_size;
+  int m_editions_count;
+};
+
+class Tags {
+  Tags(const Tags&);
+  Tags& operator=(const Tags&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  Tags(Segment*, long long payload_start, long long payload_size,
+       long long element_start, long long element_size);
+
+  ~Tags();
+
+  long Parse();
+
+  class Tag;
+  class SimpleTag;
+
+  class SimpleTag {
+    friend class Tag;
+    SimpleTag();
+    SimpleTag(const SimpleTag&);
+    ~SimpleTag();
+    SimpleTag& operator=(const SimpleTag&);
+
+   public:
+    const char* GetTagName() const;
+    const char* GetTagString() const;
+
+   private:
+    void Init();
+    void ShallowCopy(SimpleTag&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    char* m_tag_name;
+    char* m_tag_string;
+  };
+
+  class Tag {
+    friend class Tags;
+    Tag();
+    Tag(const Tag&);
+    ~Tag();
+    Tag& operator=(const Tag&);
+
+   public:
+    int GetSimpleTagCount() const;
+    const SimpleTag* GetSimpleTag(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Tag&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    long ParseSimpleTag(IMkvReader*, long long pos, long long size);
+    bool ExpandSimpleTagsArray();
+
+    SimpleTag* m_simple_tags;
+    int m_simple_tags_size;
+    int m_simple_tags_count;
+  };
+
+  int GetTagCount() const;
+  const Tag* GetTag(int index) const;
+
+ private:
+  long ParseTag(long long pos, long long size);
+  bool ExpandTagsArray();
+
+  Tag* m_tags;
+  int m_tags_size;
+  int m_tags_count;
+};
+
+class SegmentInfo {
+  SegmentInfo(const SegmentInfo&);
+  SegmentInfo& operator=(const SegmentInfo&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  SegmentInfo(Segment*, long long start, long long size,
+              long long element_start, long long element_size);
+
+  ~SegmentInfo();
+
+  long Parse();
+
+  long long GetTimeCodeScale() const;
+  long long GetDuration() const;  // scaled
+  const char* GetMuxingAppAsUTF8() const;
+  const char* GetWritingAppAsUTF8() const;
+  const char* GetTitleAsUTF8() const;
+
+ private:
+  long long m_timecodeScale;
+  double m_duration;
+  char* m_pMuxingAppAsUTF8;
+  char* m_pWritingAppAsUTF8;
+  char* m_pTitleAsUTF8;
+};
+
+class SeekHead {
+  SeekHead(const SeekHead&);
+  SeekHead& operator=(const SeekHead&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  SeekHead(Segment*, long long start, long long size, long long element_start,
+           long long element_size);
+
+  ~SeekHead();
+
+  long Parse();
+
+  struct Entry {
+    // the SeekHead entry payload
+    long long id;
+    long long pos;
+
+    // absolute pos of SeekEntry ID
+    long long element_start;
+
+    // SeekEntry ID size + size size + payload
+    long long element_size;
+  };
+
+  int GetCount() const;
+  const Entry* GetEntry(int idx) const;
+
+  struct VoidElement {
+    // absolute pos of Void ID
+    long long element_start;
+
+    // ID size + size size + payload size
+    long long element_size;
+  };
+
+  int GetVoidElementCount() const;
+  const VoidElement* GetVoidElement(int idx) const;
+
+ private:
+  Entry* m_entries;
+  int m_entry_count;
+
+  VoidElement* m_void_elements;
+  int m_void_element_count;
+
+  static bool ParseEntry(IMkvReader*,
+                         long long pos,  // payload
+                         long long size, Entry*);
+};
+
+class Cues;
+class CuePoint {
+  friend class Cues;
+
+  CuePoint(long, long long);
+  ~CuePoint();
+
+  CuePoint(const CuePoint&);
+  CuePoint& operator=(const CuePoint&);
+
+ public:
+  long long m_element_start;
+  long long m_element_size;
+
+  bool Load(IMkvReader*);
+
+  long long GetTimeCode() const;  // absolute but unscaled
+  long long GetTime(const Segment*) const;  // absolute and scaled (ns units)
+
+  struct TrackPosition {
+    long long m_track;
+    long long m_pos;  // of cluster
+    long long m_block;
+    // codec_state  //defaults to 0
+    // reference = clusters containing req'd referenced blocks
+    //  reftime = timecode of the referenced block
+
+    bool Parse(IMkvReader*, long long, long long);
+  };
+
+  const TrackPosition* Find(const Track*) const;
+
+ private:
+  const long m_index;
+  long long m_timecode;
+  TrackPosition* m_track_positions;
+  size_t m_track_positions_count;
+};
+
+class Cues {
+  friend class Segment;
+
+  Cues(Segment*, long long start, long long size, long long element_start,
+       long long element_size);
+  ~Cues();
+
+  Cues(const Cues&);
+  Cues& operator=(const Cues&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  bool Find(  // lower bound of time_ns
+      long long time_ns, const Track*, const CuePoint*&,
+      const CuePoint::TrackPosition*&) const;
+
+  const CuePoint* GetFirst() const;
+  const CuePoint* GetLast() const;
+  const CuePoint* GetNext(const CuePoint*) const;
+
+  const BlockEntry* GetBlock(const CuePoint*,
+                             const CuePoint::TrackPosition*) const;
+
+  bool LoadCuePoint() const;
+  long GetCount() const;  // loaded only
+  // long GetTotal() const;  //loaded + preloaded
+  bool DoneParsing() const;
+
+ private:
+  bool Init() const;
+  bool PreloadCuePoint(long&, long long) const;
+
+  mutable CuePoint** m_cue_points;
+  mutable long m_count;
+  mutable long m_preload_count;
+  mutable long long m_pos;
+};
+
+class Cluster {
+  friend class Segment;
+
+  Cluster(const Cluster&);
+  Cluster& operator=(const Cluster&);
+
+ public:
+  Segment* const m_pSegment;
+
+ public:
+  static Cluster* Create(Segment*,
+                         long index,  // index in segment
+                         long long off);  // offset relative to segment
+  // long long element_size);
+
+  Cluster();  // EndOfStream
+  ~Cluster();
+
+  bool EOS() const;
+
+  long long GetTimeCode() const;  // absolute, but not scaled
+  long long GetTime() const;  // absolute, and scaled (nanosecond units)
+  long long GetFirstTime() const;  // time (ns) of first (earliest) block
+  long long GetLastTime() const;  // time (ns) of last (latest) block
+
+  long GetFirst(const BlockEntry*&) const;
+  long GetLast(const BlockEntry*&) const;
+  long GetNext(const BlockEntry* curr, const BlockEntry*& next) const;
+
+  const BlockEntry* GetEntry(const Track*, long long ns = -1) const;
+  const BlockEntry* GetEntry(const CuePoint&,
+                             const CuePoint::TrackPosition&) const;
+  // const BlockEntry* GetMaxKey(const VideoTrack*) const;
+
+  //    static bool HasBlockEntries(const Segment*, long long);
+
+  static long HasBlockEntries(const Segment*, long long idoff, long long& pos,
+                              long& size);
+
+  long GetEntryCount() const;
+
+  long Load(long long& pos, long& size) const;
+
+  long Parse(long long& pos, long& size) const;
+  long GetEntry(long index, const mkvparser::BlockEntry*&) const;
+
+ protected:
+  Cluster(Segment*, long index, long long element_start);
+  // long long element_size);
+
+ public:
+  const long long m_element_start;
+  long long GetPosition() const;  // offset relative to segment
+
+  long GetIndex() const;
+  long long GetElementSize() const;
+  // long long GetPayloadSize() const;
+
+  // long long Unparsed() const;
+
+ private:
+  long m_index;
+  mutable long long m_pos;
+  // mutable long long m_size;
+  mutable long long m_element_size;
+  mutable long long m_timecode;
+  mutable BlockEntry** m_entries;
+  mutable long m_entries_size;
+  mutable long m_entries_count;
+
+  long ParseSimpleBlock(long long, long long&, long&);
+  long ParseBlockGroup(long long, long long&, long&);
+
+  long CreateBlock(long long id, long long pos, long long size,
+                   long long discard_padding);
+  long CreateBlockGroup(long long start_offset, long long size,
+                        long long discard_padding);
+  long CreateSimpleBlock(long long, long long);
+};
+
+class Segment {
+  friend class Cues;
+  friend class Track;
+  friend class VideoTrack;
+
+  Segment(const Segment&);
+  Segment& operator=(const Segment&);
+
+ private:
+  Segment(IMkvReader*, long long elem_start,
+          // long long elem_size,
+          long long pos, long long size);
+
+ public:
+  IMkvReader* const m_pReader;
+  const long long m_element_start;
+  // const long long m_element_size;
+  const long long m_start;  // posn of segment payload
+  const long long m_size;  // size of segment payload
+  Cluster m_eos;  // TODO: make private?
+
+  static long long CreateInstance(IMkvReader*, long long, Segment*&);
+  ~Segment();
+
+  long Load();  // loads headers and all clusters
+
+  // for incremental loading
+  // long long Unparsed() const;
+  bool DoneParsing() const;
+  long long ParseHeaders();  // stops when first cluster is found
+  // long FindNextCluster(long long& pos, long& size) const;
+  long LoadCluster(long long& pos, long& size);  // load one cluster
+  long LoadCluster();
+
+  long ParseNext(const Cluster* pCurr, const Cluster*& pNext, long long& pos,
+                 long& size);
+
+  const SeekHead* GetSeekHead() const;
+  const Tracks* GetTracks() const;
+  const SegmentInfo* GetInfo() const;
+  const Cues* GetCues() const;
+  const Chapters* GetChapters() const;
+  const Tags* GetTags() const;
+
+  long long GetDuration() const;
+
+  unsigned long GetCount() const;
+  const Cluster* GetFirst() const;
+  const Cluster* GetLast() const;
+  const Cluster* GetNext(const Cluster*);
+
+  const Cluster* FindCluster(long long time_nanoseconds) const;
+  // const BlockEntry* Seek(long long time_nanoseconds, const Track*) const;
+
+  const Cluster* FindOrPreloadCluster(long long pos);
+
+  long ParseCues(long long cues_off,  // offset relative to start of segment
+                 long long& parse_pos, long& parse_len);
+
+ private:
+  long long m_pos;  // absolute file posn; what has been consumed so far
+  Cluster* m_pUnknownSize;
+
+  SeekHead* m_pSeekHead;
+  SegmentInfo* m_pInfo;
+  Tracks* m_pTracks;
+  Cues* m_pCues;
+  Chapters* m_pChapters;
+  Tags* m_pTags;
+  Cluster** m_clusters;
+  long m_clusterCount;  // number of entries for which m_index >= 0
+  long m_clusterPreloadCount;  // number of entries for which m_index < 0
+  long m_clusterSize;  // array size
+
+  long DoLoadCluster(long long&, long&);
+  long DoLoadClusterUnknownSize(long long&, long&);
+  long DoParseNext(const Cluster*&, long long&, long&);
+
+  bool AppendCluster(Cluster*);
+  bool PreloadCluster(Cluster*, ptrdiff_t);
+
+  // void ParseSeekHead(long long pos, long long size);
+  // void ParseSeekEntry(long long pos, long long size);
+  // void ParseCues(long long);
+
+  const BlockEntry* GetBlock(const CuePoint&, const CuePoint::TrackPosition&);
+};
+
+}  // namespace mkvparser
+
+inline long mkvparser::Segment::LoadCluster() {
+  long long pos;
+  long size;
+
+  return LoadCluster(pos, size);
+}
+
+#endif  // MKVPARSER_MKVPARSER_H_
diff --git a/libvpx/libvpx/third_party/libwebm/mkvparser/mkvreader.cc b/libvpx/libvpx/third_party/libwebm/mkvparser/mkvreader.cc
new file mode 100644
index 0000000..9f90d8c
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvparser/mkvreader.cc
@@ -0,0 +1,131 @@
+// Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#include "mkvparser/mkvreader.h"
+
+#include <cassert>
+
+namespace mkvparser {
+
+MkvReader::MkvReader() : m_file(NULL), reader_owns_file_(true) {}
+
+MkvReader::MkvReader(FILE* fp) : m_file(fp), reader_owns_file_(false) {
+  GetFileSize();
+}
+
+MkvReader::~MkvReader() {
+  if (reader_owns_file_)
+    Close();
+  m_file = NULL;
+}
+
+int MkvReader::Open(const char* fileName) {
+  if (fileName == NULL)
+    return -1;
+
+  if (m_file)
+    return -1;
+
+#ifdef _MSC_VER
+  const errno_t e = fopen_s(&m_file, fileName, "rb");
+
+  if (e)
+    return -1;  // error
+#else
+  m_file = fopen(fileName, "rb");
+
+  if (m_file == NULL)
+    return -1;
+#endif
+  return !GetFileSize();
+}
+
+bool MkvReader::GetFileSize() {
+  if (m_file == NULL)
+    return false;
+#ifdef _MSC_VER
+  int status = _fseeki64(m_file, 0L, SEEK_END);
+
+  if (status)
+    return false;  // error
+
+  m_length = _ftelli64(m_file);
+#else
+  fseek(m_file, 0L, SEEK_END);
+  m_length = ftell(m_file);
+#endif
+  assert(m_length >= 0);
+
+  if (m_length < 0)
+    return false;
+
+#ifdef _MSC_VER
+  status = _fseeki64(m_file, 0L, SEEK_SET);
+
+  if (status)
+    return false;  // error
+#else
+  fseek(m_file, 0L, SEEK_SET);
+#endif
+
+  return true;
+}
+
+void MkvReader::Close() {
+  if (m_file != NULL) {
+    fclose(m_file);
+    m_file = NULL;
+  }
+}
+
+int MkvReader::Length(long long* total, long long* available) {
+  if (m_file == NULL)
+    return -1;
+
+  if (total)
+    *total = m_length;
+
+  if (available)
+    *available = m_length;
+
+  return 0;
+}
+
+int MkvReader::Read(long long offset, long len, unsigned char* buffer) {
+  if (m_file == NULL)
+    return -1;
+
+  if (offset < 0)
+    return -1;
+
+  if (len < 0)
+    return -1;
+
+  if (len == 0)
+    return 0;
+
+  if (offset >= m_length)
+    return -1;
+
+#ifdef _MSC_VER
+  const int status = _fseeki64(m_file, offset, SEEK_SET);
+
+  if (status)
+    return -1;  // error
+#else
+  fseek(m_file, offset, SEEK_SET);
+#endif
+
+  const size_t size = fread(buffer, 1, len, m_file);
+
+  if (size < size_t(len))
+    return -1;  // error
+
+  return 0;  // success
+}
+
+}  // namespace mkvparser
\ No newline at end of file
diff --git a/libvpx/libvpx/third_party/libwebm/mkvparser/mkvreader.h b/libvpx/libvpx/third_party/libwebm/mkvparser/mkvreader.h
new file mode 100644
index 0000000..9831ecf
--- /dev/null
+++ b/libvpx/libvpx/third_party/libwebm/mkvparser/mkvreader.h
@@ -0,0 +1,45 @@
+// Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef MKVPARSER_MKVREADER_H_
+#define MKVPARSER_MKVREADER_H_
+
+#include <cstdio>
+
+#include "mkvparser/mkvparser.h"
+
+namespace mkvparser {
+
+class MkvReader : public IMkvReader {
+ public:
+  MkvReader();
+  explicit MkvReader(FILE* fp);
+  virtual ~MkvReader();
+
+  int Open(const char*);
+  void Close();
+
+  virtual int Read(long long position, long length, unsigned char* buffer);
+  virtual int Length(long long* total, long long* available);
+
+ private:
+  MkvReader(const MkvReader&);
+  MkvReader& operator=(const MkvReader&);
+
+  // Determines the size of the file. This is called either by the constructor
+  // or by the Open function depending on file ownership. Returns true on
+  // success.
+  bool GetFileSize();
+
+  long long m_length;
+  FILE* m_file;
+  bool reader_owns_file_;
+};
+
+}  // namespace mkvparser
+
+#endif  // MKVPARSER_MKVREADER_H_
diff --git a/libvpx/libvpx/third_party/libyuv/README.libvpx b/libvpx/libvpx/third_party/libyuv/README.libvpx
new file mode 100644
index 0000000..09693c1
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/README.libvpx
@@ -0,0 +1,15 @@
+Name: libyuv
+URL: http://code.google.com/p/libyuv/
+Version: 1456
+License: BSD
+License File: LICENSE
+
+Description:
+libyuv is an open source project that includes YUV conversion and scaling
+functionality.
+
+The optimized scaler in libyuv is used in multiple resolution encoder example,
+which down-samples the original input video (f.g. 1280x720) a number of times
+in order to encode multiple resolution bit streams.
+
+Local Modifications:
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/basic_types.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/basic_types.h
new file mode 100644
index 0000000..beb750b
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/basic_types.h
@@ -0,0 +1,118 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_BASIC_TYPES_H_  // NOLINT
+#define INCLUDE_LIBYUV_BASIC_TYPES_H_
+
+#include <stddef.h>  // for NULL, size_t
+
+#if defined(__ANDROID__) || (defined(_MSC_VER) && (_MSC_VER < 1600))
+#include <sys/types.h>  // for uintptr_t on x86
+#else
+#include <stdint.h>  // for uintptr_t
+#endif
+
+#ifndef GG_LONGLONG
+#ifndef INT_TYPES_DEFINED
+#define INT_TYPES_DEFINED
+#ifdef COMPILER_MSVC
+typedef unsigned __int64 uint64;
+typedef __int64 int64;
+#ifndef INT64_C
+#define INT64_C(x) x ## I64
+#endif
+#ifndef UINT64_C
+#define UINT64_C(x) x ## UI64
+#endif
+#define INT64_F "I64"
+#else  // COMPILER_MSVC
+#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long uint64;  // NOLINT
+typedef long int64;  // NOLINT
+#ifndef INT64_C
+#define INT64_C(x) x ## L
+#endif
+#ifndef UINT64_C
+#define UINT64_C(x) x ## UL
+#endif
+#define INT64_F "l"
+#else  // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long long uint64;  // NOLINT
+typedef long long int64;  // NOLINT
+#ifndef INT64_C
+#define INT64_C(x) x ## LL
+#endif
+#ifndef UINT64_C
+#define UINT64_C(x) x ## ULL
+#endif
+#define INT64_F "ll"
+#endif  // __LP64__
+#endif  // COMPILER_MSVC
+typedef unsigned int uint32;
+typedef int int32;
+typedef unsigned short uint16;  // NOLINT
+typedef short int16;  // NOLINT
+typedef unsigned char uint8;
+typedef signed char int8;
+#endif  // INT_TYPES_DEFINED
+#endif  // GG_LONGLONG
+
+// Detect compiler is for x86 or x64.
+#if defined(__x86_64__) || defined(_M_X64) || \
+    defined(__i386__) || defined(_M_IX86)
+#define CPU_X86 1
+#endif
+// Detect compiler is for ARM.
+#if defined(__arm__) || defined(_M_ARM)
+#define CPU_ARM 1
+#endif
+
+#ifndef ALIGNP
+#ifdef __cplusplus
+#define ALIGNP(p, t) \
+    (reinterpret_cast<uint8*>(((reinterpret_cast<uintptr_t>(p) + \
+    ((t) - 1)) & ~((t) - 1))))
+#else
+#define ALIGNP(p, t) \
+    ((uint8*)((((uintptr_t)(p) + ((t) - 1)) & ~((t) - 1))))  /* NOLINT */
+#endif
+#endif
+
+#if !defined(LIBYUV_API)
+#if defined(_WIN32) || defined(__CYGWIN__)
+#if defined(LIBYUV_BUILDING_SHARED_LIBRARY)
+#define LIBYUV_API __declspec(dllexport)
+#elif defined(LIBYUV_USING_SHARED_LIBRARY)
+#define LIBYUV_API __declspec(dllimport)
+#else
+#define LIBYUV_API
+#endif  // LIBYUV_BUILDING_SHARED_LIBRARY
+#elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__APPLE__) && \
+    (defined(LIBYUV_BUILDING_SHARED_LIBRARY) || \
+    defined(LIBYUV_USING_SHARED_LIBRARY))
+#define LIBYUV_API __attribute__ ((visibility ("default")))
+#else
+#define LIBYUV_API
+#endif  // __GNUC__
+#endif  // LIBYUV_API
+
+#define LIBYUV_BOOL int
+#define LIBYUV_FALSE 0
+#define LIBYUV_TRUE 1
+
+// Visual C x86 or GCC little endian.
+#if defined(__x86_64__) || defined(_M_X64) || \
+  defined(__i386__) || defined(_M_IX86) || \
+  defined(__arm__) || defined(_M_ARM) || \
+  (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#define LIBYUV_LITTLE_ENDIAN
+#endif
+
+#endif  // INCLUDE_LIBYUV_BASIC_TYPES_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/compare.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/compare.h
new file mode 100644
index 0000000..08b2bb2
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/compare.h
@@ -0,0 +1,78 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_COMPARE_H_  // NOLINT
+#define INCLUDE_LIBYUV_COMPARE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Compute a hash for specified memory. Seed of 5381 recommended.
+LIBYUV_API
+uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed);
+
+// Scan an opaque argb image and return fourcc based on alpha offset.
+// Returns FOURCC_ARGB, FOURCC_BGRA, or 0 if unknown.
+LIBYUV_API
+uint32 ARGBDetect(const uint8* argb, int stride_argb, int width, int height);
+
+// Sum Square Error - used to compute Mean Square Error or PSNR.
+LIBYUV_API
+uint64 ComputeSumSquareError(const uint8* src_a,
+                             const uint8* src_b, int count);
+
+LIBYUV_API
+uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
+                                  const uint8* src_b, int stride_b,
+                                  int width, int height);
+
+static const int kMaxPsnr = 128;
+
+LIBYUV_API
+double SumSquareErrorToPsnr(uint64 sse, uint64 count);
+
+LIBYUV_API
+double CalcFramePsnr(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height);
+
+LIBYUV_API
+double I420Psnr(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height);
+
+LIBYUV_API
+double CalcFrameSsim(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height);
+
+LIBYUV_API
+double I420Ssim(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_COMPARE_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/convert.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/convert.h
new file mode 100644
index 0000000..a8d3fa0
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/convert.h
@@ -0,0 +1,245 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_H_
+
+#include "libyuv/basic_types.h"
+// TODO(fbarchard): Remove the following headers includes.
+#include "libyuv/convert_from.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert I444 to I420.
+LIBYUV_API
+int I444ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I422 to I420.
+LIBYUV_API
+int I422ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I411 to I420.
+LIBYUV_API
+int I411ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Copy I420 to I420.
+#define I420ToI420 I420Copy
+LIBYUV_API
+int I420Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Convert I400 (grey) to I420.
+LIBYUV_API
+int I400ToI420(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+#define J400ToJ420 I400ToI420
+
+// Convert NV12 to I420.
+LIBYUV_API
+int NV12ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert NV21 to I420.
+LIBYUV_API
+int NV21ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert YUY2 to I420.
+LIBYUV_API
+int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert UYVY to I420.
+LIBYUV_API
+int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert M420 to I420.
+LIBYUV_API
+int M420ToI420(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// ARGB little endian (bgra in memory) to I420.
+LIBYUV_API
+int ARGBToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// BGRA little endian (argb in memory) to I420.
+LIBYUV_API
+int BGRAToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// ABGR little endian (rgba in memory) to I420.
+LIBYUV_API
+int ABGRToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// RGBA little endian (abgr in memory) to I420.
+LIBYUV_API
+int RGBAToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// RGB little endian (bgr in memory) to I420.
+LIBYUV_API
+int RGB24ToI420(const uint8* src_frame, int src_stride_frame,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height);
+
+// RGB big endian (rgb in memory) to I420.
+LIBYUV_API
+int RAWToI420(const uint8* src_frame, int src_stride_frame,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int width, int height);
+
+// RGB16 (RGBP fourcc) little endian to I420.
+LIBYUV_API
+int RGB565ToI420(const uint8* src_frame, int src_stride_frame,
+                 uint8* dst_y, int dst_stride_y,
+                 uint8* dst_u, int dst_stride_u,
+                 uint8* dst_v, int dst_stride_v,
+                 int width, int height);
+
+// RGB15 (RGBO fourcc) little endian to I420.
+LIBYUV_API
+int ARGB1555ToI420(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height);
+
+// RGB12 (R444 fourcc) little endian to I420.
+LIBYUV_API
+int ARGB4444ToI420(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height);
+
+#ifdef HAVE_JPEG
+// src_width/height provided by capture.
+// dst_width/height for clipping determine final size.
+LIBYUV_API
+int MJPGToI420(const uint8* sample, size_t sample_size,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int src_width, int src_height,
+               int dst_width, int dst_height);
+
+// Query size of MJPG in pixels.
+LIBYUV_API
+int MJPGSize(const uint8* sample, size_t sample_size,
+             int* width, int* height);
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// "src_size" is needed to parse MJPG.
+// "dst_stride_y" number of bytes in a row of the dst_y plane.
+//   Normally this would be the same as dst_width, with recommended alignment
+//   to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected. The caller should
+//   allocate the I420 buffer according to rotation.
+// "dst_stride_u" number of bytes in a row of the dst_u plane.
+//   Normally this would be the same as (dst_width + 1) / 2, with
+//   recommended alignment to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected.
+// "crop_x" and "crop_y" are starting position for cropping.
+//   To center, crop_x = (src_width - dst_width) / 2
+//              crop_y = (src_height - dst_height) / 2
+// "src_width" / "src_height" is size of src_frame in pixels.
+//   "src_height" can be negative indicating a vertically flipped image source.
+// "crop_width" / "crop_height" is the size to crop the src to.
+//    Must be less than or equal to src_width/src_height
+//    Cropping parameters are pre-rotation.
+// "rotation" can be 0, 90, 180 or 270.
+// "format" is a fourcc. ie 'I420', 'YUY2'
+// Returns 0 for successful; -1 for invalid parameter. Non-zero for failure.
+LIBYUV_API
+int ConvertToI420(const uint8* src_frame, size_t src_size,
+                  uint8* dst_y, int dst_stride_y,
+                  uint8* dst_u, int dst_stride_u,
+                  uint8* dst_v, int dst_stride_v,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/convert_argb.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/convert_argb.h
new file mode 100644
index 0000000..360c6d3
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/convert_argb.h
@@ -0,0 +1,231 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_ARGB_H_
+
+#include "libyuv/basic_types.h"
+// TODO(fbarchard): Remove the following headers includes
+#include "libyuv/convert_from.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+
+// TODO(fbarchard): This set of functions should exactly match convert.h
+// TODO(fbarchard): Add tests. Create random content of right size and convert
+// with C vs Opt and or to I420 and compare.
+// TODO(fbarchard): Some of these functions lack parameter setting.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Alias.
+#define ARGBToARGB ARGBCopy
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Convert I420 to ARGB.
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I422 to ARGB.
+LIBYUV_API
+int I422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I444 to ARGB.
+LIBYUV_API
+int I444ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I411 to ARGB.
+LIBYUV_API
+int I411ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I400 (grey) to ARGB.  Reverse of ARGBToI400.
+LIBYUV_API
+int I400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert J400 (jpeg grey) to ARGB.
+LIBYUV_API
+int J400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Alias.
+#define YToARGB I400ToARGB
+
+// Convert NV12 to ARGB.
+LIBYUV_API
+int NV12ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert NV21 to ARGB.
+LIBYUV_API
+int NV21ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert M420 to ARGB.
+LIBYUV_API
+int M420ToARGB(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert YUY2 to ARGB.
+LIBYUV_API
+int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert UYVY to ARGB.
+LIBYUV_API
+int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert J420 to ARGB.
+LIBYUV_API
+int J420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert J422 to ARGB.
+LIBYUV_API
+int J422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// BGRA little endian (argb in memory) to ARGB.
+LIBYUV_API
+int BGRAToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// ABGR little endian (rgba in memory) to ARGB.
+LIBYUV_API
+int ABGRToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// RGBA little endian (abgr in memory) to ARGB.
+LIBYUV_API
+int RGBAToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Deprecated function name.
+#define BG24ToARGB RGB24ToARGB
+
+// RGB little endian (bgr in memory) to ARGB.
+LIBYUV_API
+int RGB24ToARGB(const uint8* src_frame, int src_stride_frame,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height);
+
+// RGB big endian (rgb in memory) to ARGB.
+LIBYUV_API
+int RAWToARGB(const uint8* src_frame, int src_stride_frame,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// RGB16 (RGBP fourcc) little endian to ARGB.
+LIBYUV_API
+int RGB565ToARGB(const uint8* src_frame, int src_stride_frame,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// RGB15 (RGBO fourcc) little endian to ARGB.
+LIBYUV_API
+int ARGB1555ToARGB(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height);
+
+// RGB12 (R444 fourcc) little endian to ARGB.
+LIBYUV_API
+int ARGB4444ToARGB(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height);
+
+#ifdef HAVE_JPEG
+// src_width/height provided by capture
+// dst_width/height for clipping determine final size.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample, size_t sample_size,
+               uint8* dst_argb, int dst_stride_argb,
+               int src_width, int src_height,
+               int dst_width, int dst_height);
+#endif
+
+// Convert camera sample to ARGB with cropping, rotation and vertical flip.
+// "src_size" is needed to parse MJPG.
+// "dst_stride_argb" number of bytes in a row of the dst_argb plane.
+//   Normally this would be the same as dst_width, with recommended alignment
+//   to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected. The caller should
+//   allocate the I420 buffer according to rotation.
+// "dst_stride_u" number of bytes in a row of the dst_u plane.
+//   Normally this would be the same as (dst_width + 1) / 2, with
+//   recommended alignment to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected.
+// "crop_x" and "crop_y" are starting position for cropping.
+//   To center, crop_x = (src_width - dst_width) / 2
+//              crop_y = (src_height - dst_height) / 2
+// "src_width" / "src_height" is size of src_frame in pixels.
+//   "src_height" can be negative indicating a vertically flipped image source.
+// "crop_width" / "crop_height" is the size to crop the src to.
+//    Must be less than or equal to src_width/src_height
+//    Cropping parameters are pre-rotation.
+// "rotation" can be 0, 90, 180 or 270.
+// "format" is a fourcc. ie 'I420', 'YUY2'
+// Returns 0 for successful; -1 for invalid parameter. Non-zero for failure.
+LIBYUV_API
+int ConvertToARGB(const uint8* src_frame, size_t src_size,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_ARGB_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/convert_from.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/convert_from.h
new file mode 100644
index 0000000..9fd8d4d
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/convert_from.h
@@ -0,0 +1,181 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_FROM_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_FROM_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/rotate.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// See Also convert.h for conversions from formats to I420.
+
+// I420Copy in convert to I420ToI420.
+
+LIBYUV_API
+int I420ToI422(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+LIBYUV_API
+int I420ToI444(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+LIBYUV_API
+int I420ToI411(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Copy to I400. Source can be I420, I422, I444, I400, NV12 or NV21.
+LIBYUV_API
+int I400Copy(const uint8* src_y, int src_stride_y,
+             uint8* dst_y, int dst_stride_y,
+             int width, int height);
+
+// TODO(fbarchard): I420ToM420
+
+LIBYUV_API
+int I420ToNV12(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+LIBYUV_API
+int I420ToNV21(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+LIBYUV_API
+int I420ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+LIBYUV_API
+int I420ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+LIBYUV_API
+int I420ToRGB24(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_frame, int dst_stride_frame,
+                int width, int height);
+
+LIBYUV_API
+int I420ToRAW(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              uint8* dst_frame, int dst_stride_frame,
+              int width, int height);
+
+LIBYUV_API
+int I420ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_u, int src_stride_u,
+                 const uint8* src_v, int src_stride_v,
+                 uint8* dst_frame, int dst_stride_frame,
+                 int width, int height);
+
+// Convert I420 To RGB565 with 4x4 dither matrix (16 bytes).
+// Values in dither matrix from 0 to 7 recommended.
+// The order of the dither matrix is first byte is upper left.
+
+LIBYUV_API
+int I420ToRGB565Dither(const uint8* src_y, int src_stride_y,
+                       const uint8* src_u, int src_stride_u,
+                       const uint8* src_v, int src_stride_v,
+                       uint8* dst_frame, int dst_stride_frame,
+                       const uint8* dither4x4, int width, int height);
+
+LIBYUV_API
+int I420ToARGB1555(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_frame, int dst_stride_frame,
+                   int width, int height);
+
+LIBYUV_API
+int I420ToARGB4444(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_frame, int dst_stride_frame,
+                   int width, int height);
+
+// Convert I420 to specified format.
+// "dst_sample_stride" is bytes in a row for the destination. Pass 0 if the
+//    buffer has contiguous rows. Can be negative. A multiple of 16 is optimal.
+LIBYUV_API
+int ConvertFromI420(const uint8* y, int y_stride,
+                    const uint8* u, int u_stride,
+                    const uint8* v, int v_stride,
+                    uint8* dst_sample, int dst_sample_stride,
+                    int width, int height,
+                    uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_FROM_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/convert_from_argb.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/convert_from_argb.h
new file mode 100644
index 0000000..1df5320
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/convert_from_argb.h
@@ -0,0 +1,190 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy ARGB to ARGB.
+#define ARGBToARGB ARGBCopy
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Convert ARGB To BGRA.
+LIBYUV_API
+int ARGBToBGRA(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height);
+
+// Convert ARGB To ABGR.
+LIBYUV_API
+int ARGBToABGR(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height);
+
+// Convert ARGB To RGBA.
+LIBYUV_API
+int ARGBToRGBA(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+// Convert ARGB To RGB24.
+LIBYUV_API
+int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height);
+
+// Convert ARGB To RAW.
+LIBYUV_API
+int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_rgb, int dst_stride_rgb,
+              int width, int height);
+
+// Convert ARGB To RGB565.
+LIBYUV_API
+int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// Convert ARGB To RGB565 with 4x4 dither matrix (16 bytes).
+// Values in dither matrix from 0 to 7 recommended.
+// The order of the dither matrix is first byte is upper left.
+// TODO(fbarchard): Consider pointer to 2d array for dither4x4.
+// const uint8(*dither)[4][4];
+LIBYUV_API
+int ARGBToRGB565Dither(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_rgb565, int dst_stride_rgb565,
+                       const uint8* dither4x4, int width, int height);
+
+// Convert ARGB To ARGB1555.
+LIBYUV_API
+int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height);
+
+// Convert ARGB To ARGB4444.
+LIBYUV_API
+int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height);
+
+// Convert ARGB To I444.
+LIBYUV_API
+int ARGBToI444(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I422.
+LIBYUV_API
+int ARGBToI422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I420. (also in convert.h)
+LIBYUV_API
+int ARGBToI420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB to J420. (JPeg full range I420).
+LIBYUV_API
+int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB to J422.
+LIBYUV_API
+int ARGBToJ422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I411.
+LIBYUV_API
+int ARGBToI411(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB to J400. (JPeg full range).
+LIBYUV_API
+int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               int width, int height);
+
+// Convert ARGB to I400.
+LIBYUV_API
+int ARGBToI400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Convert ARGB to G. (Reverse of J400toARGB, which replicates G back to ARGB)
+LIBYUV_API
+int ARGBToG(const uint8* src_argb, int src_stride_argb,
+            uint8* dst_g, int dst_stride_g,
+            int width, int height);
+
+// Convert ARGB To NV12.
+LIBYUV_API
+int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+// Convert ARGB To NV21.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+// Convert ARGB To NV21.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+// Convert ARGB To YUY2.
+LIBYUV_API
+int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height);
+
+// Convert ARGB To UYVY.
+LIBYUV_API
+int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/cpu_id.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/cpu_id.h
new file mode 100644
index 0000000..dc858a8
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/cpu_id.h
@@ -0,0 +1,81 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CPU_ID_H_  // NOLINT
+#define INCLUDE_LIBYUV_CPU_ID_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// TODO(fbarchard): Consider overlapping bits for different architectures.
+// Internal flag to indicate cpuid requires initialization.
+#define kCpuInit 0x1
+
+// These flags are only valid on ARM processors.
+static const int kCpuHasARM = 0x2;
+static const int kCpuHasNEON = 0x4;
+// 0x8 reserved for future ARM flag.
+
+// These flags are only valid on x86 processors.
+static const int kCpuHasX86 = 0x10;
+static const int kCpuHasSSE2 = 0x20;
+static const int kCpuHasSSSE3 = 0x40;
+static const int kCpuHasSSE41 = 0x80;
+static const int kCpuHasSSE42 = 0x100;
+static const int kCpuHasAVX = 0x200;
+static const int kCpuHasAVX2 = 0x400;
+static const int kCpuHasERMS = 0x800;
+static const int kCpuHasFMA3 = 0x1000;
+// 0x2000, 0x4000, 0x8000 reserved for future X86 flags.
+
+// These flags are only valid on MIPS processors.
+static const int kCpuHasMIPS = 0x10000;
+static const int kCpuHasMIPS_DSP = 0x20000;
+static const int kCpuHasMIPS_DSPR2 = 0x40000;
+
+// Internal function used to auto-init.
+LIBYUV_API
+int InitCpuFlags(void);
+
+// Internal function for parsing /proc/cpuinfo.
+LIBYUV_API
+int ArmCpuCaps(const char* cpuinfo_name);
+
+// Detect CPU has SSE2 etc.
+// Test_flag parameter should be one of kCpuHas constants above.
+// returns non-zero if instruction set is detected
+static __inline int TestCpuFlag(int test_flag) {
+  LIBYUV_API extern int cpu_info_;
+  return (cpu_info_ == kCpuInit ? InitCpuFlags() : cpu_info_) & test_flag;
+}
+
+// For testing, allow CPU flags to be disabled.
+// ie MaskCpuFlags(~kCpuHasSSSE3) to disable SSSE3.
+// MaskCpuFlags(-1) to enable all cpu specific optimizations.
+// MaskCpuFlags(0) to disable all cpu specific optimizations.
+LIBYUV_API
+void MaskCpuFlags(int enable_flags);
+
+// Low level cpuid for X86. Returns zeros on other CPUs.
+// eax is the info type that you want.
+// ecx is typically the cpu number, and should normally be zero.
+LIBYUV_API
+void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CPU_ID_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/mjpeg_decoder.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/mjpeg_decoder.h
new file mode 100644
index 0000000..8423121
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/mjpeg_decoder.h
@@ -0,0 +1,192 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_MJPEG_DECODER_H_  // NOLINT
+#define INCLUDE_LIBYUV_MJPEG_DECODER_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+// NOTE: For a simplified public API use convert.h MJPGToI420().
+
+struct jpeg_common_struct;
+struct jpeg_decompress_struct;
+struct jpeg_source_mgr;
+
+namespace libyuv {
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+static const uint32 kUnknownDataSize = 0xFFFFFFFF;
+
+enum JpegSubsamplingType {
+  kJpegYuv420,
+  kJpegYuv422,
+  kJpegYuv411,
+  kJpegYuv444,
+  kJpegYuv400,
+  kJpegUnknown
+};
+
+struct Buffer {
+  const uint8* data;
+  int len;
+};
+
+struct BufferVector {
+  Buffer* buffers;
+  int len;
+  int pos;
+};
+
+struct SetJmpErrorMgr;
+
+// MJPEG ("Motion JPEG") is a pseudo-standard video codec where the frames are
+// simply independent JPEG images with a fixed huffman table (which is omitted).
+// It is rarely used in video transmission, but is common as a camera capture
+// format, especially in Logitech devices. This class implements a decoder for
+// MJPEG frames.
+//
+// See http://tools.ietf.org/html/rfc2435
+class LIBYUV_API MJpegDecoder {
+ public:
+  typedef void (*CallbackFunction)(void* opaque,
+                                   const uint8* const* data,
+                                   const int* strides,
+                                   int rows);
+
+  static const int kColorSpaceUnknown;
+  static const int kColorSpaceGrayscale;
+  static const int kColorSpaceRgb;
+  static const int kColorSpaceYCbCr;
+  static const int kColorSpaceCMYK;
+  static const int kColorSpaceYCCK;
+
+  MJpegDecoder();
+  ~MJpegDecoder();
+
+  // Loads a new frame, reads its headers, and determines the uncompressed
+  // image format.
+  // Returns LIBYUV_TRUE if image looks valid and format is supported.
+  // If return value is LIBYUV_TRUE, then the values for all the following
+  // getters are populated.
+  // src_len is the size of the compressed mjpeg frame in bytes.
+  LIBYUV_BOOL LoadFrame(const uint8* src, size_t src_len);
+
+  // Returns width of the last loaded frame in pixels.
+  int GetWidth();
+
+  // Returns height of the last loaded frame in pixels.
+  int GetHeight();
+
+  // Returns format of the last loaded frame. The return value is one of the
+  // kColorSpace* constants.
+  int GetColorSpace();
+
+  // Number of color components in the color space.
+  int GetNumComponents();
+
+  // Sample factors of the n-th component.
+  int GetHorizSampFactor(int component);
+
+  int GetVertSampFactor(int component);
+
+  int GetHorizSubSampFactor(int component);
+
+  int GetVertSubSampFactor(int component);
+
+  // Public for testability.
+  int GetImageScanlinesPerImcuRow();
+
+  // Public for testability.
+  int GetComponentScanlinesPerImcuRow(int component);
+
+  // Width of a component in bytes.
+  int GetComponentWidth(int component);
+
+  // Height of a component.
+  int GetComponentHeight(int component);
+
+  // Width of a component in bytes with padding for DCTSIZE. Public for testing.
+  int GetComponentStride(int component);
+
+  // Size of a component in bytes.
+  int GetComponentSize(int component);
+
+  // Call this after LoadFrame() if you decide you don't want to decode it
+  // after all.
+  LIBYUV_BOOL UnloadFrame();
+
+  // Decodes the entire image into a one-buffer-per-color-component format.
+  // dst_width must match exactly. dst_height must be <= to image height; if
+  // less, the image is cropped. "planes" must have size equal to at least
+  // GetNumComponents() and they must point to non-overlapping buffers of size
+  // at least GetComponentSize(i). The pointers in planes are incremented
+  // to point to after the end of the written data.
+  // TODO(fbarchard): Add dst_x, dst_y to allow specific rect to be decoded.
+  LIBYUV_BOOL DecodeToBuffers(uint8** planes, int dst_width, int dst_height);
+
+  // Decodes the entire image and passes the data via repeated calls to a
+  // callback function. Each call will get the data for a whole number of
+  // image scanlines.
+  // TODO(fbarchard): Add dst_x, dst_y to allow specific rect to be decoded.
+  LIBYUV_BOOL DecodeToCallback(CallbackFunction fn, void* opaque,
+                        int dst_width, int dst_height);
+
+  // The helper function which recognizes the jpeg sub-sampling type.
+  static JpegSubsamplingType JpegSubsamplingTypeHelper(
+     int* subsample_x, int* subsample_y, int number_of_components);
+
+ private:
+  void AllocOutputBuffers(int num_outbufs);
+  void DestroyOutputBuffers();
+
+  LIBYUV_BOOL StartDecode();
+  LIBYUV_BOOL FinishDecode();
+
+  void SetScanlinePointers(uint8** data);
+  LIBYUV_BOOL DecodeImcuRow();
+
+  int GetComponentScanlinePadding(int component);
+
+  // A buffer holding the input data for a frame.
+  Buffer buf_;
+  BufferVector buf_vec_;
+
+  jpeg_decompress_struct* decompress_struct_;
+  jpeg_source_mgr* source_mgr_;
+  SetJmpErrorMgr* error_mgr_;
+
+  // LIBYUV_TRUE iff at least one component has scanline padding. (i.e.,
+  // GetComponentScanlinePadding() != 0.)
+  LIBYUV_BOOL has_scanline_padding_;
+
+  // Temporaries used to point to scanline outputs.
+  int num_outbufs_;  // Outermost size of all arrays below.
+  uint8*** scanlines_;
+  int* scanlines_sizes_;
+  // Temporary buffer used for decoding when we can't decode directly to the
+  // output buffers. Large enough for just one iMCU row.
+  uint8** databuf_;
+  int* databuf_strides_;
+};
+
+}  // namespace libyuv
+
+#endif  //  __cplusplus
+#endif  // INCLUDE_LIBYUV_MJPEG_DECODER_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/planar_functions.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/planar_functions.h
new file mode 100644
index 0000000..ae994db
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/planar_functions.h
@@ -0,0 +1,453 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  // NOLINT
+#define INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_
+
+#include "libyuv/basic_types.h"
+
+// TODO(fbarchard): Remove the following headers includes.
+#include "libyuv/convert.h"
+#include "libyuv/convert_argb.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy a plane of data.
+LIBYUV_API
+void CopyPlane(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+LIBYUV_API
+void CopyPlane_16(const uint16* src_y, int src_stride_y,
+                  uint16* dst_y, int dst_stride_y,
+                  int width, int height);
+
+// Set a plane of data to a 32 bit value.
+LIBYUV_API
+void SetPlane(uint8* dst_y, int dst_stride_y,
+              int width, int height,
+              uint32 value);
+
+// Copy I400.  Supports inverting.
+LIBYUV_API
+int I400ToI400(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+#define J400ToJ400 I400ToI400
+
+// Copy I422 to I422.
+#define I422ToI422 I422Copy
+LIBYUV_API
+int I422Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Copy I444 to I444.
+#define I444ToI444 I444Copy
+LIBYUV_API
+int I444Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Convert YUY2 to I422.
+LIBYUV_API
+int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert UYVY to I422.
+LIBYUV_API
+int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+LIBYUV_API
+int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+LIBYUV_API
+int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+// Convert I420 to I400. (calls CopyPlane ignoring u/v).
+LIBYUV_API
+int I420ToI400(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Alias
+#define J420ToJ400 I420ToI400
+#define I420ToI420Mirror I420Mirror
+
+// I420 mirror.
+LIBYUV_API
+int I420Mirror(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Alias
+#define I400ToI400Mirror I400Mirror
+
+// I400 mirror.  A single plane is mirrored horizontally.
+// Pass negative height to achieve 180 degree rotation.
+LIBYUV_API
+int I400Mirror(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Alias
+#define ARGBToARGBMirror ARGBMirror
+
+// ARGB mirror.
+LIBYUV_API
+int ARGBMirror(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// Convert NV21 to RGB565.
+LIBYUV_API
+int NV21ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// I422ToARGB is in convert_argb.h
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height);
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height);
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+// Draw a rectangle into I420.
+LIBYUV_API
+int I420Rect(uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int x, int y, int width, int height,
+             int value_y, int value_u, int value_v);
+
+// Draw a rectangle into ARGB.
+LIBYUV_API
+int ARGBRect(uint8* dst_argb, int dst_stride_argb,
+             int x, int y, int width, int height, uint32 value);
+
+// Convert ARGB to gray scale ARGB.
+LIBYUV_API
+int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Make a rectangle of ARGB gray scale.
+LIBYUV_API
+int ARGBGray(uint8* dst_argb, int dst_stride_argb,
+             int x, int y, int width, int height);
+
+// Make a rectangle of ARGB Sepia tone.
+LIBYUV_API
+int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
+              int x, int y, int width, int height);
+
+// Apply a matrix rotation to each ARGB pixel.
+// matrix_argb is 4 signed ARGB values. -128 to 127 representing -2 to 2.
+// The first 4 coefficients apply to B, G, R, A and produce B of the output.
+// The next 4 coefficients apply to B, G, R, A and produce G of the output.
+// The next 4 coefficients apply to B, G, R, A and produce R of the output.
+// The last 4 coefficients apply to B, G, R, A and produce A of the output.
+LIBYUV_API
+int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    const int8* matrix_argb,
+                    int width, int height);
+
+// Deprecated. Use ARGBColorMatrix instead.
+// Apply a matrix rotation to each ARGB pixel.
+// matrix_argb is 3 signed ARGB values. -128 to 127 representing -1 to 1.
+// The first 4 coefficients apply to B, G, R, A and produce B of the output.
+// The next 4 coefficients apply to B, G, R, A and produce G of the output.
+// The last 4 coefficients apply to B, G, R, A and produce R of the output.
+LIBYUV_API
+int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
+                   const int8* matrix_rgb,
+                   int x, int y, int width, int height);
+
+// Apply a color table each ARGB pixel.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                   const uint8* table_argb,
+                   int x, int y, int width, int height);
+
+// Apply a color table each ARGB pixel but preserve destination alpha.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                  const uint8* table_argb,
+                  int x, int y, int width, int height);
+
+// Apply a luma/color table each ARGB pixel but preserve destination alpha.
+// Table contains 32768 values indexed by [Y][C] where 7 it 7 bit luma from
+// RGB (YJ style) and C is an 8 bit color component (R, G or B).
+LIBYUV_API
+int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_argb, int dst_stride_argb,
+                       const uint8* luma_rgb_table,
+                       int width, int height);
+
+// Apply a 3 term polynomial to ARGB values.
+// poly points to a 4x4 matrix.  The first row is constants.  The 2nd row is
+// coefficients for b, g, r and a.  The 3rd row is coefficients for b squared,
+// g squared, r squared and a squared.  The 4rd row is coefficients for b to
+// the 3, g to the 3, r to the 3 and a to the 3.  The values are summed and
+// result clamped to 0 to 255.
+// A polynomial approximation can be dirived using software such as 'R'.
+
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height);
+
+// Quantize a rectangle of ARGB. Alpha unaffected.
+// scale is a 16 bit fractional fixed point scaler between 0 and 65535.
+// interval_size should be a value between 1 and 255.
+// interval_offset should be a value between 0 and 255.
+LIBYUV_API
+int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
+                 int scale, int interval_size, int interval_offset,
+                 int x, int y, int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
+                     uint8* dst_argb, int dst_stride_argb,
+                     int width, int height);
+
+typedef void (*ARGBBlendRow)(const uint8* src_argb0, const uint8* src_argb1,
+                             uint8* dst_argb, int width);
+
+// Get function to Alpha Blend ARGB pixels and store to destination.
+LIBYUV_API
+ARGBBlendRow GetARGBBlend();
+
+// Alpha Blend ARGB images and store to destination.
+// Alpha of destination is set to 255.
+LIBYUV_API
+int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
+              const uint8* src_argb1, int src_stride_argb1,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// Multiply ARGB image by ARGB image. Shifted down by 8. Saturates to 255.
+LIBYUV_API
+int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// Add ARGB image with ARGB image. Saturates to 255.
+LIBYUV_API
+int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
+            const uint8* src_argb1, int src_stride_argb1,
+            uint8* dst_argb, int dst_stride_argb,
+            int width, int height);
+
+// Subtract ARGB image (argb1) from ARGB image (argb0). Saturates to 0.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// Convert I422 to YUY2.
+LIBYUV_API
+int I422ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+// Convert I422 to UYVY.
+LIBYUV_API
+int I422ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+// Convert unattentuated ARGB to preattenuated ARGB.
+LIBYUV_API
+int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
+// Convert preattentuated ARGB to unattenuated ARGB.
+LIBYUV_API
+int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+// Convert MJPG to ARGB.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample, size_t sample_size,
+               uint8* argb, int argb_stride,
+               int w, int h, int dw, int dh);
+
+// Internal function - do not call directly.
+// Computes table of cumulative sum for image where the value is the sum
+// of all values above and to the left of the entry. Used by ARGBBlur.
+LIBYUV_API
+int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
+                             int32* dst_cumsum, int dst_stride32_cumsum,
+                             int width, int height);
+
+// Blur ARGB image.
+// dst_cumsum table of width * (height + 1) * 16 bytes aligned to
+//   16 byte boundary.
+// dst_stride32_cumsum is number of ints in a row (width * 4).
+// radius is number of pixels around the center.  e.g. 1 = 3x3. 2=5x5.
+// Blur is optimized for radius of 5 (11x11) or less.
+LIBYUV_API
+int ARGBBlur(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int32* dst_cumsum, int dst_stride32_cumsum,
+             int width, int height, int radius);
+
+// Multiply ARGB image by ARGB value.
+LIBYUV_API
+int ARGBShade(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height, uint32 value);
+
+// Interpolate between two ARGB images using specified amount of interpolation
+// (0 to 255) and store to destination.
+// 'interpolation' is specified as 8 bit fraction where 0 means 100% src_argb0
+// and 255 means 1% src_argb0 and 99% src_argb1.
+// Internally uses ARGBScale bilinear filtering.
+// Caveat: This function will write up to 16 bytes beyond the end of dst_argb.
+LIBYUV_API
+int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
+                    const uint8* src_argb1, int src_stride_argb1,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height, int interpolation);
+
+#if defined(__pnacl__) || defined(__CLR_VER) || \
+    (defined(__i386__) && !defined(__SSE2__))
+#define LIBYUV_DISABLE_X86
+#endif
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_ARGBAFFINEROW_SSE2
+#endif
+
+// Row function for copying pixels from a source with a slope to a row
+// of destination. Useful for scaling, rotation, mirror, texture mapping.
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width);
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width);
+
+// Shuffle ARGB channel order.  e.g. BGRA to ARGB.
+// shuffler is 16 bytes and must be aligned.
+LIBYUV_API
+int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
+                uint8* dst_argb, int dst_stride_argb,
+                const uint8* shuffler, int width, int height);
+
+// Sobel ARGB effect with planar output.
+LIBYUV_API
+int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_y, int dst_stride_y,
+                     int width, int height);
+
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// Sobel ARGB effect w/ Sobel X, Sobel, Sobel Y in ARGB.
+LIBYUV_API
+int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/rotate.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/rotate.h
new file mode 100644
index 0000000..8af60b8
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/rotate.h
@@ -0,0 +1,117 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROTATE_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROTATE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Supported rotation.
+typedef enum RotationMode {
+  kRotate0 = 0,  // No rotation.
+  kRotate90 = 90,  // Rotate 90 degrees clockwise.
+  kRotate180 = 180,  // Rotate 180 degrees.
+  kRotate270 = 270,  // Rotate 270 degrees clockwise.
+
+  // Deprecated.
+  kRotateNone = 0,
+  kRotateClockwise = 90,
+  kRotateCounterClockwise = 270,
+} RotationModeEnum;
+
+// Rotate I420 frame.
+LIBYUV_API
+int I420Rotate(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int src_width, int src_height, enum RotationMode mode);
+
+// Rotate NV12 input and store in I420.
+LIBYUV_API
+int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
+                     const uint8* src_uv, int src_stride_uv,
+                     uint8* dst_y, int dst_stride_y,
+                     uint8* dst_u, int dst_stride_u,
+                     uint8* dst_v, int dst_stride_v,
+                     int src_width, int src_height, enum RotationMode mode);
+
+// Rotate a plane by 0, 90, 180, or 270.
+LIBYUV_API
+int RotatePlane(const uint8* src, int src_stride,
+                uint8* dst, int dst_stride,
+                int src_width, int src_height, enum RotationMode mode);
+
+// Rotate planes by 90, 180, 270. Deprecated.
+LIBYUV_API
+void RotatePlane90(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride,
+                   int width, int height);
+
+LIBYUV_API
+void RotatePlane180(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void RotatePlane270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void RotateUV90(const uint8* src, int src_stride,
+                uint8* dst_a, int dst_stride_a,
+                uint8* dst_b, int dst_stride_b,
+                int width, int height);
+
+// Rotations for when U and V are interleaved.
+// These functions take one input pointer and
+// split the data into two buffers while
+// rotating them. Deprecated.
+LIBYUV_API
+void RotateUV180(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+LIBYUV_API
+void RotateUV270(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+// The 90 and 270 functions are based on transposes.
+// Doing a transpose with reversing the read/write
+// order will result in a rotation by +- 90 degrees.
+// Deprecated.
+LIBYUV_API
+void TransposePlane(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void TransposeUV(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROTATE_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/rotate_argb.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/rotate_argb.h
new file mode 100644
index 0000000..660ff55
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/rotate_argb.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROTATE_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROTATE_ARGB_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/rotate.h"  // For RotationMode.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Rotate ARGB frame
+LIBYUV_API
+int ARGBRotate(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int src_width, int src_height, enum RotationMode mode);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROTATE_ARGB_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/rotate_row.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/rotate_row.h
new file mode 100644
index 0000000..c41cf32
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/rotate_row.h
@@ -0,0 +1,138 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROTATE_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROTATE_ROW_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if defined(__pnacl__) || defined(__CLR_VER) || \
+    (defined(__i386__) && !defined(__SSE2__))
+#define LIBYUV_DISABLE_X86
+#endif
+
+// Visual C 2012 required for AVX2.
+#if defined(_M_IX86) && !defined(__clang__) && \
+    defined(_MSC_VER) && _MSC_VER >= 1700
+#define VISUALC_HAS_AVX2 1
+#endif  // VisualStudio >= 2012
+
+// TODO(fbarchard): switch to standard form of inline; fails on clangcl.
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#if defined(__APPLE__) && defined(__i386__)
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".private_extern _" #name "                \n"                             \
+    ".align 4,0x90                             \n"                             \
+"_" #name ":                                   \n"
+#elif defined(__MINGW32__) || defined(__CYGWIN__) && defined(__i386__)
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".align 4,0x90                             \n"                             \
+"_" #name ":                                   \n"
+#else
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".align 4,0x90                             \n"                             \
+#name ":                                       \n"
+#endif
+#endif
+
+// The following are available for Visual C:
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
+    defined(_MSC_VER) && !defined(__clang__)
+#define HAS_TRANSPOSEWX8_SSSE3
+#define HAS_TRANSPOSEUVWX8_SSE2
+#endif
+
+// The following are available for GCC but not NaCL:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__i386__) || (defined(__x86_64__) && !defined(__native_client__)))
+#define HAS_TRANSPOSEWX8_SSSE3
+#endif
+
+// The following are available for 32 bit GCC:
+#if !defined(LIBYUV_DISABLE_X86) && defined(__i386__)  && !defined(__clang__)
+#define HAS_TRANSPOSEUVWX8_SSE2
+#endif
+
+// The following are available for 64 bit GCC but not NaCL:
+#if !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \
+    defined(__x86_64__)
+#define HAS_TRANSPOSEWX8_FAST_SSSE3
+#define HAS_TRANSPOSEUVWX8_SSE2
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
+#define HAS_TRANSPOSEWX8_NEON
+#define HAS_TRANSPOSEUVWX8_NEON
+#endif
+
+#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
+    defined(__mips__) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_TRANSPOSEWX8_MIPS_DSPR2
+#define HAS_TRANSPOSEUVWx8_MIPS_DSPR2
+#endif  // defined(__mips__)
+
+void TransposeWxH_C(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride, int width, int height);
+
+void TransposeWx8_C(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride, int width);
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride, int width);
+void TransposeWx8_SSSE3(const uint8* src, int src_stride,
+                        uint8* dst, int dst_stride, int width);
+void TransposeWx8_Fast_SSSE3(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride, int width);
+void TransposeWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride, int width);
+
+void TransposeWx8_Any_NEON(const uint8* src, int src_stride,
+                           uint8* dst, int dst_stride, int width);
+void TransposeWx8_Any_SSSE3(const uint8* src, int src_stride,
+                            uint8* dst, int dst_stride, int width);
+void TransposeWx8_Fast_Any_SSSE3(const uint8* src, int src_stride,
+                                 uint8* dst, int dst_stride, int width);
+void TransposeWx8_Any_MIPS_DSPR2(const uint8* src, int src_stride,
+                                 uint8* dst, int dst_stride, int width);
+
+void TransposeUVWxH_C(const uint8* src, int src_stride,
+                      uint8* dst_a, int dst_stride_a,
+                      uint8* dst_b, int dst_stride_b,
+                      int width, int height);
+
+void TransposeUVWx8_C(const uint8* src, int src_stride,
+                      uint8* dst_a, int dst_stride_a,
+                      uint8* dst_b, int dst_stride_b, int width);
+void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b, int width);
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b, int width);
+void TransposeUVWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                               uint8* dst_a, int dst_stride_a,
+                               uint8* dst_b, int dst_stride_b, int width);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROTATE_ROW_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/row.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/row.h
new file mode 100644
index 0000000..ebae3e7
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/row.h
@@ -0,0 +1,1856 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROW_H_
+
+#include <stdlib.h>  // For malloc.
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a) - 1)))
+
+#ifdef __cplusplus
+#define align_buffer_64(var, size)                                             \
+  uint8* var##_mem = reinterpret_cast<uint8*>(malloc((size) + 63));            \
+  uint8* var = reinterpret_cast<uint8*>                                        \
+      ((reinterpret_cast<intptr_t>(var##_mem) + 63) & ~63)
+#else
+#define align_buffer_64(var, size)                                             \
+  uint8* var##_mem = (uint8*)(malloc((size) + 63));               /* NOLINT */ \
+  uint8* var = (uint8*)(((intptr_t)(var##_mem) + 63) & ~63)       /* NOLINT */
+#endif
+
+#define free_aligned_buffer_64(var) \
+  free(var##_mem);  \
+  var = 0
+
+#if defined(__pnacl__) || defined(__CLR_VER) || \
+    (defined(__i386__) && !defined(__SSE2__))
+#define LIBYUV_DISABLE_X86
+#endif
+// True if compiling for SSSE3 as a requirement.
+#if defined(__SSSE3__) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 3))
+#define LIBYUV_SSSE3_ONLY
+#endif
+
+#if defined(__native_client__)
+#define LIBYUV_DISABLE_NEON
+#endif
+// clang >= 3.5.0 required for Arm64.
+#if defined(__clang__) && defined(__aarch64__) && !defined(LIBYUV_DISABLE_NEON)
+#if (__clang_major__ < 3) || (__clang_major__ == 3 && (__clang_minor__ < 5))
+#define LIBYUV_DISABLE_NEON
+#endif  // clang >= 3.5
+#endif  // __clang__
+
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+// Conversions:
+#define HAS_ABGRTOUVROW_SSSE3
+#define HAS_ABGRTOYROW_SSSE3
+#define HAS_ARGB1555TOARGBROW_SSE2
+#define HAS_ARGB4444TOARGBROW_SSE2
+#define HAS_ARGBSETROW_X86
+#define HAS_ARGBSHUFFLEROW_SSE2
+#define HAS_ARGBSHUFFLEROW_SSSE3
+#define HAS_ARGBTOARGB1555ROW_SSE2
+#define HAS_ARGBTOARGB4444ROW_SSE2
+#define HAS_ARGBTORAWROW_SSSE3
+#define HAS_ARGBTORGB24ROW_SSSE3
+#define HAS_ARGBTORGB565ROW_SSE2
+#define HAS_ARGBTOUV422ROW_SSSE3
+#define HAS_ARGBTOUV444ROW_SSSE3
+#define HAS_ARGBTOUVJROW_SSSE3
+#define HAS_ARGBTOUVROW_SSSE3
+#define HAS_ARGBTOYJROW_SSSE3
+#define HAS_ARGBTOYROW_SSSE3
+#define HAS_BGRATOUVROW_SSSE3
+#define HAS_BGRATOYROW_SSSE3
+#define HAS_COPYROW_ERMS
+#define HAS_COPYROW_SSE2
+#define HAS_I400TOARGBROW_SSE2
+#define HAS_I411TOARGBROW_SSSE3
+#define HAS_I422TOABGRROW_SSSE3
+#define HAS_I422TOARGB1555ROW_SSSE3
+#define HAS_I422TOARGB4444ROW_SSSE3
+#define HAS_I422TOARGBROW_SSSE3
+#define HAS_I422TOBGRAROW_SSSE3
+#define HAS_I422TORAWROW_SSSE3
+#define HAS_I422TORGB24ROW_SSSE3
+#define HAS_I422TORGB565ROW_SSSE3
+#define HAS_I422TORGBAROW_SSSE3
+#define HAS_I422TOUYVYROW_SSE2
+#define HAS_I422TOYUY2ROW_SSE2
+#define HAS_I444TOARGBROW_SSSE3
+#define HAS_J400TOARGBROW_SSE2
+#define HAS_J422TOARGBROW_SSSE3
+#define HAS_MERGEUVROW_SSE2
+#define HAS_MIRRORROW_SSE2
+#define HAS_MIRRORROW_SSSE3
+#define HAS_MIRRORROW_UV_SSSE3
+#define HAS_MIRRORUVROW_SSSE3
+#define HAS_NV12TOARGBROW_SSSE3
+#define HAS_NV12TORGB565ROW_SSSE3
+#define HAS_NV21TOARGBROW_SSSE3
+#define HAS_NV21TORGB565ROW_SSSE3
+#define HAS_RAWTOARGBROW_SSSE3
+#define HAS_RAWTOYROW_SSSE3
+#define HAS_RGB24TOARGBROW_SSSE3
+#define HAS_RGB24TOYROW_SSSE3
+#define HAS_RGB565TOARGBROW_SSE2
+#define HAS_RGBATOUVROW_SSSE3
+#define HAS_RGBATOYROW_SSSE3
+#define HAS_SETROW_ERMS
+#define HAS_SETROW_X86
+#define HAS_SPLITUVROW_SSE2
+#define HAS_UYVYTOARGBROW_SSSE3
+#define HAS_UYVYTOUV422ROW_SSE2
+#define HAS_UYVYTOUVROW_SSE2
+#define HAS_UYVYTOYROW_SSE2
+#define HAS_YUY2TOARGBROW_SSSE3
+#define HAS_YUY2TOUV422ROW_SSE2
+#define HAS_YUY2TOUVROW_SSE2
+#define HAS_YUY2TOYROW_SSE2
+
+// Effects:
+#define HAS_ARGBADDROW_SSE2
+#define HAS_ARGBAFFINEROW_SSE2
+#define HAS_ARGBATTENUATEROW_SSSE3
+#define HAS_ARGBBLENDROW_SSSE3
+#define HAS_ARGBCOLORMATRIXROW_SSSE3
+#define HAS_ARGBCOLORTABLEROW_X86
+#define HAS_ARGBCOPYALPHAROW_SSE2
+#define HAS_ARGBCOPYYTOALPHAROW_SSE2
+#define HAS_ARGBGRAYROW_SSSE3
+#define HAS_ARGBLUMACOLORTABLEROW_SSSE3
+#define HAS_ARGBMIRRORROW_SSE2
+#define HAS_ARGBMULTIPLYROW_SSE2
+#define HAS_ARGBPOLYNOMIALROW_SSE2
+#define HAS_ARGBQUANTIZEROW_SSE2
+#define HAS_ARGBSEPIAROW_SSSE3
+#define HAS_ARGBSHADEROW_SSE2
+#define HAS_ARGBSUBTRACTROW_SSE2
+#define HAS_ARGBUNATTENUATEROW_SSE2
+#define HAS_COMPUTECUMULATIVESUMROW_SSE2
+#define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+#define HAS_INTERPOLATEROW_SSE2
+#define HAS_INTERPOLATEROW_SSSE3
+#define HAS_RGBCOLORTABLEROW_X86
+#define HAS_SOBELROW_SSE2
+#define HAS_SOBELTOPLANEROW_SSE2
+#define HAS_SOBELXROW_SSE2
+#define HAS_SOBELXYROW_SSE2
+#define HAS_SOBELYROW_SSE2
+#endif
+
+// The following are available on x64 Visual C and clangcl.
+#if !defined(LIBYUV_DISABLE_X86) && defined (_M_X64) && \
+    (!defined(__clang__) || defined(__SSSE3__))
+#define HAS_I422TOARGBROW_SSSE3
+#endif
+
+// GCC >= 4.7.0 required for AVX2.
+#if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
+#if (__GNUC__ > 4) || (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7))
+#define GCC_HAS_AVX2 1
+#endif  // GNUC >= 4.7
+#endif  // __GNUC__
+
+// clang >= 3.4.0 required for AVX2.
+#if defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
+#if (__clang_major__ > 3) || (__clang_major__ == 3 && (__clang_minor__ >= 4))
+#define CLANG_HAS_AVX2 1
+#endif  // clang >= 3.4
+#endif  // __clang__
+
+// Visual C 2012 required for AVX2.
+#if defined(_M_IX86) && !defined(__clang__) && \
+    defined(_MSC_VER) && _MSC_VER >= 1700
+#define VISUALC_HAS_AVX2 1
+#endif  // VisualStudio >= 2012
+
+// The following are available require VS2012.  Port to GCC.
+#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
+#define HAS_ARGB1555TOARGBROW_AVX2
+#define HAS_ARGB4444TOARGBROW_AVX2
+#define HAS_ARGBTOARGB1555ROW_AVX2
+#define HAS_ARGBTOARGB4444ROW_AVX2
+#define HAS_ARGBTORGB565DITHERROW_AVX2
+#define HAS_ARGBTORGB565DITHERROW_SSE2
+#define HAS_ARGBTORGB565ROW_AVX2
+#define HAS_I411TOARGBROW_AVX2
+#define HAS_I422TOARGB1555ROW_AVX2
+#define HAS_I422TOARGB4444ROW_AVX2
+#define HAS_I422TORGB565ROW_AVX2
+#define HAS_I444TOARGBROW_AVX2
+#define HAS_J400TOARGBROW_AVX2
+#define HAS_NV12TOARGBROW_AVX2
+#define HAS_NV12TORGB565ROW_AVX2
+#define HAS_NV21TOARGBROW_AVX2
+#define HAS_NV21TORGB565ROW_AVX2
+#define HAS_RGB565TOARGBROW_AVX2
+#endif
+
+// The following are available on all x86 platforms, but
+// require VS2012, clang 3.4 or gcc 4.7.
+// The code supports NaCL but requires a new compiler and validator.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(VISUALC_HAS_AVX2) || \
+    defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
+#define HAS_ARGBCOPYALPHAROW_AVX2
+#define HAS_ARGBCOPYYTOALPHAROW_AVX2
+#define HAS_ARGBMIRRORROW_AVX2
+#define HAS_ARGBPOLYNOMIALROW_AVX2
+#define HAS_ARGBSHUFFLEROW_AVX2
+#define HAS_ARGBTOUVROW_AVX2
+#define HAS_ARGBTOYJROW_AVX2
+#define HAS_ARGBTOYROW_AVX2
+#define HAS_COPYROW_AVX
+#define HAS_I400TOARGBROW_AVX2
+#define HAS_I422TOABGRROW_AVX2
+#define HAS_I422TOARGBROW_AVX2
+#define HAS_I422TOBGRAROW_AVX2
+#define HAS_I422TORAWROW_AVX2
+#define HAS_I422TORGB24ROW_AVX2
+#define HAS_I422TORGBAROW_AVX2
+#define HAS_INTERPOLATEROW_AVX2
+#define HAS_J422TOARGBROW_AVX2
+#define HAS_MERGEUVROW_AVX2
+#define HAS_MIRRORROW_AVX2
+#define HAS_SPLITUVROW_AVX2
+#define HAS_UYVYTOARGBROW_AVX2
+#define HAS_UYVYTOUV422ROW_AVX2
+#define HAS_UYVYTOUVROW_AVX2
+#define HAS_UYVYTOYROW_AVX2
+#define HAS_YUY2TOARGBROW_AVX2
+#define HAS_YUY2TOUV422ROW_AVX2
+#define HAS_YUY2TOUVROW_AVX2
+#define HAS_YUY2TOYROW_AVX2
+
+// Effects:
+#define HAS_ARGBADDROW_AVX2
+#define HAS_ARGBATTENUATEROW_AVX2
+#define HAS_ARGBMULTIPLYROW_AVX2
+#define HAS_ARGBSUBTRACTROW_AVX2
+#define HAS_ARGBUNATTENUATEROW_AVX2
+#endif
+
+// The following are disabled when SSSE3 is available:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) && \
+    !defined(LIBYUV_SSSE3_ONLY)
+#define HAS_ARGBATTENUATEROW_SSE2
+#define HAS_ARGBBLENDROW_SSE2
+#define HAS_MIRRORROW_SSE2
+#endif
+
+// The following are available on Neon platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_ABGRTOUVROW_NEON
+#define HAS_ABGRTOYROW_NEON
+#define HAS_ARGB1555TOARGBROW_NEON
+#define HAS_ARGB1555TOUVROW_NEON
+#define HAS_ARGB1555TOYROW_NEON
+#define HAS_ARGB4444TOARGBROW_NEON
+#define HAS_ARGB4444TOUVROW_NEON
+#define HAS_ARGB4444TOYROW_NEON
+#define HAS_ARGBTOARGB1555ROW_NEON
+#define HAS_ARGBTOARGB4444ROW_NEON
+#define HAS_ARGBTORAWROW_NEON
+#define HAS_ARGBTORGB24ROW_NEON
+#define HAS_ARGBTORGB565ROW_NEON
+#define HAS_ARGBTOUV411ROW_NEON
+#define HAS_ARGBTOUV422ROW_NEON
+#define HAS_ARGBTOUV444ROW_NEON
+#define HAS_ARGBTOUVJROW_NEON
+#define HAS_ARGBTOUVROW_NEON
+#define HAS_ARGBTOYJROW_NEON
+#define HAS_ARGBTOYROW_NEON
+#define HAS_BGRATOUVROW_NEON
+#define HAS_BGRATOYROW_NEON
+#define HAS_COPYROW_NEON
+#define HAS_J400TOARGBROW_NEON
+#define HAS_I411TOARGBROW_NEON
+#define HAS_I422TOABGRROW_NEON
+#define HAS_I422TOARGB1555ROW_NEON
+#define HAS_I422TOARGB4444ROW_NEON
+#define HAS_I422TOARGBROW_NEON
+#define HAS_I422TOBGRAROW_NEON
+#define HAS_I422TORAWROW_NEON
+#define HAS_I422TORGB24ROW_NEON
+#define HAS_I422TORGB565ROW_NEON
+#define HAS_I422TORGBAROW_NEON
+#define HAS_I422TOUYVYROW_NEON
+#define HAS_I422TOYUY2ROW_NEON
+#define HAS_I444TOARGBROW_NEON
+#define HAS_MERGEUVROW_NEON
+#define HAS_MIRRORROW_NEON
+#define HAS_MIRRORUVROW_NEON
+#define HAS_NV12TOARGBROW_NEON
+#define HAS_NV12TORGB565ROW_NEON
+#define HAS_NV21TOARGBROW_NEON
+#define HAS_NV21TORGB565ROW_NEON
+#define HAS_RAWTOARGBROW_NEON
+#define HAS_RAWTOUVROW_NEON
+#define HAS_RAWTOYROW_NEON
+#define HAS_RGB24TOARGBROW_NEON
+#define HAS_RGB24TOUVROW_NEON
+#define HAS_RGB24TOYROW_NEON
+#define HAS_RGB565TOARGBROW_NEON
+#define HAS_RGB565TOUVROW_NEON
+#define HAS_RGB565TOYROW_NEON
+#define HAS_RGBATOUVROW_NEON
+#define HAS_RGBATOYROW_NEON
+#define HAS_SETROW_NEON
+#define HAS_ARGBSETROW_NEON
+#define HAS_SPLITUVROW_NEON
+#define HAS_UYVYTOARGBROW_NEON
+#define HAS_UYVYTOUV422ROW_NEON
+#define HAS_UYVYTOUVROW_NEON
+#define HAS_UYVYTOYROW_NEON
+#define HAS_I400TOARGBROW_NEON
+#define HAS_YUY2TOARGBROW_NEON
+#define HAS_YUY2TOUV422ROW_NEON
+#define HAS_YUY2TOUVROW_NEON
+#define HAS_YUY2TOYROW_NEON
+#define HAS_ARGBTORGB565DITHERROW_NEON
+
+// Effects:
+#define HAS_ARGBADDROW_NEON
+#define HAS_ARGBATTENUATEROW_NEON
+#define HAS_ARGBBLENDROW_NEON
+#define HAS_ARGBGRAYROW_NEON
+#define HAS_ARGBMIRRORROW_NEON
+#define HAS_ARGBMULTIPLYROW_NEON
+#define HAS_ARGBQUANTIZEROW_NEON
+#define HAS_ARGBSEPIAROW_NEON
+#define HAS_ARGBSHADEROW_NEON
+#define HAS_ARGBSUBTRACTROW_NEON
+#define HAS_INTERPOLATEROW_NEON
+#define HAS_SOBELROW_NEON
+#define HAS_SOBELTOPLANEROW_NEON
+#define HAS_SOBELXROW_NEON
+#define HAS_SOBELXYROW_NEON
+#define HAS_SOBELYROW_NEON
+#define HAS_ARGBCOLORMATRIXROW_NEON
+#define HAS_ARGBSHUFFLEROW_NEON
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32) && (__mips_isa_rev < 6)
+#define HAS_COPYROW_MIPS
+#if defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_I422TOABGRROW_MIPS_DSPR2
+#define HAS_I422TOARGBROW_MIPS_DSPR2
+#define HAS_I422TOBGRAROW_MIPS_DSPR2
+#define HAS_INTERPOLATEROW_MIPS_DSPR2
+#define HAS_MIRRORROW_MIPS_DSPR2
+#define HAS_MIRRORUVROW_MIPS_DSPR2
+#define HAS_SPLITUVROW_MIPS_DSPR2
+#endif
+#endif
+
+#if defined(_MSC_VER) && !defined(__CLR_VER)
+#define SIMD_ALIGNED(var) __declspec(align(16)) var
+#define SIMD_ALIGNED32(var) __declspec(align(64)) var
+typedef __declspec(align(16)) int16 vec16[8];
+typedef __declspec(align(16)) int32 vec32[4];
+typedef __declspec(align(16)) int8 vec8[16];
+typedef __declspec(align(16)) uint16 uvec16[8];
+typedef __declspec(align(16)) uint32 uvec32[4];
+typedef __declspec(align(16)) uint8 uvec8[16];
+typedef __declspec(align(32)) int16 lvec16[16];
+typedef __declspec(align(32)) int32 lvec32[8];
+typedef __declspec(align(32)) int8 lvec8[32];
+typedef __declspec(align(32)) uint16 ulvec16[16];
+typedef __declspec(align(32)) uint32 ulvec32[8];
+typedef __declspec(align(32)) uint8 ulvec8[32];
+#elif defined(__GNUC__)
+// Caveat GCC 4.2 to 4.7 have a known issue using vectors with const.
+#define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
+#define SIMD_ALIGNED32(var) var __attribute__((aligned(64)))
+typedef int16 __attribute__((vector_size(16))) vec16;
+typedef int32 __attribute__((vector_size(16))) vec32;
+typedef int8 __attribute__((vector_size(16))) vec8;
+typedef uint16 __attribute__((vector_size(16))) uvec16;
+typedef uint32 __attribute__((vector_size(16))) uvec32;
+typedef uint8 __attribute__((vector_size(16))) uvec8;
+typedef int16 __attribute__((vector_size(32))) lvec16;
+typedef int32 __attribute__((vector_size(32))) lvec32;
+typedef int8 __attribute__((vector_size(32))) lvec8;
+typedef uint16 __attribute__((vector_size(32))) ulvec16;
+typedef uint32 __attribute__((vector_size(32))) ulvec32;
+typedef uint8 __attribute__((vector_size(32))) ulvec8;
+#else
+#define SIMD_ALIGNED(var) var
+#define SIMD_ALIGNED32(var) var
+typedef int16 vec16[8];
+typedef int32 vec32[4];
+typedef int8 vec8[16];
+typedef uint16 uvec16[8];
+typedef uint32 uvec32[4];
+typedef uint8 uvec8[16];
+typedef int16 lvec16[16];
+typedef int32 lvec32[8];
+typedef int8 lvec8[32];
+typedef uint16 ulvec16[16];
+typedef uint32 ulvec32[8];
+typedef uint8 ulvec8[32];
+#endif
+
+#if defined(__APPLE__) || defined(__x86_64__) || defined(__llvm__)
+#define OMITFP
+#else
+#define OMITFP __attribute__((optimize("omit-frame-pointer")))
+#endif
+
+// NaCL macros for GCC x86 and x64.
+#if defined(__native_client__)
+#define LABELALIGN ".p2align 5\n"
+#else
+#define LABELALIGN
+#endif
+#if defined(__native_client__) && defined(__x86_64__)
+// r14 is used for MEMOP macros.
+#define NACL_R14 "r14",
+#define BUNDLELOCK ".bundle_lock\n"
+#define BUNDLEUNLOCK ".bundle_unlock\n"
+#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
+#define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
+#define MEMLEA(offset, base) #offset "(%q" #base ")"
+#define MEMLEA3(offset, index, scale) \
+    #offset "(,%q" #index "," #scale ")"
+#define MEMLEA4(offset, base, index, scale) \
+    #offset "(%q" #base ",%q" #index "," #scale ")"
+#define MEMMOVESTRING(s, d) "%%nacl:(%q" #s "),%%nacl:(%q" #d "), %%r15"
+#define MEMSTORESTRING(reg, d) "%%" #reg ",%%nacl:(%q" #d "), %%r15"
+#define MEMOPREG(opcode, offset, base, index, scale, reg) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%%" #reg "\n" \
+    BUNDLEUNLOCK
+#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " %%" #reg ",(%%r15,%%r14)\n" \
+    BUNDLEUNLOCK
+#define MEMOPARG(opcode, offset, base, index, scale, arg) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%" #arg "\n" \
+    BUNDLEUNLOCK
+#define VMEMOPREG(opcode, offset, base, index, scale, reg1, reg2) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%%" #reg1 ",%%" #reg2 "\n" \
+    BUNDLEUNLOCK
+#define VEXTOPMEM(op, sel, reg, offset, base, index, scale) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #op " $" #sel ",%%" #reg ",(%%r15,%%r14)\n" \
+    BUNDLEUNLOCK
+#else  // defined(__native_client__) && defined(__x86_64__)
+#define NACL_R14
+#define BUNDLEALIGN
+#define MEMACCESS(base) "(%" #base ")"
+#define MEMACCESS2(offset, base) #offset "(%" #base ")"
+#define MEMLEA(offset, base) #offset "(%" #base ")"
+#define MEMLEA3(offset, index, scale) \
+    #offset "(,%" #index "," #scale ")"
+#define MEMLEA4(offset, base, index, scale) \
+    #offset "(%" #base ",%" #index "," #scale ")"
+#define MEMMOVESTRING(s, d)
+#define MEMSTORESTRING(reg, d)
+#define MEMOPREG(opcode, offset, base, index, scale, reg) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
+#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
+    #opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
+#define MEMOPARG(opcode, offset, base, index, scale, arg) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%" #arg "\n"
+#define VMEMOPREG(opcode, offset, base, index, scale, reg1, reg2) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg1 ",%%" \
+    #reg2 "\n"
+#define VEXTOPMEM(op, sel, reg, offset, base, index, scale) \
+    #op " $" #sel ",%%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
+#endif  // defined(__native_client__) && defined(__x86_64__)
+
+#if defined(__arm__) || defined(__aarch64__)
+#undef MEMACCESS
+#if defined(__native_client__)
+#define MEMACCESS(base) ".p2align 3\nbic %" #base ", #0xc0000000\n"
+#else
+#define MEMACCESS(base)
+#endif
+#endif
+
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width);
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width);
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width);
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width);
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width);
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width);
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width);
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width);
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width);
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_vu,
+                        uint8* dst_argb,
+                        int width);
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width);
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_vu,
+                          uint8* dst_rgb565,
+                          int width);
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width);
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width);
+
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix);
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix);
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix);
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
+void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_C(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_C(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_C(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_C(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_C(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_C(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int pix);
+void ARGBToYRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Any_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Any_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Any_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Any_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Any_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Any_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Any_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Any_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Any_NEON(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_Any_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_Any_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_Any_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
+
+void ARGBToUVRow_AVX2(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Any_AVX2(const uint8* src_argb, int src_stride_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_SSSE3(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_SSSE3(const uint8* src_argb, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                       uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
+                            uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_Any_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_Any_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                           uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_Any_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUV422Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUV411Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUVRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVJRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
+                           uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToUVRow_Any_NEON(const uint8* src_bgra, int src_stride_bgra,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ABGRToUVRow_Any_NEON(const uint8* src_abgr, int src_stride_abgr,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void RGBAToUVRow_Any_NEON(const uint8* src_rgba, int src_stride_rgba,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void RGB24ToUVRow_Any_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                           uint8* dst_u, uint8* dst_v, int pix);
+void RAWToUVRow_Any_NEON(const uint8* src_raw, int src_stride_raw,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void RGB565ToUVRow_Any_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                            uint8* dst_u, uint8* dst_v, int pix);
+void ARGB1555ToUVRow_Any_NEON(const uint8* src_argb1555,
+                              int src_stride_argb1555,
+                              uint8* dst_u, uint8* dst_v, int pix);
+void ARGB4444ToUVRow_Any_NEON(const uint8* src_argb4444,
+                              int src_stride_argb4444,
+                              uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVRow_C(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_C(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_C(const uint8* src_bgra, int src_stride_bgra,
+                   uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_C(const uint8* src_abgr, int src_stride_abgr,
+                   uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_C(const uint8* src_rgba, int src_stride_rgba,
+                   uint8* dst_u, uint8* dst_v, int width);
+void RGB24ToUVRow_C(const uint8* src_rgb24, int src_stride_rgb24,
+                    uint8* dst_u, uint8* dst_v, int width);
+void RAWToUVRow_C(const uint8* src_raw, int src_stride_raw,
+                  uint8* dst_u, uint8* dst_v, int width);
+void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
+                     uint8* dst_u, uint8* dst_v, int width);
+void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
+                       uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV444Row_SSSE3(const uint8* src_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Any_SSSE3(const uint8* src_argb,
+                              uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV422Row_SSSE3(const uint8* src_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_Any_SSSE3(const uint8* src_argb,
+                              uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV444Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV411Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJ422Row_C(const uint8* src_argb,
+                       uint8* dst_u, uint8* dst_v, int width);
+
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width);
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width);
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
+void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width);
+void MirrorRow_C(const uint8* src, uint8* dst, int width);
+void MirrorRow_Any_AVX2(const uint8* src, uint8* dst, int width);
+void MirrorRow_Any_SSSE3(const uint8* src, uint8* dst, int width);
+void MirrorRow_Any_SSE2(const uint8* src, uint8* dst, int width);
+void MirrorRow_Any_NEON(const uint8* src, uint8* dst, int width);
+
+void MirrorUVRow_SSSE3(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                       int width);
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width);
+void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                            int width);
+void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                   int width);
+
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_Any_AVX2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_Any_SSE2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_Any_NEON(const uint8* src, uint8* dst, int width);
+
+void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                           int pix);
+void SplitUVRow_Any_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix);
+
+void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                  int width);
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_Any_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+void MergeUVRow_Any_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+void MergeUVRow_Any_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count);
+void CopyRow_AVX(const uint8* src, uint8* dst, int count);
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count);
+void CopyRow_NEON(const uint8* src, uint8* dst, int count);
+void CopyRow_MIPS(const uint8* src, uint8* dst, int count);
+void CopyRow_C(const uint8* src, uint8* dst, int count);
+void CopyRow_Any_SSE2(const uint8* src, uint8* dst, int count);
+void CopyRow_Any_AVX(const uint8* src, uint8* dst, int count);
+void CopyRow_Any_NEON(const uint8* src, uint8* dst, int count);
+
+void CopyRow_16_C(const uint16* src, uint16* dst, int count);
+
+void ARGBCopyAlphaRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBCopyAlphaRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBCopyAlphaRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+
+void ARGBCopyYToAlphaRow_C(const uint8* src_y, uint8* dst_argb, int width);
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
+
+void SetRow_C(uint8* dst, uint8 v8, int count);
+void SetRow_X86(uint8* dst, uint8 v8, int count);
+void SetRow_ERMS(uint8* dst, uint8 v8, int count);
+void SetRow_NEON(uint8* dst, uint8 v8, int count);
+void SetRow_Any_X86(uint8* dst, uint8 v8, int count);
+void SetRow_Any_NEON(uint8* dst, uint8 v8, int count);
+
+void ARGBSetRow_C(uint8* dst_argb, uint32 v32, int count);
+void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count);
+void ARGBSetRow_NEON(uint8* dst_argb, uint32 v32, int count);
+void ARGBSetRow_Any_NEON(uint8* dst_argb, uint32 v32, int count);
+
+// ARGBShufflers for BGRAToARGB etc.
+void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
+                      const uint8* shuffler, int pix);
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix);
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+void RGB565ToARGBRow_AVX2(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_AVX2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_AVX2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_C(const uint8* src_rgb, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
+void ARGB4444ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
+void RGB24ToARGBRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_Any_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
+
+void RGB565ToARGBRow_Any_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+void RGB565ToARGBRow_Any_AVX2(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_AVX2(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_AVX2(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+
+void RGB24ToARGBRow_Any_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_Any_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_Any_NEON(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+
+void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB565DitherRow_C(const uint8* src_argb, uint8* dst_rgb,
+                             const uint32 dither4, int pix);
+void ARGBToRGB565DitherRow_SSE2(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int pix);
+void ARGBToRGB565DitherRow_AVX2(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int pix);
+
+void ARGBToRGB565Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int width);
+
+void ARGBToRGBARow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_Any_AVX2(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int pix);
+
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I411ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void NV12ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_uv,
+                     uint8* dst_argb,
+                     int width);
+void NV21ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_vu,
+                       uint8* dst_argb,
+                       int width);
+void NV12ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_uv,
+                       uint8* dst_argb,
+                       int width);
+void NV21ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_vu,
+                     uint8* dst_argb,
+                     int width);
+void YUY2ToARGBRow_C(const uint8* src_yuy2,
+                     uint8* dst_argb,
+                     int width);
+void UYVYToARGBRow_C(const uint8* src_uyvy,
+                     uint8* dst_argb,
+                     int width);
+void J422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I422ToBGRARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_bgra,
+                     int width);
+void I422ToABGRRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_abgr,
+                     int width);
+void I422ToRGBARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_rgba,
+                     int width);
+void I422ToRGB24Row_C(const uint8* src_y,
+                      const uint8* src_u,
+                      const uint8* src_v,
+                      uint8* dst_rgb24,
+                      int width);
+void I422ToRAWRow_C(const uint8* src_y,
+                    const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* dst_raw,
+                    int width);
+void I422ToARGB4444Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width);
+void I422ToARGB1555Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width);
+void I422ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_rgb565,
+                       int width);
+void I422ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToBGRARow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToRGBARow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToABGRRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I444ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I444ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I411ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I411ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void NV12ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_uv,
+                         uint8* dst_argb,
+                         int width);
+void NV21ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_vu,
+                         uint8* dst_argb,
+                         int width);
+void NV12ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width);
+void NV21ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_vu,
+                        uint8* dst_argb,
+                        int width);
+void NV12ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_uv,
+                           uint8* dst_argb,
+                           int width);
+void NV21ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_vu,
+                           uint8* dst_argb,
+                           int width);
+void NV12ToRGB565Row_AVX2(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_argb,
+                          int width);
+void NV21ToRGB565Row_AVX2(const uint8* src_y,
+                          const uint8* src_vu,
+                          uint8* dst_argb,
+                          int width);
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
+                         uint8* dst_argb,
+                         int width);
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
+                         uint8* dst_argb,
+                         int width);
+void YUY2ToARGBRow_AVX2(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width);
+void UYVYToARGBRow_AVX2(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width);
+void J422ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void J422ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToBGRARow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_bgra,
+                         int width);
+void I422ToABGRRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_abgr,
+                         int width);
+void I422ToRGBARow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgba,
+                         int width);
+void I422ToARGB4444Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToARGB4444Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGB1555Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToARGB1555Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+void I422ToRGB565Row_AVX2(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_argb,
+                          int width);
+void I422ToRGB24Row_SSSE3(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb24,
+                          int width);
+void I422ToRGB24Row_AVX2(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width);
+void I422ToRAWRow_SSSE3(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_raw,
+                        int width);
+void I422ToRAWRow_AVX2(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width);
+void I422ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToBGRARow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGBARow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToABGRRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I444ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I444ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I411ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I411ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void NV12ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_uv,
+                             uint8* dst_argb,
+                             int width);
+void NV21ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_vu,
+                             uint8* dst_argb,
+                             int width);
+void NV12ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV21ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_vu,
+                            uint8* dst_argb,
+                            int width);
+void NV12ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_uv,
+                               uint8* dst_argb,
+                               int width);
+void NV21ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_vu,
+                               uint8* dst_argb,
+                               int width);
+void NV12ToRGB565Row_Any_AVX2(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void NV21ToRGB565Row_Any_AVX2(const uint8* src_y,
+                              const uint8* src_vu,
+                              uint8* dst_argb,
+                              int width);
+void YUY2ToARGBRow_Any_SSSE3(const uint8* src_yuy2,
+                             uint8* dst_argb,
+                             int width);
+void UYVYToARGBRow_Any_SSSE3(const uint8* src_uyvy,
+                             uint8* dst_argb,
+                             int width);
+void YUY2ToARGBRow_Any_AVX2(const uint8* src_yuy2,
+                            uint8* dst_argb,
+                            int width);
+void UYVYToARGBRow_Any_AVX2(const uint8* src_uyvy,
+                            uint8* dst_argb,
+                            int width);
+void J422ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void J422ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToBGRARow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_bgra,
+                             int width);
+void I422ToABGRRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_abgr,
+                             int width);
+void I422ToRGBARow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_rgba,
+                             int width);
+void I422ToARGB4444Row_Any_SSSE3(const uint8* src_y,
+                                 const uint8* src_u,
+                                 const uint8* src_v,
+                                 uint8* dst_rgba,
+                                 int width);
+void I422ToARGB4444Row_Any_AVX2(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_rgba,
+                                int width);
+void I422ToARGB1555Row_Any_SSSE3(const uint8* src_y,
+                                 const uint8* src_u,
+                                 const uint8* src_v,
+                                 uint8* dst_rgba,
+                                 int width);
+void I422ToARGB1555Row_Any_AVX2(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_rgba,
+                                int width);
+void I422ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_u,
+                               const uint8* src_v,
+                               uint8* dst_rgba,
+                               int width);
+void I422ToRGB565Row_Any_AVX2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_rgba,
+                              int width);
+void I422ToRGB24Row_Any_SSSE3(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToRGB24Row_Any_AVX2(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToRAWRow_Any_SSSE3(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRAWRow_Any_AVX2(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+
+void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_Any_AVX2(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int width);
+
+// ARGB preattenuated alpha blend.
+void ARGBBlendRow_SSSE3(const uint8* src_argb, const uint8* src_argb1,
+                        uint8* dst_argb, int width);
+void ARGBBlendRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBBlendRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1,
+                    uint8* dst_argb, int width);
+
+// ARGB multiply images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBMultiplyRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBMultiplyRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
+// ARGB add images.
+void ARGBAddRow_C(const uint8* src_argb, const uint8* src_argb1,
+                  uint8* dst_argb, int width);
+void ARGBAddRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+void ARGBAddRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+void ARGBAddRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+
+// ARGB subtract images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBSubtractRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBSubtractRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
+void ARGBToRGB24Row_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB565DitherRow_Any_SSE2(const uint8* src_argb, uint8* dst_rgb,
+                                    const uint32 dither4, int pix);
+void ARGBToRGB565DitherRow_Any_AVX2(const uint8* src_argb, uint8* dst_rgb,
+                                    const uint32 dither4, int pix);
+
+void ARGBToRGB565Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB24Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565DitherRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb,
+                                    const uint32 dither4, int width);
+
+void I444ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I411ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToBGRARow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToABGRRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGBARow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGB24Row_Any_NEON(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToRAWRow_Any_NEON(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+void I422ToARGB4444Row_Any_NEON(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_argb,
+                                int width);
+void I422ToARGB1555Row_Any_NEON(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_argb,
+                                int width);
+void I422ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void NV12ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV21ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV12ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void NV21ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void YUY2ToARGBRow_Any_NEON(const uint8* src_yuy2,
+                            uint8* dst_argb,
+                            int width);
+void UYVYToARGBRow_Any_NEON(const uint8* src_uyvy,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+
+void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_C(const uint8* src_yuy2, int stride_yuy2,
+                   uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_C(const uint8* src_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_AVX2(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_SSE2(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_NEON(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_NEON(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_NEON(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+
+void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_C(const uint8* src_uyvy, int stride_uyvy,
+                   uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_C(const uint8* src_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_AVX2(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_SSE2(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_NEON(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_NEON(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+
+void I422ToYUY2Row_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_yuy2, int width);
+void I422ToUYVYRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_uyvy, int width);
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width);
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width);
+void I422ToYUY2Row_Any_SSE2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_yuy2, int width);
+void I422ToUYVYRow_Any_SSE2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_uyvy, int width);
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width);
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width);
+void I422ToYUY2Row_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_yuy2, int width);
+void I422ToUYVYRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_uyvy, int width);
+
+// Effects related row functions.
+void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+void ARGBAttenuateRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                int width);
+void ARGBAttenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+void ARGBAttenuateRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+
+// Inverse table for unattenuate, shared by C and SSE2.
+extern const uint32 fixed_invtbl8[256];
+void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+void ARGBUnattenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+
+void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
+
+void ARGBSepiaRow_C(uint8* dst_argb, int width);
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width);
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width);
+
+void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
+                          const int8* matrix_argb, int width);
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width);
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width);
+
+void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
+
+void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
+
+void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
+                       int interval_offset, int width);
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
+
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value);
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+
+// Used for blur.
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst, int count);
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width);
+
+void CumulativeSumToAverageRow_C(const int32* topleft, const int32* botleft,
+                                 int width, int area, uint8* dst, int count);
+void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
+                               const int32* previous_cumsum, int width);
+
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width);
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width);
+
+// Used for I420Scale, ARGBScale, and ARGBInterpolate.
+void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+                      ptrdiff_t src_stride_ptr,
+                      int width, int source_y_fraction);
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride_ptr, int width,
+                          int source_y_fraction);
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                               ptrdiff_t src_stride_ptr, int width,
+                               int source_y_fraction);
+void InterpolateRow_Any_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                              ptrdiff_t src_stride_ptr, int width,
+                              int source_y_fraction);
+void InterpolateRow_Any_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride_ptr, int width,
+                                   int source_y_fraction);
+
+void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         ptrdiff_t src_stride_ptr,
+                         int width, int source_y_fraction);
+
+// Sobel images.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width);
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width);
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width);
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width);
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width);
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width);
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width);
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
+void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_y, int width);
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width);
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width);
+void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                  uint8* dst_argb, int width);
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width);
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width);
+void SobelRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_argb, int width);
+void SobelRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_argb, int width);
+void SobelToPlaneRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                              uint8* dst_y, int width);
+void SobelToPlaneRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                              uint8* dst_y, int width);
+void SobelXYRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                         uint8* dst_argb, int width);
+void SobelXYRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                         uint8* dst_argb, int width);
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width);
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+
+void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                             const uint8* luma, uint32 lumacoeff);
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROW_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/scale.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/scale.h
new file mode 100644
index 0000000..102158d
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/scale.h
@@ -0,0 +1,103 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Supported filtering.
+typedef enum FilterMode {
+  kFilterNone = 0,  // Point sample; Fastest.
+  kFilterLinear = 1,  // Filter horizontally only.
+  kFilterBilinear = 2,  // Faster than box, but lower quality scaling down.
+  kFilterBox = 3  // Highest quality.
+} FilterModeEnum;
+
+// Scale a YUV plane.
+LIBYUV_API
+void ScalePlane(const uint8* src, int src_stride,
+                int src_width, int src_height,
+                uint8* dst, int dst_stride,
+                int dst_width, int dst_height,
+                enum FilterMode filtering);
+
+LIBYUV_API
+void ScalePlane_16(const uint16* src, int src_stride,
+                   int src_width, int src_height,
+                   uint16* dst, int dst_stride,
+                   int dst_width, int dst_height,
+                   enum FilterMode filtering);
+
+// Scales a YUV 4:2:0 image from the src width and height to the
+// dst width and height.
+// If filtering is kFilterNone, a simple nearest-neighbor algorithm is
+// used. This produces basic (blocky) quality at the fastest speed.
+// If filtering is kFilterBilinear, interpolation is used to produce a better
+// quality image, at the expense of speed.
+// If filtering is kFilterBox, averaging is used to produce ever better
+// quality image, at further expense of speed.
+// Returns 0 if successful.
+
+LIBYUV_API
+int I420Scale(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              int src_width, int src_height,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int dst_width, int dst_height,
+              enum FilterMode filtering);
+
+LIBYUV_API
+int I420Scale_16(const uint16* src_y, int src_stride_y,
+                 const uint16* src_u, int src_stride_u,
+                 const uint16* src_v, int src_stride_v,
+                 int src_width, int src_height,
+                 uint16* dst_y, int dst_stride_y,
+                 uint16* dst_u, int dst_stride_u,
+                 uint16* dst_v, int dst_stride_v,
+                 int dst_width, int dst_height,
+                 enum FilterMode filtering);
+
+#ifdef __cplusplus
+// Legacy API.  Deprecated.
+LIBYUV_API
+int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
+          int src_stride_y, int src_stride_u, int src_stride_v,
+          int src_width, int src_height,
+          uint8* dst_y, uint8* dst_u, uint8* dst_v,
+          int dst_stride_y, int dst_stride_u, int dst_stride_v,
+          int dst_width, int dst_height,
+          LIBYUV_BOOL interpolate);
+
+// Legacy API.  Deprecated.
+LIBYUV_API
+int ScaleOffset(const uint8* src_i420, int src_width, int src_height,
+                uint8* dst_i420, int dst_width, int dst_height, int dst_yoffset,
+                LIBYUV_BOOL interpolate);
+
+// For testing, allow disabling of specialized scalers.
+LIBYUV_API
+void SetUseReferenceImpl(LIBYUV_BOOL use);
+#endif  // __cplusplus
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/scale_argb.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/scale_argb.h
new file mode 100644
index 0000000..0c9b362
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/scale_argb.h
@@ -0,0 +1,57 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_ARGB_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/scale.h"  // For FilterMode
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+LIBYUV_API
+int ARGBScale(const uint8* src_argb, int src_stride_argb,
+              int src_width, int src_height,
+              uint8* dst_argb, int dst_stride_argb,
+              int dst_width, int dst_height,
+              enum FilterMode filtering);
+
+// Clipped scale takes destination rectangle coordinates for clip values.
+LIBYUV_API
+int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
+                  int src_width, int src_height,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int dst_width, int dst_height,
+                  int clip_x, int clip_y, int clip_width, int clip_height,
+                  enum FilterMode filtering);
+
+// TODO(fbarchard): Implement this.
+// Scale with YUV conversion to ARGB and clipping.
+LIBYUV_API
+int YUVToARGBScaleClip(const uint8* src_y, int src_stride_y,
+                       const uint8* src_u, int src_stride_u,
+                       const uint8* src_v, int src_stride_v,
+                       uint32 src_fourcc,
+                       int src_width, int src_height,
+                       uint8* dst_argb, int dst_stride_argb,
+                       uint32 dst_fourcc,
+                       int dst_width, int dst_height,
+                       int clip_x, int clip_y, int clip_width, int clip_height,
+                       enum FilterMode filtering);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_ARGB_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/scale_row.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/scale_row.h
new file mode 100644
index 0000000..94ad9cf
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/scale_row.h
@@ -0,0 +1,478 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_ROW_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/scale.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if defined(__pnacl__) || defined(__CLR_VER) || \
+    (defined(__i386__) && !defined(__SSE2__))
+#define LIBYUV_DISABLE_X86
+#endif
+
+// Visual C 2012 required for AVX2.
+#if defined(_M_IX86) && !defined(__clang__) && \
+    defined(_MSC_VER) && _MSC_VER >= 1700
+#define VISUALC_HAS_AVX2 1
+#endif  // VisualStudio >= 2012
+
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_FIXEDDIV1_X86
+#define HAS_FIXEDDIV_X86
+#define HAS_SCALEARGBCOLS_SSE2
+#define HAS_SCALEARGBCOLSUP2_SSE2
+#define HAS_SCALEARGBFILTERCOLS_SSSE3
+#define HAS_SCALEARGBROWDOWN2_SSE2
+#define HAS_SCALEARGBROWDOWNEVEN_SSE2
+#define HAS_SCALECOLSUP2_SSE2
+#define HAS_SCALEFILTERCOLS_SSSE3
+#define HAS_SCALEROWDOWN2_SSE2
+#define HAS_SCALEROWDOWN34_SSSE3
+#define HAS_SCALEROWDOWN38_SSSE3
+#define HAS_SCALEROWDOWN4_SSE2
+#endif
+
+// The following are available on VS2012:
+#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
+#define HAS_SCALEADDROW_AVX2
+#define HAS_SCALEROWDOWN2_AVX2
+#define HAS_SCALEROWDOWN4_AVX2
+#endif
+
+// The following are available on Visual C:
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && !defined(__clang__)
+#define HAS_SCALEADDROW_SSE2
+#endif
+
+// The following are available on Neon platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
+#define HAS_SCALEARGBCOLS_NEON
+#define HAS_SCALEARGBROWDOWN2_NEON
+#define HAS_SCALEARGBROWDOWNEVEN_NEON
+#define HAS_SCALEFILTERCOLS_NEON
+#define HAS_SCALEROWDOWN2_NEON
+#define HAS_SCALEROWDOWN34_NEON
+#define HAS_SCALEROWDOWN38_NEON
+#define HAS_SCALEROWDOWN4_NEON
+#define HAS_SCALEARGBFILTERCOLS_NEON
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
+    defined(__mips__) && defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_SCALEROWDOWN2_MIPS_DSPR2
+#define HAS_SCALEROWDOWN4_MIPS_DSPR2
+#define HAS_SCALEROWDOWN34_MIPS_DSPR2
+#define HAS_SCALEROWDOWN38_MIPS_DSPR2
+#endif
+
+// Scale ARGB vertically with bilinear interpolation.
+void ScalePlaneVertical(int src_height,
+                        int dst_width, int dst_height,
+                        int src_stride, int dst_stride,
+                        const uint8* src_argb, uint8* dst_argb,
+                        int x, int y, int dy,
+                        int bpp, enum FilterMode filtering);
+
+void ScalePlaneVertical_16(int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint16* src_argb, uint16* dst_argb,
+                           int x, int y, int dy,
+                           int wpp, enum FilterMode filtering);
+
+// Simplify the filtering based on scale factors.
+enum FilterMode ScaleFilterReduce(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  enum FilterMode filtering);
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_C(int num, int div);
+int FixedDiv_X86(int num, int div);
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
+int FixedDiv1_C(int num, int div);
+int FixedDiv1_X86(int num, int div);
+#ifdef HAS_FIXEDDIV_X86
+#define FixedDiv FixedDiv_X86
+#define FixedDiv1 FixedDiv1_X86
+#else
+#define FixedDiv FixedDiv_C
+#define FixedDiv1 FixedDiv1_C
+#endif
+
+// Compute slope values for stepping.
+void ScaleSlope(int src_width, int src_height,
+                int dst_width, int dst_height,
+                enum FilterMode filtering,
+                int* x, int* y, int* dx, int* dy);
+
+void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width);
+void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width);
+void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width);
+void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                              uint16* dst, int dst_width);
+void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width);
+void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width);
+void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width);
+void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width);
+void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width);
+void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width);
+void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width);
+void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width);
+void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width);
+void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width);
+void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                 int dst_width, int x, int dx);
+void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                    int dst_width, int x, int dx);
+void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
+                    int dst_width, int, int);
+void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int, int);
+void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx);
+void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                          int dst_width, int x, int dx);
+void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
+                         int dst_width, int x, int dx);
+void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                            int dst_width, int x, int dx);
+void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width);
+void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width);
+void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width);
+void ScaleAddRow_C(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+void ScaleAddRow_16_C(const uint16* src_ptr, uint32* dst_ptr, int src_width);
+void ScaleARGBRowDown2_C(const uint8* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            int src_stepx,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
+                     int dst_width, int x, int dx);
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x, int dx);
+void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int, int);
+void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx);
+void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
+                             int dst_width, int x, int dx);
+
+// Specialized scalers for x86.
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+void ScaleRowDown34_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+
+void ScaleAddRow_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+void ScaleAddRow_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+void ScaleAddRow_Any_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+void ScaleAddRow_Any_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx);
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx);
+
+
+// ARGB Column functions
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx);
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx);
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx);
+void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                              int dst_width, int x, int dx);
+void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx);
+void ScaleARGBFilterCols_Any_NEON(uint8* dst_argb, const uint8* src_argb,
+                                  int dst_width, int x, int dx);
+void ScaleARGBCols_Any_NEON(uint8* dst_argb, const uint8* src_argb,
+                            int dst_width, int x, int dx);
+
+// ARGB Row functions
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width);
+void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleARGBRowDown2_Any_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_Any_SSE2(const uint8* src_argb,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_Any_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                   uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                uint8* dst, int dst_width);
+void ScaleARGBRowDown2Linear_Any_NEON(const uint8* src_argb,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                   uint8* dst, int dst_width);
+
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_Any_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                   int src_stepx,
+                                   uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_Any_SSE2(const uint8* src_argb,
+                                      ptrdiff_t src_stride,
+                                      int src_stepx,
+                                      uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_Any_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                   int src_stepx,
+                                   uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_Any_NEON(const uint8* src_argb,
+                                      ptrdiff_t src_stride,
+                                      int src_stepx,
+                                      uint8* dst_argb, int dst_width);
+
+// ScaleRowDown2Box also used by planar functions
+// NEON downscalers with interpolation.
+
+// Note - not static due to reuse in convert for 444 to 420.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width);
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+//  to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width);
+// 32x3 -> 12x1
+void ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+void ScaleRowDown2_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width);
+void ScaleRowDown2Linear_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst, int dst_width);
+void ScaleRowDown2Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown4_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                             uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                   uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                   uint8* dst_ptr, int dst_width);
+// 32 -> 12
+void ScaleRowDown38_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                             uint8* dst_ptr, int dst_width);
+// 32x3 -> 12x1
+void ScaleRowDown38_3_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+void ScaleAddRow_NEON(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+void ScaleAddRow_Any_NEON(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+
+void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                          int dst_width, int x, int dx);
+
+void ScaleFilterCols_Any_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                              int dst_width, int x, int dx);
+
+
+void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width);
+void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width);
+void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width);
+void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width);
+void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_ROW_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/version.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/version.h
new file mode 100644
index 0000000..9d1d746
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/version.h
@@ -0,0 +1,16 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
+#define INCLUDE_LIBYUV_VERSION_H_
+
+#define LIBYUV_VERSION 1456
+
+#endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/include/libyuv/video_common.h b/libvpx/libvpx/third_party/libyuv/include/libyuv/video_common.h
new file mode 100644
index 0000000..cb6582f
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/include/libyuv/video_common.h
@@ -0,0 +1,182 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Common definitions for video, including fourcc and VideoFormat.
+
+#ifndef INCLUDE_LIBYUV_VIDEO_COMMON_H_  // NOLINT
+#define INCLUDE_LIBYUV_VIDEO_COMMON_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+// Definition of FourCC codes
+//////////////////////////////////////////////////////////////////////////////
+
+// Convert four characters to a FourCC code.
+// Needs to be a macro otherwise the OS X compiler complains when the kFormat*
+// constants are used in a switch.
+#ifdef __cplusplus
+#define FOURCC(a, b, c, d) ( \
+    (static_cast<uint32>(a)) | (static_cast<uint32>(b) << 8) | \
+    (static_cast<uint32>(c) << 16) | (static_cast<uint32>(d) << 24))
+#else
+#define FOURCC(a, b, c, d) ( \
+    ((uint32)(a)) | ((uint32)(b) << 8) | /* NOLINT */ \
+    ((uint32)(c) << 16) | ((uint32)(d) << 24))  /* NOLINT */
+#endif
+
+// Some pages discussing FourCC codes:
+//   http://www.fourcc.org/yuv.php
+//   http://v4l2spec.bytesex.org/spec/book1.htm
+//   http://developer.apple.com/quicktime/icefloe/dispatch020.html
+//   http://msdn.microsoft.com/library/windows/desktop/dd206750.aspx#nv12
+//   http://people.xiph.org/~xiphmont/containers/nut/nut4cc.txt
+
+// FourCC codes grouped according to implementation efficiency.
+// Primary formats should convert in 1 efficient step.
+// Secondary formats are converted in 2 steps.
+// Auxilliary formats call primary converters.
+enum FourCC {
+  // 9 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
+  FOURCC_I420 = FOURCC('I', '4', '2', '0'),
+  FOURCC_I422 = FOURCC('I', '4', '2', '2'),
+  FOURCC_I444 = FOURCC('I', '4', '4', '4'),
+  FOURCC_I411 = FOURCC('I', '4', '1', '1'),
+  FOURCC_I400 = FOURCC('I', '4', '0', '0'),
+  FOURCC_NV21 = FOURCC('N', 'V', '2', '1'),
+  FOURCC_NV12 = FOURCC('N', 'V', '1', '2'),
+  FOURCC_YUY2 = FOURCC('Y', 'U', 'Y', '2'),
+  FOURCC_UYVY = FOURCC('U', 'Y', 'V', 'Y'),
+
+  // 2 Secondary YUV formats: row biplanar.
+  FOURCC_M420 = FOURCC('M', '4', '2', '0'),
+  FOURCC_Q420 = FOURCC('Q', '4', '2', '0'), // deprecated.
+
+  // 9 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp.
+  FOURCC_ARGB = FOURCC('A', 'R', 'G', 'B'),
+  FOURCC_BGRA = FOURCC('B', 'G', 'R', 'A'),
+  FOURCC_ABGR = FOURCC('A', 'B', 'G', 'R'),
+  FOURCC_24BG = FOURCC('2', '4', 'B', 'G'),
+  FOURCC_RAW  = FOURCC('r', 'a', 'w', ' '),
+  FOURCC_RGBA = FOURCC('R', 'G', 'B', 'A'),
+  FOURCC_RGBP = FOURCC('R', 'G', 'B', 'P'),  // rgb565 LE.
+  FOURCC_RGBO = FOURCC('R', 'G', 'B', 'O'),  // argb1555 LE.
+  FOURCC_R444 = FOURCC('R', '4', '4', '4'),  // argb4444 LE.
+
+  // 4 Secondary RGB formats: 4 Bayer Patterns. deprecated.
+  FOURCC_RGGB = FOURCC('R', 'G', 'G', 'B'),
+  FOURCC_BGGR = FOURCC('B', 'G', 'G', 'R'),
+  FOURCC_GRBG = FOURCC('G', 'R', 'B', 'G'),
+  FOURCC_GBRG = FOURCC('G', 'B', 'R', 'G'),
+
+  // 1 Primary Compressed YUV format.
+  FOURCC_MJPG = FOURCC('M', 'J', 'P', 'G'),
+
+  // 5 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
+  FOURCC_YV12 = FOURCC('Y', 'V', '1', '2'),
+  FOURCC_YV16 = FOURCC('Y', 'V', '1', '6'),
+  FOURCC_YV24 = FOURCC('Y', 'V', '2', '4'),
+  FOURCC_YU12 = FOURCC('Y', 'U', '1', '2'),  // Linux version of I420.
+  FOURCC_J420 = FOURCC('J', '4', '2', '0'),
+  FOURCC_J400 = FOURCC('J', '4', '0', '0'),
+
+  // 14 Auxiliary aliases.  CanonicalFourCC() maps these to canonical fourcc.
+  FOURCC_IYUV = FOURCC('I', 'Y', 'U', 'V'),  // Alias for I420.
+  FOURCC_YU16 = FOURCC('Y', 'U', '1', '6'),  // Alias for I422.
+  FOURCC_YU24 = FOURCC('Y', 'U', '2', '4'),  // Alias for I444.
+  FOURCC_YUYV = FOURCC('Y', 'U', 'Y', 'V'),  // Alias for YUY2.
+  FOURCC_YUVS = FOURCC('y', 'u', 'v', 's'),  // Alias for YUY2 on Mac.
+  FOURCC_HDYC = FOURCC('H', 'D', 'Y', 'C'),  // Alias for UYVY.
+  FOURCC_2VUY = FOURCC('2', 'v', 'u', 'y'),  // Alias for UYVY on Mac.
+  FOURCC_JPEG = FOURCC('J', 'P', 'E', 'G'),  // Alias for MJPG.
+  FOURCC_DMB1 = FOURCC('d', 'm', 'b', '1'),  // Alias for MJPG on Mac.
+  FOURCC_BA81 = FOURCC('B', 'A', '8', '1'),  // Alias for BGGR.
+  FOURCC_RGB3 = FOURCC('R', 'G', 'B', '3'),  // Alias for RAW.
+  FOURCC_BGR3 = FOURCC('B', 'G', 'R', '3'),  // Alias for 24BG.
+  FOURCC_CM32 = FOURCC(0, 0, 0, 32),  // Alias for BGRA kCMPixelFormat_32ARGB
+  FOURCC_CM24 = FOURCC(0, 0, 0, 24),  // Alias for RAW kCMPixelFormat_24RGB
+  FOURCC_L555 = FOURCC('L', '5', '5', '5'),  // Alias for RGBO.
+  FOURCC_L565 = FOURCC('L', '5', '6', '5'),  // Alias for RGBP.
+  FOURCC_5551 = FOURCC('5', '5', '5', '1'),  // Alias for RGBO.
+
+  // 1 Auxiliary compressed YUV format set aside for capturer.
+  FOURCC_H264 = FOURCC('H', '2', '6', '4'),
+
+  // Match any fourcc.
+  FOURCC_ANY = -1,
+};
+
+enum FourCCBpp {
+  // Canonical fourcc codes used in our code.
+  FOURCC_BPP_I420 = 12,
+  FOURCC_BPP_I422 = 16,
+  FOURCC_BPP_I444 = 24,
+  FOURCC_BPP_I411 = 12,
+  FOURCC_BPP_I400 = 8,
+  FOURCC_BPP_NV21 = 12,
+  FOURCC_BPP_NV12 = 12,
+  FOURCC_BPP_YUY2 = 16,
+  FOURCC_BPP_UYVY = 16,
+  FOURCC_BPP_M420 = 12,
+  FOURCC_BPP_Q420 = 12,
+  FOURCC_BPP_ARGB = 32,
+  FOURCC_BPP_BGRA = 32,
+  FOURCC_BPP_ABGR = 32,
+  FOURCC_BPP_RGBA = 32,
+  FOURCC_BPP_24BG = 24,
+  FOURCC_BPP_RAW  = 24,
+  FOURCC_BPP_RGBP = 16,
+  FOURCC_BPP_RGBO = 16,
+  FOURCC_BPP_R444 = 16,
+  FOURCC_BPP_RGGB = 8,
+  FOURCC_BPP_BGGR = 8,
+  FOURCC_BPP_GRBG = 8,
+  FOURCC_BPP_GBRG = 8,
+  FOURCC_BPP_YV12 = 12,
+  FOURCC_BPP_YV16 = 16,
+  FOURCC_BPP_YV24 = 24,
+  FOURCC_BPP_YU12 = 12,
+  FOURCC_BPP_J420 = 12,
+  FOURCC_BPP_J400 = 8,
+  FOURCC_BPP_MJPG = 0,  // 0 means unknown.
+  FOURCC_BPP_H264 = 0,
+  FOURCC_BPP_IYUV = 12,
+  FOURCC_BPP_YU16 = 16,
+  FOURCC_BPP_YU24 = 24,
+  FOURCC_BPP_YUYV = 16,
+  FOURCC_BPP_YUVS = 16,
+  FOURCC_BPP_HDYC = 16,
+  FOURCC_BPP_2VUY = 16,
+  FOURCC_BPP_JPEG = 1,
+  FOURCC_BPP_DMB1 = 1,
+  FOURCC_BPP_BA81 = 8,
+  FOURCC_BPP_RGB3 = 24,
+  FOURCC_BPP_BGR3 = 24,
+  FOURCC_BPP_CM32 = 32,
+  FOURCC_BPP_CM24 = 24,
+
+  // Match any fourcc.
+  FOURCC_BPP_ANY  = 0,  // 0 means unknown.
+};
+
+// Converts fourcc aliases into canonical ones.
+LIBYUV_API uint32 CanonicalFourCC(uint32 fourcc);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_VIDEO_COMMON_H_  NOLINT
diff --git a/libvpx/libvpx/third_party/libyuv/source/compare.cc b/libvpx/libvpx/third_party/libyuv/source/compare.cc
new file mode 100644
index 0000000..46aa847
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/compare.cc
@@ -0,0 +1,373 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/compare.h"
+
+#include <float.h>
+#include <math.h>
+#ifdef _OPENMP
+#include <omp.h>
+#endif
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// hash seed of 5381 recommended.
+// Internal C version of HashDjb2 with int sized count for efficiency.
+uint32 HashDjb2_C(const uint8* src, int count, uint32 seed);
+
+// This module is for Visual C x86
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__))))
+#define HAS_HASHDJB2_SSE41
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed);
+
+#ifdef VISUALC_HAS_AVX2
+#define HAS_HASHDJB2_AVX2
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed);
+#endif
+
+#endif  // HAS_HASHDJB2_SSE41
+
+// hash seed of 5381 recommended.
+LIBYUV_API
+uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
+  const int kBlockSize = 1 << 15;  // 32768;
+  int remainder;
+  uint32 (*HashDjb2_SSE)(const uint8* src, int count, uint32 seed) = HashDjb2_C;
+#if defined(HAS_HASHDJB2_SSE41)
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    HashDjb2_SSE = HashDjb2_SSE41;
+  }
+#endif
+#if defined(HAS_HASHDJB2_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    HashDjb2_SSE = HashDjb2_AVX2;
+  }
+#endif
+
+  while (count >= (uint64)(kBlockSize)) {
+    seed = HashDjb2_SSE(src, kBlockSize, seed);
+    src += kBlockSize;
+    count -= kBlockSize;
+  }
+  remainder = (int)(count) & ~15;
+  if (remainder) {
+    seed = HashDjb2_SSE(src, remainder, seed);
+    src += remainder;
+    count -= remainder;
+  }
+  remainder = (int)(count) & 15;
+  if (remainder) {
+    seed = HashDjb2_C(src, remainder, seed);
+  }
+  return seed;
+}
+
+static uint32 ARGBDetectRow_C(const uint8* argb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    if (argb[0] != 255) {  // First byte is not Alpha of 255, so not ARGB.
+      return FOURCC_BGRA;
+    }
+    if (argb[3] != 255) {  // 4th byte is not Alpha of 255, so not BGRA.
+      return FOURCC_ARGB;
+    }
+    if (argb[4] != 255) {  // Second pixel first byte is not Alpha of 255.
+      return FOURCC_BGRA;
+    }
+    if (argb[7] != 255) {  // Second pixel 4th byte is not Alpha of 255.
+      return FOURCC_ARGB;
+    }
+    argb += 8;
+  }
+  if (width & 1) {
+    if (argb[0] != 255) {  // First byte is not Alpha of 255, so not ARGB.
+      return FOURCC_BGRA;
+    }
+    if (argb[3] != 255) {  // 4th byte is not Alpha of 255, so not BGRA.
+      return FOURCC_ARGB;
+    }
+  }
+  return 0;
+}
+
+// Scan an opaque argb image and return fourcc based on alpha offset.
+// Returns FOURCC_ARGB, FOURCC_BGRA, or 0 if unknown.
+LIBYUV_API
+uint32 ARGBDetect(const uint8* argb, int stride_argb, int width, int height) {
+  uint32 fourcc = 0;
+  int h;
+
+  // Coalesce rows.
+  if (stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    stride_argb = 0;
+  }
+  for (h = 0; h < height && fourcc == 0; ++h) {
+    fourcc = ARGBDetectRow_C(argb, width);
+    argb += stride_argb;
+  }
+  return fourcc;
+}
+
+uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count);
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
+#define HAS_SUMSQUAREERROR_NEON
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count);
+#endif
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_SUMSQUAREERROR_SSE2
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count);
+#endif
+
+#ifdef VISUALC_HAS_AVX2
+#define HAS_SUMSQUAREERROR_AVX2
+uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count);
+#endif
+
+// TODO(fbarchard): Refactor into row function.
+LIBYUV_API
+uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b,
+                             int count) {
+  // SumSquareError returns values 0 to 65535 for each squared difference.
+  // Up to 65536 of those can be summed and remain within a uint32.
+  // After each block of 65536 pixels, accumulate into a uint64.
+  const int kBlockSize = 65536;
+  int remainder = count & (kBlockSize - 1) & ~31;
+  uint64 sse = 0;
+  int i;
+  uint32 (*SumSquareError)(const uint8* src_a, const uint8* src_b, int count) =
+      SumSquareError_C;
+#if defined(HAS_SUMSQUAREERROR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SumSquareError = SumSquareError_NEON;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    // Note only used for multiples of 16 so count is not checked.
+    SumSquareError = SumSquareError_SSE2;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    // Note only used for multiples of 32 so count is not checked.
+    SumSquareError = SumSquareError_AVX2;
+  }
+#endif
+#ifdef _OPENMP
+#pragma omp parallel for reduction(+: sse)
+#endif
+  for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
+    sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
+  }
+  src_a += count & ~(kBlockSize - 1);
+  src_b += count & ~(kBlockSize - 1);
+  if (remainder) {
+    sse += SumSquareError(src_a, src_b, remainder);
+    src_a += remainder;
+    src_b += remainder;
+  }
+  remainder = count & 31;
+  if (remainder) {
+    sse += SumSquareError_C(src_a, src_b, remainder);
+  }
+  return sse;
+}
+
+LIBYUV_API
+uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
+                                  const uint8* src_b, int stride_b,
+                                  int width, int height) {
+  uint64 sse = 0;
+  int h;
+  // Coalesce rows.
+  if (stride_a == width &&
+      stride_b == width) {
+    width *= height;
+    height = 1;
+    stride_a = stride_b = 0;
+  }
+  for (h = 0; h < height; ++h) {
+    sse += ComputeSumSquareError(src_a, src_b, width);
+    src_a += stride_a;
+    src_b += stride_b;
+  }
+  return sse;
+}
+
+LIBYUV_API
+double SumSquareErrorToPsnr(uint64 sse, uint64 count) {
+  double psnr;
+  if (sse > 0) {
+    double mse = (double)(count) / (double)(sse);
+    psnr = 10.0 * log10(255.0 * 255.0 * mse);
+  } else {
+    psnr = kMaxPsnr;      // Limit to prevent divide by 0
+  }
+
+  if (psnr > kMaxPsnr)
+    psnr = kMaxPsnr;
+
+  return psnr;
+}
+
+LIBYUV_API
+double CalcFramePsnr(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height) {
+  const uint64 samples = width * height;
+  const uint64 sse = ComputeSumSquareErrorPlane(src_a, stride_a,
+                                                src_b, stride_b,
+                                                width, height);
+  return SumSquareErrorToPsnr(sse, samples);
+}
+
+LIBYUV_API
+double I420Psnr(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height) {
+  const uint64 sse_y = ComputeSumSquareErrorPlane(src_y_a, stride_y_a,
+                                                  src_y_b, stride_y_b,
+                                                  width, height);
+  const int width_uv = (width + 1) >> 1;
+  const int height_uv = (height + 1) >> 1;
+  const uint64 sse_u = ComputeSumSquareErrorPlane(src_u_a, stride_u_a,
+                                                  src_u_b, stride_u_b,
+                                                  width_uv, height_uv);
+  const uint64 sse_v = ComputeSumSquareErrorPlane(src_v_a, stride_v_a,
+                                                  src_v_b, stride_v_b,
+                                                  width_uv, height_uv);
+  const uint64 samples = width * height + 2 * (width_uv * height_uv);
+  const uint64 sse = sse_y + sse_u + sse_v;
+  return SumSquareErrorToPsnr(sse, samples);
+}
+
+static const int64 cc1 =  26634;  // (64^2*(.01*255)^2
+static const int64 cc2 = 239708;  // (64^2*(.03*255)^2
+
+static double Ssim8x8_C(const uint8* src_a, int stride_a,
+                        const uint8* src_b, int stride_b) {
+  int64 sum_a = 0;
+  int64 sum_b = 0;
+  int64 sum_sq_a = 0;
+  int64 sum_sq_b = 0;
+  int64 sum_axb = 0;
+
+  int i;
+  for (i = 0; i < 8; ++i) {
+    int j;
+    for (j = 0; j < 8; ++j) {
+      sum_a += src_a[j];
+      sum_b += src_b[j];
+      sum_sq_a += src_a[j] * src_a[j];
+      sum_sq_b += src_b[j] * src_b[j];
+      sum_axb += src_a[j] * src_b[j];
+    }
+
+    src_a += stride_a;
+    src_b += stride_b;
+  }
+
+  {
+    const int64 count = 64;
+    // scale the constants by number of pixels
+    const int64 c1 = (cc1 * count * count) >> 12;
+    const int64 c2 = (cc2 * count * count) >> 12;
+
+    const int64 sum_a_x_sum_b = sum_a * sum_b;
+
+    const int64 ssim_n = (2 * sum_a_x_sum_b + c1) *
+                         (2 * count * sum_axb - 2 * sum_a_x_sum_b + c2);
+
+    const int64 sum_a_sq = sum_a*sum_a;
+    const int64 sum_b_sq = sum_b*sum_b;
+
+    const int64 ssim_d = (sum_a_sq + sum_b_sq + c1) *
+                         (count * sum_sq_a - sum_a_sq +
+                          count * sum_sq_b - sum_b_sq + c2);
+
+    if (ssim_d == 0.0) {
+      return DBL_MAX;
+    }
+    return ssim_n * 1.0 / ssim_d;
+  }
+}
+
+// We are using a 8x8 moving window with starting location of each 8x8 window
+// on the 4x4 pixel grid. Such arrangement allows the windows to overlap
+// block boundaries to penalize blocking artifacts.
+LIBYUV_API
+double CalcFrameSsim(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height) {
+  int samples = 0;
+  double ssim_total = 0;
+  double (*Ssim8x8)(const uint8* src_a, int stride_a,
+                    const uint8* src_b, int stride_b) = Ssim8x8_C;
+
+  // sample point start with each 4x4 location
+  int i;
+  for (i = 0; i < height - 8; i += 4) {
+    int j;
+    for (j = 0; j < width - 8; j += 4) {
+      ssim_total += Ssim8x8(src_a + j, stride_a, src_b + j, stride_b);
+      samples++;
+    }
+
+    src_a += stride_a * 4;
+    src_b += stride_b * 4;
+  }
+
+  ssim_total /= samples;
+  return ssim_total;
+}
+
+LIBYUV_API
+double I420Ssim(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height) {
+  const double ssim_y = CalcFrameSsim(src_y_a, stride_y_a,
+                                      src_y_b, stride_y_b, width, height);
+  const int width_uv = (width + 1) >> 1;
+  const int height_uv = (height + 1) >> 1;
+  const double ssim_u = CalcFrameSsim(src_u_a, stride_u_a,
+                                      src_u_b, stride_u_b,
+                                      width_uv, height_uv);
+  const double ssim_v = CalcFrameSsim(src_v_a, stride_v_a,
+                                      src_v_b, stride_v_b,
+                                      width_uv, height_uv);
+  return ssim_y * 0.8 + 0.1 * (ssim_u + ssim_v);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/compare_common.cc b/libvpx/libvpx/third_party/libyuv/source/compare_common.cc
new file mode 100644
index 0000000..c546b51
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/compare_common.cc
@@ -0,0 +1,42 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse = 0u;
+  int i;
+  for (i = 0; i < count; ++i) {
+    int diff = src_a[i] - src_b[i];
+    sse += (uint32)(diff * diff);
+  }
+  return sse;
+}
+
+// hash seed of 5381 recommended.
+// Internal C version of HashDjb2 with int sized count for efficiency.
+uint32 HashDjb2_C(const uint8* src, int count, uint32 seed) {
+  uint32 hash = seed;
+  int i;
+  for (i = 0; i < count; ++i) {
+    hash += (hash << 5) + src[i];
+  }
+  return hash;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/compare_gcc.cc b/libvpx/libvpx/third_party/libyuv/source/compare_gcc.cc
new file mode 100644
index 0000000..247cb33
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/compare_gcc.cc
@@ -0,0 +1,152 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse;
+  asm volatile (  // NOLINT
+    "pxor      %%xmm0,%%xmm0                   \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10, 0) ",%0          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10, 1) ",%1          \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psubusb   %%xmm2,%%xmm1                   \n"
+    "psubusb   %%xmm3,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpckhbw %%xmm5,%%xmm2                   \n"
+    "pmaddwd   %%xmm1,%%xmm1                   \n"
+    "pmaddwd   %%xmm2,%%xmm2                   \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+
+    "pshufd    $0xee,%%xmm0,%%xmm1             \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "pshufd    $0x1,%%xmm0,%%xmm1              \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "movd      %%xmm0,%3                       \n"
+
+  : "+r"(src_a),      // %0
+    "+r"(src_b),      // %1
+    "+r"(count),      // %2
+    "=g"(sse)         // %3
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );  // NOLINT
+  return sse;
+}
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
+#define HAS_HASHDJB2_SSE41
+static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
+static uvec32 kHashMul0 = {
+  0x0c3525e1,  // 33 ^ 15
+  0xa3476dc1,  // 33 ^ 14
+  0x3b4039a1,  // 33 ^ 13
+  0x4f5f0981,  // 33 ^ 12
+};
+static uvec32 kHashMul1 = {
+  0x30f35d61,  // 33 ^ 11
+  0x855cb541,  // 33 ^ 10
+  0x040a9121,  // 33 ^ 9
+  0x747c7101,  // 33 ^ 8
+};
+static uvec32 kHashMul2 = {
+  0xec41d4e1,  // 33 ^ 7
+  0x4cfa3cc1,  // 33 ^ 6
+  0x025528a1,  // 33 ^ 5
+  0x00121881,  // 33 ^ 4
+};
+static uvec32 kHashMul3 = {
+  0x00008c61,  // 33 ^ 3
+  0x00000441,  // 33 ^ 2
+  0x00000021,  // 33 ^ 1
+  0x00000001,  // 33 ^ 0
+};
+
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+  uint32 hash;
+  asm volatile (  // NOLINT
+    "movd      %2,%%xmm0                       \n"
+    "pxor      %%xmm7,%%xmm7                   \n"
+    "movdqa    %4,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10, 0) ",%0          \n"
+    "pmulld    %%xmm6,%%xmm0                   \n"
+    "movdqa    %5,%%xmm5                       \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm7,%%xmm3                   \n"
+    "pmulld    %%xmm5,%%xmm3                   \n"
+    "movdqa    %6,%%xmm5                       \n"
+    "movdqa    %%xmm2,%%xmm4                   \n"
+    "punpckhwd %%xmm7,%%xmm4                   \n"
+    "pmulld    %%xmm5,%%xmm4                   \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "punpckhbw %%xmm7,%%xmm1                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm7,%%xmm2                   \n"
+    "pmulld    %%xmm5,%%xmm2                   \n"
+    "movdqa    %8,%%xmm5                       \n"
+    "punpckhwd %%xmm7,%%xmm1                   \n"
+    "pmulld    %%xmm5,%%xmm1                   \n"
+    "paddd     %%xmm4,%%xmm3                   \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm1                   \n"
+    "pshufd    $0xe,%%xmm1,%%xmm2              \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "pshufd    $0x1,%%xmm1,%%xmm2              \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%1                        \n"
+    "jg        1b                              \n"
+    "movd      %%xmm0,%3                       \n"
+  : "+r"(src),        // %0
+    "+r"(count),      // %1
+    "+rm"(seed),      // %2
+    "=g"(hash)        // %3
+  : "m"(kHash16x33),  // %4
+    "m"(kHashMul0),   // %5
+    "m"(kHashMul1),   // %6
+    "m"(kHashMul2),   // %7
+    "m"(kHashMul3)    // %8
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );  // NOLINT
+  return hash;
+}
+#endif  // defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/libvpx/libvpx/third_party/libyuv/source/compare_neon.cc b/libvpx/libvpx/third_party/libyuv/source/compare_neon.cc
new file mode 100644
index 0000000..ef006ec
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/compare_neon.cc
@@ -0,0 +1,65 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  volatile uint32 sse;
+  asm volatile (
+    "vmov.u8    q8, #0                         \n"
+    "vmov.u8    q10, #0                        \n"
+    "vmov.u8    q9, #0                         \n"
+    "vmov.u8    q11, #0                        \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"
+    "subs       %2, %2, #16                    \n"
+    "vsubl.u8   q2, d0, d2                     \n"
+    "vsubl.u8   q3, d1, d3                     \n"
+    "vmlal.s16  q8, d4, d4                     \n"
+    "vmlal.s16  q9, d6, d6                     \n"
+    "vmlal.s16  q10, d5, d5                    \n"
+    "vmlal.s16  q11, d7, d7                    \n"
+    "bgt        1b                             \n"
+
+    "vadd.u32   q8, q8, q9                     \n"
+    "vadd.u32   q10, q10, q11                  \n"
+    "vadd.u32   q11, q8, q10                   \n"
+    "vpaddl.u32 q1, q11                        \n"
+    "vadd.u64   d0, d2, d3                     \n"
+    "vmov.32    %3, d0[0]                      \n"
+    : "+r"(src_a),
+      "+r"(src_b),
+      "+r"(count),
+      "=r"(sse)
+    :
+    : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
+  return sse;
+}
+
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/compare_neon64.cc b/libvpx/libvpx/third_party/libyuv/source/compare_neon64.cc
new file mode 100644
index 0000000..6d1e5e1
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/compare_neon64.cc
@@ -0,0 +1,63 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  volatile uint32 sse;
+  asm volatile (
+    "eor        v16.16b, v16.16b, v16.16b      \n"
+    "eor        v18.16b, v18.16b, v18.16b      \n"
+    "eor        v17.16b, v17.16b, v17.16b      \n"
+    "eor        v19.16b, v19.16b, v19.16b      \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"
+    MEMACCESS(1)
+    "ld1        {v1.16b}, [%1], #16            \n"
+    "subs       %w2, %w2, #16                  \n"
+    "usubl      v2.8h, v0.8b, v1.8b            \n"
+    "usubl2     v3.8h, v0.16b, v1.16b          \n"
+    "smlal      v16.4s, v2.4h, v2.4h           \n"
+    "smlal      v17.4s, v3.4h, v3.4h           \n"
+    "smlal2     v18.4s, v2.8h, v2.8h           \n"
+    "smlal2     v19.4s, v3.8h, v3.8h           \n"
+    "b.gt       1b                             \n"
+
+    "add        v16.4s, v16.4s, v17.4s         \n"
+    "add        v18.4s, v18.4s, v19.4s         \n"
+    "add        v19.4s, v16.4s, v18.4s         \n"
+    "addv       s0, v19.4s                     \n"
+    "fmov       %w3, s0                        \n"
+    : "+r"(src_a),
+      "+r"(src_b),
+      "+r"(count),
+      "=r"(sse)
+    :
+    : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
+  return sse;
+}
+
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/compare_win.cc b/libvpx/libvpx/third_party/libyuv/source/compare_win.cc
new file mode 100644
index 0000000..19806f2
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/compare_win.cc
@@ -0,0 +1,229 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
+    defined(_MSC_VER) && !defined(__clang__)
+
+__declspec(naked)
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+  __asm {
+    mov        eax, [esp + 4]    // src_a
+    mov        edx, [esp + 8]    // src_b
+    mov        ecx, [esp + 12]   // count
+    pxor       xmm0, xmm0
+    pxor       xmm5, xmm5
+
+  wloop:
+    movdqu     xmm1, [eax]
+    lea        eax,  [eax + 16]
+    movdqu     xmm2, [edx]
+    lea        edx,  [edx + 16]
+    movdqa     xmm3, xmm1  // abs trick
+    psubusb    xmm1, xmm2
+    psubusb    xmm2, xmm3
+    por        xmm1, xmm2
+    movdqa     xmm2, xmm1
+    punpcklbw  xmm1, xmm5
+    punpckhbw  xmm2, xmm5
+    pmaddwd    xmm1, xmm1
+    pmaddwd    xmm2, xmm2
+    paddd      xmm0, xmm1
+    paddd      xmm0, xmm2
+    sub        ecx, 16
+    jg         wloop
+
+    pshufd     xmm1, xmm0, 0xee
+    paddd      xmm0, xmm1
+    pshufd     xmm1, xmm0, 0x01
+    paddd      xmm0, xmm1
+    movd       eax, xmm0
+    ret
+  }
+}
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+// C4752: found Intel(R) Advanced Vector Extensions; consider using /arch:AVX.
+#pragma warning(disable: 4752)
+__declspec(naked)
+uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) {
+  __asm {
+    mov        eax, [esp + 4]    // src_a
+    mov        edx, [esp + 8]    // src_b
+    mov        ecx, [esp + 12]   // count
+    vpxor      ymm0, ymm0, ymm0  // sum
+    vpxor      ymm5, ymm5, ymm5  // constant 0 for unpck
+    sub        edx, eax
+
+  wloop:
+    vmovdqu    ymm1, [eax]
+    vmovdqu    ymm2, [eax + edx]
+    lea        eax,  [eax + 32]
+    vpsubusb   ymm3, ymm1, ymm2  // abs difference trick
+    vpsubusb   ymm2, ymm2, ymm1
+    vpor       ymm1, ymm2, ymm3
+    vpunpcklbw ymm2, ymm1, ymm5  // u16.  mutates order.
+    vpunpckhbw ymm1, ymm1, ymm5
+    vpmaddwd   ymm2, ymm2, ymm2  // square + hadd to u32.
+    vpmaddwd   ymm1, ymm1, ymm1
+    vpaddd     ymm0, ymm0, ymm1
+    vpaddd     ymm0, ymm0, ymm2
+    sub        ecx, 32
+    jg         wloop
+
+    vpshufd    ymm1, ymm0, 0xee  // 3, 2 + 1, 0 both lanes.
+    vpaddd     ymm0, ymm0, ymm1
+    vpshufd    ymm1, ymm0, 0x01  // 1 + 0 both lanes.
+    vpaddd     ymm0, ymm0, ymm1
+    vpermq     ymm1, ymm0, 0x02  // high + low lane.
+    vpaddd     ymm0, ymm0, ymm1
+    vmovd      eax, xmm0
+    vzeroupper
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+
+#define HAS_HASHDJB2_SSE41
+static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
+static uvec32 kHashMul0 = {
+  0x0c3525e1,  // 33 ^ 15
+  0xa3476dc1,  // 33 ^ 14
+  0x3b4039a1,  // 33 ^ 13
+  0x4f5f0981,  // 33 ^ 12
+};
+static uvec32 kHashMul1 = {
+  0x30f35d61,  // 33 ^ 11
+  0x855cb541,  // 33 ^ 10
+  0x040a9121,  // 33 ^ 9
+  0x747c7101,  // 33 ^ 8
+};
+static uvec32 kHashMul2 = {
+  0xec41d4e1,  // 33 ^ 7
+  0x4cfa3cc1,  // 33 ^ 6
+  0x025528a1,  // 33 ^ 5
+  0x00121881,  // 33 ^ 4
+};
+static uvec32 kHashMul3 = {
+  0x00008c61,  // 33 ^ 3
+  0x00000441,  // 33 ^ 2
+  0x00000021,  // 33 ^ 1
+  0x00000001,  // 33 ^ 0
+};
+
+// 27: 66 0F 38 40 C6     pmulld      xmm0,xmm6
+// 44: 66 0F 38 40 DD     pmulld      xmm3,xmm5
+// 59: 66 0F 38 40 E5     pmulld      xmm4,xmm5
+// 72: 66 0F 38 40 D5     pmulld      xmm2,xmm5
+// 83: 66 0F 38 40 CD     pmulld      xmm1,xmm5
+#define pmulld(reg) _asm _emit 0x66 _asm _emit 0x0F _asm _emit 0x38 \
+    _asm _emit 0x40 _asm _emit reg
+
+__declspec(naked)
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+
+    pxor       xmm7, xmm7        // constant 0 for unpck
+    movdqa     xmm6, kHash16x33
+
+  wloop:
+    movdqu     xmm1, [eax]       // src[0-15]
+    lea        eax, [eax + 16]
+    pmulld(0xc6)                 // pmulld      xmm0,xmm6  hash *= 33 ^ 16
+    movdqa     xmm5, kHashMul0
+    movdqa     xmm2, xmm1
+    punpcklbw  xmm2, xmm7        // src[0-7]
+    movdqa     xmm3, xmm2
+    punpcklwd  xmm3, xmm7        // src[0-3]
+    pmulld(0xdd)                 // pmulld     xmm3, xmm5
+    movdqa     xmm5, kHashMul1
+    movdqa     xmm4, xmm2
+    punpckhwd  xmm4, xmm7        // src[4-7]
+    pmulld(0xe5)                 // pmulld     xmm4, xmm5
+    movdqa     xmm5, kHashMul2
+    punpckhbw  xmm1, xmm7        // src[8-15]
+    movdqa     xmm2, xmm1
+    punpcklwd  xmm2, xmm7        // src[8-11]
+    pmulld(0xd5)                 // pmulld     xmm2, xmm5
+    movdqa     xmm5, kHashMul3
+    punpckhwd  xmm1, xmm7        // src[12-15]
+    pmulld(0xcd)                 // pmulld     xmm1, xmm5
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    paddd      xmm1, xmm3
+
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    sub        ecx, 16
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+__declspec(naked)
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+    movdqa     xmm6, kHash16x33
+
+  wloop:
+    vpmovzxbd  xmm3, dword ptr [eax]  // src[0-3]
+    pmulld     xmm0, xmm6  // hash *= 33 ^ 16
+    vpmovzxbd  xmm4, dword ptr [eax + 4]  // src[4-7]
+    pmulld     xmm3, kHashMul0
+    vpmovzxbd  xmm2, dword ptr [eax + 8]  // src[8-11]
+    pmulld     xmm4, kHashMul1
+    vpmovzxbd  xmm1, dword ptr [eax + 12]  // src[12-15]
+    pmulld     xmm2, kHashMul2
+    lea        eax, [eax + 16]
+    pmulld     xmm1, kHashMul3
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    paddd      xmm1, xmm3
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    sub        ecx, 16
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/convert.cc b/libvpx/libvpx/third_party/libyuv/source/convert.cc
new file mode 100644
index 0000000..3ad6bd7
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/convert.cc
@@ -0,0 +1,1389 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+#include "libyuv/scale.h"  // For ScalePlane()
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// Any I4xx To I420 format with mirroring.
+static int I4xxToI420(const uint8* src_y, int src_stride_y,
+                      const uint8* src_u, int src_stride_u,
+                      const uint8* src_v, int src_stride_v,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int src_y_width, int src_y_height,
+                      int src_uv_width, int src_uv_height) {
+  const int dst_y_width = Abs(src_y_width);
+  const int dst_y_height = Abs(src_y_height);
+  const int dst_uv_width = SUBSAMPLE(dst_y_width, 1, 1);
+  const int dst_uv_height = SUBSAMPLE(dst_y_height, 1, 1);
+  if (src_y_width == 0 || src_y_height == 0 ||
+      src_uv_width == 0 || src_uv_height == 0) {
+    return -1;
+  }
+  ScalePlane(src_y, src_stride_y, src_y_width, src_y_height,
+             dst_y, dst_stride_y, dst_y_width, dst_y_height,
+             kFilterBilinear);
+  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height,
+             dst_u, dst_stride_u, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height,
+             dst_v, dst_stride_v, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  return 0;
+}
+
+// Copy I420 with optional flipping
+// TODO(fbarchard): Use Scale plane which supports mirroring, but ensure
+// is does row coalescing.
+LIBYUV_API
+int I420Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  // Copy UV planes.
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
+  return 0;
+}
+
+// 422 chroma is 1/2 width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I422ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int src_uv_width = SUBSAMPLE(width, 1, 1);
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    src_uv_width, height);
+}
+
+// 444 chroma is 1x width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I444ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    width, height);
+}
+
+// 411 chroma is 1/4 width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I411ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int src_uv_width = SUBSAMPLE(width, 3, 2);
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    src_uv_width, height);
+}
+
+// I400 is greyscale typically used in MJPG
+LIBYUV_API
+int I400ToI420(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  SetPlane(dst_u, dst_stride_u, halfwidth, halfheight, 128);
+  SetPlane(dst_v, dst_stride_v, halfwidth, halfheight, 128);
+  return 0;
+}
+
+static void CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1,
+                       uint8* dst, int dst_stride,
+                       int width, int height) {
+  int y;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height - 1; y += 2) {
+    CopyRow(src, dst, width);
+    CopyRow(src + src_stride_0, dst + dst_stride, width);
+    src += src_stride_0 + src_stride_1;
+    dst += dst_stride * 2;
+  }
+  if (height & 1) {
+    CopyRow(src, dst, width);
+  }
+}
+
+// Support converting from FOURCC_M420
+// Useful for bandwidth constrained transports like USB 1.0 and 2.0 and for
+// easy conversion to I420.
+// M420 format description:
+// M420 is row biplanar 420: 2 rows of Y and 1 row of UV.
+// Chroma is half width / half height. (420)
+// src_stride_m420 is row planar. Normally this will be the width in pixels.
+//   The UV plane is half width, but 2 values, so src_stride_m420 applies to
+//   this as well as the two Y planes.
+static int X420ToI420(const uint8* src_y,
+                      int src_stride_y0, int src_stride_y1,
+                      const uint8* src_uv, int src_stride_uv,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) =
+      SplitUVRow_C;
+  if (!src_y || !src_uv ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
+    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+  // Coalesce rows.
+  if (src_stride_y0 == width &&
+      src_stride_y1 == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y0 = src_stride_y1 = dst_stride_y = 0;
+  }
+  // Coalesce rows.
+  if (src_stride_uv == halfwidth * 2 &&
+      dst_stride_u == halfwidth &&
+      dst_stride_v == halfwidth) {
+    halfwidth *= halfheight;
+    halfheight = 1;
+    src_stride_uv = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_SPLITUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitUVRow = SplitUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitUVRow = SplitUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      SplitUVRow = SplitUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitUVRow = SplitUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_uv, 4) && IS_ALIGNED(src_stride_uv, 4) &&
+      IS_ALIGNED(dst_u, 4) && IS_ALIGNED(dst_stride_u, 4) &&
+      IS_ALIGNED(dst_v, 4) && IS_ALIGNED(dst_stride_v, 4)) {
+    SplitUVRow = SplitUVRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (dst_y) {
+    if (src_stride_y0 == src_stride_y1) {
+      CopyPlane(src_y, src_stride_y0, dst_y, dst_stride_y, width, height);
+    } else {
+      CopyPlane2(src_y, src_stride_y0, src_stride_y1, dst_y, dst_stride_y,
+                 width, height);
+    }
+  }
+
+  for (y = 0; y < halfheight; ++y) {
+    // Copy a row of UV.
+    SplitUVRow(src_uv, dst_u, dst_v, halfwidth);
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+    src_uv += src_stride_uv;
+  }
+  return 0;
+}
+
+// Convert NV12 to I420.
+LIBYUV_API
+int NV12ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_y, src_stride_y, src_stride_y,
+                    src_uv, src_stride_uv,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height);
+}
+
+// Convert NV21 to I420.  Same as NV12 but u and v pointers swapped.
+LIBYUV_API
+int NV21ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_y, src_stride_y, src_stride_y,
+                    src_vu, src_stride_vu,
+                    dst_y, dst_stride_y,
+                    dst_v, dst_stride_v,
+                    dst_u, dst_stride_u,
+                    width, height);
+}
+
+// Convert M420 to I420.
+LIBYUV_API
+int M420ToI420(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_m420, src_stride_m420, src_stride_m420 * 2,
+                    src_m420 + src_stride_m420 * 2, src_stride_m420 * 3,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height);
+}
+
+// Convert YUY2 to I420.
+LIBYUV_API
+int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*YUY2ToUVRow)(const uint8* src_yuy2, int src_stride_yuy2,
+      uint8* dst_u, uint8* dst_v, int pix) = YUY2ToUVRow_C;
+  void (*YUY2ToYRow)(const uint8* src_yuy2,
+      uint8* dst_y, int pix) = YUY2ToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    YUY2ToUVRow = YUY2ToUVRow_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUVRow = YUY2ToUVRow_SSE2;
+      YUY2ToYRow = YUY2ToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YUY2ToUVRow = YUY2ToUVRow_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUVRow = YUY2ToUVRow_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    YUY2ToUVRow = YUY2ToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUVRow = YUY2ToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    YUY2ToUVRow(src_yuy2, src_stride_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    YUY2ToYRow(src_yuy2 + src_stride_yuy2, dst_y + dst_stride_y, width);
+    src_yuy2 += src_stride_yuy2 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    YUY2ToUVRow(src_yuy2, 0, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert UYVY to I420.
+LIBYUV_API
+int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*UYVYToUVRow)(const uint8* src_uyvy, int src_stride_uyvy,
+      uint8* dst_u, uint8* dst_v, int pix) = UYVYToUVRow_C;
+  void (*UYVYToYRow)(const uint8* src_uyvy,
+      uint8* dst_y, int pix) = UYVYToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    UYVYToUVRow = UYVYToUVRow_Any_SSE2;
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToUVRow = UYVYToUVRow_SSE2;
+      UYVYToYRow = UYVYToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    UYVYToUVRow = UYVYToUVRow_Any_AVX2;
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToUVRow = UYVYToUVRow_AVX2;
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    UYVYToUVRow = UYVYToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+      UYVYToUVRow = UYVYToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    UYVYToUVRow(src_uyvy, src_stride_uyvy, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+    UYVYToYRow(src_uyvy + src_stride_uyvy, dst_y + dst_stride_y, width);
+    src_uyvy += src_stride_uyvy * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    UYVYToUVRow(src_uyvy, 0, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert ARGB to I420.
+LIBYUV_API
+int ARGBToI420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ARGBToUVRow(src_argb, src_stride_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+    src_argb += src_stride_argb * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ARGBToUVRow(src_argb, 0, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert BGRA to I420.
+LIBYUV_API
+int BGRAToI420(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*BGRAToUVRow)(const uint8* src_bgra0, int src_stride_bgra,
+      uint8* dst_u, uint8* dst_v, int width) = BGRAToUVRow_C;
+  void (*BGRAToYRow)(const uint8* src_bgra, uint8* dst_y, int pix) =
+      BGRAToYRow_C;
+  if (!src_bgra ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_bgra = src_bgra + (height - 1) * src_stride_bgra;
+    src_stride_bgra = -src_stride_bgra;
+  }
+#if defined(HAS_BGRATOYROW_SSSE3) && defined(HAS_BGRATOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    BGRAToUVRow = BGRAToUVRow_Any_SSSE3;
+    BGRAToYRow = BGRAToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      BGRAToUVRow = BGRAToUVRow_SSSE3;
+      BGRAToYRow = BGRAToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_BGRATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    BGRAToYRow = BGRAToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      BGRAToYRow = BGRAToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_BGRATOUVROW_NEON)
+    if (TestCpuFlag(kCpuHasNEON)) {
+      BGRAToUVRow = BGRAToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        BGRAToUVRow = BGRAToUVRow_NEON;
+      }
+    }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    BGRAToUVRow(src_bgra, src_stride_bgra, dst_u, dst_v, width);
+    BGRAToYRow(src_bgra, dst_y, width);
+    BGRAToYRow(src_bgra + src_stride_bgra, dst_y + dst_stride_y, width);
+    src_bgra += src_stride_bgra * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    BGRAToUVRow(src_bgra, 0, dst_u, dst_v, width);
+    BGRAToYRow(src_bgra, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert ABGR to I420.
+LIBYUV_API
+int ABGRToI420(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ABGRToUVRow)(const uint8* src_abgr0, int src_stride_abgr,
+      uint8* dst_u, uint8* dst_v, int width) = ABGRToUVRow_C;
+  void (*ABGRToYRow)(const uint8* src_abgr, uint8* dst_y, int pix) =
+      ABGRToYRow_C;
+  if (!src_abgr ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
+  }
+#if defined(HAS_ABGRTOYROW_SSSE3) && defined(HAS_ABGRTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
+    ABGRToYRow = ABGRToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_SSSE3;
+      ABGRToYRow = ABGRToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToYRow = ABGRToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ABGRToYRow = ABGRToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToUVRow = ABGRToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ABGRToUVRow(src_abgr, src_stride_abgr, dst_u, dst_v, width);
+    ABGRToYRow(src_abgr, dst_y, width);
+    ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width);
+    src_abgr += src_stride_abgr * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ABGRToUVRow(src_abgr, 0, dst_u, dst_v, width);
+    ABGRToYRow(src_abgr, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert RGBA to I420.
+LIBYUV_API
+int RGBAToI420(const uint8* src_rgba, int src_stride_rgba,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*RGBAToUVRow)(const uint8* src_rgba0, int src_stride_rgba,
+      uint8* dst_u, uint8* dst_v, int width) = RGBAToUVRow_C;
+  void (*RGBAToYRow)(const uint8* src_rgba, uint8* dst_y, int pix) =
+      RGBAToYRow_C;
+  if (!src_rgba ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgba = src_rgba + (height - 1) * src_stride_rgba;
+    src_stride_rgba = -src_stride_rgba;
+  }
+#if defined(HAS_RGBATOYROW_SSSE3) && defined(HAS_RGBATOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
+    RGBAToYRow = RGBAToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToUVRow = RGBAToUVRow_SSSE3;
+      RGBAToYRow = RGBAToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGBAToYRow = RGBAToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGBAToYRow = RGBAToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGBAToUVRow = RGBAToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToUVRow = RGBAToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    RGBAToUVRow(src_rgba, src_stride_rgba, dst_u, dst_v, width);
+    RGBAToYRow(src_rgba, dst_y, width);
+    RGBAToYRow(src_rgba + src_stride_rgba, dst_y + dst_stride_y, width);
+    src_rgba += src_stride_rgba * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    RGBAToUVRow(src_rgba, 0, dst_u, dst_v, width);
+    RGBAToYRow(src_rgba, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert RGB24 to I420.
+LIBYUV_API
+int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height) {
+  int y;
+#if defined(HAS_RGB24TOYROW_NEON)
+  void (*RGB24ToUVRow)(const uint8* src_rgb24, int src_stride_rgb24,
+      uint8* dst_u, uint8* dst_v, int width) = RGB24ToUVRow_C;
+  void (*RGB24ToYRow)(const uint8* src_rgb24, uint8* dst_y, int pix) =
+      RGB24ToYRow_C;
+#else
+  void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB24ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#endif
+  if (!src_rgb24 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+
+// Neon version does direct RGB24 to YUV.
+#if defined(HAS_RGB24TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB24ToUVRow = RGB24ToUVRow_Any_NEON;
+    RGB24ToYRow = RGB24ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB24ToYRow = RGB24ToYRow_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RGB24ToUVRow = RGB24ToUVRow_NEON;
+      }
+    }
+  }
+// Other platforms do intermediate conversion from RGB24 to ARGB.
+#else
+#if defined(HAS_RGB24TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RGB24TOYROW_NEON)
+      RGB24ToUVRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
+      RGB24ToYRow(src_rgb24, dst_y, width);
+      RGB24ToYRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
+#else
+      RGB24ToARGBRow(src_rgb24, row, width);
+      RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + kRowSize, width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_rgb24 += src_stride_rgb24 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_RGB24TOYROW_NEON)
+      RGB24ToUVRow(src_rgb24, 0, dst_u, dst_v, width);
+      RGB24ToYRow(src_rgb24, dst_y, width);
+#else
+      RGB24ToARGBRow(src_rgb24, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RGB24TOYROW_NEON)
+    free_aligned_buffer_64(row);
+  }
+#endif
+  return 0;
+}
+
+// Convert RAW to I420.
+LIBYUV_API
+int RAWToI420(const uint8* src_raw, int src_stride_raw,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int width, int height) {
+  int y;
+#if defined(HAS_RAWTOYROW_NEON)
+  void (*RAWToUVRow)(const uint8* src_raw, int src_stride_raw,
+      uint8* dst_u, uint8* dst_v, int width) = RAWToUVRow_C;
+  void (*RAWToYRow)(const uint8* src_raw, uint8* dst_y, int pix) =
+      RAWToYRow_C;
+#else
+  void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RAWToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#endif
+  if (!src_raw || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+
+// Neon version does direct RAW to YUV.
+#if defined(HAS_RAWTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToUVRow = RAWToUVRow_Any_NEON;
+    RAWToYRow = RAWToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RAWToYRow = RAWToYRow_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RAWToUVRow = RAWToUVRow_NEON;
+      }
+    }
+  }
+// Other platforms do intermediate conversion from RAW to ARGB.
+#else
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RAWTOYROW_NEON)
+      RAWToUVRow(src_raw, src_stride_raw, dst_u, dst_v, width);
+      RAWToYRow(src_raw, dst_y, width);
+      RAWToYRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      RAWToARGBRow(src_raw + src_stride_raw, row + kRowSize, width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_raw += src_stride_raw * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_RAWTOYROW_NEON)
+      RAWToUVRow(src_raw, 0, dst_u, dst_v, width);
+      RAWToYRow(src_raw, dst_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RAWTOYROW_NEON)
+    free_aligned_buffer_64(row);
+  }
+#endif
+  return 0;
+}
+
+// Convert RGB565 to I420.
+LIBYUV_API
+int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
+                 uint8* dst_y, int dst_stride_y,
+                 uint8* dst_u, int dst_stride_u,
+                 uint8* dst_v, int dst_stride_v,
+                 int width, int height) {
+  int y;
+#if defined(HAS_RGB565TOYROW_NEON)
+  void (*RGB565ToUVRow)(const uint8* src_rgb565, int src_stride_rgb565,
+      uint8* dst_u, uint8* dst_v, int width) = RGB565ToUVRow_C;
+  void (*RGB565ToYRow)(const uint8* src_rgb565, uint8* dst_y, int pix) =
+      RGB565ToYRow_C;
+#else
+  void (*RGB565ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB565ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#endif
+  if (!src_rgb565 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb565 = src_rgb565 + (height - 1) * src_stride_rgb565;
+    src_stride_rgb565 = -src_stride_rgb565;
+  }
+
+// Neon version does direct RGB565 to YUV.
+#if defined(HAS_RGB565TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB565ToUVRow = RGB565ToUVRow_Any_NEON;
+    RGB565ToYRow = RGB565ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToYRow = RGB565ToYRow_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RGB565ToUVRow = RGB565ToUVRow_NEON;
+      }
+    }
+  }
+// Other platforms do intermediate conversion from RGB565 to ARGB.
+#else
+#if defined(HAS_RGB565TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RGB565TOYROW_NEON)
+      RGB565ToUVRow(src_rgb565, src_stride_rgb565, dst_u, dst_v, width);
+      RGB565ToYRow(src_rgb565, dst_y, width);
+      RGB565ToYRow(src_rgb565 + src_stride_rgb565, dst_y + dst_stride_y, width);
+#else
+      RGB565ToARGBRow(src_rgb565, row, width);
+      RGB565ToARGBRow(src_rgb565 + src_stride_rgb565, row + kRowSize, width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_rgb565 += src_stride_rgb565 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_RGB565TOYROW_NEON)
+      RGB565ToUVRow(src_rgb565, 0, dst_u, dst_v, width);
+      RGB565ToYRow(src_rgb565, dst_y, width);
+#else
+      RGB565ToARGBRow(src_rgb565, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RGB565TOYROW_NEON)
+    free_aligned_buffer_64(row);
+  }
+#endif
+  return 0;
+}
+
+// Convert ARGB1555 to I420.
+LIBYUV_API
+int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height) {
+  int y;
+#if defined(HAS_ARGB1555TOYROW_NEON)
+  void (*ARGB1555ToUVRow)(const uint8* src_argb1555, int src_stride_argb1555,
+      uint8* dst_u, uint8* dst_v, int width) = ARGB1555ToUVRow_C;
+  void (*ARGB1555ToYRow)(const uint8* src_argb1555, uint8* dst_y, int pix) =
+      ARGB1555ToYRow_C;
+#else
+  void (*ARGB1555ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      ARGB1555ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#endif
+  if (!src_argb1555 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb1555 = src_argb1555 + (height - 1) * src_stride_argb1555;
+    src_stride_argb1555 = -src_stride_argb1555;
+  }
+
+// Neon version does direct ARGB1555 to YUV.
+#if defined(HAS_ARGB1555TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGB1555ToUVRow = ARGB1555ToUVRow_Any_NEON;
+    ARGB1555ToYRow = ARGB1555ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToYRow = ARGB1555ToYRow_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGB1555ToUVRow = ARGB1555ToUVRow_NEON;
+      }
+    }
+  }
+// Other platforms do intermediate conversion from ARGB1555 to ARGB.
+#else
+#if defined(HAS_ARGB1555TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_ARGB1555TOYROW_NEON)
+      ARGB1555ToUVRow(src_argb1555, src_stride_argb1555, dst_u, dst_v, width);
+      ARGB1555ToYRow(src_argb1555, dst_y, width);
+      ARGB1555ToYRow(src_argb1555 + src_stride_argb1555, dst_y + dst_stride_y,
+                     width);
+#else
+      ARGB1555ToARGBRow(src_argb1555, row, width);
+      ARGB1555ToARGBRow(src_argb1555 + src_stride_argb1555, row + kRowSize,
+                        width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_argb1555 += src_stride_argb1555 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_ARGB1555TOYROW_NEON)
+      ARGB1555ToUVRow(src_argb1555, 0, dst_u, dst_v, width);
+      ARGB1555ToYRow(src_argb1555, dst_y, width);
+#else
+      ARGB1555ToARGBRow(src_argb1555, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_ARGB1555TOYROW_NEON)
+    free_aligned_buffer_64(row);
+  }
+#endif
+  return 0;
+}
+
+// Convert ARGB4444 to I420.
+LIBYUV_API
+int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height) {
+  int y;
+#if defined(HAS_ARGB4444TOYROW_NEON)
+  void (*ARGB4444ToUVRow)(const uint8* src_argb4444, int src_stride_argb4444,
+      uint8* dst_u, uint8* dst_v, int width) = ARGB4444ToUVRow_C;
+  void (*ARGB4444ToYRow)(const uint8* src_argb4444, uint8* dst_y, int pix) =
+      ARGB4444ToYRow_C;
+#else
+  void (*ARGB4444ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      ARGB4444ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#endif
+  if (!src_argb4444 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb4444 = src_argb4444 + (height - 1) * src_stride_argb4444;
+    src_stride_argb4444 = -src_stride_argb4444;
+  }
+
+// Neon version does direct ARGB4444 to YUV.
+#if defined(HAS_ARGB4444TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGB4444ToUVRow = ARGB4444ToUVRow_Any_NEON;
+    ARGB4444ToYRow = ARGB4444ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToYRow = ARGB4444ToYRow_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGB4444ToUVRow = ARGB4444ToUVRow_NEON;
+      }
+    }
+  }
+// Other platforms do intermediate conversion from ARGB4444 to ARGB.
+#else
+#if defined(HAS_ARGB4444TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_ARGB4444TOYROW_NEON)
+      ARGB4444ToUVRow(src_argb4444, src_stride_argb4444, dst_u, dst_v, width);
+      ARGB4444ToYRow(src_argb4444, dst_y, width);
+      ARGB4444ToYRow(src_argb4444 + src_stride_argb4444, dst_y + dst_stride_y,
+                     width);
+#else
+      ARGB4444ToARGBRow(src_argb4444, row, width);
+      ARGB4444ToARGBRow(src_argb4444 + src_stride_argb4444, row + kRowSize,
+                        width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_argb4444 += src_stride_argb4444 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_ARGB4444TOYROW_NEON)
+      ARGB4444ToUVRow(src_argb4444, 0, dst_u, dst_v, width);
+      ARGB4444ToYRow(src_argb4444, dst_y, width);
+#else
+      ARGB4444ToARGBRow(src_argb4444, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_ARGB4444TOYROW_NEON)
+    free_aligned_buffer_64(row);
+  }
+#endif
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/convert_argb.cc b/libvpx/libvpx/third_party/libyuv/source/convert_argb.cc
new file mode 100644
index 0000000..44756bc
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/convert_argb.cc
@@ -0,0 +1,1155 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_argb.h"
+
+#include "libyuv/cpu_id.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/rotate_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy ARGB with optional flipping
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height) {
+  if (!src_argb || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+  CopyPlane(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+            width * 4, height);
+  return 0;
+}
+
+// Convert I444 to ARGB.
+LIBYUV_API
+int I444ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I444ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I444ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u == width &&
+      src_stride_v == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_I444TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444ToARGBRow = I444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToARGBRow = I444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444ToARGBRow = I444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I444ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to ARGB.
+LIBYUV_API
+int I422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I411 to ARGB.
+LIBYUV_API
+int I411ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I411ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I411ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 4 == width &&
+      src_stride_v * 4 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_I411TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I411ToARGBRow = I411ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I411ToARGBRow = I411ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I411TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I411ToARGBRow = I411ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I411ToARGBRow = I411ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I411TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I411ToARGBRow = I411ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I411ToARGBRow = I411ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I411ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I400 to ARGB.
+LIBYUV_API
+int I400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I400ToARGBRow)(const uint8* y_buf,
+                     uint8* rgb_buf,
+                     int width) = I400ToARGBRow_C;
+  if (!src_y || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_I400TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I400ToARGBRow = I400ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      I400ToARGBRow = I400ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I400TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I400ToARGBRow = I400ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I400ToARGBRow = I400ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I400TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I400ToARGBRow = I400ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I400ToARGBRow = I400ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I400ToARGBRow(src_y, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+  }
+  return 0;
+}
+
+// Convert J400 to ARGB.
+LIBYUV_API
+int J400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*J400ToARGBRow)(const uint8* src_y, uint8* dst_argb, int pix) =
+      J400ToARGBRow_C;
+  if (!src_y || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_J400TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    J400ToARGBRow = J400ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      J400ToARGBRow = J400ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_J400TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    J400ToARGBRow = J400ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      J400ToARGBRow = J400ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_J400TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    J400ToARGBRow = J400ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      J400ToARGBRow = J400ToARGBRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    J400ToARGBRow(src_y, dst_argb, width);
+    src_y += src_stride_y;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Shuffle table for converting BGRA to ARGB.
+static uvec8 kShuffleMaskBGRAToARGB = {
+  3u, 2u, 1u, 0u, 7u, 6u, 5u, 4u, 11u, 10u, 9u, 8u, 15u, 14u, 13u, 12u
+};
+
+// Shuffle table for converting ABGR to ARGB.
+static uvec8 kShuffleMaskABGRToARGB = {
+  2u, 1u, 0u, 3u, 6u, 5u, 4u, 7u, 10u, 9u, 8u, 11u, 14u, 13u, 12u, 15u
+};
+
+// Shuffle table for converting RGBA to ARGB.
+static uvec8 kShuffleMaskRGBAToARGB = {
+  1u, 2u, 3u, 0u, 5u, 6u, 7u, 4u, 9u, 10u, 11u, 8u, 13u, 14u, 15u, 12u
+};
+
+// Convert BGRA to ARGB.
+LIBYUV_API
+int BGRAToARGB(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_bgra, src_stride_bgra,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskBGRAToARGB),
+                     width, height);
+}
+
+// Convert ARGB to BGRA (same as BGRAToARGB).
+LIBYUV_API
+int ARGBToBGRA(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_bgra, src_stride_bgra,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskBGRAToARGB),
+                     width, height);
+}
+
+// Convert ABGR to ARGB.
+LIBYUV_API
+int ABGRToARGB(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_abgr, src_stride_abgr,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskABGRToARGB),
+                     width, height);
+}
+
+// Convert ARGB to ABGR to (same as ABGRToARGB).
+LIBYUV_API
+int ARGBToABGR(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_abgr, src_stride_abgr,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskABGRToARGB),
+                     width, height);
+}
+
+// Convert RGBA to ARGB.
+LIBYUV_API
+int RGBAToARGB(const uint8* src_rgba, int src_stride_rgba,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_rgba, src_stride_rgba,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskRGBAToARGB),
+                     width, height);
+}
+
+// Convert RGB24 to ARGB.
+LIBYUV_API
+int RGB24ToARGB(const uint8* src_rgb24, int src_stride_rgb24,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height) {
+  int y;
+  void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB24ToARGBRow_C;
+  if (!src_rgb24 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+  // Coalesce rows.
+  if (src_stride_rgb24 == width * 3 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_rgb24 = dst_stride_argb = 0;
+  }
+#if defined(HAS_RGB24TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RGB24ToARGBRow(src_rgb24, dst_argb, width);
+    src_rgb24 += src_stride_rgb24;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert RAW to ARGB.
+LIBYUV_API
+int RAWToARGB(const uint8* src_raw, int src_stride_raw,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  int y;
+  void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RAWToARGBRow_C;
+  if (!src_raw || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+  // Coalesce rows.
+  if (src_stride_raw == width * 3 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_raw = dst_stride_argb = 0;
+  }
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToARGBRow = RAWToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RAWToARGBRow = RAWToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RAWToARGBRow(src_raw, dst_argb, width);
+    src_raw += src_stride_raw;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert RGB565 to ARGB.
+LIBYUV_API
+int RGB565ToARGB(const uint8* src_rgb565, int src_stride_rgb565,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*RGB565ToARGBRow)(const uint8* src_rgb565, uint8* dst_argb, int pix) =
+      RGB565ToARGBRow_C;
+  if (!src_rgb565 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb565 = src_rgb565 + (height - 1) * src_stride_rgb565;
+    src_stride_rgb565 = -src_stride_rgb565;
+  }
+  // Coalesce rows.
+  if (src_stride_rgb565 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_rgb565 = dst_stride_argb = 0;
+  }
+#if defined(HAS_RGB565TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RGB565ToARGBRow(src_rgb565, dst_argb, width);
+    src_rgb565 += src_stride_rgb565;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB1555 to ARGB.
+LIBYUV_API
+int ARGB1555ToARGB(const uint8* src_argb1555, int src_stride_argb1555,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height) {
+  int y;
+  void (*ARGB1555ToARGBRow)(const uint8* src_argb1555, uint8* dst_argb,
+      int pix) = ARGB1555ToARGBRow_C;
+  if (!src_argb1555 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb1555 = src_argb1555 + (height - 1) * src_stride_argb1555;
+    src_stride_argb1555 = -src_stride_argb1555;
+  }
+  // Coalesce rows.
+  if (src_stride_argb1555 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb1555 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGB1555TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGB1555ToARGBRow(src_argb1555, dst_argb, width);
+    src_argb1555 += src_stride_argb1555;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB4444 to ARGB.
+LIBYUV_API
+int ARGB4444ToARGB(const uint8* src_argb4444, int src_stride_argb4444,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height) {
+  int y;
+  void (*ARGB4444ToARGBRow)(const uint8* src_argb4444, uint8* dst_argb,
+      int pix) = ARGB4444ToARGBRow_C;
+  if (!src_argb4444 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb4444 = src_argb4444 + (height - 1) * src_stride_argb4444;
+    src_stride_argb4444 = -src_stride_argb4444;
+  }
+  // Coalesce rows.
+  if (src_stride_argb4444 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb4444 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGB4444TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGB4444ToARGBRow(src_argb4444, dst_argb, width);
+    src_argb4444 += src_stride_argb4444;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert NV12 to ARGB.
+LIBYUV_API
+int NV12ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*NV12ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV12ToARGBRow_C;
+  if (!src_y || !src_uv || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_NV12TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToARGBRow = NV12ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToARGBRow(src_y, src_uv, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV21 to ARGB.
+LIBYUV_API
+int NV21ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*NV21ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV21ToARGBRow_C;
+  if (!src_y || !src_uv || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_NV21TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV21TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToARGBRow = NV21ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV21TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV21ToARGBRow(src_y, src_uv, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert M420 to ARGB.
+LIBYUV_API
+int M420ToARGB(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*NV12ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV12ToARGBRow_C;
+  if (!src_m420 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_NV12TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToARGBRow = NV12ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb, width);
+    NV12ToARGBRow(src_m420 + src_stride_m420, src_m420 + src_stride_m420 * 2,
+                  dst_argb + dst_stride_argb, width);
+    dst_argb += dst_stride_argb * 2;
+    src_m420 += src_stride_m420 * 3;
+  }
+  if (height & 1) {
+    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb, width);
+  }
+  return 0;
+}
+
+// Convert YUY2 to ARGB.
+LIBYUV_API
+int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*YUY2ToARGBRow)(const uint8* src_yuy2, uint8* dst_argb, int pix) =
+      YUY2ToARGBRow_C;
+  if (!src_yuy2 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_yuy2 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_yuy2 = dst_stride_argb = 0;
+  }
+#if defined(HAS_YUY2TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    YUY2ToARGBRow(src_yuy2, dst_argb, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert UYVY to ARGB.
+LIBYUV_API
+int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*UYVYToARGBRow)(const uint8* src_uyvy, uint8* dst_argb, int pix) =
+      UYVYToARGBRow_C;
+  if (!src_uyvy || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_argb = 0;
+  }
+#if defined(HAS_UYVYTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToARGBRow = UYVYToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToARGBRow = UYVYToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      UYVYToARGBRow = UYVYToARGBRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    UYVYToARGBRow(src_uyvy, dst_argb, width);
+    src_uyvy += src_stride_uyvy;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert J420 to ARGB.
+LIBYUV_API
+int J420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*J422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = J422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_J422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    J422ToARGBRow = J422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    J422ToARGBRow = J422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      J422ToARGBRow = J422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    J422ToARGBRow = J422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    J422ToARGBRow = J422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    J422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert J422 to ARGB.
+LIBYUV_API
+int J422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*J422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = J422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_J422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    J422ToARGBRow = J422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    J422ToARGBRow = J422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      J422ToARGBRow = J422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    J422ToARGBRow = J422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    J422ToARGBRow = J422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    J422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/convert_from.cc b/libvpx/libvpx/third_party/libyuv/source/convert_from.cc
new file mode 100644
index 0000000..31f1ac9
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/convert_from.cc
@@ -0,0 +1,1348 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_from.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/convert.h"  // For I420Copy
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+#include "libyuv/scale.h"  // For ScalePlane()
+#include "libyuv/video_common.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// I420 To any I4xx YUV format with mirroring.
+static int I420ToI4xx(const uint8* src_y, int src_stride_y,
+                      const uint8* src_u, int src_stride_u,
+                      const uint8* src_v, int src_stride_v,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int src_y_width, int src_y_height,
+                      int dst_uv_width, int dst_uv_height) {
+  const int dst_y_width = Abs(src_y_width);
+  const int dst_y_height = Abs(src_y_height);
+  const int src_uv_width = SUBSAMPLE(src_y_width, 1, 1);
+  const int src_uv_height = SUBSAMPLE(src_y_height, 1, 1);
+  if (src_y_width == 0 || src_y_height == 0 ||
+      dst_uv_width <= 0 || dst_uv_height <= 0) {
+    return -1;
+  }
+  ScalePlane(src_y, src_stride_y, src_y_width, src_y_height,
+             dst_y, dst_stride_y, dst_y_width, dst_y_height,
+             kFilterBilinear);
+  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height,
+             dst_u, dst_stride_u, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height,
+             dst_v, dst_stride_v, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  return 0;
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 422 chroma is 1/2 width, 1x height
+LIBYUV_API
+int I420ToI422(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = (Abs(width) + 1) >> 1;
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 444 chroma is 1x width, 1x height
+LIBYUV_API
+int I420ToI444(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = Abs(width);
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 411 chroma is 1/4 width, 1x height
+LIBYUV_API
+int I420ToI411(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = (Abs(width) + 3) >> 2;
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// Copy to I400. Source can be I420,422,444,400,NV12,NV21
+LIBYUV_API
+int I400Copy(const uint8* src_y, int src_stride_y,
+             uint8* dst_y, int dst_stride_y,
+             int width, int height) {
+  if (!src_y || !dst_y ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+LIBYUV_API
+int I422ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  int y;
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_yuy2, int width) =
+      I422ToYUY2Row_C;
+  if (!src_y || !src_u || !src_v || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_yuy2 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_yuy2 = 0;
+  }
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_yuy2 += dst_stride_yuy2;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  int y;
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_yuy2, int width) =
+      I422ToYUY2Row_C;
+  if (!src_y || !src_u || !src_v || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+    I422ToYUY2Row(src_y + src_stride_y, src_u, src_v,
+                  dst_yuy2 + dst_stride_yuy2, width);
+    src_y += src_stride_y * 2;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_yuy2 += dst_stride_yuy2 * 2;
+  }
+  if (height & 1) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I422ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  int y;
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_uyvy, int width) =
+      I422ToUYVYRow_C;
+  if (!src_y || !src_u || !src_v || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_uyvy == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_uyvy = 0;
+  }
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uyvy += dst_stride_uyvy;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  int y;
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_uyvy, int width) =
+      I422ToUYVYRow_C;
+  if (!src_y || !src_u || !src_v || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+    I422ToUYVYRow(src_y + src_stride_y, src_u, src_v,
+                  dst_uyvy + dst_stride_uyvy, width);
+    src_y += src_stride_y * 2;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uyvy += dst_stride_uyvy * 2;
+  }
+  if (height & 1) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToNV12(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+      int width) = MergeUVRow_C;
+  // Coalesce rows.
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_uv = dst_uv + (halfheight - 1) * dst_stride_uv;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_uv = -dst_stride_uv;
+  }
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+  // Coalesce rows.
+  if (src_stride_u == halfwidth &&
+      src_stride_v == halfwidth &&
+      dst_stride_uv == halfwidth * 2) {
+    halfwidth *= halfheight;
+    halfheight = 1;
+    src_stride_u = src_stride_v = dst_stride_uv = 0;
+  }
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  for (y = 0; y < halfheight; ++y) {
+    // Merge a row of U and V into a row of UV.
+    MergeUVRow_(src_u, src_v, dst_uv, halfwidth);
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uv += dst_stride_uv;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToNV21(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height) {
+  return I420ToNV12(src_y, src_stride_y,
+                    src_v, src_stride_v,
+                    src_u, src_stride_u,
+                    dst_y, src_stride_y,
+                    dst_vu, dst_stride_vu,
+                    width, height);
+}
+
+// Convert I420 to ARGB.
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to BGRA.
+LIBYUV_API
+int I420ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height) {
+  int y;
+  void (*I422ToBGRARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToBGRARow_C;
+  if (!src_y || !src_u || !src_v || !dst_bgra ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra;
+    dst_stride_bgra = -dst_stride_bgra;
+  }
+#if defined(HAS_I422TOBGRAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToBGRARow = I422ToBGRARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToBGRARow = I422ToBGRARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) {
+    I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width);
+    dst_bgra += dst_stride_bgra;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to ABGR.
+LIBYUV_API
+int I420ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height) {
+  int y;
+  void (*I422ToABGRRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToABGRRow_C;
+  if (!src_y || !src_u || !src_v || !dst_abgr ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr;
+    dst_stride_abgr = -dst_stride_abgr;
+  }
+#if defined(HAS_I422TOABGRROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOABGRROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToABGRRow = I422ToABGRRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOABGRROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToABGRRow = I422ToABGRRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width);
+    dst_abgr += dst_stride_abgr;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGBA.
+LIBYUV_API
+int I420ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRGBARow_C;
+  if (!src_y || !src_u || !src_v || !dst_rgba ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB24.
+LIBYUV_API
+int I420ToRGB24(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height) {
+  int y;
+  void (*I422ToRGB24Row)(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* rgb_buf,
+                         int width) = I422ToRGB24Row_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb24 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
+  }
+#if defined(HAS_I422TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB24Row = I422ToRGB24Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB24Row(src_y, src_u, src_v, dst_rgb24, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RAW.
+LIBYUV_API
+int I420ToRAW(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_raw, int dst_stride_raw,
+                int width, int height) {
+  int y;
+  void (*I422ToRAWRow)(const uint8* y_buf,
+                       const uint8* u_buf,
+                       const uint8* v_buf,
+                       uint8* rgb_buf,
+                       int width) = I422ToRAWRow_C;
+  if (!src_y || !src_u || !src_v || !dst_raw ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_raw = dst_raw + (height - 1) * dst_stride_raw;
+    dst_stride_raw = -dst_stride_raw;
+  }
+#if defined(HAS_I422TORAWROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRAWRow = I422ToRAWRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRAWRow = I422ToRAWRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORAWROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRAWRow = I422ToRAWRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRAWRow = I422ToRAWRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORAWROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRAWRow = I422ToRAWRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRAWRow = I422ToRAWRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRAWRow(src_y, src_u, src_v, dst_raw, width);
+    dst_raw += dst_stride_raw;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to ARGB1555.
+LIBYUV_API
+int I420ToARGB1555(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height) {
+  int y;
+  void (*I422ToARGB1555Row)(const uint8* y_buf,
+                            const uint8* u_buf,
+                            const uint8* v_buf,
+                            uint8* rgb_buf,
+                            int width) = I422ToARGB1555Row_C;
+  if (!src_y || !src_u || !src_v || !dst_argb1555 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb1555 = dst_argb1555 + (height - 1) * dst_stride_argb1555;
+    dst_stride_argb1555 = -dst_stride_argb1555;
+  }
+#if defined(HAS_I422TOARGB1555ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGB1555Row(src_y, src_u, src_v, dst_argb1555, width);
+    dst_argb1555 += dst_stride_argb1555;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+
+// Convert I420 to ARGB4444.
+LIBYUV_API
+int I420ToARGB4444(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height) {
+  int y;
+  void (*I422ToARGB4444Row)(const uint8* y_buf,
+                            const uint8* u_buf,
+                            const uint8* v_buf,
+                            uint8* rgb_buf,
+                            int width) = I422ToARGB4444Row_C;
+  if (!src_y || !src_u || !src_v || !dst_argb4444 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb4444 = dst_argb4444 + (height - 1) * dst_stride_argb4444;
+    dst_stride_argb4444 = -dst_stride_argb4444;
+  }
+#if defined(HAS_I422TOARGB4444ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGB4444Row(src_y, src_u, src_v, dst_argb4444, width);
+    dst_argb4444 += dst_stride_argb4444;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB565.
+LIBYUV_API
+int I420ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_u, int src_stride_u,
+                 const uint8* src_v, int src_stride_v,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*I422ToRGB565Row)(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* rgb_buf,
+                          int width) = I422ToRGB565Row_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_I422TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB565Row = I422ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Ordered 8x8 dither for 888 to 565.  Values from 0 to 7.
+static const uint8 kDither565_4x4[16] = {
+  0, 4, 1, 5,
+  6, 2, 7, 3,
+  1, 5, 0, 4,
+  7, 3, 6, 2,
+};
+
+// Convert I420 to RGB565 with dithering.
+LIBYUV_API
+int I420ToRGB565Dither(const uint8* src_y, int src_stride_y,
+                       const uint8* src_u, int src_stride_u,
+                       const uint8* src_v, int src_stride_v,
+                       uint8* dst_rgb565, int dst_stride_rgb565,
+                       const uint8* dither4x4, int width, int height) {
+  int y;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+  void (*ARGBToRGB565DitherRow)(const uint8* src_argb, uint8* dst_rgb,
+      const uint32 dither4, int pix) = ARGBToRGB565DitherRow_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+  if (!dither4x4) {
+    dither4x4 = kDither565_4x4;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_NEON;
+    }
+  }
+#endif
+  {
+    // Allocate a row of argb.
+    align_buffer_64(row_argb, width * 4);
+    for (y = 0; y < height; ++y) {
+      I422ToARGBRow(src_y, src_u, src_v, row_argb, width);
+      ARGBToRGB565DitherRow(row_argb, dst_rgb565,
+                            *(uint32*)(dither4x4 + ((y & 3) << 2)), width);
+      dst_rgb565 += dst_stride_rgb565;
+      src_y += src_stride_y;
+      if (y & 1) {
+        src_u += src_stride_u;
+        src_v += src_stride_v;
+      }
+    }
+    free_aligned_buffer_64(row_argb);
+  }
+  return 0;
+}
+
+// Convert I420 to specified format
+LIBYUV_API
+int ConvertFromI420(const uint8* y, int y_stride,
+                    const uint8* u, int u_stride,
+                    const uint8* v, int v_stride,
+                    uint8* dst_sample, int dst_sample_stride,
+                    int width, int height,
+                    uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int r = 0;
+  if (!y || !u|| !v || !dst_sample ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      r = I420ToYUY2(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 2,
+                     width, height);
+      break;
+    case FOURCC_UYVY:
+      r = I420ToUYVY(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 2,
+                     width, height);
+      break;
+    case FOURCC_RGBP:
+      r = I420ToRGB565(y, y_stride,
+                       u, u_stride,
+                       v, v_stride,
+                       dst_sample,
+                       dst_sample_stride ? dst_sample_stride : width * 2,
+                       width, height);
+      break;
+    case FOURCC_RGBO:
+      r = I420ToARGB1555(y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         dst_sample,
+                         dst_sample_stride ? dst_sample_stride : width * 2,
+                         width, height);
+      break;
+    case FOURCC_R444:
+      r = I420ToARGB4444(y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         dst_sample,
+                         dst_sample_stride ? dst_sample_stride : width * 2,
+                         width, height);
+      break;
+    case FOURCC_24BG:
+      r = I420ToRGB24(y, y_stride,
+                      u, u_stride,
+                      v, v_stride,
+                      dst_sample,
+                      dst_sample_stride ? dst_sample_stride : width * 3,
+                      width, height);
+      break;
+    case FOURCC_RAW:
+      r = I420ToRAW(y, y_stride,
+                    u, u_stride,
+                    v, v_stride,
+                    dst_sample,
+                    dst_sample_stride ? dst_sample_stride : width * 3,
+                    width, height);
+      break;
+    case FOURCC_ARGB:
+      r = I420ToARGB(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_BGRA:
+      r = I420ToBGRA(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_ABGR:
+      r = I420ToABGR(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_RGBA:
+      r = I420ToRGBA(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_I400:
+      r = I400Copy(y, y_stride,
+                   dst_sample,
+                   dst_sample_stride ? dst_sample_stride : width,
+                   width, height);
+      break;
+    case FOURCC_NV12: {
+      uint8* dst_uv = dst_sample + width * height;
+      r = I420ToNV12(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     dst_uv,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     width, height);
+      break;
+    }
+    case FOURCC_NV21: {
+      uint8* dst_vu = dst_sample + width * height;
+      r = I420ToNV21(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     dst_vu,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     width, height);
+      break;
+    }
+    // TODO(fbarchard): Add M420.
+    // Triplanar formats
+    // TODO(fbarchard): halfstride instead of halfwidth
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      int halfwidth = (width + 1) / 2;
+      int halfheight = (height + 1) / 2;
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV12) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + halfwidth * halfheight;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + halfwidth * halfheight;
+      }
+      r = I420Copy(y, y_stride,
+                   u, u_stride,
+                   v, v_stride,
+                   dst_sample, width,
+                   dst_u, halfwidth,
+                   dst_v, halfwidth,
+                   width, height);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      int halfwidth = (width + 1) / 2;
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV16) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + halfwidth * height;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + halfwidth * height;
+      }
+      r = I420ToI422(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, halfwidth,
+                     dst_v, halfwidth,
+                     width, height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV24) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + width * height;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + width * height;
+      }
+      r = I420ToI444(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, width,
+                     dst_v, width,
+                     width, height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (width + 3) / 4;
+      uint8* dst_u = dst_sample + width * height;
+      uint8* dst_v = dst_u + quarterwidth * height;
+      r = I420ToI411(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, quarterwidth,
+                     dst_v, quarterwidth,
+                     width, height);
+      break;
+    }
+
+    // Formats not supported - MJPG, biplanar, some rgb formats.
+    default:
+      return -1;  // unknown fourcc - return failure code.
+  }
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/convert_from_argb.cc b/libvpx/libvpx/third_party/libyuv/source/convert_from_argb.cc
new file mode 100644
index 0000000..8d1e97a
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/convert_from_argb.cc
@@ -0,0 +1,1301 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_from_argb.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// ARGB little endian (bgra in memory) to I444
+LIBYUV_API
+int ARGBToI444(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*ARGBToUV444Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV444Row_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u == width &&
+      dst_stride_v == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOUV444ROW_SSSE3)
+    if (TestCpuFlag(kCpuHasSSSE3)) {
+      ARGBToUV444Row = ARGBToUV444Row_Any_SSSE3;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUV444Row = ARGBToUV444Row_SSSE3;
+      }
+  }
+#endif
+#if defined(HAS_ARGBTOUV444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUV444Row = ARGBToUV444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToUV444Row = ARGBToUV444Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUV444Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// ARGB little endian (bgra in memory) to I422
+LIBYUV_API
+int ARGBToI422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV422Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUV422Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// ARGB little endian (bgra in memory) to I411
+LIBYUV_API
+int ARGBToI411(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV411Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV411Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 4 == width &&
+      dst_stride_v * 4 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUV411ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUV411Row = ARGBToUV411Row_Any_NEON;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUV411Row = ARGBToUV411Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUV411Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                      int width) = MergeUVRow_C;
+  if (!src_argb ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+  {
+    // Allocate a rows of uv.
+    align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2);
+    uint8* row_v = row_u + ((halfwidth + 31) & ~31);
+
+    for (y = 0; y < height - 1; y += 2) {
+      ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
+      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      ARGBToYRow(src_argb, dst_y, width);
+      ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+      src_argb += src_stride_argb * 2;
+      dst_y += dst_stride_y * 2;
+      dst_uv += dst_stride_uv;
+    }
+    if (height & 1) {
+      ARGBToUVRow(src_argb, 0, row_u, row_v, width);
+      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      ARGBToYRow(src_argb, dst_y, width);
+    }
+    free_aligned_buffer_64(row_u);
+  }
+  return 0;
+}
+
+// Same as NV12 but U and V swapped.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                      int width) = MergeUVRow_C;
+  if (!src_argb ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+  {
+    // Allocate a rows of uv.
+    align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2);
+    uint8* row_v = row_u + ((halfwidth + 31) & ~31);
+
+    for (y = 0; y < height - 1; y += 2) {
+      ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
+      MergeUVRow_(row_v, row_u, dst_uv, halfwidth);
+      ARGBToYRow(src_argb, dst_y, width);
+      ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+      src_argb += src_stride_argb * 2;
+      dst_y += dst_stride_y * 2;
+      dst_uv += dst_stride_uv;
+    }
+    if (height & 1) {
+      ARGBToUVRow(src_argb, 0, row_u, row_v, width);
+      MergeUVRow_(row_v, row_u, dst_uv, halfwidth);
+      ARGBToYRow(src_argb, dst_y, width);
+    }
+    free_aligned_buffer_64(row_u);
+  }
+  return 0;
+}
+
+// Convert ARGB to YUY2.
+LIBYUV_API
+int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV422Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+      const uint8* src_v, uint8* dst_yuy2, int width) = I422ToYUY2Row_C;
+
+  if (!src_argb || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_yuy2 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_yuy2 = 0;
+  }
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  {
+    // Allocate a rows of yuv.
+    align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+    uint8* row_u = row_y + ((width + 63) & ~63);
+    uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+
+    for (y = 0; y < height; ++y) {
+      ARGBToUV422Row(src_argb, row_u, row_v, width);
+      ARGBToYRow(src_argb, row_y, width);
+      I422ToYUY2Row(row_y, row_u, row_v, dst_yuy2, width);
+      src_argb += src_stride_argb;
+      dst_yuy2 += dst_stride_yuy2;
+    }
+
+    free_aligned_buffer_64(row_y);
+  }
+  return 0;
+}
+
+// Convert ARGB to UYVY.
+LIBYUV_API
+int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV422Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+      const uint8* src_v, uint8* dst_uyvy, int width) = I422ToUYVYRow_C;
+
+  if (!src_argb || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_uyvy == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_uyvy = 0;
+  }
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  {
+    // Allocate a rows of yuv.
+    align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+    uint8* row_u = row_y + ((width + 63) & ~63);
+    uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+
+    for (y = 0; y < height; ++y) {
+      ARGBToUV422Row(src_argb, row_u, row_v, width);
+      ARGBToYRow(src_argb, row_y, width);
+      I422ToUYVYRow(row_y, row_u, row_v, dst_uyvy, width);
+      src_argb += src_stride_argb;
+      dst_uyvy += dst_stride_uyvy;
+    }
+
+    free_aligned_buffer_64(row_y);
+  }
+  return 0;
+}
+
+// Convert ARGB to I400.
+LIBYUV_API
+int ARGBToI400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  int y;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = 0;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+  }
+  return 0;
+}
+
+// Shuffle table for converting ARGB to RGBA.
+static uvec8 kShuffleMaskARGBToRGBA = {
+  3u, 0u, 1u, 2u, 7u, 4u, 5u, 6u, 11u, 8u, 9u, 10u, 15u, 12u, 13u, 14u
+};
+
+// Convert ARGB to RGBA.
+LIBYUV_API
+int ARGBToRGBA(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  return ARGBShuffle(src_argb, src_stride_argb,
+                     dst_rgba, dst_stride_rgba,
+                     (const uint8*)(&kShuffleMaskARGBToRGBA),
+                     width, height);
+}
+
+// Convert ARGB To RGB24.
+LIBYUV_API
+int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height) {
+  int y;
+  void (*ARGBToRGB24Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRGB24Row_C;
+  if (!src_argb || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_rgb24 == width * 3) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_rgb24 = 0;
+  }
+#if defined(HAS_ARGBTORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToRGB24Row(src_argb, dst_rgb24, width);
+    src_argb += src_stride_argb;
+    dst_rgb24 += dst_stride_rgb24;
+  }
+  return 0;
+}
+
+// Convert ARGB To RAW.
+LIBYUV_API
+int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_raw, int dst_stride_raw,
+              int width, int height) {
+  int y;
+  void (*ARGBToRAWRow)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRAWRow_C;
+  if (!src_argb || !dst_raw || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_raw == width * 3) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_raw = 0;
+  }
+#if defined(HAS_ARGBTORAWROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRAWRow = ARGBToRAWRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORAWROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRAWRow = ARGBToRAWRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToRAWRow(src_argb, dst_raw, width);
+    src_argb += src_stride_argb;
+    dst_raw += dst_stride_raw;
+  }
+  return 0;
+}
+
+// Ordered 8x8 dither for 888 to 565.  Values from 0 to 7.
+static const uint8 kDither565_4x4[16] = {
+  0, 4, 1, 5,
+  6, 2, 7, 3,
+  1, 5, 0, 4,
+  7, 3, 6, 2,
+};
+
+// Convert ARGB To RGB565 with 4x4 dither matrix (16 bytes).
+LIBYUV_API
+int ARGBToRGB565Dither(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_rgb565, int dst_stride_rgb565,
+                       const uint8* dither4x4, int width, int height) {
+  int y;
+  void (*ARGBToRGB565DitherRow)(const uint8* src_argb, uint8* dst_rgb,
+      const uint32 dither4, int pix) = ARGBToRGB565DitherRow_C;
+  if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  if (!dither4x4) {
+    dither4x4 = kDither565_4x4;
+  }
+#if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBToRGB565DitherRow(src_argb, dst_rgb565,
+                          *(uint32*)(dither4x4 + ((y & 3) << 2)), width);
+    src_argb += src_stride_argb;
+    dst_rgb565 += dst_stride_rgb565;
+  }
+  return 0;
+}
+
+// Convert ARGB To RGB565.
+// TODO(fbarchard): Consider using dither function low level with zeros.
+LIBYUV_API
+int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*ARGBToRGB565Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRGB565Row_C;
+  if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_rgb565 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_rgb565 = 0;
+  }
+#if defined(HAS_ARGBTORGB565ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToRGB565Row(src_argb, dst_rgb565, width);
+    src_argb += src_stride_argb;
+    dst_rgb565 += dst_stride_rgb565;
+  }
+  return 0;
+}
+
+// Convert ARGB To ARGB1555.
+LIBYUV_API
+int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height) {
+  int y;
+  void (*ARGBToARGB1555Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToARGB1555Row_C;
+  if (!src_argb || !dst_argb1555 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb1555 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb1555 = 0;
+  }
+#if defined(HAS_ARGBTOARGB1555ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB1555ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToARGB1555Row(src_argb, dst_argb1555, width);
+    src_argb += src_stride_argb;
+    dst_argb1555 += dst_stride_argb1555;
+  }
+  return 0;
+}
+
+// Convert ARGB To ARGB4444.
+LIBYUV_API
+int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height) {
+  int y;
+  void (*ARGBToARGB4444Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToARGB4444Row_C;
+  if (!src_argb || !dst_argb4444 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb4444 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb4444 = 0;
+  }
+#if defined(HAS_ARGBTOARGB4444ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB4444ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToARGB4444Row(src_argb, dst_argb4444, width);
+    src_argb += src_stride_argb;
+    dst_argb4444 += dst_stride_argb4444;
+  }
+  return 0;
+}
+
+// Convert ARGB to J420. (JPeg full range I420).
+LIBYUV_API
+int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUVJRow)(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) = ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int pix) =
+      ARGBToYJRow_C;
+  if (!src_argb ||
+      !dst_yj || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ARGBToUVJRow(src_argb, src_stride_argb, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_yj, width);
+    ARGBToYJRow(src_argb + src_stride_argb, dst_yj + dst_stride_yj, width);
+    src_argb += src_stride_argb * 2;
+    dst_yj += dst_stride_yj * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ARGBToUVJRow(src_argb, 0, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_yj, width);
+  }
+  return 0;
+}
+
+// ARGB little endian (bgra in memory) to J422
+LIBYUV_API
+int ARGBToJ422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUVJ422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUVJ422Row_C;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYJRow_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOUVJ422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJ422Row = ARGBToUVJ422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJ422Row = ARGBToUVJ422Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJ422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVJ422Row = ARGBToUVJ422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJ422Row = ARGBToUVJ422Row_NEON;
+    }
+  }
+#endif
+
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUVJ422Row(src_argb, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Convert ARGB to J400.
+LIBYUV_API
+int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               int width, int height) {
+  int y;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int pix) =
+      ARGBToYJRow_C;
+  if (!src_argb || !dst_yj || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_yj = 0;
+  }
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToYJRow(src_argb, dst_yj, width);
+    src_argb += src_stride_argb;
+    dst_yj += dst_stride_yj;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/convert_jpeg.cc b/libvpx/libvpx/third_party/libyuv/source/convert_jpeg.cc
new file mode 100644
index 0000000..bcb980f
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/convert_jpeg.cc
@@ -0,0 +1,392 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert.h"
+
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#ifdef HAVE_JPEG
+struct I420Buffers {
+  uint8* y;
+  int y_stride;
+  uint8* u;
+  int u_stride;
+  uint8* v;
+  int v_stride;
+  int w;
+  int h;
+};
+
+static void JpegCopyI420(void* opaque,
+                         const uint8* const* data,
+                         const int* strides,
+                         int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I420Copy(data[0], strides[0],
+           data[1], strides[1],
+           data[2], strides[2],
+           dest->y, dest->y_stride,
+           dest->u, dest->u_stride,
+           dest->v, dest->v_stride,
+           dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I422ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I444ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI411ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I411ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I400ToI420(data[0], strides[0],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+// Query size of MJPG in pixels.
+LIBYUV_API
+int MJPGSize(const uint8* sample, size_t sample_size,
+             int* width, int* height) {
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret) {
+    *width = mjpeg_decoder.GetWidth();
+    *height = mjpeg_decoder.GetHeight();
+  }
+  mjpeg_decoder.UnloadFrame();
+  return ret ? 0 : -1;  // -1 for runtime failure.
+}
+
+// MJPG (Motion JPeg) to I420
+// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+LIBYUV_API
+int MJPGToI420(const uint8* sample,
+               size_t sample_size,
+               uint8* y, int y_stride,
+               uint8* u, int u_stride,
+               uint8* v, int v_stride,
+               int w, int h,
+               int dw, int dh) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret && (mjpeg_decoder.GetWidth() != w ||
+              mjpeg_decoder.GetHeight() != h)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    I420Buffers bufs = { y, y_stride, u, u_stride, v, v_stride, dw, dh };
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() ==
+            MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegCopyI420, &bufs, dw, dh);
+    // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToI420, &bufs, dw, dh);
+    // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToI420, &bufs, dw, dh);
+    // YUV411
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 4 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToI420, &bufs, dw, dh);
+    // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToI420, &bufs, dw, dh);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other colorspace/sample
+      // factors that occur in practice. 411 is supported by libjpeg
+      // ERROR: Unable to convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+
+#ifdef HAVE_JPEG
+struct ARGBBuffers {
+  uint8* argb;
+  int argb_stride;
+  int w;
+  int h;
+};
+
+static void JpegI420ToARGB(void* opaque,
+                         const uint8* const* data,
+                         const int* strides,
+                         int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I420ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I422ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I444ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI411ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I411ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I400ToARGB(data[0], strides[0],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+// MJPG (Motion JPeg) to ARGB
+// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample,
+               size_t sample_size,
+               uint8* argb, int argb_stride,
+               int w, int h,
+               int dw, int dh) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret && (mjpeg_decoder.GetWidth() != w ||
+              mjpeg_decoder.GetHeight() != h)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    ARGBBuffers bufs = { argb, argb_stride, dw, dh };
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() ==
+            MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToARGB, &bufs, dw, dh);
+    // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToARGB, &bufs, dw, dh);
+    // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToARGB, &bufs, dw, dh);
+    // YUV411
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 4 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToARGB, &bufs, dw, dh);
+    // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToARGB, &bufs, dw, dh);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other colorspace/sample
+      // factors that occur in practice. 411 is supported by libjpeg
+      // ERROR: Unable to convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+#endif
+
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/convert_to_argb.cc b/libvpx/libvpx/third_party/libyuv/source/convert_to_argb.cc
new file mode 100644
index 0000000..af829fb
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/convert_to_argb.cc
@@ -0,0 +1,306 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_argb.h"
+
+#include "libyuv/cpu_id.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/rotate_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// src_width is used for source stride computation
+// src_height is used to compute location of planes, and indicate inversion
+// sample_size is measured in bytes and is the size of the frame.
+//   With MJPEG it is the compressed size of the frame.
+LIBYUV_API
+int ConvertToARGB(const uint8* sample, size_t sample_size,
+                  uint8* crop_argb, int argb_stride,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int aligned_src_width = (src_width + 1) & ~1;
+  const uint8* src;
+  const uint8* src_uv;
+  int abs_src_height = (src_height < 0) ? -src_height : src_height;
+  int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+  int r = 0;
+
+  // One pass rotation is available for some formats. For the rest, convert
+  // to I420 (with optional vertical flipping) into a temporary I420 buffer,
+  // and then rotate the I420 to the final destination buffer.
+  // For in-place conversion, if destination crop_argb is same as source sample,
+  // also enable temporary buffer.
+  LIBYUV_BOOL need_buf = (rotation && format != FOURCC_ARGB) ||
+      crop_argb == sample;
+  uint8* tmp_argb = crop_argb;
+  int tmp_argb_stride = argb_stride;
+  uint8* rotate_buffer = NULL;
+  int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+
+  if (crop_argb == NULL || sample == NULL ||
+      src_width <= 0 || crop_width <= 0 ||
+      src_height == 0 || crop_height == 0) {
+    return -1;
+  }
+  if (src_height < 0) {
+    inv_crop_height = -inv_crop_height;
+  }
+
+  if (need_buf) {
+    int argb_size = crop_width * abs_crop_height * 4;
+    rotate_buffer = (uint8*)malloc(argb_size);
+    if (!rotate_buffer) {
+      return 1;  // Out of memory runtime error.
+    }
+    crop_argb = rotate_buffer;
+    argb_stride = crop_width;
+  }
+
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = YUY2ToARGB(src, aligned_src_width * 2,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_UYVY:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = UYVYToARGB(src, aligned_src_width * 2,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_24BG:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RGB24ToARGB(src, src_width * 3,
+                      crop_argb, argb_stride,
+                      crop_width, inv_crop_height);
+      break;
+    case FOURCC_RAW:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RAWToARGB(src, src_width * 3,
+                    crop_argb, argb_stride,
+                    crop_width, inv_crop_height);
+      break;
+    case FOURCC_ARGB:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ARGBToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_BGRA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = BGRAToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_ABGR:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ABGRToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = RGBAToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBP:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = RGB565ToARGB(src, src_width * 2,
+                       crop_argb, argb_stride,
+                       crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBO:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB1555ToARGB(src, src_width * 2,
+                         crop_argb, argb_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_R444:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB4444ToARGB(src, src_width * 2,
+                         crop_argb, argb_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_I400:
+      src = sample + src_width * crop_y + crop_x;
+      r = I400ToARGB(src, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+
+    // Biplanar formats
+    case FOURCC_NV12:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      r = NV12ToARGB(src, src_width,
+                     src_uv, aligned_src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_NV21:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      // Call NV12 but with u and v parameters swapped.
+      r = NV21ToARGB(src, src_width,
+                     src_uv, aligned_src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_M420:
+      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
+      r = M420ToARGB(src, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // Triplanar formats
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      if (format == FOURCC_YV12) {
+        src_v = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      }
+      r = I420ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_J420: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      src_u = sample + src_width * abs_src_height +
+          (halfwidth * crop_y + crop_x) / 2;
+      src_v = sample + src_width * abs_src_height +
+          halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      r = J420ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      if (format == FOURCC_YV16) {
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      }
+      r = I422ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      if (format == FOURCC_YV24) {
+        src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      } else {
+        src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      }
+      r = I444ToARGB(src_y, src_width,
+                     src_u, src_width,
+                     src_v, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (src_width + 3) / 4;
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u = sample + src_width * abs_src_height +
+          quarterwidth * crop_y + crop_x / 4;
+      const uint8* src_v = sample + src_width * abs_src_height +
+          quarterwidth * (abs_src_height + crop_y) + crop_x / 4;
+      r = I411ToARGB(src_y, src_width,
+                     src_u, quarterwidth,
+                     src_v, quarterwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+#ifdef HAVE_JPEG
+    case FOURCC_MJPG:
+      r = MJPGToARGB(sample, sample_size,
+                     crop_argb, argb_stride,
+                     src_width, abs_src_height, crop_width, inv_crop_height);
+      break;
+#endif
+    default:
+      r = -1;  // unknown fourcc - return failure code.
+  }
+
+  if (need_buf) {
+    if (!r) {
+      r = ARGBRotate(crop_argb, argb_stride,
+                     tmp_argb, tmp_argb_stride,
+                     crop_width, abs_crop_height, rotation);
+    }
+    free(rotate_buffer);
+  }
+
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/convert_to_i420.cc b/libvpx/libvpx/third_party/libyuv/source/convert_to_i420.cc
new file mode 100644
index 0000000..5e75369
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/convert_to_i420.cc
@@ -0,0 +1,339 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "libyuv/convert.h"
+
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// src_width is used for source stride computation
+// src_height is used to compute location of planes, and indicate inversion
+// sample_size is measured in bytes and is the size of the frame.
+//   With MJPEG it is the compressed size of the frame.
+LIBYUV_API
+int ConvertToI420(const uint8* sample,
+                  size_t sample_size,
+                  uint8* y, int y_stride,
+                  uint8* u, int u_stride,
+                  uint8* v, int v_stride,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int aligned_src_width = (src_width + 1) & ~1;
+  const uint8* src;
+  const uint8* src_uv;
+  int abs_src_height = (src_height < 0) ? -src_height : src_height;
+  int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+  int r = 0;
+  LIBYUV_BOOL need_buf = (rotation && format != FOURCC_I420 &&
+      format != FOURCC_NV12 && format != FOURCC_NV21 &&
+      format != FOURCC_YU12 && format != FOURCC_YV12) || y == sample;
+  uint8* tmp_y = y;
+  uint8* tmp_u = u;
+  uint8* tmp_v = v;
+  int tmp_y_stride = y_stride;
+  int tmp_u_stride = u_stride;
+  int tmp_v_stride = v_stride;
+  uint8* rotate_buffer = NULL;
+  int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+
+  if (!y || !u || !v || !sample ||
+      src_width <= 0 || crop_width <= 0  ||
+      src_height == 0 || crop_height == 0) {
+    return -1;
+  }
+  if (src_height < 0) {
+    inv_crop_height = -inv_crop_height;
+  }
+
+  // One pass rotation is available for some formats. For the rest, convert
+  // to I420 (with optional vertical flipping) into a temporary I420 buffer,
+  // and then rotate the I420 to the final destination buffer.
+  // For in-place conversion, if destination y is same as source sample,
+  // also enable temporary buffer.
+  if (need_buf) {
+    int y_size = crop_width * abs_crop_height;
+    int uv_size = ((crop_width + 1) / 2) * ((abs_crop_height + 1) / 2);
+    rotate_buffer = (uint8*)malloc(y_size + uv_size * 2);
+    if (!rotate_buffer) {
+      return 1;  // Out of memory runtime error.
+    }
+    y = rotate_buffer;
+    u = y + y_size;
+    v = u + uv_size;
+    y_stride = crop_width;
+    u_stride = v_stride = ((crop_width + 1) / 2);
+  }
+
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = YUY2ToI420(src, aligned_src_width * 2,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_UYVY:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = UYVYToI420(src, aligned_src_width * 2,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBP:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = RGB565ToI420(src, src_width * 2,
+                       y, y_stride,
+                       u, u_stride,
+                       v, v_stride,
+                       crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBO:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB1555ToI420(src, src_width * 2,
+                         y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_R444:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB4444ToI420(src, src_width * 2,
+                         y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_24BG:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RGB24ToI420(src, src_width * 3,
+                      y, y_stride,
+                      u, u_stride,
+                      v, v_stride,
+                      crop_width, inv_crop_height);
+      break;
+    case FOURCC_RAW:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RAWToI420(src, src_width * 3,
+                    y, y_stride,
+                    u, u_stride,
+                    v, v_stride,
+                    crop_width, inv_crop_height);
+      break;
+    case FOURCC_ARGB:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ARGBToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_BGRA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = BGRAToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_ABGR:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ABGRToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = RGBAToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_I400:
+      src = sample + src_width * crop_y + crop_x;
+      r = I400ToI420(src, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // Biplanar formats
+    case FOURCC_NV12:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + (src_width * src_height) +
+        ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
+      r = NV12ToI420Rotate(src, src_width,
+                           src_uv, aligned_src_width,
+                           y, y_stride,
+                           u, u_stride,
+                           v, v_stride,
+                           crop_width, inv_crop_height, rotation);
+      break;
+    case FOURCC_NV21:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + (src_width * src_height) +
+        ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
+      // Call NV12 but with u and v parameters swapped.
+      r = NV12ToI420Rotate(src, src_width,
+                           src_uv, aligned_src_width,
+                           y, y_stride,
+                           v, v_stride,
+                           u, u_stride,
+                           crop_width, inv_crop_height, rotation);
+      break;
+    case FOURCC_M420:
+      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
+      r = M420ToI420(src, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // Triplanar formats
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      if (format == FOURCC_YV12) {
+        src_v = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      }
+      r = I420Rotate(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height, rotation);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      if (format == FOURCC_YV16) {
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      }
+      r = I422ToI420(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      if (format == FOURCC_YV24) {
+        src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      } else {
+        src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      }
+      r = I444ToI420(src_y, src_width,
+                     src_u, src_width,
+                     src_v, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (src_width + 3) / 4;
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u = sample + src_width * abs_src_height +
+          quarterwidth * crop_y + crop_x / 4;
+      const uint8* src_v = sample + src_width * abs_src_height +
+          quarterwidth * (abs_src_height + crop_y) + crop_x / 4;
+      r = I411ToI420(src_y, src_width,
+                     src_u, quarterwidth,
+                     src_v, quarterwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+#ifdef HAVE_JPEG
+    case FOURCC_MJPG:
+      r = MJPGToI420(sample, sample_size,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     src_width, abs_src_height, crop_width, inv_crop_height);
+      break;
+#endif
+    default:
+      r = -1;  // unknown fourcc - return failure code.
+  }
+
+  if (need_buf) {
+    if (!r) {
+      r = I420Rotate(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     tmp_y, tmp_y_stride,
+                     tmp_u, tmp_u_stride,
+                     tmp_v, tmp_v_stride,
+                     crop_width, abs_crop_height, rotation);
+    }
+    free(rotate_buffer);
+  }
+
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/cpu_id.cc b/libvpx/libvpx/third_party/libyuv/source/cpu_id.cc
new file mode 100644
index 0000000..8a10b00
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/cpu_id.cc
@@ -0,0 +1,307 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/cpu_id.h"
+
+#if (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+#include <intrin.h>  // For __cpuidex()
+#endif
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    !defined(__native_client__) && (defined(_M_IX86) || defined(_M_X64)) && \
+    defined(_MSC_VER) && !defined(__clang__) && (_MSC_FULL_VER >= 160040219)
+#include <immintrin.h>  // For _xgetbv()
+#endif
+
+#if !defined(__native_client__)
+#include <stdlib.h>  // For getenv()
+#endif
+
+// For ArmCpuCaps() but unittested on all platforms
+#include <stdio.h>
+#include <string.h>
+
+#include "libyuv/basic_types.h"  // For CPU_X86
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// For functions that use the stack and have runtime checks for overflow,
+// use SAFEBUFFERS to avoid additional check.
+#if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219)
+#define SAFEBUFFERS __declspec(safebuffers)
+#else
+#define SAFEBUFFERS
+#endif
+
+// Low level cpuid for X86.
+#if (defined(_M_IX86) || defined(_M_X64) || \
+    defined(__i386__) || defined(__x86_64__)) && \
+    !defined(__pnacl__) && !defined(__CLR_VER)
+LIBYUV_API
+void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
+#if (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+// Visual C version uses intrinsic or inline x86 assembly.
+#if (_MSC_FULL_VER >= 160040219)
+  __cpuidex((int*)(cpu_info), info_eax, info_ecx);
+#elif defined(_M_IX86)
+  __asm {
+    mov        eax, info_eax
+    mov        ecx, info_ecx
+    mov        edi, cpu_info
+    cpuid
+    mov        [edi], eax
+    mov        [edi + 4], ebx
+    mov        [edi + 8], ecx
+    mov        [edi + 12], edx
+  }
+#else
+  if (info_ecx == 0) {
+    __cpuid((int*)(cpu_info), info_eax);
+  } else {
+    cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0;
+  }
+#endif
+// GCC version uses inline x86 assembly.
+#else  // (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+  uint32 info_ebx, info_edx;
+  asm volatile (  // NOLINT
+#if defined( __i386__) && defined(__PIC__)
+    // Preserve ebx for fpic 32 bit.
+    "mov %%ebx, %%edi                          \n"
+    "cpuid                                     \n"
+    "xchg %%edi, %%ebx                         \n"
+    : "=D" (info_ebx),
+#else
+    "cpuid                                     \n"
+    : "=b" (info_ebx),
+#endif  //  defined( __i386__) && defined(__PIC__)
+      "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx));
+  cpu_info[0] = info_eax;
+  cpu_info[1] = info_ebx;
+  cpu_info[2] = info_ecx;
+  cpu_info[3] = info_edx;
+#endif  // (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+}
+#else  // (defined(_M_IX86) || defined(_M_X64) ...
+LIBYUV_API
+void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
+  cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
+}
+#endif
+
+// TODO(fbarchard): Enable xgetbv when validator supports it.
+#if (defined(_M_IX86) || defined(_M_X64) || \
+    defined(__i386__) || defined(__x86_64__)) && \
+    !defined(__pnacl__) && !defined(__CLR_VER) && !defined(__native_client__)
+#define HAS_XGETBV
+// X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
+int TestOsSaveYmm() {
+  uint32 xcr0 = 0u;
+#if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219)
+  xcr0 = (uint32)(_xgetbv(0));  // VS2010 SP1 required.
+#elif defined(_M_IX86) && defined(_MSC_VER) && !defined(__clang__)
+  __asm {
+    xor        ecx, ecx    // xcr 0
+    _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
+    mov        xcr0, eax
+  }
+#elif defined(__i386__) || defined(__x86_64__)
+  asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
+#endif  // defined(__i386__) || defined(__x86_64__)
+  return((xcr0 & 6) == 6);  // Is ymm saved?
+}
+#endif  // defined(_M_IX86) || defined(_M_X64) ..
+
+// based on libvpx arm_cpudetect.c
+// For Arm, but public to allow testing on any CPU
+LIBYUV_API SAFEBUFFERS
+int ArmCpuCaps(const char* cpuinfo_name) {
+  char cpuinfo_line[512];
+  FILE* f = fopen(cpuinfo_name, "r");
+  if (!f) {
+    // Assume Neon if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasNEON;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
+    if (memcmp(cpuinfo_line, "Features", 8) == 0) {
+      char* p = strstr(cpuinfo_line, " neon");
+      if (p && (p[5] == ' ' || p[5] == '\n')) {
+        fclose(f);
+        return kCpuHasNEON;
+      }
+      // aarch64 uses asimd for Neon.
+      p = strstr(cpuinfo_line, " asimd");
+      if (p && (p[6] == ' ' || p[6] == '\n')) {
+        fclose(f);
+        return kCpuHasNEON;
+      }
+    }
+  }
+  fclose(f);
+  return 0;
+}
+
+#if defined(__mips__) && defined(__linux__)
+static int MipsCpuCaps(const char* search_string) {
+  char cpuinfo_line[512];
+  const char* file_name = "/proc/cpuinfo";
+  FILE* f = fopen(file_name, "r");
+  if (!f) {
+    // Assume DSP if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasMIPS_DSP;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f) != NULL) {
+    if (strstr(cpuinfo_line, search_string) != NULL) {
+      fclose(f);
+      return kCpuHasMIPS_DSP;
+    }
+  }
+  fclose(f);
+  return 0;
+}
+#endif
+
+// CPU detect function for SIMD instruction sets.
+LIBYUV_API
+int cpu_info_ = kCpuInit;  // cpu_info is not initialized yet.
+
+// Test environment variable for disabling CPU features. Any non-zero value
+// to disable. Zero ignored to make it easy to set the variable on/off.
+#if !defined(__native_client__) && !defined(_M_ARM)
+
+static LIBYUV_BOOL TestEnv(const char* name) {
+  const char* var = getenv(name);
+  if (var) {
+    if (var[0] != '0') {
+      return LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_FALSE;
+}
+#else  // nacl does not support getenv().
+static LIBYUV_BOOL TestEnv(const char*) {
+  return LIBYUV_FALSE;
+}
+#endif
+
+LIBYUV_API SAFEBUFFERS
+int InitCpuFlags(void) {
+#if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86)
+
+  uint32 cpu_info0[4] = { 0, 0, 0, 0 };
+  uint32 cpu_info1[4] = { 0, 0, 0, 0 };
+  uint32 cpu_info7[4] = { 0, 0, 0, 0 };
+  CpuId(0, 0, cpu_info0);
+  CpuId(1, 0, cpu_info1);
+  if (cpu_info0[0] >= 7) {
+    CpuId(7, 0, cpu_info7);
+  }
+  cpu_info_ = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
+              ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
+              ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
+              ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
+              ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) |
+              ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
+              kCpuHasX86;
+
+#ifdef HAS_XGETBV
+  if ((cpu_info1[2] & 0x18000000) == 0x18000000 &&  // AVX and OSSave
+      TestOsSaveYmm()) {  // Saves YMM.
+    cpu_info_ |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) |
+                 kCpuHasAVX;
+  }
+#endif
+  // Environment variable overrides for testing.
+  if (TestEnv("LIBYUV_DISABLE_X86")) {
+    cpu_info_ &= ~kCpuHasX86;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE2")) {
+    cpu_info_ &= ~kCpuHasSSE2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
+    cpu_info_ &= ~kCpuHasSSSE3;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE41")) {
+    cpu_info_ &= ~kCpuHasSSE41;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE42")) {
+    cpu_info_ &= ~kCpuHasSSE42;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX")) {
+    cpu_info_ &= ~kCpuHasAVX;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX2")) {
+    cpu_info_ &= ~kCpuHasAVX2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_ERMS")) {
+    cpu_info_ &= ~kCpuHasERMS;
+  }
+  if (TestEnv("LIBYUV_DISABLE_FMA3")) {
+    cpu_info_ &= ~kCpuHasFMA3;
+  }
+#endif
+#if defined(__mips__) && defined(__linux__)
+  // Linux mips parse text file for dsp detect.
+  cpu_info_ = MipsCpuCaps("dsp");  // set kCpuHasMIPS_DSP.
+#if defined(__mips_dspr2)
+  cpu_info_ |= kCpuHasMIPS_DSPR2;
+#endif
+  cpu_info_ |= kCpuHasMIPS;
+
+  if (getenv("LIBYUV_DISABLE_MIPS")) {
+    cpu_info_ &= ~kCpuHasMIPS;
+  }
+  if (getenv("LIBYUV_DISABLE_MIPS_DSP")) {
+    cpu_info_ &= ~kCpuHasMIPS_DSP;
+  }
+  if (getenv("LIBYUV_DISABLE_MIPS_DSPR2")) {
+    cpu_info_ &= ~kCpuHasMIPS_DSPR2;
+  }
+#endif
+#if defined(__arm__) || defined(__aarch64__)
+// gcc -mfpu=neon defines __ARM_NEON__
+// __ARM_NEON__ generates code that requires Neon.  NaCL also requires Neon.
+// For Linux, /proc/cpuinfo can be tested but without that assume Neon.
+#if defined(__ARM_NEON__) || defined(__native_client__) || !defined(__linux__)
+  cpu_info_ = kCpuHasNEON;
+// For aarch64(arm64), /proc/cpuinfo's feature is not complete, e.g. no neon
+// flag in it.
+// So for aarch64, neon enabling is hard coded here.
+#endif
+#if defined(__aarch64__)
+  cpu_info_ = kCpuHasNEON;
+#else
+  // Linux arm parse text file for neon detect.
+  cpu_info_ = ArmCpuCaps("/proc/cpuinfo");
+#endif
+  cpu_info_ |= kCpuHasARM;
+  if (TestEnv("LIBYUV_DISABLE_NEON")) {
+    cpu_info_ &= ~kCpuHasNEON;
+  }
+#endif  // __arm__
+  if (TestEnv("LIBYUV_DISABLE_ASM")) {
+    cpu_info_ = 0;
+  }
+  return cpu_info_;
+}
+
+LIBYUV_API
+void MaskCpuFlags(int enable_flags) {
+  cpu_info_ = InitCpuFlags() & enable_flags;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/mjpeg_decoder.cc b/libvpx/libvpx/third_party/libyuv/source/mjpeg_decoder.cc
new file mode 100644
index 0000000..75f8a61
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/mjpeg_decoder.cc
@@ -0,0 +1,572 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/mjpeg_decoder.h"
+
+#ifdef HAVE_JPEG
+#include <assert.h>
+
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    !defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR)
+// Must be included before jpeglib.
+#include <setjmp.h>
+#define HAVE_SETJMP
+
+#if defined(_MSC_VER)
+// disable warning 4324: structure was padded due to __declspec(align())
+#pragma warning(disable:4324)
+#endif
+
+#endif
+struct FILE;  // For jpeglib.h.
+
+// C++ build requires extern C for jpeg internals.
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <jpeglib.h>
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#include "libyuv/planar_functions.h"  // For CopyPlane().
+
+namespace libyuv {
+
+#ifdef HAVE_SETJMP
+struct SetJmpErrorMgr {
+  jpeg_error_mgr base;  // Must be at the top
+  jmp_buf setjmp_buffer;
+};
+#endif
+
+const int MJpegDecoder::kColorSpaceUnknown = JCS_UNKNOWN;
+const int MJpegDecoder::kColorSpaceGrayscale = JCS_GRAYSCALE;
+const int MJpegDecoder::kColorSpaceRgb = JCS_RGB;
+const int MJpegDecoder::kColorSpaceYCbCr = JCS_YCbCr;
+const int MJpegDecoder::kColorSpaceCMYK = JCS_CMYK;
+const int MJpegDecoder::kColorSpaceYCCK = JCS_YCCK;
+
+// Methods that are passed to jpeglib.
+boolean fill_input_buffer(jpeg_decompress_struct* cinfo);
+void init_source(jpeg_decompress_struct* cinfo);
+void skip_input_data(jpeg_decompress_struct* cinfo,
+                     long num_bytes);  // NOLINT
+void term_source(jpeg_decompress_struct* cinfo);
+void ErrorHandler(jpeg_common_struct* cinfo);
+
+MJpegDecoder::MJpegDecoder()
+    : has_scanline_padding_(LIBYUV_FALSE),
+      num_outbufs_(0),
+      scanlines_(NULL),
+      scanlines_sizes_(NULL),
+      databuf_(NULL),
+      databuf_strides_(NULL) {
+  decompress_struct_ = new jpeg_decompress_struct;
+  source_mgr_ = new jpeg_source_mgr;
+#ifdef HAVE_SETJMP
+  error_mgr_ = new SetJmpErrorMgr;
+  decompress_struct_->err = jpeg_std_error(&error_mgr_->base);
+  // Override standard exit()-based error handler.
+  error_mgr_->base.error_exit = &ErrorHandler;
+#endif
+  decompress_struct_->client_data = NULL;
+  source_mgr_->init_source = &init_source;
+  source_mgr_->fill_input_buffer = &fill_input_buffer;
+  source_mgr_->skip_input_data = &skip_input_data;
+  source_mgr_->resync_to_restart = &jpeg_resync_to_restart;
+  source_mgr_->term_source = &term_source;
+  jpeg_create_decompress(decompress_struct_);
+  decompress_struct_->src = source_mgr_;
+  buf_vec_.buffers = &buf_;
+  buf_vec_.len = 1;
+}
+
+MJpegDecoder::~MJpegDecoder() {
+  jpeg_destroy_decompress(decompress_struct_);
+  delete decompress_struct_;
+  delete source_mgr_;
+#ifdef HAVE_SETJMP
+  delete error_mgr_;
+#endif
+  DestroyOutputBuffers();
+}
+
+LIBYUV_BOOL MJpegDecoder::LoadFrame(const uint8* src, size_t src_len) {
+  if (!ValidateJpeg(src, src_len)) {
+    return LIBYUV_FALSE;
+  }
+
+  buf_.data = src;
+  buf_.len = static_cast<int>(src_len);
+  buf_vec_.pos = 0;
+  decompress_struct_->client_data = &buf_vec_;
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called jpeg_read_header, it experienced an error, and we called
+    // longjmp() and rewound the stack to here. Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (jpeg_read_header(decompress_struct_, TRUE) != JPEG_HEADER_OK) {
+    // ERROR: Bad MJPEG header
+    return LIBYUV_FALSE;
+  }
+  AllocOutputBuffers(GetNumComponents());
+  for (int i = 0; i < num_outbufs_; ++i) {
+    int scanlines_size = GetComponentScanlinesPerImcuRow(i);
+    if (scanlines_sizes_[i] != scanlines_size) {
+      if (scanlines_[i]) {
+        delete scanlines_[i];
+      }
+      scanlines_[i] = new uint8* [scanlines_size];
+      scanlines_sizes_[i] = scanlines_size;
+    }
+
+    // We allocate padding for the final scanline to pad it up to DCTSIZE bytes
+    // to avoid memory errors, since jpeglib only reads full MCUs blocks. For
+    // the preceding scanlines, the padding is not needed/wanted because the
+    // following addresses will already be valid (they are the initial bytes of
+    // the next scanline) and will be overwritten when jpeglib writes out that
+    // next scanline.
+    int databuf_stride = GetComponentStride(i);
+    int databuf_size = scanlines_size * databuf_stride;
+    if (databuf_strides_[i] != databuf_stride) {
+      if (databuf_[i]) {
+        delete databuf_[i];
+      }
+      databuf_[i] = new uint8[databuf_size];
+      databuf_strides_[i] = databuf_stride;
+    }
+
+    if (GetComponentStride(i) != GetComponentWidth(i)) {
+      has_scanline_padding_ = LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_TRUE;
+}
+
+static int DivideAndRoundUp(int numerator, int denominator) {
+  return (numerator + denominator - 1) / denominator;
+}
+
+static int DivideAndRoundDown(int numerator, int denominator) {
+  return numerator / denominator;
+}
+
+// Returns width of the last loaded frame.
+int MJpegDecoder::GetWidth() {
+  return decompress_struct_->image_width;
+}
+
+// Returns height of the last loaded frame.
+int MJpegDecoder::GetHeight() {
+  return decompress_struct_->image_height;
+}
+
+// Returns format of the last loaded frame. The return value is one of the
+// kColorSpace* constants.
+int MJpegDecoder::GetColorSpace() {
+  return decompress_struct_->jpeg_color_space;
+}
+
+// Number of color components in the color space.
+int MJpegDecoder::GetNumComponents() {
+  return decompress_struct_->num_components;
+}
+
+// Sample factors of the n-th component.
+int MJpegDecoder::GetHorizSampFactor(int component) {
+  return decompress_struct_->comp_info[component].h_samp_factor;
+}
+
+int MJpegDecoder::GetVertSampFactor(int component) {
+  return decompress_struct_->comp_info[component].v_samp_factor;
+}
+
+int MJpegDecoder::GetHorizSubSampFactor(int component) {
+  return decompress_struct_->max_h_samp_factor /
+      GetHorizSampFactor(component);
+}
+
+int MJpegDecoder::GetVertSubSampFactor(int component) {
+  return decompress_struct_->max_v_samp_factor /
+      GetVertSampFactor(component);
+}
+
+int MJpegDecoder::GetImageScanlinesPerImcuRow() {
+  return decompress_struct_->max_v_samp_factor * DCTSIZE;
+}
+
+int MJpegDecoder::GetComponentScanlinesPerImcuRow(int component) {
+  int vs = GetVertSubSampFactor(component);
+  return DivideAndRoundUp(GetImageScanlinesPerImcuRow(), vs);
+}
+
+int MJpegDecoder::GetComponentWidth(int component) {
+  int hs = GetHorizSubSampFactor(component);
+  return DivideAndRoundUp(GetWidth(), hs);
+}
+
+int MJpegDecoder::GetComponentHeight(int component) {
+  int vs = GetVertSubSampFactor(component);
+  return DivideAndRoundUp(GetHeight(), vs);
+}
+
+// Get width in bytes padded out to a multiple of DCTSIZE
+int MJpegDecoder::GetComponentStride(int component) {
+  return (GetComponentWidth(component) + DCTSIZE - 1) & ~(DCTSIZE - 1);
+}
+
+int MJpegDecoder::GetComponentSize(int component) {
+  return GetComponentWidth(component) * GetComponentHeight(component);
+}
+
+LIBYUV_BOOL MJpegDecoder::UnloadFrame() {
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called jpeg_abort_decompress, it experienced an error, and we called
+    // longjmp() and rewound the stack to here. Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  jpeg_abort_decompress(decompress_struct_);
+  return LIBYUV_TRUE;
+}
+
+// TODO(fbarchard): Allow rectangle to be specified: x, y, width, height.
+LIBYUV_BOOL MJpegDecoder::DecodeToBuffers(
+    uint8** planes, int dst_width, int dst_height) {
+  if (dst_width != GetWidth() ||
+      dst_height > GetHeight()) {
+    // ERROR: Bad dimensions
+    return LIBYUV_FALSE;
+  }
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called into jpeglib, it experienced an error sometime during this
+    // function call, and we called longjmp() and rewound the stack to here.
+    // Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (!StartDecode()) {
+    return LIBYUV_FALSE;
+  }
+  SetScanlinePointers(databuf_);
+  int lines_left = dst_height;
+  // Compute amount of lines to skip to implement vertical crop.
+  // TODO(fbarchard): Ensure skip is a multiple of maximum component
+  // subsample. ie 2
+  int skip = (GetHeight() - dst_height) / 2;
+  if (skip > 0) {
+    // There is no API to skip lines in the output data, so we read them
+    // into the temp buffer.
+    while (skip >= GetImageScanlinesPerImcuRow()) {
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      skip -= GetImageScanlinesPerImcuRow();
+    }
+    if (skip > 0) {
+      // Have a partial iMCU row left over to skip. Must read it and then
+      // copy the parts we want into the destination.
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      for (int i = 0; i < num_outbufs_; ++i) {
+        // TODO(fbarchard): Compute skip to avoid this
+        assert(skip % GetVertSubSampFactor(i) == 0);
+        int rows_to_skip =
+            DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i) -
+                                rows_to_skip;
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        CopyPlane(databuf_[i] + data_to_skip, GetComponentStride(i),
+                  planes[i], GetComponentWidth(i),
+                  GetComponentWidth(i), scanlines_to_copy);
+        planes[i] += scanlines_to_copy * GetComponentWidth(i);
+      }
+      lines_left -= (GetImageScanlinesPerImcuRow() - skip);
+    }
+  }
+
+  // Read full MCUs but cropped horizontally
+  for (; lines_left > GetImageScanlinesPerImcuRow();
+         lines_left -= GetImageScanlinesPerImcuRow()) {
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    for (int i = 0; i < num_outbufs_; ++i) {
+      int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i);
+      CopyPlane(databuf_[i], GetComponentStride(i),
+                planes[i], GetComponentWidth(i),
+                GetComponentWidth(i), scanlines_to_copy);
+      planes[i] += scanlines_to_copy * GetComponentWidth(i);
+    }
+  }
+
+  if (lines_left > 0) {
+    // Have a partial iMCU row left over to decode.
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    for (int i = 0; i < num_outbufs_; ++i) {
+      int scanlines_to_copy =
+          DivideAndRoundUp(lines_left, GetVertSubSampFactor(i));
+      CopyPlane(databuf_[i], GetComponentStride(i),
+                planes[i], GetComponentWidth(i),
+                GetComponentWidth(i), scanlines_to_copy);
+      planes[i] += scanlines_to_copy * GetComponentWidth(i);
+    }
+  }
+  return FinishDecode();
+}
+
+LIBYUV_BOOL MJpegDecoder::DecodeToCallback(CallbackFunction fn, void* opaque,
+    int dst_width, int dst_height) {
+  if (dst_width != GetWidth() ||
+      dst_height > GetHeight()) {
+    // ERROR: Bad dimensions
+    return LIBYUV_FALSE;
+  }
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called into jpeglib, it experienced an error sometime during this
+    // function call, and we called longjmp() and rewound the stack to here.
+    // Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (!StartDecode()) {
+    return LIBYUV_FALSE;
+  }
+  SetScanlinePointers(databuf_);
+  int lines_left = dst_height;
+  // TODO(fbarchard): Compute amount of lines to skip to implement vertical crop
+  int skip = (GetHeight() - dst_height) / 2;
+  if (skip > 0) {
+    while (skip >= GetImageScanlinesPerImcuRow()) {
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      skip -= GetImageScanlinesPerImcuRow();
+    }
+    if (skip > 0) {
+      // Have a partial iMCU row left over to skip.
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      for (int i = 0; i < num_outbufs_; ++i) {
+        // TODO(fbarchard): Compute skip to avoid this
+        assert(skip % GetVertSubSampFactor(i) == 0);
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        // Change our own data buffer pointers so we can pass them to the
+        // callback.
+        databuf_[i] += data_to_skip;
+      }
+      int scanlines_to_copy = GetImageScanlinesPerImcuRow() - skip;
+      (*fn)(opaque, databuf_, databuf_strides_, scanlines_to_copy);
+      // Now change them back.
+      for (int i = 0; i < num_outbufs_; ++i) {
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        databuf_[i] -= data_to_skip;
+      }
+      lines_left -= scanlines_to_copy;
+    }
+  }
+  // Read full MCUs until we get to the crop point.
+  for (; lines_left >= GetImageScanlinesPerImcuRow();
+         lines_left -= GetImageScanlinesPerImcuRow()) {
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    (*fn)(opaque, databuf_, databuf_strides_, GetImageScanlinesPerImcuRow());
+  }
+  if (lines_left > 0) {
+    // Have a partial iMCU row left over to decode.
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    (*fn)(opaque, databuf_, databuf_strides_, lines_left);
+  }
+  return FinishDecode();
+}
+
+void init_source(j_decompress_ptr cinfo) {
+  fill_input_buffer(cinfo);
+}
+
+boolean fill_input_buffer(j_decompress_ptr cinfo) {
+  BufferVector* buf_vec = reinterpret_cast<BufferVector*>(cinfo->client_data);
+  if (buf_vec->pos >= buf_vec->len) {
+    assert(0 && "No more data");
+    // ERROR: No more data
+    return FALSE;
+  }
+  cinfo->src->next_input_byte = buf_vec->buffers[buf_vec->pos].data;
+  cinfo->src->bytes_in_buffer = buf_vec->buffers[buf_vec->pos].len;
+  ++buf_vec->pos;
+  return TRUE;
+}
+
+void skip_input_data(j_decompress_ptr cinfo,
+                     long num_bytes) {  // NOLINT
+  cinfo->src->next_input_byte += num_bytes;
+}
+
+void term_source(j_decompress_ptr cinfo) {
+  // Nothing to do.
+}
+
+#ifdef HAVE_SETJMP
+void ErrorHandler(j_common_ptr cinfo) {
+  // This is called when a jpeglib command experiences an error. Unfortunately
+  // jpeglib's error handling model is not very flexible, because it expects the
+  // error handler to not return--i.e., it wants the program to terminate. To
+  // recover from errors we use setjmp() as shown in their example. setjmp() is
+  // C's implementation for the "call with current continuation" functionality
+  // seen in some functional programming languages.
+  // A formatted message can be output, but is unsafe for release.
+#ifdef DEBUG
+  char buf[JMSG_LENGTH_MAX];
+  (*cinfo->err->format_message)(cinfo, buf);
+  // ERROR: Error in jpeglib: buf
+#endif
+
+  SetJmpErrorMgr* mgr = reinterpret_cast<SetJmpErrorMgr*>(cinfo->err);
+  // This rewinds the call stack to the point of the corresponding setjmp()
+  // and causes it to return (for a second time) with value 1.
+  longjmp(mgr->setjmp_buffer, 1);
+}
+#endif
+
+void MJpegDecoder::AllocOutputBuffers(int num_outbufs) {
+  if (num_outbufs != num_outbufs_) {
+    // We could perhaps optimize this case to resize the output buffers without
+    // necessarily having to delete and recreate each one, but it's not worth
+    // it.
+    DestroyOutputBuffers();
+
+    scanlines_ = new uint8** [num_outbufs];
+    scanlines_sizes_ = new int[num_outbufs];
+    databuf_ = new uint8* [num_outbufs];
+    databuf_strides_ = new int[num_outbufs];
+
+    for (int i = 0; i < num_outbufs; ++i) {
+      scanlines_[i] = NULL;
+      scanlines_sizes_[i] = 0;
+      databuf_[i] = NULL;
+      databuf_strides_[i] = 0;
+    }
+
+    num_outbufs_ = num_outbufs;
+  }
+}
+
+void MJpegDecoder::DestroyOutputBuffers() {
+  for (int i = 0; i < num_outbufs_; ++i) {
+    delete [] scanlines_[i];
+    delete [] databuf_[i];
+  }
+  delete [] scanlines_;
+  delete [] databuf_;
+  delete [] scanlines_sizes_;
+  delete [] databuf_strides_;
+  scanlines_ = NULL;
+  databuf_ = NULL;
+  scanlines_sizes_ = NULL;
+  databuf_strides_ = NULL;
+  num_outbufs_ = 0;
+}
+
+// JDCT_IFAST and do_block_smoothing improve performance substantially.
+LIBYUV_BOOL MJpegDecoder::StartDecode() {
+  decompress_struct_->raw_data_out = TRUE;
+  decompress_struct_->dct_method = JDCT_IFAST;  // JDCT_ISLOW is default
+  decompress_struct_->dither_mode = JDITHER_NONE;
+  // Not applicable to 'raw':
+  decompress_struct_->do_fancy_upsampling = (boolean)(LIBYUV_FALSE);
+  // Only for buffered mode:
+  decompress_struct_->enable_2pass_quant = (boolean)(LIBYUV_FALSE);
+  // Blocky but fast:
+  decompress_struct_->do_block_smoothing = (boolean)(LIBYUV_FALSE);
+
+  if (!jpeg_start_decompress(decompress_struct_)) {
+    // ERROR: Couldn't start JPEG decompressor";
+    return LIBYUV_FALSE;
+  }
+  return LIBYUV_TRUE;
+}
+
+LIBYUV_BOOL MJpegDecoder::FinishDecode() {
+  // jpeglib considers it an error if we finish without decoding the whole
+  // image, so we call "abort" rather than "finish".
+  jpeg_abort_decompress(decompress_struct_);
+  return LIBYUV_TRUE;
+}
+
+void MJpegDecoder::SetScanlinePointers(uint8** data) {
+  for (int i = 0; i < num_outbufs_; ++i) {
+    uint8* data_i = data[i];
+    for (int j = 0; j < scanlines_sizes_[i]; ++j) {
+      scanlines_[i][j] = data_i;
+      data_i += GetComponentStride(i);
+    }
+  }
+}
+
+inline LIBYUV_BOOL MJpegDecoder::DecodeImcuRow() {
+  return (unsigned int)(GetImageScanlinesPerImcuRow()) ==
+      jpeg_read_raw_data(decompress_struct_,
+                         scanlines_,
+                         GetImageScanlinesPerImcuRow());
+}
+
+// The helper function which recognizes the jpeg sub-sampling type.
+JpegSubsamplingType MJpegDecoder::JpegSubsamplingTypeHelper(
+    int* subsample_x, int* subsample_y, int number_of_components) {
+  if (number_of_components == 3) {  // Color images.
+    if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 2 && subsample_y[1] == 2 &&
+        subsample_x[2] == 2 && subsample_y[2] == 2) {
+      return kJpegYuv420;
+    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 2 && subsample_y[1] == 1 &&
+        subsample_x[2] == 2 && subsample_y[2] == 1) {
+      return kJpegYuv422;
+    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 1 && subsample_y[1] == 1 &&
+        subsample_x[2] == 1 && subsample_y[2] == 1) {
+      return kJpegYuv444;
+    }
+  } else if (number_of_components == 1) {  // Grey-scale images.
+    if (subsample_x[0] == 1 && subsample_y[0] == 1) {
+      return kJpegYuv400;
+    }
+  }
+  return kJpegUnknown;
+}
+
+}  // namespace libyuv
+#endif  // HAVE_JPEG
+
diff --git a/libvpx/libvpx/third_party/libyuv/source/mjpeg_validate.cc b/libvpx/libvpx/third_party/libyuv/source/mjpeg_validate.cc
new file mode 100644
index 0000000..8edfbe1
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/mjpeg_validate.cc
@@ -0,0 +1,101 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/mjpeg_decoder.h"
+
+#include <string.h>  // For memchr.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Enable this to try scasb implementation.
+// #define ENABLE_SCASB 1
+
+#ifdef ENABLE_SCASB
+
+// Multiple of 1.
+__declspec(naked)
+const uint8* ScanRow_ERMS(const uint8* src, uint32 val, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // src
+    mov        eax, [esp + 8]   // val
+    mov        ecx, [esp + 12]  // count
+    repne scasb
+    jne        sr99
+    mov        eax, edi
+    sub        eax, 1
+    mov        edi, edx
+    ret
+
+  sr99:
+    mov        eax, 0
+    mov        edi, edx
+    ret
+  }
+}
+#endif
+
+// Helper function to scan for EOI marker.
+static LIBYUV_BOOL ScanEOI(const uint8* sample, size_t sample_size) {
+  const uint8* end = sample + sample_size - 1;
+  const uint8* it = sample;
+  for (;;) {
+#ifdef ENABLE_SCASB
+    it = ScanRow_ERMS(it, 0xff, end - it);
+#else
+    it = static_cast<const uint8*>(memchr(it, 0xff, end - it));
+#endif
+    if (it == NULL) {
+      break;
+    }
+    if (it[1] == 0xd9) {
+      return LIBYUV_TRUE;  // Success: Valid jpeg.
+    }
+    ++it;  // Skip over current 0xff.
+  }
+  // ERROR: Invalid jpeg end code not found. Size sample_size
+  return LIBYUV_FALSE;
+}
+
+// Helper function to validate the jpeg appears intact.
+LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size) {
+  const size_t kBackSearchSize = 1024;
+  if (sample_size < 64) {
+    // ERROR: Invalid jpeg size: sample_size
+    return LIBYUV_FALSE;
+  }
+  if (sample[0] != 0xff || sample[1] != 0xd8) {  // Start Of Image
+    // ERROR: Invalid jpeg initial start code
+    return LIBYUV_FALSE;
+  }
+  // Step over SOI marker.
+  sample += 2;
+  sample_size -= 2;
+
+  // Look for the End Of Image (EOI) marker in the end kilobyte of the buffer.
+  if (sample_size > kBackSearchSize) {
+    if (ScanEOI(sample + sample_size - kBackSearchSize, kBackSearchSize)) {
+      return LIBYUV_TRUE;  // Success: Valid jpeg.
+    }
+    // Reduce search size for forward search.
+    sample_size = sample_size - kBackSearchSize + 1;
+  }
+  return ScanEOI(sample, sample_size);
+
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/libvpx/libvpx/third_party/libyuv/source/planar_functions.cc b/libvpx/libvpx/third_party/libyuv/source/planar_functions.cc
new file mode 100644
index 0000000..b96bd50
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/planar_functions.cc
@@ -0,0 +1,2555 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/planar_functions.h"
+
+#include <string.h>  // for memset()
+
+#include "libyuv/cpu_id.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy a plane of data
+LIBYUV_API
+void CopyPlane(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  int y;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+  // Nothing to do.
+  if (src_y == dst_y && src_stride_y == dst_stride_y) {
+    return;
+  }
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height; ++y) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+LIBYUV_API
+void CopyPlane_16(const uint16* src_y, int src_stride_y,
+                  uint16* dst_y, int dst_stride_y,
+                  int width, int height) {
+  int y;
+  void (*CopyRow)(const uint16* src, uint16* dst, int width) = CopyRow_16_C;
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+#if defined(HAS_COPYROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_16_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_16_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_16_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_16_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_16_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_16_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height; ++y) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Copy I422.
+LIBYUV_API
+int I422Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+  return 0;
+}
+
+// Copy I444.
+LIBYUV_API
+int I444Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+  return 0;
+}
+
+// Copy I400.
+LIBYUV_API
+int I400ToI400(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Convert I420 to I400.
+LIBYUV_API
+int I420ToI400(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Mirror a plane of data.
+void MirrorPlane(const uint8* src_y, int src_stride_y,
+                 uint8* dst_y, int dst_stride_y,
+                 int width, int height) {
+  int y;
+  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorRow = MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MirrorRow = MirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorRow = MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorRow = MirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Mirror on mips handle unaligned memory.
+#if defined(HAS_MIRRORROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(dst_y, 4) && IS_ALIGNED(dst_stride_y, 4)) {
+    MirrorRow = MirrorRow_MIPS_DSPR2;
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    MirrorRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Convert YUY2 to I422.
+LIBYUV_API
+int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*YUY2ToUV422Row)(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) =
+      YUY2ToUV422Row_C;
+  void (*YUY2ToYRow)(const uint8* src_yuy2, uint8* dst_y, int pix) =
+      YUY2ToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_yuy2 == width * 2 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_yuy2 = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_SSE2;
+      YUY2ToYRow = YUY2ToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    if (width >= 16) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUV422Row = YUY2ToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Convert UYVY to I422.
+LIBYUV_API
+int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*UYVYToUV422Row)(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) =
+      UYVYToUV422Row_C;
+  void (*UYVYToYRow)(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) = UYVYToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    UYVYToUV422Row = UYVYToUV422Row_Any_SSE2;
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToUV422Row = UYVYToUV422Row_SSE2;
+      UYVYToYRow = UYVYToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    UYVYToUV422Row = UYVYToUV422Row_Any_AVX2;
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToUV422Row = UYVYToUV422Row_AVX2;
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    if (width >= 16) {
+      UYVYToUV422Row = UYVYToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+      UYVYToUV422Row = UYVYToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    UYVYToUV422Row(src_uyvy, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+    src_uyvy += src_stride_uyvy;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Mirror I400 with optional flipping
+LIBYUV_API
+int I400Mirror(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+
+  MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Mirror I420 with optional flipping
+LIBYUV_API
+int I420Mirror(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  MirrorPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
+  MirrorPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
+  return 0;
+}
+
+// ARGB mirror.
+LIBYUV_API
+int ARGBMirror(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+      ARGBMirrorRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMirrorRow = ARGBMirrorRow_AVX2;
+    }
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    ARGBMirrorRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Get a blender that optimized for the CPU and pixel count.
+// As there are 6 blenders to choose from, the caller should try to use
+// the same blend function for all pixels if possible.
+LIBYUV_API
+ARGBBlendRow GetARGBBlend() {
+  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width) = ARGBBlendRow_C;
+#if defined(HAS_ARGBBLENDROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBBlendRow = ARGBBlendRow_SSSE3;
+    return ARGBBlendRow;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBBlendRow = ARGBBlendRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBBlendRow = ARGBBlendRow_NEON;
+  }
+#endif
+  return ARGBBlendRow;
+}
+
+// Alpha Blend 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
+              const uint8* src_argb1, int src_stride_argb1,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  int y;
+  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width) = GetARGBBlend();
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+
+  for (y = 0; y < height; ++y) {
+    ARGBBlendRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Multiply 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*ARGBMultiplyRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBMultiplyRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBMULTIPLYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_NEON;
+    }
+  }
+#endif
+
+  // Multiply plane
+  for (y = 0; y < height; ++y) {
+    ARGBMultiplyRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Add 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
+            const uint8* src_argb1, int src_stride_argb1,
+            uint8* dst_argb, int dst_stride_argb,
+            int width, int height) {
+  int y;
+  void (*ARGBAddRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                     int width) = ARGBAddRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBADDROW_SSE2) && (defined(_MSC_VER) && !defined(__clang__))
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBAddRow = ARGBAddRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBADDROW_SSE2) && !(defined(_MSC_VER) && !defined(__clang__))
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBAddRow = ARGBAddRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAddRow = ARGBAddRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAddRow = ARGBAddRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAddRow = ARGBAddRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_NEON;
+    }
+  }
+#endif
+
+  // Add plane
+  for (y = 0; y < height; ++y) {
+    ARGBAddRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Subtract 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*ARGBSubtractRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBSubtractRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSUBTRACTROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSubtractRow = ARGBSubtractRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSUBTRACTROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSUBTRACTROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_NEON;
+    }
+  }
+#endif
+
+  // Subtract plane
+  for (y = 0; y < height; ++y) {
+    ARGBSubtractRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height) {
+  int y;
+  void (*I422ToBGRARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToBGRARow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_bgra ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra;
+    dst_stride_bgra = -dst_stride_bgra;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_bgra == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_bgra = 0;
+  }
+#if defined(HAS_I422TOBGRAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToBGRARow = I422ToBGRARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToBGRARow = I422ToBGRARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) {
+    I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width);
+    dst_bgra += dst_stride_bgra;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height) {
+  int y;
+  void (*I422ToABGRRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToABGRRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_abgr ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr;
+    dst_stride_abgr = -dst_stride_abgr;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_abgr == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_abgr = 0;
+  }
+#if defined(HAS_I422TOABGRROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToABGRRow = I422ToABGRRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOABGRROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOABGRROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToABGRRow = I422ToABGRRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_AVX2;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width);
+    dst_abgr += dst_stride_abgr;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRGBARow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_rgba ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_rgba == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_rgba = 0;
+  }
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_AVX2;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*NV12ToRGB565Row)(const uint8* y_buf,
+                          const uint8* uv_buf,
+                          uint8* rgb_buf,
+                          int width) = NV12ToRGB565Row_C;
+  if (!src_y || !src_uv || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV12TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToRGB565Row(src_y, src_uv, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV21 to RGB565.
+LIBYUV_API
+int NV21ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_vu, int src_stride_vu,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*NV21ToRGB565Row)(const uint8* y_buf,
+                          const uint8* src_vu,
+                          uint8* rgb_buf,
+                          int width) = NV21ToRGB565Row_C;
+  if (!src_y || !src_vu || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV21TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV21TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV21TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV21ToRGB565Row(src_y, src_vu, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_vu += src_stride_vu;
+    }
+  }
+  return 0;
+}
+
+LIBYUV_API
+void SetPlane(uint8* dst_y, int dst_stride_y,
+              int width, int height,
+              uint32 value) {
+  int y;
+  void (*SetRow)(uint8* dst, uint8 value, int pix) = SetRow_C;
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
+  // Coalesce rows.
+  if (dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    dst_stride_y = 0;
+  }
+#if defined(HAS_SETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SetRow = SetRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SetRow = SetRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SETROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    SetRow = SetRow_Any_X86;
+    if (IS_ALIGNED(width, 4)) {
+      SetRow = SetRow_X86;
+    }
+  }
+#endif
+#if defined(HAS_SETROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    SetRow = SetRow_ERMS;
+  }
+#endif
+
+  // Set plane
+  for (y = 0; y < height; ++y) {
+    SetRow(dst_y, value, width);
+    dst_y += dst_stride_y;
+  }
+}
+
+// Draw a rectangle into I420
+LIBYUV_API
+int I420Rect(uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int x, int y,
+             int width, int height,
+             int value_y, int value_u, int value_v) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  uint8* start_y = dst_y + y * dst_stride_y + x;
+  uint8* start_u = dst_u + (y / 2) * dst_stride_u + (x / 2);
+  uint8* start_v = dst_v + (y / 2) * dst_stride_v + (x / 2);
+  if (!dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0 ||
+      x < 0 || y < 0 ||
+      value_y < 0 || value_y > 255 ||
+      value_u < 0 || value_u > 255 ||
+      value_v < 0 || value_v > 255) {
+    return -1;
+  }
+
+  SetPlane(start_y, dst_stride_y, width, height, value_y);
+  SetPlane(start_u, dst_stride_u, halfwidth, halfheight, value_u);
+  SetPlane(start_v, dst_stride_v, halfwidth, halfheight, value_v);
+  return 0;
+}
+
+// Draw a rectangle into ARGB
+LIBYUV_API
+int ARGBRect(uint8* dst_argb, int dst_stride_argb,
+             int dst_x, int dst_y,
+             int width, int height,
+             uint32 value) {
+  int y;
+  void (*ARGBSetRow)(uint8* dst_argb, uint32 value, int pix) = ARGBSetRow_C;
+  if (!dst_argb ||
+      width <= 0 || height == 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  dst_argb += dst_y * dst_stride_argb + dst_x * 4;
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+
+#if defined(HAS_ARGBSETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBSetRow = ARGBSetRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSetRow = ARGBSetRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSETROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    ARGBSetRow = ARGBSetRow_X86;
+  }
+#endif
+
+  // Set plane
+  for (y = 0; y < height; ++y) {
+    ARGBSetRow(dst_argb, value, width);
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert unattentuated ARGB to preattenuated ARGB.
+// An unattenutated ARGB alpha blend uses the formula
+// p = a * f + (1 - a) * b
+// where
+//   p is output pixel
+//   f is foreground pixel
+//   b is background pixel
+//   a is alpha value from foreground pixel
+// An preattenutated ARGB alpha blend uses the formula
+// p = f + (1 - a) * b
+// where
+//   f is foreground pixel premultiplied by alpha
+
+LIBYUV_API
+int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  int y;
+  void (*ARGBAttenuateRow)(const uint8* src_argb, uint8* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBATTENUATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBAttenuateRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert preattentuated ARGB to unattenuated ARGB.
+LIBYUV_API
+int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height) {
+  int y;
+  void (*ARGBUnattenuateRow)(const uint8* src_argb, uint8* dst_argb,
+                             int width) = ARGBUnattenuateRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBUNATTENUATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBUNATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Neon version.
+
+  for (y = 0; y < height; ++y) {
+    ARGBUnattenuateRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB to Grayed ARGB.
+LIBYUV_API
+int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
+                      int width) = ARGBGrayRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBGRAYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBGrayRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Make a rectangle of ARGB gray scale.
+LIBYUV_API
+int ARGBGray(uint8* dst_argb, int dst_stride_argb,
+             int dst_x, int dst_y,
+             int width, int height) {
+  int y;
+  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
+                      int width) = ARGBGrayRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBGRAYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBGrayRow(dst, dst, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Make a rectangle of ARGB Sepia tone.
+LIBYUV_API
+int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
+              int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBSepiaRow)(uint8* dst_argb, int width) = ARGBSepiaRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSEPIAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
+    ARGBSepiaRow = ARGBSepiaRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBSEPIAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBSepiaRow = ARGBSepiaRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBSepiaRow(dst, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a 4x4 matrix to each ARGB pixel.
+// Note: Normally for shading, but can be used to swizzle or invert.
+LIBYUV_API
+int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    const int8* matrix_argb,
+                    int width, int height) {
+  int y;
+  void (*ARGBColorMatrixRow)(const uint8* src_argb, uint8* dst_argb,
+      const int8* matrix_argb, int width) = ARGBColorMatrixRow_C;
+  if (!src_argb || !dst_argb || !matrix_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOLORMATRIXROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBCOLORMATRIXROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBColorMatrixRow(src_argb, dst_argb, matrix_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a 4x3 matrix to each ARGB pixel.
+// Deprecated.
+LIBYUV_API
+int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
+                   const int8* matrix_rgb,
+                   int dst_x, int dst_y, int width, int height) {
+  SIMD_ALIGNED(int8 matrix_argb[16]);
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !matrix_rgb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+
+  // Convert 4x3 7 bit matrix to 4x4 6 bit matrix.
+  matrix_argb[0] = matrix_rgb[0] / 2;
+  matrix_argb[1] = matrix_rgb[1] / 2;
+  matrix_argb[2] = matrix_rgb[2] / 2;
+  matrix_argb[3] = matrix_rgb[3] / 2;
+  matrix_argb[4] = matrix_rgb[4] / 2;
+  matrix_argb[5] = matrix_rgb[5] / 2;
+  matrix_argb[6] = matrix_rgb[6] / 2;
+  matrix_argb[7] = matrix_rgb[7] / 2;
+  matrix_argb[8] = matrix_rgb[8] / 2;
+  matrix_argb[9] = matrix_rgb[9] / 2;
+  matrix_argb[10] = matrix_rgb[10] / 2;
+  matrix_argb[11] = matrix_rgb[11] / 2;
+  matrix_argb[14] = matrix_argb[13] = matrix_argb[12] = 0;
+  matrix_argb[15] = 64;  // 1.0
+
+  return ARGBColorMatrix((const uint8*)(dst), dst_stride_argb,
+                         dst, dst_stride_argb,
+                         &matrix_argb[0], width, height);
+}
+
+// Apply a color table each ARGB pixel.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                   const uint8* table_argb,
+                   int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+                            int width) = ARGBColorTableRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOLORTABLEROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    ARGBColorTableRow = ARGBColorTableRow_X86;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBColorTableRow(dst, table_argb, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a color table each ARGB pixel but preserve destination alpha.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                  const uint8* table_argb,
+                  int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*RGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+                           int width) = RGBColorTableRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_RGBCOLORTABLEROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    RGBColorTableRow = RGBColorTableRow_X86;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    RGBColorTableRow(dst, table_argb, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// ARGBQuantize is used to posterize art.
+// e.g. rgb / qvalue * qvalue + qvalue / 2
+// But the low levels implement efficiently with 3 parameters, and could be
+// used for other high level operations.
+// dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
+// where scale is 1 / interval_size as a fixed point value.
+// The divide is replaces with a multiply by reciprocal fixed point multiply.
+// Caveat - although SSE2 saturates, the C function does not and should be used
+// with care if doing anything but quantization.
+LIBYUV_API
+int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
+                 int scale, int interval_size, int interval_offset,
+                 int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBQuantizeRow)(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) = ARGBQuantizeRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0 ||
+      interval_size < 1 || interval_size > 255) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBQUANTIZEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBQUANTIZEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBQuantizeRow(dst, scale, interval_size, interval_offset, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Computes table of cumulative sum for image where the value is the sum
+// of all values above and to the left of the entry. Used by ARGBBlur.
+LIBYUV_API
+int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
+                             int32* dst_cumsum, int dst_stride32_cumsum,
+                             int width, int height) {
+  int y;
+  void (*ComputeCumulativeSumRow)(const uint8* row, int32* cumsum,
+      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
+  int32* previous_cumsum = dst_cumsum;
+  if (!dst_cumsum || !src_argb || width <= 0 || height <= 0) {
+    return -1;
+  }
+#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
+  }
+#endif
+  memset(dst_cumsum, 0, width * sizeof(dst_cumsum[0]) * 4);  // 4 int per pixel.
+  for (y = 0; y < height; ++y) {
+    ComputeCumulativeSumRow(src_argb, dst_cumsum, previous_cumsum, width);
+    previous_cumsum = dst_cumsum;
+    dst_cumsum += dst_stride32_cumsum;
+    src_argb += src_stride_argb;
+  }
+  return 0;
+}
+
+// Blur ARGB image.
+// Caller should allocate CumulativeSum table of width * height * 16 bytes
+// aligned to 16 byte boundary. height can be radius * 2 + 2 to save memory
+// as the buffer is treated as circular.
+LIBYUV_API
+int ARGBBlur(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int32* dst_cumsum, int dst_stride32_cumsum,
+             int width, int height, int radius) {
+  int y;
+  void (*ComputeCumulativeSumRow)(const uint8 *row, int32 *cumsum,
+      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
+  void (*CumulativeSumToAverageRow)(const int32* topleft, const int32* botleft,
+      int width, int area, uint8* dst, int count) = CumulativeSumToAverageRow_C;
+  int32* cumsum_bot_row;
+  int32* max_cumsum_bot_row;
+  int32* cumsum_top_row;
+
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  if (radius > height) {
+    radius = height;
+  }
+  if (radius > (width / 2 - 1)) {
+    radius = width / 2 - 1;
+  }
+  if (radius <= 0) {
+    return -1;
+  }
+#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
+    CumulativeSumToAverageRow = CumulativeSumToAverageRow_SSE2;
+  }
+#endif
+  // Compute enough CumulativeSum for first row to be blurred. After this
+  // one row of CumulativeSum is updated at a time.
+  ARGBComputeCumulativeSum(src_argb, src_stride_argb,
+                           dst_cumsum, dst_stride32_cumsum,
+                           width, radius);
+
+  src_argb = src_argb + radius * src_stride_argb;
+  cumsum_bot_row = &dst_cumsum[(radius - 1) * dst_stride32_cumsum];
+
+  max_cumsum_bot_row = &dst_cumsum[(radius * 2 + 2) * dst_stride32_cumsum];
+  cumsum_top_row = &dst_cumsum[0];
+
+  for (y = 0; y < height; ++y) {
+    int top_y = ((y - radius - 1) >= 0) ? (y - radius - 1) : 0;
+    int bot_y = ((y + radius) < height) ? (y + radius) : (height - 1);
+    int area = radius * (bot_y - top_y);
+    int boxwidth = radius * 4;
+    int x;
+    int n;
+
+    // Increment cumsum_top_row pointer with circular buffer wrap around.
+    if (top_y) {
+      cumsum_top_row += dst_stride32_cumsum;
+      if (cumsum_top_row >= max_cumsum_bot_row) {
+        cumsum_top_row = dst_cumsum;
+      }
+    }
+    // Increment cumsum_bot_row pointer with circular buffer wrap around and
+    // then fill in a row of CumulativeSum.
+    if ((y + radius) < height) {
+      const int32* prev_cumsum_bot_row = cumsum_bot_row;
+      cumsum_bot_row += dst_stride32_cumsum;
+      if (cumsum_bot_row >= max_cumsum_bot_row) {
+        cumsum_bot_row = dst_cumsum;
+      }
+      ComputeCumulativeSumRow(src_argb, cumsum_bot_row, prev_cumsum_bot_row,
+                              width);
+      src_argb += src_stride_argb;
+    }
+
+    // Left clipped.
+    for (x = 0; x < radius + 1; ++x) {
+      CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
+                                boxwidth, area, &dst_argb[x * 4], 1);
+      area += (bot_y - top_y);
+      boxwidth += 4;
+    }
+
+    // Middle unclipped.
+    n = (width - 1) - radius - x + 1;
+    CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
+                              boxwidth, area, &dst_argb[x * 4], n);
+
+    // Right clipped.
+    for (x += n; x <= width - 1; ++x) {
+      area -= (bot_y - top_y);
+      boxwidth -= 4;
+      CumulativeSumToAverageRow(cumsum_top_row + (x - radius - 1) * 4,
+                                cumsum_bot_row + (x - radius - 1) * 4,
+                                boxwidth, area, &dst_argb[x * 4], 1);
+    }
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Multiply ARGB image by a specified ARGB value.
+LIBYUV_API
+int ARGBShade(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height, uint32 value) {
+  int y;
+  void (*ARGBShadeRow)(const uint8* src_argb, uint8* dst_argb,
+                       int width, uint32 value) = ARGBShadeRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0 || value == 0u) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSHADEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4)) {
+    ARGBShadeRow = ARGBShadeRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBSHADEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBShadeRow = ARGBShadeRow_NEON;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBShadeRow(src_argb, dst_argb, width, value);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Interpolate 2 ARGB images by specified amount (0 to 255).
+LIBYUV_API
+int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
+                    const uint8* src_argb1, int src_stride_argb1,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height, int interpolation) {
+  int y;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_argb0, 4) && IS_ALIGNED(src_stride_argb0, 4) &&
+      IS_ALIGNED(src_argb1, 4) && IS_ALIGNED(src_stride_argb1, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    InterpolateRow(dst_argb, src_argb0, src_argb1 - src_argb0,
+                   width * 4, interpolation);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Shuffle ARGB channel order.  e.g. BGRA to ARGB.
+LIBYUV_API
+int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
+                uint8* dst_argb, int dst_stride_argb,
+                const uint8* shuffler, int width, int height) {
+  int y;
+  void (*ARGBShuffleRow)(const uint8* src_bgra, uint8* dst_argb,
+                         const uint8* shuffler, int pix) = ARGBShuffleRow_C;
+  if (!src_bgra || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_bgra = src_bgra + (height - 1) * src_stride_bgra;
+    src_stride_bgra = -src_stride_bgra;
+  }
+  // Coalesce rows.
+  if (src_stride_bgra == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_bgra = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSHUFFLEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBShuffleRow = ARGBShuffleRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBShuffleRow = ARGBShuffleRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBShuffleRow = ARGBShuffleRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBShuffleRow = ARGBShuffleRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBShuffleRow(src_bgra, dst_argb, shuffler, width);
+    src_bgra += src_stride_bgra;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Sobel ARGB effect.
+static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
+                        uint8* dst_argb, int dst_stride_argb,
+                        int width, int height,
+                        void (*SobelRow)(const uint8* src_sobelx,
+                                         const uint8* src_sobely,
+                                         uint8* dst, int width)) {
+  int y;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_g, int pix) =
+      ARGBToYJRow_C;
+  void (*SobelYRow)(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) = SobelYRow_C;
+  void (*SobelXRow)(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobely, int width) =
+      SobelXRow_C;
+  const int kEdge = 16;  // Extra pixels at start of row for extrude/align.
+  if (!src_argb  || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+
+#if defined(HAS_SOBELYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelYRow = SobelYRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelYRow = SobelYRow_NEON;
+  }
+#endif
+#if defined(HAS_SOBELXROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelXRow = SobelXRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELXROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelXRow = SobelXRow_NEON;
+  }
+#endif
+  {
+    // 3 rows with edges before/after.
+    const int kRowSize = (width + kEdge + 31) & ~31;
+    align_buffer_64(rows, kRowSize * 2 + (kEdge + kRowSize * 3 + kEdge));
+    uint8* row_sobelx = rows;
+    uint8* row_sobely = rows + kRowSize;
+    uint8* row_y = rows + kRowSize * 2;
+
+    // Convert first row.
+    uint8* row_y0 = row_y + kEdge;
+    uint8* row_y1 = row_y0 + kRowSize;
+    uint8* row_y2 = row_y1 + kRowSize;
+    ARGBToYJRow(src_argb, row_y0, width);
+    row_y0[-1] = row_y0[0];
+    memset(row_y0 + width, row_y0[width - 1], 16);  // Extrude 16 for valgrind.
+    ARGBToYJRow(src_argb, row_y1, width);
+    row_y1[-1] = row_y1[0];
+    memset(row_y1 + width, row_y1[width - 1], 16);
+    memset(row_y2 + width, 0, 16);
+
+    for (y = 0; y < height; ++y) {
+      // Convert next row of ARGB to G.
+      if (y < (height - 1)) {
+        src_argb += src_stride_argb;
+      }
+      ARGBToYJRow(src_argb, row_y2, width);
+      row_y2[-1] = row_y2[0];
+      row_y2[width] = row_y2[width - 1];
+
+      SobelXRow(row_y0 - 1, row_y1 - 1, row_y2 - 1, row_sobelx, width);
+      SobelYRow(row_y0 - 1, row_y2 - 1, row_sobely, width);
+      SobelRow(row_sobelx, row_sobely, dst_argb, width);
+
+      // Cycle thru circular queue of 3 row_y buffers.
+      {
+        uint8* row_yt = row_y0;
+        row_y0 = row_y1;
+        row_y1 = row_y2;
+        row_y2 = row_yt;
+      }
+
+      dst_argb += dst_stride_argb;
+    }
+    free_aligned_buffer_64(rows);
+  }
+  return 0;
+}
+
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) = SobelRow_C;
+#if defined(HAS_SOBELROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelRow = SobelRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SobelRow = SobelRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SOBELROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelRow = SobelRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      SobelRow = SobelRow_NEON;
+    }
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height, SobelRow);
+}
+
+// Sobel ARGB effect with planar output.
+LIBYUV_API
+int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_y, int dst_stride_y,
+                     int width, int height) {
+  void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_, int width) = SobelToPlaneRow_C;
+#if defined(HAS_SOBELTOPLANEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelToPlaneRow = SobelToPlaneRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SobelToPlaneRow = SobelToPlaneRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SOBELTOPLANEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelToPlaneRow = SobelToPlaneRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SobelToPlaneRow = SobelToPlaneRow_NEON;
+    }
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_y, dst_stride_y,
+                      width, height, SobelToPlaneRow);
+}
+
+// SobelXY ARGB effect.
+// Similar to Sobel, but also stores Sobel X in R and Sobel Y in B.  G = Sobel.
+LIBYUV_API
+int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height) {
+  void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) = SobelXYRow_C;
+#if defined(HAS_SOBELXYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelXYRow = SobelXYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SobelXYRow = SobelXYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SOBELXYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelXYRow = SobelXYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      SobelXYRow = SobelXYRow_NEON;
+    }
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height, SobelXYRow);
+}
+
+// Apply a 4x4 polynomial to each ARGB pixel.
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height) {
+  int y;
+  void (*ARGBPolynomialRow)(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) = ARGBPolynomialRow_C;
+  if (!src_argb || !dst_argb || !poly || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBPOLYNOMIALROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBPOLYNOMIALROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && TestCpuFlag(kCpuHasFMA3) &&
+      IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBPolynomialRow(src_argb, dst_argb, poly, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a lumacolortable to each ARGB pixel.
+LIBYUV_API
+int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_argb, int dst_stride_argb,
+                       const uint8* luma,
+                       int width, int height) {
+  int y;
+  void (*ARGBLumaColorTableRow)(const uint8* src_argb, uint8* dst_argb,
+      int width, const uint8* luma, const uint32 lumacoeff) =
+      ARGBLumaColorTableRow_C;
+  if (!src_argb || !dst_argb || !luma || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBLUMACOLORTABLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4)) {
+    ARGBLumaColorTableRow = ARGBLumaColorTableRow_SSSE3;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBLumaColorTableRow(src_argb, dst_argb, width, luma, 0x00264b0f);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Copy Alpha from one ARGB image to another.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  int y;
+  void (*ARGBCopyAlphaRow)(const uint8* src_argb, uint8* dst_argb, int width) =
+      ARGBCopyAlphaRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOPYALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBCOPYALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBCopyAlphaRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Copy a planar Y channel to the alpha channel of a destination ARGB image.
+LIBYUV_API
+int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
+                     uint8* dst_argb, int dst_stride_argb,
+                     int width, int height) {
+  int y;
+  void (*ARGBCopyYToAlphaRow)(const uint8* src_y, uint8* dst_argb, int width) =
+      ARGBCopyYToAlphaRow_C;
+  if (!src_y || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOPYYTOALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
+    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBCOPYYTOALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
+    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBCopyYToAlphaRow(src_y, dst_argb, width);
+    src_y += src_stride_y;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) =
+      SplitUVRow_C;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  if (!src_yuy2 ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+#if defined(HAS_SPLITUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitUVRow = SplitUVRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow = SplitUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitUVRow = SplitUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitUVRow = SplitUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow = SplitUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+
+  {
+    int awidth = halfwidth * 2;
+    // 2 rows of uv
+    align_buffer_64(rows, awidth * 2);
+
+    for (y = 0; y < height - 1; y += 2) {
+      // Split Y from UV.
+      SplitUVRow(src_yuy2, dst_y, rows, awidth);
+      SplitUVRow(src_yuy2 + src_stride_yuy2, dst_y + dst_stride_y,
+                 rows + awidth, awidth);
+      InterpolateRow(dst_uv, rows, awidth, awidth, 128);
+      src_yuy2 += src_stride_yuy2 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_uv += dst_stride_uv;
+    }
+    if (height & 1) {
+      // Split Y from UV.
+      SplitUVRow(src_yuy2, dst_y, dst_uv, width);
+    }
+    free_aligned_buffer_64(rows);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) =
+      SplitUVRow_C;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  if (!src_uyvy ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+#if defined(HAS_SPLITUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitUVRow = SplitUVRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow = SplitUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitUVRow = SplitUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitUVRow = SplitUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow = SplitUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+
+  {
+    int awidth = halfwidth * 2;
+    // 2 rows of uv
+    align_buffer_64(rows, awidth * 2);
+
+    for (y = 0; y < height - 1; y += 2) {
+      // Split Y from UV.
+      SplitUVRow(src_uyvy, rows, dst_y, awidth);
+      SplitUVRow(src_uyvy + src_stride_uyvy, rows + awidth,
+                 dst_y + dst_stride_y, awidth);
+      InterpolateRow(dst_uv, rows, awidth, awidth, 128);
+      src_uyvy += src_stride_uyvy * 2;
+      dst_y += dst_stride_y * 2;
+      dst_uv += dst_stride_uv;
+    }
+    if (height & 1) {
+      // Split Y from UV.
+      SplitUVRow(src_uyvy, dst_y, dst_uv, width);
+    }
+    free_aligned_buffer_64(rows);
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/rotate.cc b/libvpx/libvpx/third_party/libyuv/source/rotate.cc
new file mode 100644
index 0000000..be3d589
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/rotate.cc
@@ -0,0 +1,496 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/convert.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+LIBYUV_API
+void TransposePlane(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  int i = height;
+  void (*TransposeWx8)(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride, int width) = TransposeWx8_C;
+#if defined(HAS_TRANSPOSEWX8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    TransposeWx8 = TransposeWx8_NEON;
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX8_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    TransposeWx8 = TransposeWx8_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeWx8 = TransposeWx8_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX8_FAST_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    TransposeWx8 = TransposeWx8_Fast_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      TransposeWx8 = TransposeWx8_Fast_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX8_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    if (IS_ALIGNED(width, 4) &&
+        IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+      TransposeWx8 = TransposeWx8_Fast_MIPS_DSPR2;
+    } else {
+      TransposeWx8 = TransposeWx8_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  // Work across the source in 8x8 tiles
+  while (i >= 8) {
+    TransposeWx8(src, src_stride, dst, dst_stride, width);
+    src += 8 * src_stride;    // Go down 8 rows.
+    dst += 8;                 // Move over 8 columns.
+    i -= 8;
+  }
+
+  if (i > 0) {
+    TransposeWxH_C(src, src_stride, dst, dst_stride, width, i);
+  }
+}
+
+LIBYUV_API
+void RotatePlane90(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride,
+                   int width, int height) {
+  // Rotate by 90 is a transpose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  TransposePlane(src, src_stride, dst, dst_stride, width, height);
+}
+
+LIBYUV_API
+void RotatePlane270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  // Rotate by 270 is a transpose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  TransposePlane(src, src_stride, dst, dst_stride, width, height);
+}
+
+LIBYUV_API
+void RotatePlane180(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64(row, width);
+  const uint8* src_bot = src + src_stride * (height - 1);
+  uint8* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorRow = MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MirrorRow = MirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorRow = MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorRow = MirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Mirror on mips handle unaligned memory.
+#if defined(HAS_MIRRORROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst, 4) && IS_ALIGNED(dst_stride, 4)) {
+    MirrorRow = MirrorRow_MIPS_DSPR2;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    MirrorRow(src, row, width);  // Mirror first row into a buffer
+    src += src_stride;
+    MirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    dst += dst_stride;
+    CopyRow(row, dst_bot, width);  // Copy first mirrored row into last
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+LIBYUV_API
+void TransposeUV(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  int i = height;
+  void (*TransposeUVWx8)(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) = TransposeUVWx8_C;
+#if defined(HAS_TRANSPOSEUVWX8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    TransposeUVWx8 = TransposeUVWx8_NEON;
+  }
+#endif
+#if defined(HAS_TRANSPOSEUVWX8_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
+    TransposeUVWx8 = TransposeUVWx8_SSE2;
+  }
+#endif
+#if defined(HAS_TRANSPOSEUVWx8_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+    TransposeUVWx8 = TransposeUVWx8_MIPS_DSPR2;
+  }
+#endif
+
+  // Work through the source in 8x8 tiles.
+  while (i >= 8) {
+    TransposeUVWx8(src, src_stride,
+                   dst_a, dst_stride_a,
+                   dst_b, dst_stride_b,
+                   width);
+    src += 8 * src_stride;    // Go down 8 rows.
+    dst_a += 8;               // Move over 8 columns.
+    dst_b += 8;               // Move over 8 columns.
+    i -= 8;
+  }
+
+  if (i > 0) {
+    TransposeUVWxH_C(src, src_stride,
+                     dst_a, dst_stride_a,
+                     dst_b, dst_stride_b,
+                     width, i);
+  }
+}
+
+LIBYUV_API
+void RotateUV90(const uint8* src, int src_stride,
+                uint8* dst_a, int dst_stride_a,
+                uint8* dst_b, int dst_stride_b,
+                int width, int height) {
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+
+  TransposeUV(src, src_stride,
+              dst_a, dst_stride_a,
+              dst_b, dst_stride_b,
+              width, height);
+}
+
+LIBYUV_API
+void RotateUV270(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  dst_a += dst_stride_a * (width - 1);
+  dst_b += dst_stride_b * (width - 1);
+  dst_stride_a = -dst_stride_a;
+  dst_stride_b = -dst_stride_b;
+
+  TransposeUV(src, src_stride,
+              dst_a, dst_stride_a,
+              dst_b, dst_stride_b,
+              width, height);
+}
+
+// Rotate 180 is a horizontal and vertical flip.
+LIBYUV_API
+void RotateUV180(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  int i;
+  void (*MirrorRowUV)(const uint8* src, uint8* dst_u, uint8* dst_v, int width) =
+      MirrorUVRow_C;
+#if defined(HAS_MIRRORUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    MirrorRowUV = MirrorUVRow_NEON;
+  }
+#endif
+#if defined(HAS_MIRRORROW_UV_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16)) {
+    MirrorRowUV = MirrorUVRow_SSSE3;
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+    MirrorRowUV = MirrorUVRow_MIPS_DSPR2;
+  }
+#endif
+
+  dst_a += dst_stride_a * (height - 1);
+  dst_b += dst_stride_b * (height - 1);
+
+  for (i = 0; i < height; ++i) {
+    MirrorRowUV(src, dst_a, dst_b, width);
+    src += src_stride;
+    dst_a -= dst_stride_a;
+    dst_b -= dst_stride_b;
+  }
+}
+
+LIBYUV_API
+int RotatePlane(const uint8* src, int src_stride,
+                uint8* dst, int dst_stride,
+                int width, int height,
+                enum RotationMode mode) {
+  if (!src || width <= 0 || height == 0 || !dst) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src = src + (height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane(src, src_stride,
+                dst, dst_stride,
+                width, height);
+      return 0;
+    case kRotate90:
+      RotatePlane90(src, src_stride,
+                    dst, dst_stride,
+                    width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src, src_stride,
+                     dst, dst_stride,
+                     width, height);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src, src_stride,
+                     dst, dst_stride,
+                     width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int I420Rotate(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 ||
+      !dst_y || !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return I420Copy(src_y, src_stride_y,
+                      src_u, src_stride_u,
+                      src_v, src_stride_v,
+                      dst_y, dst_stride_y,
+                      dst_u, dst_stride_u,
+                      dst_v, dst_stride_v,
+                      width, height);
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y,
+                    dst_y, dst_stride_y,
+                    width, height);
+      RotatePlane90(src_u, src_stride_u,
+                    dst_u, dst_stride_u,
+                    halfwidth, halfheight);
+      RotatePlane90(src_v, src_stride_v,
+                    dst_v, dst_stride_v,
+                    halfwidth, halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotatePlane270(src_u, src_stride_u,
+                     dst_u, dst_stride_u,
+                     halfwidth, halfheight);
+      RotatePlane270(src_v, src_stride_v,
+                     dst_v, dst_stride_v,
+                     halfwidth, halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotatePlane180(src_u, src_stride_u,
+                     dst_u, dst_stride_u,
+                     halfwidth, halfheight);
+      RotatePlane180(src_v, src_stride_v,
+                     dst_v, dst_stride_v,
+                     halfwidth, halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
+                     const uint8* src_uv, int src_stride_uv,
+                     uint8* dst_y, int dst_stride_y,
+                     uint8* dst_u, int dst_stride_u,
+                     uint8* dst_v, int dst_stride_v,
+                     int width, int height,
+                     enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_uv || width <= 0 || height == 0 ||
+      !dst_y || !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
+    src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return NV12ToI420(src_y, src_stride_y,
+                        src_uv, src_stride_uv,
+                        dst_y, dst_stride_y,
+                        dst_u, dst_stride_u,
+                        dst_v, dst_stride_v,
+                        width, height);
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y,
+                    dst_y, dst_stride_y,
+                    width, height);
+      RotateUV90(src_uv, src_stride_uv,
+                 dst_u, dst_stride_u,
+                 dst_v, dst_stride_v,
+                 halfwidth, halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotateUV270(src_uv, src_stride_uv,
+                  dst_u, dst_stride_u,
+                  dst_v, dst_stride_v,
+                  halfwidth, halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotateUV180(src_uv, src_stride_uv,
+                  dst_u, dst_stride_u,
+                  dst_v, dst_stride_v,
+                  halfwidth, halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/rotate_any.cc b/libvpx/libvpx/third_party/libyuv/source/rotate_any.cc
new file mode 100644
index 0000000..4d6eb34
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/rotate_any.cc
@@ -0,0 +1,55 @@
+/*
+ *  Copyright 2015 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+#include "libyuv/rotate_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define TANY(NAMEANY, TPOS_SIMD, TPOS_C, MASK)                                 \
+    void NAMEANY(const uint8* src, int src_stride,                             \
+                 uint8* dst, int dst_stride, int width) {                      \
+      int r = width & MASK;                                                    \
+      int n = width - r;                                                       \
+      if (n > 0) {                                                             \
+        TPOS_SIMD(src, src_stride, dst, dst_stride, n);                        \
+      }                                                                        \
+      TPOS_C(src + n, src_stride, dst + n * dst_stride, dst_stride, r);        \
+    }
+
+#ifdef HAS_TRANSPOSEWX8_NEON
+TANY(TransposeWx8_Any_NEON, TransposeWx8_NEON, TransposeWx8_C, 7)
+#endif
+#ifdef HAS_TRANSPOSEWX8_SSSE3
+TANY(TransposeWx8_Any_SSSE3, TransposeWx8_SSSE3, TransposeWx8_C, 7)
+#endif
+#ifdef HAS_TRANSPOSEWX8_FAST_SSSE3
+TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, TransposeWx8_C, 15)
+#endif
+#ifdef HAS_TRANSPOSEWX8_MIPS_DSPR2
+TANY(TransposeWx8_Any_MIPS_DSPR2, TransposeWx8_MIPS_DSPR2, TransposeWx8_C, 7)
+#endif
+
+#undef TANY
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+
+
+
+
diff --git a/libvpx/libvpx/third_party/libyuv/source/rotate_argb.cc b/libvpx/libvpx/third_party/libyuv/source/rotate_argb.cc
new file mode 100644
index 0000000..787c0ad
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/rotate_argb.cc
@@ -0,0 +1,205 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/convert.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// ARGBScale has a function to copy pixels to a row, striding each source
+// pixel by a constant.
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || \
+    (defined(__x86_64__) && !defined(__native_client__)) || defined(__i386__))
+#define HAS_SCALEARGBROWDOWNEVEN_SSE2
+void ScaleARGBRowDownEven_SSE2(const uint8* src_ptr, int src_stride,
+                               int src_stepx, uint8* dst_ptr, int dst_width);
+#endif
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
+#define HAS_SCALEARGBROWDOWNEVEN_NEON
+void ScaleARGBRowDownEven_NEON(const uint8* src_ptr, int src_stride,
+                               int src_stepx, uint8* dst_ptr, int dst_width);
+#endif
+
+void ScaleARGBRowDownEven_C(const uint8* src_ptr, int,
+                            int src_stepx, uint8* dst_ptr, int dst_width);
+
+static void ARGBTranspose(const uint8* src, int src_stride,
+                          uint8* dst, int dst_stride, int width, int height) {
+  int i;
+  int src_pixel_step = src_stride >> 2;
+  void (*ScaleARGBRowDownEven)(const uint8* src_ptr, int src_stride,
+      int src_step, uint8* dst_ptr, int dst_width) = ScaleARGBRowDownEven_C;
+#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(height, 4)) {  // Width of dest.
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(height, 4)) {  // Width of dest.
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_NEON;
+  }
+#endif
+
+  for (i = 0; i < width; ++i) {  // column of source to row of dest.
+    ScaleARGBRowDownEven(src, 0, src_pixel_step, dst, height);
+    dst += dst_stride;
+    src += 4;
+  }
+}
+
+void ARGBRotate90(const uint8* src, int src_stride,
+                  uint8* dst, int dst_stride, int width, int height) {
+  // Rotate by 90 is a ARGBTranspose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+}
+
+void ARGBRotate270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride, int width, int height) {
+  // Rotate by 270 is a ARGBTranspose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+}
+
+void ARGBRotate180(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride, int width, int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64(row, width * 4);
+  const uint8* src_bot = src + src_stride * (height - 1);
+  uint8* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+      ARGBMirrorRow_C;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMirrorRow = ARGBMirrorRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width * 4, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    ARGBMirrorRow(src, row, width);  // Mirror first row into a buffer
+    ARGBMirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    CopyRow(row, dst_bot, width * 4);  // Copy first mirrored row into last
+    src += src_stride;
+    dst += dst_stride;
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+LIBYUV_API
+int ARGBRotate(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb, int width, int height,
+               enum RotationMode mode) {
+  if (!src_argb || width <= 0 || height == 0 || !dst_argb) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return ARGBCopy(src_argb, src_stride_argb,
+                      dst_argb, dst_stride_argb,
+                      width, height);
+    case kRotate90:
+      ARGBRotate90(src_argb, src_stride_argb,
+                   dst_argb, dst_stride_argb,
+                   width, height);
+      return 0;
+    case kRotate270:
+      ARGBRotate270(src_argb, src_stride_argb,
+                    dst_argb, dst_stride_argb,
+                    width, height);
+      return 0;
+    case kRotate180:
+      ARGBRotate180(src_argb, src_stride_argb,
+                    dst_argb, dst_stride_argb,
+                    width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/rotate_common.cc b/libvpx/libvpx/third_party/libyuv/source/rotate_common.cc
new file mode 100644
index 0000000..b33a9a0
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/rotate_common.cc
@@ -0,0 +1,92 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+void TransposeWx8_C(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst[0] = src[0 * src_stride];
+    dst[1] = src[1 * src_stride];
+    dst[2] = src[2 * src_stride];
+    dst[3] = src[3 * src_stride];
+    dst[4] = src[4 * src_stride];
+    dst[5] = src[5 * src_stride];
+    dst[6] = src[6 * src_stride];
+    dst[7] = src[7 * src_stride];
+    ++src;
+    dst += dst_stride;
+  }
+}
+
+void TransposeUVWx8_C(const uint8* src, int src_stride,
+                      uint8* dst_a, int dst_stride_a,
+                      uint8* dst_b, int dst_stride_b, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst_a[0] = src[0 * src_stride + 0];
+    dst_b[0] = src[0 * src_stride + 1];
+    dst_a[1] = src[1 * src_stride + 0];
+    dst_b[1] = src[1 * src_stride + 1];
+    dst_a[2] = src[2 * src_stride + 0];
+    dst_b[2] = src[2 * src_stride + 1];
+    dst_a[3] = src[3 * src_stride + 0];
+    dst_b[3] = src[3 * src_stride + 1];
+    dst_a[4] = src[4 * src_stride + 0];
+    dst_b[4] = src[4 * src_stride + 1];
+    dst_a[5] = src[5 * src_stride + 0];
+    dst_b[5] = src[5 * src_stride + 1];
+    dst_a[6] = src[6 * src_stride + 0];
+    dst_b[6] = src[6 * src_stride + 1];
+    dst_a[7] = src[7 * src_stride + 0];
+    dst_b[7] = src[7 * src_stride + 1];
+    src += 2;
+    dst_a += dst_stride_a;
+    dst_b += dst_stride_b;
+  }
+}
+
+void TransposeWxH_C(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst[i * dst_stride + j] = src[j * src_stride + i];
+    }
+  }
+}
+
+void TransposeUVWxH_C(const uint8* src, int src_stride,
+                      uint8* dst_a, int dst_stride_a,
+                      uint8* dst_b, int dst_stride_b,
+                      int width, int height) {
+  int i;
+  for (i = 0; i < width * 2; i += 2) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst_a[j + ((i >> 1) * dst_stride_a)] = src[i + (j * src_stride)];
+      dst_b[j + ((i >> 1) * dst_stride_b)] = src[i + (j * src_stride) + 1];
+    }
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/rotate_gcc.cc b/libvpx/libvpx/third_party/libyuv/source/rotate_gcc.cc
new file mode 100644
index 0000000..fd385bc
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/rotate_gcc.cc
@@ -0,0 +1,493 @@
+/*
+ *  Copyright 2015 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__i386__) || (defined(__x86_64__) && !defined(__native_client__)))
+void TransposeWx8_SSSE3(const uint8* src, int src_stride,
+                        uint8* dst, int dst_stride, int width) {
+  asm volatile (
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    ".p2align  2                                 \n"
+  "1:                                            \n"
+    "movq       (%0),%%xmm0                      \n"
+    "movq       (%0,%3),%%xmm1                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm1,%%xmm0                    \n"
+    "movq       (%0),%%xmm2                      \n"
+    "movdqa     %%xmm0,%%xmm1                    \n"
+    "palignr    $0x8,%%xmm1,%%xmm1               \n"
+    "movq       (%0,%3),%%xmm3                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm3,%%xmm2                    \n"
+    "movdqa     %%xmm2,%%xmm3                    \n"
+    "movq       (%0),%%xmm4                      \n"
+    "palignr    $0x8,%%xmm3,%%xmm3               \n"
+    "movq       (%0,%3),%%xmm5                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm5,%%xmm4                    \n"
+    "movdqa     %%xmm4,%%xmm5                    \n"
+    "movq       (%0),%%xmm6                      \n"
+    "palignr    $0x8,%%xmm5,%%xmm5               \n"
+    "movq       (%0,%3),%%xmm7                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm7,%%xmm6                    \n"
+    "neg        %3                               \n"
+    "movdqa     %%xmm6,%%xmm7                    \n"
+    "lea        0x8(%0,%3,8),%0                  \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    "neg        %3                               \n"
+     // Second round of bit swap.
+    "punpcklwd  %%xmm2,%%xmm0                    \n"
+    "punpcklwd  %%xmm3,%%xmm1                    \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "palignr    $0x8,%%xmm2,%%xmm2               \n"
+    "palignr    $0x8,%%xmm3,%%xmm3               \n"
+    "punpcklwd  %%xmm6,%%xmm4                    \n"
+    "punpcklwd  %%xmm7,%%xmm5                    \n"
+    "movdqa     %%xmm4,%%xmm6                    \n"
+    "movdqa     %%xmm5,%%xmm7                    \n"
+    "palignr    $0x8,%%xmm6,%%xmm6               \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    "punpckldq  %%xmm4,%%xmm0                    \n"
+    "movq       %%xmm0,(%1)                      \n"
+    "movdqa     %%xmm0,%%xmm4                    \n"
+    "palignr    $0x8,%%xmm4,%%xmm4               \n"
+    "movq       %%xmm4,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "punpckldq  %%xmm6,%%xmm2                    \n"
+    "movdqa     %%xmm2,%%xmm6                    \n"
+    "movq       %%xmm2,(%1)                      \n"
+    "palignr    $0x8,%%xmm6,%%xmm6               \n"
+    "punpckldq  %%xmm5,%%xmm1                    \n"
+    "movq       %%xmm6,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "movdqa     %%xmm1,%%xmm5                    \n"
+    "movq       %%xmm1,(%1)                      \n"
+    "palignr    $0x8,%%xmm5,%%xmm5               \n"
+    "movq       %%xmm5,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "punpckldq  %%xmm7,%%xmm3                    \n"
+    "movq       %%xmm3,(%1)                      \n"
+    "movdqa     %%xmm3,%%xmm7                    \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    "sub        $0x8,%2                          \n"
+    "movq       %%xmm7,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "jg         1b                               \n"
+    : "+r"(src),    // %0
+      "+r"(dst),    // %1
+      "+r"(width)   // %2
+    : "r"((intptr_t)(src_stride)),  // %3
+      "r"((intptr_t)(dst_stride))   // %4
+    : "memory", "cc",
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(__i386__)  && !defined(__clang__)
+void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b, int width);
+  asm (
+    DECLARE_FUNCTION(TransposeUVWx8_SSE2)
+    "push   %ebx                               \n"
+    "push   %esi                               \n"
+    "push   %edi                               \n"
+    "push   %ebp                               \n"
+    "mov    0x14(%esp),%eax                    \n"
+    "mov    0x18(%esp),%edi                    \n"
+    "mov    0x1c(%esp),%edx                    \n"
+    "mov    0x20(%esp),%esi                    \n"
+    "mov    0x24(%esp),%ebx                    \n"
+    "mov    0x28(%esp),%ebp                    \n"
+    "mov    %esp,%ecx                          \n"
+    "sub    $0x14,%esp                         \n"
+    "and    $0xfffffff0,%esp                   \n"
+    "mov    %ecx,0x10(%esp)                    \n"
+    "mov    0x2c(%ecx),%ecx                    \n"
+
+"1:                                            \n"
+    "movdqu (%eax),%xmm0                       \n"
+    "movdqu (%eax,%edi,1),%xmm1                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm0,%xmm7                        \n"
+    "punpcklbw %xmm1,%xmm0                     \n"
+    "punpckhbw %xmm1,%xmm7                     \n"
+    "movdqa %xmm7,%xmm1                        \n"
+    "movdqu (%eax),%xmm2                       \n"
+    "movdqu (%eax,%edi,1),%xmm3                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm2,%xmm7                        \n"
+    "punpcklbw %xmm3,%xmm2                     \n"
+    "punpckhbw %xmm3,%xmm7                     \n"
+    "movdqa %xmm7,%xmm3                        \n"
+    "movdqu (%eax),%xmm4                       \n"
+    "movdqu (%eax,%edi,1),%xmm5                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm4,%xmm7                        \n"
+    "punpcklbw %xmm5,%xmm4                     \n"
+    "punpckhbw %xmm5,%xmm7                     \n"
+    "movdqa %xmm7,%xmm5                        \n"
+    "movdqu (%eax),%xmm6                       \n"
+    "movdqu (%eax,%edi,1),%xmm7                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqu %xmm5,(%esp)                       \n"
+    "neg    %edi                               \n"
+    "movdqa %xmm6,%xmm5                        \n"
+    "punpcklbw %xmm7,%xmm6                     \n"
+    "punpckhbw %xmm7,%xmm5                     \n"
+    "movdqa %xmm5,%xmm7                        \n"
+    "lea    0x10(%eax,%edi,8),%eax             \n"
+    "neg    %edi                               \n"
+    "movdqa %xmm0,%xmm5                        \n"
+    "punpcklwd %xmm2,%xmm0                     \n"
+    "punpckhwd %xmm2,%xmm5                     \n"
+    "movdqa %xmm5,%xmm2                        \n"
+    "movdqa %xmm1,%xmm5                        \n"
+    "punpcklwd %xmm3,%xmm1                     \n"
+    "punpckhwd %xmm3,%xmm5                     \n"
+    "movdqa %xmm5,%xmm3                        \n"
+    "movdqa %xmm4,%xmm5                        \n"
+    "punpcklwd %xmm6,%xmm4                     \n"
+    "punpckhwd %xmm6,%xmm5                     \n"
+    "movdqa %xmm5,%xmm6                        \n"
+    "movdqu (%esp),%xmm5                       \n"
+    "movdqu %xmm6,(%esp)                       \n"
+    "movdqa %xmm5,%xmm6                        \n"
+    "punpcklwd %xmm7,%xmm5                     \n"
+    "punpckhwd %xmm7,%xmm6                     \n"
+    "movdqa %xmm6,%xmm7                        \n"
+    "movdqa %xmm0,%xmm6                        \n"
+    "punpckldq %xmm4,%xmm0                     \n"
+    "punpckhdq %xmm4,%xmm6                     \n"
+    "movdqa %xmm6,%xmm4                        \n"
+    "movdqu (%esp),%xmm6                       \n"
+    "movlpd %xmm0,(%edx)                       \n"
+    "movhpd %xmm0,(%ebx)                       \n"
+    "movlpd %xmm4,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm4,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm2,%xmm0                        \n"
+    "punpckldq %xmm6,%xmm2                     \n"
+    "movlpd %xmm2,(%edx)                       \n"
+    "movhpd %xmm2,(%ebx)                       \n"
+    "punpckhdq %xmm6,%xmm0                     \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm1,%xmm0                        \n"
+    "punpckldq %xmm5,%xmm1                     \n"
+    "movlpd %xmm1,(%edx)                       \n"
+    "movhpd %xmm1,(%ebx)                       \n"
+    "punpckhdq %xmm5,%xmm0                     \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm3,%xmm0                        \n"
+    "punpckldq %xmm7,%xmm3                     \n"
+    "movlpd %xmm3,(%edx)                       \n"
+    "movhpd %xmm3,(%ebx)                       \n"
+    "punpckhdq %xmm7,%xmm0                     \n"
+    "sub    $0x8,%ecx                          \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "jg     1b                                 \n"
+    "mov    0x10(%esp),%esp                    \n"
+    "pop    %ebp                               \n"
+    "pop    %edi                               \n"
+    "pop    %esi                               \n"
+    "pop    %ebx                               \n"
+#if defined(__native_client__)
+    "pop    %ecx                               \n"
+    "and    $0xffffffe0,%ecx                   \n"
+    "jmp    *%ecx                              \n"
+#else
+    "ret                                       \n"
+#endif
+);
+#endif
+#if !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \
+    defined(__x86_64__)
+// 64 bit version has enough registers to do 16x8 to 8x16 at a time.
+void TransposeWx8_Fast_SSSE3(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride, int width) {
+  asm volatile (
+  // Read in the data from the source pointer.
+  // First round of bit swap.
+  ".p2align  2                                 \n"
+"1:                                            \n"
+  "movdqu     (%0),%%xmm0                      \n"
+  "movdqu     (%0,%3),%%xmm1                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpcklbw  %%xmm1,%%xmm0                    \n"
+  "punpckhbw  %%xmm1,%%xmm8                    \n"
+  "movdqu     (%0),%%xmm2                      \n"
+  "movdqa     %%xmm0,%%xmm1                    \n"
+  "movdqa     %%xmm8,%%xmm9                    \n"
+  "palignr    $0x8,%%xmm1,%%xmm1               \n"
+  "palignr    $0x8,%%xmm9,%%xmm9               \n"
+  "movdqu     (%0,%3),%%xmm3                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm2,%%xmm10                   \n"
+  "punpcklbw  %%xmm3,%%xmm2                    \n"
+  "punpckhbw  %%xmm3,%%xmm10                   \n"
+  "movdqa     %%xmm2,%%xmm3                    \n"
+  "movdqa     %%xmm10,%%xmm11                  \n"
+  "movdqu     (%0),%%xmm4                      \n"
+  "palignr    $0x8,%%xmm3,%%xmm3               \n"
+  "palignr    $0x8,%%xmm11,%%xmm11             \n"
+  "movdqu     (%0,%3),%%xmm5                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm4,%%xmm12                   \n"
+  "punpcklbw  %%xmm5,%%xmm4                    \n"
+  "punpckhbw  %%xmm5,%%xmm12                   \n"
+  "movdqa     %%xmm4,%%xmm5                    \n"
+  "movdqa     %%xmm12,%%xmm13                  \n"
+  "movdqu     (%0),%%xmm6                      \n"
+  "palignr    $0x8,%%xmm5,%%xmm5               \n"
+  "palignr    $0x8,%%xmm13,%%xmm13             \n"
+  "movdqu     (%0,%3),%%xmm7                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm6,%%xmm14                   \n"
+  "punpcklbw  %%xmm7,%%xmm6                    \n"
+  "punpckhbw  %%xmm7,%%xmm14                   \n"
+  "neg        %3                               \n"
+  "movdqa     %%xmm6,%%xmm7                    \n"
+  "movdqa     %%xmm14,%%xmm15                  \n"
+  "lea        0x10(%0,%3,8),%0                 \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  "neg        %3                               \n"
+   // Second round of bit swap.
+  "punpcklwd  %%xmm2,%%xmm0                    \n"
+  "punpcklwd  %%xmm3,%%xmm1                    \n"
+  "movdqa     %%xmm0,%%xmm2                    \n"
+  "movdqa     %%xmm1,%%xmm3                    \n"
+  "palignr    $0x8,%%xmm2,%%xmm2               \n"
+  "palignr    $0x8,%%xmm3,%%xmm3               \n"
+  "punpcklwd  %%xmm6,%%xmm4                    \n"
+  "punpcklwd  %%xmm7,%%xmm5                    \n"
+  "movdqa     %%xmm4,%%xmm6                    \n"
+  "movdqa     %%xmm5,%%xmm7                    \n"
+  "palignr    $0x8,%%xmm6,%%xmm6               \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "punpcklwd  %%xmm10,%%xmm8                   \n"
+  "punpcklwd  %%xmm11,%%xmm9                   \n"
+  "movdqa     %%xmm8,%%xmm10                   \n"
+  "movdqa     %%xmm9,%%xmm11                   \n"
+  "palignr    $0x8,%%xmm10,%%xmm10             \n"
+  "palignr    $0x8,%%xmm11,%%xmm11             \n"
+  "punpcklwd  %%xmm14,%%xmm12                  \n"
+  "punpcklwd  %%xmm15,%%xmm13                  \n"
+  "movdqa     %%xmm12,%%xmm14                  \n"
+  "movdqa     %%xmm13,%%xmm15                  \n"
+  "palignr    $0x8,%%xmm14,%%xmm14             \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  // Third round of bit swap.
+  // Write to the destination pointer.
+  "punpckldq  %%xmm4,%%xmm0                    \n"
+  "movq       %%xmm0,(%1)                      \n"
+  "movdqa     %%xmm0,%%xmm4                    \n"
+  "palignr    $0x8,%%xmm4,%%xmm4               \n"
+  "movq       %%xmm4,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm6,%%xmm2                    \n"
+  "movdqa     %%xmm2,%%xmm6                    \n"
+  "movq       %%xmm2,(%1)                      \n"
+  "palignr    $0x8,%%xmm6,%%xmm6               \n"
+  "punpckldq  %%xmm5,%%xmm1                    \n"
+  "movq       %%xmm6,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "movdqa     %%xmm1,%%xmm5                    \n"
+  "movq       %%xmm1,(%1)                      \n"
+  "palignr    $0x8,%%xmm5,%%xmm5               \n"
+  "movq       %%xmm5,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm7,%%xmm3                    \n"
+  "movq       %%xmm3,(%1)                      \n"
+  "movdqa     %%xmm3,%%xmm7                    \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "movq       %%xmm7,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm12,%%xmm8                   \n"
+  "movq       %%xmm8,(%1)                      \n"
+  "movdqa     %%xmm8,%%xmm12                   \n"
+  "palignr    $0x8,%%xmm12,%%xmm12             \n"
+  "movq       %%xmm12,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm14,%%xmm10                  \n"
+  "movdqa     %%xmm10,%%xmm14                  \n"
+  "movq       %%xmm10,(%1)                     \n"
+  "palignr    $0x8,%%xmm14,%%xmm14             \n"
+  "punpckldq  %%xmm13,%%xmm9                   \n"
+  "movq       %%xmm14,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "movdqa     %%xmm9,%%xmm13                   \n"
+  "movq       %%xmm9,(%1)                      \n"
+  "palignr    $0x8,%%xmm13,%%xmm13             \n"
+  "movq       %%xmm13,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm15,%%xmm11                  \n"
+  "movq       %%xmm11,(%1)                     \n"
+  "movdqa     %%xmm11,%%xmm15                  \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  "sub        $0x10,%2                         \n"
+  "movq       %%xmm15,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "jg         1b                               \n"
+  : "+r"(src),    // %0
+    "+r"(dst),    // %1
+    "+r"(width)   // %2
+  : "r"((intptr_t)(src_stride)),  // %3
+    "r"((intptr_t)(dst_stride))   // %4
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
+    "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13",  "xmm14",  "xmm15"
+);
+}
+
+void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b, int width) {
+  asm volatile (
+  // Read in the data from the source pointer.
+  // First round of bit swap.
+  ".p2align  2                                 \n"
+"1:                                            \n"
+  "movdqu     (%0),%%xmm0                      \n"
+  "movdqu     (%0,%4),%%xmm1                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpcklbw  %%xmm1,%%xmm0                    \n"
+  "punpckhbw  %%xmm1,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm1                    \n"
+  "movdqu     (%0),%%xmm2                      \n"
+  "movdqu     (%0,%4),%%xmm3                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm2,%%xmm8                    \n"
+  "punpcklbw  %%xmm3,%%xmm2                    \n"
+  "punpckhbw  %%xmm3,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm3                    \n"
+  "movdqu     (%0),%%xmm4                      \n"
+  "movdqu     (%0,%4),%%xmm5                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm4,%%xmm8                    \n"
+  "punpcklbw  %%xmm5,%%xmm4                    \n"
+  "punpckhbw  %%xmm5,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm5                    \n"
+  "movdqu     (%0),%%xmm6                      \n"
+  "movdqu     (%0,%4),%%xmm7                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm6,%%xmm8                    \n"
+  "punpcklbw  %%xmm7,%%xmm6                    \n"
+  "neg        %4                               \n"
+  "lea        0x10(%0,%4,8),%0                 \n"
+  "punpckhbw  %%xmm7,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm7                    \n"
+  "neg        %4                               \n"
+   // Second round of bit swap.
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "movdqa     %%xmm1,%%xmm9                    \n"
+  "punpckhwd  %%xmm2,%%xmm8                    \n"
+  "punpckhwd  %%xmm3,%%xmm9                    \n"
+  "punpcklwd  %%xmm2,%%xmm0                    \n"
+  "punpcklwd  %%xmm3,%%xmm1                    \n"
+  "movdqa     %%xmm8,%%xmm2                    \n"
+  "movdqa     %%xmm9,%%xmm3                    \n"
+  "movdqa     %%xmm4,%%xmm8                    \n"
+  "movdqa     %%xmm5,%%xmm9                    \n"
+  "punpckhwd  %%xmm6,%%xmm8                    \n"
+  "punpckhwd  %%xmm7,%%xmm9                    \n"
+  "punpcklwd  %%xmm6,%%xmm4                    \n"
+  "punpcklwd  %%xmm7,%%xmm5                    \n"
+  "movdqa     %%xmm8,%%xmm6                    \n"
+  "movdqa     %%xmm9,%%xmm7                    \n"
+  // Third round of bit swap.
+  // Write to the destination pointer.
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpckldq  %%xmm4,%%xmm0                    \n"
+  "movlpd     %%xmm0,(%1)                      \n"  // Write back U channel
+  "movhpd     %%xmm0,(%2)                      \n"  // Write back V channel
+  "punpckhdq  %%xmm4,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm2,%%xmm8                    \n"
+  "punpckldq  %%xmm6,%%xmm2                    \n"
+  "movlpd     %%xmm2,(%1)                      \n"
+  "movhpd     %%xmm2,(%2)                      \n"
+  "punpckhdq  %%xmm6,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm1,%%xmm8                    \n"
+  "punpckldq  %%xmm5,%%xmm1                    \n"
+  "movlpd     %%xmm1,(%1)                      \n"
+  "movhpd     %%xmm1,(%2)                      \n"
+  "punpckhdq  %%xmm5,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm3,%%xmm8                    \n"
+  "punpckldq  %%xmm7,%%xmm3                    \n"
+  "movlpd     %%xmm3,(%1)                      \n"
+  "movhpd     %%xmm3,(%2)                      \n"
+  "punpckhdq  %%xmm7,%%xmm8                    \n"
+  "sub        $0x8,%3                          \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "jg         1b                               \n"
+  : "+r"(src),    // %0
+    "+r"(dst_a),  // %1
+    "+r"(dst_b),  // %2
+    "+r"(width)   // %3
+  : "r"((intptr_t)(src_stride)),    // %4
+    "r"((intptr_t)(dst_stride_a)),  // %5
+    "r"((intptr_t)(dst_stride_b))   // %6
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
+    "xmm8", "xmm9"
+);
+}
+#endif
+#endif
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/rotate_mips.cc b/libvpx/libvpx/third_party/libyuv/source/rotate_mips.cc
new file mode 100644
index 0000000..efe6bd9
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/rotate_mips.cc
@@ -0,0 +1,484 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_MIPS) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+void TransposeWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride, int width) {
+   __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "sll              $t2, %[src_stride], 0x1          \n" // src_stride x 2
+      "sll              $t4, %[src_stride], 0x2          \n" // src_stride x 4
+      "sll              $t9, %[src_stride], 0x3          \n" // src_stride x 8
+      "addu             $t3, $t2, %[src_stride]          \n"
+      "addu             $t5, $t4, %[src_stride]          \n"
+      "addu             $t6, $t2, $t4                    \n"
+      "andi             $t0, %[dst], 0x3                 \n"
+      "andi             $t1, %[dst_stride], 0x3          \n"
+      "or               $t0, $t0, $t1                    \n"
+      "bnez             $t0, 11f                         \n"
+      " subu            $t7, $t9, %[src_stride]          \n"
+//dst + dst_stride word aligned
+    "1:                                                  \n"
+      "lbu              $t0, 0(%[src])                   \n"
+      "lbux             $t1, %[src_stride](%[src])       \n"
+      "lbux             $t8, $t2(%[src])                 \n"
+      "lbux             $t9, $t3(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s0, $t8, $t0                    \n"
+      "lbux             $t0, $t4(%[src])                 \n"
+      "lbux             $t1, $t5(%[src])                 \n"
+      "lbux             $t8, $t6(%[src])                 \n"
+      "lbux             $t9, $t7(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s1, $t8, $t0                    \n"
+      "sw               $s0, 0(%[dst])                   \n"
+      "addiu            %[width], -1                     \n"
+      "addiu            %[src], 1                        \n"
+      "sw               $s1, 4(%[dst])                   \n"
+      "bnez             %[width], 1b                     \n"
+      " addu            %[dst], %[dst], %[dst_stride]    \n"
+      "b                2f                               \n"
+//dst + dst_stride unaligned
+   "11:                                                  \n"
+      "lbu              $t0, 0(%[src])                   \n"
+      "lbux             $t1, %[src_stride](%[src])       \n"
+      "lbux             $t8, $t2(%[src])                 \n"
+      "lbux             $t9, $t3(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s0, $t8, $t0                    \n"
+      "lbux             $t0, $t4(%[src])                 \n"
+      "lbux             $t1, $t5(%[src])                 \n"
+      "lbux             $t8, $t6(%[src])                 \n"
+      "lbux             $t9, $t7(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s1, $t8, $t0                    \n"
+      "swr              $s0, 0(%[dst])                   \n"
+      "swl              $s0, 3(%[dst])                   \n"
+      "addiu            %[width], -1                     \n"
+      "addiu            %[src], 1                        \n"
+      "swr              $s1, 4(%[dst])                   \n"
+      "swl              $s1, 7(%[dst])                   \n"
+      "bnez             %[width], 11b                    \n"
+       "addu             %[dst], %[dst], %[dst_stride]   \n"
+    "2:                                                  \n"
+      ".set pop                                          \n"
+      :[src] "+r" (src),
+       [dst] "+r" (dst),
+       [width] "+r" (width)
+      :[src_stride] "r" (src_stride),
+       [dst_stride] "r" (dst_stride)
+      : "t0", "t1",  "t2", "t3", "t4", "t5",
+        "t6", "t7", "t8", "t9",
+        "s0", "s1"
+  );
+}
+
+void TransposeWx8_Fast_MIPS_DSPR2(const uint8* src, int src_stride,
+                                  uint8* dst, int dst_stride, int width) {
+  __asm__ __volatile__ (
+      ".set noat                                         \n"
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "beqz             %[width], 2f                     \n"
+      " sll             $t2, %[src_stride], 0x1          \n"  // src_stride x 2
+      "sll              $t4, %[src_stride], 0x2          \n"  // src_stride x 4
+      "sll              $t9, %[src_stride], 0x3          \n"  // src_stride x 8
+      "addu             $t3, $t2, %[src_stride]          \n"
+      "addu             $t5, $t4, %[src_stride]          \n"
+      "addu             $t6, $t2, $t4                    \n"
+
+      "srl              $AT, %[width], 0x2               \n"
+      "andi             $t0, %[dst], 0x3                 \n"
+      "andi             $t1, %[dst_stride], 0x3          \n"
+      "or               $t0, $t0, $t1                    \n"
+      "bnez             $t0, 11f                         \n"
+      " subu            $t7, $t9, %[src_stride]          \n"
+//dst + dst_stride word aligned
+      "1:                                                \n"
+      "lw               $t0, 0(%[src])                   \n"
+      "lwx              $t1, %[src_stride](%[src])       \n"
+      "lwx              $t8, $t2(%[src])                 \n"
+      "lwx              $t9, $t3(%[src])                 \n"
+
+// t0 = | 30 | 20 | 10 | 00 |
+// t1 = | 31 | 21 | 11 | 01 |
+// t8 = | 32 | 22 | 12 | 02 |
+// t9 = | 33 | 23 | 13 | 03 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 21 | 01 | 20 | 00 |
+  // s1 = | 23 | 03 | 22 | 02 |
+  // s2 = | 31 | 11 | 30 | 10 |
+  // s3 = | 33 | 13 | 32 | 12 |
+
+      "precr.qb.ph     $s4, $s1, $s0                     \n"
+      "precrq.qb.ph    $s5, $s1, $s0                     \n"
+      "precr.qb.ph     $s6, $s3, $s2                     \n"
+      "precrq.qb.ph    $s7, $s3, $s2                     \n"
+
+  // s4 = | 03 | 02 | 01 | 00 |
+  // s5 = | 23 | 22 | 21 | 20 |
+  // s6 = | 13 | 12 | 11 | 10 |
+  // s7 = | 33 | 32 | 31 | 30 |
+
+      "lwx              $t0, $t4(%[src])                 \n"
+      "lwx              $t1, $t5(%[src])                 \n"
+      "lwx              $t8, $t6(%[src])                 \n"
+      "lwx              $t9, $t7(%[src])                 \n"
+
+// t0 = | 34 | 24 | 14 | 04 |
+// t1 = | 35 | 25 | 15 | 05 |
+// t8 = | 36 | 26 | 16 | 06 |
+// t9 = | 37 | 27 | 17 | 07 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 25 | 05 | 24 | 04 |
+  // s1 = | 27 | 07 | 26 | 06 |
+  // s2 = | 35 | 15 | 34 | 14 |
+  // s3 = | 37 | 17 | 36 | 16 |
+
+      "precr.qb.ph     $t0, $s1, $s0                     \n"
+      "precrq.qb.ph    $t1, $s1, $s0                     \n"
+      "precr.qb.ph     $t8, $s3, $s2                     \n"
+      "precrq.qb.ph    $t9, $s3, $s2                     \n"
+
+  // t0 = | 07 | 06 | 05 | 04 |
+  // t1 = | 27 | 26 | 25 | 24 |
+  // t8 = | 17 | 16 | 15 | 14 |
+  // t9 = | 37 | 36 | 35 | 34 |
+
+      "addu            $s0, %[dst], %[dst_stride]        \n"
+      "addu            $s1, $s0, %[dst_stride]           \n"
+      "addu            $s2, $s1, %[dst_stride]           \n"
+
+      "sw              $s4, 0(%[dst])                    \n"
+      "sw              $t0, 4(%[dst])                    \n"
+      "sw              $s6, 0($s0)                       \n"
+      "sw              $t8, 4($s0)                       \n"
+      "sw              $s5, 0($s1)                       \n"
+      "sw              $t1, 4($s1)                       \n"
+      "sw              $s7, 0($s2)                       \n"
+      "sw              $t9, 4($s2)                       \n"
+
+      "addiu            $AT, -1                          \n"
+      "addiu            %[src], 4                        \n"
+
+      "bnez             $AT, 1b                          \n"
+      " addu            %[dst], $s2, %[dst_stride]       \n"
+      "b                2f                               \n"
+//dst + dst_stride unaligned
+      "11:                                               \n"
+      "lw               $t0, 0(%[src])                   \n"
+      "lwx              $t1, %[src_stride](%[src])       \n"
+      "lwx              $t8, $t2(%[src])                 \n"
+      "lwx              $t9, $t3(%[src])                 \n"
+
+// t0 = | 30 | 20 | 10 | 00 |
+// t1 = | 31 | 21 | 11 | 01 |
+// t8 = | 32 | 22 | 12 | 02 |
+// t9 = | 33 | 23 | 13 | 03 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 21 | 01 | 20 | 00 |
+  // s1 = | 23 | 03 | 22 | 02 |
+  // s2 = | 31 | 11 | 30 | 10 |
+  // s3 = | 33 | 13 | 32 | 12 |
+
+      "precr.qb.ph     $s4, $s1, $s0                     \n"
+      "precrq.qb.ph    $s5, $s1, $s0                     \n"
+      "precr.qb.ph     $s6, $s3, $s2                     \n"
+      "precrq.qb.ph    $s7, $s3, $s2                     \n"
+
+  // s4 = | 03 | 02 | 01 | 00 |
+  // s5 = | 23 | 22 | 21 | 20 |
+  // s6 = | 13 | 12 | 11 | 10 |
+  // s7 = | 33 | 32 | 31 | 30 |
+
+      "lwx              $t0, $t4(%[src])                 \n"
+      "lwx              $t1, $t5(%[src])                 \n"
+      "lwx              $t8, $t6(%[src])                 \n"
+      "lwx              $t9, $t7(%[src])                 \n"
+
+// t0 = | 34 | 24 | 14 | 04 |
+// t1 = | 35 | 25 | 15 | 05 |
+// t8 = | 36 | 26 | 16 | 06 |
+// t9 = | 37 | 27 | 17 | 07 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 25 | 05 | 24 | 04 |
+  // s1 = | 27 | 07 | 26 | 06 |
+  // s2 = | 35 | 15 | 34 | 14 |
+  // s3 = | 37 | 17 | 36 | 16 |
+
+      "precr.qb.ph     $t0, $s1, $s0                     \n"
+      "precrq.qb.ph    $t1, $s1, $s0                     \n"
+      "precr.qb.ph     $t8, $s3, $s2                     \n"
+      "precrq.qb.ph    $t9, $s3, $s2                     \n"
+
+  // t0 = | 07 | 06 | 05 | 04 |
+  // t1 = | 27 | 26 | 25 | 24 |
+  // t8 = | 17 | 16 | 15 | 14 |
+  // t9 = | 37 | 36 | 35 | 34 |
+
+      "addu            $s0, %[dst], %[dst_stride]        \n"
+      "addu            $s1, $s0, %[dst_stride]           \n"
+      "addu            $s2, $s1, %[dst_stride]           \n"
+
+      "swr              $s4, 0(%[dst])                   \n"
+      "swl              $s4, 3(%[dst])                   \n"
+      "swr              $t0, 4(%[dst])                   \n"
+      "swl              $t0, 7(%[dst])                   \n"
+      "swr              $s6, 0($s0)                      \n"
+      "swl              $s6, 3($s0)                      \n"
+      "swr              $t8, 4($s0)                      \n"
+      "swl              $t8, 7($s0)                      \n"
+      "swr              $s5, 0($s1)                      \n"
+      "swl              $s5, 3($s1)                      \n"
+      "swr              $t1, 4($s1)                      \n"
+      "swl              $t1, 7($s1)                      \n"
+      "swr              $s7, 0($s2)                      \n"
+      "swl              $s7, 3($s2)                      \n"
+      "swr              $t9, 4($s2)                      \n"
+      "swl              $t9, 7($s2)                      \n"
+
+      "addiu            $AT, -1                          \n"
+      "addiu            %[src], 4                        \n"
+
+      "bnez             $AT, 11b                         \n"
+      " addu            %[dst], $s2, %[dst_stride]       \n"
+      "2:                                                \n"
+      ".set pop                                          \n"
+      ".set at                                           \n"
+      :[src] "+r" (src),
+       [dst] "+r" (dst),
+       [width] "+r" (width)
+      :[src_stride] "r" (src_stride),
+       [dst_stride] "r" (dst_stride)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9",
+        "s0", "s1", "s2", "s3", "s4", "s5", "s6", "s7"
+  );
+}
+
+void TransposeUVWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                               uint8* dst_a, int dst_stride_a,
+                               uint8* dst_b, int dst_stride_b,
+                               int width) {
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "beqz            %[width], 2f                      \n"
+      " sll            $t2, %[src_stride], 0x1           \n" // src_stride x 2
+      "sll             $t4, %[src_stride], 0x2           \n" // src_stride x 4
+      "sll             $t9, %[src_stride], 0x3           \n" // src_stride x 8
+      "addu            $t3, $t2, %[src_stride]           \n"
+      "addu            $t5, $t4, %[src_stride]           \n"
+      "addu            $t6, $t2, $t4                     \n"
+      "subu            $t7, $t9, %[src_stride]           \n"
+      "srl             $t1, %[width], 1                  \n"
+
+// check word aligment for dst_a, dst_b, dst_stride_a and dst_stride_b
+      "andi            $t0, %[dst_a], 0x3                \n"
+      "andi            $t8, %[dst_b], 0x3                \n"
+      "or              $t0, $t0, $t8                     \n"
+      "andi            $t8, %[dst_stride_a], 0x3         \n"
+      "andi            $s5, %[dst_stride_b], 0x3         \n"
+      "or              $t8, $t8, $s5                     \n"
+      "or              $t0, $t0, $t8                     \n"
+      "bnez            $t0, 11f                          \n"
+      " nop                                              \n"
+// dst + dst_stride word aligned (both, a & b dst addresses)
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src])                    \n" // |B0|A0|b0|a0|
+      "lwx             $t8, %[src_stride](%[src])        \n" // |B1|A1|b1|a1|
+      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
+      "lwx             $t9, $t2(%[src])                  \n" // |B2|A2|b2|a2|
+      "lwx             $s0, $t3(%[src])                  \n" // |B3|A3|b3|a3|
+      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B1|A1|B0|A0|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B3|A3|B2|A2|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A3|A2|A1|A0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B3|B2|B1|B0|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b1|a1|b0|a0|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b3|a3|b2|a2|
+
+      "sw              $s3, 0($s5)                       \n"
+      "sw              $s4, 0($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a3|a2|a1|a0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b3|b2|b1|b0|
+
+      "lwx             $t0, $t4(%[src])                  \n" // |B4|A4|b4|a4|
+      "lwx             $t8, $t5(%[src])                  \n" // |B5|A5|b5|a5|
+      "lwx             $t9, $t6(%[src])                  \n" // |B6|A6|b6|a6|
+      "lwx             $s0, $t7(%[src])                  \n" // |B7|A7|b7|a7|
+      "sw              $s3, 0(%[dst_a])                  \n"
+      "sw              $s4, 0(%[dst_b])                  \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B5|A5|B4|A4|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B6|A6|B7|A7|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A7|A6|A5|A4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B7|B6|B5|B4|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b5|a5|b4|a4|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b7|a7|b6|a6|
+      "sw              $s3, 4($s5)                       \n"
+      "sw              $s4, 4($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a7|a6|a5|a4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b7|b6|b5|b4|
+
+      "addiu           %[src], 4                         \n"
+      "addiu           $t1, -1                           \n"
+      "sll             $t0, %[dst_stride_a], 1           \n"
+      "sll             $t8, %[dst_stride_b], 1           \n"
+      "sw              $s3, 4(%[dst_a])                  \n"
+      "sw              $s4, 4(%[dst_b])                  \n"
+      "addu            %[dst_a], %[dst_a], $t0           \n"
+      "bnez            $t1, 1b                           \n"
+      " addu           %[dst_b], %[dst_b], $t8           \n"
+      "b               2f                                \n"
+      " nop                                              \n"
+
+// dst_a or dst_b or dst_stride_a or dst_stride_b not word aligned
+   "11:                                                  \n"
+      "lw              $t0, 0(%[src])                    \n" // |B0|A0|b0|a0|
+      "lwx             $t8, %[src_stride](%[src])        \n" // |B1|A1|b1|a1|
+      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
+      "lwx             $t9, $t2(%[src])                  \n" // |B2|A2|b2|a2|
+      "lwx             $s0, $t3(%[src])                  \n" // |B3|A3|b3|a3|
+      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B1|A1|B0|A0|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B3|A3|B2|A2|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A3|A2|A1|A0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B3|B2|B1|B0|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b1|a1|b0|a0|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b3|a3|b2|a2|
+
+      "swr             $s3, 0($s5)                       \n"
+      "swl             $s3, 3($s5)                       \n"
+      "swr             $s4, 0($s6)                       \n"
+      "swl             $s4, 3($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a3|a2|a1|a0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b3|b2|b1|b0|
+
+      "lwx             $t0, $t4(%[src])                  \n" // |B4|A4|b4|a4|
+      "lwx             $t8, $t5(%[src])                  \n" // |B5|A5|b5|a5|
+      "lwx             $t9, $t6(%[src])                  \n" // |B6|A6|b6|a6|
+      "lwx             $s0, $t7(%[src])                  \n" // |B7|A7|b7|a7|
+      "swr             $s3, 0(%[dst_a])                  \n"
+      "swl             $s3, 3(%[dst_a])                  \n"
+      "swr             $s4, 0(%[dst_b])                  \n"
+      "swl             $s4, 3(%[dst_b])                  \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B5|A5|B4|A4|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B6|A6|B7|A7|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A7|A6|A5|A4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B7|B6|B5|B4|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b5|a5|b4|a4|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b7|a7|b6|a6|
+
+      "swr             $s3, 4($s5)                       \n"
+      "swl             $s3, 7($s5)                       \n"
+      "swr             $s4, 4($s6)                       \n"
+      "swl             $s4, 7($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a7|a6|a5|a4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b7|b6|b5|b4|
+
+      "addiu           %[src], 4                         \n"
+      "addiu           $t1, -1                           \n"
+      "sll             $t0, %[dst_stride_a], 1           \n"
+      "sll             $t8, %[dst_stride_b], 1           \n"
+      "swr             $s3, 4(%[dst_a])                  \n"
+      "swl             $s3, 7(%[dst_a])                  \n"
+      "swr             $s4, 4(%[dst_b])                  \n"
+      "swl             $s4, 7(%[dst_b])                  \n"
+      "addu            %[dst_a], %[dst_a], $t0           \n"
+      "bnez            $t1, 11b                          \n"
+      " addu           %[dst_b], %[dst_b], $t8           \n"
+
+      "2:                                                \n"
+      ".set pop                                          \n"
+      : [src] "+r" (src),
+        [dst_a] "+r" (dst_a),
+        [dst_b] "+r" (dst_b),
+        [width] "+r" (width),
+        [src_stride] "+r" (src_stride)
+      : [dst_stride_a] "r" (dst_stride_a),
+        [dst_stride_b] "r" (dst_stride_b)
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+        "t6", "t7", "t8", "t9",
+        "s0", "s1", "s2", "s3",
+        "s4", "s5", "s6"
+  );
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/rotate_neon.cc b/libvpx/libvpx/third_party/libyuv/source/rotate_neon.cc
new file mode 100644
index 0000000..76043b3
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/rotate_neon.cc
@@ -0,0 +1,535 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+static uvec8 kVTbl4x4Transpose =
+  { 0,  4,  8, 12,  1,  5,  9, 13,  2,  6, 10, 14,  3,  7, 11, 15 };
+
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride,
+                       int width) {
+  const uint8* src_temp = NULL;
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub         %5, #8                        \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    ".p2align  2                               \n"
+    "1:                                        \n"
+      "mov         %0, %1                      \n"
+
+      MEMACCESS(0)
+      "vld1.8      {d0}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d1}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d2}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d3}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d4}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d5}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d6}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d7}, [%0]                  \n"
+
+      "vtrn.8      d1, d0                      \n"
+      "vtrn.8      d3, d2                      \n"
+      "vtrn.8      d5, d4                      \n"
+      "vtrn.8      d7, d6                      \n"
+
+      "vtrn.16     d1, d3                      \n"
+      "vtrn.16     d0, d2                      \n"
+      "vtrn.16     d5, d7                      \n"
+      "vtrn.16     d4, d6                      \n"
+
+      "vtrn.32     d1, d5                      \n"
+      "vtrn.32     d0, d4                      \n"
+      "vtrn.32     d3, d7                      \n"
+      "vtrn.32     d2, d6                      \n"
+
+      "vrev16.8    q0, q0                      \n"
+      "vrev16.8    q1, q1                      \n"
+      "vrev16.8    q2, q2                      \n"
+      "vrev16.8    q3, q3                      \n"
+
+      "mov         %0, %3                      \n"
+
+    MEMACCESS(0)
+      "vst1.8      {d1}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d0}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d3}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d2}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d5}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d4}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d7}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d6}, [%0]                  \n"
+
+      "add         %1, #8                      \n"  // src += 8
+      "add         %3, %3, %4, lsl #3          \n"  // dst += 8 * dst_stride
+      "subs        %5,  #8                     \n"  // w   -= 8
+      "bge         1b                          \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %5, #8                        \n"
+    "beq         4f                            \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %5, #2                        \n"
+    "blt         3f                            \n"
+
+    "cmp         %5, #4                        \n"
+    "blt         2f                            \n"
+
+    // 4x8 block
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld1.32     {d0[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d0[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d1[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d1[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d2[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d2[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d3[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d3[1]}, [%0]                 \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(6)
+    "vld1.8      {q3}, [%6]                    \n"
+
+    "vtbl.8      d4, {d0, d1}, d6              \n"
+    "vtbl.8      d5, {d0, d1}, d7              \n"
+    "vtbl.8      d0, {d2, d3}, d6              \n"
+    "vtbl.8      d1, {d2, d3}, d7              \n"
+
+    // TODO(frkoenig): Rework shuffle above to
+    // write out with 4 instead of 8 writes.
+    MEMACCESS(0)
+    "vst1.32     {d4[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d4[1]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d5[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d5[1]}, [%0]                 \n"
+
+    "add         %0, %3, #4                    \n"
+    MEMACCESS(0)
+    "vst1.32     {d0[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d0[1]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d1[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d1[1]}, [%0]                 \n"
+
+    "add         %1, #4                        \n"  // src += 4
+    "add         %3, %3, %4, lsl #2            \n"  // dst += 4 * dst_stride
+    "subs        %5,  #4                       \n"  // w   -= 4
+    "beq         4f                            \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %5, #2                        \n"
+    "blt         3f                            \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[2]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[2]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[3]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[3]}, [%0]                 \n"
+
+    "vtrn.8      d0, d1                        \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(0)
+    "vst1.64     {d0}, [%0], %4                \n"
+    MEMACCESS(0)
+    "vst1.64     {d1}, [%0]                    \n"
+
+    "add         %1, #2                        \n"  // src += 2
+    "add         %3, %3, %4, lsl #1            \n"  // dst += 2 * dst_stride
+    "subs        %5,  #2                       \n"  // w   -= 2
+    "beq         4f                            \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[0]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[1]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[2]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[3]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[4]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[5]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[6]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[7]}, [%1]                 \n"
+
+    MEMACCESS(3)
+    "vst1.64     {d0}, [%3]                    \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),          // %0
+      "+r"(src),               // %1
+      "+r"(src_stride),        // %2
+      "+r"(dst),               // %3
+      "+r"(dst_stride),        // %4
+      "+r"(width)              // %5
+    : "r"(&kVTbl4x4Transpose)  // %6
+    : "memory", "cc", "q0", "q1", "q2", "q3"
+  );
+}
+
+static uvec8 kVTbl4x4TransposeDi =
+  { 0,  8,  1,  9,  2, 10,  3, 11,  4, 12,  5, 13,  6, 14,  7, 15 };
+
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) {
+  const uint8* src_temp = NULL;
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub         %7, #8                        \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    ".p2align  2                               \n"
+    "1:                                        \n"
+      "mov         %0, %1                      \n"
+
+      MEMACCESS(0)
+      "vld2.8      {d0,  d1},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d2,  d3},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d4,  d5},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d6,  d7},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d16, d17}, [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d18, d19}, [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d20, d21}, [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d22, d23}, [%0]            \n"
+
+      "vtrn.8      q1, q0                      \n"
+      "vtrn.8      q3, q2                      \n"
+      "vtrn.8      q9, q8                      \n"
+      "vtrn.8      q11, q10                    \n"
+
+      "vtrn.16     q1, q3                      \n"
+      "vtrn.16     q0, q2                      \n"
+      "vtrn.16     q9, q11                     \n"
+      "vtrn.16     q8, q10                     \n"
+
+      "vtrn.32     q1, q9                      \n"
+      "vtrn.32     q0, q8                      \n"
+      "vtrn.32     q3, q11                     \n"
+      "vtrn.32     q2, q10                     \n"
+
+      "vrev16.8    q0, q0                      \n"
+      "vrev16.8    q1, q1                      \n"
+      "vrev16.8    q2, q2                      \n"
+      "vrev16.8    q3, q3                      \n"
+      "vrev16.8    q8, q8                      \n"
+      "vrev16.8    q9, q9                      \n"
+      "vrev16.8    q10, q10                    \n"
+      "vrev16.8    q11, q11                    \n"
+
+      "mov         %0, %3                      \n"
+
+    MEMACCESS(0)
+      "vst1.8      {d2},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d0},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d6},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d4},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d18}, [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d16}, [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d22}, [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d20}, [%0]                 \n"
+
+      "mov         %0, %5                      \n"
+
+    MEMACCESS(0)
+      "vst1.8      {d3},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d1},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d7},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d5},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d19}, [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d17}, [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d23}, [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d21}, [%0]                 \n"
+
+      "add         %1, #8*2                    \n"  // src   += 8*2
+      "add         %3, %3, %4, lsl #3          \n"  // dst_a += 8 * dst_stride_a
+      "add         %5, %5, %6, lsl #3          \n"  // dst_b += 8 * dst_stride_b
+      "subs        %7,  #8                     \n"  // w     -= 8
+      "bge         1b                          \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %7, #8                        \n"
+    "beq         4f                            \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %7, #2                        \n"
+    "blt         3f                            \n"
+
+    "cmp         %7, #4                        \n"
+    "blt         2f                            \n"
+
+    // TODO(frkoenig): Clean this up
+    // 4x8 block
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld1.64     {d0}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d1}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d2}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d3}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d4}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d5}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d6}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d7}, [%0]                    \n"
+
+    MEMACCESS(8)
+    "vld1.8      {q15}, [%8]                   \n"
+
+    "vtrn.8      q0, q1                        \n"
+    "vtrn.8      q2, q3                        \n"
+
+    "vtbl.8      d16, {d0, d1}, d30            \n"
+    "vtbl.8      d17, {d0, d1}, d31            \n"
+    "vtbl.8      d18, {d2, d3}, d30            \n"
+    "vtbl.8      d19, {d2, d3}, d31            \n"
+    "vtbl.8      d20, {d4, d5}, d30            \n"
+    "vtbl.8      d21, {d4, d5}, d31            \n"
+    "vtbl.8      d22, {d6, d7}, d30            \n"
+    "vtbl.8      d23, {d6, d7}, d31            \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(0)
+    "vst1.32     {d16[0]},  [%0], %4           \n"
+    MEMACCESS(0)
+    "vst1.32     {d16[1]},  [%0], %4           \n"
+    MEMACCESS(0)
+    "vst1.32     {d17[0]},  [%0], %4           \n"
+    MEMACCESS(0)
+    "vst1.32     {d17[1]},  [%0], %4           \n"
+
+    "add         %0, %3, #4                    \n"
+    MEMACCESS(0)
+    "vst1.32     {d20[0]}, [%0], %4            \n"
+    MEMACCESS(0)
+    "vst1.32     {d20[1]}, [%0], %4            \n"
+    MEMACCESS(0)
+    "vst1.32     {d21[0]}, [%0], %4            \n"
+    MEMACCESS(0)
+    "vst1.32     {d21[1]}, [%0]                \n"
+
+    "mov         %0, %5                        \n"
+
+    MEMACCESS(0)
+    "vst1.32     {d18[0]}, [%0], %6            \n"
+    MEMACCESS(0)
+    "vst1.32     {d18[1]}, [%0], %6            \n"
+    MEMACCESS(0)
+    "vst1.32     {d19[0]}, [%0], %6            \n"
+    MEMACCESS(0)
+    "vst1.32     {d19[1]}, [%0], %6            \n"
+
+    "add         %0, %5, #4                    \n"
+    MEMACCESS(0)
+    "vst1.32     {d22[0]},  [%0], %6           \n"
+    MEMACCESS(0)
+    "vst1.32     {d22[1]},  [%0], %6           \n"
+    MEMACCESS(0)
+    "vst1.32     {d23[0]},  [%0], %6           \n"
+    MEMACCESS(0)
+    "vst1.32     {d23[1]},  [%0]               \n"
+
+    "add         %1, #4*2                      \n"  // src   += 4 * 2
+    "add         %3, %3, %4, lsl #2            \n"  // dst_a += 4 * dst_stride_a
+    "add         %5, %5, %6, lsl #2            \n"  // dst_b += 4 * dst_stride_b
+    "subs        %7,  #4                       \n"  // w     -= 4
+    "beq         4f                            \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %7, #2                        \n"
+    "blt         3f                            \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[0], d2[0]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[0], d3[0]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[1], d2[1]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[1], d3[1]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[2], d2[2]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[2], d3[2]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[3], d2[3]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[3], d3[3]}, [%0]          \n"
+
+    "vtrn.8      d0, d1                        \n"
+    "vtrn.8      d2, d3                        \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(0)
+    "vst1.64     {d0}, [%0], %4                \n"
+    MEMACCESS(0)
+    "vst1.64     {d2}, [%0]                    \n"
+
+    "mov         %0, %5                        \n"
+
+    MEMACCESS(0)
+    "vst1.64     {d1}, [%0], %6                \n"
+    MEMACCESS(0)
+    "vst1.64     {d3}, [%0]                    \n"
+
+    "add         %1, #2*2                      \n"  // src   += 2 * 2
+    "add         %3, %3, %4, lsl #1            \n"  // dst_a += 2 * dst_stride_a
+    "add         %5, %5, %6, lsl #1            \n"  // dst_b += 2 * dst_stride_b
+    "subs        %7,  #2                       \n"  // w     -= 2
+    "beq         4f                            \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[0], d1[0]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[1], d1[1]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[2], d1[2]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[3], d1[3]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[4], d1[4]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[5], d1[5]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[6], d1[6]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[7], d1[7]}, [%1]          \n"
+
+    MEMACCESS(3)
+    "vst1.64     {d0}, [%3]                    \n"
+    MEMACCESS(5)
+    "vst1.64     {d1}, [%5]                    \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),            // %0
+      "+r"(src),                 // %1
+      "+r"(src_stride),          // %2
+      "+r"(dst_a),               // %3
+      "+r"(dst_stride_a),        // %4
+      "+r"(dst_b),               // %5
+      "+r"(dst_stride_b),        // %6
+      "+r"(width)                // %7
+    : "r"(&kVTbl4x4TransposeDi)  // %8
+    : "memory", "cc",
+      "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  );
+}
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/rotate_neon64.cc b/libvpx/libvpx/third_party/libyuv/source/rotate_neon64.cc
new file mode 100644
index 0000000..f52c082
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/rotate_neon64.cc
@@ -0,0 +1,543 @@
+/*
+ *  Copyright 2014 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon armv8 64 bit.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+static uvec8 kVTbl4x4Transpose =
+  { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 };
+
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride, int width) {
+  const uint8* src_temp = NULL;
+  int64 width64 = (int64) width;  // Work around clang 3.4 warning.
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub         %3, %3, #8                      \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    "1:                                          \n"
+      "mov         %0, %1                        \n"
+
+      MEMACCESS(0)
+      "ld1        {v0.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v1.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v2.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v3.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v4.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v5.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v6.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v7.8b}, [%0]                  \n"
+
+      "trn2     v16.8b, v0.8b, v1.8b             \n"
+      "trn1     v17.8b, v0.8b, v1.8b             \n"
+      "trn2     v18.8b, v2.8b, v3.8b             \n"
+      "trn1     v19.8b, v2.8b, v3.8b             \n"
+      "trn2     v20.8b, v4.8b, v5.8b             \n"
+      "trn1     v21.8b, v4.8b, v5.8b             \n"
+      "trn2     v22.8b, v6.8b, v7.8b             \n"
+      "trn1     v23.8b, v6.8b, v7.8b             \n"
+
+      "trn2     v3.4h, v17.4h, v19.4h            \n"
+      "trn1     v1.4h, v17.4h, v19.4h            \n"
+      "trn2     v2.4h, v16.4h, v18.4h            \n"
+      "trn1     v0.4h, v16.4h, v18.4h            \n"
+      "trn2     v7.4h, v21.4h, v23.4h            \n"
+      "trn1     v5.4h, v21.4h, v23.4h            \n"
+      "trn2     v6.4h, v20.4h, v22.4h            \n"
+      "trn1     v4.4h, v20.4h, v22.4h            \n"
+
+      "trn2     v21.2s, v1.2s, v5.2s             \n"
+      "trn1     v17.2s, v1.2s, v5.2s             \n"
+      "trn2     v20.2s, v0.2s, v4.2s             \n"
+      "trn1     v16.2s, v0.2s, v4.2s             \n"
+      "trn2     v23.2s, v3.2s, v7.2s             \n"
+      "trn1     v19.2s, v3.2s, v7.2s             \n"
+      "trn2     v22.2s, v2.2s, v6.2s             \n"
+      "trn1     v18.2s, v2.2s, v6.2s             \n"
+
+      "mov         %0, %2                        \n"
+
+    MEMACCESS(0)
+      "st1      {v17.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v16.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v19.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v18.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v21.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v20.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v23.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v22.8b}, [%0]                   \n"
+
+      "add         %1, %1, #8                    \n"  // src += 8
+      "add         %2, %2, %6, lsl #3            \n"  // dst += 8 * dst_stride
+      "subs        %3, %3, #8                    \n"  // w   -= 8
+      "b.ge        1b                            \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %3, %3, #8                      \n"
+    "b.eq        4f                              \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %3, #2                          \n"
+    "b.lt        3f                              \n"
+
+    "cmp         %3, #4                          \n"
+    "b.lt        2f                              \n"
+
+    // 4x8 block
+    "mov         %0, %1                          \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[3], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[3], [%0]                     \n"
+
+    "mov         %0, %2                          \n"
+
+    MEMACCESS(4)
+    "ld1      {v2.16b}, [%4]                     \n"
+
+    "tbl      v3.16b, {v0.16b}, v2.16b           \n"
+    "tbl      v0.16b, {v1.16b}, v2.16b           \n"
+
+    // TODO(frkoenig): Rework shuffle above to
+    // write out with 4 instead of 8 writes.
+    MEMACCESS(0)
+    "st1 {v3.s}[0], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v3.s}[1], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v3.s}[2], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v3.s}[3], [%0]                         \n"
+
+    "add         %0, %2, #4                      \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[0], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[1], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[2], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[3], [%0]                         \n"
+
+    "add         %1, %1, #4                      \n"  // src += 4
+    "add         %2, %2, %6, lsl #2              \n"  // dst += 4 * dst_stride
+    "subs        %3, %3, #4                      \n"  // w   -= 4
+    "b.eq        4f                              \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %3, #2                          \n"
+    "b.lt        3f                              \n"
+
+    // 2x8 block
+    "2:                                          \n"
+    "mov         %0, %1                          \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[3], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[3], [%0]                     \n"
+
+    "trn2    v2.8b, v0.8b, v1.8b                 \n"
+    "trn1    v3.8b, v0.8b, v1.8b                 \n"
+
+    "mov         %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1     {v3.8b}, [%0], %6                   \n"
+    MEMACCESS(0)
+    "st1     {v2.8b}, [%0]                       \n"
+
+    "add         %1, %1, #2                      \n"  // src += 2
+    "add         %2, %2, %6, lsl #1              \n"  // dst += 2 * dst_stride
+    "subs        %3, %3,  #2                     \n"  // w   -= 2
+    "b.eq        4f                              \n"
+
+    // 1x8 block
+    "3:                                          \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[0], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[1], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[2], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[3], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[4], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[5], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[6], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[7], [%1]                 \n"
+
+    MEMACCESS(2)
+    "st1         {v0.8b}, [%2]                   \n"
+
+    "4:                                          \n"
+
+    : "+r"(src_temp),                             // %0
+      "+r"(src),                                  // %1
+      "+r"(dst),                                  // %2
+      "+r"(width64)                               // %3
+    : "r"(&kVTbl4x4Transpose),                    // %4
+      "r"(static_cast<ptrdiff_t>(src_stride)),    // %5
+      "r"(static_cast<ptrdiff_t>(dst_stride))     // %6
+    : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+      "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+  );
+}
+
+static uint8 kVTbl4x4TransposeDi[32] =
+  { 0,  16, 32, 48,  2, 18, 34, 50,  4, 20, 36, 52,  6, 22, 38, 54,
+    1,  17, 33, 49,  3, 19, 35, 51,  5, 21, 37, 53,  7, 23, 39, 55};
+
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) {
+  const uint8* src_temp = NULL;
+  int64 width64 = (int64) width;  // Work around clang 3.4 warning.
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub       %4, %4, #8                      \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    "1:                                        \n"
+    "mov       %0, %1                          \n"
+
+    MEMACCESS(0)
+    "ld1       {v0.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v1.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v2.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v3.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v4.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v5.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v6.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v7.16b}, [%0]                  \n"
+
+    "trn1      v16.16b, v0.16b, v1.16b         \n"
+    "trn2      v17.16b, v0.16b, v1.16b         \n"
+    "trn1      v18.16b, v2.16b, v3.16b         \n"
+    "trn2      v19.16b, v2.16b, v3.16b         \n"
+    "trn1      v20.16b, v4.16b, v5.16b         \n"
+    "trn2      v21.16b, v4.16b, v5.16b         \n"
+    "trn1      v22.16b, v6.16b, v7.16b         \n"
+    "trn2      v23.16b, v6.16b, v7.16b         \n"
+
+    "trn1      v0.8h, v16.8h, v18.8h           \n"
+    "trn2      v1.8h, v16.8h, v18.8h           \n"
+    "trn1      v2.8h, v20.8h, v22.8h           \n"
+    "trn2      v3.8h, v20.8h, v22.8h           \n"
+    "trn1      v4.8h, v17.8h, v19.8h           \n"
+    "trn2      v5.8h, v17.8h, v19.8h           \n"
+    "trn1      v6.8h, v21.8h, v23.8h           \n"
+    "trn2      v7.8h, v21.8h, v23.8h           \n"
+
+    "trn1      v16.4s, v0.4s, v2.4s            \n"
+    "trn2      v17.4s, v0.4s, v2.4s            \n"
+    "trn1      v18.4s, v1.4s, v3.4s            \n"
+    "trn2      v19.4s, v1.4s, v3.4s            \n"
+    "trn1      v20.4s, v4.4s, v6.4s            \n"
+    "trn2      v21.4s, v4.4s, v6.4s            \n"
+    "trn1      v22.4s, v5.4s, v7.4s            \n"
+    "trn2      v23.4s, v5.4s, v7.4s            \n"
+
+    "mov       %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1       {v16.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v17.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v19.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v16.d}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.d}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v17.d}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v19.d}[1], [%0]                \n"
+
+    "mov       %0, %3                          \n"
+
+    MEMACCESS(0)
+    "st1       {v20.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v22.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v21.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v23.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v20.d}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v22.d}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v21.d}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v23.d}[1], [%0]                \n"
+
+    "add       %1, %1, #16                     \n"  // src   += 8*2
+    "add       %2, %2, %6, lsl #3              \n"  // dst_a += 8 * dst_stride_a
+    "add       %3, %3, %7, lsl #3              \n"  // dst_b += 8 * dst_stride_b
+    "subs      %4, %4,  #8                     \n"  // w     -= 8
+    "b.ge      1b                              \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds      %4, %4, #8                      \n"
+    "b.eq      4f                              \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp       %4, #2                          \n"
+    "b.lt      3f                              \n"
+
+    "cmp       %4, #4                          \n"
+    "b.lt      2f                              \n"
+
+    // TODO(frkoenig): Clean this up
+    // 4x8 block
+    "mov       %0, %1                          \n"
+    MEMACCESS(0)
+    "ld1       {v0.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v1.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v2.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v3.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v4.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v5.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v6.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v7.8b}, [%0]                   \n"
+
+    MEMACCESS(8)
+    "ld1       {v30.16b}, [%8], #16            \n"
+    "ld1       {v31.16b}, [%8]                 \n"
+
+    "tbl       v16.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v30.16b  \n"
+    "tbl       v17.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v31.16b  \n"
+    "tbl       v18.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v30.16b  \n"
+    "tbl       v19.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v31.16b  \n"
+
+    "mov       %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1       {v16.s}[0],  [%0], %6           \n"
+    MEMACCESS(0)
+    "st1       {v16.s}[1],  [%0], %6           \n"
+    MEMACCESS(0)
+    "st1       {v16.s}[2],  [%0], %6           \n"
+    MEMACCESS(0)
+    "st1       {v16.s}[3],  [%0], %6           \n"
+
+    "add       %0, %2, #4                      \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[2], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[3], [%0]                \n"
+
+    "mov       %0, %3                          \n"
+
+    MEMACCESS(0)
+    "st1       {v17.s}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v17.s}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v17.s}[2], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v17.s}[3], [%0], %7            \n"
+
+    "add       %0, %3, #4                      \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[0],  [%0], %7           \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[1],  [%0], %7           \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[2],  [%0], %7           \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[3],  [%0]               \n"
+
+    "add       %1, %1, #8                      \n"  // src   += 4 * 2
+    "add       %2, %2, %6, lsl #2              \n"  // dst_a += 4 * dst_stride_a
+    "add       %3, %3, %7, lsl #2              \n"  // dst_b += 4 * dst_stride_b
+    "subs      %4,  %4,  #4                    \n"  // w     -= 4
+    "b.eq      4f                              \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp       %4, #2                          \n"
+    "b.lt      3f                              \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov       %0, %1                          \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[0], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[0], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[1], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[1], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[2], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[2], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[3], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[3], [%0]           \n"
+
+    "trn1      v4.8b, v0.8b, v2.8b             \n"
+    "trn2      v5.8b, v0.8b, v2.8b             \n"
+    "trn1      v6.8b, v1.8b, v3.8b             \n"
+    "trn2      v7.8b, v1.8b, v3.8b             \n"
+
+    "mov       %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1       {v4.d}[0], [%0], %6             \n"
+    MEMACCESS(0)
+    "st1       {v6.d}[0], [%0]                 \n"
+
+    "mov       %0, %3                          \n"
+
+    MEMACCESS(0)
+    "st1       {v5.d}[0], [%0], %7             \n"
+    MEMACCESS(0)
+    "st1       {v7.d}[0], [%0]                 \n"
+
+    "add       %1, %1, #4                      \n"  // src   += 2 * 2
+    "add       %2, %2, %6, lsl #1              \n"  // dst_a += 2 * dst_stride_a
+    "add       %3, %3, %7, lsl #1              \n"  // dst_b += 2 * dst_stride_b
+    "subs      %4,  %4,  #2                    \n"  // w     -= 2
+    "b.eq      4f                              \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[0], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[1], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[2], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[3], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[4], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[5], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[6], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[7], [%1]           \n"
+
+    MEMACCESS(2)
+    "st1       {v0.d}[0], [%2]                 \n"
+    MEMACCESS(3)
+    "st1       {v1.d}[0], [%3]                 \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),                             // %0
+      "+r"(src),                                  // %1
+      "+r"(dst_a),                                // %2
+      "+r"(dst_b),                                // %3
+      "+r"(width64)                               // %4
+    : "r"(static_cast<ptrdiff_t>(src_stride)),    // %5
+      "r"(static_cast<ptrdiff_t>(dst_stride_a)),  // %6
+      "r"(static_cast<ptrdiff_t>(dst_stride_b)),  // %7
+      "r"(&kVTbl4x4TransposeDi)                   // %8
+    : "memory", "cc",
+      "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+      "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
+      "v30", "v31"
+  );
+}
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/rotate_win.cc b/libvpx/libvpx/third_party/libyuv/source/rotate_win.cc
new file mode 100644
index 0000000..2760066
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/rotate_win.cc
@@ -0,0 +1,248 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
+    defined(_MSC_VER) && !defined(__clang__)
+
+__declspec(naked)
+void TransposeWx8_SSSE3(const uint8* src, int src_stride,
+                        uint8* dst, int dst_stride, int width) {
+  __asm {
+    push      edi
+    push      esi
+    push      ebp
+    mov       eax, [esp + 12 + 4]   // src
+    mov       edi, [esp + 12 + 8]   // src_stride
+    mov       edx, [esp + 12 + 12]  // dst
+    mov       esi, [esp + 12 + 16]  // dst_stride
+    mov       ecx, [esp + 12 + 20]  // width
+
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    align      4
+ convertloop:
+    movq      xmm0, qword ptr [eax]
+    lea       ebp, [eax + 8]
+    movq      xmm1, qword ptr [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    punpcklbw xmm0, xmm1
+    movq      xmm2, qword ptr [eax]
+    movdqa    xmm1, xmm0
+    palignr   xmm1, xmm1, 8
+    movq      xmm3, qword ptr [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    punpcklbw xmm2, xmm3
+    movdqa    xmm3, xmm2
+    movq      xmm4, qword ptr [eax]
+    palignr   xmm3, xmm3, 8
+    movq      xmm5, qword ptr [eax + edi]
+    punpcklbw xmm4, xmm5
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm5, xmm4
+    movq      xmm6, qword ptr [eax]
+    palignr   xmm5, xmm5, 8
+    movq      xmm7, qword ptr [eax + edi]
+    punpcklbw xmm6, xmm7
+    mov       eax, ebp
+    movdqa    xmm7, xmm6
+    palignr   xmm7, xmm7, 8
+    // Second round of bit swap.
+    punpcklwd xmm0, xmm2
+    punpcklwd xmm1, xmm3
+    movdqa    xmm2, xmm0
+    movdqa    xmm3, xmm1
+    palignr   xmm2, xmm2, 8
+    palignr   xmm3, xmm3, 8
+    punpcklwd xmm4, xmm6
+    punpcklwd xmm5, xmm7
+    movdqa    xmm6, xmm4
+    movdqa    xmm7, xmm5
+    palignr   xmm6, xmm6, 8
+    palignr   xmm7, xmm7, 8
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    punpckldq xmm0, xmm4
+    movq      qword ptr [edx], xmm0
+    movdqa    xmm4, xmm0
+    palignr   xmm4, xmm4, 8
+    movq      qword ptr [edx + esi], xmm4
+    lea       edx, [edx + 2 * esi]
+    punpckldq xmm2, xmm6
+    movdqa    xmm6, xmm2
+    palignr   xmm6, xmm6, 8
+    movq      qword ptr [edx], xmm2
+    punpckldq xmm1, xmm5
+    movq      qword ptr [edx + esi], xmm6
+    lea       edx, [edx + 2 * esi]
+    movdqa    xmm5, xmm1
+    movq      qword ptr [edx], xmm1
+    palignr   xmm5, xmm5, 8
+    punpckldq xmm3, xmm7
+    movq      qword ptr [edx + esi], xmm5
+    lea       edx, [edx + 2 * esi]
+    movq      qword ptr [edx], xmm3
+    movdqa    xmm7, xmm3
+    palignr   xmm7, xmm7, 8
+    sub       ecx, 8
+    movq      qword ptr [edx + esi], xmm7
+    lea       edx, [edx + 2 * esi]
+    jg        convertloop
+
+    pop       ebp
+    pop       esi
+    pop       edi
+    ret
+  }
+}
+
+__declspec(naked)
+void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int w) {
+  __asm {
+    push      ebx
+    push      esi
+    push      edi
+    push      ebp
+    mov       eax, [esp + 16 + 4]   // src
+    mov       edi, [esp + 16 + 8]   // src_stride
+    mov       edx, [esp + 16 + 12]  // dst_a
+    mov       esi, [esp + 16 + 16]  // dst_stride_a
+    mov       ebx, [esp + 16 + 20]  // dst_b
+    mov       ebp, [esp + 16 + 24]  // dst_stride_b
+    mov       ecx, esp
+    sub       esp, 4 + 16
+    and       esp, ~15
+    mov       [esp + 16], ecx
+    mov       ecx, [ecx + 16 + 28]  // w
+
+    align      4
+ convertloop:
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm0  // use xmm7 as temp register.
+    punpcklbw xmm0, xmm1
+    punpckhbw xmm7, xmm1
+    movdqa    xmm1, xmm7
+    movdqu    xmm2, [eax]
+    movdqu    xmm3, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm2
+    punpcklbw xmm2, xmm3
+    punpckhbw xmm7, xmm3
+    movdqa    xmm3, xmm7
+    movdqu    xmm4, [eax]
+    movdqu    xmm5, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm4
+    punpcklbw xmm4, xmm5
+    punpckhbw xmm7, xmm5
+    movdqa    xmm5, xmm7
+    movdqu    xmm6, [eax]
+    movdqu    xmm7, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqu    [esp], xmm5  // backup xmm5
+    neg       edi
+    movdqa    xmm5, xmm6   // use xmm5 as temp register.
+    punpcklbw xmm6, xmm7
+    punpckhbw xmm5, xmm7
+    movdqa    xmm7, xmm5
+    lea       eax, [eax + 8 * edi + 16]
+    neg       edi
+    // Second round of bit swap.
+    movdqa    xmm5, xmm0
+    punpcklwd xmm0, xmm2
+    punpckhwd xmm5, xmm2
+    movdqa    xmm2, xmm5
+    movdqa    xmm5, xmm1
+    punpcklwd xmm1, xmm3
+    punpckhwd xmm5, xmm3
+    movdqa    xmm3, xmm5
+    movdqa    xmm5, xmm4
+    punpcklwd xmm4, xmm6
+    punpckhwd xmm5, xmm6
+    movdqa    xmm6, xmm5
+    movdqu    xmm5, [esp]  // restore xmm5
+    movdqu    [esp], xmm6  // backup xmm6
+    movdqa    xmm6, xmm5    // use xmm6 as temp register.
+    punpcklwd xmm5, xmm7
+    punpckhwd xmm6, xmm7
+    movdqa    xmm7, xmm6
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    movdqa    xmm6, xmm0
+    punpckldq xmm0, xmm4
+    punpckhdq xmm6, xmm4
+    movdqa    xmm4, xmm6
+    movdqu    xmm6, [esp]  // restore xmm6
+    movlpd    qword ptr [edx], xmm0
+    movhpd    qword ptr [ebx], xmm0
+    movlpd    qword ptr [edx + esi], xmm4
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm4
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm2   // use xmm0 as the temp register.
+    punpckldq xmm2, xmm6
+    movlpd    qword ptr [edx], xmm2
+    movhpd    qword ptr [ebx], xmm2
+    punpckhdq xmm0, xmm6
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm1   // use xmm0 as the temp register.
+    punpckldq xmm1, xmm5
+    movlpd    qword ptr [edx], xmm1
+    movhpd    qword ptr [ebx], xmm1
+    punpckhdq xmm0, xmm5
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm3   // use xmm0 as the temp register.
+    punpckldq xmm3, xmm7
+    movlpd    qword ptr [edx], xmm3
+    movhpd    qword ptr [ebx], xmm3
+    punpckhdq xmm0, xmm7
+    sub       ecx, 8
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    jg        convertloop
+
+    mov       esp, [esp + 16]
+    pop       ebp
+    pop       edi
+    pop       esi
+    pop       ebx
+    ret
+  }
+}
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/row_any.cc b/libvpx/libvpx/third_party/libyuv/source/row_any.cc
new file mode 100644
index 0000000..1cb1f6b
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/row_any.cc
@@ -0,0 +1,680 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include <string.h>  // For memset.
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Subsampled source needs to be increase by 1 of not even.
+#define SS(width, shift) (((width) + (1 << (shift)) - 1) >> (shift))
+
+// Any 3 planes to 1.
+#define ANY31(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)                 \
+    void NAMEANY(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf,   \
+                 uint8* dst_ptr, int width) {                                  \
+      SIMD_ALIGNED(uint8 temp[64 * 4]);                                        \
+      memset(temp, 0, 64 * 3);  /* for YUY2 and msan */                        \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(y_buf, u_buf, v_buf, dst_ptr, n);                             \
+      }                                                                        \
+      memcpy(temp, y_buf + n, r);                                              \
+      memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));               \
+      memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));              \
+      ANY_SIMD(temp, temp + 64, temp + 128, temp + 192, MASK + 1);             \
+      memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 192,                      \
+             SS(r, DUVSHIFT) * BPP);                                           \
+    }
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+ANY31(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_SSSE3, 1, 0, 4, 7)
+#endif
+#ifdef HAS_I444TOARGBROW_SSSE3
+ANY31(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_SSSE3, 0, 0, 4, 7)
+ANY31(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_SSSE3, 2, 0, 4, 7)
+ANY31(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_SSSE3, 1, 0, 4, 7)
+ANY31(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_SSSE3, 1, 0, 4, 7)
+ANY31(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_SSSE3, 1, 0, 4, 7)
+ANY31(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, 1, 0, 2, 7)
+ANY31(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, 1, 0, 2, 7)
+ANY31(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, 1, 0, 2, 7)
+ANY31(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, 1, 0, 3, 7)
+ANY31(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, 1, 0, 3, 7)
+ANY31(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, 1, 1, 4, 15)
+ANY31(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, 1, 1, 4, 15)
+#endif  // HAS_I444TOARGBROW_SSSE3
+#ifdef HAS_I422TORGB24ROW_AVX2
+ANY31(I422ToRGB24Row_Any_AVX2, I422ToRGB24Row_AVX2, 1, 0, 3, 15)
+#endif
+#ifdef HAS_I422TORAWROW_AVX2
+ANY31(I422ToRAWRow_Any_AVX2, I422ToRAWRow_AVX2, 1, 0, 3, 15)
+#endif
+#ifdef HAS_J422TOARGBROW_SSSE3
+ANY31(J422ToARGBRow_Any_SSSE3, J422ToARGBRow_SSSE3, 1, 0, 4, 7)
+#endif
+#ifdef HAS_J422TOARGBROW_AVX2
+ANY31(J422ToARGBRow_Any_AVX2, J422ToARGBRow_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I422TOARGBROW_AVX2
+ANY31(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I422TOBGRAROW_AVX2
+ANY31(I422ToBGRARow_Any_AVX2, I422ToBGRARow_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I422TORGBAROW_AVX2
+ANY31(I422ToRGBARow_Any_AVX2, I422ToRGBARow_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I422TOABGRROW_AVX2
+ANY31(I422ToABGRRow_Any_AVX2, I422ToABGRRow_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I444TOARGBROW_AVX2
+ANY31(I444ToARGBRow_Any_AVX2, I444ToARGBRow_AVX2, 0, 0, 4, 15)
+#endif
+#ifdef HAS_I411TOARGBROW_AVX2
+ANY31(I411ToARGBRow_Any_AVX2, I411ToARGBRow_AVX2, 2, 0, 4, 15)
+#endif
+#ifdef HAS_I422TOARGB4444ROW_AVX2
+ANY31(I422ToARGB4444Row_Any_AVX2, I422ToARGB4444Row_AVX2, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TOARGB1555ROW_AVX2
+ANY31(I422ToARGB1555Row_Any_AVX2, I422ToARGB1555Row_AVX2, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TORGB565ROW_AVX2
+ANY31(I422ToRGB565Row_Any_AVX2, I422ToRGB565Row_AVX2, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TOARGBROW_NEON
+ANY31(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, 0, 0, 4, 7)
+ANY31(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, 1, 0, 4, 7)
+ANY31(I411ToARGBRow_Any_NEON, I411ToARGBRow_NEON, 2, 0, 4, 7)
+ANY31(I422ToBGRARow_Any_NEON, I422ToBGRARow_NEON, 1, 0, 4, 7)
+ANY31(I422ToABGRRow_Any_NEON, I422ToABGRRow_NEON, 1, 0, 4, 7)
+ANY31(I422ToRGBARow_Any_NEON, I422ToRGBARow_NEON, 1, 0, 4, 7)
+ANY31(I422ToRGB24Row_Any_NEON, I422ToRGB24Row_NEON, 1, 0, 3, 7)
+ANY31(I422ToRAWRow_Any_NEON, I422ToRAWRow_NEON, 1, 0, 3, 7)
+ANY31(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, 1, 0, 2, 7)
+ANY31(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, 1, 0, 2, 7)
+ANY31(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TOYUY2ROW_NEON
+ANY31(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, 1, 1, 4, 15)
+#endif
+#ifdef HAS_I422TOUYVYROW_NEON
+ANY31(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, 1, 1, 4, 15)
+#endif
+#undef ANY31
+
+// Any 2 planes to 1.
+#define ANY21(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, SBPP2, BPP, MASK)              \
+    void NAMEANY(const uint8* y_buf, const uint8* uv_buf,                      \
+                 uint8* dst_ptr, int width) {                                  \
+      SIMD_ALIGNED(uint8 temp[64 * 3]);                                        \
+      memset(temp, 0, 64 * 2);  /* for msan */                                 \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(y_buf, uv_buf, dst_ptr, n);                                   \
+      }                                                                        \
+      memcpy(temp, y_buf + n * SBPP, r * SBPP);                                \
+      memcpy(temp + 64, uv_buf + (n >> UVSHIFT) * SBPP2,                       \
+             SS(r, UVSHIFT) * SBPP2);                                          \
+      ANY_SIMD(temp, temp + 64, temp + 128, MASK + 1);                         \
+      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
+    }
+
+// Biplanar to RGB.
+#ifdef HAS_NV12TOARGBROW_SSSE3
+ANY21(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_SSSE3, 1, 1, 2, 4, 7)
+ANY21(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_SSSE3, 1, 1, 2, 4, 7)
+#endif
+#ifdef HAS_NV12TOARGBROW_AVX2
+ANY21(NV12ToARGBRow_Any_AVX2, NV12ToARGBRow_AVX2, 1, 1, 2, 4, 15)
+ANY21(NV21ToARGBRow_Any_AVX2, NV21ToARGBRow_AVX2, 1, 1, 2, 4, 15)
+#endif
+#ifdef HAS_NV12TOARGBROW_NEON
+ANY21(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, 1, 1, 2, 4, 7)
+ANY21(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, 1, 1, 2, 4, 7)
+#endif
+#ifdef HAS_NV12TORGB565ROW_SSSE3
+ANY21(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, 1, 1, 2, 2, 7)
+ANY21(NV21ToRGB565Row_Any_SSSE3, NV21ToRGB565Row_SSSE3, 1, 1, 2, 2, 7)
+#endif
+#ifdef HAS_NV12TORGB565ROW_AVX2
+ANY21(NV12ToRGB565Row_Any_AVX2, NV12ToRGB565Row_AVX2, 1, 1, 2, 2, 15)
+ANY21(NV21ToRGB565Row_Any_AVX2, NV21ToRGB565Row_AVX2, 1, 1, 2, 2, 15)
+#endif
+#ifdef HAS_NV12TORGB565ROW_NEON
+ANY21(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, 1, 1, 2, 2, 7)
+ANY21(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, 1, 1, 2, 2, 7)
+#endif
+
+// Merge functions.
+#ifdef HAS_MERGEUVROW_SSE2
+ANY21(MergeUVRow_Any_SSE2, MergeUVRow_SSE2, 0, 1, 1, 2, 15)
+#endif
+#ifdef HAS_MERGEUVROW_AVX2
+ANY21(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, 0, 1, 1, 2, 31)
+#endif
+#ifdef HAS_MERGEUVROW_NEON
+ANY21(MergeUVRow_Any_NEON, MergeUVRow_NEON, 0, 1, 1, 2, 15)
+#endif
+
+// Math functions.
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+ANY21(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, 0, 4, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBADDROW_SSE2
+ANY21(ARGBAddRow_Any_SSE2, ARGBAddRow_SSE2, 0, 4, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+ANY21(ARGBSubtractRow_Any_SSE2, ARGBSubtractRow_SSE2, 0, 4, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+ANY21(ARGBMultiplyRow_Any_AVX2, ARGBMultiplyRow_AVX2, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBADDROW_AVX2
+ANY21(ARGBAddRow_Any_AVX2, ARGBAddRow_AVX2, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+ANY21(ARGBSubtractRow_Any_AVX2, ARGBSubtractRow_AVX2, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+ANY21(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBADDROW_NEON
+ANY21(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_NEON
+ANY21(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_SOBELROW_SSE2
+ANY21(SobelRow_Any_SSE2, SobelRow_SSE2, 0, 1, 1, 4, 15)
+#endif
+#ifdef HAS_SOBELROW_NEON
+ANY21(SobelRow_Any_NEON, SobelRow_NEON, 0, 1, 1, 4, 7)
+#endif
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+ANY21(SobelToPlaneRow_Any_SSE2, SobelToPlaneRow_SSE2, 0, 1, 1, 1, 15)
+#endif
+#ifdef HAS_SOBELTOPLANEROW_NEON
+ANY21(SobelToPlaneRow_Any_NEON, SobelToPlaneRow_NEON, 0, 1, 1, 1, 15)
+#endif
+#ifdef HAS_SOBELXYROW_SSE2
+ANY21(SobelXYRow_Any_SSE2, SobelXYRow_SSE2, 0, 1, 1, 4, 15)
+#endif
+#ifdef HAS_SOBELXYROW_NEON
+ANY21(SobelXYRow_Any_NEON, SobelXYRow_NEON, 0, 1, 1, 4, 7)
+#endif
+#undef ANY21
+
+// Any 1 to 1.
+#define ANY11(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK)                     \
+    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr, int width) {            \
+      SIMD_ALIGNED(uint8 temp[128 * 2]);                                       \
+      memset(temp, 0, 128);  /* for YUY2 and msan */                           \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(src_ptr, dst_ptr, n);                                         \
+      }                                                                        \
+      memcpy(temp, src_ptr + (n >> UVSHIFT) * SBPP, SS(r, UVSHIFT) * SBPP);    \
+      ANY_SIMD(temp, temp + 128, MASK + 1);                                    \
+      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
+    }
+
+#ifdef HAS_COPYROW_AVX
+ANY11(CopyRow_Any_AVX, CopyRow_AVX, 0, 1, 1, 63)
+#endif
+#ifdef HAS_COPYROW_SSE2
+ANY11(CopyRow_Any_SSE2, CopyRow_SSE2, 0, 1, 1, 31)
+#endif
+#ifdef HAS_COPYROW_NEON
+ANY11(CopyRow_Any_NEON, CopyRow_NEON, 0, 1, 1, 31)
+#endif
+#if defined(HAS_ARGBTORGB24ROW_SSSE3)
+ANY11(ARGBToRGB24Row_Any_SSSE3, ARGBToRGB24Row_SSSE3, 0, 4, 3, 15)
+ANY11(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, 0, 4, 3, 15)
+ANY11(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, 0, 4, 2, 3)
+ANY11(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, 0, 4, 2, 3)
+ANY11(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, 0, 4, 2, 3)
+#endif
+#if defined(HAS_ARGBTOARGB4444ROW_AVX2)
+ANY11(ARGBToRGB565Row_Any_AVX2, ARGBToRGB565Row_AVX2, 0, 4, 2, 7)
+ANY11(ARGBToARGB1555Row_Any_AVX2, ARGBToARGB1555Row_AVX2, 0, 4, 2, 7)
+ANY11(ARGBToARGB4444Row_Any_AVX2, ARGBToARGB4444Row_AVX2, 0, 4, 2, 7)
+#endif
+#if defined(HAS_J400TOARGBROW_SSE2)
+ANY11(J400ToARGBRow_Any_SSE2, J400ToARGBRow_SSE2, 0, 1, 4, 7)
+#endif
+#if defined(HAS_J400TOARGBROW_AVX2)
+ANY11(J400ToARGBRow_Any_AVX2, J400ToARGBRow_AVX2, 0, 1, 4, 15)
+#endif
+#if defined(HAS_I400TOARGBROW_SSE2)
+ANY11(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, 0, 1, 4, 7)
+#endif
+#if defined(HAS_I400TOARGBROW_AVX2)
+ANY11(I400ToARGBRow_Any_AVX2, I400ToARGBRow_AVX2, 0, 1, 4, 15)
+#endif
+#if defined(HAS_YUY2TOARGBROW_SSSE3)
+ANY11(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_SSSE3, 1, 4, 4, 15)
+ANY11(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_SSSE3, 1, 4, 4, 15)
+ANY11(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, 0, 3, 4, 15)
+ANY11(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, 0, 3, 4, 15)
+ANY11(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, 0, 2, 4, 7)
+ANY11(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, 0, 2, 4, 7)
+ANY11(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, 0, 2, 4, 7)
+#endif
+#if defined(HAS_RGB565TOARGBROW_AVX2)
+ANY11(RGB565ToARGBRow_Any_AVX2, RGB565ToARGBRow_AVX2, 0, 2, 4, 15)
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_AVX2)
+ANY11(ARGB1555ToARGBRow_Any_AVX2, ARGB1555ToARGBRow_AVX2, 0, 2, 4, 15)
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_AVX2)
+ANY11(ARGB4444ToARGBRow_Any_AVX2, ARGB4444ToARGBRow_AVX2, 0, 2, 4, 15)
+#endif
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+ANY11(YUY2ToARGBRow_Any_AVX2, YUY2ToARGBRow_AVX2, 1, 4, 4, 31)
+ANY11(UYVYToARGBRow_Any_AVX2, UYVYToARGBRow_AVX2, 1, 4, 4, 31)
+#endif
+#if defined(HAS_ARGBTORGB24ROW_NEON)
+ANY11(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, 0, 4, 3, 7)
+ANY11(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, 0, 4, 3, 7)
+ANY11(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, 0, 4, 2, 7)
+ANY11(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, 0, 4, 2, 7)
+ANY11(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, 0, 4, 2, 7)
+ANY11(J400ToARGBRow_Any_NEON, J400ToARGBRow_NEON, 0, 1, 4, 7)
+ANY11(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, 0, 1, 4, 7)
+ANY11(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, 1, 4, 4, 7)
+ANY11(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, 1, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBTOYROW_AVX2
+ANY11(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 0, 4, 1, 31)
+#endif
+#ifdef HAS_ARGBTOYJROW_AVX2
+ANY11(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 0, 4, 1, 31)
+#endif
+#ifdef HAS_UYVYTOYROW_AVX2
+ANY11(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 0, 2, 1, 31)
+#endif
+#ifdef HAS_YUY2TOYROW_AVX2
+ANY11(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 1, 4, 1, 31)
+#endif
+#ifdef HAS_ARGBTOYROW_SSSE3
+ANY11(ARGBToYRow_Any_SSSE3, ARGBToYRow_SSSE3, 0, 4, 1, 15)
+#endif
+#ifdef HAS_BGRATOYROW_SSSE3
+ANY11(BGRAToYRow_Any_SSSE3, BGRAToYRow_SSSE3, 0, 4, 1, 15)
+ANY11(ABGRToYRow_Any_SSSE3, ABGRToYRow_SSSE3, 0, 4, 1, 15)
+ANY11(RGBAToYRow_Any_SSSE3, RGBAToYRow_SSSE3, 0, 4, 1, 15)
+ANY11(YUY2ToYRow_Any_SSE2, YUY2ToYRow_SSE2, 1, 4, 1, 15)
+ANY11(UYVYToYRow_Any_SSE2, UYVYToYRow_SSE2, 1, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBTOYJROW_SSSE3
+ANY11(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_SSSE3, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBTOYROW_NEON
+ANY11(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 0, 4, 1, 7)
+#endif
+#ifdef HAS_ARGBTOYJROW_NEON
+ANY11(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 0, 4, 1, 7)
+#endif
+#ifdef HAS_BGRATOYROW_NEON
+ANY11(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 0, 4, 1, 7)
+#endif
+#ifdef HAS_ABGRTOYROW_NEON
+ANY11(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 0, 4, 1, 7)
+#endif
+#ifdef HAS_RGBATOYROW_NEON
+ANY11(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 0, 4, 1, 7)
+#endif
+#ifdef HAS_RGB24TOYROW_NEON
+ANY11(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 0, 3, 1, 7)
+#endif
+#ifdef HAS_RAWTOYROW_NEON
+ANY11(RAWToYRow_Any_NEON, RAWToYRow_NEON, 0, 3, 1, 7)
+#endif
+#ifdef HAS_RGB565TOYROW_NEON
+ANY11(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 0, 2, 1, 7)
+#endif
+#ifdef HAS_ARGB1555TOYROW_NEON
+ANY11(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 0, 2, 1, 7)
+#endif
+#ifdef HAS_ARGB4444TOYROW_NEON
+ANY11(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 0, 2, 1, 7)
+#endif
+#ifdef HAS_YUY2TOYROW_NEON
+ANY11(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 1, 4, 1, 15)
+#endif
+#ifdef HAS_UYVYTOYROW_NEON
+ANY11(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 0, 2, 1, 15)
+#endif
+#ifdef HAS_RGB24TOARGBROW_NEON
+ANY11(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 0, 3, 4, 7)
+#endif
+#ifdef HAS_RAWTOARGBROW_NEON
+ANY11(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 0, 3, 4, 7)
+#endif
+#ifdef HAS_RGB565TOARGBROW_NEON
+ANY11(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 0, 2, 4, 7)
+#endif
+#ifdef HAS_ARGB1555TOARGBROW_NEON
+ANY11(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 0, 2, 4, 7)
+#endif
+#ifdef HAS_ARGB4444TOARGBROW_NEON
+ANY11(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 0, 2, 4, 7)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+ANY11(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, 0, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+ANY11(ARGBAttenuateRow_Any_SSE2, ARGBAttenuateRow_SSE2, 0, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+ANY11(ARGBUnattenuateRow_Any_SSE2, ARGBUnattenuateRow_SSE2, 0, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+ANY11(ARGBAttenuateRow_Any_AVX2, ARGBAttenuateRow_AVX2, 0, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+ANY11(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, 0, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_NEON
+ANY11(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, 0, 4, 4, 7)
+#endif
+#undef ANY11
+
+// Any 1 to 1 with parameter.
+#define ANY11P(NAMEANY, ANY_SIMD, T, SBPP, BPP, MASK)                          \
+    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr,                         \
+                 T shuffler, int width) {                                      \
+      SIMD_ALIGNED(uint8 temp[64 * 2]);                                        \
+      memset(temp, 0, 64);  /* for msan */                                     \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(src_ptr, dst_ptr, shuffler, n);                               \
+      }                                                                        \
+      memcpy(temp, src_ptr + n * SBPP, r * SBPP);                              \
+      ANY_SIMD(temp, temp + 64, shuffler, MASK + 1);                           \
+      memcpy(dst_ptr + n * BPP, temp + 64, r * BPP);                           \
+    }
+
+#if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
+ANY11P(ARGBToRGB565DitherRow_Any_SSE2, ARGBToRGB565DitherRow_SSE2,
+       const uint32, 4, 2, 3)
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
+ANY11P(ARGBToRGB565DitherRow_Any_AVX2, ARGBToRGB565DitherRow_AVX2,
+       const uint32, 4, 2, 7)
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_NEON)
+ANY11P(ARGBToRGB565DitherRow_Any_NEON, ARGBToRGB565DitherRow_NEON,
+       const uint32, 4, 2, 7)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_SSE2
+ANY11P(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2, const uint8*, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_SSSE3
+ANY11P(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_SSSE3, const uint8*, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+ANY11P(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2, const uint8*, 4, 4, 15)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_NEON
+ANY11P(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON, const uint8*, 4, 4, 3)
+#endif
+#undef ANY11P
+
+// Any 1 to 1 interpolate.  Takes 2 rows of source via stride.
+#define ANY11T(NAMEANY, ANY_SIMD, SBPP, BPP, MASK)                             \
+    void NAMEANY(uint8* dst_ptr, const uint8* src_ptr,                         \
+                 ptrdiff_t src_stride_ptr, int width,                          \
+                 int source_y_fraction) {                                      \
+      SIMD_ALIGNED(uint8 temp[64 * 3]);                                        \
+      memset(temp, 0, 64 * 2);  /* for msan */                                 \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(dst_ptr, src_ptr, src_stride_ptr, n, source_y_fraction);      \
+      }                                                                        \
+      memcpy(temp, src_ptr + n * SBPP, r * SBPP);                              \
+      memcpy(temp + 64, src_ptr + src_stride_ptr + n * SBPP, r * SBPP);        \
+      ANY_SIMD(temp + 128, temp, 64, MASK + 1, source_y_fraction);             \
+      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
+    }
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+ANY11T(InterpolateRow_Any_AVX2, InterpolateRow_AVX2, 1, 1, 31)
+#endif
+#ifdef HAS_INTERPOLATEROW_SSSE3
+ANY11T(InterpolateRow_Any_SSSE3, InterpolateRow_SSSE3, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_SSE2
+ANY11T(InterpolateRow_Any_SSE2, InterpolateRow_SSE2, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_NEON
+ANY11T(InterpolateRow_Any_NEON, InterpolateRow_NEON, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_MIPS_DSPR2
+ANY11T(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2, 1, 1, 3)
+#endif
+#undef ANY11T
+
+// Any 1 to 1 mirror.
+#define ANY11M(NAMEANY, ANY_SIMD, BPP, MASK)                                   \
+    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr, int width) {            \
+      SIMD_ALIGNED(uint8 temp[64 * 2]);                                        \
+      memset(temp, 0, 64);  /* for msan */                                     \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(src_ptr + r * BPP, dst_ptr, n);                               \
+      }                                                                        \
+      memcpy(temp, src_ptr, r * BPP);                                          \
+      ANY_SIMD(temp, temp + 64, MASK + 1);                                     \
+      memcpy(dst_ptr + n * BPP, temp + 64 + (MASK + 1 - r) * BPP, r * BPP);    \
+    }
+
+#ifdef HAS_MIRRORROW_AVX2
+ANY11M(MirrorRow_Any_AVX2, MirrorRow_AVX2, 1, 31)
+#endif
+#ifdef HAS_MIRRORROW_SSSE3
+ANY11M(MirrorRow_Any_SSSE3, MirrorRow_SSSE3, 1, 15)
+#endif
+#ifdef HAS_MIRRORROW_SSE2
+ANY11M(MirrorRow_Any_SSE2, MirrorRow_SSE2, 1, 15)
+#endif
+#ifdef HAS_MIRRORROW_NEON
+ANY11M(MirrorRow_Any_NEON, MirrorRow_NEON, 1, 15)
+#endif
+#ifdef HAS_ARGBMIRRORROW_AVX2
+ANY11M(ARGBMirrorRow_Any_AVX2, ARGBMirrorRow_AVX2, 4, 7)
+#endif
+#ifdef HAS_ARGBMIRRORROW_SSE2
+ANY11M(ARGBMirrorRow_Any_SSE2, ARGBMirrorRow_SSE2, 4, 3)
+#endif
+#ifdef HAS_ARGBMIRRORROW_NEON
+ANY11M(ARGBMirrorRow_Any_NEON, ARGBMirrorRow_NEON, 4, 3)
+#endif
+#undef ANY11M
+
+// Any 1 plane. (memset)
+#define ANY1(NAMEANY, ANY_SIMD, T, BPP, MASK)                                  \
+    void NAMEANY(uint8* dst_ptr, T v32, int width) {                           \
+      SIMD_ALIGNED(uint8 temp[64]);                                            \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(dst_ptr, v32, n);                                             \
+      }                                                                        \
+      ANY_SIMD(temp, v32, MASK + 1);                                           \
+      memcpy(dst_ptr + n * BPP, temp, r * BPP);                                \
+    }
+
+#ifdef HAS_SETROW_X86
+ANY1(SetRow_Any_X86, SetRow_X86, uint8, 1, 3)
+#endif
+#ifdef HAS_SETROW_NEON
+ANY1(SetRow_Any_NEON, SetRow_NEON, uint8, 1, 15)
+#endif
+#ifdef HAS_ARGBSETROW_NEON
+ANY1(ARGBSetRow_Any_NEON, ARGBSetRow_NEON, uint32, 4, 3)
+#endif
+#undef ANY1
+
+// Any 1 to 2.  Outputs UV planes.
+#define ANY12(NAMEANY, ANY_SIMD, UVSHIFT, BPP, DUVSHIFT, MASK)                 \
+    void NAMEANY(const uint8* src_ptr, uint8* dst_u, uint8* dst_v, int width) {\
+      SIMD_ALIGNED(uint8 temp[128 * 3]);                                       \
+      memset(temp, 0, 128);  /* for msan */                                    \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(src_ptr, dst_u, dst_v, n);                                    \
+      }                                                                        \
+      memcpy(temp, src_ptr  + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP);     \
+      if ((width & 1) && BPP == 4) {  /* repeat last 4 bytes for subsampler */ \
+        memcpy(temp + SS(r, UVSHIFT) * BPP,                                    \
+               temp + SS(r, UVSHIFT) * BPP - BPP, 4);                          \
+      }                                                                        \
+      ANY_SIMD(temp, temp + 128, temp + 256, MASK + 1);                        \
+      memcpy(dst_u + (n >> DUVSHIFT), temp + 128, SS(r, DUVSHIFT));            \
+      memcpy(dst_v + (n >> DUVSHIFT), temp + 256, SS(r, DUVSHIFT));            \
+    }
+
+#ifdef HAS_SPLITUVROW_SSE2
+ANY12(SplitUVRow_Any_SSE2, SplitUVRow_SSE2, 0, 2, 0, 15)
+#endif
+#ifdef HAS_SPLITUVROW_AVX2
+ANY12(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, 0, 2, 0, 31)
+#endif
+#ifdef HAS_SPLITUVROW_NEON
+ANY12(SplitUVRow_Any_NEON, SplitUVRow_NEON, 0, 2, 0, 15)
+#endif
+#ifdef HAS_SPLITUVROW_MIPS_DSPR2
+ANY12(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_MIPS_DSPR2, 0, 2, 0, 15)
+#endif
+#ifdef HAS_ARGBTOUV444ROW_SSSE3
+ANY12(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_SSSE3, 0, 4, 0, 15)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_AVX2
+ANY12(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2, 1, 4, 1, 31)
+ANY12(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2, 1, 4, 1, 31)
+#endif
+#ifdef HAS_ARGBTOUV422ROW_SSSE3
+ANY12(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_SSSE3, 0, 4, 1, 15)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_SSE2
+ANY12(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_SSE2, 1, 4, 1, 15)
+ANY12(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_SSE2, 1, 4, 1, 15)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_NEON
+ANY12(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON, 0, 4, 0, 7)
+ANY12(ARGBToUV422Row_Any_NEON, ARGBToUV422Row_NEON, 0, 4, 1, 15)
+ANY12(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON, 0, 4, 2, 31)
+ANY12(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON, 1, 4, 1, 15)
+ANY12(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON, 1, 4, 1, 15)
+#endif
+#undef ANY12
+
+// Any 1 to 2 with source stride (2 rows of source).  Outputs UV planes.
+// 128 byte row allows for 32 avx ARGB pixels.
+#define ANY12S(NAMEANY, ANY_SIMD, UVSHIFT, BPP, MASK)                          \
+    void NAMEANY(const uint8* src_ptr, int src_stride_ptr,                     \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      SIMD_ALIGNED(uint8 temp[128 * 4]);                                       \
+      memset(temp, 0, 128 * 2);  /* for msan */                                \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(src_ptr, src_stride_ptr, dst_u, dst_v, n);                    \
+      }                                                                        \
+      memcpy(temp, src_ptr  + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP);     \
+      memcpy(temp + 128, src_ptr  + src_stride_ptr + (n >> UVSHIFT) * BPP,     \
+             SS(r, UVSHIFT) * BPP);                                            \
+      if ((width & 1) && BPP == 4) {  /* repeat last 4 bytes for subsampler */ \
+        memcpy(temp + SS(r, UVSHIFT) * BPP,                                    \
+               temp + SS(r, UVSHIFT) * BPP - BPP, 4);                          \
+        memcpy(temp + 128 + SS(r, UVSHIFT) * BPP,                              \
+               temp + 128 + SS(r, UVSHIFT) * BPP - BPP, 4);                    \
+      }                                                                        \
+      ANY_SIMD(temp, 128, temp + 256, temp + 384, MASK + 1);                   \
+      memcpy(dst_u + (n >> 1), temp + 256, SS(r, 1));                          \
+      memcpy(dst_v + (n >> 1), temp + 384, SS(r, 1));                          \
+    }
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+ANY12S(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, 0, 4, 31)
+#endif
+#ifdef HAS_ARGBTOUVROW_SSSE3
+ANY12S(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_SSSE3, 0, 4, 15)
+ANY12S(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_SSSE3, 0, 4, 15)
+ANY12S(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_SSSE3, 0, 4, 15)
+ANY12S(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_SSSE3, 0, 4, 15)
+ANY12S(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_SSSE3, 0, 4, 15)
+#endif
+#ifdef HAS_YUY2TOUVROW_AVX2
+ANY12S(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, 1, 4, 31)
+ANY12S(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, 1, 4, 31)
+#endif
+#ifdef HAS_YUY2TOUVROW_SSE2
+ANY12S(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_SSE2, 1, 4, 15)
+ANY12S(UYVYToUVRow_Any_SSE2, UYVYToUVRow_SSE2, 1, 4, 15)
+#endif
+#ifdef HAS_ARGBTOUVROW_NEON
+ANY12S(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_ARGBTOUVJROW_NEON
+ANY12S(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_BGRATOUVROW_NEON
+ANY12S(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_ABGRTOUVROW_NEON
+ANY12S(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_RGBATOUVROW_NEON
+ANY12S(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_RGB24TOUVROW_NEON
+ANY12S(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, 0, 3, 15)
+#endif
+#ifdef HAS_RAWTOUVROW_NEON
+ANY12S(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, 0, 3, 15)
+#endif
+#ifdef HAS_RGB565TOUVROW_NEON
+ANY12S(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, 0, 2, 15)
+#endif
+#ifdef HAS_ARGB1555TOUVROW_NEON
+ANY12S(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, 0, 2, 15)
+#endif
+#ifdef HAS_ARGB4444TOUVROW_NEON
+ANY12S(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, 0, 2, 15)
+#endif
+#ifdef HAS_YUY2TOUVROW_NEON
+ANY12S(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, 1, 4, 15)
+#endif
+#ifdef HAS_UYVYTOUVROW_NEON
+ANY12S(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, 1, 4, 15)
+#endif
+#undef ANY12S
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/row_common.cc b/libvpx/libvpx/third_party/libyuv/source/row_common.cc
new file mode 100644
index 0000000..4987589
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/row_common.cc
@@ -0,0 +1,2576 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include <string.h>  // For memcpy and memset.
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// llvm x86 is poor at ternary operator, so use branchless min/max.
+
+#define USE_BRANCHLESS 1
+#if USE_BRANCHLESS
+static __inline int32 clamp0(int32 v) {
+  return ((-(v) >> 31) & (v));
+}
+
+static __inline int32 clamp255(int32 v) {
+  return (((255 - (v)) >> 31) | (v)) & 255;
+}
+
+static __inline uint32 Clamp(int32 val) {
+  int v = clamp0(val);
+  return (uint32)(clamp255(v));
+}
+
+static __inline uint32 Abs(int32 v) {
+  int m = v >> 31;
+  return (v + m) ^ m;
+}
+#else  // USE_BRANCHLESS
+static __inline int32 clamp0(int32 v) {
+  return (v < 0) ? 0 : v;
+}
+
+static __inline int32 clamp255(int32 v) {
+  return (v > 255) ? 255 : v;
+}
+
+static __inline uint32 Clamp(int32 val) {
+  int v = clamp0(val);
+  return (uint32)(clamp255(v));
+}
+
+static __inline uint32 Abs(int32 v) {
+  return (v < 0) ? -v : v;
+}
+#endif  // USE_BRANCHLESS
+
+#ifdef LIBYUV_LITTLE_ENDIAN
+#define WRITEWORD(p, v) *(uint32*)(p) = v
+#else
+static inline void WRITEWORD(uint8* p, uint32 v) {
+  p[0] = (uint8)(v & 255);
+  p[1] = (uint8)((v >> 8) & 255);
+  p[2] = (uint8)((v >> 16) & 255);
+  p[3] = (uint8)((v >> 24) & 255);
+}
+#endif
+
+void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb24[0];
+    uint8 g = src_rgb24[1];
+    uint8 r = src_rgb24[2];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_rgb24 += 3;
+  }
+}
+
+void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 r = src_raw[0];
+    uint8 g = src_raw[1];
+    uint8 b = src_raw[2];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_raw += 3;
+  }
+}
+
+void RGB565ToARGBRow_C(const uint8* src_rgb565, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb565[0] & 0x1f;
+    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r = src_rgb565[1] >> 3;
+    dst_argb[0] = (b << 3) | (b >> 2);
+    dst_argb[1] = (g << 2) | (g >> 4);
+    dst_argb[2] = (r << 3) | (r >> 2);
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_rgb565 += 2;
+  }
+}
+
+void ARGB1555ToARGBRow_C(const uint8* src_argb1555, uint8* dst_argb,
+                         int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb1555[0] & 0x1f;
+    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 a = src_argb1555[1] >> 7;
+    dst_argb[0] = (b << 3) | (b >> 2);
+    dst_argb[1] = (g << 3) | (g >> 2);
+    dst_argb[2] = (r << 3) | (r >> 2);
+    dst_argb[3] = -a;
+    dst_argb += 4;
+    src_argb1555 += 2;
+  }
+}
+
+void ARGB4444ToARGBRow_C(const uint8* src_argb4444, uint8* dst_argb,
+                         int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb4444[0] & 0x0f;
+    uint8 g = src_argb4444[0] >> 4;
+    uint8 r = src_argb4444[1] & 0x0f;
+    uint8 a = src_argb4444[1] >> 4;
+    dst_argb[0] = (b << 4) | b;
+    dst_argb[1] = (g << 4) | g;
+    dst_argb[2] = (r << 4) | r;
+    dst_argb[3] = (a << 4) | a;
+    dst_argb += 4;
+    src_argb4444 += 2;
+  }
+}
+
+void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb[0];
+    uint8 g = src_argb[1];
+    uint8 r = src_argb[2];
+    dst_rgb[0] = b;
+    dst_rgb[1] = g;
+    dst_rgb[2] = r;
+    dst_rgb += 3;
+    src_argb += 4;
+  }
+}
+
+void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb[0];
+    uint8 g = src_argb[1];
+    uint8 r = src_argb[2];
+    dst_rgb[0] = r;
+    dst_rgb[1] = g;
+    dst_rgb[2] = b;
+    dst_rgb += 3;
+    src_argb += 4;
+  }
+}
+
+void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 2;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 b1 = src_argb[4] >> 3;
+    uint8 g1 = src_argb[5] >> 2;
+    uint8 r1 = src_argb[6] >> 3;
+    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) |
+              (b1 << 16) | (g1 << 21) | (r1 << 27));
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 2;
+    uint8 r0 = src_argb[2] >> 3;
+    *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+// dither4 is a row of 4 values from 4x4 dither matrix.
+// The 4x4 matrix contains values to increase RGB.  When converting to
+// fewer bits (565) this provides an ordered dither.
+// The order in the 4x4 matrix in first byte is upper left.
+// The 4 values are passed as an int, then referenced as an array, so
+// endian will not affect order of the original matrix.  But the dither4
+// will containing the first pixel in the lower byte for little endian
+// or the upper byte for big endian.
+void ARGBToRGB565DitherRow_C(const uint8* src_argb, uint8* dst_rgb,
+                             const uint32 dither4, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    int dither0 = ((const unsigned char*)(&dither4))[x & 3];
+    int dither1 = ((const unsigned char*)(&dither4))[(x + 1) & 3];
+    uint8 b0 = clamp255(src_argb[0] + dither0) >> 3;
+    uint8 g0 = clamp255(src_argb[1] + dither0) >> 2;
+    uint8 r0 = clamp255(src_argb[2] + dither0) >> 3;
+    uint8 b1 = clamp255(src_argb[4] + dither1) >> 3;
+    uint8 g1 = clamp255(src_argb[5] + dither1) >> 2;
+    uint8 r1 = clamp255(src_argb[6] + dither1) >> 3;
+    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) |
+              (b1 << 16) | (g1 << 21) | (r1 << 27));
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    int dither0 = ((const unsigned char*)(&dither4))[(width - 1) & 3];
+    uint8 b0 = clamp255(src_argb[0] + dither0) >> 3;
+    uint8 g0 = clamp255(src_argb[1] + dither0) >> 2;
+    uint8 r0 = clamp255(src_argb[2] + dither0) >> 3;
+    *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 3;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 a0 = src_argb[3] >> 7;
+    uint8 b1 = src_argb[4] >> 3;
+    uint8 g1 = src_argb[5] >> 3;
+    uint8 r1 = src_argb[6] >> 3;
+    uint8 a1 = src_argb[7] >> 7;
+    *(uint32*)(dst_rgb) =
+        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15) |
+        (b1 << 16) | (g1 << 21) | (r1 << 26) | (a1 << 31);
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 3;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 a0 = src_argb[3] >> 7;
+    *(uint16*)(dst_rgb) =
+        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15);
+  }
+}
+
+void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 4;
+    uint8 g0 = src_argb[1] >> 4;
+    uint8 r0 = src_argb[2] >> 4;
+    uint8 a0 = src_argb[3] >> 4;
+    uint8 b1 = src_argb[4] >> 4;
+    uint8 g1 = src_argb[5] >> 4;
+    uint8 r1 = src_argb[6] >> 4;
+    uint8 a1 = src_argb[7] >> 4;
+    *(uint32*)(dst_rgb) =
+        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12) |
+        (b1 << 16) | (g1 << 20) | (r1 << 24) | (a1 << 28);
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 4;
+    uint8 g0 = src_argb[1] >> 4;
+    uint8 r0 = src_argb[2] >> 4;
+    uint8 a0 = src_argb[3] >> 4;
+    *(uint16*)(dst_rgb) =
+        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12);
+  }
+}
+
+static __inline int RGBToY(uint8 r, uint8 g, uint8 b) {
+  return (66 * r + 129 * g +  25 * b + 0x1080) >> 8;
+}
+
+static __inline int RGBToU(uint8 r, uint8 g, uint8 b) {
+  return (112 * b - 74 * g - 38 * r + 0x8080) >> 8;
+}
+static __inline int RGBToV(uint8 r, uint8 g, uint8 b) {
+  return (112 * r - 94 * g - 18 * b + 0x8080) >> 8;
+}
+
+#define MAKEROWY(NAME, R, G, B, BPP) \
+void NAME ## ToYRow_C(const uint8* src_argb0, uint8* dst_y, int width) {       \
+  int x;                                                                       \
+  for (x = 0; x < width; ++x) {                                                \
+    dst_y[0] = RGBToY(src_argb0[R], src_argb0[G], src_argb0[B]);               \
+    src_argb0 += BPP;                                                          \
+    dst_y += 1;                                                                \
+  }                                                                            \
+}                                                                              \
+void NAME ## ToUVRow_C(const uint8* src_rgb0, int src_stride_rgb,              \
+                       uint8* dst_u, uint8* dst_v, int width) {                \
+  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
+  int x;                                                                       \
+  for (x = 0; x < width - 1; x += 2) {                                         \
+    uint8 ab = (src_rgb0[B] + src_rgb0[B + BPP] +                              \
+               src_rgb1[B] + src_rgb1[B + BPP]) >> 2;                          \
+    uint8 ag = (src_rgb0[G] + src_rgb0[G + BPP] +                              \
+               src_rgb1[G] + src_rgb1[G + BPP]) >> 2;                          \
+    uint8 ar = (src_rgb0[R] + src_rgb0[R + BPP] +                              \
+               src_rgb1[R] + src_rgb1[R + BPP]) >> 2;                          \
+    dst_u[0] = RGBToU(ar, ag, ab);                                             \
+    dst_v[0] = RGBToV(ar, ag, ab);                                             \
+    src_rgb0 += BPP * 2;                                                       \
+    src_rgb1 += BPP * 2;                                                       \
+    dst_u += 1;                                                                \
+    dst_v += 1;                                                                \
+  }                                                                            \
+  if (width & 1) {                                                             \
+    uint8 ab = (src_rgb0[B] + src_rgb1[B]) >> 1;                               \
+    uint8 ag = (src_rgb0[G] + src_rgb1[G]) >> 1;                               \
+    uint8 ar = (src_rgb0[R] + src_rgb1[R]) >> 1;                               \
+    dst_u[0] = RGBToU(ar, ag, ab);                                             \
+    dst_v[0] = RGBToV(ar, ag, ab);                                             \
+  }                                                                            \
+}
+
+MAKEROWY(ARGB, 2, 1, 0, 4)
+MAKEROWY(BGRA, 1, 2, 3, 4)
+MAKEROWY(ABGR, 0, 1, 2, 4)
+MAKEROWY(RGBA, 3, 2, 1, 4)
+MAKEROWY(RGB24, 2, 1, 0, 3)
+MAKEROWY(RAW, 0, 1, 2, 3)
+#undef MAKEROWY
+
+// JPeg uses a variation on BT.601-1 full range
+// y =  0.29900 * r + 0.58700 * g + 0.11400 * b
+// u = -0.16874 * r - 0.33126 * g + 0.50000 * b  + center
+// v =  0.50000 * r - 0.41869 * g - 0.08131 * b  + center
+// BT.601 Mpeg range uses:
+// b 0.1016 * 255 = 25.908 = 25
+// g 0.5078 * 255 = 129.489 = 129
+// r 0.2578 * 255 = 65.739 = 66
+// JPeg 8 bit Y (not used):
+// b 0.11400 * 256 = 29.184 = 29
+// g 0.58700 * 256 = 150.272 = 150
+// r 0.29900 * 256 = 76.544 = 77
+// JPeg 7 bit Y:
+// b 0.11400 * 128 = 14.592 = 15
+// g 0.58700 * 128 = 75.136 = 75
+// r 0.29900 * 128 = 38.272 = 38
+// JPeg 8 bit U:
+// b  0.50000 * 255 = 127.5 = 127
+// g -0.33126 * 255 = -84.4713 = -84
+// r -0.16874 * 255 = -43.0287 = -43
+// JPeg 8 bit V:
+// b -0.08131 * 255 = -20.73405 = -20
+// g -0.41869 * 255 = -106.76595 = -107
+// r  0.50000 * 255 = 127.5 = 127
+
+static __inline int RGBToYJ(uint8 r, uint8 g, uint8 b) {
+  return (38 * r + 75 * g +  15 * b + 64) >> 7;
+}
+
+static __inline int RGBToUJ(uint8 r, uint8 g, uint8 b) {
+  return (127 * b - 84 * g - 43 * r + 0x8080) >> 8;
+}
+static __inline int RGBToVJ(uint8 r, uint8 g, uint8 b) {
+  return (127 * r - 107 * g - 20 * b + 0x8080) >> 8;
+}
+
+#define AVGB(a, b) (((a) + (b) + 1) >> 1)
+
+#define MAKEROWYJ(NAME, R, G, B, BPP) \
+void NAME ## ToYJRow_C(const uint8* src_argb0, uint8* dst_y, int width) {      \
+  int x;                                                                       \
+  for (x = 0; x < width; ++x) {                                                \
+    dst_y[0] = RGBToYJ(src_argb0[R], src_argb0[G], src_argb0[B]);              \
+    src_argb0 += BPP;                                                          \
+    dst_y += 1;                                                                \
+  }                                                                            \
+}                                                                              \
+void NAME ## ToUVJRow_C(const uint8* src_rgb0, int src_stride_rgb,             \
+                        uint8* dst_u, uint8* dst_v, int width) {               \
+  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
+  int x;                                                                       \
+  for (x = 0; x < width - 1; x += 2) {                                         \
+    uint8 ab = AVGB(AVGB(src_rgb0[B], src_rgb1[B]),                            \
+                    AVGB(src_rgb0[B + BPP], src_rgb1[B + BPP]));               \
+    uint8 ag = AVGB(AVGB(src_rgb0[G], src_rgb1[G]),                            \
+                    AVGB(src_rgb0[G + BPP], src_rgb1[G + BPP]));               \
+    uint8 ar = AVGB(AVGB(src_rgb0[R], src_rgb1[R]),                            \
+                    AVGB(src_rgb0[R + BPP], src_rgb1[R + BPP]));               \
+    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
+    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
+    src_rgb0 += BPP * 2;                                                       \
+    src_rgb1 += BPP * 2;                                                       \
+    dst_u += 1;                                                                \
+    dst_v += 1;                                                                \
+  }                                                                            \
+  if (width & 1) {                                                             \
+    uint8 ab = AVGB(src_rgb0[B], src_rgb1[B]);                                 \
+    uint8 ag = AVGB(src_rgb0[G], src_rgb1[G]);                                 \
+    uint8 ar = AVGB(src_rgb0[R], src_rgb1[R]);                                 \
+    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
+    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
+  }                                                                            \
+}
+
+MAKEROWYJ(ARGB, 2, 1, 0, 4)
+#undef MAKEROWYJ
+
+void ARGBToUVJ422Row_C(const uint8* src_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToUJ(ar, ag, ab);
+    dst_v[0] = RGBToVJ(ar, ag, ab);
+    src_argb += 8;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToUJ(ar, ag, ab);
+    dst_v[0] = RGBToVJ(ar, ag, ab);
+  }
+}
+
+void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb565[0] & 0x1f;
+    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r = src_rgb565[1] >> 3;
+    b = (b << 3) | (b >> 2);
+    g = (g << 2) | (g >> 4);
+    r = (r << 3) | (r >> 2);
+    dst_y[0] = RGBToY(r, g, b);
+    src_rgb565 += 2;
+    dst_y += 1;
+  }
+}
+
+void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb1555[0] & 0x1f;
+    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
+    b = (b << 3) | (b >> 2);
+    g = (g << 3) | (g >> 2);
+    r = (r << 3) | (r >> 2);
+    dst_y[0] = RGBToY(r, g, b);
+    src_argb1555 += 2;
+    dst_y += 1;
+  }
+}
+
+void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb4444[0] & 0x0f;
+    uint8 g = src_argb4444[0] >> 4;
+    uint8 r = src_argb4444[1] & 0x0f;
+    b = (b << 4) | b;
+    g = (g << 4) | g;
+    r = (r << 4) | r;
+    dst_y[0] = RGBToY(r, g, b);
+    src_argb4444 += 2;
+    dst_y += 1;
+  }
+}
+
+void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
+                     uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_rgb565 = src_rgb565 + src_stride_rgb565;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_rgb565[0] & 0x1f;
+    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r0 = src_rgb565[1] >> 3;
+    uint8 b1 = src_rgb565[2] & 0x1f;
+    uint8 g1 = (src_rgb565[2] >> 5) | ((src_rgb565[3] & 0x07) << 3);
+    uint8 r1 = src_rgb565[3] >> 3;
+    uint8 b2 = next_rgb565[0] & 0x1f;
+    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
+    uint8 r2 = next_rgb565[1] >> 3;
+    uint8 b3 = next_rgb565[2] & 0x1f;
+    uint8 g3 = (next_rgb565[2] >> 5) | ((next_rgb565[3] & 0x07) << 3);
+    uint8 r3 = next_rgb565[3] >> 3;
+    uint8 b = (b0 + b1 + b2 + b3);  // 565 * 4 = 787.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 1) | (b >> 6);  // 787 -> 888.
+    r = (r << 1) | (r >> 6);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_rgb565 += 4;
+    next_rgb565 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_rgb565[0] & 0x1f;
+    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r0 = src_rgb565[1] >> 3;
+    uint8 b2 = next_rgb565[0] & 0x1f;
+    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
+    uint8 r2 = next_rgb565[1] >> 3;
+    uint8 b = (b0 + b2);  // 565 * 2 = 676.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 2) | (b >> 4);  // 676 -> 888
+    g = (g << 1) | (g >> 6);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_argb1555 = src_argb1555 + src_stride_argb1555;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb1555[0] & 0x1f;
+    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 b1 = src_argb1555[2] & 0x1f;
+    uint8 g1 = (src_argb1555[2] >> 5) | ((src_argb1555[3] & 0x03) << 3);
+    uint8 r1 = (src_argb1555[3] & 0x7c) >> 2;
+    uint8 b2 = next_argb1555[0] & 0x1f;
+    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
+    uint8 r2 = (next_argb1555[1] & 0x7c) >> 2;
+    uint8 b3 = next_argb1555[2] & 0x1f;
+    uint8 g3 = (next_argb1555[2] >> 5) | ((next_argb1555[3] & 0x03) << 3);
+    uint8 r3 = (next_argb1555[3] & 0x7c) >> 2;
+    uint8 b = (b0 + b1 + b2 + b3);  // 555 * 4 = 777.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 1) | (b >> 6);  // 777 -> 888.
+    g = (g << 1) | (g >> 6);
+    r = (r << 1) | (r >> 6);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_argb1555 += 4;
+    next_argb1555 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb1555[0] & 0x1f;
+    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 b2 = next_argb1555[0] & 0x1f;
+    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
+    uint8 r2 = next_argb1555[1] >> 3;
+    uint8 b = (b0 + b2);  // 555 * 2 = 666.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 2) | (b >> 4);  // 666 -> 888.
+    g = (g << 2) | (g >> 4);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_argb4444 = src_argb4444 + src_stride_argb4444;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb4444[0] & 0x0f;
+    uint8 g0 = src_argb4444[0] >> 4;
+    uint8 r0 = src_argb4444[1] & 0x0f;
+    uint8 b1 = src_argb4444[2] & 0x0f;
+    uint8 g1 = src_argb4444[2] >> 4;
+    uint8 r1 = src_argb4444[3] & 0x0f;
+    uint8 b2 = next_argb4444[0] & 0x0f;
+    uint8 g2 = next_argb4444[0] >> 4;
+    uint8 r2 = next_argb4444[1] & 0x0f;
+    uint8 b3 = next_argb4444[2] & 0x0f;
+    uint8 g3 = next_argb4444[2] >> 4;
+    uint8 r3 = next_argb4444[3] & 0x0f;
+    uint8 b = (b0 + b1 + b2 + b3);  // 444 * 4 = 666.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 2) | (b >> 4);  // 666 -> 888.
+    g = (g << 2) | (g >> 4);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_argb4444 += 4;
+    next_argb4444 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb4444[0] & 0x0f;
+    uint8 g0 = src_argb4444[0] >> 4;
+    uint8 r0 = src_argb4444[1] & 0x0f;
+    uint8 b2 = next_argb4444[0] & 0x0f;
+    uint8 g2 = next_argb4444[0] >> 4;
+    uint8 r2 = next_argb4444[1] & 0x0f;
+    uint8 b = (b0 + b2);  // 444 * 2 = 555.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 3) | (b >> 2);  // 555 -> 888.
+    g = (g << 3) | (g >> 2);
+    r = (r << 3) | (r >> 2);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGBToUV444Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+void ARGBToUV422Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 8;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  }
+}
+
+void ARGBToUV411Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 3; x += 4) {
+    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8] + src_argb[12]) >> 2;
+    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9] + src_argb[13]) >> 2;
+    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10] + src_argb[14]) >> 2;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 16;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if ((width & 3) == 3) {
+    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8]) / 3;
+    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9]) / 3;
+    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10]) / 3;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  } else if ((width & 3) == 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  } else if ((width & 3) == 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  }
+}
+
+void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 y = RGBToYJ(src_argb[2], src_argb[1], src_argb[0]);
+    dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
+    dst_argb[3] = src_argb[3];
+    dst_argb += 4;
+    src_argb += 4;
+  }
+}
+
+// Convert a row of image to Sepia tone.
+void ARGBSepiaRow_C(uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    int sb = (b * 17 + g * 68 + r * 35) >> 7;
+    int sg = (b * 22 + g * 88 + r * 45) >> 7;
+    int sr = (b * 24 + g * 98 + r * 50) >> 7;
+    // b does not over flow. a is preserved from original.
+    dst_argb[0] = sb;
+    dst_argb[1] = clamp255(sg);
+    dst_argb[2] = clamp255(sr);
+    dst_argb += 4;
+  }
+}
+
+// Apply color matrix to a row of image. Matrix is signed.
+// TODO(fbarchard): Consider adding rounding (+32).
+void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
+                          const int8* matrix_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = src_argb[0];
+    int g = src_argb[1];
+    int r = src_argb[2];
+    int a = src_argb[3];
+    int sb = (b * matrix_argb[0] + g * matrix_argb[1] +
+              r * matrix_argb[2] + a * matrix_argb[3]) >> 6;
+    int sg = (b * matrix_argb[4] + g * matrix_argb[5] +
+              r * matrix_argb[6] + a * matrix_argb[7]) >> 6;
+    int sr = (b * matrix_argb[8] + g * matrix_argb[9] +
+              r * matrix_argb[10] + a * matrix_argb[11]) >> 6;
+    int sa = (b * matrix_argb[12] + g * matrix_argb[13] +
+              r * matrix_argb[14] + a * matrix_argb[15]) >> 6;
+    dst_argb[0] = Clamp(sb);
+    dst_argb[1] = Clamp(sg);
+    dst_argb[2] = Clamp(sr);
+    dst_argb[3] = Clamp(sa);
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+// Apply color table to a row of image.
+void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    int a = dst_argb[3];
+    dst_argb[0] = table_argb[b * 4 + 0];
+    dst_argb[1] = table_argb[g * 4 + 1];
+    dst_argb[2] = table_argb[r * 4 + 2];
+    dst_argb[3] = table_argb[a * 4 + 3];
+    dst_argb += 4;
+  }
+}
+
+// Apply color table to a row of image.
+void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    dst_argb[0] = table_argb[b * 4 + 0];
+    dst_argb[1] = table_argb[g * 4 + 1];
+    dst_argb[2] = table_argb[r * 4 + 2];
+    dst_argb += 4;
+  }
+}
+
+void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
+                       int interval_offset, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
+    dst_argb[1] = (g * scale >> 16) * interval_size + interval_offset;
+    dst_argb[2] = (r * scale >> 16) * interval_size + interval_offset;
+    dst_argb += 4;
+  }
+}
+
+#define REPEAT8(v) (v) | ((v) << 8)
+#define SHADE(f, v) v * f >> 24
+
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value) {
+  const uint32 b_scale = REPEAT8(value & 0xff);
+  const uint32 g_scale = REPEAT8((value >> 8) & 0xff);
+  const uint32 r_scale = REPEAT8((value >> 16) & 0xff);
+  const uint32 a_scale = REPEAT8(value >> 24);
+
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32 b = REPEAT8(src_argb[0]);
+    const uint32 g = REPEAT8(src_argb[1]);
+    const uint32 r = REPEAT8(src_argb[2]);
+    const uint32 a = REPEAT8(src_argb[3]);
+    dst_argb[0] = SHADE(b, b_scale);
+    dst_argb[1] = SHADE(g, g_scale);
+    dst_argb[2] = SHADE(r, r_scale);
+    dst_argb[3] = SHADE(a, a_scale);
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+#undef REPEAT8
+#undef SHADE
+
+#define REPEAT8(v) (v) | ((v) << 8)
+#define SHADE(f, v) v * f >> 16
+
+void ARGBMultiplyRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32 b = REPEAT8(src_argb0[0]);
+    const uint32 g = REPEAT8(src_argb0[1]);
+    const uint32 r = REPEAT8(src_argb0[2]);
+    const uint32 a = REPEAT8(src_argb0[3]);
+    const uint32 b_scale = src_argb1[0];
+    const uint32 g_scale = src_argb1[1];
+    const uint32 r_scale = src_argb1[2];
+    const uint32 a_scale = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_scale);
+    dst_argb[1] = SHADE(g, g_scale);
+    dst_argb[2] = SHADE(r, r_scale);
+    dst_argb[3] = SHADE(a, a_scale);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef REPEAT8
+#undef SHADE
+
+#define SHADE(f, v) clamp255(v + f)
+
+void ARGBAddRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                  uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const int b = src_argb0[0];
+    const int g = src_argb0[1];
+    const int r = src_argb0[2];
+    const int a = src_argb0[3];
+    const int b_add = src_argb1[0];
+    const int g_add = src_argb1[1];
+    const int r_add = src_argb1[2];
+    const int a_add = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_add);
+    dst_argb[1] = SHADE(g, g_add);
+    dst_argb[2] = SHADE(r, r_add);
+    dst_argb[3] = SHADE(a, a_add);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
+#define SHADE(f, v) clamp0(f - v)
+
+void ARGBSubtractRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const int b = src_argb0[0];
+    const int g = src_argb0[1];
+    const int r = src_argb0[2];
+    const int a = src_argb0[3];
+    const int b_sub = src_argb1[0];
+    const int g_sub = src_argb1[1];
+    const int r_sub = src_argb1[2];
+    const int a_sub = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_sub);
+    dst_argb[1] = SHADE(g, g_sub);
+    dst_argb[2] = SHADE(r, r_sub);
+    dst_argb[3] = SHADE(a, a_sub);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
+// Sobel functions which mimics SSSE3.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int a = src_y0[i];
+    int b = src_y1[i];
+    int c = src_y2[i];
+    int a_sub = src_y0[i + 2];
+    int b_sub = src_y1[i + 2];
+    int c_sub = src_y2[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = Abs(a_diff + b_diff * 2 + c_diff);
+    dst_sobelx[i] = (uint8)(clamp255(sobel));
+  }
+}
+
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int a = src_y0[i + 0];
+    int b = src_y0[i + 1];
+    int c = src_y0[i + 2];
+    int a_sub = src_y1[i + 0];
+    int b_sub = src_y1[i + 1];
+    int c_sub = src_y1[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = Abs(a_diff + b_diff * 2 + c_diff);
+    dst_sobely[i] = (uint8)(clamp255(sobel));
+  }
+}
+
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = clamp255(r + b);
+    dst_argb[0] = (uint8)(s);
+    dst_argb[1] = (uint8)(s);
+    dst_argb[2] = (uint8)(s);
+    dst_argb[3] = (uint8)(255u);
+    dst_argb += 4;
+  }
+}
+
+void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_y, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = clamp255(r + b);
+    dst_y[i] = (uint8)(s);
+  }
+}
+
+void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                  uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int g = clamp255(r + b);
+    dst_argb[0] = (uint8)(b);
+    dst_argb[1] = (uint8)(g);
+    dst_argb[2] = (uint8)(r);
+    dst_argb[3] = (uint8)(255u);
+    dst_argb += 4;
+  }
+}
+
+void J400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) {
+  // Copy a Y to RGB.
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 y = src_y[0];
+    dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    ++src_y;
+  }
+}
+
+// BT.601 YUV to RGB reference
+//  R = (Y - 16) * 1.164              - V * -1.596
+//  G = (Y - 16) * 1.164 - U *  0.391 - V *  0.813
+//  B = (Y - 16) * 1.164 - U * -2.018
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* max(-128, round(-2.018 * 64)) */
+#define UG 25 /* round(0.391 * 64) */
+#define VG 52 /* round(0.813 * 64) */
+#define VR -102 /* round(-1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128 + YGB)
+#define BG (UG * 128 + VG * 128 + YGB)
+#define BR (VR * 128 + YGB)
+
+// C reference code that mimics the YUV assembly.
+static __inline void YuvPixel(uint8 y, uint8 u, uint8 v,
+                              uint8* b, uint8* g, uint8* r) {
+  uint32 y1 = (uint32)(y * 0x0101 * YG) >> 16;
+  *b = Clamp((int32)(-(u * UB) + y1 + BB) >> 6);
+  *g = Clamp((int32)(-(v * VG + u * UG) + y1 + BG) >> 6);
+  *r = Clamp((int32)(-(v * VR)+ y1 + BR) >> 6);
+}
+
+// C reference code that mimics the YUV assembly.
+static __inline void YPixel(uint8 y, uint8* b, uint8* g, uint8* r) {
+  uint32 y1 = (uint32)(y * 0x0101 * YG) >> 16;
+  *b = Clamp((int32)(y1 + YGB) >> 6);
+  *g = Clamp((int32)(y1 + YGB) >> 6);
+  *r = Clamp((int32)(y1 + YGB) >> 6);
+}
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+// JPEG YUV to RGB reference
+// *  R = Y                - V * -1.40200
+// *  G = Y - U *  0.34414 - V *  0.71414
+// *  B = Y - U * -1.77200
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YGJ 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+#define YGBJ 32  /* 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UBJ -113 /* round(-1.77200 * 64) */
+#define UGJ 22 /* round(0.34414 * 64) */
+#define VGJ 46 /* round(0.71414  * 64) */
+#define VRJ -90 /* round(-1.40200 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BBJ (UBJ * 128 + YGBJ)
+#define BGJ (UGJ * 128 + VGJ * 128 + YGBJ)
+#define BRJ (VRJ * 128 + YGBJ)
+
+// C reference code that mimics the YUV assembly.
+static __inline void YuvJPixel(uint8 y, uint8 u, uint8 v,
+                               uint8* b, uint8* g, uint8* r) {
+  uint32 y1 = (uint32)(y * 0x0101 * YGJ) >> 16;
+  *b = Clamp((int32)(-(u * UBJ) + y1 + BBJ) >> 6);
+  *g = Clamp((int32)(-(v * VGJ + u * UGJ) + y1 + BGJ) >> 6);
+  *r = Clamp((int32)(-(v * VRJ) + y1 + BRJ) >> 6);
+}
+
+#undef YGJ
+#undef YGBJ
+#undef UBJ
+#undef UGJ
+#undef VGJ
+#undef VRJ
+#undef BBJ
+#undef BGJ
+#undef BRJ
+
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(__aarch64__) || defined(LIBYUV_NEON))
+// C mimic assembly.
+// TODO(fbarchard): Remove subsampling from Neon.
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 u = (src_u[0] + src_u[1] + 1) >> 1;
+    uint8 v = (src_v[0] + src_v[1] + 1) >> 1;
+    YuvPixel(src_y[0], u, v, rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], u, v, rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 2;
+    src_v += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+  }
+}
+#else
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 4;  // Advance 1 pixel.
+  }
+}
+#endif
+
+// Also used for 420
+void I422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void J422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvJPixel(src_y[0], src_u[0], src_v[0],
+              rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvJPixel(src_y[1], src_u[0], src_v[0],
+              rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvJPixel(src_y[0], src_u[0], src_v[0],
+              rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToRGB24Row_C(const uint8* src_y,
+                      const uint8* src_u,
+                      const uint8* src_v,
+                      uint8* rgb_buf,
+                      int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 4, rgb_buf + 5);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+  }
+}
+
+void I422ToRAWRow_C(const uint8* src_y,
+                    const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* rgb_buf,
+                    int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 5, rgb_buf + 4, rgb_buf + 3);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+  }
+}
+
+void I422ToARGB4444Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 4;
+    g0 = g0 >> 4;
+    r0 = r0 >> 4;
+    b1 = b1 >> 4;
+    g1 = g1 >> 4;
+    r1 = r1 >> 4;
+    *(uint32*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
+        (b1 << 16) | (g1 << 20) | (r1 << 24) | 0xf000f000;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_argb4444 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 4;
+    g0 = g0 >> 4;
+    r0 = r0 >> 4;
+    *(uint16*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
+        0xf000;
+  }
+}
+
+void I422ToARGB1555Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb1555,
+                         int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 3;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 3;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
+        (b1 << 16) | (g1 << 21) | (r1 << 26) | 0x80008000;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_argb1555 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 3;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
+        0x8000;
+  }
+}
+
+void I422ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void I411ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 3; x += 4) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    YuvPixel(src_y[2], src_u[0], src_v[0],
+             rgb_buf + 8, rgb_buf + 9, rgb_buf + 10);
+    rgb_buf[11] = 255;
+    YuvPixel(src_y[3], src_u[0], src_v[0],
+             rgb_buf + 12, rgb_buf + 13, rgb_buf + 14);
+    rgb_buf[15] = 255;
+    src_y += 4;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 16;  // Advance 4 pixels.
+  }
+  if (width & 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV12ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_uv,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_uv[0], src_uv[1],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_uv[0], src_uv[1],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_uv += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_uv[0], src_uv[1],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV21ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_vu,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+
+    YuvPixel(src_y[1], src_vu[1], src_vu[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+
+    src_y += 2;
+    src_vu += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV12ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_uv,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_uv[0], src_uv[1], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_uv[0], src_uv[1], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    src_uv += 2;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_uv[0], src_uv[1], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void NV21ToRGB565Row_C(const uint8* src_y,
+                       const uint8* vsrc_u,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], vsrc_u[1], vsrc_u[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    vsrc_u += 2;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void YUY2ToARGBRow_C(const uint8* src_yuy2,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_yuy2[2], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_yuy2 += 4;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void UYVYToARGBRow_C(const uint8* src_uyvy,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_uyvy[3], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_uyvy += 4;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToBGRARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 2, rgb_buf + 1);
+    rgb_buf[0] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 7, rgb_buf + 6, rgb_buf + 5);
+    rgb_buf[4] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 2, rgb_buf + 1);
+    rgb_buf[0] = 255;
+  }
+}
+
+void I422ToABGRRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 6, rgb_buf + 5, rgb_buf + 4);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToRGBARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3);
+    rgb_buf[0] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 5, rgb_buf + 6, rgb_buf + 7);
+    rgb_buf[4] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3);
+    rgb_buf[0] = 255;
+  }
+}
+
+void I400ToARGBRow_C(const uint8* src_y, uint8* rgb_buf, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YPixel(src_y[1], rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void MirrorRow_C(const uint8* src, uint8* dst, int width) {
+  int x;
+  src += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst[x] = src[0];
+    dst[x + 1] = src[-1];
+    src -= 2;
+  }
+  if (width & 1) {
+    dst[width - 1] = src[0];
+  }
+}
+
+void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  src_uv += (width - 1) << 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_u[x] = src_uv[0];
+    dst_u[x + 1] = src_uv[-2];
+    dst_v[x] = src_uv[1];
+    dst_v[x + 1] = src_uv[-2 + 1];
+    src_uv -= 4;
+  }
+  if (width & 1) {
+    dst_u[width - 1] = src_uv[0];
+    dst_v[width - 1] = src_uv[1];
+  }
+}
+
+void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width) {
+  int x;
+  const uint32* src32 = (const uint32*)(src);
+  uint32* dst32 = (uint32*)(dst);
+  src32 += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst32[x] = src32[0];
+    dst32[x + 1] = src32[-1];
+    src32 -= 2;
+  }
+  if (width & 1) {
+    dst32[width - 1] = src32[0];
+  }
+}
+
+void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_u[x] = src_uv[0];
+    dst_u[x + 1] = src_uv[2];
+    dst_v[x] = src_uv[1];
+    dst_v[x + 1] = src_uv[3];
+    src_uv += 4;
+  }
+  if (width & 1) {
+    dst_u[width - 1] = src_uv[0];
+    dst_v[width - 1] = src_uv[1];
+  }
+}
+
+void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                  int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_uv[0] = src_u[x];
+    dst_uv[1] = src_v[x];
+    dst_uv[2] = src_u[x + 1];
+    dst_uv[3] = src_v[x + 1];
+    dst_uv += 4;
+  }
+  if (width & 1) {
+    dst_uv[0] = src_u[width - 1];
+    dst_uv[1] = src_v[width - 1];
+  }
+}
+
+void CopyRow_C(const uint8* src, uint8* dst, int count) {
+  memcpy(dst, src, count);
+}
+
+void CopyRow_16_C(const uint16* src, uint16* dst, int count) {
+  memcpy(dst, src, count * 2);
+}
+
+void SetRow_C(uint8* dst, uint8 v8, int width) {
+  memset(dst, v8, width);
+}
+
+void ARGBSetRow_C(uint8* dst_argb, uint32 v32, int width) {
+  uint32* d = (uint32*)(dst_argb);
+  int x;
+  for (x = 0; x < width; ++x) {
+    d[x] = v32;
+  }
+}
+
+// Filter 2 rows of YUY2 UV's (422) into U and V (420).
+void YUY2ToUVRow_C(const uint8* src_yuy2, int src_stride_yuy2,
+                   uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values, filtering 2 rows of YUY2.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = (src_yuy2[1] + src_yuy2[src_stride_yuy2 + 1] + 1) >> 1;
+    dst_v[0] = (src_yuy2[3] + src_yuy2[src_stride_yuy2 + 3] + 1) >> 1;
+    src_yuy2 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of YUY2 UV's (422) into U and V (422).
+void YUY2ToUV422Row_C(const uint8* src_yuy2,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = src_yuy2[1];
+    dst_v[0] = src_yuy2[3];
+    src_yuy2 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of YUY2 Y's (422) into Y (420/422).
+void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_y[x] = src_yuy2[0];
+    dst_y[x + 1] = src_yuy2[2];
+    src_yuy2 += 4;
+  }
+  if (width & 1) {
+    dst_y[width - 1] = src_yuy2[0];
+  }
+}
+
+// Filter 2 rows of UYVY UV's (422) into U and V (420).
+void UYVYToUVRow_C(const uint8* src_uyvy, int src_stride_uyvy,
+                   uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = (src_uyvy[0] + src_uyvy[src_stride_uyvy + 0] + 1) >> 1;
+    dst_v[0] = (src_uyvy[2] + src_uyvy[src_stride_uyvy + 2] + 1) >> 1;
+    src_uyvy += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of UYVY UV's (422) into U and V (422).
+void UYVYToUV422Row_C(const uint8* src_uyvy,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = src_uyvy[0];
+    dst_v[0] = src_uyvy[2];
+    src_uyvy += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of UYVY Y's (422) into Y (420/422).
+void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_y[x] = src_uyvy[1];
+    dst_y[x + 1] = src_uyvy[3];
+    src_uyvy += 4;
+  }
+  if (width & 1) {
+    dst_y[width - 1] = src_uyvy[1];
+  }
+}
+
+#define BLEND(f, b, a) (((256 - a) * b) >> 8) + f
+
+// Blend src_argb0 over src_argb1 and store to dst_argb.
+// dst_argb may be src_argb0 or src_argb1.
+// This code mimics the SSSE3 version for better testability.
+void ARGBBlendRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                    uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint32 fb = src_argb0[0];
+    uint32 fg = src_argb0[1];
+    uint32 fr = src_argb0[2];
+    uint32 a = src_argb0[3];
+    uint32 bb = src_argb1[0];
+    uint32 bg = src_argb1[1];
+    uint32 br = src_argb1[2];
+    dst_argb[0] = BLEND(fb, bb, a);
+    dst_argb[1] = BLEND(fg, bg, a);
+    dst_argb[2] = BLEND(fr, br, a);
+    dst_argb[3] = 255u;
+
+    fb = src_argb0[4 + 0];
+    fg = src_argb0[4 + 1];
+    fr = src_argb0[4 + 2];
+    a = src_argb0[4 + 3];
+    bb = src_argb1[4 + 0];
+    bg = src_argb1[4 + 1];
+    br = src_argb1[4 + 2];
+    dst_argb[4 + 0] = BLEND(fb, bb, a);
+    dst_argb[4 + 1] = BLEND(fg, bg, a);
+    dst_argb[4 + 2] = BLEND(fr, br, a);
+    dst_argb[4 + 3] = 255u;
+    src_argb0 += 8;
+    src_argb1 += 8;
+    dst_argb += 8;
+  }
+
+  if (width & 1) {
+    uint32 fb = src_argb0[0];
+    uint32 fg = src_argb0[1];
+    uint32 fr = src_argb0[2];
+    uint32 a = src_argb0[3];
+    uint32 bb = src_argb1[0];
+    uint32 bg = src_argb1[1];
+    uint32 br = src_argb1[2];
+    dst_argb[0] = BLEND(fb, bb, a);
+    dst_argb[1] = BLEND(fg, bg, a);
+    dst_argb[2] = BLEND(fr, br, a);
+    dst_argb[3] = 255u;
+  }
+}
+#undef BLEND
+#define ATTENUATE(f, a) (a | (a << 8)) * (f | (f << 8)) >> 24
+
+// Multiply source RGB by alpha and store to destination.
+// This code mimics the SSSE3 version for better testability.
+void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    uint32 b = src_argb[0];
+    uint32 g = src_argb[1];
+    uint32 r = src_argb[2];
+    uint32 a = src_argb[3];
+    dst_argb[0] = ATTENUATE(b, a);
+    dst_argb[1] = ATTENUATE(g, a);
+    dst_argb[2] = ATTENUATE(r, a);
+    dst_argb[3] = a;
+    b = src_argb[4];
+    g = src_argb[5];
+    r = src_argb[6];
+    a = src_argb[7];
+    dst_argb[4] = ATTENUATE(b, a);
+    dst_argb[5] = ATTENUATE(g, a);
+    dst_argb[6] = ATTENUATE(r, a);
+    dst_argb[7] = a;
+    src_argb += 8;
+    dst_argb += 8;
+  }
+
+  if (width & 1) {
+    const uint32 b = src_argb[0];
+    const uint32 g = src_argb[1];
+    const uint32 r = src_argb[2];
+    const uint32 a = src_argb[3];
+    dst_argb[0] = ATTENUATE(b, a);
+    dst_argb[1] = ATTENUATE(g, a);
+    dst_argb[2] = ATTENUATE(r, a);
+    dst_argb[3] = a;
+  }
+}
+#undef ATTENUATE
+
+// Divide source RGB by alpha and store to destination.
+// b = (b * 255 + (a / 2)) / a;
+// g = (g * 255 + (a / 2)) / a;
+// r = (r * 255 + (a / 2)) / a;
+// Reciprocal method is off by 1 on some values. ie 125
+// 8.8 fixed point inverse table with 1.0 in upper short and 1 / a in lower.
+#define T(a) 0x01000000 + (0x10000 / a)
+const uint32 fixed_invtbl8[256] = {
+  0x01000000, 0x0100ffff, T(0x02), T(0x03), T(0x04), T(0x05), T(0x06), T(0x07),
+  T(0x08), T(0x09), T(0x0a), T(0x0b), T(0x0c), T(0x0d), T(0x0e), T(0x0f),
+  T(0x10), T(0x11), T(0x12), T(0x13), T(0x14), T(0x15), T(0x16), T(0x17),
+  T(0x18), T(0x19), T(0x1a), T(0x1b), T(0x1c), T(0x1d), T(0x1e), T(0x1f),
+  T(0x20), T(0x21), T(0x22), T(0x23), T(0x24), T(0x25), T(0x26), T(0x27),
+  T(0x28), T(0x29), T(0x2a), T(0x2b), T(0x2c), T(0x2d), T(0x2e), T(0x2f),
+  T(0x30), T(0x31), T(0x32), T(0x33), T(0x34), T(0x35), T(0x36), T(0x37),
+  T(0x38), T(0x39), T(0x3a), T(0x3b), T(0x3c), T(0x3d), T(0x3e), T(0x3f),
+  T(0x40), T(0x41), T(0x42), T(0x43), T(0x44), T(0x45), T(0x46), T(0x47),
+  T(0x48), T(0x49), T(0x4a), T(0x4b), T(0x4c), T(0x4d), T(0x4e), T(0x4f),
+  T(0x50), T(0x51), T(0x52), T(0x53), T(0x54), T(0x55), T(0x56), T(0x57),
+  T(0x58), T(0x59), T(0x5a), T(0x5b), T(0x5c), T(0x5d), T(0x5e), T(0x5f),
+  T(0x60), T(0x61), T(0x62), T(0x63), T(0x64), T(0x65), T(0x66), T(0x67),
+  T(0x68), T(0x69), T(0x6a), T(0x6b), T(0x6c), T(0x6d), T(0x6e), T(0x6f),
+  T(0x70), T(0x71), T(0x72), T(0x73), T(0x74), T(0x75), T(0x76), T(0x77),
+  T(0x78), T(0x79), T(0x7a), T(0x7b), T(0x7c), T(0x7d), T(0x7e), T(0x7f),
+  T(0x80), T(0x81), T(0x82), T(0x83), T(0x84), T(0x85), T(0x86), T(0x87),
+  T(0x88), T(0x89), T(0x8a), T(0x8b), T(0x8c), T(0x8d), T(0x8e), T(0x8f),
+  T(0x90), T(0x91), T(0x92), T(0x93), T(0x94), T(0x95), T(0x96), T(0x97),
+  T(0x98), T(0x99), T(0x9a), T(0x9b), T(0x9c), T(0x9d), T(0x9e), T(0x9f),
+  T(0xa0), T(0xa1), T(0xa2), T(0xa3), T(0xa4), T(0xa5), T(0xa6), T(0xa7),
+  T(0xa8), T(0xa9), T(0xaa), T(0xab), T(0xac), T(0xad), T(0xae), T(0xaf),
+  T(0xb0), T(0xb1), T(0xb2), T(0xb3), T(0xb4), T(0xb5), T(0xb6), T(0xb7),
+  T(0xb8), T(0xb9), T(0xba), T(0xbb), T(0xbc), T(0xbd), T(0xbe), T(0xbf),
+  T(0xc0), T(0xc1), T(0xc2), T(0xc3), T(0xc4), T(0xc5), T(0xc6), T(0xc7),
+  T(0xc8), T(0xc9), T(0xca), T(0xcb), T(0xcc), T(0xcd), T(0xce), T(0xcf),
+  T(0xd0), T(0xd1), T(0xd2), T(0xd3), T(0xd4), T(0xd5), T(0xd6), T(0xd7),
+  T(0xd8), T(0xd9), T(0xda), T(0xdb), T(0xdc), T(0xdd), T(0xde), T(0xdf),
+  T(0xe0), T(0xe1), T(0xe2), T(0xe3), T(0xe4), T(0xe5), T(0xe6), T(0xe7),
+  T(0xe8), T(0xe9), T(0xea), T(0xeb), T(0xec), T(0xed), T(0xee), T(0xef),
+  T(0xf0), T(0xf1), T(0xf2), T(0xf3), T(0xf4), T(0xf5), T(0xf6), T(0xf7),
+  T(0xf8), T(0xf9), T(0xfa), T(0xfb), T(0xfc), T(0xfd), T(0xfe), 0x01000100 };
+#undef T
+
+void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    uint32 b = src_argb[0];
+    uint32 g = src_argb[1];
+    uint32 r = src_argb[2];
+    const uint32 a = src_argb[3];
+    const uint32 ia = fixed_invtbl8[a] & 0xffff;  // 8.8 fixed point
+    b = (b * ia) >> 8;
+    g = (g * ia) >> 8;
+    r = (r * ia) >> 8;
+    // Clamping should not be necessary but is free in assembly.
+    dst_argb[0] = clamp255(b);
+    dst_argb[1] = clamp255(g);
+    dst_argb[2] = clamp255(r);
+    dst_argb[3] = a;
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
+                               const int32* previous_cumsum, int width) {
+  int32 row_sum[4] = {0, 0, 0, 0};
+  int x;
+  for (x = 0; x < width; ++x) {
+    row_sum[0] += row[x * 4 + 0];
+    row_sum[1] += row[x * 4 + 1];
+    row_sum[2] += row[x * 4 + 2];
+    row_sum[3] += row[x * 4 + 3];
+    cumsum[x * 4 + 0] = row_sum[0]  + previous_cumsum[x * 4 + 0];
+    cumsum[x * 4 + 1] = row_sum[1]  + previous_cumsum[x * 4 + 1];
+    cumsum[x * 4 + 2] = row_sum[2]  + previous_cumsum[x * 4 + 2];
+    cumsum[x * 4 + 3] = row_sum[3]  + previous_cumsum[x * 4 + 3];
+  }
+}
+
+void CumulativeSumToAverageRow_C(const int32* tl, const int32* bl,
+                                int w, int area, uint8* dst, int count) {
+  float ooa = 1.0f / area;
+  int i;
+  for (i = 0; i < count; ++i) {
+    dst[0] = (uint8)((bl[w + 0] + tl[0] - bl[0] - tl[w + 0]) * ooa);
+    dst[1] = (uint8)((bl[w + 1] + tl[1] - bl[1] - tl[w + 1]) * ooa);
+    dst[2] = (uint8)((bl[w + 2] + tl[2] - bl[2] - tl[w + 2]) * ooa);
+    dst[3] = (uint8)((bl[w + 3] + tl[3] - bl[3] - tl[w + 3]) * ooa);
+    dst += 4;
+    tl += 4;
+    bl += 4;
+  }
+}
+
+// Copy pixels from rotated source to destination row with a slope.
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width) {
+  int i;
+  // Render a row of pixels from source into a buffer.
+  float uv[2];
+  uv[0] = uv_dudv[0];
+  uv[1] = uv_dudv[1];
+  for (i = 0; i < width; ++i) {
+    int x = (int)(uv[0]);
+    int y = (int)(uv[1]);
+    *(uint32*)(dst_argb) =
+        *(const uint32*)(src_argb + y * src_argb_stride +
+                                         x * 4);
+    dst_argb += 4;
+    uv[0] += uv_dudv[2];
+    uv[1] += uv_dudv[3];
+  }
+}
+
+// Blend 2 rows into 1.
+static void HalfRow_C(const uint8* src_uv, int src_uv_stride,
+                      uint8* dst_uv, int pix) {
+  int x;
+  for (x = 0; x < pix; ++x) {
+    dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
+  }
+}
+
+static void HalfRow_16_C(const uint16* src_uv, int src_uv_stride,
+                         uint16* dst_uv, int pix) {
+  int x;
+  for (x = 0; x < pix; ++x) {
+    dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
+  }
+}
+
+// C version 2x2 -> 2x1.
+void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+                      ptrdiff_t src_stride,
+                      int width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  int x;
+  if (source_y_fraction == 0) {
+    memcpy(dst_ptr, src_ptr, width);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_C(src_ptr, (int)(src_stride), dst_ptr, width);
+    return;
+  }
+  for (x = 0; x < width - 1; x += 2) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
+    src_ptr += 2;
+    src_ptr1 += 2;
+    dst_ptr += 2;
+  }
+  if (width & 1) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+  }
+}
+
+void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         ptrdiff_t src_stride,
+                         int width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16* src_ptr1 = src_ptr + src_stride;
+  int x;
+  if (source_y_fraction == 0) {
+    memcpy(dst_ptr, src_ptr, width * 2);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_16_C(src_ptr, (int)(src_stride), dst_ptr, width);
+    return;
+  }
+  for (x = 0; x < width - 1; x += 2) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
+    src_ptr += 2;
+    src_ptr1 += 2;
+    dst_ptr += 2;
+  }
+  if (width & 1) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+  }
+}
+
+// Use first 4 shuffler values to reorder ARGB channels.
+void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
+                      const uint8* shuffler, int pix) {
+  int index0 = shuffler[0];
+  int index1 = shuffler[1];
+  int index2 = shuffler[2];
+  int index3 = shuffler[3];
+  // Shuffle a row of ARGB.
+  int x;
+  for (x = 0; x < pix; ++x) {
+    // To support in-place conversion.
+    uint8 b = src_argb[index0];
+    uint8 g = src_argb[index1];
+    uint8 r = src_argb[index2];
+    uint8 a = src_argb[index3];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = a;
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void I422ToYUY2Row_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_frame, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_frame[0] = src_y[0];
+    dst_frame[1] = src_u[0];
+    dst_frame[2] = src_y[1];
+    dst_frame[3] = src_v[0];
+    dst_frame += 4;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+  }
+  if (width & 1) {
+    dst_frame[0] = src_y[0];
+    dst_frame[1] = src_u[0];
+    dst_frame[2] = 0;
+    dst_frame[3] = src_v[0];
+  }
+}
+
+void I422ToUYVYRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_frame, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_frame[0] = src_u[0];
+    dst_frame[1] = src_y[0];
+    dst_frame[2] = src_v[0];
+    dst_frame[3] = src_y[1];
+    dst_frame += 4;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+  }
+  if (width & 1) {
+    dst_frame[0] = src_u[0];
+    dst_frame[1] = src_y[0];
+    dst_frame[2] = src_v[0];
+    dst_frame[3] = 0;
+  }
+}
+
+// Maximum temporary width for wrappers to process at a time, in pixels.
+#define MAXTWIDTH 2048
+
+#if !(defined(_MSC_VER) && !defined(__clang__)) && \
+    defined(HAS_I422TORGB565ROW_SSSE3)
+// row_win.cc has asm version, but GCC uses 2 step wrapper.
+void I422ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_rgb565,
+                           int width) {
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth);
+    ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TOARGB1555ROW_SSSE3)
+void I422ToARGB1555Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb1555,
+                             int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB1555Row_SSE2(row, dst_argb1555, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb1555 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TOARGB4444ROW_SSSE3)
+void I422ToARGB4444Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb4444,
+                             int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB4444Row_SSE2(row, dst_argb4444, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb4444 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV12TORGB565ROW_SSSE3)
+void NV12ToRGB565Row_SSSE3(const uint8* src_y, const uint8* src_uv,
+                           uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV12ToARGBRow_SSSE3(src_y, src_uv, row, twidth);
+    ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_uv += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV21TORGB565ROW_SSSE3)
+void NV21ToRGB565Row_SSSE3(const uint8* src_y, const uint8* src_vu,
+                           uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV21ToARGBRow_SSSE3(src_y, src_vu, row, twidth);
+    ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_vu += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_YUY2TOARGBROW_SSSE3)
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    YUY2ToUV422Row_SSE2(src_yuy2, row_u, row_v, twidth);
+    YUY2ToYRow_SSE2(src_yuy2, row_y, twidth);
+    I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, twidth);
+    src_yuy2 += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_UYVYTOARGBROW_SSSE3)
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    UYVYToUV422Row_SSE2(src_uyvy, row_u, row_v, twidth);
+    UYVYToYRow_SSE2(src_uyvy, row_y, twidth);
+    I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, twidth);
+    src_uyvy += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif  // !defined(LIBYUV_DISABLE_X86)
+
+#if defined(HAS_I422TORGB565ROW_AVX2)
+void I422ToRGB565Row_AVX2(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TOARGB1555ROW_AVX2)
+void I422ToARGB1555Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB1555Row_AVX2(row, dst_argb1555, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb1555 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TOARGB4444ROW_AVX2)
+void I422ToARGB4444Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB4444Row_AVX2(row, dst_argb4444, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb4444 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TORGB24ROW_AVX2)
+void I422ToRGB24Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_rgb24,
+                            int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    // TODO(fbarchard): ARGBToRGB24Row_AVX2
+    ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_rgb24 += twidth * 3;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TORAWROW_AVX2)
+void I422ToRAWRow_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_raw,
+                            int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    // TODO(fbarchard): ARGBToRAWRow_AVX2
+    ARGBToRAWRow_SSSE3(row, dst_raw, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_raw += twidth * 3;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV12TORGB565ROW_AVX2)
+void NV12ToRGB565Row_AVX2(const uint8* src_y, const uint8* src_uv,
+                          uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV12ToARGBRow_AVX2(src_y, src_uv, row, twidth);
+    ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_uv += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV21TORGB565ROW_AVX2)
+void NV21ToRGB565Row_AVX2(const uint8* src_y, const uint8* src_vu,
+                          uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV21ToARGBRow_AVX2(src_y, src_vu, row, twidth);
+    ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_vu += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+void YUY2ToARGBRow_AVX2(const uint8* src_yuy2, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED32(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED32(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED32(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    YUY2ToUV422Row_AVX2(src_yuy2, row_u, row_v, twidth);
+    YUY2ToYRow_AVX2(src_yuy2, row_y, twidth);
+    I422ToARGBRow_AVX2(row_y, row_u, row_v, dst_argb, twidth);
+    src_yuy2 += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_UYVYTOARGBROW_AVX2)
+void UYVYToARGBRow_AVX2(const uint8* src_uyvy, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED32(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED32(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED32(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    UYVYToUV422Row_AVX2(src_uyvy, row_u, row_v, twidth);
+    UYVYToYRow_AVX2(src_uyvy, row_y, twidth);
+    I422ToARGBRow_AVX2(row_y, row_u, row_v, dst_argb, twidth);
+    src_uyvy += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif  // !defined(LIBYUV_DISABLE_X86)
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    float b = (float)(src_argb[0]);
+    float g = (float)(src_argb[1]);
+    float r = (float)(src_argb[2]);
+    float a = (float)(src_argb[3]);
+    float b2 = b * b;
+    float g2 = g * g;
+    float r2 = r * r;
+    float a2 = a * a;
+    float db = poly[0] + poly[4] * b;
+    float dg = poly[1] + poly[5] * g;
+    float dr = poly[2] + poly[6] * r;
+    float da = poly[3] + poly[7] * a;
+    float b3 = b2 * b;
+    float g3 = g2 * g;
+    float r3 = r2 * r;
+    float a3 = a2 * a;
+    db += poly[8] * b2;
+    dg += poly[9] * g2;
+    dr += poly[10] * r2;
+    da += poly[11] * a2;
+    db += poly[12] * b3;
+    dg += poly[13] * g3;
+    dr += poly[14] * r3;
+    da += poly[15] * a3;
+
+    dst_argb[0] = Clamp((int32)(db));
+    dst_argb[1] = Clamp((int32)(dg));
+    dst_argb[2] = Clamp((int32)(dr));
+    dst_argb[3] = Clamp((int32)(da));
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                             const uint8* luma, uint32 lumacoeff) {
+  uint32 bc = lumacoeff & 0xff;
+  uint32 gc = (lumacoeff >> 8) & 0xff;
+  uint32 rc = (lumacoeff >> 16) & 0xff;
+
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
+                           src_argb[2] * rc) & 0x7F00u) + luma;
+    const uint8* luma1;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+    luma1 = ((src_argb[4] * bc + src_argb[5] * gc +
+              src_argb[6] * rc) & 0x7F00u) + luma;
+    dst_argb[4] = luma1[src_argb[4]];
+    dst_argb[5] = luma1[src_argb[5]];
+    dst_argb[6] = luma1[src_argb[6]];
+    dst_argb[7] = src_argb[7];
+    src_argb += 8;
+    dst_argb += 8;
+  }
+  if (width & 1) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
+                           src_argb[2] * rc) & 0x7F00u) + luma;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+  }
+}
+
+void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst[3] = src[3];
+    dst[7] = src[7];
+    dst += 8;
+    src += 8;
+  }
+  if (width & 1) {
+    dst[3] = src[3];
+  }
+}
+
+void ARGBCopyYToAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst[3] = src[0];
+    dst[7] = src[1];
+    dst += 8;
+    src += 2;
+  }
+  if (width & 1) {
+    dst[3] = src[0];
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/row_gcc.cc b/libvpx/libvpx/third_party/libyuv/source/row_gcc.cc
new file mode 100644
index 0000000..820de0a
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/row_gcc.cc
@@ -0,0 +1,5475 @@
+// VERSION 2
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
+
+// Constants for ARGB
+static vec8 kARGBToY = {
+  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
+};
+
+// JPeg full range.
+static vec8 kARGBToYJ = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+#endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
+
+#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
+
+static vec8 kARGBToU = {
+  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
+};
+
+static vec8 kARGBToUJ = {
+  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
+};
+
+static vec8 kARGBToV = {
+  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+};
+
+static vec8 kARGBToVJ = {
+  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
+};
+
+// Constants for BGRA
+static vec8 kBGRAToY = {
+  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
+};
+
+static vec8 kBGRAToU = {
+  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
+};
+
+static vec8 kBGRAToV = {
+  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
+};
+
+// Constants for ABGR
+static vec8 kABGRToY = {
+  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
+};
+
+static vec8 kABGRToU = {
+  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
+};
+
+static vec8 kABGRToV = {
+  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
+};
+
+// Constants for RGBA.
+static vec8 kRGBAToY = {
+  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
+};
+
+static vec8 kRGBAToU = {
+  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
+};
+
+static vec8 kRGBAToV = {
+  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
+};
+
+static uvec8 kAddY16 = {
+  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
+};
+
+// 7 bit fixed point 0.5.
+static vec16 kAddYJ64 = {
+  64, 64, 64, 64, 64, 64, 64, 64
+};
+
+static uvec8 kAddUV128 = {
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
+};
+
+static uvec16 kAddUVJ128 = {
+  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
+};
+#endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
+
+#ifdef HAS_RGB24TOARGBROW_SSSE3
+
+// Shuffle table for converting RGB24 to ARGB.
+static uvec8 kShuffleMaskRGB24ToARGB = {
+  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
+};
+
+// Shuffle table for converting RAW to ARGB.
+static uvec8 kShuffleMaskRAWToARGB = {
+  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
+};
+
+// Shuffle table for converting ARGB to RGB24.
+static uvec8 kShuffleMaskARGBToRGB24 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static uvec8 kShuffleMaskARGBToRAW = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
+static uvec8 kShuffleMaskARGBToRGB24_0 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static uvec8 kShuffleMaskARGBToRAW_0 = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u
+};
+#endif  // HAS_RGB24TOARGBROW_SSSE3
+
+#if defined(TESTING) && defined(__x86_64__)
+void TestRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  asm volatile (
+    ".p2align  5                               \n"
+    "mov       %%eax,%%eax                     \n"
+    "mov       %%ebx,%%ebx                     \n"
+    "mov       %%ecx,%%ecx                     \n"
+    "mov       %%edx,%%edx                     \n"
+    "mov       %%esi,%%esi                     \n"
+    "mov       %%edi,%%edi                     \n"
+    "mov       %%ebp,%%ebp                     \n"
+    "mov       %%esp,%%esp                     \n"
+    ".p2align  5                               \n"
+    "mov       %%r8d,%%r8d                     \n"
+    "mov       %%r9d,%%r9d                     \n"
+    "mov       %%r10d,%%r10d                   \n"
+    "mov       %%r11d,%%r11d                   \n"
+    "mov       %%r12d,%%r12d                   \n"
+    "mov       %%r13d,%%r13d                   \n"
+    "mov       %%r14d,%%r14d                   \n"
+    "mov       %%r15d,%%r15d                   \n"
+    ".p2align  5                               \n"
+    "lea       (%%rax),%%eax                   \n"
+    "lea       (%%rbx),%%ebx                   \n"
+    "lea       (%%rcx),%%ecx                   \n"
+    "lea       (%%rdx),%%edx                   \n"
+    "lea       (%%rsi),%%esi                   \n"
+    "lea       (%%rdi),%%edi                   \n"
+    "lea       (%%rbp),%%ebp                   \n"
+    "lea       (%%rsp),%%esp                   \n"
+    ".p2align  5                               \n"
+    "lea       (%%r8),%%r8d                    \n"
+    "lea       (%%r9),%%r9d                    \n"
+    "lea       (%%r10),%%r10d                  \n"
+    "lea       (%%r11),%%r11d                  \n"
+    "lea       (%%r12),%%r12d                  \n"
+    "lea       (%%r13),%%r13d                  \n"
+    "lea       (%%r14),%%r14d                  \n"
+    "lea       (%%r15),%%r15d                  \n"
+
+    ".p2align  5                               \n"
+    "lea       0x10(%%rax),%%eax               \n"
+    "lea       0x10(%%rbx),%%ebx               \n"
+    "lea       0x10(%%rcx),%%ecx               \n"
+    "lea       0x10(%%rdx),%%edx               \n"
+    "lea       0x10(%%rsi),%%esi               \n"
+    "lea       0x10(%%rdi),%%edi               \n"
+    "lea       0x10(%%rbp),%%ebp               \n"
+    "lea       0x10(%%rsp),%%esp               \n"
+    ".p2align  5                               \n"
+    "lea       0x10(%%r8),%%r8d                \n"
+    "lea       0x10(%%r9),%%r9d                \n"
+    "lea       0x10(%%r10),%%r10d              \n"
+    "lea       0x10(%%r11),%%r11d              \n"
+    "lea       0x10(%%r12),%%r12d              \n"
+    "lea       0x10(%%r13),%%r13d              \n"
+    "lea       0x10(%%r14),%%r14d              \n"
+    "lea       0x10(%%r15),%%r15d              \n"
+
+    ".p2align  5                               \n"
+    "add       0x10,%%eax                      \n"
+    "add       0x10,%%ebx                      \n"
+    "add       0x10,%%ecx                      \n"
+    "add       0x10,%%edx                      \n"
+    "add       0x10,%%esi                      \n"
+    "add       0x10,%%edi                      \n"
+    "add       0x10,%%ebp                      \n"
+    "add       0x10,%%esp                      \n"
+    ".p2align  5                               \n"
+    "add       0x10,%%r8d                      \n"
+    "add       0x10,%%r9d                      \n"
+    "add       0x10,%%r10d                     \n"
+    "add       0x10,%%r11d                     \n"
+    "add       0x10,%%r12d                     \n"
+    "add       0x10,%%r13d                     \n"
+    "add       0x10,%%r14d                     \n"
+    "add       0x10,%%r15d                     \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc", "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // TESTING
+
+#ifdef HAS_J400TOARGBROW_SSE2
+void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_J400TOARGBROW_SSE2
+
+#ifdef HAS_RGB24TOARGBROW_SSSE3
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
+    "pslld     $0x18,%%xmm5                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x30,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm2                   \n"
+    "palignr   $0x8,%%xmm1,%%xmm2              \n"
+    "pshufb    %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "palignr   $0x4,%%xmm3,%%xmm3              \n"
+    "pshufb    %%xmm4,%%xmm3                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "m"(kShuffleMaskRGB24ToARGB)  // %3
+  : "memory", "cc" , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
+    "pslld     $0x18,%%xmm5                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x30,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm2                   \n"
+    "palignr   $0x8,%%xmm1,%%xmm2              \n"
+    "pshufb    %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "palignr   $0x4,%%xmm3,%%xmm3              \n"
+    "pshufb    %%xmm4,%%xmm3                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "m"(kShuffleMaskRAWToARGB)  // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void RGB565ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0x1080108,%%eax                \n"
+    "movd      %%eax,%%xmm5                    \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "mov       $0x20802080,%%eax               \n"
+    "movd      %%eax,%%xmm6                    \n"
+    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psllw     $0xb,%%xmm3                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0xa,%%xmm4                     \n"
+    "psrlw     $0x5,%%xmm4                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psllw     $0x8,%%xmm7                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "psllw     $0xb,%%xmm2                     \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "psllw     $0x8,%%xmm1                     \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "pmulhuw   %%xmm6,%%xmm0                   \n"
+    "por       %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpckhbw %%xmm0,%%xmm2                   \n"
+    MEMOPMEM(movdqu,xmm1,0x00,1,0,2)           //  movdqu  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqu,xmm2,0x10,1,0,2)           //  movdqu  %%xmm2,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void ARGB1555ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0x1080108,%%eax                \n"
+    "movd      %%eax,%%xmm5                    \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "mov       $0x42004200,%%eax               \n"
+    "movd      %%eax,%%xmm6                    \n"
+    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psllw     $0xb,%%xmm3                     \n"
+    "movdqa    %%xmm3,%%xmm4                   \n"
+    "psrlw     $0x6,%%xmm4                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psllw     $0x8,%%xmm7                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psllw     $0x1,%%xmm1                     \n"
+    "psllw     $0xb,%%xmm2                     \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "psllw     $0x8,%%xmm1                     \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "pmulhuw   %%xmm6,%%xmm0                   \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpckhbw %%xmm0,%%xmm2                   \n"
+    MEMOPMEM(movdqu,xmm1,0x00,1,0,2)           //  movdqu  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqu,xmm2,0x10,1,0,2)           //  movdqu  %%xmm2,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void ARGB4444ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0xf0f0f0f,%%eax                \n"
+    "movd      %%eax,%%xmm4                    \n"
+    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "pslld     $0x4,%%xmm5                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "psllw     $0x4,%%xmm1                     \n"
+    "psrlw     $0x4,%%xmm3                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,2)           //  movdqu  %%xmm0,(%1,%0,2)
+    MEMOPMEM(movdqu,xmm1,0x10,1,0,2)           //  movdqu  %%xmm1,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void ARGBToRGB24Row_SSSE3(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "pshufb    %%xmm6,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm6,%%xmm2                   \n"
+    "pshufb    %%xmm6,%%xmm3                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "psrldq    $0x4,%%xmm1                     \n"
+    "pslldq    $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm2,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pslldq    $0x8,%%xmm5                     \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "psrldq    $0x8,%%xmm2                     \n"
+    "pslldq    $0x4,%%xmm3                     \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x30,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "m"(kShuffleMaskARGBToRGB24)  // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+void ARGBToRAWRow_SSSE3(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "pshufb    %%xmm6,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm6,%%xmm2                   \n"
+    "pshufb    %%xmm6,%%xmm3                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "psrldq    $0x4,%%xmm1                     \n"
+    "pslldq    $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm2,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pslldq    $0x8,%%xmm5                     \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "psrldq    $0x8,%%xmm2                     \n"
+    "pslldq    $0x4,%%xmm3                     \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x30,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "m"(kShuffleMaskARGBToRAW)  // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+void ARGBToRGB565Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psrld     $0x1b,%%xmm3                    \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psrld     $0x1a,%%xmm4                    \n"
+    "pslld     $0x5,%%xmm4                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0xb,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pslld     $0x8,%%xmm0                     \n"
+    "psrld     $0x3,%%xmm1                     \n"
+    "psrld     $0x5,%%xmm2                     \n"
+    "psrad     $0x10,%%xmm0                    \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm4,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psrld     $0x1b,%%xmm4                    \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "pslld     $0x5,%%xmm5                     \n"
+    "movdqa    %%xmm4,%%xmm6                   \n"
+    "pslld     $0xa,%%xmm6                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "pslld     $0xf,%%xmm7                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "psrad     $0x10,%%xmm0                    \n"
+    "psrld     $0x3,%%xmm1                     \n"
+    "psrld     $0x6,%%xmm2                     \n"
+    "psrld     $0x9,%%xmm3                     \n"
+    "pand      %%xmm7,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm6,%%xmm3                   \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :: "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void ARGBToARGB4444Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm4,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm3,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm1                   \n"
+    "psrlq     $0x4,%%xmm0                     \n"
+    "psrlq     $0x8,%%xmm1                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+#endif  // HAS_RGB24TOARGBROW_SSSE3
+
+#ifdef HAS_ARGBTOYROW_SSSE3
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBTOYROW_SSSE3
+
+#ifdef HAS_ARGBTOYJROW_SSSE3
+// Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
+// Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),  // %3
+    "m"(kAddYJ64)    // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBTOYJROW_SSSE3
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// vpermd for vphaddw + vpackuswb vpermd.
+static const lvec32 kPermdARGBToY_AVX = {
+  0, 4, 1, 5, 2, 6, 3, 7
+};
+
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vbroadcastf128 %3,%%ymm4                  \n"
+    "vbroadcastf128 %4,%%ymm5                  \n"
+    "vmovdqu    %5,%%ymm6                      \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
+    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
+    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
+    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
+    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
+    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
+    "lea       " MEMLEA(0x80,0) ",%0           \n"
+    "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"  // mutates.
+    "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"
+    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x7,%%ymm2,%%ymm2             \n"
+    "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
+    "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
+    "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"  // add 16 for Y
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16),    // %4
+    "m"(kPermdARGBToY_AVX)  // %5
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBTOYROW_AVX2
+
+#ifdef HAS_ARGBTOYJROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vbroadcastf128 %3,%%ymm4                  \n"
+    "vbroadcastf128 %4,%%ymm5                  \n"
+    "vmovdqu    %5,%%ymm6                      \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
+    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
+    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
+    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
+    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
+    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
+    "lea       " MEMLEA(0x80,0) ",%0           \n"
+    "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"  // mutates.
+    "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"
+    "vpaddw     %%ymm5,%%ymm0,%%ymm0           \n"  // Add .5 for rounding.
+    "vpaddw     %%ymm5,%%ymm2,%%ymm2           \n"
+    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x7,%%ymm2,%%ymm2             \n"
+    "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
+    "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),   // %3
+    "m"(kAddYJ64),    // %4
+    "m"(kPermdARGBToY_AVX)  // %5
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBTOYJROW_AVX2
+
+#ifdef HAS_ARGBTOUVROW_SSSE3
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps    %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)), // %4
+    "m"(kARGBToV),  // %5
+    "m"(kARGBToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBTOUVROW_SSSE3
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+// vpshufb for vphaddw + vpackuswb packed to shorts.
+static const lvec8 kShufARGBToUV_AVX = {
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15
+};
+void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "vbroadcastf128 %5,%%ymm5                  \n"
+    "vbroadcastf128 %6,%%ymm6                  \n"
+    "vbroadcastf128 %7,%%ymm7                  \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
+    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
+    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
+    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
+    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
+    VMEMOPREG(vpavgb,0x40,0,4,1,ymm2,ymm2)
+    VMEMOPREG(vpavgb,0x60,0,4,1,ymm3,ymm3)
+    "lea       " MEMLEA(0x80,0) ",%0           \n"
+    "vshufps    $0x88,%%ymm1,%%ymm0,%%ymm4     \n"
+    "vshufps    $0xdd,%%ymm1,%%ymm0,%%ymm0     \n"
+    "vpavgb     %%ymm4,%%ymm0,%%ymm0           \n"
+    "vshufps    $0x88,%%ymm3,%%ymm2,%%ymm4     \n"
+    "vshufps    $0xdd,%%ymm3,%%ymm2,%%ymm2     \n"
+    "vpavgb     %%ymm4,%%ymm2,%%ymm2           \n"
+
+    "vpmaddubsw %%ymm7,%%ymm0,%%ymm1           \n"
+    "vpmaddubsw %%ymm7,%%ymm2,%%ymm3           \n"
+    "vpmaddubsw %%ymm6,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm6,%%ymm2,%%ymm2           \n"
+    "vphaddw    %%ymm3,%%ymm1,%%ymm1           \n"
+    "vphaddw    %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpsraw     $0x8,%%ymm1,%%ymm1             \n"
+    "vpsraw     $0x8,%%ymm0,%%ymm0             \n"
+    "vpacksswb  %%ymm0,%%ymm1,%%ymm0           \n"
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpshufb    %8,%%ymm0,%%ymm0               \n"
+    "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"
+
+    "vextractf128 $0x0,%%ymm0," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,1,ymm0,0x0,1,2,1) // vextractf128 $1,%%ymm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x20,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)), // %4
+    "m"(kAddUV128),  // %5
+    "m"(kARGBToV),   // %6
+    "m"(kARGBToU),   // %7
+    "m"(kShufARGBToUV_AVX)  // %8
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBTOUVROW_AVX2
+
+#ifdef HAS_ARGBTOUVJROW_SSSE3
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)), // %4
+    "m"(kARGBToVJ),  // %5
+    "m"(kARGBToUJ),  // %6
+    "m"(kAddUVJ128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBTOUVJROW_SSSE3
+
+#ifdef HAS_ARGBTOUV444ROW_SSSE3
+void ARGBToUV444Row_SSSE3(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                          int width) {
+  asm volatile (
+    "movdqa    %4,%%xmm3                       \n"
+    "movdqa    %5,%%xmm4                       \n"
+    "movdqa    %6,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,2,1)           //  movdqu  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),        // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "m"(kARGBToV),  // %4
+    "m"(kARGBToU),  // %5
+    "m"(kAddUV128)  // %6
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6"
+  );
+}
+#endif  // HAS_ARGBTOUV444ROW_SSSE3
+
+#ifdef HAS_ARGBTOUV422ROW_SSSE3
+void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %4,%%xmm3                       \n"
+    "movdqa    %5,%%xmm4                       \n"
+    "movdqa    %6,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "m"(kARGBToV),  // %4
+    "m"(kARGBToU),  // %5
+    "m"(kAddUV128)  // %6
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBTOUV422ROW_SSSE3
+
+void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kBGRAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void BGRAToUVRow_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_bgra)), // %4
+    "m"(kBGRAToV),  // %5
+    "m"(kBGRAToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+
+void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kABGRToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kRGBAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void ABGRToUVRow_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_abgr)), // %4
+    "m"(kABGRToV),  // %5
+    "m"(kABGRToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+
+void RGBAToUVRow_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_rgba)), // %4
+    "m"(kRGBAToV),  // %5
+    "m"(kRGBAToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+
+#if defined(HAS_I422TOARGBROW_SSSE3) || defined(HAS_I422TOARGBROW_AVX2)
+
+struct YuvConstants {
+  lvec8 kUVToB;     // 0
+  lvec8 kUVToG;     // 32
+  lvec8 kUVToR;     // 64
+  lvec16 kUVBiasB;  // 96
+  lvec16 kUVBiasG;  // 128
+  lvec16 kUVBiasR;  // 160
+  lvec16 kYToRgb;   // 192
+};
+
+// BT.601 YUV to RGB reference
+//  R = (Y - 16) * 1.164              - V * -1.596
+//  G = (Y - 16) * 1.164 - U *  0.391 - V *  0.813
+//  B = (Y - 16) * 1.164 - U * -2.018
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* max(-128, round(-2.018 * 64)) */
+#define UG 25 /* round(0.391 * 64) */
+#define VG 52 /* round(0.813 * 64) */
+#define VR -102 /* round(-1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            + YGB)
+#define BG (UG * 128 + VG * 128 + YGB)
+#define BR            (VR * 128 + YGB)
+
+// BT601 constants for YUV to RGB.
+static YuvConstants SIMD_ALIGNED(kYuvConstants) = {
+  { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
+    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 },
+  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
+    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
+  { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
+    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
+};
+
+// BT601 constants for NV21 where chroma plane is VU instead of UV.
+static YuvConstants SIMD_ALIGNED(kYvuConstants) = {
+  { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
+    0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB },
+  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
+    VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
+  { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
+    VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
+};
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+// JPEG YUV to RGB reference
+// *  R = Y                - V * -1.40200
+// *  G = Y - U *  0.34414 - V *  0.71414
+// *  B = Y - U * -1.77200
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YGJ 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+#define YGBJ 32  /* 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UBJ -113 /* round(-1.77200 * 64) */
+#define UGJ 22 /* round(0.34414 * 64) */
+#define VGJ 46 /* round(0.71414  * 64) */
+#define VRJ -90 /* round(-1.40200 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BBJ (UBJ * 128             + YGBJ)
+#define BGJ (UGJ * 128 + VGJ * 128 + YGBJ)
+#define BRJ             (VRJ * 128 + YGBJ)
+
+// JPEG constants for YUV to RGB.
+YuvConstants SIMD_ALIGNED(kYuvJConstants) = {
+  { UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0,
+    UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0 },
+  { UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ },
+  { 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ,
+    0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ },
+  { BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ,
+    BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ },
+  { BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ,
+    BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ },
+  { BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ,
+    BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ },
+  { YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ,
+    YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ }
+};
+
+#undef YGJ
+#undef YGBJ
+#undef UBJ
+#undef UGJ
+#undef VGJ
+#undef VRJ
+#undef BBJ
+#undef BGJ
+#undef BRJ
+
+// Read 8 UV from 411
+#define READYUV444                                                             \
+    "movq       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    MEMOPREG(movq, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"
+
+// Read 4 UV from 422, upsample to 8 UV
+#define READYUV422                                                             \
+    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x4, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"
+
+// Read 2 UV from 411, upsample to 8 UV
+#define READYUV411                                                             \
+    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x2, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"            \
+    "punpckldq  %%xmm0,%%xmm0                                   \n"
+
+// Read 4 UV from NV12, upsample to 8 UV
+#define READNV12                                                               \
+    "movq       " MEMACCESS([uv_buf]) ",%%xmm0                  \n"            \
+    "lea        " MEMLEA(0x8, [uv_buf]) ",%[uv_buf]             \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"
+
+// Convert 8 pixels: 8 UV and 8 Y
+#define YUVTORGB(YuvConstants)                                                 \
+    "movdqa     %%xmm0,%%xmm1                                   \n"            \
+    "movdqa     %%xmm0,%%xmm2                                   \n"            \
+    "movdqa     %%xmm0,%%xmm3                                   \n"            \
+    "movdqa     " MEMACCESS2(96, [YuvConstants]) ",%%xmm0       \n"            \
+    "pmaddubsw  " MEMACCESS([YuvConstants]) ",%%xmm1            \n"            \
+    "psubw      %%xmm1,%%xmm0                                   \n"            \
+    "movdqa     " MEMACCESS2(128, [YuvConstants]) ",%%xmm1      \n"            \
+    "pmaddubsw  " MEMACCESS2(32, [YuvConstants]) ",%%xmm2       \n"            \
+    "psubw      %%xmm2,%%xmm1                                   \n"            \
+    "movdqa     " MEMACCESS2(160, [YuvConstants]) ",%%xmm2      \n"            \
+    "pmaddubsw  " MEMACCESS2(64, [YuvConstants]) ",%%xmm3       \n"            \
+    "psubw      %%xmm3,%%xmm2                                   \n"            \
+    "movq       " MEMACCESS([y_buf]) ",%%xmm3                   \n"            \
+    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
+    "punpcklbw  %%xmm3,%%xmm3                                   \n"            \
+    "pmulhuw    " MEMACCESS2(192, [YuvConstants]) ",%%xmm3      \n"            \
+    "paddsw     %%xmm3,%%xmm0                                   \n"            \
+    "paddsw     %%xmm3,%%xmm1                                   \n"            \
+    "paddsw     %%xmm3,%%xmm2                                   \n"            \
+    "psraw      $0x6,%%xmm0                                     \n"            \
+    "psraw      $0x6,%%xmm1                                     \n"            \
+    "psraw      $0x6,%%xmm2                                     \n"            \
+    "packuswb   %%xmm0,%%xmm0                                   \n"            \
+    "packuswb   %%xmm1,%%xmm1                                   \n"            \
+    "packuswb   %%xmm2,%%xmm2                                   \n"
+
+// Store 8 ARGB values. Assumes XMM5 is zero.
+#define STOREARGB                                                              \
+    "punpcklbw  %%xmm1,%%xmm0                                    \n"           \
+    "punpcklbw  %%xmm5,%%xmm2                                    \n"           \
+    "movdqa     %%xmm0,%%xmm1                                    \n"           \
+    "punpcklwd  %%xmm2,%%xmm0                                    \n"           \
+    "punpckhwd  %%xmm2,%%xmm1                                    \n"           \
+    "movdqu     %%xmm0," MEMACCESS([dst_argb]) "                 \n"           \
+    "movdqu     %%xmm1," MEMACCESS2(0x10, [dst_argb]) "          \n"           \
+    "lea        " MEMLEA(0x20, [dst_argb]) ", %[dst_argb]        \n"
+
+// Store 8 BGRA values. Assumes XMM5 is zero.
+#define STOREBGRA                                                              \
+    "pcmpeqb   %%xmm5,%%xmm5                                     \n"           \
+    "punpcklbw %%xmm0,%%xmm1                                     \n"           \
+    "punpcklbw %%xmm2,%%xmm5                                     \n"           \
+    "movdqa    %%xmm5,%%xmm0                                     \n"           \
+    "punpcklwd %%xmm1,%%xmm5                                     \n"           \
+    "punpckhwd %%xmm1,%%xmm0                                     \n"           \
+    "movdqu    %%xmm5," MEMACCESS([dst_bgra]) "                  \n"           \
+    "movdqu    %%xmm0," MEMACCESS2(0x10, [dst_bgra]) "           \n"           \
+    "lea       " MEMLEA(0x20, [dst_bgra]) ", %[dst_bgra]         \n"
+
+// Store 8 ABGR values. Assumes XMM5 is zero.
+#define STOREABGR                                                              \
+    "punpcklbw %%xmm1,%%xmm2                                     \n"           \
+    "punpcklbw %%xmm5,%%xmm0                                     \n"           \
+    "movdqa    %%xmm2,%%xmm1                                     \n"           \
+    "punpcklwd %%xmm0,%%xmm2                                     \n"           \
+    "punpckhwd %%xmm0,%%xmm1                                     \n"           \
+    "movdqu    %%xmm2," MEMACCESS([dst_abgr]) "                  \n"           \
+    "movdqu    %%xmm1," MEMACCESS2(0x10, [dst_abgr]) "           \n"           \
+    "lea       " MEMLEA(0x20, [dst_abgr]) ", %[dst_abgr]         \n"
+
+// Store 8 RGBA values. Assumes XMM5 is zero.
+#define STORERGBA                                                              \
+    "pcmpeqb   %%xmm5,%%xmm5                                     \n"           \
+    "punpcklbw %%xmm2,%%xmm1                                     \n"           \
+    "punpcklbw %%xmm0,%%xmm5                                     \n"           \
+    "movdqa    %%xmm5,%%xmm0                                     \n"           \
+    "punpcklwd %%xmm1,%%xmm5                                     \n"           \
+    "punpckhwd %%xmm1,%%xmm0                                     \n"           \
+    "movdqu    %%xmm5," MEMACCESS([dst_rgba]) "                  \n"           \
+    "movdqu    %%xmm0," MEMACCESS2(0x10, [dst_rgba]) "           \n"           \
+    "lea       " MEMLEA(0x20, [dst_rgba]) ",%[dst_rgba]          \n"
+
+void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV444
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+// TODO(fbarchard): Consider putting masks into constants.
+void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf,
+                                 const uint8* u_buf,
+                                 const uint8* v_buf,
+                                 uint8* dst_rgb24,
+                                 int width) {
+  asm volatile (
+    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
+    "sub       %[u_buf],%[v_buf]               \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "movq      %%xmm0," MEMACCESS([dst_rgb24]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_rgb24]) "\n"
+    "lea       " MEMLEA(0x18,[dst_rgb24]) ",%[dst_rgb24] \n"
+    "subl      $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgb24]"+r"(dst_rgb24),  // %[dst_rgb24]
+// TODO(fbarchard): Make width a register for 32 bit.
+#if defined(__i386__) && defined(__pic__)
+    [width]"+m"(width)     // %[width]
+#else
+    [width]"+rm"(width)    // %[width]
+#endif
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB),
+    [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
+    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24)
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6"
+  );
+}
+
+void OMITFP I422ToRAWRow_SSSE3(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_raw,
+                               int width) {
+  asm volatile (
+    "movdqa    %[kShuffleMaskARGBToRAW_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRAW],%%xmm6   \n"
+    "sub       %[u_buf],%[v_buf]               \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "movq      %%xmm0," MEMACCESS([dst_raw]) " \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_raw]) "\n"
+    "lea       " MEMLEA(0x18,[dst_raw]) ",%[dst_raw] \n"
+    "subl      $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_raw]"+r"(dst_raw),  // %[dst_raw]
+// TODO(fbarchard): Make width a register for 32 bit.
+#if defined(__i386__) && defined(__pic__)
+    [width]"+m"(width)    // %[width]
+#else
+    [width]"+rm"(width)    // %[width]
+#endif
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB),
+    [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
+    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW)
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6"
+  );
+}
+
+void OMITFP I422ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP J422ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvJConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP I411ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV411
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* uv_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  // Does not use r14.
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP NV21ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* uv_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYvuConstants.kUVToB) // %[kYuvConstants]
+  // Does not use r14.
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP I422ToBGRARow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_bgra,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREBGRA
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP I422ToABGRRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_abgr,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREABGR
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_rgba,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STORERGBA
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+// Read 8 UV from 422, upsample to 16 UV.
+#define READYUV422_AVX2                                                        \
+    "vmovq       " MEMACCESS([u_buf]) ",%%xmm0                      \n"        \
+    MEMOPREG(vmovq, 0x00, [u_buf], [v_buf], 1, xmm1)                           \
+    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]                   \n"        \
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n"        \
+    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
+    "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n"
+
+// Convert 16 pixels: 16 UV and 16 Y.
+#define YUVTORGB_AVX2(YuvConstants)                                            \
+    "vpmaddubsw  " MEMACCESS2(64, [YuvConstants]) ",%%ymm0,%%ymm2   \n"        \
+    "vpmaddubsw  " MEMACCESS2(32, [YuvConstants]) ",%%ymm0,%%ymm1   \n"        \
+    "vpmaddubsw  " MEMACCESS([YuvConstants]) ",%%ymm0,%%ymm0        \n"        \
+    "vmovdqu     " MEMACCESS2(160, [YuvConstants]) ",%%ymm3         \n"        \
+    "vpsubw      %%ymm2,%%ymm3,%%ymm2                               \n"        \
+    "vmovdqu     " MEMACCESS2(128, [YuvConstants]) ",%%ymm3         \n"        \
+    "vpsubw      %%ymm1,%%ymm3,%%ymm1                               \n"        \
+    "vmovdqu     " MEMACCESS2(96, [YuvConstants]) ",%%ymm3          \n"        \
+    "vpsubw      %%ymm0,%%ymm3,%%ymm0                               \n"        \
+    "vmovdqu     " MEMACCESS([y_buf]) ",%%xmm3                      \n"        \
+    "lea         " MEMLEA(0x10, [y_buf]) ",%[y_buf]                 \n"        \
+    "vpermq      $0xd8,%%ymm3,%%ymm3                                \n"        \
+    "vpunpcklbw  %%ymm3,%%ymm3,%%ymm3                               \n"        \
+    "vpmulhuw    " MEMACCESS2(192, [YuvConstants]) ",%%ymm3,%%ymm3  \n"        \
+    "vpaddsw     %%ymm3,%%ymm0,%%ymm0           \n"                            \
+    "vpaddsw     %%ymm3,%%ymm1,%%ymm1           \n"                            \
+    "vpaddsw     %%ymm3,%%ymm2,%%ymm2           \n"                            \
+    "vpsraw      $0x6,%%ymm0,%%ymm0             \n"                            \
+    "vpsraw      $0x6,%%ymm1,%%ymm1             \n"                            \
+    "vpsraw      $0x6,%%ymm2,%%ymm2             \n"                            \
+    "vpackuswb   %%ymm0,%%ymm0,%%ymm0           \n"                            \
+    "vpackuswb   %%ymm1,%%ymm1,%%ymm1           \n"                            \
+    "vpackuswb   %%ymm2,%%ymm2,%%ymm2           \n"
+
+#if defined(HAS_I422TOBGRAROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 BGRA (64 bytes).
+void OMITFP I422ToBGRARow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_bgra,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into BGRA
+    "vpunpcklbw %%ymm0,%%ymm1,%%ymm1           \n"  // GB
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklbw %%ymm2,%%ymm5,%%ymm2           \n"  // AR
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm1,%%ymm2,%%ymm0           \n"  // ARGB first 8 pixels
+    "vpunpckhwd %%ymm1,%%ymm2,%%ymm2           \n"  // ARGB next 8 pixels
+
+    "vmovdqu    %%ymm0," MEMACCESS([dst_bgra]) "\n"
+    "vmovdqu    %%ymm2," MEMACCESS2(0x20,[dst_bgra]) "\n"
+    "lea       " MEMLEA(0x40,[dst_bgra]) ",%[dst_bgra] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_I422TOBGRAROW_AVX2
+
+#if defined(HAS_I422TOARGBROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I422ToARGBRow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ARGB
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0           \n"  // BG
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpunpcklbw %%ymm5,%%ymm2,%%ymm2           \n"  // RA
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm2,%%ymm0,%%ymm1           \n"  // BGRA first 8 pixels
+    "vpunpckhwd %%ymm2,%%ymm0,%%ymm0           \n"  // BGRA next 8 pixels
+
+    "vmovdqu    %%ymm1," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm0," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_I422TOARGBROW_AVX2
+
+#if defined(HAS_J422TOARGBROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP J422ToARGBRow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ARGB
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0           \n"  // BG
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpunpcklbw %%ymm5,%%ymm2,%%ymm2           \n"  // RA
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm2,%%ymm0,%%ymm1           \n"  // BGRA first 8 pixels
+    "vpunpckhwd %%ymm2,%%ymm0,%%ymm0           \n"  // BGRA next 8 pixels
+
+    "vmovdqu    %%ymm1," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm0," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvJConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_J422TOARGBROW_AVX2
+
+#if defined(HAS_I422TOABGRROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ABGR (64 bytes).
+void OMITFP I422ToABGRRow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ABGR
+    "vpunpcklbw %%ymm1,%%ymm2,%%ymm1           \n"  // RG
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklbw %%ymm5,%%ymm0,%%ymm2           \n"  // BA
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm2,%%ymm1,%%ymm0           \n"  // RGBA first 8 pixels
+    "vpunpckhwd %%ymm2,%%ymm1,%%ymm1           \n"  // RGBA next 8 pixels
+    "vmovdqu    %%ymm0," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm1," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_I422TOABGRROW_AVX2
+
+#if defined(HAS_I422TORGBAROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
+void OMITFP I422ToRGBARow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into RGBA
+    "vpunpcklbw %%ymm2,%%ymm1,%%ymm1           \n"
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklbw %%ymm0,%%ymm5,%%ymm2           \n"
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm1,%%ymm2,%%ymm0           \n"
+    "vpunpckhwd %%ymm1,%%ymm2,%%ymm1           \n"
+    "vmovdqu    %%ymm0," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm1," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_I422TORGBAROW_AVX2
+
+#ifdef HAS_I400TOARGBROW_SSE2
+void I400ToARGBRow_SSE2(const uint8* y_buf, uint8* dst_argb, int width) {
+  asm volatile (
+    "mov       $0x4a354a35,%%eax               \n"  // 4a35 = 18997 = 1.164
+    "movd      %%eax,%%xmm2                    \n"
+    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    "mov       $0x04880488,%%eax               \n"  // 0488 = 1160 = 1.164 * 16
+    "movd      %%eax,%%xmm3                    \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "psubusw   %%xmm3,%%xmm0                   \n"
+    "psrlw     $6, %%xmm0                      \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+
+    // Step 2: Weave into ARGB
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "por       %%xmm4,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(y_buf),     // %0
+    "+r"(dst_argb),  // %1
+    "+rm"(width)     // %2
+  :
+  : "memory", "cc", "eax"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+#endif  // HAS_I400TOARGBROW_SSE2
+
+#ifdef HAS_I400TOARGBROW_AVX2
+// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
+// note: vpunpcklbw mutates and vpackuswb unmutates.
+void I400ToARGBRow_AVX2(const uint8* y_buf, uint8* dst_argb, int width) {
+  asm volatile (
+    "mov        $0x4a354a35,%%eax              \n" // 0488 = 1160 = 1.164 * 16
+    "vmovd      %%eax,%%xmm2                   \n"
+    "vbroadcastss %%xmm2,%%ymm2                \n"
+    "mov        $0x4880488,%%eax               \n" // 4a35 = 18997 = 1.164
+    "vmovd      %%eax,%%xmm3                   \n"
+    "vbroadcastss %%xmm3,%%ymm3                \n"
+    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
+    "vpslld     $0x18,%%ymm4,%%ymm4            \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    // Step 1: Scale Y contribution to 16 G values. G = (y - 16) * 1.164
+    "vmovdqu    " MEMACCESS(0) ",%%xmm0        \n"
+    "lea        " MEMLEA(0x10,0) ",%0          \n"
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpunpcklbw %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpsubusw   %%ymm3,%%ymm0,%%ymm0           \n"
+    "vpsrlw     $0x6,%%ymm0,%%ymm0             \n"
+    "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpunpcklbw %%ymm0,%%ymm0,%%ymm1           \n"
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklwd %%ymm1,%%ymm1,%%ymm0           \n"
+    "vpunpckhwd %%ymm1,%%ymm1,%%ymm1           \n"
+    "vpor       %%ymm4,%%ymm0,%%ymm0           \n"
+    "vpor       %%ymm4,%%ymm1,%%ymm1           \n"
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "vmovdqu    %%ymm1," MEMACCESS2(0x20,1) "  \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub        $0x10,%2                       \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(y_buf),     // %0
+    "+r"(dst_argb),  // %1
+    "+rm"(width)     // %2
+  :
+  : "memory", "cc", "eax"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+#endif  // HAS_I400TOARGBROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static uvec8 kShuffleMirror = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "movdqa    %3,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(movdqu,-0x10,0,2,1,xmm0)          //  movdqu -0x10(%0,%2),%%xmm0
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kShuffleMirror) // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm5"
+  );
+}
+#endif  // HAS_MIRRORROW_SSSE3
+
+#ifdef HAS_MIRRORROW_AVX2
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "vbroadcastf128 %3,%%ymm5                  \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(vmovdqu,-0x20,0,2,1,ymm0)         //  vmovdqu -0x20(%0,%2),%%ymm0
+    "vpshufb    %%ymm5,%%ymm0,%%ymm0           \n"
+    "vpermq     $0x4e,%%ymm0,%%ymm0            \n"
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kShuffleMirror) // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm5"
+  );
+}
+#endif  // HAS_MIRRORROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSE2
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(movdqu,-0x10,0,2,1,xmm0)          //  movdqu -0x10(%0,%2),%%xmm0
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "psllw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufd    $0x4e,%%xmm0,%%xmm0             \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1)",%1            \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_MIRRORROW_SSE2
+
+#ifdef HAS_MIRRORROW_UV_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static uvec8 kShuffleMirrorUV = {
+  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
+};
+void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
+                       int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "movdqa    %4,%%xmm1                       \n"
+    "lea       " MEMLEA4(-0x10,0,3,2) ",%0     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    "pshufb    %%xmm1,%%xmm0                   \n"
+    "movlpd    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhpd,xmm0,0x00,1,2,1)           //  movhpd    %%xmm0,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $8,%3                           \n"
+    "jg        1b                              \n"
+  : "+r"(src),      // %0
+    "+r"(dst_u),    // %1
+    "+r"(dst_v),    // %2
+    "+r"(temp_width)  // %3
+  : "m"(kShuffleMirrorUV)  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_MIRRORROW_UV_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_SSE2
+
+void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "lea       " MEMLEA4(-0x10,0,2,4) ",%0     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufd    $0x1b,%%xmm0,%%xmm0             \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0"
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_SSE2
+
+#ifdef HAS_ARGBMIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec32 kARGBShuffleMirror_AVX2 = {
+  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "vmovdqu    %3,%%ymm5                      \n"
+    LABELALIGN
+  "1:                                          \n"
+    VMEMOPREG(vpermd,-0x20,0,2,4,ymm5,ymm0) // vpermd -0x20(%0,%2,4),ymm5,ymm0
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "sub        $0x8,%2                        \n"
+    "jg         1b                             \n"
+    "vzeroupper                                \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kARGBShuffleMirror_AVX2) // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm5"
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_AVX2
+
+#ifdef HAS_SPLITUVROW_AVX2
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5             \n"
+    "vpsrlw     $0x8,%%ymm5,%%ymm5               \n"
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0          \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1    \n"
+    "lea        " MEMLEA(0x40,0) ",%0            \n"
+    "vpsrlw     $0x8,%%ymm0,%%ymm2               \n"
+    "vpsrlw     $0x8,%%ymm1,%%ymm3               \n"
+    "vpand      %%ymm5,%%ymm0,%%ymm0             \n"
+    "vpand      %%ymm5,%%ymm1,%%ymm1             \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0             \n"
+    "vpackuswb  %%ymm3,%%ymm2,%%ymm2             \n"
+    "vpermq     $0xd8,%%ymm0,%%ymm0              \n"
+    "vpermq     $0xd8,%%ymm2,%%ymm2              \n"
+    "vmovdqu    %%ymm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(vmovdqu,ymm2,0x00,1,2,1)             //  vmovdqu %%ymm2,(%1,%2)
+    "lea        " MEMLEA(0x20,1) ",%1            \n"
+    "sub        $0x20,%3                         \n"
+    "jg         1b                               \n"
+    "vzeroupper                                  \n"
+  : "+r"(src_uv),     // %0
+    "+r"(dst_u),      // %1
+    "+r"(dst_v),      // %2
+    "+r"(pix)         // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_SPLITUVROW_AVX2
+
+#ifdef HAS_SPLITUVROW_SSE2
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb    %%xmm5,%%xmm5                    \n"
+    "psrlw      $0x8,%%xmm5                      \n"
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqu     " MEMACCESS(0) ",%%xmm0          \n"
+    "movdqu     " MEMACCESS2(0x10,0) ",%%xmm1    \n"
+    "lea        " MEMLEA(0x20,0) ",%0            \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "pand       %%xmm5,%%xmm0                    \n"
+    "pand       %%xmm5,%%xmm1                    \n"
+    "packuswb   %%xmm1,%%xmm0                    \n"
+    "psrlw      $0x8,%%xmm2                      \n"
+    "psrlw      $0x8,%%xmm3                      \n"
+    "packuswb   %%xmm3,%%xmm2                    \n"
+    "movdqu     %%xmm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(movdqu,xmm2,0x00,1,2,1)             //  movdqu     %%xmm2,(%1,%2)
+    "lea        " MEMLEA(0x10,1) ",%1            \n"
+    "sub        $0x10,%3                         \n"
+    "jg         1b                               \n"
+  : "+r"(src_uv),     // %0
+    "+r"(dst_u),      // %1
+    "+r"(dst_v),      // %2
+    "+r"(pix)         // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_SPLITUVROW_SSE2
+
+#ifdef HAS_MERGEUVROW_AVX2
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    "sub       %0,%1                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0           \n"
+    MEMOPREG(vmovdqu,0x00,0,1,1,ymm1)             //  vmovdqu (%0,%1,1),%%ymm1
+    "lea       " MEMLEA(0x20,0) ",%0             \n"
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm2             \n"
+    "vpunpckhbw %%ymm1,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm2," MEMACCESS(2) "   \n"
+    "vextractf128 $0x0,%%ymm0," MEMACCESS2(0x10,2) "\n"
+    "vextractf128 $0x1,%%ymm2," MEMACCESS2(0x20,2) "\n"
+    "vextractf128 $0x1,%%ymm0," MEMACCESS2(0x30,2) "\n"
+    "lea       " MEMLEA(0x40,2) ",%2             \n"
+    "sub       $0x20,%3                          \n"
+    "jg        1b                                \n"
+    "vzeroupper                                  \n"
+  : "+r"(src_u),     // %0
+    "+r"(src_v),     // %1
+    "+r"(dst_uv),    // %2
+    "+r"(width)      // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2"
+  );
+}
+#endif  // HAS_MERGEUVROW_AVX2
+
+#ifdef HAS_MERGEUVROW_SSE2
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    "sub       %0,%1                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)             //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm2                     \n"
+    "punpcklbw %%xmm1,%%xmm0                     \n"
+    "punpckhbw %%xmm1,%%xmm2                     \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "           \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,2) "     \n"
+    "lea       " MEMLEA(0x20,2) ",%2             \n"
+    "sub       $0x10,%3                          \n"
+    "jg        1b                                \n"
+  : "+r"(src_u),     // %0
+    "+r"(src_v),     // %1
+    "+r"(dst_uv),    // %2
+    "+r"(width)      // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2"
+  );
+}
+#endif  // HAS_MERGEUVROW_SSE2
+
+#ifdef HAS_COPYROW_SSE2
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_COPYROW_SSE2
+
+#ifdef HAS_COPYROW_AVX
+void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x40,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_COPYROW_AVX
+
+#ifdef HAS_COPYROW_ERMS
+// Multiple of 1.
+void CopyRow_ERMS(const uint8* src, uint8* dst, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep movsb " MEMMOVESTRING(0,1) "          \n"
+  : "+S"(src),  // %0
+    "+D"(dst),  // %1
+    "+c"(width_tmp) // %2
+  :
+  : "memory", "cc"
+  );
+}
+#endif  // HAS_COPYROW_ERMS
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm0,%%xmm0                   \n"
+    "pslld     $0x18,%%xmm0                    \n"
+    "pcmpeqb   %%xmm1,%%xmm1                   \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqu    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "pand      %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm0,%%xmm3                   \n"
+    "pand      %%xmm1,%%xmm4                   \n"
+    "pand      %%xmm1,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYALPHAROW_AVX2
+// width in pixels
+void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm1         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm2   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
+    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
+    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2"
+  );
+}
+#endif  // HAS_ARGBCOPYALPHAROW_AVX2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
+// width in pixels
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm0,%%xmm0                   \n"
+    "pslld     $0x18,%%xmm0                    \n"
+    "pcmpeqb   %%xmm1,%%xmm1                   \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpckhwd %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm2,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqu    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "pand      %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm0,%%xmm3                   \n"
+    "pand      %%xmm1,%%xmm4                   \n"
+    "pand      %%xmm1,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
+// width in pixels
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vpmovzxbd " MEMACCESS(0) ",%%ymm1         \n"
+    "vpmovzxbd " MEMACCESS2(0x8,0) ",%%ymm2    \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "vpslld    $0x18,%%ymm1,%%ymm1             \n"
+    "vpslld    $0x18,%%ymm2,%%ymm2             \n"
+    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
+    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
+    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2"
+  );
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
+
+#ifdef HAS_SETROW_X86
+void SetRow_X86(uint8* dst, uint8 v8, int width) {
+  size_t width_tmp = (size_t)(width >> 2);
+  const uint32 v32 = v8 * 0x01010101;  // Duplicate byte to all bytes.
+  asm volatile (
+    "rep stosl " MEMSTORESTRING(eax,0) "       \n"
+    : "+D"(dst),       // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v32)         // %2
+    : "memory", "cc");
+}
+
+void SetRow_ERMS(uint8* dst, uint8 v8, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep stosb " MEMSTORESTRING(al,0) "        \n"
+    : "+D"(dst),       // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v8)          // %2
+    : "memory", "cc");
+}
+
+void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep stosl " MEMSTORESTRING(eax,0) "       \n"
+    : "+D"(dst_argb),  // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v32)         // %2
+    : "memory", "cc");
+}
+#endif  // HAS_SETROW_X86
+
+#ifdef HAS_YUY2TOYROW_SSE2
+void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_YUY2TOYROW_SSE2
+
+#ifdef HAS_YUY2TOYROW_AVX2
+void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "lea      " MEMLEA(0x20,1) ",%1            \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
+    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "lea      " MEMLEA(0x20,1) ",%1            \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm5"
+  );
+}
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
+    "sub       %1,%2                           \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
+    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    "vpsrlw     $0x8,%%ymm5,%%ymm5             \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_YUY2TOYROW_AVX2
+
+#ifdef HAS_ARGBBLENDROW_SSE2
+// Blend 8 pixels at a time.
+void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0xf,%%xmm7                     \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x8,%%xmm6                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psllw     $0x8,%%xmm5                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "41:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       41b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop.
+  "91:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
+    "jge       91b                             \n"
+  "99:                                         \n"
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBBLENDROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSSE3
+// Shuffle table for isolating alpha.
+static uvec8 kShuffleAlpha = {
+  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
+  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
+};
+
+// Blend 8 pixels at a time
+// Shuffle table for reversing the bytes.
+
+// Same as SSE2, but replaces
+//    psrlw      xmm3, 8          // alpha
+//    pshufhw    xmm3, xmm3,0F5h  // 8 alpha words
+//    pshuflw    xmm3, xmm3,0F5h
+// with..
+//    pshufb     xmm3, kShuffleAlpha // alpha
+
+void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
+                        uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0xf,%%xmm7                     \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x8,%%xmm6                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psllw     $0x8,%%xmm5                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "40:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop.
+  "91:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
+    "jge       91b                             \n"
+  "99:                                         \n"
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  : "m"(kShuffleAlpha)  // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBBLENDROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+// Attenuate 4 pixels at a time.
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x8,%%xmm5                     \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "pshufhw   $0xff,%%xmm0,%%xmm2             \n"
+    "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pshufhw   $0xff,%%xmm1,%%xmm2             \n"
+    "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "pand      %%xmm4,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+// Shuffle table duplicating alpha
+static uvec8 kShuffleAlpha0 = {
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u
+};
+static uvec8 kShuffleAlpha1 = {
+  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u
+};
+// Attenuate 4 pixels at a time.
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "pslld     $0x18,%%xmm3                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpcklbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm1,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "punpckhbw %%xmm2,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "pand      %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  : "m"(kShuffleAlpha0),  // %3
+    "m"(kShuffleAlpha1)  // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha_AVX2 = {
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
+};
+// Attenuate 8 pixels at a time.
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "vbroadcastf128 %3,%%ymm4                  \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    "vpslld     $0x18,%%ymm5,%%ymm5            \n"
+    "sub        %0,%1                          \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm6        \n"
+    "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
+    "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
+    "vpshufb    %%ymm4,%%ymm0,%%ymm2           \n"
+    "vpshufb    %%ymm4,%%ymm1,%%ymm3           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpand      %%ymm5,%%ymm6,%%ymm6           \n"
+    "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x8,%%ymm1,%%ymm1             \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpor       %%ymm6,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1)          //  vmovdqu %%ymm0,(%0,%1)
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub        $0x8,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  : "m"(kShuffleAlpha_AVX2)  // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+// Unattenuate 4 pixels at a time.
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  uintptr_t alpha = 0;
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "movlhps   %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "movlhps   %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width),       // %2
+    "+r"(alpha)        // %3
+  : "r"(fixed_invtbl8)  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBUNATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kUnattenShuffleAlpha_AVX2 = {
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u
+};
+// Unattenuate 8 pixels at a time.
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  uintptr_t alpha = 0;
+  asm volatile (
+    "sub        %0,%1                          \n"
+    "vbroadcastf128 %5,%%ymm5                  \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    // replace VPGATHER
+    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm0)             //  vmovd 0x0(%4,%3,4),%%xmm0
+    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm1)             //  vmovd 0x0(%4,%3,4),%%xmm1
+    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
+    "vpunpckldq %%xmm1,%%xmm0,%%xmm6           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm2)             //  vmovd 0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm3)             //  vmovd 0x0(%4,%3,4),%%xmm3
+    "movzb     " MEMACCESS2(0x13,0) ",%3       \n"
+    "vpunpckldq %%xmm3,%%xmm2,%%xmm7           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm0)             //  vmovd 0x0(%4,%3,4),%%xmm0
+    "movzb     " MEMACCESS2(0x17,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm1)             //  vmovd 0x0(%4,%3,4),%%xmm1
+    "movzb     " MEMACCESS2(0x1b,0) ",%3       \n"
+    "vpunpckldq %%xmm1,%%xmm0,%%xmm0           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm2)             //  vmovd 0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x1f,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm3)             //  vmovd 0x0(%4,%3,4),%%xmm3
+    "vpunpckldq %%xmm3,%%xmm2,%%xmm2           \n"
+    "vpunpcklqdq %%xmm7,%%xmm6,%%xmm3          \n"
+    "vpunpcklqdq %%xmm2,%%xmm0,%%xmm0          \n"
+    "vinserti128 $0x1,%%xmm0,%%ymm3,%%ymm3     \n"
+    // end of VPGATHER
+
+    "vmovdqu    " MEMACCESS(0) ",%%ymm6        \n"
+    "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
+    "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
+    "vpunpcklwd %%ymm3,%%ymm3,%%ymm2           \n"
+    "vpunpckhwd %%ymm3,%%ymm3,%%ymm3           \n"
+    "vpshufb    %%ymm5,%%ymm2,%%ymm2           \n"
+    "vpshufb    %%ymm5,%%ymm3,%%ymm3           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1)          //  vmovdqu %%ymm0,(%0,%1)
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub        $0x8,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width),       // %2
+    "+r"(alpha)        // %3
+  : "r"(fixed_invtbl8),  // %4
+    "m"(kUnattenShuffleAlpha_AVX2)  // %5
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBUNATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBGRAYROW_SSSE3
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrld     $0x18,%%xmm2                    \n"
+    "psrld     $0x18,%%xmm3                    \n"
+    "packuswb  %%xmm3,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm3                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "m"(kARGBToYJ),   // %3
+    "m"(kAddYJ64)     // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBGRAYROW_SSSE3
+
+#ifdef HAS_ARGBSEPIAROW_SSSE3
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+// Constant for ARGB color to sepia tone
+static vec8 kARGBToSepiaB = {
+  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
+};
+
+static vec8 kARGBToSepiaG = {
+  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
+};
+
+static vec8 kARGBToSepiaR = {
+  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
+};
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+  asm volatile (
+    "movdqa    %2,%%xmm2                       \n"
+    "movdqa    %3,%%xmm3                       \n"
+    "movdqa    %4,%%xmm4                       \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "pmaddubsw %%xmm2,%%xmm6                   \n"
+    "phaddw    %%xmm6,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm3,%%xmm5                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm5                   \n"
+    "psrlw     $0x7,%%xmm5                     \n"
+    "packuswb  %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm4,%%xmm5                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm5                   \n"
+    "psrlw     $0x7,%%xmm5                     \n"
+    "packuswb  %%xmm5,%%xmm5                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "psrld     $0x18,%%xmm6                    \n"
+    "psrld     $0x18,%%xmm1                    \n"
+    "packuswb  %%xmm1,%%xmm6                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm5                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm5,%%xmm0                   \n"
+    "punpckhwd %%xmm5,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub       $0x8,%1                         \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),      // %0
+    "+r"(width)          // %1
+  : "m"(kARGBToSepiaB),  // %2
+    "m"(kARGBToSepiaG),  // %3
+    "m"(kARGBToSepiaR)   // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBSEPIAROW_SSSE3
+
+#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// Same as Sepia except matrix is provided.
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width) {
+  asm volatile (
+    "movdqu    " MEMACCESS(3) ",%%xmm5         \n"
+    "pshufd    $0x00,%%xmm5,%%xmm2             \n"
+    "pshufd    $0x55,%%xmm5,%%xmm3             \n"
+    "pshufd    $0xaa,%%xmm5,%%xmm4             \n"
+    "pshufd    $0xff,%%xmm5,%%xmm5             \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "pmaddubsw %%xmm2,%%xmm7                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "phaddsw   %%xmm7,%%xmm0                   \n"
+    "phaddsw   %%xmm1,%%xmm6                   \n"
+    "psraw     $0x6,%%xmm0                     \n"
+    "psraw     $0x6,%%xmm6                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm7                   \n"
+    "phaddsw   %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm7                   \n"
+    "phaddsw   %%xmm7,%%xmm6                   \n"
+    "psraw     $0x6,%%xmm1                     \n"
+    "psraw     $0x6,%%xmm6                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "punpcklwd %%xmm1,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm6                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb),      // %1
+    "+r"(width)          // %2
+  : "r"(matrix_argb)     // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
+
+#ifdef HAS_ARGBQUANTIZEROW_SSE2
+// Quantize 4 ARGB pixels (16 bytes).
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "movd      %2,%%xmm2                       \n"
+    "movd      %3,%%xmm3                       \n"
+    "movd      %4,%%xmm4                       \n"
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshufd    $0x44,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "pshufd    $0x44,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0x40,%%xmm4,%%xmm4             \n"
+    "pshufd    $0x44,%%xmm4,%%xmm4             \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "pslld     $0x18,%%xmm6                    \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "pmullw    %%xmm3,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm7         \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm6,%%xmm7                   \n"
+    "paddw     %%xmm4,%%xmm0                   \n"
+    "paddw     %%xmm4,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "por       %%xmm7,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x4,%1                         \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBQUANTIZEROW_SSE2
+
+#ifdef HAS_ARGBSHADEROW_SSE2
+// Shade 4 pixels at a time by specified value.
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "movd      %3,%%xmm2                       \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm2                  \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(value)       // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2"
+  );
+}
+#endif  // HAS_ARGBSHADEROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+// Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                  \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpckhbw %%xmm5,%%xmm3                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm1        \n"
+    "lea        " MEMLEA(0x20,0) ",%0          \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm3        \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "vpunpcklbw %%ymm1,%%ymm1,%%ymm0           \n"
+    "vpunpckhbw %%ymm1,%%ymm1,%%ymm1           \n"
+    "vpunpcklbw %%ymm5,%%ymm3,%%ymm2           \n"
+    "vpunpckhbw %%ymm5,%%ymm3,%%ymm3           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
+    "lea       " MEMLEA(0x20,2) ",%2           \n"
+    "sub        $0x8,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__AVX2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_AVX2
+
+#ifdef HAS_ARGBADDROW_SSE2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_ARGBADDROW_SSE2
+
+#ifdef HAS_ARGBADDROW_AVX2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "lea        " MEMLEA(0x20,0) ",%0          \n"
+    "vpaddusb   " MEMACCESS(1) ",%%ymm0,%%ymm0 \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
+    "lea        " MEMLEA(0x20,2) ",%2          \n"
+    "sub        $0x8,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0"
+  );
+}
+#endif  // HAS_ARGBADDROW_AVX2
+
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels, 4 pixels at a time.
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psubusb   %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "lea        " MEMLEA(0x20,0) ",%0          \n"
+    "vpsubusb   " MEMACCESS(1) ",%%ymm0,%%ymm0 \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
+    "lea        " MEMLEA(0x20,2) ",%2          \n"
+    "sub        $0x8,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0"
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_AVX2
+
+#ifdef HAS_SOBELXROW_SSE2
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+    "sub       %0,%3                           \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "movq      " MEMACCESS2(0x2,0) ",%%xmm1    \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "psubw     %%xmm1,%%xmm0                   \n"
+    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
+    MEMOPREG(movq,0x02,0,1,1,xmm2)             //  movq      0x2(%0,%1,1),%%xmm2
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "psubw     %%xmm2,%%xmm1                   \n"
+    MEMOPREG(movq,0x00,0,2,1,xmm2)             //  movq      (%0,%2,1),%%xmm2
+    MEMOPREG(movq,0x02,0,2,1,xmm3)             //  movq      0x2(%0,%2,1),%%xmm3
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "psubw     %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm1                   \n"
+    "pmaxsw    %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    MEMOPMEM(movq,xmm0,0x00,0,3,1)             //  movq      %%xmm0,(%0,%3,1)
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x8,%4                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_SOBELXROW_SSE2
+
+#ifdef HAS_SOBELYROW_SSE2
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "psubw     %%xmm1,%%xmm0                   \n"
+    "movq      " MEMACCESS2(0x1,0) ",%%xmm1    \n"
+    MEMOPREG(movq,0x01,0,1,1,xmm2)             //  movq      0x1(%0,%1,1),%%xmm2
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "psubw     %%xmm2,%%xmm1                   \n"
+    "movq      " MEMACCESS2(0x2,0) ",%%xmm2    \n"
+    MEMOPREG(movq,0x02,0,1,1,xmm3)             //  movq      0x2(%0,%1,1),%%xmm3
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "psubw     %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm1                   \n"
+    "pmaxsw    %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    MEMOPMEM(movq,xmm0,0x00,0,2,1)             //  movq      %%xmm0,(%0,%2,1)
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x8,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_SOBELYROW_SSE2
+
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm2                   \n"
+    "punpckhbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm1                   \n"
+    "punpckhwd %%xmm2,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklwd %%xmm0,%%xmm3                   \n"
+    "punpckhwd %%xmm0,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm1," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,2) "   \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x20,2) "   \n"
+    "movdqu    %%xmm0," MEMACCESS2(0x30,2) "   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_SOBELROW_SSE2
+
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into a plane.
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_SOBELTOPLANEROW_SSE2
+
+#ifdef HAS_SOBELXYROW_SSE2
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "paddusb   %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "punpckhbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "punpcklbw %%xmm2,%%xmm4                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm6                   \n"
+    "punpcklwd %%xmm3,%%xmm6                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "movdqa    %%xmm1,%%xmm7                   \n"
+    "punpcklwd %%xmm0,%%xmm7                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm6," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm4," MEMACCESS2(0x10,2) "   \n"
+    "movdqu    %%xmm7," MEMACCESS2(0x20,2) "   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x30,2) "   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_SOBELXYROW_SSE2
+
+#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
+// Creates a table of cumulative sums where each value is a sum of all values
+// above and to the left of the value, inclusive of the value.
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width) {
+  asm volatile (
+    "pxor      %%xmm0,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+    "test      $0xf,%1                         \n"
+    "jne       49f                             \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm2,%%xmm4                   \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
+    "punpckhwd %%xmm1,%%xmm3                   \n"
+    "punpckhbw %%xmm1,%%xmm4                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "punpcklwd %%xmm1,%%xmm4                   \n"
+    "punpckhwd %%xmm1,%%xmm5                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm2         \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,2) ",%%xmm3   \n"
+    "paddd     %%xmm0,%%xmm3                   \n"
+    "paddd     %%xmm4,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x20,2) ",%%xmm4   \n"
+    "paddd     %%xmm0,%%xmm4                   \n"
+    "paddd     %%xmm5,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x30,2) ",%%xmm5   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "paddd     %%xmm0,%%xmm5                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm4," MEMACCESS2(0x20,1) "   \n"
+    "movdqu    %%xmm5," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x1,%3                         \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+  : "+r"(row),  // %0
+    "+r"(cumsum),  // %1
+    "+r"(previous_cumsum),  // %2
+    "+r"(width)  // %3
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
+
+#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst,
+                                    int count) {
+  asm volatile (
+    "movd      %5,%%xmm5                       \n"
+    "cvtdq2ps  %%xmm5,%%xmm5                   \n"
+    "rcpss     %%xmm5,%%xmm4                   \n"
+    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+    "cmpl      $0x80,%5                        \n"
+    "ja        40f                             \n"
+
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrld     $0x10,%%xmm6                    \n"
+    "cvtdq2ps  %%xmm6,%%xmm6                   \n"
+    "addps     %%xmm6,%%xmm5                   \n"
+    "mulps     %%xmm4,%%xmm5                   \n"
+    "cvtps2dq  %%xmm5,%%xmm5                   \n"
+    "packssdw  %%xmm5,%%xmm5                   \n"
+
+  // 4 pixel small loop                        \n"
+    LABELALIGN
+  "4:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
+    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
+    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
+    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
+    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
+    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
+    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm5,%%xmm0                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       4b                              \n"
+    "jmp       49f                             \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
+    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
+    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
+    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
+    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
+    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
+    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm1,%%xmm1                   \n"
+    "mulps     %%xmm4,%%xmm0                   \n"
+    "mulps     %%xmm4,%%xmm1                   \n"
+    "cvtdq2ps  %%xmm2,%%xmm2                   \n"
+    "cvtdq2ps  %%xmm3,%%xmm3                   \n"
+    "mulps     %%xmm4,%%xmm2                   \n"
+    "mulps     %%xmm4,%%xmm3                   \n"
+    "cvtps2dq  %%xmm0,%%xmm0                   \n"
+    "cvtps2dq  %%xmm1,%%xmm1                   \n"
+    "cvtps2dq  %%xmm2,%%xmm2                   \n"
+    "cvtps2dq  %%xmm3,%%xmm3                   \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm3,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "mulps     %%xmm4,%%xmm0                   \n"
+    "cvtps2dq  %%xmm0,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
+    "jge       10b                             \n"
+  "19:                                         \n"
+  : "+r"(topleft),  // %0
+    "+r"(botleft),  // %1
+    "+r"(dst),      // %2
+    "+rm"(count)    // %3
+  : "r"((intptr_t)(width)),  // %4
+    "rm"(area)     // %5
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+
+#ifdef HAS_ARGBAFFINEROW_SSE2
+// Copy ARGB pixels from source image with slope to a row of destination.
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* src_dudv, int width) {
+  intptr_t src_argb_stride_temp = src_argb_stride;
+  intptr_t temp = 0;
+  asm volatile (
+    "movq      " MEMACCESS(3) ",%%xmm2         \n"
+    "movq      " MEMACCESS2(0x08,3) ",%%xmm7   \n"
+    "shl       $0x10,%1                        \n"
+    "add       $0x4,%1                         \n"
+    "movd      %1,%%xmm5                       \n"
+    "sub       $0x4,%4                         \n"
+    "jl        49f                             \n"
+
+    "pshufd    $0x44,%%xmm7,%%xmm7             \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm7,%%xmm0                   \n"
+    "movlhps   %%xmm0,%%xmm2                   \n"
+    "movdqa    %%xmm7,%%xmm4                   \n"
+    "addps     %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "addps     %%xmm4,%%xmm3                   \n"
+    "addps     %%xmm4,%%xmm4                   \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "cvttps2dq %%xmm2,%%xmm0                   \n"  // x, y float to int first 2
+    "cvttps2dq %%xmm3,%%xmm1                   \n"  // x, y float to int next 2
+    "packssdw  %%xmm1,%%xmm0                   \n"  // x, y as 8 shorts
+    "pmaddwd   %%xmm5,%%xmm0                   \n"  // off = x * 4 + y * stride
+    "movd      %%xmm0,%k1                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    "movd      %%xmm0,%k5                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
+    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
+    "punpckldq %%xmm6,%%xmm1                   \n"
+    "addps     %%xmm4,%%xmm2                   \n"
+    "movq      %%xmm1," MEMACCESS(2) "         \n"
+    "movd      %%xmm0,%k1                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    "movd      %%xmm0,%k5                      \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
+    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
+    "punpckldq %%xmm6,%%xmm0                   \n"
+    "addps     %%xmm4,%%xmm3                   \n"
+    "movq      %%xmm0," MEMACCESS2(0x08,2) "   \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%4                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%4                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "cvttps2dq %%xmm2,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "pmaddwd   %%xmm5,%%xmm0                   \n"
+    "addps     %%xmm7,%%xmm2                   \n"
+    "movd      %%xmm0,%k1                      \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x04,2) ",%2           \n"
+    "sub       $0x1,%4                         \n"
+    "jge       10b                             \n"
+  "19:                                         \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_argb_stride_temp),  // %1
+    "+r"(dst_argb),  // %2
+    "+r"(src_dudv),  // %3
+    "+rm"(width),    // %4
+    "+r"(temp)   // %5
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBAFFINEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "pmaddubsw %%xmm5,%%xmm0                   \n"
+    "pmaddubsw %%xmm5,%%xmm1                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm5"
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+// Bilinear filter 32x2 -> 32x1
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  asm volatile (
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "sub       %1,%0                           \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "vmovd      %3,%%xmm0                      \n"
+    "neg        %3                             \n"
+    "add        $0x80,%3                       \n"
+    "vmovd      %3,%%xmm5                      \n"
+    "vpunpcklbw %%xmm0,%%xmm5,%%xmm5           \n"
+    "vpunpcklwd %%xmm5,%%xmm5,%%xmm5           \n"
+    "vpxor      %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpermd     %%ymm5,%%ymm0,%%ymm5           \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
+    MEMOPREG(vmovdqu,0x00,1,4,1,ymm2)
+    "vpunpckhbw %%ymm2,%%ymm0,%%ymm1           \n"
+    "vpunpcklbw %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm5,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm5,%%ymm1,%%ymm1           \n"
+    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x7,%%ymm1,%%ymm1             \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
+    MEMOPREG(vmovdqu,0x00,1,4,1,ymm1)
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
+    VMEMOPREG(vpavgb,0x00,1,4,1,ymm0,ymm0)     // vpavgb (%1,%4,1),%%ymm0,%%ymm0
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm1        \n"
+    MEMOPREG(vmovdqu,0x00,1,4,1,ymm0)
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "rep movsb " MEMMOVESTRING(1,0) "          \n"
+    "jmp       999f                            \n"
+
+  "99:                                         \n"
+    "vzeroupper                                \n"
+  "999:                                        \n"
+  : "+D"(dst_ptr),    // %0
+    "+S"(src_ptr),    // %1
+    "+c"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm5"
+  );
+}
+#endif  // HAS_INTERPOLATEROW_AVX2
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)           //  movdqu    (%1,%4,1),%%xmm2
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm2                   \n"
+    "psubw     %%xmm1,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm2                   \n"
+    "paddw     %%xmm3,%%xmm3                   \n"
+    "pmulhw    %%xmm5,%%xmm2                   \n"
+    "pmulhw    %%xmm5,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)           //  movdqu    (%1,%4,1),%%xmm0
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+#ifdef HAS_ARGBSHUFFLEROW_SSSE3
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix) {
+  asm volatile (
+    "movdqu    " MEMACCESS(3) ",%%xmm5         \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_SSSE3
+
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    "vbroadcastf128 " MEMACCESS(3) ",%%ymm5    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpshufb   %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpshufb   %%ymm5,%%ymm1,%%ymm1            \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_AVX2
+
+#ifdef HAS_ARGBSHUFFLEROW_SSE2
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "mov       " MEMACCESS(4) ",%k2            \n"
+    "cmp       $0x3000102,%k2                  \n"
+    "je        3012f                           \n"
+    "cmp       $0x10203,%k2                    \n"
+    "je        123f                            \n"
+    "cmp       $0x30201,%k2                    \n"
+    "je        321f                            \n"
+    "cmp       $0x2010003,%k2                  \n"
+    "je        2103f                           \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(4) ",%2             \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS(1) "            \n"
+    "movzb     " MEMACCESS2(0x1,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x1,1) "       \n"
+    "movzb     " MEMACCESS2(0x2,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x2,1) "       \n"
+    "movzb     " MEMACCESS2(0x3,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x3,1) "       \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "sub       $0x1,%3                         \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "123:                                        \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x1b,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x1b,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        123b                            \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "321:                                        \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x39,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x39,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x39,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x39,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        321b                            \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "2103:                                       \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x93,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x93,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x93,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x93,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        2103b                           \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "3012:                                       \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0xc6,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0xc6,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0xc6,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0xc6,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        3012b                           \n"
+
+  "99:                                         \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+d"(pixel_temp),  // %2
+    "+r"(pix)         // %3
+  : "r"(shuffler)      // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_SSE2
+
+#ifdef HAS_I422TOYUY2ROW_SSE2
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+ asm volatile (
+    "sub       %1,%2                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movq      " MEMACCESS(1) ",%%xmm2           \n"
+    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
+    "lea       " MEMLEA(0x8,1) ",%1              \n"
+    "punpcklbw %%xmm3,%%xmm2                     \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm1                     \n"
+    "punpcklbw %%xmm2,%%xmm0                     \n"
+    "punpckhbw %%xmm2,%%xmm1                     \n"
+    "movdqu    %%xmm0," MEMACCESS(3) "           \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,3) "     \n"
+    "lea       " MEMLEA(0x20,3) ",%3             \n"
+    "sub       $0x10,%4                          \n"
+    "jg         1b                               \n"
+    : "+r"(src_y),  // %0
+      "+r"(src_u),  // %1
+      "+r"(src_v),  // %2
+      "+r"(dst_frame),  // %3
+      "+rm"(width)  // %4
+    :
+    : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
+  );
+}
+#endif  // HAS_I422TOYUY2ROW_SSE2
+
+#ifdef HAS_I422TOUYVYROW_SSE2
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+ asm volatile (
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movq      " MEMACCESS(1) ",%%xmm2           \n"
+    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
+    "lea       " MEMLEA(0x8,1) ",%1              \n"
+    "punpcklbw %%xmm3,%%xmm2                     \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    "movdqa    %%xmm2,%%xmm1                     \n"
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "punpcklbw %%xmm0,%%xmm1                     \n"
+    "punpckhbw %%xmm0,%%xmm2                     \n"
+    "movdqu    %%xmm1," MEMACCESS(3) "           \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,3) "     \n"
+    "lea       " MEMLEA(0x20,3) ",%3             \n"
+    "sub       $0x10,%4                          \n"
+    "jg         1b                               \n"
+    : "+r"(src_y),  // %0
+      "+r"(src_u),  // %1
+      "+r"(src_v),  // %2
+      "+r"(dst_frame),  // %3
+      "+rm"(width)  // %4
+    :
+    : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
+  );
+}
+#endif  // HAS_I422TOUYVYROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "pxor      %%xmm3,%%xmm3                   \n"
+
+    // 2 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm3,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm4                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "mulps     " MEMACCESS2(0x10,3) ",%%xmm0   \n"
+    "mulps     " MEMACCESS2(0x10,3) ",%%xmm4   \n"
+    "addps     " MEMACCESS(3) ",%%xmm0         \n"
+    "addps     " MEMACCESS(3) ",%%xmm4         \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm1,%%xmm2                   \n"
+    "mulps     %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm2,%%xmm1                   \n"
+    "mulps     %%xmm6,%%xmm5                   \n"
+    "mulps     " MEMACCESS2(0x20,3) ",%%xmm2   \n"
+    "mulps     " MEMACCESS2(0x20,3) ",%%xmm6   \n"
+    "mulps     " MEMACCESS2(0x30,3) ",%%xmm1   \n"
+    "mulps     " MEMACCESS2(0x30,3) ",%%xmm5   \n"
+    "addps     %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm6,%%xmm4                   \n"
+    "addps     %%xmm1,%%xmm0                   \n"
+    "addps     %%xmm5,%%xmm4                   \n"
+    "cvttps2dq %%xmm0,%%xmm0                   \n"
+    "cvttps2dq %%xmm4,%%xmm4                   \n"
+    "packuswb  %%xmm4,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x2,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "vbroadcastf128 " MEMACCESS(3) ",%%ymm4     \n"
+    "vbroadcastf128 " MEMACCESS2(0x10,3) ",%%ymm5 \n"
+    "vbroadcastf128 " MEMACCESS2(0x20,3) ",%%ymm6 \n"
+    "vbroadcastf128 " MEMACCESS2(0x30,3) ",%%ymm7 \n"
+
+    // 2 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vpmovzxbd   " MEMACCESS(0) ",%%ymm0       \n"  // 2 ARGB pixels
+    "lea         " MEMLEA(0x8,0) ",%0          \n"
+    "vcvtdq2ps   %%ymm0,%%ymm0                 \n"  // X 8 floats
+    "vmulps      %%ymm0,%%ymm0,%%ymm2          \n"  // X * X
+    "vmulps      %%ymm7,%%ymm0,%%ymm3          \n"  // C3 * X
+    "vfmadd132ps %%ymm5,%%ymm4,%%ymm0          \n"  // result = C0 + C1 * X
+    "vfmadd231ps %%ymm6,%%ymm2,%%ymm0          \n"  // result += C2 * X * X
+    "vfmadd231ps %%ymm3,%%ymm2,%%ymm0          \n"  // result += C3 * X * X * X
+    "vcvttps2dq  %%ymm0,%%ymm0                 \n"
+    "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
+    "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+    "vpackuswb   %%xmm0,%%xmm0,%%xmm0          \n"
+    "vmovq       %%xmm0," MEMACCESS(1) "       \n"
+    "lea         " MEMLEA(0x8,1) ",%1          \n"
+    "sub         $0x2,%2                       \n"
+    "jg          1b                            \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
+#ifdef HAS_ARGBCOLORTABLEROW_X86
+// Tranform ARGB pixels with color table.
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+                           int width) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    // 1 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(0) ",%1             \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
+    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
+    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
+    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
+    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
+    "movzb     " MEMACCESS2(-0x1,0) ",%1       \n"
+    MEMOPARG(movzb,0x03,3,1,4,1) "             \n"  // movzb 0x3(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x1,0) "      \n"
+    "dec       %2                              \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),   // %0
+    "+d"(pixel_temp), // %1
+    "+r"(width)       // %2
+  : "r"(table_argb)   // %3
+  : "memory", "cc");
+}
+#endif  // HAS_ARGBCOLORTABLEROW_X86
+
+#ifdef HAS_RGBCOLORTABLEROW_X86
+// Tranform RGB pixels with color table.
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    // 1 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(0) ",%1             \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
+    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
+    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
+    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
+    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
+    "dec       %2                              \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),   // %0
+    "+d"(pixel_temp), // %1
+    "+r"(width)       // %2
+  : "r"(table_argb)   // %3
+  : "memory", "cc");
+}
+#endif  // HAS_RGBCOLORTABLEROW_X86
+
+#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// Tranform RGB pixels with luma table.
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff) {
+  uintptr_t pixel_temp = 0u;
+  uintptr_t table_temp = 0u;
+  asm volatile (
+    "movd      %6,%%xmm3                       \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0x8,%%xmm4                     \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm0         \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "phaddw    %%xmm0,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "punpcklwd %%xmm5,%%xmm0                   \n"
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS(2) ",%0             \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS(3) "            \n"
+    "movzb     " MEMACCESS2(0x1,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x1,3) "       \n"
+    "movzb     " MEMACCESS2(0x2,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x2,3) "       \n"
+    "movzb     " MEMACCESS2(0x3,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0x3,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS2(0x4,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x4,3) "       \n"
+    "movzb     " MEMACCESS2(0x5,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x5,3) "       \n"
+    "movzb     " MEMACCESS2(0x6,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x6,3) "       \n"
+    "movzb     " MEMACCESS2(0x7,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0x7,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS2(0x8,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x8,3) "       \n"
+    "movzb     " MEMACCESS2(0x9,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x9,3) "       \n"
+    "movzb     " MEMACCESS2(0xa,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xa,3) "       \n"
+    "movzb     " MEMACCESS2(0xb,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0xb,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+
+    "movzb     " MEMACCESS2(0xc,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xc,3) "       \n"
+    "movzb     " MEMACCESS2(0xd,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xd,3) "       \n"
+    "movzb     " MEMACCESS2(0xe,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xe,3) "       \n"
+    "movzb     " MEMACCESS2(0xf,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0xf,3) "       \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "lea       " MEMLEA(0x10,3) ",%3           \n"
+    "sub       $0x4,%4                         \n"
+    "jg        1b                              \n"
+  : "+d"(pixel_temp),  // %0
+    "+a"(table_temp),  // %1
+    "+r"(src_argb),    // %2
+    "+r"(dst_argb),    // %3
+    "+rm"(width)       // %4
+  : "r"(luma),         // %5
+    "rm"(lumacoeff)    // %6
+  : "memory", "cc", "xmm0", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/row_mips.cc b/libvpx/libvpx/third_party/libyuv/source/row_mips.cc
new file mode 100644
index 0000000..cfc9ffe
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/row_mips.cc
@@ -0,0 +1,911 @@
+/*
+ *  Copyright (c) 2012 The LibYuv project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+#ifdef HAS_COPYROW_MIPS
+void CopyRow_MIPS(const uint8* src, uint8* dst, int count) {
+  __asm__ __volatile__ (
+    ".set      noreorder                         \n"
+    ".set      noat                              \n"
+    "slti      $at, %[count], 8                  \n"
+    "bne       $at ,$zero, $last8                \n"
+    "xor       $t8, %[src], %[dst]               \n"
+    "andi      $t8, $t8, 0x3                     \n"
+
+    "bne       $t8, $zero, unaligned             \n"
+    "negu      $a3, %[dst]                       \n"
+    // make dst/src aligned
+    "andi      $a3, $a3, 0x3                     \n"
+    "beq       $a3, $zero, $chk16w               \n"
+    // word-aligned now count is the remining bytes count
+    "subu     %[count], %[count], $a3            \n"
+
+    "lwr       $t8, 0(%[src])                    \n"
+    "addu      %[src], %[src], $a3               \n"
+    "swr       $t8, 0(%[dst])                    \n"
+    "addu      %[dst], %[dst], $a3               \n"
+
+    // Now the dst/src are mutually word-aligned with word-aligned addresses
+    "$chk16w:                                    \n"
+    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
+    // t8 is the byte count after 64-byte chunks
+    "beq       %[count], $t8, chk8w              \n"
+    // There will be at most 1 32-byte chunk after it
+    "subu      $a3, %[count], $t8                \n"  // the reminder
+    // Here a3 counts bytes in 16w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // Now a3 is the final dst after 64-byte chunks
+    "addu      $t0, %[dst], %[count]             \n"
+    // t0 is the "past the end" address
+
+    // When in the loop we exercise "pref 30,x(a1)", the a1+x should not be past
+    // the "t0-32" address
+    // This means: for x=128 the last "safe" a1 address is "t0-160"
+    // Alternatively, for x=64 the last "safe" a1 address is "t0-96"
+    // we will use "pref 30,128(a1)", so "t0-160" is the limit
+    "subu      $t9, $t0, 160                     \n"
+    // t9 is the "last safe pref 30,128(a1)" address
+    "pref      0, 0(%[src])                      \n"  // first line of src
+    "pref      0, 32(%[src])                     \n"  // second line of src
+    "pref      0, 64(%[src])                     \n"
+    "pref      30, 32(%[dst])                    \n"
+    // In case the a1 > t9 don't use "pref 30" at all
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bgtz      $v1, $loop16w                     \n"
+    "nop                                         \n"
+    // otherwise, start with using pref30
+    "pref      30, 64(%[dst])                    \n"
+    "$loop16w:                                    \n"
+    "pref      0, 96(%[src])                     \n"
+    "lw        $t0, 0(%[src])                    \n"
+    "bgtz      $v1, $skip_pref30_96              \n"  // skip
+    "lw        $t1, 4(%[src])                    \n"
+    "pref      30, 96(%[dst])                    \n"  // continue
+    "$skip_pref30_96:                            \n"
+    "lw        $t2, 8(%[src])                    \n"
+    "lw        $t3, 12(%[src])                   \n"
+    "lw        $t4, 16(%[src])                   \n"
+    "lw        $t5, 20(%[src])                   \n"
+    "lw        $t6, 24(%[src])                   \n"
+    "lw        $t7, 28(%[src])                   \n"
+    "pref      0, 128(%[src])                    \n"
+    //  bring the next lines of src, addr 128
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "lw        $t0, 32(%[src])                   \n"
+    "bgtz      $v1, $skip_pref30_128             \n"  // skip pref 30,128(a1)
+    "lw        $t1, 36(%[src])                   \n"
+    "pref      30, 128(%[dst])                   \n"  // set dest, addr 128
+    "$skip_pref30_128:                           \n"
+    "lw        $t2, 40(%[src])                   \n"
+    "lw        $t3, 44(%[src])                   \n"
+    "lw        $t4, 48(%[src])                   \n"
+    "lw        $t5, 52(%[src])                   \n"
+    "lw        $t6, 56(%[src])                   \n"
+    "lw        $t7, 60(%[src])                   \n"
+    "pref      0, 160(%[src])                    \n"
+    // bring the next lines of src, addr 160
+    "sw        $t0, 32(%[dst])                   \n"
+    "sw        $t1, 36(%[dst])                   \n"
+    "sw        $t2, 40(%[dst])                   \n"
+    "sw        $t3, 44(%[dst])                   \n"
+    "sw        $t4, 48(%[dst])                   \n"
+    "sw        $t5, 52(%[dst])                   \n"
+    "sw        $t6, 56(%[dst])                   \n"
+    "sw        $t7, 60(%[dst])                   \n"
+
+    "addiu     %[dst], %[dst], 64                \n"  // adding 64 to dest
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bne       %[dst], $a3, $loop16w             \n"
+    " addiu    %[src], %[src], 64                \n"  // adding 64 to src
+    "move      %[count], $t8                     \n"
+
+    // Here we have src and dest word-aligned but less than 64-bytes to go
+
+    "chk8w:                                      \n"
+    "pref      0, 0x0(%[src])                    \n"
+    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
+    // the t8 is the reminder count past 32-bytes
+    "beq       %[count], $t8, chk1w              \n"
+    // count=t8,no 32-byte chunk
+    " nop                                        \n"
+
+    "lw        $t0, 0(%[src])                    \n"
+    "lw        $t1, 4(%[src])                    \n"
+    "lw        $t2, 8(%[src])                    \n"
+    "lw        $t3, 12(%[src])                   \n"
+    "lw        $t4, 16(%[src])                   \n"
+    "lw        $t5, 20(%[src])                   \n"
+    "lw        $t6, 24(%[src])                   \n"
+    "lw        $t7, 28(%[src])                   \n"
+    "addiu     %[src], %[src], 32                \n"
+
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "addiu     %[dst], %[dst], 32                \n"
+
+    "chk1w:                                      \n"
+    "andi      %[count], $t8, 0x3                \n"
+    // now count is the reminder past 1w chunks
+    "beq       %[count], $t8, $last8             \n"
+    " subu     $a3, $t8, %[count]                \n"
+    // a3 is count of bytes in 1w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // now a3 is the dst address past the 1w chunks
+    // copying in words (4-byte chunks)
+    "$wordCopy_loop:                             \n"
+    "lw        $t3, 0(%[src])                    \n"
+    // the first t3 may be equal t0 ... optimize?
+    "addiu     %[src], %[src],4                  \n"
+    "addiu     %[dst], %[dst],4                  \n"
+    "bne       %[dst], $a3,$wordCopy_loop        \n"
+    " sw       $t3, -4(%[dst])                   \n"
+
+    // For the last (<8) bytes
+    "$last8:                                     \n"
+    "blez      %[count], leave                   \n"
+    " addu     $a3, %[dst], %[count]             \n"  // a3 -last dst address
+    "$last8loop:                                 \n"
+    "lb        $v1, 0(%[src])                    \n"
+    "addiu     %[src], %[src], 1                 \n"
+    "addiu     %[dst], %[dst], 1                 \n"
+    "bne       %[dst], $a3, $last8loop           \n"
+    " sb       $v1, -1(%[dst])                   \n"
+
+    "leave:                                      \n"
+    "  j       $ra                               \n"
+    "  nop                                       \n"
+
+    //
+    // UNALIGNED case
+    //
+
+    "unaligned:                                  \n"
+    // got here with a3="negu a1"
+    "andi      $a3, $a3, 0x3                     \n"  // a1 is word aligned?
+    "beqz      $a3, $ua_chk16w                   \n"
+    " subu     %[count], %[count], $a3           \n"
+    // bytes left after initial a3 bytes
+    "lwr       $v1, 0(%[src])                    \n"
+    "lwl       $v1, 3(%[src])                    \n"
+    "addu      %[src], %[src], $a3               \n"  // a3 may be 1, 2 or 3
+    "swr       $v1, 0(%[dst])                    \n"
+    "addu      %[dst], %[dst], $a3               \n"
+    // below the dst will be word aligned (NOTE1)
+    "$ua_chk16w:                                 \n"
+    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
+    // t8 is the byte count after 64-byte chunks
+    "beq       %[count], $t8, ua_chk8w           \n"
+    // if a2==t8, no 64-byte chunks
+    // There will be at most 1 32-byte chunk after it
+    "subu      $a3, %[count], $t8                \n"  // the reminder
+    // Here a3 counts bytes in 16w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // Now a3 is the final dst after 64-byte chunks
+    "addu      $t0, %[dst], %[count]             \n"  // t0 "past the end"
+    "subu      $t9, $t0, 160                     \n"
+    // t9 is the "last safe pref 30,128(a1)" address
+    "pref      0, 0(%[src])                      \n"  // first line of src
+    "pref      0, 32(%[src])                     \n"  // second line  addr 32
+    "pref      0, 64(%[src])                     \n"
+    "pref      30, 32(%[dst])                    \n"
+    // safe, as we have at least 64 bytes ahead
+    // In case the a1 > t9 don't use "pref 30" at all
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bgtz      $v1, $ua_loop16w                  \n"
+    // skip "pref 30,64(a1)" for too short arrays
+    " nop                                        \n"
+    // otherwise, start with using pref30
+    "pref      30, 64(%[dst])                    \n"
+    "$ua_loop16w:                                \n"
+    "pref      0, 96(%[src])                     \n"
+    "lwr       $t0, 0(%[src])                    \n"
+    "lwl       $t0, 3(%[src])                    \n"
+    "lwr       $t1, 4(%[src])                    \n"
+    "bgtz      $v1, $ua_skip_pref30_96           \n"
+    " lwl      $t1, 7(%[src])                    \n"
+    "pref      30, 96(%[dst])                    \n"
+    // continue setting up the dest, addr 96
+    "$ua_skip_pref30_96:                         \n"
+    "lwr       $t2, 8(%[src])                    \n"
+    "lwl       $t2, 11(%[src])                   \n"
+    "lwr       $t3, 12(%[src])                   \n"
+    "lwl       $t3, 15(%[src])                   \n"
+    "lwr       $t4, 16(%[src])                   \n"
+    "lwl       $t4, 19(%[src])                   \n"
+    "lwr       $t5, 20(%[src])                   \n"
+    "lwl       $t5, 23(%[src])                   \n"
+    "lwr       $t6, 24(%[src])                   \n"
+    "lwl       $t6, 27(%[src])                   \n"
+    "lwr       $t7, 28(%[src])                   \n"
+    "lwl       $t7, 31(%[src])                   \n"
+    "pref      0, 128(%[src])                    \n"
+    // bring the next lines of src, addr 128
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "lwr       $t0, 32(%[src])                   \n"
+    "lwl       $t0, 35(%[src])                   \n"
+    "lwr       $t1, 36(%[src])                   \n"
+    "bgtz      $v1, ua_skip_pref30_128           \n"
+    " lwl      $t1, 39(%[src])                   \n"
+    "pref      30, 128(%[dst])                   \n"
+    // continue setting up the dest, addr 128
+    "ua_skip_pref30_128:                         \n"
+
+    "lwr       $t2, 40(%[src])                   \n"
+    "lwl       $t2, 43(%[src])                   \n"
+    "lwr       $t3, 44(%[src])                   \n"
+    "lwl       $t3, 47(%[src])                   \n"
+    "lwr       $t4, 48(%[src])                   \n"
+    "lwl       $t4, 51(%[src])                   \n"
+    "lwr       $t5, 52(%[src])                   \n"
+    "lwl       $t5, 55(%[src])                   \n"
+    "lwr       $t6, 56(%[src])                   \n"
+    "lwl       $t6, 59(%[src])                   \n"
+    "lwr       $t7, 60(%[src])                   \n"
+    "lwl       $t7, 63(%[src])                   \n"
+    "pref      0, 160(%[src])                    \n"
+    // bring the next lines of src, addr 160
+    "sw        $t0, 32(%[dst])                   \n"
+    "sw        $t1, 36(%[dst])                   \n"
+    "sw        $t2, 40(%[dst])                   \n"
+    "sw        $t3, 44(%[dst])                   \n"
+    "sw        $t4, 48(%[dst])                   \n"
+    "sw        $t5, 52(%[dst])                   \n"
+    "sw        $t6, 56(%[dst])                   \n"
+    "sw        $t7, 60(%[dst])                   \n"
+
+    "addiu     %[dst],%[dst],64                  \n"  // adding 64 to dest
+    "sgtu      $v1,%[dst],$t9                    \n"
+    "bne       %[dst],$a3,$ua_loop16w            \n"
+    " addiu    %[src],%[src],64                  \n"  // adding 64 to src
+    "move      %[count],$t8                      \n"
+
+    // Here we have src and dest word-aligned but less than 64-bytes to go
+
+    "ua_chk8w:                                   \n"
+    "pref      0, 0x0(%[src])                    \n"
+    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
+    // the t8 is the reminder count
+    "beq       %[count], $t8, $ua_chk1w          \n"
+    // when count==t8, no 32-byte chunk
+
+    "lwr       $t0, 0(%[src])                    \n"
+    "lwl       $t0, 3(%[src])                    \n"
+    "lwr       $t1, 4(%[src])                    \n"
+    "lwl       $t1, 7(%[src])                    \n"
+    "lwr       $t2, 8(%[src])                    \n"
+    "lwl       $t2, 11(%[src])                   \n"
+    "lwr       $t3, 12(%[src])                   \n"
+    "lwl       $t3, 15(%[src])                   \n"
+    "lwr       $t4, 16(%[src])                   \n"
+    "lwl       $t4, 19(%[src])                   \n"
+    "lwr       $t5, 20(%[src])                   \n"
+    "lwl       $t5, 23(%[src])                   \n"
+    "lwr       $t6, 24(%[src])                   \n"
+    "lwl       $t6, 27(%[src])                   \n"
+    "lwr       $t7, 28(%[src])                   \n"
+    "lwl       $t7, 31(%[src])                   \n"
+    "addiu     %[src], %[src], 32                \n"
+
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "addiu     %[dst], %[dst], 32                \n"
+
+    "$ua_chk1w:                                  \n"
+    "andi      %[count], $t8, 0x3                \n"
+    // now count is the reminder past 1w chunks
+    "beq       %[count], $t8, ua_smallCopy       \n"
+    "subu      $a3, $t8, %[count]                \n"
+    // a3 is count of bytes in 1w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // now a3 is the dst address past the 1w chunks
+
+    // copying in words (4-byte chunks)
+    "$ua_wordCopy_loop:                          \n"
+    "lwr       $v1, 0(%[src])                    \n"
+    "lwl       $v1, 3(%[src])                    \n"
+    "addiu     %[src], %[src], 4                 \n"
+    "addiu     %[dst], %[dst], 4                 \n"
+    // note: dst=a1 is word aligned here, see NOTE1
+    "bne       %[dst], $a3, $ua_wordCopy_loop    \n"
+    " sw       $v1,-4(%[dst])                    \n"
+
+    // Now less than 4 bytes (value in count) left to copy
+    "ua_smallCopy:                               \n"
+    "beqz      %[count], leave                   \n"
+    " addu     $a3, %[dst], %[count]             \n" // a3 = last dst address
+    "$ua_smallCopy_loop:                         \n"
+    "lb        $v1, 0(%[src])                    \n"
+    "addiu     %[src], %[src], 1                 \n"
+    "addiu     %[dst], %[dst], 1                 \n"
+    "bne       %[dst],$a3,$ua_smallCopy_loop     \n"
+    " sb       $v1, -1(%[dst])                   \n"
+
+    "j         $ra                               \n"
+    " nop                                        \n"
+    ".set      at                                \n"
+    ".set      reorder                           \n"
+       : [dst] "+r" (dst), [src] "+r" (src)
+       : [count] "r" (count)
+       : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7",
+       "t8", "t9", "a3", "v1", "at"
+  );
+}
+#endif  // HAS_COPYROW_MIPS
+
+// MIPS DSPR2 functions
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_dsp) && \
+    (__mips_dsp_rev >= 2) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32) && (__mips_isa_rev < 6)
+
+void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                           int width) {
+  __asm__ __volatile__ (
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+    "srl             $t4, %[width], 4              \n"  // multiplies of 16
+    "blez            $t4, 2f                       \n"
+    " andi           %[width], %[width], 0xf       \n"  // residual
+
+    ".p2align        2                             \n"
+  "1:                                              \n"
+    "addiu           $t4, $t4, -1                  \n"
+    "lw              $t0, 0(%[src_uv])             \n"  // V1 | U1 | V0 | U0
+    "lw              $t1, 4(%[src_uv])             \n"  // V3 | U3 | V2 | U2
+    "lw              $t2, 8(%[src_uv])             \n"  // V5 | U5 | V4 | U4
+    "lw              $t3, 12(%[src_uv])            \n"  // V7 | U7 | V6 | U6
+    "lw              $t5, 16(%[src_uv])            \n"  // V9 | U9 | V8 | U8
+    "lw              $t6, 20(%[src_uv])            \n"  // V11 | U11 | V10 | U10
+    "lw              $t7, 24(%[src_uv])            \n"  // V13 | U13 | V12 | U12
+    "lw              $t8, 28(%[src_uv])            \n"  // V15 | U15 | V14 | U14
+    "addiu           %[src_uv], %[src_uv], 32      \n"
+    "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
+    "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
+    "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
+    "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
+    "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
+    "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
+    "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 | V12
+    "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 | U12
+    "sw              $t9, 0(%[dst_v])              \n"
+    "sw              $t0, 0(%[dst_u])              \n"
+    "sw              $t1, 4(%[dst_v])              \n"
+    "sw              $t2, 4(%[dst_u])              \n"
+    "sw              $t3, 8(%[dst_v])              \n"
+    "sw              $t5, 8(%[dst_u])              \n"
+    "sw              $t6, 12(%[dst_v])             \n"
+    "sw              $t7, 12(%[dst_u])             \n"
+    "addiu           %[dst_v], %[dst_v], 16        \n"
+    "bgtz            $t4, 1b                       \n"
+    " addiu          %[dst_u], %[dst_u], 16        \n"
+
+    "beqz            %[width], 3f                  \n"
+    " nop                                          \n"
+
+  "2:                                              \n"
+    "lbu             $t0, 0(%[src_uv])             \n"
+    "lbu             $t1, 1(%[src_uv])             \n"
+    "addiu           %[src_uv], %[src_uv], 2       \n"
+    "addiu           %[width], %[width], -1        \n"
+    "sb              $t0, 0(%[dst_u])              \n"
+    "sb              $t1, 0(%[dst_v])              \n"
+    "addiu           %[dst_u], %[dst_u], 1         \n"
+    "bgtz            %[width], 2b                  \n"
+    " addiu          %[dst_v], %[dst_v], 1         \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+     : [src_uv] "+r" (src_uv),
+       [width] "+r" (width),
+       [dst_u] "+r" (dst_u),
+       [dst_v] "+r" (dst_v)
+     :
+     : "t0", "t1", "t2", "t3",
+     "t4", "t5", "t6", "t7", "t8", "t9"
+  );
+}
+
+void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width) {
+  __asm__ __volatile__ (
+    ".set push                             \n"
+    ".set noreorder                        \n"
+
+    "srl       $t4, %[width], 4            \n"  // multiplies of 16
+    "andi      $t5, %[width], 0xf          \n"
+    "blez      $t4, 2f                     \n"
+    " addu     %[src], %[src], %[width]    \n"  // src += width
+
+    ".p2align  2                           \n"
+   "1:                                     \n"
+    "lw        $t0, -16(%[src])            \n"  // |3|2|1|0|
+    "lw        $t1, -12(%[src])            \n"  // |7|6|5|4|
+    "lw        $t2, -8(%[src])             \n"  // |11|10|9|8|
+    "lw        $t3, -4(%[src])             \n"  // |15|14|13|12|
+    "wsbh      $t0, $t0                    \n"  // |2|3|0|1|
+    "wsbh      $t1, $t1                    \n"  // |6|7|4|5|
+    "wsbh      $t2, $t2                    \n"  // |10|11|8|9|
+    "wsbh      $t3, $t3                    \n"  // |14|15|12|13|
+    "rotr      $t0, $t0, 16                \n"  // |0|1|2|3|
+    "rotr      $t1, $t1, 16                \n"  // |4|5|6|7|
+    "rotr      $t2, $t2, 16                \n"  // |8|9|10|11|
+    "rotr      $t3, $t3, 16                \n"  // |12|13|14|15|
+    "addiu     %[src], %[src], -16         \n"
+    "addiu     $t4, $t4, -1                \n"
+    "sw        $t3, 0(%[dst])              \n"  // |15|14|13|12|
+    "sw        $t2, 4(%[dst])              \n"  // |11|10|9|8|
+    "sw        $t1, 8(%[dst])              \n"  // |7|6|5|4|
+    "sw        $t0, 12(%[dst])             \n"  // |3|2|1|0|
+    "bgtz      $t4, 1b                     \n"
+    " addiu    %[dst], %[dst], 16          \n"
+    "beqz      $t5, 3f                     \n"
+    " nop                                  \n"
+
+   "2:                                     \n"
+    "lbu       $t0, -1(%[src])             \n"
+    "addiu     $t5, $t5, -1                \n"
+    "addiu     %[src], %[src], -1          \n"
+    "sb        $t0, 0(%[dst])              \n"
+    "bgez      $t5, 2b                     \n"
+    " addiu    %[dst], %[dst], 1           \n"
+
+   "3:                                     \n"
+    ".set pop                              \n"
+      : [src] "+r" (src), [dst] "+r" (dst)
+      : [width] "r" (width)
+      : "t0", "t1", "t2", "t3", "t4", "t5"
+  );
+}
+
+void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                            int width) {
+  int x = 0;
+  int y = 0;
+  __asm__ __volatile__ (
+    ".set push                                    \n"
+    ".set noreorder                               \n"
+
+    "addu            $t4, %[width], %[width]      \n"
+    "srl             %[x], %[width], 4            \n"
+    "andi            %[y], %[width], 0xf          \n"
+    "blez            %[x], 2f                     \n"
+    " addu           %[src_uv], %[src_uv], $t4    \n"
+
+    ".p2align        2                            \n"
+   "1:                                            \n"
+    "lw              $t0, -32(%[src_uv])          \n"  // |3|2|1|0|
+    "lw              $t1, -28(%[src_uv])          \n"  // |7|6|5|4|
+    "lw              $t2, -24(%[src_uv])          \n"  // |11|10|9|8|
+    "lw              $t3, -20(%[src_uv])          \n"  // |15|14|13|12|
+    "lw              $t4, -16(%[src_uv])          \n"  // |19|18|17|16|
+    "lw              $t6, -12(%[src_uv])          \n"  // |23|22|21|20|
+    "lw              $t7, -8(%[src_uv])           \n"  // |27|26|25|24|
+    "lw              $t8, -4(%[src_uv])           \n"  // |31|30|29|28|
+
+    "rotr            $t0, $t0, 16                 \n"  // |1|0|3|2|
+    "rotr            $t1, $t1, 16                 \n"  // |5|4|7|6|
+    "rotr            $t2, $t2, 16                 \n"  // |9|8|11|10|
+    "rotr            $t3, $t3, 16                 \n"  // |13|12|15|14|
+    "rotr            $t4, $t4, 16                 \n"  // |17|16|19|18|
+    "rotr            $t6, $t6, 16                 \n"  // |21|20|23|22|
+    "rotr            $t7, $t7, 16                 \n"  // |25|24|27|26|
+    "rotr            $t8, $t8, 16                 \n"  // |29|28|31|30|
+    "precr.qb.ph     $t9, $t0, $t1                \n"  // |0|2|4|6|
+    "precrq.qb.ph    $t5, $t0, $t1                \n"  // |1|3|5|7|
+    "precr.qb.ph     $t0, $t2, $t3                \n"  // |8|10|12|14|
+    "precrq.qb.ph    $t1, $t2, $t3                \n"  // |9|11|13|15|
+    "precr.qb.ph     $t2, $t4, $t6                \n"  // |16|18|20|22|
+    "precrq.qb.ph    $t3, $t4, $t6                \n"  // |17|19|21|23|
+    "precr.qb.ph     $t4, $t7, $t8                \n"  // |24|26|28|30|
+    "precrq.qb.ph    $t6, $t7, $t8                \n"  // |25|27|29|31|
+    "addiu           %[src_uv], %[src_uv], -32    \n"
+    "addiu           %[x], %[x], -1               \n"
+    "swr             $t4, 0(%[dst_u])             \n"
+    "swl             $t4, 3(%[dst_u])             \n"  // |30|28|26|24|
+    "swr             $t6, 0(%[dst_v])             \n"
+    "swl             $t6, 3(%[dst_v])             \n"  // |31|29|27|25|
+    "swr             $t2, 4(%[dst_u])             \n"
+    "swl             $t2, 7(%[dst_u])             \n"  // |22|20|18|16|
+    "swr             $t3, 4(%[dst_v])             \n"
+    "swl             $t3, 7(%[dst_v])             \n"  // |23|21|19|17|
+    "swr             $t0, 8(%[dst_u])             \n"
+    "swl             $t0, 11(%[dst_u])            \n"  // |14|12|10|8|
+    "swr             $t1, 8(%[dst_v])             \n"
+    "swl             $t1, 11(%[dst_v])            \n"  // |15|13|11|9|
+    "swr             $t9, 12(%[dst_u])            \n"
+    "swl             $t9, 15(%[dst_u])            \n"  // |6|4|2|0|
+    "swr             $t5, 12(%[dst_v])            \n"
+    "swl             $t5, 15(%[dst_v])            \n"  // |7|5|3|1|
+    "addiu           %[dst_v], %[dst_v], 16       \n"
+    "bgtz            %[x], 1b                     \n"
+    " addiu          %[dst_u], %[dst_u], 16       \n"
+    "beqz            %[y], 3f                     \n"
+    " nop                                         \n"
+    "b               2f                           \n"
+    " nop                                         \n"
+
+   "2:                                            \n"
+    "lbu             $t0, -2(%[src_uv])           \n"
+    "lbu             $t1, -1(%[src_uv])           \n"
+    "addiu           %[src_uv], %[src_uv], -2     \n"
+    "addiu           %[y], %[y], -1               \n"
+    "sb              $t0, 0(%[dst_u])             \n"
+    "sb              $t1, 0(%[dst_v])             \n"
+    "addiu           %[dst_u], %[dst_u], 1        \n"
+    "bgtz            %[y], 2b                     \n"
+    " addiu          %[dst_v], %[dst_v], 1        \n"
+
+   "3:                                            \n"
+    ".set pop                                     \n"
+      : [src_uv] "+r" (src_uv),
+        [dst_u] "+r" (dst_u),
+        [dst_v] "+r" (dst_v),
+        [x] "=&r" (x),
+        [y] "+r" (y)
+      : [width] "r" (width)
+      : "t0", "t1", "t2", "t3", "t4",
+      "t5", "t7", "t8", "t9"
+  );
+}
+
+// Convert (4 Y and 2 VU) I422 and arrange RGB values into
+// t5 = | 0 | B0 | 0 | b0 |
+// t4 = | 0 | B1 | 0 | b1 |
+// t9 = | 0 | G0 | 0 | g0 |
+// t8 = | 0 | G1 | 0 | g1 |
+// t2 = | 0 | R0 | 0 | r0 |
+// t1 = | 0 | R1 | 0 | r1 |
+#define I422ToTransientMipsRGB                                                 \
+      "lw                $t0, 0(%[y_buf])       \n"                            \
+      "lhu               $t1, 0(%[u_buf])       \n"                            \
+      "lhu               $t2, 0(%[v_buf])       \n"                            \
+      "preceu.ph.qbr     $t1, $t1               \n"                            \
+      "preceu.ph.qbr     $t2, $t2               \n"                            \
+      "preceu.ph.qbra    $t3, $t0               \n"                            \
+      "preceu.ph.qbla    $t0, $t0               \n"                            \
+      "subu.ph           $t1, $t1, $s5          \n"                            \
+      "subu.ph           $t2, $t2, $s5          \n"                            \
+      "subu.ph           $t3, $t3, $s4          \n"                            \
+      "subu.ph           $t0, $t0, $s4          \n"                            \
+      "mul.ph            $t3, $t3, $s0          \n"                            \
+      "mul.ph            $t0, $t0, $s0          \n"                            \
+      "shll.ph           $t4, $t1, 0x7          \n"                            \
+      "subu.ph           $t4, $t4, $t1          \n"                            \
+      "mul.ph            $t6, $t1, $s1          \n"                            \
+      "mul.ph            $t1, $t2, $s2          \n"                            \
+      "addq_s.ph         $t5, $t4, $t3          \n"                            \
+      "addq_s.ph         $t4, $t4, $t0          \n"                            \
+      "shra.ph           $t5, $t5, 6            \n"                            \
+      "shra.ph           $t4, $t4, 6            \n"                            \
+      "addiu             %[u_buf], 2            \n"                            \
+      "addiu             %[v_buf], 2            \n"                            \
+      "addu.ph           $t6, $t6, $t1          \n"                            \
+      "mul.ph            $t1, $t2, $s3          \n"                            \
+      "addu.ph           $t9, $t6, $t3          \n"                            \
+      "addu.ph           $t8, $t6, $t0          \n"                            \
+      "shra.ph           $t9, $t9, 6            \n"                            \
+      "shra.ph           $t8, $t8, 6            \n"                            \
+      "addu.ph           $t2, $t1, $t3          \n"                            \
+      "addu.ph           $t1, $t1, $t0          \n"                            \
+      "shra.ph           $t2, $t2, 6            \n"                            \
+      "shra.ph           $t1, $t1, 6            \n"                            \
+      "subu.ph           $t5, $t5, $s5          \n"                            \
+      "subu.ph           $t4, $t4, $s5          \n"                            \
+      "subu.ph           $t9, $t9, $s5          \n"                            \
+      "subu.ph           $t8, $t8, $s5          \n"                            \
+      "subu.ph           $t2, $t2, $s5          \n"                            \
+      "subu.ph           $t1, $t1, $s5          \n"                            \
+      "shll_s.ph         $t5, $t5, 8            \n"                            \
+      "shll_s.ph         $t4, $t4, 8            \n"                            \
+      "shll_s.ph         $t9, $t9, 8            \n"                            \
+      "shll_s.ph         $t8, $t8, 8            \n"                            \
+      "shll_s.ph         $t2, $t2, 8            \n"                            \
+      "shll_s.ph         $t1, $t1, 8            \n"                            \
+      "shra.ph           $t5, $t5, 8            \n"                            \
+      "shra.ph           $t4, $t4, 8            \n"                            \
+      "shra.ph           $t9, $t9, 8            \n"                            \
+      "shra.ph           $t8, $t8, 8            \n"                            \
+      "shra.ph           $t2, $t2, 8            \n"                            \
+      "shra.ph           $t1, $t1, 8            \n"                            \
+      "addu.ph           $t5, $t5, $s5          \n"                            \
+      "addu.ph           $t4, $t4, $s5          \n"                            \
+      "addu.ph           $t9, $t9, $s5          \n"                            \
+      "addu.ph           $t8, $t8, $s5          \n"                            \
+      "addu.ph           $t2, $t2, $s5          \n"                            \
+      "addu.ph           $t1, $t1, $s5          \n"
+
+void I422ToARGBRow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128| // clipping
+    "lui               $s6, 0xff00            \n"
+    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|ff|
+
+    ".p2align          2                      \n"
+   "1:                                        \n"
+      I422ToTransientMipsRGB
+// Arranging into argb format
+    "precr.qb.ph       $t4, $t8, $t4          \n"  // |G1|g1|B1|b1|
+    "precr.qb.ph       $t5, $t9, $t5          \n"  // |G0|g0|B0|b0|
+    "addiu             %[width], -4           \n"
+    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |G1|B1|G0|B0|
+    "precr.qb.ph       $t9, $t4, $t5          \n"  // |g1|b1|g0|b0|
+    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
+
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
+    "or                $t1, $t1, $s6          \n"  // |ff|R1|ff|R0|
+    "or                $t2, $t2, $s6          \n"  // |ff|r1|ff|r0|
+    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|r1|g1|b1|
+    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|R1|G1|B1|
+    "sll               $t9, $t9, 16           \n"
+    "sll               $t8, $t8, 16           \n"
+    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|r0|g0|b0|
+    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|R0|G0|B0|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+void I422ToABGRRow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128|
+    "lui               $s6, 0xff00            \n"
+    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|
+
+    ".p2align          2                       \n"
+   "1:                                         \n"
+      I422ToTransientMipsRGB
+// Arranging into abgr format
+    "precr.qb.ph      $t0, $t8, $t1           \n"  // |G1|g1|R1|r1|
+    "precr.qb.ph      $t3, $t9, $t2           \n"  // |G0|g0|R0|r0|
+    "precrq.qb.ph     $t8, $t0, $t3           \n"  // |G1|R1|G0|R0|
+    "precr.qb.ph      $t9, $t0, $t3           \n"  // |g1|r1|g0|r0|
+
+    "precr.qb.ph       $t2, $t4, $t5          \n"  // |B1|b1|B0|b0|
+    "addiu             %[width], -4           \n"
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |B1|0 |B0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |b1|0 |b0|
+    "or                $t1, $t1, $s6          \n"  // |ff|B1|ff|B0|
+    "or                $t2, $t2, $s6          \n"  // |ff|b1|ff|b0|
+    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|b1|g1|r1|
+    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|B1|G1|R1|
+    "sll               $t9, $t9, 16           \n"
+    "sll               $t8, $t8, 16           \n"
+    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|b0|g0|r0|
+    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|B0|G0|R0|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+void I422ToBGRARow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74 |74 |
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128|
+    "lui               $s6, 0xff              \n"
+    "ori               $s6, 0xff              \n"  // |00|ff|00|ff|
+
+    ".p2align          2                      \n"
+   "1:                                        \n"
+      I422ToTransientMipsRGB
+      // Arranging into bgra format
+    "precr.qb.ph       $t4, $t4, $t8          \n"  // |B1|b1|G1|g1|
+    "precr.qb.ph       $t5, $t5, $t9          \n"  // |B0|b0|G0|g0|
+    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |B1|G1|B0|G0|
+    "precr.qb.ph       $t9, $t4, $t5          \n"  // |b1|g1|b0|g0|
+
+    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
+    "addiu             %[width], -4           \n"
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
+    "sll               $t1, $t1, 8            \n"  // |R1|0 |R0|0 |
+    "sll               $t2, $t2, 8            \n"  // |r1|0 |r0|0 |
+    "or                $t1, $t1, $s6          \n"  // |R1|ff|R0|ff|
+    "or                $t2, $t2, $s6          \n"  // |r1|ff|r0|ff|
+    "precrq.ph.w       $t0, $t9, $t2          \n"  // |b1|g1|r1|ff|
+    "precrq.ph.w       $t3, $t8, $t1          \n"  // |B1|G1|R1|ff|
+    "sll               $t1, $t1, 16           \n"
+    "sll               $t2, $t2, 16           \n"
+    "packrl.ph         $t2, $t9, $t2          \n"  // |b0|g0|r0|ff|
+    "packrl.ph         $t1, $t8, $t1          \n"  // |B0|G0|R0|ff|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+// Bilinear filter 8x2 -> 8x1
+void InterpolateRow_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                               ptrdiff_t src_stride, int dst_width,
+                               int source_y_fraction) {
+    int y0_fraction = 256 - source_y_fraction;
+    const uint8* src_ptr1 = src_ptr + src_stride;
+
+  __asm__ __volatile__ (
+     ".set push                                           \n"
+     ".set noreorder                                      \n"
+
+     "replv.ph          $t0, %[y0_fraction]               \n"
+     "replv.ph          $t1, %[source_y_fraction]         \n"
+
+    ".p2align           2                                 \n"
+   "1:                                                    \n"
+     "lw                $t2, 0(%[src_ptr])                \n"
+     "lw                $t3, 0(%[src_ptr1])               \n"
+     "lw                $t4, 4(%[src_ptr])                \n"
+     "lw                $t5, 4(%[src_ptr1])               \n"
+     "muleu_s.ph.qbl    $t6, $t2, $t0                     \n"
+     "muleu_s.ph.qbr    $t7, $t2, $t0                     \n"
+     "muleu_s.ph.qbl    $t8, $t3, $t1                     \n"
+     "muleu_s.ph.qbr    $t9, $t3, $t1                     \n"
+     "muleu_s.ph.qbl    $t2, $t4, $t0                     \n"
+     "muleu_s.ph.qbr    $t3, $t4, $t0                     \n"
+     "muleu_s.ph.qbl    $t4, $t5, $t1                     \n"
+     "muleu_s.ph.qbr    $t5, $t5, $t1                     \n"
+     "addq.ph           $t6, $t6, $t8                     \n"
+     "addq.ph           $t7, $t7, $t9                     \n"
+     "addq.ph           $t2, $t2, $t4                     \n"
+     "addq.ph           $t3, $t3, $t5                     \n"
+     "shra.ph           $t6, $t6, 8                       \n"
+     "shra.ph           $t7, $t7, 8                       \n"
+     "shra.ph           $t2, $t2, 8                       \n"
+     "shra.ph           $t3, $t3, 8                       \n"
+     "precr.qb.ph       $t6, $t6, $t7                     \n"
+     "precr.qb.ph       $t2, $t2, $t3                     \n"
+     "addiu             %[src_ptr], %[src_ptr], 8         \n"
+     "addiu             %[src_ptr1], %[src_ptr1], 8       \n"
+     "addiu             %[dst_width], %[dst_width], -8    \n"
+     "sw                $t6, 0(%[dst_ptr])                \n"
+     "sw                $t2, 4(%[dst_ptr])                \n"
+     "bgtz              %[dst_width], 1b                  \n"
+     " addiu            %[dst_ptr], %[dst_ptr], 8         \n"
+
+     ".set pop                                            \n"
+  : [dst_ptr] "+r" (dst_ptr),
+    [src_ptr1] "+r" (src_ptr1),
+    [src_ptr] "+r" (src_ptr),
+    [dst_width] "+r" (dst_width)
+  : [source_y_fraction] "r" (source_y_fraction),
+    [y0_fraction] "r" (y0_fraction),
+    [src_stride] "r" (src_stride)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+#endif  // __mips_dsp_rev >= 2
+
+#endif  // defined(__mips__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/row_neon.cc b/libvpx/libvpx/third_party/libyuv/source/row_neon.cc
new file mode 100644
index 0000000..1a72eb9
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/row_neon.cc
@@ -0,0 +1,3084 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+// Read 8 Y, 4 U and 4 V from 422
+#define READYUV422                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.32    {d2[0]}, [%1]!                 \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.32    {d2[1]}, [%2]!                 \n"
+
+// Read 8 Y, 2 U and 2 V from 422
+#define READYUV411                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.16    {d2[0]}, [%1]!                 \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.16    {d2[1]}, [%2]!                 \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vzip.u8    d2, d3                         \n"
+
+// Read 8 Y, 8 U and 8 V from 444
+#define READYUV444                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.8     {d3}, [%2]!                    \n"                             \
+    "vpaddl.u8  q1, q1                         \n"                             \
+    "vrshrn.u16 d2, q1, #1                     \n"
+
+// Read 8 Y, and set 4 U and 4 V to 128
+#define READYUV400                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vmov.u8    d2, #128                       \n"
+
+// Read 8 Y and 4 UV from NV12
+#define READNV12                                                               \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 Y and 4 VU from NV21
+#define READNV21                                                               \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d3, d2                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 YUY2
+#define READYUY2                                                               \
+    MEMACCESS(0)                                                               \
+    "vld2.8     {d0, d2}, [%0]!                \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 UYVY
+#define READUYVY                                                               \
+    MEMACCESS(0)                                                               \
+    "vld2.8     {d2, d3}, [%0]!                \n"                             \
+    "vmov.u8    d0, d3                         \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+#define YUV422TORGB_SETUP_REG                                                  \
+    MEMACCESS([kUVToRB])                                                       \
+    "vld1.8     {d24}, [%[kUVToRB]]            \n"                             \
+    MEMACCESS([kUVToG])                                                        \
+    "vld1.8     {d25}, [%[kUVToG]]             \n"                             \
+    MEMACCESS([kUVBiasBGR])                                                    \
+    "vld1.16    {d26[], d27[]}, [%[kUVBiasBGR]]! \n"                           \
+    MEMACCESS([kUVBiasBGR])                                                    \
+    "vld1.16    {d8[], d9[]}, [%[kUVBiasBGR]]!   \n"                           \
+    MEMACCESS([kUVBiasBGR])                                                    \
+    "vld1.16    {d28[], d29[]}, [%[kUVBiasBGR]]  \n"                           \
+    MEMACCESS([kYToRgb])                                                       \
+    "vld1.32    {d30[], d31[]}, [%[kYToRgb]]     \n"
+
+#define YUV422TORGB                                                            \
+    "vmull.u8   q8, d2, d24                    \n" /* u/v B/R component      */\
+    "vmull.u8   q9, d2, d25                    \n" /* u/v G component        */\
+    "vmovl.u8   q0, d0                         \n" /* Y                      */\
+    "vmovl.s16  q10, d1                        \n"                             \
+    "vmovl.s16  q0, d0                         \n"                             \
+    "vmul.s32   q10, q10, q15                  \n"                             \
+    "vmul.s32   q0, q0, q15                    \n"                             \
+    "vqshrun.s32 d0, q0, #16                   \n"                             \
+    "vqshrun.s32 d1, q10, #16                  \n" /* Y                      */\
+    "vadd.s16   d18, d19                       \n"                             \
+    "vshll.u16  q1, d16, #16                   \n" /* Replicate u * UB       */\
+    "vshll.u16  q10, d17, #16                  \n" /* Replicate v * VR       */\
+    "vshll.u16  q3, d18, #16                   \n" /* Replicate (v*VG + u*UG)*/\
+    "vaddw.u16  q1, q1, d16                    \n"                             \
+    "vaddw.u16  q10, q10, d17                  \n"                             \
+    "vaddw.u16  q3, q3, d18                    \n"                             \
+    "vqadd.s16  q8, q0, q13                    \n" /* B */                     \
+    "vqadd.s16  q9, q0, q14                    \n" /* R */                     \
+    "vqadd.s16  q0, q0, q4                     \n" /* G */                     \
+    "vqadd.s16  q8, q8, q1                     \n" /* B */                     \
+    "vqadd.s16  q9, q9, q10                    \n" /* R */                     \
+    "vqsub.s16  q0, q0, q3                     \n" /* G */                     \
+    "vqshrun.s16 d20, q8, #6                   \n" /* B */                     \
+    "vqshrun.s16 d22, q9, #6                   \n" /* R */                     \
+    "vqshrun.s16 d21, q0, #6                   \n" /* G */
+
+// YUV to RGB conversion constants.
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* -min(128, round(2.018 * 64)) */
+#define UG 25 /* -round(-0.391 * 64) */
+#define VG 52 /* -round(-0.813 * 64) */
+#define VR -102 /* -round(1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            - YGB)
+#define BG (UG * 128 + VG * 128 - YGB)
+#define BR            (VR * 128 - YGB)
+
+static uvec8 kUVToRB  = { 128, 128, 128, 128, 102, 102, 102, 102,
+                          0, 0, 0, 0, 0, 0, 0, 0 };
+static uvec8 kUVToG = { 25, 25, 25, 25, 52, 52, 52, 52,
+                        0, 0, 0, 0, 0, 0, 0, 0 };
+static vec16 kUVBiasBGR = { BB, BG, BR, 0, 0, 0, 0, 0 };
+static vec32 kYToRgb = { 0x0101 * YG, 0, 0, 0 };
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV444
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV411
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d19, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_bgra),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_abgr),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d19, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgba),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    MEMACCESS(3)
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),      // %0
+      "+r"(src_u),      // %1
+      "+r"(src_v),      // %2
+      "+r"(dst_rgb24),  // %3
+      "+r"(width)       // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    MEMACCESS(3)
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_raw),  // %3
+      "+r"(width)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTORGB565                                                           \
+    "vshr.u8    d20, d20, #3                   \n"  /* B                    */ \
+    "vshr.u8    d21, d21, #2                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #11                  \n"  /* R                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q0, q0, q10                    \n"  /* BGR                  */
+
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_rgb565),  // %3
+      "+r"(width)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTOARGB1555                                                         \
+    "vshr.u8    q10, q10, #3                   \n"  /* B                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vshr.u8    d23, d23, #7                   \n"  /* A                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vmovl.u8   q11, d23                       \n"  /* A                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #10                  \n"  /* R                    */ \
+    "vshl.u16   q11, q11, #15                  \n"  /* A                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q1, q10, q11                   \n"  /* RA                   */ \
+    "vorr       q0, q0, q1                     \n"  /* BGRA                 */
+
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB1555
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb1555),  // %3
+      "+r"(width)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTOARGB4444                                                         \
+    "vshr.u8    d20, d20, #4                   \n"  /* B                    */ \
+    "vbic.32    d21, d21, d4                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #4                   \n"  /* R                    */ \
+    "vbic.32    d23, d23, d4                   \n"  /* A                    */ \
+    "vorr       d0, d20, d21                   \n"  /* BG                   */ \
+    "vorr       d1, d22, d23                   \n"  /* RA                   */ \
+    "vzip.u8    d0, d1                         \n"  /* BGRA                 */
+
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB4444
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb4444),  // %3
+      "+r"(width)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV400
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : [kUVToRB]"r"(&kUVToRB),   // %3
+      [kUVToG]"r"(&kUVToG),     // %4
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void J400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vmov.u8    d23, #255                      \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d20}, [%0]!                   \n"
+    "vmov       d21, d20                       \n"
+    "vmov       d22, d20                       \n"
+    "subs       %2, %2, #8                     \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    :
+    : "cc", "memory", "d20", "d21", "d22", "d23"
+  );
+}
+
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(2)
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(2)
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUY2
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_yuy2),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : [kUVToRB]"r"(&kUVToRB),   // %3
+      [kUVToG]"r"(&kUVToG),     // %4
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READUYVY
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_uyvy),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : [kUVToRB]"r"(&kUVToRB),   // %3
+      [kUVToG]"r"(&kUVToG),     // %4
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pairs of UV
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store U
+    MEMACCESS(2)
+    "vst1.8     {q1}, [%2]!                    \n"  // store V
+    "bgt        1b                             \n"
+    : "+r"(src_uv),  // %0
+      "+r"(dst_u),   // %1
+      "+r"(dst_v),   // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load U
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"  // load V
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    MEMACCESS(2)
+    "vst2.u8    {q0, q1}, [%2]!                \n"  // store 16 pairs of UV
+    "bgt        1b                             \n"
+    :
+      "+r"(src_u),   // %0
+      "+r"(src_v),   // %1
+      "+r"(dst_uv),  // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
+void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 32
+    "subs       %2, %2, #32                    \n"  // 32 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 32
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2  // Output registers
+  :                     // Input registers
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SetRow writes 'count' bytes using an 8 bit value repeated.
+void SetRow_NEON(uint8* dst, uint8 v8, int count) {
+  asm volatile (
+    "vdup.8    q0, %2                          \n"  // duplicate 16 bytes
+  "1:                                          \n"
+    "subs      %1, %1, #16                     \n"  // 16 bytes per loop
+    MEMACCESS(0)
+    "vst1.8    {q0}, [%0]!                     \n"  // store
+    "bgt       1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v8)      // %2
+  : "cc", "memory", "q0"
+  );
+}
+
+// ARGBSetRow writes 'count' pixels using an 32 bit value repeated.
+void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
+  "1:                                          \n"
+    "subs      %1, %1, #4                      \n"  // 4 pixels per loop
+    MEMACCESS(0)
+    "vst1.8    {q0}, [%0]!                     \n"  // store
+    "bgt       1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "q0"
+  );
+}
+
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2                     \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #16                        \n"  // 16 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r12, #-16                      \n"
+    "add        %0, %0, %3, lsl #1             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d0, d1}, [%0], r12            \n"  // src -= 16
+    "subs       %3, #8                         \n"  // 8 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // dst += 8
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_u),   // %1
+    "+r"(dst_v),   // %2
+    "+r"(width)    // %3
+  :
+  : "cc", "memory", "r12", "q0"
+  );
+}
+
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2, lsl #2             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #4                         \n"  // 4 pixels per loop.
+    "vrev64.32  q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),   // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    MEMACCESS(1)
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+#define RGB565TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxGGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB RRRRRxxx */ \
+    "vshl.u8    d6, d6, #2                     \n"  /* G GGGGGG00 upper 6   */ \
+    "vshr.u8    d1, d1, #3                     \n"  /* R 000RRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #6                     \n"  /* G 000000GG lower 2   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB1555TOARGB                                                         \
+    "vshrn.u16  d7, q0, #8                     \n"  /* A Arrrrrxx           */ \
+    "vshr.u8    d6, d7, #2                     \n"  /* R xxxRRRRR           */ \
+    "vshrn.u16  d5, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vmovn.u16  d4, q0                         \n"  /* B xxxBBBBB           */ \
+    "vshr.u8    d7, d7, #7                     \n"  /* A 0000000A           */ \
+    "vneg.s8    d7, d7                         \n"  /* A AAAAAAAA upper 8   */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* R RRRRR000 upper 5   */ \
+    "vshr.u8    q1, q3, #5                     \n"  /* R,A 00000RRR lower 3 */ \
+    "vshl.u8    q0, q2, #3                     \n"  /* B,G BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,G 00000BBB lower 3 */ \
+    "vorr.u8    q1, q1, q3                     \n"  /* R,A                  */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,G                  */ \
+
+// RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
+#define RGB555TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB xRRRRRxx */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* G GGGGG000 upper 5   */ \
+    "vshr.u8    d1, d1, #2                     \n"  /* R 00xRRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #5                     \n"  /* G 00000GGG lower 3   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB4444TOARGB                                                         \
+    "vuzp.u8    d0, d1                         \n"  /* d0 BG, d1 RA         */ \
+    "vshl.u8    q2, q0, #4                     \n"  /* B,R BBBB0000         */ \
+    "vshr.u8    q1, q0, #4                     \n"  /* G,A 0000GGGG         */ \
+    "vshr.u8    q0, q2, #4                     \n"  /* B,R 0000BBBB         */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,R BBBBBBBB         */ \
+    "vshl.u8    q2, q1, #4                     \n"  /* G,A GGGG0000         */ \
+    "vorr.u8    q1, q1, q2                     \n"  /* G,A GGGGGGGG         */ \
+    "vswp.u8    d1, d2                         \n"  /* B,R,G,A -> B,G,R,A   */
+
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RGB24.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_rgb24),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    MEMACCESS(1)
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RAW.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_raw),   // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of YUY2.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of UYVY.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // store 8 U.
+    MEMACCESS(2)
+    "vst1.8     {d3}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 U.
+    MEMACCESS(2)
+    "vst1.8     {d2}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_yuy2
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row YUY2.
+    "vrhadd.u8  d1, d1, d5                     \n"  // average rows of U
+    "vrhadd.u8  d3, d3, d7                     \n"  // average rows of V
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 U.
+    MEMACCESS(3)
+    "vst1.8     {d3}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),     // %0
+    "+r"(stride_yuy2),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_uyvy
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row UYVY.
+    "vrhadd.u8  d0, d0, d4                     \n"  // average rows of U
+    "vrhadd.u8  d2, d2, d6                     \n"  // average rows of V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 U.
+    MEMACCESS(3)
+    "vst1.8     {d2}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),     // %0
+    "+r"(stride_uyvy),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q2}, [%3]                     \n"  // shuffler
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 4 pixels.
+    "subs       %2, %2, #4                     \n"  // 4 processed per loop
+    "vtbl.8     d2, {d0, d1}, d4               \n"  // look up 2 first pixels
+    "vtbl.8     d3, {d0, d1}, d5               \n"  // look up 2 next pixels
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store 4.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d0, d2}, [%0]!                \n"  // load 16 Ys
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 Us
+    MEMACCESS(2)
+    "vld1.8     {d3}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    MEMACCESS(3)
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 YUY2/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_yuy2),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d1, d3}, [%0]!                \n"  // load 16 Ys
+    MEMACCESS(1)
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 Us
+    MEMACCESS(2)
+    "vld1.8     {d2}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    MEMACCESS(3)
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 UYVY/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_uyvy),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTORGB565
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_rgb565),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+    "vdup.32    d2, %2                         \n"  // dither4
+  "1:                                          \n"
+    MEMACCESS(1)
+    "vld4.8     {d20, d21, d22, d23}, [%1]!    \n"  // load 8 pixels of ARGB.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d20, d20, d2                   \n"
+    "vqadd.u8   d21, d21, d2                   \n"
+    "vqadd.u8   d22, d22, d2                   \n"
+    ARGBTORGB565
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+  : "+r"(dst_rgb)    // %0
+  : "r"(src_argb),   // %1
+    "r"(dither4),    // %2
+    "r"(width)       // %3
+  : "cc", "memory", "q0", "q1", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
+                            int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB1555
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb1555),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB4444
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb4444),  // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit Y
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+// 8x1 pixels.
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.u8    d24, #112                      \n"  // UB / VR 0.875 coefficient
+    "vmov.u8    d25, #74                       \n"  // UG -0.5781 coefficient
+    "vmov.u8    d26, #38                       \n"  // UR -0.2969 coefficient
+    "vmov.u8    d27, #18                       \n"  // VB -0.1406 coefficient
+    "vmov.u8    d28, #94                       \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlsl.u8   q2, d1, d25                    \n"  // G
+    "vmlsl.u8   q2, d2, d26                    \n"  // R
+    "vadd.u16   q2, q2, q15                    \n"  // +128 -> unsigned
+
+    "vmull.u8   q3, d2, d24                    \n"  // R
+    "vmlsl.u8   q3, d1, d28                    \n"  // G
+    "vmlsl.u8   q3, d0, d27                    \n"  // B
+    "vadd.u16   q3, q3, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q2, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q3, #8                    \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 32x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 32.
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(0)
+    "vld4.8     {d8, d10, d12, d14}, [%0]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d9, d11, d13, d15}, [%0]!     \n"  // load last 8 ARGB pixels.
+    "vpaddl.u8  q4, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q5, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q6, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vpadd.u16  d0, d0, d1                     \n"  // B 16 shorts -> 8 shorts.
+    "vpadd.u16  d1, d8, d9                     \n"  // B
+    "vpadd.u16  d2, d2, d3                     \n"  // G 16 shorts -> 8 shorts.
+    "vpadd.u16  d3, d10, d11                   \n"  // G
+    "vpadd.u16  d4, d4, d5                     \n"  // R 16 shorts -> 8 shorts.
+    "vpadd.u16  d5, d12, d13                   \n"  // R
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %3, %3, #32                    \n"  // 32 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#define RGBTOUV(QB, QG, QR) \
+    "vmul.s16   q8, " #QB ", q10               \n"  /* B                    */ \
+    "vmls.s16   q8, " #QG ", q11               \n"  /* G                    */ \
+    "vmls.s16   q8, " #QR ", q12               \n"  /* R                    */ \
+    "vadd.u16   q8, q8, q15                    \n"  /* +128 -> unsigned     */ \
+    "vmul.s16   q9, " #QR ", q10               \n"  /* R                    */ \
+    "vmls.s16   q9, " #QG ", q14               \n"  /* G                    */ \
+    "vmls.s16   q9, " #QB ", q13               \n"  /* B                    */ \
+    "vadd.u16   q9, q9, q15                    \n"  /* +128 -> unsigned     */ \
+    "vqshrn.u16  d0, q8, #8                    \n"  /* 16 bit to 8 bit U    */ \
+    "vqshrn.u16  d1, q9, #8                    \n"  /* 16 bit to 8 bit V    */
+
+// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): Subsample match C code.
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #127 / 2                  \n"  // UB / VR 0.500 coefficient
+    "vmov.s16   q11, #84 / 2                   \n"  // UG -0.33126 coefficient
+    "vmov.s16   q12, #43 / 2                   \n"  // UR -0.16874 coefficient
+    "vmov.s16   q13, #20 / 2                   \n"  // VB -0.08131 coefficient
+    "vmov.s16   q14, #107 / 2                  \n"  // VG -0.41869 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_bgra
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 BGRA pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 BGRA pixels.
+    "vpaddl.u8  q3, q3                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more BGRA pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 BGRA pixels.
+    "vpadal.u8  q3, q7                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q1, q1, #1                     \n"  // 2x average
+    "vrshr.u16  q2, q2, #1                     \n"
+    "vrshr.u16  q3, q3, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q3, q2, q1)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(src_stride_bgra),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_abgr
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ABGR pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ABGR pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ABGR pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ABGR pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_stride_abgr),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgba
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 RGBA pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 RGBA pixels.
+    "vpaddl.u8  q0, q1                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q3                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more RGBA pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 RGBA pixels.
+    "vpadal.u8  q0, q5                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q7                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(src_stride_rgba),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgb24
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RGB24 pixels.
+    MEMACCESS(0)
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RGB24 pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RGB24 pixels.
+    MEMACCESS(1)
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RGB24 pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_stride_rgb24),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_raw
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RAW pixels.
+    MEMACCESS(0)
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RAW pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RAW pixels.
+    MEMACCESS(1)
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RAW pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_stride_raw),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(src_stride_rgb565),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(src_stride_argb1555),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(src_stride_argb4444),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_y),       // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of BGRA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // R
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ABGR.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // R
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of RGBA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // B
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_NEON(uint8* dst_ptr,
+                         const uint8* src_ptr, ptrdiff_t src_stride,
+                         int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp        %4, #0                         \n"
+    "beq        100f                           \n"
+    "add        %2, %1                         \n"
+    "cmp        %4, #64                        \n"
+    "beq        75f                            \n"
+    "cmp        %4, #128                       \n"
+    "beq        50f                            \n"
+    "cmp        %4, #192                       \n"
+    "beq        25f                            \n"
+
+    "vdup.8     d5, %4                         \n"
+    "rsb        %4, #256                       \n"
+    "vdup.8     d4, %4                         \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vmull.u8   q13, d0, d4                    \n"
+    "vmull.u8   q14, d1, d4                    \n"
+    "vmlal.u8   q13, d2, d5                    \n"
+    "vmlal.u8   q14, d3, d5                    \n"
+    "vrshrn.u16 d0, q13, #8                    \n"
+    "vrshrn.u16 d1, q14, #8                    \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        1b                             \n"
+    "b          99f                            \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        25b                            \n"
+    "b          99f                            \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        50b                            \n"
+    "b          99f                            \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q0}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        75b                            \n"
+    "b          99f                            \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        100b                           \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14"
+  );
+}
+
+// dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
+void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "subs       %3, #8                         \n"
+    "blt        89f                            \n"
+    // Blend 8 pixels.
+  "8:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB0.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 pixels of ARGB1.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 pixels of ARGB.
+    "bge        8b                             \n"
+
+  "89:                                         \n"
+    "adds       %3, #8-1                       \n"
+    "blt        99f                            \n"
+
+    // Blend 1 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
+    MEMACCESS(1)
+    "vld4.8     {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
+    "subs       %3, %3, #1                     \n"  // 1 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    MEMACCESS(2)
+    "vst4.8     {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
+    "bge        1b                             \n"
+
+  "99:                                         \n"
+
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12"
+  );
+}
+
+// Attenuate 8 pixels at a time.
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    // Attenuate 8 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d0, d3                    \n"  // b * a
+    "vmull.u8   q11, d1, d3                    \n"  // g * a
+    "vmull.u8   q12, d2, d3                    \n"  // r * a
+    "vqrshrn.u16 d0, q10, #8                   \n"  // b >>= 8
+    "vqrshrn.u16 d1, q11, #8                   \n"  // g >>= 8
+    "vqrshrn.u16 d2, q12, #8                   \n"  // r >>= 8
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  :
+  : "cc", "memory", "q0", "q1", "q10", "q11", "q12"
+  );
+}
+
+// Quantize 8 ARGB pixels (32 bytes).
+// dst = (dst * scale >> 16) * interval_size + interval_offset;
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "vdup.u16   q8, %2                         \n"
+    "vshr.u16   q8, q8, #1                     \n"  // scale >>= 1
+    "vdup.u16   q9, %3                         \n"  // interval multiply.
+    "vdup.u16   q10, %4                        \n"  // interval add
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]         \n"  // load 8 pixels of ARGB.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q0, d0                         \n"  // b (0 .. 255)
+    "vmovl.u8   q1, d2                         \n"
+    "vmovl.u8   q2, d4                         \n"
+    "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
+    "vqdmulh.s16 q1, q1, q8                    \n"  // g
+    "vqdmulh.s16 q2, q2, q8                    \n"  // r
+    "vmul.u16   q0, q0, q9                     \n"  // b * interval_size
+    "vmul.u16   q1, q1, q9                     \n"  // g
+    "vmul.u16   q2, q2, q9                     \n"  // r
+    "vadd.u16   q0, q0, q10                    \n"  // b + interval_offset
+    "vadd.u16   q1, q1, q10                    \n"  // g
+    "vadd.u16   q2, q2, q10                    \n"  // r
+    "vqmovn.u16 d0, q0                         \n"
+    "vqmovn.u16 d2, q1                         \n"
+    "vqmovn.u16 d4, q2                         \n"
+    MEMACCESS(0)
+    "vst4.8     {d0, d2, d4, d6}, [%0]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10"
+  );
+}
+
+// Shade 8 pixels at a time by specified value.
+// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
+// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "vdup.u32   q0, %3                         \n"  // duplicate scale value.
+    "vzip.u8    d0, d1                         \n"  // d0 aarrggbb.
+    "vshr.u16   q0, q0, #1                     \n"  // scale / 2.
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d22, d24, d26}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q10, d20                       \n"  // b (0 .. 255)
+    "vmovl.u8   q11, d22                       \n"
+    "vmovl.u8   q12, d24                       \n"
+    "vmovl.u8   q13, d26                       \n"
+    "vqrdmulh.s16 q10, q10, d0[0]              \n"  // b * scale * 2
+    "vqrdmulh.s16 q11, q11, d0[1]              \n"  // g
+    "vqrdmulh.s16 q12, q12, d0[2]              \n"  // r
+    "vqrdmulh.s16 q13, q13, d0[3]              \n"  // a
+    "vqmovn.u16 d20, q10                       \n"
+    "vqmovn.u16 d22, q11                       \n"
+    "vqmovn.u16 d24, q12                       \n"
+    "vqmovn.u16 d26, q13                       \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d22, d24, d26}, [%1]!    \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),       // %0
+    "+r"(dst_argb),       // %1
+    "+r"(width)           // %2
+  : "r"(value)            // %3
+  : "cc", "memory", "q0", "q10", "q11", "q12", "q13"
+  );
+}
+
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+// Similar to ARGBToYJ but stores ARGB.
+// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit B
+    "vmov       d1, d0                         \n"  // G
+    "vmov       d2, d0                         \n"  // R
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d20, #17                       \n"  // BB coefficient
+    "vmov.u8    d21, #68                       \n"  // BG coefficient
+    "vmov.u8    d22, #35                       \n"  // BR coefficient
+    "vmov.u8    d24, #22                       \n"  // GB coefficient
+    "vmov.u8    d25, #88                       \n"  // GG coefficient
+    "vmov.u8    d26, #45                       \n"  // GR coefficient
+    "vmov.u8    d28, #24                       \n"  // BB coefficient
+    "vmov.u8    d29, #98                       \n"  // BG coefficient
+    "vmov.u8    d30, #50                       \n"  // BR coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]         \n"  // load 8 ARGB pixels.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d20                    \n"  // B to Sepia B
+    "vmlal.u8   q2, d1, d21                    \n"  // G
+    "vmlal.u8   q2, d2, d22                    \n"  // R
+    "vmull.u8   q3, d0, d24                    \n"  // B to Sepia G
+    "vmlal.u8   q3, d1, d25                    \n"  // G
+    "vmlal.u8   q3, d2, d26                    \n"  // R
+    "vmull.u8   q8, d0, d28                    \n"  // B to Sepia R
+    "vmlal.u8   q8, d1, d29                    \n"  // G
+    "vmlal.u8   q8, d2, d30                    \n"  // R
+    "vqshrn.u16 d0, q2, #7                     \n"  // 16 bit to 8 bit B
+    "vqshrn.u16 d1, q3, #7                     \n"  // 16 bit to 8 bit G
+    "vqshrn.u16 d2, q8, #7                     \n"  // 16 bit to 8 bit R
+    MEMACCESS(0)
+    "vst4.8     {d0, d1, d2, d3}, [%0]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),  // %0
+    "+r"(width)      // %1
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// TODO(fbarchard): Was same as Sepia except matrix is provided.  This function
+// needs to saturate.  Consider doing a non-saturating version.
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q2}, [%3]                     \n"  // load 3 ARGB vectors.
+    "vmovl.s8   q0, d4                         \n"  // B,G coefficients s16.
+    "vmovl.s8   q1, d5                         \n"  // R,A coefficients s16.
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d16, d18, d20, d22}, [%0]!    \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q8, d16                        \n"  // b (0 .. 255) 16 bit
+    "vmovl.u8   q9, d18                        \n"  // g
+    "vmovl.u8   q10, d20                       \n"  // r
+    "vmovl.u8   q11, d22                       \n"  // a
+    "vmul.s16   q12, q8, d0[0]                 \n"  // B = B * Matrix B
+    "vmul.s16   q13, q8, d1[0]                 \n"  // G = B * Matrix G
+    "vmul.s16   q14, q8, d2[0]                 \n"  // R = B * Matrix R
+    "vmul.s16   q15, q8, d3[0]                 \n"  // A = B * Matrix A
+    "vmul.s16   q4, q9, d0[1]                  \n"  // B += G * Matrix B
+    "vmul.s16   q5, q9, d1[1]                  \n"  // G += G * Matrix G
+    "vmul.s16   q6, q9, d2[1]                  \n"  // R += G * Matrix R
+    "vmul.s16   q7, q9, d3[1]                  \n"  // A += G * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q10, d0[2]                 \n"  // B += R * Matrix B
+    "vmul.s16   q5, q10, d1[2]                 \n"  // G += R * Matrix G
+    "vmul.s16   q6, q10, d2[2]                 \n"  // R += R * Matrix R
+    "vmul.s16   q7, q10, d3[2]                 \n"  // A += R * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q11, d0[3]                 \n"  // B += A * Matrix B
+    "vmul.s16   q5, q11, d1[3]                 \n"  // G += A * Matrix G
+    "vmul.s16   q6, q11, d2[3]                 \n"  // R += A * Matrix R
+    "vmul.s16   q7, q11, d3[3]                 \n"  // A += A * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vqshrun.s16 d16, q12, #6                  \n"  // 16 bit to 8 bit B
+    "vqshrun.s16 d18, q13, #6                  \n"  // 16 bit to 8 bit G
+    "vqshrun.s16 d20, q14, #6                  \n"  // 16 bit to 8 bit R
+    "vqshrun.s16 d22, q15, #6                  \n"  // 16 bit to 8 bit A
+    MEMACCESS(1)
+    "vst4.8     {d16, d18, d20, d22}, [%1]!    \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "r"(matrix_argb)  // %3
+  : "cc", "memory", "q0", "q1", "q2", "q4", "q5", "q6", "q7", "q8", "q9",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable.
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d1, d3, d5, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q0, d0, d1                     \n"  // multiply B
+    "vmull.u8   q1, d2, d3                     \n"  // multiply G
+    "vmull.u8   q2, d4, d5                     \n"  // multiply R
+    "vmull.u8   q3, d6, d7                     \n"  // multiply A
+    "vrshrn.u16 d0, q0, #8                     \n"  // 16 bit to 8 bit B
+    "vrshrn.u16 d1, q1, #8                     \n"  // 16 bit to 8 bit G
+    "vrshrn.u16 d2, q2, #8                     \n"  // 16 bit to 8 bit R
+    "vrshrn.u16 d3, q3, #8                     \n"  // 16 bit to 8 bit A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_NEON
+
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   q0, q0, q2                     \n"  // add B, G
+    "vqadd.u8   q1, q1, q3                     \n"  // add R, A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqsub.u8   q0, q0, q2                     \n"  // subtract B, G
+    "vqsub.u8   q1, q1, q3                     \n"  // subtract R, A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0]!                    \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d0, d0, d1                     \n"  // add
+    "vmov.u8    d1, d0                         \n"
+    "vmov.u8    d2, d0                         \n"
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// Adds Sobel X and Sobel Y and stores Sobel into plane.
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    // 16 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 16 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"  // load 16 sobely.
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vqadd.u8   q0, q0, q1                     \n"  // add
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0]!                    \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d1, d0, d2                     \n"  // add
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0],%5                  \n"  // top
+    MEMACCESS(0)
+    "vld1.8     {d1}, [%0],%6                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    MEMACCESS(1)
+    "vld1.8     {d2}, [%1],%5                  \n"  // center * 2
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%6                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    MEMACCESS(2)
+    "vld1.8     {d2}, [%2],%5                  \n"  // bottom
+    MEMACCESS(2)
+    "vld1.8     {d3}, [%2],%6                  \n"
+    "subs       %4, %4, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    MEMACCESS(3)
+    "vst1.8     {d0}, [%3]!                    \n"  // store 8 sobelx
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  : "r"(2),            // %5
+    "r"(6)             // %6
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0],%4                  \n"  // left
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1],%4                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0],%4                  \n"  // center * 2
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%4                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0],%5                  \n"  // right
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%5                  \n"
+    "subs       %3, %3, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 sobely
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  : "r"(1),            // %4
+    "r"(6)             // %5
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/row_neon64.cc b/libvpx/libvpx/third_party/libyuv/source/row_neon64.cc
new file mode 100644
index 0000000..5d01545
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/row_neon64.cc
@@ -0,0 +1,3087 @@
+/*
+ *  Copyright 2014 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon armv8 64 bit.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+// Read 8 Y, 4 U and 4 V from 422
+#define READYUV422                                                             \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    MEMACCESS(1)                                                               \
+    "ld1        {v1.s}[0], [%1], #4            \n"                             \
+    MEMACCESS(2)                                                               \
+    "ld1        {v1.s}[1], [%2], #4            \n"
+
+// Read 8 Y, 2 U and 2 V from 422
+#define READYUV411                                                             \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    MEMACCESS(1)                                                               \
+    "ld1        {v2.h}[0], [%1], #2            \n"                             \
+    MEMACCESS(2)                                                               \
+    "ld1        {v2.h}[1], [%2], #2            \n"                             \
+    "zip1       v1.8b, v2.8b, v2.8b            \n"
+
+// Read 8 Y, 8 U and 8 V from 444
+#define READYUV444                                                             \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    MEMACCESS(1)                                                               \
+    "ld1        {v1.d}[0], [%1], #8            \n"                             \
+    MEMACCESS(2)                                                               \
+    "ld1        {v1.d}[1], [%2], #8            \n"                             \
+    "uaddlp     v1.8h, v1.16b                  \n"                             \
+    "rshrn      v1.8b, v1.8h, #1               \n"
+
+// Read 8 Y, and set 4 U and 4 V to 128
+#define READYUV400                                                             \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    "movi       v1.8b , #128                   \n"
+
+// Read 8 Y and 4 UV from NV12
+#define READNV12                                                               \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    MEMACCESS(1)                                                               \
+    "ld1        {v2.8b}, [%1], #8              \n"                             \
+    "uzp1       v1.8b, v2.8b, v2.8b            \n"                             \
+    "uzp2       v3.8b, v2.8b, v2.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
+
+// Read 8 Y and 4 VU from NV21
+#define READNV21                                                               \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    MEMACCESS(1)                                                               \
+    "ld1        {v2.8b}, [%1], #8              \n"                             \
+    "uzp1       v3.8b, v2.8b, v2.8b            \n"                             \
+    "uzp2       v1.8b, v2.8b, v2.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
+
+// Read 8 YUY2
+#define READYUY2                                                               \
+    MEMACCESS(0)                                                               \
+    "ld2        {v0.8b, v1.8b}, [%0], #16      \n"                             \
+    "uzp2       v3.8b, v1.8b, v1.8b            \n"                             \
+    "uzp1       v1.8b, v1.8b, v1.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
+
+// Read 8 UYVY
+#define READUYVY                                                               \
+    MEMACCESS(0)                                                               \
+    "ld2        {v2.8b, v3.8b}, [%0], #16      \n"                             \
+    "orr        v0.8b, v3.8b, v3.8b            \n"                             \
+    "uzp1       v1.8b, v2.8b, v2.8b            \n"                             \
+    "uzp2       v3.8b, v2.8b, v2.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
+
+#define YUV422TORGB_SETUP_REG                                                  \
+    "ld1r       {v24.8h}, [%[kUVBiasBGR]], #2  \n"                             \
+    "ld1r       {v25.8h}, [%[kUVBiasBGR]], #2  \n"                             \
+    "ld1r       {v26.8h}, [%[kUVBiasBGR]]      \n"                             \
+    "ld1r       {v31.4s}, [%[kYToRgb]]         \n"                             \
+    "movi       v27.8h, #128                   \n"                             \
+    "movi       v28.8h, #102                   \n"                             \
+    "movi       v29.8h, #25                    \n"                             \
+    "movi       v30.8h, #52                    \n"
+
+#define YUV422TORGB(vR, vG, vB)                                                \
+    "uxtl       v0.8h, v0.8b                   \n" /* Extract Y    */          \
+    "shll       v2.8h, v1.8b, #8               \n" /* Replicate UV */          \
+    "ushll2     v3.4s, v0.8h, #0               \n" /* Y */                     \
+    "ushll      v0.4s, v0.4h, #0               \n"                             \
+    "mul        v3.4s, v3.4s, v31.4s           \n"                             \
+    "mul        v0.4s, v0.4s, v31.4s           \n"                             \
+    "sqshrun    v0.4h, v0.4s, #16              \n"                             \
+    "sqshrun2   v0.8h, v3.4s, #16              \n" /* Y */                     \
+    "uaddw      v1.8h, v2.8h, v1.8b            \n" /* Replicate UV */          \
+    "mov        v2.d[0], v1.d[1]               \n" /* Extract V */             \
+    "uxtl       v2.8h, v2.8b                   \n"                             \
+    "uxtl       v1.8h, v1.8b                   \n" /* Extract U */             \
+    "mul        v3.8h, v1.8h, v27.8h           \n"                             \
+    "mul        v5.8h, v1.8h, v29.8h           \n"                             \
+    "mul        v6.8h, v2.8h, v30.8h           \n"                             \
+    "mul        v7.8h, v2.8h, v28.8h           \n"                             \
+    "sqadd      v6.8h, v6.8h, v5.8h            \n"                             \
+    "sqadd      " #vB ".8h, v24.8h, v0.8h      \n" /* B */                     \
+    "sqadd      " #vG ".8h, v25.8h, v0.8h      \n" /* G */                     \
+    "sqadd      " #vR ".8h, v26.8h, v0.8h      \n" /* R */                     \
+    "sqadd      " #vB ".8h, " #vB ".8h, v3.8h  \n" /* B */                     \
+    "sqsub      " #vG ".8h, " #vG ".8h, v6.8h  \n" /* G */                     \
+    "sqadd      " #vR ".8h, " #vR ".8h, v7.8h  \n" /* R */                     \
+    "sqshrun    " #vB ".8b, " #vB ".8h, #6     \n" /* B */                     \
+    "sqshrun    " #vG ".8b, " #vG ".8h, #6     \n" /* G */                     \
+    "sqshrun    " #vR ".8b, " #vR ".8h, #6     \n" /* R */                     \
+
+// YUV to RGB conversion constants.
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* -min(128, round(2.018 * 64)) */
+#define UG 25 /* -round(-0.391 * 64) */
+#define VG 52 /* -round(-0.813 * 64) */
+#define VR -102 /* -round(1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            - YGB)
+#define BG (UG * 128 + VG * 128 - YGB)
+#define BR            (VR * 128 - YGB)
+
+static vec16 kUVBiasBGR = { BB, BG, BR, 0, 0, 0, 0, 0 };
+static vec32 kYToRgb = { 0x0101 * YG, 0, 0, 0 };
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+#define RGBTOUV_SETUP_REG                                                      \
+    "movi       v20.8h, #56, lsl #0  \n"  /* UB/VR coefficient (0.875) / 2 */  \
+    "movi       v21.8h, #37, lsl #0  \n"  /* UG coefficient (-0.5781) / 2  */  \
+    "movi       v22.8h, #19, lsl #0  \n"  /* UR coefficient (-0.2969) / 2  */  \
+    "movi       v23.8h, #9,  lsl #0  \n"  /* VB coefficient (-0.1406) / 2  */  \
+    "movi       v24.8h, #47, lsl #0  \n"  /* VG coefficient (-0.7344) / 2  */  \
+    "movi       v25.16b, #0x80       \n"  /* 128.5 (0x8080 in 16-bit)      */
+
+
+#ifdef HAS_I444TOARGBROW_NEON
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV444
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                 \n"
+    "movi       v23.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I444TOARGBROW_NEON
+
+#ifdef HAS_I422TOARGBROW_NEON
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v23.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TOARGBROW_NEON
+
+#ifdef HAS_I411TOARGBROW_NEON
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV411
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v23.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I411TOARGBROW_NEON
+
+#ifdef HAS_I422TOBGRAROW_NEON
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v21, v22, v23)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v20.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_bgra),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TOBGRAROW_NEON
+
+#ifdef HAS_I422TOABGRROW_NEON
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v20, v21, v22)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v23.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_abgr),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TOABGRROW_NEON
+
+#ifdef HAS_I422TORGBAROW_NEON
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v23, v22, v21)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v20.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgba),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TORGBAROW_NEON
+
+#ifdef HAS_I422TORGB24ROW_NEON
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    MEMACCESS(3)
+    "st3        {v20.8b,v21.8b,v22.8b}, [%3], #24     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgb24), // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TORGB24ROW_NEON
+
+#ifdef HAS_I422TORAWROW_NEON
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v20, v21, v22)
+    "subs       %w4, %w4, #8                   \n"
+    MEMACCESS(3)
+    "st3        {v20.8b,v21.8b,v22.8b}, [%3], #24     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_raw),   // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TORAWROW_NEON
+
+#define ARGBTORGB565                                                           \
+    "shll       v0.8h,  v22.8b, #8             \n"  /* R                    */ \
+    "shll       v20.8h, v20.8b, #8             \n"  /* B                    */ \
+    "shll       v21.8h, v21.8b, #8             \n"  /* G                    */ \
+    "sri        v0.8h,  v21.8h, #5             \n"  /* RG                   */ \
+    "sri        v0.8h,  v20.8h, #11            \n"  /* RGB                  */
+
+#ifdef HAS_I422TORGB565ROW_NEON
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    ARGBTORGB565
+    MEMACCESS(3)
+    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_rgb565),  // %3
+      "+r"(width)     // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TORGB565ROW_NEON
+
+#define ARGBTOARGB1555                                                         \
+    "shll       v0.8h,  v23.8b, #8             \n"  /* A                    */ \
+    "shll       v22.8h, v22.8b, #8             \n"  /* R                    */ \
+    "shll       v20.8h, v20.8b, #8             \n"  /* B                    */ \
+    "shll       v21.8h, v21.8b, #8             \n"  /* G                    */ \
+    "sri        v0.8h,  v22.8h, #1             \n"  /* AR                   */ \
+    "sri        v0.8h,  v21.8h, #6             \n"  /* ARG                  */ \
+    "sri        v0.8h,  v20.8h, #11            \n"  /* ARGB                 */
+
+#ifdef HAS_I422TOARGB1555ROW_NEON
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    ARGBTOARGB1555
+    MEMACCESS(3)
+    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb1555),  // %3
+      "+r"(width)     // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TOARGB1555ROW_NEON
+
+#define ARGBTOARGB4444                                                         \
+    /* Input v20.8b<=B, v21.8b<=G, v22.8b<=R, v23.8b<=A, v4.8b<=0x0f        */ \
+    "ushr       v20.8b, v20.8b, #4             \n"  /* B                    */ \
+    "bic        v21.8b, v21.8b, v4.8b          \n"  /* G                    */ \
+    "ushr       v22.8b, v22.8b, #4             \n"  /* R                    */ \
+    "bic        v23.8b, v23.8b, v4.8b          \n"  /* A                    */ \
+    "orr        v0.8b,  v20.8b, v21.8b         \n"  /* BG                   */ \
+    "orr        v1.8b,  v22.8b, v23.8b         \n"  /* RA                   */ \
+    "zip1       v0.16b, v0.16b, v1.16b         \n"  /* BGRA                 */
+
+#ifdef HAS_I422TOARGB4444ROW_NEON
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    "movi       v4.16b, #0x0f                  \n"  // bits to clear with vbic.
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    ARGBTOARGB4444
+    MEMACCESS(3)
+    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels ARGB4444.
+    "b.gt       1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb4444),  // %3
+      "+r"(width)     // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TOARGB4444ROW_NEON
+
+#ifdef HAS_I400TOARGBROW_NEON
+void I400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  int64 width64 = (int64)(width);
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV400
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w2, %w2, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    MEMACCESS(1)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width64)    // %2
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I400TOARGBROW_NEON
+
+#ifdef HAS_J400TOARGBROW_NEON
+void J400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "movi       v23.8b, #255                   \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v20.8b}, [%0], #8             \n"
+    "orr        v21.8b, v20.8b, v20.8b         \n"
+    "orr        v22.8b, v20.8b, v20.8b         \n"
+    "subs       %w2, %w2, #8                   \n"
+    MEMACCESS(1)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    :
+    : "cc", "memory", "v20", "v21", "v22", "v23"
+  );
+}
+#endif  // HAS_J400TOARGBROW_NEON
+
+#ifdef HAS_NV12TOARGBROW_NEON
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w3, %w3, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    MEMACCESS(2)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_NV12TOARGBROW_NEON
+
+#ifdef HAS_NV21TOARGBROW_NEON
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w3, %w3, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    MEMACCESS(2)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_NV21TOARGBROW_NEON
+
+#ifdef HAS_NV12TORGB565ROW_NEON
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w3, %w3, #8                   \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "st1        {v0.8h}, [%2], 16              \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_NV12TORGB565ROW_NEON
+
+#ifdef HAS_NV21TORGB565ROW_NEON
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w3, %w3, #8                   \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "st1        {v0.8h}, [%2], 16              \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_NV21TORGB565ROW_NEON
+
+#ifdef HAS_YUY2TOARGBROW_NEON
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width) {
+  int64 width64 = (int64)(width);
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUY2
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w2, %w2, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    MEMACCESS(1)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32      \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_yuy2),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width64)    // %2
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_YUY2TOARGBROW_NEON
+
+#ifdef HAS_UYVYTOARGBROW_NEON
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width) {
+  int64 width64 = (int64)(width);
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READUYVY
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w2, %w2, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    MEMACCESS(1)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], 32      \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_uyvy),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width64)    // %2
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_UYVYTOARGBROW_NEON
+
+// Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
+#ifdef HAS_SPLITUVROW_NEON
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                     int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pairs of UV
+    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"  // store U
+    MEMACCESS(2)
+    "st1        {v1.16b}, [%2], #16            \n"  // store V
+    "b.gt       1b                             \n"
+    : "+r"(src_uv),  // %0
+      "+r"(dst_u),   // %1
+      "+r"(dst_v),   // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+#endif  // HAS_SPLITUVROW_NEON
+
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+#ifdef HAS_MERGEUVROW_NEON
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load U
+    MEMACCESS(1)
+    "ld1        {v1.16b}, [%1], #16            \n"  // load V
+    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
+    MEMACCESS(2)
+    "st2        {v0.16b,v1.16b}, [%2], #32     \n"  // store 16 pairs of UV
+    "b.gt       1b                             \n"
+    :
+      "+r"(src_u),   // %0
+      "+r"(src_v),   // %1
+      "+r"(dst_uv),  // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+#endif  // HAS_MERGEUVROW_NEON
+
+// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
+#ifdef HAS_COPYROW_NEON
+void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32       \n"  // load 32
+    "subs       %w2, %w2, #32                  \n"  // 32 processed per loop
+    MEMACCESS(1)
+    "st1        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32       \n"  // store 32
+    "b.gt       1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2  // Output registers
+  :                     // Input registers
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_COPYROW_NEON
+
+// SetRow writes 'count' bytes using an 8 bit value repeated.
+void SetRow_NEON(uint8* dst, uint8 v8, int count) {
+  asm volatile (
+    "dup        v0.16b, %w2                    \n"  // duplicate 16 bytes
+  "1:                                          \n"
+    "subs      %w1, %w1, #16                   \n"  // 16 bytes per loop
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"  // store
+    "b.gt      1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v8)      // %2
+  : "cc", "memory", "v0"
+  );
+}
+
+void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "dup        v0.4s, %w2                     \n"  // duplicate 4 ints
+  "1:                                          \n"
+    "subs      %w1, %w1, #4                    \n"  // 4 ints per loop
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"  // store
+    "b.gt      1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "v0"
+  );
+}
+
+#ifdef HAS_MIRRORROW_NEON
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  int64 width64 = (int64) width;
+  asm volatile (
+    // Start at end of source row.
+    "add        %0, %0, %2                     \n"
+    "sub        %0, %0, #16                    \n"
+
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], %3             \n"  // src -= 16
+    "subs       %2, %2, #16                   \n"  // 16 pixels per loop.
+    "rev64      v0.16b, v0.16b                 \n"
+    MEMACCESS(1)
+    "st1        {v0.D}[1], [%1], #8            \n"  // dst += 16
+    MEMACCESS(1)
+    "st1        {v0.D}[0], [%1], #8            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width64)  // %2
+  : "r"((ptrdiff_t)-16)    // %3
+  : "cc", "memory", "v0"
+  );
+}
+#endif  // HAS_MIRRORROW_NEON
+
+#ifdef HAS_MIRRORUVROW_NEON
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width) {
+  int64 width64 = (int64) width;
+  asm volatile (
+    // Start at end of source row.
+    "add        %0, %0, %3, lsl #1             \n"
+    "sub        %0, %0, #16                    \n"
+
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v0.8b, v1.8b}, [%0], %4       \n"  // src -= 16
+    "subs       %3, %3, #8                     \n"  // 8 pixels per loop.
+    "rev64      v0.8b, v0.8b                   \n"
+    "rev64      v1.8b, v1.8b                   \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // dst += 8
+    MEMACCESS(2)
+    "st1        {v1.8b}, [%2], #8              \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_u),   // %1
+    "+r"(dst_v),   // %2
+    "+r"(width64)    // %3
+  : "r"((ptrdiff_t)-16)      // %4
+  : "cc", "memory", "v0", "v1"
+  );
+}
+#endif  // HAS_MIRRORUVROW_NEON
+
+#ifdef HAS_ARGBMIRRORROW_NEON
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  int64 width64 = (int64) width;
+  asm volatile (
+    // Start at end of source row.
+    "add        %0, %0, %2, lsl #2             \n"
+    "sub        %0, %0, #16                    \n"
+
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], %3             \n"  // src -= 16
+    "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
+    "rev64      v0.4s, v0.4s                   \n"
+    MEMACCESS(1)
+    "st1        {v0.D}[1], [%1], #8            \n"  // dst += 16
+    MEMACCESS(1)
+    "st1        {v0.D}[0], [%1], #8            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width64)  // %2
+  : "r"((ptrdiff_t)-16)    // %3
+  : "cc", "memory", "v0"
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_NEON
+
+#ifdef HAS_RGB24TOARGBROW_NEON
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "movi       v4.8b, #255                    \n"  // Alpha
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v1.8b,v2.8b,v3.8b}, [%0], #24 \n"  // load 8 pixels of RGB24.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "st4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),   // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "v1", "v2", "v3", "v4"  // Clobber List
+  );
+}
+#endif  // HAS_RGB24TOARGBROW_NEON
+
+#ifdef HAS_RAWTOARGBROW_NEON
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "movi       v5.8b, #255                    \n"  // Alpha
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "orr        v3.8b, v1.8b, v1.8b            \n"  // move g
+    "orr        v4.8b, v0.8b, v0.8b            \n"  // move r
+    MEMACCESS(1)
+    "st4        {v2.8b,v3.8b,v4.8b,v5.8b}, [%1], #32 \n"  // store b g r a
+    "b.gt       1b                             \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5"  // Clobber List
+  );
+}
+#endif  // HAS_RAWTOARGBROW_NEON
+
+#define RGB565TOARGB                                                           \
+    "shrn       v6.8b, v0.8h, #5               \n"  /* G xxGGGGGG           */ \
+    "shl        v6.8b, v6.8b, #2               \n"  /* G GGGGGG00 upper 6   */ \
+    "ushr       v4.8b, v6.8b, #6               \n"  /* G 000000GG lower 2   */ \
+    "orr        v1.8b, v4.8b, v6.8b            \n"  /* G                    */ \
+    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
+    "ushr       v0.8h, v0.8h, #11              \n"  /* R 000RRRRR           */ \
+    "xtn2       v2.16b,v0.8h                   \n"  /* R in upper part      */ \
+    "shl        v2.16b, v2.16b, #3             \n"  /* R,B BBBBB000 upper 5 */ \
+    "ushr       v0.16b, v2.16b, #5             \n"  /* R,B 00000BBB lower 3 */ \
+    "orr        v0.16b, v0.16b, v2.16b         \n"  /* R,B                  */ \
+    "dup        v2.2D, v0.D[1]                 \n"  /* R                    */
+
+#ifdef HAS_RGB565TOARGBROW_NEON
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) {
+  asm volatile (
+    "movi       v3.8b, #255                    \n"  // Alpha
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    RGB565TOARGB
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6"  // Clobber List
+  );
+}
+#endif  // HAS_RGB565TOARGBROW_NEON
+
+#define ARGB1555TOARGB                                                         \
+    "ushr       v2.8h, v0.8h, #10              \n"  /* R xxxRRRRR           */ \
+    "shl        v2.8h, v2.8h, #3               \n"  /* R RRRRR000 upper 5   */ \
+    "xtn        v3.8b, v2.8h                   \n"  /* RRRRR000 AAAAAAAA    */ \
+                                                                               \
+    "sshr       v2.8h, v0.8h, #15              \n"  /* A AAAAAAAA           */ \
+    "xtn2       v3.16b, v2.8h                  \n"                             \
+                                                                               \
+    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
+    "shrn2      v2.16b,v0.8h, #5               \n"  /* G xxxGGGGG           */ \
+                                                                               \
+    "ushr       v1.16b, v3.16b, #5             \n"  /* R,A 00000RRR lower 3 */ \
+    "shl        v0.16b, v2.16b, #3             \n"  /* B,G BBBBB000 upper 5 */ \
+    "ushr       v2.16b, v0.16b, #5             \n"  /* B,G 00000BBB lower 3 */ \
+                                                                               \
+    "orr        v0.16b, v0.16b, v2.16b         \n"  /* B,G                  */ \
+    "orr        v2.16b, v1.16b, v3.16b         \n"  /* R,A                  */ \
+    "dup        v1.2D, v0.D[1]                 \n"                             \
+    "dup        v3.2D, v2.D[1]                 \n"
+
+// RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
+#define RGB555TOARGB                                                           \
+    "ushr       v2.8h, v0.8h, #10              \n"  /* R xxxRRRRR           */ \
+    "shl        v2.8h, v2.8h, #3               \n"  /* R RRRRR000 upper 5   */ \
+    "xtn        v3.8b, v2.8h                   \n"  /* RRRRR000             */ \
+                                                                               \
+    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
+    "shrn2      v2.16b,v0.8h, #5               \n"  /* G xxxGGGGG           */ \
+                                                                               \
+    "ushr       v1.16b, v3.16b, #5             \n"  /* R   00000RRR lower 3 */ \
+    "shl        v0.16b, v2.16b, #3             \n"  /* B,G BBBBB000 upper 5 */ \
+    "ushr       v2.16b, v0.16b, #5             \n"  /* B,G 00000BBB lower 3 */ \
+                                                                               \
+    "orr        v0.16b, v0.16b, v2.16b         \n"  /* B,G                  */ \
+    "orr        v2.16b, v1.16b, v3.16b         \n"  /* R                    */ \
+    "dup        v1.2D, v0.D[1]                 \n"  /* G */                    \
+
+#ifdef HAS_ARGB1555TOARGBROW_NEON
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "movi       v3.8b, #255                    \n"  // Alpha
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_ARGB1555TOARGBROW_NEON
+
+#define ARGB4444TOARGB                                                         \
+    "shrn       v1.8b,  v0.8h, #8              \n"  /* v1(l) AR             */ \
+    "xtn2       v1.16b, v0.8h                  \n"  /* v1(h) GB             */ \
+    "shl        v2.16b, v1.16b, #4             \n"  /* B,R BBBB0000         */ \
+    "ushr       v3.16b, v1.16b, #4             \n"  /* G,A 0000GGGG         */ \
+    "ushr       v0.16b, v2.16b, #4             \n"  /* B,R 0000BBBB         */ \
+    "shl        v1.16b, v3.16b, #4             \n"  /* G,A GGGG0000         */ \
+    "orr        v2.16b, v0.16b, v2.16b         \n"  /* B,R BBBBBBBB         */ \
+    "orr        v3.16b, v1.16b, v3.16b         \n"  /* G,A GGGGGGGG         */ \
+    "dup        v0.2D, v2.D[1]                 \n"                             \
+    "dup        v1.2D, v3.D[1]                 \n"
+
+#ifdef HAS_ARGB4444TOARGBROW_NEON
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4"  // Clobber List
+  );
+}
+#endif  // HAS_ARGB4444TOARGBROW_NEON
+
+#ifdef HAS_ARGBTORGB24ROW_NEON
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load 8 ARGB pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "st3        {v1.8b,v2.8b,v3.8b}, [%1], #24 \n"  // store 8 pixels of RGB24.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_rgb24),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "v1", "v2", "v3", "v4"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBTORGB24ROW_NEON
+
+#ifdef HAS_ARGBTORAWROW_NEON
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load b g r a
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "orr        v4.8b, v2.8b, v2.8b            \n"  // mov g
+    "orr        v5.8b, v1.8b, v1.8b            \n"  // mov b
+    MEMACCESS(1)
+    "st3        {v3.8b,v4.8b,v5.8b}, [%1], #24 \n"  // store r g b
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_raw),   // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v1", "v2", "v3", "v4", "v5"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBTORAWROW_NEON
+
+#ifdef HAS_YUY2TOYROW_NEON
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pixels of YUY2.
+    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"  // store 16 pixels of Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+#endif  // HAS_YUY2TOYROW_NEON
+
+#ifdef HAS_UYVYTOYROW_NEON
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pixels of UYVY.
+    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "st1        {v1.16b}, [%1], #16            \n"  // store 16 pixels of Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+#endif  // HAS_UYVYTOYROW_NEON
+
+#ifdef HAS_YUY2TOUV422ROW_NEON
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 YUY2 pixels
+    "subs       %w3, %w3, #16                  \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "st1        {v1.8b}, [%1], #8              \n"  // store 8 U.
+    MEMACCESS(2)
+    "st1        {v3.8b}, [%2], #8              \n"  // store 8 V.
+    "b.gt       1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_YUY2TOUV422ROW_NEON
+
+#ifdef HAS_UYVYTOUV422ROW_NEON
+void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 UYVY pixels
+    "subs       %w3, %w3, #16                  \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 U.
+    MEMACCESS(2)
+    "st1        {v2.8b}, [%2], #8              \n"  // store 8 V.
+    "b.gt       1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_UYVYTOUV422ROW_NEON
+
+#ifdef HAS_YUY2TOUVROW_NEON
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_yuy2b = src_yuy2 + stride_yuy2;
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
+    "subs       %w4, %w4, #16                  \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
+    "urhadd     v1.8b, v1.8b, v5.8b            \n"  // average rows of U
+    "urhadd     v3.8b, v3.8b, v7.8b            \n"  // average rows of V
+    MEMACCESS(2)
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 U.
+    MEMACCESS(3)
+    "st1        {v3.8b}, [%3], #8              \n"  // store 8 V.
+    "b.gt       1b                             \n"
+  : "+r"(src_yuy2),     // %0
+    "+r"(src_yuy2b),    // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
+    "v5", "v6", "v7"  // Clobber List
+  );
+}
+#endif  // HAS_YUY2TOUVROW_NEON
+
+#ifdef HAS_UYVYTOUVROW_NEON
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_uyvyb = src_uyvy + stride_uyvy;
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
+    "subs       %w4, %w4, #16                  \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
+    "urhadd     v0.8b, v0.8b, v4.8b            \n"  // average rows of U
+    "urhadd     v2.8b, v2.8b, v6.8b            \n"  // average rows of V
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 U.
+    MEMACCESS(3)
+    "st1        {v2.8b}, [%3], #8              \n"  // store 8 V.
+    "b.gt       1b                             \n"
+  : "+r"(src_uyvy),     // %0
+    "+r"(src_uyvyb),    // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
+    "v5", "v6", "v7"  // Clobber List
+  );
+}
+#endif  // HAS_UYVYTOUVROW_NEON
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+#ifdef HAS_ARGBSHUFFLEROW_NEON
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    MEMACCESS(3)
+    "ld1        {v2.16b}, [%3]                 \n"  // shuffler
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 4 pixels.
+    "subs       %w2, %w2, #4                   \n"  // 4 processed per loop
+    "tbl        v1.16b, {v0.16b}, v2.16b       \n"  // look up 4 pixels
+    MEMACCESS(1)
+    "st1        {v1.16b}, [%1], #16            \n"  // store 4.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "cc", "memory", "v0", "v1", "v2"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_NEON
+
+#ifdef HAS_I422TOYUY2ROW_NEON
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v0.8b, v1.8b}, [%0], #16      \n"  // load 16 Ys
+    "orr        v2.8b, v1.8b, v1.8b            \n"
+    MEMACCESS(1)
+    "ld1        {v1.8b}, [%1], #8              \n"  // load 8 Us
+    MEMACCESS(2)
+    "ld1        {v3.8b}, [%2], #8              \n"  // load 8 Vs
+    "subs       %w4, %w4, #16                  \n"  // 16 pixels
+    MEMACCESS(3)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_yuy2),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"
+  );
+}
+#endif  // HAS_I422TOYUY2ROW_NEON
+
+#ifdef HAS_I422TOUYVYROW_NEON
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v1.8b,v2.8b}, [%0], #16       \n"  // load 16 Ys
+    "orr        v3.8b, v2.8b, v2.8b            \n"
+    MEMACCESS(1)
+    "ld1        {v0.8b}, [%1], #8              \n"  // load 8 Us
+    MEMACCESS(2)
+    "ld1        {v2.8b}, [%2], #8              \n"  // load 8 Vs
+    "subs       %w4, %w4, #16                  \n"  // 16 pixels
+    MEMACCESS(3)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_uyvy),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"
+  );
+}
+#endif  // HAS_I422TOUYVYROW_NEON
+
+#ifdef HAS_ARGBTORGB565ROW_NEON
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGBTORGB565
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_rgb565),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v20", "v21", "v22", "v23"
+  );
+}
+#endif  // HAS_ARGBTORGB565ROW_NEON
+
+#ifdef HAS_ARGBTORGB565DITHERROW_NEON
+void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int width) {
+  asm volatile (
+    "dup        v1.4s, %w2                     \n"  // dither4
+  "1:                                          \n"
+    MEMACCESS(1)
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n"  // load 8 pixels
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uqadd      v20.8b, v20.8b, v1.8b          \n"
+    "uqadd      v21.8b, v21.8b, v1.8b          \n"
+    "uqadd      v22.8b, v22.8b, v1.8b          \n"
+    ARGBTORGB565
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+  : "+r"(dst_rgb)    // %0
+  : "r"(src_argb),   // %1
+    "r"(dither4),    // %2
+    "r"(width)       // %3
+  : "cc", "memory", "v0", "v1", "v20", "v21", "v22", "v23"
+  );
+}
+#endif  // HAS_ARGBTORGB565ROW_NEON
+
+#ifdef HAS_ARGBTOARGB1555ROW_NEON
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
+                            int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGBTOARGB1555
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels ARGB1555.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb1555),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v20", "v21", "v22", "v23"
+  );
+}
+#endif  // HAS_ARGBTOARGB1555ROW_NEON
+
+#ifdef HAS_ARGBTOARGB4444ROW_NEON
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
+                            int pix) {
+  asm volatile (
+    "movi       v4.16b, #0x0f                  \n"  // bits to clear with vbic.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGBTOARGB4444
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels ARGB4444.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb4444),  // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "v0", "v1", "v4", "v20", "v21", "v22", "v23"
+  );
+}
+#endif  // HAS_ARGBTOARGB4444ROW_NEON
+
+#ifdef HAS_ARGBTOYROW_NEON
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
+    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
+    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGBTOYROW_NEON
+
+#ifdef HAS_ARGBTOYJROW_NEON
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #15                     \n"  // B * 0.11400 coefficient
+    "movi       v5.8b, #75                     \n"  // G * 0.58700 coefficient
+    "movi       v6.8b, #38                     \n"  // R * 0.29900 coefficient
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
+    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
+    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 15 bit to 8 bit Y
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
+  );
+}
+#endif  // HAS_ARGBTOYJROW_NEON
+
+// 8x1 pixels.
+#ifdef HAS_ARGBTOUV444ROW_NEON
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "movi       v24.8b, #112                   \n"  // UB / VR 0.875 coefficient
+    "movi       v25.8b, #74                    \n"  // UG -0.5781 coefficient
+    "movi       v26.8b, #38                    \n"  // UR -0.2969 coefficient
+    "movi       v27.8b, #18                    \n"  // VB -0.1406 coefficient
+    "movi       v28.8b, #94                    \n"  // VG -0.7344 coefficient
+    "movi       v29.16b,#0x80                  \n"  // 128.5
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "umull      v4.8h, v0.8b, v24.8b           \n"  // B
+    "umlsl      v4.8h, v1.8b, v25.8b           \n"  // G
+    "umlsl      v4.8h, v2.8b, v26.8b           \n"  // R
+    "add        v4.8h, v4.8h, v29.8h           \n"  // +128 -> unsigned
+
+    "umull      v3.8h, v2.8b, v24.8b           \n"  // R
+    "umlsl      v3.8h, v1.8b, v28.8b           \n"  // G
+    "umlsl      v3.8h, v0.8b, v27.8b           \n"  // B
+    "add        v3.8h, v3.8h, v29.8h           \n"  // +128 -> unsigned
+
+    "uqshrn     v0.8b, v4.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
+    "v24", "v25", "v26", "v27", "v28", "v29"
+  );
+}
+#endif  // HAS_ARGBTOUV444ROW_NEON
+
+// 16x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_ARGBTOUV422ROW_NEON
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop.
+    "mul        v3.8h, v0.8h, v20.8h           \n"  // B
+    "mls        v3.8h, v1.8h, v21.8h           \n"  // G
+    "mls        v3.8h, v2.8h, v22.8h           \n"  // R
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+
+    "mul        v4.8h, v2.8h, v20.8h           \n"  // R
+    "mls        v4.8h, v1.8h, v24.8h           \n"  // G
+    "mls        v4.8h, v0.8h, v23.8h           \n"  // B
+    "add        v4.8h, v4.8h, v25.8h           \n"  // +128 -> unsigned
+
+    "uqshrn     v0.8b, v3.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v4.8h, #8               \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ARGBTOUV422ROW_NEON
+
+// 32x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 32.
+#ifdef HAS_ARGBTOUV411ROW_NEON
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(0)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%0], #64 \n"  // load next 16.
+    "uaddlp     v4.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v5.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v6.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "addp       v0.8h, v0.8h, v4.8h            \n"  // B 16 shorts -> 8 shorts.
+    "addp       v1.8h, v1.8h, v5.8h            \n"  // G 16 shorts -> 8 shorts.
+    "addp       v2.8h, v2.8h, v6.8h            \n"  // R 16 shorts -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w3, %w3, #32                  \n"  // 32 processed per loop.
+    "mul        v3.8h, v0.8h, v20.8h           \n"  // B
+    "mls        v3.8h, v1.8h, v21.8h           \n"  // G
+    "mls        v3.8h, v2.8h, v22.8h           \n"  // R
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+    "mul        v4.8h, v2.8h, v20.8h           \n"  // R
+    "mls        v4.8h, v1.8h, v24.8h           \n"  // G
+    "mls        v4.8h, v0.8h, v23.8h           \n"  // B
+    "add        v4.8h, v4.8h, v25.8h           \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v3.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v4.8h, #8               \n"  // 16 bit to 8 bit V
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ARGBTOUV411ROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#define RGBTOUV(QB, QG, QR) \
+    "mul        v3.8h, " #QB ",v20.8h          \n"  /* B                    */ \
+    "mul        v4.8h, " #QR ",v20.8h          \n"  /* R                    */ \
+    "mls        v3.8h, " #QG ",v21.8h          \n"  /* G                    */ \
+    "mls        v4.8h, " #QG ",v24.8h          \n"  /* G                    */ \
+    "mls        v3.8h, " #QR ",v22.8h          \n"  /* R                    */ \
+    "mls        v4.8h, " #QB ",v23.8h          \n"  /* B                    */ \
+    "add        v3.8h, v3.8h, v25.8h           \n"  /* +128 -> unsigned     */ \
+    "add        v4.8h, v4.8h, v25.8h           \n"  /* +128 -> unsigned     */ \
+    "uqshrn     v0.8b, v3.8h, #8               \n"  /* 16 bit to 8 bit U    */ \
+    "uqshrn     v1.8b, v4.8h, #8               \n"  /* 16 bit to 8 bit V    */
+
+// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
+// TODO(fbarchard): consider ptrdiff_t for all strides.
+
+#ifdef HAS_ARGBTOUVROW_NEON
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb_1 = src_argb + src_stride_argb;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    MEMACCESS(1)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_argb_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ARGBTOUVROW_NEON
+
+// TODO(fbarchard): Subsample match C code.
+#ifdef HAS_ARGBTOUVJROW_NEON
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb_1 = src_argb + src_stride_argb;
+  asm volatile (
+    "movi       v20.8h, #63, lsl #0            \n"  // UB/VR coeff (0.500) / 2
+    "movi       v21.8h, #42, lsl #0            \n"  // UG coeff (-0.33126) / 2
+    "movi       v22.8h, #21, lsl #0            \n"  // UR coeff (-0.16874) / 2
+    "movi       v23.8h, #10, lsl #0            \n"  // VB coeff (-0.08131) / 2
+    "movi       v24.8h, #53, lsl #0            \n"  // VG coeff (-0.41869) / 2
+    "movi       v25.16b, #0x80                 \n"  // 128.5 (0x8080 in 16-bit)
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64  \n"  // load next 16
+    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_argb_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ARGBTOUVJROW_NEON
+
+#ifdef HAS_BGRATOUVROW_NEON
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_bgra_1 = src_bgra + src_stride_bgra;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v3.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v3.8h, v2.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v1.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more
+    "uadalp     v0.8h, v7.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v3.8h, v6.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v5.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v3.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(src_bgra_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_BGRATOUVROW_NEON
+
+#ifdef HAS_ABGRTOUVROW_NEON
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_abgr_1 = src_abgr + src_stride_abgr;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v3.8h, v2.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v0.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
+    "uadalp     v3.8h, v6.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v4.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v3.8h, #1               \n"  // 2x average
+    "urshr      v2.8h, v2.8h, #1               \n"
+    "urshr      v1.8h, v1.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v2.8h, v1.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_abgr_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ABGRTOUVROW_NEON
+
+#ifdef HAS_RGBATOUVROW_NEON
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_rgba_1 = src_rgba + src_stride_rgba;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v1.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v2.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v3.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
+    "uadalp     v0.8h, v5.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v6.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v7.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(src_rgba_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_RGBATOUVROW_NEON
+
+#ifdef HAS_RGB24TOUVROW_NEON
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_rgb24_1 = src_rgb24 + src_stride_rgb24;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld3        {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 16 more.
+    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_rgb24_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_RGB24TOUVROW_NEON
+
+#ifdef HAS_RAWTOUVROW_NEON
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_raw_1 = src_raw + src_stride_raw;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 8 RAW pixels.
+    "uaddlp     v2.8h, v2.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v0.8h, v0.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld3        {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 8 more RAW pixels
+    "uadalp     v2.8h, v6.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v0.8h, v4.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v2.8h, v2.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v0.8h, v0.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v2.8h, v1.8h, v0.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_raw_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_RAWTOUVROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_RGB565TOUVROW_NEON
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_rgb565_1 = src_rgb565 + src_stride_rgb565;
+  asm volatile (
+    "movi       v22.8h, #56, lsl #0            \n"  // UB / VR coeff (0.875) / 2
+    "movi       v23.8h, #37, lsl #0            \n"  // UG coeff (-0.5781) / 2
+    "movi       v24.8h, #19, lsl #0            \n"  // UR coeff (-0.2969) / 2
+    "movi       v25.8h, #9 , lsl #0            \n"  // VB coeff (-0.1406) / 2
+    "movi       v26.8h, #47, lsl #0            \n"  // VG coeff (-0.7344) / 2
+    "movi       v27.16b, #0x80                 \n"  // 128.5 (0x8080 in 16-bit)
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v18.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v20.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "uaddlp     v17.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v19.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v21.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v18.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v20.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "uadalp     v17.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v19.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v21.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    "ins        v16.D[1], v17.D[0]             \n"
+    "ins        v18.D[1], v19.D[0]             \n"
+    "ins        v20.D[1], v21.D[0]             \n"
+
+    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
+    "urshr      v5.8h, v18.8h, #1              \n"
+    "urshr      v6.8h, v20.8h, #1              \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
+    "mul        v16.8h, v4.8h, v22.8h          \n"  // B
+    "mls        v16.8h, v5.8h, v23.8h          \n"  // G
+    "mls        v16.8h, v6.8h, v24.8h          \n"  // R
+    "add        v16.8h, v16.8h, v27.8h         \n"  // +128 -> unsigned
+    "mul        v17.8h, v6.8h, v22.8h          \n"  // R
+    "mls        v17.8h, v5.8h, v26.8h          \n"  // G
+    "mls        v17.8h, v4.8h, v25.8h          \n"  // B
+    "add        v17.8h, v17.8h, v27.8h         \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v16.8h, #8              \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v17.8h, #8              \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(src_rgb565_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24",
+    "v25", "v26", "v27"
+  );
+}
+#endif  // HAS_RGB565TOUVROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_ARGB1555TOUVROW_NEON
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb1555_1 = src_argb1555 + src_stride_argb1555;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "uaddlp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "uadalp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    "ins        v16.D[1], v26.D[0]             \n"
+    "ins        v17.D[1], v27.D[0]             \n"
+    "ins        v18.D[1], v28.D[0]             \n"
+
+    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
+    "urshr      v5.8h, v17.8h, #1              \n"
+    "urshr      v6.8h, v18.8h, #1              \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
+    "mul        v2.8h, v4.8h, v20.8h           \n"  // B
+    "mls        v2.8h, v5.8h, v21.8h           \n"  // G
+    "mls        v2.8h, v6.8h, v22.8h           \n"  // R
+    "add        v2.8h, v2.8h, v25.8h           \n"  // +128 -> unsigned
+    "mul        v3.8h, v6.8h, v20.8h           \n"  // R
+    "mls        v3.8h, v5.8h, v24.8h           \n"  // G
+    "mls        v3.8h, v4.8h, v23.8h           \n"  // B
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v2.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(src_argb1555_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25",
+    "v26", "v27", "v28"
+  );
+}
+#endif  // HAS_ARGB1555TOUVROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_ARGB4444TOUVROW_NEON
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb4444_1 = src_argb4444 + src_stride_argb4444;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "uaddlp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "uadalp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    "ins        v16.D[1], v26.D[0]             \n"
+    "ins        v17.D[1], v27.D[0]             \n"
+    "ins        v18.D[1], v28.D[0]             \n"
+
+    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
+    "urshr      v5.8h, v17.8h, #1              \n"
+    "urshr      v6.8h, v18.8h, #1              \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
+    "mul        v2.8h, v4.8h, v20.8h           \n"  // B
+    "mls        v2.8h, v5.8h, v21.8h           \n"  // G
+    "mls        v2.8h, v6.8h, v22.8h           \n"  // R
+    "add        v2.8h, v2.8h, v25.8h           \n"  // +128 -> unsigned
+    "mul        v3.8h, v6.8h, v20.8h           \n"  // R
+    "mls        v3.8h, v5.8h, v24.8h           \n"  // G
+    "mls        v3.8h, v4.8h, v23.8h           \n"  // B
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v2.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(src_argb4444_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25",
+    "v26", "v27", "v28"
+
+  );
+}
+#endif  // HAS_ARGB4444TOUVROW_NEON
+
+#ifdef HAS_RGB565TOYROW_NEON
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v24.8b, #13                    \n"  // B * 0.1016 coefficient
+    "movi       v25.8b, #65                    \n"  // G * 0.5078 coefficient
+    "movi       v26.8b, #33                    \n"  // R * 0.2578 coefficient
+    "movi       v27.8b, #16                    \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    RGB565TOARGB
+    "umull      v3.8h, v0.8b, v24.8b           \n"  // B
+    "umlal      v3.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v3.8h, v2.8b, v26.8b           \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v27.8b           \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_y),       // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6",
+    "v24", "v25", "v26", "v27"
+  );
+}
+#endif  // HAS_RGB565TOYROW_NEON
+
+#ifdef HAS_ARGB1555TOYROW_NEON
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
+    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
+    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGB1555TOYROW_NEON
+
+#ifdef HAS_ARGB4444TOYROW_NEON
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v24.8b, #13                    \n"  // B * 0.1016 coefficient
+    "movi       v25.8b, #65                    \n"  // G * 0.5078 coefficient
+    "movi       v26.8b, #33                    \n"  // R * 0.2578 coefficient
+    "movi       v27.8b, #16                    \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    "umull      v3.8h, v0.8b, v24.8b           \n"  // B
+    "umlal      v3.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v3.8h, v2.8b, v26.8b           \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v27.8b           \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27"
+  );
+}
+#endif  // HAS_ARGB4444TOYROW_NEON
+
+#ifdef HAS_BGRATOYROW_NEON
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v1.8b, v4.8b           \n"  // R
+    "umlal      v16.8h, v2.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v3.8b, v6.8b           \n"  // B
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+#endif  // HAS_BGRATOYROW_NEON
+
+#ifdef HAS_ABGRTOYROW_NEON
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v0.8b, v4.8b           \n"  // R
+    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v2.8b, v6.8b           \n"  // B
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+#endif  // HAS_ABGRTOYROW_NEON
+
+#ifdef HAS_RGBATOYROW_NEON
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v1.8b, v4.8b           \n"  // B
+    "umlal      v16.8h, v2.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v3.8b, v6.8b           \n"  // R
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+#endif  // HAS_RGBATOYROW_NEON
+
+#ifdef HAS_RGB24TOYROW_NEON
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v0.8b, v4.8b           \n"  // B
+    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v2.8b, v6.8b           \n"  // R
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_y),      // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+#endif  // HAS_RGB24TOYROW_NEON
+
+#ifdef HAS_RAWTOYROW_NEON
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v0.8b, v4.8b           \n"  // B
+    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v2.8b, v6.8b           \n"  // R
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(dst_y),    // %1
+    "+r"(pix)       // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+#endif  // HAS_RAWTOYROW_NEON
+
+// Bilinear filter 16x2 -> 16x1
+#ifdef HAS_INTERPOLATEROW_NEON
+void InterpolateRow_NEON(uint8* dst_ptr,
+                         const uint8* src_ptr, ptrdiff_t src_stride,
+                         int dst_width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  asm volatile (
+    "cmp        %w4, #0                        \n"
+    "b.eq       100f                           \n"
+    "cmp        %w4, #64                       \n"
+    "b.eq       75f                            \n"
+    "cmp        %w4, #128                      \n"
+    "b.eq       50f                            \n"
+    "cmp        %w4, #192                      \n"
+    "b.eq       25f                            \n"
+
+    "dup        v5.16b, %w4                    \n"
+    "dup        v4.16b, %w5                    \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"
+    MEMACCESS(2)
+    "ld1        {v1.16b}, [%2], #16            \n"
+    "subs       %w3, %w3, #16                  \n"
+    "umull      v2.8h, v0.8b,  v4.8b           \n"
+    "umull2     v3.8h, v0.16b, v4.16b          \n"
+    "umlal      v2.8h, v1.8b,  v5.8b           \n"
+    "umlal2     v3.8h, v1.16b, v5.16b          \n"
+    "rshrn      v0.8b,  v2.8h, #8              \n"
+    "rshrn2     v0.16b, v3.8h, #8              \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       1b                             \n"
+    "b          99f                            \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"
+    MEMACCESS(2)
+    "ld1        {v1.16b}, [%2], #16            \n"
+    "subs       %w3, %w3, #16                  \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       25b                            \n"
+    "b          99f                            \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"
+    MEMACCESS(2)
+    "ld1        {v1.16b}, [%2], #16            \n"
+    "subs       %w3, %w3, #16                  \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       50b                            \n"
+    "b          99f                            \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "ld1        {v1.16b}, [%1], #16            \n"
+    MEMACCESS(2)
+    "ld1        {v0.16b}, [%2], #16            \n"
+    "subs       %w3, %w3, #16                  \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       75b                            \n"
+    "b          99f                            \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"
+    "subs       %w3, %w3, #16                  \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       100b                           \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_ptr1),         // %2
+    "+r"(dst_width),        // %3
+    "+r"(y1_fraction),      // %4
+    "+r"(y0_fraction)       // %5
+  :
+  : "cc", "memory", "v0", "v1", "v3", "v4", "v5"
+  );
+}
+#endif  // HAS_INTERPOLATEROW_NEON
+
+// dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
+#ifdef HAS_ARGBBLENDROW_NEON
+void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "subs       %w3, %w3, #8                   \n"
+    "b.lt       89f                            \n"
+    // Blend 8 pixels.
+  "8:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB0 pixels
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 ARGB1 pixels
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v4.8b, v3.8b           \n"  // db * a
+    "umull      v17.8h, v5.8b, v3.8b           \n"  // dg * a
+    "umull      v18.8h, v6.8b, v3.8b           \n"  // dr * a
+    "uqrshrn    v16.8b, v16.8h, #8             \n"  // db >>= 8
+    "uqrshrn    v17.8b, v17.8h, #8             \n"  // dg >>= 8
+    "uqrshrn    v18.8b, v18.8h, #8             \n"  // dr >>= 8
+    "uqsub      v4.8b, v4.8b, v16.8b           \n"  // db - (db * a / 256)
+    "uqsub      v5.8b, v5.8b, v17.8b           \n"  // dg - (dg * a / 256)
+    "uqsub      v6.8b, v6.8b, v18.8b           \n"  // dr - (dr * a / 256)
+    "uqadd      v0.8b, v0.8b, v4.8b            \n"  // + sb
+    "uqadd      v1.8b, v1.8b, v5.8b            \n"  // + sg
+    "uqadd      v2.8b, v2.8b, v6.8b            \n"  // + sr
+    "movi       v3.8b, #255                    \n"  // a = 255
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.ge       8b                             \n"
+
+  "89:                                         \n"
+    "adds       %w3, %w3, #8-1                 \n"
+    "b.lt       99f                            \n"
+
+    // Blend 1 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.b,v1.b,v2.b,v3.b}[0], [%0], #4 \n"  // load 1 pixel ARGB0.
+    MEMACCESS(1)
+    "ld4        {v4.b,v5.b,v6.b,v7.b}[0], [%1], #4 \n"  // load 1 pixel ARGB1.
+    "subs       %w3, %w3, #1                   \n"  // 1 processed per loop.
+    "umull      v16.8h, v4.8b, v3.8b           \n"  // db * a
+    "umull      v17.8h, v5.8b, v3.8b           \n"  // dg * a
+    "umull      v18.8h, v6.8b, v3.8b           \n"  // dr * a
+    "uqrshrn    v16.8b, v16.8h, #8             \n"  // db >>= 8
+    "uqrshrn    v17.8b, v17.8h, #8             \n"  // dg >>= 8
+    "uqrshrn    v18.8b, v18.8h, #8             \n"  // dr >>= 8
+    "uqsub      v4.8b, v4.8b, v16.8b           \n"  // db - (db * a / 256)
+    "uqsub      v5.8b, v5.8b, v17.8b           \n"  // dg - (dg * a / 256)
+    "uqsub      v6.8b, v6.8b, v18.8b           \n"  // dr - (dr * a / 256)
+    "uqadd      v0.8b, v0.8b, v4.8b            \n"  // + sb
+    "uqadd      v1.8b, v1.8b, v5.8b            \n"  // + sg
+    "uqadd      v2.8b, v2.8b, v6.8b            \n"  // + sr
+    "movi       v3.8b, #255                    \n"  // a = 255
+    MEMACCESS(2)
+    "st4        {v0.b,v1.b,v2.b,v3.b}[0], [%2], #4 \n"  // store 1 pixel.
+    "b.ge       1b                             \n"
+
+  "99:                                         \n"
+
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v16", "v17", "v18"
+  );
+}
+#endif  // HAS_ARGBBLENDROW_NEON
+
+// Attenuate 8 pixels at a time.
+#ifdef HAS_ARGBATTENUATEROW_NEON
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    // Attenuate 8 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v4.8h, v0.8b, v3.8b            \n"  // b * a
+    "umull      v5.8h, v1.8b, v3.8b            \n"  // g * a
+    "umull      v6.8h, v2.8b, v3.8b            \n"  // r * a
+    "uqrshrn    v0.8b, v4.8h, #8               \n"  // b >>= 8
+    "uqrshrn    v1.8b, v5.8h, #8               \n"  // g >>= 8
+    "uqrshrn    v2.8b, v6.8h, #8               \n"  // r >>= 8
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_NEON
+
+// Quantize 8 ARGB pixels (32 bytes).
+// dst = (dst * scale >> 16) * interval_size + interval_offset;
+#ifdef HAS_ARGBQUANTIZEROW_NEON
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "dup        v4.8h, %w2                     \n"
+    "ushr       v4.8h, v4.8h, #1               \n"  // scale >>= 1
+    "dup        v5.8h, %w3                     \n"  // interval multiply.
+    "dup        v6.8h, %w4                     \n"  // interval add
+
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0]  \n"  // load 8 pixels of ARGB.
+    "subs       %w1, %w1, #8                   \n"  // 8 processed per loop.
+    "uxtl       v0.8h, v0.8b                   \n"  // b (0 .. 255)
+    "uxtl       v1.8h, v1.8b                   \n"
+    "uxtl       v2.8h, v2.8b                   \n"
+    "sqdmulh    v0.8h, v0.8h, v4.8h            \n"  // b * scale
+    "sqdmulh    v1.8h, v1.8h, v4.8h            \n"  // g
+    "sqdmulh    v2.8h, v2.8h, v4.8h            \n"  // r
+    "mul        v0.8h, v0.8h, v5.8h            \n"  // b * interval_size
+    "mul        v1.8h, v1.8h, v5.8h            \n"  // g
+    "mul        v2.8h, v2.8h, v5.8h            \n"  // r
+    "add        v0.8h, v0.8h, v6.8h            \n"  // b + interval_offset
+    "add        v1.8h, v1.8h, v6.8h            \n"  // g
+    "add        v2.8h, v2.8h, v6.8h            \n"  // r
+    "uqxtn      v0.8b, v0.8h                   \n"
+    "uqxtn      v1.8b, v1.8h                   \n"
+    "uqxtn      v2.8b, v2.8h                   \n"
+    MEMACCESS(0)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
+  );
+}
+#endif  // HAS_ARGBQUANTIZEROW_NEON
+
+// Shade 8 pixels at a time by specified value.
+// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
+// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
+#ifdef HAS_ARGBSHADEROW_NEON
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "dup        v0.4s, %w3                     \n"  // duplicate scale value.
+    "zip1       v0.8b, v0.8b, v0.8b            \n"  // v0.8b aarrggbb.
+    "ushr       v0.8h, v0.8h, #1               \n"  // scale / 2.
+
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "uxtl       v4.8h, v4.8b                   \n"  // b (0 .. 255)
+    "uxtl       v5.8h, v5.8b                   \n"
+    "uxtl       v6.8h, v6.8b                   \n"
+    "uxtl       v7.8h, v7.8b                   \n"
+    "sqrdmulh   v4.8h, v4.8h, v0.h[0]          \n"  // b * scale * 2
+    "sqrdmulh   v5.8h, v5.8h, v0.h[1]          \n"  // g
+    "sqrdmulh   v6.8h, v6.8h, v0.h[2]          \n"  // r
+    "sqrdmulh   v7.8h, v7.8h, v0.h[3]          \n"  // a
+    "uqxtn      v4.8b, v4.8h                   \n"
+    "uqxtn      v5.8b, v5.8h                   \n"
+    "uqxtn      v6.8b, v6.8h                   \n"
+    "uqxtn      v7.8b, v7.8h                   \n"
+    MEMACCESS(1)
+    "st4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),       // %0
+    "+r"(dst_argb),       // %1
+    "+r"(width)           // %2
+  : "r"(value)            // %3
+  : "cc", "memory", "v0", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGBSHADEROW_NEON
+
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+// Similar to ARGBToYJ but stores ARGB.
+// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
+#ifdef HAS_ARGBGRAYROW_NEON
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "movi       v24.8b, #15                    \n"  // B * 0.11400 coefficient
+    "movi       v25.8b, #75                    \n"  // G * 0.58700 coefficient
+    "movi       v26.8b, #38                    \n"  // R * 0.29900 coefficient
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v4.8h, v0.8b, v24.8b           \n"  // B
+    "umlal      v4.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v4.8h, v2.8b, v26.8b           \n"  // R
+    "sqrshrun   v0.8b, v4.8h, #7               \n"  // 15 bit to 8 bit B
+    "orr        v1.8b, v0.8b, v0.8b            \n"  // G
+    "orr        v2.8b, v0.8b, v0.8b            \n"  // R
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v24", "v25", "v26"
+  );
+}
+#endif  // HAS_ARGBGRAYROW_NEON
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+
+#ifdef HAS_ARGBSEPIAROW_NEON
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
+  asm volatile (
+    "movi       v20.8b, #17                    \n"  // BB coefficient
+    "movi       v21.8b, #68                    \n"  // BG coefficient
+    "movi       v22.8b, #35                    \n"  // BR coefficient
+    "movi       v24.8b, #22                    \n"  // GB coefficient
+    "movi       v25.8b, #88                    \n"  // GG coefficient
+    "movi       v26.8b, #45                    \n"  // GR coefficient
+    "movi       v28.8b, #24                    \n"  // BB coefficient
+    "movi       v29.8b, #98                    \n"  // BG coefficient
+    "movi       v30.8b, #50                    \n"  // BR coefficient
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n"  // load 8 ARGB pixels.
+    "subs       %w1, %w1, #8                   \n"  // 8 processed per loop.
+    "umull      v4.8h, v0.8b, v20.8b           \n"  // B to Sepia B
+    "umlal      v4.8h, v1.8b, v21.8b           \n"  // G
+    "umlal      v4.8h, v2.8b, v22.8b           \n"  // R
+    "umull      v5.8h, v0.8b, v24.8b           \n"  // B to Sepia G
+    "umlal      v5.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v5.8h, v2.8b, v26.8b           \n"  // R
+    "umull      v6.8h, v0.8b, v28.8b           \n"  // B to Sepia R
+    "umlal      v6.8h, v1.8b, v29.8b           \n"  // G
+    "umlal      v6.8h, v2.8b, v30.8b           \n"  // R
+    "uqshrn     v0.8b, v4.8h, #7               \n"  // 16 bit to 8 bit B
+    "uqshrn     v1.8b, v5.8h, #7               \n"  // 16 bit to 8 bit G
+    "uqshrn     v2.8b, v6.8h, #7               \n"  // 16 bit to 8 bit R
+    MEMACCESS(0)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(dst_argb),  // %0
+    "+r"(width)      // %1
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v24", "v25", "v26", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_ARGBSEPIAROW_NEON
+
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// TODO(fbarchard): Was same as Sepia except matrix is provided.  This function
+// needs to saturate.  Consider doing a non-saturating version.
+#ifdef HAS_ARGBCOLORMATRIXROW_NEON
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "ld1        {v2.16b}, [%3]                 \n"  // load 3 ARGB vectors.
+    "sxtl       v0.8h, v2.8b                   \n"  // B,G coefficients s16.
+    "sxtl2      v1.8h, v2.16b                  \n"  // R,A coefficients s16.
+
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "uxtl       v16.8h, v16.8b                 \n"  // b (0 .. 255) 16 bit
+    "uxtl       v17.8h, v17.8b                 \n"  // g
+    "uxtl       v18.8h, v18.8b                 \n"  // r
+    "uxtl       v19.8h, v19.8b                 \n"  // a
+    "mul        v22.8h, v16.8h, v0.h[0]        \n"  // B = B * Matrix B
+    "mul        v23.8h, v16.8h, v0.h[4]        \n"  // G = B * Matrix G
+    "mul        v24.8h, v16.8h, v1.h[0]        \n"  // R = B * Matrix R
+    "mul        v25.8h, v16.8h, v1.h[4]        \n"  // A = B * Matrix A
+    "mul        v4.8h, v17.8h, v0.h[1]         \n"  // B += G * Matrix B
+    "mul        v5.8h, v17.8h, v0.h[5]         \n"  // G += G * Matrix G
+    "mul        v6.8h, v17.8h, v1.h[1]         \n"  // R += G * Matrix R
+    "mul        v7.8h, v17.8h, v1.h[5]         \n"  // A += G * Matrix A
+    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
+    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
+    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
+    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
+    "mul        v4.8h, v18.8h, v0.h[2]         \n"  // B += R * Matrix B
+    "mul        v5.8h, v18.8h, v0.h[6]         \n"  // G += R * Matrix G
+    "mul        v6.8h, v18.8h, v1.h[2]         \n"  // R += R * Matrix R
+    "mul        v7.8h, v18.8h, v1.h[6]         \n"  // A += R * Matrix A
+    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
+    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
+    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
+    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
+    "mul        v4.8h, v19.8h, v0.h[3]         \n"  // B += A * Matrix B
+    "mul        v5.8h, v19.8h, v0.h[7]         \n"  // G += A * Matrix G
+    "mul        v6.8h, v19.8h, v1.h[3]         \n"  // R += A * Matrix R
+    "mul        v7.8h, v19.8h, v1.h[7]         \n"  // A += A * Matrix A
+    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
+    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
+    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
+    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
+    "sqshrun    v16.8b, v22.8h, #6             \n"  // 16 bit to 8 bit B
+    "sqshrun    v17.8b, v23.8h, #6             \n"  // 16 bit to 8 bit G
+    "sqshrun    v18.8b, v24.8h, #6             \n"  // 16 bit to 8 bit R
+    "sqshrun    v19.8b, v25.8h, #6             \n"  // 16 bit to 8 bit A
+    MEMACCESS(1)
+    "st4        {v16.8b,v17.8b,v18.8b,v19.8b}, [%1], #32 \n"  // store 8 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "r"(matrix_argb)  // %3
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
+    "v18", "v19", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_NEON
+
+// TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable.
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "umull      v0.8h, v0.8b, v4.8b            \n"  // multiply B
+    "umull      v1.8h, v1.8b, v5.8b            \n"  // multiply G
+    "umull      v2.8h, v2.8b, v6.8b            \n"  // multiply R
+    "umull      v3.8h, v3.8b, v7.8b            \n"  // multiply A
+    "rshrn      v0.8b, v0.8h, #8               \n"  // 16 bit to 8 bit B
+    "rshrn      v1.8b, v1.8h, #8               \n"  // 16 bit to 8 bit G
+    "rshrn      v2.8b, v2.8h, #8               \n"  // 16 bit to 8 bit R
+    "rshrn      v3.8b, v3.8h, #8               \n"  // 16 bit to 8 bit A
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_NEON
+
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+#ifdef HAS_ARGBADDROW_NEON
+void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uqadd      v0.8b, v0.8b, v4.8b            \n"
+    "uqadd      v1.8b, v1.8b, v5.8b            \n"
+    "uqadd      v2.8b, v2.8b, v6.8b            \n"
+    "uqadd      v3.8b, v3.8b, v7.8b            \n"
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGBADDROW_NEON
+
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+#ifdef HAS_ARGBSUBTRACTROW_NEON
+void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uqsub      v0.8b, v0.8b, v4.8b            \n"
+    "uqsub      v1.8b, v1.8b, v5.8b            \n"
+    "uqsub      v2.8b, v2.8b, v6.8b            \n"
+    "uqsub      v3.8b, v3.8b, v7.8b            \n"
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_NEON
+
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+#ifdef HAS_SOBELROW_NEON
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "movi       v3.8b, #255                    \n"  // alpha
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.8b}, [%0], #8              \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "ld1        {v1.8b}, [%1], #8              \n"  // load 8 sobely.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uqadd      v0.8b, v0.8b, v1.8b            \n"  // add
+    "orr        v1.8b, v0.8b, v0.8b            \n"
+    "orr        v2.8b, v0.8b, v0.8b            \n"
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"
+  );
+}
+#endif  // HAS_SOBELROW_NEON
+
+// Adds Sobel X and Sobel Y and stores Sobel into plane.
+#ifdef HAS_SOBELTOPLANEROW_NEON
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    // 16 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 16 sobelx.
+    MEMACCESS(1)
+    "ld1        {v1.16b}, [%1], #16            \n"  // load 16 sobely.
+    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop.
+    "uqadd      v0.16b, v0.16b, v1.16b         \n"  // add
+    MEMACCESS(2)
+    "st1        {v0.16b}, [%2], #16            \n"  // store 16 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "v0", "v1"
+  );
+}
+#endif  // HAS_SOBELTOPLANEROW_NEON
+
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+#ifdef HAS_SOBELXYROW_NEON
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "movi       v3.8b, #255                    \n"  // alpha
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v2.8b}, [%0], #8              \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "ld1        {v0.8b}, [%1], #8              \n"  // load 8 sobely.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uqadd      v1.8b, v0.8b, v2.8b            \n"  // add
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"
+  );
+}
+#endif  // HAS_SOBELXYROW_NEON
+
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+#ifdef HAS_SOBELXROW_NEON
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.8b}, [%0],%5               \n"  // top
+    MEMACCESS(0)
+    "ld1        {v1.8b}, [%0],%6               \n"
+    "usubl      v0.8h, v0.8b, v1.8b            \n"
+    MEMACCESS(1)
+    "ld1        {v2.8b}, [%1],%5               \n"  // center * 2
+    MEMACCESS(1)
+    "ld1        {v3.8b}, [%1],%6               \n"
+    "usubl      v1.8h, v2.8b, v3.8b            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    MEMACCESS(2)
+    "ld1        {v2.8b}, [%2],%5               \n"  // bottom
+    MEMACCESS(2)
+    "ld1        {v3.8b}, [%2],%6               \n"
+    "subs       %w4, %w4, #8                   \n"  // 8 pixels
+    "usubl      v1.8h, v2.8b, v3.8b            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    "abs        v0.8h, v0.8h                   \n"
+    "uqxtn      v0.8b, v0.8h                   \n"
+    MEMACCESS(3)
+    "st1        {v0.8b}, [%3], #8              \n"  // store 8 sobelx
+    "b.gt       1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  : "r"(2LL),          // %5
+    "r"(6LL)           // %6
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_SOBELXROW_NEON
+
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+#ifdef HAS_SOBELYROW_NEON
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.8b}, [%0],%4               \n"  // left
+    MEMACCESS(1)
+    "ld1        {v1.8b}, [%1],%4               \n"
+    "usubl      v0.8h, v0.8b, v1.8b            \n"
+    MEMACCESS(0)
+    "ld1        {v2.8b}, [%0],%4               \n"  // center * 2
+    MEMACCESS(1)
+    "ld1        {v3.8b}, [%1],%4               \n"
+    "usubl      v1.8h, v2.8b, v3.8b            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    MEMACCESS(0)
+    "ld1        {v2.8b}, [%0],%5               \n"  // right
+    MEMACCESS(1)
+    "ld1        {v3.8b}, [%1],%5               \n"
+    "subs       %w3, %w3, #8                   \n"  // 8 pixels
+    "usubl      v1.8h, v2.8b, v3.8b            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    "abs        v0.8h, v0.8h                   \n"
+    "uqxtn      v0.8b, v0.8h                   \n"
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 sobely
+    "b.gt       1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  : "r"(1LL),          // %4
+    "r"(6LL)           // %5
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_SOBELYROW_NEON
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/row_win.cc b/libvpx/libvpx/third_party/libyuv/source/row_win.cc
new file mode 100644
index 0000000..71be268
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/row_win.cc
@@ -0,0 +1,6331 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_X64) && \
+    defined(_MSC_VER) && !defined(__clang__)
+#include <emmintrin.h>
+#include <tmmintrin.h>  // For _mm_maddubs_epi16
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(_M_IX86) || defined(_M_X64)) && \
+    defined(_MSC_VER) && !defined(__clang__)
+
+struct YuvConstants {
+  lvec8 kUVToB;     // 0
+  lvec8 kUVToG;     // 32
+  lvec8 kUVToR;     // 64
+  lvec16 kUVBiasB;  // 96
+  lvec16 kUVBiasG;  // 128
+  lvec16 kUVBiasR;  // 160
+  lvec16 kYToRgb;   // 192
+};
+
+// BT.601 YUV to RGB reference
+//  R = (Y - 16) * 1.164              - V * -1.596
+//  G = (Y - 16) * 1.164 - U *  0.391 - V *  0.813
+//  B = (Y - 16) * 1.164 - U * -2.018
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* max(-128, round(-2.018 * 64)) */
+#define UG 25 /* round(0.391 * 64) */
+#define VG 52 /* round(0.813 * 64) */
+#define VR -102 /* round(-1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            + YGB)
+#define BG (UG * 128 + VG * 128 + YGB)
+#define BR            (VR * 128 + YGB)
+
+// BT601 constants for YUV to RGB.
+static YuvConstants SIMD_ALIGNED(kYuvConstants) = {
+  { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
+    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 },
+  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
+    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
+  { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
+    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
+};
+
+// BT601 constants for NV21 where chroma plane is VU instead of UV.
+static YuvConstants SIMD_ALIGNED(kYvuConstants) = {
+  { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
+    0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB },
+  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
+    VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
+  { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
+    VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
+};
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+// JPEG YUV to RGB reference
+// *  R = Y                - V * -1.40200
+// *  G = Y - U *  0.34414 - V *  0.71414
+// *  B = Y - U * -1.77200
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YGJ 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+#define YGBJ 32  /* 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UBJ -113 /* round(-1.77200 * 64) */
+#define UGJ 22 /* round(0.34414 * 64) */
+#define VGJ 46 /* round(0.71414  * 64) */
+#define VRJ -90 /* round(-1.40200 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BBJ (UBJ * 128             + YGBJ)
+#define BGJ (UGJ * 128 + VGJ * 128 + YGBJ)
+#define BRJ             (VRJ * 128 + YGBJ)
+
+// JPEG constants for YUV to RGB.
+static YuvConstants SIMD_ALIGNED(kYuvJConstants) = {
+  { UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0,
+    UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0 },
+  { UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ },
+  { 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ,
+    0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ },
+  { BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ,
+    BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ },
+  { BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ,
+    BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ },
+  { BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ,
+    BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ },
+  { YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ,
+    YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ }
+};
+
+#undef YGJ
+#undef YGBJ
+#undef UBJ
+#undef UGJ
+#undef VGJ
+#undef VRJ
+#undef BBJ
+#undef BGJ
+#undef BRJ
+
+// 64 bit
+#if defined(_M_X64)
+#if defined(HAS_I422TOARGBROW_SSSE3)
+void I422ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __m128i xmm0, xmm1, xmm2, xmm3;
+  const __m128i xmm5 = _mm_set1_epi8(-1);
+  const ptrdiff_t offset = (uint8*)v_buf - (uint8*)u_buf;
+
+  while (width > 0) {
+    xmm0 = _mm_cvtsi32_si128(*(uint32*)u_buf);
+    xmm1 = _mm_cvtsi32_si128(*(uint32*)(u_buf + offset));
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
+    xmm0 = _mm_unpacklo_epi16(xmm0, xmm0);
+    xmm1 = _mm_loadu_si128(&xmm0);
+    xmm2 = _mm_loadu_si128(&xmm0);
+    xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)kYuvConstants.kUVToB);
+    xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)kYuvConstants.kUVToG);
+    xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)kYuvConstants.kUVToR);
+    xmm0 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasB, xmm0);
+    xmm1 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasG, xmm1);
+    xmm2 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasR, xmm2);
+    xmm3 = _mm_loadl_epi64((__m128i*)y_buf);
+    xmm3 = _mm_unpacklo_epi8(xmm3, xmm3);
+    xmm3 = _mm_mulhi_epu16(xmm3, *(__m128i*)kYuvConstants.kYToRgb);
+    xmm0 = _mm_adds_epi16(xmm0, xmm3);
+    xmm1 = _mm_adds_epi16(xmm1, xmm3);
+    xmm2 = _mm_adds_epi16(xmm2, xmm3);
+    xmm0 = _mm_srai_epi16(xmm0, 6);
+    xmm1 = _mm_srai_epi16(xmm1, 6);
+    xmm2 = _mm_srai_epi16(xmm2, 6);
+    xmm0 = _mm_packus_epi16(xmm0, xmm0);
+    xmm1 = _mm_packus_epi16(xmm1, xmm1);
+    xmm2 = _mm_packus_epi16(xmm2, xmm2);
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm5);
+    xmm1 = _mm_loadu_si128(&xmm0);
+    xmm0 = _mm_unpacklo_epi16(xmm0, xmm2);
+    xmm1 = _mm_unpackhi_epi16(xmm1, xmm2);
+
+    _mm_storeu_si128((__m128i *)dst_argb, xmm0);
+    _mm_storeu_si128((__m128i *)(dst_argb + 16), xmm1);
+
+    y_buf += 8;
+    u_buf += 4;
+    dst_argb += 32;
+    width -= 8;
+  }
+}
+#endif
+// 32 bit
+#else  // defined(_M_X64)
+#ifdef HAS_ARGBTOYROW_SSSE3
+
+// Constants for ARGB.
+static const vec8 kARGBToY = {
+  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
+};
+
+// JPeg full range.
+static const vec8 kARGBToYJ = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+
+static const vec8 kARGBToU = {
+  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
+};
+
+static const vec8 kARGBToUJ = {
+  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
+};
+
+static const vec8 kARGBToV = {
+  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+};
+
+static const vec8 kARGBToVJ = {
+  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
+};
+
+// vpshufb for vphaddw + vpackuswb packed to shorts.
+static const lvec8 kShufARGBToUV_AVX = {
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15
+};
+
+// Constants for BGRA.
+static const vec8 kBGRAToY = {
+  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
+};
+
+static const vec8 kBGRAToU = {
+  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
+};
+
+static const vec8 kBGRAToV = {
+  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
+};
+
+// Constants for ABGR.
+static const vec8 kABGRToY = {
+  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
+};
+
+static const vec8 kABGRToU = {
+  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
+};
+
+static const vec8 kABGRToV = {
+  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
+};
+
+// Constants for RGBA.
+static const vec8 kRGBAToY = {
+  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
+};
+
+static const vec8 kRGBAToU = {
+  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
+};
+
+static const vec8 kRGBAToV = {
+  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
+};
+
+static const uvec8 kAddY16 = {
+  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
+};
+
+// 7 bit fixed point 0.5.
+static const vec16 kAddYJ64 = {
+  64, 64, 64, 64, 64, 64, 64, 64
+};
+
+static const uvec8 kAddUV128 = {
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
+};
+
+static const uvec16 kAddUVJ128 = {
+  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
+};
+
+// Shuffle table for converting RGB24 to ARGB.
+static const uvec8 kShuffleMaskRGB24ToARGB = {
+  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
+};
+
+// Shuffle table for converting RAW to ARGB.
+static const uvec8 kShuffleMaskRAWToARGB = {
+  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
+};
+
+// Shuffle table for converting ARGB to RGB24.
+static const uvec8 kShuffleMaskARGBToRGB24 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static const uvec8 kShuffleMaskARGBToRAW = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
+static const uvec8 kShuffleMaskARGBToRGB24_0 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static const uvec8 kShuffleMaskARGBToRAW_0 = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u
+};
+
+// Duplicates gray value 3 times and fills in alpha opaque.
+__declspec(naked)
+void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]        // src_y
+    mov        edx, [esp + 8]        // dst_argb
+    mov        ecx, [esp + 12]       // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
+    pslld      xmm5, 24
+
+  convertloop:
+    movq       xmm0, qword ptr [eax]
+    lea        eax,  [eax + 8]
+    punpcklbw  xmm0, xmm0
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0
+    punpckhwd  xmm1, xmm1
+    por        xmm0, xmm5
+    por        xmm1, xmm5
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+
+#ifdef HAS_J400TOARGBROW_AVX2
+// Duplicates gray value 3 times and fills in alpha opaque.
+__declspec(naked)
+void J400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int pix) {
+  __asm {
+    mov         eax, [esp + 4]        // src_y
+    mov         edx, [esp + 8]        // dst_argb
+    mov         ecx, [esp + 12]       // pix
+    vpcmpeqb    ymm5, ymm5, ymm5      // generate mask 0xff000000
+    vpslld      ymm5, ymm5, 24
+
+  convertloop:
+    vmovdqu     xmm0, [eax]
+    lea         eax,  [eax + 16]
+    vpermq      ymm0, ymm0, 0xd8
+    vpunpcklbw  ymm0, ymm0, ymm0
+    vpermq      ymm0, ymm0, 0xd8
+    vpunpckhwd  ymm1, ymm0, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm0
+    vpor        ymm0, ymm0, ymm5
+    vpor        ymm1, ymm1, ymm5
+    vmovdqu     [edx], ymm0
+    vmovdqu     [edx + 32], ymm1
+    lea         edx, [edx + 64]
+    sub         ecx, 16
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_J400TOARGBROW_AVX2
+
+__declspec(naked)
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_rgb24
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pslld     xmm5, 24
+    movdqa    xmm4, kShuffleMaskRGB24ToARGB
+
+ convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm3, [eax + 32]
+    lea       eax, [eax + 48]
+    movdqa    xmm2, xmm3
+    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    pshufb    xmm2, xmm4
+    por       xmm2, xmm5
+    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    pshufb    xmm0, xmm4
+    movdqu    [edx + 32], xmm2
+    por       xmm0, xmm5
+    pshufb    xmm1, xmm4
+    movdqu    [edx], xmm0
+    por       xmm1, xmm5
+    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    pshufb    xmm3, xmm4
+    movdqu    [edx + 16], xmm1
+    por       xmm3, xmm5
+    movdqu    [edx + 48], xmm3
+    lea       edx, [edx + 64]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
+                        int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_raw
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pslld     xmm5, 24
+    movdqa    xmm4, kShuffleMaskRAWToARGB
+
+ convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm3, [eax + 32]
+    lea       eax, [eax + 48]
+    movdqa    xmm2, xmm3
+    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    pshufb    xmm2, xmm4
+    por       xmm2, xmm5
+    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    pshufb    xmm0, xmm4
+    movdqu    [edx + 32], xmm2
+    por       xmm0, xmm5
+    pshufb    xmm1, xmm4
+    movdqu    [edx], xmm0
+    por       xmm1, xmm5
+    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    pshufb    xmm3, xmm4
+    movdqu    [edx + 16], xmm1
+    por       xmm3, xmm5
+    movdqu    [edx + 48], xmm3
+    lea       edx, [edx + 64]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+// pmul method to replicate bits.
+// Math to replicate bits:
+// (v << 8) | (v << 3)
+// v * 256 + v * 8
+// v * (256 + 8)
+// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
+// 20 instructions.
+__declspec(naked)
+void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+                          int pix) {
+  __asm {
+    mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
+    movd      xmm5, eax
+    pshufd    xmm5, xmm5, 0
+    mov       eax, 0x20802080  // multiplier shift by 5 and then repeat 6 bits
+    movd      xmm6, eax
+    pshufd    xmm6, xmm6, 0
+    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    psllw     xmm3, 11
+    pcmpeqb   xmm4, xmm4       // generate mask 0x07e007e0 for Green
+    psllw     xmm4, 10
+    psrlw     xmm4, 5
+    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    psllw     xmm7, 8
+
+    mov       eax, [esp + 4]   // src_rgb565
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of bgr565
+    movdqa    xmm1, xmm0
+    movdqa    xmm2, xmm0
+    pand      xmm1, xmm3    // R in upper 5 bits
+    psllw     xmm2, 11      // B in upper 5 bits
+    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    psllw     xmm1, 8
+    por       xmm1, xmm2    // RB
+    pand      xmm0, xmm4    // G in middle 6 bits
+    pmulhuw   xmm0, xmm6    // << 5 * (256 + 4)
+    por       xmm0, xmm7    // AG
+    movdqa    xmm2, xmm1
+    punpcklbw xmm1, xmm0
+    punpckhbw xmm2, xmm0
+    movdqu    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqu    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+#ifdef HAS_RGB565TOARGBROW_AVX2
+// pmul method to replicate bits.
+// Math to replicate bits:
+// (v << 8) | (v << 3)
+// v * 256 + v * 8
+// v * (256 + 8)
+// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
+__declspec(naked)
+void RGB565ToARGBRow_AVX2(const uint8* src_rgb565, uint8* dst_argb,
+                          int pix) {
+  __asm {
+    mov        eax, 0x01080108  // generate multiplier to repeat 5 bits
+    vmovd      xmm5, eax
+    vbroadcastss ymm5, xmm5
+    mov        eax, 0x20802080  // multiplier shift by 5 and then repeat 6 bits
+    movd       xmm6, eax
+    vbroadcastss ymm6, xmm6
+    vpcmpeqb   ymm3, ymm3, ymm3       // generate mask 0xf800f800 for Red
+    vpsllw     ymm3, ymm3, 11
+    vpcmpeqb   ymm4, ymm4, ymm4       // generate mask 0x07e007e0 for Green
+    vpsllw     ymm4, ymm4, 10
+    vpsrlw     ymm4, ymm4, 5
+    vpcmpeqb   ymm7, ymm7, ymm7       // generate mask 0xff00ff00 for Alpha
+    vpsllw     ymm7, ymm7, 8
+
+    mov        eax, [esp + 4]   // src_rgb565
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // pix
+    sub        edx, eax
+    sub        edx, eax
+
+ convertloop:
+    vmovdqu    ymm0, [eax]   // fetch 16 pixels of bgr565
+    vpand      ymm1, ymm0, ymm3    // R in upper 5 bits
+    vpsllw     ymm2, ymm0, 11      // B in upper 5 bits
+    vpmulhuw   ymm1, ymm1, ymm5    // * (256 + 8)
+    vpmulhuw   ymm2, ymm2, ymm5    // * (256 + 8)
+    vpsllw     ymm1, ymm1, 8
+    vpor       ymm1, ymm1, ymm2    // RB
+    vpand      ymm0, ymm0, ymm4    // G in middle 6 bits
+    vpmulhuw   ymm0, ymm0, ymm6    // << 5 * (256 + 4)
+    vpor       ymm0, ymm0, ymm7    // AG
+    vpermq     ymm0, ymm0, 0xd8    // mutate for unpack
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpckhbw ymm2, ymm1, ymm0
+    vpunpcklbw ymm1, ymm1, ymm0
+    vmovdqu    [eax * 2 + edx], ymm1  // store 4 pixels of ARGB
+    vmovdqu    [eax * 2 + edx + 32], ymm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 32]
+    sub       ecx, 16
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_RGB565TOARGBROW_AVX2
+
+#ifdef HAS_ARGB1555TOARGBROW_AVX2
+__declspec(naked)
+void ARGB1555ToARGBRow_AVX2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov        eax, 0x01080108  // generate multiplier to repeat 5 bits
+    vmovd      xmm5, eax
+    vbroadcastss ymm5, xmm5
+    mov        eax, 0x42004200  // multiplier shift by 6 and then repeat 5 bits
+    movd       xmm6, eax
+    vbroadcastss ymm6, xmm6
+    vpcmpeqb   ymm3, ymm3, ymm3 // generate mask 0xf800f800 for Red
+    vpsllw     ymm3, ymm3, 11
+    vpsrlw     ymm4, ymm3, 6    // generate mask 0x03e003e0 for Green
+    vpcmpeqb   ymm7, ymm7, ymm7 // generate mask 0xff00ff00 for Alpha
+    vpsllw     ymm7, ymm7, 8
+
+    mov        eax,  [esp + 4]   // src_argb1555
+    mov        edx,  [esp + 8]   // dst_argb
+    mov        ecx,  [esp + 12]  // pix
+    sub        edx,  eax
+    sub        edx,  eax
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 16 pixels of 1555
+    vpsllw     ymm1, ymm0, 1       // R in upper 5 bits
+    vpsllw     ymm2, ymm0, 11      // B in upper 5 bits
+    vpand      ymm1, ymm1, ymm3
+    vpmulhuw   ymm2, ymm2, ymm5    // * (256 + 8)
+    vpmulhuw   ymm1, ymm1, ymm5    // * (256 + 8)
+    vpsllw     ymm1, ymm1, 8
+    vpor       ymm1, ymm1, ymm2    // RB
+    vpsraw     ymm2, ymm0, 8       // A
+    vpand      ymm0, ymm0, ymm4    // G in middle 5 bits
+    vpmulhuw   ymm0, ymm0, ymm6    // << 6 * (256 + 8)
+    vpand      ymm2, ymm2, ymm7
+    vpor       ymm0, ymm0, ymm2    // AG
+    vpermq     ymm0, ymm0, 0xd8    // mutate for unpack
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpckhbw ymm2, ymm1, ymm0
+    vpunpcklbw ymm1, ymm1, ymm0
+    vmovdqu    [eax * 2 + edx], ymm1  // store 8 pixels of ARGB
+    vmovdqu    [eax * 2 + edx + 32], ymm2  // store next 8 pixels of ARGB
+    lea       eax, [eax + 32]
+    sub       ecx, 16
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGB1555TOARGBROW_AVX2
+
+#ifdef HAS_ARGB4444TOARGBROW_AVX2
+__declspec(naked)
+void ARGB4444ToARGBRow_AVX2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax,  0x0f0f0f0f  // generate mask 0x0f0f0f0f
+    vmovd     xmm4, eax
+    vbroadcastss ymm4, xmm4
+    vpslld    ymm5, ymm4, 4     // 0xf0f0f0f0 for high nibbles
+    mov       eax,  [esp + 4]   // src_argb4444
+    mov       edx,  [esp + 8]   // dst_argb
+    mov       ecx,  [esp + 12]  // pix
+    sub       edx,  eax
+    sub       edx,  eax
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 16 pixels of bgra4444
+    vpand      ymm2, ymm0, ymm5    // mask high nibbles
+    vpand      ymm0, ymm0, ymm4    // mask low nibbles
+    vpsrlw     ymm3, ymm2, 4
+    vpsllw     ymm1, ymm0, 4
+    vpor       ymm2, ymm2, ymm3
+    vpor       ymm0, ymm0, ymm1
+    vpermq     ymm0, ymm0, 0xd8    // mutate for unpack
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpckhbw ymm1, ymm0, ymm2
+    vpunpcklbw ymm0, ymm0, ymm2
+    vmovdqu    [eax * 2 + edx], ymm0  // store 8 pixels of ARGB
+    vmovdqu    [eax * 2 + edx + 32], ymm1  // store next 8 pixels of ARGB
+    lea       eax, [eax + 32]
+    sub       ecx, 16
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGB4444TOARGBROW_AVX2
+
+// 24 instructions
+__declspec(naked)
+void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
+    movd      xmm5, eax
+    pshufd    xmm5, xmm5, 0
+    mov       eax, 0x42004200  // multiplier shift by 6 and then repeat 5 bits
+    movd      xmm6, eax
+    pshufd    xmm6, xmm6, 0
+    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    psllw     xmm3, 11
+    movdqa    xmm4, xmm3       // generate mask 0x03e003e0 for Green
+    psrlw     xmm4, 6
+    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    psllw     xmm7, 8
+
+    mov       eax, [esp + 4]   // src_argb1555
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of 1555
+    movdqa    xmm1, xmm0
+    movdqa    xmm2, xmm0
+    psllw     xmm1, 1       // R in upper 5 bits
+    psllw     xmm2, 11      // B in upper 5 bits
+    pand      xmm1, xmm3
+    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    psllw     xmm1, 8
+    por       xmm1, xmm2    // RB
+    movdqa    xmm2, xmm0
+    pand      xmm0, xmm4    // G in middle 5 bits
+    psraw     xmm2, 8       // A
+    pmulhuw   xmm0, xmm6    // << 6 * (256 + 8)
+    pand      xmm2, xmm7
+    por       xmm0, xmm2    // AG
+    movdqa    xmm2, xmm1
+    punpcklbw xmm1, xmm0
+    punpckhbw xmm2, xmm0
+    movdqu    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqu    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+// 18 instructions.
+__declspec(naked)
+void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax, 0x0f0f0f0f  // generate mask 0x0f0f0f0f
+    movd      xmm4, eax
+    pshufd    xmm4, xmm4, 0
+    movdqa    xmm5, xmm4       // 0xf0f0f0f0 for high nibbles
+    pslld     xmm5, 4
+    mov       eax, [esp + 4]   // src_argb4444
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of bgra4444
+    movdqa    xmm2, xmm0
+    pand      xmm0, xmm4    // mask low nibbles
+    pand      xmm2, xmm5    // mask high nibbles
+    movdqa    xmm1, xmm0
+    movdqa    xmm3, xmm2
+    psllw     xmm1, 4
+    psrlw     xmm3, 4
+    por       xmm0, xmm1
+    por       xmm2, xmm3
+    movdqa    xmm1, xmm0
+    punpcklbw xmm0, xmm2
+    punpckhbw xmm1, xmm2
+    movdqu    [eax * 2 + edx], xmm0  // store 4 pixels of ARGB
+    movdqu    [eax * 2 + edx + 16], xmm1  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    movdqa    xmm6, kShuffleMaskARGBToRGB24
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm2, [eax + 32]
+    movdqu    xmm3, [eax + 48]
+    lea       eax, [eax + 64]
+    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm1, xmm6
+    pshufb    xmm2, xmm6
+    pshufb    xmm3, xmm6
+    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
+    psrldq    xmm1, 4      // 8 bytes from 1
+    pslldq    xmm4, 12     // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
+    por       xmm0, xmm4   // 4 bytes from 1 for 0
+    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqu    [edx], xmm0  // store 0
+    por       xmm1, xmm5   // 8 bytes from 2 for 1
+    psrldq    xmm2, 8      // 4 bytes from 2
+    pslldq    xmm3, 4      // 12 bytes from 3 for 2
+    por       xmm2, xmm3   // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1   // store 1
+    movdqu    [edx + 32], xmm2   // store 2
+    lea       edx, [edx + 48]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    movdqa    xmm6, kShuffleMaskARGBToRAW
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm2, [eax + 32]
+    movdqu    xmm3, [eax + 48]
+    lea       eax, [eax + 64]
+    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm1, xmm6
+    pshufb    xmm2, xmm6
+    pshufb    xmm3, xmm6
+    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
+    psrldq    xmm1, 4      // 8 bytes from 1
+    pslldq    xmm4, 12     // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
+    por       xmm0, xmm4   // 4 bytes from 1 for 0
+    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqu    [edx], xmm0  // store 0
+    por       xmm1, xmm5   // 8 bytes from 2 for 1
+    psrldq    xmm2, 8      // 4 bytes from 2
+    pslldq    xmm3, 4      // 12 bytes from 3 for 2
+    por       xmm2, xmm3   // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1   // store 1
+    movdqu    [edx + 32], xmm2   // store 2
+    lea       edx, [edx + 48]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+// 4 pixels
+__declspec(naked)
+void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm3, xmm3       // generate mask 0x0000001f
+    psrld     xmm3, 27
+    pcmpeqb   xmm4, xmm4       // generate mask 0x000007e0
+    psrld     xmm4, 26
+    pslld     xmm4, 5
+    pcmpeqb   xmm5, xmm5       // generate mask 0xfffff800
+    pslld     xmm5, 11
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    pslld     xmm0, 8       // R
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 5       // G
+    psrad     xmm0, 16      // R
+    pand      xmm1, xmm3    // B
+    pand      xmm2, xmm4    // G
+    pand      xmm0, xmm5    // R
+    por       xmm1, xmm2    // BG
+    por       xmm0, xmm1    // BGR
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of RGB565
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+// 8 pixels
+__declspec(naked)
+void ARGBToRGB565DitherRow_SSE2(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int pix) {
+  __asm {
+
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    movd      xmm6, [esp + 12] // dither4
+    mov       ecx, [esp + 16]  // pix
+    punpcklbw xmm6, xmm6       // make dither 16 bytes
+    movdqa    xmm7, xmm6
+    punpcklwd xmm6, xmm6
+    punpckhwd xmm7, xmm7
+    pcmpeqb   xmm3, xmm3       // generate mask 0x0000001f
+    psrld     xmm3, 27
+    pcmpeqb   xmm4, xmm4       // generate mask 0x000007e0
+    psrld     xmm4, 26
+    pslld     xmm4, 5
+    pcmpeqb   xmm5, xmm5       // generate mask 0xfffff800
+    pslld     xmm5, 11
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
+    paddusb   xmm0, xmm6    // add dither
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    pslld     xmm0, 8       // R
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 5       // G
+    psrad     xmm0, 16      // R
+    pand      xmm1, xmm3    // B
+    pand      xmm2, xmm4    // G
+    pand      xmm0, xmm5    // R
+    por       xmm1, xmm2    // BG
+    por       xmm0, xmm1    // BGR
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of RGB565
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTORGB565DITHERROW_AVX2
+__declspec(naked)
+void ARGBToRGB565DitherRow_AVX2(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int pix) {
+  __asm {
+    mov        eax, [esp + 4]      // src_argb
+    mov        edx, [esp + 8]      // dst_rgb
+    vbroadcastss xmm6, [esp + 12]  // dither4
+    mov        ecx, [esp + 16]     // pix
+    vpunpcklbw xmm6, xmm6, xmm6    // make dither 32 bytes
+    vpermq     ymm6, ymm6, 0xd8
+    vpunpcklwd ymm6, ymm6, ymm6
+    vpcmpeqb   ymm3, ymm3, ymm3    // generate mask 0x0000001f
+    vpsrld     ymm3, ymm3, 27
+    vpcmpeqb   ymm4, ymm4, ymm4    // generate mask 0x000007e0
+    vpsrld     ymm4, ymm4, 26
+    vpslld     ymm4, ymm4, 5
+    vpslld     ymm5, ymm3, 11      // generate mask 0x0000f800
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpaddusb   ymm0, ymm0, ymm6    // add dither
+    vpsrld     ymm2, ymm0, 5       // G
+    vpsrld     ymm1, ymm0, 3       // B
+    vpsrld     ymm0, ymm0, 8       // R
+    vpand      ymm2, ymm2, ymm4    // G
+    vpand      ymm1, ymm1, ymm3    // B
+    vpand      ymm0, ymm0, ymm5    // R
+    vpor       ymm1, ymm1, ymm2    // BG
+    vpor       ymm0, ymm0, ymm1    // BGR
+    vpackusdw  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of RGB565
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTORGB565DITHERROW_AVX2
+
+// TODO(fbarchard): Improve sign extension/packing.
+__declspec(naked)
+void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm4, xmm4       // generate mask 0x0000001f
+    psrld     xmm4, 27
+    movdqa    xmm5, xmm4       // generate mask 0x000003e0
+    pslld     xmm5, 5
+    movdqa    xmm6, xmm4       // generate mask 0x00007c00
+    pslld     xmm6, 10
+    pcmpeqb   xmm7, xmm7       // generate mask 0xffff8000
+    pslld     xmm7, 15
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    movdqa    xmm3, xmm0    // R
+    psrad     xmm0, 16      // A
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 6       // G
+    psrld     xmm3, 9       // R
+    pand      xmm0, xmm7    // A
+    pand      xmm1, xmm4    // B
+    pand      xmm2, xmm5    // G
+    pand      xmm3, xmm6    // R
+    por       xmm0, xmm1    // BA
+    por       xmm2, xmm3    // GR
+    por       xmm0, xmm2    // BGRA
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB1555
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm4, xmm4       // generate mask 0xf000f000
+    psllw     xmm4, 12
+    movdqa    xmm3, xmm4       // generate mask 0x00f000f0
+    psrlw     xmm3, 8
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0
+    pand      xmm0, xmm3    // low nibble
+    pand      xmm1, xmm4    // high nibble
+    psrld     xmm0, 4
+    psrld     xmm1, 8
+    por       xmm0, xmm1
+    packuswb  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB4444
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTORGB565ROW_AVX2
+__declspec(naked)
+void ARGBToRGB565Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]      // src_argb
+    mov        edx, [esp + 8]      // dst_rgb
+    mov        ecx, [esp + 12]     // pix
+    vpcmpeqb   ymm3, ymm3, ymm3    // generate mask 0x0000001f
+    vpsrld     ymm3, ymm3, 27
+    vpcmpeqb   ymm4, ymm4, ymm4    // generate mask 0x000007e0
+    vpsrld     ymm4, ymm4, 26
+    vpslld     ymm4, ymm4, 5
+    vpslld     ymm5, ymm3, 11      // generate mask 0x0000f800
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpsrld     ymm2, ymm0, 5       // G
+    vpsrld     ymm1, ymm0, 3       // B
+    vpsrld     ymm0, ymm0, 8       // R
+    vpand      ymm2, ymm2, ymm4    // G
+    vpand      ymm1, ymm1, ymm3    // B
+    vpand      ymm0, ymm0, ymm5    // R
+    vpor       ymm1, ymm1, ymm2    // BG
+    vpor       ymm0, ymm0, ymm1    // BGR
+    vpackusdw  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of RGB565
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTORGB565ROW_AVX2
+
+#ifdef HAS_ARGBTOARGB1555ROW_AVX2
+__declspec(naked)
+void ARGBToARGB1555Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]      // src_argb
+    mov        edx, [esp + 8]      // dst_rgb
+    mov        ecx, [esp + 12]     // pix
+    vpcmpeqb   ymm4, ymm4, ymm4
+    vpsrld     ymm4, ymm4, 27      // generate mask 0x0000001f
+    vpslld     ymm5, ymm4, 5       // generate mask 0x000003e0
+    vpslld     ymm6, ymm4, 10      // generate mask 0x00007c00
+    vpcmpeqb   ymm7, ymm7, ymm7    // generate mask 0xffff8000
+    vpslld     ymm7, ymm7, 15
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpsrld     ymm3, ymm0, 9       // R
+    vpsrld     ymm2, ymm0, 6       // G
+    vpsrld     ymm1, ymm0, 3       // B
+    vpsrad     ymm0, ymm0, 16      // A
+    vpand      ymm3, ymm3, ymm6    // R
+    vpand      ymm2, ymm2, ymm5    // G
+    vpand      ymm1, ymm1, ymm4    // B
+    vpand      ymm0, ymm0, ymm7    // A
+    vpor       ymm0, ymm0, ymm1    // BA
+    vpor       ymm2, ymm2, ymm3    // GR
+    vpor       ymm0, ymm0, ymm2    // BGRA
+    vpackssdw  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of ARGB1555
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOARGB1555ROW_AVX2
+
+#ifdef HAS_ARGBTOARGB4444ROW_AVX2
+__declspec(naked)
+void ARGBToARGB4444Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb
+    mov        edx, [esp + 8]   // dst_rgb
+    mov        ecx, [esp + 12]  // pix
+    vpcmpeqb   ymm4, ymm4, ymm4   // generate mask 0xf000f000
+    vpsllw     ymm4, ymm4, 12
+    vpsrlw     ymm3, ymm4, 8      // generate mask 0x00f000f0
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpand      ymm1, ymm0, ymm4    // high nibble
+    vpand      ymm0, ymm0, ymm3    // low nibble
+    vpsrld     ymm1, ymm1, 8
+    vpsrld     ymm0, ymm0, 4
+    vpor       ymm0, ymm0, ymm1
+    vpackuswb  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of ARGB4444
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOARGB4444ROW_AVX2
+
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+__declspec(naked)
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kARGBToY
+    movdqa     xmm5, kAddY16
+
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+// Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
+// Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
+__declspec(naked)
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    paddw      xmm0, xmm5  // Add .5 for rounding.
+    paddw      xmm2, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// vpermd for vphaddw + vpackuswb vpermd.
+static const lvec32 kPermdARGBToY_AVX = {
+  0, 4, 1, 5, 2, 6, 3, 7
+};
+
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked)
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vbroadcastf128 ymm4, kARGBToY
+    vbroadcastf128 ymm5, kAddY16
+    vmovdqu    ymm6, kPermdARGBToY_AVX
+
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    vpaddb     ymm0, ymm0, ymm5  // add 16 for Y
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYROW_AVX2
+
+#ifdef HAS_ARGBTOYJROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked)
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vbroadcastf128 ymm4, kARGBToYJ
+    vbroadcastf128 ymm5, kAddYJ64
+    vmovdqu    ymm6, kPermdARGBToY_AVX
+
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpaddw     ymm0, ymm0, ymm5  // Add .5 for rounding.
+    vpaddw     ymm2, ymm2, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYJROW_AVX2
+
+__declspec(naked)
+void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kBGRAToY
+    movdqa     xmm5, kAddY16
+
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kABGRToY
+    movdqa     xmm5, kAddY16
+
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void RGBAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kRGBAToY
+    movdqa     xmm5, kAddY16
+
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm5, kAddUVJ128
+    movdqa     xmm6, kARGBToVJ
+    movdqa     xmm7, kARGBToUJ
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    paddw      xmm0, xmm5            // +.5 rounding -> unsigned
+    paddw      xmm1, xmm5
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+__declspec(naked)
+void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    vbroadcastf128 ymm5, kAddUV128
+    vbroadcastf128 ymm6, kARGBToV
+    vbroadcastf128 ymm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 32x2 argb pixels to 16x1 */
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    vpavgb     ymm2, ymm2, [eax + esi + 64]
+    vpavgb     ymm3, ymm3, [eax + esi + 96]
+    lea        eax,  [eax + 128]
+    vshufps    ymm4, ymm0, ymm1, 0x88
+    vshufps    ymm0, ymm0, ymm1, 0xdd
+    vpavgb     ymm0, ymm0, ymm4  // mutated by vshufps
+    vshufps    ymm4, ymm2, ymm3, 0x88
+    vshufps    ymm2, ymm2, ymm3, 0xdd
+    vpavgb     ymm2, ymm2, ymm4  // mutated by vshufps
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 32 different pixels, its 16 pixels of U and 16 of V
+    vpmaddubsw ymm1, ymm0, ymm7  // U
+    vpmaddubsw ymm3, ymm2, ymm7
+    vpmaddubsw ymm0, ymm0, ymm6  // V
+    vpmaddubsw ymm2, ymm2, ymm6
+    vphaddw    ymm1, ymm1, ymm3  // mutates
+    vphaddw    ymm0, ymm0, ymm2
+    vpsraw     ymm1, ymm1, 8
+    vpsraw     ymm0, ymm0, 8
+    vpacksswb  ymm0, ymm1, ymm0  // mutates
+    vpermq     ymm0, ymm0, 0xd8  // For vpacksswb
+    vpshufb    ymm0, ymm0, kShufARGBToUV_AVX  // For vshufps + vphaddw
+    vpaddb     ymm0, ymm0, ymm5  // -> unsigned
+
+    // step 3 - store 16 U and 16 V values
+    vextractf128 [edx], ymm0, 0 // U
+    vextractf128 [edx + edi], ymm0, 1 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOUVROW_AVX2
+
+__declspec(naked)
+void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* convert to U and V */
+    movdqu     xmm0, [eax]          // U
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm7
+    pmaddubsw  xmm1, xmm7
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm3, xmm7
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    movdqu     [edx], xmm0
+
+    movdqu     xmm0, [eax]          // V
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm6
+    pmaddubsw  xmm1, xmm6
+    pmaddubsw  xmm2, xmm6
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    lea        eax,  [eax + 64]
+    movdqu     [edx + edi], xmm0
+    lea        edx,  [edx + 16]
+    sub        ecx,  16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked)
+void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kBGRAToV
+    movdqa     xmm7, kBGRAToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kABGRToV
+    movdqa     xmm7, kABGRToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kRGBAToV
+    movdqa     xmm7, kRGBAToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBTOYROW_SSSE3
+
+// Read 16 UV from 444
+#define READYUV444_AVX2 __asm {                                                \
+    __asm vmovdqu    xmm0, [esi]                  /* U */         /* NOLINT */ \
+    __asm vmovdqu    xmm1, [esi + edi]            /* V */         /* NOLINT */ \
+    __asm lea        esi,  [esi + 16]                                          \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpermq     ymm1, ymm1, 0xd8                                          \
+    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
+  }
+
+// Read 8 UV from 422, upsample to 16 UV.
+#define READYUV422_AVX2 __asm {                                                \
+    __asm vmovq      xmm0, qword ptr [esi]        /* U */         /* NOLINT */ \
+    __asm vmovq      xmm1, qword ptr [esi + edi]  /* V */         /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
+  }
+
+// Read 4 UV from 411, upsample to 16 UV.
+#define READYUV411_AVX2 __asm {                                                \
+    __asm vmovd      xmm0, dword ptr [esi]        /* U */         /* NOLINT */ \
+    __asm vmovd      xmm1, dword ptr [esi + edi]  /* V */         /* NOLINT */ \
+    __asm lea        esi,  [esi + 4]                                           \
+    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
+    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpckldq ymm0, ymm0, ymm0             /* UVUVUVUV (upsample) */    \
+  }
+
+// Read 8 UV from NV12, upsample to 16 UV.
+#define READNV12_AVX2 __asm {                                                  \
+    __asm vmovdqu    xmm0, [esi]                  /* UV */                     \
+    __asm lea        esi,  [esi + 16]                                          \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
+  }
+
+// Convert 16 pixels: 16 UV and 16 Y.
+#define YUVTORGB_AVX2(YuvConstants) __asm {                                    \
+    /* Step 1: Find 8 UV contributions to 16 R,G,B values */                   \
+    __asm vpmaddubsw ymm2, ymm0, YuvConstants.kUVToR        /* scale R UV */   \
+    __asm vpmaddubsw ymm1, ymm0, YuvConstants.kUVToG        /* scale G UV */   \
+    __asm vpmaddubsw ymm0, ymm0, YuvConstants.kUVToB        /* scale B UV */   \
+    __asm vmovdqu    ymm3, YuvConstants.kUVBiasR                               \
+    __asm vpsubw     ymm2, ymm3, ymm2                                          \
+    __asm vmovdqu    ymm3, YuvConstants.kUVBiasG                               \
+    __asm vpsubw     ymm1, ymm3, ymm1                                          \
+    __asm vmovdqu    ymm3, YuvConstants.kUVBiasB                               \
+    __asm vpsubw     ymm0, ymm3, ymm0                                          \
+    /* Step 2: Find Y contribution to 16 R,G,B values */                       \
+    __asm vmovdqu    xmm3, [eax]                  /* NOLINT */                 \
+    __asm lea        eax, [eax + 16]                                           \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpunpcklbw ymm3, ymm3, ymm3                                          \
+    __asm vpmulhuw   ymm3, ymm3, YuvConstants.kYToRgb                          \
+    __asm vpaddsw    ymm0, ymm0, ymm3           /* B += Y */                   \
+    __asm vpaddsw    ymm1, ymm1, ymm3           /* G += Y */                   \
+    __asm vpaddsw    ymm2, ymm2, ymm3           /* R += Y */                   \
+    __asm vpsraw     ymm0, ymm0, 6                                             \
+    __asm vpsraw     ymm1, ymm1, 6                                             \
+    __asm vpsraw     ymm2, ymm2, 6                                             \
+    __asm vpackuswb  ymm0, ymm0, ymm0           /* B */                        \
+    __asm vpackuswb  ymm1, ymm1, ymm1           /* G */                        \
+    __asm vpackuswb  ymm2, ymm2, ymm2           /* R */                        \
+  }
+
+// Store 16 ARGB values.
+#define STOREARGB_AVX2 __asm {                                                 \
+    /* Step 3: Weave into ARGB */                                              \
+    __asm vpunpcklbw ymm0, ymm0, ymm1           /* BG */                       \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpcklbw ymm2, ymm2, ymm5           /* RA */                       \
+    __asm vpermq     ymm2, ymm2, 0xd8                                          \
+    __asm vpunpcklwd ymm1, ymm0, ymm2           /* BGRA first 8 pixels */      \
+    __asm vpunpckhwd ymm0, ymm0, ymm2           /* BGRA next 8 pixels */       \
+    __asm vmovdqu    0[edx], ymm1                                              \
+    __asm vmovdqu    32[edx], ymm0                                             \
+    __asm lea        edx,  [edx + 64]                                          \
+  }
+
+#ifdef HAS_I422TOARGBROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void I422ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I422TOARGBROW_AVX2
+
+#ifdef HAS_J422TOARGBROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void J422ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvJConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_J422TOARGBROW_AVX2
+
+#ifdef HAS_I444TOARGBROW_AVX2
+// 16 pixels
+// 16 UV values with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void I444ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV444_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I444TOARGBROW_AVX2
+
+#ifdef HAS_I411TOARGBROW_AVX2
+// 16 pixels
+// 4 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void I411ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV411_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I411TOARGBROW_AVX2
+
+#ifdef HAS_NV12TOARGBROW_AVX2
+// 16 pixels.
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void NV12ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READNV12_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_NV12TOARGBROW_AVX2
+
+#ifdef HAS_NV21TOARGBROW_AVX2
+// 16 pixels.
+// 8 VU values upsampled to 16 VU, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void NV21ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READNV12_AVX2
+    YUVTORGB_AVX2(kYvuConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_NV21TOARGBROW_AVX2
+
+#ifdef HAS_I422TOBGRAROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 BGRA (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
+__declspec(naked)
+void I422ToBGRARow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into BGRA
+    vpunpcklbw ymm1, ymm1, ymm0           // GB
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm2, ymm5, ymm2           // AR
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklwd ymm0, ymm2, ymm1           // ARGB first 8 pixels
+    vpunpckhwd ymm2, ymm2, ymm1           // ARGB next 8 pixels
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm2
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I422TOBGRAROW_AVX2
+
+#ifdef HAS_I422TORGBAROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
+__declspec(naked)
+void I422ToRGBARow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into RGBA
+    vpunpcklbw ymm1, ymm1, ymm2           // GR
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm2, ymm5, ymm0           // AB
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklwd ymm0, ymm2, ymm1           // ABGR first 8 pixels
+    vpunpckhwd ymm1, ymm2, ymm1           // ABGR next 8 pixels
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I422TORGBAROW_AVX2
+
+#ifdef HAS_I422TOABGRROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ABGR (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
+__declspec(naked)
+void I422ToABGRRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ABGR
+    vpunpcklbw ymm1, ymm2, ymm1           // RG
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm2, ymm0, ymm5           // BA
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklwd ymm0, ymm1, ymm2           // RGBA first 8 pixels
+    vpunpckhwd ymm1, ymm1, ymm2           // RGBA next 8 pixels
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I422TOABGRROW_AVX2
+
+#if defined(HAS_I422TOARGBROW_SSSE3)
+// TODO(fbarchard): Read that does half size on Y and treats 420 as 444.
+
+// Read 8 UV from 444.
+#define READYUV444 __asm {                                                     \
+    __asm movq       xmm0, qword ptr [esi] /* U */                /* NOLINT */ \
+    __asm movq       xmm1, qword ptr [esi + edi] /* V */          /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+  }
+
+// Read 4 UV from 422, upsample to 8 UV.
+#define READYUV422 __asm {                                                     \
+    __asm movd       xmm0, [esi]          /* U */                              \
+    __asm movd       xmm1, [esi + edi]    /* V */                              \
+    __asm lea        esi,  [esi + 4]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Read 2 UV from 411, upsample to 8 UV.
+#define READYUV411 __asm {                                                     \
+    __asm movzx      ebx, word ptr [esi]        /* U */           /* NOLINT */ \
+    __asm movd       xmm0, ebx                                                 \
+    __asm movzx      ebx, word ptr [esi + edi]  /* V */           /* NOLINT */ \
+    __asm movd       xmm1, ebx                                                 \
+    __asm lea        esi,  [esi + 2]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+    __asm punpckldq  xmm0, xmm0           /* UVUVUVUV (upsample) */            \
+  }
+
+// Read 4 UV from NV12, upsample to 8 UV.
+#define READNV12 __asm {                                                       \
+    __asm movq       xmm0, qword ptr [esi] /* UV */               /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Convert 8 pixels: 8 UV and 8 Y.
+#define YUVTORGB(YuvConstants) __asm {                                         \
+    /* Step 1: Find 4 UV contributions to 8 R,G,B values */                    \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm movdqa     xmm2, xmm0                                                \
+    __asm movdqa     xmm3, xmm0                                                \
+    __asm movdqa     xmm0, YuvConstants.kUVBiasB /* unbias back to signed */   \
+    __asm pmaddubsw  xmm1, YuvConstants.kUVToB   /* scale B UV */              \
+    __asm psubw      xmm0, xmm1                                                \
+    __asm movdqa     xmm1, YuvConstants.kUVBiasG                               \
+    __asm pmaddubsw  xmm2, YuvConstants.kUVToG   /* scale G UV */              \
+    __asm psubw      xmm1, xmm2                                                \
+    __asm movdqa     xmm2, YuvConstants.kUVBiasR                               \
+    __asm pmaddubsw  xmm3, YuvConstants.kUVToR   /* scale R UV */              \
+    __asm psubw      xmm2, xmm3                                                \
+    /* Step 2: Find Y contribution to 8 R,G,B values */                        \
+    __asm movq       xmm3, qword ptr [eax]                        /* NOLINT */ \
+    __asm lea        eax, [eax + 8]                                            \
+    __asm punpcklbw  xmm3, xmm3                                                \
+    __asm pmulhuw    xmm3, YuvConstants.kYToRgb                                \
+    __asm paddsw     xmm0, xmm3           /* B += Y */                         \
+    __asm paddsw     xmm1, xmm3           /* G += Y */                         \
+    __asm paddsw     xmm2, xmm3           /* R += Y */                         \
+    __asm psraw      xmm0, 6                                                   \
+    __asm psraw      xmm1, 6                                                   \
+    __asm psraw      xmm2, 6                                                   \
+    __asm packuswb   xmm0, xmm0           /* B */                              \
+    __asm packuswb   xmm1, xmm1           /* G */                              \
+    __asm packuswb   xmm2, xmm2           /* R */                              \
+  }
+
+// Store 8 ARGB values.
+#define STOREARGB __asm {                                                      \
+    /* Step 3: Weave into ARGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm5           /* RA */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRA first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRA next 4 pixels */             \
+    __asm movdqu     0[edx], xmm0                                              \
+    __asm movdqu     16[edx], xmm1                                             \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 BGRA values.
+#define STOREBGRA __asm {                                                      \
+    /* Step 3: Weave into BGRA */                                              \
+    __asm pcmpeqb    xmm5, xmm5           /* generate 0xffffffff for alpha */  \
+    __asm punpcklbw  xmm1, xmm0           /* GB */                             \
+    __asm punpcklbw  xmm5, xmm2           /* AR */                             \
+    __asm movdqa     xmm0, xmm5                                                \
+    __asm punpcklwd  xmm5, xmm1           /* BGRA first 4 pixels */            \
+    __asm punpckhwd  xmm0, xmm1           /* BGRA next 4 pixels */             \
+    __asm movdqu     0[edx], xmm5                                              \
+    __asm movdqu     16[edx], xmm0                                             \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 ABGR values.
+#define STOREABGR __asm {                                                      \
+    /* Step 3: Weave into ABGR */                                              \
+    __asm punpcklbw  xmm2, xmm1           /* RG */                             \
+    __asm punpcklbw  xmm0, xmm5           /* BA */                             \
+    __asm movdqa     xmm1, xmm2                                                \
+    __asm punpcklwd  xmm2, xmm0           /* RGBA first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm0           /* RGBA next 4 pixels */             \
+    __asm movdqu     0[edx], xmm2                                              \
+    __asm movdqu     16[edx], xmm1                                             \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 RGBA values.
+#define STORERGBA __asm {                                                      \
+    /* Step 3: Weave into RGBA */                                              \
+    __asm pcmpeqb    xmm5, xmm5           /* generate 0xffffffff for alpha */  \
+    __asm punpcklbw  xmm1, xmm2           /* GR */                             \
+    __asm punpcklbw  xmm5, xmm0           /* AB */                             \
+    __asm movdqa     xmm0, xmm5                                                \
+    __asm punpcklwd  xmm5, xmm1           /* RGBA first 4 pixels */            \
+    __asm punpckhwd  xmm0, xmm1           /* RGBA next 4 pixels */             \
+    __asm movdqu     0[edx], xmm5                                              \
+    __asm movdqu     16[edx], xmm0                                             \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 RGB24 values.
+#define STORERGB24 __asm {                                                     \
+    /* Step 3: Weave into RRGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
+    /* Step 4: RRGB -> RGB24 */                                                \
+    __asm pshufb     xmm0, xmm5           /* Pack first 8 and last 4 bytes. */ \
+    __asm pshufb     xmm1, xmm6           /* Pack first 12 bytes. */           \
+    __asm palignr    xmm1, xmm0, 12       /* last 4 bytes of xmm0 + 12 xmm1 */ \
+    __asm movq       qword ptr 0[edx], xmm0  /* First 8 bytes */               \
+    __asm movdqu     8[edx], xmm1         /* Last 16 bytes */                  \
+    __asm lea        edx,  [edx + 24]                                          \
+  }
+
+// Store 8 RAW values.
+#define STORERAW __asm {                                                       \
+    /* Step 3: Weave into RRGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
+    /* Step 4: RRGB -> RAW */                                                  \
+    __asm pshufb     xmm0, xmm5           /* Pack first 8 and last 4 bytes. */ \
+    __asm pshufb     xmm1, xmm6           /* Pack first 12 bytes. */           \
+    __asm palignr    xmm1, xmm0, 12       /* last 4 bytes of xmm0 + 12 xmm1 */ \
+    __asm movq       qword ptr 0[edx], xmm0  /* First 8 bytes */               \
+    __asm movdqu     8[edx], xmm1         /* Last 16 bytes */                  \
+    __asm lea        edx,  [edx + 24]                                          \
+  }
+
+// Store 8 RGB565 values.
+#define STORERGB565 __asm {                                                    \
+    /* Step 3: Weave into RRGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
+    /* Step 4: RRGB -> RGB565 */                                               \
+    __asm movdqa     xmm3, xmm0    /* B  first 4 pixels of argb */             \
+    __asm movdqa     xmm2, xmm0    /* G */                                     \
+    __asm pslld      xmm0, 8       /* R */                                     \
+    __asm psrld      xmm3, 3       /* B */                                     \
+    __asm psrld      xmm2, 5       /* G */                                     \
+    __asm psrad      xmm0, 16      /* R */                                     \
+    __asm pand       xmm3, xmm5    /* B */                                     \
+    __asm pand       xmm2, xmm6    /* G */                                     \
+    __asm pand       xmm0, xmm7    /* R */                                     \
+    __asm por        xmm3, xmm2    /* BG */                                    \
+    __asm por        xmm0, xmm3    /* BGR */                                   \
+    __asm movdqa     xmm3, xmm1    /* B  next 4 pixels of argb */              \
+    __asm movdqa     xmm2, xmm1    /* G */                                     \
+    __asm pslld      xmm1, 8       /* R */                                     \
+    __asm psrld      xmm3, 3       /* B */                                     \
+    __asm psrld      xmm2, 5       /* G */                                     \
+    __asm psrad      xmm1, 16      /* R */                                     \
+    __asm pand       xmm3, xmm5    /* B */                                     \
+    __asm pand       xmm2, xmm6    /* G */                                     \
+    __asm pand       xmm1, xmm7    /* R */                                     \
+    __asm por        xmm3, xmm2    /* BG */                                    \
+    __asm por        xmm1, xmm3    /* BGR */                                   \
+    __asm packssdw   xmm0, xmm1                                                \
+    __asm movdqu     0[edx], xmm0  /* store 8 pixels of RGB565 */              \
+    __asm lea        edx, [edx + 16]                                           \
+  }
+
+// 8 pixels.
+// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked)
+void I444ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READYUV444
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB24 (24 bytes).
+__declspec(naked)
+void I422ToRGB24Row_SSSE3(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* dst_rgb24,
+                          int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgb24
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    movdqa     xmm5, kShuffleMaskARGBToRGB24_0
+    movdqa     xmm6, kShuffleMaskARGBToRGB24
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STORERGB24
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RAW (24 bytes).
+__declspec(naked)
+void I422ToRAWRow_SSSE3(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_raw,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // raw
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    movdqa     xmm5, kShuffleMaskARGBToRAW_0
+    movdqa     xmm6, kShuffleMaskARGBToRAW
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STORERAW
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB565 (16 bytes).
+__declspec(naked)
+void I422ToRGB565Row_SSSE3(const uint8* y_buf,
+                           const uint8* u_buf,
+                           const uint8* v_buf,
+                           uint8* rgb565_buf,
+                           int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgb565
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5       // generate mask 0x0000001f
+    psrld      xmm5, 27
+    pcmpeqb    xmm6, xmm6       // generate mask 0x000007e0
+    psrld      xmm6, 26
+    pslld      xmm6, 5
+    pcmpeqb    xmm7, xmm7       // generate mask 0xfffff800
+    pslld      xmm7, 11
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STORERGB565
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked)
+void I422ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// JPeg color space version of I422ToARGB
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked)
+void J422ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvJConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// Similar to I420 but duplicate UV once more.
+__declspec(naked)
+void I411ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]   // Y
+    mov        esi, [esp + 12 + 8]   // U
+    mov        edi, [esp + 12 + 12]  // V
+    mov        edx, [esp + 12 + 16]  // argb
+    mov        ecx, [esp + 12 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5            // generate 0xffffffff for alpha
+
+ convertloop:
+    READYUV411  // modifies EBX
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// 8 pixels.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked)
+void NV12ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* uv_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READNV12
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// 4 VU values upsampled to 8 VU, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked)
+void NV21ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* uv_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READNV12
+    YUVTORGB(kYvuConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void I422ToBGRARow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_bgra,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // bgra
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREBGRA
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void I422ToABGRRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_abgr,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // abgr
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREABGR
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void I422ToRGBARow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_rgba,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgba
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STORERGBA
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+#ifdef HAS_I400TOARGBROW_SSE2
+// 8 pixels of Y converted to 8 pixels of ARGB (32 bytes).
+__declspec(naked)
+void I400ToARGBRow_SSE2(const uint8* y_buf,
+                        uint8* rgb_buf,
+                        int width) {
+  __asm {
+    mov        eax, 0x4a354a35      // 4a35 = 18997 = round(1.164 * 64 * 256)
+    movd       xmm2, eax
+    pshufd     xmm2, xmm2,0
+    mov        eax, 0x04880488      // 0488 = 1160 = round(1.164 * 64 * 16)
+    movd       xmm3, eax
+    pshufd     xmm3, xmm3, 0
+    pcmpeqb    xmm4, xmm4           // generate mask 0xff000000
+    pslld      xmm4, 24
+
+    mov        eax, [esp + 4]       // Y
+    mov        edx, [esp + 8]       // rgb
+    mov        ecx, [esp + 12]      // width
+
+ convertloop:
+    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+    movq       xmm0, qword ptr [eax]
+    lea        eax, [eax + 8]
+    punpcklbw  xmm0, xmm0           // Y.Y
+    pmulhuw    xmm0, xmm2
+    psubusw    xmm0, xmm3
+    psrlw      xmm0, 6
+    packuswb   xmm0, xmm0           // G
+
+    // Step 2: Weave into ARGB
+    punpcklbw  xmm0, xmm0           // GG
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm1           // BGRA next 4 pixels
+    por        xmm0, xmm4
+    por        xmm1, xmm4
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_I400TOARGBROW_SSE2
+
+#ifdef HAS_I400TOARGBROW_AVX2
+// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
+// note: vpunpcklbw mutates and vpackuswb unmutates.
+__declspec(naked)
+void I400ToARGBRow_AVX2(const uint8* y_buf,
+                        uint8* rgb_buf,
+                        int width) {
+  __asm {
+    mov        eax, 0x4a354a35      // 4a35 = 18997 = round(1.164 * 64 * 256)
+    vmovd      xmm2, eax
+    vbroadcastss ymm2, xmm2
+    mov        eax, 0x04880488      // 0488 = 1160 = round(1.164 * 64 * 16)
+    vmovd      xmm3, eax
+    vbroadcastss ymm3, xmm3
+    vpcmpeqb   ymm4, ymm4, ymm4     // generate mask 0xff000000
+    vpslld     ymm4, ymm4, 24
+
+    mov        eax, [esp + 4]       // Y
+    mov        edx, [esp + 8]       // rgb
+    mov        ecx, [esp + 12]      // width
+
+ convertloop:
+    // Step 1: Scale Y contriportbution to 16 G values. G = (y - 16) * 1.164
+    vmovdqu    xmm0, [eax]
+    lea        eax, [eax + 16]
+    vpermq     ymm0, ymm0, 0xd8           // vpunpcklbw mutates
+    vpunpcklbw ymm0, ymm0, ymm0           // Y.Y
+    vpmulhuw   ymm0, ymm0, ymm2
+    vpsubusw   ymm0, ymm0, ymm3
+    vpsrlw     ymm0, ymm0, 6
+    vpackuswb  ymm0, ymm0, ymm0           // G.  still mutated: 3120
+
+    // TODO(fbarchard): Weave alpha with unpack.
+    // Step 2: Weave into ARGB
+    vpunpcklbw ymm1, ymm0, ymm0           // GG - mutates
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklwd ymm0, ymm1, ymm1           // GGGG first 8 pixels
+    vpunpckhwd ymm1, ymm1, ymm1           // GGGG next 8 pixels
+    vpor       ymm0, ymm0, ymm4
+    vpor       ymm1, ymm1, ymm4
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I400TOARGBROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static const uvec8 kShuffleMirror = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+// TODO(fbarchard): Replace lea with -16 offset.
+__declspec(naked)
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    movdqa    xmm5, kShuffleMirror
+
+ convertloop:
+    movdqu    xmm0, [eax - 16 + ecx]
+    pshufb    xmm0, xmm5
+    movdqu    [edx], xmm0
+    lea       edx, [edx + 16]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_SSSE3
+
+#ifdef HAS_MIRRORROW_AVX2
+__declspec(naked)
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    vbroadcastf128 ymm5, kShuffleMirror
+
+ convertloop:
+    vmovdqu   ymm0, [eax - 32 + ecx]
+    vpshufb   ymm0, ymm0, ymm5
+    vpermq    ymm0, ymm0, 0x4e  // swap high and low halfs
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    sub       ecx, 32
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSE2
+__declspec(naked)
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+
+ convertloop:
+    movdqu    xmm0, [eax - 16 + ecx]
+    movdqa    xmm1, xmm0        // swap bytes
+    psllw     xmm0, 8
+    psrlw     xmm1, 8
+    por       xmm0, xmm1
+    pshuflw   xmm0, xmm0, 0x1b  // swap words
+    pshufhw   xmm0, xmm0, 0x1b
+    pshufd    xmm0, xmm0, 0x4e  // swap qwords
+    movdqu    [edx], xmm0
+    lea       edx, [edx + 16]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_SSE2
+
+#ifdef HAS_MIRRORROW_UV_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static const uvec8 kShuffleMirrorUV = {
+  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
+};
+
+__declspec(naked)
+void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
+                       int width) {
+  __asm {
+    push      edi
+    mov       eax, [esp + 4 + 4]   // src
+    mov       edx, [esp + 4 + 8]   // dst_u
+    mov       edi, [esp + 4 + 12]  // dst_v
+    mov       ecx, [esp + 4 + 16]  // width
+    movdqa    xmm1, kShuffleMirrorUV
+    lea       eax, [eax + ecx * 2 - 16]
+    sub       edi, edx
+
+ convertloop:
+    movdqu    xmm0, [eax]
+    lea       eax, [eax - 16]
+    pshufb    xmm0, xmm1
+    movlpd    qword ptr [edx], xmm0
+    movhpd    qword ptr [edx + edi], xmm0
+    lea       edx, [edx + 8]
+    sub       ecx, 8
+    jg        convertloop
+
+    pop       edi
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_UV_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_SSE2
+__declspec(naked)
+void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 16 + ecx * 4]  // last 4 pixels.
+
+ convertloop:
+    movdqu    xmm0, [eax]
+    lea       eax, [eax - 16]
+    pshufd    xmm0, xmm0, 0x1b
+    movdqu    [edx], xmm0
+    lea       edx, [edx + 16]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_SSE2
+
+#ifdef HAS_ARGBMIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec32 kARGBShuffleMirror_AVX2 = {
+  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked)
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    vmovdqu   ymm5, kARGBShuffleMirror_AVX2
+
+ convertloop:
+    vpermd    ymm0, ymm5, [eax - 32 + ecx * 4]  // permute dword order
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    sub       ecx, 8
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_AVX2
+
+#ifdef HAS_SPLITUVROW_SSE2
+__declspec(naked)
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    movdqa     xmm3, xmm1
+    pand       xmm0, xmm5   // even bytes
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm2, 8      // odd bytes
+    psrlw      xmm3, 8
+    packuswb   xmm2, xmm3
+    movdqu     [edx], xmm0
+    movdqu     [edx + edi], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+#endif  // HAS_SPLITUVROW_SSE2
+
+#ifdef HAS_SPLITUVROW_AVX2
+__declspec(naked)
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm2, ymm0, 8      // odd bytes
+    vpsrlw     ymm3, ymm1, 8
+    vpand      ymm0, ymm0, ymm5   // even bytes
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1
+    vpackuswb  ymm2, ymm2, ymm3
+    vpermq     ymm0, ymm0, 0xd8
+    vpermq     ymm2, ymm2, 0xd8
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + edi], ymm2
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SPLITUVROW_AVX2
+
+#ifdef HAS_MERGEUVROW_SSE2
+__declspec(naked)
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+  convertloop:
+    movdqu     xmm0, [eax]      // read 16 U's
+    movdqu     xmm1, [eax + edx]  // and 16 V's
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, xmm0
+    punpcklbw  xmm0, xmm1       // first 8 UV pairs
+    punpckhbw  xmm2, xmm1       // next 8 UV pairs
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  //  HAS_MERGEUVROW_SSE2
+
+#ifdef HAS_MERGEUVROW_AVX2
+__declspec(naked)
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+  convertloop:
+    vmovdqu    ymm0, [eax]           // read 32 U's
+    vmovdqu    ymm1, [eax + edx]     // and 32 V's
+    lea        eax,  [eax + 32]
+    vpunpcklbw ymm2, ymm0, ymm1      // low 16 UV pairs. mutated qqword 0,2
+    vpunpckhbw ymm0, ymm0, ymm1      // high 16 UV pairs. mutated qqword 1,3
+    vextractf128 [edi], ymm2, 0       // bytes 0..15
+    vextractf128 [edi + 16], ymm0, 0  // bytes 16..31
+    vextractf128 [edi + 32], ymm2, 1  // bytes 32..47
+    vextractf128 [edi + 48], ymm0, 1  // bytes 47..63
+    lea        edi, [edi + 64]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_MERGEUVROW_AVX2
+
+#ifdef HAS_COPYROW_SSE2
+// CopyRow copys 'count' bytes using a 16 byte load/store, 32 bytes at time.
+__declspec(naked)
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_COPYROW_SSE2
+
+#ifdef HAS_COPYROW_AVX
+// CopyRow copys 'count' bytes using a 32 byte load/store, 64 bytes at time.
+__declspec(naked)
+void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax, [eax + 64]
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx, [edx + 64]
+    sub        ecx, 64
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_COPYROW_AVX
+
+// Multiple of 1.
+__declspec(naked)
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, esi
+    mov        edx, edi
+    mov        esi, [esp + 4]   // src
+    mov        edi, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    rep movsb
+    mov        edi, edx
+    mov        esi, eax
+    ret
+  }
+}
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+__declspec(naked)
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    pslld      xmm0, 24
+    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    psrld      xmm1, 8
+
+  convertloop:
+    movdqu     xmm2, [eax]
+    movdqu     xmm3, [eax + 16]
+    lea        eax, [eax + 32]
+    movdqu     xmm4, [edx]
+    movdqu     xmm5, [edx + 16]
+    pand       xmm2, xmm0
+    pand       xmm3, xmm0
+    pand       xmm4, xmm1
+    pand       xmm5, xmm1
+    por        xmm2, xmm4
+    por        xmm3, xmm5
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm3
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYALPHAROW_AVX2
+// width in pixels
+__declspec(naked)
+void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+
+  convertloop:
+    vmovdqu    ymm1, [eax]
+    vmovdqu    ymm2, [eax + 32]
+    lea        eax, [eax + 64]
+    vpblendvb  ymm1, ymm1, [edx], ymm0
+    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_AVX2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
+// width in pixels
+__declspec(naked)
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    pslld      xmm0, 24
+    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    psrld      xmm1, 8
+
+  convertloop:
+    movq       xmm2, qword ptr [eax]  // 8 Y's
+    lea        eax, [eax + 8]
+    punpcklbw  xmm2, xmm2
+    punpckhwd  xmm3, xmm2
+    punpcklwd  xmm2, xmm2
+    movdqu     xmm4, [edx]
+    movdqu     xmm5, [edx + 16]
+    pand       xmm2, xmm0
+    pand       xmm3, xmm0
+    pand       xmm4, xmm1
+    pand       xmm5, xmm1
+    por        xmm2, xmm4
+    por        xmm3, xmm5
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm3
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
+// width in pixels
+__declspec(naked)
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+
+  convertloop:
+    vpmovzxbd  ymm1, qword ptr [eax]
+    vpmovzxbd  ymm2, qword ptr [eax + 8]
+    lea        eax, [eax + 16]
+    vpslld     ymm1, ymm1, 24
+    vpslld     ymm2, ymm2, 24
+    vpblendvb  ymm1, ymm1, [edx], ymm0
+    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
+
+#ifdef HAS_SETROW_X86
+// Write 'count' bytes using an 8 bit value repeated.
+// Count should be multiple of 4.
+__declspec(naked)
+void SetRow_X86(uint8* dst, uint8 v8, int count) {
+  __asm {
+    movzx      eax, byte ptr [esp + 8]    // v8
+    mov        edx, 0x01010101  // Duplicate byte to all bytes.
+    mul        edx              // overwrites edx with upper part of result.
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        ecx, [esp + 12]  // count
+    shr        ecx, 2
+    rep stosd
+    mov        edi, edx
+    ret
+  }
+}
+
+// Write 'count' bytes using an 8 bit value repeated.
+__declspec(naked)
+void SetRow_ERMS(uint8* dst, uint8 v8, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        eax, [esp + 8]   // v8
+    mov        ecx, [esp + 12]  // count
+    rep stosb
+    mov        edi, edx
+    ret
+  }
+}
+
+// Write 'count' 32 bit values.
+__declspec(naked)
+void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        eax, [esp + 8]   // v32
+    mov        ecx, [esp + 12]  // count
+    rep stosd
+    mov        edi, edx
+    ret
+  }
+}
+#endif  // HAS_SETROW_X86
+
+#ifdef HAS_YUY2TOYROW_AVX2
+__declspec(naked)
+void YUY2ToYRow_AVX2(const uint8* src_yuy2,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // even bytes are Y
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked)
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked)
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToYRow_AVX2(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // odd bytes are Y
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_YUY2TOYROW_AVX2
+
+#ifdef HAS_YUY2TOYROW_SSE2
+__declspec(naked)
+void YUY2ToYRow_SSE2(const uint8* src_yuy2,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // even bytes are Y
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToYRow_SSE2(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8    // odd bytes are Y
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  // HAS_YUY2TOYROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSE2
+// Blend 8 pixels at a time.
+__declspec(naked)
+void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm7, xmm7       // generate constant 1
+    psrlw      xmm7, 15
+    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    psrlw      xmm6, 8
+    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    psllw      xmm5, 8
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+    sub        ecx, 4
+    jl         convertloop4b    // less than 4 pixels?
+
+    // 4 pixel loop.
+  convertloop4:
+    movdqu     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqu     xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqu     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jge        convertloop4
+
+  convertloop4b:
+    add        ecx, 4 - 1
+    jl         convertloop1b
+
+    // 1 pixel loop.
+  convertloop1:
+    movd       xmm3, [eax]      // src argb
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jge        convertloop1
+
+  convertloop1b:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBBLENDROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSSE3
+// Shuffle table for isolating alpha.
+static const uvec8 kShuffleAlpha = {
+  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
+  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
+};
+// Same as SSE2, but replaces:
+//    psrlw      xmm3, 8          // alpha
+//    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+//    pshuflw    xmm3, xmm3, 0F5h
+// with..
+//    pshufb     xmm3, kShuffleAlpha // alpha
+// Blend 8 pixels at a time.
+
+__declspec(naked)
+void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
+                        uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm7, xmm7       // generate constant 0x0001
+    psrlw      xmm7, 15
+    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    psrlw      xmm6, 8
+    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    psllw      xmm5, 8
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+    sub        ecx, 4
+    jl         convertloop4b    // less than 4 pixels?
+
+    // 4 pixel loop.
+  convertloop4:
+    movdqu     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqu     xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqu     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jge        convertloop4
+
+  convertloop4b:
+    add        ecx, 4 - 1
+    jl         convertloop1b
+
+    // 1 pixel loop.
+  convertloop1:
+    movd       xmm3, [eax]      // src argb
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jge        convertloop1
+
+  convertloop1b:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBBLENDROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+// Attenuate 4 pixels at a time.
+__declspec(naked)
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+    pcmpeqb    xmm5, xmm5       // generate mask 0x00ffffff
+    psrld      xmm5, 8
+
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    punpcklbw  xmm0, xmm0       // first 2
+    pshufhw    xmm2, xmm0, 0FFh // 8 alpha words
+    pshuflw    xmm2, xmm2, 0FFh
+    pmulhuw    xmm0, xmm2       // rgb * a
+    movdqu     xmm1, [eax]      // read 4 pixels
+    punpckhbw  xmm1, xmm1       // next 2 pixels
+    pshufhw    xmm2, xmm1, 0FFh // 8 alpha words
+    pshuflw    xmm2, xmm2, 0FFh
+    pmulhuw    xmm1, xmm2       // rgb * a
+    movdqu     xmm2, [eax]      // alphas
+    lea        eax, [eax + 16]
+    psrlw      xmm0, 8
+    pand       xmm2, xmm4
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    pand       xmm0, xmm5       // keep original alphas
+    por        xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha0 = {
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+};
+static const uvec8 kShuffleAlpha1 = {
+  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+};
+__declspec(naked)
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm3, xmm3       // generate mask 0xff000000
+    pslld      xmm3, 24
+    movdqa     xmm4, kShuffleAlpha0
+    movdqa     xmm5, kShuffleAlpha1
+
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    pshufb     xmm0, xmm4       // isolate first 2 alphas
+    movdqu     xmm1, [eax]      // read 4 pixels
+    punpcklbw  xmm1, xmm1       // first 2 pixel rgbs
+    pmulhuw    xmm0, xmm1       // rgb * a
+    movdqu     xmm1, [eax]      // read 4 pixels
+    pshufb     xmm1, xmm5       // isolate next 2 alphas
+    movdqu     xmm2, [eax]      // read 4 pixels
+    punpckhbw  xmm2, xmm2       // next 2 pixel rgbs
+    pmulhuw    xmm1, xmm2       // rgb * a
+    movdqu     xmm2, [eax]      // mask original alpha
+    lea        eax, [eax + 16]
+    pand       xmm2, xmm3
+    psrlw      xmm0, 8
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    por        xmm0, xmm2       // copy original alpha
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha_AVX2 = {
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
+};
+__declspec(naked)
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vbroadcastf128 ymm4,kShuffleAlpha_AVX2
+    vpcmpeqb   ymm5, ymm5, ymm5 // generate mask 0xff000000
+    vpslld     ymm5, ymm5, 24
+
+ convertloop:
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpshufb    ymm2, ymm0, ymm4  // low 4 alphas
+    vpshufb    ymm3, ymm1, ymm4  // high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * a
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * a
+    vpand      ymm6, ymm6, ymm5  // isolate alpha
+    vpsrlw     ymm0, ymm0, 8
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    vpor       ymm0, ymm0, ymm6  // copy original alpha
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+// Unattenuate 4 pixels at a time.
+__declspec(naked)
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb0
+    mov        edx, [esp + 8 + 8]   // dst_argb
+    mov        ecx, [esp + 8 + 12]  // width
+
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    movzx      esi, byte ptr [eax + 3]  // first alpha
+    movzx      edi, byte ptr [eax + 7]  // second alpha
+    punpcklbw  xmm0, xmm0       // first 2
+    movd       xmm2, dword ptr fixed_invtbl8[esi * 4]
+    movd       xmm3, dword ptr fixed_invtbl8[edi * 4]
+    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words.  1, a, a, a
+    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    movlhps    xmm2, xmm3
+    pmulhuw    xmm0, xmm2       // rgb * a
+
+    movdqu     xmm1, [eax]      // read 4 pixels
+    movzx      esi, byte ptr [eax + 11]  // third alpha
+    movzx      edi, byte ptr [eax + 15]  // forth alpha
+    punpckhbw  xmm1, xmm1       // next 2
+    movd       xmm2, dword ptr fixed_invtbl8[esi * 4]
+    movd       xmm3, dword ptr fixed_invtbl8[edi * 4]
+    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words
+    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    movlhps    xmm2, xmm3
+    pmulhuw    xmm1, xmm2       // rgb * a
+    lea        eax, [eax + 16]
+
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBUNATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kUnattenShuffleAlpha_AVX2 = {
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u
+};
+// TODO(fbarchard): Enable USE_GATHER for future hardware if faster.
+// USE_GATHER is not on by default, due to being a slow instruction.
+#ifdef USE_GATHER
+__declspec(naked)
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vbroadcastf128 ymm4, kUnattenShuffleAlpha_AVX2
+
+ convertloop:
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0xffffffff for gather.
+    vpsrld     ymm2, ymm6, 24    // alpha in low 8 bits.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpgatherdd ymm3, [ymm2 * 4 + fixed_invtbl8], ymm5  // ymm5 cleared.  1, a
+    vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
+    vpunpckhwd ymm3, ymm3, ymm3  // high 4 inverted alphas. mutated.
+    vpshufb    ymm2, ymm2, ymm4  // replicate low 4 alphas. 1, a, a, a
+    vpshufb    ymm3, ymm3, ymm4  // replicate high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#else  // USE_GATHER
+__declspec(naked)
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vbroadcastf128 ymm5, kUnattenShuffleAlpha_AVX2
+
+    push       esi
+    push       edi
+
+ convertloop:
+    // replace VPGATHER
+    movzx      esi, byte ptr [eax + 3]                 // alpha0
+    movzx      edi, byte ptr [eax + 7]                 // alpha1
+    vmovd      xmm0, dword ptr fixed_invtbl8[esi * 4]  // [1,a0]
+    vmovd      xmm1, dword ptr fixed_invtbl8[edi * 4]  // [1,a1]
+    movzx      esi, byte ptr [eax + 11]                // alpha2
+    movzx      edi, byte ptr [eax + 15]                // alpha3
+    vpunpckldq xmm6, xmm0, xmm1                        // [1,a1,1,a0]
+    vmovd      xmm2, dword ptr fixed_invtbl8[esi * 4]  // [1,a2]
+    vmovd      xmm3, dword ptr fixed_invtbl8[edi * 4]  // [1,a3]
+    movzx      esi, byte ptr [eax + 19]                // alpha4
+    movzx      edi, byte ptr [eax + 23]                // alpha5
+    vpunpckldq xmm7, xmm2, xmm3                        // [1,a3,1,a2]
+    vmovd      xmm0, dword ptr fixed_invtbl8[esi * 4]  // [1,a4]
+    vmovd      xmm1, dword ptr fixed_invtbl8[edi * 4]  // [1,a5]
+    movzx      esi, byte ptr [eax + 27]                // alpha6
+    movzx      edi, byte ptr [eax + 31]                // alpha7
+    vpunpckldq xmm0, xmm0, xmm1                        // [1,a5,1,a4]
+    vmovd      xmm2, dword ptr fixed_invtbl8[esi * 4]  // [1,a6]
+    vmovd      xmm3, dword ptr fixed_invtbl8[edi * 4]  // [1,a7]
+    vpunpckldq xmm2, xmm2, xmm3                        // [1,a7,1,a6]
+    vpunpcklqdq xmm3, xmm6, xmm7                       // [1,a3,1,a2,1,a1,1,a0]
+    vpunpcklqdq xmm0, xmm0, xmm2                       // [1,a7,1,a6,1,a5,1,a4]
+    vinserti128 ymm3, ymm3, xmm0, 1 // [1,a7,1,a6,1,a5,1,a4,1,a3,1,a2,1,a1,1,a0]
+    // end of VPGATHER
+
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
+    vpunpckhwd ymm3, ymm3, ymm3  // high 4 inverted alphas. mutated.
+    vpshufb    ymm2, ymm2, ymm5  // replicate low 4 alphas. 1, a, a, a
+    vpshufb    ymm3, ymm3, ymm5  // replicate high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // USE_GATHER
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBGRAYROW_SSSE3
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels.
+__declspec(naked)
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]  /* width */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+ convertloop:
+    movdqu     xmm0, [eax]  // G
+    movdqu     xmm1, [eax + 16]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    phaddw     xmm0, xmm1
+    paddw      xmm0, xmm5  // Add .5 for rounding.
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0   // 8 G bytes
+    movdqu     xmm2, [eax]  // A
+    movdqu     xmm3, [eax + 16]
+    lea        eax, [eax + 32]
+    psrld      xmm2, 24
+    psrld      xmm3, 24
+    packuswb   xmm2, xmm3
+    packuswb   xmm2, xmm2   // 8 A bytes
+    movdqa     xmm3, xmm0   // Weave into GG, GA, then GGGA
+    punpcklbw  xmm0, xmm0   // 8 GG words
+    punpcklbw  xmm3, xmm2   // 8 GA words
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm3   // GGGA first 4
+    punpckhwd  xmm1, xmm3   // GGGA next 4
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBGRAYROW_SSSE3
+
+#ifdef HAS_ARGBSEPIAROW_SSSE3
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+// Constant for ARGB color to sepia tone.
+static const vec8 kARGBToSepiaB = {
+  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
+};
+
+static const vec8 kARGBToSepiaG = {
+  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
+};
+
+static const vec8 kARGBToSepiaR = {
+  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
+};
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+__declspec(naked)
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* dst_argb */
+    mov        ecx, [esp + 8]   /* width */
+    movdqa     xmm2, kARGBToSepiaB
+    movdqa     xmm3, kARGBToSepiaG
+    movdqa     xmm4, kARGBToSepiaR
+
+ convertloop:
+    movdqu     xmm0, [eax]  // B
+    movdqu     xmm6, [eax + 16]
+    pmaddubsw  xmm0, xmm2
+    pmaddubsw  xmm6, xmm2
+    phaddw     xmm0, xmm6
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0   // 8 B values
+    movdqu     xmm5, [eax]  // G
+    movdqu     xmm1, [eax + 16]
+    pmaddubsw  xmm5, xmm3
+    pmaddubsw  xmm1, xmm3
+    phaddw     xmm5, xmm1
+    psrlw      xmm5, 7
+    packuswb   xmm5, xmm5   // 8 G values
+    punpcklbw  xmm0, xmm5   // 8 BG values
+    movdqu     xmm5, [eax]  // R
+    movdqu     xmm1, [eax + 16]
+    pmaddubsw  xmm5, xmm4
+    pmaddubsw  xmm1, xmm4
+    phaddw     xmm5, xmm1
+    psrlw      xmm5, 7
+    packuswb   xmm5, xmm5   // 8 R values
+    movdqu     xmm6, [eax]  // A
+    movdqu     xmm1, [eax + 16]
+    psrld      xmm6, 24
+    psrld      xmm1, 24
+    packuswb   xmm6, xmm1
+    packuswb   xmm6, xmm6   // 8 A values
+    punpcklbw  xmm5, xmm6   // 8 RA values
+    movdqa     xmm1, xmm0   // Weave BG, RA together
+    punpcklwd  xmm0, xmm5   // BGRA first 4
+    punpckhwd  xmm1, xmm5   // BGRA next 4
+    movdqu     [eax], xmm0
+    movdqu     [eax + 16], xmm1
+    lea        eax, [eax + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBSEPIAROW_SSSE3
+
+#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// Same as Sepia except matrix is provided.
+// TODO(fbarchard): packuswbs only use half of the reg. To make RGBA, combine R
+// and B into a high and low, then G/A, unpackl/hbw and then unpckl/hwd.
+__declspec(naked)
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]  /* matrix_argb */
+    movdqu     xmm5, [ecx]
+    pshufd     xmm2, xmm5, 0x00
+    pshufd     xmm3, xmm5, 0x55
+    pshufd     xmm4, xmm5, 0xaa
+    pshufd     xmm5, xmm5, 0xff
+    mov        ecx, [esp + 16]  /* width */
+
+ convertloop:
+    movdqu     xmm0, [eax]  // B
+    movdqu     xmm7, [eax + 16]
+    pmaddubsw  xmm0, xmm2
+    pmaddubsw  xmm7, xmm2
+    movdqu     xmm6, [eax]  // G
+    movdqu     xmm1, [eax + 16]
+    pmaddubsw  xmm6, xmm3
+    pmaddubsw  xmm1, xmm3
+    phaddsw    xmm0, xmm7   // B
+    phaddsw    xmm6, xmm1   // G
+    psraw      xmm0, 6      // B
+    psraw      xmm6, 6      // G
+    packuswb   xmm0, xmm0   // 8 B values
+    packuswb   xmm6, xmm6   // 8 G values
+    punpcklbw  xmm0, xmm6   // 8 BG values
+    movdqu     xmm1, [eax]  // R
+    movdqu     xmm7, [eax + 16]
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm7, xmm4
+    phaddsw    xmm1, xmm7   // R
+    movdqu     xmm6, [eax]  // A
+    movdqu     xmm7, [eax + 16]
+    pmaddubsw  xmm6, xmm5
+    pmaddubsw  xmm7, xmm5
+    phaddsw    xmm6, xmm7   // A
+    psraw      xmm1, 6      // R
+    psraw      xmm6, 6      // A
+    packuswb   xmm1, xmm1   // 8 R values
+    packuswb   xmm6, xmm6   // 8 A values
+    punpcklbw  xmm1, xmm6   // 8 RA values
+    movdqa     xmm6, xmm0   // Weave BG, RA together
+    punpcklwd  xmm0, xmm1   // BGRA first 4
+    punpckhwd  xmm6, xmm1   // BGRA next 4
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm6
+    lea        eax, [eax + 32]
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
+
+#ifdef HAS_ARGBQUANTIZEROW_SSE2
+// Quantize 4 ARGB pixels (16 bytes).
+__declspec(naked)
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  __asm {
+    mov        eax, [esp + 4]    /* dst_argb */
+    movd       xmm2, [esp + 8]   /* scale */
+    movd       xmm3, [esp + 12]  /* interval_size */
+    movd       xmm4, [esp + 16]  /* interval_offset */
+    mov        ecx, [esp + 20]   /* width */
+    pshuflw    xmm2, xmm2, 040h
+    pshufd     xmm2, xmm2, 044h
+    pshuflw    xmm3, xmm3, 040h
+    pshufd     xmm3, xmm3, 044h
+    pshuflw    xmm4, xmm4, 040h
+    pshufd     xmm4, xmm4, 044h
+    pxor       xmm5, xmm5  // constant 0
+    pcmpeqb    xmm6, xmm6  // generate mask 0xff000000
+    pslld      xmm6, 24
+
+ convertloop:
+    movdqu     xmm0, [eax]  // read 4 pixels
+    punpcklbw  xmm0, xmm5   // first 2 pixels
+    pmulhuw    xmm0, xmm2   // pixel * scale >> 16
+    movdqu     xmm1, [eax]  // read 4 pixels
+    punpckhbw  xmm1, xmm5   // next 2 pixels
+    pmulhuw    xmm1, xmm2
+    pmullw     xmm0, xmm3   // * interval_size
+    movdqu     xmm7, [eax]  // read 4 pixels
+    pmullw     xmm1, xmm3
+    pand       xmm7, xmm6   // mask alpha
+    paddw      xmm0, xmm4   // + interval_size / 2
+    paddw      xmm1, xmm4
+    packuswb   xmm0, xmm1
+    por        xmm0, xmm7
+    movdqu     [eax], xmm0
+    lea        eax, [eax + 16]
+    sub        ecx, 4
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBQUANTIZEROW_SSE2
+
+#ifdef HAS_ARGBSHADEROW_SSE2
+// Shade 4 pixels at a time by specified value.
+__declspec(naked)
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    movd       xmm2, [esp + 16]  // value
+    punpcklbw  xmm2, xmm2
+    punpcklqdq xmm2, xmm2
+
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm0       // first 2
+    punpckhbw  xmm1, xmm1       // next 2
+    pmulhuw    xmm0, xmm2       // argb * value
+    pmulhuw    xmm1, xmm2       // argb * value
+    psrlw      xmm0, 8
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBSHADEROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+// Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
+__declspec(naked)
+void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pxor       xmm5, xmm5  // constant 0
+
+ convertloop:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    movdqu     xmm2, [esi]        // read 4 pixels from src_argb1
+    movdqu     xmm1, xmm0
+    movdqu     xmm3, xmm2
+    punpcklbw  xmm0, xmm0         // first 2
+    punpckhbw  xmm1, xmm1         // next 2
+    punpcklbw  xmm2, xmm5         // first 2
+    punpckhbw  xmm3, xmm5         // next 2
+    pmulhuw    xmm0, xmm2         // src_argb0 * src_argb1 first 2
+    pmulhuw    xmm1, xmm3         // src_argb0 * src_argb1 next 2
+    lea        eax, [eax + 16]
+    lea        esi, [esi + 16]
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBMULTIPLYROW_SSE2
+
+#ifdef HAS_ARGBADDROW_SSE2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+// TODO(fbarchard): Port this to posix, neon and other math functions.
+__declspec(naked)
+void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    sub        ecx, 4
+    jl         convertloop49
+
+ convertloop4:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    lea        eax, [eax + 16]
+    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    lea        esi, [esi + 16]
+    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jge        convertloop4
+
+ convertloop49:
+    add        ecx, 4 - 1
+    jl         convertloop19
+
+ convertloop1:
+    movd       xmm0, [eax]        // read 1 pixels from src_argb0
+    lea        eax, [eax + 4]
+    movd       xmm1, [esi]        // read 1 pixels from src_argb1
+    lea        esi, [esi + 4]
+    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jge        convertloop1
+
+ convertloop19:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBADDROW_SSE2
+
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels together, 4 pixels at a time.
+__declspec(naked)
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+ convertloop:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    lea        eax, [eax + 16]
+    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    lea        esi, [esi + 16]
+    psubusb    xmm0, xmm1         // src_argb0 - src_argb1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked)
+void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpxor      ymm5, ymm5, ymm5     // constant 0
+
+ convertloop:
+    vmovdqu    ymm1, [eax]        // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vmovdqu    ymm3, [esi]        // read 8 pixels from src_argb1
+    lea        esi, [esi + 32]
+    vpunpcklbw ymm0, ymm1, ymm1   // low 4
+    vpunpckhbw ymm1, ymm1, ymm1   // high 4
+    vpunpcklbw ymm2, ymm3, ymm5   // low 4
+    vpunpckhbw ymm3, ymm3, ymm5   // high 4
+    vpmulhuw   ymm0, ymm0, ymm2   // src_argb0 * src_argb1 low 4
+    vpmulhuw   ymm1, ymm1, ymm3   // src_argb0 * src_argb1 high 4
+    vpackuswb  ymm0, ymm0, ymm1
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBMULTIPLYROW_AVX2
+
+#ifdef HAS_ARGBADDROW_AVX2
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked)
+void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+ convertloop:
+    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vpaddusb   ymm0, ymm0, [esi]        // add 8 pixels from src_argb1
+    lea        esi, [esi + 32]
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBADDROW_AVX2
+
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+// Subtract 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked)
+void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+ convertloop:
+    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vpsubusb   ymm0, ymm0, [esi]        // src_argb0 - src_argb1
+    lea        esi, [esi + 32]
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_AVX2
+
+#ifdef HAS_SOBELXROW_SSE2
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+__declspec(naked)
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_y0
+    mov        esi, [esp + 8 + 8]   // src_y1
+    mov        edi, [esp + 8 + 12]  // src_y2
+    mov        edx, [esp + 8 + 16]  // dst_sobelx
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        esi, eax
+    sub        edi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    movq       xmm2, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + edi]      // read 8 pixels from src_y2[0]
+    movq       xmm3, qword ptr [eax + edi + 2]  // read 8 pixels from src_y2[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    psubw      xmm1, xmm0
+    pmaxsw     xmm0, xmm1
+    packuswb   xmm0, xmm0
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXROW_SSE2
+
+#ifdef HAS_SOBELYROW_SSE2
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+__declspec(naked)
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_y0
+    mov        esi, [esp + 4 + 8]   // src_y1
+    mov        edx, [esp + 4 + 12]  // dst_sobely
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + 1]        // read 8 pixels from src_y0[1]
+    movq       xmm2, qword ptr [eax + esi + 1]  // read 8 pixels from src_y1[1]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    movq       xmm3, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    psubw      xmm1, xmm0
+    pmaxsw     xmm0, xmm1
+    packuswb   xmm0, xmm0
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELYROW_SSE2
+
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+__declspec(naked)
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+    pslld      xmm5, 24             // 0xff000000
+
+ convertloop:
+    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm2, xmm0             // GG
+    punpcklbw  xmm2, xmm0             // First 8
+    punpckhbw  xmm0, xmm0             // Next 8
+    movdqa     xmm1, xmm2             // GGGG
+    punpcklwd  xmm1, xmm2             // First 4
+    punpckhwd  xmm2, xmm2             // Next 4
+    por        xmm1, xmm5             // GGGA
+    por        xmm2, xmm5
+    movdqa     xmm3, xmm0             // GGGG
+    punpcklwd  xmm3, xmm0             // Next 4
+    punpckhwd  xmm0, xmm0             // Last 4
+    por        xmm3, xmm5             // GGGA
+    por        xmm0, xmm5
+    movdqu     [edx], xmm1
+    movdqu     [edx + 16], xmm2
+    movdqu     [edx + 32], xmm3
+    movdqu     [edx + 48], xmm0
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELROW_SSE2
+
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into a plane.
+__declspec(naked)
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+
+ convertloop:
+    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELTOPLANEROW_SSE2
+
+#ifdef HAS_SOBELXYROW_SSE2
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+__declspec(naked)
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+
+ convertloop:
+    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    movdqa     xmm2, xmm0
+    paddusb    xmm2, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm3, xmm0             // XA
+    punpcklbw  xmm3, xmm5
+    punpckhbw  xmm0, xmm5
+    movdqa     xmm4, xmm1             // YS
+    punpcklbw  xmm4, xmm2
+    punpckhbw  xmm1, xmm2
+    movdqa     xmm6, xmm4             // YSXA
+    punpcklwd  xmm6, xmm3             // First 4
+    punpckhwd  xmm4, xmm3             // Next 4
+    movdqa     xmm7, xmm1             // YSXA
+    punpcklwd  xmm7, xmm0             // Next 4
+    punpckhwd  xmm1, xmm0             // Last 4
+    movdqu     [edx], xmm6
+    movdqu     [edx + 16], xmm4
+    movdqu     [edx + 32], xmm7
+    movdqu     [edx + 48], xmm1
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXYROW_SSE2
+
+#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+// Consider float CumulativeSum.
+// Consider calling CumulativeSum one row at time as needed.
+// Consider circular CumulativeSum buffer of radius * 2 + 1 height.
+// Convert cumulative sum for an area to an average for 1 pixel.
+// topleft is pointer to top left of CumulativeSum buffer for area.
+// botleft is pointer to bottom left of CumulativeSum buffer.
+// width is offset from left to right of area in CumulativeSum buffer measured
+//   in number of ints.
+// area is the number of pixels in the area being averaged.
+// dst points to pixel to store result to.
+// count is number of averaged pixels to produce.
+// Does 4 pixels at a time.
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst,
+                                    int count) {
+  __asm {
+    mov        eax, topleft  // eax topleft
+    mov        esi, botleft  // esi botleft
+    mov        edx, width
+    movd       xmm5, area
+    mov        edi, dst
+    mov        ecx, count
+    cvtdq2ps   xmm5, xmm5
+    rcpss      xmm4, xmm5  // 1.0f / area
+    pshufd     xmm4, xmm4, 0
+    sub        ecx, 4
+    jl         l4b
+
+    cmp        area, 128  // 128 pixels will not overflow 15 bits.
+    ja         l4
+
+    pshufd     xmm5, xmm5, 0        // area
+    pcmpeqb    xmm6, xmm6           // constant of 65536.0 - 1 = 65535.0
+    psrld      xmm6, 16
+    cvtdq2ps   xmm6, xmm6
+    addps      xmm5, xmm6           // (65536.0 + area - 1)
+    mulps      xmm5, xmm4           // (65536.0 + area - 1) * 1 / area
+    cvtps2dq   xmm5, xmm5           // 0.16 fixed point
+    packssdw   xmm5, xmm5           // 16 bit shorts
+
+    // 4 pixel loop small blocks.
+  s4:
+    // top left
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+
+    // - top right
+    psubd      xmm0, [eax + edx * 4]
+    psubd      xmm1, [eax + edx * 4 + 16]
+    psubd      xmm2, [eax + edx * 4 + 32]
+    psubd      xmm3, [eax + edx * 4 + 48]
+    lea        eax, [eax + 64]
+
+    // - bottom left
+    psubd      xmm0, [esi]
+    psubd      xmm1, [esi + 16]
+    psubd      xmm2, [esi + 32]
+    psubd      xmm3, [esi + 48]
+
+    // + bottom right
+    paddd      xmm0, [esi + edx * 4]
+    paddd      xmm1, [esi + edx * 4 + 16]
+    paddd      xmm2, [esi + edx * 4 + 32]
+    paddd      xmm3, [esi + edx * 4 + 48]
+    lea        esi, [esi + 64]
+
+    packssdw   xmm0, xmm1  // pack 4 pixels into 2 registers
+    packssdw   xmm2, xmm3
+
+    pmulhuw    xmm0, xmm5
+    pmulhuw    xmm2, xmm5
+
+    packuswb   xmm0, xmm2
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jge        s4
+
+    jmp        l4b
+
+    // 4 pixel loop
+  l4:
+    // top left
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+
+    // - top right
+    psubd      xmm0, [eax + edx * 4]
+    psubd      xmm1, [eax + edx * 4 + 16]
+    psubd      xmm2, [eax + edx * 4 + 32]
+    psubd      xmm3, [eax + edx * 4 + 48]
+    lea        eax, [eax + 64]
+
+    // - bottom left
+    psubd      xmm0, [esi]
+    psubd      xmm1, [esi + 16]
+    psubd      xmm2, [esi + 32]
+    psubd      xmm3, [esi + 48]
+
+    // + bottom right
+    paddd      xmm0, [esi + edx * 4]
+    paddd      xmm1, [esi + edx * 4 + 16]
+    paddd      xmm2, [esi + edx * 4 + 32]
+    paddd      xmm3, [esi + edx * 4 + 48]
+    lea        esi, [esi + 64]
+
+    cvtdq2ps   xmm0, xmm0   // Average = Sum * 1 / Area
+    cvtdq2ps   xmm1, xmm1
+    mulps      xmm0, xmm4
+    mulps      xmm1, xmm4
+    cvtdq2ps   xmm2, xmm2
+    cvtdq2ps   xmm3, xmm3
+    mulps      xmm2, xmm4
+    mulps      xmm3, xmm4
+    cvtps2dq   xmm0, xmm0
+    cvtps2dq   xmm1, xmm1
+    cvtps2dq   xmm2, xmm2
+    cvtps2dq   xmm3, xmm3
+    packssdw   xmm0, xmm1
+    packssdw   xmm2, xmm3
+    packuswb   xmm0, xmm2
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+  l1:
+    movdqu     xmm0, [eax]
+    psubd      xmm0, [eax + edx * 4]
+    lea        eax, [eax + 16]
+    psubd      xmm0, [esi]
+    paddd      xmm0, [esi + edx * 4]
+    lea        esi, [esi + 16]
+    cvtdq2ps   xmm0, xmm0
+    mulps      xmm0, xmm4
+    cvtps2dq   xmm0, xmm0
+    packssdw   xmm0, xmm0
+    packuswb   xmm0, xmm0
+    movd       dword ptr [edi], xmm0
+    lea        edi, [edi + 4]
+    sub        ecx, 1
+    jge        l1
+  l1b:
+  }
+}
+#endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+
+#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
+// Creates a table of cumulative sums where each value is a sum of all values
+// above and to the left of the value.
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width) {
+  __asm {
+    mov        eax, row
+    mov        edx, cumsum
+    mov        esi, previous_cumsum
+    mov        ecx, width
+    pxor       xmm0, xmm0
+    pxor       xmm1, xmm1
+
+    sub        ecx, 4
+    jl         l4b
+    test       edx, 15
+    jne        l4b
+
+    // 4 pixel loop
+  l4:
+    movdqu     xmm2, [eax]  // 4 argb pixels 16 bytes.
+    lea        eax, [eax + 16]
+    movdqa     xmm4, xmm2
+
+    punpcklbw  xmm2, xmm1
+    movdqa     xmm3, xmm2
+    punpcklwd  xmm2, xmm1
+    punpckhwd  xmm3, xmm1
+
+    punpckhbw  xmm4, xmm1
+    movdqa     xmm5, xmm4
+    punpcklwd  xmm4, xmm1
+    punpckhwd  xmm5, xmm1
+
+    paddd      xmm0, xmm2
+    movdqu     xmm2, [esi]  // previous row above.
+    paddd      xmm2, xmm0
+
+    paddd      xmm0, xmm3
+    movdqu     xmm3, [esi + 16]
+    paddd      xmm3, xmm0
+
+    paddd      xmm0, xmm4
+    movdqu     xmm4, [esi + 32]
+    paddd      xmm4, xmm0
+
+    paddd      xmm0, xmm5
+    movdqu     xmm5, [esi + 48]
+    lea        esi, [esi + 64]
+    paddd      xmm5, xmm0
+
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm3
+    movdqu     [edx + 32], xmm4
+    movdqu     [edx + 48], xmm5
+
+    lea        edx, [edx + 64]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+  l1:
+    movd       xmm2, dword ptr [eax]  // 1 argb pixel 4 bytes.
+    lea        eax, [eax + 4]
+    punpcklbw  xmm2, xmm1
+    punpcklwd  xmm2, xmm1
+    paddd      xmm0, xmm2
+    movdqu     xmm2, [esi]
+    lea        esi, [esi + 16]
+    paddd      xmm2, xmm0
+    movdqu     [edx], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 1
+    jge        l1
+
+ l1b:
+  }
+}
+#endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
+
+#ifdef HAS_ARGBAFFINEROW_SSE2
+// Copy ARGB pixels from source image with slope to a row of destination.
+__declspec(naked)
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 12]  // src_argb
+    mov        esi, [esp + 16]  // stride
+    mov        edx, [esp + 20]  // dst_argb
+    mov        ecx, [esp + 24]  // pointer to uv_dudv
+    movq       xmm2, qword ptr [ecx]  // uv
+    movq       xmm7, qword ptr [ecx + 8]  // dudv
+    mov        ecx, [esp + 28]  // width
+    shl        esi, 16          // 4, stride
+    add        esi, 4
+    movd       xmm5, esi
+    sub        ecx, 4
+    jl         l4b
+
+    // setup for 4 pixel loop
+    pshufd     xmm7, xmm7, 0x44  // dup dudv
+    pshufd     xmm5, xmm5, 0  // dup 4, stride
+    movdqa     xmm0, xmm2    // x0, y0, x1, y1
+    addps      xmm0, xmm7
+    movlhps    xmm2, xmm0
+    movdqa     xmm4, xmm7
+    addps      xmm4, xmm4    // dudv *= 2
+    movdqa     xmm3, xmm2    // x2, y2, x3, y3
+    addps      xmm3, xmm4
+    addps      xmm4, xmm4    // dudv *= 4
+
+    // 4 pixel loop
+  l4:
+    cvttps2dq  xmm0, xmm2    // x, y float to int first 2
+    cvttps2dq  xmm1, xmm3    // x, y float to int next 2
+    packssdw   xmm0, xmm1    // x, y as 8 shorts
+    pmaddwd    xmm0, xmm5    // offsets = x * 4 + y * stride.
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       edi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       xmm1, [eax + esi]  // read pixel 0
+    movd       xmm6, [eax + edi]  // read pixel 1
+    punpckldq  xmm1, xmm6     // combine pixel 0 and 1
+    addps      xmm2, xmm4    // x, y += dx, dy first 2
+    movq       qword ptr [edx], xmm1
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       edi, xmm0
+    movd       xmm6, [eax + esi]  // read pixel 2
+    movd       xmm0, [eax + edi]  // read pixel 3
+    punpckldq  xmm6, xmm0     // combine pixel 2 and 3
+    addps      xmm3, xmm4    // x, y += dx, dy next 2
+    movq       qword ptr 8[edx], xmm6
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+  l1:
+    cvttps2dq  xmm0, xmm2    // x, y float to int
+    packssdw   xmm0, xmm0    // x, y as shorts
+    pmaddwd    xmm0, xmm5    // offset = x * 4 + y * stride
+    addps      xmm2, xmm7    // x, y += dx, dy
+    movd       esi, xmm0
+    movd       xmm0, [eax + esi]  // copy a pixel
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jge        l1
+  l1b:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBAFFINEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+// Bilinear filter 32x2 -> 32x1
+__declspec(naked)
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    sub        edi, esi
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    vmovd      xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    vmovd      xmm5, eax  // low fraction 128..1
+    vpunpcklbw xmm5, xmm5, xmm0
+    vpunpcklwd xmm5, xmm5, xmm5
+    vpxor      ymm0, ymm0, ymm0
+    vpermd     ymm5, ymm0, ymm5
+
+  xloop:
+    vmovdqu    ymm0, [esi]
+    vmovdqu    ymm2, [esi + edx]
+    vpunpckhbw ymm1, ymm0, ymm2  // mutates
+    vpunpcklbw ymm0, ymm0, ymm2  // mutates
+    vpmaddubsw ymm0, ymm0, ymm5
+    vpmaddubsw ymm1, ymm1, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm1, ymm1, 7
+    vpackuswb  ymm0, ymm0, ymm1  // unmutates
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    sub        ecx, 32
+    jg         xloop
+    jmp        xloop99
+
+   // Blend 25 / 75.
+ xloop25:
+   vmovdqu    ymm0, [esi]
+   vmovdqu    ymm1, [esi + edx]
+   vpavgb     ymm0, ymm0, ymm1
+   vpavgb     ymm0, ymm0, ymm1
+   vmovdqu    [esi + edi], ymm0
+   lea        esi, [esi + 32]
+   sub        ecx, 32
+   jg         xloop25
+   jmp        xloop99
+
+   // Blend 50 / 50.
+ xloop50:
+   vmovdqu    ymm0, [esi]
+   vpavgb     ymm0, ymm0, [esi + edx]
+   vmovdqu    [esi + edi], ymm0
+   lea        esi, [esi + 32]
+   sub        ecx, 32
+   jg         xloop50
+   jmp        xloop99
+
+   // Blend 75 / 25.
+ xloop75:
+   vmovdqu    ymm1, [esi]
+   vmovdqu    ymm0, [esi + edx]
+   vpavgb     ymm0, ymm0, ymm1
+   vpavgb     ymm0, ymm0, ymm1
+   vmovdqu    [esi + edi], ymm0
+   lea        esi, [esi + 32]
+   sub        ecx, 32
+   jg         xloop75
+   jmp        xloop99
+
+   // Blend 100 / 0 - Copy row unchanged.
+ xloop100:
+   rep movsb
+
+  xloop99:
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_AVX2
+
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked)
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    movd       xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    movd       xmm5, eax  // low fraction 128..1
+    punpcklbw  xmm5, xmm0
+    punpcklwd  xmm5, xmm5
+    pshufd     xmm5, xmm5, 0
+
+  xloop:
+    movdqu     xmm0, [esi]
+    movdqu     xmm2, [esi + edx]
+    movdqu     xmm1, xmm0
+    punpcklbw  xmm0, xmm2
+    punpckhbw  xmm1, xmm2
+    pmaddubsw  xmm0, xmm5
+    pmaddubsw  xmm1, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm1, 7
+    packuswb   xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+  xloop25:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+  xloop50:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+  xloop75:
+    movdqu     xmm1, [esi]
+    movdqu     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+  xloop100:
+    movdqu     xmm0, [esi]
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked)
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 256.  Blend 100 / 0.
+    cmp        eax, 64
+    je         xloop75   // 64 / 256 is 0.25.  Blend 75 / 25.
+    cmp        eax, 128
+    je         xloop50   // 128 / 256 is 0.50.  Blend 50 / 50.
+    cmp        eax, 192
+    je         xloop25   // 192 / 256 is 0.75.  Blend 25 / 75.
+
+    movd       xmm5, eax            // xmm5 = y fraction
+    punpcklbw  xmm5, xmm5
+    psrlw      xmm5, 1
+    punpcklwd  xmm5, xmm5
+    punpckldq  xmm5, xmm5
+    punpcklqdq xmm5, xmm5
+    pxor       xmm4, xmm4
+
+  xloop:
+    movdqu     xmm0, [esi]  // row0
+    movdqu     xmm2, [esi + edx]  // row1
+    movdqu     xmm1, xmm0
+    movdqu     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    psubw      xmm2, xmm0  // row1 - row0
+    psubw      xmm3, xmm1
+    paddw      xmm2, xmm2  // 9 bits * 15 bits = 8.16
+    paddw      xmm3, xmm3
+    pmulhw     xmm2, xmm5  // scale diff
+    pmulhw     xmm3, xmm5
+    paddw      xmm0, xmm2  // sum rows
+    paddw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+  xloop25:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+  xloop50:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+  xloop75:
+    movdqu     xmm1, [esi]
+    movdqu     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+  xloop100:
+    movdqu     xmm0, [esi]
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+__declspec(naked)
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_argb
+    mov        ecx, [esp + 12]   // shuffler
+    movdqu     xmm5, [ecx]
+    mov        ecx, [esp + 16]   // pix
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         wloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+__declspec(naked)
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]     // src_argb
+    mov        edx, [esp + 8]     // dst_argb
+    mov        ecx, [esp + 12]    // shuffler
+    vbroadcastf128 ymm5, [ecx]    // same shuffle in high as low.
+    mov        ecx, [esp + 16]    // pix
+
+  wloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax, [eax + 64]
+    vpshufb    ymm0, ymm0, ymm5
+    vpshufb    ymm1, ymm1, ymm5
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         wloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBSHUFFLEROW_AVX2
+
+__declspec(naked)
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  __asm {
+    push       ebx
+    push       esi
+    mov        eax, [esp + 8 + 4]    // src_argb
+    mov        edx, [esp + 8 + 8]    // dst_argb
+    mov        esi, [esp + 8 + 12]   // shuffler
+    mov        ecx, [esp + 8 + 16]   // pix
+    pxor       xmm5, xmm5
+
+    mov        ebx, [esi]   // shuffler
+    cmp        ebx, 0x03000102
+    je         shuf_3012
+    cmp        ebx, 0x00010203
+    je         shuf_0123
+    cmp        ebx, 0x00030201
+    je         shuf_0321
+    cmp        ebx, 0x02010003
+    je         shuf_2103
+
+  // TODO(fbarchard): Use one source pointer and 3 offsets.
+  shuf_any1:
+    movzx      ebx, byte ptr [esi]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx], bl
+    movzx      ebx, byte ptr [esi + 1]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 1], bl
+    movzx      ebx, byte ptr [esi + 2]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 2], bl
+    movzx      ebx, byte ptr [esi + 3]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 3], bl
+    lea        eax, [eax + 4]
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jg         shuf_any1
+    jmp        shuf99
+
+  shuf_0123:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 01Bh   // 1B = 00011011 = 0x0123 = BGRAToARGB
+    pshuflw    xmm0, xmm0, 01Bh
+    pshufhw    xmm1, xmm1, 01Bh
+    pshuflw    xmm1, xmm1, 01Bh
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         shuf_0123
+    jmp        shuf99
+
+  shuf_0321:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 039h   // 39 = 00111001 = 0x0321 = RGBAToARGB
+    pshuflw    xmm0, xmm0, 039h
+    pshufhw    xmm1, xmm1, 039h
+    pshuflw    xmm1, xmm1, 039h
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         shuf_0321
+    jmp        shuf99
+
+  shuf_2103:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 093h   // 93 = 10010011 = 0x2103 = ARGBToRGBA
+    pshuflw    xmm0, xmm0, 093h
+    pshufhw    xmm1, xmm1, 093h
+    pshuflw    xmm1, xmm1, 093h
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         shuf_2103
+    jmp        shuf99
+
+  shuf_3012:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 0C6h   // C6 = 11000110 = 0x3012 = ABGRToARGB
+    pshuflw    xmm0, xmm0, 0C6h
+    pshufhw    xmm1, xmm1, 0C6h
+    pshuflw    xmm1, xmm1, 0C6h
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         shuf_3012
+
+  shuf99:
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// YUY2 - Macro-pixel = 2 image pixels
+// Y0U0Y1V0....Y2U2Y3V2...Y4U4Y5V4....
+
+// UYVY - Macro-pixel = 2 image pixels
+// U0Y0V0Y1
+
+__declspec(naked)
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_y
+    mov        esi, [esp + 8 + 8]    // src_u
+    mov        edx, [esp + 8 + 12]   // src_v
+    mov        edi, [esp + 8 + 16]   // dst_frame
+    mov        ecx, [esp + 8 + 20]   // width
+    sub        edx, esi
+
+  convertloop:
+    movq       xmm2, qword ptr [esi] // U
+    movq       xmm3, qword ptr [esi + edx] // V
+    lea        esi, [esi + 8]
+    punpcklbw  xmm2, xmm3 // UV
+    movdqu     xmm0, [eax] // Y
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm2 // YUYV
+    punpckhbw  xmm1, xmm2
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm1
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_y
+    mov        esi, [esp + 8 + 8]    // src_u
+    mov        edx, [esp + 8 + 12]   // src_v
+    mov        edi, [esp + 8 + 16]   // dst_frame
+    mov        ecx, [esp + 8 + 20]   // width
+    sub        edx, esi
+
+  convertloop:
+    movq       xmm2, qword ptr [esi] // U
+    movq       xmm3, qword ptr [esi + edx] // V
+    lea        esi, [esi + 8]
+    punpcklbw  xmm2, xmm3 // UV
+    movdqu     xmm0, [eax] // Y
+    movdqa     xmm1, xmm2
+    lea        eax, [eax + 16]
+    punpcklbw  xmm1, xmm0 // UYVY
+    punpckhbw  xmm2, xmm0
+    movdqu     [edi], xmm1
+    movdqu     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+__declspec(naked)
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* src_argb */
+    mov        edx, [esp + 4 + 8]   /* dst_argb */
+    mov        esi, [esp + 4 + 12]  /* poly */
+    mov        ecx, [esp + 4 + 16]  /* width */
+    pxor       xmm3, xmm3  // 0 constant for zero extending bytes to ints.
+
+    // 2 pixel loop.
+ convertloop:
+//    pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
+//    pmovzxbd  xmm4, dword ptr [eax + 4]  // BGRA pixel
+    movq       xmm0, qword ptr [eax]  // BGRABGRA
+    lea        eax, [eax + 8]
+    punpcklbw  xmm0, xmm3
+    movdqa     xmm4, xmm0
+    punpcklwd  xmm0, xmm3  // pixel 0
+    punpckhwd  xmm4, xmm3  // pixel 1
+    cvtdq2ps   xmm0, xmm0  // 4 floats
+    cvtdq2ps   xmm4, xmm4
+    movdqa     xmm1, xmm0  // X
+    movdqa     xmm5, xmm4
+    mulps      xmm0, [esi + 16]  // C1 * X
+    mulps      xmm4, [esi + 16]
+    addps      xmm0, [esi]  // result = C0 + C1 * X
+    addps      xmm4, [esi]
+    movdqa     xmm2, xmm1
+    movdqa     xmm6, xmm5
+    mulps      xmm2, xmm1  // X * X
+    mulps      xmm6, xmm5
+    mulps      xmm1, xmm2  // X * X * X
+    mulps      xmm5, xmm6
+    mulps      xmm2, [esi + 32]  // C2 * X * X
+    mulps      xmm6, [esi + 32]
+    mulps      xmm1, [esi + 48]  // C3 * X * X * X
+    mulps      xmm5, [esi + 48]
+    addps      xmm0, xmm2  // result += C2 * X * X
+    addps      xmm4, xmm6
+    addps      xmm0, xmm1  // result += C3 * X * X * X
+    addps      xmm4, xmm5
+    cvttps2dq  xmm0, xmm0
+    cvttps2dq  xmm4, xmm4
+    packuswb   xmm0, xmm4
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    sub        ecx, 2
+    jg         convertloop
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+__declspec(naked)
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]   /* poly */
+    vbroadcastf128 ymm4, [ecx]       // C0
+    vbroadcastf128 ymm5, [ecx + 16]  // C1
+    vbroadcastf128 ymm6, [ecx + 32]  // C2
+    vbroadcastf128 ymm7, [ecx + 48]  // C3
+    mov        ecx, [esp + 16]  /* width */
+
+    // 2 pixel loop.
+ convertloop:
+    vpmovzxbd   ymm0, qword ptr [eax]  // 2 BGRA pixels
+    lea         eax, [eax + 8]
+    vcvtdq2ps   ymm0, ymm0        // X 8 floats
+    vmulps      ymm2, ymm0, ymm0  // X * X
+    vmulps      ymm3, ymm0, ymm7  // C3 * X
+    vfmadd132ps ymm0, ymm4, ymm5  // result = C0 + C1 * X
+    vfmadd231ps ymm0, ymm2, ymm6  // result += C2 * X * X
+    vfmadd231ps ymm0, ymm2, ymm3  // result += C3 * X * X * X
+    vcvttps2dq  ymm0, ymm0
+    vpackusdw   ymm0, ymm0, ymm0  // b0g0r0a0_00000000_b0g0r0a0_00000000
+    vpermq      ymm0, ymm0, 0xd8  // b0g0r0a0_b0g0r0a0_00000000_00000000
+    vpackuswb   xmm0, xmm0, xmm0  // bgrabgra_00000000_00000000_00000000
+    vmovq       qword ptr [edx], xmm0
+    lea         edx, [edx + 8]
+    sub         ecx, 2
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
+#ifdef HAS_ARGBCOLORTABLEROW_X86
+// Tranform ARGB pixels with color table.
+__declspec(naked)
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+                           int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* dst_argb */
+    mov        esi, [esp + 4 + 8]   /* table_argb */
+    mov        ecx, [esp + 4 + 12]  /* width */
+
+    // 1 pixel loop.
+  convertloop:
+    movzx      edx, byte ptr [eax]
+    lea        eax, [eax + 4]
+    movzx      edx, byte ptr [esi + edx * 4]
+    mov        byte ptr [eax - 4], dl
+    movzx      edx, byte ptr [eax - 4 + 1]
+    movzx      edx, byte ptr [esi + edx * 4 + 1]
+    mov        byte ptr [eax - 4 + 1], dl
+    movzx      edx, byte ptr [eax - 4 + 2]
+    movzx      edx, byte ptr [esi + edx * 4 + 2]
+    mov        byte ptr [eax - 4 + 2], dl
+    movzx      edx, byte ptr [eax - 4 + 3]
+    movzx      edx, byte ptr [esi + edx * 4 + 3]
+    mov        byte ptr [eax - 4 + 3], dl
+    dec        ecx
+    jg         convertloop
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBCOLORTABLEROW_X86
+
+#ifdef HAS_RGBCOLORTABLEROW_X86
+// Tranform RGB pixels with color table.
+__declspec(naked)
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* dst_argb */
+    mov        esi, [esp + 4 + 8]   /* table_argb */
+    mov        ecx, [esp + 4 + 12]  /* width */
+
+    // 1 pixel loop.
+  convertloop:
+    movzx      edx, byte ptr [eax]
+    lea        eax, [eax + 4]
+    movzx      edx, byte ptr [esi + edx * 4]
+    mov        byte ptr [eax - 4], dl
+    movzx      edx, byte ptr [eax - 4 + 1]
+    movzx      edx, byte ptr [esi + edx * 4 + 1]
+    mov        byte ptr [eax - 4 + 1], dl
+    movzx      edx, byte ptr [eax - 4 + 2]
+    movzx      edx, byte ptr [esi + edx * 4 + 2]
+    mov        byte ptr [eax - 4 + 2], dl
+    dec        ecx
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_RGBCOLORTABLEROW_X86
+
+#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// Tranform RGB pixels with luma table.
+__declspec(naked)
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   /* src_argb */
+    mov        edi, [esp + 8 + 8]   /* dst_argb */
+    mov        ecx, [esp + 8 + 12]  /* width */
+    movd       xmm2, dword ptr [esp + 8 + 16]  // luma table
+    movd       xmm3, dword ptr [esp + 8 + 20]  // lumacoeff
+    pshufd     xmm2, xmm2, 0
+    pshufd     xmm3, xmm3, 0
+    pcmpeqb    xmm4, xmm4        // generate mask 0xff00ff00
+    psllw      xmm4, 8
+    pxor       xmm5, xmm5
+
+    // 4 pixel loop.
+  convertloop:
+    movdqu     xmm0, qword ptr [eax]      // generate luma ptr
+    pmaddubsw  xmm0, xmm3
+    phaddw     xmm0, xmm0
+    pand       xmm0, xmm4  // mask out low bits
+    punpcklwd  xmm0, xmm5
+    paddd      xmm0, xmm2  // add table base
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi], dl
+    movzx      edx, byte ptr [eax + 1]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 1], dl
+    movzx      edx, byte ptr [eax + 2]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 2], dl
+    movzx      edx, byte ptr [eax + 3]  // copy alpha.
+    mov        byte ptr [edi + 3], dl
+
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax + 4]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 4], dl
+    movzx      edx, byte ptr [eax + 5]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 5], dl
+    movzx      edx, byte ptr [eax + 6]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 6], dl
+    movzx      edx, byte ptr [eax + 7]  // copy alpha.
+    mov        byte ptr [edi + 7], dl
+
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax + 8]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 8], dl
+    movzx      edx, byte ptr [eax + 9]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 9], dl
+    movzx      edx, byte ptr [eax + 10]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 10], dl
+    movzx      edx, byte ptr [eax + 11]  // copy alpha.
+    mov        byte ptr [edi + 11], dl
+
+    movd       esi, xmm0
+
+    movzx      edx, byte ptr [eax + 12]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 12], dl
+    movzx      edx, byte ptr [eax + 13]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 13], dl
+    movzx      edx, byte ptr [eax + 14]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 14], dl
+    movzx      edx, byte ptr [eax + 15]  // copy alpha.
+    mov        byte ptr [edi + 15], dl
+
+    lea        eax, [eax + 16]
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
+
+#endif  // defined(_M_X64)
+#endif  // !defined(LIBYUV_DISABLE_X86) && (defined(_M_IX86) || defined(_M_X64))
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/row_x86.asm b/libvpx/libvpx/third_party/libyuv/source/row_x86.asm
new file mode 100644
index 0000000..0cb326f
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/row_x86.asm
@@ -0,0 +1,146 @@
+;
+; Copyright 2012 The LibYuv Project Authors. All rights reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+%ifdef __YASM_VERSION_ID__
+%if __YASM_VERSION_ID__ < 01020000h
+%error AVX2 is supported only by yasm 1.2.0 or later.
+%endif
+%endif
+%include "x86inc.asm"
+
+SECTION .text
+
+; cglobal numeric constants are parameters, gpr regs, mm regs
+
+; void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix)
+
+%macro YUY2TOYROW 2-3
+cglobal %1ToYRow%3, 3, 3, 3, src_yuy2, dst_y, pix
+%ifidn %1,YUY2
+    pcmpeqb    m2, m2, m2        ; generate mask 0x00ff00ff
+    psrlw      m2, m2, 8
+%endif
+
+    ALIGN      4
+.convertloop:
+    mov%2      m0, [src_yuy2q]
+    mov%2      m1, [src_yuy2q + mmsize]
+    lea        src_yuy2q, [src_yuy2q + mmsize * 2]
+%ifidn %1,YUY2
+    pand       m0, m0, m2   ; YUY2 even bytes are Y
+    pand       m1, m1, m2
+%else
+    psrlw      m0, m0, 8    ; UYVY odd bytes are Y
+    psrlw      m1, m1, 8
+%endif
+    packuswb   m0, m0, m1
+%if cpuflag(AVX2)
+    vpermq     m0, m0, 0xd8
+%endif
+    sub        pixd, mmsize
+    mov%2      [dst_yq], m0
+    lea        dst_yq, [dst_yq + mmsize]
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+; TODO(fbarchard): Remove MMX.  Add SSSE3 pshufb version.
+INIT_MMX MMX
+YUY2TOYROW YUY2,a,
+YUY2TOYROW YUY2,u,_Unaligned
+YUY2TOYROW UYVY,a,
+YUY2TOYROW UYVY,u,_Unaligned
+INIT_XMM SSE2
+YUY2TOYROW YUY2,a,
+YUY2TOYROW YUY2,u,_Unaligned
+YUY2TOYROW UYVY,a,
+YUY2TOYROW UYVY,u,_Unaligned
+INIT_YMM AVX2
+YUY2TOYROW YUY2,a,
+YUY2TOYROW UYVY,a,
+
+; void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix)
+
+%macro SplitUVRow 1-2
+cglobal SplitUVRow%2, 4, 4, 5, src_uv, dst_u, dst_v, pix
+    pcmpeqb    m4, m4, m4        ; generate mask 0x00ff00ff
+    psrlw      m4, m4, 8
+    sub        dst_vq, dst_uq
+
+    ALIGN      4
+.convertloop:
+    mov%1      m0, [src_uvq]
+    mov%1      m1, [src_uvq + mmsize]
+    lea        src_uvq, [src_uvq + mmsize * 2]
+    psrlw      m2, m0, 8         ; odd bytes
+    psrlw      m3, m1, 8
+    pand       m0, m0, m4        ; even bytes
+    pand       m1, m1, m4
+    packuswb   m0, m0, m1
+    packuswb   m2, m2, m3
+%if cpuflag(AVX2)
+    vpermq     m0, m0, 0xd8
+    vpermq     m2, m2, 0xd8
+%endif
+    mov%1      [dst_uq], m0
+    mov%1      [dst_uq + dst_vq], m2
+    lea        dst_uq, [dst_uq + mmsize]
+    sub        pixd, mmsize
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+INIT_MMX MMX
+SplitUVRow a,
+SplitUVRow u,_Unaligned
+INIT_XMM SSE2
+SplitUVRow a,
+SplitUVRow u,_Unaligned
+INIT_YMM AVX2
+SplitUVRow a,
+
+; void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+;                      int width);
+
+%macro MergeUVRow_ 1-2
+cglobal MergeUVRow_%2, 4, 4, 3, src_u, src_v, dst_uv, pix
+    sub        src_vq, src_uq
+
+    ALIGN      4
+.convertloop:
+    mov%1      m0, [src_uq]
+    mov%1      m1, [src_vq]
+    lea        src_uq, [src_uq + mmsize]
+    punpcklbw  m2, m0, m1       // first 8 UV pairs
+    punpckhbw  m0, m0, m1       // next 8 UV pairs
+%if cpuflag(AVX2)
+    vperm2i128 m1, m2, m0, 0x20  // low 128 of ymm2 and low 128 of ymm0
+    vperm2i128 m2, m2, m0, 0x31  // high 128 of ymm2 and high 128 of ymm0
+    mov%1      [dst_uvq], m1
+    mov%1      [dst_uvq + mmsize], m2
+%else
+    mov%1      [dst_uvq], m2
+    mov%1      [dst_uvq + mmsize], m0
+%endif
+    lea        dst_uvq, [dst_uvq + mmsize * 2]
+    sub        pixd, mmsize
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+INIT_MMX MMX
+MergeUVRow_ a,
+MergeUVRow_ u,_Unaligned
+INIT_XMM SSE2
+MergeUVRow_ a,
+MergeUVRow_ u,_Unaligned
+INIT_YMM AVX2
+MergeUVRow_ a,
+
diff --git a/libvpx/libvpx/third_party/libyuv/source/scale.cc b/libvpx/libvpx/third_party/libyuv/source/scale.cc
new file mode 100644
index 0000000..0a01304
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/scale.cc
@@ -0,0 +1,1689 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyPlane
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+
+// Scale plane, 1/2
+// This is an optimized version for scaling down a plane to 1/2 of
+// its original size.
+
+static void ScalePlaneDown2(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) =
+      filtering == kFilterNone ? ScaleRowDown2_C :
+      (filtering == kFilterLinear ? ScaleRowDown2Linear_C : ScaleRowDown2Box_C);
+  int row_stride = src_stride << 1;
+  if (!filtering) {
+    src_ptr += src_stride;  // Point to odd rows.
+    src_stride = 0;
+  }
+
+#if defined(HAS_SCALEROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_NEON :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_NEON :
+        ScaleRowDown2Box_Any_NEON);
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_NEON :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_NEON :
+          ScaleRowDown2Box_NEON);
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_SSE2 :
+        ScaleRowDown2Box_Any_SSE2);
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 :
+          ScaleRowDown2Box_SSE2);
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_AVX2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_AVX2 :
+        ScaleRowDown2Box_Any_AVX2);
+    if (IS_ALIGNED(dst_width, 32)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_AVX2 :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_AVX2 :
+          ScaleRowDown2Box_AVX2);
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
+      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown2 = filtering ?
+        ScaleRowDown2Box_MIPS_DSPR2 : ScaleRowDown2_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  // TODO(fbarchard): Loop through source height to allow odd height.
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void ScalePlaneDown2_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown2)(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst_ptr, int dst_width) =
+    filtering == kFilterNone ? ScaleRowDown2_16_C :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_16_C :
+        ScaleRowDown2Box_16_C);
+  int row_stride = src_stride << 1;
+  if (!filtering) {
+    src_ptr += src_stride;  // Point to odd rows.
+    src_stride = 0;
+  }
+
+#if defined(HAS_SCALEROWDOWN2_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering ? ScaleRowDown2Box_16_NEON :
+        ScaleRowDown2_16_NEON;
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_16_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_16_SSE2 :
+        ScaleRowDown2Box_16_SSE2);
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
+      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown2 = filtering ?
+        ScaleRowDown2Box_16_MIPS_DSPR2 : ScaleRowDown2_16_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  // TODO(fbarchard): Loop through source height to allow odd height.
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+// Scale plane, 1/4
+// This is an optimized version for scaling down a plane to 1/4 of
+// its original size.
+
+static void ScalePlaneDown4(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown4)(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) =
+      filtering ? ScaleRowDown4Box_C : ScaleRowDown4_C;
+  int row_stride = src_stride << 2;
+  if (!filtering) {
+    src_ptr += src_stride * 2;  // Point to row 2.
+    src_stride = 0;
+  }
+#if defined(HAS_SCALEROWDOWN4_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_Any_NEON : ScaleRowDown4_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleRowDown4 = filtering ? ScaleRowDown4Box_NEON : ScaleRowDown4_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_Any_SSE2 : ScaleRowDown4_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleRowDown4 = filtering ? ScaleRowDown4Box_SSE2 : ScaleRowDown4_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_Any_AVX2 : ScaleRowDown4_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown4 = filtering ? ScaleRowDown4Box_AVX2 : ScaleRowDown4_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_MIPS_DSPR2 : ScaleRowDown4_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void ScalePlaneDown4_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown4)(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst_ptr, int dst_width) =
+      filtering ? ScaleRowDown4Box_16_C : ScaleRowDown4_16_C;
+  int row_stride = src_stride << 2;
+  if (!filtering) {
+    src_ptr += src_stride * 2;  // Point to row 2.
+    src_stride = 0;
+  }
+#if defined(HAS_SCALEROWDOWN4_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_NEON :
+        ScaleRowDown4_16_NEON;
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_SSE2 :
+        ScaleRowDown4_16_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_16_MIPS_DSPR2 : ScaleRowDown4_16_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+// Scale plane down, 3/4
+
+static void ScalePlaneDown34(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr,
+                             enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown34_0)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown34_0 = ScaleRowDown34_C;
+    ScaleRowDown34_1 = ScaleRowDown34_C;
+  } else {
+    ScaleRowDown34_0 = ScaleRowDown34_0_Box_C;
+    ScaleRowDown34_1 = ScaleRowDown34_1_Box_C;
+  }
+#if defined(HAS_SCALEROWDOWN34_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_Any_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_Any_NEON;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_NEON;
+    }
+    if (dst_width % 24 == 0) {
+      if (!filtering) {
+        ScaleRowDown34_0 = ScaleRowDown34_NEON;
+        ScaleRowDown34_1 = ScaleRowDown34_NEON;
+      } else {
+        ScaleRowDown34_0 = ScaleRowDown34_0_Box_NEON;
+        ScaleRowDown34_1 = ScaleRowDown34_1_Box_NEON;
+      }
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_Any_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_Any_SSSE3;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_SSSE3;
+    }
+    if (dst_width % 24 == 0) {
+      if (!filtering) {
+        ScaleRowDown34_0 = ScaleRowDown34_SSSE3;
+        ScaleRowDown34_1 = ScaleRowDown34_SSSE3;
+      } else {
+        ScaleRowDown34_0 = ScaleRowDown34_0_Box_SSSE3;
+        ScaleRowDown34_1 = ScaleRowDown34_1_Box_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_MIPS_DSPR2;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
+                     dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static void ScalePlaneDown34_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr,
+                                enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown34_0)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown34_0 = ScaleRowDown34_16_C;
+    ScaleRowDown34_1 = ScaleRowDown34_16_C;
+  } else {
+    ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_C;
+    ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_C;
+  }
+#if defined(HAS_SCALEROWDOWN34_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_16_NEON;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_16_SSSE3;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_16_MIPS_DSPR2;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
+                     dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+
+// Scale plane, 3/8
+// This is an optimized version for scaling down a plane to 3/8
+// of its original size.
+//
+// Uses box filter arranges like this
+// aaabbbcc -> abc
+// aaabbbcc    def
+// aaabbbcc    ghi
+// dddeeeff
+// dddeeeff
+// dddeeeff
+// ggghhhii
+// ggghhhii
+// Boxes are 3x3, 2x3, 3x2 and 2x2
+
+static void ScalePlaneDown38(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr,
+                             enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown38_3)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown38_3 = ScaleRowDown38_C;
+    ScaleRowDown38_2 = ScaleRowDown38_C;
+  } else {
+    ScaleRowDown38_3 = ScaleRowDown38_3_Box_C;
+    ScaleRowDown38_2 = ScaleRowDown38_2_Box_C;
+  }
+
+#if defined(HAS_SCALEROWDOWN38_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_Any_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_Any_NEON;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_NEON;
+    }
+    if (dst_width % 12 == 0) {
+      if (!filtering) {
+        ScaleRowDown38_3 = ScaleRowDown38_NEON;
+        ScaleRowDown38_2 = ScaleRowDown38_NEON;
+      } else {
+        ScaleRowDown38_3 = ScaleRowDown38_3_Box_NEON;
+        ScaleRowDown38_2 = ScaleRowDown38_2_Box_NEON;
+      }
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN38_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_Any_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_Any_SSSE3;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_SSSE3;
+    }
+    if (dst_width % 12 == 0 && !filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_SSSE3;
+    }
+    if (dst_width % 6 == 0 && filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN38_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_MIPS_DSPR2;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static void ScalePlaneDown38_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr,
+                                enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown38_3)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown38_3 = ScaleRowDown38_16_C;
+    ScaleRowDown38_2 = ScaleRowDown38_16_C;
+  } else {
+    ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_C;
+    ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_C;
+  }
+#if defined(HAS_SCALEROWDOWN38_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 12 == 0)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_16_NEON;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN38_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_16_SSSE3;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN38_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_16_MIPS_DSPR2;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+#define MIN1(x) ((x) < 1 ? 1 : (x))
+
+static __inline uint32 SumPixels(int iboxwidth, const uint16* src_ptr) {
+  uint32 sum = 0u;
+  int x;
+  assert(iboxwidth > 0);
+  for (x = 0; x < iboxwidth; ++x) {
+    sum += src_ptr[x];
+  }
+  return sum;
+}
+
+static __inline uint32 SumPixels_16(int iboxwidth, const uint32* src_ptr) {
+  uint32 sum = 0u;
+  int x;
+  assert(iboxwidth > 0);
+  for (x = 0; x < iboxwidth; ++x) {
+    sum += src_ptr[x];
+  }
+  return sum;
+}
+
+static void ScaleAddCols2_C(int dst_width, int boxheight, int x, int dx,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int i;
+  int scaletbl[2];
+  int minboxwidth = dx >> 16;
+  int* scaleptr = scaletbl - minboxwidth;
+  int boxwidth;
+  scaletbl[0] = 65536 / (MIN1(minboxwidth) * boxheight);
+  scaletbl[1] = 65536 / (MIN1(minboxwidth + 1) * boxheight);
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = MIN1((x >> 16) - ix);
+    *dst_ptr++ = SumPixels(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16;
+  }
+}
+
+static void ScaleAddCols2_16_C(int dst_width, int boxheight, int x, int dx,
+                               const uint32* src_ptr, uint16* dst_ptr) {
+  int i;
+  int scaletbl[2];
+  int minboxwidth = dx >> 16;
+  int* scaleptr = scaletbl - minboxwidth;
+  int boxwidth;
+  scaletbl[0] = 65536 / (MIN1(minboxwidth) * boxheight);
+  scaletbl[1] = 65536 / (MIN1(minboxwidth + 1) * boxheight);
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = MIN1((x >> 16) - ix);
+    *dst_ptr++ =
+        SumPixels_16(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16;
+  }
+}
+
+static void ScaleAddCols0_C(int dst_width, int boxheight, int x, int,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int scaleval = 65536 / boxheight;
+  int i;
+  src_ptr += (x >> 16);
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = src_ptr[i] * scaleval >> 16;
+  }
+}
+
+static void ScaleAddCols1_C(int dst_width, int boxheight, int x, int dx,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int boxwidth = MIN1(dx >> 16);
+  int scaleval = 65536 / (boxwidth * boxheight);
+  int i;
+  x >>= 16;
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = SumPixels(boxwidth, src_ptr + x) * scaleval >> 16;
+    x += boxwidth;
+  }
+}
+
+static void ScaleAddCols1_16_C(int dst_width, int boxheight, int x, int dx,
+                               const uint32* src_ptr, uint16* dst_ptr) {
+  int boxwidth = MIN1(dx >> 16);
+  int scaleval = 65536 / (boxwidth * boxheight);
+  int i;
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = SumPixels_16(boxwidth, src_ptr + x) * scaleval >> 16;
+    x += boxwidth;
+  }
+}
+
+// Scale plane down to any dimensions, with interpolation.
+// (boxfilter).
+//
+// Same method as SimpleScale, which is fixed point, outputting
+// one pixel of destination using fixed point (16.16) to step
+// through source, sampling a box of pixel with simple
+// averaging.
+static void ScalePlaneBox(int src_width, int src_height,
+                          int dst_width, int dst_height,
+                          int src_stride, int dst_stride,
+                          const uint8* src_ptr, uint8* dst_ptr) {
+  int j, k;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height << 16);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  {
+    // Allocate a row buffer of uint16.
+    align_buffer_64(row16, src_width * 2);
+    void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
+        const uint16* src_ptr, uint8* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_C:
+        ((dx != 0x10000) ? ScaleAddCols1_C : ScaleAddCols0_C);
+    void (*ScaleAddRow)(const uint8* src_ptr, uint16* dst_ptr, int src_width) =
+        ScaleAddRow_C;
+#if defined(HAS_SCALEADDROW_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2)) {
+      ScaleAddRow = ScaleAddRow_Any_SSE2;
+      if (IS_ALIGNED(src_width, 16)) {
+        ScaleAddRow = ScaleAddRow_SSE2;
+      }
+    }
+#endif
+#if defined(HAS_SCALEADDROW_AVX2)
+    if (TestCpuFlag(kCpuHasAVX2)) {
+      ScaleAddRow = ScaleAddRow_Any_AVX2;
+      if (IS_ALIGNED(src_width, 32)) {
+        ScaleAddRow = ScaleAddRow_AVX2;
+      }
+    }
+#endif
+#if defined(HAS_SCALEADDROW_NEON)
+    if (TestCpuFlag(kCpuHasNEON)) {
+      ScaleAddRow = ScaleAddRow_Any_NEON;
+      if (IS_ALIGNED(src_width, 16)) {
+        ScaleAddRow = ScaleAddRow_NEON;
+      }
+    }
+#endif
+
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint8* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+      boxheight = MIN1((y >> 16) - iy);
+      memset(row16, 0, src_width * 2);
+      for (k = 0; k < boxheight; ++k) {
+        ScaleAddRow(src, (uint16 *)(row16), src_width);
+        src += src_stride;
+      }
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint16*)(row16), dst_ptr);
+      dst_ptr += dst_stride;
+    }
+    free_aligned_buffer_64(row16);
+  }
+}
+
+static void ScalePlaneBox_16(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint16* src_ptr, uint16* dst_ptr) {
+  int j, k;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height << 16);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  {
+    // Allocate a row buffer of uint32.
+    align_buffer_64(row32, src_width * 4);
+    void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
+        const uint32* src_ptr, uint16* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_16_C: ScaleAddCols1_16_C;
+    void (*ScaleAddRow)(const uint16* src_ptr, uint32* dst_ptr, int src_width) =
+        ScaleAddRow_16_C;
+
+#if defined(HAS_SCALEADDROW_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(src_width, 16)) {
+      ScaleAddRow = ScaleAddRow_16_SSE2;
+    }
+#endif
+
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint16* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+      boxheight = MIN1((y >> 16) - iy);
+      memset(row32, 0, src_width * 4);
+      for (k = 0; k < boxheight; ++k) {
+        ScaleAddRow(src, (uint32 *)(row32), src_width);
+        src += src_stride;
+      }
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint32*)(row32), dst_ptr);
+      dst_ptr += dst_stride;
+    }
+    free_aligned_buffer_64(row32);
+  }
+}
+
+// Scale plane down with bilinear interpolation.
+void ScalePlaneBilinearDown(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row buffer.
+  align_buffer_64(row, src_width);
+
+  const int max_y = (src_height - 1) << 16;
+  int j;
+  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleFilterCols64_C : ScaleFilterCols_C;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(src_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+
+#if defined(HAS_SCALEFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEFILTERCOLS_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleFilterCols = ScaleFilterCols_NEON;
+    }
+  }
+#endif
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    int yi = y >> 16;
+    const uint8* src = src_ptr + yi * src_stride;
+    if (filtering == kFilterLinear) {
+      ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow(row, src, src_stride, src_width, yf);
+      ScaleFilterCols(dst_ptr, row, dst_width, x, dx);
+    }
+    dst_ptr += dst_stride;
+    y += dy;
+    if (y > max_y) {
+      y = max_y;
+    }
+  }
+  free_aligned_buffer_64(row);
+}
+
+void ScalePlaneBilinearDown_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row buffer.
+  align_buffer_64(row, src_width * 2);
+
+  const int max_y = (src_height - 1) << 16;
+  int j;
+  void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleFilterCols64_16_C : ScaleFilterCols_16_C;
+  void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_16_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(src_width, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+
+#if defined(HAS_SCALEFILTERCOLS_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_16_SSSE3;
+  }
+#endif
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    int yi = y >> 16;
+    const uint16* src = src_ptr + yi * src_stride;
+    if (filtering == kFilterLinear) {
+      ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow((uint16*)row, src, src_stride, src_width, yf);
+      ScaleFilterCols(dst_ptr, (uint16*)row, dst_width, x, dx);
+    }
+    dst_ptr += dst_stride;
+    y += dy;
+    if (y > max_y) {
+      y = max_y;
+    }
+  }
+  free_aligned_buffer_64(row);
+}
+
+// Scale up down with bilinear interpolation.
+void ScalePlaneBilinearUp(int src_width, int src_height,
+                          int dst_width, int dst_height,
+                          int src_stride, int dst_stride,
+                          const uint8* src_ptr, uint8* dst_ptr,
+                          enum FilterMode filtering) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height - 1) << 16;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleFilterCols_C : ScaleCols_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (filtering && src_width >= 32768) {
+    ScaleFilterCols = ScaleFilterCols64_C;
+  }
+#if defined(HAS_SCALEFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEFILTERCOLS_NEON)
+  if (filtering && TestCpuFlag(kCpuHasNEON) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleFilterCols = ScaleFilterCols_NEON;
+    }
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleFilterCols = ScaleColsUp2_C;
+#if defined(HAS_SCALECOLS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleFilterCols = ScaleColsUp2_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+  {
+    int yi = y >> 16;
+    const uint8* src = src_ptr + yi * src_stride;
+
+    // Allocate 2 row buffers.
+    const int kRowSize = (dst_width + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+
+    uint8* rowptr = row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_ptr + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf);
+      }
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+void ScalePlaneBilinearUp_16(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint16* src_ptr, uint16* dst_ptr,
+                             enum FilterMode filtering) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height - 1) << 16;
+  void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleFilterCols_16_C : ScaleCols_16_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_16_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (filtering && src_width >= 32768) {
+    ScaleFilterCols = ScaleFilterCols64_16_C;
+  }
+#if defined(HAS_SCALEFILTERCOLS_16_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_16_SSSE3;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleFilterCols = ScaleColsUp2_16_C;
+#if defined(HAS_SCALECOLS_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleFilterCols = ScaleColsUp2_16_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+  {
+    int yi = y >> 16;
+    const uint16* src = src_ptr + yi * src_stride;
+
+    // Allocate 2 row buffers.
+    const int kRowSize = (dst_width + 31) & ~31;
+    align_buffer_64(row, kRowSize * 4);
+
+    uint16* rowptr = (uint16*)row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_ptr + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf);
+      }
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+// Scale Plane to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScalePlaneSimple(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr) {
+  int i;
+  void (*ScaleCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) = ScaleCols_C;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleCols = ScaleColsUp2_C;
+#if defined(HAS_SCALECOLS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleCols = ScaleColsUp2_SSE2;
+    }
+#endif
+  }
+
+  for (i = 0; i < dst_height; ++i) {
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride, dst_width, x, dx);
+    dst_ptr += dst_stride;
+    y += dy;
+  }
+}
+
+static void ScalePlaneSimple_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr) {
+  int i;
+  void (*ScaleCols)(uint16* dst_ptr, const uint16* src_ptr,
+      int dst_width, int x, int dx) = ScaleCols_16_C;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleCols = ScaleColsUp2_16_C;
+#if defined(HAS_SCALECOLS_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleCols = ScaleColsUp2_16_SSE2;
+    }
+#endif
+  }
+
+  for (i = 0; i < dst_height; ++i) {
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride,
+              dst_width, x, dx);
+    dst_ptr += dst_stride;
+    y += dy;
+  }
+}
+
+// Scale a plane.
+// This function dispatches to a specialized scaler based on scale factor.
+
+LIBYUV_API
+void ScalePlane(const uint8* src, int src_stride,
+                int src_width, int src_height,
+                uint8* dst, int dst_stride,
+                int dst_width, int dst_height,
+                enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height, filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  // Use specialized scales to improve performance for common resolutions.
+  // For example, all the 1/2 scalings will use ScalePlaneDown2()
+  if (dst_width == src_width && dst_height == src_height) {
+    // Straight copy.
+    CopyPlane(src, src_stride, dst, dst_stride, dst_width, dst_height);
+    return;
+  }
+  if (dst_width == src_width && filtering != kFilterBox) {
+    int dy = FixedDiv(src_height, dst_height);
+    // Arbitrary scale vertically, but unscaled horizontally.
+    ScalePlaneVertical(src_height,
+                       dst_width, dst_height,
+                       src_stride, dst_stride, src, dst,
+                       0, 0, dy, 1, filtering);
+    return;
+  }
+  if (dst_width <= Abs(src_width) && dst_height <= src_height) {
+    // Scale down.
+    if (4 * dst_width == 3 * src_width &&
+        4 * dst_height == 3 * src_height) {
+      // optimized, 3/4
+      ScalePlaneDown34(src_width, src_height, dst_width, dst_height,
+                       src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (2 * dst_width == src_width && 2 * dst_height == src_height) {
+      // optimized, 1/2
+      ScalePlaneDown2(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    // 3/8 rounded up for odd sized chroma height.
+    if (8 * dst_width == 3 * src_width &&
+        dst_height == ((src_height * 3 + 7) / 8)) {
+      // optimized, 3/8
+      ScalePlaneDown38(src_width, src_height, dst_width, dst_height,
+                       src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (4 * dst_width == src_width && 4 * dst_height == src_height &&
+        (filtering == kFilterBox || filtering == kFilterNone)) {
+      // optimized, 1/4
+      ScalePlaneDown4(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+  }
+  if (filtering == kFilterBox && dst_height * 2 < src_height) {
+    ScalePlaneBox(src_width, src_height, dst_width, dst_height,
+                  src_stride, dst_stride, src, dst);
+    return;
+  }
+  if (filtering && dst_height > src_height) {
+    ScalePlaneBilinearUp(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  if (filtering) {
+    ScalePlaneBilinearDown(src_width, src_height, dst_width, dst_height,
+                           src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
+                   src_stride, dst_stride, src, dst);
+}
+
+LIBYUV_API
+void ScalePlane_16(const uint16* src, int src_stride,
+                  int src_width, int src_height,
+                  uint16* dst, int dst_stride,
+                  int dst_width, int dst_height,
+                  enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height, filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  // Use specialized scales to improve performance for common resolutions.
+  // For example, all the 1/2 scalings will use ScalePlaneDown2()
+  if (dst_width == src_width && dst_height == src_height) {
+    // Straight copy.
+    CopyPlane_16(src, src_stride, dst, dst_stride, dst_width, dst_height);
+    return;
+  }
+  if (dst_width == src_width) {
+    int dy = FixedDiv(src_height, dst_height);
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical_16(src_height,
+                          dst_width, dst_height,
+                          src_stride, dst_stride, src, dst,
+                          0, 0, dy, 1, filtering);
+    return;
+  }
+  if (dst_width <= Abs(src_width) && dst_height <= src_height) {
+    // Scale down.
+    if (4 * dst_width == 3 * src_width &&
+        4 * dst_height == 3 * src_height) {
+      // optimized, 3/4
+      ScalePlaneDown34_16(src_width, src_height, dst_width, dst_height,
+                          src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (2 * dst_width == src_width && 2 * dst_height == src_height) {
+      // optimized, 1/2
+      ScalePlaneDown2_16(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    // 3/8 rounded up for odd sized chroma height.
+    if (8 * dst_width == 3 * src_width &&
+        dst_height == ((src_height * 3 + 7) / 8)) {
+      // optimized, 3/8
+      ScalePlaneDown38_16(src_width, src_height, dst_width, dst_height,
+                          src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (4 * dst_width == src_width && 4 * dst_height == src_height &&
+               filtering != kFilterBilinear) {
+      // optimized, 1/4
+      ScalePlaneDown4_16(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+  }
+  if (filtering == kFilterBox && dst_height * 2 < src_height) {
+    ScalePlaneBox_16(src_width, src_height, dst_width, dst_height,
+                     src_stride, dst_stride, src, dst);
+    return;
+  }
+  if (filtering && dst_height > src_height) {
+    ScalePlaneBilinearUp_16(src_width, src_height, dst_width, dst_height,
+                            src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  if (filtering) {
+    ScalePlaneBilinearDown_16(src_width, src_height, dst_width, dst_height,
+                              src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  ScalePlaneSimple_16(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst);
+}
+
+// Scale an I420 image.
+// This function in turn calls a scaling function for each plane.
+
+LIBYUV_API
+int I420Scale(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              int src_width, int src_height,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int dst_width, int dst_height,
+              enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 ||
+      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane(src_y, src_stride_y, src_width, src_height,
+             dst_y, dst_stride_y, dst_width, dst_height,
+             filtering);
+  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight,
+             dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
+             filtering);
+  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight,
+             dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
+             filtering);
+  return 0;
+}
+
+LIBYUV_API
+int I420Scale_16(const uint16* src_y, int src_stride_y,
+                 const uint16* src_u, int src_stride_u,
+                 const uint16* src_v, int src_stride_v,
+                 int src_width, int src_height,
+                 uint16* dst_y, int dst_stride_y,
+                 uint16* dst_u, int dst_stride_u,
+                 uint16* dst_v, int dst_stride_v,
+                 int dst_width, int dst_height,
+                 enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 ||
+      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_16(src_y, src_stride_y, src_width, src_height,
+                dst_y, dst_stride_y, dst_width, dst_height,
+                filtering);
+  ScalePlane_16(src_u, src_stride_u, src_halfwidth, src_halfheight,
+                dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
+                filtering);
+  ScalePlane_16(src_v, src_stride_v, src_halfwidth, src_halfheight,
+                dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
+                filtering);
+  return 0;
+}
+
+// Deprecated api
+LIBYUV_API
+int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
+          int src_stride_y, int src_stride_u, int src_stride_v,
+          int src_width, int src_height,
+          uint8* dst_y, uint8* dst_u, uint8* dst_v,
+          int dst_stride_y, int dst_stride_u, int dst_stride_v,
+          int dst_width, int dst_height,
+          LIBYUV_BOOL interpolate) {
+  return I420Scale(src_y, src_stride_y,
+                   src_u, src_stride_u,
+                   src_v, src_stride_v,
+                   src_width, src_height,
+                   dst_y, dst_stride_y,
+                   dst_u, dst_stride_u,
+                   dst_v, dst_stride_v,
+                   dst_width, dst_height,
+                   interpolate ? kFilterBox : kFilterNone);
+}
+
+// Deprecated api
+LIBYUV_API
+int ScaleOffset(const uint8* src, int src_width, int src_height,
+                uint8* dst, int dst_width, int dst_height, int dst_yoffset,
+                LIBYUV_BOOL interpolate) {
+  // Chroma requires offset to multiple of 2.
+  int dst_yoffset_even = dst_yoffset & ~1;
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  int aheight = dst_height - dst_yoffset_even * 2;  // actual output height
+  const uint8* src_y = src;
+  const uint8* src_u = src + src_width * src_height;
+  const uint8* src_v = src + src_width * src_height +
+                             src_halfwidth * src_halfheight;
+  uint8* dst_y = dst + dst_yoffset_even * dst_width;
+  uint8* dst_u = dst + dst_width * dst_height +
+                 (dst_yoffset_even >> 1) * dst_halfwidth;
+  uint8* dst_v = dst + dst_width * dst_height + dst_halfwidth * dst_halfheight +
+                 (dst_yoffset_even >> 1) * dst_halfwidth;
+  if (!src || src_width <= 0 || src_height <= 0 ||
+      !dst || dst_width <= 0 || dst_height <= 0 || dst_yoffset_even < 0 ||
+      dst_yoffset_even >= dst_height) {
+    return -1;
+  }
+  return I420Scale(src_y, src_width,
+                   src_u, src_halfwidth,
+                   src_v, src_halfwidth,
+                   src_width, src_height,
+                   dst_y, dst_width,
+                   dst_u, dst_halfwidth,
+                   dst_v, dst_halfwidth,
+                   dst_width, aheight,
+                   interpolate ? kFilterBox : kFilterNone);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/scale_any.cc b/libvpx/libvpx/third_party/libyuv/source/scale_any.cc
new file mode 100644
index 0000000..2f6a2c8
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/scale_any.cc
@@ -0,0 +1,200 @@
+/*
+ *  Copyright 2015 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+#include "libyuv/scale_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Definition for ScaleFilterCols, ScaleARGBCols and ScaleARGBFilterCols
+#define CANY(NAMEANY, TERP_SIMD, TERP_C, BPP, MASK)                            \
+    void NAMEANY(uint8* dst_ptr, const uint8* src_ptr,                         \
+                 int dst_width, int x, int dx) {                               \
+      int n = dst_width & ~MASK;                                               \
+      if (n > 0) {                                                             \
+        TERP_SIMD(dst_ptr, src_ptr, n, x, dx);                                 \
+      }                                                                        \
+      TERP_C(dst_ptr + n * BPP, src_ptr,                                       \
+             dst_width & MASK, x + n * dx, dx);                                \
+    }
+
+#ifdef HAS_SCALEFILTERCOLS_NEON
+CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7)
+#endif
+#ifdef HAS_SCALEARGBCOLS_NEON
+CANY(ScaleARGBCols_Any_NEON, ScaleARGBCols_NEON, ScaleARGBCols_C, 4, 7)
+#endif
+#ifdef HAS_SCALEARGBFILTERCOLS_NEON
+CANY(ScaleARGBFilterCols_Any_NEON, ScaleARGBFilterCols_NEON,
+     ScaleARGBFilterCols_C, 4, 3)
+#endif
+#undef CANY
+
+// Fixed scale down.
+#define SDANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK)   \
+    void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride,                   \
+                 uint8* dst_ptr, int dst_width) {                              \
+      int r = (int)((unsigned int)dst_width % (MASK + 1));                     \
+      int n = dst_width - r;                                                   \
+      if (n > 0) {                                                             \
+        SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                    \
+      }                                                                        \
+      SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                 \
+                     dst_ptr + n * BPP, r);                                    \
+    }
+
+#ifdef HAS_SCALEROWDOWN2_SSE2
+SDANY(ScaleRowDown2_Any_SSE2, ScaleRowDown2_SSE2, ScaleRowDown2_C, 2, 1, 15)
+SDANY(ScaleRowDown2Linear_Any_SSE2, ScaleRowDown2Linear_SSE2,
+      ScaleRowDown2Linear_C, 2, 1, 15)
+SDANY(ScaleRowDown2Box_Any_SSE2, ScaleRowDown2Box_SSE2, ScaleRowDown2Box_C,
+      2, 1, 15)
+#endif
+#ifdef HAS_SCALEROWDOWN2_AVX2
+SDANY(ScaleRowDown2_Any_AVX2, ScaleRowDown2_AVX2, ScaleRowDown2_C, 2, 1, 31)
+SDANY(ScaleRowDown2Linear_Any_AVX2, ScaleRowDown2Linear_AVX2,
+      ScaleRowDown2Linear_C, 2, 1, 31)
+SDANY(ScaleRowDown2Box_Any_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_C,
+      2, 1, 31)
+#endif
+#ifdef HAS_SCALEROWDOWN2_NEON
+SDANY(ScaleRowDown2_Any_NEON, ScaleRowDown2_NEON, ScaleRowDown2_C, 2, 1, 15)
+SDANY(ScaleRowDown2Linear_Any_NEON, ScaleRowDown2Linear_NEON,
+      ScaleRowDown2Linear_C, 2, 1, 15)
+SDANY(ScaleRowDown2Box_Any_NEON, ScaleRowDown2Box_NEON,
+      ScaleRowDown2Box_C, 2, 1, 15)
+#endif
+#ifdef HAS_SCALEROWDOWN4_SSE2
+SDANY(ScaleRowDown4_Any_SSE2, ScaleRowDown4_SSE2, ScaleRowDown4_C, 4, 1, 7)
+SDANY(ScaleRowDown4Box_Any_SSE2, ScaleRowDown4Box_SSE2, ScaleRowDown4Box_C,
+      4, 1, 7)
+#endif
+#ifdef HAS_SCALEROWDOWN4_AVX2
+SDANY(ScaleRowDown4_Any_AVX2, ScaleRowDown4_AVX2, ScaleRowDown4_C, 4, 1, 15)
+SDANY(ScaleRowDown4Box_Any_AVX2, ScaleRowDown4Box_AVX2, ScaleRowDown4Box_C,
+      4, 1, 15)
+#endif
+#ifdef HAS_SCALEROWDOWN4_NEON
+SDANY(ScaleRowDown4_Any_NEON, ScaleRowDown4_NEON, ScaleRowDown4_C, 4, 1, 7)
+SDANY(ScaleRowDown4Box_Any_NEON, ScaleRowDown4Box_NEON, ScaleRowDown4Box_C,
+      4, 1, 7)
+#endif
+#ifdef HAS_SCALEROWDOWN34_SSSE3
+SDANY(ScaleRowDown34_Any_SSSE3, ScaleRowDown34_SSSE3,
+      ScaleRowDown34_C, 4 / 3, 1, 23)
+SDANY(ScaleRowDown34_0_Box_Any_SSSE3, ScaleRowDown34_0_Box_SSSE3,
+      ScaleRowDown34_0_Box_C, 4 / 3, 1, 23)
+SDANY(ScaleRowDown34_1_Box_Any_SSSE3, ScaleRowDown34_1_Box_SSSE3,
+      ScaleRowDown34_1_Box_C, 4 / 3, 1, 23)
+#endif
+#ifdef HAS_SCALEROWDOWN34_NEON
+SDANY(ScaleRowDown34_Any_NEON, ScaleRowDown34_NEON,
+      ScaleRowDown34_C, 4 / 3, 1, 23)
+SDANY(ScaleRowDown34_0_Box_Any_NEON, ScaleRowDown34_0_Box_NEON,
+      ScaleRowDown34_0_Box_C, 4 / 3, 1, 23)
+SDANY(ScaleRowDown34_1_Box_Any_NEON, ScaleRowDown34_1_Box_NEON,
+      ScaleRowDown34_1_Box_C, 4 / 3, 1, 23)
+#endif
+#ifdef HAS_SCALEROWDOWN38_SSSE3
+SDANY(ScaleRowDown38_Any_SSSE3, ScaleRowDown38_SSSE3,
+      ScaleRowDown38_C, 8 / 3, 1, 11)
+SDANY(ScaleRowDown38_3_Box_Any_SSSE3, ScaleRowDown38_3_Box_SSSE3,
+      ScaleRowDown38_3_Box_C, 8 / 3, 1, 5)
+SDANY(ScaleRowDown38_2_Box_Any_SSSE3, ScaleRowDown38_2_Box_SSSE3,
+      ScaleRowDown38_2_Box_C, 8 / 3, 1, 5)
+#endif
+#ifdef HAS_SCALEROWDOWN38_NEON
+SDANY(ScaleRowDown38_Any_NEON, ScaleRowDown38_NEON,
+      ScaleRowDown38_C, 8 / 3, 1, 11)
+SDANY(ScaleRowDown38_3_Box_Any_NEON, ScaleRowDown38_3_Box_NEON,
+      ScaleRowDown38_3_Box_C, 8 / 3, 1, 11)
+SDANY(ScaleRowDown38_2_Box_Any_NEON, ScaleRowDown38_2_Box_NEON,
+      ScaleRowDown38_2_Box_C, 8 / 3, 1, 11)
+#endif
+
+#ifdef HAS_SCALEARGBROWDOWN2_SSE2
+SDANY(ScaleARGBRowDown2_Any_SSE2, ScaleARGBRowDown2_SSE2,
+      ScaleARGBRowDown2_C, 2, 4, 3)
+SDANY(ScaleARGBRowDown2Linear_Any_SSE2, ScaleARGBRowDown2Linear_SSE2,
+      ScaleARGBRowDown2Linear_C, 2, 4, 3)
+SDANY(ScaleARGBRowDown2Box_Any_SSE2, ScaleARGBRowDown2Box_SSE2,
+      ScaleARGBRowDown2Box_C, 2, 4, 3)
+#endif
+#ifdef HAS_SCALEARGBROWDOWN2_NEON
+SDANY(ScaleARGBRowDown2_Any_NEON, ScaleARGBRowDown2_NEON,
+      ScaleARGBRowDown2_C, 2, 4, 7)
+SDANY(ScaleARGBRowDown2Linear_Any_NEON, ScaleARGBRowDown2Linear_NEON,
+      ScaleARGBRowDown2Linear_C, 2, 4, 7)
+SDANY(ScaleARGBRowDown2Box_Any_NEON, ScaleARGBRowDown2Box_NEON,
+      ScaleARGBRowDown2Box_C, 2, 4, 7)
+#endif
+#undef SDANY
+
+// Scale down by even scale factor.
+#define SDAANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, BPP, MASK)          \
+    void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, int src_stepx,    \
+                 uint8* dst_ptr, int dst_width) {                              \
+      int r = (int)((unsigned int)dst_width % (MASK + 1));                     \
+      int n = dst_width - r;                                                   \
+      if (n > 0) {                                                             \
+        SCALEROWDOWN_SIMD(src_ptr, src_stride, src_stepx, dst_ptr, n);         \
+      }                                                                        \
+      SCALEROWDOWN_C(src_ptr + (n * src_stepx) * BPP, src_stride,              \
+                     src_stepx, dst_ptr + n * BPP, r);                         \
+    }
+
+#ifdef HAS_SCALEARGBROWDOWNEVEN_SSE2
+SDAANY(ScaleARGBRowDownEven_Any_SSE2, ScaleARGBRowDownEven_SSE2,
+       ScaleARGBRowDownEven_C, 4, 3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_SSE2, ScaleARGBRowDownEvenBox_SSE2,
+       ScaleARGBRowDownEvenBox_C, 4, 3)
+#endif
+#ifdef HAS_SCALEARGBROWDOWNEVEN_NEON
+SDAANY(ScaleARGBRowDownEven_Any_NEON, ScaleARGBRowDownEven_NEON,
+       ScaleARGBRowDownEven_C, 4, 3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_NEON, ScaleARGBRowDownEvenBox_NEON,
+       ScaleARGBRowDownEvenBox_C, 4, 3)
+#endif
+
+// Add rows box filter scale down.
+#define SAANY(NAMEANY, SCALEADDROW_SIMD, SCALEADDROW_C, MASK)                  \
+  void NAMEANY(const uint8* src_ptr, uint16* dst_ptr, int src_width) {         \
+      int n = src_width & ~MASK;                                               \
+      if (n > 0) {                                                             \
+        SCALEADDROW_SIMD(src_ptr, dst_ptr, n);                                 \
+      }                                                                        \
+      SCALEADDROW_C(src_ptr + n, dst_ptr + n, src_width & MASK);               \
+    }
+
+#ifdef HAS_SCALEADDROW_SSE2
+SAANY(ScaleAddRow_Any_SSE2, ScaleAddRow_SSE2, ScaleAddRow_C, 15)
+#endif
+#ifdef HAS_SCALEADDROW_AVX2
+SAANY(ScaleAddRow_Any_AVX2, ScaleAddRow_AVX2, ScaleAddRow_C, 31)
+#endif
+#ifdef HAS_SCALEADDROW_NEON
+SAANY(ScaleAddRow_Any_NEON, ScaleAddRow_NEON, ScaleAddRow_C, 15)
+#endif
+#undef SAANY
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+
+
+
+
diff --git a/libvpx/libvpx/third_party/libyuv/source/scale_argb.cc b/libvpx/libvpx/third_party/libyuv/source/scale_argb.cc
new file mode 100644
index 0000000..40a2d1a
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/scale_argb.cc
@@ -0,0 +1,853 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyARGB
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// ScaleARGB ARGB, 1/2
+// This is an optimized version for scaling down a ARGB to 1/2 of
+// its original size.
+static void ScaleARGBDown2(int src_width, int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint8* src_argb, uint8* dst_argb,
+                           int x, int dx, int y, int dy,
+                           enum FilterMode filtering) {
+  int j;
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) =
+    filtering == kFilterNone ? ScaleARGBRowDown2_C :
+        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_C :
+        ScaleARGBRowDown2Box_C);
+  assert(dx == 65536 * 2);  // Test scale factor of 2.
+  assert((dy & 0x1ffff) == 0);  // Test vertical scale is multiple of 2.
+  // Advance to odd row, even column.
+  if (filtering == kFilterBilinear) {
+    src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  } else {
+    src_argb += (y >> 16) * src_stride + ((x >> 16) - 1) * 4;
+  }
+
+#if defined(HAS_SCALEARGBROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_Any_SSE2 :
+        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_SSE2 :
+        ScaleARGBRowDown2Box_Any_SSE2);
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_SSE2 :
+          (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_SSE2 :
+          ScaleARGBRowDown2Box_SSE2);
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_Any_NEON :
+        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_NEON :
+        ScaleARGBRowDown2Box_Any_NEON);
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_NEON :
+          (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_NEON :
+          ScaleARGBRowDown2Box_NEON);
+    }
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDown2(src_argb, src_stride, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+}
+
+// ScaleARGB ARGB, 1/4
+// This is an optimized version for scaling down a ARGB to 1/4 of
+// its original size.
+static void ScaleARGBDown4Box(int src_width, int src_height,
+                              int dst_width, int dst_height,
+                              int src_stride, int dst_stride,
+                              const uint8* src_argb, uint8* dst_argb,
+                              int x, int dx, int y, int dy) {
+  int j;
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (dst_width * 2 * 4 + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2);
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
+    uint8* dst_argb, int dst_width) = ScaleARGBRowDown2Box_C;
+  // Advance to odd row, even column.
+  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  assert(dx == 65536 * 4);  // Test scale factor of 4.
+  assert((dy & 0x3ffff) == 0);  // Test vertical scale is multiple of 4.
+#if defined(HAS_SCALEARGBROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleARGBRowDown2 = ScaleARGBRowDown2Box_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDown2 = ScaleARGBRowDown2Box_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBRowDown2 = ScaleARGBRowDown2Box_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBRowDown2 = ScaleARGBRowDown2Box_NEON;
+    }
+  }
+#endif
+
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDown2(src_argb, src_stride, row, dst_width * 2);
+    ScaleARGBRowDown2(src_argb + src_stride * 2, src_stride,
+                      row + kRowSize, dst_width * 2);
+    ScaleARGBRowDown2(row, kRowSize, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+// ScaleARGB ARGB Even
+// This is an optimized version for scaling down a ARGB to even
+// multiple of its original size.
+static void ScaleARGBDownEven(int src_width, int src_height,
+                              int dst_width, int dst_height,
+                              int src_stride, int dst_stride,
+                              const uint8* src_argb, uint8* dst_argb,
+                              int x, int dx, int y, int dy,
+                              enum FilterMode filtering) {
+  int j;
+  int col_step = dx >> 16;
+  int row_stride = (dy >> 16) * src_stride;
+  void (*ScaleARGBRowDownEven)(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_step, uint8* dst_argb, int dst_width) =
+      filtering ? ScaleARGBRowDownEvenBox_C : ScaleARGBRowDownEven_C;
+  assert(IS_ALIGNED(src_width, 2));
+  assert(IS_ALIGNED(src_height, 2));
+  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_SSE2 :
+        ScaleARGBRowDownEven_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_SSE2 :
+          ScaleARGBRowDownEven_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_NEON :
+        ScaleARGBRowDownEven_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_NEON :
+          ScaleARGBRowDownEven_NEON;
+    }
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDownEven(src_argb, src_stride, col_step, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+}
+
+// Scale ARGB down with bilinear interpolation.
+static void ScaleARGBBilinearDown(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  int src_stride, int dst_stride,
+                                  const uint8* src_argb, uint8* dst_argb,
+                                  int x, int dx, int y, int dy,
+                                  enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleARGBFilterCols64_C : ScaleARGBFilterCols_C;
+  int64 xlast = x + (int64)(dst_width - 1) * dx;
+  int64 xl = (dx >= 0) ? x : xlast;
+  int64 xr = (dx >= 0) ? xlast : x;
+  int clip_src_width;
+  xl = (xl >> 16) & ~3;  // Left edge aligned.
+  xr = (xr >> 16) + 1;  // Right most pixel used.  Bilinear uses 2 pixels.
+  xr = (xr + 1 + 3) & ~3;  // 1 beyond 4 pixel aligned right most pixel.
+  if (xr > src_width) {
+    xr = src_width;
+  }
+  clip_src_width = (int)(xr - xl) * 4;  // Width aligned to 4.
+  src_argb += xl * 4;
+  x -= (int)(xl << 16);
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(clip_src_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_NEON;
+    }
+  }
+#endif
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row of ARGB.
+  {
+    align_buffer_64(row, clip_src_width * 4);
+
+    const int max_y = (src_height - 1) << 16;
+    if (y > max_y) {
+      y = max_y;
+    }
+    for (j = 0; j < dst_height; ++j) {
+      int yi = y >> 16;
+      const uint8* src = src_argb + yi * src_stride;
+      if (filtering == kFilterLinear) {
+        ScaleARGBFilterCols(dst_argb, src, dst_width, x, dx);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(row, src, src_stride, clip_src_width, yf);
+        ScaleARGBFilterCols(dst_argb, row, dst_width, x, dx);
+      }
+      dst_argb += dst_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+// Scale ARGB up with bilinear interpolation.
+static void ScaleARGBBilinearUp(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint8* src_argb, uint8* dst_argb,
+                                int x, int dx, int y, int dy,
+                                enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
+  const int max_y = (src_height - 1) << 16;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+  if (src_width >= 32768) {
+    ScaleARGBFilterCols = filtering ?
+        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
+  }
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_NEON)
+  if (filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBCols_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_NEON)
+  if (!filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBCols_NEON;
+    }
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBFilterCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  {
+    int yi = y >> 16;
+    const uint8* src = src_argb + yi * src_stride;
+
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (dst_width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+
+    uint8* rowptr = row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleARGBFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_argb + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
+      }
+      dst_argb += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+#ifdef YUVSCALEUP
+// Scale YUV to ARGB up with bilinear interpolation.
+static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
+                                     int dst_width, int dst_height,
+                                     int src_stride_y,
+                                     int src_stride_u,
+                                     int src_stride_v,
+                                     int dst_stride_argb,
+                                     const uint8* src_y,
+                                     const uint8* src_u,
+                                     const uint8* src_v,
+                                     uint8* dst_argb,
+                                     int x, int dx, int y, int dy,
+                                     enum FilterMode filtering) {
+  int j;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(src_width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(src_width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(src_width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
+  if (src_width >= 32768) {
+    ScaleARGBFilterCols = filtering ?
+        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
+  }
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_NEON)
+  if (filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBCols_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_NEON)
+  if (!filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBCols_NEON;
+    }
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBFilterCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  const int max_y = (src_height - 1) << 16;
+  if (y > max_y) {
+    y = max_y;
+  }
+  const int kYShift = 1;  // Shift Y by 1 to convert Y plane to UV coordinate.
+  int yi = y >> 16;
+  int uv_yi = yi >> kYShift;
+  const uint8* src_row_y = src_y + yi * src_stride_y;
+  const uint8* src_row_u = src_u + uv_yi * src_stride_u;
+  const uint8* src_row_v = src_v + uv_yi * src_stride_v;
+
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (dst_width * 4 + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2);
+
+  // Allocate 1 row of ARGB for source conversion.
+  align_buffer_64(argb_row, src_width * 4);
+
+  uint8* rowptr = row;
+  int rowstride = kRowSize;
+  int lasty = yi;
+
+  // TODO(fbarchard): Convert first 2 rows of YUV to ARGB.
+  ScaleARGBFilterCols(rowptr, src_row_y, dst_width, x, dx);
+  if (src_height > 1) {
+    src_row_y += src_stride_y;
+    if (yi & 1) {
+      src_row_u += src_stride_u;
+      src_row_v += src_stride_v;
+    }
+  }
+  ScaleARGBFilterCols(rowptr + rowstride, src_row_y, dst_width, x, dx);
+  if (src_height > 2) {
+    src_row_y += src_stride_y;
+    if (!(yi & 1)) {
+      src_row_u += src_stride_u;
+      src_row_v += src_stride_v;
+    }
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    yi = y >> 16;
+    if (yi != lasty) {
+      if (y > max_y) {
+        y = max_y;
+        yi = y >> 16;
+        uv_yi = yi >> kYShift;
+        src_row_y = src_y + yi * src_stride_y;
+        src_row_u = src_u + uv_yi * src_stride_u;
+        src_row_v = src_v + uv_yi * src_stride_v;
+      }
+      if (yi != lasty) {
+        // TODO(fbarchard): Convert the clipped region of row.
+        I422ToARGBRow(src_row_y, src_row_u, src_row_v, argb_row, src_width);
+        ScaleARGBFilterCols(rowptr, argb_row, dst_width, x, dx);
+        rowptr += rowstride;
+        rowstride = -rowstride;
+        lasty = yi;
+        src_row_y += src_stride_y;
+        if (yi & 1) {
+          src_row_u += src_stride_u;
+          src_row_v += src_stride_v;
+        }
+      }
+    }
+    if (filtering == kFilterLinear) {
+      InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
+    }
+    dst_argb += dst_stride_argb;
+    y += dy;
+  }
+  free_aligned_buffer_64(row);
+  free_aligned_buffer_64(row_argb);
+}
+#endif
+
+// Scale ARGB to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScaleARGBSimple(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_argb, uint8* dst_argb,
+                            int x, int dx, int y, int dy) {
+  int j;
+  void (*ScaleARGBCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleARGBCols64_C : ScaleARGBCols_C;
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBCols = ScaleARGBCols_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBCols = ScaleARGBCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBCols = ScaleARGBCols_NEON;
+    }
+  }
+#endif
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBCols(dst_argb, src_argb + (y >> 16) * src_stride,
+                  dst_width, x, dx);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
+// ScaleARGB a ARGB.
+// This function in turn calls a scaling function
+// suitable for handling the desired resolutions.
+static void ScaleARGB(const uint8* src, int src_stride,
+                      int src_width, int src_height,
+                      uint8* dst, int dst_stride,
+                      int dst_width, int dst_height,
+                      int clip_x, int clip_y, int clip_width, int clip_height,
+                      enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // ARGB does not support box filter yet, but allow the user to pass it.
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
+  // Negative src_height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  if (clip_x) {
+    int64 clipf = (int64)(clip_x) * dx;
+    x += (clipf & 0xffff);
+    src += (clipf >> 16) * 4;
+    dst += clip_x * 4;
+  }
+  if (clip_y) {
+    int64 clipf = (int64)(clip_y) * dy;
+    y += (clipf & 0xffff);
+    src += (clipf >> 16) * src_stride;
+    dst += clip_y * dst_stride;
+  }
+
+  // Special case for integer step values.
+  if (((dx | dy) & 0xffff) == 0) {
+    if (!dx || !dy) {  // 1 pixel wide and/or tall.
+      filtering = kFilterNone;
+    } else {
+      // Optimized even scale down. ie 2, 4, 6, 8, 10x.
+      if (!(dx & 0x10000) && !(dy & 0x10000)) {
+        if (dx == 0x20000) {
+          // Optimized 1/2 downsample.
+          ScaleARGBDown2(src_width, src_height,
+                         clip_width, clip_height,
+                         src_stride, dst_stride, src, dst,
+                         x, dx, y, dy, filtering);
+          return;
+        }
+        if (dx == 0x40000 && filtering == kFilterBox) {
+          // Optimized 1/4 box downsample.
+          ScaleARGBDown4Box(src_width, src_height,
+                            clip_width, clip_height,
+                            src_stride, dst_stride, src, dst,
+                            x, dx, y, dy);
+          return;
+        }
+        ScaleARGBDownEven(src_width, src_height,
+                          clip_width, clip_height,
+                          src_stride, dst_stride, src, dst,
+                          x, dx, y, dy, filtering);
+        return;
+      }
+      // Optimized odd scale down. ie 3, 5, 7, 9x.
+      if ((dx & 0x10000) && (dy & 0x10000)) {
+        filtering = kFilterNone;
+        if (dx == 0x10000 && dy == 0x10000) {
+          // Straight copy.
+          ARGBCopy(src + (y >> 16) * src_stride + (x >> 16) * 4, src_stride,
+                   dst, dst_stride, clip_width, clip_height);
+          return;
+        }
+      }
+    }
+  }
+  if (dx == 0x10000 && (x & 0xffff) == 0) {
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical(src_height,
+                       clip_width, clip_height,
+                       src_stride, dst_stride, src, dst,
+                       x, y, dy, 4, filtering);
+    return;
+  }
+  if (filtering && dy < 65536) {
+    ScaleARGBBilinearUp(src_width, src_height,
+                        clip_width, clip_height,
+                        src_stride, dst_stride, src, dst,
+                        x, dx, y, dy, filtering);
+    return;
+  }
+  if (filtering) {
+    ScaleARGBBilinearDown(src_width, src_height,
+                          clip_width, clip_height,
+                          src_stride, dst_stride, src, dst,
+                          x, dx, y, dy, filtering);
+    return;
+  }
+  ScaleARGBSimple(src_width, src_height, clip_width, clip_height,
+                  src_stride, dst_stride, src, dst,
+                  x, dx, y, dy);
+}
+
+LIBYUV_API
+int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
+                  int src_width, int src_height,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int dst_width, int dst_height,
+                  int clip_x, int clip_y, int clip_width, int clip_height,
+                  enum FilterMode filtering) {
+  if (!src_argb || src_width == 0 || src_height == 0 ||
+      !dst_argb || dst_width <= 0 || dst_height <= 0 ||
+      clip_x < 0 || clip_y < 0 ||
+      clip_width > 32768 || clip_height > 32768 ||
+      (clip_x + clip_width) > dst_width ||
+      (clip_y + clip_height) > dst_height) {
+    return -1;
+  }
+  ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
+            dst_argb, dst_stride_argb, dst_width, dst_height,
+            clip_x, clip_y, clip_width, clip_height, filtering);
+  return 0;
+}
+
+// Scale an ARGB image.
+LIBYUV_API
+int ARGBScale(const uint8* src_argb, int src_stride_argb,
+              int src_width, int src_height,
+              uint8* dst_argb, int dst_stride_argb,
+              int dst_width, int dst_height,
+              enum FilterMode filtering) {
+  if (!src_argb || src_width == 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 ||
+      !dst_argb || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+  ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
+            dst_argb, dst_stride_argb, dst_width, dst_height,
+            0, 0, dst_width, dst_height, filtering);
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/scale_common.cc b/libvpx/libvpx/third_party/libyuv/source/scale_common.cc
new file mode 100644
index 0000000..1711f3d
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/scale_common.cc
@@ -0,0 +1,1137 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyARGB
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// CPU agnostic row functions
+void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[1];
+    dst[1] = src_ptr[3];
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[1];
+  }
+}
+
+void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[1];
+    dst[1] = src_ptr[3];
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[1];
+  }
+}
+
+void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+    dst[1] = (s[2] + s[3] + 1) >> 1;
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+  }
+}
+
+void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                              uint16* dst, int dst_width) {
+  const uint16* s = src_ptr;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+    dst[1] = (s[2] + s[3] + 1) >> 1;
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+  }
+}
+
+void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+  }
+}
+
+void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+  }
+}
+
+void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[2];
+    dst[1] = src_ptr[6];
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[2];
+  }
+}
+
+void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[2];
+    dst[1] = src_ptr[6];
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[2];
+  }
+}
+
+void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+    dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
+             src_ptr[stride + 4] + src_ptr[stride + 5] +
+             src_ptr[stride + 6] + src_ptr[stride + 7] +
+             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
+             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
+             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
+             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
+             8) >> 4;
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+  }
+}
+
+void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+    dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
+             src_ptr[stride + 4] + src_ptr[stride + 5] +
+             src_ptr[stride + 6] + src_ptr[stride + 7] +
+             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
+             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
+             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
+             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
+             8) >> 4;
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+  }
+}
+
+void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width) {
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[1];
+    dst[2] = src_ptr[3];
+    dst += 3;
+    src_ptr += 4;
+  }
+}
+
+void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width) {
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[1];
+    dst[2] = src_ptr[3];
+    dst += 3;
+    src_ptr += 4;
+  }
+}
+
+// Filter rows 0 and 1 together, 3 : 1
+void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 * 3 + b0 + 2) >> 2;
+    d[1] = (a1 * 3 + b1 + 2) >> 2;
+    d[2] = (a2 * 3 + b2 + 2) >> 2;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 * 3 + b0 + 2) >> 2;
+    d[1] = (a1 * 3 + b1 + 2) >> 2;
+    d[2] = (a2 * 3 + b2 + 2) >> 2;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+// Filter rows 1 and 2 together, 1 : 1
+void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 + b0 + 1) >> 1;
+    d[1] = (a1 + b1 + 1) >> 1;
+    d[2] = (a2 + b2 + 1) >> 1;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 + b0 + 1) >> 1;
+    d[1] = (a1 + b1 + 1) >> 1;
+    d[2] = (a2 + b2 + 1) >> 1;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                 int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[0] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr[1] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[x >> 16];
+  }
+}
+
+void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                    int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[0] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr[1] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
+                    int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[1] = dst_ptr[0] = src_ptr[0];
+    src_ptr += 1;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[0];
+  }
+}
+
+void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[1] = dst_ptr[0] = src_ptr[0];
+    src_ptr += 1;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[0];
+  }
+}
+
+// (1-f)a + fb can be replaced with a + f(b-a)
+#define BLENDER(a, b, f) (uint8)((int)(a) + \
+    ((int)(f) * ((int)(b) - (int)(a)) >> 16))
+
+void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+
+void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
+                         int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+#undef BLENDER
+
+#define BLENDER(a, b, f) (uint16)((int)(a) + \
+    ((int)(f) * ((int)(b) - (int)(a)) >> 16))
+
+void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+
+void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+#undef BLENDER
+
+void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width) {
+  int x;
+  assert(dst_width % 3 == 0);
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[3];
+    dst[2] = src_ptr[6];
+    dst += 3;
+    src_ptr += 8;
+  }
+}
+
+void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width) {
+  int x;
+  assert(dst_width % 3 == 0);
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[3];
+    dst[2] = src_ptr[6];
+    dst += 3;
+    src_ptr += 8;
+  }
+}
+
+// 8x3 -> 3x1
+void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+        (65536 / 9) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+        (65536 / 9) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7] +
+        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+        (65536 / 6) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+        (65536 / 9) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+        (65536 / 9) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7] +
+        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+        (65536 / 6) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+// 8x2 -> 3x1
+void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2]) * (65536 / 6) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5]) * (65536 / 6) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7]) *
+        (65536 / 4) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2]) * (65536 / 6) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5]) * (65536 / 6) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7]) *
+        (65536 / 4) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleAddRow_C(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
+  int x;
+  assert(src_width > 0);
+  for (x = 0; x < src_width - 1; x += 2) {
+    dst_ptr[0] += src_ptr[0];
+    dst_ptr[1] += src_ptr[1];
+    src_ptr += 2;
+    dst_ptr += 2;
+  }
+  if (src_width & 1) {
+    dst_ptr[0] += src_ptr[0];
+  }
+}
+
+void ScaleAddRow_16_C(const uint16* src_ptr, uint32* dst_ptr, int src_width) {
+  int x;
+  assert(src_width > 0);
+  for (x = 0; x < src_width - 1; x += 2) {
+    dst_ptr[0] += src_ptr[0];
+    dst_ptr[1] += src_ptr[1];
+    src_ptr += 2;
+    dst_ptr += 2;
+  }
+  if (src_width & 1) {
+    dst_ptr[0] += src_ptr[0];
+  }
+}
+
+void ScaleARGBRowDown2_C(const uint8* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8* dst_argb, int dst_width) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[1];
+    dst[1] = src[3];
+    src += 4;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[1];
+  }
+}
+
+void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] + 1) >> 1;
+    dst_argb[1] = (src_argb[1] + src_argb[5] + 1) >> 1;
+    dst_argb[2] = (src_argb[2] + src_argb[6] + 1) >> 1;
+    dst_argb[3] = (src_argb[3] + src_argb[7] + 1) >> 1;
+    src_argb += 8;
+    dst_argb += 4;
+  }
+}
+
+void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] +
+                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] +
+                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] +
+                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] +
+                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    src_argb += 8;
+    dst_argb += 4;
+  }
+}
+
+void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            int src_stepx,
+                            uint8* dst_argb, int dst_width) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[0];
+    dst[1] = src[src_stepx];
+    src += src_stepx * 2;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] +
+                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] +
+                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] +
+                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] +
+                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    src_argb += src_stepx * 4;
+    dst_argb += 4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
+                     int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[1] = dst[0] = src[0];
+    src += 1;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+// Mimics SSSE3 blender
+#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b) * f) >> 7
+#define BLENDERC(a, b, f, s) (uint32)( \
+    BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
+#define BLENDER(a, b, f) \
+    BLENDERC(a, b, f, 24) | BLENDERC(a, b, f, 16) | \
+    BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0)
+
+void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+
+void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
+                             int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+#undef BLENDER1
+#undef BLENDERC
+#undef BLENDER
+
+// Scale plane vertically with bilinear interpolation.
+void ScalePlaneVertical(int src_height,
+                        int dst_width, int dst_height,
+                        int src_stride, int dst_stride,
+                        const uint8* src_argb, uint8* dst_argb,
+                        int x, int y, int dy,
+                        int bpp, enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher bpp.
+  int dst_width_bytes = dst_width * bpp;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(bpp >= 1 && bpp <= 4);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * bpp;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width_bytes, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow(dst_argb, src_argb + yi * src_stride,
+                   src_stride, dst_width_bytes, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+void ScalePlaneVertical_16(int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint16* src_argb, uint16* dst_argb,
+                           int x, int y, int dy,
+                           int wpp, enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher wpp.
+  int dst_width_words = dst_width * wpp;
+  void (*InterpolateRow)(uint16* dst_argb, const uint16* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(wpp >= 1 && wpp <= 2);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * wpp;
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_16_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width_bytes, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow(dst_argb, src_argb + yi * src_stride,
+                   src_stride, dst_width_words, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
+// Simplify the filtering based on scale factors.
+enum FilterMode ScaleFilterReduce(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  enum FilterMode filtering) {
+  if (src_width < 0) {
+    src_width = -src_width;
+  }
+  if (src_height < 0) {
+    src_height = -src_height;
+  }
+  if (filtering == kFilterBox) {
+    // If scaling both axis to 0.5 or larger, switch from Box to Bilinear.
+    if (dst_width * 2 >= src_width && dst_height * 2 >= src_height) {
+      filtering = kFilterBilinear;
+    }
+  }
+  if (filtering == kFilterBilinear) {
+    if (src_height == 1) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to Linear.
+    if (dst_height == src_height || dst_height * 3 == src_height) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Remove 1 pixel wide filter restriction, which is to
+    // avoid reading 2 pixels horizontally that causes memory exception.
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+  }
+  if (filtering == kFilterLinear) {
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to None.
+    if (dst_width == src_width || dst_width * 3 == src_width) {
+      filtering = kFilterNone;
+    }
+  }
+  return filtering;
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_C(int num, int div) {
+  return (int)(((int64)(num) << 16) / div);
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv1_C(int num, int div) {
+  return (int)((((int64)(num) << 16) - 0x00010001) /
+                          (div - 1));
+}
+
+#define CENTERSTART(dx, s) (dx < 0) ? -((-dx >> 1) + s) : ((dx >> 1) + s)
+
+// Compute slope values for stepping.
+void ScaleSlope(int src_width, int src_height,
+                int dst_width, int dst_height,
+                enum FilterMode filtering,
+                int* x, int* y, int* dx, int* dy) {
+  assert(x != NULL);
+  assert(y != NULL);
+  assert(dx != NULL);
+  assert(dy != NULL);
+  assert(src_width != 0);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  // Check for 1 pixel and avoid FixedDiv overflow.
+  if (dst_width == 1 && src_width >= 32768) {
+    dst_width = src_width;
+  }
+  if (dst_height == 1 && src_height >= 32768) {
+    dst_height = src_height;
+  }
+  if (filtering == kFilterBox) {
+    // Scale step for point sampling duplicates all pixels equally.
+    *dx = FixedDiv(Abs(src_width), dst_width);
+    *dy = FixedDiv(src_height, dst_height);
+    *x = 0;
+    *y = 0;
+  } else if (filtering == kFilterBilinear) {
+    // Scale step for bilinear sampling renders last pixel once for upsample.
+    if (dst_width <= Abs(src_width)) {
+      *dx = FixedDiv(Abs(src_width), dst_width);
+      *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_width > 1) {
+      *dx = FixedDiv1(Abs(src_width), dst_width);
+      *x = 0;
+    }
+    if (dst_height <= src_height) {
+      *dy = FixedDiv(src_height,  dst_height);
+      *y = CENTERSTART(*dy, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_height > 1) {
+      *dy = FixedDiv1(src_height, dst_height);
+      *y = 0;
+    }
+  } else if (filtering == kFilterLinear) {
+    // Scale step for bilinear sampling renders last pixel once for upsample.
+    if (dst_width <= Abs(src_width)) {
+      *dx = FixedDiv(Abs(src_width), dst_width);
+      *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_width > 1) {
+      *dx = FixedDiv1(Abs(src_width), dst_width);
+      *x = 0;
+    }
+    *dy = FixedDiv(src_height, dst_height);
+    *y = *dy >> 1;
+  } else {
+    // Scale step for point sampling duplicates all pixels equally.
+    *dx = FixedDiv(Abs(src_width), dst_width);
+    *dy = FixedDiv(src_height, dst_height);
+    *x = CENTERSTART(*dx, 0);
+    *y = CENTERSTART(*dy, 0);
+  }
+  // Negative src_width means horizontally mirror.
+  if (src_width < 0) {
+    *x += (dst_width - 1) * *dx;
+    *dx = -*dx;
+    // src_width = -src_width;   // Caller must do this.
+  }
+}
+#undef CENTERSTART
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/scale_gcc.cc b/libvpx/libvpx/third_party/libyuv/source/scale_gcc.cc
new file mode 100644
index 0000000..8a6ac54
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/scale_gcc.cc
@@ -0,0 +1,1089 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+// Offsets for source bytes 0 to 9
+static uvec8 kShuf0 =
+  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
+static uvec8 kShuf1 =
+  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf2 =
+  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 0 to 10
+static uvec8 kShuf01 =
+  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+
+// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
+static uvec8 kShuf11 =
+  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf21 =
+  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+
+// Coefficients for source bytes 0 to 10
+static uvec8 kMadd01 =
+  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+
+// Coefficients for source bytes 10 to 21
+static uvec8 kMadd11 =
+  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+
+// Coefficients for source bytes 21 to 31
+static uvec8 kMadd21 =
+  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+
+// Coefficients for source bytes 21 to 31
+static vec16 kRound34 =
+  { 2, 2, 2, 2, 2, 2, 2, 2 };
+
+static uvec8 kShuf38a =
+  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+static uvec8 kShuf38b =
+  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 0,1,2
+static uvec8 kShufAc =
+  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 3,4,5
+static uvec8 kShufAc3 =
+  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x3 and 2x3
+static uvec16 kScaleAc33 =
+  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+
+// Arrange first value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb0 =
+  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+
+// Arrange second value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb1 =
+  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+
+// Arrange third value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb2 =
+  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x2 and 2x2
+static uvec16 kScaleAb2 =
+  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+
+// GCC versions of row functions are verbatim conversions from Visual C.
+// Generated using gcc disassembly on Visual C object file:
+// objdump -D yuvscaler.obj >yuvscaler.txt
+
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :: "memory", "cc", "xmm0", "xmm1"
+  );
+}
+
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10, 0) ",%%xmm1  \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x18,%%xmm5                    \n"
+    "pslld     $0x10,%%xmm5                    \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  intptr_t stridex3 = 0;
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0x8,%%xmm7                     \n"
+    "lea       " MEMLEA4(0x00,4,4,2) ",%3      \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x00,0,4,2,xmm2)           //  movdqu  (%0,%4,2),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,2,xmm3)           //  movdqu  0x10(%0,%4,2),%%xmm3
+    MEMOPREG(movdqu,0x00,0,3,1,xmm4)           //  movdqu  (%0,%3,1),%%xmm4
+    MEMOPREG(movdqu,0x10,0,3,1,xmm5)           //  movdqu  0x10(%0,%3,1),%%xmm5
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm4,%%xmm2                   \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm5,%%xmm3                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "pand      %%xmm7,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(dst_width),   // %2
+    "+r"(stridex3)     // %3
+  : "r"((intptr_t)(src_stride))    // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm7"
+  );
+}
+
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm3                       \n"
+    "movdqa    %1,%%xmm4                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kShuf0),  // %0
+    "m"(kShuf1),  // %1
+    "m"(kShuf2)   // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm2   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "palignr   $0x8,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm3,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "movq      %%xmm1," MEMACCESS2(0x8,1) "    \n"
+    "movq      %%xmm2," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"  // kShuf01
+    "movdqa    %1,%%xmm3                       \n"  // kShuf11
+    "movdqa    %2,%%xmm4                       \n"  // kShuf21
+  :
+  : "m"(kShuf01),  // %0
+    "m"(kShuf11),  // %1
+    "m"(kShuf21)   // %2
+  );
+  asm volatile (
+    "movdqa    %0,%%xmm5                       \n"  // kMadd01
+    "movdqa    %1,%%xmm0                       \n"  // kMadd11
+    "movdqa    %2,%%xmm1                       \n"  // kRound34
+  :
+  : "m"(kMadd01),  // %0
+    "m"(kMadd11),  // %1
+    "m"(kRound34)  // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm7)           //  movdqu  (%0,%3),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
+    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm0,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x10,0,3,1,xmm7)           //  movdqu  0x10(%0,%3),%%xmm7
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm4,%%xmm6                   \n"
+    "pmaddubsw %4,%%xmm6                       \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  : "r"((intptr_t)(src_stride)),  // %3
+    "m"(kMadd21)     // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"  // kShuf01
+    "movdqa    %1,%%xmm3                       \n"  // kShuf11
+    "movdqa    %2,%%xmm4                       \n"  // kShuf21
+  :
+  : "m"(kShuf01),  // %0
+    "m"(kShuf11),  // %1
+    "m"(kShuf21)   // %2
+  );
+  asm volatile (
+    "movdqa    %0,%%xmm5                       \n"  // kMadd01
+    "movdqa    %1,%%xmm0                       \n"  // kMadd11
+    "movdqa    %2,%%xmm1                       \n"  // kRound34
+  :
+  : "m"(kMadd01),  // %0
+    "m"(kMadd11),  // %1
+    "m"(kRound34)  // %2
+  );
+
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm7)           //  movdqu  (%0,%3,1),%%xmm7
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
+    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3,1),%%xmm7
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm0,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x10,0,3,1,xmm7)           //  movdqu  0x10(%0,%3,1),%%xmm7
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm4,%%xmm6                   \n"
+    "pmaddubsw %4,%%xmm6                       \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+    : "+r"(src_ptr),   // %0
+      "+r"(dst_ptr),   // %1
+      "+r"(dst_width)  // %2
+    : "r"((intptr_t)(src_stride)),  // %3
+      "m"(kMadd21)     // %4
+    : "memory", "cc", NACL_R14
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "movhlps   %%xmm0,%%xmm1                   \n"
+    "movd      %%xmm1," MEMACCESS2(0x8,1) "    \n"
+    "lea       " MEMLEA(0xc,1) ",%1            \n"
+    "sub       $0xc,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  : "m"(kShuf38a),   // %3
+    "m"(kShuf38b)    // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5"
+  );
+}
+
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm4                       \n"
+    "movdqa    %3,%%xmm5                       \n"
+  :
+  : "m"(kShufAb0),   // %0
+    "m"(kShufAb1),   // %1
+    "m"(kShufAb2),   // %2
+    "m"(kScaleAb2)   // %3
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm1)           //  movdqu  (%0,%3,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pshufb    %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "paddusw   %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "paddusw   %%xmm0,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movd      %%xmm1," MEMACCESS(1) "         \n"
+    "psrlq     $0x10,%%xmm1                    \n"
+    "movd      %%xmm1," MEMACCESS2(0x2,1) "    \n"
+    "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "sub       $0x6,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(dst_width)    // %2
+  : "r"((intptr_t)(src_stride))  // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm4                       \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+  :
+  : "m"(kShufAc),    // %0
+    "m"(kShufAc3),   // %1
+    "m"(kScaleAc33)  // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm6)           //  movdqu  (%0,%3,1),%%xmm6
+    "movhlps   %%xmm0,%%xmm1                   \n"
+    "movhlps   %%xmm6,%%xmm7                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm6                   \n"
+    "punpcklbw %%xmm5,%%xmm7                   \n"
+    "paddusw   %%xmm6,%%xmm0                   \n"
+    "paddusw   %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x00,0,3,2,xmm6)           //  movdqu  (%0,%3,2),%%xmm6
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movhlps   %%xmm6,%%xmm7                   \n"
+    "punpcklbw %%xmm5,%%xmm6                   \n"
+    "punpcklbw %%xmm5,%%xmm7                   \n"
+    "paddusw   %%xmm6,%%xmm0                   \n"
+    "paddusw   %%xmm7,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "psrldq    $0x2,%%xmm0                     \n"
+    "paddusw   %%xmm0,%%xmm6                   \n"
+    "psrldq    $0x2,%%xmm0                     \n"
+    "paddusw   %%xmm0,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "movdqa    %%xmm1,%%xmm7                   \n"
+    "psrldq    $0x2,%%xmm1                     \n"
+    "paddusw   %%xmm1,%%xmm7                   \n"
+    "psrldq    $0x2,%%xmm1                     \n"
+    "paddusw   %%xmm1,%%xmm7                   \n"
+    "pshufb    %%xmm3,%%xmm7                   \n"
+    "paddusw   %%xmm7,%%xmm6                   \n"
+    "pmulhuw   %%xmm4,%%xmm6                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movd      %%xmm6," MEMACCESS(1) "         \n"
+    "psrlq     $0x10,%%xmm6                    \n"
+    "movd      %%xmm6," MEMACCESS2(0x2,1) "    \n"
+    "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "sub       $0x6,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+// Reads 16xN bytes and produces 16 shorts at a time.
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width, int src_height) {
+  int tmp_height = 0;
+  intptr_t tmp_src = 0;
+  asm volatile (
+    "mov       %0,%3                           \n"  // row pointer
+    "mov       %5,%2                           \n"  // height
+    "pxor      %%xmm0,%%xmm0                   \n"  // clear accumulators
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(3) ",%%xmm2         \n"
+    "add       %6,%3                           \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "paddusw   %%xmm2,%%xmm0                   \n"
+    "paddusw   %%xmm3,%%xmm1                   \n"
+    "sub       $0x1,%2                         \n"
+    "jg        1b                              \n"
+
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"  // src_ptr += 16
+    "mov       %0,%3                           \n"  // row pointer
+    "mov       %5,%2                           \n"  // height
+    "pxor      %%xmm0,%%xmm0                   \n"  // clear accumulators
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "sub       $0x10,%4                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(tmp_height),  // %2
+    "+r"(tmp_src),     // %3
+    "+r"(src_width),   // %4
+    "+rm"(src_height)  // %5
+  : "rm"((intptr_t)(src_stride))  // %6
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+
+// Bilinear column filtering. SSSE3 version.
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0, temp_pixel = 0;
+  asm volatile (
+    "movd      %6,%%xmm2                       \n"
+    "movd      %7,%%xmm3                       \n"
+    "movl      $0x04040000,%k2                 \n"
+    "movd      %k2,%%xmm5                      \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x9,%%xmm6                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "subl      $0x2,%5                         \n"
+    "jl        29f                             \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "punpckldq %%xmm0,%%xmm2                   \n"
+    "punpckldq %%xmm3,%%xmm3                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
+    "movd      %k2,%%xmm0                      \n"
+    "psrlw     $0x9,%%xmm1                     \n"
+    MEMOPARG(movzwl,0x00,1,4,1,k2)             //  movzwl  (%1,%4,1),%k2
+    "movd      %k2,%%xmm4                      \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "punpcklwd %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm1                   \n"
+    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0,%k2                      \n"
+    "mov       %w2," MEMACCESS(0) "            \n"
+    "lea       " MEMLEA(0x2,0) ",%0            \n"
+    "sub       $0x2,%5                         \n"
+    "jge       2b                              \n"
+
+    LABELALIGN
+  "29:                                         \n"
+    "addl      $0x1,%5                         \n"
+    "jl        99f                             \n"
+    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
+    "movd      %k2,%%xmm0                      \n"
+    "psrlw     $0x9,%%xmm2                     \n"
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "pxor      %%xmm6,%%xmm2                   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0,%k2                      \n"
+    "mov       %b2," MEMACCESS(0) "            \n"
+  "99:                                         \n"
+  : "+r"(dst_ptr),     // %0
+    "+r"(src_ptr),     // %1
+    "+a"(temp_pixel),  // %2
+    "+r"(x0),          // %3
+    "+r"(x1),          // %4
+    "+rm"(dst_width)   // %5
+  : "rm"(x),           // %6
+    "rm"(dx)           // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub       $0x20,%2                         \n"
+    "jg        1b                              \n"
+
+  : "+r"(dst_ptr),     // %0
+    "+r"(src_ptr),     // %1
+    "+r"(dst_width)    // %2
+  :: "memory", "cc", "xmm0", "xmm1"
+  );
+}
+
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "shufps    $0xdd,%%xmm1,%%xmm0             \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(dst_width)  // %2
+  :: "memory", "cc", "xmm0", "xmm1"
+  );
+}
+
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(dst_width)  // %2
+  :: "memory", "cc", "xmm0", "xmm1"
+  );
+}
+
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu   (%0,%3,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu   0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: dst_argb 16 byte aligned.
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width) {
+  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
+  intptr_t src_stepx_x12 = 0;
+  asm volatile (
+    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
+    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movd      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,0,1,2,xmm2)             //  movd      (%0,%1,2),%%xmm2
+    MEMOPREG(movd,0x00,0,4,1,xmm3)             //  movd      (%0,%4,1),%%xmm3
+    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
+    "punpckldq %%xmm3,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm0                  \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),      // %0
+    "+r"(src_stepx_x4),  // %1
+    "+r"(dst_argb),      // %2
+    "+r"(dst_width),     // %3
+    "+r"(src_stepx_x12)  // %4
+  :: "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
+  );
+}
+
+// Blends four 2x2 to 4x1.
+// Alignment requirement: dst_argb 16 byte aligned.
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride, int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
+  intptr_t src_stepx_x12 = 0;
+  intptr_t row1 = (intptr_t)(src_stride);
+  asm volatile (
+    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
+    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
+    "lea       " MEMLEA4(0x00,0,5,1) ",%5      \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movhps,0x00,0,1,1,xmm0)           //  movhps    (%0,%1,1),%%xmm0
+    MEMOPREG(movq,0x00,0,1,2,xmm1)             //  movq      (%0,%1,2),%%xmm1
+    MEMOPREG(movhps,0x00,0,4,1,xmm1)           //  movhps    (%0,%4,1),%%xmm1
+    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
+    "movq      " MEMACCESS(5) ",%%xmm2         \n"
+    MEMOPREG(movhps,0x00,5,1,1,xmm2)           //  movhps    (%5,%1,1),%%xmm2
+    MEMOPREG(movq,0x00,5,1,2,xmm3)             //  movq      (%5,%1,2),%%xmm3
+    MEMOPREG(movhps,0x00,5,4,1,xmm3)           //  movhps    (%5,%4,1),%%xmm3
+    "lea       " MEMLEA4(0x00,5,1,4) ",%5      \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),       // %0
+    "+r"(src_stepx_x4),   // %1
+    "+r"(dst_argb),       // %2
+    "+rm"(dst_width),     // %3
+    "+r"(src_stepx_x12),  // %4
+    "+r"(row1)            // %5
+  :: "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
+  );
+}
+
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0;
+  asm volatile (
+    "movd      %5,%%xmm2                       \n"
+    "movd      %6,%%xmm3                       \n"
+    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    "pshufd    $0x11,%%xmm3,%%xmm0             \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pshufd    $0x5,%%xmm3,%%xmm0              \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pextrw    $0x1,%%xmm2,%k0                 \n"
+    "pextrw    $0x3,%%xmm2,%k1                 \n"
+    "cmp       $0x0,%4                         \n"
+    "jl        99f                             \n"
+    "sub       $0x4,%4                         \n"
+    "jl        49f                             \n"
+
+    LABELALIGN
+  "40:                                         \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
+    "pextrw    $0x5,%%xmm2,%k0                 \n"
+    "pextrw    $0x7,%%xmm2,%k1                 \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm1)             //  movd      (%3,%0,4),%%xmm1
+    MEMOPREG(movd,0x00,3,1,4,xmm4)             //  movd      (%3,%1,4),%%xmm4
+    "pextrw    $0x1,%%xmm2,%k0                 \n"
+    "pextrw    $0x3,%%xmm2,%k1                 \n"
+    "punpckldq %%xmm4,%%xmm1                   \n"
+    "punpcklqdq %%xmm1,%%xmm0                  \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%4                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "test      $0x2,%4                         \n"
+    "je        29f                             \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
+    "pextrw    $0x5,%%xmm2,%k0                 \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x8,2) ",%2            \n"
+  "29:                                         \n"
+    "test      $0x1,%4                         \n"
+    "je        99f                             \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+  "99:                                         \n"
+  : "+a"(x0),          // %0
+    "+d"(x1),          // %1
+    "+r"(dst_argb),    // %2
+    "+r"(src_argb),    // %3
+    "+r"(dst_width)    // %4
+  : "rm"(x),           // %5
+    "rm"(dx)           // %6
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpckldq %%xmm0,%%xmm0                   \n"
+    "punpckhdq %%xmm1,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+
+  : "+r"(dst_argb),    // %0
+    "+r"(src_argb),    // %1
+    "+r"(dst_width)    // %2
+  :: "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
+  );
+}
+
+// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
+static uvec8 kShuffleColARGB = {
+  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
+  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+};
+
+// Shuffle table for duplicating 2 fractions into 8 bytes each
+static uvec8 kShuffleFractions = {
+  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+};
+
+// Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0;
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm5                       \n"
+  :
+  : "m"(kShuffleColARGB),  // %0
+    "m"(kShuffleFractions)  // %1
+  );
+
+  asm volatile (
+    "movd      %5,%%xmm2                       \n"
+    "movd      %6,%%xmm3                       \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x9,%%xmm6                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "sub       $0x2,%2                         \n"
+    "jl        29f                             \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "punpckldq %%xmm0,%%xmm2                   \n"
+    "punpckldq %%xmm3,%%xmm3                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
+    "psrlw     $0x9,%%xmm1                     \n"
+    MEMOPREG(movhps,0x00,1,4,4,xmm0)           //  movhps    (%1,%4,4),%%xmm0
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm1                   \n"
+    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(0) "         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x2,%2                         \n"
+    "jge       2b                              \n"
+
+    LABELALIGN
+  "29:                                         \n"
+    "add       $0x1,%2                         \n"
+    "jl        99f                             \n"
+    "psrlw     $0x9,%%xmm2                     \n"
+    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm2                   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(0) "         \n"
+
+    LABELALIGN
+  "99:                                         \n"
+  : "+r"(dst_argb),    // %0
+    "+r"(src_argb),    // %1
+    "+rm"(dst_width),  // %2
+    "+r"(x0),          // %3
+    "+r"(x1)           // %4
+  : "rm"(x),           // %5
+    "rm"(dx)           // %6
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_X86(int num, int div) {
+  asm volatile (
+    "cdq                                       \n"
+    "shld      $0x10,%%eax,%%edx               \n"
+    "shl       $0x10,%%eax                     \n"
+    "idiv      %1                              \n"
+    "mov       %0, %%eax                       \n"
+    : "+a"(num)  // %0
+    : "c"(div)   // %1
+    : "memory", "cc", "edx"
+  );
+  return num;
+}
+
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
+int FixedDiv1_X86(int num, int div) {
+  asm volatile (
+    "cdq                                       \n"
+    "shld      $0x10,%%eax,%%edx               \n"
+    "shl       $0x10,%%eax                     \n"
+    "sub       $0x10001,%%eax                  \n"
+    "sbb       $0x0,%%edx                      \n"
+    "sub       $0x1,%1                         \n"
+    "idiv      %1                              \n"
+    "mov       %0, %%eax                       \n"
+    : "+a"(num)  // %0
+    : "c"(div)   // %1
+    : "memory", "cc", "edx"
+  );
+  return num;
+}
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/scale_mips.cc b/libvpx/libvpx/third_party/libyuv/source/scale_mips.cc
new file mode 100644
index 0000000..3eb4f27
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/scale_mips.cc
@@ -0,0 +1,654 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC MIPS DSPR2
+#if !defined(LIBYUV_DISABLE_MIPS) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width) {
+  __asm__ __volatile__(
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+
+    "srl            $t9, %[dst_width], 4           \n"  // iterations -> by 16
+    "beqz           $t9, 2f                        \n"
+    " nop                                          \n"
+
+    ".p2align       2                              \n"
+  "1:                                              \n"
+    "lw             $t0, 0(%[src_ptr])             \n"  // |3|2|1|0|
+    "lw             $t1, 4(%[src_ptr])             \n"  // |7|6|5|4|
+    "lw             $t2, 8(%[src_ptr])             \n"  // |11|10|9|8|
+    "lw             $t3, 12(%[src_ptr])            \n"  // |15|14|13|12|
+    "lw             $t4, 16(%[src_ptr])            \n"  // |19|18|17|16|
+    "lw             $t5, 20(%[src_ptr])            \n"  // |23|22|21|20|
+    "lw             $t6, 24(%[src_ptr])            \n"  // |27|26|25|24|
+    "lw             $t7, 28(%[src_ptr])            \n"  // |31|30|29|28|
+    // TODO(fbarchard): Use odd pixels instead of even.
+    "precr.qb.ph    $t8, $t1, $t0                  \n"  // |6|4|2|0|
+    "precr.qb.ph    $t0, $t3, $t2                  \n"  // |14|12|10|8|
+    "precr.qb.ph    $t1, $t5, $t4                  \n"  // |22|20|18|16|
+    "precr.qb.ph    $t2, $t7, $t6                  \n"  // |30|28|26|24|
+    "addiu          %[src_ptr], %[src_ptr], 32     \n"
+    "addiu          $t9, $t9, -1                   \n"
+    "sw             $t8, 0(%[dst])                 \n"
+    "sw             $t0, 4(%[dst])                 \n"
+    "sw             $t1, 8(%[dst])                 \n"
+    "sw             $t2, 12(%[dst])                \n"
+    "bgtz           $t9, 1b                        \n"
+    " addiu         %[dst], %[dst], 16             \n"
+
+  "2:                                              \n"
+    "andi           $t9, %[dst_width], 0xf         \n"  // residue
+    "beqz           $t9, 3f                        \n"
+    " nop                                          \n"
+
+  "21:                                             \n"
+    "lbu            $t0, 0(%[src_ptr])             \n"
+    "addiu          %[src_ptr], %[src_ptr], 2      \n"
+    "addiu          $t9, $t9, -1                   \n"
+    "sb             $t0, 0(%[dst])                 \n"
+    "bgtz           $t9, 21b                       \n"
+    " addiu         %[dst], %[dst], 1              \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+  : [src_ptr] "+r" (src_ptr),
+    [dst] "+r" (dst)
+  : [dst_width] "r" (dst_width)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width) {
+  const uint8* t = src_ptr + src_stride;
+
+  __asm__ __volatile__ (
+    ".set push                                    \n"
+    ".set noreorder                               \n"
+
+    "srl            $t9, %[dst_width], 3          \n"  // iterations -> step 8
+    "bltz           $t9, 2f                       \n"
+    " nop                                         \n"
+
+    ".p2align       2                             \n"
+  "1:                                             \n"
+    "lw             $t0, 0(%[src_ptr])            \n"  // |3|2|1|0|
+    "lw             $t1, 4(%[src_ptr])            \n"  // |7|6|5|4|
+    "lw             $t2, 8(%[src_ptr])            \n"  // |11|10|9|8|
+    "lw             $t3, 12(%[src_ptr])           \n"  // |15|14|13|12|
+    "lw             $t4, 0(%[t])                  \n"  // |19|18|17|16|
+    "lw             $t5, 4(%[t])                  \n"  // |23|22|21|20|
+    "lw             $t6, 8(%[t])                  \n"  // |27|26|25|24|
+    "lw             $t7, 12(%[t])                 \n"  // |31|30|29|28|
+    "addiu          $t9, $t9, -1                  \n"
+    "srl            $t8, $t0, 16                  \n"  // |X|X|3|2|
+    "ins            $t0, $t4, 16, 16              \n"  // |17|16|1|0|
+    "ins            $t4, $t8, 0, 16               \n"  // |19|18|3|2|
+    "raddu.w.qb     $t0, $t0                      \n"  // |17+16+1+0|
+    "raddu.w.qb     $t4, $t4                      \n"  // |19+18+3+2|
+    "shra_r.w       $t0, $t0, 2                   \n"  // |t0+2|>>2
+    "shra_r.w       $t4, $t4, 2                   \n"  // |t4+2|>>2
+    "srl            $t8, $t1, 16                  \n"  // |X|X|7|6|
+    "ins            $t1, $t5, 16, 16              \n"  // |21|20|5|4|
+    "ins            $t5, $t8, 0, 16               \n"  // |22|23|7|6|
+    "raddu.w.qb     $t1, $t1                      \n"  // |21+20+5+4|
+    "raddu.w.qb     $t5, $t5                      \n"  // |23+22+7+6|
+    "shra_r.w       $t1, $t1, 2                   \n"  // |t1+2|>>2
+    "shra_r.w       $t5, $t5, 2                   \n"  // |t5+2|>>2
+    "srl            $t8, $t2, 16                  \n"  // |X|X|11|10|
+    "ins            $t2, $t6, 16, 16              \n"  // |25|24|9|8|
+    "ins            $t6, $t8, 0, 16               \n"  // |27|26|11|10|
+    "raddu.w.qb     $t2, $t2                      \n"  // |25+24+9+8|
+    "raddu.w.qb     $t6, $t6                      \n"  // |27+26+11+10|
+    "shra_r.w       $t2, $t2, 2                   \n"  // |t2+2|>>2
+    "shra_r.w       $t6, $t6, 2                   \n"  // |t5+2|>>2
+    "srl            $t8, $t3, 16                  \n"  // |X|X|15|14|
+    "ins            $t3, $t7, 16, 16              \n"  // |29|28|13|12|
+    "ins            $t7, $t8, 0, 16               \n"  // |31|30|15|14|
+    "raddu.w.qb     $t3, $t3                      \n"  // |29+28+13+12|
+    "raddu.w.qb     $t7, $t7                      \n"  // |31+30+15+14|
+    "shra_r.w       $t3, $t3, 2                   \n"  // |t3+2|>>2
+    "shra_r.w       $t7, $t7, 2                   \n"  // |t7+2|>>2
+    "addiu          %[src_ptr], %[src_ptr], 16    \n"
+    "addiu          %[t], %[t], 16                \n"
+    "sb             $t0, 0(%[dst])                \n"
+    "sb             $t4, 1(%[dst])                \n"
+    "sb             $t1, 2(%[dst])                \n"
+    "sb             $t5, 3(%[dst])                \n"
+    "sb             $t2, 4(%[dst])                \n"
+    "sb             $t6, 5(%[dst])                \n"
+    "sb             $t3, 6(%[dst])                \n"
+    "sb             $t7, 7(%[dst])                \n"
+    "bgtz           $t9, 1b                       \n"
+    " addiu         %[dst], %[dst], 8             \n"
+
+  "2:                                             \n"
+    "andi           $t9, %[dst_width], 0x7        \n"  // x = residue
+    "beqz           $t9, 3f                       \n"
+    " nop                                         \n"
+
+    "21:                                          \n"
+    "lwr            $t1, 0(%[src_ptr])            \n"
+    "lwl            $t1, 3(%[src_ptr])            \n"
+    "lwr            $t2, 0(%[t])                  \n"
+    "lwl            $t2, 3(%[t])                  \n"
+    "srl            $t8, $t1, 16                  \n"
+    "ins            $t1, $t2, 16, 16              \n"
+    "ins            $t2, $t8, 0, 16               \n"
+    "raddu.w.qb     $t1, $t1                      \n"
+    "raddu.w.qb     $t2, $t2                      \n"
+    "shra_r.w       $t1, $t1, 2                   \n"
+    "shra_r.w       $t2, $t2, 2                   \n"
+    "sb             $t1, 0(%[dst])                \n"
+    "sb             $t2, 1(%[dst])                \n"
+    "addiu          %[src_ptr], %[src_ptr], 4     \n"
+    "addiu          $t9, $t9, -2                  \n"
+    "addiu          %[t], %[t], 4                 \n"
+    "bgtz           $t9, 21b                      \n"
+    " addiu         %[dst], %[dst], 2             \n"
+
+  "3:                                             \n"
+    ".set pop                                     \n"
+
+  : [src_ptr] "+r" (src_ptr),
+    [dst] "+r" (dst), [t] "+r" (t)
+  : [dst_width] "r" (dst_width)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                    \n"
+      ".set noreorder                               \n"
+
+      "srl            $t9, %[dst_width], 3          \n"
+      "beqz           $t9, 2f                       \n"
+      " nop                                         \n"
+
+      ".p2align       2                             \n"
+     "1:                                            \n"
+      "lw             $t1, 0(%[src_ptr])            \n"  // |3|2|1|0|
+      "lw             $t2, 4(%[src_ptr])            \n"  // |7|6|5|4|
+      "lw             $t3, 8(%[src_ptr])            \n"  // |11|10|9|8|
+      "lw             $t4, 12(%[src_ptr])           \n"  // |15|14|13|12|
+      "lw             $t5, 16(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw             $t6, 20(%[src_ptr])           \n"  // |23|22|21|20|
+      "lw             $t7, 24(%[src_ptr])           \n"  // |27|26|25|24|
+      "lw             $t8, 28(%[src_ptr])           \n"  // |31|30|29|28|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |6|4|2|0|
+      "precr.qb.ph    $t2, $t4, $t3                 \n"  // |14|12|10|8|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |22|20|18|16|
+      "precr.qb.ph    $t6, $t8, $t7                 \n"  // |30|28|26|24|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |12|8|4|0|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |28|24|20|16|
+      "addiu          %[src_ptr], %[src_ptr], 32    \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sw             $t1, 0(%[dst])                \n"
+      "sw             $t5, 4(%[dst])                \n"
+      "bgtz           $t9, 1b                       \n"
+      " addiu         %[dst], %[dst], 8             \n"
+
+    "2:                                             \n"
+      "andi           $t9, %[dst_width], 7          \n"  // residue
+      "beqz           $t9, 3f                       \n"
+      " nop                                         \n"
+
+    "21:                                            \n"
+      "lbu            $t1, 0(%[src_ptr])            \n"
+      "addiu          %[src_ptr], %[src_ptr], 4     \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sb             $t1, 0(%[dst])                \n"
+      "bgtz           $t9, 21b                      \n"
+      " addiu         %[dst], %[dst], 1             \n"
+
+    "3:                                             \n"
+      ".set pop                                     \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst)
+      : [dst_width] "r" (dst_width)
+      : "t1", "t2", "t3", "t4", "t5",
+        "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  const uint8* s2 = s1 + stride;
+  const uint8* s3 = s2 + stride;
+
+  __asm__ __volatile__ (
+      ".set push                                  \n"
+      ".set noreorder                             \n"
+
+      "srl           $t9, %[dst_width], 1         \n"
+      "andi          $t8, %[dst_width], 1         \n"
+
+      ".p2align      2                            \n"
+     "1:                                          \n"
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "lw            $t4, 4(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw            $t5, 4(%[s1])                \n"  // |23|22|21|20|
+      "lw            $t6, 4(%[s2])                \n"  // |27|26|25|24|
+      "lw            $t7, 4(%[s3])                \n"  // |31|30|29|28|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "raddu.w.qb    $t4, $t4                     \n"  // |19 + 18 + 17 + 16|
+      "raddu.w.qb    $t5, $t5                     \n"  // |23 + 22 + 21 + 20|
+      "raddu.w.qb    $t6, $t6                     \n"  // |27 + 26 + 25 + 24|
+      "raddu.w.qb    $t7, $t7                     \n"  // |31 + 30 + 29 + 28|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "add           $t4, $t4, $t5                \n"
+      "add           $t6, $t6, $t7                \n"
+      "add           $t4, $t4, $t6                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "shra_r.w      $t4, $t4, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+      "sb            $t4, 1(%[dst])               \n"
+      "addiu         %[src_ptr], %[src_ptr], 8    \n"
+      "addiu         %[s1], %[s1], 8              \n"
+      "addiu         %[s2], %[s2], 8              \n"
+      "addiu         %[s3], %[s3], 8              \n"
+      "addiu         $t9, $t9, -1                 \n"
+      "bgtz          $t9, 1b                      \n"
+      " addiu        %[dst], %[dst], 2            \n"
+      "beqz          $t8, 2f                      \n"
+      " nop                                       \n"
+
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+
+      "2:                                         \n"
+      ".set pop                                   \n"
+
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [s1] "+r" (s1),
+        [s2] "+r" (s2),
+        [s3] "+r" (s3)
+      : [dst_width] "r" (dst_width)
+      : "t0", "t1", "t2", "t3", "t4", "t5",
+        "t6","t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                          \n"
+      ".set noreorder                                     \n"
+      ".p2align        2                                  \n"
+    "1:                                                   \n"
+      "lw              $t1, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw              $t2, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw              $t3, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw              $t4, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw              $t5, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw              $t6, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw              $t7, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw              $t8, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "precrq.qb.ph    $t0, $t2, $t4                      \n"  // |7|5|15|13|
+      "precrq.qb.ph    $t9, $t6, $t8                      \n"  // |23|21|31|30|
+      "addiu           %[dst_width], %[dst_width], -24    \n"
+      "ins             $t1, $t1, 8, 16                    \n"  // |3|1|0|X|
+      "ins             $t4, $t0, 8, 16                    \n"  // |X|15|13|12|
+      "ins             $t5, $t5, 8, 16                    \n"  // |19|17|16|X|
+      "ins             $t8, $t9, 8, 16                    \n"  // |X|31|29|28|
+      "addiu           %[src_ptr], %[src_ptr], 32         \n"
+      "packrl.ph       $t0, $t3, $t0                      \n"  // |9|8|7|5|
+      "packrl.ph       $t9, $t7, $t9                      \n"  // |25|24|23|21|
+      "prepend         $t1, $t2, 8                        \n"  // |4|3|1|0|
+      "prepend         $t3, $t4, 24                       \n"  // |15|13|12|11|
+      "prepend         $t5, $t6, 8                        \n"  // |20|19|17|16|
+      "prepend         $t7, $t8, 24                       \n"  // |31|29|28|27|
+      "sw              $t1, 0(%[dst])                     \n"
+      "sw              $t0, 4(%[dst])                     \n"
+      "sw              $t3, 8(%[dst])                     \n"
+      "sw              $t5, 12(%[dst])                    \n"
+      "sw              $t9, 16(%[dst])                    \n"
+      "sw              $t7, 20(%[dst])                    \n"
+      "bnez            %[dst_width], 1b                   \n"
+      " addiu          %[dst], %[dst], 24                 \n"
+      ".set pop                                           \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4", "t5",
+        "t6","t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "repl.ph           $t3, 3                          \n"  // 0x00030003
+
+     ".p2align           2                               \n"
+    "1:                                                  \n"
+      "lw                $t0, 0(%[src_ptr])              \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])  \n"  // |T3|T2|T1|T0|
+      "rotr              $t2, $t0, 8                     \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                     \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t4, $t2, $t3                   \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t3                   \n"  // |T0*3|T3*3|
+      "andi              $t0, $t2, 0xFFFF                \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                        \n"
+      "raddu.w.qb        $t1, $t1                        \n"
+      "shra_r.w          $t0, $t0, 1                     \n"
+      "shra_r.w          $t1, $t1, 1                     \n"
+      "preceu.ph.qbr     $t2, $t2                        \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                        \n"  // |0|T2|0|T1|
+      "rotr              $t2, $t2, 16                    \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                    \n"  // |0|T1|0|T2|
+      "addu.ph           $t2, $t2, $t4                   \n"
+      "addu.ph           $t6, $t6, $t5                   \n"
+      "sll               $t5, $t0, 1                     \n"
+      "add               $t0, $t5, $t0                   \n"
+      "shra_r.ph         $t2, $t2, 2                     \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "shll.ph           $t4, $t2, 1                     \n"
+      "addq.ph           $t4, $t4, $t2                   \n"
+      "addu              $t0, $t0, $t1                   \n"
+      "addiu             %[src_ptr], %[src_ptr], 4       \n"
+      "shra_r.w          $t0, $t0, 2                     \n"
+      "addu.ph           $t6, $t6, $t4                   \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "srl               $t1, $t6, 16                    \n"
+      "addiu             %[dst_width], %[dst_width], -3  \n"
+      "sb                $t1, 0(%[d])                    \n"
+      "sb                $t0, 1(%[d])                    \n"
+      "sb                $t6, 2(%[d])                    \n"
+      "bgtz              %[dst_width], 1b                \n"
+      " addiu            %[d], %[d], 3                   \n"
+    "3:                                                  \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [src_stride] "+r" (src_stride),
+        [d] "+r" (d),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3",
+        "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                           \n"
+      ".set noreorder                                      \n"
+      "repl.ph           $t2, 3                            \n"  // 0x00030003
+
+      ".p2align          2                                 \n"
+    "1:                                                    \n"
+      "lw                $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])    \n"  // |T3|T2|T1|T0|
+      "rotr              $t4, $t0, 8                       \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                       \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t3, $t4, $t2                     \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t2                     \n"  // |T0*3|T3*3|
+      "andi              $t0, $t4, 0xFFFF                  \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                  \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                          \n"
+      "raddu.w.qb        $t1, $t1                          \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "shra_r.w          $t1, $t1, 1                       \n"
+      "preceu.ph.qbr     $t4, $t4                          \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                          \n"  // |0|T2|0|T1|
+      "rotr              $t4, $t4, 16                      \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                      \n"  // |0|T1|0|T2|
+      "addu.ph           $t4, $t4, $t3                     \n"
+      "addu.ph           $t6, $t6, $t5                     \n"
+      "shra_r.ph         $t6, $t6, 2                       \n"
+      "shra_r.ph         $t4, $t4, 2                       \n"
+      "addu.ph           $t6, $t6, $t4                     \n"
+      "addiu             %[src_ptr], %[src_ptr], 4         \n"
+      "shra_r.ph         $t6, $t6, 1                       \n"
+      "addu              $t0, $t0, $t1                     \n"
+      "addiu             %[dst_width], %[dst_width], -3    \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "srl               $t1, $t6, 16                      \n"
+      "sb                $t1, 0(%[d])                      \n"
+      "sb                $t0, 1(%[d])                      \n"
+      "sb                $t6, 2(%[d])                      \n"
+      "bgtz              %[dst_width], 1b                  \n"
+      " addiu            %[d], %[d], 3                     \n"
+    "3:                                                    \n"
+      ".set pop                                            \n"
+      : [src_ptr] "+r" (src_ptr),
+        [src_stride] "+r" (src_stride),
+        [d] "+r" (d),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3",
+        "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                     \n"
+      ".set noreorder                                \n"
+
+      ".p2align   2                                  \n"
+    "1:                                              \n"
+      "lw         $t0, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw         $t1, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw         $t2, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw         $t3, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw         $t4, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw         $t5, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw         $t6, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw         $t7, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "wsbh       $t0, $t0                           \n"  // |2|3|0|1|
+      "wsbh       $t6, $t6                           \n"  // |26|27|24|25|
+      "srl        $t0, $t0, 8                        \n"  // |X|2|3|0|
+      "srl        $t3, $t3, 16                       \n"  // |X|X|15|14|
+      "srl        $t5, $t5, 16                       \n"  // |X|X|23|22|
+      "srl        $t7, $t7, 16                       \n"  // |X|X|31|30|
+      "ins        $t1, $t2, 24, 8                    \n"  // |8|6|5|4|
+      "ins        $t6, $t5, 0, 8                     \n"  // |26|27|24|22|
+      "ins        $t1, $t0, 0, 16                    \n"  // |8|6|3|0|
+      "ins        $t6, $t7, 24, 8                    \n"  // |30|27|24|22|
+      "prepend    $t2, $t3, 24                       \n"  // |X|15|14|11|
+      "ins        $t4, $t4, 16, 8                    \n"  // |19|16|17|X|
+      "ins        $t4, $t2, 0, 16                    \n"  // |19|16|14|11|
+      "addiu      %[src_ptr], %[src_ptr], 32         \n"
+      "addiu      %[dst_width], %[dst_width], -12    \n"
+      "addiu      $t8,%[dst_width], -12              \n"
+      "sw         $t1, 0(%[dst])                     \n"
+      "sw         $t4, 4(%[dst])                     \n"
+      "sw         $t6, 8(%[dst])                     \n"
+      "bgez       $t8, 1b                            \n"
+      " addiu     %[dst], %[dst], 12                 \n"
+      ".set pop                                      \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4",
+        "t5", "t6", "t7", "t8"
+  );
+}
+
+void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* t = src_ptr + stride;
+  const int c = 0x2AAA;
+
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      ".p2align        2                                 \n"
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[t])                      \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[t])                      \n"  // |T7|T6|T5|T4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t4, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "packrl.ph       $t5, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t4, $t4                          \n"  // S7+S6+T7+T6
+      "raddu.w.qb      $t5, $t5                          \n"  // T5+T4+S5+S4
+      "precrq.qb.ph    $t6, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t6, $t6, $t6                     \n"  // |S3|T3|S3|T3|
+      "srl             $t4, $t4, 2                       \n"  // t4 / 4
+      "srl             $t6, $t6, 16                      \n"  // |0|0|S3|T3|
+      "raddu.w.qb      $t6, $t6                          \n"  // 0+0+S3+T3
+      "addu            $t6, $t5, $t6                     \n"
+      "mul             $t6, $t6, %[c]                    \n"  // t6 * 0x2AAA
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "raddu.w.qb      $t0, $t0                          \n"  // S2+S1+S0+0
+      "raddu.w.qb      $t2, $t2                          \n"  // T2+T1+T0+0
+      "addu            $t0, $t0, $t2                     \n"
+      "mul             $t0, $t0, %[c]                    \n"  // t0 * 0x2AAA
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[t], %[t], 8                     \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t4, -1(%[dst_ptr])               \n"
+      "sb              $t6, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst_ptr] "+r" (dst_ptr),
+        [t] "+r" (t),
+        [dst_width] "+r" (dst_width)
+      : [c] "r" (c)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  stride += stride;
+  const uint8* s2 = src_ptr + stride;
+  const int c1 = 0x1C71;
+  const int c2 = 0x2AAA;
+
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      ".p2align        2                                 \n"
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[s1])                     \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[s1])                     \n"  // |T7|T6|T5|T4|
+      "lw              $t4, 0(%[s2])                     \n"  // |R3|R2|R1|R0|
+      "lw              $t5, 4(%[s2])                     \n"  // |R7|R6|R5|R4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t6, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "raddu.w.qb      $t6, $t6                          \n"  // S7+S6+T7+T6
+      "packrl.ph       $t7, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t7, $t7                          \n"  // T5+T4+S5+S4
+      "sll             $t8, $t5, 16                      \n"  // |R5|R4|0|0|
+      "raddu.w.qb      $t8, $t8                          \n"  // R5+R4
+      "addu            $t7, $t7, $t8                     \n"
+      "srl             $t8, $t5, 16                      \n"  // |0|0|R7|R6|
+      "raddu.w.qb      $t8, $t8                          \n"  // R7 + R6
+      "addu            $t6, $t6, $t8                     \n"
+      "mul             $t6, $t6, %[c2]                   \n"  // t6 * 0x2AAA
+      "precrq.qb.ph    $t8, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t8, $t8, $t4                     \n"  // |S3|T3|R3|R1|
+      "srl             $t8, $t8, 8                       \n"  // |0|S3|T3|R3|
+      "raddu.w.qb      $t8, $t8                          \n"  // S3 + T3 + R3
+      "addu            $t7, $t7, $t8                     \n"
+      "mul             $t7, $t7, %[c1]                   \n"  // t7 * 0x1C71
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "sll             $t4, $t4, 8                       \n"  // |R2|R1|R0|0|
+      "raddu.w.qb      $t0, $t0                          \n"
+      "raddu.w.qb      $t2, $t2                          \n"
+      "raddu.w.qb      $t4, $t4                          \n"
+      "addu            $t0, $t0, $t2                     \n"
+      "addu            $t0, $t0, $t4                     \n"
+      "mul             $t0, $t0, %[c1]                   \n"  // t0 * 0x1C71
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[s1], %[s1], 8                   \n"
+      "addiu           %[s2], %[s2], 8                   \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t7, $t7, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t6, -1(%[dst_ptr])               \n"
+      "sb              $t7, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst_ptr] "+r" (dst_ptr),
+        [s1] "+r" (s1),
+        [s2] "+r" (s2),
+        [dst_width] "+r" (dst_width)
+      : [c1] "r" (c1), [c2] "r" (c2)
+      : "t0", "t1", "t2", "t3", "t4",
+        "t5", "t6", "t7", "t8"
+  );
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/libvpx/libvpx/third_party/libyuv/source/scale_neon.cc b/libvpx/libvpx/third_party/libyuv/source/scale_neon.cc
new file mode 100644
index 0000000..7825878
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/scale_neon.cc
@@ -0,0 +1,1037 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+// NEON downscalers with interpolation.
+// Provided by Fritz Koenig
+
+// Read 32x1 throw away even pixels, and write 16x1.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1"              // Clobber List
+  );
+}
+
+// Read 32x1 average down and write 16x1.
+void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load pixels and post inc
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop
+    "vpaddl.u8  q0, q0                         \n"  // add adjacent
+    "vpaddl.u8  q1, q1                         \n"
+    "vrshrn.u16 d0, q0, #1                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #1                     \n"
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1"     // Clobber List
+  );
+}
+
+// Read 32x2 average down and write 16x1.
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 1 and post inc
+    MEMACCESS(1)
+    "vld1.8     {q2, q3}, [%1]!                \n"  // load row 2 and post inc
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
+    "vpaddl.u8  q1, q1                         \n"
+    "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent + row1
+    "vpadal.u8  q1, q3                         \n"
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "q0", "q1", "q2", "q3"     // Clobber List
+  );
+}
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n" // src line 0
+    "subs       %2, %2, #8                     \n" // 8 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {d2}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1", "memory", "cc"
+  );
+}
+
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  const uint8* src_ptr2 = src_ptr + src_stride * 2;
+  const uint8* src_ptr3 = src_ptr + src_stride * 3;
+asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"   // load up 16x4
+    MEMACCESS(3)
+    "vld1.8     {q1}, [%3]!                    \n"
+    MEMACCESS(4)
+    "vld1.8     {q2}, [%4]!                    \n"
+    MEMACCESS(5)
+    "vld1.8     {q3}, [%5]!                    \n"
+    "subs       %2, %2, #4                     \n"
+    "vpaddl.u8  q0, q0                         \n"
+    "vpadal.u8  q0, q1                         \n"
+    "vpadal.u8  q0, q2                         \n"
+    "vpadal.u8  q0, q3                         \n"
+    "vpaddl.u16 q0, q0                         \n"
+    "vrshrn.u32 d0, q0, #4                     \n"   // divide by 16 w/rounding
+    "vmovn.u16  d0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.32    {d0[0]}, [%1]!                 \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width), // %2
+    "+r"(src_ptr1),  // %3
+    "+r"(src_ptr2),  // %4
+    "+r"(src_ptr3)   // %5
+  :
+  : "q0", "q1", "q2", "q3", "memory", "cc"
+  );
+}
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+// to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "subs       %2, %2, #24                  \n"
+    "vmov       d2, d3                       \n" // order d0, d1, d2
+    MEMACCESS(1)
+    "vst3.8     {d0, d1, d2}, [%1]!          \n"
+    "bgt        1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "d0", "d1", "d2", "d3", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vmov.u8    d24, #3                        \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
+    "subs         %2, %2, #24                  \n"
+
+    // filter src line 0 with src line 1
+    // expand chars to shorts to allow for room
+    // when adding lines together
+    "vmovl.u8     q8, d4                       \n"
+    "vmovl.u8     q9, d5                       \n"
+    "vmovl.u8     q10, d6                      \n"
+    "vmovl.u8     q11, d7                      \n"
+
+    // 3 * line_0 + line_1
+    "vmlal.u8     q8, d0, d24                  \n"
+    "vmlal.u8     q9, d1, d24                  \n"
+    "vmlal.u8     q10, d2, d24                 \n"
+    "vmlal.u8     q11, d3, d24                 \n"
+
+    // (3 * line_0 + line_1) >> 2
+    "vqrshrn.u16  d0, q8, #2                   \n"
+    "vqrshrn.u16  d1, q9, #2                   \n"
+    "vqrshrn.u16  d2, q10, #2                  \n"
+    "vqrshrn.u16  d3, q11, #2                  \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "vmovl.u8     q8, d1                       \n"
+    "vmlal.u8     q8, d0, d24                  \n"
+    "vqrshrn.u16  d0, q8, #2                   \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "vrhadd.u8    d1, d1, d2                   \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "vmovl.u8     q8, d2                       \n"
+    "vmlal.u8     q8, d3, d24                  \n"
+    "vqrshrn.u16  d2, q8, #2                   \n"
+
+    MEMACCESS(1)
+    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vmov.u8    d24, #3                        \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
+    "subs         %2, %2, #24                  \n"
+    // average src line 0 with src line 1
+    "vrhadd.u8    q0, q0, q2                   \n"
+    "vrhadd.u8    q1, q1, q3                   \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "vmovl.u8     q3, d1                       \n"
+    "vmlal.u8     q3, d0, d24                  \n"
+    "vqrshrn.u16  d0, q3, #2                   \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "vrhadd.u8    d1, d1, d2                   \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "vmovl.u8     q3, d2                       \n"
+    "vmlal.u8     q3, d3, d24                  \n"
+    "vqrshrn.u16  d2, q3, #2                   \n"
+
+    MEMACCESS(1)
+    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
+  );
+}
+
+#define HAS_SCALEROWDOWN38_NEON
+static uvec8 kShuf38 =
+  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
+static uvec8 kShuf38_2 =
+  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
+static vec16 kMult38_Div6 =
+  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
+    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
+static vec16 kMult38_Div9 =
+  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
+    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q3}, [%3]                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"
+    "subs       %2, %2, #12                    \n"
+    "vtbl.u8    d4, {d0, d1, d2, d3}, d6       \n"
+    "vtbl.u8    d5, {d0, d1, d2, d3}, d7       \n"
+    MEMACCESS(1)
+    "vst1.8     {d4}, [%1]!                    \n"
+    MEMACCESS(1)
+    "vst1.32    {d5[0]}, [%1]!                 \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  : "r"(&kShuf38)           // %3
+  : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
+  );
+}
+
+// 32x3 -> 12x1
+void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_ptr, int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride * 2;
+
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.16    {q13}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {q14}, [%6]                    \n"
+    MEMACCESS(7)
+    "vld1.8     {q15}, [%7]                    \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+
+    // d0 = 00 40 01 41 02 42 03 43
+    // d1 = 10 50 11 51 12 52 13 53
+    // d2 = 20 60 21 61 22 62 23 63
+    // d3 = 30 70 31 71 32 72 33 73
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    MEMACCESS(4)
+    "vld4.8       {d16, d17, d18, d19}, [%4]!  \n"
+    "subs         %2, %2, #12                  \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // d0 = 00 10 01 11 02 12 03 13
+    // d1 = 40 50 41 51 42 52 43 53
+    "vtrn.u8      d0, d1                       \n"
+    "vtrn.u8      d4, d5                       \n"
+    "vtrn.u8      d16, d17                     \n"
+
+    // d2 = 20 30 21 31 22 32 23 33
+    // d3 = 60 70 61 71 62 72 63 73
+    "vtrn.u8      d2, d3                       \n"
+    "vtrn.u8      d6, d7                       \n"
+    "vtrn.u8      d18, d19                     \n"
+
+    // d0 = 00+10 01+11 02+12 03+13
+    // d2 = 40+50 41+51 42+52 43+53
+    "vpaddl.u8    q0, q0                       \n"
+    "vpaddl.u8    q2, q2                       \n"
+    "vpaddl.u8    q8, q8                       \n"
+
+    // d3 = 60+70 61+71 62+72 63+73
+    "vpaddl.u8    d3, d3                       \n"
+    "vpaddl.u8    d7, d7                       \n"
+    "vpaddl.u8    d19, d19                     \n"
+
+    // combine source lines
+    "vadd.u16     q0, q2                       \n"
+    "vadd.u16     q0, q8                       \n"
+    "vadd.u16     d4, d3, d7                   \n"
+    "vadd.u16     d4, d19                      \n"
+
+    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
+    //             + s[6 + st * 1] + s[7 + st * 1]
+    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
+    "vqrdmulh.s16 q2, q2, q13                  \n"
+    "vmovn.u16    d4, q2                       \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "vmovl.u8     q1, d2                       \n"
+    "vmovl.u8     q3, d6                       \n"
+    "vmovl.u8     q9, d18                      \n"
+
+    // combine source lines
+    "vadd.u16     q1, q3                       \n"
+    "vadd.u16     q1, q9                       \n"
+
+    // d4 = xx 20 xx 30 xx 22 xx 32
+    // d5 = xx 21 xx 31 xx 23 xx 33
+    "vtrn.u32     d2, d3                       \n"
+
+    // d4 = xx 20 xx 21 xx 22 xx 23
+    // d5 = xx 30 xx 31 xx 32 xx 33
+    "vtrn.u16     d2, d3                       \n"
+
+    // 0+1+2, 3+4+5
+    "vadd.u16     q0, q1                       \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "vqrdmulh.s16 q0, q0, q15                  \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "vmov.u8      d2, d4                       \n"
+
+    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+    MEMACCESS(1)
+    "vst1.8       {d3}, [%1]!                  \n"
+    MEMACCESS(1)
+    "vst1.32      {d4[0]}, [%1]!               \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride),       // %3
+    "+r"(src_ptr1)          // %4
+  : "r"(&kMult38_Div6),     // %5
+    "r"(&kShuf38_2),        // %6
+    "r"(&kMult38_Div9)      // %7
+  : "q0", "q1", "q2", "q3", "q8", "q9", "q13", "q14", "q15", "memory", "cc"
+  );
+}
+
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.16    {q13}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {q14}, [%5]                    \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+
+    // d0 = 00 40 01 41 02 42 03 43
+    // d1 = 10 50 11 51 12 52 13 53
+    // d2 = 20 60 21 61 22 62 23 63
+    // d3 = 30 70 31 71 32 72 33 73
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    "subs         %2, %2, #12                  \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // d0 = 00 10 01 11 02 12 03 13
+    // d1 = 40 50 41 51 42 52 43 53
+    "vtrn.u8      d0, d1                       \n"
+    "vtrn.u8      d4, d5                       \n"
+
+    // d2 = 20 30 21 31 22 32 23 33
+    // d3 = 60 70 61 71 62 72 63 73
+    "vtrn.u8      d2, d3                       \n"
+    "vtrn.u8      d6, d7                       \n"
+
+    // d0 = 00+10 01+11 02+12 03+13
+    // d2 = 40+50 41+51 42+52 43+53
+    "vpaddl.u8    q0, q0                       \n"
+    "vpaddl.u8    q2, q2                       \n"
+
+    // d3 = 60+70 61+71 62+72 63+73
+    "vpaddl.u8    d3, d3                       \n"
+    "vpaddl.u8    d7, d7                       \n"
+
+    // combine source lines
+    "vadd.u16     q0, q2                       \n"
+    "vadd.u16     d4, d3, d7                   \n"
+
+    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
+    "vqrshrn.u16  d4, q2, #2                   \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "vmovl.u8     q1, d2                       \n"
+    "vmovl.u8     q3, d6                       \n"
+
+    // combine source lines
+    "vadd.u16     q1, q3                       \n"
+
+    // d4 = xx 20 xx 30 xx 22 xx 32
+    // d5 = xx 21 xx 31 xx 23 xx 33
+    "vtrn.u32     d2, d3                       \n"
+
+    // d4 = xx 20 xx 21 xx 22 xx 23
+    // d5 = xx 30 xx 31 xx 32 xx 33
+    "vtrn.u16     d2, d3                       \n"
+
+    // 0+1+2, 3+4+5
+    "vadd.u16     q0, q1                       \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "vqrdmulh.s16 q0, q0, q13                  \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "vmov.u8      d2, d4                       \n"
+
+    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+    MEMACCESS(1)
+    "vst1.8       {d3}, [%1]!                  \n"
+    MEMACCESS(1)
+    "vst1.32      {d4[0]}, [%1]!               \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),       // %0
+    "+r"(dst_ptr),       // %1
+    "+r"(dst_width),     // %2
+    "+r"(src_stride)     // %3
+  : "r"(&kMult38_Div6),  // %4
+    "r"(&kShuf38_2)      // %5
+  : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc"
+  );
+}
+
+void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height) {
+  const uint8* src_tmp = NULL;
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "mov       %0, %1                          \n"
+    "mov       r12, %5                         \n"
+    "veor      q2, q2, q2                      \n"
+    "veor      q3, q3, q3                      \n"
+  "2:                                          \n"
+    // load 16 pixels into q0
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], %3                 \n"
+    "vaddw.u8   q3, q3, d1                     \n"
+    "vaddw.u8   q2, q2, d0                     \n"
+    "subs       r12, r12, #1                   \n"
+    "bgt        2b                             \n"
+    MEMACCESS(2)
+    "vst1.16    {q2, q3}, [%2]!                \n"  // store pixels
+    "add        %1, %1, #16                    \n"
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop
+    "bgt        1b                             \n"
+  : "+r"(src_tmp),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_ptr),          // %2
+    "+r"(src_stride),       // %3
+    "+r"(src_width),        // %4
+    "+r"(src_height)        // %5
+  :
+  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD2_DATA8_LANE(n)                                    \
+    "lsr        %5, %3, #16                    \n"             \
+    "add        %6, %1, %5                     \n"             \
+    "add        %3, %3, %4                     \n"             \
+    MEMACCESS(6)                                               \
+    "vld2.8     {d6["#n"], d7["#n"]}, [%6]     \n"
+
+void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                          int dst_width, int x, int dx) {
+  int dx_offset[4] = {0, 1, 2, 3};
+  int* tmp = dx_offset;
+  const uint8* src_tmp = src_ptr;
+  asm volatile (
+    ".p2align   2                              \n"
+    "vdup.32    q0, %3                         \n"  // x
+    "vdup.32    q1, %4                         \n"  // dx
+    "vld1.32    {q2}, [%5]                     \n"  // 0 1 2 3
+    "vshl.i32   q3, q1, #2                     \n"  // 4 * dx
+    "vmul.s32   q1, q1, q2                     \n"
+    // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
+    "vadd.s32   q1, q1, q0                     \n"
+    // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx
+    "vadd.s32   q2, q1, q3                     \n"
+    "vshl.i32   q0, q3, #1                     \n"  // 8 * dx
+  "1:                                          \n"
+    LOAD2_DATA8_LANE(0)
+    LOAD2_DATA8_LANE(1)
+    LOAD2_DATA8_LANE(2)
+    LOAD2_DATA8_LANE(3)
+    LOAD2_DATA8_LANE(4)
+    LOAD2_DATA8_LANE(5)
+    LOAD2_DATA8_LANE(6)
+    LOAD2_DATA8_LANE(7)
+    "vmov       q10, q1                        \n"
+    "vmov       q11, q2                        \n"
+    "vuzp.16    q10, q11                       \n"
+    "vmovl.u8   q8, d6                         \n"
+    "vmovl.u8   q9, d7                         \n"
+    "vsubl.s16  q11, d18, d16                  \n"
+    "vsubl.s16  q12, d19, d17                  \n"
+    "vmovl.u16  q13, d20                       \n"
+    "vmovl.u16  q10, d21                       \n"
+    "vmul.s32   q11, q11, q13                  \n"
+    "vmul.s32   q12, q12, q10                  \n"
+    "vshrn.s32  d18, q11, #16                  \n"
+    "vshrn.s32  d19, q12, #16                  \n"
+    "vadd.s16   q8, q8, q9                     \n"
+    "vmovn.s16  d6, q8                         \n"
+
+    MEMACCESS(0)
+    "vst1.8     {d6}, [%0]!                    \n"  // store pixels
+    "vadd.s32   q1, q1, q0                     \n"
+    "vadd.s32   q2, q2, q0                     \n"
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "bgt        1b                             \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(x),                // %3
+    "+r"(dx),               // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3",
+    "q8", "q9", "q10", "q11", "q12", "q13"
+  );
+}
+
+#undef LOAD2_DATA8_LANE
+
+// 16x2 -> 16x1
+void ScaleFilterRows_NEON(uint8* dst_ptr,
+                          const uint8* src_ptr, ptrdiff_t src_stride,
+                          int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp          %4, #0                       \n"
+    "beq          100f                         \n"
+    "add          %2, %1                       \n"
+    "cmp          %4, #64                      \n"
+    "beq          75f                          \n"
+    "cmp          %4, #128                     \n"
+    "beq          50f                          \n"
+    "cmp          %4, #192                     \n"
+    "beq          25f                          \n"
+
+    "vdup.8       d5, %4                       \n"
+    "rsb          %4, #256                     \n"
+    "vdup.8       d4, %4                       \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vmull.u8     q13, d0, d4                  \n"
+    "vmull.u8     q14, d1, d4                  \n"
+    "vmlal.u8     q13, d2, d5                  \n"
+    "vmlal.u8     q14, d3, d5                  \n"
+    "vrshrn.u16   d0, q13, #8                  \n"
+    "vrshrn.u16   d1, q14, #8                  \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          1b                           \n"
+    "b            99f                          \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vrhadd.u8    q0, q1                       \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          25b                          \n"
+    "b            99f                          \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          50b                          \n"
+    "b            99f                          \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "vld1.8       {q1}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q0}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vrhadd.u8    q0, q1                       \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          75b                          \n"
+    "b            99f                          \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          100b                         \n"
+
+  "99:                                         \n"
+    MEMACCESS(0)
+    "vst1.8       {d1[7]}, [%0]                \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc"
+  );
+}
+
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    MEMACCESS(0)
+    "vld2.32    {q0, q1}, [%0]!                \n"
+    MEMACCESS(0)
+    "vld2.32    {q2, q3}, [%0]!                \n"
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    MEMACCESS(1)
+    "vst1.8     {q3}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
+    "vrshrn.u16 d0, q0, #1                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #1                     \n"
+    "vrshrn.u16 d2, q2, #1                     \n"
+    "vrshrn.u16 d3, q3, #1                     \n"
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"
+    "bgt       1b                              \n"
+  : "+r"(src_argb),         // %0
+    "+r"(dst_argb),         // %1
+    "+r"(dst_width)         // %2
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3"     // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %1, %0                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
+    "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    "vrshrn.u16 d2, q2, #2                     \n"
+    "vrshrn.u16 d3, q3, #2                     \n"
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %3, lsl #2                \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.32    {d0[0]}, [%0], r12             \n"
+    MEMACCESS(0)
+    "vld1.32    {d0[1]}, [%0], r12             \n"
+    MEMACCESS(0)
+    "vld1.32    {d1[0]}, [%0], r12             \n"
+    MEMACCESS(0)
+    "vld1.32    {d1[1]}, [%0], r12             \n"
+    "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(dst_width)    // %2
+  : "r"(src_stepx)     // %3
+  : "memory", "cc", "r12", "q0"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %4, lsl #2                \n"
+    "add        %1, %1, %0                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0], r12                \n"  // Read 4 2x2 blocks -> 2x1
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1], r12                \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0], r12                \n"
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1], r12                \n"
+    MEMACCESS(0)
+    "vld1.8     {d4}, [%0], r12                \n"
+    MEMACCESS(1)
+    "vld1.8     {d5}, [%1], r12                \n"
+    MEMACCESS(0)
+    "vld1.8     {d6}, [%0], r12                \n"
+    MEMACCESS(1)
+    "vld1.8     {d7}, [%1], r12                \n"
+    "vaddl.u8   q0, d0, d1                     \n"
+    "vaddl.u8   q1, d2, d3                     \n"
+    "vaddl.u8   q2, d4, d5                     \n"
+    "vaddl.u8   q3, d6, d7                     \n"
+    "vswp.8     d1, d2                         \n"  // ab_cd -> ac_bd
+    "vswp.8     d5, d6                         \n"  // ef_gh -> eg_fh
+    "vadd.u16   q0, q0, q1                     \n"  // (a+b)_(c+d)
+    "vadd.u16   q2, q2, q3                     \n"  // (e+f)_(g+h)
+    "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
+    "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
+    "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(src_stride),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(dst_width)    // %3
+  : "r"(src_stepx)     // %4
+  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"
+  );
+}
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD1_DATA32_LANE(dn, n)                               \
+    "lsr        %5, %3, #16                    \n"             \
+    "add        %6, %1, %5, lsl #2             \n"             \
+    "add        %3, %3, %4                     \n"             \
+    MEMACCESS(6)                                               \
+    "vld1.32    {"#dn"["#n"]}, [%6]            \n"
+
+void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  int tmp = 0;
+  const uint8* src_tmp = src_argb;
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    LOAD1_DATA32_LANE(d0, 0)
+    LOAD1_DATA32_LANE(d0, 1)
+    LOAD1_DATA32_LANE(d1, 0)
+    LOAD1_DATA32_LANE(d1, 1)
+    LOAD1_DATA32_LANE(d2, 0)
+    LOAD1_DATA32_LANE(d2, 1)
+    LOAD1_DATA32_LANE(d3, 0)
+    LOAD1_DATA32_LANE(d3, 1)
+
+    MEMACCESS(0)
+    "vst1.32     {q0, q1}, [%0]!               \n"  // store pixels
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),         // %0
+    "+r"(src_argb),         // %1
+    "+r"(dst_width),        // %2
+    "+r"(x),                // %3
+    "+r"(dx),               // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "q0", "q1"
+  );
+}
+
+#undef LOAD1_DATA32_LANE
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD2_DATA32_LANE(dn1, dn2, n)                         \
+    "lsr        %5, %3, #16                           \n"      \
+    "add        %6, %1, %5, lsl #2                    \n"      \
+    "add        %3, %3, %4                            \n"      \
+    MEMACCESS(6)                                               \
+    "vld2.32    {"#dn1"["#n"], "#dn2"["#n"]}, [%6]    \n"
+
+void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                              int dst_width, int x, int dx) {
+  int dx_offset[4] = {0, 1, 2, 3};
+  int* tmp = dx_offset;
+  const uint8* src_tmp = src_argb;
+  asm volatile (
+    ".p2align   2                              \n"
+    "vdup.32    q0, %3                         \n"  // x
+    "vdup.32    q1, %4                         \n"  // dx
+    "vld1.32    {q2}, [%5]                     \n"  // 0 1 2 3
+    "vshl.i32   q9, q1, #2                     \n"  // 4 * dx
+    "vmul.s32   q1, q1, q2                     \n"
+    "vmov.i8    q3, #0x7f                      \n"  // 0x7F
+    "vmov.i16   q15, #0x7f                     \n"  // 0x7F
+    // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
+    "vadd.s32   q8, q1, q0                     \n"
+  "1:                                          \n"
+    // d0, d1: a
+    // d2, d3: b
+    LOAD2_DATA32_LANE(d0, d2, 0)
+    LOAD2_DATA32_LANE(d0, d2, 1)
+    LOAD2_DATA32_LANE(d1, d3, 0)
+    LOAD2_DATA32_LANE(d1, d3, 1)
+    "vshrn.i32   d22, q8, #9                   \n"
+    "vand.16     d22, d22, d30                 \n"
+    "vdup.8      d24, d22[0]                   \n"
+    "vdup.8      d25, d22[2]                   \n"
+    "vdup.8      d26, d22[4]                   \n"
+    "vdup.8      d27, d22[6]                   \n"
+    "vext.8      d4, d24, d25, #4              \n"
+    "vext.8      d5, d26, d27, #4              \n"  // f
+    "veor.8      q10, q2, q3                   \n"  // 0x7f ^ f
+    "vmull.u8    q11, d0, d20                  \n"
+    "vmull.u8    q12, d1, d21                  \n"
+    "vmull.u8    q13, d2, d4                   \n"
+    "vmull.u8    q14, d3, d5                   \n"
+    "vadd.i16    q11, q11, q13                 \n"
+    "vadd.i16    q12, q12, q14                 \n"
+    "vshrn.i16   d0, q11, #7                   \n"
+    "vshrn.i16   d1, q12, #7                   \n"
+
+    MEMACCESS(0)
+    "vst1.32     {d0, d1}, [%0]!               \n"  // store pixels
+    "vadd.s32    q8, q8, q9                    \n"
+    "subs        %2, %2, #4                    \n"  // 4 processed per loop
+    "bgt         1b                            \n"
+  : "+r"(dst_argb),         // %0
+    "+r"(src_argb),         // %1
+    "+r"(dst_width),        // %2
+    "+r"(x),                // %3
+    "+r"(dx),               // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#undef LOAD2_DATA32_LANE
+
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/scale_neon64.cc b/libvpx/libvpx/third_party/libyuv/source/scale_neon64.cc
new file mode 100644
index 0000000..1d55193
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/scale_neon64.cc
@@ -0,0 +1,1042 @@
+/*
+ *  Copyright 2014 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon armv8 64 bit.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+// Read 32x1 throw away even pixels, and write 16x1.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    // load even pixels into v0, odd into v1
+    MEMACCESS(0)
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"
+    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "st1        {v1.16b}, [%1], #16            \n"  // store odd pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "v0", "v1"              // Clobber List
+  );
+}
+
+// Read 32x1 average down and write 16x1.
+void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b,v1.16b}, [%0], #32     \n"  // load pixels and post inc
+    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
+    "uaddlp     v0.8h, v0.16b                  \n"  // add adjacent
+    "uaddlp     v1.8h, v1.16b                  \n"
+    "rshrn      v0.8b, v0.8h, #1               \n"  // downshift, round and pack
+    "rshrn2     v0.16b, v1.8h, #1              \n"
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "v0", "v1"     // Clobber List
+  );
+}
+
+// Read 32x2 average down and write 16x1.
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %1, %0                     \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b,v1.16b}, [%0], #32    \n"  // load row 1 and post inc
+    MEMACCESS(1)
+    "ld1        {v2.16b, v3.16b}, [%1], #32    \n"  // load row 2 and post inc
+    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
+    "uaddlp     v0.8h, v0.16b                  \n"  // row 1 add adjacent
+    "uaddlp     v1.8h, v1.16b                  \n"
+    "uadalp     v0.8h, v2.16b                  \n"  // row 2 add adjacent + row1
+    "uadalp     v1.8h, v3.16b                  \n"
+    "rshrn      v0.8b, v0.8h, #2               \n"  // downshift, round and pack
+    "rshrn2     v0.16b, v1.8h, #2              \n"
+    MEMACCESS(2)
+    "st1        {v0.16b}, [%2], #16            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "v0", "v1", "v2", "v3"     // Clobber List
+  );
+}
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4     {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32          \n"  // src line 0
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
+    MEMACCESS(1)
+    "st1     {v2.8b}, [%1], #8                 \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "v0", "v1", "v2", "v3", "memory", "cc"
+  );
+}
+
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  const uint8* src_ptr2 = src_ptr + src_stride * 2;
+  const uint8* src_ptr3 = src_ptr + src_stride * 3;
+asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1     {v0.16b}, [%0], #16               \n"   // load up 16x4
+    MEMACCESS(3)
+    "ld1     {v1.16b}, [%2], #16               \n"
+    MEMACCESS(4)
+    "ld1     {v2.16b}, [%3], #16               \n"
+    MEMACCESS(5)
+    "ld1     {v3.16b}, [%4], #16               \n"
+    "subs    %w5, %w5, #4                      \n"
+    "uaddlp  v0.8h, v0.16b                     \n"
+    "uadalp  v0.8h, v1.16b                     \n"
+    "uadalp  v0.8h, v2.16b                     \n"
+    "uadalp  v0.8h, v3.16b                     \n"
+    "addp    v0.8h, v0.8h, v0.8h               \n"
+    "rshrn   v0.8b, v0.8h, #4                  \n"   // divide by 16 w/rounding
+    MEMACCESS(1)
+    "st1    {v0.s}[0], [%1], #4                \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(src_ptr1),  // %2
+    "+r"(src_ptr2),  // %3
+    "+r"(src_ptr3),  // %4
+    "+r"(dst_width)  // %5
+  :
+  : "v0", "v1", "v2", "v3", "memory", "cc"
+  );
+}
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+// to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+  "1:                                                  \n"
+    MEMACCESS(0)
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
+    "subs      %w2, %w2, #24                           \n"
+    "orr       v2.16b, v3.16b, v3.16b                  \n"  // order v0, v1, v2
+    MEMACCESS(1)
+    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "v0", "v1", "v2", "v3", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movi      v20.8b, #3                              \n"
+    "add       %3, %3, %0                              \n"
+  "1:                                                  \n"
+    MEMACCESS(0)
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
+    MEMACCESS(3)
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32                \n"  // src line 1
+    "subs         %w2, %w2, #24                        \n"
+
+    // filter src line 0 with src line 1
+    // expand chars to shorts to allow for room
+    // when adding lines together
+    "ushll     v16.8h, v4.8b, #0                       \n"
+    "ushll     v17.8h, v5.8b, #0                       \n"
+    "ushll     v18.8h, v6.8b, #0                       \n"
+    "ushll     v19.8h, v7.8b, #0                       \n"
+
+    // 3 * line_0 + line_1
+    "umlal     v16.8h, v0.8b, v20.8b                   \n"
+    "umlal     v17.8h, v1.8b, v20.8b                   \n"
+    "umlal     v18.8h, v2.8b, v20.8b                   \n"
+    "umlal     v19.8h, v3.8b, v20.8b                   \n"
+
+    // (3 * line_0 + line_1) >> 2
+    "uqrshrn   v0.8b, v16.8h, #2                       \n"
+    "uqrshrn   v1.8b, v17.8h, #2                       \n"
+    "uqrshrn   v2.8b, v18.8h, #2                       \n"
+    "uqrshrn   v3.8b, v19.8h, #2                       \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "ushll     v16.8h, v1.8b, #0                       \n"
+    "umlal     v16.8h, v0.8b, v20.8b                   \n"
+    "uqrshrn   v0.8b, v16.8h, #2                       \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "urhadd    v1.8b, v1.8b, v2.8b                     \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "ushll     v16.8h, v2.8b, #0                       \n"
+    "umlal     v16.8h, v3.8b, v20.8b                   \n"
+    "uqrshrn   v2.8b, v16.8h, #2                       \n"
+
+    MEMACCESS(1)
+    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
+
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19",
+    "v20", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movi      v20.8b, #3                              \n"
+    "add       %3, %3, %0                              \n"
+  "1:                                                  \n"
+    MEMACCESS(0)
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
+    MEMACCESS(3)
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32                \n"  // src line 1
+    "subs         %w2, %w2, #24                        \n"
+    // average src line 0 with src line 1
+    "urhadd    v0.8b, v0.8b, v4.8b                     \n"
+    "urhadd    v1.8b, v1.8b, v5.8b                     \n"
+    "urhadd    v2.8b, v2.8b, v6.8b                     \n"
+    "urhadd    v3.8b, v3.8b, v7.8b                     \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "ushll     v4.8h, v1.8b, #0                        \n"
+    "umlal     v4.8h, v0.8b, v20.8b                    \n"
+    "uqrshrn   v0.8b, v4.8h, #2                        \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "urhadd    v1.8b, v1.8b, v2.8b                     \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "ushll     v4.8h, v2.8b, #0                        \n"
+    "umlal     v4.8h, v3.8b, v20.8b                    \n"
+    "uqrshrn   v2.8b, v4.8h, #2                        \n"
+
+    MEMACCESS(1)
+    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", "memory", "cc"
+  );
+}
+
+static uvec8 kShuf38 =
+  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
+static uvec8 kShuf38_2 =
+  { 0, 16, 32, 2, 18, 33, 4, 20, 34, 6, 22, 35, 0, 0, 0, 0 };
+static vec16 kMult38_Div6 =
+  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
+    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
+static vec16 kMult38_Div9 =
+  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
+    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    MEMACCESS(3)
+    "ld1       {v3.16b}, [%3]                          \n"
+  "1:                                                  \n"
+    MEMACCESS(0)
+    "ld1       {v0.16b,v1.16b}, [%0], #32             \n"
+    "subs      %w2, %w2, #12                           \n"
+    "tbl       v2.16b, {v0.16b,v1.16b}, v3.16b        \n"
+    MEMACCESS(1)
+    "st1       {v2.8b}, [%1], #8                       \n"
+    MEMACCESS(1)
+    "st1       {v2.s}[2], [%1], #4                     \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  : "r"(&kShuf38)           // %3
+  : "v0", "v1", "v2", "v3", "memory", "cc"
+  );
+}
+
+// 32x3 -> 12x1
+void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_ptr, int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride * 2;
+  ptrdiff_t tmp_src_stride = src_stride;
+
+  asm volatile (
+    MEMACCESS(5)
+    "ld1       {v29.8h}, [%5]                          \n"
+    MEMACCESS(6)
+    "ld1       {v30.16b}, [%6]                         \n"
+    MEMACCESS(7)
+    "ld1       {v31.8h}, [%7]                          \n"
+    "add       %2, %2, %0                              \n"
+  "1:                                                  \n"
+
+    // 00 40 01 41 02 42 03 43
+    // 10 50 11 51 12 52 13 53
+    // 20 60 21 61 22 62 23 63
+    // 30 70 31 71 32 72 33 73
+    MEMACCESS(0)
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"
+    MEMACCESS(3)
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32                \n"
+    MEMACCESS(4)
+    "ld4       {v16.8b,v17.8b,v18.8b,v19.8b}, [%3], #32              \n"
+    "subs      %w4, %w4, #12                           \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // 00 10 01 11 02 12 03 13
+    // 40 50 41 51 42 52 43 53
+    "trn1      v20.8b, v0.8b, v1.8b                    \n"
+    "trn2      v21.8b, v0.8b, v1.8b                    \n"
+    "trn1      v22.8b, v4.8b, v5.8b                    \n"
+    "trn2      v23.8b, v4.8b, v5.8b                    \n"
+    "trn1      v24.8b, v16.8b, v17.8b                  \n"
+    "trn2      v25.8b, v16.8b, v17.8b                  \n"
+
+    // 20 30 21 31 22 32 23 33
+    // 60 70 61 71 62 72 63 73
+    "trn1      v0.8b, v2.8b, v3.8b                     \n"
+    "trn2      v1.8b, v2.8b, v3.8b                     \n"
+    "trn1      v4.8b, v6.8b, v7.8b                     \n"
+    "trn2      v5.8b, v6.8b, v7.8b                     \n"
+    "trn1      v16.8b, v18.8b, v19.8b                  \n"
+    "trn2      v17.8b, v18.8b, v19.8b                  \n"
+
+    // 00+10 01+11 02+12 03+13
+    // 40+50 41+51 42+52 43+53
+    "uaddlp    v20.4h, v20.8b                          \n"
+    "uaddlp    v21.4h, v21.8b                          \n"
+    "uaddlp    v22.4h, v22.8b                          \n"
+    "uaddlp    v23.4h, v23.8b                          \n"
+    "uaddlp    v24.4h, v24.8b                          \n"
+    "uaddlp    v25.4h, v25.8b                          \n"
+
+    // 60+70 61+71 62+72 63+73
+    "uaddlp    v1.4h, v1.8b                            \n"
+    "uaddlp    v5.4h, v5.8b                            \n"
+    "uaddlp    v17.4h, v17.8b                          \n"
+
+    // combine source lines
+    "add       v20.4h, v20.4h, v22.4h                  \n"
+    "add       v21.4h, v21.4h, v23.4h                  \n"
+    "add       v20.4h, v20.4h, v24.4h                  \n"
+    "add       v21.4h, v21.4h, v25.4h                  \n"
+    "add       v2.4h, v1.4h, v5.4h                     \n"
+    "add       v2.4h, v2.4h, v17.4h                    \n"
+
+    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
+    //             + s[6 + st * 1] + s[7 + st * 1]
+    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
+    "sqrdmulh  v2.8h, v2.8h, v29.8h                    \n"
+    "xtn       v2.8b,  v2.8h                           \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "ushll     v16.8h, v16.8b, #0                      \n"
+    "uaddl     v0.8h, v0.8b, v4.8b                     \n"
+
+    // combine source lines
+    "add       v0.8h, v0.8h, v16.8h                    \n"
+
+    // xx 20 xx 21 xx 22 xx 23
+    // xx 30 xx 31 xx 32 xx 33
+    "trn1      v1.8h, v0.8h, v0.8h                     \n"
+    "trn2      v4.8h, v0.8h, v0.8h                     \n"
+    "xtn       v0.4h, v1.4s                            \n"
+    "xtn       v4.4h, v4.4s                            \n"
+
+    // 0+1+2, 3+4+5
+    "add       v20.8h, v20.8h, v0.8h                   \n"
+    "add       v21.8h, v21.8h, v4.8h                   \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "sqrdmulh  v0.8h, v20.8h, v31.8h                   \n"
+    "sqrdmulh  v1.8h, v21.8h, v31.8h                   \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "tbl       v3.16b, {v0.16b, v1.16b, v2.16b}, v30.16b \n"
+
+    MEMACCESS(1)
+    "st1       {v3.8b}, [%1], #8                       \n"
+    MEMACCESS(1)
+    "st1       {v3.s}[2], [%1], #4                     \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(tmp_src_stride),   // %2
+    "+r"(src_ptr1),         // %3
+    "+r"(dst_width)         // %4
+  : "r"(&kMult38_Div6),     // %5
+    "r"(&kShuf38_2),        // %6
+    "r"(&kMult38_Div9)      // %7
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
+    "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v29",
+    "v30", "v31", "memory", "cc"
+  );
+}
+
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  // TODO(fbarchard): use src_stride directly for clang 3.5+.
+  ptrdiff_t tmp_src_stride = src_stride;
+  asm volatile (
+    MEMACCESS(4)
+    "ld1       {v30.8h}, [%4]                          \n"
+    MEMACCESS(5)
+    "ld1       {v31.16b}, [%5]                         \n"
+    "add       %2, %2, %0                              \n"
+  "1:                                                  \n"
+
+    // 00 40 01 41 02 42 03 43
+    // 10 50 11 51 12 52 13 53
+    // 20 60 21 61 22 62 23 63
+    // 30 70 31 71 32 72 33 73
+    MEMACCESS(0)
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"
+    MEMACCESS(3)
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32                \n"
+    "subs      %w3, %w3, #12                           \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // 00 10 01 11 02 12 03 13
+    // 40 50 41 51 42 52 43 53
+    "trn1      v16.8b, v0.8b, v1.8b                    \n"
+    "trn2      v17.8b, v0.8b, v1.8b                    \n"
+    "trn1      v18.8b, v4.8b, v5.8b                    \n"
+    "trn2      v19.8b, v4.8b, v5.8b                    \n"
+
+    // 20 30 21 31 22 32 23 33
+    // 60 70 61 71 62 72 63 73
+    "trn1      v0.8b, v2.8b, v3.8b                     \n"
+    "trn2      v1.8b, v2.8b, v3.8b                     \n"
+    "trn1      v4.8b, v6.8b, v7.8b                     \n"
+    "trn2      v5.8b, v6.8b, v7.8b                     \n"
+
+    // 00+10 01+11 02+12 03+13
+    // 40+50 41+51 42+52 43+53
+    "uaddlp    v16.4h, v16.8b                          \n"
+    "uaddlp    v17.4h, v17.8b                          \n"
+    "uaddlp    v18.4h, v18.8b                          \n"
+    "uaddlp    v19.4h, v19.8b                          \n"
+
+    // 60+70 61+71 62+72 63+73
+    "uaddlp    v1.4h, v1.8b                            \n"
+    "uaddlp    v5.4h, v5.8b                            \n"
+
+    // combine source lines
+    "add       v16.4h, v16.4h, v18.4h                  \n"
+    "add       v17.4h, v17.4h, v19.4h                  \n"
+    "add       v2.4h, v1.4h, v5.4h                     \n"
+
+    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
+    "uqrshrn   v2.8b, v2.8h, #2                        \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+
+    // combine source lines
+    "uaddl     v0.8h, v0.8b, v4.8b                     \n"
+
+    // xx 20 xx 21 xx 22 xx 23
+    // xx 30 xx 31 xx 32 xx 33
+    "trn1      v1.8h, v0.8h, v0.8h                     \n"
+    "trn2      v4.8h, v0.8h, v0.8h                     \n"
+    "xtn       v0.4h, v1.4s                            \n"
+    "xtn       v4.4h, v4.4s                            \n"
+
+    // 0+1+2, 3+4+5
+    "add       v16.8h, v16.8h, v0.8h                   \n"
+    "add       v17.8h, v17.8h, v4.8h                   \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "sqrdmulh  v0.8h, v16.8h, v30.8h                   \n"
+    "sqrdmulh  v1.8h, v17.8h, v30.8h                   \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+
+    "tbl       v3.16b, {v0.16b, v1.16b, v2.16b}, v31.16b \n"
+
+    MEMACCESS(1)
+    "st1       {v3.8b}, [%1], #8                       \n"
+    MEMACCESS(1)
+    "st1       {v3.s}[2], [%1], #4                     \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),         // %0
+    "+r"(dst_ptr),         // %1
+    "+r"(tmp_src_stride),  // %2
+    "+r"(dst_width)        // %3
+  : "r"(&kMult38_Div6),    // %4
+    "r"(&kShuf38_2)        // %5
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
+    "v18", "v19", "v30", "v31", "memory", "cc"
+  );
+}
+
+void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height) {
+  const uint8* src_tmp = NULL;
+  asm volatile (
+  "1:                                          \n"
+    "mov       %0, %1                          \n"
+    "mov       w12, %w5                        \n"
+    "eor       v2.16b, v2.16b, v2.16b          \n"
+    "eor       v3.16b, v3.16b, v3.16b          \n"
+  "2:                                          \n"
+    // load 16 pixels into q0
+    MEMACCESS(0)
+    "ld1       {v0.16b}, [%0], %3              \n"
+    "uaddw2    v3.8h, v3.8h, v0.16b            \n"
+    "uaddw     v2.8h, v2.8h, v0.8b             \n"
+    "subs      w12, w12, #1                    \n"
+    "b.gt      2b                              \n"
+    MEMACCESS(2)
+    "st1      {v2.8h, v3.8h}, [%2], #32        \n"  // store pixels
+    "add      %1, %1, #16                      \n"
+    "subs     %w4, %w4, #16                    \n"  // 16 processed per loop
+    "b.gt     1b                               \n"
+  : "+r"(src_tmp),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_ptr),          // %2
+    "+r"(src_stride),       // %3
+    "+r"(src_width),        // %4
+    "+r"(src_height)        // %5
+  :
+  : "memory", "cc", "w12", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD2_DATA8_LANE(n)                                    \
+    "lsr        %5, %3, #16                    \n"             \
+    "add        %6, %1, %5                    \n"              \
+    "add        %3, %3, %4                     \n"             \
+    MEMACCESS(6)                                               \
+    "ld2        {v4.b, v5.b}["#n"], [%6]      \n"
+
+void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                          int dst_width, int x, int dx) {
+  int dx_offset[4] = {0, 1, 2, 3};
+  int* tmp = dx_offset;
+  const uint8* src_tmp = src_ptr;
+  int64 dst_width64 = (int64) dst_width;  // Work around ios 64 bit warning.
+  int64 x64 = (int64) x;
+  int64 dx64 = (int64) dx;
+  asm volatile (
+    "dup        v0.4s, %w3                     \n"  // x
+    "dup        v1.4s, %w4                     \n"  // dx
+    "ld1        {v2.4s}, [%5]                  \n"  // 0 1 2 3
+    "shl        v3.4s, v1.4s, #2               \n"  // 4 * dx
+    "mul        v1.4s, v1.4s, v2.4s            \n"
+    // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
+    "add        v1.4s, v1.4s, v0.4s            \n"
+    // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx
+    "add        v2.4s, v1.4s, v3.4s            \n"
+    "shl        v0.4s, v3.4s, #1               \n"  // 8 * dx
+  "1:                                          \n"
+    LOAD2_DATA8_LANE(0)
+    LOAD2_DATA8_LANE(1)
+    LOAD2_DATA8_LANE(2)
+    LOAD2_DATA8_LANE(3)
+    LOAD2_DATA8_LANE(4)
+    LOAD2_DATA8_LANE(5)
+    LOAD2_DATA8_LANE(6)
+    LOAD2_DATA8_LANE(7)
+    "mov       v6.16b, v1.16b                  \n"
+    "mov       v7.16b, v2.16b                  \n"
+    "uzp1      v6.8h, v6.8h, v7.8h             \n"
+    "ushll     v4.8h, v4.8b, #0                \n"
+    "ushll     v5.8h, v5.8b, #0                \n"
+    "ssubl     v16.4s, v5.4h, v4.4h            \n"
+    "ssubl2    v17.4s, v5.8h, v4.8h            \n"
+    "ushll     v7.4s, v6.4h, #0                \n"
+    "ushll2    v6.4s, v6.8h, #0                \n"
+    "mul       v16.4s, v16.4s, v7.4s           \n"
+    "mul       v17.4s, v17.4s, v6.4s           \n"
+    "shrn      v6.4h, v16.4s, #16              \n"
+    "shrn2     v6.8h, v17.4s, #16              \n"
+    "add       v4.8h, v4.8h, v6.8h             \n"
+    "xtn       v4.8b, v4.8h                    \n"
+
+    MEMACCESS(0)
+    "st1       {v4.8b}, [%0], #8               \n"  // store pixels
+    "add       v1.4s, v1.4s, v0.4s             \n"
+    "add       v2.4s, v2.4s, v0.4s             \n"
+    "subs      %w2, %w2, #8                    \n"  // 8 processed per loop
+    "b.gt      1b                              \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_width64),      // %2
+    "+r"(x64),              // %3
+    "+r"(dx64),             // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "v0", "v1", "v2", "v3",
+    "v4", "v5", "v6", "v7", "v16", "v17"
+  );
+}
+
+#undef LOAD2_DATA8_LANE
+
+// 16x2 -> 16x1
+void ScaleFilterRows_NEON(uint8* dst_ptr,
+                          const uint8* src_ptr, ptrdiff_t src_stride,
+                          int dst_width, int source_y_fraction) {
+    int y_fraction = 256 - source_y_fraction;
+  asm volatile (
+    "cmp          %w4, #0                      \n"
+    "b.eq         100f                         \n"
+    "add          %2, %2, %1                   \n"
+    "cmp          %w4, #64                     \n"
+    "b.eq         75f                          \n"
+    "cmp          %w4, #128                    \n"
+    "b.eq         50f                          \n"
+    "cmp          %w4, #192                    \n"
+    "b.eq         25f                          \n"
+
+    "dup          v5.8b, %w4                   \n"
+    "dup          v4.8b, %w5                   \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "ld1          {v0.16b}, [%1], #16          \n"
+    MEMACCESS(2)
+    "ld1          {v1.16b}, [%2], #16          \n"
+    "subs         %w3, %w3, #16                \n"
+    "umull        v6.8h, v0.8b, v4.8b          \n"
+    "umull2       v7.8h, v0.16b, v4.16b        \n"
+    "umlal        v6.8h, v1.8b, v5.8b          \n"
+    "umlal2       v7.8h, v1.16b, v5.16b        \n"
+    "rshrn        v0.8b, v6.8h, #8             \n"
+    "rshrn2       v0.16b, v7.8h, #8            \n"
+    MEMACCESS(0)
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         1b                           \n"
+    "b            99f                          \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "ld1          {v0.16b}, [%1], #16          \n"
+    MEMACCESS(2)
+    "ld1          {v1.16b}, [%2], #16          \n"
+    "subs         %w3, %w3, #16                \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    MEMACCESS(0)
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         25b                          \n"
+    "b            99f                          \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "ld1          {v0.16b}, [%1], #16          \n"
+    MEMACCESS(2)
+    "ld1          {v1.16b}, [%2], #16          \n"
+    "subs         %w3, %w3, #16                \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    MEMACCESS(0)
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         50b                          \n"
+    "b            99f                          \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "ld1          {v1.16b}, [%1], #16          \n"
+    MEMACCESS(2)
+    "ld1          {v0.16b}, [%2], #16          \n"
+    "subs         %w3, %w3, #16                \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    MEMACCESS(0)
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         75b                          \n"
+    "b            99f                          \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "ld1          {v0.16b}, [%1], #16          \n"
+    "subs         %w3, %w3, #16                \n"
+    MEMACCESS(0)
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         100b                         \n"
+
+  "99:                                         \n"
+    MEMACCESS(0)
+    "st1          {v0.b}[15], [%0]             \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction),// %4
+    "+r"(y_fraction)        // %5
+  :
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "memory", "cc"
+  );
+}
+
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    MEMACCESS (0)
+    "ld2        {v0.4s, v1.4s}, [%0], #32      \n"
+    MEMACCESS (0)
+    "ld2        {v2.4s, v3.4s}, [%0], #32      \n"
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
+    MEMACCESS (1)
+    "st1        {v1.16b}, [%1], #16            \n"  // store odd pixels
+    MEMACCESS (1)
+    "st1        {v3.16b}, [%1], #16            \n"
+    "b.gt       1b                             \n"
+  : "+r" (src_ptr),          // %0
+    "+r" (dst),              // %1
+    "+r" (dst_width)         // %2
+  :
+  : "memory", "cc", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS (0)
+    // load 8 ARGB pixels.
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64   \n"
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    "uaddlp     v3.8h, v3.16b                  \n"  // A 16 bytes -> 8 shorts.
+    "rshrn      v0.8b, v0.8h, #1               \n"  // downshift, round and pack
+    "rshrn      v1.8b, v1.8h, #1               \n"
+    "rshrn      v2.8b, v2.8h, #1               \n"
+    "rshrn      v3.8b, v3.8h, #1               \n"
+    MEMACCESS (1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32     \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),         // %0
+    "+r"(dst_argb),         // %1
+    "+r"(dst_width)         // %2
+  :
+  : "memory", "cc", "v0", "v1", "v2", "v3"    // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %1, %0                     \n"
+  "1:                                          \n"
+    MEMACCESS (0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64   \n"  // load 8 ARGB pixels.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    "uaddlp     v3.8h, v3.16b                  \n"  // A 16 bytes -> 8 shorts.
+    MEMACCESS (1)
+    "ld4        {v16.16b,v17.16b,v18.16b,v19.16b}, [%1], #64 \n"  // load 8 more ARGB pixels.
+    "uadalp     v0.8h, v16.16b                 \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v17.16b                 \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v18.16b                 \n"  // R 16 bytes -> 8 shorts.
+    "uadalp     v3.8h, v19.16b                 \n"  // A 16 bytes -> 8 shorts.
+    "rshrn      v0.8b, v0.8h, #2               \n"  // downshift, round and pack
+    "rshrn      v1.8b, v1.8h, #2               \n"
+    "rshrn      v2.8b, v2.8h, #2               \n"
+    "rshrn      v3.8b, v3.8h, #2               \n"
+    MEMACCESS (2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32     \n"
+    "b.gt       1b                             \n"
+  : "+r" (src_ptr),          // %0
+    "+r" (src_stride),       // %1
+    "+r" (dst),              // %2
+    "+r" (dst_width)         // %3
+  :
+  : "memory", "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.s}[0], [%0], %3            \n"
+    MEMACCESS(0)
+    "ld1        {v0.s}[1], [%0], %3            \n"
+    MEMACCESS(0)
+    "ld1        {v0.s}[2], [%0], %3            \n"
+    MEMACCESS(0)
+    "ld1        {v0.s}[3], [%0], %3            \n"
+    "subs       %w2, %w2, #4                   \n"  // 4 pixels per loop.
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(dst_width)    // %2
+  : "r"((int64)(src_stepx * 4)) // %3
+  : "memory", "cc", "v0"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+// TODO(Yang Zhang): Might be worth another optimization pass in future.
+// It could be upgraded to 8 pixels at a time to start with.
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "add        %1, %1, %0                     \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.8b}, [%0], %4              \n"  // Read 4 2x2 blocks -> 2x1
+    MEMACCESS(1)
+    "ld1        {v1.8b}, [%1], %4              \n"
+    MEMACCESS(0)
+    "ld1        {v2.8b}, [%0], %4              \n"
+    MEMACCESS(1)
+    "ld1        {v3.8b}, [%1], %4              \n"
+    MEMACCESS(0)
+    "ld1        {v4.8b}, [%0], %4              \n"
+    MEMACCESS(1)
+    "ld1        {v5.8b}, [%1], %4              \n"
+    MEMACCESS(0)
+    "ld1        {v6.8b}, [%0], %4              \n"
+    MEMACCESS(1)
+    "ld1        {v7.8b}, [%1], %4              \n"
+    "uaddl      v0.8h, v0.8b, v1.8b            \n"
+    "uaddl      v2.8h, v2.8b, v3.8b            \n"
+    "uaddl      v4.8h, v4.8b, v5.8b            \n"
+    "uaddl      v6.8h, v6.8b, v7.8b            \n"
+    "mov        v16.d[1], v0.d[1]              \n"  // ab_cd -> ac_bd
+    "mov        v0.d[1], v2.d[0]               \n"
+    "mov        v2.d[0], v16.d[1]              \n"
+    "mov        v16.d[1], v4.d[1]              \n"  // ef_gh -> eg_fh
+    "mov        v4.d[1], v6.d[0]               \n"
+    "mov        v6.d[0], v16.d[1]              \n"
+    "add        v0.8h, v0.8h, v2.8h            \n"  // (a+b)_(c+d)
+    "add        v4.8h, v4.8h, v6.8h            \n"  // (e+f)_(g+h)
+    "rshrn      v0.8b, v0.8h, #2               \n"  // first 2 pixels.
+    "rshrn2     v0.16b, v4.8h, #2              \n"  // next 2 pixels.
+    "subs       %w3, %w3, #4                   \n"  // 4 pixels per loop.
+    MEMACCESS(2)
+    "st1     {v0.16b}, [%2], #16               \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(src_stride),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(dst_width)    // %3
+  : "r"((int64)(src_stepx * 4)) // %4
+  : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD1_DATA32_LANE(vn, n)                               \
+    "lsr        %5, %3, #16                    \n"             \
+    "add        %6, %1, %5, lsl #2             \n"             \
+    "add        %3, %3, %4                     \n"             \
+    MEMACCESS(6)                                               \
+    "ld1        {"#vn".s}["#n"], [%6]          \n"
+
+void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  const uint8* src_tmp = src_argb;
+  int64 dst_width64 = (int64) dst_width;  // Work around ios 64 bit warning.
+  int64 x64 = (int64) x;
+  int64 dx64 = (int64) dx;
+  int64 tmp64 = 0;
+  asm volatile (
+  "1:                                          \n"
+    LOAD1_DATA32_LANE(v0, 0)
+    LOAD1_DATA32_LANE(v0, 1)
+    LOAD1_DATA32_LANE(v0, 2)
+    LOAD1_DATA32_LANE(v0, 3)
+    LOAD1_DATA32_LANE(v1, 0)
+    LOAD1_DATA32_LANE(v1, 1)
+    LOAD1_DATA32_LANE(v1, 2)
+    LOAD1_DATA32_LANE(v1, 3)
+
+    MEMACCESS(0)
+    "st1        {v0.4s, v1.4s}, [%0], #32      \n"  // store pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
+    "b.gt        1b                            \n"
+  : "+r"(dst_argb),         // %0
+    "+r"(src_argb),         // %1
+    "+r"(dst_width64),      // %2
+    "+r"(x64),              // %3
+    "+r"(dx64),             // %4
+    "+r"(tmp64),            // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "v0", "v1"
+  );
+}
+
+#undef LOAD1_DATA32_LANE
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD2_DATA32_LANE(vn1, vn2, n)                         \
+    "lsr        %5, %3, #16                           \n"      \
+    "add        %6, %1, %5, lsl #2                    \n"      \
+    "add        %3, %3, %4                            \n"      \
+    MEMACCESS(6)                                               \
+    "ld2        {"#vn1".s, "#vn2".s}["#n"], [%6]      \n"
+
+void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                              int dst_width, int x, int dx) {
+  int dx_offset[4] = {0, 1, 2, 3};
+  int* tmp = dx_offset;
+  const uint8* src_tmp = src_argb;
+  int64 dst_width64 = (int64) dst_width;  // Work around ios 64 bit warning.
+  int64 x64 = (int64) x;
+  int64 dx64 = (int64) dx;
+  asm volatile (
+    "dup        v0.4s, %w3                     \n"  // x
+    "dup        v1.4s, %w4                     \n"  // dx
+    "ld1        {v2.4s}, [%5]                  \n"  // 0 1 2 3
+    "shl        v6.4s, v1.4s, #2               \n"  // 4 * dx
+    "mul        v1.4s, v1.4s, v2.4s            \n"
+    "movi       v3.16b, #0x7f                  \n"  // 0x7F
+    "movi       v4.8h, #0x7f                   \n"  // 0x7F
+    // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
+    "add        v5.4s, v1.4s, v0.4s            \n"
+  "1:                                          \n"
+    // d0, d1: a
+    // d2, d3: b
+    LOAD2_DATA32_LANE(v0, v1, 0)
+    LOAD2_DATA32_LANE(v0, v1, 1)
+    LOAD2_DATA32_LANE(v0, v1, 2)
+    LOAD2_DATA32_LANE(v0, v1, 3)
+    "shrn       v2.4h, v5.4s, #9               \n"
+    "and        v2.8b, v2.8b, v4.8b            \n"
+    "dup        v16.8b, v2.b[0]                \n"
+    "dup        v17.8b, v2.b[2]                \n"
+    "dup        v18.8b, v2.b[4]                \n"
+    "dup        v19.8b, v2.b[6]                \n"
+    "ext        v2.8b, v16.8b, v17.8b, #4      \n"
+    "ext        v17.8b, v18.8b, v19.8b, #4     \n"
+    "ins        v2.d[1], v17.d[0]              \n"  // f
+    "eor        v7.16b, v2.16b, v3.16b         \n"  // 0x7f ^ f
+    "umull      v16.8h, v0.8b, v7.8b           \n"
+    "umull2     v17.8h, v0.16b, v7.16b         \n"
+    "umull      v18.8h, v1.8b, v2.8b           \n"
+    "umull2     v19.8h, v1.16b, v2.16b         \n"
+    "add        v16.8h, v16.8h, v18.8h         \n"
+    "add        v17.8h, v17.8h, v19.8h         \n"
+    "shrn       v0.8b, v16.8h, #7              \n"
+    "shrn2      v0.16b, v17.8h, #7             \n"
+
+    MEMACCESS(0)
+    "st1     {v0.4s}, [%0], #16                \n"  // store pixels
+    "add     v5.4s, v5.4s, v6.4s               \n"
+    "subs    %w2, %w2, #4                      \n"  // 4 processed per loop
+    "b.gt    1b                                \n"
+  : "+r"(dst_argb),         // %0
+    "+r"(src_argb),         // %1
+    "+r"(dst_width64),      // %2
+    "+r"(x64),              // %3
+    "+r"(dx64),             // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5",
+    "v6", "v7", "v16", "v17", "v18", "v19"
+  );
+}
+
+#undef LOAD2_DATA32_LANE
+
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/scale_win.cc b/libvpx/libvpx/third_party/libyuv/source/scale_win.cc
new file mode 100644
index 0000000..c3896eb
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/scale_win.cc
@@ -0,0 +1,1354 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
+    defined(_MSC_VER) && !defined(__clang__)
+
+// Offsets for source bytes 0 to 9
+static uvec8 kShuf0 =
+  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
+static uvec8 kShuf1 =
+  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf2 =
+  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 0 to 10
+static uvec8 kShuf01 =
+  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+
+// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
+static uvec8 kShuf11 =
+  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf21 =
+  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+
+// Coefficients for source bytes 0 to 10
+static uvec8 kMadd01 =
+  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+
+// Coefficients for source bytes 10 to 21
+static uvec8 kMadd11 =
+  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+
+// Coefficients for source bytes 21 to 31
+static uvec8 kMadd21 =
+  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+
+// Coefficients for source bytes 21 to 31
+static vec16 kRound34 =
+  { 2, 2, 2, 2, 2, 2, 2, 2 };
+
+static uvec8 kShuf38a =
+  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+static uvec8 kShuf38b =
+  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 0,1,2
+static uvec8 kShufAc =
+  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 3,4,5
+static uvec8 kShufAc3 =
+  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x3 and 2x3
+static uvec16 kScaleAc33 =
+  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+
+// Arrange first value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb0 =
+  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+
+// Arrange second value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb1 =
+  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+
+// Arrange third value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb2 =
+  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x2 and 2x2
+static uvec16 kScaleAb2 =
+  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+
+// Reads 32 pixels, throws half away and writes 16 pixels.
+__declspec(naked)
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8               // isolate odd pixels.
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x1 rectangle to 16x1.
+__declspec(naked)
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 16x1.
+__declspec(naked)
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_SCALEROWDOWN2_AVX2
+// Reads 64 pixels, throws half away and writes 32 pixels.
+__declspec(naked)
+void ScaleRowDown2_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+
+  wloop:
+    vmovdqu     ymm0, [eax]
+    vmovdqu     ymm1, [eax + 32]
+    lea         eax,  [eax + 64]
+    vpsrlw      ymm0, ymm0, 8        // isolate odd pixels.
+    vpsrlw      ymm1, ymm1, 8
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8     // unmutate vpackuswb
+    vmovdqu     [edx], ymm0
+    lea         edx, [edx + 32]
+    sub         ecx, 32
+    jg          wloop
+
+    vzeroupper
+    ret
+  }
+}
+
+// Blends 64x1 rectangle to 32x1.
+__declspec(naked)
+void ScaleRowDown2Linear_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov         eax, [esp + 4]        // src_ptr
+                                      // src_stride
+    mov         edx, [esp + 12]       // dst_ptr
+    mov         ecx, [esp + 16]       // dst_width
+
+    vpcmpeqb    ymm4, ymm4, ymm4      // '1' constant, 8b
+    vpsrlw      ymm4, ymm4, 15
+    vpackuswb   ymm4, ymm4, ymm4
+    vpxor       ymm5, ymm5, ymm5      // constant 0
+
+  wloop:
+    vmovdqu     ymm0, [eax]
+    vmovdqu     ymm1, [eax + 32]
+    lea         eax,  [eax + 64]
+
+    vpmaddubsw  ymm0, ymm0, ymm4      // average horizontally
+    vpmaddubsw  ymm1, ymm1, ymm4
+    vpavgw      ymm0, ymm0, ymm5      // (x + 1) / 2
+    vpavgw      ymm1, ymm1, ymm5
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+
+    vmovdqu     [edx], ymm0
+    lea         edx, [edx + 32]
+    sub         ecx, 32
+    jg          wloop
+
+    vzeroupper
+    ret
+  }
+}
+
+// Blends 64x2 rectangle to 32x1.
+__declspec(naked)
+void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push        esi
+    mov         eax, [esp + 4 + 4]    // src_ptr
+    mov         esi, [esp + 4 + 8]    // src_stride
+    mov         edx, [esp + 4 + 12]   // dst_ptr
+    mov         ecx, [esp + 4 + 16]   // dst_width
+
+    vpcmpeqb    ymm4, ymm4, ymm4      // '1' constant, 8b
+    vpsrlw      ymm4, ymm4, 15
+    vpackuswb   ymm4, ymm4, ymm4
+    vpxor       ymm5, ymm5, ymm5      // constant 0
+
+  wloop:
+    vmovdqu     ymm0, [eax]           // average rows
+    vmovdqu     ymm1, [eax + 32]
+    vpavgb      ymm0, ymm0, [eax + esi]
+    vpavgb      ymm1, ymm1, [eax + esi + 32]
+    lea         eax,  [eax + 64]
+
+    vpmaddubsw  ymm0, ymm0, ymm4      // average horizontally
+    vpmaddubsw  ymm1, ymm1, ymm4
+    vpavgw      ymm0, ymm0, ymm5      // (x + 1) / 2
+    vpavgw      ymm1, ymm1, ymm5
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+
+    vmovdqu     [edx], ymm0
+    lea         edx, [edx + 32]
+    sub         ecx, 32
+    jg          wloop
+
+    pop         esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SCALEROWDOWN2_AVX2
+
+// Point samples 32 pixels to 8 pixels.
+__declspec(naked)
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff0000
+    psrld      xmm5, 24
+    pslld      xmm5, 16
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm0, 8
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    sub        ecx, 8
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x4 rectangle to 8x1.
+__declspec(naked)
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_ptr
+    mov        esi, [esp + 8 + 8]    // src_stride
+    mov        edx, [esp + 8 + 12]   // dst_ptr
+    mov        ecx, [esp + 8 + 16]   // dst_width
+    lea        edi, [esi + esi * 2]  // src_stride * 3
+    pcmpeqb    xmm7, xmm7            // generate mask 0x00ff00ff
+    psrlw      xmm7, 8
+
+  wloop:
+    movdqu     xmm0, [eax]           // average rows
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    movdqu     xmm2, [eax + esi * 2]
+    movdqu     xmm3, [eax + esi * 2 + 16]
+    movdqu     xmm4, [eax + edi]
+    movdqu     xmm5, [eax + edi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm2, xmm4
+    pavgb      xmm3, xmm5
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm7
+    pand       xmm3, xmm7
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    movdqa     xmm2, xmm0            // average columns (16 to 8 pixels)
+    psrlw      xmm0, 8
+    pand       xmm2, xmm7
+    pavgw      xmm0, xmm2
+    packuswb   xmm0, xmm0
+
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    sub        ecx, 8
+    jg         wloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_SCALEROWDOWN4_AVX2
+// Point samples 64 pixels to 16 pixels.
+__declspec(naked)
+void ScaleRowDown4_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov         eax, [esp + 4]        // src_ptr
+                                      // src_stride ignored
+    mov         edx, [esp + 12]       // dst_ptr
+    mov         ecx, [esp + 16]       // dst_width
+    vpcmpeqb    ymm5, ymm5, ymm5      // generate mask 0x00ff0000
+    vpsrld      ymm5, ymm5, 24
+    vpslld      ymm5, ymm5, 16
+
+  wloop:
+    vmovdqu     ymm0, [eax]
+    vmovdqu     ymm1, [eax + 32]
+    lea         eax,  [eax + 64]
+    vpand       ymm0, ymm0, ymm5
+    vpand       ymm1, ymm1, ymm5
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+    vpsrlw      ymm0, ymm0, 8
+    vpackuswb   ymm0, ymm0, ymm0
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+    vmovdqu     [edx], xmm0
+    lea         edx, [edx + 16]
+    sub         ecx, 16
+    jg          wloop
+
+    vzeroupper
+    ret
+  }
+}
+
+// Blends 64x4 rectangle to 16x1.
+__declspec(naked)
+void ScaleRowDown4Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push        esi
+    push        edi
+    mov         eax, [esp + 8 + 4]    // src_ptr
+    mov         esi, [esp + 8 + 8]    // src_stride
+    mov         edx, [esp + 8 + 12]   // dst_ptr
+    mov         ecx, [esp + 8 + 16]   // dst_width
+    lea         edi, [esi + esi * 2]  // src_stride * 3
+    vpcmpeqb    ymm7, ymm7, ymm7      // generate mask 0x00ff00ff
+    vpsrlw      ymm7, ymm7, 8
+
+  wloop:
+    vmovdqu     ymm0, [eax]           // average rows
+    vmovdqu     ymm1, [eax + 32]
+    vpavgb      ymm0, ymm0, [eax + esi]
+    vpavgb      ymm1, ymm1, [eax + esi + 32]
+    vmovdqu     ymm2, [eax + esi * 2]
+    vmovdqu     ymm3, [eax + esi * 2 + 32]
+    vpavgb      ymm2, ymm2, [eax + edi]
+    vpavgb      ymm3, ymm3, [eax + edi + 32]
+    lea         eax, [eax + 64]
+    vpavgb      ymm0, ymm0, ymm2
+    vpavgb      ymm1, ymm1, ymm3
+
+    vpand       ymm2, ymm0, ymm7      // average columns (64 to 32 pixels)
+    vpand       ymm3, ymm1, ymm7
+    vpsrlw      ymm0, ymm0, 8
+    vpsrlw      ymm1, ymm1, 8
+    vpavgw      ymm0, ymm0, ymm2
+    vpavgw      ymm1, ymm1, ymm3
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+
+    vpand       ymm2, ymm0, ymm7      // average columns (32 to 16 pixels)
+    vpsrlw      ymm0, ymm0, 8
+    vpavgw      ymm0, ymm0, ymm2
+    vpackuswb   ymm0, ymm0, ymm0
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+
+    vmovdqu     [edx], xmm0
+    lea         edx, [edx + 16]
+    sub         ecx, 16
+    jg          wloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SCALEROWDOWN4_AVX2
+
+// Point samples 32 pixels to 24 pixels.
+// Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
+// Then shuffled to do the scaling.
+
+__declspec(naked)
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    movdqa     xmm3, kShuf0
+    movdqa     xmm4, kShuf1
+    movdqa     xmm5, kShuf2
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm1
+    palignr    xmm1, xmm0, 8
+    pshufb     xmm0, xmm3
+    pshufb     xmm1, xmm4
+    pshufb     xmm2, xmm5
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + 8], xmm1
+    movq       qword ptr [edx + 16], xmm2
+    lea        edx, [edx + 24]
+    sub        ecx, 24
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 24x1
+// Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
+// Then shuffled to do the scaling.
+
+// Register usage:
+// xmm0 src_row 0
+// xmm1 src_row 1
+// xmm2 shuf 0
+// xmm3 shuf 1
+// xmm4 shuf 2
+// xmm5 madd 0
+// xmm6 madd 1
+// xmm7 kRound34
+
+// Note that movdqa+palign may be better than movdqu.
+__declspec(naked)
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShuf01
+    movdqa     xmm3, kShuf11
+    movdqa     xmm4, kShuf21
+    movdqa     xmm5, kMadd01
+    movdqa     xmm6, kMadd11
+    movdqa     xmm7, kRound34
+
+  wloop:
+    movdqu     xmm0, [eax]           // pixels 0..7
+    movdqu     xmm1, [eax + esi]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm2
+    pmaddubsw  xmm0, xmm5
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm1, [eax + esi + 8]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm3
+    pmaddubsw  xmm0, xmm6
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 8], xmm0
+    movdqu     xmm0, [eax + 16]      // pixels 16..23
+    movdqu     xmm1, [eax + esi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm4
+    movdqa     xmm1, kMadd21
+    pmaddubsw  xmm0, xmm1
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 16], xmm0
+    lea        edx, [edx + 24]
+    sub        ecx, 24
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Note that movdqa+palign may be better than movdqu.
+__declspec(naked)
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShuf01
+    movdqa     xmm3, kShuf11
+    movdqa     xmm4, kShuf21
+    movdqa     xmm5, kMadd01
+    movdqa     xmm6, kMadd11
+    movdqa     xmm7, kRound34
+
+  wloop:
+    movdqu     xmm0, [eax]           // pixels 0..7
+    movdqu     xmm1, [eax + esi]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm2
+    pmaddubsw  xmm0, xmm5
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm1, [eax + esi + 8]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm3
+    pmaddubsw  xmm0, xmm6
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 8], xmm0
+    movdqu     xmm0, [eax + 16]      // pixels 16..23
+    movdqu     xmm1, [eax + esi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm4
+    movdqa     xmm1, kMadd21
+    pmaddubsw  xmm0, xmm1
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 16], xmm0
+    lea        edx, [edx+24]
+    sub        ecx, 24
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 3/8 point sampler
+
+// Scale 32 pixels to 12
+__declspec(naked)
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    movdqa     xmm4, kShuf38a
+    movdqa     xmm5, kShuf38b
+
+  xloop:
+    movdqu     xmm0, [eax]           // 16 pixels -> 0,1,2,3,4,5
+    movdqu     xmm1, [eax + 16]      // 16 pixels -> 6,7,8,9,10,11
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm4
+    pshufb     xmm1, xmm5
+    paddusb    xmm0, xmm1
+
+    movq       qword ptr [edx], xmm0  // write 12 pixels
+    movhlps    xmm1, xmm0
+    movd       [edx + 8], xmm1
+    lea        edx, [edx + 12]
+    sub        ecx, 12
+    jg         xloop
+
+    ret
+  }
+}
+
+// Scale 16x3 pixels to 6x1 with interpolation
+__declspec(naked)
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShufAc
+    movdqa     xmm3, kShufAc3
+    movdqa     xmm4, kScaleAc33
+    pxor       xmm5, xmm5
+
+  xloop:
+    movdqu     xmm0, [eax]           // sum up 3 rows into xmm0/1
+    movdqu     xmm6, [eax + esi]
+    movhlps    xmm1, xmm0
+    movhlps    xmm7, xmm6
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm6, xmm5
+    punpcklbw  xmm7, xmm5
+    paddusw    xmm0, xmm6
+    paddusw    xmm1, xmm7
+    movdqu     xmm6, [eax + esi * 2]
+    lea        eax, [eax + 16]
+    movhlps    xmm7, xmm6
+    punpcklbw  xmm6, xmm5
+    punpcklbw  xmm7, xmm5
+    paddusw    xmm0, xmm6
+    paddusw    xmm1, xmm7
+
+    movdqa     xmm6, xmm0            // 8 pixels -> 0,1,2 of xmm6
+    psrldq     xmm0, 2
+    paddusw    xmm6, xmm0
+    psrldq     xmm0, 2
+    paddusw    xmm6, xmm0
+    pshufb     xmm6, xmm2
+
+    movdqa     xmm7, xmm1            // 8 pixels -> 3,4,5 of xmm6
+    psrldq     xmm1, 2
+    paddusw    xmm7, xmm1
+    psrldq     xmm1, 2
+    paddusw    xmm7, xmm1
+    pshufb     xmm7, xmm3
+    paddusw    xmm6, xmm7
+
+    pmulhuw    xmm6, xmm4            // divide by 9,9,6, 9,9,6
+    packuswb   xmm6, xmm6
+
+    movd       [edx], xmm6           // write 6 pixels
+    psrlq      xmm6, 16
+    movd       [edx + 2], xmm6
+    lea        edx, [edx + 6]
+    sub        ecx, 6
+    jg         xloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Scale 16x2 pixels to 6x1 with interpolation
+__declspec(naked)
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShufAb0
+    movdqa     xmm3, kShufAb1
+    movdqa     xmm4, kShufAb2
+    movdqa     xmm5, kScaleAb2
+
+  xloop:
+    movdqu     xmm0, [eax]           // average 2 rows into xmm0
+    movdqu     xmm1, [eax + esi]
+    lea        eax, [eax + 16]
+    pavgb      xmm0, xmm1
+
+    movdqa     xmm1, xmm0            // 16 pixels -> 0,1,2,3,4,5 of xmm1
+    pshufb     xmm1, xmm2
+    movdqa     xmm6, xmm0
+    pshufb     xmm6, xmm3
+    paddusw    xmm1, xmm6
+    pshufb     xmm0, xmm4
+    paddusw    xmm1, xmm0
+
+    pmulhuw    xmm1, xmm5            // divide by 3,3,2, 3,3,2
+    packuswb   xmm1, xmm1
+
+    movd       [edx], xmm1           // write 6 pixels
+    psrlq      xmm1, 16
+    movd       [edx + 2], xmm1
+    lea        edx, [edx + 6]
+    sub        ecx, 6
+    jg         xloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 16 bytes and accumulates to 16 shorts at a time.
+__declspec(naked)
+void ScaleAddRow_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_ptr
+    mov        edx, [esp + 8]   // dst_ptr
+    mov        ecx, [esp + 12]  // src_width
+    pxor       xmm5, xmm5
+
+  // sum rows
+  xloop:
+    movdqu     xmm3, [eax]       // read 16 bytes
+    lea        eax, [eax + 16]
+    movdqu     xmm0, [edx]       // read 16 words from destination
+    movdqu     xmm1, [edx + 16]
+    movdqa     xmm2, xmm3
+    punpcklbw  xmm2, xmm5
+    punpckhbw  xmm3, xmm5
+    paddusw    xmm0, xmm2        // sum 16 words
+    paddusw    xmm1, xmm3
+    movdqu     [edx], xmm0       // write 16 words to destination
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 16
+    jg         xloop
+    ret
+  }
+}
+
+#ifdef HAS_SCALEADDROW_AVX2
+// Reads 32 bytes and accumulates to 32 shorts at a time.
+__declspec(naked)
+void ScaleAddRow_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
+  __asm {
+    mov         eax, [esp + 4]   // src_ptr
+    mov         edx, [esp + 8]   // dst_ptr
+    mov         ecx, [esp + 12]  // src_width
+    vpxor       ymm5, ymm5, ymm5
+
+  // sum rows
+  xloop:
+    vmovdqu     ymm3, [eax]       // read 32 bytes
+    lea         eax, [eax + 32]
+    vpermq      ymm3, ymm3, 0xd8  // unmutate for vpunpck
+    vpunpcklbw  ymm2, ymm3, ymm5
+    vpunpckhbw  ymm3, ymm3, ymm5
+    vpaddusw    ymm0, ymm2, [edx] // sum 16 words
+    vpaddusw    ymm1, ymm3, [edx + 32]
+    vmovdqu     [edx], ymm0       // write 32 words to destination
+    vmovdqu     [edx + 32], ymm1
+    lea         edx, [edx + 64]
+    sub         ecx, 32
+    jg          xloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SCALEADDROW_AVX2
+
+// Bilinear column filtering. SSSE3 version.
+__declspec(naked)
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        edi, [esp + 12 + 4]    // dst_ptr
+    mov        esi, [esp + 12 + 8]    // src_ptr
+    mov        ecx, [esp + 12 + 12]   // dst_width
+    movd       xmm2, [esp + 12 + 16]  // x
+    movd       xmm3, [esp + 12 + 20]  // dx
+    mov        eax, 0x04040000      // shuffle to line up fractions with pixel.
+    movd       xmm5, eax
+    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    psrlw      xmm6, 9
+    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    sub        ecx, 2
+    jl         xloop29
+
+    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    paddd      xmm0, xmm3
+    punpckldq  xmm2, xmm0           // x0 x1
+    punpckldq  xmm3, xmm3           // dx dx
+    paddd      xmm3, xmm3           // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+
+    // 2 Pixel loop.
+  xloop2:
+    movdqa     xmm1, xmm2           // x0, x1 fractions.
+    paddd      xmm2, xmm3           // x += dx
+    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+    movd       xmm0, ebx
+    psrlw      xmm1, 9              // 7 bit fractions.
+    movzx      ebx, word ptr [esi + edx]  // 2 source x1 pixels
+    movd       xmm4, ebx
+    pshufb     xmm1, xmm5           // 0011
+    punpcklwd  xmm0, xmm4
+    pxor       xmm1, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1           // 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    psrlw      xmm0, 7              // 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // 8 bits, 2 pixels.
+    movd       ebx, xmm0
+    mov        [edi], bx
+    lea        edi, [edi + 2]
+    sub        ecx, 2               // 2 pixels
+    jge        xloop2
+
+ xloop29:
+
+    add        ecx, 2 - 1
+    jl         xloop99
+
+    // 1 pixel remainder
+    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+    movd       xmm0, ebx
+    psrlw      xmm2, 9              // 7 bit fractions.
+    pshufb     xmm2, xmm5           // 0011
+    pxor       xmm2, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2           // 16 bit
+    psrlw      xmm0, 7              // 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // 8 bits
+    movd       ebx, xmm0
+    mov        [edi], bl
+
+ xloop99:
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// Reads 16 pixels, duplicates them and writes 32 pixels.
+__declspec(naked)
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  __asm {
+    mov        edx, [esp + 4]    // dst_ptr
+    mov        eax, [esp + 8]    // src_ptr
+    mov        ecx, [esp + 12]   // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm0
+    punpckhbw  xmm1, xmm1
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         wloop
+
+    ret
+  }
+}
+
+// Reads 8 pixels, throws half away and writes 4 even pixels (0, 2, 4, 6)
+__declspec(naked)
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_argb
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_argb
+    mov        ecx, [esp + 16]       // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    shufps     xmm0, xmm1, 0xdd
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 8x1 rectangle to 4x1.
+__declspec(naked)
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_argb
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_argb
+    mov        ecx, [esp + 16]       // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 8x2 rectangle to 4x1.
+__declspec(naked)
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_argb
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_argb
+    mov        ecx, [esp + 4 + 16]   // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 4 pixels at a time.
+__declspec(naked)
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  __asm {
+    push       ebx
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_argb
+                                     // src_stride ignored
+    mov        ebx, [esp + 8 + 12]   // src_stepx
+    mov        edx, [esp + 8 + 16]   // dst_argb
+    mov        ecx, [esp + 8 + 20]   // dst_width
+    lea        ebx, [ebx * 4]
+    lea        edi, [ebx + ebx * 2]
+
+  wloop:
+    movd       xmm0, [eax]
+    movd       xmm1, [eax + ebx]
+    punpckldq  xmm0, xmm1
+    movd       xmm2, [eax + ebx * 2]
+    movd       xmm3, [eax + edi]
+    lea        eax,  [eax + ebx * 4]
+    punpckldq  xmm2, xmm3
+    punpcklqdq xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         wloop
+
+    pop        edi
+    pop        ebx
+    ret
+  }
+}
+
+// Blends four 2x2 to 4x1.
+__declspec(naked)
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]    // src_argb
+    mov        esi, [esp + 12 + 8]    // src_stride
+    mov        ebx, [esp + 12 + 12]   // src_stepx
+    mov        edx, [esp + 12 + 16]   // dst_argb
+    mov        ecx, [esp + 12 + 20]   // dst_width
+    lea        esi, [eax + esi]       // row1 pointer
+    lea        ebx, [ebx * 4]
+    lea        edi, [ebx + ebx * 2]
+
+  wloop:
+    movq       xmm0, qword ptr [eax]  // row0 4 pairs
+    movhps     xmm0, qword ptr [eax + ebx]
+    movq       xmm1, qword ptr [eax + ebx * 2]
+    movhps     xmm1, qword ptr [eax + edi]
+    lea        eax,  [eax + ebx * 4]
+    movq       xmm2, qword ptr [esi]  // row1 4 pairs
+    movhps     xmm2, qword ptr [esi + ebx]
+    movq       xmm3, qword ptr [esi + ebx * 2]
+    movhps     xmm3, qword ptr [esi + edi]
+    lea        esi,  [esi + ebx * 4]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         wloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// Column scaling unfiltered. SSE2 version.
+__declspec(naked)
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  __asm {
+    push       edi
+    push       esi
+    mov        edi, [esp + 8 + 4]    // dst_argb
+    mov        esi, [esp + 8 + 8]    // src_argb
+    mov        ecx, [esp + 8 + 12]   // dst_width
+    movd       xmm2, [esp + 8 + 16]  // x
+    movd       xmm3, [esp + 8 + 20]  // dx
+
+    pshufd     xmm2, xmm2, 0         // x0 x0 x0 x0
+    pshufd     xmm0, xmm3, 0x11      // dx  0 dx  0
+    paddd      xmm2, xmm0
+    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 2
+    pshufd     xmm0, xmm3, 0x05      // dx * 2, dx * 2, 0, 0
+    paddd      xmm2, xmm0            // x3 x2 x1 x0
+    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 4
+    pshufd     xmm3, xmm3, 0         // dx * 4, dx * 4, dx * 4, dx * 4
+
+    pextrw     eax, xmm2, 1          // get x0 integer.
+    pextrw     edx, xmm2, 3          // get x1 integer.
+
+    cmp        ecx, 0
+    jle        xloop99
+    sub        ecx, 4
+    jl         xloop49
+
+    // 4 Pixel loop.
+ xloop4:
+    movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
+    movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
+    pextrw     eax, xmm2, 5           // get x2 integer.
+    pextrw     edx, xmm2, 7           // get x3 integer.
+    paddd      xmm2, xmm3             // x += dx
+    punpckldq  xmm0, xmm1             // x0 x1
+
+    movd       xmm1, [esi + eax * 4]  // 1 source x2 pixels
+    movd       xmm4, [esi + edx * 4]  // 1 source x3 pixels
+    pextrw     eax, xmm2, 1           // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3           // get x1 integer. next iteration.
+    punpckldq  xmm1, xmm4             // x2 x3
+    punpcklqdq xmm0, xmm1             // x0 x1 x2 x3
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4                 // 4 pixels
+    jge        xloop4
+
+ xloop49:
+    test       ecx, 2
+    je         xloop29
+
+    // 2 Pixels.
+    movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
+    movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
+    pextrw     eax, xmm2, 5           // get x2 integer.
+    punpckldq  xmm0, xmm1             // x0 x1
+
+    movq       qword ptr [edi], xmm0
+    lea        edi, [edi + 8]
+
+ xloop29:
+    test       ecx, 1
+    je         xloop99
+
+    // 1 Pixels.
+    movd       xmm0, [esi + eax * 4]  // 1 source x2 pixels
+    movd       dword ptr [edi], xmm0
+ xloop99:
+
+    pop        esi
+    pop        edi
+    ret
+  }
+}
+
+// Bilinear row filtering combines 2x1 -> 1x1. SSSE3 version.
+// TODO(fbarchard): Port to Neon
+
+// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
+static uvec8 kShuffleColARGB = {
+  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
+  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+};
+
+// Shuffle table for duplicating 2 fractions into 8 bytes each
+static uvec8 kShuffleFractions = {
+  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+};
+
+__declspec(naked)
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]    // dst_argb
+    mov        esi, [esp + 8 + 8]    // src_argb
+    mov        ecx, [esp + 8 + 12]   // dst_width
+    movd       xmm2, [esp + 8 + 16]  // x
+    movd       xmm3, [esp + 8 + 20]  // dx
+    movdqa     xmm4, kShuffleColARGB
+    movdqa     xmm5, kShuffleFractions
+    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    psrlw      xmm6, 9
+    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    sub        ecx, 2
+    jl         xloop29
+
+    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    paddd      xmm0, xmm3
+    punpckldq  xmm2, xmm0           // x0 x1
+    punpckldq  xmm3, xmm3           // dx dx
+    paddd      xmm3, xmm3           // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+
+    // 2 Pixel loop.
+  xloop2:
+    movdqa     xmm1, xmm2           // x0, x1 fractions.
+    paddd      xmm2, xmm3           // x += dx
+    movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
+    psrlw      xmm1, 9              // 7 bit fractions.
+    movhps     xmm0, qword ptr [esi + edx * 4]  // 2 source x1 pixels
+    pshufb     xmm1, xmm5           // 0000000011111111
+    pshufb     xmm0, xmm4           // arrange pixels into pairs
+    pxor       xmm1, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1           // argb_argb 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    psrlw      xmm0, 7              // argb 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // argb_argb 8 bits, 2 pixels.
+    movq       qword ptr [edi], xmm0
+    lea        edi, [edi + 8]
+    sub        ecx, 2               // 2 pixels
+    jge        xloop2
+
+ xloop29:
+
+    add        ecx, 2 - 1
+    jl         xloop99
+
+    // 1 pixel remainder
+    psrlw      xmm2, 9              // 7 bit fractions.
+    movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
+    pshufb     xmm2, xmm5           // 00000000
+    pshufb     xmm0, xmm4           // arrange pixels into pairs
+    pxor       xmm2, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2           // argb 16 bit, 1 pixel.
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0           // argb 8 bits, 1 pixel.
+    movd       [edi], xmm0
+
+ xloop99:
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+__declspec(naked)
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  __asm {
+    mov        edx, [esp + 4]    // dst_argb
+    mov        eax, [esp + 8]    // src_argb
+    mov        ecx, [esp + 12]   // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm1, xmm0
+    punpckldq  xmm0, xmm0
+    punpckhdq  xmm1, xmm1
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         wloop
+
+    ret
+  }
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+__declspec(naked)
+int FixedDiv_X86(int num, int div) {
+  __asm {
+    mov        eax, [esp + 4]    // num
+    cdq                          // extend num to 64 bits
+    shld       edx, eax, 16      // 32.16
+    shl        eax, 16
+    idiv       dword ptr [esp + 8]
+    ret
+  }
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+__declspec(naked)
+int FixedDiv1_X86(int num, int div) {
+  __asm {
+    mov        eax, [esp + 4]    // num
+    mov        ecx, [esp + 8]    // denom
+    cdq                          // extend num to 64 bits
+    shld       edx, eax, 16      // 32.16
+    shl        eax, 16
+    sub        eax, 0x00010001
+    sbb        edx, 0
+    sub        ecx, 1
+    idiv       ecx
+    ret
+  }
+}
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/libvpx/libvpx/third_party/libyuv/source/video_common.cc b/libvpx/libvpx/third_party/libyuv/source/video_common.cc
new file mode 100644
index 0000000..379a066
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/video_common.cc
@@ -0,0 +1,64 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ARRAY_SIZE(x) (int)(sizeof(x) / sizeof(x[0]))
+
+struct FourCCAliasEntry {
+  uint32 alias;
+  uint32 canonical;
+};
+
+static const struct FourCCAliasEntry kFourCCAliases[] = {
+  {FOURCC_IYUV, FOURCC_I420},
+  {FOURCC_YU16, FOURCC_I422},
+  {FOURCC_YU24, FOURCC_I444},
+  {FOURCC_YUYV, FOURCC_YUY2},
+  {FOURCC_YUVS, FOURCC_YUY2},  // kCMPixelFormat_422YpCbCr8_yuvs
+  {FOURCC_HDYC, FOURCC_UYVY},
+  {FOURCC_2VUY, FOURCC_UYVY},  // kCMPixelFormat_422YpCbCr8
+  {FOURCC_JPEG, FOURCC_MJPG},  // Note: JPEG has DHT while MJPG does not.
+  {FOURCC_DMB1, FOURCC_MJPG},
+  {FOURCC_BA81, FOURCC_BGGR},  // deprecated.
+  {FOURCC_RGB3, FOURCC_RAW },
+  {FOURCC_BGR3, FOURCC_24BG},
+  {FOURCC_CM32, FOURCC_BGRA},  // kCMPixelFormat_32ARGB
+  {FOURCC_CM24, FOURCC_RAW },  // kCMPixelFormat_24RGB
+  {FOURCC_L555, FOURCC_RGBO},  // kCMPixelFormat_16LE555
+  {FOURCC_L565, FOURCC_RGBP},  // kCMPixelFormat_16LE565
+  {FOURCC_5551, FOURCC_RGBO},  // kCMPixelFormat_16LE5551
+};
+// TODO(fbarchard): Consider mapping kCMPixelFormat_32BGRA to FOURCC_ARGB.
+//  {FOURCC_BGRA, FOURCC_ARGB},  // kCMPixelFormat_32BGRA
+
+LIBYUV_API
+uint32 CanonicalFourCC(uint32 fourcc) {
+  int i;
+  for (i = 0; i < ARRAY_SIZE(kFourCCAliases); ++i) {
+    if (kFourCCAliases[i].alias == fourcc) {
+      return kFourCCAliases[i].canonical;
+    }
+  }
+  // Not an alias, so return it as-is.
+  return fourcc;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/libvpx/libvpx/third_party/libyuv/source/x86inc.asm b/libvpx/libvpx/third_party/libyuv/source/x86inc.asm
new file mode 100644
index 0000000..cb5c32d
--- /dev/null
+++ b/libvpx/libvpx/third_party/libyuv/source/x86inc.asm
@@ -0,0 +1,1136 @@
+;*****************************************************************************
+;* x86inc.asm: x264asm abstraction layer
+;*****************************************************************************
+;* Copyright (C) 2005-2012 x264 project
+;*
+;* Authors: Loren Merritt <lorenm@u.washington.edu>
+;*          Anton Mitrofanov <BugMaster@narod.ru>
+;*          Jason Garrett-Glaser <darkshikari@gmail.com>
+;*          Henrik Gramner <hengar-6@student.ltu.se>
+;*
+;* Permission to use, copy, modify, and/or distribute this software for any
+;* purpose with or without fee is hereby granted, provided that the above
+;* copyright notice and this permission notice appear in all copies.
+;*
+;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+;*****************************************************************************
+
+; This is a header file for the x264ASM assembly language, which uses
+; NASM/YASM syntax combined with a large number of macros to provide easy
+; abstraction between different calling conventions (x86_32, win64, linux64).
+; It also has various other useful features to simplify writing the kind of
+; DSP functions that are most often used in x264.
+
+; Unlike the rest of x264, this file is available under an ISC license, as it
+; has significant usefulness outside of x264 and we want it to be available
+; to the largest audience possible.  Of course, if you modify it for your own
+; purposes to add a new feature, we strongly encourage contributing a patch
+; as this feature might be useful for others as well.  Send patches or ideas
+; to x264-devel@videolan.org .
+
+; Local changes for libyuv:
+; remove %define program_name and references in labels
+; rename cpus to uppercase
+
+%define WIN64  0
+%define UNIX64 0
+%if ARCH_X86_64
+    %ifidn __OUTPUT_FORMAT__,win32
+        %define WIN64  1
+    %elifidn __OUTPUT_FORMAT__,win64
+        %define WIN64  1
+    %else
+        %define UNIX64 1
+    %endif
+%endif
+
+%ifdef PREFIX
+    %define mangle(x) _ %+ x
+%else
+    %define mangle(x) x
+%endif
+
+; Name of the .rodata section.
+; Kludge: Something on OS X fails to align .rodata even given an align attribute,
+; so use a different read-only section.
+%macro SECTION_RODATA 0-1 16
+    %ifidn __OUTPUT_FORMAT__,macho64
+        SECTION .text align=%1
+    %elifidn __OUTPUT_FORMAT__,macho
+        SECTION .text align=%1
+        fakegot:
+    %elifidn __OUTPUT_FORMAT__,aout
+        section .text
+    %else
+        SECTION .rodata align=%1
+    %endif
+%endmacro
+
+; aout does not support align=
+%macro SECTION_TEXT 0-1 16
+    %ifidn __OUTPUT_FORMAT__,aout
+        SECTION .text
+    %else
+        SECTION .text align=%1
+    %endif
+%endmacro
+
+%if WIN64
+    %define PIC
+%elif ARCH_X86_64 == 0
+; x86_32 doesn't require PIC.
+; Some distros prefer shared objects to be PIC, but nothing breaks if
+; the code contains a few textrels, so we'll skip that complexity.
+    %undef PIC
+%endif
+%ifdef PIC
+    default rel
+%endif
+
+; Always use long nops (reduces 0x90 spam in disassembly on x86_32)
+CPU amdnop
+
+; Macros to eliminate most code duplication between x86_32 and x86_64:
+; Currently this works only for leaf functions which load all their arguments
+; into registers at the start, and make no other use of the stack. Luckily that
+; covers most of x264's asm.
+
+; PROLOGUE:
+; %1 = number of arguments. loads them from stack if needed.
+; %2 = number of registers used. pushes callee-saved regs if needed.
+; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
+; %4 = list of names to define to registers
+; PROLOGUE can also be invoked by adding the same options to cglobal
+
+; e.g.
+; cglobal foo, 2,3,0, dst, src, tmp
+; declares a function (foo), taking two args (dst and src) and one local variable (tmp)
+
+; TODO Some functions can use some args directly from the stack. If they're the
+; last args then you can just not declare them, but if they're in the middle
+; we need more flexible macro.
+
+; RET:
+; Pops anything that was pushed by PROLOGUE, and returns.
+
+; REP_RET:
+; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons
+; which are slow when a normal ret follows a branch.
+
+; registers:
+; rN and rNq are the native-size register holding function argument N
+; rNd, rNw, rNb are dword, word, and byte size
+; rNh is the high 8 bits of the word size
+; rNm is the original location of arg N (a register or on the stack), dword
+; rNmp is native size
+
+%macro DECLARE_REG 2-3
+    %define r%1q %2
+    %define r%1d %2d
+    %define r%1w %2w
+    %define r%1b %2b
+    %define r%1h %2h
+    %if %0 == 2
+        %define r%1m  %2d
+        %define r%1mp %2
+    %elif ARCH_X86_64 ; memory
+        %define r%1m [rsp + stack_offset + %3]
+        %define r%1mp qword r %+ %1m
+    %else
+        %define r%1m [esp + stack_offset + %3]
+        %define r%1mp dword r %+ %1m
+    %endif
+    %define r%1  %2
+%endmacro
+
+%macro DECLARE_REG_SIZE 3
+    %define r%1q r%1
+    %define e%1q r%1
+    %define r%1d e%1
+    %define e%1d e%1
+    %define r%1w %1
+    %define e%1w %1
+    %define r%1h %3
+    %define e%1h %3
+    %define r%1b %2
+    %define e%1b %2
+%if ARCH_X86_64 == 0
+    %define r%1  e%1
+%endif
+%endmacro
+
+DECLARE_REG_SIZE ax, al, ah
+DECLARE_REG_SIZE bx, bl, bh
+DECLARE_REG_SIZE cx, cl, ch
+DECLARE_REG_SIZE dx, dl, dh
+DECLARE_REG_SIZE si, sil, null
+DECLARE_REG_SIZE di, dil, null
+DECLARE_REG_SIZE bp, bpl, null
+
+; t# defines for when per-arch register allocation is more complex than just function arguments
+
+%macro DECLARE_REG_TMP 1-*
+    %assign %%i 0
+    %rep %0
+        CAT_XDEFINE t, %%i, r%1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro DECLARE_REG_TMP_SIZE 0-*
+    %rep %0
+        %define t%1q t%1 %+ q
+        %define t%1d t%1 %+ d
+        %define t%1w t%1 %+ w
+        %define t%1h t%1 %+ h
+        %define t%1b t%1 %+ b
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
+
+%if ARCH_X86_64
+    %define gprsize 8
+%else
+    %define gprsize 4
+%endif
+
+%macro PUSH 1
+    push %1
+    %assign stack_offset stack_offset+gprsize
+%endmacro
+
+%macro POP 1
+    pop %1
+    %assign stack_offset stack_offset-gprsize
+%endmacro
+
+%macro PUSH_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            PUSH r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro POP_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            pop r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro LOAD_IF_USED 1-*
+    %rep %0
+        %if %1 < num_args
+            mov r%1, r %+ %1 %+ mp
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro SUB 2
+    sub %1, %2
+    %ifidn %1, rsp
+        %assign stack_offset stack_offset+(%2)
+    %endif
+%endmacro
+
+%macro ADD 2
+    add %1, %2
+    %ifidn %1, rsp
+        %assign stack_offset stack_offset-(%2)
+    %endif
+%endmacro
+
+%macro movifnidn 2
+    %ifnidn %1, %2
+        mov %1, %2
+    %endif
+%endmacro
+
+%macro movsxdifnidn 2
+    %ifnidn %1, %2
+        movsxd %1, %2
+    %endif
+%endmacro
+
+%macro ASSERT 1
+    %if (%1) == 0
+        %error assert failed
+    %endif
+%endmacro
+
+%macro DEFINE_ARGS 0-*
+    %ifdef n_arg_names
+        %assign %%i 0
+        %rep n_arg_names
+            CAT_UNDEF arg_name %+ %%i, q
+            CAT_UNDEF arg_name %+ %%i, d
+            CAT_UNDEF arg_name %+ %%i, w
+            CAT_UNDEF arg_name %+ %%i, h
+            CAT_UNDEF arg_name %+ %%i, b
+            CAT_UNDEF arg_name %+ %%i, m
+            CAT_UNDEF arg_name %+ %%i, mp
+            CAT_UNDEF arg_name, %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+
+    %xdefine %%stack_offset stack_offset
+    %undef stack_offset ; so that the current value of stack_offset doesn't get baked in by xdefine
+    %assign %%i 0
+    %rep %0
+        %xdefine %1q r %+ %%i %+ q
+        %xdefine %1d r %+ %%i %+ d
+        %xdefine %1w r %+ %%i %+ w
+        %xdefine %1h r %+ %%i %+ h
+        %xdefine %1b r %+ %%i %+ b
+        %xdefine %1m r %+ %%i %+ m
+        %xdefine %1mp r %+ %%i %+ mp
+        CAT_XDEFINE arg_name, %%i, %1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+    %xdefine stack_offset %%stack_offset
+    %assign n_arg_names %0
+%endmacro
+
+%if WIN64 ; Windows x64 ;=================================================
+
+DECLARE_REG 0,  rcx
+DECLARE_REG 1,  rdx
+DECLARE_REG 2,  R8
+DECLARE_REG 3,  R9
+DECLARE_REG 4,  R10, 40
+DECLARE_REG 5,  R11, 48
+DECLARE_REG 6,  rax, 56
+DECLARE_REG 7,  rdi, 64
+DECLARE_REG 8,  rsi, 72
+DECLARE_REG 9,  rbx, 80
+DECLARE_REG 10, rbp, 88
+DECLARE_REG 11, R12, 96
+DECLARE_REG 12, R13, 104
+DECLARE_REG 13, R14, 112
+DECLARE_REG 14, R15, 120
+
+%macro PROLOGUE 2-4+ 0 ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 7, 8, 9, 10, 11, 12, 13, 14
+    %if mmsize == 8
+        %assign xmm_regs_used 0
+    %else
+        WIN64_SPILL_XMM %3
+    %endif
+    LOAD_IF_USED 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS %4
+%endmacro
+
+%macro WIN64_SPILL_XMM 1
+    %assign xmm_regs_used %1
+    ASSERT xmm_regs_used <= 16
+    %if xmm_regs_used > 6
+        SUB rsp, (xmm_regs_used-6)*16+16
+        %assign %%i xmm_regs_used
+        %rep (xmm_regs_used-6)
+            %assign %%i %%i-1
+            movdqa [rsp + (%%i-6)*16+(~stack_offset&8)], xmm %+ %%i
+        %endrep
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM_INTERNAL 1
+    %if xmm_regs_used > 6
+        %assign %%i xmm_regs_used
+        %rep (xmm_regs_used-6)
+            %assign %%i %%i-1
+            movdqa xmm %+ %%i, [%1 + (%%i-6)*16+(~stack_offset&8)]
+        %endrep
+        add %1, (xmm_regs_used-6)*16+16
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM 1
+    WIN64_RESTORE_XMM_INTERNAL %1
+    %assign stack_offset stack_offset-(xmm_regs_used-6)*16+16
+    %assign xmm_regs_used 0
+%endmacro
+
+%define has_epilogue regs_used > 7 || xmm_regs_used > 6 || mmsize == 32
+
+%macro RET 0
+    WIN64_RESTORE_XMM_INTERNAL rsp
+    POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%elif ARCH_X86_64 ; *nix x64 ;=============================================
+
+DECLARE_REG 0,  rdi
+DECLARE_REG 1,  rsi
+DECLARE_REG 2,  rdx
+DECLARE_REG 3,  rcx
+DECLARE_REG 4,  R8
+DECLARE_REG 5,  R9
+DECLARE_REG 6,  rax, 8
+DECLARE_REG 7,  R10, 16
+DECLARE_REG 8,  R11, 24
+DECLARE_REG 9,  rbx, 32
+DECLARE_REG 10, rbp, 40
+DECLARE_REG 11, R12, 48
+DECLARE_REG 12, R13, 56
+DECLARE_REG 13, R14, 64
+DECLARE_REG 14, R15, 72
+
+%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 9, 10, 11, 12, 13, 14
+    LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS %4
+%endmacro
+
+%define has_epilogue regs_used > 9 || mmsize == 32
+
+%macro RET 0
+    POP_IF_USED 14, 13, 12, 11, 10, 9
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%else ; X86_32 ;==============================================================
+
+DECLARE_REG 0, eax, 4
+DECLARE_REG 1, ecx, 8
+DECLARE_REG 2, edx, 12
+DECLARE_REG 3, ebx, 16
+DECLARE_REG 4, esi, 20
+DECLARE_REG 5, edi, 24
+DECLARE_REG 6, ebp, 28
+%define rsp esp
+
+%macro DECLARE_ARG 1-*
+    %rep %0
+        %define r%1m [esp + stack_offset + 4*%1 + 4]
+        %define r%1mp dword r%1m
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
+
+%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    %if regs_used > 7
+        %assign regs_used 7
+    %endif
+    ASSERT regs_used >= num_args
+    PUSH_IF_USED 3, 4, 5, 6
+    LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6
+    DEFINE_ARGS %4
+%endmacro
+
+%define has_epilogue regs_used > 3 || mmsize == 32
+
+%macro RET 0
+    POP_IF_USED 6, 5, 4, 3
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%endif ;======================================================================
+
+%if WIN64 == 0
+%macro WIN64_SPILL_XMM 1
+%endmacro
+%macro WIN64_RESTORE_XMM 1
+%endmacro
+%endif
+
+%macro REP_RET 0
+    %if has_epilogue
+        RET
+    %else
+        rep ret
+    %endif
+%endmacro
+
+%macro TAIL_CALL 2 ; callee, is_nonadjacent
+    %if has_epilogue
+        call %1
+        RET
+    %elif %2
+        jmp %1
+    %endif
+%endmacro
+
+;=============================================================================
+; arch-independent part
+;=============================================================================
+
+%assign function_align 16
+
+; Begin a function.
+; Applies any symbol mangling needed for C linkage, and sets up a define such that
+; subsequent uses of the function name automatically refer to the mangled version.
+; Appends cpuflags to the function name if cpuflags has been specified.
+%macro cglobal 1-2+ ; name, [PROLOGUE args]
+%if %0 == 1
+    cglobal_internal %1 %+ SUFFIX
+%else
+    cglobal_internal %1 %+ SUFFIX, %2
+%endif
+%endmacro
+%macro cglobal_internal 1-2+
+    %ifndef cglobaled_%1
+        %xdefine %1 mangle(%1)
+        %xdefine %1.skip_prologue %1 %+ .skip_prologue
+        CAT_XDEFINE cglobaled_, %1, 1
+    %endif
+    %xdefine current_function %1
+    %ifidn __OUTPUT_FORMAT__,elf
+        global %1:function hidden
+    %else
+        global %1
+    %endif
+    align function_align
+    %1:
+    RESET_MM_PERMUTATION ; not really needed, but makes disassembly somewhat nicer
+    %assign stack_offset 0
+    %if %0 > 1
+        PROLOGUE %2
+    %endif
+%endmacro
+
+%macro cextern 1
+    %xdefine %1 mangle(%1)
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+; like cextern, but without the prefix
+%macro cextern_naked 1
+    %xdefine %1 mangle(%1)
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+%macro const 2+
+    %xdefine %1 mangle(%1)
+    global %1
+    %1: %2
+%endmacro
+
+; This is needed for ELF, otherwise the GNU linker assumes the stack is
+; executable by default.
+%ifidn __OUTPUT_FORMAT__,elf
+SECTION .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+%ifidn __OUTPUT_FORMAT__,elf32
+section .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+%ifidn __OUTPUT_FORMAT__,elf64
+section .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+
+; cpuflags
+
+%assign cpuflags_MMX      (1<<0)
+%assign cpuflags_MMX2     (1<<1) | cpuflags_MMX
+%assign cpuflags_3dnow    (1<<2) | cpuflags_MMX
+%assign cpuflags_3dnow2   (1<<3) | cpuflags_3dnow
+%assign cpuflags_SSE      (1<<4) | cpuflags_MMX2
+%assign cpuflags_SSE2     (1<<5) | cpuflags_SSE
+%assign cpuflags_SSE2slow (1<<6) | cpuflags_SSE2
+%assign cpuflags_SSE3     (1<<7) | cpuflags_SSE2
+%assign cpuflags_SSSE3    (1<<8) | cpuflags_SSE3
+%assign cpuflags_SSE4     (1<<9) | cpuflags_SSSE3
+%assign cpuflags_SSE42    (1<<10)| cpuflags_SSE4
+%assign cpuflags_AVX      (1<<11)| cpuflags_SSE42
+%assign cpuflags_xop      (1<<12)| cpuflags_AVX
+%assign cpuflags_fma4     (1<<13)| cpuflags_AVX
+%assign cpuflags_AVX2     (1<<14)| cpuflags_AVX
+%assign cpuflags_fma3     (1<<15)| cpuflags_AVX
+
+%assign cpuflags_cache32  (1<<16)
+%assign cpuflags_cache64  (1<<17)
+%assign cpuflags_slowctz  (1<<18)
+%assign cpuflags_lzcnt    (1<<19)
+%assign cpuflags_misalign (1<<20)
+%assign cpuflags_aligned  (1<<21) ; not a cpu feature, but a function variant
+%assign cpuflags_atom     (1<<22)
+%assign cpuflags_bmi1     (1<<23)
+%assign cpuflags_bmi2     (1<<24)|cpuflags_bmi1
+%assign cpuflags_tbm      (1<<25)|cpuflags_bmi1
+
+%define    cpuflag(x) ((cpuflags & (cpuflags_ %+ x)) == (cpuflags_ %+ x))
+%define notcpuflag(x) ((cpuflags & (cpuflags_ %+ x)) != (cpuflags_ %+ x))
+
+; Takes up to 2 cpuflags from the above list.
+; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu.
+; You shouldn't need to invoke this macro directly, it's a subroutine for INIT_MMX &co.
+%macro INIT_CPUFLAGS 0-2
+    %if %0 >= 1
+        %xdefine cpuname %1
+        %assign cpuflags cpuflags_%1
+        %if %0 >= 2
+            %xdefine cpuname %1_%2
+            %assign cpuflags cpuflags | cpuflags_%2
+        %endif
+        %xdefine SUFFIX _ %+ cpuname
+        %if cpuflag(AVX)
+            %assign AVX_enabled 1
+        %endif
+        %if mmsize == 16 && notcpuflag(SSE2)
+            %define mova movaps
+            %define movu movups
+            %define movnta movntps
+        %endif
+        %if cpuflag(aligned)
+            %define movu mova
+        %elifidn %1, SSE3
+            %define movu lddqu
+        %endif
+    %else
+        %xdefine SUFFIX
+        %undef cpuname
+        %undef cpuflags
+    %endif
+%endmacro
+
+; merge MMX and SSE*
+
+%macro CAT_XDEFINE 3
+    %xdefine %1%2 %3
+%endmacro
+
+%macro CAT_UNDEF 2
+    %undef %1%2
+%endmacro
+
+%macro INIT_MMX 0-1+
+    %assign AVX_enabled 0
+    %define RESET_MM_PERMUTATION INIT_MMX %1
+    %define mmsize 8
+    %define num_mmregs 8
+    %define mova movq
+    %define movu movq
+    %define movh movd
+    %define movnta movntq
+    %assign %%i 0
+    %rep 8
+    CAT_XDEFINE m, %%i, mm %+ %%i
+    CAT_XDEFINE nmm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    %rep 8
+    CAT_UNDEF m, %%i
+    CAT_UNDEF nmm, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_XMM 0-1+
+    %assign AVX_enabled 0
+    %define RESET_MM_PERMUTATION INIT_XMM %1
+    %define mmsize 16
+    %define num_mmregs 8
+    %if ARCH_X86_64
+    %define num_mmregs 16
+    %endif
+    %define mova movdqa
+    %define movu movdqu
+    %define movh movq
+    %define movnta movntdq
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, xmm %+ %%i
+    CAT_XDEFINE nxmm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_YMM 0-1+
+    %assign AVX_enabled 1
+    %define RESET_MM_PERMUTATION INIT_YMM %1
+    %define mmsize 32
+    %define num_mmregs 8
+    %if ARCH_X86_64
+    %define num_mmregs 16
+    %endif
+    %define mova vmovaps
+    %define movu vmovups
+    %undef movh
+    %define movnta vmovntps
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, ymm %+ %%i
+    CAT_XDEFINE nymm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+INIT_XMM
+
+; I often want to use macros that permute their arguments. e.g. there's no
+; efficient way to implement butterfly or transpose or dct without swapping some
+; arguments.
+;
+; I would like to not have to manually keep track of the permutations:
+; If I insert a permutation in the middle of a function, it should automatically
+; change everything that follows. For more complex macros I may also have multiple
+; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
+;
+; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
+; permutes its arguments. It's equivalent to exchanging the contents of the
+; registers, except that this way you exchange the register names instead, so it
+; doesn't cost any cycles.
+
+%macro PERMUTE 2-* ; takes a list of pairs to swap
+%rep %0/2
+    %xdefine tmp%2 m%2
+    %xdefine ntmp%2 nm%2
+    %rotate 2
+%endrep
+%rep %0/2
+    %xdefine m%1 tmp%2
+    %xdefine nm%1 ntmp%2
+    %undef tmp%2
+    %undef ntmp%2
+    %rotate 2
+%endrep
+%endmacro
+
+%macro SWAP 2-* ; swaps a single chain (sometimes more concise than pairs)
+%rep %0-1
+%ifdef m%1
+    %xdefine tmp m%1
+    %xdefine m%1 m%2
+    %xdefine m%2 tmp
+    CAT_XDEFINE n, m%1, %1
+    CAT_XDEFINE n, m%2, %2
+%else
+    ; If we were called as "SWAP m0,m1" rather than "SWAP 0,1" infer the original numbers here.
+    ; Be careful using this mode in nested macros though, as in some cases there may be
+    ; other copies of m# that have already been dereferenced and don't get updated correctly.
+    %xdefine %%n1 n %+ %1
+    %xdefine %%n2 n %+ %2
+    %xdefine tmp m %+ %%n1
+    CAT_XDEFINE m, %%n1, m %+ %%n2
+    CAT_XDEFINE m, %%n2, tmp
+    CAT_XDEFINE n, m %+ %%n1, %%n1
+    CAT_XDEFINE n, m %+ %%n2, %%n2
+%endif
+    %undef tmp
+    %rotate 1
+%endrep
+%endmacro
+
+; If SAVE_MM_PERMUTATION is placed at the end of a function, then any later
+; calls to that function will automatically load the permutation, so values can
+; be returned in mmregs.
+%macro SAVE_MM_PERMUTATION 0-1
+    %if %0
+        %xdefine %%f %1_m
+    %else
+        %xdefine %%f current_function %+ _m
+    %endif
+    %assign %%i 0
+    %rep num_mmregs
+        CAT_XDEFINE %%f, %%i, m %+ %%i
+    %assign %%i %%i+1
+    %endrep
+%endmacro
+
+%macro LOAD_MM_PERMUTATION 1 ; name to load from
+    %ifdef %1_m0
+        %assign %%i 0
+        %rep num_mmregs
+            CAT_XDEFINE m, %%i, %1_m %+ %%i
+            CAT_XDEFINE n, m %+ %%i, %%i
+        %assign %%i %%i+1
+        %endrep
+    %endif
+%endmacro
+
+; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't
+%macro call 1
+    call_internal %1, %1 %+ SUFFIX
+%endmacro
+%macro call_internal 2
+    %xdefine %%i %1
+    %ifndef cglobaled_%1
+        %ifdef cglobaled_%2
+            %xdefine %%i %2
+        %endif
+    %endif
+    call %%i
+    LOAD_MM_PERMUTATION %%i
+%endmacro
+
+; Substitutions that reduce instruction size but are functionally equivalent
+%macro add 2
+    %ifnum %2
+        %if %2==128
+            sub %1, -128
+        %else
+            add %1, %2
+        %endif
+    %else
+        add %1, %2
+    %endif
+%endmacro
+
+%macro sub 2
+    %ifnum %2
+        %if %2==128
+            add %1, -128
+        %else
+            sub %1, %2
+        %endif
+    %else
+        sub %1, %2
+    %endif
+%endmacro
+
+;=============================================================================
+; AVX abstraction layer
+;=============================================================================
+
+%assign i 0
+%rep 16
+    %if i < 8
+        CAT_XDEFINE sizeofmm, i, 8
+    %endif
+    CAT_XDEFINE sizeofxmm, i, 16
+    CAT_XDEFINE sizeofymm, i, 32
+%assign i i+1
+%endrep
+%undef i
+
+%macro CHECK_AVX_INSTR_EMU 3-*
+    %xdefine %%opcode %1
+    %xdefine %%dst %2
+    %rep %0-2
+        %ifidn %%dst, %3
+            %error non-AVX emulation of ``%%opcode'' is not supported
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+;%1 == instruction
+;%2 == 1 if float, 0 if int
+;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm)
+;%4 == number of operands given
+;%5+: operands
+%macro RUN_AVX_INSTR 6-7+
+    %ifid %6
+        %define %%sizeofreg sizeof%6
+    %elifid %5
+        %define %%sizeofreg sizeof%5
+    %else
+        %define %%sizeofreg mmsize
+    %endif
+    %if %%sizeofreg==32
+        %if %4>=3
+            v%1 %5, %6, %7
+        %else
+            v%1 %5, %6
+        %endif
+    %else
+        %if %%sizeofreg==8
+            %define %%regmov movq
+        %elif %2
+            %define %%regmov movaps
+        %else
+            %define %%regmov movdqa
+        %endif
+
+        %if %4>=3+%3
+            %ifnidn %5, %6
+                %if AVX_enabled && %%sizeofreg==16
+                    v%1 %5, %6, %7
+                %else
+                    CHECK_AVX_INSTR_EMU {%1 %5, %6, %7}, %5, %7
+                    %%regmov %5, %6
+                    %1 %5, %7
+                %endif
+            %else
+                %1 %5, %7
+            %endif
+        %elif %4>=3
+            %1 %5, %6, %7
+        %else
+            %1 %5, %6
+        %endif
+    %endif
+%endmacro
+
+; 3arg AVX ops with a memory arg can only have it in src2,
+; whereas SSE emulation of 3arg prefers to have it in src1 (i.e. the mov).
+; So, if the op is symmetric and the wrong one is memory, swap them.
+%macro RUN_AVX_INSTR1 8
+    %assign %%swap 0
+    %if AVX_enabled
+        %ifnid %6
+            %assign %%swap 1
+        %endif
+    %elifnidn %5, %6
+        %ifnid %7
+            %assign %%swap 1
+        %endif
+    %endif
+    %if %%swap && %3 == 0 && %8 == 1
+        RUN_AVX_INSTR %1, %2, %3, %4, %5, %7, %6
+    %else
+        RUN_AVX_INSTR %1, %2, %3, %4, %5, %6, %7
+    %endif
+%endmacro
+
+;%1 == instruction
+;%2 == 1 if float, 0 if int
+;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm)
+;%4 == 1 if symmetric (i.e. doesn't matter which src arg is which), 0 if not
+%macro AVX_INSTR 4
+    %macro %1 2-9 fnord, fnord, fnord, %1, %2, %3, %4
+        %ifidn %3, fnord
+            RUN_AVX_INSTR %6, %7, %8, 2, %1, %2
+        %elifidn %4, fnord
+            RUN_AVX_INSTR1 %6, %7, %8, 3, %1, %2, %3, %9
+        %elifidn %5, fnord
+            RUN_AVX_INSTR %6, %7, %8, 4, %1, %2, %3, %4
+        %else
+            RUN_AVX_INSTR %6, %7, %8, 5, %1, %2, %3, %4, %5
+        %endif
+    %endmacro
+%endmacro
+
+AVX_INSTR addpd, 1, 0, 1
+AVX_INSTR addps, 1, 0, 1
+AVX_INSTR addsd, 1, 0, 1
+AVX_INSTR addss, 1, 0, 1
+AVX_INSTR addsubpd, 1, 0, 0
+AVX_INSTR addsubps, 1, 0, 0
+AVX_INSTR andpd, 1, 0, 1
+AVX_INSTR andps, 1, 0, 1
+AVX_INSTR andnpd, 1, 0, 0
+AVX_INSTR andnps, 1, 0, 0
+AVX_INSTR blendpd, 1, 0, 0
+AVX_INSTR blendps, 1, 0, 0
+AVX_INSTR blendvpd, 1, 0, 0
+AVX_INSTR blendvps, 1, 0, 0
+AVX_INSTR cmppd, 1, 0, 0
+AVX_INSTR cmpps, 1, 0, 0
+AVX_INSTR cmpsd, 1, 0, 0
+AVX_INSTR cmpss, 1, 0, 0
+AVX_INSTR cvtdq2ps, 1, 0, 0
+AVX_INSTR cvtps2dq, 1, 0, 0
+AVX_INSTR divpd, 1, 0, 0
+AVX_INSTR divps, 1, 0, 0
+AVX_INSTR divsd, 1, 0, 0
+AVX_INSTR divss, 1, 0, 0
+AVX_INSTR dppd, 1, 1, 0
+AVX_INSTR dpps, 1, 1, 0
+AVX_INSTR haddpd, 1, 0, 0
+AVX_INSTR haddps, 1, 0, 0
+AVX_INSTR hsubpd, 1, 0, 0
+AVX_INSTR hsubps, 1, 0, 0
+AVX_INSTR maxpd, 1, 0, 1
+AVX_INSTR maxps, 1, 0, 1
+AVX_INSTR maxsd, 1, 0, 1
+AVX_INSTR maxss, 1, 0, 1
+AVX_INSTR minpd, 1, 0, 1
+AVX_INSTR minps, 1, 0, 1
+AVX_INSTR minsd, 1, 0, 1
+AVX_INSTR minss, 1, 0, 1
+AVX_INSTR movhlps, 1, 0, 0
+AVX_INSTR movlhps, 1, 0, 0
+AVX_INSTR movsd, 1, 0, 0
+AVX_INSTR movss, 1, 0, 0
+AVX_INSTR mpsadbw, 0, 1, 0
+AVX_INSTR mulpd, 1, 0, 1
+AVX_INSTR mulps, 1, 0, 1
+AVX_INSTR mulsd, 1, 0, 1
+AVX_INSTR mulss, 1, 0, 1
+AVX_INSTR orpd, 1, 0, 1
+AVX_INSTR orps, 1, 0, 1
+AVX_INSTR pabsb, 0, 0, 0
+AVX_INSTR pabsw, 0, 0, 0
+AVX_INSTR pabsd, 0, 0, 0
+AVX_INSTR packsswb, 0, 0, 0
+AVX_INSTR packssdw, 0, 0, 0
+AVX_INSTR packuswb, 0, 0, 0
+AVX_INSTR packusdw, 0, 0, 0
+AVX_INSTR paddb, 0, 0, 1
+AVX_INSTR paddw, 0, 0, 1
+AVX_INSTR paddd, 0, 0, 1
+AVX_INSTR paddq, 0, 0, 1
+AVX_INSTR paddsb, 0, 0, 1
+AVX_INSTR paddsw, 0, 0, 1
+AVX_INSTR paddusb, 0, 0, 1
+AVX_INSTR paddusw, 0, 0, 1
+AVX_INSTR palignr, 0, 1, 0
+AVX_INSTR pand, 0, 0, 1
+AVX_INSTR pandn, 0, 0, 0
+AVX_INSTR pavgb, 0, 0, 1
+AVX_INSTR pavgw, 0, 0, 1
+AVX_INSTR pblendvb, 0, 0, 0
+AVX_INSTR pblendw, 0, 1, 0
+AVX_INSTR pcmpestri, 0, 0, 0
+AVX_INSTR pcmpestrm, 0, 0, 0
+AVX_INSTR pcmpistri, 0, 0, 0
+AVX_INSTR pcmpistrm, 0, 0, 0
+AVX_INSTR pcmpeqb, 0, 0, 1
+AVX_INSTR pcmpeqw, 0, 0, 1
+AVX_INSTR pcmpeqd, 0, 0, 1
+AVX_INSTR pcmpeqq, 0, 0, 1
+AVX_INSTR pcmpgtb, 0, 0, 0
+AVX_INSTR pcmpgtw, 0, 0, 0
+AVX_INSTR pcmpgtd, 0, 0, 0
+AVX_INSTR pcmpgtq, 0, 0, 0
+AVX_INSTR phaddw, 0, 0, 0
+AVX_INSTR phaddd, 0, 0, 0
+AVX_INSTR phaddsw, 0, 0, 0
+AVX_INSTR phsubw, 0, 0, 0
+AVX_INSTR phsubd, 0, 0, 0
+AVX_INSTR phsubsw, 0, 0, 0
+AVX_INSTR pmaddwd, 0, 0, 1
+AVX_INSTR pmaddubsw, 0, 0, 0
+AVX_INSTR pmaxsb, 0, 0, 1
+AVX_INSTR pmaxsw, 0, 0, 1
+AVX_INSTR pmaxsd, 0, 0, 1
+AVX_INSTR pmaxub, 0, 0, 1
+AVX_INSTR pmaxuw, 0, 0, 1
+AVX_INSTR pmaxud, 0, 0, 1
+AVX_INSTR pminsb, 0, 0, 1
+AVX_INSTR pminsw, 0, 0, 1
+AVX_INSTR pminsd, 0, 0, 1
+AVX_INSTR pminub, 0, 0, 1
+AVX_INSTR pminuw, 0, 0, 1
+AVX_INSTR pminud, 0, 0, 1
+AVX_INSTR pmovmskb, 0, 0, 0
+AVX_INSTR pmulhuw, 0, 0, 1
+AVX_INSTR pmulhrsw, 0, 0, 1
+AVX_INSTR pmulhw, 0, 0, 1
+AVX_INSTR pmullw, 0, 0, 1
+AVX_INSTR pmulld, 0, 0, 1
+AVX_INSTR pmuludq, 0, 0, 1
+AVX_INSTR pmuldq, 0, 0, 1
+AVX_INSTR por, 0, 0, 1
+AVX_INSTR psadbw, 0, 0, 1
+AVX_INSTR pshufb, 0, 0, 0
+AVX_INSTR pshufd, 0, 1, 0
+AVX_INSTR pshufhw, 0, 1, 0
+AVX_INSTR pshuflw, 0, 1, 0
+AVX_INSTR psignb, 0, 0, 0
+AVX_INSTR psignw, 0, 0, 0
+AVX_INSTR psignd, 0, 0, 0
+AVX_INSTR psllw, 0, 0, 0
+AVX_INSTR pslld, 0, 0, 0
+AVX_INSTR psllq, 0, 0, 0
+AVX_INSTR pslldq, 0, 0, 0
+AVX_INSTR psraw, 0, 0, 0
+AVX_INSTR psrad, 0, 0, 0
+AVX_INSTR psrlw, 0, 0, 0
+AVX_INSTR psrld, 0, 0, 0
+AVX_INSTR psrlq, 0, 0, 0
+AVX_INSTR psrldq, 0, 0, 0
+AVX_INSTR psubb, 0, 0, 0
+AVX_INSTR psubw, 0, 0, 0
+AVX_INSTR psubd, 0, 0, 0
+AVX_INSTR psubq, 0, 0, 0
+AVX_INSTR psubsb, 0, 0, 0
+AVX_INSTR psubsw, 0, 0, 0
+AVX_INSTR psubusb, 0, 0, 0
+AVX_INSTR psubusw, 0, 0, 0
+AVX_INSTR ptest, 0, 0, 0
+AVX_INSTR punpckhbw, 0, 0, 0
+AVX_INSTR punpckhwd, 0, 0, 0
+AVX_INSTR punpckhdq, 0, 0, 0
+AVX_INSTR punpckhqdq, 0, 0, 0
+AVX_INSTR punpcklbw, 0, 0, 0
+AVX_INSTR punpcklwd, 0, 0, 0
+AVX_INSTR punpckldq, 0, 0, 0
+AVX_INSTR punpcklqdq, 0, 0, 0
+AVX_INSTR pxor, 0, 0, 1
+AVX_INSTR shufps, 1, 1, 0
+AVX_INSTR subpd, 1, 0, 0
+AVX_INSTR subps, 1, 0, 0
+AVX_INSTR subsd, 1, 0, 0
+AVX_INSTR subss, 1, 0, 0
+AVX_INSTR unpckhpd, 1, 0, 0
+AVX_INSTR unpckhps, 1, 0, 0
+AVX_INSTR unpcklpd, 1, 0, 0
+AVX_INSTR unpcklps, 1, 0, 0
+AVX_INSTR xorpd, 1, 0, 1
+AVX_INSTR xorps, 1, 0, 1
+
+; 3DNow instructions, for sharing code between AVX, SSE and 3DN
+AVX_INSTR pfadd, 1, 0, 1
+AVX_INSTR pfsub, 1, 0, 0
+AVX_INSTR pfmul, 1, 0, 1
+
+; base-4 constants for shuffles
+%assign i 0
+%rep 256
+    %assign j ((i>>6)&3)*1000 + ((i>>4)&3)*100 + ((i>>2)&3)*10 + (i&3)
+    %if j < 10
+        CAT_XDEFINE q000, j, i
+    %elif j < 100
+        CAT_XDEFINE q00, j, i
+    %elif j < 1000
+        CAT_XDEFINE q0, j, i
+    %else
+        CAT_XDEFINE q, j, i
+    %endif
+%assign i i+1
+%endrep
+%undef i
+%undef j
+
+%macro FMA_INSTR 3
+    %macro %1 4-7 %1, %2, %3
+        %if cpuflag(xop)
+            v%5 %1, %2, %3, %4
+        %else
+            %6 %1, %2, %3
+            %7 %1, %4
+        %endif
+    %endmacro
+%endmacro
+
+FMA_INSTR  pmacsdd,  pmulld, paddd
+FMA_INSTR  pmacsww,  pmullw, paddw
+FMA_INSTR pmadcswd, pmaddwd, paddd
+
+; tzcnt is equivalent to "rep bsf" and is backwards-compatible with bsf.
+; This lets us use tzcnt without bumping the yasm version requirement yet.
+%define tzcnt rep bsf
diff --git a/libvpx/libvpx/third_party/x86inc/LICENSE b/libvpx/libvpx/third_party/x86inc/LICENSE
new file mode 100644
index 0000000..7d07645
--- /dev/null
+++ b/libvpx/libvpx/third_party/x86inc/LICENSE
@@ -0,0 +1,18 @@
+Copyright (C) 2005-2012 x264 project
+
+Authors: Loren Merritt <lorenm@u.washington.edu>
+         Anton Mitrofanov <BugMaster@narod.ru>
+         Jason Garrett-Glaser <darkshikari@gmail.com>
+         Henrik Gramner <hengar-6@student.ltu.se>
+
+Permission to use, copy, modify, and/or distribute this software for any
+purpose with or without fee is hereby granted, provided that the above
+copyright notice and this permission notice appear in all copies.
+
+THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
diff --git a/libvpx/libvpx/third_party/x86inc/README.libvpx b/libvpx/libvpx/third_party/x86inc/README.libvpx
new file mode 100644
index 0000000..8d3cd96
--- /dev/null
+++ b/libvpx/libvpx/third_party/x86inc/README.libvpx
@@ -0,0 +1,20 @@
+URL: https://git.videolan.org/git/x264.git
+Version: d23d18655249944c1ca894b451e2c82c7a584c62
+License: ISC
+License File: LICENSE
+
+Description:
+x264/libav's framework for x86 assembly. Contains a variety of macros and
+defines that help automatically allow assembly to work cross-platform.
+
+Local Modifications:
+Get configuration from vpx_config.asm.
+Prefix functions with vpx by default.
+Manage name mangling (prefixing with '_') manually because 'PREFIX' does not
+  exist in libvpx.
+Expand PIC default to macho64 and respect CONFIG_PIC from libvpx
+Set 'private_extern' visibility for macho targets.
+Copy PIC 'GLOBAL' macros from x86_abi_support.asm
+Use .text instead of .rodata on macho to avoid broken tables in PIC mode.
+Use .text with no alignment for aout
+Only use 'hidden' visibility with Chromium
diff --git a/libvpx/libvpx/third_party/x86inc/x86inc.asm b/libvpx/libvpx/third_party/x86inc/x86inc.asm
new file mode 100644
index 0000000..b647dff
--- /dev/null
+++ b/libvpx/libvpx/third_party/x86inc/x86inc.asm
@@ -0,0 +1,1649 @@
+;*****************************************************************************
+;* x86inc.asm: x264asm abstraction layer
+;*****************************************************************************
+;* Copyright (C) 2005-2016 x264 project
+;*
+;* Authors: Loren Merritt <lorenm@u.washington.edu>
+;*          Anton Mitrofanov <BugMaster@narod.ru>
+;*          Fiona Glaser <fiona@x264.com>
+;*          Henrik Gramner <henrik@gramner.com>
+;*
+;* Permission to use, copy, modify, and/or distribute this software for any
+;* purpose with or without fee is hereby granted, provided that the above
+;* copyright notice and this permission notice appear in all copies.
+;*
+;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+;*****************************************************************************
+
+; This is a header file for the x264ASM assembly language, which uses
+; NASM/YASM syntax combined with a large number of macros to provide easy
+; abstraction between different calling conventions (x86_32, win64, linux64).
+; It also has various other useful features to simplify writing the kind of
+; DSP functions that are most often used in x264.
+
+; Unlike the rest of x264, this file is available under an ISC license, as it
+; has significant usefulness outside of x264 and we want it to be available
+; to the largest audience possible.  Of course, if you modify it for your own
+; purposes to add a new feature, we strongly encourage contributing a patch
+; as this feature might be useful for others as well.  Send patches or ideas
+; to x264-devel@videolan.org .
+
+%include "vpx_config.asm"
+
+%ifndef private_prefix
+    %define private_prefix vpx
+%endif
+
+%ifndef public_prefix
+    %define public_prefix private_prefix
+%endif
+
+%ifndef STACK_ALIGNMENT
+    %if ARCH_X86_64
+        %define STACK_ALIGNMENT 16
+    %else
+        %define STACK_ALIGNMENT 4
+    %endif
+%endif
+
+%define WIN64  0
+%define UNIX64 0
+%if ARCH_X86_64
+    %ifidn __OUTPUT_FORMAT__,win32
+        %define WIN64  1
+    %elifidn __OUTPUT_FORMAT__,win64
+        %define WIN64  1
+    %elifidn __OUTPUT_FORMAT__,x64
+        %define WIN64  1
+    %else
+        %define UNIX64 1
+    %endif
+%endif
+
+%define FORMAT_ELF 0
+%ifidn __OUTPUT_FORMAT__,elf
+    %define FORMAT_ELF 1
+%elifidn __OUTPUT_FORMAT__,elf32
+    %define FORMAT_ELF 1
+%elifidn __OUTPUT_FORMAT__,elf64
+    %define FORMAT_ELF 1
+%endif
+
+%define FORMAT_MACHO 0
+%ifidn __OUTPUT_FORMAT__,macho32
+     %define FORMAT_MACHO 1
+%elifidn __OUTPUT_FORMAT__,macho64
+     %define FORMAT_MACHO 1
+%endif
+
+; Set PREFIX for libvpx builds.
+%if FORMAT_ELF
+    %undef PREFIX
+%elif WIN64
+    %undef PREFIX
+%else
+    %define PREFIX
+%endif
+
+%ifdef PREFIX
+    %define mangle(x) _ %+ x
+%else
+    %define mangle(x) x
+%endif
+
+; In some instances macho32 tables get misaligned when using .rodata.
+; When looking at the disassembly it appears that the offset is either
+; correct or consistently off by 90. Placing them in the .text section
+; works around the issue. It appears to be specific to the way libvpx
+; handles the tables.
+%macro SECTION_RODATA 0-1 16
+    %ifidn __OUTPUT_FORMAT__,macho32
+        SECTION .text align=%1
+        fakegot:
+    %elifidn __OUTPUT_FORMAT__,aout
+        SECTION .text
+    %else
+        SECTION .rodata align=%1
+    %endif
+%endmacro
+
+; PIC macros are copied from vpx_ports/x86_abi_support.asm. The "define PIC"
+; from original code is added in for 64bit.
+%ifidn __OUTPUT_FORMAT__,elf32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,macho32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,win32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,aout
+%define ABI_IS_32BIT 1
+%else
+%define ABI_IS_32BIT 0
+%endif
+
+%if ABI_IS_32BIT
+    %if CONFIG_PIC=1
+        %ifidn __OUTPUT_FORMAT__,elf32
+            %define GET_GOT_DEFINED 1
+            %define WRT_PLT wrt ..plt
+            %macro GET_GOT 1
+                extern _GLOBAL_OFFSET_TABLE_
+                push %1
+                call %%get_got
+                %%sub_offset:
+                jmp %%exitGG
+                %%get_got:
+                mov %1, [esp]
+                add %1, _GLOBAL_OFFSET_TABLE_ + $$ - %%sub_offset wrt ..gotpc
+                ret
+                %%exitGG:
+                %undef GLOBAL
+                %define GLOBAL(x) x + %1 wrt ..gotoff
+                %undef RESTORE_GOT
+                %define RESTORE_GOT pop %1
+            %endmacro
+        %elifidn __OUTPUT_FORMAT__,macho32
+            %define GET_GOT_DEFINED 1
+            %macro GET_GOT 1
+                push %1
+                call %%get_got
+                %%get_got:
+                pop  %1
+                %undef GLOBAL
+                %define GLOBAL(x) x + %1 - %%get_got
+                %undef RESTORE_GOT
+                %define RESTORE_GOT pop %1
+            %endmacro
+        %else
+            %define GET_GOT_DEFINED 0
+        %endif
+    %endif
+
+    %if ARCH_X86_64 == 0
+        %undef PIC
+    %endif
+
+%else
+    %macro GET_GOT 1
+    %endmacro
+    %define GLOBAL(x) rel x
+    %define WRT_PLT wrt ..plt
+
+    %if WIN64
+        %define PIC
+    %elifidn __OUTPUT_FORMAT__,macho64
+        %define PIC
+    %elif CONFIG_PIC
+        %define PIC
+    %endif
+%endif
+
+%ifnmacro GET_GOT
+    %macro GET_GOT 1
+    %endmacro
+    %define GLOBAL(x) x
+%endif
+%ifndef RESTORE_GOT
+    %define RESTORE_GOT
+%endif
+%ifndef WRT_PLT
+    %define WRT_PLT
+%endif
+
+%ifdef PIC
+    default rel
+%endif
+
+%ifndef GET_GOT_DEFINED
+    %define GET_GOT_DEFINED 0
+%endif
+; Done with PIC macros
+
+%ifdef __NASM_VER__
+    %use smartalign
+%endif
+
+; Macros to eliminate most code duplication between x86_32 and x86_64:
+; Currently this works only for leaf functions which load all their arguments
+; into registers at the start, and make no other use of the stack. Luckily that
+; covers most of x264's asm.
+
+; PROLOGUE:
+; %1 = number of arguments. loads them from stack if needed.
+; %2 = number of registers used. pushes callee-saved regs if needed.
+; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
+; %4 = (optional) stack size to be allocated. The stack will be aligned before
+;      allocating the specified stack size. If the required stack alignment is
+;      larger than the known stack alignment the stack will be manually aligned
+;      and an extra register will be allocated to hold the original stack
+;      pointer (to not invalidate r0m etc.). To prevent the use of an extra
+;      register as stack pointer, request a negative stack size.
+; %4+/%5+ = list of names to define to registers
+; PROLOGUE can also be invoked by adding the same options to cglobal
+
+; e.g.
+; cglobal foo, 2,3,7,0x40, dst, src, tmp
+; declares a function (foo) that automatically loads two arguments (dst and
+; src) into registers, uses one additional register (tmp) plus 7 vector
+; registers (m0-m6) and allocates 0x40 bytes of stack space.
+
+; TODO Some functions can use some args directly from the stack. If they're the
+; last args then you can just not declare them, but if they're in the middle
+; we need more flexible macro.
+
+; RET:
+; Pops anything that was pushed by PROLOGUE, and returns.
+
+; REP_RET:
+; Use this instead of RET if it's a branch target.
+
+; registers:
+; rN and rNq are the native-size register holding function argument N
+; rNd, rNw, rNb are dword, word, and byte size
+; rNh is the high 8 bits of the word size
+; rNm is the original location of arg N (a register or on the stack), dword
+; rNmp is native size
+
+%macro DECLARE_REG 2-3
+    %define r%1q %2
+    %define r%1d %2d
+    %define r%1w %2w
+    %define r%1b %2b
+    %define r%1h %2h
+    %define %2q %2
+    %if %0 == 2
+        %define r%1m  %2d
+        %define r%1mp %2
+    %elif ARCH_X86_64 ; memory
+        %define r%1m [rstk + stack_offset + %3]
+        %define r%1mp qword r %+ %1 %+ m
+    %else
+        %define r%1m [rstk + stack_offset + %3]
+        %define r%1mp dword r %+ %1 %+ m
+    %endif
+    %define r%1  %2
+%endmacro
+
+%macro DECLARE_REG_SIZE 3
+    %define r%1q r%1
+    %define e%1q r%1
+    %define r%1d e%1
+    %define e%1d e%1
+    %define r%1w %1
+    %define e%1w %1
+    %define r%1h %3
+    %define e%1h %3
+    %define r%1b %2
+    %define e%1b %2
+    %if ARCH_X86_64 == 0
+        %define r%1 e%1
+    %endif
+%endmacro
+
+DECLARE_REG_SIZE ax, al, ah
+DECLARE_REG_SIZE bx, bl, bh
+DECLARE_REG_SIZE cx, cl, ch
+DECLARE_REG_SIZE dx, dl, dh
+DECLARE_REG_SIZE si, sil, null
+DECLARE_REG_SIZE di, dil, null
+DECLARE_REG_SIZE bp, bpl, null
+
+; t# defines for when per-arch register allocation is more complex than just function arguments
+
+%macro DECLARE_REG_TMP 1-*
+    %assign %%i 0
+    %rep %0
+        CAT_XDEFINE t, %%i, r%1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro DECLARE_REG_TMP_SIZE 0-*
+    %rep %0
+        %define t%1q t%1 %+ q
+        %define t%1d t%1 %+ d
+        %define t%1w t%1 %+ w
+        %define t%1h t%1 %+ h
+        %define t%1b t%1 %+ b
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
+
+%if ARCH_X86_64
+    %define gprsize 8
+%else
+    %define gprsize 4
+%endif
+
+%macro PUSH 1
+    push %1
+    %ifidn rstk, rsp
+        %assign stack_offset stack_offset+gprsize
+    %endif
+%endmacro
+
+%macro POP 1
+    pop %1
+    %ifidn rstk, rsp
+        %assign stack_offset stack_offset-gprsize
+    %endif
+%endmacro
+
+%macro PUSH_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            PUSH r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro POP_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            pop r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro LOAD_IF_USED 1-*
+    %rep %0
+        %if %1 < num_args
+            mov r%1, r %+ %1 %+ mp
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro SUB 2
+    sub %1, %2
+    %ifidn %1, rstk
+        %assign stack_offset stack_offset+(%2)
+    %endif
+%endmacro
+
+%macro ADD 2
+    add %1, %2
+    %ifidn %1, rstk
+        %assign stack_offset stack_offset-(%2)
+    %endif
+%endmacro
+
+%macro movifnidn 2
+    %ifnidn %1, %2
+        mov %1, %2
+    %endif
+%endmacro
+
+%macro movsxdifnidn 2
+    %ifnidn %1, %2
+        movsxd %1, %2
+    %endif
+%endmacro
+
+%macro ASSERT 1
+    %if (%1) == 0
+        %error assertion ``%1'' failed
+    %endif
+%endmacro
+
+%macro DEFINE_ARGS 0-*
+    %ifdef n_arg_names
+        %assign %%i 0
+        %rep n_arg_names
+            CAT_UNDEF arg_name %+ %%i, q
+            CAT_UNDEF arg_name %+ %%i, d
+            CAT_UNDEF arg_name %+ %%i, w
+            CAT_UNDEF arg_name %+ %%i, h
+            CAT_UNDEF arg_name %+ %%i, b
+            CAT_UNDEF arg_name %+ %%i, m
+            CAT_UNDEF arg_name %+ %%i, mp
+            CAT_UNDEF arg_name, %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+
+    %xdefine %%stack_offset stack_offset
+    %undef stack_offset ; so that the current value of stack_offset doesn't get baked in by xdefine
+    %assign %%i 0
+    %rep %0
+        %xdefine %1q r %+ %%i %+ q
+        %xdefine %1d r %+ %%i %+ d
+        %xdefine %1w r %+ %%i %+ w
+        %xdefine %1h r %+ %%i %+ h
+        %xdefine %1b r %+ %%i %+ b
+        %xdefine %1m r %+ %%i %+ m
+        %xdefine %1mp r %+ %%i %+ mp
+        CAT_XDEFINE arg_name, %%i, %1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+    %xdefine stack_offset %%stack_offset
+    %assign n_arg_names %0
+%endmacro
+
+%define required_stack_alignment ((mmsize + 15) & ~15)
+
+%macro ALLOC_STACK 1-2 0 ; stack_size, n_xmm_regs (for win64 only)
+    %ifnum %1
+        %if %1 != 0
+            %assign %%pad 0
+            %assign stack_size %1
+            %if stack_size < 0
+                %assign stack_size -stack_size
+            %endif
+            %if WIN64
+                %assign %%pad %%pad + 32 ; shadow space
+                %if mmsize != 8
+                    %assign xmm_regs_used %2
+                    %if xmm_regs_used > 8
+                        %assign %%pad %%pad + (xmm_regs_used-8)*16 ; callee-saved xmm registers
+                    %endif
+                %endif
+            %endif
+            %if required_stack_alignment <= STACK_ALIGNMENT
+                ; maintain the current stack alignment
+                %assign stack_size_padded stack_size + %%pad + ((-%%pad-stack_offset-gprsize) & (STACK_ALIGNMENT-1))
+                SUB rsp, stack_size_padded
+            %else
+                %assign %%reg_num (regs_used - 1)
+                %xdefine rstk r %+ %%reg_num
+                ; align stack, and save original stack location directly above
+                ; it, i.e. in [rsp+stack_size_padded], so we can restore the
+                ; stack in a single instruction (i.e. mov rsp, rstk or mov
+                ; rsp, [rsp+stack_size_padded])
+                %if %1 < 0 ; need to store rsp on stack
+                    %xdefine rstkm [rsp + stack_size + %%pad]
+                    %assign %%pad %%pad + gprsize
+                %else ; can keep rsp in rstk during whole function
+                    %xdefine rstkm rstk
+                %endif
+                %assign stack_size_padded stack_size + ((%%pad + required_stack_alignment-1) & ~(required_stack_alignment-1))
+                mov rstk, rsp
+                and rsp, ~(required_stack_alignment-1)
+                sub rsp, stack_size_padded
+                movifnidn rstkm, rstk
+            %endif
+            WIN64_PUSH_XMM
+        %endif
+    %endif
+%endmacro
+
+%macro SETUP_STACK_POINTER 1
+    %ifnum %1
+        %if %1 != 0 && required_stack_alignment > STACK_ALIGNMENT
+            %if %1 > 0
+                %assign regs_used (regs_used + 1)
+            %endif
+            %if ARCH_X86_64 && regs_used < 5 + UNIX64 * 3
+                ; Ensure that we don't clobber any registers containing arguments
+                %assign regs_used 5 + UNIX64 * 3
+            %endif
+        %endif
+    %endif
+%endmacro
+
+%macro DEFINE_ARGS_INTERNAL 3+
+    %ifnum %2
+        DEFINE_ARGS %3
+    %elif %1 == 4
+        DEFINE_ARGS %2
+    %elif %1 > 4
+        DEFINE_ARGS %2, %3
+    %endif
+%endmacro
+
+%if WIN64 ; Windows x64 ;=================================================
+
+DECLARE_REG 0,  rcx
+DECLARE_REG 1,  rdx
+DECLARE_REG 2,  R8
+DECLARE_REG 3,  R9
+DECLARE_REG 4,  R10, 40
+DECLARE_REG 5,  R11, 48
+DECLARE_REG 6,  rax, 56
+DECLARE_REG 7,  rdi, 64
+DECLARE_REG 8,  rsi, 72
+DECLARE_REG 9,  rbx, 80
+DECLARE_REG 10, rbp, 88
+DECLARE_REG 11, R12, 96
+DECLARE_REG 12, R13, 104
+DECLARE_REG 13, R14, 112
+DECLARE_REG 14, R15, 120
+
+%macro PROLOGUE 2-5+ 0 ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    SETUP_STACK_POINTER %4
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 7, 8, 9, 10, 11, 12, 13, 14
+    ALLOC_STACK %4, %3
+    %if mmsize != 8 && stack_size == 0
+        WIN64_SPILL_XMM %3
+    %endif
+    LOAD_IF_USED 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS_INTERNAL %0, %4, %5
+%endmacro
+
+%macro WIN64_PUSH_XMM 0
+    ; Use the shadow space to store XMM6 and XMM7, the rest needs stack space allocated.
+    %if xmm_regs_used > 6
+        movaps [rstk + stack_offset +  8], xmm6
+    %endif
+    %if xmm_regs_used > 7
+        movaps [rstk + stack_offset + 24], xmm7
+    %endif
+    %if xmm_regs_used > 8
+        %assign %%i 8
+        %rep xmm_regs_used-8
+            movaps [rsp + (%%i-8)*16 + stack_size + 32], xmm %+ %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+%endmacro
+
+%macro WIN64_SPILL_XMM 1
+    %assign xmm_regs_used %1
+    ASSERT xmm_regs_used <= 16
+    %if xmm_regs_used > 8
+        ; Allocate stack space for callee-saved xmm registers plus shadow space and align the stack.
+        %assign %%pad (xmm_regs_used-8)*16 + 32
+        %assign stack_size_padded %%pad + ((-%%pad-stack_offset-gprsize) & (STACK_ALIGNMENT-1))
+        SUB rsp, stack_size_padded
+    %endif
+    WIN64_PUSH_XMM
+%endmacro
+
+%macro WIN64_RESTORE_XMM_INTERNAL 1
+    %assign %%pad_size 0
+    %if xmm_regs_used > 8
+        %assign %%i xmm_regs_used
+        %rep xmm_regs_used-8
+            %assign %%i %%i-1
+            movaps xmm %+ %%i, [%1 + (%%i-8)*16 + stack_size + 32]
+        %endrep
+    %endif
+    %if stack_size_padded > 0
+        %if stack_size > 0 && required_stack_alignment > STACK_ALIGNMENT
+            mov rsp, rstkm
+        %else
+            add %1, stack_size_padded
+            %assign %%pad_size stack_size_padded
+        %endif
+    %endif
+    %if xmm_regs_used > 7
+        movaps xmm7, [%1 + stack_offset - %%pad_size + 24]
+    %endif
+    %if xmm_regs_used > 6
+        movaps xmm6, [%1 + stack_offset - %%pad_size +  8]
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM 1
+    WIN64_RESTORE_XMM_INTERNAL %1
+    %assign stack_offset (stack_offset-stack_size_padded)
+    %assign xmm_regs_used 0
+%endmacro
+
+%define has_epilogue regs_used > 7 || xmm_regs_used > 6 || mmsize == 32 || stack_size > 0
+
+%macro RET 0
+    WIN64_RESTORE_XMM_INTERNAL rsp
+    POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7
+    %if mmsize == 32
+        vzeroupper
+    %endif
+    AUTO_REP_RET
+%endmacro
+
+%elif ARCH_X86_64 ; *nix x64 ;=============================================
+
+DECLARE_REG 0,  rdi
+DECLARE_REG 1,  rsi
+DECLARE_REG 2,  rdx
+DECLARE_REG 3,  rcx
+DECLARE_REG 4,  R8
+DECLARE_REG 5,  R9
+DECLARE_REG 6,  rax, 8
+DECLARE_REG 7,  R10, 16
+DECLARE_REG 8,  R11, 24
+DECLARE_REG 9,  rbx, 32
+DECLARE_REG 10, rbp, 40
+DECLARE_REG 11, R12, 48
+DECLARE_REG 12, R13, 56
+DECLARE_REG 13, R14, 64
+DECLARE_REG 14, R15, 72
+
+%macro PROLOGUE 2-5+ ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    SETUP_STACK_POINTER %4
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 9, 10, 11, 12, 13, 14
+    ALLOC_STACK %4
+    LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS_INTERNAL %0, %4, %5
+%endmacro
+
+%define has_epilogue regs_used > 9 || mmsize == 32 || stack_size > 0
+
+%macro RET 0
+    %if stack_size_padded > 0
+        %if required_stack_alignment > STACK_ALIGNMENT
+            mov rsp, rstkm
+        %else
+            add rsp, stack_size_padded
+        %endif
+    %endif
+    POP_IF_USED 14, 13, 12, 11, 10, 9
+    %if mmsize == 32
+        vzeroupper
+    %endif
+    AUTO_REP_RET
+%endmacro
+
+%else ; X86_32 ;==============================================================
+
+DECLARE_REG 0, eax, 4
+DECLARE_REG 1, ecx, 8
+DECLARE_REG 2, edx, 12
+DECLARE_REG 3, ebx, 16
+DECLARE_REG 4, esi, 20
+DECLARE_REG 5, edi, 24
+DECLARE_REG 6, ebp, 28
+%define rsp esp
+
+%macro DECLARE_ARG 1-*
+    %rep %0
+        %define r%1m [rstk + stack_offset + 4*%1 + 4]
+        %define r%1mp dword r%1m
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
+
+%macro PROLOGUE 2-5+ ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    %if num_args > 7
+        %assign num_args 7
+    %endif
+    %if regs_used > 7
+        %assign regs_used 7
+    %endif
+    SETUP_STACK_POINTER %4
+    ASSERT regs_used <= 7
+    PUSH_IF_USED 3, 4, 5, 6
+    ALLOC_STACK %4
+    LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6
+    DEFINE_ARGS_INTERNAL %0, %4, %5
+%endmacro
+
+%define has_epilogue regs_used > 3 || mmsize == 32 || stack_size > 0
+
+%macro RET 0
+    %if stack_size_padded > 0
+        %if required_stack_alignment > STACK_ALIGNMENT
+            mov rsp, rstkm
+        %else
+            add rsp, stack_size_padded
+        %endif
+    %endif
+    POP_IF_USED 6, 5, 4, 3
+    %if mmsize == 32
+        vzeroupper
+    %endif
+    AUTO_REP_RET
+%endmacro
+
+%endif ;======================================================================
+
+%if WIN64 == 0
+    %macro WIN64_SPILL_XMM 1
+    %endmacro
+    %macro WIN64_RESTORE_XMM 1
+    %endmacro
+    %macro WIN64_PUSH_XMM 0
+    %endmacro
+%endif
+
+; On AMD cpus <=K10, an ordinary ret is slow if it immediately follows either
+; a branch or a branch target. So switch to a 2-byte form of ret in that case.
+; We can automatically detect "follows a branch", but not a branch target.
+; (SSSE3 is a sufficient condition to know that your cpu doesn't have this problem.)
+%macro REP_RET 0
+    %if has_epilogue
+        RET
+    %else
+        rep ret
+    %endif
+    annotate_function_size
+%endmacro
+
+%define last_branch_adr $$
+%macro AUTO_REP_RET 0
+    %if notcpuflag(ssse3)
+        times ((last_branch_adr-$)>>31)+1 rep ; times 1 iff $ == last_branch_adr.
+    %endif
+    ret
+    annotate_function_size
+%endmacro
+
+%macro BRANCH_INSTR 0-*
+    %rep %0
+        %macro %1 1-2 %1
+            %2 %1
+            %if notcpuflag(ssse3)
+                %%branch_instr equ $
+                %xdefine last_branch_adr %%branch_instr
+            %endif
+        %endmacro
+        %rotate 1
+    %endrep
+%endmacro
+
+BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae, jna, jnae, jb, jbe, jnb, jnbe, jc, jnc, js, jns, jo, jno, jp, jnp
+
+%macro TAIL_CALL 2 ; callee, is_nonadjacent
+    %if has_epilogue
+        call %1
+        RET
+    %elif %2
+        jmp %1
+    %endif
+    annotate_function_size
+%endmacro
+
+;=============================================================================
+; arch-independent part
+;=============================================================================
+
+%assign function_align 16
+
+; Begin a function.
+; Applies any symbol mangling needed for C linkage, and sets up a define such that
+; subsequent uses of the function name automatically refer to the mangled version.
+; Appends cpuflags to the function name if cpuflags has been specified.
+; The "" empty default parameter is a workaround for nasm, which fails if SUFFIX
+; is empty and we call cglobal_internal with just %1 %+ SUFFIX (without %2).
+%macro cglobal 1-2+ "" ; name, [PROLOGUE args]
+    cglobal_internal 1, %1 %+ SUFFIX, %2
+%endmacro
+%macro cvisible 1-2+ "" ; name, [PROLOGUE args]
+    cglobal_internal 0, %1 %+ SUFFIX, %2
+%endmacro
+%macro cglobal_internal 2-3+
+    annotate_function_size
+    %if %1
+        %xdefine %%FUNCTION_PREFIX private_prefix
+        ; libvpx explicitly sets visibility in shared object builds. Avoid
+        ; setting visibility to hidden as it may break builds that split
+        ; sources on e.g., directory boundaries.
+        %ifdef CHROMIUM
+            %xdefine %%VISIBILITY hidden
+        %else
+            %xdefine %%VISIBILITY
+        %endif
+    %else
+        %xdefine %%FUNCTION_PREFIX public_prefix
+        %xdefine %%VISIBILITY
+    %endif
+    %ifndef cglobaled_%2
+        %xdefine %2 mangle(%%FUNCTION_PREFIX %+ _ %+ %2)
+        %xdefine %2.skip_prologue %2 %+ .skip_prologue
+        CAT_XDEFINE cglobaled_, %2, 1
+    %endif
+    %xdefine current_function %2
+    %xdefine current_function_section __SECT__
+    %if FORMAT_ELF
+        global %2:function %%VISIBILITY
+    %elif FORMAT_MACHO
+        %ifdef __NASM_VER__
+            global %2
+        %else
+            global %2:private_extern
+        %endif
+    %else
+        global %2
+    %endif
+    align function_align
+    %2:
+    RESET_MM_PERMUTATION        ; needed for x86-64, also makes disassembly somewhat nicer
+    %xdefine rstk rsp           ; copy of the original stack pointer, used when greater alignment than the known stack alignment is required
+    %assign stack_offset 0      ; stack pointer offset relative to the return address
+    %assign stack_size 0        ; amount of stack space that can be freely used inside a function
+    %assign stack_size_padded 0 ; total amount of allocated stack space, including space for callee-saved xmm registers on WIN64 and alignment padding
+    %assign xmm_regs_used 0     ; number of XMM registers requested, used for dealing with callee-saved registers on WIN64
+    %ifnidn %3, ""
+        PROLOGUE %3
+    %endif
+%endmacro
+
+%macro cextern 1
+    %xdefine %1 mangle(private_prefix %+ _ %+ %1)
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+; like cextern, but without the prefix
+%macro cextern_naked 1
+    %ifdef PREFIX
+        %xdefine %1 mangle(%1)
+    %endif
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+%macro const 1-2+
+    %xdefine %1 mangle(private_prefix %+ _ %+ %1)
+    %if FORMAT_ELF
+        global %1:data hidden
+    %else
+        global %1
+    %endif
+    %1: %2
+%endmacro
+
+; This is needed for ELF, otherwise the GNU linker assumes the stack is executable by default.
+%if FORMAT_ELF
+    [SECTION .note.GNU-stack noalloc noexec nowrite progbits]
+%endif
+
+; Tell debuggers how large the function was.
+; This may be invoked multiple times per function; we rely on later instances overriding earlier ones.
+; This is invoked by RET and similar macros, and also cglobal does it for the previous function,
+; but if the last function in a source file doesn't use any of the standard macros for its epilogue,
+; then its size might be unspecified.
+%macro annotate_function_size 0
+    %ifdef __YASM_VER__
+        %ifdef current_function
+            %if FORMAT_ELF
+                current_function_section
+                %%ecf equ $
+                size current_function %%ecf - current_function
+                __SECT__
+            %endif
+        %endif
+    %endif
+%endmacro
+
+; cpuflags
+
+%assign cpuflags_mmx      (1<<0)
+%assign cpuflags_mmx2     (1<<1) | cpuflags_mmx
+%assign cpuflags_3dnow    (1<<2) | cpuflags_mmx
+%assign cpuflags_3dnowext (1<<3) | cpuflags_3dnow
+%assign cpuflags_sse      (1<<4) | cpuflags_mmx2
+%assign cpuflags_sse2     (1<<5) | cpuflags_sse
+%assign cpuflags_sse2slow (1<<6) | cpuflags_sse2
+%assign cpuflags_sse3     (1<<7) | cpuflags_sse2
+%assign cpuflags_ssse3    (1<<8) | cpuflags_sse3
+%assign cpuflags_sse4     (1<<9) | cpuflags_ssse3
+%assign cpuflags_sse42    (1<<10)| cpuflags_sse4
+%assign cpuflags_avx      (1<<11)| cpuflags_sse42
+%assign cpuflags_xop      (1<<12)| cpuflags_avx
+%assign cpuflags_fma4     (1<<13)| cpuflags_avx
+%assign cpuflags_fma3     (1<<14)| cpuflags_avx
+%assign cpuflags_avx2     (1<<15)| cpuflags_fma3
+
+%assign cpuflags_cache32  (1<<16)
+%assign cpuflags_cache64  (1<<17)
+%assign cpuflags_slowctz  (1<<18)
+%assign cpuflags_lzcnt    (1<<19)
+%assign cpuflags_aligned  (1<<20) ; not a cpu feature, but a function variant
+%assign cpuflags_atom     (1<<21)
+%assign cpuflags_bmi1     (1<<22)|cpuflags_lzcnt
+%assign cpuflags_bmi2     (1<<23)|cpuflags_bmi1
+
+; Returns a boolean value expressing whether or not the specified cpuflag is enabled.
+%define    cpuflag(x) (((((cpuflags & (cpuflags_ %+ x)) ^ (cpuflags_ %+ x)) - 1) >> 31) & 1)
+%define notcpuflag(x) (cpuflag(x) ^ 1)
+
+; Takes an arbitrary number of cpuflags from the above list.
+; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu.
+; You shouldn't need to invoke this macro directly, it's a subroutine for INIT_MMX &co.
+%macro INIT_CPUFLAGS 0-*
+    %xdefine SUFFIX
+    %undef cpuname
+    %assign cpuflags 0
+
+    %if %0 >= 1
+        %rep %0
+            %ifdef cpuname
+                %xdefine cpuname cpuname %+ _%1
+            %else
+                %xdefine cpuname %1
+            %endif
+            %assign cpuflags cpuflags | cpuflags_%1
+            %rotate 1
+        %endrep
+        %xdefine SUFFIX _ %+ cpuname
+
+        %if cpuflag(avx)
+            %assign avx_enabled 1
+        %endif
+        %if (mmsize == 16 && notcpuflag(sse2)) || (mmsize == 32 && notcpuflag(avx2))
+            %define mova movaps
+            %define movu movups
+            %define movnta movntps
+        %endif
+        %if cpuflag(aligned)
+            %define movu mova
+        %elif cpuflag(sse3) && notcpuflag(ssse3)
+            %define movu lddqu
+        %endif
+    %endif
+
+    %if ARCH_X86_64 || cpuflag(sse2)
+        %ifdef __NASM_VER__
+            ALIGNMODE k8
+        %else
+            CPU amdnop
+        %endif
+    %else
+        %ifdef __NASM_VER__
+            ALIGNMODE nop
+        %else
+            CPU basicnop
+        %endif
+    %endif
+%endmacro
+
+; Merge mmx and sse*
+; m# is a simd register of the currently selected size
+; xm# is the corresponding xmm register if mmsize >= 16, otherwise the same as m#
+; ym# is the corresponding ymm register if mmsize >= 32, otherwise the same as m#
+; (All 3 remain in sync through SWAP.)
+
+%macro CAT_XDEFINE 3
+    %xdefine %1%2 %3
+%endmacro
+
+%macro CAT_UNDEF 2
+    %undef %1%2
+%endmacro
+
+%macro INIT_MMX 0-1+
+    %assign avx_enabled 0
+    %define RESET_MM_PERMUTATION INIT_MMX %1
+    %define mmsize 8
+    %define num_mmregs 8
+    %define mova movq
+    %define movu movq
+    %define movh movd
+    %define movnta movntq
+    %assign %%i 0
+    %rep 8
+        CAT_XDEFINE m, %%i, mm %+ %%i
+        CAT_XDEFINE nnmm, %%i, %%i
+        %assign %%i %%i+1
+    %endrep
+    %rep 8
+        CAT_UNDEF m, %%i
+        CAT_UNDEF nnmm, %%i
+        %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_XMM 0-1+
+    %assign avx_enabled 0
+    %define RESET_MM_PERMUTATION INIT_XMM %1
+    %define mmsize 16
+    %define num_mmregs 8
+    %if ARCH_X86_64
+        %define num_mmregs 16
+    %endif
+    %define mova movdqa
+    %define movu movdqu
+    %define movh movq
+    %define movnta movntdq
+    %assign %%i 0
+    %rep num_mmregs
+        CAT_XDEFINE m, %%i, xmm %+ %%i
+        CAT_XDEFINE nnxmm, %%i, %%i
+        %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_YMM 0-1+
+    %assign avx_enabled 1
+    %define RESET_MM_PERMUTATION INIT_YMM %1
+    %define mmsize 32
+    %define num_mmregs 8
+    %if ARCH_X86_64
+        %define num_mmregs 16
+    %endif
+    %define mova movdqa
+    %define movu movdqu
+    %undef movh
+    %define movnta movntdq
+    %assign %%i 0
+    %rep num_mmregs
+        CAT_XDEFINE m, %%i, ymm %+ %%i
+        CAT_XDEFINE nnymm, %%i, %%i
+        %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+INIT_XMM
+
+%macro DECLARE_MMCAST 1
+    %define  mmmm%1   mm%1
+    %define  mmxmm%1  mm%1
+    %define  mmymm%1  mm%1
+    %define xmmmm%1   mm%1
+    %define xmmxmm%1 xmm%1
+    %define xmmymm%1 xmm%1
+    %define ymmmm%1   mm%1
+    %define ymmxmm%1 xmm%1
+    %define ymmymm%1 ymm%1
+    %define xm%1 xmm %+ m%1
+    %define ym%1 ymm %+ m%1
+%endmacro
+
+%assign i 0
+%rep 16
+    DECLARE_MMCAST i
+    %assign i i+1
+%endrep
+
+; I often want to use macros that permute their arguments. e.g. there's no
+; efficient way to implement butterfly or transpose or dct without swapping some
+; arguments.
+;
+; I would like to not have to manually keep track of the permutations:
+; If I insert a permutation in the middle of a function, it should automatically
+; change everything that follows. For more complex macros I may also have multiple
+; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
+;
+; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
+; permutes its arguments. It's equivalent to exchanging the contents of the
+; registers, except that this way you exchange the register names instead, so it
+; doesn't cost any cycles.
+
+%macro PERMUTE 2-* ; takes a list of pairs to swap
+    %rep %0/2
+        %xdefine %%tmp%2 m%2
+        %rotate 2
+    %endrep
+    %rep %0/2
+        %xdefine m%1 %%tmp%2
+        CAT_XDEFINE nn, m%1, %1
+        %rotate 2
+    %endrep
+%endmacro
+
+%macro SWAP 2+ ; swaps a single chain (sometimes more concise than pairs)
+    %ifnum %1 ; SWAP 0, 1, ...
+        SWAP_INTERNAL_NUM %1, %2
+    %else ; SWAP m0, m1, ...
+        SWAP_INTERNAL_NAME %1, %2
+    %endif
+%endmacro
+
+%macro SWAP_INTERNAL_NUM 2-*
+    %rep %0-1
+        %xdefine %%tmp m%1
+        %xdefine m%1 m%2
+        %xdefine m%2 %%tmp
+        CAT_XDEFINE nn, m%1, %1
+        CAT_XDEFINE nn, m%2, %2
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro SWAP_INTERNAL_NAME 2-*
+    %xdefine %%args nn %+ %1
+    %rep %0-1
+        %xdefine %%args %%args, nn %+ %2
+        %rotate 1
+    %endrep
+    SWAP_INTERNAL_NUM %%args
+%endmacro
+
+; If SAVE_MM_PERMUTATION is placed at the end of a function, then any later
+; calls to that function will automatically load the permutation, so values can
+; be returned in mmregs.
+%macro SAVE_MM_PERMUTATION 0-1
+    %if %0
+        %xdefine %%f %1_m
+    %else
+        %xdefine %%f current_function %+ _m
+    %endif
+    %assign %%i 0
+    %rep num_mmregs
+        CAT_XDEFINE %%f, %%i, m %+ %%i
+        %assign %%i %%i+1
+    %endrep
+%endmacro
+
+%macro LOAD_MM_PERMUTATION 1 ; name to load from
+    %ifdef %1_m0
+        %assign %%i 0
+        %rep num_mmregs
+            CAT_XDEFINE m, %%i, %1_m %+ %%i
+            CAT_XDEFINE nn, m %+ %%i, %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+%endmacro
+
+; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't
+%macro call 1
+    call_internal %1 %+ SUFFIX, %1
+%endmacro
+%macro call_internal 2
+    %xdefine %%i %2
+    %ifndef cglobaled_%2
+        %ifdef cglobaled_%1
+            %xdefine %%i %1
+        %endif
+    %endif
+    call %%i
+    LOAD_MM_PERMUTATION %%i
+%endmacro
+
+; Substitutions that reduce instruction size but are functionally equivalent
+%macro add 2
+    %ifnum %2
+        %if %2==128
+            sub %1, -128
+        %else
+            add %1, %2
+        %endif
+    %else
+        add %1, %2
+    %endif
+%endmacro
+
+%macro sub 2
+    %ifnum %2
+        %if %2==128
+            add %1, -128
+        %else
+            sub %1, %2
+        %endif
+    %else
+        sub %1, %2
+    %endif
+%endmacro
+
+;=============================================================================
+; AVX abstraction layer
+;=============================================================================
+
+%assign i 0
+%rep 16
+    %if i < 8
+        CAT_XDEFINE sizeofmm, i, 8
+    %endif
+    CAT_XDEFINE sizeofxmm, i, 16
+    CAT_XDEFINE sizeofymm, i, 32
+    %assign i i+1
+%endrep
+%undef i
+
+%macro CHECK_AVX_INSTR_EMU 3-*
+    %xdefine %%opcode %1
+    %xdefine %%dst %2
+    %rep %0-2
+        %ifidn %%dst, %3
+            %error non-avx emulation of ``%%opcode'' is not supported
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+;%1 == instruction
+;%2 == minimal instruction set
+;%3 == 1 if float, 0 if int
+;%4 == 1 if non-destructive or 4-operand (xmm, xmm, xmm, imm), 0 otherwise
+;%5 == 1 if commutative (i.e. doesn't matter which src arg is which), 0 if not
+;%6+: operands
+%macro RUN_AVX_INSTR 6-9+
+    %ifnum sizeof%7
+        %assign __sizeofreg sizeof%7
+    %elifnum sizeof%6
+        %assign __sizeofreg sizeof%6
+    %else
+        %assign __sizeofreg mmsize
+    %endif
+    %assign __emulate_avx 0
+    %if avx_enabled && __sizeofreg >= 16
+        %xdefine __instr v%1
+    %else
+        %xdefine __instr %1
+        %if %0 >= 8+%4
+            %assign __emulate_avx 1
+        %endif
+    %endif
+    %ifnidn %2, fnord
+        %ifdef cpuname
+            %if notcpuflag(%2)
+                %error use of ``%1'' %2 instruction in cpuname function: current_function
+            %elif cpuflags_%2 < cpuflags_sse && notcpuflag(sse2) && __sizeofreg > 8
+                %error use of ``%1'' sse2 instruction in cpuname function: current_function
+            %endif
+        %endif
+    %endif
+
+    %if __emulate_avx
+        %xdefine __src1 %7
+        %xdefine __src2 %8
+        %ifnidn %6, %7
+            %if %0 >= 9
+                CHECK_AVX_INSTR_EMU {%1 %6, %7, %8, %9}, %6, %8, %9
+            %else
+                CHECK_AVX_INSTR_EMU {%1 %6, %7, %8}, %6, %8
+            %endif
+            %if %5 && %4 == 0
+                %ifnid %8
+                    ; 3-operand AVX instructions with a memory arg can only have it in src2,
+                    ; whereas SSE emulation prefers to have it in src1 (i.e. the mov).
+                    ; So, if the instruction is commutative with a memory arg, swap them.
+                    %xdefine __src1 %8
+                    %xdefine __src2 %7
+                %endif
+            %endif
+            %if __sizeofreg == 8
+                MOVQ %6, __src1
+            %elif %3
+                MOVAPS %6, __src1
+            %else
+                MOVDQA %6, __src1
+            %endif
+        %endif
+        %if %0 >= 9
+            %1 %6, __src2, %9
+        %else
+            %1 %6, __src2
+        %endif
+    %elif %0 >= 9
+        __instr %6, %7, %8, %9
+    %elif %0 == 8
+        __instr %6, %7, %8
+    %elif %0 == 7
+        __instr %6, %7
+    %else
+        __instr %6
+    %endif
+%endmacro
+
+;%1 == instruction
+;%2 == minimal instruction set
+;%3 == 1 if float, 0 if int
+;%4 == 1 if non-destructive or 4-operand (xmm, xmm, xmm, imm), 0 otherwise
+;%5 == 1 if commutative (i.e. doesn't matter which src arg is which), 0 if not
+%macro AVX_INSTR 1-5 fnord, 0, 1, 0
+    %macro %1 1-10 fnord, fnord, fnord, fnord, %1, %2, %3, %4, %5
+        %ifidn %2, fnord
+            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1
+        %elifidn %3, fnord
+            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2
+        %elifidn %4, fnord
+            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2, %3
+        %elifidn %5, fnord
+            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2, %3, %4
+        %else
+            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2, %3, %4, %5
+        %endif
+    %endmacro
+%endmacro
+
+; Instructions with both VEX and non-VEX encodings
+; Non-destructive instructions are written without parameters
+AVX_INSTR addpd, sse2, 1, 0, 1
+AVX_INSTR addps, sse, 1, 0, 1
+AVX_INSTR addsd, sse2, 1, 0, 1
+AVX_INSTR addss, sse, 1, 0, 1
+AVX_INSTR addsubpd, sse3, 1, 0, 0
+AVX_INSTR addsubps, sse3, 1, 0, 0
+AVX_INSTR aesdec, fnord, 0, 0, 0
+AVX_INSTR aesdeclast, fnord, 0, 0, 0
+AVX_INSTR aesenc, fnord, 0, 0, 0
+AVX_INSTR aesenclast, fnord, 0, 0, 0
+AVX_INSTR aesimc
+AVX_INSTR aeskeygenassist
+AVX_INSTR andnpd, sse2, 1, 0, 0
+AVX_INSTR andnps, sse, 1, 0, 0
+AVX_INSTR andpd, sse2, 1, 0, 1
+AVX_INSTR andps, sse, 1, 0, 1
+AVX_INSTR blendpd, sse4, 1, 0, 0
+AVX_INSTR blendps, sse4, 1, 0, 0
+AVX_INSTR blendvpd, sse4, 1, 0, 0
+AVX_INSTR blendvps, sse4, 1, 0, 0
+AVX_INSTR cmppd, sse2, 1, 1, 0
+AVX_INSTR cmpps, sse, 1, 1, 0
+AVX_INSTR cmpsd, sse2, 1, 1, 0
+AVX_INSTR cmpss, sse, 1, 1, 0
+AVX_INSTR comisd, sse2
+AVX_INSTR comiss, sse
+AVX_INSTR cvtdq2pd, sse2
+AVX_INSTR cvtdq2ps, sse2
+AVX_INSTR cvtpd2dq, sse2
+AVX_INSTR cvtpd2ps, sse2
+AVX_INSTR cvtps2dq, sse2
+AVX_INSTR cvtps2pd, sse2
+AVX_INSTR cvtsd2si, sse2
+AVX_INSTR cvtsd2ss, sse2
+AVX_INSTR cvtsi2sd, sse2
+AVX_INSTR cvtsi2ss, sse
+AVX_INSTR cvtss2sd, sse2
+AVX_INSTR cvtss2si, sse
+AVX_INSTR cvttpd2dq, sse2
+AVX_INSTR cvttps2dq, sse2
+AVX_INSTR cvttsd2si, sse2
+AVX_INSTR cvttss2si, sse
+AVX_INSTR divpd, sse2, 1, 0, 0
+AVX_INSTR divps, sse, 1, 0, 0
+AVX_INSTR divsd, sse2, 1, 0, 0
+AVX_INSTR divss, sse, 1, 0, 0
+AVX_INSTR dppd, sse4, 1, 1, 0
+AVX_INSTR dpps, sse4, 1, 1, 0
+AVX_INSTR extractps, sse4
+AVX_INSTR haddpd, sse3, 1, 0, 0
+AVX_INSTR haddps, sse3, 1, 0, 0
+AVX_INSTR hsubpd, sse3, 1, 0, 0
+AVX_INSTR hsubps, sse3, 1, 0, 0
+AVX_INSTR insertps, sse4, 1, 1, 0
+AVX_INSTR lddqu, sse3
+AVX_INSTR ldmxcsr, sse
+AVX_INSTR maskmovdqu, sse2
+AVX_INSTR maxpd, sse2, 1, 0, 1
+AVX_INSTR maxps, sse, 1, 0, 1
+AVX_INSTR maxsd, sse2, 1, 0, 1
+AVX_INSTR maxss, sse, 1, 0, 1
+AVX_INSTR minpd, sse2, 1, 0, 1
+AVX_INSTR minps, sse, 1, 0, 1
+AVX_INSTR minsd, sse2, 1, 0, 1
+AVX_INSTR minss, sse, 1, 0, 1
+AVX_INSTR movapd, sse2
+AVX_INSTR movaps, sse
+AVX_INSTR movd, mmx
+AVX_INSTR movddup, sse3
+AVX_INSTR movdqa, sse2
+AVX_INSTR movdqu, sse2
+AVX_INSTR movhlps, sse, 1, 0, 0
+AVX_INSTR movhpd, sse2, 1, 0, 0
+AVX_INSTR movhps, sse, 1, 0, 0
+AVX_INSTR movlhps, sse, 1, 0, 0
+AVX_INSTR movlpd, sse2, 1, 0, 0
+AVX_INSTR movlps, sse, 1, 0, 0
+AVX_INSTR movmskpd, sse2
+AVX_INSTR movmskps, sse
+AVX_INSTR movntdq, sse2
+AVX_INSTR movntdqa, sse4
+AVX_INSTR movntpd, sse2
+AVX_INSTR movntps, sse
+AVX_INSTR movq, mmx
+AVX_INSTR movsd, sse2, 1, 0, 0
+AVX_INSTR movshdup, sse3
+AVX_INSTR movsldup, sse3
+AVX_INSTR movss, sse, 1, 0, 0
+AVX_INSTR movupd, sse2
+AVX_INSTR movups, sse
+AVX_INSTR mpsadbw, sse4
+AVX_INSTR mulpd, sse2, 1, 0, 1
+AVX_INSTR mulps, sse, 1, 0, 1
+AVX_INSTR mulsd, sse2, 1, 0, 1
+AVX_INSTR mulss, sse, 1, 0, 1
+AVX_INSTR orpd, sse2, 1, 0, 1
+AVX_INSTR orps, sse, 1, 0, 1
+AVX_INSTR pabsb, ssse3
+AVX_INSTR pabsd, ssse3
+AVX_INSTR pabsw, ssse3
+AVX_INSTR packsswb, mmx, 0, 0, 0
+AVX_INSTR packssdw, mmx, 0, 0, 0
+AVX_INSTR packuswb, mmx, 0, 0, 0
+AVX_INSTR packusdw, sse4, 0, 0, 0
+AVX_INSTR paddb, mmx, 0, 0, 1
+AVX_INSTR paddw, mmx, 0, 0, 1
+AVX_INSTR paddd, mmx, 0, 0, 1
+AVX_INSTR paddq, sse2, 0, 0, 1
+AVX_INSTR paddsb, mmx, 0, 0, 1
+AVX_INSTR paddsw, mmx, 0, 0, 1
+AVX_INSTR paddusb, mmx, 0, 0, 1
+AVX_INSTR paddusw, mmx, 0, 0, 1
+AVX_INSTR palignr, ssse3
+AVX_INSTR pand, mmx, 0, 0, 1
+AVX_INSTR pandn, mmx, 0, 0, 0
+AVX_INSTR pavgb, mmx2, 0, 0, 1
+AVX_INSTR pavgw, mmx2, 0, 0, 1
+AVX_INSTR pblendvb, sse4, 0, 0, 0
+AVX_INSTR pblendw, sse4
+AVX_INSTR pclmulqdq
+AVX_INSTR pcmpestri, sse42
+AVX_INSTR pcmpestrm, sse42
+AVX_INSTR pcmpistri, sse42
+AVX_INSTR pcmpistrm, sse42
+AVX_INSTR pcmpeqb, mmx, 0, 0, 1
+AVX_INSTR pcmpeqw, mmx, 0, 0, 1
+AVX_INSTR pcmpeqd, mmx, 0, 0, 1
+AVX_INSTR pcmpeqq, sse4, 0, 0, 1
+AVX_INSTR pcmpgtb, mmx, 0, 0, 0
+AVX_INSTR pcmpgtw, mmx, 0, 0, 0
+AVX_INSTR pcmpgtd, mmx, 0, 0, 0
+AVX_INSTR pcmpgtq, sse42, 0, 0, 0
+AVX_INSTR pextrb, sse4
+AVX_INSTR pextrd, sse4
+AVX_INSTR pextrq, sse4
+AVX_INSTR pextrw, mmx2
+AVX_INSTR phaddw, ssse3, 0, 0, 0
+AVX_INSTR phaddd, ssse3, 0, 0, 0
+AVX_INSTR phaddsw, ssse3, 0, 0, 0
+AVX_INSTR phminposuw, sse4
+AVX_INSTR phsubw, ssse3, 0, 0, 0
+AVX_INSTR phsubd, ssse3, 0, 0, 0
+AVX_INSTR phsubsw, ssse3, 0, 0, 0
+AVX_INSTR pinsrb, sse4
+AVX_INSTR pinsrd, sse4
+AVX_INSTR pinsrq, sse4
+AVX_INSTR pinsrw, mmx2
+AVX_INSTR pmaddwd, mmx, 0, 0, 1
+AVX_INSTR pmaddubsw, ssse3, 0, 0, 0
+AVX_INSTR pmaxsb, sse4, 0, 0, 1
+AVX_INSTR pmaxsw, mmx2, 0, 0, 1
+AVX_INSTR pmaxsd, sse4, 0, 0, 1
+AVX_INSTR pmaxub, mmx2, 0, 0, 1
+AVX_INSTR pmaxuw, sse4, 0, 0, 1
+AVX_INSTR pmaxud, sse4, 0, 0, 1
+AVX_INSTR pminsb, sse4, 0, 0, 1
+AVX_INSTR pminsw, mmx2, 0, 0, 1
+AVX_INSTR pminsd, sse4, 0, 0, 1
+AVX_INSTR pminub, mmx2, 0, 0, 1
+AVX_INSTR pminuw, sse4, 0, 0, 1
+AVX_INSTR pminud, sse4, 0, 0, 1
+AVX_INSTR pmovmskb, mmx2
+AVX_INSTR pmovsxbw, sse4
+AVX_INSTR pmovsxbd, sse4
+AVX_INSTR pmovsxbq, sse4
+AVX_INSTR pmovsxwd, sse4
+AVX_INSTR pmovsxwq, sse4
+AVX_INSTR pmovsxdq, sse4
+AVX_INSTR pmovzxbw, sse4
+AVX_INSTR pmovzxbd, sse4
+AVX_INSTR pmovzxbq, sse4
+AVX_INSTR pmovzxwd, sse4
+AVX_INSTR pmovzxwq, sse4
+AVX_INSTR pmovzxdq, sse4
+AVX_INSTR pmuldq, sse4, 0, 0, 1
+AVX_INSTR pmulhrsw, ssse3, 0, 0, 1
+AVX_INSTR pmulhuw, mmx2, 0, 0, 1
+AVX_INSTR pmulhw, mmx, 0, 0, 1
+AVX_INSTR pmullw, mmx, 0, 0, 1
+AVX_INSTR pmulld, sse4, 0, 0, 1
+AVX_INSTR pmuludq, sse2, 0, 0, 1
+AVX_INSTR por, mmx, 0, 0, 1
+AVX_INSTR psadbw, mmx2, 0, 0, 1
+AVX_INSTR pshufb, ssse3, 0, 0, 0
+AVX_INSTR pshufd, sse2
+AVX_INSTR pshufhw, sse2
+AVX_INSTR pshuflw, sse2
+AVX_INSTR psignb, ssse3, 0, 0, 0
+AVX_INSTR psignw, ssse3, 0, 0, 0
+AVX_INSTR psignd, ssse3, 0, 0, 0
+AVX_INSTR psllw, mmx, 0, 0, 0
+AVX_INSTR pslld, mmx, 0, 0, 0
+AVX_INSTR psllq, mmx, 0, 0, 0
+AVX_INSTR pslldq, sse2, 0, 0, 0
+AVX_INSTR psraw, mmx, 0, 0, 0
+AVX_INSTR psrad, mmx, 0, 0, 0
+AVX_INSTR psrlw, mmx, 0, 0, 0
+AVX_INSTR psrld, mmx, 0, 0, 0
+AVX_INSTR psrlq, mmx, 0, 0, 0
+AVX_INSTR psrldq, sse2, 0, 0, 0
+AVX_INSTR psubb, mmx, 0, 0, 0
+AVX_INSTR psubw, mmx, 0, 0, 0
+AVX_INSTR psubd, mmx, 0, 0, 0
+AVX_INSTR psubq, sse2, 0, 0, 0
+AVX_INSTR psubsb, mmx, 0, 0, 0
+AVX_INSTR psubsw, mmx, 0, 0, 0
+AVX_INSTR psubusb, mmx, 0, 0, 0
+AVX_INSTR psubusw, mmx, 0, 0, 0
+AVX_INSTR ptest, sse4
+AVX_INSTR punpckhbw, mmx, 0, 0, 0
+AVX_INSTR punpckhwd, mmx, 0, 0, 0
+AVX_INSTR punpckhdq, mmx, 0, 0, 0
+AVX_INSTR punpckhqdq, sse2, 0, 0, 0
+AVX_INSTR punpcklbw, mmx, 0, 0, 0
+AVX_INSTR punpcklwd, mmx, 0, 0, 0
+AVX_INSTR punpckldq, mmx, 0, 0, 0
+AVX_INSTR punpcklqdq, sse2, 0, 0, 0
+AVX_INSTR pxor, mmx, 0, 0, 1
+AVX_INSTR rcpps, sse, 1, 0, 0
+AVX_INSTR rcpss, sse, 1, 0, 0
+AVX_INSTR roundpd, sse4
+AVX_INSTR roundps, sse4
+AVX_INSTR roundsd, sse4
+AVX_INSTR roundss, sse4
+AVX_INSTR rsqrtps, sse, 1, 0, 0
+AVX_INSTR rsqrtss, sse, 1, 0, 0
+AVX_INSTR shufpd, sse2, 1, 1, 0
+AVX_INSTR shufps, sse, 1, 1, 0
+AVX_INSTR sqrtpd, sse2, 1, 0, 0
+AVX_INSTR sqrtps, sse, 1, 0, 0
+AVX_INSTR sqrtsd, sse2, 1, 0, 0
+AVX_INSTR sqrtss, sse, 1, 0, 0
+AVX_INSTR stmxcsr, sse
+AVX_INSTR subpd, sse2, 1, 0, 0
+AVX_INSTR subps, sse, 1, 0, 0
+AVX_INSTR subsd, sse2, 1, 0, 0
+AVX_INSTR subss, sse, 1, 0, 0
+AVX_INSTR ucomisd, sse2
+AVX_INSTR ucomiss, sse
+AVX_INSTR unpckhpd, sse2, 1, 0, 0
+AVX_INSTR unpckhps, sse, 1, 0, 0
+AVX_INSTR unpcklpd, sse2, 1, 0, 0
+AVX_INSTR unpcklps, sse, 1, 0, 0
+AVX_INSTR xorpd, sse2, 1, 0, 1
+AVX_INSTR xorps, sse, 1, 0, 1
+
+; 3DNow instructions, for sharing code between AVX, SSE and 3DN
+AVX_INSTR pfadd, 3dnow, 1, 0, 1
+AVX_INSTR pfsub, 3dnow, 1, 0, 0
+AVX_INSTR pfmul, 3dnow, 1, 0, 1
+
+; base-4 constants for shuffles
+%assign i 0
+%rep 256
+    %assign j ((i>>6)&3)*1000 + ((i>>4)&3)*100 + ((i>>2)&3)*10 + (i&3)
+    %if j < 10
+        CAT_XDEFINE q000, j, i
+    %elif j < 100
+        CAT_XDEFINE q00, j, i
+    %elif j < 1000
+        CAT_XDEFINE q0, j, i
+    %else
+        CAT_XDEFINE q, j, i
+    %endif
+    %assign i i+1
+%endrep
+%undef i
+%undef j
+
+%macro FMA_INSTR 3
+    %macro %1 4-7 %1, %2, %3
+        %if cpuflag(xop)
+            v%5 %1, %2, %3, %4
+        %elifnidn %1, %4
+            %6 %1, %2, %3
+            %7 %1, %4
+        %else
+            %error non-xop emulation of ``%5 %1, %2, %3, %4'' is not supported
+        %endif
+    %endmacro
+%endmacro
+
+FMA_INSTR  pmacsww,  pmullw, paddw
+FMA_INSTR  pmacsdd,  pmulld, paddd ; sse4 emulation
+FMA_INSTR pmacsdql,  pmuldq, paddq ; sse4 emulation
+FMA_INSTR pmadcswd, pmaddwd, paddd
+
+; Macros for consolidating FMA3 and FMA4 using 4-operand (dst, src1, src2, src3) syntax.
+; FMA3 is only possible if dst is the same as one of the src registers.
+; Either src2 or src3 can be a memory operand.
+%macro FMA4_INSTR 2-*
+    %push fma4_instr
+    %xdefine %$prefix %1
+    %rep %0 - 1
+        %macro %$prefix%2 4-6 %$prefix, %2
+            %if notcpuflag(fma3) && notcpuflag(fma4)
+                %error use of ``%5%6'' fma instruction in cpuname function: current_function
+            %elif cpuflag(fma4)
+                v%5%6 %1, %2, %3, %4
+            %elifidn %1, %2
+                ; If %3 or %4 is a memory operand it needs to be encoded as the last operand.
+                %ifid %3
+                    v%{5}213%6 %2, %3, %4
+                %else
+                    v%{5}132%6 %2, %4, %3
+                %endif
+            %elifidn %1, %3
+                v%{5}213%6 %3, %2, %4
+            %elifidn %1, %4
+                v%{5}231%6 %4, %2, %3
+            %else
+                %error fma3 emulation of ``%5%6 %1, %2, %3, %4'' is not supported
+            %endif
+        %endmacro
+        %rotate 1
+    %endrep
+    %pop
+%endmacro
+
+FMA4_INSTR fmadd,    pd, ps, sd, ss
+FMA4_INSTR fmaddsub, pd, ps
+FMA4_INSTR fmsub,    pd, ps, sd, ss
+FMA4_INSTR fmsubadd, pd, ps
+FMA4_INSTR fnmadd,   pd, ps, sd, ss
+FMA4_INSTR fnmsub,   pd, ps, sd, ss
+
+; workaround: vpbroadcastq is broken in x86_32 due to a yasm bug (fixed in 1.3.0)
+%ifdef __YASM_VER__
+    %if __YASM_VERSION_ID__ < 0x01030000 && ARCH_X86_64 == 0
+        %macro vpbroadcastq 2
+            %if sizeof%1 == 16
+                movddup %1, %2
+            %else
+                vbroadcastsd %1, %2
+            %endif
+        %endmacro
+    %endif
+%endif
diff --git a/libvpx/libvpx/tools/all_builds.py b/libvpx/libvpx/tools/all_builds.py
new file mode 100755
index 0000000..d1f0c80
--- /dev/null
+++ b/libvpx/libvpx/tools/all_builds.py
@@ -0,0 +1,72 @@
+#!/usr/bin/python
+
+import getopt
+import subprocess
+import sys
+
+LONG_OPTIONS = ["shard=", "shards="]
+BASE_COMMAND = "./configure --enable-internal-stats --enable-experimental"
+
+def RunCommand(command):
+  run = subprocess.Popen(command, shell=True)
+  output = run.communicate()
+  if run.returncode:
+    print "Non-zero return code: " + str(run.returncode) + " => exiting!"
+    sys.exit(1)
+
+def list_of_experiments():
+  experiments = []
+  configure_file = open("configure")
+  list_start = False
+  for line in configure_file.read().split("\n"):
+    if line == 'EXPERIMENT_LIST="':
+      list_start = True
+    elif line == '"':
+      list_start = False
+    elif list_start:
+      currently_broken = ["csm"]
+      experiment = line[4:]
+      if experiment not in currently_broken:
+        experiments.append(experiment)
+  return experiments
+
+def main(argv):
+  # Parse arguments
+  options = {"--shard": 0, "--shards": 1}
+  if "--" in argv:
+    opt_end_index = argv.index("--")
+  else:
+    opt_end_index = len(argv)
+  try:
+    o, _ = getopt.getopt(argv[1:opt_end_index], None, LONG_OPTIONS)
+  except getopt.GetoptError, err:
+    print str(err)
+    print "Usage: %s [--shard=<n> --shards=<n>] -- [configure flag ...]"%argv[0]
+    sys.exit(2)
+
+  options.update(o)
+  extra_args = argv[opt_end_index + 1:]
+
+  # Shard experiment list
+  shard = int(options["--shard"])
+  shards = int(options["--shards"])
+  experiments = list_of_experiments()
+  base_command = " ".join([BASE_COMMAND] + extra_args)
+  configs = [base_command]
+  configs += ["%s --enable-%s" % (base_command, e) for e in experiments]
+  my_configs = zip(configs, range(len(configs)))
+  my_configs = filter(lambda x: x[1] % shards == shard, my_configs)
+  my_configs = [e[0] for e in my_configs]
+
+  # Run configs for this shard
+  for config in my_configs:
+    test_build(config)
+
+def test_build(configure_command):
+  print "\033[34m\033[47mTesting %s\033[0m" % (configure_command)
+  RunCommand(configure_command)
+  RunCommand("make clean")
+  RunCommand("make")
+
+if __name__ == "__main__":
+  main(sys.argv)
diff --git a/libvpx/libvpx/tools/author_first_release.sh b/libvpx/libvpx/tools/author_first_release.sh
new file mode 100755
index 0000000..7b0b797
--- /dev/null
+++ b/libvpx/libvpx/tools/author_first_release.sh
@@ -0,0 +1,15 @@
+#!/bin/bash
+##
+## List the release each author first contributed to.
+##
+## Usage: author_first_release.sh [TAGS]
+##
+## If the TAGS arguments are unspecified, all tags reported by `git tag`
+## will be considered.
+##
+tags=${@:-$(git tag)}
+for tag in $tags; do
+  git shortlog -n -e -s $tag |
+      cut -f2- |
+      awk "{print \"${tag#v}\t\"\$0}"
+done | sort -k2  | uniq -f2
diff --git a/libvpx/libvpx/tools/cpplint.py b/libvpx/libvpx/tools/cpplint.py
new file mode 100755
index 0000000..25fbef7
--- /dev/null
+++ b/libvpx/libvpx/tools/cpplint.py
@@ -0,0 +1,4756 @@
+#!/usr/bin/python
+#
+# Copyright (c) 2009 Google Inc. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#    * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#    * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#    * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Does google-lint on c++ files.
+
+The goal of this script is to identify places in the code that *may*
+be in non-compliance with google style.  It does not attempt to fix
+up these problems -- the point is to educate.  It does also not
+attempt to find all problems, or to ensure that everything it does
+find is legitimately a problem.
+
+In particular, we can get very confused by /* and // inside strings!
+We do a small hack, which is to ignore //'s with "'s after them on the
+same line, but it is far from perfect (in either direction).
+"""
+
+import codecs
+import copy
+import getopt
+import math  # for log
+import os
+import re
+import sre_compile
+import string
+import sys
+import unicodedata
+
+
+_USAGE = """
+Syntax: cpplint.py [--verbose=#] [--output=vs7] [--filter=-x,+y,...]
+                   [--counting=total|toplevel|detailed] [--root=subdir]
+                   [--linelength=digits]
+        <file> [file] ...
+
+  The style guidelines this tries to follow are those in
+    http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml
+
+  Every problem is given a confidence score from 1-5, with 5 meaning we are
+  certain of the problem, and 1 meaning it could be a legitimate construct.
+  This will miss some errors, and is not a substitute for a code review.
+
+  To suppress false-positive errors of a certain category, add a
+  'NOLINT(category)' comment to the line.  NOLINT or NOLINT(*)
+  suppresses errors of all categories on that line.
+
+  The files passed in will be linted; at least one file must be provided.
+  Default linted extensions are .cc, .cpp, .cu, .cuh and .h.  Change the
+  extensions with the --extensions flag.
+
+  Flags:
+
+    output=vs7
+      By default, the output is formatted to ease emacs parsing.  Visual Studio
+      compatible output (vs7) may also be used.  Other formats are unsupported.
+
+    verbose=#
+      Specify a number 0-5 to restrict errors to certain verbosity levels.
+
+    filter=-x,+y,...
+      Specify a comma-separated list of category-filters to apply: only
+      error messages whose category names pass the filters will be printed.
+      (Category names are printed with the message and look like
+      "[whitespace/indent]".)  Filters are evaluated left to right.
+      "-FOO" and "FOO" means "do not print categories that start with FOO".
+      "+FOO" means "do print categories that start with FOO".
+
+      Examples: --filter=-whitespace,+whitespace/braces
+                --filter=whitespace,runtime/printf,+runtime/printf_format
+                --filter=-,+build/include_what_you_use
+
+      To see a list of all the categories used in cpplint, pass no arg:
+         --filter=
+
+    counting=total|toplevel|detailed
+      The total number of errors found is always printed. If
+      'toplevel' is provided, then the count of errors in each of
+      the top-level categories like 'build' and 'whitespace' will
+      also be printed. If 'detailed' is provided, then a count
+      is provided for each category like 'build/class'.
+
+    root=subdir
+      The root directory used for deriving header guard CPP variable.
+      By default, the header guard CPP variable is calculated as the relative
+      path to the directory that contains .git, .hg, or .svn.  When this flag
+      is specified, the relative path is calculated from the specified
+      directory. If the specified directory does not exist, this flag is
+      ignored.
+
+      Examples:
+        Assuing that src/.git exists, the header guard CPP variables for
+        src/chrome/browser/ui/browser.h are:
+
+        No flag => CHROME_BROWSER_UI_BROWSER_H_
+        --root=chrome => BROWSER_UI_BROWSER_H_
+        --root=chrome/browser => UI_BROWSER_H_
+
+    linelength=digits
+      This is the allowed line length for the project. The default value is
+      80 characters.
+
+      Examples:
+        --linelength=120
+
+    extensions=extension,extension,...
+      The allowed file extensions that cpplint will check
+
+      Examples:
+        --extensions=hpp,cpp
+"""
+
+# We categorize each error message we print.  Here are the categories.
+# We want an explicit list so we can list them all in cpplint --filter=.
+# If you add a new error message with a new category, add it to the list
+# here!  cpplint_unittest.py should tell you if you forget to do this.
+_ERROR_CATEGORIES = [
+  'build/class',
+  'build/deprecated',
+  'build/endif_comment',
+  'build/explicit_make_pair',
+  'build/forward_decl',
+  'build/header_guard',
+  'build/include',
+  'build/include_alpha',
+  'build/include_order',
+  'build/include_what_you_use',
+  'build/namespaces',
+  'build/printf_format',
+  'build/storage_class',
+  'legal/copyright',
+  'readability/alt_tokens',
+  'readability/braces',
+  'readability/casting',
+  'readability/check',
+  'readability/constructors',
+  'readability/fn_size',
+  'readability/function',
+  'readability/multiline_comment',
+  'readability/multiline_string',
+  'readability/namespace',
+  'readability/nolint',
+  'readability/nul',
+  'readability/streams',
+  'readability/todo',
+  'readability/utf8',
+  'runtime/arrays',
+  'runtime/casting',
+  'runtime/explicit',
+  'runtime/int',
+  'runtime/init',
+  'runtime/invalid_increment',
+  'runtime/member_string_references',
+  'runtime/memset',
+  'runtime/operator',
+  'runtime/printf',
+  'runtime/printf_format',
+  'runtime/references',
+  'runtime/sizeof',
+  'runtime/string',
+  'runtime/threadsafe_fn',
+  'runtime/vlog',
+  'whitespace/blank_line',
+  'whitespace/braces',
+  'whitespace/comma',
+  'whitespace/comments',
+  'whitespace/empty_conditional_body',
+  'whitespace/empty_loop_body',
+  'whitespace/end_of_line',
+  'whitespace/ending_newline',
+  'whitespace/forcolon',
+  'whitespace/indent',
+  'whitespace/line_length',
+  'whitespace/newline',
+  'whitespace/operators',
+  'whitespace/parens',
+  'whitespace/semicolon',
+  'whitespace/tab',
+  'whitespace/todo'
+  ]
+
+# The default state of the category filter. This is overrided by the --filter=
+# flag. By default all errors are on, so only add here categories that should be
+# off by default (i.e., categories that must be enabled by the --filter= flags).
+# All entries here should start with a '-' or '+', as in the --filter= flag.
+_DEFAULT_FILTERS = ['-build/include_alpha']
+
+# We used to check for high-bit characters, but after much discussion we
+# decided those were OK, as long as they were in UTF-8 and didn't represent
+# hard-coded international strings, which belong in a separate i18n file.
+
+
+# C++ headers
+_CPP_HEADERS = frozenset([
+    # Legacy
+    'algobase.h',
+    'algo.h',
+    'alloc.h',
+    'builtinbuf.h',
+    'bvector.h',
+    'complex.h',
+    'defalloc.h',
+    'deque.h',
+    'editbuf.h',
+    'fstream.h',
+    'function.h',
+    'hash_map',
+    'hash_map.h',
+    'hash_set',
+    'hash_set.h',
+    'hashtable.h',
+    'heap.h',
+    'indstream.h',
+    'iomanip.h',
+    'iostream.h',
+    'istream.h',
+    'iterator.h',
+    'list.h',
+    'map.h',
+    'multimap.h',
+    'multiset.h',
+    'ostream.h',
+    'pair.h',
+    'parsestream.h',
+    'pfstream.h',
+    'procbuf.h',
+    'pthread_alloc',
+    'pthread_alloc.h',
+    'rope',
+    'rope.h',
+    'ropeimpl.h',
+    'set.h',
+    'slist',
+    'slist.h',
+    'stack.h',
+    'stdiostream.h',
+    'stl_alloc.h',
+    'stl_relops.h',
+    'streambuf.h',
+    'stream.h',
+    'strfile.h',
+    'strstream.h',
+    'tempbuf.h',
+    'tree.h',
+    'type_traits.h',
+    'vector.h',
+    # 17.6.1.2 C++ library headers
+    'algorithm',
+    'array',
+    'atomic',
+    'bitset',
+    'chrono',
+    'codecvt',
+    'complex',
+    'condition_variable',
+    'deque',
+    'exception',
+    'forward_list',
+    'fstream',
+    'functional',
+    'future',
+    'initializer_list',
+    'iomanip',
+    'ios',
+    'iosfwd',
+    'iostream',
+    'istream',
+    'iterator',
+    'limits',
+    'list',
+    'locale',
+    'map',
+    'memory',
+    'mutex',
+    'new',
+    'numeric',
+    'ostream',
+    'queue',
+    'random',
+    'ratio',
+    'regex',
+    'set',
+    'sstream',
+    'stack',
+    'stdexcept',
+    'streambuf',
+    'string',
+    'strstream',
+    'system_error',
+    'thread',
+    'tuple',
+    'typeindex',
+    'typeinfo',
+    'type_traits',
+    'unordered_map',
+    'unordered_set',
+    'utility',
+    'valarray',
+    'vector',
+    # 17.6.1.2 C++ headers for C library facilities
+    'cassert',
+    'ccomplex',
+    'cctype',
+    'cerrno',
+    'cfenv',
+    'cfloat',
+    'cinttypes',
+    'ciso646',
+    'climits',
+    'clocale',
+    'cmath',
+    'csetjmp',
+    'csignal',
+    'cstdalign',
+    'cstdarg',
+    'cstdbool',
+    'cstddef',
+    'cstdint',
+    'cstdio',
+    'cstdlib',
+    'cstring',
+    'ctgmath',
+    'ctime',
+    'cuchar',
+    'cwchar',
+    'cwctype',
+    ])
+
+# Assertion macros.  These are defined in base/logging.h and
+# testing/base/gunit.h.  Note that the _M versions need to come first
+# for substring matching to work.
+_CHECK_MACROS = [
+    'DCHECK', 'CHECK',
+    'EXPECT_TRUE_M', 'EXPECT_TRUE',
+    'ASSERT_TRUE_M', 'ASSERT_TRUE',
+    'EXPECT_FALSE_M', 'EXPECT_FALSE',
+    'ASSERT_FALSE_M', 'ASSERT_FALSE',
+    ]
+
+# Replacement macros for CHECK/DCHECK/EXPECT_TRUE/EXPECT_FALSE
+_CHECK_REPLACEMENT = dict([(m, {}) for m in _CHECK_MACROS])
+
+for op, replacement in [('==', 'EQ'), ('!=', 'NE'),
+                        ('>=', 'GE'), ('>', 'GT'),
+                        ('<=', 'LE'), ('<', 'LT')]:
+  _CHECK_REPLACEMENT['DCHECK'][op] = 'DCHECK_%s' % replacement
+  _CHECK_REPLACEMENT['CHECK'][op] = 'CHECK_%s' % replacement
+  _CHECK_REPLACEMENT['EXPECT_TRUE'][op] = 'EXPECT_%s' % replacement
+  _CHECK_REPLACEMENT['ASSERT_TRUE'][op] = 'ASSERT_%s' % replacement
+  _CHECK_REPLACEMENT['EXPECT_TRUE_M'][op] = 'EXPECT_%s_M' % replacement
+  _CHECK_REPLACEMENT['ASSERT_TRUE_M'][op] = 'ASSERT_%s_M' % replacement
+
+for op, inv_replacement in [('==', 'NE'), ('!=', 'EQ'),
+                            ('>=', 'LT'), ('>', 'LE'),
+                            ('<=', 'GT'), ('<', 'GE')]:
+  _CHECK_REPLACEMENT['EXPECT_FALSE'][op] = 'EXPECT_%s' % inv_replacement
+  _CHECK_REPLACEMENT['ASSERT_FALSE'][op] = 'ASSERT_%s' % inv_replacement
+  _CHECK_REPLACEMENT['EXPECT_FALSE_M'][op] = 'EXPECT_%s_M' % inv_replacement
+  _CHECK_REPLACEMENT['ASSERT_FALSE_M'][op] = 'ASSERT_%s_M' % inv_replacement
+
+# Alternative tokens and their replacements.  For full list, see section 2.5
+# Alternative tokens [lex.digraph] in the C++ standard.
+#
+# Digraphs (such as '%:') are not included here since it's a mess to
+# match those on a word boundary.
+_ALT_TOKEN_REPLACEMENT = {
+    'and': '&&',
+    'bitor': '|',
+    'or': '||',
+    'xor': '^',
+    'compl': '~',
+    'bitand': '&',
+    'and_eq': '&=',
+    'or_eq': '|=',
+    'xor_eq': '^=',
+    'not': '!',
+    'not_eq': '!='
+    }
+
+# Compile regular expression that matches all the above keywords.  The "[ =()]"
+# bit is meant to avoid matching these keywords outside of boolean expressions.
+#
+# False positives include C-style multi-line comments and multi-line strings
+# but those have always been troublesome for cpplint.
+_ALT_TOKEN_REPLACEMENT_PATTERN = re.compile(
+    r'[ =()](' + ('|'.join(_ALT_TOKEN_REPLACEMENT.keys())) + r')(?=[ (]|$)')
+
+
+# These constants define types of headers for use with
+# _IncludeState.CheckNextIncludeOrder().
+_C_SYS_HEADER = 1
+_CPP_SYS_HEADER = 2
+_LIKELY_MY_HEADER = 3
+_POSSIBLE_MY_HEADER = 4
+_OTHER_HEADER = 5
+
+# These constants define the current inline assembly state
+_NO_ASM = 0       # Outside of inline assembly block
+_INSIDE_ASM = 1   # Inside inline assembly block
+_END_ASM = 2      # Last line of inline assembly block
+_BLOCK_ASM = 3    # The whole block is an inline assembly block
+
+# Match start of assembly blocks
+_MATCH_ASM = re.compile(r'^\s*(?:asm|_asm|__asm|__asm__)'
+                        r'(?:\s+(volatile|__volatile__))?'
+                        r'\s*[{(]')
+
+
+_regexp_compile_cache = {}
+
+# Finds occurrences of NOLINT or NOLINT(...).
+_RE_SUPPRESSION = re.compile(r'\bNOLINT\b(\([^)]*\))?')
+
+# {str, set(int)}: a map from error categories to sets of linenumbers
+# on which those errors are expected and should be suppressed.
+_error_suppressions = {}
+
+# The root directory used for deriving header guard CPP variable.
+# This is set by --root flag.
+_root = None
+
+# The allowed line length of files.
+# This is set by --linelength flag.
+_line_length = 80
+
+# The allowed extensions for file names
+# This is set by --extensions flag.
+_valid_extensions = set(['cc', 'h', 'cpp', 'cu', 'cuh'])
+
+def ParseNolintSuppressions(filename, raw_line, linenum, error):
+  """Updates the global list of error-suppressions.
+
+  Parses any NOLINT comments on the current line, updating the global
+  error_suppressions store.  Reports an error if the NOLINT comment
+  was malformed.
+
+  Args:
+    filename: str, the name of the input file.
+    raw_line: str, the line of input text, with comments.
+    linenum: int, the number of the current line.
+    error: function, an error handler.
+  """
+  # FIXME(adonovan): "NOLINT(" is misparsed as NOLINT(*).
+  matched = _RE_SUPPRESSION.search(raw_line)
+  if matched:
+    category = matched.group(1)
+    if category in (None, '(*)'):  # => "suppress all"
+      _error_suppressions.setdefault(None, set()).add(linenum)
+    else:
+      if category.startswith('(') and category.endswith(')'):
+        category = category[1:-1]
+        if category in _ERROR_CATEGORIES:
+          _error_suppressions.setdefault(category, set()).add(linenum)
+        else:
+          error(filename, linenum, 'readability/nolint', 5,
+                'Unknown NOLINT error category: %s' % category)
+
+
+def ResetNolintSuppressions():
+  "Resets the set of NOLINT suppressions to empty."
+  _error_suppressions.clear()
+
+
+def IsErrorSuppressedByNolint(category, linenum):
+  """Returns true if the specified error category is suppressed on this line.
+
+  Consults the global error_suppressions map populated by
+  ParseNolintSuppressions/ResetNolintSuppressions.
+
+  Args:
+    category: str, the category of the error.
+    linenum: int, the current line number.
+  Returns:
+    bool, True iff the error should be suppressed due to a NOLINT comment.
+  """
+  return (linenum in _error_suppressions.get(category, set()) or
+          linenum in _error_suppressions.get(None, set()))
+
+def Match(pattern, s):
+  """Matches the string with the pattern, caching the compiled regexp."""
+  # The regexp compilation caching is inlined in both Match and Search for
+  # performance reasons; factoring it out into a separate function turns out
+  # to be noticeably expensive.
+  if pattern not in _regexp_compile_cache:
+    _regexp_compile_cache[pattern] = sre_compile.compile(pattern)
+  return _regexp_compile_cache[pattern].match(s)
+
+
+def ReplaceAll(pattern, rep, s):
+  """Replaces instances of pattern in a string with a replacement.
+
+  The compiled regex is kept in a cache shared by Match and Search.
+
+  Args:
+    pattern: regex pattern
+    rep: replacement text
+    s: search string
+
+  Returns:
+    string with replacements made (or original string if no replacements)
+  """
+  if pattern not in _regexp_compile_cache:
+    _regexp_compile_cache[pattern] = sre_compile.compile(pattern)
+  return _regexp_compile_cache[pattern].sub(rep, s)
+
+
+def Search(pattern, s):
+  """Searches the string for the pattern, caching the compiled regexp."""
+  if pattern not in _regexp_compile_cache:
+    _regexp_compile_cache[pattern] = sre_compile.compile(pattern)
+  return _regexp_compile_cache[pattern].search(s)
+
+
+class _IncludeState(dict):
+  """Tracks line numbers for includes, and the order in which includes appear.
+
+  As a dict, an _IncludeState object serves as a mapping between include
+  filename and line number on which that file was included.
+
+  Call CheckNextIncludeOrder() once for each header in the file, passing
+  in the type constants defined above. Calls in an illegal order will
+  raise an _IncludeError with an appropriate error message.
+
+  """
+  # self._section will move monotonically through this set. If it ever
+  # needs to move backwards, CheckNextIncludeOrder will raise an error.
+  _INITIAL_SECTION = 0
+  _MY_H_SECTION = 1
+  _C_SECTION = 2
+  _CPP_SECTION = 3
+  _OTHER_H_SECTION = 4
+
+  _TYPE_NAMES = {
+      _C_SYS_HEADER: 'C system header',
+      _CPP_SYS_HEADER: 'C++ system header',
+      _LIKELY_MY_HEADER: 'header this file implements',
+      _POSSIBLE_MY_HEADER: 'header this file may implement',
+      _OTHER_HEADER: 'other header',
+      }
+  _SECTION_NAMES = {
+      _INITIAL_SECTION: "... nothing. (This can't be an error.)",
+      _MY_H_SECTION: 'a header this file implements',
+      _C_SECTION: 'C system header',
+      _CPP_SECTION: 'C++ system header',
+      _OTHER_H_SECTION: 'other header',
+      }
+
+  def __init__(self):
+    dict.__init__(self)
+    self.ResetSection()
+
+  def ResetSection(self):
+    # The name of the current section.
+    self._section = self._INITIAL_SECTION
+    # The path of last found header.
+    self._last_header = ''
+
+  def SetLastHeader(self, header_path):
+    self._last_header = header_path
+
+  def CanonicalizeAlphabeticalOrder(self, header_path):
+    """Returns a path canonicalized for alphabetical comparison.
+
+    - replaces "-" with "_" so they both cmp the same.
+    - removes '-inl' since we don't require them to be after the main header.
+    - lowercase everything, just in case.
+
+    Args:
+      header_path: Path to be canonicalized.
+
+    Returns:
+      Canonicalized path.
+    """
+    return header_path.replace('-inl.h', '.h').replace('-', '_').lower()
+
+  def IsInAlphabeticalOrder(self, clean_lines, linenum, header_path):
+    """Check if a header is in alphabetical order with the previous header.
+
+    Args:
+      clean_lines: A CleansedLines instance containing the file.
+      linenum: The number of the line to check.
+      header_path: Canonicalized header to be checked.
+
+    Returns:
+      Returns true if the header is in alphabetical order.
+    """
+    # If previous section is different from current section, _last_header will
+    # be reset to empty string, so it's always less than current header.
+    #
+    # If previous line was a blank line, assume that the headers are
+    # intentionally sorted the way they are.
+    if (self._last_header > header_path and
+        not Match(r'^\s*$', clean_lines.elided[linenum - 1])):
+      return False
+    return True
+
+  def CheckNextIncludeOrder(self, header_type):
+    """Returns a non-empty error message if the next header is out of order.
+
+    This function also updates the internal state to be ready to check
+    the next include.
+
+    Args:
+      header_type: One of the _XXX_HEADER constants defined above.
+
+    Returns:
+      The empty string if the header is in the right order, or an
+      error message describing what's wrong.
+
+    """
+    error_message = ('Found %s after %s' %
+                     (self._TYPE_NAMES[header_type],
+                      self._SECTION_NAMES[self._section]))
+
+    last_section = self._section
+
+    if header_type == _C_SYS_HEADER:
+      if self._section <= self._C_SECTION:
+        self._section = self._C_SECTION
+      else:
+        self._last_header = ''
+        return error_message
+    elif header_type == _CPP_SYS_HEADER:
+      if self._section <= self._CPP_SECTION:
+        self._section = self._CPP_SECTION
+      else:
+        self._last_header = ''
+        return error_message
+    elif header_type == _LIKELY_MY_HEADER:
+      if self._section <= self._MY_H_SECTION:
+        self._section = self._MY_H_SECTION
+      else:
+        self._section = self._OTHER_H_SECTION
+    elif header_type == _POSSIBLE_MY_HEADER:
+      if self._section <= self._MY_H_SECTION:
+        self._section = self._MY_H_SECTION
+      else:
+        # This will always be the fallback because we're not sure
+        # enough that the header is associated with this file.
+        self._section = self._OTHER_H_SECTION
+    else:
+      assert header_type == _OTHER_HEADER
+      self._section = self._OTHER_H_SECTION
+
+    if last_section != self._section:
+      self._last_header = ''
+
+    return ''
+
+
+class _CppLintState(object):
+  """Maintains module-wide state.."""
+
+  def __init__(self):
+    self.verbose_level = 1  # global setting.
+    self.error_count = 0    # global count of reported errors
+    # filters to apply when emitting error messages
+    self.filters = _DEFAULT_FILTERS[:]
+    self.counting = 'total'  # In what way are we counting errors?
+    self.errors_by_category = {}  # string to int dict storing error counts
+
+    # output format:
+    # "emacs" - format that emacs can parse (default)
+    # "vs7" - format that Microsoft Visual Studio 7 can parse
+    self.output_format = 'emacs'
+
+  def SetOutputFormat(self, output_format):
+    """Sets the output format for errors."""
+    self.output_format = output_format
+
+  def SetVerboseLevel(self, level):
+    """Sets the module's verbosity, and returns the previous setting."""
+    last_verbose_level = self.verbose_level
+    self.verbose_level = level
+    return last_verbose_level
+
+  def SetCountingStyle(self, counting_style):
+    """Sets the module's counting options."""
+    self.counting = counting_style
+
+  def SetFilters(self, filters):
+    """Sets the error-message filters.
+
+    These filters are applied when deciding whether to emit a given
+    error message.
+
+    Args:
+      filters: A string of comma-separated filters (eg "+whitespace/indent").
+               Each filter should start with + or -; else we die.
+
+    Raises:
+      ValueError: The comma-separated filters did not all start with '+' or '-'.
+                  E.g. "-,+whitespace,-whitespace/indent,whitespace/badfilter"
+    """
+    # Default filters always have less priority than the flag ones.
+    self.filters = _DEFAULT_FILTERS[:]
+    for filt in filters.split(','):
+      clean_filt = filt.strip()
+      if clean_filt:
+        self.filters.append(clean_filt)
+    for filt in self.filters:
+      if not (filt.startswith('+') or filt.startswith('-')):
+        raise ValueError('Every filter in --filters must start with + or -'
+                         ' (%s does not)' % filt)
+
+  def ResetErrorCounts(self):
+    """Sets the module's error statistic back to zero."""
+    self.error_count = 0
+    self.errors_by_category = {}
+
+  def IncrementErrorCount(self, category):
+    """Bumps the module's error statistic."""
+    self.error_count += 1
+    if self.counting in ('toplevel', 'detailed'):
+      if self.counting != 'detailed':
+        category = category.split('/')[0]
+      if category not in self.errors_by_category:
+        self.errors_by_category[category] = 0
+      self.errors_by_category[category] += 1
+
+  def PrintErrorCounts(self):
+    """Print a summary of errors by category, and the total."""
+    for category, count in self.errors_by_category.iteritems():
+      sys.stderr.write('Category \'%s\' errors found: %d\n' %
+                       (category, count))
+    sys.stderr.write('Total errors found: %d\n' % self.error_count)
+
+_cpplint_state = _CppLintState()
+
+
+def _OutputFormat():
+  """Gets the module's output format."""
+  return _cpplint_state.output_format
+
+
+def _SetOutputFormat(output_format):
+  """Sets the module's output format."""
+  _cpplint_state.SetOutputFormat(output_format)
+
+
+def _VerboseLevel():
+  """Returns the module's verbosity setting."""
+  return _cpplint_state.verbose_level
+
+
+def _SetVerboseLevel(level):
+  """Sets the module's verbosity, and returns the previous setting."""
+  return _cpplint_state.SetVerboseLevel(level)
+
+
+def _SetCountingStyle(level):
+  """Sets the module's counting options."""
+  _cpplint_state.SetCountingStyle(level)
+
+
+def _Filters():
+  """Returns the module's list of output filters, as a list."""
+  return _cpplint_state.filters
+
+
+def _SetFilters(filters):
+  """Sets the module's error-message filters.
+
+  These filters are applied when deciding whether to emit a given
+  error message.
+
+  Args:
+    filters: A string of comma-separated filters (eg "whitespace/indent").
+             Each filter should start with + or -; else we die.
+  """
+  _cpplint_state.SetFilters(filters)
+
+
+class _FunctionState(object):
+  """Tracks current function name and the number of lines in its body."""
+
+  _NORMAL_TRIGGER = 250  # for --v=0, 500 for --v=1, etc.
+  _TEST_TRIGGER = 400    # about 50% more than _NORMAL_TRIGGER.
+
+  def __init__(self):
+    self.in_a_function = False
+    self.lines_in_function = 0
+    self.current_function = ''
+
+  def Begin(self, function_name):
+    """Start analyzing function body.
+
+    Args:
+      function_name: The name of the function being tracked.
+    """
+    self.in_a_function = True
+    self.lines_in_function = 0
+    self.current_function = function_name
+
+  def Count(self):
+    """Count line in current function body."""
+    if self.in_a_function:
+      self.lines_in_function += 1
+
+  def Check(self, error, filename, linenum):
+    """Report if too many lines in function body.
+
+    Args:
+      error: The function to call with any errors found.
+      filename: The name of the current file.
+      linenum: The number of the line to check.
+    """
+    if Match(r'T(EST|est)', self.current_function):
+      base_trigger = self._TEST_TRIGGER
+    else:
+      base_trigger = self._NORMAL_TRIGGER
+    trigger = base_trigger * 2**_VerboseLevel()
+
+    if self.lines_in_function > trigger:
+      error_level = int(math.log(self.lines_in_function / base_trigger, 2))
+      # 50 => 0, 100 => 1, 200 => 2, 400 => 3, 800 => 4, 1600 => 5, ...
+      if error_level > 5:
+        error_level = 5
+      error(filename, linenum, 'readability/fn_size', error_level,
+            'Small and focused functions are preferred:'
+            ' %s has %d non-comment lines'
+            ' (error triggered by exceeding %d lines).'  % (
+                self.current_function, self.lines_in_function, trigger))
+
+  def End(self):
+    """Stop analyzing function body."""
+    self.in_a_function = False
+
+
+class _IncludeError(Exception):
+  """Indicates a problem with the include order in a file."""
+  pass
+
+
+class FileInfo:
+  """Provides utility functions for filenames.
+
+  FileInfo provides easy access to the components of a file's path
+  relative to the project root.
+  """
+
+  def __init__(self, filename):
+    self._filename = filename
+
+  def FullName(self):
+    """Make Windows paths like Unix."""
+    return os.path.abspath(self._filename).replace('\\', '/')
+
+  def RepositoryName(self):
+    """FullName after removing the local path to the repository.
+
+    If we have a real absolute path name here we can try to do something smart:
+    detecting the root of the checkout and truncating /path/to/checkout from
+    the name so that we get header guards that don't include things like
+    "C:\Documents and Settings\..." or "/home/username/..." in them and thus
+    people on different computers who have checked the source out to different
+    locations won't see bogus errors.
+    """
+    fullname = self.FullName()
+
+    if os.path.exists(fullname):
+      project_dir = os.path.dirname(fullname)
+
+      if os.path.exists(os.path.join(project_dir, ".svn")):
+        # If there's a .svn file in the current directory, we recursively look
+        # up the directory tree for the top of the SVN checkout
+        root_dir = project_dir
+        one_up_dir = os.path.dirname(root_dir)
+        while os.path.exists(os.path.join(one_up_dir, ".svn")):
+          root_dir = os.path.dirname(root_dir)
+          one_up_dir = os.path.dirname(one_up_dir)
+
+        prefix = os.path.commonprefix([root_dir, project_dir])
+        return fullname[len(prefix) + 1:]
+
+      # Not SVN <= 1.6? Try to find a git, hg, or svn top level directory by
+      # searching up from the current path.
+      root_dir = os.path.dirname(fullname)
+      while (root_dir != os.path.dirname(root_dir) and
+             not os.path.exists(os.path.join(root_dir, ".git")) and
+             not os.path.exists(os.path.join(root_dir, ".hg")) and
+             not os.path.exists(os.path.join(root_dir, ".svn"))):
+        root_dir = os.path.dirname(root_dir)
+
+      if (os.path.exists(os.path.join(root_dir, ".git")) or
+          os.path.exists(os.path.join(root_dir, ".hg")) or
+          os.path.exists(os.path.join(root_dir, ".svn"))):
+        prefix = os.path.commonprefix([root_dir, project_dir])
+        return fullname[len(prefix) + 1:]
+
+    # Don't know what to do; header guard warnings may be wrong...
+    return fullname
+
+  def Split(self):
+    """Splits the file into the directory, basename, and extension.
+
+    For 'chrome/browser/browser.cc', Split() would
+    return ('chrome/browser', 'browser', '.cc')
+
+    Returns:
+      A tuple of (directory, basename, extension).
+    """
+
+    googlename = self.RepositoryName()
+    project, rest = os.path.split(googlename)
+    return (project,) + os.path.splitext(rest)
+
+  def BaseName(self):
+    """File base name - text after the final slash, before the final period."""
+    return self.Split()[1]
+
+  def Extension(self):
+    """File extension - text following the final period."""
+    return self.Split()[2]
+
+  def NoExtension(self):
+    """File has no source file extension."""
+    return '/'.join(self.Split()[0:2])
+
+  def IsSource(self):
+    """File has a source file extension."""
+    return self.Extension()[1:] in ('c', 'cc', 'cpp', 'cxx')
+
+
+def _ShouldPrintError(category, confidence, linenum):
+  """If confidence >= verbose, category passes filter and is not suppressed."""
+
+  # There are three ways we might decide not to print an error message:
+  # a "NOLINT(category)" comment appears in the source,
+  # the verbosity level isn't high enough, or the filters filter it out.
+  if IsErrorSuppressedByNolint(category, linenum):
+    return False
+  if confidence < _cpplint_state.verbose_level:
+    return False
+
+  is_filtered = False
+  for one_filter in _Filters():
+    if one_filter.startswith('-'):
+      if category.startswith(one_filter[1:]):
+        is_filtered = True
+    elif one_filter.startswith('+'):
+      if category.startswith(one_filter[1:]):
+        is_filtered = False
+    else:
+      assert False  # should have been checked for in SetFilter.
+  if is_filtered:
+    return False
+
+  return True
+
+
+def Error(filename, linenum, category, confidence, message):
+  """Logs the fact we've found a lint error.
+
+  We log where the error was found, and also our confidence in the error,
+  that is, how certain we are this is a legitimate style regression, and
+  not a misidentification or a use that's sometimes justified.
+
+  False positives can be suppressed by the use of
+  "cpplint(category)"  comments on the offending line.  These are
+  parsed into _error_suppressions.
+
+  Args:
+    filename: The name of the file containing the error.
+    linenum: The number of the line containing the error.
+    category: A string used to describe the "category" this bug
+      falls under: "whitespace", say, or "runtime".  Categories
+      may have a hierarchy separated by slashes: "whitespace/indent".
+    confidence: A number from 1-5 representing a confidence score for
+      the error, with 5 meaning that we are certain of the problem,
+      and 1 meaning that it could be a legitimate construct.
+    message: The error message.
+  """
+  if _ShouldPrintError(category, confidence, linenum):
+    _cpplint_state.IncrementErrorCount(category)
+    if _cpplint_state.output_format == 'vs7':
+      sys.stderr.write('%s(%s):  %s  [%s] [%d]\n' % (
+          filename, linenum, message, category, confidence))
+    elif _cpplint_state.output_format == 'eclipse':
+      sys.stderr.write('%s:%s: warning: %s  [%s] [%d]\n' % (
+          filename, linenum, message, category, confidence))
+    else:
+      sys.stderr.write('%s:%s:  %s  [%s] [%d]\n' % (
+          filename, linenum, message, category, confidence))
+
+
+# Matches standard C++ escape sequences per 2.13.2.3 of the C++ standard.
+_RE_PATTERN_CLEANSE_LINE_ESCAPES = re.compile(
+    r'\\([abfnrtv?"\\\']|\d+|x[0-9a-fA-F]+)')
+# Matches strings.  Escape codes should already be removed by ESCAPES.
+_RE_PATTERN_CLEANSE_LINE_DOUBLE_QUOTES = re.compile(r'"[^"]*"')
+# Matches characters.  Escape codes should already be removed by ESCAPES.
+_RE_PATTERN_CLEANSE_LINE_SINGLE_QUOTES = re.compile(r"'.'")
+# Matches multi-line C++ comments.
+# This RE is a little bit more complicated than one might expect, because we
+# have to take care of space removals tools so we can handle comments inside
+# statements better.
+# The current rule is: We only clear spaces from both sides when we're at the
+# end of the line. Otherwise, we try to remove spaces from the right side,
+# if this doesn't work we try on left side but only if there's a non-character
+# on the right.
+_RE_PATTERN_CLEANSE_LINE_C_COMMENTS = re.compile(
+    r"""(\s*/\*.*\*/\s*$|
+            /\*.*\*/\s+|
+         \s+/\*.*\*/(?=\W)|
+            /\*.*\*/)""", re.VERBOSE)
+
+
+def IsCppString(line):
+  """Does line terminate so, that the next symbol is in string constant.
+
+  This function does not consider single-line nor multi-line comments.
+
+  Args:
+    line: is a partial line of code starting from the 0..n.
+
+  Returns:
+    True, if next character appended to 'line' is inside a
+    string constant.
+  """
+
+  line = line.replace(r'\\', 'XX')  # after this, \\" does not match to \"
+  return ((line.count('"') - line.count(r'\"') - line.count("'\"'")) & 1) == 1
+
+
+def CleanseRawStrings(raw_lines):
+  """Removes C++11 raw strings from lines.
+
+    Before:
+      static const char kData[] = R"(
+          multi-line string
+          )";
+
+    After:
+      static const char kData[] = ""
+          (replaced by blank line)
+          "";
+
+  Args:
+    raw_lines: list of raw lines.
+
+  Returns:
+    list of lines with C++11 raw strings replaced by empty strings.
+  """
+
+  delimiter = None
+  lines_without_raw_strings = []
+  for line in raw_lines:
+    if delimiter:
+      # Inside a raw string, look for the end
+      end = line.find(delimiter)
+      if end >= 0:
+        # Found the end of the string, match leading space for this
+        # line and resume copying the original lines, and also insert
+        # a "" on the last line.
+        leading_space = Match(r'^(\s*)\S', line)
+        line = leading_space.group(1) + '""' + line[end + len(delimiter):]
+        delimiter = None
+      else:
+        # Haven't found the end yet, append a blank line.
+        line = ''
+
+    else:
+      # Look for beginning of a raw string.
+      # See 2.14.15 [lex.string] for syntax.
+      matched = Match(r'^(.*)\b(?:R|u8R|uR|UR|LR)"([^\s\\()]*)\((.*)$', line)
+      if matched:
+        delimiter = ')' + matched.group(2) + '"'
+
+        end = matched.group(3).find(delimiter)
+        if end >= 0:
+          # Raw string ended on same line
+          line = (matched.group(1) + '""' +
+                  matched.group(3)[end + len(delimiter):])
+          delimiter = None
+        else:
+          # Start of a multi-line raw string
+          line = matched.group(1) + '""'
+
+    lines_without_raw_strings.append(line)
+
+  # TODO(unknown): if delimiter is not None here, we might want to
+  # emit a warning for unterminated string.
+  return lines_without_raw_strings
+
+
+def FindNextMultiLineCommentStart(lines, lineix):
+  """Find the beginning marker for a multiline comment."""
+  while lineix < len(lines):
+    if lines[lineix].strip().startswith('/*'):
+      # Only return this marker if the comment goes beyond this line
+      if lines[lineix].strip().find('*/', 2) < 0:
+        return lineix
+    lineix += 1
+  return len(lines)
+
+
+def FindNextMultiLineCommentEnd(lines, lineix):
+  """We are inside a comment, find the end marker."""
+  while lineix < len(lines):
+    if lines[lineix].strip().endswith('*/'):
+      return lineix
+    lineix += 1
+  return len(lines)
+
+
+def RemoveMultiLineCommentsFromRange(lines, begin, end):
+  """Clears a range of lines for multi-line comments."""
+  # Having // dummy comments makes the lines non-empty, so we will not get
+  # unnecessary blank line warnings later in the code.
+  for i in range(begin, end):
+    lines[i] = '// dummy'
+
+
+def RemoveMultiLineComments(filename, lines, error):
+  """Removes multiline (c-style) comments from lines."""
+  lineix = 0
+  while lineix < len(lines):
+    lineix_begin = FindNextMultiLineCommentStart(lines, lineix)
+    if lineix_begin >= len(lines):
+      return
+    lineix_end = FindNextMultiLineCommentEnd(lines, lineix_begin)
+    if lineix_end >= len(lines):
+      error(filename, lineix_begin + 1, 'readability/multiline_comment', 5,
+            'Could not find end of multi-line comment')
+      return
+    RemoveMultiLineCommentsFromRange(lines, lineix_begin, lineix_end + 1)
+    lineix = lineix_end + 1
+
+
+def CleanseComments(line):
+  """Removes //-comments and single-line C-style /* */ comments.
+
+  Args:
+    line: A line of C++ source.
+
+  Returns:
+    The line with single-line comments removed.
+  """
+  commentpos = line.find('//')
+  if commentpos != -1 and not IsCppString(line[:commentpos]):
+    line = line[:commentpos].rstrip()
+  # get rid of /* ... */
+  return _RE_PATTERN_CLEANSE_LINE_C_COMMENTS.sub('', line)
+
+
+class CleansedLines(object):
+  """Holds 3 copies of all lines with different preprocessing applied to them.
+
+  1) elided member contains lines without strings and comments,
+  2) lines member contains lines without comments, and
+  3) raw_lines member contains all the lines without processing.
+  All these three members are of <type 'list'>, and of the same length.
+  """
+
+  def __init__(self, lines):
+    self.elided = []
+    self.lines = []
+    self.raw_lines = lines
+    self.num_lines = len(lines)
+    self.lines_without_raw_strings = CleanseRawStrings(lines)
+    for linenum in range(len(self.lines_without_raw_strings)):
+      self.lines.append(CleanseComments(
+          self.lines_without_raw_strings[linenum]))
+      elided = self._CollapseStrings(self.lines_without_raw_strings[linenum])
+      self.elided.append(CleanseComments(elided))
+
+  def NumLines(self):
+    """Returns the number of lines represented."""
+    return self.num_lines
+
+  @staticmethod
+  def _CollapseStrings(elided):
+    """Collapses strings and chars on a line to simple "" or '' blocks.
+
+    We nix strings first so we're not fooled by text like '"http://"'
+
+    Args:
+      elided: The line being processed.
+
+    Returns:
+      The line with collapsed strings.
+    """
+    if not _RE_PATTERN_INCLUDE.match(elided):
+      # Remove escaped characters first to make quote/single quote collapsing
+      # basic.  Things that look like escaped characters shouldn't occur
+      # outside of strings and chars.
+      elided = _RE_PATTERN_CLEANSE_LINE_ESCAPES.sub('', elided)
+      elided = _RE_PATTERN_CLEANSE_LINE_SINGLE_QUOTES.sub("''", elided)
+      elided = _RE_PATTERN_CLEANSE_LINE_DOUBLE_QUOTES.sub('""', elided)
+    return elided
+
+
+def FindEndOfExpressionInLine(line, startpos, depth, startchar, endchar):
+  """Find the position just after the matching endchar.
+
+  Args:
+    line: a CleansedLines line.
+    startpos: start searching at this position.
+    depth: nesting level at startpos.
+    startchar: expression opening character.
+    endchar: expression closing character.
+
+  Returns:
+    On finding matching endchar: (index just after matching endchar, 0)
+    Otherwise: (-1, new depth at end of this line)
+  """
+  for i in xrange(startpos, len(line)):
+    if line[i] == startchar:
+      depth += 1
+    elif line[i] == endchar:
+      depth -= 1
+      if depth == 0:
+        return (i + 1, 0)
+  return (-1, depth)
+
+
+def CloseExpression(clean_lines, linenum, pos):
+  """If input points to ( or { or [ or <, finds the position that closes it.
+
+  If lines[linenum][pos] points to a '(' or '{' or '[' or '<', finds the
+  linenum/pos that correspond to the closing of the expression.
+
+  Args:
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    pos: A position on the line.
+
+  Returns:
+    A tuple (line, linenum, pos) pointer *past* the closing brace, or
+    (line, len(lines), -1) if we never find a close.  Note we ignore
+    strings and comments when matching; and the line we return is the
+    'cleansed' line at linenum.
+  """
+
+  line = clean_lines.elided[linenum]
+  startchar = line[pos]
+  if startchar not in '({[<':
+    return (line, clean_lines.NumLines(), -1)
+  if startchar == '(': endchar = ')'
+  if startchar == '[': endchar = ']'
+  if startchar == '{': endchar = '}'
+  if startchar == '<': endchar = '>'
+
+  # Check first line
+  (end_pos, num_open) = FindEndOfExpressionInLine(
+      line, pos, 0, startchar, endchar)
+  if end_pos > -1:
+    return (line, linenum, end_pos)
+
+  # Continue scanning forward
+  while linenum < clean_lines.NumLines() - 1:
+    linenum += 1
+    line = clean_lines.elided[linenum]
+    (end_pos, num_open) = FindEndOfExpressionInLine(
+        line, 0, num_open, startchar, endchar)
+    if end_pos > -1:
+      return (line, linenum, end_pos)
+
+  # Did not find endchar before end of file, give up
+  return (line, clean_lines.NumLines(), -1)
+
+
+def FindStartOfExpressionInLine(line, endpos, depth, startchar, endchar):
+  """Find position at the matching startchar.
+
+  This is almost the reverse of FindEndOfExpressionInLine, but note
+  that the input position and returned position differs by 1.
+
+  Args:
+    line: a CleansedLines line.
+    endpos: start searching at this position.
+    depth: nesting level at endpos.
+    startchar: expression opening character.
+    endchar: expression closing character.
+
+  Returns:
+    On finding matching startchar: (index at matching startchar, 0)
+    Otherwise: (-1, new depth at beginning of this line)
+  """
+  for i in xrange(endpos, -1, -1):
+    if line[i] == endchar:
+      depth += 1
+    elif line[i] == startchar:
+      depth -= 1
+      if depth == 0:
+        return (i, 0)
+  return (-1, depth)
+
+
+def ReverseCloseExpression(clean_lines, linenum, pos):
+  """If input points to ) or } or ] or >, finds the position that opens it.
+
+  If lines[linenum][pos] points to a ')' or '}' or ']' or '>', finds the
+  linenum/pos that correspond to the opening of the expression.
+
+  Args:
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    pos: A position on the line.
+
+  Returns:
+    A tuple (line, linenum, pos) pointer *at* the opening brace, or
+    (line, 0, -1) if we never find the matching opening brace.  Note
+    we ignore strings and comments when matching; and the line we
+    return is the 'cleansed' line at linenum.
+  """
+  line = clean_lines.elided[linenum]
+  endchar = line[pos]
+  if endchar not in ')}]>':
+    return (line, 0, -1)
+  if endchar == ')': startchar = '('
+  if endchar == ']': startchar = '['
+  if endchar == '}': startchar = '{'
+  if endchar == '>': startchar = '<'
+
+  # Check last line
+  (start_pos, num_open) = FindStartOfExpressionInLine(
+      line, pos, 0, startchar, endchar)
+  if start_pos > -1:
+    return (line, linenum, start_pos)
+
+  # Continue scanning backward
+  while linenum > 0:
+    linenum -= 1
+    line = clean_lines.elided[linenum]
+    (start_pos, num_open) = FindStartOfExpressionInLine(
+        line, len(line) - 1, num_open, startchar, endchar)
+    if start_pos > -1:
+      return (line, linenum, start_pos)
+
+  # Did not find startchar before beginning of file, give up
+  return (line, 0, -1)
+
+
+def CheckForCopyright(filename, lines, error):
+  """Logs an error if no Copyright message appears at the top of the file."""
+
+  # We'll say it should occur by line 10. Don't forget there's a
+  # dummy line at the front.
+  for line in xrange(1, min(len(lines), 11)):
+    if re.search(r'Copyright', lines[line], re.I): break
+  else:                       # means no copyright line was found
+    error(filename, 0, 'legal/copyright', 5,
+          'No copyright message found.  '
+          'You should have a line: "Copyright [year] <Copyright Owner>"')
+
+
+def GetHeaderGuardCPPVariable(filename):
+  """Returns the CPP variable that should be used as a header guard.
+
+  Args:
+    filename: The name of a C++ header file.
+
+  Returns:
+    The CPP variable that should be used as a header guard in the
+    named file.
+
+  """
+
+  # Restores original filename in case that cpplint is invoked from Emacs's
+  # flymake.
+  filename = re.sub(r'_flymake\.h$', '.h', filename)
+  filename = re.sub(r'/\.flymake/([^/]*)$', r'/\1', filename)
+
+  fileinfo = FileInfo(filename)
+  file_path_from_root = fileinfo.RepositoryName()
+  if _root:
+    file_path_from_root = re.sub('^' + _root + os.sep, '', file_path_from_root)
+  return re.sub(r'[-./\s]', '_', file_path_from_root).upper() + '_'
+
+
+def CheckForHeaderGuard(filename, lines, error):
+  """Checks that the file contains a header guard.
+
+  Logs an error if no #ifndef header guard is present.  For other
+  headers, checks that the full pathname is used.
+
+  Args:
+    filename: The name of the C++ header file.
+    lines: An array of strings, each representing a line of the file.
+    error: The function to call with any errors found.
+  """
+
+  cppvar = GetHeaderGuardCPPVariable(filename)
+
+  ifndef = None
+  ifndef_linenum = 0
+  define = None
+  endif = None
+  endif_linenum = 0
+  for linenum, line in enumerate(lines):
+    linesplit = line.split()
+    if len(linesplit) >= 2:
+      # find the first occurrence of #ifndef and #define, save arg
+      if not ifndef and linesplit[0] == '#ifndef':
+        # set ifndef to the header guard presented on the #ifndef line.
+        ifndef = linesplit[1]
+        ifndef_linenum = linenum
+      if not define and linesplit[0] == '#define':
+        define = linesplit[1]
+    # find the last occurrence of #endif, save entire line
+    if line.startswith('#endif'):
+      endif = line
+      endif_linenum = linenum
+
+  if not ifndef:
+    error(filename, 0, 'build/header_guard', 5,
+          'No #ifndef header guard found, suggested CPP variable is: %s' %
+          cppvar)
+    return
+
+  if not define:
+    error(filename, 0, 'build/header_guard', 5,
+          'No #define header guard found, suggested CPP variable is: %s' %
+          cppvar)
+    return
+
+  # The guard should be PATH_FILE_H_, but we also allow PATH_FILE_H__
+  # for backward compatibility.
+  if ifndef != cppvar:
+    error_level = 0
+    if ifndef != cppvar + '_':
+      error_level = 5
+
+    ParseNolintSuppressions(filename, lines[ifndef_linenum], ifndef_linenum,
+                            error)
+    error(filename, ifndef_linenum, 'build/header_guard', error_level,
+          '#ifndef header guard has wrong style, please use: %s' % cppvar)
+
+  if define != ifndef:
+    error(filename, 0, 'build/header_guard', 5,
+          '#ifndef and #define don\'t match, suggested CPP variable is: %s' %
+          cppvar)
+    return
+
+  if endif != ('#endif  // %s' % cppvar):
+    error_level = 0
+    if endif != ('#endif  // %s' % (cppvar + '_')):
+      error_level = 5
+
+    ParseNolintSuppressions(filename, lines[endif_linenum], endif_linenum,
+                            error)
+    error(filename, endif_linenum, 'build/header_guard', error_level,
+          '#endif line should be "#endif  // %s"' % cppvar)
+
+
+def CheckForBadCharacters(filename, lines, error):
+  """Logs an error for each line containing bad characters.
+
+  Two kinds of bad characters:
+
+  1. Unicode replacement characters: These indicate that either the file
+  contained invalid UTF-8 (likely) or Unicode replacement characters (which
+  it shouldn't).  Note that it's possible for this to throw off line
+  numbering if the invalid UTF-8 occurred adjacent to a newline.
+
+  2. NUL bytes.  These are problematic for some tools.
+
+  Args:
+    filename: The name of the current file.
+    lines: An array of strings, each representing a line of the file.
+    error: The function to call with any errors found.
+  """
+  for linenum, line in enumerate(lines):
+    if u'\ufffd' in line:
+      error(filename, linenum, 'readability/utf8', 5,
+            'Line contains invalid UTF-8 (or Unicode replacement character).')
+    if '\0' in line:
+      error(filename, linenum, 'readability/nul', 5, 'Line contains NUL byte.')
+
+
+def CheckForNewlineAtEOF(filename, lines, error):
+  """Logs an error if there is no newline char at the end of the file.
+
+  Args:
+    filename: The name of the current file.
+    lines: An array of strings, each representing a line of the file.
+    error: The function to call with any errors found.
+  """
+
+  # The array lines() was created by adding two newlines to the
+  # original file (go figure), then splitting on \n.
+  # To verify that the file ends in \n, we just have to make sure the
+  # last-but-two element of lines() exists and is empty.
+  if len(lines) < 3 or lines[-2]:
+    error(filename, len(lines) - 2, 'whitespace/ending_newline', 5,
+          'Could not find a newline character at the end of the file.')
+
+
+def CheckForMultilineCommentsAndStrings(filename, clean_lines, linenum, error):
+  """Logs an error if we see /* ... */ or "..." that extend past one line.
+
+  /* ... */ comments are legit inside macros, for one line.
+  Otherwise, we prefer // comments, so it's ok to warn about the
+  other.  Likewise, it's ok for strings to extend across multiple
+  lines, as long as a line continuation character (backslash)
+  terminates each line. Although not currently prohibited by the C++
+  style guide, it's ugly and unnecessary. We don't do well with either
+  in this lint program, so we warn about both.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+
+  # Remove all \\ (escaped backslashes) from the line. They are OK, and the
+  # second (escaped) slash may trigger later \" detection erroneously.
+  line = line.replace('\\\\', '')
+
+  if line.count('/*') > line.count('*/'):
+    error(filename, linenum, 'readability/multiline_comment', 5,
+          'Complex multi-line /*...*/-style comment found. '
+          'Lint may give bogus warnings.  '
+          'Consider replacing these with //-style comments, '
+          'with #if 0...#endif, '
+          'or with more clearly structured multi-line comments.')
+
+  if (line.count('"') - line.count('\\"')) % 2:
+    error(filename, linenum, 'readability/multiline_string', 5,
+          'Multi-line string ("...") found.  This lint script doesn\'t '
+          'do well with such strings, and may give bogus warnings.  '
+          'Use C++11 raw strings or concatenation instead.')
+
+
+threading_list = (
+    ('asctime(', 'asctime_r('),
+    ('ctime(', 'ctime_r('),
+    ('getgrgid(', 'getgrgid_r('),
+    ('getgrnam(', 'getgrnam_r('),
+    ('getlogin(', 'getlogin_r('),
+    ('getpwnam(', 'getpwnam_r('),
+    ('getpwuid(', 'getpwuid_r('),
+    ('gmtime(', 'gmtime_r('),
+    ('localtime(', 'localtime_r('),
+    ('rand(', 'rand_r('),
+    ('strtok(', 'strtok_r('),
+    ('ttyname(', 'ttyname_r('),
+    )
+
+
+def CheckPosixThreading(filename, clean_lines, linenum, error):
+  """Checks for calls to thread-unsafe functions.
+
+  Much code has been originally written without consideration of
+  multi-threading. Also, engineers are relying on their old experience;
+  they have learned posix before threading extensions were added. These
+  tests guide the engineers to use thread-safe functions (when using
+  posix directly).
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+  for single_thread_function, multithread_safe_function in threading_list:
+    ix = line.find(single_thread_function)
+    # Comparisons made explicit for clarity -- pylint: disable=g-explicit-bool-comparison
+    if ix >= 0 and (ix == 0 or (not line[ix - 1].isalnum() and
+                                line[ix - 1] not in ('_', '.', '>'))):
+      error(filename, linenum, 'runtime/threadsafe_fn', 2,
+            'Consider using ' + multithread_safe_function +
+            '...) instead of ' + single_thread_function +
+            '...) for improved thread safety.')
+
+
+def CheckVlogArguments(filename, clean_lines, linenum, error):
+  """Checks that VLOG() is only used for defining a logging level.
+
+  For example, VLOG(2) is correct. VLOG(INFO), VLOG(WARNING), VLOG(ERROR), and
+  VLOG(FATAL) are not.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+  if Search(r'\bVLOG\((INFO|ERROR|WARNING|DFATAL|FATAL)\)', line):
+    error(filename, linenum, 'runtime/vlog', 5,
+          'VLOG() should be used with numeric verbosity level.  '
+          'Use LOG() if you want symbolic severity levels.')
+
+
+# Matches invalid increment: *count++, which moves pointer instead of
+# incrementing a value.
+_RE_PATTERN_INVALID_INCREMENT = re.compile(
+    r'^\s*\*\w+(\+\+|--);')
+
+
+def CheckInvalidIncrement(filename, clean_lines, linenum, error):
+  """Checks for invalid increment *count++.
+
+  For example following function:
+  void increment_counter(int* count) {
+    *count++;
+  }
+  is invalid, because it effectively does count++, moving pointer, and should
+  be replaced with ++*count, (*count)++ or *count += 1.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+  if _RE_PATTERN_INVALID_INCREMENT.match(line):
+    error(filename, linenum, 'runtime/invalid_increment', 5,
+          'Changing pointer instead of value (or unused value of operator*).')
+
+
+class _BlockInfo(object):
+  """Stores information about a generic block of code."""
+
+  def __init__(self, seen_open_brace):
+    self.seen_open_brace = seen_open_brace
+    self.open_parentheses = 0
+    self.inline_asm = _NO_ASM
+
+  def CheckBegin(self, filename, clean_lines, linenum, error):
+    """Run checks that applies to text up to the opening brace.
+
+    This is mostly for checking the text after the class identifier
+    and the "{", usually where the base class is specified.  For other
+    blocks, there isn't much to check, so we always pass.
+
+    Args:
+      filename: The name of the current file.
+      clean_lines: A CleansedLines instance containing the file.
+      linenum: The number of the line to check.
+      error: The function to call with any errors found.
+    """
+    pass
+
+  def CheckEnd(self, filename, clean_lines, linenum, error):
+    """Run checks that applies to text after the closing brace.
+
+    This is mostly used for checking end of namespace comments.
+
+    Args:
+      filename: The name of the current file.
+      clean_lines: A CleansedLines instance containing the file.
+      linenum: The number of the line to check.
+      error: The function to call with any errors found.
+    """
+    pass
+
+
+class _ClassInfo(_BlockInfo):
+  """Stores information about a class."""
+
+  def __init__(self, name, class_or_struct, clean_lines, linenum):
+    _BlockInfo.__init__(self, False)
+    self.name = name
+    self.starting_linenum = linenum
+    self.is_derived = False
+    if class_or_struct == 'struct':
+      self.access = 'public'
+      self.is_struct = True
+    else:
+      self.access = 'private'
+      self.is_struct = False
+
+    # Remember initial indentation level for this class.  Using raw_lines here
+    # instead of elided to account for leading comments.
+    initial_indent = Match(r'^( *)\S', clean_lines.raw_lines[linenum])
+    if initial_indent:
+      self.class_indent = len(initial_indent.group(1))
+    else:
+      self.class_indent = 0
+
+    # Try to find the end of the class.  This will be confused by things like:
+    #   class A {
+    #   } *x = { ...
+    #
+    # But it's still good enough for CheckSectionSpacing.
+    self.last_line = 0
+    depth = 0
+    for i in range(linenum, clean_lines.NumLines()):
+      line = clean_lines.elided[i]
+      depth += line.count('{') - line.count('}')
+      if not depth:
+        self.last_line = i
+        break
+
+  def CheckBegin(self, filename, clean_lines, linenum, error):
+    # Look for a bare ':'
+    if Search('(^|[^:]):($|[^:])', clean_lines.elided[linenum]):
+      self.is_derived = True
+
+  def CheckEnd(self, filename, clean_lines, linenum, error):
+    # Check that closing brace is aligned with beginning of the class.
+    # Only do this if the closing brace is indented by only whitespaces.
+    # This means we will not check single-line class definitions.
+    indent = Match(r'^( *)\}', clean_lines.elided[linenum])
+    if indent and len(indent.group(1)) != self.class_indent:
+      if self.is_struct:
+        parent = 'struct ' + self.name
+      else:
+        parent = 'class ' + self.name
+      error(filename, linenum, 'whitespace/indent', 3,
+            'Closing brace should be aligned with beginning of %s' % parent)
+
+
+class _NamespaceInfo(_BlockInfo):
+  """Stores information about a namespace."""
+
+  def __init__(self, name, linenum):
+    _BlockInfo.__init__(self, False)
+    self.name = name or ''
+    self.starting_linenum = linenum
+
+  def CheckEnd(self, filename, clean_lines, linenum, error):
+    """Check end of namespace comments."""
+    line = clean_lines.raw_lines[linenum]
+
+    # Check how many lines is enclosed in this namespace.  Don't issue
+    # warning for missing namespace comments if there aren't enough
+    # lines.  However, do apply checks if there is already an end of
+    # namespace comment and it's incorrect.
+    #
+    # TODO(unknown): We always want to check end of namespace comments
+    # if a namespace is large, but sometimes we also want to apply the
+    # check if a short namespace contained nontrivial things (something
+    # other than forward declarations).  There is currently no logic on
+    # deciding what these nontrivial things are, so this check is
+    # triggered by namespace size only, which works most of the time.
+    if (linenum - self.starting_linenum < 10
+        and not Match(r'};*\s*(//|/\*).*\bnamespace\b', line)):
+      return
+
+    # Look for matching comment at end of namespace.
+    #
+    # Note that we accept C style "/* */" comments for terminating
+    # namespaces, so that code that terminate namespaces inside
+    # preprocessor macros can be cpplint clean.
+    #
+    # We also accept stuff like "// end of namespace <name>." with the
+    # period at the end.
+    #
+    # Besides these, we don't accept anything else, otherwise we might
+    # get false negatives when existing comment is a substring of the
+    # expected namespace.
+    if self.name:
+      # Named namespace
+      if not Match((r'};*\s*(//|/\*).*\bnamespace\s+' + re.escape(self.name) +
+                    r'[\*/\.\\\s]*$'),
+                   line):
+        error(filename, linenum, 'readability/namespace', 5,
+              'Namespace should be terminated with "// namespace %s"' %
+              self.name)
+    else:
+      # Anonymous namespace
+      if not Match(r'};*\s*(//|/\*).*\bnamespace[\*/\.\\\s]*$', line):
+        error(filename, linenum, 'readability/namespace', 5,
+              'Namespace should be terminated with "// namespace"')
+
+
+class _PreprocessorInfo(object):
+  """Stores checkpoints of nesting stacks when #if/#else is seen."""
+
+  def __init__(self, stack_before_if):
+    # The entire nesting stack before #if
+    self.stack_before_if = stack_before_if
+
+    # The entire nesting stack up to #else
+    self.stack_before_else = []
+
+    # Whether we have already seen #else or #elif
+    self.seen_else = False
+
+
+class _NestingState(object):
+  """Holds states related to parsing braces."""
+
+  def __init__(self):
+    # Stack for tracking all braces.  An object is pushed whenever we
+    # see a "{", and popped when we see a "}".  Only 3 types of
+    # objects are possible:
+    # - _ClassInfo: a class or struct.
+    # - _NamespaceInfo: a namespace.
+    # - _BlockInfo: some other type of block.
+    self.stack = []
+
+    # Stack of _PreprocessorInfo objects.
+    self.pp_stack = []
+
+  def SeenOpenBrace(self):
+    """Check if we have seen the opening brace for the innermost block.
+
+    Returns:
+      True if we have seen the opening brace, False if the innermost
+      block is still expecting an opening brace.
+    """
+    return (not self.stack) or self.stack[-1].seen_open_brace
+
+  def InNamespaceBody(self):
+    """Check if we are currently one level inside a namespace body.
+
+    Returns:
+      True if top of the stack is a namespace block, False otherwise.
+    """
+    return self.stack and isinstance(self.stack[-1], _NamespaceInfo)
+
+  def UpdatePreprocessor(self, line):
+    """Update preprocessor stack.
+
+    We need to handle preprocessors due to classes like this:
+      #ifdef SWIG
+      struct ResultDetailsPageElementExtensionPoint {
+      #else
+      struct ResultDetailsPageElementExtensionPoint : public Extension {
+      #endif
+
+    We make the following assumptions (good enough for most files):
+    - Preprocessor condition evaluates to true from #if up to first
+      #else/#elif/#endif.
+
+    - Preprocessor condition evaluates to false from #else/#elif up
+      to #endif.  We still perform lint checks on these lines, but
+      these do not affect nesting stack.
+
+    Args:
+      line: current line to check.
+    """
+    if Match(r'^\s*#\s*(if|ifdef|ifndef)\b', line):
+      # Beginning of #if block, save the nesting stack here.  The saved
+      # stack will allow us to restore the parsing state in the #else case.
+      self.pp_stack.append(_PreprocessorInfo(copy.deepcopy(self.stack)))
+    elif Match(r'^\s*#\s*(else|elif)\b', line):
+      # Beginning of #else block
+      if self.pp_stack:
+        if not self.pp_stack[-1].seen_else:
+          # This is the first #else or #elif block.  Remember the
+          # whole nesting stack up to this point.  This is what we
+          # keep after the #endif.
+          self.pp_stack[-1].seen_else = True
+          self.pp_stack[-1].stack_before_else = copy.deepcopy(self.stack)
+
+        # Restore the stack to how it was before the #if
+        self.stack = copy.deepcopy(self.pp_stack[-1].stack_before_if)
+      else:
+        # TODO(unknown): unexpected #else, issue warning?
+        pass
+    elif Match(r'^\s*#\s*endif\b', line):
+      # End of #if or #else blocks.
+      if self.pp_stack:
+        # If we saw an #else, we will need to restore the nesting
+        # stack to its former state before the #else, otherwise we
+        # will just continue from where we left off.
+        if self.pp_stack[-1].seen_else:
+          # Here we can just use a shallow copy since we are the last
+          # reference to it.
+          self.stack = self.pp_stack[-1].stack_before_else
+        # Drop the corresponding #if
+        self.pp_stack.pop()
+      else:
+        # TODO(unknown): unexpected #endif, issue warning?
+        pass
+
+  def Update(self, filename, clean_lines, linenum, error):
+    """Update nesting state with current line.
+
+    Args:
+      filename: The name of the current file.
+      clean_lines: A CleansedLines instance containing the file.
+      linenum: The number of the line to check.
+      error: The function to call with any errors found.
+    """
+    line = clean_lines.elided[linenum]
+
+    # Update pp_stack first
+    self.UpdatePreprocessor(line)
+
+    # Count parentheses.  This is to avoid adding struct arguments to
+    # the nesting stack.
+    if self.stack:
+      inner_block = self.stack[-1]
+      depth_change = line.count('(') - line.count(')')
+      inner_block.open_parentheses += depth_change
+
+      # Also check if we are starting or ending an inline assembly block.
+      if inner_block.inline_asm in (_NO_ASM, _END_ASM):
+        if (depth_change != 0 and
+            inner_block.open_parentheses == 1 and
+            _MATCH_ASM.match(line)):
+          # Enter assembly block
+          inner_block.inline_asm = _INSIDE_ASM
+        else:
+          # Not entering assembly block.  If previous line was _END_ASM,
+          # we will now shift to _NO_ASM state.
+          inner_block.inline_asm = _NO_ASM
+      elif (inner_block.inline_asm == _INSIDE_ASM and
+            inner_block.open_parentheses == 0):
+        # Exit assembly block
+        inner_block.inline_asm = _END_ASM
+
+    # Consume namespace declaration at the beginning of the line.  Do
+    # this in a loop so that we catch same line declarations like this:
+    #   namespace proto2 { namespace bridge { class MessageSet; } }
+    while True:
+      # Match start of namespace.  The "\b\s*" below catches namespace
+      # declarations even if it weren't followed by a whitespace, this
+      # is so that we don't confuse our namespace checker.  The
+      # missing spaces will be flagged by CheckSpacing.
+      namespace_decl_match = Match(r'^\s*namespace\b\s*([:\w]+)?(.*)$', line)
+      if not namespace_decl_match:
+        break
+
+      new_namespace = _NamespaceInfo(namespace_decl_match.group(1), linenum)
+      self.stack.append(new_namespace)
+
+      line = namespace_decl_match.group(2)
+      if line.find('{') != -1:
+        new_namespace.seen_open_brace = True
+        line = line[line.find('{') + 1:]
+
+    # Look for a class declaration in whatever is left of the line
+    # after parsing namespaces.  The regexp accounts for decorated classes
+    # such as in:
+    #   class LOCKABLE API Object {
+    #   };
+    #
+    # Templates with class arguments may confuse the parser, for example:
+    #   template <class T
+    #             class Comparator = less<T>,
+    #             class Vector = vector<T> >
+    #   class HeapQueue {
+    #
+    # Because this parser has no nesting state about templates, by the
+    # time it saw "class Comparator", it may think that it's a new class.
+    # Nested templates have a similar problem:
+    #   template <
+    #       typename ExportedType,
+    #       typename TupleType,
+    #       template <typename, typename> class ImplTemplate>
+    #
+    # To avoid these cases, we ignore classes that are followed by '=' or '>'
+    class_decl_match = Match(
+        r'\s*(template\s*<[\w\s<>,:]*>\s*)?'
+        r'(class|struct)\s+([A-Z_]+\s+)*(\w+(?:::\w+)*)'
+        r'(([^=>]|<[^<>]*>|<[^<>]*<[^<>]*>\s*>)*)$', line)
+    if (class_decl_match and
+        (not self.stack or self.stack[-1].open_parentheses == 0)):
+      self.stack.append(_ClassInfo(
+          class_decl_match.group(4), class_decl_match.group(2),
+          clean_lines, linenum))
+      line = class_decl_match.group(5)
+
+    # If we have not yet seen the opening brace for the innermost block,
+    # run checks here.
+    if not self.SeenOpenBrace():
+      self.stack[-1].CheckBegin(filename, clean_lines, linenum, error)
+
+    # Update access control if we are inside a class/struct
+    if self.stack and isinstance(self.stack[-1], _ClassInfo):
+      classinfo = self.stack[-1]
+      access_match = Match(
+          r'^(.*)\b(public|private|protected|signals)(\s+(?:slots\s*)?)?'
+          r':(?:[^:]|$)',
+          line)
+      if access_match:
+        classinfo.access = access_match.group(2)
+
+        # Check that access keywords are indented +1 space.  Skip this
+        # check if the keywords are not preceded by whitespaces.
+        indent = access_match.group(1)
+        if (len(indent) != classinfo.class_indent + 1 and
+            Match(r'^\s*$', indent)):
+          if classinfo.is_struct:
+            parent = 'struct ' + classinfo.name
+          else:
+            parent = 'class ' + classinfo.name
+          slots = ''
+          if access_match.group(3):
+            slots = access_match.group(3)
+          error(filename, linenum, 'whitespace/indent', 3,
+                '%s%s: should be indented +1 space inside %s' % (
+                    access_match.group(2), slots, parent))
+
+    # Consume braces or semicolons from what's left of the line
+    while True:
+      # Match first brace, semicolon, or closed parenthesis.
+      matched = Match(r'^[^{;)}]*([{;)}])(.*)$', line)
+      if not matched:
+        break
+
+      token = matched.group(1)
+      if token == '{':
+        # If namespace or class hasn't seen a opening brace yet, mark
+        # namespace/class head as complete.  Push a new block onto the
+        # stack otherwise.
+        if not self.SeenOpenBrace():
+          self.stack[-1].seen_open_brace = True
+        else:
+          self.stack.append(_BlockInfo(True))
+          if _MATCH_ASM.match(line):
+            self.stack[-1].inline_asm = _BLOCK_ASM
+      elif token == ';' or token == ')':
+        # If we haven't seen an opening brace yet, but we already saw
+        # a semicolon, this is probably a forward declaration.  Pop
+        # the stack for these.
+        #
+        # Similarly, if we haven't seen an opening brace yet, but we
+        # already saw a closing parenthesis, then these are probably
+        # function arguments with extra "class" or "struct" keywords.
+        # Also pop these stack for these.
+        if not self.SeenOpenBrace():
+          self.stack.pop()
+      else:  # token == '}'
+        # Perform end of block checks and pop the stack.
+        if self.stack:
+          self.stack[-1].CheckEnd(filename, clean_lines, linenum, error)
+          self.stack.pop()
+      line = matched.group(2)
+
+  def InnermostClass(self):
+    """Get class info on the top of the stack.
+
+    Returns:
+      A _ClassInfo object if we are inside a class, or None otherwise.
+    """
+    for i in range(len(self.stack), 0, -1):
+      classinfo = self.stack[i - 1]
+      if isinstance(classinfo, _ClassInfo):
+        return classinfo
+    return None
+
+  def CheckCompletedBlocks(self, filename, error):
+    """Checks that all classes and namespaces have been completely parsed.
+
+    Call this when all lines in a file have been processed.
+    Args:
+      filename: The name of the current file.
+      error: The function to call with any errors found.
+    """
+    # Note: This test can result in false positives if #ifdef constructs
+    # get in the way of brace matching. See the testBuildClass test in
+    # cpplint_unittest.py for an example of this.
+    for obj in self.stack:
+      if isinstance(obj, _ClassInfo):
+        error(filename, obj.starting_linenum, 'build/class', 5,
+              'Failed to find complete declaration of class %s' %
+              obj.name)
+      elif isinstance(obj, _NamespaceInfo):
+        error(filename, obj.starting_linenum, 'build/namespaces', 5,
+              'Failed to find complete declaration of namespace %s' %
+              obj.name)
+
+
+def CheckForNonStandardConstructs(filename, clean_lines, linenum,
+                                  nesting_state, error):
+  r"""Logs an error if we see certain non-ANSI constructs ignored by gcc-2.
+
+  Complain about several constructs which gcc-2 accepts, but which are
+  not standard C++.  Warning about these in lint is one way to ease the
+  transition to new compilers.
+  - put storage class first (e.g. "static const" instead of "const static").
+  - "%lld" instead of %qd" in printf-type functions.
+  - "%1$d" is non-standard in printf-type functions.
+  - "\%" is an undefined character escape sequence.
+  - text after #endif is not allowed.
+  - invalid inner-style forward declaration.
+  - >? and <? operators, and their >?= and <?= cousins.
+
+  Additionally, check for constructor/destructor style violations and reference
+  members, as it is very convenient to do so while checking for
+  gcc-2 compliance.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: A callable to which errors are reported, which takes 4 arguments:
+           filename, line number, error level, and message
+  """
+
+  # Remove comments from the line, but leave in strings for now.
+  line = clean_lines.lines[linenum]
+
+  if Search(r'printf\s*\(.*".*%[-+ ]?\d*q', line):
+    error(filename, linenum, 'runtime/printf_format', 3,
+          '%q in format strings is deprecated.  Use %ll instead.')
+
+  if Search(r'printf\s*\(.*".*%\d+\$', line):
+    error(filename, linenum, 'runtime/printf_format', 2,
+          '%N$ formats are unconventional.  Try rewriting to avoid them.')
+
+  # Remove escaped backslashes before looking for undefined escapes.
+  line = line.replace('\\\\', '')
+
+  if Search(r'("|\').*\\(%|\[|\(|{)', line):
+    error(filename, linenum, 'build/printf_format', 3,
+          '%, [, (, and { are undefined character escapes.  Unescape them.')
+
+  # For the rest, work with both comments and strings removed.
+  line = clean_lines.elided[linenum]
+
+  if Search(r'\b(const|volatile|void|char|short|int|long'
+            r'|float|double|signed|unsigned'
+            r'|schar|u?int8|u?int16|u?int32|u?int64)'
+            r'\s+(register|static|extern|typedef)\b',
+            line):
+    error(filename, linenum, 'build/storage_class', 5,
+          'Storage class (static, extern, typedef, etc) should be first.')
+
+  if Match(r'\s*#\s*endif\s*[^/\s]+', line):
+    error(filename, linenum, 'build/endif_comment', 5,
+          'Uncommented text after #endif is non-standard.  Use a comment.')
+
+  if Match(r'\s*class\s+(\w+\s*::\s*)+\w+\s*;', line):
+    error(filename, linenum, 'build/forward_decl', 5,
+          'Inner-style forward declarations are invalid.  Remove this line.')
+
+  if Search(r'(\w+|[+-]?\d+(\.\d*)?)\s*(<|>)\?=?\s*(\w+|[+-]?\d+)(\.\d*)?',
+            line):
+    error(filename, linenum, 'build/deprecated', 3,
+          '>? and <? (max and min) operators are non-standard and deprecated.')
+
+  if Search(r'^\s*const\s*string\s*&\s*\w+\s*;', line):
+    # TODO(unknown): Could it be expanded safely to arbitrary references,
+    # without triggering too many false positives? The first
+    # attempt triggered 5 warnings for mostly benign code in the regtest, hence
+    # the restriction.
+    # Here's the original regexp, for the reference:
+    # type_name = r'\w+((\s*::\s*\w+)|(\s*<\s*\w+?\s*>))?'
+    # r'\s*const\s*' + type_name + '\s*&\s*\w+\s*;'
+    error(filename, linenum, 'runtime/member_string_references', 2,
+          'const string& members are dangerous. It is much better to use '
+          'alternatives, such as pointers or simple constants.')
+
+  # Everything else in this function operates on class declarations.
+  # Return early if the top of the nesting stack is not a class, or if
+  # the class head is not completed yet.
+  classinfo = nesting_state.InnermostClass()
+  if not classinfo or not classinfo.seen_open_brace:
+    return
+
+  # The class may have been declared with namespace or classname qualifiers.
+  # The constructor and destructor will not have those qualifiers.
+  base_classname = classinfo.name.split('::')[-1]
+
+  # Look for single-argument constructors that aren't marked explicit.
+  # Technically a valid construct, but against style.
+  args = Match(r'\s+(?:inline\s+)?%s\s*\(([^,()]+)\)'
+               % re.escape(base_classname),
+               line)
+  if (args and
+      args.group(1) != 'void' and
+      not Match(r'(const\s+)?%s(\s+const)?\s*(?:<\w+>\s*)?&'
+                % re.escape(base_classname), args.group(1).strip())):
+    error(filename, linenum, 'runtime/explicit', 5,
+          'Single-argument constructors should be marked explicit.')
+
+
+def CheckSpacingForFunctionCall(filename, line, linenum, error):
+  """Checks for the correctness of various spacing around function calls.
+
+  Args:
+    filename: The name of the current file.
+    line: The text of the line to check.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+
+  # Since function calls often occur inside if/for/while/switch
+  # expressions - which have their own, more liberal conventions - we
+  # first see if we should be looking inside such an expression for a
+  # function call, to which we can apply more strict standards.
+  fncall = line    # if there's no control flow construct, look at whole line
+  for pattern in (r'\bif\s*\((.*)\)\s*{',
+                  r'\bfor\s*\((.*)\)\s*{',
+                  r'\bwhile\s*\((.*)\)\s*[{;]',
+                  r'\bswitch\s*\((.*)\)\s*{'):
+    match = Search(pattern, line)
+    if match:
+      fncall = match.group(1)    # look inside the parens for function calls
+      break
+
+  # Except in if/for/while/switch, there should never be space
+  # immediately inside parens (eg "f( 3, 4 )").  We make an exception
+  # for nested parens ( (a+b) + c ).  Likewise, there should never be
+  # a space before a ( when it's a function argument.  I assume it's a
+  # function argument when the char before the whitespace is legal in
+  # a function name (alnum + _) and we're not starting a macro. Also ignore
+  # pointers and references to arrays and functions coz they're too tricky:
+  # we use a very simple way to recognize these:
+  # " (something)(maybe-something)" or
+  # " (something)(maybe-something," or
+  # " (something)[something]"
+  # Note that we assume the contents of [] to be short enough that
+  # they'll never need to wrap.
+  if (  # Ignore control structures.
+      not Search(r'\b(if|for|while|switch|return|new|delete|catch|sizeof)\b',
+                 fncall) and
+      # Ignore pointers/references to functions.
+      not Search(r' \([^)]+\)\([^)]*(\)|,$)', fncall) and
+      # Ignore pointers/references to arrays.
+      not Search(r' \([^)]+\)\[[^\]]+\]', fncall)):
+    if Search(r'\w\s*\(\s(?!\s*\\$)', fncall):      # a ( used for a fn call
+      error(filename, linenum, 'whitespace/parens', 4,
+            'Extra space after ( in function call')
+    elif Search(r'\(\s+(?!(\s*\\)|\()', fncall):
+      error(filename, linenum, 'whitespace/parens', 2,
+            'Extra space after (')
+    if (Search(r'\w\s+\(', fncall) and
+        not Search(r'#\s*define|typedef', fncall) and
+        not Search(r'\w\s+\((\w+::)*\*\w+\)\(', fncall)):
+      error(filename, linenum, 'whitespace/parens', 4,
+            'Extra space before ( in function call')
+    # If the ) is followed only by a newline or a { + newline, assume it's
+    # part of a control statement (if/while/etc), and don't complain
+    if Search(r'[^)]\s+\)\s*[^{\s]', fncall):
+      # If the closing parenthesis is preceded by only whitespaces,
+      # try to give a more descriptive error message.
+      if Search(r'^\s+\)', fncall):
+        error(filename, linenum, 'whitespace/parens', 2,
+              'Closing ) should be moved to the previous line')
+      else:
+        error(filename, linenum, 'whitespace/parens', 2,
+              'Extra space before )')
+
+
+def IsBlankLine(line):
+  """Returns true if the given line is blank.
+
+  We consider a line to be blank if the line is empty or consists of
+  only white spaces.
+
+  Args:
+    line: A line of a string.
+
+  Returns:
+    True, if the given line is blank.
+  """
+  return not line or line.isspace()
+
+
+def CheckForFunctionLengths(filename, clean_lines, linenum,
+                            function_state, error):
+  """Reports for long function bodies.
+
+  For an overview why this is done, see:
+  http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml#Write_Short_Functions
+
+  Uses a simplistic algorithm assuming other style guidelines
+  (especially spacing) are followed.
+  Only checks unindented functions, so class members are unchecked.
+  Trivial bodies are unchecked, so constructors with huge initializer lists
+  may be missed.
+  Blank/comment lines are not counted so as to avoid encouraging the removal
+  of vertical space and comments just to get through a lint check.
+  NOLINT *on the last line of a function* disables this check.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    function_state: Current function name and lines in body so far.
+    error: The function to call with any errors found.
+  """
+  lines = clean_lines.lines
+  line = lines[linenum]
+  raw = clean_lines.raw_lines
+  raw_line = raw[linenum]
+  joined_line = ''
+
+  starting_func = False
+  regexp = r'(\w(\w|::|\*|\&|\s)*)\('  # decls * & space::name( ...
+  match_result = Match(regexp, line)
+  if match_result:
+    # If the name is all caps and underscores, figure it's a macro and
+    # ignore it, unless it's TEST or TEST_F.
+    function_name = match_result.group(1).split()[-1]
+    if function_name == 'TEST' or function_name == 'TEST_F' or (
+        not Match(r'[A-Z_]+$', function_name)):
+      starting_func = True
+
+  if starting_func:
+    body_found = False
+    for start_linenum in xrange(linenum, clean_lines.NumLines()):
+      start_line = lines[start_linenum]
+      joined_line += ' ' + start_line.lstrip()
+      if Search(r'(;|})', start_line):  # Declarations and trivial functions
+        body_found = True
+        break                              # ... ignore
+      elif Search(r'{', start_line):
+        body_found = True
+        function = Search(r'((\w|:)*)\(', line).group(1)
+        if Match(r'TEST', function):    # Handle TEST... macros
+          parameter_regexp = Search(r'(\(.*\))', joined_line)
+          if parameter_regexp:             # Ignore bad syntax
+            function += parameter_regexp.group(1)
+        else:
+          function += '()'
+        function_state.Begin(function)
+        break
+    if not body_found:
+      # No body for the function (or evidence of a non-function) was found.
+      error(filename, linenum, 'readability/fn_size', 5,
+            'Lint failed to find start of function body.')
+  elif Match(r'^\}\s*$', line):  # function end
+    function_state.Check(error, filename, linenum)
+    function_state.End()
+  elif not Match(r'^\s*$', line):
+    function_state.Count()  # Count non-blank/non-comment lines.
+
+
+_RE_PATTERN_TODO = re.compile(r'^//(\s*)TODO(\(.+?\))?:?(\s|$)?')
+
+
+def CheckComment(comment, filename, linenum, error):
+  """Checks for common mistakes in TODO comments.
+
+  Args:
+    comment: The text of the comment from the line in question.
+    filename: The name of the current file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  match = _RE_PATTERN_TODO.match(comment)
+  if match:
+    # One whitespace is correct; zero whitespace is handled elsewhere.
+    leading_whitespace = match.group(1)
+    if len(leading_whitespace) > 1:
+      error(filename, linenum, 'whitespace/todo', 2,
+            'Too many spaces before TODO')
+
+    username = match.group(2)
+    if not username:
+      error(filename, linenum, 'readability/todo', 2,
+            'Missing username in TODO; it should look like '
+            '"// TODO(my_username): Stuff."')
+
+    middle_whitespace = match.group(3)
+    # Comparisons made explicit for correctness -- pylint: disable=g-explicit-bool-comparison
+    if middle_whitespace != ' ' and middle_whitespace != '':
+      error(filename, linenum, 'whitespace/todo', 2,
+            'TODO(my_username) should be followed by a space')
+
+def CheckAccess(filename, clean_lines, linenum, nesting_state, error):
+  """Checks for improper use of DISALLOW* macros.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]  # get rid of comments and strings
+
+  matched = Match((r'\s*(DISALLOW_COPY_AND_ASSIGN|'
+                   r'DISALLOW_EVIL_CONSTRUCTORS|'
+                   r'DISALLOW_IMPLICIT_CONSTRUCTORS)'), line)
+  if not matched:
+    return
+  if nesting_state.stack and isinstance(nesting_state.stack[-1], _ClassInfo):
+    if nesting_state.stack[-1].access != 'private':
+      error(filename, linenum, 'readability/constructors', 3,
+            '%s must be in the private: section' % matched.group(1))
+
+  else:
+    # Found DISALLOW* macro outside a class declaration, or perhaps it
+    # was used inside a function when it should have been part of the
+    # class declaration.  We could issue a warning here, but it
+    # probably resulted in a compiler error already.
+    pass
+
+
+def FindNextMatchingAngleBracket(clean_lines, linenum, init_suffix):
+  """Find the corresponding > to close a template.
+
+  Args:
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: Current line number.
+    init_suffix: Remainder of the current line after the initial <.
+
+  Returns:
+    True if a matching bracket exists.
+  """
+  line = init_suffix
+  nesting_stack = ['<']
+  while True:
+    # Find the next operator that can tell us whether < is used as an
+    # opening bracket or as a less-than operator.  We only want to
+    # warn on the latter case.
+    #
+    # We could also check all other operators and terminate the search
+    # early, e.g. if we got something like this "a<b+c", the "<" is
+    # most likely a less-than operator, but then we will get false
+    # positives for default arguments and other template expressions.
+    match = Search(r'^[^<>(),;\[\]]*([<>(),;\[\]])(.*)$', line)
+    if match:
+      # Found an operator, update nesting stack
+      operator = match.group(1)
+      line = match.group(2)
+
+      if nesting_stack[-1] == '<':
+        # Expecting closing angle bracket
+        if operator in ('<', '(', '['):
+          nesting_stack.append(operator)
+        elif operator == '>':
+          nesting_stack.pop()
+          if not nesting_stack:
+            # Found matching angle bracket
+            return True
+        elif operator == ',':
+          # Got a comma after a bracket, this is most likely a template
+          # argument.  We have not seen a closing angle bracket yet, but
+          # it's probably a few lines later if we look for it, so just
+          # return early here.
+          return True
+        else:
+          # Got some other operator.
+          return False
+
+      else:
+        # Expecting closing parenthesis or closing bracket
+        if operator in ('<', '(', '['):
+          nesting_stack.append(operator)
+        elif operator in (')', ']'):
+          # We don't bother checking for matching () or [].  If we got
+          # something like (] or [), it would have been a syntax error.
+          nesting_stack.pop()
+
+    else:
+      # Scan the next line
+      linenum += 1
+      if linenum >= len(clean_lines.elided):
+        break
+      line = clean_lines.elided[linenum]
+
+  # Exhausted all remaining lines and still no matching angle bracket.
+  # Most likely the input was incomplete, otherwise we should have
+  # seen a semicolon and returned early.
+  return True
+
+
+def FindPreviousMatchingAngleBracket(clean_lines, linenum, init_prefix):
+  """Find the corresponding < that started a template.
+
+  Args:
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: Current line number.
+    init_prefix: Part of the current line before the initial >.
+
+  Returns:
+    True if a matching bracket exists.
+  """
+  line = init_prefix
+  nesting_stack = ['>']
+  while True:
+    # Find the previous operator
+    match = Search(r'^(.*)([<>(),;\[\]])[^<>(),;\[\]]*$', line)
+    if match:
+      # Found an operator, update nesting stack
+      operator = match.group(2)
+      line = match.group(1)
+
+      if nesting_stack[-1] == '>':
+        # Expecting opening angle bracket
+        if operator in ('>', ')', ']'):
+          nesting_stack.append(operator)
+        elif operator == '<':
+          nesting_stack.pop()
+          if not nesting_stack:
+            # Found matching angle bracket
+            return True
+        elif operator == ',':
+          # Got a comma before a bracket, this is most likely a
+          # template argument.  The opening angle bracket is probably
+          # there if we look for it, so just return early here.
+          return True
+        else:
+          # Got some other operator.
+          return False
+
+      else:
+        # Expecting opening parenthesis or opening bracket
+        if operator in ('>', ')', ']'):
+          nesting_stack.append(operator)
+        elif operator in ('(', '['):
+          nesting_stack.pop()
+
+    else:
+      # Scan the previous line
+      linenum -= 1
+      if linenum < 0:
+        break
+      line = clean_lines.elided[linenum]
+
+  # Exhausted all earlier lines and still no matching angle bracket.
+  return False
+
+
+def CheckSpacing(filename, clean_lines, linenum, nesting_state, error):
+  """Checks for the correctness of various spacing issues in the code.
+
+  Things we check for: spaces around operators, spaces after
+  if/for/while/switch, no spaces around parens in function calls, two
+  spaces between code and comment, don't start a block with a blank
+  line, don't end a function with a blank line, don't add a blank line
+  after public/protected/private, don't have too many blank lines in a row.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: The function to call with any errors found.
+  """
+
+  # Don't use "elided" lines here, otherwise we can't check commented lines.
+  # Don't want to use "raw" either, because we don't want to check inside C++11
+  # raw strings,
+  raw = clean_lines.lines_without_raw_strings
+  line = raw[linenum]
+
+  # Before nixing comments, check if the line is blank for no good
+  # reason.  This includes the first line after a block is opened, and
+  # blank lines at the end of a function (ie, right before a line like '}'
+  #
+  # Skip all the blank line checks if we are immediately inside a
+  # namespace body.  In other words, don't issue blank line warnings
+  # for this block:
+  #   namespace {
+  #
+  #   }
+  #
+  # A warning about missing end of namespace comments will be issued instead.
+  if IsBlankLine(line) and not nesting_state.InNamespaceBody():
+    elided = clean_lines.elided
+    prev_line = elided[linenum - 1]
+    prevbrace = prev_line.rfind('{')
+    # TODO(unknown): Don't complain if line before blank line, and line after,
+    #                both start with alnums and are indented the same amount.
+    #                This ignores whitespace at the start of a namespace block
+    #                because those are not usually indented.
+    if prevbrace != -1 and prev_line[prevbrace:].find('}') == -1:
+      # OK, we have a blank line at the start of a code block.  Before we
+      # complain, we check if it is an exception to the rule: The previous
+      # non-empty line has the parameters of a function header that are indented
+      # 4 spaces (because they did not fit in a 80 column line when placed on
+      # the same line as the function name).  We also check for the case where
+      # the previous line is indented 6 spaces, which may happen when the
+      # initializers of a constructor do not fit into a 80 column line.
+      exception = False
+      if Match(r' {6}\w', prev_line):  # Initializer list?
+        # We are looking for the opening column of initializer list, which
+        # should be indented 4 spaces to cause 6 space indentation afterwards.
+        search_position = linenum-2
+        while (search_position >= 0
+               and Match(r' {6}\w', elided[search_position])):
+          search_position -= 1
+        exception = (search_position >= 0
+                     and elided[search_position][:5] == '    :')
+      else:
+        # Search for the function arguments or an initializer list.  We use a
+        # simple heuristic here: If the line is indented 4 spaces; and we have a
+        # closing paren, without the opening paren, followed by an opening brace
+        # or colon (for initializer lists) we assume that it is the last line of
+        # a function header.  If we have a colon indented 4 spaces, it is an
+        # initializer list.
+        exception = (Match(r' {4}\w[^\(]*\)\s*(const\s*)?(\{\s*$|:)',
+                           prev_line)
+                     or Match(r' {4}:', prev_line))
+
+      if not exception:
+        error(filename, linenum, 'whitespace/blank_line', 2,
+              'Redundant blank line at the start of a code block '
+              'should be deleted.')
+    # Ignore blank lines at the end of a block in a long if-else
+    # chain, like this:
+    #   if (condition1) {
+    #     // Something followed by a blank line
+    #
+    #   } else if (condition2) {
+    #     // Something else
+    #   }
+    if linenum + 1 < clean_lines.NumLines():
+      next_line = raw[linenum + 1]
+      if (next_line
+          and Match(r'\s*}', next_line)
+          and next_line.find('} else ') == -1):
+        error(filename, linenum, 'whitespace/blank_line', 3,
+              'Redundant blank line at the end of a code block '
+              'should be deleted.')
+
+    matched = Match(r'\s*(public|protected|private):', prev_line)
+    if matched:
+      error(filename, linenum, 'whitespace/blank_line', 3,
+            'Do not leave a blank line after "%s:"' % matched.group(1))
+
+  # Next, we complain if there's a comment too near the text
+  commentpos = line.find('//')
+  if commentpos != -1:
+    # Check if the // may be in quotes.  If so, ignore it
+    # Comparisons made explicit for clarity -- pylint: disable=g-explicit-bool-comparison
+    if (line.count('"', 0, commentpos) -
+        line.count('\\"', 0, commentpos)) % 2 == 0:   # not in quotes
+      # Allow one space for new scopes, two spaces otherwise:
+      if (not Match(r'^\s*{ //', line) and
+          ((commentpos >= 1 and
+            line[commentpos-1] not in string.whitespace) or
+           (commentpos >= 2 and
+            line[commentpos-2] not in string.whitespace))):
+        error(filename, linenum, 'whitespace/comments', 2,
+              'At least two spaces is best between code and comments')
+      # There should always be a space between the // and the comment
+      commentend = commentpos + 2
+      if commentend < len(line) and not line[commentend] == ' ':
+        # but some lines are exceptions -- e.g. if they're big
+        # comment delimiters like:
+        # //----------------------------------------------------------
+        # or are an empty C++ style Doxygen comment, like:
+        # ///
+        # or C++ style Doxygen comments placed after the variable:
+        # ///<  Header comment
+        # //!<  Header comment
+        # or they begin with multiple slashes followed by a space:
+        # //////// Header comment
+        match = (Search(r'[=/-]{4,}\s*$', line[commentend:]) or
+                 Search(r'^/$', line[commentend:]) or
+                 Search(r'^!< ', line[commentend:]) or
+                 Search(r'^/< ', line[commentend:]) or
+                 Search(r'^/+ ', line[commentend:]))
+        if not match:
+          error(filename, linenum, 'whitespace/comments', 4,
+                'Should have a space between // and comment')
+      CheckComment(line[commentpos:], filename, linenum, error)
+
+  line = clean_lines.elided[linenum]  # get rid of comments and strings
+
+  # Don't try to do spacing checks for operator methods
+  line = re.sub(r'operator(==|!=|<|<<|<=|>=|>>|>)\(', 'operator\(', line)
+
+  # We allow no-spaces around = within an if: "if ( (a=Foo()) == 0 )".
+  # Otherwise not.  Note we only check for non-spaces on *both* sides;
+  # sometimes people put non-spaces on one side when aligning ='s among
+  # many lines (not that this is behavior that I approve of...)
+  if Search(r'[\w.]=[\w.]', line) and not Search(r'\b(if|while) ', line):
+    error(filename, linenum, 'whitespace/operators', 4,
+          'Missing spaces around =')
+
+  # It's ok not to have spaces around binary operators like + - * /, but if
+  # there's too little whitespace, we get concerned.  It's hard to tell,
+  # though, so we punt on this one for now.  TODO.
+
+  # You should always have whitespace around binary operators.
+  #
+  # Check <= and >= first to avoid false positives with < and >, then
+  # check non-include lines for spacing around < and >.
+  match = Search(r'[^<>=!\s](==|!=|<=|>=)[^<>=!\s]', line)
+  if match:
+    error(filename, linenum, 'whitespace/operators', 3,
+          'Missing spaces around %s' % match.group(1))
+  # We allow no-spaces around << when used like this: 10<<20, but
+  # not otherwise (particularly, not when used as streams)
+  # Also ignore using ns::operator<<;
+  match = Search(r'(operator|\S)(?:L|UL|ULL|l|ul|ull)?<<(\S)', line)
+  if (match and
+      not (match.group(1).isdigit() and match.group(2).isdigit()) and
+      not (match.group(1) == 'operator' and match.group(2) == ';')):
+    error(filename, linenum, 'whitespace/operators', 3,
+          'Missing spaces around <<')
+  elif not Match(r'#.*include', line):
+    # Avoid false positives on ->
+    reduced_line = line.replace('->', '')
+
+    # Look for < that is not surrounded by spaces.  This is only
+    # triggered if both sides are missing spaces, even though
+    # technically should should flag if at least one side is missing a
+    # space.  This is done to avoid some false positives with shifts.
+    match = Search(r'[^\s<]<([^\s=<].*)', reduced_line)
+    if (match and
+        not FindNextMatchingAngleBracket(clean_lines, linenum, match.group(1))):
+      error(filename, linenum, 'whitespace/operators', 3,
+            'Missing spaces around <')
+
+    # Look for > that is not surrounded by spaces.  Similar to the
+    # above, we only trigger if both sides are missing spaces to avoid
+    # false positives with shifts.
+    match = Search(r'^(.*[^\s>])>[^\s=>]', reduced_line)
+    if (match and
+        not FindPreviousMatchingAngleBracket(clean_lines, linenum,
+                                             match.group(1))):
+      error(filename, linenum, 'whitespace/operators', 3,
+            'Missing spaces around >')
+
+  # We allow no-spaces around >> for almost anything.  This is because
+  # C++11 allows ">>" to close nested templates, which accounts for
+  # most cases when ">>" is not followed by a space.
+  #
+  # We still warn on ">>" followed by alpha character, because that is
+  # likely due to ">>" being used for right shifts, e.g.:
+  #   value >> alpha
+  #
+  # When ">>" is used to close templates, the alphanumeric letter that
+  # follows would be part of an identifier, and there should still be
+  # a space separating the template type and the identifier.
+  #   type<type<type>> alpha
+  match = Search(r'>>[a-zA-Z_]', line)
+  if match:
+    error(filename, linenum, 'whitespace/operators', 3,
+          'Missing spaces around >>')
+
+  # There shouldn't be space around unary operators
+  match = Search(r'(!\s|~\s|[\s]--[\s;]|[\s]\+\+[\s;])', line)
+  if match:
+    error(filename, linenum, 'whitespace/operators', 4,
+          'Extra space for operator %s' % match.group(1))
+
+  # A pet peeve of mine: no spaces after an if, while, switch, or for
+  match = Search(r' (if\(|for\(|while\(|switch\()', line)
+  if match:
+    error(filename, linenum, 'whitespace/parens', 5,
+          'Missing space before ( in %s' % match.group(1))
+
+  # For if/for/while/switch, the left and right parens should be
+  # consistent about how many spaces are inside the parens, and
+  # there should either be zero or one spaces inside the parens.
+  # We don't want: "if ( foo)" or "if ( foo   )".
+  # Exception: "for ( ; foo; bar)" and "for (foo; bar; )" are allowed.
+  match = Search(r'\b(if|for|while|switch)\s*'
+                 r'\(([ ]*)(.).*[^ ]+([ ]*)\)\s*{\s*$',
+                 line)
+  if match:
+    if len(match.group(2)) != len(match.group(4)):
+      if not (match.group(3) == ';' and
+              len(match.group(2)) == 1 + len(match.group(4)) or
+              not match.group(2) and Search(r'\bfor\s*\(.*; \)', line)):
+        error(filename, linenum, 'whitespace/parens', 5,
+              'Mismatching spaces inside () in %s' % match.group(1))
+    if len(match.group(2)) not in [0, 1]:
+      error(filename, linenum, 'whitespace/parens', 5,
+            'Should have zero or one spaces inside ( and ) in %s' %
+            match.group(1))
+
+  # You should always have a space after a comma (either as fn arg or operator)
+  #
+  # This does not apply when the non-space character following the
+  # comma is another comma, since the only time when that happens is
+  # for empty macro arguments.
+  #
+  # We run this check in two passes: first pass on elided lines to
+  # verify that lines contain missing whitespaces, second pass on raw
+  # lines to confirm that those missing whitespaces are not due to
+  # elided comments.
+  if Search(r',[^,\s]', line) and Search(r',[^,\s]', raw[linenum]):
+    error(filename, linenum, 'whitespace/comma', 3,
+          'Missing space after ,')
+
+  # You should always have a space after a semicolon
+  # except for few corner cases
+  # TODO(unknown): clarify if 'if (1) { return 1;}' is requires one more
+  # space after ;
+  if Search(r';[^\s};\\)/]', line):
+    error(filename, linenum, 'whitespace/semicolon', 3,
+          'Missing space after ;')
+
+  # Next we will look for issues with function calls.
+  CheckSpacingForFunctionCall(filename, line, linenum, error)
+
+  # Except after an opening paren, or after another opening brace (in case of
+  # an initializer list, for instance), you should have spaces before your
+  # braces. And since you should never have braces at the beginning of a line,
+  # this is an easy test.
+  match = Match(r'^(.*[^ ({]){', line)
+  if match:
+    # Try a bit harder to check for brace initialization.  This
+    # happens in one of the following forms:
+    #   Constructor() : initializer_list_{} { ... }
+    #   Constructor{}.MemberFunction()
+    #   Type variable{};
+    #   FunctionCall(type{}, ...);
+    #   LastArgument(..., type{});
+    #   LOG(INFO) << type{} << " ...";
+    #   map_of_type[{...}] = ...;
+    #
+    # We check for the character following the closing brace, and
+    # silence the warning if it's one of those listed above, i.e.
+    # "{.;,)<]".
+    #
+    # To account for nested initializer list, we allow any number of
+    # closing braces up to "{;,)<".  We can't simply silence the
+    # warning on first sight of closing brace, because that would
+    # cause false negatives for things that are not initializer lists.
+    #   Silence this:         But not this:
+    #     Outer{                if (...) {
+    #       Inner{...}            if (...){  // Missing space before {
+    #     };                    }
+    #
+    # There is a false negative with this approach if people inserted
+    # spurious semicolons, e.g. "if (cond){};", but we will catch the
+    # spurious semicolon with a separate check.
+    (endline, endlinenum, endpos) = CloseExpression(
+        clean_lines, linenum, len(match.group(1)))
+    trailing_text = ''
+    if endpos > -1:
+      trailing_text = endline[endpos:]
+    for offset in xrange(endlinenum + 1,
+                         min(endlinenum + 3, clean_lines.NumLines() - 1)):
+      trailing_text += clean_lines.elided[offset]
+    if not Match(r'^[\s}]*[{.;,)<\]]', trailing_text):
+      error(filename, linenum, 'whitespace/braces', 5,
+            'Missing space before {')
+
+  # Make sure '} else {' has spaces.
+  if Search(r'}else', line):
+    error(filename, linenum, 'whitespace/braces', 5,
+          'Missing space before else')
+
+  # You shouldn't have spaces before your brackets, except maybe after
+  # 'delete []' or 'new char * []'.
+  if Search(r'\w\s+\[', line) and not Search(r'delete\s+\[', line):
+    error(filename, linenum, 'whitespace/braces', 5,
+          'Extra space before [')
+
+  # You shouldn't have a space before a semicolon at the end of the line.
+  # There's a special case for "for" since the style guide allows space before
+  # the semicolon there.
+  if Search(r':\s*;\s*$', line):
+    error(filename, linenum, 'whitespace/semicolon', 5,
+          'Semicolon defining empty statement. Use {} instead.')
+  elif Search(r'^\s*;\s*$', line):
+    error(filename, linenum, 'whitespace/semicolon', 5,
+          'Line contains only semicolon. If this should be an empty statement, '
+          'use {} instead.')
+  elif (Search(r'\s+;\s*$', line) and
+        not Search(r'\bfor\b', line)):
+    error(filename, linenum, 'whitespace/semicolon', 5,
+          'Extra space before last semicolon. If this should be an empty '
+          'statement, use {} instead.')
+
+  # In range-based for, we wanted spaces before and after the colon, but
+  # not around "::" tokens that might appear.
+  if (Search('for *\(.*[^:]:[^: ]', line) or
+      Search('for *\(.*[^: ]:[^:]', line)):
+    error(filename, linenum, 'whitespace/forcolon', 2,
+          'Missing space around colon in range-based for loop')
+
+
+def CheckSectionSpacing(filename, clean_lines, class_info, linenum, error):
+  """Checks for additional blank line issues related to sections.
+
+  Currently the only thing checked here is blank line before protected/private.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    class_info: A _ClassInfo objects.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  # Skip checks if the class is small, where small means 25 lines or less.
+  # 25 lines seems like a good cutoff since that's the usual height of
+  # terminals, and any class that can't fit in one screen can't really
+  # be considered "small".
+  #
+  # Also skip checks if we are on the first line.  This accounts for
+  # classes that look like
+  #   class Foo { public: ... };
+  #
+  # If we didn't find the end of the class, last_line would be zero,
+  # and the check will be skipped by the first condition.
+  if (class_info.last_line - class_info.starting_linenum <= 24 or
+      linenum <= class_info.starting_linenum):
+    return
+
+  matched = Match(r'\s*(public|protected|private):', clean_lines.lines[linenum])
+  if matched:
+    # Issue warning if the line before public/protected/private was
+    # not a blank line, but don't do this if the previous line contains
+    # "class" or "struct".  This can happen two ways:
+    #  - We are at the beginning of the class.
+    #  - We are forward-declaring an inner class that is semantically
+    #    private, but needed to be public for implementation reasons.
+    # Also ignores cases where the previous line ends with a backslash as can be
+    # common when defining classes in C macros.
+    prev_line = clean_lines.lines[linenum - 1]
+    if (not IsBlankLine(prev_line) and
+        not Search(r'\b(class|struct)\b', prev_line) and
+        not Search(r'\\$', prev_line)):
+      # Try a bit harder to find the beginning of the class.  This is to
+      # account for multi-line base-specifier lists, e.g.:
+      #   class Derived
+      #       : public Base {
+      end_class_head = class_info.starting_linenum
+      for i in range(class_info.starting_linenum, linenum):
+        if Search(r'\{\s*$', clean_lines.lines[i]):
+          end_class_head = i
+          break
+      if end_class_head < linenum - 1:
+        error(filename, linenum, 'whitespace/blank_line', 3,
+              '"%s:" should be preceded by a blank line' % matched.group(1))
+
+
+def GetPreviousNonBlankLine(clean_lines, linenum):
+  """Return the most recent non-blank line and its line number.
+
+  Args:
+    clean_lines: A CleansedLines instance containing the file contents.
+    linenum: The number of the line to check.
+
+  Returns:
+    A tuple with two elements.  The first element is the contents of the last
+    non-blank line before the current line, or the empty string if this is the
+    first non-blank line.  The second is the line number of that line, or -1
+    if this is the first non-blank line.
+  """
+
+  prevlinenum = linenum - 1
+  while prevlinenum >= 0:
+    prevline = clean_lines.elided[prevlinenum]
+    if not IsBlankLine(prevline):     # if not a blank line...
+      return (prevline, prevlinenum)
+    prevlinenum -= 1
+  return ('', -1)
+
+
+def CheckBraces(filename, clean_lines, linenum, error):
+  """Looks for misplaced braces (e.g. at the end of line).
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+
+  line = clean_lines.elided[linenum]        # get rid of comments and strings
+
+  if Match(r'\s*{\s*$', line):
+    # We allow an open brace to start a line in the case where someone is using
+    # braces in a block to explicitly create a new scope, which is commonly used
+    # to control the lifetime of stack-allocated variables.  Braces are also
+    # used for brace initializers inside function calls.  We don't detect this
+    # perfectly: we just don't complain if the last non-whitespace character on
+    # the previous non-blank line is ',', ';', ':', '(', '{', or '}', or if the
+    # previous line starts a preprocessor block.
+    prevline = GetPreviousNonBlankLine(clean_lines, linenum)[0]
+    if (not Search(r'[,;:}{(]\s*$', prevline) and
+        not Match(r'\s*#', prevline)):
+      error(filename, linenum, 'whitespace/braces', 4,
+            '{ should almost always be at the end of the previous line')
+
+  # An else clause should be on the same line as the preceding closing brace.
+  if Match(r'\s*else\s*', line):
+    prevline = GetPreviousNonBlankLine(clean_lines, linenum)[0]
+    if Match(r'\s*}\s*$', prevline):
+      error(filename, linenum, 'whitespace/newline', 4,
+            'An else should appear on the same line as the preceding }')
+
+  # If braces come on one side of an else, they should be on both.
+  # However, we have to worry about "else if" that spans multiple lines!
+  if Search(r'}\s*else[^{]*$', line) or Match(r'[^}]*else\s*{', line):
+    if Search(r'}\s*else if([^{]*)$', line):       # could be multi-line if
+      # find the ( after the if
+      pos = line.find('else if')
+      pos = line.find('(', pos)
+      if pos > 0:
+        (endline, _, endpos) = CloseExpression(clean_lines, linenum, pos)
+        if endline[endpos:].find('{') == -1:    # must be brace after if
+          error(filename, linenum, 'readability/braces', 5,
+                'If an else has a brace on one side, it should have it on both')
+    else:            # common case: else not followed by a multi-line if
+      error(filename, linenum, 'readability/braces', 5,
+            'If an else has a brace on one side, it should have it on both')
+
+  # Likewise, an else should never have the else clause on the same line
+  if Search(r'\belse [^\s{]', line) and not Search(r'\belse if\b', line):
+    error(filename, linenum, 'whitespace/newline', 4,
+          'Else clause should never be on same line as else (use 2 lines)')
+
+  # In the same way, a do/while should never be on one line
+  if Match(r'\s*do [^\s{]', line):
+    error(filename, linenum, 'whitespace/newline', 4,
+          'do/while clauses should not be on a single line')
+
+  # Block bodies should not be followed by a semicolon.  Due to C++11
+  # brace initialization, there are more places where semicolons are
+  # required than not, so we use a whitelist approach to check these
+  # rather than a blacklist.  These are the places where "};" should
+  # be replaced by just "}":
+  # 1. Some flavor of block following closing parenthesis:
+  #    for (;;) {};
+  #    while (...) {};
+  #    switch (...) {};
+  #    Function(...) {};
+  #    if (...) {};
+  #    if (...) else if (...) {};
+  #
+  # 2. else block:
+  #    if (...) else {};
+  #
+  # 3. const member function:
+  #    Function(...) const {};
+  #
+  # 4. Block following some statement:
+  #    x = 42;
+  #    {};
+  #
+  # 5. Block at the beginning of a function:
+  #    Function(...) {
+  #      {};
+  #    }
+  #
+  #    Note that naively checking for the preceding "{" will also match
+  #    braces inside multi-dimensional arrays, but this is fine since
+  #    that expression will not contain semicolons.
+  #
+  # 6. Block following another block:
+  #    while (true) {}
+  #    {};
+  #
+  # 7. End of namespaces:
+  #    namespace {};
+  #
+  #    These semicolons seems far more common than other kinds of
+  #    redundant semicolons, possibly due to people converting classes
+  #    to namespaces.  For now we do not warn for this case.
+  #
+  # Try matching case 1 first.
+  match = Match(r'^(.*\)\s*)\{', line)
+  if match:
+    # Matched closing parenthesis (case 1).  Check the token before the
+    # matching opening parenthesis, and don't warn if it looks like a
+    # macro.  This avoids these false positives:
+    #  - macro that defines a base class
+    #  - multi-line macro that defines a base class
+    #  - macro that defines the whole class-head
+    #
+    # But we still issue warnings for macros that we know are safe to
+    # warn, specifically:
+    #  - TEST, TEST_F, TEST_P, MATCHER, MATCHER_P
+    #  - TYPED_TEST
+    #  - INTERFACE_DEF
+    #  - EXCLUSIVE_LOCKS_REQUIRED, SHARED_LOCKS_REQUIRED, LOCKS_EXCLUDED:
+    #
+    # We implement a whitelist of safe macros instead of a blacklist of
+    # unsafe macros, even though the latter appears less frequently in
+    # google code and would have been easier to implement.  This is because
+    # the downside for getting the whitelist wrong means some extra
+    # semicolons, while the downside for getting the blacklist wrong
+    # would result in compile errors.
+    #
+    # In addition to macros, we also don't want to warn on compound
+    # literals.
+    closing_brace_pos = match.group(1).rfind(')')
+    opening_parenthesis = ReverseCloseExpression(
+        clean_lines, linenum, closing_brace_pos)
+    if opening_parenthesis[2] > -1:
+      line_prefix = opening_parenthesis[0][0:opening_parenthesis[2]]
+      macro = Search(r'\b([A-Z_]+)\s*$', line_prefix)
+      if ((macro and
+           macro.group(1) not in (
+               'TEST', 'TEST_F', 'MATCHER', 'MATCHER_P', 'TYPED_TEST',
+               'EXCLUSIVE_LOCKS_REQUIRED', 'SHARED_LOCKS_REQUIRED',
+               'LOCKS_EXCLUDED', 'INTERFACE_DEF')) or
+          Search(r'\s+=\s*$', line_prefix)):
+        match = None
+
+  else:
+    # Try matching cases 2-3.
+    match = Match(r'^(.*(?:else|\)\s*const)\s*)\{', line)
+    if not match:
+      # Try matching cases 4-6.  These are always matched on separate lines.
+      #
+      # Note that we can't simply concatenate the previous line to the
+      # current line and do a single match, otherwise we may output
+      # duplicate warnings for the blank line case:
+      #   if (cond) {
+      #     // blank line
+      #   }
+      prevline = GetPreviousNonBlankLine(clean_lines, linenum)[0]
+      if prevline and Search(r'[;{}]\s*$', prevline):
+        match = Match(r'^(\s*)\{', line)
+
+  # Check matching closing brace
+  if match:
+    (endline, endlinenum, endpos) = CloseExpression(
+        clean_lines, linenum, len(match.group(1)))
+    if endpos > -1 and Match(r'^\s*;', endline[endpos:]):
+      # Current {} pair is eligible for semicolon check, and we have found
+      # the redundant semicolon, output warning here.
+      #
+      # Note: because we are scanning forward for opening braces, and
+      # outputting warnings for the matching closing brace, if there are
+      # nested blocks with trailing semicolons, we will get the error
+      # messages in reversed order.
+      error(filename, endlinenum, 'readability/braces', 4,
+            "You don't need a ; after a }")
+
+
+def CheckEmptyBlockBody(filename, clean_lines, linenum, error):
+  """Look for empty loop/conditional body with only a single semicolon.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+
+  # Search for loop keywords at the beginning of the line.  Because only
+  # whitespaces are allowed before the keywords, this will also ignore most
+  # do-while-loops, since those lines should start with closing brace.
+  #
+  # We also check "if" blocks here, since an empty conditional block
+  # is likely an error.
+  line = clean_lines.elided[linenum]
+  matched = Match(r'\s*(for|while|if)\s*\(', line)
+  if matched:
+    # Find the end of the conditional expression
+    (end_line, end_linenum, end_pos) = CloseExpression(
+        clean_lines, linenum, line.find('('))
+
+    # Output warning if what follows the condition expression is a semicolon.
+    # No warning for all other cases, including whitespace or newline, since we
+    # have a separate check for semicolons preceded by whitespace.
+    if end_pos >= 0 and Match(r';', end_line[end_pos:]):
+      if matched.group(1) == 'if':
+        error(filename, end_linenum, 'whitespace/empty_conditional_body', 5,
+              'Empty conditional bodies should use {}')
+      else:
+        error(filename, end_linenum, 'whitespace/empty_loop_body', 5,
+              'Empty loop bodies should use {} or continue')
+
+
+def CheckCheck(filename, clean_lines, linenum, error):
+  """Checks the use of CHECK and EXPECT macros.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+
+  # Decide the set of replacement macros that should be suggested
+  lines = clean_lines.elided
+  check_macro = None
+  start_pos = -1
+  for macro in _CHECK_MACROS:
+    i = lines[linenum].find(macro)
+    if i >= 0:
+      check_macro = macro
+
+      # Find opening parenthesis.  Do a regular expression match here
+      # to make sure that we are matching the expected CHECK macro, as
+      # opposed to some other macro that happens to contain the CHECK
+      # substring.
+      matched = Match(r'^(.*\b' + check_macro + r'\s*)\(', lines[linenum])
+      if not matched:
+        continue
+      start_pos = len(matched.group(1))
+      break
+  if not check_macro or start_pos < 0:
+    # Don't waste time here if line doesn't contain 'CHECK' or 'EXPECT'
+    return
+
+  # Find end of the boolean expression by matching parentheses
+  (last_line, end_line, end_pos) = CloseExpression(
+      clean_lines, linenum, start_pos)
+  if end_pos < 0:
+    return
+  if linenum == end_line:
+    expression = lines[linenum][start_pos + 1:end_pos - 1]
+  else:
+    expression = lines[linenum][start_pos + 1:]
+    for i in xrange(linenum + 1, end_line):
+      expression += lines[i]
+    expression += last_line[0:end_pos - 1]
+
+  # Parse expression so that we can take parentheses into account.
+  # This avoids false positives for inputs like "CHECK((a < 4) == b)",
+  # which is not replaceable by CHECK_LE.
+  lhs = ''
+  rhs = ''
+  operator = None
+  while expression:
+    matched = Match(r'^\s*(<<|<<=|>>|>>=|->\*|->|&&|\|\||'
+                    r'==|!=|>=|>|<=|<|\()(.*)$', expression)
+    if matched:
+      token = matched.group(1)
+      if token == '(':
+        # Parenthesized operand
+        expression = matched.group(2)
+        (end, _) = FindEndOfExpressionInLine(expression, 0, 1, '(', ')')
+        if end < 0:
+          return  # Unmatched parenthesis
+        lhs += '(' + expression[0:end]
+        expression = expression[end:]
+      elif token in ('&&', '||'):
+        # Logical and/or operators.  This means the expression
+        # contains more than one term, for example:
+        #   CHECK(42 < a && a < b);
+        #
+        # These are not replaceable with CHECK_LE, so bail out early.
+        return
+      elif token in ('<<', '<<=', '>>', '>>=', '->*', '->'):
+        # Non-relational operator
+        lhs += token
+        expression = matched.group(2)
+      else:
+        # Relational operator
+        operator = token
+        rhs = matched.group(2)
+        break
+    else:
+      # Unparenthesized operand.  Instead of appending to lhs one character
+      # at a time, we do another regular expression match to consume several
+      # characters at once if possible.  Trivial benchmark shows that this
+      # is more efficient when the operands are longer than a single
+      # character, which is generally the case.
+      matched = Match(r'^([^-=!<>()&|]+)(.*)$', expression)
+      if not matched:
+        matched = Match(r'^(\s*\S)(.*)$', expression)
+        if not matched:
+          break
+      lhs += matched.group(1)
+      expression = matched.group(2)
+
+  # Only apply checks if we got all parts of the boolean expression
+  if not (lhs and operator and rhs):
+    return
+
+  # Check that rhs do not contain logical operators.  We already know
+  # that lhs is fine since the loop above parses out && and ||.
+  if rhs.find('&&') > -1 or rhs.find('||') > -1:
+    return
+
+  # At least one of the operands must be a constant literal.  This is
+  # to avoid suggesting replacements for unprintable things like
+  # CHECK(variable != iterator)
+  #
+  # The following pattern matches decimal, hex integers, strings, and
+  # characters (in that order).
+  lhs = lhs.strip()
+  rhs = rhs.strip()
+  match_constant = r'^([-+]?(\d+|0[xX][0-9a-fA-F]+)[lLuU]{0,3}|".*"|\'.*\')$'
+  if Match(match_constant, lhs) or Match(match_constant, rhs):
+    # Note: since we know both lhs and rhs, we can provide a more
+    # descriptive error message like:
+    #   Consider using CHECK_EQ(x, 42) instead of CHECK(x == 42)
+    # Instead of:
+    #   Consider using CHECK_EQ instead of CHECK(a == b)
+    #
+    # We are still keeping the less descriptive message because if lhs
+    # or rhs gets long, the error message might become unreadable.
+    error(filename, linenum, 'readability/check', 2,
+          'Consider using %s instead of %s(a %s b)' % (
+              _CHECK_REPLACEMENT[check_macro][operator],
+              check_macro, operator))
+
+
+def CheckAltTokens(filename, clean_lines, linenum, error):
+  """Check alternative keywords being used in boolean expressions.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+
+  # Avoid preprocessor lines
+  if Match(r'^\s*#', line):
+    return
+
+  # Last ditch effort to avoid multi-line comments.  This will not help
+  # if the comment started before the current line or ended after the
+  # current line, but it catches most of the false positives.  At least,
+  # it provides a way to workaround this warning for people who use
+  # multi-line comments in preprocessor macros.
+  #
+  # TODO(unknown): remove this once cpplint has better support for
+  # multi-line comments.
+  if line.find('/*') >= 0 or line.find('*/') >= 0:
+    return
+
+  for match in _ALT_TOKEN_REPLACEMENT_PATTERN.finditer(line):
+    error(filename, linenum, 'readability/alt_tokens', 2,
+          'Use operator %s instead of %s' % (
+              _ALT_TOKEN_REPLACEMENT[match.group(1)], match.group(1)))
+
+
+def GetLineWidth(line):
+  """Determines the width of the line in column positions.
+
+  Args:
+    line: A string, which may be a Unicode string.
+
+  Returns:
+    The width of the line in column positions, accounting for Unicode
+    combining characters and wide characters.
+  """
+  if isinstance(line, unicode):
+    width = 0
+    for uc in unicodedata.normalize('NFC', line):
+      if unicodedata.east_asian_width(uc) in ('W', 'F'):
+        width += 2
+      elif not unicodedata.combining(uc):
+        width += 1
+    return width
+  else:
+    return len(line)
+
+
+def CheckStyle(filename, clean_lines, linenum, file_extension, nesting_state,
+               error):
+  """Checks rules from the 'C++ style rules' section of cppguide.html.
+
+  Most of these rules are hard to test (naming, comment style), but we
+  do what we can.  In particular we check for 2-space indents, line lengths,
+  tab usage, spaces inside code, etc.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    file_extension: The extension (without the dot) of the filename.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: The function to call with any errors found.
+  """
+
+  # Don't use "elided" lines here, otherwise we can't check commented lines.
+  # Don't want to use "raw" either, because we don't want to check inside C++11
+  # raw strings,
+  raw_lines = clean_lines.lines_without_raw_strings
+  line = raw_lines[linenum]
+
+  if line.find('\t') != -1:
+    error(filename, linenum, 'whitespace/tab', 1,
+          'Tab found; better to use spaces')
+
+  # One or three blank spaces at the beginning of the line is weird; it's
+  # hard to reconcile that with 2-space indents.
+  # NOTE: here are the conditions rob pike used for his tests.  Mine aren't
+  # as sophisticated, but it may be worth becoming so:  RLENGTH==initial_spaces
+  # if(RLENGTH > 20) complain = 0;
+  # if(match($0, " +(error|private|public|protected):")) complain = 0;
+  # if(match(prev, "&& *$")) complain = 0;
+  # if(match(prev, "\\|\\| *$")) complain = 0;
+  # if(match(prev, "[\",=><] *$")) complain = 0;
+  # if(match($0, " <<")) complain = 0;
+  # if(match(prev, " +for \\(")) complain = 0;
+  # if(prevodd && match(prevprev, " +for \\(")) complain = 0;
+  initial_spaces = 0
+  cleansed_line = clean_lines.elided[linenum]
+  while initial_spaces < len(line) and line[initial_spaces] == ' ':
+    initial_spaces += 1
+  if line and line[-1].isspace():
+    error(filename, linenum, 'whitespace/end_of_line', 4,
+          'Line ends in whitespace.  Consider deleting these extra spaces.')
+  # There are certain situations we allow one space, notably for section labels
+  elif ((initial_spaces == 1 or initial_spaces == 3) and
+        not Match(r'\s*\w+\s*:\s*$', cleansed_line)):
+    error(filename, linenum, 'whitespace/indent', 3,
+          'Weird number of spaces at line-start.  '
+          'Are you using a 2-space indent?')
+
+  # Check if the line is a header guard.
+  is_header_guard = False
+  if file_extension == 'h':
+    cppvar = GetHeaderGuardCPPVariable(filename)
+    if (line.startswith('#ifndef %s' % cppvar) or
+        line.startswith('#define %s' % cppvar) or
+        line.startswith('#endif  // %s' % cppvar)):
+      is_header_guard = True
+  # #include lines and header guards can be long, since there's no clean way to
+  # split them.
+  #
+  # URLs can be long too.  It's possible to split these, but it makes them
+  # harder to cut&paste.
+  #
+  # The "$Id:...$" comment may also get very long without it being the
+  # developers fault.
+  if (not line.startswith('#include') and not is_header_guard and
+      not Match(r'^\s*//.*http(s?)://\S*$', line) and
+      not Match(r'^// \$Id:.*#[0-9]+ \$$', line)):
+    line_width = GetLineWidth(line)
+    extended_length = int((_line_length * 1.25))
+    if line_width > extended_length:
+      error(filename, linenum, 'whitespace/line_length', 4,
+            'Lines should very rarely be longer than %i characters' %
+            extended_length)
+    elif line_width > _line_length:
+      error(filename, linenum, 'whitespace/line_length', 2,
+            'Lines should be <= %i characters long' % _line_length)
+
+  if (cleansed_line.count(';') > 1 and
+      # for loops are allowed two ;'s (and may run over two lines).
+      cleansed_line.find('for') == -1 and
+      (GetPreviousNonBlankLine(clean_lines, linenum)[0].find('for') == -1 or
+       GetPreviousNonBlankLine(clean_lines, linenum)[0].find(';') != -1) and
+      # It's ok to have many commands in a switch case that fits in 1 line
+      not ((cleansed_line.find('case ') != -1 or
+            cleansed_line.find('default:') != -1) and
+           cleansed_line.find('break;') != -1)):
+    error(filename, linenum, 'whitespace/newline', 0,
+          'More than one command on the same line')
+
+  # Some more style checks
+  CheckBraces(filename, clean_lines, linenum, error)
+  CheckEmptyBlockBody(filename, clean_lines, linenum, error)
+  CheckAccess(filename, clean_lines, linenum, nesting_state, error)
+  CheckSpacing(filename, clean_lines, linenum, nesting_state, error)
+  CheckCheck(filename, clean_lines, linenum, error)
+  CheckAltTokens(filename, clean_lines, linenum, error)
+  classinfo = nesting_state.InnermostClass()
+  if classinfo:
+    CheckSectionSpacing(filename, clean_lines, classinfo, linenum, error)
+
+
+_RE_PATTERN_INCLUDE_NEW_STYLE = re.compile(r'#include +"[^/]+\.h"')
+_RE_PATTERN_INCLUDE = re.compile(r'^\s*#\s*include\s*([<"])([^>"]*)[>"].*$')
+# Matches the first component of a filename delimited by -s and _s. That is:
+#  _RE_FIRST_COMPONENT.match('foo').group(0) == 'foo'
+#  _RE_FIRST_COMPONENT.match('foo.cc').group(0) == 'foo'
+#  _RE_FIRST_COMPONENT.match('foo-bar_baz.cc').group(0) == 'foo'
+#  _RE_FIRST_COMPONENT.match('foo_bar-baz.cc').group(0) == 'foo'
+_RE_FIRST_COMPONENT = re.compile(r'^[^-_.]+')
+
+
+def _DropCommonSuffixes(filename):
+  """Drops common suffixes like _test.cc or -inl.h from filename.
+
+  For example:
+    >>> _DropCommonSuffixes('foo/foo-inl.h')
+    'foo/foo'
+    >>> _DropCommonSuffixes('foo/bar/foo.cc')
+    'foo/bar/foo'
+    >>> _DropCommonSuffixes('foo/foo_internal.h')
+    'foo/foo'
+    >>> _DropCommonSuffixes('foo/foo_unusualinternal.h')
+    'foo/foo_unusualinternal'
+
+  Args:
+    filename: The input filename.
+
+  Returns:
+    The filename with the common suffix removed.
+  """
+  for suffix in ('test.cc', 'regtest.cc', 'unittest.cc',
+                 'inl.h', 'impl.h', 'internal.h'):
+    if (filename.endswith(suffix) and len(filename) > len(suffix) and
+        filename[-len(suffix) - 1] in ('-', '_')):
+      return filename[:-len(suffix) - 1]
+  return os.path.splitext(filename)[0]
+
+
+def _IsTestFilename(filename):
+  """Determines if the given filename has a suffix that identifies it as a test.
+
+  Args:
+    filename: The input filename.
+
+  Returns:
+    True if 'filename' looks like a test, False otherwise.
+  """
+  if (filename.endswith('_test.cc') or
+      filename.endswith('_unittest.cc') or
+      filename.endswith('_regtest.cc')):
+    return True
+  else:
+    return False
+
+
+def _ClassifyInclude(fileinfo, include, is_system):
+  """Figures out what kind of header 'include' is.
+
+  Args:
+    fileinfo: The current file cpplint is running over. A FileInfo instance.
+    include: The path to a #included file.
+    is_system: True if the #include used <> rather than "".
+
+  Returns:
+    One of the _XXX_HEADER constants.
+
+  For example:
+    >>> _ClassifyInclude(FileInfo('foo/foo.cc'), 'stdio.h', True)
+    _C_SYS_HEADER
+    >>> _ClassifyInclude(FileInfo('foo/foo.cc'), 'string', True)
+    _CPP_SYS_HEADER
+    >>> _ClassifyInclude(FileInfo('foo/foo.cc'), 'foo/foo.h', False)
+    _LIKELY_MY_HEADER
+    >>> _ClassifyInclude(FileInfo('foo/foo_unknown_extension.cc'),
+    ...                  'bar/foo_other_ext.h', False)
+    _POSSIBLE_MY_HEADER
+    >>> _ClassifyInclude(FileInfo('foo/foo.cc'), 'foo/bar.h', False)
+    _OTHER_HEADER
+  """
+  # This is a list of all standard c++ header files, except
+  # those already checked for above.
+  is_cpp_h = include in _CPP_HEADERS
+
+  if is_system:
+    if is_cpp_h:
+      return _CPP_SYS_HEADER
+    else:
+      return _C_SYS_HEADER
+
+  # If the target file and the include we're checking share a
+  # basename when we drop common extensions, and the include
+  # lives in . , then it's likely to be owned by the target file.
+  target_dir, target_base = (
+      os.path.split(_DropCommonSuffixes(fileinfo.RepositoryName())))
+  include_dir, include_base = os.path.split(_DropCommonSuffixes(include))
+  if target_base == include_base and (
+      include_dir == target_dir or
+      include_dir == os.path.normpath(target_dir + '/../public')):
+    return _LIKELY_MY_HEADER
+
+  # If the target and include share some initial basename
+  # component, it's possible the target is implementing the
+  # include, so it's allowed to be first, but we'll never
+  # complain if it's not there.
+  target_first_component = _RE_FIRST_COMPONENT.match(target_base)
+  include_first_component = _RE_FIRST_COMPONENT.match(include_base)
+  if (target_first_component and include_first_component and
+      target_first_component.group(0) ==
+      include_first_component.group(0)):
+    return _POSSIBLE_MY_HEADER
+
+  return _OTHER_HEADER
+
+
+
+def CheckIncludeLine(filename, clean_lines, linenum, include_state, error):
+  """Check rules that are applicable to #include lines.
+
+  Strings on #include lines are NOT removed from elided line, to make
+  certain tasks easier. However, to prevent false positives, checks
+  applicable to #include lines in CheckLanguage must be put here.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    include_state: An _IncludeState instance in which the headers are inserted.
+    error: The function to call with any errors found.
+  """
+  fileinfo = FileInfo(filename)
+
+  line = clean_lines.lines[linenum]
+
+  # "include" should use the new style "foo/bar.h" instead of just "bar.h"
+  if _RE_PATTERN_INCLUDE_NEW_STYLE.search(line):
+    error(filename, linenum, 'build/include', 4,
+          'Include the directory when naming .h files')
+
+  # we shouldn't include a file more than once. actually, there are a
+  # handful of instances where doing so is okay, but in general it's
+  # not.
+  match = _RE_PATTERN_INCLUDE.search(line)
+  if match:
+    include = match.group(2)
+    is_system = (match.group(1) == '<')
+    if include in include_state:
+      error(filename, linenum, 'build/include', 4,
+            '"%s" already included at %s:%s' %
+            (include, filename, include_state[include]))
+    else:
+      include_state[include] = linenum
+
+      # We want to ensure that headers appear in the right order:
+      # 1) for foo.cc, foo.h  (preferred location)
+      # 2) c system files
+      # 3) cpp system files
+      # 4) for foo.cc, foo.h  (deprecated location)
+      # 5) other google headers
+      #
+      # We classify each include statement as one of those 5 types
+      # using a number of techniques. The include_state object keeps
+      # track of the highest type seen, and complains if we see a
+      # lower type after that.
+      error_message = include_state.CheckNextIncludeOrder(
+          _ClassifyInclude(fileinfo, include, is_system))
+      if error_message:
+        error(filename, linenum, 'build/include_order', 4,
+              '%s. Should be: %s.h, c system, c++ system, other.' %
+              (error_message, fileinfo.BaseName()))
+      canonical_include = include_state.CanonicalizeAlphabeticalOrder(include)
+      if not include_state.IsInAlphabeticalOrder(
+          clean_lines, linenum, canonical_include):
+        error(filename, linenum, 'build/include_alpha', 4,
+              'Include "%s" not in alphabetical order' % include)
+      include_state.SetLastHeader(canonical_include)
+
+  # Look for any of the stream classes that are part of standard C++.
+  match = _RE_PATTERN_INCLUDE.match(line)
+  if match:
+    include = match.group(2)
+    if Match(r'(f|ind|io|i|o|parse|pf|stdio|str|)?stream$', include):
+      # Many unit tests use cout, so we exempt them.
+      if not _IsTestFilename(filename):
+        error(filename, linenum, 'readability/streams', 3,
+              'Streams are highly discouraged.')
+
+
+def _GetTextInside(text, start_pattern):
+  r"""Retrieves all the text between matching open and close parentheses.
+
+  Given a string of lines and a regular expression string, retrieve all the text
+  following the expression and between opening punctuation symbols like
+  (, [, or {, and the matching close-punctuation symbol. This properly nested
+  occurrences of the punctuations, so for the text like
+    printf(a(), b(c()));
+  a call to _GetTextInside(text, r'printf\(') will return 'a(), b(c())'.
+  start_pattern must match string having an open punctuation symbol at the end.
+
+  Args:
+    text: The lines to extract text. Its comments and strings must be elided.
+           It can be single line and can span multiple lines.
+    start_pattern: The regexp string indicating where to start extracting
+                   the text.
+  Returns:
+    The extracted text.
+    None if either the opening string or ending punctuation could not be found.
+  """
+  # TODO(sugawarayu): Audit cpplint.py to see what places could be profitably
+  # rewritten to use _GetTextInside (and use inferior regexp matching today).
+
+  # Give opening punctuations to get the matching close-punctuations.
+  matching_punctuation = {'(': ')', '{': '}', '[': ']'}
+  closing_punctuation = set(matching_punctuation.itervalues())
+
+  # Find the position to start extracting text.
+  match = re.search(start_pattern, text, re.M)
+  if not match:  # start_pattern not found in text.
+    return None
+  start_position = match.end(0)
+
+  assert start_position > 0, (
+      'start_pattern must ends with an opening punctuation.')
+  assert text[start_position - 1] in matching_punctuation, (
+      'start_pattern must ends with an opening punctuation.')
+  # Stack of closing punctuations we expect to have in text after position.
+  punctuation_stack = [matching_punctuation[text[start_position - 1]]]
+  position = start_position
+  while punctuation_stack and position < len(text):
+    if text[position] == punctuation_stack[-1]:
+      punctuation_stack.pop()
+    elif text[position] in closing_punctuation:
+      # A closing punctuation without matching opening punctuations.
+      return None
+    elif text[position] in matching_punctuation:
+      punctuation_stack.append(matching_punctuation[text[position]])
+    position += 1
+  if punctuation_stack:
+    # Opening punctuations left without matching close-punctuations.
+    return None
+  # punctuations match.
+  return text[start_position:position - 1]
+
+
+# Patterns for matching call-by-reference parameters.
+#
+# Supports nested templates up to 2 levels deep using this messy pattern:
+#   < (?: < (?: < [^<>]*
+#               >
+#           |   [^<>] )*
+#         >
+#     |   [^<>] )*
+#   >
+_RE_PATTERN_IDENT = r'[_a-zA-Z]\w*'  # =~ [[:alpha:]][[:alnum:]]*
+_RE_PATTERN_TYPE = (
+    r'(?:const\s+)?(?:typename\s+|class\s+|struct\s+|union\s+|enum\s+)?'
+    r'(?:\w|'
+    r'\s*<(?:<(?:<[^<>]*>|[^<>])*>|[^<>])*>|'
+    r'::)+')
+# A call-by-reference parameter ends with '& identifier'.
+_RE_PATTERN_REF_PARAM = re.compile(
+    r'(' + _RE_PATTERN_TYPE + r'(?:\s*(?:\bconst\b|[*]))*\s*'
+    r'&\s*' + _RE_PATTERN_IDENT + r')\s*(?:=[^,()]+)?[,)]')
+# A call-by-const-reference parameter either ends with 'const& identifier'
+# or looks like 'const type& identifier' when 'type' is atomic.
+_RE_PATTERN_CONST_REF_PARAM = (
+    r'(?:.*\s*\bconst\s*&\s*' + _RE_PATTERN_IDENT +
+    r'|const\s+' + _RE_PATTERN_TYPE + r'\s*&\s*' + _RE_PATTERN_IDENT + r')')
+
+
+def CheckLanguage(filename, clean_lines, linenum, file_extension,
+                  include_state, nesting_state, error):
+  """Checks rules from the 'C++ language rules' section of cppguide.html.
+
+  Some of these rules are hard to test (function overloading, using
+  uint32 inappropriately), but we do the best we can.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    file_extension: The extension (without the dot) of the filename.
+    include_state: An _IncludeState instance in which the headers are inserted.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: The function to call with any errors found.
+  """
+  # If the line is empty or consists of entirely a comment, no need to
+  # check it.
+  line = clean_lines.elided[linenum]
+  if not line:
+    return
+
+  match = _RE_PATTERN_INCLUDE.search(line)
+  if match:
+    CheckIncludeLine(filename, clean_lines, linenum, include_state, error)
+    return
+
+  # Reset include state across preprocessor directives.  This is meant
+  # to silence warnings for conditional includes.
+  if Match(r'^\s*#\s*(?:ifdef|elif|else|endif)\b', line):
+    include_state.ResetSection()
+
+  # Make Windows paths like Unix.
+  fullname = os.path.abspath(filename).replace('\\', '/')
+
+  # TODO(unknown): figure out if they're using default arguments in fn proto.
+
+  # Check to see if they're using an conversion function cast.
+  # I just try to capture the most common basic types, though there are more.
+  # Parameterless conversion functions, such as bool(), are allowed as they are
+  # probably a member operator declaration or default constructor.
+  match = Search(
+      r'(\bnew\s+)?\b'  # Grab 'new' operator, if it's there
+      r'(int|float|double|bool|char|int32|uint32|int64|uint64)'
+      r'(\([^)].*)', line)
+  if match:
+    matched_new = match.group(1)
+    matched_type = match.group(2)
+    matched_funcptr = match.group(3)
+
+    # gMock methods are defined using some variant of MOCK_METHODx(name, type)
+    # where type may be float(), int(string), etc.  Without context they are
+    # virtually indistinguishable from int(x) casts. Likewise, gMock's
+    # MockCallback takes a template parameter of the form return_type(arg_type),
+    # which looks much like the cast we're trying to detect.
+    #
+    # std::function<> wrapper has a similar problem.
+    #
+    # Return types for function pointers also look like casts if they
+    # don't have an extra space.
+    if (matched_new is None and  # If new operator, then this isn't a cast
+        not (Match(r'^\s*MOCK_(CONST_)?METHOD\d+(_T)?\(', line) or
+             Search(r'\bMockCallback<.*>', line) or
+             Search(r'\bstd::function<.*>', line)) and
+        not (matched_funcptr and
+             Match(r'\((?:[^() ]+::\s*\*\s*)?[^() ]+\)\s*\(',
+                   matched_funcptr))):
+      # Try a bit harder to catch gmock lines: the only place where
+      # something looks like an old-style cast is where we declare the
+      # return type of the mocked method, and the only time when we
+      # are missing context is if MOCK_METHOD was split across
+      # multiple lines.  The missing MOCK_METHOD is usually one or two
+      # lines back, so scan back one or two lines.
+      #
+      # It's not possible for gmock macros to appear in the first 2
+      # lines, since the class head + section name takes up 2 lines.
+      if (linenum < 2 or
+          not (Match(r'^\s*MOCK_(?:CONST_)?METHOD\d+(?:_T)?\((?:\S+,)?\s*$',
+                     clean_lines.elided[linenum - 1]) or
+               Match(r'^\s*MOCK_(?:CONST_)?METHOD\d+(?:_T)?\(\s*$',
+                     clean_lines.elided[linenum - 2]))):
+        error(filename, linenum, 'readability/casting', 4,
+              'Using deprecated casting style.  '
+              'Use static_cast<%s>(...) instead' %
+              matched_type)
+
+  CheckCStyleCast(filename, linenum, line, clean_lines.raw_lines[linenum],
+                  'static_cast',
+                  r'\((int|float|double|bool|char|u?int(16|32|64))\)', error)
+
+  # This doesn't catch all cases. Consider (const char * const)"hello".
+  #
+  # (char *) "foo" should always be a const_cast (reinterpret_cast won't
+  # compile).
+  if CheckCStyleCast(filename, linenum, line, clean_lines.raw_lines[linenum],
+                     'const_cast', r'\((char\s?\*+\s?)\)\s*"', error):
+    pass
+  else:
+    # Check pointer casts for other than string constants
+    CheckCStyleCast(filename, linenum, line, clean_lines.raw_lines[linenum],
+                    'reinterpret_cast', r'\((\w+\s?\*+\s?)\)', error)
+
+  # In addition, we look for people taking the address of a cast.  This
+  # is dangerous -- casts can assign to temporaries, so the pointer doesn't
+  # point where you think.
+  match = Search(
+      r'(?:&\(([^)]+)\)[\w(])|'
+      r'(?:&(static|dynamic|down|reinterpret)_cast\b)', line)
+  if match and match.group(1) != '*':
+    error(filename, linenum, 'runtime/casting', 4,
+          ('Are you taking an address of a cast?  '
+           'This is dangerous: could be a temp var.  '
+           'Take the address before doing the cast, rather than after'))
+
+  # Create an extended_line, which is the concatenation of the current and
+  # next lines, for more effective checking of code that may span more than one
+  # line.
+  if linenum + 1 < clean_lines.NumLines():
+    extended_line = line + clean_lines.elided[linenum + 1]
+  else:
+    extended_line = line
+
+  # Check for people declaring static/global STL strings at the top level.
+  # This is dangerous because the C++ language does not guarantee that
+  # globals with constructors are initialized before the first access.
+  match = Match(
+      r'((?:|static +)(?:|const +))string +([a-zA-Z0-9_:]+)\b(.*)',
+      line)
+  # Make sure it's not a function.
+  # Function template specialization looks like: "string foo<Type>(...".
+  # Class template definitions look like: "string Foo<Type>::Method(...".
+  #
+  # Also ignore things that look like operators.  These are matched separately
+  # because operator names cross non-word boundaries.  If we change the pattern
+  # above, we would decrease the accuracy of matching identifiers.
+  if (match and
+      not Search(r'\boperator\W', line) and
+      not Match(r'\s*(<.*>)?(::[a-zA-Z0-9_]+)?\s*\(([^"]|$)', match.group(3))):
+    error(filename, linenum, 'runtime/string', 4,
+          'For a static/global string constant, use a C style string instead: '
+          '"%schar %s[]".' %
+          (match.group(1), match.group(2)))
+
+  if Search(r'\b([A-Za-z0-9_]*_)\(\1\)', line):
+    error(filename, linenum, 'runtime/init', 4,
+          'You seem to be initializing a member variable with itself.')
+
+  if file_extension == 'h':
+    # TODO(unknown): check that 1-arg constructors are explicit.
+    #                How to tell it's a constructor?
+    #                (handled in CheckForNonStandardConstructs for now)
+    # TODO(unknown): check that classes have DISALLOW_EVIL_CONSTRUCTORS
+    #                (level 1 error)
+    pass
+
+  # Check if people are using the verboten C basic types.  The only exception
+  # we regularly allow is "unsigned short port" for port.
+  if Search(r'\bshort port\b', line):
+    if not Search(r'\bunsigned short port\b', line):
+      error(filename, linenum, 'runtime/int', 4,
+            'Use "unsigned short" for ports, not "short"')
+  else:
+    match = Search(r'\b(short|long(?! +double)|long long)\b', line)
+    if match:
+      error(filename, linenum, 'runtime/int', 4,
+            'Use int16/int64/etc, rather than the C type %s' % match.group(1))
+
+  # When snprintf is used, the second argument shouldn't be a literal.
+  match = Search(r'snprintf\s*\(([^,]*),\s*([0-9]*)\s*,', line)
+  if match and match.group(2) != '0':
+    # If 2nd arg is zero, snprintf is used to calculate size.
+    error(filename, linenum, 'runtime/printf', 3,
+          'If you can, use sizeof(%s) instead of %s as the 2nd arg '
+          'to snprintf.' % (match.group(1), match.group(2)))
+
+  # Check if some verboten C functions are being used.
+  if Search(r'\bsprintf\b', line):
+    error(filename, linenum, 'runtime/printf', 5,
+          'Never use sprintf.  Use snprintf instead.')
+  match = Search(r'\b(strcpy|strcat)\b', line)
+  if match:
+    error(filename, linenum, 'runtime/printf', 4,
+          'Almost always, snprintf is better than %s' % match.group(1))
+
+  # Check if some verboten operator overloading is going on
+  # TODO(unknown): catch out-of-line unary operator&:
+  #   class X {};
+  #   int operator&(const X& x) { return 42; }  // unary operator&
+  # The trick is it's hard to tell apart from binary operator&:
+  #   class Y { int operator&(const Y& x) { return 23; } }; // binary operator&
+  if Search(r'\boperator\s*&\s*\(\s*\)', line):
+    error(filename, linenum, 'runtime/operator', 4,
+          'Unary operator& is dangerous.  Do not use it.')
+
+  # Check for suspicious usage of "if" like
+  # } if (a == b) {
+  if Search(r'\}\s*if\s*\(', line):
+    error(filename, linenum, 'readability/braces', 4,
+          'Did you mean "else if"? If not, start a new line for "if".')
+
+  # Check for potential format string bugs like printf(foo).
+  # We constrain the pattern not to pick things like DocidForPrintf(foo).
+  # Not perfect but it can catch printf(foo.c_str()) and printf(foo->c_str())
+  # TODO(sugawarayu): Catch the following case. Need to change the calling
+  # convention of the whole function to process multiple line to handle it.
+  #   printf(
+  #       boy_this_is_a_really_long_variable_that_cannot_fit_on_the_prev_line);
+  printf_args = _GetTextInside(line, r'(?i)\b(string)?printf\s*\(')
+  if printf_args:
+    match = Match(r'([\w.\->()]+)$', printf_args)
+    if match and match.group(1) != '__VA_ARGS__':
+      function_name = re.search(r'\b((?:string)?printf)\s*\(',
+                                line, re.I).group(1)
+      error(filename, linenum, 'runtime/printf', 4,
+            'Potential format string bug. Do %s("%%s", %s) instead.'
+            % (function_name, match.group(1)))
+
+  # Check for potential memset bugs like memset(buf, sizeof(buf), 0).
+  match = Search(r'memset\s*\(([^,]*),\s*([^,]*),\s*0\s*\)', line)
+  if match and not Match(r"^''|-?[0-9]+|0x[0-9A-Fa-f]$", match.group(2)):
+    error(filename, linenum, 'runtime/memset', 4,
+          'Did you mean "memset(%s, 0, %s)"?'
+          % (match.group(1), match.group(2)))
+
+  if Search(r'\busing namespace\b', line):
+    error(filename, linenum, 'build/namespaces', 5,
+          'Do not use namespace using-directives.  '
+          'Use using-declarations instead.')
+
+  # Detect variable-length arrays.
+  match = Match(r'\s*(.+::)?(\w+) [a-z]\w*\[(.+)];', line)
+  if (match and match.group(2) != 'return' and match.group(2) != 'delete' and
+      match.group(3).find(']') == -1):
+    # Split the size using space and arithmetic operators as delimiters.
+    # If any of the resulting tokens are not compile time constants then
+    # report the error.
+    tokens = re.split(r'\s|\+|\-|\*|\/|<<|>>]', match.group(3))
+    is_const = True
+    skip_next = False
+    for tok in tokens:
+      if skip_next:
+        skip_next = False
+        continue
+
+      if Search(r'sizeof\(.+\)', tok): continue
+      if Search(r'arraysize\(\w+\)', tok): continue
+
+      tok = tok.lstrip('(')
+      tok = tok.rstrip(')')
+      if not tok: continue
+      if Match(r'\d+', tok): continue
+      if Match(r'0[xX][0-9a-fA-F]+', tok): continue
+      if Match(r'k[A-Z0-9]\w*', tok): continue
+      if Match(r'(.+::)?k[A-Z0-9]\w*', tok): continue
+      if Match(r'(.+::)?[A-Z][A-Z0-9_]*', tok): continue
+      # A catch all for tricky sizeof cases, including 'sizeof expression',
+      # 'sizeof(*type)', 'sizeof(const type)', 'sizeof(struct StructName)'
+      # requires skipping the next token because we split on ' ' and '*'.
+      if tok.startswith('sizeof'):
+        skip_next = True
+        continue
+      is_const = False
+      break
+    if not is_const:
+      error(filename, linenum, 'runtime/arrays', 1,
+            'Do not use variable-length arrays.  Use an appropriately named '
+            "('k' followed by CamelCase) compile-time constant for the size.")
+
+  # If DISALLOW_EVIL_CONSTRUCTORS, DISALLOW_COPY_AND_ASSIGN, or
+  # DISALLOW_IMPLICIT_CONSTRUCTORS is present, then it should be the last thing
+  # in the class declaration.
+  match = Match(
+      (r'\s*'
+       r'(DISALLOW_(EVIL_CONSTRUCTORS|COPY_AND_ASSIGN|IMPLICIT_CONSTRUCTORS))'
+       r'\(.*\);$'),
+      line)
+  if match and linenum + 1 < clean_lines.NumLines():
+    next_line = clean_lines.elided[linenum + 1]
+    # We allow some, but not all, declarations of variables to be present
+    # in the statement that defines the class.  The [\w\*,\s]* fragment of
+    # the regular expression below allows users to declare instances of
+    # the class or pointers to instances, but not less common types such
+    # as function pointers or arrays.  It's a tradeoff between allowing
+    # reasonable code and avoiding trying to parse more C++ using regexps.
+    if not Search(r'^\s*}[\w\*,\s]*;', next_line):
+      error(filename, linenum, 'readability/constructors', 3,
+            match.group(1) + ' should be the last thing in the class')
+
+  # Check for use of unnamed namespaces in header files.  Registration
+  # macros are typically OK, so we allow use of "namespace {" on lines
+  # that end with backslashes.
+  if (file_extension == 'h'
+      and Search(r'\bnamespace\s*{', line)
+      and line[-1] != '\\'):
+    error(filename, linenum, 'build/namespaces', 4,
+          'Do not use unnamed namespaces in header files.  See '
+          'http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml#Namespaces'
+          ' for more information.')
+
+def CheckForNonConstReference(filename, clean_lines, linenum,
+                              nesting_state, error):
+  """Check for non-const references.
+
+  Separate from CheckLanguage since it scans backwards from current
+  line, instead of scanning forward.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: The function to call with any errors found.
+  """
+  # Do nothing if there is no '&' on current line.
+  line = clean_lines.elided[linenum]
+  if '&' not in line:
+    return
+
+  # Long type names may be broken across multiple lines, usually in one
+  # of these forms:
+  #   LongType
+  #       ::LongTypeContinued &identifier
+  #   LongType::
+  #       LongTypeContinued &identifier
+  #   LongType<
+  #       ...>::LongTypeContinued &identifier
+  #
+  # If we detected a type split across two lines, join the previous
+  # line to current line so that we can match const references
+  # accordingly.
+  #
+  # Note that this only scans back one line, since scanning back
+  # arbitrary number of lines would be expensive.  If you have a type
+  # that spans more than 2 lines, please use a typedef.
+  if linenum > 1:
+    previous = None
+    if Match(r'\s*::(?:[\w<>]|::)+\s*&\s*\S', line):
+      # previous_line\n + ::current_line
+      previous = Search(r'\b((?:const\s*)?(?:[\w<>]|::)+[\w<>])\s*$',
+                        clean_lines.elided[linenum - 1])
+    elif Match(r'\s*[a-zA-Z_]([\w<>]|::)+\s*&\s*\S', line):
+      # previous_line::\n + current_line
+      previous = Search(r'\b((?:const\s*)?(?:[\w<>]|::)+::)\s*$',
+                        clean_lines.elided[linenum - 1])
+    if previous:
+      line = previous.group(1) + line.lstrip()
+    else:
+      # Check for templated parameter that is split across multiple lines
+      endpos = line.rfind('>')
+      if endpos > -1:
+        (_, startline, startpos) = ReverseCloseExpression(
+            clean_lines, linenum, endpos)
+        if startpos > -1 and startline < linenum:
+          # Found the matching < on an earlier line, collect all
+          # pieces up to current line.
+          line = ''
+          for i in xrange(startline, linenum + 1):
+            line += clean_lines.elided[i].strip()
+
+  # Check for non-const references in function parameters.  A single '&' may
+  # found in the following places:
+  #   inside expression: binary & for bitwise AND
+  #   inside expression: unary & for taking the address of something
+  #   inside declarators: reference parameter
+  # We will exclude the first two cases by checking that we are not inside a
+  # function body, including one that was just introduced by a trailing '{'.
+  # TODO(unknwon): Doesn't account for preprocessor directives.
+  # TODO(unknown): Doesn't account for 'catch(Exception& e)' [rare].
+  check_params = False
+  if not nesting_state.stack:
+    check_params = True  # top level
+  elif (isinstance(nesting_state.stack[-1], _ClassInfo) or
+        isinstance(nesting_state.stack[-1], _NamespaceInfo)):
+    check_params = True  # within class or namespace
+  elif Match(r'.*{\s*$', line):
+    if (len(nesting_state.stack) == 1 or
+        isinstance(nesting_state.stack[-2], _ClassInfo) or
+        isinstance(nesting_state.stack[-2], _NamespaceInfo)):
+      check_params = True  # just opened global/class/namespace block
+  # We allow non-const references in a few standard places, like functions
+  # called "swap()" or iostream operators like "<<" or ">>".  Do not check
+  # those function parameters.
+  #
+  # We also accept & in static_assert, which looks like a function but
+  # it's actually a declaration expression.
+  whitelisted_functions = (r'(?:[sS]wap(?:<\w:+>)?|'
+                           r'operator\s*[<>][<>]|'
+                           r'static_assert|COMPILE_ASSERT'
+                           r')\s*\(')
+  if Search(whitelisted_functions, line):
+    check_params = False
+  elif not Search(r'\S+\([^)]*$', line):
+    # Don't see a whitelisted function on this line.  Actually we
+    # didn't see any function name on this line, so this is likely a
+    # multi-line parameter list.  Try a bit harder to catch this case.
+    for i in xrange(2):
+      if (linenum > i and
+          Search(whitelisted_functions, clean_lines.elided[linenum - i - 1])):
+        check_params = False
+        break
+
+  if check_params:
+    decls = ReplaceAll(r'{[^}]*}', ' ', line)  # exclude function body
+    for parameter in re.findall(_RE_PATTERN_REF_PARAM, decls):
+      if not Match(_RE_PATTERN_CONST_REF_PARAM, parameter):
+        error(filename, linenum, 'runtime/references', 2,
+              'Is this a non-const reference? '
+              'If so, make const or use a pointer: ' +
+              ReplaceAll(' *<', '<', parameter))
+
+
+def CheckCStyleCast(filename, linenum, line, raw_line, cast_type, pattern,
+                    error):
+  """Checks for a C-style cast by looking for the pattern.
+
+  Args:
+    filename: The name of the current file.
+    linenum: The number of the line to check.
+    line: The line of code to check.
+    raw_line: The raw line of code to check, with comments.
+    cast_type: The string for the C++ cast to recommend.  This is either
+      reinterpret_cast, static_cast, or const_cast, depending.
+    pattern: The regular expression used to find C-style casts.
+    error: The function to call with any errors found.
+
+  Returns:
+    True if an error was emitted.
+    False otherwise.
+  """
+  match = Search(pattern, line)
+  if not match:
+    return False
+
+  # e.g., sizeof(int)
+  sizeof_match = Match(r'.*sizeof\s*$', line[0:match.start(1) - 1])
+  if sizeof_match:
+    error(filename, linenum, 'runtime/sizeof', 1,
+          'Using sizeof(type).  Use sizeof(varname) instead if possible')
+    return True
+
+  # operator++(int) and operator--(int)
+  if (line[0:match.start(1) - 1].endswith(' operator++') or
+      line[0:match.start(1) - 1].endswith(' operator--')):
+    return False
+
+  # A single unnamed argument for a function tends to look like old
+  # style cast.  If we see those, don't issue warnings for deprecated
+  # casts, instead issue warnings for unnamed arguments where
+  # appropriate.
+  #
+  # These are things that we want warnings for, since the style guide
+  # explicitly require all parameters to be named:
+  #   Function(int);
+  #   Function(int) {
+  #   ConstMember(int) const;
+  #   ConstMember(int) const {
+  #   ExceptionMember(int) throw (...);
+  #   ExceptionMember(int) throw (...) {
+  #   PureVirtual(int) = 0;
+  #
+  # These are functions of some sort, where the compiler would be fine
+  # if they had named parameters, but people often omit those
+  # identifiers to reduce clutter:
+  #   (FunctionPointer)(int);
+  #   (FunctionPointer)(int) = value;
+  #   Function((function_pointer_arg)(int))
+  #   <TemplateArgument(int)>;
+  #   <(FunctionPointerTemplateArgument)(int)>;
+  remainder = line[match.end(0):]
+  if Match(r'^\s*(?:;|const\b|throw\b|=|>|\{|\))', remainder):
+    # Looks like an unnamed parameter.
+
+    # Don't warn on any kind of template arguments.
+    if Match(r'^\s*>', remainder):
+      return False
+
+    # Don't warn on assignments to function pointers, but keep warnings for
+    # unnamed parameters to pure virtual functions.  Note that this pattern
+    # will also pass on assignments of "0" to function pointers, but the
+    # preferred values for those would be "nullptr" or "NULL".
+    matched_zero = Match(r'^\s=\s*(\S+)\s*;', remainder)
+    if matched_zero and matched_zero.group(1) != '0':
+      return False
+
+    # Don't warn on function pointer declarations.  For this we need
+    # to check what came before the "(type)" string.
+    if Match(r'.*\)\s*$', line[0:match.start(0)]):
+      return False
+
+    # Don't warn if the parameter is named with block comments, e.g.:
+    #  Function(int /*unused_param*/);
+    if '/*' in raw_line:
+      return False
+
+    # Passed all filters, issue warning here.
+    error(filename, linenum, 'readability/function', 3,
+          'All parameters should be named in a function')
+    return True
+
+  # At this point, all that should be left is actual casts.
+  error(filename, linenum, 'readability/casting', 4,
+        'Using C-style cast.  Use %s<%s>(...) instead' %
+        (cast_type, match.group(1)))
+
+  return True
+
+
+_HEADERS_CONTAINING_TEMPLATES = (
+    ('<deque>', ('deque',)),
+    ('<functional>', ('unary_function', 'binary_function',
+                      'plus', 'minus', 'multiplies', 'divides', 'modulus',
+                      'negate',
+                      'equal_to', 'not_equal_to', 'greater', 'less',
+                      'greater_equal', 'less_equal',
+                      'logical_and', 'logical_or', 'logical_not',
+                      'unary_negate', 'not1', 'binary_negate', 'not2',
+                      'bind1st', 'bind2nd',
+                      'pointer_to_unary_function',
+                      'pointer_to_binary_function',
+                      'ptr_fun',
+                      'mem_fun_t', 'mem_fun', 'mem_fun1_t', 'mem_fun1_ref_t',
+                      'mem_fun_ref_t',
+                      'const_mem_fun_t', 'const_mem_fun1_t',
+                      'const_mem_fun_ref_t', 'const_mem_fun1_ref_t',
+                      'mem_fun_ref',
+                     )),
+    ('<limits>', ('numeric_limits',)),
+    ('<list>', ('list',)),
+    ('<map>', ('map', 'multimap',)),
+    ('<memory>', ('allocator',)),
+    ('<queue>', ('queue', 'priority_queue',)),
+    ('<set>', ('set', 'multiset',)),
+    ('<stack>', ('stack',)),
+    ('<string>', ('char_traits', 'basic_string',)),
+    ('<utility>', ('pair',)),
+    ('<vector>', ('vector',)),
+
+    # gcc extensions.
+    # Note: std::hash is their hash, ::hash is our hash
+    ('<hash_map>', ('hash_map', 'hash_multimap',)),
+    ('<hash_set>', ('hash_set', 'hash_multiset',)),
+    ('<slist>', ('slist',)),
+    )
+
+_RE_PATTERN_STRING = re.compile(r'\bstring\b')
+
+_re_pattern_algorithm_header = []
+for _template in ('copy', 'max', 'min', 'min_element', 'sort', 'swap',
+                  'transform'):
+  # Match max<type>(..., ...), max(..., ...), but not foo->max, foo.max or
+  # type::max().
+  _re_pattern_algorithm_header.append(
+      (re.compile(r'[^>.]\b' + _template + r'(<.*?>)?\([^\)]'),
+       _template,
+       '<algorithm>'))
+
+_re_pattern_templates = []
+for _header, _templates in _HEADERS_CONTAINING_TEMPLATES:
+  for _template in _templates:
+    _re_pattern_templates.append(
+        (re.compile(r'(\<|\b)' + _template + r'\s*\<'),
+         _template + '<>',
+         _header))
+
+
+def FilesBelongToSameModule(filename_cc, filename_h):
+  """Check if these two filenames belong to the same module.
+
+  The concept of a 'module' here is a as follows:
+  foo.h, foo-inl.h, foo.cc, foo_test.cc and foo_unittest.cc belong to the
+  same 'module' if they are in the same directory.
+  some/path/public/xyzzy and some/path/internal/xyzzy are also considered
+  to belong to the same module here.
+
+  If the filename_cc contains a longer path than the filename_h, for example,
+  '/absolute/path/to/base/sysinfo.cc', and this file would include
+  'base/sysinfo.h', this function also produces the prefix needed to open the
+  header. This is used by the caller of this function to more robustly open the
+  header file. We don't have access to the real include paths in this context,
+  so we need this guesswork here.
+
+  Known bugs: tools/base/bar.cc and base/bar.h belong to the same module
+  according to this implementation. Because of this, this function gives
+  some false positives. This should be sufficiently rare in practice.
+
+  Args:
+    filename_cc: is the path for the .cc file
+    filename_h: is the path for the header path
+
+  Returns:
+    Tuple with a bool and a string:
+    bool: True if filename_cc and filename_h belong to the same module.
+    string: the additional prefix needed to open the header file.
+  """
+
+  if not filename_cc.endswith('.cc'):
+    return (False, '')
+  filename_cc = filename_cc[:-len('.cc')]
+  if filename_cc.endswith('_unittest'):
+    filename_cc = filename_cc[:-len('_unittest')]
+  elif filename_cc.endswith('_test'):
+    filename_cc = filename_cc[:-len('_test')]
+  filename_cc = filename_cc.replace('/public/', '/')
+  filename_cc = filename_cc.replace('/internal/', '/')
+
+  if not filename_h.endswith('.h'):
+    return (False, '')
+  filename_h = filename_h[:-len('.h')]
+  if filename_h.endswith('-inl'):
+    filename_h = filename_h[:-len('-inl')]
+  filename_h = filename_h.replace('/public/', '/')
+  filename_h = filename_h.replace('/internal/', '/')
+
+  files_belong_to_same_module = filename_cc.endswith(filename_h)
+  common_path = ''
+  if files_belong_to_same_module:
+    common_path = filename_cc[:-len(filename_h)]
+  return files_belong_to_same_module, common_path
+
+
+def UpdateIncludeState(filename, include_state, io=codecs):
+  """Fill up the include_state with new includes found from the file.
+
+  Args:
+    filename: the name of the header to read.
+    include_state: an _IncludeState instance in which the headers are inserted.
+    io: The io factory to use to read the file. Provided for testability.
+
+  Returns:
+    True if a header was succesfully added. False otherwise.
+  """
+  headerfile = None
+  try:
+    headerfile = io.open(filename, 'r', 'utf8', 'replace')
+  except IOError:
+    return False
+  linenum = 0
+  for line in headerfile:
+    linenum += 1
+    clean_line = CleanseComments(line)
+    match = _RE_PATTERN_INCLUDE.search(clean_line)
+    if match:
+      include = match.group(2)
+      # The value formatting is cute, but not really used right now.
+      # What matters here is that the key is in include_state.
+      include_state.setdefault(include, '%s:%d' % (filename, linenum))
+  return True
+
+
+def CheckForIncludeWhatYouUse(filename, clean_lines, include_state, error,
+                              io=codecs):
+  """Reports for missing stl includes.
+
+  This function will output warnings to make sure you are including the headers
+  necessary for the stl containers and functions that you use. We only give one
+  reason to include a header. For example, if you use both equal_to<> and
+  less<> in a .h file, only one (the latter in the file) of these will be
+  reported as a reason to include the <functional>.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    include_state: An _IncludeState instance.
+    error: The function to call with any errors found.
+    io: The IO factory to use to read the header file. Provided for unittest
+        injection.
+  """
+  required = {}  # A map of header name to linenumber and the template entity.
+                 # Example of required: { '<functional>': (1219, 'less<>') }
+
+  for linenum in xrange(clean_lines.NumLines()):
+    line = clean_lines.elided[linenum]
+    if not line or line[0] == '#':
+      continue
+
+    # String is special -- it is a non-templatized type in STL.
+    matched = _RE_PATTERN_STRING.search(line)
+    if matched:
+      # Don't warn about strings in non-STL namespaces:
+      # (We check only the first match per line; good enough.)
+      prefix = line[:matched.start()]
+      if prefix.endswith('std::') or not prefix.endswith('::'):
+        required['<string>'] = (linenum, 'string')
+
+    for pattern, template, header in _re_pattern_algorithm_header:
+      if pattern.search(line):
+        required[header] = (linenum, template)
+
+    # The following function is just a speed up, no semantics are changed.
+    if not '<' in line:  # Reduces the cpu time usage by skipping lines.
+      continue
+
+    for pattern, template, header in _re_pattern_templates:
+      if pattern.search(line):
+        required[header] = (linenum, template)
+
+  # The policy is that if you #include something in foo.h you don't need to
+  # include it again in foo.cc. Here, we will look at possible includes.
+  # Let's copy the include_state so it is only messed up within this function.
+  include_state = include_state.copy()
+
+  # Did we find the header for this file (if any) and succesfully load it?
+  header_found = False
+
+  # Use the absolute path so that matching works properly.
+  abs_filename = FileInfo(filename).FullName()
+
+  # For Emacs's flymake.
+  # If cpplint is invoked from Emacs's flymake, a temporary file is generated
+  # by flymake and that file name might end with '_flymake.cc'. In that case,
+  # restore original file name here so that the corresponding header file can be
+  # found.
+  # e.g. If the file name is 'foo_flymake.cc', we should search for 'foo.h'
+  # instead of 'foo_flymake.h'
+  abs_filename = re.sub(r'_flymake\.cc$', '.cc', abs_filename)
+
+  # include_state is modified during iteration, so we iterate over a copy of
+  # the keys.
+  header_keys = include_state.keys()
+  for header in header_keys:
+    (same_module, common_path) = FilesBelongToSameModule(abs_filename, header)
+    fullpath = common_path + header
+    if same_module and UpdateIncludeState(fullpath, include_state, io):
+      header_found = True
+
+  # If we can't find the header file for a .cc, assume it's because we don't
+  # know where to look. In that case we'll give up as we're not sure they
+  # didn't include it in the .h file.
+  # TODO(unknown): Do a better job of finding .h files so we are confident that
+  # not having the .h file means there isn't one.
+  if filename.endswith('.cc') and not header_found:
+    return
+
+  # All the lines have been processed, report the errors found.
+  for required_header_unstripped in required:
+    template = required[required_header_unstripped][1]
+    if required_header_unstripped.strip('<>"') not in include_state:
+      error(filename, required[required_header_unstripped][0],
+            'build/include_what_you_use', 4,
+            'Add #include ' + required_header_unstripped + ' for ' + template)
+
+
+_RE_PATTERN_EXPLICIT_MAKEPAIR = re.compile(r'\bmake_pair\s*<')
+
+
+def CheckMakePairUsesDeduction(filename, clean_lines, linenum, error):
+  """Check that make_pair's template arguments are deduced.
+
+  G++ 4.6 in C++0x mode fails badly if make_pair's template arguments are
+  specified explicitly, and such use isn't intended in any case.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+  match = _RE_PATTERN_EXPLICIT_MAKEPAIR.search(line)
+  if match:
+    error(filename, linenum, 'build/explicit_make_pair',
+          4,  # 4 = high confidence
+          'For C++11-compatibility, omit template arguments from make_pair'
+          ' OR use pair directly OR if appropriate, construct a pair directly')
+
+
+def ProcessLine(filename, file_extension, clean_lines, line,
+                include_state, function_state, nesting_state, error,
+                extra_check_functions=[]):
+  """Processes a single line in the file.
+
+  Args:
+    filename: Filename of the file that is being processed.
+    file_extension: The extension (dot not included) of the file.
+    clean_lines: An array of strings, each representing a line of the file,
+                 with comments stripped.
+    line: Number of line being processed.
+    include_state: An _IncludeState instance in which the headers are inserted.
+    function_state: A _FunctionState instance which counts function lines, etc.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: A callable to which errors are reported, which takes 4 arguments:
+           filename, line number, error level, and message
+    extra_check_functions: An array of additional check functions that will be
+                           run on each source line. Each function takes 4
+                           arguments: filename, clean_lines, line, error
+  """
+  raw_lines = clean_lines.raw_lines
+  ParseNolintSuppressions(filename, raw_lines[line], line, error)
+  nesting_state.Update(filename, clean_lines, line, error)
+  if nesting_state.stack and nesting_state.stack[-1].inline_asm != _NO_ASM:
+    return
+  CheckForFunctionLengths(filename, clean_lines, line, function_state, error)
+  CheckForMultilineCommentsAndStrings(filename, clean_lines, line, error)
+  CheckStyle(filename, clean_lines, line, file_extension, nesting_state, error)
+  CheckLanguage(filename, clean_lines, line, file_extension, include_state,
+                nesting_state, error)
+  CheckForNonConstReference(filename, clean_lines, line, nesting_state, error)
+  CheckForNonStandardConstructs(filename, clean_lines, line,
+                                nesting_state, error)
+  CheckVlogArguments(filename, clean_lines, line, error)
+  CheckPosixThreading(filename, clean_lines, line, error)
+  CheckInvalidIncrement(filename, clean_lines, line, error)
+  CheckMakePairUsesDeduction(filename, clean_lines, line, error)
+  for check_fn in extra_check_functions:
+    check_fn(filename, clean_lines, line, error)
+
+def ProcessFileData(filename, file_extension, lines, error,
+                    extra_check_functions=[]):
+  """Performs lint checks and reports any errors to the given error function.
+
+  Args:
+    filename: Filename of the file that is being processed.
+    file_extension: The extension (dot not included) of the file.
+    lines: An array of strings, each representing a line of the file, with the
+           last element being empty if the file is terminated with a newline.
+    error: A callable to which errors are reported, which takes 4 arguments:
+           filename, line number, error level, and message
+    extra_check_functions: An array of additional check functions that will be
+                           run on each source line. Each function takes 4
+                           arguments: filename, clean_lines, line, error
+  """
+  lines = (['// marker so line numbers and indices both start at 1'] + lines +
+           ['// marker so line numbers end in a known way'])
+
+  include_state = _IncludeState()
+  function_state = _FunctionState()
+  nesting_state = _NestingState()
+
+  ResetNolintSuppressions()
+
+  CheckForCopyright(filename, lines, error)
+
+  if file_extension == 'h':
+    CheckForHeaderGuard(filename, lines, error)
+
+  RemoveMultiLineComments(filename, lines, error)
+  clean_lines = CleansedLines(lines)
+  for line in xrange(clean_lines.NumLines()):
+    ProcessLine(filename, file_extension, clean_lines, line,
+                include_state, function_state, nesting_state, error,
+                extra_check_functions)
+  nesting_state.CheckCompletedBlocks(filename, error)
+
+  CheckForIncludeWhatYouUse(filename, clean_lines, include_state, error)
+
+  # We check here rather than inside ProcessLine so that we see raw
+  # lines rather than "cleaned" lines.
+  CheckForBadCharacters(filename, lines, error)
+
+  CheckForNewlineAtEOF(filename, lines, error)
+
+def ProcessFile(filename, vlevel, extra_check_functions=[]):
+  """Does google-lint on a single file.
+
+  Args:
+    filename: The name of the file to parse.
+
+    vlevel: The level of errors to report.  Every error of confidence
+    >= verbose_level will be reported.  0 is a good default.
+
+    extra_check_functions: An array of additional check functions that will be
+                           run on each source line. Each function takes 4
+                           arguments: filename, clean_lines, line, error
+  """
+
+  _SetVerboseLevel(vlevel)
+
+  try:
+    # Support the UNIX convention of using "-" for stdin.  Note that
+    # we are not opening the file with universal newline support
+    # (which codecs doesn't support anyway), so the resulting lines do
+    # contain trailing '\r' characters if we are reading a file that
+    # has CRLF endings.
+    # If after the split a trailing '\r' is present, it is removed
+    # below. If it is not expected to be present (i.e. os.linesep !=
+    # '\r\n' as in Windows), a warning is issued below if this file
+    # is processed.
+
+    if filename == '-':
+      lines = codecs.StreamReaderWriter(sys.stdin,
+                                        codecs.getreader('utf8'),
+                                        codecs.getwriter('utf8'),
+                                        'replace').read().split('\n')
+    else:
+      lines = codecs.open(filename, 'r', 'utf8', 'replace').read().split('\n')
+
+    carriage_return_found = False
+    # Remove trailing '\r'.
+    for linenum in range(len(lines)):
+      if lines[linenum].endswith('\r'):
+        lines[linenum] = lines[linenum].rstrip('\r')
+        carriage_return_found = True
+
+  except IOError:
+    sys.stderr.write(
+        "Skipping input '%s': Can't open for reading\n" % filename)
+    return
+
+  # Note, if no dot is found, this will give the entire filename as the ext.
+  file_extension = filename[filename.rfind('.') + 1:]
+
+  # When reading from stdin, the extension is unknown, so no cpplint tests
+  # should rely on the extension.
+  if filename != '-' and file_extension not in _valid_extensions:
+    sys.stderr.write('Ignoring %s; not a valid file name '
+                     '(%s)\n' % (filename, ', '.join(_valid_extensions)))
+  else:
+    ProcessFileData(filename, file_extension, lines, Error,
+                    extra_check_functions)
+    if carriage_return_found and os.linesep != '\r\n':
+      # Use 0 for linenum since outputting only one error for potentially
+      # several lines.
+      Error(filename, 0, 'whitespace/newline', 1,
+            'One or more unexpected \\r (^M) found;'
+            'better to use only a \\n')
+
+  sys.stderr.write('Done processing %s\n' % filename)
+
+
+def PrintUsage(message):
+  """Prints a brief usage string and exits, optionally with an error message.
+
+  Args:
+    message: The optional error message.
+  """
+  sys.stderr.write(_USAGE)
+  if message:
+    sys.exit('\nFATAL ERROR: ' + message)
+  else:
+    sys.exit(1)
+
+
+def PrintCategories():
+  """Prints a list of all the error-categories used by error messages.
+
+  These are the categories used to filter messages via --filter.
+  """
+  sys.stderr.write(''.join('  %s\n' % cat for cat in _ERROR_CATEGORIES))
+  sys.exit(0)
+
+
+def ParseArguments(args):
+  """Parses the command line arguments.
+
+  This may set the output format and verbosity level as side-effects.
+
+  Args:
+    args: The command line arguments:
+
+  Returns:
+    The list of filenames to lint.
+  """
+  try:
+    (opts, filenames) = getopt.getopt(args, '', ['help', 'output=', 'verbose=',
+                                                 'counting=',
+                                                 'filter=',
+                                                 'root=',
+                                                 'linelength=',
+                                                 'extensions='])
+  except getopt.GetoptError:
+    PrintUsage('Invalid arguments.')
+
+  verbosity = _VerboseLevel()
+  output_format = _OutputFormat()
+  filters = ''
+  counting_style = ''
+
+  for (opt, val) in opts:
+    if opt == '--help':
+      PrintUsage(None)
+    elif opt == '--output':
+      if val not in ('emacs', 'vs7', 'eclipse'):
+        PrintUsage('The only allowed output formats are emacs, vs7 and eclipse.')
+      output_format = val
+    elif opt == '--verbose':
+      verbosity = int(val)
+    elif opt == '--filter':
+      filters = val
+      if not filters:
+        PrintCategories()
+    elif opt == '--counting':
+      if val not in ('total', 'toplevel', 'detailed'):
+        PrintUsage('Valid counting options are total, toplevel, and detailed')
+      counting_style = val
+    elif opt == '--root':
+      global _root
+      _root = val
+    elif opt == '--linelength':
+      global _line_length
+      try:
+          _line_length = int(val)
+      except ValueError:
+          PrintUsage('Line length must be digits.')
+    elif opt == '--extensions':
+      global _valid_extensions
+      try:
+          _valid_extensions = set(val.split(','))
+      except ValueError:
+          PrintUsage('Extensions must be comma seperated list.')
+
+  if not filenames:
+    PrintUsage('No files were specified.')
+
+  _SetOutputFormat(output_format)
+  _SetVerboseLevel(verbosity)
+  _SetFilters(filters)
+  _SetCountingStyle(counting_style)
+
+  return filenames
+
+
+def main():
+  filenames = ParseArguments(sys.argv[1:])
+
+  # Change stderr to write with replacement characters so we don't die
+  # if we try to print something containing non-ASCII characters.
+  sys.stderr = codecs.StreamReaderWriter(sys.stderr,
+                                         codecs.getreader('utf8'),
+                                         codecs.getwriter('utf8'),
+                                         'replace')
+
+  _cpplint_state.ResetErrorCounts()
+  for filename in filenames:
+    ProcessFile(filename, _cpplint_state.verbose_level)
+  _cpplint_state.PrintErrorCounts()
+
+  sys.exit(_cpplint_state.error_count > 0)
+
+
+if __name__ == '__main__':
+  main()
diff --git a/libvpx/libvpx/tools/diff.py b/libvpx/libvpx/tools/diff.py
new file mode 100644
index 0000000..a96c7db
--- /dev/null
+++ b/libvpx/libvpx/tools/diff.py
@@ -0,0 +1,130 @@
+#!/usr/bin/env python
+##  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+"""Classes for representing diff pieces."""
+
+__author__ = "jkoleszar@google.com"
+
+import re
+
+
+class DiffLines(object):
+    """A container for one half of a diff."""
+
+    def __init__(self, filename, offset, length):
+        self.filename = filename
+        self.offset = offset
+        self.length = length
+        self.lines = []
+        self.delta_line_nums = []
+
+    def Append(self, line):
+        l = len(self.lines)
+        if line[0] != " ":
+            self.delta_line_nums.append(self.offset + l)
+        self.lines.append(line[1:])
+        assert l+1 <= self.length
+
+    def Complete(self):
+        return len(self.lines) == self.length
+
+    def __contains__(self, item):
+        return item >= self.offset and item <= self.offset + self.length - 1
+
+
+class DiffHunk(object):
+    """A container for one diff hunk, consisting of two DiffLines."""
+
+    def __init__(self, header, file_a, file_b, start_a, len_a, start_b, len_b):
+        self.header = header
+        self.left = DiffLines(file_a, start_a, len_a)
+        self.right = DiffLines(file_b, start_b, len_b)
+        self.lines = []
+
+    def Append(self, line):
+        """Adds a line to the DiffHunk and its DiffLines children."""
+        if line[0] == "-":
+            self.left.Append(line)
+        elif line[0] == "+":
+            self.right.Append(line)
+        elif line[0] == " ":
+            self.left.Append(line)
+            self.right.Append(line)
+        elif line[0] == "\\":
+            # Ignore newline messages from git diff.
+            pass
+        else:
+            assert False, ("Unrecognized character at start of diff line "
+                           "%r" % line[0])
+        self.lines.append(line)
+
+    def Complete(self):
+        return self.left.Complete() and self.right.Complete()
+
+    def __repr__(self):
+        return "DiffHunk(%s, %s, len %d)" % (
+            self.left.filename, self.right.filename,
+            max(self.left.length, self.right.length))
+
+
+def ParseDiffHunks(stream):
+    """Walk a file-like object, yielding DiffHunks as they're parsed."""
+
+    file_regex = re.compile(r"(\+\+\+|---) (\S+)")
+    range_regex = re.compile(r"@@ -(\d+)(,(\d+))? \+(\d+)(,(\d+))?")
+    hunk = None
+    while True:
+        line = stream.readline()
+        if not line:
+            break
+
+        if hunk is None:
+            # Parse file names
+            diff_file = file_regex.match(line)
+            if diff_file:
+              if line.startswith("---"):
+                  a_line = line
+                  a = diff_file.group(2)
+                  continue
+              if line.startswith("+++"):
+                  b_line = line
+                  b = diff_file.group(2)
+                  continue
+
+            # Parse offset/lengths
+            diffrange = range_regex.match(line)
+            if diffrange:
+                if diffrange.group(2):
+                    start_a = int(diffrange.group(1))
+                    len_a = int(diffrange.group(3))
+                else:
+                    start_a = 1
+                    len_a = int(diffrange.group(1))
+
+                if diffrange.group(5):
+                    start_b = int(diffrange.group(4))
+                    len_b = int(diffrange.group(6))
+                else:
+                    start_b = 1
+                    len_b = int(diffrange.group(4))
+
+                header = [a_line, b_line, line]
+                hunk = DiffHunk(header, a, b, start_a, len_a, start_b, len_b)
+        else:
+            # Add the current line to the hunk
+            hunk.Append(line)
+
+            # See if the whole hunk has been parsed. If so, yield it and prepare
+            # for the next hunk.
+            if hunk.Complete():
+                yield hunk
+                hunk = None
+
+    # Partial hunks are a parse error
+    assert hunk is None
diff --git a/libvpx/libvpx/tools/ftfy.sh b/libvpx/libvpx/tools/ftfy.sh
new file mode 100755
index 0000000..29ae95e
--- /dev/null
+++ b/libvpx/libvpx/tools/ftfy.sh
@@ -0,0 +1,159 @@
+#!/bin/sh
+self="$0"
+dirname_self=$(dirname "$self")
+
+usage() {
+  cat <<EOF >&2
+Usage: $self [option]
+
+This script applies a whitespace transformation to the commit at HEAD. If no
+options are given, then the modified files are left in the working tree.
+
+Options:
+  -h, --help     Shows this message
+  -n, --dry-run  Shows a diff of the changes to be made.
+  --amend        Squashes the changes into the commit at HEAD
+                     This option will also reformat the commit message.
+  --commit       Creates a new commit containing only the whitespace changes
+  --msg-only     Reformat the commit message only, ignore the patch itself.
+
+EOF
+  rm -f ${CLEAN_FILES}
+  exit 1
+}
+
+
+log() {
+  echo "${self##*/}: $@" >&2
+}
+
+
+vpx_style() {
+  for f; do
+    case "$f" in
+      *.h|*.c|*.cc)
+        "${dirname_self}"/vpx-astyle.sh "$f"
+        ;;
+    esac
+  done
+}
+
+
+apply() {
+  [ $INTERSECT_RESULT -ne 0 ] && patch -p1 < "$1"
+}
+
+
+commit() {
+  LAST_CHANGEID=$(git show | awk '/Change-Id:/{print $2}')
+  if [ -z "$LAST_CHANGEID" ]; then
+    log "HEAD doesn't have a Change-Id, unable to generate a new commit"
+    exit 1
+  fi
+
+  # Build a deterministic Change-Id from the parent's
+  NEW_CHANGEID=${LAST_CHANGEID}-styled
+  NEW_CHANGEID=I$(echo $NEW_CHANGEID | git hash-object --stdin)
+
+  # Commit, preserving authorship from the parent commit.
+  git commit -a -C HEAD > /dev/null
+  git commit --amend -F- << EOF
+Cosmetic: Fix whitespace in change ${LAST_CHANGEID:0:9}
+
+Change-Id: ${NEW_CHANGEID}
+EOF
+}
+
+
+show_commit_msg_diff() {
+  if [ $DIFF_MSG_RESULT -ne 0 ]; then
+    log "Modified commit message:"
+    diff -u "$ORIG_COMMIT_MSG" "$NEW_COMMIT_MSG" | tail -n +3
+  fi
+}
+
+
+amend() {
+  show_commit_msg_diff
+  if [ $DIFF_MSG_RESULT -ne 0 ] || [ $INTERSECT_RESULT -ne 0 ]; then
+    git commit -a --amend -F "$NEW_COMMIT_MSG"
+  fi
+}
+
+
+diff_msg() {
+  git log -1 --format=%B > "$ORIG_COMMIT_MSG"
+  "${dirname_self}"/wrap-commit-msg.py \
+      < "$ORIG_COMMIT_MSG" > "$NEW_COMMIT_MSG"
+  cmp -s "$ORIG_COMMIT_MSG" "$NEW_COMMIT_MSG"
+  DIFF_MSG_RESULT=$?
+}
+
+
+# Temporary files
+ORIG_DIFF=orig.diff.$$
+MODIFIED_DIFF=modified.diff.$$
+FINAL_DIFF=final.diff.$$
+ORIG_COMMIT_MSG=orig.commit-msg.$$
+NEW_COMMIT_MSG=new.commit-msg.$$
+CLEAN_FILES="${ORIG_DIFF} ${MODIFIED_DIFF} ${FINAL_DIFF}"
+CLEAN_FILES="${CLEAN_FILES} ${ORIG_COMMIT_MSG} ${NEW_COMMIT_MSG}"
+
+# Preconditions
+[ $# -lt 2 ] || usage
+
+# Check that astyle supports pad-header and align-pointer=name
+if ! astyle --pad-header --align-pointer=name < /dev/null; then
+  log "Install astyle v1.24 or newer"
+  exit 1
+fi
+
+if ! git diff --quiet HEAD; then
+  log "Working tree is dirty, commit your changes first"
+  exit 1
+fi
+
+# Need to be in the root
+cd "$(git rev-parse --show-toplevel)"
+
+# Collect the original diff
+git show > "${ORIG_DIFF}"
+
+# Apply the style guide on new and modified files and collect its diff
+for f in $(git diff HEAD^ --name-only -M90 --diff-filter=AM); do
+  case "$f" in
+    third_party/*) continue;;
+  esac
+  vpx_style "$f"
+done
+git diff --no-color --no-ext-diff > "${MODIFIED_DIFF}"
+
+# Intersect the two diffs
+"${dirname_self}"/intersect-diffs.py \
+    "${ORIG_DIFF}" "${MODIFIED_DIFF}" > "${FINAL_DIFF}"
+INTERSECT_RESULT=$?
+git reset --hard >/dev/null
+
+# Fixup the commit message
+diff_msg
+
+# Handle options
+if [ -n "$1" ]; then
+  case "$1" in
+    -h|--help) usage;;
+    -n|--dry-run) cat "${FINAL_DIFF}"; show_commit_msg_diff;;
+    --commit) apply "${FINAL_DIFF}"; commit;;
+    --amend) apply "${FINAL_DIFF}"; amend;;
+    --msg-only) amend;;
+    *) usage;;
+  esac
+else
+  apply "${FINAL_DIFF}"
+  if ! git diff --quiet; then
+    log "Formatting changes applied, verify and commit."
+    log "See also: http://www.webmproject.org/code/contribute/conventions/"
+    git diff --stat
+  fi
+fi
+
+rm -f ${CLEAN_FILES}
diff --git a/libvpx/libvpx/tools/gen_authors.sh b/libvpx/libvpx/tools/gen_authors.sh
new file mode 100755
index 0000000..4cfd81e
--- /dev/null
+++ b/libvpx/libvpx/tools/gen_authors.sh
@@ -0,0 +1,13 @@
+#!/bin/bash
+
+# Add organization names manually.
+
+cat <<EOF
+# This file is automatically generated from the git commit history
+# by tools/gen_authors.sh.
+
+$(git log --pretty=format:"%aN <%aE>" | sort | uniq | grep -v corp.google)
+Google Inc.
+The Mozilla Foundation
+The Xiph.Org Foundation
+EOF
diff --git a/libvpx/libvpx/tools/intersect-diffs.py b/libvpx/libvpx/tools/intersect-diffs.py
new file mode 100755
index 0000000..4dbafa9
--- /dev/null
+++ b/libvpx/libvpx/tools/intersect-diffs.py
@@ -0,0 +1,76 @@
+#!/usr/bin/env python
+##  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+"""Calculates the "intersection" of two unified diffs.
+
+Given two diffs, A and B, it finds all hunks in B that had non-context lines
+in A and prints them to stdout. This is useful to determine the hunks in B that
+are relevant to A. The resulting file can be applied with patch(1) on top of A.
+"""
+
+__author__ = "jkoleszar@google.com"
+
+import sys
+
+import diff
+
+
+def FormatDiffHunks(hunks):
+    """Re-serialize a list of DiffHunks."""
+    r = []
+    last_header = None
+    for hunk in hunks:
+        this_header = hunk.header[0:2]
+        if last_header != this_header:
+            r.extend(hunk.header)
+            last_header = this_header
+        else:
+            r.extend(hunk.header[2])
+        r.extend(hunk.lines)
+        r.append("\n")
+    return "".join(r)
+
+
+def ZipHunks(rhs_hunks, lhs_hunks):
+    """Join two hunk lists on filename."""
+    for rhs_hunk in rhs_hunks:
+        rhs_file = rhs_hunk.right.filename.split("/")[1:]
+
+        for lhs_hunk in lhs_hunks:
+            lhs_file = lhs_hunk.left.filename.split("/")[1:]
+            if lhs_file != rhs_file:
+                continue
+            yield (rhs_hunk, lhs_hunk)
+
+
+def main():
+    old_hunks = [x for x in diff.ParseDiffHunks(open(sys.argv[1], "r"))]
+    new_hunks = [x for x in diff.ParseDiffHunks(open(sys.argv[2], "r"))]
+    out_hunks = []
+
+    # Join the right hand side of the older diff with the left hand side of the
+    # newer diff.
+    for old_hunk, new_hunk in ZipHunks(old_hunks, new_hunks):
+        if new_hunk in out_hunks:
+            continue
+        old_lines = old_hunk.right
+        new_lines = new_hunk.left
+
+        # Determine if this hunk overlaps any non-context line from the other
+        for i in old_lines.delta_line_nums:
+            if i in new_lines:
+                out_hunks.append(new_hunk)
+                break
+
+    if out_hunks:
+        print FormatDiffHunks(out_hunks)
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/libvpx/libvpx/tools/lint-hunks.py b/libvpx/libvpx/tools/lint-hunks.py
new file mode 100755
index 0000000..6e25d93
--- /dev/null
+++ b/libvpx/libvpx/tools/lint-hunks.py
@@ -0,0 +1,144 @@
+#!/usr/bin/python
+##  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+"""Performs style checking on each diff hunk."""
+import getopt
+import os
+import StringIO
+import subprocess
+import sys
+
+import diff
+
+
+SHORT_OPTIONS = "h"
+LONG_OPTIONS = ["help"]
+
+TOPLEVEL_CMD = ["git", "rev-parse", "--show-toplevel"]
+DIFF_CMD = ["git", "diff"]
+DIFF_INDEX_CMD = ["git", "diff-index", "-u", "HEAD", "--"]
+SHOW_CMD = ["git", "show"]
+CPPLINT_FILTERS = ["-readability/casting"]
+
+
+class Usage(Exception):
+    pass
+
+
+class SubprocessException(Exception):
+    def __init__(self, args):
+        msg = "Failed to execute '%s'"%(" ".join(args))
+        super(SubprocessException, self).__init__(msg)
+
+
+class Subprocess(subprocess.Popen):
+    """Adds the notion of an expected returncode to Popen."""
+
+    def __init__(self, args, expected_returncode=0, **kwargs):
+        self._args = args
+        self._expected_returncode = expected_returncode
+        super(Subprocess, self).__init__(args, **kwargs)
+
+    def communicate(self, *args, **kwargs):
+        result = super(Subprocess, self).communicate(*args, **kwargs)
+        if self._expected_returncode is not None:
+            try:
+                ok = self.returncode in self._expected_returncode
+            except TypeError:
+                ok = self.returncode == self._expected_returncode
+            if not ok:
+                raise SubprocessException(self._args)
+        return result
+
+
+def main(argv=None):
+    if argv is None:
+        argv = sys.argv
+    try:
+        try:
+            opts, args = getopt.getopt(argv[1:], SHORT_OPTIONS, LONG_OPTIONS)
+        except getopt.error, msg:
+            raise Usage(msg)
+
+        # process options
+        for o, _ in opts:
+            if o in ("-h", "--help"):
+                print __doc__
+                sys.exit(0)
+
+        if args and len(args) > 1:
+            print __doc__
+            sys.exit(0)
+
+        # Find the fully qualified path to the root of the tree
+        tl = Subprocess(TOPLEVEL_CMD, stdout=subprocess.PIPE)
+        tl = tl.communicate()[0].strip()
+
+        # See if we're working on the index or not.
+        if args:
+            diff_cmd = DIFF_CMD + [args[0] + "^!"]
+        else:
+            diff_cmd = DIFF_INDEX_CMD
+
+        # Build the command line to execute cpplint
+        cpplint_cmd = [os.path.join(tl, "tools", "cpplint.py"),
+                       "--filter=" + ",".join(CPPLINT_FILTERS),
+                       "-"]
+
+        # Get a list of all affected lines
+        file_affected_line_map = {}
+        p = Subprocess(diff_cmd, stdout=subprocess.PIPE)
+        stdout = p.communicate()[0]
+        for hunk in diff.ParseDiffHunks(StringIO.StringIO(stdout)):
+            filename = hunk.right.filename[2:]
+            if filename not in file_affected_line_map:
+                file_affected_line_map[filename] = set()
+            file_affected_line_map[filename].update(hunk.right.delta_line_nums)
+
+        # Run each affected file through cpplint
+        lint_failed = False
+        for filename, affected_lines in file_affected_line_map.iteritems():
+            if filename.split(".")[-1] not in ("c", "h", "cc"):
+                continue
+
+            if args:
+                # File contents come from git
+                show_cmd = SHOW_CMD + [args[0] + ":" + filename]
+                show = Subprocess(show_cmd, stdout=subprocess.PIPE)
+                lint = Subprocess(cpplint_cmd, expected_returncode=(0, 1),
+                                  stdin=show.stdout, stderr=subprocess.PIPE)
+                lint_out = lint.communicate()[1]
+            else:
+                # File contents come from the working tree
+                lint = Subprocess(cpplint_cmd, expected_returncode=(0, 1),
+                                  stdin=subprocess.PIPE, stderr=subprocess.PIPE)
+                stdin = open(os.path.join(tl, filename)).read()
+                lint_out = lint.communicate(stdin)[1]
+
+            for line in lint_out.split("\n"):
+                fields = line.split(":")
+                if fields[0] != "-":
+                    continue
+                warning_line_num = int(fields[1])
+                if warning_line_num in affected_lines:
+                    print "%s:%d:%s"%(filename, warning_line_num,
+                                      ":".join(fields[2:]))
+                    lint_failed = True
+
+        # Set exit code if any relevant lint errors seen
+        if lint_failed:
+            return 1
+
+    except Usage, err:
+        print >>sys.stderr, err
+        print >>sys.stderr, "for help use --help"
+        return 2
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/libvpx/libvpx/tools/vpx-astyle.sh b/libvpx/libvpx/tools/vpx-astyle.sh
new file mode 100755
index 0000000..6340426
--- /dev/null
+++ b/libvpx/libvpx/tools/vpx-astyle.sh
@@ -0,0 +1,27 @@
+#!/bin/sh
+set -e
+astyle --style=java --indent=spaces=2 --indent-switches\
+       --min-conditional-indent=0 \
+       --pad-oper --pad-header --unpad-paren \
+       --align-pointer=name \
+       --indent-preprocessor --convert-tabs --indent-labels \
+       --suffix=none --quiet --max-instatement-indent=80 "$@"
+# Disabled, too greedy?
+#sed -i 's;[[:space:]]\{1,\}\[;[;g' "$@"
+
+sed_i() {
+  # Incompatible sed parameter parsing.
+  if sed -i 2>&1 | grep -q 'requires an argument'; then
+    sed -i '' "$@"
+  else
+    sed -i "$@"
+  fi
+}
+
+sed_i -e 's/[[:space:]]\{1,\}\([,;]\)/\1/g' \
+      -e 's/[[:space:]]\{1,\}\([+-]\{2\};\)/\1/g' \
+      -e 's/,[[:space:]]*}/}/g' \
+      -e 's;//\([^/[:space:]].*$\);// \1;g' \
+      -e 's/^\(public\|private\|protected\):$/ \1:/g' \
+      -e 's/[[:space:]]\{1,\}$//g' \
+      "$@"
diff --git a/libvpx/libvpx/tools/wrap-commit-msg.py b/libvpx/libvpx/tools/wrap-commit-msg.py
new file mode 100755
index 0000000..d5b4b04
--- /dev/null
+++ b/libvpx/libvpx/tools/wrap-commit-msg.py
@@ -0,0 +1,70 @@
+#!/usr/bin/env python
+##  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+"""Wraps paragraphs of text, preserving manual formatting
+
+This is like fold(1), but has the special convention of not modifying lines
+that start with whitespace. This allows you to intersperse blocks with
+special formatting, like code blocks, with written prose. The prose will
+be wordwrapped, and the manual formatting will be preserved.
+
+ * This won't handle the case of a bulleted (or ordered) list specially, so
+   manual wrapping must be done.
+
+Occasionally it's useful to put something with explicit formatting that
+doesn't look at all like a block of text inline.
+
+  indicator = has_leading_whitespace(line);
+  if (indicator)
+    preserve_formatting(line);
+
+The intent is that this docstring would make it through the transform
+and still be legible and presented as it is in the source. If additional
+cases are handled, update this doc to describe the effect.
+"""
+
+__author__ = "jkoleszar@google.com"
+import textwrap
+import sys
+
+def wrap(text):
+    if text:
+        return textwrap.fill(text, break_long_words=False) + '\n'
+    return ""
+
+
+def main(fileobj):
+    text = ""
+    output = ""
+    while True:
+        line = fileobj.readline()
+        if not line:
+            break
+
+        if line.lstrip() == line:
+            text += line
+        else:
+            output += wrap(text)
+            text=""
+            output += line
+    output += wrap(text)
+
+    # Replace the file or write to stdout.
+    if fileobj == sys.stdin:
+        fileobj = sys.stdout
+    else:
+        fileobj.seek(0)
+        fileobj.truncate(0)
+    fileobj.write(output)
+
+if __name__ == "__main__":
+    if len(sys.argv) > 1:
+        main(open(sys.argv[1], "r+"))
+    else:
+        main(sys.stdin)
diff --git a/libvpx/libvpx/tools_common.c b/libvpx/libvpx/tools_common.c
new file mode 100644
index 0000000..17c0d44
--- /dev/null
+++ b/libvpx/libvpx/tools_common.c
@@ -0,0 +1,494 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "./tools_common.h"
+
+#if CONFIG_VP8_ENCODER || CONFIG_VP9_ENCODER
+#include "vpx/vp8cx.h"
+#endif
+
+#if CONFIG_VP8_DECODER || CONFIG_VP9_DECODER
+#include "vpx/vp8dx.h"
+#endif
+
+#if defined(_WIN32) || defined(__OS2__)
+#include <io.h>
+#include <fcntl.h>
+
+#ifdef __OS2__
+#define _setmode    setmode
+#define _fileno     fileno
+#define _O_BINARY   O_BINARY
+#endif
+#endif
+
+#define LOG_ERROR(label) do {\
+  const char *l = label;\
+  va_list ap;\
+  va_start(ap, fmt);\
+  if (l)\
+    fprintf(stderr, "%s: ", l);\
+  vfprintf(stderr, fmt, ap);\
+  fprintf(stderr, "\n");\
+  va_end(ap);\
+} while (0)
+
+
+FILE *set_binary_mode(FILE *stream) {
+  (void)stream;
+#if defined(_WIN32) || defined(__OS2__)
+  _setmode(_fileno(stream), _O_BINARY);
+#endif
+  return stream;
+}
+
+void die(const char *fmt, ...) {
+  LOG_ERROR(NULL);
+  usage_exit();
+}
+
+void fatal(const char *fmt, ...) {
+  LOG_ERROR("Fatal");
+  exit(EXIT_FAILURE);
+}
+
+void warn(const char *fmt, ...) {
+  LOG_ERROR("Warning");
+}
+
+void die_codec(vpx_codec_ctx_t *ctx, const char *s) {
+  const char *detail = vpx_codec_error_detail(ctx);
+
+  printf("%s: %s\n", s, vpx_codec_error(ctx));
+  if (detail)
+    printf("    %s\n", detail);
+  exit(EXIT_FAILURE);
+}
+
+int read_yuv_frame(struct VpxInputContext *input_ctx, vpx_image_t *yuv_frame) {
+  FILE *f = input_ctx->file;
+  struct FileTypeDetectionBuffer *detect = &input_ctx->detect;
+  int plane = 0;
+  int shortread = 0;
+  const int bytespp = (yuv_frame->fmt & VPX_IMG_FMT_HIGHBITDEPTH) ? 2 : 1;
+
+  for (plane = 0; plane < 3; ++plane) {
+    uint8_t *ptr;
+    const int w = vpx_img_plane_width(yuv_frame, plane);
+    const int h = vpx_img_plane_height(yuv_frame, plane);
+    int r;
+
+    /* Determine the correct plane based on the image format. The for-loop
+     * always counts in Y,U,V order, but this may not match the order of
+     * the data on disk.
+     */
+    switch (plane) {
+      case 1:
+        ptr = yuv_frame->planes[
+            yuv_frame->fmt == VPX_IMG_FMT_YV12 ? VPX_PLANE_V : VPX_PLANE_U];
+        break;
+      case 2:
+        ptr = yuv_frame->planes[
+            yuv_frame->fmt == VPX_IMG_FMT_YV12 ? VPX_PLANE_U : VPX_PLANE_V];
+        break;
+      default:
+        ptr = yuv_frame->planes[plane];
+    }
+
+    for (r = 0; r < h; ++r) {
+      size_t needed = w * bytespp;
+      size_t buf_position = 0;
+      const size_t left = detect->buf_read - detect->position;
+      if (left > 0) {
+        const size_t more = (left < needed) ? left : needed;
+        memcpy(ptr, detect->buf + detect->position, more);
+        buf_position = more;
+        needed -= more;
+        detect->position += more;
+      }
+      if (needed > 0) {
+        shortread |= (fread(ptr + buf_position, 1, needed, f) < needed);
+      }
+
+      ptr += yuv_frame->stride[plane];
+    }
+  }
+
+  return shortread;
+}
+
+#if CONFIG_ENCODERS
+
+static const VpxInterface vpx_encoders[] = {
+#if CONFIG_VP8_ENCODER
+  {"vp8", VP8_FOURCC, &vpx_codec_vp8_cx},
+#endif
+
+#if CONFIG_VP9_ENCODER
+  {"vp9", VP9_FOURCC, &vpx_codec_vp9_cx},
+#endif
+};
+
+int get_vpx_encoder_count(void) {
+  return sizeof(vpx_encoders) / sizeof(vpx_encoders[0]);
+}
+
+const VpxInterface *get_vpx_encoder_by_index(int i) {
+  return &vpx_encoders[i];
+}
+
+const VpxInterface *get_vpx_encoder_by_name(const char *name) {
+  int i;
+
+  for (i = 0; i < get_vpx_encoder_count(); ++i) {
+    const VpxInterface *encoder = get_vpx_encoder_by_index(i);
+    if (strcmp(encoder->name, name) == 0)
+      return encoder;
+  }
+
+  return NULL;
+}
+
+#endif  // CONFIG_ENCODERS
+
+#if CONFIG_DECODERS
+
+static const VpxInterface vpx_decoders[] = {
+#if CONFIG_VP8_DECODER
+  {"vp8", VP8_FOURCC, &vpx_codec_vp8_dx},
+#endif
+
+#if CONFIG_VP9_DECODER
+  {"vp9", VP9_FOURCC, &vpx_codec_vp9_dx},
+#endif
+};
+
+int get_vpx_decoder_count(void) {
+  return sizeof(vpx_decoders) / sizeof(vpx_decoders[0]);
+}
+
+const VpxInterface *get_vpx_decoder_by_index(int i) {
+  return &vpx_decoders[i];
+}
+
+const VpxInterface *get_vpx_decoder_by_name(const char *name) {
+  int i;
+
+  for (i = 0; i < get_vpx_decoder_count(); ++i) {
+     const VpxInterface *const decoder = get_vpx_decoder_by_index(i);
+     if (strcmp(decoder->name, name) == 0)
+       return decoder;
+  }
+
+  return NULL;
+}
+
+const VpxInterface *get_vpx_decoder_by_fourcc(uint32_t fourcc) {
+  int i;
+
+  for (i = 0; i < get_vpx_decoder_count(); ++i) {
+    const VpxInterface *const decoder = get_vpx_decoder_by_index(i);
+    if (decoder->fourcc == fourcc)
+      return decoder;
+  }
+
+  return NULL;
+}
+
+#endif  // CONFIG_DECODERS
+
+// TODO(dkovalev): move this function to vpx_image.{c, h}, so it will be part
+// of vpx_image_t support
+int vpx_img_plane_width(const vpx_image_t *img, int plane) {
+  if (plane > 0 && img->x_chroma_shift > 0)
+    return (img->d_w + 1) >> img->x_chroma_shift;
+  else
+    return img->d_w;
+}
+
+int vpx_img_plane_height(const vpx_image_t *img, int plane) {
+  if (plane > 0 &&  img->y_chroma_shift > 0)
+    return (img->d_h + 1) >> img->y_chroma_shift;
+  else
+    return img->d_h;
+}
+
+void vpx_img_write(const vpx_image_t *img, FILE *file) {
+  int plane;
+
+  for (plane = 0; plane < 3; ++plane) {
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int w = vpx_img_plane_width(img, plane) *
+        ((img->fmt & VPX_IMG_FMT_HIGHBITDEPTH) ? 2 : 1);
+    const int h = vpx_img_plane_height(img, plane);
+    int y;
+
+    for (y = 0; y < h; ++y) {
+      fwrite(buf, 1, w, file);
+      buf += stride;
+    }
+  }
+}
+
+int vpx_img_read(vpx_image_t *img, FILE *file) {
+  int plane;
+
+  for (plane = 0; plane < 3; ++plane) {
+    unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int w = vpx_img_plane_width(img, plane) *
+        ((img->fmt & VPX_IMG_FMT_HIGHBITDEPTH) ? 2 : 1);
+    const int h = vpx_img_plane_height(img, plane);
+    int y;
+
+    for (y = 0; y < h; ++y) {
+      if (fread(buf, 1, w, file) != (size_t)w)
+        return 0;
+      buf += stride;
+    }
+  }
+
+  return 1;
+}
+
+// TODO(dkovalev) change sse_to_psnr signature: double -> int64_t
+double sse_to_psnr(double samples, double peak, double sse) {
+  static const double kMaxPSNR = 100.0;
+
+  if (sse > 0.0) {
+    const double psnr = 10.0 * log10(samples * peak * peak / sse);
+    return psnr > kMaxPSNR ? kMaxPSNR : psnr;
+  } else {
+    return kMaxPSNR;
+  }
+}
+
+// TODO(debargha): Consolidate the functions below into a separate file.
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_img_upshift(vpx_image_t *dst, vpx_image_t *src,
+                               int input_shift) {
+  // Note the offset is 1 less than half.
+  const int offset = input_shift > 0 ? (1 << (input_shift - 1)) - 1 : 0;
+  int plane;
+  if (dst->d_w != src->d_w || dst->d_h != src->d_h ||
+      dst->x_chroma_shift != src->x_chroma_shift ||
+      dst->y_chroma_shift != src->y_chroma_shift ||
+      dst->fmt != src->fmt || input_shift < 0) {
+    fatal("Unsupported image conversion");
+  }
+  switch (src->fmt) {
+    case VPX_IMG_FMT_I42016:
+    case VPX_IMG_FMT_I42216:
+    case VPX_IMG_FMT_I44416:
+    case VPX_IMG_FMT_I44016:
+      break;
+    default:
+      fatal("Unsupported image conversion");
+      break;
+  }
+  for (plane = 0; plane < 3; plane++) {
+    int w = src->d_w;
+    int h = src->d_h;
+    int x, y;
+    if (plane) {
+      w = (w + src->x_chroma_shift) >> src->x_chroma_shift;
+      h = (h + src->y_chroma_shift) >> src->y_chroma_shift;
+    }
+    for (y = 0; y < h; y++) {
+      uint16_t *p_src =
+          (uint16_t *)(src->planes[plane] + y * src->stride[plane]);
+      uint16_t *p_dst =
+          (uint16_t *)(dst->planes[plane] + y * dst->stride[plane]);
+      for (x = 0; x < w; x++)
+        *p_dst++ = (*p_src++ << input_shift) + offset;
+    }
+  }
+}
+
+static void lowbd_img_upshift(vpx_image_t *dst, vpx_image_t *src,
+                              int input_shift) {
+  // Note the offset is 1 less than half.
+  const int offset = input_shift > 0 ? (1 << (input_shift - 1)) - 1 : 0;
+  int plane;
+  if (dst->d_w != src->d_w || dst->d_h != src->d_h ||
+      dst->x_chroma_shift != src->x_chroma_shift ||
+      dst->y_chroma_shift != src->y_chroma_shift ||
+      dst->fmt != src->fmt + VPX_IMG_FMT_HIGHBITDEPTH ||
+      input_shift < 0) {
+    fatal("Unsupported image conversion");
+  }
+  switch (src->fmt) {
+    case VPX_IMG_FMT_I420:
+    case VPX_IMG_FMT_I422:
+    case VPX_IMG_FMT_I444:
+    case VPX_IMG_FMT_I440:
+      break;
+    default:
+      fatal("Unsupported image conversion");
+      break;
+  }
+  for (plane = 0; plane < 3; plane++) {
+    int w = src->d_w;
+    int h = src->d_h;
+    int x, y;
+    if (plane) {
+      w = (w + src->x_chroma_shift) >> src->x_chroma_shift;
+      h = (h + src->y_chroma_shift) >> src->y_chroma_shift;
+    }
+    for (y = 0; y < h; y++) {
+      uint8_t *p_src = src->planes[plane] + y * src->stride[plane];
+      uint16_t *p_dst =
+          (uint16_t *)(dst->planes[plane] + y * dst->stride[plane]);
+      for (x = 0; x < w; x++) {
+        *p_dst++ = (*p_src++ << input_shift) + offset;
+      }
+    }
+  }
+}
+
+void vpx_img_upshift(vpx_image_t *dst, vpx_image_t *src,
+                     int input_shift) {
+  if (src->fmt & VPX_IMG_FMT_HIGHBITDEPTH) {
+    highbd_img_upshift(dst, src, input_shift);
+  } else {
+    lowbd_img_upshift(dst, src, input_shift);
+  }
+}
+
+void vpx_img_truncate_16_to_8(vpx_image_t *dst, vpx_image_t *src) {
+  int plane;
+  if (dst->fmt + VPX_IMG_FMT_HIGHBITDEPTH != src->fmt ||
+      dst->d_w != src->d_w || dst->d_h != src->d_h ||
+      dst->x_chroma_shift != src->x_chroma_shift ||
+      dst->y_chroma_shift != src->y_chroma_shift) {
+    fatal("Unsupported image conversion");
+  }
+  switch (dst->fmt) {
+    case VPX_IMG_FMT_I420:
+    case VPX_IMG_FMT_I422:
+    case VPX_IMG_FMT_I444:
+    case VPX_IMG_FMT_I440:
+      break;
+    default:
+      fatal("Unsupported image conversion");
+      break;
+  }
+  for (plane = 0; plane < 3; plane++) {
+    int w = src->d_w;
+    int h = src->d_h;
+    int x, y;
+    if (plane) {
+      w = (w + src->x_chroma_shift) >> src->x_chroma_shift;
+      h = (h + src->y_chroma_shift) >> src->y_chroma_shift;
+    }
+    for (y = 0; y < h; y++) {
+      uint16_t *p_src =
+          (uint16_t *)(src->planes[plane] + y * src->stride[plane]);
+      uint8_t *p_dst = dst->planes[plane] + y * dst->stride[plane];
+      for (x = 0; x < w; x++) {
+        *p_dst++ = (uint8_t)(*p_src++);
+      }
+    }
+  }
+}
+
+static void highbd_img_downshift(vpx_image_t *dst, vpx_image_t *src,
+                                 int down_shift) {
+  int plane;
+  if (dst->d_w != src->d_w || dst->d_h != src->d_h ||
+      dst->x_chroma_shift != src->x_chroma_shift ||
+      dst->y_chroma_shift != src->y_chroma_shift ||
+      dst->fmt != src->fmt || down_shift < 0) {
+    fatal("Unsupported image conversion");
+  }
+  switch (src->fmt) {
+    case VPX_IMG_FMT_I42016:
+    case VPX_IMG_FMT_I42216:
+    case VPX_IMG_FMT_I44416:
+    case VPX_IMG_FMT_I44016:
+      break;
+    default:
+      fatal("Unsupported image conversion");
+      break;
+  }
+  for (plane = 0; plane < 3; plane++) {
+    int w = src->d_w;
+    int h = src->d_h;
+    int x, y;
+    if (plane) {
+      w = (w + src->x_chroma_shift) >> src->x_chroma_shift;
+      h = (h + src->y_chroma_shift) >> src->y_chroma_shift;
+    }
+    for (y = 0; y < h; y++) {
+      uint16_t *p_src =
+          (uint16_t *)(src->planes[plane] + y * src->stride[plane]);
+      uint16_t *p_dst =
+          (uint16_t *)(dst->planes[plane] + y * dst->stride[plane]);
+      for (x = 0; x < w; x++)
+        *p_dst++ = *p_src++ >> down_shift;
+    }
+  }
+}
+
+static void lowbd_img_downshift(vpx_image_t *dst, vpx_image_t *src,
+                                int down_shift) {
+  int plane;
+  if (dst->d_w != src->d_w || dst->d_h != src->d_h ||
+      dst->x_chroma_shift != src->x_chroma_shift ||
+      dst->y_chroma_shift != src->y_chroma_shift ||
+      src->fmt != dst->fmt + VPX_IMG_FMT_HIGHBITDEPTH ||
+      down_shift < 0) {
+    fatal("Unsupported image conversion");
+  }
+  switch (dst->fmt) {
+    case VPX_IMG_FMT_I420:
+    case VPX_IMG_FMT_I422:
+    case VPX_IMG_FMT_I444:
+    case VPX_IMG_FMT_I440:
+      break;
+    default:
+      fatal("Unsupported image conversion");
+      break;
+  }
+  for (plane = 0; plane < 3; plane++) {
+    int w = src->d_w;
+    int h = src->d_h;
+    int x, y;
+    if (plane) {
+      w = (w + src->x_chroma_shift) >> src->x_chroma_shift;
+      h = (h + src->y_chroma_shift) >> src->y_chroma_shift;
+    }
+    for (y = 0; y < h; y++) {
+      uint16_t *p_src =
+          (uint16_t *)(src->planes[plane] + y * src->stride[plane]);
+      uint8_t *p_dst = dst->planes[plane] + y * dst->stride[plane];
+      for (x = 0; x < w; x++) {
+        *p_dst++ = *p_src++ >> down_shift;
+      }
+    }
+  }
+}
+
+void vpx_img_downshift(vpx_image_t *dst, vpx_image_t *src,
+                       int down_shift) {
+  if (dst->fmt & VPX_IMG_FMT_HIGHBITDEPTH) {
+    highbd_img_downshift(dst, src, down_shift);
+  } else {
+    lowbd_img_downshift(dst, src, down_shift);
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/tools_common.h b/libvpx/libvpx/tools_common.h
new file mode 100644
index 0000000..310b569
--- /dev/null
+++ b/libvpx/libvpx/tools_common.h
@@ -0,0 +1,163 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef TOOLS_COMMON_H_
+#define TOOLS_COMMON_H_
+
+#include <stdio.h>
+
+#include "./vpx_config.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_image.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/msvc.h"
+
+#if CONFIG_ENCODERS
+#include "./y4minput.h"
+#endif
+
+#if defined(_MSC_VER)
+/* MSVS uses _f{seek,tell}i64. */
+#define fseeko _fseeki64
+#define ftello _ftelli64
+#elif defined(_WIN32)
+/* MinGW uses f{seek,tell}o64 for large files. */
+#define fseeko fseeko64
+#define ftello ftello64
+#endif  /* _WIN32 */
+
+#if CONFIG_OS_SUPPORT
+#if defined(_MSC_VER)
+#include <io.h>  /* NOLINT */
+#define isatty   _isatty
+#define fileno   _fileno
+#else
+#include <unistd.h>  /* NOLINT */
+#endif  /* _MSC_VER */
+#endif  /* CONFIG_OS_SUPPORT */
+
+/* Use 32-bit file operations in WebM file format when building ARM
+ * executables (.axf) with RVCT. */
+#if !CONFIG_OS_SUPPORT
+#define fseeko fseek
+#define ftello ftell
+#endif  /* CONFIG_OS_SUPPORT */
+
+#define LITERALU64(hi, lo) ((((uint64_t)hi) << 32) | lo)
+
+#ifndef PATH_MAX
+#define PATH_MAX 512
+#endif
+
+#define IVF_FRAME_HDR_SZ (4 + 8)  /* 4 byte size + 8 byte timestamp */
+#define IVF_FILE_HDR_SZ 32
+
+#define RAW_FRAME_HDR_SZ sizeof(uint32_t)
+
+#define VP8_FOURCC 0x30385056
+#define VP9_FOURCC 0x30395056
+
+enum VideoFileType {
+  FILE_TYPE_RAW,
+  FILE_TYPE_IVF,
+  FILE_TYPE_Y4M,
+  FILE_TYPE_WEBM
+};
+
+struct FileTypeDetectionBuffer {
+  char buf[4];
+  size_t buf_read;
+  size_t position;
+};
+
+struct VpxRational {
+  int numerator;
+  int denominator;
+};
+
+struct VpxInputContext {
+  const char *filename;
+  FILE *file;
+  int64_t length;
+  struct FileTypeDetectionBuffer detect;
+  enum VideoFileType file_type;
+  uint32_t width;
+  uint32_t height;
+  struct VpxRational pixel_aspect_ratio;
+  vpx_img_fmt_t fmt;
+  vpx_bit_depth_t bit_depth;
+  int only_i420;
+  uint32_t fourcc;
+  struct VpxRational framerate;
+#if CONFIG_ENCODERS
+  y4m_input y4m;
+#endif
+};
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(__GNUC__)
+#define VPX_NO_RETURN __attribute__((noreturn))
+#else
+#define VPX_NO_RETURN
+#endif
+
+/* Sets a stdio stream into binary mode */
+FILE *set_binary_mode(FILE *stream);
+
+void die(const char *fmt, ...) VPX_NO_RETURN;
+void fatal(const char *fmt, ...) VPX_NO_RETURN;
+void warn(const char *fmt, ...);
+
+void die_codec(vpx_codec_ctx_t *ctx, const char *s) VPX_NO_RETURN;
+
+/* The tool including this file must define usage_exit() */
+void usage_exit(void) VPX_NO_RETURN;
+
+#undef VPX_NO_RETURN
+
+int read_yuv_frame(struct VpxInputContext *input_ctx, vpx_image_t *yuv_frame);
+
+typedef struct VpxInterface {
+  const char *const name;
+  const uint32_t fourcc;
+  vpx_codec_iface_t *(*const codec_interface)();
+} VpxInterface;
+
+int get_vpx_encoder_count(void);
+const VpxInterface *get_vpx_encoder_by_index(int i);
+const VpxInterface *get_vpx_encoder_by_name(const char *name);
+
+int get_vpx_decoder_count(void);
+const VpxInterface *get_vpx_decoder_by_index(int i);
+const VpxInterface *get_vpx_decoder_by_name(const char *name);
+const VpxInterface *get_vpx_decoder_by_fourcc(uint32_t fourcc);
+
+// TODO(dkovalev): move this function to vpx_image.{c, h}, so it will be part
+// of vpx_image_t support
+int vpx_img_plane_width(const vpx_image_t *img, int plane);
+int vpx_img_plane_height(const vpx_image_t *img, int plane);
+void vpx_img_write(const vpx_image_t *img, FILE *file);
+int vpx_img_read(vpx_image_t *img, FILE *file);
+
+double sse_to_psnr(double samples, double peak, double mse);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_img_upshift(vpx_image_t *dst, vpx_image_t *src, int input_shift);
+void vpx_img_downshift(vpx_image_t *dst, vpx_image_t *src, int down_shift);
+void vpx_img_truncate_16_to_8(vpx_image_t *dst, vpx_image_t *src);
+#endif
+
+#ifdef __cplusplus
+}  /* extern "C" */
+#endif
+
+#endif  // TOOLS_COMMON_H_
diff --git a/libvpx/libvpx/usage.dox b/libvpx/libvpx/usage.dox
new file mode 100644
index 0000000..8823520
--- /dev/null
+++ b/libvpx/libvpx/usage.dox
@@ -0,0 +1,136 @@
+/*!\page usage Usage
+
+    The vpx multi-format codec SDK provides a unified interface amongst its
+    supported codecs. This abstraction allows applications using this SDK to
+    easily support multiple video formats with minimal code duplication or
+    "special casing." This section describes the interface common to all codecs.
+    For codec-specific details, see the \ref codecs page.
+
+    The following sections are common to all codecs:
+    - \ref usage_types
+    - \ref usage_features
+    - \ref usage_init
+    - \ref usage_errors
+
+    For more information on decoder and encoder specific usage, see the
+    following pages:
+    \if decoder
+    \li \subpage usage_decode
+    \endif
+    \if encoder
+    \li \subpage usage_encode
+    \endif
+
+    \section usage_types Important Data Types
+    There are two important data structures to consider in this interface.
+
+    \subsection usage_ctxs Contexts
+    A context is a storage area allocated by the calling application that the
+    codec may write into to store details about a single instance of that codec.
+    Most of the context is implementation specific, and thus opaque to the
+    application. The context structure as seen by the application is of fixed
+    size, and thus can be allocated with automatic storage or dynamically
+    on the heap.
+
+    Most operations require an initialized codec context. Codec context
+    instances are codec specific. That is, the codec to be used for the encoded
+    video must be known at initialization time. See #vpx_codec_ctx_t for further
+    information.
+
+    \subsection usage_ifaces Interfaces
+    A codec interface is an opaque structure that controls how function calls
+    into the generic interface are dispatched to their codec-specific
+    implementations. Applications \ref MUSTNOT attempt to examine or override
+    this storage, as it contains internal implementation details likely to
+    change from release to release.
+
+    Each supported codec will expose an interface structure to the application
+    as an <code>extern</code> reference to a structure of the incomplete type
+    #vpx_codec_iface_t.
+
+    \section usage_features Features
+    Several "features" are defined that are optionally implemented by codec
+    algorithms. Indeed, the same algorithm may support different features on
+    different platforms. The purpose of defining these features is that when
+    they are implemented, they conform to a common interface. The features, or
+    capabilities, of an algorithm can be queried from it's interface by using
+    the vpx_codec_get_caps() method. Attempts to invoke features not supported
+    by an algorithm will generally result in #VPX_CODEC_INCAPABLE.
+
+    \if decoder
+    Currently defined decoder features include:
+    - \ref usage_cb
+    - \ref usage_postproc
+    \endif
+
+    \section usage_init Initialization
+    To initialize a codec instance, the address of the codec context
+    and interface structures are passed to an initialization function. Depending
+    on the \ref usage_features that the codec supports, the codec could be
+    initialized in different modes.
+
+    To prevent cases of confusion where the ABI of the library changes,
+    the ABI is versioned. The ABI version number must be passed at
+    initialization time to ensure the application is using a header file that
+    matches the library. The current ABI version number is stored in the
+    preprocessor macros #VPX_CODEC_ABI_VERSION, #VPX_ENCODER_ABI_VERSION, and
+    #VPX_DECODER_ABI_VERSION. For convenience, each initialization function has
+    a wrapper macro that inserts the correct version number. These macros are
+    named like the initialization methods, but without the _ver suffix.
+
+
+    The available initialization methods are:
+    \if encoder
+    \li #vpx_codec_enc_init (calls vpx_codec_enc_init_ver())
+    \li #vpx_codec_enc_init_multi (calls vpx_codec_enc_init_multi_ver())
+    \endif
+    \if decoder
+    \li #vpx_codec_dec_init (calls vpx_codec_dec_init_ver())
+    \endif
+
+
+    \section usage_errors Error Handling
+    Almost all codec functions return an error status of type #vpx_codec_err_t.
+    The semantics of how each error condition should be processed is clearly
+    defined in the definitions of each enumerated value. Error values can be
+    converted into ASCII strings with the vpx_codec_error() and
+    vpx_codec_err_to_string() methods. The difference between these two methods is
+    that vpx_codec_error() returns the error state from an initialized context,
+    whereas vpx_codec_err_to_string() can be used in cases where an error occurs
+    outside any context. The enumerated value returned from the last call can be
+    retrieved from the <code>err</code> member of the decoder context as well.
+    Finally, more detailed error information may be able to be obtained by using
+    the vpx_codec_error_detail() method. Not all errors produce detailed error
+    information.
+
+    In addition to error information, the codec library's build configuration
+    is available at runtime on some platforms. This information can be returned
+    by calling vpx_codec_build_config(), and is formatted as a base64 coded string
+    (comprised of characters in the set [a-z_a-Z0-9+/]). This information is not
+    useful to an application at runtime, but may be of use to vpx for support.
+
+
+    \section usage_deadline Deadline
+    Both the encoding and decoding functions have a <code>deadline</code>
+    parameter. This parameter indicates the amount of time, in microseconds
+    (us), that the application wants the codec to spend processing before
+    returning. This is a soft deadline -- that is, the semantics of the
+    requested operation take precedence over meeting the deadline. If, for
+    example, an application sets a <code>deadline</code> of 1000us, and the
+    frame takes 2000us to decode, the call to vpx_codec_decode() will return
+    after 2000us. In this case the deadline is not met, but the semantics of the
+    function are preserved. If, for the same frame, an application instead sets
+    a <code>deadline</code> of 5000us, the decoder will see that it has 3000us
+    remaining in its time slice when decoding completes. It could then choose to
+    run a set of \ref usage_postproc filters, and perhaps would return after
+    4000us (instead of the allocated 5000us). In this case the deadline is met,
+    and the semantics of the call are preserved, as before.
+
+    The special value <code>0</code> is reserved to represent an infinite
+    deadline. In this case, the codec will perform as much processing as
+    possible to yield the highest quality frame.
+
+    By convention, the value <code>1</code> is used to mean "return as fast as
+    possible."
+
+*/
diff --git a/libvpx/libvpx/usage_cx.dox b/libvpx/libvpx/usage_cx.dox
new file mode 100644
index 0000000..92b0d34
--- /dev/null
+++ b/libvpx/libvpx/usage_cx.dox
@@ -0,0 +1,13 @@
+/*! \page usage_encode Encoding
+
+    The vpx_codec_encode() function is at the core of the encode loop. It
+    processes raw images passed by the application, producing packets of
+    compressed data. The <code>deadline</code> parameter controls the amount
+    of time in microseconds the encoder should spend working on the frame. For
+    more information on the <code>deadline</code> parameter, see
+    \ref usage_deadline.
+
+
+    \ref samples
+
+*/
diff --git a/libvpx/libvpx/usage_dx.dox b/libvpx/libvpx/usage_dx.dox
new file mode 100644
index 0000000..883ce24
--- /dev/null
+++ b/libvpx/libvpx/usage_dx.dox
@@ -0,0 +1,62 @@
+/*! \page usage_decode Decoding
+
+    The vpx_codec_decode() function is at the core of the decode loop. It
+    processes packets of compressed data passed by the application, producing
+    decoded images. The decoder expects packets to comprise exactly one image
+    frame of data. Packets \ref MUST be passed in decode order. If the
+    application wishes to associate some data with the frame, the
+    <code>user_priv</code> member may be set. The <code>deadline</code>
+    parameter controls the amount of time in microseconds the decoder should
+    spend working on the frame. This is typically used to support adaptive
+    \ref usage_postproc based on the amount of free CPU time. For more
+    information on the <code>deadline</code> parameter, see \ref usage_deadline.
+
+    \ref samples
+
+
+    \section usage_cb Callback Based Decoding
+    There are two methods for the application to access decoded frame data. Some
+    codecs support asynchronous (callback-based) decoding \ref usage_features
+    that allow the application to register a callback to be invoked by the
+    decoder when decoded data becomes available. Decoders are not required to
+    support this feature, however. Like all \ref usage_features, support can be
+    determined by calling vpx_codec_get_caps(). Callbacks are available in both
+    frame-based and slice-based variants. Frame based callbacks conform to the
+    signature of #vpx_codec_put_frame_cb_fn_t and are invoked once the entire
+    frame has been decoded. Slice based callbacks conform to the signature of
+    #vpx_codec_put_slice_cb_fn_t and are invoked after a subsection of the frame
+    is decoded. For example, a slice callback could be issued for each
+    macroblock row. However, the number and size of slices to return is
+    implementation specific. Also, the image data passed in a slice callback is
+    not necessarily in the same memory segment as the data will be when it is
+    assembled into a full frame. For this reason, the application \ref MUST
+    examine the rectangles that describe what data is valid to access and what
+    data has been updated in this call. For all their additional complexity,
+    slice based decoding callbacks provide substantial speed gains to the
+    overall application in some cases, due to improved cache behavior.
+
+
+    \section usage_frame_iter Frame Iterator Based Decoding
+    If the codec does not support callback based decoding, or the application
+    chooses not to make use of that feature, decoded frames are made available
+    through the vpx_codec_get_frame() iterator. The application initializes the
+    iterator storage (of type #vpx_codec_iter_t) to NULL, then calls
+    vpx_codec_get_frame repeatedly until it returns NULL, indicating that all
+    images have been returned. This process may result in zero, one, or many
+    frames that are ready for display, depending on the codec.
+
+
+    \section usage_postproc Postprocessing
+    Postprocessing is a process that is applied after a frame is decoded to
+    enhance the image's appearance by removing artifacts introduced in the
+    compression process. It is not required to properly decode the frame, and
+    is generally done only when there is enough spare CPU time to execute
+    the required filters. Codecs may support a number of different
+    postprocessing filters, and the available filters may differ from platform
+    to platform. Embedded devices often do not have enough CPU to implement
+    postprocessing in software. The filter selection is generally handled
+    automatically by the codec, depending on the amount of time remaining before
+    hitting the user-specified \ref usage_deadline after decoding the frame.
+
+
+*/
diff --git a/libvpx/libvpx/video_common.h b/libvpx/libvpx/video_common.h
new file mode 100644
index 0000000..44b27a8
--- /dev/null
+++ b/libvpx/libvpx/video_common.h
@@ -0,0 +1,23 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VIDEO_COMMON_H_
+#define VIDEO_COMMON_H_
+
+#include "./tools_common.h"
+
+typedef struct {
+  uint32_t codec_fourcc;
+  int frame_width;
+  int frame_height;
+  struct VpxRational time_base;
+} VpxVideoInfo;
+
+#endif  // VIDEO_COMMON_H_
diff --git a/libvpx/libvpx/video_reader.c b/libvpx/libvpx/video_reader.c
new file mode 100644
index 0000000..39c7edb
--- /dev/null
+++ b/libvpx/libvpx/video_reader.c
@@ -0,0 +1,83 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "./ivfdec.h"
+#include "./video_reader.h"
+
+#include "vpx_ports/mem_ops.h"
+
+static const char *const kIVFSignature = "DKIF";
+
+struct VpxVideoReaderStruct {
+  VpxVideoInfo info;
+  FILE *file;
+  uint8_t *buffer;
+  size_t buffer_size;
+  size_t frame_size;
+};
+
+VpxVideoReader *vpx_video_reader_open(const char *filename) {
+  char header[32];
+  VpxVideoReader *reader = NULL;
+  FILE *const file = fopen(filename, "rb");
+  if (!file)
+    return NULL;  // Can't open file
+
+  if (fread(header, 1, 32, file) != 32)
+    return NULL;  // Can't read file header
+
+  if (memcmp(kIVFSignature, header, 4) != 0)
+    return NULL;  // Wrong IVF signature
+
+  if (mem_get_le16(header + 4) != 0)
+    return NULL;  // Wrong IVF version
+
+  reader = calloc(1, sizeof(*reader));
+  if (!reader)
+    return NULL;  // Can't allocate VpxVideoReader
+
+  reader->file = file;
+  reader->info.codec_fourcc = mem_get_le32(header + 8);
+  reader->info.frame_width = mem_get_le16(header + 12);
+  reader->info.frame_height = mem_get_le16(header + 14);
+  reader->info.time_base.numerator = mem_get_le32(header + 16);
+  reader->info.time_base.denominator = mem_get_le32(header + 20);
+
+  return reader;
+}
+
+void vpx_video_reader_close(VpxVideoReader *reader) {
+  if (reader) {
+    fclose(reader->file);
+    free(reader->buffer);
+    free(reader);
+  }
+}
+
+int vpx_video_reader_read_frame(VpxVideoReader *reader) {
+  return !ivf_read_frame(reader->file, &reader->buffer, &reader->frame_size,
+                         &reader->buffer_size);
+}
+
+const uint8_t *vpx_video_reader_get_frame(VpxVideoReader *reader,
+                                          size_t *size) {
+  if (size)
+    *size = reader->frame_size;
+
+  return reader->buffer;
+}
+
+const VpxVideoInfo *vpx_video_reader_get_info(VpxVideoReader *reader) {
+  return &reader->info;
+}
+
diff --git a/libvpx/libvpx/video_reader.h b/libvpx/libvpx/video_reader.h
new file mode 100644
index 0000000..a62c6d7
--- /dev/null
+++ b/libvpx/libvpx/video_reader.h
@@ -0,0 +1,52 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VIDEO_READER_H_
+#define VIDEO_READER_H_
+
+#include "./video_common.h"
+
+// The following code is work in progress. It is going to  support transparent
+// reading of input files. Right now only IVF format is supported for
+// simplicity. The main goal the API is to be simple and easy to use in example
+// code and in vpxenc/vpxdec later. All low-level details like memory
+// buffer management are hidden from API users.
+struct VpxVideoReaderStruct;
+typedef struct VpxVideoReaderStruct VpxVideoReader;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Opens the input file for reading and inspects it to determine file type.
+// Returns an opaque VpxVideoReader* upon success, or NULL upon failure.
+// Right now only IVF format is supported.
+VpxVideoReader *vpx_video_reader_open(const char *filename);
+
+// Frees all resources associated with VpxVideoReader* returned from
+// vpx_video_reader_open() call.
+void vpx_video_reader_close(VpxVideoReader *reader);
+
+// Reads frame from the file and stores it in internal buffer.
+int vpx_video_reader_read_frame(VpxVideoReader *reader);
+
+// Returns the pointer to memory buffer with frame data read by last call to
+// vpx_video_reader_read_frame().
+const uint8_t *vpx_video_reader_get_frame(VpxVideoReader *reader,
+                                          size_t *size);
+
+// Fills VpxVideoInfo with information from opened video file.
+const VpxVideoInfo *vpx_video_reader_get_info(VpxVideoReader *reader);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VIDEO_READER_H_
diff --git a/libvpx/libvpx/video_writer.c b/libvpx/libvpx/video_writer.c
new file mode 100644
index 0000000..3695236
--- /dev/null
+++ b/libvpx/libvpx/video_writer.c
@@ -0,0 +1,80 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "./ivfenc.h"
+#include "./video_writer.h"
+#include "vpx/vpx_encoder.h"
+
+struct VpxVideoWriterStruct {
+  VpxVideoInfo info;
+  FILE *file;
+  int frame_count;
+};
+
+static void write_header(FILE *file, const VpxVideoInfo *info,
+                         int frame_count) {
+  struct vpx_codec_enc_cfg cfg;
+  cfg.g_w = info->frame_width;
+  cfg.g_h = info->frame_height;
+  cfg.g_timebase.num = info->time_base.numerator;
+  cfg.g_timebase.den = info->time_base.denominator;
+
+  ivf_write_file_header(file, &cfg, info->codec_fourcc, frame_count);
+}
+
+VpxVideoWriter *vpx_video_writer_open(const char *filename,
+                                      VpxContainer container,
+                                      const VpxVideoInfo *info) {
+  if (container == kContainerIVF) {
+    VpxVideoWriter *writer = NULL;
+    FILE *const file = fopen(filename, "wb");
+    if (!file)
+      return NULL;
+
+    writer = malloc(sizeof(*writer));
+    if (!writer)
+      return NULL;
+
+    writer->frame_count = 0;
+    writer->info = *info;
+    writer->file = file;
+
+    write_header(writer->file, info, 0);
+
+    return writer;
+  }
+
+  return NULL;
+}
+
+void vpx_video_writer_close(VpxVideoWriter *writer) {
+  if (writer) {
+    // Rewriting frame header with real frame count
+    rewind(writer->file);
+    write_header(writer->file, &writer->info, writer->frame_count);
+
+    fclose(writer->file);
+    free(writer);
+  }
+}
+
+int vpx_video_writer_write_frame(VpxVideoWriter *writer,
+                                 const uint8_t *buffer, size_t size,
+                                 int64_t pts) {
+  ivf_write_frame_header(writer->file, pts, size);
+  if (fwrite(buffer, 1, size, writer->file) != size)
+    return 0;
+
+  ++writer->frame_count;
+
+  return 1;
+}
diff --git a/libvpx/libvpx/video_writer.h b/libvpx/libvpx/video_writer.h
new file mode 100644
index 0000000..5dbfe52
--- /dev/null
+++ b/libvpx/libvpx/video_writer.h
@@ -0,0 +1,47 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VIDEO_WRITER_H_
+#define VIDEO_WRITER_H_
+
+#include "./video_common.h"
+
+typedef enum {
+  kContainerIVF
+} VpxContainer;
+
+struct VpxVideoWriterStruct;
+typedef struct VpxVideoWriterStruct VpxVideoWriter;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Finds and opens writer for specified container format.
+// Returns an opaque VpxVideoWriter* upon success, or NULL upon failure.
+// Right now only IVF format is supported.
+VpxVideoWriter *vpx_video_writer_open(const char *filename,
+                                      VpxContainer container,
+                                      const VpxVideoInfo *info);
+
+// Frees all resources associated with VpxVideoWriter* returned from
+// vpx_video_writer_open() call.
+void vpx_video_writer_close(VpxVideoWriter *writer);
+
+// Writes frame bytes to the file.
+int vpx_video_writer_write_frame(VpxVideoWriter *writer,
+                                 const uint8_t *buffer, size_t size,
+                                 int64_t pts);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VIDEO_WRITER_H_
diff --git a/libvpx/libvpx/vp8/common/alloccommon.c b/libvpx/libvpx/vp8/common/alloccommon.c
new file mode 100644
index 0000000..8dfd4ce
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/alloccommon.c
@@ -0,0 +1,190 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "alloccommon.h"
+#include "blockd.h"
+#include "vpx_mem/vpx_mem.h"
+#include "onyxc_int.h"
+#include "findnearmv.h"
+#include "entropymode.h"
+#include "systemdependent.h"
+
+void vp8_de_alloc_frame_buffers(VP8_COMMON *oci)
+{
+    int i;
+    for (i = 0; i < NUM_YV12_BUFFERS; i++)
+        vp8_yv12_de_alloc_frame_buffer(&oci->yv12_fb[i]);
+
+    vp8_yv12_de_alloc_frame_buffer(&oci->temp_scale_frame);
+#if CONFIG_POSTPROC
+    vp8_yv12_de_alloc_frame_buffer(&oci->post_proc_buffer);
+    if (oci->post_proc_buffer_int_used)
+        vp8_yv12_de_alloc_frame_buffer(&oci->post_proc_buffer_int);
+
+    vpx_free(oci->pp_limits_buffer);
+    oci->pp_limits_buffer = NULL;
+#endif
+
+    vpx_free(oci->above_context);
+    vpx_free(oci->mip);
+#if CONFIG_ERROR_CONCEALMENT
+    vpx_free(oci->prev_mip);
+    oci->prev_mip = NULL;
+#endif
+
+    oci->above_context = NULL;
+    oci->mip = NULL;
+}
+
+int vp8_alloc_frame_buffers(VP8_COMMON *oci, int width, int height)
+{
+    int i;
+
+    vp8_de_alloc_frame_buffers(oci);
+
+    /* our internal buffers are always multiples of 16 */
+    if ((width & 0xf) != 0)
+        width += 16 - (width & 0xf);
+
+    if ((height & 0xf) != 0)
+        height += 16 - (height & 0xf);
+
+
+    for (i = 0; i < NUM_YV12_BUFFERS; i++)
+    {
+        oci->fb_idx_ref_cnt[i] = 0;
+        oci->yv12_fb[i].flags = 0;
+        if (vp8_yv12_alloc_frame_buffer(&oci->yv12_fb[i], width, height, VP8BORDERINPIXELS) < 0)
+            goto allocation_fail;
+    }
+
+    oci->new_fb_idx = 0;
+    oci->lst_fb_idx = 1;
+    oci->gld_fb_idx = 2;
+    oci->alt_fb_idx = 3;
+
+    oci->fb_idx_ref_cnt[0] = 1;
+    oci->fb_idx_ref_cnt[1] = 1;
+    oci->fb_idx_ref_cnt[2] = 1;
+    oci->fb_idx_ref_cnt[3] = 1;
+
+    if (vp8_yv12_alloc_frame_buffer(&oci->temp_scale_frame,   width, 16, VP8BORDERINPIXELS) < 0)
+        goto allocation_fail;
+
+    oci->mb_rows = height >> 4;
+    oci->mb_cols = width >> 4;
+    oci->MBs = oci->mb_rows * oci->mb_cols;
+    oci->mode_info_stride = oci->mb_cols + 1;
+    oci->mip = vpx_calloc((oci->mb_cols + 1) * (oci->mb_rows + 1), sizeof(MODE_INFO));
+
+    if (!oci->mip)
+        goto allocation_fail;
+
+    oci->mi = oci->mip + oci->mode_info_stride + 1;
+
+    /* Allocation of previous mode info will be done in vp8_decode_frame()
+     * as it is a decoder only data */
+
+    oci->above_context = vpx_calloc(sizeof(ENTROPY_CONTEXT_PLANES) * oci->mb_cols, 1);
+
+    if (!oci->above_context)
+        goto allocation_fail;
+
+#if CONFIG_POSTPROC
+    if (vp8_yv12_alloc_frame_buffer(&oci->post_proc_buffer, width, height, VP8BORDERINPIXELS) < 0)
+        goto allocation_fail;
+
+    oci->post_proc_buffer_int_used = 0;
+    memset(&oci->postproc_state, 0, sizeof(oci->postproc_state));
+    memset(oci->post_proc_buffer.buffer_alloc, 128,
+           oci->post_proc_buffer.frame_size);
+
+    /* Allocate buffer to store post-processing filter coefficients.
+     *
+     * Note: Round up mb_cols to support SIMD reads
+     */
+    oci->pp_limits_buffer = vpx_memalign(16, 24 * ((oci->mb_cols + 1) & ~1));
+    if (!oci->pp_limits_buffer)
+        goto allocation_fail;
+#endif
+
+    return 0;
+
+allocation_fail:
+    vp8_de_alloc_frame_buffers(oci);
+    return 1;
+}
+
+void vp8_setup_version(VP8_COMMON *cm)
+{
+    switch (cm->version)
+    {
+    case 0:
+        cm->no_lpf = 0;
+        cm->filter_type = NORMAL_LOOPFILTER;
+        cm->use_bilinear_mc_filter = 0;
+        cm->full_pixel = 0;
+        break;
+    case 1:
+        cm->no_lpf = 0;
+        cm->filter_type = SIMPLE_LOOPFILTER;
+        cm->use_bilinear_mc_filter = 1;
+        cm->full_pixel = 0;
+        break;
+    case 2:
+        cm->no_lpf = 1;
+        cm->filter_type = NORMAL_LOOPFILTER;
+        cm->use_bilinear_mc_filter = 1;
+        cm->full_pixel = 0;
+        break;
+    case 3:
+        cm->no_lpf = 1;
+        cm->filter_type = SIMPLE_LOOPFILTER;
+        cm->use_bilinear_mc_filter = 1;
+        cm->full_pixel = 1;
+        break;
+    default:
+        /*4,5,6,7 are reserved for future use*/
+        cm->no_lpf = 0;
+        cm->filter_type = NORMAL_LOOPFILTER;
+        cm->use_bilinear_mc_filter = 0;
+        cm->full_pixel = 0;
+        break;
+    }
+}
+void vp8_create_common(VP8_COMMON *oci)
+{
+    vp8_machine_specific_config(oci);
+
+    vp8_init_mbmode_probs(oci);
+    vp8_default_bmode_probs(oci->fc.bmode_prob);
+
+    oci->mb_no_coeff_skip = 1;
+    oci->no_lpf = 0;
+    oci->filter_type = NORMAL_LOOPFILTER;
+    oci->use_bilinear_mc_filter = 0;
+    oci->full_pixel = 0;
+    oci->multi_token_partition = ONE_PARTITION;
+    oci->clamp_type = RECON_CLAMP_REQUIRED;
+
+    /* Initialize reference frame sign bias structure to defaults */
+    memset(oci->ref_frame_sign_bias, 0, sizeof(oci->ref_frame_sign_bias));
+
+    /* Default disable buffer to buffer copying */
+    oci->copy_buffer_to_gf = 0;
+    oci->copy_buffer_to_arf = 0;
+}
+
+void vp8_remove_common(VP8_COMMON *oci)
+{
+    vp8_de_alloc_frame_buffers(oci);
+}
diff --git a/libvpx/libvpx/vp8/common/alloccommon.h b/libvpx/libvpx/vp8/common/alloccommon.h
new file mode 100644
index 0000000..93e99d7
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/alloccommon.h
@@ -0,0 +1,31 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_ALLOCCOMMON_H_
+#define VP8_COMMON_ALLOCCOMMON_H_
+
+#include "onyxc_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_create_common(VP8_COMMON *oci);
+void vp8_remove_common(VP8_COMMON *oci);
+void vp8_de_alloc_frame_buffers(VP8_COMMON *oci);
+int vp8_alloc_frame_buffers(VP8_COMMON *oci, int width, int height);
+void vp8_setup_version(VP8_COMMON *oci);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_ALLOCCOMMON_H_
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/bilinearfilter_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/bilinearfilter_v6.asm
new file mode 100644
index 0000000..9704b42
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/bilinearfilter_v6.asm
@@ -0,0 +1,237 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vp8_filter_block2d_bil_first_pass_armv6|
+    EXPORT  |vp8_filter_block2d_bil_second_pass_armv6|
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+
+;-------------------------------------
+; r0    unsigned char  *src_ptr,
+; r1    unsigned short *dst_ptr,
+; r2    unsigned int    src_pitch,
+; r3    unsigned int    height,
+; stack unsigned int    width,
+; stack const short    *vp8_filter
+;-------------------------------------
+; The output is transposed stroed in output array to make it easy for second pass filtering.
+|vp8_filter_block2d_bil_first_pass_armv6| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; vp8_filter address
+    ldr     r4, [sp, #36]                   ; width
+
+    mov     r12, r3                         ; outer-loop counter
+
+    add     r7, r2, r4                      ; preload next row
+    pld     [r0, r7]
+
+    sub     r2, r2, r4                      ; src increment for height loop
+
+    ldr     r5, [r11]                       ; load up filter coefficients
+
+    mov     r3, r3, lsl #1                  ; height*2
+    add     r3, r3, #2                      ; plus 2 to make output buffer 4-bit aligned since height is actually (height+1)
+
+    mov     r11, r1                         ; save dst_ptr for each row
+
+    cmp     r5, #128                        ; if filter coef = 128, then skip the filter
+    beq     bil_null_1st_filter
+
+|bil_height_loop_1st_v6|
+    ldrb    r6, [r0]                        ; load source data
+    ldrb    r7, [r0, #1]
+    ldrb    r8, [r0, #2]
+    mov     lr, r4, lsr #2                  ; 4-in-parellel loop counter
+
+|bil_width_loop_1st_v6|
+    ldrb    r9, [r0, #3]
+    ldrb    r10, [r0, #4]
+
+    pkhbt   r6, r6, r7, lsl #16             ; src[1] | src[0]
+    pkhbt   r7, r7, r8, lsl #16             ; src[2] | src[1]
+
+    smuad   r6, r6, r5                      ; apply the filter
+    pkhbt   r8, r8, r9, lsl #16             ; src[3] | src[2]
+    smuad   r7, r7, r5
+    pkhbt   r9, r9, r10, lsl #16            ; src[4] | src[3]
+
+    smuad   r8, r8, r5
+    smuad   r9, r9, r5
+
+    add     r0, r0, #4
+    subs    lr, lr, #1
+
+    add     r6, r6, #0x40                   ; round_shift_and_clamp
+    add     r7, r7, #0x40
+    usat    r6, #16, r6, asr #7
+    usat    r7, #16, r7, asr #7
+
+    strh    r6, [r1], r3                    ; result is transposed and stored
+
+    add     r8, r8, #0x40                   ; round_shift_and_clamp
+    strh    r7, [r1], r3
+    add     r9, r9, #0x40
+    usat    r8, #16, r8, asr #7
+    usat    r9, #16, r9, asr #7
+
+    strh    r8, [r1], r3                    ; result is transposed and stored
+
+    ldrneb  r6, [r0]                        ; load source data
+    strh    r9, [r1], r3
+
+    ldrneb  r7, [r0, #1]
+    ldrneb  r8, [r0, #2]
+
+    bne     bil_width_loop_1st_v6
+
+    add     r0, r0, r2                      ; move to next input row
+    subs    r12, r12, #1
+
+    add     r9, r2, r4, lsl #1              ; adding back block width
+    pld     [r0, r9]                        ; preload next row
+
+    add     r11, r11, #2                    ; move over to next column
+    mov     r1, r11
+
+    bne     bil_height_loop_1st_v6
+
+    ldmia   sp!, {r4 - r11, pc}
+
+|bil_null_1st_filter|
+|bil_height_loop_null_1st|
+    mov     lr, r4, lsr #2                  ; loop counter
+
+|bil_width_loop_null_1st|
+    ldrb    r6, [r0]                        ; load data
+    ldrb    r7, [r0, #1]
+    ldrb    r8, [r0, #2]
+    ldrb    r9, [r0, #3]
+
+    strh    r6, [r1], r3                    ; store it to immediate buffer
+    add     r0, r0, #4
+    strh    r7, [r1], r3
+    subs    lr, lr, #1
+    strh    r8, [r1], r3
+    strh    r9, [r1], r3
+
+    bne     bil_width_loop_null_1st
+
+    subs    r12, r12, #1
+    add     r0, r0, r2                      ; move to next input line
+    add     r11, r11, #2                    ; move over to next column
+    mov     r1, r11
+
+    bne     bil_height_loop_null_1st
+
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP  ; |vp8_filter_block2d_bil_first_pass_armv6|
+
+
+;---------------------------------
+; r0    unsigned short *src_ptr,
+; r1    unsigned char  *dst_ptr,
+; r2    int             dst_pitch,
+; r3    unsigned int    height,
+; stack unsigned int    width,
+; stack const short    *vp8_filter
+;---------------------------------
+|vp8_filter_block2d_bil_second_pass_armv6| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; vp8_filter address
+    ldr     r4, [sp, #36]                   ; width
+
+    ldr     r5, [r11]                       ; load up filter coefficients
+    mov     r12, r4                         ; outer-loop counter = width, since we work on transposed data matrix
+    mov     r11, r1
+
+    cmp     r5, #128                        ; if filter coef = 128, then skip the filter
+    beq     bil_null_2nd_filter
+
+|bil_height_loop_2nd|
+    ldr     r6, [r0]                        ; load the data
+    ldr     r8, [r0, #4]
+    ldrh    r10, [r0, #8]
+    mov     lr, r3, lsr #2                  ; loop counter
+
+|bil_width_loop_2nd|
+    pkhtb   r7, r6, r8                      ; src[1] | src[2]
+    pkhtb   r9, r8, r10                     ; src[3] | src[4]
+
+    smuad   r6, r6, r5                      ; apply filter
+    smuad   r8, r8, r5                      ; apply filter
+
+    subs    lr, lr, #1
+
+    smuadx  r7, r7, r5                      ; apply filter
+    smuadx  r9, r9, r5                      ; apply filter
+
+    add     r0, r0, #8
+
+    add     r6, r6, #0x40                   ; round_shift_and_clamp
+    add     r7, r7, #0x40
+    usat    r6, #8, r6, asr #7
+    usat    r7, #8, r7, asr #7
+    strb    r6, [r1], r2                    ; the result is transposed back and stored
+
+    add     r8, r8, #0x40                   ; round_shift_and_clamp
+    strb    r7, [r1], r2
+    add     r9, r9, #0x40
+    usat    r8, #8, r8, asr #7
+    usat    r9, #8, r9, asr #7
+    strb    r8, [r1], r2                    ; the result is transposed back and stored
+
+    ldrne   r6, [r0]                        ; load data
+    strb    r9, [r1], r2
+    ldrne   r8, [r0, #4]
+    ldrneh  r10, [r0, #8]
+
+    bne     bil_width_loop_2nd
+
+    subs    r12, r12, #1
+    add     r0, r0, #4                      ; update src for next row
+    add     r11, r11, #1
+    mov     r1, r11
+
+    bne     bil_height_loop_2nd
+    ldmia   sp!, {r4 - r11, pc}
+
+|bil_null_2nd_filter|
+|bil_height_loop_null_2nd|
+    mov     lr, r3, lsr #2
+
+|bil_width_loop_null_2nd|
+    ldr     r6, [r0], #4                    ; load data
+    subs    lr, lr, #1
+    ldr     r8, [r0], #4
+
+    strb    r6, [r1], r2                    ; store data
+    mov     r7, r6, lsr #16
+    strb    r7, [r1], r2
+    mov     r9, r8, lsr #16
+    strb    r8, [r1], r2
+    strb    r9, [r1], r2
+
+    bne     bil_width_loop_null_2nd
+
+    subs    r12, r12, #1
+    add     r0, r0, #4
+    add     r11, r11, #1
+    mov     r1, r11
+
+    bne     bil_height_loop_null_2nd
+
+    ldmia   sp!, {r4 - r11, pc}
+    ENDP  ; |vp8_filter_block2d_second_pass_armv6|
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/copymem16x16_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/copymem16x16_v6.asm
new file mode 100644
index 0000000..abf048c
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/copymem16x16_v6.asm
@@ -0,0 +1,186 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vp8_copy_mem16x16_v6|
+    ; ARM
+    ; REQUIRE8
+    ; PRESERVE8
+
+    AREA    Block, CODE, READONLY ; name this block of code
+;void copy_mem16x16_v6( unsigned char *src, int src_stride, unsigned char *dst, int dst_stride)
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
+|vp8_copy_mem16x16_v6| PROC
+    stmdb       sp!, {r4 - r7}
+    ;push   {r4-r7}
+
+    ;preload
+    pld     [r0, #31]                ; preload for next 16x16 block
+
+    ands    r4, r0, #15
+    beq     copy_mem16x16_fast
+
+    ands    r4, r0, #7
+    beq     copy_mem16x16_8
+
+    ands    r4, r0, #3
+    beq     copy_mem16x16_4
+
+    ;copy one byte each time
+    ldrb    r4, [r0]
+    ldrb    r5, [r0, #1]
+    ldrb    r6, [r0, #2]
+    ldrb    r7, [r0, #3]
+
+    mov     r12, #16
+
+copy_mem16x16_1_loop
+    strb    r4, [r2]
+    strb    r5, [r2, #1]
+    strb    r6, [r2, #2]
+    strb    r7, [r2, #3]
+
+    ldrb    r4, [r0, #4]
+    ldrb    r5, [r0, #5]
+    ldrb    r6, [r0, #6]
+    ldrb    r7, [r0, #7]
+
+    subs    r12, r12, #1
+
+    strb    r4, [r2, #4]
+    strb    r5, [r2, #5]
+    strb    r6, [r2, #6]
+    strb    r7, [r2, #7]
+
+    ldrb    r4, [r0, #8]
+    ldrb    r5, [r0, #9]
+    ldrb    r6, [r0, #10]
+    ldrb    r7, [r0, #11]
+
+    strb    r4, [r2, #8]
+    strb    r5, [r2, #9]
+    strb    r6, [r2, #10]
+    strb    r7, [r2, #11]
+
+    ldrb    r4, [r0, #12]
+    ldrb    r5, [r0, #13]
+    ldrb    r6, [r0, #14]
+    ldrb    r7, [r0, #15]
+
+    add     r0, r0, r1
+
+    strb    r4, [r2, #12]
+    strb    r5, [r2, #13]
+    strb    r6, [r2, #14]
+    strb    r7, [r2, #15]
+
+    add     r2, r2, r3
+
+    ldrneb  r4, [r0]
+    ldrneb  r5, [r0, #1]
+    ldrneb  r6, [r0, #2]
+    ldrneb  r7, [r0, #3]
+
+    pld     [r0, #31]               ; preload for next 16x16 block
+
+    bne     copy_mem16x16_1_loop
+
+    ldmia       sp!, {r4 - r7}
+    ;pop        {r4-r7}
+    mov     pc, lr
+
+;copy 4 bytes each time
+copy_mem16x16_4
+    ldr     r4, [r0]
+    ldr     r5, [r0, #4]
+    ldr     r6, [r0, #8]
+    ldr     r7, [r0, #12]
+
+    mov     r12, #16
+
+copy_mem16x16_4_loop
+    subs    r12, r12, #1
+    add     r0, r0, r1
+
+    str     r4, [r2]
+    str     r5, [r2, #4]
+    str     r6, [r2, #8]
+    str     r7, [r2, #12]
+
+    add     r2, r2, r3
+
+    ldrne   r4, [r0]
+    ldrne   r5, [r0, #4]
+    ldrne   r6, [r0, #8]
+    ldrne   r7, [r0, #12]
+
+    pld     [r0, #31]               ; preload for next 16x16 block
+
+    bne     copy_mem16x16_4_loop
+
+    ldmia       sp!, {r4 - r7}
+    ;pop        {r4-r7}
+    mov     pc, lr
+
+;copy 8 bytes each time
+copy_mem16x16_8
+    sub     r1, r1, #16
+    sub     r3, r3, #16
+
+    mov     r12, #16
+
+copy_mem16x16_8_loop
+    ldmia   r0!, {r4-r5}
+    ;ldm        r0, {r4-r5}
+    ldmia   r0!, {r6-r7}
+
+    add     r0, r0, r1
+
+    stmia   r2!, {r4-r5}
+    subs    r12, r12, #1
+    ;stm        r2, {r4-r5}
+    stmia   r2!, {r6-r7}
+
+    add     r2, r2, r3
+
+    pld     [r0, #31]               ; preload for next 16x16 block
+    bne     copy_mem16x16_8_loop
+
+    ldmia       sp!, {r4 - r7}
+    ;pop        {r4-r7}
+    mov     pc, lr
+
+;copy 16 bytes each time
+copy_mem16x16_fast
+    ;sub        r1, r1, #16
+    ;sub        r3, r3, #16
+
+    mov     r12, #16
+
+copy_mem16x16_fast_loop
+    ldmia   r0, {r4-r7}
+    ;ldm        r0, {r4-r7}
+    add     r0, r0, r1
+
+    subs    r12, r12, #1
+    stmia   r2, {r4-r7}
+    ;stm        r2, {r4-r7}
+    add     r2, r2, r3
+
+    pld     [r0, #31]               ; preload for next 16x16 block
+    bne     copy_mem16x16_fast_loop
+
+    ldmia       sp!, {r4 - r7}
+    ;pop        {r4-r7}
+    mov     pc, lr
+
+    ENDP  ; |vp8_copy_mem16x16_v6|
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/copymem8x4_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/copymem8x4_v6.asm
new file mode 100644
index 0000000..d8362ef
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/copymem8x4_v6.asm
@@ -0,0 +1,128 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vp8_copy_mem8x4_v6|
+    ; ARM
+    ; REQUIRE8
+    ; PRESERVE8
+
+    AREA    Block, CODE, READONLY ; name this block of code
+;void vp8_copy_mem8x4_v6( unsigned char *src, int src_stride, unsigned char *dst, int dst_stride)
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
+|vp8_copy_mem8x4_v6| PROC
+    ;push   {r4-r5}
+    stmdb  sp!, {r4-r5}
+
+    ;preload
+    pld     [r0]
+    pld     [r0, r1]
+    pld     [r0, r1, lsl #1]
+
+    ands    r4, r0, #7
+    beq     copy_mem8x4_fast
+
+    ands    r4, r0, #3
+    beq     copy_mem8x4_4
+
+    ;copy 1 byte each time
+    ldrb    r4, [r0]
+    ldrb    r5, [r0, #1]
+
+    mov     r12, #4
+
+copy_mem8x4_1_loop
+    strb    r4, [r2]
+    strb    r5, [r2, #1]
+
+    ldrb    r4, [r0, #2]
+    ldrb    r5, [r0, #3]
+
+    subs    r12, r12, #1
+
+    strb    r4, [r2, #2]
+    strb    r5, [r2, #3]
+
+    ldrb    r4, [r0, #4]
+    ldrb    r5, [r0, #5]
+
+    strb    r4, [r2, #4]
+    strb    r5, [r2, #5]
+
+    ldrb    r4, [r0, #6]
+    ldrb    r5, [r0, #7]
+
+    add     r0, r0, r1
+
+    strb    r4, [r2, #6]
+    strb    r5, [r2, #7]
+
+    add     r2, r2, r3
+
+    ldrneb  r4, [r0]
+    ldrneb  r5, [r0, #1]
+
+    bne     copy_mem8x4_1_loop
+
+    ldmia       sp!, {r4 - r5}
+    ;pop        {r4-r5}
+    mov     pc, lr
+
+;copy 4 bytes each time
+copy_mem8x4_4
+    ldr     r4, [r0]
+    ldr     r5, [r0, #4]
+
+    mov     r12, #4
+
+copy_mem8x4_4_loop
+    subs    r12, r12, #1
+    add     r0, r0, r1
+
+    str     r4, [r2]
+    str     r5, [r2, #4]
+
+    add     r2, r2, r3
+
+    ldrne   r4, [r0]
+    ldrne   r5, [r0, #4]
+
+    bne     copy_mem8x4_4_loop
+
+    ldmia  sp!, {r4-r5}
+    ;pop        {r4-r5}
+    mov     pc, lr
+
+;copy 8 bytes each time
+copy_mem8x4_fast
+    ;sub        r1, r1, #8
+    ;sub        r3, r3, #8
+
+    mov     r12, #4
+
+copy_mem8x4_fast_loop
+    ldmia   r0, {r4-r5}
+    ;ldm        r0, {r4-r5}
+    add     r0, r0, r1
+
+    subs    r12, r12, #1
+    stmia   r2, {r4-r5}
+    ;stm        r2, {r4-r5}
+    add     r2, r2, r3
+
+    bne     copy_mem8x4_fast_loop
+
+    ldmia  sp!, {r4-r5}
+    ;pop        {r4-r5}
+    mov     pc, lr
+
+    ENDP  ; |vp8_copy_mem8x4_v6|
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/copymem8x8_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/copymem8x8_v6.asm
new file mode 100644
index 0000000..c6a60c6
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/copymem8x8_v6.asm
@@ -0,0 +1,128 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vp8_copy_mem8x8_v6|
+    ; ARM
+    ; REQUIRE8
+    ; PRESERVE8
+
+    AREA    Block, CODE, READONLY ; name this block of code
+;void copy_mem8x8_v6( unsigned char *src, int src_stride, unsigned char *dst, int dst_stride)
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
+|vp8_copy_mem8x8_v6| PROC
+    ;push   {r4-r5}
+    stmdb  sp!, {r4-r5}
+
+    ;preload
+    pld     [r0]
+    pld     [r0, r1]
+    pld     [r0, r1, lsl #1]
+
+    ands    r4, r0, #7
+    beq     copy_mem8x8_fast
+
+    ands    r4, r0, #3
+    beq     copy_mem8x8_4
+
+    ;copy 1 byte each time
+    ldrb    r4, [r0]
+    ldrb    r5, [r0, #1]
+
+    mov     r12, #8
+
+copy_mem8x8_1_loop
+    strb    r4, [r2]
+    strb    r5, [r2, #1]
+
+    ldrb    r4, [r0, #2]
+    ldrb    r5, [r0, #3]
+
+    subs    r12, r12, #1
+
+    strb    r4, [r2, #2]
+    strb    r5, [r2, #3]
+
+    ldrb    r4, [r0, #4]
+    ldrb    r5, [r0, #5]
+
+    strb    r4, [r2, #4]
+    strb    r5, [r2, #5]
+
+    ldrb    r4, [r0, #6]
+    ldrb    r5, [r0, #7]
+
+    add     r0, r0, r1
+
+    strb    r4, [r2, #6]
+    strb    r5, [r2, #7]
+
+    add     r2, r2, r3
+
+    ldrneb  r4, [r0]
+    ldrneb  r5, [r0, #1]
+
+    bne     copy_mem8x8_1_loop
+
+    ldmia       sp!, {r4 - r5}
+    ;pop        {r4-r5}
+    mov     pc, lr
+
+;copy 4 bytes each time
+copy_mem8x8_4
+    ldr     r4, [r0]
+    ldr     r5, [r0, #4]
+
+    mov     r12, #8
+
+copy_mem8x8_4_loop
+    subs    r12, r12, #1
+    add     r0, r0, r1
+
+    str     r4, [r2]
+    str     r5, [r2, #4]
+
+    add     r2, r2, r3
+
+    ldrne   r4, [r0]
+    ldrne   r5, [r0, #4]
+
+    bne     copy_mem8x8_4_loop
+
+    ldmia       sp!, {r4 - r5}
+    ;pop        {r4-r5}
+    mov     pc, lr
+
+;copy 8 bytes each time
+copy_mem8x8_fast
+    ;sub        r1, r1, #8
+    ;sub        r3, r3, #8
+
+    mov     r12, #8
+
+copy_mem8x8_fast_loop
+    ldmia   r0, {r4-r5}
+    ;ldm        r0, {r4-r5}
+    add     r0, r0, r1
+
+    subs    r12, r12, #1
+    stmia   r2, {r4-r5}
+    ;stm        r2, {r4-r5}
+    add     r2, r2, r3
+
+    bne     copy_mem8x8_fast_loop
+
+    ldmia  sp!, {r4-r5}
+    ;pop        {r4-r5}
+    mov     pc, lr
+
+    ENDP  ; |vp8_copy_mem8x8_v6|
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/dc_only_idct_add_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/dc_only_idct_add_v6.asm
new file mode 100644
index 0000000..9aa659f
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/dc_only_idct_add_v6.asm
@@ -0,0 +1,70 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+    EXPORT  |vp8_dc_only_idct_add_v6|
+
+    AREA    |.text|, CODE, READONLY
+
+;void vp8_dc_only_idct_add_c(short input_dc, unsigned char *pred_ptr,
+;                            int pred_stride, unsigned char *dst_ptr,
+;                            int dst_stride)
+; r0  input_dc
+; r1  pred_ptr
+; r2  pred_stride
+; r3  dst_ptr
+; sp  dst_stride
+
+|vp8_dc_only_idct_add_v6| PROC
+    stmdb       sp!, {r4 - r7}
+
+    add         r0, r0, #4                ; input_dc += 4
+    ldr         r12, c0x0000FFFF
+    ldr         r4, [r1], r2
+    and         r0, r12, r0, asr #3       ; input_dc >> 3 + mask
+    ldr         r6, [r1], r2
+    orr         r0, r0, r0, lsl #16       ; a1 | a1
+
+    ldr         r12, [sp, #16]            ; dst stride
+
+    uxtab16     r5, r0, r4                ; a1+2 | a1+0
+    uxtab16     r4, r0, r4, ror #8        ; a1+3 | a1+1
+    uxtab16     r7, r0, r6
+    uxtab16     r6, r0, r6, ror #8
+    usat16      r5, #8, r5
+    usat16      r4, #8, r4
+    usat16      r7, #8, r7
+    usat16      r6, #8, r6
+    orr         r5, r5, r4, lsl #8
+    orr         r7, r7, r6, lsl #8
+    ldr         r4, [r1], r2
+    str         r5, [r3], r12
+    ldr         r6, [r1]
+    str         r7, [r3], r12
+
+    uxtab16     r5, r0, r4
+    uxtab16     r4, r0, r4, ror #8
+    uxtab16     r7, r0, r6
+    uxtab16     r6, r0, r6, ror #8
+    usat16      r5, #8, r5
+    usat16      r4, #8, r4
+    usat16      r7, #8, r7
+    usat16      r6, #8, r6
+    orr         r5, r5, r4, lsl #8
+    orr         r7, r7, r6, lsl #8
+    str         r5, [r3], r12
+    str         r7, [r3]
+
+    ldmia       sp!, {r4 - r7}
+    bx          lr
+
+    ENDP  ; |vp8_dc_only_idct_add_v6|
+
+; Constant Pool
+c0x0000FFFF DCD 0x0000FFFF
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/dequant_idct_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/dequant_idct_v6.asm
new file mode 100644
index 0000000..db48ded
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/dequant_idct_v6.asm
@@ -0,0 +1,190 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+    EXPORT |vp8_dequant_idct_add_v6|
+
+    AREA |.text|, CODE, READONLY
+;void vp8_dequant_idct_v6(short *input, short *dq,
+;                         unsigned char *dest, int stride)
+; r0 = q
+; r1 = dq
+; r2 = dst
+; r3 = stride
+
+|vp8_dequant_idct_add_v6| PROC
+    stmdb   sp!, {r4-r11, lr}
+
+    ldr     r4, [r0]                ;input
+    ldr     r5, [r1], #4            ;dq
+
+    sub     sp, sp, #4
+    str     r3, [sp]
+
+    mov     r12, #4
+
+vp8_dequant_add_loop
+    smulbb  r6, r4, r5
+    smultt  r7, r4, r5
+
+    ldr     r4, [r0, #4]            ;input
+    ldr     r5, [r1], #4            ;dq
+
+    strh    r6, [r0], #2
+    strh    r7, [r0], #2
+
+    smulbb  r6, r4, r5
+    smultt  r7, r4, r5
+
+    subs    r12, r12, #1
+
+    ldrne   r4, [r0, #4]
+    ldrne   r5, [r1], #4
+
+    strh    r6, [r0], #2
+    strh    r7, [r0], #2
+
+    bne     vp8_dequant_add_loop
+
+    sub     r0, r0, #32
+    mov     r1, r0
+
+; short_idct4x4llm_v6_dual
+    ldr     r3, cospi8sqrt2minus1
+    ldr     r4, sinpi8sqrt2
+    ldr     r6, [r0, #8]
+    mov     r5, #2
+vp8_dequant_idct_loop1_v6
+    ldr     r12, [r0, #24]
+    ldr     r14, [r0, #16]
+    smulwt  r9, r3, r6
+    smulwb  r7, r3, r6
+    smulwt  r10, r4, r6
+    smulwb  r8, r4, r6
+    pkhbt   r7, r7, r9, lsl #16
+    smulwt  r11, r3, r12
+    pkhbt   r8, r8, r10, lsl #16
+    uadd16  r6, r6, r7
+    smulwt  r7, r4, r12
+    smulwb  r9, r3, r12
+    smulwb  r10, r4, r12
+    subs    r5, r5, #1
+    pkhbt   r9, r9, r11, lsl #16
+    ldr     r11, [r0], #4
+    pkhbt   r10, r10, r7, lsl #16
+    uadd16  r7, r12, r9
+    usub16  r7, r8, r7
+    uadd16  r6, r6, r10
+    uadd16  r10, r11, r14
+    usub16  r8, r11, r14
+    uadd16  r9, r10, r6
+    usub16  r10, r10, r6
+    uadd16  r6, r8, r7
+    usub16  r7, r8, r7
+    str     r6, [r1, #8]
+    ldrne   r6, [r0, #8]
+    str     r7, [r1, #16]
+    str     r10, [r1, #24]
+    str     r9, [r1], #4
+    bne     vp8_dequant_idct_loop1_v6
+
+    mov     r5, #2
+    sub     r0, r1, #8
+vp8_dequant_idct_loop2_v6
+    ldr     r6, [r0], #4
+    ldr     r7, [r0], #4
+    ldr     r8, [r0], #4
+    ldr     r9, [r0], #4
+    smulwt  r1, r3, r6
+    smulwt  r12, r4, r6
+    smulwt  lr, r3, r8
+    smulwt  r10, r4, r8
+    pkhbt   r11, r8, r6, lsl #16
+    pkhbt   r1, lr, r1, lsl #16
+    pkhbt   r12, r10, r12, lsl #16
+    pkhtb   r6, r6, r8, asr #16
+    uadd16  r6, r1, r6
+    pkhbt   lr, r9, r7, lsl #16
+    uadd16  r10, r11, lr
+    usub16  lr, r11, lr
+    pkhtb   r8, r7, r9, asr #16
+    subs    r5, r5, #1
+    smulwt  r1, r3, r8
+    smulwb  r7, r3, r8
+    smulwt  r11, r4, r8
+    smulwb  r9, r4, r8
+    pkhbt   r1, r7, r1, lsl #16
+    uadd16  r8, r1, r8
+    pkhbt   r11, r9, r11, lsl #16
+    usub16  r1, r12, r8
+    uadd16  r8, r11, r6
+    ldr     r9, c0x00040004
+    ldr     r12, [sp]               ; get stride from stack
+    uadd16  r6, r10, r8
+    usub16  r7, r10, r8
+    uadd16  r7, r7, r9
+    uadd16  r6, r6, r9
+    uadd16  r10, r14, r1
+    usub16  r1, r14, r1
+    uadd16  r10, r10, r9
+    uadd16  r1, r1, r9
+    ldr     r11, [r2]               ; load input from dst
+    mov     r8, r7, asr #3
+    pkhtb   r9, r8, r10, asr #19
+    mov     r8, r1, asr #3
+    pkhtb   r8, r8, r6, asr #19
+    uxtb16  lr, r11, ror #8
+    qadd16  r9, r9, lr
+    uxtb16  lr, r11
+    qadd16  r8, r8, lr
+    usat16  r9, #8, r9
+    usat16  r8, #8, r8
+    orr     r9, r8, r9, lsl #8
+    ldr     r11, [r2, r12]          ; load input from dst
+    mov     r7, r7, lsl #16
+    mov     r1, r1, lsl #16
+    mov     r10, r10, lsl #16
+    mov     r6, r6, lsl #16
+    mov     r7, r7, asr #3
+    pkhtb   r7, r7, r10, asr #19
+    mov     r1, r1, asr #3
+    pkhtb   r1, r1, r6, asr #19
+    uxtb16  r8, r11, ror #8
+    qadd16  r7, r7, r8
+    uxtb16  r8, r11
+    qadd16  r1, r1, r8
+    usat16  r7, #8, r7
+    usat16  r1, #8, r1
+    orr     r1, r1, r7, lsl #8
+    str     r9, [r2], r12           ; store output to dst
+    str     r1, [r2], r12           ; store output to dst
+    bne     vp8_dequant_idct_loop2_v6
+
+; memset
+    sub     r0, r0, #32
+    add     sp, sp, #4
+
+    mov     r12, #0
+    str     r12, [r0]
+    str     r12, [r0, #4]
+    str     r12, [r0, #8]
+    str     r12, [r0, #12]
+    str     r12, [r0, #16]
+    str     r12, [r0, #20]
+    str     r12, [r0, #24]
+    str     r12, [r0, #28]
+
+    ldmia   sp!, {r4 - r11, pc}
+    ENDP    ; |vp8_dequant_idct_add_v6|
+
+; Constant Pool
+cospi8sqrt2minus1 DCD 0x00004E7B
+sinpi8sqrt2       DCD 0x00008A8C
+c0x00040004       DCD 0x00040004
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/dequantize_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/dequantize_v6.asm
new file mode 100644
index 0000000..72f7e0e
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/dequantize_v6.asm
@@ -0,0 +1,69 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vp8_dequantize_b_loop_v6|
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+;-------------------------------
+;void   vp8_dequantize_b_loop_v6(short *Q, short *DQC, short *DQ);
+; r0    short *Q,
+; r1    short *DQC
+; r2    short *DQ
+|vp8_dequantize_b_loop_v6| PROC
+    stmdb   sp!, {r4-r9, lr}
+
+    ldr     r3, [r0]                ;load Q
+    ldr     r4, [r1]                ;load DQC
+    ldr     r5, [r0, #4]
+    ldr     r6, [r1, #4]
+
+    mov     r12, #2                 ;loop counter
+
+dequant_loop
+    smulbb  r7, r3, r4              ;multiply
+    smultt  r8, r3, r4
+    smulbb  r9, r5, r6
+    smultt  lr, r5, r6
+
+    ldr     r3, [r0, #8]
+    ldr     r4, [r1, #8]
+    ldr     r5, [r0, #12]
+    ldr     r6, [r1, #12]
+
+    strh    r7, [r2], #2            ;store result
+    smulbb  r7, r3, r4              ;multiply
+    strh    r8, [r2], #2
+    smultt  r8, r3, r4
+    strh    r9, [r2], #2
+    smulbb  r9, r5, r6
+    strh    lr, [r2], #2
+    smultt  lr, r5, r6
+
+    subs    r12, r12, #1
+
+    add     r0, r0, #16
+    add     r1, r1, #16
+
+    ldrne       r3, [r0]
+    strh    r7, [r2], #2            ;store result
+    ldrne       r4, [r1]
+    strh    r8, [r2], #2
+    ldrne       r5, [r0, #4]
+    strh    r9, [r2], #2
+    ldrne       r6, [r1, #4]
+    strh    lr, [r2], #2
+
+    bne     dequant_loop
+
+    ldmia   sp!, {r4-r9, pc}
+    ENDP    ;|vp8_dequantize_b_loop_v6|
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/filter_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/filter_v6.asm
new file mode 100644
index 0000000..eb4b75b
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/filter_v6.asm
@@ -0,0 +1,624 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vp8_filter_block2d_first_pass_armv6|
+    EXPORT  |vp8_filter_block2d_first_pass_16x16_armv6|
+    EXPORT  |vp8_filter_block2d_first_pass_8x8_armv6|
+    EXPORT  |vp8_filter_block2d_second_pass_armv6|
+    EXPORT  |vp8_filter4_block2d_second_pass_armv6|
+    EXPORT  |vp8_filter_block2d_first_pass_only_armv6|
+    EXPORT  |vp8_filter_block2d_second_pass_only_armv6|
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+;-------------------------------------
+; r0    unsigned char *src_ptr
+; r1    short         *output_ptr
+; r2    unsigned int src_pixels_per_line
+; r3    unsigned int output_width
+; stack unsigned int output_height
+; stack const short *vp8_filter
+;-------------------------------------
+; vp8_filter the input and put in the output array.  Apply the 6 tap FIR filter with
+; the output being a 2 byte value and the intput being a 1 byte value.
+|vp8_filter_block2d_first_pass_armv6| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; vp8_filter address
+    ldr     r7, [sp, #36]                   ; output height
+
+    sub     r2, r2, r3                      ; inside loop increments input array,
+                                            ; so the height loop only needs to add
+                                            ; r2 - width to the input pointer
+
+    mov     r3, r3, lsl #1                  ; multiply width by 2 because using shorts
+    add     r12, r3, #16                    ; square off the output
+    sub     sp, sp, #4
+
+    ldr     r4, [r11]                       ; load up packed filter coefficients
+    ldr     r5, [r11, #4]
+    ldr     r6, [r11, #8]
+
+    str     r1, [sp]                        ; push destination to stack
+    mov     r7, r7, lsl #16                 ; height is top part of counter
+
+; six tap filter
+|height_loop_1st_6|
+    ldrb    r8, [r0, #-2]                   ; load source data
+    ldrb    r9, [r0, #-1]
+    ldrb    r10, [r0], #2
+    orr     r7, r7, r3, lsr #2              ; construct loop counter
+
+|width_loop_1st_6|
+    ldrb    r11, [r0, #-1]
+
+    pkhbt   lr, r8, r9, lsl #16             ; r9 | r8
+    pkhbt   r8, r9, r10, lsl #16            ; r10 | r9
+
+    ldrb    r9, [r0]
+
+    smuad   lr, lr, r4                      ; apply the filter
+    pkhbt   r10, r10, r11, lsl #16          ; r11 | r10
+    smuad   r8, r8, r4
+    pkhbt   r11, r11, r9, lsl #16           ; r9 | r11
+
+    smlad   lr, r10, r5, lr
+    ldrb    r10, [r0, #1]
+    smlad   r8, r11, r5, r8
+    ldrb    r11, [r0, #2]
+
+    sub     r7, r7, #1
+
+    pkhbt   r9, r9, r10, lsl #16            ; r10 | r9
+    pkhbt   r10, r10, r11, lsl #16          ; r11 | r10
+
+    smlad   lr, r9, r6, lr
+    smlad   r11, r10, r6, r8
+
+    ands    r10, r7, #0xff                  ; test loop counter
+
+    add     lr, lr, #0x40                   ; round_shift_and_clamp
+    ldrneb  r8, [r0, #-2]                   ; load data for next loop
+    usat    lr, #8, lr, asr #7
+    add     r11, r11, #0x40
+    ldrneb  r9, [r0, #-1]
+    usat    r11, #8, r11, asr #7
+
+    strh    lr, [r1], r12                   ; result is transposed and stored, which
+                                            ; will make second pass filtering easier.
+    ldrneb  r10, [r0], #2
+    strh    r11, [r1], r12
+
+    bne     width_loop_1st_6
+
+    ldr     r1, [sp]                        ; load and update dst address
+    subs    r7, r7, #0x10000
+    add     r0, r0, r2                      ; move to next input line
+
+    add     r1, r1, #2                      ; move over to next column
+    str     r1, [sp]
+
+    bne     height_loop_1st_6
+
+    add     sp, sp, #4
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP
+
+; --------------------------
+; 16x16 version
+; -----------------------------
+|vp8_filter_block2d_first_pass_16x16_armv6| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; vp8_filter address
+    ldr     r7, [sp, #36]                   ; output height
+
+    add     r4, r2, #18                     ; preload next low
+    pld     [r0, r4]
+
+    sub     r2, r2, r3                      ; inside loop increments input array,
+                                            ; so the height loop only needs to add
+                                            ; r2 - width to the input pointer
+
+    mov     r3, r3, lsl #1                  ; multiply width by 2 because using shorts
+    add     r12, r3, #16                    ; square off the output
+    sub     sp, sp, #4
+
+    ldr     r4, [r11]                       ; load up packed filter coefficients
+    ldr     r5, [r11, #4]
+    ldr     r6, [r11, #8]
+
+    str     r1, [sp]                        ; push destination to stack
+    mov     r7, r7, lsl #16                 ; height is top part of counter
+
+; six tap filter
+|height_loop_1st_16_6|
+    ldrb    r8, [r0, #-2]                   ; load source data
+    ldrb    r9, [r0, #-1]
+    ldrb    r10, [r0], #2
+    orr     r7, r7, r3, lsr #2              ; construct loop counter
+
+|width_loop_1st_16_6|
+    ldrb    r11, [r0, #-1]
+
+    pkhbt   lr, r8, r9, lsl #16             ; r9 | r8
+    pkhbt   r8, r9, r10, lsl #16            ; r10 | r9
+
+    ldrb    r9, [r0]
+
+    smuad   lr, lr, r4                      ; apply the filter
+    pkhbt   r10, r10, r11, lsl #16          ; r11 | r10
+    smuad   r8, r8, r4
+    pkhbt   r11, r11, r9, lsl #16           ; r9 | r11
+
+    smlad   lr, r10, r5, lr
+    ldrb    r10, [r0, #1]
+    smlad   r8, r11, r5, r8
+    ldrb    r11, [r0, #2]
+
+    sub     r7, r7, #1
+
+    pkhbt   r9, r9, r10, lsl #16            ; r10 | r9
+    pkhbt   r10, r10, r11, lsl #16          ; r11 | r10
+
+    smlad   lr, r9, r6, lr
+    smlad   r11, r10, r6, r8
+
+    ands    r10, r7, #0xff                  ; test loop counter
+
+    add     lr, lr, #0x40                   ; round_shift_and_clamp
+    ldrneb  r8, [r0, #-2]                   ; load data for next loop
+    usat    lr, #8, lr, asr #7
+    add     r11, r11, #0x40
+    ldrneb  r9, [r0, #-1]
+    usat    r11, #8, r11, asr #7
+
+    strh    lr, [r1], r12                   ; result is transposed and stored, which
+                                            ; will make second pass filtering easier.
+    ldrneb  r10, [r0], #2
+    strh    r11, [r1], r12
+
+    bne     width_loop_1st_16_6
+
+    ldr     r1, [sp]                        ; load and update dst address
+    subs    r7, r7, #0x10000
+    add     r0, r0, r2                      ; move to next input line
+
+    add     r11, r2, #34                    ; adding back block width(=16)
+    pld     [r0, r11]                       ; preload next low
+
+    add     r1, r1, #2                      ; move over to next column
+    str     r1, [sp]
+
+    bne     height_loop_1st_16_6
+
+    add     sp, sp, #4
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP
+
+; --------------------------
+; 8x8 version
+; -----------------------------
+|vp8_filter_block2d_first_pass_8x8_armv6| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; vp8_filter address
+    ldr     r7, [sp, #36]                   ; output height
+
+    add     r4, r2, #10                     ; preload next low
+    pld     [r0, r4]
+
+    sub     r2, r2, r3                      ; inside loop increments input array,
+                                            ; so the height loop only needs to add
+                                            ; r2 - width to the input pointer
+
+    mov     r3, r3, lsl #1                  ; multiply width by 2 because using shorts
+    add     r12, r3, #16                    ; square off the output
+    sub     sp, sp, #4
+
+    ldr     r4, [r11]                       ; load up packed filter coefficients
+    ldr     r5, [r11, #4]
+    ldr     r6, [r11, #8]
+
+    str     r1, [sp]                        ; push destination to stack
+    mov     r7, r7, lsl #16                 ; height is top part of counter
+
+; six tap filter
+|height_loop_1st_8_6|
+    ldrb    r8, [r0, #-2]                   ; load source data
+    ldrb    r9, [r0, #-1]
+    ldrb    r10, [r0], #2
+    orr     r7, r7, r3, lsr #2              ; construct loop counter
+
+|width_loop_1st_8_6|
+    ldrb    r11, [r0, #-1]
+
+    pkhbt   lr, r8, r9, lsl #16             ; r9 | r8
+    pkhbt   r8, r9, r10, lsl #16            ; r10 | r9
+
+    ldrb    r9, [r0]
+
+    smuad   lr, lr, r4                      ; apply the filter
+    pkhbt   r10, r10, r11, lsl #16          ; r11 | r10
+    smuad   r8, r8, r4
+    pkhbt   r11, r11, r9, lsl #16           ; r9 | r11
+
+    smlad   lr, r10, r5, lr
+    ldrb    r10, [r0, #1]
+    smlad   r8, r11, r5, r8
+    ldrb    r11, [r0, #2]
+
+    sub     r7, r7, #1
+
+    pkhbt   r9, r9, r10, lsl #16            ; r10 | r9
+    pkhbt   r10, r10, r11, lsl #16          ; r11 | r10
+
+    smlad   lr, r9, r6, lr
+    smlad   r11, r10, r6, r8
+
+    ands    r10, r7, #0xff                  ; test loop counter
+
+    add     lr, lr, #0x40                   ; round_shift_and_clamp
+    ldrneb  r8, [r0, #-2]                   ; load data for next loop
+    usat    lr, #8, lr, asr #7
+    add     r11, r11, #0x40
+    ldrneb  r9, [r0, #-1]
+    usat    r11, #8, r11, asr #7
+
+    strh    lr, [r1], r12                   ; result is transposed and stored, which
+                                            ; will make second pass filtering easier.
+    ldrneb  r10, [r0], #2
+    strh    r11, [r1], r12
+
+    bne     width_loop_1st_8_6
+
+    ldr     r1, [sp]                        ; load and update dst address
+    subs    r7, r7, #0x10000
+    add     r0, r0, r2                      ; move to next input line
+
+    add     r11, r2, #18                    ; adding back block width(=8)
+    pld     [r0, r11]                       ; preload next low
+
+    add     r1, r1, #2                      ; move over to next column
+    str     r1, [sp]
+
+    bne     height_loop_1st_8_6
+
+    add     sp, sp, #4
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP
+
+;---------------------------------
+; r0    short         *src_ptr,
+; r1    unsigned char *output_ptr,
+; r2    unsigned int output_pitch,
+; r3    unsigned int cnt,
+; stack const short *vp8_filter
+;---------------------------------
+|vp8_filter_block2d_second_pass_armv6| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #36]                  ; vp8_filter address
+    sub     sp, sp, #4
+    mov     r7, r3, lsl #16                 ; height is top part of counter
+    str     r1, [sp]                        ; push destination to stack
+
+    ldr     r4, [r11]                       ; load up packed filter coefficients
+    ldr     r5, [r11, #4]
+    ldr     r6, [r11, #8]
+
+    pkhbt   r12, r5, r4                     ; pack the filter differently
+    pkhbt   r11, r6, r5
+
+    sub     r0, r0, #4                      ; offset input buffer
+
+|height_loop_2nd|
+    ldr     r8, [r0]                        ; load the data
+    ldr     r9, [r0, #4]
+    orr     r7, r7, r3, lsr #1              ; loop counter
+
+|width_loop_2nd|
+    smuad   lr, r4, r8                      ; apply filter
+    sub     r7, r7, #1
+    smulbt  r8, r4, r8
+
+    ldr     r10, [r0, #8]
+
+    smlad   lr, r5, r9, lr
+    smladx  r8, r12, r9, r8
+
+    ldrh    r9, [r0, #12]
+
+    smlad   lr, r6, r10, lr
+    smladx  r8, r11, r10, r8
+
+    add     r0, r0, #4
+    smlatb  r10, r6, r9, r8
+
+    add     lr, lr, #0x40                   ; round_shift_and_clamp
+    ands    r8, r7, #0xff
+    usat    lr, #8, lr, asr #7
+    add     r10, r10, #0x40
+    strb    lr, [r1], r2                    ; the result is transposed back and stored
+    usat    r10, #8, r10, asr #7
+
+    ldrne   r8, [r0]                        ; load data for next loop
+    ldrne   r9, [r0, #4]
+    strb    r10, [r1], r2
+
+    bne     width_loop_2nd
+
+    ldr     r1, [sp]                        ; update dst for next loop
+    subs    r7, r7, #0x10000
+    add     r0, r0, #16                     ; updata src for next loop
+    add     r1, r1, #1
+    str     r1, [sp]
+
+    bne     height_loop_2nd
+
+    add     sp, sp, #4
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP
+
+;---------------------------------
+; r0    short         *src_ptr,
+; r1    unsigned char *output_ptr,
+; r2    unsigned int output_pitch,
+; r3    unsigned int cnt,
+; stack const short *vp8_filter
+;---------------------------------
+|vp8_filter4_block2d_second_pass_armv6| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #36]                  ; vp8_filter address
+    mov     r7, r3, lsl #16                 ; height is top part of counter
+
+    ldr     r4, [r11]                       ; load up packed filter coefficients
+    add     lr, r1, r3                      ; save final destination pointer
+    ldr     r5, [r11, #4]
+    ldr     r6, [r11, #8]
+
+    pkhbt   r12, r5, r4                     ; pack the filter differently
+    pkhbt   r11, r6, r5
+    mov     r4, #0x40                       ; rounding factor (for smlad{x})
+
+|height_loop_2nd_4|
+    ldrd    r8, r9, [r0, #-4]               ; load the data
+    orr     r7, r7, r3, lsr #1              ; loop counter
+
+|width_loop_2nd_4|
+    ldr     r10, [r0, #4]!
+    smladx  r6, r9, r12, r4                 ; apply filter
+    pkhbt   r8, r9, r8
+    smlad   r5, r8, r12, r4
+    pkhbt   r8, r10, r9
+    smladx  r6, r10, r11, r6
+    sub     r7, r7, #1
+    smlad   r5, r8, r11, r5
+
+    mov     r8, r9                          ; shift the data for the next loop
+    mov     r9, r10
+
+    usat    r6, #8, r6, asr #7              ; shift and clamp
+    usat    r5, #8, r5, asr #7
+
+    strb    r5, [r1], r2                    ; the result is transposed back and stored
+    tst     r7, #0xff
+    strb    r6, [r1], r2
+
+    bne     width_loop_2nd_4
+
+    subs    r7, r7, #0x10000
+    add     r0, r0, #16                     ; update src for next loop
+    sub     r1, lr, r7, lsr #16             ; update dst for next loop
+
+    bne     height_loop_2nd_4
+
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP
+
+;------------------------------------
+; r0    unsigned char *src_ptr
+; r1    unsigned char *output_ptr,
+; r2    unsigned int src_pixels_per_line
+; r3    unsigned int cnt,
+; stack unsigned int output_pitch,
+; stack const short *vp8_filter
+;------------------------------------
+|vp8_filter_block2d_first_pass_only_armv6| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    add     r7, r2, r3                      ; preload next low
+    add     r7, r7, #2
+    pld     [r0, r7]
+
+    ldr     r4, [sp, #36]                   ; output pitch
+    ldr     r11, [sp, #40]                  ; HFilter address
+    sub     sp, sp, #8
+
+    mov     r7, r3
+    sub     r2, r2, r3                      ; inside loop increments input array,
+                                            ; so the height loop only needs to add
+                                            ; r2 - width to the input pointer
+
+    sub     r4, r4, r3
+    str     r4, [sp]                        ; save modified output pitch
+    str     r2, [sp, #4]
+
+    mov     r2, #0x40
+
+    ldr     r4, [r11]                       ; load up packed filter coefficients
+    ldr     r5, [r11, #4]
+    ldr     r6, [r11, #8]
+
+; six tap filter
+|height_loop_1st_only_6|
+    ldrb    r8, [r0, #-2]                   ; load data
+    ldrb    r9, [r0, #-1]
+    ldrb    r10, [r0], #2
+
+    mov     r12, r3, lsr #1                 ; loop counter
+
+|width_loop_1st_only_6|
+    ldrb    r11, [r0, #-1]
+
+    pkhbt   lr, r8, r9, lsl #16             ; r9 | r8
+    pkhbt   r8, r9, r10, lsl #16            ; r10 | r9
+
+    ldrb    r9, [r0]
+
+;;  smuad   lr, lr, r4
+    smlad   lr, lr, r4, r2
+    pkhbt   r10, r10, r11, lsl #16          ; r11 | r10
+;;  smuad   r8, r8, r4
+    smlad   r8, r8, r4, r2
+    pkhbt   r11, r11, r9, lsl #16           ; r9 | r11
+
+    smlad   lr, r10, r5, lr
+    ldrb    r10, [r0, #1]
+    smlad   r8, r11, r5, r8
+    ldrb    r11, [r0, #2]
+
+    subs    r12, r12, #1
+
+    pkhbt   r9, r9, r10, lsl #16            ; r10 | r9
+    pkhbt   r10, r10, r11, lsl #16          ; r11 | r10
+
+    smlad   lr, r9, r6, lr
+    smlad   r10, r10, r6, r8
+
+;;  add     lr, lr, #0x40                   ; round_shift_and_clamp
+    ldrneb  r8, [r0, #-2]                   ; load data for next loop
+    usat    lr, #8, lr, asr #7
+;;  add     r10, r10, #0x40
+    strb    lr, [r1], #1                    ; store the result
+    usat    r10, #8, r10, asr #7
+
+    ldrneb  r9, [r0, #-1]
+    strb    r10, [r1], #1
+    ldrneb  r10, [r0], #2
+
+    bne     width_loop_1st_only_6
+
+    ldr     lr, [sp]                        ; load back output pitch
+    ldr     r12, [sp, #4]                   ; load back output pitch
+    subs    r7, r7, #1
+    add     r0, r0, r12                     ; updata src for next loop
+
+    add     r11, r12, r3                    ; preload next low
+    add     r11, r11, #2
+    pld     [r0, r11]
+
+    add     r1, r1, lr                      ; update dst for next loop
+
+    bne     height_loop_1st_only_6
+
+    add     sp, sp, #8
+    ldmia   sp!, {r4 - r11, pc}
+    ENDP  ; |vp8_filter_block2d_first_pass_only_armv6|
+
+
+;------------------------------------
+; r0    unsigned char *src_ptr,
+; r1    unsigned char *output_ptr,
+; r2    unsigned int src_pixels_per_line
+; r3    unsigned int cnt,
+; stack unsigned int output_pitch,
+; stack const short *vp8_filter
+;------------------------------------
+|vp8_filter_block2d_second_pass_only_armv6| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; VFilter address
+    ldr     r12, [sp, #36]                  ; output pitch
+
+    mov     r7, r3, lsl #16                 ; height is top part of counter
+    sub     r0, r0, r2, lsl #1              ; need 6 elements for filtering, 2 before, 3 after
+
+    sub     sp, sp, #8
+
+    ldr     r4, [r11]                       ; load up packed filter coefficients
+    ldr     r5, [r11, #4]
+    ldr     r6, [r11, #8]
+
+    str     r0, [sp]                        ; save r0 to stack
+    str     r1, [sp, #4]                    ; save dst to stack
+
+; six tap filter
+|width_loop_2nd_only_6|
+    ldrb    r8, [r0], r2                    ; load data
+    orr     r7, r7, r3                      ; loop counter
+    ldrb    r9, [r0], r2
+    ldrb    r10, [r0], r2
+
+|height_loop_2nd_only_6|
+    ; filter first column in this inner loop, than, move to next colum.
+    ldrb    r11, [r0], r2
+
+    pkhbt   lr, r8, r9, lsl #16             ; r9 | r8
+    pkhbt   r8, r9, r10, lsl #16            ; r10 | r9
+
+    ldrb    r9, [r0], r2
+
+    smuad   lr, lr, r4
+    pkhbt   r10, r10, r11, lsl #16          ; r11 | r10
+    smuad   r8, r8, r4
+    pkhbt   r11, r11, r9, lsl #16           ; r9 | r11
+
+    smlad   lr, r10, r5, lr
+    ldrb    r10, [r0], r2
+    smlad   r8, r11, r5, r8
+    ldrb    r11, [r0]
+
+    sub     r7, r7, #2
+    sub     r0, r0, r2, lsl #2
+
+    pkhbt   r9, r9, r10, lsl #16            ; r10 | r9
+    pkhbt   r10, r10, r11, lsl #16          ; r11 | r10
+
+    smlad   lr, r9, r6, lr
+    smlad   r10, r10, r6, r8
+
+    ands    r9, r7, #0xff
+
+    add     lr, lr, #0x40                   ; round_shift_and_clamp
+    ldrneb  r8, [r0], r2                    ; load data for next loop
+    usat    lr, #8, lr, asr #7
+    add     r10, r10, #0x40
+    strb    lr, [r1], r12                   ; store the result for the column
+    usat    r10, #8, r10, asr #7
+
+    ldrneb  r9, [r0], r2
+    strb    r10, [r1], r12
+    ldrneb  r10, [r0], r2
+
+    bne     height_loop_2nd_only_6
+
+    ldr     r0, [sp]
+    ldr     r1, [sp, #4]
+    subs    r7, r7, #0x10000
+    add     r0, r0, #1                      ; move to filter next column
+    str     r0, [sp]
+    add     r1, r1, #1
+    str     r1, [sp, #4]
+
+    bne     width_loop_2nd_only_6
+
+    add     sp, sp, #8
+
+    ldmia   sp!, {r4 - r11, pc}
+    ENDP  ; |vp8_filter_block2d_second_pass_only_armv6|
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/idct_blk_v6.c b/libvpx/libvpx/vp8/common/arm/armv6/idct_blk_v6.c
new file mode 100644
index 0000000..c94f84a
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/idct_blk_v6.c
@@ -0,0 +1,115 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+
+
+void vp8_dequant_idct_add_y_block_v6(short *q, short *dq,
+                                     unsigned char *dst,
+                                     int stride, char *eobs)
+{
+    int i;
+
+    for (i = 0; i < 4; i++)
+    {
+        if (eobs[0] > 1)
+            vp8_dequant_idct_add_v6 (q, dq, dst, stride);
+        else if (eobs[0] == 1)
+        {
+            vp8_dc_only_idct_add_v6 (q[0]*dq[0], dst, stride, dst, stride);
+            ((int *)q)[0] = 0;
+        }
+
+        if (eobs[1] > 1)
+            vp8_dequant_idct_add_v6 (q+16, dq, dst+4, stride);
+        else if (eobs[1] == 1)
+        {
+            vp8_dc_only_idct_add_v6 (q[16]*dq[0], dst+4, stride, dst+4, stride);
+            ((int *)(q+16))[0] = 0;
+        }
+
+        if (eobs[2] > 1)
+            vp8_dequant_idct_add_v6 (q+32, dq, dst+8, stride);
+        else if (eobs[2] == 1)
+        {
+            vp8_dc_only_idct_add_v6 (q[32]*dq[0], dst+8, stride, dst+8, stride);
+            ((int *)(q+32))[0] = 0;
+        }
+
+        if (eobs[3] > 1)
+            vp8_dequant_idct_add_v6 (q+48, dq, dst+12, stride);
+        else if (eobs[3] == 1)
+        {
+            vp8_dc_only_idct_add_v6 (q[48]*dq[0], dst+12, stride,dst+12,stride);
+            ((int *)(q+48))[0] = 0;
+        }
+
+        q    += 64;
+        dst  += 4*stride;
+        eobs += 4;
+    }
+}
+
+void vp8_dequant_idct_add_uv_block_v6(short *q, short *dq,
+                                      unsigned char *dstu,
+                                      unsigned char *dstv,
+                                      int stride, char *eobs)
+{
+    int i;
+
+    for (i = 0; i < 2; i++)
+    {
+        if (eobs[0] > 1)
+            vp8_dequant_idct_add_v6 (q, dq, dstu, stride);
+        else if (eobs[0] == 1)
+        {
+            vp8_dc_only_idct_add_v6 (q[0]*dq[0], dstu, stride, dstu, stride);
+            ((int *)q)[0] = 0;
+        }
+
+        if (eobs[1] > 1)
+            vp8_dequant_idct_add_v6 (q+16, dq, dstu+4, stride);
+        else if (eobs[1] == 1)
+        {
+            vp8_dc_only_idct_add_v6 (q[16]*dq[0], dstu+4, stride,
+                                                  dstu+4, stride);
+            ((int *)(q+16))[0] = 0;
+        }
+
+        q    += 32;
+        dstu += 4*stride;
+        eobs += 2;
+    }
+
+    for (i = 0; i < 2; i++)
+    {
+        if (eobs[0] > 1)
+            vp8_dequant_idct_add_v6 (q, dq, dstv, stride);
+        else if (eobs[0] == 1)
+        {
+            vp8_dc_only_idct_add_v6 (q[0]*dq[0], dstv, stride, dstv, stride);
+            ((int *)q)[0] = 0;
+        }
+
+        if (eobs[1] > 1)
+            vp8_dequant_idct_add_v6 (q+16, dq, dstv+4, stride);
+        else if (eobs[1] == 1)
+        {
+            vp8_dc_only_idct_add_v6 (q[16]*dq[0], dstv+4, stride,
+                                                  dstv+4, stride);
+            ((int *)(q+16))[0] = 0;
+        }
+
+        q    += 32;
+        dstv += 4*stride;
+        eobs += 2;
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/idct_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/idct_v6.asm
new file mode 100644
index 0000000..b4d44cb
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/idct_v6.asm
@@ -0,0 +1,202 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vp8_short_idct4x4llm_v6_dual|
+
+    AREA    |.text|, CODE, READONLY
+
+
+; void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch,
+;                             unsigned char *dst, int stride)
+; r0    short* input
+; r1    unsigned char* pred
+; r2    int pitch
+; r3    unsigned char* dst
+; sp    int stride
+
+|vp8_short_idct4x4llm_v6_dual| PROC
+    stmdb   sp!, {r4-r11, lr}
+
+    sub     sp, sp, #4
+
+    mov     r4, #0x00008A00         ; sin
+    orr     r4, r4, #0x0000008C     ; sinpi8sqrt2
+
+    mov     r5, #0x00004E00         ; cos
+    orr     r5, r5, #0x0000007B     ; cospi8sqrt2minus1
+    orr     r5, r5, #1<<31          ; loop counter on top bit
+
+loop1_dual
+    ldr     r6, [r0, #(4*2)]        ; i5 | i4
+    ldr     r12, [r0, #(12*2)]      ; i13|i12
+    ldr     r14, [r0, #(8*2)]       ; i9 | i8
+
+    smulbt  r9, r5, r6              ; (ip[5] * cospi8sqrt2minus1) >> 16
+    smulbb  r7, r5, r6              ; (ip[4] * cospi8sqrt2minus1) >> 16
+    smulwt  r10, r4, r6             ; (ip[5] * sinpi8sqrt2) >> 16
+    smulwb  r8, r4, r6              ; (ip[4] * sinpi8sqrt2) >> 16
+
+    smulbt  r11, r5, r12            ; (ip[13] * cospi8sqrt2minus1) >> 16
+    pkhtb   r7, r9, r7, asr #16     ; 5c | 4c
+    pkhbt   r8, r8, r10, lsl #16    ; 5s | 4s
+    uadd16  r6, r6, r7              ; 5c+5 | 4c+4
+
+    smulwt  r7, r4, r12             ; (ip[13] * sinpi8sqrt2) >> 16
+    smulbb  r9, r5, r12             ; (ip[12] * cospi8sqrt2minus1) >> 16
+    smulwb  r10, r4, r12            ; (ip[12] * sinpi8sqrt2) >> 16
+
+    subs    r5, r5, #1<<31          ; i--
+
+    pkhtb   r9, r11, r9, asr #16    ; 13c | 12c
+    ldr     r11, [r0]               ; i1 | i0
+    pkhbt   r10, r10, r7, lsl #16   ; 13s | 12s
+    uadd16  r7, r12, r9             ; 13c+13 | 12c+12
+
+    usub16  r7, r8, r7              ; c
+    uadd16  r6, r6, r10             ; d
+    uadd16  r10, r11, r14           ; a
+    usub16  r8, r11, r14            ; b
+
+    uadd16  r9, r10, r6             ; a+d
+    usub16  r10, r10, r6            ; a-d
+    uadd16  r6, r8, r7              ; b+c
+    usub16  r7, r8, r7              ; b-c
+
+    ; use input buffer to store intermediate results
+    str      r6, [r0, #(4*2)]       ; o5 | o4
+    str      r7, [r0, #(8*2)]       ; o9 | o8
+    str      r10,[r0, #(12*2)]      ; o13|o12
+    str      r9, [r0], #4           ; o1 | o0
+
+    bcs loop1_dual
+
+    sub     r0, r0, #8              ; reset input/output
+    str     r0, [sp]
+
+loop2_dual
+
+    ldr     r6, [r0, #(4*2)]        ; i5 | i4
+    ldr     r12,[r0, #(2*2)]        ; i3 | i2
+    ldr     r14,[r0, #(6*2)]        ; i7 | i6
+    ldr     r0, [r0, #(0*2)]        ; i1 | i0
+
+    smulbt  r9, r5, r6              ; (ip[5] * cospi8sqrt2minus1) >> 16
+    smulbt  r7, r5, r0              ; (ip[1] * cospi8sqrt2minus1) >> 16
+    smulwt  r10, r4, r6             ; (ip[5] * sinpi8sqrt2) >> 16
+    smulwt  r8, r4, r0              ; (ip[1] * sinpi8sqrt2) >> 16
+
+    pkhbt   r11, r6, r0, lsl #16    ; i0 | i4
+    pkhtb   r7, r7, r9, asr #16     ; 1c | 5c
+    pkhtb   r0, r0, r6, asr #16     ; i1 | i5
+    pkhbt   r8, r10, r8, lsl #16    ; 1s | 5s = temp1
+
+    uadd16  r0, r7, r0              ; 1c+1 | 5c+5 = temp2
+    pkhbt   r9, r14, r12, lsl #16   ; i2 | i6
+    uadd16  r10, r11, r9            ; a
+    usub16  r9, r11, r9             ; b
+    pkhtb   r6, r12, r14, asr #16   ; i3 | i7
+
+    subs    r5, r5, #1<<31          ; i--
+
+    smulbt  r7, r5, r6              ; (ip[3] * cospi8sqrt2minus1) >> 16
+    smulwt  r11, r4, r6             ; (ip[3] * sinpi8sqrt2) >> 16
+    smulbb  r12, r5, r6             ; (ip[7] * cospi8sqrt2minus1) >> 16
+    smulwb  r14, r4, r6             ; (ip[7] * sinpi8sqrt2) >> 16
+
+    pkhtb   r7, r7, r12, asr #16    ; 3c | 7c
+    pkhbt   r11, r14, r11, lsl #16  ; 3s | 7s = temp1
+
+    uadd16  r6, r7, r6              ; 3c+3 | 7c+7 = temp2
+    usub16  r12, r8, r6             ; c (o1 | o5)
+    uadd16  r6, r11, r0             ; d (o3 | o7)
+    uadd16  r7, r10, r6             ; a+d
+
+    mov     r8, #4                  ; set up 4's
+    orr     r8, r8, #0x40000        ; 4|4
+
+    usub16  r6, r10, r6             ; a-d
+    uadd16  r6, r6, r8              ; a-d+4, 3|7
+    uadd16  r7, r7, r8              ; a+d+4, 0|4
+    uadd16  r10, r9, r12            ; b+c
+    usub16  r0, r9, r12             ; b-c
+    uadd16  r10, r10, r8            ; b+c+4, 1|5
+    uadd16  r8, r0, r8              ; b-c+4, 2|6
+
+    ldr     lr, [sp, #40]           ; dst stride
+
+    ldrb    r0, [r1]                ; pred p0
+    ldrb    r11, [r1, #1]           ; pred p1
+    ldrb    r12, [r1, #2]           ; pred p2
+
+    add     r0, r0, r7, asr #19     ; p0 + o0
+    add     r11, r11, r10, asr #19  ; p1 + o1
+    add     r12, r12, r8, asr #19   ; p2 + o2
+
+    usat    r0, #8, r0              ; d0 = clip8(p0 + o0)
+    usat    r11, #8, r11            ; d1 = clip8(p1 + o1)
+    usat    r12, #8, r12            ; d2 = clip8(p2 + o2)
+
+    add     r0, r0, r11, lsl #8     ; |--|--|d1|d0|
+
+    ldrb    r11, [r1, #3]           ; pred p3
+
+    add     r0, r0, r12, lsl #16    ; |--|d2|d1|d0|
+
+    add     r11, r11, r6, asr #19   ; p3 + o3
+
+    sxth    r7, r7                  ;
+    sxth    r10, r10                ;
+
+    usat    r11, #8, r11            ; d3 = clip8(p3 + o3)
+
+    sxth    r8, r8                  ;
+    sxth    r6, r6                  ;
+
+    add     r0, r0, r11, lsl #24    ; |d3|d2|d1|d0|
+
+    ldrb    r12, [r1, r2]!          ; pred p4
+    str     r0, [r3], lr
+    ldrb    r11, [r1, #1]           ; pred p5
+
+    add     r12, r12, r7, asr #3    ; p4 + o4
+    add     r11, r11, r10, asr #3   ; p5 + o5
+
+    usat    r12, #8, r12            ; d4 = clip8(p4 + o4)
+    usat    r11, #8, r11            ; d5 = clip8(p5 + o5)
+
+    ldrb    r7, [r1, #2]            ; pred p6
+    ldrb    r10, [r1, #3]           ; pred p6
+
+    add     r12, r12, r11, lsl #8   ; |--|--|d5|d4|
+
+    add     r7, r7, r8, asr #3      ; p6 + o6
+    add     r10, r10, r6, asr #3    ; p7 + o7
+
+    ldr     r0, [sp]                ; load input pointer
+
+    usat    r7, #8, r7              ; d6 = clip8(p6 + o6)
+    usat    r10, #8, r10            ; d7 = clip8(p7 + o7)
+
+    add     r12, r12, r7, lsl #16   ; |--|d6|d5|d4|
+    add     r12, r12, r10, lsl #24  ; |d7|d6|d5|d4|
+
+    str     r12, [r3], lr
+    add     r0, r0, #16
+    add     r1, r1, r2              ; pred + pitch
+
+    bcs loop2_dual
+
+    add     sp, sp, #4              ; idct_output buffer
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/iwalsh_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/iwalsh_v6.asm
new file mode 100644
index 0000000..31ef09c
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/iwalsh_v6.asm
@@ -0,0 +1,136 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT |vp8_short_inv_walsh4x4_v6|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+
+;short vp8_short_inv_walsh4x4_v6(short *input, short *mb_dqcoeff)
+|vp8_short_inv_walsh4x4_v6| PROC
+
+    stmdb       sp!, {r4 - r12, lr}
+
+    ldr         r2, [r0, #0]         ; [1  |  0]
+    ldr         r3, [r0, #4]         ; [3  |  2]
+    ldr         r4, [r0, #8]         ; [5  |  4]
+    ldr         r5, [r0, #12]        ; [7  |  6]
+    ldr         r6, [r0, #16]        ; [9  |  8]
+    ldr         r7, [r0, #20]        ; [11 | 10]
+    ldr         r8, [r0, #24]        ; [13 | 12]
+    ldr         r9, [r0, #28]        ; [15 | 14]
+
+    qadd16      r10, r2, r8          ; a1 [1+13  |  0+12]
+    qadd16      r11, r4, r6          ; b1 [5+9   |  4+8]
+    qsub16      r12, r4, r6          ; c1 [5-9   |  4-8]
+    qsub16      lr, r2, r8           ; d1 [1-13  |  0-12]
+
+    qadd16      r2, r10, r11         ; a1 + b1 [1  |  0]
+    qadd16      r4, r12, lr          ; c1 + d1 [5  |  4]
+    qsub16      r6, r10, r11         ; a1 - b1 [9  |  8]
+    qsub16      r8, lr, r12          ; d1 - c1 [13 | 12]
+
+    qadd16      r10, r3, r9          ; a1 [3+15  |  2+14]
+    qadd16      r11, r5, r7          ; b1 [7+11  |  6+10]
+    qsub16      r12, r5, r7          ; c1 [7-11  |  6-10]
+    qsub16      lr, r3, r9           ; d1 [3-15  |  2-14]
+
+    qadd16      r3, r10, r11         ; a1 + b1 [3  |  2]
+    qadd16      r5, r12, lr          ; c1 + d1 [7  |  6]
+    qsub16      r7, r10, r11         ; a1 - b1 [11 | 10]
+    qsub16      r9, lr, r12          ; d1 - c1 [15 | 14]
+
+    ; first transform complete
+
+    qsubaddx    r10, r2, r3          ; [c1|a1] [1-2   |   0+3]
+    qaddsubx    r11, r2, r3          ; [b1|d1] [1+2   |   0-3]
+    qsubaddx    r12, r4, r5          ; [c1|a1] [5-6   |   4+7]
+    qaddsubx    lr, r4, r5           ; [b1|d1] [5+6   |   4-7]
+
+    qaddsubx    r2, r10, r11         ; [b2|c2] [c1+d1 | a1-b1]
+    qaddsubx    r3, r11, r10         ; [a2|d2] [b1+a1 | d1-c1]
+    ldr         r10, c0x00030003
+    qaddsubx    r4, r12, lr          ; [b2|c2] [c1+d1 | a1-b1]
+    qaddsubx    r5, lr, r12          ; [a2|d2] [b1+a1 | d1-c1]
+
+    qadd16      r2, r2, r10          ; [b2+3|c2+3]
+    qadd16      r3, r3, r10          ; [a2+3|d2+3]
+    qadd16      r4, r4, r10          ; [b2+3|c2+3]
+    qadd16      r5, r5, r10          ; [a2+3|d2+3]
+
+    asr         r12, r3, #19         ; [0]
+    strh        r12, [r1], #32
+    asr         lr, r2, #19          ; [1]
+    strh        lr, [r1], #32
+    sxth        r2, r2
+    sxth        r3, r3
+    asr         r2, r2, #3           ; [2]
+    strh        r2, [r1], #32
+    asr         r3, r3, #3           ; [3]
+    strh        r3, [r1], #32
+
+    asr         r12, r5, #19         ; [4]
+    strh        r12, [r1], #32
+    asr         lr, r4, #19          ; [5]
+    strh        lr, [r1], #32
+    sxth        r4, r4
+    sxth        r5, r5
+    asr         r4, r4, #3           ; [6]
+    strh        r4, [r1], #32
+    asr         r5, r5, #3           ; [7]
+    strh        r5, [r1], #32
+
+    qsubaddx    r2, r6, r7           ; [c1|a1] [9-10  |  8+11]
+    qaddsubx    r3, r6, r7           ; [b1|d1] [9+10  |  8-11]
+    qsubaddx    r4, r8, r9           ; [c1|a1] [13-14 | 12+15]
+    qaddsubx    r5, r8, r9           ; [b1|d1] [13+14 | 12-15]
+
+    qaddsubx    r6, r2, r3           ; [b2|c2] [c1+d1 | a1-b1]
+    qaddsubx    r7, r3, r2           ; [a2|d2] [b1+a1 | d1-c1]
+    qaddsubx    r8, r4, r5           ; [b2|c2] [c1+d1 | a1-b1]
+    qaddsubx    r9, r5, r4           ; [a2|d2] [b1+a1 | d1-c1]
+
+    qadd16      r6, r6, r10          ; [b2+3|c2+3]
+    qadd16      r7, r7, r10          ; [a2+3|d2+3]
+    qadd16      r8, r8, r10          ; [b2+3|c2+3]
+    qadd16      r9, r9, r10          ; [a2+3|d2+3]
+
+    asr         r12, r7, #19         ; [8]
+    strh        r12, [r1], #32
+    asr         lr, r6, #19          ; [9]
+    strh        lr, [r1], #32
+    sxth        r6, r6
+    sxth        r7, r7
+    asr         r6, r6, #3           ; [10]
+    strh        r6, [r1], #32
+    asr         r7, r7, #3           ; [11]
+    strh        r7, [r1], #32
+
+    asr         r12, r9, #19         ; [12]
+    strh        r12, [r1], #32
+    asr         lr, r8, #19          ; [13]
+    strh        lr, [r1], #32
+    sxth        r8, r8
+    sxth        r9, r9
+    asr         r8, r8, #3           ; [14]
+    strh        r8, [r1], #32
+    asr         r9, r9, #3           ; [15]
+    strh        r9, [r1], #32
+
+    ldmia       sp!, {r4 - r12, pc}
+    ENDP        ; |vp8_short_inv_walsh4x4_v6|
+
+
+; Constant Pool
+c0x00030003 DCD 0x00030003
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/loopfilter_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/loopfilter_v6.asm
new file mode 100644
index 0000000..1cbbbcd
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/loopfilter_v6.asm
@@ -0,0 +1,1282 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT |vp8_loop_filter_horizontal_edge_armv6|
+    EXPORT |vp8_mbloop_filter_horizontal_edge_armv6|
+    EXPORT |vp8_loop_filter_vertical_edge_armv6|
+    EXPORT |vp8_mbloop_filter_vertical_edge_armv6|
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+
+    MACRO
+    TRANSPOSE_MATRIX $a0, $a1, $a2, $a3, $b0, $b1, $b2, $b3
+    ; input: $a0, $a1, $a2, $a3; output: $b0, $b1, $b2, $b3
+    ; a0: 03 02 01 00
+    ; a1: 13 12 11 10
+    ; a2: 23 22 21 20
+    ; a3: 33 32 31 30
+    ;     b3 b2 b1 b0
+
+    uxtb16      $b1, $a1                    ; xx 12 xx 10
+    uxtb16      $b0, $a0                    ; xx 02 xx 00
+    uxtb16      $b3, $a3                    ; xx 32 xx 30
+    uxtb16      $b2, $a2                    ; xx 22 xx 20
+    orr         $b1, $b0, $b1, lsl #8       ; 12 02 10 00
+    orr         $b3, $b2, $b3, lsl #8       ; 32 22 30 20
+
+    uxtb16      $a1, $a1, ror #8            ; xx 13 xx 11
+    uxtb16      $a3, $a3, ror #8            ; xx 33 xx 31
+    uxtb16      $a0, $a0, ror #8            ; xx 03 xx 01
+    uxtb16      $a2, $a2, ror #8            ; xx 23 xx 21
+    orr         $a0, $a0, $a1, lsl #8       ; 13 03 11 01
+    orr         $a2, $a2, $a3, lsl #8       ; 33 23 31 21
+
+    pkhtb       $b2, $b3, $b1, asr #16      ; 32 22 12 02   -- p1
+    pkhbt       $b0, $b1, $b3, lsl #16      ; 30 20 10 00   -- p3
+
+    pkhtb       $b3, $a2, $a0, asr #16      ; 33 23 13 03   -- p0
+    pkhbt       $b1, $a0, $a2, lsl #16      ; 31 21 11 01   -- p2
+    MEND
+
+
+src         RN  r0
+pstep       RN  r1
+count       RN  r5
+
+;r0     unsigned char *src_ptr,
+;r1     int src_pixel_step,
+;r2     const char *blimit,
+;r3     const char *limit,
+;stack  const char *thresh,
+;stack  int  count
+
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+|vp8_loop_filter_horizontal_edge_armv6| PROC
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+    stmdb       sp!, {r4 - r11, lr}
+
+    sub         src, src, pstep, lsl #2     ; move src pointer down by 4 lines
+    ldr         count, [sp, #40]            ; count for 8-in-parallel
+    ldr         r6, [sp, #36]               ; load thresh address
+    sub         sp, sp, #16                 ; create temp buffer
+
+    ldr         r9, [src], pstep            ; p3
+    ldrb        r4, [r2]                    ; blimit
+    ldr         r10, [src], pstep           ; p2
+    ldrb        r2, [r3]                    ; limit
+    ldr         r11, [src], pstep           ; p1
+    orr         r4, r4, r4, lsl #8
+    ldrb        r3, [r6]                    ; thresh
+    orr         r2, r2, r2, lsl #8
+    mov         count, count, lsl #1        ; 4-in-parallel
+    orr         r4, r4, r4, lsl #16
+    orr         r3, r3, r3, lsl #8
+    orr         r2, r2, r2, lsl #16
+    orr         r3, r3, r3, lsl #16
+
+|Hnext8|
+    ; vp8_filter_mask() function
+    ; calculate breakout conditions
+    ldr         r12, [src], pstep           ; p0
+
+    uqsub8      r6, r9, r10                 ; p3 - p2
+    uqsub8      r7, r10, r9                 ; p2 - p3
+    uqsub8      r8, r10, r11                ; p2 - p1
+    uqsub8      r10, r11, r10               ; p1 - p2
+
+    orr         r6, r6, r7                  ; abs (p3-p2)
+    orr         r8, r8, r10                 ; abs (p2-p1)
+    uqsub8      lr, r6, r2                  ; compare to limit. lr: vp8_filter_mask
+    uqsub8      r8, r8, r2                  ; compare to limit
+    uqsub8      r6, r11, r12                ; p1 - p0
+    orr         lr, lr, r8
+    uqsub8      r7, r12, r11                ; p0 - p1
+    ldr         r9, [src], pstep            ; q0
+    ldr         r10, [src], pstep           ; q1
+    orr         r6, r6, r7                  ; abs (p1-p0)
+    uqsub8      r7, r6, r2                  ; compare to limit
+    uqsub8      r8, r6, r3                  ; compare to thresh  -- save r8 for later
+    orr         lr, lr, r7
+
+    uqsub8      r6, r11, r10                ; p1 - q1
+    uqsub8      r7, r10, r11                ; q1 - p1
+    uqsub8      r11, r12, r9                ; p0 - q0
+    uqsub8      r12, r9, r12                ; q0 - p0
+    orr         r6, r6, r7                  ; abs (p1-q1)
+    ldr         r7, c0x7F7F7F7F
+    orr         r12, r11, r12               ; abs (p0-q0)
+    ldr         r11, [src], pstep           ; q2
+    uqadd8      r12, r12, r12               ; abs (p0-q0) * 2
+    and         r6, r7, r6, lsr #1          ; abs (p1-q1) / 2
+    uqsub8      r7, r9, r10                 ; q0 - q1
+    uqadd8      r12, r12, r6                ; abs (p0-q0)*2 + abs (p1-q1)/2
+    uqsub8      r6, r10, r9                 ; q1 - q0
+    uqsub8      r12, r12, r4                ; compare to flimit
+    uqsub8      r9, r11, r10                ; q2 - q1
+
+    orr         lr, lr, r12
+
+    ldr         r12, [src], pstep           ; q3
+    uqsub8      r10, r10, r11               ; q1 - q2
+    orr         r6, r7, r6                  ; abs (q1-q0)
+    orr         r10, r9, r10                ; abs (q2-q1)
+    uqsub8      r7, r6, r2                  ; compare to limit
+    uqsub8      r10, r10, r2                ; compare to limit
+    uqsub8      r6, r6, r3                  ; compare to thresh -- save r6 for later
+    orr         lr, lr, r7
+    orr         lr, lr, r10
+
+    uqsub8      r10, r12, r11               ; q3 - q2
+    uqsub8      r9, r11, r12                ; q2 - q3
+
+    mvn         r11, #0                     ; r11 == -1
+
+    orr         r10, r10, r9                ; abs (q3-q2)
+    uqsub8      r10, r10, r2                ; compare to limit
+
+    mov         r12, #0
+    orr         lr, lr, r10
+    sub         src, src, pstep, lsl #2
+
+    usub8       lr, r12, lr                 ; use usub8 instead of ssub8
+    sel         lr, r11, r12                ; filter mask: lr
+
+    cmp         lr, #0
+    beq         hskip_filter                 ; skip filtering
+
+    sub         src, src, pstep, lsl #1     ; move src pointer down by 6 lines
+
+    ;vp8_hevmask() function
+    ;calculate high edge variance
+    orr         r10, r6, r8                 ; calculate vp8_hevmask
+
+    ldr         r7, [src], pstep            ; p1
+
+    usub8       r10, r12, r10               ; use usub8 instead of ssub8
+    sel         r6, r12, r11                ; obtain vp8_hevmask: r6
+
+    ;vp8_filter() function
+    ldr         r8, [src], pstep            ; p0
+    ldr         r12, c0x80808080
+    ldr         r9, [src], pstep            ; q0
+    ldr         r10, [src], pstep           ; q1
+
+    eor         r7, r7, r12                 ; p1 offset to convert to a signed value
+    eor         r8, r8, r12                 ; p0 offset to convert to a signed value
+    eor         r9, r9, r12                 ; q0 offset to convert to a signed value
+    eor         r10, r10, r12               ; q1 offset to convert to a signed value
+
+    str         r9, [sp]                    ; store qs0 temporarily
+    str         r8, [sp, #4]                ; store ps0 temporarily
+    str         r10, [sp, #8]               ; store qs1 temporarily
+    str         r7, [sp, #12]               ; store ps1 temporarily
+
+    qsub8       r7, r7, r10                 ; vp8_signed_char_clamp(ps1-qs1)
+    qsub8       r8, r9, r8                  ; vp8_signed_char_clamp(vp8_filter + 3 * ( qs0 - ps0))
+
+    and         r7, r7, r6                  ; vp8_filter (r7) &= hev
+
+    qadd8       r7, r7, r8
+    ldr         r9, c0x03030303             ; r9 = 3 --modified for vp8
+
+    qadd8       r7, r7, r8
+    ldr         r10, c0x04040404
+
+    qadd8       r7, r7, r8
+    and         r7, r7, lr                  ; vp8_filter &= mask;
+
+    ;modify code for vp8 -- Filter1 = vp8_filter (r7)
+    qadd8       r8 , r7 , r9                ; Filter2 (r8) = vp8_signed_char_clamp(vp8_filter+3)
+    qadd8       r7 , r7 , r10               ; vp8_filter = vp8_signed_char_clamp(vp8_filter+4)
+
+    mov         r9, #0
+    shadd8      r8 , r8 , r9                ; Filter2 >>= 3
+    shadd8      r7 , r7 , r9                ; vp8_filter >>= 3
+    shadd8      r8 , r8 , r9
+    shadd8      r7 , r7 , r9
+    shadd8      lr , r8 , r9                ; lr: Filter2
+    shadd8      r7 , r7 , r9                ; r7: filter
+
+    ;usub8      lr, r8, r10                 ; s = (s==4)*-1
+    ;sel        lr, r11, r9
+    ;usub8      r8, r10, r8
+    ;sel        r8, r11, r9
+    ;and        r8, r8, lr                  ; -1 for each element that equals 4
+
+    ;calculate output
+    ;qadd8      lr, r8, r7                  ; u = vp8_signed_char_clamp(s + vp8_filter)
+
+    ldr         r8, [sp]                    ; load qs0
+    ldr         r9, [sp, #4]                ; load ps0
+
+    ldr         r10, c0x01010101
+
+    qsub8       r8 ,r8, r7                  ; u = vp8_signed_char_clamp(qs0 - vp8_filter)
+    qadd8       r9, r9, lr                  ; u = vp8_signed_char_clamp(ps0 + Filter2)
+
+    ;end of modification for vp8
+
+    mov         lr, #0
+    sadd8       r7, r7 , r10                ; vp8_filter += 1
+    shadd8      r7, r7, lr                  ; vp8_filter >>= 1
+
+    ldr         r11, [sp, #12]              ; load ps1
+    ldr         r10, [sp, #8]               ; load qs1
+
+    bic         r7, r7, r6                  ; vp8_filter &= ~hev
+    sub         src, src, pstep, lsl #2
+
+    qadd8       r11, r11, r7                ; u = vp8_signed_char_clamp(ps1 + vp8_filter)
+    qsub8       r10, r10,r7                 ; u = vp8_signed_char_clamp(qs1 - vp8_filter)
+
+    eor         r11, r11, r12               ; *op1 = u^0x80
+    str         r11, [src], pstep           ; store op1
+    eor         r9, r9, r12                 ; *op0 = u^0x80
+    str         r9, [src], pstep            ; store op0 result
+    eor         r8, r8, r12                 ; *oq0 = u^0x80
+    str         r8, [src], pstep            ; store oq0 result
+    eor         r10, r10, r12               ; *oq1 = u^0x80
+    str         r10, [src], pstep           ; store oq1
+
+    sub         src, src, pstep, lsl #1
+
+|hskip_filter|
+    add         src, src, #4
+    sub         src, src, pstep, lsl #2
+
+    subs        count, count, #1
+
+    ldrne       r9, [src], pstep            ; p3
+    ldrne       r10, [src], pstep           ; p2
+    ldrne       r11, [src], pstep           ; p1
+
+    bne         Hnext8
+
+    add         sp, sp, #16
+    ldmia       sp!, {r4 - r11, pc}
+    ENDP        ; |vp8_loop_filter_horizontal_edge_armv6|
+
+
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+|vp8_mbloop_filter_horizontal_edge_armv6| PROC
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+    stmdb       sp!, {r4 - r11, lr}
+
+    sub         src, src, pstep, lsl #2     ; move src pointer down by 4 lines
+    ldr         count, [sp, #40]            ; count for 8-in-parallel
+    ldr         r6, [sp, #36]               ; load thresh address
+    sub         sp, sp, #16                 ; create temp buffer
+
+    ldr         r9, [src], pstep            ; p3
+    ldrb        r4, [r2]                    ; blimit
+    ldr         r10, [src], pstep           ; p2
+    ldrb        r2, [r3]                    ; limit
+    ldr         r11, [src], pstep           ; p1
+    orr         r4, r4, r4, lsl #8
+    ldrb        r3, [r6]                    ; thresh
+    orr         r2, r2, r2, lsl #8
+    mov         count, count, lsl #1        ; 4-in-parallel
+    orr         r4, r4, r4, lsl #16
+    orr         r3, r3, r3, lsl #8
+    orr         r2, r2, r2, lsl #16
+    orr         r3, r3, r3, lsl #16
+
+|MBHnext8|
+
+    ; vp8_filter_mask() function
+    ; calculate breakout conditions
+    ldr         r12, [src], pstep           ; p0
+
+    uqsub8      r6, r9, r10                 ; p3 - p2
+    uqsub8      r7, r10, r9                 ; p2 - p3
+    uqsub8      r8, r10, r11                ; p2 - p1
+    uqsub8      r10, r11, r10               ; p1 - p2
+
+    orr         r6, r6, r7                  ; abs (p3-p2)
+    orr         r8, r8, r10                 ; abs (p2-p1)
+    uqsub8      lr, r6, r2                  ; compare to limit. lr: vp8_filter_mask
+    uqsub8      r8, r8, r2                  ; compare to limit
+
+    uqsub8      r6, r11, r12                ; p1 - p0
+    orr         lr, lr, r8
+    uqsub8      r7, r12, r11                ; p0 - p1
+    ldr         r9, [src], pstep            ; q0
+    ldr         r10, [src], pstep           ; q1
+    orr         r6, r6, r7                  ; abs (p1-p0)
+    uqsub8      r7, r6, r2                  ; compare to limit
+    uqsub8      r8, r6, r3                  ; compare to thresh  -- save r8 for later
+    orr         lr, lr, r7
+
+    uqsub8      r6, r11, r10                ; p1 - q1
+    uqsub8      r7, r10, r11                ; q1 - p1
+    uqsub8      r11, r12, r9                ; p0 - q0
+    uqsub8      r12, r9, r12                ; q0 - p0
+    orr         r6, r6, r7                  ; abs (p1-q1)
+    ldr         r7, c0x7F7F7F7F
+    orr         r12, r11, r12               ; abs (p0-q0)
+    ldr         r11, [src], pstep           ; q2
+    uqadd8      r12, r12, r12               ; abs (p0-q0) * 2
+    and         r6, r7, r6, lsr #1          ; abs (p1-q1) / 2
+    uqsub8      r7, r9, r10                 ; q0 - q1
+    uqadd8      r12, r12, r6                ; abs (p0-q0)*2 + abs (p1-q1)/2
+    uqsub8      r6, r10, r9                 ; q1 - q0
+    uqsub8      r12, r12, r4                ; compare to flimit
+    uqsub8      r9, r11, r10                ; q2 - q1
+
+    orr         lr, lr, r12
+
+    ldr         r12, [src], pstep           ; q3
+
+    uqsub8      r10, r10, r11               ; q1 - q2
+    orr         r6, r7, r6                  ; abs (q1-q0)
+    orr         r10, r9, r10                ; abs (q2-q1)
+    uqsub8      r7, r6, r2                  ; compare to limit
+    uqsub8      r10, r10, r2                ; compare to limit
+    uqsub8      r6, r6, r3                  ; compare to thresh -- save r6 for later
+    orr         lr, lr, r7
+    orr         lr, lr, r10
+
+    uqsub8      r10, r12, r11               ; q3 - q2
+    uqsub8      r9, r11, r12                ; q2 - q3
+
+    mvn         r11, #0                     ; r11 == -1
+
+    orr         r10, r10, r9                ; abs (q3-q2)
+    uqsub8      r10, r10, r2                ; compare to limit
+
+    mov         r12, #0
+
+    orr         lr, lr, r10
+
+    usub8       lr, r12, lr                 ; use usub8 instead of ssub8
+    sel         lr, r11, r12                ; filter mask: lr
+
+    cmp         lr, #0
+    beq         mbhskip_filter               ; skip filtering
+
+    ;vp8_hevmask() function
+    ;calculate high edge variance
+    sub         src, src, pstep, lsl #2     ; move src pointer down by 6 lines
+    sub         src, src, pstep, lsl #1
+
+    orr         r10, r6, r8
+    ldr         r7, [src], pstep            ; p1
+
+    usub8       r10, r12, r10
+    sel         r6, r12, r11                ; hev mask: r6
+
+    ;vp8_mbfilter() function
+    ;p2, q2 are only needed at the end. Don't need to load them in now.
+    ldr         r8, [src], pstep            ; p0
+    ldr         r12, c0x80808080
+    ldr         r9, [src], pstep            ; q0
+    ldr         r10, [src]                  ; q1
+
+    eor         r7, r7, r12                 ; ps1
+    eor         r8, r8, r12                 ; ps0
+    eor         r9, r9, r12                 ; qs0
+    eor         r10, r10, r12               ; qs1
+
+    qsub8       r12, r9, r8                 ; vp8_signed_char_clamp(vp8_filter + 3 * ( qs0 - ps0))
+    str         r7, [sp, #12]               ; store ps1 temporarily
+    qsub8       r7, r7, r10                 ; vp8_signed_char_clamp(ps1-qs1)
+    str         r10, [sp, #8]               ; store qs1 temporarily
+    qadd8       r7, r7, r12
+    str         r9, [sp]                    ; store qs0 temporarily
+    qadd8       r7, r7, r12
+    str         r8, [sp, #4]                ; store ps0 temporarily
+    qadd8       r7, r7, r12                 ; vp8_filter: r7
+
+    ldr         r10, c0x03030303            ; r10 = 3 --modified for vp8
+    ldr         r9, c0x04040404
+
+    and         r7, r7, lr                  ; vp8_filter &= mask (lr is free)
+
+    mov         r12, r7                     ; Filter2: r12
+    and         r12, r12, r6                ; Filter2 &= hev
+
+    ;modify code for vp8
+    ;save bottom 3 bits so that we round one side +4 and the other +3
+    qadd8       r8 , r12 , r9               ; Filter1 (r8) = vp8_signed_char_clamp(Filter2+4)
+    qadd8       r12 , r12 , r10             ; Filter2 (r12) = vp8_signed_char_clamp(Filter2+3)
+
+    mov         r10, #0
+    shadd8      r8 , r8 , r10               ; Filter1 >>= 3
+    shadd8      r12 , r12 , r10             ; Filter2 >>= 3
+    shadd8      r8 , r8 , r10
+    shadd8      r12 , r12 , r10
+    shadd8      r8 , r8 , r10               ; r8: Filter1
+    shadd8      r12 , r12 , r10             ; r12: Filter2
+
+    ldr         r9, [sp]                    ; load qs0
+    ldr         r11, [sp, #4]               ; load ps0
+
+    qsub8       r9 , r9, r8                 ; qs0 = vp8_signed_char_clamp(qs0 - Filter1)
+    qadd8       r11, r11, r12               ; ps0 = vp8_signed_char_clamp(ps0 + Filter2)
+
+    ;save bottom 3 bits so that we round one side +4 and the other +3
+    ;and            r8, r12, r10                ; s = Filter2 & 7 (s: r8)
+    ;qadd8      r12 , r12 , r9              ; Filter2 = vp8_signed_char_clamp(Filter2+4)
+    ;mov            r10, #0
+    ;shadd8     r12 , r12 , r10             ; Filter2 >>= 3
+    ;usub8      lr, r8, r9                  ; s = (s==4)*-1
+    ;sel            lr, r11, r10
+    ;shadd8     r12 , r12 , r10
+    ;usub8      r8, r9, r8
+    ;sel            r8, r11, r10
+    ;ldr            r9, [sp]                    ; load qs0
+    ;ldr            r11, [sp, #4]               ; load ps0
+    ;shadd8     r12 , r12 , r10
+    ;and            r8, r8, lr                  ; -1 for each element that equals 4
+    ;qadd8      r10, r8, r12                ; u = vp8_signed_char_clamp(s + Filter2)
+    ;qsub8      r9 , r9, r12                ; qs0 = vp8_signed_char_clamp(qs0 - Filter2)
+    ;qadd8      r11, r11, r10               ; ps0 = vp8_signed_char_clamp(ps0 + u)
+
+    ;end of modification for vp8
+
+    bic         r12, r7, r6                 ; vp8_filter &= ~hev    ( r6 is free)
+    ;mov        r12, r7
+
+    ;roughly 3/7th difference across boundary
+    mov         lr, #0x1b                   ; 27
+    mov         r7, #0x3f                   ; 63
+
+    sxtb16      r6, r12
+    sxtb16      r10, r12, ror #8
+    smlabb      r8, r6, lr, r7
+    smlatb      r6, r6, lr, r7
+    smlabb      r7, r10, lr, r7
+    smultb      r10, r10, lr
+    ssat        r8, #8, r8, asr #7
+    ssat        r6, #8, r6, asr #7
+    add         r10, r10, #63
+    ssat        r7, #8, r7, asr #7
+    ssat        r10, #8, r10, asr #7
+
+    ldr         lr, c0x80808080
+
+    pkhbt       r6, r8, r6, lsl #16
+    pkhbt       r10, r7, r10, lsl #16
+    uxtb16      r6, r6
+    uxtb16      r10, r10
+
+    sub         src, src, pstep
+
+    orr         r10, r6, r10, lsl #8        ; u = vp8_signed_char_clamp((63 + Filter2 * 27)>>7)
+
+    qsub8       r8, r9, r10                 ; s = vp8_signed_char_clamp(qs0 - u)
+    qadd8       r10, r11, r10               ; s = vp8_signed_char_clamp(ps0 + u)
+    eor         r8, r8, lr                  ; *oq0 = s^0x80
+    str         r8, [src]                   ; store *oq0
+    sub         src, src, pstep
+    eor         r10, r10, lr                ; *op0 = s^0x80
+    str         r10, [src]                  ; store *op0
+
+    ;roughly 2/7th difference across boundary
+    mov         lr, #0x12                   ; 18
+    mov         r7, #0x3f                   ; 63
+
+    sxtb16      r6, r12
+    sxtb16      r10, r12, ror #8
+    smlabb      r8, r6, lr, r7
+    smlatb      r6, r6, lr, r7
+    smlabb      r9, r10, lr, r7
+    smlatb      r10, r10, lr, r7
+    ssat        r8, #8, r8, asr #7
+    ssat        r6, #8, r6, asr #7
+    ssat        r9, #8, r9, asr #7
+    ssat        r10, #8, r10, asr #7
+
+    ldr         lr, c0x80808080
+
+    pkhbt       r6, r8, r6, lsl #16
+    pkhbt       r10, r9, r10, lsl #16
+
+    ldr         r9, [sp, #8]                ; load qs1
+    ldr         r11, [sp, #12]              ; load ps1
+
+    uxtb16      r6, r6
+    uxtb16      r10, r10
+
+    sub         src, src, pstep
+
+    orr         r10, r6, r10, lsl #8        ; u = vp8_signed_char_clamp((63 + Filter2 * 18)>>7)
+
+    qadd8       r11, r11, r10               ; s = vp8_signed_char_clamp(ps1 + u)
+    qsub8       r8, r9, r10                 ; s = vp8_signed_char_clamp(qs1 - u)
+    eor         r11, r11, lr                ; *op1 = s^0x80
+    str         r11, [src], pstep           ; store *op1
+    eor         r8, r8, lr                  ; *oq1 = s^0x80
+    add         src, src, pstep, lsl #1
+
+    mov         r7, #0x3f                   ; 63
+
+    str         r8, [src], pstep            ; store *oq1
+
+    ;roughly 1/7th difference across boundary
+    mov         lr, #0x9                    ; 9
+    ldr         r9, [src]                   ; load q2
+
+    sxtb16      r6, r12
+    sxtb16      r10, r12, ror #8
+    smlabb      r8, r6, lr, r7
+    smlatb      r6, r6, lr, r7
+    smlabb      r12, r10, lr, r7
+    smlatb      r10, r10, lr, r7
+    ssat        r8, #8, r8, asr #7
+    ssat        r6, #8, r6, asr #7
+    ssat        r12, #8, r12, asr #7
+    ssat        r10, #8, r10, asr #7
+
+    sub         src, src, pstep, lsl #2
+
+    pkhbt       r6, r8, r6, lsl #16
+    pkhbt       r10, r12, r10, lsl #16
+
+    sub         src, src, pstep
+    ldr         lr, c0x80808080
+
+    ldr         r11, [src]                  ; load p2
+
+    uxtb16      r6, r6
+    uxtb16      r10, r10
+
+    eor         r9, r9, lr
+    eor         r11, r11, lr
+
+    orr         r10, r6, r10, lsl #8        ; u = vp8_signed_char_clamp((63 + Filter2 * 9)>>7)
+
+    qadd8       r8, r11, r10                ; s = vp8_signed_char_clamp(ps2 + u)
+    qsub8       r10, r9, r10                ; s = vp8_signed_char_clamp(qs2 - u)
+    eor         r8, r8, lr                  ; *op2 = s^0x80
+    str         r8, [src], pstep, lsl #2    ; store *op2
+    add         src, src, pstep
+    eor         r10, r10, lr                ; *oq2 = s^0x80
+    str         r10, [src], pstep, lsl #1   ; store *oq2
+
+|mbhskip_filter|
+    add         src, src, #4
+    sub         src, src, pstep, lsl #3
+    subs        count, count, #1
+
+    ldrne       r9, [src], pstep            ; p3
+    ldrne       r10, [src], pstep           ; p2
+    ldrne       r11, [src], pstep           ; p1
+
+    bne         MBHnext8
+
+    add         sp, sp, #16
+    ldmia       sp!, {r4 - r11, pc}
+    ENDP        ; |vp8_mbloop_filter_horizontal_edge_armv6|
+
+
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+|vp8_loop_filter_vertical_edge_armv6| PROC
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+    stmdb       sp!, {r4 - r11, lr}
+
+    sub         src, src, #4                ; move src pointer down by 4
+    ldr         count, [sp, #40]            ; count for 8-in-parallel
+    ldr         r12, [sp, #36]              ; load thresh address
+    sub         sp, sp, #16                 ; create temp buffer
+
+    ldr         r6, [src], pstep            ; load source data
+    ldrb        r4, [r2]                    ; blimit
+    ldr         r7, [src], pstep
+    ldrb        r2, [r3]                    ; limit
+    ldr         r8, [src], pstep
+    orr         r4, r4, r4, lsl #8
+    ldrb        r3, [r12]                   ; thresh
+    orr         r2, r2, r2, lsl #8
+    ldr         lr, [src], pstep
+    mov         count, count, lsl #1        ; 4-in-parallel
+    orr         r4, r4, r4, lsl #16
+    orr         r3, r3, r3, lsl #8
+    orr         r2, r2, r2, lsl #16
+    orr         r3, r3, r3, lsl #16
+
+|Vnext8|
+
+    ; vp8_filter_mask() function
+    ; calculate breakout conditions
+    ; transpose the source data for 4-in-parallel operation
+    TRANSPOSE_MATRIX r6, r7, r8, lr, r9, r10, r11, r12
+
+    uqsub8      r7, r9, r10                 ; p3 - p2
+    uqsub8      r8, r10, r9                 ; p2 - p3
+    uqsub8      r9, r10, r11                ; p2 - p1
+    uqsub8      r10, r11, r10               ; p1 - p2
+    orr         r7, r7, r8                  ; abs (p3-p2)
+    orr         r10, r9, r10                ; abs (p2-p1)
+    uqsub8      lr, r7, r2                  ; compare to limit. lr: vp8_filter_mask
+    uqsub8      r10, r10, r2                ; compare to limit
+
+    sub         src, src, pstep, lsl #2     ; move src pointer down by 4 lines
+
+    orr         lr, lr, r10
+
+    uqsub8      r6, r11, r12                ; p1 - p0
+    uqsub8      r7, r12, r11                ; p0 - p1
+    add         src, src, #4                ; move src pointer up by 4
+    orr         r6, r6, r7                  ; abs (p1-p0)
+    str         r11, [sp, #12]              ; save p1
+    uqsub8      r10, r6, r2                 ; compare to limit
+    uqsub8      r11, r6, r3                 ; compare to thresh
+    orr         lr, lr, r10
+
+    ; transpose uses 8 regs(r6 - r12 and lr). Need to save reg value now
+    ; transpose the source data for 4-in-parallel operation
+    ldr         r6, [src], pstep            ; load source data
+    str         r11, [sp]                   ; push r11 to stack
+    ldr         r7, [src], pstep
+    str         r12, [sp, #4]               ; save current reg before load q0 - q3 data
+    ldr         r8, [src], pstep
+    str         lr, [sp, #8]
+    ldr         lr, [src], pstep
+
+    TRANSPOSE_MATRIX r6, r7, r8, lr, r9, r10, r11, r12
+
+    ldr         lr, [sp, #8]                ; load back (f)limit accumulator
+
+    uqsub8      r6, r12, r11                ; q3 - q2
+    uqsub8      r7, r11, r12                ; q2 - q3
+    uqsub8      r12, r11, r10               ; q2 - q1
+    uqsub8      r11, r10, r11               ; q1 - q2
+    orr         r6, r6, r7                  ; abs (q3-q2)
+    orr         r7, r12, r11                ; abs (q2-q1)
+    uqsub8      r6, r6, r2                  ; compare to limit
+    uqsub8      r7, r7, r2                  ; compare to limit
+    ldr         r11, [sp, #4]               ; load back p0
+    ldr         r12, [sp, #12]              ; load back p1
+    orr         lr, lr, r6
+    orr         lr, lr, r7
+
+    uqsub8      r6, r11, r9                 ; p0 - q0
+    uqsub8      r7, r9, r11                 ; q0 - p0
+    uqsub8      r8, r12, r10                ; p1 - q1
+    uqsub8      r11, r10, r12               ; q1 - p1
+    orr         r6, r6, r7                  ; abs (p0-q0)
+    ldr         r7, c0x7F7F7F7F
+    orr         r8, r8, r11                 ; abs (p1-q1)
+    uqadd8      r6, r6, r6                  ; abs (p0-q0) * 2
+    and         r8, r7, r8, lsr #1          ; abs (p1-q1) / 2
+    uqsub8      r11, r10, r9                ; q1 - q0
+    uqadd8      r6, r8, r6                  ; abs (p0-q0)*2 + abs (p1-q1)/2
+    uqsub8      r12, r9, r10                ; q0 - q1
+    uqsub8      r6, r6, r4                  ; compare to flimit
+
+    orr         r9, r11, r12                ; abs (q1-q0)
+    uqsub8      r8, r9, r2                  ; compare to limit
+    uqsub8      r10, r9, r3                 ; compare to thresh
+    orr         lr, lr, r6
+    orr         lr, lr, r8
+
+    mvn         r11, #0                     ; r11 == -1
+    mov         r12, #0
+
+    usub8       lr, r12, lr
+    ldr         r9, [sp]                    ; load the compared result
+    sel         lr, r11, r12                ; filter mask: lr
+
+    cmp         lr, #0
+    beq         vskip_filter                 ; skip filtering
+
+    ;vp8_hevmask() function
+    ;calculate high edge variance
+
+    sub         src, src, pstep, lsl #2     ; move src pointer down by 4 lines
+
+    orr         r9, r9, r10
+
+    ldrh        r7, [src, #-2]
+    ldrh        r8, [src], pstep
+
+    usub8       r9, r12, r9
+    sel         r6, r12, r11                ; hev mask: r6
+
+    ;vp8_filter() function
+    ; load soure data to r6, r11, r12, lr
+    ldrh        r9, [src, #-2]
+    ldrh        r10, [src], pstep
+
+    pkhbt       r12, r7, r8, lsl #16
+
+    ldrh        r7, [src, #-2]
+    ldrh        r8, [src], pstep
+
+    pkhbt       r11, r9, r10, lsl #16
+
+    ldrh        r9, [src, #-2]
+    ldrh        r10, [src], pstep
+
+    ; Transpose needs 8 regs(r6 - r12, and lr). Save r6 and lr first
+    str         r6, [sp]
+    str         lr, [sp, #4]
+
+    pkhbt       r6, r7, r8, lsl #16
+    pkhbt       lr, r9, r10, lsl #16
+
+    ;transpose r12, r11, r6, lr to r7, r8, r9, r10
+    TRANSPOSE_MATRIX r12, r11, r6, lr, r7, r8, r9, r10
+
+    ;load back hev_mask r6 and filter_mask lr
+    ldr         r12, c0x80808080
+    ldr         r6, [sp]
+    ldr         lr, [sp, #4]
+
+    eor         r7, r7, r12                 ; p1 offset to convert to a signed value
+    eor         r8, r8, r12                 ; p0 offset to convert to a signed value
+    eor         r9, r9, r12                 ; q0 offset to convert to a signed value
+    eor         r10, r10, r12               ; q1 offset to convert to a signed value
+
+    str         r9, [sp]                    ; store qs0 temporarily
+    str         r8, [sp, #4]                ; store ps0 temporarily
+    str         r10, [sp, #8]               ; store qs1 temporarily
+    str         r7, [sp, #12]               ; store ps1 temporarily
+
+    qsub8       r7, r7, r10                 ; vp8_signed_char_clamp(ps1-qs1)
+    qsub8       r8, r9, r8                  ; vp8_signed_char_clamp(vp8_filter + 3 * ( qs0 - ps0))
+
+    and         r7, r7, r6                  ;  vp8_filter (r7) &= hev (r7 : filter)
+
+    qadd8       r7, r7, r8
+    ldr         r9, c0x03030303             ; r9 = 3 --modified for vp8
+
+    qadd8       r7, r7, r8
+    ldr         r10, c0x04040404
+
+    qadd8       r7, r7, r8
+    ;mvn         r11, #0                     ; r11 == -1
+
+    and         r7, r7, lr                  ; vp8_filter &= mask
+
+    ;modify code for vp8 -- Filter1 = vp8_filter (r7)
+    qadd8       r8 , r7 , r9                ; Filter2 (r8) = vp8_signed_char_clamp(vp8_filter+3)
+    qadd8       r7 , r7 , r10               ; vp8_filter = vp8_signed_char_clamp(vp8_filter+4)
+
+    mov         r9, #0
+    shadd8      r8 , r8 , r9                ; Filter2 >>= 3
+    shadd8      r7 , r7 , r9                ; vp8_filter >>= 3
+    shadd8      r8 , r8 , r9
+    shadd8      r7 , r7 , r9
+    shadd8      lr , r8 , r9                ; lr: filter2
+    shadd8      r7 , r7 , r9                ; r7: filter
+
+    ;usub8      lr, r8, r10                 ; s = (s==4)*-1
+    ;sel            lr, r11, r9
+    ;usub8      r8, r10, r8
+    ;sel            r8, r11, r9
+    ;and            r8, r8, lr                  ; -1 for each element that equals 4 -- r8: s
+
+    ;calculate output
+    ;qadd8      lr, r8, r7                  ; u = vp8_signed_char_clamp(s + vp8_filter)
+
+    ldr         r8, [sp]                    ; load qs0
+    ldr         r9, [sp, #4]                ; load ps0
+
+    ldr         r10, c0x01010101
+
+    qsub8       r8, r8, r7                  ; u = vp8_signed_char_clamp(qs0 - vp8_filter)
+    qadd8       r9, r9, lr                  ; u = vp8_signed_char_clamp(ps0 + Filter2)
+    ;end of modification for vp8
+
+    eor         r8, r8, r12
+    eor         r9, r9, r12
+
+    mov         lr, #0
+
+    sadd8       r7, r7, r10
+    shadd8      r7, r7, lr
+
+    ldr         r10, [sp, #8]               ; load qs1
+    ldr         r11, [sp, #12]              ; load ps1
+
+    bic         r7, r7, r6                  ; r7: vp8_filter
+
+    qsub8       r10 , r10, r7               ; u = vp8_signed_char_clamp(qs1 - vp8_filter)
+    qadd8       r11, r11, r7                ; u = vp8_signed_char_clamp(ps1 + vp8_filter)
+    eor         r10, r10, r12
+    eor         r11, r11, r12
+
+    sub         src, src, pstep, lsl #2
+
+    ;we can use TRANSPOSE_MATRIX macro to transpose output - input: q1, q0, p0, p1
+    ;output is b0, b1, b2, b3
+    ;b0: 03 02 01 00
+    ;b1: 13 12 11 10
+    ;b2: 23 22 21 20
+    ;b3: 33 32 31 30
+    ;    p1 p0 q0 q1
+    ;   (a3 a2 a1 a0)
+    TRANSPOSE_MATRIX r11, r9, r8, r10, r6, r7, r12, lr
+
+    strh        r6, [src, #-2]              ; store the result
+    mov         r6, r6, lsr #16
+    strh        r6, [src], pstep
+
+    strh        r7, [src, #-2]
+    mov         r7, r7, lsr #16
+    strh        r7, [src], pstep
+
+    strh        r12, [src, #-2]
+    mov         r12, r12, lsr #16
+    strh        r12, [src], pstep
+
+    strh        lr, [src, #-2]
+    mov         lr, lr, lsr #16
+    strh        lr, [src], pstep
+
+|vskip_filter|
+    sub         src, src, #4
+    subs        count, count, #1
+
+    ldrne       r6, [src], pstep            ; load source data
+    ldrne       r7, [src], pstep
+    ldrne       r8, [src], pstep
+    ldrne       lr, [src], pstep
+
+    bne         Vnext8
+
+    add         sp, sp, #16
+
+    ldmia       sp!, {r4 - r11, pc}
+    ENDP        ; |vp8_loop_filter_vertical_edge_armv6|
+
+
+
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+|vp8_mbloop_filter_vertical_edge_armv6| PROC
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+    stmdb       sp!, {r4 - r11, lr}
+
+    sub         src, src, #4                ; move src pointer down by 4
+    ldr         count, [sp, #40]            ; count for 8-in-parallel
+    ldr         r12, [sp, #36]              ; load thresh address
+    pld         [src, #23]                  ; preload for next block
+    sub         sp, sp, #16                 ; create temp buffer
+
+    ldr         r6, [src], pstep            ; load source data
+    ldrb        r4, [r2]                    ; blimit
+    pld         [src, #23]
+    ldr         r7, [src], pstep
+    ldrb        r2, [r3]                    ; limit
+    pld         [src, #23]
+    ldr         r8, [src], pstep
+    orr         r4, r4, r4, lsl #8
+    ldrb        r3, [r12]                   ; thresh
+    orr         r2, r2, r2, lsl #8
+    pld         [src, #23]
+    ldr         lr, [src], pstep
+    mov         count, count, lsl #1        ; 4-in-parallel
+    orr         r4, r4, r4, lsl #16
+    orr         r3, r3, r3, lsl #8
+    orr         r2, r2, r2, lsl #16
+    orr         r3, r3, r3, lsl #16
+
+|MBVnext8|
+    ; vp8_filter_mask() function
+    ; calculate breakout conditions
+    ; transpose the source data for 4-in-parallel operation
+    TRANSPOSE_MATRIX r6, r7, r8, lr, r9, r10, r11, r12
+
+    uqsub8      r7, r9, r10                 ; p3 - p2
+    uqsub8      r8, r10, r9                 ; p2 - p3
+    uqsub8      r9, r10, r11                ; p2 - p1
+    uqsub8      r10, r11, r10               ; p1 - p2
+    orr         r7, r7, r8                  ; abs (p3-p2)
+    orr         r10, r9, r10                ; abs (p2-p1)
+    uqsub8      lr, r7, r2                  ; compare to limit. lr: vp8_filter_mask
+    uqsub8      r10, r10, r2                ; compare to limit
+
+    sub         src, src, pstep, lsl #2     ; move src pointer down by 4 lines
+
+    orr         lr, lr, r10
+
+    uqsub8      r6, r11, r12                ; p1 - p0
+    uqsub8      r7, r12, r11                ; p0 - p1
+    add         src, src, #4                ; move src pointer up by 4
+    orr         r6, r6, r7                  ; abs (p1-p0)
+    str         r11, [sp, #12]              ; save p1
+    uqsub8      r10, r6, r2                 ; compare to limit
+    uqsub8      r11, r6, r3                 ; compare to thresh
+    orr         lr, lr, r10
+
+    ; transpose uses 8 regs(r6 - r12 and lr). Need to save reg value now
+    ; transpose the source data for 4-in-parallel operation
+    ldr         r6, [src], pstep            ; load source data
+    str         r11, [sp]                   ; push r11 to stack
+    ldr         r7, [src], pstep
+    str         r12, [sp, #4]               ; save current reg before load q0 - q3 data
+    ldr         r8, [src], pstep
+    str         lr, [sp, #8]
+    ldr         lr, [src], pstep
+
+
+    TRANSPOSE_MATRIX r6, r7, r8, lr, r9, r10, r11, r12
+
+    ldr         lr, [sp, #8]                ; load back (f)limit accumulator
+
+    uqsub8      r6, r12, r11                ; q3 - q2
+    uqsub8      r7, r11, r12                ; q2 - q3
+    uqsub8      r12, r11, r10               ; q2 - q1
+    uqsub8      r11, r10, r11               ; q1 - q2
+    orr         r6, r6, r7                  ; abs (q3-q2)
+    orr         r7, r12, r11                ; abs (q2-q1)
+    uqsub8      r6, r6, r2                  ; compare to limit
+    uqsub8      r7, r7, r2                  ; compare to limit
+    ldr         r11, [sp, #4]               ; load back p0
+    ldr         r12, [sp, #12]              ; load back p1
+    orr         lr, lr, r6
+    orr         lr, lr, r7
+
+    uqsub8      r6, r11, r9                 ; p0 - q0
+    uqsub8      r7, r9, r11                 ; q0 - p0
+    uqsub8      r8, r12, r10                ; p1 - q1
+    uqsub8      r11, r10, r12               ; q1 - p1
+    orr         r6, r6, r7                  ; abs (p0-q0)
+    ldr         r7, c0x7F7F7F7F
+    orr         r8, r8, r11                 ; abs (p1-q1)
+    uqadd8      r6, r6, r6                  ; abs (p0-q0) * 2
+    and         r8, r7, r8, lsr #1          ; abs (p1-q1) / 2
+    uqsub8      r11, r10, r9                ; q1 - q0
+    uqadd8      r6, r8, r6                  ; abs (p0-q0)*2 + abs (p1-q1)/2
+    uqsub8      r12, r9, r10                ; q0 - q1
+    uqsub8      r6, r6, r4                  ; compare to flimit
+
+    orr         r9, r11, r12                ; abs (q1-q0)
+    uqsub8      r8, r9, r2                  ; compare to limit
+    uqsub8      r10, r9, r3                 ; compare to thresh
+    orr         lr, lr, r6
+    orr         lr, lr, r8
+
+    mvn         r11, #0                     ; r11 == -1
+    mov         r12, #0
+
+    usub8       lr, r12, lr
+    ldr         r9, [sp]                    ; load the compared result
+    sel         lr, r11, r12                ; filter mask: lr
+
+    cmp         lr, #0
+    beq         mbvskip_filter               ; skip filtering
+
+
+
+    ;vp8_hevmask() function
+    ;calculate high edge variance
+
+    sub         src, src, pstep, lsl #2     ; move src pointer down by 4 lines
+
+    orr         r9, r9, r10
+
+    ldrh        r7, [src, #-2]
+    ldrh        r8, [src], pstep
+
+    usub8       r9, r12, r9
+    sel         r6, r12, r11                ; hev mask: r6
+
+
+    ; vp8_mbfilter() function
+    ; p2, q2 are only needed at the end. Don't need to load them in now.
+    ; Transpose needs 8 regs(r6 - r12, and lr). Save r6 and lr first
+    ; load soure data to r6, r11, r12, lr
+    ldrh        r9, [src, #-2]
+    ldrh        r10, [src], pstep
+
+    pkhbt       r12, r7, r8, lsl #16
+
+    ldrh        r7, [src, #-2]
+    ldrh        r8, [src], pstep
+
+    pkhbt       r11, r9, r10, lsl #16
+
+    ldrh        r9, [src, #-2]
+    ldrh        r10, [src], pstep
+
+    str         r6, [sp]                    ; save r6
+    str         lr, [sp, #4]                ; save lr
+
+    pkhbt       r6, r7, r8, lsl #16
+    pkhbt       lr, r9, r10, lsl #16
+
+    ;transpose r12, r11, r6, lr to p1, p0, q0, q1
+    TRANSPOSE_MATRIX r12, r11, r6, lr, r7, r8, r9, r10
+
+    ;load back hev_mask r6 and filter_mask lr
+    ldr         r12, c0x80808080
+    ldr         r6, [sp]
+    ldr         lr, [sp, #4]
+
+    eor         r7, r7, r12                 ; ps1
+    eor         r8, r8, r12                 ; ps0
+    eor         r9, r9, r12                 ; qs0
+    eor         r10, r10, r12               ; qs1
+
+    qsub8       r12, r9, r8                 ; vp8_signed_char_clamp(vp8_filter + 3 * ( qs0 - ps0))
+    str         r7, [sp, #12]               ; store ps1 temporarily
+    qsub8       r7, r7, r10                 ; vp8_signed_char_clamp(ps1-qs1)
+    str         r10, [sp, #8]               ; store qs1 temporarily
+    qadd8       r7, r7, r12
+    str         r9, [sp]                    ; store qs0 temporarily
+    qadd8       r7, r7, r12
+    str         r8, [sp, #4]                ; store ps0 temporarily
+    qadd8       r7, r7, r12                 ; vp8_filter: r7
+
+    ldr         r10, c0x03030303            ; r10 = 3 --modified for vp8
+    ldr         r9, c0x04040404
+    ;mvn         r11, #0                     ; r11 == -1
+
+    and         r7, r7, lr                  ; vp8_filter &= mask (lr is free)
+
+    mov         r12, r7                     ; Filter2: r12
+    and         r12, r12, r6                ; Filter2 &= hev
+
+    ;modify code for vp8
+    ;save bottom 3 bits so that we round one side +4 and the other +3
+    qadd8       r8 , r12 , r9               ; Filter1 (r8) = vp8_signed_char_clamp(Filter2+4)
+    qadd8       r12 , r12 , r10             ; Filter2 (r12) = vp8_signed_char_clamp(Filter2+3)
+
+    mov         r10, #0
+    shadd8      r8 , r8 , r10               ; Filter1 >>= 3
+    shadd8      r12 , r12 , r10             ; Filter2 >>= 3
+    shadd8      r8 , r8 , r10
+    shadd8      r12 , r12 , r10
+    shadd8      r8 , r8 , r10               ; r8: Filter1
+    shadd8      r12 , r12 , r10             ; r12: Filter2
+
+    ldr         r9, [sp]                    ; load qs0
+    ldr         r11, [sp, #4]               ; load ps0
+
+    qsub8       r9 , r9, r8                 ; qs0 = vp8_signed_char_clamp(qs0 - Filter1)
+    qadd8       r11, r11, r12               ; ps0 = vp8_signed_char_clamp(ps0 + Filter2)
+
+    ;save bottom 3 bits so that we round one side +4 and the other +3
+    ;and            r8, r12, r10                ; s = Filter2 & 7 (s: r8)
+    ;qadd8      r12 , r12 , r9              ; Filter2 = vp8_signed_char_clamp(Filter2+4)
+    ;mov            r10, #0
+    ;shadd8     r12 , r12 , r10             ; Filter2 >>= 3
+    ;usub8      lr, r8, r9                  ; s = (s==4)*-1
+    ;sel            lr, r11, r10
+    ;shadd8     r12 , r12 , r10
+    ;usub8      r8, r9, r8
+    ;sel            r8, r11, r10
+    ;ldr            r9, [sp]                    ; load qs0
+    ;ldr            r11, [sp, #4]               ; load ps0
+    ;shadd8     r12 , r12 , r10
+    ;and            r8, r8, lr                  ; -1 for each element that equals 4
+    ;qadd8      r10, r8, r12                ; u = vp8_signed_char_clamp(s + Filter2)
+    ;qsub8      r9 , r9, r12                ; qs0 = vp8_signed_char_clamp(qs0 - Filter2)
+    ;qadd8      r11, r11, r10               ; ps0 = vp8_signed_char_clamp(ps0 + u)
+
+    ;end of modification for vp8
+
+    bic         r12, r7, r6                 ;vp8_filter &= ~hev    ( r6 is free)
+    ;mov            r12, r7
+
+    ;roughly 3/7th difference across boundary
+    mov         lr, #0x1b                   ; 27
+    mov         r7, #0x3f                   ; 63
+
+    sxtb16      r6, r12
+    sxtb16      r10, r12, ror #8
+    smlabb      r8, r6, lr, r7
+    smlatb      r6, r6, lr, r7
+    smlabb      r7, r10, lr, r7
+    smultb      r10, r10, lr
+    ssat        r8, #8, r8, asr #7
+    ssat        r6, #8, r6, asr #7
+    add         r10, r10, #63
+    ssat        r7, #8, r7, asr #7
+    ssat        r10, #8, r10, asr #7
+
+    ldr         lr, c0x80808080
+
+    pkhbt       r6, r8, r6, lsl #16
+    pkhbt       r10, r7, r10, lsl #16
+    uxtb16      r6, r6
+    uxtb16      r10, r10
+
+    sub         src, src, pstep, lsl #2     ; move src pointer down by 4 lines
+
+    orr         r10, r6, r10, lsl #8        ; u = vp8_signed_char_clamp((63 + Filter2 * 27)>>7)
+
+    qsub8       r8, r9, r10                 ; s = vp8_signed_char_clamp(qs0 - u)
+    qadd8       r10, r11, r10               ; s = vp8_signed_char_clamp(ps0 + u)
+    eor         r8, r8, lr                  ; *oq0 = s^0x80
+    eor         r10, r10, lr                ; *op0 = s^0x80
+
+    strb        r10, [src, #-1]             ; store op0 result
+    strb        r8, [src], pstep            ; store oq0 result
+    mov         r10, r10, lsr #8
+    mov         r8, r8, lsr #8
+    strb        r10, [src, #-1]
+    strb        r8, [src], pstep
+    mov         r10, r10, lsr #8
+    mov         r8, r8, lsr #8
+    strb        r10, [src, #-1]
+    strb        r8, [src], pstep
+    mov         r10, r10, lsr #8
+    mov         r8, r8, lsr #8
+    strb        r10, [src, #-1]
+    strb        r8, [src], pstep
+
+    ;roughly 2/7th difference across boundary
+    mov         lr, #0x12                   ; 18
+    mov         r7, #0x3f                   ; 63
+
+    sxtb16      r6, r12
+    sxtb16      r10, r12, ror #8
+    smlabb      r8, r6, lr, r7
+    smlatb      r6, r6, lr, r7
+    smlabb      r9, r10, lr, r7
+
+    smlatb      r10, r10, lr, r7
+    ssat        r8, #8, r8, asr #7
+    ssat        r6, #8, r6, asr #7
+    ssat        r9, #8, r9, asr #7
+    ssat        r10, #8, r10, asr #7
+
+    sub         src, src, pstep, lsl #2     ; move src pointer down by 4 lines
+
+    pkhbt       r6, r8, r6, lsl #16
+    pkhbt       r10, r9, r10, lsl #16
+
+    ldr         r9, [sp, #8]                ; load qs1
+    ldr         r11, [sp, #12]              ; load ps1
+    ldr         lr, c0x80808080
+
+    uxtb16      r6, r6
+    uxtb16      r10, r10
+
+    add         src, src, #2
+
+    orr         r10, r6, r10, lsl #8        ; u = vp8_signed_char_clamp((63 + Filter2 * 18)>>7)
+
+    qsub8       r8, r9, r10                 ; s = vp8_signed_char_clamp(qs1 - u)
+    qadd8       r10, r11, r10               ; s = vp8_signed_char_clamp(ps1 + u)
+    eor         r8, r8, lr                  ; *oq1 = s^0x80
+    eor         r10, r10, lr                ; *op1 = s^0x80
+
+    ldrb        r11, [src, #-5]             ; load p2 for 1/7th difference across boundary
+    strb        r10, [src, #-4]             ; store op1
+    strb        r8, [src, #-1]              ; store oq1
+    ldrb        r9, [src], pstep            ; load q2 for 1/7th difference across boundary
+
+    mov         r10, r10, lsr #8
+    mov         r8, r8, lsr #8
+
+    ldrb        r6, [src, #-5]
+    strb        r10, [src, #-4]
+    strb        r8, [src, #-1]
+    ldrb        r7, [src], pstep
+
+    mov         r10, r10, lsr #8
+    mov         r8, r8, lsr #8
+    orr         r11, r11, r6, lsl #8
+    orr         r9, r9, r7, lsl #8
+
+    ldrb        r6, [src, #-5]
+    strb        r10, [src, #-4]
+    strb        r8, [src, #-1]
+    ldrb        r7, [src], pstep
+
+    mov         r10, r10, lsr #8
+    mov         r8, r8, lsr #8
+    orr         r11, r11, r6, lsl #16
+    orr         r9, r9, r7, lsl #16
+
+    ldrb        r6, [src, #-5]
+    strb        r10, [src, #-4]
+    strb        r8, [src, #-1]
+    ldrb        r7, [src], pstep
+    orr         r11, r11, r6, lsl #24
+    orr         r9, r9, r7, lsl #24
+
+    ;roughly 1/7th difference across boundary
+    eor         r9, r9, lr
+    eor         r11, r11, lr
+
+    mov         lr, #0x9                    ; 9
+    mov         r7, #0x3f                   ; 63
+
+    sxtb16      r6, r12
+    sxtb16      r10, r12, ror #8
+    smlabb      r8, r6, lr, r7
+    smlatb      r6, r6, lr, r7
+    smlabb      r12, r10, lr, r7
+    smlatb      r10, r10, lr, r7
+    ssat        r8, #8, r8, asr #7
+    ssat        r6, #8, r6, asr #7
+    ssat        r12, #8, r12, asr #7
+    ssat        r10, #8, r10, asr #7
+
+    sub         src, src, pstep, lsl #2
+
+    pkhbt       r6, r8, r6, lsl #16
+    pkhbt       r10, r12, r10, lsl #16
+
+    uxtb16      r6, r6
+    uxtb16      r10, r10
+
+    ldr         lr, c0x80808080
+
+    orr         r10, r6, r10, lsl #8        ; u = vp8_signed_char_clamp((63 + Filter2 * 9)>>7)
+
+    qadd8       r8, r11, r10                ; s = vp8_signed_char_clamp(ps2 + u)
+    qsub8       r10, r9, r10                ; s = vp8_signed_char_clamp(qs2 - u)
+    eor         r8, r8, lr                  ; *op2 = s^0x80
+    eor         r10, r10, lr                ; *oq2 = s^0x80
+
+    strb        r8, [src, #-5]              ; store *op2
+    strb        r10, [src], pstep           ; store *oq2
+    mov         r8, r8, lsr #8
+    mov         r10, r10, lsr #8
+    strb        r8, [src, #-5]
+    strb        r10, [src], pstep
+    mov         r8, r8, lsr #8
+    mov         r10, r10, lsr #8
+    strb        r8, [src, #-5]
+    strb        r10, [src], pstep
+    mov         r8, r8, lsr #8
+    mov         r10, r10, lsr #8
+    strb        r8, [src, #-5]
+    strb        r10, [src], pstep
+
+    ;adjust src pointer for next loop
+    sub         src, src, #2
+
+|mbvskip_filter|
+    sub         src, src, #4
+    subs        count, count, #1
+
+    pld         [src, #23]                  ; preload for next block
+    ldrne       r6, [src], pstep            ; load source data
+    pld         [src, #23]
+    ldrne       r7, [src], pstep
+    pld         [src, #23]
+    ldrne       r8, [src], pstep
+    pld         [src, #23]
+    ldrne       lr, [src], pstep
+
+    bne         MBVnext8
+
+    add         sp, sp, #16
+
+    ldmia       sp!, {r4 - r11, pc}
+    ENDP        ; |vp8_mbloop_filter_vertical_edge_armv6|
+
+; Constant Pool
+c0x80808080 DCD     0x80808080
+c0x03030303 DCD     0x03030303
+c0x04040404 DCD     0x04040404
+c0x01010101 DCD     0x01010101
+c0x7F7F7F7F DCD     0x7F7F7F7F
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/simpleloopfilter_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/simpleloopfilter_v6.asm
new file mode 100644
index 0000000..5e00cf0
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/simpleloopfilter_v6.asm
@@ -0,0 +1,286 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT |vp8_loop_filter_simple_horizontal_edge_armv6|
+    EXPORT |vp8_loop_filter_simple_vertical_edge_armv6|
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+
+    MACRO
+    TRANSPOSE_MATRIX $a0, $a1, $a2, $a3, $b0, $b1, $b2, $b3
+    ; input: $a0, $a1, $a2, $a3; output: $b0, $b1, $b2, $b3
+    ; a0: 03 02 01 00
+    ; a1: 13 12 11 10
+    ; a2: 23 22 21 20
+    ; a3: 33 32 31 30
+    ;     b3 b2 b1 b0
+
+    uxtb16      $b1, $a1                    ; xx 12 xx 10
+    uxtb16      $b0, $a0                    ; xx 02 xx 00
+    uxtb16      $b3, $a3                    ; xx 32 xx 30
+    uxtb16      $b2, $a2                    ; xx 22 xx 20
+    orr         $b1, $b0, $b1, lsl #8       ; 12 02 10 00
+    orr         $b3, $b2, $b3, lsl #8       ; 32 22 30 20
+
+    uxtb16      $a1, $a1, ror #8            ; xx 13 xx 11
+    uxtb16      $a3, $a3, ror #8            ; xx 33 xx 31
+    uxtb16      $a0, $a0, ror #8            ; xx 03 xx 01
+    uxtb16      $a2, $a2, ror #8            ; xx 23 xx 21
+    orr         $a0, $a0, $a1, lsl #8       ; 13 03 11 01
+    orr         $a2, $a2, $a3, lsl #8       ; 33 23 31 21
+
+    pkhtb       $b2, $b3, $b1, asr #16      ; 32 22 12 02   -- p1
+    pkhbt       $b0, $b1, $b3, lsl #16      ; 30 20 10 00   -- p3
+
+    pkhtb       $b3, $a2, $a0, asr #16      ; 33 23 13 03   -- p0
+    pkhbt       $b1, $a0, $a2, lsl #16      ; 31 21 11 01   -- p2
+    MEND
+
+
+
+src         RN  r0
+pstep       RN  r1
+
+;r0     unsigned char *src_ptr,
+;r1     int src_pixel_step,
+;r2     const char *blimit
+
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+|vp8_loop_filter_simple_horizontal_edge_armv6| PROC
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+    stmdb       sp!, {r4 - r11, lr}
+
+    ldrb        r12, [r2]                   ; blimit
+    ldr         r3, [src, -pstep, lsl #1]   ; p1
+    ldr         r4, [src, -pstep]           ; p0
+    ldr         r5, [src]                   ; q0
+    ldr         r6, [src, pstep]            ; q1
+    orr         r12, r12, r12, lsl #8       ; blimit
+    ldr         r2, c0x80808080
+    orr         r12, r12, r12, lsl #16      ; blimit
+    mov         r9, #4                      ; double the count. we're doing 4 at a time
+    mov         lr, #0                      ; need 0 in a couple places
+
+|simple_hnext8|
+    ; vp8_simple_filter_mask()
+
+    uqsub8      r7, r3, r6                  ; p1 - q1
+    uqsub8      r8, r6, r3                  ; q1 - p1
+    uqsub8      r10, r4, r5                 ; p0 - q0
+    uqsub8      r11, r5, r4                 ; q0 - p0
+    orr         r8, r8, r7                  ; abs(p1 - q1)
+    orr         r10, r10, r11               ; abs(p0 - q0)
+    uqadd8      r10, r10, r10               ; abs(p0 - q0) * 2
+    uhadd8      r8, r8, lr                  ; abs(p1 - q2) >> 1
+    uqadd8      r10, r10, r8                ; abs(p0 - q0)*2 + abs(p1 - q1)/2
+    mvn         r8, #0
+    usub8       r10, r12, r10               ; compare to flimit. usub8 sets GE flags
+    sel         r10, r8, lr                 ; filter mask: F or 0
+    cmp         r10, #0
+    beq         simple_hskip_filter         ; skip filtering if all masks are 0x00
+
+    ;vp8_simple_filter()
+
+    eor         r3, r3, r2                  ; p1 offset to convert to a signed value
+    eor         r6, r6, r2                  ; q1 offset to convert to a signed value
+    eor         r4, r4, r2                  ; p0 offset to convert to a signed value
+    eor         r5, r5, r2                  ; q0 offset to convert to a signed value
+
+    qsub8       r3, r3, r6                  ; vp8_filter = p1 - q1
+    qsub8       r6, r5, r4                  ; q0 - p0
+    qadd8       r3, r3, r6                  ; += q0 - p0
+    ldr         r7, c0x04040404
+    qadd8       r3, r3, r6                  ; += q0 - p0
+    ldr         r8, c0x03030303
+    qadd8       r3, r3, r6                  ; vp8_filter = p1-q1 + 3*(q0-p0))
+    ;STALL
+    and         r3, r3, r10                 ; vp8_filter &= mask
+
+    qadd8       r7 , r3 , r7                ; Filter1 = vp8_filter + 4
+    qadd8       r8 , r3 , r8                ; Filter2 = vp8_filter + 3
+
+    shadd8      r7 , r7 , lr
+    shadd8      r8 , r8 , lr
+    shadd8      r7 , r7 , lr
+    shadd8      r8 , r8 , lr
+    shadd8      r7 , r7 , lr                ; Filter1 >>= 3
+    shadd8      r8 , r8 , lr                ; Filter2 >>= 3
+
+    qsub8       r5 ,r5, r7                  ; u = q0 - Filter1
+    qadd8       r4, r4, r8                  ; u = p0 + Filter2
+    eor         r5, r5, r2                  ; *oq0 = u^0x80
+    str         r5, [src]                   ; store oq0 result
+    eor         r4, r4, r2                  ; *op0 = u^0x80
+    str         r4, [src, -pstep]           ; store op0 result
+
+|simple_hskip_filter|
+    subs        r9, r9, #1
+    addne       src, src, #4                ; next row
+
+    ldrne       r3, [src, -pstep, lsl #1]   ; p1
+    ldrne       r4, [src, -pstep]           ; p0
+    ldrne       r5, [src]                   ; q0
+    ldrne       r6, [src, pstep]            ; q1
+
+    bne         simple_hnext8
+
+    ldmia       sp!, {r4 - r11, pc}
+    ENDP        ; |vp8_loop_filter_simple_horizontal_edge_armv6|
+
+
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+|vp8_loop_filter_simple_vertical_edge_armv6| PROC
+;-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
+    stmdb       sp!, {r4 - r11, lr}
+
+    ldrb        r12, [r2]                   ; r12: blimit
+    ldr         r2, c0x80808080
+    orr         r12, r12, r12, lsl #8
+
+    ; load soure data to r7, r8, r9, r10
+    ldrh        r3, [src, #-2]
+    pld         [src, #23]                  ; preload for next block
+    ldrh        r4, [src], pstep
+    orr         r12, r12, r12, lsl #16
+
+    ldrh        r5, [src, #-2]
+    pld         [src, #23]
+    ldrh        r6, [src], pstep
+
+    pkhbt       r7, r3, r4, lsl #16
+
+    ldrh        r3, [src, #-2]
+    pld         [src, #23]
+    ldrh        r4, [src], pstep
+
+    pkhbt       r8, r5, r6, lsl #16
+
+    ldrh        r5, [src, #-2]
+    pld         [src, #23]
+    ldrh        r6, [src], pstep
+    mov         r11, #4                     ; double the count. we're doing 4 at a time
+
+|simple_vnext8|
+    ; vp8_simple_filter_mask() function
+    pkhbt       r9, r3, r4, lsl #16
+    pkhbt       r10, r5, r6, lsl #16
+
+    ;transpose r7, r8, r9, r10 to r3, r4, r5, r6
+    TRANSPOSE_MATRIX r7, r8, r9, r10, r3, r4, r5, r6
+
+    uqsub8      r7, r3, r6                  ; p1 - q1
+    uqsub8      r8, r6, r3                  ; q1 - p1
+    uqsub8      r9, r4, r5                  ; p0 - q0
+    uqsub8      r10, r5, r4                 ; q0 - p0
+    orr         r7, r7, r8                  ; abs(p1 - q1)
+    orr         r9, r9, r10                 ; abs(p0 - q0)
+    mov         r8, #0
+    uqadd8      r9, r9, r9                  ; abs(p0 - q0) * 2
+    uhadd8      r7, r7, r8                  ; abs(p1 - q1) / 2
+    uqadd8      r7, r7, r9                  ; abs(p0 - q0)*2 + abs(p1 - q1)/2
+    mvn         r10, #0                     ; r10 == -1
+
+    usub8       r7, r12, r7                 ; compare to flimit
+    sel         lr, r10, r8                 ; filter mask
+
+    cmp         lr, #0
+    beq         simple_vskip_filter         ; skip filtering
+
+    ;vp8_simple_filter() function
+    eor         r3, r3, r2                  ; p1 offset to convert to a signed value
+    eor         r6, r6, r2                  ; q1 offset to convert to a signed value
+    eor         r4, r4, r2                  ; p0 offset to convert to a signed value
+    eor         r5, r5, r2                  ; q0 offset to convert to a signed value
+
+    qsub8       r3, r3, r6                  ; vp8_filter = p1 - q1
+    qsub8       r6, r5, r4                  ; q0 - p0
+
+    qadd8       r3, r3, r6                  ; vp8_filter += q0 - p0
+    ldr         r9, c0x03030303             ; r9 = 3
+
+    qadd8       r3, r3, r6                  ; vp8_filter += q0 - p0
+    ldr         r7, c0x04040404
+
+    qadd8       r3, r3, r6                  ; vp8_filter = p1-q1 + 3*(q0-p0))
+    ;STALL
+    and         r3, r3, lr                  ; vp8_filter &= mask
+
+    qadd8       r9 , r3 , r9                ; Filter2 = vp8_filter + 3
+    qadd8       r3 , r3 , r7                ; Filter1 = vp8_filter + 4
+
+    shadd8      r9 , r9 , r8
+    shadd8      r3 , r3 , r8
+    shadd8      r9 , r9 , r8
+    shadd8      r3 , r3 , r8
+    shadd8      r9 , r9 , r8                ; Filter2 >>= 3
+    shadd8      r3 , r3 , r8                ; Filter1 >>= 3
+
+    ;calculate output
+    sub         src, src, pstep, lsl #2
+
+    qadd8       r4, r4, r9                  ; u = p0 + Filter2
+    qsub8       r5, r5, r3                  ; u = q0 - Filter1
+    eor         r4, r4, r2                  ; *op0 = u^0x80
+    eor         r5, r5, r2                  ; *oq0 = u^0x80
+
+    strb        r4, [src, #-1]              ; store the result
+    mov         r4, r4, lsr #8
+    strb        r5, [src], pstep
+    mov         r5, r5, lsr #8
+
+    strb        r4, [src, #-1]
+    mov         r4, r4, lsr #8
+    strb        r5, [src], pstep
+    mov         r5, r5, lsr #8
+
+    strb        r4, [src, #-1]
+    mov         r4, r4, lsr #8
+    strb        r5, [src], pstep
+    mov         r5, r5, lsr #8
+
+    strb        r4, [src, #-1]
+    strb        r5, [src], pstep
+
+|simple_vskip_filter|
+    subs        r11, r11, #1
+
+    ; load soure data to r7, r8, r9, r10
+    ldrneh      r3, [src, #-2]
+    pld         [src, #23]                  ; preload for next block
+    ldrneh      r4, [src], pstep
+
+    ldrneh      r5, [src, #-2]
+    pld         [src, #23]
+    ldrneh      r6, [src], pstep
+
+    pkhbt       r7, r3, r4, lsl #16
+
+    ldrneh      r3, [src, #-2]
+    pld         [src, #23]
+    ldrneh      r4, [src], pstep
+
+    pkhbt       r8, r5, r6, lsl #16
+
+    ldrneh      r5, [src, #-2]
+    pld         [src, #23]
+    ldrneh      r6, [src], pstep
+
+    bne         simple_vnext8
+
+    ldmia       sp!, {r4 - r11, pc}
+    ENDP        ; |vp8_loop_filter_simple_vertical_edge_armv6|
+
+; Constant Pool
+c0x80808080 DCD     0x80808080
+c0x03030303 DCD     0x03030303
+c0x04040404 DCD     0x04040404
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/armv6/sixtappredict8x4_v6.asm b/libvpx/libvpx/vp8/common/arm/armv6/sixtappredict8x4_v6.asm
new file mode 100644
index 0000000..e81aef5
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/armv6/sixtappredict8x4_v6.asm
@@ -0,0 +1,273 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vp8_sixtap_predict8x4_armv6|
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+;-------------------------------------
+; r0    unsigned char *src_ptr,
+; r1    int  src_pixels_per_line,
+; r2    int  xoffset,
+; r3    int  yoffset,
+; stack unsigned char *dst_ptr,
+; stack int  dst_pitch
+;-------------------------------------
+;note: In first pass, store the result in transpose(8linesx9columns) on stack. Temporary stack size is 184.
+;Line width is 20 that is 9 short data plus 2 to make it 4bytes aligned. In second pass, load data from stack,
+;and the result is stored in transpose.
+|vp8_sixtap_predict8x4_armv6| PROC
+    stmdb       sp!, {r4 - r11, lr}
+    str         r3, [sp, #-184]!            ;reserve space on stack for temporary storage, store yoffset
+
+    cmp         r2, #0                      ;skip first_pass filter if xoffset=0
+    add         lr, sp, #4                  ;point to temporary buffer
+    beq         skip_firstpass_filter
+
+;first-pass filter
+    adr         r12, filter8_coeff
+    sub         r0, r0, r1, lsl #1
+
+    add         r3, r1, #10                 ; preload next low
+    pld         [r0, r3]
+
+    add         r2, r12, r2, lsl #4         ;calculate filter location
+    add         r0, r0, #3                  ;adjust src only for loading convinience
+
+    ldr         r3, [r2]                    ; load up packed filter coefficients
+    ldr         r4, [r2, #4]
+    ldr         r5, [r2, #8]
+
+    mov         r2, #0x90000                ; height=9 is top part of counter
+
+    sub         r1, r1, #8
+
+|first_pass_hloop_v6|
+    ldrb        r6, [r0, #-5]               ; load source data
+    ldrb        r7, [r0, #-4]
+    ldrb        r8, [r0, #-3]
+    ldrb        r9, [r0, #-2]
+    ldrb        r10, [r0, #-1]
+
+    orr         r2, r2, #0x4                ; construct loop counter. width=8=4x2
+
+    pkhbt       r6, r6, r7, lsl #16         ; r7 | r6
+    pkhbt       r7, r7, r8, lsl #16         ; r8 | r7
+
+    pkhbt       r8, r8, r9, lsl #16         ; r9 | r8
+    pkhbt       r9, r9, r10, lsl #16        ; r10 | r9
+
+|first_pass_wloop_v6|
+    smuad       r11, r6, r3                 ; vp8_filter[0], vp8_filter[1]
+    smuad       r12, r7, r3
+
+    ldrb        r6, [r0], #1
+
+    smlad       r11, r8, r4, r11            ; vp8_filter[2], vp8_filter[3]
+    ldrb        r7, [r0], #1
+    smlad       r12, r9, r4, r12
+
+    pkhbt       r10, r10, r6, lsl #16       ; r10 | r9
+    pkhbt       r6, r6, r7, lsl #16         ; r11 | r10
+    smlad       r11, r10, r5, r11           ; vp8_filter[4], vp8_filter[5]
+    smlad       r12, r6, r5, r12
+
+    sub         r2, r2, #1
+
+    add         r11, r11, #0x40             ; round_shift_and_clamp
+    tst         r2, #0xff                   ; test loop counter
+    usat        r11, #8, r11, asr #7
+    add         r12, r12, #0x40
+    strh        r11, [lr], #20              ; result is transposed and stored, which
+    usat        r12, #8, r12, asr #7
+
+    strh        r12, [lr], #20
+
+    movne       r11, r6
+    movne       r12, r7
+
+    movne       r6, r8
+    movne       r7, r9
+    movne       r8, r10
+    movne       r9, r11
+    movne       r10, r12
+
+    bne         first_pass_wloop_v6
+
+    ;;add       r9, ppl, #30                ; attempt to load 2 adjacent cache lines
+    ;;IF ARCHITECTURE=6
+    ;pld        [src, ppl]
+    ;;pld       [src, r9]
+    ;;ENDIF
+
+    subs        r2, r2, #0x10000
+
+    sub         lr, lr, #158
+
+    add         r0, r0, r1                  ; move to next input line
+
+    add         r11, r1, #18                ; preload next low. adding back block width(=8), which is subtracted earlier
+    pld         [r0, r11]
+
+    bne         first_pass_hloop_v6
+
+;second pass filter
+secondpass_filter
+    ldr         r3, [sp], #4                ; load back yoffset
+    ldr         r0, [sp, #216]              ; load dst address from stack 180+36
+    ldr         r1, [sp, #220]              ; load dst stride from stack 180+40
+
+    cmp         r3, #0
+    beq         skip_secondpass_filter
+
+    adr         r12, filter8_coeff
+    add         lr, r12, r3, lsl #4         ;calculate filter location
+
+    mov         r2, #0x00080000
+
+    ldr         r3, [lr]                    ; load up packed filter coefficients
+    ldr         r4, [lr, #4]
+    ldr         r5, [lr, #8]
+
+    pkhbt       r12, r4, r3                 ; pack the filter differently
+    pkhbt       r11, r5, r4
+
+second_pass_hloop_v6
+    ldr         r6, [sp]                    ; load the data
+    ldr         r7, [sp, #4]
+
+    orr         r2, r2, #2                  ; loop counter
+
+second_pass_wloop_v6
+    smuad       lr, r3, r6                  ; apply filter
+    smulbt      r10, r3, r6
+
+    ldr         r8, [sp, #8]
+
+    smlad       lr, r4, r7, lr
+    smladx      r10, r12, r7, r10
+
+    ldrh        r9, [sp, #12]
+
+    smlad       lr, r5, r8, lr
+    smladx      r10, r11, r8, r10
+
+    add         sp, sp, #4
+    smlatb      r10, r5, r9, r10
+
+    sub         r2, r2, #1
+
+    add         lr, lr, #0x40               ; round_shift_and_clamp
+    tst         r2, #0xff
+    usat        lr, #8, lr, asr #7
+    add         r10, r10, #0x40
+    strb        lr, [r0], r1                ; the result is transposed back and stored
+    usat        r10, #8, r10, asr #7
+
+    strb        r10, [r0],r1
+
+    movne       r6, r7
+    movne       r7, r8
+
+    bne         second_pass_wloop_v6
+
+    subs        r2, r2, #0x10000
+    add         sp, sp, #12                 ; updata src for next loop (20-8)
+    sub         r0, r0, r1, lsl #2
+    add         r0, r0, #1
+
+    bne         second_pass_hloop_v6
+
+    add         sp, sp, #20
+    ldmia       sp!, {r4 - r11, pc}
+
+;--------------------
+skip_firstpass_filter
+    sub         r0, r0, r1, lsl #1
+    sub         r1, r1, #8
+    mov         r2, #9
+
+skip_firstpass_hloop
+    ldrb        r4, [r0], #1                ; load data
+    subs        r2, r2, #1
+    ldrb        r5, [r0], #1
+    strh        r4, [lr], #20               ; store it to immediate buffer
+    ldrb        r6, [r0], #1                ; load data
+    strh        r5, [lr], #20
+    ldrb        r7, [r0], #1
+    strh        r6, [lr], #20
+    ldrb        r8, [r0], #1
+    strh        r7, [lr], #20
+    ldrb        r9, [r0], #1
+    strh        r8, [lr], #20
+    ldrb        r10, [r0], #1
+    strh        r9, [lr], #20
+    ldrb        r11, [r0], #1
+    strh        r10, [lr], #20
+    add         r0, r0, r1                  ; move to next input line
+    strh        r11, [lr], #20
+
+    sub         lr, lr, #158                ; move over to next column
+    bne         skip_firstpass_hloop
+
+    b           secondpass_filter
+
+;--------------------
+skip_secondpass_filter
+    mov         r2, #8
+    add         sp, sp, #4                  ;start from src[0] instead of src[-2]
+
+skip_secondpass_hloop
+    ldr         r6, [sp], #4
+    subs        r2, r2, #1
+    ldr         r8, [sp], #4
+
+    mov         r7, r6, lsr #16             ; unpack
+    strb        r6, [r0], r1
+    mov         r9, r8, lsr #16
+    strb        r7, [r0], r1
+    add         sp, sp, #12                 ; 20-8
+    strb        r8, [r0], r1
+    strb        r9, [r0], r1
+
+    sub         r0, r0, r1, lsl #2
+    add         r0, r0, #1
+
+    bne         skip_secondpass_hloop
+
+    add         sp, sp, #16                 ; 180 - (160 +4)
+
+    ldmia       sp!, {r4 - r11, pc}
+
+    ENDP
+
+;-----------------
+;One word each is reserved. Label filter_coeff can be used to access the data.
+;Data address: filter_coeff, filter_coeff+4, filter_coeff+8 ...
+filter8_coeff
+    DCD     0x00000000,     0x00000080,     0x00000000,     0x00000000
+    DCD     0xfffa0000,     0x000c007b,     0x0000ffff,     0x00000000
+    DCD     0xfff50002,     0x0024006c,     0x0001fff8,     0x00000000
+    DCD     0xfff70000,     0x0032005d,     0x0000fffa,     0x00000000
+    DCD     0xfff00003,     0x004d004d,     0x0003fff0,     0x00000000
+    DCD     0xfffa0000,     0x005d0032,     0x0000fff7,     0x00000000
+    DCD     0xfff80001,     0x006c0024,     0x0002fff5,     0x00000000
+    DCD     0xffff0000,     0x007b000c,     0x0000fffa,     0x00000000
+
+    ;DCD        0,  0,  128,    0,   0,  0
+    ;DCD        0, -6,  123,   12,  -1,  0
+    ;DCD        2, -11, 108,   36,  -8,  1
+    ;DCD        0, -9,   93,   50,  -6,  0
+    ;DCD        3, -16,  77,   77, -16,  3
+    ;DCD        0, -6,   50,   93,  -9,  0
+    ;DCD        1, -8,   36,  108, -11,  2
+    ;DCD        0, -1,   12,  123,  -6,  0
+
+    END
diff --git a/libvpx/libvpx/vp8/common/arm/bilinearfilter_arm.c b/libvpx/libvpx/vp8/common/arm/bilinearfilter_arm.c
new file mode 100644
index 0000000..799c8bd
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/bilinearfilter_arm.c
@@ -0,0 +1,113 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include <math.h>
+#include "vp8/common/filter.h"
+#include "bilinearfilter_arm.h"
+
+void vp8_filter_block2d_bil_armv6
+(
+    unsigned char *src_ptr,
+    unsigned char *dst_ptr,
+    unsigned int   src_pitch,
+    unsigned int   dst_pitch,
+    const short   *HFilter,
+    const short   *VFilter,
+    int            Width,
+    int            Height
+)
+{
+    unsigned short FData[36*16]; /* Temp data buffer used in filtering */
+
+    /* First filter 1-D horizontally... */
+    vp8_filter_block2d_bil_first_pass_armv6(src_ptr, FData, src_pitch, Height + 1, Width, HFilter);
+
+    /* then 1-D vertically... */
+    vp8_filter_block2d_bil_second_pass_armv6(FData, dst_ptr, dst_pitch, Height, Width, VFilter);
+}
+
+
+void vp8_bilinear_predict4x4_armv6
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+
+    HFilter = vp8_bilinear_filters[xoffset];
+    VFilter = vp8_bilinear_filters[yoffset];
+
+    vp8_filter_block2d_bil_armv6(src_ptr, dst_ptr, src_pixels_per_line, dst_pitch, HFilter, VFilter, 4, 4);
+}
+
+void vp8_bilinear_predict8x8_armv6
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+
+    HFilter = vp8_bilinear_filters[xoffset];
+    VFilter = vp8_bilinear_filters[yoffset];
+
+    vp8_filter_block2d_bil_armv6(src_ptr, dst_ptr, src_pixels_per_line, dst_pitch, HFilter, VFilter, 8, 8);
+}
+
+void vp8_bilinear_predict8x4_armv6
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+
+    HFilter = vp8_bilinear_filters[xoffset];
+    VFilter = vp8_bilinear_filters[yoffset];
+
+    vp8_filter_block2d_bil_armv6(src_ptr, dst_ptr, src_pixels_per_line, dst_pitch, HFilter, VFilter, 8, 4);
+}
+
+void vp8_bilinear_predict16x16_armv6
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+
+    HFilter = vp8_bilinear_filters[xoffset];
+    VFilter = vp8_bilinear_filters[yoffset];
+
+    vp8_filter_block2d_bil_armv6(src_ptr, dst_ptr, src_pixels_per_line, dst_pitch, HFilter, VFilter, 16, 16);
+}
diff --git a/libvpx/libvpx/vp8/common/arm/bilinearfilter_arm.h b/libvpx/libvpx/vp8/common/arm/bilinearfilter_arm.h
new file mode 100644
index 0000000..6b84e6f
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/bilinearfilter_arm.h
@@ -0,0 +1,43 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_ARM_BILINEARFILTER_ARM_H_
+#define VP8_COMMON_ARM_BILINEARFILTER_ARM_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern void vp8_filter_block2d_bil_first_pass_armv6
+(
+    const unsigned char  *src_ptr,
+    unsigned short       *dst_ptr,
+    unsigned int          src_pitch,
+    unsigned int          height,
+    unsigned int          width,
+    const short          *vp8_filter
+);
+
+extern void vp8_filter_block2d_bil_second_pass_armv6
+(
+    const unsigned short *src_ptr,
+    unsigned char        *dst_ptr,
+    int                   dst_pitch,
+    unsigned int          height,
+    unsigned int          width,
+    const short         *vp8_filter
+);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_ARM_BILINEARFILTER_ARM_H_
diff --git a/libvpx/libvpx/vp8/common/arm/dequantize_arm.c b/libvpx/libvpx/vp8/common/arm/dequantize_arm.c
new file mode 100644
index 0000000..1f8157f
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/dequantize_arm.c
@@ -0,0 +1,25 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8/common/blockd.h"
+
+#if HAVE_MEDIA
+extern void vp8_dequantize_b_loop_v6(short *Q, short *DQC, short *DQ);
+
+void vp8_dequantize_b_v6(BLOCKD *d, short *DQC)
+{
+    short *DQ  = d->dqcoeff;
+    short *Q   = d->qcoeff;
+
+    vp8_dequantize_b_loop_v6(Q, DQC, DQ);
+}
+#endif
diff --git a/libvpx/libvpx/vp8/common/arm/filter_arm.c b/libvpx/libvpx/vp8/common/arm/filter_arm.c
new file mode 100644
index 0000000..d6a6781
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/filter_arm.c
@@ -0,0 +1,221 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include <math.h>
+#include "vp8/common/filter.h"
+#include "vpx_ports/mem.h"
+
+extern void vp8_filter_block2d_first_pass_armv6
+(
+    unsigned char *src_ptr,
+    short         *output_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int output_width,
+    unsigned int output_height,
+    const short *vp8_filter
+);
+
+// 8x8
+extern void vp8_filter_block2d_first_pass_8x8_armv6
+(
+    unsigned char *src_ptr,
+    short         *output_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int output_width,
+    unsigned int output_height,
+    const short *vp8_filter
+);
+
+// 16x16
+extern void vp8_filter_block2d_first_pass_16x16_armv6
+(
+    unsigned char *src_ptr,
+    short         *output_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int output_width,
+    unsigned int output_height,
+    const short *vp8_filter
+);
+
+extern void vp8_filter_block2d_second_pass_armv6
+(
+    short         *src_ptr,
+    unsigned char *output_ptr,
+    unsigned int output_pitch,
+    unsigned int cnt,
+    const short *vp8_filter
+);
+
+extern void vp8_filter4_block2d_second_pass_armv6
+(
+    short         *src_ptr,
+    unsigned char *output_ptr,
+    unsigned int output_pitch,
+    unsigned int cnt,
+    const short *vp8_filter
+);
+
+extern void vp8_filter_block2d_first_pass_only_armv6
+(
+    unsigned char *src_ptr,
+    unsigned char *output_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int cnt,
+    unsigned int output_pitch,
+    const short *vp8_filter
+);
+
+
+extern void vp8_filter_block2d_second_pass_only_armv6
+(
+    unsigned char *src_ptr,
+    unsigned char *output_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int cnt,
+    unsigned int output_pitch,
+    const short *vp8_filter
+);
+
+#if HAVE_MEDIA
+void vp8_sixtap_predict4x4_armv6
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+    DECLARE_ALIGNED(4, short, FData[12*4]); /* Temp data buffer used in filtering */
+
+
+    HFilter = vp8_sub_pel_filters[xoffset];   /* 6 tap */
+    VFilter = vp8_sub_pel_filters[yoffset];   /* 6 tap */
+
+    /* Vfilter is null. First pass only */
+    if (xoffset && !yoffset)
+    {
+        /*vp8_filter_block2d_first_pass_armv6 ( src_ptr, FData+2, src_pixels_per_line, 4, 4, HFilter );
+        vp8_filter_block2d_second_pass_armv6 ( FData+2, dst_ptr, dst_pitch, 4, VFilter );*/
+
+        vp8_filter_block2d_first_pass_only_armv6(src_ptr, dst_ptr, src_pixels_per_line, 4, dst_pitch, HFilter);
+    }
+    /* Hfilter is null. Second pass only */
+    else if (!xoffset && yoffset)
+    {
+        vp8_filter_block2d_second_pass_only_armv6(src_ptr, dst_ptr, src_pixels_per_line, 4, dst_pitch, VFilter);
+    }
+    else
+    {
+        /* Vfilter is a 4 tap filter */
+        if (yoffset & 0x1)
+        {
+            vp8_filter_block2d_first_pass_armv6(src_ptr - src_pixels_per_line, FData + 1, src_pixels_per_line, 4, 7, HFilter);
+            vp8_filter4_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 4, VFilter);
+        }
+        /* Vfilter is 6 tap filter */
+        else
+        {
+            vp8_filter_block2d_first_pass_armv6(src_ptr - (2 * src_pixels_per_line), FData, src_pixels_per_line, 4, 9, HFilter);
+            vp8_filter_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 4, VFilter);
+        }
+    }
+}
+
+void vp8_sixtap_predict8x8_armv6
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+    DECLARE_ALIGNED(4, short, FData[16*8]); /* Temp data buffer used in filtering */
+
+    HFilter = vp8_sub_pel_filters[xoffset];   /* 6 tap */
+    VFilter = vp8_sub_pel_filters[yoffset];   /* 6 tap */
+
+    if (xoffset && !yoffset)
+    {
+        vp8_filter_block2d_first_pass_only_armv6(src_ptr, dst_ptr, src_pixels_per_line, 8, dst_pitch, HFilter);
+    }
+    /* Hfilter is null. Second pass only */
+    else if (!xoffset && yoffset)
+    {
+        vp8_filter_block2d_second_pass_only_armv6(src_ptr, dst_ptr, src_pixels_per_line, 8, dst_pitch, VFilter);
+    }
+    else
+    {
+        if (yoffset & 0x1)
+        {
+            vp8_filter_block2d_first_pass_8x8_armv6(src_ptr - src_pixels_per_line, FData + 1, src_pixels_per_line, 8, 11, HFilter);
+            vp8_filter4_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 8, VFilter);
+        }
+        else
+        {
+            vp8_filter_block2d_first_pass_8x8_armv6(src_ptr - (2 * src_pixels_per_line), FData, src_pixels_per_line, 8, 13, HFilter);
+            vp8_filter_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 8, VFilter);
+        }
+    }
+}
+
+
+void vp8_sixtap_predict16x16_armv6
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+    DECLARE_ALIGNED(4, short, FData[24*16]);    /* Temp data buffer used in filtering */
+
+    HFilter = vp8_sub_pel_filters[xoffset];   /* 6 tap */
+    VFilter = vp8_sub_pel_filters[yoffset];   /* 6 tap */
+
+    if (xoffset && !yoffset)
+    {
+        vp8_filter_block2d_first_pass_only_armv6(src_ptr, dst_ptr, src_pixels_per_line, 16, dst_pitch, HFilter);
+    }
+    /* Hfilter is null. Second pass only */
+    else if (!xoffset && yoffset)
+    {
+        vp8_filter_block2d_second_pass_only_armv6(src_ptr, dst_ptr, src_pixels_per_line, 16, dst_pitch, VFilter);
+    }
+    else
+    {
+        if (yoffset & 0x1)
+        {
+            vp8_filter_block2d_first_pass_16x16_armv6(src_ptr - src_pixels_per_line, FData + 1, src_pixels_per_line, 16, 19, HFilter);
+            vp8_filter4_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 16, VFilter);
+        }
+        else
+        {
+            vp8_filter_block2d_first_pass_16x16_armv6(src_ptr - (2 * src_pixels_per_line), FData, src_pixels_per_line, 16, 21, HFilter);
+            vp8_filter_block2d_second_pass_armv6(FData + 2, dst_ptr, dst_pitch, 16, VFilter);
+        }
+    }
+
+}
+#endif
diff --git a/libvpx/libvpx/vp8/common/arm/loopfilter_arm.c b/libvpx/libvpx/vp8/common/arm/loopfilter_arm.c
new file mode 100644
index 0000000..5840c2b
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/loopfilter_arm.c
@@ -0,0 +1,181 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vp8/common/loopfilter.h"
+#include "vp8/common/onyxc_int.h"
+
+#define prototype_loopfilter(sym) \
+    void sym(unsigned char *src, int pitch, const unsigned char *blimit,\
+             const unsigned char *limit, const unsigned char *thresh, int count)
+
+#if HAVE_MEDIA
+extern prototype_loopfilter(vp8_loop_filter_horizontal_edge_armv6);
+extern prototype_loopfilter(vp8_loop_filter_vertical_edge_armv6);
+extern prototype_loopfilter(vp8_mbloop_filter_horizontal_edge_armv6);
+extern prototype_loopfilter(vp8_mbloop_filter_vertical_edge_armv6);
+#endif
+
+#if HAVE_NEON
+typedef void loopfilter_y_neon(unsigned char *src, int pitch,
+        unsigned char blimit, unsigned char limit, unsigned char thresh);
+typedef void loopfilter_uv_neon(unsigned char *u, int pitch,
+        unsigned char blimit, unsigned char limit, unsigned char thresh,
+        unsigned char *v);
+
+extern loopfilter_y_neon vp8_loop_filter_horizontal_edge_y_neon;
+extern loopfilter_y_neon vp8_loop_filter_vertical_edge_y_neon;
+extern loopfilter_uv_neon vp8_loop_filter_horizontal_edge_uv_neon;
+extern loopfilter_uv_neon vp8_loop_filter_vertical_edge_uv_neon;
+
+extern loopfilter_y_neon vp8_mbloop_filter_horizontal_edge_y_neon;
+extern loopfilter_y_neon vp8_mbloop_filter_vertical_edge_y_neon;
+extern loopfilter_uv_neon vp8_mbloop_filter_horizontal_edge_uv_neon;
+extern loopfilter_uv_neon vp8_mbloop_filter_vertical_edge_uv_neon;
+#endif
+
+#if HAVE_MEDIA
+/* ARMV6/MEDIA loopfilter functions*/
+/* Horizontal MB filtering */
+void vp8_loop_filter_mbh_armv6(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                               int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    vp8_mbloop_filter_horizontal_edge_armv6(y_ptr, y_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_mbloop_filter_horizontal_edge_armv6(u_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_mbloop_filter_horizontal_edge_armv6(v_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+}
+
+/* Vertical MB Filtering */
+void vp8_loop_filter_mbv_armv6(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                               int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    vp8_mbloop_filter_vertical_edge_armv6(y_ptr, y_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_mbloop_filter_vertical_edge_armv6(u_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_mbloop_filter_vertical_edge_armv6(v_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+}
+
+/* Horizontal B Filtering */
+void vp8_loop_filter_bh_armv6(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                              int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    vp8_loop_filter_horizontal_edge_armv6(y_ptr + 4 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_horizontal_edge_armv6(y_ptr + 8 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_horizontal_edge_armv6(y_ptr + 12 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_loop_filter_horizontal_edge_armv6(u_ptr + 4 * uv_stride, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_loop_filter_horizontal_edge_armv6(v_ptr + 4 * uv_stride, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+}
+
+void vp8_loop_filter_bhs_armv6(unsigned char *y_ptr, int y_stride,
+                               const unsigned char *blimit)
+{
+    vp8_loop_filter_simple_horizontal_edge_armv6(y_ptr + 4 * y_stride, y_stride, blimit);
+    vp8_loop_filter_simple_horizontal_edge_armv6(y_ptr + 8 * y_stride, y_stride, blimit);
+    vp8_loop_filter_simple_horizontal_edge_armv6(y_ptr + 12 * y_stride, y_stride, blimit);
+}
+
+/* Vertical B Filtering */
+void vp8_loop_filter_bv_armv6(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                              int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    vp8_loop_filter_vertical_edge_armv6(y_ptr + 4, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_vertical_edge_armv6(y_ptr + 8, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_vertical_edge_armv6(y_ptr + 12, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_loop_filter_vertical_edge_armv6(u_ptr + 4, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_loop_filter_vertical_edge_armv6(v_ptr + 4, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+}
+
+void vp8_loop_filter_bvs_armv6(unsigned char *y_ptr, int y_stride,
+                               const unsigned char *blimit)
+{
+    vp8_loop_filter_simple_vertical_edge_armv6(y_ptr + 4, y_stride, blimit);
+    vp8_loop_filter_simple_vertical_edge_armv6(y_ptr + 8, y_stride, blimit);
+    vp8_loop_filter_simple_vertical_edge_armv6(y_ptr + 12, y_stride, blimit);
+}
+#endif
+
+#if HAVE_NEON
+/* NEON loopfilter functions */
+/* Horizontal MB filtering */
+void vp8_loop_filter_mbh_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                              int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    unsigned char mblim = *lfi->mblim;
+    unsigned char lim = *lfi->lim;
+    unsigned char hev_thr = *lfi->hev_thr;
+    vp8_mbloop_filter_horizontal_edge_y_neon(y_ptr, y_stride, mblim, lim, hev_thr);
+
+    if (u_ptr)
+        vp8_mbloop_filter_horizontal_edge_uv_neon(u_ptr, uv_stride, mblim, lim, hev_thr, v_ptr);
+}
+
+/* Vertical MB Filtering */
+void vp8_loop_filter_mbv_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                              int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    unsigned char mblim = *lfi->mblim;
+    unsigned char lim = *lfi->lim;
+    unsigned char hev_thr = *lfi->hev_thr;
+
+    vp8_mbloop_filter_vertical_edge_y_neon(y_ptr, y_stride, mblim, lim, hev_thr);
+
+    if (u_ptr)
+        vp8_mbloop_filter_vertical_edge_uv_neon(u_ptr, uv_stride, mblim, lim, hev_thr, v_ptr);
+}
+
+/* Horizontal B Filtering */
+void vp8_loop_filter_bh_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                             int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    unsigned char blim = *lfi->blim;
+    unsigned char lim = *lfi->lim;
+    unsigned char hev_thr = *lfi->hev_thr;
+
+    vp8_loop_filter_horizontal_edge_y_neon(y_ptr + 4 * y_stride, y_stride, blim, lim, hev_thr);
+    vp8_loop_filter_horizontal_edge_y_neon(y_ptr + 8 * y_stride, y_stride, blim, lim, hev_thr);
+    vp8_loop_filter_horizontal_edge_y_neon(y_ptr + 12 * y_stride, y_stride, blim, lim, hev_thr);
+
+    if (u_ptr)
+        vp8_loop_filter_horizontal_edge_uv_neon(u_ptr + 4 * uv_stride, uv_stride, blim, lim, hev_thr, v_ptr + 4 * uv_stride);
+}
+
+/* Vertical B Filtering */
+void vp8_loop_filter_bv_neon(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                             int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    unsigned char blim = *lfi->blim;
+    unsigned char lim = *lfi->lim;
+    unsigned char hev_thr = *lfi->hev_thr;
+
+    vp8_loop_filter_vertical_edge_y_neon(y_ptr + 4, y_stride, blim, lim, hev_thr);
+    vp8_loop_filter_vertical_edge_y_neon(y_ptr + 8, y_stride, blim, lim, hev_thr);
+    vp8_loop_filter_vertical_edge_y_neon(y_ptr + 12, y_stride, blim, lim, hev_thr);
+
+    if (u_ptr)
+        vp8_loop_filter_vertical_edge_uv_neon(u_ptr + 4, uv_stride, blim, lim, hev_thr, v_ptr + 4);
+}
+#endif
diff --git a/libvpx/libvpx/vp8/common/arm/neon/bilinearpredict_neon.c b/libvpx/libvpx/vp8/common/arm/neon/bilinearpredict_neon.c
new file mode 100644
index 0000000..bb6ea76
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/bilinearpredict_neon.c
@@ -0,0 +1,591 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+static const uint8_t bifilter4_coeff[8][2] = {
+    {128,   0},
+    {112,  16},
+    { 96,  32},
+    { 80,  48},
+    { 64,  64},
+    { 48,  80},
+    { 32,  96},
+    { 16, 112}
+};
+
+void vp8_bilinear_predict8x4_neon(
+        unsigned char *src_ptr,
+        int src_pixels_per_line,
+        int xoffset,
+        int yoffset,
+        unsigned char *dst_ptr,
+        int dst_pitch) {
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8;
+    uint8x8_t d7u8, d9u8, d11u8, d22u8, d23u8, d24u8, d25u8, d26u8;
+    uint8x16_t q1u8, q2u8, q3u8, q4u8, q5u8;
+    uint16x8_t q1u16, q2u16, q3u16, q4u16;
+    uint16x8_t q6u16, q7u16, q8u16, q9u16, q10u16;
+
+    if (xoffset == 0) {  // skip_1stpass_filter
+        d22u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d23u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d24u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d25u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d26u8 = vld1_u8(src_ptr);
+    } else {
+        q1u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q2u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q3u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q4u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q5u8 = vld1q_u8(src_ptr);
+
+        d0u8 = vdup_n_u8(bifilter4_coeff[xoffset][0]);
+        d1u8 = vdup_n_u8(bifilter4_coeff[xoffset][1]);
+
+        q6u16 = vmull_u8(vget_low_u8(q1u8), d0u8);
+        q7u16 = vmull_u8(vget_low_u8(q2u8), d0u8);
+        q8u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
+        q9u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
+        q10u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
+
+        d3u8 = vext_u8(vget_low_u8(q1u8), vget_high_u8(q1u8), 1);
+        d5u8 = vext_u8(vget_low_u8(q2u8), vget_high_u8(q2u8), 1);
+        d7u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+        d9u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+        d11u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
+
+        q6u16 = vmlal_u8(q6u16, d3u8, d1u8);
+        q7u16 = vmlal_u8(q7u16, d5u8, d1u8);
+        q8u16 = vmlal_u8(q8u16, d7u8, d1u8);
+        q9u16 = vmlal_u8(q9u16, d9u8, d1u8);
+        q10u16 = vmlal_u8(q10u16, d11u8, d1u8);
+
+        d22u8 = vqrshrn_n_u16(q6u16, 7);
+        d23u8 = vqrshrn_n_u16(q7u16, 7);
+        d24u8 = vqrshrn_n_u16(q8u16, 7);
+        d25u8 = vqrshrn_n_u16(q9u16, 7);
+        d26u8 = vqrshrn_n_u16(q10u16, 7);
+    }
+
+    // secondpass_filter
+    if (yoffset == 0) {  // skip_2ndpass_filter
+        vst1_u8((uint8_t *)dst_ptr, d22u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d23u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d24u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d25u8);
+    } else {
+        d0u8 = vdup_n_u8(bifilter4_coeff[yoffset][0]);
+        d1u8 = vdup_n_u8(bifilter4_coeff[yoffset][1]);
+
+        q1u16 = vmull_u8(d22u8, d0u8);
+        q2u16 = vmull_u8(d23u8, d0u8);
+        q3u16 = vmull_u8(d24u8, d0u8);
+        q4u16 = vmull_u8(d25u8, d0u8);
+
+        q1u16 = vmlal_u8(q1u16, d23u8, d1u8);
+        q2u16 = vmlal_u8(q2u16, d24u8, d1u8);
+        q3u16 = vmlal_u8(q3u16, d25u8, d1u8);
+        q4u16 = vmlal_u8(q4u16, d26u8, d1u8);
+
+        d2u8 = vqrshrn_n_u16(q1u16, 7);
+        d3u8 = vqrshrn_n_u16(q2u16, 7);
+        d4u8 = vqrshrn_n_u16(q3u16, 7);
+        d5u8 = vqrshrn_n_u16(q4u16, 7);
+
+        vst1_u8((uint8_t *)dst_ptr, d2u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d3u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d4u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d5u8);
+    }
+    return;
+}
+
+void vp8_bilinear_predict8x8_neon(
+        unsigned char *src_ptr,
+        int src_pixels_per_line,
+        int xoffset,
+        int yoffset,
+        unsigned char *dst_ptr,
+        int dst_pitch) {
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8, d11u8;
+    uint8x8_t d22u8, d23u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8, d30u8;
+    uint8x16_t q1u8, q2u8, q3u8, q4u8, q5u8;
+    uint16x8_t q1u16, q2u16, q3u16, q4u16, q5u16;
+    uint16x8_t q6u16, q7u16, q8u16, q9u16, q10u16;
+
+    if (xoffset == 0) {  // skip_1stpass_filter
+        d22u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d23u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d24u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d25u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d26u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d27u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d28u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d29u8 = vld1_u8(src_ptr); src_ptr += src_pixels_per_line;
+        d30u8 = vld1_u8(src_ptr);
+    } else {
+        q1u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q2u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q3u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q4u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+
+        d0u8 = vdup_n_u8(bifilter4_coeff[xoffset][0]);
+        d1u8 = vdup_n_u8(bifilter4_coeff[xoffset][1]);
+
+        q6u16 = vmull_u8(vget_low_u8(q1u8), d0u8);
+        q7u16 = vmull_u8(vget_low_u8(q2u8), d0u8);
+        q8u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
+        q9u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
+
+        d3u8 = vext_u8(vget_low_u8(q1u8), vget_high_u8(q1u8), 1);
+        d5u8 = vext_u8(vget_low_u8(q2u8), vget_high_u8(q2u8), 1);
+        d7u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+        d9u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+
+        q6u16 = vmlal_u8(q6u16, d3u8, d1u8);
+        q7u16 = vmlal_u8(q7u16, d5u8, d1u8);
+        q8u16 = vmlal_u8(q8u16, d7u8, d1u8);
+        q9u16 = vmlal_u8(q9u16, d9u8, d1u8);
+
+        d22u8 = vqrshrn_n_u16(q6u16, 7);
+        d23u8 = vqrshrn_n_u16(q7u16, 7);
+        d24u8 = vqrshrn_n_u16(q8u16, 7);
+        d25u8 = vqrshrn_n_u16(q9u16, 7);
+
+        // first_pass filtering on the rest 5-line data
+        q1u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q2u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q3u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q4u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+        q5u8 = vld1q_u8(src_ptr);
+
+        q6u16 = vmull_u8(vget_low_u8(q1u8), d0u8);
+        q7u16 = vmull_u8(vget_low_u8(q2u8), d0u8);
+        q8u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
+        q9u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
+        q10u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
+
+        d3u8 = vext_u8(vget_low_u8(q1u8), vget_high_u8(q1u8), 1);
+        d5u8 = vext_u8(vget_low_u8(q2u8), vget_high_u8(q2u8), 1);
+        d7u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+        d9u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+        d11u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
+
+        q6u16 = vmlal_u8(q6u16, d3u8, d1u8);
+        q7u16 = vmlal_u8(q7u16, d5u8, d1u8);
+        q8u16 = vmlal_u8(q8u16, d7u8, d1u8);
+        q9u16 = vmlal_u8(q9u16, d9u8, d1u8);
+        q10u16 = vmlal_u8(q10u16, d11u8, d1u8);
+
+        d26u8 = vqrshrn_n_u16(q6u16, 7);
+        d27u8 = vqrshrn_n_u16(q7u16, 7);
+        d28u8 = vqrshrn_n_u16(q8u16, 7);
+        d29u8 = vqrshrn_n_u16(q9u16, 7);
+        d30u8 = vqrshrn_n_u16(q10u16, 7);
+    }
+
+    // secondpass_filter
+    if (yoffset == 0) {  // skip_2ndpass_filter
+        vst1_u8((uint8_t *)dst_ptr, d22u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d23u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d24u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d25u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d26u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d27u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d28u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d29u8);
+    } else {
+        d0u8 = vdup_n_u8(bifilter4_coeff[yoffset][0]);
+        d1u8 = vdup_n_u8(bifilter4_coeff[yoffset][1]);
+
+        q1u16 = vmull_u8(d22u8, d0u8);
+        q2u16 = vmull_u8(d23u8, d0u8);
+        q3u16 = vmull_u8(d24u8, d0u8);
+        q4u16 = vmull_u8(d25u8, d0u8);
+        q5u16 = vmull_u8(d26u8, d0u8);
+        q6u16 = vmull_u8(d27u8, d0u8);
+        q7u16 = vmull_u8(d28u8, d0u8);
+        q8u16 = vmull_u8(d29u8, d0u8);
+
+        q1u16 = vmlal_u8(q1u16, d23u8, d1u8);
+        q2u16 = vmlal_u8(q2u16, d24u8, d1u8);
+        q3u16 = vmlal_u8(q3u16, d25u8, d1u8);
+        q4u16 = vmlal_u8(q4u16, d26u8, d1u8);
+        q5u16 = vmlal_u8(q5u16, d27u8, d1u8);
+        q6u16 = vmlal_u8(q6u16, d28u8, d1u8);
+        q7u16 = vmlal_u8(q7u16, d29u8, d1u8);
+        q8u16 = vmlal_u8(q8u16, d30u8, d1u8);
+
+        d2u8 = vqrshrn_n_u16(q1u16, 7);
+        d3u8 = vqrshrn_n_u16(q2u16, 7);
+        d4u8 = vqrshrn_n_u16(q3u16, 7);
+        d5u8 = vqrshrn_n_u16(q4u16, 7);
+        d6u8 = vqrshrn_n_u16(q5u16, 7);
+        d7u8 = vqrshrn_n_u16(q6u16, 7);
+        d8u8 = vqrshrn_n_u16(q7u16, 7);
+        d9u8 = vqrshrn_n_u16(q8u16, 7);
+
+        vst1_u8((uint8_t *)dst_ptr, d2u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d3u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d4u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d5u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d6u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d7u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d8u8); dst_ptr += dst_pitch;
+        vst1_u8((uint8_t *)dst_ptr, d9u8);
+    }
+    return;
+}
+
+void vp8_bilinear_predict16x16_neon(
+        unsigned char *src_ptr,
+        int src_pixels_per_line,
+        int xoffset,
+        int yoffset,
+        unsigned char *dst_ptr,
+        int dst_pitch) {
+    int i;
+    unsigned char tmp[272];
+    unsigned char *tmpp;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
+    uint8x8_t d10u8, d11u8, d12u8, d13u8, d14u8, d15u8, d16u8, d17u8, d18u8;
+    uint8x8_t d19u8, d20u8, d21u8;
+    uint8x16_t q1u8, q2u8, q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8;
+    uint8x16_t q11u8, q12u8, q13u8, q14u8, q15u8;
+    uint16x8_t q1u16, q2u16, q3u16, q4u16, q5u16, q6u16, q7u16, q8u16;
+    uint16x8_t q9u16, q10u16, q11u16, q12u16, q13u16, q14u16;
+
+    if (xoffset == 0) {  // secondpass_bfilter16x16_only
+        d0u8 = vdup_n_u8(bifilter4_coeff[yoffset][0]);
+        d1u8 = vdup_n_u8(bifilter4_coeff[yoffset][1]);
+
+        q11u8 = vld1q_u8(src_ptr);
+        src_ptr += src_pixels_per_line;
+        for (i = 4; i > 0; i--) {
+            q12u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+            q13u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+            q14u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+            q15u8 = vld1q_u8(src_ptr); src_ptr += src_pixels_per_line;
+
+            q1u16 = vmull_u8(vget_low_u8(q11u8), d0u8);
+            q2u16 = vmull_u8(vget_high_u8(q11u8), d0u8);
+            q3u16 = vmull_u8(vget_low_u8(q12u8), d0u8);
+            q4u16 = vmull_u8(vget_high_u8(q12u8), d0u8);
+            q5u16 = vmull_u8(vget_low_u8(q13u8), d0u8);
+            q6u16 = vmull_u8(vget_high_u8(q13u8), d0u8);
+            q7u16 = vmull_u8(vget_low_u8(q14u8), d0u8);
+            q8u16 = vmull_u8(vget_high_u8(q14u8), d0u8);
+
+            q1u16 = vmlal_u8(q1u16, vget_low_u8(q12u8), d1u8);
+            q2u16 = vmlal_u8(q2u16, vget_high_u8(q12u8), d1u8);
+            q3u16 = vmlal_u8(q3u16, vget_low_u8(q13u8), d1u8);
+            q4u16 = vmlal_u8(q4u16, vget_high_u8(q13u8), d1u8);
+            q5u16 = vmlal_u8(q5u16, vget_low_u8(q14u8), d1u8);
+            q6u16 = vmlal_u8(q6u16, vget_high_u8(q14u8), d1u8);
+            q7u16 = vmlal_u8(q7u16, vget_low_u8(q15u8), d1u8);
+            q8u16 = vmlal_u8(q8u16, vget_high_u8(q15u8), d1u8);
+
+            d2u8 = vqrshrn_n_u16(q1u16, 7);
+            d3u8 = vqrshrn_n_u16(q2u16, 7);
+            d4u8 = vqrshrn_n_u16(q3u16, 7);
+            d5u8 = vqrshrn_n_u16(q4u16, 7);
+            d6u8 = vqrshrn_n_u16(q5u16, 7);
+            d7u8 = vqrshrn_n_u16(q6u16, 7);
+            d8u8 = vqrshrn_n_u16(q7u16, 7);
+            d9u8 = vqrshrn_n_u16(q8u16, 7);
+
+            q1u8 = vcombine_u8(d2u8, d3u8);
+            q2u8 = vcombine_u8(d4u8, d5u8);
+            q3u8 = vcombine_u8(d6u8, d7u8);
+            q4u8 = vcombine_u8(d8u8, d9u8);
+
+            q11u8 = q15u8;
+
+            vst1q_u8((uint8_t *)dst_ptr, q1u8); dst_ptr += dst_pitch;
+            vst1q_u8((uint8_t *)dst_ptr, q2u8); dst_ptr += dst_pitch;
+            vst1q_u8((uint8_t *)dst_ptr, q3u8); dst_ptr += dst_pitch;
+            vst1q_u8((uint8_t *)dst_ptr, q4u8); dst_ptr += dst_pitch;
+        }
+        return;
+    }
+
+    if (yoffset == 0) {  // firstpass_bfilter16x16_only
+        d0u8 = vdup_n_u8(bifilter4_coeff[xoffset][0]);
+        d1u8 = vdup_n_u8(bifilter4_coeff[xoffset][1]);
+
+        for (i = 4; i > 0 ; i--) {
+            d2u8 = vld1_u8(src_ptr);
+            d3u8 = vld1_u8(src_ptr + 8);
+            d4u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+            d5u8 = vld1_u8(src_ptr);
+            d6u8 = vld1_u8(src_ptr + 8);
+            d7u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+            d8u8 = vld1_u8(src_ptr);
+            d9u8 = vld1_u8(src_ptr + 8);
+            d10u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+            d11u8 = vld1_u8(src_ptr);
+            d12u8 = vld1_u8(src_ptr + 8);
+            d13u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+
+            q7u16  = vmull_u8(d2u8, d0u8);
+            q8u16  = vmull_u8(d3u8, d0u8);
+            q9u16  = vmull_u8(d5u8, d0u8);
+            q10u16 = vmull_u8(d6u8, d0u8);
+            q11u16 = vmull_u8(d8u8, d0u8);
+            q12u16 = vmull_u8(d9u8, d0u8);
+            q13u16 = vmull_u8(d11u8, d0u8);
+            q14u16 = vmull_u8(d12u8, d0u8);
+
+            d2u8  = vext_u8(d2u8, d3u8, 1);
+            d5u8  = vext_u8(d5u8, d6u8, 1);
+            d8u8  = vext_u8(d8u8, d9u8, 1);
+            d11u8 = vext_u8(d11u8, d12u8, 1);
+
+            q7u16  = vmlal_u8(q7u16, d2u8, d1u8);
+            q9u16  = vmlal_u8(q9u16, d5u8, d1u8);
+            q11u16 = vmlal_u8(q11u16, d8u8, d1u8);
+            q13u16 = vmlal_u8(q13u16, d11u8, d1u8);
+
+            d3u8  = vext_u8(d3u8, d4u8, 1);
+            d6u8  = vext_u8(d6u8, d7u8, 1);
+            d9u8  = vext_u8(d9u8, d10u8, 1);
+            d12u8 = vext_u8(d12u8, d13u8, 1);
+
+            q8u16  = vmlal_u8(q8u16,  d3u8, d1u8);
+            q10u16 = vmlal_u8(q10u16, d6u8, d1u8);
+            q12u16 = vmlal_u8(q12u16, d9u8, d1u8);
+            q14u16 = vmlal_u8(q14u16, d12u8, d1u8);
+
+            d14u8 = vqrshrn_n_u16(q7u16, 7);
+            d15u8 = vqrshrn_n_u16(q8u16, 7);
+            d16u8 = vqrshrn_n_u16(q9u16, 7);
+            d17u8 = vqrshrn_n_u16(q10u16, 7);
+            d18u8 = vqrshrn_n_u16(q11u16, 7);
+            d19u8 = vqrshrn_n_u16(q12u16, 7);
+            d20u8 = vqrshrn_n_u16(q13u16, 7);
+            d21u8 = vqrshrn_n_u16(q14u16, 7);
+
+            q7u8 = vcombine_u8(d14u8, d15u8);
+            q8u8 = vcombine_u8(d16u8, d17u8);
+            q9u8 = vcombine_u8(d18u8, d19u8);
+            q10u8 =vcombine_u8(d20u8, d21u8);
+
+            vst1q_u8((uint8_t *)dst_ptr, q7u8); dst_ptr += dst_pitch;
+            vst1q_u8((uint8_t *)dst_ptr, q8u8); dst_ptr += dst_pitch;
+            vst1q_u8((uint8_t *)dst_ptr, q9u8); dst_ptr += dst_pitch;
+            vst1q_u8((uint8_t *)dst_ptr, q10u8); dst_ptr += dst_pitch;
+        }
+        return;
+    }
+
+    d0u8 = vdup_n_u8(bifilter4_coeff[xoffset][0]);
+    d1u8 = vdup_n_u8(bifilter4_coeff[xoffset][1]);
+
+    d2u8 = vld1_u8(src_ptr);
+    d3u8 = vld1_u8(src_ptr + 8);
+    d4u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+    d5u8 = vld1_u8(src_ptr);
+    d6u8 = vld1_u8(src_ptr + 8);
+    d7u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+    d8u8 = vld1_u8(src_ptr);
+    d9u8 = vld1_u8(src_ptr + 8);
+    d10u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+    d11u8 = vld1_u8(src_ptr);
+    d12u8 = vld1_u8(src_ptr + 8);
+    d13u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+
+    // First Pass: output_height lines x output_width columns (17x16)
+    tmpp = tmp;
+    for (i = 3; i > 0; i--) {
+        q7u16  = vmull_u8(d2u8, d0u8);
+        q8u16  = vmull_u8(d3u8, d0u8);
+        q9u16  = vmull_u8(d5u8, d0u8);
+        q10u16 = vmull_u8(d6u8, d0u8);
+        q11u16 = vmull_u8(d8u8, d0u8);
+        q12u16 = vmull_u8(d9u8, d0u8);
+        q13u16 = vmull_u8(d11u8, d0u8);
+        q14u16 = vmull_u8(d12u8, d0u8);
+
+        d2u8  = vext_u8(d2u8, d3u8, 1);
+        d5u8  = vext_u8(d5u8, d6u8, 1);
+        d8u8  = vext_u8(d8u8, d9u8, 1);
+        d11u8 = vext_u8(d11u8, d12u8, 1);
+
+        q7u16  = vmlal_u8(q7u16, d2u8, d1u8);
+        q9u16  = vmlal_u8(q9u16, d5u8, d1u8);
+        q11u16 = vmlal_u8(q11u16, d8u8, d1u8);
+        q13u16 = vmlal_u8(q13u16, d11u8, d1u8);
+
+        d3u8  = vext_u8(d3u8, d4u8, 1);
+        d6u8  = vext_u8(d6u8, d7u8, 1);
+        d9u8  = vext_u8(d9u8, d10u8, 1);
+        d12u8 = vext_u8(d12u8, d13u8, 1);
+
+        q8u16  = vmlal_u8(q8u16,  d3u8, d1u8);
+        q10u16 = vmlal_u8(q10u16, d6u8, d1u8);
+        q12u16 = vmlal_u8(q12u16, d9u8, d1u8);
+        q14u16 = vmlal_u8(q14u16, d12u8, d1u8);
+
+        d14u8 = vqrshrn_n_u16(q7u16, 7);
+        d15u8 = vqrshrn_n_u16(q8u16, 7);
+        d16u8 = vqrshrn_n_u16(q9u16, 7);
+        d17u8 = vqrshrn_n_u16(q10u16, 7);
+        d18u8 = vqrshrn_n_u16(q11u16, 7);
+        d19u8 = vqrshrn_n_u16(q12u16, 7);
+        d20u8 = vqrshrn_n_u16(q13u16, 7);
+        d21u8 = vqrshrn_n_u16(q14u16, 7);
+
+        d2u8 = vld1_u8(src_ptr);
+        d3u8 = vld1_u8(src_ptr + 8);
+        d4u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+        d5u8 = vld1_u8(src_ptr);
+        d6u8 = vld1_u8(src_ptr + 8);
+        d7u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+        d8u8 = vld1_u8(src_ptr);
+        d9u8 = vld1_u8(src_ptr + 8);
+        d10u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+        d11u8 = vld1_u8(src_ptr);
+        d12u8 = vld1_u8(src_ptr + 8);
+        d13u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+
+        q7u8 = vcombine_u8(d14u8, d15u8);
+        q8u8 = vcombine_u8(d16u8, d17u8);
+        q9u8 = vcombine_u8(d18u8, d19u8);
+        q10u8 = vcombine_u8(d20u8, d21u8);
+
+        vst1q_u8((uint8_t *)tmpp, q7u8); tmpp += 16;
+        vst1q_u8((uint8_t *)tmpp, q8u8); tmpp += 16;
+        vst1q_u8((uint8_t *)tmpp, q9u8); tmpp += 16;
+        vst1q_u8((uint8_t *)tmpp, q10u8); tmpp += 16;
+    }
+
+    // First-pass filtering for rest 5 lines
+    d14u8 = vld1_u8(src_ptr);
+    d15u8 = vld1_u8(src_ptr + 8);
+    d16u8 = vld1_u8(src_ptr + 16); src_ptr += src_pixels_per_line;
+
+    q9u16  = vmull_u8(d2u8, d0u8);
+    q10u16 = vmull_u8(d3u8, d0u8);
+    q11u16 = vmull_u8(d5u8, d0u8);
+    q12u16 = vmull_u8(d6u8, d0u8);
+    q13u16 = vmull_u8(d8u8, d0u8);
+    q14u16 = vmull_u8(d9u8, d0u8);
+
+    d2u8  = vext_u8(d2u8, d3u8, 1);
+    d5u8  = vext_u8(d5u8, d6u8, 1);
+    d8u8  = vext_u8(d8u8, d9u8, 1);
+
+    q9u16  = vmlal_u8(q9u16, d2u8, d1u8);
+    q11u16 = vmlal_u8(q11u16, d5u8, d1u8);
+    q13u16 = vmlal_u8(q13u16, d8u8, d1u8);
+
+    d3u8  = vext_u8(d3u8, d4u8, 1);
+    d6u8  = vext_u8(d6u8, d7u8, 1);
+    d9u8  = vext_u8(d9u8, d10u8, 1);
+
+    q10u16 = vmlal_u8(q10u16, d3u8, d1u8);
+    q12u16 = vmlal_u8(q12u16, d6u8, d1u8);
+    q14u16 = vmlal_u8(q14u16, d9u8, d1u8);
+
+    q1u16 = vmull_u8(d11u8, d0u8);
+    q2u16 = vmull_u8(d12u8, d0u8);
+    q3u16 = vmull_u8(d14u8, d0u8);
+    q4u16 = vmull_u8(d15u8, d0u8);
+
+    d11u8 = vext_u8(d11u8, d12u8, 1);
+    d14u8 = vext_u8(d14u8, d15u8, 1);
+
+    q1u16 = vmlal_u8(q1u16, d11u8, d1u8);
+    q3u16 = vmlal_u8(q3u16, d14u8, d1u8);
+
+    d12u8 = vext_u8(d12u8, d13u8, 1);
+    d15u8 = vext_u8(d15u8, d16u8, 1);
+
+    q2u16 = vmlal_u8(q2u16, d12u8, d1u8);
+    q4u16 = vmlal_u8(q4u16, d15u8, d1u8);
+
+    d10u8 = vqrshrn_n_u16(q9u16, 7);
+    d11u8 = vqrshrn_n_u16(q10u16, 7);
+    d12u8 = vqrshrn_n_u16(q11u16, 7);
+    d13u8 = vqrshrn_n_u16(q12u16, 7);
+    d14u8 = vqrshrn_n_u16(q13u16, 7);
+    d15u8 = vqrshrn_n_u16(q14u16, 7);
+    d16u8 = vqrshrn_n_u16(q1u16, 7);
+    d17u8 = vqrshrn_n_u16(q2u16, 7);
+    d18u8 = vqrshrn_n_u16(q3u16, 7);
+    d19u8 = vqrshrn_n_u16(q4u16, 7);
+
+    q5u8 = vcombine_u8(d10u8, d11u8);
+    q6u8 = vcombine_u8(d12u8, d13u8);
+    q7u8 = vcombine_u8(d14u8, d15u8);
+    q8u8 = vcombine_u8(d16u8, d17u8);
+    q9u8 = vcombine_u8(d18u8, d19u8);
+
+    vst1q_u8((uint8_t *)tmpp, q5u8); tmpp += 16;
+    vst1q_u8((uint8_t *)tmpp, q6u8); tmpp += 16;
+    vst1q_u8((uint8_t *)tmpp, q7u8); tmpp += 16;
+    vst1q_u8((uint8_t *)tmpp, q8u8); tmpp += 16;
+    vst1q_u8((uint8_t *)tmpp, q9u8);
+
+    // secondpass_filter
+    d0u8 = vdup_n_u8(bifilter4_coeff[yoffset][0]);
+    d1u8 = vdup_n_u8(bifilter4_coeff[yoffset][1]);
+
+    tmpp = tmp;
+    q11u8 = vld1q_u8(tmpp);
+    tmpp += 16;
+    for (i = 4; i > 0; i--) {
+        q12u8 = vld1q_u8(tmpp); tmpp += 16;
+        q13u8 = vld1q_u8(tmpp); tmpp += 16;
+        q14u8 = vld1q_u8(tmpp); tmpp += 16;
+        q15u8 = vld1q_u8(tmpp); tmpp += 16;
+
+        q1u16 = vmull_u8(vget_low_u8(q11u8), d0u8);
+        q2u16 = vmull_u8(vget_high_u8(q11u8), d0u8);
+        q3u16 = vmull_u8(vget_low_u8(q12u8), d0u8);
+        q4u16 = vmull_u8(vget_high_u8(q12u8), d0u8);
+        q5u16 = vmull_u8(vget_low_u8(q13u8), d0u8);
+        q6u16 = vmull_u8(vget_high_u8(q13u8), d0u8);
+        q7u16 = vmull_u8(vget_low_u8(q14u8), d0u8);
+        q8u16 = vmull_u8(vget_high_u8(q14u8), d0u8);
+
+        q1u16 = vmlal_u8(q1u16, vget_low_u8(q12u8), d1u8);
+        q2u16 = vmlal_u8(q2u16, vget_high_u8(q12u8), d1u8);
+        q3u16 = vmlal_u8(q3u16, vget_low_u8(q13u8), d1u8);
+        q4u16 = vmlal_u8(q4u16, vget_high_u8(q13u8), d1u8);
+        q5u16 = vmlal_u8(q5u16, vget_low_u8(q14u8), d1u8);
+        q6u16 = vmlal_u8(q6u16, vget_high_u8(q14u8), d1u8);
+        q7u16 = vmlal_u8(q7u16, vget_low_u8(q15u8), d1u8);
+        q8u16 = vmlal_u8(q8u16, vget_high_u8(q15u8), d1u8);
+
+        d2u8 = vqrshrn_n_u16(q1u16, 7);
+        d3u8 = vqrshrn_n_u16(q2u16, 7);
+        d4u8 = vqrshrn_n_u16(q3u16, 7);
+        d5u8 = vqrshrn_n_u16(q4u16, 7);
+        d6u8 = vqrshrn_n_u16(q5u16, 7);
+        d7u8 = vqrshrn_n_u16(q6u16, 7);
+        d8u8 = vqrshrn_n_u16(q7u16, 7);
+        d9u8 = vqrshrn_n_u16(q8u16, 7);
+
+        q1u8 = vcombine_u8(d2u8, d3u8);
+        q2u8 = vcombine_u8(d4u8, d5u8);
+        q3u8 = vcombine_u8(d6u8, d7u8);
+        q4u8 = vcombine_u8(d8u8, d9u8);
+
+        q11u8 = q15u8;
+
+        vst1q_u8((uint8_t *)dst_ptr, q1u8); dst_ptr += dst_pitch;
+        vst1q_u8((uint8_t *)dst_ptr, q2u8); dst_ptr += dst_pitch;
+        vst1q_u8((uint8_t *)dst_ptr, q3u8); dst_ptr += dst_pitch;
+        vst1q_u8((uint8_t *)dst_ptr, q4u8); dst_ptr += dst_pitch;
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/copymem_neon.c b/libvpx/libvpx/vp8/common/arm/neon/copymem_neon.c
new file mode 100644
index 0000000..deced11
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/copymem_neon.c
@@ -0,0 +1,59 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+void vp8_copy_mem8x4_neon(
+        unsigned char *src,
+        int src_stride,
+        unsigned char *dst,
+        int dst_stride) {
+    uint8x8_t vtmp;
+    int r;
+
+    for (r = 0; r < 4; r++) {
+        vtmp = vld1_u8(src);
+        vst1_u8(dst, vtmp);
+        src += src_stride;
+        dst += dst_stride;
+    }
+}
+
+void vp8_copy_mem8x8_neon(
+        unsigned char *src,
+        int src_stride,
+        unsigned char *dst,
+        int dst_stride) {
+    uint8x8_t vtmp;
+    int r;
+
+    for (r = 0; r < 8; r++) {
+        vtmp = vld1_u8(src);
+        vst1_u8(dst, vtmp);
+        src += src_stride;
+        dst += dst_stride;
+    }
+}
+
+void vp8_copy_mem16x16_neon(
+        unsigned char *src,
+        int src_stride,
+        unsigned char *dst,
+        int dst_stride) {
+    int r;
+    uint8x16_t qtmp;
+
+    for (r = 0; r < 16; r++) {
+        qtmp = vld1q_u8(src);
+        vst1q_u8(dst, qtmp);
+        src += src_stride;
+        dst += dst_stride;
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/dc_only_idct_add_neon.c b/libvpx/libvpx/vp8/common/arm/neon/dc_only_idct_add_neon.c
new file mode 100644
index 0000000..ad5f41d
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/dc_only_idct_add_neon.c
@@ -0,0 +1,42 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+void vp8_dc_only_idct_add_neon(
+        int16_t input_dc,
+        unsigned char *pred_ptr,
+        int pred_stride,
+        unsigned char *dst_ptr,
+        int dst_stride) {
+    int i;
+    uint16_t a1 = ((input_dc + 4) >> 3);
+    uint32x2_t d2u32 = vdup_n_u32(0);
+    uint8x8_t d2u8;
+    uint16x8_t q1u16;
+    uint16x8_t qAdd;
+
+    qAdd = vdupq_n_u16(a1);
+
+    for (i = 0; i < 2; i++) {
+        d2u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d2u32, 0);
+        pred_ptr += pred_stride;
+        d2u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d2u32, 1);
+        pred_ptr += pred_stride;
+
+        q1u16 = vaddw_u8(qAdd, vreinterpret_u8_u32(d2u32));
+        d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q1u16));
+
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d2u8), 0);
+        dst_ptr += dst_stride;
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d2u8), 1);
+        dst_ptr += dst_stride;
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/dequant_idct_neon.c b/libvpx/libvpx/vp8/common/arm/neon/dequant_idct_neon.c
new file mode 100644
index 0000000..58e1192
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/dequant_idct_neon.c
@@ -0,0 +1,142 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+static const int16_t cospi8sqrt2minus1 = 20091;
+static const int16_t sinpi8sqrt2       = 35468;
+
+void vp8_dequant_idct_add_neon(
+        int16_t *input,
+        int16_t *dq,
+        unsigned char *dst,
+        int stride) {
+    unsigned char *dst0;
+    int32x2_t d14, d15;
+    int16x4_t d2, d3, d4, d5, d10, d11, d12, d13;
+    int16x8_t q1, q2, q3, q4, q5, q6;
+    int16x8_t qEmpty = vdupq_n_s16(0);
+    int32x2x2_t d2tmp0, d2tmp1;
+    int16x4x2_t d2tmp2, d2tmp3;
+
+    d14 = d15 = vdup_n_s32(0);
+
+    // load input
+    q3 = vld1q_s16(input);
+    vst1q_s16(input, qEmpty);
+    input += 8;
+    q4 = vld1q_s16(input);
+    vst1q_s16(input, qEmpty);
+
+    // load dq
+    q5 = vld1q_s16(dq);
+    dq += 8;
+    q6 = vld1q_s16(dq);
+
+    // load src from dst
+    dst0 = dst;
+    d14 = vld1_lane_s32((const int32_t *)dst0, d14, 0);
+    dst0 += stride;
+    d14 = vld1_lane_s32((const int32_t *)dst0, d14, 1);
+    dst0 += stride;
+    d15 = vld1_lane_s32((const int32_t *)dst0, d15, 0);
+    dst0 += stride;
+    d15 = vld1_lane_s32((const int32_t *)dst0, d15, 1);
+
+    q1 = vreinterpretq_s16_u16(vmulq_u16(vreinterpretq_u16_s16(q3),
+                                         vreinterpretq_u16_s16(q5)));
+    q2 = vreinterpretq_s16_u16(vmulq_u16(vreinterpretq_u16_s16(q4),
+                                         vreinterpretq_u16_s16(q6)));
+
+    d12 = vqadd_s16(vget_low_s16(q1), vget_low_s16(q2));
+    d13 = vqsub_s16(vget_low_s16(q1), vget_low_s16(q2));
+
+    q2 = vcombine_s16(vget_high_s16(q1), vget_high_s16(q2));
+
+    q3 = vqdmulhq_n_s16(q2, sinpi8sqrt2);
+    q4 = vqdmulhq_n_s16(q2, cospi8sqrt2minus1);
+
+    q3 = vshrq_n_s16(q3, 1);
+    q4 = vshrq_n_s16(q4, 1);
+
+    q3 = vqaddq_s16(q3, q2);
+    q4 = vqaddq_s16(q4, q2);
+
+    d10 = vqsub_s16(vget_low_s16(q3), vget_high_s16(q4));
+    d11 = vqadd_s16(vget_high_s16(q3), vget_low_s16(q4));
+
+    d2 = vqadd_s16(d12, d11);
+    d3 = vqadd_s16(d13, d10);
+    d4 = vqsub_s16(d13, d10);
+    d5 = vqsub_s16(d12, d11);
+
+    d2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
+    d2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
+    d2tmp2 = vtrn_s16(vreinterpret_s16_s32(d2tmp0.val[0]),
+                      vreinterpret_s16_s32(d2tmp1.val[0]));
+    d2tmp3 = vtrn_s16(vreinterpret_s16_s32(d2tmp0.val[1]),
+                      vreinterpret_s16_s32(d2tmp1.val[1]));
+
+    // loop 2
+    q2 = vcombine_s16(d2tmp2.val[1], d2tmp3.val[1]);
+
+    q3 = vqdmulhq_n_s16(q2, sinpi8sqrt2);
+    q4 = vqdmulhq_n_s16(q2, cospi8sqrt2minus1);
+
+    d12 = vqadd_s16(d2tmp2.val[0], d2tmp3.val[0]);
+    d13 = vqsub_s16(d2tmp2.val[0], d2tmp3.val[0]);
+
+    q3 = vshrq_n_s16(q3, 1);
+    q4 = vshrq_n_s16(q4, 1);
+
+    q3 = vqaddq_s16(q3, q2);
+    q4 = vqaddq_s16(q4, q2);
+
+    d10 = vqsub_s16(vget_low_s16(q3), vget_high_s16(q4));
+    d11 = vqadd_s16(vget_high_s16(q3), vget_low_s16(q4));
+
+    d2 = vqadd_s16(d12, d11);
+    d3 = vqadd_s16(d13, d10);
+    d4 = vqsub_s16(d13, d10);
+    d5 = vqsub_s16(d12, d11);
+
+    d2 = vrshr_n_s16(d2, 3);
+    d3 = vrshr_n_s16(d3, 3);
+    d4 = vrshr_n_s16(d4, 3);
+    d5 = vrshr_n_s16(d5, 3);
+
+    d2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
+    d2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
+    d2tmp2 = vtrn_s16(vreinterpret_s16_s32(d2tmp0.val[0]),
+                      vreinterpret_s16_s32(d2tmp1.val[0]));
+    d2tmp3 = vtrn_s16(vreinterpret_s16_s32(d2tmp0.val[1]),
+                      vreinterpret_s16_s32(d2tmp1.val[1]));
+
+    q1 = vcombine_s16(d2tmp2.val[0], d2tmp2.val[1]);
+    q2 = vcombine_s16(d2tmp3.val[0], d2tmp3.val[1]);
+
+    q1 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q1),
+                                        vreinterpret_u8_s32(d14)));
+    q2 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2),
+                                        vreinterpret_u8_s32(d15)));
+
+    d14 = vreinterpret_s32_u8(vqmovun_s16(q1));
+    d15 = vreinterpret_s32_u8(vqmovun_s16(q2));
+
+    dst0 = dst;
+    vst1_lane_s32((int32_t *)dst0, d14, 0);
+    dst0 += stride;
+    vst1_lane_s32((int32_t *)dst0, d14, 1);
+    dst0 += stride;
+    vst1_lane_s32((int32_t *)dst0, d15, 0);
+    dst0 += stride;
+    vst1_lane_s32((int32_t *)dst0, d15, 1);
+    return;
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/dequantizeb_neon.c b/libvpx/libvpx/vp8/common/arm/neon/dequantizeb_neon.c
new file mode 100644
index 0000000..54e709d
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/dequantizeb_neon.c
@@ -0,0 +1,25 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "vp8/common/blockd.h"
+
+void vp8_dequantize_b_neon(BLOCKD *d, short *DQC) {
+    int16x8x2_t qQ, qDQC, qDQ;
+
+    qQ   = vld2q_s16(d->qcoeff);
+    qDQC = vld2q_s16(DQC);
+
+    qDQ.val[0] = vmulq_s16(qQ.val[0], qDQC.val[0]);
+    qDQ.val[1] = vmulq_s16(qQ.val[1], qDQC.val[1]);
+
+    vst2q_s16(d->dqcoeff, qDQ);
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/idct_blk_neon.c b/libvpx/libvpx/vp8/common/arm/neon/idct_blk_neon.c
new file mode 100644
index 0000000..fb327a7
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/idct_blk_neon.c
@@ -0,0 +1,96 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+
+/* place these declarations here because we don't want to maintain them
+ * outside of this scope
+ */
+void idct_dequant_full_2x_neon(short *q, short *dq,
+                               unsigned char *dst, int stride);
+void idct_dequant_0_2x_neon(short *q, short dq,
+                            unsigned char *dst, int stride);
+
+
+void vp8_dequant_idct_add_y_block_neon(short *q, short *dq,
+                                       unsigned char *dst,
+                                       int stride, char *eobs)
+{
+    int i;
+
+    for (i = 0; i < 4; i++)
+    {
+        if (((short *)(eobs))[0])
+        {
+            if (((short *)eobs)[0] & 0xfefe)
+                idct_dequant_full_2x_neon (q, dq, dst, stride);
+            else
+                idct_dequant_0_2x_neon (q, dq[0], dst, stride);
+        }
+
+        if (((short *)(eobs))[1])
+        {
+            if (((short *)eobs)[1] & 0xfefe)
+                idct_dequant_full_2x_neon (q+32, dq, dst+8, stride);
+            else
+                idct_dequant_0_2x_neon (q+32, dq[0], dst+8, stride);
+        }
+        q    += 64;
+        dst  += 4*stride;
+        eobs += 4;
+    }
+}
+
+void vp8_dequant_idct_add_uv_block_neon(short *q, short *dq,
+                                        unsigned char *dstu,
+                                        unsigned char *dstv,
+                                        int stride, char *eobs)
+{
+    if (((short *)(eobs))[0])
+    {
+        if (((short *)eobs)[0] & 0xfefe)
+            idct_dequant_full_2x_neon (q, dq, dstu, stride);
+        else
+            idct_dequant_0_2x_neon (q, dq[0], dstu, stride);
+    }
+
+    q    += 32;
+    dstu += 4*stride;
+
+    if (((short *)(eobs))[1])
+    {
+        if (((short *)eobs)[1] & 0xfefe)
+            idct_dequant_full_2x_neon (q, dq, dstu, stride);
+        else
+            idct_dequant_0_2x_neon (q, dq[0], dstu, stride);
+    }
+
+    q += 32;
+
+    if (((short *)(eobs))[2])
+    {
+        if (((short *)eobs)[2] & 0xfefe)
+            idct_dequant_full_2x_neon (q, dq, dstv, stride);
+        else
+            idct_dequant_0_2x_neon (q, dq[0], dstv, stride);
+    }
+
+    q    += 32;
+    dstv += 4*stride;
+
+    if (((short *)(eobs))[3])
+    {
+        if (((short *)eobs)[3] & 0xfefe)
+            idct_dequant_full_2x_neon (q, dq, dstv, stride);
+        else
+            idct_dequant_0_2x_neon (q, dq[0], dstv, stride);
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/idct_dequant_0_2x_neon.c b/libvpx/libvpx/vp8/common/arm/neon/idct_dequant_0_2x_neon.c
new file mode 100644
index 0000000..e6f862f
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/idct_dequant_0_2x_neon.c
@@ -0,0 +1,63 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+void idct_dequant_0_2x_neon(
+        int16_t *q,
+        int16_t dq,
+        unsigned char *dst,
+        int stride) {
+    unsigned char *dst0;
+    int i, a0, a1;
+    int16x8x2_t q2Add;
+    int32x2_t d2s32 = vdup_n_s32(0),
+              d4s32 = vdup_n_s32(0);
+    uint8x8_t d2u8, d4u8;
+    uint16x8_t q1u16, q2u16;
+
+    a0 = ((q[0] * dq) + 4) >> 3;
+    a1 = ((q[16] * dq) + 4) >> 3;
+    q[0] = q[16] = 0;
+    q2Add.val[0] = vdupq_n_s16((int16_t)a0);
+    q2Add.val[1] = vdupq_n_s16((int16_t)a1);
+
+    for (i = 0; i < 2; i++, dst += 4) {
+        dst0 = dst;
+        d2s32 = vld1_lane_s32((const int32_t *)dst0, d2s32, 0);
+        dst0 += stride;
+        d2s32 = vld1_lane_s32((const int32_t *)dst0, d2s32, 1);
+        dst0 += stride;
+        d4s32 = vld1_lane_s32((const int32_t *)dst0, d4s32, 0);
+        dst0 += stride;
+        d4s32 = vld1_lane_s32((const int32_t *)dst0, d4s32, 1);
+
+        q1u16 = vaddw_u8(vreinterpretq_u16_s16(q2Add.val[i]),
+                         vreinterpret_u8_s32(d2s32));
+        q2u16 = vaddw_u8(vreinterpretq_u16_s16(q2Add.val[i]),
+                         vreinterpret_u8_s32(d4s32));
+
+        d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q1u16));
+        d4u8 = vqmovun_s16(vreinterpretq_s16_u16(q2u16));
+
+        d2s32 = vreinterpret_s32_u8(d2u8);
+        d4s32 = vreinterpret_s32_u8(d4u8);
+
+        dst0 = dst;
+        vst1_lane_s32((int32_t *)dst0, d2s32, 0);
+        dst0 += stride;
+        vst1_lane_s32((int32_t *)dst0, d2s32, 1);
+        dst0 += stride;
+        vst1_lane_s32((int32_t *)dst0, d4s32, 0);
+        dst0 += stride;
+        vst1_lane_s32((int32_t *)dst0, d4s32, 1);
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/idct_dequant_full_2x_neon.c b/libvpx/libvpx/vp8/common/arm/neon/idct_dequant_full_2x_neon.c
new file mode 100644
index 0000000..a60ed46
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/idct_dequant_full_2x_neon.c
@@ -0,0 +1,185 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+static const int16_t cospi8sqrt2minus1 = 20091;
+static const int16_t sinpi8sqrt2       = 17734;
+// because the lowest bit in 0x8a8c is 0, we can pre-shift this
+
+void idct_dequant_full_2x_neon(
+        int16_t *q,
+        int16_t *dq,
+        unsigned char *dst,
+        int stride) {
+    unsigned char *dst0, *dst1;
+    int32x2_t d28, d29, d30, d31;
+    int16x8_t q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11;
+    int16x8_t qEmpty = vdupq_n_s16(0);
+    int32x4x2_t q2tmp0, q2tmp1;
+    int16x8x2_t q2tmp2, q2tmp3;
+    int16x4_t dLow0, dLow1, dHigh0, dHigh1;
+
+    d28 = d29 = d30 = d31 = vdup_n_s32(0);
+
+    // load dq
+    q0 = vld1q_s16(dq);
+    dq += 8;
+    q1 = vld1q_s16(dq);
+
+    // load q
+    q2 = vld1q_s16(q);
+    vst1q_s16(q, qEmpty);
+    q += 8;
+    q3 = vld1q_s16(q);
+    vst1q_s16(q, qEmpty);
+    q += 8;
+    q4 = vld1q_s16(q);
+    vst1q_s16(q, qEmpty);
+    q += 8;
+    q5 = vld1q_s16(q);
+    vst1q_s16(q, qEmpty);
+
+    // load src from dst
+    dst0 = dst;
+    dst1 = dst + 4;
+    d28 = vld1_lane_s32((const int32_t *)dst0, d28, 0);
+    dst0 += stride;
+    d28 = vld1_lane_s32((const int32_t *)dst1, d28, 1);
+    dst1 += stride;
+    d29 = vld1_lane_s32((const int32_t *)dst0, d29, 0);
+    dst0 += stride;
+    d29 = vld1_lane_s32((const int32_t *)dst1, d29, 1);
+    dst1 += stride;
+
+    d30 = vld1_lane_s32((const int32_t *)dst0, d30, 0);
+    dst0 += stride;
+    d30 = vld1_lane_s32((const int32_t *)dst1, d30, 1);
+    dst1 += stride;
+    d31 = vld1_lane_s32((const int32_t *)dst0, d31, 0);
+    d31 = vld1_lane_s32((const int32_t *)dst1, d31, 1);
+
+    q2 = vmulq_s16(q2, q0);
+    q3 = vmulq_s16(q3, q1);
+    q4 = vmulq_s16(q4, q0);
+    q5 = vmulq_s16(q5, q1);
+
+    // vswp
+    dLow0 = vget_low_s16(q2);
+    dHigh0 = vget_high_s16(q2);
+    dLow1 = vget_low_s16(q4);
+    dHigh1 = vget_high_s16(q4);
+    q2 = vcombine_s16(dLow0, dLow1);
+    q4 = vcombine_s16(dHigh0, dHigh1);
+
+    dLow0 = vget_low_s16(q3);
+    dHigh0 = vget_high_s16(q3);
+    dLow1 = vget_low_s16(q5);
+    dHigh1 = vget_high_s16(q5);
+    q3 = vcombine_s16(dLow0, dLow1);
+    q5 = vcombine_s16(dHigh0, dHigh1);
+
+    q6 = vqdmulhq_n_s16(q4, sinpi8sqrt2);
+    q7 = vqdmulhq_n_s16(q5, sinpi8sqrt2);
+    q8 = vqdmulhq_n_s16(q4, cospi8sqrt2minus1);
+    q9 = vqdmulhq_n_s16(q5, cospi8sqrt2minus1);
+
+    q10 = vqaddq_s16(q2, q3);
+    q11 = vqsubq_s16(q2, q3);
+
+    q8 = vshrq_n_s16(q8, 1);
+    q9 = vshrq_n_s16(q9, 1);
+
+    q4 = vqaddq_s16(q4, q8);
+    q5 = vqaddq_s16(q5, q9);
+
+    q2 = vqsubq_s16(q6, q5);
+    q3 = vqaddq_s16(q7, q4);
+
+    q4 = vqaddq_s16(q10, q3);
+    q5 = vqaddq_s16(q11, q2);
+    q6 = vqsubq_s16(q11, q2);
+    q7 = vqsubq_s16(q10, q3);
+
+    q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6));
+    q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7));
+    q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]),
+                       vreinterpretq_s16_s32(q2tmp1.val[0]));
+    q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]),
+                       vreinterpretq_s16_s32(q2tmp1.val[1]));
+
+    // loop 2
+    q8  = vqdmulhq_n_s16(q2tmp2.val[1], sinpi8sqrt2);
+    q9  = vqdmulhq_n_s16(q2tmp3.val[1], sinpi8sqrt2);
+    q10 = vqdmulhq_n_s16(q2tmp2.val[1], cospi8sqrt2minus1);
+    q11 = vqdmulhq_n_s16(q2tmp3.val[1], cospi8sqrt2minus1);
+
+    q2 = vqaddq_s16(q2tmp2.val[0], q2tmp3.val[0]);
+    q3 = vqsubq_s16(q2tmp2.val[0], q2tmp3.val[0]);
+
+    q10 = vshrq_n_s16(q10, 1);
+    q11 = vshrq_n_s16(q11, 1);
+
+    q10 = vqaddq_s16(q2tmp2.val[1], q10);
+    q11 = vqaddq_s16(q2tmp3.val[1], q11);
+
+    q8 = vqsubq_s16(q8, q11);
+    q9 = vqaddq_s16(q9, q10);
+
+    q4 = vqaddq_s16(q2, q9);
+    q5 = vqaddq_s16(q3, q8);
+    q6 = vqsubq_s16(q3, q8);
+    q7 = vqsubq_s16(q2, q9);
+
+    q4 = vrshrq_n_s16(q4, 3);
+    q5 = vrshrq_n_s16(q5, 3);
+    q6 = vrshrq_n_s16(q6, 3);
+    q7 = vrshrq_n_s16(q7, 3);
+
+    q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6));
+    q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7));
+    q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]),
+                       vreinterpretq_s16_s32(q2tmp1.val[0]));
+    q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]),
+                       vreinterpretq_s16_s32(q2tmp1.val[1]));
+
+    q4 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[0]),
+                                          vreinterpret_u8_s32(d28)));
+    q5 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[1]),
+                                          vreinterpret_u8_s32(d29)));
+    q6 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[0]),
+                                          vreinterpret_u8_s32(d30)));
+    q7 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[1]),
+                                          vreinterpret_u8_s32(d31)));
+
+    d28 = vreinterpret_s32_u8(vqmovun_s16(q4));
+    d29 = vreinterpret_s32_u8(vqmovun_s16(q5));
+    d30 = vreinterpret_s32_u8(vqmovun_s16(q6));
+    d31 = vreinterpret_s32_u8(vqmovun_s16(q7));
+
+    dst0 = dst;
+    dst1 = dst + 4;
+    vst1_lane_s32((int32_t *)dst0, d28, 0);
+    dst0 += stride;
+    vst1_lane_s32((int32_t *)dst1, d28, 1);
+    dst1 += stride;
+    vst1_lane_s32((int32_t *)dst0, d29, 0);
+    dst0 += stride;
+    vst1_lane_s32((int32_t *)dst1, d29, 1);
+    dst1 += stride;
+
+    vst1_lane_s32((int32_t *)dst0, d30, 0);
+    dst0 += stride;
+    vst1_lane_s32((int32_t *)dst1, d30, 1);
+    dst1 += stride;
+    vst1_lane_s32((int32_t *)dst0, d31, 0);
+    vst1_lane_s32((int32_t *)dst1, d31, 1);
+    return;
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/iwalsh_neon.c b/libvpx/libvpx/vp8/common/arm/neon/iwalsh_neon.c
new file mode 100644
index 0000000..6ea9dd7
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/iwalsh_neon.c
@@ -0,0 +1,102 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+void vp8_short_inv_walsh4x4_neon(
+        int16_t *input,
+        int16_t *mb_dqcoeff) {
+    int16x8_t q0s16, q1s16, q2s16, q3s16;
+    int16x4_t d4s16, d5s16, d6s16, d7s16;
+    int16x4x2_t v2tmp0, v2tmp1;
+    int32x2x2_t v2tmp2, v2tmp3;
+    int16x8_t qAdd3;
+
+    q0s16 = vld1q_s16(input);
+    q1s16 = vld1q_s16(input + 8);
+
+    // 1st for loop
+    d4s16 = vadd_s16(vget_low_s16(q0s16), vget_high_s16(q1s16));
+    d6s16 = vadd_s16(vget_high_s16(q0s16), vget_low_s16(q1s16));
+    d5s16 = vsub_s16(vget_low_s16(q0s16), vget_high_s16(q1s16));
+    d7s16 = vsub_s16(vget_high_s16(q0s16), vget_low_s16(q1s16));
+
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+
+    q0s16 = vaddq_s16(q2s16, q3s16);
+    q1s16 = vsubq_s16(q2s16, q3s16);
+
+    v2tmp2 = vtrn_s32(vreinterpret_s32_s16(vget_low_s16(q0s16)),
+                      vreinterpret_s32_s16(vget_low_s16(q1s16)));
+    v2tmp3 = vtrn_s32(vreinterpret_s32_s16(vget_high_s16(q0s16)),
+                      vreinterpret_s32_s16(vget_high_s16(q1s16)));
+    v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),
+                      vreinterpret_s16_s32(v2tmp3.val[0]));
+    v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),
+                      vreinterpret_s16_s32(v2tmp3.val[1]));
+
+    // 2nd for loop
+    d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
+    d6s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    d5s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
+    d7s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+
+    qAdd3 = vdupq_n_s16(3);
+
+    q0s16 = vaddq_s16(q2s16, q3s16);
+    q1s16 = vsubq_s16(q2s16, q3s16);
+
+    q0s16 = vaddq_s16(q0s16, qAdd3);
+    q1s16 = vaddq_s16(q1s16, qAdd3);
+
+    q0s16 = vshrq_n_s16(q0s16, 3);
+    q1s16 = vshrq_n_s16(q1s16, 3);
+
+    // store
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  0);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 0);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  0);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 0);
+    mb_dqcoeff += 16;
+
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  1);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 1);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  1);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 1);
+    mb_dqcoeff += 16;
+
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  2);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 2);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  2);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 2);
+    mb_dqcoeff += 16;
+
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  3);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 3);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  3);
+    mb_dqcoeff += 16;
+    vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 3);
+    mb_dqcoeff += 16;
+    return;
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.c b/libvpx/libvpx/vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.c
new file mode 100644
index 0000000..b25686f
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/loopfiltersimplehorizontaledge_neon.c
@@ -0,0 +1,111 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vpx_config.h"
+
+static INLINE void vp8_loop_filter_simple_horizontal_edge_neon(
+        unsigned char *s,
+        int p,
+        const unsigned char *blimit) {
+    uint8_t *sp;
+    uint8x16_t qblimit, q0u8;
+    uint8x16_t q5u8, q6u8, q7u8, q8u8, q9u8, q10u8, q14u8, q15u8;
+    int16x8_t q2s16, q3s16, q13s16;
+    int8x8_t d8s8, d9s8;
+    int8x16_t q2s8, q3s8, q4s8, q10s8, q11s8, q14s8;
+
+    qblimit = vdupq_n_u8(*blimit);
+
+    sp = s - (p << 1);
+    q5u8 = vld1q_u8(sp);
+    sp += p;
+    q6u8 = vld1q_u8(sp);
+    sp += p;
+    q7u8 = vld1q_u8(sp);
+    sp += p;
+    q8u8 = vld1q_u8(sp);
+
+    q15u8 = vabdq_u8(q6u8, q7u8);
+    q14u8 = vabdq_u8(q5u8, q8u8);
+
+    q15u8 = vqaddq_u8(q15u8, q15u8);
+    q14u8 = vshrq_n_u8(q14u8, 1);
+    q0u8 = vdupq_n_u8(0x80);
+    q13s16 = vdupq_n_s16(3);
+    q15u8 = vqaddq_u8(q15u8, q14u8);
+
+    q5u8 = veorq_u8(q5u8, q0u8);
+    q6u8 = veorq_u8(q6u8, q0u8);
+    q7u8 = veorq_u8(q7u8, q0u8);
+    q8u8 = veorq_u8(q8u8, q0u8);
+
+    q15u8 = vcgeq_u8(qblimit, q15u8);
+
+    q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7u8)),
+                     vget_low_s8(vreinterpretq_s8_u8(q6u8)));
+    q3s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7u8)),
+                     vget_high_s8(vreinterpretq_s8_u8(q6u8)));
+
+    q4s8 = vqsubq_s8(vreinterpretq_s8_u8(q5u8),
+                     vreinterpretq_s8_u8(q8u8));
+
+    q2s16 = vmulq_s16(q2s16, q13s16);
+    q3s16 = vmulq_s16(q3s16, q13s16);
+
+    q10u8 = vdupq_n_u8(3);
+    q9u8 = vdupq_n_u8(4);
+
+    q2s16 = vaddw_s8(q2s16, vget_low_s8(q4s8));
+    q3s16 = vaddw_s8(q3s16, vget_high_s8(q4s8));
+
+    d8s8 = vqmovn_s16(q2s16);
+    d9s8 = vqmovn_s16(q3s16);
+    q4s8 = vcombine_s8(d8s8, d9s8);
+
+    q14s8 = vandq_s8(q4s8, vreinterpretq_s8_u8(q15u8));
+
+    q2s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q10u8));
+    q3s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q9u8));
+    q2s8 = vshrq_n_s8(q2s8, 3);
+    q3s8 = vshrq_n_s8(q3s8, 3);
+
+    q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6u8), q2s8);
+    q10s8 = vqsubq_s8(vreinterpretq_s8_u8(q7u8), q3s8);
+
+    q6u8 = veorq_u8(vreinterpretq_u8_s8(q11s8), q0u8);
+    q7u8 = veorq_u8(vreinterpretq_u8_s8(q10s8), q0u8);
+
+    vst1q_u8(s, q7u8);
+    s -= p;
+    vst1q_u8(s, q6u8);
+    return;
+}
+
+void vp8_loop_filter_bhs_neon(
+        unsigned char *y_ptr,
+        int y_stride,
+        const unsigned char *blimit) {
+    y_ptr += y_stride * 4;
+    vp8_loop_filter_simple_horizontal_edge_neon(y_ptr, y_stride, blimit);
+    y_ptr += y_stride * 4;
+    vp8_loop_filter_simple_horizontal_edge_neon(y_ptr, y_stride, blimit);
+    y_ptr += y_stride * 4;
+    vp8_loop_filter_simple_horizontal_edge_neon(y_ptr, y_stride, blimit);
+    return;
+}
+
+void vp8_loop_filter_mbhs_neon(
+        unsigned char *y_ptr,
+        int y_stride,
+        const unsigned char *blimit) {
+    vp8_loop_filter_simple_horizontal_edge_neon(y_ptr, y_stride, blimit);
+    return;
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.c b/libvpx/libvpx/vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.c
new file mode 100644
index 0000000..921bcad
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/loopfiltersimpleverticaledge_neon.c
@@ -0,0 +1,283 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vpx_config.h"
+#include "vpx_ports/arm.h"
+
+#ifdef VPX_INCOMPATIBLE_GCC
+static INLINE void write_2x4(unsigned char *dst, int pitch,
+                             const uint8x8x2_t result) {
+    /*
+     * uint8x8x2_t result
+    00 01 02 03 | 04 05 06 07
+    10 11 12 13 | 14 15 16 17
+    ---
+    * after vtrn_u8
+    00 10 02 12 | 04 14 06 16
+    01 11 03 13 | 05 15 07 17
+    */
+    const uint8x8x2_t r01_u8 = vtrn_u8(result.val[0],
+                                       result.val[1]);
+    const uint16x4_t x_0_4 = vreinterpret_u16_u8(r01_u8.val[0]);
+    const uint16x4_t x_1_5 = vreinterpret_u16_u8(r01_u8.val[1]);
+    vst1_lane_u16((uint16_t *)dst, x_0_4, 0);
+    dst += pitch;
+    vst1_lane_u16((uint16_t *)dst, x_1_5, 0);
+    dst += pitch;
+    vst1_lane_u16((uint16_t *)dst, x_0_4, 1);
+    dst += pitch;
+    vst1_lane_u16((uint16_t *)dst, x_1_5, 1);
+    dst += pitch;
+    vst1_lane_u16((uint16_t *)dst, x_0_4, 2);
+    dst += pitch;
+    vst1_lane_u16((uint16_t *)dst, x_1_5, 2);
+    dst += pitch;
+    vst1_lane_u16((uint16_t *)dst, x_0_4, 3);
+    dst += pitch;
+    vst1_lane_u16((uint16_t *)dst, x_1_5, 3);
+}
+
+static INLINE void write_2x8(unsigned char *dst, int pitch,
+                             const uint8x8x2_t result,
+                             const uint8x8x2_t result2) {
+  write_2x4(dst, pitch, result);
+  dst += pitch * 8;
+  write_2x4(dst, pitch, result2);
+}
+#else
+static INLINE void write_2x8(unsigned char *dst, int pitch,
+                             const uint8x8x2_t result,
+                             const uint8x8x2_t result2) {
+  vst2_lane_u8(dst, result, 0);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 1);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 2);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 3);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 4);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 5);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 6);
+  dst += pitch;
+  vst2_lane_u8(dst, result, 7);
+  dst += pitch;
+
+  vst2_lane_u8(dst, result2, 0);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 1);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 2);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 3);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 4);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 5);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 6);
+  dst += pitch;
+  vst2_lane_u8(dst, result2, 7);
+}
+#endif  // VPX_INCOMPATIBLE_GCC
+
+
+#ifdef VPX_INCOMPATIBLE_GCC
+static INLINE
+uint8x8x4_t read_4x8(unsigned char *src, int pitch) {
+    uint8x8x4_t x;
+    const uint8x8_t a = vld1_u8(src);
+    const uint8x8_t b = vld1_u8(src + pitch * 1);
+    const uint8x8_t c = vld1_u8(src + pitch * 2);
+    const uint8x8_t d = vld1_u8(src + pitch * 3);
+    const uint8x8_t e = vld1_u8(src + pitch * 4);
+    const uint8x8_t f = vld1_u8(src + pitch * 5);
+    const uint8x8_t g = vld1_u8(src + pitch * 6);
+    const uint8x8_t h = vld1_u8(src + pitch * 7);
+    const uint32x2x2_t r04_u32 = vtrn_u32(vreinterpret_u32_u8(a),
+                                          vreinterpret_u32_u8(e));
+    const uint32x2x2_t r15_u32 = vtrn_u32(vreinterpret_u32_u8(b),
+                                          vreinterpret_u32_u8(f));
+    const uint32x2x2_t r26_u32 = vtrn_u32(vreinterpret_u32_u8(c),
+                                          vreinterpret_u32_u8(g));
+    const uint32x2x2_t r37_u32 = vtrn_u32(vreinterpret_u32_u8(d),
+                                          vreinterpret_u32_u8(h));
+    const uint16x4x2_t r02_u16 = vtrn_u16(vreinterpret_u16_u32(r04_u32.val[0]),
+                                          vreinterpret_u16_u32(r26_u32.val[0]));
+    const uint16x4x2_t r13_u16 = vtrn_u16(vreinterpret_u16_u32(r15_u32.val[0]),
+                                          vreinterpret_u16_u32(r37_u32.val[0]));
+    const uint8x8x2_t r01_u8 = vtrn_u8(vreinterpret_u8_u16(r02_u16.val[0]),
+                                       vreinterpret_u8_u16(r13_u16.val[0]));
+    const uint8x8x2_t r23_u8 = vtrn_u8(vreinterpret_u8_u16(r02_u16.val[1]),
+                                       vreinterpret_u8_u16(r13_u16.val[1]));
+    /*
+     * after vtrn_u32
+    00 01 02 03 | 40 41 42 43
+    10 11 12 13 | 50 51 52 53
+    20 21 22 23 | 60 61 62 63
+    30 31 32 33 | 70 71 72 73
+    ---
+    * after vtrn_u16
+    00 01 20 21 | 40 41 60 61
+    02 03 22 23 | 42 43 62 63
+    10 11 30 31 | 50 51 70 71
+    12 13 32 33 | 52 52 72 73
+
+    00 01 20 21 | 40 41 60 61
+    10 11 30 31 | 50 51 70 71
+    02 03 22 23 | 42 43 62 63
+    12 13 32 33 | 52 52 72 73
+    ---
+    * after vtrn_u8
+    00 10 20 30 | 40 50 60 70
+    01 11 21 31 | 41 51 61 71
+    02 12 22 32 | 42 52 62 72
+    03 13 23 33 | 43 53 63 73
+    */
+    x.val[0] = r01_u8.val[0];
+    x.val[1] = r01_u8.val[1];
+    x.val[2] = r23_u8.val[0];
+    x.val[3] = r23_u8.val[1];
+
+    return x;
+}
+#else
+static INLINE
+uint8x8x4_t read_4x8(unsigned char *src, int pitch) {
+    uint8x8x4_t x;
+    x.val[0] = x.val[1] = x.val[2] = x.val[3] = vdup_n_u8(0);
+    x = vld4_lane_u8(src, x, 0);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 1);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 2);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 3);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 4);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 5);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 6);
+    src += pitch;
+    x = vld4_lane_u8(src, x, 7);
+    return x;
+}
+#endif  // VPX_INCOMPATIBLE_GCC
+
+static INLINE void vp8_loop_filter_simple_vertical_edge_neon(
+        unsigned char *s,
+        int p,
+        const unsigned char *blimit) {
+    unsigned char *src1;
+    uint8x16_t qblimit, q0u8;
+    uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q11u8, q12u8, q14u8, q15u8;
+    int16x8_t q2s16, q13s16, q11s16;
+    int8x8_t d28s8, d29s8;
+    int8x16_t q2s8, q3s8, q10s8, q11s8, q14s8;
+    uint8x8x4_t d0u8x4;  // d6, d7, d8, d9
+    uint8x8x4_t d1u8x4;  // d10, d11, d12, d13
+    uint8x8x2_t d2u8x2;  // d12, d13
+    uint8x8x2_t d3u8x2;  // d14, d15
+
+    qblimit = vdupq_n_u8(*blimit);
+
+    src1 = s - 2;
+    d0u8x4 = read_4x8(src1, p);
+    src1 += p * 8;
+    d1u8x4 = read_4x8(src1, p);
+
+    q3u8 = vcombine_u8(d0u8x4.val[0], d1u8x4.val[0]);  // d6 d10
+    q4u8 = vcombine_u8(d0u8x4.val[2], d1u8x4.val[2]);  // d8 d12
+    q5u8 = vcombine_u8(d0u8x4.val[1], d1u8x4.val[1]);  // d7 d11
+    q6u8 = vcombine_u8(d0u8x4.val[3], d1u8x4.val[3]);  // d9 d13
+
+    q15u8 = vabdq_u8(q5u8, q4u8);
+    q14u8 = vabdq_u8(q3u8, q6u8);
+
+    q15u8 = vqaddq_u8(q15u8, q15u8);
+    q14u8 = vshrq_n_u8(q14u8, 1);
+    q0u8 = vdupq_n_u8(0x80);
+    q11s16 = vdupq_n_s16(3);
+    q15u8 = vqaddq_u8(q15u8, q14u8);
+
+    q3u8 = veorq_u8(q3u8, q0u8);
+    q4u8 = veorq_u8(q4u8, q0u8);
+    q5u8 = veorq_u8(q5u8, q0u8);
+    q6u8 = veorq_u8(q6u8, q0u8);
+
+    q15u8 = vcgeq_u8(qblimit, q15u8);
+
+    q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q4u8)),
+                     vget_low_s8(vreinterpretq_s8_u8(q5u8)));
+    q13s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q4u8)),
+                      vget_high_s8(vreinterpretq_s8_u8(q5u8)));
+
+    q14s8 = vqsubq_s8(vreinterpretq_s8_u8(q3u8),
+                      vreinterpretq_s8_u8(q6u8));
+
+    q2s16 = vmulq_s16(q2s16, q11s16);
+    q13s16 = vmulq_s16(q13s16, q11s16);
+
+    q11u8 = vdupq_n_u8(3);
+    q12u8 = vdupq_n_u8(4);
+
+    q2s16 = vaddw_s8(q2s16, vget_low_s8(q14s8));
+    q13s16 = vaddw_s8(q13s16, vget_high_s8(q14s8));
+
+    d28s8 = vqmovn_s16(q2s16);
+    d29s8 = vqmovn_s16(q13s16);
+    q14s8 = vcombine_s8(d28s8, d29s8);
+
+    q14s8 = vandq_s8(q14s8, vreinterpretq_s8_u8(q15u8));
+
+    q2s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q11u8));
+    q3s8 = vqaddq_s8(q14s8, vreinterpretq_s8_u8(q12u8));
+    q2s8 = vshrq_n_s8(q2s8, 3);
+    q14s8 = vshrq_n_s8(q3s8, 3);
+
+    q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q5u8), q2s8);
+    q10s8 = vqsubq_s8(vreinterpretq_s8_u8(q4u8), q14s8);
+
+    q6u8 = veorq_u8(vreinterpretq_u8_s8(q11s8), q0u8);
+    q7u8 = veorq_u8(vreinterpretq_u8_s8(q10s8), q0u8);
+
+    d2u8x2.val[0] = vget_low_u8(q6u8);   // d12
+    d2u8x2.val[1] = vget_low_u8(q7u8);   // d14
+    d3u8x2.val[0] = vget_high_u8(q6u8);  // d13
+    d3u8x2.val[1] = vget_high_u8(q7u8);  // d15
+
+    src1 = s - 1;
+    write_2x8(src1, p, d2u8x2, d3u8x2);
+}
+
+void vp8_loop_filter_bvs_neon(
+        unsigned char *y_ptr,
+        int y_stride,
+        const unsigned char *blimit) {
+    y_ptr += 4;
+    vp8_loop_filter_simple_vertical_edge_neon(y_ptr, y_stride, blimit);
+    y_ptr += 4;
+    vp8_loop_filter_simple_vertical_edge_neon(y_ptr, y_stride, blimit);
+    y_ptr += 4;
+    vp8_loop_filter_simple_vertical_edge_neon(y_ptr, y_stride, blimit);
+    return;
+}
+
+void vp8_loop_filter_mbvs_neon(
+        unsigned char *y_ptr,
+        int y_stride,
+        const unsigned char *blimit) {
+    vp8_loop_filter_simple_vertical_edge_neon(y_ptr, y_stride, blimit);
+    return;
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/mbloopfilter_neon.c b/libvpx/libvpx/vp8/common/arm/neon/mbloopfilter_neon.c
new file mode 100644
index 0000000..5351f4b
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/mbloopfilter_neon.c
@@ -0,0 +1,625 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vpx_config.h"
+
+static INLINE void vp8_mbloop_filter_neon(
+        uint8x16_t qblimit,  // mblimit
+        uint8x16_t qlimit,   // limit
+        uint8x16_t qthresh,  // thresh
+        uint8x16_t q3,       // p2
+        uint8x16_t q4,       // p2
+        uint8x16_t q5,       // p1
+        uint8x16_t q6,       // p0
+        uint8x16_t q7,       // q0
+        uint8x16_t q8,       // q1
+        uint8x16_t q9,       // q2
+        uint8x16_t q10,      // q3
+        uint8x16_t *q4r,     // p1
+        uint8x16_t *q5r,     // p1
+        uint8x16_t *q6r,     // p0
+        uint8x16_t *q7r,     // q0
+        uint8x16_t *q8r,     // q1
+        uint8x16_t *q9r) {   // q1
+    uint8x16_t q0u8, q1u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+    int16x8_t q0s16, q2s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int8x16_t q1s8, q6s8, q7s8, q2s8, q11s8, q13s8;
+    uint16x8_t q0u16, q11u16, q12u16, q13u16, q14u16, q15u16;
+    int8x16_t q0s8, q12s8, q14s8, q15s8;
+    int8x8_t d0, d1, d2, d3, d4, d5, d24, d25, d28, d29;
+
+    q11u8 = vabdq_u8(q3, q4);
+    q12u8 = vabdq_u8(q4, q5);
+    q13u8 = vabdq_u8(q5, q6);
+    q14u8 = vabdq_u8(q8, q7);
+    q1u8  = vabdq_u8(q9, q8);
+    q0u8  = vabdq_u8(q10, q9);
+
+    q11u8 = vmaxq_u8(q11u8, q12u8);
+    q12u8 = vmaxq_u8(q13u8, q14u8);
+    q1u8  = vmaxq_u8(q1u8, q0u8);
+    q15u8 = vmaxq_u8(q11u8, q12u8);
+
+    q12u8 = vabdq_u8(q6, q7);
+
+    // vp8_hevmask
+    q13u8 = vcgtq_u8(q13u8, qthresh);
+    q14u8 = vcgtq_u8(q14u8, qthresh);
+    q15u8 = vmaxq_u8(q15u8, q1u8);
+
+    q15u8 = vcgeq_u8(qlimit, q15u8);
+
+    q1u8 = vabdq_u8(q5, q8);
+    q12u8 = vqaddq_u8(q12u8, q12u8);
+
+    // vp8_filter() function
+    // convert to signed
+    q0u8 = vdupq_n_u8(0x80);
+    q9 = veorq_u8(q9, q0u8);
+    q8 = veorq_u8(q8, q0u8);
+    q7 = veorq_u8(q7, q0u8);
+    q6 = veorq_u8(q6, q0u8);
+    q5 = veorq_u8(q5, q0u8);
+    q4 = veorq_u8(q4, q0u8);
+
+    q1u8 = vshrq_n_u8(q1u8, 1);
+    q12u8 = vqaddq_u8(q12u8, q1u8);
+
+    q14u8 = vorrq_u8(q13u8, q14u8);
+    q12u8 = vcgeq_u8(qblimit, q12u8);
+
+    q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
+                     vget_low_s8(vreinterpretq_s8_u8(q6)));
+    q13s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
+                      vget_high_s8(vreinterpretq_s8_u8(q6)));
+
+    q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5),
+                     vreinterpretq_s8_u8(q8));
+
+    q11s16 = vdupq_n_s16(3);
+    q2s16  = vmulq_s16(q2s16, q11s16);
+    q13s16 = vmulq_s16(q13s16, q11s16);
+
+    q15u8 = vandq_u8(q15u8, q12u8);
+
+    q2s16  = vaddw_s8(q2s16, vget_low_s8(q1s8));
+    q13s16 = vaddw_s8(q13s16, vget_high_s8(q1s8));
+
+    q12u8 = vdupq_n_u8(3);
+    q11u8 = vdupq_n_u8(4);
+    // vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
+    d2 = vqmovn_s16(q2s16);
+    d3 = vqmovn_s16(q13s16);
+    q1s8 = vcombine_s8(d2, d3);
+    q1s8 = vandq_s8(q1s8, vreinterpretq_s8_u8(q15u8));
+    q13s8 = vandq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+    q2s8 = vqaddq_s8(q13s8, vreinterpretq_s8_u8(q11u8));
+    q13s8 = vqaddq_s8(q13s8, vreinterpretq_s8_u8(q12u8));
+    q2s8 = vshrq_n_s8(q2s8, 3);
+    q13s8 = vshrq_n_s8(q13s8, 3);
+
+    q7s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q2s8);
+    q6s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q13s8);
+
+    q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+    q0u16 = q11u16 = q12u16 = q13u16 = q14u16 = q15u16 = vdupq_n_u16(63);
+    d5 = vdup_n_s8(9);
+    d4 = vdup_n_s8(18);
+
+    q0s16  = vmlal_s8(vreinterpretq_s16_u16(q0u16),  vget_low_s8(q1s8),  d5);
+    q11s16 = vmlal_s8(vreinterpretq_s16_u16(q11u16), vget_high_s8(q1s8), d5);
+    d5 = vdup_n_s8(27);
+    q12s16 = vmlal_s8(vreinterpretq_s16_u16(q12u16), vget_low_s8(q1s8),  d4);
+    q13s16 = vmlal_s8(vreinterpretq_s16_u16(q13u16), vget_high_s8(q1s8), d4);
+    q14s16 = vmlal_s8(vreinterpretq_s16_u16(q14u16), vget_low_s8(q1s8),  d5);
+    q15s16 = vmlal_s8(vreinterpretq_s16_u16(q15u16), vget_high_s8(q1s8), d5);
+
+    d0  = vqshrn_n_s16(q0s16 , 7);
+    d1  = vqshrn_n_s16(q11s16, 7);
+    d24 = vqshrn_n_s16(q12s16, 7);
+    d25 = vqshrn_n_s16(q13s16, 7);
+    d28 = vqshrn_n_s16(q14s16, 7);
+    d29 = vqshrn_n_s16(q15s16, 7);
+
+    q0s8  = vcombine_s8(d0, d1);
+    q12s8 = vcombine_s8(d24, d25);
+    q14s8 = vcombine_s8(d28, d29);
+
+    q11s8 = vqsubq_s8(vreinterpretq_s8_u8(q9), q0s8);
+    q0s8  = vqaddq_s8(vreinterpretq_s8_u8(q4), q0s8);
+    q13s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q12s8);
+    q12s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q12s8);
+    q15s8 = vqsubq_s8((q7s8), q14s8);
+    q14s8 = vqaddq_s8((q6s8), q14s8);
+
+    q1u8 = vdupq_n_u8(0x80);
+    *q9r = veorq_u8(vreinterpretq_u8_s8(q11s8), q1u8);
+    *q8r = veorq_u8(vreinterpretq_u8_s8(q13s8), q1u8);
+    *q7r = veorq_u8(vreinterpretq_u8_s8(q15s8), q1u8);
+    *q6r = veorq_u8(vreinterpretq_u8_s8(q14s8), q1u8);
+    *q5r = veorq_u8(vreinterpretq_u8_s8(q12s8), q1u8);
+    *q4r = veorq_u8(vreinterpretq_u8_s8(q0s8), q1u8);
+    return;
+}
+
+void vp8_mbloop_filter_horizontal_edge_y_neon(
+        unsigned char *src,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh) {
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    src -= (pitch << 2);
+
+    q3 = vld1q_u8(src);
+    src += pitch;
+    q4 = vld1q_u8(src);
+    src += pitch;
+    q5 = vld1q_u8(src);
+    src += pitch;
+    q6 = vld1q_u8(src);
+    src += pitch;
+    q7 = vld1q_u8(src);
+    src += pitch;
+    q8 = vld1q_u8(src);
+    src += pitch;
+    q9 = vld1q_u8(src);
+    src += pitch;
+    q10 = vld1q_u8(src);
+
+    vp8_mbloop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q4, &q5, &q6, &q7, &q8, &q9);
+
+    src -= (pitch * 6);
+    vst1q_u8(src, q4);
+    src += pitch;
+    vst1q_u8(src, q5);
+    src += pitch;
+    vst1q_u8(src, q6);
+    src += pitch;
+    vst1q_u8(src, q7);
+    src += pitch;
+    vst1q_u8(src, q8);
+    src += pitch;
+    vst1q_u8(src, q9);
+    return;
+}
+
+void vp8_mbloop_filter_horizontal_edge_uv_neon(
+        unsigned char *u,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh,
+        unsigned char *v) {
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+    uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+    uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    u -= (pitch << 2);
+    v -= (pitch << 2);
+
+    d6 = vld1_u8(u);
+    u += pitch;
+    d7 = vld1_u8(v);
+    v += pitch;
+    d8 = vld1_u8(u);
+    u += pitch;
+    d9 = vld1_u8(v);
+    v += pitch;
+    d10 = vld1_u8(u);
+    u += pitch;
+    d11 = vld1_u8(v);
+    v += pitch;
+    d12 = vld1_u8(u);
+    u += pitch;
+    d13 = vld1_u8(v);
+    v += pitch;
+    d14 = vld1_u8(u);
+    u += pitch;
+    d15 = vld1_u8(v);
+    v += pitch;
+    d16 = vld1_u8(u);
+    u += pitch;
+    d17 = vld1_u8(v);
+    v += pitch;
+    d18 = vld1_u8(u);
+    u += pitch;
+    d19 = vld1_u8(v);
+    v += pitch;
+    d20 = vld1_u8(u);
+    d21 = vld1_u8(v);
+
+    q3 = vcombine_u8(d6, d7);
+    q4 = vcombine_u8(d8, d9);
+    q5 = vcombine_u8(d10, d11);
+    q6 = vcombine_u8(d12, d13);
+    q7 = vcombine_u8(d14, d15);
+    q8 = vcombine_u8(d16, d17);
+    q9 = vcombine_u8(d18, d19);
+    q10 = vcombine_u8(d20, d21);
+
+    vp8_mbloop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q4, &q5, &q6, &q7, &q8, &q9);
+
+    u -= (pitch * 6);
+    v -= (pitch * 6);
+    vst1_u8(u, vget_low_u8(q4));
+    u += pitch;
+    vst1_u8(v, vget_high_u8(q4));
+    v += pitch;
+    vst1_u8(u, vget_low_u8(q5));
+    u += pitch;
+    vst1_u8(v, vget_high_u8(q5));
+    v += pitch;
+    vst1_u8(u, vget_low_u8(q6));
+    u += pitch;
+    vst1_u8(v, vget_high_u8(q6));
+    v += pitch;
+    vst1_u8(u, vget_low_u8(q7));
+    u += pitch;
+    vst1_u8(v, vget_high_u8(q7));
+    v += pitch;
+    vst1_u8(u, vget_low_u8(q8));
+    u += pitch;
+    vst1_u8(v, vget_high_u8(q8));
+    v += pitch;
+    vst1_u8(u, vget_low_u8(q9));
+    vst1_u8(v, vget_high_u8(q9));
+    return;
+}
+
+void vp8_mbloop_filter_vertical_edge_y_neon(
+        unsigned char *src,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh) {
+    unsigned char *s1, *s2;
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+    uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+    uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+    uint32x4x2_t q2tmp0, q2tmp1, q2tmp2, q2tmp3;
+    uint16x8x2_t q2tmp4, q2tmp5, q2tmp6, q2tmp7;
+    uint8x16x2_t q2tmp8, q2tmp9, q2tmp10, q2tmp11;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    s1 = src - 4;
+    s2 = s1 + 8 * pitch;
+    d6  = vld1_u8(s1);
+    s1 += pitch;
+    d7  = vld1_u8(s2);
+    s2 += pitch;
+    d8  = vld1_u8(s1);
+    s1 += pitch;
+    d9  = vld1_u8(s2);
+    s2 += pitch;
+    d10 = vld1_u8(s1);
+    s1 += pitch;
+    d11 = vld1_u8(s2);
+    s2 += pitch;
+    d12 = vld1_u8(s1);
+    s1 += pitch;
+    d13 = vld1_u8(s2);
+    s2 += pitch;
+    d14 = vld1_u8(s1);
+    s1 += pitch;
+    d15 = vld1_u8(s2);
+    s2 += pitch;
+    d16 = vld1_u8(s1);
+    s1 += pitch;
+    d17 = vld1_u8(s2);
+    s2 += pitch;
+    d18 = vld1_u8(s1);
+    s1 += pitch;
+    d19 = vld1_u8(s2);
+    s2 += pitch;
+    d20 = vld1_u8(s1);
+    d21 = vld1_u8(s2);
+
+    q3 = vcombine_u8(d6, d7);
+    q4 = vcombine_u8(d8, d9);
+    q5 = vcombine_u8(d10, d11);
+    q6 = vcombine_u8(d12, d13);
+    q7 = vcombine_u8(d14, d15);
+    q8 = vcombine_u8(d16, d17);
+    q9 = vcombine_u8(d18, d19);
+    q10 = vcombine_u8(d20, d21);
+
+    q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+    q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+    q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+    q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+    q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+                       vreinterpretq_u16_u32(q2tmp2.val[0]));
+    q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+                       vreinterpretq_u16_u32(q2tmp3.val[0]));
+    q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+                       vreinterpretq_u16_u32(q2tmp2.val[1]));
+    q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+                       vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+    q2tmp8  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+                       vreinterpretq_u8_u16(q2tmp5.val[0]));
+    q2tmp9  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+                       vreinterpretq_u8_u16(q2tmp5.val[1]));
+    q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+                       vreinterpretq_u8_u16(q2tmp7.val[0]));
+    q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+                       vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+    q3 = q2tmp8.val[0];
+    q4 = q2tmp8.val[1];
+    q5 = q2tmp9.val[0];
+    q6 = q2tmp9.val[1];
+    q7 = q2tmp10.val[0];
+    q8 = q2tmp10.val[1];
+    q9 = q2tmp11.val[0];
+    q10 = q2tmp11.val[1];
+
+    vp8_mbloop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q4, &q5, &q6, &q7, &q8, &q9);
+
+    q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+    q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+    q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+    q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+    q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+                       vreinterpretq_u16_u32(q2tmp2.val[0]));
+    q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+                       vreinterpretq_u16_u32(q2tmp3.val[0]));
+    q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+                       vreinterpretq_u16_u32(q2tmp2.val[1]));
+    q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+                       vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+    q2tmp8  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+                       vreinterpretq_u8_u16(q2tmp5.val[0]));
+    q2tmp9  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+                       vreinterpretq_u8_u16(q2tmp5.val[1]));
+    q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+                       vreinterpretq_u8_u16(q2tmp7.val[0]));
+    q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+                       vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+    q3 = q2tmp8.val[0];
+    q4 = q2tmp8.val[1];
+    q5 = q2tmp9.val[0];
+    q6 = q2tmp9.val[1];
+    q7 = q2tmp10.val[0];
+    q8 = q2tmp10.val[1];
+    q9 = q2tmp11.val[0];
+    q10 = q2tmp11.val[1];
+
+    s1 -= 7 * pitch;
+    s2 -= 7 * pitch;
+
+    vst1_u8(s1, vget_low_u8(q3));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q3));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q4));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q4));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q5));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q5));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q6));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q6));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q7));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q7));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q8));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q8));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q9));
+    s1 += pitch;
+    vst1_u8(s2, vget_high_u8(q9));
+    s2 += pitch;
+    vst1_u8(s1, vget_low_u8(q10));
+    vst1_u8(s2, vget_high_u8(q10));
+    return;
+}
+
+void vp8_mbloop_filter_vertical_edge_uv_neon(
+        unsigned char *u,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh,
+        unsigned char *v) {
+    unsigned char *us, *ud;
+    unsigned char *vs, *vd;
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+    uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+    uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+    uint32x4x2_t q2tmp0, q2tmp1, q2tmp2, q2tmp3;
+    uint16x8x2_t q2tmp4, q2tmp5, q2tmp6, q2tmp7;
+    uint8x16x2_t q2tmp8, q2tmp9, q2tmp10, q2tmp11;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    us = u - 4;
+    vs = v - 4;
+    d6 = vld1_u8(us);
+    us += pitch;
+    d7 = vld1_u8(vs);
+    vs += pitch;
+    d8 = vld1_u8(us);
+    us += pitch;
+    d9 = vld1_u8(vs);
+    vs += pitch;
+    d10 = vld1_u8(us);
+    us += pitch;
+    d11 = vld1_u8(vs);
+    vs += pitch;
+    d12 = vld1_u8(us);
+    us += pitch;
+    d13 = vld1_u8(vs);
+    vs += pitch;
+    d14 = vld1_u8(us);
+    us += pitch;
+    d15 = vld1_u8(vs);
+    vs += pitch;
+    d16 = vld1_u8(us);
+    us += pitch;
+    d17 = vld1_u8(vs);
+    vs += pitch;
+    d18 = vld1_u8(us);
+    us += pitch;
+    d19 = vld1_u8(vs);
+    vs += pitch;
+    d20 = vld1_u8(us);
+    d21 = vld1_u8(vs);
+
+    q3 = vcombine_u8(d6, d7);
+    q4 = vcombine_u8(d8, d9);
+    q5 = vcombine_u8(d10, d11);
+    q6 = vcombine_u8(d12, d13);
+    q7 = vcombine_u8(d14, d15);
+    q8 = vcombine_u8(d16, d17);
+    q9 = vcombine_u8(d18, d19);
+    q10 = vcombine_u8(d20, d21);
+
+    q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+    q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+    q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+    q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+    q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+                       vreinterpretq_u16_u32(q2tmp2.val[0]));
+    q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+                       vreinterpretq_u16_u32(q2tmp3.val[0]));
+    q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+                       vreinterpretq_u16_u32(q2tmp2.val[1]));
+    q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+                       vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+    q2tmp8  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+                       vreinterpretq_u8_u16(q2tmp5.val[0]));
+    q2tmp9  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+                       vreinterpretq_u8_u16(q2tmp5.val[1]));
+    q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+                       vreinterpretq_u8_u16(q2tmp7.val[0]));
+    q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+                       vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+    q3 = q2tmp8.val[0];
+    q4 = q2tmp8.val[1];
+    q5 = q2tmp9.val[0];
+    q6 = q2tmp9.val[1];
+    q7 = q2tmp10.val[0];
+    q8 = q2tmp10.val[1];
+    q9 = q2tmp11.val[0];
+    q10 = q2tmp11.val[1];
+
+    vp8_mbloop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q4, &q5, &q6, &q7, &q8, &q9);
+
+    q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+    q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+    q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+    q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+    q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+                       vreinterpretq_u16_u32(q2tmp2.val[0]));
+    q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+                       vreinterpretq_u16_u32(q2tmp3.val[0]));
+    q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+                       vreinterpretq_u16_u32(q2tmp2.val[1]));
+    q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+                       vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+    q2tmp8  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+                       vreinterpretq_u8_u16(q2tmp5.val[0]));
+    q2tmp9  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+                       vreinterpretq_u8_u16(q2tmp5.val[1]));
+    q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+                       vreinterpretq_u8_u16(q2tmp7.val[0]));
+    q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+                       vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+    q3 = q2tmp8.val[0];
+    q4 = q2tmp8.val[1];
+    q5 = q2tmp9.val[0];
+    q6 = q2tmp9.val[1];
+    q7 = q2tmp10.val[0];
+    q8 = q2tmp10.val[1];
+    q9 = q2tmp11.val[0];
+    q10 = q2tmp11.val[1];
+
+    ud = u - 4;
+    vst1_u8(ud, vget_low_u8(q3));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q4));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q5));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q6));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q7));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q8));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q9));
+    ud += pitch;
+    vst1_u8(ud, vget_low_u8(q10));
+
+    vd = v - 4;
+    vst1_u8(vd, vget_high_u8(q3));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q4));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q5));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q6));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q7));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q8));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q9));
+    vd += pitch;
+    vst1_u8(vd, vget_high_u8(q10));
+    return;
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/shortidct4x4llm_neon.c b/libvpx/libvpx/vp8/common/arm/neon/shortidct4x4llm_neon.c
new file mode 100644
index 0000000..373afa6
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/shortidct4x4llm_neon.c
@@ -0,0 +1,123 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+static const int16_t cospi8sqrt2minus1 = 20091;
+static const int16_t sinpi8sqrt2       = 35468;
+
+void vp8_short_idct4x4llm_neon(
+        int16_t *input,
+        unsigned char *pred_ptr,
+        int pred_stride,
+        unsigned char *dst_ptr,
+        int dst_stride) {
+    int i;
+    uint32x2_t d6u32 = vdup_n_u32(0);
+    uint8x8_t d1u8;
+    int16x4_t d2, d3, d4, d5, d10, d11, d12, d13;
+    uint16x8_t q1u16;
+    int16x8_t q1s16, q2s16, q3s16, q4s16;
+    int32x2x2_t v2tmp0, v2tmp1;
+    int16x4x2_t v2tmp2, v2tmp3;
+
+    d2 = vld1_s16(input);
+    d3 = vld1_s16(input + 4);
+    d4 = vld1_s16(input + 8);
+    d5 = vld1_s16(input + 12);
+
+    // 1st for loop
+    q1s16 = vcombine_s16(d2, d4);  // Swap d3 d4 here
+    q2s16 = vcombine_s16(d3, d5);
+
+    q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2);
+    q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1);
+
+    d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // a1
+    d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // b1
+
+    q3s16 = vshrq_n_s16(q3s16, 1);
+    q4s16 = vshrq_n_s16(q4s16, 1);
+
+    q3s16 = vqaddq_s16(q3s16, q2s16);
+    q4s16 = vqaddq_s16(q4s16, q2s16);
+
+    d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16));  // c1
+    d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16));  // d1
+
+    d2 = vqadd_s16(d12, d11);
+    d3 = vqadd_s16(d13, d10);
+    d4 = vqsub_s16(d13, d10);
+    d5 = vqsub_s16(d12, d11);
+
+    v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
+    v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
+    v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]),
+                      vreinterpret_s16_s32(v2tmp1.val[0]));
+    v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]),
+                      vreinterpret_s16_s32(v2tmp1.val[1]));
+
+    // 2nd for loop
+    q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp3.val[0]);
+    q2s16 = vcombine_s16(v2tmp2.val[1], v2tmp3.val[1]);
+
+    q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2);
+    q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1);
+
+    d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // a1
+    d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // b1
+
+    q3s16 = vshrq_n_s16(q3s16, 1);
+    q4s16 = vshrq_n_s16(q4s16, 1);
+
+    q3s16 = vqaddq_s16(q3s16, q2s16);
+    q4s16 = vqaddq_s16(q4s16, q2s16);
+
+    d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16));  // c1
+    d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16));  // d1
+
+    d2 = vqadd_s16(d12, d11);
+    d3 = vqadd_s16(d13, d10);
+    d4 = vqsub_s16(d13, d10);
+    d5 = vqsub_s16(d12, d11);
+
+    d2 = vrshr_n_s16(d2, 3);
+    d3 = vrshr_n_s16(d3, 3);
+    d4 = vrshr_n_s16(d4, 3);
+    d5 = vrshr_n_s16(d5, 3);
+
+    v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
+    v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
+    v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]),
+                      vreinterpret_s16_s32(v2tmp1.val[0]));
+    v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]),
+                      vreinterpret_s16_s32(v2tmp1.val[1]));
+
+    q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp2.val[1]);
+    q2s16 = vcombine_s16(v2tmp3.val[0], v2tmp3.val[1]);
+
+    // dc_only_idct_add
+    for (i = 0; i < 2; i++, q1s16 = q2s16) {
+        d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 0);
+        pred_ptr += pred_stride;
+        d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 1);
+        pred_ptr += pred_stride;
+
+        q1u16 = vaddw_u8(vreinterpretq_u16_s16(q1s16),
+                         vreinterpret_u8_u32(d6u32));
+        d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q1u16));
+
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 0);
+        dst_ptr += dst_stride;
+        vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 1);
+        dst_ptr += dst_stride;
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/sixtappredict_neon.c b/libvpx/libvpx/vp8/common/arm/neon/sixtappredict_neon.c
new file mode 100644
index 0000000..49d8d22
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/sixtappredict_neon.c
@@ -0,0 +1,1377 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "vpx_ports/mem.h"
+
+static const int8_t vp8_sub_pel_filters[8][8] = {
+    {0,  0,  128,   0,   0, 0, 0, 0},  /* note that 1/8 pel positionyys are */
+    {0, -6,  123,  12,  -1, 0, 0, 0},  /*    just as per alpha -0.5 bicubic */
+    {2, -11, 108,  36,  -8, 1, 0, 0},  /* New 1/4 pel 6 tap filter */
+    {0, -9,   93,  50,  -6, 0, 0, 0},
+    {3, -16,  77,  77, -16, 3, 0, 0},  /* New 1/2 pel 6 tap filter */
+    {0, -6,   50,  93,  -9, 0, 0, 0},
+    {1, -8,   36, 108, -11, 2, 0, 0},  /* New 1/4 pel 6 tap filter */
+    {0, -1,   12, 123,  -6, 0, 0, 0},
+};
+
+void vp8_sixtap_predict8x4_neon(
+        unsigned char *src_ptr,
+        int src_pixels_per_line,
+        int xoffset,
+        int yoffset,
+        unsigned char *dst_ptr,
+        int dst_pitch) {
+    unsigned char *src;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
+    uint8x8_t d22u8, d23u8, d24u8, d25u8, d26u8;
+    uint8x8_t d27u8, d28u8, d29u8, d30u8, d31u8;
+    int8x8_t dtmps8, d0s8, d1s8, d2s8, d3s8, d4s8, d5s8;
+    uint16x8_t q3u16, q4u16, q5u16, q6u16, q7u16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16, q12u16;
+    int16x8_t q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16;
+    uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8;
+
+    if (xoffset == 0) {  // secondpass_filter8x4_only
+        // load second_pass filter
+        dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+        d0s8 = vdup_lane_s8(dtmps8, 0);
+        d1s8 = vdup_lane_s8(dtmps8, 1);
+        d2s8 = vdup_lane_s8(dtmps8, 2);
+        d3s8 = vdup_lane_s8(dtmps8, 3);
+        d4s8 = vdup_lane_s8(dtmps8, 4);
+        d5s8 = vdup_lane_s8(dtmps8, 5);
+        d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+        d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+        d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+        d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+        d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+        d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+        // load src data
+        src = src_ptr - src_pixels_per_line * 2;
+        d22u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d23u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d24u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d25u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d26u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d27u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d28u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d29u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d30u8 = vld1_u8(src);
+
+        q3u16 = vmull_u8(d22u8, d0u8);
+        q4u16 = vmull_u8(d23u8, d0u8);
+        q5u16 = vmull_u8(d24u8, d0u8);
+        q6u16 = vmull_u8(d25u8, d0u8);
+
+        q3u16 = vmlsl_u8(q3u16, d23u8, d1u8);
+        q4u16 = vmlsl_u8(q4u16, d24u8, d1u8);
+        q5u16 = vmlsl_u8(q5u16, d25u8, d1u8);
+        q6u16 = vmlsl_u8(q6u16, d26u8, d1u8);
+
+        q3u16 = vmlsl_u8(q3u16, d26u8, d4u8);
+        q4u16 = vmlsl_u8(q4u16, d27u8, d4u8);
+        q5u16 = vmlsl_u8(q5u16, d28u8, d4u8);
+        q6u16 = vmlsl_u8(q6u16, d29u8, d4u8);
+
+        q3u16 = vmlal_u8(q3u16, d24u8, d2u8);
+        q4u16 = vmlal_u8(q4u16, d25u8, d2u8);
+        q5u16 = vmlal_u8(q5u16, d26u8, d2u8);
+        q6u16 = vmlal_u8(q6u16, d27u8, d2u8);
+
+        q3u16 = vmlal_u8(q3u16, d27u8, d5u8);
+        q4u16 = vmlal_u8(q4u16, d28u8, d5u8);
+        q5u16 = vmlal_u8(q5u16, d29u8, d5u8);
+        q6u16 = vmlal_u8(q6u16, d30u8, d5u8);
+
+        q7u16 = vmull_u8(d25u8, d3u8);
+        q8u16 = vmull_u8(d26u8, d3u8);
+        q9u16 = vmull_u8(d27u8, d3u8);
+        q10u16 = vmull_u8(d28u8, d3u8);
+
+        q3s16 = vreinterpretq_s16_u16(q3u16);
+        q4s16 = vreinterpretq_s16_u16(q4u16);
+        q5s16 = vreinterpretq_s16_u16(q5u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+        q7s16 = vreinterpretq_s16_u16(q7u16);
+        q8s16 = vreinterpretq_s16_u16(q8u16);
+        q9s16 = vreinterpretq_s16_u16(q9u16);
+        q10s16 = vreinterpretq_s16_u16(q10u16);
+
+        q7s16 = vqaddq_s16(q7s16, q3s16);
+        q8s16 = vqaddq_s16(q8s16, q4s16);
+        q9s16 = vqaddq_s16(q9s16, q5s16);
+        q10s16 = vqaddq_s16(q10s16, q6s16);
+
+        d6u8 = vqrshrun_n_s16(q7s16, 7);
+        d7u8 = vqrshrun_n_s16(q8s16, 7);
+        d8u8 = vqrshrun_n_s16(q9s16, 7);
+        d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+        vst1_u8(dst_ptr, d6u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d7u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d8u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d9u8);
+        return;
+    }
+
+    // load first_pass filter
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[xoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    // First pass: output_height lines x output_width columns (9x4)
+    if (yoffset == 0)  // firstpass_filter4x4_only
+        src = src_ptr - 2;
+    else
+        src = src_ptr - 2 - (src_pixels_per_line * 2);
+    q3u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q4u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q5u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q6u8 = vld1q_u8(src);
+
+    q7u16  = vmull_u8(vget_low_u8(q3u8), d0u8);
+    q8u16  = vmull_u8(vget_low_u8(q4u8), d0u8);
+    q9u16  = vmull_u8(vget_low_u8(q5u8), d0u8);
+    q10u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
+
+    d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+    d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+    d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
+    d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
+
+    q7u16  = vmlsl_u8(q7u16, d28u8, d1u8);
+    q8u16  = vmlsl_u8(q8u16, d29u8, d1u8);
+    q9u16  = vmlsl_u8(q9u16, d30u8, d1u8);
+    q10u16 = vmlsl_u8(q10u16, d31u8, d1u8);
+
+    d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
+    d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
+    d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
+    d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
+
+    q7u16  = vmlsl_u8(q7u16, d28u8, d4u8);
+    q8u16  = vmlsl_u8(q8u16, d29u8, d4u8);
+    q9u16  = vmlsl_u8(q9u16, d30u8, d4u8);
+    q10u16 = vmlsl_u8(q10u16, d31u8, d4u8);
+
+    d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
+    d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
+    d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
+    d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
+
+    q7u16  = vmlal_u8(q7u16, d28u8, d2u8);
+    q8u16  = vmlal_u8(q8u16, d29u8, d2u8);
+    q9u16  = vmlal_u8(q9u16, d30u8, d2u8);
+    q10u16 = vmlal_u8(q10u16, d31u8, d2u8);
+
+    d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
+    d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
+    d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
+    d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
+
+    q7u16 = vmlal_u8(q7u16, d28u8, d5u8);
+    q8u16 = vmlal_u8(q8u16, d29u8, d5u8);
+    q9u16 = vmlal_u8(q9u16, d30u8, d5u8);
+    q10u16 = vmlal_u8(q10u16, d31u8, d5u8);
+
+    d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
+    d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
+    d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
+    d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
+
+    q3u16 = vmull_u8(d28u8, d3u8);
+    q4u16 = vmull_u8(d29u8, d3u8);
+    q5u16 = vmull_u8(d30u8, d3u8);
+    q6u16 = vmull_u8(d31u8, d3u8);
+
+    q3s16 = vreinterpretq_s16_u16(q3u16);
+    q4s16 = vreinterpretq_s16_u16(q4u16);
+    q5s16 = vreinterpretq_s16_u16(q5u16);
+    q6s16 = vreinterpretq_s16_u16(q6u16);
+    q7s16 = vreinterpretq_s16_u16(q7u16);
+    q8s16 = vreinterpretq_s16_u16(q8u16);
+    q9s16 = vreinterpretq_s16_u16(q9u16);
+    q10s16 = vreinterpretq_s16_u16(q10u16);
+
+    q7s16 = vqaddq_s16(q7s16, q3s16);
+    q8s16 = vqaddq_s16(q8s16, q4s16);
+    q9s16 = vqaddq_s16(q9s16, q5s16);
+    q10s16 = vqaddq_s16(q10s16, q6s16);
+
+    d22u8 = vqrshrun_n_s16(q7s16, 7);
+    d23u8 = vqrshrun_n_s16(q8s16, 7);
+    d24u8 = vqrshrun_n_s16(q9s16, 7);
+    d25u8 = vqrshrun_n_s16(q10s16, 7);
+
+    if (yoffset == 0) {  // firstpass_filter8x4_only
+        vst1_u8(dst_ptr, d22u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d23u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d24u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d25u8);
+        return;
+    }
+
+    // First Pass on rest 5-line data
+    src += src_pixels_per_line;
+    q3u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q4u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q5u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q6u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q7u8 = vld1q_u8(src);
+
+    q8u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
+    q9u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
+    q10u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
+    q11u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
+    q12u16 = vmull_u8(vget_low_u8(q7u8), d0u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 1);
+
+    q8u16  = vmlsl_u8(q8u16, d27u8, d1u8);
+    q9u16  = vmlsl_u8(q9u16, d28u8, d1u8);
+    q10u16 = vmlsl_u8(q10u16, d29u8, d1u8);
+    q11u16 = vmlsl_u8(q11u16, d30u8, d1u8);
+    q12u16 = vmlsl_u8(q12u16, d31u8, d1u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 4);
+
+    q8u16  = vmlsl_u8(q8u16, d27u8, d4u8);
+    q9u16  = vmlsl_u8(q9u16, d28u8, d4u8);
+    q10u16 = vmlsl_u8(q10u16, d29u8, d4u8);
+    q11u16 = vmlsl_u8(q11u16, d30u8, d4u8);
+    q12u16 = vmlsl_u8(q12u16, d31u8, d4u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 2);
+
+    q8u16  = vmlal_u8(q8u16, d27u8, d2u8);
+    q9u16  = vmlal_u8(q9u16, d28u8, d2u8);
+    q10u16 = vmlal_u8(q10u16, d29u8, d2u8);
+    q11u16 = vmlal_u8(q11u16, d30u8, d2u8);
+    q12u16 = vmlal_u8(q12u16, d31u8, d2u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 5);
+
+    q8u16  = vmlal_u8(q8u16, d27u8, d5u8);
+    q9u16  = vmlal_u8(q9u16, d28u8, d5u8);
+    q10u16 = vmlal_u8(q10u16, d29u8, d5u8);
+    q11u16 = vmlal_u8(q11u16, d30u8, d5u8);
+    q12u16 = vmlal_u8(q12u16, d31u8, d5u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 3);
+
+    q3u16 = vmull_u8(d27u8, d3u8);
+    q4u16 = vmull_u8(d28u8, d3u8);
+    q5u16 = vmull_u8(d29u8, d3u8);
+    q6u16 = vmull_u8(d30u8, d3u8);
+    q7u16 = vmull_u8(d31u8, d3u8);
+
+    q3s16 = vreinterpretq_s16_u16(q3u16);
+    q4s16 = vreinterpretq_s16_u16(q4u16);
+    q5s16 = vreinterpretq_s16_u16(q5u16);
+    q6s16 = vreinterpretq_s16_u16(q6u16);
+    q7s16 = vreinterpretq_s16_u16(q7u16);
+    q8s16 = vreinterpretq_s16_u16(q8u16);
+    q9s16 = vreinterpretq_s16_u16(q9u16);
+    q10s16 = vreinterpretq_s16_u16(q10u16);
+    q11s16 = vreinterpretq_s16_u16(q11u16);
+    q12s16 = vreinterpretq_s16_u16(q12u16);
+
+    q8s16 = vqaddq_s16(q8s16, q3s16);
+    q9s16 = vqaddq_s16(q9s16, q4s16);
+    q10s16 = vqaddq_s16(q10s16, q5s16);
+    q11s16 = vqaddq_s16(q11s16, q6s16);
+    q12s16 = vqaddq_s16(q12s16, q7s16);
+
+    d26u8 = vqrshrun_n_s16(q8s16, 7);
+    d27u8 = vqrshrun_n_s16(q9s16, 7);
+    d28u8 = vqrshrun_n_s16(q10s16, 7);
+    d29u8 = vqrshrun_n_s16(q11s16, 7);
+    d30u8 = vqrshrun_n_s16(q12s16, 7);
+
+    // Second pass: 8x4
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    q3u16 = vmull_u8(d22u8, d0u8);
+    q4u16 = vmull_u8(d23u8, d0u8);
+    q5u16 = vmull_u8(d24u8, d0u8);
+    q6u16 = vmull_u8(d25u8, d0u8);
+
+    q3u16 = vmlsl_u8(q3u16, d23u8, d1u8);
+    q4u16 = vmlsl_u8(q4u16, d24u8, d1u8);
+    q5u16 = vmlsl_u8(q5u16, d25u8, d1u8);
+    q6u16 = vmlsl_u8(q6u16, d26u8, d1u8);
+
+    q3u16 = vmlsl_u8(q3u16, d26u8, d4u8);
+    q4u16 = vmlsl_u8(q4u16, d27u8, d4u8);
+    q5u16 = vmlsl_u8(q5u16, d28u8, d4u8);
+    q6u16 = vmlsl_u8(q6u16, d29u8, d4u8);
+
+    q3u16 = vmlal_u8(q3u16, d24u8, d2u8);
+    q4u16 = vmlal_u8(q4u16, d25u8, d2u8);
+    q5u16 = vmlal_u8(q5u16, d26u8, d2u8);
+    q6u16 = vmlal_u8(q6u16, d27u8, d2u8);
+
+    q3u16 = vmlal_u8(q3u16, d27u8, d5u8);
+    q4u16 = vmlal_u8(q4u16, d28u8, d5u8);
+    q5u16 = vmlal_u8(q5u16, d29u8, d5u8);
+    q6u16 = vmlal_u8(q6u16, d30u8, d5u8);
+
+    q7u16 = vmull_u8(d25u8, d3u8);
+    q8u16 = vmull_u8(d26u8, d3u8);
+    q9u16 = vmull_u8(d27u8, d3u8);
+    q10u16 = vmull_u8(d28u8, d3u8);
+
+    q3s16 = vreinterpretq_s16_u16(q3u16);
+    q4s16 = vreinterpretq_s16_u16(q4u16);
+    q5s16 = vreinterpretq_s16_u16(q5u16);
+    q6s16 = vreinterpretq_s16_u16(q6u16);
+    q7s16 = vreinterpretq_s16_u16(q7u16);
+    q8s16 = vreinterpretq_s16_u16(q8u16);
+    q9s16 = vreinterpretq_s16_u16(q9u16);
+    q10s16 = vreinterpretq_s16_u16(q10u16);
+
+    q7s16 = vqaddq_s16(q7s16, q3s16);
+    q8s16 = vqaddq_s16(q8s16, q4s16);
+    q9s16 = vqaddq_s16(q9s16, q5s16);
+    q10s16 = vqaddq_s16(q10s16, q6s16);
+
+    d6u8 = vqrshrun_n_s16(q7s16, 7);
+    d7u8 = vqrshrun_n_s16(q8s16, 7);
+    d8u8 = vqrshrun_n_s16(q9s16, 7);
+    d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+    vst1_u8(dst_ptr, d6u8);
+    dst_ptr += dst_pitch;
+    vst1_u8(dst_ptr, d7u8);
+    dst_ptr += dst_pitch;
+    vst1_u8(dst_ptr, d8u8);
+    dst_ptr += dst_pitch;
+    vst1_u8(dst_ptr, d9u8);
+    return;
+}
+
+void vp8_sixtap_predict8x8_neon(
+        unsigned char *src_ptr,
+        int src_pixels_per_line,
+        int xoffset,
+        int yoffset,
+        unsigned char *dst_ptr,
+        int dst_pitch) {
+    unsigned char *src, *tmpp;
+    unsigned char tmp[64];
+    int i;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
+    uint8x8_t d18u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8, d25u8;
+    uint8x8_t d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
+    int8x8_t dtmps8, d0s8, d1s8, d2s8, d3s8, d4s8, d5s8;
+    uint16x8_t q3u16, q4u16, q5u16, q6u16, q7u16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16, q12u16;
+    int16x8_t q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16;
+    uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q9u8, q10u8, q11u8, q12u8;
+
+    if (xoffset == 0) {  // secondpass_filter8x8_only
+        // load second_pass filter
+        dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+        d0s8 = vdup_lane_s8(dtmps8, 0);
+        d1s8 = vdup_lane_s8(dtmps8, 1);
+        d2s8 = vdup_lane_s8(dtmps8, 2);
+        d3s8 = vdup_lane_s8(dtmps8, 3);
+        d4s8 = vdup_lane_s8(dtmps8, 4);
+        d5s8 = vdup_lane_s8(dtmps8, 5);
+        d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+        d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+        d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+        d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+        d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+        d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+        // load src data
+        src = src_ptr - src_pixels_per_line * 2;
+        d18u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d19u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d20u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d21u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d22u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d23u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d24u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d25u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d26u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d27u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d28u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d29u8 = vld1_u8(src);
+        src += src_pixels_per_line;
+        d30u8 = vld1_u8(src);
+
+        for (i = 2; i > 0; i--) {
+            q3u16 = vmull_u8(d18u8, d0u8);
+            q4u16 = vmull_u8(d19u8, d0u8);
+            q5u16 = vmull_u8(d20u8, d0u8);
+            q6u16 = vmull_u8(d21u8, d0u8);
+
+            q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
+            q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
+            q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
+            q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
+
+            q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
+            q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
+            q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
+            q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
+
+            q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
+            q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
+            q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
+            q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
+
+            q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
+            q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
+            q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
+            q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
+
+            q7u16 = vmull_u8(d21u8, d3u8);
+            q8u16 = vmull_u8(d22u8, d3u8);
+            q9u16 = vmull_u8(d23u8, d3u8);
+            q10u16 = vmull_u8(d24u8, d3u8);
+
+            q3s16 = vreinterpretq_s16_u16(q3u16);
+            q4s16 = vreinterpretq_s16_u16(q4u16);
+            q5s16 = vreinterpretq_s16_u16(q5u16);
+            q6s16 = vreinterpretq_s16_u16(q6u16);
+            q7s16 = vreinterpretq_s16_u16(q7u16);
+            q8s16 = vreinterpretq_s16_u16(q8u16);
+            q9s16 = vreinterpretq_s16_u16(q9u16);
+            q10s16 = vreinterpretq_s16_u16(q10u16);
+
+            q7s16 = vqaddq_s16(q7s16, q3s16);
+            q8s16 = vqaddq_s16(q8s16, q4s16);
+            q9s16 = vqaddq_s16(q9s16, q5s16);
+            q10s16 = vqaddq_s16(q10s16, q6s16);
+
+            d6u8 = vqrshrun_n_s16(q7s16, 7);
+            d7u8 = vqrshrun_n_s16(q8s16, 7);
+            d8u8 = vqrshrun_n_s16(q9s16, 7);
+            d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+            d18u8 = d22u8;
+            d19u8 = d23u8;
+            d20u8 = d24u8;
+            d21u8 = d25u8;
+            d22u8 = d26u8;
+            d23u8 = d27u8;
+            d24u8 = d28u8;
+            d25u8 = d29u8;
+            d26u8 = d30u8;
+
+            vst1_u8(dst_ptr, d6u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d7u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d8u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d9u8);
+            dst_ptr += dst_pitch;
+        }
+        return;
+    }
+
+    // load first_pass filter
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[xoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    // First pass: output_height lines x output_width columns (9x4)
+    if (yoffset == 0)  // firstpass_filter4x4_only
+        src = src_ptr - 2;
+    else
+        src = src_ptr - 2 - (src_pixels_per_line * 2);
+
+    tmpp = tmp;
+    for (i = 2; i > 0; i--) {
+        q3u8 = vld1q_u8(src);
+        src += src_pixels_per_line;
+        q4u8 = vld1q_u8(src);
+        src += src_pixels_per_line;
+        q5u8 = vld1q_u8(src);
+        src += src_pixels_per_line;
+        q6u8 = vld1q_u8(src);
+        src += src_pixels_per_line;
+
+        __builtin_prefetch(src);
+        __builtin_prefetch(src + src_pixels_per_line);
+        __builtin_prefetch(src + src_pixels_per_line * 2);
+
+        q7u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
+        q8u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
+        q9u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
+        q10u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
+
+        d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+        d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+        d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
+        d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
+
+        q7u16 = vmlsl_u8(q7u16, d28u8, d1u8);
+        q8u16 = vmlsl_u8(q8u16, d29u8, d1u8);
+        q9u16 = vmlsl_u8(q9u16, d30u8, d1u8);
+        q10u16 = vmlsl_u8(q10u16, d31u8, d1u8);
+
+        d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
+        d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
+        d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
+        d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
+
+        q7u16 = vmlsl_u8(q7u16, d28u8, d4u8);
+        q8u16 = vmlsl_u8(q8u16, d29u8, d4u8);
+        q9u16 = vmlsl_u8(q9u16, d30u8, d4u8);
+        q10u16 = vmlsl_u8(q10u16, d31u8, d4u8);
+
+        d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
+        d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
+        d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
+        d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
+
+        q7u16 = vmlal_u8(q7u16, d28u8, d2u8);
+        q8u16 = vmlal_u8(q8u16, d29u8, d2u8);
+        q9u16 = vmlal_u8(q9u16, d30u8, d2u8);
+        q10u16 = vmlal_u8(q10u16, d31u8, d2u8);
+
+        d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
+        d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
+        d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
+        d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
+
+        q7u16 = vmlal_u8(q7u16, d28u8, d5u8);
+        q8u16 = vmlal_u8(q8u16, d29u8, d5u8);
+        q9u16 = vmlal_u8(q9u16, d30u8, d5u8);
+        q10u16 = vmlal_u8(q10u16, d31u8, d5u8);
+
+        d28u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
+        d29u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
+        d30u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
+        d31u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
+
+        q3u16 = vmull_u8(d28u8, d3u8);
+        q4u16 = vmull_u8(d29u8, d3u8);
+        q5u16 = vmull_u8(d30u8, d3u8);
+        q6u16 = vmull_u8(d31u8, d3u8);
+
+        q3s16 = vreinterpretq_s16_u16(q3u16);
+        q4s16 = vreinterpretq_s16_u16(q4u16);
+        q5s16 = vreinterpretq_s16_u16(q5u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+        q7s16 = vreinterpretq_s16_u16(q7u16);
+        q8s16 = vreinterpretq_s16_u16(q8u16);
+        q9s16 = vreinterpretq_s16_u16(q9u16);
+        q10s16 = vreinterpretq_s16_u16(q10u16);
+
+        q7s16 = vqaddq_s16(q7s16, q3s16);
+        q8s16 = vqaddq_s16(q8s16, q4s16);
+        q9s16 = vqaddq_s16(q9s16, q5s16);
+        q10s16 = vqaddq_s16(q10s16, q6s16);
+
+        d22u8 = vqrshrun_n_s16(q7s16, 7);
+        d23u8 = vqrshrun_n_s16(q8s16, 7);
+        d24u8 = vqrshrun_n_s16(q9s16, 7);
+        d25u8 = vqrshrun_n_s16(q10s16, 7);
+
+        if (yoffset == 0) {  // firstpass_filter8x4_only
+            vst1_u8(dst_ptr, d22u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d23u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d24u8);
+            dst_ptr += dst_pitch;
+            vst1_u8(dst_ptr, d25u8);
+            dst_ptr += dst_pitch;
+        } else {
+            vst1_u8(tmpp, d22u8);
+            tmpp += 8;
+            vst1_u8(tmpp, d23u8);
+            tmpp += 8;
+            vst1_u8(tmpp, d24u8);
+            tmpp += 8;
+            vst1_u8(tmpp, d25u8);
+            tmpp += 8;
+        }
+    }
+    if (yoffset == 0)
+        return;
+
+    // First Pass on rest 5-line data
+    q3u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q4u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q5u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q6u8 = vld1q_u8(src);
+    src += src_pixels_per_line;
+    q7u8 = vld1q_u8(src);
+
+    q8u16 = vmull_u8(vget_low_u8(q3u8), d0u8);
+    q9u16 = vmull_u8(vget_low_u8(q4u8), d0u8);
+    q10u16 = vmull_u8(vget_low_u8(q5u8), d0u8);
+    q11u16 = vmull_u8(vget_low_u8(q6u8), d0u8);
+    q12u16 = vmull_u8(vget_low_u8(q7u8), d0u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 1);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 1);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 1);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 1);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 1);
+
+    q8u16 = vmlsl_u8(q8u16, d27u8, d1u8);
+    q9u16 = vmlsl_u8(q9u16, d28u8, d1u8);
+    q10u16 = vmlsl_u8(q10u16, d29u8, d1u8);
+    q11u16 = vmlsl_u8(q11u16, d30u8, d1u8);
+    q12u16 = vmlsl_u8(q12u16, d31u8, d1u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 4);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 4);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 4);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 4);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 4);
+
+    q8u16 = vmlsl_u8(q8u16, d27u8, d4u8);
+    q9u16 = vmlsl_u8(q9u16, d28u8, d4u8);
+    q10u16 = vmlsl_u8(q10u16, d29u8, d4u8);
+    q11u16 = vmlsl_u8(q11u16, d30u8, d4u8);
+    q12u16 = vmlsl_u8(q12u16, d31u8, d4u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 2);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 2);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 2);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 2);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 2);
+
+    q8u16 = vmlal_u8(q8u16, d27u8, d2u8);
+    q9u16 = vmlal_u8(q9u16, d28u8, d2u8);
+    q10u16 = vmlal_u8(q10u16, d29u8, d2u8);
+    q11u16 = vmlal_u8(q11u16, d30u8, d2u8);
+    q12u16 = vmlal_u8(q12u16, d31u8, d2u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 5);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 5);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 5);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 5);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 5);
+
+    q8u16 = vmlal_u8(q8u16, d27u8, d5u8);
+    q9u16 = vmlal_u8(q9u16, d28u8, d5u8);
+    q10u16 = vmlal_u8(q10u16, d29u8, d5u8);
+    q11u16 = vmlal_u8(q11u16, d30u8, d5u8);
+    q12u16 = vmlal_u8(q12u16, d31u8, d5u8);
+
+    d27u8 = vext_u8(vget_low_u8(q3u8), vget_high_u8(q3u8), 3);
+    d28u8 = vext_u8(vget_low_u8(q4u8), vget_high_u8(q4u8), 3);
+    d29u8 = vext_u8(vget_low_u8(q5u8), vget_high_u8(q5u8), 3);
+    d30u8 = vext_u8(vget_low_u8(q6u8), vget_high_u8(q6u8), 3);
+    d31u8 = vext_u8(vget_low_u8(q7u8), vget_high_u8(q7u8), 3);
+
+    q3u16 = vmull_u8(d27u8, d3u8);
+    q4u16 = vmull_u8(d28u8, d3u8);
+    q5u16 = vmull_u8(d29u8, d3u8);
+    q6u16 = vmull_u8(d30u8, d3u8);
+    q7u16 = vmull_u8(d31u8, d3u8);
+
+    q3s16 = vreinterpretq_s16_u16(q3u16);
+    q4s16 = vreinterpretq_s16_u16(q4u16);
+    q5s16 = vreinterpretq_s16_u16(q5u16);
+    q6s16 = vreinterpretq_s16_u16(q6u16);
+    q7s16 = vreinterpretq_s16_u16(q7u16);
+    q8s16 = vreinterpretq_s16_u16(q8u16);
+    q9s16 = vreinterpretq_s16_u16(q9u16);
+    q10s16 = vreinterpretq_s16_u16(q10u16);
+    q11s16 = vreinterpretq_s16_u16(q11u16);
+    q12s16 = vreinterpretq_s16_u16(q12u16);
+
+    q8s16 = vqaddq_s16(q8s16, q3s16);
+    q9s16 = vqaddq_s16(q9s16, q4s16);
+    q10s16 = vqaddq_s16(q10s16, q5s16);
+    q11s16 = vqaddq_s16(q11s16, q6s16);
+    q12s16 = vqaddq_s16(q12s16, q7s16);
+
+    d26u8 = vqrshrun_n_s16(q8s16, 7);
+    d27u8 = vqrshrun_n_s16(q9s16, 7);
+    d28u8 = vqrshrun_n_s16(q10s16, 7);
+    d29u8 = vqrshrun_n_s16(q11s16, 7);
+    d30u8 = vqrshrun_n_s16(q12s16, 7);
+
+    // Second pass: 8x8
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    tmpp = tmp;
+    q9u8 = vld1q_u8(tmpp);
+    tmpp += 16;
+    q10u8 = vld1q_u8(tmpp);
+    tmpp += 16;
+    q11u8 = vld1q_u8(tmpp);
+    tmpp += 16;
+    q12u8 = vld1q_u8(tmpp);
+
+    d18u8 = vget_low_u8(q9u8);
+    d19u8 = vget_high_u8(q9u8);
+    d20u8 = vget_low_u8(q10u8);
+    d21u8 = vget_high_u8(q10u8);
+    d22u8 = vget_low_u8(q11u8);
+    d23u8 = vget_high_u8(q11u8);
+    d24u8 = vget_low_u8(q12u8);
+    d25u8 = vget_high_u8(q12u8);
+
+    for (i = 2; i > 0; i--) {
+        q3u16 = vmull_u8(d18u8, d0u8);
+        q4u16 = vmull_u8(d19u8, d0u8);
+        q5u16 = vmull_u8(d20u8, d0u8);
+        q6u16 = vmull_u8(d21u8, d0u8);
+
+        q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
+        q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
+        q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
+        q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
+
+        q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
+        q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
+        q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
+        q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
+
+        q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
+        q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
+        q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
+        q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
+
+        q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
+        q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
+        q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
+        q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
+
+        q7u16 = vmull_u8(d21u8, d3u8);
+        q8u16 = vmull_u8(d22u8, d3u8);
+        q9u16 = vmull_u8(d23u8, d3u8);
+        q10u16 = vmull_u8(d24u8, d3u8);
+
+        q3s16 = vreinterpretq_s16_u16(q3u16);
+        q4s16 = vreinterpretq_s16_u16(q4u16);
+        q5s16 = vreinterpretq_s16_u16(q5u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+        q7s16 = vreinterpretq_s16_u16(q7u16);
+        q8s16 = vreinterpretq_s16_u16(q8u16);
+        q9s16 = vreinterpretq_s16_u16(q9u16);
+        q10s16 = vreinterpretq_s16_u16(q10u16);
+
+        q7s16 = vqaddq_s16(q7s16, q3s16);
+        q8s16 = vqaddq_s16(q8s16, q4s16);
+        q9s16 = vqaddq_s16(q9s16, q5s16);
+        q10s16 = vqaddq_s16(q10s16, q6s16);
+
+        d6u8 = vqrshrun_n_s16(q7s16, 7);
+        d7u8 = vqrshrun_n_s16(q8s16, 7);
+        d8u8 = vqrshrun_n_s16(q9s16, 7);
+        d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+        d18u8 = d22u8;
+        d19u8 = d23u8;
+        d20u8 = d24u8;
+        d21u8 = d25u8;
+        d22u8 = d26u8;
+        d23u8 = d27u8;
+        d24u8 = d28u8;
+        d25u8 = d29u8;
+        d26u8 = d30u8;
+
+        vst1_u8(dst_ptr, d6u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d7u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d8u8);
+        dst_ptr += dst_pitch;
+        vst1_u8(dst_ptr, d9u8);
+        dst_ptr += dst_pitch;
+    }
+    return;
+}
+
+void vp8_sixtap_predict16x16_neon(
+        unsigned char *src_ptr,
+        int src_pixels_per_line,
+        int xoffset,
+        int yoffset,
+        unsigned char *dst_ptr,
+        int dst_pitch) {
+    unsigned char *src, *src_tmp, *dst, *tmpp;
+    unsigned char tmp[336];
+    int i, j;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8, d8u8, d9u8;
+    uint8x8_t d10u8, d11u8, d12u8, d13u8, d14u8, d15u8, d18u8, d19u8;
+    uint8x8_t d20u8, d21u8, d22u8, d23u8, d24u8, d25u8, d26u8, d27u8;
+    uint8x8_t d28u8, d29u8, d30u8, d31u8;
+    int8x8_t dtmps8, d0s8, d1s8, d2s8, d3s8, d4s8, d5s8;
+    uint8x16_t q3u8, q4u8;
+    uint16x8_t q3u16, q4u16, q5u16, q6u16, q7u16, q8u16, q9u16, q10u16;
+    uint16x8_t q11u16, q12u16, q13u16, q15u16;
+    int16x8_t q3s16, q4s16, q5s16, q6s16, q7s16, q8s16, q9s16, q10s16;
+    int16x8_t q11s16, q12s16, q13s16, q15s16;
+
+    if (xoffset == 0) {  // secondpass_filter8x8_only
+        // load second_pass filter
+        dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+        d0s8 = vdup_lane_s8(dtmps8, 0);
+        d1s8 = vdup_lane_s8(dtmps8, 1);
+        d2s8 = vdup_lane_s8(dtmps8, 2);
+        d3s8 = vdup_lane_s8(dtmps8, 3);
+        d4s8 = vdup_lane_s8(dtmps8, 4);
+        d5s8 = vdup_lane_s8(dtmps8, 5);
+        d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+        d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+        d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+        d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+        d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+        d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+        // load src data
+        src_tmp = src_ptr - src_pixels_per_line * 2;
+        for (i = 0; i < 2; i++) {
+            src = src_tmp + i * 8;
+            dst = dst_ptr + i * 8;
+            d18u8 = vld1_u8(src);
+            src += src_pixels_per_line;
+            d19u8 = vld1_u8(src);
+            src += src_pixels_per_line;
+            d20u8 = vld1_u8(src);
+            src += src_pixels_per_line;
+            d21u8 = vld1_u8(src);
+            src += src_pixels_per_line;
+            d22u8 = vld1_u8(src);
+            src += src_pixels_per_line;
+            for (j = 0; j < 4; j++) {
+                d23u8 = vld1_u8(src);
+                src += src_pixels_per_line;
+                d24u8 = vld1_u8(src);
+                src += src_pixels_per_line;
+                d25u8 = vld1_u8(src);
+                src += src_pixels_per_line;
+                d26u8 = vld1_u8(src);
+                src += src_pixels_per_line;
+
+                q3u16 = vmull_u8(d18u8, d0u8);
+                q4u16 = vmull_u8(d19u8, d0u8);
+                q5u16 = vmull_u8(d20u8, d0u8);
+                q6u16 = vmull_u8(d21u8, d0u8);
+
+                q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
+                q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
+                q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
+                q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
+
+                q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
+                q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
+                q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
+                q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
+
+                q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
+                q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
+                q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
+                q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
+
+                q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
+                q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
+                q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
+                q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
+
+                q7u16 = vmull_u8(d21u8, d3u8);
+                q8u16 = vmull_u8(d22u8, d3u8);
+                q9u16 = vmull_u8(d23u8, d3u8);
+                q10u16 = vmull_u8(d24u8, d3u8);
+
+                q3s16 = vreinterpretq_s16_u16(q3u16);
+                q4s16 = vreinterpretq_s16_u16(q4u16);
+                q5s16 = vreinterpretq_s16_u16(q5u16);
+                q6s16 = vreinterpretq_s16_u16(q6u16);
+                q7s16 = vreinterpretq_s16_u16(q7u16);
+                q8s16 = vreinterpretq_s16_u16(q8u16);
+                q9s16 = vreinterpretq_s16_u16(q9u16);
+                q10s16 = vreinterpretq_s16_u16(q10u16);
+
+                q7s16 = vqaddq_s16(q7s16, q3s16);
+                q8s16 = vqaddq_s16(q8s16, q4s16);
+                q9s16 = vqaddq_s16(q9s16, q5s16);
+                q10s16 = vqaddq_s16(q10s16, q6s16);
+
+                d6u8 = vqrshrun_n_s16(q7s16, 7);
+                d7u8 = vqrshrun_n_s16(q8s16, 7);
+                d8u8 = vqrshrun_n_s16(q9s16, 7);
+                d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+                d18u8 = d22u8;
+                d19u8 = d23u8;
+                d20u8 = d24u8;
+                d21u8 = d25u8;
+                d22u8 = d26u8;
+
+                vst1_u8(dst, d6u8);
+                dst += dst_pitch;
+                vst1_u8(dst, d7u8);
+                dst += dst_pitch;
+                vst1_u8(dst, d8u8);
+                dst += dst_pitch;
+                vst1_u8(dst, d9u8);
+                dst += dst_pitch;
+            }
+        }
+        return;
+    }
+
+    // load first_pass filter
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[xoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    // First pass: output_height lines x output_width columns (9x4)
+    if (yoffset == 0) {  // firstpass_filter4x4_only
+        src = src_ptr - 2;
+        dst = dst_ptr;
+        for (i = 0; i < 8; i++) {
+            d6u8 = vld1_u8(src);
+            d7u8 = vld1_u8(src + 8);
+            d8u8 = vld1_u8(src + 16);
+            src += src_pixels_per_line;
+            d9u8 = vld1_u8(src);
+            d10u8 = vld1_u8(src + 8);
+            d11u8 = vld1_u8(src + 16);
+            src += src_pixels_per_line;
+
+            __builtin_prefetch(src);
+            __builtin_prefetch(src + src_pixels_per_line);
+
+            q6u16 = vmull_u8(d6u8, d0u8);
+            q7u16 = vmull_u8(d7u8, d0u8);
+            q8u16 = vmull_u8(d9u8, d0u8);
+            q9u16 = vmull_u8(d10u8, d0u8);
+
+            d20u8 = vext_u8(d6u8, d7u8, 1);
+            d21u8 = vext_u8(d9u8, d10u8, 1);
+            d22u8 = vext_u8(d7u8, d8u8, 1);
+            d23u8 = vext_u8(d10u8, d11u8, 1);
+            d24u8 = vext_u8(d6u8, d7u8, 4);
+            d25u8 = vext_u8(d9u8, d10u8, 4);
+            d26u8 = vext_u8(d7u8, d8u8, 4);
+            d27u8 = vext_u8(d10u8, d11u8, 4);
+            d28u8 = vext_u8(d6u8, d7u8, 5);
+            d29u8 = vext_u8(d9u8, d10u8, 5);
+
+            q6u16 = vmlsl_u8(q6u16, d20u8, d1u8);
+            q8u16 = vmlsl_u8(q8u16, d21u8, d1u8);
+            q7u16 = vmlsl_u8(q7u16, d22u8, d1u8);
+            q9u16 = vmlsl_u8(q9u16, d23u8, d1u8);
+            q6u16 = vmlsl_u8(q6u16, d24u8, d4u8);
+            q8u16 = vmlsl_u8(q8u16, d25u8, d4u8);
+            q7u16 = vmlsl_u8(q7u16, d26u8, d4u8);
+            q9u16 = vmlsl_u8(q9u16, d27u8, d4u8);
+            q6u16 = vmlal_u8(q6u16, d28u8, d5u8);
+            q8u16 = vmlal_u8(q8u16, d29u8, d5u8);
+
+            d20u8 = vext_u8(d7u8, d8u8, 5);
+            d21u8 = vext_u8(d10u8, d11u8, 5);
+            d22u8 = vext_u8(d6u8, d7u8, 2);
+            d23u8 = vext_u8(d9u8, d10u8, 2);
+            d24u8 = vext_u8(d7u8, d8u8, 2);
+            d25u8 = vext_u8(d10u8, d11u8, 2);
+            d26u8 = vext_u8(d6u8, d7u8, 3);
+            d27u8 = vext_u8(d9u8, d10u8, 3);
+            d28u8 = vext_u8(d7u8, d8u8, 3);
+            d29u8 = vext_u8(d10u8, d11u8, 3);
+
+            q7u16 = vmlal_u8(q7u16, d20u8, d5u8);
+            q9u16 = vmlal_u8(q9u16, d21u8, d5u8);
+            q6u16 = vmlal_u8(q6u16, d22u8, d2u8);
+            q8u16 = vmlal_u8(q8u16, d23u8, d2u8);
+            q7u16 = vmlal_u8(q7u16, d24u8, d2u8);
+            q9u16 = vmlal_u8(q9u16, d25u8, d2u8);
+
+            q10u16 = vmull_u8(d26u8, d3u8);
+            q11u16 = vmull_u8(d27u8, d3u8);
+            q12u16 = vmull_u8(d28u8, d3u8);
+            q15u16 = vmull_u8(d29u8, d3u8);
+
+            q6s16 = vreinterpretq_s16_u16(q6u16);
+            q7s16 = vreinterpretq_s16_u16(q7u16);
+            q8s16 = vreinterpretq_s16_u16(q8u16);
+            q9s16 = vreinterpretq_s16_u16(q9u16);
+            q10s16 = vreinterpretq_s16_u16(q10u16);
+            q11s16 = vreinterpretq_s16_u16(q11u16);
+            q12s16 = vreinterpretq_s16_u16(q12u16);
+            q15s16 = vreinterpretq_s16_u16(q15u16);
+
+            q6s16 = vqaddq_s16(q6s16, q10s16);
+            q8s16 = vqaddq_s16(q8s16, q11s16);
+            q7s16 = vqaddq_s16(q7s16, q12s16);
+            q9s16 = vqaddq_s16(q9s16, q15s16);
+
+            d6u8 = vqrshrun_n_s16(q6s16, 7);
+            d7u8 = vqrshrun_n_s16(q7s16, 7);
+            d8u8 = vqrshrun_n_s16(q8s16, 7);
+            d9u8 = vqrshrun_n_s16(q9s16, 7);
+
+            q3u8 = vcombine_u8(d6u8, d7u8);
+            q4u8 = vcombine_u8(d8u8, d9u8);
+            vst1q_u8(dst, q3u8);
+            dst += dst_pitch;
+            vst1q_u8(dst, q4u8);
+            dst += dst_pitch;
+        }
+        return;
+    }
+
+    src = src_ptr - 2 - src_pixels_per_line * 2;
+    tmpp = tmp;
+    for (i = 0; i < 7; i++) {
+        d6u8 = vld1_u8(src);
+        d7u8 = vld1_u8(src + 8);
+        d8u8 = vld1_u8(src + 16);
+        src += src_pixels_per_line;
+        d9u8 = vld1_u8(src);
+        d10u8 = vld1_u8(src + 8);
+        d11u8 = vld1_u8(src + 16);
+        src += src_pixels_per_line;
+        d12u8 = vld1_u8(src);
+        d13u8 = vld1_u8(src + 8);
+        d14u8 = vld1_u8(src + 16);
+        src += src_pixels_per_line;
+
+        __builtin_prefetch(src);
+        __builtin_prefetch(src + src_pixels_per_line);
+        __builtin_prefetch(src + src_pixels_per_line * 2);
+
+        q8u16 = vmull_u8(d6u8, d0u8);
+        q9u16 = vmull_u8(d7u8, d0u8);
+        q10u16 = vmull_u8(d9u8, d0u8);
+        q11u16 = vmull_u8(d10u8, d0u8);
+        q12u16 = vmull_u8(d12u8, d0u8);
+        q13u16 = vmull_u8(d13u8, d0u8);
+
+        d28u8 = vext_u8(d6u8, d7u8, 1);
+        d29u8 = vext_u8(d9u8, d10u8, 1);
+        d30u8 = vext_u8(d12u8, d13u8, 1);
+        q8u16 = vmlsl_u8(q8u16, d28u8, d1u8);
+        q10u16 = vmlsl_u8(q10u16, d29u8, d1u8);
+        q12u16 = vmlsl_u8(q12u16, d30u8, d1u8);
+        d28u8 = vext_u8(d7u8, d8u8, 1);
+        d29u8 = vext_u8(d10u8, d11u8, 1);
+        d30u8 = vext_u8(d13u8, d14u8, 1);
+        q9u16  = vmlsl_u8(q9u16, d28u8, d1u8);
+        q11u16 = vmlsl_u8(q11u16, d29u8, d1u8);
+        q13u16 = vmlsl_u8(q13u16, d30u8, d1u8);
+
+        d28u8 = vext_u8(d6u8, d7u8, 4);
+        d29u8 = vext_u8(d9u8, d10u8, 4);
+        d30u8 = vext_u8(d12u8, d13u8, 4);
+        q8u16 = vmlsl_u8(q8u16, d28u8, d4u8);
+        q10u16 = vmlsl_u8(q10u16, d29u8, d4u8);
+        q12u16 = vmlsl_u8(q12u16, d30u8, d4u8);
+        d28u8 = vext_u8(d7u8, d8u8, 4);
+        d29u8 = vext_u8(d10u8, d11u8, 4);
+        d30u8 = vext_u8(d13u8, d14u8, 4);
+        q9u16 = vmlsl_u8(q9u16, d28u8, d4u8);
+        q11u16 = vmlsl_u8(q11u16, d29u8, d4u8);
+        q13u16 = vmlsl_u8(q13u16, d30u8, d4u8);
+
+        d28u8 = vext_u8(d6u8, d7u8, 5);
+        d29u8 = vext_u8(d9u8, d10u8, 5);
+        d30u8 = vext_u8(d12u8, d13u8, 5);
+        q8u16 = vmlal_u8(q8u16, d28u8, d5u8);
+        q10u16 = vmlal_u8(q10u16, d29u8, d5u8);
+        q12u16 = vmlal_u8(q12u16, d30u8, d5u8);
+        d28u8 = vext_u8(d7u8, d8u8, 5);
+        d29u8 = vext_u8(d10u8, d11u8, 5);
+        d30u8 = vext_u8(d13u8, d14u8, 5);
+        q9u16 = vmlal_u8(q9u16, d28u8, d5u8);
+        q11u16 = vmlal_u8(q11u16, d29u8, d5u8);
+        q13u16 = vmlal_u8(q13u16, d30u8, d5u8);
+
+        d28u8 = vext_u8(d6u8, d7u8, 2);
+        d29u8 = vext_u8(d9u8, d10u8, 2);
+        d30u8 = vext_u8(d12u8, d13u8, 2);
+        q8u16 = vmlal_u8(q8u16, d28u8, d2u8);
+        q10u16 = vmlal_u8(q10u16, d29u8, d2u8);
+        q12u16 = vmlal_u8(q12u16, d30u8, d2u8);
+        d28u8 = vext_u8(d7u8, d8u8, 2);
+        d29u8 = vext_u8(d10u8, d11u8, 2);
+        d30u8 = vext_u8(d13u8, d14u8, 2);
+        q9u16 = vmlal_u8(q9u16, d28u8, d2u8);
+        q11u16 = vmlal_u8(q11u16, d29u8, d2u8);
+        q13u16 = vmlal_u8(q13u16, d30u8, d2u8);
+
+        d28u8 = vext_u8(d6u8, d7u8, 3);
+        d29u8 = vext_u8(d9u8, d10u8, 3);
+        d30u8 = vext_u8(d12u8, d13u8, 3);
+        d15u8 = vext_u8(d7u8, d8u8, 3);
+        d31u8 = vext_u8(d10u8, d11u8, 3);
+        d6u8  = vext_u8(d13u8, d14u8, 3);
+        q4u16 = vmull_u8(d28u8, d3u8);
+        q5u16 = vmull_u8(d29u8, d3u8);
+        q6u16 = vmull_u8(d30u8, d3u8);
+        q4s16 = vreinterpretq_s16_u16(q4u16);
+        q5s16 = vreinterpretq_s16_u16(q5u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+        q8s16 = vreinterpretq_s16_u16(q8u16);
+        q10s16 = vreinterpretq_s16_u16(q10u16);
+        q12s16 = vreinterpretq_s16_u16(q12u16);
+        q8s16 = vqaddq_s16(q8s16, q4s16);
+        q10s16 = vqaddq_s16(q10s16, q5s16);
+        q12s16 = vqaddq_s16(q12s16, q6s16);
+
+        q6u16 = vmull_u8(d15u8, d3u8);
+        q7u16 = vmull_u8(d31u8, d3u8);
+        q3u16 = vmull_u8(d6u8, d3u8);
+        q3s16 = vreinterpretq_s16_u16(q3u16);
+        q6s16 = vreinterpretq_s16_u16(q6u16);
+        q7s16 = vreinterpretq_s16_u16(q7u16);
+        q9s16 = vreinterpretq_s16_u16(q9u16);
+        q11s16 = vreinterpretq_s16_u16(q11u16);
+        q13s16 = vreinterpretq_s16_u16(q13u16);
+        q9s16 = vqaddq_s16(q9s16, q6s16);
+        q11s16 = vqaddq_s16(q11s16, q7s16);
+        q13s16 = vqaddq_s16(q13s16, q3s16);
+
+        d6u8 = vqrshrun_n_s16(q8s16, 7);
+        d7u8 = vqrshrun_n_s16(q9s16, 7);
+        d8u8 = vqrshrun_n_s16(q10s16, 7);
+        d9u8 = vqrshrun_n_s16(q11s16, 7);
+        d10u8 = vqrshrun_n_s16(q12s16, 7);
+        d11u8 = vqrshrun_n_s16(q13s16, 7);
+
+        vst1_u8(tmpp, d6u8);
+        tmpp += 8;
+        vst1_u8(tmpp, d7u8);
+        tmpp += 8;
+        vst1_u8(tmpp, d8u8);
+        tmpp += 8;
+        vst1_u8(tmpp, d9u8);
+        tmpp += 8;
+        vst1_u8(tmpp, d10u8);
+        tmpp += 8;
+        vst1_u8(tmpp, d11u8);
+        tmpp += 8;
+    }
+
+    // Second pass: 16x16
+    dtmps8 = vld1_s8(vp8_sub_pel_filters[yoffset]);
+    d0s8 = vdup_lane_s8(dtmps8, 0);
+    d1s8 = vdup_lane_s8(dtmps8, 1);
+    d2s8 = vdup_lane_s8(dtmps8, 2);
+    d3s8 = vdup_lane_s8(dtmps8, 3);
+    d4s8 = vdup_lane_s8(dtmps8, 4);
+    d5s8 = vdup_lane_s8(dtmps8, 5);
+    d0u8 = vreinterpret_u8_s8(vabs_s8(d0s8));
+    d1u8 = vreinterpret_u8_s8(vabs_s8(d1s8));
+    d2u8 = vreinterpret_u8_s8(vabs_s8(d2s8));
+    d3u8 = vreinterpret_u8_s8(vabs_s8(d3s8));
+    d4u8 = vreinterpret_u8_s8(vabs_s8(d4s8));
+    d5u8 = vreinterpret_u8_s8(vabs_s8(d5s8));
+
+    for (i = 0; i < 2; i++) {
+        dst = dst_ptr + 8 * i;
+        tmpp = tmp + 8 * i;
+        d18u8 = vld1_u8(tmpp);
+        tmpp += 16;
+        d19u8 = vld1_u8(tmpp);
+        tmpp += 16;
+        d20u8 = vld1_u8(tmpp);
+        tmpp += 16;
+        d21u8 = vld1_u8(tmpp);
+        tmpp += 16;
+        d22u8 = vld1_u8(tmpp);
+        tmpp += 16;
+        for (j = 0; j < 4; j++) {
+            d23u8 = vld1_u8(tmpp);
+            tmpp += 16;
+            d24u8 = vld1_u8(tmpp);
+            tmpp += 16;
+            d25u8 = vld1_u8(tmpp);
+            tmpp += 16;
+            d26u8 = vld1_u8(tmpp);
+            tmpp += 16;
+
+            q3u16 = vmull_u8(d18u8, d0u8);
+            q4u16 = vmull_u8(d19u8, d0u8);
+            q5u16 = vmull_u8(d20u8, d0u8);
+            q6u16 = vmull_u8(d21u8, d0u8);
+
+            q3u16 = vmlsl_u8(q3u16, d19u8, d1u8);
+            q4u16 = vmlsl_u8(q4u16, d20u8, d1u8);
+            q5u16 = vmlsl_u8(q5u16, d21u8, d1u8);
+            q6u16 = vmlsl_u8(q6u16, d22u8, d1u8);
+
+            q3u16 = vmlsl_u8(q3u16, d22u8, d4u8);
+            q4u16 = vmlsl_u8(q4u16, d23u8, d4u8);
+            q5u16 = vmlsl_u8(q5u16, d24u8, d4u8);
+            q6u16 = vmlsl_u8(q6u16, d25u8, d4u8);
+
+            q3u16 = vmlal_u8(q3u16, d20u8, d2u8);
+            q4u16 = vmlal_u8(q4u16, d21u8, d2u8);
+            q5u16 = vmlal_u8(q5u16, d22u8, d2u8);
+            q6u16 = vmlal_u8(q6u16, d23u8, d2u8);
+
+            q3u16 = vmlal_u8(q3u16, d23u8, d5u8);
+            q4u16 = vmlal_u8(q4u16, d24u8, d5u8);
+            q5u16 = vmlal_u8(q5u16, d25u8, d5u8);
+            q6u16 = vmlal_u8(q6u16, d26u8, d5u8);
+
+            q7u16 = vmull_u8(d21u8, d3u8);
+            q8u16 = vmull_u8(d22u8, d3u8);
+            q9u16 = vmull_u8(d23u8, d3u8);
+            q10u16 = vmull_u8(d24u8, d3u8);
+
+            q3s16 = vreinterpretq_s16_u16(q3u16);
+            q4s16 = vreinterpretq_s16_u16(q4u16);
+            q5s16 = vreinterpretq_s16_u16(q5u16);
+            q6s16 = vreinterpretq_s16_u16(q6u16);
+            q7s16 = vreinterpretq_s16_u16(q7u16);
+            q8s16 = vreinterpretq_s16_u16(q8u16);
+            q9s16 = vreinterpretq_s16_u16(q9u16);
+            q10s16 = vreinterpretq_s16_u16(q10u16);
+
+            q7s16 = vqaddq_s16(q7s16, q3s16);
+            q8s16 = vqaddq_s16(q8s16, q4s16);
+            q9s16 = vqaddq_s16(q9s16, q5s16);
+            q10s16 = vqaddq_s16(q10s16, q6s16);
+
+            d6u8 = vqrshrun_n_s16(q7s16, 7);
+            d7u8 = vqrshrun_n_s16(q8s16, 7);
+            d8u8 = vqrshrun_n_s16(q9s16, 7);
+            d9u8 = vqrshrun_n_s16(q10s16, 7);
+
+            d18u8 = d22u8;
+            d19u8 = d23u8;
+            d20u8 = d24u8;
+            d21u8 = d25u8;
+            d22u8 = d26u8;
+
+            vst1_u8(dst, d6u8);
+            dst += dst_pitch;
+            vst1_u8(dst, d7u8);
+            dst += dst_pitch;
+            vst1_u8(dst, d8u8);
+            dst += dst_pitch;
+            vst1_u8(dst, d9u8);
+            dst += dst_pitch;
+        }
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vp8/common/arm/neon/vp8_loopfilter_neon.c b/libvpx/libvpx/vp8/common/arm/neon/vp8_loopfilter_neon.c
new file mode 100644
index 0000000..9d6807a
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/arm/neon/vp8_loopfilter_neon.c
@@ -0,0 +1,550 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vpx_config.h"
+#include "vpx_ports/arm.h"
+
+static INLINE void vp8_loop_filter_neon(
+        uint8x16_t qblimit,  // flimit
+        uint8x16_t qlimit,   // limit
+        uint8x16_t qthresh,  // thresh
+        uint8x16_t q3,       // p3
+        uint8x16_t q4,       // p2
+        uint8x16_t q5,       // p1
+        uint8x16_t q6,       // p0
+        uint8x16_t q7,       // q0
+        uint8x16_t q8,       // q1
+        uint8x16_t q9,       // q2
+        uint8x16_t q10,      // q3
+        uint8x16_t *q5r,     // p1
+        uint8x16_t *q6r,     // p0
+        uint8x16_t *q7r,     // q0
+        uint8x16_t *q8r) {   // q1
+    uint8x16_t q0u8, q1u8, q2u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+    int16x8_t q2s16, q11s16;
+    uint16x8_t q4u16;
+    int8x16_t q1s8, q2s8, q10s8, q11s8, q12s8, q13s8;
+    int8x8_t d2s8, d3s8;
+
+    q11u8 = vabdq_u8(q3, q4);
+    q12u8 = vabdq_u8(q4, q5);
+    q13u8 = vabdq_u8(q5, q6);
+    q14u8 = vabdq_u8(q8, q7);
+    q3    = vabdq_u8(q9, q8);
+    q4    = vabdq_u8(q10, q9);
+
+    q11u8 = vmaxq_u8(q11u8, q12u8);
+    q12u8 = vmaxq_u8(q13u8, q14u8);
+    q3    = vmaxq_u8(q3, q4);
+    q15u8 = vmaxq_u8(q11u8, q12u8);
+
+    q9 = vabdq_u8(q6, q7);
+
+    // vp8_hevmask
+    q13u8 = vcgtq_u8(q13u8, qthresh);
+    q14u8 = vcgtq_u8(q14u8, qthresh);
+    q15u8 = vmaxq_u8(q15u8, q3);
+
+    q2u8 = vabdq_u8(q5, q8);
+    q9 = vqaddq_u8(q9, q9);
+
+    q15u8 = vcgeq_u8(qlimit, q15u8);
+
+    // vp8_filter() function
+    // convert to signed
+    q10 = vdupq_n_u8(0x80);
+    q8 = veorq_u8(q8, q10);
+    q7 = veorq_u8(q7, q10);
+    q6 = veorq_u8(q6, q10);
+    q5 = veorq_u8(q5, q10);
+
+    q2u8 = vshrq_n_u8(q2u8, 1);
+    q9 = vqaddq_u8(q9, q2u8);
+
+    q10 = vdupq_n_u8(3);
+
+    q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
+                     vget_low_s8(vreinterpretq_s8_u8(q6)));
+    q11s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
+                      vget_high_s8(vreinterpretq_s8_u8(q6)));
+
+    q9 = vcgeq_u8(qblimit, q9);
+
+    q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5),
+                    vreinterpretq_s8_u8(q8));
+
+    q14u8 = vorrq_u8(q13u8, q14u8);
+
+    q4u16 = vmovl_u8(vget_low_u8(q10));
+    q2s16 = vmulq_s16(q2s16, vreinterpretq_s16_u16(q4u16));
+    q11s16 = vmulq_s16(q11s16, vreinterpretq_s16_u16(q4u16));
+
+    q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q14u8);
+    q15u8 = vandq_u8(q15u8, q9);
+
+    q1s8 = vreinterpretq_s8_u8(q1u8);
+    q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
+    q11s16 = vaddw_s8(q11s16, vget_high_s8(q1s8));
+
+    q9 = vdupq_n_u8(4);
+    // vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
+    d2s8 = vqmovn_s16(q2s16);
+    d3s8 = vqmovn_s16(q11s16);
+    q1s8 = vcombine_s8(d2s8, d3s8);
+    q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q15u8);
+    q1s8 = vreinterpretq_s8_u8(q1u8);
+
+    q2s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q10));
+    q1s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q9));
+    q2s8 = vshrq_n_s8(q2s8, 3);
+    q1s8 = vshrq_n_s8(q1s8, 3);
+
+    q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8);
+    q10s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q1s8);
+
+    q1s8 = vrshrq_n_s8(q1s8, 1);
+    q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+    q13s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q1s8);
+    q12s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q1s8);
+
+    q0u8 = vdupq_n_u8(0x80);
+    *q8r = veorq_u8(vreinterpretq_u8_s8(q12s8), q0u8);
+    *q7r = veorq_u8(vreinterpretq_u8_s8(q10s8), q0u8);
+    *q6r = veorq_u8(vreinterpretq_u8_s8(q11s8), q0u8);
+    *q5r = veorq_u8(vreinterpretq_u8_s8(q13s8), q0u8);
+    return;
+}
+
+void vp8_loop_filter_horizontal_edge_y_neon(
+        unsigned char *src,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh) {
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit  = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+    src -= (pitch << 2);
+
+    q3 = vld1q_u8(src);
+    src += pitch;
+    q4 = vld1q_u8(src);
+    src += pitch;
+    q5 = vld1q_u8(src);
+    src += pitch;
+    q6 = vld1q_u8(src);
+    src += pitch;
+    q7 = vld1q_u8(src);
+    src += pitch;
+    q8 = vld1q_u8(src);
+    src += pitch;
+    q9 = vld1q_u8(src);
+    src += pitch;
+    q10 = vld1q_u8(src);
+
+    vp8_loop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q5, &q6, &q7, &q8);
+
+    src -= (pitch * 5);
+    vst1q_u8(src, q5);
+    src += pitch;
+    vst1q_u8(src, q6);
+    src += pitch;
+    vst1q_u8(src, q7);
+    src += pitch;
+    vst1q_u8(src, q8);
+    return;
+}
+
+void vp8_loop_filter_horizontal_edge_uv_neon(
+        unsigned char *u,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh,
+        unsigned char *v) {
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+    uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+    uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit  = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    u -= (pitch << 2);
+    v -= (pitch << 2);
+
+    d6  = vld1_u8(u);
+    u += pitch;
+    d7  = vld1_u8(v);
+    v += pitch;
+    d8  = vld1_u8(u);
+    u += pitch;
+    d9  = vld1_u8(v);
+    v += pitch;
+    d10 = vld1_u8(u);
+    u += pitch;
+    d11 = vld1_u8(v);
+    v += pitch;
+    d12 = vld1_u8(u);
+    u += pitch;
+    d13 = vld1_u8(v);
+    v += pitch;
+    d14 = vld1_u8(u);
+    u += pitch;
+    d15 = vld1_u8(v);
+    v += pitch;
+    d16 = vld1_u8(u);
+    u += pitch;
+    d17 = vld1_u8(v);
+    v += pitch;
+    d18 = vld1_u8(u);
+    u += pitch;
+    d19 = vld1_u8(v);
+    v += pitch;
+    d20 = vld1_u8(u);
+    d21 = vld1_u8(v);
+
+    q3 = vcombine_u8(d6, d7);
+    q4 = vcombine_u8(d8, d9);
+    q5 = vcombine_u8(d10, d11);
+    q6 = vcombine_u8(d12, d13);
+    q7 = vcombine_u8(d14, d15);
+    q8 = vcombine_u8(d16, d17);
+    q9 = vcombine_u8(d18, d19);
+    q10 = vcombine_u8(d20, d21);
+
+    vp8_loop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q5, &q6, &q7, &q8);
+
+    u -= (pitch * 5);
+    vst1_u8(u, vget_low_u8(q5));
+    u += pitch;
+    vst1_u8(u, vget_low_u8(q6));
+    u += pitch;
+    vst1_u8(u, vget_low_u8(q7));
+    u += pitch;
+    vst1_u8(u, vget_low_u8(q8));
+
+    v -= (pitch * 5);
+    vst1_u8(v, vget_high_u8(q5));
+    v += pitch;
+    vst1_u8(v, vget_high_u8(q6));
+    v += pitch;
+    vst1_u8(v, vget_high_u8(q7));
+    v += pitch;
+    vst1_u8(v, vget_high_u8(q8));
+    return;
+}
+
+static INLINE void write_4x8(unsigned char *dst, int pitch,
+                             const uint8x8x4_t result) {
+#ifdef VPX_INCOMPATIBLE_GCC
+    /*
+     * uint8x8x4_t result
+    00 01 02 03 | 04 05 06 07
+    10 11 12 13 | 14 15 16 17
+    20 21 22 23 | 24 25 26 27
+    30 31 32 33 | 34 35 36 37
+    ---
+    * after vtrn_u16
+    00 01 20 21 | 04 05 24 25
+    02 03 22 23 | 06 07 26 27
+    10 11 30 31 | 14 15 34 35
+    12 13 32 33 | 16 17 36 37
+    ---
+    * after vtrn_u8
+    00 10 20 30 | 04 14 24 34
+    01 11 21 31 | 05 15 25 35
+    02 12 22 32 | 06 16 26 36
+    03 13 23 33 | 07 17 27 37
+    */
+    const uint16x4x2_t r02_u16 = vtrn_u16(vreinterpret_u16_u8(result.val[0]),
+                                          vreinterpret_u16_u8(result.val[2]));
+    const uint16x4x2_t r13_u16 = vtrn_u16(vreinterpret_u16_u8(result.val[1]),
+                                          vreinterpret_u16_u8(result.val[3]));
+    const uint8x8x2_t r01_u8 = vtrn_u8(vreinterpret_u8_u16(r02_u16.val[0]),
+                                       vreinterpret_u8_u16(r13_u16.val[0]));
+    const uint8x8x2_t r23_u8 = vtrn_u8(vreinterpret_u8_u16(r02_u16.val[1]),
+                                       vreinterpret_u8_u16(r13_u16.val[1]));
+    const uint32x2_t x_0_4 = vreinterpret_u32_u8(r01_u8.val[0]);
+    const uint32x2_t x_1_5 = vreinterpret_u32_u8(r01_u8.val[1]);
+    const uint32x2_t x_2_6 = vreinterpret_u32_u8(r23_u8.val[0]);
+    const uint32x2_t x_3_7 = vreinterpret_u32_u8(r23_u8.val[1]);
+    vst1_lane_u32((uint32_t *)dst, x_0_4, 0);
+    dst += pitch;
+    vst1_lane_u32((uint32_t *)dst, x_1_5, 0);
+    dst += pitch;
+    vst1_lane_u32((uint32_t *)dst, x_2_6, 0);
+    dst += pitch;
+    vst1_lane_u32((uint32_t *)dst, x_3_7, 0);
+    dst += pitch;
+    vst1_lane_u32((uint32_t *)dst, x_0_4, 1);
+    dst += pitch;
+    vst1_lane_u32((uint32_t *)dst, x_1_5, 1);
+    dst += pitch;
+    vst1_lane_u32((uint32_t *)dst, x_2_6, 1);
+    dst += pitch;
+    vst1_lane_u32((uint32_t *)dst, x_3_7, 1);
+#else
+    vst4_lane_u8(dst, result, 0);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 1);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 2);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 3);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 4);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 5);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 6);
+    dst += pitch;
+    vst4_lane_u8(dst, result, 7);
+#endif  // VPX_INCOMPATIBLE_GCC
+}
+
+void vp8_loop_filter_vertical_edge_y_neon(
+        unsigned char *src,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh) {
+    unsigned char *s, *d;
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+    uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+    uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+    uint32x4x2_t q2tmp0, q2tmp1, q2tmp2, q2tmp3;
+    uint16x8x2_t q2tmp4, q2tmp5, q2tmp6, q2tmp7;
+    uint8x16x2_t q2tmp8, q2tmp9, q2tmp10, q2tmp11;
+    uint8x8x4_t q4ResultH, q4ResultL;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit  = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    s = src - 4;
+    d6  = vld1_u8(s);
+    s += pitch;
+    d8  = vld1_u8(s);
+    s += pitch;
+    d10 = vld1_u8(s);
+    s += pitch;
+    d12 = vld1_u8(s);
+    s += pitch;
+    d14 = vld1_u8(s);
+    s += pitch;
+    d16 = vld1_u8(s);
+    s += pitch;
+    d18 = vld1_u8(s);
+    s += pitch;
+    d20 = vld1_u8(s);
+    s += pitch;
+    d7  = vld1_u8(s);
+    s += pitch;
+    d9  = vld1_u8(s);
+    s += pitch;
+    d11 = vld1_u8(s);
+    s += pitch;
+    d13 = vld1_u8(s);
+    s += pitch;
+    d15 = vld1_u8(s);
+    s += pitch;
+    d17 = vld1_u8(s);
+    s += pitch;
+    d19 = vld1_u8(s);
+    s += pitch;
+    d21 = vld1_u8(s);
+
+    q3 = vcombine_u8(d6, d7);
+    q4 = vcombine_u8(d8, d9);
+    q5 = vcombine_u8(d10, d11);
+    q6 = vcombine_u8(d12, d13);
+    q7 = vcombine_u8(d14, d15);
+    q8 = vcombine_u8(d16, d17);
+    q9 = vcombine_u8(d18, d19);
+    q10 = vcombine_u8(d20, d21);
+
+    q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+    q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+    q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+    q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+    q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+                       vreinterpretq_u16_u32(q2tmp2.val[0]));
+    q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+                       vreinterpretq_u16_u32(q2tmp3.val[0]));
+    q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+                       vreinterpretq_u16_u32(q2tmp2.val[1]));
+    q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+                       vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+    q2tmp8  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+                       vreinterpretq_u8_u16(q2tmp5.val[0]));
+    q2tmp9  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+                       vreinterpretq_u8_u16(q2tmp5.val[1]));
+    q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+                       vreinterpretq_u8_u16(q2tmp7.val[0]));
+    q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+                       vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+    q3 = q2tmp8.val[0];
+    q4 = q2tmp8.val[1];
+    q5 = q2tmp9.val[0];
+    q6 = q2tmp9.val[1];
+    q7 = q2tmp10.val[0];
+    q8 = q2tmp10.val[1];
+    q9 = q2tmp11.val[0];
+    q10 = q2tmp11.val[1];
+
+    vp8_loop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q5, &q6, &q7, &q8);
+
+    q4ResultL.val[0] = vget_low_u8(q5);   // d10
+    q4ResultL.val[1] = vget_low_u8(q6);   // d12
+    q4ResultL.val[2] = vget_low_u8(q7);   // d14
+    q4ResultL.val[3] = vget_low_u8(q8);   // d16
+    q4ResultH.val[0] = vget_high_u8(q5);  // d11
+    q4ResultH.val[1] = vget_high_u8(q6);  // d13
+    q4ResultH.val[2] = vget_high_u8(q7);  // d15
+    q4ResultH.val[3] = vget_high_u8(q8);  // d17
+
+    d = src - 2;
+    write_4x8(d, pitch, q4ResultL);
+    d += pitch * 8;
+    write_4x8(d, pitch, q4ResultH);
+}
+
+void vp8_loop_filter_vertical_edge_uv_neon(
+        unsigned char *u,
+        int pitch,
+        unsigned char blimit,
+        unsigned char limit,
+        unsigned char thresh,
+        unsigned char *v) {
+    unsigned char *us, *ud;
+    unsigned char *vs, *vd;
+    uint8x16_t qblimit, qlimit, qthresh, q3, q4;
+    uint8x16_t q5, q6, q7, q8, q9, q10;
+    uint8x8_t d6, d7, d8, d9, d10, d11, d12, d13, d14;
+    uint8x8_t d15, d16, d17, d18, d19, d20, d21;
+    uint32x4x2_t q2tmp0, q2tmp1, q2tmp2, q2tmp3;
+    uint16x8x2_t q2tmp4, q2tmp5, q2tmp6, q2tmp7;
+    uint8x16x2_t q2tmp8, q2tmp9, q2tmp10, q2tmp11;
+    uint8x8x4_t q4ResultH, q4ResultL;
+
+    qblimit = vdupq_n_u8(blimit);
+    qlimit  = vdupq_n_u8(limit);
+    qthresh = vdupq_n_u8(thresh);
+
+    us = u - 4;
+    d6 = vld1_u8(us);
+    us += pitch;
+    d8 = vld1_u8(us);
+    us += pitch;
+    d10 = vld1_u8(us);
+    us += pitch;
+    d12 = vld1_u8(us);
+    us += pitch;
+    d14 = vld1_u8(us);
+    us += pitch;
+    d16 = vld1_u8(us);
+    us += pitch;
+    d18 = vld1_u8(us);
+    us += pitch;
+    d20 = vld1_u8(us);
+
+    vs = v - 4;
+    d7 = vld1_u8(vs);
+    vs += pitch;
+    d9 = vld1_u8(vs);
+    vs += pitch;
+    d11 = vld1_u8(vs);
+    vs += pitch;
+    d13 = vld1_u8(vs);
+    vs += pitch;
+    d15 = vld1_u8(vs);
+    vs += pitch;
+    d17 = vld1_u8(vs);
+    vs += pitch;
+    d19 = vld1_u8(vs);
+    vs += pitch;
+    d21 = vld1_u8(vs);
+
+    q3 = vcombine_u8(d6, d7);
+    q4 = vcombine_u8(d8, d9);
+    q5 = vcombine_u8(d10, d11);
+    q6 = vcombine_u8(d12, d13);
+    q7 = vcombine_u8(d14, d15);
+    q8 = vcombine_u8(d16, d17);
+    q9 = vcombine_u8(d18, d19);
+    q10 = vcombine_u8(d20, d21);
+
+    q2tmp0 = vtrnq_u32(vreinterpretq_u32_u8(q3), vreinterpretq_u32_u8(q7));
+    q2tmp1 = vtrnq_u32(vreinterpretq_u32_u8(q4), vreinterpretq_u32_u8(q8));
+    q2tmp2 = vtrnq_u32(vreinterpretq_u32_u8(q5), vreinterpretq_u32_u8(q9));
+    q2tmp3 = vtrnq_u32(vreinterpretq_u32_u8(q6), vreinterpretq_u32_u8(q10));
+
+    q2tmp4 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[0]),
+                       vreinterpretq_u16_u32(q2tmp2.val[0]));
+    q2tmp5 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[0]),
+                       vreinterpretq_u16_u32(q2tmp3.val[0]));
+    q2tmp6 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp0.val[1]),
+                       vreinterpretq_u16_u32(q2tmp2.val[1]));
+    q2tmp7 = vtrnq_u16(vreinterpretq_u16_u32(q2tmp1.val[1]),
+                       vreinterpretq_u16_u32(q2tmp3.val[1]));
+
+    q2tmp8  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[0]),
+                       vreinterpretq_u8_u16(q2tmp5.val[0]));
+    q2tmp9  = vtrnq_u8(vreinterpretq_u8_u16(q2tmp4.val[1]),
+                       vreinterpretq_u8_u16(q2tmp5.val[1]));
+    q2tmp10 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[0]),
+                       vreinterpretq_u8_u16(q2tmp7.val[0]));
+    q2tmp11 = vtrnq_u8(vreinterpretq_u8_u16(q2tmp6.val[1]),
+                       vreinterpretq_u8_u16(q2tmp7.val[1]));
+
+    q3 = q2tmp8.val[0];
+    q4 = q2tmp8.val[1];
+    q5 = q2tmp9.val[0];
+    q6 = q2tmp9.val[1];
+    q7 = q2tmp10.val[0];
+    q8 = q2tmp10.val[1];
+    q9 = q2tmp11.val[0];
+    q10 = q2tmp11.val[1];
+
+    vp8_loop_filter_neon(qblimit, qlimit, qthresh, q3, q4,
+                         q5, q6, q7, q8, q9, q10,
+                         &q5, &q6, &q7, &q8);
+
+    q4ResultL.val[0] = vget_low_u8(q5);   // d10
+    q4ResultL.val[1] = vget_low_u8(q6);   // d12
+    q4ResultL.val[2] = vget_low_u8(q7);   // d14
+    q4ResultL.val[3] = vget_low_u8(q8);   // d16
+    ud = u - 2;
+    write_4x8(ud, pitch, q4ResultL);
+
+    q4ResultH.val[0] = vget_high_u8(q5);  // d11
+    q4ResultH.val[1] = vget_high_u8(q6);  // d13
+    q4ResultH.val[2] = vget_high_u8(q7);  // d15
+    q4ResultH.val[3] = vget_high_u8(q8);  // d17
+    vd = v - 2;
+    write_4x8(vd, pitch, q4ResultH);
+}
diff --git a/libvpx/libvpx/vp8/common/blockd.c b/libvpx/libvpx/vp8/common/blockd.c
new file mode 100644
index 0000000..1fc3cd0
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/blockd.c
@@ -0,0 +1,22 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "blockd.h"
+#include "vpx_mem/vpx_mem.h"
+
+const unsigned char vp8_block2left[25] =
+{
+    0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
+};
+const unsigned char vp8_block2above[25] =
+{
+    0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 4, 5, 4, 5, 6, 7, 6, 7, 8
+};
diff --git a/libvpx/libvpx/vp8/common/blockd.h b/libvpx/libvpx/vp8/common/blockd.h
new file mode 100644
index 0000000..192108a
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/blockd.h
@@ -0,0 +1,312 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_BLOCKD_H_
+#define VP8_COMMON_BLOCKD_H_
+
+void vpx_log(const char *format, ...);
+
+#include "vpx_config.h"
+#include "vpx_scale/yv12config.h"
+#include "mv.h"
+#include "treecoder.h"
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*#define DCPRED 1*/
+#define DCPREDSIMTHRESH 0
+#define DCPREDCNTTHRESH 3
+
+#define MB_FEATURE_TREE_PROBS   3
+#define MAX_MB_SEGMENTS         4
+
+#define MAX_REF_LF_DELTAS       4
+#define MAX_MODE_LF_DELTAS      4
+
+/* Segment Feature Masks */
+#define SEGMENT_DELTADATA   0
+#define SEGMENT_ABSDATA     1
+
+typedef struct
+{
+    int r, c;
+} POS;
+
+#define PLANE_TYPE_Y_NO_DC    0
+#define PLANE_TYPE_Y2         1
+#define PLANE_TYPE_UV         2
+#define PLANE_TYPE_Y_WITH_DC  3
+
+
+typedef char ENTROPY_CONTEXT;
+typedef struct
+{
+    ENTROPY_CONTEXT y1[4];
+    ENTROPY_CONTEXT u[2];
+    ENTROPY_CONTEXT v[2];
+    ENTROPY_CONTEXT y2;
+} ENTROPY_CONTEXT_PLANES;
+
+extern const unsigned char vp8_block2left[25];
+extern const unsigned char vp8_block2above[25];
+
+#define VP8_COMBINEENTROPYCONTEXTS( Dest, A, B) \
+    Dest = (A)+(B);
+
+
+typedef enum
+{
+    KEY_FRAME = 0,
+    INTER_FRAME = 1
+} FRAME_TYPE;
+
+typedef enum
+{
+    DC_PRED,            /* average of above and left pixels */
+    V_PRED,             /* vertical prediction */
+    H_PRED,             /* horizontal prediction */
+    TM_PRED,            /* Truemotion prediction */
+    B_PRED,             /* block based prediction, each block has its own prediction mode */
+
+    NEARESTMV,
+    NEARMV,
+    ZEROMV,
+    NEWMV,
+    SPLITMV,
+
+    MB_MODE_COUNT
+} MB_PREDICTION_MODE;
+
+/* Macroblock level features */
+typedef enum
+{
+    MB_LVL_ALT_Q = 0,               /* Use alternate Quantizer .... */
+    MB_LVL_ALT_LF = 1,              /* Use alternate loop filter value... */
+    MB_LVL_MAX = 2                  /* Number of MB level features supported */
+
+} MB_LVL_FEATURES;
+
+/* Segment Feature Masks */
+#define SEGMENT_ALTQ    0x01
+#define SEGMENT_ALT_LF  0x02
+
+#define VP8_YMODES  (B_PRED + 1)
+#define VP8_UV_MODES (TM_PRED + 1)
+
+#define VP8_MVREFS (1 + SPLITMV - NEARESTMV)
+
+typedef enum
+{
+    B_DC_PRED,          /* average of above and left pixels */
+    B_TM_PRED,
+
+    B_VE_PRED,           /* vertical prediction */
+    B_HE_PRED,           /* horizontal prediction */
+
+    B_LD_PRED,
+    B_RD_PRED,
+
+    B_VR_PRED,
+    B_VL_PRED,
+    B_HD_PRED,
+    B_HU_PRED,
+
+    LEFT4X4,
+    ABOVE4X4,
+    ZERO4X4,
+    NEW4X4,
+
+    B_MODE_COUNT
+} B_PREDICTION_MODE;
+
+#define VP8_BINTRAMODES (B_HU_PRED + 1)  /* 10 */
+#define VP8_SUBMVREFS (1 + NEW4X4 - LEFT4X4)
+
+/* For keyframes, intra block modes are predicted by the (already decoded)
+   modes for the Y blocks to the left and above us; for interframes, there
+   is a single probability table. */
+
+union b_mode_info
+{
+    B_PREDICTION_MODE as_mode;
+    int_mv mv;
+};
+
+typedef enum
+{
+    INTRA_FRAME = 0,
+    LAST_FRAME = 1,
+    GOLDEN_FRAME = 2,
+    ALTREF_FRAME = 3,
+    MAX_REF_FRAMES = 4
+} MV_REFERENCE_FRAME;
+
+typedef struct
+{
+    uint8_t mode, uv_mode;
+    uint8_t ref_frame;
+    uint8_t is_4x4;
+    int_mv mv;
+
+    uint8_t partitioning;
+    uint8_t mb_skip_coeff;                                /* does this mb has coefficients at all, 1=no coefficients, 0=need decode tokens */
+    uint8_t need_to_clamp_mvs;
+    uint8_t segment_id;                  /* Which set of segmentation parameters should be used for this MB */
+} MB_MODE_INFO;
+
+typedef struct modeinfo
+{
+    MB_MODE_INFO mbmi;
+    union b_mode_info bmi[16];
+} MODE_INFO;
+
+#if CONFIG_MULTI_RES_ENCODING
+/* The mb-level information needed to be stored for higher-resolution encoder */
+typedef struct
+{
+    MB_PREDICTION_MODE mode;
+    MV_REFERENCE_FRAME ref_frame;
+    int_mv mv;
+    int dissim;    /* dissimilarity level of the macroblock */
+} LOWER_RES_MB_INFO;
+
+/* The frame-level information needed to be stored for higher-resolution
+ *  encoder */
+typedef struct
+{
+    FRAME_TYPE frame_type;
+    int is_frame_dropped;
+    // The frame rate for the lowest resolution.
+    double low_res_framerate;
+    /* The frame number of each reference frames */
+    unsigned int low_res_ref_frames[MAX_REF_FRAMES];
+    // The video frame counter value for the key frame, for lowest resolution.
+    unsigned int key_frame_counter_value;
+    LOWER_RES_MB_INFO *mb_info;
+} LOWER_RES_FRAME_INFO;
+#endif
+
+typedef struct blockd
+{
+    short *qcoeff;
+    short *dqcoeff;
+    unsigned char  *predictor;
+    short *dequant;
+
+    int offset;
+    char *eob;
+
+    union b_mode_info bmi;
+} BLOCKD;
+
+typedef void (*vp8_subpix_fn_t)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+typedef struct macroblockd
+{
+    DECLARE_ALIGNED(16, unsigned char,  predictor[384]);
+    DECLARE_ALIGNED(16, short, qcoeff[400]);
+    DECLARE_ALIGNED(16, short, dqcoeff[400]);
+    DECLARE_ALIGNED(16, char,  eobs[25]);
+
+    DECLARE_ALIGNED(16, short,  dequant_y1[16]);
+    DECLARE_ALIGNED(16, short,  dequant_y1_dc[16]);
+    DECLARE_ALIGNED(16, short,  dequant_y2[16]);
+    DECLARE_ALIGNED(16, short,  dequant_uv[16]);
+
+    /* 16 Y blocks, 4 U, 4 V, 1 DC 2nd order block, each with 16 entries. */
+    BLOCKD block[25];
+    int fullpixel_mask;
+
+    YV12_BUFFER_CONFIG pre; /* Filtered copy of previous frame reconstruction */
+    YV12_BUFFER_CONFIG dst;
+
+    MODE_INFO *mode_info_context;
+    int mode_info_stride;
+
+    FRAME_TYPE frame_type;
+
+    int up_available;
+    int left_available;
+
+    unsigned char *recon_above[3];
+    unsigned char *recon_left[3];
+    int recon_left_stride[2];
+
+    /* Y,U,V,Y2 */
+    ENTROPY_CONTEXT_PLANES *above_context;
+    ENTROPY_CONTEXT_PLANES *left_context;
+
+    /* 0 indicates segmentation at MB level is not enabled. Otherwise the individual bits indicate which features are active. */
+    unsigned char segmentation_enabled;
+
+    /* 0 (do not update) 1 (update) the macroblock segmentation map. */
+    unsigned char update_mb_segmentation_map;
+
+    /* 0 (do not update) 1 (update) the macroblock segmentation feature data. */
+    unsigned char update_mb_segmentation_data;
+
+    /* 0 (do not update) 1 (update) the macroblock segmentation feature data. */
+    unsigned char mb_segement_abs_delta;
+
+    /* Per frame flags that define which MB level features (such as quantizer or loop filter level) */
+    /* are enabled and when enabled the proabilities used to decode the per MB flags in MB_MODE_INFO */
+    vp8_prob mb_segment_tree_probs[MB_FEATURE_TREE_PROBS];         /* Probability Tree used to code Segment number */
+
+    signed char segment_feature_data[MB_LVL_MAX][MAX_MB_SEGMENTS];            /* Segment parameters */
+
+    /* mode_based Loop filter adjustment */
+    unsigned char mode_ref_lf_delta_enabled;
+    unsigned char mode_ref_lf_delta_update;
+
+    /* Delta values have the range +/- MAX_LOOP_FILTER */
+    signed char last_ref_lf_deltas[MAX_REF_LF_DELTAS];                /* 0 = Intra, Last, GF, ARF */
+    signed char ref_lf_deltas[MAX_REF_LF_DELTAS];                     /* 0 = Intra, Last, GF, ARF */
+    signed char last_mode_lf_deltas[MAX_MODE_LF_DELTAS];                      /* 0 = BPRED, ZERO_MV, MV, SPLIT */
+    signed char mode_lf_deltas[MAX_MODE_LF_DELTAS];                           /* 0 = BPRED, ZERO_MV, MV, SPLIT */
+
+    /* Distance of MB away from frame edges */
+    int mb_to_left_edge;
+    int mb_to_right_edge;
+    int mb_to_top_edge;
+    int mb_to_bottom_edge;
+
+
+
+    vp8_subpix_fn_t  subpixel_predict;
+    vp8_subpix_fn_t  subpixel_predict8x4;
+    vp8_subpix_fn_t  subpixel_predict8x8;
+    vp8_subpix_fn_t  subpixel_predict16x16;
+
+    void *current_bc;
+
+    int corrupted;
+
+#if ARCH_X86 || ARCH_X86_64
+    /* This is an intermediate buffer currently used in sub-pixel motion search
+     * to keep a copy of the reference area. This buffer can be used for other
+     * purpose.
+     */
+    DECLARE_ALIGNED(32, unsigned char, y_buf[22*32]);
+#endif
+} MACROBLOCKD;
+
+
+extern void vp8_build_block_doffsets(MACROBLOCKD *x);
+extern void vp8_setup_block_dptrs(MACROBLOCKD *x);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_BLOCKD_H_
diff --git a/libvpx/libvpx/vp8/common/coefupdateprobs.h b/libvpx/libvpx/vp8/common/coefupdateprobs.h
new file mode 100644
index 0000000..d96a19e
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/coefupdateprobs.h
@@ -0,0 +1,197 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_COMMON_COEFUPDATEPROBS_H_
+#define VP8_COMMON_COEFUPDATEPROBS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Update probabilities for the nodes in the token entropy tree.
+   Generated file included by entropy.c */
+
+const vp8_prob vp8_coef_update_probs [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES] =
+{
+    {
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {176, 246, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {223, 241, 252, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {249, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 244, 252, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {234, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {253, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 246, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {239, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 248, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {251, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {251, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 253, 255, 254, 255, 255, 255, 255, 255, 255, },
+            {250, 255, 254, 255, 254, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+    },
+    {
+        {
+            {217, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {225, 252, 241, 253, 255, 255, 254, 255, 255, 255, 255, },
+            {234, 250, 241, 250, 253, 255, 253, 254, 255, 255, 255, },
+        },
+        {
+            {255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {223, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {238, 253, 254, 254, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 248, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {249, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 253, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {247, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {252, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {253, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {250, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+    },
+    {
+        {
+            {186, 251, 250, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {234, 251, 244, 254, 255, 255, 255, 255, 255, 255, 255, },
+            {251, 251, 243, 253, 254, 255, 254, 255, 255, 255, 255, },
+        },
+        {
+            {255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {236, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {251, 253, 253, 254, 254, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+    },
+    {
+        {
+            {248, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {250, 254, 252, 254, 255, 255, 255, 255, 255, 255, 255, },
+            {248, 254, 249, 253, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {246, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {252, 254, 251, 254, 254, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 252, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {248, 254, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {253, 255, 254, 254, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 251, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {245, 251, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {253, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 251, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {252, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 252, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {249, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {250, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+    },
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_COEFUPDATEPROBS_H_
diff --git a/libvpx/libvpx/vp8/common/common.h b/libvpx/libvpx/vp8/common/common.h
new file mode 100644
index 0000000..e58a9cc
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/common.h
@@ -0,0 +1,48 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_COMMON_H_
+#define VP8_COMMON_COMMON_H_
+
+#include <assert.h>
+
+/* Interface header for common constant data structures and lookup tables */
+
+#include "vpx_mem/vpx_mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Only need this for fixed-size arrays, for structs just assign. */
+
+#define vp8_copy( Dest, Src) { \
+        assert( sizeof( Dest) == sizeof( Src)); \
+        memcpy( Dest, Src, sizeof( Src)); \
+    }
+
+/* Use this for variably-sized arrays. */
+
+#define vp8_copy_array( Dest, Src, N) { \
+        assert( sizeof( *Dest) == sizeof( *Src)); \
+        memcpy( Dest, Src, N * sizeof( *Src)); \
+    }
+
+#define vp8_zero( Dest)  memset( &Dest, 0, sizeof( Dest));
+
+#define vp8_zero_array( Dest, N)  memset( Dest, 0, N * sizeof( *Dest));
+
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_COMMON_H_
diff --git a/libvpx/libvpx/vp8/common/context.c b/libvpx/libvpx/vp8/common/context.c
new file mode 100644
index 0000000..99e95d3
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/context.c
@@ -0,0 +1,399 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "entropy.h"
+
+/* *** GENERATED FILE: DO NOT EDIT *** */
+
+#if 0
+int Contexts[vp8_coef_counter_dimen];
+
+const int default_contexts[vp8_coef_counter_dimen] =
+{
+    {
+        // Block Type ( 0 )
+        {
+            // Coeff Band ( 0 )
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+        },
+        {
+            // Coeff Band ( 1 )
+            {30190, 26544, 225,  24,   4,   0,   0,   0,   0,   0,   0, 4171593,},
+            {26846, 25157, 1241, 130,  26,   6,   1,   0,   0,   0,   0, 149987,},
+            {10484, 9538, 1006, 160,  36,  18,   0,   0,   0,   0,   0, 15104,},
+        },
+        {
+            // Coeff Band ( 2 )
+            {25842, 40456, 1126,  83,  11,   2,   0,   0,   0,   0,   0,   0,},
+            {9338, 8010, 512,  73,   7,   3,   2,   0,   0,   0,   0, 43294,},
+            {1047, 751, 149,  31,  13,   6,   1,   0,   0,   0,   0, 879,},
+        },
+        {
+            // Coeff Band ( 3 )
+            {26136, 9826, 252,  13,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {8134, 5574, 191,  14,   2,   0,   0,   0,   0,   0,   0, 35302,},
+            { 605, 677, 116,   9,   1,   0,   0,   0,   0,   0,   0, 611,},
+        },
+        {
+            // Coeff Band ( 4 )
+            {10263, 15463, 283,  17,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {2773, 2191, 128,   9,   2,   2,   0,   0,   0,   0,   0, 10073,},
+            { 134, 125,  32,   4,   0,   2,   0,   0,   0,   0,   0,  50,},
+        },
+        {
+            // Coeff Band ( 5 )
+            {10483, 2663,  23,   1,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {2137, 1251,  27,   1,   1,   0,   0,   0,   0,   0,   0, 14362,},
+            { 116, 156,  14,   2,   1,   0,   0,   0,   0,   0,   0, 190,},
+        },
+        {
+            // Coeff Band ( 6 )
+            {40977, 27614, 412,  28,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {6113, 5213, 261,  22,   3,   0,   0,   0,   0,   0,   0, 26164,},
+            { 382, 312,  50,  14,   2,   0,   0,   0,   0,   0,   0, 345,},
+        },
+        {
+            // Coeff Band ( 7 )
+            {   0,  26,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,  13,   0,   0,   0,   0,   0,   0,   0,   0,   0, 319,},
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   8,},
+        },
+    },
+    {
+        // Block Type ( 1 )
+        {
+            // Coeff Band ( 0 )
+            {3268, 19382, 1043, 250,  93,  82,  49,  26,  17,   8,  25, 82289,},
+            {8758, 32110, 5436, 1832, 827, 668, 420, 153,  24,   0,   3, 52914,},
+            {9337, 23725, 8487, 3954, 2107, 1836, 1069, 399,  59,   0,   0, 18620,},
+        },
+        {
+            // Coeff Band ( 1 )
+            {12419, 8420, 452,  62,   9,   1,   0,   0,   0,   0,   0,   0,},
+            {11715, 8705, 693,  92,  15,   7,   2,   0,   0,   0,   0, 53988,},
+            {7603, 8585, 2306, 778, 270, 145,  39,   5,   0,   0,   0, 9136,},
+        },
+        {
+            // Coeff Band ( 2 )
+            {15938, 14335, 1207, 184,  55,  13,   4,   1,   0,   0,   0,   0,},
+            {7415, 6829, 1138, 244,  71,  26,   7,   0,   0,   0,   0, 9980,},
+            {1580, 1824, 655, 241,  89,  46,  10,   2,   0,   0,   0, 429,},
+        },
+        {
+            // Coeff Band ( 3 )
+            {19453, 5260, 201,  19,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {9173, 3758, 213,  22,   1,   1,   0,   0,   0,   0,   0, 9820,},
+            {1689, 1277, 276,  51,  17,   4,   0,   0,   0,   0,   0, 679,},
+        },
+        {
+            // Coeff Band ( 4 )
+            {12076, 10667, 620,  85,  19,   9,   5,   0,   0,   0,   0,   0,},
+            {4665, 3625, 423,  55,  19,   9,   0,   0,   0,   0,   0, 5127,},
+            { 415, 440, 143,  34,  20,   7,   2,   0,   0,   0,   0, 101,},
+        },
+        {
+            // Coeff Band ( 5 )
+            {12183, 4846, 115,  11,   1,   0,   0,   0,   0,   0,   0,   0,},
+            {4226, 3149, 177,  21,   2,   0,   0,   0,   0,   0,   0, 7157,},
+            { 375, 621, 189,  51,  11,   4,   1,   0,   0,   0,   0, 198,},
+        },
+        {
+            // Coeff Band ( 6 )
+            {61658, 37743, 1203,  94,  10,   3,   0,   0,   0,   0,   0,   0,},
+            {15514, 11563, 903, 111,  14,   5,   0,   0,   0,   0,   0, 25195,},
+            { 929, 1077, 291,  78,  14,   7,   1,   0,   0,   0,   0, 507,},
+        },
+        {
+            // Coeff Band ( 7 )
+            {   0, 990,  15,   3,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0, 412,  13,   0,   0,   0,   0,   0,   0,   0,   0, 1641,},
+            {   0,  18,   7,   1,   0,   0,   0,   0,   0,   0,   0,  30,},
+        },
+    },
+    {
+        // Block Type ( 2 )
+        {
+            // Coeff Band ( 0 )
+            { 953, 24519, 628, 120,  28,  12,   4,   0,   0,   0,   0, 2248798,},
+            {1525, 25654, 2647, 617, 239, 143,  42,   5,   0,   0,   0, 66837,},
+            {1180, 11011, 3001, 1237, 532, 448, 239,  54,   5,   0,   0, 7122,},
+        },
+        {
+            // Coeff Band ( 1 )
+            {1356, 2220,  67,  10,   4,   1,   0,   0,   0,   0,   0,   0,},
+            {1450, 2544, 102,  18,   4,   3,   0,   0,   0,   0,   0, 57063,},
+            {1182, 2110, 470, 130,  41,  21,   0,   0,   0,   0,   0, 6047,},
+        },
+        {
+            // Coeff Band ( 2 )
+            { 370, 3378, 200,  30,   5,   4,   1,   0,   0,   0,   0,   0,},
+            { 293, 1006, 131,  29,  11,   0,   0,   0,   0,   0,   0, 5404,},
+            { 114, 387,  98,  23,   4,   8,   1,   0,   0,   0,   0, 236,},
+        },
+        {
+            // Coeff Band ( 3 )
+            { 579, 194,   4,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            { 395, 213,   5,   1,   0,   0,   0,   0,   0,   0,   0, 4157,},
+            { 119, 122,   4,   0,   0,   0,   0,   0,   0,   0,   0, 300,},
+        },
+        {
+            // Coeff Band ( 4 )
+            {  38, 557,  19,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {  21, 114,  12,   1,   0,   0,   0,   0,   0,   0,   0, 427,},
+            {   0,   5,   0,   0,   0,   0,   0,   0,   0,   0,   0,   7,},
+        },
+        {
+            // Coeff Band ( 5 )
+            {  52,   7,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {  18,   6,   0,   0,   0,   0,   0,   0,   0,   0,   0, 652,},
+            {   1,   1,   0,   0,   0,   0,   0,   0,   0,   0,   0,  30,},
+        },
+        {
+            // Coeff Band ( 6 )
+            { 640, 569,  10,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {  25,  77,   2,   0,   0,   0,   0,   0,   0,   0,   0, 517,},
+            {   4,   7,   0,   0,   0,   0,   0,   0,   0,   0,   0,   3,},
+        },
+        {
+            // Coeff Band ( 7 )
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+        },
+    },
+    {
+        // Block Type ( 3 )
+        {
+            // Coeff Band ( 0 )
+            {2506, 20161, 2707, 767, 261, 178, 107,  30,  14,   3,   0, 100694,},
+            {8806, 36478, 8817, 3268, 1280, 850, 401, 114,  42,   0,   0, 58572,},
+            {11003, 27214, 11798, 5716, 2482, 2072, 1048, 175,  32,   0,   0, 19284,},
+        },
+        {
+            // Coeff Band ( 1 )
+            {9738, 11313, 959, 205,  70,  18,  11,   1,   0,   0,   0,   0,},
+            {12628, 15085, 1507, 273,  52,  19,   9,   0,   0,   0,   0, 54280,},
+            {10701, 15846, 5561, 1926, 813, 570, 249,  36,   0,   0,   0, 6460,},
+        },
+        {
+            // Coeff Band ( 2 )
+            {6781, 22539, 2784, 634, 182, 123,  20,   4,   0,   0,   0,   0,},
+            {6263, 11544, 2649, 790, 259, 168,  27,   5,   0,   0,   0, 20539,},
+            {3109, 4075, 2031, 896, 457, 386, 158,  29,   0,   0,   0, 1138,},
+        },
+        {
+            // Coeff Band ( 3 )
+            {11515, 4079, 465,  73,   5,  14,   2,   0,   0,   0,   0,   0,},
+            {9361, 5834, 650,  96,  24,   8,   4,   0,   0,   0,   0, 22181,},
+            {4343, 3974, 1360, 415, 132,  96,  14,   1,   0,   0,   0, 1267,},
+        },
+        {
+            // Coeff Band ( 4 )
+            {4787, 9297, 823, 168,  44,  12,   4,   0,   0,   0,   0,   0,},
+            {3619, 4472, 719, 198,  60,  31,   3,   0,   0,   0,   0, 8401,},
+            {1157, 1175, 483, 182,  88,  31,   8,   0,   0,   0,   0, 268,},
+        },
+        {
+            // Coeff Band ( 5 )
+            {8299, 1226,  32,   5,   1,   0,   0,   0,   0,   0,   0,   0,},
+            {3502, 1568,  57,   4,   1,   1,   0,   0,   0,   0,   0, 9811,},
+            {1055, 1070, 166,  29,   6,   1,   0,   0,   0,   0,   0, 527,},
+        },
+        {
+            // Coeff Band ( 6 )
+            {27414, 27927, 1989, 347,  69,  26,   0,   0,   0,   0,   0,   0,},
+            {5876, 10074, 1574, 341,  91,  24,   4,   0,   0,   0,   0, 21954,},
+            {1571, 2171, 778, 324, 124,  65,  16,   0,   0,   0,   0, 979,},
+        },
+        {
+            // Coeff Band ( 7 )
+            {   0,  29,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,  23,   0,   0,   0,   0,   0,   0,   0,   0,   0, 459,},
+            {   0,   1,   0,   0,   0,   0,   0,   0,   0,   0,   0,  13,},
+        },
+    },
+};
+
+//Update probabilities for the nodes in the token entropy tree.
+const vp8_prob tree_update_probs[vp8_coef_tree_dimen] =
+{
+    {
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {176, 246, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {223, 241, 252, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {249, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 244, 252, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {234, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {253, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 246, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {239, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 248, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {251, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {251, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 253, 255, 254, 255, 255, 255, 255, 255, 255, },
+            {250, 255, 254, 255, 254, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+    },
+    {
+        {
+            {217, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {225, 252, 241, 253, 255, 255, 254, 255, 255, 255, 255, },
+            {234, 250, 241, 250, 253, 255, 253, 254, 255, 255, 255, },
+        },
+        {
+            {255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {223, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {238, 253, 254, 254, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 248, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {249, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 253, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {247, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {252, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {253, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {250, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+    },
+    {
+        {
+            {186, 251, 250, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {234, 251, 244, 254, 255, 255, 255, 255, 255, 255, 255, },
+            {251, 251, 243, 253, 254, 255, 254, 255, 255, 255, 255, },
+        },
+        {
+            {255, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {236, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {251, 253, 253, 254, 254, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+    },
+    {
+        {
+            {248, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {250, 254, 252, 254, 255, 255, 255, 255, 255, 255, 255, },
+            {248, 254, 249, 253, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {246, 253, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {252, 254, 251, 254, 254, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 254, 252, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {248, 254, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {253, 255, 254, 254, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 251, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {245, 251, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {253, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 251, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {252, 253, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 252, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {249, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 254, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 253, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {250, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+        {
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+            {255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, },
+        },
+    },
+};
+#endif
diff --git a/libvpx/libvpx/vp8/common/copy_c.c b/libvpx/libvpx/vp8/common/copy_c.c
new file mode 100644
index 0000000..e339291
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/copy_c.c
@@ -0,0 +1,32 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <string.h>
+
+#include "./vp8_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+/* Copy 2 macroblocks to a buffer */
+void vp8_copy32xn_c(const unsigned char *src_ptr, int src_stride,
+                    unsigned char *dst_ptr, int dst_stride,
+                    int height)
+{
+    int r;
+
+    for (r = 0; r < height; r++)
+    {
+        memcpy(dst_ptr, src_ptr, 32);
+
+        src_ptr += src_stride;
+        dst_ptr += dst_stride;
+
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/debugmodes.c b/libvpx/libvpx/vp8/common/debugmodes.c
new file mode 100644
index 0000000..159fddc
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/debugmodes.c
@@ -0,0 +1,155 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <stdio.h>
+#include "blockd.h"
+
+
+void vp8_print_modes_and_motion_vectors(MODE_INFO *mi, int rows, int cols, int frame)
+{
+
+    int mb_row;
+    int mb_col;
+    int mb_index = 0;
+    FILE *mvs = fopen("mvs.stt", "a");
+
+    /* print out the macroblock Y modes */
+    mb_index = 0;
+    fprintf(mvs, "Mb Modes for Frame %d\n", frame);
+
+    for (mb_row = 0; mb_row < rows; mb_row++)
+    {
+        for (mb_col = 0; mb_col < cols; mb_col++)
+        {
+
+            fprintf(mvs, "%2d ", mi[mb_index].mbmi.mode);
+
+            mb_index++;
+        }
+
+        fprintf(mvs, "\n");
+        mb_index++;
+    }
+
+    fprintf(mvs, "\n");
+
+    mb_index = 0;
+    fprintf(mvs, "Mb mv ref for Frame %d\n", frame);
+
+    for (mb_row = 0; mb_row < rows; mb_row++)
+    {
+        for (mb_col = 0; mb_col < cols; mb_col++)
+        {
+
+            fprintf(mvs, "%2d ", mi[mb_index].mbmi.ref_frame);
+
+            mb_index++;
+        }
+
+        fprintf(mvs, "\n");
+        mb_index++;
+    }
+
+    fprintf(mvs, "\n");
+
+    /* print out the macroblock UV modes */
+    mb_index = 0;
+    fprintf(mvs, "UV Modes for Frame %d\n", frame);
+
+    for (mb_row = 0; mb_row < rows; mb_row++)
+    {
+        for (mb_col = 0; mb_col < cols; mb_col++)
+        {
+
+            fprintf(mvs, "%2d ", mi[mb_index].mbmi.uv_mode);
+
+            mb_index++;
+        }
+
+        mb_index++;
+        fprintf(mvs, "\n");
+    }
+
+    fprintf(mvs, "\n");
+
+    /* print out the block modes */
+    fprintf(mvs, "Mbs for Frame %d\n", frame);
+    {
+        int b_row;
+
+        for (b_row = 0; b_row < 4 * rows; b_row++)
+        {
+            int b_col;
+            int bindex;
+
+            for (b_col = 0; b_col < 4 * cols; b_col++)
+            {
+                mb_index = (b_row >> 2) * (cols + 1) + (b_col >> 2);
+                bindex = (b_row & 3) * 4 + (b_col & 3);
+
+                if (mi[mb_index].mbmi.mode == B_PRED)
+                    fprintf(mvs, "%2d ", mi[mb_index].bmi[bindex].as_mode);
+                else
+                    fprintf(mvs, "xx ");
+
+            }
+
+            fprintf(mvs, "\n");
+        }
+    }
+    fprintf(mvs, "\n");
+
+    /* print out the macroblock mvs */
+    mb_index = 0;
+    fprintf(mvs, "MVs for Frame %d\n", frame);
+
+    for (mb_row = 0; mb_row < rows; mb_row++)
+    {
+        for (mb_col = 0; mb_col < cols; mb_col++)
+        {
+            fprintf(mvs, "%5d:%-5d", mi[mb_index].mbmi.mv.as_mv.row / 2, mi[mb_index].mbmi.mv.as_mv.col / 2);
+
+            mb_index++;
+        }
+
+        mb_index++;
+        fprintf(mvs, "\n");
+    }
+
+    fprintf(mvs, "\n");
+
+
+    /* print out the block modes */
+    fprintf(mvs, "MVs for Frame %d\n", frame);
+    {
+        int b_row;
+
+        for (b_row = 0; b_row < 4 * rows; b_row++)
+        {
+            int b_col;
+            int bindex;
+
+            for (b_col = 0; b_col < 4 * cols; b_col++)
+            {
+                mb_index = (b_row >> 2) * (cols + 1) + (b_col >> 2);
+                bindex = (b_row & 3) * 4 + (b_col & 3);
+                fprintf(mvs, "%3d:%-3d ", mi[mb_index].bmi[bindex].mv.as_mv.row, mi[mb_index].bmi[bindex].mv.as_mv.col);
+
+            }
+
+            fprintf(mvs, "\n");
+        }
+    }
+    fprintf(mvs, "\n");
+
+
+    fclose(mvs);
+}
diff --git a/libvpx/libvpx/vp8/common/default_coef_probs.h b/libvpx/libvpx/vp8/common/default_coef_probs.h
new file mode 100644
index 0000000..4d69e4b
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/default_coef_probs.h
@@ -0,0 +1,200 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+*/
+
+#ifndef VP8_COMMON_DEFAULT_COEF_PROBS_H_
+#define VP8_COMMON_DEFAULT_COEF_PROBS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*Generated file, included by entropy.c*/
+
+
+static const vp8_prob default_coef_probs [BLOCK_TYPES]
+                                         [COEF_BANDS]
+                                         [PREV_COEF_CONTEXTS]
+                                         [ENTROPY_NODES] =
+{
+    { /* Block Type ( 0 ) */
+        { /* Coeff Band ( 0 )*/
+            { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+            { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+            { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 1 )*/
+            { 253, 136, 254, 255, 228, 219, 128, 128, 128, 128, 128 },
+            { 189, 129, 242, 255, 227, 213, 255, 219, 128, 128, 128 },
+            { 106, 126, 227, 252, 214, 209, 255, 255, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 2 )*/
+            {   1,  98, 248, 255, 236, 226, 255, 255, 128, 128, 128 },
+            { 181, 133, 238, 254, 221, 234, 255, 154, 128, 128, 128 },
+            {  78, 134, 202, 247, 198, 180, 255, 219, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 3 )*/
+            {   1, 185, 249, 255, 243, 255, 128, 128, 128, 128, 128 },
+            { 184, 150, 247, 255, 236, 224, 128, 128, 128, 128, 128 },
+            {  77, 110, 216, 255, 236, 230, 128, 128, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 4 )*/
+            {   1, 101, 251, 255, 241, 255, 128, 128, 128, 128, 128 },
+            { 170, 139, 241, 252, 236, 209, 255, 255, 128, 128, 128 },
+            {  37, 116, 196, 243, 228, 255, 255, 255, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 5 )*/
+            {   1, 204, 254, 255, 245, 255, 128, 128, 128, 128, 128 },
+            { 207, 160, 250, 255, 238, 128, 128, 128, 128, 128, 128 },
+            { 102, 103, 231, 255, 211, 171, 128, 128, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 6 )*/
+            {   1, 152, 252, 255, 240, 255, 128, 128, 128, 128, 128 },
+            { 177, 135, 243, 255, 234, 225, 128, 128, 128, 128, 128 },
+            {  80, 129, 211, 255, 194, 224, 128, 128, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 7 )*/
+            {   1,   1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+            { 246,   1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+            { 255, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+        }
+    },
+    { /* Block Type ( 1 ) */
+        { /* Coeff Band ( 0 )*/
+            { 198,  35, 237, 223, 193, 187, 162, 160, 145, 155,  62 },
+            { 131,  45, 198, 221, 172, 176, 220, 157, 252, 221,   1 },
+            {  68,  47, 146, 208, 149, 167, 221, 162, 255, 223, 128 }
+        },
+        { /* Coeff Band ( 1 )*/
+            {   1, 149, 241, 255, 221, 224, 255, 255, 128, 128, 128 },
+            { 184, 141, 234, 253, 222, 220, 255, 199, 128, 128, 128 },
+            {  81,  99, 181, 242, 176, 190, 249, 202, 255, 255, 128 }
+        },
+        { /* Coeff Band ( 2 )*/
+            {   1, 129, 232, 253, 214, 197, 242, 196, 255, 255, 128 },
+            {  99, 121, 210, 250, 201, 198, 255, 202, 128, 128, 128 },
+            {  23,  91, 163, 242, 170, 187, 247, 210, 255, 255, 128 }
+        },
+        { /* Coeff Band ( 3 )*/
+            {   1, 200, 246, 255, 234, 255, 128, 128, 128, 128, 128 },
+            { 109, 178, 241, 255, 231, 245, 255, 255, 128, 128, 128 },
+            {  44, 130, 201, 253, 205, 192, 255, 255, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 4 )*/
+            {   1, 132, 239, 251, 219, 209, 255, 165, 128, 128, 128 },
+            {  94, 136, 225, 251, 218, 190, 255, 255, 128, 128, 128 },
+            {  22, 100, 174, 245, 186, 161, 255, 199, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 5 )*/
+            {   1, 182, 249, 255, 232, 235, 128, 128, 128, 128, 128 },
+            { 124, 143, 241, 255, 227, 234, 128, 128, 128, 128, 128 },
+            {  35,  77, 181, 251, 193, 211, 255, 205, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 6 )*/
+            {   1, 157, 247, 255, 236, 231, 255, 255, 128, 128, 128 },
+            { 121, 141, 235, 255, 225, 227, 255, 255, 128, 128, 128 },
+            {  45,  99, 188, 251, 195, 217, 255, 224, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 7 )*/
+            {   1,   1, 251, 255, 213, 255, 128, 128, 128, 128, 128 },
+            { 203,   1, 248, 255, 255, 128, 128, 128, 128, 128, 128 },
+            { 137,   1, 177, 255, 224, 255, 128, 128, 128, 128, 128 }
+        }
+    },
+    { /* Block Type ( 2 ) */
+        { /* Coeff Band ( 0 )*/
+            { 253,   9, 248, 251, 207, 208, 255, 192, 128, 128, 128 },
+            { 175,  13, 224, 243, 193, 185, 249, 198, 255, 255, 128 },
+            {  73,  17, 171, 221, 161, 179, 236, 167, 255, 234, 128 }
+        },
+        { /* Coeff Band ( 1 )*/
+            {   1,  95, 247, 253, 212, 183, 255, 255, 128, 128, 128 },
+            { 239,  90, 244, 250, 211, 209, 255, 255, 128, 128, 128 },
+            { 155,  77, 195, 248, 188, 195, 255, 255, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 2 )*/
+            {   1,  24, 239, 251, 218, 219, 255, 205, 128, 128, 128 },
+            { 201,  51, 219, 255, 196, 186, 128, 128, 128, 128, 128 },
+            {  69,  46, 190, 239, 201, 218, 255, 228, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 3 )*/
+            {   1, 191, 251, 255, 255, 128, 128, 128, 128, 128, 128 },
+            { 223, 165, 249, 255, 213, 255, 128, 128, 128, 128, 128 },
+            { 141, 124, 248, 255, 255, 128, 128, 128, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 4 )*/
+            {   1,  16, 248, 255, 255, 128, 128, 128, 128, 128, 128 },
+            { 190,  36, 230, 255, 236, 255, 128, 128, 128, 128, 128 },
+            { 149,   1, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 5 )*/
+            {   1, 226, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+            { 247, 192, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+            { 240, 128, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 6 )*/
+            {   1, 134, 252, 255, 255, 128, 128, 128, 128, 128, 128 },
+            { 213,  62, 250, 255, 255, 128, 128, 128, 128, 128, 128 },
+            {  55,  93, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 7 )*/
+            { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+            { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+            { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }
+        }
+    },
+    { /* Block Type ( 3 ) */
+        { /* Coeff Band ( 0 )*/
+            { 202,  24, 213, 235, 186, 191, 220, 160, 240, 175, 255 },
+            { 126,  38, 182, 232, 169, 184, 228, 174, 255, 187, 128 },
+            {  61,  46, 138, 219, 151, 178, 240, 170, 255, 216, 128 }
+        },
+        { /* Coeff Band ( 1 )*/
+            {   1, 112, 230, 250, 199, 191, 247, 159, 255, 255, 128 },
+            { 166, 109, 228, 252, 211, 215, 255, 174, 128, 128, 128 },
+            {  39,  77, 162, 232, 172, 180, 245, 178, 255, 255, 128 }
+        },
+        { /* Coeff Band ( 2 )*/
+            {   1,  52, 220, 246, 198, 199, 249, 220, 255, 255, 128 },
+            { 124,  74, 191, 243, 183, 193, 250, 221, 255, 255, 128 },
+            {  24,  71, 130, 219, 154, 170, 243, 182, 255, 255, 128 }
+        },
+        { /* Coeff Band ( 3 )*/
+            {   1, 182, 225, 249, 219, 240, 255, 224, 128, 128, 128 },
+            { 149, 150, 226, 252, 216, 205, 255, 171, 128, 128, 128 },
+            {  28, 108, 170, 242, 183, 194, 254, 223, 255, 255, 128 }
+        },
+        { /* Coeff Band ( 4 )*/
+            {   1,  81, 230, 252, 204, 203, 255, 192, 128, 128, 128 },
+            { 123, 102, 209, 247, 188, 196, 255, 233, 128, 128, 128 },
+            {  20,  95, 153, 243, 164, 173, 255, 203, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 5 )*/
+            {   1, 222, 248, 255, 216, 213, 128, 128, 128, 128, 128 },
+            { 168, 175, 246, 252, 235, 205, 255, 255, 128, 128, 128 },
+            {  47, 116, 215, 255, 211, 212, 255, 255, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 6 )*/
+            {   1, 121, 236, 253, 212, 214, 255, 255, 128, 128, 128 },
+            { 141,  84, 213, 252, 201, 202, 255, 219, 128, 128, 128 },
+            {  42,  80, 160, 240, 162, 185, 255, 205, 128, 128, 128 }
+        },
+        { /* Coeff Band ( 7 )*/
+            {   1,   1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+            { 244,   1, 255, 128, 128, 128, 128, 128, 128, 128, 128 },
+            { 238,   1, 255, 128, 128, 128, 128, 128, 128, 128, 128 }
+        }
+    }
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_DEFAULT_COEF_PROBS_H_
diff --git a/libvpx/libvpx/vp8/common/dequantize.c b/libvpx/libvpx/vp8/common/dequantize.c
new file mode 100644
index 0000000..f8b04fa
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/dequantize.c
@@ -0,0 +1,43 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vp8/common/blockd.h"
+#include "vpx_mem/vpx_mem.h"
+
+void vp8_dequantize_b_c(BLOCKD *d, short *DQC)
+{
+    int i;
+    short *DQ  = d->dqcoeff;
+    short *Q   = d->qcoeff;
+
+    for (i = 0; i < 16; i++)
+    {
+        DQ[i] = Q[i] * DQC[i];
+    }
+}
+
+void vp8_dequant_idct_add_c(short *input, short *dq,
+                            unsigned char *dest, int stride)
+{
+    int i;
+
+    for (i = 0; i < 16; i++)
+    {
+        input[i] = dq[i] * input[i];
+    }
+
+    vp8_short_idct4x4llm_c(input, dest, stride, dest, stride);
+
+    memset(input, 0, 32);
+
+}
diff --git a/libvpx/libvpx/vp8/common/entropy.c b/libvpx/libvpx/vp8/common/entropy.c
new file mode 100644
index 0000000..c00e565
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/entropy.c
@@ -0,0 +1,188 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "entropy.h"
+#include "blockd.h"
+#include "onyxc_int.h"
+#include "vpx_mem/vpx_mem.h"
+
+#include "coefupdateprobs.h"
+
+DECLARE_ALIGNED(16, const unsigned char, vp8_norm[256]) =
+{
+    0, 7, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4,
+    3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
+    2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+    2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
+};
+
+DECLARE_ALIGNED(16, const unsigned char, vp8_coef_bands[16]) =
+{ 0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7};
+
+DECLARE_ALIGNED(16, const unsigned char,
+                vp8_prev_token_class[MAX_ENTROPY_TOKENS]) =
+{ 0, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 0};
+
+DECLARE_ALIGNED(16, const int, vp8_default_zig_zag1d[16]) =
+{
+    0,  1,  4,  8,
+    5,  2,  3,  6,
+    9, 12, 13, 10,
+    7, 11, 14, 15,
+};
+
+DECLARE_ALIGNED(16, const short, vp8_default_inv_zig_zag[16]) =
+{
+    1,  2,  6,  7,
+    3,  5,  8, 13,
+    4,  9, 12, 14,
+   10, 11, 15, 16
+};
+
+/* vp8_default_zig_zag_mask generated with:
+
+    void vp8_init_scan_order_mask()
+    {
+        int i;
+
+        for (i = 0; i < 16; i++)
+        {
+            vp8_default_zig_zag_mask[vp8_default_zig_zag1d[i]] = 1 << i;
+        }
+
+    }
+*/
+DECLARE_ALIGNED(16, const short, vp8_default_zig_zag_mask[16]) =
+{
+     1,    2,    32,     64,
+     4,   16,   128,   4096,
+     8,  256,  2048,   8192,
+   512, 1024, 16384, -32768
+};
+
+const int vp8_mb_feature_data_bits[MB_LVL_MAX] = {7, 6};
+
+/* Array indices are identical to previously-existing CONTEXT_NODE indices */
+
+const vp8_tree_index vp8_coef_tree[ 22] =     /* corresponding _CONTEXT_NODEs */
+{
+    -DCT_EOB_TOKEN, 2,                             /* 0 = EOB */
+    -ZERO_TOKEN, 4,                               /* 1 = ZERO */
+    -ONE_TOKEN, 6,                               /* 2 = ONE */
+    8, 12,                                      /* 3 = LOW_VAL */
+    -TWO_TOKEN, 10,                            /* 4 = TWO */
+    -THREE_TOKEN, -FOUR_TOKEN,                /* 5 = THREE */
+    14, 16,                                    /* 6 = HIGH_LOW */
+    -DCT_VAL_CATEGORY1, -DCT_VAL_CATEGORY2,   /* 7 = CAT_ONE */
+    18, 20,                                   /* 8 = CAT_THREEFOUR */
+    -DCT_VAL_CATEGORY3, -DCT_VAL_CATEGORY4,  /* 9 = CAT_THREE */
+    -DCT_VAL_CATEGORY5, -DCT_VAL_CATEGORY6   /* 10 = CAT_FIVE */
+};
+
+/* vp8_coef_encodings generated with:
+    vp8_tokens_from_tree(vp8_coef_encodings, vp8_coef_tree);
+*/
+vp8_token vp8_coef_encodings[MAX_ENTROPY_TOKENS] =
+{
+    {2, 2},
+    {6, 3},
+    {28, 5},
+    {58, 6},
+    {59, 6},
+    {60, 6},
+    {61, 6},
+    {124, 7},
+    {125, 7},
+    {126, 7},
+    {127, 7},
+    {0, 1}
+};
+
+/* Trees for extra bits.  Probabilities are constant and
+   do not depend on previously encoded bits */
+
+static const vp8_prob Pcat1[] = { 159};
+static const vp8_prob Pcat2[] = { 165, 145};
+static const vp8_prob Pcat3[] = { 173, 148, 140};
+static const vp8_prob Pcat4[] = { 176, 155, 140, 135};
+static const vp8_prob Pcat5[] = { 180, 157, 141, 134, 130};
+static const vp8_prob Pcat6[] =
+{ 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129};
+
+
+/* tree index tables generated with:
+
+    void init_bit_tree(vp8_tree_index *p, int n)
+    {
+        int i = 0;
+
+        while (++i < n)
+        {
+            p[0] = p[1] = i << 1;
+            p += 2;
+        }
+
+        p[0] = p[1] = 0;
+    }
+
+    void init_bit_trees()
+    {
+        init_bit_tree(cat1, 1);
+        init_bit_tree(cat2, 2);
+        init_bit_tree(cat3, 3);
+        init_bit_tree(cat4, 4);
+        init_bit_tree(cat5, 5);
+        init_bit_tree(cat6, 11);
+    }
+*/
+
+static const vp8_tree_index cat1[2] = { 0, 0 };
+static const vp8_tree_index cat2[4] = { 2, 2, 0, 0 };
+static const vp8_tree_index cat3[6] = { 2, 2, 4, 4, 0, 0 };
+static const vp8_tree_index cat4[8] = { 2, 2, 4, 4, 6, 6, 0, 0 };
+static const vp8_tree_index cat5[10] = { 2, 2, 4, 4, 6, 6, 8, 8, 0, 0 };
+static const vp8_tree_index cat6[22] = { 2, 2, 4, 4, 6, 6, 8, 8, 10, 10, 12, 12,
+                                        14, 14, 16, 16, 18, 18, 20, 20, 0, 0 };
+
+const vp8_extra_bit_struct vp8_extra_bits[12] =
+{
+    { 0, 0, 0, 0},
+    { 0, 0, 0, 1},
+    { 0, 0, 0, 2},
+    { 0, 0, 0, 3},
+    { 0, 0, 0, 4},
+    { cat1, Pcat1, 1, 5},
+    { cat2, Pcat2, 2, 7},
+    { cat3, Pcat3, 3, 11},
+    { cat4, Pcat4, 4, 19},
+    { cat5, Pcat5, 5, 35},
+    { cat6, Pcat6, 11, 67},
+    { 0, 0, 0, 0}
+};
+
+#include "default_coef_probs.h"
+
+void vp8_default_coef_probs(VP8_COMMON *pc)
+{
+    memcpy(pc->fc.coef_probs, default_coef_probs, sizeof(default_coef_probs));
+}
+
diff --git a/libvpx/libvpx/vp8/common/entropy.h b/libvpx/libvpx/vp8/common/entropy.h
new file mode 100644
index 0000000..a90bab4
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/entropy.h
@@ -0,0 +1,109 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_ENTROPY_H_
+#define VP8_COMMON_ENTROPY_H_
+
+#include "treecoder.h"
+#include "blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Coefficient token alphabet */
+
+#define ZERO_TOKEN              0       /* 0         Extra Bits 0+0 */
+#define ONE_TOKEN               1       /* 1         Extra Bits 0+1 */
+#define TWO_TOKEN               2       /* 2         Extra Bits 0+1 */
+#define THREE_TOKEN             3       /* 3         Extra Bits 0+1 */
+#define FOUR_TOKEN              4       /* 4         Extra Bits 0+1 */
+#define DCT_VAL_CATEGORY1       5       /* 5-6       Extra Bits 1+1 */
+#define DCT_VAL_CATEGORY2       6       /* 7-10      Extra Bits 2+1 */
+#define DCT_VAL_CATEGORY3       7       /* 11-18     Extra Bits 3+1 */
+#define DCT_VAL_CATEGORY4       8       /* 19-34     Extra Bits 4+1 */
+#define DCT_VAL_CATEGORY5       9       /* 35-66     Extra Bits 5+1 */
+#define DCT_VAL_CATEGORY6       10      /* 67+       Extra Bits 11+1 */
+#define DCT_EOB_TOKEN           11      /* EOB       Extra Bits 0+0 */
+
+#define MAX_ENTROPY_TOKENS 12
+#define ENTROPY_NODES 11
+
+extern const vp8_tree_index vp8_coef_tree[];
+
+extern const struct vp8_token_struct vp8_coef_encodings[MAX_ENTROPY_TOKENS];
+
+typedef struct
+{
+    vp8_tree_p tree;
+    const vp8_prob *prob;
+    int Len;
+    int base_val;
+} vp8_extra_bit_struct;
+
+extern const vp8_extra_bit_struct vp8_extra_bits[12];    /* indexed by token value */
+
+#define PROB_UPDATE_BASELINE_COST   7
+
+#define MAX_PROB                255
+#define DCT_MAX_VALUE           2048
+
+
+/* Coefficients are predicted via a 3-dimensional probability table. */
+
+/* Outside dimension.  0 = Y no DC, 1 = Y2, 2 = UV, 3 = Y with DC */
+
+#define BLOCK_TYPES 4
+
+/* Middle dimension is a coarsening of the coefficient's
+   position within the 4x4 DCT. */
+
+#define COEF_BANDS 8
+extern DECLARE_ALIGNED(16, const unsigned char, vp8_coef_bands[16]);
+
+/* Inside dimension is 3-valued measure of nearby complexity, that is,
+   the extent to which nearby coefficients are nonzero.  For the first
+   coefficient (DC, unless block type is 0), we look at the (already encoded)
+   blocks above and to the left of the current block.  The context index is
+   then the number (0,1,or 2) of these blocks having nonzero coefficients.
+   After decoding a coefficient, the measure is roughly the size of the
+   most recently decoded coefficient (0 for 0, 1 for 1, 2 for >1).
+   Note that the intuitive meaning of this measure changes as coefficients
+   are decoded, e.g., prior to the first token, a zero means that my neighbors
+   are empty while, after the first token, because of the use of end-of-block,
+   a zero means we just decoded a zero and hence guarantees that a non-zero
+   coefficient will appear later in this block.  However, this shift
+   in meaning is perfectly OK because our context depends also on the
+   coefficient band (and since zigzag positions 0, 1, and 2 are in
+   distinct bands). */
+
+/*# define DC_TOKEN_CONTEXTS        3*/ /* 00, 0!0, !0!0 */
+#   define PREV_COEF_CONTEXTS       3
+
+extern DECLARE_ALIGNED(16, const unsigned char, vp8_prev_token_class[MAX_ENTROPY_TOKENS]);
+
+extern const vp8_prob vp8_coef_update_probs [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES];
+
+
+struct VP8Common;
+void vp8_default_coef_probs(struct VP8Common *);
+
+extern DECLARE_ALIGNED(16, const int, vp8_default_zig_zag1d[16]);
+extern DECLARE_ALIGNED(16, const short, vp8_default_inv_zig_zag[16]);
+extern DECLARE_ALIGNED(16, const short, vp8_default_zig_zag_mask[16]);
+extern const int vp8_mb_feature_data_bits[MB_LVL_MAX];
+
+void vp8_coef_tree_initialize(void);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_ENTROPY_H_
diff --git a/libvpx/libvpx/vp8/common/entropymode.c b/libvpx/libvpx/vp8/common/entropymode.c
new file mode 100644
index 0000000..8981a8d
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/entropymode.c
@@ -0,0 +1,171 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#define USE_PREBUILT_TABLES
+
+#include "entropymode.h"
+#include "entropy.h"
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp8_entropymodedata.h"
+
+int vp8_mv_cont(const int_mv *l, const int_mv *a)
+{
+    int lez = (l->as_int == 0);
+    int aez = (a->as_int == 0);
+    int lea = (l->as_int == a->as_int);
+
+    if (lea && lez)
+        return SUBMVREF_LEFT_ABOVE_ZED;
+
+    if (lea)
+        return SUBMVREF_LEFT_ABOVE_SAME;
+
+    if (aez)
+        return SUBMVREF_ABOVE_ZED;
+
+    if (lez)
+        return SUBMVREF_LEFT_ZED;
+
+    return SUBMVREF_NORMAL;
+}
+
+static const vp8_prob sub_mv_ref_prob [VP8_SUBMVREFS-1] = { 180, 162, 25};
+
+const vp8_prob vp8_sub_mv_ref_prob2 [SUBMVREF_COUNT][VP8_SUBMVREFS-1] =
+{
+    { 147, 136, 18 },
+    { 106, 145, 1  },
+    { 179, 121, 1  },
+    { 223, 1  , 34 },
+    { 208, 1  , 1  }
+};
+
+
+
+const vp8_mbsplit vp8_mbsplits [VP8_NUMMBSPLITS] =
+{
+    {
+        0,  0,  0,  0,
+        0,  0,  0,  0,
+        1,  1,  1,  1,
+        1,  1,  1,  1,
+    },
+    {
+        0,  0,  1,  1,
+        0,  0,  1,  1,
+        0,  0,  1,  1,
+        0,  0,  1,  1,
+    },
+    {
+        0,  0,  1,  1,
+        0,  0,  1,  1,
+        2,  2,  3,  3,
+        2,  2,  3,  3,
+    },
+    {
+        0,  1,  2,  3,
+        4,  5,  6,  7,
+        8,  9,  10, 11,
+        12, 13, 14, 15,
+    }
+};
+
+const int vp8_mbsplit_count [VP8_NUMMBSPLITS] = { 2, 2, 4, 16};
+
+const vp8_prob vp8_mbsplit_probs [VP8_NUMMBSPLITS-1] = { 110, 111, 150};
+
+
+/* Array indices are identical to previously-existing INTRAMODECONTEXTNODES. */
+
+const vp8_tree_index vp8_bmode_tree[18] =     /* INTRAMODECONTEXTNODE value */
+{
+    -B_DC_PRED, 2,                             /* 0 = DC_NODE */
+    -B_TM_PRED, 4,                            /* 1 = TM_NODE */
+    -B_VE_PRED, 6,                           /* 2 = VE_NODE */
+    8, 12,                                  /* 3 = COM_NODE */
+    -B_HE_PRED, 10,                        /* 4 = HE_NODE */
+    -B_RD_PRED, -B_VR_PRED,               /* 5 = RD_NODE */
+    -B_LD_PRED, 14,                        /* 6 = LD_NODE */
+    -B_VL_PRED, 16,                      /* 7 = VL_NODE */
+    -B_HD_PRED, -B_HU_PRED             /* 8 = HD_NODE */
+};
+
+/* Again, these trees use the same probability indices as their
+   explicitly-programmed predecessors. */
+
+const vp8_tree_index vp8_ymode_tree[8] =
+{
+    -DC_PRED, 2,
+    4, 6,
+    -V_PRED, -H_PRED,
+    -TM_PRED, -B_PRED
+};
+
+const vp8_tree_index vp8_kf_ymode_tree[8] =
+{
+    -B_PRED, 2,
+    4, 6,
+    -DC_PRED, -V_PRED,
+    -H_PRED, -TM_PRED
+};
+
+const vp8_tree_index vp8_uv_mode_tree[6] =
+{
+    -DC_PRED, 2,
+    -V_PRED, 4,
+    -H_PRED, -TM_PRED
+};
+
+const vp8_tree_index vp8_mbsplit_tree[6] =
+{
+    -3, 2,
+    -2, 4,
+    -0, -1
+};
+
+const vp8_tree_index vp8_mv_ref_tree[8] =
+{
+    -ZEROMV, 2,
+    -NEARESTMV, 4,
+    -NEARMV, 6,
+    -NEWMV, -SPLITMV
+};
+
+const vp8_tree_index vp8_sub_mv_ref_tree[6] =
+{
+    -LEFT4X4, 2,
+    -ABOVE4X4, 4,
+    -ZERO4X4, -NEW4X4
+};
+
+const vp8_tree_index vp8_small_mvtree [14] =
+{
+    2, 8,
+    4, 6,
+    -0, -1,
+    -2, -3,
+    10, 12,
+    -4, -5,
+    -6, -7
+};
+
+void vp8_init_mbmode_probs(VP8_COMMON *x)
+{
+    memcpy(x->fc.ymode_prob, vp8_ymode_prob, sizeof(vp8_ymode_prob));
+    memcpy(x->fc.uv_mode_prob, vp8_uv_mode_prob, sizeof(vp8_uv_mode_prob));
+    memcpy(x->fc.sub_mv_ref_prob, sub_mv_ref_prob, sizeof(sub_mv_ref_prob));
+}
+
+void vp8_default_bmode_probs(vp8_prob p [VP8_BINTRAMODES-1])
+{
+    memcpy(p, vp8_bmode_prob, sizeof(vp8_bmode_prob));
+}
+
diff --git a/libvpx/libvpx/vp8/common/entropymode.h b/libvpx/libvpx/vp8/common/entropymode.h
new file mode 100644
index 0000000..81bdfc4
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/entropymode.h
@@ -0,0 +1,88 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_ENTROPYMODE_H_
+#define VP8_COMMON_ENTROPYMODE_H_
+
+#include "onyxc_int.h"
+#include "treecoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef enum
+{
+    SUBMVREF_NORMAL,
+    SUBMVREF_LEFT_ZED,
+    SUBMVREF_ABOVE_ZED,
+    SUBMVREF_LEFT_ABOVE_SAME,
+    SUBMVREF_LEFT_ABOVE_ZED
+} sumvfref_t;
+
+typedef int vp8_mbsplit[16];
+
+#define VP8_NUMMBSPLITS 4
+
+extern const vp8_mbsplit vp8_mbsplits [VP8_NUMMBSPLITS];
+
+extern const int vp8_mbsplit_count [VP8_NUMMBSPLITS];    /* # of subsets */
+
+extern const vp8_prob vp8_mbsplit_probs [VP8_NUMMBSPLITS-1];
+
+extern int vp8_mv_cont(const int_mv *l, const int_mv *a);
+#define SUBMVREF_COUNT 5
+extern const vp8_prob vp8_sub_mv_ref_prob2 [SUBMVREF_COUNT][VP8_SUBMVREFS-1];
+
+
+extern const unsigned int vp8_kf_default_bmode_counts [VP8_BINTRAMODES] [VP8_BINTRAMODES] [VP8_BINTRAMODES];
+
+
+extern const vp8_tree_index vp8_bmode_tree[];
+
+extern const vp8_tree_index  vp8_ymode_tree[];
+extern const vp8_tree_index  vp8_kf_ymode_tree[];
+extern const vp8_tree_index  vp8_uv_mode_tree[];
+
+extern const vp8_tree_index  vp8_mbsplit_tree[];
+extern const vp8_tree_index  vp8_mv_ref_tree[];
+extern const vp8_tree_index  vp8_sub_mv_ref_tree[];
+
+extern const struct vp8_token_struct vp8_bmode_encodings[VP8_BINTRAMODES];
+extern const struct vp8_token_struct vp8_ymode_encodings[VP8_YMODES];
+extern const struct vp8_token_struct vp8_kf_ymode_encodings[VP8_YMODES];
+extern const struct vp8_token_struct vp8_uv_mode_encodings[VP8_UV_MODES];
+extern const struct vp8_token_struct vp8_mbsplit_encodings[VP8_NUMMBSPLITS];
+
+/* Inter mode values do not start at zero */
+
+extern const struct vp8_token_struct vp8_mv_ref_encoding_array[VP8_MVREFS];
+extern const struct vp8_token_struct vp8_sub_mv_ref_encoding_array[VP8_SUBMVREFS];
+
+extern const vp8_tree_index vp8_small_mvtree[];
+
+extern const struct vp8_token_struct vp8_small_mvencodings[8];
+
+/* Key frame default mode probs */
+extern const vp8_prob vp8_kf_bmode_prob[VP8_BINTRAMODES][VP8_BINTRAMODES]
+[VP8_BINTRAMODES-1];
+extern const vp8_prob vp8_kf_uv_mode_prob[VP8_UV_MODES-1];
+extern const vp8_prob vp8_kf_ymode_prob[VP8_YMODES-1];
+
+void vp8_init_mbmode_probs(VP8_COMMON *x);
+void vp8_default_bmode_probs(vp8_prob dest [VP8_BINTRAMODES-1]);
+void vp8_kf_default_bmode_probs(vp8_prob dest [VP8_BINTRAMODES] [VP8_BINTRAMODES] [VP8_BINTRAMODES-1]);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_ENTROPYMODE_H_
diff --git a/libvpx/libvpx/vp8/common/entropymv.c b/libvpx/libvpx/vp8/common/entropymv.c
new file mode 100644
index 0000000..e5df1f0
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/entropymv.c
@@ -0,0 +1,49 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "entropymv.h"
+
+const MV_CONTEXT vp8_mv_update_probs[2] =
+{
+    {{
+        237,
+        246,
+        253, 253, 254, 254, 254, 254, 254,
+        254, 254, 254, 254, 254, 250, 250, 252, 254, 254
+    }},
+    {{
+        231,
+        243,
+        245, 253, 254, 254, 254, 254, 254,
+        254, 254, 254, 254, 254, 251, 251, 254, 254, 254
+    }}
+};
+const MV_CONTEXT vp8_default_mv_context[2] =
+{
+    {{
+        /* row */
+        162,                                        /* is short */
+        128,                                        /* sign */
+        225, 146, 172, 147, 214,  39, 156,          /* short tree */
+        128, 129, 132,  75, 145, 178, 206, 239, 254, 254 /* long bits */
+    }},
+
+
+
+    {{
+        /* same for column */
+        164,                                        /* is short */
+        128,
+        204, 170, 119, 235, 140, 230, 228,
+        128, 130, 130,  74, 148, 180, 203, 236, 254, 254 /* long bits */
+
+    }}
+};
diff --git a/libvpx/libvpx/vp8/common/entropymv.h b/libvpx/libvpx/vp8/common/entropymv.h
new file mode 100644
index 0000000..42840d5
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/entropymv.h
@@ -0,0 +1,52 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_ENTROPYMV_H_
+#define VP8_COMMON_ENTROPYMV_H_
+
+#include "treecoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum
+{
+    mv_max  = 1023,              /* max absolute value of a MV component */
+    MVvals = (2 * mv_max) + 1,   /* # possible values "" */
+    mvfp_max  = 255,              /* max absolute value of a full pixel MV component */
+    MVfpvals = (2 * mvfp_max) +1, /* # possible full pixel MV values */
+
+    mvlong_width = 10,       /* Large MVs have 9 bit magnitudes */
+    mvnum_short = 8,         /* magnitudes 0 through 7 */
+
+    /* probability offsets for coding each MV component */
+
+    mvpis_short = 0,         /* short (<= 7) vs long (>= 8) */
+    MVPsign,                /* sign for non-zero */
+    MVPshort,               /* 8 short values = 7-position tree */
+
+    MVPbits = MVPshort + mvnum_short - 1, /* mvlong_width long value bits */
+    MVPcount = MVPbits + mvlong_width    /* (with independent probabilities) */
+};
+
+typedef struct mv_context
+{
+    vp8_prob prob[MVPcount];  /* often come in row, col pairs */
+} MV_CONTEXT;
+
+extern const MV_CONTEXT vp8_mv_update_probs[2], vp8_default_mv_context[2];
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_ENTROPYMV_H_
diff --git a/libvpx/libvpx/vp8/common/extend.c b/libvpx/libvpx/vp8/common/extend.c
new file mode 100644
index 0000000..2d938ad
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/extend.c
@@ -0,0 +1,188 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "extend.h"
+#include "vpx_mem/vpx_mem.h"
+
+
+static void copy_and_extend_plane
+(
+    unsigned char *s, /* source */
+    int sp,           /* source pitch */
+    unsigned char *d, /* destination */
+    int dp,           /* destination pitch */
+    int h,            /* height */
+    int w,            /* width */
+    int et,           /* extend top border */
+    int el,           /* extend left border */
+    int eb,           /* extend bottom border */
+    int er            /* extend right border */
+)
+{
+    int i;
+    unsigned char *src_ptr1, *src_ptr2;
+    unsigned char *dest_ptr1, *dest_ptr2;
+    int linesize;
+
+    /* copy the left and right most columns out */
+    src_ptr1 = s;
+    src_ptr2 = s + w - 1;
+    dest_ptr1 = d - el;
+    dest_ptr2 = d + w;
+
+    for (i = 0; i < h; i++)
+    {
+        memset(dest_ptr1, src_ptr1[0], el);
+        memcpy(dest_ptr1 + el, src_ptr1, w);
+        memset(dest_ptr2, src_ptr2[0], er);
+        src_ptr1  += sp;
+        src_ptr2  += sp;
+        dest_ptr1 += dp;
+        dest_ptr2 += dp;
+    }
+
+    /* Now copy the top and bottom lines into each line of the respective
+     * borders
+     */
+    src_ptr1 = d - el;
+    src_ptr2 = d + dp * (h - 1) - el;
+    dest_ptr1 = d + dp * (-et) - el;
+    dest_ptr2 = d + dp * (h) - el;
+    linesize = el + er + w;
+
+    for (i = 0; i < et; i++)
+    {
+        memcpy(dest_ptr1, src_ptr1, linesize);
+        dest_ptr1 += dp;
+    }
+
+    for (i = 0; i < eb; i++)
+    {
+        memcpy(dest_ptr2, src_ptr2, linesize);
+        dest_ptr2 += dp;
+    }
+}
+
+
+void vp8_copy_and_extend_frame(YV12_BUFFER_CONFIG *src,
+                               YV12_BUFFER_CONFIG *dst)
+{
+    int et = dst->border;
+    int el = dst->border;
+    int eb = dst->border + dst->y_height - src->y_height;
+    int er = dst->border + dst->y_width - src->y_width;
+
+    copy_and_extend_plane(src->y_buffer, src->y_stride,
+                          dst->y_buffer, dst->y_stride,
+                          src->y_height, src->y_width,
+                          et, el, eb, er);
+
+    et = dst->border >> 1;
+    el = dst->border >> 1;
+    eb = (dst->border >> 1) + dst->uv_height - src->uv_height;
+    er = (dst->border >> 1) + dst->uv_width - src->uv_width;
+
+    copy_and_extend_plane(src->u_buffer, src->uv_stride,
+                          dst->u_buffer, dst->uv_stride,
+                          src->uv_height, src->uv_width,
+                          et, el, eb, er);
+
+    copy_and_extend_plane(src->v_buffer, src->uv_stride,
+                          dst->v_buffer, dst->uv_stride,
+                          src->uv_height, src->uv_width,
+                          et, el, eb, er);
+}
+
+
+void vp8_copy_and_extend_frame_with_rect(YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst,
+                                         int srcy, int srcx,
+                                         int srch, int srcw)
+{
+    int et = dst->border;
+    int el = dst->border;
+    int eb = dst->border + dst->y_height - src->y_height;
+    int er = dst->border + dst->y_width - src->y_width;
+    int src_y_offset = srcy * src->y_stride + srcx;
+    int dst_y_offset = srcy * dst->y_stride + srcx;
+    int src_uv_offset = ((srcy * src->uv_stride) >> 1) + (srcx >> 1);
+    int dst_uv_offset = ((srcy * dst->uv_stride) >> 1) + (srcx >> 1);
+
+    /* If the side is not touching the bounder then don't extend. */
+    if (srcy)
+      et = 0;
+    if (srcx)
+      el = 0;
+    if (srcy + srch != src->y_height)
+      eb = 0;
+    if (srcx + srcw != src->y_width)
+      er = 0;
+
+    copy_and_extend_plane(src->y_buffer + src_y_offset,
+                          src->y_stride,
+                          dst->y_buffer + dst_y_offset,
+                          dst->y_stride,
+                          srch, srcw,
+                          et, el, eb, er);
+
+    et = (et + 1) >> 1;
+    el = (el + 1) >> 1;
+    eb = (eb + 1) >> 1;
+    er = (er + 1) >> 1;
+    srch = (srch + 1) >> 1;
+    srcw = (srcw + 1) >> 1;
+
+    copy_and_extend_plane(src->u_buffer + src_uv_offset,
+                          src->uv_stride,
+                          dst->u_buffer + dst_uv_offset,
+                          dst->uv_stride,
+                          srch, srcw,
+                          et, el, eb, er);
+
+    copy_and_extend_plane(src->v_buffer + src_uv_offset,
+                          src->uv_stride,
+                          dst->v_buffer + dst_uv_offset,
+                          dst->uv_stride,
+                          srch, srcw,
+                          et, el, eb, er);
+}
+
+
+/* note the extension is only for the last row, for intra prediction purpose */
+void vp8_extend_mb_row(YV12_BUFFER_CONFIG *ybf,
+                       unsigned char *YPtr,
+                       unsigned char *UPtr,
+                       unsigned char *VPtr)
+{
+    int i;
+
+    YPtr += ybf->y_stride * 14;
+    UPtr += ybf->uv_stride * 6;
+    VPtr += ybf->uv_stride * 6;
+
+    for (i = 0; i < 4; i++)
+    {
+        YPtr[i] = YPtr[-1];
+        UPtr[i] = UPtr[-1];
+        VPtr[i] = VPtr[-1];
+    }
+
+    YPtr += ybf->y_stride;
+    UPtr += ybf->uv_stride;
+    VPtr += ybf->uv_stride;
+
+    for (i = 0; i < 4; i++)
+    {
+        YPtr[i] = YPtr[-1];
+        UPtr[i] = UPtr[-1];
+        VPtr[i] = VPtr[-1];
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/extend.h b/libvpx/libvpx/vp8/common/extend.h
new file mode 100644
index 0000000..068f4ac
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/extend.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_EXTEND_H_
+#define VP8_COMMON_EXTEND_H_
+
+#include "vpx_scale/yv12config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_extend_mb_row(YV12_BUFFER_CONFIG *ybf, unsigned char *YPtr, unsigned char *UPtr, unsigned char *VPtr);
+void vp8_copy_and_extend_frame(YV12_BUFFER_CONFIG *src,
+                               YV12_BUFFER_CONFIG *dst);
+void vp8_copy_and_extend_frame_with_rect(YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst,
+                                         int srcy, int srcx,
+                                         int srch, int srcw);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_EXTEND_H_
diff --git a/libvpx/libvpx/vp8/common/filter.c b/libvpx/libvpx/vp8/common/filter.c
new file mode 100644
index 0000000..84c608e
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/filter.c
@@ -0,0 +1,493 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "filter.h"
+#include "./vp8_rtcd.h"
+
+DECLARE_ALIGNED(16, const short, vp8_bilinear_filters[8][2]) =
+{
+    { 128,   0 },
+    { 112,  16 },
+    {  96,  32 },
+    {  80,  48 },
+    {  64,  64 },
+    {  48,  80 },
+    {  32,  96 },
+    {  16, 112 }
+};
+
+DECLARE_ALIGNED(16, const short, vp8_sub_pel_filters[8][6]) =
+{
+
+    { 0,  0,  128,    0,   0,  0 },         /* note that 1/8 pel positions are just as per alpha -0.5 bicubic */
+    { 0, -6,  123,   12,  -1,  0 },
+    { 2, -11, 108,   36,  -8,  1 },         /* New 1/4 pel 6 tap filter */
+    { 0, -9,   93,   50,  -6,  0 },
+    { 3, -16,  77,   77, -16,  3 },         /* New 1/2 pel 6 tap filter */
+    { 0, -6,   50,   93,  -9,  0 },
+    { 1, -8,   36,  108, -11,  2 },         /* New 1/4 pel 6 tap filter */
+    { 0, -1,   12,  123,  -6,  0 },
+};
+
+static void filter_block2d_first_pass
+(
+    unsigned char *src_ptr,
+    int *output_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int pixel_step,
+    unsigned int output_height,
+    unsigned int output_width,
+    const short *vp8_filter
+)
+{
+    unsigned int i, j;
+    int  Temp;
+
+    for (i = 0; i < output_height; i++)
+    {
+        for (j = 0; j < output_width; j++)
+        {
+            Temp = ((int)src_ptr[-2 * (int)pixel_step] * vp8_filter[0]) +
+                   ((int)src_ptr[-1 * (int)pixel_step] * vp8_filter[1]) +
+                   ((int)src_ptr[0]                 * vp8_filter[2]) +
+                   ((int)src_ptr[pixel_step]         * vp8_filter[3]) +
+                   ((int)src_ptr[2*pixel_step]       * vp8_filter[4]) +
+                   ((int)src_ptr[3*pixel_step]       * vp8_filter[5]) +
+                   (VP8_FILTER_WEIGHT >> 1);      /* Rounding */
+
+            /* Normalize back to 0-255 */
+            Temp = Temp >> VP8_FILTER_SHIFT;
+
+            if (Temp < 0)
+                Temp = 0;
+            else if (Temp > 255)
+                Temp = 255;
+
+            output_ptr[j] = Temp;
+            src_ptr++;
+        }
+
+        /* Next row... */
+        src_ptr    += src_pixels_per_line - output_width;
+        output_ptr += output_width;
+    }
+}
+
+static void filter_block2d_second_pass
+(
+    int *src_ptr,
+    unsigned char *output_ptr,
+    int output_pitch,
+    unsigned int src_pixels_per_line,
+    unsigned int pixel_step,
+    unsigned int output_height,
+    unsigned int output_width,
+    const short *vp8_filter
+)
+{
+    unsigned int i, j;
+    int  Temp;
+
+    for (i = 0; i < output_height; i++)
+    {
+        for (j = 0; j < output_width; j++)
+        {
+            /* Apply filter */
+            Temp = ((int)src_ptr[-2 * (int)pixel_step] * vp8_filter[0]) +
+                   ((int)src_ptr[-1 * (int)pixel_step] * vp8_filter[1]) +
+                   ((int)src_ptr[0]                 * vp8_filter[2]) +
+                   ((int)src_ptr[pixel_step]         * vp8_filter[3]) +
+                   ((int)src_ptr[2*pixel_step]       * vp8_filter[4]) +
+                   ((int)src_ptr[3*pixel_step]       * vp8_filter[5]) +
+                   (VP8_FILTER_WEIGHT >> 1);   /* Rounding */
+
+            /* Normalize back to 0-255 */
+            Temp = Temp >> VP8_FILTER_SHIFT;
+
+            if (Temp < 0)
+                Temp = 0;
+            else if (Temp > 255)
+                Temp = 255;
+
+            output_ptr[j] = (unsigned char)Temp;
+            src_ptr++;
+        }
+
+        /* Start next row */
+        src_ptr    += src_pixels_per_line - output_width;
+        output_ptr += output_pitch;
+    }
+}
+
+
+static void filter_block2d
+(
+    unsigned char  *src_ptr,
+    unsigned char  *output_ptr,
+    unsigned int src_pixels_per_line,
+    int output_pitch,
+    const short  *HFilter,
+    const short  *VFilter
+)
+{
+    int FData[9*4]; /* Temp data buffer used in filtering */
+
+    /* First filter 1-D horizontally... */
+    filter_block2d_first_pass(src_ptr - (2 * src_pixels_per_line), FData, src_pixels_per_line, 1, 9, 4, HFilter);
+
+    /* then filter verticaly... */
+    filter_block2d_second_pass(FData + 8, output_ptr, output_pitch, 4, 4, 4, 4, VFilter);
+}
+
+
+void vp8_sixtap_predict4x4_c
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+
+    HFilter = vp8_sub_pel_filters[xoffset];   /* 6 tap */
+    VFilter = vp8_sub_pel_filters[yoffset];   /* 6 tap */
+
+    filter_block2d(src_ptr, dst_ptr, src_pixels_per_line, dst_pitch, HFilter, VFilter);
+}
+void vp8_sixtap_predict8x8_c
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+    int FData[13*16];   /* Temp data buffer used in filtering */
+
+    HFilter = vp8_sub_pel_filters[xoffset];   /* 6 tap */
+    VFilter = vp8_sub_pel_filters[yoffset];   /* 6 tap */
+
+    /* First filter 1-D horizontally... */
+    filter_block2d_first_pass(src_ptr - (2 * src_pixels_per_line), FData, src_pixels_per_line, 1, 13, 8, HFilter);
+
+
+    /* then filter verticaly... */
+    filter_block2d_second_pass(FData + 16, dst_ptr, dst_pitch, 8, 8, 8, 8, VFilter);
+
+}
+
+void vp8_sixtap_predict8x4_c
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+    int FData[13*16];   /* Temp data buffer used in filtering */
+
+    HFilter = vp8_sub_pel_filters[xoffset];   /* 6 tap */
+    VFilter = vp8_sub_pel_filters[yoffset];   /* 6 tap */
+
+    /* First filter 1-D horizontally... */
+    filter_block2d_first_pass(src_ptr - (2 * src_pixels_per_line), FData, src_pixels_per_line, 1, 9, 8, HFilter);
+
+
+    /* then filter verticaly... */
+    filter_block2d_second_pass(FData + 16, dst_ptr, dst_pitch, 8, 8, 4, 8, VFilter);
+
+}
+
+void vp8_sixtap_predict16x16_c
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short  *HFilter;
+    const short  *VFilter;
+    int FData[21*24];   /* Temp data buffer used in filtering */
+
+
+    HFilter = vp8_sub_pel_filters[xoffset];   /* 6 tap */
+    VFilter = vp8_sub_pel_filters[yoffset];   /* 6 tap */
+
+    /* First filter 1-D horizontally... */
+    filter_block2d_first_pass(src_ptr - (2 * src_pixels_per_line), FData, src_pixels_per_line, 1, 21, 16, HFilter);
+
+    /* then filter verticaly... */
+    filter_block2d_second_pass(FData + 32, dst_ptr, dst_pitch, 16, 16, 16, 16, VFilter);
+
+}
+
+
+/****************************************************************************
+ *
+ *  ROUTINE       : filter_block2d_bil_first_pass
+ *
+ *  INPUTS        : UINT8  *src_ptr    : Pointer to source block.
+ *                  UINT32  src_stride : Stride of source block.
+ *                  UINT32  height     : Block height.
+ *                  UINT32  width      : Block width.
+ *                  INT32  *vp8_filter : Array of 2 bi-linear filter taps.
+ *
+ *  OUTPUTS       : INT32  *dst_ptr    : Pointer to filtered block.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Applies a 1-D 2-tap bi-linear filter to the source block
+ *                  in the horizontal direction to produce the filtered output
+ *                  block. Used to implement first-pass of 2-D separable filter.
+ *
+ *  SPECIAL NOTES : Produces INT32 output to retain precision for next pass.
+ *                  Two filter taps should sum to VP8_FILTER_WEIGHT.
+ *
+ ****************************************************************************/
+static void filter_block2d_bil_first_pass
+(
+    unsigned char  *src_ptr,
+    unsigned short *dst_ptr,
+    unsigned int    src_stride,
+    unsigned int    height,
+    unsigned int    width,
+    const short    *vp8_filter
+)
+{
+    unsigned int i, j;
+
+    for (i = 0; i < height; i++)
+    {
+        for (j = 0; j < width; j++)
+        {
+            /* Apply bilinear filter */
+            dst_ptr[j] = (((int)src_ptr[0] * vp8_filter[0]) +
+                          ((int)src_ptr[1] * vp8_filter[1]) +
+                          (VP8_FILTER_WEIGHT / 2)) >> VP8_FILTER_SHIFT;
+            src_ptr++;
+        }
+
+        /* Next row... */
+        src_ptr += src_stride - width;
+        dst_ptr += width;
+    }
+}
+
+/****************************************************************************
+ *
+ *  ROUTINE       : filter_block2d_bil_second_pass
+ *
+ *  INPUTS        : INT32  *src_ptr    : Pointer to source block.
+ *                  UINT32  dst_pitch  : Destination block pitch.
+ *                  UINT32  height     : Block height.
+ *                  UINT32  width      : Block width.
+ *                  INT32  *vp8_filter : Array of 2 bi-linear filter taps.
+ *
+ *  OUTPUTS       : UINT16 *dst_ptr    : Pointer to filtered block.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Applies a 1-D 2-tap bi-linear filter to the source block
+ *                  in the vertical direction to produce the filtered output
+ *                  block. Used to implement second-pass of 2-D separable filter.
+ *
+ *  SPECIAL NOTES : Requires 32-bit input as produced by filter_block2d_bil_first_pass.
+ *                  Two filter taps should sum to VP8_FILTER_WEIGHT.
+ *
+ ****************************************************************************/
+static void filter_block2d_bil_second_pass
+(
+    unsigned short *src_ptr,
+    unsigned char  *dst_ptr,
+    int             dst_pitch,
+    unsigned int    height,
+    unsigned int    width,
+    const short    *vp8_filter
+)
+{
+    unsigned int  i, j;
+    int  Temp;
+
+    for (i = 0; i < height; i++)
+    {
+        for (j = 0; j < width; j++)
+        {
+            /* Apply filter */
+            Temp = ((int)src_ptr[0]     * vp8_filter[0]) +
+                   ((int)src_ptr[width] * vp8_filter[1]) +
+                   (VP8_FILTER_WEIGHT / 2);
+            dst_ptr[j] = (unsigned int)(Temp >> VP8_FILTER_SHIFT);
+            src_ptr++;
+        }
+
+        /* Next row... */
+        dst_ptr += dst_pitch;
+    }
+}
+
+
+/****************************************************************************
+ *
+ *  ROUTINE       : filter_block2d_bil
+ *
+ *  INPUTS        : UINT8  *src_ptr          : Pointer to source block.
+ *                  UINT32  src_pitch        : Stride of source block.
+ *                  UINT32  dst_pitch        : Stride of destination block.
+ *                  INT32  *HFilter          : Array of 2 horizontal filter taps.
+ *                  INT32  *VFilter          : Array of 2 vertical filter taps.
+ *                  INT32  Width             : Block width
+ *                  INT32  Height            : Block height
+ *
+ *  OUTPUTS       : UINT16 *dst_ptr       : Pointer to filtered block.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : 2-D filters an input block by applying a 2-tap
+ *                  bi-linear filter horizontally followed by a 2-tap
+ *                  bi-linear filter vertically on the result.
+ *
+ *  SPECIAL NOTES : The largest block size can be handled here is 16x16
+ *
+ ****************************************************************************/
+static void filter_block2d_bil
+(
+    unsigned char *src_ptr,
+    unsigned char *dst_ptr,
+    unsigned int   src_pitch,
+    unsigned int   dst_pitch,
+    const short   *HFilter,
+    const short   *VFilter,
+    int            Width,
+    int            Height
+)
+{
+
+    unsigned short FData[17*16];    /* Temp data buffer used in filtering */
+
+    /* First filter 1-D horizontally... */
+    filter_block2d_bil_first_pass(src_ptr, FData, src_pitch, Height + 1, Width, HFilter);
+
+    /* then 1-D vertically... */
+    filter_block2d_bil_second_pass(FData, dst_ptr, dst_pitch, Height, Width, VFilter);
+}
+
+
+void vp8_bilinear_predict4x4_c
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+    const short *HFilter;
+    const short *VFilter;
+
+    HFilter = vp8_bilinear_filters[xoffset];
+    VFilter = vp8_bilinear_filters[yoffset];
+#if 0
+    {
+        int i;
+        unsigned char temp1[16];
+        unsigned char temp2[16];
+
+        bilinear_predict4x4_mmx(src_ptr, src_pixels_per_line, xoffset, yoffset, temp1, 4);
+        filter_block2d_bil(src_ptr, temp2, src_pixels_per_line, 4, HFilter, VFilter, 4, 4);
+
+        for (i = 0; i < 16; i++)
+        {
+            if (temp1[i] != temp2[i])
+            {
+                bilinear_predict4x4_mmx(src_ptr, src_pixels_per_line, xoffset, yoffset, temp1, 4);
+                filter_block2d_bil(src_ptr, temp2, src_pixels_per_line, 4, HFilter, VFilter, 4, 4);
+            }
+        }
+    }
+#endif
+    filter_block2d_bil(src_ptr, dst_ptr, src_pixels_per_line, dst_pitch, HFilter, VFilter, 4, 4);
+
+}
+
+void vp8_bilinear_predict8x8_c
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short *HFilter;
+    const short *VFilter;
+
+    HFilter = vp8_bilinear_filters[xoffset];
+    VFilter = vp8_bilinear_filters[yoffset];
+
+    filter_block2d_bil(src_ptr, dst_ptr, src_pixels_per_line, dst_pitch, HFilter, VFilter, 8, 8);
+
+}
+
+void vp8_bilinear_predict8x4_c
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short *HFilter;
+    const short *VFilter;
+
+    HFilter = vp8_bilinear_filters[xoffset];
+    VFilter = vp8_bilinear_filters[yoffset];
+
+    filter_block2d_bil(src_ptr, dst_ptr, src_pixels_per_line, dst_pitch, HFilter, VFilter, 8, 4);
+
+}
+
+void vp8_bilinear_predict16x16_c
+(
+    unsigned char  *src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int  dst_pitch
+)
+{
+    const short *HFilter;
+    const short *VFilter;
+
+    HFilter = vp8_bilinear_filters[xoffset];
+    VFilter = vp8_bilinear_filters[yoffset];
+
+    filter_block2d_bil(src_ptr, dst_ptr, src_pixels_per_line, dst_pitch, HFilter, VFilter, 16, 16);
+}
diff --git a/libvpx/libvpx/vp8/common/filter.h b/libvpx/libvpx/vp8/common/filter.h
new file mode 100644
index 0000000..cfba775
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/filter.h
@@ -0,0 +1,32 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_FILTER_H_
+#define VP8_COMMON_FILTER_H_
+
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define BLOCK_HEIGHT_WIDTH 4
+#define VP8_FILTER_WEIGHT 128
+#define VP8_FILTER_SHIFT  7
+
+extern DECLARE_ALIGNED(16, const short, vp8_bilinear_filters[8][2]);
+extern DECLARE_ALIGNED(16, const short, vp8_sub_pel_filters[8][6]);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_FILTER_H_
diff --git a/libvpx/libvpx/vp8/common/findnearmv.c b/libvpx/libvpx/vp8/common/findnearmv.c
new file mode 100644
index 0000000..e8ee40f
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/findnearmv.c
@@ -0,0 +1,193 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "findnearmv.h"
+
+const unsigned char vp8_mbsplit_offset[4][16] = {
+    { 0,  8,  0,  0,  0,  0,  0,  0,  0,  0,   0,  0,  0,  0,  0,  0},
+    { 0,  2,  0,  0,  0,  0,  0,  0,  0,  0,   0,  0,  0,  0,  0,  0},
+    { 0,  2,  8, 10,  0,  0,  0,  0,  0,  0,   0,  0,  0,  0,  0,  0},
+    { 0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15}
+};
+
+/* Predict motion vectors using those from already-decoded nearby blocks.
+   Note that we only consider one 4x4 subblock from each candidate 16x16
+   macroblock.   */
+void vp8_find_near_mvs
+(
+    MACROBLOCKD *xd,
+    const MODE_INFO *here,
+    int_mv *nearest,
+    int_mv *nearby,
+    int_mv *best_mv,
+    int cnt[4],
+    int refframe,
+    int *ref_frame_sign_bias
+)
+{
+    const MODE_INFO *above = here - xd->mode_info_stride;
+    const MODE_INFO *left = here - 1;
+    const MODE_INFO *aboveleft = above - 1;
+    int_mv            near_mvs[4];
+    int_mv           *mv = near_mvs;
+    int             *cntx = cnt;
+    enum {CNT_INTRA, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV};
+
+    /* Zero accumulators */
+    mv[0].as_int = mv[1].as_int = mv[2].as_int = 0;
+    cnt[0] = cnt[1] = cnt[2] = cnt[3] = 0;
+
+    /* Process above */
+    if (above->mbmi.ref_frame != INTRA_FRAME)
+    {
+        if (above->mbmi.mv.as_int)
+        {
+            (++mv)->as_int = above->mbmi.mv.as_int;
+            mv_bias(ref_frame_sign_bias[above->mbmi.ref_frame], refframe, mv, ref_frame_sign_bias);
+            ++cntx;
+        }
+
+        *cntx += 2;
+    }
+
+    /* Process left */
+    if (left->mbmi.ref_frame != INTRA_FRAME)
+    {
+        if (left->mbmi.mv.as_int)
+        {
+            int_mv this_mv;
+
+            this_mv.as_int = left->mbmi.mv.as_int;
+            mv_bias(ref_frame_sign_bias[left->mbmi.ref_frame], refframe, &this_mv, ref_frame_sign_bias);
+
+            if (this_mv.as_int != mv->as_int)
+            {
+                (++mv)->as_int = this_mv.as_int;
+                ++cntx;
+            }
+
+            *cntx += 2;
+        }
+        else
+            cnt[CNT_INTRA] += 2;
+    }
+
+    /* Process above left */
+    if (aboveleft->mbmi.ref_frame != INTRA_FRAME)
+    {
+        if (aboveleft->mbmi.mv.as_int)
+        {
+            int_mv this_mv;
+
+            this_mv.as_int = aboveleft->mbmi.mv.as_int;
+            mv_bias(ref_frame_sign_bias[aboveleft->mbmi.ref_frame], refframe, &this_mv, ref_frame_sign_bias);
+
+            if (this_mv.as_int != mv->as_int)
+            {
+                (++mv)->as_int = this_mv.as_int;
+                ++cntx;
+            }
+
+            *cntx += 1;
+        }
+        else
+            cnt[CNT_INTRA] += 1;
+    }
+
+    /* If we have three distinct MV's ... */
+    if (cnt[CNT_SPLITMV])
+    {
+        /* See if above-left MV can be merged with NEAREST */
+        if (mv->as_int == near_mvs[CNT_NEAREST].as_int)
+            cnt[CNT_NEAREST] += 1;
+    }
+
+    cnt[CNT_SPLITMV] = ((above->mbmi.mode == SPLITMV)
+                        + (left->mbmi.mode == SPLITMV)) * 2
+                       + (aboveleft->mbmi.mode == SPLITMV);
+
+    /* Swap near and nearest if necessary */
+    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST])
+    {
+        int tmp;
+        tmp = cnt[CNT_NEAREST];
+        cnt[CNT_NEAREST] = cnt[CNT_NEAR];
+        cnt[CNT_NEAR] = tmp;
+        tmp = near_mvs[CNT_NEAREST].as_int;
+        near_mvs[CNT_NEAREST].as_int = near_mvs[CNT_NEAR].as_int;
+        near_mvs[CNT_NEAR].as_int = tmp;
+    }
+
+    /* Use near_mvs[0] to store the "best" MV */
+    if (cnt[CNT_NEAREST] >= cnt[CNT_INTRA])
+        near_mvs[CNT_INTRA] = near_mvs[CNT_NEAREST];
+
+    /* Set up return values */
+    best_mv->as_int = near_mvs[0].as_int;
+    nearest->as_int = near_mvs[CNT_NEAREST].as_int;
+    nearby->as_int = near_mvs[CNT_NEAR].as_int;
+}
+
+
+static void invert_and_clamp_mvs(int_mv *inv, int_mv *src, MACROBLOCKD *xd)
+{
+    inv->as_mv.row = src->as_mv.row * -1;
+    inv->as_mv.col = src->as_mv.col * -1;
+    vp8_clamp_mv2(inv, xd);
+    vp8_clamp_mv2(src, xd);
+}
+
+
+int vp8_find_near_mvs_bias
+(
+    MACROBLOCKD *xd,
+    const MODE_INFO *here,
+    int_mv mode_mv_sb[2][MB_MODE_COUNT],
+    int_mv best_mv_sb[2],
+    int cnt[4],
+    int refframe,
+    int *ref_frame_sign_bias
+)
+{
+    int sign_bias = ref_frame_sign_bias[refframe];
+
+    vp8_find_near_mvs(xd,
+                      here,
+                      &mode_mv_sb[sign_bias][NEARESTMV],
+                      &mode_mv_sb[sign_bias][NEARMV],
+                      &best_mv_sb[sign_bias],
+                      cnt,
+                      refframe,
+                      ref_frame_sign_bias);
+
+    invert_and_clamp_mvs(&mode_mv_sb[!sign_bias][NEARESTMV],
+                         &mode_mv_sb[sign_bias][NEARESTMV], xd);
+    invert_and_clamp_mvs(&mode_mv_sb[!sign_bias][NEARMV],
+                         &mode_mv_sb[sign_bias][NEARMV], xd);
+    invert_and_clamp_mvs(&best_mv_sb[!sign_bias],
+                         &best_mv_sb[sign_bias], xd);
+
+    return sign_bias;
+}
+
+
+vp8_prob *vp8_mv_ref_probs(
+    vp8_prob p[VP8_MVREFS-1], const int near_mv_ref_ct[4]
+)
+{
+    p[0] = vp8_mode_contexts [near_mv_ref_ct[0]] [0];
+    p[1] = vp8_mode_contexts [near_mv_ref_ct[1]] [1];
+    p[2] = vp8_mode_contexts [near_mv_ref_ct[2]] [2];
+    p[3] = vp8_mode_contexts [near_mv_ref_ct[3]] [3];
+    /*p[3] = vp8_mode_contexts [near_mv_ref_ct[1] + near_mv_ref_ct[2] + near_mv_ref_ct[3]] [3];*/
+    return p;
+}
+
diff --git a/libvpx/libvpx/vp8/common/findnearmv.h b/libvpx/libvpx/vp8/common/findnearmv.h
new file mode 100644
index 0000000..472a7b5
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/findnearmv.h
@@ -0,0 +1,195 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_FINDNEARMV_H_
+#define VP8_COMMON_FINDNEARMV_H_
+
+#include "./vpx_config.h"
+#include "mv.h"
+#include "blockd.h"
+#include "modecont.h"
+#include "treecoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+static INLINE void mv_bias(int refmb_ref_frame_sign_bias, int refframe,
+                           int_mv *mvp, const int *ref_frame_sign_bias)
+{
+    if (refmb_ref_frame_sign_bias != ref_frame_sign_bias[refframe])
+    {
+        mvp->as_mv.row *= -1;
+        mvp->as_mv.col *= -1;
+    }
+}
+
+#define LEFT_TOP_MARGIN (16 << 3)
+#define RIGHT_BOTTOM_MARGIN (16 << 3)
+static INLINE void vp8_clamp_mv2(int_mv *mv, const MACROBLOCKD *xd)
+{
+    if (mv->as_mv.col < (xd->mb_to_left_edge - LEFT_TOP_MARGIN))
+        mv->as_mv.col = xd->mb_to_left_edge - LEFT_TOP_MARGIN;
+    else if (mv->as_mv.col > xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN)
+        mv->as_mv.col = xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN;
+
+    if (mv->as_mv.row < (xd->mb_to_top_edge - LEFT_TOP_MARGIN))
+        mv->as_mv.row = xd->mb_to_top_edge - LEFT_TOP_MARGIN;
+    else if (mv->as_mv.row > xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN)
+        mv->as_mv.row = xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN;
+}
+
+static INLINE void vp8_clamp_mv(int_mv *mv, int mb_to_left_edge,
+                                int mb_to_right_edge, int mb_to_top_edge,
+                                int mb_to_bottom_edge)
+{
+    mv->as_mv.col = (mv->as_mv.col < mb_to_left_edge) ?
+        mb_to_left_edge : mv->as_mv.col;
+    mv->as_mv.col = (mv->as_mv.col > mb_to_right_edge) ?
+        mb_to_right_edge : mv->as_mv.col;
+    mv->as_mv.row = (mv->as_mv.row < mb_to_top_edge) ?
+        mb_to_top_edge : mv->as_mv.row;
+    mv->as_mv.row = (mv->as_mv.row > mb_to_bottom_edge) ?
+        mb_to_bottom_edge : mv->as_mv.row;
+}
+static INLINE unsigned int vp8_check_mv_bounds(int_mv *mv, int mb_to_left_edge,
+                                               int mb_to_right_edge,
+                                               int mb_to_top_edge,
+                                               int mb_to_bottom_edge)
+{
+    unsigned int need_to_clamp;
+    need_to_clamp = (mv->as_mv.col < mb_to_left_edge);
+    need_to_clamp |= (mv->as_mv.col > mb_to_right_edge);
+    need_to_clamp |= (mv->as_mv.row < mb_to_top_edge);
+    need_to_clamp |= (mv->as_mv.row > mb_to_bottom_edge);
+    return need_to_clamp;
+}
+
+void vp8_find_near_mvs
+(
+    MACROBLOCKD *xd,
+    const MODE_INFO *here,
+    int_mv *nearest, int_mv *nearby, int_mv *best,
+    int near_mv_ref_cts[4],
+    int refframe,
+    int *ref_frame_sign_bias
+);
+
+
+int vp8_find_near_mvs_bias
+(
+    MACROBLOCKD *xd,
+    const MODE_INFO *here,
+    int_mv mode_mv_sb[2][MB_MODE_COUNT],
+    int_mv best_mv_sb[2],
+    int cnt[4],
+    int refframe,
+    int *ref_frame_sign_bias
+);
+
+
+vp8_prob *vp8_mv_ref_probs(
+    vp8_prob p[VP8_MVREFS-1], const int near_mv_ref_ct[4]
+);
+
+extern const unsigned char vp8_mbsplit_offset[4][16];
+
+
+static INLINE uint32_t left_block_mv(const MODE_INFO *cur_mb, int b)
+{
+    if (!(b & 3))
+    {
+        /* On L edge, get from MB to left of us */
+        --cur_mb;
+
+        if(cur_mb->mbmi.mode != SPLITMV)
+            return cur_mb->mbmi.mv.as_int;
+        b += 4;
+    }
+
+    return (cur_mb->bmi + b - 1)->mv.as_int;
+}
+
+static INLINE uint32_t above_block_mv(const MODE_INFO *cur_mb, int b,
+                                      int mi_stride)
+{
+    if (!(b >> 2))
+    {
+        /* On top edge, get from MB above us */
+        cur_mb -= mi_stride;
+
+        if(cur_mb->mbmi.mode != SPLITMV)
+            return cur_mb->mbmi.mv.as_int;
+        b += 16;
+    }
+
+    return (cur_mb->bmi + (b - 4))->mv.as_int;
+}
+static INLINE B_PREDICTION_MODE left_block_mode(const MODE_INFO *cur_mb, int b)
+{
+    if (!(b & 3))
+    {
+        /* On L edge, get from MB to left of us */
+        --cur_mb;
+        switch (cur_mb->mbmi.mode)
+        {
+            case B_PRED:
+              return (cur_mb->bmi + b + 3)->as_mode;
+            case DC_PRED:
+                return B_DC_PRED;
+            case V_PRED:
+                return B_VE_PRED;
+            case H_PRED:
+                return B_HE_PRED;
+            case TM_PRED:
+                return B_TM_PRED;
+            default:
+                return B_DC_PRED;
+        }
+    }
+
+    return (cur_mb->bmi + b - 1)->as_mode;
+}
+
+static INLINE B_PREDICTION_MODE above_block_mode(const MODE_INFO *cur_mb, int b,
+                                                 int mi_stride)
+{
+    if (!(b >> 2))
+    {
+        /* On top edge, get from MB above us */
+        cur_mb -= mi_stride;
+
+        switch (cur_mb->mbmi.mode)
+        {
+            case B_PRED:
+              return (cur_mb->bmi + b + 12)->as_mode;
+            case DC_PRED:
+                return B_DC_PRED;
+            case V_PRED:
+                return B_VE_PRED;
+            case H_PRED:
+                return B_HE_PRED;
+            case TM_PRED:
+                return B_TM_PRED;
+            default:
+                return B_DC_PRED;
+        }
+    }
+
+    return (cur_mb->bmi + b - 4)->as_mode;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_FINDNEARMV_H_
diff --git a/libvpx/libvpx/vp8/common/generic/systemdependent.c b/libvpx/libvpx/vp8/common/generic/systemdependent.c
new file mode 100644
index 0000000..6d5f302
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/generic/systemdependent.c
@@ -0,0 +1,106 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#if ARCH_ARM
+#include "vpx_ports/arm.h"
+#elif ARCH_X86 || ARCH_X86_64
+#include "vpx_ports/x86.h"
+#endif
+#include "vp8/common/onyxc_int.h"
+#include "vp8/common/systemdependent.h"
+
+#if CONFIG_MULTITHREAD
+#if HAVE_UNISTD_H && !defined(__OS2__)
+#include <unistd.h>
+#elif defined(_WIN32)
+#include <windows.h>
+typedef void (WINAPI *PGNSI)(LPSYSTEM_INFO);
+#elif defined(__OS2__)
+#define INCL_DOS
+#define INCL_DOSSPINLOCK
+#include <os2.h>
+#endif
+#endif
+
+#if CONFIG_MULTITHREAD
+static int get_cpu_count()
+{
+    int core_count = 16;
+
+#if HAVE_UNISTD_H && !defined(__OS2__)
+#if defined(_SC_NPROCESSORS_ONLN)
+    core_count = sysconf(_SC_NPROCESSORS_ONLN);
+#elif defined(_SC_NPROC_ONLN)
+    core_count = sysconf(_SC_NPROC_ONLN);
+#endif
+#elif defined(_WIN32)
+    {
+#if _WIN32_WINNT >= 0x0501
+        SYSTEM_INFO sysinfo;
+        GetNativeSystemInfo(&sysinfo);
+#else
+        PGNSI pGNSI;
+        SYSTEM_INFO sysinfo;
+
+        /* Call GetNativeSystemInfo if supported or
+         * GetSystemInfo otherwise. */
+
+        pGNSI = (PGNSI) GetProcAddress(
+                GetModuleHandle(TEXT("kernel32.dll")), "GetNativeSystemInfo");
+        if (pGNSI != NULL)
+            pGNSI(&sysinfo);
+        else
+            GetSystemInfo(&sysinfo);
+#endif
+
+        core_count = sysinfo.dwNumberOfProcessors;
+    }
+#elif defined(__OS2__)
+    {
+        ULONG proc_id;
+        ULONG status;
+
+        core_count = 0;
+        for (proc_id = 1; ; proc_id++)
+        {
+            if (DosGetProcessorStatus(proc_id, &status))
+                break;
+
+            if (status == PROC_ONLINE)
+                core_count++;
+        }
+    }
+#else
+    /* other platforms */
+#endif
+
+    return core_count > 0 ? core_count : 1;
+}
+#endif
+
+void vp8_clear_system_state_c() {};
+
+void vp8_machine_specific_config(VP8_COMMON *ctx)
+{
+#if CONFIG_MULTITHREAD
+    ctx->processor_core_count = get_cpu_count();
+#else
+    (void)ctx;
+#endif /* CONFIG_MULTITHREAD */
+
+#if ARCH_ARM
+    ctx->cpu_caps = arm_cpu_caps();
+#elif ARCH_X86 || ARCH_X86_64
+    ctx->cpu_caps = x86_simd_caps();
+#endif
+}
diff --git a/libvpx/libvpx/vp8/common/header.h b/libvpx/libvpx/vp8/common/header.h
new file mode 100644
index 0000000..e27bca1
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/header.h
@@ -0,0 +1,51 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_HEADER_H_
+#define VP8_COMMON_HEADER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* 24 bits total */
+typedef struct
+{
+    unsigned int type: 1;
+    unsigned int version: 3;
+    unsigned int show_frame: 1;
+
+    /* Allow 2^20 bytes = 8 megabits for first partition */
+
+    unsigned int first_partition_length_in_bytes: 19;
+
+#ifdef PACKET_TESTING
+    unsigned int frame_number;
+    unsigned int update_gold: 1;
+    unsigned int uses_gold: 1;
+    unsigned int update_last: 1;
+    unsigned int uses_last: 1;
+#endif
+
+} VP8_HEADER;
+
+#ifdef PACKET_TESTING
+#define VP8_HEADER_SIZE 8
+#else
+#define VP8_HEADER_SIZE 3
+#endif
+
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_HEADER_H_
diff --git a/libvpx/libvpx/vp8/common/idct_blk.c b/libvpx/libvpx/vp8/common/idct_blk.c
new file mode 100644
index 0000000..8aa7d9b
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/idct_blk.c
@@ -0,0 +1,90 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vpx_mem/vpx_mem.h"
+
+void vp8_dequant_idct_add_c(short *input, short *dq,
+                            unsigned char *dest, int stride);
+void vp8_dc_only_idct_add_c(short input_dc, unsigned char * pred,
+                            int pred_stride, unsigned char *dst_ptr,
+                            int dst_stride);
+
+void vp8_dequant_idct_add_y_block_c
+            (short *q, short *dq,
+             unsigned char *dst, int stride, char *eobs)
+{
+    int i, j;
+
+    for (i = 0; i < 4; i++)
+    {
+        for (j = 0; j < 4; j++)
+        {
+            if (*eobs++ > 1)
+                vp8_dequant_idct_add_c (q, dq, dst, stride);
+            else
+            {
+                vp8_dc_only_idct_add_c (q[0]*dq[0], dst, stride, dst, stride);
+                memset(q, 0, 2 * sizeof(q[0]));
+            }
+
+            q   += 16;
+            dst += 4;
+        }
+
+        dst += 4*stride - 16;
+    }
+}
+
+void vp8_dequant_idct_add_uv_block_c
+            (short *q, short *dq,
+             unsigned char *dstu, unsigned char *dstv, int stride, char *eobs)
+{
+    int i, j;
+
+    for (i = 0; i < 2; i++)
+    {
+        for (j = 0; j < 2; j++)
+        {
+            if (*eobs++ > 1)
+                vp8_dequant_idct_add_c (q, dq, dstu, stride);
+            else
+            {
+                vp8_dc_only_idct_add_c (q[0]*dq[0], dstu, stride, dstu, stride);
+                memset(q, 0, 2 * sizeof(q[0]));
+            }
+
+            q    += 16;
+            dstu += 4;
+        }
+
+        dstu += 4*stride - 8;
+    }
+
+    for (i = 0; i < 2; i++)
+    {
+        for (j = 0; j < 2; j++)
+        {
+            if (*eobs++ > 1)
+                vp8_dequant_idct_add_c (q, dq, dstv, stride);
+            else
+            {
+                vp8_dc_only_idct_add_c (q[0]*dq[0], dstv, stride, dstv, stride);
+                memset(q, 0, 2 * sizeof(q[0]));
+            }
+
+            q    += 16;
+            dstv += 4;
+        }
+
+        dstv += 4*stride - 8;
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/idctllm.c b/libvpx/libvpx/vp8/common/idctllm.c
new file mode 100644
index 0000000..f5403c5
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/idctllm.c
@@ -0,0 +1,205 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+
+/****************************************************************************
+ * Notes:
+ *
+ * This implementation makes use of 16 bit fixed point verio of two multiply
+ * constants:
+ *         1.   sqrt(2) * cos (pi/8)
+ *         2.   sqrt(2) * sin (pi/8)
+ * Becuase the first constant is bigger than 1, to maintain the same 16 bit
+ * fixed point precision as the second one, we use a trick of
+ *         x * a = x + x*(a-1)
+ * so
+ *         x * sqrt(2) * cos (pi/8) = x + x * (sqrt(2) *cos(pi/8)-1).
+ **************************************************************************/
+static const int cospi8sqrt2minus1 = 20091;
+static const int sinpi8sqrt2      = 35468;
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred_ptr,
+                            int pred_stride, unsigned char *dst_ptr,
+                            int dst_stride)
+{
+    int i;
+    int r, c;
+    int a1, b1, c1, d1;
+    short output[16];
+    short *ip = input;
+    short *op = output;
+    int temp1, temp2;
+    int shortpitch = 4;
+
+    for (i = 0; i < 4; i++)
+    {
+        a1 = ip[0] + ip[8];
+        b1 = ip[0] - ip[8];
+
+        temp1 = (ip[4] * sinpi8sqrt2) >> 16;
+        temp2 = ip[12] + ((ip[12] * cospi8sqrt2minus1) >> 16);
+        c1 = temp1 - temp2;
+
+        temp1 = ip[4] + ((ip[4] * cospi8sqrt2minus1) >> 16);
+        temp2 = (ip[12] * sinpi8sqrt2) >> 16;
+        d1 = temp1 + temp2;
+
+        op[shortpitch*0] = a1 + d1;
+        op[shortpitch*3] = a1 - d1;
+
+        op[shortpitch*1] = b1 + c1;
+        op[shortpitch*2] = b1 - c1;
+
+        ip++;
+        op++;
+    }
+
+    ip = output;
+    op = output;
+
+    for (i = 0; i < 4; i++)
+    {
+        a1 = ip[0] + ip[2];
+        b1 = ip[0] - ip[2];
+
+        temp1 = (ip[1] * sinpi8sqrt2) >> 16;
+        temp2 = ip[3] + ((ip[3] * cospi8sqrt2minus1) >> 16);
+        c1 = temp1 - temp2;
+
+        temp1 = ip[1] + ((ip[1] * cospi8sqrt2minus1) >> 16);
+        temp2 = (ip[3] * sinpi8sqrt2) >> 16;
+        d1 = temp1 + temp2;
+
+
+        op[0] = (a1 + d1 + 4) >> 3;
+        op[3] = (a1 - d1 + 4) >> 3;
+
+        op[1] = (b1 + c1 + 4) >> 3;
+        op[2] = (b1 - c1 + 4) >> 3;
+
+        ip += shortpitch;
+        op += shortpitch;
+    }
+
+    ip = output;
+    for (r = 0; r < 4; r++)
+    {
+        for (c = 0; c < 4; c++)
+        {
+            int a = ip[c] + pred_ptr[c] ;
+
+            if (a < 0)
+                a = 0;
+
+            if (a > 255)
+                a = 255;
+
+            dst_ptr[c] = (unsigned char) a ;
+        }
+        ip += 4;
+        dst_ptr += dst_stride;
+        pred_ptr += pred_stride;
+    }
+}
+
+void vp8_dc_only_idct_add_c(short input_dc, unsigned char *pred_ptr,
+                            int pred_stride, unsigned char *dst_ptr,
+                            int dst_stride)
+{
+    int a1 = ((input_dc + 4) >> 3);
+    int r, c;
+
+    for (r = 0; r < 4; r++)
+    {
+        for (c = 0; c < 4; c++)
+        {
+            int a = a1 + pred_ptr[c] ;
+
+            if (a < 0)
+                a = 0;
+
+            if (a > 255)
+                a = 255;
+
+            dst_ptr[c] = (unsigned char) a ;
+        }
+
+        dst_ptr += dst_stride;
+        pred_ptr += pred_stride;
+    }
+
+}
+
+void vp8_short_inv_walsh4x4_c(short *input, short *mb_dqcoeff)
+{
+    short output[16];
+    int i;
+    int a1, b1, c1, d1;
+    int a2, b2, c2, d2;
+    short *ip = input;
+    short *op = output;
+
+    for (i = 0; i < 4; i++)
+    {
+        a1 = ip[0] + ip[12];
+        b1 = ip[4] + ip[8];
+        c1 = ip[4] - ip[8];
+        d1 = ip[0] - ip[12];
+
+        op[0] = a1 + b1;
+        op[4] = c1 + d1;
+        op[8] = a1 - b1;
+        op[12] = d1 - c1;
+        ip++;
+        op++;
+    }
+
+    ip = output;
+    op = output;
+
+    for (i = 0; i < 4; i++)
+    {
+        a1 = ip[0] + ip[3];
+        b1 = ip[1] + ip[2];
+        c1 = ip[1] - ip[2];
+        d1 = ip[0] - ip[3];
+
+        a2 = a1 + b1;
+        b2 = c1 + d1;
+        c2 = a1 - b1;
+        d2 = d1 - c1;
+
+        op[0] = (a2 + 3) >> 3;
+        op[1] = (b2 + 3) >> 3;
+        op[2] = (c2 + 3) >> 3;
+        op[3] = (d2 + 3) >> 3;
+
+        ip += 4;
+        op += 4;
+    }
+
+    for(i = 0; i < 16; i++)
+    {
+        mb_dqcoeff[i * 16] = output[i];
+    }
+}
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *mb_dqcoeff)
+{
+    int i;
+    int a1;
+
+    a1 = ((input[0] + 3) >> 3);
+    for(i = 0; i < 16; i++)
+    {
+        mb_dqcoeff[i * 16] = a1;
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/invtrans.h b/libvpx/libvpx/vp8/common/invtrans.h
new file mode 100644
index 0000000..9cfea8d
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/invtrans.h
@@ -0,0 +1,70 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_INVTRANS_H_
+#define VP8_COMMON_INVTRANS_H_
+
+#include "./vpx_config.h"
+#include "vp8_rtcd.h"
+#include "blockd.h"
+#include "onyxc_int.h"
+
+#if CONFIG_MULTITHREAD
+#include "vpx_mem/vpx_mem.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static void eob_adjust(char *eobs, short *diff)
+{
+    /* eob adjust.... the idct can only skip if both the dc and eob are zero */
+    int js;
+    for(js = 0; js < 16; js++)
+    {
+        if((eobs[js] == 0) && (diff[0] != 0))
+            eobs[js]++;
+        diff+=16;
+    }
+}
+
+static INLINE void vp8_inverse_transform_mby(MACROBLOCKD *xd)
+{
+    short *DQC = xd->dequant_y1;
+
+    if (xd->mode_info_context->mbmi.mode != SPLITMV)
+    {
+        /* do 2nd order transform on the dc block */
+        if (xd->eobs[24] > 1)
+        {
+            vp8_short_inv_walsh4x4
+                (&xd->block[24].dqcoeff[0], xd->qcoeff);
+        }
+        else
+        {
+            vp8_short_inv_walsh4x4_1
+                (&xd->block[24].dqcoeff[0], xd->qcoeff);
+        }
+        eob_adjust(xd->eobs, xd->qcoeff);
+
+        DQC = xd->dequant_y1_dc;
+    }
+    vp8_dequant_idct_add_y_block
+                    (xd->qcoeff, DQC,
+                     xd->dst.y_buffer,
+                     xd->dst.y_stride, xd->eobs);
+}
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_INVTRANS_H_
diff --git a/libvpx/libvpx/vp8/common/loopfilter.h b/libvpx/libvpx/vp8/common/loopfilter.h
new file mode 100644
index 0000000..20a6bd3
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/loopfilter.h
@@ -0,0 +1,113 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_LOOPFILTER_H_
+#define VP8_COMMON_LOOPFILTER_H_
+
+#include "vpx_ports/mem.h"
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_LOOP_FILTER             63
+/* fraction of total macroblock rows to be used in fast filter level picking */
+/* has to be > 2 */
+#define PARTIAL_FRAME_FRACTION      8
+
+typedef enum
+{
+    NORMAL_LOOPFILTER = 0,
+    SIMPLE_LOOPFILTER = 1
+} LOOPFILTERTYPE;
+
+#if ARCH_ARM
+#define SIMD_WIDTH 1
+#else
+#define SIMD_WIDTH 16
+#endif
+
+/* Need to align this structure so when it is declared and
+ * passed it can be loaded into vector registers.
+ */
+typedef struct
+{
+    DECLARE_ALIGNED(SIMD_WIDTH, unsigned char, mblim[MAX_LOOP_FILTER + 1][SIMD_WIDTH]);
+    DECLARE_ALIGNED(SIMD_WIDTH, unsigned char, blim[MAX_LOOP_FILTER + 1][SIMD_WIDTH]);
+    DECLARE_ALIGNED(SIMD_WIDTH, unsigned char, lim[MAX_LOOP_FILTER + 1][SIMD_WIDTH]);
+    DECLARE_ALIGNED(SIMD_WIDTH, unsigned char, hev_thr[4][SIMD_WIDTH]);
+    unsigned char lvl[4][4][4];
+    unsigned char hev_thr_lut[2][MAX_LOOP_FILTER + 1];
+    unsigned char mode_lf_lut[10];
+} loop_filter_info_n;
+
+typedef struct loop_filter_info
+{
+    const unsigned char * mblim;
+    const unsigned char * blim;
+    const unsigned char * lim;
+    const unsigned char * hev_thr;
+} loop_filter_info;
+
+
+typedef void loop_filter_uvfunction
+(
+    unsigned char *u,   /* source pointer */
+    int p,              /* pitch */
+    const unsigned char *blimit,
+    const unsigned char *limit,
+    const unsigned char *thresh,
+    unsigned char *v
+);
+
+/* assorted loopfilter functions which get used elsewhere */
+struct VP8Common;
+struct macroblockd;
+struct modeinfo;
+
+void vp8_loop_filter_init(struct VP8Common *cm);
+
+void vp8_loop_filter_frame_init(struct VP8Common *cm,
+                                struct macroblockd *mbd,
+                                int default_filt_lvl);
+
+void vp8_loop_filter_frame(struct VP8Common *cm, struct macroblockd *mbd,
+                           int frame_type);
+
+void vp8_loop_filter_partial_frame(struct VP8Common *cm,
+                                   struct macroblockd *mbd,
+                                   int default_filt_lvl);
+
+void vp8_loop_filter_frame_yonly(struct VP8Common *cm,
+                                 struct macroblockd *mbd,
+                                 int default_filt_lvl);
+
+void vp8_loop_filter_update_sharpness(loop_filter_info_n *lfi,
+                                      int sharpness_lvl);
+
+void vp8_loop_filter_row_normal(struct VP8Common *cm,
+                                struct modeinfo *mode_info_context,
+                                int mb_row, int post_ystride, int post_uvstride,
+                                unsigned char *y_ptr, unsigned char *u_ptr,
+                                unsigned char *v_ptr);
+
+void vp8_loop_filter_row_simple(struct VP8Common *cm,
+                                struct modeinfo *mode_info_context,
+                                int mb_row, int post_ystride, int post_uvstride,
+                                unsigned char *y_ptr, unsigned char *u_ptr,
+                                unsigned char *v_ptr);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_LOOPFILTER_H_
diff --git a/libvpx/libvpx/vp8/common/loopfilter_filters.c b/libvpx/libvpx/vp8/common/loopfilter_filters.c
new file mode 100644
index 0000000..1d51696
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/loopfilter_filters.c
@@ -0,0 +1,430 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <stdlib.h>
+#include "loopfilter.h"
+#include "onyxc_int.h"
+
+typedef unsigned char uc;
+
+static signed char vp8_signed_char_clamp(int t)
+{
+    t = (t < -128 ? -128 : t);
+    t = (t > 127 ? 127 : t);
+    return (signed char) t;
+}
+
+
+/* should we apply any filter at all ( 11111111 yes, 00000000 no) */
+static signed char vp8_filter_mask(uc limit, uc blimit,
+                            uc p3, uc p2, uc p1, uc p0,
+                            uc q0, uc q1, uc q2, uc q3)
+{
+    signed char mask = 0;
+    mask |= (abs(p3 - p2) > limit);
+    mask |= (abs(p2 - p1) > limit);
+    mask |= (abs(p1 - p0) > limit);
+    mask |= (abs(q1 - q0) > limit);
+    mask |= (abs(q2 - q1) > limit);
+    mask |= (abs(q3 - q2) > limit);
+    mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit);
+    return mask - 1;
+}
+
+/* is there high variance internal edge ( 11111111 yes, 00000000 no) */
+static signed char vp8_hevmask(uc thresh, uc p1, uc p0, uc q0, uc q1)
+{
+    signed char hev = 0;
+    hev  |= (abs(p1 - p0) > thresh) * -1;
+    hev  |= (abs(q1 - q0) > thresh) * -1;
+    return hev;
+}
+
+static void vp8_filter(signed char mask, uc hev, uc *op1,
+        uc *op0, uc *oq0, uc *oq1)
+
+{
+    signed char ps0, qs0;
+    signed char ps1, qs1;
+    signed char filter_value, Filter1, Filter2;
+    signed char u;
+
+    ps1 = (signed char) * op1 ^ 0x80;
+    ps0 = (signed char) * op0 ^ 0x80;
+    qs0 = (signed char) * oq0 ^ 0x80;
+    qs1 = (signed char) * oq1 ^ 0x80;
+
+    /* add outer taps if we have high edge variance */
+    filter_value = vp8_signed_char_clamp(ps1 - qs1);
+    filter_value &= hev;
+
+    /* inner taps */
+    filter_value = vp8_signed_char_clamp(filter_value + 3 * (qs0 - ps0));
+    filter_value &= mask;
+
+    /* save bottom 3 bits so that we round one side +4 and the other +3
+     * if it equals 4 we'll set to adjust by -1 to account for the fact
+     * we'd round 3 the other way
+     */
+    Filter1 = vp8_signed_char_clamp(filter_value + 4);
+    Filter2 = vp8_signed_char_clamp(filter_value + 3);
+    Filter1 >>= 3;
+    Filter2 >>= 3;
+    u = vp8_signed_char_clamp(qs0 - Filter1);
+    *oq0 = u ^ 0x80;
+    u = vp8_signed_char_clamp(ps0 + Filter2);
+    *op0 = u ^ 0x80;
+    filter_value = Filter1;
+
+    /* outer tap adjustments */
+    filter_value += 1;
+    filter_value >>= 1;
+    filter_value &= ~hev;
+
+    u = vp8_signed_char_clamp(qs1 - filter_value);
+    *oq1 = u ^ 0x80;
+    u = vp8_signed_char_clamp(ps1 + filter_value);
+    *op1 = u ^ 0x80;
+
+}
+void vp8_loop_filter_horizontal_edge_c
+(
+    unsigned char *s,
+    int p, /* pitch */
+    const unsigned char *blimit,
+    const unsigned char *limit,
+    const unsigned char *thresh,
+    int count
+)
+{
+    int  hev = 0; /* high edge variance */
+    signed char mask = 0;
+    int i = 0;
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+    do
+    {
+        mask = vp8_filter_mask(limit[0], blimit[0],
+                               s[-4*p], s[-3*p], s[-2*p], s[-1*p],
+                               s[0*p], s[1*p], s[2*p], s[3*p]);
+
+        hev = vp8_hevmask(thresh[0], s[-2*p], s[-1*p], s[0*p], s[1*p]);
+
+        vp8_filter(mask, hev, s - 2 * p, s - 1 * p, s, s + 1 * p);
+
+        ++s;
+    }
+    while (++i < count * 8);
+}
+
+void vp8_loop_filter_vertical_edge_c
+(
+    unsigned char *s,
+    int p,
+    const unsigned char *blimit,
+    const unsigned char *limit,
+    const unsigned char *thresh,
+    int count
+)
+{
+    int  hev = 0; /* high edge variance */
+    signed char mask = 0;
+    int i = 0;
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+    do
+    {
+        mask = vp8_filter_mask(limit[0], blimit[0],
+                               s[-4], s[-3], s[-2], s[-1], s[0], s[1], s[2], s[3]);
+
+        hev = vp8_hevmask(thresh[0], s[-2], s[-1], s[0], s[1]);
+
+        vp8_filter(mask, hev, s - 2, s - 1, s, s + 1);
+
+        s += p;
+    }
+    while (++i < count * 8);
+}
+
+static void vp8_mbfilter(signed char mask, uc hev,
+                           uc *op2, uc *op1, uc *op0, uc *oq0, uc *oq1, uc *oq2)
+{
+    signed char s, u;
+    signed char filter_value, Filter1, Filter2;
+    signed char ps2 = (signed char) * op2 ^ 0x80;
+    signed char ps1 = (signed char) * op1 ^ 0x80;
+    signed char ps0 = (signed char) * op0 ^ 0x80;
+    signed char qs0 = (signed char) * oq0 ^ 0x80;
+    signed char qs1 = (signed char) * oq1 ^ 0x80;
+    signed char qs2 = (signed char) * oq2 ^ 0x80;
+
+    /* add outer taps if we have high edge variance */
+    filter_value = vp8_signed_char_clamp(ps1 - qs1);
+    filter_value = vp8_signed_char_clamp(filter_value + 3 * (qs0 - ps0));
+    filter_value &= mask;
+
+    Filter2 = filter_value;
+    Filter2 &= hev;
+
+    /* save bottom 3 bits so that we round one side +4 and the other +3 */
+    Filter1 = vp8_signed_char_clamp(Filter2 + 4);
+    Filter2 = vp8_signed_char_clamp(Filter2 + 3);
+    Filter1 >>= 3;
+    Filter2 >>= 3;
+    qs0 = vp8_signed_char_clamp(qs0 - Filter1);
+    ps0 = vp8_signed_char_clamp(ps0 + Filter2);
+
+
+    /* only apply wider filter if not high edge variance */
+    filter_value &= ~hev;
+    Filter2 = filter_value;
+
+    /* roughly 3/7th difference across boundary */
+    u = vp8_signed_char_clamp((63 + Filter2 * 27) >> 7);
+    s = vp8_signed_char_clamp(qs0 - u);
+    *oq0 = s ^ 0x80;
+    s = vp8_signed_char_clamp(ps0 + u);
+    *op0 = s ^ 0x80;
+
+    /* roughly 2/7th difference across boundary */
+    u = vp8_signed_char_clamp((63 + Filter2 * 18) >> 7);
+    s = vp8_signed_char_clamp(qs1 - u);
+    *oq1 = s ^ 0x80;
+    s = vp8_signed_char_clamp(ps1 + u);
+    *op1 = s ^ 0x80;
+
+    /* roughly 1/7th difference across boundary */
+    u = vp8_signed_char_clamp((63 + Filter2 * 9) >> 7);
+    s = vp8_signed_char_clamp(qs2 - u);
+    *oq2 = s ^ 0x80;
+    s = vp8_signed_char_clamp(ps2 + u);
+    *op2 = s ^ 0x80;
+}
+
+void vp8_mbloop_filter_horizontal_edge_c
+(
+    unsigned char *s,
+    int p,
+    const unsigned char *blimit,
+    const unsigned char *limit,
+    const unsigned char *thresh,
+    int count
+)
+{
+    signed char hev = 0; /* high edge variance */
+    signed char mask = 0;
+    int i = 0;
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+    do
+    {
+
+        mask = vp8_filter_mask(limit[0], blimit[0],
+                               s[-4*p], s[-3*p], s[-2*p], s[-1*p],
+                               s[0*p], s[1*p], s[2*p], s[3*p]);
+
+        hev = vp8_hevmask(thresh[0], s[-2*p], s[-1*p], s[0*p], s[1*p]);
+
+        vp8_mbfilter(mask, hev, s - 3 * p, s - 2 * p, s - 1 * p, s, s + 1 * p, s + 2 * p);
+
+        ++s;
+    }
+    while (++i < count * 8);
+
+}
+
+
+void vp8_mbloop_filter_vertical_edge_c
+(
+    unsigned char *s,
+    int p,
+    const unsigned char *blimit,
+    const unsigned char *limit,
+    const unsigned char *thresh,
+    int count
+)
+{
+    signed char hev = 0; /* high edge variance */
+    signed char mask = 0;
+    int i = 0;
+
+    do
+    {
+
+        mask = vp8_filter_mask(limit[0], blimit[0],
+                               s[-4], s[-3], s[-2], s[-1], s[0], s[1], s[2], s[3]);
+
+        hev = vp8_hevmask(thresh[0], s[-2], s[-1], s[0], s[1]);
+
+        vp8_mbfilter(mask, hev, s - 3, s - 2, s - 1, s, s + 1, s + 2);
+
+        s += p;
+    }
+    while (++i < count * 8);
+
+}
+
+/* should we apply any filter at all ( 11111111 yes, 00000000 no) */
+static signed char vp8_simple_filter_mask(uc blimit, uc p1, uc p0, uc q0, uc q1)
+{
+/* Why does this cause problems for win32?
+ * error C2143: syntax error : missing ';' before 'type'
+ *  (void) limit;
+ */
+    signed char mask = (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  <= blimit) * -1;
+    return mask;
+}
+
+static void vp8_simple_filter(signed char mask, uc *op1, uc *op0, uc *oq0, uc *oq1)
+{
+    signed char filter_value, Filter1, Filter2;
+    signed char p1 = (signed char) * op1 ^ 0x80;
+    signed char p0 = (signed char) * op0 ^ 0x80;
+    signed char q0 = (signed char) * oq0 ^ 0x80;
+    signed char q1 = (signed char) * oq1 ^ 0x80;
+    signed char u;
+
+    filter_value = vp8_signed_char_clamp(p1 - q1);
+    filter_value = vp8_signed_char_clamp(filter_value + 3 * (q0 - p0));
+    filter_value &= mask;
+
+    /* save bottom 3 bits so that we round one side +4 and the other +3 */
+    Filter1 = vp8_signed_char_clamp(filter_value + 4);
+    Filter1 >>= 3;
+    u = vp8_signed_char_clamp(q0 - Filter1);
+    *oq0  = u ^ 0x80;
+
+    Filter2 = vp8_signed_char_clamp(filter_value + 3);
+    Filter2 >>= 3;
+    u = vp8_signed_char_clamp(p0 + Filter2);
+    *op0 = u ^ 0x80;
+}
+
+void vp8_loop_filter_simple_horizontal_edge_c
+(
+    unsigned char *s,
+    int p,
+    const unsigned char *blimit
+)
+{
+    signed char mask = 0;
+    int i = 0;
+
+    do
+    {
+        mask = vp8_simple_filter_mask(blimit[0], s[-2*p], s[-1*p], s[0*p], s[1*p]);
+        vp8_simple_filter(mask, s - 2 * p, s - 1 * p, s, s + 1 * p);
+        ++s;
+    }
+    while (++i < 16);
+}
+
+void vp8_loop_filter_simple_vertical_edge_c
+(
+    unsigned char *s,
+    int p,
+    const unsigned char *blimit
+)
+{
+    signed char mask = 0;
+    int i = 0;
+
+    do
+    {
+        mask = vp8_simple_filter_mask(blimit[0], s[-2], s[-1], s[0], s[1]);
+        vp8_simple_filter(mask, s - 2, s - 1, s, s + 1);
+        s += p;
+    }
+    while (++i < 16);
+
+}
+
+/* Horizontal MB filtering */
+void vp8_loop_filter_mbh_c(unsigned char *y_ptr, unsigned char *u_ptr,
+                           unsigned char *v_ptr, int y_stride, int uv_stride,
+                           loop_filter_info *lfi)
+{
+    vp8_mbloop_filter_horizontal_edge_c(y_ptr, y_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_mbloop_filter_horizontal_edge_c(u_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_mbloop_filter_horizontal_edge_c(v_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+}
+
+/* Vertical MB Filtering */
+void vp8_loop_filter_mbv_c(unsigned char *y_ptr, unsigned char *u_ptr,
+                           unsigned char *v_ptr, int y_stride, int uv_stride,
+                           loop_filter_info *lfi)
+{
+    vp8_mbloop_filter_vertical_edge_c(y_ptr, y_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_mbloop_filter_vertical_edge_c(u_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_mbloop_filter_vertical_edge_c(v_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+}
+
+/* Horizontal B Filtering */
+void vp8_loop_filter_bh_c(unsigned char *y_ptr, unsigned char *u_ptr,
+                          unsigned char *v_ptr, int y_stride, int uv_stride,
+                          loop_filter_info *lfi)
+{
+    vp8_loop_filter_horizontal_edge_c(y_ptr + 4 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_horizontal_edge_c(y_ptr + 8 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_horizontal_edge_c(y_ptr + 12 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_loop_filter_horizontal_edge_c(u_ptr + 4 * uv_stride, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_loop_filter_horizontal_edge_c(v_ptr + 4 * uv_stride, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+}
+
+void vp8_loop_filter_bhs_c(unsigned char *y_ptr, int y_stride,
+                           const unsigned char *blimit)
+{
+    vp8_loop_filter_simple_horizontal_edge_c(y_ptr + 4 * y_stride, y_stride, blimit);
+    vp8_loop_filter_simple_horizontal_edge_c(y_ptr + 8 * y_stride, y_stride, blimit);
+    vp8_loop_filter_simple_horizontal_edge_c(y_ptr + 12 * y_stride, y_stride, blimit);
+}
+
+/* Vertical B Filtering */
+void vp8_loop_filter_bv_c(unsigned char *y_ptr, unsigned char *u_ptr,
+                          unsigned char *v_ptr, int y_stride, int uv_stride,
+                          loop_filter_info *lfi)
+{
+    vp8_loop_filter_vertical_edge_c(y_ptr + 4, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_vertical_edge_c(y_ptr + 8, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_vertical_edge_c(y_ptr + 12, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_loop_filter_vertical_edge_c(u_ptr + 4, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_loop_filter_vertical_edge_c(v_ptr + 4, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+}
+
+void vp8_loop_filter_bvs_c(unsigned char *y_ptr, int y_stride,
+                           const unsigned char *blimit)
+{
+    vp8_loop_filter_simple_vertical_edge_c(y_ptr + 4, y_stride, blimit);
+    vp8_loop_filter_simple_vertical_edge_c(y_ptr + 8, y_stride, blimit);
+    vp8_loop_filter_simple_vertical_edge_c(y_ptr + 12, y_stride, blimit);
+}
diff --git a/libvpx/libvpx/vp8/common/mbpitch.c b/libvpx/libvpx/vp8/common/mbpitch.c
new file mode 100644
index 0000000..32e1b66
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mbpitch.c
@@ -0,0 +1,68 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "blockd.h"
+
+void vp8_setup_block_dptrs(MACROBLOCKD *x)
+{
+    int r, c;
+
+    for (r = 0; r < 4; r++)
+    {
+        for (c = 0; c < 4; c++)
+        {
+            x->block[r*4+c].predictor = x->predictor + r * 4 * 16 + c * 4;
+        }
+    }
+
+    for (r = 0; r < 2; r++)
+    {
+        for (c = 0; c < 2; c++)
+        {
+            x->block[16+r*2+c].predictor = x->predictor + 256 + r * 4 * 8 + c * 4;
+
+        }
+    }
+
+    for (r = 0; r < 2; r++)
+    {
+        for (c = 0; c < 2; c++)
+        {
+            x->block[20+r*2+c].predictor = x->predictor + 320 + r * 4 * 8 + c * 4;
+
+        }
+    }
+
+    for (r = 0; r < 25; r++)
+    {
+        x->block[r].qcoeff  = x->qcoeff  + r * 16;
+        x->block[r].dqcoeff = x->dqcoeff + r * 16;
+        x->block[r].eob     = x->eobs + r;
+    }
+}
+
+void vp8_build_block_doffsets(MACROBLOCKD *x)
+{
+    int block;
+
+    for (block = 0; block < 16; block++) /* y blocks */
+    {
+        x->block[block].offset =
+            (block >> 2) * 4 * x->dst.y_stride + (block & 3) * 4;
+    }
+
+    for (block = 16; block < 20; block++) /* U and V blocks */
+    {
+        x->block[block+4].offset =
+        x->block[block].offset =
+            ((block - 16) >> 1) * 4 * x->dst.uv_stride + (block & 1) * 4;
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/mfqe.c b/libvpx/libvpx/vp8/common/mfqe.c
new file mode 100644
index 0000000..2bfefb1
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mfqe.c
@@ -0,0 +1,386 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+/* MFQE: Multiframe Quality Enhancement
+ * In rate limited situations keyframes may cause significant visual artifacts
+ * commonly referred to as "popping." This file implements a postproccesing
+ * algorithm which blends data from the preceeding frame when there is no
+ * motion and the q from the previous frame is lower which indicates that it is
+ * higher quality.
+ */
+
+#include "./vp8_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vp8/common/postproc.h"
+#include "vpx_dsp/variance.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_scale/yv12config.h"
+
+#include <limits.h>
+#include <stdlib.h>
+
+static void filter_by_weight(unsigned char *src, int src_stride,
+                             unsigned char *dst, int dst_stride,
+                             int block_size, int src_weight)
+{
+    int dst_weight = (1 << MFQE_PRECISION) - src_weight;
+    int rounding_bit = 1 << (MFQE_PRECISION - 1);
+    int r, c;
+
+    for (r = 0; r < block_size; r++)
+    {
+        for (c = 0; c < block_size; c++)
+        {
+            dst[c] = (src[c] * src_weight +
+                      dst[c] * dst_weight +
+                      rounding_bit) >> MFQE_PRECISION;
+        }
+        src += src_stride;
+        dst += dst_stride;
+    }
+}
+
+void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride,
+                                 unsigned char *dst, int dst_stride,
+                                 int src_weight)
+{
+    filter_by_weight(src, src_stride, dst, dst_stride, 16, src_weight);
+}
+
+void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride,
+                               unsigned char *dst, int dst_stride,
+                               int src_weight)
+{
+    filter_by_weight(src, src_stride, dst, dst_stride, 8, src_weight);
+}
+
+void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride,
+                               unsigned char *dst, int dst_stride,
+                               int src_weight)
+{
+    filter_by_weight(src, src_stride, dst, dst_stride, 4, src_weight);
+}
+
+static void apply_ifactor(unsigned char *y_src,
+                          int y_src_stride,
+                          unsigned char *y_dst,
+                          int y_dst_stride,
+                          unsigned char *u_src,
+                          unsigned char *v_src,
+                          int uv_src_stride,
+                          unsigned char *u_dst,
+                          unsigned char *v_dst,
+                          int uv_dst_stride,
+                          int block_size,
+                          int src_weight)
+{
+    if (block_size == 16)
+    {
+        vp8_filter_by_weight16x16(y_src, y_src_stride, y_dst, y_dst_stride, src_weight);
+        vp8_filter_by_weight8x8(u_src, uv_src_stride, u_dst, uv_dst_stride, src_weight);
+        vp8_filter_by_weight8x8(v_src, uv_src_stride, v_dst, uv_dst_stride, src_weight);
+    }
+    else /* if (block_size == 8) */
+    {
+        vp8_filter_by_weight8x8(y_src, y_src_stride, y_dst, y_dst_stride, src_weight);
+        vp8_filter_by_weight4x4(u_src, uv_src_stride, u_dst, uv_dst_stride, src_weight);
+        vp8_filter_by_weight4x4(v_src, uv_src_stride, v_dst, uv_dst_stride, src_weight);
+    }
+}
+
+static unsigned int int_sqrt(unsigned int x)
+{
+    unsigned int y = x;
+    unsigned int guess;
+    int p = 1;
+    while (y>>=1) p++;
+    p>>=1;
+
+    guess=0;
+    while (p>=0)
+    {
+        guess |= (1<<p);
+        if (x<guess*guess)
+            guess -= (1<<p);
+        p--;
+    }
+    /* choose between guess or guess+1 */
+    return guess+(guess*guess+guess+1<=x);
+}
+
+#define USE_SSD
+static void multiframe_quality_enhance_block
+(
+    int blksize, /* Currently only values supported are 16, 8 */
+    int qcurr,
+    int qprev,
+    unsigned char *y,
+    unsigned char *u,
+    unsigned char *v,
+    int y_stride,
+    int uv_stride,
+    unsigned char *yd,
+    unsigned char *ud,
+    unsigned char *vd,
+    int yd_stride,
+    int uvd_stride
+)
+{
+    static const unsigned char VP8_ZEROS[16]=
+    {
+         0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+    };
+    int uvblksize = blksize >> 1;
+    int qdiff = qcurr - qprev;
+
+    int i;
+    unsigned char *up;
+    unsigned char *udp;
+    unsigned char *vp;
+    unsigned char *vdp;
+
+    unsigned int act, actd, sad, usad, vsad, sse, thr, thrsq, actrisk;
+
+    if (blksize == 16)
+    {
+        actd = (vpx_variance16x16(yd, yd_stride, VP8_ZEROS, 0, &sse)+128)>>8;
+        act = (vpx_variance16x16(y, y_stride, VP8_ZEROS, 0, &sse)+128)>>8;
+#ifdef USE_SSD
+        vpx_variance16x16(y, y_stride, yd, yd_stride, &sse);
+        sad = (sse + 128)>>8;
+        vpx_variance8x8(u, uv_stride, ud, uvd_stride, &sse);
+        usad = (sse + 32)>>6;
+        vpx_variance8x8(v, uv_stride, vd, uvd_stride, &sse);
+        vsad = (sse + 32)>>6;
+#else
+        sad = (vpx_sad16x16(y, y_stride, yd, yd_stride) + 128) >> 8;
+        usad = (vpx_sad8x8(u, uv_stride, ud, uvd_stride) + 32) >> 6;
+        vsad = (vpx_sad8x8(v, uv_stride, vd, uvd_stride)+ 32) >> 6;
+#endif
+    }
+    else /* if (blksize == 8) */
+    {
+        actd = (vpx_variance8x8(yd, yd_stride, VP8_ZEROS, 0, &sse)+32)>>6;
+        act = (vpx_variance8x8(y, y_stride, VP8_ZEROS, 0, &sse)+32)>>6;
+#ifdef USE_SSD
+        vpx_variance8x8(y, y_stride, yd, yd_stride, &sse);
+        sad = (sse + 32)>>6;
+        vpx_variance4x4(u, uv_stride, ud, uvd_stride, &sse);
+        usad = (sse + 8)>>4;
+        vpx_variance4x4(v, uv_stride, vd, uvd_stride, &sse);
+        vsad = (sse + 8)>>4;
+#else
+        sad = (vpx_sad8x8(y, y_stride, yd, yd_stride) + 32) >> 6;
+        usad = (vpx_sad4x4(u, uv_stride, ud, uvd_stride) + 8) >> 4;
+        vsad = (vpx_sad4x4(v, uv_stride, vd, uvd_stride) + 8) >> 4;
+#endif
+    }
+
+    actrisk = (actd > act * 5);
+
+    /* thr = qdiff/16 + log2(act) + log4(qprev) */
+    thr = (qdiff >> 4);
+    while (actd >>= 1) thr++;
+    while (qprev >>= 2) thr++;
+
+#ifdef USE_SSD
+    thrsq = thr * thr;
+    if (sad < thrsq &&
+        /* additional checks for color mismatch and excessive addition of
+         * high-frequencies */
+        4 * usad < thrsq && 4 * vsad < thrsq && !actrisk)
+#else
+    if (sad < thr &&
+        /* additional checks for color mismatch and excessive addition of
+         * high-frequencies */
+        2 * usad < thr && 2 * vsad < thr && !actrisk)
+#endif
+    {
+        int ifactor;
+#ifdef USE_SSD
+        /* TODO: optimize this later to not need sqr root */
+        sad = int_sqrt(sad);
+#endif
+        ifactor = (sad << MFQE_PRECISION) / thr;
+        ifactor >>= (qdiff >> 5);
+
+        if (ifactor)
+        {
+            apply_ifactor(y, y_stride, yd, yd_stride,
+                          u, v, uv_stride,
+                          ud, vd, uvd_stride,
+                          blksize, ifactor);
+        }
+    }
+    else  /* else implicitly copy from previous frame */
+    {
+        if (blksize == 16)
+        {
+            vp8_copy_mem16x16(y, y_stride, yd, yd_stride);
+            vp8_copy_mem8x8(u, uv_stride, ud, uvd_stride);
+            vp8_copy_mem8x8(v, uv_stride, vd, uvd_stride);
+        }
+        else  /* if (blksize == 8) */
+        {
+            vp8_copy_mem8x8(y, y_stride, yd, yd_stride);
+            for (up = u, udp = ud, i = 0; i < uvblksize; ++i, up += uv_stride, udp += uvd_stride)
+                memcpy(udp, up, uvblksize);
+            for (vp = v, vdp = vd, i = 0; i < uvblksize; ++i, vp += uv_stride, vdp += uvd_stride)
+                memcpy(vdp, vp, uvblksize);
+        }
+    }
+}
+
+static int qualify_inter_mb(const MODE_INFO *mode_info_context, int *map)
+{
+    if (mode_info_context->mbmi.mb_skip_coeff)
+        map[0] = map[1] = map[2] = map[3] = 1;
+    else if (mode_info_context->mbmi.mode==SPLITMV)
+    {
+        static int ndx[4][4] =
+        {
+            {0, 1, 4, 5},
+            {2, 3, 6, 7},
+            {8, 9, 12, 13},
+            {10, 11, 14, 15}
+        };
+        int i, j;
+        for (i=0; i<4; ++i)
+        {
+            map[i] = 1;
+            for (j=0; j<4 && map[j]; ++j)
+                map[i] &= (mode_info_context->bmi[ndx[i][j]].mv.as_mv.row <= 2 &&
+                           mode_info_context->bmi[ndx[i][j]].mv.as_mv.col <= 2);
+        }
+    }
+    else
+    {
+        map[0] = map[1] = map[2] = map[3] =
+            (mode_info_context->mbmi.mode > B_PRED &&
+             abs(mode_info_context->mbmi.mv.as_mv.row) <= 2 &&
+             abs(mode_info_context->mbmi.mv.as_mv.col) <= 2);
+    }
+    return (map[0]+map[1]+map[2]+map[3]);
+}
+
+void vp8_multiframe_quality_enhance
+(
+    VP8_COMMON *cm
+)
+{
+    YV12_BUFFER_CONFIG *show = cm->frame_to_show;
+    YV12_BUFFER_CONFIG *dest = &cm->post_proc_buffer;
+
+    FRAME_TYPE frame_type = cm->frame_type;
+    /* Point at base of Mb MODE_INFO list has motion vectors etc */
+    const MODE_INFO *mode_info_context = cm->show_frame_mi;
+    int mb_row;
+    int mb_col;
+    int totmap, map[4];
+    int qcurr = cm->base_qindex;
+    int qprev = cm->postproc_state.last_base_qindex;
+
+    unsigned char *y_ptr, *u_ptr, *v_ptr;
+    unsigned char *yd_ptr, *ud_ptr, *vd_ptr;
+
+    /* Set up the buffer pointers */
+    y_ptr = show->y_buffer;
+    u_ptr = show->u_buffer;
+    v_ptr = show->v_buffer;
+    yd_ptr = dest->y_buffer;
+    ud_ptr = dest->u_buffer;
+    vd_ptr = dest->v_buffer;
+
+    /* postprocess each macro block */
+    for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
+    {
+        for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
+        {
+            /* if motion is high there will likely be no benefit */
+            if (frame_type == INTER_FRAME) totmap = qualify_inter_mb(mode_info_context, map);
+            else totmap = (frame_type == KEY_FRAME ? 4 : 0);
+            if (totmap)
+            {
+                if (totmap < 4)
+                {
+                    int i, j;
+                    for (i=0; i<2; ++i)
+                        for (j=0; j<2; ++j)
+                        {
+                            if (map[i*2+j])
+                            {
+                                multiframe_quality_enhance_block(8, qcurr, qprev,
+                                                                 y_ptr + 8*(i*show->y_stride+j),
+                                                                 u_ptr + 4*(i*show->uv_stride+j),
+                                                                 v_ptr + 4*(i*show->uv_stride+j),
+                                                                 show->y_stride,
+                                                                 show->uv_stride,
+                                                                 yd_ptr + 8*(i*dest->y_stride+j),
+                                                                 ud_ptr + 4*(i*dest->uv_stride+j),
+                                                                 vd_ptr + 4*(i*dest->uv_stride+j),
+                                                                 dest->y_stride,
+                                                                 dest->uv_stride);
+                            }
+                            else
+                            {
+                                /* copy a 8x8 block */
+                                int k;
+                                unsigned char *up = u_ptr + 4*(i*show->uv_stride+j);
+                                unsigned char *udp = ud_ptr + 4*(i*dest->uv_stride+j);
+                                unsigned char *vp = v_ptr + 4*(i*show->uv_stride+j);
+                                unsigned char *vdp = vd_ptr + 4*(i*dest->uv_stride+j);
+                                vp8_copy_mem8x8(y_ptr + 8*(i*show->y_stride+j), show->y_stride,
+                                                yd_ptr + 8*(i*dest->y_stride+j), dest->y_stride);
+                                for (k = 0; k < 4; ++k, up += show->uv_stride, udp += dest->uv_stride,
+                                                        vp += show->uv_stride, vdp += dest->uv_stride)
+                                {
+                                    memcpy(udp, up, 4);
+                                    memcpy(vdp, vp, 4);
+                                }
+                            }
+                        }
+                }
+                else /* totmap = 4 */
+                {
+                    multiframe_quality_enhance_block(16, qcurr, qprev, y_ptr,
+                                                     u_ptr, v_ptr,
+                                                     show->y_stride,
+                                                     show->uv_stride,
+                                                     yd_ptr, ud_ptr, vd_ptr,
+                                                     dest->y_stride,
+                                                     dest->uv_stride);
+                }
+            }
+            else
+            {
+                vp8_copy_mem16x16(y_ptr, show->y_stride, yd_ptr, dest->y_stride);
+                vp8_copy_mem8x8(u_ptr, show->uv_stride, ud_ptr, dest->uv_stride);
+                vp8_copy_mem8x8(v_ptr, show->uv_stride, vd_ptr, dest->uv_stride);
+            }
+            y_ptr += 16;
+            u_ptr += 8;
+            v_ptr += 8;
+            yd_ptr += 16;
+            ud_ptr += 8;
+            vd_ptr += 8;
+            mode_info_context++;     /* step to next MB */
+        }
+
+        y_ptr += show->y_stride  * 16 - 16 * cm->mb_cols;
+        u_ptr += show->uv_stride *  8 - 8 * cm->mb_cols;
+        v_ptr += show->uv_stride *  8 - 8 * cm->mb_cols;
+        yd_ptr += dest->y_stride  * 16 - 16 * cm->mb_cols;
+        ud_ptr += dest->uv_stride *  8 - 8 * cm->mb_cols;
+        vd_ptr += dest->uv_stride *  8 - 8 * cm->mb_cols;
+
+        mode_info_context++;         /* Skip border mb */
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/mips/dspr2/dequantize_dspr2.c b/libvpx/libvpx/vp8/common/mips/dspr2/dequantize_dspr2.c
new file mode 100644
index 0000000..fc3bb8a
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/dspr2/dequantize_dspr2.c
@@ -0,0 +1,33 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vpx_mem/vpx_mem.h"
+
+#if HAVE_DSPR2
+void vp8_dequant_idct_add_dspr2(short *input, short *dq,
+                                unsigned char *dest, int stride)
+{
+    int i;
+
+    for (i = 0; i < 16; i++)
+    {
+        input[i] = dq[i] * input[i];
+    }
+
+    vp8_short_idct4x4llm_dspr2(input, dest, stride, dest, stride);
+
+    memset(input, 0, 32);
+
+}
+
+#endif
diff --git a/libvpx/libvpx/vp8/common/mips/dspr2/filter_dspr2.c b/libvpx/libvpx/vp8/common/mips/dspr2/filter_dspr2.c
new file mode 100644
index 0000000..ace5d40
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/dspr2/filter_dspr2.c
@@ -0,0 +1,2823 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <stdlib.h>
+#include "vp8_rtcd.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+#define CROP_WIDTH 256
+unsigned char ff_cropTbl[256 + 2 * CROP_WIDTH];
+
+static const unsigned short sub_pel_filterss[8][3] =
+{
+    {      0,      0,      0},
+    {      0, 0x0601, 0x7b0c},
+    { 0x0201, 0x0b08, 0x6c24},
+    {      0, 0x0906, 0x5d32},
+    { 0x0303, 0x1010, 0x4d4d},
+    {      0, 0x0609, 0x325d},
+    { 0x0102, 0x080b, 0x246c},
+    {      0, 0x0106, 0x0c7b},
+};
+
+
+static const int sub_pel_filters_int[8][3] =
+{
+    {          0,          0,          0},
+    { 0x0000fffa, 0x007b000c, 0xffff0000},
+    { 0x0002fff5, 0x006c0024, 0xfff80001},
+    { 0x0000fff7, 0x005d0032, 0xfffa0000},
+    { 0x0003fff0, 0x004d004d, 0xfff00003},
+    { 0x0000fffa, 0x0032005d, 0xfff70000},
+    { 0x0001fff8, 0x0024006c, 0xfff50002},
+    { 0x0000ffff, 0x000c007b, 0xfffa0000},
+};
+
+
+static const int sub_pel_filters_inv[8][3] =
+{
+    {          0,          0,          0},
+    { 0xfffa0000, 0x000c007b, 0x0000ffff},
+    { 0xfff50002, 0x0024006c, 0x0001fff8},
+    { 0xfff70000, 0x0032005d, 0x0000fffa},
+    { 0xfff00003, 0x004d004d, 0x0003fff0},
+    { 0xfffa0000, 0x005d0032, 0x0000fff7},
+    { 0xfff80001, 0x006c0024, 0x0002fff5},
+    { 0xffff0000, 0x007b000c, 0x0000fffa},
+};
+
+
+static const int sub_pel_filters_int_tap_4[8][2] =
+{
+    {          0,          0},
+    { 0xfffa007b, 0x000cffff},
+    {          0,          0},
+    { 0xfff7005d, 0x0032fffa},
+    {          0,          0},
+    { 0xfffa0032, 0x005dfff7},
+    {          0,          0},
+    { 0xffff000c, 0x007bfffa},
+};
+
+
+static const int sub_pel_filters_inv_tap_4[8][2] =
+{
+    {          0,          0},
+    { 0x007bfffa, 0xffff000c},
+    {          0,          0},
+    { 0x005dfff7, 0xfffa0032},
+    {          0,          0},
+    { 0x0032fffa, 0xfff7005d},
+    {          0,          0},
+    { 0x000cffff, 0xfffa007b},
+};
+
+inline void prefetch_load(unsigned char *src)
+{
+    __asm__ __volatile__ (
+        "pref   0,  0(%[src])   \n\t"
+        :
+        : [src] "r" (src)
+    );
+}
+
+
+inline void prefetch_store(unsigned char *dst)
+{
+    __asm__ __volatile__ (
+        "pref   1,  0(%[dst])   \n\t"
+        :
+        : [dst] "r" (dst)
+    );
+}
+
+void dsputil_static_init(void)
+{
+    int i;
+
+    for (i = 0; i < 256; i++) ff_cropTbl[i + CROP_WIDTH] = i;
+
+    for (i = 0; i < CROP_WIDTH; i++)
+    {
+        ff_cropTbl[i] = 0;
+        ff_cropTbl[i + CROP_WIDTH + 256] = 255;
+    }
+}
+
+void vp8_filter_block2d_first_pass_4
+(
+    unsigned char *RESTRICT src_ptr,
+    unsigned char *RESTRICT dst_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int output_height,
+    int xoffset,
+    int pitch
+)
+{
+    unsigned int i;
+    int Temp1, Temp2, Temp3, Temp4;
+
+    unsigned int vector4a = 64;
+    int vector1b, vector2b, vector3b;
+    unsigned int tp1, tp2, tn1, tn2;
+    unsigned int p1, p2, p3;
+    unsigned int n1, n2, n3;
+    unsigned char *cm = ff_cropTbl + CROP_WIDTH;
+
+    vector3b = sub_pel_filters_inv[xoffset][2];
+
+    /* if (xoffset == 0) we don't need any filtering */
+    if (vector3b == 0)
+    {
+        for (i = 0; i < output_height; i++)
+        {
+            /* prefetch src_ptr data to cache memory */
+            prefetch_load(src_ptr + src_pixels_per_line);
+            dst_ptr[0] = src_ptr[0];
+            dst_ptr[1] = src_ptr[1];
+            dst_ptr[2] = src_ptr[2];
+            dst_ptr[3] = src_ptr[3];
+
+            /* next row... */
+            src_ptr += src_pixels_per_line;
+            dst_ptr += 4;
+        }
+    }
+    else
+    {
+        if (vector3b > 65536)
+        {
+            /* 6 tap filter */
+
+            vector1b = sub_pel_filters_inv[xoffset][0];
+            vector2b = sub_pel_filters_inv[xoffset][1];
+
+            /* prefetch src_ptr data to cache memory */
+            prefetch_load(src_ptr + src_pixels_per_line);
+
+            for (i = output_height; i--;)
+            {
+                /* apply filter with vectors pairs */
+                __asm__ __volatile__ (
+                    "ulw              %[tp1],      -2(%[src_ptr])                 \n\t"
+                    "ulw              %[tp2],      2(%[src_ptr])                  \n\t"
+
+                    /* even 1. pixel */
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+                    "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+                    "preceu.ph.qbr    %[p3],       %[tp2]                         \n\t"
+                    "dpa.w.ph         $ac3,        %[p1],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[p2],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[p3],          %[vector3b]    \n\t"
+
+                    /* even 2. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "preceu.ph.qbl    %[p1],       %[tp2]                         \n\t"
+                    "balign           %[tp2],      %[tp1],         3              \n\t"
+                    "extp             %[Temp1],    $ac3,           9              \n\t"
+                    "dpa.w.ph         $ac2,        %[p2],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[p3],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[p1],          %[vector3b]    \n\t"
+
+                    /* odd 1. pixel */
+                    "ulw              %[tn2],      3(%[src_ptr])                  \n\t"
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "preceu.ph.qbr    %[n1],       %[tp2]                         \n\t"
+                    "preceu.ph.qbl    %[n2],       %[tp2]                         \n\t"
+                    "preceu.ph.qbr    %[n3],       %[tn2]                         \n\t"
+                    "extp             %[Temp3],    $ac2,           9              \n\t"
+                    "dpa.w.ph         $ac3,        %[n1],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[n2],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[n3],          %[vector3b]    \n\t"
+
+                    /* even 2. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "preceu.ph.qbl    %[n1],       %[tn2]                         \n\t"
+                    "extp             %[Temp2],    $ac3,           9              \n\t"
+                    "dpa.w.ph         $ac2,        %[n2],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[n3],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[n1],          %[vector3b]    \n\t"
+                    "extp             %[Temp4],    $ac2,           9              \n\t"
+
+                    /* clamp */
+                    "lbux             %[tp1],      %[Temp1](%[cm])                \n\t"
+                    "lbux             %[tn1],      %[Temp2](%[cm])                \n\t"
+                    "lbux             %[tp2],      %[Temp3](%[cm])                \n\t"
+                    "lbux             %[n2],       %[Temp4](%[cm])                \n\t"
+
+                    /* store bytes */
+                    "sb               %[tp1],      0(%[dst_ptr])                  \n\t"
+                    "sb               %[tn1],      1(%[dst_ptr])                  \n\t"
+                    "sb               %[tp2],      2(%[dst_ptr])                  \n\t"
+                    "sb               %[n2],       3(%[dst_ptr])                  \n\t"
+
+                    : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2), [tn1] "=&r" (tn1),
+                      [tn2] "=&r" (tn2), [p1] "=&r" (p1), [p2] "=&r" (p2),
+                      [p3] "=&r" (p3), [n1] "=&r" (n1), [n2] "=&r" (n2),
+                      [n3] "=&r" (n3), [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                      [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4)
+                    : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                      [vector4a] "r" (vector4a), [cm] "r" (cm), [dst_ptr] "r" (dst_ptr),
+                      [vector3b] "r" (vector3b), [src_ptr] "r" (src_ptr)
+                );
+
+                /* Next row... */
+                src_ptr += src_pixels_per_line;
+                dst_ptr += pitch;
+            }
+        }
+        else
+        {
+            /* 4 tap filter */
+
+            vector1b = sub_pel_filters_inv_tap_4[xoffset][0];
+            vector2b = sub_pel_filters_inv_tap_4[xoffset][1];
+
+            for (i = output_height; i--;)
+            {
+                /* apply filter with vectors pairs */
+                __asm__ __volatile__ (
+                    "ulw              %[tp1],      -1(%[src_ptr])                 \n\t"
+                    "ulw              %[tp2],      3(%[src_ptr])                  \n\t"
+
+                    /* even 1. pixel */
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+                    "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+                    "preceu.ph.qbr    %[p3],       %[tp2]                         \n\t"
+                    "dpa.w.ph         $ac3,        %[p1],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[p2],          %[vector2b]    \n\t"
+
+                    /* even 2. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "dpa.w.ph         $ac2,        %[p2],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[p3],          %[vector2b]    \n\t"
+                    "extp             %[Temp1],    $ac3,           9              \n\t"
+
+                    /* odd 1. pixel */
+                    "srl              %[tn1],      %[tp2],         8              \n\t"
+                    "balign           %[tp2],      %[tp1],         3              \n\t"
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "preceu.ph.qbr    %[n1],       %[tp2]                         \n\t"
+                    "preceu.ph.qbl    %[n2],       %[tp2]                         \n\t"
+                    "preceu.ph.qbr    %[n3],       %[tn1]                         \n\t"
+                    "extp             %[Temp3],    $ac2,           9              \n\t"
+                    "dpa.w.ph         $ac3,        %[n1],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[n2],          %[vector2b]    \n\t"
+
+                    /* odd 2. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "extp             %[Temp2],    $ac3,           9              \n\t"
+                    "dpa.w.ph         $ac2,        %[n2],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[n3],          %[vector2b]    \n\t"
+                    "extp             %[Temp4],    $ac2,           9              \n\t"
+
+                    /* clamp and store results */
+                    "lbux             %[tp1],      %[Temp1](%[cm])                \n\t"
+                    "lbux             %[tn1],      %[Temp2](%[cm])                \n\t"
+                    "lbux             %[tp2],      %[Temp3](%[cm])                \n\t"
+                    "sb               %[tp1],      0(%[dst_ptr])                  \n\t"
+                    "sb               %[tn1],      1(%[dst_ptr])                  \n\t"
+                    "lbux             %[n2],       %[Temp4](%[cm])                \n\t"
+                    "sb               %[tp2],      2(%[dst_ptr])                  \n\t"
+                    "sb               %[n2],       3(%[dst_ptr])                  \n\t"
+
+                    : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2), [tn1] "=&r" (tn1),
+                      [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3),
+                      [n1] "=&r" (n1), [n2] "=&r" (n2), [n3] "=&r" (n3),
+                      [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                      [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4)
+                    : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                      [vector4a] "r" (vector4a), [cm] "r" (cm), [dst_ptr] "r" (dst_ptr),
+                      [src_ptr] "r" (src_ptr)
+                );
+                /*  Next row... */
+                src_ptr += src_pixels_per_line;
+                dst_ptr += pitch;
+            }
+        }
+    }
+}
+
+void vp8_filter_block2d_first_pass_8_all
+(
+    unsigned char *RESTRICT src_ptr,
+    unsigned char *RESTRICT dst_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int output_height,
+    int xoffset,
+    int pitch
+)
+{
+    unsigned int i;
+    int Temp1, Temp2, Temp3, Temp4;
+
+    unsigned int vector4a = 64;
+    unsigned int vector1b, vector2b, vector3b;
+    unsigned int tp1, tp2, tn1, tn2;
+    unsigned int p1, p2, p3, p4;
+    unsigned int n1, n2, n3, n4;
+
+    unsigned char *cm = ff_cropTbl + CROP_WIDTH;
+
+    /* if (xoffset == 0) we don't need any filtering */
+    if (xoffset == 0)
+    {
+        for (i = 0; i < output_height; i++)
+        {
+            /* prefetch src_ptr data to cache memory */
+            prefetch_load(src_ptr + src_pixels_per_line);
+
+            dst_ptr[0] = src_ptr[0];
+            dst_ptr[1] = src_ptr[1];
+            dst_ptr[2] = src_ptr[2];
+            dst_ptr[3] = src_ptr[3];
+            dst_ptr[4] = src_ptr[4];
+            dst_ptr[5] = src_ptr[5];
+            dst_ptr[6] = src_ptr[6];
+            dst_ptr[7] = src_ptr[7];
+
+            /* next row... */
+            src_ptr += src_pixels_per_line;
+            dst_ptr += 8;
+        }
+    }
+    else
+    {
+        vector3b = sub_pel_filters_inv[xoffset][2];
+
+        if (vector3b > 65536)
+        {
+            /* 6 tap filter */
+
+            vector1b = sub_pel_filters_inv[xoffset][0];
+            vector2b = sub_pel_filters_inv[xoffset][1];
+
+            for (i = output_height; i--;)
+            {
+                /* prefetch src_ptr data to cache memory */
+                prefetch_load(src_ptr + src_pixels_per_line);
+
+                /* apply filter with vectors pairs */
+                __asm__ __volatile__ (
+                    "ulw              %[tp1],      -2(%[src_ptr])                 \n\t"
+                    "ulw              %[tp2],      2(%[src_ptr])                  \n\t"
+
+                    /* even 1. pixel */
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+                    "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+                    "preceu.ph.qbr    %[p3],       %[tp2]                         \n\t"
+                    "dpa.w.ph         $ac3,        %[p1],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[p2],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[p3],          %[vector3b]    \n\t"
+
+                    /* even 2. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "preceu.ph.qbl    %[p1],       %[tp2]                         \n\t"
+                    "dpa.w.ph         $ac2,        %[p2],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[p3],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[p1],          %[vector3b]    \n\t"
+
+                    "balign           %[tp2],      %[tp1],         3              \n\t"
+                    "extp             %[Temp1],    $ac3,           9              \n\t"
+                    "ulw              %[tn2],      3(%[src_ptr])                  \n\t"
+
+                    /* odd 1. pixel */
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "preceu.ph.qbr    %[n1],       %[tp2]                         \n\t"
+                    "preceu.ph.qbl    %[n2],       %[tp2]                         \n\t"
+                    "preceu.ph.qbr    %[n3],       %[tn2]                         \n\t"
+                    "extp             %[Temp3],    $ac2,           9              \n\t"
+                    "dpa.w.ph         $ac3,        %[n1],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[n2],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[n3],          %[vector3b]    \n\t"
+
+                    /* odd 2. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "preceu.ph.qbl    %[n1],       %[tn2]                         \n\t"
+                    "dpa.w.ph         $ac2,        %[n2],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[n3],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[n1],          %[vector3b]    \n\t"
+                    "ulw              %[tp1],      6(%[src_ptr])                  \n\t"
+                    "extp             %[Temp2],    $ac3,           9              \n\t"
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "preceu.ph.qbr    %[p2],       %[tp1]                         \n\t"
+                    "extp             %[Temp4],    $ac2,           9              \n\t"
+
+                    : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2), [tn2] "=&r" (tn2),
+                      [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3),
+                      [n1] "=&r" (n1), [n2] "=&r" (n2), [n3] "=&r" (n3),
+                      [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                      [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4)
+                    : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                      [vector4a] "r" (vector4a), [vector3b] "r" (vector3b),
+                      [src_ptr] "r" (src_ptr)
+                );
+
+                /* clamp and store results */
+                dst_ptr[0] = cm[Temp1];
+                dst_ptr[1] = cm[Temp2];
+                dst_ptr[2] = cm[Temp3];
+                dst_ptr[3] = cm[Temp4];
+
+                /* next 4 pixels */
+                __asm__ __volatile__ (
+                    /* even 3. pixel */
+                    "dpa.w.ph         $ac3,        %[p3],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[p1],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[p2],          %[vector3b]    \n\t"
+
+                    /* even 4. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "preceu.ph.qbl    %[p4],       %[tp1]                         \n\t"
+                    "dpa.w.ph         $ac2,        %[p1],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[p2],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[p4],          %[vector3b]    \n\t"
+
+                    "ulw              %[tn1],      7(%[src_ptr])                  \n\t"
+                    "extp             %[Temp1],    $ac3,           9              \n\t"
+
+                    /* odd 3. pixel */
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "preceu.ph.qbr    %[n2],       %[tn1]                         \n\t"
+                    "dpa.w.ph         $ac3,        %[n3],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[n1],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[n2],          %[vector3b]    \n\t"
+                    "extp             %[Temp3],    $ac2,           9              \n\t"
+
+                    /* odd 4. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "preceu.ph.qbl    %[n4],       %[tn1]                         \n\t"
+                    "dpa.w.ph         $ac2,        %[n1],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[n2],          %[vector2b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[n4],          %[vector3b]    \n\t"
+                    "extp             %[Temp2],    $ac3,           9              \n\t"
+                    "extp             %[Temp4],    $ac2,           9              \n\t"
+
+                    : [tn1] "=&r" (tn1), [n2] "=&r" (n2),
+                      [p4] "=&r" (p4), [n4] "=&r" (n4),
+                      [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                      [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4)
+                    : [tp1] "r" (tp1), [vector1b] "r" (vector1b), [p2] "r" (p2),
+                      [vector2b] "r" (vector2b), [n1] "r" (n1), [p1] "r" (p1),
+                      [vector4a] "r" (vector4a), [vector3b] "r" (vector3b),
+                      [p3] "r" (p3), [n3] "r" (n3), [src_ptr] "r" (src_ptr)
+                );
+
+                /* clamp and store results */
+                dst_ptr[4] = cm[Temp1];
+                dst_ptr[5] = cm[Temp2];
+                dst_ptr[6] = cm[Temp3];
+                dst_ptr[7] = cm[Temp4];
+
+                src_ptr += src_pixels_per_line;
+                dst_ptr += pitch;
+            }
+        }
+        else
+        {
+            /* 4 tap filter */
+
+            vector1b = sub_pel_filters_inv_tap_4[xoffset][0];
+            vector2b = sub_pel_filters_inv_tap_4[xoffset][1];
+
+            for (i = output_height; i--;)
+            {
+                /* prefetch src_ptr data to cache memory */
+                prefetch_load(src_ptr + src_pixels_per_line);
+
+                /* apply filter with vectors pairs */
+                __asm__ __volatile__ (
+                    "ulw              %[tp1],      -1(%[src_ptr])                 \n\t"
+
+                    /* even 1. pixel */
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+                    "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+                    "dpa.w.ph         $ac3,        %[p1],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[p2],          %[vector2b]    \n\t"
+
+                    "ulw              %[tp2],      3(%[src_ptr])                  \n\t"
+
+                    /* even 2. pixel  */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "preceu.ph.qbr    %[p3],       %[tp2]                         \n\t"
+                    "preceu.ph.qbl    %[p4],       %[tp2]                         \n\t"
+                    "dpa.w.ph         $ac2,        %[p2],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[p3],          %[vector2b]    \n\t"
+                    "extp             %[Temp1],    $ac3,           9              \n\t"
+
+                    "balign           %[tp2],      %[tp1],         3              \n\t"
+
+                    /* odd 1. pixel */
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "preceu.ph.qbr    %[n1],       %[tp2]                         \n\t"
+                    "preceu.ph.qbl    %[n2],       %[tp2]                         \n\t"
+                    "dpa.w.ph         $ac3,        %[n1],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[n2],          %[vector2b]    \n\t"
+                    "extp             %[Temp3],    $ac2,           9              \n\t"
+
+                    "ulw              %[tn2],      4(%[src_ptr])                  \n\t"
+
+                    /* odd 2. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "preceu.ph.qbr    %[n3],       %[tn2]                         \n\t"
+                    "preceu.ph.qbl    %[n4],       %[tn2]                         \n\t"
+                    "dpa.w.ph         $ac2,        %[n2],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[n3],          %[vector2b]    \n\t"
+                    "ulw              %[tp1],      7(%[src_ptr])                  \n\t"
+                    "extp             %[Temp2],    $ac3,           9              \n\t"
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "extp             %[Temp4],    $ac2,           9              \n\t"
+
+                    : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+                      [tn2] "=&r" (tn2), [p1] "=&r" (p1), [p2] "=&r" (p2),
+                      [p3] "=&r" (p3), [p4] "=&r" (p4), [n1] "=&r" (n1),
+                      [n2] "=&r" (n2), [n3] "=&r" (n3), [n4] "=&r" (n4),
+                      [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                      [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4)
+                    : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                      [vector4a] "r" (vector4a), [src_ptr] "r" (src_ptr)
+                );
+
+                /* clamp and store results */
+                dst_ptr[0] = cm[Temp1];
+                dst_ptr[1] = cm[Temp2];
+                dst_ptr[2] = cm[Temp3];
+                dst_ptr[3] = cm[Temp4];
+
+                /* next 4 pixels */
+                __asm__ __volatile__ (
+                    /* even 3. pixel */
+                    "dpa.w.ph         $ac3,        %[p3],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[p4],          %[vector2b]    \n\t"
+
+                    /* even 4. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "preceu.ph.qbr    %[p2],       %[tp1]                         \n\t"
+                    "dpa.w.ph         $ac2,        %[p4],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[p2],          %[vector2b]    \n\t"
+                    "extp             %[Temp1],    $ac3,           9              \n\t"
+
+                    /* odd 3. pixel */
+                    "mtlo             %[vector4a], $ac3                           \n\t"
+                    "dpa.w.ph         $ac3,        %[n3],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac3,        %[n4],          %[vector2b]    \n\t"
+                    "ulw              %[tn1],      8(%[src_ptr])                  \n\t"
+                    "extp             %[Temp3],    $ac2,           9              \n\t"
+
+                    /* odd 4. pixel */
+                    "mtlo             %[vector4a], $ac2                           \n\t"
+                    "preceu.ph.qbr    %[n2],       %[tn1]                         \n\t"
+                    "dpa.w.ph         $ac2,        %[n4],          %[vector1b]    \n\t"
+                    "dpa.w.ph         $ac2,        %[n2],          %[vector2b]    \n\t"
+                    "extp             %[Temp2],    $ac3,           9              \n\t"
+                    "extp             %[Temp4],    $ac2,           9              \n\t"
+
+                    : [tn1] "=&r" (tn1), [p2] "=&r" (p2), [n2] "=&r" (n2),
+                      [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                      [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4)
+                    : [tp1] "r" (tp1), [p3] "r" (p3), [p4] "r" (p4),
+                      [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                      [vector4a] "r" (vector4a), [src_ptr] "r" (src_ptr),
+                      [n3] "r" (n3), [n4] "r" (n4)
+                );
+
+                /* clamp and store results */
+                dst_ptr[4] = cm[Temp1];
+                dst_ptr[5] = cm[Temp2];
+                dst_ptr[6] = cm[Temp3];
+                dst_ptr[7] = cm[Temp4];
+
+                /* next row... */
+                src_ptr += src_pixels_per_line;
+                dst_ptr += pitch;
+            }
+        }
+    }
+}
+
+
+void vp8_filter_block2d_first_pass16_6tap
+(
+    unsigned char *RESTRICT src_ptr,
+    unsigned char *RESTRICT dst_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int output_height,
+    int xoffset,
+    int pitch
+)
+{
+    unsigned int i;
+    int Temp1, Temp2, Temp3, Temp4;
+
+    unsigned int vector4a;
+    unsigned int vector1b, vector2b, vector3b;
+    unsigned int tp1, tp2, tn1, tn2;
+    unsigned int p1, p2, p3, p4;
+    unsigned int n1, n2, n3, n4;
+    unsigned char *cm = ff_cropTbl + CROP_WIDTH;
+
+    vector1b = sub_pel_filters_inv[xoffset][0];
+    vector2b = sub_pel_filters_inv[xoffset][1];
+    vector3b = sub_pel_filters_inv[xoffset][2];
+    vector4a = 64;
+
+    for (i = output_height; i--;)
+    {
+        /* prefetch src_ptr data to cache memory */
+        prefetch_load(src_ptr + src_pixels_per_line);
+
+        /* apply filter with vectors pairs */
+        __asm__ __volatile__ (
+            "ulw                %[tp1],      -2(%[src_ptr])                 \n\t"
+            "ulw                %[tp2],      2(%[src_ptr])                  \n\t"
+
+            /* even 1. pixel */
+            "mtlo               %[vector4a], $ac3                           \n\t"
+            "preceu.ph.qbr      %[p1],       %[tp1]                         \n\t"
+            "preceu.ph.qbl      %[p2],       %[tp1]                         \n\t"
+            "preceu.ph.qbr      %[p3],       %[tp2]                         \n\t"
+            "dpa.w.ph           $ac3,        %[p1],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac3,        %[p2],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac3,        %[p3],           %[vector3b]   \n\t"
+
+            /* even 2. pixel */
+            "mtlo               %[vector4a], $ac2                           \n\t"
+            "preceu.ph.qbl      %[p1],       %[tp2]                         \n\t"
+            "dpa.w.ph           $ac2,        %[p2],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac2,        %[p3],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac2,        %[p1],           %[vector3b]   \n\t"
+
+            "balign             %[tp2],      %[tp1],          3             \n\t"
+            "ulw                %[tn2],      3(%[src_ptr])                  \n\t"
+            "extp               %[Temp1],    $ac3,            9             \n\t"
+
+            /* odd 1. pixel */
+            "mtlo               %[vector4a], $ac3                           \n\t"
+            "preceu.ph.qbr      %[n1],       %[tp2]                         \n\t"
+            "preceu.ph.qbl      %[n2],       %[tp2]                         \n\t"
+            "preceu.ph.qbr      %[n3],       %[tn2]                         \n\t"
+            "extp               %[Temp3],    $ac2,            9             \n\t"
+            "dpa.w.ph           $ac3,        %[n1],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac3,        %[n2],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac3,        %[n3],           %[vector3b]   \n\t"
+
+            /* odd 2. pixel */
+            "mtlo               %[vector4a], $ac2                           \n\t"
+            "preceu.ph.qbl      %[n1],       %[tn2]                         \n\t"
+            "dpa.w.ph           $ac2,        %[n2],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac2,        %[n3],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac2,        %[n1],           %[vector3b]   \n\t"
+            "ulw                %[tp1],      6(%[src_ptr])                  \n\t"
+            "extp               %[Temp2],    $ac3,            9             \n\t"
+            "mtlo               %[vector4a], $ac3                           \n\t"
+            "preceu.ph.qbr      %[p2],       %[tp1]                         \n\t"
+            "extp               %[Temp4],    $ac2,            9             \n\t"
+
+            : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2), [tn2] "=&r" (tn2),
+              [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3),
+              [n1] "=&r" (n1), [n2] "=&r" (n2), [n3] "=&r" (n3),
+              [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+              [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4)
+            : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+              [vector4a] "r" (vector4a), [vector3b] "r" (vector3b),
+              [src_ptr] "r" (src_ptr)
+        );
+
+        /* clamp and store results */
+        dst_ptr[0] = cm[Temp1];
+        dst_ptr[1] = cm[Temp2];
+        dst_ptr[2] = cm[Temp3];
+        dst_ptr[3] = cm[Temp4];
+
+        /* next 4 pixels */
+        __asm__ __volatile__ (
+            /* even 3. pixel */
+            "dpa.w.ph           $ac3,        %[p3],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac3,        %[p1],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac3,        %[p2],           %[vector3b]   \n\t"
+
+            /* even 4. pixel */
+            "mtlo               %[vector4a], $ac2                           \n\t"
+            "preceu.ph.qbl      %[p4],       %[tp1]                         \n\t"
+            "dpa.w.ph           $ac2,        %[p1],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac2,        %[p2],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac2,        %[p4],           %[vector3b]   \n\t"
+            "ulw                %[tn1],      7(%[src_ptr])                  \n\t"
+            "extp               %[Temp1],    $ac3,            9             \n\t"
+
+            /* odd 3. pixel */
+            "mtlo               %[vector4a], $ac3                           \n\t"
+            "preceu.ph.qbr      %[n2],       %[tn1]                         \n\t"
+            "dpa.w.ph           $ac3,        %[n3],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac3,        %[n1],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac3,        %[n2],           %[vector3b]   \n\t"
+            "extp               %[Temp3],    $ac2,            9             \n\t"
+
+            /* odd 4. pixel */
+            "mtlo               %[vector4a], $ac2                           \n\t"
+            "preceu.ph.qbl      %[n4],       %[tn1]                         \n\t"
+            "dpa.w.ph           $ac2,        %[n1],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac2,        %[n2],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac2,        %[n4],           %[vector3b]   \n\t"
+            "ulw                %[tp2],      10(%[src_ptr])                 \n\t"
+            "extp               %[Temp2],    $ac3,            9             \n\t"
+            "mtlo               %[vector4a], $ac3                           \n\t"
+            "preceu.ph.qbr      %[p1],       %[tp2]                         \n\t"
+            "extp               %[Temp4],    $ac2,            9             \n\t"
+
+            : [tn1] "=&r" (tn1), [tp2] "=&r" (tp2), [n2] "=&r" (n2),
+              [p4] "=&r" (p4), [n4] "=&r" (n4),
+              [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+              [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4)
+            : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+              [tp1] "r" (tp1), [n1] "r" (n1), [p1] "r" (p1),
+              [vector4a] "r" (vector4a), [p2] "r" (p2), [vector3b] "r" (vector3b),
+              [p3] "r" (p3), [n3] "r" (n3), [src_ptr] "r" (src_ptr)
+        );
+
+        /* clamp and store results */
+        dst_ptr[4] = cm[Temp1];
+        dst_ptr[5] = cm[Temp2];
+        dst_ptr[6] = cm[Temp3];
+        dst_ptr[7] = cm[Temp4];
+
+        /* next 4 pixels */
+        __asm__ __volatile__ (
+            /* even 5. pixel */
+            "dpa.w.ph           $ac3,        %[p2],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac3,        %[p4],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac3,        %[p1],           %[vector3b]   \n\t"
+
+            /* even 6. pixel */
+            "mtlo               %[vector4a], $ac2                           \n\t"
+            "preceu.ph.qbl      %[p3],       %[tp2]                         \n\t"
+            "dpa.w.ph           $ac2,        %[p4],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac2,        %[p1],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac2,        %[p3],           %[vector3b]   \n\t"
+
+            "ulw                %[tn1],      11(%[src_ptr])                 \n\t"
+            "extp               %[Temp1],    $ac3,            9             \n\t"
+
+            /* odd 5. pixel */
+            "mtlo               %[vector4a], $ac3                           \n\t"
+            "preceu.ph.qbr      %[n1],       %[tn1]                         \n\t"
+            "dpa.w.ph           $ac3,        %[n2],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac3,        %[n4],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac3,        %[n1],           %[vector3b]   \n\t"
+            "extp               %[Temp3],    $ac2,            9             \n\t"
+
+            /* odd 6. pixel */
+            "mtlo               %[vector4a], $ac2                           \n\t"
+            "preceu.ph.qbl      %[n3],       %[tn1]                         \n\t"
+            "dpa.w.ph           $ac2,        %[n4],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac2,        %[n1],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac2,        %[n3],           %[vector3b]   \n\t"
+            "ulw                %[tp1],      14(%[src_ptr])                 \n\t"
+            "extp               %[Temp2],    $ac3,            9             \n\t"
+            "mtlo               %[vector4a], $ac3                           \n\t"
+            "preceu.ph.qbr      %[p4],       %[tp1]                         \n\t"
+            "extp               %[Temp4],    $ac2,            9             \n\t"
+
+            : [tn1] "=&r" (tn1), [tp1] "=&r" (tp1),
+              [n1] "=&r" (n1), [p3] "=&r" (p3), [n3] "=&r" (n3),
+              [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+              [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4)
+            : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+              [tp2] "r" (tp2), [p2] "r" (p2), [n2] "r" (n2),
+              [p4] "r" (p4), [n4] "r" (n4), [p1] "r" (p1), [src_ptr] "r" (src_ptr),
+              [vector4a] "r" (vector4a), [vector3b] "r" (vector3b)
+        );
+
+        /* clamp and store results */
+        dst_ptr[8] = cm[Temp1];
+        dst_ptr[9] = cm[Temp2];
+        dst_ptr[10] = cm[Temp3];
+        dst_ptr[11] = cm[Temp4];
+
+        /* next 4 pixels */
+        __asm__ __volatile__ (
+            /* even 7. pixel */
+            "dpa.w.ph           $ac3,        %[p1],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac3,        %[p3],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac3,        %[p4],           %[vector3b]   \n\t"
+
+            /* even 8. pixel */
+            "mtlo               %[vector4a], $ac2                           \n\t"
+            "preceu.ph.qbl      %[p2],       %[tp1]                         \n\t"
+            "dpa.w.ph           $ac2,        %[p3],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac2,        %[p4],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac2,        %[p2],           %[vector3b]   \n\t"
+            "ulw                %[tn1],      15(%[src_ptr])                 \n\t"
+            "extp               %[Temp1],    $ac3,            9             \n\t"
+
+            /* odd 7. pixel */
+            "mtlo               %[vector4a], $ac3                           \n\t"
+            "preceu.ph.qbr      %[n4],       %[tn1]                         \n\t"
+            "dpa.w.ph           $ac3,        %[n1],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac3,        %[n3],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac3,        %[n4],           %[vector3b]   \n\t"
+            "extp               %[Temp3],    $ac2,            9             \n\t"
+
+            /* odd 8. pixel */
+            "mtlo               %[vector4a], $ac2                           \n\t"
+            "preceu.ph.qbl      %[n2],       %[tn1]                         \n\t"
+            "dpa.w.ph           $ac2,        %[n3],           %[vector1b]   \n\t"
+            "dpa.w.ph           $ac2,        %[n4],           %[vector2b]   \n\t"
+            "dpa.w.ph           $ac2,        %[n2],           %[vector3b]   \n\t"
+            "extp               %[Temp2],    $ac3,            9             \n\t"
+            "extp               %[Temp4],    $ac2,            9             \n\t"
+
+            /* clamp and store results */
+            "lbux               %[tp1],      %[Temp1](%[cm])                \n\t"
+            "lbux               %[tn1],      %[Temp2](%[cm])                \n\t"
+            "lbux               %[p2],       %[Temp3](%[cm])                \n\t"
+            "sb                 %[tp1],      12(%[dst_ptr])                 \n\t"
+            "sb                 %[tn1],      13(%[dst_ptr])                 \n\t"
+            "lbux               %[n2],       %[Temp4](%[cm])                \n\t"
+            "sb                 %[p2],       14(%[dst_ptr])                 \n\t"
+            "sb                 %[n2],       15(%[dst_ptr])                 \n\t"
+
+            : [tn1] "=&r" (tn1), [p2] "=&r" (p2), [n2] "=&r" (n2), [n4] "=&r" (n4),
+              [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+              [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4)
+            : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+              [tp1] "r" (tp1), [p4] "r" (p4), [n1] "r" (n1), [p1] "r" (p1),
+              [vector4a] "r" (vector4a), [vector3b] "r" (vector3b), [p3] "r" (p3),
+              [n3] "r" (n3), [src_ptr] "r" (src_ptr),
+              [cm] "r" (cm), [dst_ptr] "r" (dst_ptr)
+        );
+
+        src_ptr += src_pixels_per_line;
+        dst_ptr += pitch;
+    }
+}
+
+
+void vp8_filter_block2d_first_pass16_0
+(
+    unsigned char *RESTRICT src_ptr,
+    unsigned char *RESTRICT output_ptr,
+    unsigned int src_pixels_per_line
+)
+{
+    int Temp1, Temp2, Temp3, Temp4;
+    int i;
+
+    /* prefetch src_ptr data to cache memory */
+    prefetch_store(output_ptr + 32);
+
+    /* copy memory from src buffer to dst buffer */
+    for (i = 0; i < 7; i++)
+    {
+        __asm__ __volatile__ (
+            "ulw    %[Temp1],   0(%[src_ptr])                               \n\t"
+            "ulw    %[Temp2],   4(%[src_ptr])                               \n\t"
+            "ulw    %[Temp3],   8(%[src_ptr])                               \n\t"
+            "ulw    %[Temp4],   12(%[src_ptr])                              \n\t"
+            "sw     %[Temp1],   0(%[output_ptr])                            \n\t"
+            "sw     %[Temp2],   4(%[output_ptr])                            \n\t"
+            "sw     %[Temp3],   8(%[output_ptr])                            \n\t"
+            "sw     %[Temp4],   12(%[output_ptr])                           \n\t"
+            "addu   %[src_ptr], %[src_ptr],        %[src_pixels_per_line]   \n\t"
+
+            : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+              [Temp4] "=&r" (Temp4), [src_ptr] "+r" (src_ptr)
+            : [src_pixels_per_line] "r" (src_pixels_per_line),
+              [output_ptr] "r" (output_ptr)
+        );
+
+        __asm__ __volatile__ (
+            "ulw    %[Temp1],   0(%[src_ptr])                               \n\t"
+            "ulw    %[Temp2],   4(%[src_ptr])                               \n\t"
+            "ulw    %[Temp3],   8(%[src_ptr])                               \n\t"
+            "ulw    %[Temp4],   12(%[src_ptr])                              \n\t"
+            "sw     %[Temp1],   16(%[output_ptr])                           \n\t"
+            "sw     %[Temp2],   20(%[output_ptr])                           \n\t"
+            "sw     %[Temp3],   24(%[output_ptr])                           \n\t"
+            "sw     %[Temp4],   28(%[output_ptr])                           \n\t"
+            "addu   %[src_ptr], %[src_ptr],        %[src_pixels_per_line]   \n\t"
+
+            : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+              [Temp4] "=&r" (Temp4), [src_ptr] "+r" (src_ptr)
+            : [src_pixels_per_line] "r" (src_pixels_per_line),
+              [output_ptr] "r" (output_ptr)
+        );
+
+        __asm__ __volatile__ (
+            "ulw    %[Temp1],   0(%[src_ptr])                               \n\t"
+            "ulw    %[Temp2],   4(%[src_ptr])                               \n\t"
+            "ulw    %[Temp3],   8(%[src_ptr])                               \n\t"
+            "ulw    %[Temp4],   12(%[src_ptr])                              \n\t"
+            "sw     %[Temp1],   32(%[output_ptr])                           \n\t"
+            "sw     %[Temp2],   36(%[output_ptr])                           \n\t"
+            "sw     %[Temp3],   40(%[output_ptr])                           \n\t"
+            "sw     %[Temp4],   44(%[output_ptr])                           \n\t"
+            "addu   %[src_ptr], %[src_ptr],        %[src_pixels_per_line]   \n\t"
+
+            : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+              [Temp4] "=&r" (Temp4), [src_ptr] "+r" (src_ptr)
+            : [src_pixels_per_line] "r" (src_pixels_per_line),
+              [output_ptr] "r" (output_ptr)
+        );
+
+        output_ptr += 48;
+    }
+}
+
+
+void vp8_filter_block2d_first_pass16_4tap
+(
+    unsigned char *RESTRICT src_ptr,
+    unsigned char *RESTRICT output_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int output_width,
+    unsigned int output_height,
+    int xoffset,
+    int yoffset,
+    unsigned char *RESTRICT dst_ptr,
+    int pitch
+)
+{
+    unsigned int i, j;
+    int Temp1, Temp2, Temp3, Temp4;
+
+    unsigned int vector4a;
+    int vector1b, vector2b;
+    unsigned int tp1, tp2, tp3, tn1;
+    unsigned int p1, p2, p3;
+    unsigned int n1, n2, n3;
+    unsigned char *cm = ff_cropTbl + CROP_WIDTH;
+
+    vector4a = 64;
+
+    vector1b = sub_pel_filters_inv_tap_4[xoffset][0];
+    vector2b = sub_pel_filters_inv_tap_4[xoffset][1];
+
+    /* if (yoffset == 0) don't need temp buffer, data will be stored in dst_ptr */
+    if (yoffset == 0)
+    {
+        output_height -= 5;
+        src_ptr += (src_pixels_per_line + src_pixels_per_line);
+
+        for (i = output_height; i--;)
+        {
+            __asm__ __volatile__ (
+                "ulw     %[tp3],   -1(%[src_ptr])               \n\t"
+                : [tp3] "=&r" (tp3)
+                : [src_ptr] "r" (src_ptr)
+            );
+
+            /* processing 4 adjacent pixels */
+            for (j = 0; j < 16; j += 4)
+            {
+                /* apply filter with vectors pairs */
+                __asm__ __volatile__ (
+                    "ulw              %[tp2],      3(%[src_ptr])                    \n\t"
+                    "move             %[tp1],      %[tp3]                           \n\t"
+
+                    /* even 1. pixel */
+                    "mtlo             %[vector4a], $ac3                             \n\t"
+                    "mthi             $0,          $ac3                             \n\t"
+                    "move             %[tp3],      %[tp2]                           \n\t"
+                    "preceu.ph.qbr    %[p1],       %[tp1]                           \n\t"
+                    "preceu.ph.qbl    %[p2],       %[tp1]                           \n\t"
+                    "preceu.ph.qbr    %[p3],       %[tp2]                           \n\t"
+                    "dpa.w.ph         $ac3,        %[p1],           %[vector1b]     \n\t"
+                    "dpa.w.ph         $ac3,        %[p2],           %[vector2b]     \n\t"
+
+                    /* even 2. pixel */
+                    "mtlo             %[vector4a], $ac2                             \n\t"
+                    "mthi             $0,          $ac2                             \n\t"
+                    "dpa.w.ph         $ac2,        %[p2],           %[vector1b]     \n\t"
+                    "dpa.w.ph         $ac2,        %[p3],           %[vector2b]     \n\t"
+                    "extr.w           %[Temp1],    $ac3,            7               \n\t"
+
+                    /* odd 1. pixel */
+                    "ulw              %[tn1],      4(%[src_ptr])                    \n\t"
+                    "balign           %[tp2],      %[tp1],          3               \n\t"
+                    "mtlo             %[vector4a], $ac3                             \n\t"
+                    "mthi             $0,          $ac3                             \n\t"
+                    "preceu.ph.qbr    %[n1],       %[tp2]                           \n\t"
+                    "preceu.ph.qbl    %[n2],       %[tp2]                           \n\t"
+                    "preceu.ph.qbr    %[n3],       %[tn1]                           \n\t"
+                    "extr.w           %[Temp3],    $ac2,            7               \n\t"
+                    "dpa.w.ph         $ac3,        %[n1],           %[vector1b]     \n\t"
+                    "dpa.w.ph         $ac3,        %[n2],           %[vector2b]     \n\t"
+
+                    /* odd 2. pixel */
+                    "mtlo             %[vector4a], $ac2                             \n\t"
+                    "mthi             $0,          $ac2                             \n\t"
+                    "extr.w           %[Temp2],    $ac3,            7               \n\t"
+                    "dpa.w.ph         $ac2,        %[n2],           %[vector1b]     \n\t"
+                    "dpa.w.ph         $ac2,        %[n3],           %[vector2b]     \n\t"
+                    "extr.w           %[Temp4],    $ac2,            7               \n\t"
+
+                    /* clamp and store results */
+                    "lbux             %[tp1],      %[Temp1](%[cm])                  \n\t"
+                    "lbux             %[tn1],      %[Temp2](%[cm])                  \n\t"
+                    "lbux             %[tp2],      %[Temp3](%[cm])                  \n\t"
+                    "sb               %[tp1],      0(%[dst_ptr])                    \n\t"
+                    "sb               %[tn1],      1(%[dst_ptr])                    \n\t"
+                    "lbux             %[n2],       %[Temp4](%[cm])                  \n\t"
+                    "sb               %[tp2],      2(%[dst_ptr])                    \n\t"
+                    "sb               %[n2],       3(%[dst_ptr])                    \n\t"
+
+                    : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2), [tp3] "=&r" (tp3),
+                      [tn1] "=&r" (tn1), [p1] "=&r" (p1), [p2] "=&r" (p2),
+                      [n1] "=&r" (n1), [n2] "=&r" (n2), [n3] "=&r" (n3),
+                      [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [p3] "=&r" (p3),
+                      [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4)
+                    : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                      [vector4a] "r" (vector4a), [cm] "r" (cm), [dst_ptr] "r" (dst_ptr),
+                      [src_ptr] "r" (src_ptr)
+                );
+
+                src_ptr += 4;
+            }
+
+            /* Next row... */
+            src_ptr += src_pixels_per_line - 16;
+            dst_ptr += pitch;
+        }
+    }
+    else
+    {
+        for (i = output_height; i--;)
+        {
+            /* processing 4 adjacent pixels */
+            for (j = 0; j < 16; j += 4)
+            {
+                /* apply filter with vectors pairs */
+                __asm__ __volatile__ (
+                    "ulw              %[tp1],      -1(%[src_ptr])                   \n\t"
+                    "ulw              %[tp2],      3(%[src_ptr])                    \n\t"
+
+                    /* even 1. pixel */
+                    "mtlo             %[vector4a], $ac3                             \n\t"
+                    "mthi             $0,          $ac3                             \n\t"
+                    "preceu.ph.qbr    %[p1],       %[tp1]                           \n\t"
+                    "preceu.ph.qbl    %[p2],       %[tp1]                           \n\t"
+                    "preceu.ph.qbr    %[p3],       %[tp2]                           \n\t"
+                    "dpa.w.ph         $ac3,        %[p1],           %[vector1b]     \n\t"
+                    "dpa.w.ph         $ac3,        %[p2],           %[vector2b]     \n\t"
+
+                    /* even 2. pixel */
+                    "mtlo             %[vector4a], $ac2                             \n\t"
+                    "mthi             $0,          $ac2                             \n\t"
+                    "dpa.w.ph         $ac2,        %[p2],           %[vector1b]     \n\t"
+                    "dpa.w.ph         $ac2,        %[p3],           %[vector2b]     \n\t"
+                    "extr.w           %[Temp1],    $ac3,            7               \n\t"
+
+                    /* odd 1. pixel */
+                    "ulw              %[tn1],      4(%[src_ptr])                    \n\t"
+                    "balign           %[tp2],      %[tp1],          3               \n\t"
+                    "mtlo             %[vector4a], $ac3                             \n\t"
+                    "mthi             $0,          $ac3                             \n\t"
+                    "preceu.ph.qbr    %[n1],       %[tp2]                           \n\t"
+                    "preceu.ph.qbl    %[n2],       %[tp2]                           \n\t"
+                    "preceu.ph.qbr    %[n3],       %[tn1]                           \n\t"
+                    "extr.w           %[Temp3],    $ac2,            7               \n\t"
+                    "dpa.w.ph         $ac3,        %[n1],           %[vector1b]     \n\t"
+                    "dpa.w.ph         $ac3,        %[n2],           %[vector2b]     \n\t"
+
+                    /* odd 2. pixel */
+                    "mtlo             %[vector4a], $ac2                             \n\t"
+                    "mthi             $0,          $ac2                             \n\t"
+                    "extr.w           %[Temp2],    $ac3,            7               \n\t"
+                    "dpa.w.ph         $ac2,        %[n2],           %[vector1b]     \n\t"
+                    "dpa.w.ph         $ac2,        %[n3],           %[vector2b]     \n\t"
+                    "extr.w           %[Temp4],    $ac2,            7               \n\t"
+
+                    /* clamp and store results */
+                    "lbux             %[tp1],      %[Temp1](%[cm])                  \n\t"
+                    "lbux             %[tn1],      %[Temp2](%[cm])                  \n\t"
+                    "lbux             %[tp2],      %[Temp3](%[cm])                  \n\t"
+                    "sb               %[tp1],      0(%[output_ptr])                 \n\t"
+                    "sb               %[tn1],      1(%[output_ptr])                 \n\t"
+                    "lbux             %[n2],       %[Temp4](%[cm])                  \n\t"
+                    "sb               %[tp2],      2(%[output_ptr])                 \n\t"
+                    "sb               %[n2],       3(%[output_ptr])                 \n\t"
+
+                    : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2), [tn1] "=&r" (tn1),
+                      [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3),
+                      [n1] "=&r" (n1), [n2] "=&r" (n2), [n3] "=&r" (n3),
+                      [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                      [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4)
+                    : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                      [vector4a] "r" (vector4a), [cm] "r" (cm),
+                      [output_ptr] "r" (output_ptr), [src_ptr] "r" (src_ptr)
+                );
+
+                src_ptr += 4;
+            }
+
+            /* next row... */
+            src_ptr += src_pixels_per_line;
+            output_ptr += output_width;
+        }
+    }
+}
+
+
+void vp8_filter_block2d_second_pass4
+(
+    unsigned char *RESTRICT src_ptr,
+    unsigned char *RESTRICT output_ptr,
+    int output_pitch,
+    int yoffset
+)
+{
+    unsigned int i;
+
+    int Temp1, Temp2, Temp3, Temp4;
+    unsigned int vector1b, vector2b, vector3b, vector4a;
+
+    unsigned char src_ptr_l2;
+    unsigned char src_ptr_l1;
+    unsigned char src_ptr_0;
+    unsigned char src_ptr_r1;
+    unsigned char src_ptr_r2;
+    unsigned char src_ptr_r3;
+
+    unsigned char *cm = ff_cropTbl + CROP_WIDTH;
+
+    vector4a = 64;
+
+    /* load filter coefficients */
+    vector1b = sub_pel_filterss[yoffset][0];
+    vector2b = sub_pel_filterss[yoffset][2];
+    vector3b = sub_pel_filterss[yoffset][1];
+
+    if (vector1b)
+    {
+        /* 6 tap filter */
+
+        for (i = 2; i--;)
+        {
+            /* prefetch src_ptr data to cache memory */
+            prefetch_load(src_ptr);
+
+            /* do not allow compiler to reorder instructions */
+            __asm__ __volatile__ (
+                ".set noreorder                                                 \n\t"
+                :
+                :
+            );
+
+            /* apply filter with vectors pairs */
+            __asm__ __volatile__ (
+                "lbu            %[src_ptr_l2],  -8(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_l1],  -4(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   0(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  4(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  8(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r3],  12(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -7(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_l1],  -3(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   1(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  5(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  9(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r3],  13(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "extp           %[Temp1],       $ac2,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -6(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_l1],  -2(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   2(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  6(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  10(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  14(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "extp           %[Temp2],       $ac3,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -5(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_l1],  -1(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   3(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  7(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  11(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  15(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "extp           %[Temp3],       $ac0,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp4],       $ac1,           9               \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                  [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4),
+                  [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                  [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2),
+                  [src_ptr_l2] "=&r" (src_ptr_l2), [src_ptr_r3] "=&r" (src_ptr_r3)
+                : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                  [vector3b] "r" (vector3b), [vector4a] "r" (vector4a),
+                  [src_ptr] "r" (src_ptr)
+            );
+
+            /* clamp and store results */
+            output_ptr[0] = cm[Temp1];
+            output_ptr[1] = cm[Temp2];
+            output_ptr[2] = cm[Temp3];
+            output_ptr[3] = cm[Temp4];
+
+            output_ptr += output_pitch;
+
+            /* apply filter with vectors pairs */
+            __asm__ __volatile__ (
+                "lbu            %[src_ptr_l2],  -4(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_l1],  0(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_0],   4(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  8(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  12(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  16(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -3(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_l1],  1(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_0],   5(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  9(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  13(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  17(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "extp           %[Temp1],       $ac2,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -2(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_l1],  2(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_0],   6(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  10(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  14(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  18(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "extp           %[Temp2],       $ac3,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -1(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_l1],  3(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_0],   7(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  11(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  15(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  19(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "extp           %[Temp3],       $ac0,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp4],       $ac1,           9               \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                  [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4),
+                  [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                  [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2),
+                  [src_ptr_l2] "=&r" (src_ptr_l2), [src_ptr_r3] "=&r" (src_ptr_r3)
+                : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                  [vector3b] "r" (vector3b), [vector4a] "r" (vector4a),
+                  [src_ptr] "r" (src_ptr)
+            );
+
+            /* clamp and store results */
+            output_ptr[0] = cm[Temp1];
+            output_ptr[1] = cm[Temp2];
+            output_ptr[2] = cm[Temp3];
+            output_ptr[3] = cm[Temp4];
+
+            src_ptr += 8;
+            output_ptr += output_pitch;
+        }
+    }
+    else
+    {
+        /* 4 tap filter */
+
+        /* prefetch src_ptr data to cache memory */
+        prefetch_load(src_ptr);
+
+        for (i = 2; i--;)
+        {
+            /* do not allow compiler to reorder instructions */
+            __asm__ __volatile__ (
+                ".set noreorder                                                 \n\t"
+                :
+                :
+            );
+
+            /* apply filter with vectors pairs */
+            __asm__ __volatile__ (
+                "lbu            %[src_ptr_l1],  -4(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   0(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  4(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  8(%[src_ptr])                   \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l1],  -3(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   1(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  5(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  9(%[src_ptr])                   \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "extp           %[Temp1],       $ac2,           9               \n\t"
+
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l1],  -2(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   2(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  6(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  10(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "extp           %[Temp2],       $ac3,           9               \n\t"
+
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l1],  -1(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   3(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  7(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  11(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "extp           %[Temp3],       $ac0,           9               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp4],       $ac1,           9               \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                  [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4),
+                  [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                  [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2)
+                : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                  [vector4a] "r" (vector4a), [src_ptr] "r" (src_ptr)
+            );
+
+            /* clamp and store results */
+            output_ptr[0] = cm[Temp1];
+            output_ptr[1] = cm[Temp2];
+            output_ptr[2] = cm[Temp3];
+            output_ptr[3] = cm[Temp4];
+
+            output_ptr += output_pitch;
+
+            /* apply filter with vectors pairs */
+            __asm__ __volatile__ (
+                "lbu            %[src_ptr_l1],  0(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_0],   4(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  8(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  12(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l1],  1(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_0],   5(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  9(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  13(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "extp           %[Temp1],       $ac2,           9               \n\t"
+
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l1],  2(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_0],   6(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  10(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  14(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "extp           %[Temp2],       $ac3,           9               \n\t"
+
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l1],  3(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_0],   7(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  11(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  15(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "extp           %[Temp3],       $ac0,           9               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp4],       $ac1,           9               \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                  [Temp3] "=&r" (Temp3), [Temp4] "=r" (Temp4),
+                  [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                  [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2)
+                : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                  [vector4a] "r" (vector4a), [src_ptr] "r" (src_ptr)
+            );
+
+            /* clamp and store results */
+            output_ptr[0] = cm[Temp1];
+            output_ptr[1] = cm[Temp2];
+            output_ptr[2] = cm[Temp3];
+            output_ptr[3] = cm[Temp4];
+
+            src_ptr += 8;
+            output_ptr += output_pitch;
+        }
+    }
+}
+
+
+void vp8_filter_block2d_second_pass_8
+(
+    unsigned char *RESTRICT src_ptr,
+    unsigned char *RESTRICT output_ptr,
+    int output_pitch,
+    unsigned int output_height,
+    unsigned int output_width,
+    unsigned int yoffset
+)
+{
+    unsigned int i;
+
+    int Temp1, Temp2, Temp3, Temp4, Temp5, Temp6, Temp7, Temp8;
+    unsigned int vector1b, vector2b, vector3b, vector4a;
+
+    unsigned char src_ptr_l2;
+    unsigned char src_ptr_l1;
+    unsigned char src_ptr_0;
+    unsigned char src_ptr_r1;
+    unsigned char src_ptr_r2;
+    unsigned char src_ptr_r3;
+    unsigned char *cm = ff_cropTbl + CROP_WIDTH;
+
+    vector4a = 64;
+
+    vector1b = sub_pel_filterss[yoffset][0];
+    vector2b = sub_pel_filterss[yoffset][2];
+    vector3b = sub_pel_filterss[yoffset][1];
+
+    if (vector1b)
+    {
+        /* 6 tap filter */
+
+        /* prefetch src_ptr data to cache memory */
+        prefetch_load(src_ptr);
+
+        for (i = output_height; i--;)
+        {
+            /* apply filter with vectors pairs */
+            __asm__ __volatile__ (
+                "lbu            %[src_ptr_l2],  -16(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -8(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   0(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  8(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  16(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  24(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -15(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -7(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   1(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  9(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  17(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  25(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "extp           %[Temp1],       $ac2,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -14(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -6(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   2(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  10(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  18(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  26(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "extp           %[Temp2],       $ac3,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -13(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -5(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   3(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  11(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  19(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  27(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "extp           %[Temp3],       $ac0,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+                  [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                  [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2),
+                  [src_ptr_l2] "=&r" (src_ptr_l2), [src_ptr_r3] "=&r" (src_ptr_r3)
+                : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                  [vector3b] "r" (vector3b), [vector4a] "r" (vector4a),
+                  [src_ptr] "r" (src_ptr)
+            );
+
+            /* apply filter with vectors pairs */
+            __asm__ __volatile__ (
+                "lbu            %[src_ptr_l2],  -12(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -4(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   4(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  12(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  20(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  28(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp4],       $ac1,           9               \n\t"
+
+                "lbu            %[src_ptr_l2],  -11(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -3(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   5(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  13(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  21(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  29(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "extp           %[Temp5],       $ac2,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -10(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -2(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   6(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  14(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  22(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  30(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "extp           %[Temp6],       $ac3,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -9(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_l1],  -1(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   7(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  15(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  23(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  31(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "extp           %[Temp7],       $ac0,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp8],       $ac1,           9               \n\t"
+
+                : [Temp4] "=&r" (Temp4), [Temp5] "=&r" (Temp5),
+                  [Temp6] "=&r" (Temp6), [Temp7] "=&r" (Temp7), [Temp8] "=r" (Temp8),
+                  [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                  [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2),
+                  [src_ptr_l2] "=&r" (src_ptr_l2),[src_ptr_r3] "=&r" (src_ptr_r3)
+                : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                  [vector3b] "r" (vector3b), [vector4a] "r" (vector4a),
+                  [src_ptr] "r" (src_ptr)
+            );
+
+            /* clamp and store results */
+            output_ptr[0] = cm[Temp1];
+            output_ptr[1] = cm[Temp2];
+            output_ptr[2] = cm[Temp3];
+            output_ptr[3] = cm[Temp4];
+            output_ptr[4] = cm[Temp5];
+            output_ptr[5] = cm[Temp6];
+            output_ptr[6] = cm[Temp7];
+            output_ptr[7] = cm[Temp8];
+
+            src_ptr += 8;
+            output_ptr += output_pitch;
+        }
+    }
+    else
+    {
+        /* 4 tap filter */
+
+        /* prefetch src_ptr data to cache memory */
+        prefetch_load(src_ptr);
+
+        for (i = output_height; i--;)
+        {
+            __asm__ __volatile__ (
+                "lbu            %[src_ptr_l1],  -8(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   0(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  8(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  16(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                : [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                  [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2)
+                : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                  [vector4a] "r" (vector4a), [src_ptr] "r" (src_ptr)
+            );
+
+            __asm__ __volatile__ (
+                "lbu            %[src_ptr_l1],  -7(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   1(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  9(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r2],  17(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp1],       $ac2,           9               \n\t"
+
+                : [Temp1] "=r" (Temp1),
+                  [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                  [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2)
+                : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                  [vector4a] "r" (vector4a), [src_ptr] "r" (src_ptr)
+            );
+
+            src_ptr_l1 = src_ptr[-6];
+            src_ptr_0  = src_ptr[2];
+            src_ptr_r1 = src_ptr[10];
+            src_ptr_r2 = src_ptr[18];
+
+            __asm__ __volatile__ (
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp2],       $ac3,           9               \n\t"
+
+                : [Temp2] "=r" (Temp2)
+                : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                  [src_ptr_l1] "r" (src_ptr_l1), [src_ptr_0] "r" (src_ptr_0),
+                  [src_ptr_r1] "r" (src_ptr_r1), [src_ptr_r2] "r" (src_ptr_r2),
+                  [vector4a] "r" (vector4a)
+            );
+
+            src_ptr_l1 = src_ptr[-5];
+            src_ptr_0  = src_ptr[3];
+            src_ptr_r1 = src_ptr[11];
+            src_ptr_r2 = src_ptr[19];
+
+            __asm__ __volatile__ (
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp3],       $ac0,           9               \n\t"
+
+                : [Temp3] "=r" (Temp3)
+                : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                  [src_ptr_l1] "r" (src_ptr_l1), [src_ptr_0] "r" (src_ptr_0),
+                  [src_ptr_r1] "r" (src_ptr_r1), [src_ptr_r2] "r" (src_ptr_r2),
+                  [vector4a] "r" (vector4a)
+            );
+
+            src_ptr_l1 = src_ptr[-4];
+            src_ptr_0  = src_ptr[4];
+            src_ptr_r1 = src_ptr[12];
+            src_ptr_r2 = src_ptr[20];
+
+            __asm__ __volatile__ (
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp4],       $ac1,           9               \n\t"
+
+                : [Temp4] "=r" (Temp4)
+                : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                  [src_ptr_l1] "r" (src_ptr_l1), [src_ptr_0] "r" (src_ptr_0),
+                  [src_ptr_r1] "r" (src_ptr_r1), [src_ptr_r2] "r" (src_ptr_r2),
+                  [vector4a] "r" (vector4a)
+            );
+
+            src_ptr_l1 = src_ptr[-3];
+            src_ptr_0  = src_ptr[5];
+            src_ptr_r1 = src_ptr[13];
+            src_ptr_r2 = src_ptr[21];
+
+            __asm__ __volatile__ (
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp5],       $ac2,           9               \n\t"
+
+                : [Temp5] "=&r" (Temp5)
+                : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                  [src_ptr_l1] "r" (src_ptr_l1), [src_ptr_0] "r" (src_ptr_0),
+                  [src_ptr_r1] "r" (src_ptr_r1), [src_ptr_r2] "r" (src_ptr_r2),
+                  [vector4a] "r" (vector4a)
+            );
+
+            src_ptr_l1 = src_ptr[-2];
+            src_ptr_0  = src_ptr[6];
+            src_ptr_r1 = src_ptr[14];
+            src_ptr_r2 = src_ptr[22];
+
+            __asm__ __volatile__ (
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp6],       $ac3,           9               \n\t"
+
+                : [Temp6] "=r" (Temp6)
+                : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                  [src_ptr_l1] "r" (src_ptr_l1), [src_ptr_0] "r" (src_ptr_0),
+                  [src_ptr_r1] "r" (src_ptr_r1), [src_ptr_r2] "r" (src_ptr_r2),
+                  [vector4a] "r" (vector4a)
+            );
+
+            src_ptr_l1 = src_ptr[-1];
+            src_ptr_0  = src_ptr[7];
+            src_ptr_r1 = src_ptr[15];
+            src_ptr_r2 = src_ptr[23];
+
+            __asm__ __volatile__ (
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp7],       $ac0,           9               \n\t"
+                "extp           %[Temp8],       $ac1,           9               \n\t"
+
+                : [Temp7] "=&r" (Temp7), [Temp8] "=r" (Temp8)
+                : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                  [src_ptr_l1] "r" (src_ptr_l1), [src_ptr_0] "r" (src_ptr_0),
+                  [src_ptr_r1] "r" (src_ptr_r1), [src_ptr_r2] "r" (src_ptr_r2),
+                  [vector4a] "r" (vector4a)
+            );
+
+            /* clamp and store results */
+            output_ptr[0] = cm[Temp1];
+            output_ptr[1] = cm[Temp2];
+            output_ptr[2] = cm[Temp3];
+            output_ptr[3] = cm[Temp4];
+            output_ptr[4] = cm[Temp5];
+            output_ptr[5] = cm[Temp6];
+            output_ptr[6] = cm[Temp7];
+            output_ptr[7] = cm[Temp8];
+
+            src_ptr += 8;
+            output_ptr += output_pitch;
+        }
+    }
+}
+
+
+void vp8_filter_block2d_second_pass161
+(
+    unsigned char *RESTRICT src_ptr,
+    unsigned char *RESTRICT output_ptr,
+    int output_pitch,
+    const unsigned short *vp8_filter
+)
+{
+    unsigned int i, j;
+
+    int Temp1, Temp2, Temp3, Temp4, Temp5, Temp6, Temp7, Temp8;
+    unsigned int vector4a;
+    unsigned int vector1b, vector2b, vector3b;
+
+    unsigned char src_ptr_l2;
+    unsigned char src_ptr_l1;
+    unsigned char src_ptr_0;
+    unsigned char src_ptr_r1;
+    unsigned char src_ptr_r2;
+    unsigned char src_ptr_r3;
+    unsigned char *cm = ff_cropTbl + CROP_WIDTH;
+
+    vector4a = 64;
+
+    vector1b = vp8_filter[0];
+    vector2b = vp8_filter[2];
+    vector3b = vp8_filter[1];
+
+    if (vector1b == 0)
+    {
+        /* 4 tap filter */
+
+        /* prefetch src_ptr data to cache memory */
+        prefetch_load(src_ptr + 16);
+
+        for (i = 16; i--;)
+        {
+            /* unrolling for loop */
+            for (j = 0; j < 16; j += 8)
+            {
+                /* apply filter with vectors pairs */
+                __asm__ __volatile__ (
+                    "lbu            %[src_ptr_l1],  -16(%[src_ptr])                 \n\t"
+                    "lbu            %[src_ptr_0],   0(%[src_ptr])                   \n\t"
+                    "lbu            %[src_ptr_r1],  16(%[src_ptr])                  \n\t"
+                    "lbu            %[src_ptr_r2],  32(%[src_ptr])                  \n\t"
+                    "mtlo           %[vector4a],    $ac2                            \n\t"
+                    "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                    "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                    "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                    "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                    "lbu            %[src_ptr_l1],  -15(%[src_ptr])                 \n\t"
+                    "lbu            %[src_ptr_0],   1(%[src_ptr])                   \n\t"
+                    "lbu            %[src_ptr_r1],  17(%[src_ptr])                  \n\t"
+                    "lbu            %[src_ptr_r2],  33(%[src_ptr])                  \n\t"
+                    "mtlo           %[vector4a],    $ac3                            \n\t"
+                    "extp           %[Temp1],       $ac2,           9               \n\t"
+
+                    "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                    "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                    "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                    "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                    "lbu            %[src_ptr_l1],  -14(%[src_ptr])                 \n\t"
+                    "lbu            %[src_ptr_0],   2(%[src_ptr])                   \n\t"
+                    "lbu            %[src_ptr_r1],  18(%[src_ptr])                  \n\t"
+                    "lbu            %[src_ptr_r2],  34(%[src_ptr])                  \n\t"
+                    "mtlo           %[vector4a],    $ac1                            \n\t"
+                    "extp           %[Temp2],       $ac3,           9               \n\t"
+
+                    "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                    "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                    "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                    "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                    "lbu            %[src_ptr_l1],  -13(%[src_ptr])                 \n\t"
+                    "lbu            %[src_ptr_0],   3(%[src_ptr])                   \n\t"
+                    "lbu            %[src_ptr_r1],  19(%[src_ptr])                  \n\t"
+                    "lbu            %[src_ptr_r2],  35(%[src_ptr])                  \n\t"
+                    "mtlo           %[vector4a],    $ac3                            \n\t"
+                    "extp           %[Temp3],       $ac1,           9               \n\t"
+
+                    "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                    "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                    "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                    "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                    "lbu            %[src_ptr_l1],  -12(%[src_ptr])                 \n\t"
+                    "lbu            %[src_ptr_0],   4(%[src_ptr])                   \n\t"
+                    "lbu            %[src_ptr_r1],  20(%[src_ptr])                  \n\t"
+                    "lbu            %[src_ptr_r2],  36(%[src_ptr])                  \n\t"
+                    "mtlo           %[vector4a],    $ac2                            \n\t"
+                    "extp           %[Temp4],       $ac3,           9               \n\t"
+
+                    "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                    "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                    "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                    "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                    "lbu            %[src_ptr_l1],  -11(%[src_ptr])                 \n\t"
+                    "lbu            %[src_ptr_0],   5(%[src_ptr])                   \n\t"
+                    "lbu            %[src_ptr_r1],  21(%[src_ptr])                  \n\t"
+                    "lbu            %[src_ptr_r2],  37(%[src_ptr])                  \n\t"
+                    "mtlo           %[vector4a],    $ac3                            \n\t"
+                    "extp           %[Temp5],       $ac2,           9               \n\t"
+
+                    "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                    "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                    "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                    "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                    "lbu            %[src_ptr_l1],  -10(%[src_ptr])                 \n\t"
+                    "lbu            %[src_ptr_0],   6(%[src_ptr])                   \n\t"
+                    "lbu            %[src_ptr_r1],  22(%[src_ptr])                  \n\t"
+                    "lbu            %[src_ptr_r2],  38(%[src_ptr])                  \n\t"
+                    "mtlo           %[vector4a],    $ac1                            \n\t"
+                    "extp           %[Temp6],       $ac3,           9               \n\t"
+
+                    "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                    "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                    "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                    "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                    "lbu            %[src_ptr_l1],  -9(%[src_ptr])                  \n\t"
+                    "lbu            %[src_ptr_0],   7(%[src_ptr])                   \n\t"
+                    "lbu            %[src_ptr_r1],  23(%[src_ptr])                  \n\t"
+                    "lbu            %[src_ptr_r2],  39(%[src_ptr])                  \n\t"
+                    "mtlo           %[vector4a],    $ac3                            \n\t"
+                    "extp           %[Temp7],       $ac1,           9               \n\t"
+
+                    "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                    "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                    "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                    "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                    "extp           %[Temp8],       $ac3,           9               \n\t"
+
+                    : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                      [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4),
+                      [Temp5] "=&r" (Temp5), [Temp6] "=&r" (Temp6),
+                      [Temp7] "=&r" (Temp7), [Temp8] "=r" (Temp8),
+                      [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                      [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2)
+                    : [vector2b] "r" (vector2b), [vector3b] "r" (vector3b),
+                      [vector4a] "r" (vector4a), [src_ptr] "r" (src_ptr)
+                );
+
+                /* clamp and store results */
+                output_ptr[j] = cm[Temp1];
+                output_ptr[j + 1] = cm[Temp2];
+                output_ptr[j + 2] = cm[Temp3];
+                output_ptr[j + 3] = cm[Temp4];
+                output_ptr[j + 4] = cm[Temp5];
+                output_ptr[j + 5] = cm[Temp6];
+                output_ptr[j + 6] = cm[Temp7];
+                output_ptr[j + 7] = cm[Temp8];
+
+                src_ptr += 8;
+            }
+
+            output_ptr += output_pitch;
+        }
+    }
+    else
+    {
+        /* 4 tap filter */
+
+        /* prefetch src_ptr data to cache memory */
+        prefetch_load(src_ptr + 16);
+
+        /* unroll for loop */
+        for (i = 16; i--;)
+        {
+            /* apply filter with vectors pairs */
+            __asm__ __volatile__ (
+                "lbu            %[src_ptr_l2],  -32(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -16(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_0],   0(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  16(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  32(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  48(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -31(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -15(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_0],   1(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  17(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  33(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  49(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "extp           %[Temp1],       $ac2,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -30(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -14(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_0],   2(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  18(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  34(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  50(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "extp           %[Temp2],       $ac0,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -29(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -13(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_0],   3(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  19(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  35(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  51(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "extp           %[Temp3],       $ac1,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -28(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -12(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_0],   4(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  20(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  36(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  52(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+                "extp           %[Temp4],       $ac3,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -27(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -11(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_0],   5(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  21(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  37(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  53(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "extp           %[Temp5],       $ac2,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -26(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -10(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_0],   6(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  22(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  38(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  54(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "extp           %[Temp6],       $ac0,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -25(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -9(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   7(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  23(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  39(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  55(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "extp           %[Temp7],       $ac1,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp8],       $ac3,           9               \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                  [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4),
+                  [Temp5] "=&r" (Temp5), [Temp6] "=&r" (Temp6),
+                  [Temp7] "=&r" (Temp7), [Temp8] "=r" (Temp8),
+                  [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                  [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2),
+                  [src_ptr_l2] "=&r" (src_ptr_l2),[src_ptr_r3] "=&r" (src_ptr_r3)
+                : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                  [vector3b] "r" (vector3b), [vector4a] "r" (vector4a),
+                  [src_ptr] "r" (src_ptr)
+            );
+
+            /* clamp and store results */
+            output_ptr[0] = cm[Temp1];
+            output_ptr[1] = cm[Temp2];
+            output_ptr[2] = cm[Temp3];
+            output_ptr[3] = cm[Temp4];
+            output_ptr[4] = cm[Temp5];
+            output_ptr[5] = cm[Temp6];
+            output_ptr[6] = cm[Temp7];
+            output_ptr[7] = cm[Temp8];
+
+            /* apply filter with vectors pairs */
+            __asm__ __volatile__ (
+                "lbu            %[src_ptr_l2],  -24(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -8(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   8(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  24(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  40(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  56(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -23(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -7(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   9(%[src_ptr])                   \n\t"
+                "lbu            %[src_ptr_r1],  25(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  41(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  57(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "extp           %[Temp1],       $ac2,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -22(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -6(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   10(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r1],  26(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  42(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  58(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "extp           %[Temp2],       $ac0,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -21(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -5(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   11(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r1],  27(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  43(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  59(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "extp           %[Temp3],       $ac1,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -20(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -4(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   12(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r1],  28(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  44(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  60(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac2                            \n\t"
+                "extp           %[Temp4],       $ac3,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac2,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac2,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -19(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -3(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   13(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r1],  29(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  45(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  61(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac0                            \n\t"
+                "extp           %[Temp5],       $ac2,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac0,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac0,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -18(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -2(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   14(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r1],  30(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  46(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  62(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac1                            \n\t"
+                "extp           %[Temp6],       $ac0,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac1,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac1,           %[src_ptr_l1],  %[vector3b]     \n\t"
+
+                "lbu            %[src_ptr_l2],  -17(%[src_ptr])                 \n\t"
+                "lbu            %[src_ptr_l1],  -1(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_0],   15(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r1],  31(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r2],  47(%[src_ptr])                  \n\t"
+                "lbu            %[src_ptr_r3],  63(%[src_ptr])                  \n\t"
+                "mtlo           %[vector4a],    $ac3                            \n\t"
+                "extp           %[Temp7],       $ac1,           9               \n\t"
+
+                "append         %[src_ptr_l2],  %[src_ptr_r3],  8               \n\t"
+                "append         %[src_ptr_0],   %[src_ptr_r1],  8               \n\t"
+                "append         %[src_ptr_l1],  %[src_ptr_r2],  8               \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_l2],  %[vector1b]     \n\t"
+                "dpau.h.qbr     $ac3,           %[src_ptr_0],   %[vector2b]     \n\t"
+                "dpsu.h.qbr     $ac3,           %[src_ptr_l1],  %[vector3b]     \n\t"
+                "extp           %[Temp8],       $ac3,           9               \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+                  [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4),
+                  [Temp5] "=&r" (Temp5), [Temp6] "=&r" (Temp6),
+                  [Temp7] "=&r" (Temp7), [Temp8] "=r" (Temp8),
+                  [src_ptr_l1] "=&r" (src_ptr_l1), [src_ptr_0] "=&r" (src_ptr_0),
+                  [src_ptr_r1] "=&r" (src_ptr_r1), [src_ptr_r2] "=&r" (src_ptr_r2),
+                  [src_ptr_l2] "=&r" (src_ptr_l2), [src_ptr_r3] "=&r" (src_ptr_r3)
+                : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+                  [vector3b] "r" (vector3b), [vector4a] "r" (vector4a),
+                  [src_ptr] "r" (src_ptr)
+            );
+
+            src_ptr += 16;
+            output_ptr[8] = cm[Temp1];
+            output_ptr[9] = cm[Temp2];
+            output_ptr[10] = cm[Temp3];
+            output_ptr[11] = cm[Temp4];
+            output_ptr[12] = cm[Temp5];
+            output_ptr[13] = cm[Temp6];
+            output_ptr[14] = cm[Temp7];
+            output_ptr[15] = cm[Temp8];
+
+            output_ptr += output_pitch;
+        }
+    }
+}
+
+
+void vp8_sixtap_predict4x4_dspr2
+(
+    unsigned char *RESTRICT src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *RESTRICT dst_ptr,
+    int dst_pitch
+)
+{
+    unsigned char FData[9 * 4]; /* Temp data bufffer used in filtering */
+    unsigned int pos = 16;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__ (
+        "wrdsp      %[pos],     1           \n\t"
+        :
+        : [pos] "r" (pos)
+    );
+
+    if (yoffset)
+    {
+        /* First filter 1-D horizontally... */
+        vp8_filter_block2d_first_pass_4(src_ptr - (2 * src_pixels_per_line), FData,
+                                        src_pixels_per_line, 9, xoffset, 4);
+        /* then filter verticaly... */
+        vp8_filter_block2d_second_pass4(FData + 8, dst_ptr, dst_pitch, yoffset);
+    }
+    else
+        /* if (yoffsset == 0) vp8_filter_block2d_first_pass save data to dst_ptr */
+        vp8_filter_block2d_first_pass_4(src_ptr, dst_ptr, src_pixels_per_line,
+                                        4, xoffset, dst_pitch);
+}
+
+
+void vp8_sixtap_predict8x8_dspr2
+(
+    unsigned char   *RESTRICT src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *RESTRICT dst_ptr,
+    int  dst_pitch
+)
+{
+
+    unsigned char FData[13 * 8]; /* Temp data bufffer used in filtering */
+    unsigned int pos, Temp1, Temp2;
+
+    pos = 16;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__ (
+        "wrdsp      %[pos],     1               \n\t"
+        :
+        : [pos] "r" (pos)
+    );
+
+    if (yoffset)
+    {
+
+        src_ptr = src_ptr - (2 * src_pixels_per_line);
+
+        if (xoffset)
+            /* filter 1-D horizontally... */
+            vp8_filter_block2d_first_pass_8_all(src_ptr, FData, src_pixels_per_line,
+                                                13, xoffset, 8);
+
+        else
+        {
+            /* prefetch src_ptr data to cache memory */
+            prefetch_load(src_ptr + 2 * src_pixels_per_line);
+
+            __asm__ __volatile__ (
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   0(%[FData])                             \n\t"
+                "sw     %[Temp2],   4(%[FData])                             \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   8(%[FData])                             \n\t"
+                "sw     %[Temp2],   12(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   16(%[FData])                            \n\t"
+                "sw     %[Temp2],   20(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   24(%[FData])                            \n\t"
+                "sw     %[Temp2],   28(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   32(%[FData])                            \n\t"
+                "sw     %[Temp2],   36(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   40(%[FData])                            \n\t"
+                "sw     %[Temp2],   44(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   48(%[FData])                            \n\t"
+                "sw     %[Temp2],   52(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   56(%[FData])                            \n\t"
+                "sw     %[Temp2],   60(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   64(%[FData])                            \n\t"
+                "sw     %[Temp2],   68(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   72(%[FData])                            \n\t"
+                "sw     %[Temp2],   76(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   80(%[FData])                            \n\t"
+                "sw     %[Temp2],   84(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   88(%[FData])                            \n\t"
+                "sw     %[Temp2],   92(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   96(%[FData])                            \n\t"
+                "sw     %[Temp2],   100(%[FData])                           \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2)
+                : [FData] "r" (FData), [src_ptr] "r" (src_ptr),
+                  [src_pixels_per_line] "r" (src_pixels_per_line)
+            );
+        }
+
+        /* filter verticaly... */
+        vp8_filter_block2d_second_pass_8(FData + 16, dst_ptr, dst_pitch, 8, 8, yoffset);
+    }
+
+    /* if (yoffsset == 0) vp8_filter_block2d_first_pass save data to dst_ptr */
+    else
+    {
+        if (xoffset)
+            vp8_filter_block2d_first_pass_8_all(src_ptr, dst_ptr, src_pixels_per_line,
+                                                8, xoffset, dst_pitch);
+
+        else
+        {
+            /* copy from src buffer to dst buffer */
+            __asm__ __volatile__ (
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   0(%[dst_ptr])                           \n\t"
+                "sw     %[Temp2],   4(%[dst_ptr])                           \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   8(%[dst_ptr])                           \n\t"
+                "sw     %[Temp2],   12(%[dst_ptr])                          \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   16(%[dst_ptr])                          \n\t"
+                "sw     %[Temp2],   20(%[dst_ptr])                          \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   24(%[dst_ptr])                          \n\t"
+                "sw     %[Temp2],   28(%[dst_ptr])                          \n\t"
+                "addu   %[src_ptr], %[src_ptr],   %[src_pixels_per_line]    \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   32(%[dst_ptr])                          \n\t"
+                "sw     %[Temp2],   36(%[dst_ptr])                          \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   40(%[dst_ptr])                          \n\t"
+                "sw     %[Temp2],   44(%[dst_ptr])                          \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   48(%[dst_ptr])                          \n\t"
+                "sw     %[Temp2],   52(%[dst_ptr])                          \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   56(%[dst_ptr])                          \n\t"
+                "sw     %[Temp2],   60(%[dst_ptr])                          \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2)
+                : [dst_ptr] "r" (dst_ptr), [src_ptr] "r" (src_ptr),
+                  [src_pixels_per_line] "r" (src_pixels_per_line)
+            );
+        }
+    }
+}
+
+
+void vp8_sixtap_predict8x4_dspr2
+(
+    unsigned char   *RESTRICT src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *RESTRICT dst_ptr,
+    int  dst_pitch
+)
+{
+    unsigned char FData[9 * 8]; /* Temp data bufffer used in filtering */
+    unsigned int pos, Temp1, Temp2;
+
+    pos = 16;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__ (
+        "wrdsp      %[pos],     1           \n\t"
+        :
+        : [pos] "r" (pos)
+    );
+
+    if (yoffset)
+    {
+
+        src_ptr = src_ptr - (2 * src_pixels_per_line);
+
+        if (xoffset)
+            /* filter 1-D horizontally... */
+            vp8_filter_block2d_first_pass_8_all(src_ptr, FData, src_pixels_per_line,
+                                                9, xoffset, 8);
+
+        else
+        {
+            /* prefetch src_ptr data to cache memory */
+            prefetch_load(src_ptr + 2 * src_pixels_per_line);
+
+            __asm__ __volatile__ (
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   0(%[FData])                             \n\t"
+                "sw     %[Temp2],   4(%[FData])                             \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   8(%[FData])                             \n\t"
+                "sw     %[Temp2],   12(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   16(%[FData])                            \n\t"
+                "sw     %[Temp2],   20(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   24(%[FData])                            \n\t"
+                "sw     %[Temp2],   28(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   32(%[FData])                            \n\t"
+                "sw     %[Temp2],   36(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   40(%[FData])                            \n\t"
+                "sw     %[Temp2],   44(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   48(%[FData])                            \n\t"
+                "sw     %[Temp2],   52(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   56(%[FData])                            \n\t"
+                "sw     %[Temp2],   60(%[FData])                            \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   64(%[FData])                            \n\t"
+                "sw     %[Temp2],   68(%[FData])                            \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2)
+                : [FData] "r" (FData), [src_ptr] "r" (src_ptr),
+                  [src_pixels_per_line] "r" (src_pixels_per_line)
+            );
+        }
+
+        /* filter verticaly... */
+        vp8_filter_block2d_second_pass_8(FData + 16, dst_ptr, dst_pitch, 4, 8, yoffset);
+    }
+
+    /* if (yoffsset == 0) vp8_filter_block2d_first_pass save data to dst_ptr */
+    else
+    {
+        if (xoffset)
+            vp8_filter_block2d_first_pass_8_all(src_ptr, dst_ptr, src_pixels_per_line,
+                                                4, xoffset, dst_pitch);
+
+        else
+        {
+            /* copy from src buffer to dst buffer */
+            __asm__ __volatile__ (
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   0(%[dst_ptr])                           \n\t"
+                "sw     %[Temp2],   4(%[dst_ptr])                           \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   8(%[dst_ptr])                           \n\t"
+                "sw     %[Temp2],   12(%[dst_ptr])                          \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   16(%[dst_ptr])                          \n\t"
+                "sw     %[Temp2],   20(%[dst_ptr])                          \n\t"
+                "addu   %[src_ptr], %[src_ptr],    %[src_pixels_per_line]   \n\t"
+
+                "ulw    %[Temp1],   0(%[src_ptr])                           \n\t"
+                "ulw    %[Temp2],   4(%[src_ptr])                           \n\t"
+                "sw     %[Temp1],   24(%[dst_ptr])                          \n\t"
+                "sw     %[Temp2],   28(%[dst_ptr])                          \n\t"
+
+                : [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2)
+                : [dst_ptr] "r" (dst_ptr), [src_ptr] "r" (src_ptr),
+                  [src_pixels_per_line] "r" (src_pixels_per_line)
+            );
+        }
+    }
+}
+
+
+void vp8_sixtap_predict16x16_dspr2
+(
+    unsigned char   *RESTRICT src_ptr,
+    int  src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *RESTRICT dst_ptr,
+    int  dst_pitch
+)
+{
+    const unsigned short *VFilter;
+    unsigned char FData[21 * 16]; /* Temp data bufffer used in filtering */
+    unsigned int pos;
+
+    VFilter = sub_pel_filterss[yoffset];
+
+    pos = 16;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__ (
+        "wrdsp      %[pos],     1           \n\t"
+        :
+        : [pos] "r" (pos)
+    );
+
+    if (yoffset)
+    {
+
+        src_ptr = src_ptr - (2 * src_pixels_per_line);
+
+        switch (xoffset)
+        {
+            /* filter 1-D horizontally... */
+        case 2:
+        case 4:
+        case 6:
+            /* 6 tap filter */
+            vp8_filter_block2d_first_pass16_6tap(src_ptr, FData, src_pixels_per_line,
+                                                 21, xoffset, 16);
+            break;
+
+        case 0:
+            /* only copy buffer */
+            vp8_filter_block2d_first_pass16_0(src_ptr, FData, src_pixels_per_line);
+            break;
+
+        case 1:
+        case 3:
+        case 5:
+        case 7:
+            /* 4 tap filter */
+            vp8_filter_block2d_first_pass16_4tap(src_ptr, FData, src_pixels_per_line, 16,
+                                                 21, xoffset, yoffset, dst_ptr, dst_pitch);
+            break;
+        }
+
+        /* filter verticaly... */
+        vp8_filter_block2d_second_pass161(FData + 32, dst_ptr, dst_pitch, VFilter);
+    }
+    else
+    {
+        /* if (yoffsset == 0) vp8_filter_block2d_first_pass save data to dst_ptr */
+        switch (xoffset)
+        {
+        case 2:
+        case 4:
+        case 6:
+            /* 6 tap filter */
+            vp8_filter_block2d_first_pass16_6tap(src_ptr, dst_ptr, src_pixels_per_line,
+                                                 16, xoffset, dst_pitch);
+            break;
+
+        case 1:
+        case 3:
+        case 5:
+        case 7:
+            /* 4 tap filter */
+            vp8_filter_block2d_first_pass16_4tap(src_ptr, dst_ptr, src_pixels_per_line, 16,
+                                                 21, xoffset, yoffset, dst_ptr, dst_pitch);
+            break;
+        }
+    }
+}
+
+#endif
diff --git a/libvpx/libvpx/vp8/common/mips/dspr2/idct_blk_dspr2.c b/libvpx/libvpx/vp8/common/mips/dspr2/idct_blk_dspr2.c
new file mode 100644
index 0000000..ab938cd
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/dspr2/idct_blk_dspr2.c
@@ -0,0 +1,88 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+
+#if HAVE_DSPR2
+
+void vp8_dequant_idct_add_y_block_dspr2
+(short *q, short *dq,
+ unsigned char *dst, int stride, char *eobs)
+{
+    int i, j;
+
+    for (i = 0; i < 4; i++)
+    {
+        for (j = 0; j < 4; j++)
+        {
+            if (*eobs++ > 1)
+                vp8_dequant_idct_add_dspr2(q, dq, dst, stride);
+            else
+            {
+                vp8_dc_only_idct_add_dspr2(q[0]*dq[0], dst, stride, dst, stride);
+                ((int *)q)[0] = 0;
+            }
+
+            q   += 16;
+            dst += 4;
+        }
+
+        dst += 4 * stride - 16;
+    }
+}
+
+void vp8_dequant_idct_add_uv_block_dspr2
+(short *q, short *dq,
+ unsigned char *dstu, unsigned char *dstv, int stride, char *eobs)
+{
+    int i, j;
+
+    for (i = 0; i < 2; i++)
+    {
+        for (j = 0; j < 2; j++)
+        {
+            if (*eobs++ > 1)
+                vp8_dequant_idct_add_dspr2(q, dq, dstu, stride);
+            else
+            {
+                vp8_dc_only_idct_add_dspr2(q[0]*dq[0], dstu, stride, dstu, stride);
+                ((int *)q)[0] = 0;
+            }
+
+            q    += 16;
+            dstu += 4;
+        }
+
+        dstu += 4 * stride - 8;
+    }
+
+    for (i = 0; i < 2; i++)
+    {
+        for (j = 0; j < 2; j++)
+        {
+            if (*eobs++ > 1)
+                vp8_dequant_idct_add_dspr2(q, dq, dstv, stride);
+            else
+            {
+                vp8_dc_only_idct_add_dspr2(q[0]*dq[0], dstv, stride, dstv, stride);
+                ((int *)q)[0] = 0;
+            }
+
+            q    += 16;
+            dstv += 4;
+        }
+
+        dstv += 4 * stride - 8;
+    }
+}
+
+#endif
+
diff --git a/libvpx/libvpx/vp8/common/mips/dspr2/idctllm_dspr2.c b/libvpx/libvpx/vp8/common/mips/dspr2/idctllm_dspr2.c
new file mode 100644
index 0000000..2eff710
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/dspr2/idctllm_dspr2.c
@@ -0,0 +1,369 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp8_rtcd.h"
+
+#if HAVE_DSPR2
+#define CROP_WIDTH 256
+
+/******************************************************************************
+ * Notes:
+ *
+ * This implementation makes use of 16 bit fixed point version of two multiply
+ * constants:
+ *         1.   sqrt(2) * cos (pi/8)
+ *         2.   sqrt(2) * sin (pi/8)
+ * Since the first constant is bigger than 1, to maintain the same 16 bit
+ * fixed point precision as the second one, we use a trick of
+ *         x * a = x + x*(a-1)
+ * so
+ *         x * sqrt(2) * cos (pi/8) = x + x * (sqrt(2) *cos(pi/8)-1).
+ ****************************************************************************/
+extern unsigned char ff_cropTbl[256 + 2 * CROP_WIDTH];
+static const int cospi8sqrt2minus1 = 20091;
+static const int sinpi8sqrt2      = 35468;
+
+inline void prefetch_load_short(short *src)
+{
+    __asm__ __volatile__ (
+        "pref   0,  0(%[src])   \n\t"
+        :
+        : [src] "r" (src)
+    );
+}
+
+void vp8_short_idct4x4llm_dspr2(short *input, unsigned char *pred_ptr,
+                                int pred_stride, unsigned char *dst_ptr,
+                                int dst_stride)
+{
+    int r, c;
+    int a1, b1, c1, d1;
+    short output[16];
+    short *ip = input;
+    short *op = output;
+    int temp1, temp2;
+    int shortpitch = 4;
+
+    int c2, d2;
+    int temp3, temp4;
+    unsigned char *cm = ff_cropTbl + CROP_WIDTH;
+
+    /* prepare data for load */
+    prefetch_load_short(ip + 8);
+
+    /* first loop is unrolled */
+    a1 = ip[0] + ip[8];
+    b1 = ip[0] - ip[8];
+
+    temp1 = (ip[4] * sinpi8sqrt2) >> 16;
+    temp2 = ip[12] + ((ip[12] * cospi8sqrt2minus1) >> 16);
+    c1 = temp1 - temp2;
+
+    temp1 = ip[4] + ((ip[4] * cospi8sqrt2minus1) >> 16);
+    temp2 = (ip[12] * sinpi8sqrt2) >> 16;
+    d1 = temp1 + temp2;
+
+    temp3 = (ip[5] * sinpi8sqrt2) >> 16;
+    temp4 = ip[13] + ((ip[13] * cospi8sqrt2minus1) >> 16);
+    c2 = temp3 - temp4;
+
+    temp3 = ip[5] + ((ip[5] * cospi8sqrt2minus1) >> 16);
+    temp4 = (ip[13] * sinpi8sqrt2) >> 16;
+    d2 = temp3 + temp4;
+
+    op[0] = a1 + d1;
+    op[12] = a1 - d1;
+    op[4] = b1 + c1;
+    op[8] = b1 - c1;
+
+    a1 = ip[1] + ip[9];
+    b1 = ip[1] - ip[9];
+
+    op[1] = a1 + d2;
+    op[13] = a1 - d2;
+    op[5] = b1 + c2;
+    op[9] = b1 - c2;
+
+    a1 = ip[2] + ip[10];
+    b1 = ip[2] - ip[10];
+
+    temp1 = (ip[6] * sinpi8sqrt2) >> 16;
+    temp2 = ip[14] + ((ip[14] * cospi8sqrt2minus1) >> 16);
+    c1 = temp1 - temp2;
+
+    temp1 = ip[6] + ((ip[6] * cospi8sqrt2minus1) >> 16);
+    temp2 = (ip[14] * sinpi8sqrt2) >> 16;
+    d1 = temp1 + temp2;
+
+    temp3 = (ip[7] * sinpi8sqrt2) >> 16;
+    temp4 = ip[15] + ((ip[15] * cospi8sqrt2minus1) >> 16);
+    c2 = temp3 - temp4;
+
+    temp3 = ip[7] + ((ip[7] * cospi8sqrt2minus1) >> 16);
+    temp4 = (ip[15] * sinpi8sqrt2) >> 16;
+    d2 = temp3 + temp4;
+
+    op[2] = a1 + d1;
+    op[14] = a1 - d1;
+    op[6] = b1 + c1;
+    op[10] = b1 - c1;
+
+    a1 = ip[3] + ip[11];
+    b1 = ip[3] - ip[11];
+
+    op[3] = a1 + d2;
+    op[15] = a1 - d2;
+    op[7] = b1 + c2;
+    op[11] = b1 - c2;
+
+    ip = output;
+
+    /* prepare data for load */
+    prefetch_load_short(ip + shortpitch);
+
+    /* second loop is unrolled */
+    a1 = ip[0] + ip[2];
+    b1 = ip[0] - ip[2];
+
+    temp1 = (ip[1] * sinpi8sqrt2) >> 16;
+    temp2 = ip[3] + ((ip[3] * cospi8sqrt2minus1) >> 16);
+    c1 = temp1 - temp2;
+
+    temp1 = ip[1] + ((ip[1] * cospi8sqrt2minus1) >> 16);
+    temp2 = (ip[3] * sinpi8sqrt2) >> 16;
+    d1 = temp1 + temp2;
+
+    temp3 = (ip[5] * sinpi8sqrt2) >> 16;
+    temp4 = ip[7] + ((ip[7] * cospi8sqrt2minus1) >> 16);
+    c2 = temp3 - temp4;
+
+    temp3 = ip[5] + ((ip[5] * cospi8sqrt2minus1) >> 16);
+    temp4 = (ip[7] * sinpi8sqrt2) >> 16;
+    d2 = temp3 + temp4;
+
+    op[0] = (a1 + d1 + 4) >> 3;
+    op[3] = (a1 - d1 + 4) >> 3;
+    op[1] = (b1 + c1 + 4) >> 3;
+    op[2] = (b1 - c1 + 4) >> 3;
+
+    a1 = ip[4] + ip[6];
+    b1 = ip[4] - ip[6];
+
+    op[4] = (a1 + d2 + 4) >> 3;
+    op[7] = (a1 - d2 + 4) >> 3;
+    op[5] = (b1 + c2 + 4) >> 3;
+    op[6] = (b1 - c2 + 4) >> 3;
+
+    a1 = ip[8] + ip[10];
+    b1 = ip[8] - ip[10];
+
+    temp1 = (ip[9] * sinpi8sqrt2) >> 16;
+    temp2 = ip[11] + ((ip[11] * cospi8sqrt2minus1) >> 16);
+    c1 = temp1 - temp2;
+
+    temp1 = ip[9] + ((ip[9] * cospi8sqrt2minus1) >> 16);
+    temp2 = (ip[11] * sinpi8sqrt2) >> 16;
+    d1 = temp1 + temp2;
+
+    temp3 = (ip[13] * sinpi8sqrt2) >> 16;
+    temp4 = ip[15] + ((ip[15] * cospi8sqrt2minus1) >> 16);
+    c2 = temp3 - temp4;
+
+    temp3 = ip[13] + ((ip[13] * cospi8sqrt2minus1) >> 16);
+    temp4 = (ip[15] * sinpi8sqrt2) >> 16;
+    d2 = temp3 + temp4;
+
+    op[8] = (a1 + d1 + 4) >> 3;
+    op[11] = (a1 - d1 + 4) >> 3;
+    op[9] = (b1 + c1 + 4) >> 3;
+    op[10] = (b1 - c1 + 4) >> 3;
+
+    a1 = ip[12] + ip[14];
+    b1 = ip[12] - ip[14];
+
+    op[12] = (a1 + d2 + 4) >> 3;
+    op[15] = (a1 - d2 + 4) >> 3;
+    op[13] = (b1 + c2 + 4) >> 3;
+    op[14] = (b1 - c2 + 4) >> 3;
+
+    ip = output;
+
+    for (r = 0; r < 4; r++)
+    {
+        for (c = 0; c < 4; c++)
+        {
+            short a = ip[c] + pred_ptr[c] ;
+            dst_ptr[c] = cm[a] ;
+        }
+
+        ip += 4;
+        dst_ptr += dst_stride;
+        pred_ptr += pred_stride;
+    }
+}
+
+void vp8_dc_only_idct_add_dspr2(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride)
+{
+    int a1;
+    int i, absa1;
+    int t2, vector_a1, vector_a;
+
+    /* a1 = ((input_dc + 4) >> 3); */
+    __asm__ __volatile__ (
+        "addi  %[a1], %[input_dc], 4   \n\t"
+        "sra   %[a1], %[a1],       3   \n\t"
+        : [a1] "=r" (a1)
+        : [input_dc] "r" (input_dc)
+    );
+
+    if (a1 < 0)
+    {
+        /* use quad-byte
+         * input and output memory are four byte aligned
+         */
+        __asm__ __volatile__ (
+            "abs        %[absa1],     %[a1]         \n\t"
+            "replv.qb   %[vector_a1], %[absa1]      \n\t"
+            : [absa1] "=r" (absa1), [vector_a1] "=r" (vector_a1)
+            : [a1] "r" (a1)
+        );
+
+        /* use (a1 - predptr[c]) instead a1 + predptr[c] */
+        for (i = 4; i--;)
+        {
+            __asm__ __volatile__ (
+                "lw             %[t2],       0(%[pred_ptr])                     \n\t"
+                "add            %[pred_ptr], %[pred_ptr],    %[pred_stride]     \n\t"
+                "subu_s.qb      %[vector_a], %[t2],          %[vector_a1]       \n\t"
+                "sw             %[vector_a], 0(%[dst_ptr])                      \n\t"
+                "add            %[dst_ptr],  %[dst_ptr],     %[dst_stride]      \n\t"
+                : [t2] "=&r" (t2), [vector_a] "=&r" (vector_a),
+                  [dst_ptr] "+&r" (dst_ptr), [pred_ptr] "+&r" (pred_ptr)
+                : [dst_stride] "r" (dst_stride), [pred_stride] "r" (pred_stride), [vector_a1] "r" (vector_a1)
+            );
+        }
+    }
+    else
+    {
+        /* use quad-byte
+         * input and output memory are four byte aligned
+         */
+        __asm__ __volatile__ (
+            "replv.qb       %[vector_a1], %[a1]     \n\t"
+            : [vector_a1] "=r" (vector_a1)
+            : [a1] "r" (a1)
+        );
+
+        for (i = 4; i--;)
+        {
+            __asm__ __volatile__ (
+                "lw             %[t2],       0(%[pred_ptr])                 \n\t"
+                "add            %[pred_ptr], %[pred_ptr],    %[pred_stride] \n\t"
+                "addu_s.qb      %[vector_a], %[vector_a1],   %[t2]          \n\t"
+                "sw             %[vector_a], 0(%[dst_ptr])                  \n\t"
+                "add            %[dst_ptr],  %[dst_ptr],     %[dst_stride]  \n\t"
+                : [t2] "=&r" (t2), [vector_a] "=&r" (vector_a),
+                  [dst_ptr] "+&r" (dst_ptr), [pred_ptr] "+&r" (pred_ptr)
+                : [dst_stride] "r" (dst_stride), [pred_stride] "r" (pred_stride), [vector_a1] "r" (vector_a1)
+            );
+        }
+    }
+
+}
+
+void vp8_short_inv_walsh4x4_dspr2(short *input, short *mb_dqcoeff)
+{
+    short output[16];
+    int i;
+    int a1, b1, c1, d1;
+    int a2, b2, c2, d2;
+    short *ip = input;
+    short *op = output;
+
+    prefetch_load_short(ip);
+
+    for (i = 4; i--;)
+    {
+        a1 = ip[0] + ip[12];
+        b1 = ip[4] + ip[8];
+        c1 = ip[4] - ip[8];
+        d1 = ip[0] - ip[12];
+
+        op[0] = a1 + b1;
+        op[4] = c1 + d1;
+        op[8] = a1 - b1;
+        op[12] = d1 - c1;
+
+        ip++;
+        op++;
+    }
+
+    ip = output;
+    op = output;
+
+    prefetch_load_short(ip);
+
+    for (i = 4; i--;)
+    {
+        a1 = ip[0] + ip[3] + 3;
+        b1 = ip[1] + ip[2];
+        c1 = ip[1] - ip[2];
+        d1 = ip[0] - ip[3] + 3;
+
+        a2 = a1 + b1;
+        b2 = d1 + c1;
+        c2 = a1 - b1;
+        d2 = d1 - c1;
+
+        op[0] = a2 >> 3;
+        op[1] = b2 >> 3;
+        op[2] = c2 >> 3;
+        op[3] = d2 >> 3;
+
+        ip += 4;
+        op += 4;
+    }
+
+    for (i = 0; i < 16; i++)
+    {
+        mb_dqcoeff[i * 16] = output[i];
+    }
+}
+
+void vp8_short_inv_walsh4x4_1_dspr2(short *input, short *mb_dqcoeff)
+{
+    int a1;
+
+    a1 = ((input[0] + 3) >> 3);
+
+    __asm__ __volatile__ (
+        "sh             %[a1], 0(%[mb_dqcoeff])                    \n\t"
+        "sh             %[a1], 32(%[mb_dqcoeff])                   \n\t"
+        "sh             %[a1], 64(%[mb_dqcoeff])                   \n\t"
+        "sh             %[a1], 96(%[mb_dqcoeff])                   \n\t"
+        "sh             %[a1], 128(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 160(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 192(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 224(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 256(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 288(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 320(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 352(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 384(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 416(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 448(%[mb_dqcoeff])                  \n\t"
+        "sh             %[a1], 480(%[mb_dqcoeff])                  \n\t"
+
+        :
+        : [a1] "r" (a1), [mb_dqcoeff] "r" (mb_dqcoeff)
+    );
+}
+
+#endif
diff --git a/libvpx/libvpx/vp8/common/mips/dspr2/reconinter_dspr2.c b/libvpx/libvpx/vp8/common/mips/dspr2/reconinter_dspr2.c
new file mode 100644
index 0000000..a14b397
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/dspr2/reconinter_dspr2.c
@@ -0,0 +1,121 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+#if HAVE_DSPR2
+inline void prefetch_load_int(unsigned char *src)
+{
+    __asm__ __volatile__ (
+        "pref   0,  0(%[src])   \n\t"
+        :
+        : [src] "r" (src)
+    );
+}
+
+
+__inline void vp8_copy_mem16x16_dspr2(
+    unsigned char *RESTRICT src,
+    int src_stride,
+    unsigned char *RESTRICT dst,
+    int dst_stride)
+{
+    int r;
+    unsigned int a0, a1, a2, a3;
+
+    for (r = 16; r--;)
+    {
+        /* load src data in cache memory */
+        prefetch_load_int(src + src_stride);
+
+        /* use unaligned memory load and store */
+        __asm__ __volatile__ (
+            "ulw    %[a0], 0(%[src])            \n\t"
+            "ulw    %[a1], 4(%[src])            \n\t"
+            "ulw    %[a2], 8(%[src])            \n\t"
+            "ulw    %[a3], 12(%[src])           \n\t"
+            "sw     %[a0], 0(%[dst])            \n\t"
+            "sw     %[a1], 4(%[dst])            \n\t"
+            "sw     %[a2], 8(%[dst])            \n\t"
+            "sw     %[a3], 12(%[dst])           \n\t"
+            : [a0] "=&r" (a0), [a1] "=&r" (a1),
+              [a2] "=&r" (a2), [a3] "=&r" (a3)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+    }
+}
+
+
+__inline void vp8_copy_mem8x8_dspr2(
+    unsigned char *RESTRICT src,
+    int src_stride,
+    unsigned char *RESTRICT dst,
+    int dst_stride)
+{
+    int r;
+    unsigned int a0, a1;
+
+    /* load src data in cache memory */
+    prefetch_load_int(src + src_stride);
+
+    for (r = 8; r--;)
+    {
+        /* use unaligned memory load and store */
+        __asm__ __volatile__ (
+            "ulw    %[a0], 0(%[src])            \n\t"
+            "ulw    %[a1], 4(%[src])            \n\t"
+            "sw     %[a0], 0(%[dst])            \n\t"
+            "sw     %[a1], 4(%[dst])            \n\t"
+            : [a0] "=&r" (a0), [a1] "=&r" (a1)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+    }
+}
+
+
+__inline void vp8_copy_mem8x4_dspr2(
+    unsigned char *RESTRICT src,
+    int src_stride,
+    unsigned char *RESTRICT dst,
+    int dst_stride)
+{
+    int r;
+    unsigned int a0, a1;
+
+    /* load src data in cache memory */
+    prefetch_load_int(src + src_stride);
+
+    for (r = 4; r--;)
+    {
+        /* use unaligned memory load and store */
+        __asm__ __volatile__ (
+            "ulw    %[a0], 0(%[src])            \n\t"
+            "ulw    %[a1], 4(%[src])            \n\t"
+            "sw     %[a0], 0(%[dst])            \n\t"
+            "sw     %[a1], 4(%[dst])            \n\t"
+           : [a0] "=&r" (a0), [a1] "=&r" (a1)
+           : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+    }
+}
+
+#endif
diff --git a/libvpx/libvpx/vp8/common/mips/dspr2/vp8_loopfilter_filters_dspr2.c b/libvpx/libvpx/vp8/common/mips/dspr2/vp8_loopfilter_filters_dspr2.c
new file mode 100644
index 0000000..9ae6bc8
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/dspr2/vp8_loopfilter_filters_dspr2.c
@@ -0,0 +1,2622 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <stdlib.h>
+#include "vp8_rtcd.h"
+#include "vp8/common/onyxc_int.h"
+
+#if HAVE_DSPR2
+typedef unsigned char uc;
+
+/* prefetch data for load */
+inline void prefetch_load_lf(unsigned char *src)
+{
+    __asm__ __volatile__ (
+        "pref   0,  0(%[src])   \n\t"
+        :
+        : [src] "r" (src)
+    );
+}
+
+
+/* prefetch data for store */
+inline void prefetch_store_lf(unsigned char *dst)
+{
+    __asm__ __volatile__ (
+        "pref   1,  0(%[dst])   \n\t"
+        :
+        : [dst] "r" (dst)
+    );
+}
+
+/* processing 4 pixels at the same time
+ * compute hev and mask in the same function
+ */
+static __inline void vp8_filter_mask_vec_mips
+(
+    uint32_t limit,
+    uint32_t flimit,
+    uint32_t p1,
+    uint32_t p0,
+    uint32_t p3,
+    uint32_t p2,
+    uint32_t q0,
+    uint32_t q1,
+    uint32_t q2,
+    uint32_t q3,
+    uint32_t thresh,
+    uint32_t *hev,
+    uint32_t *mask
+)
+{
+    uint32_t c, r, r3, r_k;
+    uint32_t s1, s2, s3;
+    uint32_t ones = 0xFFFFFFFF;
+    uint32_t hev1;
+
+    __asm__ __volatile__ (
+        /* mask |= (abs(p3 - p2) > limit) */
+        "subu_s.qb      %[c],   %[p3],     %[p2]        \n\t"
+        "subu_s.qb      %[r_k], %[p2],     %[p3]        \n\t"
+        "or             %[r_k], %[r_k],    %[c]         \n\t"
+        "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+        "or             %[r],   $0,        %[c]         \n\t"
+
+        /* mask |= (abs(p2 - p1) > limit) */
+        "subu_s.qb      %[c],   %[p2],     %[p1]        \n\t"
+        "subu_s.qb      %[r_k], %[p1],     %[p2]        \n\t"
+        "or             %[r_k], %[r_k],    %[c]         \n\t"
+        "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+        "or             %[r],   %[r],      %[c]         \n\t"
+
+        /* mask |= (abs(p1 - p0) > limit)
+         * hev  |= (abs(p1 - p0) > thresh)
+         */
+        "subu_s.qb      %[c],   %[p1],     %[p0]        \n\t"
+        "subu_s.qb      %[r_k], %[p0],     %[p1]        \n\t"
+        "or             %[r_k], %[r_k],    %[c]         \n\t"
+        "cmpgu.lt.qb    %[c],   %[thresh], %[r_k]       \n\t"
+        "or             %[r3],  $0,        %[c]         \n\t"
+        "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+        "or             %[r],   %[r],      %[c]         \n\t"
+
+        /* mask |= (abs(q1 - q0) > limit)
+         * hev  |= (abs(q1 - q0) > thresh)
+         */
+        "subu_s.qb      %[c],   %[q1],     %[q0]        \n\t"
+        "subu_s.qb      %[r_k], %[q0],     %[q1]        \n\t"
+        "or             %[r_k], %[r_k],    %[c]         \n\t"
+        "cmpgu.lt.qb    %[c],   %[thresh], %[r_k]       \n\t"
+        "or             %[r3],  %[r3],     %[c]         \n\t"
+        "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+        "or             %[r],   %[r],      %[c]         \n\t"
+
+        /* mask |= (abs(q2 - q1) > limit) */
+        "subu_s.qb      %[c],   %[q2],     %[q1]        \n\t"
+        "subu_s.qb      %[r_k], %[q1],     %[q2]        \n\t"
+        "or             %[r_k], %[r_k],    %[c]         \n\t"
+        "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+        "or             %[r],   %[r],      %[c]         \n\t"
+        "sll            %[r3],    %[r3],    24          \n\t"
+
+        /* mask |= (abs(q3 - q2) > limit) */
+        "subu_s.qb      %[c],   %[q3],     %[q2]        \n\t"
+        "subu_s.qb      %[r_k], %[q2],     %[q3]        \n\t"
+        "or             %[r_k], %[r_k],    %[c]         \n\t"
+        "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+        "or             %[r],   %[r],      %[c]         \n\t"
+
+        : [c] "=&r" (c), [r_k] "=&r" (r_k),
+          [r] "=&r" (r), [r3] "=&r" (r3)
+        : [limit] "r" (limit), [p3] "r" (p3), [p2] "r" (p2),
+          [p1] "r" (p1), [p0] "r" (p0), [q1] "r" (q1), [q0] "r" (q0),
+          [q2] "r" (q2), [q3] "r" (q3), [thresh] "r" (thresh)
+    );
+
+    __asm__ __volatile__ (
+        /* abs(p0 - q0) */
+        "subu_s.qb      %[c],   %[p0],     %[q0]        \n\t"
+        "subu_s.qb      %[r_k], %[q0],     %[p0]        \n\t"
+        "wrdsp          %[r3]                           \n\t"
+        "or             %[s1],  %[r_k],    %[c]         \n\t"
+
+        /* abs(p1 - q1) */
+        "subu_s.qb      %[c],    %[p1],    %[q1]        \n\t"
+        "addu_s.qb      %[s3],   %[s1],    %[s1]        \n\t"
+        "pick.qb        %[hev1], %[ones],  $0           \n\t"
+        "subu_s.qb      %[r_k],  %[q1],    %[p1]        \n\t"
+        "or             %[s2],   %[r_k],   %[c]         \n\t"
+
+        /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > flimit * 2 + limit */
+        "shrl.qb        %[s2],   %[s2],     1           \n\t"
+        "addu_s.qb      %[s1],   %[s2],     %[s3]       \n\t"
+        "cmpgu.lt.qb    %[c],    %[flimit], %[s1]       \n\t"
+        "or             %[r],    %[r],      %[c]        \n\t"
+        "sll            %[r],    %[r],      24          \n\t"
+
+        "wrdsp          %[r]                            \n\t"
+        "pick.qb        %[s2],  $0,         %[ones]     \n\t"
+
+        : [c] "=&r" (c), [r_k] "=&r" (r_k), [s1] "=&r" (s1), [hev1] "=&r" (hev1),
+          [s2] "=&r" (s2), [r] "+r" (r), [s3] "=&r" (s3)
+        : [p0] "r" (p0), [q0] "r" (q0), [p1] "r" (p1), [r3] "r" (r3),
+          [q1] "r" (q1), [ones] "r" (ones), [flimit] "r" (flimit)
+    );
+
+    *hev = hev1;
+    *mask = s2;
+}
+
+
+/* inputs & outputs are quad-byte vectors */
+static __inline void vp8_filter_mips
+(
+    uint32_t mask,
+    uint32_t hev,
+    uint32_t *ps1,
+    uint32_t *ps0,
+    uint32_t *qs0,
+    uint32_t *qs1
+)
+{
+    int32_t vp8_filter_l, vp8_filter_r;
+    int32_t Filter1_l, Filter1_r, Filter2_l, Filter2_r;
+    int32_t subr_r, subr_l;
+    uint32_t t1, t2, HWM, t3;
+    uint32_t hev_l, hev_r, mask_l, mask_r, invhev_l, invhev_r;
+
+    int32_t vps1, vps0, vqs0, vqs1;
+    int32_t vps1_l, vps1_r, vps0_l, vps0_r, vqs0_l, vqs0_r, vqs1_l, vqs1_r;
+    uint32_t N128;
+
+    N128 = 0x80808080;
+    t1  = 0x03000300;
+    t2  = 0x04000400;
+    t3  = 0x01000100;
+    HWM = 0xFF00FF00;
+
+    vps0 = (*ps0) ^ N128;
+    vps1 = (*ps1) ^ N128;
+    vqs0 = (*qs0) ^ N128;
+    vqs1 = (*qs1) ^ N128;
+
+    /* use halfword pairs instead quad-bytes because of accuracy */
+    vps0_l = vps0 & HWM;
+    vps0_r = vps0 << 8;
+    vps0_r = vps0_r & HWM;
+
+    vps1_l = vps1 & HWM;
+    vps1_r = vps1 << 8;
+    vps1_r = vps1_r & HWM;
+
+    vqs0_l = vqs0 & HWM;
+    vqs0_r = vqs0 << 8;
+    vqs0_r = vqs0_r & HWM;
+
+    vqs1_l = vqs1 & HWM;
+    vqs1_r = vqs1 << 8;
+    vqs1_r = vqs1_r & HWM;
+
+    mask_l = mask & HWM;
+    mask_r = mask << 8;
+    mask_r = mask_r & HWM;
+
+    hev_l = hev & HWM;
+    hev_r = hev << 8;
+    hev_r = hev_r & HWM;
+
+    __asm__ __volatile__ (
+        /* vp8_filter = vp8_signed_char_clamp(ps1 - qs1); */
+        "subq_s.ph    %[vp8_filter_l], %[vps1_l],       %[vqs1_l]       \n\t"
+        "subq_s.ph    %[vp8_filter_r], %[vps1_r],       %[vqs1_r]       \n\t"
+
+        /* qs0 - ps0 */
+        "subq_s.ph    %[subr_l],       %[vqs0_l],       %[vps0_l]       \n\t"
+        "subq_s.ph    %[subr_r],       %[vqs0_r],       %[vps0_r]       \n\t"
+
+        /* vp8_filter &= hev; */
+        "and          %[vp8_filter_l], %[vp8_filter_l], %[hev_l]        \n\t"
+        "and          %[vp8_filter_r], %[vp8_filter_r], %[hev_r]        \n\t"
+
+        /* vp8_filter = vp8_signed_char_clamp(vp8_filter + 3 * (qs0 - ps0)); */
+        "addq_s.ph    %[vp8_filter_l], %[vp8_filter_l], %[subr_l]       \n\t"
+        "addq_s.ph    %[vp8_filter_r], %[vp8_filter_r], %[subr_r]       \n\t"
+        "xor          %[invhev_l],     %[hev_l],        %[HWM]          \n\t"
+        "addq_s.ph    %[vp8_filter_l], %[vp8_filter_l], %[subr_l]       \n\t"
+        "addq_s.ph    %[vp8_filter_r], %[vp8_filter_r], %[subr_r]       \n\t"
+        "xor          %[invhev_r],     %[hev_r],        %[HWM]          \n\t"
+        "addq_s.ph    %[vp8_filter_l], %[vp8_filter_l], %[subr_l]       \n\t"
+        "addq_s.ph    %[vp8_filter_r], %[vp8_filter_r], %[subr_r]       \n\t"
+
+        /* vp8_filter &= mask; */
+        "and          %[vp8_filter_l], %[vp8_filter_l], %[mask_l]       \n\t"
+        "and          %[vp8_filter_r], %[vp8_filter_r], %[mask_r]       \n\t"
+
+        : [vp8_filter_l] "=&r" (vp8_filter_l), [vp8_filter_r] "=&r" (vp8_filter_r),
+          [subr_l] "=&r" (subr_l), [subr_r] "=&r" (subr_r),
+          [invhev_l] "=&r" (invhev_l), [invhev_r] "=&r" (invhev_r)
+
+        : [vps0_l] "r" (vps0_l), [vps0_r] "r" (vps0_r), [vps1_l] "r" (vps1_l),
+          [vps1_r] "r" (vps1_r), [vqs0_l] "r" (vqs0_l), [vqs0_r] "r" (vqs0_r),
+          [vqs1_l] "r" (vqs1_l), [vqs1_r] "r" (vqs1_r),
+          [mask_l] "r" (mask_l), [mask_r] "r" (mask_r),
+          [hev_l] "r" (hev_l), [hev_r] "r" (hev_r),
+          [HWM] "r" (HWM)
+    );
+
+    /* save bottom 3 bits so that we round one side +4 and the other +3 */
+    __asm__ __volatile__ (
+        /* Filter2 = vp8_signed_char_clamp(vp8_filter + 3) >>= 3; */
+        "addq_s.ph    %[Filter1_l],    %[vp8_filter_l], %[t2]           \n\t"
+        "addq_s.ph    %[Filter1_r],    %[vp8_filter_r], %[t2]           \n\t"
+
+        /* Filter1 = vp8_signed_char_clamp(vp8_filter + 4) >>= 3; */
+        "addq_s.ph    %[Filter2_l],    %[vp8_filter_l], %[t1]           \n\t"
+        "addq_s.ph    %[Filter2_r],    %[vp8_filter_r], %[t1]           \n\t"
+        "shra.ph      %[Filter1_r],    %[Filter1_r],    3               \n\t"
+        "shra.ph      %[Filter1_l],    %[Filter1_l],    3               \n\t"
+
+        "shra.ph      %[Filter2_l],    %[Filter2_l],    3               \n\t"
+        "shra.ph      %[Filter2_r],    %[Filter2_r],    3               \n\t"
+
+        "and          %[Filter1_l],    %[Filter1_l],    %[HWM]          \n\t"
+        "and          %[Filter1_r],    %[Filter1_r],    %[HWM]          \n\t"
+
+        /* vps0 = vp8_signed_char_clamp(ps0 + Filter2); */
+        "addq_s.ph    %[vps0_l],       %[vps0_l],       %[Filter2_l]    \n\t"
+        "addq_s.ph    %[vps0_r],       %[vps0_r],       %[Filter2_r]    \n\t"
+
+        /* vqs0 = vp8_signed_char_clamp(qs0 - Filter1); */
+        "subq_s.ph    %[vqs0_l],       %[vqs0_l],       %[Filter1_l]    \n\t"
+        "subq_s.ph    %[vqs0_r],       %[vqs0_r],       %[Filter1_r]    \n\t"
+
+        : [Filter1_l] "=&r" (Filter1_l), [Filter1_r] "=&r" (Filter1_r),
+          [Filter2_l] "=&r" (Filter2_l), [Filter2_r] "=&r" (Filter2_r),
+          [vps0_l] "+r" (vps0_l), [vps0_r] "+r" (vps0_r),
+          [vqs0_l] "+r" (vqs0_l), [vqs0_r] "+r" (vqs0_r)
+
+        : [t1] "r" (t1), [t2] "r" (t2),
+          [vp8_filter_l] "r" (vp8_filter_l), [vp8_filter_r] "r" (vp8_filter_r),
+          [HWM] "r" (HWM)
+    );
+
+    __asm__ __volatile__ (
+        /* (vp8_filter += 1) >>= 1 */
+        "addqh.ph    %[Filter1_l],    %[Filter1_l],     %[t3]           \n\t"
+        "addqh.ph    %[Filter1_r],    %[Filter1_r],     %[t3]           \n\t"
+
+        /* vp8_filter &= ~hev; */
+        "and          %[Filter1_l],    %[Filter1_l],    %[invhev_l]     \n\t"
+        "and          %[Filter1_r],    %[Filter1_r],    %[invhev_r]     \n\t"
+
+        /* vps1 = vp8_signed_char_clamp(ps1 + vp8_filter); */
+        "addq_s.ph    %[vps1_l],       %[vps1_l],       %[Filter1_l]    \n\t"
+        "addq_s.ph    %[vps1_r],       %[vps1_r],       %[Filter1_r]    \n\t"
+
+        /* vqs1 = vp8_signed_char_clamp(qs1 - vp8_filter); */
+        "subq_s.ph    %[vqs1_l],       %[vqs1_l],       %[Filter1_l]    \n\t"
+        "subq_s.ph    %[vqs1_r],       %[vqs1_r],       %[Filter1_r]    \n\t"
+
+        : [Filter1_l] "+r" (Filter1_l), [Filter1_r] "+r" (Filter1_r),
+          [vps1_l] "+r" (vps1_l), [vps1_r] "+r" (vps1_r),
+          [vqs1_l] "+r" (vqs1_l), [vqs1_r] "+r" (vqs1_r)
+
+        : [t3] "r" (t3), [invhev_l] "r" (invhev_l), [invhev_r] "r" (invhev_r)
+    );
+
+    /* Create quad-bytes from halfword pairs */
+    vqs0_l = vqs0_l & HWM;
+    vqs1_l = vqs1_l & HWM;
+    vps0_l = vps0_l & HWM;
+    vps1_l = vps1_l & HWM;
+
+    __asm__ __volatile__ (
+        "shrl.ph      %[vqs0_r],       %[vqs0_r],       8               \n\t"
+        "shrl.ph      %[vps0_r],       %[vps0_r],       8               \n\t"
+        "shrl.ph      %[vqs1_r],       %[vqs1_r],       8               \n\t"
+        "shrl.ph      %[vps1_r],       %[vps1_r],       8               \n\t"
+
+        : [vps1_r] "+r" (vps1_r), [vqs1_r] "+r" (vqs1_r),
+          [vps0_r] "+r" (vps0_r), [vqs0_r] "+r" (vqs0_r)
+        :
+    );
+
+    vqs0 = vqs0_l | vqs0_r;
+    vqs1 = vqs1_l | vqs1_r;
+    vps0 = vps0_l | vps0_r;
+    vps1 = vps1_l | vps1_r;
+
+    *ps0 = vps0 ^ N128;
+    *ps1 = vps1 ^ N128;
+    *qs0 = vqs0 ^ N128;
+    *qs1 = vqs1 ^ N128;
+}
+
+void vp8_loop_filter_horizontal_edge_mips
+(
+    unsigned char *s,
+    int p,
+    unsigned int flimit,
+    unsigned int limit,
+    unsigned int thresh,
+    int count
+)
+{
+    uint32_t mask;
+    uint32_t hev;
+    uint32_t pm1, p0, p1, p2, p3, p4, p5, p6;
+    unsigned char *sm1, *s0, *s1, *s2, *s3, *s4, *s5, *s6;
+
+    mask = 0;
+    hev = 0;
+    p1 = 0;
+    p2 = 0;
+    p3 = 0;
+    p4 = 0;
+
+    /* prefetch data for store */
+    prefetch_store_lf(s);
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+
+    sm1 = s - (p << 2);
+    s0 = s - p - p - p;
+    s1 = s - p - p ;
+    s2 = s - p;
+    s3 = s;
+    s4 = s + p;
+    s5 = s + p + p;
+    s6 = s + p + p + p;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p1 = *((uint32_t *)(s1));
+    p2 = *((uint32_t *)(s2));
+    p3 = *((uint32_t *)(s3));
+    p4 = *((uint32_t *)(s4));
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        pm1 = *((uint32_t *)(sm1));
+        p0  = *((uint32_t *)(s0));
+        p5  = *((uint32_t *)(s5));
+        p6  = *((uint32_t *)(s6));
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_filter_mips(mask, hev, &p1, &p2, &p3, &p4);
+
+            /* unpack processed 4x4 neighborhood */
+            *((uint32_t *)s1) = p1;
+            *((uint32_t *)s2) = p2;
+            *((uint32_t *)s3) = p3;
+            *((uint32_t *)s4) = p4;
+        }
+    }
+
+    sm1 += 4;
+    s0  += 4;
+    s1  += 4;
+    s2  += 4;
+    s3  += 4;
+    s4  += 4;
+    s5  += 4;
+    s6  += 4;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p1 = *((uint32_t *)(s1));
+    p2 = *((uint32_t *)(s2));
+    p3 = *((uint32_t *)(s3));
+    p4 = *((uint32_t *)(s4));
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        pm1 = *((uint32_t *)(sm1));
+        p0  = *((uint32_t *)(s0));
+        p5  = *((uint32_t *)(s5));
+        p6  = *((uint32_t *)(s6));
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_filter_mips(mask, hev, &p1, &p2, &p3, &p4);
+
+            /* unpack processed 4x4 neighborhood */
+            *((uint32_t *)s1) = p1;
+            *((uint32_t *)s2) = p2;
+            *((uint32_t *)s3) = p3;
+            *((uint32_t *)s4) = p4;
+        }
+    }
+
+    sm1 += 4;
+    s0  += 4;
+    s1  += 4;
+    s2  += 4;
+    s3  += 4;
+    s4  += 4;
+    s5  += 4;
+    s6  += 4;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p1 = *((uint32_t *)(s1));
+    p2 = *((uint32_t *)(s2));
+    p3 = *((uint32_t *)(s3));
+    p4 = *((uint32_t *)(s4));
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        pm1 = *((uint32_t *)(sm1));
+        p0  = *((uint32_t *)(s0));
+        p5  = *((uint32_t *)(s5));
+        p6  = *((uint32_t *)(s6));
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_filter_mips(mask, hev, &p1, &p2, &p3, &p4);
+
+            /* unpack processed 4x4 neighborhood */
+            *((uint32_t *)s1) = p1;
+            *((uint32_t *)s2) = p2;
+            *((uint32_t *)s3) = p3;
+            *((uint32_t *)s4) = p4;
+        }
+    }
+
+    sm1 += 4;
+    s0  += 4;
+    s1  += 4;
+    s2  += 4;
+    s3  += 4;
+    s4  += 4;
+    s5  += 4;
+    s6  += 4;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p1 = *((uint32_t *)(s1));
+    p2 = *((uint32_t *)(s2));
+    p3 = *((uint32_t *)(s3));
+    p4 = *((uint32_t *)(s4));
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        pm1 = *((uint32_t *)(sm1));
+        p0  = *((uint32_t *)(s0));
+        p5  = *((uint32_t *)(s5));
+        p6  = *((uint32_t *)(s6));
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_filter_mips(mask, hev, &p1, &p2, &p3, &p4);
+
+            /* unpack processed 4x4 neighborhood */
+            *((uint32_t *)s1) = p1;
+            *((uint32_t *)s2) = p2;
+            *((uint32_t *)s3) = p3;
+            *((uint32_t *)s4) = p4;
+        }
+    }
+}
+
+void vp8_loop_filter_uvhorizontal_edge_mips
+(
+    unsigned char *s,
+    int p,
+    unsigned int flimit,
+    unsigned int limit,
+    unsigned int thresh,
+    int count
+)
+{
+    uint32_t mask;
+    uint32_t hev;
+    uint32_t pm1, p0, p1, p2, p3, p4, p5, p6;
+    unsigned char *sm1, *s0, *s1, *s2, *s3, *s4, *s5, *s6;
+
+    mask = 0;
+    hev = 0;
+    p1 = 0;
+    p2 = 0;
+    p3 = 0;
+    p4 = 0;
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+
+    sm1 = s - (p << 2);
+    s0  = s - p - p - p;
+    s1  = s - p - p ;
+    s2  = s - p;
+    s3  = s;
+    s4  = s + p;
+    s5  = s + p + p;
+    s6  = s + p + p + p;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p1 = *((uint32_t *)(s1));
+    p2 = *((uint32_t *)(s2));
+    p3 = *((uint32_t *)(s3));
+    p4 = *((uint32_t *)(s4));
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        pm1 = *((uint32_t *)(sm1));
+        p0  = *((uint32_t *)(s0));
+        p5  = *((uint32_t *)(s5));
+        p6  = *((uint32_t *)(s6));
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_filter_mips(mask, hev, &p1, &p2, &p3, &p4);
+
+            /* unpack processed 4x4 neighborhood */
+            *((uint32_t *)s1) = p1;
+            *((uint32_t *)s2) = p2;
+            *((uint32_t *)s3) = p3;
+            *((uint32_t *)s4) = p4;
+        }
+    }
+
+    sm1 += 4;
+    s0  += 4;
+    s1  += 4;
+    s2  += 4;
+    s3  += 4;
+    s4  += 4;
+    s5  += 4;
+    s6  += 4;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p1 = *((uint32_t *)(s1));
+    p2 = *((uint32_t *)(s2));
+    p3 = *((uint32_t *)(s3));
+    p4 = *((uint32_t *)(s4));
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        pm1 = *((uint32_t *)(sm1));
+        p0  = *((uint32_t *)(s0));
+        p5  = *((uint32_t *)(s5));
+        p6  = *((uint32_t *)(s6));
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_filter_mips(mask, hev, &p1, &p2, &p3, &p4);
+
+            /* unpack processed 4x4 neighborhood */
+            *((uint32_t *)s1) = p1;
+            *((uint32_t *)s2) = p2;
+            *((uint32_t *)s3) = p3;
+            *((uint32_t *)s4) = p4;
+        }
+    }
+}
+
+void vp8_loop_filter_vertical_edge_mips
+(
+    unsigned char *s,
+    int p,
+    const unsigned int flimit,
+    const unsigned int limit,
+    const unsigned int thresh,
+    int count
+)
+{
+    int i;
+    uint32_t mask, hev;
+    uint32_t pm1, p0, p1, p2, p3, p4, p5, p6;
+    unsigned char *s1, *s2, *s3, *s4;
+    uint32_t prim1, prim2, sec3, sec4, prim3, prim4;
+
+    hev = 0;
+    mask = 0;
+    i = 0;
+    pm1 = 0;
+    p0 = 0;
+    p1 = 0;
+    p2 = 0;
+    p3 = 0;
+    p4 = 0;
+    p5 = 0;
+    p6 = 0;
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+
+    /* apply filter on 4 pixesl at the same time */
+    do
+    {
+
+        /* prefetch data for store */
+        prefetch_store_lf(s + p);
+
+        s1 = s;
+        s2 = s + p;
+        s3 = s2 + p;
+        s4 = s3 + p;
+        s  = s4 + p;
+
+        /* load quad-byte vectors
+         * memory is 4 byte aligned
+         */
+        p2  = *((uint32_t *)(s1 - 4));
+        p6  = *((uint32_t *)(s1));
+        p1  = *((uint32_t *)(s2 - 4));
+        p5  = *((uint32_t *)(s2));
+        p0  = *((uint32_t *)(s3 - 4));
+        p4  = *((uint32_t *)(s3));
+        pm1 = *((uint32_t *)(s4 - 4));
+        p3  = *((uint32_t *)(s4));
+
+        /* transpose pm1, p0, p1, p2 */
+        __asm__ __volatile__ (
+            "precrq.qb.ph   %[prim1],   %[p2],      %[p1]       \n\t"
+            "precr.qb.ph    %[prim2],   %[p2],      %[p1]       \n\t"
+            "precrq.qb.ph   %[prim3],   %[p0],      %[pm1]      \n\t"
+            "precr.qb.ph    %[prim4],   %[p0],      %[pm1]      \n\t"
+
+            "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+            "precr.qb.ph    %[pm1],     %[prim1],   %[prim2]    \n\t"
+            "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+            "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+            "precrq.ph.w    %[p2],      %[p1],      %[sec3]     \n\t"
+            "precrq.ph.w    %[p0],      %[pm1],     %[sec4]     \n\t"
+            "append         %[p1],      %[sec3],    16          \n\t"
+            "append         %[pm1],     %[sec4],    16          \n\t"
+
+            : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+              [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+              [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0), [pm1] "+r" (pm1),
+              [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+            :
+        );
+
+        /* transpose p3, p4, p5, p6 */
+        __asm__ __volatile__ (
+            "precrq.qb.ph   %[prim1],   %[p6],      %[p5]       \n\t"
+            "precr.qb.ph    %[prim2],   %[p6],      %[p5]       \n\t"
+            "precrq.qb.ph   %[prim3],   %[p4],      %[p3]       \n\t"
+            "precr.qb.ph    %[prim4],   %[p4],      %[p3]       \n\t"
+
+            "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+            "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+            "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+            "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+            "precrq.ph.w    %[p6],      %[p5],      %[sec3]     \n\t"
+            "precrq.ph.w    %[p4],      %[p3],      %[sec4]     \n\t"
+            "append         %[p5],      %[sec3],    16          \n\t"
+            "append         %[p3],      %[sec4],    16          \n\t"
+
+            : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+              [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+              [p6] "+r" (p6), [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+              [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+            :
+        );
+
+        /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+         * mask will be zero and filtering is not needed
+         */
+        if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+        {
+
+            vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                     thresh, &hev, &mask);
+
+            /* if mask == 0 do filtering is not needed */
+            if (mask)
+            {
+                /* filtering */
+                vp8_filter_mips(mask, hev, &p1, &p2, &p3, &p4);
+
+                /* unpack processed 4x4 neighborhood
+                 * don't use transpose on output data
+                 * because memory isn't aligned
+                 */
+                __asm__ __volatile__ (
+                    "sb         %[p4],  1(%[s4])    \n\t"
+                    "sb         %[p3],  0(%[s4])    \n\t"
+                    "sb         %[p2], -1(%[s4])    \n\t"
+                    "sb         %[p1], -2(%[s4])    \n\t"
+                    :
+                    : [p4] "r" (p4), [p3] "r" (p3), [s4] "r" (s4),
+                      [p2] "r" (p2), [p1] "r" (p1)
+                );
+
+                __asm__ __volatile__ (
+                    "srl        %[p4], %[p4], 8     \n\t"
+                    "srl        %[p3], %[p3], 8     \n\t"
+                    "srl        %[p2], %[p2], 8     \n\t"
+                    "srl        %[p1], %[p1], 8     \n\t"
+                    : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                    :
+                );
+
+                __asm__ __volatile__ (
+                    "sb         %[p4],  1(%[s3])    \n\t"
+                    "sb         %[p3],  0(%[s3])    \n\t"
+                    "sb         %[p2], -1(%[s3])    \n\t"
+                    "sb         %[p1], -2(%[s3])    \n\t"
+                    : [p1] "+r" (p1)
+                    : [p4] "r" (p4), [p3] "r" (p3), [s3] "r" (s3), [p2] "r" (p2)
+                );
+
+                __asm__ __volatile__ (
+                    "srl        %[p4], %[p4], 8     \n\t"
+                    "srl        %[p3], %[p3], 8     \n\t"
+                    "srl        %[p2], %[p2], 8     \n\t"
+                    "srl        %[p1], %[p1], 8     \n\t"
+                    : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                    :
+                );
+
+                __asm__ __volatile__ (
+                    "sb         %[p4],  1(%[s2])    \n\t"
+                    "sb         %[p3],  0(%[s2])    \n\t"
+                    "sb         %[p2], -1(%[s2])    \n\t"
+                    "sb         %[p1], -2(%[s2])    \n\t"
+                    :
+                    : [p4] "r" (p4), [p3] "r" (p3), [s2] "r" (s2),
+                      [p2] "r" (p2), [p1] "r" (p1)
+                );
+
+                __asm__ __volatile__ (
+                    "srl        %[p4], %[p4], 8     \n\t"
+                    "srl        %[p3], %[p3], 8     \n\t"
+                    "srl        %[p2], %[p2], 8     \n\t"
+                    "srl        %[p1], %[p1], 8     \n\t"
+                    : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                    :
+                );
+
+                __asm__ __volatile__ (
+                    "sb         %[p4],  1(%[s1])    \n\t"
+                    "sb         %[p3],  0(%[s1])    \n\t"
+                    "sb         %[p2], -1(%[s1])    \n\t"
+                    "sb         %[p1], -2(%[s1])    \n\t"
+                    :
+                    : [p4] "r" (p4), [p3] "r" (p3), [s1] "r" (s1),
+                      [p2] "r" (p2), [p1] "r" (p1)
+                );
+            }
+        }
+
+        s1 = s;
+        s2 = s + p;
+        s3 = s2 + p;
+        s4 = s3 + p;
+        s  = s4 + p;
+
+        /* load quad-byte vectors
+         * memory is 4 byte aligned
+         */
+        p2  = *((uint32_t *)(s1 - 4));
+        p6  = *((uint32_t *)(s1));
+        p1  = *((uint32_t *)(s2 - 4));
+        p5  = *((uint32_t *)(s2));
+        p0  = *((uint32_t *)(s3 - 4));
+        p4  = *((uint32_t *)(s3));
+        pm1 = *((uint32_t *)(s4 - 4));
+        p3  = *((uint32_t *)(s4));
+
+        /* transpose pm1, p0, p1, p2 */
+        __asm__ __volatile__ (
+            "precrq.qb.ph   %[prim1],   %[p2],      %[p1]       \n\t"
+            "precr.qb.ph    %[prim2],   %[p2],      %[p1]       \n\t"
+            "precrq.qb.ph   %[prim3],   %[p0],      %[pm1]      \n\t"
+            "precr.qb.ph    %[prim4],   %[p0],      %[pm1]      \n\t"
+
+            "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+            "precr.qb.ph    %[pm1],     %[prim1],   %[prim2]    \n\t"
+            "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+            "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+            "precrq.ph.w    %[p2],      %[p1],      %[sec3]     \n\t"
+            "precrq.ph.w    %[p0],      %[pm1],     %[sec4]     \n\t"
+            "append         %[p1],      %[sec3],    16          \n\t"
+            "append         %[pm1],     %[sec4],    16          \n\t"
+
+            : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+              [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+              [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0), [pm1] "+r" (pm1),
+              [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+            :
+        );
+
+        /* transpose p3, p4, p5, p6 */
+        __asm__ __volatile__ (
+            "precrq.qb.ph   %[prim1],   %[p6],      %[p5]       \n\t"
+            "precr.qb.ph    %[prim2],   %[p6],      %[p5]       \n\t"
+            "precrq.qb.ph   %[prim3],   %[p4],      %[p3]       \n\t"
+            "precr.qb.ph    %[prim4],   %[p4],      %[p3]       \n\t"
+
+            "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+            "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+            "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+            "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+            "precrq.ph.w    %[p6],      %[p5],      %[sec3]     \n\t"
+            "precrq.ph.w    %[p4],      %[p3],      %[sec4]     \n\t"
+            "append         %[p5],      %[sec3],    16          \n\t"
+            "append         %[p3],      %[sec4],    16          \n\t"
+
+            : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+              [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+              [p6] "+r" (p6), [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+              [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+            :
+        );
+
+        /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+         * mask will be zero and filtering is not needed
+         */
+        if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+        {
+
+            vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                     thresh, &hev, &mask);
+
+            /* if mask == 0 do filtering is not needed */
+            if (mask)
+            {
+                /* filtering */
+                vp8_filter_mips(mask, hev, &p1, &p2, &p3, &p4);
+
+                /* unpack processed 4x4 neighborhood
+                 * don't use transpose on output data
+                 * because memory isn't aligned
+                 */
+                __asm__ __volatile__ (
+                    "sb         %[p4],  1(%[s4])    \n\t"
+                    "sb         %[p3],  0(%[s4])    \n\t"
+                    "sb         %[p2], -1(%[s4])    \n\t"
+                    "sb         %[p1], -2(%[s4])    \n\t"
+                    :
+                    : [p4] "r" (p4), [p3] "r" (p3), [s4] "r" (s4),
+                      [p2] "r" (p2), [p1] "r" (p1)
+                );
+
+                __asm__ __volatile__ (
+                    "srl        %[p4], %[p4], 8     \n\t"
+                    "srl        %[p3], %[p3], 8     \n\t"
+                    "srl        %[p2], %[p2], 8     \n\t"
+                    "srl        %[p1], %[p1], 8     \n\t"
+                    : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                    :
+                );
+
+                __asm__ __volatile__ (
+                    "sb         %[p4],  1(%[s3])    \n\t"
+                    "sb         %[p3],  0(%[s3])    \n\t"
+                    "sb         %[p2], -1(%[s3])    \n\t"
+                    "sb         %[p1], -2(%[s3])    \n\t"
+                    : [p1] "+r" (p1)
+                    : [p4] "r" (p4), [p3] "r" (p3), [s3] "r" (s3), [p2] "r" (p2)
+                );
+
+                __asm__ __volatile__ (
+                    "srl        %[p4], %[p4], 8     \n\t"
+                    "srl        %[p3], %[p3], 8     \n\t"
+                    "srl        %[p2], %[p2], 8     \n\t"
+                    "srl        %[p1], %[p1], 8     \n\t"
+                    : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                    :
+                );
+
+                __asm__ __volatile__ (
+                    "sb         %[p4],  1(%[s2])    \n\t"
+                    "sb         %[p3],  0(%[s2])    \n\t"
+                    "sb         %[p2], -1(%[s2])    \n\t"
+                    "sb         %[p1], -2(%[s2])    \n\t"
+                    :
+                    : [p4] "r" (p4), [p3] "r" (p3), [s2] "r" (s2),
+                      [p2] "r" (p2), [p1] "r" (p1)
+                );
+
+                __asm__ __volatile__ (
+                    "srl        %[p4], %[p4], 8     \n\t"
+                    "srl        %[p3], %[p3], 8     \n\t"
+                    "srl        %[p2], %[p2], 8     \n\t"
+                    "srl        %[p1], %[p1], 8     \n\t"
+                    : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                    :
+                );
+
+                __asm__ __volatile__ (
+                    "sb         %[p4],  1(%[s1])    \n\t"
+                    "sb         %[p3],  0(%[s1])    \n\t"
+                    "sb         %[p2], -1(%[s1])    \n\t"
+                    "sb         %[p1], -2(%[s1])    \n\t"
+                    :
+                    : [p4] "r" (p4), [p3] "r" (p3), [s1] "r" (s1),
+                      [p2] "r" (p2), [p1] "r" (p1)
+                );
+            }
+        }
+
+        i += 8;
+    }
+
+    while (i < count);
+}
+
+void vp8_loop_filter_uvvertical_edge_mips
+(
+    unsigned char *s,
+    int p,
+    unsigned int flimit,
+    unsigned int limit,
+    unsigned int thresh,
+    int count
+)
+{
+    uint32_t mask, hev;
+    uint32_t pm1, p0, p1, p2, p3, p4, p5, p6;
+    unsigned char *s1, *s2, *s3, *s4;
+    uint32_t prim1, prim2, sec3, sec4, prim3, prim4;
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+
+    /* apply filter on 4 pixesl at the same time */
+
+    s1 = s;
+    s2 = s + p;
+    s3 = s2 + p;
+    s4 = s3 + p;
+
+    /* load quad-byte vectors
+    * memory is 4 byte aligned
+    */
+    p2  = *((uint32_t *)(s1 - 4));
+    p6  = *((uint32_t *)(s1));
+    p1  = *((uint32_t *)(s2 - 4));
+    p5  = *((uint32_t *)(s2));
+    p0  = *((uint32_t *)(s3 - 4));
+    p4  = *((uint32_t *)(s3));
+    pm1 = *((uint32_t *)(s4 - 4));
+    p3  = *((uint32_t *)(s4));
+
+    /* transpose pm1, p0, p1, p2 */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p2],      %[p1]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p2],      %[p1]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p0],      %[pm1]      \n\t"
+        "precr.qb.ph    %[prim4],   %[p0],      %[pm1]      \n\t"
+
+        "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[pm1],     %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p2],      %[p1],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p0],      %[pm1],     %[sec4]     \n\t"
+        "append         %[p1],      %[sec3],    16          \n\t"
+        "append         %[pm1],     %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0), [pm1] "+r" (pm1),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* transpose p3, p4, p5, p6 */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p6],      %[p5]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p6],      %[p5]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p4],      %[p3]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p4],      %[p3]       \n\t"
+
+        "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p6],      %[p5],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p4],      %[p3],      %[sec4]     \n\t"
+        "append         %[p5],      %[sec3],    16          \n\t"
+        "append         %[p3],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p6] "+r" (p6), [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+    * mask will be zero and filtering is not needed
+    */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_filter_mips(mask, hev, &p1, &p2, &p3, &p4);
+
+            /* unpack processed 4x4 neighborhood
+             * don't use transpose on output data
+             * because memory isn't aligned
+             */
+            __asm__ __volatile__ (
+                "sb         %[p4],  1(%[s4])    \n\t"
+                "sb         %[p3],  0(%[s4])    \n\t"
+                "sb         %[p2], -1(%[s4])    \n\t"
+                "sb         %[p1], -2(%[s4])    \n\t"
+                :
+                : [p4] "r" (p4), [p3] "r" (p3), [s4] "r" (s4),
+                  [p2] "r" (p2), [p1] "r" (p1)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p4], %[p4], 8     \n\t"
+                "srl        %[p3], %[p3], 8     \n\t"
+                "srl        %[p2], %[p2], 8     \n\t"
+                "srl        %[p1], %[p1], 8     \n\t"
+                : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p4],  1(%[s3])    \n\t"
+                "sb         %[p3],  0(%[s3])    \n\t"
+                "sb         %[p2], -1(%[s3])    \n\t"
+                "sb         %[p1], -2(%[s3])    \n\t"
+                : [p1] "+r" (p1)
+                : [p4] "r" (p4), [p3] "r" (p3), [s3] "r" (s3), [p2] "r" (p2)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p4], %[p4], 8     \n\t"
+                "srl        %[p3], %[p3], 8     \n\t"
+                "srl        %[p2], %[p2], 8     \n\t"
+                "srl        %[p1], %[p1], 8     \n\t"
+                : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p4],  1(%[s2])    \n\t"
+                "sb         %[p3],  0(%[s2])    \n\t"
+                "sb         %[p2], -1(%[s2])    \n\t"
+                "sb         %[p1], -2(%[s2])    \n\t"
+                :
+                : [p4] "r" (p4), [p3] "r" (p3), [s2] "r" (s2),
+                  [p2] "r" (p2), [p1] "r" (p1)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p4], %[p4], 8     \n\t"
+                "srl        %[p3], %[p3], 8     \n\t"
+                "srl        %[p2], %[p2], 8     \n\t"
+                "srl        %[p1], %[p1], 8     \n\t"
+                : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p4],  1(%[s1])    \n\t"
+                "sb         %[p3],  0(%[s1])    \n\t"
+                "sb         %[p2], -1(%[s1])    \n\t"
+                "sb         %[p1], -2(%[s1])    \n\t"
+                :
+                : [p4] "r" (p4), [p3] "r" (p3), [s1] "r" (s1), [p2] "r" (p2), [p1] "r" (p1)
+            );
+        }
+    }
+
+    s1 = s4 + p;
+    s2 = s1 + p;
+    s3 = s2 + p;
+    s4 = s3 + p;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p2  = *((uint32_t *)(s1 - 4));
+    p6  = *((uint32_t *)(s1));
+    p1  = *((uint32_t *)(s2 - 4));
+    p5  = *((uint32_t *)(s2));
+    p0  = *((uint32_t *)(s3 - 4));
+    p4  = *((uint32_t *)(s3));
+    pm1 = *((uint32_t *)(s4 - 4));
+    p3  = *((uint32_t *)(s4));
+
+    /* transpose pm1, p0, p1, p2 */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p2],      %[p1]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p2],      %[p1]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p0],      %[pm1]      \n\t"
+        "precr.qb.ph    %[prim4],   %[p0],      %[pm1]      \n\t"
+
+        "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[pm1],     %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p2],      %[p1],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p0],      %[pm1],     %[sec4]     \n\t"
+        "append         %[p1],      %[sec3],    16          \n\t"
+        "append         %[pm1],     %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0), [pm1] "+r" (pm1),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* transpose p3, p4, p5, p6 */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p6],      %[p5]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p6],      %[p5]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p4],      %[p3]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p4],      %[p3]       \n\t"
+
+        "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p6],      %[p5],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p4],      %[p3],      %[sec4]     \n\t"
+        "append         %[p5],      %[sec3],    16          \n\t"
+        "append         %[p3],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p6] "+r" (p6), [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_filter_mips(mask, hev, &p1, &p2, &p3, &p4);
+
+            /* unpack processed 4x4 neighborhood
+             * don't use transpose on output data
+             * because memory isn't aligned
+             */
+            __asm__ __volatile__ (
+                "sb         %[p4],  1(%[s4])    \n\t"
+                "sb         %[p3],  0(%[s4])    \n\t"
+                "sb         %[p2], -1(%[s4])    \n\t"
+                "sb         %[p1], -2(%[s4])    \n\t"
+                :
+                : [p4] "r" (p4), [p3] "r" (p3), [s4] "r" (s4),
+                  [p2] "r" (p2), [p1] "r" (p1)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p4], %[p4], 8     \n\t"
+                "srl        %[p3], %[p3], 8     \n\t"
+                "srl        %[p2], %[p2], 8     \n\t"
+                "srl        %[p1], %[p1], 8     \n\t"
+                : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p4],  1(%[s3])    \n\t"
+                "sb         %[p3],  0(%[s3])    \n\t"
+                "sb         %[p2], -1(%[s3])    \n\t"
+                "sb         %[p1], -2(%[s3])    \n\t"
+                : [p1] "+r" (p1)
+                : [p4] "r" (p4), [p3] "r" (p3), [s3] "r" (s3), [p2] "r" (p2)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p4], %[p4], 8     \n\t"
+                "srl        %[p3], %[p3], 8     \n\t"
+                "srl        %[p2], %[p2], 8     \n\t"
+                "srl        %[p1], %[p1], 8     \n\t"
+                : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p4],  1(%[s2])    \n\t"
+                "sb         %[p3],  0(%[s2])    \n\t"
+                "sb         %[p2], -1(%[s2])    \n\t"
+                "sb         %[p1], -2(%[s2])    \n\t"
+                :
+                : [p4] "r" (p4), [p3] "r" (p3), [s2] "r" (s2),
+                  [p2] "r" (p2), [p1] "r" (p1)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p4], %[p4], 8     \n\t"
+                "srl        %[p3], %[p3], 8     \n\t"
+                "srl        %[p2], %[p2], 8     \n\t"
+                "srl        %[p1], %[p1], 8     \n\t"
+                : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p4],  1(%[s1])    \n\t"
+                "sb         %[p3],  0(%[s1])    \n\t"
+                "sb         %[p2], -1(%[s1])    \n\t"
+                "sb         %[p1], -2(%[s1])    \n\t"
+                :
+                : [p4] "r" (p4), [p3] "r" (p3), [s1] "r" (s1),
+                  [p2] "r" (p2), [p1] "r" (p1)
+            );
+        }
+    }
+}
+
+/* inputs & outputs are quad-byte vectors */
+static __inline void vp8_mbfilter_mips
+(
+    uint32_t mask,
+    uint32_t hev,
+    uint32_t *ps2,
+    uint32_t *ps1,
+    uint32_t *ps0,
+    uint32_t *qs0,
+    uint32_t *qs1,
+    uint32_t *qs2
+)
+{
+    int32_t vps2, vps1, vps0, vqs0, vqs1, vqs2;
+    int32_t vps2_l, vps1_l, vps0_l, vqs0_l, vqs1_l, vqs2_l;
+    int32_t vps2_r, vps1_r, vps0_r, vqs0_r, vqs1_r, vqs2_r;
+    uint32_t HWM, vp8_filter_l, vp8_filter_r, mask_l, mask_r, hev_l, hev_r, subr_r, subr_l;
+    uint32_t Filter2_l, Filter2_r, t1, t2, Filter1_l, Filter1_r, invhev_l, invhev_r;
+    uint32_t N128, R63;
+    uint32_t u1_l, u1_r, u2_l, u2_r, u3_l, u3_r;
+
+    R63  = 0x003F003F;
+    HWM  = 0xFF00FF00;
+    N128 = 0x80808080;
+    t1   = 0x03000300;
+    t2   = 0x04000400;
+
+    vps0 = (*ps0) ^ N128;
+    vps1 = (*ps1) ^ N128;
+    vps2 = (*ps2) ^ N128;
+    vqs0 = (*qs0) ^ N128;
+    vqs1 = (*qs1) ^ N128;
+    vqs2 = (*qs2) ^ N128;
+
+    /* use halfword pairs instead quad-bytes because of accuracy */
+    vps0_l = vps0 & HWM;
+    vps0_r = vps0 << 8;
+    vps0_r = vps0_r & HWM;
+
+    vqs0_l = vqs0 & HWM;
+    vqs0_r = vqs0 << 8;
+    vqs0_r = vqs0_r & HWM;
+
+    vps1_l = vps1 & HWM;
+    vps1_r = vps1 << 8;
+    vps1_r = vps1_r & HWM;
+
+    vqs1_l = vqs1 & HWM;
+    vqs1_r = vqs1 << 8;
+    vqs1_r = vqs1_r & HWM;
+
+    vqs2_l = vqs2 & HWM;
+    vqs2_r = vqs2 << 8;
+    vqs2_r = vqs2_r & HWM;
+
+    __asm__ __volatile__ (
+        /* qs0 - ps0 */
+        "subq_s.ph    %[subr_l],       %[vqs0_l],       %[vps0_l]       \n\t"
+        "subq_s.ph    %[subr_r],       %[vqs0_r],       %[vps0_r]       \n\t"
+
+        /* vp8_filter = vp8_signed_char_clamp(ps1 - qs1); */
+        "subq_s.ph    %[vp8_filter_l], %[vps1_l],       %[vqs1_l]       \n\t"
+        "subq_s.ph    %[vp8_filter_r], %[vps1_r],       %[vqs1_r]       \n\t"
+
+        : [vp8_filter_l] "=&r" (vp8_filter_l), [vp8_filter_r] "=r" (vp8_filter_r),
+          [subr_l] "=&r" (subr_l), [subr_r] "=&r" (subr_r)
+        : [vps0_l] "r" (vps0_l), [vps0_r] "r" (vps0_r), [vps1_l] "r" (vps1_l),
+          [vps1_r] "r" (vps1_r), [vqs0_l] "r" (vqs0_l), [vqs0_r] "r" (vqs0_r),
+          [vqs1_l] "r" (vqs1_l), [vqs1_r] "r" (vqs1_r)
+    );
+
+    vps2_l = vps2 & HWM;
+    vps2_r = vps2 << 8;
+    vps2_r = vps2_r & HWM;
+
+    /* add outer taps if we have high edge variance */
+    __asm__ __volatile__ (
+        /* vp8_filter = vp8_signed_char_clamp(vp8_filter + 3 * (qs0 - ps0)); */
+        "addq_s.ph    %[vp8_filter_l], %[vp8_filter_l], %[subr_l]       \n\t"
+        "addq_s.ph    %[vp8_filter_r], %[vp8_filter_r], %[subr_r]       \n\t"
+        "and          %[mask_l],       %[HWM],          %[mask]         \n\t"
+        "sll          %[mask_r],       %[mask],         8               \n\t"
+        "and          %[mask_r],       %[HWM],          %[mask_r]       \n\t"
+        "addq_s.ph    %[vp8_filter_l], %[vp8_filter_l], %[subr_l]       \n\t"
+        "addq_s.ph    %[vp8_filter_r], %[vp8_filter_r], %[subr_r]       \n\t"
+        "and          %[hev_l],        %[HWM],          %[hev]          \n\t"
+        "sll          %[hev_r],        %[hev],          8               \n\t"
+        "and          %[hev_r],        %[HWM],          %[hev_r]        \n\t"
+        "addq_s.ph    %[vp8_filter_l], %[vp8_filter_l], %[subr_l]       \n\t"
+        "addq_s.ph    %[vp8_filter_r], %[vp8_filter_r], %[subr_r]       \n\t"
+
+        /* vp8_filter &= mask; */
+        "and          %[vp8_filter_l], %[vp8_filter_l], %[mask_l]       \n\t"
+        "and          %[vp8_filter_r], %[vp8_filter_r], %[mask_r]       \n\t"
+
+        /* Filter2 = vp8_filter & hev; */
+        "and          %[Filter2_l],    %[vp8_filter_l], %[hev_l]        \n\t"
+        "and          %[Filter2_r],    %[vp8_filter_r], %[hev_r]        \n\t"
+
+        : [vp8_filter_l] "+r" (vp8_filter_l), [vp8_filter_r] "+r" (vp8_filter_r),
+          [hev_l] "=&r" (hev_l), [hev_r] "=&r" (hev_r),
+          [mask_l] "=&r" (mask_l), [mask_r] "=&r" (mask_r),
+          [Filter2_l] "=&r" (Filter2_l), [Filter2_r] "=&r" (Filter2_r)
+        : [subr_l] "r" (subr_l), [subr_r] "r" (subr_r),
+          [HWM] "r" (HWM), [hev]  "r" (hev), [mask] "r" (mask)
+    );
+
+    /* save bottom 3 bits so that we round one side +4 and the other +3 */
+    __asm__ __volatile__ (
+        /* Filter1 = vp8_signed_char_clamp(Filter2 + 4) >>= 3; */
+        "addq_s.ph    %[Filter1_l],    %[Filter2_l],    %[t2]           \n\t"
+        "xor          %[invhev_l],     %[hev_l],        %[HWM]          \n\t"
+        "addq_s.ph    %[Filter1_r],    %[Filter2_r],    %[t2]           \n\t"
+
+        /* Filter2 = vp8_signed_char_clamp(Filter2 + 3) >>= 3; */
+        "addq_s.ph    %[Filter2_l],    %[Filter2_l],    %[t1]           \n\t"
+        "addq_s.ph    %[Filter2_r],    %[Filter2_r],    %[t1]           \n\t"
+
+        "shra.ph      %[Filter1_l],    %[Filter1_l],    3               \n\t"
+        "shra.ph      %[Filter1_r],    %[Filter1_r],    3               \n\t"
+
+        "shra.ph      %[Filter2_l],    %[Filter2_l],    3               \n\t"
+        "shra.ph      %[Filter2_r],    %[Filter2_r],    3               \n\t"
+        "and          %[Filter1_l],    %[Filter1_l],    %[HWM]          \n\t"
+        "and          %[Filter1_r],    %[Filter1_r],    %[HWM]          \n\t"
+        "xor          %[invhev_r],     %[hev_r],        %[HWM]          \n\t"
+
+        /* qs0 = vp8_signed_char_clamp(qs0 - Filter1); */
+        "subq_s.ph    %[vqs0_l],       %[vqs0_l],       %[Filter1_l]    \n\t"
+        "subq_s.ph    %[vqs0_r],       %[vqs0_r],       %[Filter1_r]    \n\t"
+
+        /* ps0 = vp8_signed_char_clamp(ps0 + Filter2); */
+        "addq_s.ph    %[vps0_l],       %[vps0_l],       %[Filter2_l]    \n\t"
+        "addq_s.ph    %[vps0_r],       %[vps0_r],       %[Filter2_r]    \n\t"
+
+        : [invhev_l] "=&r" (invhev_l), [invhev_r] "=&r" (invhev_r),
+          [Filter1_l] "=&r" (Filter1_l), [Filter1_r] "=&r" (Filter1_r),
+          [Filter2_l] "+r" (Filter2_l), [Filter2_r] "+r" (Filter2_r),
+          [vps0_l] "+r" (vps0_l), [vps0_r] "+r" (vps0_r),
+          [vqs0_l] "+r" (vqs0_l), [vqs0_r] "+r" (vqs0_r)
+        : [t1] "r" (t1), [t2] "r" (t2), [HWM] "r" (HWM),
+          [hev_l] "r" (hev_l), [hev_r] "r" (hev_r)
+    );
+
+    /* only apply wider filter if not high edge variance */
+    __asm__ __volatile__ (
+        /* vp8_filter &= ~hev; */
+        "and          %[Filter2_l],    %[vp8_filter_l], %[invhev_l]     \n\t"
+        "and          %[Filter2_r],    %[vp8_filter_r], %[invhev_r]     \n\t"
+
+        "shra.ph      %[Filter2_l],    %[Filter2_l],    8               \n\t"
+        "shra.ph      %[Filter2_r],    %[Filter2_r],    8               \n\t"
+
+        : [Filter2_l] "=&r" (Filter2_l), [Filter2_r] "=&r" (Filter2_r)
+        : [vp8_filter_l] "r" (vp8_filter_l), [vp8_filter_r] "r" (vp8_filter_r),
+          [invhev_l] "r" (invhev_l), [invhev_r] "r" (invhev_r)
+    );
+
+    /* roughly 3/7th difference across boundary */
+    __asm__ __volatile__ (
+        "shll.ph      %[u3_l],         %[Filter2_l],    3               \n\t"
+        "shll.ph      %[u3_r],         %[Filter2_r],    3               \n\t"
+
+        "addq.ph      %[u3_l],         %[u3_l],         %[Filter2_l]    \n\t"
+        "addq.ph      %[u3_r],         %[u3_r],         %[Filter2_r]    \n\t"
+
+        "shll.ph      %[u2_l],         %[u3_l],         1               \n\t"
+        "shll.ph      %[u2_r],         %[u3_r],         1               \n\t"
+
+        "addq.ph      %[u1_l],         %[u3_l],         %[u2_l]         \n\t"
+        "addq.ph      %[u1_r],         %[u3_r],         %[u2_r]         \n\t"
+
+        "addq.ph      %[u2_l],         %[u2_l],         %[R63]          \n\t"
+        "addq.ph      %[u2_r],         %[u2_r],         %[R63]          \n\t"
+
+        "addq.ph      %[u3_l],         %[u3_l],         %[R63]          \n\t"
+        "addq.ph      %[u3_r],         %[u3_r],         %[R63]          \n\t"
+
+        /* vp8_signed_char_clamp((63 + Filter2 * 27) >> 7)
+         * vp8_signed_char_clamp((63 + Filter2 * 18) >> 7)
+         */
+        "addq.ph      %[u1_l],         %[u1_l],         %[R63]          \n\t"
+        "addq.ph      %[u1_r],         %[u1_r],         %[R63]          \n\t"
+        "shra.ph      %[u1_l],         %[u1_l],         7               \n\t"
+        "shra.ph      %[u1_r],         %[u1_r],         7               \n\t"
+        "shra.ph      %[u2_l],         %[u2_l],         7               \n\t"
+        "shra.ph      %[u2_r],         %[u2_r],         7               \n\t"
+        "shll.ph      %[u1_l],         %[u1_l],         8               \n\t"
+        "shll.ph      %[u1_r],         %[u1_r],         8               \n\t"
+        "shll.ph      %[u2_l],         %[u2_l],         8               \n\t"
+        "shll.ph      %[u2_r],         %[u2_r],         8               \n\t"
+
+        /* vqs0 = vp8_signed_char_clamp(qs0 - u); */
+        "subq_s.ph    %[vqs0_l],       %[vqs0_l],       %[u1_l]         \n\t"
+        "subq_s.ph    %[vqs0_r],       %[vqs0_r],       %[u1_r]         \n\t"
+
+        /* vps0 = vp8_signed_char_clamp(ps0 + u); */
+        "addq_s.ph    %[vps0_l],       %[vps0_l],       %[u1_l]         \n\t"
+        "addq_s.ph    %[vps0_r],       %[vps0_r],       %[u1_r]         \n\t"
+
+        : [u1_l] "=&r" (u1_l), [u1_r] "=&r" (u1_r), [u2_l] "=&r" (u2_l),
+          [u2_r] "=&r" (u2_r), [u3_l] "=&r" (u3_l), [u3_r] "=&r" (u3_r),
+          [vps0_l] "+r" (vps0_l), [vps0_r] "+r" (vps0_r),
+          [vqs0_l] "+r" (vqs0_l), [vqs0_r] "+r" (vqs0_r)
+        : [R63]  "r" (R63),
+          [Filter2_l] "r" (Filter2_l), [Filter2_r] "r" (Filter2_r)
+    );
+
+    __asm__ __volatile__ (
+        /* vqs1 = vp8_signed_char_clamp(qs1 - u); */
+        "subq_s.ph    %[vqs1_l],       %[vqs1_l],       %[u2_l]         \n\t"
+        "addq_s.ph    %[vps1_l],       %[vps1_l],       %[u2_l]         \n\t"
+
+        /* vps1 = vp8_signed_char_clamp(ps1 + u); */
+        "addq_s.ph    %[vps1_r],       %[vps1_r],       %[u2_r]         \n\t"
+        "subq_s.ph    %[vqs1_r],       %[vqs1_r],       %[u2_r]         \n\t"
+
+        : [vps1_l] "+r" (vps1_l), [vps1_r] "+r" (vps1_r),
+          [vqs1_l] "+r" (vqs1_l), [vqs1_r] "+r" (vqs1_r)
+        : [u2_l] "r" (u2_l), [u2_r] "r" (u2_r)
+    );
+
+    /* roughly 1/7th difference across boundary */
+    __asm__ __volatile__ (
+        /* u = vp8_signed_char_clamp((63 + Filter2 * 9) >> 7); */
+        "shra.ph      %[u3_l],         %[u3_l],         7               \n\t"
+        "shra.ph      %[u3_r],         %[u3_r],         7               \n\t"
+        "shll.ph      %[u3_l],         %[u3_l],         8               \n\t"
+        "shll.ph      %[u3_r],         %[u3_r],         8               \n\t"
+
+        /* vqs2 = vp8_signed_char_clamp(qs2 - u); */
+        "subq_s.ph    %[vqs2_l],       %[vqs2_l],       %[u3_l]         \n\t"
+        "subq_s.ph    %[vqs2_r],       %[vqs2_r],       %[u3_r]         \n\t"
+
+        /* vps2 = vp8_signed_char_clamp(ps2 + u); */
+        "addq_s.ph    %[vps2_l],       %[vps2_l],       %[u3_l]         \n\t"
+        "addq_s.ph    %[vps2_r],       %[vps2_r],       %[u3_r]         \n\t"
+
+        : [u3_l] "+r" (u3_l), [u3_r] "+r" (u3_r), [vps2_l] "+r" (vps2_l),
+          [vps2_r] "+r" (vps2_r), [vqs2_l] "+r" (vqs2_l), [vqs2_r] "+r" (vqs2_r)
+        :
+    );
+
+    /* Create quad-bytes from halfword pairs */
+    __asm__ __volatile__ (
+        "and          %[vqs0_l],       %[vqs0_l],       %[HWM]          \n\t"
+        "shrl.ph      %[vqs0_r],       %[vqs0_r],       8               \n\t"
+
+        "and          %[vps0_l],       %[vps0_l],       %[HWM]          \n\t"
+        "shrl.ph      %[vps0_r],       %[vps0_r],       8               \n\t"
+
+        "and          %[vqs1_l],       %[vqs1_l],       %[HWM]          \n\t"
+        "shrl.ph      %[vqs1_r],       %[vqs1_r],       8               \n\t"
+
+        "and          %[vps1_l],       %[vps1_l],       %[HWM]          \n\t"
+        "shrl.ph      %[vps1_r],       %[vps1_r],       8               \n\t"
+
+        "and          %[vqs2_l],       %[vqs2_l],       %[HWM]          \n\t"
+        "shrl.ph      %[vqs2_r],       %[vqs2_r],       8               \n\t"
+
+        "and          %[vps2_l],       %[vps2_l],       %[HWM]          \n\t"
+        "shrl.ph      %[vps2_r],       %[vps2_r],       8               \n\t"
+
+        "or           %[vqs0_r],       %[vqs0_l],       %[vqs0_r]       \n\t"
+        "or           %[vps0_r],       %[vps0_l],       %[vps0_r]       \n\t"
+        "or           %[vqs1_r],       %[vqs1_l],       %[vqs1_r]       \n\t"
+        "or           %[vps1_r],       %[vps1_l],       %[vps1_r]       \n\t"
+        "or           %[vqs2_r],       %[vqs2_l],       %[vqs2_r]       \n\t"
+        "or           %[vps2_r],       %[vps2_l],       %[vps2_r]       \n\t"
+
+        : [vps1_l] "+r" (vps1_l), [vps1_r] "+r" (vps1_r), [vqs1_l] "+r" (vqs1_l),
+          [vqs1_r] "+r" (vqs1_r), [vps0_l] "+r" (vps0_l), [vps0_r] "+r" (vps0_r),
+          [vqs0_l] "+r" (vqs0_l), [vqs0_r] "+r" (vqs0_r), [vqs2_l] "+r" (vqs2_l),
+          [vqs2_r] "+r" (vqs2_r), [vps2_r] "+r" (vps2_r), [vps2_l] "+r" (vps2_l)
+        : [HWM] "r" (HWM)
+    );
+
+    *ps0 = vps0_r ^ N128;
+    *ps1 = vps1_r ^ N128;
+    *ps2 = vps2_r ^ N128;
+    *qs0 = vqs0_r ^ N128;
+    *qs1 = vqs1_r ^ N128;
+    *qs2 = vqs2_r ^ N128;
+}
+
+void vp8_mbloop_filter_horizontal_edge_mips
+(
+    unsigned char *s,
+    int p,
+    unsigned int flimit,
+    unsigned int limit,
+    unsigned int thresh,
+    int count
+)
+{
+    int i;
+    uint32_t mask, hev;
+    uint32_t pm1, p0, p1, p2, p3, p4, p5, p6;
+    unsigned char *sm1, *s0, *s1, *s2, *s3, *s4, *s5, *s6;
+
+    mask = 0;
+    hev = 0;
+    i = 0;
+    p1 = 0;
+    p2 = 0;
+    p3 = 0;
+    p4 = 0;
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+
+    sm1 = s - (p << 2);
+    s0  = s - p - p - p;
+    s1  = s - p - p;
+    s2  = s - p;
+    s3  = s;
+    s4  = s + p;
+    s5  = s + p + p;
+    s6  = s + p + p + p;
+
+    /* prefetch data for load */
+    prefetch_load_lf(s + p);
+
+    /* apply filter on 4 pixesl at the same time */
+    do
+    {
+        /* load quad-byte vectors
+         * memory is 4 byte aligned
+         */
+        p1 = *((uint32_t *)(s1));
+        p2 = *((uint32_t *)(s2));
+        p3 = *((uint32_t *)(s3));
+        p4 = *((uint32_t *)(s4));
+
+        /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+         * mask will be zero and filtering is not needed
+         */
+        if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+        {
+
+            pm1 = *((uint32_t *)(sm1));
+            p0  = *((uint32_t *)(s0));
+            p5  = *((uint32_t *)(s5));
+            p6  = *((uint32_t *)(s6));
+
+            vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                     thresh, &hev, &mask);
+
+            /* if mask == 0 do filtering is not needed */
+            if (mask)
+            {
+                /* filtering */
+                vp8_mbfilter_mips(mask, hev, &p0, &p1, &p2, &p3, &p4, &p5);
+
+                /* unpack processed 4x4 neighborhood
+                 * memory is 4 byte aligned
+                 */
+                *((uint32_t *)s0) = p0;
+                *((uint32_t *)s1) = p1;
+                *((uint32_t *)s2) = p2;
+                *((uint32_t *)s3) = p3;
+                *((uint32_t *)s4) = p4;
+                *((uint32_t *)s5) = p5;
+            }
+        }
+
+        sm1 += 4;
+        s0  += 4;
+        s1  += 4;
+        s2  += 4;
+        s3  += 4;
+        s4  += 4;
+        s5  += 4;
+        s6  += 4;
+
+        /* load quad-byte vectors
+         * memory is 4 byte aligned
+         */
+        p1 = *((uint32_t *)(s1));
+        p2 = *((uint32_t *)(s2));
+        p3 = *((uint32_t *)(s3));
+        p4 = *((uint32_t *)(s4));
+
+        /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+         * mask will be zero and filtering is not needed
+         */
+        if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+        {
+
+            pm1 = *((uint32_t *)(sm1));
+            p0  = *((uint32_t *)(s0));
+            p5  = *((uint32_t *)(s5));
+            p6  = *((uint32_t *)(s6));
+
+            vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                     thresh, &hev, &mask);
+
+            /* if mask == 0 do filtering is not needed */
+            if (mask)
+            {
+                /* filtering */
+                vp8_mbfilter_mips(mask, hev, &p0, &p1, &p2, &p3, &p4, &p5);
+
+                /* unpack processed 4x4 neighborhood
+                 * memory is 4 byte aligned
+                 */
+                *((uint32_t *)s0) = p0;
+                *((uint32_t *)s1) = p1;
+                *((uint32_t *)s2) = p2;
+                *((uint32_t *)s3) = p3;
+                *((uint32_t *)s4) = p4;
+                *((uint32_t *)s5) = p5;
+            }
+        }
+
+        sm1 += 4;
+        s0  += 4;
+        s1  += 4;
+        s2  += 4;
+        s3  += 4;
+        s4  += 4;
+        s5  += 4;
+        s6  += 4;
+
+        i += 8;
+    }
+
+    while (i < count);
+}
+
+void vp8_mbloop_filter_uvhorizontal_edge_mips
+(
+    unsigned char *s,
+    int p,
+    unsigned int flimit,
+    unsigned int limit,
+    unsigned int thresh,
+    int count
+)
+{
+    uint32_t mask, hev;
+    uint32_t pm1, p0, p1, p2, p3, p4, p5, p6;
+    unsigned char *sm1, *s0, *s1, *s2, *s3, *s4, *s5, *s6;
+
+    mask = 0;
+    hev = 0;
+    p1 = 0;
+    p2 = 0;
+    p3 = 0;
+    p4 = 0;
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+
+    sm1 = s - (p << 2);
+    s0  = s - p - p - p;
+    s1  = s - p - p;
+    s2  = s - p;
+    s3  = s;
+    s4  = s + p;
+    s5  = s + p + p;
+    s6  = s + p + p + p;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p1 = *((uint32_t *)(s1));
+    p2 = *((uint32_t *)(s2));
+    p3 = *((uint32_t *)(s3));
+    p4 = *((uint32_t *)(s4));
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        pm1 = *((uint32_t *)(sm1));
+        p0  = *((uint32_t *)(s0));
+        p5  = *((uint32_t *)(s5));
+        p6  = *((uint32_t *)(s6));
+
+        /* if mask == 0 do filtering is not needed */
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        if (mask)
+        {
+            /* filtering */
+            vp8_mbfilter_mips(mask, hev, &p0, &p1, &p2, &p3, &p4, &p5);
+
+            /* unpack processed 4x4 neighborhood
+             * memory is 4 byte aligned
+             */
+            *((uint32_t *)s0) = p0;
+            *((uint32_t *)s1) = p1;
+            *((uint32_t *)s2) = p2;
+            *((uint32_t *)s3) = p3;
+            *((uint32_t *)s4) = p4;
+            *((uint32_t *)s5) = p5;
+        }
+    }
+
+    sm1 += 4;
+    s0  += 4;
+    s1  += 4;
+    s2  += 4;
+    s3  += 4;
+    s4  += 4;
+    s5  += 4;
+    s6  += 4;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p1 = *((uint32_t *)(s1));
+    p2 = *((uint32_t *)(s2));
+    p3 = *((uint32_t *)(s3));
+    p4 = *((uint32_t *)(s4));
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        pm1 = *((uint32_t *)(sm1));
+        p0  = *((uint32_t *)(s0));
+        p5  = *((uint32_t *)(s5));
+        p6  = *((uint32_t *)(s6));
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_mbfilter_mips(mask, hev, &p0, &p1, &p2, &p3, &p4, &p5);
+
+            /* unpack processed 4x4 neighborhood
+             * memory is 4 byte aligned
+             */
+            *((uint32_t *)s0) = p0;
+            *((uint32_t *)s1) = p1;
+            *((uint32_t *)s2) = p2;
+            *((uint32_t *)s3) = p3;
+            *((uint32_t *)s4) = p4;
+            *((uint32_t *)s5) = p5;
+        }
+    }
+}
+
+
+void vp8_mbloop_filter_vertical_edge_mips
+(
+    unsigned char *s,
+    int p,
+    unsigned int flimit,
+    unsigned int limit,
+    unsigned int thresh,
+    int count
+)
+{
+
+    int i;
+    uint32_t mask, hev;
+    uint32_t pm1, p0, p1, p2, p3, p4, p5, p6;
+    unsigned char *s1, *s2, *s3, *s4;
+    uint32_t prim1, prim2, sec3, sec4, prim3, prim4;
+
+    mask = 0;
+    hev = 0;
+    i = 0;
+    pm1 = 0;
+    p0 = 0;
+    p1 = 0;
+    p2 = 0;
+    p3 = 0;
+    p4 = 0;
+    p5 = 0;
+    p6 = 0;
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+
+    /* apply filter on 4 pixesl at the same time */
+    do
+    {
+        s1 = s;
+        s2 = s + p;
+        s3 = s2 + p;
+        s4 = s3 + p;
+        s  = s4 + p;
+
+        /* load quad-byte vectors
+         * memory is 4 byte aligned
+         */
+        p2  = *((uint32_t *)(s1 - 4));
+        p6  = *((uint32_t *)(s1));
+        p1  = *((uint32_t *)(s2 - 4));
+        p5  = *((uint32_t *)(s2));
+        p0  = *((uint32_t *)(s3 - 4));
+        p4  = *((uint32_t *)(s3));
+        pm1 = *((uint32_t *)(s4 - 4));
+        p3  = *((uint32_t *)(s4));
+
+        /* transpose pm1, p0, p1, p2 */
+        __asm__ __volatile__ (
+            "precrq.qb.ph   %[prim1],   %[p2],      %[p1]       \n\t"
+            "precr.qb.ph    %[prim2],   %[p2],      %[p1]       \n\t"
+            "precrq.qb.ph   %[prim3],   %[p0],      %[pm1]      \n\t"
+            "precr.qb.ph    %[prim4],   %[p0],      %[pm1]      \n\t"
+
+            "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+            "precr.qb.ph    %[pm1],     %[prim1],   %[prim2]    \n\t"
+            "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+            "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+            "precrq.ph.w    %[p2],      %[p1],      %[sec3]     \n\t"
+            "precrq.ph.w    %[p0],      %[pm1],     %[sec4]     \n\t"
+            "append         %[p1],      %[sec3],    16          \n\t"
+            "append         %[pm1],     %[sec4],    16          \n\t"
+
+            : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+              [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+              [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0), [pm1] "+r" (pm1),
+              [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+            :
+        );
+
+        /* transpose p3, p4, p5, p6 */
+        __asm__ __volatile__ (
+            "precrq.qb.ph   %[prim1],   %[p6],      %[p5]       \n\t"
+            "precr.qb.ph    %[prim2],   %[p6],      %[p5]       \n\t"
+            "precrq.qb.ph   %[prim3],   %[p4],      %[p3]       \n\t"
+            "precr.qb.ph    %[prim4],   %[p4],      %[p3]       \n\t"
+
+            "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+            "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+            "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+            "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+            "precrq.ph.w    %[p6],      %[p5],      %[sec3]     \n\t"
+            "precrq.ph.w    %[p4],      %[p3],      %[sec4]     \n\t"
+            "append         %[p5],      %[sec3],    16          \n\t"
+            "append         %[p3],      %[sec4],    16          \n\t"
+
+            : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+              [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+              [p6] "+r" (p6), [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+              [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+            :
+        );
+
+        /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+         * mask will be zero and filtering is not needed
+         */
+        if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+        {
+
+            vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                     thresh, &hev, &mask);
+
+            /* if mask == 0 do filtering is not needed */
+            if (mask)
+            {
+                /* filtering */
+                vp8_mbfilter_mips(mask, hev, &p0, &p1, &p2, &p3, &p4, &p5);
+
+                /* don't use transpose on output data
+                 * because memory isn't aligned
+                 */
+                __asm__ __volatile__ (
+                    "sb         %[p5],  2(%[s4])        \n\t"
+                    "sb         %[p4],  1(%[s4])        \n\t"
+                    "sb         %[p3],  0(%[s4])        \n\t"
+                    "sb         %[p2], -1(%[s4])        \n\t"
+                    "sb         %[p1], -2(%[s4])        \n\t"
+                    "sb         %[p0], -3(%[s4])        \n\t"
+                    :
+                    : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s4] "r" (s4),
+                      [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+                );
+
+                __asm__ __volatile__ (
+                    "srl        %[p5], %[p5], 8         \n\t"
+                    "srl        %[p4], %[p4], 8         \n\t"
+                    "srl        %[p3], %[p3], 8         \n\t"
+                    "srl        %[p2], %[p2], 8         \n\t"
+                    "srl        %[p1], %[p1], 8         \n\t"
+                    "srl        %[p0], %[p0], 8         \n\t"
+                    : [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+                      [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0)
+                    :
+                );
+
+                __asm__ __volatile__ (
+                    "sb         %[p5],  2(%[s3])        \n\t"
+                    "sb         %[p4],  1(%[s3])        \n\t"
+                    "sb         %[p3],  0(%[s3])        \n\t"
+                    "sb         %[p2], -1(%[s3])        \n\t"
+                    "sb         %[p1], -2(%[s3])        \n\t"
+                    "sb         %[p0], -3(%[s3])        \n\t"
+                    :
+                    : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s3] "r" (s3),
+                      [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+                );
+
+                __asm__ __volatile__ (
+                    "srl        %[p5], %[p5], 8         \n\t"
+                    "srl        %[p4], %[p4], 8         \n\t"
+                    "srl        %[p3], %[p3], 8         \n\t"
+                    "srl        %[p2], %[p2], 8         \n\t"
+                    "srl        %[p1], %[p1], 8         \n\t"
+                    "srl        %[p0], %[p0], 8         \n\t"
+                    : [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+                      [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0)
+                    :
+                );
+
+                __asm__ __volatile__ (
+                    "sb         %[p5],  2(%[s2])        \n\t"
+                    "sb         %[p4],  1(%[s2])        \n\t"
+                    "sb         %[p3],  0(%[s2])        \n\t"
+                    "sb         %[p2], -1(%[s2])        \n\t"
+                    "sb         %[p1], -2(%[s2])        \n\t"
+                    "sb         %[p0], -3(%[s2])        \n\t"
+                    :
+                    : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s2] "r" (s2),
+                      [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+                );
+
+                __asm__ __volatile__ (
+                    "srl        %[p5], %[p5], 8         \n\t"
+                    "srl        %[p4], %[p4], 8         \n\t"
+                    "srl        %[p3], %[p3], 8         \n\t"
+                    "srl        %[p2], %[p2], 8         \n\t"
+                    "srl        %[p1], %[p1], 8         \n\t"
+                    "srl        %[p0], %[p0], 8         \n\t"
+                    : [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+                      [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0)
+                    :
+                );
+
+                __asm__ __volatile__ (
+                    "sb         %[p5],  2(%[s1])        \n\t"
+                    "sb         %[p4],  1(%[s1])        \n\t"
+                    "sb         %[p3],  0(%[s1])        \n\t"
+                    "sb         %[p2], -1(%[s1])        \n\t"
+                    "sb         %[p1], -2(%[s1])        \n\t"
+                    "sb         %[p0], -3(%[s1])        \n\t"
+                    :
+                    : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s1] "r" (s1),
+                      [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+                );
+            }
+        }
+
+        i += 4;
+    }
+
+    while (i < count);
+}
+
+void vp8_mbloop_filter_uvvertical_edge_mips
+(
+    unsigned char *s,
+    int p,
+    unsigned int flimit,
+    unsigned int limit,
+    unsigned int thresh,
+    int count
+)
+{
+    uint32_t mask, hev;
+    uint32_t pm1, p0, p1, p2, p3, p4, p5, p6;
+    unsigned char *s1, *s2, *s3, *s4;
+    uint32_t prim1, prim2, sec3, sec4, prim3, prim4;
+
+    mask = 0;
+    hev = 0;
+    pm1 = 0;
+    p0 = 0;
+    p1 = 0;
+    p2 = 0;
+    p3 = 0;
+    p4 = 0;
+    p5 = 0;
+    p6 = 0;
+
+    /* loop filter designed to work using chars so that we can make maximum use
+     * of 8 bit simd instructions.
+     */
+
+    /* apply filter on 4 pixesl at the same time */
+
+    s1 = s;
+    s2 = s + p;
+    s3 = s2 + p;
+    s4 = s3 + p;
+
+    /* prefetch data for load */
+    prefetch_load_lf(s + 2 * p);
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p2  = *((uint32_t *)(s1 - 4));
+    p6  = *((uint32_t *)(s1));
+    p1  = *((uint32_t *)(s2 - 4));
+    p5  = *((uint32_t *)(s2));
+    p0  = *((uint32_t *)(s3 - 4));
+    p4  = *((uint32_t *)(s3));
+    pm1 = *((uint32_t *)(s4 - 4));
+    p3  = *((uint32_t *)(s4));
+
+    /* transpose pm1, p0, p1, p2 */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p2],      %[p1]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p2],      %[p1]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p0],      %[pm1]      \n\t"
+        "precr.qb.ph    %[prim4],   %[p0],      %[pm1]      \n\t"
+
+        "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[pm1],     %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p2],      %[p1],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p0],      %[pm1],     %[sec4]     \n\t"
+        "append         %[p1],      %[sec3],    16          \n\t"
+        "append         %[pm1],     %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0), [pm1] "+r" (pm1),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* transpose p3, p4, p5, p6 */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p6],      %[p5]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p6],      %[p5]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p4],      %[p3]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p4],      %[p3]       \n\t"
+
+        "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p6],      %[p5],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p4],      %[p3],      %[sec4]     \n\t"
+        "append         %[p5],      %[sec3],    16          \n\t"
+        "append         %[p3],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p6] "+r" (p6), [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6,
+                                 thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_mbfilter_mips(mask, hev, &p0, &p1, &p2, &p3, &p4, &p5);
+
+            /* don't use transpose on output data
+             * because memory isn't aligned
+             */
+            __asm__ __volatile__ (
+                "sb         %[p5],  2(%[s4])        \n\t"
+                "sb         %[p4],  1(%[s4])        \n\t"
+                "sb         %[p3],  0(%[s4])        \n\t"
+                "sb         %[p2], -1(%[s4])        \n\t"
+                "sb         %[p1], -2(%[s4])        \n\t"
+                "sb         %[p0], -3(%[s4])        \n\t"
+                :
+                : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s4] "r" (s4),
+                  [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p5], %[p5], 8         \n\t"
+                "srl        %[p4], %[p4], 8         \n\t"
+                "srl        %[p3], %[p3], 8         \n\t"
+                "srl        %[p2], %[p2], 8         \n\t"
+                "srl        %[p1], %[p1], 8         \n\t"
+                "srl        %[p0], %[p0], 8         \n\t"
+                : [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+                  [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p5],  2(%[s3])        \n\t"
+                "sb         %[p4],  1(%[s3])        \n\t"
+                "sb         %[p3],  0(%[s3])        \n\t"
+                "sb         %[p2], -1(%[s3])        \n\t"
+                "sb         %[p1], -2(%[s3])        \n\t"
+                "sb         %[p0], -3(%[s3])        \n\t"
+                :
+                : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s3] "r" (s3),
+                  [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p5], %[p5], 8         \n\t"
+                "srl        %[p4], %[p4], 8         \n\t"
+                "srl        %[p3], %[p3], 8         \n\t"
+                "srl        %[p2], %[p2], 8         \n\t"
+                "srl        %[p1], %[p1], 8         \n\t"
+                "srl        %[p0], %[p0], 8         \n\t"
+                : [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+                  [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p5],  2(%[s2])        \n\t"
+                "sb         %[p4],  1(%[s2])        \n\t"
+                "sb         %[p3],  0(%[s2])        \n\t"
+                "sb         %[p2], -1(%[s2])        \n\t"
+                "sb         %[p1], -2(%[s2])        \n\t"
+                "sb         %[p0], -3(%[s2])        \n\t"
+                :
+                : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s2] "r" (s2),
+                  [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p5], %[p5], 8         \n\t"
+                "srl        %[p4], %[p4], 8         \n\t"
+                "srl        %[p3], %[p3], 8         \n\t"
+                "srl        %[p2], %[p2], 8         \n\t"
+                "srl        %[p1], %[p1], 8         \n\t"
+                "srl        %[p0], %[p0], 8         \n\t"
+                : [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+                  [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p5],  2(%[s1])        \n\t"
+                "sb         %[p4],  1(%[s1])        \n\t"
+                "sb         %[p3],  0(%[s1])        \n\t"
+                "sb         %[p2], -1(%[s1])        \n\t"
+                "sb         %[p1], -2(%[s1])        \n\t"
+                "sb         %[p0], -3(%[s1])        \n\t"
+                :
+                : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s1] "r" (s1),
+                  [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+            );
+        }
+    }
+
+    s1 = s4 + p;
+    s2 = s1 + p;
+    s3 = s2 + p;
+    s4 = s3 + p;
+
+    /* load quad-byte vectors
+    * memory is 4 byte aligned
+    */
+    p2  = *((uint32_t *)(s1 - 4));
+    p6  = *((uint32_t *)(s1));
+    p1  = *((uint32_t *)(s2 - 4));
+    p5  = *((uint32_t *)(s2));
+    p0  = *((uint32_t *)(s3 - 4));
+    p4  = *((uint32_t *)(s3));
+    pm1 = *((uint32_t *)(s4 - 4));
+    p3  = *((uint32_t *)(s4));
+
+    /* transpose pm1, p0, p1, p2 */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p2],      %[p1]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p2],      %[p1]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p0],      %[pm1]      \n\t"
+        "precr.qb.ph    %[prim4],   %[p0],      %[pm1]      \n\t"
+
+        "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[pm1],     %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p2],      %[p1],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p0],      %[pm1],     %[sec4]     \n\t"
+        "append         %[p1],      %[sec3],    16          \n\t"
+        "append         %[pm1],     %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0), [pm1] "+r" (pm1),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* transpose p3, p4, p5, p6 */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p6],      %[p5]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p6],      %[p5]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p4],      %[p3]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p4],      %[p3]       \n\t"
+
+        "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p6],      %[p5],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p4],      %[p3],      %[sec4]     \n\t"
+        "append         %[p5],      %[sec3],    16          \n\t"
+        "append         %[p3],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p6] "+r" (p6), [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0)))
+    {
+
+        vp8_filter_mask_vec_mips(limit, flimit, p1, p2, pm1, p0, p3, p4, p5, p6, thresh, &hev, &mask);
+
+        /* if mask == 0 do filtering is not needed */
+        if (mask)
+        {
+            /* filtering */
+            vp8_mbfilter_mips(mask, hev, &p0, &p1, &p2, &p3, &p4, &p5);
+
+            /* don't use transpose on output data
+             * because memory isn't aligned
+             */
+            __asm__ __volatile__ (
+                "sb         %[p5],  2(%[s4])        \n\t"
+                "sb         %[p4],  1(%[s4])        \n\t"
+                "sb         %[p3],  0(%[s4])        \n\t"
+                "sb         %[p2], -1(%[s4])        \n\t"
+                "sb         %[p1], -2(%[s4])        \n\t"
+                "sb         %[p0], -3(%[s4])        \n\t"
+                :
+                : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s4] "r" (s4),
+                  [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p5], %[p5], 8         \n\t"
+                "srl        %[p4], %[p4], 8         \n\t"
+                "srl        %[p3], %[p3], 8         \n\t"
+                "srl        %[p2], %[p2], 8         \n\t"
+                "srl        %[p1], %[p1], 8         \n\t"
+                "srl        %[p0], %[p0], 8         \n\t"
+                : [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+                  [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p5],  2(%[s3])        \n\t"
+                "sb         %[p4],  1(%[s3])        \n\t"
+                "sb         %[p3],  0(%[s3])        \n\t"
+                "sb         %[p2], -1(%[s3])        \n\t"
+                "sb         %[p1], -2(%[s3])        \n\t"
+                "sb         %[p0], -3(%[s3])        \n\t"
+                :
+                : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s3] "r" (s3),
+                  [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p5], %[p5], 8         \n\t"
+                "srl        %[p4], %[p4], 8         \n\t"
+                "srl        %[p3], %[p3], 8         \n\t"
+                "srl        %[p2], %[p2], 8         \n\t"
+                "srl        %[p1], %[p1], 8         \n\t"
+                "srl        %[p0], %[p0], 8         \n\t"
+                : [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+                  [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p5],  2(%[s2])        \n\t"
+                "sb         %[p4],  1(%[s2])        \n\t"
+                "sb         %[p3],  0(%[s2])        \n\t"
+                "sb         %[p2], -1(%[s2])        \n\t"
+                "sb         %[p1], -2(%[s2])        \n\t"
+                "sb         %[p0], -3(%[s2])        \n\t"
+                :
+                : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s2] "r" (s2),
+                  [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+            );
+
+            __asm__ __volatile__ (
+                "srl        %[p5], %[p5], 8         \n\t"
+                "srl        %[p4], %[p4], 8         \n\t"
+                "srl        %[p3], %[p3], 8         \n\t"
+                "srl        %[p2], %[p2], 8         \n\t"
+                "srl        %[p1], %[p1], 8         \n\t"
+                "srl        %[p0], %[p0], 8         \n\t"
+                : [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+                  [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0)
+                :
+            );
+
+            __asm__ __volatile__ (
+                "sb         %[p5],  2(%[s1])        \n\t"
+                "sb         %[p4],  1(%[s1])        \n\t"
+                "sb         %[p3],  0(%[s1])        \n\t"
+                "sb         %[p2], -1(%[s1])        \n\t"
+                "sb         %[p1], -2(%[s1])        \n\t"
+                "sb         %[p0], -3(%[s1])        \n\t"
+                :
+                : [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3), [s1] "r" (s1),
+                  [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0)
+            );
+        }
+    }
+}
+
+/* Horizontal MB filtering */
+void vp8_loop_filter_mbh_dspr2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                               int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    unsigned int thresh_vec, flimit_vec, limit_vec;
+    unsigned char thresh, flimit, limit, flimit_temp;
+
+    /* use direct value instead pointers */
+    limit = *(lfi->lim);
+    flimit_temp = *(lfi->mblim);
+    thresh = *(lfi->hev_thr);
+    flimit = flimit_temp;
+
+    /* create quad-byte */
+    __asm__ __volatile__ (
+        "replv.qb       %[thresh_vec], %[thresh]    \n\t"
+        "replv.qb       %[flimit_vec], %[flimit]    \n\t"
+        "replv.qb       %[limit_vec],  %[limit]     \n\t"
+        : [thresh_vec] "=&r" (thresh_vec), [flimit_vec] "=&r" (flimit_vec), [limit_vec] "=r" (limit_vec)
+        : [thresh] "r" (thresh), [flimit] "r" (flimit), [limit] "r" (limit)
+    );
+
+    vp8_mbloop_filter_horizontal_edge_mips(y_ptr, y_stride, flimit_vec, limit_vec, thresh_vec, 16);
+
+    if (u_ptr)
+    {
+        vp8_mbloop_filter_uvhorizontal_edge_mips(u_ptr, uv_stride, flimit_vec, limit_vec, thresh_vec, 0);
+    }
+
+    if (v_ptr)
+    {
+        vp8_mbloop_filter_uvhorizontal_edge_mips(v_ptr, uv_stride, flimit_vec, limit_vec, thresh_vec, 0);
+    }
+}
+
+
+/* Vertical MB Filtering */
+void vp8_loop_filter_mbv_dspr2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                               int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    unsigned int thresh_vec, flimit_vec, limit_vec;
+    unsigned char thresh, flimit, limit, flimit_temp;
+
+    /* use direct value instead pointers */
+    limit = *(lfi->lim);
+    flimit_temp = *(lfi->mblim);
+    thresh = *(lfi->hev_thr);
+    flimit = flimit_temp;
+
+    /* create quad-byte */
+    __asm__ __volatile__ (
+        "replv.qb       %[thresh_vec], %[thresh]    \n\t"
+        "replv.qb       %[flimit_vec], %[flimit]    \n\t"
+        "replv.qb       %[limit_vec],  %[limit]     \n\t"
+        : [thresh_vec] "=&r" (thresh_vec), [flimit_vec] "=&r" (flimit_vec), [limit_vec] "=r" (limit_vec)
+        : [thresh] "r" (thresh), [flimit] "r" (flimit), [limit] "r" (limit)
+    );
+
+    vp8_mbloop_filter_vertical_edge_mips(y_ptr, y_stride, flimit_vec, limit_vec, thresh_vec, 16);
+
+    if (u_ptr)
+        vp8_mbloop_filter_uvvertical_edge_mips(u_ptr, uv_stride, flimit_vec, limit_vec, thresh_vec, 0);
+
+    if (v_ptr)
+        vp8_mbloop_filter_uvvertical_edge_mips(v_ptr, uv_stride, flimit_vec, limit_vec, thresh_vec, 0);
+}
+
+
+/* Horizontal B Filtering */
+void vp8_loop_filter_bh_dspr2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                              int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    unsigned int thresh_vec, flimit_vec, limit_vec;
+    unsigned char thresh, flimit, limit, flimit_temp;
+
+    /* use direct value instead pointers */
+    limit = *(lfi->lim);
+    flimit_temp = *(lfi->blim);
+    thresh = *(lfi->hev_thr);
+    flimit = flimit_temp;
+
+    /* create quad-byte */
+    __asm__ __volatile__ (
+        "replv.qb       %[thresh_vec], %[thresh]    \n\t"
+        "replv.qb       %[flimit_vec], %[flimit]    \n\t"
+        "replv.qb       %[limit_vec],  %[limit]     \n\t"
+        : [thresh_vec] "=&r" (thresh_vec), [flimit_vec] "=&r" (flimit_vec), [limit_vec] "=r" (limit_vec)
+        : [thresh] "r" (thresh), [flimit] "r" (flimit), [limit] "r" (limit)
+    );
+
+    vp8_loop_filter_horizontal_edge_mips(y_ptr + 4 * y_stride, y_stride, flimit_vec, limit_vec, thresh_vec, 16);
+    vp8_loop_filter_horizontal_edge_mips(y_ptr + 8 * y_stride, y_stride, flimit_vec, limit_vec, thresh_vec, 16);
+    vp8_loop_filter_horizontal_edge_mips(y_ptr + 12 * y_stride, y_stride, flimit_vec, limit_vec, thresh_vec, 16);
+
+    if (u_ptr)
+        vp8_loop_filter_uvhorizontal_edge_mips(u_ptr + 4 * uv_stride, uv_stride, flimit_vec, limit_vec, thresh_vec, 0);
+
+    if (v_ptr)
+        vp8_loop_filter_uvhorizontal_edge_mips(v_ptr + 4 * uv_stride, uv_stride, flimit_vec, limit_vec, thresh_vec, 0);
+}
+
+
+/* Vertical B Filtering */
+void vp8_loop_filter_bv_dspr2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                              int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    unsigned int thresh_vec, flimit_vec, limit_vec;
+    unsigned char thresh, flimit, limit, flimit_temp;
+
+    /* use direct value instead pointers */
+    limit = *(lfi->lim);
+    flimit_temp = *(lfi->blim);
+    thresh = *(lfi->hev_thr);
+    flimit = flimit_temp;
+
+    /* create quad-byte */
+    __asm__ __volatile__ (
+        "replv.qb       %[thresh_vec], %[thresh]    \n\t"
+        "replv.qb       %[flimit_vec], %[flimit]    \n\t"
+        "replv.qb       %[limit_vec],  %[limit]     \n\t"
+        : [thresh_vec] "=&r" (thresh_vec), [flimit_vec] "=&r" (flimit_vec), [limit_vec] "=r" (limit_vec)
+        : [thresh] "r" (thresh), [flimit] "r" (flimit), [limit] "r" (limit)
+    );
+
+    vp8_loop_filter_vertical_edge_mips(y_ptr + 4, y_stride, flimit_vec, limit_vec, thresh_vec, 16);
+    vp8_loop_filter_vertical_edge_mips(y_ptr + 8, y_stride, flimit_vec, limit_vec, thresh_vec, 16);
+    vp8_loop_filter_vertical_edge_mips(y_ptr + 12, y_stride, flimit_vec, limit_vec, thresh_vec, 16);
+
+    if (u_ptr)
+        vp8_loop_filter_uvvertical_edge_mips(u_ptr + 4, uv_stride, flimit_vec, limit_vec, thresh_vec, 0);
+
+    if (v_ptr)
+        vp8_loop_filter_uvvertical_edge_mips(v_ptr + 4, uv_stride, flimit_vec, limit_vec, thresh_vec, 0);
+}
+
+#endif
diff --git a/libvpx/libvpx/vp8/common/mips/msa/bilinear_filter_msa.c b/libvpx/libvpx/vp8/common/mips/msa/bilinear_filter_msa.c
new file mode 100644
index 0000000..1054ed3
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/msa/bilinear_filter_msa.c
@@ -0,0 +1,911 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vp8/common/filter.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+
+DECLARE_ALIGNED(16, static const int8_t, vp8_bilinear_filters_msa[7][2]) =
+{
+    { 112, 16 },
+    { 96, 32 },
+    { 80, 48 },
+    { 64, 64 },
+    { 48, 80 },
+    { 32, 96 },
+    { 16, 112 }
+};
+
+static const uint8_t vp8_mc_filt_mask_arr[16 * 3] =
+{
+    /* 8 width cases */
+    0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+    /* 4 width cases */
+    0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20,
+    /* 4 width cases */
+    8, 9, 9, 10, 10, 11, 11, 12, 24, 25, 25, 26, 26, 27, 27, 28
+};
+
+static void common_hz_2t_4x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter)
+{
+    v16i8 src0, src1, src2, src3, mask;
+    v16u8 filt0, vec0, vec1, res0, res1;
+    v8u16 vec2, vec3, filt;
+
+    mask = LD_SB(&vp8_mc_filt_mask_arr[16]);
+
+    filt = LD_UH(filter);
+    filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, vec2, vec3);
+    SRARI_H2_UH(vec2, vec3, VP8_FILTER_SHIFT);
+    PCKEV_B2_UB(vec2, vec2, vec3, vec3, res0, res1);
+    ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_2t_4x8_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter)
+{
+    v16u8 vec0, vec1, vec2, vec3, filt0;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+    v16i8 res0, res1, res2, res3;
+    v8u16 vec4, vec5, vec6, vec7, filt;
+
+    mask = LD_SB(&vp8_mc_filt_mask_arr[16]);
+
+    filt = LD_UH(filter);
+    filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+    LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src4, src5, src6, src7, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                vec4, vec5, vec6, vec7);
+    SRARI_H4_UH(vec4, vec5, vec6, vec7, VP8_FILTER_SHIFT);
+    PCKEV_B4_SB(vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7,
+                res0, res1, res2, res3);
+    ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+    dst += (4 * dst_stride);
+    ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_2t_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    if (4 == height)
+    {
+        common_hz_2t_4x4_msa(src, src_stride, dst, dst_stride, filter);
+    }
+    else if (8 == height)
+    {
+        common_hz_2t_4x8_msa(src, src_stride, dst, dst_stride, filter);
+    }
+}
+
+static void common_hz_2t_8x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter)
+{
+    v16u8 filt0;
+    v16i8 src0, src1, src2, src3, mask;
+    v8u16 vec0, vec1, vec2, vec3, filt;
+
+    mask = LD_SB(&vp8_mc_filt_mask_arr[0]);
+
+    filt = LD_UH(filter);
+    filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                vec0, vec1, vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, VP8_FILTER_SHIFT);
+    PCKEV_B2_SB(vec1, vec0, vec3, vec2, src0, src1);
+    ST8x4_UB(src0, src1, dst, dst_stride);
+}
+
+static void common_hz_2t_8x8mult_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter, int32_t height)
+{
+    v16u8 filt0;
+    v16i8 src0, src1, src2, src3, mask, out0, out1;
+    v8u16 vec0, vec1, vec2, vec3, filt;
+
+    mask = LD_SB(&vp8_mc_filt_mask_arr[0]);
+
+    filt = LD_UH(filter);
+    filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                vec0, vec1, vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, VP8_FILTER_SHIFT);
+
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+
+    PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                vec0, vec1, vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, VP8_FILTER_SHIFT);
+    PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    if (16 == height)
+    {
+        LD_SB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+
+        VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+        VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+        DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                    vec0, vec1, vec2, vec3);
+        SRARI_H4_UH(vec0, vec1, vec2, vec3, VP8_FILTER_SHIFT);
+        LD_SB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+
+        PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+        ST8x4_UB(out0, out1, dst, dst_stride);
+
+        VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+        VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+        DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                    vec0, vec1, vec2, vec3);
+        SRARI_H4_UH(vec0, vec1, vec2, vec3, VP8_FILTER_SHIFT);
+        PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+        ST8x4_UB(out0, out1, dst + 4 * dst_stride, dst_stride);
+    }
+}
+
+static void common_hz_2t_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    if (4 == height)
+    {
+        common_hz_2t_8x4_msa(src, src_stride, dst, dst_stride, filter);
+    }
+    else
+    {
+        common_hz_2t_8x8mult_msa(src, src_stride, dst, dst_stride, filter, height);
+    }
+}
+
+static void common_hz_2t_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+    v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+    v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt;
+
+    mask = LD_SB(&vp8_mc_filt_mask_arr[0]);
+
+    loop_cnt = (height >> 2) - 1;
+
+    filt = LD_UH(filter);
+    filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                out0, out1, out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0,
+                out4, out5, out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, VP8_FILTER_SHIFT);
+    SRARI_H4_UH(out4, out5, out6, out7, VP8_FILTER_SHIFT);
+    PCKEV_ST_SB(out0, out1, dst);
+    dst += dst_stride;
+    PCKEV_ST_SB(out2, out3, dst);
+    dst += dst_stride;
+    PCKEV_ST_SB(out4, out5, dst);
+    dst += dst_stride;
+    PCKEV_ST_SB(out6, out7, dst);
+    dst += dst_stride;
+
+    for (; loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src0, src2, src4, src6);
+        LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+        src += (4 * src_stride);
+
+        VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+        VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+        VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+        VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+        DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                    out0, out1, out2, out3);
+        DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0,
+                    out4, out5, out6, out7);
+        SRARI_H4_UH(out0, out1, out2, out3, VP8_FILTER_SHIFT);
+        SRARI_H4_UH(out4, out5, out6, out7, VP8_FILTER_SHIFT);
+        PCKEV_ST_SB(out0, out1, dst);
+        dst += dst_stride;
+        PCKEV_ST_SB(out2, out3, dst);
+        dst += dst_stride;
+        PCKEV_ST_SB(out4, out5, dst);
+        dst += dst_stride;
+        PCKEV_ST_SB(out6, out7, dst);
+        dst += dst_stride;
+    }
+}
+
+static void common_vt_2t_4x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter)
+{
+    v16i8 src0, src1, src2, src3, src4;
+    v16i8 src10_r, src32_r, src21_r, src43_r, src2110, src4332;
+    v16u8 filt0;
+    v8i16 filt;
+    v8u16 tmp0, tmp1;
+
+    filt = LD_SH(filter);
+    filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+    LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+    src += (5 * src_stride);
+
+    ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3,
+               src10_r, src21_r, src32_r, src43_r);
+    ILVR_D2_SB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+    DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, VP8_FILTER_SHIFT);
+    src2110 = __msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    ST4x4_UB(src2110, src2110, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_vt_2t_4x8_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter)
+{
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+    v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r;
+    v16i8 src65_r, src87_r, src2110, src4332, src6554, src8776;
+    v8u16 tmp0, tmp1, tmp2, tmp3;
+    v16u8 filt0;
+    v8i16 filt;
+
+    filt = LD_SH(filter);
+    filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+    LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    src += (8 * src_stride);
+
+    src8 = LD_SB(src);
+    src += src_stride;
+
+    ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+               src32_r, src43_r);
+    ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r, src65_r,
+               src76_r, src87_r);
+    ILVR_D4_SB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r,
+               src87_r, src76_r, src2110, src4332, src6554, src8776);
+    DOTP_UB4_UH(src2110, src4332, src6554, src8776, filt0, filt0, filt0, filt0,
+                tmp0, tmp1, tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, VP8_FILTER_SHIFT);
+    PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, src2110, src4332);
+    ST4x4_UB(src2110, src2110, 0, 1, 2, 3, dst, dst_stride);
+    ST4x4_UB(src4332, src4332, 0, 1, 2, 3, dst + 4 * dst_stride, dst_stride);
+}
+
+static void common_vt_2t_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    if (4 == height)
+    {
+        common_vt_2t_4x4_msa(src, src_stride, dst, dst_stride, filter);
+    }
+    else if (8 == height)
+    {
+        common_vt_2t_4x8_msa(src, src_stride, dst, dst_stride, filter);
+    }
+}
+
+static void common_vt_2t_8x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter)
+{
+    v16u8 src0, src1, src2, src3, src4, vec0, vec1, vec2, vec3, filt0;
+    v16i8 out0, out1;
+    v8u16 tmp0, tmp1, tmp2, tmp3;
+    v8i16 filt;
+
+    filt = LD_SH(filter);
+    filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+    LD_UB5(src, src_stride, src0, src1, src2, src3, src4);
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec1);
+    ILVR_B2_UB(src3, src2, src4, src3, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                tmp0, tmp1, tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, VP8_FILTER_SHIFT);
+    PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+}
+
+static void common_vt_2t_8x8mult_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+    v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+    v16i8 out0, out1;
+    v8u16 tmp0, tmp1, tmp2, tmp3;
+    v8i16 filt;
+
+    filt = LD_SH(filter);
+    filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+    src0 = LD_UB(src);
+    src += src_stride;
+
+    for (loop_cnt = (height >> 3); loop_cnt--;)
+    {
+        LD_UB8(src, src_stride, src1, src2, src3, src4, src5, src6, src7, src8);
+        src += (8 * src_stride);
+
+        ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3,
+                   vec0, vec1, vec2, vec3);
+        ILVR_B4_UB(src5, src4, src6, src5, src7, src6, src8, src7,
+                   vec4, vec5, vec6, vec7);
+        DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                    tmp0, tmp1, tmp2, tmp3);
+        SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, VP8_FILTER_SHIFT);
+        PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+        ST8x4_UB(out0, out1, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0,
+                    tmp0, tmp1, tmp2, tmp3);
+        SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, VP8_FILTER_SHIFT);
+        PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+        ST8x4_UB(out0, out1, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        src0 = src8;
+    }
+}
+
+static void common_vt_2t_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    if (4 == height)
+    {
+        common_vt_2t_8x4_msa(src, src_stride, dst, dst_stride, filter);
+    }
+    else
+    {
+        common_vt_2t_8x8mult_msa(src, src_stride, dst, dst_stride, filter,
+                                 height);
+    }
+}
+
+static void common_vt_2t_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16u8 src0, src1, src2, src3, src4;
+    v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+    v8u16 tmp0, tmp1, tmp2, tmp3;
+    v8i16 filt;
+
+    filt = LD_SH(filter);
+    filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+    src0 = LD_UB(src);
+    src += src_stride;
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_UB4(src, src_stride, src1, src2, src3, src4);
+        src += (4 * src_stride);
+
+        ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+        ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+        DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+        SRARI_H2_UH(tmp0, tmp1, VP8_FILTER_SHIFT);
+        PCKEV_ST_SB(tmp0, tmp1, dst);
+        dst += dst_stride;
+
+        ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+        ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+        DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+        SRARI_H2_UH(tmp2, tmp3, VP8_FILTER_SHIFT);
+        PCKEV_ST_SB(tmp2, tmp3, dst);
+        dst += dst_stride;
+
+        DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+        SRARI_H2_UH(tmp0, tmp1, VP8_FILTER_SHIFT);
+        PCKEV_ST_SB(tmp0, tmp1, dst);
+        dst += dst_stride;
+
+        DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+        SRARI_H2_UH(tmp2, tmp3, VP8_FILTER_SHIFT);
+        PCKEV_ST_SB(tmp2, tmp3, dst);
+        dst += dst_stride;
+
+        src0 = src4;
+    }
+}
+
+static void common_hv_2ht_2vt_4x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                      uint8_t *RESTRICT dst, int32_t dst_stride,
+                                      const int8_t *filter_horiz,
+                                      const int8_t *filter_vert)
+{
+    v16i8 src0, src1, src2, src3, src4, mask;
+    v16u8 filt_vt, filt_hz, vec0, vec1, res0, res1;
+    v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, filt, tmp0, tmp1;
+
+    mask = LD_SB(&vp8_mc_filt_mask_arr[16]);
+
+    filt = LD_UH(filter_horiz);
+    filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0);
+    filt = LD_UH(filter_vert);
+    filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+    LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, VP8_FILTER_SHIFT);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, VP8_FILTER_SHIFT);
+    hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, VP8_FILTER_SHIFT);
+    hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+    hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2);
+
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, VP8_FILTER_SHIFT);
+    PCKEV_B2_UB(tmp0, tmp0, tmp1, tmp1, res0, res1);
+    ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_4x8_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                      uint8_t *RESTRICT dst, int32_t dst_stride,
+                                      const int8_t *filter_horiz,
+                                      const int8_t *filter_vert)
+{
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, mask;
+    v16i8 res0, res1, res2, res3;
+    v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3;
+    v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+    v8u16 hz_out7, hz_out8, vec4, vec5, vec6, vec7, filt;
+
+    mask = LD_SB(&vp8_mc_filt_mask_arr[16]);
+
+    filt = LD_UH(filter_horiz);
+    filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0);
+    filt = LD_UH(filter_vert);
+    filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+    LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    src += (8 * src_stride);
+    src8 = LD_SB(src);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, VP8_FILTER_SHIFT);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, VP8_FILTER_SHIFT);
+    hz_out4 = HORIZ_2TAP_FILT_UH(src4, src5, mask, filt_hz, VP8_FILTER_SHIFT);
+    hz_out6 = HORIZ_2TAP_FILT_UH(src6, src7, mask, filt_hz, VP8_FILTER_SHIFT);
+    hz_out8 = HORIZ_2TAP_FILT_UH(src8, src8, mask, filt_hz, VP8_FILTER_SHIFT);
+    SLDI_B3_UH(hz_out2, hz_out4, hz_out6, hz_out0, hz_out2, hz_out4, hz_out1,
+               hz_out3, hz_out5, 8);
+    hz_out7 = (v8u16)__msa_pckod_d((v2i64)hz_out8, (v2i64)hz_out6);
+
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    ILVEV_B2_UB(hz_out4, hz_out5, hz_out6, hz_out7, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt_vt, filt_vt, filt_vt, filt_vt,
+                vec4, vec5, vec6, vec7);
+    SRARI_H4_UH(vec4, vec5, vec6, vec7, VP8_FILTER_SHIFT);
+    PCKEV_B4_SB(vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7,
+                res0, res1, res2, res3);
+    ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+    dst += (4 * dst_stride);
+    ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height)
+{
+    if (4 == height)
+    {
+        common_hv_2ht_2vt_4x4_msa(src, src_stride, dst, dst_stride,
+                                  filter_horiz, filter_vert);
+    }
+    else if (8 == height)
+    {
+        common_hv_2ht_2vt_4x8_msa(src, src_stride, dst, dst_stride,
+                                  filter_horiz, filter_vert);
+    }
+}
+
+static void common_hv_2ht_2vt_8x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                      uint8_t *RESTRICT dst, int32_t dst_stride,
+                                      const int8_t *filter_horiz,
+                                      const int8_t *filter_vert)
+{
+    v16i8 src0, src1, src2, src3, src4, mask, out0, out1;
+    v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3;
+    v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3;
+    v8i16 filt;
+
+    mask = LD_SB(&vp8_mc_filt_mask_arr[0]);
+
+    filt = LD_SH(filter_horiz);
+    filt_hz = (v16u8)__msa_splati_h(filt, 0);
+    filt = LD_SH(filter_vert);
+    filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+    LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, VP8_FILTER_SHIFT);
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, VP8_FILTER_SHIFT);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, VP8_FILTER_SHIFT);
+    vec1 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp1 = __msa_dotp_u_h(vec1, filt_vt);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, VP8_FILTER_SHIFT);
+    vec2 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp2 = __msa_dotp_u_h(vec2, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, VP8_FILTER_SHIFT);
+    vec3 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp3 = __msa_dotp_u_h(vec3, filt_vt);
+
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, VP8_FILTER_SHIFT);
+    PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_8x8mult_msa(uint8_t *RESTRICT src,
+                                          int32_t src_stride,
+                                          uint8_t *RESTRICT dst,
+                                          int32_t dst_stride,
+                                          const int8_t *filter_horiz,
+                                          const int8_t *filter_vert,
+                                          int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, mask, out0, out1;
+    v16u8 filt_hz, filt_vt, vec0;
+    v8u16 hz_out0, hz_out1, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
+    v8i16 filt;
+
+    mask = LD_SB(&vp8_mc_filt_mask_arr[0]);
+
+    filt = LD_SH(filter_horiz);
+    filt_hz = (v16u8)__msa_splati_h(filt, 0);
+    filt = LD_SH(filter_vert);
+    filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+    src0 = LD_SB(src);
+    src += src_stride;
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, VP8_FILTER_SHIFT);
+
+    for (loop_cnt = (height >> 3); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src1, src2, src3, src4);
+        src += (4 * src_stride);
+
+        hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+        tmp1 = __msa_dotp_u_h(vec0, filt_vt);
+
+        hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+        tmp2 = __msa_dotp_u_h(vec0, filt_vt);
+
+        SRARI_H2_UH(tmp1, tmp2, VP8_FILTER_SHIFT);
+
+        hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+        tmp3 = __msa_dotp_u_h(vec0, filt_vt);
+
+        hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        LD_SB4(src, src_stride, src1, src2, src3, src4);
+        src += (4 * src_stride);
+        vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+        tmp4 = __msa_dotp_u_h(vec0, filt_vt);
+
+        SRARI_H2_UH(tmp3, tmp4, VP8_FILTER_SHIFT);
+        PCKEV_B2_SB(tmp2, tmp1, tmp4, tmp3, out0, out1);
+        ST8x4_UB(out0, out1, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+        tmp5 = __msa_dotp_u_h(vec0, filt_vt);
+
+        hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+        tmp6 = __msa_dotp_u_h(vec0, filt_vt);
+
+        hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+        tmp7 = __msa_dotp_u_h(vec0, filt_vt);
+
+        hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+        tmp8 = __msa_dotp_u_h(vec0, filt_vt);
+
+        SRARI_H4_UH(tmp5, tmp6, tmp7, tmp8, VP8_FILTER_SHIFT);
+        PCKEV_B2_SB(tmp6, tmp5, tmp8, tmp7, out0, out1);
+        ST8x4_UB(out0, out1, dst, dst_stride);
+        dst += (4 * dst_stride);
+    }
+}
+
+static void common_hv_2ht_2vt_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height)
+{
+    if (4 == height)
+    {
+        common_hv_2ht_2vt_8x4_msa(src, src_stride, dst, dst_stride,
+                                  filter_horiz, filter_vert);
+    }
+    else
+    {
+        common_hv_2ht_2vt_8x8mult_msa(src, src_stride, dst, dst_stride,
+                                      filter_horiz, filter_vert, height);
+    }
+}
+
+static void common_hv_2ht_2vt_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                      uint8_t *RESTRICT dst, int32_t dst_stride,
+                                      const int8_t *filter_horiz,
+                                      const int8_t *filter_vert,
+                                      int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+    v16u8 filt_hz, filt_vt, vec0, vec1;
+    v8u16 tmp1, tmp2, hz_out0, hz_out1, hz_out2, hz_out3;
+    v8i16 filt;
+
+    mask = LD_SB(&vp8_mc_filt_mask_arr[0]);
+
+    /* rearranging filter */
+    filt = LD_SH(filter_horiz);
+    filt_hz = (v16u8)__msa_splati_h(filt, 0);
+    filt = LD_SH(filter_vert);
+    filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+    LD_SB2(src, 8, src0, src1);
+    src += src_stride;
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, VP8_FILTER_SHIFT);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, VP8_FILTER_SHIFT);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src0, src2, src4, src6);
+        LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+        src += (4 * src_stride);
+
+        hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+        DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+        SRARI_H2_UH(tmp1, tmp2, VP8_FILTER_SHIFT);
+        PCKEV_ST_SB(tmp1, tmp2, dst);
+        dst += dst_stride;
+
+        hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+        DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+        SRARI_H2_UH(tmp1, tmp2, VP8_FILTER_SHIFT);
+        PCKEV_ST_SB(tmp1, tmp2, dst);
+        dst += dst_stride;
+
+        hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+        DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+        SRARI_H2_UH(tmp1, tmp2, VP8_FILTER_SHIFT);
+        PCKEV_ST_SB(tmp1, tmp2, dst);
+        dst += dst_stride;
+
+        hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz,
+                                     VP8_FILTER_SHIFT);
+        ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+        DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+        SRARI_H2_UH(tmp1, tmp2, VP8_FILTER_SHIFT);
+        PCKEV_ST_SB(tmp1, tmp2, dst);
+        dst += dst_stride;
+    }
+}
+
+void vp8_bilinear_predict4x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 int32_t xoffset, int32_t yoffset,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride)
+{
+    const int8_t *h_filter = vp8_bilinear_filters_msa[xoffset - 1];
+    const int8_t *v_filter = vp8_bilinear_filters_msa[yoffset - 1];
+
+    if (yoffset)
+    {
+        if (xoffset)
+        {
+            common_hv_2ht_2vt_4w_msa(src, src_stride, dst, dst_stride,
+                                     h_filter, v_filter, 4);
+        }
+        else
+        {
+            common_vt_2t_4w_msa(src, src_stride, dst, dst_stride, v_filter, 4);
+        }
+    }
+    else
+    {
+        if (xoffset)
+        {
+            common_hz_2t_4w_msa(src, src_stride, dst, dst_stride, h_filter, 4);
+        }
+        else
+        {
+            uint32_t tp0, tp1, tp2, tp3;
+
+            LW4(src, src_stride, tp0, tp1, tp2, tp3);
+            SW4(tp0, tp1, tp2, tp3, dst, dst_stride);
+        }
+    }
+}
+
+void vp8_bilinear_predict8x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 int32_t xoffset, int32_t yoffset,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride)
+{
+    const int8_t *h_filter = vp8_bilinear_filters_msa[xoffset - 1];
+    const int8_t *v_filter = vp8_bilinear_filters_msa[yoffset - 1];
+
+    if (yoffset)
+    {
+        if (xoffset)
+        {
+            common_hv_2ht_2vt_8w_msa(src, src_stride, dst, dst_stride,
+                                     h_filter, v_filter, 4);
+        }
+        else
+        {
+            common_vt_2t_8w_msa(src, src_stride, dst, dst_stride, v_filter, 4);
+        }
+    }
+    else
+    {
+        if (xoffset)
+        {
+            common_hz_2t_8w_msa(src, src_stride, dst, dst_stride, h_filter, 4);
+        }
+        else
+        {
+            vp8_copy_mem8x4(src, src_stride, dst, dst_stride);
+        }
+    }
+}
+
+void vp8_bilinear_predict8x8_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 int32_t xoffset, int32_t yoffset,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride)
+{
+    const int8_t *h_filter = vp8_bilinear_filters_msa[xoffset - 1];
+    const int8_t *v_filter = vp8_bilinear_filters_msa[yoffset - 1];
+
+    if (yoffset)
+    {
+        if (xoffset)
+        {
+            common_hv_2ht_2vt_8w_msa(src, src_stride, dst, dst_stride,
+                                     h_filter, v_filter, 8);
+        }
+        else
+        {
+            common_vt_2t_8w_msa(src, src_stride, dst, dst_stride, v_filter, 8);
+        }
+    }
+    else
+    {
+        if (xoffset)
+        {
+            common_hz_2t_8w_msa(src, src_stride, dst, dst_stride, h_filter, 8);
+        }
+        else
+        {
+            vp8_copy_mem8x8(src, src_stride, dst, dst_stride);
+        }
+    }
+}
+
+void vp8_bilinear_predict16x16_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                   int32_t xoffset, int32_t yoffset,
+                                   uint8_t *RESTRICT dst, int32_t dst_stride)
+{
+    const int8_t *h_filter = vp8_bilinear_filters_msa[xoffset - 1];
+    const int8_t *v_filter = vp8_bilinear_filters_msa[yoffset - 1];
+
+    if (yoffset)
+    {
+        if (xoffset)
+        {
+            common_hv_2ht_2vt_16w_msa(src, src_stride, dst, dst_stride,
+                                      h_filter, v_filter, 16);
+        }
+        else
+        {
+            common_vt_2t_16w_msa(src, src_stride, dst, dst_stride, v_filter,
+                                 16);
+        }
+    }
+    else
+    {
+        if (xoffset)
+        {
+            common_hz_2t_16w_msa(src, src_stride, dst, dst_stride, h_filter,
+                                 16);
+        }
+        else
+        {
+            vp8_copy_mem16x16(src, src_stride, dst, dst_stride);
+        }
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/mips/msa/copymem_msa.c b/libvpx/libvpx/vp8/common/mips/msa/copymem_msa.c
new file mode 100644
index 0000000..002a5ed
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/msa/copymem_msa.c
@@ -0,0 +1,70 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+
+static void copy_8x4_msa(uint8_t *src, int32_t src_stride,
+                         uint8_t *dst, int32_t dst_stride)
+{
+    uint64_t src0, src1, src2, src3;
+
+    LD4(src, src_stride, src0, src1, src2, src3);
+    SD4(src0, src1, src2, src3, dst, dst_stride);
+}
+
+static void copy_8x8_msa(uint8_t *src, int32_t src_stride,
+                         uint8_t *dst, int32_t dst_stride)
+{
+    uint64_t src0, src1, src2, src3;
+
+    LD4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    SD4(src0, src1, src2, src3, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    LD4(src, src_stride, src0, src1, src2, src3);
+    SD4(src0, src1, src2, src3, dst, dst_stride);
+}
+
+static void copy_16x16_msa(uint8_t *src, int32_t src_stride,
+                           uint8_t *dst, int32_t dst_stride)
+{
+    v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+    v16u8 src8, src9, src10, src11, src12, src13, src14, src15;
+
+    LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    src += (8 * src_stride);
+    LD_UB8(src, src_stride, src8, src9, src10, src11, src12, src13, src14,
+           src15);
+
+    ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
+    dst += (8 * dst_stride);
+    ST_UB8(src8, src9, src10, src11, src12, src13, src14, src15, dst,
+           dst_stride);
+}
+
+void vp8_copy_mem16x16_msa(uint8_t *src, int32_t src_stride,
+                           uint8_t *dst, int32_t dst_stride)
+{
+    copy_16x16_msa(src, src_stride, dst, dst_stride);
+}
+
+void vp8_copy_mem8x8_msa(uint8_t *src, int32_t src_stride,
+                         uint8_t *dst, int32_t dst_stride)
+{
+    copy_8x8_msa(src, src_stride, dst, dst_stride);
+}
+
+void vp8_copy_mem8x4_msa(uint8_t *src, int32_t src_stride,
+                         uint8_t *dst, int32_t dst_stride)
+{
+    copy_8x4_msa(src, src_stride, dst, dst_stride);
+}
diff --git a/libvpx/libvpx/vp8/common/mips/msa/idct_msa.c b/libvpx/libvpx/vp8/common/mips/msa/idct_msa.c
new file mode 100644
index 0000000..e537a3f
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/msa/idct_msa.c
@@ -0,0 +1,457 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+#include "vp8/common/blockd.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+
+static const int32_t cospi8sqrt2minus1 = 20091;
+static const int32_t sinpi8sqrt2 = 35468;
+
+#define TRANSPOSE_TWO_4x4_H(in0, in1, in2, in3, out0, out1, out2, out3)  \
+{                                                                        \
+    v8i16 s4_m, s5_m, s6_m, s7_m;                                        \
+                                                                         \
+    TRANSPOSE8X4_SH_SH(in0, in1, in2, in3, s4_m, s5_m, s6_m, s7_m);      \
+    ILVR_D2_SH(s6_m, s4_m, s7_m, s5_m, out0, out2);                      \
+    out1 = (v8i16)__msa_ilvl_d((v2i64)s6_m, (v2i64)s4_m);                \
+    out3 = (v8i16)__msa_ilvl_d((v2i64)s7_m, (v2i64)s5_m);                \
+}
+
+#define EXPAND_TO_H_MULTIPLY_SINPI8SQRT2_PCK_TO_W(in)     \
+({                                                        \
+    v8i16 out_m;                                          \
+    v8i16 zero_m = { 0 };                                 \
+    v4i32 tmp1_m, tmp2_m;                                 \
+    v4i32 sinpi8_sqrt2_m = __msa_fill_w(sinpi8sqrt2);     \
+                                                          \
+    ILVRL_H2_SW(in, zero_m, tmp1_m, tmp2_m);              \
+    tmp1_m >>= 16;                                        \
+    tmp2_m >>= 16;                                        \
+    tmp1_m = (tmp1_m * sinpi8_sqrt2_m) >> 16;             \
+    tmp2_m = (tmp2_m * sinpi8_sqrt2_m) >> 16;             \
+    out_m = __msa_pckev_h((v8i16)tmp2_m, (v8i16)tmp1_m);  \
+                                                          \
+    out_m;                                                \
+})
+
+#define VP8_IDCT_1D_H(in0, in1, in2, in3, out0, out1, out2, out3)  \
+{                                                                  \
+    v8i16 a1_m, b1_m, c1_m, d1_m;                                  \
+    v8i16 c_tmp1_m, c_tmp2_m, d_tmp1_m, d_tmp2_m;                  \
+    v8i16 const_cospi8sqrt2minus1_m;                               \
+                                                                   \
+    const_cospi8sqrt2minus1_m = __msa_fill_h(cospi8sqrt2minus1);   \
+    a1_m = in0 + in2;                                              \
+    b1_m = in0 - in2;                                              \
+    c_tmp1_m = EXPAND_TO_H_MULTIPLY_SINPI8SQRT2_PCK_TO_W(in1);     \
+    c_tmp2_m = __msa_mul_q_h(in3, const_cospi8sqrt2minus1_m);      \
+    c_tmp2_m = c_tmp2_m >> 1;                                      \
+    c_tmp2_m = in3 + c_tmp2_m;                                     \
+    c1_m = c_tmp1_m - c_tmp2_m;                                    \
+    d_tmp1_m = __msa_mul_q_h(in1, const_cospi8sqrt2minus1_m);      \
+    d_tmp1_m = d_tmp1_m >> 1;                                      \
+    d_tmp1_m = in1 + d_tmp1_m;                                     \
+    d_tmp2_m = EXPAND_TO_H_MULTIPLY_SINPI8SQRT2_PCK_TO_W(in3);     \
+    d1_m = d_tmp1_m + d_tmp2_m;                                    \
+    BUTTERFLY_4(a1_m, b1_m, c1_m, d1_m, out0, out1, out2, out3);   \
+}
+
+#define VP8_IDCT_1D_W(in0, in1, in2, in3, out0, out1, out2, out3)  \
+{                                                                  \
+    v4i32 a1_m, b1_m, c1_m, d1_m;                                  \
+    v4i32 c_tmp1_m, c_tmp2_m, d_tmp1_m, d_tmp2_m;                  \
+    v4i32 const_cospi8sqrt2minus1_m, sinpi8_sqrt2_m;               \
+                                                                   \
+    const_cospi8sqrt2minus1_m = __msa_fill_w(cospi8sqrt2minus1);   \
+    sinpi8_sqrt2_m = __msa_fill_w(sinpi8sqrt2);                    \
+    a1_m = in0 + in2;                                              \
+    b1_m = in0 - in2;                                              \
+    c_tmp1_m = (in1 * sinpi8_sqrt2_m) >> 16;                       \
+    c_tmp2_m = in3 + ((in3 * const_cospi8sqrt2minus1_m) >> 16);    \
+    c1_m = c_tmp1_m - c_tmp2_m;                                    \
+    d_tmp1_m = in1 + ((in1 * const_cospi8sqrt2minus1_m) >> 16);    \
+    d_tmp2_m = (in3 * sinpi8_sqrt2_m) >> 16;                       \
+    d1_m = d_tmp1_m + d_tmp2_m;                                    \
+    BUTTERFLY_4(a1_m, b1_m, c1_m, d1_m, out0, out1, out2, out3);   \
+}
+
+static void idct4x4_addblk_msa(int16_t *input, uint8_t *pred,
+                               int32_t pred_stride,
+                               uint8_t *dest, int32_t dest_stride)
+{
+    v8i16 input0, input1;
+    v4i32 in0, in1, in2, in3, hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3;
+    v4i32 res0, res1, res2, res3;
+    v16i8 zero = { 0 };
+    v16i8 pred0, pred1, pred2, pred3, dest0, dest1, dest2, dest3;
+    v16i8 mask = { 0, 4, 8, 12, 20, 21, 22, 23, 24,
+                   25, 26, 27, 28, 29, 30, 31 };
+
+    LD_SH2(input, 8, input0, input1);
+    UNPCK_SH_SW(input0, in0, in1);
+    UNPCK_SH_SW(input1, in2, in3);
+    VP8_IDCT_1D_W(in0, in1, in2, in3, hz0, hz1, hz2, hz3);
+    TRANSPOSE4x4_SW_SW(hz0, hz1, hz2, hz3, hz0, hz1, hz2, hz3);
+    VP8_IDCT_1D_W(hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3);
+    SRARI_W4_SW(vt0, vt1, vt2, vt3, 3);
+    TRANSPOSE4x4_SW_SW(vt0, vt1, vt2, vt3, vt0, vt1, vt2, vt3);
+    LD_SB4(pred, pred_stride, pred0, pred1, pred2, pred3);
+    ILVR_B4_SW(zero, pred0, zero, pred1, zero, pred2, zero, pred3, res0, res1,
+               res2, res3);
+    ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3, res0, res1,
+               res2, res3);
+    ADD4(res0, vt0, res1, vt1, res2, vt2, res3, vt3, res0, res1, res2, res3);
+    res0 = CLIP_SW_0_255(res0);
+    res1 = CLIP_SW_0_255(res1);
+    res2 = CLIP_SW_0_255(res2);
+    res3 = CLIP_SW_0_255(res3);
+    LD_SB4(dest, dest_stride, dest0, dest1, dest2, dest3);
+    VSHF_B2_SB(res0, dest0, res1, dest1, mask, mask, dest0, dest1);
+    VSHF_B2_SB(res2, dest2, res3, dest3, mask, mask, dest2, dest3);
+    ST_SB4(dest0, dest1, dest2, dest3, dest, dest_stride);
+}
+
+static void idct4x4_addconst_msa(int16_t in_dc, uint8_t *pred,
+                                 int32_t pred_stride,
+                                 uint8_t *dest, int32_t dest_stride)
+{
+    v8i16 vec;
+    v8i16 res0, res1, res2, res3;
+    v16i8 zero = { 0 };
+    v16i8 pred0, pred1, pred2, pred3, dest0, dest1, dest2, dest3;
+    v16i8 mask = { 0, 2, 4, 6, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 };
+
+    vec = __msa_fill_h(in_dc);
+    vec = __msa_srari_h(vec, 3);
+    LD_SB4(pred, pred_stride, pred0, pred1, pred2, pred3);
+    ILVR_B4_SH(zero, pred0, zero, pred1, zero, pred2, zero, pred3, res0, res1,
+               res2, res3);
+    ADD4(res0, vec, res1, vec, res2, vec, res3, vec, res0, res1, res2, res3);
+    CLIP_SH4_0_255(res0, res1, res2, res3);
+    LD_SB4(dest, dest_stride, dest0, dest1, dest2, dest3);
+    VSHF_B2_SB(res0, dest0, res1, dest1, mask, mask, dest0, dest1);
+    VSHF_B2_SB(res2, dest2, res3, dest3, mask, mask, dest2, dest3);
+    ST_SB4(dest0, dest1, dest2, dest3, dest, dest_stride);
+}
+
+void vp8_short_inv_walsh4x4_msa(int16_t *input, int16_t *mb_dq_coeff)
+{
+    v8i16 input0, input1;
+    v4i32 in0, in1, in2, in3, a1, b1, c1, d1;
+    v4i32 hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3;
+
+    LD_SH2(input, 8, input0, input1);
+    UNPCK_SH_SW(input0, in0, in1);
+    UNPCK_SH_SW(input1, in2, in3);
+    BUTTERFLY_4(in0, in1, in2, in3, a1, b1, c1, d1);
+    BUTTERFLY_4(a1, d1, c1, b1, hz0, hz1, hz3, hz2);
+    TRANSPOSE4x4_SW_SW(hz0, hz1, hz2, hz3, hz0, hz1, hz2, hz3);
+    BUTTERFLY_4(hz0, hz1, hz2, hz3, a1, b1, c1, d1);
+    BUTTERFLY_4(a1, d1, c1, b1, vt0, vt1, vt3, vt2);
+    ADD4(vt0, 3, vt1, 3, vt2, 3, vt3, 3, vt0, vt1, vt2, vt3);
+    SRA_4V(vt0, vt1, vt2, vt3, 3);
+    mb_dq_coeff[0] = __msa_copy_s_h((v8i16)vt0, 0);
+    mb_dq_coeff[16] = __msa_copy_s_h((v8i16)vt1, 0);
+    mb_dq_coeff[32] = __msa_copy_s_h((v8i16)vt2, 0);
+    mb_dq_coeff[48] = __msa_copy_s_h((v8i16)vt3, 0);
+    mb_dq_coeff[64] = __msa_copy_s_h((v8i16)vt0, 2);
+    mb_dq_coeff[80] = __msa_copy_s_h((v8i16)vt1, 2);
+    mb_dq_coeff[96] = __msa_copy_s_h((v8i16)vt2, 2);
+    mb_dq_coeff[112] = __msa_copy_s_h((v8i16)vt3, 2);
+    mb_dq_coeff[128] = __msa_copy_s_h((v8i16)vt0, 4);
+    mb_dq_coeff[144] = __msa_copy_s_h((v8i16)vt1, 4);
+    mb_dq_coeff[160] = __msa_copy_s_h((v8i16)vt2, 4);
+    mb_dq_coeff[176] = __msa_copy_s_h((v8i16)vt3, 4);
+    mb_dq_coeff[192] = __msa_copy_s_h((v8i16)vt0, 6);
+    mb_dq_coeff[208] = __msa_copy_s_h((v8i16)vt1, 6);
+    mb_dq_coeff[224] = __msa_copy_s_h((v8i16)vt2, 6);
+    mb_dq_coeff[240] = __msa_copy_s_h((v8i16)vt3, 6);
+}
+
+static void dequant_idct4x4_addblk_msa(int16_t *input, int16_t *dequant_input,
+                                       uint8_t *dest, int32_t dest_stride)
+{
+    v8i16 input0, input1, dequant_in0, dequant_in1, mul0, mul1;
+    v8i16 in0, in1, in2, in3;
+    v8i16 hz0_h, hz1_h, hz2_h, hz3_h;
+    v16i8 dest0, dest1, dest2, dest3;
+    v4i32 hz0_w, hz1_w, hz2_w, hz3_w;
+    v4i32 vt0, vt1, vt2, vt3, res0, res1, res2, res3;
+    v2i64 zero = { 0 };
+    v16i8 mask = { 0, 4, 8, 12, 20, 21, 22, 23, 24,
+                   25, 26, 27, 28, 29, 30, 31 };
+
+    LD_SH2(input, 8, input0, input1);
+    LD_SH2(dequant_input, 8, dequant_in0, dequant_in1);
+    MUL2(input0, dequant_in0, input1, dequant_in1, mul0, mul1);
+    PCKEV_D2_SH(zero, mul0, zero, mul1, in0, in2);
+    PCKOD_D2_SH(zero, mul0, zero, mul1, in1, in3);
+    VP8_IDCT_1D_H(in0, in1, in2, in3, hz0_h, hz1_h, hz2_h, hz3_h);
+    PCKEV_D2_SH(hz1_h, hz0_h, hz3_h, hz2_h, mul0, mul1);
+    UNPCK_SH_SW(mul0, hz0_w, hz1_w);
+    UNPCK_SH_SW(mul1, hz2_w, hz3_w);
+    TRANSPOSE4x4_SW_SW(hz0_w, hz1_w, hz2_w, hz3_w, hz0_w, hz1_w, hz2_w, hz3_w);
+    VP8_IDCT_1D_W(hz0_w, hz1_w, hz2_w, hz3_w, vt0, vt1, vt2, vt3);
+    SRARI_W4_SW(vt0, vt1, vt2, vt3, 3);
+    TRANSPOSE4x4_SW_SW(vt0, vt1, vt2, vt3, vt0, vt1, vt2, vt3);
+    LD_SB4(dest, dest_stride, dest0, dest1, dest2, dest3);
+    ILVR_B4_SW(zero, dest0, zero, dest1, zero, dest2, zero, dest3, res0, res1,
+               res2, res3);
+    ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3, res0, res1,
+               res2, res3);
+    ADD4(res0, vt0, res1, vt1, res2, vt2, res3, vt3, res0, res1, res2, res3);
+    res0 = CLIP_SW_0_255(res0);
+    res1 = CLIP_SW_0_255(res1);
+    res2 = CLIP_SW_0_255(res2);
+    res3 = CLIP_SW_0_255(res3);
+    VSHF_B2_SB(res0, dest0, res1, dest1, mask, mask, dest0, dest1);
+    VSHF_B2_SB(res2, dest2, res3, dest3, mask, mask, dest2, dest3);
+    ST_SB4(dest0, dest1, dest2, dest3, dest, dest_stride);
+}
+
+static void dequant_idct4x4_addblk_2x_msa(int16_t *input,
+                                          int16_t *dequant_input,
+                                          uint8_t *dest, int32_t dest_stride)
+{
+    v16u8 dest0, dest1, dest2, dest3;
+    v8i16 in0, in1, in2, in3;
+    v8i16 mul0, mul1, mul2, mul3, dequant_in0, dequant_in1;
+    v8i16 hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3;
+    v8i16 res0, res1, res2, res3;
+    v4i32 hz0l, hz1l, hz2l, hz3l, hz0r, hz1r, hz2r, hz3r;
+    v4i32 vt0l, vt1l, vt2l, vt3l, vt0r, vt1r, vt2r, vt3r;
+    v16i8 zero = { 0 };
+
+    LD_SH4(input, 8, in0, in1, in2, in3);
+    LD_SH2(dequant_input, 8, dequant_in0, dequant_in1);
+    MUL4(in0, dequant_in0, in1, dequant_in1, in2, dequant_in0, in3, dequant_in1,
+         mul0, mul1, mul2, mul3);
+    PCKEV_D2_SH(mul2, mul0, mul3, mul1, in0, in2);
+    PCKOD_D2_SH(mul2, mul0, mul3, mul1, in1, in3);
+    VP8_IDCT_1D_H(in0, in1, in2, in3, hz0, hz1, hz2, hz3);
+    TRANSPOSE_TWO_4x4_H(hz0, hz1, hz2, hz3, hz0, hz1, hz2, hz3);
+    UNPCK_SH_SW(hz0, hz0r, hz0l);
+    UNPCK_SH_SW(hz1, hz1r, hz1l);
+    UNPCK_SH_SW(hz2, hz2r, hz2l);
+    UNPCK_SH_SW(hz3, hz3r, hz3l);
+    VP8_IDCT_1D_W(hz0l, hz1l, hz2l, hz3l, vt0l, vt1l, vt2l, vt3l);
+    SRARI_W4_SW(vt0l, vt1l, vt2l, vt3l, 3);
+    VP8_IDCT_1D_W(hz0r, hz1r, hz2r, hz3r, vt0r, vt1r, vt2r, vt3r);
+    SRARI_W4_SW(vt0r, vt1r, vt2r, vt3r, 3);
+    PCKEV_H4_SH(vt0l, vt0r, vt1l, vt1r, vt2l, vt2r, vt3l, vt3r, vt0, vt1, vt2,
+                vt3);
+    TRANSPOSE_TWO_4x4_H(vt0, vt1, vt2, vt3, vt0, vt1, vt2, vt3);
+    LD_UB4(dest, dest_stride, dest0, dest1, dest2, dest3);
+    ILVR_B4_SH(zero, dest0, zero, dest1, zero, dest2, zero, dest3, res0, res1,
+               res2, res3);
+    ADD4(res0, vt0, res1, vt1, res2, vt2, res3, vt3, res0, res1, res2, res3);
+    CLIP_SH4_0_255(res0, res1, res2, res3);
+    PCKEV_B4_SH(res0, res0, res1, res1, res2, res2, res3, res3, res0, res1,
+                res2, res3);
+    PCKOD_D2_UB(dest0, res0, dest1, res1, dest0, dest1);
+    PCKOD_D2_UB(dest2, res2, dest3, res3, dest2, dest3);
+    ST_UB4(dest0, dest1, dest2, dest3, dest, dest_stride);
+
+    __asm__ __volatile__(
+        "sw   $zero,    0(%[input])  \n\t"
+        "sw   $zero,    4(%[input])  \n\t"
+        "sw   $zero,    8(%[input])  \n\t"
+        "sw   $zero,   12(%[input])  \n\t"
+        "sw   $zero,   16(%[input])  \n\t"
+        "sw   $zero,   20(%[input])  \n\t"
+        "sw   $zero,   24(%[input])  \n\t"
+        "sw   $zero,   28(%[input])  \n\t"
+        "sw   $zero,   32(%[input])  \n\t"
+        "sw   $zero,   36(%[input])  \n\t"
+        "sw   $zero,   40(%[input])  \n\t"
+        "sw   $zero,   44(%[input])  \n\t"
+        "sw   $zero,   48(%[input])  \n\t"
+        "sw   $zero,   52(%[input])  \n\t"
+        "sw   $zero,   56(%[input])  \n\t"
+        "sw   $zero,   60(%[input])  \n\t"::
+
+        [input] "r"(input)
+    );
+}
+
+static void dequant_idct_addconst_2x_msa(int16_t *input, int16_t *dequant_input,
+                                         uint8_t *dest, int32_t dest_stride)
+{
+    v8i16 input_dc0, input_dc1, vec;
+    v16u8 dest0, dest1, dest2, dest3;
+    v16i8 zero = { 0 };
+    v8i16 res0, res1, res2, res3;
+
+    input_dc0 = __msa_fill_h(input[0] * dequant_input[0]);
+    input_dc1 = __msa_fill_h(input[16] * dequant_input[0]);
+    SRARI_H2_SH(input_dc0, input_dc1, 3);
+    vec = (v8i16)__msa_pckev_d((v2i64)input_dc1, (v2i64)input_dc0);
+    input[0] = 0;
+    input[16] = 0;
+    LD_UB4(dest, dest_stride, dest0, dest1, dest2, dest3);
+    ILVR_B4_SH(zero, dest0, zero, dest1, zero, dest2, zero, dest3, res0,
+               res1, res2, res3);
+    ADD4(res0, vec, res1, vec, res2, vec, res3, vec, res0, res1, res2, res3);
+    CLIP_SH4_0_255(res0, res1, res2, res3);
+    PCKEV_B4_SH(res0, res0, res1, res1, res2, res2, res3, res3, res0, res1,
+                res2, res3);
+    PCKOD_D2_UB(dest0, res0, dest1, res1, dest0, dest1);
+    PCKOD_D2_UB(dest2, res2, dest3, res3, dest2, dest3);
+    ST_UB4(dest0, dest1, dest2, dest3, dest, dest_stride);
+}
+
+void vp8_short_idct4x4llm_msa(int16_t *input, uint8_t *pred_ptr,
+                              int32_t pred_stride, uint8_t *dst_ptr,
+                              int32_t dst_stride)
+{
+    idct4x4_addblk_msa(input, pred_ptr, pred_stride, dst_ptr, dst_stride);
+}
+
+void vp8_dc_only_idct_add_msa(int16_t input_dc, uint8_t *pred_ptr,
+                              int32_t pred_stride, uint8_t *dst_ptr,
+                              int32_t dst_stride)
+{
+    idct4x4_addconst_msa(input_dc, pred_ptr, pred_stride, dst_ptr, dst_stride);
+}
+
+void vp8_dequantize_b_msa(BLOCKD *d, int16_t *DQC)
+{
+    v8i16 dqc0, dqc1, q0, q1, dq0, dq1;
+
+    LD_SH2(DQC, 8, dqc0, dqc1);
+    LD_SH2(d->qcoeff, 8, q0, q1);
+    MUL2(dqc0, q0, dqc1, q1, dq0, dq1);
+    ST_SH2(dq0, dq1, d->dqcoeff, 8);
+}
+
+void vp8_dequant_idct_add_msa(int16_t *input, int16_t *dq,
+                              uint8_t *dest, int32_t stride)
+{
+    dequant_idct4x4_addblk_msa(input, dq, dest, stride);
+
+    __asm__ __volatile__ (
+        "sw     $zero,    0(%[input])     \n\t"
+        "sw     $zero,    4(%[input])     \n\t"
+        "sw     $zero,    8(%[input])     \n\t"
+        "sw     $zero,   12(%[input])     \n\t"
+        "sw     $zero,   16(%[input])     \n\t"
+        "sw     $zero,   20(%[input])     \n\t"
+        "sw     $zero,   24(%[input])     \n\t"
+        "sw     $zero,   28(%[input])     \n\t"
+
+        :
+        : [input] "r" (input)
+    );
+}
+
+void vp8_dequant_idct_add_y_block_msa(int16_t *q, int16_t *dq,
+                                      uint8_t *dst, int32_t stride,
+                                      char *eobs)
+{
+    int16_t *eobs_h = (int16_t *)eobs;
+    uint8_t i;
+
+    for (i = 4; i--;)
+    {
+        if (eobs_h[0])
+        {
+            if (eobs_h[0] & 0xfefe)
+            {
+                dequant_idct4x4_addblk_2x_msa(q, dq, dst, stride);
+            }
+            else
+            {
+                dequant_idct_addconst_2x_msa(q, dq, dst, stride);
+            }
+        }
+
+        q += 32;
+
+        if (eobs_h[1])
+        {
+            if (eobs_h[1] & 0xfefe)
+            {
+                dequant_idct4x4_addblk_2x_msa(q, dq, dst + 8, stride);
+            }
+            else
+            {
+                dequant_idct_addconst_2x_msa(q, dq, dst + 8, stride);
+            }
+        }
+
+        q += 32;
+        dst += (4 * stride);
+        eobs_h += 2;
+    }
+}
+
+void vp8_dequant_idct_add_uv_block_msa(int16_t *q, int16_t *dq,
+                                       uint8_t *dstu, uint8_t *dstv,
+                                       int32_t stride, char *eobs)
+{
+    int16_t *eobs_h = (int16_t *)eobs;
+
+    if (eobs_h[0])
+    {
+        if (eobs_h[0] & 0xfefe)
+        {
+            dequant_idct4x4_addblk_2x_msa(q, dq, dstu, stride);
+        }
+        else
+        {
+            dequant_idct_addconst_2x_msa(q, dq, dstu, stride);
+        }
+    }
+
+    q += 32;
+    dstu += (stride * 4);
+
+    if (eobs_h[1])
+    {
+        if (eobs_h[1] & 0xfefe)
+        {
+            dequant_idct4x4_addblk_2x_msa(q, dq, dstu, stride);
+        }
+        else
+        {
+            dequant_idct_addconst_2x_msa(q, dq, dstu, stride);
+        }
+    }
+
+    q += 32;
+
+    if (eobs_h[2])
+    {
+        if (eobs_h[2] & 0xfefe)
+        {
+            dequant_idct4x4_addblk_2x_msa(q, dq, dstv, stride);
+        }
+        else
+        {
+            dequant_idct_addconst_2x_msa(q, dq, dstv, stride);
+        }
+    }
+
+    q += 32;
+    dstv += (stride * 4);
+
+    if (eobs_h[3])
+    {
+        if (eobs_h[3] & 0xfefe)
+        {
+            dequant_idct4x4_addblk_2x_msa(q, dq, dstv, stride);
+        }
+        else
+        {
+            dequant_idct_addconst_2x_msa(q, dq, dstv, stride);
+        }
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/mips/msa/loopfilter_filters_msa.c b/libvpx/libvpx/vp8/common/mips/msa/loopfilter_filters_msa.c
new file mode 100644
index 0000000..a40f378
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/msa/loopfilter_filters_msa.c
@@ -0,0 +1,826 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+#include "vp8/common/loopfilter.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+
+#define VP8_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask)         \
+{                                                              \
+    v16u8 p1_a_sub_q1, p0_a_sub_q0;                            \
+                                                               \
+    p0_a_sub_q0 = __msa_asub_u_b(p0, q0);                      \
+    p1_a_sub_q1 = __msa_asub_u_b(p1, q1);                      \
+    p1_a_sub_q1 = (v16u8)__msa_srli_b((v16i8)p1_a_sub_q1, 1);  \
+    p0_a_sub_q0 = __msa_adds_u_b(p0_a_sub_q0, p0_a_sub_q0);    \
+    mask = __msa_adds_u_b(p0_a_sub_q0, p1_a_sub_q1);           \
+    mask = ((v16u8)mask <= b_limit);                           \
+}
+
+#define VP8_LPF_FILTER4_4W(p1_in_out, p0_in_out, q0_in_out, q1_in_out,  \
+                           mask_in, hev_in)                             \
+{                                                                       \
+    v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign;                 \
+    v16i8 filt, filt1, filt2, cnst4b, cnst3b;                           \
+    v8i16 q0_sub_p0_r, q0_sub_p0_l, filt_l, filt_r, cnst3h;             \
+                                                                        \
+    p1_m = (v16i8)__msa_xori_b(p1_in_out, 0x80);                        \
+    p0_m = (v16i8)__msa_xori_b(p0_in_out, 0x80);                        \
+    q0_m = (v16i8)__msa_xori_b(q0_in_out, 0x80);                        \
+    q1_m = (v16i8)__msa_xori_b(q1_in_out, 0x80);                        \
+                                                                        \
+    filt = __msa_subs_s_b(p1_m, q1_m);                                  \
+                                                                        \
+    filt = filt & (v16i8)hev_in;                                        \
+                                                                        \
+    q0_sub_p0 = q0_m - p0_m;                                            \
+    filt_sign = __msa_clti_s_b(filt, 0);                                \
+                                                                        \
+    cnst3h = __msa_ldi_h(3);                                            \
+    q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0);            \
+    q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h);    \
+    filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt);                      \
+    filt_r += q0_sub_p0_r;                                              \
+    filt_r = __msa_sat_s_h(filt_r, 7);                                  \
+                                                                        \
+    q0_sub_p0_l = (v8i16)__msa_ilvl_b(q0_sub_p0, q0_sub_p0);            \
+    q0_sub_p0_l = __msa_dotp_s_h((v16i8)q0_sub_p0_l, (v16i8)cnst3h);    \
+    filt_l = (v8i16)__msa_ilvl_b(filt_sign, filt);                      \
+    filt_l += q0_sub_p0_l;                                              \
+    filt_l = __msa_sat_s_h(filt_l, 7);                                  \
+                                                                        \
+    filt = __msa_pckev_b((v16i8)filt_l, (v16i8)filt_r);                 \
+    filt = filt & (v16i8)mask_in;                                       \
+                                                                        \
+    cnst4b = __msa_ldi_b(4);                                            \
+    filt1 = __msa_adds_s_b(filt, cnst4b);                               \
+    filt1 >>= 3;                                                        \
+                                                                        \
+    cnst3b = __msa_ldi_b(3);                                            \
+    filt2 = __msa_adds_s_b(filt, cnst3b);                               \
+    filt2 >>= 3;                                                        \
+                                                                        \
+    q0_m = __msa_subs_s_b(q0_m, filt1);                                 \
+    q0_in_out = __msa_xori_b((v16u8)q0_m, 0x80);                        \
+    p0_m = __msa_adds_s_b(p0_m, filt2);                                 \
+    p0_in_out = __msa_xori_b((v16u8)p0_m, 0x80);                        \
+                                                                        \
+    filt = __msa_srari_b(filt1, 1);                                     \
+    hev_in = __msa_xori_b((v16u8)hev_in, 0xff);                         \
+    filt = filt & (v16i8)hev_in;                                        \
+                                                                        \
+    q1_m = __msa_subs_s_b(q1_m, filt);                                  \
+    q1_in_out = __msa_xori_b((v16u8)q1_m, 0x80);                        \
+    p1_m = __msa_adds_s_b(p1_m, filt);                                  \
+    p1_in_out = __msa_xori_b((v16u8)p1_m, 0x80);                        \
+}
+
+#define VP8_SIMPLE_FILT(p1_in, p0_in, q0_in, q1_in, mask)          \
+{                                                                  \
+    v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, q0_sub_p0_sign;       \
+    v16i8 filt, filt1, filt2, cnst4b, cnst3b, filt_sign;           \
+    v8i16 q0_sub_p0_r, q0_sub_p0_l, filt_l, filt_r, cnst3h;        \
+                                                                   \
+    p1_m = (v16i8)__msa_xori_b(p1_in, 0x80);                       \
+    p0_m = (v16i8)__msa_xori_b(p0_in, 0x80);                       \
+    q0_m = (v16i8)__msa_xori_b(q0_in, 0x80);                       \
+    q1_m = (v16i8)__msa_xori_b(q1_in, 0x80);                       \
+                                                                   \
+    filt = __msa_subs_s_b(p1_m, q1_m);                             \
+                                                                   \
+    q0_sub_p0 = q0_m - p0_m;                                       \
+    filt_sign = __msa_clti_s_b(filt, 0);                           \
+                                                                   \
+    cnst3h = __msa_ldi_h(3);                                       \
+    q0_sub_p0_sign = __msa_clti_s_b(q0_sub_p0, 0);                 \
+    q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0_sign, q0_sub_p0);  \
+    q0_sub_p0_r *= cnst3h;                                         \
+    filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt);                 \
+    filt_r += q0_sub_p0_r;                                         \
+    filt_r = __msa_sat_s_h(filt_r, 7);                             \
+                                                                   \
+    q0_sub_p0_l = (v8i16)__msa_ilvl_b(q0_sub_p0_sign, q0_sub_p0);  \
+    q0_sub_p0_l *= cnst3h;                                         \
+    filt_l = (v8i16)__msa_ilvl_b(filt_sign, filt);                 \
+    filt_l += q0_sub_p0_l;                                         \
+    filt_l = __msa_sat_s_h(filt_l, 7);                             \
+                                                                   \
+    filt = __msa_pckev_b((v16i8)filt_l, (v16i8)filt_r);            \
+    filt = filt & (v16i8)(mask);                                   \
+                                                                   \
+    cnst4b = __msa_ldi_b(4);                                       \
+    filt1 = __msa_adds_s_b(filt, cnst4b);                          \
+    filt1 >>= 3;                                                   \
+                                                                   \
+    cnst3b = __msa_ldi_b(3);                                       \
+    filt2 = __msa_adds_s_b(filt, cnst3b);                          \
+    filt2 >>= 3;                                                   \
+                                                                   \
+    q0_m = __msa_subs_s_b(q0_m, filt1);                            \
+    p0_m = __msa_adds_s_b(p0_m, filt2);                            \
+    q0_in = __msa_xori_b((v16u8)q0_m, 0x80);                       \
+    p0_in = __msa_xori_b((v16u8)p0_m, 0x80);                       \
+}
+
+#define VP8_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev)            \
+{                                                                  \
+    v16i8 p2_m, p1_m, p0_m, q2_m, q1_m, q0_m;                      \
+    v16i8 filt, q0_sub_p0, cnst4b, cnst3b;                         \
+    v16i8 u, filt1, filt2, filt_sign, q0_sub_p0_sign;              \
+    v8i16 q0_sub_p0_r, q0_sub_p0_l, filt_r, u_r, u_l, filt_l;      \
+    v8i16 cnst3h, cnst27h, cnst18h, cnst63h;                       \
+                                                                   \
+    cnst3h = __msa_ldi_h(3);                                       \
+                                                                   \
+    p2_m = (v16i8)__msa_xori_b(p2, 0x80);                          \
+    p1_m = (v16i8)__msa_xori_b(p1, 0x80);                          \
+    p0_m = (v16i8)__msa_xori_b(p0, 0x80);                          \
+    q0_m = (v16i8)__msa_xori_b(q0, 0x80);                          \
+    q1_m = (v16i8)__msa_xori_b(q1, 0x80);                          \
+    q2_m = (v16i8)__msa_xori_b(q2, 0x80);                          \
+                                                                   \
+    filt = __msa_subs_s_b(p1_m, q1_m);                             \
+    q0_sub_p0 = q0_m - p0_m;                                       \
+    q0_sub_p0_sign = __msa_clti_s_b(q0_sub_p0, 0);                 \
+    filt_sign = __msa_clti_s_b(filt, 0);                           \
+                                                                   \
+    q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0_sign, q0_sub_p0);  \
+    q0_sub_p0_r *= cnst3h;                                         \
+    filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt);                 \
+    filt_r = filt_r + q0_sub_p0_r;                                 \
+    filt_r = __msa_sat_s_h(filt_r, 7);                             \
+                                                                   \
+    q0_sub_p0_l = (v8i16)__msa_ilvl_b(q0_sub_p0_sign, q0_sub_p0);  \
+    q0_sub_p0_l *= cnst3h;                                         \
+    filt_l = (v8i16)__msa_ilvl_b(filt_sign, filt);                 \
+    filt_l = filt_l + q0_sub_p0_l;                                 \
+    filt_l = __msa_sat_s_h(filt_l, 7);                             \
+                                                                   \
+    filt = __msa_pckev_b((v16i8)filt_l, (v16i8)filt_r);            \
+    filt = filt & (v16i8)mask;                                     \
+    filt2 = filt & (v16i8)hev;                                     \
+                                                                   \
+    hev = __msa_xori_b(hev, 0xff);                                 \
+    filt = filt & (v16i8)hev;                                      \
+    cnst4b = __msa_ldi_b(4);                                       \
+    filt1 = __msa_adds_s_b(filt2, cnst4b);                         \
+    filt1 >>= 3;                                                   \
+    cnst3b = __msa_ldi_b(3);                                       \
+    filt2 = __msa_adds_s_b(filt2, cnst3b);                         \
+    filt2 >>= 3;                                                   \
+    q0_m = __msa_subs_s_b(q0_m, filt1);                            \
+    p0_m = __msa_adds_s_b(p0_m, filt2);                            \
+                                                                   \
+    filt_sign = __msa_clti_s_b(filt, 0);                           \
+    ILVRL_B2_SH(filt_sign, filt, filt_r, filt_l);                  \
+                                                                   \
+    cnst27h = __msa_ldi_h(27);                                     \
+    cnst63h = __msa_ldi_h(63);                                     \
+                                                                   \
+    u_r = filt_r * cnst27h;                                        \
+    u_r += cnst63h;                                                \
+    u_r >>= 7;                                                     \
+    u_r = __msa_sat_s_h(u_r, 7);                                   \
+    u_l = filt_l * cnst27h;                                        \
+    u_l += cnst63h;                                                \
+    u_l >>= 7;                                                     \
+    u_l = __msa_sat_s_h(u_l, 7);                                   \
+    u = __msa_pckev_b((v16i8)u_l, (v16i8)u_r);                     \
+    q0_m = __msa_subs_s_b(q0_m, u);                                \
+    q0 = __msa_xori_b((v16u8)q0_m, 0x80);                          \
+    p0_m = __msa_adds_s_b(p0_m, u);                                \
+    p0 = __msa_xori_b((v16u8)p0_m, 0x80);                          \
+    cnst18h = __msa_ldi_h(18);                                     \
+    u_r = filt_r * cnst18h;                                        \
+    u_r += cnst63h;                                                \
+    u_r >>= 7;                                                     \
+    u_r = __msa_sat_s_h(u_r, 7);                                   \
+                                                                   \
+    u_l = filt_l * cnst18h;                                        \
+    u_l += cnst63h;                                                \
+    u_l >>= 7;                                                     \
+    u_l = __msa_sat_s_h(u_l, 7);                                   \
+    u = __msa_pckev_b((v16i8)u_l, (v16i8)u_r);                     \
+    q1_m = __msa_subs_s_b(q1_m, u);                                \
+    q1 = __msa_xori_b((v16u8)q1_m, 0x80);                          \
+    p1_m = __msa_adds_s_b(p1_m, u);                                \
+    p1 = __msa_xori_b((v16u8)p1_m, 0x80);                          \
+    u_r = filt_r << 3;                                             \
+    u_r += filt_r + cnst63h;                                       \
+    u_r >>= 7;                                                     \
+    u_r = __msa_sat_s_h(u_r, 7);                                   \
+                                                                   \
+    u_l = filt_l << 3;                                             \
+    u_l += filt_l + cnst63h;                                       \
+    u_l >>= 7;                                                     \
+    u_l = __msa_sat_s_h(u_l, 7);                                   \
+    u = __msa_pckev_b((v16i8)u_l, (v16i8)u_r);                     \
+    q2_m = __msa_subs_s_b(q2_m, u);                                \
+    q2 = __msa_xori_b((v16u8)q2_m, 0x80);                          \
+    p2_m = __msa_adds_s_b(p2_m, u);                                \
+    p2 = __msa_xori_b((v16u8)p2_m, 0x80);                          \
+}
+
+#define LPF_MASK_HEV(p3_in, p2_in, p1_in, p0_in,                   \
+                     q0_in, q1_in, q2_in, q3_in,                   \
+                     limit_in, b_limit_in, thresh_in,              \
+                     hev_out, mask_out, flat_out)                  \
+{                                                                  \
+    v16u8 p3_asub_p2_m, p2_asub_p1_m, p1_asub_p0_m, q1_asub_q0_m;  \
+    v16u8 p1_asub_q1_m, p0_asub_q0_m, q3_asub_q2_m, q2_asub_q1_m;  \
+                                                                   \
+    p3_asub_p2_m = __msa_asub_u_b((p3_in), (p2_in));               \
+    p2_asub_p1_m = __msa_asub_u_b((p2_in), (p1_in));               \
+    p1_asub_p0_m = __msa_asub_u_b((p1_in), (p0_in));               \
+    q1_asub_q0_m = __msa_asub_u_b((q1_in), (q0_in));               \
+    q2_asub_q1_m = __msa_asub_u_b((q2_in), (q1_in));               \
+    q3_asub_q2_m = __msa_asub_u_b((q3_in), (q2_in));               \
+    p0_asub_q0_m = __msa_asub_u_b((p0_in), (q0_in));               \
+    p1_asub_q1_m = __msa_asub_u_b((p1_in), (q1_in));               \
+    flat_out = __msa_max_u_b(p1_asub_p0_m, q1_asub_q0_m);          \
+    hev_out = (thresh_in) < (v16u8)flat_out;                       \
+    p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p0_asub_q0_m);     \
+    p1_asub_q1_m >>= 1;                                            \
+    p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p1_asub_q1_m);     \
+    mask_out = (b_limit_in) < p0_asub_q0_m;                        \
+    mask_out = __msa_max_u_b(flat_out, mask_out);                  \
+    p3_asub_p2_m = __msa_max_u_b(p3_asub_p2_m, p2_asub_p1_m);      \
+    mask_out = __msa_max_u_b(p3_asub_p2_m, mask_out);              \
+    q2_asub_q1_m = __msa_max_u_b(q2_asub_q1_m, q3_asub_q2_m);      \
+    mask_out = __msa_max_u_b(q2_asub_q1_m, mask_out);              \
+    mask_out = (limit_in) < (v16u8)mask_out;                       \
+    mask_out = __msa_xori_b(mask_out, 0xff);                       \
+}
+
+#define VP8_ST6x1_UB(in0, in0_idx, in1, in1_idx, pdst, stride)  \
+{                                                               \
+    uint16_t tmp0_h;                                            \
+    uint32_t tmp0_w;                                            \
+                                                                \
+    tmp0_w = __msa_copy_u_w((v4i32)in0, in0_idx);               \
+    tmp0_h = __msa_copy_u_h((v8i16)in1, in1_idx);               \
+    SW(tmp0_w, pdst);                                           \
+    SH(tmp0_h, pdst + stride);                                  \
+}
+
+
+static void loop_filter_horizontal_4_dual_msa(uint8_t *src, int32_t pitch,
+                                              const uint8_t *b_limit0_ptr,
+                                              const uint8_t *limit0_ptr,
+                                              const uint8_t *thresh0_ptr,
+                                              const uint8_t *b_limit1_ptr,
+                                              const uint8_t *limit1_ptr,
+                                              const uint8_t *thresh1_ptr)
+{
+    v16u8 mask, hev, flat;
+    v16u8 thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
+    v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+
+    LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+    thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
+    thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
+    thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
+
+    b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
+    b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
+    b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
+
+    limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
+    limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
+    limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
+
+    LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
+                 hev, mask, flat);
+    VP8_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
+
+    ST_UB4(p1, p0, q0, q1, (src - 2 * pitch), pitch);
+}
+
+static void loop_filter_vertical_4_dual_msa(uint8_t *src, int32_t pitch,
+                                            const uint8_t *b_limit0_ptr,
+                                            const uint8_t *limit0_ptr,
+                                            const uint8_t *thresh0_ptr,
+                                            const uint8_t *b_limit1_ptr,
+                                            const uint8_t *limit1_ptr,
+                                            const uint8_t *thresh1_ptr)
+{
+    v16u8 mask, hev, flat;
+    v16u8 thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
+    v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+    v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
+    v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
+    v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+
+    LD_UB8(src - 4, pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+    LD_UB8(src - 4 + (8 * pitch), pitch,
+           row8, row9, row10, row11, row12, row13, row14, row15);
+    TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
+                        row8, row9, row10, row11, row12, row13, row14, row15,
+                        p3, p2, p1, p0, q0, q1, q2, q3);
+
+    thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
+    thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
+    thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
+
+    b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
+    b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
+    b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
+
+    limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
+    limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
+    limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
+
+    LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
+                 hev, mask, flat);
+    VP8_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
+    ILVR_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
+    ILVRL_H2_SH(tmp1, tmp0, tmp2, tmp3);
+    ILVL_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
+    ILVRL_H2_SH(tmp1, tmp0, tmp4, tmp5);
+
+    src -= 2;
+    ST4x8_UB(tmp2, tmp3, src, pitch);
+    src += (8 * pitch);
+    ST4x8_UB(tmp4, tmp5, src, pitch);
+}
+
+static void mbloop_filter_horizontal_edge_y_msa(uint8_t *src, int32_t pitch,
+                                                const uint8_t b_limit_in,
+                                                const uint8_t limit_in,
+                                                const uint8_t thresh_in)
+{
+    uint8_t *temp_src;
+    v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+    v16u8 mask, hev, flat, thresh, limit, b_limit;
+
+    b_limit = (v16u8)__msa_fill_b(b_limit_in);
+    limit = (v16u8)__msa_fill_b(limit_in);
+    thresh = (v16u8)__msa_fill_b(thresh_in);
+    temp_src = src - (pitch << 2);
+    LD_UB8(temp_src, pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+    LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+                 hev, mask, flat);
+    VP8_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
+    temp_src = src - 3 * pitch;
+    ST_UB4(p2, p1, p0, q0, temp_src, pitch);
+    temp_src += (4 * pitch);
+    ST_UB2(q1, q2, temp_src, pitch);
+}
+
+static void mbloop_filter_horizontal_edge_uv_msa(uint8_t *src_u, uint8_t *src_v,
+                                                 int32_t pitch,
+                                                 const uint8_t b_limit_in,
+                                                 const uint8_t limit_in,
+                                                 const uint8_t thresh_in)
+{
+    uint8_t *temp_src;
+    uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
+    v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+    v16u8 mask, hev, flat, thresh, limit, b_limit;
+    v16u8 p3_u, p2_u, p1_u, p0_u, q3_u, q2_u, q1_u, q0_u;
+    v16u8 p3_v, p2_v, p1_v, p0_v, q3_v, q2_v, q1_v, q0_v;
+
+    b_limit = (v16u8)__msa_fill_b(b_limit_in);
+    limit = (v16u8)__msa_fill_b(limit_in);
+    thresh = (v16u8)__msa_fill_b(thresh_in);
+
+    temp_src = src_u - (pitch << 2);
+    LD_UB8(temp_src, pitch, p3_u, p2_u, p1_u, p0_u, q0_u, q1_u, q2_u, q3_u);
+    temp_src = src_v - (pitch << 2);
+    LD_UB8(temp_src, pitch, p3_v, p2_v, p1_v, p0_v, q0_v, q1_v, q2_v, q3_v);
+
+    ILVR_D4_UB(p3_v, p3_u, p2_v, p2_u, p1_v, p1_u, p0_v, p0_u, p3, p2, p1, p0);
+    ILVR_D4_UB(q0_v, q0_u, q1_v, q1_u, q2_v, q2_u, q3_v, q3_u, q0, q1, q2, q3);
+    LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+                 hev, mask, flat);
+    VP8_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
+
+    p2_d = __msa_copy_u_d((v2i64)p2, 0);
+    p1_d = __msa_copy_u_d((v2i64)p1, 0);
+    p0_d = __msa_copy_u_d((v2i64)p0, 0);
+    q0_d = __msa_copy_u_d((v2i64)q0, 0);
+    q1_d = __msa_copy_u_d((v2i64)q1, 0);
+    q2_d = __msa_copy_u_d((v2i64)q2, 0);
+    src_u -= (pitch * 3);
+    SD4(p2_d, p1_d, p0_d, q0_d, src_u, pitch);
+    src_u += 4 * pitch;
+    SD(q1_d, src_u);
+    src_u += pitch;
+    SD(q2_d, src_u);
+
+    p2_d = __msa_copy_u_d((v2i64)p2, 1);
+    p1_d = __msa_copy_u_d((v2i64)p1, 1);
+    p0_d = __msa_copy_u_d((v2i64)p0, 1);
+    q0_d = __msa_copy_u_d((v2i64)q0, 1);
+    q1_d = __msa_copy_u_d((v2i64)q1, 1);
+    q2_d = __msa_copy_u_d((v2i64)q2, 1);
+    src_v -= (pitch * 3);
+    SD4(p2_d, p1_d, p0_d, q0_d, src_v, pitch);
+    src_v += 4 * pitch;
+    SD(q1_d, src_v);
+    src_v += pitch;
+    SD(q2_d, src_v);
+}
+
+static void mbloop_filter_vertical_edge_y_msa(uint8_t *src, int32_t pitch,
+                                              const uint8_t b_limit_in,
+                                              const uint8_t limit_in,
+                                              const uint8_t thresh_in)
+{
+    uint8_t *temp_src;
+    v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+    v16u8 mask, hev, flat, thresh, limit, b_limit;
+    v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
+    v16u8 row9, row10, row11, row12, row13, row14, row15;
+    v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+    b_limit = (v16u8)__msa_fill_b(b_limit_in);
+    limit = (v16u8)__msa_fill_b(limit_in);
+    thresh = (v16u8)__msa_fill_b(thresh_in);
+    temp_src = src - 4;
+    LD_UB8(temp_src, pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+    temp_src += (8 * pitch);
+    LD_UB8(temp_src, pitch,
+           row8, row9, row10, row11, row12, row13, row14, row15);
+    TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
+                        row8, row9, row10, row11, row12, row13, row14, row15,
+                        p3, p2, p1, p0, q0, q1, q2, q3);
+
+    LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+                 hev, mask, flat);
+    VP8_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
+    ILVR_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
+    ILVRL_H2_SH(tmp1, tmp0, tmp3, tmp4);
+    ILVL_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
+    ILVRL_H2_SH(tmp1, tmp0, tmp6, tmp7);
+    ILVRL_B2_SH(q2, q1, tmp2, tmp5);
+
+    temp_src = src - 3;
+    VP8_ST6x1_UB(tmp3, 0, tmp2, 0, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp3, 1, tmp2, 1, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp3, 2, tmp2, 2, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp3, 3, tmp2, 3, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp4, 0, tmp2, 4, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp4, 1, tmp2, 5, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp4, 2, tmp2, 6, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp4, 3, tmp2, 7, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp6, 0, tmp5, 0, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp6, 1, tmp5, 1, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp6, 2, tmp5, 2, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp6, 3, tmp5, 3, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp7, 0, tmp5, 4, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp7, 1, tmp5, 5, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp7, 2, tmp5, 6, temp_src, 4);
+    temp_src += pitch;
+    VP8_ST6x1_UB(tmp7, 3, tmp5, 7, temp_src, 4);
+}
+
+static void mbloop_filter_vertical_edge_uv_msa(uint8_t *src_u, uint8_t *src_v,
+                                               int32_t pitch,
+                                               const uint8_t b_limit_in,
+                                               const uint8_t limit_in,
+                                               const uint8_t thresh_in)
+{
+    v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+    v16u8 mask, hev, flat, thresh, limit, b_limit;
+    v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
+    v16u8 row9, row10, row11, row12, row13, row14, row15;
+    v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+    b_limit = (v16u8)__msa_fill_b(b_limit_in);
+    limit = (v16u8)__msa_fill_b(limit_in);
+    thresh = (v16u8)__msa_fill_b(thresh_in);
+
+    LD_UB8(src_u - 4, pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+    LD_UB8(src_v - 4, pitch,
+           row8, row9, row10, row11, row12, row13, row14, row15);
+    TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
+                        row8, row9, row10, row11, row12, row13, row14, row15,
+                        p3, p2, p1, p0, q0, q1, q2, q3);
+
+    LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+                 hev, mask, flat);
+    VP8_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
+
+    ILVR_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
+    ILVRL_H2_SH(tmp1, tmp0, tmp3, tmp4);
+    ILVL_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
+    ILVRL_H2_SH(tmp1, tmp0, tmp6, tmp7);
+    ILVRL_B2_SH(q2, q1, tmp2, tmp5);
+
+    src_u -= 3;
+    VP8_ST6x1_UB(tmp3, 0, tmp2, 0, src_u, 4);
+    src_u += pitch;
+    VP8_ST6x1_UB(tmp3, 1, tmp2, 1, src_u, 4);
+    src_u += pitch;
+    VP8_ST6x1_UB(tmp3, 2, tmp2, 2, src_u, 4);
+    src_u += pitch;
+    VP8_ST6x1_UB(tmp3, 3, tmp2, 3, src_u, 4);
+    src_u += pitch;
+    VP8_ST6x1_UB(tmp4, 0, tmp2, 4, src_u, 4);
+    src_u += pitch;
+    VP8_ST6x1_UB(tmp4, 1, tmp2, 5, src_u, 4);
+    src_u += pitch;
+    VP8_ST6x1_UB(tmp4, 2, tmp2, 6, src_u, 4);
+    src_u += pitch;
+    VP8_ST6x1_UB(tmp4, 3, tmp2, 7, src_u, 4);
+
+    src_v -= 3;
+    VP8_ST6x1_UB(tmp6, 0, tmp5, 0, src_v, 4);
+    src_v += pitch;
+    VP8_ST6x1_UB(tmp6, 1, tmp5, 1, src_v, 4);
+    src_v += pitch;
+    VP8_ST6x1_UB(tmp6, 2, tmp5, 2, src_v, 4);
+    src_v += pitch;
+    VP8_ST6x1_UB(tmp6, 3, tmp5, 3, src_v, 4);
+    src_v += pitch;
+    VP8_ST6x1_UB(tmp7, 0, tmp5, 4, src_v, 4);
+    src_v += pitch;
+    VP8_ST6x1_UB(tmp7, 1, tmp5, 5, src_v, 4);
+    src_v += pitch;
+    VP8_ST6x1_UB(tmp7, 2, tmp5, 6, src_v, 4);
+    src_v += pitch;
+    VP8_ST6x1_UB(tmp7, 3, tmp5, 7, src_v, 4);
+}
+
+void vp8_loop_filter_simple_horizontal_edge_msa(uint8_t *src, int32_t pitch,
+                                                const uint8_t *b_limit_ptr)
+{
+    v16u8 p1, p0, q1, q0;
+    v16u8 mask, b_limit;
+
+    b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+    LD_UB4(src - (pitch << 1), pitch, p1, p0, q0, q1);
+    VP8_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask);
+    VP8_SIMPLE_FILT(p1, p0, q0, q1, mask);
+    ST_UB2(p0, q0, (src - pitch), pitch);
+}
+
+void vp8_loop_filter_simple_vertical_edge_msa(uint8_t *src, int32_t pitch,
+                                              const uint8_t *b_limit_ptr)
+{
+    uint8_t *temp_src;
+    v16u8 p1, p0, q1, q0;
+    v16u8 mask, b_limit;
+    v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
+    v16u8 row9, row10, row11, row12, row13, row14, row15;
+    v8i16 tmp0, tmp1;
+
+    b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+    temp_src = src - 2;
+    LD_UB8(temp_src, pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+    temp_src += (8 * pitch);
+    LD_UB8(temp_src, pitch,
+           row8, row9, row10, row11, row12, row13, row14, row15);
+    TRANSPOSE16x4_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
+                        row8, row9, row10, row11, row12, row13, row14, row15,
+                        p1, p0, q0, q1);
+    VP8_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask);
+    VP8_SIMPLE_FILT(p1, p0, q0, q1, mask);
+    ILVRL_B2_SH(q0, p0, tmp1, tmp0);
+
+    src -= 1;
+    ST2x4_UB(tmp1, 0, src, pitch);
+    src += 4 * pitch;
+    ST2x4_UB(tmp1, 4, src, pitch);
+    src += 4 * pitch;
+    ST2x4_UB(tmp0, 0, src, pitch);
+    src += 4 * pitch;
+    ST2x4_UB(tmp0, 4, src, pitch);
+    src += 4 * pitch;
+}
+
+static void loop_filter_horizontal_edge_uv_msa(uint8_t *src_u, uint8_t *src_v,
+                                               int32_t pitch,
+                                               const uint8_t b_limit_in,
+                                               const uint8_t limit_in,
+                                               const uint8_t thresh_in)
+{
+    uint64_t p1_d, p0_d, q0_d, q1_d;
+    v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+    v16u8 mask, hev, flat, thresh, limit, b_limit;
+    v16u8 p3_u, p2_u, p1_u, p0_u, q3_u, q2_u, q1_u, q0_u;
+    v16u8 p3_v, p2_v, p1_v, p0_v, q3_v, q2_v, q1_v, q0_v;
+
+    thresh = (v16u8)__msa_fill_b(thresh_in);
+    limit = (v16u8)__msa_fill_b(limit_in);
+    b_limit = (v16u8)__msa_fill_b(b_limit_in);
+
+    src_u = src_u - (pitch << 2);
+    LD_UB8(src_u, pitch, p3_u, p2_u, p1_u, p0_u, q0_u, q1_u, q2_u, q3_u);
+    src_u += (5 * pitch);
+    src_v = src_v - (pitch << 2);
+    LD_UB8(src_v, pitch, p3_v, p2_v, p1_v, p0_v, q0_v, q1_v, q2_v, q3_v);
+    src_v += (5 * pitch);
+
+    /* right 8 element of p3 are u pixel and
+       left 8 element of p3 are v pixel */
+    ILVR_D4_UB(p3_v, p3_u, p2_v, p2_u, p1_v, p1_u, p0_v, p0_u, p3, p2, p1, p0);
+    ILVR_D4_UB(q0_v, q0_u, q1_v, q1_u, q2_v, q2_u, q3_v, q3_u, q0, q1, q2, q3);
+    LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+                 hev, mask, flat);
+    VP8_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
+
+    p1_d = __msa_copy_u_d((v2i64)p1, 0);
+    p0_d = __msa_copy_u_d((v2i64)p0, 0);
+    q0_d = __msa_copy_u_d((v2i64)q0, 0);
+    q1_d = __msa_copy_u_d((v2i64)q1, 0);
+    SD4(q1_d, q0_d, p0_d, p1_d, src_u, (- pitch));
+
+    p1_d = __msa_copy_u_d((v2i64)p1, 1);
+    p0_d = __msa_copy_u_d((v2i64)p0, 1);
+    q0_d = __msa_copy_u_d((v2i64)q0, 1);
+    q1_d = __msa_copy_u_d((v2i64)q1, 1);
+    SD4(q1_d, q0_d, p0_d, p1_d, src_v, (- pitch));
+}
+
+static void loop_filter_vertical_edge_uv_msa(uint8_t *src_u, uint8_t *src_v,
+                                             int32_t pitch,
+                                             const uint8_t b_limit_in,
+                                             const uint8_t limit_in,
+                                             const uint8_t thresh_in)
+{
+    uint8_t *temp_src_u, *temp_src_v;
+    v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+    v16u8 mask, hev, flat, thresh, limit, b_limit;
+    v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
+    v16u8 row9, row10, row11, row12, row13, row14, row15;
+    v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+
+    thresh = (v16u8)__msa_fill_b(thresh_in);
+    limit = (v16u8)__msa_fill_b(limit_in);
+    b_limit = (v16u8)__msa_fill_b(b_limit_in);
+
+    LD_UB8(src_u - 4, pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+    LD_UB8(src_v - 4, pitch,
+           row8, row9, row10, row11, row12, row13, row14, row15);
+    TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
+                        row8, row9, row10, row11, row12, row13, row14, row15,
+                        p3, p2, p1, p0, q0, q1, q2, q3);
+
+    LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+                 hev, mask, flat);
+    VP8_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
+    ILVR_B2_SW(p0, p1, q1, q0, tmp0, tmp1);
+    ILVRL_H2_SW(tmp1, tmp0, tmp2, tmp3);
+    tmp0 = (v4i32)__msa_ilvl_b((v16i8)p0, (v16i8)p1);
+    tmp1 = (v4i32)__msa_ilvl_b((v16i8)q1, (v16i8)q0);
+    ILVRL_H2_SW(tmp1, tmp0, tmp4, tmp5);
+
+    temp_src_u = src_u - 2;
+    ST4x4_UB(tmp2, tmp2, 0, 1, 2, 3, temp_src_u, pitch);
+    temp_src_u += 4 * pitch;
+    ST4x4_UB(tmp3, tmp3, 0, 1, 2, 3, temp_src_u, pitch);
+
+    temp_src_v = src_v - 2;
+    ST4x4_UB(tmp4, tmp4, 0, 1, 2, 3, temp_src_v, pitch);
+    temp_src_v += 4 * pitch;
+    ST4x4_UB(tmp5, tmp5, 0, 1, 2, 3, temp_src_v, pitch);
+}
+
+void vp8_loop_filter_mbh_msa(uint8_t *src_y, uint8_t *src_u,
+                             uint8_t *src_v, int32_t pitch_y,
+                             int32_t pitch_u_v,
+                             loop_filter_info *lpf_info_ptr)
+{
+    mbloop_filter_horizontal_edge_y_msa(src_y, pitch_y,
+                                        *lpf_info_ptr->mblim,
+                                        *lpf_info_ptr->lim,
+                                        *lpf_info_ptr->hev_thr);
+    if (src_u)
+    {
+        mbloop_filter_horizontal_edge_uv_msa(src_u, src_v, pitch_u_v,
+                                             *lpf_info_ptr->mblim,
+                                             *lpf_info_ptr->lim,
+                                             *lpf_info_ptr->hev_thr);
+    }
+}
+
+void vp8_loop_filter_mbv_msa(uint8_t *src_y, uint8_t *src_u,
+                             uint8_t *src_v, int32_t pitch_y,
+                             int32_t pitch_u_v,
+                             loop_filter_info *lpf_info_ptr)
+{
+    mbloop_filter_vertical_edge_y_msa(src_y, pitch_y,
+                                      *lpf_info_ptr->mblim,
+                                      *lpf_info_ptr->lim,
+                                      *lpf_info_ptr->hev_thr);
+    if (src_u)
+    {
+        mbloop_filter_vertical_edge_uv_msa(src_u, src_v, pitch_u_v,
+                                           *lpf_info_ptr->mblim,
+                                           *lpf_info_ptr->lim,
+                                           *lpf_info_ptr->hev_thr);
+    }
+}
+
+void vp8_loop_filter_bh_msa(uint8_t *src_y, uint8_t *src_u,
+                            uint8_t *src_v, int32_t pitch_y,
+                            int32_t pitch_u_v,
+                            loop_filter_info *lpf_info_ptr)
+{
+    loop_filter_horizontal_4_dual_msa(src_y + 4 * pitch_y, pitch_y,
+                                      lpf_info_ptr->blim,
+                                      lpf_info_ptr->lim,
+                                      lpf_info_ptr->hev_thr,
+                                      lpf_info_ptr->blim,
+                                      lpf_info_ptr->lim,
+                                      lpf_info_ptr->hev_thr);
+    loop_filter_horizontal_4_dual_msa(src_y + 8 * pitch_y, pitch_y,
+                                      lpf_info_ptr->blim,
+                                      lpf_info_ptr->lim,
+                                      lpf_info_ptr->hev_thr,
+                                      lpf_info_ptr->blim,
+                                      lpf_info_ptr->lim,
+                                      lpf_info_ptr->hev_thr);
+    loop_filter_horizontal_4_dual_msa(src_y + 12 * pitch_y, pitch_y,
+                                      lpf_info_ptr->blim,
+                                      lpf_info_ptr->lim,
+                                      lpf_info_ptr->hev_thr,
+                                      lpf_info_ptr->blim,
+                                      lpf_info_ptr->lim,
+                                      lpf_info_ptr->hev_thr);
+    if (src_u)
+    {
+        loop_filter_horizontal_edge_uv_msa(src_u + (4 * pitch_u_v),
+                                           src_v + (4 * pitch_u_v),
+                                           pitch_u_v,
+                                           *lpf_info_ptr->blim,
+                                           *lpf_info_ptr->lim,
+                                           *lpf_info_ptr->hev_thr);
+    }
+}
+
+void vp8_loop_filter_bv_msa(uint8_t *src_y, uint8_t *src_u,
+                            uint8_t *src_v, int32_t pitch_y,
+                            int32_t pitch_u_v,
+                            loop_filter_info *lpf_info_ptr)
+{
+    loop_filter_vertical_4_dual_msa(src_y + 4, pitch_y, lpf_info_ptr->blim,
+                                    lpf_info_ptr->lim,
+                                    lpf_info_ptr->hev_thr,
+                                    lpf_info_ptr->blim,
+                                    lpf_info_ptr->lim,
+                                    lpf_info_ptr->hev_thr);
+    loop_filter_vertical_4_dual_msa(src_y + 8, pitch_y,
+                                    lpf_info_ptr->blim,
+                                    lpf_info_ptr->lim,
+                                    lpf_info_ptr->hev_thr,
+                                    lpf_info_ptr->blim,
+                                    lpf_info_ptr->lim,
+                                    lpf_info_ptr->hev_thr);
+    loop_filter_vertical_4_dual_msa(src_y + 12, pitch_y,
+                                    lpf_info_ptr->blim,
+                                    lpf_info_ptr->lim,
+                                    lpf_info_ptr->hev_thr,
+                                    lpf_info_ptr->blim,
+                                    lpf_info_ptr->lim,
+                                    lpf_info_ptr->hev_thr);
+    if (src_u)
+    {
+        loop_filter_vertical_edge_uv_msa(src_u + 4, src_v + 4, pitch_u_v,
+                                         *lpf_info_ptr->blim,
+                                         *lpf_info_ptr->lim,
+                                         *lpf_info_ptr->hev_thr);
+    }
+}
+
+void vp8_loop_filter_bhs_msa(uint8_t *src_y, int32_t pitch_y,
+                             const uint8_t *b_limit_ptr)
+{
+    vp8_loop_filter_simple_horizontal_edge_msa(src_y + (4 * pitch_y),
+                                               pitch_y, b_limit_ptr);
+    vp8_loop_filter_simple_horizontal_edge_msa(src_y + (8 * pitch_y),
+                                               pitch_y, b_limit_ptr);
+    vp8_loop_filter_simple_horizontal_edge_msa(src_y + (12 * pitch_y),
+                                               pitch_y, b_limit_ptr);
+}
+
+void vp8_loop_filter_bvs_msa(uint8_t *src_y, int32_t pitch_y,
+                             const uint8_t *b_limit_ptr)
+{
+    vp8_loop_filter_simple_vertical_edge_msa(src_y + 4, pitch_y, b_limit_ptr);
+    vp8_loop_filter_simple_vertical_edge_msa(src_y + 8, pitch_y, b_limit_ptr);
+    vp8_loop_filter_simple_vertical_edge_msa(src_y + 12, pitch_y, b_limit_ptr);
+}
diff --git a/libvpx/libvpx/vp8/common/mips/msa/mfqe_msa.c b/libvpx/libvpx/vp8/common/mips/msa/mfqe_msa.c
new file mode 100644
index 0000000..3e7629f
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/msa/mfqe_msa.c
@@ -0,0 +1,146 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+#include "vp8/common/postproc.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+
+static void filter_by_weight8x8_msa(uint8_t *src_ptr, int32_t src_stride,
+                                    uint8_t *dst_ptr, int32_t dst_stride,
+                                    int32_t src_weight)
+{
+    int32_t dst_weight = (1 << MFQE_PRECISION) - src_weight;
+    int32_t row;
+    uint64_t src0_d, src1_d, dst0_d, dst1_d;
+    v16i8 src0 = { 0 };
+    v16i8 src1 = { 0 };
+    v16i8 dst0 = { 0 };
+    v16i8 dst1 = { 0 };
+    v8i16 src_wt, dst_wt, res_h_r, res_h_l, src_r, src_l, dst_r, dst_l;
+
+    src_wt = __msa_fill_h(src_weight);
+    dst_wt = __msa_fill_h(dst_weight);
+
+    for (row = 2; row--;)
+    {
+        LD2(src_ptr, src_stride, src0_d, src1_d);
+        src_ptr += (2 * src_stride);
+        LD2(dst_ptr, dst_stride, dst0_d, dst1_d);
+        INSERT_D2_SB(src0_d, src1_d, src0);
+        INSERT_D2_SB(dst0_d, dst1_d, dst0);
+
+        LD2(src_ptr, src_stride, src0_d, src1_d);
+        src_ptr += (2 * src_stride);
+        LD2((dst_ptr + 2 * dst_stride), dst_stride, dst0_d, dst1_d);
+        INSERT_D2_SB(src0_d, src1_d, src1);
+        INSERT_D2_SB(dst0_d, dst1_d, dst1);
+
+        UNPCK_UB_SH(src0, src_r, src_l);
+        UNPCK_UB_SH(dst0, dst_r, dst_l);
+        res_h_r = (src_r * src_wt);
+        res_h_r += (dst_r * dst_wt);
+        res_h_l = (src_l * src_wt);
+        res_h_l += (dst_l * dst_wt);
+        SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+        dst0 = (v16i8)__msa_pckev_b((v16i8)res_h_l, (v16i8)res_h_r);
+        ST8x2_UB(dst0, dst_ptr, dst_stride);
+        dst_ptr += (2 * dst_stride);
+
+        UNPCK_UB_SH(src1, src_r, src_l);
+        UNPCK_UB_SH(dst1, dst_r, dst_l);
+        res_h_r = (src_r * src_wt);
+        res_h_r += (dst_r * dst_wt);
+        res_h_l = (src_l * src_wt);
+        res_h_l += (dst_l * dst_wt);
+        SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+        dst1 = (v16i8)__msa_pckev_b((v16i8)res_h_l, (v16i8)res_h_r);
+        ST8x2_UB(dst1, dst_ptr, dst_stride);
+        dst_ptr += (2 * dst_stride);
+    }
+}
+
+static void filter_by_weight16x16_msa(uint8_t *src_ptr, int32_t src_stride,
+                                      uint8_t *dst_ptr, int32_t dst_stride,
+                                      int32_t src_weight)
+{
+    int32_t dst_weight = (1 << MFQE_PRECISION) - src_weight;
+    int32_t row;
+    v16i8 src0, src1, src2, src3;
+    v16i8 dst0, dst1, dst2, dst3;
+    v8i16 src_wt, dst_wt;
+    v8i16 res_h_r, res_h_l;
+    v8i16 src_r, src_l, dst_r, dst_l;
+
+    src_wt = __msa_fill_h(src_weight);
+    dst_wt = __msa_fill_h(dst_weight);
+
+    for (row = 4; row--;)
+    {
+        LD_SB4(src_ptr, src_stride, src0, src1, src2, src3);
+        src_ptr += (4 * src_stride);
+        LD_SB4(dst_ptr, dst_stride, dst0, dst1, dst2, dst3);
+
+        UNPCK_UB_SH(src0, src_r, src_l);
+        UNPCK_UB_SH(dst0, dst_r, dst_l);
+        res_h_r = (src_r * src_wt);
+        res_h_r += (dst_r * dst_wt);
+        res_h_l = (src_l * src_wt);
+        res_h_l += (dst_l * dst_wt);
+        SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+        PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+        dst_ptr += dst_stride;
+
+        UNPCK_UB_SH(src1, src_r, src_l);
+        UNPCK_UB_SH(dst1, dst_r, dst_l);
+        res_h_r = (src_r * src_wt);
+        res_h_r += (dst_r * dst_wt);
+        res_h_l = (src_l * src_wt);
+        res_h_l += (dst_l * dst_wt);
+        SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+        PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+        dst_ptr += dst_stride;
+
+        UNPCK_UB_SH(src2, src_r, src_l);
+        UNPCK_UB_SH(dst2, dst_r, dst_l);
+        res_h_r = (src_r * src_wt);
+        res_h_r += (dst_r * dst_wt);
+        res_h_l = (src_l * src_wt);
+        res_h_l += (dst_l * dst_wt);
+        SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+        PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+        dst_ptr += dst_stride;
+
+        UNPCK_UB_SH(src3, src_r, src_l);
+        UNPCK_UB_SH(dst3, dst_r, dst_l);
+        res_h_r = (src_r * src_wt);
+        res_h_r += (dst_r * dst_wt);
+        res_h_l = (src_l * src_wt);
+        res_h_l += (dst_l * dst_wt);
+        SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+        PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+        dst_ptr += dst_stride;
+    }
+}
+
+void vp8_filter_by_weight16x16_msa(uint8_t *src_ptr, int32_t src_stride,
+                                   uint8_t *dst_ptr, int32_t dst_stride,
+                                   int32_t src_weight)
+{
+    filter_by_weight16x16_msa(src_ptr, src_stride, dst_ptr, dst_stride,
+                              src_weight);
+}
+
+void vp8_filter_by_weight8x8_msa(uint8_t *src_ptr, int32_t src_stride,
+                                 uint8_t *dst_ptr, int32_t dst_stride,
+                                 int32_t src_weight)
+{
+    filter_by_weight8x8_msa(src_ptr, src_stride, dst_ptr, dst_stride,
+                            src_weight);
+}
diff --git a/libvpx/libvpx/vp8/common/mips/msa/postproc_msa.c b/libvpx/libvpx/vp8/common/mips/msa/postproc_msa.c
new file mode 100644
index 0000000..23dcde2
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/msa/postproc_msa.c
@@ -0,0 +1,801 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include "./vp8_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+
+static const int16_t vp8_rv_msa[] =
+{
+    8, 5, 2, 2, 8, 12, 4, 9, 8, 3,
+    0, 3, 9, 0, 0, 0, 8, 3, 14, 4,
+    10, 1, 11, 14, 1, 14, 9, 6, 12, 11,
+    8, 6, 10, 0, 0, 8, 9, 0, 3, 14,
+    8, 11, 13, 4, 2, 9, 0, 3, 9, 6,
+    1, 2, 3, 14, 13, 1, 8, 2, 9, 7,
+    3, 3, 1, 13, 13, 6, 6, 5, 2, 7,
+    11, 9, 11, 8, 7, 3, 2, 0, 13, 13,
+    14, 4, 12, 5, 12, 10, 8, 10, 13, 10,
+    4, 14, 4, 10, 0, 8, 11, 1, 13, 7,
+    7, 14, 6, 14, 13, 2, 13, 5, 4, 4,
+    0, 10, 0, 5, 13, 2, 12, 7, 11, 13,
+    8, 0, 4, 10, 7, 2, 7, 2, 2, 5,
+    3, 4, 7, 3, 3, 14, 14, 5, 9, 13,
+    3, 14, 3, 6, 3, 0, 11, 8, 13, 1,
+    13, 1, 12, 0, 10, 9, 7, 6, 2, 8,
+    5, 2, 13, 7, 1, 13, 14, 7, 6, 7,
+    9, 6, 10, 11, 7, 8, 7, 5, 14, 8,
+    4, 4, 0, 8, 7, 10, 0, 8, 14, 11,
+    3, 12, 5, 7, 14, 3, 14, 5, 2, 6,
+    11, 12, 12, 8, 0, 11, 13, 1, 2, 0,
+    5, 10, 14, 7, 8, 0, 4, 11, 0, 8,
+    0, 3, 10, 5, 8, 0, 11, 6, 7, 8,
+    10, 7, 13, 9, 2, 5, 1, 5, 10, 2,
+    4, 3, 5, 6, 10, 8, 9, 4, 11, 14,
+    0, 10, 0, 5, 13, 2, 12, 7, 11, 13,
+    8, 0, 4, 10, 7, 2, 7, 2, 2, 5,
+    3, 4, 7, 3, 3, 14, 14, 5, 9, 13,
+    3, 14, 3, 6, 3, 0, 11, 8, 13, 1,
+    13, 1, 12, 0, 10, 9, 7, 6, 2, 8,
+    5, 2, 13, 7, 1, 13, 14, 7, 6, 7,
+    9, 6, 10, 11, 7, 8, 7, 5, 14, 8,
+    4, 4, 0, 8, 7, 10, 0, 8, 14, 11,
+    3, 12, 5, 7, 14, 3, 14, 5, 2, 6,
+    11, 12, 12, 8, 0, 11, 13, 1, 2, 0,
+    5, 10, 14, 7, 8, 0, 4, 11, 0, 8,
+    0, 3, 10, 5, 8, 0, 11, 6, 7, 8,
+    10, 7, 13, 9, 2, 5, 1, 5, 10, 2,
+    4, 3, 5, 6, 10, 8, 9, 4, 11, 14,
+    3, 8, 3, 7, 8, 5, 11, 4, 12, 3,
+    11, 9, 14, 8, 14, 13, 4, 3, 1, 2,
+    14, 6, 5, 4, 4, 11, 4, 6, 2, 1,
+    5, 8, 8, 12, 13, 5, 14, 10, 12, 13,
+    0, 9, 5, 5, 11, 10, 13, 9, 10, 13,
+};
+
+#define VP8_TRANSPOSE8x16_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7,  \
+                                out0, out1, out2, out3,                  \
+                                out4, out5, out6, out7,                  \
+                                out8, out9, out10, out11,                \
+                                out12, out13, out14, out15)              \
+{                                                                        \
+    v8i16 temp0, temp1, temp2, temp3, temp4;                             \
+    v8i16 temp5, temp6, temp7, temp8, temp9;                             \
+                                                                         \
+    ILVR_B4_SH(in1, in0, in3, in2, in5, in4, in7, in6,                   \
+               temp0, temp1, temp2, temp3);                              \
+    ILVR_H2_SH(temp1, temp0, temp3, temp2, temp4, temp5);                \
+    ILVRL_W2_SH(temp5, temp4, temp6, temp7);                             \
+    ILVL_H2_SH(temp1, temp0, temp3, temp2, temp4, temp5);                \
+    ILVRL_W2_SH(temp5, temp4, temp8, temp9);                             \
+    ILVL_B4_SH(in1, in0, in3, in2, in5, in4, in7, in6,                   \
+               temp0, temp1, temp2, temp3);                              \
+    ILVR_H2_SH(temp1, temp0, temp3, temp2, temp4, temp5);                \
+    ILVRL_W2_UB(temp5, temp4, out8, out10);                              \
+    ILVL_H2_SH(temp1, temp0, temp3, temp2, temp4, temp5);                \
+    ILVRL_W2_UB(temp5, temp4, out12, out14);                             \
+    out0 = (v16u8)temp6;                                                 \
+    out2 = (v16u8)temp7;                                                 \
+    out4 = (v16u8)temp8;                                                 \
+    out6 = (v16u8)temp9;                                                 \
+    out9 = (v16u8)__msa_ilvl_d((v2i64)out8, (v2i64)out8);                \
+    out11 = (v16u8)__msa_ilvl_d((v2i64)out10, (v2i64)out10);             \
+    out13 = (v16u8)__msa_ilvl_d((v2i64)out12, (v2i64)out12);             \
+    out15 = (v16u8)__msa_ilvl_d((v2i64)out14, (v2i64)out14);             \
+    out1 = (v16u8)__msa_ilvl_d((v2i64)out0, (v2i64)out0);                \
+    out3 = (v16u8)__msa_ilvl_d((v2i64)out2, (v2i64)out2);                \
+    out5 = (v16u8)__msa_ilvl_d((v2i64)out4, (v2i64)out4);                \
+    out7 = (v16u8)__msa_ilvl_d((v2i64)out6, (v2i64)out6);                \
+}
+
+#define VP8_AVER_IF_RETAIN(above2_in, above1_in, src_in,    \
+                           below1_in, below2_in, ref, out)  \
+{                                                           \
+    v16u8 temp0, temp1;                                     \
+                                                            \
+    temp1 = __msa_aver_u_b(above2_in, above1_in);           \
+    temp0 = __msa_aver_u_b(below2_in, below1_in);           \
+    temp1 = __msa_aver_u_b(temp1, temp0);                   \
+    out = __msa_aver_u_b(src_in, temp1);                    \
+    temp0 = __msa_asub_u_b(src_in, above2_in);              \
+    temp1 = __msa_asub_u_b(src_in, above1_in);              \
+    temp0 = (temp0 < ref);                                  \
+    temp1 = (temp1 < ref);                                  \
+    temp0 = temp0 & temp1;                                  \
+    temp1 = __msa_asub_u_b(src_in, below1_in);              \
+    temp1 = (temp1 < ref);                                  \
+    temp0 = temp0 & temp1;                                  \
+    temp1 = __msa_asub_u_b(src_in, below2_in);              \
+    temp1 = (temp1 < ref);                                  \
+    temp0 = temp0 & temp1;                                  \
+    out = __msa_bmz_v(out, src_in, temp0);                  \
+}
+
+#define TRANSPOSE12x16_B(in0, in1, in2, in3, in4, in5, in6, in7,        \
+                         in8, in9, in10, in11, in12, in13, in14, in15)  \
+{                                                                       \
+    v8i16 temp0, temp1, temp2, temp3, temp4;                            \
+    v8i16 temp5, temp6, temp7, temp8, temp9;                            \
+                                                                        \
+    ILVR_B2_SH(in1, in0, in3, in2, temp0, temp1);                       \
+    ILVRL_H2_SH(temp1, temp0, temp2, temp3);                            \
+    ILVR_B2_SH(in5, in4, in7, in6, temp0, temp1);                       \
+    ILVRL_H2_SH(temp1, temp0, temp4, temp5);                            \
+    ILVRL_W2_SH(temp4, temp2, temp0, temp1);                            \
+    ILVRL_W2_SH(temp5, temp3, temp2, temp3);                            \
+    ILVR_B2_SH(in9, in8, in11, in10, temp4, temp5);                     \
+    ILVR_B2_SH(in9, in8, in11, in10, temp4, temp5);                     \
+    ILVRL_H2_SH(temp5, temp4, temp6, temp7);                            \
+    ILVR_B2_SH(in13, in12, in15, in14, temp4, temp5);                   \
+    ILVRL_H2_SH(temp5, temp4, temp8, temp9);                            \
+    ILVRL_W2_SH(temp8, temp6, temp4, temp5);                            \
+    ILVRL_W2_SH(temp9, temp7, temp6, temp7);                            \
+    ILVL_B2_SH(in1, in0, in3, in2, temp8, temp9);                       \
+    ILVR_D2_UB(temp4, temp0, temp5, temp1, in0, in2);                   \
+    in1 = (v16u8)__msa_ilvl_d((v2i64)temp4, (v2i64)temp0);              \
+    in3 = (v16u8)__msa_ilvl_d((v2i64)temp5, (v2i64)temp1);              \
+    ILVL_B2_SH(in5, in4, in7, in6, temp0, temp1);                       \
+    ILVR_D2_UB(temp6, temp2, temp7, temp3, in4, in6);                   \
+    in5 = (v16u8)__msa_ilvl_d((v2i64)temp6, (v2i64)temp2);              \
+    in7 = (v16u8)__msa_ilvl_d((v2i64)temp7, (v2i64)temp3);              \
+    ILVL_B4_SH(in9, in8, in11, in10, in13, in12, in15, in14,            \
+               temp2, temp3, temp4, temp5);                             \
+    ILVR_H4_SH(temp9, temp8, temp1, temp0, temp3, temp2, temp5, temp4,  \
+               temp6, temp7, temp8, temp9);                             \
+    ILVR_W2_SH(temp7, temp6, temp9, temp8, temp0, temp1);               \
+    in8 = (v16u8)__msa_ilvr_d((v2i64)temp1, (v2i64)temp0);              \
+    in9 = (v16u8)__msa_ilvl_d((v2i64)temp1, (v2i64)temp0);              \
+    ILVL_W2_SH(temp7, temp6, temp9, temp8, temp2, temp3);               \
+    in10 = (v16u8)__msa_ilvr_d((v2i64)temp3, (v2i64)temp2);             \
+    in11 = (v16u8)__msa_ilvl_d((v2i64)temp3, (v2i64)temp2);             \
+}
+
+#define VP8_TRANSPOSE12x8_UB_UB(in0, in1, in2, in3, in4, in5,    \
+                                in6, in7, in8, in9, in10, in11)  \
+{                                                                \
+    v8i16 temp0, temp1, temp2, temp3;                            \
+    v8i16 temp4, temp5, temp6, temp7;                            \
+                                                                 \
+    ILVR_B2_SH(in1, in0, in3, in2, temp0, temp1);                \
+    ILVRL_H2_SH(temp1, temp0, temp2, temp3);                     \
+    ILVR_B2_SH(in5, in4, in7, in6, temp0, temp1);                \
+    ILVRL_H2_SH(temp1, temp0, temp4, temp5);                     \
+    ILVRL_W2_SH(temp4, temp2, temp0, temp1);                     \
+    ILVRL_W2_SH(temp5, temp3, temp2, temp3);                     \
+    ILVL_B2_SH(in1, in0, in3, in2, temp4, temp5);                \
+    temp4 = __msa_ilvr_h(temp5, temp4);                          \
+    ILVL_B2_SH(in5, in4, in7, in6, temp6, temp7);                \
+    temp5 = __msa_ilvr_h(temp7, temp6);                          \
+    ILVRL_W2_SH(temp5, temp4, temp6, temp7);                     \
+    in0 = (v16u8)temp0;                                          \
+    in2 = (v16u8)temp1;                                          \
+    in4 = (v16u8)temp2;                                          \
+    in6 = (v16u8)temp3;                                          \
+    in8 = (v16u8)temp6;                                          \
+    in10 = (v16u8)temp7;                                         \
+    in1 = (v16u8)__msa_ilvl_d((v2i64)temp0, (v2i64)temp0);       \
+    in3 = (v16u8)__msa_ilvl_d((v2i64)temp1, (v2i64)temp1);       \
+    in5 = (v16u8)__msa_ilvl_d((v2i64)temp2, (v2i64)temp2);       \
+    in7 = (v16u8)__msa_ilvl_d((v2i64)temp3, (v2i64)temp3);       \
+    in9 = (v16u8)__msa_ilvl_d((v2i64)temp6, (v2i64)temp6);       \
+    in11 = (v16u8)__msa_ilvl_d((v2i64)temp7, (v2i64)temp7);      \
+}
+
+static void postproc_down_across_chroma_msa(uint8_t *src_ptr, uint8_t *dst_ptr,
+                                            int32_t src_stride,
+                                            int32_t dst_stride,
+                                            int32_t cols, uint8_t *f)
+{
+    uint8_t *p_src = src_ptr;
+    uint8_t *p_dst = dst_ptr;
+    uint8_t *f_orig = f;
+    uint8_t *p_dst_st = dst_ptr;
+    uint16_t col;
+    uint64_t out0, out1, out2, out3;
+    v16u8 above2, above1, below2, below1, src, ref, ref_temp;
+    v16u8 inter0, inter1, inter2, inter3, inter4, inter5;
+    v16u8 inter6, inter7, inter8, inter9, inter10, inter11;
+
+    for (col = (cols / 16); col--;)
+    {
+        ref = LD_UB(f);
+        LD_UB2(p_src - 2 * src_stride, src_stride, above2, above1);
+        src = LD_UB(p_src);
+        LD_UB2(p_src + 1 * src_stride, src_stride, below1, below2);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2, ref, inter0);
+        above2 = LD_UB(p_src + 3 * src_stride);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2, ref, inter1);
+        above1 = LD_UB(p_src + 4 * src_stride);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1, ref, inter2);
+        src = LD_UB(p_src + 5 * src_stride);
+        VP8_AVER_IF_RETAIN(below1, below2, above2, above1, src, ref, inter3);
+        below1 = LD_UB(p_src + 6 * src_stride);
+        VP8_AVER_IF_RETAIN(below2, above2, above1, src, below1, ref, inter4);
+        below2 = LD_UB(p_src + 7 * src_stride);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2, ref, inter5);
+        above2 = LD_UB(p_src + 8 * src_stride);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2, ref, inter6);
+        above1 = LD_UB(p_src + 9 * src_stride);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1, ref, inter7);
+        ST_UB8(inter0, inter1, inter2, inter3, inter4, inter5, inter6, inter7,
+               p_dst, dst_stride);
+
+        p_dst += 16;
+        p_src += 16;
+        f += 16;
+    }
+
+    if (0 != (cols / 16))
+    {
+        ref = LD_UB(f);
+        LD_UB2(p_src - 2 * src_stride, src_stride, above2, above1);
+        src = LD_UB(p_src);
+        LD_UB2(p_src + 1 * src_stride, src_stride, below1, below2);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2, ref, inter0);
+        above2 = LD_UB(p_src + 3 * src_stride);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2, ref, inter1);
+        above1 = LD_UB(p_src + 4 * src_stride);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1, ref, inter2);
+        src = LD_UB(p_src + 5 * src_stride);
+        VP8_AVER_IF_RETAIN(below1, below2, above2, above1, src, ref, inter3);
+        below1 = LD_UB(p_src + 6 * src_stride);
+        VP8_AVER_IF_RETAIN(below2, above2, above1, src, below1, ref, inter4);
+        below2 = LD_UB(p_src + 7 * src_stride);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2, ref, inter5);
+        above2 = LD_UB(p_src + 8 * src_stride);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2, ref, inter6);
+        above1 = LD_UB(p_src + 9 * src_stride);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1, ref, inter7);
+        out0 = __msa_copy_u_d((v2i64)inter0, 0);
+        out1 = __msa_copy_u_d((v2i64)inter1, 0);
+        out2 = __msa_copy_u_d((v2i64)inter2, 0);
+        out3 = __msa_copy_u_d((v2i64)inter3, 0);
+        SD4(out0, out1, out2, out3, p_dst, dst_stride);
+
+        out0 = __msa_copy_u_d((v2i64)inter4, 0);
+        out1 = __msa_copy_u_d((v2i64)inter5, 0);
+        out2 = __msa_copy_u_d((v2i64)inter6, 0);
+        out3 = __msa_copy_u_d((v2i64)inter7, 0);
+        SD4(out0, out1, out2, out3, p_dst + 4 * dst_stride, dst_stride);
+    }
+
+    f = f_orig;
+    p_dst = dst_ptr - 2;
+    LD_UB8(p_dst, dst_stride,
+           inter0, inter1, inter2, inter3, inter4, inter5, inter6, inter7);
+
+    for (col = 0; col < (cols / 8); ++col)
+    {
+        ref = LD_UB(f);
+        f += 8;
+        VP8_TRANSPOSE12x8_UB_UB(inter0, inter1, inter2, inter3,
+                                inter4, inter5, inter6, inter7,
+                                inter8, inter9, inter10, inter11);
+        if (0 == col)
+        {
+            above2 = inter2;
+            above1 = inter2;
+        }
+        else
+        {
+            above2 = inter0;
+            above1 = inter1;
+        }
+        src = inter2;
+        below1 = inter3;
+        below2 = inter4;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 0);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2,
+                           ref_temp, inter2);
+        above2 = inter5;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 1);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2,
+                           ref_temp, inter3);
+        above1 = inter6;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 2);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1,
+                           ref_temp, inter4);
+        src = inter7;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 3);
+        VP8_AVER_IF_RETAIN(below1, below2, above2, above1, src,
+                           ref_temp, inter5);
+        below1 = inter8;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 4);
+        VP8_AVER_IF_RETAIN(below2, above2, above1, src, below1,
+                           ref_temp, inter6);
+        below2 = inter9;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 5);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2,
+                           ref_temp, inter7);
+        if (col == (cols / 8 - 1))
+        {
+            above2 = inter9;
+        }
+        else
+        {
+            above2 = inter10;
+        }
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 6);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2,
+                           ref_temp, inter8);
+        if (col == (cols / 8 - 1))
+        {
+            above1 = inter9;
+        }
+        else
+        {
+            above1 = inter11;
+        }
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 7);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1,
+                           ref_temp, inter9);
+        TRANSPOSE8x8_UB_UB(inter2, inter3, inter4, inter5, inter6, inter7,
+                           inter8, inter9, inter2, inter3, inter4, inter5,
+                           inter6, inter7, inter8, inter9);
+        p_dst += 8;
+        LD_UB2(p_dst, dst_stride, inter0, inter1);
+        ST8x1_UB(inter2, p_dst_st);
+        ST8x1_UB(inter3, (p_dst_st + 1 * dst_stride));
+        LD_UB2(p_dst + 2 * dst_stride, dst_stride, inter2, inter3);
+        ST8x1_UB(inter4, (p_dst_st + 2 * dst_stride));
+        ST8x1_UB(inter5, (p_dst_st + 3 * dst_stride));
+        LD_UB2(p_dst + 4 * dst_stride, dst_stride, inter4, inter5);
+        ST8x1_UB(inter6, (p_dst_st + 4 * dst_stride));
+        ST8x1_UB(inter7, (p_dst_st + 5 * dst_stride));
+        LD_UB2(p_dst + 6 * dst_stride, dst_stride, inter6, inter7);
+        ST8x1_UB(inter8, (p_dst_st + 6 * dst_stride));
+        ST8x1_UB(inter9, (p_dst_st + 7 * dst_stride));
+        p_dst_st += 8;
+    }
+}
+
+static void postproc_down_across_luma_msa(uint8_t *src_ptr, uint8_t *dst_ptr,
+                                          int32_t src_stride,
+                                          int32_t dst_stride,
+                                          int32_t cols, uint8_t *f)
+{
+    uint8_t *p_src = src_ptr;
+    uint8_t *p_dst = dst_ptr;
+    uint8_t *p_dst_st = dst_ptr;
+    uint8_t *f_orig = f;
+    uint16_t col;
+    v16u8 above2, above1, below2, below1;
+    v16u8 src, ref, ref_temp;
+    v16u8 inter0, inter1, inter2, inter3, inter4, inter5, inter6;
+    v16u8 inter7, inter8, inter9, inter10, inter11;
+    v16u8 inter12, inter13, inter14, inter15;
+
+    for (col = (cols / 16); col--;)
+    {
+        ref = LD_UB(f);
+        LD_UB2(p_src - 2 * src_stride, src_stride, above2, above1);
+        src = LD_UB(p_src);
+        LD_UB2(p_src + 1 * src_stride, src_stride, below1, below2);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2, ref, inter0);
+        above2 = LD_UB(p_src + 3 * src_stride);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2, ref, inter1);
+        above1 = LD_UB(p_src + 4 * src_stride);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1, ref, inter2);
+        src = LD_UB(p_src + 5 * src_stride);
+        VP8_AVER_IF_RETAIN(below1, below2, above2, above1, src, ref, inter3);
+        below1 = LD_UB(p_src + 6 * src_stride);
+        VP8_AVER_IF_RETAIN(below2, above2, above1, src, below1, ref, inter4);
+        below2 = LD_UB(p_src + 7 * src_stride);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2, ref, inter5);
+        above2 = LD_UB(p_src + 8 * src_stride);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2, ref, inter6);
+        above1 = LD_UB(p_src + 9 * src_stride);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1, ref, inter7);
+        src = LD_UB(p_src + 10 * src_stride);
+        VP8_AVER_IF_RETAIN(below1, below2, above2, above1, src, ref, inter8);
+        below1 = LD_UB(p_src + 11 * src_stride);
+        VP8_AVER_IF_RETAIN(below2, above2, above1, src, below1, ref, inter9);
+        below2 = LD_UB(p_src + 12 * src_stride);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2, ref, inter10);
+        above2 = LD_UB(p_src + 13 * src_stride);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2, ref, inter11);
+        above1 = LD_UB(p_src + 14 * src_stride);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1, ref, inter12);
+        src = LD_UB(p_src + 15 * src_stride);
+        VP8_AVER_IF_RETAIN(below1, below2, above2, above1, src, ref, inter13);
+        below1 = LD_UB(p_src + 16 * src_stride);
+        VP8_AVER_IF_RETAIN(below2, above2, above1, src, below1, ref, inter14);
+        below2 = LD_UB(p_src + 17 * src_stride);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2, ref, inter15);
+        ST_UB8(inter0, inter1, inter2, inter3, inter4, inter5, inter6, inter7,
+               p_dst, dst_stride);
+        ST_UB8(inter8, inter9, inter10, inter11, inter12, inter13,
+               inter14, inter15, p_dst + 8 * dst_stride, dst_stride);
+        p_src += 16;
+        p_dst += 16;
+        f += 16;
+    }
+
+    f = f_orig;
+    p_dst = dst_ptr - 2;
+    LD_UB8(p_dst, dst_stride,
+           inter0, inter1, inter2, inter3, inter4, inter5, inter6, inter7);
+    LD_UB8(p_dst + 8 * dst_stride, dst_stride,
+           inter8, inter9, inter10, inter11, inter12, inter13,
+           inter14, inter15);
+
+    for (col = 0; col < cols / 8; ++col)
+    {
+        ref = LD_UB(f);
+        f += 8;
+        TRANSPOSE12x16_B(inter0, inter1, inter2, inter3, inter4, inter5,
+                         inter6, inter7, inter8, inter9, inter10, inter11,
+                         inter12, inter13, inter14, inter15);
+        if (0 == col)
+        {
+            above2 = inter2;
+            above1 = inter2;
+        }
+        else
+        {
+            above2 = inter0;
+            above1 = inter1;
+        }
+
+        src = inter2;
+        below1 = inter3;
+        below2 = inter4;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 0);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2,
+                           ref_temp, inter2);
+        above2 = inter5;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 1);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2,
+                           ref_temp, inter3);
+        above1 = inter6;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 2);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1,
+                           ref_temp, inter4);
+        src = inter7;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 3);
+        VP8_AVER_IF_RETAIN(below1, below2, above2, above1, src,
+                           ref_temp, inter5);
+        below1 = inter8;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 4);
+        VP8_AVER_IF_RETAIN(below2, above2, above1, src, below1,
+                           ref_temp, inter6);
+        below2 = inter9;
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 5);
+        VP8_AVER_IF_RETAIN(above2, above1, src, below1, below2,
+                           ref_temp, inter7);
+        if (col == (cols / 8 - 1))
+        {
+            above2 = inter9;
+        }
+        else
+        {
+            above2 = inter10;
+        }
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 6);
+        VP8_AVER_IF_RETAIN(above1, src, below1, below2, above2,
+                           ref_temp, inter8);
+        if (col == (cols / 8 - 1))
+        {
+            above1 = inter9;
+        }
+        else
+        {
+            above1 = inter11;
+        }
+        ref_temp = (v16u8)__msa_splati_b((v16i8)ref, 7);
+        VP8_AVER_IF_RETAIN(src, below1, below2, above2, above1,
+                           ref_temp, inter9);
+        VP8_TRANSPOSE8x16_UB_UB(inter2, inter3, inter4, inter5,
+                                inter6, inter7, inter8, inter9,
+                                inter2, inter3, inter4, inter5,
+                                inter6, inter7, inter8, inter9,
+                                inter10, inter11, inter12, inter13,
+                                inter14, inter15, above2, above1);
+
+        p_dst += 8;
+        LD_UB2(p_dst, dst_stride, inter0, inter1);
+        ST8x1_UB(inter2, p_dst_st);
+        ST8x1_UB(inter3, (p_dst_st + 1 * dst_stride));
+        LD_UB2(p_dst + 2 * dst_stride, dst_stride, inter2, inter3);
+        ST8x1_UB(inter4, (p_dst_st + 2 * dst_stride));
+        ST8x1_UB(inter5, (p_dst_st + 3 * dst_stride));
+        LD_UB2(p_dst + 4 * dst_stride, dst_stride, inter4, inter5);
+        ST8x1_UB(inter6, (p_dst_st + 4 * dst_stride));
+        ST8x1_UB(inter7, (p_dst_st + 5 * dst_stride));
+        LD_UB2(p_dst + 6 * dst_stride, dst_stride, inter6, inter7);
+        ST8x1_UB(inter8, (p_dst_st + 6 * dst_stride));
+        ST8x1_UB(inter9, (p_dst_st + 7 * dst_stride));
+        LD_UB2(p_dst + 8 * dst_stride, dst_stride, inter8, inter9);
+        ST8x1_UB(inter10, (p_dst_st + 8 * dst_stride));
+        ST8x1_UB(inter11, (p_dst_st + 9 * dst_stride));
+        LD_UB2(p_dst + 10 * dst_stride, dst_stride, inter10, inter11);
+        ST8x1_UB(inter12, (p_dst_st + 10 * dst_stride));
+        ST8x1_UB(inter13, (p_dst_st + 11 * dst_stride));
+        LD_UB2(p_dst + 12 * dst_stride, dst_stride, inter12, inter13);
+        ST8x1_UB(inter14, (p_dst_st + 12 * dst_stride));
+        ST8x1_UB(inter15, (p_dst_st + 13 * dst_stride));
+        LD_UB2(p_dst + 14 * dst_stride, dst_stride, inter14, inter15);
+        ST8x1_UB(above2, (p_dst_st + 14 * dst_stride));
+        ST8x1_UB(above1, (p_dst_st + 15 * dst_stride));
+        p_dst_st += 8;
+    }
+}
+
+void vp8_post_proc_down_and_across_mb_row_msa(uint8_t *src, uint8_t *dst,
+                                              int32_t src_stride,
+                                              int32_t dst_stride,
+                                              int32_t cols, uint8_t *f,
+                                              int32_t size)
+{
+    if (8 == size)
+    {
+        postproc_down_across_chroma_msa(src, dst, src_stride, dst_stride,
+                                        cols, f);
+    }
+    else if (16 == size)
+    {
+        postproc_down_across_luma_msa(src, dst, src_stride, dst_stride,
+                                      cols, f);
+    }
+}
+
+void vp8_mbpost_proc_across_ip_msa(uint8_t *src_ptr, int32_t pitch,
+                                   int32_t rows, int32_t cols, int32_t flimit)
+{
+    int32_t row, col, cnt;
+    uint8_t *src_dup = src_ptr;
+    v16u8 src0, src, tmp_orig;
+    v16u8 tmp = { 0 };
+    v16i8 zero = { 0 };
+    v8u16 sum_h, src_r_h, src_l_h;
+    v4u32 src_r_w, src_l_w;
+    v4i32 flimit_vec;
+
+    flimit_vec = __msa_fill_w(flimit);
+    for (row = rows; row--;)
+    {
+        int32_t sum_sq = 0;
+        int32_t sum = 0;
+        src0 = (v16u8)__msa_fill_b(src_dup[0]);
+        ST8x1_UB(src0, (src_dup - 8));
+
+        src0 = (v16u8)__msa_fill_b(src_dup[cols - 1]);
+        ST_UB(src0, src_dup + cols);
+        src_dup[cols + 16] = src_dup[cols - 1];
+        tmp_orig = (v16u8)__msa_ldi_b(0);
+        tmp_orig[15] = tmp[15];
+        src = LD_UB(src_dup - 8);
+        src[15] = 0;
+        ILVRL_B2_UH(zero, src, src_r_h, src_l_h);
+        src_r_w = __msa_dotp_u_w(src_r_h, src_r_h);
+        src_l_w = __msa_dotp_u_w(src_l_h, src_l_h);
+        sum_sq = HADD_SW_S32(src_r_w);
+        sum_sq += HADD_SW_S32(src_l_w);
+        sum_h = __msa_hadd_u_h(src, src);
+        sum = HADD_UH_U32(sum_h);
+        {
+            v16u8 src7, src8, src_r, src_l;
+            v16i8 mask;
+            v8u16 add_r, add_l;
+            v8i16 sub_r, sub_l, sum_r, sum_l, mask0, mask1;
+            v4i32 sum_sq0, sum_sq1, sum_sq2, sum_sq3;
+            v4i32 sub0, sub1, sub2, sub3;
+            v4i32 sum0_w, sum1_w, sum2_w, sum3_w;
+            v4i32 mul0, mul1, mul2, mul3;
+            v4i32 total0, total1, total2, total3;
+            v8i16 const8 = __msa_fill_h(8);
+
+            src7 = LD_UB(src_dup + 7);
+            src8 = LD_UB(src_dup - 8);
+            for (col = 0; col < (cols >> 4); ++col)
+            {
+                ILVRL_B2_UB(src7, src8, src_r, src_l);
+                HSUB_UB2_SH(src_r, src_l, sub_r, sub_l);
+
+                sum_r[0] = sum + sub_r[0];
+                for (cnt = 0; cnt < 7; ++cnt)
+                {
+                    sum_r[cnt + 1] = sum_r[cnt] + sub_r[cnt + 1];
+                }
+                sum_l[0] = sum_r[7] + sub_l[0];
+                for (cnt = 0; cnt < 7; ++cnt)
+                {
+                    sum_l[cnt + 1] = sum_l[cnt] + sub_l[cnt + 1];
+                }
+                sum = sum_l[7];
+                src = LD_UB(src_dup + 16 * col);
+                ILVRL_B2_UH(zero, src, src_r_h, src_l_h);
+                src7 = (v16u8)((const8 + sum_r + (v8i16)src_r_h) >> 4);
+                src8 = (v16u8)((const8 + sum_l + (v8i16)src_l_h) >> 4);
+                tmp = (v16u8)__msa_pckev_b((v16i8)src8, (v16i8)src7);
+
+                HADD_UB2_UH(src_r, src_l, add_r, add_l);
+                UNPCK_SH_SW(sub_r, sub0, sub1);
+                UNPCK_SH_SW(sub_l, sub2, sub3);
+                ILVR_H2_SW(zero, add_r, zero, add_l, sum0_w, sum2_w);
+                ILVL_H2_SW(zero, add_r, zero, add_l, sum1_w, sum3_w);
+                MUL4(sum0_w, sub0, sum1_w, sub1, sum2_w, sub2, sum3_w, sub3,
+                     mul0, mul1, mul2, mul3);
+                sum_sq0[0] = sum_sq + mul0[0];
+                for (cnt = 0; cnt < 3; ++cnt)
+                {
+                    sum_sq0[cnt + 1] = sum_sq0[cnt] + mul0[cnt + 1];
+                }
+                sum_sq1[0] = sum_sq0[3] + mul1[0];
+                for (cnt = 0; cnt < 3; ++cnt)
+                {
+                    sum_sq1[cnt + 1] = sum_sq1[cnt] + mul1[cnt + 1];
+                }
+                sum_sq2[0] = sum_sq1[3] + mul2[0];
+                for (cnt = 0; cnt < 3; ++cnt)
+                {
+                    sum_sq2[cnt + 1] = sum_sq2[cnt] + mul2[cnt + 1];
+                }
+                sum_sq3[0] = sum_sq2[3] + mul3[0];
+                for (cnt = 0; cnt < 3; ++cnt)
+                {
+                    sum_sq3[cnt + 1] = sum_sq3[cnt] + mul3[cnt + 1];
+                }
+                sum_sq = sum_sq3[3];
+
+                UNPCK_SH_SW(sum_r, sum0_w, sum1_w);
+                UNPCK_SH_SW(sum_l, sum2_w, sum3_w);
+                total0 = sum_sq0 * __msa_ldi_w(15);
+                total0 -= sum0_w * sum0_w;
+                total1 = sum_sq1 * __msa_ldi_w(15);
+                total1 -= sum1_w * sum1_w;
+                total2 = sum_sq2 * __msa_ldi_w(15);
+                total2 -= sum2_w * sum2_w;
+                total3 = sum_sq3 * __msa_ldi_w(15);
+                total3 -= sum3_w * sum3_w;
+                total0 = (total0 < flimit_vec);
+                total1 = (total1 < flimit_vec);
+                total2 = (total2 < flimit_vec);
+                total3 = (total3 < flimit_vec);
+                PCKEV_H2_SH(total1, total0, total3, total2, mask0, mask1);
+                mask = __msa_pckev_b((v16i8)mask1, (v16i8)mask0);
+                tmp = __msa_bmz_v(tmp, src, (v16u8)mask);
+
+                if (col == 0)
+                {
+                    uint64_t src_d;
+
+                    src_d = __msa_copy_u_d((v2i64)tmp_orig, 1);
+                    SD(src_d, (src_dup - 8));
+                }
+
+                src7 = LD_UB(src_dup + 16 * (col + 1) + 7);
+                src8 = LD_UB(src_dup + 16 * (col + 1) - 8);
+                ST_UB(tmp, (src_dup + (16 * col)));
+            }
+
+            src_dup += pitch;
+        }
+    }
+}
+
+void vp8_mbpost_proc_down_msa(uint8_t *dst_ptr, int32_t pitch, int32_t rows,
+                              int32_t cols, int32_t flimit)
+{
+    int32_t row, col, cnt, i;
+    const int16_t *rv3 = &vp8_rv_msa[63 & rand()];
+    v4i32 flimit_vec;
+    v16u8 dst7, dst8, dst_r_b, dst_l_b;
+    v16i8 mask;
+    v8u16 add_r, add_l;
+    v8i16 dst_r_h, dst_l_h, sub_r, sub_l, mask0, mask1;
+    v4i32 sub0, sub1, sub2, sub3, total0, total1, total2, total3;
+
+    flimit_vec = __msa_fill_w(flimit);
+
+    for (col = 0; col < (cols >> 4); ++col)
+    {
+        uint8_t *dst_tmp = &dst_ptr[col << 4];
+        v16u8 dst;
+        v16i8 zero = { 0 };
+        v16u8 tmp[16];
+        v8i16 mult0, mult1, rv2_0, rv2_1;
+        v8i16 sum0_h = { 0 };
+        v8i16 sum1_h = { 0 };
+        v4i32 mul0 = { 0 };
+        v4i32 mul1 = { 0 };
+        v4i32 mul2 = { 0 };
+        v4i32 mul3 = { 0 };
+        v4i32 sum0_w, sum1_w, sum2_w, sum3_w;
+        v4i32 add0, add1, add2, add3;
+        const int16_t *rv2[16];
+
+        dst = LD_UB(dst_tmp);
+        for (cnt = (col << 4), i = 0; i < 16; ++cnt)
+        {
+            rv2[i] = rv3 + ((cnt * 17) & 127);
+            ++i;
+        }
+        for (cnt = -8; cnt < 0; ++cnt)
+        {
+            ST_UB(dst, dst_tmp + cnt * pitch);
+        }
+
+        dst = LD_UB((dst_tmp + (rows - 1) * pitch));
+        for (cnt = rows; cnt < rows + 17; ++cnt)
+        {
+            ST_UB(dst, dst_tmp + cnt * pitch);
+        }
+        for (cnt = -8; cnt <= 6; ++cnt)
+        {
+            dst = LD_UB(dst_tmp + (cnt * pitch));
+            UNPCK_UB_SH(dst, dst_r_h, dst_l_h);
+            MUL2(dst_r_h, dst_r_h, dst_l_h, dst_l_h, mult0, mult1);
+            mul0 += (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)mult0);
+            mul1 += (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)mult0);
+            mul2 += (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)mult1);
+            mul3 += (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)mult1);
+            ADD2(sum0_h, dst_r_h, sum1_h, dst_l_h, sum0_h, sum1_h);
+        }
+
+        for (row = 0; row < (rows + 8); ++row)
+        {
+            for (i = 0; i < 8; ++i)
+            {
+                rv2_0[i] = *(rv2[i] + (row & 127));
+                rv2_1[i] = *(rv2[i + 8] + (row & 127));
+            }
+            dst7 = LD_UB(dst_tmp + (7 * pitch));
+            dst8 = LD_UB(dst_tmp - (8 * pitch));
+            ILVRL_B2_UB(dst7, dst8, dst_r_b, dst_l_b);
+
+            HSUB_UB2_SH(dst_r_b, dst_l_b, sub_r, sub_l);
+            UNPCK_SH_SW(sub_r, sub0, sub1);
+            UNPCK_SH_SW(sub_l, sub2, sub3);
+            sum0_h += sub_r;
+            sum1_h += sub_l;
+
+            HADD_UB2_UH(dst_r_b, dst_l_b, add_r, add_l);
+
+            ILVRL_H2_SW(zero, add_r, add0, add1);
+            ILVRL_H2_SW(zero, add_l, add2, add3);
+            mul0 += add0 * sub0;
+            mul1 += add1 * sub1;
+            mul2 += add2 * sub2;
+            mul3 += add3 * sub3;
+            dst = LD_UB(dst_tmp);
+            ILVRL_B2_SH(zero, dst, dst_r_h, dst_l_h);
+            dst7 = (v16u8)((rv2_0 + sum0_h + dst_r_h) >> 4);
+            dst8 = (v16u8)((rv2_1 + sum1_h + dst_l_h) >> 4);
+            tmp[row & 15] = (v16u8)__msa_pckev_b((v16i8)dst8, (v16i8)dst7);
+
+            UNPCK_SH_SW(sum0_h, sum0_w, sum1_w);
+            UNPCK_SH_SW(sum1_h, sum2_w, sum3_w);
+            total0 = mul0 * __msa_ldi_w(15);
+            total0 -= sum0_w * sum0_w;
+            total1 = mul1 * __msa_ldi_w(15);
+            total1 -= sum1_w * sum1_w;
+            total2 = mul2 * __msa_ldi_w(15);
+            total2 -= sum2_w * sum2_w;
+            total3 = mul3 * __msa_ldi_w(15);
+            total3 -= sum3_w * sum3_w;
+            total0 = (total0 < flimit_vec);
+            total1 = (total1 < flimit_vec);
+            total2 = (total2 < flimit_vec);
+            total3 = (total3 < flimit_vec);
+            PCKEV_H2_SH(total1, total0, total3, total2, mask0, mask1);
+            mask = __msa_pckev_b((v16i8)mask1, (v16i8)mask0);
+            tmp[row & 15] = __msa_bmz_v(tmp[row & 15], dst, (v16u8)mask);
+
+            if (row >= 8)
+            {
+                ST_UB(tmp[(row - 8) & 15], (dst_tmp - 8 * pitch));
+            }
+
+            dst_tmp += pitch;
+        }
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/mips/msa/sixtap_filter_msa.c b/libvpx/libvpx/vp8/common/mips/msa/sixtap_filter_msa.c
new file mode 100644
index 0000000..fb60fc1
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/msa/sixtap_filter_msa.c
@@ -0,0 +1,1850 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vp8/common/filter.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+
+DECLARE_ALIGNED(16, static const int8_t, vp8_subpel_filters_msa[7][8]) =
+{
+    { 0, -6, 123, 12, -1, 0, 0, 0 },
+    { 2, -11, 108, 36, -8, 1, 0, 0 },  /* New 1/4 pel 6 tap filter */
+    { 0, -9, 93, 50, -6, 0, 0, 0 },
+    { 3, -16, 77, 77, -16, 3, 0, 0 },  /* New 1/2 pel 6 tap filter */
+    { 0, -6, 50, 93, -9, 0, 0, 0 },
+    { 1, -8, 36, 108, -11, 2, 0, 0 },  /* New 1/4 pel 6 tap filter */
+    { 0, -1, 12, 123, -6, 0, 0, 0 },
+};
+
+static const uint8_t vp8_mc_filt_mask_arr[16 * 3] =
+{
+    /* 8 width cases */
+    0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+    /* 4 width cases */
+    0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20,
+    /* 4 width cases */
+    8, 9, 9, 10, 10, 11, 11, 12, 24, 25, 25, 26, 26, 27, 27, 28
+};
+
+#define HORIZ_6TAP_FILT(src0, src1, mask0, mask1, mask2,                 \
+                        filt_h0, filt_h1, filt_h2)                       \
+({                                                                       \
+    v16i8 vec0_m, vec1_m, vec2_m;                                        \
+    v8i16 hz_out_m;                                                      \
+                                                                         \
+    VSHF_B3_SB(src0, src1, src0, src1, src0, src1, mask0, mask1, mask2,  \
+               vec0_m, vec1_m, vec2_m);                                  \
+    hz_out_m = DPADD_SH3_SH(vec0_m, vec1_m, vec2_m,                      \
+                            filt_h0, filt_h1, filt_h2);                  \
+                                                                         \
+    hz_out_m = __msa_srari_h(hz_out_m, VP8_FILTER_SHIFT);                \
+    hz_out_m = __msa_sat_s_h(hz_out_m, 7);                               \
+                                                                         \
+    hz_out_m;                                                            \
+})
+
+#define HORIZ_6TAP_4WID_4VECS_FILT(src0, src1, src2, src3,             \
+                                   mask0, mask1, mask2,                \
+                                   filt0, filt1, filt2,                \
+                                   out0, out1)                         \
+{                                                                      \
+    v16i8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m;              \
+                                                                       \
+    VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0_m, vec1_m);  \
+    DOTP_SB2_SH(vec0_m, vec1_m, filt0, filt0, out0, out1);             \
+    VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2_m, vec3_m);  \
+    DPADD_SB2_SH(vec2_m, vec3_m, filt1, filt1, out0, out1);            \
+    VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4_m, vec5_m);  \
+    DPADD_SB2_SH(vec4_m, vec5_m, filt2, filt2, out0, out1);            \
+}
+
+#define HORIZ_6TAP_8WID_4VECS_FILT(src0, src1, src2, src3,                    \
+                                   mask0, mask1, mask2,                       \
+                                   filt0, filt1, filt2,                       \
+                                   out0, out1, out2, out3)                    \
+{                                                                             \
+    v16i8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;     \
+                                                                              \
+    VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0_m, vec1_m);         \
+    VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2_m, vec3_m);         \
+    DOTP_SB4_SH(vec0_m, vec1_m, vec2_m, vec3_m, filt0, filt0, filt0, filt0,   \
+                out0, out1, out2, out3);                                      \
+    VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec0_m, vec1_m);         \
+    VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec2_m, vec3_m);         \
+    VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec4_m, vec5_m);         \
+    VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec6_m, vec7_m);         \
+    DPADD_SB4_SH(vec0_m, vec1_m, vec2_m, vec3_m, filt1, filt1, filt1, filt1,  \
+                 out0, out1, out2, out3);                                     \
+    DPADD_SB4_SH(vec4_m, vec5_m, vec6_m, vec7_m, filt2, filt2, filt2, filt2,  \
+                 out0, out1, out2, out3);                                     \
+}
+
+#define FILT_4TAP_DPADD_S_H(vec0, vec1, filt0, filt1)         \
+({                                                            \
+    v8i16 tmp0;                                               \
+                                                              \
+    tmp0 = __msa_dotp_s_h((v16i8)vec0, (v16i8)filt0);         \
+    tmp0 = __msa_dpadd_s_h(tmp0, (v16i8)vec1, (v16i8)filt1);  \
+                                                              \
+    tmp0;                                                     \
+})
+
+#define HORIZ_4TAP_FILT(src0, src1, mask0, mask1, filt_h0, filt_h1)    \
+({                                                                     \
+    v16i8 vec0_m, vec1_m;                                              \
+    v8i16 hz_out_m;                                                    \
+                                                                       \
+    VSHF_B2_SB(src0, src1, src0, src1, mask0, mask1, vec0_m, vec1_m);  \
+    hz_out_m = FILT_4TAP_DPADD_S_H(vec0_m, vec1_m, filt_h0, filt_h1);  \
+                                                                       \
+    hz_out_m = __msa_srari_h(hz_out_m, VP8_FILTER_SHIFT);              \
+    hz_out_m = __msa_sat_s_h(hz_out_m, 7);                             \
+                                                                       \
+    hz_out_m;                                                          \
+})
+
+#define HORIZ_4TAP_4WID_4VECS_FILT(src0, src1, src2, src3,             \
+                                   mask0, mask1, filt0, filt1,         \
+                                   out0, out1)                         \
+{                                                                      \
+    v16i8 vec0_m, vec1_m, vec2_m, vec3_m;                              \
+                                                                       \
+    VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0_m, vec1_m);  \
+    DOTP_SB2_SH(vec0_m, vec1_m, filt0, filt0, out0, out1);             \
+    VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2_m, vec3_m);  \
+    DPADD_SB2_SH(vec2_m, vec3_m, filt1, filt1, out0, out1);            \
+}
+
+#define HORIZ_4TAP_8WID_4VECS_FILT(src0, src1, src2, src3,                    \
+                                   mask0, mask1, filt0, filt1,                \
+                                   out0, out1, out2, out3)                    \
+{                                                                             \
+    v16i8 vec0_m, vec1_m, vec2_m, vec3_m;                                     \
+                                                                              \
+    VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0_m, vec1_m);         \
+    VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2_m, vec3_m);         \
+    DOTP_SB4_SH(vec0_m, vec1_m, vec2_m, vec3_m, filt0, filt0, filt0, filt0,   \
+                out0, out1, out2, out3);                                      \
+    VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec0_m, vec1_m);         \
+    VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec2_m, vec3_m);         \
+    DPADD_SB4_SH(vec0_m, vec1_m, vec2_m, vec3_m, filt1, filt1, filt1, filt1,  \
+                 out0, out1, out2, out3);                                     \
+}
+
+static void common_hz_6t_4x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter)
+{
+    v16i8 src0, src1, src2, src3, filt0, filt1, filt2;
+    v16u8 mask0, mask1, mask2, out;
+    v8i16 filt, out0, out1;
+
+    mask0 = LD_UB(&vp8_mc_filt_mask_arr[16]);
+    src -= 2;
+
+    filt = LD_SH(filter);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt0, filt1, filt2);
+
+    mask1 = mask0 + 2;
+    mask2 = mask0 + 4;
+
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_6TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               filt0, filt1, filt2, out0, out1);
+    SRARI_H2_SH(out0, out1, VP8_FILTER_SHIFT);
+    SAT_SH2_SH(out0, out1, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_hz_6t_4x8_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter)
+{
+    v16i8 src0, src1, src2, src3, filt0, filt1, filt2;
+    v16u8 mask0, mask1, mask2, out;
+    v8i16 filt, out0, out1, out2, out3;
+
+    mask0 = LD_UB(&vp8_mc_filt_mask_arr[16]);
+    src -= 2;
+
+    filt = LD_SH(filter);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt0, filt1, filt2);
+
+    mask1 = mask0 + 2;
+    mask2 = mask0 + 4;
+
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    src += (4 * src_stride);
+    HORIZ_6TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               filt0, filt1, filt2, out0, out1);
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_6TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               filt0, filt1, filt2, out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, VP8_FILTER_SHIFT);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+    dst += (4 * dst_stride);
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_hz_6t_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    if (4 == height)
+    {
+        common_hz_6t_4x4_msa(src, src_stride, dst, dst_stride, filter);
+    }
+    else if (8 == height)
+    {
+        common_hz_6t_4x8_msa(src, src_stride, dst, dst_stride, filter);
+    }
+}
+
+static void common_hz_6t_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, filt0, filt1, filt2;
+    v16u8 mask0, mask1, mask2, tmp0, tmp1;
+    v8i16 filt, out0, out1, out2, out3;
+
+    mask0 = LD_UB(&vp8_mc_filt_mask_arr[0]);
+    src -= 2;
+
+    filt = LD_SH(filter);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt0, filt1, filt2);
+
+    mask1 = mask0 + 2;
+    mask2 = mask0 + 4;
+
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    src += (4 * src_stride);
+    HORIZ_6TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               filt0, filt1, filt2, out0, out1, out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, VP8_FILTER_SHIFT);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    tmp0 = PCKEV_XORI128_UB(out0, out1);
+    tmp1 = PCKEV_XORI128_UB(out2, out3);
+    ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    for (loop_cnt = (height >> 2) - 1; loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src0, src1, src2, src3);
+        XORI_B4_128_SB(src0, src1, src2, src3);
+        src += (4 * src_stride);
+        HORIZ_6TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                                   filt0, filt1, filt2, out0, out1, out2, out3);
+        SRARI_H4_SH(out0, out1, out2, out3, VP8_FILTER_SHIFT);
+        SAT_SH4_SH(out0, out1, out2, out3, 7);
+        tmp0 = PCKEV_XORI128_UB(out0, out1);
+        tmp1 = PCKEV_XORI128_UB(out2, out3);
+        ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+        dst += (4 * dst_stride);
+    }
+}
+
+static void common_hz_6t_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, filt0, filt1, filt2;
+    v16u8 mask0, mask1, mask2, out;
+    v8i16 filt, out0, out1, out2, out3, out4, out5, out6, out7;
+
+    mask0 = LD_UB(&vp8_mc_filt_mask_arr[0]);
+    src -= 2;
+
+    filt = LD_SH(filter);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt0, filt1, filt2);
+
+    mask1 = mask0 + 2;
+    mask2 = mask0 + 4;
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src0, src2, src4, src6);
+        LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+        XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7);
+        src += (4 * src_stride);
+
+        HORIZ_6TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                                   filt0, filt1, filt2, out0, out1, out2, out3);
+        HORIZ_6TAP_8WID_4VECS_FILT(src4, src5, src6, src7, mask0, mask1, mask2,
+                                   filt0, filt1, filt2, out4, out5, out6, out7);
+        SRARI_H4_SH(out0, out1, out2, out3, VP8_FILTER_SHIFT);
+        SRARI_H4_SH(out4, out5, out6, out7, VP8_FILTER_SHIFT);
+        SAT_SH4_SH(out0, out1, out2, out3, 7);
+        SAT_SH4_SH(out4, out5, out6, out7, 7);
+        out = PCKEV_XORI128_UB(out0, out1);
+        ST_UB(out, dst);
+        dst += dst_stride;
+        out = PCKEV_XORI128_UB(out2, out3);
+        ST_UB(out, dst);
+        dst += dst_stride;
+        out = PCKEV_XORI128_UB(out4, out5);
+        ST_UB(out, dst);
+        dst += dst_stride;
+        out = PCKEV_XORI128_UB(out6, out7);
+        ST_UB(out, dst);
+        dst += dst_stride;
+    }
+}
+
+static void common_vt_6t_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+    v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r;
+    v16i8 src87_r, src2110, src4332, src6554, src8776, filt0, filt1, filt2;
+    v16u8 out;
+    v8i16 filt, out10, out32;
+
+    src -= (2 * src_stride);
+
+    filt = LD_SH(filter);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt0, filt1, filt2);
+
+    LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+    src += (5 * src_stride);
+
+    ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+               src32_r, src43_r);
+    ILVR_D2_SB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+    XORI_B2_128_SB(src2110, src4332);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src5, src6, src7, src8);
+        src += (4 * src_stride);
+
+        ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r,
+                   src65_r, src76_r, src87_r);
+        ILVR_D2_SB(src65_r, src54_r, src87_r, src76_r, src6554, src8776);
+        XORI_B2_128_SB(src6554, src8776);
+        out10 = DPADD_SH3_SH(src2110, src4332, src6554, filt0, filt1, filt2);
+        out32 = DPADD_SH3_SH(src4332, src6554, src8776, filt0, filt1, filt2);
+        SRARI_H2_SH(out10, out32, VP8_FILTER_SHIFT);
+        SAT_SH2_SH(out10, out32, 7);
+        out = PCKEV_XORI128_UB(out10, out32);
+        ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        src2110 = src6554;
+        src4332 = src8776;
+        src4 = src8;
+    }
+}
+
+static void common_vt_6t_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src7, src8, src9, src10;
+    v16i8 src10_r, src32_r, src76_r, src98_r, src21_r, src43_r, src87_r;
+    v16i8 src109_r, filt0, filt1, filt2;
+    v16u8 tmp0, tmp1;
+    v8i16 filt, out0_r, out1_r, out2_r, out3_r;
+
+    src -= (2 * src_stride);
+
+    filt = LD_SH(filter);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt0, filt1, filt2);
+
+    LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+    src += (5 * src_stride);
+
+    XORI_B5_128_SB(src0, src1, src2, src3, src4);
+    ILVR_B4_SB(src1, src0, src3, src2, src2, src1, src4, src3, src10_r, src32_r,
+               src21_r, src43_r);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src7, src8, src9, src10);
+        XORI_B4_128_SB(src7, src8, src9, src10);
+        src += (4 * src_stride);
+
+        ILVR_B4_SB(src7, src4, src8, src7, src9, src8, src10, src9, src76_r,
+                   src87_r, src98_r, src109_r);
+        out0_r = DPADD_SH3_SH(src10_r, src32_r, src76_r, filt0, filt1, filt2);
+        out1_r = DPADD_SH3_SH(src21_r, src43_r, src87_r, filt0, filt1, filt2);
+        out2_r = DPADD_SH3_SH(src32_r, src76_r, src98_r, filt0, filt1, filt2);
+        out3_r = DPADD_SH3_SH(src43_r, src87_r, src109_r, filt0, filt1, filt2);
+        SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, VP8_FILTER_SHIFT);
+        SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+        tmp0 = PCKEV_XORI128_UB(out0_r, out1_r);
+        tmp1 = PCKEV_XORI128_UB(out2_r, out3_r);
+        ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        src10_r = src76_r;
+        src32_r = src98_r;
+        src21_r = src87_r;
+        src43_r = src109_r;
+        src4 = src10;
+    }
+}
+
+static void common_vt_6t_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+    v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r;
+    v16i8 src87_r, src10_l, src32_l, src54_l, src76_l, src21_l, src43_l;
+    v16i8 src65_l, src87_l, filt0, filt1, filt2;
+    v16u8 tmp0, tmp1, tmp2, tmp3;
+    v8i16 out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l, filt;
+
+    src -= (2 * src_stride);
+
+    filt = LD_SH(filter);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt0, filt1, filt2);
+
+    LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+    src += (5 * src_stride);
+
+    XORI_B5_128_SB(src0, src1, src2, src3, src4);
+    ILVR_B4_SB(src1, src0, src3, src2, src4, src3, src2, src1, src10_r,
+               src32_r, src43_r, src21_r);
+    ILVL_B4_SB(src1, src0, src3, src2, src4, src3, src2, src1, src10_l,
+               src32_l, src43_l, src21_l);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src5, src6, src7, src8);
+        src += (4 * src_stride);
+
+        XORI_B4_128_SB(src5, src6, src7, src8);
+        ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r,
+                   src65_r, src76_r, src87_r);
+        ILVL_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_l,
+                   src65_l, src76_l, src87_l);
+        out0_r = DPADD_SH3_SH(src10_r, src32_r, src54_r, filt0, filt1, filt2);
+        out1_r = DPADD_SH3_SH(src21_r, src43_r, src65_r, filt0, filt1, filt2);
+        out2_r = DPADD_SH3_SH(src32_r, src54_r, src76_r, filt0, filt1, filt2);
+        out3_r = DPADD_SH3_SH(src43_r, src65_r, src87_r, filt0, filt1, filt2);
+        out0_l = DPADD_SH3_SH(src10_l, src32_l, src54_l, filt0, filt1, filt2);
+        out1_l = DPADD_SH3_SH(src21_l, src43_l, src65_l, filt0, filt1, filt2);
+        out2_l = DPADD_SH3_SH(src32_l, src54_l, src76_l, filt0, filt1, filt2);
+        out3_l = DPADD_SH3_SH(src43_l, src65_l, src87_l, filt0, filt1, filt2);
+        SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, VP8_FILTER_SHIFT);
+        SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, VP8_FILTER_SHIFT);
+        SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+        SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7);
+        PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l,
+                    out3_r, tmp0, tmp1, tmp2, tmp3);
+        XORI_B4_128_UB(tmp0, tmp1, tmp2, tmp3);
+        ST_UB4(tmp0, tmp1, tmp2, tmp3, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        src10_r = src54_r;
+        src32_r = src76_r;
+        src21_r = src65_r;
+        src43_r = src87_r;
+        src10_l = src54_l;
+        src32_l = src76_l;
+        src21_l = src65_l;
+        src43_l = src87_l;
+        src4 = src8;
+    }
+}
+
+static void common_hv_6ht_6vt_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+    v16i8 filt_hz0, filt_hz1, filt_hz2;
+    v16u8 mask0, mask1, mask2, out;
+    v8i16 tmp0, tmp1;
+    v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+    v8i16 hz_out7, filt, filt_vt0, filt_vt1, filt_vt2, out0, out1, out2, out3;
+
+    mask0 = LD_UB(&vp8_mc_filt_mask_arr[16]);
+    src -= (2 + 2 * src_stride);
+
+    filt = LD_SH(filter_horiz);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt_hz0, filt_hz1, filt_hz2);
+    filt = LD_SH(filter_vert);
+    SPLATI_H3_SH(filt, 0, 1, 2, filt_vt0, filt_vt1, filt_vt2);
+
+    mask1 = mask0 + 2;
+    mask2 = mask0 + 4;
+
+    LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+    src += (5 * src_stride);
+
+    XORI_B5_128_SB(src0, src1, src2, src3, src4);
+    hz_out0 = HORIZ_6TAP_FILT(src0, src1, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out2 = HORIZ_6TAP_FILT(src2, src3, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out1 = (v8i16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+    hz_out3 = HORIZ_6TAP_FILT(src3, src4, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB2(src, src_stride, src5, src6);
+        src += (2 * src_stride);
+
+        XORI_B2_128_SB(src5, src6);
+        hz_out5 = HORIZ_6TAP_FILT(src5, src6, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        hz_out4 = (v8i16)__msa_sldi_b((v16i8)hz_out5, (v16i8)hz_out3, 8);
+
+        LD_SB2(src, src_stride, src7, src8);
+        src += (2 * src_stride);
+
+        XORI_B2_128_SB(src7, src8);
+        hz_out7 = HORIZ_6TAP_FILT(src7, src8, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        hz_out6 = (v8i16)__msa_sldi_b((v16i8)hz_out7, (v16i8)hz_out5, 8);
+
+        out2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
+        tmp0 = DPADD_SH3_SH(out0, out1, out2, filt_vt0, filt_vt1, filt_vt2);
+
+        out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+        tmp1 = DPADD_SH3_SH(out1, out2, out3, filt_vt0, filt_vt1, filt_vt2);
+
+        SRARI_H2_SH(tmp0, tmp1, 7);
+        SAT_SH2_SH(tmp0, tmp1, 7);
+        out = PCKEV_XORI128_UB(tmp0, tmp1);
+        ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        hz_out3 = hz_out7;
+        out0 = out2;
+        out1 = out3;
+    }
+}
+
+static void common_hv_6ht_6vt_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+    v16i8 filt_hz0, filt_hz1, filt_hz2;
+    v16u8 mask0, mask1, mask2, vec0, vec1;
+    v8i16 filt, filt_vt0, filt_vt1, filt_vt2;
+    v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4,  hz_out5, hz_out6;
+    v8i16 hz_out7, hz_out8, out0, out1, out2, out3, out4, out5, out6, out7;
+    v8i16 tmp0, tmp1, tmp2, tmp3;
+
+    mask0 = LD_UB(&vp8_mc_filt_mask_arr[0]);
+    src -= (2 + 2 * src_stride);
+
+    filt = LD_SH(filter_horiz);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt_hz0, filt_hz1, filt_hz2);
+
+    mask1 = mask0 + 2;
+    mask2 = mask0 + 4;
+
+    LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+    src += (5 * src_stride);
+
+    XORI_B5_128_SB(src0, src1, src2, src3, src4);
+    hz_out0 = HORIZ_6TAP_FILT(src0, src0, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out1 = HORIZ_6TAP_FILT(src1, src1, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out2 = HORIZ_6TAP_FILT(src2, src2, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out3 = HORIZ_6TAP_FILT(src3, src3, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out4 = HORIZ_6TAP_FILT(src4, src4, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+
+    filt = LD_SH(filter_vert);
+    SPLATI_H3_SH(filt, 0, 1, 2, filt_vt0, filt_vt1, filt_vt2);
+
+    ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
+    ILVEV_B2_SH(hz_out1, hz_out2, hz_out3, hz_out4, out3, out4);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src5, src6, src7, src8);
+        src += (4 * src_stride);
+
+        XORI_B4_128_SB(src5, src6, src7, src8);
+        hz_out5 = HORIZ_6TAP_FILT(src5, src5, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        out2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
+        tmp0 = DPADD_SH3_SH(out0, out1, out2, filt_vt0, filt_vt1, filt_vt2);
+
+        hz_out6 = HORIZ_6TAP_FILT(src6, src6, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        out5 = (v8i16)__msa_ilvev_b((v16i8)hz_out6, (v16i8)hz_out5);
+        tmp1 = DPADD_SH3_SH(out3, out4, out5, filt_vt0, filt_vt1, filt_vt2);
+
+        hz_out7 = HORIZ_6TAP_FILT(src7, src7, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        out7 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+        tmp2 = DPADD_SH3_SH(out1, out2, out7, filt_vt0, filt_vt1, filt_vt2);
+
+        hz_out8 = HORIZ_6TAP_FILT(src8, src8, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        out6 = (v8i16)__msa_ilvev_b((v16i8)hz_out8, (v16i8)hz_out7);
+        tmp3 = DPADD_SH3_SH(out4, out5, out6, filt_vt0, filt_vt1, filt_vt2);
+
+        SRARI_H4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+        SAT_SH4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+        vec0 = PCKEV_XORI128_UB(tmp0, tmp1);
+        vec1 = PCKEV_XORI128_UB(tmp2, tmp3);
+        ST8x4_UB(vec0, vec1, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        hz_out4 = hz_out8;
+        out0 = out2;
+        out1 = out7;
+        out3 = out5;
+        out4 = out6;
+    }
+}
+
+static void common_hv_6ht_6vt_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                      uint8_t *RESTRICT dst, int32_t dst_stride,
+                                      const int8_t *filter_horiz,
+                                      const int8_t *filter_vert,
+                                      int32_t height)
+{
+    int32_t multiple8_cnt;
+    for (multiple8_cnt = 2; multiple8_cnt--;)
+    {
+        common_hv_6ht_6vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                                 filter_vert, height);
+        src += 8;
+        dst += 8;
+    }
+}
+
+static void common_hz_4t_4x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter)
+{
+    v16i8 src0, src1, src2, src3, filt0, filt1, mask0, mask1;
+    v8i16 filt, out0, out1;
+    v16u8 out;
+
+    mask0 = LD_SB(&vp8_mc_filt_mask_arr[16]);
+    src -= 1;
+
+    filt = LD_SH(filter);
+    SPLATI_H2_SB(filt, 0, 1, filt0, filt1);
+
+    mask1 = mask0 + 2;
+
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_4TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,
+                               filt0, filt1, out0, out1);
+    SRARI_H2_SH(out0, out1, VP8_FILTER_SHIFT);
+    SAT_SH2_SH(out0, out1, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_hz_4t_4x8_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter)
+{
+    v16i8 src0, src1, src2, src3, filt0, filt1, mask0, mask1;
+    v16u8 out;
+    v8i16 filt, out0, out1, out2, out3;
+
+    mask0 = LD_SB(&vp8_mc_filt_mask_arr[16]);
+    src -= 1;
+
+    filt = LD_SH(filter);
+    SPLATI_H2_SB(filt, 0, 1, filt0, filt1);
+
+    mask1 = mask0 + 2;
+
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_4TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,
+                               filt0, filt1, out0, out1);
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_4TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,
+                               filt0, filt1, out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, VP8_FILTER_SHIFT);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+    dst += (4 * dst_stride);
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_hz_4t_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    if (4 == height)
+    {
+        common_hz_4t_4x4_msa(src, src_stride, dst, dst_stride, filter);
+    }
+    else if (8 == height)
+    {
+        common_hz_4t_4x8_msa(src, src_stride, dst, dst_stride, filter);
+    }
+}
+
+static void common_hz_4t_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, filt0, filt1, mask0, mask1;
+    v16u8 tmp0, tmp1;
+    v8i16 filt, out0, out1, out2, out3;
+
+    mask0 = LD_SB(&vp8_mc_filt_mask_arr[0]);
+    src -= 1;
+
+    filt = LD_SH(filter);
+    SPLATI_H2_SB(filt, 0, 1, filt0, filt1);
+
+    mask1 = mask0 + 2;
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src0, src1, src2, src3);
+        src += (4 * src_stride);
+
+        XORI_B4_128_SB(src0, src1, src2, src3);
+        HORIZ_4TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, filt0,
+                                   filt1, out0, out1, out2, out3);
+        SRARI_H4_SH(out0, out1, out2, out3, VP8_FILTER_SHIFT);
+        SAT_SH4_SH(out0, out1, out2, out3, 7);
+        tmp0 = PCKEV_XORI128_UB(out0, out1);
+        tmp1 = PCKEV_XORI128_UB(out2, out3);
+        ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+        dst += (4 * dst_stride);
+    }
+}
+
+static void common_hz_4t_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+    v16i8 filt0, filt1, mask0, mask1;
+    v8i16 filt, out0, out1, out2, out3, out4, out5, out6, out7;
+    v16u8 out;
+
+    mask0 = LD_SB(&vp8_mc_filt_mask_arr[0]);
+    src -= 1;
+
+    filt = LD_SH(filter);
+    SPLATI_H2_SB(filt, 0, 1, filt0, filt1);
+
+    mask1 = mask0 + 2;
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src0, src2, src4, src6);
+        LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+        src += (4 * src_stride);
+
+        XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7);
+        HORIZ_4TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, filt0,
+                                   filt1, out0, out1, out2, out3);
+        HORIZ_4TAP_8WID_4VECS_FILT(src4, src5, src6, src7, mask0, mask1, filt0,
+                                   filt1, out4, out5, out6, out7);
+        SRARI_H4_SH(out0, out1, out2, out3, VP8_FILTER_SHIFT);
+        SRARI_H4_SH(out4, out5, out6, out7, VP8_FILTER_SHIFT);
+        SAT_SH4_SH(out0, out1, out2, out3, 7);
+        SAT_SH4_SH(out4, out5, out6, out7, 7);
+        out = PCKEV_XORI128_UB(out0, out1);
+        ST_UB(out, dst);
+        dst += dst_stride;
+        out = PCKEV_XORI128_UB(out2, out3);
+        ST_UB(out, dst);
+        dst += dst_stride;
+        out = PCKEV_XORI128_UB(out4, out5);
+        ST_UB(out, dst);
+        dst += dst_stride;
+        out = PCKEV_XORI128_UB(out6, out7);
+        ST_UB(out, dst);
+        dst += dst_stride;
+    }
+}
+
+static void common_vt_4t_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5;
+    v16i8 src10_r, src32_r, src54_r, src21_r, src43_r, src65_r;
+    v16i8 src2110, src4332, filt0, filt1;
+    v8i16 filt, out10, out32;
+    v16u8 out;
+
+    src -= src_stride;
+
+    filt = LD_SH(filter);
+    SPLATI_H2_SB(filt, 0, 1, filt0, filt1);
+
+    LD_SB3(src, src_stride, src0, src1, src2);
+    src += (3 * src_stride);
+
+    ILVR_B2_SB(src1, src0, src2, src1, src10_r, src21_r);
+
+    src2110 = (v16i8)__msa_ilvr_d((v2i64)src21_r, (v2i64)src10_r);
+    src2110 = (v16i8)__msa_xori_b((v16u8)src2110, 128);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB3(src, src_stride, src3, src4, src5);
+        src += (3 * src_stride);
+        ILVR_B2_SB(src3, src2, src4, src3, src32_r, src43_r);
+        src4332 = (v16i8)__msa_ilvr_d((v2i64)src43_r, (v2i64)src32_r);
+        src4332 = (v16i8)__msa_xori_b((v16u8)src4332, 128);
+        out10 = FILT_4TAP_DPADD_S_H(src2110, src4332, filt0, filt1);
+
+        src2 = LD_SB(src);
+        src += (src_stride);
+        ILVR_B2_SB(src5, src4, src2, src5, src54_r, src65_r);
+        src2110 = (v16i8)__msa_ilvr_d((v2i64)src65_r, (v2i64)src54_r);
+        src2110 = (v16i8)__msa_xori_b((v16u8)src2110, 128);
+        out32 = FILT_4TAP_DPADD_S_H(src4332, src2110, filt0, filt1);
+        SRARI_H2_SH(out10, out32, VP8_FILTER_SHIFT);
+        SAT_SH2_SH(out10, out32, 7);
+        out = PCKEV_XORI128_UB(out10, out32);
+        ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+        dst += (4 * dst_stride);
+    }
+}
+
+static void common_vt_4t_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                uint8_t *RESTRICT dst, int32_t dst_stride,
+                                const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src7, src8, src9, src10;
+    v16i8 src10_r, src72_r, src98_r, src21_r, src87_r, src109_r, filt0, filt1;
+    v16u8 tmp0, tmp1;
+    v8i16 filt, out0_r, out1_r, out2_r, out3_r;
+
+    src -= src_stride;
+
+    filt = LD_SH(filter);
+    SPLATI_H2_SB(filt, 0, 1, filt0, filt1);
+
+    LD_SB3(src, src_stride, src0, src1, src2);
+    src += (3 * src_stride);
+
+    XORI_B3_128_SB(src0, src1, src2);
+    ILVR_B2_SB(src1, src0, src2, src1, src10_r, src21_r);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src7, src8, src9, src10);
+        src += (4 * src_stride);
+
+        XORI_B4_128_SB(src7, src8, src9, src10);
+        ILVR_B4_SB(src7, src2, src8, src7, src9, src8, src10, src9,
+                   src72_r, src87_r, src98_r, src109_r);
+        out0_r = FILT_4TAP_DPADD_S_H(src10_r, src72_r, filt0, filt1);
+        out1_r = FILT_4TAP_DPADD_S_H(src21_r, src87_r, filt0, filt1);
+        out2_r = FILT_4TAP_DPADD_S_H(src72_r, src98_r, filt0, filt1);
+        out3_r = FILT_4TAP_DPADD_S_H(src87_r, src109_r, filt0, filt1);
+        SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, VP8_FILTER_SHIFT);
+        SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+        tmp0 = PCKEV_XORI128_UB(out0_r, out1_r);
+        tmp1 = PCKEV_XORI128_UB(out2_r, out3_r);
+        ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        src10_r = src98_r;
+        src21_r = src109_r;
+        src2 = src10;
+    }
+}
+
+static void common_vt_4t_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride,
+                                 const int8_t *filter, int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6;
+    v16i8 src10_r, src32_r, src54_r, src21_r, src43_r, src65_r, src10_l;
+    v16i8 src32_l, src54_l, src21_l, src43_l, src65_l, filt0, filt1;
+    v16u8 tmp0, tmp1, tmp2, tmp3;
+    v8i16 filt, out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l;
+
+    src -= src_stride;
+
+    filt = LD_SH(filter);
+    SPLATI_H2_SB(filt, 0, 1, filt0, filt1);
+
+    LD_SB3(src, src_stride, src0, src1, src2);
+    src += (3 * src_stride);
+
+    XORI_B3_128_SB(src0, src1, src2);
+    ILVR_B2_SB(src1, src0, src2, src1, src10_r, src21_r);
+    ILVL_B2_SB(src1, src0, src2, src1, src10_l, src21_l);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src3, src4, src5, src6);
+        src += (4 * src_stride);
+
+        XORI_B4_128_SB(src3, src4, src5, src6);
+        ILVR_B4_SB(src3, src2, src4, src3, src5, src4, src6, src5,
+                   src32_r, src43_r, src54_r, src65_r);
+        ILVL_B4_SB(src3, src2, src4, src3, src5, src4, src6, src5,
+                   src32_l, src43_l, src54_l, src65_l);
+        out0_r = FILT_4TAP_DPADD_S_H(src10_r, src32_r, filt0, filt1);
+        out1_r = FILT_4TAP_DPADD_S_H(src21_r, src43_r, filt0, filt1);
+        out2_r = FILT_4TAP_DPADD_S_H(src32_r, src54_r, filt0, filt1);
+        out3_r = FILT_4TAP_DPADD_S_H(src43_r, src65_r, filt0, filt1);
+        out0_l = FILT_4TAP_DPADD_S_H(src10_l, src32_l, filt0, filt1);
+        out1_l = FILT_4TAP_DPADD_S_H(src21_l, src43_l, filt0, filt1);
+        out2_l = FILT_4TAP_DPADD_S_H(src32_l, src54_l, filt0, filt1);
+        out3_l = FILT_4TAP_DPADD_S_H(src43_l, src65_l, filt0, filt1);
+        SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, VP8_FILTER_SHIFT);
+        SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, VP8_FILTER_SHIFT);
+        SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+        SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7);
+        PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l,
+                    out3_r, tmp0, tmp1, tmp2, tmp3);
+        XORI_B4_128_UB(tmp0, tmp1, tmp2, tmp3);
+        ST_UB4(tmp0, tmp1, tmp2, tmp3, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        src10_r = src54_r;
+        src21_r = src65_r;
+        src10_l = src54_l;
+        src21_l = src65_l;
+        src2 = src6;
+    }
+}
+
+static void common_hv_4ht_4vt_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, filt_hz0, filt_hz1;
+    v16u8 mask0, mask1, out;
+    v8i16 filt, filt_vt0, filt_vt1, tmp0, tmp1, vec0, vec1, vec2;
+    v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5;
+
+    mask0 = LD_UB(&vp8_mc_filt_mask_arr[16]);
+    src -= (1 + 1 * src_stride);
+
+    filt = LD_SH(filter_horiz);
+    SPLATI_H2_SB(filt, 0, 1, filt_hz0, filt_hz1);
+
+    mask1 = mask0 + 2;
+
+    LD_SB3(src, src_stride, src0, src1, src2);
+    src += (3 * src_stride);
+
+    XORI_B3_128_SB(src0, src1, src2);
+    hz_out0 = HORIZ_4TAP_FILT(src0, src1, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out1 = HORIZ_4TAP_FILT(src1, src2, mask0, mask1, filt_hz0, filt_hz1);
+    vec0 = (v8i16)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+
+    filt = LD_SH(filter_vert);
+    SPLATI_H2_SH(filt, 0, 1, filt_vt0, filt_vt1);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src3, src4, src5, src6);
+        src += (4 * src_stride);
+
+        XORI_B2_128_SB(src3, src4);
+        hz_out3 = HORIZ_4TAP_FILT(src3, src4, mask0, mask1, filt_hz0, filt_hz1);
+        hz_out2 = (v8i16)__msa_sldi_b((v16i8)hz_out3, (v16i8)hz_out1, 8);
+        vec1 = (v8i16)__msa_ilvev_b((v16i8)hz_out3, (v16i8)hz_out2);
+        tmp0 = FILT_4TAP_DPADD_S_H(vec0, vec1, filt_vt0, filt_vt1);
+
+        XORI_B2_128_SB(src5, src6);
+        hz_out5 = HORIZ_4TAP_FILT(src5, src6, mask0, mask1, filt_hz0, filt_hz1);
+        hz_out4 = (v8i16)__msa_sldi_b((v16i8)hz_out5, (v16i8)hz_out3, 8);
+        vec2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
+        tmp1 = FILT_4TAP_DPADD_S_H(vec1, vec2, filt_vt0, filt_vt1);
+
+        SRARI_H2_SH(tmp0, tmp1, 7);
+        SAT_SH2_SH(tmp0, tmp1, 7);
+        out = PCKEV_XORI128_UB(tmp0, tmp1);
+        ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        hz_out1 = hz_out5;
+        vec0 = vec2;
+    }
+}
+
+static void common_hv_4ht_4vt_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, filt_hz0, filt_hz1;
+    v16u8 mask0, mask1, out0, out1;
+    v8i16 filt, filt_vt0, filt_vt1, tmp0, tmp1, tmp2, tmp3;
+    v8i16 hz_out0, hz_out1, hz_out2, hz_out3;
+    v8i16 vec0, vec1, vec2, vec3, vec4;
+
+    mask0 = LD_UB(&vp8_mc_filt_mask_arr[0]);
+    src -= (1 + 1 * src_stride);
+
+    filt = LD_SH(filter_horiz);
+    SPLATI_H2_SB(filt, 0, 1, filt_hz0, filt_hz1);
+
+    mask1 = mask0 + 2;
+
+    LD_SB3(src, src_stride, src0, src1, src2);
+    src += (3 * src_stride);
+
+    XORI_B3_128_SB(src0, src1, src2);
+    hz_out0 = HORIZ_4TAP_FILT(src0, src0, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out1 = HORIZ_4TAP_FILT(src1, src1, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out2 = HORIZ_4TAP_FILT(src2, src2, mask0, mask1, filt_hz0, filt_hz1);
+    ILVEV_B2_SH(hz_out0, hz_out1, hz_out1, hz_out2, vec0, vec2);
+
+    filt = LD_SH(filter_vert);
+    SPLATI_H2_SH(filt, 0, 1, filt_vt0, filt_vt1);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src3, src4, src5, src6);
+        src += (4 * src_stride);
+
+        XORI_B4_128_SB(src3, src4, src5, src6);
+        hz_out3 = HORIZ_4TAP_FILT(src3, src3, mask0, mask1, filt_hz0, filt_hz1);
+        vec1 = (v8i16)__msa_ilvev_b((v16i8)hz_out3, (v16i8)hz_out2);
+        tmp0 = FILT_4TAP_DPADD_S_H(vec0, vec1, filt_vt0, filt_vt1);
+
+        hz_out0 = HORIZ_4TAP_FILT(src4, src4, mask0, mask1, filt_hz0, filt_hz1);
+        vec3 = (v8i16)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out3);
+        tmp1 = FILT_4TAP_DPADD_S_H(vec2, vec3, filt_vt0, filt_vt1);
+
+        hz_out1 = HORIZ_4TAP_FILT(src5, src5, mask0, mask1, filt_hz0, filt_hz1);
+        vec4 = (v8i16)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+        tmp2 = FILT_4TAP_DPADD_S_H(vec1, vec4, filt_vt0, filt_vt1);
+
+        hz_out2 = HORIZ_4TAP_FILT(src6, src6, mask0, mask1, filt_hz0, filt_hz1);
+        ILVEV_B2_SH(hz_out3, hz_out0, hz_out1, hz_out2, vec0, vec1);
+        tmp3 = FILT_4TAP_DPADD_S_H(vec0, vec1, filt_vt0, filt_vt1);
+
+        SRARI_H4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+        SAT_SH4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+        out0 = PCKEV_XORI128_UB(tmp0, tmp1);
+        out1 = PCKEV_XORI128_UB(tmp2, tmp3);
+        ST8x4_UB(out0, out1, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        vec0 = vec4;
+        vec2 = vec1;
+    }
+}
+
+static void common_hv_4ht_4vt_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                      uint8_t *RESTRICT dst, int32_t dst_stride,
+                                      const int8_t *filter_horiz,
+                                      const int8_t *filter_vert,
+                                      int32_t height)
+{
+    int32_t multiple8_cnt;
+    for (multiple8_cnt = 2; multiple8_cnt--;)
+    {
+        common_hv_4ht_4vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                                 filter_vert, height);
+        src += 8;
+        dst += 8;
+    }
+}
+
+static void common_hv_6ht_4vt_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6;
+    v16i8 filt_hz0, filt_hz1, filt_hz2;
+    v16u8 res0, res1, mask0, mask1, mask2;
+    v8i16 filt, filt_vt0, filt_vt1, tmp0, tmp1, vec0, vec1, vec2;
+    v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5;
+
+    mask0 = LD_UB(&vp8_mc_filt_mask_arr[16]);
+    src -= (2 + 1 * src_stride);
+
+    filt = LD_SH(filter_horiz);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt_hz0, filt_hz1, filt_hz2);
+
+    mask1 = mask0 + 2;
+    mask2 = mask0 + 4;
+
+    LD_SB3(src, src_stride, src0, src1, src2);
+    src += (3 * src_stride);
+
+    XORI_B3_128_SB(src0, src1, src2);
+    hz_out0 = HORIZ_6TAP_FILT(src0, src1, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out1 = HORIZ_6TAP_FILT(src1, src2, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    vec0 = (v8i16)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+
+    filt = LD_SH(filter_vert);
+    SPLATI_H2_SH(filt, 0, 1, filt_vt0, filt_vt1);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src3, src4, src5, src6);
+        src += (4 * src_stride);
+
+        XORI_B4_128_SB(src3, src4, src5, src6);
+        hz_out3 = HORIZ_6TAP_FILT(src3, src4, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        hz_out2 = (v8i16)__msa_sldi_b((v16i8)hz_out3, (v16i8)hz_out1, 8);
+        vec1 = (v8i16)__msa_ilvev_b((v16i8)hz_out3, (v16i8)hz_out2);
+        tmp0 = FILT_4TAP_DPADD_S_H(vec0, vec1, filt_vt0, filt_vt1);
+
+        hz_out5 = HORIZ_6TAP_FILT(src5, src6, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        hz_out4 = (v8i16)__msa_sldi_b((v16i8)hz_out5, (v16i8)hz_out3, 8);
+        vec2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
+        tmp1 = FILT_4TAP_DPADD_S_H(vec1, vec2, filt_vt0, filt_vt1);
+
+        SRARI_H2_SH(tmp0, tmp1, 7);
+        SAT_SH2_SH(tmp0, tmp1, 7);
+        PCKEV_B2_UB(tmp0, tmp0, tmp1, tmp1, res0, res1);
+        XORI_B2_128_UB(res0, res1);
+        ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        hz_out1 = hz_out5;
+        vec0 = vec2;
+    }
+}
+
+static void common_hv_6ht_4vt_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6;
+    v16i8 filt_hz0, filt_hz1, filt_hz2, mask0, mask1, mask2;
+    v8i16 filt, filt_vt0, filt_vt1, hz_out0, hz_out1, hz_out2, hz_out3;
+    v8i16 tmp0, tmp1, tmp2, tmp3, vec0, vec1, vec2, vec3;
+    v16u8 out0, out1;
+
+    mask0 = LD_SB(&vp8_mc_filt_mask_arr[0]);
+    src -= (2 + src_stride);
+
+    filt = LD_SH(filter_horiz);
+    SPLATI_H3_SB(filt, 0, 1, 2, filt_hz0, filt_hz1, filt_hz2);
+
+    mask1 = mask0 + 2;
+    mask2 = mask0 + 4;
+
+    LD_SB3(src, src_stride, src0, src1, src2);
+    src += (3 * src_stride);
+
+    XORI_B3_128_SB(src0, src1, src2);
+    hz_out0 = HORIZ_6TAP_FILT(src0, src0, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out1 = HORIZ_6TAP_FILT(src1, src1, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    hz_out2 = HORIZ_6TAP_FILT(src2, src2, mask0, mask1, mask2, filt_hz0,
+                              filt_hz1, filt_hz2);
+    ILVEV_B2_SH(hz_out0, hz_out1, hz_out1, hz_out2, vec0, vec2);
+
+    filt = LD_SH(filter_vert);
+    SPLATI_H2_SH(filt, 0, 1, filt_vt0, filt_vt1);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src3, src4, src5, src6);
+        src += (4 * src_stride);
+
+        XORI_B4_128_SB(src3, src4, src5, src6);
+
+        hz_out3 = HORIZ_6TAP_FILT(src3, src3, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        vec1 = (v8i16)__msa_ilvev_b((v16i8)hz_out3, (v16i8)hz_out2);
+        tmp0 = FILT_4TAP_DPADD_S_H(vec0, vec1, filt_vt0, filt_vt1);
+
+        hz_out0 = HORIZ_6TAP_FILT(src4, src4, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        vec3 = (v8i16)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out3);
+        tmp1 = FILT_4TAP_DPADD_S_H(vec2, vec3, filt_vt0, filt_vt1);
+
+        hz_out1 = HORIZ_6TAP_FILT(src5, src5, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        vec0 = (v8i16)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+        tmp2 = FILT_4TAP_DPADD_S_H(vec1, vec0, filt_vt0, filt_vt1);
+
+        hz_out2 = HORIZ_6TAP_FILT(src6, src6, mask0, mask1, mask2, filt_hz0,
+                                  filt_hz1, filt_hz2);
+        ILVEV_B2_SH(hz_out3, hz_out0, hz_out1, hz_out2, vec1, vec2);
+        tmp3 = FILT_4TAP_DPADD_S_H(vec1, vec2, filt_vt0, filt_vt1);
+
+        SRARI_H4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+        SAT_SH4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+        out0 = PCKEV_XORI128_UB(tmp0, tmp1);
+        out1 = PCKEV_XORI128_UB(tmp2, tmp3);
+        ST8x4_UB(out0, out1, dst, dst_stride);
+        dst += (4 * dst_stride);
+    }
+}
+
+static void common_hv_6ht_4vt_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                      uint8_t *RESTRICT dst, int32_t dst_stride,
+                                      const int8_t *filter_horiz,
+                                      const int8_t *filter_vert,
+                                      int32_t height)
+{
+    int32_t multiple8_cnt;
+    for (multiple8_cnt = 2; multiple8_cnt--;)
+    {
+        common_hv_6ht_4vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                                 filter_vert, height);
+        src += 8;
+        dst += 8;
+    }
+}
+
+static void common_hv_4ht_6vt_4w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+    v16i8 filt_hz0, filt_hz1, mask0, mask1;
+    v16u8 out;
+    v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+    v8i16 hz_out7, tmp0, tmp1, out0, out1, out2, out3;
+    v8i16 filt, filt_vt0, filt_vt1, filt_vt2;
+
+    mask0 = LD_SB(&vp8_mc_filt_mask_arr[16]);
+
+    src -= (1 + 2 * src_stride);
+
+    filt = LD_SH(filter_horiz);
+    SPLATI_H2_SB(filt, 0, 1, filt_hz0, filt_hz1);
+
+    mask1 = mask0 + 2;
+
+    LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+    src += (5 * src_stride);
+
+    XORI_B5_128_SB(src0, src1, src2, src3, src4);
+    hz_out0 = HORIZ_4TAP_FILT(src0, src1, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out2 = HORIZ_4TAP_FILT(src2, src3, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out3 = HORIZ_4TAP_FILT(src3, src4, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out1 = (v8i16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+    ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
+
+    filt = LD_SH(filter_vert);
+    SPLATI_H3_SH(filt, 0, 1, 2, filt_vt0, filt_vt1, filt_vt2);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src5, src6, src7, src8);
+        XORI_B4_128_SB(src5, src6, src7, src8);
+        src += (4 * src_stride);
+
+        hz_out5 = HORIZ_4TAP_FILT(src5, src6, mask0, mask1, filt_hz0, filt_hz1);
+        hz_out4 = (v8i16)__msa_sldi_b((v16i8)hz_out5, (v16i8)hz_out3, 8);
+        out2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
+        tmp0 = DPADD_SH3_SH(out0, out1, out2, filt_vt0, filt_vt1, filt_vt2);
+
+        hz_out7 = HORIZ_4TAP_FILT(src7, src8, mask0, mask1, filt_hz0, filt_hz1);
+        hz_out6 = (v8i16)__msa_sldi_b((v16i8)hz_out7, (v16i8)hz_out5, 8);
+        out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+        tmp1 = DPADD_SH3_SH(out1, out2, out3, filt_vt0, filt_vt1, filt_vt2);
+
+        SRARI_H2_SH(tmp0, tmp1, 7);
+        SAT_SH2_SH(tmp0, tmp1, 7);
+        out = PCKEV_XORI128_UB(tmp0, tmp1);
+        ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        hz_out3 = hz_out7;
+        out0 = out2;
+        out1 = out3;
+    }
+}
+
+static void common_hv_4ht_6vt_8w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                     uint8_t *RESTRICT dst, int32_t dst_stride,
+                                     const int8_t *filter_horiz,
+                                     const int8_t *filter_vert,
+                                     int32_t height)
+{
+    uint32_t loop_cnt;
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+    v16i8 filt_hz0, filt_hz1, mask0, mask1;
+    v8i16 filt, filt_vt0, filt_vt1, filt_vt2, tmp0, tmp1, tmp2, tmp3;
+    v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+    v8i16 hz_out7, hz_out8, out0, out1, out2, out3, out4, out5, out6, out7;
+    v16u8 vec0, vec1;
+
+    mask0 = LD_SB(&vp8_mc_filt_mask_arr[0]);
+    src -= (1 + 2 * src_stride);
+
+    filt = LD_SH(filter_horiz);
+    SPLATI_H2_SB(filt, 0, 1, filt_hz0, filt_hz1);
+
+    mask1 = mask0 + 2;
+
+    LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+    src += (5 * src_stride);
+
+    XORI_B5_128_SB(src0, src1, src2, src3, src4);
+    hz_out0 = HORIZ_4TAP_FILT(src0, src0, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out1 = HORIZ_4TAP_FILT(src1, src1, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out2 = HORIZ_4TAP_FILT(src2, src2, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out3 = HORIZ_4TAP_FILT(src3, src3, mask0, mask1, filt_hz0, filt_hz1);
+    hz_out4 = HORIZ_4TAP_FILT(src4, src4, mask0, mask1, filt_hz0, filt_hz1);
+    ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
+    ILVEV_B2_SH(hz_out1, hz_out2, hz_out3, hz_out4, out3, out4);
+
+    filt = LD_SH(filter_vert);
+    SPLATI_H3_SH(filt, 0, 1, 2, filt_vt0, filt_vt1, filt_vt2);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;)
+    {
+        LD_SB4(src, src_stride, src5, src6, src7, src8);
+        src += (4 * src_stride);
+
+        XORI_B4_128_SB(src5, src6, src7, src8);
+
+        hz_out5 = HORIZ_4TAP_FILT(src5, src5, mask0, mask1, filt_hz0, filt_hz1);
+        out2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
+        tmp0 = DPADD_SH3_SH(out0, out1, out2, filt_vt0, filt_vt1, filt_vt2);
+
+        hz_out6 = HORIZ_4TAP_FILT(src6, src6, mask0, mask1, filt_hz0, filt_hz1);
+        out5 = (v8i16)__msa_ilvev_b((v16i8)hz_out6, (v16i8)hz_out5);
+        tmp1 = DPADD_SH3_SH(out3, out4, out5, filt_vt0, filt_vt1, filt_vt2);
+
+        hz_out7 = HORIZ_4TAP_FILT(src7, src7, mask0, mask1, filt_hz0, filt_hz1);
+        out6 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+        tmp2 = DPADD_SH3_SH(out1, out2, out6, filt_vt0, filt_vt1, filt_vt2);
+
+        hz_out8 = HORIZ_4TAP_FILT(src8, src8, mask0, mask1, filt_hz0, filt_hz1);
+        out7 = (v8i16)__msa_ilvev_b((v16i8)hz_out8, (v16i8)hz_out7);
+        tmp3 = DPADD_SH3_SH(out4, out5, out7, filt_vt0, filt_vt1, filt_vt2);
+
+        SRARI_H4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+        SAT_SH4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+        vec0 = PCKEV_XORI128_UB(tmp0, tmp1);
+        vec1 = PCKEV_XORI128_UB(tmp2, tmp3);
+        ST8x4_UB(vec0, vec1, dst, dst_stride);
+        dst += (4 * dst_stride);
+
+        hz_out4 = hz_out8;
+        out0 = out2;
+        out1 = out6;
+        out3 = out5;
+        out4 = out7;
+    }
+}
+
+static void common_hv_4ht_6vt_16w_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                      uint8_t *RESTRICT dst, int32_t dst_stride,
+                                      const int8_t *filter_horiz,
+                                      const int8_t *filter_vert,
+                                      int32_t height)
+{
+    int32_t multiple8_cnt;
+    for (multiple8_cnt = 2; multiple8_cnt--;)
+    {
+        common_hv_4ht_6vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                                 filter_vert, height);
+        src += 8;
+        dst += 8;
+    }
+}
+
+void vp8_sixtap_predict4x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                               int32_t xoffset, int32_t yoffset,
+                               uint8_t *RESTRICT dst, int32_t dst_stride)
+{
+    const int8_t *h_filter = vp8_subpel_filters_msa[xoffset - 1];
+    const int8_t *v_filter = vp8_subpel_filters_msa[yoffset - 1];
+
+    if (yoffset)
+    {
+        if (xoffset)
+        {
+            switch (xoffset)
+            {
+                case 2:
+                case 4:
+                case 6:
+                    switch (yoffset)
+                    {
+                        case 2:
+                        case 4:
+                        case 6:
+                            common_hv_6ht_6vt_4w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter,
+                                                     v_filter, 4);
+                            break;
+
+                        case 1:
+                        case 3:
+                        case 5:
+                        case 7:
+                            common_hv_6ht_4vt_4w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter,
+                                                     v_filter + 1, 4);
+                            break;
+                    }
+                    break;
+
+                case 1:
+                case 3:
+                case 5:
+                case 7:
+                    switch (yoffset)
+                    {
+                        case 2:
+                        case 4:
+                        case 6:
+                            common_hv_4ht_6vt_4w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter + 1,
+                                                     v_filter, 4);
+                            break;
+
+                        case 1:
+                        case 3:
+                        case 5:
+                        case 7:
+                            common_hv_4ht_4vt_4w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter + 1,
+                                                     v_filter + 1, 4);
+                            break;
+                    }
+                    break;
+            }
+        }
+        else
+        {
+            switch (yoffset)
+            {
+                case 2:
+                case 4:
+                case 6:
+                    common_vt_6t_4w_msa(src, src_stride, dst, dst_stride,
+                                        v_filter, 4);
+                    break;
+
+                case 1:
+                case 3:
+                case 5:
+                case 7:
+                    common_vt_4t_4w_msa(src, src_stride, dst, dst_stride,
+                                        v_filter + 1, 4);
+                    break;
+            }
+        }
+    }
+    else
+    {
+        switch (xoffset)
+        {
+            case 0:
+                {
+                uint32_t tp0, tp1, tp2, tp3;
+
+                LW4(src, src_stride, tp0, tp1, tp2, tp3);
+                SW4(tp0, tp1, tp2, tp3, dst, dst_stride);
+                break;
+                }
+            case 2:
+            case 4:
+            case 6:
+                common_hz_6t_4w_msa(src, src_stride, dst, dst_stride,
+                                    h_filter, 4);
+                break;
+
+            case 1:
+            case 3:
+            case 5:
+            case 7:
+                common_hz_4t_4w_msa(src, src_stride, dst, dst_stride,
+                                    h_filter + 1, 4);
+                break;
+        }
+    }
+}
+
+void vp8_sixtap_predict8x4_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                               int32_t xoffset, int32_t yoffset,
+                               uint8_t *RESTRICT dst, int32_t dst_stride)
+{
+    const int8_t *h_filter = vp8_subpel_filters_msa[xoffset - 1];
+    const int8_t *v_filter = vp8_subpel_filters_msa[yoffset - 1];
+
+    if (yoffset)
+    {
+        if (xoffset)
+        {
+            switch (xoffset)
+            {
+                case 2:
+                case 4:
+                case 6:
+                    switch (yoffset)
+                    {
+                        case 2:
+                        case 4:
+                        case 6:
+                            common_hv_6ht_6vt_8w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter,
+                                                     v_filter, 4);
+                            break;
+
+                        case 1:
+                        case 3:
+                        case 5:
+                        case 7:
+                            common_hv_6ht_4vt_8w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter,
+                                                     v_filter + 1, 4);
+                            break;
+                    }
+                    break;
+
+                case 1:
+                case 3:
+                case 5:
+                case 7:
+                    switch (yoffset)
+                    {
+                        case 2:
+                        case 4:
+                        case 6:
+                            common_hv_4ht_6vt_8w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter + 1,
+                                                     v_filter, 4);
+                            break;
+
+                        case 1:
+                        case 3:
+                        case 5:
+                        case 7:
+                            common_hv_4ht_4vt_8w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter + 1,
+                                                     v_filter + 1, 4);
+                            break;
+                    }
+                    break;
+            }
+        }
+        else
+        {
+            switch (yoffset)
+            {
+                case 2:
+                case 4:
+                case 6:
+                    common_vt_6t_8w_msa(src, src_stride, dst, dst_stride,
+                                        v_filter, 4);
+                    break;
+
+                case 1:
+                case 3:
+                case 5:
+                case 7:
+                    common_vt_4t_8w_msa(src, src_stride, dst, dst_stride,
+                                        v_filter + 1, 4);
+                    break;
+            }
+        }
+    }
+    else
+    {
+        switch (xoffset)
+        {
+            case 0:
+                vp8_copy_mem8x4(src, src_stride, dst, dst_stride);
+                break;
+            case 2:
+            case 4:
+            case 6:
+                common_hz_6t_8w_msa(src, src_stride, dst, dst_stride,
+                                    h_filter, 4);
+                break;
+
+            case 1:
+            case 3:
+            case 5:
+            case 7:
+                common_hz_4t_8w_msa(src, src_stride, dst, dst_stride,
+                                    h_filter + 1, 4);
+                break;
+        }
+    }
+}
+
+void vp8_sixtap_predict8x8_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                               int32_t xoffset, int32_t yoffset,
+                               uint8_t *RESTRICT dst, int32_t dst_stride)
+{
+    const int8_t *h_filter = vp8_subpel_filters_msa[xoffset - 1];
+    const int8_t *v_filter = vp8_subpel_filters_msa[yoffset - 1];
+
+    if (yoffset)
+    {
+        if (xoffset)
+        {
+            switch (xoffset)
+            {
+                case 2:
+                case 4:
+                case 6:
+                    switch (yoffset)
+                    {
+                        case 2:
+                        case 4:
+                        case 6:
+                            common_hv_6ht_6vt_8w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter,
+                                                     v_filter, 8);
+                            break;
+
+                        case 1:
+                        case 3:
+                        case 5:
+                        case 7:
+                            common_hv_6ht_4vt_8w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter,
+                                                     v_filter + 1, 8);
+                            break;
+                    }
+                    break;
+
+                case 1:
+                case 3:
+                case 5:
+                case 7:
+                    switch (yoffset)
+                    {
+                        case 2:
+                        case 4:
+                        case 6:
+                            common_hv_4ht_6vt_8w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter + 1,
+                                                     v_filter, 8);
+                            break;
+
+                        case 1:
+                        case 3:
+                        case 5:
+                        case 7:
+                            common_hv_4ht_4vt_8w_msa(src, src_stride, dst,
+                                                     dst_stride, h_filter + 1,
+                                                     v_filter + 1, 8);
+                            break;
+                    }
+                    break;
+                }
+        }
+        else
+        {
+            switch (yoffset)
+            {
+                case 2:
+                case 4:
+                case 6:
+                    common_vt_6t_8w_msa(src, src_stride, dst, dst_stride,
+                                        v_filter, 8);
+                    break;
+
+                case 1:
+                case 3:
+                case 5:
+                case 7:
+                    common_vt_4t_8w_msa(src, src_stride, dst, dst_stride,
+                                        v_filter + 1, 8);
+                    break;
+            }
+        }
+    }
+    else
+    {
+        switch (xoffset)
+        {
+            case 0:
+                vp8_copy_mem8x8(src, src_stride, dst, dst_stride);
+                break;
+            case 2:
+            case 4:
+            case 6:
+                common_hz_6t_8w_msa(src, src_stride, dst, dst_stride, h_filter,
+                                    8);
+                break;
+
+            case 1:
+            case 3:
+            case 5:
+            case 7:
+                common_hz_4t_8w_msa(src, src_stride, dst, dst_stride,
+                                    h_filter + 1, 8);
+                break;
+        }
+    }
+}
+
+void vp8_sixtap_predict16x16_msa(uint8_t *RESTRICT src, int32_t src_stride,
+                                 int32_t xoffset, int32_t yoffset,
+                                 uint8_t *RESTRICT dst, int32_t dst_stride)
+{
+    const int8_t *h_filter = vp8_subpel_filters_msa[xoffset - 1];
+    const int8_t *v_filter = vp8_subpel_filters_msa[yoffset - 1];
+
+    if (yoffset)
+    {
+        if (xoffset)
+        {
+            switch (xoffset)
+            {
+                case 2:
+                case 4:
+                case 6:
+                    switch (yoffset)
+                    {
+                        case 2:
+                        case 4:
+                        case 6:
+                            common_hv_6ht_6vt_16w_msa(src, src_stride, dst,
+                                                      dst_stride, h_filter,
+                                                      v_filter, 16);
+                            break;
+
+                        case 1:
+                        case 3:
+                        case 5:
+                        case 7:
+                            common_hv_6ht_4vt_16w_msa(src, src_stride, dst,
+                                                      dst_stride, h_filter,
+                                                      v_filter + 1, 16);
+                            break;
+                    }
+                    break;
+
+                case 1:
+                case 3:
+                case 5:
+                case 7:
+                    switch (yoffset)
+                    {
+                        case 2:
+                        case 4:
+                        case 6:
+                            common_hv_4ht_6vt_16w_msa(src, src_stride, dst,
+                                                      dst_stride, h_filter + 1,
+                                                      v_filter, 16);
+                            break;
+
+                        case 1:
+                        case 3:
+                        case 5:
+                        case 7:
+                            common_hv_4ht_4vt_16w_msa(src, src_stride, dst,
+                                                      dst_stride, h_filter + 1,
+                                                      v_filter + 1, 16);
+                            break;
+                    }
+                    break;
+            }
+        }
+        else
+        {
+            switch (yoffset)
+            {
+                case 2:
+                case 4:
+                case 6:
+                    common_vt_6t_16w_msa(src, src_stride, dst, dst_stride,
+                                         v_filter, 16);
+                    break;
+
+                case 1:
+                case 3:
+                case 5:
+                case 7:
+                    common_vt_4t_16w_msa(src, src_stride, dst, dst_stride,
+                                         v_filter + 1, 16);
+                    break;
+            }
+        }
+    }
+    else
+    {
+        switch (xoffset)
+        {
+            case 0:
+                vp8_copy_mem16x16(src, src_stride, dst, dst_stride);
+                break;
+            case 2:
+            case 4:
+            case 6:
+                common_hz_6t_16w_msa(src, src_stride, dst, dst_stride,
+                                     h_filter, 16);
+                break;
+
+            case 1:
+            case 3:
+            case 5:
+            case 7:
+                common_hz_4t_16w_msa(src, src_stride, dst, dst_stride,
+                                     h_filter + 1, 16);
+                break;
+        }
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/mips/msa/vp8_macros_msa.h b/libvpx/libvpx/vp8/common/mips/msa/vp8_macros_msa.h
new file mode 100644
index 0000000..27d5929
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mips/msa/vp8_macros_msa.h
@@ -0,0 +1,1783 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_COMMON_MIPS_MSA_VP8_MACROS_MSA_H_
+#define VP8_COMMON_MIPS_MSA_VP8_MACROS_MSA_H_
+
+#include <msa.h>
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+#define LD_B(RTYPE, psrc) *((const RTYPE *)(psrc))
+#define LD_UB(...) LD_B(v16u8, __VA_ARGS__)
+#define LD_SB(...) LD_B(v16i8, __VA_ARGS__)
+
+#define LD_H(RTYPE, psrc) *((const RTYPE *)(psrc))
+#define LD_UH(...) LD_H(v8u16, __VA_ARGS__)
+#define LD_SH(...) LD_H(v8i16, __VA_ARGS__)
+
+#define LD_W(RTYPE, psrc) *((const RTYPE *)(psrc))
+#define LD_UW(...) LD_W(v4u32, __VA_ARGS__)
+#define LD_SW(...) LD_W(v4i32, __VA_ARGS__)
+
+#define ST_B(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_UB(...) ST_B(v16u8, __VA_ARGS__)
+#define ST_SB(...) ST_B(v16i8, __VA_ARGS__)
+
+#define ST_H(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_UH(...) ST_H(v8u16, __VA_ARGS__)
+#define ST_SH(...) ST_H(v8i16, __VA_ARGS__)
+
+#define ST_W(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_SW(...) ST_W(v4i32, __VA_ARGS__)
+
+#if (__mips_isa_rev >= 6)
+#define LW(psrc)                                      \
+({                                                    \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);  \
+    uint32_t val_m;                                   \
+                                                      \
+    asm volatile (                                    \
+        "lw  %[val_m],  %[psrc_m]  \n\t"              \
+                                                      \
+        : [val_m] "=r" (val_m)                        \
+        : [psrc_m] "m" (*psrc_m)                      \
+    );                                                \
+                                                      \
+    val_m;                                            \
+})
+
+#if (__mips == 64)
+#define LD(psrc)                                      \
+({                                                    \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);  \
+    uint64_t val_m = 0;                               \
+                                                      \
+    asm volatile (                                    \
+        "ld  %[val_m],  %[psrc_m]  \n\t"              \
+                                                      \
+        : [val_m] "=r" (val_m)                        \
+        : [psrc_m] "m" (*psrc_m)                      \
+    );                                                \
+                                                      \
+    val_m;                                            \
+})
+#else  // !(__mips == 64)
+#define LD(psrc)                                             \
+({                                                           \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);         \
+    uint32_t val0_m, val1_m;                                 \
+    uint64_t val_m = 0;                                      \
+                                                             \
+    val0_m = LW(psrc_m);                                     \
+    val1_m = LW(psrc_m + 4);                                 \
+                                                             \
+    val_m = (uint64_t)(val1_m);                              \
+    val_m = (uint64_t)((val_m << 32) & 0xFFFFFFFF00000000);  \
+    val_m = (uint64_t)(val_m | (uint64_t)val0_m);            \
+                                                             \
+    val_m;                                                   \
+})
+#endif  // (__mips == 64)
+
+#define SH(val, pdst)                     \
+{                                         \
+    uint8_t *pdst_m = (uint8_t *)(pdst);  \
+    const uint16_t val_m = (val);         \
+                                          \
+    asm volatile (                        \
+        "sh  %[val_m],  %[pdst_m]  \n\t"  \
+                                          \
+        : [pdst_m] "=m" (*pdst_m)         \
+        : [val_m] "r" (val_m)             \
+    );                                    \
+}
+
+#define SW(val, pdst)                     \
+{                                         \
+    uint8_t *pdst_m = (uint8_t *)(pdst);  \
+    const uint32_t val_m = (val);         \
+                                          \
+    asm volatile (                        \
+        "sw  %[val_m],  %[pdst_m]  \n\t"  \
+                                          \
+        : [pdst_m] "=m" (*pdst_m)         \
+        : [val_m] "r" (val_m)             \
+    );                                    \
+}
+
+#define SD(val, pdst)                     \
+{                                         \
+    uint8_t *pdst_m = (uint8_t *)(pdst);  \
+    const uint64_t val_m = (val);         \
+                                          \
+    asm volatile (                        \
+        "sd  %[val_m],  %[pdst_m]  \n\t"  \
+                                          \
+        : [pdst_m] "=m" (*pdst_m)         \
+        : [val_m] "r" (val_m)             \
+    );                                    \
+}
+#else  // !(__mips_isa_rev >= 6)
+#define LW(psrc)                                      \
+({                                                    \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);  \
+    uint32_t val_m;                                   \
+                                                      \
+    asm volatile (                                    \
+        "ulw  %[val_m],  %[psrc_m]  \n\t"             \
+                                                      \
+        : [val_m] "=r" (val_m)                        \
+        : [psrc_m] "m" (*psrc_m)                      \
+    );                                                \
+                                                      \
+    val_m;                                            \
+})
+
+#if (__mips == 64)
+#define LD(psrc)                                      \
+({                                                    \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);  \
+    uint64_t val_m = 0;                               \
+                                                      \
+    asm volatile (                                    \
+        "uld  %[val_m],  %[psrc_m]  \n\t"             \
+                                                      \
+        : [val_m] "=r" (val_m)                        \
+        : [psrc_m] "m" (*psrc_m)                      \
+    );                                                \
+                                                      \
+    val_m;                                            \
+})
+#else  // !(__mips == 64)
+#define LD(psrc)                                             \
+({                                                           \
+    const uint8_t *psrc_m1 = (const uint8_t *)(psrc);        \
+    uint32_t val0_m, val1_m;                                 \
+    uint64_t val_m = 0;                                      \
+                                                             \
+    val0_m = LW(psrc_m1);                                    \
+    val1_m = LW(psrc_m1 + 4);                                \
+                                                             \
+    val_m = (uint64_t)(val1_m);                              \
+    val_m = (uint64_t)((val_m << 32) & 0xFFFFFFFF00000000);  \
+    val_m = (uint64_t)(val_m | (uint64_t)val0_m);            \
+                                                             \
+    val_m;                                                   \
+})
+#endif  // (__mips == 64)
+#define SH(val, pdst)                      \
+{                                          \
+    uint8_t *pdst_m = (uint8_t *)(pdst);   \
+    const uint16_t val_m = (val);          \
+                                           \
+    asm volatile (                         \
+        "ush  %[val_m],  %[pdst_m]  \n\t"  \
+                                           \
+        : [pdst_m] "=m" (*pdst_m)          \
+        : [val_m] "r" (val_m)              \
+    );                                     \
+}
+
+#define SW(val, pdst)                      \
+{                                          \
+    uint8_t *pdst_m = (uint8_t *)(pdst);   \
+    const uint32_t val_m = (val);          \
+                                           \
+    asm volatile (                         \
+        "usw  %[val_m],  %[pdst_m]  \n\t"  \
+                                           \
+        : [pdst_m] "=m" (*pdst_m)          \
+        : [val_m] "r" (val_m)              \
+    );                                     \
+}
+
+#define SD(val, pdst)                                         \
+{                                                             \
+    uint8_t *pdst_m1 = (uint8_t *)(pdst);                     \
+    uint32_t val0_m, val1_m;                                  \
+                                                              \
+    val0_m = (uint32_t)((val) & 0x00000000FFFFFFFF);          \
+    val1_m = (uint32_t)(((val) >> 32) & 0x00000000FFFFFFFF);  \
+                                                              \
+    SW(val0_m, pdst_m1);                                      \
+    SW(val1_m, pdst_m1 + 4);                                  \
+}
+#endif  // (__mips_isa_rev >= 6)
+
+/* Description : Load 4 words with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1, out2, out3
+   Details     : Load word in 'out0' from (psrc)
+                 Load word in 'out1' from (psrc + stride)
+                 Load word in 'out2' from (psrc + 2 * stride)
+                 Load word in 'out3' from (psrc + 3 * stride)
+*/
+#define LW4(psrc, stride, out0, out1, out2, out3)  \
+{                                                  \
+    out0 = LW((psrc));                             \
+    out1 = LW((psrc) + stride);                    \
+    out2 = LW((psrc) + 2 * stride);                \
+    out3 = LW((psrc) + 3 * stride);                \
+}
+
+/* Description : Load double words with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+   Details     : Load double word in 'out0' from (psrc)
+                 Load double word in 'out1' from (psrc + stride)
+*/
+#define LD2(psrc, stride, out0, out1)  \
+{                                      \
+    out0 = LD((psrc));                 \
+    out1 = LD((psrc) + stride);        \
+}
+#define LD4(psrc, stride, out0, out1, out2, out3)  \
+{                                                  \
+    LD2((psrc), stride, out0, out1);               \
+    LD2((psrc) + 2 * stride, stride, out2, out3);  \
+}
+
+/* Description : Store 4 words with stride
+   Arguments   : Inputs - in0, in1, in2, in3, pdst, stride
+   Details     : Store word from 'in0' to (pdst)
+                 Store word from 'in1' to (pdst + stride)
+                 Store word from 'in2' to (pdst + 2 * stride)
+                 Store word from 'in3' to (pdst + 3 * stride)
+*/
+#define SW4(in0, in1, in2, in3, pdst, stride)  \
+{                                              \
+    SW(in0, (pdst));                           \
+    SW(in1, (pdst) + stride);                  \
+    SW(in2, (pdst) + 2 * stride);              \
+    SW(in3, (pdst) + 3 * stride);              \
+}
+
+/* Description : Store 4 double words with stride
+   Arguments   : Inputs - in0, in1, in2, in3, pdst, stride
+   Details     : Store double word from 'in0' to (pdst)
+                 Store double word from 'in1' to (pdst + stride)
+                 Store double word from 'in2' to (pdst + 2 * stride)
+                 Store double word from 'in3' to (pdst + 3 * stride)
+*/
+#define SD4(in0, in1, in2, in3, pdst, stride)  \
+{                                              \
+    SD(in0, (pdst));                           \
+    SD(in1, (pdst) + stride);                  \
+    SD(in2, (pdst) + 2 * stride);              \
+    SD(in3, (pdst) + 3 * stride);              \
+}
+
+/* Description : Load vectors with 16 byte elements with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Load 16 byte elements in 'out0' from (psrc)
+                 Load 16 byte elements in 'out1' from (psrc + stride)
+*/
+#define LD_B2(RTYPE, psrc, stride, out0, out1)  \
+{                                               \
+    out0 = LD_B(RTYPE, (psrc));                 \
+    out1 = LD_B(RTYPE, (psrc) + stride);        \
+}
+#define LD_UB2(...) LD_B2(v16u8, __VA_ARGS__)
+#define LD_SB2(...) LD_B2(v16i8, __VA_ARGS__)
+
+#define LD_B3(RTYPE, psrc, stride, out0, out1, out2)  \
+{                                                     \
+    LD_B2(RTYPE, (psrc), stride, out0, out1);         \
+    out2 = LD_B(RTYPE, (psrc) + 2 * stride);          \
+}
+#define LD_UB3(...) LD_B3(v16u8, __VA_ARGS__)
+#define LD_SB3(...) LD_B3(v16i8, __VA_ARGS__)
+
+#define LD_B4(RTYPE, psrc, stride, out0, out1, out2, out3)   \
+{                                                            \
+    LD_B2(RTYPE, (psrc), stride, out0, out1);                \
+    LD_B2(RTYPE, (psrc) + 2 * stride , stride, out2, out3);  \
+}
+#define LD_UB4(...) LD_B4(v16u8, __VA_ARGS__)
+#define LD_SB4(...) LD_B4(v16i8, __VA_ARGS__)
+
+#define LD_B5(RTYPE, psrc, stride, out0, out1, out2, out3, out4)  \
+{                                                                 \
+    LD_B4(RTYPE, (psrc), stride, out0, out1, out2, out3);         \
+    out4 = LD_B(RTYPE, (psrc) + 4 * stride);                      \
+}
+#define LD_UB5(...) LD_B5(v16u8, __VA_ARGS__)
+#define LD_SB5(...) LD_B5(v16i8, __VA_ARGS__)
+
+#define LD_B8(RTYPE, psrc, stride,                                      \
+              out0, out1, out2, out3, out4, out5, out6, out7)           \
+{                                                                       \
+    LD_B4(RTYPE, (psrc), stride, out0, out1, out2, out3);               \
+    LD_B4(RTYPE, (psrc) + 4 * stride, stride, out4, out5, out6, out7);  \
+}
+#define LD_UB8(...) LD_B8(v16u8, __VA_ARGS__)
+#define LD_SB8(...) LD_B8(v16i8, __VA_ARGS__)
+
+/* Description : Load vectors with 8 halfword elements with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+   Details     : Load 8 halfword elements in 'out0' from (psrc)
+                 Load 8 halfword elements in 'out1' from (psrc + stride)
+*/
+#define LD_H2(RTYPE, psrc, stride, out0, out1)  \
+{                                               \
+    out0 = LD_H(RTYPE, (psrc));                 \
+    out1 = LD_H(RTYPE, (psrc) + (stride));      \
+}
+#define LD_SH2(...) LD_H2(v8i16, __VA_ARGS__)
+
+#define LD_H4(RTYPE, psrc, stride, out0, out1, out2, out3)  \
+{                                                           \
+    LD_H2(RTYPE, (psrc), stride, out0, out1);               \
+    LD_H2(RTYPE, (psrc) + 2 * stride, stride, out2, out3);  \
+}
+#define LD_SH4(...) LD_H4(v8i16, __VA_ARGS__)
+
+/* Description : Load 2 vectors of signed word elements with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+                 Return Type - signed word
+*/
+#define LD_SW2(psrc, stride, out0, out1)  \
+{                                         \
+    out0 = LD_SW((psrc));                 \
+    out1 = LD_SW((psrc) + stride);        \
+}
+
+/* Description : Store vectors of 16 byte elements with stride
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 16 byte elements from 'in0' to (pdst)
+                 Store 16 byte elements from 'in1' to (pdst + stride)
+*/
+#define ST_B2(RTYPE, in0, in1, pdst, stride)  \
+{                                             \
+    ST_B(RTYPE, in0, (pdst));                 \
+    ST_B(RTYPE, in1, (pdst) + stride);        \
+}
+#define ST_UB2(...) ST_B2(v16u8, __VA_ARGS__)
+
+#define ST_B4(RTYPE, in0, in1, in2, in3, pdst, stride)    \
+{                                                         \
+    ST_B2(RTYPE, in0, in1, (pdst), stride);               \
+    ST_B2(RTYPE, in2, in3, (pdst) + 2 * stride, stride);  \
+}
+#define ST_UB4(...) ST_B4(v16u8, __VA_ARGS__)
+#define ST_SB4(...) ST_B4(v16i8, __VA_ARGS__)
+
+#define ST_B8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,        \
+              pdst, stride)                                         \
+{                                                                   \
+    ST_B4(RTYPE, in0, in1, in2, in3, pdst, stride);                 \
+    ST_B4(RTYPE, in4, in5, in6, in7, (pdst) + 4 * stride, stride);  \
+}
+#define ST_UB8(...) ST_B8(v16u8, __VA_ARGS__)
+
+/* Description : Store vectors of 8 halfword elements with stride
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 8 halfword elements from 'in0' to (pdst)
+                 Store 8 halfword elements from 'in1' to (pdst + stride)
+*/
+#define ST_H2(RTYPE, in0, in1, pdst, stride)  \
+{                                             \
+    ST_H(RTYPE, in0, (pdst));                 \
+    ST_H(RTYPE, in1, (pdst) + stride);        \
+}
+#define ST_SH2(...) ST_H2(v8i16, __VA_ARGS__)
+
+/* Description : Store vectors of word elements with stride
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 4 word elements from 'in0' to (pdst)
+                 Store 4 word elements from 'in1' to (pdst + stride)
+*/
+#define ST_SW2(in0, in1, pdst, stride)  \
+{                                       \
+    ST_SW(in0, (pdst));                 \
+    ST_SW(in1, (pdst) + stride);        \
+}
+
+/* Description : Store 2x4 byte block to destination memory from input vector
+   Arguments   : Inputs - in, stidx, pdst, stride
+   Details     : Index 'stidx' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst)
+                 Index 'stidx+1' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst + stride)
+                 Index 'stidx+2' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst + 2 * stride)
+                 Index 'stidx+3' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst + 3 * stride)
+*/
+#define ST2x4_UB(in, stidx, pdst, stride)             \
+{                                                     \
+    uint16_t out0_m, out1_m, out2_m, out3_m;          \
+    uint8_t *pblk_2x4_m = (uint8_t *)(pdst);          \
+                                                      \
+    out0_m = __msa_copy_u_h((v8i16)in, (stidx));      \
+    out1_m = __msa_copy_u_h((v8i16)in, (stidx + 1));  \
+    out2_m = __msa_copy_u_h((v8i16)in, (stidx + 2));  \
+    out3_m = __msa_copy_u_h((v8i16)in, (stidx + 3));  \
+                                                      \
+    SH(out0_m, pblk_2x4_m);                           \
+    SH(out1_m, pblk_2x4_m + stride);                  \
+    SH(out2_m, pblk_2x4_m + 2 * stride);              \
+    SH(out3_m, pblk_2x4_m + 3 * stride);              \
+}
+
+/* Description : Store 4x4 byte block to destination memory from input vector
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : 'Idx0' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst)
+                 'Idx1' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst + stride)
+                 'Idx2' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst + 2 * stride)
+                 'Idx3' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst + 3 * stride)
+*/
+#define ST4x4_UB(in0, in1, idx0, idx1, idx2, idx3, pdst, stride)  \
+{                                                                 \
+    uint32_t out0_m, out1_m, out2_m, out3_m;                      \
+    uint8_t *pblk_4x4_m = (uint8_t *)(pdst);                      \
+                                                                  \
+    out0_m = __msa_copy_u_w((v4i32)in0, idx0);                    \
+    out1_m = __msa_copy_u_w((v4i32)in0, idx1);                    \
+    out2_m = __msa_copy_u_w((v4i32)in1, idx2);                    \
+    out3_m = __msa_copy_u_w((v4i32)in1, idx3);                    \
+                                                                  \
+    SW4(out0_m, out1_m, out2_m, out3_m, pblk_4x4_m, stride);      \
+}
+#define ST4x8_UB(in0, in1, pdst, stride)                            \
+{                                                                   \
+    uint8_t *pblk_4x8 = (uint8_t *)(pdst);                          \
+                                                                    \
+    ST4x4_UB(in0, in0, 0, 1, 2, 3, pblk_4x8, stride);               \
+    ST4x4_UB(in1, in1, 0, 1, 2, 3, pblk_4x8 + 4 * stride, stride);  \
+}
+
+/* Description : Store 8x1 byte block to destination memory from input vector
+   Arguments   : Inputs - in, pdst
+   Details     : Index 0 double word element from 'in' vector is copied to the
+                 GP register and stored to (pdst)
+*/
+#define ST8x1_UB(in, pdst)                  \
+{                                           \
+    uint64_t out0_m;                        \
+                                            \
+    out0_m = __msa_copy_u_d((v2i64)in, 0);  \
+    SD(out0_m, pdst);                       \
+}
+
+/* Description : Store 8x2 byte block to destination memory from input vector
+   Arguments   : Inputs - in, pdst, stride
+   Details     : Index 0 double word element from 'in' vector is copied to the
+                 GP register and stored to (pdst)
+                 Index 1 double word element from 'in' vector is copied to the
+                 GP register and stored to (pdst + stride)
+*/
+#define ST8x2_UB(in, pdst, stride)            \
+{                                             \
+    uint64_t out0_m, out1_m;                  \
+    uint8_t *pblk_8x2_m = (uint8_t *)(pdst);  \
+                                              \
+    out0_m = __msa_copy_u_d((v2i64)in, 0);    \
+    out1_m = __msa_copy_u_d((v2i64)in, 1);    \
+                                              \
+    SD(out0_m, pblk_8x2_m);                   \
+    SD(out1_m, pblk_8x2_m + stride);          \
+}
+
+/* Description : Store 8x4 byte block to destination memory from input
+                 vectors
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Index 0 double word element from 'in0' vector is copied to the
+                 GP register and stored to (pdst)
+                 Index 1 double word element from 'in0' vector is copied to the
+                 GP register and stored to (pdst + stride)
+                 Index 0 double word element from 'in1' vector is copied to the
+                 GP register and stored to (pdst + 2 * stride)
+                 Index 1 double word element from 'in1' vector is copied to the
+                 GP register and stored to (pdst + 3 * stride)
+*/
+#define ST8x4_UB(in0, in1, pdst, stride)                      \
+{                                                             \
+    uint64_t out0_m, out1_m, out2_m, out3_m;                  \
+    uint8_t *pblk_8x4_m = (uint8_t *)(pdst);                  \
+                                                              \
+    out0_m = __msa_copy_u_d((v2i64)in0, 0);                   \
+    out1_m = __msa_copy_u_d((v2i64)in0, 1);                   \
+    out2_m = __msa_copy_u_d((v2i64)in1, 0);                   \
+    out3_m = __msa_copy_u_d((v2i64)in1, 1);                   \
+                                                              \
+    SD4(out0_m, out1_m, out2_m, out3_m, pblk_8x4_m, stride);  \
+}
+
+/* Description : Immediate number of elements to slide with zero
+   Arguments   : Inputs  - in0, in1, slide_val
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'zero_m' vector are slid into 'in0' by
+                 value specified in the 'slide_val'
+*/
+#define SLDI_B2_0(RTYPE, in0, in1, out0, out1, slide_val)              \
+{                                                                      \
+    v16i8 zero_m = { 0 };                                              \
+                                                                       \
+    out0 = (RTYPE)__msa_sldi_b((v16i8)zero_m, (v16i8)in0, slide_val);  \
+    out1 = (RTYPE)__msa_sldi_b((v16i8)zero_m, (v16i8)in1, slide_val);  \
+}
+#define SLDI_B2_0_UB(...) SLDI_B2_0(v16u8, __VA_ARGS__)
+
+/* Description : Immediate number of elements to slide
+   Arguments   : Inputs  - in0_0, in0_1, in1_0, in1_1, slide_val
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'in0_0' vector are slid into 'in1_0' by
+                 value specified in the 'slide_val'
+*/
+#define SLDI_B2(RTYPE, in0_0, in0_1, in1_0, in1_1, out0, out1, slide_val)  \
+{                                                                          \
+    out0 = (RTYPE)__msa_sldi_b((v16i8)in0_0, (v16i8)in1_0, slide_val);     \
+    out1 = (RTYPE)__msa_sldi_b((v16i8)in0_1, (v16i8)in1_1, slide_val);     \
+}
+
+#define SLDI_B3(RTYPE, in0_0, in0_1, in0_2, in1_0, in1_1, in1_2,        \
+                out0, out1, out2, slide_val)                            \
+{                                                                       \
+    SLDI_B2(RTYPE, in0_0, in0_1, in1_0, in1_1, out0, out1, slide_val);  \
+    out2 = (RTYPE)__msa_sldi_b((v16i8)in0_2, (v16i8)in1_2, slide_val);  \
+}
+#define SLDI_B3_UH(...) SLDI_B3(v8u16, __VA_ARGS__)
+
+/* Description : Shuffle byte vector elements as per mask vector
+   Arguments   : Inputs  - in0, in1, in2, in3, mask0, mask1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'in0' & 'in1' are copied selectively to
+                 'out0' as per control vector 'mask0'
+*/
+#define VSHF_B2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1)   \
+{                                                                      \
+    out0 = (RTYPE)__msa_vshf_b((v16i8)mask0, (v16i8)in1, (v16i8)in0);  \
+    out1 = (RTYPE)__msa_vshf_b((v16i8)mask1, (v16i8)in3, (v16i8)in2);  \
+}
+#define VSHF_B2_UB(...) VSHF_B2(v16u8, __VA_ARGS__)
+#define VSHF_B2_SB(...) VSHF_B2(v16i8, __VA_ARGS__)
+#define VSHF_B2_UH(...) VSHF_B2(v8u16, __VA_ARGS__)
+
+#define VSHF_B3(RTYPE, in0, in1, in2, in3, in4, in5, mask0, mask1, mask2,  \
+                out0, out1, out2)                                          \
+{                                                                          \
+    VSHF_B2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1);          \
+    out2 = (RTYPE)__msa_vshf_b((v16i8)mask2, (v16i8)in5, (v16i8)in4);      \
+}
+#define VSHF_B3_SB(...) VSHF_B3(v16i8, __VA_ARGS__)
+
+/* Description : Shuffle halfword vector elements as per mask vector
+   Arguments   : Inputs  - in0, in1, in2, in3, mask0, mask1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : halfword elements from 'in0' & 'in1' are copied selectively to
+                 'out0' as per control vector 'mask0'
+*/
+#define VSHF_H2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1)   \
+{                                                                      \
+    out0 = (RTYPE)__msa_vshf_h((v8i16)mask0, (v8i16)in1, (v8i16)in0);  \
+    out1 = (RTYPE)__msa_vshf_h((v8i16)mask1, (v8i16)in3, (v8i16)in2);  \
+}
+#define VSHF_H2_SH(...) VSHF_H2(v8i16, __VA_ARGS__)
+
+/* Description : Dot product of byte vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Unsigned byte elements from 'mult0' are multiplied with
+                 unsigned byte elements from 'cnst0' producing a result
+                 twice the size of input i.e. unsigned halfword.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_UB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1)  \
+{                                                                \
+    out0 = (RTYPE)__msa_dotp_u_h((v16u8)mult0, (v16u8)cnst0);    \
+    out1 = (RTYPE)__msa_dotp_u_h((v16u8)mult1, (v16u8)cnst1);    \
+}
+#define DOTP_UB2_UH(...) DOTP_UB2(v8u16, __VA_ARGS__)
+
+#define DOTP_UB4(RTYPE, mult0, mult1, mult2, mult3,           \
+                 cnst0, cnst1, cnst2, cnst3,                  \
+                 out0, out1, out2, out3)                      \
+{                                                             \
+    DOTP_UB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);  \
+    DOTP_UB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);  \
+}
+#define DOTP_UB4_UH(...) DOTP_UB4(v8u16, __VA_ARGS__)
+
+/* Description : Dot product of byte vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed byte elements from 'mult0' are multiplied with
+                 signed byte elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed halfword.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1)  \
+{                                                                \
+    out0 = (RTYPE)__msa_dotp_s_h((v16i8)mult0, (v16i8)cnst0);    \
+    out1 = (RTYPE)__msa_dotp_s_h((v16i8)mult1, (v16i8)cnst1);    \
+}
+#define DOTP_SB2_SH(...) DOTP_SB2(v8i16, __VA_ARGS__)
+
+#define DOTP_SB4(RTYPE, mult0, mult1, mult2, mult3,                   \
+                 cnst0, cnst1, cnst2, cnst3, out0, out1, out2, out3)  \
+{                                                                     \
+    DOTP_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);          \
+    DOTP_SB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);          \
+}
+#define DOTP_SB4_SH(...) DOTP_SB4(v8i16, __VA_ARGS__)
+
+/* Description : Dot product of halfword vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed halfword elements from 'mult0' are multiplied with
+                 signed halfword elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed word.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1)  \
+{                                                                \
+    out0 = (RTYPE)__msa_dotp_s_w((v8i16)mult0, (v8i16)cnst0);    \
+    out1 = (RTYPE)__msa_dotp_s_w((v8i16)mult1, (v8i16)cnst1);    \
+}
+
+#define DOTP_SH4(RTYPE, mult0, mult1, mult2, mult3,           \
+                 cnst0, cnst1, cnst2, cnst3,                  \
+                 out0, out1, out2, out3)                      \
+{                                                             \
+    DOTP_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);  \
+    DOTP_SH2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);  \
+}
+#define DOTP_SH4_SW(...) DOTP_SH4(v4i32, __VA_ARGS__)
+
+/* Description : Dot product of word vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed word elements from 'mult0' are multiplied with
+                 signed word elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed double word.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_SW2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1)  \
+{                                                                \
+    out0 = (RTYPE)__msa_dotp_s_d((v4i32)mult0, (v4i32)cnst0);    \
+    out1 = (RTYPE)__msa_dotp_s_d((v4i32)mult1, (v4i32)cnst1);    \
+}
+#define DOTP_SW2_SD(...) DOTP_SW2(v2i64, __VA_ARGS__)
+
+/* Description : Dot product & addition of byte vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed byte elements from 'mult0' are multiplied with
+                 signed byte elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed halfword.
+                 The multiplication result of adjacent odd-even elements
+                 are added to the 'out0' vector
+*/
+#define DPADD_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1)             \
+{                                                                            \
+    out0 = (RTYPE)__msa_dpadd_s_h((v8i16)out0, (v16i8)mult0, (v16i8)cnst0);  \
+    out1 = (RTYPE)__msa_dpadd_s_h((v8i16)out1, (v16i8)mult1, (v16i8)cnst1);  \
+}
+#define DPADD_SB2_SH(...) DPADD_SB2(v8i16, __VA_ARGS__)
+
+#define DPADD_SB4(RTYPE, mult0, mult1, mult2, mult3,                   \
+                  cnst0, cnst1, cnst2, cnst3, out0, out1, out2, out3)  \
+{                                                                      \
+    DPADD_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);          \
+    DPADD_SB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);          \
+}
+#define DPADD_SB4_SH(...) DPADD_SB4(v8i16, __VA_ARGS__)
+
+/* Description : Dot product & addition of halfword vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed halfword elements from 'mult0' are multiplied with
+                 signed halfword elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed word.
+                 The multiplication result of adjacent odd-even elements
+                 are added to the 'out0' vector
+*/
+#define DPADD_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1)             \
+{                                                                            \
+    out0 = (RTYPE)__msa_dpadd_s_w((v4i32)out0, (v8i16)mult0, (v8i16)cnst0);  \
+    out1 = (RTYPE)__msa_dpadd_s_w((v4i32)out1, (v8i16)mult1, (v8i16)cnst1);  \
+}
+#define DPADD_SH2_SW(...) DPADD_SH2(v4i32, __VA_ARGS__)
+
+#define DPADD_SH4(RTYPE, mult0, mult1, mult2, mult3,                   \
+                  cnst0, cnst1, cnst2, cnst3, out0, out1, out2, out3)  \
+{                                                                      \
+    DPADD_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);          \
+    DPADD_SH2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);          \
+}
+#define DPADD_SH4_SW(...) DPADD_SH4(v4i32, __VA_ARGS__)
+
+/* Description : Dot product & addition of double word vector elements
+   Arguments   : Inputs  - mult0, mult1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each signed word element from 'mult0' is multiplied with itself
+                 producing an intermediate result twice the size of it
+                 i.e. signed double word
+                 The multiplication result of adjacent odd-even elements
+                 are added to the 'out0' vector
+*/
+#define DPADD_SD2(RTYPE, mult0, mult1, out0, out1)                           \
+{                                                                            \
+    out0 = (RTYPE)__msa_dpadd_s_d((v2i64)out0, (v4i32)mult0, (v4i32)mult0);  \
+    out1 = (RTYPE)__msa_dpadd_s_d((v2i64)out1, (v4i32)mult1, (v4i32)mult1);  \
+}
+#define DPADD_SD2_SD(...) DPADD_SD2(v2i64, __VA_ARGS__)
+
+/* Description : Clips all signed halfword elements of input vector
+                 between 0 & 255
+   Arguments   : Input  - in
+                 Output - out_m
+                 Return Type - signed halfword
+*/
+#define CLIP_SH_0_255(in)                               \
+({                                                      \
+    v8i16 max_m = __msa_ldi_h(255);                     \
+    v8i16 out_m;                                        \
+                                                        \
+    out_m = __msa_maxi_s_h((v8i16)in, 0);               \
+    out_m = __msa_min_s_h((v8i16)max_m, (v8i16)out_m);  \
+    out_m;                                              \
+})
+#define CLIP_SH2_0_255(in0, in1)  \
+{                                 \
+    in0 = CLIP_SH_0_255(in0);     \
+    in1 = CLIP_SH_0_255(in1);     \
+}
+#define CLIP_SH4_0_255(in0, in1, in2, in3)  \
+{                                           \
+    CLIP_SH2_0_255(in0, in1);               \
+    CLIP_SH2_0_255(in2, in3);               \
+}
+
+/* Description : Clips all signed word elements of input vector
+                 between 0 & 255
+   Arguments   : Input  - in
+                 Output - out_m
+                 Return Type - signed word
+*/
+#define CLIP_SW_0_255(in)                               \
+({                                                      \
+    v4i32 max_m = __msa_ldi_w(255);                     \
+    v4i32 out_m;                                        \
+                                                        \
+    out_m = __msa_maxi_s_w((v4i32)in, 0);               \
+    out_m = __msa_min_s_w((v4i32)max_m, (v4i32)out_m);  \
+    out_m;                                              \
+})
+
+/* Description : Horizontal addition of 4 signed word elements of input vector
+   Arguments   : Input  - in       (signed word vector)
+                 Output - sum_m    (i32 sum)
+                 Return Type - signed word (GP)
+   Details     : 4 signed word elements of 'in' vector are added together and
+                 the resulting integer sum is returned
+*/
+#define HADD_SW_S32(in)                             \
+({                                                  \
+    v2i64 res0_m, res1_m;                           \
+    int32_t sum_m;                                  \
+                                                    \
+    res0_m = __msa_hadd_s_d((v4i32)in, (v4i32)in);  \
+    res1_m = __msa_splati_d(res0_m, 1);             \
+    res0_m = res0_m + res1_m;                       \
+    sum_m = __msa_copy_s_w((v4i32)res0_m, 0);       \
+    sum_m;                                          \
+})
+
+/* Description : Horizontal addition of 8 unsigned halfword elements
+   Arguments   : Inputs  - in       (unsigned halfword vector)
+                 Outputs - sum_m    (u32 sum)
+                 Return Type - unsigned word
+   Details     : 8 unsigned halfword elements of input vector are added
+                 together and the resulting integer sum is returned
+*/
+#define HADD_UH_U32(in)                                \
+({                                                     \
+    v4u32 res_m;                                       \
+    v2u64 res0_m, res1_m;                              \
+    uint32_t sum_m;                                    \
+                                                       \
+    res_m = __msa_hadd_u_w((v8u16)in, (v8u16)in);      \
+    res0_m = __msa_hadd_u_d(res_m, res_m);             \
+    res1_m = (v2u64)__msa_splati_d((v2i64)res0_m, 1);  \
+    res0_m = res0_m + res1_m;                          \
+    sum_m = __msa_copy_u_w((v4i32)res0_m, 0);          \
+    sum_m;                                             \
+})
+
+/* Description : Horizontal addition of unsigned byte vector elements
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each unsigned odd byte element from 'in0' is added to
+                 even unsigned byte element from 'in0' (pairwise) and the
+                 halfword result is written to 'out0'
+*/
+#define HADD_UB2(RTYPE, in0, in1, out0, out1)              \
+{                                                          \
+    out0 = (RTYPE)__msa_hadd_u_h((v16u8)in0, (v16u8)in0);  \
+    out1 = (RTYPE)__msa_hadd_u_h((v16u8)in1, (v16u8)in1);  \
+}
+#define HADD_UB2_UH(...) HADD_UB2(v8u16, __VA_ARGS__)
+
+/* Description : Horizontal subtraction of unsigned byte vector elements
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each unsigned odd byte element from 'in0' is subtracted from
+                 even unsigned byte element from 'in0' (pairwise) and the
+                 halfword result is written to 'out0'
+*/
+#define HSUB_UB2(RTYPE, in0, in1, out0, out1)              \
+{                                                          \
+    out0 = (RTYPE)__msa_hsub_u_h((v16u8)in0, (v16u8)in0);  \
+    out1 = (RTYPE)__msa_hsub_u_h((v16u8)in1, (v16u8)in1);  \
+}
+#define HSUB_UB2_SH(...) HSUB_UB2(v8i16, __VA_ARGS__)
+
+/* Description : Horizontal subtraction of signed halfword vector elements
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each signed odd halfword element from 'in0' is subtracted from
+                 even signed halfword element from 'in0' (pairwise) and the
+                 word result is written to 'out0'
+*/
+#define HSUB_UH2(RTYPE, in0, in1, out0, out1)              \
+{                                                          \
+    out0 = (RTYPE)__msa_hsub_s_w((v8i16)in0, (v8i16)in0);  \
+    out1 = (RTYPE)__msa_hsub_s_w((v8i16)in1, (v8i16)in1);  \
+}
+#define HSUB_UH2_SW(...) HSUB_UH2(v4i32, __VA_ARGS__)
+
+/* Description : Set element n input vector to GPR value
+   Arguments   : Inputs - in0, in1, in2, in3
+                 Output - out
+                 Return Type - as per RTYPE
+   Details     : Set element 0 in vector 'out' to value specified in 'in0'
+*/
+#define INSERT_D2(RTYPE, in0, in1, out)               \
+{                                                     \
+    out = (RTYPE)__msa_insert_d((v2i64)out, 0, in0);  \
+    out = (RTYPE)__msa_insert_d((v2i64)out, 1, in1);  \
+}
+#define INSERT_D2_SB(...) INSERT_D2(v16i8, __VA_ARGS__)
+
+/* Description : Interleave even byte elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even byte elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_B2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                         \
+    out0 = (RTYPE)__msa_ilvev_b((v16i8)in1, (v16i8)in0);  \
+    out1 = (RTYPE)__msa_ilvev_b((v16i8)in3, (v16i8)in2);  \
+}
+#define ILVEV_B2_UB(...) ILVEV_B2(v16u8, __VA_ARGS__)
+#define ILVEV_B2_SH(...) ILVEV_B2(v8i16, __VA_ARGS__)
+#define ILVEV_B2_SD(...) ILVEV_B2(v2i64, __VA_ARGS__)
+
+/* Description : Interleave even halfword elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even halfword elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_H2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                         \
+    out0 = (RTYPE)__msa_ilvev_h((v8i16)in1, (v8i16)in0);  \
+    out1 = (RTYPE)__msa_ilvev_h((v8i16)in3, (v8i16)in2);  \
+}
+#define ILVEV_H2_UB(...) ILVEV_H2(v16u8, __VA_ARGS__)
+#define ILVEV_H2_SH(...) ILVEV_H2(v8i16, __VA_ARGS__)
+
+/* Description : Interleave even word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even word elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_W2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                         \
+    out0 = (RTYPE)__msa_ilvev_w((v4i32)in1, (v4i32)in0);  \
+    out1 = (RTYPE)__msa_ilvev_w((v4i32)in3, (v4i32)in2);  \
+}
+#define ILVEV_W2_SD(...) ILVEV_W2(v2i64, __VA_ARGS__)
+
+/* Description : Interleave even double word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even double word elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_D2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                         \
+    out0 = (RTYPE)__msa_ilvev_d((v2i64)in1, (v2i64)in0);  \
+    out1 = (RTYPE)__msa_ilvev_d((v2i64)in3, (v2i64)in2);  \
+}
+#define ILVEV_D2_UB(...) ILVEV_D2(v16u8, __VA_ARGS__)
+
+/* Description : Interleave left half of byte elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Left half of byte elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'.
+*/
+#define ILVL_B2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                        \
+    out0 = (RTYPE)__msa_ilvl_b((v16i8)in0, (v16i8)in1);  \
+    out1 = (RTYPE)__msa_ilvl_b((v16i8)in2, (v16i8)in3);  \
+}
+#define ILVL_B2_UB(...) ILVL_B2(v16u8, __VA_ARGS__)
+#define ILVL_B2_SB(...) ILVL_B2(v16i8, __VA_ARGS__)
+#define ILVL_B2_SH(...) ILVL_B2(v8i16, __VA_ARGS__)
+
+#define ILVL_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3)                         \
+{                                                               \
+    ILVL_B2(RTYPE, in0, in1, in2, in3, out0, out1);             \
+    ILVL_B2(RTYPE, in4, in5, in6, in7, out2, out3);             \
+}
+#define ILVL_B4_SB(...) ILVL_B4(v16i8, __VA_ARGS__)
+#define ILVL_B4_SH(...) ILVL_B4(v8i16, __VA_ARGS__)
+
+/* Description : Interleave left half of halfword elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Left half of halfword elements of 'in0' and 'in1' are
+                 interleaved and written to 'out0'.
+*/
+#define ILVL_H2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                        \
+    out0 = (RTYPE)__msa_ilvl_h((v8i16)in0, (v8i16)in1);  \
+    out1 = (RTYPE)__msa_ilvl_h((v8i16)in2, (v8i16)in3);  \
+}
+#define ILVL_H2_SH(...) ILVL_H2(v8i16, __VA_ARGS__)
+#define ILVL_H2_SW(...) ILVL_H2(v4i32, __VA_ARGS__)
+
+/* Description : Interleave left half of word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Left half of word elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'.
+*/
+#define ILVL_W2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                        \
+    out0 = (RTYPE)__msa_ilvl_w((v4i32)in0, (v4i32)in1);  \
+    out1 = (RTYPE)__msa_ilvl_w((v4i32)in2, (v4i32)in3);  \
+}
+#define ILVL_W2_SH(...) ILVL_W2(v8i16, __VA_ARGS__)
+
+/* Description : Interleave right half of byte elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of byte elements of 'in0' and 'in1' are interleaved
+                 and written to out0.
+*/
+#define ILVR_B2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                        \
+    out0 = (RTYPE)__msa_ilvr_b((v16i8)in0, (v16i8)in1);  \
+    out1 = (RTYPE)__msa_ilvr_b((v16i8)in2, (v16i8)in3);  \
+}
+#define ILVR_B2_UB(...) ILVR_B2(v16u8, __VA_ARGS__)
+#define ILVR_B2_SB(...) ILVR_B2(v16i8, __VA_ARGS__)
+#define ILVR_B2_SH(...) ILVR_B2(v8i16, __VA_ARGS__)
+#define ILVR_B2_SW(...) ILVR_B2(v4i32, __VA_ARGS__)
+
+#define ILVR_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3)                         \
+{                                                               \
+    ILVR_B2(RTYPE, in0, in1, in2, in3, out0, out1);             \
+    ILVR_B2(RTYPE, in4, in5, in6, in7, out2, out3);             \
+}
+#define ILVR_B4_UB(...) ILVR_B4(v16u8, __VA_ARGS__)
+#define ILVR_B4_SB(...) ILVR_B4(v16i8, __VA_ARGS__)
+#define ILVR_B4_UH(...) ILVR_B4(v8u16, __VA_ARGS__)
+#define ILVR_B4_SH(...) ILVR_B4(v8i16, __VA_ARGS__)
+#define ILVR_B4_SW(...) ILVR_B4(v4i32, __VA_ARGS__)
+
+/* Description : Interleave right half of halfword elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of halfword elements of 'in0' and 'in1' are
+                 interleaved and written to 'out0'.
+*/
+#define ILVR_H2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                        \
+    out0 = (RTYPE)__msa_ilvr_h((v8i16)in0, (v8i16)in1);  \
+    out1 = (RTYPE)__msa_ilvr_h((v8i16)in2, (v8i16)in3);  \
+}
+#define ILVR_H2_SH(...) ILVR_H2(v8i16, __VA_ARGS__)
+#define ILVR_H2_SW(...) ILVR_H2(v4i32, __VA_ARGS__)
+
+#define ILVR_H4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3)                         \
+{                                                               \
+    ILVR_H2(RTYPE, in0, in1, in2, in3, out0, out1);             \
+    ILVR_H2(RTYPE, in4, in5, in6, in7, out2, out3);             \
+}
+#define ILVR_H4_SH(...) ILVR_H4(v8i16, __VA_ARGS__)
+#define ILVR_H4_SW(...) ILVR_H4(v4i32, __VA_ARGS__)
+
+#define ILVR_W2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                        \
+    out0 = (RTYPE)__msa_ilvr_w((v4i32)in0, (v4i32)in1);  \
+    out1 = (RTYPE)__msa_ilvr_w((v4i32)in2, (v4i32)in3);  \
+}
+#define ILVR_W2_SH(...) ILVR_W2(v8i16, __VA_ARGS__)
+
+/* Description : Interleave right half of double word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of double word elements of 'in0' and 'in1' are
+                 interleaved and written to 'out0'.
+*/
+#define ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1)       \
+{                                                            \
+    out0 = (RTYPE)__msa_ilvr_d((v2i64)(in0), (v2i64)(in1));  \
+    out1 = (RTYPE)__msa_ilvr_d((v2i64)(in2), (v2i64)(in3));  \
+}
+#define ILVR_D2_UB(...) ILVR_D2(v16u8, __VA_ARGS__)
+#define ILVR_D2_SB(...) ILVR_D2(v16i8, __VA_ARGS__)
+#define ILVR_D2_SH(...) ILVR_D2(v8i16, __VA_ARGS__)
+
+#define ILVR_D4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3)                         \
+{                                                               \
+    ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1);             \
+    ILVR_D2(RTYPE, in4, in5, in6, in7, out2, out3);             \
+}
+#define ILVR_D4_SB(...) ILVR_D4(v16i8, __VA_ARGS__)
+#define ILVR_D4_UB(...) ILVR_D4(v16u8, __VA_ARGS__)
+
+/* Description : Interleave both left and right half of input vectors
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of byte elements from 'in0' and 'in1' are
+                 interleaved and written to 'out0'
+*/
+#define ILVRL_B2(RTYPE, in0, in1, out0, out1)            \
+{                                                        \
+    out0 = (RTYPE)__msa_ilvr_b((v16i8)in0, (v16i8)in1);  \
+    out1 = (RTYPE)__msa_ilvl_b((v16i8)in0, (v16i8)in1);  \
+}
+#define ILVRL_B2_UB(...) ILVRL_B2(v16u8, __VA_ARGS__)
+#define ILVRL_B2_SB(...) ILVRL_B2(v16i8, __VA_ARGS__)
+#define ILVRL_B2_UH(...) ILVRL_B2(v8u16, __VA_ARGS__)
+#define ILVRL_B2_SH(...) ILVRL_B2(v8i16, __VA_ARGS__)
+
+#define ILVRL_H2(RTYPE, in0, in1, out0, out1)            \
+{                                                        \
+    out0 = (RTYPE)__msa_ilvr_h((v8i16)in0, (v8i16)in1);  \
+    out1 = (RTYPE)__msa_ilvl_h((v8i16)in0, (v8i16)in1);  \
+}
+#define ILVRL_H2_SH(...) ILVRL_H2(v8i16, __VA_ARGS__)
+#define ILVRL_H2_SW(...) ILVRL_H2(v4i32, __VA_ARGS__)
+
+#define ILVRL_W2(RTYPE, in0, in1, out0, out1)            \
+{                                                        \
+    out0 = (RTYPE)__msa_ilvr_w((v4i32)in0, (v4i32)in1);  \
+    out1 = (RTYPE)__msa_ilvl_w((v4i32)in0, (v4i32)in1);  \
+}
+#define ILVRL_W2_UB(...) ILVRL_W2(v16u8, __VA_ARGS__)
+#define ILVRL_W2_SH(...) ILVRL_W2(v8i16, __VA_ARGS__)
+#define ILVRL_W2_SW(...) ILVRL_W2(v4i32, __VA_ARGS__)
+
+/* Description : Maximum values between signed elements of vector and
+                 5-bit signed immediate value are copied to the output vector
+   Arguments   : Inputs  - in0, in1, in2, in3, max_val
+                 Outputs - in place operation
+                 Return Type - unsigned halfword
+   Details     : Maximum of signed halfword element values from 'in0' and
+                 'max_val' are written in place
+*/
+#define MAXI_SH2(RTYPE, in0, in1, max_val)               \
+{                                                        \
+    in0 = (RTYPE)__msa_maxi_s_h((v8i16)in0, (max_val));  \
+    in1 = (RTYPE)__msa_maxi_s_h((v8i16)in1, (max_val));  \
+}
+#define MAXI_SH2_SH(...) MAXI_SH2(v8i16, __VA_ARGS__)
+
+/* Description : Saturate the halfword element values to the max
+                 unsigned value of (sat_val + 1) bits
+                 The element data width remains unchanged
+   Arguments   : Inputs  - in0, in1, sat_val
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each unsigned halfword element from 'in0' is saturated to the
+                 value generated with (sat_val + 1) bit range.
+                 The results are written in place
+*/
+#define SAT_UH2(RTYPE, in0, in1, sat_val)             \
+{                                                     \
+    in0 = (RTYPE)__msa_sat_u_h((v8u16)in0, sat_val);  \
+    in1 = (RTYPE)__msa_sat_u_h((v8u16)in1, sat_val);  \
+}
+#define SAT_UH2_SH(...) SAT_UH2(v8i16, __VA_ARGS__)
+
+/* Description : Saturate the halfword element values to the max
+                 unsigned value of (sat_val + 1) bits
+                 The element data width remains unchanged
+   Arguments   : Inputs  - in0, in1, sat_val
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each unsigned halfword element from 'in0' is saturated to the
+                 value generated with (sat_val + 1) bit range
+                 The results are written in place
+*/
+#define SAT_SH2(RTYPE, in0, in1, sat_val)             \
+{                                                     \
+    in0 = (RTYPE)__msa_sat_s_h((v8i16)in0, sat_val);  \
+    in1 = (RTYPE)__msa_sat_s_h((v8i16)in1, sat_val);  \
+}
+#define SAT_SH2_SH(...) SAT_SH2(v8i16, __VA_ARGS__)
+
+#define SAT_SH4(RTYPE, in0, in1, in2, in3, sat_val)  \
+{                                                    \
+    SAT_SH2(RTYPE, in0, in1, sat_val);               \
+    SAT_SH2(RTYPE, in2, in3, sat_val);               \
+}
+#define SAT_SH4_SH(...) SAT_SH4(v8i16, __VA_ARGS__)
+
+/* Description : Indexed halfword element values are replicated to all
+                 elements in output vector
+   Arguments   : Inputs  - in, idx0, idx1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : 'idx0' element value from 'in' vector is replicated to all
+                  elements in 'out0' vector
+                  Valid index range for halfword operation is 0-7
+*/
+#define SPLATI_H2(RTYPE, in, idx0, idx1, out0, out1)  \
+{                                                     \
+    out0 = (RTYPE)__msa_splati_h((v8i16)in, idx0);    \
+    out1 = (RTYPE)__msa_splati_h((v8i16)in, idx1);    \
+}
+#define SPLATI_H2_SB(...) SPLATI_H2(v16i8, __VA_ARGS__)
+#define SPLATI_H2_SH(...) SPLATI_H2(v8i16, __VA_ARGS__)
+
+#define SPLATI_H3(RTYPE, in, idx0, idx1, idx2,      \
+                  out0, out1, out2)                 \
+{                                                   \
+    SPLATI_H2(RTYPE, in, idx0, idx1, out0, out1);   \
+    out2 = (RTYPE)__msa_splati_h((v8i16)in, idx2);  \
+}
+#define SPLATI_H3_SB(...) SPLATI_H3(v16i8, __VA_ARGS__)
+#define SPLATI_H3_SH(...) SPLATI_H3(v8i16, __VA_ARGS__)
+
+/* Description : Indexed word element values are replicated to all
+                 elements in output vector
+   Arguments   : Inputs  - in, stidx
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : 'stidx' element value from 'in' vector is replicated to all
+                 elements in 'out0' vector
+                 'stidx + 1' element value from 'in' vector is replicated to all
+                 elements in 'out1' vector
+                 Valid index range for word operation is 0-3
+*/
+#define SPLATI_W2(RTYPE, in, stidx, out0, out1)          \
+{                                                        \
+    out0 = (RTYPE)__msa_splati_w((v4i32)in, stidx);      \
+    out1 = (RTYPE)__msa_splati_w((v4i32)in, (stidx+1));  \
+}
+#define SPLATI_W2_SW(...) SPLATI_W2(v4i32, __VA_ARGS__)
+
+/* Description : Pack even byte elements of vector pairs
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even byte elements of 'in0' are copied to the left half of
+                 'out0' & even byte elements of 'in1' are copied to the right
+                 half of 'out0'.
+*/
+#define PCKEV_B2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                         \
+    out0 = (RTYPE)__msa_pckev_b((v16i8)in0, (v16i8)in1);  \
+    out1 = (RTYPE)__msa_pckev_b((v16i8)in2, (v16i8)in3);  \
+}
+#define PCKEV_B2_SB(...) PCKEV_B2(v16i8, __VA_ARGS__)
+#define PCKEV_B2_UB(...) PCKEV_B2(v16u8, __VA_ARGS__)
+
+#define PCKEV_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                 out0, out1, out2, out3)                         \
+{                                                                \
+    PCKEV_B2(RTYPE, in0, in1, in2, in3, out0, out1);             \
+    PCKEV_B2(RTYPE, in4, in5, in6, in7, out2, out3);             \
+}
+#define PCKEV_B4_SB(...) PCKEV_B4(v16i8, __VA_ARGS__)
+#define PCKEV_B4_UB(...) PCKEV_B4(v16u8, __VA_ARGS__)
+#define PCKEV_B4_SH(...) PCKEV_B4(v8i16, __VA_ARGS__)
+
+/* Description : Pack even halfword elements of vector pairs
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even halfword elements of 'in0' are copied to the left half of
+                 'out0' & even halfword elements of 'in1' are copied to the
+                 right half of 'out0'.
+*/
+#define PCKEV_H2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                         \
+    out0 = (RTYPE)__msa_pckev_h((v8i16)in0, (v8i16)in1);  \
+    out1 = (RTYPE)__msa_pckev_h((v8i16)in2, (v8i16)in3);  \
+}
+#define PCKEV_H2_SH(...) PCKEV_H2(v8i16, __VA_ARGS__)
+
+#define PCKEV_H4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                 out0, out1, out2, out3)                         \
+{                                                                \
+    PCKEV_H2(RTYPE, in0, in1, in2, in3, out0, out1);             \
+    PCKEV_H2(RTYPE, in4, in5, in6, in7, out2, out3);             \
+}
+#define PCKEV_H4_SH(...) PCKEV_H4(v8i16, __VA_ARGS__)
+
+/* Description : Pack even double word elements of vector pairs
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even double elements of 'in0' are copied to the left half of
+                 'out0' & even double elements of 'in1' are copied to the right
+                 half of 'out0'.
+*/
+#define PCKEV_D2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                         \
+    out0 = (RTYPE)__msa_pckev_d((v2i64)in0, (v2i64)in1);  \
+    out1 = (RTYPE)__msa_pckev_d((v2i64)in2, (v2i64)in3);  \
+}
+#define PCKEV_D2_UB(...) PCKEV_D2(v16u8, __VA_ARGS__)
+#define PCKEV_D2_SH(...) PCKEV_D2(v8i16, __VA_ARGS__)
+
+/* Description : Pack odd double word elements of vector pairs
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Odd double word elements of 'in0' are copied to the left half
+                 of 'out0' & odd double word elements of 'in1' are copied to
+                 the right half of 'out0'.
+*/
+#define PCKOD_D2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+{                                                         \
+    out0 = (RTYPE)__msa_pckod_d((v2i64)in0, (v2i64)in1);  \
+    out1 = (RTYPE)__msa_pckod_d((v2i64)in2, (v2i64)in3);  \
+}
+#define PCKOD_D2_UB(...) PCKOD_D2(v16u8, __VA_ARGS__)
+#define PCKOD_D2_SH(...) PCKOD_D2(v8i16, __VA_ARGS__)
+
+/* Description : Each byte element is logically xor'ed with immediate 128
+   Arguments   : Inputs  - in0, in1
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each unsigned byte element from input vector 'in0' is
+                 logically xor'ed with 128 and the result is stored in-place.
+*/
+#define XORI_B2_128(RTYPE, in0, in1)             \
+{                                                \
+    in0 = (RTYPE)__msa_xori_b((v16u8)in0, 128);  \
+    in1 = (RTYPE)__msa_xori_b((v16u8)in1, 128);  \
+}
+#define XORI_B2_128_UB(...) XORI_B2_128(v16u8, __VA_ARGS__)
+#define XORI_B2_128_SB(...) XORI_B2_128(v16i8, __VA_ARGS__)
+
+#define XORI_B3_128(RTYPE, in0, in1, in2)        \
+{                                                \
+    XORI_B2_128(RTYPE, in0, in1);                \
+    in2 = (RTYPE)__msa_xori_b((v16u8)in2, 128);  \
+}
+#define XORI_B3_128_SB(...) XORI_B3_128(v16i8, __VA_ARGS__)
+
+#define XORI_B4_128(RTYPE, in0, in1, in2, in3)  \
+{                                               \
+    XORI_B2_128(RTYPE, in0, in1);               \
+    XORI_B2_128(RTYPE, in2, in3);               \
+}
+#define XORI_B4_128_UB(...) XORI_B4_128(v16u8, __VA_ARGS__)
+#define XORI_B4_128_SB(...) XORI_B4_128(v16i8, __VA_ARGS__)
+
+#define XORI_B5_128(RTYPE, in0, in1, in2, in3, in4)  \
+{                                                    \
+    XORI_B3_128(RTYPE, in0, in1, in2);               \
+    XORI_B2_128(RTYPE, in3, in4);                    \
+}
+#define XORI_B5_128_SB(...) XORI_B5_128(v16i8, __VA_ARGS__)
+
+#define XORI_B8_128(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7)  \
+{                                                                   \
+    XORI_B4_128(RTYPE, in0, in1, in2, in3);                         \
+    XORI_B4_128(RTYPE, in4, in5, in6, in7);                         \
+}
+#define XORI_B8_128_SB(...) XORI_B8_128(v16i8, __VA_ARGS__)
+
+/* Description : Shift left all elements of vector (generic for all data types)
+   Arguments   : Inputs  - in0, in1, in2, in3, shift
+                 Outputs - in place operation
+                 Return Type - as per input vector RTYPE
+   Details     : Each element of vector 'in0' is left shifted by 'shift' and
+                 the result is written in-place.
+*/
+#define SLLI_4V(in0, in1, in2, in3, shift)  \
+{                                           \
+    in0 = in0 << shift;                     \
+    in1 = in1 << shift;                     \
+    in2 = in2 << shift;                     \
+    in3 = in3 << shift;                     \
+}
+
+/* Description : Arithmetic shift right all elements of vector
+                 (generic for all data types)
+   Arguments   : Inputs  - in0, in1, in2, in3, shift
+                 Outputs - in place operation
+                 Return Type - as per input vector RTYPE
+   Details     : Each element of vector 'in0' is right shifted by 'shift' and
+                 the result is written in-place. 'shift' is a GP variable.
+*/
+#define SRA_4V(in0, in1, in2, in3, shift)  \
+{                                          \
+    in0 = in0 >> shift;                    \
+    in1 = in1 >> shift;                    \
+    in2 = in2 >> shift;                    \
+    in3 = in3 >> shift;                    \
+}
+
+/* Description : Shift right arithmetic rounded words
+   Arguments   : Inputs  - in0, in1, shift
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is shifted right arithmetically by
+                 the number of bits in the corresponding element in the vector
+                 'shift'. The last discarded bit is added to shifted value for
+                 rounding and the result is written in-place.
+                 'shift' is a vector.
+*/
+#define SRAR_W2(RTYPE, in0, in1, shift)                   \
+{                                                         \
+    in0 = (RTYPE)__msa_srar_w((v4i32)in0, (v4i32)shift);  \
+    in1 = (RTYPE)__msa_srar_w((v4i32)in1, (v4i32)shift);  \
+}
+
+#define SRAR_W4(RTYPE, in0, in1, in2, in3, shift)  \
+{                                                  \
+    SRAR_W2(RTYPE, in0, in1, shift);               \
+    SRAR_W2(RTYPE, in2, in3, shift);               \
+}
+#define SRAR_W4_SW(...) SRAR_W4(v4i32, __VA_ARGS__)
+
+/* Description : Shift right arithmetic rounded (immediate)
+   Arguments   : Inputs  - in0, in1, shift
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is shifted right arithmetically by
+                 the value in 'shift'. The last discarded bit is added to the
+                 shifted value for rounding and the result is written in-place.
+                 'shift' is an immediate value.
+*/
+#define SRARI_H2(RTYPE, in0, in1, shift)            \
+{                                                   \
+    in0 = (RTYPE)__msa_srari_h((v8i16)in0, shift);  \
+    in1 = (RTYPE)__msa_srari_h((v8i16)in1, shift);  \
+}
+#define SRARI_H2_UH(...) SRARI_H2(v8u16, __VA_ARGS__)
+#define SRARI_H2_SH(...) SRARI_H2(v8i16, __VA_ARGS__)
+
+#define SRARI_H4(RTYPE, in0, in1, in2, in3, shift)  \
+{                                                   \
+    SRARI_H2(RTYPE, in0, in1, shift);               \
+    SRARI_H2(RTYPE, in2, in3, shift);               \
+}
+#define SRARI_H4_UH(...) SRARI_H4(v8u16, __VA_ARGS__)
+#define SRARI_H4_SH(...) SRARI_H4(v8i16, __VA_ARGS__)
+
+#define SRARI_W2(RTYPE, in0, in1, shift)            \
+{                                                   \
+    in0 = (RTYPE)__msa_srari_w((v4i32)in0, shift);  \
+    in1 = (RTYPE)__msa_srari_w((v4i32)in1, shift);  \
+}
+
+#define SRARI_W4(RTYPE, in0, in1, in2, in3, shift)  \
+{                                                   \
+    SRARI_W2(RTYPE, in0, in1, shift);               \
+    SRARI_W2(RTYPE, in2, in3, shift);               \
+}
+#define SRARI_W4_SW(...) SRARI_W4(v4i32, __VA_ARGS__)
+
+/* Description : Multiplication of pairs of vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+   Details     : Each element from 'in0' is multiplied with elements from 'in1'
+                 and the result is written to 'out0'
+*/
+#define MUL2(in0, in1, in2, in3, out0, out1)  \
+{                                             \
+    out0 = in0 * in1;                         \
+    out1 = in2 * in3;                         \
+}
+#define MUL4(in0, in1, in2, in3, in4, in5, in6, in7,  \
+             out0, out1, out2, out3)                  \
+{                                                     \
+    MUL2(in0, in1, in2, in3, out0, out1);             \
+    MUL2(in4, in5, in6, in7, out2, out3);             \
+}
+
+/* Description : Addition of 2 pairs of vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+   Details     : Each element in 'in0' is added to 'in1' and result is written
+                 to 'out0'.
+*/
+#define ADD2(in0, in1, in2, in3, out0, out1)  \
+{                                             \
+    out0 = in0 + in1;                         \
+    out1 = in2 + in3;                         \
+}
+#define ADD4(in0, in1, in2, in3, in4, in5, in6, in7,  \
+             out0, out1, out2, out3)                  \
+{                                                     \
+    ADD2(in0, in1, in2, in3, out0, out1);             \
+    ADD2(in4, in5, in6, in7, out2, out3);             \
+}
+
+/* Description : Subtraction of 2 pairs of vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+   Details     : Each element in 'in1' is subtracted from 'in0' and result is
+                 written to 'out0'.
+*/
+#define SUB2(in0, in1, in2, in3, out0, out1)  \
+{                                             \
+    out0 = in0 - in1;                         \
+    out1 = in2 - in3;                         \
+}
+#define SUB4(in0, in1, in2, in3, in4, in5, in6, in7,  \
+             out0, out1, out2, out3)                  \
+{                                                     \
+    out0 = in0 - in1;                                 \
+    out1 = in2 - in3;                                 \
+    out2 = in4 - in5;                                 \
+    out3 = in6 - in7;                                 \
+}
+
+/* Description : Sign extend halfword elements from right half of the vector
+   Arguments   : Input  - in    (halfword vector)
+                 Output - out   (sign extended word vector)
+                 Return Type - signed word
+   Details     : Sign bit of halfword elements from input vector 'in' is
+                 extracted and interleaved with same vector 'in0' to generate
+                 4 word elements keeping sign intact
+*/
+#define UNPCK_R_SH_SW(in, out)                     \
+{                                                  \
+    v8i16 sign_m;                                  \
+                                                   \
+    sign_m = __msa_clti_s_h((v8i16)in, 0);         \
+    out = (v4i32)__msa_ilvr_h(sign_m, (v8i16)in);  \
+}
+
+/* Description : Zero extend unsigned byte elements to halfword elements
+   Arguments   : Input   - in          (unsigned byte vector)
+                 Outputs - out0, out1  (unsigned  halfword vectors)
+                 Return Type - signed halfword
+   Details     : Zero extended right half of vector is returned in 'out0'
+                 Zero extended left half of vector is returned in 'out1'
+*/
+#define UNPCK_UB_SH(in, out0, out1)       \
+{                                         \
+    v16i8 zero_m = { 0 };                 \
+                                          \
+    ILVRL_B2_SH(zero_m, in, out0, out1);  \
+}
+
+/* Description : Sign extend halfword elements from input vector and return
+                 the result in pair of vectors
+   Arguments   : Input   - in            (halfword vector)
+                 Outputs - out0, out1   (sign extended word vectors)
+                 Return Type - signed word
+   Details     : Sign bit of halfword elements from input vector 'in' is
+                 extracted and interleaved right with same vector 'in0' to
+                 generate 4 signed word elements in 'out0'
+                 Then interleaved left with same vector 'in0' to
+                 generate 4 signed word elements in 'out1'
+*/
+#define UNPCK_SH_SW(in, out0, out1)        \
+{                                          \
+    v8i16 tmp_m;                           \
+                                           \
+    tmp_m = __msa_clti_s_h((v8i16)in, 0);  \
+    ILVRL_H2_SW(tmp_m, in, out0, out1);    \
+}
+
+/* Description : Butterfly of 4 input vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1, out2, out3
+   Details     : Butterfly operation
+*/
+#define BUTTERFLY_4(in0, in1, in2, in3, out0, out1, out2, out3)  \
+{                                                                \
+    out0 = in0 + in3;                                            \
+    out1 = in1 + in2;                                            \
+                                                                 \
+    out2 = in1 - in2;                                            \
+    out3 = in0 - in3;                                            \
+}
+
+/* Description : Transpose input 8x8 byte block
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - as per RTYPE
+*/
+#define TRANSPOSE8x8_UB(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,   \
+                        out0, out1, out2, out3, out4, out5, out6, out7)  \
+{                                                                        \
+    v16i8 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                \
+    v16i8 tmp4_m, tmp5_m, tmp6_m, tmp7_m;                                \
+                                                                         \
+    ILVR_B4_SB(in2, in0, in3, in1, in6, in4, in7, in5,                   \
+               tmp0_m, tmp1_m, tmp2_m, tmp3_m);                          \
+    ILVRL_B2_SB(tmp1_m, tmp0_m, tmp4_m, tmp5_m);                         \
+    ILVRL_B2_SB(tmp3_m, tmp2_m, tmp6_m, tmp7_m);                         \
+    ILVRL_W2(RTYPE, tmp6_m, tmp4_m, out0, out2);                         \
+    ILVRL_W2(RTYPE, tmp7_m, tmp5_m, out4, out6);                         \
+    SLDI_B2_0(RTYPE, out0, out2, out1, out3, 8);                         \
+    SLDI_B2_0(RTYPE, out4, out6, out5, out7, 8);                         \
+}
+#define TRANSPOSE8x8_UB_UB(...) TRANSPOSE8x8_UB(v16u8, __VA_ARGS__)
+
+/* Description : Transpose 16x4 block into 4x16 with byte elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7,
+                           in8, in9, in10, in11, in12, in13, in14, in15
+                 Outputs - out0, out1, out2, out3
+                 Return Type - unsigned byte
+*/
+#define TRANSPOSE16x4_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7,        \
+                            in8, in9, in10, in11, in12, in13, in14, in15,  \
+                            out0, out1, out2, out3)                        \
+{                                                                          \
+    v2i64 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                  \
+                                                                           \
+    ILVEV_W2_SD(in0, in4, in8, in12, tmp0_m, tmp1_m);                      \
+    out1 = (v16u8)__msa_ilvev_d(tmp1_m, tmp0_m);                           \
+                                                                           \
+    ILVEV_W2_SD(in1, in5, in9, in13, tmp0_m, tmp1_m);                      \
+    out3 = (v16u8)__msa_ilvev_d(tmp1_m, tmp0_m);                           \
+                                                                           \
+    ILVEV_W2_SD(in2, in6, in10, in14, tmp0_m, tmp1_m);                     \
+                                                                           \
+    tmp2_m = __msa_ilvev_d(tmp1_m, tmp0_m);                                \
+    ILVEV_W2_SD(in3, in7, in11, in15, tmp0_m, tmp1_m);                     \
+                                                                           \
+    tmp3_m = __msa_ilvev_d(tmp1_m, tmp0_m);                                \
+    ILVEV_B2_SD(out1, out3, tmp2_m, tmp3_m, tmp0_m, tmp1_m);               \
+    out0 = (v16u8)__msa_ilvev_h((v8i16)tmp1_m, (v8i16)tmp0_m);             \
+    out2 = (v16u8)__msa_ilvod_h((v8i16)tmp1_m, (v8i16)tmp0_m);             \
+                                                                           \
+    tmp0_m = (v2i64)__msa_ilvod_b((v16i8)out3, (v16i8)out1);               \
+    tmp1_m = (v2i64)__msa_ilvod_b((v16i8)tmp3_m, (v16i8)tmp2_m);           \
+    out1 = (v16u8)__msa_ilvev_h((v8i16)tmp1_m, (v8i16)tmp0_m);             \
+    out3 = (v16u8)__msa_ilvod_h((v8i16)tmp1_m, (v8i16)tmp0_m);             \
+}
+
+/* Description : Transpose 16x8 block into 8x16 with byte elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7,
+                           in8, in9, in10, in11, in12, in13, in14, in15
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - unsigned byte
+*/
+#define TRANSPOSE16x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7,          \
+                            in8, in9, in10, in11, in12, in13, in14, in15,    \
+                            out0, out1, out2, out3, out4, out5, out6, out7)  \
+{                                                                            \
+    v16u8 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                    \
+    v16u8 tmp4_m, tmp5_m, tmp6_m, tmp7_m;                                    \
+                                                                             \
+    ILVEV_D2_UB(in0, in8, in1, in9, out7, out6);                             \
+    ILVEV_D2_UB(in2, in10, in3, in11, out5, out4);                           \
+    ILVEV_D2_UB(in4, in12, in5, in13, out3, out2);                           \
+    ILVEV_D2_UB(in6, in14, in7, in15, out1, out0);                           \
+                                                                             \
+    tmp0_m = (v16u8)__msa_ilvev_b((v16i8)out6, (v16i8)out7);                 \
+    tmp4_m = (v16u8)__msa_ilvod_b((v16i8)out6, (v16i8)out7);                 \
+    tmp1_m = (v16u8)__msa_ilvev_b((v16i8)out4, (v16i8)out5);                 \
+    tmp5_m = (v16u8)__msa_ilvod_b((v16i8)out4, (v16i8)out5);                 \
+    out5 = (v16u8)__msa_ilvev_b((v16i8)out2, (v16i8)out3);                   \
+    tmp6_m = (v16u8)__msa_ilvod_b((v16i8)out2, (v16i8)out3);                 \
+    out7 = (v16u8)__msa_ilvev_b((v16i8)out0, (v16i8)out1);                   \
+    tmp7_m = (v16u8)__msa_ilvod_b((v16i8)out0, (v16i8)out1);                 \
+                                                                             \
+    ILVEV_H2_UB(tmp0_m, tmp1_m, out5, out7, tmp2_m, tmp3_m);                 \
+    out0 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);               \
+    out4 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);               \
+                                                                             \
+    tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp1_m, (v8i16)tmp0_m);             \
+    tmp3_m = (v16u8)__msa_ilvod_h((v8i16)out7, (v8i16)out5);                 \
+    out2 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);               \
+    out6 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);               \
+                                                                             \
+    ILVEV_H2_UB(tmp4_m, tmp5_m, tmp6_m, tmp7_m, tmp2_m, tmp3_m);             \
+    out1 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);               \
+    out5 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);               \
+                                                                             \
+    tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp5_m, (v8i16)tmp4_m);             \
+    tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp5_m, (v8i16)tmp4_m);             \
+    tmp3_m = (v16u8)__msa_ilvod_h((v8i16)tmp7_m, (v8i16)tmp6_m);             \
+    tmp3_m = (v16u8)__msa_ilvod_h((v8i16)tmp7_m, (v8i16)tmp6_m);             \
+    out3 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);               \
+    out7 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);               \
+}
+
+/* Description : Transpose 4x4 block with half word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1, out2, out3
+                 Return Type - signed halfword
+*/
+#define TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, out0, out1, out2, out3)  \
+{                                                                       \
+    v8i16 s0_m, s1_m;                                                   \
+                                                                        \
+    ILVR_H2_SH(in1, in0, in3, in2, s0_m, s1_m);                         \
+    ILVRL_W2_SH(s1_m, s0_m, out0, out2);                                \
+    out1 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out0);               \
+    out3 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out2);               \
+}
+
+/* Description : Transpose 8x4 block with half word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - signed halfword
+*/
+#define TRANSPOSE8X4_SH_SH(in0, in1, in2, in3, out0, out1, out2, out3)  \
+{                                                                       \
+    v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                               \
+                                                                        \
+    ILVR_H2_SH(in1, in0, in3, in2, tmp0_m, tmp1_m);                     \
+    ILVL_H2_SH(in1, in0, in3, in2, tmp2_m, tmp3_m);                     \
+    ILVR_W2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out0, out2);             \
+    ILVL_W2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out1, out3);             \
+}
+
+/* Description : Transpose 4x4 block with word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1, out2, out3
+                 Return Type - signed word
+*/
+#define TRANSPOSE4x4_SW_SW(in0, in1, in2, in3, out0, out1, out2, out3)  \
+{                                                                       \
+    v4i32 s0_m, s1_m, s2_m, s3_m;                                       \
+                                                                        \
+    ILVRL_W2_SW(in1, in0, s0_m, s1_m);                                  \
+    ILVRL_W2_SW(in3, in2, s2_m, s3_m);                                  \
+                                                                        \
+    out0 = (v4i32)__msa_ilvr_d((v2i64)s2_m, (v2i64)s0_m);               \
+    out1 = (v4i32)__msa_ilvl_d((v2i64)s2_m, (v2i64)s0_m);               \
+    out2 = (v4i32)__msa_ilvr_d((v2i64)s3_m, (v2i64)s1_m);               \
+    out3 = (v4i32)__msa_ilvl_d((v2i64)s3_m, (v2i64)s1_m);               \
+}
+
+/* Description : Dot product and addition of 3 signed halfword input vectors
+   Arguments   : Inputs - in0, in1, in2, coeff0, coeff1, coeff2
+                 Output - out0_m
+                 Return Type - signed halfword
+   Details     : Dot product of 'in0' with 'coeff0'
+                 Dot product of 'in1' with 'coeff1'
+                 Dot product of 'in2' with 'coeff2'
+                 Addition of all the 3 vector results
+                 out0_m = (in0 * coeff0) + (in1 * coeff1) + (in2 * coeff2)
+*/
+#define DPADD_SH3_SH(in0, in1, in2, coeff0, coeff1, coeff2)       \
+({                                                                \
+    v8i16 tmp1_m;                                                 \
+    v8i16 out0_m;                                                 \
+                                                                  \
+    out0_m = __msa_dotp_s_h((v16i8)in0, (v16i8)coeff0);           \
+    out0_m = __msa_dpadd_s_h(out0_m, (v16i8)in1, (v16i8)coeff1);  \
+    tmp1_m = __msa_dotp_s_h((v16i8)in2, (v16i8)coeff2);           \
+    out0_m = __msa_adds_s_h(out0_m, tmp1_m);                      \
+                                                                  \
+    out0_m;                                                       \
+})
+
+/* Description : Pack even elements of input vectors & xor with 128
+   Arguments   : Inputs - in0, in1
+                 Output - out_m
+                 Return Type - unsigned byte
+   Details     : Signed byte even elements from 'in0' and 'in1' are packed
+                 together in one vector and the resulting vector is xor'ed with
+                 128 to shift the range from signed to unsigned byte
+*/
+#define PCKEV_XORI128_UB(in0, in1)                         \
+({                                                         \
+    v16u8 out_m;                                           \
+    out_m = (v16u8)__msa_pckev_b((v16i8)in1, (v16i8)in0);  \
+    out_m = (v16u8)__msa_xori_b((v16u8)out_m, 128);        \
+    out_m;                                                 \
+})
+
+/* Description : Pack even byte elements and store byte vector in destination
+                 memory
+   Arguments   : Inputs - in0, in1, pdst
+*/
+#define PCKEV_ST_SB(in0, in1, pdst)                 \
+{                                                   \
+    v16i8 tmp_m;                                    \
+    tmp_m = __msa_pckev_b((v16i8)in1, (v16i8)in0);  \
+    ST_SB(tmp_m, (pdst));                           \
+}
+
+/* Description : Horizontal 2 tap filter kernel code
+   Arguments   : Inputs - in0, in1, mask, coeff, shift
+*/
+#define HORIZ_2TAP_FILT_UH(in0, in1, mask, coeff, shift)         \
+({                                                               \
+    v16i8 tmp0_m;                                                \
+    v8u16 tmp1_m;                                                \
+                                                                 \
+    tmp0_m = __msa_vshf_b((v16i8)mask, (v16i8)in1, (v16i8)in0);  \
+    tmp1_m = __msa_dotp_u_h((v16u8)tmp0_m, (v16u8)coeff);        \
+    tmp1_m = (v8u16)__msa_srari_h((v8i16)tmp1_m, shift);         \
+                                                                 \
+    tmp1_m;                                                      \
+})
+#endif  /* VP8_COMMON_MIPS_MSA_VP8_MACROS_MSA_H_ */
diff --git a/libvpx/libvpx/vp8/common/modecont.c b/libvpx/libvpx/vp8/common/modecont.c
new file mode 100644
index 0000000..86a74bc
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/modecont.c
@@ -0,0 +1,40 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "entropy.h"
+
+const int vp8_mode_contexts[6][4] =
+{
+    {
+        /* 0 */
+        7,     1,     1,   143,
+    },
+    {
+        /* 1 */
+        14,    18,    14,   107,
+    },
+    {
+        /* 2 */
+        135,    64,    57,    68,
+    },
+    {
+        /* 3 */
+        60,    56,   128,    65,
+    },
+    {
+        /* 4 */
+        159,   134,   128,    34,
+    },
+    {
+        /* 5 */
+        234,   188,   128,    28,
+    },
+};
diff --git a/libvpx/libvpx/vp8/common/modecont.h b/libvpx/libvpx/vp8/common/modecont.h
new file mode 100644
index 0000000..ff34c33
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/modecont.h
@@ -0,0 +1,25 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_MODECONT_H_
+#define VP8_COMMON_MODECONT_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern const int vp8_mode_contexts[6][4];
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_MODECONT_H_
diff --git a/libvpx/libvpx/vp8/common/mv.h b/libvpx/libvpx/vp8/common/mv.h
new file mode 100644
index 0000000..111ccd6
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/mv.h
@@ -0,0 +1,36 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_MV_H_
+#define VP8_COMMON_MV_H_
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct
+{
+    short row;
+    short col;
+} MV;
+
+typedef union int_mv
+{
+    uint32_t  as_int;
+    MV        as_mv;
+} int_mv;        /* facilitates faster equality tests and copies */
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_MV_H_
diff --git a/libvpx/libvpx/vp8/common/onyx.h b/libvpx/libvpx/vp8/common/onyx.h
new file mode 100644
index 0000000..febe815
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/onyx.h
@@ -0,0 +1,282 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_ONYX_H_
+#define VP8_COMMON_ONYX_H_
+
+#ifdef __cplusplus
+extern "C"
+{
+#endif
+
+#include "vpx_config.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+#include "vpx_scale/yv12config.h"
+#include "ppflags.h"
+
+    struct VP8_COMP;
+
+    /* Create/destroy static data structures. */
+
+    typedef enum
+    {
+        NORMAL      = 0,
+        FOURFIVE    = 1,
+        THREEFIVE   = 2,
+        ONETWO      = 3
+
+    } VPX_SCALING;
+
+    typedef enum
+    {
+        USAGE_LOCAL_FILE_PLAYBACK   = 0x0,
+        USAGE_STREAM_FROM_SERVER    = 0x1,
+        USAGE_CONSTRAINED_QUALITY   = 0x2,
+        USAGE_CONSTANT_QUALITY      = 0x3
+    } END_USAGE;
+
+
+    typedef enum
+    {
+        MODE_REALTIME       = 0x0,
+        MODE_GOODQUALITY    = 0x1,
+        MODE_BESTQUALITY    = 0x2,
+        MODE_FIRSTPASS      = 0x3,
+        MODE_SECONDPASS     = 0x4,
+        MODE_SECONDPASS_BEST = 0x5
+    } MODE;
+
+    typedef enum
+    {
+        FRAMEFLAGS_KEY    = 1,
+        FRAMEFLAGS_GOLDEN = 2,
+        FRAMEFLAGS_ALTREF = 4
+    } FRAMETYPE_FLAGS;
+
+
+#include <assert.h>
+    static INLINE void Scale2Ratio(int mode, int *hr, int *hs)
+    {
+        switch (mode)
+        {
+        case    NORMAL:
+            *hr = 1;
+            *hs = 1;
+            break;
+        case    FOURFIVE:
+            *hr = 4;
+            *hs = 5;
+            break;
+        case    THREEFIVE:
+            *hr = 3;
+            *hs = 5;
+            break;
+        case    ONETWO:
+            *hr = 1;
+            *hs = 2;
+            break;
+        default:
+            *hr = 1;
+            *hs = 1;
+            assert(0);
+            break;
+        }
+    }
+
+    typedef struct
+    {
+        /* 4 versions of bitstream defined:
+         *   0 best quality/slowest decode, 3 lowest quality/fastest decode
+         */
+        int Version;
+        int Width;
+        int Height;
+        struct vpx_rational  timebase;
+        unsigned int target_bandwidth;    /* kilobits per second */
+
+        /* Parameter used for applying denoiser.
+         * For temporal denoiser: noise_sensitivity = 0 means off,
+         * noise_sensitivity = 1 means temporal denoiser on for Y channel only,
+         * noise_sensitivity = 2 means temporal denoiser on for all channels.
+         * noise_sensitivity = 3 means aggressive denoising mode.
+         * noise_sensitivity >= 4 means adaptive denoising mode.
+         * Temporal denoiser is enabled via the configuration option:
+         * CONFIG_TEMPORAL_DENOISING.
+         * For spatial denoiser: noise_sensitivity controls the amount of
+         * pre-processing blur: noise_sensitivity = 0 means off.
+         * Spatial denoiser invoked under !CONFIG_TEMPORAL_DENOISING.
+         */
+        int noise_sensitivity;
+
+        /* parameter used for sharpening output: recommendation 0: */
+        int Sharpness;
+        int cpu_used;
+        unsigned int rc_max_intra_bitrate_pct;
+        unsigned int screen_content_mode;
+
+        /* mode ->
+         *(0)=Realtime/Live Encoding. This mode is optimized for realtim
+         *    encoding (for example, capturing a television signal or feed
+         *    from a live camera). ( speed setting controls how fast )
+         *(1)=Good Quality Fast Encoding. The encoder balances quality with
+         *    the amount of time it takes to encode the output. ( speed
+         *    setting controls how fast )
+         *(2)=One Pass - Best Quality. The encoder places priority on the
+         *    quality of the output over encoding speed. The output is
+         *    compressed at the highest possible quality. This option takes
+         *    the longest amount of time to encode. ( speed setting ignored
+         *    )
+         *(3)=Two Pass - First Pass. The encoder generates a file of
+         *    statistics for use in the second encoding pass. ( speed
+         *    setting controls how fast )
+         *(4)=Two Pass - Second Pass. The encoder uses the statistics that
+         *    were generated in the first encoding pass to create the
+         *    compressed output. ( speed setting controls how fast )
+         *(5)=Two Pass - Second Pass Best.  The encoder uses the statistics
+         *    that were generated in the first encoding pass to create the
+         *    compressed output using the highest possible quality, and
+         *    taking a longer amount of time to encode.. ( speed setting
+         *    ignored )
+         */
+        int Mode;
+
+        /* Key Framing Operations */
+        int auto_key;       /* automatically detect cut scenes */
+        int key_freq;       /* maximum distance to key frame. */
+
+        /* lagged compression (if allow_lag == 0 lag_in_frames is ignored) */
+        int allow_lag;
+        int lag_in_frames; /* how many frames lag before we start encoding */
+
+        /*
+         * DATARATE CONTROL OPTIONS
+         */
+
+        int end_usage; /* vbr or cbr */
+
+        /* buffer targeting aggressiveness */
+        int under_shoot_pct;
+        int over_shoot_pct;
+
+        /* buffering parameters */
+        int64_t starting_buffer_level;
+        int64_t optimal_buffer_level;
+        int64_t maximum_buffer_size;
+
+        int64_t starting_buffer_level_in_ms;
+        int64_t optimal_buffer_level_in_ms;
+        int64_t maximum_buffer_size_in_ms;
+
+        /* controlling quality */
+        int fixed_q;
+        int worst_allowed_q;
+        int best_allowed_q;
+        int cq_level;
+
+        /* allow internal resizing */
+        int allow_spatial_resampling;
+        int resample_down_water_mark;
+        int resample_up_water_mark;
+
+        /* allow internal frame rate alterations */
+        int allow_df;
+        int drop_frames_water_mark;
+
+        /* two pass datarate control */
+        int two_pass_vbrbias;
+        int two_pass_vbrmin_section;
+        int two_pass_vbrmax_section;
+
+        /*
+         * END DATARATE CONTROL OPTIONS
+         */
+
+        /* these parameters aren't to be used in final build don't use!!! */
+        int play_alternate;
+        int alt_freq;
+        int alt_q;
+        int key_q;
+        int gold_q;
+
+
+        int multi_threaded;   /* how many threads to run the encoder on */
+        int token_partitions; /* how many token partitions to create */
+
+        /* early breakout threshold: for video conf recommend 800 */
+        int encode_breakout;
+
+        /* Bitfield defining the error resiliency features to enable.
+         * Can provide decodable frames after losses in previous
+         * frames and decodable partitions after losses in the same frame.
+         */
+        unsigned int error_resilient_mode;
+
+        int arnr_max_frames;
+        int arnr_strength;
+        int arnr_type;
+
+        vpx_fixed_buf_t        two_pass_stats_in;
+        struct vpx_codec_pkt_list  *output_pkt_list;
+
+        vp8e_tuning tuning;
+
+        /* Temporal scaling parameters */
+        unsigned int number_of_layers;
+        unsigned int target_bitrate[VPX_TS_MAX_PERIODICITY];
+        unsigned int rate_decimator[VPX_TS_MAX_PERIODICITY];
+        unsigned int periodicity;
+        unsigned int layer_id[VPX_TS_MAX_PERIODICITY];
+
+#if CONFIG_MULTI_RES_ENCODING
+        /* Number of total resolutions encoded */
+        unsigned int mr_total_resolutions;
+
+        /* Current encoder ID */
+        unsigned int mr_encoder_id;
+
+        /* Down-sampling factor */
+        vpx_rational_t mr_down_sampling_factor;
+
+        /* Memory location to store low-resolution encoder's mode info */
+        void* mr_low_res_mode_info;
+#endif
+    } VP8_CONFIG;
+
+
+    void vp8_initialize();
+
+    struct VP8_COMP* vp8_create_compressor(VP8_CONFIG *oxcf);
+    void vp8_remove_compressor(struct VP8_COMP* *comp);
+
+    void vp8_init_config(struct VP8_COMP* onyx, VP8_CONFIG *oxcf);
+    void vp8_change_config(struct VP8_COMP* onyx, VP8_CONFIG *oxcf);
+
+    int vp8_receive_raw_frame(struct VP8_COMP* comp, unsigned int frame_flags, YV12_BUFFER_CONFIG *sd, int64_t time_stamp, int64_t end_time_stamp);
+    int vp8_get_compressed_data(struct VP8_COMP* comp, unsigned int *frame_flags, unsigned long *size, unsigned char *dest, unsigned char *dest_end, int64_t *time_stamp, int64_t *time_end, int flush);
+    int vp8_get_preview_raw_frame(struct VP8_COMP* comp, YV12_BUFFER_CONFIG *dest, vp8_ppflags_t *flags);
+
+    int vp8_use_as_reference(struct VP8_COMP* comp, int ref_frame_flags);
+    int vp8_update_reference(struct VP8_COMP* comp, int ref_frame_flags);
+    int vp8_get_reference(struct VP8_COMP* comp, enum vpx_ref_frame_type ref_frame_flag, YV12_BUFFER_CONFIG *sd);
+    int vp8_set_reference(struct VP8_COMP* comp, enum vpx_ref_frame_type ref_frame_flag, YV12_BUFFER_CONFIG *sd);
+    int vp8_update_entropy(struct VP8_COMP* comp, int update);
+    int vp8_set_roimap(struct VP8_COMP* comp, unsigned char *map, unsigned int rows, unsigned int cols, int delta_q[4], int delta_lf[4], unsigned int threshold[4]);
+    int vp8_set_active_map(struct VP8_COMP* comp, unsigned char *map, unsigned int rows, unsigned int cols);
+    int vp8_set_internal_size(struct VP8_COMP* comp, VPX_SCALING horiz_mode, VPX_SCALING vert_mode);
+    int vp8_get_quantizer(struct VP8_COMP* c);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // VP8_COMMON_ONYX_H_
diff --git a/libvpx/libvpx/vp8/common/onyxc_int.h b/libvpx/libvpx/vp8/common/onyxc_int.h
new file mode 100644
index 0000000..6d89865
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/onyxc_int.h
@@ -0,0 +1,185 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_ONYXC_INT_H_
+#define VP8_COMMON_ONYXC_INT_H_
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "loopfilter.h"
+#include "entropymv.h"
+#include "entropy.h"
+#if CONFIG_POSTPROC
+#include "postproc.h"
+#endif
+
+/*#ifdef PACKET_TESTING*/
+#include "header.h"
+/*#endif*/
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MINQ 0
+#define MAXQ 127
+#define QINDEX_RANGE (MAXQ + 1)
+
+#define NUM_YV12_BUFFERS 4
+
+#define MAX_PARTITIONS 9
+
+typedef struct frame_contexts
+{
+    vp8_prob bmode_prob [VP8_BINTRAMODES-1];
+    vp8_prob ymode_prob [VP8_YMODES-1];   /* interframe intra mode probs */
+    vp8_prob uv_mode_prob [VP8_UV_MODES-1];
+    vp8_prob sub_mv_ref_prob [VP8_SUBMVREFS-1];
+    vp8_prob coef_probs [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES];
+    MV_CONTEXT mvc[2];
+} FRAME_CONTEXT;
+
+typedef enum
+{
+    ONE_PARTITION  = 0,
+    TWO_PARTITION  = 1,
+    FOUR_PARTITION = 2,
+    EIGHT_PARTITION = 3
+} TOKEN_PARTITION;
+
+typedef enum
+{
+    RECON_CLAMP_REQUIRED        = 0,
+    RECON_CLAMP_NOTREQUIRED     = 1
+} CLAMP_TYPE;
+
+typedef struct VP8Common
+
+{
+    struct vpx_internal_error_info  error;
+
+    DECLARE_ALIGNED(16, short, Y1dequant[QINDEX_RANGE][2]);
+    DECLARE_ALIGNED(16, short, Y2dequant[QINDEX_RANGE][2]);
+    DECLARE_ALIGNED(16, short, UVdequant[QINDEX_RANGE][2]);
+
+    int Width;
+    int Height;
+    int horiz_scale;
+    int vert_scale;
+
+    CLAMP_TYPE  clamp_type;
+
+    YV12_BUFFER_CONFIG *frame_to_show;
+
+    YV12_BUFFER_CONFIG yv12_fb[NUM_YV12_BUFFERS];
+    int fb_idx_ref_cnt[NUM_YV12_BUFFERS];
+    int new_fb_idx, lst_fb_idx, gld_fb_idx, alt_fb_idx;
+
+    YV12_BUFFER_CONFIG temp_scale_frame;
+
+#if CONFIG_POSTPROC
+    YV12_BUFFER_CONFIG post_proc_buffer;
+    YV12_BUFFER_CONFIG post_proc_buffer_int;
+    int post_proc_buffer_int_used;
+    unsigned char *pp_limits_buffer;   /* post-processing filter coefficients */
+#endif
+
+    FRAME_TYPE last_frame_type;  /* Save last frame's frame type for motion search. */
+    FRAME_TYPE frame_type;
+
+    int show_frame;
+
+    int frame_flags;
+    int MBs;
+    int mb_rows;
+    int mb_cols;
+    int mode_info_stride;
+
+    /* profile settings */
+    int mb_no_coeff_skip;
+    int no_lpf;
+    int use_bilinear_mc_filter;
+    int full_pixel;
+
+    int base_qindex;
+
+    int y1dc_delta_q;
+    int y2dc_delta_q;
+    int y2ac_delta_q;
+    int uvdc_delta_q;
+    int uvac_delta_q;
+
+    /* We allocate a MODE_INFO struct for each macroblock, together with
+       an extra row on top and column on the left to simplify prediction. */
+
+    MODE_INFO *mip; /* Base of allocated array */
+    MODE_INFO *mi;  /* Corresponds to upper left visible macroblock */
+#if CONFIG_ERROR_CONCEALMENT
+    MODE_INFO *prev_mip; /* MODE_INFO array 'mip' from last decoded frame */
+    MODE_INFO *prev_mi;  /* 'mi' from last frame (points into prev_mip) */
+#endif
+    MODE_INFO *show_frame_mi;  /* MODE_INFO for the last decoded frame
+                                  to show */
+    LOOPFILTERTYPE filter_type;
+
+    loop_filter_info_n lf_info;
+
+    int filter_level;
+    int last_sharpness_level;
+    int sharpness_level;
+
+    int refresh_last_frame;       /* Two state 0 = NO, 1 = YES */
+    int refresh_golden_frame;     /* Two state 0 = NO, 1 = YES */
+    int refresh_alt_ref_frame;     /* Two state 0 = NO, 1 = YES */
+
+    int copy_buffer_to_gf;         /* 0 none, 1 Last to GF, 2 ARF to GF */
+    int copy_buffer_to_arf;        /* 0 none, 1 Last to ARF, 2 GF to ARF */
+
+    int refresh_entropy_probs;    /* Two state 0 = NO, 1 = YES */
+
+    int ref_frame_sign_bias[MAX_REF_FRAMES];    /* Two state 0, 1 */
+
+    /* Y,U,V,Y2 */
+    ENTROPY_CONTEXT_PLANES *above_context;   /* row of context for each plane */
+    ENTROPY_CONTEXT_PLANES left_context;  /* (up to) 4 contexts "" */
+
+    FRAME_CONTEXT lfc; /* last frame entropy */
+    FRAME_CONTEXT fc;  /* this frame entropy */
+
+    unsigned int current_video_frame;
+
+    int version;
+
+    TOKEN_PARTITION multi_token_partition;
+
+#ifdef PACKET_TESTING
+    VP8_HEADER oh;
+#endif
+#if CONFIG_POSTPROC_VISUALIZER
+    double bitrate;
+    double framerate;
+#endif
+
+#if CONFIG_MULTITHREAD
+    int processor_core_count;
+#endif
+#if CONFIG_POSTPROC
+    struct postproc_state  postproc_state;
+#endif
+    int cpu_caps;
+} VP8_COMMON;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_ONYXC_INT_H_
diff --git a/libvpx/libvpx/vp8/common/onyxd.h b/libvpx/libvpx/vp8/common/onyxd.h
new file mode 100644
index 0000000..e37b29f
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/onyxd.h
@@ -0,0 +1,63 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_ONYXD_H_
+#define VP8_COMMON_ONYXD_H_
+
+
+/* Create/destroy static data structures. */
+#ifdef __cplusplus
+extern "C"
+{
+#endif
+#include "vpx_scale/yv12config.h"
+#include "ppflags.h"
+#include "vpx_ports/mem.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/vp8.h"
+
+    struct VP8D_COMP;
+
+    typedef struct
+    {
+        int     Width;
+        int     Height;
+        int     Version;
+        int     postprocess;
+        int     max_threads;
+        int     error_concealment;
+    } VP8D_CONFIG;
+
+    typedef enum
+    {
+        VP8D_OK = 0
+    } VP8D_SETTING;
+
+    void vp8dx_initialize(void);
+
+    void vp8dx_set_setting(struct VP8D_COMP* comp, VP8D_SETTING oxst, int x);
+
+    int vp8dx_get_setting(struct VP8D_COMP* comp, VP8D_SETTING oxst);
+
+    int vp8dx_receive_compressed_data(struct VP8D_COMP* comp,
+                                      size_t size, const uint8_t *dest,
+                                      int64_t time_stamp);
+    int vp8dx_get_raw_frame(struct VP8D_COMP* comp, YV12_BUFFER_CONFIG *sd, int64_t *time_stamp, int64_t *time_end_stamp, vp8_ppflags_t *flags);
+
+    vpx_codec_err_t vp8dx_get_reference(struct VP8D_COMP* comp, enum vpx_ref_frame_type ref_frame_flag, YV12_BUFFER_CONFIG *sd);
+    vpx_codec_err_t vp8dx_set_reference(struct VP8D_COMP* comp, enum vpx_ref_frame_type ref_frame_flag, YV12_BUFFER_CONFIG *sd);
+
+#ifdef __cplusplus
+}
+#endif
+
+
+#endif  // VP8_COMMON_ONYXD_H_
diff --git a/libvpx/libvpx/vp8/common/postproc.c b/libvpx/libvpx/vp8/common/postproc.c
new file mode 100644
index 0000000..6baf00f
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/postproc.c
@@ -0,0 +1,1161 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vpx_dsp_rtcd.h"
+#include "vp8_rtcd.h"
+#include "vpx_scale_rtcd.h"
+#include "vpx_scale/yv12config.h"
+#include "postproc.h"
+#include "common.h"
+#include "vpx_scale/vpx_scale.h"
+#include "systemdependent.h"
+
+#include <limits.h>
+#include <math.h>
+#include <stdlib.h>
+#include <stdio.h>
+
+#define RGB_TO_YUV(t)                                                                       \
+    ( (0.257*(float)(t>>16)) + (0.504*(float)(t>>8&0xff)) + (0.098*(float)(t&0xff)) + 16),  \
+    (-(0.148*(float)(t>>16)) - (0.291*(float)(t>>8&0xff)) + (0.439*(float)(t&0xff)) + 128), \
+    ( (0.439*(float)(t>>16)) - (0.368*(float)(t>>8&0xff)) - (0.071*(float)(t&0xff)) + 128)
+
+/* global constants */
+#if CONFIG_POSTPROC_VISUALIZER
+static const unsigned char MB_PREDICTION_MODE_colors[MB_MODE_COUNT][3] =
+{
+    { RGB_TO_YUV(0x98FB98) },   /* PaleGreen */
+    { RGB_TO_YUV(0x00FF00) },   /* Green */
+    { RGB_TO_YUV(0xADFF2F) },   /* GreenYellow */
+    { RGB_TO_YUV(0x228B22) },   /* ForestGreen */
+    { RGB_TO_YUV(0x006400) },   /* DarkGreen */
+    { RGB_TO_YUV(0x98F5FF) },   /* Cadet Blue */
+    { RGB_TO_YUV(0x6CA6CD) },   /* Sky Blue */
+    { RGB_TO_YUV(0x00008B) },   /* Dark blue */
+    { RGB_TO_YUV(0x551A8B) },   /* Purple */
+    { RGB_TO_YUV(0xFF0000) }    /* Red */
+};
+
+static const unsigned char B_PREDICTION_MODE_colors[B_MODE_COUNT][3] =
+{
+    { RGB_TO_YUV(0x6633ff) },   /* Purple */
+    { RGB_TO_YUV(0xcc33ff) },   /* Magenta */
+    { RGB_TO_YUV(0xff33cc) },   /* Pink */
+    { RGB_TO_YUV(0xff3366) },   /* Coral */
+    { RGB_TO_YUV(0x3366ff) },   /* Blue */
+    { RGB_TO_YUV(0xed00f5) },   /* Dark Blue */
+    { RGB_TO_YUV(0x2e00b8) },   /* Dark Purple */
+    { RGB_TO_YUV(0xff6633) },   /* Orange */
+    { RGB_TO_YUV(0x33ccff) },   /* Light Blue */
+    { RGB_TO_YUV(0x8ab800) },   /* Green */
+    { RGB_TO_YUV(0xffcc33) },   /* Light Orange */
+    { RGB_TO_YUV(0x33ffcc) },   /* Aqua */
+    { RGB_TO_YUV(0x66ff33) },   /* Light Green */
+    { RGB_TO_YUV(0xccff33) },   /* Yellow */
+};
+
+static const unsigned char MV_REFERENCE_FRAME_colors[MAX_REF_FRAMES][3] =
+{
+    { RGB_TO_YUV(0x00ff00) },   /* Blue */
+    { RGB_TO_YUV(0x0000ff) },   /* Green */
+    { RGB_TO_YUV(0xffff00) },   /* Yellow */
+    { RGB_TO_YUV(0xff0000) },   /* Red */
+};
+#endif
+
+const short vp8_rv[] =
+{
+    8, 5, 2, 2, 8, 12, 4, 9, 8, 3,
+    0, 3, 9, 0, 0, 0, 8, 3, 14, 4,
+    10, 1, 11, 14, 1, 14, 9, 6, 12, 11,
+    8, 6, 10, 0, 0, 8, 9, 0, 3, 14,
+    8, 11, 13, 4, 2, 9, 0, 3, 9, 6,
+    1, 2, 3, 14, 13, 1, 8, 2, 9, 7,
+    3, 3, 1, 13, 13, 6, 6, 5, 2, 7,
+    11, 9, 11, 8, 7, 3, 2, 0, 13, 13,
+    14, 4, 12, 5, 12, 10, 8, 10, 13, 10,
+    4, 14, 4, 10, 0, 8, 11, 1, 13, 7,
+    7, 14, 6, 14, 13, 2, 13, 5, 4, 4,
+    0, 10, 0, 5, 13, 2, 12, 7, 11, 13,
+    8, 0, 4, 10, 7, 2, 7, 2, 2, 5,
+    3, 4, 7, 3, 3, 14, 14, 5, 9, 13,
+    3, 14, 3, 6, 3, 0, 11, 8, 13, 1,
+    13, 1, 12, 0, 10, 9, 7, 6, 2, 8,
+    5, 2, 13, 7, 1, 13, 14, 7, 6, 7,
+    9, 6, 10, 11, 7, 8, 7, 5, 14, 8,
+    4, 4, 0, 8, 7, 10, 0, 8, 14, 11,
+    3, 12, 5, 7, 14, 3, 14, 5, 2, 6,
+    11, 12, 12, 8, 0, 11, 13, 1, 2, 0,
+    5, 10, 14, 7, 8, 0, 4, 11, 0, 8,
+    0, 3, 10, 5, 8, 0, 11, 6, 7, 8,
+    10, 7, 13, 9, 2, 5, 1, 5, 10, 2,
+    4, 3, 5, 6, 10, 8, 9, 4, 11, 14,
+    0, 10, 0, 5, 13, 2, 12, 7, 11, 13,
+    8, 0, 4, 10, 7, 2, 7, 2, 2, 5,
+    3, 4, 7, 3, 3, 14, 14, 5, 9, 13,
+    3, 14, 3, 6, 3, 0, 11, 8, 13, 1,
+    13, 1, 12, 0, 10, 9, 7, 6, 2, 8,
+    5, 2, 13, 7, 1, 13, 14, 7, 6, 7,
+    9, 6, 10, 11, 7, 8, 7, 5, 14, 8,
+    4, 4, 0, 8, 7, 10, 0, 8, 14, 11,
+    3, 12, 5, 7, 14, 3, 14, 5, 2, 6,
+    11, 12, 12, 8, 0, 11, 13, 1, 2, 0,
+    5, 10, 14, 7, 8, 0, 4, 11, 0, 8,
+    0, 3, 10, 5, 8, 0, 11, 6, 7, 8,
+    10, 7, 13, 9, 2, 5, 1, 5, 10, 2,
+    4, 3, 5, 6, 10, 8, 9, 4, 11, 14,
+    3, 8, 3, 7, 8, 5, 11, 4, 12, 3,
+    11, 9, 14, 8, 14, 13, 4, 3, 1, 2,
+    14, 6, 5, 4, 4, 11, 4, 6, 2, 1,
+    5, 8, 8, 12, 13, 5, 14, 10, 12, 13,
+    0, 9, 5, 5, 11, 10, 13, 9, 10, 13,
+};
+
+extern void vp8_blit_text(const char *msg, unsigned char *address, const int pitch);
+extern void vp8_blit_line(int x0, int x1, int y0, int y1, unsigned char *image, const int pitch);
+/***********************************************************************************************************
+ */
+void vp8_post_proc_down_and_across_mb_row_c
+(
+    unsigned char *src_ptr,
+    unsigned char *dst_ptr,
+    int src_pixels_per_line,
+    int dst_pixels_per_line,
+    int cols,
+    unsigned char *f,
+    int size
+)
+{
+    unsigned char *p_src, *p_dst;
+    int row;
+    int col;
+    unsigned char v;
+    unsigned char d[4];
+
+    for (row = 0; row < size; row++)
+    {
+        /* post_proc_down for one row */
+        p_src = src_ptr;
+        p_dst = dst_ptr;
+
+        for (col = 0; col < cols; col++)
+        {
+            unsigned char p_above2 = p_src[col - 2 * src_pixels_per_line];
+            unsigned char p_above1 = p_src[col - src_pixels_per_line];
+            unsigned char p_below1 = p_src[col + src_pixels_per_line];
+            unsigned char p_below2 = p_src[col + 2 * src_pixels_per_line];
+
+            v = p_src[col];
+
+            if ((abs(v - p_above2) < f[col]) && (abs(v - p_above1) < f[col])
+                && (abs(v - p_below1) < f[col]) && (abs(v - p_below2) < f[col]))
+            {
+                unsigned char k1, k2, k3;
+                k1 = (p_above2 + p_above1 + 1) >> 1;
+                k2 = (p_below2 + p_below1 + 1) >> 1;
+                k3 = (k1 + k2 + 1) >> 1;
+                v = (k3 + v + 1) >> 1;
+            }
+
+            p_dst[col] = v;
+        }
+
+        /* now post_proc_across */
+        p_src = dst_ptr;
+        p_dst = dst_ptr;
+
+        p_src[-2] = p_src[-1] = p_src[0];
+        p_src[cols] = p_src[cols + 1] = p_src[cols - 1];
+
+        for (col = 0; col < cols; col++)
+        {
+            v = p_src[col];
+
+            if ((abs(v - p_src[col - 2]) < f[col])
+                && (abs(v - p_src[col - 1]) < f[col])
+                && (abs(v - p_src[col + 1]) < f[col])
+                && (abs(v - p_src[col + 2]) < f[col]))
+            {
+                unsigned char k1, k2, k3;
+                k1 = (p_src[col - 2] + p_src[col - 1] + 1) >> 1;
+                k2 = (p_src[col + 2] + p_src[col + 1] + 1) >> 1;
+                k3 = (k1 + k2 + 1) >> 1;
+                v = (k3 + v + 1) >> 1;
+            }
+
+            d[col & 3] = v;
+
+            if (col >= 2)
+                p_dst[col - 2] = d[(col - 2) & 3];
+        }
+
+        /* handle the last two pixels */
+        p_dst[col - 2] = d[(col - 2) & 3];
+        p_dst[col - 1] = d[(col - 1) & 3];
+
+        /* next row */
+        src_ptr += src_pixels_per_line;
+        dst_ptr += dst_pixels_per_line;
+    }
+}
+
+static int q2mbl(int x)
+{
+    if (x < 20) x = 20;
+
+    x = 50 + (x - 50) * 10 / 8;
+    return x * x / 3;
+}
+
+void vp8_mbpost_proc_across_ip_c(unsigned char *src, int pitch, int rows, int cols, int flimit)
+{
+    int r, c, i;
+
+    unsigned char *s = src;
+    unsigned char d[16];
+
+    for (r = 0; r < rows; r++)
+    {
+        int sumsq = 0;
+        int sum   = 0;
+
+        for (i = -8; i < 0; i++)
+          s[i]=s[0];
+
+        /* 17 avoids valgrind warning - we buffer values in c in d
+         * and only write them when we've read 8 ahead...
+         */
+        for (i = 0; i < 17; i++)
+          s[i+cols]=s[cols-1];
+
+        for (i = -8; i <= 6; i++)
+        {
+            sumsq += s[i] * s[i];
+            sum   += s[i];
+            d[i+8] = 0;
+        }
+
+        for (c = 0; c < cols + 8; c++)
+        {
+            int x = s[c+7] - s[c-8];
+            int y = s[c+7] + s[c-8];
+
+            sum  += x;
+            sumsq += x * y;
+
+            d[c&15] = s[c];
+
+            if (sumsq * 15 - sum * sum < flimit)
+            {
+                d[c&15] = (8 + sum + s[c]) >> 4;
+            }
+
+            s[c-8] = d[(c-8)&15];
+        }
+
+        s += pitch;
+    }
+}
+
+void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols, int flimit)
+{
+    int r, c, i;
+    const short *rv3 = &vp8_rv[63&rand()];
+
+    for (c = 0; c < cols; c++ )
+    {
+        unsigned char *s = &dst[c];
+        int sumsq = 0;
+        int sum   = 0;
+        unsigned char d[16];
+        const short *rv2 = rv3 + ((c * 17) & 127);
+
+        for (i = -8; i < 0; i++)
+          s[i*pitch]=s[0];
+
+        /* 17 avoids valgrind warning - we buffer values in c in d
+         * and only write them when we've read 8 ahead...
+         */
+        for (i = 0; i < 17; i++)
+          s[(i+rows)*pitch]=s[(rows-1)*pitch];
+
+        for (i = -8; i <= 6; i++)
+        {
+            sumsq += s[i*pitch] * s[i*pitch];
+            sum   += s[i*pitch];
+        }
+
+        for (r = 0; r < rows + 8; r++)
+        {
+            sumsq += s[7*pitch] * s[ 7*pitch] - s[-8*pitch] * s[-8*pitch];
+            sum  += s[7*pitch] - s[-8*pitch];
+            d[r&15] = s[0];
+
+            if (sumsq * 15 - sum * sum < flimit)
+            {
+                d[r&15] = (rv2[r&127] + sum + s[0]) >> 4;
+            }
+            if (r >= 8)
+              s[-8*pitch] = d[(r-8)&15];
+            s += pitch;
+        }
+    }
+}
+
+#if CONFIG_POSTPROC
+static void vp8_de_mblock(YV12_BUFFER_CONFIG         *post,
+                          int                         q)
+{
+    vp8_mbpost_proc_across_ip(post->y_buffer, post->y_stride, post->y_height,
+                              post->y_width, q2mbl(q));
+    vp8_mbpost_proc_down(post->y_buffer, post->y_stride, post->y_height,
+                         post->y_width, q2mbl(q));
+}
+
+void vp8_deblock(VP8_COMMON                 *cm,
+                 YV12_BUFFER_CONFIG         *source,
+                 YV12_BUFFER_CONFIG         *post,
+                 int                         q,
+                 int                         low_var_thresh,
+                 int                         flag)
+{
+    double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
+    int ppl = (int)(level + .5);
+
+    const MODE_INFO *mode_info_context = cm->show_frame_mi;
+    int mbr, mbc;
+
+    /* The pixel thresholds are adjusted according to if or not the macroblock
+     * is a skipped block.  */
+    unsigned char *ylimits = cm->pp_limits_buffer;
+    unsigned char *uvlimits = cm->pp_limits_buffer + 16 * cm->mb_cols;
+    (void) low_var_thresh;
+    (void) flag;
+
+    if (ppl > 0)
+    {
+        for (mbr = 0; mbr < cm->mb_rows; mbr++)
+        {
+            unsigned char *ylptr = ylimits;
+            unsigned char *uvlptr = uvlimits;
+            for (mbc = 0; mbc < cm->mb_cols; mbc++)
+            {
+                unsigned char mb_ppl;
+
+                if (mode_info_context->mbmi.mb_skip_coeff)
+                    mb_ppl = (unsigned char)ppl >> 1;
+                else
+                    mb_ppl = (unsigned char)ppl;
+
+                memset(ylptr, mb_ppl, 16);
+                memset(uvlptr, mb_ppl, 8);
+
+                ylptr += 16;
+                uvlptr += 8;
+                mode_info_context++;
+            }
+            mode_info_context++;
+
+            vp8_post_proc_down_and_across_mb_row(
+                source->y_buffer + 16 * mbr * source->y_stride,
+                post->y_buffer + 16 * mbr * post->y_stride, source->y_stride,
+                post->y_stride, source->y_width, ylimits, 16);
+
+            vp8_post_proc_down_and_across_mb_row(
+                source->u_buffer + 8 * mbr * source->uv_stride,
+                post->u_buffer + 8 * mbr * post->uv_stride, source->uv_stride,
+                post->uv_stride, source->uv_width, uvlimits, 8);
+            vp8_post_proc_down_and_across_mb_row(
+                source->v_buffer + 8 * mbr * source->uv_stride,
+                post->v_buffer + 8 * mbr * post->uv_stride, source->uv_stride,
+                post->uv_stride, source->uv_width, uvlimits, 8);
+        }
+    } else
+    {
+        vp8_yv12_copy_frame(source, post);
+    }
+}
+#endif
+
+void vp8_de_noise(VP8_COMMON                 *cm,
+                  YV12_BUFFER_CONFIG         *source,
+                  YV12_BUFFER_CONFIG         *post,
+                  int                         q,
+                  int                         low_var_thresh,
+                  int                         flag,
+                  int                         uvfilter)
+{
+    int mbr;
+    double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
+    int ppl = (int)(level + .5);
+    int mb_rows = cm->mb_rows;
+    int mb_cols = cm->mb_cols;
+    unsigned char *limits = cm->pp_limits_buffer;;
+    (void) post;
+    (void) low_var_thresh;
+    (void) flag;
+
+    memset(limits, (unsigned char)ppl, 16 * mb_cols);
+
+    /* TODO: The original code don't filter the 2 outer rows and columns. */
+    for (mbr = 0; mbr < mb_rows; mbr++)
+    {
+        vp8_post_proc_down_and_across_mb_row(
+            source->y_buffer + 16 * mbr * source->y_stride,
+            source->y_buffer + 16 * mbr * source->y_stride,
+            source->y_stride, source->y_stride, source->y_width, limits, 16);
+        if (uvfilter == 1) {
+          vp8_post_proc_down_and_across_mb_row(
+              source->u_buffer + 8 * mbr * source->uv_stride,
+              source->u_buffer + 8 * mbr * source->uv_stride,
+              source->uv_stride, source->uv_stride, source->uv_width, limits,
+              8);
+          vp8_post_proc_down_and_across_mb_row(
+              source->v_buffer + 8 * mbr * source->uv_stride,
+              source->v_buffer + 8 * mbr * source->uv_stride,
+              source->uv_stride, source->uv_stride, source->uv_width, limits,
+              8);
+        }
+    }
+}
+
+static double gaussian(double sigma, double mu, double x)
+{
+    return 1 / (sigma * sqrt(2.0 * 3.14159265)) *
+           (exp(-(x - mu) * (x - mu) / (2 * sigma * sigma)));
+}
+
+static void fillrd(struct postproc_state *state, int q, int a)
+{
+    char char_dist[300];
+
+    double sigma;
+    int i;
+
+    vp8_clear_system_state();
+
+
+    sigma = a + .5 + .6 * (63 - q) / 63.0;
+
+    /* set up a lookup table of 256 entries that matches
+     * a gaussian distribution with sigma determined by q.
+     */
+    {
+        int next, j;
+
+        next = 0;
+
+        for (i = -32; i < 32; i++)
+        {
+            const int v = (int)(.5 + 256 * gaussian(sigma, 0, i));
+
+            if (v)
+            {
+                for (j = 0; j < v; j++)
+                {
+                    char_dist[next+j] = (char) i;
+                }
+
+                next = next + j;
+            }
+
+        }
+
+        for (; next < 256; next++)
+            char_dist[next] = 0;
+
+    }
+
+    for (i = 0; i < 3072; i++)
+    {
+        state->noise[i] = char_dist[rand() & 0xff];
+    }
+
+    for (i = 0; i < 16; i++)
+    {
+        state->blackclamp[i] = -char_dist[0];
+        state->whiteclamp[i] = -char_dist[0];
+        state->bothclamp[i] = -2 * char_dist[0];
+    }
+
+    state->last_q = q;
+    state->last_noise = a;
+}
+
+/* Blend the macro block with a solid colored square.  Leave the
+ * edges unblended to give distinction to macro blocks in areas
+ * filled with the same color block.
+ */
+void vp8_blend_mb_inner_c (unsigned char *y, unsigned char *u, unsigned char *v,
+                        int y_1, int u_1, int v_1, int alpha, int stride)
+{
+    int i, j;
+    int y1_const = y_1*((1<<16)-alpha);
+    int u1_const = u_1*((1<<16)-alpha);
+    int v1_const = v_1*((1<<16)-alpha);
+
+    y += 2*stride + 2;
+    for (i = 0; i < 12; i++)
+    {
+        for (j = 0; j < 12; j++)
+        {
+            y[j] = (y[j]*alpha + y1_const)>>16;
+        }
+        y += stride;
+    }
+
+    stride >>= 1;
+
+    u += stride + 1;
+    v += stride + 1;
+
+    for (i = 0; i < 6; i++)
+    {
+        for (j = 0; j < 6; j++)
+        {
+            u[j] = (u[j]*alpha + u1_const)>>16;
+            v[j] = (v[j]*alpha + v1_const)>>16;
+        }
+        u += stride;
+        v += stride;
+    }
+}
+
+/* Blend only the edge of the macro block.  Leave center
+ * unblended to allow for other visualizations to be layered.
+ */
+void vp8_blend_mb_outer_c (unsigned char *y, unsigned char *u, unsigned char *v,
+                        int y_1, int u_1, int v_1, int alpha, int stride)
+{
+    int i, j;
+    int y1_const = y_1*((1<<16)-alpha);
+    int u1_const = u_1*((1<<16)-alpha);
+    int v1_const = v_1*((1<<16)-alpha);
+
+    for (i = 0; i < 2; i++)
+    {
+        for (j = 0; j < 16; j++)
+        {
+            y[j] = (y[j]*alpha + y1_const)>>16;
+        }
+        y += stride;
+    }
+
+    for (i = 0; i < 12; i++)
+    {
+        y[0]  = (y[0]*alpha  + y1_const)>>16;
+        y[1]  = (y[1]*alpha  + y1_const)>>16;
+        y[14] = (y[14]*alpha + y1_const)>>16;
+        y[15] = (y[15]*alpha + y1_const)>>16;
+        y += stride;
+    }
+
+    for (i = 0; i < 2; i++)
+    {
+        for (j = 0; j < 16; j++)
+        {
+            y[j] = (y[j]*alpha + y1_const)>>16;
+        }
+        y += stride;
+    }
+
+    stride >>= 1;
+
+    for (j = 0; j < 8; j++)
+    {
+        u[j] = (u[j]*alpha + u1_const)>>16;
+        v[j] = (v[j]*alpha + v1_const)>>16;
+    }
+    u += stride;
+    v += stride;
+
+    for (i = 0; i < 6; i++)
+    {
+        u[0] = (u[0]*alpha + u1_const)>>16;
+        v[0] = (v[0]*alpha + v1_const)>>16;
+
+        u[7] = (u[7]*alpha + u1_const)>>16;
+        v[7] = (v[7]*alpha + v1_const)>>16;
+
+        u += stride;
+        v += stride;
+    }
+
+    for (j = 0; j < 8; j++)
+    {
+        u[j] = (u[j]*alpha + u1_const)>>16;
+        v[j] = (v[j]*alpha + v1_const)>>16;
+    }
+}
+
+void vp8_blend_b_c (unsigned char *y, unsigned char *u, unsigned char *v,
+                        int y_1, int u_1, int v_1, int alpha, int stride)
+{
+    int i, j;
+    int y1_const = y_1*((1<<16)-alpha);
+    int u1_const = u_1*((1<<16)-alpha);
+    int v1_const = v_1*((1<<16)-alpha);
+
+    for (i = 0; i < 4; i++)
+    {
+        for (j = 0; j < 4; j++)
+        {
+            y[j] = (y[j]*alpha + y1_const)>>16;
+        }
+        y += stride;
+    }
+
+    stride >>= 1;
+
+    for (i = 0; i < 2; i++)
+    {
+        for (j = 0; j < 2; j++)
+        {
+            u[j] = (u[j]*alpha + u1_const)>>16;
+            v[j] = (v[j]*alpha + v1_const)>>16;
+        }
+        u += stride;
+        v += stride;
+    }
+}
+
+#if CONFIG_POSTPROC_VISUALIZER
+static void constrain_line (int x_0, int *x_1, int y_0, int *y_1, int width, int height)
+{
+    int dx;
+    int dy;
+
+    if (*x_1 > width)
+    {
+        dx = *x_1 - x_0;
+        dy = *y_1 - y_0;
+
+        *x_1 = width;
+        if (dx)
+            *y_1 = ((width-x_0)*dy)/dx + y_0;
+    }
+    if (*x_1 < 0)
+    {
+        dx = *x_1 - x_0;
+        dy = *y_1 - y_0;
+
+        *x_1 = 0;
+        if (dx)
+            *y_1 = ((0-x_0)*dy)/dx + y_0;
+    }
+    if (*y_1 > height)
+    {
+        dx = *x_1 - x_0;
+        dy = *y_1 - y_0;
+
+        *y_1 = height;
+        if (dy)
+            *x_1 = ((height-y_0)*dx)/dy + x_0;
+    }
+    if (*y_1 < 0)
+    {
+        dx = *x_1 - x_0;
+        dy = *y_1 - y_0;
+
+        *y_1 = 0;
+        if (dy)
+            *x_1 = ((0-y_0)*dx)/dy + x_0;
+    }
+}
+#endif  // CONFIG_POSTPROC_VISUALIZER
+
+#if CONFIG_POSTPROC
+int vp8_post_proc_frame(VP8_COMMON *oci, YV12_BUFFER_CONFIG *dest, vp8_ppflags_t *ppflags)
+{
+    int q = oci->filter_level * 10 / 6;
+    int flags = ppflags->post_proc_flag;
+    int deblock_level = ppflags->deblocking_level;
+    int noise_level = ppflags->noise_level;
+
+    if (!oci->frame_to_show)
+        return -1;
+
+    if (q > 63)
+        q = 63;
+
+    if (!flags)
+    {
+        *dest = *oci->frame_to_show;
+
+        /* handle problem with extending borders */
+        dest->y_width = oci->Width;
+        dest->y_height = oci->Height;
+        dest->uv_height = dest->y_height / 2;
+        oci->postproc_state.last_base_qindex = oci->base_qindex;
+        oci->postproc_state.last_frame_valid = 1;
+        return 0;
+    }
+
+    /* Allocate post_proc_buffer_int if needed */
+    if ((flags & VP8D_MFQE) && !oci->post_proc_buffer_int_used)
+    {
+        if ((flags & VP8D_DEBLOCK) || (flags & VP8D_DEMACROBLOCK))
+        {
+            int width = (oci->Width + 15) & ~15;
+            int height = (oci->Height + 15) & ~15;
+
+            if (vp8_yv12_alloc_frame_buffer(&oci->post_proc_buffer_int,
+                                            width, height, VP8BORDERINPIXELS))
+                vpx_internal_error(&oci->error, VPX_CODEC_MEM_ERROR,
+                                   "Failed to allocate MFQE framebuffer");
+
+            oci->post_proc_buffer_int_used = 1;
+
+            /* insure that postproc is set to all 0's so that post proc
+             * doesn't pull random data in from edge
+             */
+            memset((&oci->post_proc_buffer_int)->buffer_alloc,128,(&oci->post_proc_buffer)->frame_size);
+
+        }
+    }
+
+    vp8_clear_system_state();
+
+    if ((flags & VP8D_MFQE) &&
+         oci->postproc_state.last_frame_valid &&
+         oci->current_video_frame >= 2 &&
+         oci->postproc_state.last_base_qindex < 60 &&
+         oci->base_qindex - oci->postproc_state.last_base_qindex >= 20)
+    {
+        vp8_multiframe_quality_enhance(oci);
+        if (((flags & VP8D_DEBLOCK) || (flags & VP8D_DEMACROBLOCK)) &&
+            oci->post_proc_buffer_int_used)
+        {
+            vp8_yv12_copy_frame(&oci->post_proc_buffer, &oci->post_proc_buffer_int);
+            if (flags & VP8D_DEMACROBLOCK)
+            {
+                vp8_deblock(oci, &oci->post_proc_buffer_int, &oci->post_proc_buffer,
+                                               q + (deblock_level - 5) * 10, 1, 0);
+                vp8_de_mblock(&oci->post_proc_buffer,
+                              q + (deblock_level - 5) * 10);
+            }
+            else if (flags & VP8D_DEBLOCK)
+            {
+                vp8_deblock(oci, &oci->post_proc_buffer_int, &oci->post_proc_buffer,
+                            q, 1, 0);
+            }
+        }
+        /* Move partially towards the base q of the previous frame */
+        oci->postproc_state.last_base_qindex = (3*oci->postproc_state.last_base_qindex + oci->base_qindex)>>2;
+    }
+    else if (flags & VP8D_DEMACROBLOCK)
+    {
+        vp8_deblock(oci, oci->frame_to_show, &oci->post_proc_buffer,
+                                     q + (deblock_level - 5) * 10, 1, 0);
+        vp8_de_mblock(&oci->post_proc_buffer, q + (deblock_level - 5) * 10);
+
+        oci->postproc_state.last_base_qindex = oci->base_qindex;
+    }
+    else if (flags & VP8D_DEBLOCK)
+    {
+        vp8_deblock(oci, oci->frame_to_show, &oci->post_proc_buffer,
+                    q, 1, 0);
+        oci->postproc_state.last_base_qindex = oci->base_qindex;
+    }
+    else
+    {
+        vp8_yv12_copy_frame(oci->frame_to_show, &oci->post_proc_buffer);
+        oci->postproc_state.last_base_qindex = oci->base_qindex;
+    }
+    oci->postproc_state.last_frame_valid = 1;
+
+    if (flags & VP8D_ADDNOISE)
+    {
+        if (oci->postproc_state.last_q != q
+            || oci->postproc_state.last_noise != noise_level)
+        {
+            fillrd(&oci->postproc_state, 63 - q, noise_level);
+        }
+
+        vpx_plane_add_noise
+        (oci->post_proc_buffer.y_buffer,
+         oci->postproc_state.noise,
+         oci->postproc_state.blackclamp,
+         oci->postproc_state.whiteclamp,
+         oci->postproc_state.bothclamp,
+         oci->post_proc_buffer.y_width, oci->post_proc_buffer.y_height,
+         oci->post_proc_buffer.y_stride);
+    }
+
+#if CONFIG_POSTPROC_VISUALIZER
+    if (flags & VP8D_DEBUG_TXT_FRAME_INFO)
+    {
+        char message[512];
+        sprintf(message, "F%1dG%1dQ%3dF%3dP%d_s%dx%d",
+                (oci->frame_type == KEY_FRAME),
+                oci->refresh_golden_frame,
+                oci->base_qindex,
+                oci->filter_level,
+                flags,
+                oci->mb_cols, oci->mb_rows);
+        vp8_blit_text(message, oci->post_proc_buffer.y_buffer, oci->post_proc_buffer.y_stride);
+    }
+
+    if (flags & VP8D_DEBUG_TXT_MBLK_MODES)
+    {
+        int i, j;
+        unsigned char *y_ptr;
+        YV12_BUFFER_CONFIG *post = &oci->post_proc_buffer;
+        int mb_rows = post->y_height >> 4;
+        int mb_cols = post->y_width  >> 4;
+        int mb_index = 0;
+        MODE_INFO *mi = oci->mi;
+
+        y_ptr = post->y_buffer + 4 * post->y_stride + 4;
+
+        /* vp8_filter each macro block */
+        for (i = 0; i < mb_rows; i++)
+        {
+            for (j = 0; j < mb_cols; j++)
+            {
+                char zz[4];
+
+                sprintf(zz, "%c", mi[mb_index].mbmi.mode + 'a');
+
+                vp8_blit_text(zz, y_ptr, post->y_stride);
+                mb_index ++;
+                y_ptr += 16;
+            }
+
+            mb_index ++; /* border */
+            y_ptr += post->y_stride  * 16 - post->y_width;
+
+        }
+    }
+
+    if (flags & VP8D_DEBUG_TXT_DC_DIFF)
+    {
+        int i, j;
+        unsigned char *y_ptr;
+        YV12_BUFFER_CONFIG *post = &oci->post_proc_buffer;
+        int mb_rows = post->y_height >> 4;
+        int mb_cols = post->y_width  >> 4;
+        int mb_index = 0;
+        MODE_INFO *mi = oci->mi;
+
+        y_ptr = post->y_buffer + 4 * post->y_stride + 4;
+
+        /* vp8_filter each macro block */
+        for (i = 0; i < mb_rows; i++)
+        {
+            for (j = 0; j < mb_cols; j++)
+            {
+                char zz[4];
+                int dc_diff = !(mi[mb_index].mbmi.mode != B_PRED &&
+                              mi[mb_index].mbmi.mode != SPLITMV &&
+                              mi[mb_index].mbmi.mb_skip_coeff);
+
+                if (oci->frame_type == KEY_FRAME)
+                    sprintf(zz, "a");
+                else
+                    sprintf(zz, "%c", dc_diff + '0');
+
+                vp8_blit_text(zz, y_ptr, post->y_stride);
+                mb_index ++;
+                y_ptr += 16;
+            }
+
+            mb_index ++; /* border */
+            y_ptr += post->y_stride  * 16 - post->y_width;
+
+        }
+    }
+
+    if (flags & VP8D_DEBUG_TXT_RATE_INFO)
+    {
+        char message[512];
+        sprintf(message, "Bitrate: %10.2f framerate: %10.2f ", oci->bitrate, oci->framerate);
+        vp8_blit_text(message, oci->post_proc_buffer.y_buffer, oci->post_proc_buffer.y_stride);
+    }
+
+    /* Draw motion vectors */
+    if ((flags & VP8D_DEBUG_DRAW_MV) && ppflags->display_mv_flag)
+    {
+        YV12_BUFFER_CONFIG *post = &oci->post_proc_buffer;
+        int width  = post->y_width;
+        int height = post->y_height;
+        unsigned char *y_buffer = oci->post_proc_buffer.y_buffer;
+        int y_stride = oci->post_proc_buffer.y_stride;
+        MODE_INFO *mi = oci->mi;
+        int x0, y0;
+
+        for (y0 = 0; y0 < height; y0 += 16)
+        {
+            for (x0 = 0; x0 < width; x0 += 16)
+            {
+                int x1, y1;
+
+                if (!(ppflags->display_mv_flag & (1<<mi->mbmi.mode)))
+                {
+                    mi++;
+                    continue;
+                }
+
+                if (mi->mbmi.mode == SPLITMV)
+                {
+                    switch (mi->mbmi.partitioning)
+                    {
+                        case 0 :    /* mv_top_bottom */
+                        {
+                            union b_mode_info *bmi = &mi->bmi[0];
+                            MV *mv = &bmi->mv.as_mv;
+
+                            x1 = x0 + 8 + (mv->col >> 3);
+                            y1 = y0 + 4 + (mv->row >> 3);
+
+                            constrain_line (x0+8, &x1, y0+4, &y1, width, height);
+                            vp8_blit_line  (x0+8,  x1, y0+4,  y1, y_buffer, y_stride);
+
+                            bmi = &mi->bmi[8];
+
+                            x1 = x0 + 8 + (mv->col >> 3);
+                            y1 = y0 +12 + (mv->row >> 3);
+
+                            constrain_line (x0+8, &x1, y0+12, &y1, width, height);
+                            vp8_blit_line  (x0+8,  x1, y0+12,  y1, y_buffer, y_stride);
+
+                            break;
+                        }
+                        case 1 :    /* mv_left_right */
+                        {
+                            union b_mode_info *bmi = &mi->bmi[0];
+                            MV *mv = &bmi->mv.as_mv;
+
+                            x1 = x0 + 4 + (mv->col >> 3);
+                            y1 = y0 + 8 + (mv->row >> 3);
+
+                            constrain_line (x0+4, &x1, y0+8, &y1, width, height);
+                            vp8_blit_line  (x0+4,  x1, y0+8,  y1, y_buffer, y_stride);
+
+                            bmi = &mi->bmi[2];
+
+                            x1 = x0 +12 + (mv->col >> 3);
+                            y1 = y0 + 8 + (mv->row >> 3);
+
+                            constrain_line (x0+12, &x1, y0+8, &y1, width, height);
+                            vp8_blit_line  (x0+12,  x1, y0+8,  y1, y_buffer, y_stride);
+
+                            break;
+                        }
+                        case 2 :    /* mv_quarters   */
+                        {
+                            union b_mode_info *bmi = &mi->bmi[0];
+                            MV *mv = &bmi->mv.as_mv;
+
+                            x1 = x0 + 4 + (mv->col >> 3);
+                            y1 = y0 + 4 + (mv->row >> 3);
+
+                            constrain_line (x0+4, &x1, y0+4, &y1, width, height);
+                            vp8_blit_line  (x0+4,  x1, y0+4,  y1, y_buffer, y_stride);
+
+                            bmi = &mi->bmi[2];
+
+                            x1 = x0 +12 + (mv->col >> 3);
+                            y1 = y0 + 4 + (mv->row >> 3);
+
+                            constrain_line (x0+12, &x1, y0+4, &y1, width, height);
+                            vp8_blit_line  (x0+12,  x1, y0+4,  y1, y_buffer, y_stride);
+
+                            bmi = &mi->bmi[8];
+
+                            x1 = x0 + 4 + (mv->col >> 3);
+                            y1 = y0 +12 + (mv->row >> 3);
+
+                            constrain_line (x0+4, &x1, y0+12, &y1, width, height);
+                            vp8_blit_line  (x0+4,  x1, y0+12,  y1, y_buffer, y_stride);
+
+                            bmi = &mi->bmi[10];
+
+                            x1 = x0 +12 + (mv->col >> 3);
+                            y1 = y0 +12 + (mv->row >> 3);
+
+                            constrain_line (x0+12, &x1, y0+12, &y1, width, height);
+                            vp8_blit_line  (x0+12,  x1, y0+12,  y1, y_buffer, y_stride);
+                            break;
+                        }
+                        default :
+                        {
+                            union b_mode_info *bmi = mi->bmi;
+                            int bx0, by0;
+
+                            for (by0 = y0; by0 < (y0+16); by0 += 4)
+                            {
+                                for (bx0 = x0; bx0 < (x0+16); bx0 += 4)
+                                {
+                                    MV *mv = &bmi->mv.as_mv;
+
+                                    x1 = bx0 + 2 + (mv->col >> 3);
+                                    y1 = by0 + 2 + (mv->row >> 3);
+
+                                    constrain_line (bx0+2, &x1, by0+2, &y1, width, height);
+                                    vp8_blit_line  (bx0+2,  x1, by0+2,  y1, y_buffer, y_stride);
+
+                                    bmi++;
+                                }
+                            }
+                        }
+                    }
+                }
+                else if (mi->mbmi.mode >= NEARESTMV)
+                {
+                    MV *mv = &mi->mbmi.mv.as_mv;
+                    const int lx0 = x0 + 8;
+                    const int ly0 = y0 + 8;
+
+                    x1 = lx0 + (mv->col >> 3);
+                    y1 = ly0 + (mv->row >> 3);
+
+                    if (x1 != lx0 && y1 != ly0)
+                    {
+                        constrain_line (lx0, &x1, ly0-1, &y1, width, height);
+                        vp8_blit_line  (lx0,  x1, ly0-1,  y1, y_buffer, y_stride);
+
+                        constrain_line (lx0, &x1, ly0+1, &y1, width, height);
+                        vp8_blit_line  (lx0,  x1, ly0+1,  y1, y_buffer, y_stride);
+                    }
+                    else
+                        vp8_blit_line  (lx0,  x1, ly0,  y1, y_buffer, y_stride);
+                }
+
+                mi++;
+            }
+            mi++;
+        }
+    }
+
+    /* Color in block modes */
+    if ((flags & VP8D_DEBUG_CLR_BLK_MODES)
+        && (ppflags->display_mb_modes_flag || ppflags->display_b_modes_flag))
+    {
+        int y, x;
+        YV12_BUFFER_CONFIG *post = &oci->post_proc_buffer;
+        int width  = post->y_width;
+        int height = post->y_height;
+        unsigned char *y_ptr = oci->post_proc_buffer.y_buffer;
+        unsigned char *u_ptr = oci->post_proc_buffer.u_buffer;
+        unsigned char *v_ptr = oci->post_proc_buffer.v_buffer;
+        int y_stride = oci->post_proc_buffer.y_stride;
+        MODE_INFO *mi = oci->mi;
+
+        for (y = 0; y < height; y += 16)
+        {
+            for (x = 0; x < width; x += 16)
+            {
+                int Y = 0, U = 0, V = 0;
+
+                if (mi->mbmi.mode == B_PRED &&
+                    ((ppflags->display_mb_modes_flag & B_PRED) || ppflags->display_b_modes_flag))
+                {
+                    int by, bx;
+                    unsigned char *yl, *ul, *vl;
+                    union b_mode_info *bmi = mi->bmi;
+
+                    yl = y_ptr + x;
+                    ul = u_ptr + (x>>1);
+                    vl = v_ptr + (x>>1);
+
+                    for (by = 0; by < 16; by += 4)
+                    {
+                        for (bx = 0; bx < 16; bx += 4)
+                        {
+                            if ((ppflags->display_b_modes_flag & (1<<mi->mbmi.mode))
+                                || (ppflags->display_mb_modes_flag & B_PRED))
+                            {
+                                Y = B_PREDICTION_MODE_colors[bmi->as_mode][0];
+                                U = B_PREDICTION_MODE_colors[bmi->as_mode][1];
+                                V = B_PREDICTION_MODE_colors[bmi->as_mode][2];
+
+                                vp8_blend_b
+                                    (yl+bx, ul+(bx>>1), vl+(bx>>1), Y, U, V, 0xc000, y_stride);
+                            }
+                            bmi++;
+                        }
+
+                        yl += y_stride*4;
+                        ul += y_stride*1;
+                        vl += y_stride*1;
+                    }
+                }
+                else if (ppflags->display_mb_modes_flag & (1<<mi->mbmi.mode))
+                {
+                    Y = MB_PREDICTION_MODE_colors[mi->mbmi.mode][0];
+                    U = MB_PREDICTION_MODE_colors[mi->mbmi.mode][1];
+                    V = MB_PREDICTION_MODE_colors[mi->mbmi.mode][2];
+
+                    vp8_blend_mb_inner
+                        (y_ptr+x, u_ptr+(x>>1), v_ptr+(x>>1), Y, U, V, 0xc000, y_stride);
+                }
+
+                mi++;
+            }
+            y_ptr += y_stride*16;
+            u_ptr += y_stride*4;
+            v_ptr += y_stride*4;
+
+            mi++;
+        }
+    }
+
+    /* Color in frame reference blocks */
+    if ((flags & VP8D_DEBUG_CLR_FRM_REF_BLKS) && ppflags->display_ref_frame_flag)
+    {
+        int y, x;
+        YV12_BUFFER_CONFIG *post = &oci->post_proc_buffer;
+        int width  = post->y_width;
+        int height = post->y_height;
+        unsigned char *y_ptr = oci->post_proc_buffer.y_buffer;
+        unsigned char *u_ptr = oci->post_proc_buffer.u_buffer;
+        unsigned char *v_ptr = oci->post_proc_buffer.v_buffer;
+        int y_stride = oci->post_proc_buffer.y_stride;
+        MODE_INFO *mi = oci->mi;
+
+        for (y = 0; y < height; y += 16)
+        {
+            for (x = 0; x < width; x +=16)
+            {
+                int Y = 0, U = 0, V = 0;
+
+                if (ppflags->display_ref_frame_flag & (1<<mi->mbmi.ref_frame))
+                {
+                    Y = MV_REFERENCE_FRAME_colors[mi->mbmi.ref_frame][0];
+                    U = MV_REFERENCE_FRAME_colors[mi->mbmi.ref_frame][1];
+                    V = MV_REFERENCE_FRAME_colors[mi->mbmi.ref_frame][2];
+
+                    vp8_blend_mb_outer
+                        (y_ptr+x, u_ptr+(x>>1), v_ptr+(x>>1), Y, U, V, 0xc000, y_stride);
+                }
+
+                mi++;
+            }
+            y_ptr += y_stride*16;
+            u_ptr += y_stride*4;
+            v_ptr += y_stride*4;
+
+            mi++;
+        }
+    }
+#endif
+
+    *dest = oci->post_proc_buffer;
+
+    /* handle problem with extending borders */
+    dest->y_width = oci->Width;
+    dest->y_height = oci->Height;
+    dest->uv_height = dest->y_height / 2;
+    return 0;
+}
+#endif
diff --git a/libvpx/libvpx/vp8/common/postproc.h b/libvpx/libvpx/vp8/common/postproc.h
new file mode 100644
index 0000000..0fa12a7
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/postproc.h
@@ -0,0 +1,59 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_POSTPROC_H_
+#define VP8_COMMON_POSTPROC_H_
+
+#include "vpx_ports/mem.h"
+struct postproc_state
+{
+    int           last_q;
+    int           last_noise;
+    char          noise[3072];
+    int           last_base_qindex;
+    int           last_frame_valid;
+    DECLARE_ALIGNED(16, char, blackclamp[16]);
+    DECLARE_ALIGNED(16, char, whiteclamp[16]);
+    DECLARE_ALIGNED(16, char, bothclamp[16]);
+};
+#include "onyxc_int.h"
+#include "ppflags.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+int vp8_post_proc_frame(struct VP8Common *oci, YV12_BUFFER_CONFIG *dest,
+                        vp8_ppflags_t *flags);
+
+
+void vp8_de_noise(struct VP8Common           *oci,
+                  YV12_BUFFER_CONFIG         *source,
+                  YV12_BUFFER_CONFIG         *post,
+                  int                         q,
+                  int                         low_var_thresh,
+                  int                         flag,
+                  int                         uvfilter);
+
+void vp8_deblock(struct VP8Common           *oci,
+                 YV12_BUFFER_CONFIG         *source,
+                 YV12_BUFFER_CONFIG         *post,
+                 int                         q,
+                 int                         low_var_thresh,
+                 int                         flag);
+
+#define MFQE_PRECISION 4
+
+void vp8_multiframe_quality_enhance(struct VP8Common *cm);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_POSTPROC_H_
diff --git a/libvpx/libvpx/vp8/common/ppflags.h b/libvpx/libvpx/vp8/common/ppflags.h
new file mode 100644
index 0000000..768224a
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/ppflags.h
@@ -0,0 +1,49 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_PPFLAGS_H_
+#define VP8_COMMON_PPFLAGS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+enum
+{
+    VP8D_NOFILTERING            = 0,
+    VP8D_DEBLOCK                = 1<<0,
+    VP8D_DEMACROBLOCK           = 1<<1,
+    VP8D_ADDNOISE               = 1<<2,
+    VP8D_DEBUG_TXT_FRAME_INFO   = 1<<3,
+    VP8D_DEBUG_TXT_MBLK_MODES   = 1<<4,
+    VP8D_DEBUG_TXT_DC_DIFF      = 1<<5,
+    VP8D_DEBUG_TXT_RATE_INFO    = 1<<6,
+    VP8D_DEBUG_DRAW_MV          = 1<<7,
+    VP8D_DEBUG_CLR_BLK_MODES    = 1<<8,
+    VP8D_DEBUG_CLR_FRM_REF_BLKS = 1<<9,
+    VP8D_MFQE                   = 1<<10
+};
+
+typedef struct
+{
+    int post_proc_flag;
+    int deblocking_level;
+    int noise_level;
+    int display_ref_frame_flag;
+    int display_mb_modes_flag;
+    int display_b_modes_flag;
+    int display_mv_flag;
+} vp8_ppflags_t;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_PPFLAGS_H_
diff --git a/libvpx/libvpx/vp8/common/quant_common.c b/libvpx/libvpx/vp8/common/quant_common.c
new file mode 100644
index 0000000..05f9210
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/quant_common.c
@@ -0,0 +1,135 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "quant_common.h"
+
+static const int dc_qlookup[QINDEX_RANGE] =
+{
+    4,    5,    6,    7,    8,    9,   10,   10,   11,   12,   13,   14,   15,   16,   17,   17,
+    18,   19,   20,   20,   21,   21,   22,   22,   23,   23,   24,   25,   25,   26,   27,   28,
+    29,   30,   31,   32,   33,   34,   35,   36,   37,   37,   38,   39,   40,   41,   42,   43,
+    44,   45,   46,   46,   47,   48,   49,   50,   51,   52,   53,   54,   55,   56,   57,   58,
+    59,   60,   61,   62,   63,   64,   65,   66,   67,   68,   69,   70,   71,   72,   73,   74,
+    75,   76,   76,   77,   78,   79,   80,   81,   82,   83,   84,   85,   86,   87,   88,   89,
+    91,   93,   95,   96,   98,  100,  101,  102,  104,  106,  108,  110,  112,  114,  116,  118,
+    122,  124,  126,  128,  130,  132,  134,  136,  138,  140,  143,  145,  148,  151,  154,  157,
+};
+
+static const int ac_qlookup[QINDEX_RANGE] =
+{
+    4,    5,    6,    7,    8,    9,   10,   11,   12,   13,   14,   15,   16,   17,   18,   19,
+    20,   21,   22,   23,   24,   25,   26,   27,   28,   29,   30,   31,   32,   33,   34,   35,
+    36,   37,   38,   39,   40,   41,   42,   43,   44,   45,   46,   47,   48,   49,   50,   51,
+    52,   53,   54,   55,   56,   57,   58,   60,   62,   64,   66,   68,   70,   72,   74,   76,
+    78,   80,   82,   84,   86,   88,   90,   92,   94,   96,   98,  100,  102,  104,  106,  108,
+    110,  112,  114,  116,  119,  122,  125,  128,  131,  134,  137,  140,  143,  146,  149,  152,
+    155,  158,  161,  164,  167,  170,  173,  177,  181,  185,  189,  193,  197,  201,  205,  209,
+    213,  217,  221,  225,  229,  234,  239,  245,  249,  254,  259,  264,  269,  274,  279,  284,
+};
+
+
+int vp8_dc_quant(int QIndex, int Delta)
+{
+    int retval;
+
+    QIndex = QIndex + Delta;
+
+    if (QIndex > 127)
+        QIndex = 127;
+    else if (QIndex < 0)
+        QIndex = 0;
+
+    retval = dc_qlookup[ QIndex ];
+    return retval;
+}
+
+int vp8_dc2quant(int QIndex, int Delta)
+{
+    int retval;
+
+    QIndex = QIndex + Delta;
+
+    if (QIndex > 127)
+        QIndex = 127;
+    else if (QIndex < 0)
+        QIndex = 0;
+
+    retval = dc_qlookup[ QIndex ] * 2;
+    return retval;
+
+}
+int vp8_dc_uv_quant(int QIndex, int Delta)
+{
+    int retval;
+
+    QIndex = QIndex + Delta;
+
+    if (QIndex > 127)
+        QIndex = 127;
+    else if (QIndex < 0)
+        QIndex = 0;
+
+    retval = dc_qlookup[ QIndex ];
+
+    if (retval > 132)
+        retval = 132;
+
+    return retval;
+}
+
+int vp8_ac_yquant(int QIndex)
+{
+    int retval;
+
+    if (QIndex > 127)
+        QIndex = 127;
+    else if (QIndex < 0)
+        QIndex = 0;
+
+    retval = ac_qlookup[ QIndex ];
+    return retval;
+}
+
+int vp8_ac2quant(int QIndex, int Delta)
+{
+    int retval;
+
+    QIndex = QIndex + Delta;
+
+    if (QIndex > 127)
+        QIndex = 127;
+    else if (QIndex < 0)
+        QIndex = 0;
+
+    /* For all x in [0..284], x*155/100 is bitwise equal to (x*101581) >> 16.
+     * The smallest precision for that is '(x*6349) >> 12' but 16 is a good
+     * word size. */
+    retval = (ac_qlookup[ QIndex ] * 101581) >> 16;
+
+    if (retval < 8)
+        retval = 8;
+
+    return retval;
+}
+int vp8_ac_uv_quant(int QIndex, int Delta)
+{
+    int retval;
+
+    QIndex = QIndex + Delta;
+
+    if (QIndex > 127)
+        QIndex = 127;
+    else if (QIndex < 0)
+        QIndex = 0;
+
+    retval = ac_qlookup[ QIndex ];
+    return retval;
+}
diff --git a/libvpx/libvpx/vp8/common/quant_common.h b/libvpx/libvpx/vp8/common/quant_common.h
new file mode 100644
index 0000000..700b5e6
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/quant_common.h
@@ -0,0 +1,34 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_COMMON_QUANT_COMMON_H_
+#define VP8_COMMON_QUANT_COMMON_H_
+
+
+#include "string.h"
+#include "blockd.h"
+#include "onyxc_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern int vp8_ac_yquant(int QIndex);
+extern int vp8_dc_quant(int QIndex, int Delta);
+extern int vp8_dc2quant(int QIndex, int Delta);
+extern int vp8_ac2quant(int QIndex, int Delta);
+extern int vp8_dc_uv_quant(int QIndex, int Delta);
+extern int vp8_ac_uv_quant(int QIndex, int Delta);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_QUANT_COMMON_H_
diff --git a/libvpx/libvpx/vp8/common/reconinter.c b/libvpx/libvpx/vp8/common/reconinter.c
new file mode 100644
index 0000000..e302595
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/reconinter.c
@@ -0,0 +1,544 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <limits.h>
+#include <string.h>
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "blockd.h"
+#include "reconinter.h"
+#if CONFIG_RUNTIME_CPU_DETECT
+#include "onyxc_int.h"
+#endif
+
+void vp8_copy_mem16x16_c(
+    unsigned char *src,
+    int src_stride,
+    unsigned char *dst,
+    int dst_stride)
+{
+
+    int r;
+
+    for (r = 0; r < 16; r++)
+    {
+        memcpy(dst, src, 16);
+
+        src += src_stride;
+        dst += dst_stride;
+
+    }
+
+}
+
+void vp8_copy_mem8x8_c(
+    unsigned char *src,
+    int src_stride,
+    unsigned char *dst,
+    int dst_stride)
+{
+    int r;
+
+    for (r = 0; r < 8; r++)
+    {
+        memcpy(dst, src, 8);
+
+        src += src_stride;
+        dst += dst_stride;
+
+    }
+
+}
+
+void vp8_copy_mem8x4_c(
+    unsigned char *src,
+    int src_stride,
+    unsigned char *dst,
+    int dst_stride)
+{
+    int r;
+
+    for (r = 0; r < 4; r++)
+    {
+        memcpy(dst, src, 8);
+
+        src += src_stride;
+        dst += dst_stride;
+
+    }
+
+}
+
+
+void vp8_build_inter_predictors_b(BLOCKD *d, int pitch, unsigned char *base_pre, int pre_stride, vp8_subpix_fn_t sppf)
+{
+    int r;
+    unsigned char *pred_ptr = d->predictor;
+    unsigned char *ptr;
+    ptr = base_pre + d->offset + (d->bmi.mv.as_mv.row >> 3) * pre_stride + (d->bmi.mv.as_mv.col >> 3);
+
+    if (d->bmi.mv.as_mv.row & 7 || d->bmi.mv.as_mv.col & 7)
+    {
+        sppf(ptr, pre_stride, d->bmi.mv.as_mv.col & 7, d->bmi.mv.as_mv.row & 7, pred_ptr, pitch);
+    }
+    else
+    {
+        for (r = 0; r < 4; r++)
+        {
+            pred_ptr[0]  = ptr[0];
+            pred_ptr[1]  = ptr[1];
+            pred_ptr[2]  = ptr[2];
+            pred_ptr[3]  = ptr[3];
+            pred_ptr     += pitch;
+            ptr         += pre_stride;
+        }
+    }
+}
+
+static void build_inter_predictors4b(MACROBLOCKD *x, BLOCKD *d, unsigned char *dst, int dst_stride, unsigned char *base_pre, int pre_stride)
+{
+    unsigned char *ptr;
+    ptr = base_pre + d->offset + (d->bmi.mv.as_mv.row >> 3) * pre_stride + (d->bmi.mv.as_mv.col >> 3);
+
+    if (d->bmi.mv.as_mv.row & 7 || d->bmi.mv.as_mv.col & 7)
+    {
+        x->subpixel_predict8x8(ptr, pre_stride, d->bmi.mv.as_mv.col & 7, d->bmi.mv.as_mv.row & 7, dst, dst_stride);
+    }
+    else
+    {
+        vp8_copy_mem8x8(ptr, pre_stride, dst, dst_stride);
+    }
+}
+
+static void build_inter_predictors2b(MACROBLOCKD *x, BLOCKD *d, unsigned char *dst, int dst_stride, unsigned char *base_pre, int pre_stride)
+{
+    unsigned char *ptr;
+    ptr = base_pre + d->offset + (d->bmi.mv.as_mv.row >> 3) * pre_stride + (d->bmi.mv.as_mv.col >> 3);
+
+    if (d->bmi.mv.as_mv.row & 7 || d->bmi.mv.as_mv.col & 7)
+    {
+        x->subpixel_predict8x4(ptr, pre_stride, d->bmi.mv.as_mv.col & 7, d->bmi.mv.as_mv.row & 7, dst, dst_stride);
+    }
+    else
+    {
+        vp8_copy_mem8x4(ptr, pre_stride, dst, dst_stride);
+    }
+}
+
+static void build_inter_predictors_b(BLOCKD *d, unsigned char *dst, int dst_stride, unsigned char *base_pre, int pre_stride, vp8_subpix_fn_t sppf)
+{
+    int r;
+    unsigned char *ptr;
+    ptr = base_pre + d->offset + (d->bmi.mv.as_mv.row >> 3) * pre_stride + (d->bmi.mv.as_mv.col >> 3);
+
+    if (d->bmi.mv.as_mv.row & 7 || d->bmi.mv.as_mv.col & 7)
+    {
+        sppf(ptr, pre_stride, d->bmi.mv.as_mv.col & 7, d->bmi.mv.as_mv.row & 7, dst, dst_stride);
+    }
+    else
+    {
+        for (r = 0; r < 4; r++)
+        {
+          dst[0]  = ptr[0];
+          dst[1]  = ptr[1];
+          dst[2]  = ptr[2];
+          dst[3]  = ptr[3];
+          dst     += dst_stride;
+          ptr     += pre_stride;
+        }
+    }
+}
+
+
+/*encoder only*/
+void vp8_build_inter16x16_predictors_mbuv(MACROBLOCKD *x)
+{
+    unsigned char *uptr, *vptr;
+    unsigned char *upred_ptr = &x->predictor[256];
+    unsigned char *vpred_ptr = &x->predictor[320];
+
+    int mv_row = x->mode_info_context->mbmi.mv.as_mv.row;
+    int mv_col = x->mode_info_context->mbmi.mv.as_mv.col;
+    int offset;
+    int pre_stride = x->pre.uv_stride;
+
+    /* calc uv motion vectors */
+    mv_row += 1 | (mv_row >> (sizeof(int) * CHAR_BIT - 1));
+    mv_col += 1 | (mv_col >> (sizeof(int) * CHAR_BIT - 1));
+    mv_row /= 2;
+    mv_col /= 2;
+    mv_row &= x->fullpixel_mask;
+    mv_col &= x->fullpixel_mask;
+
+    offset = (mv_row >> 3) * pre_stride + (mv_col >> 3);
+    uptr = x->pre.u_buffer + offset;
+    vptr = x->pre.v_buffer + offset;
+
+    if ((mv_row | mv_col) & 7)
+    {
+        x->subpixel_predict8x8(uptr, pre_stride, mv_col & 7, mv_row & 7, upred_ptr, 8);
+        x->subpixel_predict8x8(vptr, pre_stride, mv_col & 7, mv_row & 7, vpred_ptr, 8);
+    }
+    else
+    {
+        vp8_copy_mem8x8(uptr, pre_stride, upred_ptr, 8);
+        vp8_copy_mem8x8(vptr, pre_stride, vpred_ptr, 8);
+    }
+}
+
+/*encoder only*/
+void vp8_build_inter4x4_predictors_mbuv(MACROBLOCKD *x)
+{
+    int i, j;
+    int pre_stride = x->pre.uv_stride;
+    unsigned char *base_pre;
+
+    /* build uv mvs */
+    for (i = 0; i < 2; i++)
+    {
+        for (j = 0; j < 2; j++)
+        {
+            int yoffset = i * 8 + j * 2;
+            int uoffset = 16 + i * 2 + j;
+            int voffset = 20 + i * 2 + j;
+
+            int temp;
+
+            temp = x->block[yoffset  ].bmi.mv.as_mv.row
+                   + x->block[yoffset+1].bmi.mv.as_mv.row
+                   + x->block[yoffset+4].bmi.mv.as_mv.row
+                   + x->block[yoffset+5].bmi.mv.as_mv.row;
+
+            temp += 4 + ((temp >> (sizeof(temp) * CHAR_BIT - 1)) * 8);
+
+            x->block[uoffset].bmi.mv.as_mv.row = (temp / 8) & x->fullpixel_mask;
+
+            temp = x->block[yoffset  ].bmi.mv.as_mv.col
+                   + x->block[yoffset+1].bmi.mv.as_mv.col
+                   + x->block[yoffset+4].bmi.mv.as_mv.col
+                   + x->block[yoffset+5].bmi.mv.as_mv.col;
+
+            temp += 4 + ((temp >> (sizeof(temp) * CHAR_BIT - 1)) * 8);
+
+            x->block[uoffset].bmi.mv.as_mv.col = (temp / 8) & x->fullpixel_mask;
+
+            x->block[voffset].bmi.mv.as_int = x->block[uoffset].bmi.mv.as_int;
+        }
+    }
+
+    base_pre = x->pre.u_buffer;
+    for (i = 16; i < 20; i += 2)
+    {
+        BLOCKD *d0 = &x->block[i];
+        BLOCKD *d1 = &x->block[i+1];
+
+        if (d0->bmi.mv.as_int == d1->bmi.mv.as_int)
+            build_inter_predictors2b(x, d0, d0->predictor, 8, base_pre, pre_stride);
+        else
+        {
+            vp8_build_inter_predictors_b(d0, 8, base_pre, pre_stride, x->subpixel_predict);
+            vp8_build_inter_predictors_b(d1, 8, base_pre, pre_stride, x->subpixel_predict);
+        }
+    }
+
+    base_pre = x->pre.v_buffer;
+    for (i = 20; i < 24; i += 2)
+    {
+        BLOCKD *d0 = &x->block[i];
+        BLOCKD *d1 = &x->block[i+1];
+
+        if (d0->bmi.mv.as_int == d1->bmi.mv.as_int)
+            build_inter_predictors2b(x, d0, d0->predictor, 8, base_pre, pre_stride);
+        else
+        {
+            vp8_build_inter_predictors_b(d0, 8, base_pre, pre_stride, x->subpixel_predict);
+            vp8_build_inter_predictors_b(d1, 8, base_pre, pre_stride, x->subpixel_predict);
+        }
+    }
+}
+
+
+/*encoder only*/
+void vp8_build_inter16x16_predictors_mby(MACROBLOCKD *x,
+                                         unsigned char *dst_y,
+                                         int dst_ystride)
+{
+    unsigned char *ptr_base;
+    unsigned char *ptr;
+    int mv_row = x->mode_info_context->mbmi.mv.as_mv.row;
+    int mv_col = x->mode_info_context->mbmi.mv.as_mv.col;
+    int pre_stride = x->pre.y_stride;
+
+    ptr_base = x->pre.y_buffer;
+    ptr = ptr_base + (mv_row >> 3) * pre_stride + (mv_col >> 3);
+
+    if ((mv_row | mv_col) & 7)
+    {
+        x->subpixel_predict16x16(ptr, pre_stride, mv_col & 7, mv_row & 7,
+                                 dst_y, dst_ystride);
+    }
+    else
+    {
+        vp8_copy_mem16x16(ptr, pre_stride, dst_y,
+            dst_ystride);
+    }
+}
+
+static void clamp_mv_to_umv_border(MV *mv, const MACROBLOCKD *xd)
+{
+    /* If the MV points so far into the UMV border that no visible pixels
+     * are used for reconstruction, the subpel part of the MV can be
+     * discarded and the MV limited to 16 pixels with equivalent results.
+     *
+     * This limit kicks in at 19 pixels for the top and left edges, for
+     * the 16 pixels plus 3 taps right of the central pixel when subpel
+     * filtering. The bottom and right edges use 16 pixels plus 2 pixels
+     * left of the central pixel when filtering.
+     */
+    if (mv->col < (xd->mb_to_left_edge - (19 << 3)))
+        mv->col = xd->mb_to_left_edge - (16 << 3);
+    else if (mv->col > xd->mb_to_right_edge + (18 << 3))
+        mv->col = xd->mb_to_right_edge + (16 << 3);
+
+    if (mv->row < (xd->mb_to_top_edge - (19 << 3)))
+        mv->row = xd->mb_to_top_edge - (16 << 3);
+    else if (mv->row > xd->mb_to_bottom_edge + (18 << 3))
+        mv->row = xd->mb_to_bottom_edge + (16 << 3);
+}
+
+/* A version of the above function for chroma block MVs.*/
+static void clamp_uvmv_to_umv_border(MV *mv, const MACROBLOCKD *xd)
+{
+    mv->col = (2*mv->col < (xd->mb_to_left_edge - (19 << 3))) ?
+        (xd->mb_to_left_edge - (16 << 3)) >> 1 : mv->col;
+    mv->col = (2*mv->col > xd->mb_to_right_edge + (18 << 3)) ?
+        (xd->mb_to_right_edge + (16 << 3)) >> 1 : mv->col;
+
+    mv->row = (2*mv->row < (xd->mb_to_top_edge - (19 << 3))) ?
+        (xd->mb_to_top_edge - (16 << 3)) >> 1 : mv->row;
+    mv->row = (2*mv->row > xd->mb_to_bottom_edge + (18 << 3)) ?
+        (xd->mb_to_bottom_edge + (16 << 3)) >> 1 : mv->row;
+}
+
+void vp8_build_inter16x16_predictors_mb(MACROBLOCKD *x,
+                                        unsigned char *dst_y,
+                                        unsigned char *dst_u,
+                                        unsigned char *dst_v,
+                                        int dst_ystride,
+                                        int dst_uvstride)
+{
+    int offset;
+    unsigned char *ptr;
+    unsigned char *uptr, *vptr;
+
+    int_mv _16x16mv;
+
+    unsigned char *ptr_base = x->pre.y_buffer;
+    int pre_stride = x->pre.y_stride;
+
+    _16x16mv.as_int = x->mode_info_context->mbmi.mv.as_int;
+
+    if (x->mode_info_context->mbmi.need_to_clamp_mvs)
+    {
+        clamp_mv_to_umv_border(&_16x16mv.as_mv, x);
+    }
+
+    ptr = ptr_base + ( _16x16mv.as_mv.row >> 3) * pre_stride + (_16x16mv.as_mv.col >> 3);
+
+    if ( _16x16mv.as_int & 0x00070007)
+    {
+        x->subpixel_predict16x16(ptr, pre_stride, _16x16mv.as_mv.col & 7,  _16x16mv.as_mv.row & 7, dst_y, dst_ystride);
+    }
+    else
+    {
+        vp8_copy_mem16x16(ptr, pre_stride, dst_y, dst_ystride);
+    }
+
+    /* calc uv motion vectors */
+    _16x16mv.as_mv.row += 1 | (_16x16mv.as_mv.row >> (sizeof(int) * CHAR_BIT - 1));
+    _16x16mv.as_mv.col += 1 | (_16x16mv.as_mv.col >> (sizeof(int) * CHAR_BIT - 1));
+    _16x16mv.as_mv.row /= 2;
+    _16x16mv.as_mv.col /= 2;
+    _16x16mv.as_mv.row &= x->fullpixel_mask;
+    _16x16mv.as_mv.col &= x->fullpixel_mask;
+
+    pre_stride >>= 1;
+    offset = ( _16x16mv.as_mv.row >> 3) * pre_stride + (_16x16mv.as_mv.col >> 3);
+    uptr = x->pre.u_buffer + offset;
+    vptr = x->pre.v_buffer + offset;
+
+    if ( _16x16mv.as_int & 0x00070007)
+    {
+        x->subpixel_predict8x8(uptr, pre_stride, _16x16mv.as_mv.col & 7,  _16x16mv.as_mv.row & 7, dst_u, dst_uvstride);
+        x->subpixel_predict8x8(vptr, pre_stride, _16x16mv.as_mv.col & 7,  _16x16mv.as_mv.row & 7, dst_v, dst_uvstride);
+    }
+    else
+    {
+        vp8_copy_mem8x8(uptr, pre_stride, dst_u, dst_uvstride);
+        vp8_copy_mem8x8(vptr, pre_stride, dst_v, dst_uvstride);
+    }
+}
+
+static void build_inter4x4_predictors_mb(MACROBLOCKD *x)
+{
+    int i;
+    unsigned char *base_dst = x->dst.y_buffer;
+    unsigned char *base_pre = x->pre.y_buffer;
+
+    if (x->mode_info_context->mbmi.partitioning < 3)
+    {
+        BLOCKD *b;
+        int dst_stride = x->dst.y_stride;
+
+        x->block[ 0].bmi = x->mode_info_context->bmi[ 0];
+        x->block[ 2].bmi = x->mode_info_context->bmi[ 2];
+        x->block[ 8].bmi = x->mode_info_context->bmi[ 8];
+        x->block[10].bmi = x->mode_info_context->bmi[10];
+        if (x->mode_info_context->mbmi.need_to_clamp_mvs)
+        {
+            clamp_mv_to_umv_border(&x->block[ 0].bmi.mv.as_mv, x);
+            clamp_mv_to_umv_border(&x->block[ 2].bmi.mv.as_mv, x);
+            clamp_mv_to_umv_border(&x->block[ 8].bmi.mv.as_mv, x);
+            clamp_mv_to_umv_border(&x->block[10].bmi.mv.as_mv, x);
+        }
+
+        b = &x->block[ 0];
+        build_inter_predictors4b(x, b, base_dst + b->offset, dst_stride, base_pre, dst_stride);
+        b = &x->block[ 2];
+        build_inter_predictors4b(x, b, base_dst + b->offset, dst_stride, base_pre, dst_stride);
+        b = &x->block[ 8];
+        build_inter_predictors4b(x, b, base_dst + b->offset, dst_stride, base_pre, dst_stride);
+        b = &x->block[10];
+        build_inter_predictors4b(x, b, base_dst + b->offset, dst_stride, base_pre, dst_stride);
+    }
+    else
+    {
+        for (i = 0; i < 16; i += 2)
+        {
+            BLOCKD *d0 = &x->block[i];
+            BLOCKD *d1 = &x->block[i+1];
+            int dst_stride = x->dst.y_stride;
+
+            x->block[i+0].bmi = x->mode_info_context->bmi[i+0];
+            x->block[i+1].bmi = x->mode_info_context->bmi[i+1];
+            if (x->mode_info_context->mbmi.need_to_clamp_mvs)
+            {
+                clamp_mv_to_umv_border(&x->block[i+0].bmi.mv.as_mv, x);
+                clamp_mv_to_umv_border(&x->block[i+1].bmi.mv.as_mv, x);
+            }
+
+            if (d0->bmi.mv.as_int == d1->bmi.mv.as_int)
+                build_inter_predictors2b(x, d0, base_dst + d0->offset, dst_stride, base_pre, dst_stride);
+            else
+            {
+                build_inter_predictors_b(d0, base_dst + d0->offset, dst_stride, base_pre, dst_stride, x->subpixel_predict);
+                build_inter_predictors_b(d1, base_dst + d1->offset, dst_stride, base_pre, dst_stride, x->subpixel_predict);
+            }
+
+        }
+
+    }
+    base_dst = x->dst.u_buffer;
+    base_pre = x->pre.u_buffer;
+    for (i = 16; i < 20; i += 2)
+    {
+        BLOCKD *d0 = &x->block[i];
+        BLOCKD *d1 = &x->block[i+1];
+        int dst_stride = x->dst.uv_stride;
+
+        /* Note: uv mvs already clamped in build_4x4uvmvs() */
+
+        if (d0->bmi.mv.as_int == d1->bmi.mv.as_int)
+            build_inter_predictors2b(x, d0, base_dst + d0->offset, dst_stride, base_pre, dst_stride);
+        else
+        {
+            build_inter_predictors_b(d0, base_dst + d0->offset, dst_stride, base_pre, dst_stride, x->subpixel_predict);
+            build_inter_predictors_b(d1, base_dst + d1->offset, dst_stride, base_pre, dst_stride, x->subpixel_predict);
+        }
+    }
+
+    base_dst = x->dst.v_buffer;
+    base_pre = x->pre.v_buffer;
+    for (i = 20; i < 24; i += 2)
+    {
+        BLOCKD *d0 = &x->block[i];
+        BLOCKD *d1 = &x->block[i+1];
+        int dst_stride = x->dst.uv_stride;
+
+        /* Note: uv mvs already clamped in build_4x4uvmvs() */
+
+        if (d0->bmi.mv.as_int == d1->bmi.mv.as_int)
+            build_inter_predictors2b(x, d0, base_dst + d0->offset, dst_stride, base_pre, dst_stride);
+        else
+        {
+            build_inter_predictors_b(d0, base_dst + d0->offset, dst_stride, base_pre, dst_stride, x->subpixel_predict);
+            build_inter_predictors_b(d1, base_dst + d1->offset, dst_stride, base_pre, dst_stride, x->subpixel_predict);
+        }
+    }
+}
+
+static
+void build_4x4uvmvs(MACROBLOCKD *x)
+{
+    int i, j;
+
+    for (i = 0; i < 2; i++)
+    {
+        for (j = 0; j < 2; j++)
+        {
+            int yoffset = i * 8 + j * 2;
+            int uoffset = 16 + i * 2 + j;
+            int voffset = 20 + i * 2 + j;
+
+            int temp;
+
+            temp = x->mode_info_context->bmi[yoffset + 0].mv.as_mv.row
+                 + x->mode_info_context->bmi[yoffset + 1].mv.as_mv.row
+                 + x->mode_info_context->bmi[yoffset + 4].mv.as_mv.row
+                 + x->mode_info_context->bmi[yoffset + 5].mv.as_mv.row;
+
+            temp += 4 + ((temp >> (sizeof(temp) * CHAR_BIT - 1)) * 8);
+
+            x->block[uoffset].bmi.mv.as_mv.row = (temp / 8) & x->fullpixel_mask;
+
+            temp = x->mode_info_context->bmi[yoffset + 0].mv.as_mv.col
+                 + x->mode_info_context->bmi[yoffset + 1].mv.as_mv.col
+                 + x->mode_info_context->bmi[yoffset + 4].mv.as_mv.col
+                 + x->mode_info_context->bmi[yoffset + 5].mv.as_mv.col;
+
+            temp += 4 + ((temp >> (sizeof(temp) * CHAR_BIT - 1)) * 8);
+
+            x->block[uoffset].bmi.mv.as_mv.col = (temp / 8) & x->fullpixel_mask;
+
+            if (x->mode_info_context->mbmi.need_to_clamp_mvs)
+                clamp_uvmv_to_umv_border(&x->block[uoffset].bmi.mv.as_mv, x);
+
+            x->block[voffset].bmi.mv.as_int = x->block[uoffset].bmi.mv.as_int;
+        }
+    }
+}
+
+void vp8_build_inter_predictors_mb(MACROBLOCKD *xd)
+{
+    if (xd->mode_info_context->mbmi.mode != SPLITMV)
+    {
+        vp8_build_inter16x16_predictors_mb(xd, xd->dst.y_buffer,
+                                           xd->dst.u_buffer, xd->dst.v_buffer,
+                                           xd->dst.y_stride, xd->dst.uv_stride);
+    }
+    else
+    {
+        build_4x4uvmvs(xd);
+        build_inter4x4_predictors_mb(xd);
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/reconinter.h b/libvpx/libvpx/vp8/common/reconinter.h
new file mode 100644
index 0000000..ba979b9
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/reconinter.h
@@ -0,0 +1,43 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_RECONINTER_H_
+#define VP8_COMMON_RECONINTER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern void vp8_build_inter_predictors_mb(MACROBLOCKD *x);
+extern void vp8_build_inter16x16_predictors_mb(MACROBLOCKD *x,
+                                               unsigned char *dst_y,
+                                               unsigned char *dst_u,
+                                               unsigned char *dst_v,
+                                               int dst_ystride,
+                                               int dst_uvstride);
+
+
+extern void vp8_build_inter16x16_predictors_mby(MACROBLOCKD *x,
+                                                unsigned char *dst_y,
+                                                int dst_ystride);
+extern void vp8_build_inter_predictors_b(BLOCKD *d, int pitch,
+                                         unsigned char *base_pre,
+                                         int pre_stride,
+                                         vp8_subpix_fn_t sppf);
+
+extern void vp8_build_inter16x16_predictors_mbuv(MACROBLOCKD *x);
+extern void vp8_build_inter4x4_predictors_mbuv(MACROBLOCKD *x);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_RECONINTER_H_
diff --git a/libvpx/libvpx/vp8/common/reconintra.c b/libvpx/libvpx/vp8/common/reconintra.c
new file mode 100644
index 0000000..356655d
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/reconintra.c
@@ -0,0 +1,117 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vp8_rtcd.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/vpx_once.h"
+#include "blockd.h"
+#include "vp8/common/reconintra.h"
+#include "vp8/common/reconintra4x4.h"
+
+enum {
+    SIZE_16,
+    SIZE_8,
+    NUM_SIZES,
+};
+
+typedef void (*intra_pred_fn)(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left);
+
+static intra_pred_fn pred[4][NUM_SIZES];
+static intra_pred_fn dc_pred[2][2][NUM_SIZES];
+
+static void vp8_init_intra_predictors_internal(void)
+{
+#define INIT_SIZE(sz) \
+    pred[V_PRED][SIZE_##sz] = vpx_v_predictor_##sz##x##sz; \
+    pred[H_PRED][SIZE_##sz] = vpx_h_predictor_##sz##x##sz; \
+    pred[TM_PRED][SIZE_##sz] = vpx_tm_predictor_##sz##x##sz; \
+ \
+    dc_pred[0][0][SIZE_##sz] = vpx_dc_128_predictor_##sz##x##sz; \
+    dc_pred[0][1][SIZE_##sz] = vpx_dc_top_predictor_##sz##x##sz; \
+    dc_pred[1][0][SIZE_##sz] = vpx_dc_left_predictor_##sz##x##sz; \
+    dc_pred[1][1][SIZE_##sz] = vpx_dc_predictor_##sz##x##sz
+
+    INIT_SIZE(16);
+    INIT_SIZE(8);
+    vp8_init_intra4x4_predictors_internal();
+}
+
+void vp8_build_intra_predictors_mby_s(MACROBLOCKD *x,
+                                      unsigned char * yabove_row,
+                                      unsigned char * yleft,
+                                      int left_stride,
+                                      unsigned char * ypred_ptr,
+                                      int y_stride)
+{
+    MB_PREDICTION_MODE mode = x->mode_info_context->mbmi.mode;
+    DECLARE_ALIGNED(16, uint8_t, yleft_col[16]);
+    int i;
+    intra_pred_fn fn;
+
+    for (i = 0; i < 16; i++)
+    {
+        yleft_col[i] = yleft[i* left_stride];
+    }
+
+    if (mode == DC_PRED)
+    {
+        fn = dc_pred[x->left_available][x->up_available][SIZE_16];
+    }
+    else
+    {
+        fn = pred[mode][SIZE_16];
+    }
+
+    fn(ypred_ptr, y_stride, yabove_row, yleft_col);
+}
+
+void vp8_build_intra_predictors_mbuv_s(MACROBLOCKD *x,
+                                       unsigned char * uabove_row,
+                                       unsigned char * vabove_row,
+                                       unsigned char * uleft,
+                                       unsigned char * vleft,
+                                       int left_stride,
+                                       unsigned char * upred_ptr,
+                                       unsigned char * vpred_ptr,
+                                       int pred_stride)
+{
+    MB_PREDICTION_MODE uvmode = x->mode_info_context->mbmi.uv_mode;
+    unsigned char uleft_col[8];
+    unsigned char vleft_col[8];
+    int i;
+    intra_pred_fn fn;
+
+    for (i = 0; i < 8; i++)
+    {
+        uleft_col[i] = uleft[i * left_stride];
+        vleft_col[i] = vleft[i * left_stride];
+    }
+
+    if (uvmode == DC_PRED)
+    {
+        fn = dc_pred[x->left_available][x->up_available][SIZE_8];
+    }
+    else
+    {
+        fn = pred[uvmode][SIZE_8];
+    }
+
+    fn(upred_ptr, pred_stride, uabove_row, uleft_col);
+    fn(vpred_ptr, pred_stride, vabove_row, vleft_col);
+}
+
+void vp8_init_intra_predictors(void)
+{
+    once(vp8_init_intra_predictors_internal);
+}
diff --git a/libvpx/libvpx/vp8/common/reconintra.h b/libvpx/libvpx/vp8/common/reconintra.h
new file mode 100644
index 0000000..b6225a6
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/reconintra.h
@@ -0,0 +1,44 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_RECONINTRA_H_
+#define VP8_COMMON_RECONINTRA_H_
+
+#include "vp8/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_build_intra_predictors_mby_s(MACROBLOCKD *x,
+                                      unsigned char *yabove_row,
+                                      unsigned char *yleft,
+                                      int left_stride,
+                                      unsigned char *ypred_ptr,
+                                      int y_stride);
+
+void vp8_build_intra_predictors_mbuv_s(MACROBLOCKD *x,
+                                       unsigned char * uabove_row,
+                                       unsigned char * vabove_row,
+                                       unsigned char * uleft,
+                                       unsigned char * vleft,
+                                       int left_stride,
+                                       unsigned char * upred_ptr,
+                                       unsigned char * vpred_ptr,
+                                       int pred_stride);
+
+void vp8_init_intra_predictors(void);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_RECONINTRA_H_
diff --git a/libvpx/libvpx/vp8/common/reconintra4x4.c b/libvpx/libvpx/vp8/common/reconintra4x4.c
new file mode 100644
index 0000000..35ad891
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/reconintra4x4.c
@@ -0,0 +1,54 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string.h>
+
+#include "vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vp8_rtcd.h"
+#include "blockd.h"
+
+typedef void (*intra_pred_fn)(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left);
+
+static intra_pred_fn pred[10];
+
+void vp8_init_intra4x4_predictors_internal(void)
+{
+    pred[B_DC_PRED] = vpx_dc_predictor_4x4;
+    pred[B_TM_PRED] = vpx_tm_predictor_4x4;
+    pred[B_VE_PRED] = vpx_ve_predictor_4x4;
+    pred[B_HE_PRED] = vpx_he_predictor_4x4;
+    pred[B_LD_PRED] = vpx_d45e_predictor_4x4;
+    pred[B_RD_PRED] = vpx_d135_predictor_4x4;
+    pred[B_VR_PRED] = vpx_d117_predictor_4x4;
+    pred[B_VL_PRED] = vpx_d63f_predictor_4x4;
+    pred[B_HD_PRED] = vpx_d153_predictor_4x4;
+    pred[B_HU_PRED] = vpx_d207_predictor_4x4;
+}
+
+void vp8_intra4x4_predict(unsigned char *above,
+                          unsigned char *yleft, int left_stride,
+                          B_PREDICTION_MODE b_mode,
+                          unsigned char *dst, int dst_stride,
+                          unsigned char top_left)
+{
+    unsigned char Left[4];
+    unsigned char Aboveb[12], *Above = Aboveb + 4;
+
+    Left[0] = yleft[0];
+    Left[1] = yleft[left_stride];
+    Left[2] = yleft[2 * left_stride];
+    Left[3] = yleft[3 * left_stride];
+    memcpy(Above, above, 8);
+    Above[-1] = top_left;
+
+    pred[b_mode](dst, dst_stride, Above, Left);
+}
diff --git a/libvpx/libvpx/vp8/common/reconintra4x4.h b/libvpx/libvpx/vp8/common/reconintra4x4.h
new file mode 100644
index 0000000..5dc5d13
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/reconintra4x4.h
@@ -0,0 +1,48 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_RECONINTRA4X4_H_
+#define VP8_COMMON_RECONINTRA4X4_H_
+#include "vp8/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE void intra_prediction_down_copy(MACROBLOCKD *xd,
+                                              unsigned char *above_right_src)
+{
+    int dst_stride = xd->dst.y_stride;
+    unsigned char *above_right_dst = xd->dst.y_buffer - dst_stride + 16;
+
+    unsigned int *src_ptr = (unsigned int *)above_right_src;
+    unsigned int *dst_ptr0 = (unsigned int *)(above_right_dst + 4 * dst_stride);
+    unsigned int *dst_ptr1 = (unsigned int *)(above_right_dst + 8 * dst_stride);
+    unsigned int *dst_ptr2 = (unsigned int *)(above_right_dst + 12 * dst_stride);
+
+    *dst_ptr0 = *src_ptr;
+    *dst_ptr1 = *src_ptr;
+    *dst_ptr2 = *src_ptr;
+}
+
+void vp8_intra4x4_predict(unsigned char *Above,
+                          unsigned char *yleft, int left_stride,
+                          B_PREDICTION_MODE b_mode,
+                          unsigned char *dst, int dst_stride,
+                          unsigned char top_left);
+
+void vp8_init_intra4x4_predictors_internal(void);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_RECONINTRA4X4_H_
diff --git a/libvpx/libvpx/vp8/common/rtcd.c b/libvpx/libvpx/vp8/common/rtcd.c
new file mode 100644
index 0000000..ab0e9b4
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/rtcd.c
@@ -0,0 +1,19 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_config.h"
+#define RTCD_C
+#include "./vp8_rtcd.h"
+#include "vpx_ports/vpx_once.h"
+
+
+void vp8_rtcd()
+{
+    once(setup_rtcd_internal);
+}
diff --git a/libvpx/libvpx/vp8/common/rtcd_defs.pl b/libvpx/libvpx/vp8/common/rtcd_defs.pl
new file mode 100644
index 0000000..856ede1
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/rtcd_defs.pl
@@ -0,0 +1,315 @@
+sub vp8_common_forward_decls() {
+print <<EOF
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+EOF
+}
+forward_decls qw/vp8_common_forward_decls/;
+
+#
+# system state
+#
+add_proto qw/void vp8_clear_system_state/, "";
+specialize qw/vp8_clear_system_state mmx/;
+$vp8_clear_system_state_mmx=vpx_reset_mmx_state;
+
+#
+# Dequant
+#
+add_proto qw/void vp8_dequantize_b/, "struct blockd*, short *dqc";
+specialize qw/vp8_dequantize_b mmx media neon msa/;
+$vp8_dequantize_b_media=vp8_dequantize_b_v6;
+
+add_proto qw/void vp8_dequant_idct_add/, "short *input, short *dq, unsigned char *output, int stride";
+specialize qw/vp8_dequant_idct_add mmx media neon dspr2 msa/;
+$vp8_dequant_idct_add_media=vp8_dequant_idct_add_v6;
+$vp8_dequant_idct_add_dspr2=vp8_dequant_idct_add_dspr2;
+
+add_proto qw/void vp8_dequant_idct_add_y_block/, "short *q, short *dq, unsigned char *dst, int stride, char *eobs";
+specialize qw/vp8_dequant_idct_add_y_block mmx sse2 media neon dspr2 msa/;
+$vp8_dequant_idct_add_y_block_media=vp8_dequant_idct_add_y_block_v6;
+$vp8_dequant_idct_add_y_block_dspr2=vp8_dequant_idct_add_y_block_dspr2;
+
+add_proto qw/void vp8_dequant_idct_add_uv_block/, "short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs";
+specialize qw/vp8_dequant_idct_add_uv_block mmx sse2 media neon dspr2 msa/;
+$vp8_dequant_idct_add_uv_block_media=vp8_dequant_idct_add_uv_block_v6;
+$vp8_dequant_idct_add_y_block_dspr2=vp8_dequant_idct_add_y_block_dspr2;
+
+#
+# Loopfilter
+#
+add_proto qw/void vp8_loop_filter_mbv/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
+specialize qw/vp8_loop_filter_mbv mmx sse2 media neon dspr2 msa/;
+$vp8_loop_filter_mbv_media=vp8_loop_filter_mbv_armv6;
+$vp8_loop_filter_mbv_dspr2=vp8_loop_filter_mbv_dspr2;
+
+add_proto qw/void vp8_loop_filter_bv/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
+specialize qw/vp8_loop_filter_bv mmx sse2 media neon dspr2 msa/;
+$vp8_loop_filter_bv_media=vp8_loop_filter_bv_armv6;
+$vp8_loop_filter_bv_dspr2=vp8_loop_filter_bv_dspr2;
+
+add_proto qw/void vp8_loop_filter_mbh/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
+specialize qw/vp8_loop_filter_mbh mmx sse2 media neon dspr2 msa/;
+$vp8_loop_filter_mbh_media=vp8_loop_filter_mbh_armv6;
+$vp8_loop_filter_mbh_dspr2=vp8_loop_filter_mbh_dspr2;
+
+add_proto qw/void vp8_loop_filter_bh/, "unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi";
+specialize qw/vp8_loop_filter_bh mmx sse2 media neon dspr2 msa/;
+$vp8_loop_filter_bh_media=vp8_loop_filter_bh_armv6;
+$vp8_loop_filter_bh_dspr2=vp8_loop_filter_bh_dspr2;
+
+
+add_proto qw/void vp8_loop_filter_simple_mbv/, "unsigned char *y, int ystride, const unsigned char *blimit";
+specialize qw/vp8_loop_filter_simple_mbv mmx sse2 media neon msa/;
+$vp8_loop_filter_simple_mbv_c=vp8_loop_filter_simple_vertical_edge_c;
+$vp8_loop_filter_simple_mbv_mmx=vp8_loop_filter_simple_vertical_edge_mmx;
+$vp8_loop_filter_simple_mbv_sse2=vp8_loop_filter_simple_vertical_edge_sse2;
+$vp8_loop_filter_simple_mbv_media=vp8_loop_filter_simple_vertical_edge_armv6;
+$vp8_loop_filter_simple_mbv_neon=vp8_loop_filter_mbvs_neon;
+$vp8_loop_filter_simple_mbv_msa=vp8_loop_filter_simple_vertical_edge_msa;
+
+add_proto qw/void vp8_loop_filter_simple_mbh/, "unsigned char *y, int ystride, const unsigned char *blimit";
+specialize qw/vp8_loop_filter_simple_mbh mmx sse2 media neon msa/;
+$vp8_loop_filter_simple_mbh_c=vp8_loop_filter_simple_horizontal_edge_c;
+$vp8_loop_filter_simple_mbh_mmx=vp8_loop_filter_simple_horizontal_edge_mmx;
+$vp8_loop_filter_simple_mbh_sse2=vp8_loop_filter_simple_horizontal_edge_sse2;
+$vp8_loop_filter_simple_mbh_media=vp8_loop_filter_simple_horizontal_edge_armv6;
+$vp8_loop_filter_simple_mbh_neon=vp8_loop_filter_mbhs_neon;
+$vp8_loop_filter_simple_mbh_msa=vp8_loop_filter_simple_horizontal_edge_msa;
+
+add_proto qw/void vp8_loop_filter_simple_bv/, "unsigned char *y, int ystride, const unsigned char *blimit";
+specialize qw/vp8_loop_filter_simple_bv mmx sse2 media neon msa/;
+$vp8_loop_filter_simple_bv_c=vp8_loop_filter_bvs_c;
+$vp8_loop_filter_simple_bv_mmx=vp8_loop_filter_bvs_mmx;
+$vp8_loop_filter_simple_bv_sse2=vp8_loop_filter_bvs_sse2;
+$vp8_loop_filter_simple_bv_media=vp8_loop_filter_bvs_armv6;
+$vp8_loop_filter_simple_bv_neon=vp8_loop_filter_bvs_neon;
+$vp8_loop_filter_simple_bv_msa=vp8_loop_filter_bvs_msa;
+
+add_proto qw/void vp8_loop_filter_simple_bh/, "unsigned char *y, int ystride, const unsigned char *blimit";
+specialize qw/vp8_loop_filter_simple_bh mmx sse2 media neon msa/;
+$vp8_loop_filter_simple_bh_c=vp8_loop_filter_bhs_c;
+$vp8_loop_filter_simple_bh_mmx=vp8_loop_filter_bhs_mmx;
+$vp8_loop_filter_simple_bh_sse2=vp8_loop_filter_bhs_sse2;
+$vp8_loop_filter_simple_bh_media=vp8_loop_filter_bhs_armv6;
+$vp8_loop_filter_simple_bh_neon=vp8_loop_filter_bhs_neon;
+$vp8_loop_filter_simple_bh_msa=vp8_loop_filter_bhs_msa;
+
+#
+# IDCT
+#
+#idct16
+add_proto qw/void vp8_short_idct4x4llm/, "short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride";
+specialize qw/vp8_short_idct4x4llm mmx media neon dspr2 msa/;
+$vp8_short_idct4x4llm_media=vp8_short_idct4x4llm_v6_dual;
+$vp8_short_idct4x4llm_dspr2=vp8_short_idct4x4llm_dspr2;
+
+#iwalsh1
+add_proto qw/void vp8_short_inv_walsh4x4_1/, "short *input, short *output";
+specialize qw/vp8_short_inv_walsh4x4_1 dspr2/;
+$vp8_short_inv_walsh4x4_1_dspr2=vp8_short_inv_walsh4x4_1_dspr2;
+# no asm yet
+
+#iwalsh16
+add_proto qw/void vp8_short_inv_walsh4x4/, "short *input, short *output";
+specialize qw/vp8_short_inv_walsh4x4 mmx sse2 media neon dspr2 msa/;
+$vp8_short_inv_walsh4x4_media=vp8_short_inv_walsh4x4_v6;
+$vp8_short_inv_walsh4x4_dspr2=vp8_short_inv_walsh4x4_dspr2;
+
+#idct1_scalar_add
+add_proto qw/void vp8_dc_only_idct_add/, "short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride";
+specialize qw/vp8_dc_only_idct_add	mmx media neon dspr2 msa/;
+$vp8_dc_only_idct_add_media=vp8_dc_only_idct_add_v6;
+$vp8_dc_only_idct_add_dspr2=vp8_dc_only_idct_add_dspr2;
+
+#
+# RECON
+#
+add_proto qw/void vp8_copy_mem16x16/, "unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_copy_mem16x16 mmx sse2 media neon dspr2 msa/;
+$vp8_copy_mem16x16_media=vp8_copy_mem16x16_v6;
+$vp8_copy_mem16x16_dspr2=vp8_copy_mem16x16_dspr2;
+
+add_proto qw/void vp8_copy_mem8x8/, "unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_copy_mem8x8 mmx media neon dspr2 msa/;
+$vp8_copy_mem8x8_media=vp8_copy_mem8x8_v6;
+$vp8_copy_mem8x8_dspr2=vp8_copy_mem8x8_dspr2;
+
+add_proto qw/void vp8_copy_mem8x4/, "unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_copy_mem8x4 mmx media neon dspr2 msa/;
+$vp8_copy_mem8x4_media=vp8_copy_mem8x4_v6;
+$vp8_copy_mem8x4_dspr2=vp8_copy_mem8x4_dspr2;
+
+#
+# Postproc
+#
+if (vpx_config("CONFIG_POSTPROC") eq "yes") {
+    add_proto qw/void vp8_mbpost_proc_down/, "unsigned char *dst, int pitch, int rows, int cols,int flimit";
+    specialize qw/vp8_mbpost_proc_down mmx sse2 msa/;
+    $vp8_mbpost_proc_down_sse2=vp8_mbpost_proc_down_xmm;
+
+    add_proto qw/void vp8_mbpost_proc_across_ip/, "unsigned char *dst, int pitch, int rows, int cols,int flimit";
+    specialize qw/vp8_mbpost_proc_across_ip sse2 msa/;
+    $vp8_mbpost_proc_across_ip_sse2=vp8_mbpost_proc_across_ip_xmm;
+
+    add_proto qw/void vp8_post_proc_down_and_across_mb_row/, "unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size";
+    specialize qw/vp8_post_proc_down_and_across_mb_row sse2 msa/;
+
+    add_proto qw/void vp8_blend_mb_inner/, "unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride";
+    # no asm yet
+
+    add_proto qw/void vp8_blend_mb_outer/, "unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride";
+    # no asm yet
+
+    add_proto qw/void vp8_blend_b/, "unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride";
+    # no asm yet
+
+    add_proto qw/void vp8_filter_by_weight16x16/, "unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight";
+    specialize qw/vp8_filter_by_weight16x16 sse2 msa/;
+
+    add_proto qw/void vp8_filter_by_weight8x8/, "unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight";
+    specialize qw/vp8_filter_by_weight8x8 sse2 msa/;
+
+    add_proto qw/void vp8_filter_by_weight4x4/, "unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight";
+    # no asm yet
+}
+
+#
+# Subpixel
+#
+add_proto qw/void vp8_sixtap_predict16x16/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_sixtap_predict16x16 mmx sse2 ssse3 media neon dspr2 msa/;
+$vp8_sixtap_predict16x16_media=vp8_sixtap_predict16x16_armv6;
+$vp8_sixtap_predict16x16_dspr2=vp8_sixtap_predict16x16_dspr2;
+
+add_proto qw/void vp8_sixtap_predict8x8/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_sixtap_predict8x8 mmx sse2 ssse3 media neon dspr2 msa/;
+$vp8_sixtap_predict8x8_media=vp8_sixtap_predict8x8_armv6;
+$vp8_sixtap_predict8x8_dspr2=vp8_sixtap_predict8x8_dspr2;
+
+add_proto qw/void vp8_sixtap_predict8x4/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_sixtap_predict8x4 mmx sse2 ssse3 media neon dspr2 msa/;
+$vp8_sixtap_predict8x4_media=vp8_sixtap_predict8x4_armv6;
+$vp8_sixtap_predict8x4_dspr2=vp8_sixtap_predict8x4_dspr2;
+
+add_proto qw/void vp8_sixtap_predict4x4/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_sixtap_predict4x4 mmx ssse3 media dspr2 msa/;
+$vp8_sixtap_predict4x4_media=vp8_sixtap_predict4x4_armv6;
+$vp8_sixtap_predict4x4_dspr2=vp8_sixtap_predict4x4_dspr2;
+
+add_proto qw/void vp8_bilinear_predict16x16/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_bilinear_predict16x16 mmx sse2 ssse3 media neon msa/;
+$vp8_bilinear_predict16x16_media=vp8_bilinear_predict16x16_armv6;
+
+add_proto qw/void vp8_bilinear_predict8x8/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_bilinear_predict8x8 mmx sse2 ssse3 media neon msa/;
+$vp8_bilinear_predict8x8_media=vp8_bilinear_predict8x8_armv6;
+
+add_proto qw/void vp8_bilinear_predict8x4/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_bilinear_predict8x4 mmx media neon msa/;
+$vp8_bilinear_predict8x4_media=vp8_bilinear_predict8x4_armv6;
+
+add_proto qw/void vp8_bilinear_predict4x4/, "unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch";
+specialize qw/vp8_bilinear_predict4x4 mmx media msa/;
+$vp8_bilinear_predict4x4_media=vp8_bilinear_predict4x4_armv6;
+
+#
+# Encoder functions below this point.
+#
+if (vpx_config("CONFIG_VP8_ENCODER") eq "yes") {
+
+#
+# Block copy
+#
+if ($opts{arch} =~ /x86/) {
+    add_proto qw/void vp8_copy32xn/, "const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n";
+    specialize qw/vp8_copy32xn sse2 sse3/;
+}
+
+#
+# Forward DCT
+#
+add_proto qw/void vp8_short_fdct4x4/, "short *input, short *output, int pitch";
+specialize qw/vp8_short_fdct4x4 mmx sse2 media neon msa/;
+$vp8_short_fdct4x4_media=vp8_short_fdct4x4_armv6;
+
+add_proto qw/void vp8_short_fdct8x4/, "short *input, short *output, int pitch";
+specialize qw/vp8_short_fdct8x4 mmx sse2 media neon msa/;
+$vp8_short_fdct8x4_media=vp8_short_fdct8x4_armv6;
+
+add_proto qw/void vp8_short_walsh4x4/, "short *input, short *output, int pitch";
+specialize qw/vp8_short_walsh4x4 sse2 media neon msa/;
+$vp8_short_walsh4x4_media=vp8_short_walsh4x4_armv6;
+
+#
+# Quantizer
+#
+add_proto qw/void vp8_regular_quantize_b/, "struct block *, struct blockd *";
+specialize qw/vp8_regular_quantize_b sse2 sse4_1 msa/;
+
+add_proto qw/void vp8_fast_quantize_b/, "struct block *, struct blockd *";
+specialize qw/vp8_fast_quantize_b sse2 ssse3 neon msa/;
+
+#
+# Block subtraction
+#
+add_proto qw/int vp8_block_error/, "short *coeff, short *dqcoeff";
+specialize qw/vp8_block_error mmx sse2 msa/;
+$vp8_block_error_sse2=vp8_block_error_xmm;
+
+add_proto qw/int vp8_mbblock_error/, "struct macroblock *mb, int dc";
+specialize qw/vp8_mbblock_error mmx sse2 msa/;
+$vp8_mbblock_error_sse2=vp8_mbblock_error_xmm;
+
+add_proto qw/int vp8_mbuverror/, "struct macroblock *mb";
+specialize qw/vp8_mbuverror mmx sse2 msa/;
+$vp8_mbuverror_sse2=vp8_mbuverror_xmm;
+
+#
+# Motion search
+#
+add_proto qw/int vp8_full_search_sad/, "struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv";
+specialize qw/vp8_full_search_sad sse3 sse4_1/;
+$vp8_full_search_sad_sse3=vp8_full_search_sadx3;
+$vp8_full_search_sad_sse4_1=vp8_full_search_sadx8;
+
+add_proto qw/int vp8_refining_search_sad/, "struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv";
+specialize qw/vp8_refining_search_sad sse3/;
+$vp8_refining_search_sad_sse3=vp8_refining_search_sadx4;
+
+add_proto qw/int vp8_diamond_search_sad/, "struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv";
+$vp8_diamond_search_sad_sse3=vp8_diamond_search_sadx4;
+
+#
+# Alt-ref Noise Reduction (ARNR)
+#
+if (vpx_config("CONFIG_REALTIME_ONLY") ne "yes") {
+    add_proto qw/void vp8_temporal_filter_apply/, "unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count";
+    specialize qw/vp8_temporal_filter_apply sse2 msa/;
+}
+
+#
+# Denoiser filter
+#
+if (vpx_config("CONFIG_TEMPORAL_DENOISING") eq "yes") {
+    add_proto qw/int vp8_denoiser_filter/, "unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
+    specialize qw/vp8_denoiser_filter sse2 neon msa/;
+    add_proto qw/int vp8_denoiser_filter_uv/, "unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
+    specialize qw/vp8_denoiser_filter_uv sse2 neon msa/;
+}
+
+# End of encoder only functions
+}
+1;
diff --git a/libvpx/libvpx/vp8/common/setupintrarecon.c b/libvpx/libvpx/vp8/common/setupintrarecon.c
new file mode 100644
index 0000000..669564d
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/setupintrarecon.c
@@ -0,0 +1,39 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "setupintrarecon.h"
+#include "vpx_mem/vpx_mem.h"
+
+void vp8_setup_intra_recon(YV12_BUFFER_CONFIG *ybf)
+{
+    int i;
+
+    /* set up frame new frame for intra coded blocks */
+    memset(ybf->y_buffer - 1 - ybf->y_stride, 127, ybf->y_width + 5);
+    for (i = 0; i < ybf->y_height; i++)
+        ybf->y_buffer[ybf->y_stride *i - 1] = (unsigned char) 129;
+
+    memset(ybf->u_buffer - 1 - ybf->uv_stride, 127, ybf->uv_width + 5);
+    for (i = 0; i < ybf->uv_height; i++)
+        ybf->u_buffer[ybf->uv_stride *i - 1] = (unsigned char) 129;
+
+    memset(ybf->v_buffer - 1 - ybf->uv_stride, 127, ybf->uv_width + 5);
+    for (i = 0; i < ybf->uv_height; i++)
+        ybf->v_buffer[ybf->uv_stride *i - 1] = (unsigned char) 129;
+
+}
+
+void vp8_setup_intra_recon_top_line(YV12_BUFFER_CONFIG *ybf)
+{
+    memset(ybf->y_buffer - 1 - ybf->y_stride, 127, ybf->y_width + 5);
+    memset(ybf->u_buffer - 1 - ybf->uv_stride, 127, ybf->uv_width + 5);
+    memset(ybf->v_buffer - 1 - ybf->uv_stride, 127, ybf->uv_width + 5);
+}
diff --git a/libvpx/libvpx/vp8/common/setupintrarecon.h b/libvpx/libvpx/vp8/common/setupintrarecon.h
new file mode 100644
index 0000000..1857c4e
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/setupintrarecon.h
@@ -0,0 +1,45 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_COMMON_SETUPINTRARECON_H_
+#define VP8_COMMON_SETUPINTRARECON_H_
+
+#include "./vpx_config.h"
+#include "vpx_scale/yv12config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+extern void vp8_setup_intra_recon(YV12_BUFFER_CONFIG *ybf);
+extern void vp8_setup_intra_recon_top_line(YV12_BUFFER_CONFIG *ybf);
+
+static INLINE void setup_intra_recon_left(unsigned char *y_buffer,
+                                          unsigned char *u_buffer,
+                                          unsigned char *v_buffer,
+                                          int y_stride,
+                                          int uv_stride)
+{
+    int i;
+
+    for (i = 0; i < 16; i++)
+        y_buffer[y_stride *i] = (unsigned char) 129;
+
+    for (i = 0; i < 8; i++)
+        u_buffer[uv_stride *i] = (unsigned char) 129;
+
+    for (i = 0; i < 8; i++)
+        v_buffer[uv_stride *i] = (unsigned char) 129;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_SETUPINTRARECON_H_
diff --git a/libvpx/libvpx/vp8/common/swapyv12buffer.c b/libvpx/libvpx/vp8/common/swapyv12buffer.c
new file mode 100644
index 0000000..73656b3
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/swapyv12buffer.c
@@ -0,0 +1,34 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "swapyv12buffer.h"
+
+void vp8_swap_yv12_buffer(YV12_BUFFER_CONFIG *new_frame, YV12_BUFFER_CONFIG *last_frame)
+{
+    unsigned char *temp;
+
+    temp = last_frame->buffer_alloc;
+    last_frame->buffer_alloc = new_frame->buffer_alloc;
+    new_frame->buffer_alloc = temp;
+
+    temp = last_frame->y_buffer;
+    last_frame->y_buffer = new_frame->y_buffer;
+    new_frame->y_buffer = temp;
+
+    temp = last_frame->u_buffer;
+    last_frame->u_buffer = new_frame->u_buffer;
+    new_frame->u_buffer = temp;
+
+    temp = last_frame->v_buffer;
+    last_frame->v_buffer = new_frame->v_buffer;
+    new_frame->v_buffer = temp;
+
+}
diff --git a/libvpx/libvpx/vp8/common/swapyv12buffer.h b/libvpx/libvpx/vp8/common/swapyv12buffer.h
new file mode 100644
index 0000000..1d66cd3
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/swapyv12buffer.h
@@ -0,0 +1,27 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_SWAPYV12BUFFER_H_
+#define VP8_COMMON_SWAPYV12BUFFER_H_
+
+#include "vpx_scale/yv12config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_swap_yv12_buffer(YV12_BUFFER_CONFIG *new_frame, YV12_BUFFER_CONFIG *last_frame);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_SWAPYV12BUFFER_H_
diff --git a/libvpx/libvpx/vp8/common/systemdependent.h b/libvpx/libvpx/vp8/common/systemdependent.h
new file mode 100644
index 0000000..3d44e37
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/systemdependent.h
@@ -0,0 +1,27 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_COMMON_SYSTEMDEPENDENT_H_
+#define VP8_COMMON_SYSTEMDEPENDENT_H_
+
+#include "vpx_config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP8Common;
+void vp8_machine_specific_config(struct VP8Common *);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_SYSTEMDEPENDENT_H_
diff --git a/libvpx/libvpx/vp8/common/textblit.c b/libvpx/libvpx/vp8/common/textblit.c
new file mode 100644
index 0000000..1756100
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/textblit.c
@@ -0,0 +1,130 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+
+void vp8_blit_text(const char *msg, unsigned char *address, const int pitch)
+{
+    int letter_bitmap;
+    unsigned char *output_pos = address;
+    int colpos;
+    const int font[] =
+    {
+        0x0, 0x5C00, 0x8020, 0xAFABEA, 0xD7EC0, 0x1111111, 0x1855740, 0x18000,
+        0x45C0, 0x74400, 0x51140, 0x23880, 0xC4000, 0x21080, 0x80000, 0x111110,
+        0xE9D72E, 0x87E40, 0x12AD732, 0xAAD62A, 0x4F94C4, 0x4D6B7, 0x456AA,
+        0x3E8423, 0xAAD6AA, 0xAAD6A2, 0x2800, 0x2A00, 0x8A880, 0x52940, 0x22A20,
+        0x15422, 0x6AD62E, 0x1E4A53E, 0xAAD6BF, 0x8C62E, 0xE8C63F, 0x118D6BF,
+        0x1094BF, 0xCAC62E, 0x1F2109F, 0x118FE31, 0xF8C628, 0x8A89F, 0x108421F,
+        0x1F1105F, 0x1F4105F, 0xE8C62E, 0x2294BF, 0x164C62E, 0x12694BF, 0x8AD6A2,
+        0x10FC21, 0x1F8421F, 0x744107, 0xF8220F, 0x1151151, 0x117041, 0x119D731,
+        0x47E0, 0x1041041, 0xFC400, 0x10440, 0x1084210, 0x820
+    };
+    colpos = 0;
+
+    while (msg[colpos] != 0)
+    {
+        char letter = msg[colpos];
+        int fontcol, fontrow;
+
+        if (letter <= 'Z' && letter >= ' ')
+            letter_bitmap = font[letter-' '];
+        else if (letter <= 'z' && letter >= 'a')
+            letter_bitmap = font[letter-'a'+'A' - ' '];
+        else
+            letter_bitmap = font[0];
+
+        for (fontcol = 6; fontcol >= 0 ; fontcol--)
+            for (fontrow = 0; fontrow < 5; fontrow++)
+                output_pos[fontrow *pitch + fontcol] =
+                    ((letter_bitmap >> (fontcol * 5)) & (1 << fontrow) ? 255 : 0);
+
+        output_pos += 7;
+        colpos++;
+    }
+}
+
+static void plot (const int x, const int y, unsigned char *image, const int pitch)
+{
+    image [x+y*pitch] ^= 255;
+}
+
+/* Bresenham line algorithm */
+void vp8_blit_line(int x0, int x1, int y0, int y1, unsigned char *image, const int pitch)
+{
+    int steep = abs(y1 - y0) > abs(x1 - x0);
+    int deltax, deltay;
+    int error, ystep, y, x;
+
+    if (steep)
+    {
+        int t;
+        t = x0;
+        x0 = y0;
+        y0 = t;
+
+        t = x1;
+        x1 = y1;
+        y1 = t;
+    }
+
+    if (x0 > x1)
+    {
+        int t;
+        t = x0;
+        x0 = x1;
+        x1 = t;
+
+        t = y0;
+        y0 = y1;
+        y1 = t;
+    }
+
+    deltax = x1 - x0;
+    deltay = abs(y1 - y0);
+    error  = deltax / 2;
+
+    y = y0;
+
+    if (y0 < y1)
+        ystep = 1;
+    else
+        ystep = -1;
+
+    if (steep)
+    {
+        for (x = x0; x <= x1; x++)
+        {
+            plot(y,x, image, pitch);
+
+            error = error - deltay;
+            if (error < 0)
+            {
+                y = y + ystep;
+                error = error + deltax;
+            }
+        }
+    }
+    else
+    {
+        for (x = x0; x <= x1; x++)
+        {
+            plot(x,y, image, pitch);
+
+            error = error - deltay;
+            if (error < 0)
+            {
+                y = y + ystep;
+                error = error + deltax;
+            }
+        }
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/threading.h b/libvpx/libvpx/vp8/common/threading.h
new file mode 100644
index 0000000..183b49b
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/threading.h
@@ -0,0 +1,232 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_THREADING_H_
+#define VP8_COMMON_THREADING_H_
+
+#include "./vpx_config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_OS_SUPPORT && CONFIG_MULTITHREAD
+
+/* Thread management macros */
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+/* Win32 */
+#include <process.h>
+#include <windows.h>
+#define THREAD_FUNCTION unsigned int __stdcall
+#define THREAD_FUNCTION_RETURN DWORD
+#define THREAD_SPECIFIC_INDEX DWORD
+#define pthread_t HANDLE
+#define pthread_attr_t DWORD
+#define pthread_detach(thread) if(thread!=NULL)CloseHandle(thread)
+#define thread_sleep(nms) Sleep(nms)
+#define pthread_cancel(thread) terminate_thread(thread,0)
+#define ts_key_create(ts_key, destructor) {ts_key = TlsAlloc();};
+#define pthread_getspecific(ts_key) TlsGetValue(ts_key)
+#define pthread_setspecific(ts_key, value) TlsSetValue(ts_key, (void *)value)
+#define pthread_self() GetCurrentThreadId()
+
+#elif defined(__OS2__)
+/* OS/2 */
+#define INCL_DOS
+#include <os2.h>
+
+#include <stdlib.h>
+#define THREAD_FUNCTION void *
+#define THREAD_FUNCTION_RETURN void *
+#define THREAD_SPECIFIC_INDEX PULONG
+#define pthread_t TID
+#define pthread_attr_t ULONG
+#define pthread_detach(thread) 0
+#define thread_sleep(nms) DosSleep(nms)
+#define pthread_cancel(thread) DosKillThread(thread)
+#define ts_key_create(ts_key, destructor) \
+    DosAllocThreadLocalMemory(1, &(ts_key));
+#define pthread_getspecific(ts_key) ((void *)(*(ts_key)))
+#define pthread_setspecific(ts_key, value) (*(ts_key)=(ULONG)(value))
+#define pthread_self() _gettid()
+#else
+#ifdef __APPLE__
+#include <mach/mach_init.h>
+#include <mach/semaphore.h>
+#include <mach/task.h>
+#include <time.h>
+#include <unistd.h>
+
+#else
+#include <semaphore.h>
+#endif
+
+#include <pthread.h>
+/* pthreads */
+/* Nearly everything is already defined */
+#define THREAD_FUNCTION void *
+#define THREAD_FUNCTION_RETURN void *
+#define THREAD_SPECIFIC_INDEX pthread_key_t
+#define ts_key_create(ts_key, destructor) pthread_key_create (&(ts_key), destructor);
+#endif
+
+/* Synchronization macros: Win32 and Pthreads */
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+#define sem_t HANDLE
+#define pause(voidpara) __asm PAUSE
+#define sem_init(sem, sem_attr1, sem_init_value) (int)((*sem = CreateSemaphore(NULL,0,32768,NULL))==NULL)
+#define sem_wait(sem) (int)(WAIT_OBJECT_0 != WaitForSingleObject(*sem,INFINITE))
+#define sem_post(sem) ReleaseSemaphore(*sem,1,NULL)
+#define sem_destroy(sem) if(*sem)((int)(CloseHandle(*sem))==TRUE)
+#define thread_sleep(nms) Sleep(nms)
+
+#elif defined(__OS2__)
+typedef struct
+{
+    HEV  event;
+    HMTX wait_mutex;
+    HMTX count_mutex;
+    int  count;
+} sem_t;
+
+static inline int sem_init(sem_t *sem, int pshared, unsigned int value)
+{
+    DosCreateEventSem(NULL, &sem->event, pshared ? DC_SEM_SHARED : 0,
+                      value > 0 ? TRUE : FALSE);
+    DosCreateMutexSem(NULL, &sem->wait_mutex, 0, FALSE);
+    DosCreateMutexSem(NULL, &sem->count_mutex, 0, FALSE);
+
+    sem->count = value;
+
+    return 0;
+}
+
+static inline int sem_wait(sem_t * sem)
+{
+    DosRequestMutexSem(sem->wait_mutex, -1);
+
+    DosWaitEventSem(sem->event, -1);
+
+    DosRequestMutexSem(sem->count_mutex, -1);
+
+    sem->count--;
+    if (sem->count == 0)
+    {
+        ULONG post_count;
+
+        DosResetEventSem(sem->event, &post_count);
+    }
+
+    DosReleaseMutexSem(sem->count_mutex);
+
+    DosReleaseMutexSem(sem->wait_mutex);
+
+    return 0;
+}
+
+static inline int sem_post(sem_t * sem)
+{
+    DosRequestMutexSem(sem->count_mutex, -1);
+
+    if (sem->count < 32768)
+    {
+        sem->count++;
+        DosPostEventSem(sem->event);
+    }
+
+    DosReleaseMutexSem(sem->count_mutex);
+
+    return 0;
+}
+
+static inline int sem_destroy(sem_t * sem)
+{
+    DosCloseEventSem(sem->event);
+    DosCloseMutexSem(sem->wait_mutex);
+    DosCloseMutexSem(sem->count_mutex);
+
+    return 0;
+}
+
+#define thread_sleep(nms) DosSleep(nms)
+
+#else
+
+#ifdef __APPLE__
+#define sem_t semaphore_t
+#define sem_init(X,Y,Z) semaphore_create(mach_task_self(), X, SYNC_POLICY_FIFO, Z)
+#define sem_wait(sem) (semaphore_wait(*sem) )
+#define sem_post(sem) semaphore_signal(*sem)
+#define sem_destroy(sem) semaphore_destroy(mach_task_self(),*sem)
+#define thread_sleep(nms) /* { struct timespec ts;ts.tv_sec=0; ts.tv_nsec = 1000*nms;nanosleep(&ts, NULL);} */
+#else
+#include <unistd.h>
+#include <sched.h>
+#define thread_sleep(nms) sched_yield();/* {struct timespec ts;ts.tv_sec=0; ts.tv_nsec = 1000*nms;nanosleep(&ts, NULL);} */
+#endif
+/* Not Windows. Assume pthreads */
+
+#endif
+
+#if ARCH_X86 || ARCH_X86_64
+#include "vpx_ports/x86.h"
+#else
+#define x86_pause_hint()
+#endif
+
+#include "vpx_util/vpx_thread.h"
+
+static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
+    const int kMaxTryLocks = 4000;
+    int locked = 0;
+    int i;
+
+    for (i = 0; i < kMaxTryLocks; ++i) {
+        if (!pthread_mutex_trylock(mutex)) {
+            locked = 1;
+            break;
+        }
+    }
+
+    if (!locked)
+        pthread_mutex_lock(mutex);
+}
+
+static INLINE int protected_read(pthread_mutex_t *const mutex, const int *p) {
+    int ret;
+    mutex_lock(mutex);
+    ret = *p;
+    pthread_mutex_unlock(mutex);
+    return ret;
+}
+
+static INLINE void sync_read(pthread_mutex_t *const mutex, int mb_col,
+                             const int *last_row_current_mb_col,
+                             const int nsync) {
+    while (mb_col > (protected_read(mutex, last_row_current_mb_col) - nsync)) {
+        x86_pause_hint();
+        thread_sleep(0);
+    }
+}
+
+static INLINE void protected_write(pthread_mutex_t *mutex, int *p, int v) {
+    mutex_lock(mutex);
+    *p = v;
+    pthread_mutex_unlock(mutex);
+}
+
+#endif /* CONFIG_OS_SUPPORT && CONFIG_MULTITHREAD */
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_THREADING_H_
diff --git a/libvpx/libvpx/vp8/common/treecoder.c b/libvpx/libvpx/vp8/common/treecoder.c
new file mode 100644
index 0000000..d80c64b
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/treecoder.c
@@ -0,0 +1,143 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#if CONFIG_DEBUG
+#include <assert.h>
+#endif
+#include <stdio.h>
+
+#include "treecoder.h"
+
+static void tree2tok(
+    struct vp8_token_struct *const p,
+    vp8_tree t,
+    int i,
+    int v,
+    int L
+)
+{
+    v += v;
+    ++L;
+
+    do
+    {
+        const vp8_tree_index j = t[i++];
+
+        if (j <= 0)
+        {
+            p[-j].value = v;
+            p[-j].Len = L;
+        }
+        else
+            tree2tok(p, t, j, v, L);
+    }
+    while (++v & 1);
+}
+
+void vp8_tokens_from_tree(struct vp8_token_struct *p, vp8_tree t)
+{
+    tree2tok(p, t, 0, 0, 0);
+}
+
+void vp8_tokens_from_tree_offset(struct vp8_token_struct *p, vp8_tree t,
+                                 int offset)
+{
+    tree2tok(p - offset, t, 0, 0, 0);
+}
+
+static void branch_counts(
+    int n,                      /* n = size of alphabet */
+    vp8_token tok               [ /* n */ ],
+    vp8_tree tree,
+    unsigned int branch_ct       [ /* n-1 */ ] [2],
+    const unsigned int num_events[ /* n */ ]
+)
+{
+    const int tree_len = n - 1;
+    int t = 0;
+
+#if CONFIG_DEBUG
+    assert(tree_len);
+#endif
+
+    do
+    {
+        branch_ct[t][0] = branch_ct[t][1] = 0;
+    }
+    while (++t < tree_len);
+
+    t = 0;
+
+    do
+    {
+        int L = tok[t].Len;
+        const int enc = tok[t].value;
+        const unsigned int ct = num_events[t];
+
+        vp8_tree_index i = 0;
+
+        do
+        {
+            const int b = (enc >> --L) & 1;
+            const int j = i >> 1;
+#if CONFIG_DEBUG
+            assert(j < tree_len  &&  0 <= L);
+#endif
+
+            branch_ct [j] [b] += ct;
+            i = tree[ i + b];
+        }
+        while (i > 0);
+
+#if CONFIG_DEBUG
+        assert(!L);
+#endif
+    }
+    while (++t < n);
+
+}
+
+
+void vp8_tree_probs_from_distribution(
+    int n,                      /* n = size of alphabet */
+    vp8_token tok               [ /* n */ ],
+    vp8_tree tree,
+    vp8_prob probs          [ /* n-1 */ ],
+    unsigned int branch_ct       [ /* n-1 */ ] [2],
+    const unsigned int num_events[ /* n */ ],
+    unsigned int Pfac,
+    int rd
+)
+{
+    const int tree_len = n - 1;
+    int t = 0;
+
+    branch_counts(n, tok, tree, branch_ct, num_events);
+
+    do
+    {
+        const unsigned int *const c = branch_ct[t];
+        const unsigned int tot = c[0] + c[1];
+
+#if CONFIG_DEBUG
+        assert(tot < (1 << 24));        /* no overflow below */
+#endif
+
+        if (tot)
+        {
+            const unsigned int p = ((c[0] * Pfac) + (rd ? tot >> 1 : 0)) / tot;
+            probs[t] = p < 256 ? (p ? p : 1) : 255; /* agree w/old version for now */
+        }
+        else
+            probs[t] = vp8_prob_half;
+    }
+    while (++t < tree_len);
+}
diff --git a/libvpx/libvpx/vp8/common/treecoder.h b/libvpx/libvpx/vp8/common/treecoder.h
new file mode 100644
index 0000000..d22b7c5
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/treecoder.h
@@ -0,0 +1,98 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_COMMON_TREECODER_H_
+#define VP8_COMMON_TREECODER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef unsigned char vp8bc_index_t; /* probability index */
+
+
+typedef unsigned char vp8_prob;
+
+#define vp8_prob_half ( (vp8_prob) 128)
+
+typedef signed char vp8_tree_index;
+struct bool_coder_spec;
+
+typedef struct bool_coder_spec bool_coder_spec;
+typedef struct bool_writer bool_writer;
+typedef struct bool_reader bool_reader;
+
+typedef const bool_coder_spec c_bool_coder_spec;
+typedef const bool_writer c_bool_writer;
+typedef const bool_reader c_bool_reader;
+
+
+
+# define vp8_complement( x) (255 - x)
+
+
+/* We build coding trees compactly in arrays.
+   Each node of the tree is a pair of vp8_tree_indices.
+   Array index often references a corresponding probability table.
+   Index <= 0 means done encoding/decoding and value = -Index,
+   Index > 0 means need another bit, specification at index.
+   Nonnegative indices are always even;  processing begins at node 0. */
+
+typedef const vp8_tree_index vp8_tree[], *vp8_tree_p;
+
+
+typedef const struct vp8_token_struct
+{
+    int value;
+    int Len;
+} vp8_token;
+
+/* Construct encoding array from tree. */
+
+void vp8_tokens_from_tree(struct vp8_token_struct *, vp8_tree);
+void vp8_tokens_from_tree_offset(struct vp8_token_struct *, vp8_tree,
+                                 int offset);
+
+
+/* Convert array of token occurrence counts into a table of probabilities
+   for the associated binary encoding tree.  Also writes count of branches
+   taken for each node on the tree; this facilitiates decisions as to
+   probability updates. */
+
+void vp8_tree_probs_from_distribution(
+    int n,                      /* n = size of alphabet */
+    vp8_token tok               [ /* n */ ],
+    vp8_tree tree,
+    vp8_prob probs          [ /* n-1 */ ],
+    unsigned int branch_ct       [ /* n-1 */ ] [2],
+    const unsigned int num_events[ /* n */ ],
+    unsigned int Pfactor,
+    int Round
+);
+
+/* Variant of above using coder spec rather than hardwired 8-bit probs. */
+
+void vp8bc_tree_probs_from_distribution(
+    int n,                      /* n = size of alphabet */
+    vp8_token tok               [ /* n */ ],
+    vp8_tree tree,
+    vp8_prob probs          [ /* n-1 */ ],
+    unsigned int branch_ct       [ /* n-1 */ ] [2],
+    const unsigned int num_events[ /* n */ ],
+    c_bool_coder_spec *s
+);
+
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_TREECODER_H_
diff --git a/libvpx/libvpx/vp8/common/vp8_entropymodedata.h b/libvpx/libvpx/vp8/common/vp8_entropymodedata.h
new file mode 100644
index 0000000..c4aed49
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/vp8_entropymodedata.h
@@ -0,0 +1,254 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+*/
+
+#ifndef VP8_COMMON_VP8_ENTROPYMODEDATA_H_
+#define VP8_COMMON_VP8_ENTROPYMODEDATA_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*Generated file, included by entropymode.c*/
+
+
+const struct vp8_token_struct vp8_bmode_encodings[VP8_BINTRAMODES] =
+{
+    { 0, 1 },
+    { 2, 2 },
+    { 6, 3 },
+    { 28, 5 },
+    { 30, 5 },
+    { 58, 6 },
+    { 59, 6 },
+    { 62, 6 },
+    { 126, 7 },
+    { 127, 7 }
+};
+
+const struct vp8_token_struct vp8_ymode_encodings[VP8_YMODES] =
+{
+    { 0, 1 },
+    { 4, 3 },
+    { 5, 3 },
+    { 6, 3 },
+    { 7, 3 }
+};
+
+const struct vp8_token_struct vp8_kf_ymode_encodings[VP8_YMODES] =
+{
+    { 4, 3 },
+    { 5, 3 },
+    { 6, 3 },
+    { 7, 3 },
+    { 0, 1 }
+};
+
+const struct vp8_token_struct vp8_uv_mode_encodings[VP8_UV_MODES] =
+{
+    { 0, 1 },
+    { 2, 2 },
+    { 6, 3 },
+    { 7, 3 }
+};
+
+const struct vp8_token_struct vp8_mbsplit_encodings[VP8_NUMMBSPLITS] =
+{
+    { 6, 3 },
+    { 7, 3 },
+    { 2, 2 },
+    { 0, 1 }
+};
+
+const struct vp8_token_struct vp8_mv_ref_encoding_array[VP8_MVREFS] =
+{
+    { 2, 2 },
+    { 6, 3 },
+    { 0, 1 },
+    { 14, 4 },
+    { 15, 4 }
+};
+
+const struct vp8_token_struct vp8_sub_mv_ref_encoding_array[VP8_SUBMVREFS] =
+{
+    { 0, 1 },
+    { 2, 2 },
+    { 6, 3 },
+    { 7, 3 }
+};
+
+const struct vp8_token_struct vp8_small_mvencodings[8] =
+{
+    { 0, 3 },
+    { 1, 3 },
+    { 2, 3 },
+    { 3, 3 },
+    { 4, 3 },
+    { 5, 3 },
+    { 6, 3 },
+    { 7, 3 }
+};
+
+const vp8_prob vp8_ymode_prob[VP8_YMODES-1] =
+{
+    112, 86, 140, 37
+};
+
+const vp8_prob vp8_kf_ymode_prob[VP8_YMODES-1] =
+{
+    145, 156, 163, 128
+};
+
+const vp8_prob vp8_uv_mode_prob[VP8_UV_MODES-1] =
+{
+    162, 101, 204
+};
+
+const vp8_prob vp8_kf_uv_mode_prob[VP8_UV_MODES-1] =
+{
+    142, 114, 183
+};
+
+const vp8_prob vp8_bmode_prob[VP8_BINTRAMODES-1] =
+{
+    120, 90, 79, 133, 87, 85, 80, 111, 151
+};
+
+
+
+const vp8_prob vp8_kf_bmode_prob
+[VP8_BINTRAMODES] [VP8_BINTRAMODES] [VP8_BINTRAMODES-1] =
+{
+    {
+        { 231, 120,  48,  89, 115, 113, 120, 152, 112 },
+        { 152, 179,  64, 126, 170, 118,  46,  70,  95 },
+        { 175,  69, 143,  80,  85,  82,  72, 155, 103 },
+        {  56,  58,  10, 171, 218, 189,  17,  13, 152 },
+        { 144,  71,  10,  38, 171, 213, 144,  34,  26 },
+        { 114,  26,  17, 163,  44, 195,  21,  10, 173 },
+        { 121,  24,  80, 195,  26,  62,  44,  64,  85 },
+        { 170,  46,  55,  19, 136, 160,  33, 206,  71 },
+        {  63,  20,   8, 114, 114, 208,  12,   9, 226 },
+        {  81,  40,  11,  96, 182,  84,  29,  16,  36 }
+    },
+    {
+        { 134, 183,  89, 137,  98, 101, 106, 165, 148 },
+        {  72, 187, 100, 130, 157, 111,  32,  75,  80 },
+        {  66, 102, 167,  99,  74,  62,  40, 234, 128 },
+        {  41,  53,   9, 178, 241, 141,  26,   8, 107 },
+        { 104,  79,  12,  27, 217, 255,  87,  17,   7 },
+        {  74,  43,  26, 146,  73, 166,  49,  23, 157 },
+        {  65,  38, 105, 160,  51,  52,  31, 115, 128 },
+        {  87,  68,  71,  44, 114,  51,  15, 186,  23 },
+        {  47,  41,  14, 110, 182, 183,  21,  17, 194 },
+        {  66,  45,  25, 102, 197, 189,  23,  18,  22 }
+    },
+    {
+        {  88,  88, 147, 150,  42,  46,  45, 196, 205 },
+        {  43,  97, 183, 117,  85,  38,  35, 179,  61 },
+        {  39,  53, 200,  87,  26,  21,  43, 232, 171 },
+        {  56,  34,  51, 104, 114, 102,  29,  93,  77 },
+        { 107,  54,  32,  26,  51,   1,  81,  43,  31 },
+        {  39,  28,  85, 171,  58, 165,  90,  98,  64 },
+        {  34,  22, 116, 206,  23,  34,  43, 166,  73 },
+        {  68,  25, 106,  22,  64, 171,  36, 225, 114 },
+        {  34,  19,  21, 102, 132, 188,  16,  76, 124 },
+        {  62,  18,  78,  95,  85,  57,  50,  48,  51 }
+    },
+    {
+        { 193, 101,  35, 159, 215, 111,  89,  46, 111 },
+        {  60, 148,  31, 172, 219, 228,  21,  18, 111 },
+        { 112, 113,  77,  85, 179, 255,  38, 120, 114 },
+        {  40,  42,   1, 196, 245, 209,  10,  25, 109 },
+        { 100,  80,   8,  43, 154,   1,  51,  26,  71 },
+        {  88,  43,  29, 140, 166, 213,  37,  43, 154 },
+        {  61,  63,  30, 155,  67,  45,  68,   1, 209 },
+        { 142,  78,  78,  16, 255, 128,  34, 197, 171 },
+        {  41,  40,   5, 102, 211, 183,   4,   1, 221 },
+        {  51,  50,  17, 168, 209, 192,  23,  25,  82 }
+    },
+    {
+        { 125,  98,  42,  88, 104,  85, 117, 175,  82 },
+        {  95,  84,  53,  89, 128, 100, 113, 101,  45 },
+        {  75,  79, 123,  47,  51, 128,  81, 171,   1 },
+        {  57,  17,   5,  71, 102,  57,  53,  41,  49 },
+        { 115,  21,   2,  10, 102, 255, 166,  23,   6 },
+        {  38,  33,  13, 121,  57,  73,  26,   1,  85 },
+        {  41,  10,  67, 138,  77, 110,  90,  47, 114 },
+        { 101,  29,  16,  10,  85, 128, 101, 196,  26 },
+        {  57,  18,  10, 102, 102, 213,  34,  20,  43 },
+        { 117,  20,  15,  36, 163, 128,  68,   1,  26 }
+    },
+    {
+        { 138,  31,  36, 171,  27, 166,  38,  44, 229 },
+        {  67,  87,  58, 169,  82, 115,  26,  59, 179 },
+        {  63,  59,  90, 180,  59, 166,  93,  73, 154 },
+        {  40,  40,  21, 116, 143, 209,  34,  39, 175 },
+        {  57,  46,  22,  24, 128,   1,  54,  17,  37 },
+        {  47,  15,  16, 183,  34, 223,  49,  45, 183 },
+        {  46,  17,  33, 183,   6,  98,  15,  32, 183 },
+        {  65,  32,  73, 115,  28, 128,  23, 128, 205 },
+        {  40,   3,   9, 115,  51, 192,  18,   6, 223 },
+        {  87,  37,   9, 115,  59,  77,  64,  21,  47 }
+    },
+    {
+        { 104,  55,  44, 218,   9,  54,  53, 130, 226 },
+        {  64,  90,  70, 205,  40,  41,  23,  26,  57 },
+        {  54,  57, 112, 184,   5,  41,  38, 166, 213 },
+        {  30,  34,  26, 133, 152, 116,  10,  32, 134 },
+        {  75,  32,  12,  51, 192, 255, 160,  43,  51 },
+        {  39,  19,  53, 221,  26, 114,  32,  73, 255 },
+        {  31,   9,  65, 234,   2,  15,   1, 118,  73 },
+        {  88,  31,  35,  67, 102,  85,  55, 186,  85 },
+        {  56,  21,  23, 111,  59, 205,  45,  37, 192 },
+        {  55,  38,  70, 124,  73, 102,   1,  34,  98 }
+    },
+    {
+        { 102,  61,  71,  37,  34,  53,  31, 243, 192 },
+        {  69,  60,  71,  38,  73, 119,  28, 222,  37 },
+        {  68,  45, 128,  34,   1,  47,  11, 245, 171 },
+        {  62,  17,  19,  70, 146,  85,  55,  62,  70 },
+        {  75,  15,   9,   9,  64, 255, 184, 119,  16 },
+        {  37,  43,  37, 154, 100, 163,  85, 160,   1 },
+        {  63,   9,  92, 136,  28,  64,  32, 201,  85 },
+        {  86,   6,  28,   5,  64, 255,  25, 248,   1 },
+        {  56,   8,  17, 132, 137, 255,  55, 116, 128 },
+        {  58,  15,  20,  82, 135,  57,  26, 121,  40 }
+    },
+    {
+        { 164,  50,  31, 137, 154, 133,  25,  35, 218 },
+        {  51, 103,  44, 131, 131, 123,  31,   6, 158 },
+        {  86,  40,  64, 135, 148, 224,  45, 183, 128 },
+        {  22,  26,  17, 131, 240, 154,  14,   1, 209 },
+        {  83,  12,  13,  54, 192, 255,  68,  47,  28 },
+        {  45,  16,  21,  91,  64, 222,   7,   1, 197 },
+        {  56,  21,  39, 155,  60, 138,  23, 102, 213 },
+        {  85,  26,  85,  85, 128, 128,  32, 146, 171 },
+        {  18,  11,   7,  63, 144, 171,   4,   4, 246 },
+        {  35,  27,  10, 146, 174, 171,  12,  26, 128 }
+    },
+    {
+        { 190,  80,  35,  99, 180,  80, 126,  54,  45 },
+        {  85, 126,  47,  87, 176,  51,  41,  20,  32 },
+        { 101,  75, 128, 139, 118, 146, 116, 128,  85 },
+        {  56,  41,  15, 176, 236,  85,  37,   9,  62 },
+        { 146,  36,  19,  30, 171, 255,  97,  27,  20 },
+        {  71,  30,  17, 119, 118, 255,  17,  18, 138 },
+        { 101,  38,  60, 138,  55,  70,  43,  26, 142 },
+        { 138,  45,  61,  62, 219,   1,  81, 188,  64 },
+        {  32,  41,  20, 117, 151, 142,  20,  21, 163 },
+        { 112,  19,  12,  61, 195, 128,  48,   4,  24 }
+    }
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_VP8_ENTROPYMODEDATA_H_
diff --git a/libvpx/libvpx/vp8/common/vp8_loopfilter.c b/libvpx/libvpx/vp8/common/vp8_loopfilter.c
new file mode 100644
index 0000000..756ad48
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/vp8_loopfilter.c
@@ -0,0 +1,661 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "loopfilter.h"
+#include "onyxc_int.h"
+#include "vpx_mem/vpx_mem.h"
+
+
+static void lf_init_lut(loop_filter_info_n *lfi)
+{
+    int filt_lvl;
+
+    for (filt_lvl = 0; filt_lvl <= MAX_LOOP_FILTER; filt_lvl++)
+    {
+        if (filt_lvl >= 40)
+        {
+            lfi->hev_thr_lut[KEY_FRAME][filt_lvl] = 2;
+            lfi->hev_thr_lut[INTER_FRAME][filt_lvl] = 3;
+        }
+        else if (filt_lvl >= 20)
+        {
+            lfi->hev_thr_lut[KEY_FRAME][filt_lvl] = 1;
+            lfi->hev_thr_lut[INTER_FRAME][filt_lvl] = 2;
+        }
+        else if (filt_lvl >= 15)
+        {
+            lfi->hev_thr_lut[KEY_FRAME][filt_lvl] = 1;
+            lfi->hev_thr_lut[INTER_FRAME][filt_lvl] = 1;
+        }
+        else
+        {
+            lfi->hev_thr_lut[KEY_FRAME][filt_lvl] = 0;
+            lfi->hev_thr_lut[INTER_FRAME][filt_lvl] = 0;
+        }
+    }
+
+    lfi->mode_lf_lut[DC_PRED] = 1;
+    lfi->mode_lf_lut[V_PRED] = 1;
+    lfi->mode_lf_lut[H_PRED] = 1;
+    lfi->mode_lf_lut[TM_PRED] = 1;
+    lfi->mode_lf_lut[B_PRED]  = 0;
+
+    lfi->mode_lf_lut[ZEROMV]  = 1;
+    lfi->mode_lf_lut[NEARESTMV] = 2;
+    lfi->mode_lf_lut[NEARMV] = 2;
+    lfi->mode_lf_lut[NEWMV] = 2;
+    lfi->mode_lf_lut[SPLITMV] = 3;
+
+}
+
+void vp8_loop_filter_update_sharpness(loop_filter_info_n *lfi,
+                                      int sharpness_lvl)
+{
+    int i;
+
+    /* For each possible value for the loop filter fill out limits */
+    for (i = 0; i <= MAX_LOOP_FILTER; i++)
+    {
+        int filt_lvl = i;
+        int block_inside_limit = 0;
+
+        /* Set loop filter paramaeters that control sharpness. */
+        block_inside_limit = filt_lvl >> (sharpness_lvl > 0);
+        block_inside_limit = block_inside_limit >> (sharpness_lvl > 4);
+
+        if (sharpness_lvl > 0)
+        {
+            if (block_inside_limit > (9 - sharpness_lvl))
+                block_inside_limit = (9 - sharpness_lvl);
+        }
+
+        if (block_inside_limit < 1)
+            block_inside_limit = 1;
+
+        memset(lfi->lim[i], block_inside_limit, SIMD_WIDTH);
+        memset(lfi->blim[i], (2 * filt_lvl + block_inside_limit), SIMD_WIDTH);
+        memset(lfi->mblim[i], (2 * (filt_lvl + 2) + block_inside_limit),
+               SIMD_WIDTH);
+    }
+}
+
+void vp8_loop_filter_init(VP8_COMMON *cm)
+{
+    loop_filter_info_n *lfi = &cm->lf_info;
+    int i;
+
+    /* init limits for given sharpness*/
+    vp8_loop_filter_update_sharpness(lfi, cm->sharpness_level);
+    cm->last_sharpness_level = cm->sharpness_level;
+
+    /* init LUT for lvl  and hev thr picking */
+    lf_init_lut(lfi);
+
+    /* init hev threshold const vectors */
+    for(i = 0; i < 4 ; i++)
+    {
+        memset(lfi->hev_thr[i], i, SIMD_WIDTH);
+    }
+}
+
+void vp8_loop_filter_frame_init(VP8_COMMON *cm,
+                                MACROBLOCKD *mbd,
+                                int default_filt_lvl)
+{
+    int seg,  /* segment number */
+        ref,  /* index in ref_lf_deltas */
+        mode; /* index in mode_lf_deltas */
+
+    loop_filter_info_n *lfi = &cm->lf_info;
+
+    /* update limits if sharpness has changed */
+    if(cm->last_sharpness_level != cm->sharpness_level)
+    {
+        vp8_loop_filter_update_sharpness(lfi, cm->sharpness_level);
+        cm->last_sharpness_level = cm->sharpness_level;
+    }
+
+    for(seg = 0; seg < MAX_MB_SEGMENTS; seg++)
+    {
+        int lvl_seg = default_filt_lvl;
+        int lvl_ref, lvl_mode;
+
+        /* Note the baseline filter values for each segment */
+        if (mbd->segmentation_enabled)
+        {
+            /* Abs value */
+            if (mbd->mb_segement_abs_delta == SEGMENT_ABSDATA)
+            {
+                lvl_seg = mbd->segment_feature_data[MB_LVL_ALT_LF][seg];
+            }
+            else  /* Delta Value */
+            {
+                lvl_seg += mbd->segment_feature_data[MB_LVL_ALT_LF][seg];
+            }
+            lvl_seg = (lvl_seg > 0) ? ((lvl_seg > 63) ? 63: lvl_seg) : 0;
+        }
+
+        if (!mbd->mode_ref_lf_delta_enabled)
+        {
+            /* we could get rid of this if we assume that deltas are set to
+             * zero when not in use; encoder always uses deltas
+             */
+            memset(lfi->lvl[seg][0], lvl_seg, 4 * 4 );
+            continue;
+        }
+
+        /* INTRA_FRAME */
+        ref = INTRA_FRAME;
+
+        /* Apply delta for reference frame */
+        lvl_ref = lvl_seg + mbd->ref_lf_deltas[ref];
+
+        /* Apply delta for Intra modes */
+        mode = 0; /* B_PRED */
+        /* Only the split mode BPRED has a further special case */
+        lvl_mode = lvl_ref + mbd->mode_lf_deltas[mode];
+        /* clamp */
+        lvl_mode = (lvl_mode > 0) ? (lvl_mode > 63 ? 63 : lvl_mode) : 0;
+
+        lfi->lvl[seg][ref][mode] = lvl_mode;
+
+        mode = 1; /* all the rest of Intra modes */
+        /* clamp */
+        lvl_mode = (lvl_ref > 0) ? (lvl_ref > 63 ? 63 : lvl_ref) : 0;
+        lfi->lvl[seg][ref][mode] = lvl_mode;
+
+        /* LAST, GOLDEN, ALT */
+        for(ref = 1; ref < MAX_REF_FRAMES; ref++)
+        {
+            /* Apply delta for reference frame */
+            lvl_ref = lvl_seg + mbd->ref_lf_deltas[ref];
+
+            /* Apply delta for Inter modes */
+            for (mode = 1; mode < 4; mode++)
+            {
+                lvl_mode = lvl_ref + mbd->mode_lf_deltas[mode];
+                /* clamp */
+                lvl_mode = (lvl_mode > 0) ? (lvl_mode > 63 ? 63 : lvl_mode) : 0;
+
+                lfi->lvl[seg][ref][mode] = lvl_mode;
+            }
+        }
+    }
+}
+
+
+void vp8_loop_filter_row_normal(VP8_COMMON *cm, MODE_INFO *mode_info_context,
+                         int mb_row, int post_ystride, int post_uvstride,
+                         unsigned char *y_ptr, unsigned char *u_ptr,
+                         unsigned char *v_ptr)
+{
+    int mb_col;
+    int filter_level;
+    loop_filter_info_n *lfi_n = &cm->lf_info;
+    loop_filter_info lfi;
+    FRAME_TYPE frame_type = cm->frame_type;
+
+    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
+    {
+        int skip_lf = (mode_info_context->mbmi.mode != B_PRED &&
+                        mode_info_context->mbmi.mode != SPLITMV &&
+                        mode_info_context->mbmi.mb_skip_coeff);
+
+        const int mode_index = lfi_n->mode_lf_lut[mode_info_context->mbmi.mode];
+        const int seg = mode_info_context->mbmi.segment_id;
+        const int ref_frame = mode_info_context->mbmi.ref_frame;
+
+        filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
+
+        if (filter_level)
+        {
+            const int hev_index = lfi_n->hev_thr_lut[frame_type][filter_level];
+            lfi.mblim = lfi_n->mblim[filter_level];
+            lfi.blim = lfi_n->blim[filter_level];
+            lfi.lim = lfi_n->lim[filter_level];
+            lfi.hev_thr = lfi_n->hev_thr[hev_index];
+
+            if (mb_col > 0)
+                vp8_loop_filter_mbv
+                (y_ptr, u_ptr, v_ptr, post_ystride, post_uvstride, &lfi);
+
+            if (!skip_lf)
+                vp8_loop_filter_bv
+                (y_ptr, u_ptr, v_ptr, post_ystride, post_uvstride, &lfi);
+
+            /* don't apply across umv border */
+            if (mb_row > 0)
+                vp8_loop_filter_mbh
+                (y_ptr, u_ptr, v_ptr, post_ystride, post_uvstride, &lfi);
+
+            if (!skip_lf)
+                vp8_loop_filter_bh
+                (y_ptr, u_ptr, v_ptr, post_ystride, post_uvstride, &lfi);
+        }
+
+        y_ptr += 16;
+        u_ptr += 8;
+        v_ptr += 8;
+
+        mode_info_context++;     /* step to next MB */
+    }
+
+}
+
+void vp8_loop_filter_row_simple(VP8_COMMON *cm, MODE_INFO *mode_info_context,
+                         int mb_row, int post_ystride, int post_uvstride,
+                         unsigned char *y_ptr, unsigned char *u_ptr,
+                         unsigned char *v_ptr)
+{
+    int mb_col;
+    int filter_level;
+    loop_filter_info_n *lfi_n = &cm->lf_info;
+    (void)post_uvstride;
+
+    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
+    {
+        int skip_lf = (mode_info_context->mbmi.mode != B_PRED &&
+                        mode_info_context->mbmi.mode != SPLITMV &&
+                        mode_info_context->mbmi.mb_skip_coeff);
+
+        const int mode_index = lfi_n->mode_lf_lut[mode_info_context->mbmi.mode];
+        const int seg = mode_info_context->mbmi.segment_id;
+        const int ref_frame = mode_info_context->mbmi.ref_frame;
+
+        filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
+
+        if (filter_level)
+        {
+            if (mb_col > 0)
+                vp8_loop_filter_simple_mbv
+                (y_ptr, post_ystride, lfi_n->mblim[filter_level]);
+
+            if (!skip_lf)
+                vp8_loop_filter_simple_bv
+                (y_ptr, post_ystride, lfi_n->blim[filter_level]);
+
+            /* don't apply across umv border */
+            if (mb_row > 0)
+                vp8_loop_filter_simple_mbh
+                (y_ptr, post_ystride, lfi_n->mblim[filter_level]);
+
+            if (!skip_lf)
+                vp8_loop_filter_simple_bh
+                (y_ptr, post_ystride, lfi_n->blim[filter_level]);
+        }
+
+        y_ptr += 16;
+        u_ptr += 8;
+        v_ptr += 8;
+
+        mode_info_context++;     /* step to next MB */
+    }
+
+}
+void vp8_loop_filter_frame(VP8_COMMON *cm,
+                           MACROBLOCKD *mbd,
+                           int frame_type)
+{
+    YV12_BUFFER_CONFIG *post = cm->frame_to_show;
+    loop_filter_info_n *lfi_n = &cm->lf_info;
+    loop_filter_info lfi;
+
+    int mb_row;
+    int mb_col;
+    int mb_rows = cm->mb_rows;
+    int mb_cols = cm->mb_cols;
+
+    int filter_level;
+
+    unsigned char *y_ptr, *u_ptr, *v_ptr;
+
+    /* Point at base of Mb MODE_INFO list */
+    const MODE_INFO *mode_info_context = cm->mi;
+    int post_y_stride = post->y_stride;
+    int post_uv_stride = post->uv_stride;
+
+    /* Initialize the loop filter for this frame. */
+    vp8_loop_filter_frame_init(cm, mbd, cm->filter_level);
+
+    /* Set up the buffer pointers */
+    y_ptr = post->y_buffer;
+    u_ptr = post->u_buffer;
+    v_ptr = post->v_buffer;
+
+    /* vp8_filter each macro block */
+    if (cm->filter_type == NORMAL_LOOPFILTER)
+    {
+        for (mb_row = 0; mb_row < mb_rows; mb_row++)
+        {
+            for (mb_col = 0; mb_col < mb_cols; mb_col++)
+            {
+                int skip_lf = (mode_info_context->mbmi.mode != B_PRED &&
+                                mode_info_context->mbmi.mode != SPLITMV &&
+                                mode_info_context->mbmi.mb_skip_coeff);
+
+                const int mode_index = lfi_n->mode_lf_lut[mode_info_context->mbmi.mode];
+                const int seg = mode_info_context->mbmi.segment_id;
+                const int ref_frame = mode_info_context->mbmi.ref_frame;
+
+                filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
+
+                if (filter_level)
+                {
+                    const int hev_index = lfi_n->hev_thr_lut[frame_type][filter_level];
+                    lfi.mblim = lfi_n->mblim[filter_level];
+                    lfi.blim = lfi_n->blim[filter_level];
+                    lfi.lim = lfi_n->lim[filter_level];
+                    lfi.hev_thr = lfi_n->hev_thr[hev_index];
+
+                    if (mb_col > 0)
+                        vp8_loop_filter_mbv
+                        (y_ptr, u_ptr, v_ptr, post_y_stride, post_uv_stride, &lfi);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_bv
+                        (y_ptr, u_ptr, v_ptr, post_y_stride, post_uv_stride, &lfi);
+
+                    /* don't apply across umv border */
+                    if (mb_row > 0)
+                        vp8_loop_filter_mbh
+                        (y_ptr, u_ptr, v_ptr, post_y_stride, post_uv_stride, &lfi);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_bh
+                        (y_ptr, u_ptr, v_ptr, post_y_stride, post_uv_stride, &lfi);
+                }
+
+                y_ptr += 16;
+                u_ptr += 8;
+                v_ptr += 8;
+
+                mode_info_context++;     /* step to next MB */
+            }
+            y_ptr += post_y_stride  * 16 - post->y_width;
+            u_ptr += post_uv_stride *  8 - post->uv_width;
+            v_ptr += post_uv_stride *  8 - post->uv_width;
+
+            mode_info_context++;         /* Skip border mb */
+
+        }
+    }
+    else /* SIMPLE_LOOPFILTER */
+    {
+        for (mb_row = 0; mb_row < mb_rows; mb_row++)
+        {
+            for (mb_col = 0; mb_col < mb_cols; mb_col++)
+            {
+                int skip_lf = (mode_info_context->mbmi.mode != B_PRED &&
+                                mode_info_context->mbmi.mode != SPLITMV &&
+                                mode_info_context->mbmi.mb_skip_coeff);
+
+                const int mode_index = lfi_n->mode_lf_lut[mode_info_context->mbmi.mode];
+                const int seg = mode_info_context->mbmi.segment_id;
+                const int ref_frame = mode_info_context->mbmi.ref_frame;
+
+                filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
+                if (filter_level)
+                {
+                    const unsigned char * mblim = lfi_n->mblim[filter_level];
+                    const unsigned char * blim = lfi_n->blim[filter_level];
+
+                    if (mb_col > 0)
+                        vp8_loop_filter_simple_mbv
+                        (y_ptr, post_y_stride, mblim);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_simple_bv
+                        (y_ptr, post_y_stride, blim);
+
+                    /* don't apply across umv border */
+                    if (mb_row > 0)
+                        vp8_loop_filter_simple_mbh
+                        (y_ptr, post_y_stride, mblim);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_simple_bh
+                        (y_ptr, post_y_stride, blim);
+                }
+
+                y_ptr += 16;
+                u_ptr += 8;
+                v_ptr += 8;
+
+                mode_info_context++;     /* step to next MB */
+            }
+            y_ptr += post_y_stride  * 16 - post->y_width;
+            u_ptr += post_uv_stride *  8 - post->uv_width;
+            v_ptr += post_uv_stride *  8 - post->uv_width;
+
+            mode_info_context++;         /* Skip border mb */
+
+        }
+    }
+}
+
+void vp8_loop_filter_frame_yonly
+(
+    VP8_COMMON *cm,
+    MACROBLOCKD *mbd,
+    int default_filt_lvl
+)
+{
+    YV12_BUFFER_CONFIG *post = cm->frame_to_show;
+
+    unsigned char *y_ptr;
+    int mb_row;
+    int mb_col;
+
+    loop_filter_info_n *lfi_n = &cm->lf_info;
+    loop_filter_info lfi;
+
+    int filter_level;
+    FRAME_TYPE frame_type = cm->frame_type;
+
+    /* Point at base of Mb MODE_INFO list */
+    const MODE_INFO *mode_info_context = cm->mi;
+
+#if 0
+    if(default_filt_lvl == 0) /* no filter applied */
+        return;
+#endif
+
+    /* Initialize the loop filter for this frame. */
+    vp8_loop_filter_frame_init( cm, mbd, default_filt_lvl);
+
+    /* Set up the buffer pointers */
+    y_ptr = post->y_buffer;
+
+    /* vp8_filter each macro block */
+    for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
+    {
+        for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
+        {
+            int skip_lf = (mode_info_context->mbmi.mode != B_PRED &&
+                            mode_info_context->mbmi.mode != SPLITMV &&
+                            mode_info_context->mbmi.mb_skip_coeff);
+
+            const int mode_index = lfi_n->mode_lf_lut[mode_info_context->mbmi.mode];
+            const int seg = mode_info_context->mbmi.segment_id;
+            const int ref_frame = mode_info_context->mbmi.ref_frame;
+
+            filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
+
+            if (filter_level)
+            {
+                if (cm->filter_type == NORMAL_LOOPFILTER)
+                {
+                    const int hev_index = lfi_n->hev_thr_lut[frame_type][filter_level];
+                    lfi.mblim = lfi_n->mblim[filter_level];
+                    lfi.blim = lfi_n->blim[filter_level];
+                    lfi.lim = lfi_n->lim[filter_level];
+                    lfi.hev_thr = lfi_n->hev_thr[hev_index];
+
+                    if (mb_col > 0)
+                        vp8_loop_filter_mbv
+                        (y_ptr, 0, 0, post->y_stride, 0, &lfi);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_bv
+                        (y_ptr, 0, 0, post->y_stride, 0, &lfi);
+
+                    /* don't apply across umv border */
+                    if (mb_row > 0)
+                        vp8_loop_filter_mbh
+                        (y_ptr, 0, 0, post->y_stride, 0, &lfi);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_bh
+                        (y_ptr, 0, 0, post->y_stride, 0, &lfi);
+                }
+                else
+                {
+                    if (mb_col > 0)
+                        vp8_loop_filter_simple_mbv
+                        (y_ptr, post->y_stride, lfi_n->mblim[filter_level]);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_simple_bv
+                        (y_ptr, post->y_stride, lfi_n->blim[filter_level]);
+
+                    /* don't apply across umv border */
+                    if (mb_row > 0)
+                        vp8_loop_filter_simple_mbh
+                        (y_ptr, post->y_stride, lfi_n->mblim[filter_level]);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_simple_bh
+                        (y_ptr, post->y_stride, lfi_n->blim[filter_level]);
+                }
+            }
+
+            y_ptr += 16;
+            mode_info_context ++;        /* step to next MB */
+
+        }
+
+        y_ptr += post->y_stride  * 16 - post->y_width;
+        mode_info_context ++;            /* Skip border mb */
+    }
+
+}
+
+void vp8_loop_filter_partial_frame
+(
+    VP8_COMMON *cm,
+    MACROBLOCKD *mbd,
+    int default_filt_lvl
+)
+{
+    YV12_BUFFER_CONFIG *post = cm->frame_to_show;
+
+    unsigned char *y_ptr;
+    int mb_row;
+    int mb_col;
+    int mb_cols = post->y_width >> 4;
+    int mb_rows = post->y_height >> 4;
+
+    int linestocopy;
+
+    loop_filter_info_n *lfi_n = &cm->lf_info;
+    loop_filter_info lfi;
+
+    int filter_level;
+    FRAME_TYPE frame_type = cm->frame_type;
+
+    const MODE_INFO *mode_info_context;
+
+#if 0
+    if(default_filt_lvl == 0) /* no filter applied */
+        return;
+#endif
+
+    /* Initialize the loop filter for this frame. */
+    vp8_loop_filter_frame_init( cm, mbd, default_filt_lvl);
+
+    /* number of MB rows to use in partial filtering */
+    linestocopy = mb_rows / PARTIAL_FRAME_FRACTION;
+    linestocopy = linestocopy ? linestocopy << 4 : 16;     /* 16 lines per MB */
+
+    /* Set up the buffer pointers; partial image starts at ~middle of frame */
+    y_ptr = post->y_buffer + ((post->y_height >> 5) * 16) * post->y_stride;
+    mode_info_context = cm->mi + (post->y_height >> 5) * (mb_cols + 1);
+
+    /* vp8_filter each macro block */
+    for (mb_row = 0; mb_row<(linestocopy >> 4); mb_row++)
+    {
+        for (mb_col = 0; mb_col < mb_cols; mb_col++)
+        {
+            int skip_lf = (mode_info_context->mbmi.mode != B_PRED &&
+                           mode_info_context->mbmi.mode != SPLITMV &&
+                           mode_info_context->mbmi.mb_skip_coeff);
+
+            const int mode_index =
+                lfi_n->mode_lf_lut[mode_info_context->mbmi.mode];
+            const int seg = mode_info_context->mbmi.segment_id;
+            const int ref_frame = mode_info_context->mbmi.ref_frame;
+
+            filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
+
+            if (filter_level)
+            {
+                if (cm->filter_type == NORMAL_LOOPFILTER)
+                {
+                    const int hev_index = lfi_n->hev_thr_lut[frame_type][filter_level];
+                    lfi.mblim = lfi_n->mblim[filter_level];
+                    lfi.blim = lfi_n->blim[filter_level];
+                    lfi.lim = lfi_n->lim[filter_level];
+                    lfi.hev_thr = lfi_n->hev_thr[hev_index];
+
+                    if (mb_col > 0)
+                        vp8_loop_filter_mbv
+                        (y_ptr, 0, 0, post->y_stride, 0, &lfi);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_bv
+                        (y_ptr, 0, 0, post->y_stride, 0, &lfi);
+
+                    vp8_loop_filter_mbh
+                        (y_ptr, 0, 0, post->y_stride, 0, &lfi);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_bh
+                        (y_ptr, 0, 0, post->y_stride, 0, &lfi);
+                }
+                else
+                {
+                    if (mb_col > 0)
+                        vp8_loop_filter_simple_mbv
+                        (y_ptr, post->y_stride, lfi_n->mblim[filter_level]);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_simple_bv
+                        (y_ptr, post->y_stride, lfi_n->blim[filter_level]);
+
+                    vp8_loop_filter_simple_mbh
+                        (y_ptr, post->y_stride, lfi_n->mblim[filter_level]);
+
+                    if (!skip_lf)
+                        vp8_loop_filter_simple_bh
+                        (y_ptr, post->y_stride, lfi_n->blim[filter_level]);
+                }
+            }
+
+            y_ptr += 16;
+            mode_info_context += 1;      /* step to next MB */
+        }
+
+        y_ptr += post->y_stride  * 16 - post->y_width;
+        mode_info_context += 1;          /* Skip border mb */
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/x86/copy_sse2.asm b/libvpx/libvpx/vp8/common/x86/copy_sse2.asm
new file mode 100644
index 0000000..86fae26
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/copy_sse2.asm
@@ -0,0 +1,93 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+
+;void vp8_copy32xn_sse2(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *dst_ptr,
+;    int  dst_stride,
+;    int height);
+global sym(vp8_copy32xn_sse2) PRIVATE
+sym(vp8_copy32xn_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov             rsi,        arg(0) ;src_ptr
+        mov             rdi,        arg(2) ;dst_ptr
+
+        movsxd          rax,        dword ptr arg(1) ;src_stride
+        movsxd          rdx,        dword ptr arg(3) ;dst_stride
+        movsxd          rcx,        dword ptr arg(4) ;height
+
+.block_copy_sse2_loopx4:
+        movdqu          xmm0,       XMMWORD PTR [rsi]
+        movdqu          xmm1,       XMMWORD PTR [rsi + 16]
+        movdqu          xmm2,       XMMWORD PTR [rsi + rax]
+        movdqu          xmm3,       XMMWORD PTR [rsi + rax + 16]
+
+        lea             rsi,        [rsi+rax*2]
+
+        movdqu          xmm4,       XMMWORD PTR [rsi]
+        movdqu          xmm5,       XMMWORD PTR [rsi + 16]
+        movdqu          xmm6,       XMMWORD PTR [rsi + rax]
+        movdqu          xmm7,       XMMWORD PTR [rsi + rax + 16]
+
+        lea             rsi,    [rsi+rax*2]
+
+        movdqa          XMMWORD PTR [rdi], xmm0
+        movdqa          XMMWORD PTR [rdi + 16], xmm1
+        movdqa          XMMWORD PTR [rdi + rdx], xmm2
+        movdqa          XMMWORD PTR [rdi + rdx + 16], xmm3
+
+        lea             rdi,    [rdi+rdx*2]
+
+        movdqa          XMMWORD PTR [rdi], xmm4
+        movdqa          XMMWORD PTR [rdi + 16], xmm5
+        movdqa          XMMWORD PTR [rdi + rdx], xmm6
+        movdqa          XMMWORD PTR [rdi + rdx + 16], xmm7
+
+        lea             rdi,    [rdi+rdx*2]
+
+        sub             rcx,     4
+        cmp             rcx,     4
+        jge             .block_copy_sse2_loopx4
+
+        cmp             rcx, 0
+        je              .copy_is_done
+
+.block_copy_sse2_loop:
+        movdqu          xmm0,       XMMWORD PTR [rsi]
+        movdqu          xmm1,       XMMWORD PTR [rsi + 16]
+        lea             rsi,    [rsi+rax]
+
+        movdqa          XMMWORD PTR [rdi], xmm0
+        movdqa          XMMWORD PTR [rdi + 16], xmm1
+        lea             rdi,    [rdi+rdx]
+
+        sub             rcx,     1
+        jne             .block_copy_sse2_loop
+
+.copy_is_done:
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vp8/common/x86/copy_sse3.asm b/libvpx/libvpx/vp8/common/x86/copy_sse3.asm
new file mode 100644
index 0000000..d789a40
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/copy_sse3.asm
@@ -0,0 +1,146 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro STACK_FRAME_CREATE_X3 0
+%if ABI_IS_32BIT
+  %define     src_ptr       rsi
+  %define     src_stride    rax
+  %define     ref_ptr       rdi
+  %define     ref_stride    rdx
+  %define     end_ptr       rcx
+  %define     ret_var       rbx
+  %define     result_ptr    arg(4)
+  %define     max_sad       arg(4)
+  %define     height        dword ptr arg(4)
+    push        rbp
+    mov         rbp,        rsp
+    push        rsi
+    push        rdi
+    push        rbx
+
+    mov         rsi,        arg(0)              ; src_ptr
+    mov         rdi,        arg(2)              ; ref_ptr
+
+    movsxd      rax,        dword ptr arg(1)    ; src_stride
+    movsxd      rdx,        dword ptr arg(3)    ; ref_stride
+%else
+  %if LIBVPX_YASM_WIN64
+    SAVE_XMM 7, u
+    %define     src_ptr     rcx
+    %define     src_stride  rdx
+    %define     ref_ptr     r8
+    %define     ref_stride  r9
+    %define     end_ptr     r10
+    %define     ret_var     r11
+    %define     result_ptr  [rsp+xmm_stack_space+8+4*8]
+    %define     max_sad     [rsp+xmm_stack_space+8+4*8]
+    %define     height      dword ptr [rsp+xmm_stack_space+8+4*8]
+  %else
+    %define     src_ptr     rdi
+    %define     src_stride  rsi
+    %define     ref_ptr     rdx
+    %define     ref_stride  rcx
+    %define     end_ptr     r9
+    %define     ret_var     r10
+    %define     result_ptr  r8
+    %define     max_sad     r8
+    %define     height      r8
+  %endif
+%endif
+
+%endmacro
+
+%macro STACK_FRAME_DESTROY_X3 0
+  %define     src_ptr
+  %define     src_stride
+  %define     ref_ptr
+  %define     ref_stride
+  %define     end_ptr
+  %define     ret_var
+  %define     result_ptr
+  %define     max_sad
+  %define     height
+
+%if ABI_IS_32BIT
+    pop         rbx
+    pop         rdi
+    pop         rsi
+    pop         rbp
+%else
+  %if LIBVPX_YASM_WIN64
+    RESTORE_XMM
+  %endif
+%endif
+    ret
+%endmacro
+
+
+;void vp8_copy32xn_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *dst_ptr,
+;    int  dst_stride,
+;    int height);
+global sym(vp8_copy32xn_sse3) PRIVATE
+sym(vp8_copy32xn_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+.block_copy_sse3_loopx4:
+        lea             end_ptr,    [src_ptr+src_stride*2]
+
+        movdqu          xmm0,       XMMWORD PTR [src_ptr]
+        movdqu          xmm1,       XMMWORD PTR [src_ptr + 16]
+        movdqu          xmm2,       XMMWORD PTR [src_ptr + src_stride]
+        movdqu          xmm3,       XMMWORD PTR [src_ptr + src_stride + 16]
+        movdqu          xmm4,       XMMWORD PTR [end_ptr]
+        movdqu          xmm5,       XMMWORD PTR [end_ptr + 16]
+        movdqu          xmm6,       XMMWORD PTR [end_ptr + src_stride]
+        movdqu          xmm7,       XMMWORD PTR [end_ptr + src_stride + 16]
+
+        lea             src_ptr,    [src_ptr+src_stride*4]
+
+        lea             end_ptr,    [ref_ptr+ref_stride*2]
+
+        movdqa          XMMWORD PTR [ref_ptr], xmm0
+        movdqa          XMMWORD PTR [ref_ptr + 16], xmm1
+        movdqa          XMMWORD PTR [ref_ptr + ref_stride], xmm2
+        movdqa          XMMWORD PTR [ref_ptr + ref_stride + 16], xmm3
+        movdqa          XMMWORD PTR [end_ptr], xmm4
+        movdqa          XMMWORD PTR [end_ptr + 16], xmm5
+        movdqa          XMMWORD PTR [end_ptr + ref_stride], xmm6
+        movdqa          XMMWORD PTR [end_ptr + ref_stride + 16], xmm7
+
+        lea             ref_ptr,    [ref_ptr+ref_stride*4]
+
+        sub             height,     4
+        cmp             height,     4
+        jge             .block_copy_sse3_loopx4
+
+        ;Check to see if there is more rows need to be copied.
+        cmp             height, 0
+        je              .copy_is_done
+
+.block_copy_sse3_loop:
+        movdqu          xmm0,       XMMWORD PTR [src_ptr]
+        movdqu          xmm1,       XMMWORD PTR [src_ptr + 16]
+        lea             src_ptr,    [src_ptr+src_stride]
+
+        movdqa          XMMWORD PTR [ref_ptr], xmm0
+        movdqa          XMMWORD PTR [ref_ptr + 16], xmm1
+        lea             ref_ptr,    [ref_ptr+ref_stride]
+
+        sub             height,     1
+        jne             .block_copy_sse3_loop
+
+.copy_is_done:
+    STACK_FRAME_DESTROY_X3
diff --git a/libvpx/libvpx/vp8/common/x86/dequantize_mmx.asm b/libvpx/libvpx/vp8/common/x86/dequantize_mmx.asm
new file mode 100644
index 0000000..4e551f0
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/dequantize_mmx.asm
@@ -0,0 +1,258 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+
+;void vp8_dequantize_b_impl_mmx(short *sq, short *dq, short *q)
+global sym(vp8_dequantize_b_impl_mmx) PRIVATE
+sym(vp8_dequantize_b_impl_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 3
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov       rsi, arg(0) ;sq
+        mov       rdi, arg(1) ;dq
+        mov       rax, arg(2) ;q
+
+        movq      mm1, [rsi]
+        pmullw    mm1, [rax+0]            ; mm4 *= kernel 0 modifiers.
+        movq      [rdi], mm1
+
+        movq      mm1, [rsi+8]
+        pmullw    mm1, [rax+8]            ; mm4 *= kernel 0 modifiers.
+        movq      [rdi+8], mm1
+
+        movq      mm1, [rsi+16]
+        pmullw    mm1, [rax+16]            ; mm4 *= kernel 0 modifiers.
+        movq      [rdi+16], mm1
+
+        movq      mm1, [rsi+24]
+        pmullw    mm1, [rax+24]            ; mm4 *= kernel 0 modifiers.
+        movq      [rdi+24], mm1
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void dequant_idct_add_mmx(
+;short *input,            0
+;short *dq,               1
+;unsigned char *dest,     2
+;int stride)              3
+global sym(vp8_dequant_idct_add_mmx) PRIVATE
+sym(vp8_dequant_idct_add_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 4
+    GET_GOT     rbx
+    push        rdi
+    ; end prolog
+
+        mov         rax,    arg(0) ;input
+        mov         rdx,    arg(1) ;dq
+
+
+        movq        mm0,    [rax   ]
+        pmullw      mm0,    [rdx]
+
+        movq        mm1,    [rax +8]
+        pmullw      mm1,    [rdx +8]
+
+        movq        mm2,    [rax+16]
+        pmullw      mm2,    [rdx+16]
+
+        movq        mm3,    [rax+24]
+        pmullw      mm3,    [rdx+24]
+
+        mov         rdx,    arg(2) ;dest
+
+        pxor        mm7,    mm7
+
+
+        movq        [rax],   mm7
+        movq        [rax+8], mm7
+
+        movq        [rax+16],mm7
+        movq        [rax+24],mm7
+
+
+        movsxd      rdi,            dword ptr arg(3) ;stride
+
+        psubw       mm0,            mm2             ; b1= 0-2
+        paddw       mm2,            mm2             ;
+
+        movq        mm5,            mm1
+        paddw       mm2,            mm0             ; a1 =0+2
+
+        pmulhw      mm5,            [GLOBAL(x_s1sqr2)];
+        paddw       mm5,            mm1             ; ip1 * sin(pi/8) * sqrt(2)
+
+        movq        mm7,            mm3             ;
+        pmulhw      mm7,            [GLOBAL(x_c1sqr2less1)];
+
+        paddw       mm7,            mm3             ; ip3 * cos(pi/8) * sqrt(2)
+        psubw       mm7,            mm5             ; c1
+
+        movq        mm5,            mm1
+        movq        mm4,            mm3
+
+        pmulhw      mm5,            [GLOBAL(x_c1sqr2less1)]
+        paddw       mm5,            mm1
+
+        pmulhw      mm3,            [GLOBAL(x_s1sqr2)]
+        paddw       mm3,            mm4
+
+        paddw       mm3,            mm5             ; d1
+        movq        mm6,            mm2             ; a1
+
+        movq        mm4,            mm0             ; b1
+        paddw       mm2,            mm3             ;0
+
+        paddw       mm4,            mm7             ;1
+        psubw       mm0,            mm7             ;2
+
+        psubw       mm6,            mm3             ;3
+
+        movq        mm1,            mm2             ; 03 02 01 00
+        movq        mm3,            mm4             ; 23 22 21 20
+
+        punpcklwd   mm1,            mm0             ; 11 01 10 00
+        punpckhwd   mm2,            mm0             ; 13 03 12 02
+
+        punpcklwd   mm3,            mm6             ; 31 21 30 20
+        punpckhwd   mm4,            mm6             ; 33 23 32 22
+
+        movq        mm0,            mm1             ; 11 01 10 00
+        movq        mm5,            mm2             ; 13 03 12 02
+
+        punpckldq   mm0,            mm3             ; 30 20 10 00
+        punpckhdq   mm1,            mm3             ; 31 21 11 01
+
+        punpckldq   mm2,            mm4             ; 32 22 12 02
+        punpckhdq   mm5,            mm4             ; 33 23 13 03
+
+        movq        mm3,            mm5             ; 33 23 13 03
+
+        psubw       mm0,            mm2             ; b1= 0-2
+        paddw       mm2,            mm2             ;
+
+        movq        mm5,            mm1
+        paddw       mm2,            mm0             ; a1 =0+2
+
+        pmulhw      mm5,            [GLOBAL(x_s1sqr2)];
+        paddw       mm5,            mm1             ; ip1 * sin(pi/8) * sqrt(2)
+
+        movq        mm7,            mm3             ;
+        pmulhw      mm7,            [GLOBAL(x_c1sqr2less1)];
+
+        paddw       mm7,            mm3             ; ip3 * cos(pi/8) * sqrt(2)
+        psubw       mm7,            mm5             ; c1
+
+        movq        mm5,            mm1
+        movq        mm4,            mm3
+
+        pmulhw      mm5,            [GLOBAL(x_c1sqr2less1)]
+        paddw       mm5,            mm1
+
+        pmulhw      mm3,            [GLOBAL(x_s1sqr2)]
+        paddw       mm3,            mm4
+
+        paddw       mm3,            mm5             ; d1
+        paddw       mm0,            [GLOBAL(fours)]
+
+        paddw       mm2,            [GLOBAL(fours)]
+        movq        mm6,            mm2             ; a1
+
+        movq        mm4,            mm0             ; b1
+        paddw       mm2,            mm3             ;0
+
+        paddw       mm4,            mm7             ;1
+        psubw       mm0,            mm7             ;2
+
+        psubw       mm6,            mm3             ;3
+        psraw       mm2,            3
+
+        psraw       mm0,            3
+        psraw       mm4,            3
+
+        psraw       mm6,            3
+
+        movq        mm1,            mm2             ; 03 02 01 00
+        movq        mm3,            mm4             ; 23 22 21 20
+
+        punpcklwd   mm1,            mm0             ; 11 01 10 00
+        punpckhwd   mm2,            mm0             ; 13 03 12 02
+
+        punpcklwd   mm3,            mm6             ; 31 21 30 20
+        punpckhwd   mm4,            mm6             ; 33 23 32 22
+
+        movq        mm0,            mm1             ; 11 01 10 00
+        movq        mm5,            mm2             ; 13 03 12 02
+
+        punpckldq   mm0,            mm3             ; 30 20 10 00
+        punpckhdq   mm1,            mm3             ; 31 21 11 01
+
+        punpckldq   mm2,            mm4             ; 32 22 12 02
+        punpckhdq   mm5,            mm4             ; 33 23 13 03
+
+        pxor        mm7,            mm7
+
+        movd        mm4,            [rdx]
+        punpcklbw   mm4,            mm7
+        paddsw      mm0,            mm4
+        packuswb    mm0,            mm7
+        movd        [rdx],          mm0
+
+        movd        mm4,            [rdx+rdi]
+        punpcklbw   mm4,            mm7
+        paddsw      mm1,            mm4
+        packuswb    mm1,            mm7
+        movd        [rdx+rdi],      mm1
+
+        movd        mm4,            [rdx+2*rdi]
+        punpcklbw   mm4,            mm7
+        paddsw      mm2,            mm4
+        packuswb    mm2,            mm7
+        movd        [rdx+rdi*2],    mm2
+
+        add         rdx,            rdi
+
+        movd        mm4,            [rdx+2*rdi]
+        punpcklbw   mm4,            mm7
+        paddsw      mm5,            mm4
+        packuswb    mm5,            mm7
+        movd        [rdx+rdi*2],    mm5
+
+    ; begin epilog
+    pop rdi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+x_s1sqr2:
+    times 4 dw 0x8A8C
+align 16
+x_c1sqr2less1:
+    times 4 dw 0x4E7B
+align 16
+fours:
+    times 4 dw 0x0004
diff --git a/libvpx/libvpx/vp8/common/x86/filter_x86.c b/libvpx/libvpx/vp8/common/x86/filter_x86.c
new file mode 100644
index 0000000..7f496ed
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/filter_x86.c
@@ -0,0 +1,35 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp8/common/x86/filter_x86.h"
+
+DECLARE_ALIGNED(16, const short, vp8_bilinear_filters_x86_4[8][8]) =
+{
+    { 128, 128, 128, 128,   0,   0,   0,   0 },
+    { 112, 112, 112, 112,  16,  16,  16,  16 },
+    {  96,  96,  96,  96,  32,  32,  32,  32 },
+    {  80,  80,  80,  80,  48,  48,  48,  48 },
+    {  64,  64,  64,  64,  64,  64,  64,  64 },
+    {  48,  48,  48,  48,  80,  80,  80,  80 },
+    {  32,  32,  32,  32,  96,  96,  96,  96 },
+    {  16,  16,  16,  16, 112, 112, 112, 112 }
+};
+
+DECLARE_ALIGNED(16, const short, vp8_bilinear_filters_x86_8[8][16]) =
+{
+    { 128, 128, 128, 128, 128, 128, 128, 128,   0,   0,   0,   0,   0,   0,   0,   0 },
+    { 112, 112, 112, 112, 112, 112, 112, 112,  16,  16,  16,  16,  16,  16,  16,  16 },
+    {  96,  96,  96,  96,  96,  96,  96,  96,  32,  32,  32,  32,  32,  32,  32,  32 },
+    {  80,  80,  80,  80,  80,  80,  80,  80,  48,  48,  48,  48,  48,  48,  48,  48 },
+    {  64,  64,  64,  64,  64,  64,  64,  64,  64,  64,  64,  64,  64,  64,  64,  64 },
+    {  48,  48,  48,  48,  48,  48,  48,  48,  80,  80,  80,  80,  80,  80,  80,  80 },
+    {  32,  32,  32,  32,  32,  32,  32,  32,  96,  96,  96,  96,  96,  96,  96,  96 },
+    {  16,  16,  16,  16,  16,  16,  16,  16, 112, 112, 112, 112, 112, 112, 112, 112 }
+};
diff --git a/libvpx/libvpx/vp8/common/x86/filter_x86.h b/libvpx/libvpx/vp8/common/x86/filter_x86.h
new file mode 100644
index 0000000..d282841
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/filter_x86.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_COMMON_X86_FILTER_X86_H_
+#define VP8_COMMON_X86_FILTER_X86_H_
+
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* x86 assembly specific copy of vp8/common/filter.c:vp8_bilinear_filters with
+ * duplicated values */
+
+/* duplicated 4x */
+extern DECLARE_ALIGNED(16, const short, vp8_bilinear_filters_x86_4[8][8]);
+
+/* duplicated 8x */
+extern DECLARE_ALIGNED(16, const short, vp8_bilinear_filters_x86_8[8][16]);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_COMMON_X86_FILTER_X86_H_
diff --git a/libvpx/libvpx/vp8/common/x86/idct_blk_mmx.c b/libvpx/libvpx/vp8/common/x86/idct_blk_mmx.c
new file mode 100644
index 0000000..f2532b3
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/idct_blk_mmx.c
@@ -0,0 +1,128 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vp8/common/blockd.h"
+#include "vpx_mem/vpx_mem.h"
+
+extern void vp8_dequantize_b_impl_mmx(short *sq, short *dq, short *q);
+
+void vp8_dequantize_b_mmx(BLOCKD *d, short *DQC)
+{
+    short *sq = (short *) d->qcoeff;
+    short *dq = (short *) d->dqcoeff;
+
+    vp8_dequantize_b_impl_mmx(sq, dq, DQC);
+}
+
+void vp8_dequant_idct_add_y_block_mmx
+            (short *q, short *dq,
+             unsigned char *dst, int stride, char *eobs)
+{
+    int i;
+
+    for (i = 0; i < 4; i++)
+    {
+        if (eobs[0] > 1)
+            vp8_dequant_idct_add_mmx (q, dq, dst, stride);
+        else if (eobs[0] == 1)
+        {
+            vp8_dc_only_idct_add_mmx (q[0]*dq[0], dst, stride, dst, stride);
+            memset(q, 0, 2 * sizeof(q[0]));
+        }
+
+        if (eobs[1] > 1)
+            vp8_dequant_idct_add_mmx (q+16, dq, dst+4, stride);
+        else if (eobs[1] == 1)
+        {
+            vp8_dc_only_idct_add_mmx (q[16]*dq[0], dst+4, stride,
+                                      dst+4, stride);
+            memset(q + 16, 0, 2 * sizeof(q[0]));
+        }
+
+        if (eobs[2] > 1)
+            vp8_dequant_idct_add_mmx (q+32, dq, dst+8, stride);
+        else if (eobs[2] == 1)
+        {
+            vp8_dc_only_idct_add_mmx (q[32]*dq[0], dst+8, stride,
+                                      dst+8, stride);
+            memset(q + 32, 0, 2 * sizeof(q[0]));
+        }
+
+        if (eobs[3] > 1)
+            vp8_dequant_idct_add_mmx (q+48, dq, dst+12, stride);
+        else if (eobs[3] == 1)
+        {
+            vp8_dc_only_idct_add_mmx (q[48]*dq[0], dst+12, stride,
+                                      dst+12, stride);
+            memset(q + 48, 0, 2 * sizeof(q[0]));
+        }
+
+        q    += 64;
+        dst  += 4*stride;
+        eobs += 4;
+    }
+}
+
+void vp8_dequant_idct_add_uv_block_mmx
+            (short *q, short *dq,
+             unsigned char *dstu, unsigned char *dstv, int stride, char *eobs)
+{
+    int i;
+
+    for (i = 0; i < 2; i++)
+    {
+        if (eobs[0] > 1)
+            vp8_dequant_idct_add_mmx (q, dq, dstu, stride);
+        else if (eobs[0] == 1)
+        {
+            vp8_dc_only_idct_add_mmx (q[0]*dq[0], dstu, stride, dstu, stride);
+            memset(q, 0, 2 * sizeof(q[0]));
+        }
+
+        if (eobs[1] > 1)
+            vp8_dequant_idct_add_mmx (q+16, dq, dstu+4, stride);
+        else if (eobs[1] == 1)
+        {
+            vp8_dc_only_idct_add_mmx (q[16]*dq[0], dstu+4, stride,
+                                      dstu+4, stride);
+            memset(q + 16, 0, 2 * sizeof(q[0]));
+        }
+
+        q    += 32;
+        dstu += 4*stride;
+        eobs += 2;
+    }
+
+    for (i = 0; i < 2; i++)
+    {
+        if (eobs[0] > 1)
+            vp8_dequant_idct_add_mmx (q, dq, dstv, stride);
+        else if (eobs[0] == 1)
+        {
+            vp8_dc_only_idct_add_mmx (q[0]*dq[0], dstv, stride, dstv, stride);
+            memset(q, 0, 2 * sizeof(q[0]));
+        }
+
+        if (eobs[1] > 1)
+            vp8_dequant_idct_add_mmx (q+16, dq, dstv+4, stride);
+        else if (eobs[1] == 1)
+        {
+            vp8_dc_only_idct_add_mmx (q[16]*dq[0], dstv+4, stride,
+                                      dstv+4, stride);
+            memset(q + 16, 0, 2 * sizeof(q[0]));
+        }
+
+        q    += 32;
+        dstv += 4*stride;
+        eobs += 2;
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/x86/idct_blk_sse2.c b/libvpx/libvpx/vp8/common/x86/idct_blk_sse2.c
new file mode 100644
index 0000000..ae96ec8
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/idct_blk_sse2.c
@@ -0,0 +1,89 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+
+void vp8_idct_dequant_0_2x_sse2
+            (short *q, short *dq ,
+             unsigned char *dst, int dst_stride);
+void vp8_idct_dequant_full_2x_sse2
+            (short *q, short *dq ,
+             unsigned char *dst, int dst_stride);
+
+void vp8_dequant_idct_add_y_block_sse2
+            (short *q, short *dq,
+             unsigned char *dst, int stride, char *eobs)
+{
+    int i;
+
+    for (i = 0; i < 4; i++)
+    {
+        if (((short *)(eobs))[0])
+        {
+            if (((short *)(eobs))[0] & 0xfefe)
+                vp8_idct_dequant_full_2x_sse2 (q, dq, dst, stride);
+            else
+                vp8_idct_dequant_0_2x_sse2 (q, dq, dst, stride);
+        }
+        if (((short *)(eobs))[1])
+        {
+            if (((short *)(eobs))[1] & 0xfefe)
+                vp8_idct_dequant_full_2x_sse2 (q+32, dq, dst+8, stride);
+            else
+                vp8_idct_dequant_0_2x_sse2 (q+32, dq, dst+8, stride);
+        }
+        q    += 64;
+        dst  += stride*4;
+        eobs += 4;
+    }
+}
+
+void vp8_dequant_idct_add_uv_block_sse2
+            (short *q, short *dq,
+             unsigned char *dstu, unsigned char *dstv, int stride, char *eobs)
+{
+    if (((short *)(eobs))[0])
+    {
+        if (((short *)(eobs))[0] & 0xfefe)
+            vp8_idct_dequant_full_2x_sse2 (q, dq, dstu, stride);
+        else
+            vp8_idct_dequant_0_2x_sse2 (q, dq, dstu, stride);
+    }
+    q    += 32;
+    dstu += stride*4;
+
+    if (((short *)(eobs))[1])
+    {
+        if (((short *)(eobs))[1] & 0xfefe)
+            vp8_idct_dequant_full_2x_sse2 (q, dq, dstu, stride);
+        else
+            vp8_idct_dequant_0_2x_sse2 (q, dq, dstu, stride);
+    }
+    q    += 32;
+
+    if (((short *)(eobs))[2])
+    {
+        if (((short *)(eobs))[2] & 0xfefe)
+            vp8_idct_dequant_full_2x_sse2 (q, dq, dstv, stride);
+        else
+            vp8_idct_dequant_0_2x_sse2 (q, dq, dstv, stride);
+    }
+    q    += 32;
+    dstv += stride*4;
+
+    if (((short *)(eobs))[3])
+    {
+      if (((short *)(eobs))[3] & 0xfefe)
+          vp8_idct_dequant_full_2x_sse2 (q, dq, dstv, stride);
+      else
+          vp8_idct_dequant_0_2x_sse2 (q, dq, dstv, stride);
+    }
+}
diff --git a/libvpx/libvpx/vp8/common/x86/idctllm_mmx.asm b/libvpx/libvpx/vp8/common/x86/idctllm_mmx.asm
new file mode 100644
index 0000000..96fa2c6
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/idctllm_mmx.asm
@@ -0,0 +1,295 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+; /****************************************************************************
+; * Notes:
+; *
+; * This implementation makes use of 16 bit fixed point version of two multiply
+; * constants:
+; *        1.   sqrt(2) * cos (pi/8)
+; *        2.   sqrt(2) * sin (pi/8)
+; * Because the first constant is bigger than 1, to maintain the same 16 bit
+; * fixed point precision as the second one, we use a trick of
+; *        x * a = x + x*(a-1)
+; * so
+; *        x * sqrt(2) * cos (pi/8) = x + x * (sqrt(2) *cos(pi/8)-1).
+; *
+; * For the second constant, because of the 16bit version is 35468, which
+; * is bigger than 32768, in signed 16 bit multiply, it becomes a negative
+; * number.
+; *        (x * (unsigned)35468 >> 16) = x * (signed)35468 >> 16 + x
+; *
+; **************************************************************************/
+
+
+;void vp8_short_idct4x4llm_mmx(short *input, unsigned char *pred,
+;int pitch, unsigned char *dest,int stride)
+global sym(vp8_short_idct4x4llm_mmx) PRIVATE
+sym(vp8_short_idct4x4llm_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov         rax,    arg(0)              ;input
+    mov         rsi,    arg(1)              ;pred
+
+    movq        mm0,    [rax   ]
+    movq        mm1,    [rax+ 8]
+    movq        mm2,    [rax+16]
+    movq        mm3,    [rax+24]
+
+%if 0
+    pxor        mm7,    mm7
+    movq        [rax],   mm7
+    movq        [rax+8], mm7
+    movq        [rax+16],mm7
+    movq        [rax+24],mm7
+%endif
+    movsxd      rax,    dword ptr arg(2)    ;pitch
+    mov         rdx,    arg(3)              ;dest
+    movsxd      rdi,    dword ptr arg(4)    ;stride
+
+
+    psubw       mm0,            mm2             ; b1= 0-2
+    paddw       mm2,            mm2             ;
+
+    movq        mm5,            mm1
+    paddw       mm2,            mm0             ; a1 =0+2
+
+    pmulhw      mm5,            [GLOBAL(x_s1sqr2)];
+    paddw       mm5,            mm1             ; ip1 * sin(pi/8) * sqrt(2)
+
+    movq        mm7,            mm3             ;
+    pmulhw      mm7,            [GLOBAL(x_c1sqr2less1)];
+
+    paddw       mm7,            mm3             ; ip3 * cos(pi/8) * sqrt(2)
+    psubw       mm7,            mm5             ; c1
+
+    movq        mm5,            mm1
+    movq        mm4,            mm3
+
+    pmulhw      mm5,            [GLOBAL(x_c1sqr2less1)]
+    paddw       mm5,            mm1
+
+    pmulhw      mm3,            [GLOBAL(x_s1sqr2)]
+    paddw       mm3,            mm4
+
+    paddw       mm3,            mm5             ; d1
+    movq        mm6,            mm2             ; a1
+
+    movq        mm4,            mm0             ; b1
+    paddw       mm2,            mm3             ;0
+
+    paddw       mm4,            mm7             ;1
+    psubw       mm0,            mm7             ;2
+
+    psubw       mm6,            mm3             ;3
+
+    movq        mm1,            mm2             ; 03 02 01 00
+    movq        mm3,            mm4             ; 23 22 21 20
+
+    punpcklwd   mm1,            mm0             ; 11 01 10 00
+    punpckhwd   mm2,            mm0             ; 13 03 12 02
+
+    punpcklwd   mm3,            mm6             ; 31 21 30 20
+    punpckhwd   mm4,            mm6             ; 33 23 32 22
+
+    movq        mm0,            mm1             ; 11 01 10 00
+    movq        mm5,            mm2             ; 13 03 12 02
+
+    punpckldq   mm0,            mm3             ; 30 20 10 00
+    punpckhdq   mm1,            mm3             ; 31 21 11 01
+
+    punpckldq   mm2,            mm4             ; 32 22 12 02
+    punpckhdq   mm5,            mm4             ; 33 23 13 03
+
+    movq        mm3,            mm5             ; 33 23 13 03
+
+    psubw       mm0,            mm2             ; b1= 0-2
+    paddw       mm2,            mm2             ;
+
+    movq        mm5,            mm1
+    paddw       mm2,            mm0             ; a1 =0+2
+
+    pmulhw      mm5,            [GLOBAL(x_s1sqr2)];
+    paddw       mm5,            mm1             ; ip1 * sin(pi/8) * sqrt(2)
+
+    movq        mm7,            mm3             ;
+    pmulhw      mm7,            [GLOBAL(x_c1sqr2less1)];
+
+    paddw       mm7,            mm3             ; ip3 * cos(pi/8) * sqrt(2)
+    psubw       mm7,            mm5             ; c1
+
+    movq        mm5,            mm1
+    movq        mm4,            mm3
+
+    pmulhw      mm5,            [GLOBAL(x_c1sqr2less1)]
+    paddw       mm5,            mm1
+
+    pmulhw      mm3,            [GLOBAL(x_s1sqr2)]
+    paddw       mm3,            mm4
+
+    paddw       mm3,            mm5             ; d1
+    paddw       mm0,            [GLOBAL(fours)]
+
+    paddw       mm2,            [GLOBAL(fours)]
+    movq        mm6,            mm2             ; a1
+
+    movq        mm4,            mm0             ; b1
+    paddw       mm2,            mm3             ;0
+
+    paddw       mm4,            mm7             ;1
+    psubw       mm0,            mm7             ;2
+
+    psubw       mm6,            mm3             ;3
+    psraw       mm2,            3
+
+    psraw       mm0,            3
+    psraw       mm4,            3
+
+    psraw       mm6,            3
+
+    movq        mm1,            mm2             ; 03 02 01 00
+    movq        mm3,            mm4             ; 23 22 21 20
+
+    punpcklwd   mm1,            mm0             ; 11 01 10 00
+    punpckhwd   mm2,            mm0             ; 13 03 12 02
+
+    punpcklwd   mm3,            mm6             ; 31 21 30 20
+    punpckhwd   mm4,            mm6             ; 33 23 32 22
+
+    movq        mm0,            mm1             ; 11 01 10 00
+    movq        mm5,            mm2             ; 13 03 12 02
+
+    punpckldq   mm0,            mm3             ; 30 20 10 00
+    punpckhdq   mm1,            mm3             ; 31 21 11 01
+
+    punpckldq   mm2,            mm4             ; 32 22 12 02
+    punpckhdq   mm5,            mm4             ; 33 23 13 03
+
+    pxor        mm7,            mm7
+
+    movd        mm4,            [rsi]
+    punpcklbw   mm4,            mm7
+    paddsw      mm0,            mm4
+    packuswb    mm0,            mm7
+    movd        [rdx],          mm0
+
+    movd        mm4,            [rsi+rax]
+    punpcklbw   mm4,            mm7
+    paddsw      mm1,            mm4
+    packuswb    mm1,            mm7
+    movd        [rdx+rdi],      mm1
+
+    movd        mm4,            [rsi+2*rax]
+    punpcklbw   mm4,            mm7
+    paddsw      mm2,            mm4
+    packuswb    mm2,            mm7
+    movd        [rdx+rdi*2],    mm2
+
+    add         rdx,            rdi
+    add         rsi,            rax
+
+    movd        mm4,            [rsi+2*rax]
+    punpcklbw   mm4,            mm7
+    paddsw      mm5,            mm4
+    packuswb    mm5,            mm7
+    movd        [rdx+rdi*2],    mm5
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp8_dc_only_idct_add_mmx(
+;short input_dc,
+;unsigned char *pred_ptr,
+;int pred_stride,
+;unsigned char *dst_ptr,
+;int stride)
+global sym(vp8_dc_only_idct_add_mmx) PRIVATE
+sym(vp8_dc_only_idct_add_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    GET_GOT     rbx
+    ; end prolog
+
+        movd        mm5,            arg(0) ;input_dc
+        mov         rax,            arg(1) ;pred_ptr
+        movsxd      rdx,            dword ptr arg(2) ;pred_stride
+
+        pxor        mm0,            mm0
+
+        paddw       mm5,            [GLOBAL(fours)]
+        lea         rcx,            [rdx + rdx*2]
+
+        psraw       mm5,            3
+
+        punpcklwd   mm5,            mm5
+
+        punpckldq   mm5,            mm5
+
+        movd        mm1,            [rax]
+        movd        mm2,            [rax+rdx]
+        movd        mm3,            [rax+2*rdx]
+        movd        mm4,            [rax+rcx]
+
+        mov         rax,            arg(3) ;d -- destination
+        movsxd      rdx,            dword ptr arg(4) ;dst_stride
+
+        punpcklbw   mm1,            mm0
+        paddsw      mm1,            mm5
+        packuswb    mm1,            mm0              ; pack and unpack to saturate
+        lea         rcx,            [rdx + rdx*2]
+
+        punpcklbw   mm2,            mm0
+        paddsw      mm2,            mm5
+        packuswb    mm2,            mm0              ; pack and unpack to saturate
+
+        punpcklbw   mm3,            mm0
+        paddsw      mm3,            mm5
+        packuswb    mm3,            mm0              ; pack and unpack to saturate
+
+        punpcklbw   mm4,            mm0
+        paddsw      mm4,            mm5
+        packuswb    mm4,            mm0              ; pack and unpack to saturate
+
+        movd        [rax],          mm1
+        movd        [rax+rdx],      mm2
+        movd        [rax+2*rdx],    mm3
+        movd        [rax+rcx],      mm4
+
+    ; begin epilog
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+x_s1sqr2:
+    times 4 dw 0x8A8C
+align 16
+x_c1sqr2less1:
+    times 4 dw 0x4E7B
+align 16
+fours:
+    times 4 dw 0x0004
diff --git a/libvpx/libvpx/vp8/common/x86/idctllm_sse2.asm b/libvpx/libvpx/vp8/common/x86/idctllm_sse2.asm
new file mode 100644
index 0000000..bf8e2c4
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/idctllm_sse2.asm
@@ -0,0 +1,708 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vp8_idct_dequant_0_2x_sse2
+; (
+;   short *qcoeff       - 0
+;   short *dequant      - 1
+;   unsigned char *dst  - 2
+;   int dst_stride      - 3
+; )
+
+global sym(vp8_idct_dequant_0_2x_sse2) PRIVATE
+sym(vp8_idct_dequant_0_2x_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 4
+    GET_GOT     rbx
+    ; end prolog
+
+        mov         rdx,            arg(1) ; dequant
+        mov         rax,            arg(0) ; qcoeff
+
+        movd        xmm4,           [rax]
+        movd        xmm5,           [rdx]
+
+        pinsrw      xmm4,           [rax+32],   4
+        pinsrw      xmm5,           [rdx],      4
+
+        pmullw      xmm4,           xmm5
+
+    ; Zero out xmm5, for use unpacking
+        pxor        xmm5,           xmm5
+
+    ; clear coeffs
+        movd        [rax],          xmm5
+        movd        [rax+32],       xmm5
+;pshufb
+        mov         rax,            arg(2) ; dst
+        movsxd      rdx,            dword ptr arg(3) ; dst_stride
+
+        pshuflw     xmm4,           xmm4,       00000000b
+        pshufhw     xmm4,           xmm4,       00000000b
+
+        lea         rcx,            [rdx + rdx*2]
+        paddw       xmm4,           [GLOBAL(fours)]
+
+        psraw       xmm4,           3
+
+        movq        xmm0,           [rax]
+        movq        xmm1,           [rax+rdx]
+        movq        xmm2,           [rax+2*rdx]
+        movq        xmm3,           [rax+rcx]
+
+        punpcklbw   xmm0,           xmm5
+        punpcklbw   xmm1,           xmm5
+        punpcklbw   xmm2,           xmm5
+        punpcklbw   xmm3,           xmm5
+
+
+    ; Add to predict buffer
+        paddw       xmm0,           xmm4
+        paddw       xmm1,           xmm4
+        paddw       xmm2,           xmm4
+        paddw       xmm3,           xmm4
+
+    ; pack up before storing
+        packuswb    xmm0,           xmm5
+        packuswb    xmm1,           xmm5
+        packuswb    xmm2,           xmm5
+        packuswb    xmm3,           xmm5
+
+    ; store blocks back out
+        movq        [rax],          xmm0
+        movq        [rax + rdx],    xmm1
+
+        lea         rax,            [rax + 2*rdx]
+
+        movq        [rax],          xmm2
+        movq        [rax + rdx],    xmm3
+
+    ; begin epilog
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp8_idct_dequant_full_2x_sse2
+; (
+;   short *qcoeff       - 0
+;   short *dequant      - 1
+;   unsigned char *dst  - 2
+;   int dst_stride      - 3
+; )
+global sym(vp8_idct_dequant_full_2x_sse2) PRIVATE
+sym(vp8_idct_dequant_full_2x_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 4
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ; special case when 2 blocks have 0 or 1 coeffs
+    ; dc is set as first coeff, so no need to load qcoeff
+        mov         rax,            arg(0) ; qcoeff
+        mov         rdx,            arg(1)  ; dequant
+        mov         rdi,            arg(2) ; dst
+
+
+    ; Zero out xmm7, for use unpacking
+        pxor        xmm7,           xmm7
+
+
+    ; note the transpose of xmm1 and xmm2, necessary for shuffle
+    ;   to spit out sensicle data
+        movdqa      xmm0,           [rax]
+        movdqa      xmm2,           [rax+16]
+        movdqa      xmm1,           [rax+32]
+        movdqa      xmm3,           [rax+48]
+
+    ; Clear out coeffs
+        movdqa      [rax],          xmm7
+        movdqa      [rax+16],       xmm7
+        movdqa      [rax+32],       xmm7
+        movdqa      [rax+48],       xmm7
+
+    ; dequantize qcoeff buffer
+        pmullw      xmm0,           [rdx]
+        pmullw      xmm2,           [rdx+16]
+        pmullw      xmm1,           [rdx]
+        pmullw      xmm3,           [rdx+16]
+        movsxd      rdx,            dword ptr arg(3) ; dst_stride
+
+    ; repack so block 0 row x and block 1 row x are together
+        movdqa      xmm4,           xmm0
+        punpckldq   xmm0,           xmm1
+        punpckhdq   xmm4,           xmm1
+
+        pshufd      xmm0,           xmm0,       11011000b
+        pshufd      xmm1,           xmm4,       11011000b
+
+        movdqa      xmm4,           xmm2
+        punpckldq   xmm2,           xmm3
+        punpckhdq   xmm4,           xmm3
+
+        pshufd      xmm2,           xmm2,       11011000b
+        pshufd      xmm3,           xmm4,       11011000b
+
+    ; first pass
+        psubw       xmm0,           xmm2        ; b1 = 0-2
+        paddw       xmm2,           xmm2        ;
+
+        movdqa      xmm5,           xmm1
+        paddw       xmm2,           xmm0        ; a1 = 0+2
+
+        pmulhw      xmm5,           [GLOBAL(x_s1sqr2)]
+        lea         rcx,            [rdx + rdx*2]   ;dst_stride * 3
+        paddw       xmm5,           xmm1        ; ip1 * sin(pi/8) * sqrt(2)
+
+        movdqa      xmm7,           xmm3
+        pmulhw      xmm7,           [GLOBAL(x_c1sqr2less1)]
+
+        paddw       xmm7,           xmm3        ; ip3 * cos(pi/8) * sqrt(2)
+        psubw       xmm7,           xmm5        ; c1
+
+        movdqa      xmm5,           xmm1
+        movdqa      xmm4,           xmm3
+
+        pmulhw      xmm5,           [GLOBAL(x_c1sqr2less1)]
+        paddw       xmm5,           xmm1
+
+        pmulhw      xmm3,           [GLOBAL(x_s1sqr2)]
+        paddw       xmm3,           xmm4
+
+        paddw       xmm3,           xmm5        ; d1
+        movdqa      xmm6,           xmm2        ; a1
+
+        movdqa      xmm4,           xmm0        ; b1
+        paddw       xmm2,           xmm3        ;0
+
+        paddw       xmm4,           xmm7        ;1
+        psubw       xmm0,           xmm7        ;2
+
+        psubw       xmm6,           xmm3        ;3
+
+    ; transpose for the second pass
+        movdqa      xmm7,           xmm2        ; 103 102 101 100 003 002 001 000
+        punpcklwd   xmm2,           xmm0        ; 007 003 006 002 005 001 004 000
+        punpckhwd   xmm7,           xmm0        ; 107 103 106 102 105 101 104 100
+
+        movdqa      xmm5,           xmm4        ; 111 110 109 108 011 010 009 008
+        punpcklwd   xmm4,           xmm6        ; 015 011 014 010 013 009 012 008
+        punpckhwd   xmm5,           xmm6        ; 115 111 114 110 113 109 112 108
+
+
+        movdqa      xmm1,           xmm2        ; 007 003 006 002 005 001 004 000
+        punpckldq   xmm2,           xmm4        ; 013 009 005 001 012 008 004 000
+        punpckhdq   xmm1,           xmm4        ; 015 011 007 003 014 010 006 002
+
+        movdqa      xmm6,           xmm7        ; 107 103 106 102 105 101 104 100
+        punpckldq   xmm7,           xmm5        ; 113 109 105 101 112 108 104 100
+        punpckhdq   xmm6,           xmm5        ; 115 111 107 103 114 110 106 102
+
+
+        movdqa      xmm5,           xmm2        ; 013 009 005 001 012 008 004 000
+        punpckldq   xmm2,           xmm7        ; 112 108 012 008 104 100 004 000
+        punpckhdq   xmm5,           xmm7        ; 113 109 013 009 105 101 005 001
+
+        movdqa      xmm7,           xmm1        ; 015 011 007 003 014 010 006 002
+        punpckldq   xmm1,           xmm6        ; 114 110 014 010 106 102 006 002
+        punpckhdq   xmm7,           xmm6        ; 115 111 015 011 107 103 007 003
+
+        pshufd      xmm0,           xmm2,       11011000b
+        pshufd      xmm2,           xmm1,       11011000b
+
+        pshufd      xmm1,           xmm5,       11011000b
+        pshufd      xmm3,           xmm7,       11011000b
+
+    ; second pass
+        psubw       xmm0,           xmm2            ; b1 = 0-2
+        paddw       xmm2,           xmm2
+
+        movdqa      xmm5,           xmm1
+        paddw       xmm2,           xmm0            ; a1 = 0+2
+
+        pmulhw      xmm5,           [GLOBAL(x_s1sqr2)]
+        paddw       xmm5,           xmm1            ; ip1 * sin(pi/8) * sqrt(2)
+
+        movdqa      xmm7,           xmm3
+        pmulhw      xmm7,           [GLOBAL(x_c1sqr2less1)]
+
+        paddw       xmm7,           xmm3            ; ip3 * cos(pi/8) * sqrt(2)
+        psubw       xmm7,           xmm5            ; c1
+
+        movdqa      xmm5,           xmm1
+        movdqa      xmm4,           xmm3
+
+        pmulhw      xmm5,           [GLOBAL(x_c1sqr2less1)]
+        paddw       xmm5,           xmm1
+
+        pmulhw      xmm3,           [GLOBAL(x_s1sqr2)]
+        paddw       xmm3,           xmm4
+
+        paddw       xmm3,           xmm5            ; d1
+        paddw       xmm0,           [GLOBAL(fours)]
+
+        paddw       xmm2,           [GLOBAL(fours)]
+        movdqa      xmm6,           xmm2            ; a1
+
+        movdqa      xmm4,           xmm0            ; b1
+        paddw       xmm2,           xmm3            ;0
+
+        paddw       xmm4,           xmm7            ;1
+        psubw       xmm0,           xmm7            ;2
+
+        psubw       xmm6,           xmm3            ;3
+        psraw       xmm2,           3
+
+        psraw       xmm0,           3
+        psraw       xmm4,           3
+
+        psraw       xmm6,           3
+
+    ; transpose to save
+        movdqa      xmm7,           xmm2        ; 103 102 101 100 003 002 001 000
+        punpcklwd   xmm2,           xmm0        ; 007 003 006 002 005 001 004 000
+        punpckhwd   xmm7,           xmm0        ; 107 103 106 102 105 101 104 100
+
+        movdqa      xmm5,           xmm4        ; 111 110 109 108 011 010 009 008
+        punpcklwd   xmm4,           xmm6        ; 015 011 014 010 013 009 012 008
+        punpckhwd   xmm5,           xmm6        ; 115 111 114 110 113 109 112 108
+
+
+        movdqa      xmm1,           xmm2        ; 007 003 006 002 005 001 004 000
+        punpckldq   xmm2,           xmm4        ; 013 009 005 001 012 008 004 000
+        punpckhdq   xmm1,           xmm4        ; 015 011 007 003 014 010 006 002
+
+        movdqa      xmm6,           xmm7        ; 107 103 106 102 105 101 104 100
+        punpckldq   xmm7,           xmm5        ; 113 109 105 101 112 108 104 100
+        punpckhdq   xmm6,           xmm5        ; 115 111 107 103 114 110 106 102
+
+
+        movdqa      xmm5,           xmm2        ; 013 009 005 001 012 008 004 000
+        punpckldq   xmm2,           xmm7        ; 112 108 012 008 104 100 004 000
+        punpckhdq   xmm5,           xmm7        ; 113 109 013 009 105 101 005 001
+
+        movdqa      xmm7,           xmm1        ; 015 011 007 003 014 010 006 002
+        punpckldq   xmm1,           xmm6        ; 114 110 014 010 106 102 006 002
+        punpckhdq   xmm7,           xmm6        ; 115 111 015 011 107 103 007 003
+
+        pshufd      xmm0,           xmm2,       11011000b
+        pshufd      xmm2,           xmm1,       11011000b
+
+        pshufd      xmm1,           xmm5,       11011000b
+        pshufd      xmm3,           xmm7,       11011000b
+
+        pxor        xmm7,           xmm7
+
+    ; Load up predict blocks
+        movq        xmm4,           [rdi]
+        movq        xmm5,           [rdi+rdx]
+
+        punpcklbw   xmm4,           xmm7
+        punpcklbw   xmm5,           xmm7
+
+        paddw       xmm0,           xmm4
+        paddw       xmm1,           xmm5
+
+        movq        xmm4,           [rdi+2*rdx]
+        movq        xmm5,           [rdi+rcx]
+
+        punpcklbw   xmm4,           xmm7
+        punpcklbw   xmm5,           xmm7
+
+        paddw       xmm2,           xmm4
+        paddw       xmm3,           xmm5
+
+.finish:
+
+    ; pack up before storing
+        packuswb    xmm0,           xmm7
+        packuswb    xmm1,           xmm7
+        packuswb    xmm2,           xmm7
+        packuswb    xmm3,           xmm7
+
+    ; store blocks back out
+        movq        [rdi],          xmm0
+        movq        [rdi + rdx],    xmm1
+        movq        [rdi + rdx*2],  xmm2
+        movq        [rdi + rcx],    xmm3
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp8_idct_dequant_dc_0_2x_sse2
+; (
+;   short *qcoeff       - 0
+;   short *dequant      - 1
+;   unsigned char *dst  - 2
+;   int dst_stride      - 3
+;   short *dc           - 4
+; )
+global sym(vp8_idct_dequant_dc_0_2x_sse2) PRIVATE
+sym(vp8_idct_dequant_dc_0_2x_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    GET_GOT     rbx
+    push        rdi
+    ; end prolog
+
+    ; special case when 2 blocks have 0 or 1 coeffs
+    ; dc is set as first coeff, so no need to load qcoeff
+        mov         rax,            arg(0) ; qcoeff
+
+        mov         rdi,            arg(2) ; dst
+        mov         rdx,            arg(4) ; dc
+
+    ; Zero out xmm5, for use unpacking
+        pxor        xmm5,           xmm5
+
+    ; load up 2 dc words here == 2*16 = doubleword
+        movd        xmm4,           [rdx]
+
+        movsxd      rdx,            dword ptr arg(3) ; dst_stride
+        lea         rcx, [rdx + rdx*2]
+    ; Load up predict blocks
+        movq        xmm0,           [rdi]
+        movq        xmm1,           [rdi+rdx*1]
+        movq        xmm2,           [rdi+rdx*2]
+        movq        xmm3,           [rdi+rcx]
+
+    ; Duplicate and expand dc across
+        punpcklwd   xmm4,           xmm4
+        punpckldq   xmm4,           xmm4
+
+    ; Rounding to dequant and downshift
+        paddw       xmm4,           [GLOBAL(fours)]
+        psraw       xmm4,           3
+
+    ; Predict buffer needs to be expanded from bytes to words
+        punpcklbw   xmm0,           xmm5
+        punpcklbw   xmm1,           xmm5
+        punpcklbw   xmm2,           xmm5
+        punpcklbw   xmm3,           xmm5
+
+    ; Add to predict buffer
+        paddw       xmm0,           xmm4
+        paddw       xmm1,           xmm4
+        paddw       xmm2,           xmm4
+        paddw       xmm3,           xmm4
+
+    ; pack up before storing
+        packuswb    xmm0,           xmm5
+        packuswb    xmm1,           xmm5
+        packuswb    xmm2,           xmm5
+        packuswb    xmm3,           xmm5
+
+    ; store blocks back out
+        movq        [rdi],          xmm0
+        movq        [rdi + rdx],    xmm1
+        movq        [rdi + rdx*2],  xmm2
+        movq        [rdi + rcx],    xmm3
+
+    ; begin epilog
+    pop         rdi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+;void vp8_idct_dequant_dc_full_2x_sse2
+; (
+;   short *qcoeff       - 0
+;   short *dequant      - 1
+;   unsigned char *dst  - 2
+;   int dst_stride      - 3
+;   short *dc           - 4
+; )
+global sym(vp8_idct_dequant_dc_full_2x_sse2) PRIVATE
+sym(vp8_idct_dequant_dc_full_2x_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rdi
+    ; end prolog
+
+    ; special case when 2 blocks have 0 or 1 coeffs
+    ; dc is set as first coeff, so no need to load qcoeff
+        mov         rax,            arg(0) ; qcoeff
+        mov         rdx,            arg(1)  ; dequant
+
+        mov         rdi,            arg(2) ; dst
+
+    ; Zero out xmm7, for use unpacking
+        pxor        xmm7,           xmm7
+
+
+    ; note the transpose of xmm1 and xmm2, necessary for shuffle
+    ;   to spit out sensicle data
+        movdqa      xmm0,           [rax]
+        movdqa      xmm2,           [rax+16]
+        movdqa      xmm1,           [rax+32]
+        movdqa      xmm3,           [rax+48]
+
+    ; Clear out coeffs
+        movdqa      [rax],          xmm7
+        movdqa      [rax+16],       xmm7
+        movdqa      [rax+32],       xmm7
+        movdqa      [rax+48],       xmm7
+
+    ; dequantize qcoeff buffer
+        pmullw      xmm0,           [rdx]
+        pmullw      xmm2,           [rdx+16]
+        pmullw      xmm1,           [rdx]
+        pmullw      xmm3,           [rdx+16]
+
+    ; DC component
+        mov         rdx,            arg(4)
+
+    ; repack so block 0 row x and block 1 row x are together
+        movdqa      xmm4,           xmm0
+        punpckldq   xmm0,           xmm1
+        punpckhdq   xmm4,           xmm1
+
+        pshufd      xmm0,           xmm0,       11011000b
+        pshufd      xmm1,           xmm4,       11011000b
+
+        movdqa      xmm4,           xmm2
+        punpckldq   xmm2,           xmm3
+        punpckhdq   xmm4,           xmm3
+
+        pshufd      xmm2,           xmm2,       11011000b
+        pshufd      xmm3,           xmm4,       11011000b
+
+    ; insert DC component
+        pinsrw      xmm0,           [rdx],      0
+        pinsrw      xmm0,           [rdx+2],    4
+
+    ; first pass
+        psubw       xmm0,           xmm2        ; b1 = 0-2
+        paddw       xmm2,           xmm2        ;
+
+        movdqa      xmm5,           xmm1
+        paddw       xmm2,           xmm0        ; a1 = 0+2
+
+        pmulhw      xmm5,           [GLOBAL(x_s1sqr2)]
+        paddw       xmm5,           xmm1        ; ip1 * sin(pi/8) * sqrt(2)
+
+        movdqa      xmm7,           xmm3
+        pmulhw      xmm7,           [GLOBAL(x_c1sqr2less1)]
+
+        paddw       xmm7,           xmm3        ; ip3 * cos(pi/8) * sqrt(2)
+        psubw       xmm7,           xmm5        ; c1
+
+        movdqa      xmm5,           xmm1
+        movdqa      xmm4,           xmm3
+
+        pmulhw      xmm5,           [GLOBAL(x_c1sqr2less1)]
+        paddw       xmm5,           xmm1
+
+        pmulhw      xmm3,           [GLOBAL(x_s1sqr2)]
+        paddw       xmm3,           xmm4
+
+        paddw       xmm3,           xmm5        ; d1
+        movdqa      xmm6,           xmm2        ; a1
+
+        movdqa      xmm4,           xmm0        ; b1
+        paddw       xmm2,           xmm3        ;0
+
+        paddw       xmm4,           xmm7        ;1
+        psubw       xmm0,           xmm7        ;2
+
+        psubw       xmm6,           xmm3        ;3
+
+    ; transpose for the second pass
+        movdqa      xmm7,           xmm2        ; 103 102 101 100 003 002 001 000
+        punpcklwd   xmm2,           xmm0        ; 007 003 006 002 005 001 004 000
+        punpckhwd   xmm7,           xmm0        ; 107 103 106 102 105 101 104 100
+
+        movdqa      xmm5,           xmm4        ; 111 110 109 108 011 010 009 008
+        punpcklwd   xmm4,           xmm6        ; 015 011 014 010 013 009 012 008
+        punpckhwd   xmm5,           xmm6        ; 115 111 114 110 113 109 112 108
+
+
+        movdqa      xmm1,           xmm2        ; 007 003 006 002 005 001 004 000
+        punpckldq   xmm2,           xmm4        ; 013 009 005 001 012 008 004 000
+        punpckhdq   xmm1,           xmm4        ; 015 011 007 003 014 010 006 002
+
+        movdqa      xmm6,           xmm7        ; 107 103 106 102 105 101 104 100
+        punpckldq   xmm7,           xmm5        ; 113 109 105 101 112 108 104 100
+        punpckhdq   xmm6,           xmm5        ; 115 111 107 103 114 110 106 102
+
+
+        movdqa      xmm5,           xmm2        ; 013 009 005 001 012 008 004 000
+        punpckldq   xmm2,           xmm7        ; 112 108 012 008 104 100 004 000
+        punpckhdq   xmm5,           xmm7        ; 113 109 013 009 105 101 005 001
+
+        movdqa      xmm7,           xmm1        ; 015 011 007 003 014 010 006 002
+        punpckldq   xmm1,           xmm6        ; 114 110 014 010 106 102 006 002
+        punpckhdq   xmm7,           xmm6        ; 115 111 015 011 107 103 007 003
+
+        pshufd      xmm0,           xmm2,       11011000b
+        pshufd      xmm2,           xmm1,       11011000b
+
+        pshufd      xmm1,           xmm5,       11011000b
+        pshufd      xmm3,           xmm7,       11011000b
+
+    ; second pass
+        psubw       xmm0,           xmm2            ; b1 = 0-2
+        paddw       xmm2,           xmm2
+
+        movdqa      xmm5,           xmm1
+        paddw       xmm2,           xmm0            ; a1 = 0+2
+
+        pmulhw      xmm5,           [GLOBAL(x_s1sqr2)]
+        paddw       xmm5,           xmm1            ; ip1 * sin(pi/8) * sqrt(2)
+
+        movdqa      xmm7,           xmm3
+        pmulhw      xmm7,           [GLOBAL(x_c1sqr2less1)]
+
+        paddw       xmm7,           xmm3            ; ip3 * cos(pi/8) * sqrt(2)
+        psubw       xmm7,           xmm5            ; c1
+
+        movdqa      xmm5,           xmm1
+        movdqa      xmm4,           xmm3
+
+        pmulhw      xmm5,           [GLOBAL(x_c1sqr2less1)]
+        paddw       xmm5,           xmm1
+
+        pmulhw      xmm3,           [GLOBAL(x_s1sqr2)]
+        paddw       xmm3,           xmm4
+
+        paddw       xmm3,           xmm5            ; d1
+        paddw       xmm0,           [GLOBAL(fours)]
+
+        paddw       xmm2,           [GLOBAL(fours)]
+        movdqa      xmm6,           xmm2            ; a1
+
+        movdqa      xmm4,           xmm0            ; b1
+        paddw       xmm2,           xmm3            ;0
+
+        paddw       xmm4,           xmm7            ;1
+        psubw       xmm0,           xmm7            ;2
+
+        psubw       xmm6,           xmm3            ;3
+        psraw       xmm2,           3
+
+        psraw       xmm0,           3
+        psraw       xmm4,           3
+
+        psraw       xmm6,           3
+
+    ; transpose to save
+        movdqa      xmm7,           xmm2        ; 103 102 101 100 003 002 001 000
+        punpcklwd   xmm2,           xmm0        ; 007 003 006 002 005 001 004 000
+        punpckhwd   xmm7,           xmm0        ; 107 103 106 102 105 101 104 100
+
+        movdqa      xmm5,           xmm4        ; 111 110 109 108 011 010 009 008
+        punpcklwd   xmm4,           xmm6        ; 015 011 014 010 013 009 012 008
+        punpckhwd   xmm5,           xmm6        ; 115 111 114 110 113 109 112 108
+
+
+        movdqa      xmm1,           xmm2        ; 007 003 006 002 005 001 004 000
+        punpckldq   xmm2,           xmm4        ; 013 009 005 001 012 008 004 000
+        punpckhdq   xmm1,           xmm4        ; 015 011 007 003 014 010 006 002
+
+        movdqa      xmm6,           xmm7        ; 107 103 106 102 105 101 104 100
+        punpckldq   xmm7,           xmm5        ; 113 109 105 101 112 108 104 100
+        punpckhdq   xmm6,           xmm5        ; 115 111 107 103 114 110 106 102
+
+
+        movdqa      xmm5,           xmm2        ; 013 009 005 001 012 008 004 000
+        punpckldq   xmm2,           xmm7        ; 112 108 012 008 104 100 004 000
+        punpckhdq   xmm5,           xmm7        ; 113 109 013 009 105 101 005 001
+
+        movdqa      xmm7,           xmm1        ; 015 011 007 003 014 010 006 002
+        punpckldq   xmm1,           xmm6        ; 114 110 014 010 106 102 006 002
+        punpckhdq   xmm7,           xmm6        ; 115 111 015 011 107 103 007 003
+
+        pshufd      xmm0,           xmm2,       11011000b
+        pshufd      xmm2,           xmm1,       11011000b
+
+        pshufd      xmm1,           xmm5,       11011000b
+        pshufd      xmm3,           xmm7,       11011000b
+
+        pxor        xmm7,           xmm7
+
+    ; Load up predict blocks
+        movsxd      rdx,            dword ptr arg(3) ; dst_stride
+        movq        xmm4,           [rdi]
+        movq        xmm5,           [rdi+rdx]
+        lea         rcx,            [rdx + rdx*2]
+
+        punpcklbw   xmm4,           xmm7
+        punpcklbw   xmm5,           xmm7
+
+        paddw       xmm0,           xmm4
+        paddw       xmm1,           xmm5
+
+        movq        xmm4,           [rdi+rdx*2]
+        movq        xmm5,           [rdi+rcx]
+
+        punpcklbw   xmm4,           xmm7
+        punpcklbw   xmm5,           xmm7
+
+        paddw       xmm2,           xmm4
+        paddw       xmm3,           xmm5
+
+.finish:
+
+    ; pack up before storing
+        packuswb    xmm0,           xmm7
+        packuswb    xmm1,           xmm7
+        packuswb    xmm2,           xmm7
+        packuswb    xmm3,           xmm7
+
+    ; Load destination stride before writing out,
+    ;   doesn't need to persist
+        movsxd      rdx,            dword ptr arg(3) ; dst_stride
+
+    ; store blocks back out
+        movq        [rdi],          xmm0
+        movq        [rdi + rdx],    xmm1
+
+        lea         rdi,            [rdi + 2*rdx]
+
+        movq        [rdi],          xmm2
+        movq        [rdi + rdx],    xmm3
+
+
+    ; begin epilog
+    pop         rdi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+fours:
+    times 8 dw 0x0004
+align 16
+x_s1sqr2:
+    times 8 dw 0x8A8C
+align 16
+x_c1sqr2less1:
+    times 8 dw 0x4E7B
diff --git a/libvpx/libvpx/vp8/common/x86/iwalsh_mmx.asm b/libvpx/libvpx/vp8/common/x86/iwalsh_mmx.asm
new file mode 100644
index 0000000..158c3b7
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/iwalsh_mmx.asm
@@ -0,0 +1,140 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vp8_short_inv_walsh4x4_mmx(short *input, short *output)
+global sym(vp8_short_inv_walsh4x4_mmx) PRIVATE
+sym(vp8_short_inv_walsh4x4_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 2
+    ; end prolog
+
+    mov         rdx, arg(0)
+    mov         rax, 30003h
+
+    movq        mm0, [rdx + 0]    ;ip[0]
+    movq        mm1, [rdx + 8]    ;ip[4]
+    movq        mm7, rax
+
+    movq        mm2, [rdx + 16]   ;ip[8]
+    movq        mm3, [rdx + 24]   ;ip[12]
+    punpcklwd   mm7, mm7          ;0003000300030003h
+    mov         rdx, arg(1)
+
+    movq        mm4, mm0
+    movq        mm5, mm1
+
+    paddw       mm4, mm3          ;ip[0] + ip[12] aka al
+    paddw       mm5, mm2          ;ip[4] + ip[8] aka bl
+
+    movq        mm6, mm4          ;temp al
+    paddw       mm4, mm5          ;al + bl
+    psubw       mm6, mm5          ;al - bl
+
+    psubw       mm0, mm3          ;ip[0] - ip[12] aka d1
+    psubw       mm1, mm2          ;ip[4] - ip[8] aka c1
+
+    movq        mm5, mm0          ;temp dl
+    paddw       mm0, mm1          ;dl + cl
+    psubw       mm5, mm1          ;dl - cl
+
+    ; 03 02 01 00
+    ; 13 12 11 10
+    ; 23 22 21 20
+    ; 33 32 31 30
+
+    movq        mm3, mm4          ; 03 02 01 00
+    punpcklwd   mm4, mm0          ; 11 01 10 00
+    punpckhwd   mm3, mm0          ; 13 03 12 02
+
+    movq        mm1, mm6          ; 23 22 21 20
+    punpcklwd   mm6, mm5          ; 31 21 30 20
+    punpckhwd   mm1, mm5          ; 33 23 32 22
+
+    movq        mm0, mm4          ; 11 01 10 00
+    movq        mm2, mm3          ; 13 03 12 02
+
+    punpckldq   mm0, mm6          ; 30 20 10 00 aka ip[0]
+    punpckhdq   mm4, mm6          ; 31 21 11 01 aka ip[4]
+
+    punpckldq   mm2, mm1          ; 32 22 12 02 aka ip[8]
+    punpckhdq   mm3, mm1          ; 33 23 13 03 aka ip[12]
+;~~~~~~~~~~~~~~~~~~~~~
+    movq        mm1, mm0
+    movq        mm5, mm4
+    paddw       mm1, mm3          ;ip[0] + ip[12] aka al
+    paddw       mm5, mm2          ;ip[4] + ip[8] aka bl
+
+    movq        mm6, mm1          ;temp al
+    paddw       mm1, mm5          ;al + bl
+    psubw       mm6, mm5          ;al - bl
+    paddw       mm1, mm7
+    paddw       mm6, mm7
+    psraw       mm1, 3
+    psraw       mm6, 3
+
+    psubw       mm0, mm3          ;ip[0] - ip[12] aka d1
+    psubw       mm4, mm2          ;ip[4] - ip[8] aka c1
+
+    movq        mm5, mm0          ;temp dl
+    paddw       mm0, mm4          ;dl + cl
+    psubw       mm5, mm4          ;dl - cl
+    paddw       mm0, mm7
+    paddw       mm5, mm7
+    psraw       mm0, 3
+    psraw       mm5, 3
+;~~~~~~~~~~~~~~~~~~~~~
+
+    movd        eax, mm1
+    movd        ecx, mm0
+    psrlq       mm0, 32
+    psrlq       mm1, 32
+    mov         word ptr[rdx+32*0], ax
+    mov         word ptr[rdx+32*1], cx
+    shr         eax, 16
+    shr         ecx, 16
+    mov         word ptr[rdx+32*4], ax
+    mov         word ptr[rdx+32*5], cx
+    movd        eax, mm1
+    movd        ecx, mm0
+    mov         word ptr[rdx+32*8], ax
+    mov         word ptr[rdx+32*9], cx
+    shr         eax, 16
+    shr         ecx, 16
+    mov         word ptr[rdx+32*12], ax
+    mov         word ptr[rdx+32*13], cx
+
+    movd        eax, mm6
+    movd        ecx, mm5
+    psrlq       mm5, 32
+    psrlq       mm6, 32
+    mov         word ptr[rdx+32*2], ax
+    mov         word ptr[rdx+32*3], cx
+    shr         eax, 16
+    shr         ecx, 16
+    mov         word ptr[rdx+32*6], ax
+    mov         word ptr[rdx+32*7], cx
+    movd        eax, mm6
+    movd        ecx, mm5
+    mov         word ptr[rdx+32*10], ax
+    mov         word ptr[rdx+32*11], cx
+    shr         eax, 16
+    shr         ecx, 16
+    mov         word ptr[rdx+32*14], ax
+    mov         word ptr[rdx+32*15], cx
+
+    ; begin epilog
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
diff --git a/libvpx/libvpx/vp8/common/x86/iwalsh_sse2.asm b/libvpx/libvpx/vp8/common/x86/iwalsh_sse2.asm
new file mode 100644
index 0000000..06e86a8
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/iwalsh_sse2.asm
@@ -0,0 +1,121 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vp8_short_inv_walsh4x4_sse2(short *input, short *output)
+global sym(vp8_short_inv_walsh4x4_sse2) PRIVATE
+sym(vp8_short_inv_walsh4x4_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 2
+    ; end prolog
+
+    mov         rcx, arg(0)
+    mov         rdx, arg(1)
+    mov         rax, 30003h
+
+    movdqa      xmm0, [rcx + 0]     ;ip[4] ip[0]
+    movdqa      xmm1, [rcx + 16]    ;ip[12] ip[8]
+
+
+    pshufd      xmm2, xmm1, 4eh     ;ip[8] ip[12]
+    movdqa      xmm3, xmm0          ;ip[4] ip[0]
+
+    paddw       xmm0, xmm2          ;ip[4]+ip[8] ip[0]+ip[12] aka b1 a1
+    psubw       xmm3, xmm2          ;ip[4]-ip[8] ip[0]-ip[12] aka c1 d1
+
+    movdqa      xmm4, xmm0
+    punpcklqdq  xmm0, xmm3          ;d1 a1
+    punpckhqdq  xmm4, xmm3          ;c1 b1
+
+    movdqa      xmm1, xmm4          ;c1 b1
+    paddw       xmm4, xmm0          ;dl+cl a1+b1 aka op[4] op[0]
+    psubw       xmm0, xmm1          ;d1-c1 a1-b1 aka op[12] op[8]
+
+    ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    ; 13 12 11 10 03 02 01 00
+    ;
+    ; 33 32 31 30 23 22 21 20
+    ;
+    movdqa      xmm3, xmm4          ; 13 12 11 10 03 02 01 00
+    punpcklwd   xmm4, xmm0          ; 23 03 22 02 21 01 20 00
+    punpckhwd   xmm3, xmm0          ; 33 13 32 12 31 11 30 10
+    movdqa      xmm1, xmm4          ; 23 03 22 02 21 01 20 00
+    punpcklwd   xmm4, xmm3          ; 31 21 11 01 30 20 10 00
+    punpckhwd   xmm1, xmm3          ; 33 23 13 03 32 22 12 02
+    ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    movd        xmm0, eax
+    pshufd      xmm2, xmm1, 4eh     ;ip[8] ip[12]
+    movdqa      xmm3, xmm4          ;ip[4] ip[0]
+
+    pshufd      xmm0, xmm0, 0       ;03 03 03 03 03 03 03 03
+
+    paddw       xmm4, xmm2          ;ip[4]+ip[8] ip[0]+ip[12] aka b1 a1
+    psubw       xmm3, xmm2          ;ip[4]-ip[8] ip[0]-ip[12] aka c1 d1
+
+    movdqa      xmm5, xmm4
+    punpcklqdq  xmm4, xmm3          ;d1 a1
+    punpckhqdq  xmm5, xmm3          ;c1 b1
+
+    movdqa      xmm1, xmm5          ;c1 b1
+    paddw       xmm5, xmm4          ;dl+cl a1+b1 aka op[4] op[0]
+    psubw       xmm4, xmm1          ;d1-c1 a1-b1 aka op[12] op[8]
+
+    paddw       xmm5, xmm0
+    paddw       xmm4, xmm0
+    psraw       xmm5, 3
+    psraw       xmm4, 3
+
+    movd        eax, xmm5
+    movd        ecx, xmm4
+    psrldq      xmm5, 4
+    psrldq      xmm4, 4
+    mov         word ptr[rdx+32*0], ax
+    mov         word ptr[rdx+32*2], cx
+    shr         eax, 16
+    shr         ecx, 16
+    mov         word ptr[rdx+32*4], ax
+    mov         word ptr[rdx+32*6], cx
+    movd        eax, xmm5
+    movd        ecx, xmm4
+    psrldq      xmm5, 4
+    psrldq      xmm4, 4
+    mov         word ptr[rdx+32*8], ax
+    mov         word ptr[rdx+32*10], cx
+    shr         eax, 16
+    shr         ecx, 16
+    mov         word ptr[rdx+32*12], ax
+    mov         word ptr[rdx+32*14], cx
+
+    movd        eax, xmm5
+    movd        ecx, xmm4
+    psrldq      xmm5, 4
+    psrldq      xmm4, 4
+    mov         word ptr[rdx+32*1], ax
+    mov         word ptr[rdx+32*3], cx
+    shr         eax, 16
+    shr         ecx, 16
+    mov         word ptr[rdx+32*5], ax
+    mov         word ptr[rdx+32*7], cx
+    movd        eax, xmm5
+    movd        ecx, xmm4
+    mov         word ptr[rdx+32*9], ax
+    mov         word ptr[rdx+32*11], cx
+    shr         eax, 16
+    shr         ecx, 16
+    mov         word ptr[rdx+32*13], ax
+    mov         word ptr[rdx+32*15], cx
+
+    ; begin epilog
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vp8/common/x86/loopfilter_block_sse2_x86_64.asm b/libvpx/libvpx/vp8/common/x86/loopfilter_block_sse2_x86_64.asm
new file mode 100644
index 0000000..6d5aaa1
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/loopfilter_block_sse2_x86_64.asm
@@ -0,0 +1,815 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro LF_ABS 2
+        ; %1 value not preserved
+        ; %2 value preserved
+        ; output in %1
+        movdqa      scratch1, %2            ; v2
+
+        psubusb     scratch1, %1            ; v2 - v1
+        psubusb     %1, %2                  ; v1 - v2
+        por         %1, scratch1            ; abs(v2 - v1)
+%endmacro
+
+%macro LF_FILTER_HEV_MASK 8-9
+
+        LF_ABS      %1, %2                  ; abs(p3 - p2)
+        LF_ABS      %2, %3                  ; abs(p2 - p1)
+        pmaxub      %1, %2                  ; accumulate mask
+%if %0 == 8
+        movdqa      scratch2, %3            ; save p1
+        LF_ABS      scratch2, %4            ; abs(p1 - p0)
+%endif
+        LF_ABS      %4, %5                  ; abs(p0 - q0)
+        LF_ABS      %5, %6                  ; abs(q0 - q1)
+%if %0 == 8
+        pmaxub      %5, scratch2            ; accumulate hev
+%else
+        pmaxub      %5, %9
+%endif
+        pmaxub      %1, %5                  ; accumulate mask
+
+        LF_ABS      %3, %6                  ; abs(p1 - q1)
+        LF_ABS      %6, %7                  ; abs(q1 - q2)
+        pmaxub      %1, %6                  ; accumulate mask
+        LF_ABS      %7, %8                  ; abs(q2 - q3)
+        pmaxub      %1, %7                  ; accumulate mask
+
+        paddusb     %4, %4                  ; 2 * abs(p0 - q0)
+        pand        %3, [GLOBAL(tfe)]
+        psrlw       %3, 1                   ; abs(p1 - q1) / 2
+        paddusb     %4, %3                  ; abs(p0 - q0) * 2 + abs(p1 - q1) / 2
+
+        psubusb     %1, [limit]
+        psubusb     %4, [blimit]
+        por         %1, %4
+        pcmpeqb     %1, zero                ; mask
+
+        psubusb     %5, [thresh]
+        pcmpeqb     %5, zero                ; ~hev
+%endmacro
+
+%macro LF_FILTER 6
+        ; %1-%4: p1-q1
+        ; %5: mask
+        ; %6: hev
+
+        movdqa      scratch2, %6            ; save hev
+
+        pxor        %1, [GLOBAL(t80)]       ; ps1
+        pxor        %4, [GLOBAL(t80)]       ; qs1
+        movdqa      scratch1, %1
+        psubsb      scratch1, %4            ; signed_char_clamp(ps1 - qs1)
+        pandn       scratch2, scratch1      ; vp8_filter &= hev
+
+        pxor        %2, [GLOBAL(t80)]       ; ps0
+        pxor        %3, [GLOBAL(t80)]       ; qs0
+        movdqa      scratch1, %3
+        psubsb      scratch1, %2            ; qs0 - ps0
+        paddsb      scratch2, scratch1      ; vp8_filter += (qs0 - ps0)
+        paddsb      scratch2, scratch1      ; vp8_filter += (qs0 - ps0)
+        paddsb      scratch2, scratch1      ; vp8_filter += (qs0 - ps0)
+        pand        %5, scratch2            ; &= mask
+
+        movdqa      scratch2, %5
+        paddsb      %5, [GLOBAL(t4)]        ; Filter1
+        paddsb      scratch2, [GLOBAL(t3)]  ; Filter2
+
+        ; Filter1 >> 3
+        movdqa      scratch1, zero
+        pcmpgtb     scratch1, %5
+        psrlw       %5, 3
+        pand        scratch1, [GLOBAL(te0)]
+        pand        %5, [GLOBAL(t1f)]
+        por         %5, scratch1
+
+        psubsb      %3, %5                  ; qs0 - Filter1
+        pxor        %3, [GLOBAL(t80)]
+
+        ; Filter2 >> 3
+        movdqa      scratch1, zero
+        pcmpgtb     scratch1, scratch2
+        psrlw       scratch2, 3
+        pand        scratch1, [GLOBAL(te0)]
+        pand        scratch2, [GLOBAL(t1f)]
+        por         scratch2, scratch1
+
+        paddsb      %2, scratch2            ; ps0 + Filter2
+        pxor        %2, [GLOBAL(t80)]
+
+        ; outer tap adjustments
+        paddsb      %5, [GLOBAL(t1)]
+        movdqa      scratch1, zero
+        pcmpgtb     scratch1, %5
+        psrlw       %5, 1
+        pand        scratch1, [GLOBAL(t80)]
+        pand        %5, [GLOBAL(t7f)]
+        por         %5, scratch1
+        pand        %5, %6                  ; vp8_filter &= ~hev
+
+        psubsb      %4, %5                  ; qs1 - vp8_filter
+        pxor        %4, [GLOBAL(t80)]
+
+        paddsb      %1, %5                  ; ps1 + vp8_filter
+        pxor        %1, [GLOBAL(t80)]
+%endmacro
+
+;void vp8_loop_filter_bh_y_sse2
+;(
+;    unsigned char *src_ptr,
+;    int            src_pixel_step,
+;    const char    *blimit,
+;    const char    *limit,
+;    const char    *thresh
+;)
+global sym(vp8_loop_filter_bh_y_sse2) PRIVATE
+sym(vp8_loop_filter_bh_y_sse2):
+
+%if LIBVPX_YASM_WIN64
+    %define src      rcx ; src_ptr
+    %define stride   rdx ; src_pixel_step
+    %define blimit   r8
+    %define limit    r9
+    %define thresh   r10
+
+    %define spp      rax
+    %define stride3  r11
+    %define stride5  r12
+    %define stride7  r13
+
+    push    rbp
+    mov     rbp, rsp
+    SAVE_XMM 11
+    push    r12
+    push    r13
+    mov     thresh, arg(4)
+%else
+    %define src      rdi ; src_ptr
+    %define stride   rsi ; src_pixel_step
+    %define blimit   rdx
+    %define limit    rcx
+    %define thresh   r8
+
+    %define spp      rax
+    %define stride3  r9
+    %define stride5  r10
+    %define stride7  r11
+%endif
+
+    %define scratch1 xmm5
+    %define scratch2 xmm6
+    %define zero     xmm7
+
+    %define i0       [src]
+    %define i1       [spp]
+    %define i2       [src + 2 * stride]
+    %define i3       [spp + 2 * stride]
+    %define i4       [src + 4 * stride]
+    %define i5       [spp + 4 * stride]
+    %define i6       [src + 2 * stride3]
+    %define i7       [spp + 2 * stride3]
+    %define i8       [src + 8 * stride]
+    %define i9       [spp + 8 * stride]
+    %define i10      [src + 2 * stride5]
+    %define i11      [spp + 2 * stride5]
+    %define i12      [src + 4 * stride3]
+    %define i13      [spp + 4 * stride3]
+    %define i14      [src + 2 * stride7]
+    %define i15      [spp + 2 * stride7]
+
+    ; prep work
+    lea         spp, [src + stride]
+    lea         stride3, [stride + 2 * stride]
+    lea         stride5, [stride3 + 2 * stride]
+    lea         stride7, [stride3 + 4 * stride]
+    pxor        zero, zero
+
+        ; load the first set into registers
+        movdqa       xmm0, i0
+        movdqa       xmm1, i1
+        movdqa       xmm2, i2
+        movdqa       xmm3, i3
+        movdqa       xmm4, i4
+        movdqa       xmm8, i5
+        movdqa       xmm9, i6   ; q2, will contain abs(p1-p0)
+        movdqa       xmm10, i7
+LF_FILTER_HEV_MASK xmm0, xmm1, xmm2, xmm3, xmm4, xmm8, xmm9, xmm10
+
+        movdqa       xmm1, i2
+        movdqa       xmm2, i3
+        movdqa       xmm3, i4
+        movdqa       xmm8, i5
+LF_FILTER xmm1, xmm2, xmm3, xmm8, xmm0, xmm4
+        movdqa       i2, xmm1
+        movdqa       i3, xmm2
+
+; second set
+        movdqa       i4, xmm3
+        movdqa       i5, xmm8
+
+        movdqa       xmm0, i6
+        movdqa       xmm1, i7
+        movdqa       xmm2, i8
+        movdqa       xmm4, i9
+        movdqa       xmm10, i10   ; q2, will contain abs(p1-p0)
+        movdqa       xmm11, i11
+LF_FILTER_HEV_MASK xmm3, xmm8, xmm0, xmm1, xmm2, xmm4, xmm10, xmm11, xmm9
+
+        movdqa       xmm0, i6
+        movdqa       xmm1, i7
+        movdqa       xmm4, i8
+        movdqa       xmm8, i9
+LF_FILTER xmm0, xmm1, xmm4, xmm8, xmm3, xmm2
+        movdqa       i6, xmm0
+        movdqa       i7, xmm1
+
+; last set
+        movdqa       i8, xmm4
+        movdqa       i9, xmm8
+
+        movdqa       xmm0, i10
+        movdqa       xmm1, i11
+        movdqa       xmm2, i12
+        movdqa       xmm3, i13
+        movdqa       xmm9, i14   ; q2, will contain abs(p1-p0)
+        movdqa       xmm11, i15
+LF_FILTER_HEV_MASK xmm4, xmm8, xmm0, xmm1, xmm2, xmm3, xmm9, xmm11, xmm10
+
+        movdqa       xmm0, i10
+        movdqa       xmm1, i11
+        movdqa       xmm3, i12
+        movdqa       xmm8, i13
+LF_FILTER xmm0, xmm1, xmm3, xmm8, xmm4, xmm2
+        movdqa       i10, xmm0
+        movdqa       i11, xmm1
+        movdqa       i12, xmm3
+        movdqa       i13, xmm8
+
+%if LIBVPX_YASM_WIN64
+    pop    r13
+    pop    r12
+    RESTORE_XMM
+    pop    rbp
+%endif
+
+    ret
+
+
+;void vp8_loop_filter_bv_y_sse2
+;(
+;    unsigned char *src_ptr,
+;    int            src_pixel_step,
+;    const char    *blimit,
+;    const char    *limit,
+;    const char    *thresh
+;)
+
+global sym(vp8_loop_filter_bv_y_sse2) PRIVATE
+sym(vp8_loop_filter_bv_y_sse2):
+
+%if LIBVPX_YASM_WIN64
+    %define src      rcx ; src_ptr
+    %define stride   rdx ; src_pixel_step
+    %define blimit   r8
+    %define limit    r9
+    %define thresh   r10
+
+    %define spp      rax
+    %define stride3  r11
+    %define stride5  r12
+    %define stride7  r13
+
+    push    rbp
+    mov     rbp, rsp
+    SAVE_XMM 15
+    push    r12
+    push    r13
+    mov     thresh, arg(4)
+%else
+    %define src      rdi
+    %define stride   rsi
+    %define blimit   rdx
+    %define limit    rcx
+    %define thresh   r8
+
+    %define spp      rax
+    %define stride3  r9
+    %define stride5  r10
+    %define stride7  r11
+%endif
+
+    %define scratch1 xmm5
+    %define scratch2 xmm6
+    %define zero     xmm7
+
+    %define s0       [src]
+    %define s1       [spp]
+    %define s2       [src + 2 * stride]
+    %define s3       [spp + 2 * stride]
+    %define s4       [src + 4 * stride]
+    %define s5       [spp + 4 * stride]
+    %define s6       [src + 2 * stride3]
+    %define s7       [spp + 2 * stride3]
+    %define s8       [src + 8 * stride]
+    %define s9       [spp + 8 * stride]
+    %define s10      [src + 2 * stride5]
+    %define s11      [spp + 2 * stride5]
+    %define s12      [src + 4 * stride3]
+    %define s13      [spp + 4 * stride3]
+    %define s14      [src + 2 * stride7]
+    %define s15      [spp + 2 * stride7]
+
+    %define i0       [rsp]
+    %define i1       [rsp + 16]
+    %define i2       [rsp + 32]
+    %define i3       [rsp + 48]
+    %define i4       [rsp + 64]
+    %define i5       [rsp + 80]
+    %define i6       [rsp + 96]
+    %define i7       [rsp + 112]
+    %define i8       [rsp + 128]
+    %define i9       [rsp + 144]
+    %define i10      [rsp + 160]
+    %define i11      [rsp + 176]
+    %define i12      [rsp + 192]
+    %define i13      [rsp + 208]
+    %define i14      [rsp + 224]
+    %define i15      [rsp + 240]
+
+    ALIGN_STACK 16, rax
+
+    ; reserve stack space
+    %define      temp_storage  0 ; size is 256 (16*16)
+    %define      stack_size 256
+    sub          rsp, stack_size
+
+    ; prep work
+    lea         spp, [src + stride]
+    lea         stride3, [stride + 2 * stride]
+    lea         stride5, [stride3 + 2 * stride]
+    lea         stride7, [stride3 + 4 * stride]
+
+        ; 8-f
+        movdqa      xmm0, s8
+        movdqa      xmm1, xmm0
+        punpcklbw   xmm0, s9                ; 80 90
+        punpckhbw   xmm1, s9                ; 88 98
+
+        movdqa      xmm2, s10
+        movdqa      xmm3, xmm2
+        punpcklbw   xmm2, s11 ; a0 b0
+        punpckhbw   xmm3, s11 ; a8 b8
+
+        movdqa      xmm4, xmm0
+        punpcklwd   xmm0, xmm2              ; 80 90 a0 b0
+        punpckhwd   xmm4, xmm2              ; 84 94 a4 b4
+
+        movdqa      xmm2, xmm1
+        punpcklwd   xmm1, xmm3              ; 88 98 a8 b8
+        punpckhwd   xmm2, xmm3              ; 8c 9c ac bc
+
+        ; using xmm[0124]
+        ; work on next 4 rows
+
+        movdqa      xmm3, s12
+        movdqa      xmm5, xmm3
+        punpcklbw   xmm3, s13 ; c0 d0
+        punpckhbw   xmm5, s13 ; c8 d8
+
+        movdqa      xmm6, s14
+        movdqa      xmm7, xmm6
+        punpcklbw   xmm6, s15 ; e0 f0
+        punpckhbw   xmm7, s15 ; e8 f8
+
+        movdqa      xmm8, xmm3
+        punpcklwd   xmm3, xmm6              ; c0 d0 e0 f0
+        punpckhwd   xmm8, xmm6              ; c4 d4 e4 f4
+
+        movdqa      xmm6, xmm5
+        punpcklwd   xmm5, xmm7              ; c8 d8 e8 f8
+        punpckhwd   xmm6, xmm7              ; cc dc ec fc
+
+        ; pull the third and fourth sets together
+
+        movdqa      xmm7, xmm0
+        punpckldq   xmm0, xmm3              ; 80 90 a0 b0 c0 d0 e0 f0
+        punpckhdq   xmm7, xmm3              ; 82 92 a2 b2 c2 d2 e2 f2
+
+        movdqa      xmm3, xmm4
+        punpckldq   xmm4, xmm8              ; 84 94 a4 b4 c4 d4 e4 f4
+        punpckhdq   xmm3, xmm8              ; 86 96 a6 b6 c6 d6 e6 f6
+
+        movdqa      xmm8, xmm1
+        punpckldq   xmm1, xmm5              ; 88 88 a8 b8 c8 d8 e8 f8
+        punpckhdq   xmm8, xmm5              ; 8a 9a aa ba ca da ea fa
+
+        movdqa      xmm5, xmm2
+        punpckldq   xmm2, xmm6              ; 8c 9c ac bc cc dc ec fc
+        punpckhdq   xmm5, xmm6              ; 8e 9e ae be ce de ee fe
+
+        ; save the calculations. we only have 15 registers ...
+        movdqa      i0, xmm0
+        movdqa      i1, xmm7
+        movdqa      i2, xmm4
+        movdqa      i3, xmm3
+        movdqa      i4, xmm1
+        movdqa      i5, xmm8
+        movdqa      i6, xmm2
+        movdqa      i7, xmm5
+
+        ; 0-7
+        movdqa      xmm0, s0
+        movdqa      xmm1, xmm0
+        punpcklbw   xmm0, s1 ; 00 10
+        punpckhbw   xmm1, s1 ; 08 18
+
+        movdqa      xmm2, s2
+        movdqa      xmm3, xmm2
+        punpcklbw   xmm2, s3 ; 20 30
+        punpckhbw   xmm3, s3 ; 28 38
+
+        movdqa      xmm4, xmm0
+        punpcklwd   xmm0, xmm2              ; 00 10 20 30
+        punpckhwd   xmm4, xmm2              ; 04 14 24 34
+
+        movdqa      xmm2, xmm1
+        punpcklwd   xmm1, xmm3              ; 08 18 28 38
+        punpckhwd   xmm2, xmm3              ; 0c 1c 2c 3c
+
+        ; using xmm[0124]
+        ; work on next 4 rows
+
+        movdqa      xmm3, s4
+        movdqa      xmm5, xmm3
+        punpcklbw   xmm3, s5 ; 40 50
+        punpckhbw   xmm5, s5 ; 48 58
+
+        movdqa      xmm6, s6
+        movdqa      xmm7, xmm6
+        punpcklbw   xmm6, s7   ; 60 70
+        punpckhbw   xmm7, s7   ; 68 78
+
+        movdqa      xmm8, xmm3
+        punpcklwd   xmm3, xmm6              ; 40 50 60 70
+        punpckhwd   xmm8, xmm6              ; 44 54 64 74
+
+        movdqa      xmm6, xmm5
+        punpcklwd   xmm5, xmm7              ; 48 58 68 78
+        punpckhwd   xmm6, xmm7              ; 4c 5c 6c 7c
+
+        ; pull the first two sets together
+
+        movdqa      xmm7, xmm0
+        punpckldq   xmm0, xmm3              ; 00 10 20 30 40 50 60 70
+        punpckhdq   xmm7, xmm3              ; 02 12 22 32 42 52 62 72
+
+        movdqa      xmm3, xmm4
+        punpckldq   xmm4, xmm8              ; 04 14 24 34 44 54 64 74
+        punpckhdq   xmm3, xmm8              ; 06 16 26 36 46 56 66 76
+
+        movdqa      xmm8, xmm1
+        punpckldq   xmm1, xmm5              ; 08 18 28 38 48 58 68 78
+        punpckhdq   xmm8, xmm5              ; 0a 1a 2a 3a 4a 5a 6a 7a
+
+        movdqa      xmm5, xmm2
+        punpckldq   xmm2, xmm6              ; 0c 1c 2c 3c 4c 5c 6c 7c
+        punpckhdq   xmm5, xmm6              ; 0e 1e 2e 3e 4e 5e 6e 7e
+        ; final combination
+
+        movdqa      xmm6, xmm0
+        punpcklqdq  xmm0, i0
+        punpckhqdq  xmm6, i0
+
+        movdqa      xmm9, xmm7
+        punpcklqdq  xmm7, i1
+        punpckhqdq  xmm9, i1
+
+        movdqa      xmm10, xmm4
+        punpcklqdq  xmm4, i2
+        punpckhqdq  xmm10, i2
+
+        movdqa      xmm11, xmm3
+        punpcklqdq  xmm3, i3
+        punpckhqdq  xmm11, i3
+
+        movdqa      xmm12, xmm1
+        punpcklqdq  xmm1, i4
+        punpckhqdq  xmm12, i4
+
+        movdqa      xmm13, xmm8
+        punpcklqdq  xmm8, i5
+        punpckhqdq  xmm13, i5
+
+        movdqa      xmm14, xmm2
+        punpcklqdq  xmm2, i6
+        punpckhqdq  xmm14, i6
+
+        movdqa      xmm15, xmm5
+        punpcklqdq  xmm5, i7
+        punpckhqdq  xmm15, i7
+
+        movdqa      i0, xmm0
+        movdqa      i1, xmm6
+        movdqa      i2, xmm7
+        movdqa      i3, xmm9
+        movdqa      i4, xmm4
+        movdqa      i5, xmm10
+        movdqa      i6, xmm3
+        movdqa      i7, xmm11
+        movdqa      i8, xmm1
+        movdqa      i9, xmm12
+        movdqa      i10, xmm8
+        movdqa      i11, xmm13
+        movdqa      i12, xmm2
+        movdqa      i13, xmm14
+        movdqa      i14, xmm5
+        movdqa      i15, xmm15
+
+; TRANSPOSED DATA AVAILABLE ON THE STACK
+
+        movdqa      xmm12, xmm6
+        movdqa      xmm13, xmm7
+
+        pxor        zero, zero
+
+LF_FILTER_HEV_MASK xmm0, xmm12, xmm13, xmm9, xmm4, xmm10, xmm3, xmm11
+
+        movdqa       xmm1, i2
+        movdqa       xmm2, i3
+        movdqa       xmm8, i4
+        movdqa       xmm9, i5
+LF_FILTER xmm1, xmm2, xmm8, xmm9, xmm0, xmm4
+        movdqa       i2, xmm1
+        movdqa       i3, xmm2
+
+; second set
+        movdqa       i4, xmm8
+        movdqa       i5, xmm9
+
+        movdqa       xmm0, i6
+        movdqa       xmm1, i7
+        movdqa       xmm2, i8
+        movdqa       xmm4, i9
+        movdqa       xmm10, i10   ; q2, will contain abs(p1-p0)
+        movdqa       xmm11, i11
+LF_FILTER_HEV_MASK xmm8, xmm9, xmm0, xmm1, xmm2, xmm4, xmm10, xmm11, xmm3
+
+        movdqa       xmm0, i6
+        movdqa       xmm1, i7
+        movdqa       xmm3, i8
+        movdqa       xmm4, i9
+LF_FILTER xmm0, xmm1, xmm3, xmm4, xmm8, xmm2
+        movdqa       i6, xmm0
+        movdqa       i7, xmm1
+
+; last set
+        movdqa       i8, xmm3
+        movdqa       i9, xmm4
+
+        movdqa       xmm0, i10
+        movdqa       xmm1, i11
+        movdqa       xmm2, i12
+        movdqa       xmm8, i13
+        movdqa       xmm9, i14   ; q2, will contain abs(p1-p0)
+        movdqa       xmm11, i15
+LF_FILTER_HEV_MASK xmm3, xmm4, xmm0, xmm1, xmm2, xmm8, xmm9, xmm11, xmm10
+
+        movdqa       xmm0, i10
+        movdqa       xmm1, i11
+        movdqa       xmm4, i12
+        movdqa       xmm8, i13
+LF_FILTER xmm0, xmm1, xmm4, xmm8, xmm3, xmm2
+        movdqa       i10, xmm0
+        movdqa       i11, xmm1
+        movdqa       i12, xmm4
+        movdqa       i13, xmm8
+
+
+; RESHUFFLE AND WRITE OUT
+        ; 8-f
+        movdqa      xmm0, i8
+        movdqa      xmm1, xmm0
+        punpcklbw   xmm0, i9                ; 80 90
+        punpckhbw   xmm1, i9                ; 88 98
+
+        movdqa      xmm2, i10
+        movdqa      xmm3, xmm2
+        punpcklbw   xmm2, i11               ; a0 b0
+        punpckhbw   xmm3, i11               ; a8 b8
+
+        movdqa      xmm4, xmm0
+        punpcklwd   xmm0, xmm2              ; 80 90 a0 b0
+        punpckhwd   xmm4, xmm2              ; 84 94 a4 b4
+
+        movdqa      xmm2, xmm1
+        punpcklwd   xmm1, xmm3              ; 88 98 a8 b8
+        punpckhwd   xmm2, xmm3              ; 8c 9c ac bc
+
+        ; using xmm[0124]
+        ; work on next 4 rows
+
+        movdqa      xmm3, i12
+        movdqa      xmm5, xmm3
+        punpcklbw   xmm3, i13               ; c0 d0
+        punpckhbw   xmm5, i13               ; c8 d8
+
+        movdqa      xmm6, i14
+        movdqa      xmm7, xmm6
+        punpcklbw   xmm6, i15               ; e0 f0
+        punpckhbw   xmm7, i15               ; e8 f8
+
+        movdqa      xmm8, xmm3
+        punpcklwd   xmm3, xmm6              ; c0 d0 e0 f0
+        punpckhwd   xmm8, xmm6              ; c4 d4 e4 f4
+
+        movdqa      xmm6, xmm5
+        punpcklwd   xmm5, xmm7              ; c8 d8 e8 f8
+        punpckhwd   xmm6, xmm7              ; cc dc ec fc
+
+        ; pull the third and fourth sets together
+
+        movdqa      xmm7, xmm0
+        punpckldq   xmm0, xmm3              ; 80 90 a0 b0 c0 d0 e0 f0
+        punpckhdq   xmm7, xmm3              ; 82 92 a2 b2 c2 d2 e2 f2
+
+        movdqa      xmm3, xmm4
+        punpckldq   xmm4, xmm8              ; 84 94 a4 b4 c4 d4 e4 f4
+        punpckhdq   xmm3, xmm8              ; 86 96 a6 b6 c6 d6 e6 f6
+
+        movdqa      xmm8, xmm1
+        punpckldq   xmm1, xmm5              ; 88 88 a8 b8 c8 d8 e8 f8
+        punpckhdq   xmm8, xmm5              ; 8a 9a aa ba ca da ea fa
+
+        movdqa      xmm5, xmm2
+        punpckldq   xmm2, xmm6              ; 8c 9c ac bc cc dc ec fc
+        punpckhdq   xmm5, xmm6              ; 8e 9e ae be ce de ee fe
+
+        ; save the calculations. we only have 15 registers ...
+        movdqa      i8, xmm0
+        movdqa      i9, xmm7
+        movdqa      i10, xmm4
+        movdqa      i11, xmm3
+        movdqa      i12, xmm1
+        movdqa      i13, xmm8
+        movdqa      i14, xmm2
+        movdqa      i15, xmm5
+
+        ; 0-7
+        movdqa      xmm0, i0
+        movdqa      xmm1, xmm0
+        punpcklbw   xmm0, i1                ; 00 10
+        punpckhbw   xmm1, i1                ; 08 18
+
+        movdqa      xmm2, i2
+        movdqa      xmm3, xmm2
+        punpcklbw   xmm2, i3                ; 20 30
+        punpckhbw   xmm3, i3                ; 28 38
+
+        movdqa      xmm4, xmm0
+        punpcklwd   xmm0, xmm2              ; 00 10 20 30
+        punpckhwd   xmm4, xmm2              ; 04 14 24 34
+
+        movdqa      xmm2, xmm1
+        punpcklwd   xmm1, xmm3              ; 08 18 28 38
+        punpckhwd   xmm2, xmm3              ; 0c 1c 2c 3c
+
+        ; using xmm[0124]
+        ; work on next 4 rows
+
+        movdqa      xmm3, i4
+        movdqa      xmm5, xmm3
+        punpcklbw   xmm3, i5                ; 40 50
+        punpckhbw   xmm5, i5                ; 48 58
+
+        movdqa      xmm6, i6
+        movdqa      xmm7, xmm6
+        punpcklbw   xmm6, i7                ; 60 70
+        punpckhbw   xmm7, i7                ; 68 78
+
+        movdqa      xmm8, xmm3
+        punpcklwd   xmm3, xmm6              ; 40 50 60 70
+        punpckhwd   xmm8, xmm6              ; 44 54 64 74
+
+        movdqa      xmm6, xmm5
+        punpcklwd   xmm5, xmm7              ; 48 58 68 78
+        punpckhwd   xmm6, xmm7              ; 4c 5c 6c 7c
+
+        ; pull the first two sets together
+
+        movdqa      xmm7, xmm0
+        punpckldq   xmm0, xmm3              ; 00 10 20 30 40 50 60 70
+        punpckhdq   xmm7, xmm3              ; 02 12 22 32 42 52 62 72
+
+        movdqa      xmm3, xmm4
+        punpckldq   xmm4, xmm8              ; 04 14 24 34 44 54 64 74
+        punpckhdq   xmm3, xmm8              ; 06 16 26 36 46 56 66 76
+
+        movdqa      xmm8, xmm1
+        punpckldq   xmm1, xmm5              ; 08 18 28 38 48 58 68 78
+        punpckhdq   xmm8, xmm5              ; 0a 1a 2a 3a 4a 5a 6a 7a
+
+        movdqa      xmm5, xmm2
+        punpckldq   xmm2, xmm6              ; 0c 1c 2c 3c 4c 5c 6c 7c
+        punpckhdq   xmm5, xmm6              ; 0e 1e 2e 3e 4e 5e 6e 7e
+        ; final combination
+
+        movdqa      xmm6, xmm0
+        punpcklqdq  xmm0, i8
+        punpckhqdq  xmm6, i8
+
+        movdqa      xmm9, xmm7
+        punpcklqdq  xmm7, i9
+        punpckhqdq  xmm9, i9
+
+        movdqa      xmm10, xmm4
+        punpcklqdq  xmm4, i10
+        punpckhqdq  xmm10, i10
+
+        movdqa      xmm11, xmm3
+        punpcklqdq  xmm3, i11
+        punpckhqdq  xmm11, i11
+
+        movdqa      xmm12, xmm1
+        punpcklqdq  xmm1, i12
+        punpckhqdq  xmm12, i12
+
+        movdqa      xmm13, xmm8
+        punpcklqdq  xmm8, i13
+        punpckhqdq  xmm13, i13
+
+        movdqa      xmm14, xmm2
+        punpcklqdq  xmm2, i14
+        punpckhqdq  xmm14, i14
+
+        movdqa      xmm15, xmm5
+        punpcklqdq  xmm5, i15
+        punpckhqdq  xmm15, i15
+
+        movdqa      s0, xmm0
+        movdqa      s1, xmm6
+        movdqa      s2, xmm7
+        movdqa      s3, xmm9
+        movdqa      s4, xmm4
+        movdqa      s5, xmm10
+        movdqa      s6, xmm3
+        movdqa      s7, xmm11
+        movdqa      s8, xmm1
+        movdqa      s9, xmm12
+        movdqa      s10, xmm8
+        movdqa      s11, xmm13
+        movdqa      s12, xmm2
+        movdqa      s13, xmm14
+        movdqa      s14, xmm5
+        movdqa      s15, xmm15
+
+    ; free stack space
+    add          rsp, stack_size
+
+    ; un-ALIGN_STACK
+    pop          rsp
+
+%if LIBVPX_YASM_WIN64
+    pop    r13
+    pop    r12
+    RESTORE_XMM
+    pop    rbp
+%endif
+
+    ret
+
+SECTION_RODATA
+align 16
+te0:
+    times 16 db 0xe0
+align 16
+t7f:
+    times 16 db 0x7f
+align 16
+tfe:
+    times 16 db 0xfe
+align 16
+t1f:
+    times 16 db 0x1f
+align 16
+t80:
+    times 16 db 0x80
+align 16
+t1:
+    times 16 db 0x01
+align 16
+t3:
+    times 16 db 0x03
+align 16
+t4:
+    times 16 db 0x04
diff --git a/libvpx/libvpx/vp8/common/x86/loopfilter_sse2.asm b/libvpx/libvpx/vp8/common/x86/loopfilter_sse2.asm
new file mode 100644
index 0000000..1913abc
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/loopfilter_sse2.asm
@@ -0,0 +1,1640 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+%define _t0 0
+%define _t1 _t0 + 16
+%define _p3 _t1 + 16
+%define _p2 _p3 + 16
+%define _p1 _p2 + 16
+%define _p0 _p1 + 16
+%define _q0 _p0 + 16
+%define _q1 _q0 + 16
+%define _q2 _q1 + 16
+%define _q3 _q2 + 16
+%define lf_var_size 160
+
+; Use of pmaxub instead of psubusb to compute filter mask was seen
+; in ffvp8
+
+%macro LFH_FILTER_AND_HEV_MASK 1
+%if %1
+        movdqa      xmm2,                   [rdi+2*rax]       ; q3
+        movdqa      xmm1,                   [rsi+2*rax]       ; q2
+        movdqa      xmm4,                   [rsi+rax]         ; q1
+        movdqa      xmm5,                   [rsi]             ; q0
+        neg         rax                     ; negate pitch to deal with above border
+%else
+        movlps      xmm2,                   [rsi + rcx*2]     ; q3
+        movlps      xmm1,                   [rsi + rcx]       ; q2
+        movlps      xmm4,                   [rsi]             ; q1
+        movlps      xmm5,                   [rsi + rax]       ; q0
+
+        movhps      xmm2,                   [rdi + rcx*2]
+        movhps      xmm1,                   [rdi + rcx]
+        movhps      xmm4,                   [rdi]
+        movhps      xmm5,                   [rdi + rax]
+
+        lea         rsi,                    [rsi + rax*4]
+        lea         rdi,                    [rdi + rax*4]
+
+        movdqa      [rsp+_q2],              xmm1              ; store q2
+        movdqa      [rsp+_q1],              xmm4              ; store q1
+%endif
+        movdqa      xmm7,                   [rdx]             ;limit
+
+        movdqa      xmm6,                   xmm1              ; q2
+        movdqa      xmm3,                   xmm4              ; q1
+
+        psubusb     xmm1,                   xmm2              ; q2-=q3
+        psubusb     xmm2,                   xmm6              ; q3-=q2
+
+        psubusb     xmm4,                   xmm6              ; q1-=q2
+        psubusb     xmm6,                   xmm3              ; q2-=q1
+
+        por         xmm4,                   xmm6              ; abs(q2-q1)
+        por         xmm1,                   xmm2              ; abs(q3-q2)
+
+        movdqa      xmm0,                   xmm5              ; q0
+        pmaxub      xmm1,                   xmm4
+
+        psubusb     xmm5,                   xmm3              ; q0-=q1
+        psubusb     xmm3,                   xmm0              ; q1-=q0
+
+        por         xmm5,                   xmm3              ; abs(q0-q1)
+        movdqa      [rsp+_t0],              xmm5              ; save to t0
+
+        pmaxub      xmm1,                   xmm5
+
+%if %1
+        movdqa      xmm2,                   [rsi+4*rax]       ; p3
+        movdqa      xmm4,                   [rdi+4*rax]       ; p2
+        movdqa      xmm6,                   [rsi+2*rax]       ; p1
+%else
+        movlps      xmm2,                   [rsi + rax]       ; p3
+        movlps      xmm4,                   [rsi]             ; p2
+        movlps      xmm6,                   [rsi + rcx]       ; p1
+
+        movhps      xmm2,                   [rdi + rax]
+        movhps      xmm4,                   [rdi]
+        movhps      xmm6,                   [rdi + rcx]
+
+        movdqa      [rsp+_p2],              xmm4              ; store p2
+        movdqa      [rsp+_p1],              xmm6              ; store p1
+%endif
+
+        movdqa      xmm5,                   xmm4              ; p2
+        movdqa      xmm3,                   xmm6              ; p1
+
+        psubusb     xmm4,                   xmm2              ; p2-=p3
+        psubusb     xmm2,                   xmm5              ; p3-=p2
+
+        psubusb     xmm3,                   xmm5              ; p1-=p2
+        pmaxub      xmm1,                   xmm4              ; abs(p3 - p2)
+
+        psubusb     xmm5,                   xmm6              ; p2-=p1
+        pmaxub      xmm1,                   xmm2              ; abs(p3 - p2)
+
+        pmaxub      xmm1,                   xmm5              ; abs(p2 - p1)
+        movdqa      xmm2,                   xmm6              ; p1
+
+        pmaxub      xmm1,                   xmm3              ; abs(p2 - p1)
+%if %1
+        movdqa      xmm4,                   [rsi+rax]         ; p0
+        movdqa      xmm3,                   [rdi]             ; q1
+%else
+        movlps      xmm4,                   [rsi + rcx*2]     ; p0
+        movhps      xmm4,                   [rdi + rcx*2]
+        movdqa      xmm3,                   [rsp+_q1]                ; q1
+%endif
+
+        movdqa      xmm5,                   xmm4              ; p0
+        psubusb     xmm4,                   xmm6              ; p0-=p1
+
+        psubusb     xmm6,                   xmm5              ; p1-=p0
+
+        por         xmm6,                   xmm4              ; abs(p1 - p0)
+        mov         rdx,                    arg(2)            ; get blimit
+
+        movdqa     [rsp+_t1],               xmm6              ; save to t1
+
+        movdqa      xmm4,                   xmm3              ; q1
+        pmaxub      xmm1,                   xmm6
+
+        psubusb     xmm3,                   xmm2              ; q1-=p1
+        psubusb     xmm2,                   xmm4              ; p1-=q1
+
+        psubusb     xmm1,                   xmm7
+        por         xmm2,                   xmm3              ; abs(p1-q1)
+
+        movdqa      xmm7,                   [rdx]             ; blimit
+        mov         rdx,                    arg(4)            ; hev get thresh
+
+        movdqa      xmm3,                   xmm0              ; q0
+        pand        xmm2,                   [GLOBAL(tfe)]     ; set lsb of each byte to zero
+
+        movdqa      xmm6,                   xmm5              ; p0
+        psrlw       xmm2,                   1                 ; abs(p1-q1)/2
+
+        psubusb     xmm5,                   xmm3              ; p0-=q0
+        psubusb     xmm3,                   xmm6              ; q0-=p0
+        por         xmm5,                   xmm3              ; abs(p0 - q0)
+
+        paddusb     xmm5,                   xmm5              ; abs(p0-q0)*2
+
+        movdqa      xmm4,                   [rsp+_t0]                ; hev get abs (q1 - q0)
+        movdqa      xmm3,                   [rsp+_t1]                ; get abs (p1 - p0)
+
+        paddusb     xmm5,                   xmm2              ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+        movdqa      xmm2,                   [rdx]             ; hev
+
+        psubusb     xmm5,                   xmm7              ; abs (p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        psubusb     xmm4,                   xmm2              ; hev
+
+        psubusb     xmm3,                   xmm2              ; hev
+        por         xmm1,                   xmm5
+
+        pxor        xmm7,                   xmm7
+        paddb       xmm4,                   xmm3              ; hev abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
+
+        pcmpeqb     xmm4,                   xmm5              ; hev
+        pcmpeqb     xmm3,                   xmm3              ; hev
+
+        pcmpeqb     xmm1,                   xmm7              ; mask xmm1
+        pxor        xmm4,                   xmm3              ; hev
+%endmacro
+
+%macro B_FILTER 1
+        movdqa      xmm3,                   [GLOBAL(t80)]
+%if %1 == 0
+        movdqa      xmm2,                   [rsp+_p1]                ; p1
+        movdqa      xmm7,                   [rsp+_q1]                ; q1
+%elif %1 == 1
+        movdqa      xmm2,                   [rsi+2*rax]       ; p1
+        movdqa      xmm7,                   [rdi]             ; q1
+%elif %1 == 2
+        movdqa      xmm2,                   [rsp+_p1]         ; p1
+        movdqa      xmm6,                   [rsp+_p0]         ; p0
+        movdqa      xmm0,                   [rsp+_q0]         ; q0
+        movdqa      xmm7,                   [rsp+_q1]         ; q1
+%endif
+
+        pxor        xmm2,                   xmm3              ; p1 offset to convert to signed values
+        pxor        xmm7,                   xmm3              ; q1 offset to convert to signed values
+
+        psubsb      xmm2,                   xmm7              ; p1 - q1
+        pxor        xmm6,                   xmm3              ; offset to convert to signed values
+
+        pand        xmm2,                   xmm4              ; high var mask (hvm)(p1 - q1)
+        pxor        xmm0,                   xmm3              ; offset to convert to signed values
+
+        movdqa      xmm3,                   xmm0              ; q0
+        psubsb      xmm0,                   xmm6              ; q0 - p0
+        paddsb      xmm2,                   xmm0              ; 1 * (q0 - p0) + hvm(p1 - q1)
+        paddsb      xmm2,                   xmm0              ; 2 * (q0 - p0) + hvm(p1 - q1)
+        paddsb      xmm2,                   xmm0              ; 3 * (q0 - p0) + hvm(p1 - q1)
+        pand        xmm1,                   xmm2              ; mask filter values we don't care about
+
+        movdqa      xmm2,                   xmm1
+        paddsb      xmm1,                   [GLOBAL(t4)]      ; 3* (q0 - p0) + hvm(p1 - q1) + 4
+        paddsb      xmm2,                   [GLOBAL(t3)]      ; 3* (q0 - p0) + hvm(p1 - q1) + 3
+
+        punpckhbw   xmm5,                   xmm2              ; axbxcxdx
+        punpcklbw   xmm2,                   xmm2              ; exfxgxhx
+
+        punpcklbw   xmm0,                   xmm1              ; exfxgxhx
+        psraw       xmm5,                   11                ; sign extended shift right by 3
+
+        punpckhbw   xmm1,                   xmm1              ; axbxcxdx
+        psraw       xmm2,                   11                ; sign extended shift right by 3
+
+        packsswb    xmm2,                   xmm5              ; (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
+        psraw       xmm0,                   11                ; sign extended shift right by 3
+
+        psraw       xmm1,                   11                ; sign extended shift right by 3
+        movdqa      xmm5,                   xmm0              ; save results
+
+        packsswb    xmm0,                   xmm1              ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
+
+        paddsb      xmm6,                   xmm2              ; p0+= p0 add
+
+        movdqa      xmm2,                   [GLOBAL(ones)]
+        paddsw      xmm5,                   xmm2
+        paddsw      xmm1,                   xmm2
+        psraw       xmm5,                   1                 ; partial shifted one more time for 2nd tap
+        psraw       xmm1,                   1                 ; partial shifted one more time for 2nd tap
+        packsswb    xmm5,                   xmm1              ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>4
+        movdqa      xmm2,                   [GLOBAL(t80)]
+
+%if %1 == 0
+        movdqa      xmm1,                   [rsp+_p1]         ; p1
+        lea         rsi,                    [rsi + rcx*2]
+        lea         rdi,                    [rdi + rcx*2]
+%elif %1 == 1
+        movdqa      xmm1,                   [rsi+2*rax]       ; p1
+%elif %1 == 2
+        movdqa      xmm1,                   [rsp+_p1]         ; p1
+%endif
+
+        pandn       xmm4,                   xmm5              ; high edge variance additive
+        pxor        xmm6,                   xmm2              ; unoffset
+
+        pxor        xmm1,                   xmm2              ; reoffset
+        psubsb      xmm3,                   xmm0              ; q0-= q0 add
+
+        paddsb      xmm1,                   xmm4              ; p1+= p1 add
+        pxor        xmm3,                   xmm2              ; unoffset
+
+        pxor        xmm1,                   xmm2              ; unoffset
+        psubsb      xmm7,                   xmm4              ; q1-= q1 add
+
+        pxor        xmm7,                   xmm2              ; unoffset
+%if %1 == 0
+        movq        [rsi],                  xmm6              ; p0
+        movhps      [rdi],                  xmm6
+        movq        [rsi + rax],            xmm1              ; p1
+        movhps      [rdi + rax],            xmm1
+        movq        [rsi + rcx],            xmm3              ; q0
+        movhps      [rdi + rcx],            xmm3
+        movq        [rsi + rcx*2],          xmm7              ; q1
+        movhps      [rdi + rcx*2],          xmm7
+%elif %1 == 1
+        movdqa      [rsi+rax],              xmm6              ; write back
+        movdqa      [rsi+2*rax],            xmm1              ; write back
+        movdqa      [rsi],                  xmm3              ; write back
+        movdqa      [rdi],                  xmm7              ; write back
+%endif
+
+%endmacro
+
+%if ABI_IS_32BIT
+
+;void vp8_loop_filter_horizontal_edge_sse2
+;(
+;    unsigned char *src_ptr,
+;    int            src_pixel_step,
+;    const char    *blimit,
+;    const char    *limit,
+;    const char    *thresh,
+;)
+global sym(vp8_loop_filter_horizontal_edge_sse2) PRIVATE
+sym(vp8_loop_filter_horizontal_edge_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, lf_var_size
+
+        mov         rsi,                    arg(0)           ;src_ptr
+        movsxd      rax,                    dword ptr arg(1) ;src_pixel_step
+
+        mov         rdx,                    arg(3)           ;limit
+
+        lea         rdi,                    [rsi+rax]        ; rdi points to row +1 for indirect addressing
+
+        ; calculate breakout conditions and high edge variance
+        LFH_FILTER_AND_HEV_MASK 1
+        ; filter and write back the result
+        B_FILTER 1
+
+    add rsp, lf_var_size
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%endif
+
+;void vp8_loop_filter_horizontal_edge_uv_sse2
+;(
+;    unsigned char *src_ptr,
+;    int            src_pixel_step,
+;    const char    *blimit,
+;    const char    *limit,
+;    const char    *thresh,
+;    int            count
+;)
+global sym(vp8_loop_filter_horizontal_edge_uv_sse2) PRIVATE
+sym(vp8_loop_filter_horizontal_edge_uv_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, lf_var_size
+
+        mov         rsi,                    arg(0)             ; u
+        mov         rdi,                    arg(5)             ; v
+        movsxd      rax,                    dword ptr arg(1)   ; src_pixel_step
+        mov         rcx,                    rax
+        neg         rax                     ; negate pitch to deal with above border
+
+        mov         rdx,                    arg(3)             ;limit
+
+        lea         rsi,                    [rsi + rcx]
+        lea         rdi,                    [rdi + rcx]
+
+        ; calculate breakout conditions and high edge variance
+        LFH_FILTER_AND_HEV_MASK 0
+        ; filter and write back the result
+        B_FILTER 0
+
+    add rsp, lf_var_size
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+%macro MB_FILTER_AND_WRITEBACK 1
+        movdqa      xmm3,                   [GLOBAL(t80)]
+%if %1 == 0
+        movdqa      xmm2,                   [rsp+_p1]              ; p1
+        movdqa      xmm7,                   [rsp+_q1]              ; q1
+%elif %1 == 1
+        movdqa      xmm2,                   [rsi+2*rax]     ; p1
+        movdqa      xmm7,                   [rdi]           ; q1
+
+        mov         rcx,                    rax
+        neg         rcx
+%elif %1 == 2
+        movdqa      xmm2,                   [rsp+_p1]       ; p1
+        movdqa      xmm6,                   [rsp+_p0]       ; p0
+        movdqa      xmm0,                   [rsp+_q0]       ; q0
+        movdqa      xmm7,                   [rsp+_q1]       ; q1
+%endif
+
+        pxor        xmm2,                   xmm3            ; p1 offset to convert to signed values
+        pxor        xmm7,                   xmm3            ; q1 offset to convert to signed values
+        pxor        xmm6,                   xmm3            ; offset to convert to signed values
+        pxor        xmm0,                   xmm3            ; offset to convert to signed values
+
+        psubsb      xmm2,                   xmm7            ; p1 - q1
+
+        movdqa      xmm3,                   xmm0            ; q0
+        psubsb      xmm0,                   xmm6            ; q0 - p0
+        paddsb      xmm2,                   xmm0            ; 1 * (q0 - p0) + (p1 - q1)
+        paddsb      xmm2,                   xmm0            ; 2 * (q0 - p0)
+        paddsb      xmm2,                   xmm0            ; 3 * (q0 - p0) + (p1 - q1)
+        pand        xmm1,                   xmm2            ; mask filter values we don't care about
+
+        movdqa      xmm2,                   xmm1            ; vp8_filter
+
+        pand        xmm2,                   xmm4            ; Filter2 = vp8_filter & hev
+        pxor        xmm0,                   xmm0
+
+        pandn       xmm4,                   xmm1            ; vp8_filter&=~hev
+        pxor        xmm1,                   xmm1
+
+        punpcklbw   xmm0,                   xmm4            ; Filter 2 (hi)
+        punpckhbw   xmm1,                   xmm4            ; Filter 2 (lo)
+
+        movdqa      xmm5,                   xmm2
+
+        movdqa      xmm4,                   [GLOBAL(s9)]
+        paddsb      xmm5,                   [GLOBAL(t3)]    ; vp8_signed_char_clamp(Filter2 + 3)
+        paddsb      xmm2,                   [GLOBAL(t4)]    ; vp8_signed_char_clamp(Filter2 + 4)
+
+        pmulhw      xmm1,                   xmm4            ; Filter 2 (lo) * 9
+        pmulhw      xmm0,                   xmm4            ; Filter 2 (hi) * 9
+
+        punpckhbw   xmm7,                   xmm5            ; axbxcxdx
+        punpcklbw   xmm5,                   xmm5            ; exfxgxhx
+
+        psraw       xmm7,                   11              ; sign extended shift right by 3
+
+        psraw       xmm5,                   11              ; sign extended shift right by 3
+        punpckhbw   xmm4,                   xmm2            ; axbxcxdx
+
+        punpcklbw   xmm2,                   xmm2            ; exfxgxhx
+        psraw       xmm4,                   11              ; sign extended shift right by 3
+
+        packsswb    xmm5,                   xmm7            ; Filter2 >>=3;
+        psraw       xmm2,                   11              ; sign extended shift right by 3
+
+        packsswb    xmm2,                   xmm4            ; Filter1 >>=3;
+
+        paddsb      xmm6,                   xmm5            ; ps0 =ps0 + Fitler2
+
+        psubsb      xmm3,                   xmm2            ; qs0 =qs0 - Filter1
+        movdqa      xmm7,                   xmm1
+
+        movdqa      xmm4,                   [GLOBAL(s63)]
+        movdqa      xmm5,                   xmm0
+        movdqa      xmm2,                   xmm5
+        paddw       xmm0,                   xmm4            ; Filter 2 (hi) * 9 + 63
+        paddw       xmm1,                   xmm4            ; Filter 2 (lo) * 9 + 63
+        movdqa      xmm4,                   xmm7
+
+        paddw       xmm5,                   xmm5            ; Filter 2 (hi) * 18
+
+        paddw       xmm7,                   xmm7            ; Filter 2 (lo) * 18
+        paddw       xmm5,                   xmm0            ; Filter 2 (hi) * 27 + 63
+
+        paddw       xmm7,                   xmm1            ; Filter 2 (lo) * 27 + 63
+        paddw       xmm2,                   xmm0            ; Filter 2 (hi) * 18 + 63
+        psraw       xmm0,                   7               ; (Filter 2 (hi) * 9 + 63) >> 7
+
+        paddw       xmm4,                   xmm1            ; Filter 2 (lo) * 18 + 63
+        psraw       xmm1,                   7               ; (Filter 2 (lo) * 9 + 63) >> 7
+        psraw       xmm2,                   7               ; (Filter 2 (hi) * 18 + 63) >> 7
+
+        packsswb    xmm0,                   xmm1            ; u1 = vp8_signed_char_clamp((63 + Filter2 * 9)>>7)
+
+        psraw       xmm4,                   7               ; (Filter 2 (lo) * 18 + 63) >> 7
+        psraw       xmm5,                   7               ; (Filter 2 (hi) * 27 + 63) >> 7
+        psraw       xmm7,                   7               ; (Filter 2 (lo) * 27 + 63) >> 7
+
+        packsswb    xmm5,                   xmm7            ; u3 = vp8_signed_char_clamp((63 + Filter2 * 27)>>7)
+        packsswb    xmm2,                   xmm4            ; u2 = vp8_signed_char_clamp((63 + Filter2 * 18)>>7)
+        movdqa      xmm7,                   [GLOBAL(t80)]
+
+%if %1 == 0
+        movdqa      xmm1,                   [rsp+_q1]       ; q1
+        movdqa      xmm4,                   [rsp+_p1]       ; p1
+        lea         rsi,                    [rsi+rcx*2]
+        lea         rdi,                    [rdi+rcx*2]
+
+%elif %1 == 1
+        movdqa      xmm1,                   [rdi]           ; q1
+        movdqa      xmm4,                   [rsi+rax*2]     ; p1
+%elif %1 == 2
+        movdqa      xmm4,                   [rsp+_p1]       ; p1
+        movdqa      xmm1,                   [rsp+_q1]       ; q1
+%endif
+
+        pxor        xmm1,                   xmm7
+        pxor        xmm4,                   xmm7
+
+        psubsb      xmm3,                   xmm5            ; sq = vp8_signed_char_clamp(qs0 - u3)
+        paddsb      xmm6,                   xmm5            ; sp = vp8_signed_char_clamp(ps0 - u3)
+        psubsb      xmm1,                   xmm2            ; sq = vp8_signed_char_clamp(qs1 - u2)
+        paddsb      xmm4,                   xmm2            ; sp = vp8_signed_char_clamp(ps1 - u2)
+
+%if %1 == 1
+        movdqa      xmm2,                   [rdi+rax*4]     ; p2
+        movdqa      xmm5,                   [rdi+rcx]       ; q2
+%else
+        movdqa      xmm2,                   [rsp+_p2]       ; p2
+        movdqa      xmm5,                   [rsp+_q2]       ; q2
+%endif
+
+        pxor        xmm1,                   xmm7            ; *oq1 = sq^0x80;
+        pxor        xmm4,                   xmm7            ; *op1 = sp^0x80;
+        pxor        xmm2,                   xmm7
+        pxor        xmm5,                   xmm7
+        paddsb      xmm2,                   xmm0            ; sp = vp8_signed_char_clamp(ps2 - u)
+        psubsb      xmm5,                   xmm0            ; sq = vp8_signed_char_clamp(qs2 - u)
+        pxor        xmm2,                   xmm7            ; *op2 = sp^0x80;
+        pxor        xmm5,                   xmm7            ; *oq2 = sq^0x80;
+        pxor        xmm3,                   xmm7            ; *oq0 = sq^0x80
+        pxor        xmm6,                   xmm7            ; *oq0 = sp^0x80
+%if %1 == 0
+        movq        [rsi],                  xmm6            ; p0
+        movhps      [rdi],                  xmm6
+        movq        [rsi + rcx],            xmm3            ; q0
+        movhps      [rdi + rcx],            xmm3
+        lea         rdx,                    [rcx + rcx*2]
+        movq        [rsi+rcx*2],            xmm1            ; q1
+        movhps      [rdi+rcx*2],            xmm1
+
+        movq        [rsi + rax],            xmm4            ; p1
+        movhps      [rdi + rax],            xmm4
+
+        movq        [rsi+rax*2],            xmm2            ; p2
+        movhps      [rdi+rax*2],            xmm2
+
+        movq        [rsi+rdx],              xmm5            ; q2
+        movhps      [rdi+rdx],              xmm5
+%elif %1 == 1
+        movdqa      [rdi+rcx],              xmm5            ; q2
+        movdqa      [rdi],                  xmm1            ; q1
+        movdqa      [rsi],                  xmm3            ; q0
+        movdqa      [rsi+rax  ],            xmm6            ; p0
+        movdqa      [rsi+rax*2],            xmm4            ; p1
+        movdqa      [rdi+rax*4],            xmm2            ; p2
+%elif %1 == 2
+        movdqa      [rsp+_p1],              xmm4            ; p1
+        movdqa      [rsp+_p0],              xmm6            ; p0
+        movdqa      [rsp+_q0],              xmm3            ; q0
+        movdqa      [rsp+_q1],              xmm1            ; q1
+%endif
+
+%endmacro
+
+
+;void vp8_mbloop_filter_horizontal_edge_sse2
+;(
+;    unsigned char *src_ptr,
+;    int            src_pixel_step,
+;    const char    *blimit,
+;    const char    *limit,
+;    const char    *thresh,
+;)
+global sym(vp8_mbloop_filter_horizontal_edge_sse2) PRIVATE
+sym(vp8_mbloop_filter_horizontal_edge_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, lf_var_size
+
+        mov         rsi,                    arg(0)            ;src_ptr
+        movsxd      rax,                    dword ptr arg(1)  ;src_pixel_step
+        mov         rdx,                    arg(3)            ;limit
+
+        lea         rdi,                    [rsi+rax]         ; rdi points to row +1 for indirect addressing
+
+        ; calculate breakout conditions and high edge variance
+        LFH_FILTER_AND_HEV_MASK 1
+        ; filter and write back the results
+        MB_FILTER_AND_WRITEBACK 1
+
+    add rsp, lf_var_size
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_mbloop_filter_horizontal_edge_uv_sse2
+;(
+;    unsigned char *u,
+;    int            src_pixel_step,
+;    const char    *blimit,
+;    const char    *limit,
+;    const char    *thresh,
+;    unsigned char *v
+;)
+global sym(vp8_mbloop_filter_horizontal_edge_uv_sse2) PRIVATE
+sym(vp8_mbloop_filter_horizontal_edge_uv_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, lf_var_size
+
+        mov         rsi,                    arg(0)             ; u
+        mov         rdi,                    arg(5)             ; v
+        movsxd      rax,                    dword ptr arg(1)   ; src_pixel_step
+        mov         rcx,                    rax
+        neg         rax                     ; negate pitch to deal with above border
+        mov         rdx,                    arg(3)             ;limit
+
+        lea         rsi,                    [rsi + rcx]
+        lea         rdi,                    [rdi + rcx]
+
+        ; calculate breakout conditions and high edge variance
+        LFH_FILTER_AND_HEV_MASK 0
+        ; filter and write back the results
+        MB_FILTER_AND_WRITEBACK 0
+
+    add rsp, lf_var_size
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+%macro TRANSPOSE_16X8 2
+        movq        xmm4,               [rsi]           ; xx xx xx xx xx xx xx xx 07 06 05 04 03 02 01 00
+        movq        xmm1,               [rdi]           ; xx xx xx xx xx xx xx xx 17 16 15 14 13 12 11 10
+        movq        xmm0,               [rsi+2*rax]     ; xx xx xx xx xx xx xx xx 27 26 25 24 23 22 21 20
+        movq        xmm7,               [rdi+2*rax]     ; xx xx xx xx xx xx xx xx 37 36 35 34 33 32 31 30
+        movq        xmm5,               [rsi+4*rax]     ; xx xx xx xx xx xx xx xx 47 46 45 44 43 42 41 40
+        movq        xmm2,               [rdi+4*rax]     ; xx xx xx xx xx xx xx xx 57 56 55 54 53 52 51 50
+
+        punpcklbw   xmm4,               xmm1            ; 17 07 16 06 15 05 14 04 13 03 12 02 11 01 10 00
+
+        movq        xmm1,               [rdi+2*rcx]     ; xx xx xx xx xx xx xx xx 77 76 75 74 73 72 71 70
+
+        movdqa      xmm3,               xmm4            ; 17 07 16 06 15 05 14 04 13 03 12 02 11 01 10 00
+        punpcklbw   xmm0,               xmm7            ; 37 27 36 36 35 25 34 24 33 23 32 22 31 21 30 20
+
+        movq        xmm7,               [rsi+2*rcx]     ; xx xx xx xx xx xx xx xx 67 66 65 64 63 62 61 60
+
+        punpcklbw   xmm5,               xmm2            ; 57 47 56 46 55 45 54 44 53 43 52 42 51 41 50 40
+%if %1
+        lea         rsi,                [rsi+rax*8]
+        lea         rdi,                [rdi+rax*8]
+%else
+        mov         rsi,                arg(5)          ; v_ptr
+%endif
+
+        movdqa      xmm6,               xmm5            ; 57 47 56 46 55 45 54 44 53 43 52 42 51 41 50 40
+        punpcklbw   xmm7,               xmm1            ; 77 67 76 66 75 65 74 64 73 63 72 62 71 61 70 60
+        punpcklwd   xmm5,               xmm7            ; 73 63 53 43 72 62 52 42 71 61 51 41 70 60 50 40
+        punpckhwd   xmm6,               xmm7            ; 77 67 57 47 76 66 56 46 75 65 55 45 74 64 54 44
+        punpcklwd   xmm3,               xmm0            ; 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00
+
+%if %1 == 0
+        lea         rdi,                [rsi + rax - 4] ; rdi points to row +1 for indirect addressing
+        lea         rsi,                [rsi - 4]
+%endif
+
+        movdqa      xmm2,               xmm3            ; 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00
+        punpckhwd   xmm4,               xmm0            ; 37 27 17 07 36 26 16 06 35 25 15 05 34 24 14 04
+
+        movdqa      xmm7,               xmm4            ; 37 27 17 07 36 26 16 06 35 25 15 05 34 24 14 04
+        punpckhdq   xmm3,               xmm5            ; 73 63 53 43 33 23 13 03 72 62 52 42 32 22 12 02
+
+        punpckhdq   xmm7,               xmm6            ; 77 67 57 47 37 27 17 07 76 66 56 46 36 26 16 06
+
+        punpckldq   xmm4,               xmm6            ; 75 65 55 45 35 25 15 05 74 64 54 44 34 24 14 04
+
+        punpckldq   xmm2,               xmm5            ; 71 61 51 41 31 21 11 01 70 60 50 40 30 20 10 00
+
+        movdqa      [rsp+_t0],          xmm2            ; save to free XMM2
+
+        movq        xmm2,               [rsi]           ; xx xx xx xx xx xx xx xx 87 86 85 84 83 82 81 80
+        movq        xmm6,               [rdi]           ; xx xx xx xx xx xx xx xx 97 96 95 94 93 92 91 90
+        movq        xmm0,               [rsi+2*rax]     ; xx xx xx xx xx xx xx xx a7 a6 a5 a4 a3 a2 a1 a0
+        movq        xmm5,               [rdi+2*rax]     ; xx xx xx xx xx xx xx xx b7 b6 b5 b4 b3 b2 b1 b0
+        movq        xmm1,               [rsi+4*rax]     ; xx xx xx xx xx xx xx xx c7 c6 c5 c4 c3 c2 c1 c0
+
+        punpcklbw   xmm2,               xmm6            ; 97 87 96 86 95 85 94 84 93 83 92 82 91 81 90 80
+
+        movq        xmm6,               [rdi+4*rax]     ; xx xx xx xx xx xx xx xx d7 d6 d5 d4 d3 d2 d1 d0
+
+        punpcklbw   xmm0,               xmm5            ; b7 a7 b6 a6 b5 a5 b4 a4 b3 a3 b2 a2 b1 a1 b0 a0
+
+        movq        xmm5,               [rsi+2*rcx]     ; xx xx xx xx xx xx xx xx e7 e6 e5 e4 e3 e2 e1 e0
+
+        punpcklbw   xmm1,               xmm6            ; d7 c7 d6 c6 d5 c5 d4 c4 d3 c3 d2 c2 d1 e1 d0 c0
+
+        movq        xmm6,               [rdi+2*rcx]     ; xx xx xx xx xx xx xx xx f7 f6 f5 f4 f3 f2 f1 f0
+
+        punpcklbw   xmm5,               xmm6            ; f7 e7 f6 e6 f5 e5 f4 e4 f3 e3 f2 e2 f1 e1 f0 e0
+
+        movdqa      xmm6,               xmm1            ;
+        punpckhwd   xmm6,               xmm5            ; f7 e7 d7 c7 f6 e6 d6 c6 f5 e5 d5 c5 f4 e4 d4 c4
+
+        punpcklwd   xmm1,               xmm5            ; f3 e3 d3 c3 f2 e2 d2 c2 f1 e1 d1 c1 f0 e0 d0 c0
+        movdqa      xmm5,               xmm2            ; 97 87 96 86 95 85 94 84 93 83 92 82 91 81 90 80
+
+        punpcklwd   xmm5,               xmm0            ; b3 a3 93 83 b2 a2 92 82 b1 a1 91 81 b0 a0 90 80
+
+        punpckhwd   xmm2,               xmm0            ; b7 a7 97 87 b6 a6 96 86 b5 a5 95 85 b4 a4 94 84
+
+        movdqa      xmm0,               xmm5
+        punpckldq   xmm0,               xmm1            ; f1 e1 d1 c1 b1 a1 91 81 f0 e0 d0 c0 b0 a0 90 80
+
+        punpckhdq   xmm5,               xmm1            ; f3 e3 d3 c3 b3 a3 93 83 f2 e2 d2 c2 b2 a2 92 82
+        movdqa      xmm1,               xmm2            ; b7 a7 97 87 b6 a6 96 86 b5 a5 95 85 b4 a4 94 84
+
+        punpckldq   xmm1,               xmm6            ; f5 e5 d5 c5 b5 a5 95 85 f4 e4 d4 c4 b4 a4 94 84
+
+        punpckhdq   xmm2,               xmm6            ; f7 e7 d7 c7 b7 a7 97 87 f6 e6 d6 c6 b6 a6 96 86
+        movdqa      xmm6,               xmm7            ; 77 67 57 47 37 27 17 07 76 66 56 46 36 26 16 06
+
+        punpcklqdq  xmm6,               xmm2            ; f6 e6 d6 c6 b6 a6 96 86 76 66 56 46 36 26 16 06
+
+        punpckhqdq  xmm7,               xmm2            ; f7 e7 d7 c7 b7 a7 97 87 77 67 57 47 37 27 17 07
+
+%if %2 == 0
+        movdqa      [rsp+_q3],          xmm7            ; save 7
+        movdqa      [rsp+_q2],          xmm6            ; save 6
+%endif
+        movdqa      xmm2,               xmm3            ; 73 63 53 43 33 23 13 03 72 62 52 42 32 22 12 02
+        punpckhqdq  xmm3,               xmm5            ; f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03
+        punpcklqdq  xmm2,               xmm5            ; f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
+        movdqa      [rsp+_p1],          xmm2            ; save 2
+
+        movdqa      xmm5,               xmm4            ; 75 65 55 45 35 25 15 05 74 64 54 44 34 24 14 04
+        punpcklqdq  xmm4,               xmm1            ; f4 e4 d4 c4 b4 a4 94 84 74 64 54 44 34 24 14 04
+        movdqa      [rsp+_p0],          xmm3            ; save 3
+
+        punpckhqdq  xmm5,               xmm1            ; f5 e5 d5 c5 b5 a5 95 85 75 65 55 45 35 25 15 05
+
+        movdqa      [rsp+_q0],          xmm4            ; save 4
+        movdqa      [rsp+_q1],          xmm5            ; save 5
+        movdqa      xmm1,               [rsp+_t0]
+
+        movdqa      xmm2,               xmm1            ;
+        punpckhqdq  xmm1,               xmm0            ; f1 e1 d1 c1 b1 a1 91 81 71 61 51 41 31 21 11 01
+        punpcklqdq  xmm2,               xmm0            ; f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00
+
+%if %2 == 0
+        movdqa      [rsp+_p2],          xmm1
+        movdqa      [rsp+_p3],          xmm2
+%endif
+
+%endmacro
+
+%macro LFV_FILTER_MASK_HEV_MASK 0
+        movdqa      xmm0,               xmm6            ; q2
+        psubusb     xmm0,               xmm7            ; q2-q3
+
+        psubusb     xmm7,               xmm6            ; q3-q2
+        movdqa      xmm4,               xmm5            ; q1
+
+        por         xmm7,               xmm0            ; abs (q3-q2)
+        psubusb     xmm4,               xmm6            ; q1-q2
+
+        movdqa      xmm0,               xmm1
+        psubusb     xmm6,               xmm5            ; q2-q1
+
+        por         xmm6,               xmm4            ; abs (q2-q1)
+        psubusb     xmm0,               xmm2            ; p2 - p3;
+
+        psubusb     xmm2,               xmm1            ; p3 - p2;
+        por         xmm0,               xmm2            ; abs(p2-p3)
+
+        movdqa      xmm5,               [rsp+_p1]       ; p1
+        pmaxub      xmm0,               xmm7
+
+        movdqa      xmm2,               xmm5            ; p1
+        psubusb     xmm5,               xmm1            ; p1-p2
+        psubusb     xmm1,               xmm2            ; p2-p1
+
+        movdqa      xmm7,               xmm3            ; p0
+        psubusb     xmm7,               xmm2            ; p0-p1
+
+        por         xmm1,               xmm5            ; abs(p2-p1)
+        pmaxub      xmm0,               xmm6
+
+        pmaxub      xmm0,               xmm1
+        movdqa      xmm1,               xmm2            ; p1
+
+        psubusb     xmm2,               xmm3            ; p1-p0
+
+        por         xmm2,               xmm7            ; abs(p1-p0)
+
+        pmaxub      xmm0,               xmm2
+
+        movdqa      xmm5,               [rsp+_q0]       ; q0
+        movdqa      xmm7,               [rsp+_q1]       ; q1
+
+        mov         rdx,                arg(3)          ; limit
+
+        movdqa      xmm6,               xmm5            ; q0
+        movdqa      xmm4,               xmm7            ; q1
+
+        psubusb     xmm5,               xmm7            ; q0-q1
+        psubusb     xmm7,               xmm6            ; q1-q0
+
+        por         xmm7,               xmm5            ; abs(q1-q0)
+
+        pmaxub      xmm0,               xmm7
+
+        psubusb     xmm0,               [rdx]           ; limit
+
+        mov         rdx,                arg(2)          ; blimit
+        movdqa      xmm5,               xmm4            ; q1
+
+        psubusb     xmm5,               xmm1            ; q1-=p1
+        psubusb     xmm1,               xmm4            ; p1-=q1
+
+        por         xmm5,               xmm1            ; abs(p1-q1)
+        movdqa      xmm1,               xmm3            ; p0
+
+        pand        xmm5,               [GLOBAL(tfe)]   ; set lsb of each byte to zero
+        psubusb     xmm1,               xmm6            ; p0-q0
+
+        movdqa      xmm4,               [rdx]           ; blimit
+        mov         rdx,                arg(4)          ; get thresh
+
+        psrlw       xmm5,               1               ; abs(p1-q1)/2
+        psubusb     xmm6,               xmm3            ; q0-p0
+
+        por         xmm1,               xmm6            ; abs(q0-p0)
+        paddusb     xmm1,               xmm1            ; abs(q0-p0)*2
+        movdqa      xmm3,               [rdx]
+
+        paddusb     xmm1,               xmm5            ; abs (p0 - q0) *2 + abs(p1-q1)/2
+        psubusb     xmm2,               xmm3            ; abs(q1 - q0) > thresh
+
+        psubusb     xmm7,               xmm3            ; abs(p1 - p0)> thresh
+
+        psubusb     xmm1,               xmm4            ; abs (p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        por         xmm2,               xmm7            ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
+
+        por         xmm1,               xmm0            ; mask
+        pcmpeqb     xmm2,               xmm0
+
+        pxor        xmm0,               xmm0
+        pcmpeqb     xmm4,               xmm4
+
+        pcmpeqb     xmm1,               xmm0
+        pxor        xmm4,               xmm2
+%endmacro
+
+%macro BV_TRANSPOSE 0
+        ; xmm1 =    f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
+        ; xmm6 =    f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03
+        ; xmm3 =    f4 e4 d4 c4 b4 a4 94 84 74 64 54 44 34 24 14 04
+        ; xmm7 =    f5 e5 d5 c5 b5 a5 95 85 75 65 55 45 35 25 15 05
+        movdqa      xmm2,               xmm1            ; f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
+        punpcklbw   xmm2,               xmm6            ; 73 72 63 62 53 52 43 42 33 32 23 22 13 12 03 02
+
+        movdqa      xmm4,               xmm3            ; f4 e4 d4 c4 b4 a4 94 84 74 64 54 44 34 24 14 04
+        punpckhbw   xmm1,               xmm6            ; f3 f2 e3 e2 d3 d2 c3 c2 b3 b2 a3 a2 93 92 83 82
+
+        punpcklbw   xmm4,               xmm7            ; 75 74 65 64 55 54 45 44 35 34 25 24 15 14 05 04
+
+        punpckhbw   xmm3,               xmm7            ; f5 f4 e5 e4 d5 d4 c5 c4 b5 b4 a5 a4 95 94 85 84
+
+        movdqa      xmm6,               xmm2            ; 73 72 63 62 53 52 43 42 33 32 23 22 13 12 03 02
+        punpcklwd   xmm2,               xmm4            ; 35 34 33 32 25 24 23 22 15 14 13 12 05 04 03 02
+
+        punpckhwd   xmm6,               xmm4            ; 75 74 73 72 65 64 63 62 55 54 53 52 45 44 43 42
+        movdqa      xmm5,               xmm1            ; f3 f2 e3 e2 d3 d2 c3 c2 b3 b2 a3 a2 93 92 83 82
+
+        punpcklwd   xmm1,               xmm3            ; b5 b4 b3 b2 a5 a4 a3 a2 95 94 93 92 85 84 83 82
+
+        punpckhwd   xmm5,               xmm3            ; f5 f4 f3 f2 e5 e4 e3 e2 d5 d4 d3 d2 c5 c4 c3 c2
+        ; xmm2 = 35 34 33 32 25 24 23 22 15 14 13 12 05 04 03 02
+        ; xmm6 = 75 74 73 72 65 64 63 62 55 54 53 52 45 44 43 42
+        ; xmm1 = b5 b4 b3 b2 a5 a4 a3 a2 95 94 93 92 85 84 83 82
+        ; xmm5 = f5 f4 f3 f2 e5 e4 e3 e2 d5 d4 d3 d2 c5 c4 c3 c2
+%endmacro
+
+%macro BV_WRITEBACK 2
+        movd        [rsi+2],            %1
+        movd        [rsi+4*rax+2],      %2
+        psrldq      %1,                 4
+        psrldq      %2,                 4
+        movd        [rdi+2],            %1
+        movd        [rdi+4*rax+2],      %2
+        psrldq      %1,                 4
+        psrldq      %2,                 4
+        movd        [rsi+2*rax+2],      %1
+        movd        [rsi+2*rcx+2],      %2
+        psrldq      %1,                 4
+        psrldq      %2,                 4
+        movd        [rdi+2*rax+2],      %1
+        movd        [rdi+2*rcx+2],      %2
+%endmacro
+
+%if ABI_IS_32BIT
+
+;void vp8_loop_filter_vertical_edge_sse2
+;(
+;    unsigned char *src_ptr,
+;    int            src_pixel_step,
+;    const char    *blimit,
+;    const char    *limit,
+;    const char    *thresh,
+;)
+global sym(vp8_loop_filter_vertical_edge_sse2) PRIVATE
+sym(vp8_loop_filter_vertical_edge_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub             rsp, lf_var_size
+
+        mov         rsi,        arg(0)                  ; src_ptr
+        movsxd      rax,        dword ptr arg(1)        ; src_pixel_step
+
+        lea         rsi,        [rsi - 4]
+        lea         rdi,        [rsi + rax]             ; rdi points to row +1 for indirect addressing
+        lea         rcx,        [rax*2+rax]
+
+        ;transpose 16x8 to 8x16, and store the 8-line result on stack.
+        TRANSPOSE_16X8 1, 1
+
+        ; calculate filter mask and high edge variance
+        LFV_FILTER_MASK_HEV_MASK
+
+        ; start work on filters
+        B_FILTER 2
+
+        ; transpose and write back - only work on q1, q0, p0, p1
+        BV_TRANSPOSE
+        ; store 16-line result
+
+        lea         rdx,        [rax]
+        neg         rdx
+
+        BV_WRITEBACK xmm1, xmm5
+
+        lea         rsi,        [rsi+rdx*8]
+        lea         rdi,        [rdi+rdx*8]
+        BV_WRITEBACK xmm2, xmm6
+
+    add rsp, lf_var_size
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%endif
+
+;void vp8_loop_filter_vertical_edge_uv_sse2
+;(
+;    unsigned char *u,
+;    int            src_pixel_step,
+;    const char    *blimit,
+;    const char    *limit,
+;    const char    *thresh,
+;    unsigned char *v
+;)
+global sym(vp8_loop_filter_vertical_edge_uv_sse2) PRIVATE
+sym(vp8_loop_filter_vertical_edge_uv_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub             rsp, lf_var_size
+
+        mov         rsi,        arg(0)                  ; u_ptr
+        movsxd      rax,        dword ptr arg(1)        ; src_pixel_step
+
+        lea         rsi,        [rsi - 4]
+        lea         rdi,        [rsi + rax]             ; rdi points to row +1 for indirect addressing
+        lea         rcx,        [rax+2*rax]
+
+        ;transpose 16x8 to 8x16, and store the 8-line result on stack.
+        TRANSPOSE_16X8 0, 1
+
+        ; calculate filter mask and high edge variance
+        LFV_FILTER_MASK_HEV_MASK
+
+        ; start work on filters
+        B_FILTER 2
+
+        ; transpose and write back - only work on q1, q0, p0, p1
+        BV_TRANSPOSE
+
+        lea         rdi,        [rsi + rax]             ; rdi points to row +1 for indirect addressing
+
+        ; store 16-line result
+        BV_WRITEBACK xmm1, xmm5
+
+        mov         rsi,        arg(0)                  ; u_ptr
+        lea         rsi,        [rsi - 4]
+        lea         rdi,        [rsi + rax]             ; rdi points to row +1 for indirect addressing
+        BV_WRITEBACK xmm2, xmm6
+
+    add rsp, lf_var_size
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%macro MBV_TRANSPOSE 0
+        movdqa      xmm0,               [rsp+_p3]           ; f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00
+        movdqa      xmm1,               xmm0                ; f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00
+
+        punpcklbw   xmm0,               xmm2                ; 71 70 61 60 51 50 41 40 31 30 21 20 11 10 01 00
+        punpckhbw   xmm1,               xmm2                ; f1 f0 e1 e0 d1 d0 c1 c0 b1 b0 a1 a0 91 90 81 80
+
+        movdqa      xmm7,               [rsp+_p1]           ; f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
+        movdqa      xmm6,               xmm7                ; f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
+
+        punpcklbw   xmm7,               [rsp+_p0]           ; 73 72 63 62 53 52 43 42 33 32 23 22 13 12 03 02
+        punpckhbw   xmm6,               [rsp+_p0]           ; f3 f2 e3 e2 d3 d2 c3 c2 b3 b2 a3 a2 93 92 83 82
+
+        movdqa      xmm3,               xmm0                ; 71 70 61 60 51 50 41 40 31 30 21 20 11 10 01 00
+        punpcklwd   xmm0,               xmm7                ; 33 32 31 30 23 22 21 20 13 12 11 10 03 02 01 00
+
+        punpckhwd   xmm3,               xmm7                ; 73 72 71 70 63 62 61 60 53 52 51 50 43 42 41 40
+        movdqa      xmm4,               xmm1                ; f1 f0 e1 e0 d1 d0 c1 c0 b1 b0 a1 a0 91 90 81 80
+
+        punpcklwd   xmm1,               xmm6                ; b3 b2 b1 b0 a3 a2 a1 a0 93 92 91 90 83 82 81 80
+        punpckhwd   xmm4,               xmm6                ; f3 f2 f1 f0 e3 e2 e1 e0 d3 d2 d1 d0 c3 c2 c1 c0
+
+        movdqa      xmm7,               [rsp+_q0]           ; f4 e4 d4 c4 b4 a4 94 84 74 64 54 44 34 24 14 04
+        punpcklbw   xmm7,               [rsp+_q1]           ; 75 74 65 64 55 54 45 44 35 34 25 24 15 14 05 04
+
+        movdqa      xmm6,               xmm5                ; f6 e6 d6 c6 b6 a6 96 86 76 66 56 46 36 26 16 06
+        punpcklbw   xmm6,               [rsp+_q3]           ; 77 76 67 66 57 56 47 46 37 36 27 26 17 16 07 06
+
+        movdqa      xmm2,               xmm7                ; 75 74 65 64 55 54 45 44 35 34 25 24 15 14 05 04
+        punpcklwd   xmm7,               xmm6                ; 37 36 35 34 27 26 25 24 17 16 15 14 07 06 05 04
+
+        punpckhwd   xmm2,               xmm6                ; 77 76 75 74 67 66 65 64 57 56 55 54 47 46 45 44
+        movdqa      xmm6,               xmm0                ; 33 32 31 30 23 22 21 20 13 12 11 10 03 02 01 00
+
+        punpckldq   xmm0,               xmm7                ; 17 16 15 14 13 12 11 10 07 06 05 04 03 02 01 00
+        punpckhdq   xmm6,               xmm7                ; 37 36 35 34 33 32 31 30 27 26 25 24 23 22 21 20
+%endmacro
+
+%macro MBV_WRITEBACK_1 0
+        movq        [rsi],              xmm0
+        movhps      [rdi],              xmm0
+
+        movq        [rsi+2*rax],        xmm6
+        movhps      [rdi+2*rax],        xmm6
+
+        movdqa      xmm0,               xmm3                ; 73 72 71 70 63 62 61 60 53 52 51 50 43 42 41 40
+        punpckldq   xmm0,               xmm2                ; 57 56 55 54 53 52 51 50 47 46 45 44 43 42 41 40
+        punpckhdq   xmm3,               xmm2                ; 77 76 75 74 73 72 71 70 67 66 65 64 63 62 61 60
+
+        movq        [rsi+4*rax],        xmm0
+        movhps      [rdi+4*rax],        xmm0
+
+        movq        [rsi+2*rcx],        xmm3
+        movhps      [rdi+2*rcx],        xmm3
+
+        movdqa      xmm7,               [rsp+_q0]           ; f4 e4 d4 c4 b4 a4 94 84 74 64 54 44 34 24 14 04
+        punpckhbw   xmm7,               [rsp+_q1]           ; f5 f4 e5 e4 d5 d4 c5 c4 b5 b4 a5 a4 95 94 85 84
+        punpckhbw   xmm5,               [rsp+_q3]           ; f7 f6 e7 e6 d7 d6 c7 c6 b7 b6 a7 a6 97 96 87 86
+
+        movdqa      xmm0,               xmm7
+        punpcklwd   xmm0,               xmm5                ; b7 b6 b4 b4 a7 a6 a5 a4 97 96 95 94 87 86 85 84
+        punpckhwd   xmm7,               xmm5                ; f7 f6 f5 f4 e7 e6 e5 e4 d7 d6 d5 d4 c7 c6 c5 c4
+
+        movdqa      xmm5,               xmm1                ; b3 b2 b1 b0 a3 a2 a1 a0 93 92 91 90 83 82 81 80
+        punpckldq   xmm1,               xmm0                ; 97 96 95 94 93 92 91 90 87 86 85 83 84 82 81 80
+        punpckhdq   xmm5,               xmm0                ; b7 b6 b5 b4 b3 b2 b1 b0 a7 a6 a5 a4 a3 a2 a1 a0
+%endmacro
+
+%macro MBV_WRITEBACK_2 0
+        movq        [rsi],              xmm1
+        movhps      [rdi],              xmm1
+
+        movq        [rsi+2*rax],        xmm5
+        movhps      [rdi+2*rax],        xmm5
+
+        movdqa      xmm1,               xmm4                ; f3 f2 f1 f0 e3 e2 e1 e0 d3 d2 d1 d0 c3 c2 c1 c0
+        punpckldq   xmm1,               xmm7                ; d7 d6 d5 d4 d3 d2 d1 d0 c7 c6 c5 c4 c3 c2 c1 c0
+        punpckhdq   xmm4,               xmm7                ; f7 f6 f4 f4 f3 f2 f1 f0 e7 e6 e5 e4 e3 e2 e1 e0
+
+        movq        [rsi+4*rax],        xmm1
+        movhps      [rdi+4*rax],        xmm1
+
+        movq        [rsi+2*rcx],        xmm4
+        movhps      [rdi+2*rcx],        xmm4
+%endmacro
+
+
+;void vp8_mbloop_filter_vertical_edge_sse2
+;(
+;    unsigned char *src_ptr,
+;    int            src_pixel_step,
+;    const char    *blimit,
+;    const char    *limit,
+;    const char    *thresh,
+;)
+global sym(vp8_mbloop_filter_vertical_edge_sse2) PRIVATE
+sym(vp8_mbloop_filter_vertical_edge_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub          rsp, lf_var_size
+
+        mov         rsi,                arg(0)              ; src_ptr
+        movsxd      rax,                dword ptr arg(1)    ; src_pixel_step
+
+        lea         rsi,                [rsi - 4]
+        lea         rdi,                [rsi + rax]         ; rdi points to row +1 for indirect addressing
+        lea         rcx,                [rax*2+rax]
+
+        ; Transpose
+        TRANSPOSE_16X8 1, 0
+
+        ; calculate filter mask and high edge variance
+        LFV_FILTER_MASK_HEV_MASK
+
+        neg         rax
+        ; start work on filters
+        MB_FILTER_AND_WRITEBACK 2
+
+        lea         rsi,                [rsi+rax*8]
+        lea         rdi,                [rdi+rax*8]
+
+        ; transpose and write back
+        MBV_TRANSPOSE
+
+        neg         rax
+
+        MBV_WRITEBACK_1
+
+
+        lea         rsi,                [rsi+rax*8]
+        lea         rdi,                [rdi+rax*8]
+        MBV_WRITEBACK_2
+
+    add rsp, lf_var_size
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_mbloop_filter_vertical_edge_uv_sse2
+;(
+;    unsigned char *u,
+;    int            src_pixel_step,
+;    const char    *blimit,
+;    const char    *limit,
+;    const char    *thresh,
+;    unsigned char *v
+;)
+global sym(vp8_mbloop_filter_vertical_edge_uv_sse2) PRIVATE
+sym(vp8_mbloop_filter_vertical_edge_uv_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub          rsp, lf_var_size
+
+        mov         rsi,                arg(0)              ; u_ptr
+        movsxd      rax,                dword ptr arg(1)    ; src_pixel_step
+
+        lea         rsi,                [rsi - 4]
+        lea         rdi,                [rsi + rax]         ; rdi points to row +1 for indirect addressing
+        lea         rcx,                [rax+2*rax]
+
+        ; Transpose
+        TRANSPOSE_16X8 0, 0
+
+        ; calculate filter mask and high edge variance
+        LFV_FILTER_MASK_HEV_MASK
+
+        ; start work on filters
+        MB_FILTER_AND_WRITEBACK 2
+
+        ; transpose and write back
+        MBV_TRANSPOSE
+
+        mov         rsi,                arg(0)             ;u_ptr
+        lea         rsi,                [rsi - 4]
+        lea         rdi,                [rsi + rax]
+        MBV_WRITEBACK_1
+        mov         rsi,                arg(5)             ;v_ptr
+        lea         rsi,                [rsi - 4]
+        lea         rdi,                [rsi + rax]
+        MBV_WRITEBACK_2
+
+    add rsp, lf_var_size
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_loop_filter_simple_horizontal_edge_sse2
+;(
+;    unsigned char *src_ptr,
+;    int  src_pixel_step,
+;    const char *blimit,
+;)
+global sym(vp8_loop_filter_simple_horizontal_edge_sse2) PRIVATE
+sym(vp8_loop_filter_simple_horizontal_edge_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 3
+    SAVE_XMM 7
+    GET_GOT     rbx
+    ; end prolog
+
+        mov         rcx, arg(0)             ;src_ptr
+        movsxd      rax, dword ptr arg(1)   ;src_pixel_step     ; destination pitch?
+        movdqa      xmm6, [GLOBAL(tfe)]
+        lea         rdx, [rcx + rax]
+        neg         rax
+
+        ; calculate mask
+        movdqa      xmm0, [rdx]             ; q1
+        mov         rdx, arg(2)             ;blimit
+        movdqa      xmm1, [rcx+2*rax]       ; p1
+
+        movdqa      xmm2, xmm1
+        movdqa      xmm3, xmm0
+
+        psubusb     xmm0, xmm1              ; q1-=p1
+        psubusb     xmm1, xmm3              ; p1-=q1
+        por         xmm1, xmm0              ; abs(p1-q1)
+        pand        xmm1, xmm6              ; set lsb of each byte to zero
+        psrlw       xmm1, 1                 ; abs(p1-q1)/2
+
+        movdqa      xmm7, XMMWORD PTR [rdx]
+
+        movdqa      xmm5, [rcx+rax]         ; p0
+        movdqa      xmm4, [rcx]             ; q0
+        movdqa      xmm0, xmm4              ; q0
+        movdqa      xmm6, xmm5              ; p0
+        psubusb     xmm5, xmm4              ; p0-=q0
+        psubusb     xmm4, xmm6              ; q0-=p0
+        por         xmm5, xmm4              ; abs(p0 - q0)
+
+        movdqa      xmm4, [GLOBAL(t80)]
+
+        paddusb     xmm5, xmm5              ; abs(p0-q0)*2
+        paddusb     xmm5, xmm1              ; abs (p0 - q0) *2 + abs(p1-q1)/2
+        psubusb     xmm5, xmm7              ; abs(p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        pxor        xmm7, xmm7
+        pcmpeqb     xmm5, xmm7
+
+
+        ; start work on filters
+        pxor        xmm2, xmm4     ; p1 offset to convert to signed values
+        pxor        xmm3, xmm4     ; q1 offset to convert to signed values
+        psubsb      xmm2, xmm3              ; p1 - q1
+
+        pxor        xmm6, xmm4     ; offset to convert to signed values
+        pxor        xmm0, xmm4     ; offset to convert to signed values
+        movdqa      xmm3, xmm0              ; q0
+        psubsb      xmm0, xmm6              ; q0 - p0
+        paddsb      xmm2, xmm0              ; p1 - q1 + 1 * (q0 - p0)
+        paddsb      xmm2, xmm0              ; p1 - q1 + 2 * (q0 - p0)
+        paddsb      xmm2, xmm0              ; p1 - q1 + 3 * (q0 - p0)
+        pand        xmm5, xmm2              ; mask filter values we don't care about
+
+        movdqa      xmm0, xmm5
+        paddsb      xmm5,        [GLOBAL(t3)]                  ;  3* (q0 - p0) + (p1 - q1) + 4
+        paddsb      xmm0,        [GLOBAL(t4)]                  ; +3 instead of +4
+
+        movdqa      xmm1, [GLOBAL(te0)]
+        movdqa      xmm2, [GLOBAL(t1f)]
+
+;        pxor        xmm7, xmm7
+        pcmpgtb     xmm7, xmm0              ;save sign
+        pand        xmm7, xmm1              ;preserve the upper 3 bits
+        psrlw       xmm0, 3
+        pand        xmm0, xmm2              ;clear out upper 3 bits
+        por         xmm0, xmm7              ;add sign
+        psubsb      xmm3, xmm0              ; q0-= q0sz add
+
+        pxor        xmm7, xmm7
+        pcmpgtb     xmm7, xmm5              ;save sign
+        pand        xmm7, xmm1              ;preserve the upper 3 bits
+        psrlw       xmm5, 3
+        pand        xmm5, xmm2              ;clear out upper 3 bits
+        por         xmm5, xmm7              ;add sign
+        paddsb      xmm6, xmm5              ; p0+= p0 add
+
+        pxor        xmm3, xmm4     ; unoffset
+        movdqa      [rcx], xmm3             ; write back
+
+        pxor        xmm6, xmm4     ; unoffset
+        movdqa      [rcx+rax], xmm6         ; write back
+
+    ; begin epilog
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_loop_filter_simple_vertical_edge_sse2
+;(
+;    unsigned char *src_ptr,
+;    int  src_pixel_step,
+;    const char *blimit,
+;)
+global sym(vp8_loop_filter_simple_vertical_edge_sse2) PRIVATE
+sym(vp8_loop_filter_simple_vertical_edge_sse2):
+    push        rbp         ; save old base pointer value.
+    mov         rbp, rsp    ; set new base pointer value.
+    SHADOW_ARGS_TO_STACK 3
+    SAVE_XMM 7
+    GET_GOT     rbx         ; save callee-saved reg
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 32                         ; reserve 32 bytes
+    %define t0  [rsp + 0]    ;__declspec(align(16)) char t0[16];
+    %define t1  [rsp + 16]   ;__declspec(align(16)) char t1[16];
+
+        mov         rsi, arg(0) ;src_ptr
+        movsxd      rax, dword ptr arg(1) ;src_pixel_step     ; destination pitch?
+
+        lea         rsi,        [rsi - 2 ]
+        lea         rdi,        [rsi + rax]
+        lea         rdx,        [rsi + rax*4]
+        lea         rcx,        [rdx + rax]
+
+        movd        xmm0,       [rsi]                   ; (high 96 bits unused) 03 02 01 00
+        movd        xmm1,       [rdx]                   ; (high 96 bits unused) 43 42 41 40
+        movd        xmm2,       [rdi]                   ; 13 12 11 10
+        movd        xmm3,       [rcx]                   ; 53 52 51 50
+        punpckldq   xmm0,       xmm1                    ; (high 64 bits unused) 43 42 41 40 03 02 01 00
+        punpckldq   xmm2,       xmm3                    ; 53 52 51 50 13 12 11 10
+
+        movd        xmm4,       [rsi + rax*2]           ; 23 22 21 20
+        movd        xmm5,       [rdx + rax*2]           ; 63 62 61 60
+        movd        xmm6,       [rdi + rax*2]           ; 33 32 31 30
+        movd        xmm7,       [rcx + rax*2]           ; 73 72 71 70
+        punpckldq   xmm4,       xmm5                    ; 63 62 61 60 23 22 21 20
+        punpckldq   xmm6,       xmm7                    ; 73 72 71 70 33 32 31 30
+
+        punpcklbw   xmm0,       xmm2                    ; 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00
+        punpcklbw   xmm4,       xmm6                    ; 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20
+
+        movdqa      xmm1,       xmm0
+        punpcklwd   xmm0,       xmm4                    ; 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00
+        punpckhwd   xmm1,       xmm4                    ; 73 63 53 43 72 62 52 42 71 61 51 41 70 60 50 40
+
+        movdqa      xmm2,       xmm0
+        punpckldq   xmm0,       xmm1                    ; 71 61 51 41 31 21 11 01 70 60 50 40 30 20 10 00
+        punpckhdq   xmm2,       xmm1                    ; 73 63 53 43 33 23 13 03 72 62 52 42 32 22 12 02
+
+        lea         rsi,        [rsi + rax*8]
+        lea         rdi,        [rsi + rax]
+        lea         rdx,        [rsi + rax*4]
+        lea         rcx,        [rdx + rax]
+
+        movd        xmm4,       [rsi]                   ; 83 82 81 80
+        movd        xmm1,       [rdx]                   ; c3 c2 c1 c0
+        movd        xmm6,       [rdi]                   ; 93 92 91 90
+        movd        xmm3,       [rcx]                   ; d3 d2 d1 d0
+        punpckldq   xmm4,       xmm1                    ; c3 c2 c1 c0 83 82 81 80
+        punpckldq   xmm6,       xmm3                    ; d3 d2 d1 d0 93 92 91 90
+
+        movd        xmm1,       [rsi + rax*2]           ; a3 a2 a1 a0
+        movd        xmm5,       [rdx + rax*2]           ; e3 e2 e1 e0
+        movd        xmm3,       [rdi + rax*2]           ; b3 b2 b1 b0
+        movd        xmm7,       [rcx + rax*2]           ; f3 f2 f1 f0
+        punpckldq   xmm1,       xmm5                    ; e3 e2 e1 e0 a3 a2 a1 a0
+        punpckldq   xmm3,       xmm7                    ; f3 f2 f1 f0 b3 b2 b1 b0
+
+        punpcklbw   xmm4,       xmm6                    ; d3 c3 d2 c2 d1 c1 d0 c0 93 83 92 82 91 81 90 80
+        punpcklbw   xmm1,       xmm3                    ; f3 e3 f2 e2 f1 e1 f0 e0 b3 a3 b2 a2 b1 a1 b0 a0
+
+        movdqa      xmm7,       xmm4
+        punpcklwd   xmm4,       xmm1                    ; b3 a3 93 83 b2 a2 92 82 b1 a1 91 81 b0 a0 90 80
+        punpckhwd   xmm7,       xmm1                    ; f3 e3 d3 c3 f2 e2 d2 c2 f1 e1 d1 c1 f0 e0 d0 c0
+
+        movdqa      xmm6,       xmm4
+        punpckldq   xmm4,       xmm7                    ; f1 e1 d1 c1 b1 a1 91 81 f0 e0 d0 c0 b0 a0 90 80
+        punpckhdq   xmm6,       xmm7                    ; f3 e3 d3 c3 b3 a3 93 83 f2 e2 d2 c2 b2 a2 92 82
+
+        movdqa      xmm1,       xmm0
+        movdqa      xmm3,       xmm2
+
+        punpcklqdq  xmm0,       xmm4                    ; p1  f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00
+        punpckhqdq  xmm1,       xmm4                    ; p0  f1 e1 d1 c1 b1 a1 91 81 71 61 51 41 31 21 11 01
+        punpcklqdq  xmm2,       xmm6                    ; q0  f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
+        punpckhqdq  xmm3,       xmm6                    ; q1  f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03
+
+        mov         rdx,        arg(2)                          ;blimit
+
+        ; calculate mask
+        movdqa      xmm6,       xmm0                            ; p1
+        movdqa      xmm7,       xmm3                            ; q1
+        psubusb     xmm7,       xmm0                            ; q1-=p1
+        psubusb     xmm6,       xmm3                            ; p1-=q1
+        por         xmm6,       xmm7                            ; abs(p1-q1)
+        pand        xmm6,       [GLOBAL(tfe)]                   ; set lsb of each byte to zero
+        psrlw       xmm6,       1                               ; abs(p1-q1)/2
+
+        movdqa      xmm7, [rdx]
+
+        movdqa      xmm5,       xmm1                            ; p0
+        movdqa      xmm4,       xmm2                            ; q0
+        psubusb     xmm5,       xmm2                            ; p0-=q0
+        psubusb     xmm4,       xmm1                            ; q0-=p0
+        por         xmm5,       xmm4                            ; abs(p0 - q0)
+        paddusb     xmm5,       xmm5                            ; abs(p0-q0)*2
+        paddusb     xmm5,       xmm6                            ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+        movdqa      xmm4, [GLOBAL(t80)]
+
+        psubusb     xmm5,        xmm7                           ; abs(p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        pxor        xmm7,        xmm7
+        pcmpeqb     xmm5,        xmm7                           ; mm5 = mask
+
+        ; start work on filters
+        movdqa        t0,        xmm0
+        movdqa        t1,        xmm3
+
+        pxor        xmm0,        xmm4                  ; p1 offset to convert to signed values
+        pxor        xmm3,        xmm4                  ; q1 offset to convert to signed values
+        psubsb      xmm0,        xmm3                           ; p1 - q1
+
+        pxor        xmm1,        xmm4                  ; offset to convert to signed values
+        pxor        xmm2,        xmm4                  ; offset to convert to signed values
+
+        movdqa      xmm3,        xmm2                           ; offseted ; q0
+        psubsb      xmm2,        xmm1                           ; q0 - p0
+        paddsb      xmm0,        xmm2                           ; p1 - q1 + 1 * (q0 - p0)
+        paddsb      xmm0,        xmm2                           ; p1 - q1 + 2 * (q0 - p0)
+        paddsb      xmm0,        xmm2                           ; p1 - q1 + 3 * (q0 - p0)
+        pand        xmm5,        xmm0                           ; mask filter values we don't care about
+
+        movdqa      xmm0, xmm5
+        paddsb      xmm5,        [GLOBAL(t3)]                  ;  3* (q0 - p0) + (p1 - q1) + 4
+        paddsb      xmm0,        [GLOBAL(t4)]                  ; +3 instead of +4
+
+        movdqa  xmm6, [GLOBAL(te0)]
+        movdqa  xmm2, [GLOBAL(t1f)]
+
+;        pxor        xmm7, xmm7
+        pcmpgtb     xmm7, xmm0              ;save sign
+        pand        xmm7, xmm6              ;preserve the upper 3 bits
+        psrlw       xmm0, 3
+        pand        xmm0, xmm2              ;clear out upper 3 bits
+        por         xmm0, xmm7              ;add sign
+        psubsb      xmm3, xmm0              ; q0-= q0sz add
+
+        pxor        xmm7, xmm7
+        pcmpgtb     xmm7, xmm5              ;save sign
+        pand        xmm7, xmm6              ;preserve the upper 3 bits
+        psrlw       xmm5, 3
+        pand        xmm5, xmm2              ;clear out upper 3 bits
+        por         xmm5, xmm7              ;add sign
+        paddsb      xmm1, xmm5              ; p0+= p0 add
+
+        pxor        xmm3,        xmm4                  ; unoffset   q0
+        pxor        xmm1,        xmm4                  ; unoffset   p0
+
+        movdqa      xmm0,        t0                             ; p1
+        movdqa      xmm4,        t1                             ; q1
+
+        ; write out order: xmm0 xmm2 xmm1 xmm3
+        lea         rdx,        [rsi + rax*4]
+
+        ; transpose back to write out
+        ; p1  f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00
+        ; p0  f1 e1 d1 c1 b1 a1 91 81 71 61 51 41 31 21 11 01
+        ; q0  f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
+        ; q1  f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03
+        movdqa      xmm6,       xmm0
+        punpcklbw   xmm0,       xmm1                               ; 71 70 61 60 51 50 41 40 31 30 21 20 11 10 01 00
+        punpckhbw   xmm6,       xmm1                               ; f1 f0 e1 e0 d1 d0 c1 c0 b1 b0 a1 a0 91 90 81 80
+
+        movdqa      xmm5,       xmm3
+        punpcklbw   xmm3,       xmm4                               ; 73 72 63 62 53 52 43 42 33 32 23 22 13 12 03 02
+        punpckhbw   xmm5,       xmm4                               ; f3 f2 e3 e2 d3 d2 c3 c2 b3 b2 a3 a2 93 92 83 82
+
+        movdqa      xmm2,       xmm0
+        punpcklwd   xmm0,       xmm3                               ; 33 32 31 30 23 22 21 20 13 12 11 10 03 02 01 00
+        punpckhwd   xmm2,       xmm3                               ; 73 72 71 70 63 62 61 60 53 52 51 50 43 42 41 40
+
+        movdqa      xmm3,       xmm6
+        punpcklwd   xmm6,       xmm5                               ; b3 b2 b1 b0 a3 a2 a1 a0 93 92 91 90 83 82 81 80
+        punpckhwd   xmm3,       xmm5                               ; f3 f2 f1 f0 e3 e2 e1 e0 d3 d2 d1 d0 c3 c2 c1 c0
+
+        movd        [rsi],      xmm6                               ; write the second 8-line result
+        movd        [rdx],      xmm3
+        psrldq      xmm6,       4
+        psrldq      xmm3,       4
+        movd        [rdi],      xmm6
+        movd        [rcx],      xmm3
+        psrldq      xmm6,       4
+        psrldq      xmm3,       4
+        movd        [rsi + rax*2], xmm6
+        movd        [rdx + rax*2], xmm3
+        psrldq      xmm6,       4
+        psrldq      xmm3,       4
+        movd        [rdi + rax*2], xmm6
+        movd        [rcx + rax*2], xmm3
+
+        neg         rax
+        lea         rsi,        [rsi + rax*8]
+        neg         rax
+        lea         rdi,        [rsi + rax]
+        lea         rdx,        [rsi + rax*4]
+        lea         rcx,        [rdx + rax]
+
+        movd        [rsi],      xmm0                                ; write the first 8-line result
+        movd        [rdx],      xmm2
+        psrldq      xmm0,       4
+        psrldq      xmm2,       4
+        movd        [rdi],      xmm0
+        movd        [rcx],      xmm2
+        psrldq      xmm0,       4
+        psrldq      xmm2,       4
+        movd        [rsi + rax*2], xmm0
+        movd        [rdx + rax*2], xmm2
+        psrldq      xmm0,       4
+        psrldq      xmm2,       4
+        movd        [rdi + rax*2], xmm0
+        movd        [rcx + rax*2], xmm2
+
+    add rsp, 32
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+tfe:
+    times 16 db 0xfe
+align 16
+t80:
+    times 16 db 0x80
+align 16
+t1s:
+    times 16 db 0x01
+align 16
+t3:
+    times 16 db 0x03
+align 16
+t4:
+    times 16 db 0x04
+align 16
+ones:
+    times 8 dw 0x0001
+align 16
+s9:
+    times 8 dw 0x0900
+align 16
+s63:
+    times 8 dw 0x003f
+align 16
+te0:
+    times 16 db 0xe0
+align 16
+t1f:
+    times 16 db 0x1f
diff --git a/libvpx/libvpx/vp8/common/x86/loopfilter_x86.c b/libvpx/libvpx/vp8/common/x86/loopfilter_x86.c
new file mode 100644
index 0000000..6586004
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/loopfilter_x86.c
@@ -0,0 +1,198 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8/common/loopfilter.h"
+
+#define prototype_loopfilter(sym) \
+    void sym(unsigned char *src, int pitch, const unsigned char *blimit,\
+             const unsigned char *limit, const unsigned char *thresh, int count)
+
+#define prototype_loopfilter_nc(sym) \
+    void sym(unsigned char *src, int pitch, const unsigned char *blimit,\
+             const unsigned char *limit, const unsigned char *thresh)
+
+#define prototype_simple_loopfilter(sym) \
+    void sym(unsigned char *y, int ystride, const unsigned char *blimit)
+
+prototype_loopfilter(vp8_mbloop_filter_vertical_edge_mmx);
+prototype_loopfilter(vp8_mbloop_filter_horizontal_edge_mmx);
+prototype_loopfilter(vp8_loop_filter_vertical_edge_mmx);
+prototype_loopfilter(vp8_loop_filter_horizontal_edge_mmx);
+prototype_simple_loopfilter(vp8_loop_filter_simple_horizontal_edge_mmx);
+prototype_simple_loopfilter(vp8_loop_filter_simple_vertical_edge_mmx);
+
+#if HAVE_SSE2 && ARCH_X86_64
+prototype_loopfilter(vp8_loop_filter_bv_y_sse2);
+prototype_loopfilter(vp8_loop_filter_bh_y_sse2);
+#else
+prototype_loopfilter_nc(vp8_loop_filter_vertical_edge_sse2);
+prototype_loopfilter_nc(vp8_loop_filter_horizontal_edge_sse2);
+#endif
+prototype_loopfilter_nc(vp8_mbloop_filter_vertical_edge_sse2);
+prototype_loopfilter_nc(vp8_mbloop_filter_horizontal_edge_sse2);
+
+extern loop_filter_uvfunction vp8_loop_filter_horizontal_edge_uv_sse2;
+extern loop_filter_uvfunction vp8_loop_filter_vertical_edge_uv_sse2;
+extern loop_filter_uvfunction vp8_mbloop_filter_horizontal_edge_uv_sse2;
+extern loop_filter_uvfunction vp8_mbloop_filter_vertical_edge_uv_sse2;
+
+#if HAVE_MMX
+/* Horizontal MB filtering */
+void vp8_loop_filter_mbh_mmx(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                             int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    vp8_mbloop_filter_horizontal_edge_mmx(y_ptr, y_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_mbloop_filter_horizontal_edge_mmx(u_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_mbloop_filter_horizontal_edge_mmx(v_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+}
+
+
+/* Vertical MB Filtering */
+void vp8_loop_filter_mbv_mmx(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                             int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    vp8_mbloop_filter_vertical_edge_mmx(y_ptr, y_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_mbloop_filter_vertical_edge_mmx(u_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_mbloop_filter_vertical_edge_mmx(v_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+}
+
+
+/* Horizontal B Filtering */
+void vp8_loop_filter_bh_mmx(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                            int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    vp8_loop_filter_horizontal_edge_mmx(y_ptr + 4 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_horizontal_edge_mmx(y_ptr + 8 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_horizontal_edge_mmx(y_ptr + 12 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_loop_filter_horizontal_edge_mmx(u_ptr + 4 * uv_stride, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_loop_filter_horizontal_edge_mmx(v_ptr + 4 * uv_stride, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+}
+
+
+void vp8_loop_filter_bhs_mmx(unsigned char *y_ptr, int y_stride, const unsigned char *blimit)
+{
+    vp8_loop_filter_simple_horizontal_edge_mmx(y_ptr + 4 * y_stride, y_stride, blimit);
+    vp8_loop_filter_simple_horizontal_edge_mmx(y_ptr + 8 * y_stride, y_stride, blimit);
+    vp8_loop_filter_simple_horizontal_edge_mmx(y_ptr + 12 * y_stride, y_stride, blimit);
+}
+
+
+/* Vertical B Filtering */
+void vp8_loop_filter_bv_mmx(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                            int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    vp8_loop_filter_vertical_edge_mmx(y_ptr + 4, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_vertical_edge_mmx(y_ptr + 8, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+    vp8_loop_filter_vertical_edge_mmx(y_ptr + 12, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+
+    if (u_ptr)
+        vp8_loop_filter_vertical_edge_mmx(u_ptr + 4, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+
+    if (v_ptr)
+        vp8_loop_filter_vertical_edge_mmx(v_ptr + 4, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1);
+}
+
+
+void vp8_loop_filter_bvs_mmx(unsigned char *y_ptr, int y_stride, const unsigned char *blimit)
+{
+    vp8_loop_filter_simple_vertical_edge_mmx(y_ptr + 4, y_stride, blimit);
+    vp8_loop_filter_simple_vertical_edge_mmx(y_ptr + 8, y_stride, blimit);
+    vp8_loop_filter_simple_vertical_edge_mmx(y_ptr + 12, y_stride, blimit);
+}
+#endif
+
+
+/* Horizontal MB filtering */
+#if HAVE_SSE2
+void vp8_loop_filter_mbh_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                              int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    vp8_mbloop_filter_horizontal_edge_sse2(y_ptr, y_stride, lfi->mblim, lfi->lim, lfi->hev_thr);
+
+    if (u_ptr)
+        vp8_mbloop_filter_horizontal_edge_uv_sse2(u_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, v_ptr);
+}
+
+
+/* Vertical MB Filtering */
+void vp8_loop_filter_mbv_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                              int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+    vp8_mbloop_filter_vertical_edge_sse2(y_ptr, y_stride, lfi->mblim, lfi->lim, lfi->hev_thr);
+
+    if (u_ptr)
+        vp8_mbloop_filter_vertical_edge_uv_sse2(u_ptr, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, v_ptr);
+}
+
+
+/* Horizontal B Filtering */
+void vp8_loop_filter_bh_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                             int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+#if ARCH_X86_64
+    vp8_loop_filter_bh_y_sse2(y_ptr, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+#else
+    vp8_loop_filter_horizontal_edge_sse2(y_ptr + 4 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr);
+    vp8_loop_filter_horizontal_edge_sse2(y_ptr + 8 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr);
+    vp8_loop_filter_horizontal_edge_sse2(y_ptr + 12 * y_stride, y_stride, lfi->blim, lfi->lim, lfi->hev_thr);
+#endif
+
+    if (u_ptr)
+        vp8_loop_filter_horizontal_edge_uv_sse2(u_ptr + 4 * uv_stride, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, v_ptr + 4 * uv_stride);
+}
+
+
+void vp8_loop_filter_bhs_sse2(unsigned char *y_ptr, int y_stride, const unsigned char *blimit)
+{
+    vp8_loop_filter_simple_horizontal_edge_sse2(y_ptr + 4 * y_stride, y_stride, blimit);
+    vp8_loop_filter_simple_horizontal_edge_sse2(y_ptr + 8 * y_stride, y_stride, blimit);
+    vp8_loop_filter_simple_horizontal_edge_sse2(y_ptr + 12 * y_stride, y_stride, blimit);
+}
+
+
+/* Vertical B Filtering */
+void vp8_loop_filter_bv_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr,
+                             int y_stride, int uv_stride, loop_filter_info *lfi)
+{
+#if ARCH_X86_64
+    vp8_loop_filter_bv_y_sse2(y_ptr, y_stride, lfi->blim, lfi->lim, lfi->hev_thr, 2);
+#else
+    vp8_loop_filter_vertical_edge_sse2(y_ptr + 4, y_stride, lfi->blim, lfi->lim, lfi->hev_thr);
+    vp8_loop_filter_vertical_edge_sse2(y_ptr + 8, y_stride, lfi->blim, lfi->lim, lfi->hev_thr);
+    vp8_loop_filter_vertical_edge_sse2(y_ptr + 12, y_stride, lfi->blim, lfi->lim, lfi->hev_thr);
+#endif
+
+    if (u_ptr)
+        vp8_loop_filter_vertical_edge_uv_sse2(u_ptr + 4, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, v_ptr + 4);
+}
+
+
+void vp8_loop_filter_bvs_sse2(unsigned char *y_ptr, int y_stride, const unsigned char *blimit)
+{
+    vp8_loop_filter_simple_vertical_edge_sse2(y_ptr + 4, y_stride, blimit);
+    vp8_loop_filter_simple_vertical_edge_sse2(y_ptr + 8, y_stride, blimit);
+    vp8_loop_filter_simple_vertical_edge_sse2(y_ptr + 12, y_stride, blimit);
+}
+
+#endif
diff --git a/libvpx/libvpx/vp8/common/x86/mfqe_sse2.asm b/libvpx/libvpx/vp8/common/x86/mfqe_sse2.asm
new file mode 100644
index 0000000..a8a7f56
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/mfqe_sse2.asm
@@ -0,0 +1,287 @@
+;
+;  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vp8_filter_by_weight16x16_sse2
+;(
+;    unsigned char *src,
+;    int            src_stride,
+;    unsigned char *dst,
+;    int            dst_stride,
+;    int            src_weight
+;)
+global sym(vp8_filter_by_weight16x16_sse2) PRIVATE
+sym(vp8_filter_by_weight16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movd        xmm0, arg(4)                ; src_weight
+    pshuflw     xmm0, xmm0, 0x0             ; replicate to all low words
+    punpcklqdq  xmm0, xmm0                  ; replicate to all hi words
+
+    movdqa      xmm1, [GLOBAL(tMFQE)]
+    psubw       xmm1, xmm0                  ; dst_weight
+
+    mov         rax, arg(0)                 ; src
+    mov         rsi, arg(1)                 ; src_stride
+    mov         rdx, arg(2)                 ; dst
+    mov         rdi, arg(3)                 ; dst_stride
+
+    mov         rcx, 16                     ; loop count
+    pxor        xmm6, xmm6
+
+.combine
+    movdqa      xmm2, [rax]
+    movdqa      xmm4, [rdx]
+    add         rax, rsi
+
+    ; src * src_weight
+    movdqa      xmm3, xmm2
+    punpcklbw   xmm2, xmm6
+    punpckhbw   xmm3, xmm6
+    pmullw      xmm2, xmm0
+    pmullw      xmm3, xmm0
+
+    ; dst * dst_weight
+    movdqa      xmm5, xmm4
+    punpcklbw   xmm4, xmm6
+    punpckhbw   xmm5, xmm6
+    pmullw      xmm4, xmm1
+    pmullw      xmm5, xmm1
+
+    ; sum, round and shift
+    paddw       xmm2, xmm4
+    paddw       xmm3, xmm5
+    paddw       xmm2, [GLOBAL(tMFQE_round)]
+    paddw       xmm3, [GLOBAL(tMFQE_round)]
+    psrlw       xmm2, 4
+    psrlw       xmm3, 4
+
+    packuswb    xmm2, xmm3
+    movdqa      [rdx], xmm2
+    add         rdx, rdi
+
+    dec         rcx
+    jnz         .combine
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+
+    ret
+
+;void vp8_filter_by_weight8x8_sse2
+;(
+;    unsigned char *src,
+;    int            src_stride,
+;    unsigned char *dst,
+;    int            dst_stride,
+;    int            src_weight
+;)
+global sym(vp8_filter_by_weight8x8_sse2) PRIVATE
+sym(vp8_filter_by_weight8x8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movd        xmm0, arg(4)                ; src_weight
+    pshuflw     xmm0, xmm0, 0x0             ; replicate to all low words
+    punpcklqdq  xmm0, xmm0                  ; replicate to all hi words
+
+    movdqa      xmm1, [GLOBAL(tMFQE)]
+    psubw       xmm1, xmm0                  ; dst_weight
+
+    mov         rax, arg(0)                 ; src
+    mov         rsi, arg(1)                 ; src_stride
+    mov         rdx, arg(2)                 ; dst
+    mov         rdi, arg(3)                 ; dst_stride
+
+    mov         rcx, 8                      ; loop count
+    pxor        xmm4, xmm4
+
+.combine
+    movq        xmm2, [rax]
+    movq        xmm3, [rdx]
+    add         rax, rsi
+
+    ; src * src_weight
+    punpcklbw   xmm2, xmm4
+    pmullw      xmm2, xmm0
+
+    ; dst * dst_weight
+    punpcklbw   xmm3, xmm4
+    pmullw      xmm3, xmm1
+
+    ; sum, round and shift
+    paddw       xmm2, xmm3
+    paddw       xmm2, [GLOBAL(tMFQE_round)]
+    psrlw       xmm2, 4
+
+    packuswb    xmm2, xmm4
+    movq        [rdx], xmm2
+    add         rdx, rdi
+
+    dec         rcx
+    jnz         .combine
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+
+    ret
+
+;void vp8_variance_and_sad_16x16_sse2 | arg
+;(
+;    unsigned char *src1,          0
+;    int            stride1,       1
+;    unsigned char *src2,          2
+;    int            stride2,       3
+;    unsigned int  *variance,      4
+;    unsigned int  *sad,           5
+;)
+global sym(vp8_variance_and_sad_16x16_sse2) PRIVATE
+sym(vp8_variance_and_sad_16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov         rax,        arg(0)          ; src1
+    mov         rcx,        arg(1)          ; stride1
+    mov         rdx,        arg(2)          ; src2
+    mov         rdi,        arg(3)          ; stride2
+
+    mov         rsi,        16              ; block height
+
+    ; Prep accumulator registers
+    pxor        xmm3, xmm3                  ; SAD
+    pxor        xmm4, xmm4                  ; sum of src2
+    pxor        xmm5, xmm5                  ; sum of src2^2
+
+    ; Because we're working with the actual output frames
+    ; we can't depend on any kind of data alignment.
+.accumulate
+    movdqa      xmm0, [rax]                 ; src1
+    movdqa      xmm1, [rdx]                 ; src2
+    add         rax, rcx                    ; src1 + stride1
+    add         rdx, rdi                    ; src2 + stride2
+
+    ; SAD(src1, src2)
+    psadbw      xmm0, xmm1
+    paddusw     xmm3, xmm0
+
+    ; SUM(src2)
+    pxor        xmm2, xmm2
+    psadbw      xmm2, xmm1                  ; sum src2 by misusing SAD against 0
+    paddusw     xmm4, xmm2
+
+    ; pmaddubsw would be ideal if it took two unsigned values. instead,
+    ; it expects a signed and an unsigned value. so instead we zero extend
+    ; and operate on words.
+    pxor        xmm2, xmm2
+    movdqa      xmm0, xmm1
+    punpcklbw   xmm0, xmm2
+    punpckhbw   xmm1, xmm2
+    pmaddwd     xmm0, xmm0
+    pmaddwd     xmm1, xmm1
+    paddd       xmm5, xmm0
+    paddd       xmm5, xmm1
+
+    sub         rsi,        1
+    jnz         .accumulate
+
+    ; phaddd only operates on adjacent double words.
+    ; Finalize SAD and store
+    movdqa      xmm0, xmm3
+    psrldq      xmm0, 8
+    paddusw     xmm0, xmm3
+    paddd       xmm0, [GLOBAL(t128)]
+    psrld       xmm0, 8
+
+    mov         rax,  arg(5)
+    movd        [rax], xmm0
+
+    ; Accumulate sum of src2
+    movdqa      xmm0, xmm4
+    psrldq      xmm0, 8
+    paddusw     xmm0, xmm4
+    ; Square src2. Ignore high value
+    pmuludq     xmm0, xmm0
+    psrld       xmm0, 8
+
+    ; phaddw could be used to sum adjacent values but we want
+    ; all the values summed. promote to doubles, accumulate,
+    ; shift and sum
+    pxor        xmm2, xmm2
+    movdqa      xmm1, xmm5
+    punpckldq   xmm1, xmm2
+    punpckhdq   xmm5, xmm2
+    paddd       xmm1, xmm5
+    movdqa      xmm2, xmm1
+    psrldq      xmm1, 8
+    paddd       xmm1, xmm2
+
+    psubd       xmm1, xmm0
+
+    ; (variance + 128) >> 8
+    paddd       xmm1, [GLOBAL(t128)]
+    psrld       xmm1, 8
+    mov         rax,  arg(4)
+
+    movd        [rax], xmm1
+
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+t128:
+%ifndef __NASM_VER__
+    ddq 128
+%elif CONFIG_BIG_ENDIAN
+    dq  0, 128
+%else
+    dq  128, 0
+%endif
+align 16
+tMFQE: ; 1 << MFQE_PRECISION
+    times 8 dw 0x10
+align 16
+tMFQE_round: ; 1 << (MFQE_PRECISION - 1)
+    times 8 dw 0x08
+
diff --git a/libvpx/libvpx/vp8/common/x86/postproc_mmx.asm b/libvpx/libvpx/vp8/common/x86/postproc_mmx.asm
new file mode 100644
index 0000000..1a89e7e
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/postproc_mmx.asm
@@ -0,0 +1,253 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%define VP8_FILTER_WEIGHT 128
+%define VP8_FILTER_SHIFT  7
+
+;void vp8_mbpost_proc_down_mmx(unsigned char *dst,
+;                             int pitch, int rows, int cols,int flimit)
+extern sym(vp8_rv)
+global sym(vp8_mbpost_proc_down_mmx) PRIVATE
+sym(vp8_mbpost_proc_down_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 136
+
+    ; unsigned char d[16][8] at [rsp]
+    ; create flimit2 at [rsp+128]
+    mov         eax, dword ptr arg(4) ;flimit
+    mov         [rsp+128], eax
+    mov         [rsp+128+4], eax
+%define flimit2 [rsp+128]
+
+%if ABI_IS_32BIT=0
+    lea         r8,       [GLOBAL(sym(vp8_rv))]
+%endif
+
+    ;rows +=8;
+    add         dword ptr arg(2), 8
+
+    ;for(c=0; c<cols; c+=4)
+.loop_col:
+            mov         rsi,        arg(0)  ;s
+            pxor        mm0,        mm0     ;
+
+            movsxd      rax,        dword ptr arg(1) ;pitch       ;
+
+            ; this copies the last row down into the border 8 rows
+            mov         rdi,        rsi
+            mov         rdx,        arg(2)
+            sub         rdx,        9
+            imul        rdx,        rax
+            lea         rdi,        [rdi+rdx]
+            movq        mm1,        QWORD ptr[rdi]              ; first row
+            mov         rcx,        8
+.init_borderd                                                    ; initialize borders
+            lea         rdi,        [rdi + rax]
+            movq        [rdi],      mm1
+
+            dec         rcx
+            jne         .init_borderd
+
+            neg         rax                                     ; rax = -pitch
+
+            ; this copies the first row up into the border 8 rows
+            mov         rdi,        rsi
+            movq        mm1,        QWORD ptr[rdi]              ; first row
+            mov         rcx,        8
+.init_border                                                    ; initialize borders
+            lea         rdi,        [rdi + rax]
+            movq        [rdi],      mm1
+
+            dec         rcx
+            jne         .init_border
+
+
+            lea         rsi,        [rsi + rax*8];              ; rdi = s[-pitch*8]
+            neg         rax
+
+
+            pxor        mm5,        mm5
+            pxor        mm6,        mm6     ;
+
+            pxor        mm7,        mm7     ;
+            mov         rdi,        rsi
+
+            mov         rcx,        15          ;
+
+.loop_initvar:
+            movd        mm1,        DWORD PTR [rdi];
+            punpcklbw   mm1,        mm0     ;
+
+            paddw       mm5,        mm1     ;
+            pmullw      mm1,        mm1     ;
+
+            movq        mm2,        mm1     ;
+            punpcklwd   mm1,        mm0     ;
+
+            punpckhwd   mm2,        mm0     ;
+            paddd       mm6,        mm1     ;
+
+            paddd       mm7,        mm2     ;
+            lea         rdi,        [rdi+rax]   ;
+
+            dec         rcx
+            jne         .loop_initvar
+            ;save the var and sum
+            xor         rdx,        rdx
+.loop_row:
+            movd        mm1,        DWORD PTR [rsi]     ; [s-pitch*8]
+            movd        mm2,        DWORD PTR [rdi]     ; [s+pitch*7]
+
+            punpcklbw   mm1,        mm0
+            punpcklbw   mm2,        mm0
+
+            paddw       mm5,        mm2
+            psubw       mm5,        mm1
+
+            pmullw      mm2,        mm2
+            movq        mm4,        mm2
+
+            punpcklwd   mm2,        mm0
+            punpckhwd   mm4,        mm0
+
+            paddd       mm6,        mm2
+            paddd       mm7,        mm4
+
+            pmullw      mm1,        mm1
+            movq        mm2,        mm1
+
+            punpcklwd   mm1,        mm0
+            psubd       mm6,        mm1
+
+            punpckhwd   mm2,        mm0
+            psubd       mm7,        mm2
+
+
+            movq        mm3,        mm6
+            pslld       mm3,        4
+
+            psubd       mm3,        mm6
+            movq        mm1,        mm5
+
+            movq        mm4,        mm5
+            pmullw      mm1,        mm1
+
+            pmulhw      mm4,        mm4
+            movq        mm2,        mm1
+
+            punpcklwd   mm1,        mm4
+            punpckhwd   mm2,        mm4
+
+            movq        mm4,        mm7
+            pslld       mm4,        4
+
+            psubd       mm4,        mm7
+
+            psubd       mm3,        mm1
+            psubd       mm4,        mm2
+
+            psubd       mm3,        flimit2
+            psubd       mm4,        flimit2
+
+            psrad       mm3,        31
+            psrad       mm4,        31
+
+            packssdw    mm3,        mm4
+            packsswb    mm3,        mm0
+
+            movd        mm1,        DWORD PTR [rsi+rax*8]
+
+            movq        mm2,        mm1
+            punpcklbw   mm1,        mm0
+
+            paddw       mm1,        mm5
+            mov         rcx,        rdx
+
+            and         rcx,        127
+%if ABI_IS_32BIT=1 && CONFIG_PIC=1
+            push        rax
+            lea         rax,        [GLOBAL(sym(vp8_rv))]
+            movq        mm4,        [rax + rcx*2] ;vp8_rv[rcx*2]
+            pop         rax
+%elif ABI_IS_32BIT=0
+            movq        mm4,        [r8 + rcx*2] ;vp8_rv[rcx*2]
+%else
+            movq        mm4,        [sym(vp8_rv) + rcx*2]
+%endif
+            paddw       mm1,        mm4
+            psraw       mm1,        4
+
+            packuswb    mm1,        mm0
+            pand        mm1,        mm3
+
+            pandn       mm3,        mm2
+            por         mm1,        mm3
+
+            and         rcx,        15
+            movd        DWORD PTR   [rsp+rcx*4], mm1 ;d[rcx*4]
+
+            cmp         edx,        8
+            jl          .skip_assignment
+
+            mov         rcx,        rdx
+            sub         rcx,        8
+            and         rcx,        15
+            movd        mm1,        DWORD PTR [rsp+rcx*4] ;d[rcx*4]
+            movd        [rsi],      mm1
+
+.skip_assignment
+            lea         rsi,        [rsi+rax]
+
+            lea         rdi,        [rdi+rax]
+            add         rdx,        1
+
+            cmp         edx,        dword arg(2) ;rows
+            jl          .loop_row
+
+
+        add         dword arg(0), 4 ; s += 4
+        sub         dword arg(3), 4 ; cols -= 4
+        cmp         dword arg(3), 0
+        jg          .loop_col
+
+    add         rsp, 136
+    pop         rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%undef flimit2
+
+
+SECTION_RODATA
+align 16
+Blur:
+    times 16 dw 16
+    times  8 dw 64
+    times 16 dw 16
+    times  8 dw  0
+
+rd:
+    times 4 dw 0x40
diff --git a/libvpx/libvpx/vp8/common/x86/postproc_sse2.asm b/libvpx/libvpx/vp8/common/x86/postproc_sse2.asm
new file mode 100644
index 0000000..de17afa
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/postproc_sse2.asm
@@ -0,0 +1,661 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;macro in deblock functions
+%macro FIRST_2_ROWS 0
+        movdqa      xmm4,       xmm0
+        movdqa      xmm6,       xmm0
+        movdqa      xmm5,       xmm1
+        pavgb       xmm5,       xmm3
+
+        ;calculate absolute value
+        psubusb     xmm4,       xmm1
+        psubusb     xmm1,       xmm0
+        psubusb     xmm6,       xmm3
+        psubusb     xmm3,       xmm0
+        paddusb     xmm4,       xmm1
+        paddusb     xmm6,       xmm3
+
+        ;get threshold
+        movdqa      xmm2,       flimit
+        pxor        xmm1,       xmm1
+        movdqa      xmm7,       xmm2
+
+        ;get mask
+        psubusb     xmm2,       xmm4
+        psubusb     xmm7,       xmm6
+        pcmpeqb     xmm2,       xmm1
+        pcmpeqb     xmm7,       xmm1
+        por         xmm7,       xmm2
+%endmacro
+
+%macro SECOND_2_ROWS 0
+        movdqa      xmm6,       xmm0
+        movdqa      xmm4,       xmm0
+        movdqa      xmm2,       xmm1
+        pavgb       xmm1,       xmm3
+
+        ;calculate absolute value
+        psubusb     xmm6,       xmm2
+        psubusb     xmm2,       xmm0
+        psubusb     xmm4,       xmm3
+        psubusb     xmm3,       xmm0
+        paddusb     xmm6,       xmm2
+        paddusb     xmm4,       xmm3
+
+        pavgb       xmm5,       xmm1
+
+        ;get threshold
+        movdqa      xmm2,       flimit
+        pxor        xmm1,       xmm1
+        movdqa      xmm3,       xmm2
+
+        ;get mask
+        psubusb     xmm2,       xmm6
+        psubusb     xmm3,       xmm4
+        pcmpeqb     xmm2,       xmm1
+        pcmpeqb     xmm3,       xmm1
+
+        por         xmm7,       xmm2
+        por         xmm7,       xmm3
+
+        pavgb       xmm5,       xmm0
+
+        ;decide if or not to use filtered value
+        pand        xmm0,       xmm7
+        pandn       xmm7,       xmm5
+        paddusb     xmm0,       xmm7
+%endmacro
+
+%macro UPDATE_FLIMIT 0
+        movdqa      xmm2,       XMMWORD PTR [rbx]
+        movdqa      [rsp],      xmm2
+        add         rbx,        16
+%endmacro
+
+;void vp8_post_proc_down_and_across_mb_row_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned char *dst_ptr,
+;    int src_pixels_per_line,
+;    int dst_pixels_per_line,
+;    int cols,
+;    int *flimits,
+;    int size
+;)
+global sym(vp8_post_proc_down_and_across_mb_row_sse2) PRIVATE
+sym(vp8_post_proc_down_and_across_mb_row_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+    ALIGN_STACK 16, rax
+    sub         rsp, 16
+
+        ; put flimit on stack
+        mov         rbx,        arg(5)           ;flimits ptr
+        UPDATE_FLIMIT
+
+%define flimit [rsp]
+
+        mov         rsi,        arg(0)           ;src_ptr
+        mov         rdi,        arg(1)           ;dst_ptr
+
+        movsxd      rax,        DWORD PTR arg(2) ;src_pixels_per_line
+        movsxd      rcx,        DWORD PTR arg(6) ;rows in a macroblock
+.nextrow:
+        xor         rdx,        rdx              ;col
+.nextcol:
+        ;load current and next 2 rows
+        movdqu      xmm0,       XMMWORD PTR [rsi]
+        movdqu      xmm1,       XMMWORD PTR [rsi + rax]
+        movdqu      xmm3,       XMMWORD PTR [rsi + 2*rax]
+
+        FIRST_2_ROWS
+
+        ;load above 2 rows
+        neg         rax
+        movdqu      xmm1,       XMMWORD PTR [rsi + 2*rax]
+        movdqu      xmm3,       XMMWORD PTR [rsi + rax]
+
+        SECOND_2_ROWS
+
+        movdqu      XMMWORD PTR [rdi], xmm0
+
+        neg         rax                          ; positive stride
+        add         rsi,        16
+        add         rdi,        16
+
+        add         rdx,        16
+        cmp         edx,        dword arg(4)     ;cols
+        jge         .downdone
+        UPDATE_FLIMIT
+        jmp         .nextcol
+
+.downdone:
+        ; done with the all cols, start the across filtering in place
+        sub         rsi,        rdx
+        sub         rdi,        rdx
+
+        mov         rbx,        arg(5) ; flimits
+        UPDATE_FLIMIT
+
+        ; dup the first byte into the left border 8 times
+        movq        mm1,   [rdi]
+        punpcklbw   mm1,   mm1
+        punpcklwd   mm1,   mm1
+        punpckldq   mm1,   mm1
+        mov         rdx,    -8
+        movq        [rdi+rdx], mm1
+
+        ; dup the last byte into the right border
+        movsxd      rdx,    dword arg(4)
+        movq        mm1,   [rdi + rdx + -1]
+        punpcklbw   mm1,   mm1
+        punpcklwd   mm1,   mm1
+        punpckldq   mm1,   mm1
+        movq        [rdi+rdx], mm1
+
+        xor         rdx,        rdx
+        movq        mm0,        QWORD PTR [rdi-16];
+        movq        mm1,        QWORD PTR [rdi-8];
+
+.acrossnextcol:
+        movdqu      xmm0,       XMMWORD PTR [rdi + rdx]
+        movdqu      xmm1,       XMMWORD PTR [rdi + rdx -2]
+        movdqu      xmm3,       XMMWORD PTR [rdi + rdx -1]
+
+        FIRST_2_ROWS
+
+        movdqu      xmm1,       XMMWORD PTR [rdi + rdx +1]
+        movdqu      xmm3,       XMMWORD PTR [rdi + rdx +2]
+
+        SECOND_2_ROWS
+
+        movq        QWORD PTR [rdi+rdx-16], mm0  ; store previous 8 bytes
+        movq        QWORD PTR [rdi+rdx-8], mm1   ; store previous 8 bytes
+        movdq2q     mm0,        xmm0
+        psrldq      xmm0,       8
+        movdq2q     mm1,        xmm0
+
+        add         rdx,        16
+        cmp         edx,        dword arg(4)     ;cols
+        jge         .acrossdone
+        UPDATE_FLIMIT
+        jmp         .acrossnextcol
+
+.acrossdone
+        ; last 16 pixels
+        movq        QWORD PTR [rdi+rdx-16], mm0
+
+        cmp         edx,        dword arg(4)
+        jne         .throw_last_8
+        movq        QWORD PTR [rdi+rdx-8], mm1
+.throw_last_8:
+        ; done with this rwo
+        add         rsi,rax                      ;next src line
+        mov         eax, dword arg(3)            ;dst_pixels_per_line
+        add         rdi,rax                      ;next destination
+        mov         eax, dword arg(2)            ;src_pixels_per_line
+
+        mov         rbx,        arg(5)           ;flimits
+        UPDATE_FLIMIT
+
+        dec         rcx                          ;decrement count
+        jnz         .nextrow                     ;next row
+
+    add rsp, 16
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    pop rbx
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%undef flimit
+
+;void vp8_mbpost_proc_down_xmm(unsigned char *dst,
+;                            int pitch, int rows, int cols,int flimit)
+extern sym(vp8_rv)
+global sym(vp8_mbpost_proc_down_xmm) PRIVATE
+sym(vp8_mbpost_proc_down_xmm):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 128+16
+
+    ; unsigned char d[16][8] at [rsp]
+    ; create flimit2 at [rsp+128]
+    mov         eax, dword ptr arg(4) ;flimit
+    mov         [rsp+128], eax
+    mov         [rsp+128+4], eax
+    mov         [rsp+128+8], eax
+    mov         [rsp+128+12], eax
+%define flimit4 [rsp+128]
+
+%if ABI_IS_32BIT=0
+    lea         r8,       [GLOBAL(sym(vp8_rv))]
+%endif
+
+    ;rows +=8;
+    add         dword arg(2), 8
+
+    ;for(c=0; c<cols; c+=8)
+.loop_col:
+            mov         rsi,        arg(0) ; s
+            pxor        xmm0,       xmm0        ;
+
+            movsxd      rax,        dword ptr arg(1) ;pitch       ;
+
+            ; this copies the last row down into the border 8 rows
+            mov         rdi,        rsi
+            mov         rdx,        arg(2)
+            sub         rdx,        9
+            imul        rdx,        rax
+            lea         rdi,        [rdi+rdx]
+            movq        xmm1,       QWORD ptr[rdi]              ; first row
+            mov         rcx,        8
+.init_borderd                                                    ; initialize borders
+            lea         rdi,        [rdi + rax]
+            movq        [rdi],      xmm1
+
+            dec         rcx
+            jne         .init_borderd
+
+            neg         rax                                     ; rax = -pitch
+
+            ; this copies the first row up into the border 8 rows
+            mov         rdi,        rsi
+            movq        xmm1,       QWORD ptr[rdi]              ; first row
+            mov         rcx,        8
+.init_border                                                    ; initialize borders
+            lea         rdi,        [rdi + rax]
+            movq        [rdi],      xmm1
+
+            dec         rcx
+            jne         .init_border
+
+
+
+            lea         rsi,        [rsi + rax*8];              ; rdi = s[-pitch*8]
+            neg         rax
+
+            pxor        xmm5,       xmm5
+            pxor        xmm6,       xmm6        ;
+
+            pxor        xmm7,       xmm7        ;
+            mov         rdi,        rsi
+
+            mov         rcx,        15          ;
+
+.loop_initvar:
+            movq        xmm1,       QWORD PTR [rdi];
+            punpcklbw   xmm1,       xmm0        ;
+
+            paddw       xmm5,       xmm1        ;
+            pmullw      xmm1,       xmm1        ;
+
+            movdqa      xmm2,       xmm1        ;
+            punpcklwd   xmm1,       xmm0        ;
+
+            punpckhwd   xmm2,       xmm0        ;
+            paddd       xmm6,       xmm1        ;
+
+            paddd       xmm7,       xmm2        ;
+            lea         rdi,        [rdi+rax]   ;
+
+            dec         rcx
+            jne         .loop_initvar
+            ;save the var and sum
+            xor         rdx,        rdx
+.loop_row:
+            movq        xmm1,       QWORD PTR [rsi]     ; [s-pitch*8]
+            movq        xmm2,       QWORD PTR [rdi]     ; [s+pitch*7]
+
+            punpcklbw   xmm1,       xmm0
+            punpcklbw   xmm2,       xmm0
+
+            paddw       xmm5,       xmm2
+            psubw       xmm5,       xmm1
+
+            pmullw      xmm2,       xmm2
+            movdqa      xmm4,       xmm2
+
+            punpcklwd   xmm2,       xmm0
+            punpckhwd   xmm4,       xmm0
+
+            paddd       xmm6,       xmm2
+            paddd       xmm7,       xmm4
+
+            pmullw      xmm1,       xmm1
+            movdqa      xmm2,       xmm1
+
+            punpcklwd   xmm1,       xmm0
+            psubd       xmm6,       xmm1
+
+            punpckhwd   xmm2,       xmm0
+            psubd       xmm7,       xmm2
+
+
+            movdqa      xmm3,       xmm6
+            pslld       xmm3,       4
+
+            psubd       xmm3,       xmm6
+            movdqa      xmm1,       xmm5
+
+            movdqa      xmm4,       xmm5
+            pmullw      xmm1,       xmm1
+
+            pmulhw      xmm4,       xmm4
+            movdqa      xmm2,       xmm1
+
+            punpcklwd   xmm1,       xmm4
+            punpckhwd   xmm2,       xmm4
+
+            movdqa      xmm4,       xmm7
+            pslld       xmm4,       4
+
+            psubd       xmm4,       xmm7
+
+            psubd       xmm3,       xmm1
+            psubd       xmm4,       xmm2
+
+            psubd       xmm3,       flimit4
+            psubd       xmm4,       flimit4
+
+            psrad       xmm3,       31
+            psrad       xmm4,       31
+
+            packssdw    xmm3,       xmm4
+            packsswb    xmm3,       xmm0
+
+            movq        xmm1,       QWORD PTR [rsi+rax*8]
+
+            movq        xmm2,       xmm1
+            punpcklbw   xmm1,       xmm0
+
+            paddw       xmm1,       xmm5
+            mov         rcx,        rdx
+
+            and         rcx,        127
+%if ABI_IS_32BIT=1 && CONFIG_PIC=1
+            push        rax
+            lea         rax,        [GLOBAL(sym(vp8_rv))]
+            movdqu      xmm4,       [rax + rcx*2] ;vp8_rv[rcx*2]
+            pop         rax
+%elif ABI_IS_32BIT=0
+            movdqu      xmm4,       [r8 + rcx*2] ;vp8_rv[rcx*2]
+%else
+            movdqu      xmm4,       [sym(vp8_rv) + rcx*2]
+%endif
+
+            paddw       xmm1,       xmm4
+            ;paddw     xmm1,       eight8s
+            psraw       xmm1,       4
+
+            packuswb    xmm1,       xmm0
+            pand        xmm1,       xmm3
+
+            pandn       xmm3,       xmm2
+            por         xmm1,       xmm3
+
+            and         rcx,        15
+            movq        QWORD PTR   [rsp + rcx*8], xmm1 ;d[rcx*8]
+
+            cmp         edx,        8
+            jl          .skip_assignment
+
+            mov         rcx,        rdx
+            sub         rcx,        8
+            and         rcx,        15
+            movq        mm0,        [rsp + rcx*8] ;d[rcx*8]
+            movq        [rsi],      mm0
+
+.skip_assignment
+            lea         rsi,        [rsi+rax]
+
+            lea         rdi,        [rdi+rax]
+            add         rdx,        1
+
+            cmp         edx,        dword arg(2) ;rows
+            jl          .loop_row
+
+        add         dword arg(0), 8 ; s += 8
+        sub         dword arg(3), 8 ; cols -= 8
+        cmp         dword arg(3), 0
+        jg          .loop_col
+
+    add         rsp, 128+16
+    pop         rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%undef flimit4
+
+
+;void vp8_mbpost_proc_across_ip_xmm(unsigned char *src,
+;                                int pitch, int rows, int cols,int flimit)
+global sym(vp8_mbpost_proc_across_ip_xmm) PRIVATE
+sym(vp8_mbpost_proc_across_ip_xmm):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16
+
+    ; create flimit4 at [rsp]
+    mov         eax, dword ptr arg(4) ;flimit
+    mov         [rsp], eax
+    mov         [rsp+4], eax
+    mov         [rsp+8], eax
+    mov         [rsp+12], eax
+%define flimit4 [rsp]
+
+
+    ;for(r=0;r<rows;r++)
+.ip_row_loop:
+
+        xor         rdx,    rdx ;sumsq=0;
+        xor         rcx,    rcx ;sum=0;
+        mov         rsi,    arg(0); s
+
+
+        ; dup the first byte into the left border 8 times
+        movq        mm1,   [rsi]
+        punpcklbw   mm1,   mm1
+        punpcklwd   mm1,   mm1
+        punpckldq   mm1,   mm1
+
+        mov         rdi,    -8
+        movq        [rsi+rdi], mm1
+
+        ; dup the last byte into the right border
+        movsxd      rdx,    dword arg(3)
+        movq        mm1,   [rsi + rdx + -1]
+        punpcklbw   mm1,   mm1
+        punpcklwd   mm1,   mm1
+        punpckldq   mm1,   mm1
+        movq        [rsi+rdx], mm1
+
+.ip_var_loop:
+        ;for(i=-8;i<=6;i++)
+        ;{
+        ;    sumsq += s[i]*s[i];
+        ;    sum   += s[i];
+        ;}
+        movzx       eax, byte [rsi+rdi]
+        add         ecx, eax
+        mul         al
+        add         edx, eax
+        add         rdi, 1
+        cmp         rdi, 6
+        jle         .ip_var_loop
+
+
+            ;mov         rax,    sumsq
+            ;movd        xmm7,   rax
+            movd        xmm7,   edx
+
+            ;mov         rax,    sum
+            ;movd        xmm6,   rax
+            movd        xmm6,   ecx
+
+            mov         rsi,    arg(0) ;s
+            xor         rcx,    rcx
+
+            movsxd      rdx,    dword arg(3) ;cols
+            add         rdx,    8
+            pxor        mm0,    mm0
+            pxor        mm1,    mm1
+
+            pxor        xmm0,   xmm0
+.nextcol4:
+
+            movd        xmm1,   DWORD PTR [rsi+rcx-8]   ; -8 -7 -6 -5
+            movd        xmm2,   DWORD PTR [rsi+rcx+7]   ; +7 +8 +9 +10
+
+            punpcklbw   xmm1,   xmm0                    ; expanding
+            punpcklbw   xmm2,   xmm0                    ; expanding
+
+            punpcklwd   xmm1,   xmm0                    ; expanding to dwords
+            punpcklwd   xmm2,   xmm0                    ; expanding to dwords
+
+            psubd       xmm2,   xmm1                    ; 7--8   8--7   9--6 10--5
+            paddd       xmm1,   xmm1                    ; -8*2   -7*2   -6*2 -5*2
+
+            paddd       xmm1,   xmm2                    ; 7+-8   8+-7   9+-6 10+-5
+            pmaddwd     xmm1,   xmm2                    ; squared of 7+-8   8+-7   9+-6 10+-5
+
+            paddd       xmm6,   xmm2
+            paddd       xmm7,   xmm1
+
+            pshufd      xmm6,   xmm6,   0               ; duplicate the last ones
+            pshufd      xmm7,   xmm7,   0               ; duplicate the last ones
+
+            psrldq      xmm1,       4                   ; 8--7   9--6 10--5  0000
+            psrldq      xmm2,       4                   ; 8--7   9--6 10--5  0000
+
+            pshufd      xmm3,   xmm1,   3               ; 0000  8--7   8--7   8--7 squared
+            pshufd      xmm4,   xmm2,   3               ; 0000  8--7   8--7   8--7 squared
+
+            paddd       xmm6,   xmm4
+            paddd       xmm7,   xmm3
+
+            pshufd      xmm3,   xmm1,   01011111b       ; 0000  0000   9--6   9--6 squared
+            pshufd      xmm4,   xmm2,   01011111b       ; 0000  0000   9--6   9--6 squared
+
+            paddd       xmm7,   xmm3
+            paddd       xmm6,   xmm4
+
+            pshufd      xmm3,   xmm1,   10111111b       ; 0000  0000   8--7   8--7 squared
+            pshufd      xmm4,   xmm2,   10111111b       ; 0000  0000   8--7   8--7 squared
+
+            paddd       xmm7,   xmm3
+            paddd       xmm6,   xmm4
+
+            movdqa      xmm3,   xmm6
+            pmaddwd     xmm3,   xmm3
+
+            movdqa      xmm5,   xmm7
+            pslld       xmm5,   4
+
+            psubd       xmm5,   xmm7
+            psubd       xmm5,   xmm3
+
+            psubd       xmm5,   flimit4
+            psrad       xmm5,   31
+
+            packssdw    xmm5,   xmm0
+            packsswb    xmm5,   xmm0
+
+            movd        xmm1,   DWORD PTR [rsi+rcx]
+            movq        xmm2,   xmm1
+
+            punpcklbw   xmm1,   xmm0
+            punpcklwd   xmm1,   xmm0
+
+            paddd       xmm1,   xmm6
+            paddd       xmm1,   [GLOBAL(four8s)]
+
+            psrad       xmm1,   4
+            packssdw    xmm1,   xmm0
+
+            packuswb    xmm1,   xmm0
+            pand        xmm1,   xmm5
+
+            pandn       xmm5,   xmm2
+            por         xmm5,   xmm1
+
+            movd        [rsi+rcx-8],  mm0
+            movq        mm0,    mm1
+
+            movdq2q     mm1,    xmm5
+            psrldq      xmm7,   12
+
+            psrldq      xmm6,   12
+            add         rcx,    4
+
+            cmp         rcx,    rdx
+            jl          .nextcol4
+
+        ;s+=pitch;
+        movsxd rax, dword arg(1)
+        add    arg(0), rax
+
+        sub dword arg(2), 1 ;rows-=1
+        cmp dword arg(2), 0
+        jg .ip_row_loop
+
+    add         rsp, 16
+    pop         rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%undef flimit4
+
+
+SECTION_RODATA
+align 16
+four8s:
+    times 4 dd 8
diff --git a/libvpx/libvpx/vp8/common/x86/recon_mmx.asm b/libvpx/libvpx/vp8/common/x86/recon_mmx.asm
new file mode 100644
index 0000000..15e9871
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/recon_mmx.asm
@@ -0,0 +1,274 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+
+;void copy_mem8x8_mmx(
+;    unsigned char *src,
+;    int src_stride,
+;    unsigned char *dst,
+;    int dst_stride
+;    )
+global sym(vp8_copy_mem8x8_mmx) PRIVATE
+sym(vp8_copy_mem8x8_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 4
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rsi,        arg(0) ;src;
+        movq        mm0,        [rsi]
+
+        movsxd      rax,        dword ptr arg(1) ;src_stride;
+        mov         rdi,        arg(2) ;dst;
+
+        movq        mm1,        [rsi+rax]
+        movq        mm2,        [rsi+rax*2]
+
+        movsxd      rcx,        dword ptr arg(3) ;dst_stride
+        lea         rsi,        [rsi+rax*2]
+
+        movq        [rdi],      mm0
+        add         rsi,        rax
+
+        movq        [rdi+rcx],      mm1
+        movq        [rdi+rcx*2],    mm2
+
+
+        lea         rdi,        [rdi+rcx*2]
+        movq        mm3,        [rsi]
+
+        add         rdi,        rcx
+        movq        mm4,        [rsi+rax]
+
+        movq        mm5,        [rsi+rax*2]
+        movq        [rdi],      mm3
+
+        lea         rsi,        [rsi+rax*2]
+        movq        [rdi+rcx],  mm4
+
+        movq        [rdi+rcx*2],    mm5
+        lea         rdi,        [rdi+rcx*2]
+
+        movq        mm0,        [rsi+rax]
+        movq        mm1,        [rsi+rax*2]
+
+        movq        [rdi+rcx],  mm0
+        movq        [rdi+rcx*2],mm1
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void copy_mem8x4_mmx(
+;    unsigned char *src,
+;    int src_stride,
+;    unsigned char *dst,
+;    int dst_stride
+;    )
+global sym(vp8_copy_mem8x4_mmx) PRIVATE
+sym(vp8_copy_mem8x4_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 4
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rsi,        arg(0) ;src;
+        movq        mm0,        [rsi]
+
+        movsxd      rax,        dword ptr arg(1) ;src_stride;
+        mov         rdi,        arg(2) ;dst;
+
+        movq        mm1,        [rsi+rax]
+        movq        mm2,        [rsi+rax*2]
+
+        movsxd      rcx,        dword ptr arg(3) ;dst_stride
+        lea         rsi,        [rsi+rax*2]
+
+        movq        [rdi],      mm0
+        movq        [rdi+rcx],      mm1
+
+        movq        [rdi+rcx*2],    mm2
+        lea         rdi,        [rdi+rcx*2]
+
+        movq        mm3,        [rsi+rax]
+        movq        [rdi+rcx],      mm3
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void copy_mem16x16_mmx(
+;    unsigned char *src,
+;    int src_stride,
+;    unsigned char *dst,
+;    int dst_stride
+;    )
+global sym(vp8_copy_mem16x16_mmx) PRIVATE
+sym(vp8_copy_mem16x16_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 4
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rsi,        arg(0) ;src;
+        movsxd      rax,        dword ptr arg(1) ;src_stride;
+
+        mov         rdi,        arg(2) ;dst;
+        movsxd      rcx,        dword ptr arg(3) ;dst_stride
+
+        movq        mm0,            [rsi]
+        movq        mm3,            [rsi+8];
+
+        movq        mm1,            [rsi+rax]
+        movq        mm4,            [rsi+rax+8]
+
+        movq        mm2,            [rsi+rax*2]
+        movq        mm5,            [rsi+rax*2+8]
+
+        lea         rsi,            [rsi+rax*2]
+        add         rsi,            rax
+
+        movq        [rdi],          mm0
+        movq        [rdi+8],        mm3
+
+        movq        [rdi+rcx],      mm1
+        movq        [rdi+rcx+8],    mm4
+
+        movq        [rdi+rcx*2],    mm2
+        movq        [rdi+rcx*2+8],  mm5
+
+        lea         rdi,            [rdi+rcx*2]
+        add         rdi,            rcx
+
+        movq        mm0,            [rsi]
+        movq        mm3,            [rsi+8];
+
+        movq        mm1,            [rsi+rax]
+        movq        mm4,            [rsi+rax+8]
+
+        movq        mm2,            [rsi+rax*2]
+        movq        mm5,            [rsi+rax*2+8]
+
+        lea         rsi,            [rsi+rax*2]
+        add         rsi,            rax
+
+        movq        [rdi],          mm0
+        movq        [rdi+8],        mm3
+
+        movq        [rdi+rcx],      mm1
+        movq        [rdi+rcx+8],    mm4
+
+        movq        [rdi+rcx*2],    mm2
+        movq        [rdi+rcx*2+8],  mm5
+
+        lea         rdi,            [rdi+rcx*2]
+        add         rdi,            rcx
+
+        movq        mm0,            [rsi]
+        movq        mm3,            [rsi+8];
+
+        movq        mm1,            [rsi+rax]
+        movq        mm4,            [rsi+rax+8]
+
+        movq        mm2,            [rsi+rax*2]
+        movq        mm5,            [rsi+rax*2+8]
+
+        lea         rsi,            [rsi+rax*2]
+        add         rsi,            rax
+
+        movq        [rdi],          mm0
+        movq        [rdi+8],        mm3
+
+        movq        [rdi+rcx],      mm1
+        movq        [rdi+rcx+8],    mm4
+
+        movq        [rdi+rcx*2],    mm2
+        movq        [rdi+rcx*2+8],  mm5
+
+        lea         rdi,            [rdi+rcx*2]
+        add         rdi,            rcx
+
+        movq        mm0,            [rsi]
+        movq        mm3,            [rsi+8];
+
+        movq        mm1,            [rsi+rax]
+        movq        mm4,            [rsi+rax+8]
+
+        movq        mm2,            [rsi+rax*2]
+        movq        mm5,            [rsi+rax*2+8]
+
+        lea         rsi,            [rsi+rax*2]
+        add         rsi,            rax
+
+        movq        [rdi],          mm0
+        movq        [rdi+8],        mm3
+
+        movq        [rdi+rcx],      mm1
+        movq        [rdi+rcx+8],    mm4
+
+        movq        [rdi+rcx*2],    mm2
+        movq        [rdi+rcx*2+8],  mm5
+
+        lea         rdi,            [rdi+rcx*2]
+        add         rdi,            rcx
+
+        movq        mm0,            [rsi]
+        movq        mm3,            [rsi+8];
+
+        movq        mm1,            [rsi+rax]
+        movq        mm4,            [rsi+rax+8]
+
+        movq        mm2,            [rsi+rax*2]
+        movq        mm5,            [rsi+rax*2+8]
+
+        lea         rsi,            [rsi+rax*2]
+        add         rsi,            rax
+
+        movq        [rdi],          mm0
+        movq        [rdi+8],        mm3
+
+        movq        [rdi+rcx],      mm1
+        movq        [rdi+rcx+8],    mm4
+
+        movq        [rdi+rcx*2],    mm2
+        movq        [rdi+rcx*2+8],  mm5
+
+        lea         rdi,            [rdi+rcx*2]
+        add         rdi,            rcx
+
+        movq        mm0,            [rsi]
+        movq        mm3,            [rsi+8];
+
+        movq        [rdi],          mm0
+        movq        [rdi+8],        mm3
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vp8/common/x86/recon_sse2.asm b/libvpx/libvpx/vp8/common/x86/recon_sse2.asm
new file mode 100644
index 0000000..cb89537
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/recon_sse2.asm
@@ -0,0 +1,116 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void copy_mem16x16_sse2(
+;    unsigned char *src,
+;    int src_stride,
+;    unsigned char *dst,
+;    int dst_stride
+;    )
+global sym(vp8_copy_mem16x16_sse2) PRIVATE
+sym(vp8_copy_mem16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 4
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rsi,        arg(0) ;src;
+        movdqu      xmm0,       [rsi]
+
+        movsxd      rax,        dword ptr arg(1) ;src_stride;
+        mov         rdi,        arg(2) ;dst;
+
+        movdqu      xmm1,       [rsi+rax]
+        movdqu      xmm2,       [rsi+rax*2]
+
+        movsxd      rcx,        dword ptr arg(3) ;dst_stride
+        lea         rsi,        [rsi+rax*2]
+
+        movdqa      [rdi],      xmm0
+        add         rsi,        rax
+
+        movdqa      [rdi+rcx],  xmm1
+        movdqa      [rdi+rcx*2],xmm2
+
+        lea         rdi,        [rdi+rcx*2]
+        movdqu      xmm3,       [rsi]
+
+        add         rdi,        rcx
+        movdqu      xmm4,       [rsi+rax]
+
+        movdqu      xmm5,       [rsi+rax*2]
+        lea         rsi,        [rsi+rax*2]
+
+        movdqa      [rdi],  xmm3
+        add         rsi,        rax
+
+        movdqa      [rdi+rcx],  xmm4
+        movdqa      [rdi+rcx*2],xmm5
+
+        lea         rdi,        [rdi+rcx*2]
+        movdqu      xmm0,       [rsi]
+
+        add         rdi,        rcx
+        movdqu      xmm1,       [rsi+rax]
+
+        movdqu      xmm2,       [rsi+rax*2]
+        lea         rsi,        [rsi+rax*2]
+
+        movdqa      [rdi],      xmm0
+        add         rsi,        rax
+
+        movdqa      [rdi+rcx],  xmm1
+
+        movdqa      [rdi+rcx*2],    xmm2
+        movdqu      xmm3,       [rsi]
+
+        movdqu      xmm4,       [rsi+rax]
+        lea         rdi,        [rdi+rcx*2]
+
+        add         rdi,        rcx
+        movdqu      xmm5,       [rsi+rax*2]
+
+        lea         rsi,        [rsi+rax*2]
+        movdqa      [rdi],  xmm3
+
+        add         rsi,        rax
+        movdqa      [rdi+rcx],  xmm4
+
+        movdqa      [rdi+rcx*2],xmm5
+        movdqu      xmm0,       [rsi]
+
+        lea         rdi,        [rdi+rcx*2]
+        movdqu      xmm1,       [rsi+rax]
+
+        add         rdi,        rcx
+        movdqu      xmm2,       [rsi+rax*2]
+
+        lea         rsi,        [rsi+rax*2]
+        movdqa      [rdi],      xmm0
+
+        movdqa      [rdi+rcx],  xmm1
+        movdqa      [rdi+rcx*2],xmm2
+
+        movdqu      xmm3,       [rsi+rax]
+        lea         rdi,        [rdi+rcx*2]
+
+        movdqa      [rdi+rcx],  xmm3
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vp8/common/x86/subpixel_mmx.asm b/libvpx/libvpx/vp8/common/x86/subpixel_mmx.asm
new file mode 100644
index 0000000..47dd452
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/subpixel_mmx.asm
@@ -0,0 +1,702 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+extern sym(vp8_bilinear_filters_x86_8)
+
+
+%define BLOCK_HEIGHT_WIDTH 4
+%define vp8_filter_weight 128
+%define VP8_FILTER_SHIFT  7
+
+
+;void vp8_filter_block1d_h6_mmx
+;(
+;    unsigned char   *src_ptr,
+;    unsigned short  *output_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned int    pixel_step,
+;    unsigned int    output_height,
+;    unsigned int    output_width,
+;    short           * vp8_filter
+;)
+global sym(vp8_filter_block1d_h6_mmx) PRIVATE
+sym(vp8_filter_block1d_h6_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rdx,    arg(6) ;vp8_filter
+
+        movq        mm1,    [rdx + 16]             ; do both the negative taps first!!!
+        movq        mm2,    [rdx + 32]         ;
+        movq        mm6,    [rdx + 48]        ;
+        movq        mm7,    [rdx + 64]        ;
+
+        mov         rdi,    arg(1) ;output_ptr
+        mov         rsi,    arg(0) ;src_ptr
+        movsxd      rcx,    dword ptr arg(4) ;output_height
+        movsxd      rax,    dword ptr arg(5) ;output_width      ; destination pitch?
+        pxor        mm0,    mm0              ; mm0 = 00000000
+
+.nextrow:
+        movq        mm3,    [rsi-2]          ; mm3 = p-2..p5
+        movq        mm4,    mm3              ; mm4 = p-2..p5
+        psrlq       mm3,    8                ; mm3 = p-1..p5
+        punpcklbw   mm3,    mm0              ; mm3 = p-1..p2
+        pmullw      mm3,    mm1              ; mm3 *= kernel 1 modifiers.
+
+        movq        mm5,    mm4              ; mm5 = p-2..p5
+        punpckhbw   mm4,    mm0              ; mm5 = p2..p5
+        pmullw      mm4,    mm7              ; mm5 *= kernel 4 modifiers
+        paddsw      mm3,    mm4              ; mm3 += mm5
+
+        movq        mm4,    mm5              ; mm4 = p-2..p5;
+        psrlq       mm5,    16               ; mm5 = p0..p5;
+        punpcklbw   mm5,    mm0              ; mm5 = p0..p3
+        pmullw      mm5,    mm2              ; mm5 *= kernel 2 modifiers
+        paddsw      mm3,    mm5              ; mm3 += mm5
+
+        movq        mm5,    mm4              ; mm5 = p-2..p5
+        psrlq       mm4,    24               ; mm4 = p1..p5
+        punpcklbw   mm4,    mm0              ; mm4 = p1..p4
+        pmullw      mm4,    mm6              ; mm5 *= kernel 3 modifiers
+        paddsw      mm3,    mm4              ; mm3 += mm5
+
+        ; do outer positive taps
+        movd        mm4,    [rsi+3]
+        punpcklbw   mm4,    mm0              ; mm5 = p3..p6
+        pmullw      mm4,    [rdx+80]         ; mm5 *= kernel 0 modifiers
+        paddsw      mm3,    mm4              ; mm3 += mm5
+
+        punpcklbw   mm5,    mm0              ; mm5 = p-2..p1
+        pmullw      mm5,    [rdx]            ; mm5 *= kernel 5 modifiers
+        paddsw      mm3,    mm5              ; mm3 += mm5
+
+        paddsw      mm3,    [GLOBAL(rd)]              ; mm3 += round value
+        psraw       mm3,    VP8_FILTER_SHIFT     ; mm3 /= 128
+        packuswb    mm3,    mm0              ; pack and unpack to saturate
+        punpcklbw   mm3,    mm0              ;
+
+        movq        [rdi],  mm3              ; store the results in the destination
+
+%if ABI_IS_32BIT
+        add         rsi,    dword ptr arg(2) ;src_pixels_per_line ; next line
+        add         rdi,    rax;
+%else
+        movsxd      r8,     dword ptr arg(2) ;src_pixels_per_line
+        add         rdi,    rax;
+
+        add         rsi,    r8               ; next line
+%endif
+
+        dec         rcx                      ; decrement count
+        jnz         .nextrow                 ; next row
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_filter_block1dc_v6_mmx
+;(
+;   short *src_ptr,
+;   unsigned char *output_ptr,
+;    int output_pitch,
+;   unsigned int pixels_per_line,
+;   unsigned int pixel_step,
+;   unsigned int output_height,
+;   unsigned int output_width,
+;   short * vp8_filter
+;)
+global sym(vp8_filter_block1dc_v6_mmx) PRIVATE
+sym(vp8_filter_block1dc_v6_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 8
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        movq      mm5, [GLOBAL(rd)]
+        push        rbx
+        mov         rbx, arg(7) ;vp8_filter
+        movq      mm1, [rbx + 16]             ; do both the negative taps first!!!
+        movq      mm2, [rbx + 32]         ;
+        movq      mm6, [rbx + 48]        ;
+        movq      mm7, [rbx + 64]        ;
+
+        movsxd      rdx, dword ptr arg(3) ;pixels_per_line
+        mov         rdi, arg(1) ;output_ptr
+        mov         rsi, arg(0) ;src_ptr
+        sub         rsi, rdx
+        sub         rsi, rdx
+        movsxd      rcx, DWORD PTR arg(5) ;output_height
+        movsxd      rax, DWORD PTR arg(2) ;output_pitch      ; destination pitch?
+        pxor        mm0, mm0              ; mm0 = 00000000
+
+
+.nextrow_cv:
+        movq        mm3, [rsi+rdx]        ; mm3 = p0..p8  = row -1
+        pmullw      mm3, mm1              ; mm3 *= kernel 1 modifiers.
+
+
+        movq        mm4, [rsi + 4*rdx]      ; mm4 = p0..p3  = row 2
+        pmullw      mm4, mm7              ; mm4 *= kernel 4 modifiers.
+        paddsw      mm3, mm4              ; mm3 += mm4
+
+        movq        mm4, [rsi + 2*rdx]           ; mm4 = p0..p3  = row 0
+        pmullw      mm4, mm2              ; mm4 *= kernel 2 modifiers.
+        paddsw      mm3, mm4              ; mm3 += mm4
+
+        movq        mm4, [rsi]            ; mm4 = p0..p3  = row -2
+        pmullw      mm4, [rbx]            ; mm4 *= kernel 0 modifiers.
+        paddsw      mm3, mm4              ; mm3 += mm4
+
+
+        add         rsi, rdx              ; move source forward 1 line to avoid 3 * pitch
+        movq        mm4, [rsi + 2*rdx]     ; mm4 = p0..p3  = row 1
+        pmullw      mm4, mm6              ; mm4 *= kernel 3 modifiers.
+        paddsw      mm3, mm4              ; mm3 += mm4
+
+        movq        mm4, [rsi + 4*rdx]    ; mm4 = p0..p3  = row 3
+        pmullw      mm4, [rbx +80]        ; mm4 *= kernel 3 modifiers.
+        paddsw      mm3, mm4              ; mm3 += mm4
+
+
+        paddsw      mm3, mm5               ; mm3 += round value
+        psraw       mm3, VP8_FILTER_SHIFT     ; mm3 /= 128
+        packuswb    mm3, mm0              ; pack and saturate
+
+        movd        [rdi],mm3             ; store the results in the destination
+        ; the subsequent iterations repeat 3 out of 4 of these reads.  Since the
+        ; recon block should be in cache this shouldn't cost much.  Its obviously
+        ; avoidable!!!.
+        lea         rdi,  [rdi+rax] ;
+        dec         rcx                   ; decrement count
+        jnz         .nextrow_cv           ; next row
+
+        pop         rbx
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void bilinear_predict8x8_mmx
+;(
+;    unsigned char  *src_ptr,
+;    int   src_pixels_per_line,
+;    int  xoffset,
+;    int  yoffset,
+;   unsigned char *dst_ptr,
+;    int dst_pitch
+;)
+global sym(vp8_bilinear_predict8x8_mmx) PRIVATE
+sym(vp8_bilinear_predict8x8_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ;const short *HFilter = vp8_bilinear_filters_x86_8[xoffset];
+    ;const short *VFilter = vp8_bilinear_filters_x86_8[yoffset];
+
+        movsxd      rax,        dword ptr arg(2) ;xoffset
+        mov         rdi,        arg(4) ;dst_ptr           ;
+
+        shl         rax,        5 ; offset * 32
+        lea         rcx,        [GLOBAL(sym(vp8_bilinear_filters_x86_8))]
+
+        add         rax,        rcx ; HFilter
+        mov         rsi,        arg(0) ;src_ptr              ;
+
+        movsxd      rdx,        dword ptr arg(5) ;dst_pitch
+        movq        mm1,        [rax]               ;
+
+        movq        mm2,        [rax+16]            ;
+        movsxd      rax,        dword ptr arg(3) ;yoffset
+
+        pxor        mm0,        mm0                 ;
+
+        shl         rax,        5 ; offset*32
+        add         rax,        rcx ; VFilter
+
+        lea         rcx,        [rdi+rdx*8]          ;
+        movsxd      rdx,        dword ptr arg(1) ;src_pixels_per_line    ;
+
+
+
+        ; get the first horizontal line done       ;
+        movq        mm3,        [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+        movq        mm4,        mm3                 ; make a copy of current line
+
+        punpcklbw   mm3,        mm0                 ; xx 00 01 02 03 04 05 06
+        punpckhbw   mm4,        mm0                 ;
+
+        pmullw      mm3,        mm1                 ;
+        pmullw      mm4,        mm1                 ;
+
+        movq        mm5,        [rsi+1]             ;
+        movq        mm6,        mm5                 ;
+
+        punpcklbw   mm5,        mm0                 ;
+        punpckhbw   mm6,        mm0                 ;
+
+        pmullw      mm5,        mm2                 ;
+        pmullw      mm6,        mm2                 ;
+
+        paddw       mm3,        mm5                 ;
+        paddw       mm4,        mm6                 ;
+
+        paddw       mm3,        [GLOBAL(rd)]                 ; xmm3 += round value
+        psraw       mm3,        VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       mm4,        [GLOBAL(rd)]                 ;
+        psraw       mm4,        VP8_FILTER_SHIFT        ;
+
+        movq        mm7,        mm3                 ;
+        packuswb    mm7,        mm4                 ;
+
+        add         rsi,        rdx                 ; next line
+.next_row_8x8:
+        movq        mm3,        [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+        movq        mm4,        mm3                 ; make a copy of current line
+
+        punpcklbw   mm3,        mm0                 ; xx 00 01 02 03 04 05 06
+        punpckhbw   mm4,        mm0                 ;
+
+        pmullw      mm3,        mm1                 ;
+        pmullw      mm4,        mm1                 ;
+
+        movq        mm5,        [rsi+1]             ;
+        movq        mm6,        mm5                 ;
+
+        punpcklbw   mm5,        mm0                 ;
+        punpckhbw   mm6,        mm0                 ;
+
+        pmullw      mm5,        mm2                 ;
+        pmullw      mm6,        mm2                 ;
+
+        paddw       mm3,        mm5                 ;
+        paddw       mm4,        mm6                 ;
+
+        movq        mm5,        mm7                 ;
+        movq        mm6,        mm7                 ;
+
+        punpcklbw   mm5,        mm0                 ;
+        punpckhbw   mm6,        mm0
+
+        pmullw      mm5,        [rax]               ;
+        pmullw      mm6,        [rax]               ;
+
+        paddw       mm3,        [GLOBAL(rd)]                 ; xmm3 += round value
+        psraw       mm3,        VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       mm4,        [GLOBAL(rd)]                 ;
+        psraw       mm4,        VP8_FILTER_SHIFT        ;
+
+        movq        mm7,        mm3                 ;
+        packuswb    mm7,        mm4                 ;
+
+
+        pmullw      mm3,        [rax+16]            ;
+        pmullw      mm4,        [rax+16]            ;
+
+        paddw       mm3,        mm5                 ;
+        paddw       mm4,        mm6                 ;
+
+
+        paddw       mm3,        [GLOBAL(rd)]                 ; xmm3 += round value
+        psraw       mm3,        VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       mm4,        [GLOBAL(rd)]                 ;
+        psraw       mm4,        VP8_FILTER_SHIFT        ;
+
+        packuswb    mm3,        mm4
+
+        movq        [rdi],      mm3                 ; store the results in the destination
+
+%if ABI_IS_32BIT
+        add         rsi,        rdx                 ; next line
+        add         rdi,        dword ptr arg(5) ;dst_pitch                   ;
+%else
+        movsxd      r8,         dword ptr arg(5) ;dst_pitch
+        add         rsi,        rdx                 ; next line
+        add         rdi,        r8                  ;dst_pitch
+%endif
+        cmp         rdi,        rcx                 ;
+        jne         .next_row_8x8
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void bilinear_predict8x4_mmx
+;(
+;    unsigned char  *src_ptr,
+;    int   src_pixels_per_line,
+;    int  xoffset,
+;    int  yoffset,
+;    unsigned char *dst_ptr,
+;    int dst_pitch
+;)
+global sym(vp8_bilinear_predict8x4_mmx) PRIVATE
+sym(vp8_bilinear_predict8x4_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ;const short *HFilter = vp8_bilinear_filters_x86_8[xoffset];
+    ;const short *VFilter = vp8_bilinear_filters_x86_8[yoffset];
+
+        movsxd      rax,        dword ptr arg(2) ;xoffset
+        mov         rdi,        arg(4) ;dst_ptr           ;
+
+        lea         rcx,        [GLOBAL(sym(vp8_bilinear_filters_x86_8))]
+        shl         rax,        5
+
+        mov         rsi,        arg(0) ;src_ptr              ;
+        add         rax,        rcx
+
+        movsxd      rdx,        dword ptr arg(5) ;dst_pitch
+        movq        mm1,        [rax]               ;
+
+        movq        mm2,        [rax+16]            ;
+        movsxd      rax,        dword ptr arg(3) ;yoffset
+
+        pxor        mm0,        mm0                 ;
+        shl         rax,        5
+
+        add         rax,        rcx
+        lea         rcx,        [rdi+rdx*4]          ;
+
+        movsxd      rdx,        dword ptr arg(1) ;src_pixels_per_line    ;
+
+        ; get the first horizontal line done       ;
+        movq        mm3,        [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+        movq        mm4,        mm3                 ; make a copy of current line
+
+        punpcklbw   mm3,        mm0                 ; xx 00 01 02 03 04 05 06
+        punpckhbw   mm4,        mm0                 ;
+
+        pmullw      mm3,        mm1                 ;
+        pmullw      mm4,        mm1                 ;
+
+        movq        mm5,        [rsi+1]             ;
+        movq        mm6,        mm5                 ;
+
+        punpcklbw   mm5,        mm0                 ;
+        punpckhbw   mm6,        mm0                 ;
+
+        pmullw      mm5,        mm2                 ;
+        pmullw      mm6,        mm2                 ;
+
+        paddw       mm3,        mm5                 ;
+        paddw       mm4,        mm6                 ;
+
+        paddw       mm3,        [GLOBAL(rd)]                 ; xmm3 += round value
+        psraw       mm3,        VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       mm4,        [GLOBAL(rd)]                 ;
+        psraw       mm4,        VP8_FILTER_SHIFT        ;
+
+        movq        mm7,        mm3                 ;
+        packuswb    mm7,        mm4                 ;
+
+        add         rsi,        rdx                 ; next line
+.next_row_8x4:
+        movq        mm3,        [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+        movq        mm4,        mm3                 ; make a copy of current line
+
+        punpcklbw   mm3,        mm0                 ; xx 00 01 02 03 04 05 06
+        punpckhbw   mm4,        mm0                 ;
+
+        pmullw      mm3,        mm1                 ;
+        pmullw      mm4,        mm1                 ;
+
+        movq        mm5,        [rsi+1]             ;
+        movq        mm6,        mm5                 ;
+
+        punpcklbw   mm5,        mm0                 ;
+        punpckhbw   mm6,        mm0                 ;
+
+        pmullw      mm5,        mm2                 ;
+        pmullw      mm6,        mm2                 ;
+
+        paddw       mm3,        mm5                 ;
+        paddw       mm4,        mm6                 ;
+
+        movq        mm5,        mm7                 ;
+        movq        mm6,        mm7                 ;
+
+        punpcklbw   mm5,        mm0                 ;
+        punpckhbw   mm6,        mm0
+
+        pmullw      mm5,        [rax]               ;
+        pmullw      mm6,        [rax]               ;
+
+        paddw       mm3,        [GLOBAL(rd)]                 ; xmm3 += round value
+        psraw       mm3,        VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       mm4,        [GLOBAL(rd)]                 ;
+        psraw       mm4,        VP8_FILTER_SHIFT        ;
+
+        movq        mm7,        mm3                 ;
+        packuswb    mm7,        mm4                 ;
+
+
+        pmullw      mm3,        [rax+16]            ;
+        pmullw      mm4,        [rax+16]            ;
+
+        paddw       mm3,        mm5                 ;
+        paddw       mm4,        mm6                 ;
+
+
+        paddw       mm3,        [GLOBAL(rd)]                 ; xmm3 += round value
+        psraw       mm3,        VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       mm4,        [GLOBAL(rd)]                 ;
+        psraw       mm4,        VP8_FILTER_SHIFT        ;
+
+        packuswb    mm3,        mm4
+
+        movq        [rdi],      mm3                 ; store the results in the destination
+
+%if ABI_IS_32BIT
+        add         rsi,        rdx                 ; next line
+        add         rdi,        dword ptr arg(5) ;dst_pitch                   ;
+%else
+        movsxd      r8,         dword ptr arg(5) ;dst_pitch
+        add         rsi,        rdx                 ; next line
+        add         rdi,        r8
+%endif
+        cmp         rdi,        rcx                 ;
+        jne         .next_row_8x4
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void bilinear_predict4x4_mmx
+;(
+;    unsigned char  *src_ptr,
+;    int   src_pixels_per_line,
+;    int  xoffset,
+;    int  yoffset,
+;    unsigned char *dst_ptr,
+;    int dst_pitch
+;)
+global sym(vp8_bilinear_predict4x4_mmx) PRIVATE
+sym(vp8_bilinear_predict4x4_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ;const short *HFilter = vp8_bilinear_filters_x86_8[xoffset];
+    ;const short *VFilter = vp8_bilinear_filters_x86_8[yoffset];
+
+        movsxd      rax,        dword ptr arg(2) ;xoffset
+        mov         rdi,        arg(4) ;dst_ptr           ;
+
+        lea         rcx,        [GLOBAL(sym(vp8_bilinear_filters_x86_8))]
+        shl         rax,        5
+
+        add         rax,        rcx ; HFilter
+        mov         rsi,        arg(0) ;src_ptr              ;
+
+        movsxd      rdx,        dword ptr arg(5) ;ldst_pitch
+        movq        mm1,        [rax]               ;
+
+        movq        mm2,        [rax+16]            ;
+        movsxd      rax,        dword ptr arg(3) ;yoffset
+
+        pxor        mm0,        mm0                 ;
+        shl         rax,        5
+
+        add         rax,        rcx
+        lea         rcx,        [rdi+rdx*4]          ;
+
+        movsxd      rdx,        dword ptr arg(1) ;src_pixels_per_line    ;
+
+        ; get the first horizontal line done       ;
+        movd        mm3,        [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+        punpcklbw   mm3,        mm0                 ; xx 00 01 02 03 04 05 06
+
+        pmullw      mm3,        mm1                 ;
+        movd        mm5,        [rsi+1]             ;
+
+        punpcklbw   mm5,        mm0                 ;
+        pmullw      mm5,        mm2                 ;
+
+        paddw       mm3,        mm5                 ;
+        paddw       mm3,        [GLOBAL(rd)]                 ; xmm3 += round value
+
+        psraw       mm3,        VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        movq        mm7,        mm3                 ;
+        packuswb    mm7,        mm0                 ;
+
+        add         rsi,        rdx                 ; next line
+.next_row_4x4:
+        movd        mm3,        [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+        punpcklbw   mm3,        mm0                 ; xx 00 01 02 03 04 05 06
+
+        pmullw      mm3,        mm1                 ;
+        movd        mm5,        [rsi+1]             ;
+
+        punpcklbw   mm5,        mm0                 ;
+        pmullw      mm5,        mm2                 ;
+
+        paddw       mm3,        mm5                 ;
+
+        movq        mm5,        mm7                 ;
+        punpcklbw   mm5,        mm0                 ;
+
+        pmullw      mm5,        [rax]               ;
+        paddw       mm3,        [GLOBAL(rd)]                 ; xmm3 += round value
+
+        psraw       mm3,        VP8_FILTER_SHIFT        ; xmm3 /= 128
+        movq        mm7,        mm3                 ;
+
+        packuswb    mm7,        mm0                 ;
+
+        pmullw      mm3,        [rax+16]            ;
+        paddw       mm3,        mm5                 ;
+
+
+        paddw       mm3,        [GLOBAL(rd)]                 ; xmm3 += round value
+        psraw       mm3,        VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        packuswb    mm3,        mm0
+        movd        [rdi],      mm3                 ; store the results in the destination
+
+%if ABI_IS_32BIT
+        add         rsi,        rdx                 ; next line
+        add         rdi,        dword ptr arg(5) ;dst_pitch                   ;
+%else
+        movsxd      r8,         dword ptr arg(5) ;dst_pitch                   ;
+        add         rsi,        rdx                 ; next line
+        add         rdi,        r8
+%endif
+
+        cmp         rdi,        rcx                 ;
+        jne         .next_row_4x4
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+
+SECTION_RODATA
+align 16
+rd:
+    times 4 dw 0x40
+
+align 16
+global HIDDEN_DATA(sym(vp8_six_tap_mmx))
+sym(vp8_six_tap_mmx):
+    times 8 dw 0
+    times 8 dw 0
+    times 8 dw 128
+    times 8 dw 0
+    times 8 dw 0
+    times 8 dw 0
+
+    times 8 dw 0
+    times 8 dw -6
+    times 8 dw 123
+    times 8 dw 12
+    times 8 dw -1
+    times 8 dw 0
+
+    times 8 dw 2
+    times 8 dw -11
+    times 8 dw 108
+    times 8 dw 36
+    times 8 dw -8
+    times 8 dw 1
+
+    times 8 dw 0
+    times 8 dw -9
+    times 8 dw 93
+    times 8 dw 50
+    times 8 dw -6
+    times 8 dw 0
+
+    times 8 dw 3
+    times 8 dw -16
+    times 8 dw 77
+    times 8 dw 77
+    times 8 dw -16
+    times 8 dw 3
+
+    times 8 dw 0
+    times 8 dw -6
+    times 8 dw 50
+    times 8 dw 93
+    times 8 dw -9
+    times 8 dw 0
+
+    times 8 dw 1
+    times 8 dw -8
+    times 8 dw 36
+    times 8 dw 108
+    times 8 dw -11
+    times 8 dw 2
+
+    times 8 dw 0
+    times 8 dw -1
+    times 8 dw 12
+    times 8 dw 123
+    times 8 dw -6
+    times 8 dw 0
+
+
diff --git a/libvpx/libvpx/vp8/common/x86/subpixel_sse2.asm b/libvpx/libvpx/vp8/common/x86/subpixel_sse2.asm
new file mode 100644
index 0000000..69f8d10
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/subpixel_sse2.asm
@@ -0,0 +1,1372 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+extern sym(vp8_bilinear_filters_x86_8)
+
+%define BLOCK_HEIGHT_WIDTH 4
+%define VP8_FILTER_WEIGHT 128
+%define VP8_FILTER_SHIFT  7
+
+
+;/************************************************************************************
+; Notes: filter_block1d_h6 applies a 6 tap filter horizontally to the input pixels. The
+; input pixel array has output_height rows. This routine assumes that output_height is an
+; even number. This function handles 8 pixels in horizontal direction, calculating ONE
+; rows each iteration to take advantage of the 128 bits operations.
+;*************************************************************************************/
+;void vp8_filter_block1d8_h6_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned short *output_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned int    pixel_step,
+;    unsigned int    output_height,
+;    unsigned int    output_width,
+;    short           *vp8_filter
+;)
+global sym(vp8_filter_block1d8_h6_sse2) PRIVATE
+sym(vp8_filter_block1d8_h6_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rdx,        arg(6) ;vp8_filter
+        mov         rsi,        arg(0) ;src_ptr
+
+        mov         rdi,        arg(1) ;output_ptr
+
+        movsxd      rcx,        dword ptr arg(4) ;output_height
+        movsxd      rax,        dword ptr arg(2) ;src_pixels_per_line            ; Pitch for Source
+%if ABI_IS_32BIT=0
+        movsxd      r8,         dword ptr arg(5) ;output_width
+%endif
+        pxor        xmm0,       xmm0                        ; clear xmm0 for unpack
+
+.filter_block1d8_h6_rowloop:
+        movq        xmm3,       MMWORD PTR [rsi - 2]
+        movq        xmm1,       MMWORD PTR [rsi + 6]
+
+        prefetcht2  [rsi+rax-2]
+
+        pslldq      xmm1,       8
+        por         xmm1,       xmm3
+
+        movdqa      xmm4,       xmm1
+        movdqa      xmm5,       xmm1
+
+        movdqa      xmm6,       xmm1
+        movdqa      xmm7,       xmm1
+
+        punpcklbw   xmm3,       xmm0                        ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
+        psrldq      xmm4,       1                           ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
+
+        pmullw      xmm3,       XMMWORD PTR [rdx]           ; x[-2] * H[-2]; Tap 1
+        punpcklbw   xmm4,       xmm0                        ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
+
+        psrldq      xmm5,       2                           ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
+        pmullw      xmm4,       XMMWORD PTR [rdx+16]        ; x[-1] * H[-1]; Tap 2
+
+
+        punpcklbw   xmm5,       xmm0                        ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
+        psrldq      xmm6,       3                           ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
+
+        pmullw      xmm5,       [rdx+32]                    ; x[ 0] * H[ 0]; Tap 3
+
+        punpcklbw   xmm6,       xmm0                        ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
+        psrldq      xmm7,       4                           ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
+
+        pmullw      xmm6,       [rdx+48]                    ; x[ 1] * h[ 1] ; Tap 4
+
+        punpcklbw   xmm7,       xmm0                        ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
+        psrldq      xmm1,       5                           ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
+
+
+        pmullw      xmm7,       [rdx+64]                    ; x[ 2] * h[ 2] ; Tap 5
+
+        punpcklbw   xmm1,       xmm0                        ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
+        pmullw      xmm1,       [rdx+80]                    ; x[ 3] * h[ 3] ; Tap 6
+
+
+        paddsw      xmm4,       xmm7
+        paddsw      xmm4,       xmm5
+
+        paddsw      xmm4,       xmm3
+        paddsw      xmm4,       xmm6
+
+        paddsw      xmm4,       xmm1
+        paddsw      xmm4,       [GLOBAL(rd)]
+
+        psraw       xmm4,       7
+
+        packuswb    xmm4,       xmm0
+        punpcklbw   xmm4,       xmm0
+
+        movdqa      XMMWORD Ptr [rdi],         xmm4
+        lea         rsi,        [rsi + rax]
+
+%if ABI_IS_32BIT
+        add         rdi,        DWORD Ptr arg(5) ;[output_width]
+%else
+        add         rdi,        r8
+%endif
+        dec         rcx
+
+        jnz         .filter_block1d8_h6_rowloop                ; next row
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_filter_block1d16_h6_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned short *output_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned int    pixel_step,
+;    unsigned int    output_height,
+;    unsigned int    output_width,
+;    short           *vp8_filter
+;)
+;/************************************************************************************
+; Notes: filter_block1d_h6 applies a 6 tap filter horizontally to the input pixels. The
+; input pixel array has output_height rows. This routine assumes that output_height is an
+; even number. This function handles 8 pixels in horizontal direction, calculating ONE
+; rows each iteration to take advantage of the 128 bits operations.
+;*************************************************************************************/
+global sym(vp8_filter_block1d16_h6_sse2) PRIVATE
+sym(vp8_filter_block1d16_h6_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rdx,        arg(6) ;vp8_filter
+        mov         rsi,        arg(0) ;src_ptr
+
+        mov         rdi,        arg(1) ;output_ptr
+
+        movsxd      rcx,        dword ptr arg(4) ;output_height
+        movsxd      rax,        dword ptr arg(2) ;src_pixels_per_line            ; Pitch for Source
+%if ABI_IS_32BIT=0
+        movsxd      r8,         dword ptr arg(5) ;output_width
+%endif
+
+        pxor        xmm0,       xmm0                        ; clear xmm0 for unpack
+
+.filter_block1d16_h6_sse2_rowloop:
+        movq        xmm3,       MMWORD PTR [rsi - 2]
+        movq        xmm1,       MMWORD PTR [rsi + 6]
+
+        movq        xmm2,       MMWORD PTR [rsi +14]
+        pslldq      xmm2,       8
+
+        por         xmm2,       xmm1
+        prefetcht2  [rsi+rax-2]
+
+        pslldq      xmm1,       8
+        por         xmm1,       xmm3
+
+        movdqa      xmm4,       xmm1
+        movdqa      xmm5,       xmm1
+
+        movdqa      xmm6,       xmm1
+        movdqa      xmm7,       xmm1
+
+        punpcklbw   xmm3,       xmm0                        ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
+        psrldq      xmm4,       1                           ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
+
+        pmullw      xmm3,       XMMWORD PTR [rdx]           ; x[-2] * H[-2]; Tap 1
+        punpcklbw   xmm4,       xmm0                        ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
+
+        psrldq      xmm5,       2                           ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
+        pmullw      xmm4,       XMMWORD PTR [rdx+16]        ; x[-1] * H[-1]; Tap 2
+
+
+        punpcklbw   xmm5,       xmm0                        ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
+        psrldq      xmm6,       3                           ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
+
+        pmullw      xmm5,       [rdx+32]                    ; x[ 0] * H[ 0]; Tap 3
+
+        punpcklbw   xmm6,       xmm0                        ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
+        psrldq      xmm7,       4                           ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
+
+        pmullw      xmm6,       [rdx+48]                    ; x[ 1] * h[ 1] ; Tap 4
+
+        punpcklbw   xmm7,       xmm0                        ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
+        psrldq      xmm1,       5                           ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
+
+
+        pmullw      xmm7,       [rdx+64]                    ; x[ 2] * h[ 2] ; Tap 5
+
+        punpcklbw   xmm1,       xmm0                        ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
+        pmullw      xmm1,       [rdx+80]                    ; x[ 3] * h[ 3] ; Tap 6
+
+        paddsw      xmm4,       xmm7
+        paddsw      xmm4,       xmm5
+
+        paddsw      xmm4,       xmm3
+        paddsw      xmm4,       xmm6
+
+        paddsw      xmm4,       xmm1
+        paddsw      xmm4,       [GLOBAL(rd)]
+
+        psraw       xmm4,       7
+
+        packuswb    xmm4,       xmm0
+        punpcklbw   xmm4,       xmm0
+
+        movdqa      XMMWORD Ptr [rdi],         xmm4
+
+        movdqa      xmm3,       xmm2
+        movdqa      xmm4,       xmm2
+
+        movdqa      xmm5,       xmm2
+        movdqa      xmm6,       xmm2
+
+        movdqa      xmm7,       xmm2
+
+        punpcklbw   xmm3,       xmm0                        ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
+        psrldq      xmm4,       1                           ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
+
+        pmullw      xmm3,       XMMWORD PTR [rdx]           ; x[-2] * H[-2]; Tap 1
+        punpcklbw   xmm4,       xmm0                        ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
+
+        psrldq      xmm5,       2                           ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
+        pmullw      xmm4,       XMMWORD PTR [rdx+16]        ; x[-1] * H[-1]; Tap 2
+
+
+        punpcklbw   xmm5,       xmm0                        ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
+        psrldq      xmm6,       3                           ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
+
+        pmullw      xmm5,       [rdx+32]                    ; x[ 0] * H[ 0]; Tap 3
+
+        punpcklbw   xmm6,       xmm0                        ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
+        psrldq      xmm7,       4                           ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
+
+        pmullw      xmm6,       [rdx+48]                    ; x[ 1] * h[ 1] ; Tap 4
+
+        punpcklbw   xmm7,       xmm0                        ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
+        psrldq      xmm2,       5                           ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
+
+        pmullw      xmm7,       [rdx+64]                    ; x[ 2] * h[ 2] ; Tap 5
+
+        punpcklbw   xmm2,       xmm0                        ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
+        pmullw      xmm2,       [rdx+80]                    ; x[ 3] * h[ 3] ; Tap 6
+
+
+        paddsw      xmm4,       xmm7
+        paddsw      xmm4,       xmm5
+
+        paddsw      xmm4,       xmm3
+        paddsw      xmm4,       xmm6
+
+        paddsw      xmm4,       xmm2
+        paddsw      xmm4,       [GLOBAL(rd)]
+
+        psraw       xmm4,       7
+
+        packuswb    xmm4,       xmm0
+        punpcklbw   xmm4,       xmm0
+
+        movdqa      XMMWORD Ptr [rdi+16],      xmm4
+
+        lea         rsi,        [rsi + rax]
+%if ABI_IS_32BIT
+        add         rdi,        DWORD Ptr arg(5) ;[output_width]
+%else
+        add         rdi,        r8
+%endif
+
+        dec         rcx
+        jnz         .filter_block1d16_h6_sse2_rowloop                ; next row
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_filter_block1d8_v6_sse2
+;(
+;    short *src_ptr,
+;    unsigned char *output_ptr,
+;    int dst_ptich,
+;    unsigned int pixels_per_line,
+;    unsigned int pixel_step,
+;    unsigned int output_height,
+;    unsigned int output_width,
+;    short * vp8_filter
+;)
+;/************************************************************************************
+; Notes: filter_block1d8_v6 applies a 6 tap filter vertically to the input pixels. The
+; input pixel array has output_height rows.
+;*************************************************************************************/
+global sym(vp8_filter_block1d8_v6_sse2) PRIVATE
+sym(vp8_filter_block1d8_v6_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 8
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rax,        arg(7) ;vp8_filter
+        movsxd      rdx,        dword ptr arg(3) ;pixels_per_line
+
+        mov         rdi,        arg(1) ;output_ptr
+        mov         rsi,        arg(0) ;src_ptr
+
+        sub         rsi,        rdx
+        sub         rsi,        rdx
+
+        movsxd      rcx,        DWORD PTR arg(5) ;[output_height]
+        pxor        xmm0,       xmm0                        ; clear xmm0
+
+        movdqa      xmm7,       XMMWORD PTR [GLOBAL(rd)]
+%if ABI_IS_32BIT=0
+        movsxd      r8,         dword ptr arg(2) ; dst_ptich
+%endif
+
+.vp8_filter_block1d8_v6_sse2_loop:
+        movdqa      xmm1,       XMMWORD PTR [rsi]
+        pmullw      xmm1,       [rax]
+
+        movdqa      xmm2,       XMMWORD PTR [rsi + rdx]
+        pmullw      xmm2,       [rax + 16]
+
+        movdqa      xmm3,       XMMWORD PTR [rsi + rdx * 2]
+        pmullw      xmm3,       [rax + 32]
+
+        movdqa      xmm5,       XMMWORD PTR [rsi + rdx * 4]
+        pmullw      xmm5,       [rax + 64]
+
+        add         rsi,        rdx
+        movdqa      xmm4,       XMMWORD PTR [rsi + rdx * 2]
+
+        pmullw      xmm4,       [rax + 48]
+        movdqa      xmm6,       XMMWORD PTR [rsi + rdx * 4]
+
+        pmullw      xmm6,       [rax + 80]
+
+        paddsw      xmm2,       xmm5
+        paddsw      xmm2,       xmm3
+
+        paddsw      xmm2,       xmm1
+        paddsw      xmm2,       xmm4
+
+        paddsw      xmm2,       xmm6
+        paddsw      xmm2,       xmm7
+
+        psraw       xmm2,       7
+        packuswb    xmm2,       xmm0              ; pack and saturate
+
+        movq        QWORD PTR [rdi], xmm2         ; store the results in the destination
+%if ABI_IS_32BIT
+        add         rdi,        DWORD PTR arg(2) ;[dst_ptich]
+%else
+        add         rdi,        r8
+%endif
+        dec         rcx         ; decrement count
+        jnz         .vp8_filter_block1d8_v6_sse2_loop               ; next row
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_filter_block1d16_v6_sse2
+;(
+;    unsigned short *src_ptr,
+;    unsigned char *output_ptr,
+;    int dst_ptich,
+;    unsigned int pixels_per_line,
+;    unsigned int pixel_step,
+;    unsigned int output_height,
+;    unsigned int output_width,
+;    const short    *vp8_filter
+;)
+;/************************************************************************************
+; Notes: filter_block1d16_v6 applies a 6 tap filter vertically to the input pixels. The
+; input pixel array has output_height rows.
+;*************************************************************************************/
+global sym(vp8_filter_block1d16_v6_sse2) PRIVATE
+sym(vp8_filter_block1d16_v6_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 8
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rax,        arg(7) ;vp8_filter
+        movsxd      rdx,        dword ptr arg(3) ;pixels_per_line
+
+        mov         rdi,        arg(1) ;output_ptr
+        mov         rsi,        arg(0) ;src_ptr
+
+        sub         rsi,        rdx
+        sub         rsi,        rdx
+
+        movsxd      rcx,        DWORD PTR arg(5) ;[output_height]
+%if ABI_IS_32BIT=0
+        movsxd      r8,         dword ptr arg(2) ; dst_ptich
+%endif
+
+.vp8_filter_block1d16_v6_sse2_loop:
+; The order for adding 6-tap is 2 5 3 1 4 6. Read in data in that order.
+        movdqa      xmm1,       XMMWORD PTR [rsi + rdx]       ; line 2
+        movdqa      xmm2,       XMMWORD PTR [rsi + rdx + 16]
+        pmullw      xmm1,       [rax + 16]
+        pmullw      xmm2,       [rax + 16]
+
+        movdqa      xmm3,       XMMWORD PTR [rsi + rdx * 4]       ; line 5
+        movdqa      xmm4,       XMMWORD PTR [rsi + rdx * 4 + 16]
+        pmullw      xmm3,       [rax + 64]
+        pmullw      xmm4,       [rax + 64]
+
+        movdqa      xmm5,       XMMWORD PTR [rsi + rdx * 2]       ; line 3
+        movdqa      xmm6,       XMMWORD PTR [rsi + rdx * 2 + 16]
+        pmullw      xmm5,       [rax + 32]
+        pmullw      xmm6,       [rax + 32]
+
+        movdqa      xmm7,       XMMWORD PTR [rsi]       ; line 1
+        movdqa      xmm0,       XMMWORD PTR [rsi + 16]
+        pmullw      xmm7,       [rax]
+        pmullw      xmm0,       [rax]
+
+        paddsw      xmm1,       xmm3
+        paddsw      xmm2,       xmm4
+        paddsw      xmm1,       xmm5
+        paddsw      xmm2,       xmm6
+        paddsw      xmm1,       xmm7
+        paddsw      xmm2,       xmm0
+
+        add         rsi,        rdx
+
+        movdqa      xmm3,       XMMWORD PTR [rsi + rdx * 2]       ; line 4
+        movdqa      xmm4,       XMMWORD PTR [rsi + rdx * 2 + 16]
+        pmullw      xmm3,       [rax + 48]
+        pmullw      xmm4,       [rax + 48]
+
+        movdqa      xmm5,       XMMWORD PTR [rsi + rdx * 4]       ; line 6
+        movdqa      xmm6,       XMMWORD PTR [rsi + rdx * 4 + 16]
+        pmullw      xmm5,       [rax + 80]
+        pmullw      xmm6,       [rax + 80]
+
+        movdqa      xmm7,       XMMWORD PTR [GLOBAL(rd)]
+        pxor        xmm0,       xmm0                        ; clear xmm0
+
+        paddsw      xmm1,       xmm3
+        paddsw      xmm2,       xmm4
+        paddsw      xmm1,       xmm5
+        paddsw      xmm2,       xmm6
+
+        paddsw      xmm1,       xmm7
+        paddsw      xmm2,       xmm7
+
+        psraw       xmm1,       7
+        psraw       xmm2,       7
+
+        packuswb    xmm1,       xmm2              ; pack and saturate
+        movdqa      XMMWORD PTR [rdi], xmm1       ; store the results in the destination
+%if ABI_IS_32BIT
+        add         rdi,        DWORD PTR arg(2) ;[dst_ptich]
+%else
+        add         rdi,        r8
+%endif
+        dec         rcx         ; decrement count
+        jnz         .vp8_filter_block1d16_v6_sse2_loop              ; next row
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_filter_block1d8_h6_only_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    int dst_ptich,
+;    unsigned int    output_height,
+;    const short    *vp8_filter
+;)
+; First-pass filter only when yoffset==0
+global sym(vp8_filter_block1d8_h6_only_sse2) PRIVATE
+sym(vp8_filter_block1d8_h6_only_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rdx,        arg(5) ;vp8_filter
+        mov         rsi,        arg(0) ;src_ptr
+
+        mov         rdi,        arg(2) ;output_ptr
+
+        movsxd      rcx,        dword ptr arg(4) ;output_height
+        movsxd      rax,        dword ptr arg(1) ;src_pixels_per_line            ; Pitch for Source
+%if ABI_IS_32BIT=0
+        movsxd      r8,         dword ptr arg(3) ;dst_ptich
+%endif
+        pxor        xmm0,       xmm0                        ; clear xmm0 for unpack
+
+.filter_block1d8_h6_only_rowloop:
+        movq        xmm3,       MMWORD PTR [rsi - 2]
+        movq        xmm1,       MMWORD PTR [rsi + 6]
+
+        prefetcht2  [rsi+rax-2]
+
+        pslldq      xmm1,       8
+        por         xmm1,       xmm3
+
+        movdqa      xmm4,       xmm1
+        movdqa      xmm5,       xmm1
+
+        movdqa      xmm6,       xmm1
+        movdqa      xmm7,       xmm1
+
+        punpcklbw   xmm3,       xmm0                        ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
+        psrldq      xmm4,       1                           ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
+
+        pmullw      xmm3,       XMMWORD PTR [rdx]           ; x[-2] * H[-2]; Tap 1
+        punpcklbw   xmm4,       xmm0                        ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
+
+        psrldq      xmm5,       2                           ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
+        pmullw      xmm4,       XMMWORD PTR [rdx+16]        ; x[-1] * H[-1]; Tap 2
+
+
+        punpcklbw   xmm5,       xmm0                        ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
+        psrldq      xmm6,       3                           ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
+
+        pmullw      xmm5,       [rdx+32]                    ; x[ 0] * H[ 0]; Tap 3
+
+        punpcklbw   xmm6,       xmm0                        ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
+        psrldq      xmm7,       4                           ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
+
+        pmullw      xmm6,       [rdx+48]                    ; x[ 1] * h[ 1] ; Tap 4
+
+        punpcklbw   xmm7,       xmm0                        ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
+        psrldq      xmm1,       5                           ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
+
+
+        pmullw      xmm7,       [rdx+64]                    ; x[ 2] * h[ 2] ; Tap 5
+
+        punpcklbw   xmm1,       xmm0                        ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
+        pmullw      xmm1,       [rdx+80]                    ; x[ 3] * h[ 3] ; Tap 6
+
+
+        paddsw      xmm4,       xmm7
+        paddsw      xmm4,       xmm5
+
+        paddsw      xmm4,       xmm3
+        paddsw      xmm4,       xmm6
+
+        paddsw      xmm4,       xmm1
+        paddsw      xmm4,       [GLOBAL(rd)]
+
+        psraw       xmm4,       7
+
+        packuswb    xmm4,       xmm0
+
+        movq        QWORD PTR [rdi],   xmm4       ; store the results in the destination
+        lea         rsi,        [rsi + rax]
+
+%if ABI_IS_32BIT
+        add         rdi,        DWORD Ptr arg(3) ;dst_ptich
+%else
+        add         rdi,        r8
+%endif
+        dec         rcx
+
+        jnz         .filter_block1d8_h6_only_rowloop               ; next row
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_filter_block1d16_h6_only_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    int dst_ptich,
+;    unsigned int    output_height,
+;    const short    *vp8_filter
+;)
+; First-pass filter only when yoffset==0
+global sym(vp8_filter_block1d16_h6_only_sse2) PRIVATE
+sym(vp8_filter_block1d16_h6_only_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rdx,        arg(5) ;vp8_filter
+        mov         rsi,        arg(0) ;src_ptr
+
+        mov         rdi,        arg(2) ;output_ptr
+
+        movsxd      rcx,        dword ptr arg(4) ;output_height
+        movsxd      rax,        dword ptr arg(1) ;src_pixels_per_line            ; Pitch for Source
+%if ABI_IS_32BIT=0
+        movsxd      r8,         dword ptr arg(3) ;dst_ptich
+%endif
+
+        pxor        xmm0,       xmm0                        ; clear xmm0 for unpack
+
+.filter_block1d16_h6_only_sse2_rowloop:
+        movq        xmm3,       MMWORD PTR [rsi - 2]
+        movq        xmm1,       MMWORD PTR [rsi + 6]
+
+        movq        xmm2,       MMWORD PTR [rsi +14]
+        pslldq      xmm2,       8
+
+        por         xmm2,       xmm1
+        prefetcht2  [rsi+rax-2]
+
+        pslldq      xmm1,       8
+        por         xmm1,       xmm3
+
+        movdqa      xmm4,       xmm1
+        movdqa      xmm5,       xmm1
+
+        movdqa      xmm6,       xmm1
+        movdqa      xmm7,       xmm1
+
+        punpcklbw   xmm3,       xmm0                        ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
+        psrldq      xmm4,       1                           ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
+
+        pmullw      xmm3,       XMMWORD PTR [rdx]           ; x[-2] * H[-2]; Tap 1
+        punpcklbw   xmm4,       xmm0                        ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
+
+        psrldq      xmm5,       2                           ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
+        pmullw      xmm4,       XMMWORD PTR [rdx+16]        ; x[-1] * H[-1]; Tap 2
+
+        punpcklbw   xmm5,       xmm0                        ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
+        psrldq      xmm6,       3                           ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
+
+        pmullw      xmm5,       [rdx+32]                    ; x[ 0] * H[ 0]; Tap 3
+
+        punpcklbw   xmm6,       xmm0                        ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
+        psrldq      xmm7,       4                           ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
+
+        pmullw      xmm6,       [rdx+48]                    ; x[ 1] * h[ 1] ; Tap 4
+
+        punpcklbw   xmm7,       xmm0                        ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
+        psrldq      xmm1,       5                           ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
+
+        pmullw      xmm7,       [rdx+64]                    ; x[ 2] * h[ 2] ; Tap 5
+
+        punpcklbw   xmm1,       xmm0                        ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
+        pmullw      xmm1,       [rdx+80]                    ; x[ 3] * h[ 3] ; Tap 6
+
+        paddsw      xmm4,       xmm7
+        paddsw      xmm4,       xmm5
+
+        paddsw      xmm4,       xmm3
+        paddsw      xmm4,       xmm6
+
+        paddsw      xmm4,       xmm1
+        paddsw      xmm4,       [GLOBAL(rd)]
+
+        psraw       xmm4,       7
+
+        packuswb    xmm4,       xmm0                        ; lower 8 bytes
+
+        movq        QWORD Ptr [rdi],         xmm4           ; store the results in the destination
+
+        movdqa      xmm3,       xmm2
+        movdqa      xmm4,       xmm2
+
+        movdqa      xmm5,       xmm2
+        movdqa      xmm6,       xmm2
+
+        movdqa      xmm7,       xmm2
+
+        punpcklbw   xmm3,       xmm0                        ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
+        psrldq      xmm4,       1                           ; xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1
+
+        pmullw      xmm3,       XMMWORD PTR [rdx]           ; x[-2] * H[-2]; Tap 1
+        punpcklbw   xmm4,       xmm0                        ; xx06 xx05 xx04 xx03 xx02 xx01 xx00 xx-1
+
+        psrldq      xmm5,       2                           ; xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00
+        pmullw      xmm4,       XMMWORD PTR [rdx+16]        ; x[-1] * H[-1]; Tap 2
+
+        punpcklbw   xmm5,       xmm0                        ; xx07 xx06 xx05 xx04 xx03 xx02 xx01 xx00
+        psrldq      xmm6,       3                           ; xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01
+
+        pmullw      xmm5,       [rdx+32]                    ; x[ 0] * H[ 0]; Tap 3
+
+        punpcklbw   xmm6,       xmm0                        ; xx08 xx07 xx06 xx05 xx04 xx03 xx02 xx01
+        psrldq      xmm7,       4                           ; xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03 02
+
+        pmullw      xmm6,       [rdx+48]                    ; x[ 1] * h[ 1] ; Tap 4
+
+        punpcklbw   xmm7,       xmm0                        ; xx09 xx08 xx07 xx06 xx05 xx04 xx03 xx02
+        psrldq      xmm2,       5                           ; xx xx xx xx xx 0d 0c 0b 0a 09 08 07 06 05 04 03
+
+        pmullw      xmm7,       [rdx+64]                    ; x[ 2] * h[ 2] ; Tap 5
+
+        punpcklbw   xmm2,       xmm0                        ; xx0a xx09 xx08 xx07 xx06 xx05 xx04 xx03
+        pmullw      xmm2,       [rdx+80]                    ; x[ 3] * h[ 3] ; Tap 6
+
+        paddsw      xmm4,       xmm7
+        paddsw      xmm4,       xmm5
+
+        paddsw      xmm4,       xmm3
+        paddsw      xmm4,       xmm6
+
+        paddsw      xmm4,       xmm2
+        paddsw      xmm4,       [GLOBAL(rd)]
+
+        psraw       xmm4,       7
+
+        packuswb    xmm4,       xmm0                        ; higher 8 bytes
+
+        movq        QWORD Ptr [rdi+8],      xmm4            ; store the results in the destination
+
+        lea         rsi,        [rsi + rax]
+%if ABI_IS_32BIT
+        add         rdi,        DWORD Ptr arg(3) ;dst_ptich
+%else
+        add         rdi,        r8
+%endif
+
+        dec         rcx
+        jnz         .filter_block1d16_h6_only_sse2_rowloop               ; next row
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_filter_block1d8_v6_only_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char *output_ptr,
+;    int dst_ptich,
+;    unsigned int output_height,
+;    const short    *vp8_filter
+;)
+; Second-pass filter only when xoffset==0
+global sym(vp8_filter_block1d8_v6_only_sse2) PRIVATE
+sym(vp8_filter_block1d8_v6_only_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rsi,        arg(0) ;src_ptr
+        mov         rdi,        arg(2) ;output_ptr
+
+        movsxd      rcx,        dword ptr arg(4) ;output_height
+        movsxd      rdx,        dword ptr arg(1) ;src_pixels_per_line
+
+        mov         rax,        arg(5) ;vp8_filter
+
+        pxor        xmm0,       xmm0                        ; clear xmm0
+
+        movdqa      xmm7,       XMMWORD PTR [GLOBAL(rd)]
+%if ABI_IS_32BIT=0
+        movsxd      r8,         dword ptr arg(3) ; dst_ptich
+%endif
+
+.vp8_filter_block1d8_v6_only_sse2_loop:
+        movq        xmm1,       MMWORD PTR [rsi]
+        movq        xmm2,       MMWORD PTR [rsi + rdx]
+        movq        xmm3,       MMWORD PTR [rsi + rdx * 2]
+        movq        xmm5,       MMWORD PTR [rsi + rdx * 4]
+        add         rsi,        rdx
+        movq        xmm4,       MMWORD PTR [rsi + rdx * 2]
+        movq        xmm6,       MMWORD PTR [rsi + rdx * 4]
+
+        punpcklbw   xmm1,       xmm0
+        pmullw      xmm1,       [rax]
+
+        punpcklbw   xmm2,       xmm0
+        pmullw      xmm2,       [rax + 16]
+
+        punpcklbw   xmm3,       xmm0
+        pmullw      xmm3,       [rax + 32]
+
+        punpcklbw   xmm5,       xmm0
+        pmullw      xmm5,       [rax + 64]
+
+        punpcklbw   xmm4,       xmm0
+        pmullw      xmm4,       [rax + 48]
+
+        punpcklbw   xmm6,       xmm0
+        pmullw      xmm6,       [rax + 80]
+
+        paddsw      xmm2,       xmm5
+        paddsw      xmm2,       xmm3
+
+        paddsw      xmm2,       xmm1
+        paddsw      xmm2,       xmm4
+
+        paddsw      xmm2,       xmm6
+        paddsw      xmm2,       xmm7
+
+        psraw       xmm2,       7
+        packuswb    xmm2,       xmm0              ; pack and saturate
+
+        movq        QWORD PTR [rdi], xmm2         ; store the results in the destination
+%if ABI_IS_32BIT
+        add         rdi,        DWORD PTR arg(3) ;[dst_ptich]
+%else
+        add         rdi,        r8
+%endif
+        dec         rcx         ; decrement count
+        jnz         .vp8_filter_block1d8_v6_only_sse2_loop              ; next row
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_unpack_block1d16_h6_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned short *output_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned int    output_height,
+;    unsigned int    output_width
+;)
+global sym(vp8_unpack_block1d16_h6_sse2) PRIVATE
+sym(vp8_unpack_block1d16_h6_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rsi,        arg(0) ;src_ptr
+        mov         rdi,        arg(1) ;output_ptr
+
+        movsxd      rcx,        dword ptr arg(3) ;output_height
+        movsxd      rax,        dword ptr arg(2) ;src_pixels_per_line            ; Pitch for Source
+
+        pxor        xmm0,       xmm0                        ; clear xmm0 for unpack
+%if ABI_IS_32BIT=0
+        movsxd      r8,         dword ptr arg(4) ;output_width            ; Pitch for Source
+%endif
+
+.unpack_block1d16_h6_sse2_rowloop:
+        movq        xmm1,       MMWORD PTR [rsi]            ; 0d 0c 0b 0a 09 08 07 06 05 04 03 02 01 00 -1 -2
+        movq        xmm3,       MMWORD PTR [rsi+8]          ; make copy of xmm1
+
+        punpcklbw   xmm3,       xmm0                        ; xx05 xx04 xx03 xx02 xx01 xx01 xx-1 xx-2
+        punpcklbw   xmm1,       xmm0
+
+        movdqa      XMMWORD Ptr [rdi],         xmm1
+        movdqa      XMMWORD Ptr [rdi + 16],    xmm3
+
+        lea         rsi,        [rsi + rax]
+%if ABI_IS_32BIT
+        add         rdi,        DWORD Ptr arg(4) ;[output_width]
+%else
+        add         rdi,        r8
+%endif
+        dec         rcx
+        jnz         .unpack_block1d16_h6_sse2_rowloop               ; next row
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_bilinear_predict16x16_sse2
+;(
+;    unsigned char  *src_ptr,
+;    int   src_pixels_per_line,
+;    int  xoffset,
+;    int  yoffset,
+;    unsigned char *dst_ptr,
+;    int dst_pitch
+;)
+extern sym(vp8_bilinear_filters_x86_8)
+global sym(vp8_bilinear_predict16x16_sse2) PRIVATE
+sym(vp8_bilinear_predict16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ;const short *HFilter = vp8_bilinear_filters_x86_8[xoffset]
+    ;const short *VFilter = vp8_bilinear_filters_x86_8[yoffset]
+
+        lea         rcx,        [GLOBAL(sym(vp8_bilinear_filters_x86_8))]
+        movsxd      rax,        dword ptr arg(2) ;xoffset
+
+        cmp         rax,        0      ;skip first_pass filter if xoffset=0
+        je          .b16x16_sp_only
+
+        shl         rax,        5
+        add         rax,        rcx    ;HFilter
+
+        mov         rdi,        arg(4) ;dst_ptr
+        mov         rsi,        arg(0) ;src_ptr
+        movsxd      rdx,        dword ptr arg(5) ;dst_pitch
+
+        movdqa      xmm1,       [rax]
+        movdqa      xmm2,       [rax+16]
+
+        movsxd      rax,        dword ptr arg(3) ;yoffset
+
+        cmp         rax,        0      ;skip second_pass filter if yoffset=0
+        je          .b16x16_fp_only
+
+        shl         rax,        5
+        add         rax,        rcx    ;VFilter
+
+        lea         rcx,        [rdi+rdx*8]
+        lea         rcx,        [rcx+rdx*8]
+        movsxd      rdx,        dword ptr arg(1) ;src_pixels_per_line
+
+        pxor        xmm0,       xmm0
+
+%if ABI_IS_32BIT=0
+        movsxd      r8,         dword ptr arg(5) ;dst_pitch
+%endif
+        ; get the first horizontal line done
+        movdqu      xmm3,       [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+        movdqa      xmm4,       xmm3                 ; make a copy of current line
+
+        punpcklbw   xmm3,       xmm0                 ; xx 00 01 02 03 04 05 06
+        punpckhbw   xmm4,       xmm0
+
+        pmullw      xmm3,       xmm1
+        pmullw      xmm4,       xmm1
+
+        movdqu      xmm5,       [rsi+1]
+        movdqa      xmm6,       xmm5
+
+        punpcklbw   xmm5,       xmm0
+        punpckhbw   xmm6,       xmm0
+
+        pmullw      xmm5,       xmm2
+        pmullw      xmm6,       xmm2
+
+        paddw       xmm3,       xmm5
+        paddw       xmm4,       xmm6
+
+        paddw       xmm3,       [GLOBAL(rd)]        ; xmm3 += round value
+        psraw       xmm3,       VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       xmm4,       [GLOBAL(rd)]
+        psraw       xmm4,       VP8_FILTER_SHIFT
+
+        movdqa      xmm7,       xmm3
+        packuswb    xmm7,       xmm4
+
+        add         rsi,        rdx                 ; next line
+.next_row:
+        movdqu      xmm3,       [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+        movdqa      xmm4,       xmm3                 ; make a copy of current line
+
+        punpcklbw   xmm3,       xmm0                 ; xx 00 01 02 03 04 05 06
+        punpckhbw   xmm4,       xmm0
+
+        pmullw      xmm3,       xmm1
+        pmullw      xmm4,       xmm1
+
+        movdqu      xmm5,       [rsi+1]
+        movdqa      xmm6,       xmm5
+
+        punpcklbw   xmm5,       xmm0
+        punpckhbw   xmm6,       xmm0
+
+        pmullw      xmm5,       xmm2
+        pmullw      xmm6,       xmm2
+
+        paddw       xmm3,       xmm5
+        paddw       xmm4,       xmm6
+
+        movdqa      xmm5,       xmm7
+        movdqa      xmm6,       xmm7
+
+        punpcklbw   xmm5,       xmm0
+        punpckhbw   xmm6,       xmm0
+
+        pmullw      xmm5,       [rax]
+        pmullw      xmm6,       [rax]
+
+        paddw       xmm3,       [GLOBAL(rd)]        ; xmm3 += round value
+        psraw       xmm3,       VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       xmm4,       [GLOBAL(rd)]
+        psraw       xmm4,       VP8_FILTER_SHIFT
+
+        movdqa      xmm7,       xmm3
+        packuswb    xmm7,       xmm4
+
+        pmullw      xmm3,       [rax+16]
+        pmullw      xmm4,       [rax+16]
+
+        paddw       xmm3,       xmm5
+        paddw       xmm4,       xmm6
+
+        paddw       xmm3,       [GLOBAL(rd)]        ; xmm3 += round value
+        psraw       xmm3,       VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       xmm4,       [GLOBAL(rd)]
+        psraw       xmm4,       VP8_FILTER_SHIFT
+
+        packuswb    xmm3,       xmm4
+        movdqa      [rdi],      xmm3                 ; store the results in the destination
+
+        add         rsi,        rdx                 ; next line
+%if ABI_IS_32BIT
+        add         rdi,        DWORD PTR arg(5) ;dst_pitch
+%else
+        add         rdi,        r8
+%endif
+
+        cmp         rdi,        rcx
+        jne         .next_row
+
+        jmp         .done
+
+.b16x16_sp_only:
+        movsxd      rax,        dword ptr arg(3) ;yoffset
+        shl         rax,        5
+        add         rax,        rcx    ;VFilter
+
+        mov         rdi,        arg(4) ;dst_ptr
+        mov         rsi,        arg(0) ;src_ptr
+        movsxd      rdx,        dword ptr arg(5) ;dst_pitch
+
+        movdqa      xmm1,       [rax]
+        movdqa      xmm2,       [rax+16]
+
+        lea         rcx,        [rdi+rdx*8]
+        lea         rcx,        [rcx+rdx*8]
+        movsxd      rax,        dword ptr arg(1) ;src_pixels_per_line
+
+        pxor        xmm0,       xmm0
+
+        ; get the first horizontal line done
+        movdqu      xmm7,       [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+
+        add         rsi,        rax                 ; next line
+.next_row_spo:
+        movdqu      xmm3,       [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+
+        movdqa      xmm5,       xmm7
+        movdqa      xmm6,       xmm7
+
+        movdqa      xmm4,       xmm3                 ; make a copy of current line
+        movdqa      xmm7,       xmm3
+
+        punpcklbw   xmm5,       xmm0
+        punpckhbw   xmm6,       xmm0
+        punpcklbw   xmm3,       xmm0                 ; xx 00 01 02 03 04 05 06
+        punpckhbw   xmm4,       xmm0
+
+        pmullw      xmm5,       xmm1
+        pmullw      xmm6,       xmm1
+        pmullw      xmm3,       xmm2
+        pmullw      xmm4,       xmm2
+
+        paddw       xmm3,       xmm5
+        paddw       xmm4,       xmm6
+
+        paddw       xmm3,       [GLOBAL(rd)]        ; xmm3 += round value
+        psraw       xmm3,       VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       xmm4,       [GLOBAL(rd)]
+        psraw       xmm4,       VP8_FILTER_SHIFT
+
+        packuswb    xmm3,       xmm4
+        movdqa      [rdi],      xmm3                 ; store the results in the destination
+
+        add         rsi,        rax                 ; next line
+        add         rdi,        rdx                 ;dst_pitch
+        cmp         rdi,        rcx
+        jne         .next_row_spo
+
+        jmp         .done
+
+.b16x16_fp_only:
+        lea         rcx,        [rdi+rdx*8]
+        lea         rcx,        [rcx+rdx*8]
+        movsxd      rax,        dword ptr arg(1) ;src_pixels_per_line
+        pxor        xmm0,       xmm0
+
+.next_row_fpo:
+        movdqu      xmm3,       [rsi]               ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
+        movdqa      xmm4,       xmm3                 ; make a copy of current line
+
+        punpcklbw   xmm3,       xmm0                 ; xx 00 01 02 03 04 05 06
+        punpckhbw   xmm4,       xmm0
+
+        pmullw      xmm3,       xmm1
+        pmullw      xmm4,       xmm1
+
+        movdqu      xmm5,       [rsi+1]
+        movdqa      xmm6,       xmm5
+
+        punpcklbw   xmm5,       xmm0
+        punpckhbw   xmm6,       xmm0
+
+        pmullw      xmm5,       xmm2
+        pmullw      xmm6,       xmm2
+
+        paddw       xmm3,       xmm5
+        paddw       xmm4,       xmm6
+
+        paddw       xmm3,       [GLOBAL(rd)]        ; xmm3 += round value
+        psraw       xmm3,       VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        paddw       xmm4,       [GLOBAL(rd)]
+        psraw       xmm4,       VP8_FILTER_SHIFT
+
+        packuswb    xmm3,       xmm4
+        movdqa      [rdi],      xmm3                 ; store the results in the destination
+
+        add         rsi,        rax                 ; next line
+        add         rdi,        rdx                 ; dst_pitch
+        cmp         rdi,        rcx
+        jne         .next_row_fpo
+
+.done:
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_bilinear_predict8x8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    int   src_pixels_per_line,
+;    int  xoffset,
+;    int  yoffset,
+;    unsigned char *dst_ptr,
+;    int dst_pitch
+;)
+global sym(vp8_bilinear_predict8x8_sse2) PRIVATE
+sym(vp8_bilinear_predict8x8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 144                         ; reserve 144 bytes
+
+    ;const short *HFilter = vp8_bilinear_filters_x86_8[xoffset]
+    ;const short *VFilter = vp8_bilinear_filters_x86_8[yoffset]
+        lea         rcx,        [GLOBAL(sym(vp8_bilinear_filters_x86_8))]
+
+        mov         rsi,        arg(0) ;src_ptr
+        movsxd      rdx,        dword ptr arg(1) ;src_pixels_per_line
+
+    ;Read 9-line unaligned data in and put them on stack. This gives a big
+    ;performance boost.
+        movdqu      xmm0,       [rsi]
+        lea         rax,        [rdx + rdx*2]
+        movdqu      xmm1,       [rsi+rdx]
+        movdqu      xmm2,       [rsi+rdx*2]
+        add         rsi,        rax
+        movdqu      xmm3,       [rsi]
+        movdqu      xmm4,       [rsi+rdx]
+        movdqu      xmm5,       [rsi+rdx*2]
+        add         rsi,        rax
+        movdqu      xmm6,       [rsi]
+        movdqu      xmm7,       [rsi+rdx]
+
+        movdqa      XMMWORD PTR [rsp],            xmm0
+
+        movdqu      xmm0,       [rsi+rdx*2]
+
+        movdqa      XMMWORD PTR [rsp+16],         xmm1
+        movdqa      XMMWORD PTR [rsp+32],         xmm2
+        movdqa      XMMWORD PTR [rsp+48],         xmm3
+        movdqa      XMMWORD PTR [rsp+64],         xmm4
+        movdqa      XMMWORD PTR [rsp+80],         xmm5
+        movdqa      XMMWORD PTR [rsp+96],         xmm6
+        movdqa      XMMWORD PTR [rsp+112],        xmm7
+        movdqa      XMMWORD PTR [rsp+128],        xmm0
+
+        movsxd      rax,        dword ptr arg(2) ;xoffset
+        shl         rax,        5
+        add         rax,        rcx    ;HFilter
+
+        mov         rdi,        arg(4) ;dst_ptr
+        movsxd      rdx,        dword ptr arg(5) ;dst_pitch
+
+        movdqa      xmm1,       [rax]
+        movdqa      xmm2,       [rax+16]
+
+        movsxd      rax,        dword ptr arg(3) ;yoffset
+        shl         rax,        5
+        add         rax,        rcx    ;VFilter
+
+        lea         rcx,        [rdi+rdx*8]
+
+        movdqa      xmm5,       [rax]
+        movdqa      xmm6,       [rax+16]
+
+        pxor        xmm0,       xmm0
+
+        ; get the first horizontal line done
+        movdqa      xmm3,       XMMWORD PTR [rsp]
+        movdqa      xmm4,       xmm3                 ; make a copy of current line
+        psrldq      xmm4,       1
+
+        punpcklbw   xmm3,       xmm0                 ; 00 01 02 03 04 05 06 07
+        punpcklbw   xmm4,       xmm0                 ; 01 02 03 04 05 06 07 08
+
+        pmullw      xmm3,       xmm1
+        pmullw      xmm4,       xmm2
+
+        paddw       xmm3,       xmm4
+
+        paddw       xmm3,       [GLOBAL(rd)]        ; xmm3 += round value
+        psraw       xmm3,       VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        movdqa      xmm7,       xmm3
+        add         rsp,        16                 ; next line
+.next_row8x8:
+        movdqa      xmm3,       XMMWORD PTR [rsp]               ; 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
+        movdqa      xmm4,       xmm3                 ; make a copy of current line
+        psrldq      xmm4,       1
+
+        punpcklbw   xmm3,       xmm0                 ; 00 01 02 03 04 05 06 07
+        punpcklbw   xmm4,       xmm0                 ; 01 02 03 04 05 06 07 08
+
+        pmullw      xmm3,       xmm1
+        pmullw      xmm4,       xmm2
+
+        paddw       xmm3,       xmm4
+        pmullw      xmm7,       xmm5
+
+        paddw       xmm3,       [GLOBAL(rd)]        ; xmm3 += round value
+        psraw       xmm3,       VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        movdqa      xmm4,       xmm3
+
+        pmullw      xmm3,       xmm6
+        paddw       xmm3,       xmm7
+
+        movdqa      xmm7,       xmm4
+
+        paddw       xmm3,       [GLOBAL(rd)]        ; xmm3 += round value
+        psraw       xmm3,       VP8_FILTER_SHIFT        ; xmm3 /= 128
+
+        packuswb    xmm3,       xmm0
+        movq        [rdi],      xmm3                 ; store the results in the destination
+
+        add         rsp,        16                 ; next line
+        add         rdi,        rdx
+
+        cmp         rdi,        rcx
+        jne         .next_row8x8
+
+    ;add rsp, 144
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+SECTION_RODATA
+align 16
+rd:
+    times 8 dw 0x40
diff --git a/libvpx/libvpx/vp8/common/x86/subpixel_ssse3.asm b/libvpx/libvpx/vp8/common/x86/subpixel_ssse3.asm
new file mode 100644
index 0000000..c06f245
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/subpixel_ssse3.asm
@@ -0,0 +1,1508 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%define BLOCK_HEIGHT_WIDTH 4
+%define VP8_FILTER_WEIGHT 128
+%define VP8_FILTER_SHIFT  7
+
+
+;/************************************************************************************
+; Notes: filter_block1d_h6 applies a 6 tap filter horizontally to the input pixels. The
+; input pixel array has output_height rows. This routine assumes that output_height is an
+; even number. This function handles 8 pixels in horizontal direction, calculating ONE
+; rows each iteration to take advantage of the 128 bits operations.
+;
+; This is an implementation of some of the SSE optimizations first seen in ffvp8
+;
+;*************************************************************************************/
+;void vp8_filter_block1d8_h6_ssse3
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    unsigned int    vp8_filter_index
+;)
+global sym(vp8_filter_block1d8_h6_ssse3) PRIVATE
+sym(vp8_filter_block1d8_h6_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movsxd      rdx, DWORD PTR arg(5)   ;table index
+    xor         rsi, rsi
+    shl         rdx, 4
+
+    movdqa      xmm7, [GLOBAL(rd)]
+
+    lea         rax, [GLOBAL(k0_k5)]
+    add         rax, rdx
+    mov         rdi, arg(2)             ;output_ptr
+
+    cmp         esi, DWORD PTR [rax]
+    je          vp8_filter_block1d8_h4_ssse3
+
+    movdqa      xmm4, XMMWORD PTR [rax]         ;k0_k5
+    movdqa      xmm5, XMMWORD PTR [rax+256]     ;k2_k4
+    movdqa      xmm6, XMMWORD PTR [rax+128]     ;k1_k3
+
+    mov         rsi, arg(0)             ;src_ptr
+    movsxd      rax, dword ptr arg(1)   ;src_pixels_per_line
+    movsxd      rcx, dword ptr arg(4)   ;output_height
+
+    movsxd      rdx, dword ptr arg(3)   ;output_pitch
+
+    sub         rdi, rdx
+;xmm3 free
+.filter_block1d8_h6_rowloop_ssse3:
+    movq        xmm0,   MMWORD PTR [rsi - 2]    ; -2 -1  0  1  2  3  4  5
+
+    movq        xmm2,   MMWORD PTR [rsi + 3]    ;  3  4  5  6  7  8  9 10
+
+    punpcklbw   xmm0,   xmm2                    ; -2  3 -1  4  0  5  1  6  2  7  3  8  4  9  5 10
+
+    movdqa      xmm1,   xmm0
+    pmaddubsw   xmm0,   xmm4
+
+    movdqa      xmm2,   xmm1
+    pshufb      xmm1,   [GLOBAL(shuf2bfrom1)]
+
+    pshufb      xmm2,   [GLOBAL(shuf3bfrom1)]
+    pmaddubsw   xmm1,   xmm5
+
+    lea         rdi,    [rdi + rdx]
+    pmaddubsw   xmm2,   xmm6
+
+    lea         rsi,    [rsi + rax]
+    dec         rcx
+
+    paddsw      xmm0,   xmm1
+    paddsw      xmm2,   xmm7
+
+    paddsw      xmm0,   xmm2
+
+    psraw       xmm0,   7
+
+    packuswb    xmm0,   xmm0
+
+    movq        MMWORD Ptr [rdi], xmm0
+    jnz         .filter_block1d8_h6_rowloop_ssse3
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+vp8_filter_block1d8_h4_ssse3:
+    movdqa      xmm5, XMMWORD PTR [rax+256]     ;k2_k4
+    movdqa      xmm6, XMMWORD PTR [rax+128]     ;k1_k3
+
+    movdqa      xmm3, XMMWORD PTR [GLOBAL(shuf2bfrom1)]
+    movdqa      xmm4, XMMWORD PTR [GLOBAL(shuf3bfrom1)]
+
+    mov         rsi, arg(0)             ;src_ptr
+
+    movsxd      rax, dword ptr arg(1)   ;src_pixels_per_line
+    movsxd      rcx, dword ptr arg(4)   ;output_height
+
+    movsxd      rdx, dword ptr arg(3)   ;output_pitch
+
+    sub         rdi, rdx
+
+.filter_block1d8_h4_rowloop_ssse3:
+    movq        xmm0,   MMWORD PTR [rsi - 2]    ; -2 -1  0  1  2  3  4  5
+
+    movq        xmm1,   MMWORD PTR [rsi + 3]    ;  3  4  5  6  7  8  9 10
+
+    punpcklbw   xmm0,   xmm1                    ; -2  3 -1  4  0  5  1  6  2  7  3  8  4  9  5 10
+
+    movdqa      xmm2,   xmm0
+    pshufb      xmm0,   xmm3
+
+    pshufb      xmm2,   xmm4
+    pmaddubsw   xmm0,   xmm5
+
+    lea         rdi,    [rdi + rdx]
+    pmaddubsw   xmm2,   xmm6
+
+    lea         rsi,    [rsi + rax]
+    dec         rcx
+
+    paddsw      xmm0,   xmm7
+
+    paddsw      xmm0,   xmm2
+
+    psraw       xmm0,   7
+
+    packuswb    xmm0,   xmm0
+
+    movq        MMWORD Ptr [rdi], xmm0
+
+    jnz         .filter_block1d8_h4_rowloop_ssse3
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+;void vp8_filter_block1d16_h6_ssse3
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    unsigned int    vp8_filter_index
+;)
+global sym(vp8_filter_block1d16_h6_ssse3) PRIVATE
+sym(vp8_filter_block1d16_h6_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movsxd      rdx, DWORD PTR arg(5)           ;table index
+    xor         rsi, rsi
+    shl         rdx, 4      ;
+
+    lea         rax, [GLOBAL(k0_k5)]
+    add         rax, rdx
+
+    mov         rdi, arg(2)                     ;output_ptr
+
+    mov         rsi, arg(0)                     ;src_ptr
+
+    movdqa      xmm4, XMMWORD PTR [rax]         ;k0_k5
+    movdqa      xmm5, XMMWORD PTR [rax+256]     ;k2_k4
+    movdqa      xmm6, XMMWORD PTR [rax+128]     ;k1_k3
+
+    movsxd      rax, dword ptr arg(1)           ;src_pixels_per_line
+    movsxd      rcx, dword ptr arg(4)           ;output_height
+    movsxd      rdx, dword ptr arg(3)           ;output_pitch
+
+.filter_block1d16_h6_rowloop_ssse3:
+    movq        xmm0,   MMWORD PTR [rsi - 2]    ; -2 -1  0  1  2  3  4  5
+
+    movq        xmm3,   MMWORD PTR [rsi + 3]    ;  3  4  5  6  7  8  9 10
+
+    punpcklbw   xmm0,   xmm3                    ; -2  3 -1  4  0  5  1  6  2  7  3  8  4  9  5 10
+
+    movdqa      xmm1,   xmm0
+    pmaddubsw   xmm0,   xmm4
+
+    movdqa      xmm2,   xmm1
+    pshufb      xmm1,   [GLOBAL(shuf2bfrom1)]
+
+    pshufb      xmm2,   [GLOBAL(shuf3bfrom1)]
+    movq        xmm3,   MMWORD PTR [rsi +  6]
+
+    pmaddubsw   xmm1,   xmm5
+    movq        xmm7,   MMWORD PTR [rsi + 11]
+
+    pmaddubsw   xmm2,   xmm6
+    punpcklbw   xmm3,   xmm7
+
+    paddsw      xmm0,   xmm1
+    movdqa      xmm1,   xmm3
+
+    pmaddubsw   xmm3,   xmm4
+    paddsw      xmm0,   xmm2
+
+    movdqa      xmm2,   xmm1
+    paddsw      xmm0,   [GLOBAL(rd)]
+
+    pshufb      xmm1,   [GLOBAL(shuf2bfrom1)]
+    pshufb      xmm2,   [GLOBAL(shuf3bfrom1)]
+
+    psraw       xmm0,   7
+    pmaddubsw   xmm1,   xmm5
+
+    pmaddubsw   xmm2,   xmm6
+    packuswb    xmm0,   xmm0
+
+    lea         rsi,    [rsi + rax]
+    paddsw      xmm3,   xmm1
+
+    paddsw      xmm3,   xmm2
+
+    paddsw      xmm3,   [GLOBAL(rd)]
+
+    psraw       xmm3,   7
+
+    packuswb    xmm3,   xmm3
+
+    punpcklqdq  xmm0,   xmm3
+
+    movdqa      XMMWORD Ptr [rdi], xmm0
+
+    lea         rdi,    [rdi + rdx]
+    dec         rcx
+    jnz         .filter_block1d16_h6_rowloop_ssse3
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp8_filter_block1d4_h6_ssse3
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    unsigned int    vp8_filter_index
+;)
+global sym(vp8_filter_block1d4_h6_ssse3) PRIVATE
+sym(vp8_filter_block1d4_h6_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movsxd      rdx, DWORD PTR arg(5)   ;table index
+    xor         rsi, rsi
+    shl         rdx, 4      ;
+
+    lea         rax, [GLOBAL(k0_k5)]
+    add         rax, rdx
+    movdqa      xmm7, [GLOBAL(rd)]
+
+    cmp         esi, DWORD PTR [rax]
+    je          .vp8_filter_block1d4_h4_ssse3
+
+    movdqa      xmm4, XMMWORD PTR [rax]         ;k0_k5
+    movdqa      xmm5, XMMWORD PTR [rax+256]     ;k2_k4
+    movdqa      xmm6, XMMWORD PTR [rax+128]     ;k1_k3
+
+    mov         rsi, arg(0)             ;src_ptr
+    mov         rdi, arg(2)             ;output_ptr
+    movsxd      rax, dword ptr arg(1)   ;src_pixels_per_line
+    movsxd      rcx, dword ptr arg(4)   ;output_height
+
+    movsxd      rdx, dword ptr arg(3)   ;output_pitch
+
+;xmm3 free
+.filter_block1d4_h6_rowloop_ssse3:
+    movdqu      xmm0,   XMMWORD PTR [rsi - 2]
+
+    movdqa      xmm1, xmm0
+    pshufb      xmm0, [GLOBAL(shuf1b)]
+
+    movdqa      xmm2, xmm1
+    pshufb      xmm1, [GLOBAL(shuf2b)]
+    pmaddubsw   xmm0, xmm4
+    pshufb      xmm2, [GLOBAL(shuf3b)]
+    pmaddubsw   xmm1, xmm5
+
+;--
+    pmaddubsw   xmm2, xmm6
+
+    lea         rsi,    [rsi + rax]
+;--
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm7
+    pxor        xmm1, xmm1
+    paddsw      xmm0, xmm2
+    psraw       xmm0, 7
+    packuswb    xmm0, xmm0
+
+    movd        DWORD PTR [rdi], xmm0
+
+    add         rdi, rdx
+    dec         rcx
+    jnz         .filter_block1d4_h6_rowloop_ssse3
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+.vp8_filter_block1d4_h4_ssse3:
+    movdqa      xmm5, XMMWORD PTR [rax+256]     ;k2_k4
+    movdqa      xmm6, XMMWORD PTR [rax+128]     ;k1_k3
+    movdqa      xmm0, XMMWORD PTR [GLOBAL(shuf2b)]
+    movdqa      xmm3, XMMWORD PTR [GLOBAL(shuf3b)]
+
+    mov         rsi, arg(0)             ;src_ptr
+    mov         rdi, arg(2)             ;output_ptr
+    movsxd      rax, dword ptr arg(1)   ;src_pixels_per_line
+    movsxd      rcx, dword ptr arg(4)   ;output_height
+
+    movsxd      rdx, dword ptr arg(3)   ;output_pitch
+
+.filter_block1d4_h4_rowloop_ssse3:
+    movdqu      xmm1,   XMMWORD PTR [rsi - 2]
+
+    movdqa      xmm2, xmm1
+    pshufb      xmm1, xmm0 ;;[GLOBAL(shuf2b)]
+    pshufb      xmm2, xmm3 ;;[GLOBAL(shuf3b)]
+    pmaddubsw   xmm1, xmm5
+
+;--
+    pmaddubsw   xmm2, xmm6
+
+    lea         rsi,    [rsi + rax]
+;--
+    paddsw      xmm1, xmm7
+    paddsw      xmm1, xmm2
+    psraw       xmm1, 7
+    packuswb    xmm1, xmm1
+
+    movd        DWORD PTR [rdi], xmm1
+
+    add         rdi, rdx
+    dec         rcx
+    jnz         .filter_block1d4_h4_rowloop_ssse3
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+
+;void vp8_filter_block1d16_v6_ssse3
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    unsigned int   vp8_filter_index
+;)
+global sym(vp8_filter_block1d16_v6_ssse3) PRIVATE
+sym(vp8_filter_block1d16_v6_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movsxd      rdx, DWORD PTR arg(5)   ;table index
+    xor         rsi, rsi
+    shl         rdx, 4      ;
+
+    lea         rax, [GLOBAL(k0_k5)]
+    add         rax, rdx
+
+    cmp         esi, DWORD PTR [rax]
+    je          .vp8_filter_block1d16_v4_ssse3
+
+    movdqa      xmm5, XMMWORD PTR [rax]         ;k0_k5
+    movdqa      xmm6, XMMWORD PTR [rax+256]     ;k2_k4
+    movdqa      xmm7, XMMWORD PTR [rax+128]     ;k1_k3
+
+    mov         rsi, arg(0)             ;src_ptr
+    movsxd      rdx, DWORD PTR arg(1)   ;pixels_per_line
+    mov         rdi, arg(2)             ;output_ptr
+
+%if ABI_IS_32BIT=0
+    movsxd      r8, DWORD PTR arg(3)    ;out_pitch
+%endif
+    mov         rax, rsi
+    movsxd      rcx, DWORD PTR arg(4)   ;output_height
+    add         rax, rdx
+
+
+.vp8_filter_block1d16_v6_ssse3_loop:
+    movq        xmm1, MMWORD PTR [rsi]                  ;A
+    movq        xmm2, MMWORD PTR [rsi + rdx]            ;B
+    movq        xmm3, MMWORD PTR [rsi + rdx * 2]        ;C
+    movq        xmm4, MMWORD PTR [rax + rdx * 2]        ;D
+    movq        xmm0, MMWORD PTR [rsi + rdx * 4]        ;E
+
+    punpcklbw   xmm2, xmm4                  ;B D
+    punpcklbw   xmm3, xmm0                  ;C E
+
+    movq        xmm0, MMWORD PTR [rax + rdx * 4]        ;F
+
+    pmaddubsw   xmm3, xmm6
+    punpcklbw   xmm1, xmm0                  ;A F
+    pmaddubsw   xmm2, xmm7
+    pmaddubsw   xmm1, xmm5
+
+    paddsw      xmm2, xmm3
+    paddsw      xmm2, xmm1
+    paddsw      xmm2, [GLOBAL(rd)]
+    psraw       xmm2, 7
+    packuswb    xmm2, xmm2
+
+    movq        MMWORD PTR [rdi], xmm2          ;store the results
+
+    movq        xmm1, MMWORD PTR [rsi + 8]                  ;A
+    movq        xmm2, MMWORD PTR [rsi + rdx + 8]            ;B
+    movq        xmm3, MMWORD PTR [rsi + rdx * 2 + 8]        ;C
+    movq        xmm4, MMWORD PTR [rax + rdx * 2 + 8]        ;D
+    movq        xmm0, MMWORD PTR [rsi + rdx * 4 + 8]        ;E
+
+    punpcklbw   xmm2, xmm4                  ;B D
+    punpcklbw   xmm3, xmm0                  ;C E
+
+    movq        xmm0, MMWORD PTR [rax + rdx * 4 + 8]        ;F
+    pmaddubsw   xmm3, xmm6
+    punpcklbw   xmm1, xmm0                  ;A F
+    pmaddubsw   xmm2, xmm7
+    pmaddubsw   xmm1, xmm5
+
+    add         rsi,  rdx
+    add         rax,  rdx
+;--
+;--
+    paddsw      xmm2, xmm3
+    paddsw      xmm2, xmm1
+    paddsw      xmm2, [GLOBAL(rd)]
+    psraw       xmm2, 7
+    packuswb    xmm2, xmm2
+
+    movq        MMWORD PTR [rdi+8], xmm2
+
+%if ABI_IS_32BIT
+    add         rdi,        DWORD PTR arg(3) ;out_pitch
+%else
+    add         rdi,        r8
+%endif
+    dec         rcx
+    jnz         .vp8_filter_block1d16_v6_ssse3_loop
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+.vp8_filter_block1d16_v4_ssse3:
+    movdqa      xmm6, XMMWORD PTR [rax+256]     ;k2_k4
+    movdqa      xmm7, XMMWORD PTR [rax+128]     ;k1_k3
+
+    mov         rsi, arg(0)             ;src_ptr
+    movsxd      rdx, DWORD PTR arg(1)   ;pixels_per_line
+    mov         rdi, arg(2)             ;output_ptr
+
+%if ABI_IS_32BIT=0
+    movsxd      r8, DWORD PTR arg(3)    ;out_pitch
+%endif
+    mov         rax, rsi
+    movsxd      rcx, DWORD PTR arg(4)   ;output_height
+    add         rax, rdx
+
+.vp8_filter_block1d16_v4_ssse3_loop:
+    movq        xmm2, MMWORD PTR [rsi + rdx]            ;B
+    movq        xmm3, MMWORD PTR [rsi + rdx * 2]        ;C
+    movq        xmm4, MMWORD PTR [rax + rdx * 2]        ;D
+    movq        xmm0, MMWORD PTR [rsi + rdx * 4]        ;E
+
+    punpcklbw   xmm2, xmm4                  ;B D
+    punpcklbw   xmm3, xmm0                  ;C E
+
+    pmaddubsw   xmm3, xmm6
+    pmaddubsw   xmm2, xmm7
+    movq        xmm5, MMWORD PTR [rsi + rdx + 8]            ;B
+    movq        xmm1, MMWORD PTR [rsi + rdx * 2 + 8]        ;C
+    movq        xmm4, MMWORD PTR [rax + rdx * 2 + 8]        ;D
+    movq        xmm0, MMWORD PTR [rsi + rdx * 4 + 8]        ;E
+
+    paddsw      xmm2, [GLOBAL(rd)]
+    paddsw      xmm2, xmm3
+    psraw       xmm2, 7
+    packuswb    xmm2, xmm2
+
+    punpcklbw   xmm5, xmm4                  ;B D
+    punpcklbw   xmm1, xmm0                  ;C E
+
+    pmaddubsw   xmm1, xmm6
+    pmaddubsw   xmm5, xmm7
+
+    movdqa      xmm4, [GLOBAL(rd)]
+    add         rsi,  rdx
+    add         rax,  rdx
+;--
+;--
+    paddsw      xmm5, xmm1
+    paddsw      xmm5, xmm4
+    psraw       xmm5, 7
+    packuswb    xmm5, xmm5
+
+    punpcklqdq  xmm2, xmm5
+
+    movdqa       XMMWORD PTR [rdi], xmm2
+
+%if ABI_IS_32BIT
+    add         rdi,        DWORD PTR arg(3) ;out_pitch
+%else
+    add         rdi,        r8
+%endif
+    dec         rcx
+    jnz         .vp8_filter_block1d16_v4_ssse3_loop
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp8_filter_block1d8_v6_ssse3
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    unsigned int   vp8_filter_index
+;)
+global sym(vp8_filter_block1d8_v6_ssse3) PRIVATE
+sym(vp8_filter_block1d8_v6_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movsxd      rdx, DWORD PTR arg(5)   ;table index
+    xor         rsi, rsi
+    shl         rdx, 4      ;
+
+    lea         rax, [GLOBAL(k0_k5)]
+    add         rax, rdx
+
+    movsxd      rdx, DWORD PTR arg(1)   ;pixels_per_line
+    mov         rdi, arg(2)             ;output_ptr
+%if ABI_IS_32BIT=0
+    movsxd      r8, DWORD PTR arg(3)    ; out_pitch
+%endif
+    movsxd      rcx, DWORD PTR arg(4)   ;[output_height]
+
+    cmp         esi, DWORD PTR [rax]
+    je          .vp8_filter_block1d8_v4_ssse3
+
+    movdqa      xmm5, XMMWORD PTR [rax]         ;k0_k5
+    movdqa      xmm6, XMMWORD PTR [rax+256]     ;k2_k4
+    movdqa      xmm7, XMMWORD PTR [rax+128]     ;k1_k3
+
+    mov         rsi, arg(0)             ;src_ptr
+
+    mov         rax, rsi
+    add         rax, rdx
+
+.vp8_filter_block1d8_v6_ssse3_loop:
+    movq        xmm1, MMWORD PTR [rsi]                  ;A
+    movq        xmm2, MMWORD PTR [rsi + rdx]            ;B
+    movq        xmm3, MMWORD PTR [rsi + rdx * 2]        ;C
+    movq        xmm4, MMWORD PTR [rax + rdx * 2]        ;D
+    movq        xmm0, MMWORD PTR [rsi + rdx * 4]        ;E
+
+    punpcklbw   xmm2, xmm4                  ;B D
+    punpcklbw   xmm3, xmm0                  ;C E
+
+    movq        xmm0, MMWORD PTR [rax + rdx * 4]        ;F
+    movdqa      xmm4, [GLOBAL(rd)]
+
+    pmaddubsw   xmm3, xmm6
+    punpcklbw   xmm1, xmm0                  ;A F
+    pmaddubsw   xmm2, xmm7
+    pmaddubsw   xmm1, xmm5
+    add         rsi,  rdx
+    add         rax,  rdx
+;--
+;--
+    paddsw      xmm2, xmm3
+    paddsw      xmm2, xmm1
+    paddsw      xmm2, xmm4
+    psraw       xmm2, 7
+    packuswb    xmm2, xmm2
+
+    movq        MMWORD PTR [rdi], xmm2
+
+%if ABI_IS_32BIT
+    add         rdi,        DWORD PTR arg(3) ;[out_pitch]
+%else
+    add         rdi,        r8
+%endif
+    dec         rcx
+    jnz         .vp8_filter_block1d8_v6_ssse3_loop
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+.vp8_filter_block1d8_v4_ssse3:
+    movdqa      xmm6, XMMWORD PTR [rax+256]     ;k2_k4
+    movdqa      xmm7, XMMWORD PTR [rax+128]     ;k1_k3
+    movdqa      xmm5, [GLOBAL(rd)]
+
+    mov         rsi, arg(0)             ;src_ptr
+
+    mov         rax, rsi
+    add         rax, rdx
+
+.vp8_filter_block1d8_v4_ssse3_loop:
+    movq        xmm2, MMWORD PTR [rsi + rdx]            ;B
+    movq        xmm3, MMWORD PTR [rsi + rdx * 2]        ;C
+    movq        xmm4, MMWORD PTR [rax + rdx * 2]        ;D
+    movq        xmm0, MMWORD PTR [rsi + rdx * 4]        ;E
+
+    punpcklbw   xmm2, xmm4                  ;B D
+    punpcklbw   xmm3, xmm0                  ;C E
+
+    pmaddubsw   xmm3, xmm6
+    pmaddubsw   xmm2, xmm7
+    add         rsi,  rdx
+    add         rax,  rdx
+;--
+;--
+    paddsw      xmm2, xmm3
+    paddsw      xmm2, xmm5
+    psraw       xmm2, 7
+    packuswb    xmm2, xmm2
+
+    movq        MMWORD PTR [rdi], xmm2
+
+%if ABI_IS_32BIT
+    add         rdi,        DWORD PTR arg(3) ;[out_pitch]
+%else
+    add         rdi,        r8
+%endif
+    dec         rcx
+    jnz         .vp8_filter_block1d8_v4_ssse3_loop
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+;void vp8_filter_block1d4_v6_ssse3
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    unsigned int   vp8_filter_index
+;)
+global sym(vp8_filter_block1d4_v6_ssse3) PRIVATE
+sym(vp8_filter_block1d4_v6_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movsxd      rdx, DWORD PTR arg(5)   ;table index
+    xor         rsi, rsi
+    shl         rdx, 4      ;
+
+    lea         rax, [GLOBAL(k0_k5)]
+    add         rax, rdx
+
+    movsxd      rdx, DWORD PTR arg(1)   ;pixels_per_line
+    mov         rdi, arg(2)             ;output_ptr
+%if ABI_IS_32BIT=0
+    movsxd      r8, DWORD PTR arg(3)    ; out_pitch
+%endif
+    movsxd      rcx, DWORD PTR arg(4)   ;[output_height]
+
+    cmp         esi, DWORD PTR [rax]
+    je          .vp8_filter_block1d4_v4_ssse3
+
+    movq        mm5, MMWORD PTR [rax]         ;k0_k5
+    movq        mm6, MMWORD PTR [rax+256]     ;k2_k4
+    movq        mm7, MMWORD PTR [rax+128]     ;k1_k3
+
+    mov         rsi, arg(0)             ;src_ptr
+
+    mov         rax, rsi
+    add         rax, rdx
+
+.vp8_filter_block1d4_v6_ssse3_loop:
+    movd        mm1, DWORD PTR [rsi]                  ;A
+    movd        mm2, DWORD PTR [rsi + rdx]            ;B
+    movd        mm3, DWORD PTR [rsi + rdx * 2]        ;C
+    movd        mm4, DWORD PTR [rax + rdx * 2]        ;D
+    movd        mm0, DWORD PTR [rsi + rdx * 4]        ;E
+
+    punpcklbw   mm2, mm4                  ;B D
+    punpcklbw   mm3, mm0                  ;C E
+
+    movd        mm0, DWORD PTR [rax + rdx * 4]        ;F
+
+    movq        mm4, [GLOBAL(rd)]
+
+    pmaddubsw   mm3, mm6
+    punpcklbw   mm1, mm0                  ;A F
+    pmaddubsw   mm2, mm7
+    pmaddubsw   mm1, mm5
+    add         rsi,  rdx
+    add         rax,  rdx
+;--
+;--
+    paddsw      mm2, mm3
+    paddsw      mm2, mm1
+    paddsw      mm2, mm4
+    psraw       mm2, 7
+    packuswb    mm2, mm2
+
+    movd        DWORD PTR [rdi], mm2
+
+%if ABI_IS_32BIT
+    add         rdi,        DWORD PTR arg(3) ;[out_pitch]
+%else
+    add         rdi,        r8
+%endif
+    dec         rcx
+    jnz         .vp8_filter_block1d4_v6_ssse3_loop
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+.vp8_filter_block1d4_v4_ssse3:
+    movq        mm6, MMWORD PTR [rax+256]     ;k2_k4
+    movq        mm7, MMWORD PTR [rax+128]     ;k1_k3
+    movq        mm5, MMWORD PTR [GLOBAL(rd)]
+
+    mov         rsi, arg(0)             ;src_ptr
+
+    mov         rax, rsi
+    add         rax, rdx
+
+.vp8_filter_block1d4_v4_ssse3_loop:
+    movd        mm2, DWORD PTR [rsi + rdx]            ;B
+    movd        mm3, DWORD PTR [rsi + rdx * 2]        ;C
+    movd        mm4, DWORD PTR [rax + rdx * 2]        ;D
+    movd        mm0, DWORD PTR [rsi + rdx * 4]        ;E
+
+    punpcklbw   mm2, mm4                  ;B D
+    punpcklbw   mm3, mm0                  ;C E
+
+    pmaddubsw   mm3, mm6
+    pmaddubsw   mm2, mm7
+    add         rsi,  rdx
+    add         rax,  rdx
+;--
+;--
+    paddsw      mm2, mm3
+    paddsw      mm2, mm5
+    psraw       mm2, 7
+    packuswb    mm2, mm2
+
+    movd        DWORD PTR [rdi], mm2
+
+%if ABI_IS_32BIT
+    add         rdi,        DWORD PTR arg(3) ;[out_pitch]
+%else
+    add         rdi,        r8
+%endif
+    dec         rcx
+    jnz         .vp8_filter_block1d4_v4_ssse3_loop
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp8_bilinear_predict16x16_ssse3
+;(
+;    unsigned char  *src_ptr,
+;    int   src_pixels_per_line,
+;    int  xoffset,
+;    int  yoffset,
+;    unsigned char *dst_ptr,
+;    int dst_pitch
+;)
+global sym(vp8_bilinear_predict16x16_ssse3) PRIVATE
+sym(vp8_bilinear_predict16x16_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        lea         rcx,        [GLOBAL(vp8_bilinear_filters_ssse3)]
+        movsxd      rax,        dword ptr arg(2)    ; xoffset
+
+        cmp         rax,        0                   ; skip first_pass filter if xoffset=0
+        je          .b16x16_sp_only
+
+        shl         rax,        4
+        lea         rax,        [rax + rcx]         ; HFilter
+
+        mov         rdi,        arg(4)              ; dst_ptr
+        mov         rsi,        arg(0)              ; src_ptr
+        movsxd      rdx,        dword ptr arg(5)    ; dst_pitch
+
+        movdqa      xmm1,       [rax]
+
+        movsxd      rax,        dword ptr arg(3)    ; yoffset
+
+        cmp         rax,        0                   ; skip second_pass filter if yoffset=0
+        je          .b16x16_fp_only
+
+        shl         rax,        4
+        lea         rax,        [rax + rcx]         ; VFilter
+
+        lea         rcx,        [rdi+rdx*8]
+        lea         rcx,        [rcx+rdx*8]
+        movsxd      rdx,        dword ptr arg(1)    ; src_pixels_per_line
+
+        movdqa      xmm2,       [rax]
+
+%if ABI_IS_32BIT=0
+        movsxd      r8,         dword ptr arg(5)    ; dst_pitch
+%endif
+        movq        xmm3,       [rsi]               ; 00 01 02 03 04 05 06 07
+        movq        xmm5,       [rsi+1]             ; 01 02 03 04 05 06 07 08
+
+        punpcklbw   xmm3,       xmm5                ; 00 01 01 02 02 03 03 04 04 05 05 06 06 07 07 08
+        movq        xmm4,       [rsi+8]             ; 08 09 10 11 12 13 14 15
+
+        movq        xmm5,       [rsi+9]             ; 09 10 11 12 13 14 15 16
+
+        lea         rsi,        [rsi + rdx]         ; next line
+
+        pmaddubsw   xmm3,       xmm1                ; 00 02 04 06 08 10 12 14
+
+        punpcklbw   xmm4,       xmm5                ; 08 09 09 10 10 11 11 12 12 13 13 14 14 15 15 16
+        pmaddubsw   xmm4,       xmm1                ; 01 03 05 07 09 11 13 15
+
+        paddw       xmm3,       [GLOBAL(rd)]        ; xmm3 += round value
+        psraw       xmm3,       VP8_FILTER_SHIFT    ; xmm3 /= 128
+
+        paddw       xmm4,       [GLOBAL(rd)]        ; xmm4 += round value
+        psraw       xmm4,       VP8_FILTER_SHIFT    ; xmm4 /= 128
+
+        movdqa      xmm7,       xmm3
+        packuswb    xmm7,       xmm4                ; 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
+
+.next_row:
+        movq        xmm6,       [rsi]               ; 00 01 02 03 04 05 06 07
+        movq        xmm5,       [rsi+1]             ; 01 02 03 04 05 06 07 08
+
+        punpcklbw   xmm6,       xmm5
+        movq        xmm4,       [rsi+8]             ; 08 09 10 11 12 13 14 15
+
+        movq        xmm5,       [rsi+9]             ; 09 10 11 12 13 14 15 16
+        lea         rsi,        [rsi + rdx]         ; next line
+
+        pmaddubsw   xmm6,       xmm1
+
+        punpcklbw   xmm4,       xmm5
+        pmaddubsw   xmm4,       xmm1
+
+        paddw       xmm6,       [GLOBAL(rd)]        ; xmm6 += round value
+        psraw       xmm6,       VP8_FILTER_SHIFT    ; xmm6 /= 128
+
+        paddw       xmm4,       [GLOBAL(rd)]        ; xmm4 += round value
+        psraw       xmm4,       VP8_FILTER_SHIFT    ; xmm4 /= 128
+
+        packuswb    xmm6,       xmm4
+        movdqa      xmm5,       xmm7
+
+        punpcklbw   xmm5,       xmm6
+        pmaddubsw   xmm5,       xmm2
+
+        punpckhbw   xmm7,       xmm6
+        pmaddubsw   xmm7,       xmm2
+
+        paddw       xmm5,       [GLOBAL(rd)]        ; xmm5 += round value
+        psraw       xmm5,       VP8_FILTER_SHIFT    ; xmm5 /= 128
+
+        paddw       xmm7,       [GLOBAL(rd)]        ; xmm7 += round value
+        psraw       xmm7,       VP8_FILTER_SHIFT    ; xmm7 /= 128
+
+        packuswb    xmm5,       xmm7
+        movdqa      xmm7,       xmm6
+
+        movdqa      [rdi],      xmm5                ; store the results in the destination
+%if ABI_IS_32BIT
+        add         rdi,        DWORD PTR arg(5)    ; dst_pitch
+%else
+        add         rdi,        r8
+%endif
+
+        cmp         rdi,        rcx
+        jne         .next_row
+
+        jmp         .done
+
+.b16x16_sp_only:
+        movsxd      rax,        dword ptr arg(3)    ; yoffset
+        shl         rax,        4
+        lea         rax,        [rax + rcx]         ; VFilter
+
+        mov         rdi,        arg(4)              ; dst_ptr
+        mov         rsi,        arg(0)              ; src_ptr
+        movsxd      rdx,        dword ptr arg(5)    ; dst_pitch
+
+        movdqa      xmm1,       [rax]               ; VFilter
+
+        lea         rcx,        [rdi+rdx*8]
+        lea         rcx,        [rcx+rdx*8]
+        movsxd      rax,        dword ptr arg(1)    ; src_pixels_per_line
+
+        ; get the first horizontal line done
+        movq        xmm4,       [rsi]               ; load row 0
+        movq        xmm2,       [rsi + 8]           ; load row 0
+
+        lea         rsi,        [rsi + rax]         ; next line
+.next_row_sp:
+        movq        xmm3,       [rsi]               ; load row + 1
+        movq        xmm5,       [rsi + 8]           ; load row + 1
+
+        punpcklbw   xmm4,       xmm3
+        punpcklbw   xmm2,       xmm5
+
+        pmaddubsw   xmm4,       xmm1
+        movq        xmm7,       [rsi + rax]         ; load row + 2
+
+        pmaddubsw   xmm2,       xmm1
+        movq        xmm6,       [rsi + rax + 8]     ; load row + 2
+
+        punpcklbw   xmm3,       xmm7
+        punpcklbw   xmm5,       xmm6
+
+        pmaddubsw   xmm3,       xmm1
+        paddw       xmm4,       [GLOBAL(rd)]
+
+        pmaddubsw   xmm5,       xmm1
+        paddw       xmm2,       [GLOBAL(rd)]
+
+        psraw       xmm4,       VP8_FILTER_SHIFT
+        psraw       xmm2,       VP8_FILTER_SHIFT
+
+        packuswb    xmm4,       xmm2
+        paddw       xmm3,       [GLOBAL(rd)]
+
+        movdqa      [rdi],      xmm4                ; store row 0
+        paddw       xmm5,       [GLOBAL(rd)]
+
+        psraw       xmm3,       VP8_FILTER_SHIFT
+        psraw       xmm5,       VP8_FILTER_SHIFT
+
+        packuswb    xmm3,       xmm5
+        movdqa      xmm4,       xmm7
+
+        movdqa      [rdi + rdx],xmm3                ; store row 1
+        lea         rsi,        [rsi + 2*rax]
+
+        movdqa      xmm2,       xmm6
+        lea         rdi,        [rdi + 2*rdx]
+
+        cmp         rdi,        rcx
+        jne         .next_row_sp
+
+        jmp         .done
+
+.b16x16_fp_only:
+        lea         rcx,        [rdi+rdx*8]
+        lea         rcx,        [rcx+rdx*8]
+        movsxd      rax,        dword ptr arg(1)    ; src_pixels_per_line
+
+.next_row_fp:
+        movq        xmm2,       [rsi]               ; 00 01 02 03 04 05 06 07
+        movq        xmm4,       [rsi+1]             ; 01 02 03 04 05 06 07 08
+
+        punpcklbw   xmm2,       xmm4
+        movq        xmm3,       [rsi+8]             ; 08 09 10 11 12 13 14 15
+
+        pmaddubsw   xmm2,       xmm1
+        movq        xmm4,       [rsi+9]             ; 09 10 11 12 13 14 15 16
+
+        lea         rsi,        [rsi + rax]         ; next line
+        punpcklbw   xmm3,       xmm4
+
+        pmaddubsw   xmm3,       xmm1
+        movq        xmm5,       [rsi]
+
+        paddw       xmm2,       [GLOBAL(rd)]
+        movq        xmm7,       [rsi+1]
+
+        movq        xmm6,       [rsi+8]
+        psraw       xmm2,       VP8_FILTER_SHIFT
+
+        punpcklbw   xmm5,       xmm7
+        movq        xmm7,       [rsi+9]
+
+        paddw       xmm3,       [GLOBAL(rd)]
+        pmaddubsw   xmm5,       xmm1
+
+        psraw       xmm3,       VP8_FILTER_SHIFT
+        punpcklbw   xmm6,       xmm7
+
+        packuswb    xmm2,       xmm3
+        pmaddubsw   xmm6,       xmm1
+
+        movdqa      [rdi],      xmm2                ; store the results in the destination
+        paddw       xmm5,       [GLOBAL(rd)]
+
+        lea         rdi,        [rdi + rdx]         ; dst_pitch
+        psraw       xmm5,       VP8_FILTER_SHIFT
+
+        paddw       xmm6,       [GLOBAL(rd)]
+        psraw       xmm6,       VP8_FILTER_SHIFT
+
+        packuswb    xmm5,       xmm6
+        lea         rsi,        [rsi + rax]         ; next line
+
+        movdqa      [rdi],      xmm5                ; store the results in the destination
+        lea         rdi,        [rdi + rdx]         ; dst_pitch
+
+        cmp         rdi,        rcx
+
+        jne         .next_row_fp
+
+.done:
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp8_bilinear_predict8x8_ssse3
+;(
+;    unsigned char  *src_ptr,
+;    int   src_pixels_per_line,
+;    int  xoffset,
+;    int  yoffset,
+;    unsigned char *dst_ptr,
+;    int dst_pitch
+;)
+global sym(vp8_bilinear_predict8x8_ssse3) PRIVATE
+sym(vp8_bilinear_predict8x8_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 144                         ; reserve 144 bytes
+
+        lea         rcx,        [GLOBAL(vp8_bilinear_filters_ssse3)]
+
+        mov         rsi,        arg(0) ;src_ptr
+        movsxd      rdx,        dword ptr arg(1) ;src_pixels_per_line
+
+    ;Read 9-line unaligned data in and put them on stack. This gives a big
+    ;performance boost.
+        movdqu      xmm0,       [rsi]
+        lea         rax,        [rdx + rdx*2]
+        movdqu      xmm1,       [rsi+rdx]
+        movdqu      xmm2,       [rsi+rdx*2]
+        add         rsi,        rax
+        movdqu      xmm3,       [rsi]
+        movdqu      xmm4,       [rsi+rdx]
+        movdqu      xmm5,       [rsi+rdx*2]
+        add         rsi,        rax
+        movdqu      xmm6,       [rsi]
+        movdqu      xmm7,       [rsi+rdx]
+
+        movdqa      XMMWORD PTR [rsp],            xmm0
+
+        movdqu      xmm0,       [rsi+rdx*2]
+
+        movdqa      XMMWORD PTR [rsp+16],         xmm1
+        movdqa      XMMWORD PTR [rsp+32],         xmm2
+        movdqa      XMMWORD PTR [rsp+48],         xmm3
+        movdqa      XMMWORD PTR [rsp+64],         xmm4
+        movdqa      XMMWORD PTR [rsp+80],         xmm5
+        movdqa      XMMWORD PTR [rsp+96],         xmm6
+        movdqa      XMMWORD PTR [rsp+112],        xmm7
+        movdqa      XMMWORD PTR [rsp+128],        xmm0
+
+        movsxd      rax,        dword ptr arg(2)    ; xoffset
+        cmp         rax,        0                   ; skip first_pass filter if xoffset=0
+        je          .b8x8_sp_only
+
+        shl         rax,        4
+        add         rax,        rcx                 ; HFilter
+
+        mov         rdi,        arg(4)              ; dst_ptr
+        movsxd      rdx,        dword ptr arg(5)    ; dst_pitch
+
+        movdqa      xmm0,       [rax]
+
+        movsxd      rax,        dword ptr arg(3)    ; yoffset
+        cmp         rax,        0                   ; skip second_pass filter if yoffset=0
+        je          .b8x8_fp_only
+
+        shl         rax,        4
+        lea         rax,        [rax + rcx]         ; VFilter
+
+        lea         rcx,        [rdi+rdx*8]
+
+        movdqa      xmm1,       [rax]
+
+        ; get the first horizontal line done
+        movdqa      xmm3,       [rsp]               ; 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
+        movdqa      xmm5,       xmm3                ; 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 xx
+
+        psrldq      xmm5,       1
+        lea         rsp,        [rsp + 16]          ; next line
+
+        punpcklbw   xmm3,       xmm5                ; 00 01 01 02 02 03 03 04 04 05 05 06 06 07 07 08
+        pmaddubsw   xmm3,       xmm0                ; 00 02 04 06 08 10 12 14
+
+        paddw       xmm3,       [GLOBAL(rd)]        ; xmm3 += round value
+        psraw       xmm3,       VP8_FILTER_SHIFT    ; xmm3 /= 128
+
+        movdqa      xmm7,       xmm3
+        packuswb    xmm7,       xmm7                ; 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
+
+.next_row:
+        movdqa      xmm6,       [rsp]               ; 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15
+        lea         rsp,        [rsp + 16]          ; next line
+
+        movdqa      xmm5,       xmm6
+
+        psrldq      xmm5,       1
+
+        punpcklbw   xmm6,       xmm5
+        pmaddubsw   xmm6,       xmm0
+
+        paddw       xmm6,       [GLOBAL(rd)]        ; xmm6 += round value
+        psraw       xmm6,       VP8_FILTER_SHIFT    ; xmm6 /= 128
+
+        packuswb    xmm6,       xmm6
+
+        punpcklbw   xmm7,       xmm6
+        pmaddubsw   xmm7,       xmm1
+
+        paddw       xmm7,       [GLOBAL(rd)]        ; xmm7 += round value
+        psraw       xmm7,       VP8_FILTER_SHIFT    ; xmm7 /= 128
+
+        packuswb    xmm7,       xmm7
+
+        movq        [rdi],      xmm7                ; store the results in the destination
+        lea         rdi,        [rdi + rdx]
+
+        movdqa      xmm7,       xmm6
+
+        cmp         rdi,        rcx
+        jne         .next_row
+
+        jmp         .done8x8
+
+.b8x8_sp_only:
+        movsxd      rax,        dword ptr arg(3)    ; yoffset
+        shl         rax,        4
+        lea         rax,        [rax + rcx]         ; VFilter
+
+        mov         rdi,        arg(4) ;dst_ptr
+        movsxd      rdx,        dword ptr arg(5)    ; dst_pitch
+
+        movdqa      xmm0,       [rax]               ; VFilter
+
+        movq        xmm1,       XMMWORD PTR [rsp]
+        movq        xmm2,       XMMWORD PTR [rsp+16]
+
+        movq        xmm3,       XMMWORD PTR [rsp+32]
+        punpcklbw   xmm1,       xmm2
+
+        movq        xmm4,       XMMWORD PTR [rsp+48]
+        punpcklbw   xmm2,       xmm3
+
+        movq        xmm5,       XMMWORD PTR [rsp+64]
+        punpcklbw   xmm3,       xmm4
+
+        movq        xmm6,       XMMWORD PTR [rsp+80]
+        punpcklbw   xmm4,       xmm5
+
+        movq        xmm7,       XMMWORD PTR [rsp+96]
+        punpcklbw   xmm5,       xmm6
+
+        pmaddubsw   xmm1,       xmm0
+        pmaddubsw   xmm2,       xmm0
+
+        pmaddubsw   xmm3,       xmm0
+        pmaddubsw   xmm4,       xmm0
+
+        pmaddubsw   xmm5,       xmm0
+        punpcklbw   xmm6,       xmm7
+
+        pmaddubsw   xmm6,       xmm0
+        paddw       xmm1,       [GLOBAL(rd)]
+
+        paddw       xmm2,       [GLOBAL(rd)]
+        psraw       xmm1,       VP8_FILTER_SHIFT
+
+        paddw       xmm3,       [GLOBAL(rd)]
+        psraw       xmm2,       VP8_FILTER_SHIFT
+
+        paddw       xmm4,       [GLOBAL(rd)]
+        psraw       xmm3,       VP8_FILTER_SHIFT
+
+        paddw       xmm5,       [GLOBAL(rd)]
+        psraw       xmm4,       VP8_FILTER_SHIFT
+
+        paddw       xmm6,       [GLOBAL(rd)]
+        psraw       xmm5,       VP8_FILTER_SHIFT
+
+        psraw       xmm6,       VP8_FILTER_SHIFT
+        packuswb    xmm1,       xmm1
+
+        packuswb    xmm2,       xmm2
+        movq        [rdi],      xmm1
+
+        packuswb    xmm3,       xmm3
+        movq        [rdi+rdx],  xmm2
+
+        packuswb    xmm4,       xmm4
+        movq        xmm1,       XMMWORD PTR [rsp+112]
+
+        lea         rdi,        [rdi + 2*rdx]
+        movq        xmm2,       XMMWORD PTR [rsp+128]
+
+        packuswb    xmm5,       xmm5
+        movq        [rdi],      xmm3
+
+        packuswb    xmm6,       xmm6
+        movq        [rdi+rdx],  xmm4
+
+        lea         rdi,        [rdi + 2*rdx]
+        punpcklbw   xmm7,       xmm1
+
+        movq        [rdi],      xmm5
+        pmaddubsw   xmm7,       xmm0
+
+        movq        [rdi+rdx],  xmm6
+        punpcklbw   xmm1,       xmm2
+
+        pmaddubsw   xmm1,       xmm0
+        paddw       xmm7,       [GLOBAL(rd)]
+
+        psraw       xmm7,       VP8_FILTER_SHIFT
+        paddw       xmm1,       [GLOBAL(rd)]
+
+        psraw       xmm1,       VP8_FILTER_SHIFT
+        packuswb    xmm7,       xmm7
+
+        packuswb    xmm1,       xmm1
+        lea         rdi,        [rdi + 2*rdx]
+
+        movq        [rdi],      xmm7
+
+        movq        [rdi+rdx],  xmm1
+        lea         rsp,        [rsp + 144]
+
+        jmp         .done8x8
+
+.b8x8_fp_only:
+        lea         rcx,        [rdi+rdx*8]
+
+.next_row_fp:
+        movdqa      xmm1,       XMMWORD PTR [rsp]
+        movdqa      xmm3,       XMMWORD PTR [rsp+16]
+
+        movdqa      xmm2,       xmm1
+        movdqa      xmm5,       XMMWORD PTR [rsp+32]
+
+        psrldq      xmm2,       1
+        movdqa      xmm7,       XMMWORD PTR [rsp+48]
+
+        movdqa      xmm4,       xmm3
+        psrldq      xmm4,       1
+
+        movdqa      xmm6,       xmm5
+        psrldq      xmm6,       1
+
+        punpcklbw   xmm1,       xmm2
+        pmaddubsw   xmm1,       xmm0
+
+        punpcklbw   xmm3,       xmm4
+        pmaddubsw   xmm3,       xmm0
+
+        punpcklbw   xmm5,       xmm6
+        pmaddubsw   xmm5,       xmm0
+
+        movdqa      xmm2,       xmm7
+        psrldq      xmm2,       1
+
+        punpcklbw   xmm7,       xmm2
+        pmaddubsw   xmm7,       xmm0
+
+        paddw       xmm1,       [GLOBAL(rd)]
+        psraw       xmm1,       VP8_FILTER_SHIFT
+
+        paddw       xmm3,       [GLOBAL(rd)]
+        psraw       xmm3,       VP8_FILTER_SHIFT
+
+        paddw       xmm5,       [GLOBAL(rd)]
+        psraw       xmm5,       VP8_FILTER_SHIFT
+
+        paddw       xmm7,       [GLOBAL(rd)]
+        psraw       xmm7,       VP8_FILTER_SHIFT
+
+        packuswb    xmm1,       xmm1
+        packuswb    xmm3,       xmm3
+
+        packuswb    xmm5,       xmm5
+        movq        [rdi],      xmm1
+
+        packuswb    xmm7,       xmm7
+        movq        [rdi+rdx],  xmm3
+
+        lea         rdi,        [rdi + 2*rdx]
+        movq        [rdi],      xmm5
+
+        lea         rsp,        [rsp + 4*16]
+        movq        [rdi+rdx],  xmm7
+
+        lea         rdi,        [rdi + 2*rdx]
+        cmp         rdi,        rcx
+
+        jne         .next_row_fp
+
+        lea         rsp,        [rsp + 16]
+
+.done8x8:
+    ;add rsp, 144
+    pop         rsp
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+shuf1b:
+    db 0, 5, 1, 6, 2, 7, 3, 8, 4, 9, 5, 10, 6, 11, 7, 12
+shuf2b:
+    db 2, 4, 3, 5, 4, 6, 5, 7, 6, 8, 7, 9, 8, 10, 9, 11
+shuf3b:
+    db 1, 3, 2, 4, 3, 5, 4, 6, 5, 7, 6, 8, 7, 9, 8, 10
+
+align 16
+shuf2bfrom1:
+    db  4, 8, 6, 1, 8, 3, 1, 5, 3, 7, 5, 9, 7,11, 9,13
+align 16
+shuf3bfrom1:
+    db  2, 6, 4, 8, 6, 1, 8, 3, 1, 5, 3, 7, 5, 9, 7,11
+
+align 16
+rd:
+    times 8 dw 0x40
+
+align 16
+k0_k5:
+    times 8 db 0, 0             ;placeholder
+    times 8 db 0, 0
+    times 8 db 2, 1
+    times 8 db 0, 0
+    times 8 db 3, 3
+    times 8 db 0, 0
+    times 8 db 1, 2
+    times 8 db 0, 0
+k1_k3:
+    times 8 db  0,    0         ;placeholder
+    times 8 db  -6,  12
+    times 8 db -11,  36
+    times 8 db  -9,  50
+    times 8 db -16,  77
+    times 8 db  -6,  93
+    times 8 db  -8, 108
+    times 8 db  -1, 123
+k2_k4:
+    times 8 db 128,    0        ;placeholder
+    times 8 db 123,   -1
+    times 8 db 108,   -8
+    times 8 db  93,   -6
+    times 8 db  77,  -16
+    times 8 db  50,   -9
+    times 8 db  36,  -11
+    times 8 db  12,   -6
+align 16
+vp8_bilinear_filters_ssse3:
+    times 8 db 128, 0
+    times 8 db 112, 16
+    times 8 db 96,  32
+    times 8 db 80,  48
+    times 8 db 64,  64
+    times 8 db 48,  80
+    times 8 db 32,  96
+    times 8 db 16,  112
+
diff --git a/libvpx/libvpx/vp8/common/x86/vp8_asm_stubs.c b/libvpx/libvpx/vp8/common/x86/vp8_asm_stubs.c
new file mode 100644
index 0000000..fb0b57e
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/vp8_asm_stubs.c
@@ -0,0 +1,625 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "filter_x86.h"
+
+extern const short vp8_six_tap_mmx[8][6*8];
+
+extern void vp8_filter_block1d_h6_mmx
+(
+    unsigned char   *src_ptr,
+    unsigned short  *output_ptr,
+    unsigned int    src_pixels_per_line,
+    unsigned int    pixel_step,
+    unsigned int    output_height,
+    unsigned int    output_width,
+    const short      *vp8_filter
+);
+extern void vp8_filter_block1dc_v6_mmx
+(
+    unsigned short *src_ptr,
+    unsigned char  *output_ptr,
+    int             output_pitch,
+    unsigned int    pixels_per_line,
+    unsigned int    pixel_step,
+    unsigned int    output_height,
+    unsigned int    output_width,
+    const short    *vp8_filter
+);
+extern void vp8_filter_block1d8_h6_sse2
+(
+    unsigned char  *src_ptr,
+    unsigned short *output_ptr,
+    unsigned int    src_pixels_per_line,
+    unsigned int    pixel_step,
+    unsigned int    output_height,
+    unsigned int    output_width,
+    const short    *vp8_filter
+);
+extern void vp8_filter_block1d16_h6_sse2
+(
+    unsigned char  *src_ptr,
+    unsigned short *output_ptr,
+    unsigned int    src_pixels_per_line,
+    unsigned int    pixel_step,
+    unsigned int    output_height,
+    unsigned int    output_width,
+    const short    *vp8_filter
+);
+extern void vp8_filter_block1d8_v6_sse2
+(
+    unsigned short *src_ptr,
+    unsigned char *output_ptr,
+    int dst_ptich,
+    unsigned int pixels_per_line,
+    unsigned int pixel_step,
+    unsigned int output_height,
+    unsigned int output_width,
+    const short    *vp8_filter
+);
+extern void vp8_filter_block1d16_v6_sse2
+(
+    unsigned short *src_ptr,
+    unsigned char *output_ptr,
+    int dst_ptich,
+    unsigned int pixels_per_line,
+    unsigned int pixel_step,
+    unsigned int output_height,
+    unsigned int output_width,
+    const short    *vp8_filter
+);
+extern void vp8_unpack_block1d16_h6_sse2
+(
+    unsigned char  *src_ptr,
+    unsigned short *output_ptr,
+    unsigned int    src_pixels_per_line,
+    unsigned int    output_height,
+    unsigned int    output_width
+);
+extern void vp8_filter_block1d8_h6_only_sse2
+(
+    unsigned char  *src_ptr,
+    unsigned int    src_pixels_per_line,
+    unsigned char  *output_ptr,
+    int dst_ptich,
+    unsigned int    output_height,
+    const short    *vp8_filter
+);
+extern void vp8_filter_block1d16_h6_only_sse2
+(
+    unsigned char  *src_ptr,
+    unsigned int    src_pixels_per_line,
+    unsigned char  *output_ptr,
+    int dst_ptich,
+    unsigned int    output_height,
+    const short    *vp8_filter
+);
+extern void vp8_filter_block1d8_v6_only_sse2
+(
+    unsigned char *src_ptr,
+    unsigned int   src_pixels_per_line,
+    unsigned char *output_ptr,
+    int dst_ptich,
+    unsigned int   output_height,
+    const short   *vp8_filter
+);
+
+
+#if HAVE_MMX
+void vp8_sixtap_predict4x4_mmx
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+    DECLARE_ALIGNED(16, unsigned short, FData2[16*16]);  /* Temp data bufffer used in filtering */
+    const short *HFilter, *VFilter;
+    HFilter = vp8_six_tap_mmx[xoffset];
+    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line), FData2, src_pixels_per_line, 1, 9, 8, HFilter);
+    VFilter = vp8_six_tap_mmx[yoffset];
+    vp8_filter_block1dc_v6_mmx(FData2 + 8, dst_ptr, dst_pitch, 8, 4 , 4, 4, VFilter);
+
+}
+
+
+void vp8_sixtap_predict16x16_mmx
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+
+    DECLARE_ALIGNED(16, unsigned short, FData2[24*24]);  /* Temp data bufffer used in filtering */
+
+    const short *HFilter, *VFilter;
+
+
+    HFilter = vp8_six_tap_mmx[xoffset];
+
+    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line),    FData2,   src_pixels_per_line, 1, 21, 32, HFilter);
+    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 4,  FData2 + 4, src_pixels_per_line, 1, 21, 32, HFilter);
+    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 8,  FData2 + 8, src_pixels_per_line, 1, 21, 32, HFilter);
+    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 12, FData2 + 12, src_pixels_per_line, 1, 21, 32, HFilter);
+
+    VFilter = vp8_six_tap_mmx[yoffset];
+    vp8_filter_block1dc_v6_mmx(FData2 + 32, dst_ptr,   dst_pitch, 32, 16 , 16, 16, VFilter);
+    vp8_filter_block1dc_v6_mmx(FData2 + 36, dst_ptr + 4, dst_pitch, 32, 16 , 16, 16, VFilter);
+    vp8_filter_block1dc_v6_mmx(FData2 + 40, dst_ptr + 8, dst_pitch, 32, 16 , 16, 16, VFilter);
+    vp8_filter_block1dc_v6_mmx(FData2 + 44, dst_ptr + 12, dst_pitch, 32, 16 , 16, 16, VFilter);
+
+}
+
+
+void vp8_sixtap_predict8x8_mmx
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+
+    DECLARE_ALIGNED(16, unsigned short, FData2[256]);    /* Temp data bufffer used in filtering */
+
+    const short *HFilter, *VFilter;
+
+    HFilter = vp8_six_tap_mmx[xoffset];
+    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line),    FData2,   src_pixels_per_line, 1, 13, 16, HFilter);
+    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 4,  FData2 + 4, src_pixels_per_line, 1, 13, 16, HFilter);
+
+    VFilter = vp8_six_tap_mmx[yoffset];
+    vp8_filter_block1dc_v6_mmx(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 8, 8, VFilter);
+    vp8_filter_block1dc_v6_mmx(FData2 + 20, dst_ptr + 4, dst_pitch, 16, 8 , 8, 8, VFilter);
+
+}
+
+
+void vp8_sixtap_predict8x4_mmx
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+
+    DECLARE_ALIGNED(16, unsigned short, FData2[256]);    /* Temp data bufffer used in filtering */
+
+    const short *HFilter, *VFilter;
+
+    HFilter = vp8_six_tap_mmx[xoffset];
+    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line),    FData2,   src_pixels_per_line, 1, 9, 16, HFilter);
+    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 4,  FData2 + 4, src_pixels_per_line, 1, 9, 16, HFilter);
+
+    VFilter = vp8_six_tap_mmx[yoffset];
+    vp8_filter_block1dc_v6_mmx(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 4, 8, VFilter);
+    vp8_filter_block1dc_v6_mmx(FData2 + 20, dst_ptr + 4, dst_pitch, 16, 8 , 4, 8, VFilter);
+
+}
+
+
+
+void vp8_bilinear_predict16x16_mmx
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+    vp8_bilinear_predict8x8_mmx(src_ptr,   src_pixels_per_line, xoffset, yoffset, dst_ptr,   dst_pitch);
+    vp8_bilinear_predict8x8_mmx(src_ptr + 8, src_pixels_per_line, xoffset, yoffset, dst_ptr + 8, dst_pitch);
+    vp8_bilinear_predict8x8_mmx(src_ptr + 8 * src_pixels_per_line,   src_pixels_per_line, xoffset, yoffset, dst_ptr + dst_pitch * 8,   dst_pitch);
+    vp8_bilinear_predict8x8_mmx(src_ptr + 8 * src_pixels_per_line + 8, src_pixels_per_line, xoffset, yoffset, dst_ptr + dst_pitch * 8 + 8, dst_pitch);
+}
+#endif
+
+
+#if HAVE_SSE2
+void vp8_sixtap_predict16x16_sse2
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+
+)
+{
+    DECLARE_ALIGNED(16, unsigned short, FData2[24*24]);    /* Temp data bufffer used in filtering */
+
+    const short *HFilter, *VFilter;
+
+    if (xoffset)
+    {
+        if (yoffset)
+        {
+            HFilter = vp8_six_tap_mmx[xoffset];
+            vp8_filter_block1d16_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 1, 21, 32, HFilter);
+            VFilter = vp8_six_tap_mmx[yoffset];
+            vp8_filter_block1d16_v6_sse2(FData2 + 32, dst_ptr,   dst_pitch, 32, 16 , 16, dst_pitch, VFilter);
+        }
+        else
+        {
+            /* First-pass only */
+            HFilter = vp8_six_tap_mmx[xoffset];
+            vp8_filter_block1d16_h6_only_sse2(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 16, HFilter);
+        }
+    }
+    else
+    {
+        /* Second-pass only */
+        VFilter = vp8_six_tap_mmx[yoffset];
+        vp8_unpack_block1d16_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 21, 32);
+        vp8_filter_block1d16_v6_sse2(FData2 + 32, dst_ptr,   dst_pitch, 32, 16 , 16, dst_pitch, VFilter);
+    }
+}
+
+
+void vp8_sixtap_predict8x8_sse2
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+    DECLARE_ALIGNED(16, unsigned short, FData2[256]);  /* Temp data bufffer used in filtering */
+    const short *HFilter, *VFilter;
+
+    if (xoffset)
+    {
+        if (yoffset)
+        {
+            HFilter = vp8_six_tap_mmx[xoffset];
+            vp8_filter_block1d8_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 1, 13, 16, HFilter);
+            VFilter = vp8_six_tap_mmx[yoffset];
+            vp8_filter_block1d8_v6_sse2(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 8, dst_pitch, VFilter);
+        }
+        else
+        {
+            /* First-pass only */
+            HFilter = vp8_six_tap_mmx[xoffset];
+            vp8_filter_block1d8_h6_only_sse2(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 8, HFilter);
+        }
+    }
+    else
+    {
+        /* Second-pass only */
+        VFilter = vp8_six_tap_mmx[yoffset];
+        vp8_filter_block1d8_v6_only_sse2(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, dst_ptr, dst_pitch, 8, VFilter);
+    }
+}
+
+
+void vp8_sixtap_predict8x4_sse2
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+    DECLARE_ALIGNED(16, unsigned short, FData2[256]);  /* Temp data bufffer used in filtering */
+    const short *HFilter, *VFilter;
+
+    if (xoffset)
+    {
+        if (yoffset)
+        {
+            HFilter = vp8_six_tap_mmx[xoffset];
+            vp8_filter_block1d8_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 1, 9, 16, HFilter);
+            VFilter = vp8_six_tap_mmx[yoffset];
+            vp8_filter_block1d8_v6_sse2(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 4, dst_pitch, VFilter);
+        }
+        else
+        {
+            /* First-pass only */
+            HFilter = vp8_six_tap_mmx[xoffset];
+            vp8_filter_block1d8_h6_only_sse2(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 4, HFilter);
+        }
+    }
+    else
+    {
+        /* Second-pass only */
+        VFilter = vp8_six_tap_mmx[yoffset];
+        vp8_filter_block1d8_v6_only_sse2(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, dst_ptr, dst_pitch, 4, VFilter);
+    }
+}
+
+#endif
+
+#if HAVE_SSSE3
+
+extern void vp8_filter_block1d8_h6_ssse3
+(
+    unsigned char  *src_ptr,
+    unsigned int    src_pixels_per_line,
+    unsigned char  *output_ptr,
+    unsigned int    output_pitch,
+    unsigned int    output_height,
+    unsigned int    vp8_filter_index
+);
+
+extern void vp8_filter_block1d16_h6_ssse3
+(
+    unsigned char  *src_ptr,
+    unsigned int    src_pixels_per_line,
+    unsigned char  *output_ptr,
+    unsigned int    output_pitch,
+    unsigned int    output_height,
+    unsigned int    vp8_filter_index
+);
+
+extern void vp8_filter_block1d16_v6_ssse3
+(
+    unsigned char *src_ptr,
+    unsigned int   src_pitch,
+    unsigned char *output_ptr,
+    unsigned int   out_pitch,
+    unsigned int   output_height,
+    unsigned int   vp8_filter_index
+);
+
+extern void vp8_filter_block1d8_v6_ssse3
+(
+    unsigned char *src_ptr,
+    unsigned int   src_pitch,
+    unsigned char *output_ptr,
+    unsigned int   out_pitch,
+    unsigned int   output_height,
+    unsigned int   vp8_filter_index
+);
+
+extern void vp8_filter_block1d4_h6_ssse3
+(
+    unsigned char  *src_ptr,
+    unsigned int    src_pixels_per_line,
+    unsigned char  *output_ptr,
+    unsigned int    output_pitch,
+    unsigned int    output_height,
+    unsigned int    vp8_filter_index
+);
+
+extern void vp8_filter_block1d4_v6_ssse3
+(
+    unsigned char *src_ptr,
+    unsigned int   src_pitch,
+    unsigned char *output_ptr,
+    unsigned int   out_pitch,
+    unsigned int   output_height,
+    unsigned int   vp8_filter_index
+);
+
+void vp8_sixtap_predict16x16_ssse3
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+
+)
+{
+    DECLARE_ALIGNED(16, unsigned char, FData2[24*24]);
+
+    if (xoffset)
+    {
+        if (yoffset)
+        {
+            vp8_filter_block1d16_h6_ssse3(src_ptr - (2 * src_pixels_per_line),
+                                          src_pixels_per_line, FData2,
+                                          16, 21, xoffset);
+            vp8_filter_block1d16_v6_ssse3(FData2 , 16, dst_ptr, dst_pitch,
+                                          16, yoffset);
+        }
+        else
+        {
+            /* First-pass only */
+            vp8_filter_block1d16_h6_ssse3(src_ptr, src_pixels_per_line,
+                                          dst_ptr, dst_pitch, 16, xoffset);
+        }
+    }
+    else
+    {
+        if (yoffset)
+        {
+            /* Second-pass only */
+            vp8_filter_block1d16_v6_ssse3(src_ptr - (2 * src_pixels_per_line),
+                                          src_pixels_per_line,
+                                          dst_ptr, dst_pitch, 16, yoffset);
+        }
+        else
+        {
+            /* ssse3 second-pass only function couldn't handle (xoffset==0 &&
+             * yoffset==0) case correctly. Add copy function here to guarantee
+             * six-tap function handles all possible offsets. */
+            vp8_copy_mem16x16(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch);
+        }
+    }
+}
+
+void vp8_sixtap_predict8x8_ssse3
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+    DECLARE_ALIGNED(16, unsigned char, FData2[256]);
+
+    if (xoffset)
+    {
+        if (yoffset)
+        {
+            vp8_filter_block1d8_h6_ssse3(src_ptr - (2 * src_pixels_per_line),
+                                         src_pixels_per_line, FData2,
+                                         8, 13, xoffset);
+            vp8_filter_block1d8_v6_ssse3(FData2, 8, dst_ptr, dst_pitch,
+                                         8, yoffset);
+        }
+        else
+        {
+            vp8_filter_block1d8_h6_ssse3(src_ptr, src_pixels_per_line,
+                                         dst_ptr, dst_pitch, 8, xoffset);
+        }
+    }
+    else
+    {
+        if (yoffset)
+        {
+            /* Second-pass only */
+            vp8_filter_block1d8_v6_ssse3(src_ptr - (2 * src_pixels_per_line),
+                                         src_pixels_per_line,
+                                         dst_ptr, dst_pitch, 8, yoffset);
+        }
+        else
+        {
+            /* ssse3 second-pass only function couldn't handle (xoffset==0 &&
+             * yoffset==0) case correctly. Add copy function here to guarantee
+             * six-tap function handles all possible offsets. */
+            vp8_copy_mem8x8(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch);
+        }
+    }
+}
+
+
+void vp8_sixtap_predict8x4_ssse3
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+    DECLARE_ALIGNED(16, unsigned char, FData2[256]);
+
+    if (xoffset)
+    {
+        if (yoffset)
+        {
+            vp8_filter_block1d8_h6_ssse3(src_ptr - (2 * src_pixels_per_line),
+                                         src_pixels_per_line, FData2,
+                                         8, 9, xoffset);
+            vp8_filter_block1d8_v6_ssse3(FData2, 8, dst_ptr, dst_pitch,
+                                         4, yoffset);
+        }
+        else
+        {
+            /* First-pass only */
+            vp8_filter_block1d8_h6_ssse3(src_ptr, src_pixels_per_line,
+                                         dst_ptr, dst_pitch, 4, xoffset);
+        }
+    }
+    else
+    {
+        if (yoffset)
+        {
+            /* Second-pass only */
+            vp8_filter_block1d8_v6_ssse3(src_ptr - (2 * src_pixels_per_line),
+                                         src_pixels_per_line,
+                                         dst_ptr, dst_pitch, 4, yoffset);
+        }
+        else
+        {
+            /* ssse3 second-pass only function couldn't handle (xoffset==0 &&
+             * yoffset==0) case correctly. Add copy function here to guarantee
+             * six-tap function handles all possible offsets. */
+            vp8_copy_mem8x4(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch);
+        }
+    }
+}
+
+void vp8_sixtap_predict4x4_ssse3
+(
+    unsigned char  *src_ptr,
+    int   src_pixels_per_line,
+    int  xoffset,
+    int  yoffset,
+    unsigned char *dst_ptr,
+    int dst_pitch
+)
+{
+  DECLARE_ALIGNED(16, unsigned char, FData2[4*9]);
+
+  if (xoffset)
+  {
+      if (yoffset)
+      {
+          vp8_filter_block1d4_h6_ssse3(src_ptr - (2 * src_pixels_per_line),
+                                       src_pixels_per_line,
+                                       FData2, 4, 9, xoffset);
+          vp8_filter_block1d4_v6_ssse3(FData2, 4, dst_ptr, dst_pitch,
+                                       4, yoffset);
+      }
+      else
+      {
+          vp8_filter_block1d4_h6_ssse3(src_ptr, src_pixels_per_line,
+                                       dst_ptr, dst_pitch, 4, xoffset);
+      }
+  }
+  else
+  {
+      if (yoffset)
+      {
+          vp8_filter_block1d4_v6_ssse3(src_ptr - (2 * src_pixels_per_line),
+                                       src_pixels_per_line,
+                                       dst_ptr, dst_pitch, 4, yoffset);
+      }
+      else
+      {
+        /* ssse3 second-pass only function couldn't handle (xoffset==0 &&
+          * yoffset==0) case correctly. Add copy function here to guarantee
+          * six-tap function handles all possible offsets. */
+          int r;
+
+          for (r = 0; r < 4; r++)
+          {
+            dst_ptr[0]  = src_ptr[0];
+            dst_ptr[1]  = src_ptr[1];
+            dst_ptr[2]  = src_ptr[2];
+            dst_ptr[3]  = src_ptr[3];
+            dst_ptr     += dst_pitch;
+            src_ptr     += src_pixels_per_line;
+          }
+      }
+  }
+}
+
+#endif
diff --git a/libvpx/libvpx/vp8/common/x86/vp8_loopfilter_mmx.asm b/libvpx/libvpx/vp8/common/x86/vp8_loopfilter_mmx.asm
new file mode 100644
index 0000000..88a07b9
--- /dev/null
+++ b/libvpx/libvpx/vp8/common/x86/vp8_loopfilter_mmx.asm
@@ -0,0 +1,1753 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+
+;void vp8_loop_filter_horizontal_edge_mmx
+;(
+;    unsigned char *src_ptr,
+;    int src_pixel_step,
+;    const char *blimit,
+;    const char *limit,
+;    const char *thresh,
+;    int  count
+;)
+global sym(vp8_loop_filter_horizontal_edge_mmx) PRIVATE
+sym(vp8_loop_filter_horizontal_edge_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 32                         ; reserve 32 bytes
+    %define t0 [rsp + 0]    ;__declspec(align(16)) char t0[8];
+    %define t1 [rsp + 16]   ;__declspec(align(16)) char t1[8];
+
+        mov         rsi, arg(0) ;src_ptr
+        movsxd      rax, dword ptr arg(1) ;src_pixel_step     ; destination pitch?
+
+        movsxd      rcx, dword ptr arg(5) ;count
+.next8_h:
+        mov         rdx, arg(3) ;limit
+        movq        mm7, [rdx]
+        mov         rdi, rsi              ; rdi points to row +1 for indirect addressing
+        add         rdi, rax
+
+        ; calculate breakout conditions
+        movq        mm2, [rdi+2*rax]      ; q3
+        movq        mm1, [rsi+2*rax]      ; q2
+        movq        mm6, mm1              ; q2
+        psubusb     mm1, mm2              ; q2-=q3
+        psubusb     mm2, mm6              ; q3-=q2
+        por         mm1, mm2              ; abs(q3-q2)
+        psubusb     mm1, mm7              ;
+
+
+        movq        mm4, [rsi+rax]        ; q1
+        movq        mm3, mm4              ; q1
+        psubusb     mm4, mm6              ; q1-=q2
+        psubusb     mm6, mm3              ; q2-=q1
+        por         mm4, mm6              ; abs(q2-q1)
+
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+        movq        mm4, [rsi]            ; q0
+        movq        mm0, mm4              ; q0
+        psubusb     mm4, mm3              ; q0-=q1
+        psubusb     mm3, mm0              ; q1-=q0
+        por         mm4, mm3              ; abs(q0-q1)
+        movq        t0, mm4               ; save to t0
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+
+        neg         rax                   ; negate pitch to deal with above border
+
+        movq        mm2, [rsi+4*rax]      ; p3
+        movq        mm4, [rdi+4*rax]      ; p2
+        movq        mm5, mm4              ; p2
+        psubusb     mm4, mm2              ; p2-=p3
+        psubusb     mm2, mm5              ; p3-=p2
+        por         mm4, mm2              ; abs(p3 - p2)
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+
+        movq        mm4, [rsi+2*rax]      ; p1
+        movq        mm3, mm4              ; p1
+        psubusb     mm4, mm5              ; p1-=p2
+        psubusb     mm5, mm3              ; p2-=p1
+        por         mm4, mm5              ; abs(p2 - p1)
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+        movq        mm2, mm3              ; p1
+
+        movq        mm4, [rsi+rax]        ; p0
+        movq        mm5, mm4              ; p0
+        psubusb     mm4, mm3              ; p0-=p1
+        psubusb     mm3, mm5              ; p1-=p0
+        por         mm4, mm3              ; abs(p1 - p0)
+        movq        t1, mm4               ; save to t1
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+        movq        mm3, [rdi]            ; q1
+        movq        mm4, mm3              ; q1
+        psubusb     mm3, mm2              ; q1-=p1
+        psubusb     mm2, mm4              ; p1-=q1
+        por         mm2, mm3              ; abs(p1-q1)
+        pand        mm2, [GLOBAL(tfe)]    ; set lsb of each byte to zero
+        psrlw       mm2, 1                ; abs(p1-q1)/2
+
+        movq        mm6, mm5              ; p0
+        movq        mm3, [rsi]            ; q0
+        psubusb     mm5, mm3              ; p0-=q0
+        psubusb     mm3, mm6              ; q0-=p0
+        por         mm5, mm3              ; abs(p0 - q0)
+        paddusb     mm5, mm5              ; abs(p0-q0)*2
+        paddusb     mm5, mm2              ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+        mov         rdx, arg(2) ;blimit           ; get blimit
+        movq        mm7, [rdx]            ; blimit
+
+        psubusb     mm5,    mm7           ; abs (p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        por         mm1,    mm5
+        pxor        mm5,    mm5
+        pcmpeqb     mm1,    mm5           ; mask mm1
+
+        ; calculate high edge variance
+        mov         rdx, arg(4) ;thresh           ; get thresh
+        movq        mm7, [rdx]            ;
+        movq        mm4, t0               ; get abs (q1 - q0)
+        psubusb     mm4, mm7
+        movq        mm3, t1               ; get abs (p1 - p0)
+        psubusb     mm3, mm7
+        paddb       mm4, mm3              ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
+
+        pcmpeqb     mm4,        mm5
+
+        pcmpeqb     mm5,        mm5
+        pxor        mm4,        mm5
+
+
+        ; start work on filters
+        movq        mm2, [rsi+2*rax]      ; p1
+        movq        mm7, [rdi]            ; q1
+        pxor        mm2, [GLOBAL(t80)]    ; p1 offset to convert to signed values
+        pxor        mm7, [GLOBAL(t80)]    ; q1 offset to convert to signed values
+        psubsb      mm2, mm7              ; p1 - q1
+        pand        mm2, mm4              ; high var mask (hvm)(p1 - q1)
+        pxor        mm6, [GLOBAL(t80)]    ; offset to convert to signed values
+        pxor        mm0, [GLOBAL(t80)]    ; offset to convert to signed values
+        movq        mm3, mm0              ; q0
+        psubsb      mm0, mm6              ; q0 - p0
+        paddsb      mm2, mm0              ; 1 * (q0 - p0) + hvm(p1 - q1)
+        paddsb      mm2, mm0              ; 2 * (q0 - p0) + hvm(p1 - q1)
+        paddsb      mm2, mm0              ; 3 * (q0 - p0) + hvm(p1 - q1)
+        pand        mm1, mm2                  ; mask filter values we don't care about
+        movq        mm2, mm1
+        paddsb      mm1, [GLOBAL(t4)]     ; 3* (q0 - p0) + hvm(p1 - q1) + 4
+        paddsb      mm2, [GLOBAL(t3)]     ; 3* (q0 - p0) + hvm(p1 - q1) + 3
+
+        pxor        mm0, mm0             ;
+        pxor        mm5, mm5
+        punpcklbw   mm0, mm2            ;
+        punpckhbw   mm5, mm2            ;
+        psraw       mm0, 11             ;
+        psraw       mm5, 11
+        packsswb    mm0, mm5
+        movq        mm2, mm0            ;  (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
+
+        pxor        mm0, mm0              ; 0
+        movq        mm5, mm1              ; abcdefgh
+        punpcklbw   mm0, mm1              ; e0f0g0h0
+        psraw       mm0, 11               ; sign extended shift right by 3
+        pxor        mm1, mm1              ; 0
+        punpckhbw   mm1, mm5              ; a0b0c0d0
+        psraw       mm1, 11               ; sign extended shift right by 3
+        movq        mm5, mm0              ; save results
+
+        packsswb    mm0, mm1              ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
+        paddsw      mm5, [GLOBAL(ones)]
+        paddsw      mm1, [GLOBAL(ones)]
+        psraw       mm5, 1                ; partial shifted one more time for 2nd tap
+        psraw       mm1, 1                ; partial shifted one more time for 2nd tap
+        packsswb    mm5, mm1              ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>4
+        pandn       mm4, mm5              ; high edge variance additive
+
+        paddsb      mm6, mm2              ; p0+= p0 add
+        pxor        mm6, [GLOBAL(t80)]    ; unoffset
+        movq        [rsi+rax], mm6        ; write back
+
+        movq        mm6, [rsi+2*rax]      ; p1
+        pxor        mm6, [GLOBAL(t80)]    ; reoffset
+        paddsb      mm6, mm4              ; p1+= p1 add
+        pxor        mm6, [GLOBAL(t80)]    ; unoffset
+        movq        [rsi+2*rax], mm6      ; write back
+
+        psubsb      mm3, mm0              ; q0-= q0 add
+        pxor        mm3, [GLOBAL(t80)]    ; unoffset
+        movq        [rsi], mm3            ; write back
+
+        psubsb      mm7, mm4              ; q1-= q1 add
+        pxor        mm7, [GLOBAL(t80)]    ; unoffset
+        movq        [rdi], mm7            ; write back
+
+        add         rsi,8
+        neg         rax
+        dec         rcx
+        jnz         .next8_h
+
+    add rsp, 32
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_loop_filter_vertical_edge_mmx
+;(
+;    unsigned char *src_ptr,
+;    int  src_pixel_step,
+;    const char *blimit,
+;    const char *limit,
+;    const char *thresh,
+;    int count
+;)
+global sym(vp8_loop_filter_vertical_edge_mmx) PRIVATE
+sym(vp8_loop_filter_vertical_edge_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub          rsp, 64      ; reserve 64 bytes
+    %define t0   [rsp + 0]    ;__declspec(align(16)) char t0[8];
+    %define t1   [rsp + 16]   ;__declspec(align(16)) char t1[8];
+    %define srct [rsp + 32]   ;__declspec(align(16)) char srct[32];
+
+        mov         rsi,        arg(0) ;src_ptr
+        movsxd      rax,        dword ptr arg(1) ;src_pixel_step     ; destination pitch?
+
+        lea         rsi,        [rsi + rax*4 - 4]
+
+        movsxd      rcx,        dword ptr arg(5) ;count
+.next8_v:
+        mov         rdi,        rsi           ; rdi points to row +1 for indirect addressing
+        add         rdi,        rax
+
+
+        ;transpose
+        movq        mm6,        [rsi+2*rax]                 ; 67 66 65 64 63 62 61 60
+        movq        mm7,        mm6                         ; 77 76 75 74 73 72 71 70
+
+        punpckhbw   mm7,        [rdi+2*rax]                 ; 77 67 76 66 75 65 74 64
+        punpcklbw   mm6,        [rdi+2*rax]                 ; 73 63 72 62 71 61 70 60
+
+        movq        mm4,        [rsi]                       ; 47 46 45 44 43 42 41 40
+        movq        mm5,        mm4                         ; 47 46 45 44 43 42 41 40
+
+        punpckhbw   mm5,        [rsi+rax]                   ; 57 47 56 46 55 45 54 44
+        punpcklbw   mm4,        [rsi+rax]                   ; 53 43 52 42 51 41 50 40
+
+        movq        mm3,        mm5                         ; 57 47 56 46 55 45 54 44
+        punpckhwd   mm5,        mm7                         ; 77 67 57 47 76 66 56 46
+
+        punpcklwd   mm3,        mm7                         ; 75 65 55 45 74 64 54 44
+        movq        mm2,        mm4                         ; 53 43 52 42 51 41 50 40
+
+        punpckhwd   mm4,        mm6                         ; 73 63 53 43 72 62 52 42
+        punpcklwd   mm2,        mm6                         ; 71 61 51 41 70 60 50 40
+
+        neg         rax
+        movq        mm6,        [rsi+rax*2]                 ; 27 26 25 24 23 22 21 20
+
+        movq        mm1,        mm6                         ; 27 26 25 24 23 22 21 20
+        punpckhbw   mm6,        [rsi+rax]                   ; 37 27 36 36 35 25 34 24
+
+        punpcklbw   mm1,        [rsi+rax]                   ; 33 23 32 22 31 21 30 20
+        movq        mm7,        [rsi+rax*4];                ; 07 06 05 04 03 02 01 00
+
+        punpckhbw   mm7,        [rdi+rax*4]                 ; 17 07 16 06 15 05 14 04
+        movq        mm0,        mm7                         ; 17 07 16 06 15 05 14 04
+
+        punpckhwd   mm7,        mm6                         ; 37 27 17 07 36 26 16 06
+        punpcklwd   mm0,        mm6                         ; 35 25 15 05 34 24 14 04
+
+        movq        mm6,        mm7                         ; 37 27 17 07 36 26 16 06
+        punpckhdq   mm7,        mm5                         ; 77 67 57 47 37 27 17 07  = q3
+
+        punpckldq   mm6,        mm5                         ; 76 66 56 46 36 26 16 06  = q2
+
+        movq        mm5,        mm6                         ; 76 66 56 46 36 26 16 06
+        psubusb     mm5,        mm7                         ; q2-q3
+
+        psubusb     mm7,        mm6                         ; q3-q2
+        por         mm7,        mm5;                        ; mm7=abs (q3-q2)
+
+        movq        mm5,        mm0                         ; 35 25 15 05 34 24 14 04
+        punpckhdq   mm5,        mm3                         ; 75 65 55 45 35 25 15 05 = q1
+
+        punpckldq   mm0,        mm3                         ; 74 64 54 44 34 24 15 04 = q0
+        movq        mm3,        mm5                         ; 75 65 55 45 35 25 15 05 = q1
+
+        psubusb     mm3,        mm6                         ; q1-q2
+        psubusb     mm6,        mm5                         ; q2-q1
+
+        por         mm6,        mm3                         ; mm6=abs(q2-q1)
+        lea         rdx,        srct
+
+        movq        [rdx+24],   mm5                         ; save q1
+        movq        [rdx+16],   mm0                         ; save q0
+
+        movq        mm3,        [rsi+rax*4]                 ; 07 06 05 04 03 02 01 00
+        punpcklbw   mm3,        [rdi+rax*4]                 ; 13 03 12 02 11 01 10 00
+
+        movq        mm0,        mm3                         ; 13 03 12 02 11 01 10 00
+        punpcklwd   mm0,        mm1                         ; 31 21 11 01 30 20 10 00
+
+        punpckhwd   mm3,        mm1                         ; 33 23 13 03 32 22 12 02
+        movq        mm1,        mm0                         ; 31 21 11 01 30 20 10 00
+
+        punpckldq   mm0,        mm2                         ; 70 60 50 40 30 20 10 00  =p3
+        punpckhdq   mm1,        mm2                         ; 71 61 51 41 31 21 11 01  =p2
+
+        movq        mm2,        mm1                         ; 71 61 51 41 31 21 11 01  =p2
+        psubusb     mm2,        mm0                         ; p2-p3
+
+        psubusb     mm0,        mm1                         ; p3-p2
+        por         mm0,        mm2                         ; mm0=abs(p3-p2)
+
+        movq        mm2,        mm3                         ; 33 23 13 03 32 22 12 02
+        punpckldq   mm2,        mm4                         ; 72 62 52 42 32 22 12 02 = p1
+
+        punpckhdq   mm3,        mm4                         ; 73 63 53 43 33 23 13 03 = p0
+        movq        [rdx+8],    mm3                         ; save p0
+
+        movq        [rdx],      mm2                         ; save p1
+        movq        mm5,        mm2                         ; mm5 = p1
+
+        psubusb     mm2,        mm1                         ; p1-p2
+        psubusb     mm1,        mm5                         ; p2-p1
+
+        por         mm1,        mm2                         ; mm1=abs(p2-p1)
+        mov         rdx,        arg(3) ;limit
+
+        movq        mm4,        [rdx]                       ; mm4 = limit
+        psubusb     mm7,        mm4
+
+        psubusb     mm0,        mm4
+        psubusb     mm1,        mm4
+
+        psubusb     mm6,        mm4
+        por         mm7,        mm6
+
+        por         mm0,        mm1
+        por         mm0,        mm7                         ;   abs(q3-q2) > limit || abs(p3-p2) > limit ||abs(p2-p1) > limit || abs(q2-q1) > limit
+
+        movq        mm1,        mm5                         ; p1
+
+        movq        mm7,        mm3                         ; mm3=mm7=p0
+        psubusb     mm7,        mm5                         ; p0 - p1
+
+        psubusb     mm5,        mm3                         ; p1 - p0
+        por         mm5,        mm7                         ; abs(p1-p0)
+
+        movq        t0,         mm5                         ; save abs(p1-p0)
+        lea         rdx,        srct
+
+        psubusb     mm5,        mm4
+        por         mm0,        mm5                         ; mm0=mask
+
+        movq        mm5,        [rdx+16]                    ; mm5=q0
+        movq        mm7,        [rdx+24]                    ; mm7=q1
+
+        movq        mm6,        mm5                         ; mm6=q0
+        movq        mm2,        mm7                         ; q1
+        psubusb     mm5,        mm7                         ; q0-q1
+
+        psubusb     mm7,        mm6                         ; q1-q0
+        por         mm7,        mm5                         ; abs(q1-q0)
+
+        movq        t1,         mm7                         ; save abs(q1-q0)
+        psubusb     mm7,        mm4
+
+        por         mm0,        mm7                         ; mask
+
+        movq        mm5,        mm2                         ; q1
+        psubusb     mm5,        mm1                         ; q1-=p1
+        psubusb     mm1,        mm2                         ; p1-=q1
+        por         mm5,        mm1                         ; abs(p1-q1)
+        pand        mm5,        [GLOBAL(tfe)]               ; set lsb of each byte to zero
+        psrlw       mm5,        1                           ; abs(p1-q1)/2
+
+        mov         rdx,        arg(2) ;blimit                      ;
+
+        movq        mm4,        [rdx]                       ;blimit
+        movq        mm1,        mm3                         ; mm1=mm3=p0
+
+        movq        mm7,        mm6                         ; mm7=mm6=q0
+        psubusb     mm1,        mm7                         ; p0-q0
+
+        psubusb     mm7,        mm3                         ; q0-p0
+        por         mm1,        mm7                         ; abs(q0-p0)
+        paddusb     mm1,        mm1                         ; abs(q0-p0)*2
+        paddusb     mm1,        mm5                         ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+        psubusb     mm1,        mm4                         ; abs (p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        por         mm1,        mm0;                        ; mask
+
+        pxor        mm0,        mm0
+        pcmpeqb     mm1,        mm0
+
+        ; calculate high edge variance
+        mov         rdx,        arg(4) ;thresh            ; get thresh
+        movq        mm7,        [rdx]
+        ;
+        movq        mm4,        t0              ; get abs (q1 - q0)
+        psubusb     mm4,        mm7
+
+        movq        mm3,        t1              ; get abs (p1 - p0)
+        psubusb     mm3,        mm7
+
+        por         mm4,        mm3             ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
+        pcmpeqb     mm4,        mm0
+
+        pcmpeqb     mm0,        mm0
+        pxor        mm4,        mm0
+
+
+
+        ; start work on filters
+        lea         rdx,        srct
+
+        movq        mm2,        [rdx]           ; p1
+        movq        mm7,        [rdx+24]        ; q1
+
+        movq        mm6,        [rdx+8]         ; p0
+        movq        mm0,        [rdx+16]        ; q0
+
+        pxor        mm2,        [GLOBAL(t80)]   ; p1 offset to convert to signed values
+        pxor        mm7,        [GLOBAL(t80)]   ; q1 offset to convert to signed values
+
+        psubsb      mm2,        mm7             ; p1 - q1
+        pand        mm2,        mm4             ; high var mask (hvm)(p1 - q1)
+
+        pxor        mm6,        [GLOBAL(t80)]   ; offset to convert to signed values
+        pxor        mm0,        [GLOBAL(t80)]   ; offset to convert to signed values
+
+        movq        mm3,        mm0             ; q0
+        psubsb      mm0,        mm6             ; q0 - p0
+
+        paddsb      mm2,        mm0             ; 1 * (q0 - p0) + hvm(p1 - q1)
+        paddsb      mm2,        mm0             ; 2 * (q0 - p0) + hvm(p1 - q1)
+
+        paddsb      mm2,        mm0             ; 3 * (q0 - p0) + hvm(p1 - q1)
+        pand       mm1,        mm2              ; mask filter values we don't care about
+
+        movq        mm2,        mm1
+        paddsb      mm1,        [GLOBAL(t4)]      ; 3* (q0 - p0) + hvm(p1 - q1) + 4
+
+        paddsb      mm2,        [GLOBAL(t3)]      ; 3* (q0 - p0) + hvm(p1 - q1) + 3
+        pxor        mm0,        mm0          ;
+
+        pxor        mm5,        mm5
+        punpcklbw   mm0,        mm2         ;
+
+        punpckhbw   mm5,        mm2         ;
+        psraw       mm0,        11              ;
+
+        psraw       mm5,        11
+        packsswb    mm0,        mm5
+
+        movq        mm2,        mm0         ;  (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
+
+        pxor        mm0,        mm0           ; 0
+        movq        mm5,        mm1           ; abcdefgh
+
+        punpcklbw   mm0,        mm1           ; e0f0g0h0
+        psraw       mm0,        11                ; sign extended shift right by 3
+
+        pxor        mm1,        mm1           ; 0
+        punpckhbw   mm1,        mm5           ; a0b0c0d0
+
+        psraw       mm1,        11                ; sign extended shift right by 3
+        movq        mm5,        mm0              ; save results
+
+        packsswb    mm0,        mm1           ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
+        paddsw      mm5,        [GLOBAL(ones)]
+
+        paddsw      mm1,        [GLOBAL(ones)]
+        psraw       mm5,        1                 ; partial shifted one more time for 2nd tap
+
+        psraw       mm1,        1                 ; partial shifted one more time for 2nd tap
+        packsswb    mm5,        mm1           ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>4
+
+        pandn       mm4,        mm5             ; high edge variance additive
+
+        paddsb      mm6,        mm2             ; p0+= p0 add
+        pxor        mm6,        [GLOBAL(t80)]   ; unoffset
+
+        ; mm6=p0                               ;
+        movq        mm1,        [rdx]           ; p1
+        pxor        mm1,        [GLOBAL(t80)]   ; reoffset
+
+        paddsb      mm1,        mm4                 ; p1+= p1 add
+        pxor        mm1,        [GLOBAL(t80)]       ; unoffset
+        ; mm6 = p0 mm1 = p1
+
+        psubsb      mm3,        mm0                 ; q0-= q0 add
+        pxor        mm3,        [GLOBAL(t80)]       ; unoffset
+
+        ; mm3 = q0
+        psubsb      mm7,        mm4                 ; q1-= q1 add
+        pxor        mm7,        [GLOBAL(t80)]       ; unoffset
+        ; mm7 = q1
+
+        ; transpose and write back
+        ; mm1 =    72 62 52 42 32 22 12 02
+        ; mm6 =    73 63 53 43 33 23 13 03
+        ; mm3 =    74 64 54 44 34 24 14 04
+        ; mm7 =    75 65 55 45 35 25 15 05
+
+        movq        mm2,        mm1             ; 72 62 52 42 32 22 12 02
+        punpcklbw   mm2,        mm6             ; 33 32 23 22 13 12 03 02
+
+        movq        mm4,        mm3             ; 74 64 54 44 34 24 14 04
+        punpckhbw   mm1,        mm6             ; 73 72 63 62 53 52 43 42
+
+        punpcklbw   mm4,        mm7             ; 35 34 25 24 15 14 05 04
+        punpckhbw   mm3,        mm7             ; 75 74 65 64 55 54 45 44
+
+        movq        mm6,        mm2             ; 33 32 23 22 13 12 03 02
+        punpcklwd   mm2,        mm4             ; 15 14 13 12 05 04 03 02
+
+        punpckhwd   mm6,        mm4             ; 35 34 33 32 25 24 23 22
+        movq        mm5,        mm1             ; 73 72 63 62 53 52 43 42
+
+        punpcklwd   mm1,        mm3             ; 55 54 53 52 45 44 43 42
+        punpckhwd   mm5,        mm3             ; 75 74 73 72 65 64 63 62
+
+
+        ; mm2 = 15 14 13 12 05 04 03 02
+        ; mm6 = 35 34 33 32 25 24 23 22
+        ; mm5 = 55 54 53 52 45 44 43 42
+        ; mm1 = 75 74 73 72 65 64 63 62
+
+
+
+        movd        [rsi+rax*4+2], mm2
+        psrlq       mm2,        32
+
+        movd        [rdi+rax*4+2], mm2
+        movd        [rsi+rax*2+2], mm6
+
+        psrlq       mm6,        32
+        movd        [rsi+rax+2],mm6
+
+        movd        [rsi+2],    mm1
+        psrlq       mm1,        32
+
+        movd        [rdi+2],    mm1
+        neg         rax
+
+        movd        [rdi+rax+2],mm5
+        psrlq       mm5,        32
+
+        movd        [rdi+rax*2+2], mm5
+
+        lea         rsi,        [rsi+rax*8]
+        dec         rcx
+        jnz         .next8_v
+
+    add rsp, 64
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_mbloop_filter_horizontal_edge_mmx
+;(
+;    unsigned char *src_ptr,
+;    int  src_pixel_step,
+;    const char *blimit,
+;    const char *limit,
+;    const char *thresh,
+;    int count
+;)
+global sym(vp8_mbloop_filter_horizontal_edge_mmx) PRIVATE
+sym(vp8_mbloop_filter_horizontal_edge_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub          rsp, 32      ; reserve 32 bytes
+    %define t0   [rsp + 0]    ;__declspec(align(16)) char t0[8];
+    %define t1   [rsp + 16]   ;__declspec(align(16)) char t1[8];
+
+        mov         rsi, arg(0) ;src_ptr
+        movsxd      rax, dword ptr arg(1) ;src_pixel_step     ; destination pitch?
+
+        movsxd      rcx, dword ptr arg(5) ;count
+.next8_mbh:
+        mov         rdx, arg(3) ;limit
+        movq        mm7, [rdx]
+        mov         rdi, rsi              ; rdi points to row +1 for indirect addressing
+        add         rdi, rax
+
+        ; calculate breakout conditions
+        movq        mm2, [rdi+2*rax]      ; q3
+
+        movq        mm1, [rsi+2*rax]      ; q2
+        movq        mm6, mm1              ; q2
+        psubusb     mm1, mm2              ; q2-=q3
+        psubusb     mm2, mm6              ; q3-=q2
+        por         mm1, mm2              ; abs(q3-q2)
+        psubusb     mm1, mm7
+
+
+        ; mm1 = abs(q3-q2), mm6 =q2, mm7 = limit
+        movq        mm4, [rsi+rax]        ; q1
+        movq        mm3, mm4              ; q1
+        psubusb     mm4, mm6              ; q1-=q2
+        psubusb     mm6, mm3              ; q2-=q1
+        por         mm4, mm6              ; abs(q2-q1)
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+
+        ; mm1 = mask,      mm3=q1, mm7 = limit
+
+        movq        mm4, [rsi]            ; q0
+        movq        mm0, mm4              ; q0
+        psubusb     mm4, mm3              ; q0-=q1
+        psubusb     mm3, mm0              ; q1-=q0
+        por         mm4, mm3              ; abs(q0-q1)
+        movq        t0, mm4               ; save to t0
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+
+        ; mm1 = mask, mm0=q0,  mm7 = limit, t0 = abs(q0-q1)
+
+        neg         rax                   ; negate pitch to deal with above border
+
+        movq        mm2, [rsi+4*rax]      ; p3
+        movq        mm4, [rdi+4*rax]      ; p2
+        movq        mm5, mm4              ; p2
+        psubusb     mm4, mm2              ; p2-=p3
+        psubusb     mm2, mm5              ; p3-=p2
+        por         mm4, mm2              ; abs(p3 - p2)
+        psubusb     mm4, mm7
+        por        mm1, mm4
+        ; mm1 = mask, mm0=q0,  mm7 = limit, t0 = abs(q0-q1)
+
+        movq        mm4, [rsi+2*rax]      ; p1
+        movq        mm3, mm4              ; p1
+        psubusb     mm4, mm5              ; p1-=p2
+        psubusb     mm5, mm3              ; p2-=p1
+        por         mm4, mm5              ; abs(p2 - p1)
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+        movq        mm2, mm3              ; p1
+
+
+        ; mm1 = mask, mm0=q0,  mm7 = limit, t0 = abs(q0-q1)
+
+        movq        mm4, [rsi+rax]        ; p0
+        movq        mm5, mm4              ; p0
+        psubusb     mm4, mm3              ; p0-=p1
+        psubusb     mm3, mm5              ; p1-=p0
+        por         mm4, mm3              ; abs(p1 - p0)
+        movq        t1, mm4               ; save to t1
+        psubusb     mm4, mm7
+        por        mm1, mm4
+        ; mm1 = mask, mm0=q0,  mm7 = limit, t0 = abs(q0-q1) t1 = abs(p1-p0)
+        ; mm5 = p0
+        movq        mm3, [rdi]            ; q1
+        movq        mm4, mm3              ; q1
+        psubusb     mm3, mm2              ; q1-=p1
+        psubusb     mm2, mm4              ; p1-=q1
+        por         mm2, mm3              ; abs(p1-q1)
+        pand        mm2, [GLOBAL(tfe)]    ; set lsb of each byte to zero
+        psrlw       mm2, 1                ; abs(p1-q1)/2
+
+        movq        mm6, mm5              ; p0
+        movq        mm3, mm0              ; q0
+        psubusb     mm5, mm3              ; p0-=q0
+        psubusb     mm3, mm6              ; q0-=p0
+        por         mm5, mm3              ; abs(p0 - q0)
+        paddusb     mm5, mm5              ; abs(p0-q0)*2
+        paddusb     mm5, mm2              ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+        mov         rdx, arg(2) ;blimit           ; get blimit
+        movq        mm7, [rdx]            ; blimit
+
+        psubusb     mm5,    mm7           ; abs (p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        por         mm1,    mm5
+        pxor        mm5,    mm5
+        pcmpeqb     mm1,    mm5           ; mask mm1
+
+        ; mm1 = mask, mm0=q0,  mm7 = blimit, t0 = abs(q0-q1) t1 = abs(p1-p0)
+        ; mm6 = p0,
+
+        ; calculate high edge variance
+        mov         rdx, arg(4) ;thresh           ; get thresh
+        movq        mm7, [rdx]            ;
+        movq        mm4, t0               ; get abs (q1 - q0)
+        psubusb     mm4, mm7
+        movq        mm3, t1               ; get abs (p1 - p0)
+        psubusb     mm3, mm7
+        paddb       mm4, mm3              ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
+
+        pcmpeqb     mm4,        mm5
+
+        pcmpeqb     mm5,        mm5
+        pxor        mm4,        mm5
+
+
+
+        ; mm1 = mask, mm0=q0,  mm7 = thresh, t0 = abs(q0-q1) t1 = abs(p1-p0)
+        ; mm6 = p0, mm4=hev
+        ; start work on filters
+        movq        mm2, [rsi+2*rax]      ; p1
+        movq        mm7, [rdi]            ; q1
+        pxor        mm2, [GLOBAL(t80)]    ; p1 offset to convert to signed values
+        pxor        mm7, [GLOBAL(t80)]    ; q1 offset to convert to signed values
+        psubsb      mm2, mm7              ; p1 - q1
+
+        pxor        mm6, [GLOBAL(t80)]    ; offset to convert to signed values
+        pxor        mm0, [GLOBAL(t80)]    ; offset to convert to signed values
+        movq        mm3, mm0              ; q0
+        psubsb      mm0, mm6              ; q0 - p0
+        paddsb      mm2, mm0              ; 1 * (q0 - p0) + (p1 - q1)
+        paddsb      mm2, mm0              ; 2 * (q0 - p0)
+        paddsb      mm2, mm0              ; 3 * (q0 - p0) + (p1 - q1)
+        pand        mm1, mm2              ; mask filter values we don't care about
+
+
+        ; mm1 = vp8_filter, mm4=hev, mm6=ps0, mm3=qs0
+        movq        mm2, mm1              ; vp8_filter
+        pand        mm2, mm4;             ; Filter2 = vp8_filter & hev
+
+        movq        mm5,        mm2       ;
+        paddsb      mm5,        [GLOBAL(t3)];
+
+        pxor        mm0, mm0              ; 0
+        pxor        mm7, mm7              ; 0
+
+        punpcklbw   mm0, mm5              ; e0f0g0h0
+        psraw       mm0, 11               ; sign extended shift right by 3
+        punpckhbw   mm7, mm5              ; a0b0c0d0
+        psraw       mm7, 11               ; sign extended shift right by 3
+        packsswb    mm0, mm7              ; Filter2 >>=3;
+
+        movq        mm5, mm0              ; Filter2
+
+        paddsb      mm2, [GLOBAL(t4)]     ; vp8_signed_char_clamp(Filter2 + 4)
+        pxor        mm0, mm0              ; 0
+        pxor        mm7, mm7              ; 0
+
+        punpcklbw   mm0, mm2              ; e0f0g0h0
+        psraw       mm0, 11               ; sign extended shift right by 3
+        punpckhbw   mm7, mm2              ; a0b0c0d0
+        psraw       mm7, 11               ; sign extended shift right by 3
+        packsswb    mm0, mm7              ; Filter2 >>=3;
+
+        ; mm0= filter2 mm1 = vp8_filter,  mm3 =qs0 mm5=s mm4 =hev mm6=ps0
+        psubsb      mm3, mm0              ; qs0 =qs0 - filter1
+        paddsb      mm6, mm5              ; ps0 =ps0 + Fitler2
+
+        ; mm1=vp8_filter, mm3=qs0, mm4 =hev mm6=ps0
+        ; vp8_filter &= ~hev;
+        ; Filter2 = vp8_filter;
+        pandn       mm4, mm1              ; vp8_filter&=~hev
+
+
+        ; mm3=qs0, mm4=filter2, mm6=ps0
+
+        ; u = vp8_signed_char_clamp((63 + Filter2 * 27)>>7);
+        ; s = vp8_signed_char_clamp(qs0 - u);
+        ; *oq0 = s^0x80;
+        ; s = vp8_signed_char_clamp(ps0 + u);
+        ; *op0 = s^0x80;
+        pxor        mm0, mm0
+
+        pxor        mm1, mm1
+        pxor        mm2, mm2
+        punpcklbw   mm1, mm4
+        punpckhbw   mm2, mm4
+        pmulhw      mm1, [GLOBAL(s27)]
+        pmulhw      mm2, [GLOBAL(s27)]
+        paddw       mm1, [GLOBAL(s63)]
+        paddw       mm2, [GLOBAL(s63)]
+        psraw       mm1, 7
+        psraw       mm2, 7
+        packsswb    mm1, mm2
+
+        psubsb      mm3, mm1
+        paddsb      mm6, mm1
+
+        pxor        mm3, [GLOBAL(t80)]
+        pxor        mm6, [GLOBAL(t80)]
+        movq        [rsi+rax], mm6
+        movq        [rsi],     mm3
+
+        ; roughly 2/7th difference across boundary
+        ; u = vp8_signed_char_clamp((63 + Filter2 * 18)>>7);
+        ; s = vp8_signed_char_clamp(qs1 - u);
+        ; *oq1 = s^0x80;
+        ; s = vp8_signed_char_clamp(ps1 + u);
+        ; *op1 = s^0x80;
+        pxor        mm1, mm1
+        pxor        mm2, mm2
+        punpcklbw   mm1, mm4
+        punpckhbw   mm2, mm4
+        pmulhw      mm1, [GLOBAL(s18)]
+        pmulhw      mm2, [GLOBAL(s18)]
+        paddw       mm1, [GLOBAL(s63)]
+        paddw       mm2, [GLOBAL(s63)]
+        psraw       mm1, 7
+        psraw       mm2, 7
+        packsswb    mm1, mm2
+
+        movq        mm3, [rdi]
+        movq        mm6, [rsi+rax*2]       ; p1
+
+        pxor        mm3, [GLOBAL(t80)]
+        pxor        mm6, [GLOBAL(t80)]
+
+        paddsb      mm6, mm1
+        psubsb      mm3, mm1
+
+        pxor        mm6, [GLOBAL(t80)]
+        pxor        mm3, [GLOBAL(t80)]
+        movq        [rdi], mm3
+        movq        [rsi+rax*2], mm6
+
+        ; roughly 1/7th difference across boundary
+        ; u = vp8_signed_char_clamp((63 + Filter2 * 9)>>7);
+        ; s = vp8_signed_char_clamp(qs2 - u);
+        ; *oq2 = s^0x80;
+        ; s = vp8_signed_char_clamp(ps2 + u);
+        ; *op2 = s^0x80;
+        pxor        mm1, mm1
+        pxor        mm2, mm2
+        punpcklbw   mm1, mm4
+        punpckhbw   mm2, mm4
+        pmulhw      mm1, [GLOBAL(s9)]
+        pmulhw      mm2, [GLOBAL(s9)]
+        paddw       mm1, [GLOBAL(s63)]
+        paddw       mm2, [GLOBAL(s63)]
+        psraw       mm1, 7
+        psraw       mm2, 7
+        packsswb    mm1, mm2
+
+
+        movq        mm6, [rdi+rax*4]
+        neg         rax
+        movq        mm3, [rdi+rax  ]
+
+        pxor        mm6, [GLOBAL(t80)]
+        pxor        mm3, [GLOBAL(t80)]
+
+        paddsb      mm6, mm1
+        psubsb      mm3, mm1
+
+        pxor        mm6, [GLOBAL(t80)]
+        pxor        mm3, [GLOBAL(t80)]
+        movq        [rdi+rax  ], mm3
+        neg         rax
+        movq        [rdi+rax*4], mm6
+
+;EARLY_BREAK_OUT:
+        neg         rax
+        add         rsi,8
+        dec         rcx
+        jnz         .next8_mbh
+
+    add rsp, 32
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_mbloop_filter_vertical_edge_mmx
+;(
+;    unsigned char *src_ptr,
+;    int  src_pixel_step,
+;    const char *blimit,
+;    const char *limit,
+;    const char *thresh,
+;    int count
+;)
+global sym(vp8_mbloop_filter_vertical_edge_mmx) PRIVATE
+sym(vp8_mbloop_filter_vertical_edge_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub          rsp, 96      ; reserve 96 bytes
+    %define t0   [rsp + 0]    ;__declspec(align(16)) char t0[8];
+    %define t1   [rsp + 16]   ;__declspec(align(16)) char t1[8];
+    %define srct [rsp + 32]   ;__declspec(align(16)) char srct[64];
+
+        mov         rsi,        arg(0) ;src_ptr
+        movsxd      rax,        dword ptr arg(1) ;src_pixel_step     ; destination pitch?
+
+        lea         rsi,        [rsi + rax*4 - 4]
+
+        movsxd      rcx,        dword ptr arg(5) ;count
+.next8_mbv:
+        lea         rdi,        [rsi + rax]  ; rdi points to row +1 for indirect addressing
+
+        ;transpose
+        movq        mm0,        [rdi+2*rax]                 ; 77 76 75 74 73 72 71 70
+        movq        mm6,        [rsi+2*rax]                 ; 67 66 65 64 63 62 61 60
+
+        movq        mm7,        mm6                         ; 77 76 75 74 73 72 71 70
+        punpckhbw   mm7,        mm0                         ; 77 67 76 66 75 65 74 64
+
+        punpcklbw   mm6,        mm0                         ; 73 63 72 62 71 61 70 60
+        movq        mm0,        [rsi+rax]                   ; 57 56 55 54 53 52 51 50
+
+        movq        mm4,        [rsi]                       ; 47 46 45 44 43 42 41 40
+        movq        mm5,        mm4                         ; 47 46 45 44 43 42 41 40
+
+        punpckhbw   mm5,        mm0                         ; 57 47 56 46 55 45 54 44
+        punpcklbw   mm4,        mm0                         ; 53 43 52 42 51 41 50 40
+
+        movq        mm3,        mm5                         ; 57 47 56 46 55 45 54 44
+        punpckhwd   mm5,        mm7                         ; 77 67 57 47 76 66 56 46
+
+        punpcklwd   mm3,        mm7                         ; 75 65 55 45 74 64 54 44
+        movq        mm2,        mm4                         ; 53 43 52 42 51 41 50 40
+
+        punpckhwd   mm4,        mm6                         ; 73 63 53 43 72 62 52 42
+        punpcklwd   mm2,        mm6                         ; 71 61 51 41 70 60 50 40
+
+        neg         rax
+
+        movq        mm7,        [rsi+rax]                   ; 37 36 35 34 33 32 31 30
+        movq        mm6,        [rsi+rax*2]                 ; 27 26 25 24 23 22 21 20
+
+        movq        mm1,        mm6                         ; 27 26 25 24 23 22 21 20
+        punpckhbw   mm6,        mm7                         ; 37 27 36 36 35 25 34 24
+
+        punpcklbw   mm1,        mm7                         ; 33 23 32 22 31 21 30 20
+
+        movq        mm7,        [rsi+rax*4];                ; 07 06 05 04 03 02 01 00
+        punpckhbw   mm7,        [rdi+rax*4]                 ; 17 07 16 06 15 05 14 04
+
+        movq        mm0,        mm7                         ; 17 07 16 06 15 05 14 04
+        punpckhwd   mm7,        mm6                         ; 37 27 17 07 36 26 16 06
+
+        punpcklwd   mm0,        mm6                         ; 35 25 15 05 34 24 14 04
+        movq        mm6,        mm7                         ; 37 27 17 07 36 26 16 06
+
+        punpckhdq   mm7,        mm5                         ; 77 67 57 47 37 27 17 07  = q3
+        punpckldq   mm6,        mm5                         ; 76 66 56 46 36 26 16 06  = q2
+
+        lea         rdx,        srct
+        movq        mm5,        mm6                         ; 76 66 56 46 36 26 16 06
+
+        movq        [rdx+56],   mm7
+        psubusb     mm5,        mm7                         ; q2-q3
+
+
+        movq        [rdx+48],   mm6
+        psubusb     mm7,        mm6                         ; q3-q2
+
+        por         mm7,        mm5;                        ; mm7=abs (q3-q2)
+        movq        mm5,        mm0                         ; 35 25 15 05 34 24 14 04
+
+        punpckhdq   mm5,        mm3                         ; 75 65 55 45 35 25 15 05 = q1
+        punpckldq   mm0,        mm3                         ; 74 64 54 44 34 24 15 04 = q0
+
+        movq        mm3,        mm5                         ; 75 65 55 45 35 25 15 05 = q1
+        psubusb     mm3,        mm6                         ; q1-q2
+
+        psubusb     mm6,        mm5                         ; q2-q1
+        por         mm6,        mm3                         ; mm6=abs(q2-q1)
+
+        movq        [rdx+40],   mm5                         ; save q1
+        movq        [rdx+32],   mm0                         ; save q0
+
+        movq        mm3,        [rsi+rax*4]                 ; 07 06 05 04 03 02 01 00
+        punpcklbw   mm3,        [rdi+rax*4]                 ; 13 03 12 02 11 01 10 00
+
+        movq        mm0,        mm3                         ; 13 03 12 02 11 01 10 00
+        punpcklwd   mm0,        mm1                         ; 31 21 11 01 30 20 10 00
+
+        punpckhwd   mm3,        mm1                         ; 33 23 13 03 32 22 12 02
+        movq        mm1,        mm0                         ; 31 21 11 01 30 20 10 00
+
+        punpckldq   mm0,        mm2                         ; 70 60 50 40 30 20 10 00  =p3
+        punpckhdq   mm1,        mm2                         ; 71 61 51 41 31 21 11 01  =p2
+
+        movq        [rdx],      mm0                         ; save p3
+        movq        [rdx+8],    mm1                         ; save p2
+
+        movq        mm2,        mm1                         ; 71 61 51 41 31 21 11 01  =p2
+        psubusb     mm2,        mm0                         ; p2-p3
+
+        psubusb     mm0,        mm1                         ; p3-p2
+        por         mm0,        mm2                         ; mm0=abs(p3-p2)
+
+        movq        mm2,        mm3                         ; 33 23 13 03 32 22 12 02
+        punpckldq   mm2,        mm4                         ; 72 62 52 42 32 22 12 02 = p1
+
+        punpckhdq   mm3,        mm4                         ; 73 63 53 43 33 23 13 03 = p0
+        movq        [rdx+24],   mm3                         ; save p0
+
+        movq        [rdx+16],   mm2                         ; save p1
+        movq        mm5,        mm2                         ; mm5 = p1
+
+        psubusb     mm2,        mm1                         ; p1-p2
+        psubusb     mm1,        mm5                         ; p2-p1
+
+        por         mm1,        mm2                         ; mm1=abs(p2-p1)
+        mov         rdx,        arg(3) ;limit
+
+        movq        mm4,        [rdx]                       ; mm4 = limit
+        psubusb     mm7,        mm4                         ; abs(q3-q2) > limit
+
+        psubusb     mm0,        mm4                         ; abs(p3-p2) > limit
+        psubusb     mm1,        mm4                         ; abs(p2-p1) > limit
+
+        psubusb     mm6,        mm4                         ; abs(q2-q1) > limit
+        por         mm7,        mm6                         ; or
+
+        por         mm0,        mm1                         ;
+        por         mm0,        mm7                         ; abs(q3-q2) > limit || abs(p3-p2) > limit ||abs(p2-p1) > limit || abs(q2-q1) > limit
+
+        movq        mm1,        mm5                         ; p1
+
+        movq        mm7,        mm3                         ; mm3=mm7=p0
+        psubusb     mm7,        mm5                         ; p0 - p1
+
+        psubusb     mm5,        mm3                         ; p1 - p0
+        por         mm5,        mm7                         ; abs(p1-p0)
+
+        movq        t0,         mm5                         ; save abs(p1-p0)
+        lea         rdx,        srct
+
+        psubusb     mm5,        mm4                         ; mm5 = abs(p1-p0) > limit
+        por         mm0,        mm5                         ; mm0=mask
+
+        movq        mm5,        [rdx+32]                    ; mm5=q0
+        movq        mm7,        [rdx+40]                    ; mm7=q1
+
+        movq        mm6,        mm5                         ; mm6=q0
+        movq        mm2,        mm7                         ; q1
+        psubusb     mm5,        mm7                         ; q0-q1
+
+        psubusb     mm7,        mm6                         ; q1-q0
+        por         mm7,        mm5                         ; abs(q1-q0)
+
+        movq        t1,         mm7                         ; save abs(q1-q0)
+        psubusb     mm7,        mm4                         ; mm7=abs(q1-q0)> limit
+
+        por         mm0,        mm7                         ; mask
+
+        movq        mm5,        mm2                         ; q1
+        psubusb     mm5,        mm1                         ; q1-=p1
+        psubusb     mm1,        mm2                         ; p1-=q1
+        por         mm5,        mm1                         ; abs(p1-q1)
+        pand        mm5,        [GLOBAL(tfe)]               ; set lsb of each byte to zero
+        psrlw       mm5,        1                           ; abs(p1-q1)/2
+
+        mov         rdx,        arg(2) ;blimit                      ;
+
+        movq        mm4,        [rdx]                       ;blimit
+        movq        mm1,        mm3                         ; mm1=mm3=p0
+
+        movq        mm7,        mm6                         ; mm7=mm6=q0
+        psubusb     mm1,        mm7                         ; p0-q0
+
+        psubusb     mm7,        mm3                         ; q0-p0
+        por         mm1,        mm7                         ; abs(q0-p0)
+        paddusb     mm1,        mm1                         ; abs(q0-p0)*2
+        paddusb     mm1,        mm5                         ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+        psubusb     mm1,        mm4                         ; abs (p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        por         mm1,        mm0;                        ; mask
+
+        pxor        mm0,        mm0
+        pcmpeqb     mm1,        mm0
+
+        ; calculate high edge variance
+        mov         rdx,        arg(4) ;thresh            ; get thresh
+        movq        mm7,        [rdx]
+        ;
+        movq        mm4,        t0              ; get abs (q1 - q0)
+        psubusb     mm4,        mm7             ; abs(q1 - q0) > thresh
+
+        movq        mm3,        t1              ; get abs (p1 - p0)
+        psubusb     mm3,        mm7             ; abs(p1 - p0)> thresh
+
+        por         mm4,        mm3             ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
+        pcmpeqb     mm4,        mm0
+
+        pcmpeqb     mm0,        mm0
+        pxor        mm4,        mm0
+
+
+
+
+        ; start work on filters
+        lea         rdx,        srct
+
+        ; start work on filters
+        movq        mm2, [rdx+16]         ; p1
+        movq        mm7, [rdx+40]         ; q1
+        pxor        mm2, [GLOBAL(t80)]    ; p1 offset to convert to signed values
+        pxor        mm7, [GLOBAL(t80)]    ; q1 offset to convert to signed values
+        psubsb      mm2, mm7              ; p1 - q1
+
+        movq        mm6, [rdx+24]         ; p0
+        movq        mm0, [rdx+32]         ; q0
+        pxor        mm6, [GLOBAL(t80)]    ; offset to convert to signed values
+        pxor        mm0, [GLOBAL(t80)]    ; offset to convert to signed values
+
+        movq        mm3, mm0              ; q0
+        psubsb      mm0, mm6              ; q0 - p0
+        paddsb      mm2, mm0              ; 1 * (q0 - p0) + (p1 - q1)
+        paddsb      mm2, mm0              ; 2 * (q0 - p0)
+        paddsb      mm2, mm0              ; 3 * (q0 - p0) + (p1 - q1)
+        pand       mm1, mm2           ; mask filter values we don't care about
+
+        ; mm1 = vp8_filter, mm4=hev, mm6=ps0, mm3=qs0
+        movq        mm2, mm1              ; vp8_filter
+        pand        mm2, mm4;             ; Filter2 = vp8_filter & hev
+
+        movq        mm5,        mm2       ;
+        paddsb      mm5,        [GLOBAL(t3)];
+
+        pxor        mm0, mm0              ; 0
+        pxor        mm7, mm7              ; 0
+
+        punpcklbw   mm0, mm5              ; e0f0g0h0
+        psraw       mm0, 11               ; sign extended shift right by 3
+        punpckhbw   mm7, mm5              ; a0b0c0d0
+        psraw       mm7, 11               ; sign extended shift right by 3
+        packsswb    mm0, mm7              ; Filter2 >>=3;
+
+        movq        mm5, mm0              ; Filter2
+
+        paddsb      mm2, [GLOBAL(t4)]     ; vp8_signed_char_clamp(Filter2 + 4)
+        pxor        mm0, mm0              ; 0
+        pxor        mm7, mm7              ; 0
+
+        punpcklbw   mm0, mm2              ; e0f0g0h0
+        psraw       mm0, 11               ; sign extended shift right by 3
+        punpckhbw   mm7, mm2              ; a0b0c0d0
+        psraw       mm7, 11               ; sign extended shift right by 3
+        packsswb    mm0, mm7              ; Filter2 >>=3;
+
+        ; mm0= filter2 mm1 = vp8_filter,  mm3 =qs0 mm5=s mm4 =hev mm6=ps0
+        psubsb      mm3, mm0              ; qs0 =qs0 - filter1
+        paddsb      mm6, mm5              ; ps0 =ps0 + Fitler2
+
+        ; mm1=vp8_filter, mm3=qs0, mm4 =hev mm6=ps0
+        ; vp8_filter &= ~hev;
+        ; Filter2 = vp8_filter;
+        pandn       mm4, mm1              ; vp8_filter&=~hev
+
+
+        ; mm3=qs0, mm4=filter2, mm6=ps0
+
+        ; u = vp8_signed_char_clamp((63 + Filter2 * 27)>>7);
+        ; s = vp8_signed_char_clamp(qs0 - u);
+        ; *oq0 = s^0x80;
+        ; s = vp8_signed_char_clamp(ps0 + u);
+        ; *op0 = s^0x80;
+        pxor        mm0, mm0
+
+        pxor        mm1, mm1
+        pxor        mm2, mm2
+        punpcklbw   mm1, mm4
+        punpckhbw   mm2, mm4
+        pmulhw      mm1, [GLOBAL(s27)]
+        pmulhw      mm2, [GLOBAL(s27)]
+        paddw       mm1, [GLOBAL(s63)]
+        paddw       mm2, [GLOBAL(s63)]
+        psraw       mm1, 7
+        psraw       mm2, 7
+        packsswb    mm1, mm2
+
+        psubsb      mm3, mm1
+        paddsb      mm6, mm1
+
+        pxor        mm3, [GLOBAL(t80)]
+        pxor        mm6, [GLOBAL(t80)]
+        movq        [rdx+24], mm6
+        movq        [rdx+32], mm3
+
+        ; roughly 2/7th difference across boundary
+        ; u = vp8_signed_char_clamp((63 + Filter2 * 18)>>7);
+        ; s = vp8_signed_char_clamp(qs1 - u);
+        ; *oq1 = s^0x80;
+        ; s = vp8_signed_char_clamp(ps1 + u);
+        ; *op1 = s^0x80;
+        pxor        mm1, mm1
+        pxor        mm2, mm2
+        punpcklbw   mm1, mm4
+        punpckhbw   mm2, mm4
+        pmulhw      mm1, [GLOBAL(s18)]
+        pmulhw      mm2, [GLOBAL(s18)]
+        paddw       mm1, [GLOBAL(s63)]
+        paddw       mm2, [GLOBAL(s63)]
+        psraw       mm1, 7
+        psraw       mm2, 7
+        packsswb    mm1, mm2
+
+        movq        mm3, [rdx + 40]
+        movq        mm6, [rdx + 16]       ; p1
+        pxor        mm3, [GLOBAL(t80)]
+        pxor        mm6, [GLOBAL(t80)]
+
+        paddsb      mm6, mm1
+        psubsb      mm3, mm1
+
+        pxor        mm6, [GLOBAL(t80)]
+        pxor        mm3, [GLOBAL(t80)]
+        movq        [rdx + 40], mm3
+        movq        [rdx + 16], mm6
+
+        ; roughly 1/7th difference across boundary
+        ; u = vp8_signed_char_clamp((63 + Filter2 * 9)>>7);
+        ; s = vp8_signed_char_clamp(qs2 - u);
+        ; *oq2 = s^0x80;
+        ; s = vp8_signed_char_clamp(ps2 + u);
+        ; *op2 = s^0x80;
+        pxor        mm1, mm1
+        pxor        mm2, mm2
+        punpcklbw   mm1, mm4
+        punpckhbw   mm2, mm4
+        pmulhw      mm1, [GLOBAL(s9)]
+        pmulhw      mm2, [GLOBAL(s9)]
+        paddw       mm1, [GLOBAL(s63)]
+        paddw       mm2, [GLOBAL(s63)]
+        psraw       mm1, 7
+        psraw       mm2, 7
+        packsswb    mm1, mm2
+
+        movq        mm6, [rdx+ 8]
+        movq        mm3, [rdx+48]
+
+        pxor        mm6, [GLOBAL(t80)]
+        pxor        mm3, [GLOBAL(t80)]
+
+        paddsb      mm6, mm1
+        psubsb      mm3, mm1
+
+        pxor        mm6, [GLOBAL(t80)]          ; mm6 = 71 61 51 41 31 21 11 01
+        pxor        mm3, [GLOBAL(t80)]          ; mm3 = 76 66 56 46 36 26 15 06
+
+        ; transpose and write back
+        movq        mm0,    [rdx]               ; mm0 = 70 60 50 40 30 20 10 00
+        movq        mm1,    mm0                 ; mm0 = 70 60 50 40 30 20 10 00
+
+        punpcklbw   mm0,    mm6                 ; mm0 = 31 30 21 20 11 10 01 00
+        punpckhbw   mm1,    mm6                 ; mm3 = 71 70 61 60 51 50 41 40
+
+        movq        mm2,    [rdx+16]            ; mm2 = 72 62 52 42 32 22 12 02
+        movq        mm6,    mm2                 ; mm3 = 72 62 52 42 32 22 12 02
+
+        punpcklbw   mm2,    [rdx+24]            ; mm2 = 33 32 23 22 13 12 03 02
+        punpckhbw   mm6,    [rdx+24]            ; mm3 = 73 72 63 62 53 52 43 42
+
+        movq        mm5,    mm0                 ; mm5 = 31 30 21 20 11 10 01 00
+        punpcklwd   mm0,    mm2                 ; mm0 = 13 12 11 10 03 02 01 00
+
+        punpckhwd   mm5,    mm2                 ; mm5 = 33 32 31 30 23 22 21 20
+        movq        mm4,    mm1                 ; mm4 = 71 70 61 60 51 50 41 40
+
+        punpcklwd   mm1,    mm6                 ; mm1 = 53 52 51 50 43 42 41 40
+        punpckhwd   mm4,    mm6                 ; mm4 = 73 72 71 70 63 62 61 60
+
+        movq        mm2,    [rdx+32]            ; mm2 = 74 64 54 44 34 24 14 04
+        punpcklbw   mm2,    [rdx+40]            ; mm2 = 35 34 25 24 15 14 05 04
+
+        movq        mm6,    mm3                 ; mm6 = 76 66 56 46 36 26 15 06
+        punpcklbw   mm6,    [rdx+56]            ; mm6 = 37 36 27 26 17 16 07 06
+
+        movq        mm7,    mm2                 ; mm7 = 35 34 25 24 15 14 05 04
+        punpcklwd   mm2,    mm6                 ; mm2 = 17 16 15 14 07 06 05 04
+
+        punpckhwd   mm7,    mm6                 ; mm7 = 37 36 35 34 27 26 25 24
+        movq        mm6,    mm0                 ; mm6 = 13 12 11 10 03 02 01 00
+
+        punpckldq   mm0,    mm2                 ; mm0 = 07 06 05 04 03 02 01 00
+        punpckhdq   mm6,    mm2                 ; mm6 = 17 16 15 14 13 12 11 10
+
+        movq        [rsi+rax*4], mm0            ; write out
+        movq        [rdi+rax*4], mm6            ; write out
+
+        movq        mm0,    mm5                 ; mm0 = 33 32 31 30 23 22 21 20
+        punpckldq   mm0,    mm7                 ; mm0 = 27 26 25 24 23 22 20 20
+
+        punpckhdq   mm5,    mm7                 ; mm5 = 37 36 35 34 33 32 31 30
+        movq        [rsi+rax*2], mm0            ; write out
+
+        movq        [rdi+rax*2], mm5            ; write out
+        movq        mm2,    [rdx+32]            ; mm2 = 74 64 54 44 34 24 14 04
+
+        punpckhbw   mm2,    [rdx+40]            ; mm2 = 75 74 65 64 54 54 45 44
+        punpckhbw   mm3,    [rdx+56]            ; mm3 = 77 76 67 66 57 56 47 46
+
+        movq        mm5,    mm2                 ; mm5 = 75 74 65 64 54 54 45 44
+        punpcklwd   mm2,    mm3                 ; mm2 = 57 56 55 54 47 46 45 44
+
+        punpckhwd   mm5,    mm3                 ; mm5 = 77 76 75 74 67 66 65 64
+        movq        mm0,    mm1                 ; mm0=  53 52 51 50 43 42 41 40
+
+        movq        mm3,    mm4                 ; mm4 = 73 72 71 70 63 62 61 60
+        punpckldq   mm0,    mm2                 ; mm0 = 47 46 45 44 43 42 41 40
+
+        punpckhdq   mm1,    mm2                 ; mm1 = 57 56 55 54 53 52 51 50
+        movq        [rsi],  mm0                 ; write out
+
+        movq        [rdi],  mm1                 ; write out
+        neg         rax
+
+        punpckldq   mm3,    mm5                 ; mm3 = 67 66 65 64 63 62 61 60
+        punpckhdq   mm4,    mm5                 ; mm4 = 77 76 75 74 73 72 71 60
+
+        movq        [rsi+rax*2], mm3
+        movq        [rdi+rax*2], mm4
+
+        lea         rsi,        [rsi+rax*8]
+        dec         rcx
+
+        jnz         .next8_mbv
+
+    add rsp, 96
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_loop_filter_simple_horizontal_edge_mmx
+;(
+;    unsigned char *src_ptr,
+;    int  src_pixel_step,
+;    const char *blimit
+;)
+global sym(vp8_loop_filter_simple_horizontal_edge_mmx) PRIVATE
+sym(vp8_loop_filter_simple_horizontal_edge_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 3
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rsi, arg(0) ;src_ptr
+        movsxd      rax, dword ptr arg(1) ;src_pixel_step     ; destination pitch?
+
+        mov         rcx, 2                ; count
+.nexts8_h:
+        mov         rdx, arg(2) ;blimit           ; get blimit
+        movq        mm3, [rdx]            ;
+
+        mov         rdi, rsi              ; rdi points to row +1 for indirect addressing
+        add         rdi, rax
+        neg         rax
+
+        ; calculate mask
+        movq        mm1, [rsi+2*rax]      ; p1
+        movq        mm0, [rdi]            ; q1
+        movq        mm2, mm1
+        movq        mm7, mm0
+        movq        mm4, mm0
+        psubusb     mm0, mm1              ; q1-=p1
+        psubusb     mm1, mm4              ; p1-=q1
+        por         mm1, mm0              ; abs(p1-q1)
+        pand        mm1, [GLOBAL(tfe)]    ; set lsb of each byte to zero
+        psrlw       mm1, 1                ; abs(p1-q1)/2
+
+        movq        mm5, [rsi+rax]        ; p0
+        movq        mm4, [rsi]            ; q0
+        movq        mm0, mm4              ; q0
+        movq        mm6, mm5              ; p0
+        psubusb     mm5, mm4              ; p0-=q0
+        psubusb     mm4, mm6              ; q0-=p0
+        por         mm5, mm4              ; abs(p0 - q0)
+        paddusb     mm5, mm5              ; abs(p0-q0)*2
+        paddusb     mm5, mm1              ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+        psubusb     mm5, mm3              ; abs(p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        pxor        mm3, mm3
+        pcmpeqb     mm5, mm3
+
+        ; start work on filters
+        pxor        mm2, [GLOBAL(t80)]    ; p1 offset to convert to signed values
+        pxor        mm7, [GLOBAL(t80)]    ; q1 offset to convert to signed values
+        psubsb      mm2, mm7              ; p1 - q1
+
+        pxor        mm6, [GLOBAL(t80)]    ; offset to convert to signed values
+        pxor        mm0, [GLOBAL(t80)]    ; offset to convert to signed values
+        movq        mm3, mm0              ; q0
+        psubsb      mm0, mm6              ; q0 - p0
+        paddsb      mm2, mm0              ; p1 - q1 + 1 * (q0 - p0)
+        paddsb      mm2, mm0              ; p1 - q1 + 2 * (q0 - p0)
+        paddsb      mm2, mm0              ; p1 - q1 + 3 * (q0 - p0)
+        pand        mm5, mm2              ; mask filter values we don't care about
+
+        ; do + 4 side
+        paddsb      mm5, [GLOBAL(t4)]     ; 3* (q0 - p0) + (p1 - q1) + 4
+
+        movq        mm0, mm5              ; get a copy of filters
+        psllw       mm0, 8                ; shift left 8
+        psraw       mm0, 3                ; arithmetic shift right 11
+        psrlw       mm0, 8
+        movq        mm1, mm5              ; get a copy of filters
+        psraw       mm1, 11               ; arithmetic shift right 11
+        psllw       mm1, 8                ; shift left 8 to put it back
+
+        por         mm0, mm1              ; put the two together to get result
+
+        psubsb      mm3, mm0              ; q0-= q0 add
+        pxor        mm3, [GLOBAL(t80)]    ; unoffset
+        movq        [rsi], mm3            ; write back
+
+
+        ; now do +3 side
+        psubsb      mm5, [GLOBAL(t1s)]     ; +3 instead of +4
+
+        movq        mm0, mm5              ; get a copy of filters
+        psllw       mm0, 8                ; shift left 8
+        psraw       mm0, 3                ; arithmetic shift right 11
+        psrlw       mm0, 8
+        psraw       mm5, 11               ; arithmetic shift right 11
+        psllw       mm5, 8                ; shift left 8 to put it back
+        por         mm0, mm5              ; put the two together to get result
+
+
+        paddsb      mm6, mm0              ; p0+= p0 add
+        pxor        mm6, [GLOBAL(t80)]    ; unoffset
+        movq        [rsi+rax], mm6        ; write back
+
+        add         rsi,8
+        neg         rax
+        dec         rcx
+        jnz         .nexts8_h
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp8_loop_filter_simple_vertical_edge_mmx
+;(
+;    unsigned char *src_ptr,
+;    int  src_pixel_step,
+;    const char *blimit
+;)
+global sym(vp8_loop_filter_simple_vertical_edge_mmx) PRIVATE
+sym(vp8_loop_filter_simple_vertical_edge_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 3
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub          rsp, 32      ; reserve 32 bytes
+    %define t0   [rsp + 0]    ;__declspec(align(16)) char t0[8];
+    %define t1   [rsp + 16]   ;__declspec(align(16)) char t1[8];
+
+        mov         rsi, arg(0) ;src_ptr
+        movsxd      rax, dword ptr arg(1) ;src_pixel_step     ; destination pitch?
+
+        lea         rsi, [rsi + rax*4- 2];  ;
+        mov         rcx, 2                                      ; count
+.nexts8_v:
+
+        lea         rdi,        [rsi + rax];
+        movd        mm0,        [rdi + rax * 2]                 ; xx xx xx xx 73 72 71 70
+
+        movd        mm6,        [rsi + rax * 2]                 ; xx xx xx xx 63 62 61 60
+        punpcklbw   mm6,        mm0                             ; 73 63 72 62 71 61 70 60
+
+        movd        mm0,        [rsi + rax]                     ; xx xx xx xx 53 52 51 50
+        movd        mm4,        [rsi]                           ; xx xx xx xx 43 42 41 40
+
+        punpcklbw   mm4,        mm0                             ; 53 43 52 42 51 41 50 40
+        movq        mm5,        mm4                             ; 53 43 52 42 51 41 50 40
+
+        punpcklwd   mm4,        mm6                             ; 71 61 51 41 70 60 50 40
+        punpckhwd   mm5,        mm6                             ; 73 63 53 43 72 62 52 42
+
+        neg         rax
+
+        movd        mm7,        [rsi + rax]                     ; xx xx xx xx 33 32 31 30
+        movd        mm6,        [rsi + rax * 2]                 ; xx xx xx xx 23 22 21 20
+
+        punpcklbw   mm6,        mm7                             ; 33 23 32 22 31 21 30 20
+        movd        mm1,        [rdi + rax * 4]                 ; xx xx xx xx 13 12 11 10
+
+        movd        mm0,        [rsi + rax * 4]                 ; xx xx xx xx 03 02 01 00
+        punpcklbw   mm0,        mm1                             ; 13 03 12 02 11 01 10 00
+
+        movq        mm2,        mm0                             ; 13 03 12 02 11 01 10 00
+        punpcklwd   mm0,        mm6                             ; 31 21 11 01 30 20 10 00
+
+        punpckhwd   mm2,        mm6                             ; 33 23 13 03 32 22 12 02
+        movq        mm1,        mm0                             ; 13 03 12 02 11 01 10 00
+
+        punpckldq   mm0,        mm4                             ; 70 60 50 40 30 20 10 00       = p1
+        movq        mm3,        mm2                             ; 33 23 13 03 32 22 12 02
+
+        punpckhdq   mm1,        mm4                             ; 71 61 51 41 31 21 11 01       = p0
+        punpckldq   mm2,        mm5                             ; 72 62 52 42 32 22 12 02       = q0
+
+        punpckhdq   mm3,        mm5                             ; 73 63 53 43 33 23 13 03       = q1
+
+
+        ; calculate mask
+        movq        mm6,        mm0                             ; p1
+        movq        mm7,        mm3                             ; q1
+        psubusb     mm7,        mm6                             ; q1-=p1
+        psubusb     mm6,        mm3                             ; p1-=q1
+        por         mm6,        mm7                             ; abs(p1-q1)
+        pand        mm6,        [GLOBAL(tfe)]                   ; set lsb of each byte to zero
+        psrlw       mm6,        1                               ; abs(p1-q1)/2
+
+        movq        mm5,        mm1                             ; p0
+        movq        mm4,        mm2                             ; q0
+
+        psubusb     mm5,        mm2                             ; p0-=q0
+        psubusb     mm4,        mm1                             ; q0-=p0
+
+        por         mm5,        mm4                             ; abs(p0 - q0)
+        paddusb     mm5,        mm5                             ; abs(p0-q0)*2
+        paddusb     mm5,        mm6                             ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+        mov         rdx,        arg(2) ;blimit                          ; get blimit
+        movq        mm7,        [rdx]
+
+        psubusb     mm5,        mm7                             ; abs(p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        pxor        mm7,        mm7
+        pcmpeqb     mm5,        mm7                             ; mm5 = mask
+
+        ; start work on filters
+        movq        t0,         mm0
+        movq        t1,         mm3
+
+        pxor        mm0,        [GLOBAL(t80)]                   ; p1 offset to convert to signed values
+        pxor        mm3,        [GLOBAL(t80)]                   ; q1 offset to convert to signed values
+
+        psubsb      mm0,        mm3                             ; p1 - q1
+        movq        mm6,        mm1                             ; p0
+
+        movq        mm7,        mm2                             ; q0
+        pxor        mm6,        [GLOBAL(t80)]                   ; offset to convert to signed values
+
+        pxor        mm7,        [GLOBAL(t80)]                   ; offset to convert to signed values
+        movq        mm3,        mm7                             ; offseted ; q0
+
+        psubsb      mm7,        mm6                             ; q0 - p0
+        paddsb      mm0,        mm7                             ; p1 - q1 + 1 * (q0 - p0)
+
+        paddsb      mm0,        mm7                             ; p1 - q1 + 2 * (q0 - p0)
+        paddsb      mm0,        mm7                             ; p1 - q1 + 3 * (q0 - p0)
+
+        pand        mm5,        mm0                             ; mask filter values we don't care about
+
+        paddsb      mm5,        [GLOBAL(t4)]                    ;  3* (q0 - p0) + (p1 - q1) + 4
+
+        movq        mm0,        mm5                             ; get a copy of filters
+        psllw       mm0,        8                               ; shift left 8
+        psraw       mm0,        3                               ; arithmetic shift right 11
+        psrlw       mm0,        8
+
+        movq        mm7,        mm5                             ; get a copy of filters
+        psraw       mm7,        11                              ; arithmetic shift right 11
+        psllw       mm7,        8                               ; shift left 8 to put it back
+
+        por         mm0,        mm7                             ; put the two together to get result
+
+        psubsb      mm3,        mm0                             ; q0-= q0sz add
+        pxor        mm3,        [GLOBAL(t80)]                   ; unoffset
+
+        ; now do +3 side
+        psubsb      mm5, [GLOBAL(t1s)]                          ; +3 instead of +4
+
+        movq        mm0, mm5                                    ; get a copy of filters
+        psllw       mm0, 8                                      ; shift left 8
+        psraw       mm0, 3                                      ; arithmetic shift right 11
+        psrlw       mm0, 8
+
+        psraw       mm5, 11                                     ; arithmetic shift right 11
+        psllw       mm5, 8                                      ; shift left 8 to put it back
+        por         mm0, mm5                                    ; put the two together to get result
+
+        paddsb      mm6, mm0                                    ; p0+= p0 add
+        pxor        mm6, [GLOBAL(t80)]                          ; unoffset
+
+
+        movq        mm0,        t0
+        movq        mm4,        t1
+
+        ; mm0 = 70 60 50 40 30 20 10 00
+        ; mm6 = 71 61 51 41 31 21 11 01
+        ; mm3 = 72 62 52 42 32 22 12 02
+        ; mm4 = 73 63 53 43 33 23 13 03
+        ; transpose back to write out
+
+        movq        mm1,        mm0                         ;
+        punpcklbw   mm0,        mm6                         ; 31 30 21 20 11 10 01 00
+
+        punpckhbw   mm1,        mm6                         ; 71 70 61 60 51 50 41 40
+        movq        mm2,        mm3                         ;
+
+        punpcklbw   mm2,        mm4                         ; 33 32 23 22 13 12 03 02
+        movq        mm5,        mm1                         ; 71 70 61 60 51 50 41 40
+
+        punpckhbw   mm3,        mm4                         ; 73 72 63 62 53 52 43 42
+        movq        mm6,        mm0                         ; 31 30 21 20 11 10 01 00
+
+        punpcklwd   mm0,        mm2                         ; 13 12 11 10 03 02 01 00
+        punpckhwd   mm6,        mm2                         ; 33 32 31 30 23 22 21 20
+
+        movd        [rsi+rax*4], mm0                        ; write 03 02 01 00
+        punpcklwd   mm1,        mm3                         ; 53 52 51 50 43 42 41 40
+
+        psrlq       mm0,        32                          ; xx xx xx xx 13 12 11 10
+        punpckhwd   mm5,        mm3                         ; 73 72 71 70 63 62 61 60
+
+        movd        [rdi+rax*4], mm0                        ; write 13 12 11 10
+        movd        [rsi+rax*2], mm6                        ; write 23 22 21 20
+
+        psrlq       mm6,        32                          ; 33 32 31 30
+        movd        [rsi],      mm1                         ; write 43 42 41 40
+
+        movd        [rsi + rax], mm6                        ; write 33 32 31 30
+        neg         rax
+
+        movd        [rsi + rax*2], mm5                      ; write 63 62 61 60
+        psrlq       mm1,        32                          ; 53 52 51 50
+
+        movd        [rdi],      mm1                         ; write out 53 52 51 50
+        psrlq       mm5,        32                          ; 73 72 71 70
+
+        movd        [rdi + rax*2], mm5                      ; write 73 72 71 70
+
+        lea         rsi,        [rsi+rax*8]                 ; next 8
+
+        dec         rcx
+        jnz         .nexts8_v
+
+    add rsp, 32
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+
+;void fast_loop_filter_vertical_edges_mmx(unsigned char *y_ptr,
+;                  int y_stride,
+;                  loop_filter_info *lfi)
+;{
+;
+;
+;    vp8_loop_filter_simple_vertical_edge_mmx(y_ptr+4, y_stride, lfi->flim,lfi->lim,lfi->thr,2);
+;    vp8_loop_filter_simple_vertical_edge_mmx(y_ptr+8, y_stride, lfi->flim,lfi->lim,lfi->thr,2);
+;    vp8_loop_filter_simple_vertical_edge_mmx(y_ptr+12, y_stride, lfi->flim,lfi->lim,lfi->thr,2);
+;}
+
+SECTION_RODATA
+align 16
+tfe:
+    times 8 db 0xfe
+align 16
+t80:
+    times 8 db 0x80
+align 16
+t1s:
+    times 8 db 0x01
+align 16
+t3:
+    times 8 db 0x03
+align 16
+t4:
+    times 8 db 0x04
+align 16
+ones:
+    times 4 dw 0x0001
+align 16
+s27:
+    times 4 dw 0x1b00
+align 16
+s18:
+    times 4 dw 0x1200
+align 16
+s9:
+    times 4 dw 0x0900
+align 16
+s63:
+    times 4 dw 0x003f
diff --git a/libvpx/libvpx/vp8/decoder/dboolhuff.c b/libvpx/libvpx/vp8/decoder/dboolhuff.c
new file mode 100644
index 0000000..5cdd2a2
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/dboolhuff.c
@@ -0,0 +1,77 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "dboolhuff.h"
+#include "vp8/common/common.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+int vp8dx_start_decode(BOOL_DECODER *br,
+                       const unsigned char *source,
+                       unsigned int source_sz,
+                       vpx_decrypt_cb decrypt_cb,
+                       void *decrypt_state)
+{
+    br->user_buffer_end = source+source_sz;
+    br->user_buffer     = source;
+    br->value    = 0;
+    br->count    = -8;
+    br->range    = 255;
+    br->decrypt_cb = decrypt_cb;
+    br->decrypt_state = decrypt_state;
+
+    if (source_sz && !source)
+        return 1;
+
+    /* Populate the buffer */
+    vp8dx_bool_decoder_fill(br);
+
+    return 0;
+}
+
+void vp8dx_bool_decoder_fill(BOOL_DECODER *br)
+{
+    const unsigned char *bufptr = br->user_buffer;
+    VP8_BD_VALUE value = br->value;
+    int count = br->count;
+    int shift = VP8_BD_VALUE_SIZE - CHAR_BIT - (count + CHAR_BIT);
+    size_t bytes_left = br->user_buffer_end - bufptr;
+    size_t bits_left = bytes_left * CHAR_BIT;
+    int x = shift + CHAR_BIT - (int)bits_left;
+    int loop_end = 0;
+    unsigned char decrypted[sizeof(VP8_BD_VALUE) + 1];
+
+    if (br->decrypt_cb) {
+        size_t n = VPXMIN(sizeof(decrypted), bytes_left);
+        br->decrypt_cb(br->decrypt_state, bufptr, decrypted, (int)n);
+        bufptr = decrypted;
+    }
+
+    if(x >= 0)
+    {
+        count += VP8_LOTS_OF_BITS;
+        loop_end = x;
+    }
+
+    if (x < 0 || bits_left)
+    {
+        while(shift >= loop_end)
+        {
+            count += CHAR_BIT;
+            value |= (VP8_BD_VALUE)*bufptr << shift;
+            ++bufptr;
+            ++br->user_buffer;
+            shift -= CHAR_BIT;
+        }
+    }
+
+    br->value = value;
+    br->count = count;
+}
diff --git a/libvpx/libvpx/vp8/decoder/dboolhuff.h b/libvpx/libvpx/vp8/decoder/dboolhuff.h
new file mode 100644
index 0000000..1b1bbf8
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/dboolhuff.h
@@ -0,0 +1,141 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_DECODER_DBOOLHUFF_H_
+#define VP8_DECODER_DBOOLHUFF_H_
+
+#include <stddef.h>
+#include <limits.h>
+
+#include "./vpx_config.h"
+#include "vpx_ports/mem.h"
+#include "vpx/vp8dx.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef size_t VP8_BD_VALUE;
+
+#define VP8_BD_VALUE_SIZE ((int)sizeof(VP8_BD_VALUE)*CHAR_BIT)
+
+/*This is meant to be a large, positive constant that can still be efficiently
+   loaded as an immediate (on platforms like ARM, for example).
+  Even relatively modest values like 100 would work fine.*/
+#define VP8_LOTS_OF_BITS (0x40000000)
+
+typedef struct
+{
+    const unsigned char *user_buffer_end;
+    const unsigned char *user_buffer;
+    VP8_BD_VALUE         value;
+    int                  count;
+    unsigned int         range;
+    vpx_decrypt_cb       decrypt_cb;
+    void                *decrypt_state;
+} BOOL_DECODER;
+
+DECLARE_ALIGNED(16, extern const unsigned char, vp8_norm[256]);
+
+int vp8dx_start_decode(BOOL_DECODER *br,
+                       const unsigned char *source,
+                       unsigned int source_sz,
+                       vpx_decrypt_cb decrypt_cb,
+                       void *decrypt_state);
+
+void vp8dx_bool_decoder_fill(BOOL_DECODER *br);
+
+
+static int vp8dx_decode_bool(BOOL_DECODER *br, int probability) {
+    unsigned int bit = 0;
+    VP8_BD_VALUE value;
+    unsigned int split;
+    VP8_BD_VALUE bigsplit;
+    int count;
+    unsigned int range;
+
+    split = 1 + (((br->range - 1) * probability) >> 8);
+
+    if(br->count < 0)
+        vp8dx_bool_decoder_fill(br);
+
+    value = br->value;
+    count = br->count;
+
+    bigsplit = (VP8_BD_VALUE)split << (VP8_BD_VALUE_SIZE - 8);
+
+    range = split;
+
+    if (value >= bigsplit)
+    {
+        range = br->range - split;
+        value = value - bigsplit;
+        bit = 1;
+    }
+
+    {
+        register int shift = vp8_norm[range];
+        range <<= shift;
+        value <<= shift;
+        count -= shift;
+    }
+    br->value = value;
+    br->count = count;
+    br->range = range;
+
+    return bit;
+}
+
+static INLINE int vp8_decode_value(BOOL_DECODER *br, int bits)
+{
+    int z = 0;
+    int bit;
+
+    for (bit = bits - 1; bit >= 0; bit--)
+    {
+        z |= (vp8dx_decode_bool(br, 0x80) << bit);
+    }
+
+    return z;
+}
+
+static INLINE int vp8dx_bool_error(BOOL_DECODER *br)
+{
+    /* Check if we have reached the end of the buffer.
+     *
+     * Variable 'count' stores the number of bits in the 'value' buffer, minus
+     * 8. The top byte is part of the algorithm, and the remainder is buffered
+     * to be shifted into it. So if count == 8, the top 16 bits of 'value' are
+     * occupied, 8 for the algorithm and 8 in the buffer.
+     *
+     * When reading a byte from the user's buffer, count is filled with 8 and
+     * one byte is filled into the value buffer. When we reach the end of the
+     * data, count is additionally filled with VP8_LOTS_OF_BITS. So when
+     * count == VP8_LOTS_OF_BITS - 1, the user's data has been exhausted.
+     */
+    if ((br->count > VP8_BD_VALUE_SIZE) && (br->count < VP8_LOTS_OF_BITS))
+    {
+       /* We have tried to decode bits after the end of
+        * stream was encountered.
+        */
+        return 1;
+    }
+
+    /* No error. */
+    return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_DECODER_DBOOLHUFF_H_
diff --git a/libvpx/libvpx/vp8/decoder/decodeframe.c b/libvpx/libvpx/vp8/decoder/decodeframe.c
new file mode 100644
index 0000000..51acdbb
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/decodeframe.c
@@ -0,0 +1,1397 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+#include "onyxd_int.h"
+#include "vp8/common/header.h"
+#include "vp8/common/reconintra4x4.h"
+#include "vp8/common/reconinter.h"
+#include "detokenize.h"
+#include "vp8/common/common.h"
+#include "vp8/common/invtrans.h"
+#include "vp8/common/alloccommon.h"
+#include "vp8/common/entropymode.h"
+#include "vp8/common/quant_common.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vp8/common/reconintra.h"
+#include "vp8/common/setupintrarecon.h"
+
+#include "decodemv.h"
+#include "vp8/common/extend.h"
+#if CONFIG_ERROR_CONCEALMENT
+#include "error_concealment.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+#include "vp8/common/threading.h"
+#include "decoderthreading.h"
+#include "dboolhuff.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+#include <assert.h>
+#include <stdio.h>
+
+void vp8cx_init_de_quantizer(VP8D_COMP *pbi)
+{
+    int Q;
+    VP8_COMMON *const pc = & pbi->common;
+
+    for (Q = 0; Q < QINDEX_RANGE; Q++)
+    {
+        pc->Y1dequant[Q][0] = (short)vp8_dc_quant(Q, pc->y1dc_delta_q);
+        pc->Y2dequant[Q][0] = (short)vp8_dc2quant(Q, pc->y2dc_delta_q);
+        pc->UVdequant[Q][0] = (short)vp8_dc_uv_quant(Q, pc->uvdc_delta_q);
+
+        pc->Y1dequant[Q][1] = (short)vp8_ac_yquant(Q);
+        pc->Y2dequant[Q][1] = (short)vp8_ac2quant(Q, pc->y2ac_delta_q);
+        pc->UVdequant[Q][1] = (short)vp8_ac_uv_quant(Q, pc->uvac_delta_q);
+    }
+}
+
+void vp8_mb_init_dequantizer(VP8D_COMP *pbi, MACROBLOCKD *xd)
+{
+    int i;
+    int QIndex;
+    MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi;
+    VP8_COMMON *const pc = & pbi->common;
+
+    /* Decide whether to use the default or alternate baseline Q value. */
+    if (xd->segmentation_enabled)
+    {
+        /* Abs Value */
+        if (xd->mb_segement_abs_delta == SEGMENT_ABSDATA)
+            QIndex = xd->segment_feature_data[MB_LVL_ALT_Q][mbmi->segment_id];
+
+        /* Delta Value */
+        else
+            QIndex = pc->base_qindex + xd->segment_feature_data[MB_LVL_ALT_Q][mbmi->segment_id];
+
+        QIndex = (QIndex >= 0) ? ((QIndex <= MAXQ) ? QIndex : MAXQ) : 0;    /* Clamp to valid range */
+    }
+    else
+        QIndex = pc->base_qindex;
+
+    /* Set up the macroblock dequant constants */
+    xd->dequant_y1_dc[0] = 1;
+    xd->dequant_y1[0] = pc->Y1dequant[QIndex][0];
+    xd->dequant_y2[0] = pc->Y2dequant[QIndex][0];
+    xd->dequant_uv[0] = pc->UVdequant[QIndex][0];
+
+    for (i = 1; i < 16; i++)
+    {
+        xd->dequant_y1_dc[i] =
+        xd->dequant_y1[i] = pc->Y1dequant[QIndex][1];
+        xd->dequant_y2[i] = pc->Y2dequant[QIndex][1];
+        xd->dequant_uv[i] = pc->UVdequant[QIndex][1];
+    }
+}
+
+static void decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd,
+                              unsigned int mb_idx)
+{
+    MB_PREDICTION_MODE mode;
+    int i;
+#if CONFIG_ERROR_CONCEALMENT
+    int corruption_detected = 0;
+#else
+    (void)mb_idx;
+#endif
+
+    if (xd->mode_info_context->mbmi.mb_skip_coeff)
+    {
+        vp8_reset_mb_tokens_context(xd);
+    }
+    else if (!vp8dx_bool_error(xd->current_bc))
+    {
+        int eobtotal;
+        eobtotal = vp8_decode_mb_tokens(pbi, xd);
+
+        /* Special case:  Force the loopfilter to skip when eobtotal is zero */
+        xd->mode_info_context->mbmi.mb_skip_coeff = (eobtotal==0);
+    }
+
+    mode = xd->mode_info_context->mbmi.mode;
+
+    if (xd->segmentation_enabled)
+        vp8_mb_init_dequantizer(pbi, xd);
+
+
+#if CONFIG_ERROR_CONCEALMENT
+
+    if(pbi->ec_active)
+    {
+        int throw_residual;
+        /* When we have independent partitions we can apply residual even
+         * though other partitions within the frame are corrupt.
+         */
+        throw_residual = (!pbi->independent_partitions &&
+                          pbi->frame_corrupt_residual);
+        throw_residual = (throw_residual || vp8dx_bool_error(xd->current_bc));
+
+        if ((mb_idx >= pbi->mvs_corrupt_from_mb || throw_residual))
+        {
+            /* MB with corrupt residuals or corrupt mode/motion vectors.
+             * Better to use the predictor as reconstruction.
+             */
+            pbi->frame_corrupt_residual = 1;
+            memset(xd->qcoeff, 0, sizeof(xd->qcoeff));
+
+            corruption_detected = 1;
+
+            /* force idct to be skipped for B_PRED and use the
+             * prediction only for reconstruction
+             * */
+            memset(xd->eobs, 0, 25);
+        }
+    }
+#endif
+
+    /* do prediction */
+    if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
+    {
+        vp8_build_intra_predictors_mbuv_s(xd,
+                                          xd->recon_above[1],
+                                          xd->recon_above[2],
+                                          xd->recon_left[1],
+                                          xd->recon_left[2],
+                                          xd->recon_left_stride[1],
+                                          xd->dst.u_buffer, xd->dst.v_buffer,
+                                          xd->dst.uv_stride);
+
+        if (mode != B_PRED)
+        {
+            vp8_build_intra_predictors_mby_s(xd,
+                                                 xd->recon_above[0],
+                                                 xd->recon_left[0],
+                                                 xd->recon_left_stride[0],
+                                                 xd->dst.y_buffer,
+                                                 xd->dst.y_stride);
+        }
+        else
+        {
+            short *DQC = xd->dequant_y1;
+            int dst_stride = xd->dst.y_stride;
+
+            /* clear out residual eob info */
+            if(xd->mode_info_context->mbmi.mb_skip_coeff)
+                memset(xd->eobs, 0, 25);
+
+            intra_prediction_down_copy(xd, xd->recon_above[0] + 16);
+
+            for (i = 0; i < 16; i++)
+            {
+                BLOCKD *b = &xd->block[i];
+                unsigned char *dst = xd->dst.y_buffer + b->offset;
+                B_PREDICTION_MODE b_mode =
+                    xd->mode_info_context->bmi[i].as_mode;
+                unsigned char *Above = dst - dst_stride;
+                unsigned char *yleft = dst - 1;
+                int left_stride = dst_stride;
+                unsigned char top_left = Above[-1];
+
+                vp8_intra4x4_predict(Above, yleft, left_stride, b_mode,
+                                     dst, dst_stride, top_left);
+
+                if (xd->eobs[i])
+                {
+                    if (xd->eobs[i] > 1)
+                    {
+                    vp8_dequant_idct_add(b->qcoeff, DQC, dst, dst_stride);
+                    }
+                    else
+                    {
+                        vp8_dc_only_idct_add
+                            (b->qcoeff[0] * DQC[0],
+                                dst, dst_stride,
+                                dst, dst_stride);
+                        memset(b->qcoeff, 0, 2 * sizeof(b->qcoeff[0]));
+                    }
+                }
+            }
+        }
+    }
+    else
+    {
+        vp8_build_inter_predictors_mb(xd);
+    }
+
+
+#if CONFIG_ERROR_CONCEALMENT
+    if (corruption_detected)
+    {
+        return;
+    }
+#endif
+
+    if(!xd->mode_info_context->mbmi.mb_skip_coeff)
+    {
+        /* dequantization and idct */
+        if (mode != B_PRED)
+        {
+            short *DQC = xd->dequant_y1;
+
+            if (mode != SPLITMV)
+            {
+                BLOCKD *b = &xd->block[24];
+
+                /* do 2nd order transform on the dc block */
+                if (xd->eobs[24] > 1)
+                {
+                    vp8_dequantize_b(b, xd->dequant_y2);
+
+                    vp8_short_inv_walsh4x4(&b->dqcoeff[0],
+                        xd->qcoeff);
+                    memset(b->qcoeff, 0, 16 * sizeof(b->qcoeff[0]));
+                }
+                else
+                {
+                    b->dqcoeff[0] = b->qcoeff[0] * xd->dequant_y2[0];
+                    vp8_short_inv_walsh4x4_1(&b->dqcoeff[0],
+                        xd->qcoeff);
+                    memset(b->qcoeff, 0, 2 * sizeof(b->qcoeff[0]));
+                }
+
+                /* override the dc dequant constant in order to preserve the
+                 * dc components
+                 */
+                DQC = xd->dequant_y1_dc;
+            }
+
+            vp8_dequant_idct_add_y_block
+                            (xd->qcoeff, DQC,
+                             xd->dst.y_buffer,
+                             xd->dst.y_stride, xd->eobs);
+        }
+
+        vp8_dequant_idct_add_uv_block
+                        (xd->qcoeff+16*16, xd->dequant_uv,
+                         xd->dst.u_buffer, xd->dst.v_buffer,
+                         xd->dst.uv_stride, xd->eobs+16);
+    }
+}
+
+static int get_delta_q(vp8_reader *bc, int prev, int *q_update)
+{
+    int ret_val = 0;
+
+    if (vp8_read_bit(bc))
+    {
+        ret_val = vp8_read_literal(bc, 4);
+
+        if (vp8_read_bit(bc))
+            ret_val = -ret_val;
+    }
+
+    /* Trigger a quantizer update if the delta-q value has changed */
+    if (ret_val != prev)
+        *q_update = 1;
+
+    return ret_val;
+}
+
+#ifdef PACKET_TESTING
+#include <stdio.h>
+FILE *vpxlog = 0;
+#endif
+
+static void yv12_extend_frame_top_c(YV12_BUFFER_CONFIG *ybf)
+{
+    int i;
+    unsigned char *src_ptr1;
+    unsigned char *dest_ptr1;
+
+    unsigned int Border;
+    int plane_stride;
+
+    /***********/
+    /* Y Plane */
+    /***********/
+    Border = ybf->border;
+    plane_stride = ybf->y_stride;
+    src_ptr1 = ybf->y_buffer - Border;
+    dest_ptr1 = src_ptr1 - (Border * plane_stride);
+
+    for (i = 0; i < (int)Border; i++)
+    {
+        memcpy(dest_ptr1, src_ptr1, plane_stride);
+        dest_ptr1 += plane_stride;
+    }
+
+
+    /***********/
+    /* U Plane */
+    /***********/
+    plane_stride = ybf->uv_stride;
+    Border /= 2;
+    src_ptr1 = ybf->u_buffer - Border;
+    dest_ptr1 = src_ptr1 - (Border * plane_stride);
+
+    for (i = 0; i < (int)(Border); i++)
+    {
+        memcpy(dest_ptr1, src_ptr1, plane_stride);
+        dest_ptr1 += plane_stride;
+    }
+
+    /***********/
+    /* V Plane */
+    /***********/
+
+    src_ptr1 = ybf->v_buffer - Border;
+    dest_ptr1 = src_ptr1 - (Border * plane_stride);
+
+    for (i = 0; i < (int)(Border); i++)
+    {
+        memcpy(dest_ptr1, src_ptr1, plane_stride);
+        dest_ptr1 += plane_stride;
+    }
+}
+
+static void yv12_extend_frame_bottom_c(YV12_BUFFER_CONFIG *ybf)
+{
+    int i;
+    unsigned char *src_ptr1, *src_ptr2;
+    unsigned char *dest_ptr2;
+
+    unsigned int Border;
+    int plane_stride;
+    int plane_height;
+
+    /***********/
+    /* Y Plane */
+    /***********/
+    Border = ybf->border;
+    plane_stride = ybf->y_stride;
+    plane_height = ybf->y_height;
+
+    src_ptr1 = ybf->y_buffer - Border;
+    src_ptr2 = src_ptr1 + (plane_height * plane_stride) - plane_stride;
+    dest_ptr2 = src_ptr2 + plane_stride;
+
+    for (i = 0; i < (int)Border; i++)
+    {
+        memcpy(dest_ptr2, src_ptr2, plane_stride);
+        dest_ptr2 += plane_stride;
+    }
+
+
+    /***********/
+    /* U Plane */
+    /***********/
+    plane_stride = ybf->uv_stride;
+    plane_height = ybf->uv_height;
+    Border /= 2;
+
+    src_ptr1 = ybf->u_buffer - Border;
+    src_ptr2 = src_ptr1 + (plane_height * plane_stride) - plane_stride;
+    dest_ptr2 = src_ptr2 + plane_stride;
+
+    for (i = 0; i < (int)(Border); i++)
+    {
+        memcpy(dest_ptr2, src_ptr2, plane_stride);
+        dest_ptr2 += plane_stride;
+    }
+
+    /***********/
+    /* V Plane */
+    /***********/
+
+    src_ptr1 = ybf->v_buffer - Border;
+    src_ptr2 = src_ptr1 + (plane_height * plane_stride) - plane_stride;
+    dest_ptr2 = src_ptr2 + plane_stride;
+
+    for (i = 0; i < (int)(Border); i++)
+    {
+        memcpy(dest_ptr2, src_ptr2, plane_stride);
+        dest_ptr2 += plane_stride;
+    }
+}
+
+static void yv12_extend_frame_left_right_c(YV12_BUFFER_CONFIG *ybf,
+                                           unsigned char *y_src,
+                                           unsigned char *u_src,
+                                           unsigned char *v_src)
+{
+    int i;
+    unsigned char *src_ptr1, *src_ptr2;
+    unsigned char *dest_ptr1, *dest_ptr2;
+
+    unsigned int Border;
+    int plane_stride;
+    int plane_height;
+    int plane_width;
+
+    /***********/
+    /* Y Plane */
+    /***********/
+    Border = ybf->border;
+    plane_stride = ybf->y_stride;
+    plane_height = 16;
+    plane_width = ybf->y_width;
+
+    /* copy the left and right most columns out */
+    src_ptr1 = y_src;
+    src_ptr2 = src_ptr1 + plane_width - 1;
+    dest_ptr1 = src_ptr1 - Border;
+    dest_ptr2 = src_ptr2 + 1;
+
+    for (i = 0; i < plane_height; i++)
+    {
+        memset(dest_ptr1, src_ptr1[0], Border);
+        memset(dest_ptr2, src_ptr2[0], Border);
+        src_ptr1  += plane_stride;
+        src_ptr2  += plane_stride;
+        dest_ptr1 += plane_stride;
+        dest_ptr2 += plane_stride;
+    }
+
+    /***********/
+    /* U Plane */
+    /***********/
+    plane_stride = ybf->uv_stride;
+    plane_height = 8;
+    plane_width = ybf->uv_width;
+    Border /= 2;
+
+    /* copy the left and right most columns out */
+    src_ptr1 = u_src;
+    src_ptr2 = src_ptr1 + plane_width - 1;
+    dest_ptr1 = src_ptr1 - Border;
+    dest_ptr2 = src_ptr2 + 1;
+
+    for (i = 0; i < plane_height; i++)
+    {
+        memset(dest_ptr1, src_ptr1[0], Border);
+        memset(dest_ptr2, src_ptr2[0], Border);
+        src_ptr1  += plane_stride;
+        src_ptr2  += plane_stride;
+        dest_ptr1 += plane_stride;
+        dest_ptr2 += plane_stride;
+    }
+
+    /***********/
+    /* V Plane */
+    /***********/
+
+    /* copy the left and right most columns out */
+    src_ptr1 = v_src;
+    src_ptr2 = src_ptr1 + plane_width - 1;
+    dest_ptr1 = src_ptr1 - Border;
+    dest_ptr2 = src_ptr2 + 1;
+
+    for (i = 0; i < plane_height; i++)
+    {
+        memset(dest_ptr1, src_ptr1[0], Border);
+        memset(dest_ptr2, src_ptr2[0], Border);
+        src_ptr1  += plane_stride;
+        src_ptr2  += plane_stride;
+        dest_ptr1 += plane_stride;
+        dest_ptr2 += plane_stride;
+    }
+}
+
+static void decode_mb_rows(VP8D_COMP *pbi)
+{
+    VP8_COMMON *const pc = & pbi->common;
+    MACROBLOCKD *const xd  = & pbi->mb;
+
+    MODE_INFO *lf_mic = xd->mode_info_context;
+
+    int ibc = 0;
+    int num_part = 1 << pc->multi_token_partition;
+
+    int recon_yoffset, recon_uvoffset;
+    int mb_row, mb_col;
+    int mb_idx = 0;
+
+    YV12_BUFFER_CONFIG *yv12_fb_new = pbi->dec_fb_ref[INTRA_FRAME];
+
+    int recon_y_stride = yv12_fb_new->y_stride;
+    int recon_uv_stride = yv12_fb_new->uv_stride;
+
+    unsigned char *ref_buffer[MAX_REF_FRAMES][3];
+    unsigned char *dst_buffer[3];
+    unsigned char *lf_dst[3];
+    unsigned char *eb_dst[3];
+    int i;
+    int ref_fb_corrupted[MAX_REF_FRAMES];
+
+    ref_fb_corrupted[INTRA_FRAME] = 0;
+
+    for(i = 1; i < MAX_REF_FRAMES; i++)
+    {
+        YV12_BUFFER_CONFIG *this_fb = pbi->dec_fb_ref[i];
+
+        ref_buffer[i][0] = this_fb->y_buffer;
+        ref_buffer[i][1] = this_fb->u_buffer;
+        ref_buffer[i][2] = this_fb->v_buffer;
+
+        ref_fb_corrupted[i] = this_fb->corrupted;
+    }
+
+    /* Set up the buffer pointers */
+    eb_dst[0] = lf_dst[0] = dst_buffer[0] = yv12_fb_new->y_buffer;
+    eb_dst[1] = lf_dst[1] = dst_buffer[1] = yv12_fb_new->u_buffer;
+    eb_dst[2] = lf_dst[2] = dst_buffer[2] = yv12_fb_new->v_buffer;
+
+    xd->up_available = 0;
+
+    /* Initialize the loop filter for this frame. */
+    if(pc->filter_level)
+        vp8_loop_filter_frame_init(pc, xd, pc->filter_level);
+
+    vp8_setup_intra_recon_top_line(yv12_fb_new);
+
+    /* Decode the individual macro block */
+    for (mb_row = 0; mb_row < pc->mb_rows; mb_row++)
+    {
+        if (num_part > 1)
+        {
+            xd->current_bc = & pbi->mbc[ibc];
+            ibc++;
+
+            if (ibc == num_part)
+                ibc = 0;
+        }
+
+        recon_yoffset = mb_row * recon_y_stride * 16;
+        recon_uvoffset = mb_row * recon_uv_stride * 8;
+
+        /* reset contexts */
+        xd->above_context = pc->above_context;
+        memset(xd->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES));
+
+        xd->left_available = 0;
+
+        xd->mb_to_top_edge = -((mb_row * 16) << 3);
+        xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
+
+        xd->recon_above[0] = dst_buffer[0] + recon_yoffset;
+        xd->recon_above[1] = dst_buffer[1] + recon_uvoffset;
+        xd->recon_above[2] = dst_buffer[2] + recon_uvoffset;
+
+        xd->recon_left[0] = xd->recon_above[0] - 1;
+        xd->recon_left[1] = xd->recon_above[1] - 1;
+        xd->recon_left[2] = xd->recon_above[2] - 1;
+
+        xd->recon_above[0] -= xd->dst.y_stride;
+        xd->recon_above[1] -= xd->dst.uv_stride;
+        xd->recon_above[2] -= xd->dst.uv_stride;
+
+        /* TODO: move to outside row loop */
+        xd->recon_left_stride[0] = xd->dst.y_stride;
+        xd->recon_left_stride[1] = xd->dst.uv_stride;
+
+        setup_intra_recon_left(xd->recon_left[0], xd->recon_left[1],
+                               xd->recon_left[2], xd->dst.y_stride,
+                               xd->dst.uv_stride);
+
+        for (mb_col = 0; mb_col < pc->mb_cols; mb_col++)
+        {
+            /* Distance of Mb to the various image edges.
+             * These are specified to 8th pel as they are always compared to values
+             * that are in 1/8th pel units
+             */
+            xd->mb_to_left_edge = -((mb_col * 16) << 3);
+            xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
+
+#if CONFIG_ERROR_CONCEALMENT
+            {
+                int corrupt_residual = (!pbi->independent_partitions &&
+                                       pbi->frame_corrupt_residual) ||
+                                       vp8dx_bool_error(xd->current_bc);
+                if (pbi->ec_active &&
+                    xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME &&
+                    corrupt_residual)
+                {
+                    /* We have an intra block with corrupt coefficients, better to
+                     * conceal with an inter block. Interpolate MVs from neighboring
+                     * MBs.
+                     *
+                     * Note that for the first mb with corrupt residual in a frame,
+                     * we might not discover that before decoding the residual. That
+                     * happens after this check, and therefore no inter concealment
+                     * will be done.
+                     */
+                    vp8_interpolate_motion(xd,
+                                           mb_row, mb_col,
+                                           pc->mb_rows, pc->mb_cols);
+                }
+            }
+#endif
+
+            xd->dst.y_buffer = dst_buffer[0] + recon_yoffset;
+            xd->dst.u_buffer = dst_buffer[1] + recon_uvoffset;
+            xd->dst.v_buffer = dst_buffer[2] + recon_uvoffset;
+
+            if (xd->mode_info_context->mbmi.ref_frame >= LAST_FRAME) {
+              const MV_REFERENCE_FRAME ref = xd->mode_info_context->mbmi.ref_frame;
+              xd->pre.y_buffer = ref_buffer[ref][0] + recon_yoffset;
+              xd->pre.u_buffer = ref_buffer[ref][1] + recon_uvoffset;
+              xd->pre.v_buffer = ref_buffer[ref][2] + recon_uvoffset;
+            } else {
+              // ref_frame is INTRA_FRAME, pre buffer should not be used.
+              xd->pre.y_buffer = 0;
+              xd->pre.u_buffer = 0;
+              xd->pre.v_buffer = 0;
+            }
+
+            /* propagate errors from reference frames */
+            xd->corrupted |= ref_fb_corrupted[xd->mode_info_context->mbmi.ref_frame];
+
+            decode_macroblock(pbi, xd, mb_idx);
+
+            mb_idx++;
+            xd->left_available = 1;
+
+            /* check if the boolean decoder has suffered an error */
+            xd->corrupted |= vp8dx_bool_error(xd->current_bc);
+
+            xd->recon_above[0] += 16;
+            xd->recon_above[1] += 8;
+            xd->recon_above[2] += 8;
+            xd->recon_left[0] += 16;
+            xd->recon_left[1] += 8;
+            xd->recon_left[2] += 8;
+
+            recon_yoffset += 16;
+            recon_uvoffset += 8;
+
+            ++xd->mode_info_context;  /* next mb */
+
+            xd->above_context++;
+        }
+
+        /* adjust to the next row of mbs */
+        vp8_extend_mb_row(yv12_fb_new, xd->dst.y_buffer + 16,
+                          xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
+
+        ++xd->mode_info_context;      /* skip prediction column */
+        xd->up_available = 1;
+
+        if(pc->filter_level)
+        {
+            if(mb_row > 0)
+            {
+                if (pc->filter_type == NORMAL_LOOPFILTER)
+                    vp8_loop_filter_row_normal(pc, lf_mic, mb_row-1,
+                                               recon_y_stride, recon_uv_stride,
+                                               lf_dst[0], lf_dst[1], lf_dst[2]);
+                else
+                    vp8_loop_filter_row_simple(pc, lf_mic, mb_row-1,
+                                               recon_y_stride, recon_uv_stride,
+                                               lf_dst[0], lf_dst[1], lf_dst[2]);
+                if(mb_row > 1)
+                {
+                    yv12_extend_frame_left_right_c(yv12_fb_new,
+                                                   eb_dst[0],
+                                                   eb_dst[1],
+                                                   eb_dst[2]);
+
+                    eb_dst[0] += recon_y_stride  * 16;
+                    eb_dst[1] += recon_uv_stride *  8;
+                    eb_dst[2] += recon_uv_stride *  8;
+                }
+
+                lf_dst[0] += recon_y_stride  * 16;
+                lf_dst[1] += recon_uv_stride *  8;
+                lf_dst[2] += recon_uv_stride *  8;
+                lf_mic += pc->mb_cols;
+                lf_mic++;         /* Skip border mb */
+            }
+        }
+        else
+        {
+            if(mb_row > 0)
+            {
+                /**/
+                yv12_extend_frame_left_right_c(yv12_fb_new,
+                                               eb_dst[0],
+                                               eb_dst[1],
+                                               eb_dst[2]);
+                eb_dst[0] += recon_y_stride  * 16;
+                eb_dst[1] += recon_uv_stride *  8;
+                eb_dst[2] += recon_uv_stride *  8;
+            }
+        }
+    }
+
+    if(pc->filter_level)
+    {
+        if (pc->filter_type == NORMAL_LOOPFILTER)
+            vp8_loop_filter_row_normal(pc, lf_mic, mb_row-1, recon_y_stride,
+                                       recon_uv_stride, lf_dst[0], lf_dst[1],
+                                       lf_dst[2]);
+        else
+            vp8_loop_filter_row_simple(pc, lf_mic, mb_row-1, recon_y_stride,
+                                       recon_uv_stride, lf_dst[0], lf_dst[1],
+                                       lf_dst[2]);
+
+        yv12_extend_frame_left_right_c(yv12_fb_new,
+                                       eb_dst[0],
+                                       eb_dst[1],
+                                       eb_dst[2]);
+        eb_dst[0] += recon_y_stride  * 16;
+        eb_dst[1] += recon_uv_stride *  8;
+        eb_dst[2] += recon_uv_stride *  8;
+    }
+    yv12_extend_frame_left_right_c(yv12_fb_new,
+                                   eb_dst[0],
+                                   eb_dst[1],
+                                   eb_dst[2]);
+    yv12_extend_frame_top_c(yv12_fb_new);
+    yv12_extend_frame_bottom_c(yv12_fb_new);
+
+}
+
+static unsigned int read_partition_size(VP8D_COMP *pbi,
+                                        const unsigned char *cx_size)
+{
+    unsigned char temp[3];
+    if (pbi->decrypt_cb)
+    {
+        pbi->decrypt_cb(pbi->decrypt_state, cx_size, temp, 3);
+        cx_size = temp;
+    }
+    return cx_size[0] + (cx_size[1] << 8) + (cx_size[2] << 16);
+}
+
+static int read_is_valid(const unsigned char *start,
+                         size_t               len,
+                         const unsigned char *end)
+{
+    return (start + len > start && start + len <= end);
+}
+
+static unsigned int read_available_partition_size(
+                                       VP8D_COMP *pbi,
+                                       const unsigned char *token_part_sizes,
+                                       const unsigned char *fragment_start,
+                                       const unsigned char *first_fragment_end,
+                                       const unsigned char *fragment_end,
+                                       int i,
+                                       int num_part)
+{
+    VP8_COMMON* pc = &pbi->common;
+    const unsigned char *partition_size_ptr = token_part_sizes + i * 3;
+    unsigned int partition_size = 0;
+    ptrdiff_t bytes_left = fragment_end - fragment_start;
+    /* Calculate the length of this partition. The last partition
+     * size is implicit. If the partition size can't be read, then
+     * either use the remaining data in the buffer (for EC mode)
+     * or throw an error.
+     */
+    if (i < num_part - 1)
+    {
+        if (read_is_valid(partition_size_ptr, 3, first_fragment_end))
+            partition_size = read_partition_size(pbi, partition_size_ptr);
+        else if (pbi->ec_active)
+            partition_size = (unsigned int)bytes_left;
+        else
+            vpx_internal_error(&pc->error, VPX_CODEC_CORRUPT_FRAME,
+                               "Truncated partition size data");
+    }
+    else
+        partition_size = (unsigned int)bytes_left;
+
+    /* Validate the calculated partition length. If the buffer
+     * described by the partition can't be fully read, then restrict
+     * it to the portion that can be (for EC mode) or throw an error.
+     */
+    if (!read_is_valid(fragment_start, partition_size, fragment_end))
+    {
+        if (pbi->ec_active)
+            partition_size = (unsigned int)bytes_left;
+        else
+            vpx_internal_error(&pc->error, VPX_CODEC_CORRUPT_FRAME,
+                               "Truncated packet or corrupt partition "
+                               "%d length", i + 1);
+    }
+    return partition_size;
+}
+
+
+static void setup_token_decoder(VP8D_COMP *pbi,
+                                const unsigned char* token_part_sizes)
+{
+    vp8_reader *bool_decoder = &pbi->mbc[0];
+    unsigned int partition_idx;
+    unsigned int fragment_idx;
+    unsigned int num_token_partitions;
+    const unsigned char *first_fragment_end = pbi->fragments.ptrs[0] +
+                                          pbi->fragments.sizes[0];
+
+    TOKEN_PARTITION multi_token_partition =
+            (TOKEN_PARTITION)vp8_read_literal(&pbi->mbc[8], 2);
+    if (!vp8dx_bool_error(&pbi->mbc[8]))
+        pbi->common.multi_token_partition = multi_token_partition;
+    num_token_partitions = 1 << pbi->common.multi_token_partition;
+
+    /* Check for partitions within the fragments and unpack the fragments
+     * so that each fragment pointer points to its corresponding partition. */
+    for (fragment_idx = 0; fragment_idx < pbi->fragments.count; ++fragment_idx)
+    {
+        unsigned int fragment_size = pbi->fragments.sizes[fragment_idx];
+        const unsigned char *fragment_end = pbi->fragments.ptrs[fragment_idx] +
+                                            fragment_size;
+        /* Special case for handling the first partition since we have already
+         * read its size. */
+        if (fragment_idx == 0)
+        {
+            /* Size of first partition + token partition sizes element */
+            ptrdiff_t ext_first_part_size = token_part_sizes -
+                pbi->fragments.ptrs[0] + 3 * (num_token_partitions - 1);
+            fragment_size -= (unsigned int)ext_first_part_size;
+            if (fragment_size > 0)
+            {
+                pbi->fragments.sizes[0] = (unsigned int)ext_first_part_size;
+                /* The fragment contains an additional partition. Move to
+                 * next. */
+                fragment_idx++;
+                pbi->fragments.ptrs[fragment_idx] = pbi->fragments.ptrs[0] +
+                  pbi->fragments.sizes[0];
+            }
+        }
+        /* Split the chunk into partitions read from the bitstream */
+        while (fragment_size > 0)
+        {
+            ptrdiff_t partition_size = read_available_partition_size(
+                                                 pbi,
+                                                 token_part_sizes,
+                                                 pbi->fragments.ptrs[fragment_idx],
+                                                 first_fragment_end,
+                                                 fragment_end,
+                                                 fragment_idx - 1,
+                                                 num_token_partitions);
+            pbi->fragments.sizes[fragment_idx] = (unsigned int)partition_size;
+            fragment_size -= (unsigned int)partition_size;
+            assert(fragment_idx <= num_token_partitions);
+            if (fragment_size > 0)
+            {
+                /* The fragment contains an additional partition.
+                 * Move to next. */
+                fragment_idx++;
+                pbi->fragments.ptrs[fragment_idx] =
+                    pbi->fragments.ptrs[fragment_idx - 1] + partition_size;
+            }
+        }
+    }
+
+    pbi->fragments.count = num_token_partitions + 1;
+
+    for (partition_idx = 1; partition_idx < pbi->fragments.count; ++partition_idx)
+    {
+        if (vp8dx_start_decode(bool_decoder,
+                               pbi->fragments.ptrs[partition_idx],
+                               pbi->fragments.sizes[partition_idx],
+                               pbi->decrypt_cb, pbi->decrypt_state))
+            vpx_internal_error(&pbi->common.error, VPX_CODEC_MEM_ERROR,
+                               "Failed to allocate bool decoder %d",
+                               partition_idx);
+
+        bool_decoder++;
+    }
+
+#if CONFIG_MULTITHREAD
+    /* Clamp number of decoder threads */
+    if (pbi->decoding_thread_count > num_token_partitions - 1)
+        pbi->decoding_thread_count = num_token_partitions - 1;
+#endif
+}
+
+
+static void init_frame(VP8D_COMP *pbi)
+{
+    VP8_COMMON *const pc = & pbi->common;
+    MACROBLOCKD *const xd  = & pbi->mb;
+
+    if (pc->frame_type == KEY_FRAME)
+    {
+        /* Various keyframe initializations */
+        memcpy(pc->fc.mvc, vp8_default_mv_context, sizeof(vp8_default_mv_context));
+
+        vp8_init_mbmode_probs(pc);
+
+        vp8_default_coef_probs(pc);
+
+        /* reset the segment feature data to 0 with delta coding (Default state). */
+        memset(xd->segment_feature_data, 0, sizeof(xd->segment_feature_data));
+        xd->mb_segement_abs_delta = SEGMENT_DELTADATA;
+
+        /* reset the mode ref deltasa for loop filter */
+        memset(xd->ref_lf_deltas, 0, sizeof(xd->ref_lf_deltas));
+        memset(xd->mode_lf_deltas, 0, sizeof(xd->mode_lf_deltas));
+
+        /* All buffers are implicitly updated on key frames. */
+        pc->refresh_golden_frame = 1;
+        pc->refresh_alt_ref_frame = 1;
+        pc->copy_buffer_to_gf = 0;
+        pc->copy_buffer_to_arf = 0;
+
+        /* Note that Golden and Altref modes cannot be used on a key frame so
+         * ref_frame_sign_bias[] is undefined and meaningless
+         */
+        pc->ref_frame_sign_bias[GOLDEN_FRAME] = 0;
+        pc->ref_frame_sign_bias[ALTREF_FRAME] = 0;
+    }
+    else
+    {
+        /* To enable choice of different interploation filters */
+        if (!pc->use_bilinear_mc_filter)
+        {
+            xd->subpixel_predict        = vp8_sixtap_predict4x4;
+            xd->subpixel_predict8x4     = vp8_sixtap_predict8x4;
+            xd->subpixel_predict8x8     = vp8_sixtap_predict8x8;
+            xd->subpixel_predict16x16   = vp8_sixtap_predict16x16;
+        }
+        else
+        {
+            xd->subpixel_predict        = vp8_bilinear_predict4x4;
+            xd->subpixel_predict8x4     = vp8_bilinear_predict8x4;
+            xd->subpixel_predict8x8     = vp8_bilinear_predict8x8;
+            xd->subpixel_predict16x16   = vp8_bilinear_predict16x16;
+        }
+
+        if (pbi->decoded_key_frame && pbi->ec_enabled && !pbi->ec_active)
+            pbi->ec_active = 1;
+    }
+
+    xd->left_context = &pc->left_context;
+    xd->mode_info_context = pc->mi;
+    xd->frame_type = pc->frame_type;
+    xd->mode_info_context->mbmi.mode = DC_PRED;
+    xd->mode_info_stride = pc->mode_info_stride;
+    xd->corrupted = 0; /* init without corruption */
+
+    xd->fullpixel_mask = 0xffffffff;
+    if(pc->full_pixel)
+        xd->fullpixel_mask = 0xfffffff8;
+
+}
+
+int vp8_decode_frame(VP8D_COMP *pbi)
+{
+    vp8_reader *const bc = &pbi->mbc[8];
+    VP8_COMMON *const pc = &pbi->common;
+    MACROBLOCKD *const xd  = &pbi->mb;
+    const unsigned char *data = pbi->fragments.ptrs[0];
+    const unsigned int data_sz = pbi->fragments.sizes[0];
+    const unsigned char *data_end = data + data_sz;
+    ptrdiff_t first_partition_length_in_bytes;
+
+    int i, j, k, l;
+    const int *const mb_feature_data_bits = vp8_mb_feature_data_bits;
+    int corrupt_tokens = 0;
+    int prev_independent_partitions = pbi->independent_partitions;
+
+    YV12_BUFFER_CONFIG *yv12_fb_new = pbi->dec_fb_ref[INTRA_FRAME];
+
+    /* start with no corruption of current frame */
+    xd->corrupted = 0;
+    yv12_fb_new->corrupted = 0;
+
+    if (data_end - data < 3)
+    {
+        if (!pbi->ec_active)
+        {
+            vpx_internal_error(&pc->error, VPX_CODEC_CORRUPT_FRAME,
+                               "Truncated packet");
+        }
+
+        /* Declare the missing frame as an inter frame since it will
+           be handled as an inter frame when we have estimated its
+           motion vectors. */
+        pc->frame_type = INTER_FRAME;
+        pc->version = 0;
+        pc->show_frame = 1;
+        first_partition_length_in_bytes = 0;
+    }
+    else
+    {
+        unsigned char clear_buffer[10];
+        const unsigned char *clear = data;
+        if (pbi->decrypt_cb)
+        {
+            int n = (int)VPXMIN(sizeof(clear_buffer), data_sz);
+            pbi->decrypt_cb(pbi->decrypt_state, data, clear_buffer, n);
+            clear = clear_buffer;
+        }
+
+        pc->frame_type = (FRAME_TYPE)(clear[0] & 1);
+        pc->version = (clear[0] >> 1) & 7;
+        pc->show_frame = (clear[0] >> 4) & 1;
+        first_partition_length_in_bytes =
+            (clear[0] | (clear[1] << 8) | (clear[2] << 16)) >> 5;
+
+        if (!pbi->ec_active &&
+            (data + first_partition_length_in_bytes > data_end
+            || data + first_partition_length_in_bytes < data))
+            vpx_internal_error(&pc->error, VPX_CODEC_CORRUPT_FRAME,
+                               "Truncated packet or corrupt partition 0 length");
+
+        data += 3;
+        clear += 3;
+
+        vp8_setup_version(pc);
+
+
+        if (pc->frame_type == KEY_FRAME)
+        {
+            /* vet via sync code */
+            /* When error concealment is enabled we should only check the sync
+             * code if we have enough bits available
+             */
+            if (!pbi->ec_active || data + 3 < data_end)
+            {
+                if (clear[0] != 0x9d || clear[1] != 0x01 || clear[2] != 0x2a)
+                    vpx_internal_error(&pc->error, VPX_CODEC_UNSUP_BITSTREAM,
+                                   "Invalid frame sync code");
+            }
+
+            /* If error concealment is enabled we should only parse the new size
+             * if we have enough data. Otherwise we will end up with the wrong
+             * size.
+             */
+            if (!pbi->ec_active || data + 6 < data_end)
+            {
+                pc->Width = (clear[3] | (clear[4] << 8)) & 0x3fff;
+                pc->horiz_scale = clear[4] >> 6;
+                pc->Height = (clear[5] | (clear[6] << 8)) & 0x3fff;
+                pc->vert_scale = clear[6] >> 6;
+            }
+            data += 7;
+        }
+        else
+        {
+          memcpy(&xd->pre, yv12_fb_new, sizeof(YV12_BUFFER_CONFIG));
+          memcpy(&xd->dst, yv12_fb_new, sizeof(YV12_BUFFER_CONFIG));
+        }
+    }
+    if ((!pbi->decoded_key_frame && pc->frame_type != KEY_FRAME))
+    {
+        return -1;
+    }
+
+    init_frame(pbi);
+
+    if (vp8dx_start_decode(bc, data, (unsigned int)(data_end - data),
+                           pbi->decrypt_cb, pbi->decrypt_state))
+        vpx_internal_error(&pc->error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate bool decoder 0");
+    if (pc->frame_type == KEY_FRAME) {
+        (void)vp8_read_bit(bc);  // colorspace
+        pc->clamp_type  = (CLAMP_TYPE)vp8_read_bit(bc);
+    }
+
+    /* Is segmentation enabled */
+    xd->segmentation_enabled = (unsigned char)vp8_read_bit(bc);
+
+    if (xd->segmentation_enabled)
+    {
+        /* Signal whether or not the segmentation map is being explicitly updated this frame. */
+        xd->update_mb_segmentation_map = (unsigned char)vp8_read_bit(bc);
+        xd->update_mb_segmentation_data = (unsigned char)vp8_read_bit(bc);
+
+        if (xd->update_mb_segmentation_data)
+        {
+            xd->mb_segement_abs_delta = (unsigned char)vp8_read_bit(bc);
+
+            memset(xd->segment_feature_data, 0, sizeof(xd->segment_feature_data));
+
+            /* For each segmentation feature (Quant and loop filter level) */
+            for (i = 0; i < MB_LVL_MAX; i++)
+            {
+                for (j = 0; j < MAX_MB_SEGMENTS; j++)
+                {
+                    /* Frame level data */
+                    if (vp8_read_bit(bc))
+                    {
+                        xd->segment_feature_data[i][j] = (signed char)vp8_read_literal(bc, mb_feature_data_bits[i]);
+
+                        if (vp8_read_bit(bc))
+                            xd->segment_feature_data[i][j] = -xd->segment_feature_data[i][j];
+                    }
+                    else
+                        xd->segment_feature_data[i][j] = 0;
+                }
+            }
+        }
+
+        if (xd->update_mb_segmentation_map)
+        {
+            /* Which macro block level features are enabled */
+            memset(xd->mb_segment_tree_probs, 255, sizeof(xd->mb_segment_tree_probs));
+
+            /* Read the probs used to decode the segment id for each macro block. */
+            for (i = 0; i < MB_FEATURE_TREE_PROBS; i++)
+            {
+                /* If not explicitly set value is defaulted to 255 by memset above */
+                if (vp8_read_bit(bc))
+                    xd->mb_segment_tree_probs[i] = (vp8_prob)vp8_read_literal(bc, 8);
+            }
+        }
+    }
+    else
+    {
+        /* No segmentation updates on this frame */
+        xd->update_mb_segmentation_map = 0;
+        xd->update_mb_segmentation_data = 0;
+    }
+
+    /* Read the loop filter level and type */
+    pc->filter_type = (LOOPFILTERTYPE) vp8_read_bit(bc);
+    pc->filter_level = vp8_read_literal(bc, 6);
+    pc->sharpness_level = vp8_read_literal(bc, 3);
+
+    /* Read in loop filter deltas applied at the MB level based on mode or ref frame. */
+    xd->mode_ref_lf_delta_update = 0;
+    xd->mode_ref_lf_delta_enabled = (unsigned char)vp8_read_bit(bc);
+
+    if (xd->mode_ref_lf_delta_enabled)
+    {
+        /* Do the deltas need to be updated */
+        xd->mode_ref_lf_delta_update = (unsigned char)vp8_read_bit(bc);
+
+        if (xd->mode_ref_lf_delta_update)
+        {
+            /* Send update */
+            for (i = 0; i < MAX_REF_LF_DELTAS; i++)
+            {
+                if (vp8_read_bit(bc))
+                {
+                    /*sign = vp8_read_bit( bc );*/
+                    xd->ref_lf_deltas[i] = (signed char)vp8_read_literal(bc, 6);
+
+                    if (vp8_read_bit(bc))        /* Apply sign */
+                        xd->ref_lf_deltas[i] = xd->ref_lf_deltas[i] * -1;
+                }
+            }
+
+            /* Send update */
+            for (i = 0; i < MAX_MODE_LF_DELTAS; i++)
+            {
+                if (vp8_read_bit(bc))
+                {
+                    /*sign = vp8_read_bit( bc );*/
+                    xd->mode_lf_deltas[i] = (signed char)vp8_read_literal(bc, 6);
+
+                    if (vp8_read_bit(bc))        /* Apply sign */
+                        xd->mode_lf_deltas[i] = xd->mode_lf_deltas[i] * -1;
+                }
+            }
+        }
+    }
+
+    setup_token_decoder(pbi, data + first_partition_length_in_bytes);
+
+    xd->current_bc = &pbi->mbc[0];
+
+    /* Read the default quantizers. */
+    {
+        int Q, q_update;
+
+        Q = vp8_read_literal(bc, 7);  /* AC 1st order Q = default */
+        pc->base_qindex = Q;
+        q_update = 0;
+        pc->y1dc_delta_q = get_delta_q(bc, pc->y1dc_delta_q, &q_update);
+        pc->y2dc_delta_q = get_delta_q(bc, pc->y2dc_delta_q, &q_update);
+        pc->y2ac_delta_q = get_delta_q(bc, pc->y2ac_delta_q, &q_update);
+        pc->uvdc_delta_q = get_delta_q(bc, pc->uvdc_delta_q, &q_update);
+        pc->uvac_delta_q = get_delta_q(bc, pc->uvac_delta_q, &q_update);
+
+        if (q_update)
+            vp8cx_init_de_quantizer(pbi);
+
+        /* MB level dequantizer setup */
+        vp8_mb_init_dequantizer(pbi, &pbi->mb);
+    }
+
+    /* Determine if the golden frame or ARF buffer should be updated and how.
+     * For all non key frames the GF and ARF refresh flags and sign bias
+     * flags must be set explicitly.
+     */
+    if (pc->frame_type != KEY_FRAME)
+    {
+        /* Should the GF or ARF be updated from the current frame */
+        pc->refresh_golden_frame = vp8_read_bit(bc);
+#if CONFIG_ERROR_CONCEALMENT
+        /* Assume we shouldn't refresh golden if the bit is missing */
+        xd->corrupted |= vp8dx_bool_error(bc);
+        if (pbi->ec_active && xd->corrupted)
+            pc->refresh_golden_frame = 0;
+#endif
+
+        pc->refresh_alt_ref_frame = vp8_read_bit(bc);
+#if CONFIG_ERROR_CONCEALMENT
+        /* Assume we shouldn't refresh altref if the bit is missing */
+        xd->corrupted |= vp8dx_bool_error(bc);
+        if (pbi->ec_active && xd->corrupted)
+            pc->refresh_alt_ref_frame = 0;
+#endif
+
+        /* Buffer to buffer copy flags. */
+        pc->copy_buffer_to_gf = 0;
+
+        if (!pc->refresh_golden_frame)
+            pc->copy_buffer_to_gf = vp8_read_literal(bc, 2);
+
+#if CONFIG_ERROR_CONCEALMENT
+        /* Assume we shouldn't copy to the golden if the bit is missing */
+        xd->corrupted |= vp8dx_bool_error(bc);
+        if (pbi->ec_active && xd->corrupted)
+            pc->copy_buffer_to_gf = 0;
+#endif
+
+        pc->copy_buffer_to_arf = 0;
+
+        if (!pc->refresh_alt_ref_frame)
+            pc->copy_buffer_to_arf = vp8_read_literal(bc, 2);
+
+#if CONFIG_ERROR_CONCEALMENT
+        /* Assume we shouldn't copy to the alt-ref if the bit is missing */
+        xd->corrupted |= vp8dx_bool_error(bc);
+        if (pbi->ec_active && xd->corrupted)
+            pc->copy_buffer_to_arf = 0;
+#endif
+
+
+        pc->ref_frame_sign_bias[GOLDEN_FRAME] = vp8_read_bit(bc);
+        pc->ref_frame_sign_bias[ALTREF_FRAME] = vp8_read_bit(bc);
+    }
+
+    pc->refresh_entropy_probs = vp8_read_bit(bc);
+#if CONFIG_ERROR_CONCEALMENT
+    /* Assume we shouldn't refresh the probabilities if the bit is
+     * missing */
+    xd->corrupted |= vp8dx_bool_error(bc);
+    if (pbi->ec_active && xd->corrupted)
+        pc->refresh_entropy_probs = 0;
+#endif
+    if (pc->refresh_entropy_probs == 0)
+    {
+        memcpy(&pc->lfc, &pc->fc, sizeof(pc->fc));
+    }
+
+    pc->refresh_last_frame = pc->frame_type == KEY_FRAME  ||  vp8_read_bit(bc);
+
+#if CONFIG_ERROR_CONCEALMENT
+    /* Assume we should refresh the last frame if the bit is missing */
+    xd->corrupted |= vp8dx_bool_error(bc);
+    if (pbi->ec_active && xd->corrupted)
+        pc->refresh_last_frame = 1;
+#endif
+
+    if (0)
+    {
+        FILE *z = fopen("decodestats.stt", "a");
+        fprintf(z, "%6d F:%d,G:%d,A:%d,L:%d,Q:%d\n",
+                pc->current_video_frame,
+                pc->frame_type,
+                pc->refresh_golden_frame,
+                pc->refresh_alt_ref_frame,
+                pc->refresh_last_frame,
+                pc->base_qindex);
+        fclose(z);
+    }
+
+    {
+        pbi->independent_partitions = 1;
+
+        /* read coef probability tree */
+        for (i = 0; i < BLOCK_TYPES; i++)
+            for (j = 0; j < COEF_BANDS; j++)
+                for (k = 0; k < PREV_COEF_CONTEXTS; k++)
+                    for (l = 0; l < ENTROPY_NODES; l++)
+                    {
+
+                        vp8_prob *const p = pc->fc.coef_probs [i][j][k] + l;
+
+                        if (vp8_read(bc, vp8_coef_update_probs [i][j][k][l]))
+                        {
+                            *p = (vp8_prob)vp8_read_literal(bc, 8);
+
+                        }
+                        if (k > 0 && *p != pc->fc.coef_probs[i][j][k-1][l])
+                            pbi->independent_partitions = 0;
+
+                    }
+    }
+
+    /* clear out the coeff buffer */
+    memset(xd->qcoeff, 0, sizeof(xd->qcoeff));
+
+    vp8_decode_mode_mvs(pbi);
+
+#if CONFIG_ERROR_CONCEALMENT
+    if (pbi->ec_active &&
+            pbi->mvs_corrupt_from_mb < (unsigned int)pc->mb_cols * pc->mb_rows)
+    {
+        /* Motion vectors are missing in this frame. We will try to estimate
+         * them and then continue decoding the frame as usual */
+        vp8_estimate_missing_mvs(pbi);
+    }
+#endif
+
+    memset(pc->above_context, 0, sizeof(ENTROPY_CONTEXT_PLANES) * pc->mb_cols);
+    pbi->frame_corrupt_residual = 0;
+
+#if CONFIG_MULTITHREAD
+    if (pbi->b_multithreaded_rd && pc->multi_token_partition != ONE_PARTITION)
+    {
+        unsigned int thread;
+        vp8mt_decode_mb_rows(pbi, xd);
+        vp8_yv12_extend_frame_borders(yv12_fb_new);
+        for (thread = 0; thread < pbi->decoding_thread_count; ++thread)
+            corrupt_tokens |= pbi->mb_row_di[thread].mbd.corrupted;
+    }
+    else
+#endif
+    {
+        decode_mb_rows(pbi);
+        corrupt_tokens |= xd->corrupted;
+    }
+
+    /* Collect information about decoder corruption. */
+    /* 1. Check first boolean decoder for errors. */
+    yv12_fb_new->corrupted = vp8dx_bool_error(bc);
+    /* 2. Check the macroblock information */
+    yv12_fb_new->corrupted |= corrupt_tokens;
+
+    if (!pbi->decoded_key_frame)
+    {
+        if (pc->frame_type == KEY_FRAME &&
+            !yv12_fb_new->corrupted)
+            pbi->decoded_key_frame = 1;
+        else
+            vpx_internal_error(&pbi->common.error, VPX_CODEC_CORRUPT_FRAME,
+                               "A stream must start with a complete key frame");
+    }
+
+    /* vpx_log("Decoder: Frame Decoded, Size Roughly:%d bytes  \n",bc->pos+pbi->bc2.pos); */
+
+    if (pc->refresh_entropy_probs == 0)
+    {
+        memcpy(&pc->fc, &pc->lfc, sizeof(pc->fc));
+        pbi->independent_partitions = prev_independent_partitions;
+    }
+
+#ifdef PACKET_TESTING
+    {
+        FILE *f = fopen("decompressor.VP8", "ab");
+        unsigned int size = pbi->bc2.pos + pbi->bc.pos + 8;
+        fwrite((void *) &size, 4, 1, f);
+        fwrite((void *) pbi->Source, size, 1, f);
+        fclose(f);
+    }
+#endif
+
+    return 0;
+}
diff --git a/libvpx/libvpx/vp8/decoder/decodemv.c b/libvpx/libvpx/vp8/decoder/decodemv.c
new file mode 100644
index 0000000..1d155e7
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/decodemv.c
@@ -0,0 +1,670 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "treereader.h"
+#include "vp8/common/entropymv.h"
+#include "vp8/common/entropymode.h"
+#include "onyxd_int.h"
+#include "vp8/common/findnearmv.h"
+
+#if CONFIG_DEBUG
+#include <assert.h>
+#endif
+static B_PREDICTION_MODE read_bmode(vp8_reader *bc, const vp8_prob *p)
+{
+    const int i = vp8_treed_read(bc, vp8_bmode_tree, p);
+
+    return (B_PREDICTION_MODE)i;
+}
+
+static MB_PREDICTION_MODE read_ymode(vp8_reader *bc, const vp8_prob *p)
+{
+    const int i = vp8_treed_read(bc, vp8_ymode_tree, p);
+
+    return (MB_PREDICTION_MODE)i;
+}
+
+static MB_PREDICTION_MODE read_kf_ymode(vp8_reader *bc, const vp8_prob *p)
+{
+    const int i = vp8_treed_read(bc, vp8_kf_ymode_tree, p);
+
+    return (MB_PREDICTION_MODE)i;
+}
+
+static MB_PREDICTION_MODE read_uv_mode(vp8_reader *bc, const vp8_prob *p)
+{
+    const int i = vp8_treed_read(bc, vp8_uv_mode_tree, p);
+
+    return (MB_PREDICTION_MODE)i;
+}
+
+static void read_kf_modes(VP8D_COMP *pbi, MODE_INFO *mi)
+{
+    vp8_reader *const bc = & pbi->mbc[8];
+    const int mis = pbi->common.mode_info_stride;
+
+    mi->mbmi.ref_frame = INTRA_FRAME;
+    mi->mbmi.mode = read_kf_ymode(bc, vp8_kf_ymode_prob);
+
+    if (mi->mbmi.mode == B_PRED)
+    {
+        int i = 0;
+        mi->mbmi.is_4x4 = 1;
+
+        do
+        {
+            const B_PREDICTION_MODE A = above_block_mode(mi, i, mis);
+            const B_PREDICTION_MODE L = left_block_mode(mi, i);
+
+            mi->bmi[i].as_mode =
+                read_bmode(bc, vp8_kf_bmode_prob [A] [L]);
+        }
+        while (++i < 16);
+    }
+
+    mi->mbmi.uv_mode = read_uv_mode(bc, vp8_kf_uv_mode_prob);
+}
+
+static int read_mvcomponent(vp8_reader *r, const MV_CONTEXT *mvc)
+{
+    const vp8_prob *const p = (const vp8_prob *) mvc;
+    int x = 0;
+
+    if (vp8_read(r, p [mvpis_short]))  /* Large */
+    {
+        int i = 0;
+
+        do
+        {
+            x += vp8_read(r, p [MVPbits + i]) << i;
+        }
+        while (++i < 3);
+
+        i = mvlong_width - 1;  /* Skip bit 3, which is sometimes implicit */
+
+        do
+        {
+            x += vp8_read(r, p [MVPbits + i]) << i;
+        }
+        while (--i > 3);
+
+        if (!(x & 0xFFF0)  ||  vp8_read(r, p [MVPbits + 3]))
+            x += 8;
+    }
+    else   /* small */
+        x = vp8_treed_read(r, vp8_small_mvtree, p + MVPshort);
+
+    if (x  &&  vp8_read(r, p [MVPsign]))
+        x = -x;
+
+    return x;
+}
+
+static void read_mv(vp8_reader *r, MV *mv, const MV_CONTEXT *mvc)
+{
+    mv->row = (short)(read_mvcomponent(r,   mvc) * 2);
+    mv->col = (short)(read_mvcomponent(r, ++mvc) * 2);
+}
+
+
+static void read_mvcontexts(vp8_reader *bc, MV_CONTEXT *mvc)
+{
+    int i = 0;
+
+    do
+    {
+        const vp8_prob *up = vp8_mv_update_probs[i].prob;
+        vp8_prob *p = (vp8_prob *)(mvc + i);
+        vp8_prob *const pstop = p + MVPcount;
+
+        do
+        {
+            if (vp8_read(bc, *up++))
+            {
+                const vp8_prob x = (vp8_prob)vp8_read_literal(bc, 7);
+
+                *p = x ? x << 1 : 1;
+            }
+        }
+        while (++p < pstop);
+    }
+    while (++i < 2);
+}
+
+static const unsigned char mbsplit_fill_count[4] = {8, 8, 4, 1};
+static const unsigned char mbsplit_fill_offset[4][16] = {
+    { 0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15},
+    { 0,  1,  4,  5,  8,  9, 12, 13,  2,  3,   6,  7, 10, 11, 14, 15},
+    { 0,  1,  4,  5,  2,  3,  6,  7,  8,  9,  12, 13, 10, 11, 14, 15},
+    { 0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15}
+};
+
+
+static void mb_mode_mv_init(VP8D_COMP *pbi)
+{
+    vp8_reader *const bc = & pbi->mbc[8];
+    MV_CONTEXT *const mvc = pbi->common.fc.mvc;
+
+#if CONFIG_ERROR_CONCEALMENT
+    /* Default is that no macroblock is corrupt, therefore we initialize
+     * mvs_corrupt_from_mb to something very big, which we can be sure is
+     * outside the frame. */
+    pbi->mvs_corrupt_from_mb = UINT_MAX;
+#endif
+    /* Read the mb_no_coeff_skip flag */
+    pbi->common.mb_no_coeff_skip = (int)vp8_read_bit(bc);
+
+    pbi->prob_skip_false = 0;
+    if (pbi->common.mb_no_coeff_skip)
+        pbi->prob_skip_false = (vp8_prob)vp8_read_literal(bc, 8);
+
+    if(pbi->common.frame_type != KEY_FRAME)
+    {
+        pbi->prob_intra = (vp8_prob)vp8_read_literal(bc, 8);
+        pbi->prob_last  = (vp8_prob)vp8_read_literal(bc, 8);
+        pbi->prob_gf    = (vp8_prob)vp8_read_literal(bc, 8);
+
+        if (vp8_read_bit(bc))
+        {
+            int i = 0;
+
+            do
+            {
+                pbi->common.fc.ymode_prob[i] =
+                    (vp8_prob) vp8_read_literal(bc, 8);
+            }
+            while (++i < 4);
+        }
+
+        if (vp8_read_bit(bc))
+        {
+            int i = 0;
+
+            do
+            {
+                pbi->common.fc.uv_mode_prob[i] =
+                    (vp8_prob) vp8_read_literal(bc, 8);
+            }
+            while (++i < 3);
+        }
+
+        read_mvcontexts(bc, mvc);
+    }
+}
+
+const vp8_prob vp8_sub_mv_ref_prob3 [8][VP8_SUBMVREFS-1] =
+{
+    { 147, 136, 18 },   /* SUBMVREF_NORMAL          */
+    { 223, 1  , 34 },   /* SUBMVREF_LEFT_ABOVE_SAME */
+    { 106, 145, 1  },   /* SUBMVREF_LEFT_ZED        */
+    { 208, 1  , 1  },   /* SUBMVREF_LEFT_ABOVE_ZED  */
+    { 179, 121, 1  },   /* SUBMVREF_ABOVE_ZED       */
+    { 223, 1  , 34 },   /* SUBMVREF_LEFT_ABOVE_SAME */
+    { 179, 121, 1  },   /* SUBMVREF_ABOVE_ZED       */
+    { 208, 1  , 1  }    /* SUBMVREF_LEFT_ABOVE_ZED  */
+};
+
+static
+const vp8_prob * get_sub_mv_ref_prob(const int left, const int above)
+{
+    int lez = (left == 0);
+    int aez = (above == 0);
+    int lea = (left == above);
+    const vp8_prob * prob;
+
+    prob = vp8_sub_mv_ref_prob3[(aez << 2) |
+                                (lez << 1) |
+                                (lea)];
+
+    return prob;
+}
+
+static void decode_split_mv(vp8_reader *const bc, MODE_INFO *mi,
+                        const MODE_INFO *left_mb, const MODE_INFO *above_mb,
+                        MB_MODE_INFO *mbmi, int_mv best_mv,
+                        MV_CONTEXT *const mvc, int mb_to_left_edge,
+                        int mb_to_right_edge, int mb_to_top_edge,
+                        int mb_to_bottom_edge)
+{
+    int s;      /* split configuration (16x8, 8x16, 8x8, 4x4) */
+    int num_p;  /* number of partitions in the split configuration
+                  (see vp8_mbsplit_count) */
+    int j = 0;
+
+    s = 3;
+    num_p = 16;
+    if( vp8_read(bc, 110) )
+    {
+        s = 2;
+        num_p = 4;
+        if( vp8_read(bc, 111) )
+        {
+            s = vp8_read(bc, 150);
+            num_p = 2;
+        }
+    }
+
+    do  /* for each subset j */
+    {
+        int_mv leftmv, abovemv;
+        int_mv blockmv;
+        int k;  /* first block in subset j */
+
+        const vp8_prob *prob;
+        k = vp8_mbsplit_offset[s][j];
+
+        if (!(k & 3))
+        {
+            /* On L edge, get from MB to left of us */
+            if(left_mb->mbmi.mode != SPLITMV)
+                leftmv.as_int =  left_mb->mbmi.mv.as_int;
+            else
+                leftmv.as_int =  (left_mb->bmi + k + 4 - 1)->mv.as_int;
+        }
+        else
+            leftmv.as_int =  (mi->bmi + k - 1)->mv.as_int;
+
+        if (!(k >> 2))
+        {
+            /* On top edge, get from MB above us */
+            if(above_mb->mbmi.mode != SPLITMV)
+                abovemv.as_int =  above_mb->mbmi.mv.as_int;
+            else
+                abovemv.as_int =  (above_mb->bmi + k + 16 - 4)->mv.as_int;
+        }
+        else
+            abovemv.as_int = (mi->bmi + k - 4)->mv.as_int;
+
+        prob = get_sub_mv_ref_prob(leftmv.as_int, abovemv.as_int);
+
+        if( vp8_read(bc, prob[0]) )
+        {
+            if( vp8_read(bc, prob[1]) )
+            {
+                blockmv.as_int = 0;
+                if( vp8_read(bc, prob[2]) )
+                {
+                    blockmv.as_mv.row = read_mvcomponent(bc, &mvc[0]) * 2;
+                    blockmv.as_mv.row += best_mv.as_mv.row;
+                    blockmv.as_mv.col = read_mvcomponent(bc, &mvc[1]) * 2;
+                    blockmv.as_mv.col += best_mv.as_mv.col;
+                }
+            }
+            else
+            {
+                blockmv.as_int = abovemv.as_int;
+            }
+        }
+        else
+        {
+            blockmv.as_int = leftmv.as_int;
+        }
+
+        mbmi->need_to_clamp_mvs |= vp8_check_mv_bounds(&blockmv,
+                                                  mb_to_left_edge,
+                                                  mb_to_right_edge,
+                                                  mb_to_top_edge,
+                                                  mb_to_bottom_edge);
+
+        {
+            /* Fill (uniform) modes, mvs of jth subset.
+             Must do it here because ensuing subsets can
+             refer back to us via "left" or "above". */
+            const unsigned char *fill_offset;
+            unsigned int fill_count = mbsplit_fill_count[s];
+
+            fill_offset = &mbsplit_fill_offset[s]
+                             [(unsigned char)j * mbsplit_fill_count[s]];
+
+            do {
+                mi->bmi[ *fill_offset].mv.as_int = blockmv.as_int;
+                fill_offset++;
+            }while (--fill_count);
+        }
+
+    }
+    while (++j < num_p);
+
+    mbmi->partitioning = s;
+}
+
+static void read_mb_modes_mv(VP8D_COMP *pbi, MODE_INFO *mi, MB_MODE_INFO *mbmi)
+{
+    vp8_reader *const bc = & pbi->mbc[8];
+    mbmi->ref_frame = (MV_REFERENCE_FRAME) vp8_read(bc, pbi->prob_intra);
+    if (mbmi->ref_frame)    /* inter MB */
+    {
+        enum {CNT_INTRA, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV};
+        int cnt[4];
+        int *cntx = cnt;
+        int_mv near_mvs[4];
+        int_mv *nmv = near_mvs;
+        const int mis = pbi->mb.mode_info_stride;
+        const MODE_INFO *above = mi - mis;
+        const MODE_INFO *left = mi - 1;
+        const MODE_INFO *aboveleft = above - 1;
+        int *ref_frame_sign_bias = pbi->common.ref_frame_sign_bias;
+
+        mbmi->need_to_clamp_mvs = 0;
+
+        if (vp8_read(bc, pbi->prob_last))
+        {
+            mbmi->ref_frame =
+                (MV_REFERENCE_FRAME)((int)(2 + vp8_read(bc, pbi->prob_gf)));
+        }
+
+        /* Zero accumulators */
+        nmv[0].as_int = nmv[1].as_int = nmv[2].as_int = 0;
+        cnt[0] = cnt[1] = cnt[2] = cnt[3] = 0;
+
+        /* Process above */
+        if (above->mbmi.ref_frame != INTRA_FRAME)
+        {
+            if (above->mbmi.mv.as_int)
+            {
+                (++nmv)->as_int = above->mbmi.mv.as_int;
+                mv_bias(ref_frame_sign_bias[above->mbmi.ref_frame],
+                        mbmi->ref_frame, nmv, ref_frame_sign_bias);
+                ++cntx;
+            }
+
+            *cntx += 2;
+        }
+
+        /* Process left */
+        if (left->mbmi.ref_frame != INTRA_FRAME)
+        {
+            if (left->mbmi.mv.as_int)
+            {
+                int_mv this_mv;
+
+                this_mv.as_int = left->mbmi.mv.as_int;
+                mv_bias(ref_frame_sign_bias[left->mbmi.ref_frame],
+                        mbmi->ref_frame, &this_mv, ref_frame_sign_bias);
+
+                if (this_mv.as_int != nmv->as_int)
+                {
+                    (++nmv)->as_int = this_mv.as_int;
+                    ++cntx;
+                }
+
+                *cntx += 2;
+            }
+            else
+                cnt[CNT_INTRA] += 2;
+        }
+
+        /* Process above left */
+        if (aboveleft->mbmi.ref_frame != INTRA_FRAME)
+        {
+            if (aboveleft->mbmi.mv.as_int)
+            {
+                int_mv this_mv;
+
+                this_mv.as_int = aboveleft->mbmi.mv.as_int;
+                mv_bias(ref_frame_sign_bias[aboveleft->mbmi.ref_frame],
+                        mbmi->ref_frame, &this_mv, ref_frame_sign_bias);
+
+                if (this_mv.as_int != nmv->as_int)
+                {
+                    (++nmv)->as_int = this_mv.as_int;
+                    ++cntx;
+                }
+
+                *cntx += 1;
+            }
+            else
+                cnt[CNT_INTRA] += 1;
+        }
+
+        if( vp8_read(bc, vp8_mode_contexts [cnt[CNT_INTRA]] [0]) )
+        {
+
+            /* If we have three distinct MV's ... */
+            /* See if above-left MV can be merged with NEAREST */
+            cnt[CNT_NEAREST] += ( (cnt[CNT_SPLITMV] > 0) &
+                (nmv->as_int == near_mvs[CNT_NEAREST].as_int));
+
+            /* Swap near and nearest if necessary */
+            if (cnt[CNT_NEAR] > cnt[CNT_NEAREST])
+            {
+                int tmp;
+                tmp = cnt[CNT_NEAREST];
+                cnt[CNT_NEAREST] = cnt[CNT_NEAR];
+                cnt[CNT_NEAR] = tmp;
+                tmp = near_mvs[CNT_NEAREST].as_int;
+                near_mvs[CNT_NEAREST].as_int = near_mvs[CNT_NEAR].as_int;
+                near_mvs[CNT_NEAR].as_int = tmp;
+            }
+
+            if( vp8_read(bc, vp8_mode_contexts [cnt[CNT_NEAREST]] [1]) )
+            {
+
+                if( vp8_read(bc, vp8_mode_contexts [cnt[CNT_NEAR]] [2]) )
+                {
+                    int mb_to_top_edge;
+                    int mb_to_bottom_edge;
+                    int mb_to_left_edge;
+                    int mb_to_right_edge;
+                    MV_CONTEXT *const mvc = pbi->common.fc.mvc;
+                    int near_index;
+
+                    mb_to_top_edge = pbi->mb.mb_to_top_edge;
+                    mb_to_bottom_edge = pbi->mb.mb_to_bottom_edge;
+                    mb_to_top_edge -= LEFT_TOP_MARGIN;
+                    mb_to_bottom_edge += RIGHT_BOTTOM_MARGIN;
+                    mb_to_right_edge = pbi->mb.mb_to_right_edge;
+                    mb_to_right_edge += RIGHT_BOTTOM_MARGIN;
+                    mb_to_left_edge = pbi->mb.mb_to_left_edge;
+                    mb_to_left_edge -= LEFT_TOP_MARGIN;
+
+                    /* Use near_mvs[0] to store the "best" MV */
+                    near_index = CNT_INTRA +
+                        (cnt[CNT_NEAREST] >= cnt[CNT_INTRA]);
+
+                    vp8_clamp_mv2(&near_mvs[near_index], &pbi->mb);
+
+                    cnt[CNT_SPLITMV] = ((above->mbmi.mode == SPLITMV)
+                                        + (left->mbmi.mode == SPLITMV)) * 2
+                                       + (aboveleft->mbmi.mode == SPLITMV);
+
+                    if( vp8_read(bc, vp8_mode_contexts [cnt[CNT_SPLITMV]] [3]) )
+                    {
+                        decode_split_mv(bc, mi, left, above,
+                                                    mbmi,
+                                                    near_mvs[near_index],
+                                                    mvc, mb_to_left_edge,
+                                                    mb_to_right_edge,
+                                                    mb_to_top_edge,
+                                                    mb_to_bottom_edge);
+                        mbmi->mv.as_int = mi->bmi[15].mv.as_int;
+                        mbmi->mode =  SPLITMV;
+                        mbmi->is_4x4 = 1;
+                    }
+                    else
+                    {
+                        int_mv *const mbmi_mv = & mbmi->mv;
+                        read_mv(bc, &mbmi_mv->as_mv, (const MV_CONTEXT *) mvc);
+                        mbmi_mv->as_mv.row += near_mvs[near_index].as_mv.row;
+                        mbmi_mv->as_mv.col += near_mvs[near_index].as_mv.col;
+
+                        /* Don't need to check this on NEARMV and NEARESTMV
+                         * modes since those modes clamp the MV. The NEWMV mode
+                         * does not, so signal to the prediction stage whether
+                         * special handling may be required.
+                         */
+                        mbmi->need_to_clamp_mvs =
+                            vp8_check_mv_bounds(mbmi_mv, mb_to_left_edge,
+                                                mb_to_right_edge,
+                                                mb_to_top_edge,
+                                                mb_to_bottom_edge);
+                        mbmi->mode =  NEWMV;
+                    }
+                }
+                else
+                {
+                    mbmi->mode =  NEARMV;
+                    mbmi->mv.as_int = near_mvs[CNT_NEAR].as_int;
+                    vp8_clamp_mv2(&mbmi->mv, &pbi->mb);
+                }
+            }
+            else
+            {
+                mbmi->mode =  NEARESTMV;
+                mbmi->mv.as_int = near_mvs[CNT_NEAREST].as_int;
+                vp8_clamp_mv2(&mbmi->mv, &pbi->mb);
+            }
+        }
+        else
+        {
+            mbmi->mode =  ZEROMV;
+            mbmi->mv.as_int = 0;
+        }
+
+#if CONFIG_ERROR_CONCEALMENT
+        if(pbi->ec_enabled && (mbmi->mode != SPLITMV))
+        {
+            mi->bmi[ 0].mv.as_int =
+            mi->bmi[ 1].mv.as_int =
+            mi->bmi[ 2].mv.as_int =
+            mi->bmi[ 3].mv.as_int =
+            mi->bmi[ 4].mv.as_int =
+            mi->bmi[ 5].mv.as_int =
+            mi->bmi[ 6].mv.as_int =
+            mi->bmi[ 7].mv.as_int =
+            mi->bmi[ 8].mv.as_int =
+            mi->bmi[ 9].mv.as_int =
+            mi->bmi[10].mv.as_int =
+            mi->bmi[11].mv.as_int =
+            mi->bmi[12].mv.as_int =
+            mi->bmi[13].mv.as_int =
+            mi->bmi[14].mv.as_int =
+            mi->bmi[15].mv.as_int = mbmi->mv.as_int;
+        }
+#endif
+    }
+    else
+    {
+        /* required for left and above block mv */
+        mbmi->mv.as_int = 0;
+
+        /* MB is intra coded */
+        if ((mbmi->mode = read_ymode(bc, pbi->common.fc.ymode_prob)) == B_PRED)
+        {
+            int j = 0;
+            mbmi->is_4x4 = 1;
+            do
+            {
+                mi->bmi[j].as_mode = read_bmode(bc, pbi->common.fc.bmode_prob);
+            }
+            while (++j < 16);
+        }
+
+        mbmi->uv_mode = read_uv_mode(bc, pbi->common.fc.uv_mode_prob);
+    }
+
+}
+
+static void read_mb_features(vp8_reader *r, MB_MODE_INFO *mi, MACROBLOCKD *x)
+{
+    /* Is segmentation enabled */
+    if (x->segmentation_enabled && x->update_mb_segmentation_map)
+    {
+        /* If so then read the segment id. */
+        if (vp8_read(r, x->mb_segment_tree_probs[0]))
+            mi->segment_id =
+                (unsigned char)(2 + vp8_read(r, x->mb_segment_tree_probs[2]));
+        else
+            mi->segment_id =
+                (unsigned char)(vp8_read(r, x->mb_segment_tree_probs[1]));
+    }
+}
+
+static void decode_mb_mode_mvs(VP8D_COMP *pbi, MODE_INFO *mi,
+                               MB_MODE_INFO *mbmi)
+{
+    (void)mbmi;
+
+    /* Read the Macroblock segmentation map if it is being updated explicitly
+     * this frame (reset to 0 above by default)
+     * By default on a key frame reset all MBs to segment 0
+     */
+    if (pbi->mb.update_mb_segmentation_map)
+        read_mb_features(&pbi->mbc[8], &mi->mbmi, &pbi->mb);
+    else if(pbi->common.frame_type == KEY_FRAME)
+        mi->mbmi.segment_id = 0;
+
+    /* Read the macroblock coeff skip flag if this feature is in use,
+     * else default to 0 */
+    if (pbi->common.mb_no_coeff_skip)
+        mi->mbmi.mb_skip_coeff = vp8_read(&pbi->mbc[8], pbi->prob_skip_false);
+    else
+        mi->mbmi.mb_skip_coeff = 0;
+
+    mi->mbmi.is_4x4 = 0;
+    if(pbi->common.frame_type == KEY_FRAME)
+        read_kf_modes(pbi, mi);
+    else
+        read_mb_modes_mv(pbi, mi, &mi->mbmi);
+
+}
+
+void vp8_decode_mode_mvs(VP8D_COMP *pbi)
+{
+    MODE_INFO *mi = pbi->common.mi;
+    int mb_row = -1;
+    int mb_to_right_edge_start;
+
+    mb_mode_mv_init(pbi);
+
+    pbi->mb.mb_to_top_edge = 0;
+    pbi->mb.mb_to_bottom_edge = ((pbi->common.mb_rows - 1) * 16) << 3;
+    mb_to_right_edge_start = ((pbi->common.mb_cols - 1) * 16) << 3;
+
+    while (++mb_row < pbi->common.mb_rows)
+    {
+        int mb_col = -1;
+
+        pbi->mb.mb_to_left_edge =  0;
+        pbi->mb.mb_to_right_edge = mb_to_right_edge_start;
+
+        while (++mb_col < pbi->common.mb_cols)
+        {
+#if CONFIG_ERROR_CONCEALMENT
+            int mb_num = mb_row * pbi->common.mb_cols + mb_col;
+#endif
+
+            decode_mb_mode_mvs(pbi, mi, &mi->mbmi);
+
+#if CONFIG_ERROR_CONCEALMENT
+            /* look for corruption. set mvs_corrupt_from_mb to the current
+             * mb_num if the frame is corrupt from this macroblock. */
+            if (vp8dx_bool_error(&pbi->mbc[8]) && mb_num <
+                (int)pbi->mvs_corrupt_from_mb)
+            {
+                pbi->mvs_corrupt_from_mb = mb_num;
+                /* no need to continue since the partition is corrupt from
+                 * here on.
+                 */
+                return;
+            }
+#endif
+
+            pbi->mb.mb_to_left_edge -= (16 << 3);
+            pbi->mb.mb_to_right_edge -= (16 << 3);
+            mi++;       /* next macroblock */
+        }
+        pbi->mb.mb_to_top_edge -= (16 << 3);
+        pbi->mb.mb_to_bottom_edge -= (16 << 3);
+
+        mi++;           /* skip left predictor each row */
+    }
+}
diff --git a/libvpx/libvpx/vp8/decoder/decodemv.h b/libvpx/libvpx/vp8/decoder/decodemv.h
new file mode 100644
index 0000000..f33b073
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/decodemv.h
@@ -0,0 +1,26 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_DECODER_DECODEMV_H_
+#define VP8_DECODER_DECODEMV_H_
+
+#include "onyxd_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_decode_mode_mvs(VP8D_COMP *);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_DECODER_DECODEMV_H_
diff --git a/libvpx/libvpx/vp8/decoder/decoderthreading.h b/libvpx/libvpx/vp8/decoder/decoderthreading.h
new file mode 100644
index 0000000..c563cf6
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/decoderthreading.h
@@ -0,0 +1,30 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_DECODER_DECODERTHREADING_H_
+#define VP8_DECODER_DECODERTHREADING_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_MULTITHREAD
+void vp8mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd);
+void vp8_decoder_remove_threads(VP8D_COMP *pbi);
+void vp8_decoder_create_threads(VP8D_COMP *pbi);
+void vp8mt_alloc_temp_buffers(VP8D_COMP *pbi, int width, int prev_mb_rows);
+void vp8mt_de_alloc_temp_buffers(VP8D_COMP *pbi, int mb_rows);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_DECODER_DECODERTHREADING_H_
diff --git a/libvpx/libvpx/vp8/decoder/detokenize.c b/libvpx/libvpx/vp8/decoder/detokenize.c
new file mode 100644
index 0000000..fcc7533
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/detokenize.c
@@ -0,0 +1,245 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vp8/common/blockd.h"
+#include "onyxd_int.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "detokenize.h"
+
+void vp8_reset_mb_tokens_context(MACROBLOCKD *x)
+{
+    ENTROPY_CONTEXT *a_ctx = ((ENTROPY_CONTEXT *)x->above_context);
+    ENTROPY_CONTEXT *l_ctx = ((ENTROPY_CONTEXT *)x->left_context);
+
+    memset(a_ctx, 0, sizeof(ENTROPY_CONTEXT_PLANES)-1);
+    memset(l_ctx, 0, sizeof(ENTROPY_CONTEXT_PLANES)-1);
+
+    /* Clear entropy contexts for Y2 blocks */
+    if (!x->mode_info_context->mbmi.is_4x4)
+    {
+        a_ctx[8] = l_ctx[8] = 0;
+    }
+}
+
+/*
+    ------------------------------------------------------------------------------
+    Residual decoding (Paragraph 13.2 / 13.3)
+*/
+static const uint8_t kBands[16 + 1] = {
+  0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7,
+  0  /* extra entry as sentinel */
+};
+
+static const uint8_t kCat3[] = { 173, 148, 140, 0 };
+static const uint8_t kCat4[] = { 176, 155, 140, 135, 0 };
+static const uint8_t kCat5[] = { 180, 157, 141, 134, 130, 0 };
+static const uint8_t kCat6[] =
+  { 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129, 0 };
+static const uint8_t* const kCat3456[] = { kCat3, kCat4, kCat5, kCat6 };
+static const uint8_t kZigzag[16] = {
+  0, 1, 4, 8,  5, 2, 3, 6,  9, 12, 13, 10,  7, 11, 14, 15
+};
+
+#define VP8GetBit vp8dx_decode_bool
+#define NUM_PROBAS  11
+#define NUM_CTX  3
+
+/* for const-casting */
+typedef const uint8_t (*ProbaArray)[NUM_CTX][NUM_PROBAS];
+
+static int GetSigned(BOOL_DECODER *br, int value_to_sign)
+{
+    int split = (br->range + 1) >> 1;
+    VP8_BD_VALUE bigsplit = (VP8_BD_VALUE)split << (VP8_BD_VALUE_SIZE - 8);
+    int v;
+
+    if(br->count < 0)
+        vp8dx_bool_decoder_fill(br);
+
+    if ( br->value < bigsplit )
+    {
+        br->range = split;
+        v= value_to_sign;
+    }
+    else
+    {
+        br->range = br->range-split;
+        br->value = br->value-bigsplit;
+        v = -value_to_sign;
+    }
+    br->range +=br->range;
+    br->value +=br->value;
+    br->count--;
+
+    return v;
+}
+/*
+   Returns the position of the last non-zero coeff plus one
+   (and 0 if there's no coeff at all)
+*/
+static int GetCoeffs(BOOL_DECODER *br, ProbaArray prob,
+                     int ctx, int n, int16_t* out)
+{
+    const uint8_t* p = prob[n][ctx];
+    if (!VP8GetBit(br, p[0]))
+    {   /* first EOB is more a 'CBP' bit. */
+        return 0;
+    }
+    while (1)
+    {
+        ++n;
+        if (!VP8GetBit(br, p[1]))
+        {
+            p = prob[kBands[n]][0];
+        }
+        else
+        {  /* non zero coeff */
+            int v, j;
+            if (!VP8GetBit(br, p[2]))
+            {
+                p = prob[kBands[n]][1];
+                v = 1;
+            }
+            else
+            {
+                if (!VP8GetBit(br, p[3]))
+                {
+                    if (!VP8GetBit(br, p[4]))
+                    {
+                        v = 2;
+                    }
+                    else
+                    {
+                        v = 3 + VP8GetBit(br, p[5]);
+                    }
+                }
+                else
+                {
+                    if (!VP8GetBit(br, p[6]))
+                    {
+                        if (!VP8GetBit(br, p[7]))
+                        {
+                            v = 5 + VP8GetBit(br, 159);
+                        } else
+                        {
+                            v = 7 + 2 * VP8GetBit(br, 165);
+                            v += VP8GetBit(br, 145);
+                        }
+                    }
+                    else
+                    {
+                        const uint8_t* tab;
+                        const int bit1 = VP8GetBit(br, p[8]);
+                        const int bit0 = VP8GetBit(br, p[9 + bit1]);
+                        const int cat = 2 * bit1 + bit0;
+                        v = 0;
+                        for (tab = kCat3456[cat]; *tab; ++tab)
+                        {
+                            v += v + VP8GetBit(br, *tab);
+                        }
+                        v += 3 + (8 << cat);
+                    }
+                }
+                p = prob[kBands[n]][2];
+            }
+            j = kZigzag[n - 1];
+
+            out[j] = GetSigned(br, v);
+
+            if (n == 16 || !VP8GetBit(br, p[0]))
+            {   /* EOB */
+                return n;
+            }
+        }
+        if (n == 16)
+        {
+            return 16;
+        }
+    }
+}
+
+int vp8_decode_mb_tokens(VP8D_COMP *dx, MACROBLOCKD *x)
+{
+    BOOL_DECODER *bc = x->current_bc;
+    const FRAME_CONTEXT * const fc = &dx->common.fc;
+    char *eobs = x->eobs;
+
+    int i;
+    int nonzeros;
+    int eobtotal = 0;
+
+    short *qcoeff_ptr;
+    ProbaArray coef_probs;
+    ENTROPY_CONTEXT *a_ctx = ((ENTROPY_CONTEXT *)x->above_context);
+    ENTROPY_CONTEXT *l_ctx = ((ENTROPY_CONTEXT *)x->left_context);
+    ENTROPY_CONTEXT *a;
+    ENTROPY_CONTEXT *l;
+    int skip_dc = 0;
+
+    qcoeff_ptr = &x->qcoeff[0];
+
+    if (!x->mode_info_context->mbmi.is_4x4)
+    {
+        a = a_ctx + 8;
+        l = l_ctx + 8;
+
+        coef_probs = fc->coef_probs [1];
+
+        nonzeros = GetCoeffs(bc, coef_probs, (*a + *l), 0, qcoeff_ptr + 24 * 16);
+        *a = *l = (nonzeros > 0);
+
+        eobs[24] = nonzeros;
+        eobtotal += nonzeros - 16;
+
+        coef_probs = fc->coef_probs [0];
+        skip_dc = 1;
+    }
+    else
+    {
+        coef_probs = fc->coef_probs [3];
+        skip_dc = 0;
+    }
+
+    for (i = 0; i < 16; ++i)
+    {
+        a = a_ctx + (i&3);
+        l = l_ctx + ((i&0xc)>>2);
+
+        nonzeros = GetCoeffs(bc, coef_probs, (*a + *l), skip_dc, qcoeff_ptr);
+        *a = *l = (nonzeros > 0);
+
+        nonzeros += skip_dc;
+        eobs[i] = nonzeros;
+        eobtotal += nonzeros;
+        qcoeff_ptr += 16;
+    }
+
+    coef_probs = fc->coef_probs [2];
+
+    a_ctx += 4;
+    l_ctx += 4;
+    for (i = 16; i < 24; ++i)
+    {
+        a = a_ctx + ((i > 19)<<1) + (i&1);
+        l = l_ctx + ((i > 19)<<1) + ((i&3)>1);
+
+        nonzeros = GetCoeffs(bc, coef_probs, (*a + *l), 0, qcoeff_ptr);
+        *a = *l = (nonzeros > 0);
+
+        eobs[i] = nonzeros;
+        eobtotal += nonzeros;
+        qcoeff_ptr += 16;
+    }
+
+    return eobtotal;
+}
+
diff --git a/libvpx/libvpx/vp8/decoder/detokenize.h b/libvpx/libvpx/vp8/decoder/detokenize.h
new file mode 100644
index 0000000..f0b1254
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/detokenize.h
@@ -0,0 +1,27 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_DECODER_DETOKENIZE_H_
+#define VP8_DECODER_DETOKENIZE_H_
+
+#include "onyxd_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_reset_mb_tokens_context(MACROBLOCKD *x);
+int vp8_decode_mb_tokens(VP8D_COMP *, MACROBLOCKD *);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_DECODER_DETOKENIZE_H_
diff --git a/libvpx/libvpx/vp8/decoder/ec_types.h b/libvpx/libvpx/vp8/decoder/ec_types.h
new file mode 100644
index 0000000..3af5ca8
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/ec_types.h
@@ -0,0 +1,58 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_DECODER_EC_TYPES_H_
+#define VP8_DECODER_EC_TYPES_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_OVERLAPS 16
+
+
+/* The area (pixel area in Q6) the block pointed to by bmi overlaps
+ * another block with.
+ */
+typedef struct
+{
+    int overlap;
+    union b_mode_info *bmi;
+} OVERLAP_NODE;
+
+/* Structure to keep track of overlapping blocks on a block level. */
+typedef struct
+{
+    /* TODO(holmer): This array should be exchanged for a linked list */
+    OVERLAP_NODE overlaps[MAX_OVERLAPS];
+} B_OVERLAP;
+
+/* Structure used to hold all the overlaps of a macroblock. The overlaps of a
+ * macroblock is further divided into block overlaps.
+ */
+typedef struct
+{
+    B_OVERLAP overlaps[16];
+} MB_OVERLAP;
+
+/* Structure for keeping track of motion vectors and which reference frame they
+ * refer to. Used for motion vector interpolation.
+ */
+typedef struct
+{
+    MV mv;
+    MV_REFERENCE_FRAME ref_frame;
+} EC_BLOCK;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_DECODER_EC_TYPES_H_
diff --git a/libvpx/libvpx/vp8/decoder/error_concealment.c b/libvpx/libvpx/vp8/decoder/error_concealment.c
new file mode 100644
index 0000000..a73813f
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/error_concealment.c
@@ -0,0 +1,586 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "error_concealment.h"
+#include "onyxd_int.h"
+#include "decodemv.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp8/common/findnearmv.h"
+#include "vp8/common/common.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+#define FLOOR(x,q) ((x) & -(1 << (q)))
+
+#define NUM_NEIGHBORS 20
+
+typedef struct ec_position
+{
+    int row;
+    int col;
+} EC_POS;
+
+/*
+ * Regenerate the table in Matlab with:
+ * x = meshgrid((1:4), (1:4));
+ * y = meshgrid((1:4), (1:4))';
+ * W = round((1./(sqrt(x.^2 + y.^2))*2^7));
+ * W(1,1) = 0;
+ */
+static const int weights_q7[5][5] = {
+       {  0,   128,    64,    43,    32 },
+       {128,    91,    57,    40,    31 },
+       { 64,    57,    45,    36,    29 },
+       { 43,    40,    36,    30,    26 },
+       { 32,    31,    29,    26,    23 }
+};
+
+int vp8_alloc_overlap_lists(VP8D_COMP *pbi)
+{
+    if (pbi->overlaps != NULL)
+    {
+        vpx_free(pbi->overlaps);
+        pbi->overlaps = NULL;
+    }
+
+    pbi->overlaps = vpx_calloc(pbi->common.mb_rows * pbi->common.mb_cols,
+                               sizeof(MB_OVERLAP));
+
+    if (pbi->overlaps == NULL)
+        return -1;
+
+    return 0;
+}
+
+void vp8_de_alloc_overlap_lists(VP8D_COMP *pbi)
+{
+    vpx_free(pbi->overlaps);
+    pbi->overlaps = NULL;
+}
+
+/* Inserts a new overlap area value to the list of overlaps of a block */
+static void assign_overlap(OVERLAP_NODE* overlaps,
+                           union b_mode_info *bmi,
+                           int overlap)
+{
+    int i;
+    if (overlap <= 0)
+        return;
+    /* Find and assign to the next empty overlap node in the list of overlaps.
+     * Empty is defined as bmi == NULL */
+    for (i = 0; i < MAX_OVERLAPS; i++)
+    {
+        if (overlaps[i].bmi == NULL)
+        {
+            overlaps[i].bmi = bmi;
+            overlaps[i].overlap = overlap;
+            break;
+        }
+    }
+}
+
+/* Calculates the overlap area between two 4x4 squares, where the first
+ * square has its upper-left corner at (b1_row, b1_col) and the second
+ * square has its upper-left corner at (b2_row, b2_col). Doesn't
+ * properly handle squares which do not overlap.
+ */
+static int block_overlap(int b1_row, int b1_col, int b2_row, int b2_col)
+{
+    const int int_top = VPXMAX(b1_row, b2_row); // top
+    const int int_left = VPXMAX(b1_col, b2_col); // left
+    /* Since each block is 4x4 pixels, adding 4 (Q3) to the left/top edge
+     * gives us the right/bottom edge.
+     */
+    const int int_right = VPXMIN(b1_col + (4<<3), b2_col + (4<<3)); // right
+    const int int_bottom = VPXMIN(b1_row + (4<<3), b2_row + (4<<3)); // bottom
+    return (int_bottom - int_top) * (int_right - int_left);
+}
+
+/* Calculates the overlap area for all blocks in a macroblock at position
+ * (mb_row, mb_col) in macroblocks, which are being overlapped by a given
+ * overlapping block at position (new_row, new_col) (in pixels, Q3). The
+ * first block being overlapped in the macroblock has position (first_blk_row,
+ * first_blk_col) in blocks relative the upper-left corner of the image.
+ */
+static void calculate_overlaps_mb(B_OVERLAP *b_overlaps, union b_mode_info *bmi,
+                                  int new_row, int new_col,
+                                  int mb_row, int mb_col,
+                                  int first_blk_row, int first_blk_col)
+{
+    /* Find the blocks within this MB (defined by mb_row, mb_col) which are
+     * overlapped by bmi and calculate and assign overlap for each of those
+     * blocks. */
+
+    /* Block coordinates relative the upper-left block */
+    const int rel_ol_blk_row = first_blk_row - mb_row * 4;
+    const int rel_ol_blk_col = first_blk_col - mb_col * 4;
+    /* If the block partly overlaps any previous MB, these coordinates
+     * can be < 0. We don't want to access blocks in previous MBs.
+     */
+    const int blk_idx = VPXMAX(rel_ol_blk_row,0) * 4 + VPXMAX(rel_ol_blk_col,0);
+    /* Upper left overlapping block */
+    B_OVERLAP *b_ol_ul = &(b_overlaps[blk_idx]);
+
+    /* Calculate and assign overlaps for all blocks in this MB
+     * which the motion compensated block overlaps
+     */
+    /* Avoid calculating overlaps for blocks in later MBs */
+    int end_row = VPXMIN(4 + mb_row * 4 - first_blk_row, 2);
+    int end_col = VPXMIN(4 + mb_col * 4 - first_blk_col, 2);
+    int row, col;
+
+    /* Check if new_row and new_col are evenly divisible by 4 (Q3),
+     * and if so we shouldn't check neighboring blocks
+     */
+    if (new_row >= 0 && (new_row & 0x1F) == 0)
+        end_row = 1;
+    if (new_col >= 0 && (new_col & 0x1F) == 0)
+        end_col = 1;
+
+    /* Check if the overlapping block partly overlaps a previous MB
+     * and if so, we're overlapping fewer blocks in this MB.
+     */
+    if (new_row < (mb_row*16)<<3)
+        end_row = 1;
+    if (new_col < (mb_col*16)<<3)
+        end_col = 1;
+
+    for (row = 0; row < end_row; ++row)
+    {
+        for (col = 0; col < end_col; ++col)
+        {
+            /* input in Q3, result in Q6 */
+            const int overlap = block_overlap(new_row, new_col,
+                                                  (((first_blk_row + row) *
+                                                      4) << 3),
+                                                  (((first_blk_col + col) *
+                                                      4) << 3));
+            assign_overlap(b_ol_ul[row * 4 + col].overlaps, bmi, overlap);
+        }
+    }
+}
+
+void vp8_calculate_overlaps(MB_OVERLAP *overlap_ul,
+                            int mb_rows, int mb_cols,
+                            union b_mode_info *bmi,
+                            int b_row, int b_col)
+{
+    MB_OVERLAP *mb_overlap;
+    int row, col, rel_row, rel_col;
+    int new_row, new_col;
+    int end_row, end_col;
+    int overlap_b_row, overlap_b_col;
+    int overlap_mb_row, overlap_mb_col;
+
+    /* mb subpixel position */
+    row = (4 * b_row) << 3; /* Q3 */
+    col = (4 * b_col) << 3; /* Q3 */
+
+    /* reverse compensate for motion */
+    new_row = row - bmi->mv.as_mv.row;
+    new_col = col - bmi->mv.as_mv.col;
+
+    if (new_row >= ((16*mb_rows) << 3) || new_col >= ((16*mb_cols) << 3))
+    {
+        /* the new block ended up outside the frame */
+        return;
+    }
+
+    if (new_row <= -32 || new_col <= -32)
+    {
+        /* outside the frame */
+        return;
+    }
+    /* overlapping block's position in blocks */
+    overlap_b_row = FLOOR(new_row / 4, 3) >> 3;
+    overlap_b_col = FLOOR(new_col / 4, 3) >> 3;
+
+    /* overlapping block's MB position in MBs
+     * operations are done in Q3
+     */
+    overlap_mb_row = FLOOR((overlap_b_row << 3) / 4, 3) >> 3;
+    overlap_mb_col = FLOOR((overlap_b_col << 3) / 4, 3) >> 3;
+
+    end_row = VPXMIN(mb_rows - overlap_mb_row, 2);
+    end_col = VPXMIN(mb_cols - overlap_mb_col, 2);
+
+    /* Don't calculate overlap for MBs we don't overlap */
+    /* Check if the new block row starts at the last block row of the MB */
+    if (abs(new_row - ((16*overlap_mb_row) << 3)) < ((3*4) << 3))
+        end_row = 1;
+    /* Check if the new block col starts at the last block col of the MB */
+    if (abs(new_col - ((16*overlap_mb_col) << 3)) < ((3*4) << 3))
+        end_col = 1;
+
+    /* find the MB(s) this block is overlapping */
+    for (rel_row = 0; rel_row < end_row; ++rel_row)
+    {
+        for (rel_col = 0; rel_col < end_col; ++rel_col)
+        {
+            if (overlap_mb_row + rel_row < 0 ||
+                overlap_mb_col + rel_col < 0)
+                continue;
+            mb_overlap = overlap_ul + (overlap_mb_row + rel_row) * mb_cols +
+                 overlap_mb_col + rel_col;
+
+            calculate_overlaps_mb(mb_overlap->overlaps, bmi,
+                                  new_row, new_col,
+                                  overlap_mb_row + rel_row,
+                                  overlap_mb_col + rel_col,
+                                  overlap_b_row + rel_row,
+                                  overlap_b_col + rel_col);
+        }
+    }
+}
+
+/* Estimates a motion vector given the overlapping blocks' motion vectors.
+ * Filters out all overlapping blocks which do not refer to the correct
+ * reference frame type.
+ */
+static void estimate_mv(const OVERLAP_NODE *overlaps, union b_mode_info *bmi)
+{
+    int i;
+    int overlap_sum = 0;
+    int row_acc = 0;
+    int col_acc = 0;
+
+    bmi->mv.as_int = 0;
+    for (i=0; i < MAX_OVERLAPS; ++i)
+    {
+        if (overlaps[i].bmi == NULL)
+            break;
+        col_acc += overlaps[i].overlap * overlaps[i].bmi->mv.as_mv.col;
+        row_acc += overlaps[i].overlap * overlaps[i].bmi->mv.as_mv.row;
+        overlap_sum += overlaps[i].overlap;
+    }
+    if (overlap_sum > 0)
+    {
+        /* Q9 / Q6 = Q3 */
+        bmi->mv.as_mv.col = col_acc / overlap_sum;
+        bmi->mv.as_mv.row = row_acc / overlap_sum;
+    }
+    else
+    {
+        bmi->mv.as_mv.col = 0;
+        bmi->mv.as_mv.row = 0;
+    }
+}
+
+/* Estimates all motion vectors for a macroblock given the lists of
+ * overlaps for each block. Decides whether or not the MVs must be clamped.
+ */
+static void estimate_mb_mvs(const B_OVERLAP *block_overlaps,
+                            MODE_INFO *mi,
+                            int mb_to_left_edge,
+                            int mb_to_right_edge,
+                            int mb_to_top_edge,
+                            int mb_to_bottom_edge)
+{
+    int row, col;
+    int non_zero_count = 0;
+    MV * const filtered_mv = &(mi->mbmi.mv.as_mv);
+    union b_mode_info * const bmi = mi->bmi;
+    filtered_mv->col = 0;
+    filtered_mv->row = 0;
+    mi->mbmi.need_to_clamp_mvs = 0;
+    for (row = 0; row < 4; ++row)
+    {
+        int this_b_to_top_edge = mb_to_top_edge + ((row*4)<<3);
+        int this_b_to_bottom_edge = mb_to_bottom_edge - ((row*4)<<3);
+        for (col = 0; col < 4; ++col)
+        {
+            int i = row * 4 + col;
+            int this_b_to_left_edge = mb_to_left_edge + ((col*4)<<3);
+            int this_b_to_right_edge = mb_to_right_edge - ((col*4)<<3);
+            /* Estimate vectors for all blocks which are overlapped by this */
+            /* type. Interpolate/extrapolate the rest of the block's MVs */
+            estimate_mv(block_overlaps[i].overlaps, &(bmi[i]));
+            mi->mbmi.need_to_clamp_mvs |= vp8_check_mv_bounds(
+                                                         &bmi[i].mv,
+                                                         this_b_to_left_edge,
+                                                         this_b_to_right_edge,
+                                                         this_b_to_top_edge,
+                                                         this_b_to_bottom_edge);
+            if (bmi[i].mv.as_int != 0)
+            {
+                ++non_zero_count;
+                filtered_mv->col += bmi[i].mv.as_mv.col;
+                filtered_mv->row += bmi[i].mv.as_mv.row;
+            }
+        }
+    }
+    if (non_zero_count > 0)
+    {
+        filtered_mv->col /= non_zero_count;
+        filtered_mv->row /= non_zero_count;
+    }
+}
+
+static void calc_prev_mb_overlaps(MB_OVERLAP *overlaps, MODE_INFO *prev_mi,
+                                    int mb_row, int mb_col,
+                                    int mb_rows, int mb_cols)
+{
+    int sub_row;
+    int sub_col;
+    for (sub_row = 0; sub_row < 4; ++sub_row)
+    {
+        for (sub_col = 0; sub_col < 4; ++sub_col)
+        {
+            vp8_calculate_overlaps(
+                                overlaps, mb_rows, mb_cols,
+                                &(prev_mi->bmi[sub_row * 4 + sub_col]),
+                                4 * mb_row + sub_row,
+                                4 * mb_col + sub_col);
+        }
+    }
+}
+
+/* Estimate all missing motion vectors. This function does the same as the one
+ * above, but has different input arguments. */
+static void estimate_missing_mvs(MB_OVERLAP *overlaps,
+                                 MODE_INFO *mi, MODE_INFO *prev_mi,
+                                 int mb_rows, int mb_cols,
+                                 unsigned int first_corrupt)
+{
+    int mb_row, mb_col;
+    memset(overlaps, 0, sizeof(MB_OVERLAP) * mb_rows * mb_cols);
+    /* First calculate the overlaps for all blocks */
+    for (mb_row = 0; mb_row < mb_rows; ++mb_row)
+    {
+        for (mb_col = 0; mb_col < mb_cols; ++mb_col)
+        {
+            /* We're only able to use blocks referring to the last frame
+             * when extrapolating new vectors.
+             */
+            if (prev_mi->mbmi.ref_frame == LAST_FRAME)
+            {
+                calc_prev_mb_overlaps(overlaps, prev_mi,
+                                      mb_row, mb_col,
+                                      mb_rows, mb_cols);
+            }
+            ++prev_mi;
+        }
+        ++prev_mi;
+    }
+
+    mb_row = first_corrupt / mb_cols;
+    mb_col = first_corrupt - mb_row * mb_cols;
+    mi += mb_row*(mb_cols + 1) + mb_col;
+    /* Go through all macroblocks in the current image with missing MVs
+     * and calculate new MVs using the overlaps.
+     */
+    for (; mb_row < mb_rows; ++mb_row)
+    {
+        int mb_to_top_edge = -((mb_row * 16)) << 3;
+        int mb_to_bottom_edge = ((mb_rows - 1 - mb_row) * 16) << 3;
+        for (; mb_col < mb_cols; ++mb_col)
+        {
+            int mb_to_left_edge = -((mb_col * 16) << 3);
+            int mb_to_right_edge = ((mb_cols - 1 - mb_col) * 16) << 3;
+            const B_OVERLAP *block_overlaps =
+                    overlaps[mb_row*mb_cols + mb_col].overlaps;
+            mi->mbmi.ref_frame = LAST_FRAME;
+            mi->mbmi.mode = SPLITMV;
+            mi->mbmi.uv_mode = DC_PRED;
+            mi->mbmi.partitioning = 3;
+            mi->mbmi.segment_id = 0;
+            estimate_mb_mvs(block_overlaps,
+                            mi,
+                            mb_to_left_edge,
+                            mb_to_right_edge,
+                            mb_to_top_edge,
+                            mb_to_bottom_edge);
+            ++mi;
+        }
+        mb_col = 0;
+        ++mi;
+    }
+}
+
+void vp8_estimate_missing_mvs(VP8D_COMP *pbi)
+{
+    VP8_COMMON * const pc = &pbi->common;
+    estimate_missing_mvs(pbi->overlaps,
+                         pc->mi, pc->prev_mi,
+                         pc->mb_rows, pc->mb_cols,
+                         pbi->mvs_corrupt_from_mb);
+}
+
+static void assign_neighbor(EC_BLOCK *neighbor, MODE_INFO *mi, int block_idx)
+{
+    assert(mi->mbmi.ref_frame < MAX_REF_FRAMES);
+    neighbor->ref_frame = mi->mbmi.ref_frame;
+    neighbor->mv = mi->bmi[block_idx].mv.as_mv;
+}
+
+/* Finds the neighboring blocks of a macroblocks. In the general case
+ * 20 blocks are found. If a fewer number of blocks are found due to
+ * image boundaries, those positions in the EC_BLOCK array are left "empty".
+ * The neighbors are enumerated with the upper-left neighbor as the first
+ * element, the second element refers to the neighbor to right of the previous
+ * neighbor, and so on. The last element refers to the neighbor below the first
+ * neighbor.
+ */
+static void find_neighboring_blocks(MODE_INFO *mi,
+                                    EC_BLOCK *neighbors,
+                                    int mb_row, int mb_col,
+                                    int mb_rows, int mb_cols,
+                                    int mi_stride)
+{
+    int i = 0;
+    int j;
+    if (mb_row > 0)
+    {
+        /* upper left */
+        if (mb_col > 0)
+            assign_neighbor(&neighbors[i], mi - mi_stride - 1, 15);
+        ++i;
+        /* above */
+        for (j = 12; j < 16; ++j, ++i)
+            assign_neighbor(&neighbors[i], mi - mi_stride, j);
+    }
+    else
+        i += 5;
+    if (mb_col < mb_cols - 1)
+    {
+        /* upper right */
+        if (mb_row > 0)
+            assign_neighbor(&neighbors[i], mi - mi_stride + 1, 12);
+        ++i;
+        /* right */
+        for (j = 0; j <= 12; j += 4, ++i)
+            assign_neighbor(&neighbors[i], mi + 1, j);
+    }
+    else
+        i += 5;
+    if (mb_row < mb_rows - 1)
+    {
+        /* lower right */
+        if (mb_col < mb_cols - 1)
+            assign_neighbor(&neighbors[i], mi + mi_stride + 1, 0);
+        ++i;
+        /* below */
+        for (j = 0; j < 4; ++j, ++i)
+            assign_neighbor(&neighbors[i], mi + mi_stride, j);
+    }
+    else
+        i += 5;
+    if (mb_col > 0)
+    {
+        /* lower left */
+        if (mb_row < mb_rows - 1)
+            assign_neighbor(&neighbors[i], mi + mi_stride - 1, 4);
+        ++i;
+        /* left */
+        for (j = 3; j < 16; j += 4, ++i)
+        {
+            assign_neighbor(&neighbors[i], mi - 1, j);
+        }
+    }
+    else
+        i += 5;
+    assert(i == 20);
+}
+
+/* Interpolates all motion vectors for a macroblock from the neighboring blocks'
+ * motion vectors.
+ */
+static void interpolate_mvs(MACROBLOCKD *mb,
+                         EC_BLOCK *neighbors,
+                         MV_REFERENCE_FRAME dom_ref_frame)
+{
+    int row, col, i;
+    MODE_INFO * const mi = mb->mode_info_context;
+    /* Table with the position of the neighboring blocks relative the position
+     * of the upper left block of the current MB. Starting with the upper left
+     * neighbor and going to the right.
+     */
+    const EC_POS neigh_pos[NUM_NEIGHBORS] = {
+                                        {-1,-1}, {-1,0}, {-1,1}, {-1,2}, {-1,3},
+                                        {-1,4}, {0,4}, {1,4}, {2,4}, {3,4},
+                                        {4,4}, {4,3}, {4,2}, {4,1}, {4,0},
+                                        {4,-1}, {3,-1}, {2,-1}, {1,-1}, {0,-1}
+                                      };
+    mi->mbmi.need_to_clamp_mvs = 0;
+    for (row = 0; row < 4; ++row)
+    {
+        int mb_to_top_edge = mb->mb_to_top_edge + ((row*4)<<3);
+        int mb_to_bottom_edge = mb->mb_to_bottom_edge - ((row*4)<<3);
+        for (col = 0; col < 4; ++col)
+        {
+            int mb_to_left_edge = mb->mb_to_left_edge + ((col*4)<<3);
+            int mb_to_right_edge = mb->mb_to_right_edge - ((col*4)<<3);
+            int w_sum = 0;
+            int mv_row_sum = 0;
+            int mv_col_sum = 0;
+            int_mv * const mv = &(mi->bmi[row*4 + col].mv);
+            mv->as_int = 0;
+            for (i = 0; i < NUM_NEIGHBORS; ++i)
+            {
+                /* Calculate the weighted sum of neighboring MVs referring
+                 * to the dominant frame type.
+                 */
+                const int w = weights_q7[abs(row - neigh_pos[i].row)]
+                                        [abs(col - neigh_pos[i].col)];
+                if (neighbors[i].ref_frame != dom_ref_frame)
+                    continue;
+                w_sum += w;
+                /* Q7 * Q3 = Q10 */
+                mv_row_sum += w*neighbors[i].mv.row;
+                mv_col_sum += w*neighbors[i].mv.col;
+            }
+            if (w_sum > 0)
+            {
+                /* Avoid division by zero.
+                 * Normalize with the sum of the coefficients
+                 * Q3 = Q10 / Q7
+                 */
+                mv->as_mv.row = mv_row_sum / w_sum;
+                mv->as_mv.col = mv_col_sum / w_sum;
+                mi->mbmi.need_to_clamp_mvs |= vp8_check_mv_bounds(
+                                                            mv,
+                                                            mb_to_left_edge,
+                                                            mb_to_right_edge,
+                                                            mb_to_top_edge,
+                                                            mb_to_bottom_edge);
+            }
+        }
+    }
+}
+
+void vp8_interpolate_motion(MACROBLOCKD *mb,
+                        int mb_row, int mb_col,
+                        int mb_rows, int mb_cols)
+{
+    /* Find relevant neighboring blocks */
+    EC_BLOCK neighbors[NUM_NEIGHBORS];
+    int i;
+    /* Initialize the array. MAX_REF_FRAMES is interpreted as "doesn't exist" */
+    for (i = 0; i < NUM_NEIGHBORS; ++i)
+    {
+        neighbors[i].ref_frame = MAX_REF_FRAMES;
+        neighbors[i].mv.row = neighbors[i].mv.col = 0;
+    }
+    find_neighboring_blocks(mb->mode_info_context,
+                                neighbors,
+                                mb_row, mb_col,
+                                mb_rows, mb_cols,
+                                mb->mode_info_stride);
+    /* Interpolate MVs for the missing blocks from the surrounding
+     * blocks which refer to the last frame. */
+    interpolate_mvs(mb, neighbors, LAST_FRAME);
+
+    mb->mode_info_context->mbmi.ref_frame = LAST_FRAME;
+    mb->mode_info_context->mbmi.mode = SPLITMV;
+    mb->mode_info_context->mbmi.uv_mode = DC_PRED;
+    mb->mode_info_context->mbmi.partitioning = 3;
+    mb->mode_info_context->mbmi.segment_id = 0;
+}
diff --git a/libvpx/libvpx/vp8/decoder/error_concealment.h b/libvpx/libvpx/vp8/decoder/error_concealment.h
new file mode 100644
index 0000000..b6b4972
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/error_concealment.h
@@ -0,0 +1,43 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_DECODER_ERROR_CONCEALMENT_H_
+#define VP8_DECODER_ERROR_CONCEALMENT_H_
+
+#include "onyxd_int.h"
+#include "ec_types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Allocate memory for the overlap lists */
+int vp8_alloc_overlap_lists(VP8D_COMP *pbi);
+
+/* Deallocate the overlap lists */
+void vp8_de_alloc_overlap_lists(VP8D_COMP *pbi);
+
+/* Estimate all missing motion vectors. */
+void vp8_estimate_missing_mvs(VP8D_COMP *pbi);
+
+/* Functions for spatial MV interpolation */
+
+/* Interpolates all motion vectors for a macroblock mb at position
+ * (mb_row, mb_col). */
+void vp8_interpolate_motion(MACROBLOCKD *mb,
+                            int mb_row, int mb_col,
+                            int mb_rows, int mb_cols);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_DECODER_ERROR_CONCEALMENT_H_
diff --git a/libvpx/libvpx/vp8/decoder/onyxd_if.c b/libvpx/libvpx/vp8/decoder/onyxd_if.c
new file mode 100644
index 0000000..3468268
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/onyxd_if.c
@@ -0,0 +1,521 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vp8/common/onyxc_int.h"
+#if CONFIG_POSTPROC
+#include "vp8/common/postproc.h"
+#endif
+#include "vp8/common/onyxd.h"
+#include "onyxd_int.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp8/common/alloccommon.h"
+#include "vp8/common/loopfilter.h"
+#include "vp8/common/swapyv12buffer.h"
+#include "vp8/common/threading.h"
+#include "decoderthreading.h"
+#include <stdio.h>
+#include <assert.h>
+
+#include "vp8/common/quant_common.h"
+#include "vp8/common/reconintra.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vp8/common/systemdependent.h"
+#include "vpx_ports/vpx_once.h"
+#include "vpx_ports/vpx_timer.h"
+#include "detokenize.h"
+#if CONFIG_ERROR_CONCEALMENT
+#include "error_concealment.h"
+#endif
+#if ARCH_ARM
+#include "vpx_ports/arm.h"
+#endif
+
+extern void vp8_init_loop_filter(VP8_COMMON *cm);
+extern void vp8cx_init_de_quantizer(VP8D_COMP *pbi);
+static int get_free_fb (VP8_COMMON *cm);
+static void ref_cnt_fb (int *buf, int *idx, int new_idx);
+
+static void initialize_dec(void) {
+    static volatile int init_done = 0;
+
+    if (!init_done)
+    {
+        vpx_dsp_rtcd();
+        vp8_init_intra_predictors();
+        init_done = 1;
+    }
+}
+
+static void remove_decompressor(VP8D_COMP *pbi)
+{
+#if CONFIG_ERROR_CONCEALMENT
+    vp8_de_alloc_overlap_lists(pbi);
+#endif
+    vp8_remove_common(&pbi->common);
+    vpx_free(pbi);
+}
+
+static struct VP8D_COMP * create_decompressor(VP8D_CONFIG *oxcf)
+{
+    VP8D_COMP *pbi = vpx_memalign(32, sizeof(VP8D_COMP));
+
+    if (!pbi)
+        return NULL;
+
+    memset(pbi, 0, sizeof(VP8D_COMP));
+
+    if (setjmp(pbi->common.error.jmp))
+    {
+        pbi->common.error.setjmp = 0;
+        remove_decompressor(pbi);
+        return 0;
+    }
+
+    pbi->common.error.setjmp = 1;
+
+    vp8_create_common(&pbi->common);
+
+    pbi->common.current_video_frame = 0;
+    pbi->ready_for_new_data = 1;
+
+    /* vp8cx_init_de_quantizer() is first called here. Add check in frame_init_dequantizer() to avoid
+     *  unnecessary calling of vp8cx_init_de_quantizer() for every frame.
+     */
+    vp8cx_init_de_quantizer(pbi);
+
+    vp8_loop_filter_init(&pbi->common);
+
+    pbi->common.error.setjmp = 0;
+
+#if CONFIG_ERROR_CONCEALMENT
+    pbi->ec_enabled = oxcf->error_concealment;
+    pbi->overlaps = NULL;
+#else
+    (void)oxcf;
+    pbi->ec_enabled = 0;
+#endif
+    /* Error concealment is activated after a key frame has been
+     * decoded without errors when error concealment is enabled.
+     */
+    pbi->ec_active = 0;
+
+    pbi->decoded_key_frame = 0;
+
+    /* Independent partitions is activated when a frame updates the
+     * token probability table to have equal probabilities over the
+     * PREV_COEF context.
+     */
+    pbi->independent_partitions = 0;
+
+    vp8_setup_block_dptrs(&pbi->mb);
+
+    once(initialize_dec);
+
+    return pbi;
+}
+
+vpx_codec_err_t vp8dx_get_reference(VP8D_COMP *pbi, enum vpx_ref_frame_type ref_frame_flag, YV12_BUFFER_CONFIG *sd)
+{
+    VP8_COMMON *cm = &pbi->common;
+    int ref_fb_idx;
+
+    if (ref_frame_flag == VP8_LAST_FRAME)
+        ref_fb_idx = cm->lst_fb_idx;
+    else if (ref_frame_flag == VP8_GOLD_FRAME)
+        ref_fb_idx = cm->gld_fb_idx;
+    else if (ref_frame_flag == VP8_ALTR_FRAME)
+        ref_fb_idx = cm->alt_fb_idx;
+    else{
+        vpx_internal_error(&pbi->common.error, VPX_CODEC_ERROR,
+            "Invalid reference frame");
+        return pbi->common.error.error_code;
+    }
+
+    if(cm->yv12_fb[ref_fb_idx].y_height != sd->y_height ||
+        cm->yv12_fb[ref_fb_idx].y_width != sd->y_width ||
+        cm->yv12_fb[ref_fb_idx].uv_height != sd->uv_height ||
+        cm->yv12_fb[ref_fb_idx].uv_width != sd->uv_width){
+        vpx_internal_error(&pbi->common.error, VPX_CODEC_ERROR,
+            "Incorrect buffer dimensions");
+    }
+    else
+        vp8_yv12_copy_frame(&cm->yv12_fb[ref_fb_idx], sd);
+
+    return pbi->common.error.error_code;
+}
+
+
+vpx_codec_err_t vp8dx_set_reference(VP8D_COMP *pbi, enum vpx_ref_frame_type ref_frame_flag, YV12_BUFFER_CONFIG *sd)
+{
+    VP8_COMMON *cm = &pbi->common;
+    int *ref_fb_ptr = NULL;
+    int free_fb;
+
+    if (ref_frame_flag == VP8_LAST_FRAME)
+        ref_fb_ptr = &cm->lst_fb_idx;
+    else if (ref_frame_flag == VP8_GOLD_FRAME)
+        ref_fb_ptr = &cm->gld_fb_idx;
+    else if (ref_frame_flag == VP8_ALTR_FRAME)
+        ref_fb_ptr = &cm->alt_fb_idx;
+    else{
+        vpx_internal_error(&pbi->common.error, VPX_CODEC_ERROR,
+            "Invalid reference frame");
+        return pbi->common.error.error_code;
+    }
+
+    if(cm->yv12_fb[*ref_fb_ptr].y_height != sd->y_height ||
+        cm->yv12_fb[*ref_fb_ptr].y_width != sd->y_width ||
+        cm->yv12_fb[*ref_fb_ptr].uv_height != sd->uv_height ||
+        cm->yv12_fb[*ref_fb_ptr].uv_width != sd->uv_width){
+        vpx_internal_error(&pbi->common.error, VPX_CODEC_ERROR,
+            "Incorrect buffer dimensions");
+    }
+    else{
+        /* Find an empty frame buffer. */
+        free_fb = get_free_fb(cm);
+        /* Decrease fb_idx_ref_cnt since it will be increased again in
+         * ref_cnt_fb() below. */
+        cm->fb_idx_ref_cnt[free_fb]--;
+
+        /* Manage the reference counters and copy image. */
+        ref_cnt_fb (cm->fb_idx_ref_cnt, ref_fb_ptr, free_fb);
+        vp8_yv12_copy_frame(sd, &cm->yv12_fb[*ref_fb_ptr]);
+    }
+
+   return pbi->common.error.error_code;
+}
+
+static int get_free_fb (VP8_COMMON *cm)
+{
+    int i;
+    for (i = 0; i < NUM_YV12_BUFFERS; i++)
+        if (cm->fb_idx_ref_cnt[i] == 0)
+            break;
+
+    assert(i < NUM_YV12_BUFFERS);
+    cm->fb_idx_ref_cnt[i] = 1;
+    return i;
+}
+
+static void ref_cnt_fb (int *buf, int *idx, int new_idx)
+{
+    if (buf[*idx] > 0)
+        buf[*idx]--;
+
+    *idx = new_idx;
+
+    buf[new_idx]++;
+}
+
+/* If any buffer copy / swapping is signalled it should be done here. */
+static int swap_frame_buffers (VP8_COMMON *cm)
+{
+    int err = 0;
+
+    /* The alternate reference frame or golden frame can be updated
+     *  using the new, last, or golden/alt ref frame.  If it
+     *  is updated using the newly decoded frame it is a refresh.
+     *  An update using the last or golden/alt ref frame is a copy.
+     */
+    if (cm->copy_buffer_to_arf)
+    {
+        int new_fb = 0;
+
+        if (cm->copy_buffer_to_arf == 1)
+            new_fb = cm->lst_fb_idx;
+        else if (cm->copy_buffer_to_arf == 2)
+            new_fb = cm->gld_fb_idx;
+        else
+            err = -1;
+
+        ref_cnt_fb (cm->fb_idx_ref_cnt, &cm->alt_fb_idx, new_fb);
+    }
+
+    if (cm->copy_buffer_to_gf)
+    {
+        int new_fb = 0;
+
+        if (cm->copy_buffer_to_gf == 1)
+            new_fb = cm->lst_fb_idx;
+        else if (cm->copy_buffer_to_gf == 2)
+            new_fb = cm->alt_fb_idx;
+        else
+            err = -1;
+
+        ref_cnt_fb (cm->fb_idx_ref_cnt, &cm->gld_fb_idx, new_fb);
+    }
+
+    if (cm->refresh_golden_frame)
+        ref_cnt_fb (cm->fb_idx_ref_cnt, &cm->gld_fb_idx, cm->new_fb_idx);
+
+    if (cm->refresh_alt_ref_frame)
+        ref_cnt_fb (cm->fb_idx_ref_cnt, &cm->alt_fb_idx, cm->new_fb_idx);
+
+    if (cm->refresh_last_frame)
+    {
+        ref_cnt_fb (cm->fb_idx_ref_cnt, &cm->lst_fb_idx, cm->new_fb_idx);
+
+        cm->frame_to_show = &cm->yv12_fb[cm->lst_fb_idx];
+    }
+    else
+        cm->frame_to_show = &cm->yv12_fb[cm->new_fb_idx];
+
+    cm->fb_idx_ref_cnt[cm->new_fb_idx]--;
+
+    return err;
+}
+
+static int check_fragments_for_errors(VP8D_COMP *pbi)
+{
+    if (!pbi->ec_active &&
+        pbi->fragments.count <= 1 && pbi->fragments.sizes[0] == 0)
+    {
+        VP8_COMMON *cm = &pbi->common;
+
+        /* If error concealment is disabled we won't signal missing frames
+         * to the decoder.
+         */
+        if (cm->fb_idx_ref_cnt[cm->lst_fb_idx] > 1)
+        {
+            /* The last reference shares buffer with another reference
+             * buffer. Move it to its own buffer before setting it as
+             * corrupt, otherwise we will make multiple buffers corrupt.
+             */
+            const int prev_idx = cm->lst_fb_idx;
+            cm->fb_idx_ref_cnt[prev_idx]--;
+            cm->lst_fb_idx = get_free_fb(cm);
+            vp8_yv12_copy_frame(&cm->yv12_fb[prev_idx],
+                                    &cm->yv12_fb[cm->lst_fb_idx]);
+        }
+        /* This is used to signal that we are missing frames.
+         * We do not know if the missing frame(s) was supposed to update
+         * any of the reference buffers, but we act conservative and
+         * mark only the last buffer as corrupted.
+         */
+        cm->yv12_fb[cm->lst_fb_idx].corrupted = 1;
+
+        /* Signal that we have no frame to show. */
+        cm->show_frame = 0;
+
+        /* Nothing more to do. */
+        return 0;
+    }
+
+    return 1;
+}
+
+int vp8dx_receive_compressed_data(VP8D_COMP *pbi, size_t size,
+                                  const uint8_t *source,
+                                  int64_t time_stamp)
+{
+    VP8_COMMON *cm = &pbi->common;
+    int retcode = -1;
+    (void)size;
+    (void)source;
+
+    pbi->common.error.error_code = VPX_CODEC_OK;
+
+    retcode = check_fragments_for_errors(pbi);
+    if(retcode <= 0)
+        return retcode;
+
+    cm->new_fb_idx = get_free_fb (cm);
+
+    /* setup reference frames for vp8_decode_frame */
+    pbi->dec_fb_ref[INTRA_FRAME]  = &cm->yv12_fb[cm->new_fb_idx];
+    pbi->dec_fb_ref[LAST_FRAME]   = &cm->yv12_fb[cm->lst_fb_idx];
+    pbi->dec_fb_ref[GOLDEN_FRAME] = &cm->yv12_fb[cm->gld_fb_idx];
+    pbi->dec_fb_ref[ALTREF_FRAME] = &cm->yv12_fb[cm->alt_fb_idx];
+
+    if (setjmp(pbi->common.error.jmp))
+    {
+       /* We do not know if the missing frame(s) was supposed to update
+        * any of the reference buffers, but we act conservative and
+        * mark only the last buffer as corrupted.
+        */
+        cm->yv12_fb[cm->lst_fb_idx].corrupted = 1;
+
+        if (cm->fb_idx_ref_cnt[cm->new_fb_idx] > 0)
+          cm->fb_idx_ref_cnt[cm->new_fb_idx]--;
+
+        goto decode_exit;
+    }
+
+    pbi->common.error.setjmp = 1;
+
+    retcode = vp8_decode_frame(pbi);
+
+    if (retcode < 0)
+    {
+        if (cm->fb_idx_ref_cnt[cm->new_fb_idx] > 0)
+          cm->fb_idx_ref_cnt[cm->new_fb_idx]--;
+
+        pbi->common.error.error_code = VPX_CODEC_ERROR;
+        goto decode_exit;
+    }
+
+    if (swap_frame_buffers (cm))
+    {
+        pbi->common.error.error_code = VPX_CODEC_ERROR;
+        goto decode_exit;
+    }
+
+    vp8_clear_system_state();
+
+    if (cm->show_frame)
+    {
+        cm->current_video_frame++;
+        cm->show_frame_mi = cm->mi;
+    }
+
+    #if CONFIG_ERROR_CONCEALMENT
+    /* swap the mode infos to storage for future error concealment */
+    if (pbi->ec_enabled && pbi->common.prev_mi)
+    {
+        MODE_INFO* tmp = pbi->common.prev_mi;
+        int row, col;
+        pbi->common.prev_mi = pbi->common.mi;
+        pbi->common.mi = tmp;
+
+        /* Propagate the segment_ids to the next frame */
+        for (row = 0; row < pbi->common.mb_rows; ++row)
+        {
+            for (col = 0; col < pbi->common.mb_cols; ++col)
+            {
+                const int i = row*pbi->common.mode_info_stride + col;
+                pbi->common.mi[i].mbmi.segment_id =
+                        pbi->common.prev_mi[i].mbmi.segment_id;
+            }
+        }
+    }
+#endif
+
+    pbi->ready_for_new_data = 0;
+    pbi->last_time_stamp = time_stamp;
+
+decode_exit:
+    pbi->common.error.setjmp = 0;
+    vp8_clear_system_state();
+    return retcode;
+}
+int vp8dx_get_raw_frame(VP8D_COMP *pbi, YV12_BUFFER_CONFIG *sd, int64_t *time_stamp, int64_t *time_end_stamp, vp8_ppflags_t *flags)
+{
+    int ret = -1;
+
+    if (pbi->ready_for_new_data == 1)
+        return ret;
+
+    /* ie no raw frame to show!!! */
+    if (pbi->common.show_frame == 0)
+        return ret;
+
+    pbi->ready_for_new_data = 1;
+    *time_stamp = pbi->last_time_stamp;
+    *time_end_stamp = 0;
+
+#if CONFIG_POSTPROC
+    ret = vp8_post_proc_frame(&pbi->common, sd, flags);
+#else
+    (void)flags;
+
+    if (pbi->common.frame_to_show)
+    {
+        *sd = *pbi->common.frame_to_show;
+        sd->y_width = pbi->common.Width;
+        sd->y_height = pbi->common.Height;
+        sd->uv_height = pbi->common.Height / 2;
+        ret = 0;
+    }
+    else
+    {
+        ret = -1;
+    }
+
+#endif /*!CONFIG_POSTPROC*/
+    vp8_clear_system_state();
+    return ret;
+}
+
+
+/* This function as written isn't decoder specific, but the encoder has
+ * much faster ways of computing this, so it's ok for it to live in a
+ * decode specific file.
+ */
+int vp8dx_references_buffer( VP8_COMMON *oci, int ref_frame )
+{
+    const MODE_INFO *mi = oci->mi;
+    int mb_row, mb_col;
+
+    for (mb_row = 0; mb_row < oci->mb_rows; mb_row++)
+    {
+        for (mb_col = 0; mb_col < oci->mb_cols; mb_col++,mi++)
+        {
+            if( mi->mbmi.ref_frame == ref_frame)
+              return 1;
+        }
+        mi++;
+    }
+    return 0;
+
+}
+
+int vp8_create_decoder_instances(struct frame_buffers *fb, VP8D_CONFIG *oxcf)
+{
+    if(!fb->use_frame_threads)
+    {
+        /* decoder instance for single thread mode */
+        fb->pbi[0] = create_decompressor(oxcf);
+        if(!fb->pbi[0])
+            return VPX_CODEC_ERROR;
+
+#if CONFIG_MULTITHREAD
+        /* enable row-based threading only when use_frame_threads
+         * is disabled */
+        fb->pbi[0]->max_threads = oxcf->max_threads;
+        vp8_decoder_create_threads(fb->pbi[0]);
+#endif
+    }
+    else
+    {
+        /* TODO : create frame threads and decoder instances for each
+         * thread here */
+    }
+
+    return VPX_CODEC_OK;
+}
+
+int vp8_remove_decoder_instances(struct frame_buffers *fb)
+{
+    if(!fb->use_frame_threads)
+    {
+        VP8D_COMP *pbi = fb->pbi[0];
+
+        if (!pbi)
+            return VPX_CODEC_ERROR;
+#if CONFIG_MULTITHREAD
+        if (pbi->b_multithreaded_rd)
+            vp8mt_de_alloc_temp_buffers(pbi, pbi->common.mb_rows);
+        vp8_decoder_remove_threads(pbi);
+#endif
+
+        /* decoder instance for single thread mode */
+        remove_decompressor(pbi);
+    }
+    else
+    {
+        /* TODO : remove frame threads and decoder instances for each
+         * thread here */
+    }
+
+    return VPX_CODEC_OK;
+}
diff --git a/libvpx/libvpx/vp8/decoder/onyxd_int.h b/libvpx/libvpx/vp8/decoder/onyxd_int.h
new file mode 100644
index 0000000..313fe01
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/onyxd_int.h
@@ -0,0 +1,161 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_DECODER_ONYXD_INT_H_
+#define VP8_DECODER_ONYXD_INT_H_
+
+#include "vpx_config.h"
+#include "vp8/common/onyxd.h"
+#include "treereader.h"
+#include "vp8/common/onyxc_int.h"
+#include "vp8/common/threading.h"
+
+#if CONFIG_ERROR_CONCEALMENT
+#include "ec_types.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct
+{
+    int ithread;
+    void *ptr1;
+    void *ptr2;
+} DECODETHREAD_DATA;
+
+typedef struct
+{
+    MACROBLOCKD  mbd;
+} MB_ROW_DEC;
+
+
+typedef struct
+{
+    int enabled;
+    unsigned int count;
+    const unsigned char *ptrs[MAX_PARTITIONS];
+    unsigned int sizes[MAX_PARTITIONS];
+} FRAGMENT_DATA;
+
+#define MAX_FB_MT_DEC 32
+
+struct frame_buffers
+{
+    /*
+     * this struct will be populated with frame buffer management
+     * info in future commits. */
+
+    /* enable/disable frame-based threading */
+    int     use_frame_threads;
+
+    /* decoder instances */
+    struct VP8D_COMP *pbi[MAX_FB_MT_DEC];
+
+};
+
+typedef struct VP8D_COMP
+{
+    DECLARE_ALIGNED(16, MACROBLOCKD, mb);
+
+    YV12_BUFFER_CONFIG *dec_fb_ref[NUM_YV12_BUFFERS];
+
+    DECLARE_ALIGNED(16, VP8_COMMON, common);
+
+    /* the last partition will be used for the modes/mvs */
+    vp8_reader mbc[MAX_PARTITIONS];
+
+    VP8D_CONFIG oxcf;
+
+    FRAGMENT_DATA fragments;
+
+#if CONFIG_MULTITHREAD
+    /* variable for threading */
+
+    int b_multithreaded_rd;
+    int max_threads;
+    int current_mb_col_main;
+    unsigned int decoding_thread_count;
+    int allocated_decoding_thread_count;
+
+    int mt_baseline_filter_level[MAX_MB_SEGMENTS];
+    int sync_range;
+    int *mt_current_mb_col;                  /* Each row remembers its already decoded column. */
+    pthread_mutex_t *pmutex;
+    pthread_mutex_t mt_mutex;                /* mutex for b_multithreaded_rd */
+
+    unsigned char **mt_yabove_row;           /* mb_rows x width */
+    unsigned char **mt_uabove_row;
+    unsigned char **mt_vabove_row;
+    unsigned char **mt_yleft_col;            /* mb_rows x 16 */
+    unsigned char **mt_uleft_col;            /* mb_rows x 8 */
+    unsigned char **mt_vleft_col;            /* mb_rows x 8 */
+
+    MB_ROW_DEC           *mb_row_di;
+    DECODETHREAD_DATA    *de_thread_data;
+
+    pthread_t           *h_decoding_thread;
+    sem_t               *h_event_start_decoding;
+    sem_t                h_event_end_decoding;
+    /* end of threading data */
+#endif
+
+    int64_t last_time_stamp;
+    int   ready_for_new_data;
+
+    vp8_prob prob_intra;
+    vp8_prob prob_last;
+    vp8_prob prob_gf;
+    vp8_prob prob_skip_false;
+
+#if CONFIG_ERROR_CONCEALMENT
+    MB_OVERLAP *overlaps;
+    /* the mb num from which modes and mvs (first partition) are corrupt */
+    unsigned int mvs_corrupt_from_mb;
+#endif
+    int ec_enabled;
+    int ec_active;
+    int decoded_key_frame;
+    int independent_partitions;
+    int frame_corrupt_residual;
+
+    vpx_decrypt_cb decrypt_cb;
+    void *decrypt_state;
+} VP8D_COMP;
+
+int vp8_decode_frame(VP8D_COMP *cpi);
+
+int vp8_create_decoder_instances(struct frame_buffers *fb, VP8D_CONFIG *oxcf);
+int vp8_remove_decoder_instances(struct frame_buffers *fb);
+
+#if CONFIG_DEBUG
+#define CHECK_MEM_ERROR(lval,expr) do {\
+        lval = (expr); \
+        if(!lval) \
+            vpx_internal_error(&pbi->common.error, VPX_CODEC_MEM_ERROR,\
+                               "Failed to allocate "#lval" at %s:%d", \
+                               __FILE__,__LINE__);\
+    } while(0)
+#else
+#define CHECK_MEM_ERROR(lval,expr) do {\
+        lval = (expr); \
+        if(!lval) \
+            vpx_internal_error(&pbi->common.error, VPX_CODEC_MEM_ERROR,\
+                               "Failed to allocate "#lval);\
+    } while(0)
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_DECODER_ONYXD_INT_H_
diff --git a/libvpx/libvpx/vp8/decoder/threading.c b/libvpx/libvpx/vp8/decoder/threading.c
new file mode 100644
index 0000000..3c1b838
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/threading.c
@@ -0,0 +1,928 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#if !defined(WIN32) && CONFIG_OS_SUPPORT == 1
+# include <unistd.h>
+#endif
+#include "onyxd_int.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp8/common/threading.h"
+
+#include "vp8/common/loopfilter.h"
+#include "vp8/common/extend.h"
+#include "vpx_ports/vpx_timer.h"
+#include "detokenize.h"
+#include "vp8/common/reconintra4x4.h"
+#include "vp8/common/reconinter.h"
+#include "vp8/common/reconintra.h"
+#include "vp8/common/setupintrarecon.h"
+#if CONFIG_ERROR_CONCEALMENT
+#include "error_concealment.h"
+#endif
+
+#define CALLOC_ARRAY(p, n) CHECK_MEM_ERROR((p), vpx_calloc(sizeof(*(p)), (n)))
+#define CALLOC_ARRAY_ALIGNED(p, n, algn) do {                      \
+  CHECK_MEM_ERROR((p), vpx_memalign((algn), sizeof(*(p)) * (n)));  \
+  memset((p), 0, (n) * sizeof(*(p)));                              \
+} while (0)
+
+
+void vp8_mb_init_dequantizer(VP8D_COMP *pbi, MACROBLOCKD *xd);
+
+static void setup_decoding_thread_data(VP8D_COMP *pbi, MACROBLOCKD *xd, MB_ROW_DEC *mbrd, int count)
+{
+    VP8_COMMON *const pc = & pbi->common;
+    int i;
+
+    for (i = 0; i < count; i++)
+    {
+        MACROBLOCKD *mbd = &mbrd[i].mbd;
+        mbd->subpixel_predict        = xd->subpixel_predict;
+        mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
+        mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
+        mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
+
+        mbd->frame_type = pc->frame_type;
+        mbd->pre = xd->pre;
+        mbd->dst = xd->dst;
+
+        mbd->segmentation_enabled    = xd->segmentation_enabled;
+        mbd->mb_segement_abs_delta     = xd->mb_segement_abs_delta;
+        memcpy(mbd->segment_feature_data, xd->segment_feature_data, sizeof(xd->segment_feature_data));
+
+        /*signed char ref_lf_deltas[MAX_REF_LF_DELTAS];*/
+        memcpy(mbd->ref_lf_deltas, xd->ref_lf_deltas, sizeof(xd->ref_lf_deltas));
+        /*signed char mode_lf_deltas[MAX_MODE_LF_DELTAS];*/
+        memcpy(mbd->mode_lf_deltas, xd->mode_lf_deltas, sizeof(xd->mode_lf_deltas));
+        /*unsigned char mode_ref_lf_delta_enabled;
+        unsigned char mode_ref_lf_delta_update;*/
+        mbd->mode_ref_lf_delta_enabled    = xd->mode_ref_lf_delta_enabled;
+        mbd->mode_ref_lf_delta_update    = xd->mode_ref_lf_delta_update;
+
+        mbd->current_bc = &pbi->mbc[0];
+
+        memcpy(mbd->dequant_y1_dc, xd->dequant_y1_dc, sizeof(xd->dequant_y1_dc));
+        memcpy(mbd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
+        memcpy(mbd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
+        memcpy(mbd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
+
+        mbd->fullpixel_mask = 0xffffffff;
+
+        if (pc->full_pixel)
+            mbd->fullpixel_mask = 0xfffffff8;
+
+    }
+
+    for (i = 0; i < pc->mb_rows; i++)
+        pbi->mt_current_mb_col[i] = -1;
+}
+
+static void mt_decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd,
+                                 unsigned int mb_idx)
+{
+    MB_PREDICTION_MODE mode;
+    int i;
+#if CONFIG_ERROR_CONCEALMENT
+    int corruption_detected = 0;
+#else
+    (void)mb_idx;
+#endif
+
+    if (xd->mode_info_context->mbmi.mb_skip_coeff)
+    {
+        vp8_reset_mb_tokens_context(xd);
+    }
+    else if (!vp8dx_bool_error(xd->current_bc))
+    {
+        int eobtotal;
+        eobtotal = vp8_decode_mb_tokens(pbi, xd);
+
+        /* Special case:  Force the loopfilter to skip when eobtotal is zero */
+        xd->mode_info_context->mbmi.mb_skip_coeff = (eobtotal==0);
+    }
+
+    mode = xd->mode_info_context->mbmi.mode;
+
+    if (xd->segmentation_enabled)
+        vp8_mb_init_dequantizer(pbi, xd);
+
+
+#if CONFIG_ERROR_CONCEALMENT
+
+    if(pbi->ec_active)
+    {
+        int throw_residual;
+        /* When we have independent partitions we can apply residual even
+         * though other partitions within the frame are corrupt.
+         */
+        throw_residual = (!pbi->independent_partitions &&
+                          pbi->frame_corrupt_residual);
+        throw_residual = (throw_residual || vp8dx_bool_error(xd->current_bc));
+
+        if ((mb_idx >= pbi->mvs_corrupt_from_mb || throw_residual))
+        {
+            /* MB with corrupt residuals or corrupt mode/motion vectors.
+             * Better to use the predictor as reconstruction.
+             */
+            pbi->frame_corrupt_residual = 1;
+            memset(xd->qcoeff, 0, sizeof(xd->qcoeff));
+
+            corruption_detected = 1;
+
+            /* force idct to be skipped for B_PRED and use the
+             * prediction only for reconstruction
+             * */
+            memset(xd->eobs, 0, 25);
+        }
+    }
+#endif
+
+    /* do prediction */
+    if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
+    {
+        vp8_build_intra_predictors_mbuv_s(xd,
+                                          xd->recon_above[1],
+                                          xd->recon_above[2],
+                                          xd->recon_left[1],
+                                          xd->recon_left[2],
+                                          xd->recon_left_stride[1],
+                                          xd->dst.u_buffer, xd->dst.v_buffer,
+                                          xd->dst.uv_stride);
+
+        if (mode != B_PRED)
+        {
+            vp8_build_intra_predictors_mby_s(xd,
+                                                 xd->recon_above[0],
+                                                 xd->recon_left[0],
+                                                 xd->recon_left_stride[0],
+                                                 xd->dst.y_buffer,
+                                                 xd->dst.y_stride);
+        }
+        else
+        {
+            short *DQC = xd->dequant_y1;
+            int dst_stride = xd->dst.y_stride;
+
+            /* clear out residual eob info */
+            if(xd->mode_info_context->mbmi.mb_skip_coeff)
+                memset(xd->eobs, 0, 25);
+
+            intra_prediction_down_copy(xd, xd->recon_above[0] + 16);
+
+            for (i = 0; i < 16; i++)
+            {
+                BLOCKD *b = &xd->block[i];
+                unsigned char *dst = xd->dst.y_buffer + b->offset;
+                B_PREDICTION_MODE b_mode =
+                    xd->mode_info_context->bmi[i].as_mode;
+                unsigned char *Above;
+                unsigned char *yleft;
+                int left_stride;
+                unsigned char top_left;
+
+                /*Caution: For some b_mode, it needs 8 pixels (4 above + 4 above-right).*/
+                if (i < 4 && pbi->common.filter_level)
+                    Above = xd->recon_above[0] + b->offset;
+                else
+                    Above = dst - dst_stride;
+
+                if (i%4==0 && pbi->common.filter_level)
+                {
+                    yleft = xd->recon_left[0] + i;
+                    left_stride = 1;
+                }
+                else
+                {
+                    yleft = dst - 1;
+                    left_stride = dst_stride;
+                }
+
+                if ((i==4 || i==8 || i==12) && pbi->common.filter_level)
+                    top_left = *(xd->recon_left[0] + i - 1);
+                else
+                    top_left = Above[-1];
+
+                vp8_intra4x4_predict(Above, yleft, left_stride,
+                                     b_mode, dst, dst_stride, top_left);
+
+                if (xd->eobs[i] )
+                {
+                    if (xd->eobs[i] > 1)
+                    {
+                        vp8_dequant_idct_add(b->qcoeff, DQC, dst, dst_stride);
+                    }
+                    else
+                    {
+                        vp8_dc_only_idct_add(b->qcoeff[0] * DQC[0],
+                                             dst, dst_stride, dst, dst_stride);
+                        memset(b->qcoeff, 0, 2 * sizeof(b->qcoeff[0]));
+                    }
+                }
+            }
+        }
+    }
+    else
+    {
+        vp8_build_inter_predictors_mb(xd);
+    }
+
+
+#if CONFIG_ERROR_CONCEALMENT
+    if (corruption_detected)
+    {
+        return;
+    }
+#endif
+
+    if(!xd->mode_info_context->mbmi.mb_skip_coeff)
+    {
+        /* dequantization and idct */
+        if (mode != B_PRED)
+        {
+            short *DQC = xd->dequant_y1;
+
+            if (mode != SPLITMV)
+            {
+                BLOCKD *b = &xd->block[24];
+
+                /* do 2nd order transform on the dc block */
+                if (xd->eobs[24] > 1)
+                {
+                    vp8_dequantize_b(b, xd->dequant_y2);
+
+                    vp8_short_inv_walsh4x4(&b->dqcoeff[0],
+                        xd->qcoeff);
+                    memset(b->qcoeff, 0, 16 * sizeof(b->qcoeff[0]));
+                }
+                else
+                {
+                    b->dqcoeff[0] = b->qcoeff[0] * xd->dequant_y2[0];
+                    vp8_short_inv_walsh4x4_1(&b->dqcoeff[0],
+                        xd->qcoeff);
+                    memset(b->qcoeff, 0, 2 * sizeof(b->qcoeff[0]));
+                }
+
+                /* override the dc dequant constant in order to preserve the
+                 * dc components
+                 */
+                DQC = xd->dequant_y1_dc;
+            }
+
+            vp8_dequant_idct_add_y_block
+                            (xd->qcoeff, DQC,
+                             xd->dst.y_buffer,
+                             xd->dst.y_stride, xd->eobs);
+        }
+
+        vp8_dequant_idct_add_uv_block
+                        (xd->qcoeff+16*16, xd->dequant_uv,
+                         xd->dst.u_buffer, xd->dst.v_buffer,
+                         xd->dst.uv_stride, xd->eobs+16);
+    }
+}
+
+static void mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd, int start_mb_row)
+{
+    const int *last_row_current_mb_col;
+    int *current_mb_col;
+    int mb_row;
+    VP8_COMMON *pc = &pbi->common;
+    const int nsync = pbi->sync_range;
+    const int first_row_no_sync_above = pc->mb_cols + nsync;
+    int num_part = 1 << pbi->common.multi_token_partition;
+    int last_mb_row = start_mb_row;
+
+    YV12_BUFFER_CONFIG *yv12_fb_new = pbi->dec_fb_ref[INTRA_FRAME];
+    YV12_BUFFER_CONFIG *yv12_fb_lst = pbi->dec_fb_ref[LAST_FRAME];
+
+    int recon_y_stride = yv12_fb_new->y_stride;
+    int recon_uv_stride = yv12_fb_new->uv_stride;
+
+    unsigned char *ref_buffer[MAX_REF_FRAMES][3];
+    unsigned char *dst_buffer[3];
+    int i;
+    int ref_fb_corrupted[MAX_REF_FRAMES];
+
+    ref_fb_corrupted[INTRA_FRAME] = 0;
+
+    for(i = 1; i < MAX_REF_FRAMES; i++)
+    {
+        YV12_BUFFER_CONFIG *this_fb = pbi->dec_fb_ref[i];
+
+        ref_buffer[i][0] = this_fb->y_buffer;
+        ref_buffer[i][1] = this_fb->u_buffer;
+        ref_buffer[i][2] = this_fb->v_buffer;
+
+        ref_fb_corrupted[i] = this_fb->corrupted;
+    }
+
+    dst_buffer[0] = yv12_fb_new->y_buffer;
+    dst_buffer[1] = yv12_fb_new->u_buffer;
+    dst_buffer[2] = yv12_fb_new->v_buffer;
+
+    xd->up_available = (start_mb_row != 0);
+
+    xd->mode_info_context = pc->mi + pc->mode_info_stride * start_mb_row;
+    xd->mode_info_stride = pc->mode_info_stride;
+
+    for (mb_row = start_mb_row; mb_row < pc->mb_rows; mb_row += (pbi->decoding_thread_count + 1))
+    {
+       int recon_yoffset, recon_uvoffset;
+       int mb_col;
+       int filter_level;
+       loop_filter_info_n *lfi_n = &pc->lf_info;
+
+       /* save last row processed by this thread */
+       last_mb_row = mb_row;
+       /* select bool coder for current partition */
+       xd->current_bc =  &pbi->mbc[mb_row%num_part];
+
+       if (mb_row > 0)
+           last_row_current_mb_col = &pbi->mt_current_mb_col[mb_row -1];
+       else
+           last_row_current_mb_col = &first_row_no_sync_above;
+
+       current_mb_col = &pbi->mt_current_mb_col[mb_row];
+
+       recon_yoffset = mb_row * recon_y_stride * 16;
+       recon_uvoffset = mb_row * recon_uv_stride * 8;
+
+       /* reset contexts */
+       xd->above_context = pc->above_context;
+       memset(xd->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES));
+
+       xd->left_available = 0;
+
+       xd->mb_to_top_edge = -((mb_row * 16)) << 3;
+       xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
+
+       if (pbi->common.filter_level)
+       {
+          xd->recon_above[0] = pbi->mt_yabove_row[mb_row] + 0*16 +32;
+          xd->recon_above[1] = pbi->mt_uabove_row[mb_row] + 0*8 +16;
+          xd->recon_above[2] = pbi->mt_vabove_row[mb_row] + 0*8 +16;
+
+          xd->recon_left[0] = pbi->mt_yleft_col[mb_row];
+          xd->recon_left[1] = pbi->mt_uleft_col[mb_row];
+          xd->recon_left[2] = pbi->mt_vleft_col[mb_row];
+
+          /* TODO: move to outside row loop */
+          xd->recon_left_stride[0] = 1;
+          xd->recon_left_stride[1] = 1;
+       }
+       else
+       {
+          xd->recon_above[0] = dst_buffer[0] + recon_yoffset;
+          xd->recon_above[1] = dst_buffer[1] + recon_uvoffset;
+          xd->recon_above[2] = dst_buffer[2] + recon_uvoffset;
+
+          xd->recon_left[0] = xd->recon_above[0] - 1;
+          xd->recon_left[1] = xd->recon_above[1] - 1;
+          xd->recon_left[2] = xd->recon_above[2] - 1;
+
+          xd->recon_above[0] -= xd->dst.y_stride;
+          xd->recon_above[1] -= xd->dst.uv_stride;
+          xd->recon_above[2] -= xd->dst.uv_stride;
+
+          /* TODO: move to outside row loop */
+          xd->recon_left_stride[0] = xd->dst.y_stride;
+          xd->recon_left_stride[1] = xd->dst.uv_stride;
+
+          setup_intra_recon_left(xd->recon_left[0], xd->recon_left[1],
+                                 xd->recon_left[2], xd->dst.y_stride,
+                                 xd->dst.uv_stride);
+       }
+
+       for (mb_col = 0; mb_col < pc->mb_cols; mb_col++) {
+           if (((mb_col - 1) % nsync) == 0) {
+               pthread_mutex_t *mutex = &pbi->pmutex[mb_row];
+               protected_write(mutex, current_mb_col, mb_col - 1);
+           }
+
+           if (mb_row && !(mb_col & (nsync - 1))) {
+               pthread_mutex_t *mutex = &pbi->pmutex[mb_row-1];
+               sync_read(mutex, mb_col, last_row_current_mb_col, nsync);
+           }
+
+           /* Distance of MB to the various image edges.
+            * These are specified to 8th pel as they are always
+            * compared to values that are in 1/8th pel units.
+            */
+           xd->mb_to_left_edge = -((mb_col * 16) << 3);
+           xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
+
+    #if CONFIG_ERROR_CONCEALMENT
+           {
+               int corrupt_residual =
+                           (!pbi->independent_partitions &&
+                           pbi->frame_corrupt_residual) ||
+                           vp8dx_bool_error(xd->current_bc);
+               if (pbi->ec_active &&
+                   (xd->mode_info_context->mbmi.ref_frame ==
+                                                    INTRA_FRAME) &&
+                   corrupt_residual)
+               {
+                   /* We have an intra block with corrupt
+                    * coefficients, better to conceal with an inter
+                    * block.
+                    * Interpolate MVs from neighboring MBs
+                    *
+                    * Note that for the first mb with corrupt
+                    * residual in a frame, we might not discover
+                    * that before decoding the residual. That
+                    * happens after this check, and therefore no
+                    * inter concealment will be done.
+                    */
+                   vp8_interpolate_motion(xd,
+                                          mb_row, mb_col,
+                                          pc->mb_rows, pc->mb_cols);
+               }
+           }
+    #endif
+
+
+           xd->dst.y_buffer = dst_buffer[0] + recon_yoffset;
+           xd->dst.u_buffer = dst_buffer[1] + recon_uvoffset;
+           xd->dst.v_buffer = dst_buffer[2] + recon_uvoffset;
+
+           xd->pre.y_buffer = ref_buffer[xd->mode_info_context->mbmi.ref_frame][0] + recon_yoffset;
+           xd->pre.u_buffer = ref_buffer[xd->mode_info_context->mbmi.ref_frame][1] + recon_uvoffset;
+           xd->pre.v_buffer = ref_buffer[xd->mode_info_context->mbmi.ref_frame][2] + recon_uvoffset;
+
+           /* propagate errors from reference frames */
+           xd->corrupted |= ref_fb_corrupted[xd->mode_info_context->mbmi.ref_frame];
+
+           mt_decode_macroblock(pbi, xd, 0);
+
+           xd->left_available = 1;
+
+           /* check if the boolean decoder has suffered an error */
+           xd->corrupted |= vp8dx_bool_error(xd->current_bc);
+
+           xd->recon_above[0] += 16;
+           xd->recon_above[1] += 8;
+           xd->recon_above[2] += 8;
+
+           if (!pbi->common.filter_level)
+           {
+              xd->recon_left[0] += 16;
+              xd->recon_left[1] += 8;
+              xd->recon_left[2] += 8;
+           }
+
+           if (pbi->common.filter_level)
+           {
+               int skip_lf = (xd->mode_info_context->mbmi.mode != B_PRED &&
+                               xd->mode_info_context->mbmi.mode != SPLITMV &&
+                               xd->mode_info_context->mbmi.mb_skip_coeff);
+
+               const int mode_index = lfi_n->mode_lf_lut[xd->mode_info_context->mbmi.mode];
+               const int seg = xd->mode_info_context->mbmi.segment_id;
+               const int ref_frame = xd->mode_info_context->mbmi.ref_frame;
+
+               filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
+
+               if( mb_row != pc->mb_rows-1 )
+               {
+                   /* Save decoded MB last row data for next-row decoding */
+                   memcpy((pbi->mt_yabove_row[mb_row + 1] + 32 + mb_col*16), (xd->dst.y_buffer + 15 * recon_y_stride), 16);
+                   memcpy((pbi->mt_uabove_row[mb_row + 1] + 16 + mb_col*8), (xd->dst.u_buffer + 7 * recon_uv_stride), 8);
+                   memcpy((pbi->mt_vabove_row[mb_row + 1] + 16 + mb_col*8), (xd->dst.v_buffer + 7 * recon_uv_stride), 8);
+               }
+
+               /* save left_col for next MB decoding */
+               if(mb_col != pc->mb_cols-1)
+               {
+                   MODE_INFO *next = xd->mode_info_context +1;
+
+                   if (next->mbmi.ref_frame == INTRA_FRAME)
+                   {
+                       for (i = 0; i < 16; i++)
+                           pbi->mt_yleft_col[mb_row][i] = xd->dst.y_buffer [i* recon_y_stride + 15];
+                       for (i = 0; i < 8; i++)
+                       {
+                           pbi->mt_uleft_col[mb_row][i] = xd->dst.u_buffer [i* recon_uv_stride + 7];
+                           pbi->mt_vleft_col[mb_row][i] = xd->dst.v_buffer [i* recon_uv_stride + 7];
+                       }
+                   }
+               }
+
+               /* loopfilter on this macroblock. */
+               if (filter_level)
+               {
+                   if(pc->filter_type == NORMAL_LOOPFILTER)
+                   {
+                       loop_filter_info lfi;
+                       FRAME_TYPE frame_type = pc->frame_type;
+                       const int hev_index = lfi_n->hev_thr_lut[frame_type][filter_level];
+                       lfi.mblim = lfi_n->mblim[filter_level];
+                       lfi.blim = lfi_n->blim[filter_level];
+                       lfi.lim = lfi_n->lim[filter_level];
+                       lfi.hev_thr = lfi_n->hev_thr[hev_index];
+
+                       if (mb_col > 0)
+                           vp8_loop_filter_mbv
+                           (xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi);
+
+                       if (!skip_lf)
+                           vp8_loop_filter_bv
+                           (xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi);
+
+                       /* don't apply across umv border */
+                       if (mb_row > 0)
+                           vp8_loop_filter_mbh
+                           (xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer, recon_y_stride, recon_uv_stride, &lfi);
+
+                       if (!skip_lf)
+                           vp8_loop_filter_bh
+                           (xd->dst.y_buffer, xd->dst.u_buffer, xd->dst.v_buffer,  recon_y_stride, recon_uv_stride, &lfi);
+                   }
+                   else
+                   {
+                       if (mb_col > 0)
+                           vp8_loop_filter_simple_mbv
+                           (xd->dst.y_buffer, recon_y_stride, lfi_n->mblim[filter_level]);
+
+                       if (!skip_lf)
+                           vp8_loop_filter_simple_bv
+                           (xd->dst.y_buffer, recon_y_stride, lfi_n->blim[filter_level]);
+
+                       /* don't apply across umv border */
+                       if (mb_row > 0)
+                           vp8_loop_filter_simple_mbh
+                           (xd->dst.y_buffer, recon_y_stride, lfi_n->mblim[filter_level]);
+
+                       if (!skip_lf)
+                           vp8_loop_filter_simple_bh
+                           (xd->dst.y_buffer, recon_y_stride, lfi_n->blim[filter_level]);
+                   }
+               }
+
+           }
+
+           recon_yoffset += 16;
+           recon_uvoffset += 8;
+
+           ++xd->mode_info_context;  /* next mb */
+
+           xd->above_context++;
+       }
+
+       /* adjust to the next row of mbs */
+       if (pbi->common.filter_level)
+       {
+           if(mb_row != pc->mb_rows-1)
+           {
+               int lasty = yv12_fb_lst->y_width + VP8BORDERINPIXELS;
+               int lastuv = (yv12_fb_lst->y_width>>1) + (VP8BORDERINPIXELS>>1);
+
+               for (i = 0; i < 4; i++)
+               {
+                   pbi->mt_yabove_row[mb_row +1][lasty + i] = pbi->mt_yabove_row[mb_row +1][lasty -1];
+                   pbi->mt_uabove_row[mb_row +1][lastuv + i] = pbi->mt_uabove_row[mb_row +1][lastuv -1];
+                   pbi->mt_vabove_row[mb_row +1][lastuv + i] = pbi->mt_vabove_row[mb_row +1][lastuv -1];
+               }
+           }
+       }
+       else
+           vp8_extend_mb_row(yv12_fb_new, xd->dst.y_buffer + 16,
+                             xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
+
+       /* last MB of row is ready just after extension is done */
+       protected_write(&pbi->pmutex[mb_row], current_mb_col, mb_col + nsync);
+
+       ++xd->mode_info_context;      /* skip prediction column */
+       xd->up_available = 1;
+
+       /* since we have multithread */
+       xd->mode_info_context += xd->mode_info_stride * pbi->decoding_thread_count;
+    }
+
+    /* signal end of frame decoding if this thread processed the last mb_row */
+    if (last_mb_row == (pc->mb_rows - 1))
+        sem_post(&pbi->h_event_end_decoding);
+
+}
+
+
+static THREAD_FUNCTION thread_decoding_proc(void *p_data)
+{
+    int ithread = ((DECODETHREAD_DATA *)p_data)->ithread;
+    VP8D_COMP *pbi = (VP8D_COMP *)(((DECODETHREAD_DATA *)p_data)->ptr1);
+    MB_ROW_DEC *mbrd = (MB_ROW_DEC *)(((DECODETHREAD_DATA *)p_data)->ptr2);
+    ENTROPY_CONTEXT_PLANES mb_row_left_context;
+
+    while (1)
+    {
+        if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd) == 0)
+            break;
+
+        if (sem_wait(&pbi->h_event_start_decoding[ithread]) == 0)
+        {
+            if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd) == 0)
+                break;
+            else
+            {
+                MACROBLOCKD *xd = &mbrd->mbd;
+                xd->left_context = &mb_row_left_context;
+
+                mt_decode_mb_rows(pbi, xd, ithread+1);
+            }
+        }
+    }
+
+    return 0 ;
+}
+
+
+void vp8_decoder_create_threads(VP8D_COMP *pbi)
+{
+    int core_count = 0;
+    unsigned int ithread;
+
+    pbi->b_multithreaded_rd = 0;
+    pbi->allocated_decoding_thread_count = 0;
+    pthread_mutex_init(&pbi->mt_mutex, NULL);
+
+    /* limit decoding threads to the max number of token partitions */
+    core_count = (pbi->max_threads > 8) ? 8 : pbi->max_threads;
+
+    /* limit decoding threads to the available cores */
+    if (core_count > pbi->common.processor_core_count)
+        core_count = pbi->common.processor_core_count;
+
+    if (core_count > 1)
+    {
+        pbi->b_multithreaded_rd = 1;
+        pbi->decoding_thread_count = core_count - 1;
+
+        CALLOC_ARRAY(pbi->h_decoding_thread, pbi->decoding_thread_count);
+        CALLOC_ARRAY(pbi->h_event_start_decoding, pbi->decoding_thread_count);
+        CALLOC_ARRAY_ALIGNED(pbi->mb_row_di, pbi->decoding_thread_count, 32);
+        CALLOC_ARRAY(pbi->de_thread_data, pbi->decoding_thread_count);
+
+        for (ithread = 0; ithread < pbi->decoding_thread_count; ithread++)
+        {
+            sem_init(&pbi->h_event_start_decoding[ithread], 0, 0);
+
+            vp8_setup_block_dptrs(&pbi->mb_row_di[ithread].mbd);
+
+            pbi->de_thread_data[ithread].ithread  = ithread;
+            pbi->de_thread_data[ithread].ptr1     = (void *)pbi;
+            pbi->de_thread_data[ithread].ptr2     = (void *) &pbi->mb_row_di[ithread];
+
+            pthread_create(&pbi->h_decoding_thread[ithread], 0, thread_decoding_proc, (&pbi->de_thread_data[ithread]));
+        }
+
+        sem_init(&pbi->h_event_end_decoding, 0, 0);
+
+        pbi->allocated_decoding_thread_count = pbi->decoding_thread_count;
+    }
+}
+
+
+void vp8mt_de_alloc_temp_buffers(VP8D_COMP *pbi, int mb_rows)
+{
+    int i;
+
+    if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd))
+    {
+        /* De-allocate mutex */
+        if (pbi->pmutex != NULL) {
+            for (i = 0; i < mb_rows; i++) {
+                pthread_mutex_destroy(&pbi->pmutex[i]);
+            }
+            vpx_free(pbi->pmutex);
+            pbi->pmutex = NULL;
+        }
+
+            vpx_free(pbi->mt_current_mb_col);
+            pbi->mt_current_mb_col = NULL ;
+
+        /* Free above_row buffers. */
+        if (pbi->mt_yabove_row)
+        {
+            for (i=0; i< mb_rows; i++)
+            {
+                    vpx_free(pbi->mt_yabove_row[i]);
+                    pbi->mt_yabove_row[i] = NULL ;
+            }
+            vpx_free(pbi->mt_yabove_row);
+            pbi->mt_yabove_row = NULL ;
+        }
+
+        if (pbi->mt_uabove_row)
+        {
+            for (i=0; i< mb_rows; i++)
+            {
+                    vpx_free(pbi->mt_uabove_row[i]);
+                    pbi->mt_uabove_row[i] = NULL ;
+            }
+            vpx_free(pbi->mt_uabove_row);
+            pbi->mt_uabove_row = NULL ;
+        }
+
+        if (pbi->mt_vabove_row)
+        {
+            for (i=0; i< mb_rows; i++)
+            {
+                    vpx_free(pbi->mt_vabove_row[i]);
+                    pbi->mt_vabove_row[i] = NULL ;
+            }
+            vpx_free(pbi->mt_vabove_row);
+            pbi->mt_vabove_row = NULL ;
+        }
+
+        /* Free left_col buffers. */
+        if (pbi->mt_yleft_col)
+        {
+            for (i=0; i< mb_rows; i++)
+            {
+                    vpx_free(pbi->mt_yleft_col[i]);
+                    pbi->mt_yleft_col[i] = NULL ;
+            }
+            vpx_free(pbi->mt_yleft_col);
+            pbi->mt_yleft_col = NULL ;
+        }
+
+        if (pbi->mt_uleft_col)
+        {
+            for (i=0; i< mb_rows; i++)
+            {
+                    vpx_free(pbi->mt_uleft_col[i]);
+                    pbi->mt_uleft_col[i] = NULL ;
+            }
+            vpx_free(pbi->mt_uleft_col);
+            pbi->mt_uleft_col = NULL ;
+        }
+
+        if (pbi->mt_vleft_col)
+        {
+            for (i=0; i< mb_rows; i++)
+            {
+                    vpx_free(pbi->mt_vleft_col[i]);
+                    pbi->mt_vleft_col[i] = NULL ;
+            }
+            vpx_free(pbi->mt_vleft_col);
+            pbi->mt_vleft_col = NULL ;
+        }
+    }
+}
+
+
+void vp8mt_alloc_temp_buffers(VP8D_COMP *pbi, int width, int prev_mb_rows)
+{
+    VP8_COMMON *const pc = & pbi->common;
+    int i;
+    int uv_width;
+
+    if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd))
+    {
+        vp8mt_de_alloc_temp_buffers(pbi, prev_mb_rows);
+
+        /* our internal buffers are always multiples of 16 */
+        if ((width & 0xf) != 0)
+            width += 16 - (width & 0xf);
+
+        if (width < 640) pbi->sync_range = 1;
+        else if (width <= 1280) pbi->sync_range = 8;
+        else if (width <= 2560) pbi->sync_range =16;
+        else pbi->sync_range = 32;
+
+        uv_width = width >>1;
+
+        /* Allocate mutex */
+        CHECK_MEM_ERROR(pbi->pmutex, vpx_malloc(sizeof(*pbi->pmutex) *
+                                                pc->mb_rows));
+        if (pbi->pmutex) {
+            for (i = 0; i < pc->mb_rows; i++) {
+                pthread_mutex_init(&pbi->pmutex[i], NULL);
+            }
+        }
+
+        /* Allocate an int for each mb row. */
+        CALLOC_ARRAY(pbi->mt_current_mb_col, pc->mb_rows);
+
+        /* Allocate memory for above_row buffers. */
+        CALLOC_ARRAY(pbi->mt_yabove_row, pc->mb_rows);
+        for (i = 0; i < pc->mb_rows; i++)
+            CHECK_MEM_ERROR(pbi->mt_yabove_row[i], vpx_memalign(16,sizeof(unsigned char) * (width + (VP8BORDERINPIXELS<<1))));
+
+        CALLOC_ARRAY(pbi->mt_uabove_row, pc->mb_rows);
+        for (i = 0; i < pc->mb_rows; i++)
+            CHECK_MEM_ERROR(pbi->mt_uabove_row[i], vpx_memalign(16,sizeof(unsigned char) * (uv_width + VP8BORDERINPIXELS)));
+
+        CALLOC_ARRAY(pbi->mt_vabove_row, pc->mb_rows);
+        for (i = 0; i < pc->mb_rows; i++)
+            CHECK_MEM_ERROR(pbi->mt_vabove_row[i], vpx_memalign(16,sizeof(unsigned char) * (uv_width + VP8BORDERINPIXELS)));
+
+        /* Allocate memory for left_col buffers. */
+        CALLOC_ARRAY(pbi->mt_yleft_col, pc->mb_rows);
+        for (i = 0; i < pc->mb_rows; i++)
+            CHECK_MEM_ERROR(pbi->mt_yleft_col[i], vpx_calloc(sizeof(unsigned char) * 16, 1));
+
+        CALLOC_ARRAY(pbi->mt_uleft_col, pc->mb_rows);
+        for (i = 0; i < pc->mb_rows; i++)
+            CHECK_MEM_ERROR(pbi->mt_uleft_col[i], vpx_calloc(sizeof(unsigned char) * 8, 1));
+
+        CALLOC_ARRAY(pbi->mt_vleft_col, pc->mb_rows);
+        for (i = 0; i < pc->mb_rows; i++)
+            CHECK_MEM_ERROR(pbi->mt_vleft_col[i], vpx_calloc(sizeof(unsigned char) * 8, 1));
+    }
+}
+
+
+void vp8_decoder_remove_threads(VP8D_COMP *pbi)
+{
+    /* shutdown MB Decoding thread; */
+    if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd))
+    {
+        int i;
+
+        protected_write(&pbi->mt_mutex, &pbi->b_multithreaded_rd, 0);
+
+        /* allow all threads to exit */
+        for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
+        {
+            sem_post(&pbi->h_event_start_decoding[i]);
+            pthread_join(pbi->h_decoding_thread[i], NULL);
+        }
+
+        for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
+        {
+            sem_destroy(&pbi->h_event_start_decoding[i]);
+        }
+
+        sem_destroy(&pbi->h_event_end_decoding);
+
+            vpx_free(pbi->h_decoding_thread);
+            pbi->h_decoding_thread = NULL;
+
+            vpx_free(pbi->h_event_start_decoding);
+            pbi->h_event_start_decoding = NULL;
+
+            vpx_free(pbi->mb_row_di);
+            pbi->mb_row_di = NULL ;
+
+            vpx_free(pbi->de_thread_data);
+            pbi->de_thread_data = NULL;
+    }
+    pthread_mutex_destroy(&pbi->mt_mutex);
+}
+
+void vp8mt_decode_mb_rows( VP8D_COMP *pbi, MACROBLOCKD *xd)
+{
+    VP8_COMMON *pc = &pbi->common;
+    unsigned int i;
+    int j;
+
+    int filter_level = pc->filter_level;
+    YV12_BUFFER_CONFIG *yv12_fb_new = pbi->dec_fb_ref[INTRA_FRAME];
+
+    if (filter_level)
+    {
+        /* Set above_row buffer to 127 for decoding first MB row */
+        memset(pbi->mt_yabove_row[0] + VP8BORDERINPIXELS-1, 127, yv12_fb_new->y_width + 5);
+        memset(pbi->mt_uabove_row[0] + (VP8BORDERINPIXELS>>1)-1, 127, (yv12_fb_new->y_width>>1) +5);
+        memset(pbi->mt_vabove_row[0] + (VP8BORDERINPIXELS>>1)-1, 127, (yv12_fb_new->y_width>>1) +5);
+
+        for (j=1; j<pc->mb_rows; j++)
+        {
+            memset(pbi->mt_yabove_row[j] + VP8BORDERINPIXELS-1, (unsigned char)129, 1);
+            memset(pbi->mt_uabove_row[j] + (VP8BORDERINPIXELS>>1)-1, (unsigned char)129, 1);
+            memset(pbi->mt_vabove_row[j] + (VP8BORDERINPIXELS>>1)-1, (unsigned char)129, 1);
+        }
+
+        /* Set left_col to 129 initially */
+        for (j=0; j<pc->mb_rows; j++)
+        {
+            memset(pbi->mt_yleft_col[j], (unsigned char)129, 16);
+            memset(pbi->mt_uleft_col[j], (unsigned char)129, 8);
+            memset(pbi->mt_vleft_col[j], (unsigned char)129, 8);
+        }
+
+        /* Initialize the loop filter for this frame. */
+        vp8_loop_filter_frame_init(pc, &pbi->mb, filter_level);
+    }
+    else
+        vp8_setup_intra_recon_top_line(yv12_fb_new);
+
+    setup_decoding_thread_data(pbi, xd, pbi->mb_row_di, pbi->decoding_thread_count);
+
+    for (i = 0; i < pbi->decoding_thread_count; i++)
+        sem_post(&pbi->h_event_start_decoding[i]);
+
+    mt_decode_mb_rows(pbi, xd, 0);
+
+    sem_wait(&pbi->h_event_end_decoding);   /* add back for each frame */
+}
diff --git a/libvpx/libvpx/vp8/decoder/treereader.h b/libvpx/libvpx/vp8/decoder/treereader.h
new file mode 100644
index 0000000..f7d23c3
--- /dev/null
+++ b/libvpx/libvpx/vp8/decoder/treereader.h
@@ -0,0 +1,49 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_DECODER_TREEREADER_H_
+#define VP8_DECODER_TREEREADER_H_
+
+#include "./vpx_config.h"
+#include "vp8/common/treecoder.h"
+#include "dboolhuff.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef BOOL_DECODER vp8_reader;
+
+#define vp8_read vp8dx_decode_bool
+#define vp8_read_literal vp8_decode_value
+#define vp8_read_bit(R) vp8_read(R, vp8_prob_half)
+
+
+/* Intent of tree data structure is to make decoding trivial. */
+
+static INLINE int vp8_treed_read(
+    vp8_reader *const r,        /* !!! must return a 0 or 1 !!! */
+    vp8_tree t,
+    const vp8_prob *const p
+)
+{
+    register vp8_tree_index i = 0;
+
+    while ((i = t[ i + vp8_read(r, p[i>>1])]) > 0) ;
+
+    return -i;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_DECODER_TREEREADER_H_
diff --git a/libvpx/libvpx/vp8/encoder/arm/armv6/vp8_short_fdct4x4_armv6.asm b/libvpx/libvpx/vp8/encoder/arm/armv6/vp8_short_fdct4x4_armv6.asm
new file mode 100644
index 0000000..8034c1d
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/arm/armv6/vp8_short_fdct4x4_armv6.asm
@@ -0,0 +1,262 @@
+;
+;  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT |vp8_short_fdct4x4_armv6|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA    |.text|, CODE, READONLY
+; void vp8_short_fdct4x4_c(short *input, short *output, int pitch)
+|vp8_short_fdct4x4_armv6| PROC
+
+    stmfd       sp!, {r4 - r12, lr}
+
+    ; PART 1
+
+    ; coeffs 0-3
+    ldrd        r4, r5, [r0]        ; [i1 | i0] [i3 | i2]
+
+    ldr         r10, c7500
+    ldr         r11, c14500
+    ldr         r12, c0x22a453a0    ; [2217*4 | 5352*4]
+    ldr         lr, c0x00080008
+    ror         r5, r5, #16         ; [i2 | i3]
+
+    qadd16      r6, r4, r5          ; [i1+i2 | i0+i3] = [b1 | a1] without shift
+    qsub16      r7, r4, r5          ; [i1-i2 | i0-i3] = [c1 | d1] without shift
+
+    add         r0, r0, r2          ; update input pointer
+
+    qadd16      r7, r7, r7          ; 2*[c1|d1] --> we can use smlad and smlsd
+                                    ; with 2217*4 and 5352*4 without losing the
+                                    ; sign bit (overflow)
+
+    smuad       r4, r6, lr          ; o0 = (i1+i2)*8 + (i0+i3)*8
+    smusd       r5, r6, lr          ; o2 = (i1+i2)*8 - (i0+i3)*8
+
+    smlad       r6, r7, r12, r11    ; o1 = (c1 * 2217 + d1 * 5352 +  14500)
+    smlsdx      r7, r7, r12, r10    ; o3 = (d1 * 2217 - c1 * 5352 +   7500)
+
+    ldrd        r8, r9, [r0]        ; [i5 | i4] [i7 | i6]
+
+    pkhbt       r3, r4, r6, lsl #4  ; [o1 | o0], keep in register for PART 2
+    pkhbt       r6, r5, r7, lsl #4  ; [o3 | o2]
+
+    str         r6, [r1, #4]
+
+    ; coeffs 4-7
+    ror         r9, r9, #16         ; [i6 | i7]
+
+    qadd16      r6, r8, r9          ; [i5+i6 | i4+i7] = [b1 | a1] without shift
+    qsub16      r7, r8, r9          ; [i5-i6 | i4-i7] = [c1 | d1] without shift
+
+    add         r0, r0, r2          ; update input pointer
+
+    qadd16      r7, r7, r7          ; 2x[c1|d1] --> we can use smlad and smlsd
+                                    ; with 2217*4 and 5352*4 without losing the
+                                    ; sign bit (overflow)
+
+    smuad       r9, r6, lr          ; o4 = (i5+i6)*8 + (i4+i7)*8
+    smusd       r8, r6, lr          ; o6 = (i5+i6)*8 - (i4+i7)*8
+
+    smlad       r6, r7, r12, r11    ; o5 = (c1 * 2217 + d1 * 5352 +  14500)
+    smlsdx      r7, r7, r12, r10    ; o7 = (d1 * 2217 - c1 * 5352 +   7500)
+
+    ldrd        r4, r5, [r0]        ; [i9 | i8] [i11 | i10]
+
+    pkhbt       r9, r9, r6, lsl #4  ; [o5 | o4], keep in register for PART 2
+    pkhbt       r6, r8, r7, lsl #4  ; [o7 | o6]
+
+    str         r6, [r1, #12]
+
+    ; coeffs 8-11
+    ror         r5, r5, #16         ; [i10 | i11]
+
+    qadd16      r6, r4, r5          ; [i9+i10 | i8+i11]=[b1 | a1] without shift
+    qsub16      r7, r4, r5          ; [i9-i10 | i8-i11]=[c1 | d1] without shift
+
+    add         r0, r0, r2          ; update input pointer
+
+    qadd16      r7, r7, r7          ; 2x[c1|d1] --> we can use smlad and smlsd
+                                    ; with 2217*4 and 5352*4 without losing the
+                                    ; sign bit (overflow)
+
+    smuad       r2, r6, lr          ; o8 = (i9+i10)*8 + (i8+i11)*8
+    smusd       r8, r6, lr          ; o10 = (i9+i10)*8 - (i8+i11)*8
+
+    smlad       r6, r7, r12, r11    ; o9 = (c1 * 2217 + d1 * 5352 +  14500)
+    smlsdx      r7, r7, r12, r10    ; o11 = (d1 * 2217 - c1 * 5352 +   7500)
+
+    ldrd        r4, r5, [r0]        ; [i13 | i12] [i15 | i14]
+
+    pkhbt       r2, r2, r6, lsl #4  ; [o9 | o8], keep in register for PART 2
+    pkhbt       r6, r8, r7, lsl #4  ; [o11 | o10]
+
+    str         r6, [r1, #20]
+
+    ; coeffs 12-15
+    ror         r5, r5, #16         ; [i14 | i15]
+
+    qadd16      r6, r4, r5          ; [i13+i14 | i12+i15]=[b1|a1] without shift
+    qsub16      r7, r4, r5          ; [i13-i14 | i12-i15]=[c1|d1] without shift
+
+    qadd16      r7, r7, r7          ; 2x[c1|d1] --> we can use smlad and smlsd
+                                    ; with 2217*4 and 5352*4 without losing the
+                                    ; sign bit (overflow)
+
+    smuad       r4, r6, lr          ; o12 = (i13+i14)*8 + (i12+i15)*8
+    smusd       r5, r6, lr          ; o14 = (i13+i14)*8 - (i12+i15)*8
+
+    smlad       r6, r7, r12, r11    ; o13 = (c1 * 2217 + d1 * 5352 +  14500)
+    smlsdx      r7, r7, r12, r10    ; o15 = (d1 * 2217 - c1 * 5352 +   7500)
+
+    pkhbt       r0, r4, r6, lsl #4  ; [o13 | o12], keep in register for PART 2
+    pkhbt       r6, r5, r7, lsl #4  ; [o15 | o14]
+
+    str         r6, [r1, #28]
+
+
+    ; PART 2 -------------------------------------------------
+    ldr         r11, c12000
+    ldr         r10, c51000
+    ldr         lr, c0x00070007
+
+    qadd16      r4, r3, r0          ; a1 = [i1+i13 | i0+i12]
+    qadd16      r5, r9, r2          ; b1 = [i5+i9  |  i4+i8]
+    qsub16      r6, r9, r2          ; c1 = [i5-i9  |  i4-i8]
+    qsub16      r7, r3, r0          ; d1 = [i1-i13 | i0-i12]
+
+    qadd16      r4, r4, lr          ; a1 + 7
+
+    add         r0, r11, #0x10000   ; add (d!=0)
+
+    qadd16      r2, r4, r5          ; a1 + b1 + 7
+    qsub16      r3, r4, r5          ; a1 - b1 + 7
+
+    ldr         r12, c0x08a914e8    ; [2217 | 5352]
+
+    lsl         r8, r2, #16         ; prepare bottom halfword for scaling
+    asr         r2, r2, #4          ; scale top halfword
+    lsl         r9, r3, #16         ; prepare bottom halfword for scaling
+    asr         r3, r3, #4          ; scale top halfword
+    pkhtb       r4, r2, r8, asr #20 ; pack and scale bottom halfword
+    pkhtb       r5, r3, r9, asr #20 ; pack and scale bottom halfword
+
+    smulbt      r2, r6, r12         ; [ ------ | c1*2217]
+    str         r4, [r1, #0]        ; [     o1 |      o0]
+    smultt      r3, r6, r12         ; [c1*2217 | ------ ]
+    str         r5, [r1, #16]       ; [     o9 |      o8]
+
+    smlabb      r8, r7, r12, r2     ; [ ------ | d1*5352]
+    smlatb      r9, r7, r12, r3     ; [d1*5352 | ------ ]
+
+    smulbb      r2, r6, r12         ; [ ------ | c1*5352]
+    smultb      r3, r6, r12         ; [c1*5352 | ------ ]
+
+    lsls        r6, r7, #16         ; d1 != 0 ?
+    addeq       r8, r8, r11         ; c1_b*2217+d1_b*5352+12000 + (d==0)
+    addne       r8, r8, r0          ; c1_b*2217+d1_b*5352+12000 + (d!=0)
+    asrs        r6, r7, #16
+    addeq       r9, r9, r11         ; c1_t*2217+d1_t*5352+12000 + (d==0)
+    addne       r9, r9, r0          ; c1_t*2217+d1_t*5352+12000 + (d!=0)
+
+    smlabt      r4, r7, r12, r10    ; [ ------ | d1*2217] + 51000
+    smlatt      r5, r7, r12, r10    ; [d1*2217 | ------ ] + 51000
+
+    pkhtb       r9, r9, r8, asr #16
+
+    sub         r4, r4, r2
+    sub         r5, r5, r3
+
+    ldr         r3, [r1, #4]        ; [i3 | i2]
+
+    pkhtb       r5, r5, r4, asr #16 ; [o13|o12]
+
+    str         r9, [r1, #8]        ; [o5 | 04]
+
+    ldr         r9, [r1, #12]       ; [i7 | i6]
+    ldr         r8, [r1, #28]       ; [i15|i14]
+    ldr         r2, [r1, #20]       ; [i11|i10]
+    str         r5, [r1, #24]       ; [o13|o12]
+
+    qadd16      r4, r3, r8          ; a1 = [i3+i15 | i2+i14]
+    qadd16      r5, r9, r2          ; b1 = [i7+i11 | i6+i10]
+
+    qadd16      r4, r4, lr          ; a1 + 7
+
+    qsub16      r6, r9, r2          ; c1 = [i7-i11 | i6-i10]
+    qadd16      r2, r4, r5          ; a1 + b1 + 7
+    qsub16      r7, r3, r8          ; d1 = [i3-i15 | i2-i14]
+    qsub16      r3, r4, r5          ; a1 - b1 + 7
+
+    lsl         r8, r2, #16         ; prepare bottom halfword for scaling
+    asr         r2, r2, #4          ; scale top halfword
+    lsl         r9, r3, #16         ; prepare bottom halfword for scaling
+    asr         r3, r3, #4          ; scale top halfword
+    pkhtb       r4, r2, r8, asr #20 ; pack and scale bottom halfword
+    pkhtb       r5, r3, r9, asr #20 ; pack and scale bottom halfword
+
+    smulbt      r2, r6, r12         ; [ ------ | c1*2217]
+    str         r4, [r1, #4]        ; [     o3 |      o2]
+    smultt      r3, r6, r12         ; [c1*2217 | ------ ]
+    str         r5, [r1, #20]       ; [    o11 |     o10]
+
+    smlabb      r8, r7, r12, r2     ; [ ------ | d1*5352]
+    smlatb      r9, r7, r12, r3     ; [d1*5352 | ------ ]
+
+    smulbb      r2, r6, r12         ; [ ------ | c1*5352]
+    smultb      r3, r6, r12         ; [c1*5352 | ------ ]
+
+    lsls        r6, r7, #16         ; d1 != 0 ?
+    addeq       r8, r8, r11         ; c1_b*2217+d1_b*5352+12000 + (d==0)
+    addne       r8, r8, r0          ; c1_b*2217+d1_b*5352+12000 + (d!=0)
+
+    asrs        r6, r7, #16
+    addeq       r9, r9, r11         ; c1_t*2217+d1_t*5352+12000 + (d==0)
+    addne       r9, r9, r0          ; c1_t*2217+d1_t*5352+12000 + (d!=0)
+
+    smlabt      r4, r7, r12, r10    ; [ ------ | d1*2217] + 51000
+    smlatt      r5, r7, r12, r10    ; [d1*2217 | ------ ] + 51000
+
+    pkhtb       r9, r9, r8, asr #16
+
+    sub         r4, r4, r2
+    sub         r5, r5, r3
+
+    str         r9, [r1, #12]       ; [o7 | o6]
+    pkhtb       r5, r5, r4, asr #16 ; [o15|o14]
+
+    str         r5, [r1, #28]       ; [o15|o14]
+
+    ldmfd       sp!, {r4 - r12, pc}
+
+    ENDP
+
+; Used constants
+c7500
+    DCD     7500
+c14500
+    DCD     14500
+c0x22a453a0
+    DCD     0x22a453a0
+c0x00080008
+    DCD     0x00080008
+c12000
+    DCD     12000
+c51000
+    DCD     51000
+c0x00070007
+    DCD     0x00070007
+c0x08a914e8
+    DCD     0x08a914e8
+
+    END
diff --git a/libvpx/libvpx/vp8/encoder/arm/armv6/walsh_v6.asm b/libvpx/libvpx/vp8/encoder/arm/armv6/walsh_v6.asm
new file mode 100644
index 0000000..5eaf3f2
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/arm/armv6/walsh_v6.asm
@@ -0,0 +1,212 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT |vp8_short_walsh4x4_armv6|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+
+;short vp8_short_walsh4x4_armv6(short *input, short *output, int pitch)
+; r0    short *input,
+; r1    short *output,
+; r2    int pitch
+|vp8_short_walsh4x4_armv6| PROC
+
+    stmdb       sp!, {r4 - r11, lr}
+
+    ldrd        r4, r5, [r0], r2
+    ldr         lr, c00040004
+    ldrd        r6, r7, [r0], r2
+
+    ; 0-3
+    qadd16      r3, r4, r5          ; [d1|a1] [1+3   |   0+2]
+    qsub16      r4, r4, r5          ; [c1|b1] [1-3   |   0-2]
+
+    ldrd        r8, r9, [r0], r2
+    ; 4-7
+    qadd16      r5, r6, r7          ; [d1|a1] [5+7   |   4+6]
+    qsub16      r6, r6, r7          ; [c1|b1] [5-7   |   4-6]
+
+    ldrd        r10, r11, [r0]
+    ; 8-11
+    qadd16      r7, r8, r9          ; [d1|a1] [9+11  |  8+10]
+    qsub16      r8, r8, r9          ; [c1|b1] [9-11  |  8-10]
+
+    ; 12-15
+    qadd16      r9, r10, r11        ; [d1|a1] [13+15 | 12+14]
+    qsub16      r10, r10, r11       ; [c1|b1] [13-15 | 12-14]
+
+
+    lsls        r2, r3, #16
+    smuad       r11, r3, lr         ; A0 = a1<<2 + d1<<2
+    addne       r11, r11, #1        ; A0 += (a1!=0)
+
+    lsls        r2, r7, #16
+    smuad       r12, r7, lr         ; C0 = a1<<2 + d1<<2
+    addne       r12, r12, #1        ; C0 += (a1!=0)
+
+    add         r0, r11, r12        ; a1_0 = A0 + C0
+    sub         r11, r11, r12       ; b1_0 = A0 - C0
+
+    lsls        r2, r5, #16
+    smuad       r12, r5, lr         ; B0 = a1<<2 + d1<<2
+    addne       r12, r12, #1        ; B0 += (a1!=0)
+
+    lsls        r2, r9, #16
+    smuad       r2, r9, lr          ; D0 = a1<<2 + d1<<2
+    addne       r2, r2, #1          ; D0 += (a1!=0)
+
+    add         lr, r12, r2         ; d1_0 = B0 + D0
+    sub         r12, r12, r2        ; c1_0 = B0 - D0
+
+    ; op[0,4,8,12]
+    adds        r2, r0, lr          ; a2 = a1_0 + d1_0
+    addmi       r2, r2, #1          ; += a2 < 0
+    add         r2, r2, #3          ; += 3
+    subs        r0, r0, lr          ; d2 = a1_0 - d1_0
+    mov         r2, r2, asr #3      ; >> 3
+    strh        r2, [r1]            ; op[0]
+
+    addmi       r0, r0, #1          ; += a2 < 0
+    add         r0, r0, #3          ; += 3
+    ldr         lr, c00040004
+    mov         r0, r0, asr #3      ; >> 3
+    strh        r0, [r1, #24]       ; op[12]
+
+    adds        r2, r11, r12        ; b2 = b1_0 + c1_0
+    addmi       r2, r2, #1          ; += a2 < 0
+    add         r2, r2, #3          ; += 3
+    subs        r0, r11, r12        ; c2 = b1_0 - c1_0
+    mov         r2, r2, asr #3      ; >> 3
+    strh        r2, [r1, #8]        ; op[4]
+
+    addmi       r0, r0, #1          ; += a2 < 0
+    add         r0, r0, #3          ; += 3
+    smusd       r3, r3, lr          ; A3 = a1<<2 - d1<<2
+    smusd       r7, r7, lr          ; C3 = a1<<2 - d1<<2
+    mov         r0, r0, asr #3      ; >> 3
+    strh        r0, [r1, #16]       ; op[8]
+
+
+    ; op[3,7,11,15]
+    add         r0, r3, r7          ; a1_3 = A3 + C3
+    sub         r3, r3, r7          ; b1_3 = A3 - C3
+
+    smusd       r5, r5, lr          ; B3 = a1<<2 - d1<<2
+    smusd       r9, r9, lr          ; D3 = a1<<2 - d1<<2
+    add         r7, r5, r9          ; d1_3 = B3 + D3
+    sub         r5, r5, r9          ; c1_3 = B3 - D3
+
+    adds        r2, r0, r7          ; a2 = a1_3 + d1_3
+    addmi       r2, r2, #1          ; += a2 < 0
+    add         r2, r2, #3          ; += 3
+    adds        r9, r3, r5          ; b2 = b1_3 + c1_3
+    mov         r2, r2, asr #3      ; >> 3
+    strh        r2, [r1, #6]        ; op[3]
+
+    addmi       r9, r9, #1          ; += a2 < 0
+    add         r9, r9, #3          ; += 3
+    subs        r2, r3, r5          ; c2 = b1_3 - c1_3
+    mov         r9, r9, asr #3      ; >> 3
+    strh        r9, [r1, #14]       ; op[7]
+
+    addmi       r2, r2, #1          ; += a2 < 0
+    add         r2, r2, #3          ; += 3
+    subs        r9, r0, r7          ; d2 = a1_3 - d1_3
+    mov         r2, r2, asr #3      ; >> 3
+    strh        r2, [r1, #22]       ; op[11]
+
+    addmi       r9, r9, #1          ; += a2 < 0
+    add         r9, r9, #3          ; += 3
+    smuad       r3, r4, lr          ; A1 = b1<<2 + c1<<2
+    smuad       r5, r8, lr          ; C1 = b1<<2 + c1<<2
+    mov         r9, r9, asr #3      ; >> 3
+    strh        r9, [r1, #30]       ; op[15]
+
+    ; op[1,5,9,13]
+    add         r0, r3, r5          ; a1_1 = A1 + C1
+    sub         r3, r3, r5          ; b1_1 = A1 - C1
+
+    smuad       r7, r6, lr          ; B1 = b1<<2 + c1<<2
+    smuad       r9, r10, lr         ; D1 = b1<<2 + c1<<2
+    add         r5, r7, r9          ; d1_1 = B1 + D1
+    sub         r7, r7, r9          ; c1_1 = B1 - D1
+
+    adds        r2, r0, r5          ; a2 = a1_1 + d1_1
+    addmi       r2, r2, #1          ; += a2 < 0
+    add         r2, r2, #3          ; += 3
+    adds        r9, r3, r7          ; b2 = b1_1 + c1_1
+    mov         r2, r2, asr #3      ; >> 3
+    strh        r2, [r1, #2]        ; op[1]
+
+    addmi       r9, r9, #1          ; += a2 < 0
+    add         r9, r9, #3          ; += 3
+    subs        r2, r3, r7          ; c2 = b1_1 - c1_1
+    mov         r9, r9, asr #3      ; >> 3
+    strh        r9, [r1, #10]       ; op[5]
+
+    addmi       r2, r2, #1          ; += a2 < 0
+    add         r2, r2, #3          ; += 3
+    subs        r9, r0, r5          ; d2 = a1_1 - d1_1
+    mov         r2, r2, asr #3      ; >> 3
+    strh        r2, [r1, #18]       ; op[9]
+
+    addmi       r9, r9, #1          ; += a2 < 0
+    add         r9, r9, #3          ; += 3
+    smusd       r4, r4, lr          ; A2 = b1<<2 - c1<<2
+    smusd       r8, r8, lr          ; C2 = b1<<2 - c1<<2
+    mov         r9, r9, asr #3      ; >> 3
+    strh        r9, [r1, #26]       ; op[13]
+
+
+    ; op[2,6,10,14]
+    add         r11, r4, r8         ; a1_2 = A2 + C2
+    sub         r12, r4, r8         ; b1_2 = A2 - C2
+
+    smusd       r6, r6, lr          ; B2 = b1<<2 - c1<<2
+    smusd       r10, r10, lr        ; D2 = b1<<2 - c1<<2
+    add         r4, r6, r10         ; d1_2 = B2 + D2
+    sub         r8, r6, r10         ; c1_2 = B2 - D2
+
+    adds        r2, r11, r4         ; a2 = a1_2 + d1_2
+    addmi       r2, r2, #1          ; += a2 < 0
+    add         r2, r2, #3          ; += 3
+    adds        r9, r12, r8         ; b2 = b1_2 + c1_2
+    mov         r2, r2, asr #3      ; >> 3
+    strh        r2, [r1, #4]        ; op[2]
+
+    addmi       r9, r9, #1          ; += a2 < 0
+    add         r9, r9, #3          ; += 3
+    subs        r2, r12, r8         ; c2 = b1_2 - c1_2
+    mov         r9, r9, asr #3      ; >> 3
+    strh        r9, [r1, #12]       ; op[6]
+
+    addmi       r2, r2, #1          ; += a2 < 0
+    add         r2, r2, #3          ; += 3
+    subs        r9, r11, r4         ; d2 = a1_2 - d1_2
+    mov         r2, r2, asr #3      ; >> 3
+    strh        r2, [r1, #20]       ; op[10]
+
+    addmi       r9, r9, #1          ; += a2 < 0
+    add         r9, r9, #3          ; += 3
+    mov         r9, r9, asr #3      ; >> 3
+    strh        r9, [r1, #28]       ; op[14]
+
+
+    ldmia       sp!, {r4 - r11, pc}
+    ENDP        ; |vp8_short_walsh4x4_armv6|
+
+c00040004
+    DCD         0x00040004
+
+    END
diff --git a/libvpx/libvpx/vp8/encoder/arm/dct_arm.c b/libvpx/libvpx/vp8/encoder/arm/dct_arm.c
new file mode 100644
index 0000000..f71300d
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/arm/dct_arm.c
@@ -0,0 +1,22 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+
+#if HAVE_MEDIA
+
+void vp8_short_fdct8x4_armv6(short *input, short *output, int pitch)
+{
+    vp8_short_fdct4x4_armv6(input,   output,    pitch);
+    vp8_short_fdct4x4_armv6(input + 4, output + 16, pitch);
+}
+
+#endif /* HAVE_MEDIA */
diff --git a/libvpx/libvpx/vp8/encoder/arm/neon/denoising_neon.c b/libvpx/libvpx/vp8/encoder/arm/neon/denoising_neon.c
new file mode 100644
index 0000000..08be76e
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/arm/neon/denoising_neon.c
@@ -0,0 +1,478 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "vp8/encoder/denoising.h"
+#include "vpx_mem/vpx_mem.h"
+#include "./vp8_rtcd.h"
+
+/*
+ * The filter function was modified to reduce the computational complexity.
+ *
+ * Step 1:
+ *  Instead of applying tap coefficients for each pixel, we calculated the
+ *  pixel adjustments vs. pixel diff value ahead of time.
+ *     adjustment = filtered_value - current_raw
+ *                = (filter_coefficient * diff + 128) >> 8
+ *  where
+ *     filter_coefficient = (255 << 8) / (256 + ((abs_diff * 330) >> 3));
+ *     filter_coefficient += filter_coefficient /
+ *                           (3 + motion_magnitude_adjustment);
+ *     filter_coefficient is clamped to 0 ~ 255.
+ *
+ * Step 2:
+ *  The adjustment vs. diff curve becomes flat very quick when diff increases.
+ *  This allowed us to use only several levels to approximate the curve without
+ *  changing the filtering algorithm too much.
+ *  The adjustments were further corrected by checking the motion magnitude.
+ *  The levels used are:
+ *      diff          level       adjustment w/o       adjustment w/
+ *                               motion correction    motion correction
+ *      [-255, -16]     3              -6                   -7
+ *      [-15, -8]       2              -4                   -5
+ *      [-7, -4]        1              -3                   -4
+ *      [-3, 3]         0              diff                 diff
+ *      [4, 7]          1               3                    4
+ *      [8, 15]         2               4                    5
+ *      [16, 255]       3               6                    7
+ */
+
+int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y,
+                             int mc_running_avg_y_stride,
+                             unsigned char *running_avg_y,
+                             int running_avg_y_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    /* If motion_magnitude is small, making the denoiser more aggressive by
+     * increasing the adjustment for each level, level1 adjustment is
+     * increased, the deltas stay the same.
+     */
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 1 : 0;
+    const uint8x16_t v_level1_adjustment = vmovq_n_u8(
+        (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 4 + shift_inc : 3);
+    const uint8x16_t v_delta_level_1_and_2 = vdupq_n_u8(1);
+    const uint8x16_t v_delta_level_2_and_3 = vdupq_n_u8(2);
+    const uint8x16_t v_level1_threshold = vmovq_n_u8(4 + shift_inc);
+    const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
+    const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
+    int64x2_t v_sum_diff_total = vdupq_n_s64(0);
+
+    /* Go over lines. */
+    int r;
+    for (r = 0; r < 16; ++r) {
+        /* Load inputs. */
+        const uint8x16_t v_sig = vld1q_u8(sig);
+        const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
+
+        /* Calculate absolute difference and sign masks. */
+        const uint8x16_t v_abs_diff      = vabdq_u8(v_sig, v_mc_running_avg_y);
+        const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig, v_mc_running_avg_y);
+        const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig, v_mc_running_avg_y);
+
+        /* Figure out which level that put us in. */
+        const uint8x16_t v_level1_mask = vcleq_u8(v_level1_threshold,
+                                                  v_abs_diff);
+        const uint8x16_t v_level2_mask = vcleq_u8(v_level2_threshold,
+                                                  v_abs_diff);
+        const uint8x16_t v_level3_mask = vcleq_u8(v_level3_threshold,
+                                                  v_abs_diff);
+
+        /* Calculate absolute adjustments for level 1, 2 and 3. */
+        const uint8x16_t v_level2_adjustment = vandq_u8(v_level2_mask,
+                                                        v_delta_level_1_and_2);
+        const uint8x16_t v_level3_adjustment = vandq_u8(v_level3_mask,
+                                                        v_delta_level_2_and_3);
+        const uint8x16_t v_level1and2_adjustment = vaddq_u8(v_level1_adjustment,
+            v_level2_adjustment);
+        const uint8x16_t v_level1and2and3_adjustment = vaddq_u8(
+            v_level1and2_adjustment, v_level3_adjustment);
+
+        /* Figure adjustment absolute value by selecting between the absolute
+         * difference if in level0 or the value for level 1, 2 and 3.
+         */
+        const uint8x16_t v_abs_adjustment = vbslq_u8(v_level1_mask,
+            v_level1and2and3_adjustment, v_abs_diff);
+
+        /* Calculate positive and negative adjustments. Apply them to the signal
+         * and accumulate them. Adjustments are less than eight and the maximum
+         * sum of them (7 * 16) can fit in a signed char.
+         */
+        const uint8x16_t v_pos_adjustment = vandq_u8(v_diff_pos_mask,
+                                                     v_abs_adjustment);
+        const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
+                                                     v_abs_adjustment);
+
+        uint8x16_t v_running_avg_y = vqaddq_u8(v_sig, v_pos_adjustment);
+        v_running_avg_y = vqsubq_u8(v_running_avg_y, v_neg_adjustment);
+
+        /* Store results. */
+        vst1q_u8(running_avg_y, v_running_avg_y);
+
+        /* Sum all the accumulators to have the sum of all pixel differences
+         * for this macroblock.
+         */
+        {
+            const int8x16_t v_sum_diff =
+                vqsubq_s8(vreinterpretq_s8_u8(v_pos_adjustment),
+                          vreinterpretq_s8_u8(v_neg_adjustment));
+
+            const int16x8_t fe_dc_ba_98_76_54_32_10 = vpaddlq_s8(v_sum_diff);
+
+            const int32x4_t fedc_ba98_7654_3210 =
+                vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+
+            const int64x2_t fedcba98_76543210 =
+                vpaddlq_s32(fedc_ba98_7654_3210);
+
+            v_sum_diff_total = vqaddq_s64(v_sum_diff_total, fedcba98_76543210);
+        }
+
+        /* Update pointers for next iteration. */
+        sig += sig_stride;
+        mc_running_avg_y += mc_running_avg_y_stride;
+        running_avg_y += running_avg_y_stride;
+    }
+
+    /* Too much adjustments => copy block. */
+    {
+        int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+                                      vget_low_s64(v_sum_diff_total));
+        int sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+        int sum_diff_thresh = SUM_DIFF_THRESHOLD;
+
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
+        if (sum_diff > sum_diff_thresh) {
+          // Before returning to copy the block (i.e., apply no denoising),
+          // checK if we can still apply some (weaker) temporal filtering to
+          // this block, that would otherwise not be denoised at all. Simplest
+          // is to apply an additional adjustment to running_avg_y to bring it
+          // closer to sig. The adjustment is capped by a maximum delta, and
+          // chosen such that in most cases the resulting sum_diff will be
+          // within the accceptable range given by sum_diff_thresh.
+
+          // The delta is set by the excess of absolute pixel diff over the
+          // threshold.
+          int delta = ((sum_diff - sum_diff_thresh) >> 8) + 1;
+          // Only apply the adjustment for max delta up to 3.
+          if (delta < 4) {
+            const uint8x16_t k_delta = vmovq_n_u8(delta);
+            sig -= sig_stride * 16;
+            mc_running_avg_y -= mc_running_avg_y_stride * 16;
+            running_avg_y -= running_avg_y_stride * 16;
+            for (r = 0; r < 16; ++r) {
+              uint8x16_t v_running_avg_y = vld1q_u8(running_avg_y);
+              const uint8x16_t v_sig = vld1q_u8(sig);
+              const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
+
+              /* Calculate absolute difference and sign masks. */
+              const uint8x16_t v_abs_diff      = vabdq_u8(v_sig,
+                                                          v_mc_running_avg_y);
+              const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig,
+                                                          v_mc_running_avg_y);
+              const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig,
+                                                          v_mc_running_avg_y);
+              // Clamp absolute difference to delta to get the adjustment.
+              const uint8x16_t v_abs_adjustment =
+                  vminq_u8(v_abs_diff, (k_delta));
+
+              const uint8x16_t v_pos_adjustment = vandq_u8(v_diff_pos_mask,
+                                                           v_abs_adjustment);
+              const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
+                                                           v_abs_adjustment);
+
+              v_running_avg_y = vqsubq_u8(v_running_avg_y, v_pos_adjustment);
+              v_running_avg_y = vqaddq_u8(v_running_avg_y, v_neg_adjustment);
+
+              /* Store results. */
+              vst1q_u8(running_avg_y, v_running_avg_y);
+
+              {
+                  const int8x16_t v_sum_diff =
+                      vqsubq_s8(vreinterpretq_s8_u8(v_neg_adjustment),
+                                vreinterpretq_s8_u8(v_pos_adjustment));
+
+                  const int16x8_t fe_dc_ba_98_76_54_32_10 =
+                      vpaddlq_s8(v_sum_diff);
+                  const int32x4_t fedc_ba98_7654_3210 =
+                      vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+                  const int64x2_t fedcba98_76543210 =
+                      vpaddlq_s32(fedc_ba98_7654_3210);
+
+                  v_sum_diff_total = vqaddq_s64(v_sum_diff_total,
+                                                fedcba98_76543210);
+              }
+              /* Update pointers for next iteration. */
+              sig += sig_stride;
+              mc_running_avg_y += mc_running_avg_y_stride;
+              running_avg_y += running_avg_y_stride;
+            }
+            {
+              // Update the sum of all pixel differences of this MB.
+              x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+                            vget_low_s64(v_sum_diff_total));
+              sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+
+              if (sum_diff > sum_diff_thresh) {
+                return COPY_BLOCK;
+              }
+            }
+          } else {
+            return COPY_BLOCK;
+          }
+        }
+    }
+
+    /* Tell above level that block was filtered. */
+    running_avg_y -= running_avg_y_stride * 16;
+    sig -= sig_stride * 16;
+
+    vp8_copy_mem16x16(running_avg_y, running_avg_y_stride, sig, sig_stride);
+
+    return FILTER_BLOCK;
+}
+
+int vp8_denoiser_filter_uv_neon(unsigned char *mc_running_avg,
+                             int mc_running_avg_stride,
+                             unsigned char *running_avg,
+                             int running_avg_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    /* If motion_magnitude is small, making the denoiser more aggressive by
+     * increasing the adjustment for each level, level1 adjustment is
+     * increased, the deltas stay the same.
+     */
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ? 1 : 0;
+    const uint8x16_t v_level1_adjustment = vmovq_n_u8(
+        (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ? 4 + shift_inc : 3);
+
+    const uint8x16_t v_delta_level_1_and_2 = vdupq_n_u8(1);
+    const uint8x16_t v_delta_level_2_and_3 = vdupq_n_u8(2);
+    const uint8x16_t v_level1_threshold = vmovq_n_u8(4 + shift_inc);
+    const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
+    const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
+    int64x2_t v_sum_diff_total = vdupq_n_s64(0);
+    int r;
+
+    {
+      uint16x4_t v_sum_block = vdup_n_u16(0);
+
+      // Avoid denoising color signal if its close to average level.
+      for (r = 0; r < 8; ++r) {
+        const uint8x8_t v_sig = vld1_u8(sig);
+        const uint16x4_t _76_54_32_10 = vpaddl_u8(v_sig);
+        v_sum_block = vqadd_u16(v_sum_block, _76_54_32_10);
+        sig += sig_stride;
+      }
+      sig -= sig_stride * 8;
+      {
+        const uint32x2_t _7654_3210 = vpaddl_u16(v_sum_block);
+        const uint64x1_t _76543210 = vpaddl_u32(_7654_3210);
+        const int sum_block =
+            vget_lane_s32(vreinterpret_s32_u64(_76543210), 0);
+        if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
+          return COPY_BLOCK;
+        }
+      }
+    }
+
+    /* Go over lines. */
+    for (r = 0; r < 4; ++r) {
+        /* Load inputs. */
+        const uint8x8_t v_sig_lo = vld1_u8(sig);
+        const uint8x8_t v_sig_hi = vld1_u8(&sig[sig_stride]);
+        const uint8x16_t v_sig = vcombine_u8(v_sig_lo, v_sig_hi);
+        const uint8x8_t v_mc_running_avg_lo = vld1_u8(mc_running_avg);
+        const uint8x8_t v_mc_running_avg_hi =
+            vld1_u8(&mc_running_avg[mc_running_avg_stride]);
+        const uint8x16_t v_mc_running_avg =
+            vcombine_u8(v_mc_running_avg_lo, v_mc_running_avg_hi);
+        /* Calculate absolute difference and sign masks. */
+        const uint8x16_t v_abs_diff      = vabdq_u8(v_sig, v_mc_running_avg);
+        const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig, v_mc_running_avg);
+        const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig, v_mc_running_avg);
+
+        /* Figure out which level that put us in. */
+        const uint8x16_t v_level1_mask = vcleq_u8(v_level1_threshold,
+                                                  v_abs_diff);
+        const uint8x16_t v_level2_mask = vcleq_u8(v_level2_threshold,
+                                                  v_abs_diff);
+        const uint8x16_t v_level3_mask = vcleq_u8(v_level3_threshold,
+                                                  v_abs_diff);
+
+        /* Calculate absolute adjustments for level 1, 2 and 3. */
+        const uint8x16_t v_level2_adjustment = vandq_u8(v_level2_mask,
+                                                        v_delta_level_1_and_2);
+        const uint8x16_t v_level3_adjustment = vandq_u8(v_level3_mask,
+                                                        v_delta_level_2_and_3);
+        const uint8x16_t v_level1and2_adjustment = vaddq_u8(v_level1_adjustment,
+            v_level2_adjustment);
+        const uint8x16_t v_level1and2and3_adjustment = vaddq_u8(
+            v_level1and2_adjustment, v_level3_adjustment);
+
+        /* Figure adjustment absolute value by selecting between the absolute
+         * difference if in level0 or the value for level 1, 2 and 3.
+         */
+        const uint8x16_t v_abs_adjustment = vbslq_u8(v_level1_mask,
+            v_level1and2and3_adjustment, v_abs_diff);
+
+        /* Calculate positive and negative adjustments. Apply them to the signal
+         * and accumulate them. Adjustments are less than eight and the maximum
+         * sum of them (7 * 16) can fit in a signed char.
+         */
+        const uint8x16_t v_pos_adjustment = vandq_u8(v_diff_pos_mask,
+                                                     v_abs_adjustment);
+        const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
+                                                     v_abs_adjustment);
+
+        uint8x16_t v_running_avg = vqaddq_u8(v_sig, v_pos_adjustment);
+        v_running_avg = vqsubq_u8(v_running_avg, v_neg_adjustment);
+
+        /* Store results. */
+        vst1_u8(running_avg, vget_low_u8(v_running_avg));
+        vst1_u8(&running_avg[running_avg_stride], vget_high_u8(v_running_avg));
+
+        /* Sum all the accumulators to have the sum of all pixel differences
+         * for this macroblock.
+         */
+        {
+            const int8x16_t v_sum_diff =
+                vqsubq_s8(vreinterpretq_s8_u8(v_pos_adjustment),
+                          vreinterpretq_s8_u8(v_neg_adjustment));
+
+            const int16x8_t fe_dc_ba_98_76_54_32_10 = vpaddlq_s8(v_sum_diff);
+
+            const int32x4_t fedc_ba98_7654_3210 =
+                vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+
+            const int64x2_t fedcba98_76543210 =
+                vpaddlq_s32(fedc_ba98_7654_3210);
+
+            v_sum_diff_total = vqaddq_s64(v_sum_diff_total, fedcba98_76543210);
+        }
+
+        /* Update pointers for next iteration. */
+        sig += sig_stride * 2;
+        mc_running_avg += mc_running_avg_stride * 2;
+        running_avg += running_avg_stride * 2;
+    }
+
+
+    /* Too much adjustments => copy block. */
+    {
+        int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+                                      vget_low_s64(v_sum_diff_total));
+        int sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+        int sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+        if (sum_diff > sum_diff_thresh) {
+          // Before returning to copy the block (i.e., apply no denoising),
+          // checK if we can still apply some (weaker) temporal filtering to
+          // this block, that would otherwise not be denoised at all. Simplest
+          // is to apply an additional adjustment to running_avg_y to bring it
+          // closer to sig. The adjustment is capped by a maximum delta, and
+          // chosen such that in most cases the resulting sum_diff will be
+          // within the accceptable range given by sum_diff_thresh.
+
+          // The delta is set by the excess of absolute pixel diff over the
+          // threshold.
+          int delta = ((sum_diff - sum_diff_thresh) >> 8) + 1;
+          // Only apply the adjustment for max delta up to 3.
+          if (delta < 4) {
+            const uint8x16_t k_delta = vmovq_n_u8(delta);
+            sig -= sig_stride * 8;
+            mc_running_avg -= mc_running_avg_stride * 8;
+            running_avg -= running_avg_stride * 8;
+            for (r = 0; r < 4; ++r) {
+              const uint8x8_t v_sig_lo = vld1_u8(sig);
+              const uint8x8_t v_sig_hi = vld1_u8(&sig[sig_stride]);
+              const uint8x16_t v_sig = vcombine_u8(v_sig_lo, v_sig_hi);
+              const uint8x8_t v_mc_running_avg_lo = vld1_u8(mc_running_avg);
+              const uint8x8_t v_mc_running_avg_hi =
+                  vld1_u8(&mc_running_avg[mc_running_avg_stride]);
+              const uint8x16_t v_mc_running_avg =
+                  vcombine_u8(v_mc_running_avg_lo, v_mc_running_avg_hi);
+              /* Calculate absolute difference and sign masks. */
+              const uint8x16_t v_abs_diff      = vabdq_u8(v_sig,
+                                                          v_mc_running_avg);
+              const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig,
+                                                          v_mc_running_avg);
+              const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig,
+                                                          v_mc_running_avg);
+              // Clamp absolute difference to delta to get the adjustment.
+              const uint8x16_t v_abs_adjustment =
+                  vminq_u8(v_abs_diff, (k_delta));
+
+              const uint8x16_t v_pos_adjustment = vandq_u8(v_diff_pos_mask,
+                                                           v_abs_adjustment);
+              const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
+                                                           v_abs_adjustment);
+              const uint8x8_t v_running_avg_lo = vld1_u8(running_avg);
+              const uint8x8_t v_running_avg_hi =
+                  vld1_u8(&running_avg[running_avg_stride]);
+              uint8x16_t v_running_avg =
+                  vcombine_u8(v_running_avg_lo, v_running_avg_hi);
+
+              v_running_avg = vqsubq_u8(v_running_avg, v_pos_adjustment);
+              v_running_avg = vqaddq_u8(v_running_avg, v_neg_adjustment);
+
+              /* Store results. */
+              vst1_u8(running_avg, vget_low_u8(v_running_avg));
+              vst1_u8(&running_avg[running_avg_stride],
+                      vget_high_u8(v_running_avg));
+
+              {
+                  const int8x16_t v_sum_diff =
+                      vqsubq_s8(vreinterpretq_s8_u8(v_neg_adjustment),
+                                vreinterpretq_s8_u8(v_pos_adjustment));
+
+                  const int16x8_t fe_dc_ba_98_76_54_32_10 =
+                      vpaddlq_s8(v_sum_diff);
+                  const int32x4_t fedc_ba98_7654_3210 =
+                      vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+                  const int64x2_t fedcba98_76543210 =
+                      vpaddlq_s32(fedc_ba98_7654_3210);
+
+                  v_sum_diff_total = vqaddq_s64(v_sum_diff_total,
+                                                fedcba98_76543210);
+              }
+              /* Update pointers for next iteration. */
+              sig += sig_stride * 2;
+              mc_running_avg += mc_running_avg_stride * 2;
+              running_avg += running_avg_stride * 2;
+            }
+            {
+              // Update the sum of all pixel differences of this MB.
+              x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+                            vget_low_s64(v_sum_diff_total));
+              sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+
+              if (sum_diff > sum_diff_thresh) {
+                return COPY_BLOCK;
+              }
+            }
+          } else {
+            return COPY_BLOCK;
+          }
+        }
+    }
+
+    /* Tell above level that block was filtered. */
+    running_avg -= running_avg_stride * 8;
+    sig -= sig_stride * 8;
+
+    vp8_copy_mem8x8(running_avg, running_avg_stride, sig, sig_stride);
+
+    return FILTER_BLOCK;
+}
diff --git a/libvpx/libvpx/vp8/encoder/arm/neon/fastquantizeb_neon.c b/libvpx/libvpx/vp8/encoder/arm/neon/fastquantizeb_neon.c
new file mode 100644
index 0000000..e5824bf
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/arm/neon/fastquantizeb_neon.c
@@ -0,0 +1,89 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "vp8/encoder/block.h"
+
+static const uint16_t inv_zig_zag[16] = {
+    1,  2,  6,   7,
+    3,  5,  8,  13,
+    4,  9,  12, 14,
+    10, 11, 15, 16
+};
+
+void vp8_fast_quantize_b_neon(BLOCK *b, BLOCKD *d) {
+    const int16x8_t one_q = vdupq_n_s16(-1),
+                    z0 = vld1q_s16(b->coeff),
+                    z1 = vld1q_s16(b->coeff + 8),
+                    round0 = vld1q_s16(b->round),
+                    round1 = vld1q_s16(b->round + 8),
+                    quant0 = vld1q_s16(b->quant_fast),
+                    quant1 = vld1q_s16(b->quant_fast + 8),
+                    dequant0 = vld1q_s16(d->dequant),
+                    dequant1 = vld1q_s16(d->dequant + 8);
+    const uint16x8_t zig_zag0 = vld1q_u16(inv_zig_zag),
+                     zig_zag1 = vld1q_u16(inv_zig_zag + 8);
+    int16x8_t x0, x1, sz0, sz1, y0, y1;
+    uint16x8_t eob0, eob1;
+    uint16x4_t eob_d16;
+    uint32x2_t eob_d32;
+    uint32x4_t eob_q32;
+
+    /* sign of z: z >> 15 */
+    sz0 = vshrq_n_s16(z0, 15);
+    sz1 = vshrq_n_s16(z1, 15);
+
+    /* x = abs(z) */
+    x0 = vabsq_s16(z0);
+    x1 = vabsq_s16(z1);
+
+    /* x += round */
+    x0 = vaddq_s16(x0, round0);
+    x1 = vaddq_s16(x1, round1);
+
+    /* y = 2 * (x * quant) >> 16 */
+    y0 = vqdmulhq_s16(x0, quant0);
+    y1 = vqdmulhq_s16(x1, quant1);
+
+    /* Compensate for doubling in vqdmulhq */
+    y0 = vshrq_n_s16(y0, 1);
+    y1 = vshrq_n_s16(y1, 1);
+
+    /* Restore sign bit */
+    y0 = veorq_s16(y0, sz0);
+    y1 = veorq_s16(y1, sz1);
+    x0 = vsubq_s16(y0, sz0);
+    x1 = vsubq_s16(y1, sz1);
+
+    /* find non-zero elements */
+    eob0 = vtstq_s16(x0, one_q);
+    eob1 = vtstq_s16(x1, one_q);
+
+    /* mask zig zag */
+    eob0 = vandq_u16(eob0, zig_zag0);
+    eob1 = vandq_u16(eob1, zig_zag1);
+
+    /* select the largest value */
+    eob0 = vmaxq_u16(eob0, eob1);
+    eob_d16 = vmax_u16(vget_low_u16(eob0), vget_high_u16(eob0));
+    eob_q32 = vmovl_u16(eob_d16);
+    eob_d32 = vmax_u32(vget_low_u32(eob_q32), vget_high_u32(eob_q32));
+    eob_d32 = vpmax_u32(eob_d32, eob_d32);
+
+    /* qcoeff = x */
+    vst1q_s16(d->qcoeff, x0);
+    vst1q_s16(d->qcoeff + 8, x1);
+
+    /* dqcoeff = x * dequant */
+    vst1q_s16(d->dqcoeff, vmulq_s16(dequant0, x0));
+    vst1q_s16(d->dqcoeff + 8, vmulq_s16(dequant1, x1));
+
+    vst1_lane_s8((int8_t *)d->eob, vreinterpret_s8_u32(eob_d32), 0);
+}
diff --git a/libvpx/libvpx/vp8/encoder/arm/neon/shortfdct_neon.c b/libvpx/libvpx/vp8/encoder/arm/neon/shortfdct_neon.c
new file mode 100644
index 0000000..391e5f9
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/arm/neon/shortfdct_neon.c
@@ -0,0 +1,269 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+void vp8_short_fdct4x4_neon(
+        int16_t *input,
+        int16_t *output,
+        int pitch) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+    int16x4_t d16s16, d17s16, d26s16, dEmptys16;
+    uint16x4_t d4u16;
+    int16x8_t q0s16, q1s16;
+    int32x4_t q9s32, q10s32, q11s32, q12s32;
+    int16x4x2_t v2tmp0, v2tmp1;
+    int32x2x2_t v2tmp2, v2tmp3;
+
+    d16s16 = vdup_n_s16(5352);
+    d17s16 = vdup_n_s16(2217);
+    q9s32 = vdupq_n_s32(14500);
+    q10s32 = vdupq_n_s32(7500);
+    q11s32 = vdupq_n_s32(12000);
+    q12s32 = vdupq_n_s32(51000);
+
+    // Part one
+    pitch >>= 1;
+    d0s16 = vld1_s16(input);
+    input += pitch;
+    d1s16 = vld1_s16(input);
+    input += pitch;
+    d2s16 = vld1_s16(input);
+    input += pitch;
+    d3s16 = vld1_s16(input);
+
+    v2tmp2 = vtrn_s32(vreinterpret_s32_s16(d0s16),
+                      vreinterpret_s32_s16(d2s16));
+    v2tmp3 = vtrn_s32(vreinterpret_s32_s16(d1s16),
+                      vreinterpret_s32_s16(d3s16));
+    v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),   // d0
+                      vreinterpret_s16_s32(v2tmp3.val[0]));  // d1
+    v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),   // d2
+                      vreinterpret_s16_s32(v2tmp3.val[1]));  // d3
+
+    d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
+    d5s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    d6s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    d7s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
+
+    d4s16 = vshl_n_s16(d4s16, 3);
+    d5s16 = vshl_n_s16(d5s16, 3);
+    d6s16 = vshl_n_s16(d6s16, 3);
+    d7s16 = vshl_n_s16(d7s16, 3);
+
+    d0s16 = vadd_s16(d4s16, d5s16);
+    d2s16 = vsub_s16(d4s16, d5s16);
+
+    q9s32 = vmlal_s16(q9s32, d7s16, d16s16);
+    q10s32 = vmlal_s16(q10s32, d7s16, d17s16);
+    q9s32 = vmlal_s16(q9s32, d6s16, d17s16);
+    q10s32 = vmlsl_s16(q10s32, d6s16, d16s16);
+
+    d1s16 = vshrn_n_s32(q9s32, 12);
+    d3s16 = vshrn_n_s32(q10s32, 12);
+
+    // Part two
+    v2tmp2 = vtrn_s32(vreinterpret_s32_s16(d0s16),
+                      vreinterpret_s32_s16(d2s16));
+    v2tmp3 = vtrn_s32(vreinterpret_s32_s16(d1s16),
+                      vreinterpret_s32_s16(d3s16));
+    v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),   // d0
+                      vreinterpret_s16_s32(v2tmp3.val[0]));  // d1
+    v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),   // d2
+                      vreinterpret_s16_s32(v2tmp3.val[1]));  // d3
+
+    d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
+    d5s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    d6s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    d7s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
+
+    d26s16 = vdup_n_s16(7);
+    d4s16 = vadd_s16(d4s16, d26s16);
+
+    d0s16 = vadd_s16(d4s16, d5s16);
+    d2s16 = vsub_s16(d4s16, d5s16);
+
+    q11s32 = vmlal_s16(q11s32, d7s16, d16s16);
+    q12s32 = vmlal_s16(q12s32, d7s16, d17s16);
+
+    dEmptys16 = vdup_n_s16(0);
+    d4u16 = vceq_s16(d7s16, dEmptys16);
+
+    d0s16 = vshr_n_s16(d0s16, 4);
+    d2s16 = vshr_n_s16(d2s16, 4);
+
+    q11s32 = vmlal_s16(q11s32, d6s16, d17s16);
+    q12s32 = vmlsl_s16(q12s32, d6s16, d16s16);
+
+    d4u16 = vmvn_u16(d4u16);
+    d1s16 = vshrn_n_s32(q11s32, 16);
+    d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d4u16));
+    d3s16 = vshrn_n_s32(q12s32, 16);
+
+    q0s16 = vcombine_s16(d0s16, d1s16);
+    q1s16 = vcombine_s16(d2s16, d3s16);
+
+    vst1q_s16(output, q0s16);
+    vst1q_s16(output + 8, q1s16);
+    return;
+}
+
+void vp8_short_fdct8x4_neon(
+        int16_t *input,
+        int16_t *output,
+        int pitch) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+    int16x4_t d16s16, d17s16, d26s16, d27s16, d28s16, d29s16;
+    uint16x4_t d28u16, d29u16;
+    uint16x8_t q14u16;
+    int16x8_t q0s16, q1s16, q2s16, q3s16;
+    int16x8_t q11s16, q12s16, q13s16, q14s16, q15s16, qEmptys16;
+    int32x4_t q9s32, q10s32, q11s32, q12s32;
+    int16x8x2_t v2tmp0, v2tmp1;
+    int32x4x2_t v2tmp2, v2tmp3;
+
+    d16s16 = vdup_n_s16(5352);
+    d17s16 = vdup_n_s16(2217);
+    q9s32 = vdupq_n_s32(14500);
+    q10s32 = vdupq_n_s32(7500);
+
+    // Part one
+    pitch >>= 1;
+    q0s16 = vld1q_s16(input);
+    input += pitch;
+    q1s16 = vld1q_s16(input);
+    input += pitch;
+    q2s16 = vld1q_s16(input);
+    input += pitch;
+    q3s16 = vld1q_s16(input);
+
+    v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16),
+                       vreinterpretq_s32_s16(q2s16));
+    v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16),
+                       vreinterpretq_s32_s16(q3s16));
+    v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]),   // q0
+                       vreinterpretq_s16_s32(v2tmp3.val[0]));  // q1
+    v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]),   // q2
+                       vreinterpretq_s16_s32(v2tmp3.val[1]));  // q3
+
+    q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
+    q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
+
+    q11s16 = vshlq_n_s16(q11s16, 3);
+    q12s16 = vshlq_n_s16(q12s16, 3);
+    q13s16 = vshlq_n_s16(q13s16, 3);
+    q14s16 = vshlq_n_s16(q14s16, 3);
+
+    q0s16 = vaddq_s16(q11s16, q12s16);
+    q2s16 = vsubq_s16(q11s16, q12s16);
+
+    q11s32 = q9s32;
+    q12s32 = q10s32;
+
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+
+    q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
+    q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
+    q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
+    q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
+
+    q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
+    q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
+    q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
+    q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
+
+    d2s16 = vshrn_n_s32(q9s32, 12);
+    d6s16 = vshrn_n_s32(q10s32, 12);
+    d3s16 = vshrn_n_s32(q11s32, 12);
+    d7s16 = vshrn_n_s32(q12s32, 12);
+    q1s16 = vcombine_s16(d2s16, d3s16);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+
+    // Part two
+    q9s32 = vdupq_n_s32(12000);
+    q10s32 = vdupq_n_s32(51000);
+
+    v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16),
+                       vreinterpretq_s32_s16(q2s16));
+    v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16),
+                       vreinterpretq_s32_s16(q3s16));
+    v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]),   // q0
+                       vreinterpretq_s16_s32(v2tmp3.val[0]));  // q1
+    v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]),   // q2
+                       vreinterpretq_s16_s32(v2tmp3.val[1]));  // q3
+
+    q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
+    q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
+    q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
+
+    q15s16 = vdupq_n_s16(7);
+    q11s16 = vaddq_s16(q11s16, q15s16);
+    q0s16 = vaddq_s16(q11s16, q12s16);
+    q1s16 = vsubq_s16(q11s16, q12s16);
+
+    q11s32 = q9s32;
+    q12s32 = q10s32;
+
+    d0s16 = vget_low_s16(q0s16);
+    d1s16 = vget_high_s16(q0s16);
+    d2s16 = vget_low_s16(q1s16);
+    d3s16 = vget_high_s16(q1s16);
+
+    d0s16 = vshr_n_s16(d0s16, 4);
+    d4s16 = vshr_n_s16(d1s16, 4);
+    d2s16 = vshr_n_s16(d2s16, 4);
+    d6s16 = vshr_n_s16(d3s16, 4);
+
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+
+    q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
+    q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
+    q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
+    q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
+
+    q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
+    q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
+    q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
+    q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
+
+    d1s16 = vshrn_n_s32(q9s32, 16);
+    d3s16 = vshrn_n_s32(q10s32, 16);
+    d5s16 = vshrn_n_s32(q11s32, 16);
+    d7s16 = vshrn_n_s32(q12s32, 16);
+
+    qEmptys16 = vdupq_n_s16(0);
+    q14u16 = vceqq_s16(q14s16, qEmptys16);
+    q14u16 = vmvnq_u16(q14u16);
+
+    d28u16 = vget_low_u16(q14u16);
+    d29u16 = vget_high_u16(q14u16);
+    d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d28u16));
+    d5s16 = vsub_s16(d5s16, vreinterpret_s16_u16(d29u16));
+
+    q0s16 = vcombine_s16(d0s16, d1s16);
+    q1s16 = vcombine_s16(d2s16, d3s16);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+
+    vst1q_s16(output, q0s16);
+    vst1q_s16(output + 8, q1s16);
+    vst1q_s16(output + 16, q2s16);
+    vst1q_s16(output + 24, q3s16);
+    return;
+}
diff --git a/libvpx/libvpx/vp8/encoder/arm/neon/vp8_shortwalsh4x4_neon.c b/libvpx/libvpx/vp8/encoder/arm/neon/vp8_shortwalsh4x4_neon.c
new file mode 100644
index 0000000..5ad9465
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/arm/neon/vp8_shortwalsh4x4_neon.c
@@ -0,0 +1,129 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "vpx_ports/arm.h"
+
+#ifdef VPX_INCOMPATIBLE_GCC
+#include "./vp8_rtcd.h"
+void vp8_short_walsh4x4_neon(
+        int16_t *input,
+        int16_t *output,
+        int pitch) {
+  vp8_short_walsh4x4_c(input, output, pitch);
+}
+#else
+void vp8_short_walsh4x4_neon(
+        int16_t *input,
+        int16_t *output,
+        int pitch) {
+    uint16x4_t d16u16;
+    int16x8_t q0s16, q1s16;
+    int16x4_t dEmptys16, d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+    int32x4_t qEmptys32, q0s32, q1s32, q2s32, q3s32, q8s32;
+    int32x4_t q9s32, q10s32, q11s32, q15s32;
+    uint32x4_t q8u32, q9u32, q10u32, q11u32;
+    int16x4x2_t v2tmp0, v2tmp1;
+    int32x2x2_t v2tmp2, v2tmp3;
+
+    dEmptys16 = vdup_n_s16(0);
+    qEmptys32 = vdupq_n_s32(0);
+    q15s32 = vdupq_n_s32(3);
+
+    d0s16 = vld1_s16(input);
+    input += pitch/2;
+    d1s16 = vld1_s16(input);
+    input += pitch/2;
+    d2s16 = vld1_s16(input);
+    input += pitch/2;
+    d3s16 = vld1_s16(input);
+
+    v2tmp2 = vtrn_s32(vreinterpret_s32_s16(d0s16),
+                      vreinterpret_s32_s16(d2s16));
+    v2tmp3 = vtrn_s32(vreinterpret_s32_s16(d1s16),
+                      vreinterpret_s32_s16(d3s16));
+    v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),   // d0
+                      vreinterpret_s16_s32(v2tmp3.val[0]));  // d1
+    v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),   // d2
+                      vreinterpret_s16_s32(v2tmp3.val[1]));  // d3
+
+    d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[0]);
+    d5s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[1]);
+    d6s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[1]);
+    d7s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[0]);
+
+    d4s16 = vshl_n_s16(d4s16, 2);
+    d5s16 = vshl_n_s16(d5s16, 2);
+    d6s16 = vshl_n_s16(d6s16, 2);
+    d7s16 = vshl_n_s16(d7s16, 2);
+
+    d16u16 = vceq_s16(d4s16, dEmptys16);
+    d16u16 = vmvn_u16(d16u16);
+
+    d0s16 = vadd_s16(d4s16, d5s16);
+    d3s16 = vsub_s16(d4s16, d5s16);
+    d1s16 = vadd_s16(d7s16, d6s16);
+    d2s16 = vsub_s16(d7s16, d6s16);
+
+    d0s16 = vsub_s16(d0s16, vreinterpret_s16_u16(d16u16));
+
+    // Second for-loop
+    v2tmp2 = vtrn_s32(vreinterpret_s32_s16(d1s16),
+                      vreinterpret_s32_s16(d3s16));
+    v2tmp3 = vtrn_s32(vreinterpret_s32_s16(d0s16),
+                      vreinterpret_s32_s16(d2s16));
+    v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp3.val[1]),   // d2
+                      vreinterpret_s16_s32(v2tmp2.val[1]));  // d3
+    v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp3.val[0]),   // d0
+                      vreinterpret_s16_s32(v2tmp2.val[0]));  // d1
+
+    q8s32  = vaddl_s16(v2tmp1.val[0], v2tmp0.val[0]);
+    q9s32  = vaddl_s16(v2tmp1.val[1], v2tmp0.val[1]);
+    q10s32 = vsubl_s16(v2tmp1.val[1], v2tmp0.val[1]);
+    q11s32 = vsubl_s16(v2tmp1.val[0], v2tmp0.val[0]);
+
+    q0s32 = vaddq_s32(q8s32, q9s32);
+    q1s32 = vaddq_s32(q11s32, q10s32);
+    q2s32 = vsubq_s32(q11s32, q10s32);
+    q3s32 = vsubq_s32(q8s32, q9s32);
+
+    q8u32  = vcltq_s32(q0s32, qEmptys32);
+    q9u32  = vcltq_s32(q1s32, qEmptys32);
+    q10u32 = vcltq_s32(q2s32, qEmptys32);
+    q11u32 = vcltq_s32(q3s32, qEmptys32);
+
+    q8s32  = vreinterpretq_s32_u32(q8u32);
+    q9s32  = vreinterpretq_s32_u32(q9u32);
+    q10s32 = vreinterpretq_s32_u32(q10u32);
+    q11s32 = vreinterpretq_s32_u32(q11u32);
+
+    q0s32 = vsubq_s32(q0s32, q8s32);
+    q1s32 = vsubq_s32(q1s32, q9s32);
+    q2s32 = vsubq_s32(q2s32, q10s32);
+    q3s32 = vsubq_s32(q3s32, q11s32);
+
+    q8s32  = vaddq_s32(q0s32, q15s32);
+    q9s32  = vaddq_s32(q1s32, q15s32);
+    q10s32 = vaddq_s32(q2s32, q15s32);
+    q11s32 = vaddq_s32(q3s32, q15s32);
+
+    d0s16 = vshrn_n_s32(q8s32, 3);
+    d1s16 = vshrn_n_s32(q9s32, 3);
+    d2s16 = vshrn_n_s32(q10s32, 3);
+    d3s16 = vshrn_n_s32(q11s32, 3);
+
+    q0s16 = vcombine_s16(d0s16, d1s16);
+    q1s16 = vcombine_s16(d2s16, d3s16);
+
+    vst1q_s16(output, q0s16);
+    vst1q_s16(output + 8, q1s16);
+    return;
+}
+#endif  // VPX_INCOMPATIBLE_GCC
diff --git a/libvpx/libvpx/vp8/encoder/bitstream.c b/libvpx/libvpx/vp8/encoder/bitstream.c
new file mode 100644
index 0000000..3196422
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/bitstream.c
@@ -0,0 +1,1739 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vp8/common/header.h"
+#include "encodemv.h"
+#include "vp8/common/entropymode.h"
+#include "vp8/common/findnearmv.h"
+#include "mcomp.h"
+#include "vp8/common/systemdependent.h"
+#include <assert.h>
+#include <stdio.h>
+#include <limits.h>
+#include "vpx/vpx_encoder.h"
+#include "vpx_mem/vpx_mem.h"
+#include "bitstream.h"
+
+#include "defaultcoefcounts.h"
+#include "vp8/common/common.h"
+
+const int vp8cx_base_skip_false_prob[128] =
+{
+    255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255,
+    255, 255, 255, 255, 255, 255, 255, 255,
+    251, 248, 244, 240, 236, 232, 229, 225,
+    221, 217, 213, 208, 204, 199, 194, 190,
+    187, 183, 179, 175, 172, 168, 164, 160,
+    157, 153, 149, 145, 142, 138, 134, 130,
+    127, 124, 120, 117, 114, 110, 107, 104,
+    101, 98,  95,  92,  89,  86,  83, 80,
+    77,  74,  71,  68,  65,  62,  59, 56,
+    53,  50,  47,  44,  41,  38,  35, 32,
+    30,  28,  26,  24,  22,  20,  18, 16,
+};
+
+#if defined(SECTIONBITS_OUTPUT)
+unsigned __int64 Sectionbits[500];
+#endif
+
+#ifdef VP8_ENTROPY_STATS
+int intra_mode_stats[10][10][10];
+static unsigned int tree_update_hist [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES] [2];
+extern unsigned int active_section;
+#endif
+
+#ifdef MODE_STATS
+int count_mb_seg[4] = { 0, 0, 0, 0 };
+#endif
+
+
+static void update_mode(
+    vp8_writer *const w,
+    int n,
+    vp8_token tok               [/* n */],
+    vp8_tree tree,
+    vp8_prob Pnew               [/* n-1 */],
+    vp8_prob Pcur               [/* n-1 */],
+    unsigned int bct            [/* n-1 */] [2],
+    const unsigned int num_events[/* n */]
+)
+{
+    unsigned int new_b = 0, old_b = 0;
+    int i = 0;
+
+    vp8_tree_probs_from_distribution(
+        n--, tok, tree,
+        Pnew, bct, num_events,
+        256, 1
+    );
+
+    do
+    {
+        new_b += vp8_cost_branch(bct[i], Pnew[i]);
+        old_b += vp8_cost_branch(bct[i], Pcur[i]);
+    }
+    while (++i < n);
+
+    if (new_b + (n << 8) < old_b)
+    {
+        int j = 0;
+
+        vp8_write_bit(w, 1);
+
+        do
+        {
+            const vp8_prob p = Pnew[j];
+
+            vp8_write_literal(w, Pcur[j] = p ? p : 1, 8);
+        }
+        while (++j < n);
+    }
+    else
+        vp8_write_bit(w, 0);
+}
+
+static void update_mbintra_mode_probs(VP8_COMP *cpi)
+{
+    VP8_COMMON *const x = & cpi->common;
+
+    vp8_writer *const w = cpi->bc;
+
+    {
+        vp8_prob Pnew   [VP8_YMODES-1];
+        unsigned int bct [VP8_YMODES-1] [2];
+
+        update_mode(
+            w, VP8_YMODES, vp8_ymode_encodings, vp8_ymode_tree,
+            Pnew, x->fc.ymode_prob, bct, (unsigned int *)cpi->mb.ymode_count
+        );
+    }
+    {
+        vp8_prob Pnew   [VP8_UV_MODES-1];
+        unsigned int bct [VP8_UV_MODES-1] [2];
+
+        update_mode(
+            w, VP8_UV_MODES, vp8_uv_mode_encodings, vp8_uv_mode_tree,
+            Pnew, x->fc.uv_mode_prob, bct, (unsigned int *)cpi->mb.uv_mode_count
+        );
+    }
+}
+
+static void write_ymode(vp8_writer *bc, int m, const vp8_prob *p)
+{
+    vp8_write_token(bc, vp8_ymode_tree, p, vp8_ymode_encodings + m);
+}
+
+static void kfwrite_ymode(vp8_writer *bc, int m, const vp8_prob *p)
+{
+    vp8_write_token(bc, vp8_kf_ymode_tree, p, vp8_kf_ymode_encodings + m);
+}
+
+static void write_uv_mode(vp8_writer *bc, int m, const vp8_prob *p)
+{
+    vp8_write_token(bc, vp8_uv_mode_tree, p, vp8_uv_mode_encodings + m);
+}
+
+
+static void write_bmode(vp8_writer *bc, int m, const vp8_prob *p)
+{
+    vp8_write_token(bc, vp8_bmode_tree, p, vp8_bmode_encodings + m);
+}
+
+static void write_split(vp8_writer *bc, int x)
+{
+    vp8_write_token(
+        bc, vp8_mbsplit_tree, vp8_mbsplit_probs, vp8_mbsplit_encodings + x
+    );
+}
+
+void vp8_pack_tokens(vp8_writer *w, const TOKENEXTRA *p, int xcount)
+{
+    const TOKENEXTRA *stop = p + xcount;
+    unsigned int split;
+    int shift;
+    int count = w->count;
+    unsigned int range = w->range;
+    unsigned int lowvalue = w->lowvalue;
+
+    while (p < stop)
+    {
+        const int t = p->Token;
+        vp8_token *a = vp8_coef_encodings + t;
+        const vp8_extra_bit_struct *b = vp8_extra_bits + t;
+        int i = 0;
+        const unsigned char *pp = p->context_tree;
+        int v = a->value;
+        int n = a->Len;
+
+        if (p->skip_eob_node)
+        {
+            n--;
+            i = 2;
+        }
+
+        do
+        {
+            const int bb = (v >> --n) & 1;
+            split = 1 + (((range - 1) * pp[i>>1]) >> 8);
+            i = vp8_coef_tree[i+bb];
+
+            if (bb)
+            {
+                lowvalue += split;
+                range = range - split;
+            }
+            else
+            {
+                range = split;
+            }
+
+            shift = vp8_norm[range];
+            range <<= shift;
+            count += shift;
+
+            if (count >= 0)
+            {
+                int offset = shift - count;
+
+                if ((lowvalue << (offset - 1)) & 0x80000000)
+                {
+                    int x = w->pos - 1;
+
+                    while (x >= 0 && w->buffer[x] == 0xff)
+                    {
+                        w->buffer[x] = (unsigned char)0;
+                        x--;
+                    }
+
+                    w->buffer[x] += 1;
+                }
+
+                validate_buffer(w->buffer + w->pos,
+                                1,
+                                w->buffer_end,
+                                w->error);
+
+                w->buffer[w->pos++] = (lowvalue >> (24 - offset));
+                lowvalue <<= offset;
+                shift = count;
+                lowvalue &= 0xffffff;
+                count -= 8 ;
+            }
+
+            lowvalue <<= shift;
+        }
+        while (n);
+
+
+        if (b->base_val)
+        {
+            const int e = p->Extra, L = b->Len;
+
+            if (L)
+            {
+                const unsigned char *proba = b->prob;
+                const int v2 = e >> 1;
+                int n2 = L;              /* number of bits in v2, assumed nonzero */
+                i = 0;
+
+                do
+                {
+                    const int bb = (v2 >> --n2) & 1;
+                    split = 1 + (((range - 1) * proba[i>>1]) >> 8);
+                    i = b->tree[i+bb];
+
+                    if (bb)
+                    {
+                        lowvalue += split;
+                        range = range - split;
+                    }
+                    else
+                    {
+                        range = split;
+                    }
+
+                    shift = vp8_norm[range];
+                    range <<= shift;
+                    count += shift;
+
+                    if (count >= 0)
+                    {
+                        int offset = shift - count;
+
+                        if ((lowvalue << (offset - 1)) & 0x80000000)
+                        {
+                            int x = w->pos - 1;
+
+                            while (x >= 0 && w->buffer[x] == 0xff)
+                            {
+                                w->buffer[x] = (unsigned char)0;
+                                x--;
+                            }
+
+                            w->buffer[x] += 1;
+                        }
+
+                        validate_buffer(w->buffer + w->pos,
+                                        1,
+                                        w->buffer_end,
+                                        w->error);
+
+                        w->buffer[w->pos++] = (lowvalue >> (24 - offset));
+                        lowvalue <<= offset;
+                        shift = count;
+                        lowvalue &= 0xffffff;
+                        count -= 8 ;
+                    }
+
+                    lowvalue <<= shift;
+                }
+                while (n2);
+            }
+
+
+            {
+
+                split = (range + 1) >> 1;
+
+                if (e & 1)
+                {
+                    lowvalue += split;
+                    range = range - split;
+                }
+                else
+                {
+                    range = split;
+                }
+
+                range <<= 1;
+
+                if ((lowvalue & 0x80000000))
+                {
+                    int x = w->pos - 1;
+
+                    while (x >= 0 && w->buffer[x] == 0xff)
+                    {
+                        w->buffer[x] = (unsigned char)0;
+                        x--;
+                    }
+
+                    w->buffer[x] += 1;
+
+                }
+
+                lowvalue  <<= 1;
+
+                if (!++count)
+                {
+                    count = -8;
+
+                    validate_buffer(w->buffer + w->pos,
+                                    1,
+                                    w->buffer_end,
+                                    w->error);
+
+                    w->buffer[w->pos++] = (lowvalue >> 24);
+                    lowvalue &= 0xffffff;
+                }
+            }
+
+        }
+
+        ++p;
+    }
+
+    w->count = count;
+    w->lowvalue = lowvalue;
+    w->range = range;
+
+}
+
+static void write_partition_size(unsigned char *cx_data, int size)
+{
+    signed char csize;
+
+    csize = size & 0xff;
+    *cx_data = csize;
+    csize = (size >> 8) & 0xff;
+    *(cx_data + 1) = csize;
+    csize = (size >> 16) & 0xff;
+    *(cx_data + 2) = csize;
+
+}
+
+static void pack_tokens_into_partitions(VP8_COMP *cpi, unsigned char *cx_data,
+                                          unsigned char * cx_data_end,
+                                          int num_part)
+{
+
+    int i;
+    unsigned char *ptr = cx_data;
+    unsigned char *ptr_end = cx_data_end;
+    vp8_writer * w;
+
+    for (i = 0; i < num_part; i++)
+    {
+        int mb_row;
+
+        w = cpi->bc + i + 1;
+
+        vp8_start_encode(w, ptr, ptr_end);
+
+        for (mb_row = i; mb_row < cpi->common.mb_rows; mb_row += num_part)
+        {
+            const TOKENEXTRA *p    = cpi->tplist[mb_row].start;
+            const TOKENEXTRA *stop = cpi->tplist[mb_row].stop;
+            int tokens = (int)(stop - p);
+
+            vp8_pack_tokens(w, p, tokens);
+        }
+
+        vp8_stop_encode(w);
+        ptr += w->pos;
+    }
+}
+
+
+#if CONFIG_MULTITHREAD
+static void pack_mb_row_tokens(VP8_COMP *cpi, vp8_writer *w)
+{
+    int mb_row;
+
+    for (mb_row = 0; mb_row < cpi->common.mb_rows; mb_row++)
+    {
+        const TOKENEXTRA *p    = cpi->tplist[mb_row].start;
+        const TOKENEXTRA *stop = cpi->tplist[mb_row].stop;
+        int tokens = (int)(stop - p);
+
+        vp8_pack_tokens(w, p, tokens);
+    }
+
+}
+#endif  // CONFIG_MULTITHREAD
+
+static void write_mv_ref
+(
+    vp8_writer *w, MB_PREDICTION_MODE m, const vp8_prob *p
+)
+{
+#if CONFIG_DEBUG
+    assert(NEARESTMV <= m  &&  m <= SPLITMV);
+#endif
+    vp8_write_token(w, vp8_mv_ref_tree, p,
+                    vp8_mv_ref_encoding_array + (m - NEARESTMV));
+}
+
+static void write_sub_mv_ref
+(
+    vp8_writer *w, B_PREDICTION_MODE m, const vp8_prob *p
+)
+{
+#if CONFIG_DEBUG
+    assert(LEFT4X4 <= m  &&  m <= NEW4X4);
+#endif
+    vp8_write_token(w, vp8_sub_mv_ref_tree, p,
+                    vp8_sub_mv_ref_encoding_array + (m - LEFT4X4));
+}
+
+static void write_mv
+(
+    vp8_writer *w, const MV *mv, const int_mv *ref, const MV_CONTEXT *mvc
+)
+{
+    MV e;
+    e.row = mv->row - ref->as_mv.row;
+    e.col = mv->col - ref->as_mv.col;
+
+    vp8_encode_motion_vector(w, &e, mvc);
+}
+
+static void write_mb_features(vp8_writer *w, const MB_MODE_INFO *mi, const MACROBLOCKD *x)
+{
+    /* Encode the MB segment id. */
+    if (x->segmentation_enabled && x->update_mb_segmentation_map)
+    {
+        switch (mi->segment_id)
+        {
+        case 0:
+            vp8_write(w, 0, x->mb_segment_tree_probs[0]);
+            vp8_write(w, 0, x->mb_segment_tree_probs[1]);
+            break;
+        case 1:
+            vp8_write(w, 0, x->mb_segment_tree_probs[0]);
+            vp8_write(w, 1, x->mb_segment_tree_probs[1]);
+            break;
+        case 2:
+            vp8_write(w, 1, x->mb_segment_tree_probs[0]);
+            vp8_write(w, 0, x->mb_segment_tree_probs[2]);
+            break;
+        case 3:
+            vp8_write(w, 1, x->mb_segment_tree_probs[0]);
+            vp8_write(w, 1, x->mb_segment_tree_probs[2]);
+            break;
+
+            /* TRAP.. This should not happen */
+        default:
+            vp8_write(w, 0, x->mb_segment_tree_probs[0]);
+            vp8_write(w, 0, x->mb_segment_tree_probs[1]);
+            break;
+        }
+    }
+}
+void vp8_convert_rfct_to_prob(VP8_COMP *const cpi)
+{
+    const int *const rfct = cpi->mb.count_mb_ref_frame_usage;
+    const int rf_intra = rfct[INTRA_FRAME];
+    const int rf_inter = rfct[LAST_FRAME] + rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME];
+
+    /* Calculate the probabilities used to code the ref frame based on usage */
+    if (!(cpi->prob_intra_coded = rf_intra * 255 / (rf_intra + rf_inter)))
+        cpi->prob_intra_coded = 1;
+
+    cpi->prob_last_coded = rf_inter ? (rfct[LAST_FRAME] * 255) / rf_inter : 128;
+
+    if (!cpi->prob_last_coded)
+        cpi->prob_last_coded = 1;
+
+    cpi->prob_gf_coded = (rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME])
+                  ? (rfct[GOLDEN_FRAME] * 255) / (rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME]) : 128;
+
+    if (!cpi->prob_gf_coded)
+        cpi->prob_gf_coded = 1;
+
+}
+
+static void pack_inter_mode_mvs(VP8_COMP *const cpi)
+{
+    VP8_COMMON *const pc = & cpi->common;
+    vp8_writer *const w = cpi->bc;
+    const MV_CONTEXT *mvc = pc->fc.mvc;
+
+
+    MODE_INFO *m = pc->mi;
+    const int mis = pc->mode_info_stride;
+    int mb_row = -1;
+
+    int prob_skip_false = 0;
+
+    cpi->mb.partition_info = cpi->mb.pi;
+
+    vp8_convert_rfct_to_prob(cpi);
+
+#ifdef VP8_ENTROPY_STATS
+    active_section = 1;
+#endif
+
+    if (pc->mb_no_coeff_skip)
+    {
+        int total_mbs = pc->mb_rows * pc->mb_cols;
+
+        prob_skip_false = (total_mbs - cpi->mb.skip_true_count ) * 256 / total_mbs;
+
+        if (prob_skip_false <= 1)
+            prob_skip_false = 1;
+
+        if (prob_skip_false > 255)
+            prob_skip_false = 255;
+
+        cpi->prob_skip_false = prob_skip_false;
+        vp8_write_literal(w, prob_skip_false, 8);
+    }
+
+    vp8_write_literal(w, cpi->prob_intra_coded, 8);
+    vp8_write_literal(w, cpi->prob_last_coded, 8);
+    vp8_write_literal(w, cpi->prob_gf_coded, 8);
+
+    update_mbintra_mode_probs(cpi);
+
+    vp8_write_mvprobs(cpi);
+
+    while (++mb_row < pc->mb_rows)
+    {
+        int mb_col = -1;
+
+        while (++mb_col < pc->mb_cols)
+        {
+            const MB_MODE_INFO *const mi = & m->mbmi;
+            const MV_REFERENCE_FRAME rf = mi->ref_frame;
+            const MB_PREDICTION_MODE mode = mi->mode;
+
+            MACROBLOCKD *xd = &cpi->mb.e_mbd;
+
+            /* Distance of Mb to the various image edges.
+             * These specified to 8th pel as they are always compared to MV
+             * values that are in 1/8th pel units
+             */
+            xd->mb_to_left_edge = -((mb_col * 16) << 3);
+            xd->mb_to_right_edge = ((pc->mb_cols - 1 - mb_col) * 16) << 3;
+            xd->mb_to_top_edge = -((mb_row * 16) << 3);
+            xd->mb_to_bottom_edge = ((pc->mb_rows - 1 - mb_row) * 16) << 3;
+
+#ifdef VP8_ENTROPY_STATS
+            active_section = 9;
+#endif
+
+            if (cpi->mb.e_mbd.update_mb_segmentation_map)
+                write_mb_features(w, mi, &cpi->mb.e_mbd);
+
+            if (pc->mb_no_coeff_skip)
+                vp8_encode_bool(w, m->mbmi.mb_skip_coeff, prob_skip_false);
+
+            if (rf == INTRA_FRAME)
+            {
+                vp8_write(w, 0, cpi->prob_intra_coded);
+#ifdef VP8_ENTROPY_STATS
+                active_section = 6;
+#endif
+                write_ymode(w, mode, pc->fc.ymode_prob);
+
+                if (mode == B_PRED)
+                {
+                    int j = 0;
+
+                    do
+                        write_bmode(w, m->bmi[j].as_mode, pc->fc.bmode_prob);
+                    while (++j < 16);
+                }
+
+                write_uv_mode(w, mi->uv_mode, pc->fc.uv_mode_prob);
+            }
+            else    /* inter coded */
+            {
+                int_mv best_mv;
+                vp8_prob mv_ref_p [VP8_MVREFS-1];
+
+                vp8_write(w, 1, cpi->prob_intra_coded);
+
+                if (rf == LAST_FRAME)
+                    vp8_write(w, 0, cpi->prob_last_coded);
+                else
+                {
+                    vp8_write(w, 1, cpi->prob_last_coded);
+                    vp8_write(w, (rf == GOLDEN_FRAME) ? 0 : 1, cpi->prob_gf_coded);
+                }
+
+                {
+                    int_mv n1, n2;
+                    int ct[4];
+
+                    vp8_find_near_mvs(xd, m, &n1, &n2, &best_mv, ct, rf, cpi->common.ref_frame_sign_bias);
+                    vp8_clamp_mv2(&best_mv, xd);
+
+                    vp8_mv_ref_probs(mv_ref_p, ct);
+
+#ifdef VP8_ENTROPY_STATS
+                    accum_mv_refs(mode, ct);
+#endif
+
+                }
+
+#ifdef VP8_ENTROPY_STATS
+                active_section = 3;
+#endif
+
+                write_mv_ref(w, mode, mv_ref_p);
+
+                switch (mode)   /* new, split require MVs */
+                {
+                case NEWMV:
+
+#ifdef VP8_ENTROPY_STATS
+                    active_section = 5;
+#endif
+
+                    write_mv(w, &mi->mv.as_mv, &best_mv, mvc);
+                    break;
+
+                case SPLITMV:
+                {
+                    int j = 0;
+
+#ifdef MODE_STATS
+                    ++count_mb_seg [mi->partitioning];
+#endif
+
+                    write_split(w, mi->partitioning);
+
+                    do
+                    {
+                        B_PREDICTION_MODE blockmode;
+                        int_mv blockmv;
+                        const int *const  L = vp8_mbsplits [mi->partitioning];
+                        int k = -1;  /* first block in subset j */
+                        int mv_contz;
+                        int_mv leftmv, abovemv;
+
+                        blockmode =  cpi->mb.partition_info->bmi[j].mode;
+                        blockmv =  cpi->mb.partition_info->bmi[j].mv;
+#if CONFIG_DEBUG
+                        while (j != L[++k])
+                            if (k >= 16)
+                                assert(0);
+#else
+                        while (j != L[++k]);
+#endif
+                        leftmv.as_int = left_block_mv(m, k);
+                        abovemv.as_int = above_block_mv(m, k, mis);
+                        mv_contz = vp8_mv_cont(&leftmv, &abovemv);
+
+                        write_sub_mv_ref(w, blockmode, vp8_sub_mv_ref_prob2 [mv_contz]);
+
+                        if (blockmode == NEW4X4)
+                        {
+#ifdef VP8_ENTROPY_STATS
+                            active_section = 11;
+#endif
+                            write_mv(w, &blockmv.as_mv, &best_mv, (const MV_CONTEXT *) mvc);
+                        }
+                    }
+                    while (++j < cpi->mb.partition_info->count);
+                }
+                break;
+                default:
+                    break;
+                }
+            }
+
+            ++m;
+            cpi->mb.partition_info++;
+        }
+
+        ++m;  /* skip L prediction border */
+        cpi->mb.partition_info++;
+    }
+}
+
+
+static void write_kfmodes(VP8_COMP *cpi)
+{
+    vp8_writer *const bc = cpi->bc;
+    const VP8_COMMON *const c = & cpi->common;
+    /* const */
+    MODE_INFO *m = c->mi;
+
+    int mb_row = -1;
+    int prob_skip_false = 0;
+
+    if (c->mb_no_coeff_skip)
+    {
+        int total_mbs = c->mb_rows * c->mb_cols;
+
+        prob_skip_false = (total_mbs - cpi->mb.skip_true_count ) * 256 / total_mbs;
+
+        if (prob_skip_false <= 1)
+            prob_skip_false = 1;
+
+        if (prob_skip_false >= 255)
+            prob_skip_false = 255;
+
+        cpi->prob_skip_false = prob_skip_false;
+        vp8_write_literal(bc, prob_skip_false, 8);
+    }
+
+    while (++mb_row < c->mb_rows)
+    {
+        int mb_col = -1;
+
+        while (++mb_col < c->mb_cols)
+        {
+            const int ym = m->mbmi.mode;
+
+            if (cpi->mb.e_mbd.update_mb_segmentation_map)
+                write_mb_features(bc, &m->mbmi, &cpi->mb.e_mbd);
+
+            if (c->mb_no_coeff_skip)
+                vp8_encode_bool(bc, m->mbmi.mb_skip_coeff, prob_skip_false);
+
+            kfwrite_ymode(bc, ym, vp8_kf_ymode_prob);
+
+            if (ym == B_PRED)
+            {
+                const int mis = c->mode_info_stride;
+                int i = 0;
+
+                do
+                {
+                    const B_PREDICTION_MODE A = above_block_mode(m, i, mis);
+                    const B_PREDICTION_MODE L = left_block_mode(m, i);
+                    const int bm = m->bmi[i].as_mode;
+
+#ifdef VP8_ENTROPY_STATS
+                    ++intra_mode_stats [A] [L] [bm];
+#endif
+
+                    write_bmode(bc, bm, vp8_kf_bmode_prob [A] [L]);
+                }
+                while (++i < 16);
+            }
+
+            write_uv_mode(bc, (m++)->mbmi.uv_mode, vp8_kf_uv_mode_prob);
+        }
+
+        m++;    /* skip L prediction border */
+    }
+}
+
+#if 0
+/* This function is used for debugging probability trees. */
+static void print_prob_tree(vp8_prob
+     coef_probs[BLOCK_TYPES][COEF_BANDS][PREV_COEF_CONTEXTS][ENTROPY_NODES])
+{
+    /* print coef probability tree */
+    int i,j,k,l;
+    FILE* f = fopen("enc_tree_probs.txt", "a");
+    fprintf(f, "{\n");
+    for (i = 0; i < BLOCK_TYPES; i++)
+    {
+        fprintf(f, "  {\n");
+        for (j = 0; j < COEF_BANDS; j++)
+        {
+            fprintf(f, "    {\n");
+            for (k = 0; k < PREV_COEF_CONTEXTS; k++)
+            {
+                fprintf(f, "      {");
+                for (l = 0; l < ENTROPY_NODES; l++)
+                {
+                    fprintf(f, "%3u, ",
+                            (unsigned int)(coef_probs [i][j][k][l]));
+                }
+                fprintf(f, " }\n");
+            }
+            fprintf(f, "    }\n");
+        }
+        fprintf(f, "  }\n");
+    }
+    fprintf(f, "}\n");
+    fclose(f);
+}
+#endif
+
+static void sum_probs_over_prev_coef_context(
+        const unsigned int probs[PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS],
+        unsigned int* out)
+{
+    int i, j;
+    for (i=0; i < MAX_ENTROPY_TOKENS; ++i)
+    {
+        for (j=0; j < PREV_COEF_CONTEXTS; ++j)
+        {
+            const unsigned int tmp = out[i];
+            out[i] += probs[j][i];
+            /* check for wrap */
+            if (out[i] < tmp)
+                out[i] = UINT_MAX;
+        }
+    }
+}
+
+static int prob_update_savings(const unsigned int *ct,
+                                   const vp8_prob oldp, const vp8_prob newp,
+                                   const vp8_prob upd)
+{
+    const int old_b = vp8_cost_branch(ct, oldp);
+    const int new_b = vp8_cost_branch(ct, newp);
+    const int update_b = 8 +
+                         ((vp8_cost_one(upd) - vp8_cost_zero(upd)) >> 8);
+
+    return old_b - new_b - update_b;
+}
+
+static int independent_coef_context_savings(VP8_COMP *cpi)
+{
+    MACROBLOCK *const x = & cpi->mb;
+    int savings = 0;
+    int i = 0;
+    do
+    {
+        int j = 0;
+        do
+        {
+            int k = 0;
+            unsigned int prev_coef_count_sum[MAX_ENTROPY_TOKENS] = {0};
+            int prev_coef_savings[MAX_ENTROPY_TOKENS] = {0};
+            const unsigned int (*probs)[MAX_ENTROPY_TOKENS];
+            /* Calculate new probabilities given the constraint that
+             * they must be equal over the prev coef contexts
+             */
+
+            probs = (const unsigned int (*)[MAX_ENTROPY_TOKENS])
+                x->coef_counts[i][j];
+
+            /* Reset to default probabilities at key frames */
+            if (cpi->common.frame_type == KEY_FRAME)
+                probs = default_coef_counts[i][j];
+
+            sum_probs_over_prev_coef_context(probs, prev_coef_count_sum);
+
+            do
+            {
+                /* at every context */
+
+                /* calc probs and branch cts for this frame only */
+                int t = 0;      /* token/prob index */
+
+                vp8_tree_probs_from_distribution(
+                    MAX_ENTROPY_TOKENS, vp8_coef_encodings, vp8_coef_tree,
+                    cpi->frame_coef_probs[i][j][k],
+                    cpi->frame_branch_ct [i][j][k],
+                    prev_coef_count_sum,
+                    256, 1);
+
+                do
+                {
+                    const unsigned int *ct  = cpi->frame_branch_ct [i][j][k][t];
+                    const vp8_prob newp = cpi->frame_coef_probs [i][j][k][t];
+                    const vp8_prob oldp = cpi->common.fc.coef_probs [i][j][k][t];
+                    const vp8_prob upd = vp8_coef_update_probs [i][j][k][t];
+                    const int s = prob_update_savings(ct, oldp, newp, upd);
+
+                    if (cpi->common.frame_type != KEY_FRAME ||
+                        (cpi->common.frame_type == KEY_FRAME && newp != oldp))
+                        prev_coef_savings[t] += s;
+                }
+                while (++t < ENTROPY_NODES);
+            }
+            while (++k < PREV_COEF_CONTEXTS);
+            k = 0;
+            do
+            {
+                /* We only update probabilities if we can save bits, except
+                 * for key frames where we have to update all probabilities
+                 * to get the equal probabilities across the prev coef
+                 * contexts.
+                 */
+                if (prev_coef_savings[k] > 0 ||
+                    cpi->common.frame_type == KEY_FRAME)
+                    savings += prev_coef_savings[k];
+            }
+            while (++k < ENTROPY_NODES);
+        }
+        while (++j < COEF_BANDS);
+    }
+    while (++i < BLOCK_TYPES);
+    return savings;
+}
+
+static int default_coef_context_savings(VP8_COMP *cpi)
+{
+    MACROBLOCK *const x = & cpi->mb;
+    int savings = 0;
+    int i = 0;
+    do
+    {
+        int j = 0;
+        do
+        {
+            int k = 0;
+            do
+            {
+                /* at every context */
+
+                /* calc probs and branch cts for this frame only */
+                int t = 0;      /* token/prob index */
+
+                vp8_tree_probs_from_distribution(
+                    MAX_ENTROPY_TOKENS, vp8_coef_encodings, vp8_coef_tree,
+                    cpi->frame_coef_probs [i][j][k],
+                    cpi->frame_branch_ct [i][j][k],
+                    x->coef_counts [i][j][k],
+                    256, 1
+                );
+
+                do
+                {
+                    const unsigned int *ct  = cpi->frame_branch_ct [i][j][k][t];
+                    const vp8_prob newp = cpi->frame_coef_probs [i][j][k][t];
+                    const vp8_prob oldp = cpi->common.fc.coef_probs [i][j][k][t];
+                    const vp8_prob upd = vp8_coef_update_probs [i][j][k][t];
+                    const int s = prob_update_savings(ct, oldp, newp, upd);
+
+                    if (s > 0)
+                    {
+                        savings += s;
+                    }
+                }
+                while (++t < ENTROPY_NODES);
+            }
+            while (++k < PREV_COEF_CONTEXTS);
+        }
+        while (++j < COEF_BANDS);
+    }
+    while (++i < BLOCK_TYPES);
+    return savings;
+}
+
+void vp8_calc_ref_frame_costs(int *ref_frame_cost,
+                              int prob_intra,
+                              int prob_last,
+                              int prob_garf
+                             )
+{
+    assert(prob_intra >= 0);
+    assert(prob_intra <= 255);
+    assert(prob_last >= 0);
+    assert(prob_last <= 255);
+    assert(prob_garf >= 0);
+    assert(prob_garf <= 255);
+    ref_frame_cost[INTRA_FRAME]   = vp8_cost_zero(prob_intra);
+    ref_frame_cost[LAST_FRAME]    = vp8_cost_one(prob_intra)
+                                    + vp8_cost_zero(prob_last);
+    ref_frame_cost[GOLDEN_FRAME]  = vp8_cost_one(prob_intra)
+                                    + vp8_cost_one(prob_last)
+                                    + vp8_cost_zero(prob_garf);
+    ref_frame_cost[ALTREF_FRAME]  = vp8_cost_one(prob_intra)
+                                    + vp8_cost_one(prob_last)
+                                    + vp8_cost_one(prob_garf);
+
+}
+
+int vp8_estimate_entropy_savings(VP8_COMP *cpi)
+{
+    int savings = 0;
+
+    const int *const rfct = cpi->mb.count_mb_ref_frame_usage;
+    const int rf_intra = rfct[INTRA_FRAME];
+    const int rf_inter = rfct[LAST_FRAME] + rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME];
+    int new_intra, new_last, new_garf, oldtotal, newtotal;
+    int ref_frame_cost[MAX_REF_FRAMES];
+
+    vp8_clear_system_state();
+
+    if (cpi->common.frame_type != KEY_FRAME)
+    {
+        if (!(new_intra = rf_intra * 255 / (rf_intra + rf_inter)))
+            new_intra = 1;
+
+        new_last = rf_inter ? (rfct[LAST_FRAME] * 255) / rf_inter : 128;
+
+        new_garf = (rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME])
+                  ? (rfct[GOLDEN_FRAME] * 255) / (rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME]) : 128;
+
+
+        vp8_calc_ref_frame_costs(ref_frame_cost,new_intra,new_last,new_garf);
+
+        newtotal =
+            rfct[INTRA_FRAME] * ref_frame_cost[INTRA_FRAME] +
+            rfct[LAST_FRAME] * ref_frame_cost[LAST_FRAME] +
+            rfct[GOLDEN_FRAME] * ref_frame_cost[GOLDEN_FRAME] +
+            rfct[ALTREF_FRAME] * ref_frame_cost[ALTREF_FRAME];
+
+
+        /* old costs */
+        vp8_calc_ref_frame_costs(ref_frame_cost,cpi->prob_intra_coded,
+                                 cpi->prob_last_coded,cpi->prob_gf_coded);
+
+        oldtotal =
+            rfct[INTRA_FRAME] * ref_frame_cost[INTRA_FRAME] +
+            rfct[LAST_FRAME] * ref_frame_cost[LAST_FRAME] +
+            rfct[GOLDEN_FRAME] * ref_frame_cost[GOLDEN_FRAME] +
+            rfct[ALTREF_FRAME] * ref_frame_cost[ALTREF_FRAME];
+
+        savings += (oldtotal - newtotal) / 256;
+    }
+
+
+    if (cpi->oxcf.error_resilient_mode & VPX_ERROR_RESILIENT_PARTITIONS)
+        savings += independent_coef_context_savings(cpi);
+    else
+        savings += default_coef_context_savings(cpi);
+
+
+    return savings;
+}
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+int vp8_update_coef_context(VP8_COMP *cpi)
+{
+    int savings = 0;
+
+
+    if (cpi->common.frame_type == KEY_FRAME)
+    {
+        /* Reset to default counts/probabilities at key frames */
+        vp8_copy(cpi->mb.coef_counts, default_coef_counts);
+    }
+
+    if (cpi->oxcf.error_resilient_mode & VPX_ERROR_RESILIENT_PARTITIONS)
+        savings += independent_coef_context_savings(cpi);
+    else
+        savings += default_coef_context_savings(cpi);
+
+    return savings;
+}
+#endif
+
+void vp8_update_coef_probs(VP8_COMP *cpi)
+{
+    int i = 0;
+#if !(CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+    vp8_writer *const w = cpi->bc;
+#endif
+    int savings = 0;
+
+    vp8_clear_system_state();
+
+    do
+    {
+        int j = 0;
+
+        do
+        {
+            int k = 0;
+            int prev_coef_savings[ENTROPY_NODES] = {0};
+            if (cpi->oxcf.error_resilient_mode & VPX_ERROR_RESILIENT_PARTITIONS)
+            {
+                for (k = 0; k < PREV_COEF_CONTEXTS; ++k)
+                {
+                    int t;      /* token/prob index */
+                    for (t = 0; t < ENTROPY_NODES; ++t)
+                    {
+                        const unsigned int *ct = cpi->frame_branch_ct [i][j]
+                                                                      [k][t];
+                        const vp8_prob newp = cpi->frame_coef_probs[i][j][k][t];
+                        const vp8_prob oldp = cpi->common.fc.coef_probs[i][j]
+                                                                       [k][t];
+                        const vp8_prob upd = vp8_coef_update_probs[i][j][k][t];
+
+                        prev_coef_savings[t] +=
+                                prob_update_savings(ct, oldp, newp, upd);
+                    }
+                }
+                k = 0;
+            }
+            do
+            {
+                /* note: use result from vp8_estimate_entropy_savings, so no
+                 * need to call vp8_tree_probs_from_distribution here.
+                 */
+
+                /* at every context */
+
+                /* calc probs and branch cts for this frame only */
+                int t = 0;      /* token/prob index */
+
+                do
+                {
+                    const vp8_prob newp = cpi->frame_coef_probs [i][j][k][t];
+
+                    vp8_prob *Pold = cpi->common.fc.coef_probs [i][j][k] + t;
+                    const vp8_prob upd = vp8_coef_update_probs [i][j][k][t];
+
+                    int s = prev_coef_savings[t];
+                    int u = 0;
+
+                    if (!(cpi->oxcf.error_resilient_mode &
+                            VPX_ERROR_RESILIENT_PARTITIONS))
+                    {
+                        s = prob_update_savings(
+                                cpi->frame_branch_ct [i][j][k][t],
+                                *Pold, newp, upd);
+                    }
+
+                    if (s > 0)
+                        u = 1;
+
+                    /* Force updates on key frames if the new is different,
+                     * so that we can be sure we end up with equal probabilities
+                     * over the prev coef contexts.
+                     */
+                    if ((cpi->oxcf.error_resilient_mode &
+                            VPX_ERROR_RESILIENT_PARTITIONS) &&
+                        cpi->common.frame_type == KEY_FRAME && newp != *Pold)
+                        u = 1;
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+                    cpi->update_probs[i][j][k][t] = u;
+#else
+                    vp8_write(w, u, upd);
+#endif
+
+
+#ifdef VP8_ENTROPY_STATS
+                    ++ tree_update_hist [i][j][k][t] [u];
+#endif
+
+                    if (u)
+                    {
+                        /* send/use new probability */
+
+                        *Pold = newp;
+#if !(CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+                        vp8_write_literal(w, newp, 8);
+#endif
+
+                        savings += s;
+
+                    }
+
+                }
+                while (++t < ENTROPY_NODES);
+
+                /* Accum token counts for generation of default statistics */
+#ifdef VP8_ENTROPY_STATS
+                t = 0;
+
+                do
+                {
+                    context_counters [i][j][k][t] += cpi->coef_counts [i][j][k][t];
+                }
+                while (++t < MAX_ENTROPY_TOKENS);
+
+#endif
+
+            }
+            while (++k < PREV_COEF_CONTEXTS);
+        }
+        while (++j < COEF_BANDS);
+    }
+    while (++i < BLOCK_TYPES);
+
+}
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+static void pack_coef_probs(VP8_COMP *cpi)
+{
+    int i = 0;
+    vp8_writer *const w = cpi->bc;
+
+    do
+    {
+        int j = 0;
+
+        do
+        {
+            int k = 0;
+
+            do
+            {
+                int t = 0;      /* token/prob index */
+
+                do
+                {
+                    const vp8_prob newp = cpi->common.fc.coef_probs [i][j][k][t];
+                    const vp8_prob upd = vp8_coef_update_probs [i][j][k][t];
+
+                    const char u = cpi->update_probs[i][j][k][t] ;
+
+                    vp8_write(w, u, upd);
+
+                    if (u)
+                    {
+                        /* send/use new probability */
+                        vp8_write_literal(w, newp, 8);
+                    }
+                }
+                while (++t < ENTROPY_NODES);
+            }
+            while (++k < PREV_COEF_CONTEXTS);
+        }
+        while (++j < COEF_BANDS);
+    }
+    while (++i < BLOCK_TYPES);
+}
+#endif
+
+#ifdef PACKET_TESTING
+FILE *vpxlogc = 0;
+#endif
+
+static void put_delta_q(vp8_writer *bc, int delta_q)
+{
+    if (delta_q != 0)
+    {
+        vp8_write_bit(bc, 1);
+        vp8_write_literal(bc, abs(delta_q), 4);
+
+        if (delta_q < 0)
+            vp8_write_bit(bc, 1);
+        else
+            vp8_write_bit(bc, 0);
+    }
+    else
+        vp8_write_bit(bc, 0);
+}
+
+void vp8_pack_bitstream(VP8_COMP *cpi, unsigned char *dest, unsigned char * dest_end, unsigned long *size)
+{
+    int i, j;
+    VP8_HEADER oh;
+    VP8_COMMON *const pc = & cpi->common;
+    vp8_writer *const bc = cpi->bc;
+    MACROBLOCKD *const xd = & cpi->mb.e_mbd;
+    int extra_bytes_packed = 0;
+
+    unsigned char *cx_data = dest;
+    unsigned char *cx_data_end = dest_end;
+    const int *mb_feature_data_bits;
+
+    oh.show_frame = (int) pc->show_frame;
+    oh.type = (int)pc->frame_type;
+    oh.version = pc->version;
+    oh.first_partition_length_in_bytes = 0;
+
+    mb_feature_data_bits = vp8_mb_feature_data_bits;
+
+    bc[0].error = &pc->error;
+
+    validate_buffer(cx_data, 3, cx_data_end, &cpi->common.error);
+    cx_data += 3;
+
+#if defined(SECTIONBITS_OUTPUT)
+    Sectionbits[active_section = 1] += sizeof(VP8_HEADER) * 8 * 256;
+#endif
+
+    /* every keyframe send startcode, width, height, scale factor, clamp
+     * and color type
+     */
+    if (oh.type == KEY_FRAME)
+    {
+        int v;
+
+        validate_buffer(cx_data, 7, cx_data_end, &cpi->common.error);
+
+        /* Start / synch code */
+        cx_data[0] = 0x9D;
+        cx_data[1] = 0x01;
+        cx_data[2] = 0x2a;
+
+        v = (pc->horiz_scale << 14) | pc->Width;
+        cx_data[3] = v;
+        cx_data[4] = v >> 8;
+
+        v = (pc->vert_scale << 14) | pc->Height;
+        cx_data[5] = v;
+        cx_data[6] = v >> 8;
+
+
+        extra_bytes_packed = 7;
+        cx_data += extra_bytes_packed ;
+
+        vp8_start_encode(bc, cx_data, cx_data_end);
+
+        /* signal clr type */
+        vp8_write_bit(bc, 0);
+        vp8_write_bit(bc, pc->clamp_type);
+
+    }
+    else
+        vp8_start_encode(bc, cx_data, cx_data_end);
+
+
+    /* Signal whether or not Segmentation is enabled */
+    vp8_write_bit(bc, xd->segmentation_enabled);
+
+    /*  Indicate which features are enabled */
+    if (xd->segmentation_enabled)
+    {
+        /* Signal whether or not the segmentation map is being updated. */
+        vp8_write_bit(bc, xd->update_mb_segmentation_map);
+        vp8_write_bit(bc, xd->update_mb_segmentation_data);
+
+        if (xd->update_mb_segmentation_data)
+        {
+            signed char Data;
+
+            vp8_write_bit(bc, xd->mb_segement_abs_delta);
+
+            /* For each segmentation feature (Quant and loop filter level) */
+            for (i = 0; i < MB_LVL_MAX; i++)
+            {
+                /* For each of the segments */
+                for (j = 0; j < MAX_MB_SEGMENTS; j++)
+                {
+                    Data = xd->segment_feature_data[i][j];
+
+                    /* Frame level data */
+                    if (Data)
+                    {
+                        vp8_write_bit(bc, 1);
+
+                        if (Data < 0)
+                        {
+                            Data = - Data;
+                            vp8_write_literal(bc, Data, mb_feature_data_bits[i]);
+                            vp8_write_bit(bc, 1);
+                        }
+                        else
+                        {
+                            vp8_write_literal(bc, Data, mb_feature_data_bits[i]);
+                            vp8_write_bit(bc, 0);
+                        }
+                    }
+                    else
+                        vp8_write_bit(bc, 0);
+                }
+            }
+        }
+
+        if (xd->update_mb_segmentation_map)
+        {
+            /* Write the probs used to decode the segment id for each mb */
+            for (i = 0; i < MB_FEATURE_TREE_PROBS; i++)
+            {
+                int Data = xd->mb_segment_tree_probs[i];
+
+                if (Data != 255)
+                {
+                    vp8_write_bit(bc, 1);
+                    vp8_write_literal(bc, Data, 8);
+                }
+                else
+                    vp8_write_bit(bc, 0);
+            }
+        }
+    }
+
+    vp8_write_bit(bc, pc->filter_type);
+    vp8_write_literal(bc, pc->filter_level, 6);
+    vp8_write_literal(bc, pc->sharpness_level, 3);
+
+    /* Write out loop filter deltas applied at the MB level based on mode
+     * or ref frame (if they are enabled).
+     */
+    vp8_write_bit(bc, xd->mode_ref_lf_delta_enabled);
+
+    if (xd->mode_ref_lf_delta_enabled)
+    {
+        /* Do the deltas need to be updated */
+        int send_update = xd->mode_ref_lf_delta_update
+                          || cpi->oxcf.error_resilient_mode;
+
+        vp8_write_bit(bc, send_update);
+        if (send_update)
+        {
+            int Data;
+
+            /* Send update */
+            for (i = 0; i < MAX_REF_LF_DELTAS; i++)
+            {
+                Data = xd->ref_lf_deltas[i];
+
+                /* Frame level data */
+                if (xd->ref_lf_deltas[i] != xd->last_ref_lf_deltas[i]
+                    || cpi->oxcf.error_resilient_mode)
+                {
+                    xd->last_ref_lf_deltas[i] = xd->ref_lf_deltas[i];
+                    vp8_write_bit(bc, 1);
+
+                    if (Data > 0)
+                    {
+                        vp8_write_literal(bc, (Data & 0x3F), 6);
+                        vp8_write_bit(bc, 0);    /* sign */
+                    }
+                    else
+                    {
+                        Data = -Data;
+                        vp8_write_literal(bc, (Data & 0x3F), 6);
+                        vp8_write_bit(bc, 1);    /* sign */
+                    }
+                }
+                else
+                    vp8_write_bit(bc, 0);
+            }
+
+            /* Send update */
+            for (i = 0; i < MAX_MODE_LF_DELTAS; i++)
+            {
+                Data = xd->mode_lf_deltas[i];
+
+                if (xd->mode_lf_deltas[i] != xd->last_mode_lf_deltas[i]
+                    || cpi->oxcf.error_resilient_mode)
+                {
+                    xd->last_mode_lf_deltas[i] = xd->mode_lf_deltas[i];
+                    vp8_write_bit(bc, 1);
+
+                    if (Data > 0)
+                    {
+                        vp8_write_literal(bc, (Data & 0x3F), 6);
+                        vp8_write_bit(bc, 0);    /* sign */
+                    }
+                    else
+                    {
+                        Data = -Data;
+                        vp8_write_literal(bc, (Data & 0x3F), 6);
+                        vp8_write_bit(bc, 1);    /* sign */
+                    }
+                }
+                else
+                    vp8_write_bit(bc, 0);
+            }
+        }
+    }
+
+    /* signal here is multi token partition is enabled */
+    vp8_write_literal(bc, pc->multi_token_partition, 2);
+
+    /* Frame Qbaseline quantizer index */
+    vp8_write_literal(bc, pc->base_qindex, 7);
+
+    /* Transmit Dc, Second order and Uv quantizer delta information */
+    put_delta_q(bc, pc->y1dc_delta_q);
+    put_delta_q(bc, pc->y2dc_delta_q);
+    put_delta_q(bc, pc->y2ac_delta_q);
+    put_delta_q(bc, pc->uvdc_delta_q);
+    put_delta_q(bc, pc->uvac_delta_q);
+
+    /* When there is a key frame all reference buffers are updated using
+     * the new key frame
+     */
+    if (pc->frame_type != KEY_FRAME)
+    {
+        /* Should the GF or ARF be updated using the transmitted frame
+         * or buffer
+         */
+        vp8_write_bit(bc, pc->refresh_golden_frame);
+        vp8_write_bit(bc, pc->refresh_alt_ref_frame);
+
+        /* If not being updated from current frame should either GF or ARF
+         * be updated from another buffer
+         */
+        if (!pc->refresh_golden_frame)
+            vp8_write_literal(bc, pc->copy_buffer_to_gf, 2);
+
+        if (!pc->refresh_alt_ref_frame)
+            vp8_write_literal(bc, pc->copy_buffer_to_arf, 2);
+
+        /* Indicate reference frame sign bias for Golden and ARF frames
+         * (always 0 for last frame buffer)
+         */
+        vp8_write_bit(bc, pc->ref_frame_sign_bias[GOLDEN_FRAME]);
+        vp8_write_bit(bc, pc->ref_frame_sign_bias[ALTREF_FRAME]);
+    }
+
+#if !(CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+    if (cpi->oxcf.error_resilient_mode & VPX_ERROR_RESILIENT_PARTITIONS)
+    {
+        if (pc->frame_type == KEY_FRAME)
+            pc->refresh_entropy_probs = 1;
+        else
+            pc->refresh_entropy_probs = 0;
+    }
+#endif
+
+    vp8_write_bit(bc, pc->refresh_entropy_probs);
+
+    if (pc->frame_type != KEY_FRAME)
+        vp8_write_bit(bc, pc->refresh_last_frame);
+
+#ifdef VP8_ENTROPY_STATS
+
+    if (pc->frame_type == INTER_FRAME)
+        active_section = 0;
+    else
+        active_section = 7;
+
+#endif
+
+    vp8_clear_system_state();
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+    pack_coef_probs(cpi);
+#else
+    if (pc->refresh_entropy_probs == 0)
+    {
+        /* save a copy for later refresh */
+        memcpy(&cpi->common.lfc, &cpi->common.fc, sizeof(cpi->common.fc));
+    }
+
+    vp8_update_coef_probs(cpi);
+#endif
+
+#ifdef VP8_ENTROPY_STATS
+    active_section = 2;
+#endif
+
+    /* Write out the mb_no_coeff_skip flag */
+    vp8_write_bit(bc, pc->mb_no_coeff_skip);
+
+    if (pc->frame_type == KEY_FRAME)
+    {
+        write_kfmodes(cpi);
+
+#ifdef VP8_ENTROPY_STATS
+        active_section = 8;
+#endif
+    }
+    else
+    {
+        pack_inter_mode_mvs(cpi);
+
+#ifdef VP8_ENTROPY_STATS
+        active_section = 1;
+#endif
+    }
+
+    vp8_stop_encode(bc);
+
+    cx_data += bc->pos;
+
+    oh.first_partition_length_in_bytes = cpi->bc->pos;
+
+    /* update frame tag */
+    {
+        int v = (oh.first_partition_length_in_bytes << 5) |
+                (oh.show_frame << 4) |
+                (oh.version << 1) |
+                oh.type;
+
+        dest[0] = v;
+        dest[1] = v >> 8;
+        dest[2] = v >> 16;
+    }
+
+    *size = VP8_HEADER_SIZE + extra_bytes_packed + cpi->bc->pos;
+
+    cpi->partition_sz[0] = *size;
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+    {
+        const int num_part = (1 << pc->multi_token_partition);
+        unsigned char * dp = cpi->partition_d[0] + cpi->partition_sz[0];
+
+        if (num_part > 1)
+        {
+            /* write token part sizes (all but last) if more than 1 */
+            validate_buffer(dp, 3 * (num_part - 1), cpi->partition_d_end[0],
+                            &pc->error);
+
+            cpi->partition_sz[0] += 3*(num_part-1);
+
+            for(i = 1; i < num_part; i++)
+            {
+                write_partition_size(dp, cpi->partition_sz[i]);
+                dp += 3;
+            }
+        }
+
+        if (!cpi->output_partition)
+        {
+            /* concatenate partition buffers */
+            for(i = 0; i < num_part; i++)
+            {
+                memmove(dp, cpi->partition_d[i+1], cpi->partition_sz[i+1]);
+                cpi->partition_d[i+1] = dp;
+                dp += cpi->partition_sz[i+1];
+            }
+        }
+
+        /* update total size */
+        *size = 0;
+        for(i = 0; i < num_part+1; i++)
+        {
+            *size += cpi->partition_sz[i];
+        }
+    }
+#else
+    if (pc->multi_token_partition != ONE_PARTITION)
+    {
+        int num_part = 1 << pc->multi_token_partition;
+
+        /* partition size table at the end of first partition */
+        cpi->partition_sz[0] += 3 * (num_part - 1);
+        *size += 3 * (num_part - 1);
+
+        validate_buffer(cx_data, 3 * (num_part - 1), cx_data_end,
+                        &pc->error);
+
+        for(i = 1; i < num_part + 1; i++)
+        {
+            cpi->bc[i].error = &pc->error;
+        }
+
+        pack_tokens_into_partitions(cpi, cx_data + 3 * (num_part - 1),
+                                    cx_data_end, num_part);
+
+        for(i = 1; i < num_part; i++)
+        {
+            cpi->partition_sz[i] = cpi->bc[i].pos;
+            write_partition_size(cx_data, cpi->partition_sz[i]);
+            cx_data += 3;
+            *size += cpi->partition_sz[i]; /* add to total */
+        }
+
+        /* add last partition to total size */
+        cpi->partition_sz[i] = cpi->bc[i].pos;
+        *size += cpi->partition_sz[i];
+    }
+    else
+    {
+        bc[1].error = &pc->error;
+
+        vp8_start_encode(&cpi->bc[1], cx_data, cx_data_end);
+
+#if CONFIG_MULTITHREAD
+        if (cpi->b_multi_threaded)
+            pack_mb_row_tokens(cpi, &cpi->bc[1]);
+        else
+#endif  // CONFIG_MULTITHREAD
+            vp8_pack_tokens(&cpi->bc[1], cpi->tok, cpi->tok_count);
+
+        vp8_stop_encode(&cpi->bc[1]);
+
+        *size += cpi->bc[1].pos;
+        cpi->partition_sz[1] = cpi->bc[1].pos;
+    }
+#endif
+}
+
+#ifdef VP8_ENTROPY_STATS
+void print_tree_update_probs()
+{
+    int i, j, k, l;
+    FILE *f = fopen("context.c", "a");
+    int Sum;
+    fprintf(f, "\n/* Update probabilities for token entropy tree. */\n\n");
+    fprintf(f, "const vp8_prob tree_update_probs[BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES] = {\n");
+
+    for (i = 0; i < BLOCK_TYPES; i++)
+    {
+        fprintf(f, "  { \n");
+
+        for (j = 0; j < COEF_BANDS; j++)
+        {
+            fprintf(f, "    {\n");
+
+            for (k = 0; k < PREV_COEF_CONTEXTS; k++)
+            {
+                fprintf(f, "      {");
+
+                for (l = 0; l < ENTROPY_NODES; l++)
+                {
+                    Sum = tree_update_hist[i][j][k][l][0] + tree_update_hist[i][j][k][l][1];
+
+                    if (Sum > 0)
+                    {
+                        if (((tree_update_hist[i][j][k][l][0] * 255) / Sum) > 0)
+                            fprintf(f, "%3ld, ", (tree_update_hist[i][j][k][l][0] * 255) / Sum);
+                        else
+                            fprintf(f, "%3ld, ", 1);
+                    }
+                    else
+                        fprintf(f, "%3ld, ", 128);
+                }
+
+                fprintf(f, "},\n");
+            }
+
+            fprintf(f, "    },\n");
+        }
+
+        fprintf(f, "  },\n");
+    }
+
+    fprintf(f, "};\n");
+    fclose(f);
+}
+#endif
diff --git a/libvpx/libvpx/vp8/encoder/bitstream.h b/libvpx/libvpx/vp8/encoder/bitstream.h
new file mode 100644
index 0000000..de69805
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/bitstream.h
@@ -0,0 +1,25 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_BITSTREAM_H_
+#define VP8_ENCODER_BITSTREAM_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_pack_tokens(vp8_writer *w, const TOKENEXTRA *p, int xcount);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_BITSTREAM_H_
diff --git a/libvpx/libvpx/vp8/encoder/block.h b/libvpx/libvpx/vp8/encoder/block.h
new file mode 100644
index 0000000..248e795
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/block.h
@@ -0,0 +1,175 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_BLOCK_H_
+#define VP8_ENCODER_BLOCK_H_
+
+#include "vp8/common/onyx.h"
+#include "vp8/common/blockd.h"
+#include "vp8/common/entropymv.h"
+#include "vp8/common/entropy.h"
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_MODES 20
+#define MAX_ERROR_BINS 1024
+
+/* motion search site */
+typedef struct
+{
+    MV mv;
+    int offset;
+} search_site;
+
+typedef struct block
+{
+    /* 16 Y blocks, 4 U blocks, 4 V blocks each with 16 entries */
+    short *src_diff;
+    short *coeff;
+
+    /* 16 Y blocks, 4 U blocks, 4 V blocks each with 16 entries */
+    short *quant;
+    short *quant_fast;
+    short *quant_shift;
+    short *zbin;
+    short *zrun_zbin_boost;
+    short *round;
+
+    /* Zbin Over Quant value */
+    short zbin_extra;
+
+    unsigned char **base_src;
+    int src;
+    int src_stride;
+} BLOCK;
+
+typedef struct
+{
+    int count;
+    struct
+    {
+        B_PREDICTION_MODE mode;
+        int_mv mv;
+    } bmi[16];
+} PARTITION_INFO;
+
+typedef struct macroblock
+{
+    DECLARE_ALIGNED(16, short, src_diff[400]); /* 25 blocks Y,U,V,Y2 */
+    DECLARE_ALIGNED(16, short, coeff[400]); /* 25 blocks Y,U,V,Y2 */
+    DECLARE_ALIGNED(16, unsigned char, thismb[256]);
+
+    unsigned char *thismb_ptr;
+    /* 16 Y, 4 U, 4 V, 1 DC 2nd order block */
+    BLOCK block[25];
+
+    YV12_BUFFER_CONFIG src;
+
+    MACROBLOCKD e_mbd;
+    PARTITION_INFO *partition_info; /* work pointer */
+    PARTITION_INFO *pi;   /* Corresponds to upper left visible macroblock */
+    PARTITION_INFO *pip;  /* Base of allocated array */
+
+    int ref_frame_cost[MAX_REF_FRAMES];
+
+    search_site *ss;
+    int ss_count;
+    int searches_per_step;
+
+    int errorperbit;
+    int sadperbit16;
+    int sadperbit4;
+    int rddiv;
+    int rdmult;
+    unsigned int * mb_activity_ptr;
+    int * mb_norm_activity_ptr;
+    signed int act_zbin_adj;
+    signed int last_act_zbin_adj;
+
+    int *mvcost[2];
+    int *mvsadcost[2];
+    int (*mbmode_cost)[MB_MODE_COUNT];
+    int (*intra_uv_mode_cost)[MB_MODE_COUNT];
+    int (*bmode_costs)[10][10];
+    int *inter_bmode_costs;
+    int (*token_costs)[COEF_BANDS][PREV_COEF_CONTEXTS]
+    [MAX_ENTROPY_TOKENS];
+
+    /* These define limits to motion vector components to prevent
+     * them from extending outside the UMV borders.
+     */
+    int mv_col_min;
+    int mv_col_max;
+    int mv_row_min;
+    int mv_row_max;
+
+    int skip;
+
+    unsigned int encode_breakout;
+
+    signed char *gf_active_ptr;
+
+    unsigned char *active_ptr;
+    MV_CONTEXT *mvc;
+
+    int optimize;
+    int q_index;
+    int is_skin;
+    int denoise_zeromv;
+
+#if CONFIG_TEMPORAL_DENOISING
+    int increase_denoising;
+    MB_PREDICTION_MODE best_sse_inter_mode;
+    int_mv best_sse_mv;
+    MV_REFERENCE_FRAME best_reference_frame;
+    MV_REFERENCE_FRAME best_zeromv_reference_frame;
+    unsigned char need_to_clamp_best_mvs;
+#endif
+
+    int skip_true_count;
+    unsigned int coef_counts [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [MAX_ENTROPY_TOKENS];
+    unsigned int MVcount [2] [MVvals];  /* (row,col) MV cts this frame */
+    int ymode_count [VP8_YMODES];        /* intra MB type cts this frame */
+    int uv_mode_count[VP8_UV_MODES];     /* intra MB type cts this frame */
+    int64_t prediction_error;
+    int64_t intra_error;
+    int count_mb_ref_frame_usage[MAX_REF_FRAMES];
+
+    int rd_thresh_mult[MAX_MODES];
+    int rd_threshes[MAX_MODES];
+    unsigned int mbs_tested_so_far;
+    unsigned int mode_test_hit_counts[MAX_MODES];
+    int zbin_mode_boost_enabled;
+    int zbin_mode_boost;
+    int last_zbin_mode_boost;
+
+    int last_zbin_over_quant;
+    int zbin_over_quant;
+    int error_bins[MAX_ERROR_BINS];
+
+    void (*short_fdct4x4)(short *input, short *output, int pitch);
+    void (*short_fdct8x4)(short *input, short *output, int pitch);
+    void (*short_walsh4x4)(short *input, short *output, int pitch);
+    void (*quantize_b)(BLOCK *b, BLOCKD *d);
+
+    unsigned int mbs_zero_last_dot_suppress;
+    int zero_last_dot_suppress;
+} MACROBLOCK;
+
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_BLOCK_H_
diff --git a/libvpx/libvpx/vp8/encoder/boolhuff.c b/libvpx/libvpx/vp8/encoder/boolhuff.c
new file mode 100644
index 0000000..3b0c03a
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/boolhuff.c
@@ -0,0 +1,70 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "boolhuff.h"
+
+#if defined(SECTIONBITS_OUTPUT)
+unsigned __int64 Sectionbits[500];
+
+#endif
+
+#ifdef VP8_ENTROPY_STATS
+unsigned int active_section = 0;
+#endif
+
+const unsigned int vp8_prob_cost[256] =
+{
+    2047, 2047, 1791, 1641, 1535, 1452, 1385, 1328, 1279, 1235, 1196, 1161, 1129, 1099, 1072, 1046,
+    1023, 1000,  979,  959,  940,  922,  905,  889,  873,  858,  843,  829,  816,  803,  790,  778,
+    767,  755,  744,  733,  723,  713,  703,  693,  684,  675,  666,  657,  649,  641,  633,  625,
+    617,  609,  602,  594,  587,  580,  573,  567,  560,  553,  547,  541,  534,  528,  522,  516,
+    511,  505,  499,  494,  488,  483,  477,  472,  467,  462,  457,  452,  447,  442,  437,  433,
+    428,  424,  419,  415,  410,  406,  401,  397,  393,  389,  385,  381,  377,  373,  369,  365,
+    361,  357,  353,  349,  346,  342,  338,  335,  331,  328,  324,  321,  317,  314,  311,  307,
+    304,  301,  297,  294,  291,  288,  285,  281,  278,  275,  272,  269,  266,  263,  260,  257,
+    255,  252,  249,  246,  243,  240,  238,  235,  232,  229,  227,  224,  221,  219,  216,  214,
+    211,  208,  206,  203,  201,  198,  196,  194,  191,  189,  186,  184,  181,  179,  177,  174,
+    172,  170,  168,  165,  163,  161,  159,  156,  154,  152,  150,  148,  145,  143,  141,  139,
+    137,  135,  133,  131,  129,  127,  125,  123,  121,  119,  117,  115,  113,  111,  109,  107,
+    105,  103,  101,   99,   97,   95,   93,   92,   90,   88,   86,   84,   82,   81,   79,   77,
+    75,   73,   72,   70,   68,   66,   65,   63,   61,   60,   58,   56,   55,   53,   51,   50,
+    48,   46,   45,   43,   41,   40,   38,   37,   35,   33,   32,   30,   29,   27,   25,   24,
+    22,   21,   19,   18,   16,   15,   13,   12,   10,    9,    7,    6,    4,    3,    1,   1
+};
+
+void vp8_start_encode(BOOL_CODER *br, unsigned char *source, unsigned char *source_end)
+{
+
+    br->lowvalue   = 0;
+    br->range      = 255;
+    br->count      = -24;
+    br->buffer     = source;
+    br->buffer_end = source_end;
+    br->pos        = 0;
+}
+
+void vp8_stop_encode(BOOL_CODER *br)
+{
+    int i;
+
+    for (i = 0; i < 32; i++)
+        vp8_encode_bool(br, 0, 128);
+}
+
+
+void vp8_encode_value(BOOL_CODER *br, int data, int bits)
+{
+    int bit;
+
+    for (bit = bits - 1; bit >= 0; bit--)
+        vp8_encode_bool(br, (1 & (data >> bit)), 0x80);
+
+}
diff --git a/libvpx/libvpx/vp8/encoder/boolhuff.h b/libvpx/libvpx/vp8/encoder/boolhuff.h
new file mode 100644
index 0000000..e66a2db
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/boolhuff.h
@@ -0,0 +1,132 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+/****************************************************************************
+*
+*   Module Title :     boolhuff.h
+*
+*   Description  :     Bool Coder header file.
+*
+****************************************************************************/
+#ifndef VP8_ENCODER_BOOLHUFF_H_
+#define VP8_ENCODER_BOOLHUFF_H_
+
+#include "vpx_ports/mem.h"
+#include "vpx/internal/vpx_codec_internal.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct
+{
+    unsigned int lowvalue;
+    unsigned int range;
+    int count;
+    unsigned int pos;
+    unsigned char *buffer;
+    unsigned char *buffer_end;
+    struct vpx_internal_error_info *error;
+} BOOL_CODER;
+
+extern void vp8_start_encode(BOOL_CODER *bc, unsigned char *buffer, unsigned char *buffer_end);
+
+extern void vp8_encode_value(BOOL_CODER *br, int data, int bits);
+extern void vp8_stop_encode(BOOL_CODER *bc);
+extern const unsigned int vp8_prob_cost[256];
+
+
+DECLARE_ALIGNED(16, extern const unsigned char, vp8_norm[256]);
+
+static int validate_buffer(const unsigned char *start,
+                           size_t               len,
+                           const unsigned char *end,
+                           struct vpx_internal_error_info *error)
+{
+    if (start + len > start && start + len < end)
+        return 1;
+    else
+        vpx_internal_error(error, VPX_CODEC_CORRUPT_FRAME,
+            "Truncated packet or corrupt partition ");
+
+    return 0;
+}
+static void vp8_encode_bool(BOOL_CODER *br, int bit, int probability)
+{
+    unsigned int split;
+    int count = br->count;
+    unsigned int range = br->range;
+    unsigned int lowvalue = br->lowvalue;
+    register int shift;
+
+#ifdef VP8_ENTROPY_STATS
+#if defined(SECTIONBITS_OUTPUT)
+
+    if (bit)
+        Sectionbits[active_section] += vp8_prob_cost[255-probability];
+    else
+        Sectionbits[active_section] += vp8_prob_cost[probability];
+
+#endif
+#endif
+
+    split = 1 + (((range - 1) * probability) >> 8);
+
+    range = split;
+
+    if (bit)
+    {
+        lowvalue += split;
+        range = br->range - split;
+    }
+
+    shift = vp8_norm[range];
+
+    range <<= shift;
+    count += shift;
+
+    if (count >= 0)
+    {
+        int offset = shift - count;
+
+        if ((lowvalue << (offset - 1)) & 0x80000000)
+        {
+            int x = br->pos - 1;
+
+            while (x >= 0 && br->buffer[x] == 0xff)
+            {
+                br->buffer[x] = (unsigned char)0;
+                x--;
+            }
+
+            br->buffer[x] += 1;
+        }
+
+        validate_buffer(br->buffer + br->pos, 1, br->buffer_end, br->error);
+        br->buffer[br->pos++] = (lowvalue >> (24 - offset));
+
+        lowvalue <<= offset;
+        shift = count;
+        lowvalue &= 0xffffff;
+        count -= 8 ;
+    }
+
+    lowvalue <<= shift;
+    br->count = count;
+    br->lowvalue = lowvalue;
+    br->range = range;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_BOOLHUFF_H_
diff --git a/libvpx/libvpx/vp8/encoder/dct.c b/libvpx/libvpx/vp8/encoder/dct.c
new file mode 100644
index 0000000..0c7198d
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/dct.c
@@ -0,0 +1,118 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <math.h>
+
+#include "./vp8_rtcd.h"
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch)
+{
+    int i;
+    int a1, b1, c1, d1;
+    short *ip = input;
+    short *op = output;
+
+    for (i = 0; i < 4; i++)
+    {
+        a1 = ((ip[0] + ip[3]) * 8);
+        b1 = ((ip[1] + ip[2]) * 8);
+        c1 = ((ip[1] - ip[2]) * 8);
+        d1 = ((ip[0] - ip[3]) * 8);
+
+        op[0] = a1 + b1;
+        op[2] = a1 - b1;
+
+        op[1] = (c1 * 2217 + d1 * 5352 +  14500)>>12;
+        op[3] = (d1 * 2217 - c1 * 5352 +   7500)>>12;
+
+        ip += pitch / 2;
+        op += 4;
+
+    }
+    ip = output;
+    op = output;
+    for (i = 0; i < 4; i++)
+    {
+        a1 = ip[0] + ip[12];
+        b1 = ip[4] + ip[8];
+        c1 = ip[4] - ip[8];
+        d1 = ip[0] - ip[12];
+
+        op[0]  = ( a1 + b1 + 7)>>4;
+        op[8]  = ( a1 - b1 + 7)>>4;
+
+        op[4]  =((c1 * 2217 + d1 * 5352 +  12000)>>16) + (d1!=0);
+        op[12] = (d1 * 2217 - c1 * 5352 +  51000)>>16;
+
+        ip++;
+        op++;
+    }
+}
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch)
+{
+    vp8_short_fdct4x4_c(input,   output,    pitch);
+    vp8_short_fdct4x4_c(input + 4, output + 16, pitch);
+}
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch)
+{
+    int i;
+    int a1, b1, c1, d1;
+    int a2, b2, c2, d2;
+    short *ip = input;
+    short *op = output;
+
+
+    for (i = 0; i < 4; i++)
+    {
+        a1 = ((ip[0] + ip[2]) * 4);
+        d1 = ((ip[1] + ip[3]) * 4);
+        c1 = ((ip[1] - ip[3]) * 4);
+        b1 = ((ip[0] - ip[2]) * 4);
+
+        op[0] = a1 + d1 + (a1!=0);
+        op[1] = b1 + c1;
+        op[2] = b1 - c1;
+        op[3] = a1 - d1;
+        ip += pitch / 2;
+        op += 4;
+    }
+
+    ip = output;
+    op = output;
+
+    for (i = 0; i < 4; i++)
+    {
+        a1 = ip[0] + ip[8];
+        d1 = ip[4] + ip[12];
+        c1 = ip[4] - ip[12];
+        b1 = ip[0] - ip[8];
+
+        a2 = a1 + d1;
+        b2 = b1 + c1;
+        c2 = b1 - c1;
+        d2 = a1 - d1;
+
+        a2 += a2<0;
+        b2 += b2<0;
+        c2 += c2<0;
+        d2 += d2<0;
+
+        op[0] = (a2+3) >> 3;
+        op[4] = (b2+3) >> 3;
+        op[8] = (c2+3) >> 3;
+        op[12]= (d2+3) >> 3;
+
+        ip++;
+        op++;
+    }
+}
diff --git a/libvpx/libvpx/vp8/encoder/dct_value_cost.h b/libvpx/libvpx/vp8/encoder/dct_value_cost.h
new file mode 100644
index 0000000..1cd3eec
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/dct_value_cost.h
@@ -0,0 +1,371 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_ENCODER_DCT_VALUE_COST_H_
+#define VP8_ENCODER_DCT_VALUE_COST_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Generated file, included by tokenize.c  */
+/* Values generated by fill_value_tokens() */
+
+static const short dct_value_cost[2048*2] =
+{
+    8285, 8277, 8267, 8259, 8253, 8245, 8226, 8218, 8212, 8204, 8194, 8186,
+    8180, 8172, 8150, 8142, 8136, 8128, 8118, 8110, 8104, 8096, 8077, 8069,
+    8063, 8055, 8045, 8037, 8031, 8023, 7997, 7989, 7983, 7975, 7965, 7957,
+    7951, 7943, 7924, 7916, 7910, 7902, 7892, 7884, 7878, 7870, 7848, 7840,
+    7834, 7826, 7816, 7808, 7802, 7794, 7775, 7767, 7761, 7753, 7743, 7735,
+    7729, 7721, 7923, 7915, 7909, 7901, 7891, 7883, 7877, 7869, 7850, 7842,
+    7836, 7828, 7818, 7810, 7804, 7796, 7774, 7766, 7760, 7752, 7742, 7734,
+    7728, 7720, 7701, 7693, 7687, 7679, 7669, 7661, 7655, 7647, 7621, 7613,
+    7607, 7599, 7589, 7581, 7575, 7567, 7548, 7540, 7534, 7526, 7516, 7508,
+    7502, 7494, 7472, 7464, 7458, 7450, 7440, 7432, 7426, 7418, 7399, 7391,
+    7385, 7377, 7367, 7359, 7353, 7345, 7479, 7471, 7465, 7457, 7447, 7439,
+    7433, 7425, 7406, 7398, 7392, 7384, 7374, 7366, 7360, 7352, 7330, 7322,
+    7316, 7308, 7298, 7290, 7284, 7276, 7257, 7249, 7243, 7235, 7225, 7217,
+    7211, 7203, 7177, 7169, 7163, 7155, 7145, 7137, 7131, 7123, 7104, 7096,
+    7090, 7082, 7072, 7064, 7058, 7050, 7028, 7020, 7014, 7006, 6996, 6988,
+    6982, 6974, 6955, 6947, 6941, 6933, 6923, 6915, 6909, 6901, 7632, 7624,
+    7618, 7610, 7600, 7592, 7586, 7578, 7559, 7551, 7545, 7537, 7527, 7519,
+    7513, 7505, 7483, 7475, 7469, 7461, 7451, 7443, 7437, 7429, 7410, 7402,
+    7396, 7388, 7378, 7370, 7364, 7356, 7330, 7322, 7316, 7308, 7298, 7290,
+    7284, 7276, 7257, 7249, 7243, 7235, 7225, 7217, 7211, 7203, 7181, 7173,
+    7167, 7159, 7149, 7141, 7135, 7127, 7108, 7100, 7094, 7086, 7076, 7068,
+    7062, 7054, 7188, 7180, 7174, 7166, 7156, 7148, 7142, 7134, 7115, 7107,
+    7101, 7093, 7083, 7075, 7069, 7061, 7039, 7031, 7025, 7017, 7007, 6999,
+    6993, 6985, 6966, 6958, 6952, 6944, 6934, 6926, 6920, 6912, 6886, 6878,
+    6872, 6864, 6854, 6846, 6840, 6832, 6813, 6805, 6799, 6791, 6781, 6773,
+    6767, 6759, 6737, 6729, 6723, 6715, 6705, 6697, 6691, 6683, 6664, 6656,
+    6650, 6642, 6632, 6624, 6618, 6610, 6812, 6804, 6798, 6790, 6780, 6772,
+    6766, 6758, 6739, 6731, 6725, 6717, 6707, 6699, 6693, 6685, 6663, 6655,
+    6649, 6641, 6631, 6623, 6617, 6609, 6590, 6582, 6576, 6568, 6558, 6550,
+    6544, 6536, 6510, 6502, 6496, 6488, 6478, 6470, 6464, 6456, 6437, 6429,
+    6423, 6415, 6405, 6397, 6391, 6383, 6361, 6353, 6347, 6339, 6329, 6321,
+    6315, 6307, 6288, 6280, 6274, 6266, 6256, 6248, 6242, 6234, 6368, 6360,
+    6354, 6346, 6336, 6328, 6322, 6314, 6295, 6287, 6281, 6273, 6263, 6255,
+    6249, 6241, 6219, 6211, 6205, 6197, 6187, 6179, 6173, 6165, 6146, 6138,
+    6132, 6124, 6114, 6106, 6100, 6092, 6066, 6058, 6052, 6044, 6034, 6026,
+    6020, 6012, 5993, 5985, 5979, 5971, 5961, 5953, 5947, 5939, 5917, 5909,
+    5903, 5895, 5885, 5877, 5871, 5863, 5844, 5836, 5830, 5822, 5812, 5804,
+    5798, 5790, 6697, 6689, 6683, 6675, 6665, 6657, 6651, 6643, 6624, 6616,
+    6610, 6602, 6592, 6584, 6578, 6570, 6548, 6540, 6534, 6526, 6516, 6508,
+    6502, 6494, 6475, 6467, 6461, 6453, 6443, 6435, 6429, 6421, 6395, 6387,
+    6381, 6373, 6363, 6355, 6349, 6341, 6322, 6314, 6308, 6300, 6290, 6282,
+    6276, 6268, 6246, 6238, 6232, 6224, 6214, 6206, 6200, 6192, 6173, 6165,
+    6159, 6151, 6141, 6133, 6127, 6119, 6253, 6245, 6239, 6231, 6221, 6213,
+    6207, 6199, 6180, 6172, 6166, 6158, 6148, 6140, 6134, 6126, 6104, 6096,
+    6090, 6082, 6072, 6064, 6058, 6050, 6031, 6023, 6017, 6009, 5999, 5991,
+    5985, 5977, 5951, 5943, 5937, 5929, 5919, 5911, 5905, 5897, 5878, 5870,
+    5864, 5856, 5846, 5838, 5832, 5824, 5802, 5794, 5788, 5780, 5770, 5762,
+    5756, 5748, 5729, 5721, 5715, 5707, 5697, 5689, 5683, 5675, 5877, 5869,
+    5863, 5855, 5845, 5837, 5831, 5823, 5804, 5796, 5790, 5782, 5772, 5764,
+    5758, 5750, 5728, 5720, 5714, 5706, 5696, 5688, 5682, 5674, 5655, 5647,
+    5641, 5633, 5623, 5615, 5609, 5601, 5575, 5567, 5561, 5553, 5543, 5535,
+    5529, 5521, 5502, 5494, 5488, 5480, 5470, 5462, 5456, 5448, 5426, 5418,
+    5412, 5404, 5394, 5386, 5380, 5372, 5353, 5345, 5339, 5331, 5321, 5313,
+    5307, 5299, 5433, 5425, 5419, 5411, 5401, 5393, 5387, 5379, 5360, 5352,
+    5346, 5338, 5328, 5320, 5314, 5306, 5284, 5276, 5270, 5262, 5252, 5244,
+    5238, 5230, 5211, 5203, 5197, 5189, 5179, 5171, 5165, 5157, 5131, 5123,
+    5117, 5109, 5099, 5091, 5085, 5077, 5058, 5050, 5044, 5036, 5026, 5018,
+    5012, 5004, 4982, 4974, 4968, 4960, 4950, 4942, 4936, 4928, 4909, 4901,
+    4895, 4887, 4877, 4869, 4863, 4855, 5586, 5578, 5572, 5564, 5554, 5546,
+    5540, 5532, 5513, 5505, 5499, 5491, 5481, 5473, 5467, 5459, 5437, 5429,
+    5423, 5415, 5405, 5397, 5391, 5383, 5364, 5356, 5350, 5342, 5332, 5324,
+    5318, 5310, 5284, 5276, 5270, 5262, 5252, 5244, 5238, 5230, 5211, 5203,
+    5197, 5189, 5179, 5171, 5165, 5157, 5135, 5127, 5121, 5113, 5103, 5095,
+    5089, 5081, 5062, 5054, 5048, 5040, 5030, 5022, 5016, 5008, 5142, 5134,
+    5128, 5120, 5110, 5102, 5096, 5088, 5069, 5061, 5055, 5047, 5037, 5029,
+    5023, 5015, 4993, 4985, 4979, 4971, 4961, 4953, 4947, 4939, 4920, 4912,
+    4906, 4898, 4888, 4880, 4874, 4866, 4840, 4832, 4826, 4818, 4808, 4800,
+    4794, 4786, 4767, 4759, 4753, 4745, 4735, 4727, 4721, 4713, 4691, 4683,
+    4677, 4669, 4659, 4651, 4645, 4637, 4618, 4610, 4604, 4596, 4586, 4578,
+    4572, 4564, 4766, 4758, 4752, 4744, 4734, 4726, 4720, 4712, 4693, 4685,
+    4679, 4671, 4661, 4653, 4647, 4639, 4617, 4609, 4603, 4595, 4585, 4577,
+    4571, 4563, 4544, 4536, 4530, 4522, 4512, 4504, 4498, 4490, 4464, 4456,
+    4450, 4442, 4432, 4424, 4418, 4410, 4391, 4383, 4377, 4369, 4359, 4351,
+    4345, 4337, 4315, 4307, 4301, 4293, 4283, 4275, 4269, 4261, 4242, 4234,
+    4228, 4220, 4210, 4202, 4196, 4188, 4322, 4314, 4308, 4300, 4290, 4282,
+    4276, 4268, 4249, 4241, 4235, 4227, 4217, 4209, 4203, 4195, 4173, 4165,
+    4159, 4151, 4141, 4133, 4127, 4119, 4100, 4092, 4086, 4078, 4068, 4060,
+    4054, 4046, 4020, 4012, 4006, 3998, 3988, 3980, 3974, 3966, 3947, 3939,
+    3933, 3925, 3915, 3907, 3901, 3893, 3871, 3863, 3857, 3849, 3839, 3831,
+    3825, 3817, 3798, 3790, 3784, 3776, 3766, 3758, 3752, 3744, 6697, 6689,
+    6683, 6675, 6665, 6657, 6651, 6643, 6624, 6616, 6610, 6602, 6592, 6584,
+    6578, 6570, 6548, 6540, 6534, 6526, 6516, 6508, 6502, 6494, 6475, 6467,
+    6461, 6453, 6443, 6435, 6429, 6421, 6395, 6387, 6381, 6373, 6363, 6355,
+    6349, 6341, 6322, 6314, 6308, 6300, 6290, 6282, 6276, 6268, 6246, 6238,
+    6232, 6224, 6214, 6206, 6200, 6192, 6173, 6165, 6159, 6151, 6141, 6133,
+    6127, 6119, 6253, 6245, 6239, 6231, 6221, 6213, 6207, 6199, 6180, 6172,
+    6166, 6158, 6148, 6140, 6134, 6126, 6104, 6096, 6090, 6082, 6072, 6064,
+    6058, 6050, 6031, 6023, 6017, 6009, 5999, 5991, 5985, 5977, 5951, 5943,
+    5937, 5929, 5919, 5911, 5905, 5897, 5878, 5870, 5864, 5856, 5846, 5838,
+    5832, 5824, 5802, 5794, 5788, 5780, 5770, 5762, 5756, 5748, 5729, 5721,
+    5715, 5707, 5697, 5689, 5683, 5675, 5877, 5869, 5863, 5855, 5845, 5837,
+    5831, 5823, 5804, 5796, 5790, 5782, 5772, 5764, 5758, 5750, 5728, 5720,
+    5714, 5706, 5696, 5688, 5682, 5674, 5655, 5647, 5641, 5633, 5623, 5615,
+    5609, 5601, 5575, 5567, 5561, 5553, 5543, 5535, 5529, 5521, 5502, 5494,
+    5488, 5480, 5470, 5462, 5456, 5448, 5426, 5418, 5412, 5404, 5394, 5386,
+    5380, 5372, 5353, 5345, 5339, 5331, 5321, 5313, 5307, 5299, 5433, 5425,
+    5419, 5411, 5401, 5393, 5387, 5379, 5360, 5352, 5346, 5338, 5328, 5320,
+    5314, 5306, 5284, 5276, 5270, 5262, 5252, 5244, 5238, 5230, 5211, 5203,
+    5197, 5189, 5179, 5171, 5165, 5157, 5131, 5123, 5117, 5109, 5099, 5091,
+    5085, 5077, 5058, 5050, 5044, 5036, 5026, 5018, 5012, 5004, 4982, 4974,
+    4968, 4960, 4950, 4942, 4936, 4928, 4909, 4901, 4895, 4887, 4877, 4869,
+    4863, 4855, 5586, 5578, 5572, 5564, 5554, 5546, 5540, 5532, 5513, 5505,
+    5499, 5491, 5481, 5473, 5467, 5459, 5437, 5429, 5423, 5415, 5405, 5397,
+    5391, 5383, 5364, 5356, 5350, 5342, 5332, 5324, 5318, 5310, 5284, 5276,
+    5270, 5262, 5252, 5244, 5238, 5230, 5211, 5203, 5197, 5189, 5179, 5171,
+    5165, 5157, 5135, 5127, 5121, 5113, 5103, 5095, 5089, 5081, 5062, 5054,
+    5048, 5040, 5030, 5022, 5016, 5008, 5142, 5134, 5128, 5120, 5110, 5102,
+    5096, 5088, 5069, 5061, 5055, 5047, 5037, 5029, 5023, 5015, 4993, 4985,
+    4979, 4971, 4961, 4953, 4947, 4939, 4920, 4912, 4906, 4898, 4888, 4880,
+    4874, 4866, 4840, 4832, 4826, 4818, 4808, 4800, 4794, 4786, 4767, 4759,
+    4753, 4745, 4735, 4727, 4721, 4713, 4691, 4683, 4677, 4669, 4659, 4651,
+    4645, 4637, 4618, 4610, 4604, 4596, 4586, 4578, 4572, 4564, 4766, 4758,
+    4752, 4744, 4734, 4726, 4720, 4712, 4693, 4685, 4679, 4671, 4661, 4653,
+    4647, 4639, 4617, 4609, 4603, 4595, 4585, 4577, 4571, 4563, 4544, 4536,
+    4530, 4522, 4512, 4504, 4498, 4490, 4464, 4456, 4450, 4442, 4432, 4424,
+    4418, 4410, 4391, 4383, 4377, 4369, 4359, 4351, 4345, 4337, 4315, 4307,
+    4301, 4293, 4283, 4275, 4269, 4261, 4242, 4234, 4228, 4220, 4210, 4202,
+    4196, 4188, 4322, 4314, 4308, 4300, 4290, 4282, 4276, 4268, 4249, 4241,
+    4235, 4227, 4217, 4209, 4203, 4195, 4173, 4165, 4159, 4151, 4141, 4133,
+    4127, 4119, 4100, 4092, 4086, 4078, 4068, 4060, 4054, 4046, 4020, 4012,
+    4006, 3998, 3988, 3980, 3974, 3966, 3947, 3939, 3933, 3925, 3915, 3907,
+    3901, 3893, 3871, 3863, 3857, 3849, 3839, 3831, 3825, 3817, 3798, 3790,
+    3784, 3776, 3766, 3758, 3752, 3744, 4651, 4643, 4637, 4629, 4619, 4611,
+    4605, 4597, 4578, 4570, 4564, 4556, 4546, 4538, 4532, 4524, 4502, 4494,
+    4488, 4480, 4470, 4462, 4456, 4448, 4429, 4421, 4415, 4407, 4397, 4389,
+    4383, 4375, 4349, 4341, 4335, 4327, 4317, 4309, 4303, 4295, 4276, 4268,
+    4262, 4254, 4244, 4236, 4230, 4222, 4200, 4192, 4186, 4178, 4168, 4160,
+    4154, 4146, 4127, 4119, 4113, 4105, 4095, 4087, 4081, 4073, 4207, 4199,
+    4193, 4185, 4175, 4167, 4161, 4153, 4134, 4126, 4120, 4112, 4102, 4094,
+    4088, 4080, 4058, 4050, 4044, 4036, 4026, 4018, 4012, 4004, 3985, 3977,
+    3971, 3963, 3953, 3945, 3939, 3931, 3905, 3897, 3891, 3883, 3873, 3865,
+    3859, 3851, 3832, 3824, 3818, 3810, 3800, 3792, 3786, 3778, 3756, 3748,
+    3742, 3734, 3724, 3716, 3710, 3702, 3683, 3675, 3669, 3661, 3651, 3643,
+    3637, 3629, 3831, 3823, 3817, 3809, 3799, 3791, 3785, 3777, 3758, 3750,
+    3744, 3736, 3726, 3718, 3712, 3704, 3682, 3674, 3668, 3660, 3650, 3642,
+    3636, 3628, 3609, 3601, 3595, 3587, 3577, 3569, 3563, 3555, 3529, 3521,
+    3515, 3507, 3497, 3489, 3483, 3475, 3456, 3448, 3442, 3434, 3424, 3416,
+    3410, 3402, 3380, 3372, 3366, 3358, 3348, 3340, 3334, 3326, 3307, 3299,
+    3293, 3285, 3275, 3267, 3261, 3253, 3387, 3379, 3373, 3365, 3355, 3347,
+    3341, 3333, 3314, 3306, 3300, 3292, 3282, 3274, 3268, 3260, 3238, 3230,
+    3224, 3216, 3206, 3198, 3192, 3184, 3165, 3157, 3151, 3143, 3133, 3125,
+    3119, 3111, 3085, 3077, 3071, 3063, 3053, 3045, 3039, 3031, 3012, 3004,
+    2998, 2990, 2980, 2972, 2966, 2958, 2936, 2928, 2922, 2914, 2904, 2896,
+    2890, 2882, 2863, 2855, 2849, 2841, 2831, 2823, 2817, 2809, 3540, 3532,
+    3526, 3518, 3508, 3500, 3494, 3486, 3467, 3459, 3453, 3445, 3435, 3427,
+    3421, 3413, 3391, 3383, 3377, 3369, 3359, 3351, 3345, 3337, 3318, 3310,
+    3304, 3296, 3286, 3278, 3272, 3264, 3238, 3230, 3224, 3216, 3206, 3198,
+    3192, 3184, 3165, 3157, 3151, 3143, 3133, 3125, 3119, 3111, 3089, 3081,
+    3075, 3067, 3057, 3049, 3043, 3035, 3016, 3008, 3002, 2994, 2984, 2976,
+    2970, 2962, 3096, 3088, 3082, 3074, 3064, 3056, 3050, 3042, 3023, 3015,
+    3009, 3001, 2991, 2983, 2977, 2969, 2947, 2939, 2933, 2925, 2915, 2907,
+    2901, 2893, 2874, 2866, 2860, 2852, 2842, 2834, 2828, 2820, 2794, 2786,
+    2780, 2772, 2762, 2754, 2748, 2740, 2721, 2713, 2707, 2699, 2689, 2681,
+    2675, 2667, 2645, 2637, 2631, 2623, 2613, 2605, 2599, 2591, 2572, 2564,
+    2558, 2550, 2540, 2532, 2526, 2518, 2720, 2712, 2706, 2698, 2688, 2680,
+    2674, 2666, 2647, 2639, 2633, 2625, 2615, 2607, 2601, 2593, 2571, 2563,
+    2557, 2549, 2539, 2531, 2525, 2517, 2498, 2490, 2484, 2476, 2466, 2458,
+    2452, 2444, 2418, 2410, 2404, 2396, 2386, 2378, 2372, 2364, 2345, 2337,
+    2331, 2323, 2313, 2305, 2299, 2291, 2269, 2261, 2255, 2247, 2237, 2229,
+    2223, 2215, 2196, 2188, 2182, 2174, 2164, 2156, 2150, 2142, 2276, 2268,
+    2262, 2254, 2244, 2236, 2230, 2222, 2203, 2195, 2189, 2181, 2171, 2163,
+    2157, 2149, 2127, 2119, 2113, 2105, 2095, 2087, 2081, 2073, 2054, 2046,
+    2040, 2032, 2022, 2014, 2008, 2000, 1974, 1966, 1960, 1952, 1942, 1934,
+    1928, 1920, 1901, 1893, 1887, 1879, 1869, 1861, 1855, 1847, 1825, 1817,
+    1811, 1803, 1793, 1785, 1779, 1771, 1752, 1744, 1738, 1730, 1720, 1712,
+    1706, 1698, 1897, 1883, 1860, 1846, 1819, 1805, 1782, 1768, 1723, 1709,
+    1686, 1672, 1645, 1631, 1608, 1594, 1574, 1560, 1537, 1523, 1496, 1482,
+    1459, 1445, 1400, 1386, 1363, 1349, 1322, 1308, 1285, 1271, 1608, 1565,
+    1535, 1492, 1446, 1403, 1373, 1330, 1312, 1269, 1239, 1196, 1150, 1107,
+    1077, 1034, 1291, 1218, 1171, 1098, 1015, 942, 895, 822, 953, 850,
+    729, 626, 618, 431, 257, 257, 257, 257, 0, 255, 255, 255,
+    255, 429, 616, 624, 727, 848, 951, 820, 893, 940, 1013, 1096,
+    1169, 1216, 1289, 1032, 1075, 1105, 1148, 1194, 1237, 1267, 1310, 1328,
+    1371, 1401, 1444, 1490, 1533, 1563, 1606, 1269, 1283, 1306, 1320, 1347,
+    1361, 1384, 1398, 1443, 1457, 1480, 1494, 1521, 1535, 1558, 1572, 1592,
+    1606, 1629, 1643, 1670, 1684, 1707, 1721, 1766, 1780, 1803, 1817, 1844,
+    1858, 1881, 1895, 1696, 1704, 1710, 1718, 1728, 1736, 1742, 1750, 1769,
+    1777, 1783, 1791, 1801, 1809, 1815, 1823, 1845, 1853, 1859, 1867, 1877,
+    1885, 1891, 1899, 1918, 1926, 1932, 1940, 1950, 1958, 1964, 1972, 1998,
+    2006, 2012, 2020, 2030, 2038, 2044, 2052, 2071, 2079, 2085, 2093, 2103,
+    2111, 2117, 2125, 2147, 2155, 2161, 2169, 2179, 2187, 2193, 2201, 2220,
+    2228, 2234, 2242, 2252, 2260, 2266, 2274, 2140, 2148, 2154, 2162, 2172,
+    2180, 2186, 2194, 2213, 2221, 2227, 2235, 2245, 2253, 2259, 2267, 2289,
+    2297, 2303, 2311, 2321, 2329, 2335, 2343, 2362, 2370, 2376, 2384, 2394,
+    2402, 2408, 2416, 2442, 2450, 2456, 2464, 2474, 2482, 2488, 2496, 2515,
+    2523, 2529, 2537, 2547, 2555, 2561, 2569, 2591, 2599, 2605, 2613, 2623,
+    2631, 2637, 2645, 2664, 2672, 2678, 2686, 2696, 2704, 2710, 2718, 2516,
+    2524, 2530, 2538, 2548, 2556, 2562, 2570, 2589, 2597, 2603, 2611, 2621,
+    2629, 2635, 2643, 2665, 2673, 2679, 2687, 2697, 2705, 2711, 2719, 2738,
+    2746, 2752, 2760, 2770, 2778, 2784, 2792, 2818, 2826, 2832, 2840, 2850,
+    2858, 2864, 2872, 2891, 2899, 2905, 2913, 2923, 2931, 2937, 2945, 2967,
+    2975, 2981, 2989, 2999, 3007, 3013, 3021, 3040, 3048, 3054, 3062, 3072,
+    3080, 3086, 3094, 2960, 2968, 2974, 2982, 2992, 3000, 3006, 3014, 3033,
+    3041, 3047, 3055, 3065, 3073, 3079, 3087, 3109, 3117, 3123, 3131, 3141,
+    3149, 3155, 3163, 3182, 3190, 3196, 3204, 3214, 3222, 3228, 3236, 3262,
+    3270, 3276, 3284, 3294, 3302, 3308, 3316, 3335, 3343, 3349, 3357, 3367,
+    3375, 3381, 3389, 3411, 3419, 3425, 3433, 3443, 3451, 3457, 3465, 3484,
+    3492, 3498, 3506, 3516, 3524, 3530, 3538, 2807, 2815, 2821, 2829, 2839,
+    2847, 2853, 2861, 2880, 2888, 2894, 2902, 2912, 2920, 2926, 2934, 2956,
+    2964, 2970, 2978, 2988, 2996, 3002, 3010, 3029, 3037, 3043, 3051, 3061,
+    3069, 3075, 3083, 3109, 3117, 3123, 3131, 3141, 3149, 3155, 3163, 3182,
+    3190, 3196, 3204, 3214, 3222, 3228, 3236, 3258, 3266, 3272, 3280, 3290,
+    3298, 3304, 3312, 3331, 3339, 3345, 3353, 3363, 3371, 3377, 3385, 3251,
+    3259, 3265, 3273, 3283, 3291, 3297, 3305, 3324, 3332, 3338, 3346, 3356,
+    3364, 3370, 3378, 3400, 3408, 3414, 3422, 3432, 3440, 3446, 3454, 3473,
+    3481, 3487, 3495, 3505, 3513, 3519, 3527, 3553, 3561, 3567, 3575, 3585,
+    3593, 3599, 3607, 3626, 3634, 3640, 3648, 3658, 3666, 3672, 3680, 3702,
+    3710, 3716, 3724, 3734, 3742, 3748, 3756, 3775, 3783, 3789, 3797, 3807,
+    3815, 3821, 3829, 3627, 3635, 3641, 3649, 3659, 3667, 3673, 3681, 3700,
+    3708, 3714, 3722, 3732, 3740, 3746, 3754, 3776, 3784, 3790, 3798, 3808,
+    3816, 3822, 3830, 3849, 3857, 3863, 3871, 3881, 3889, 3895, 3903, 3929,
+    3937, 3943, 3951, 3961, 3969, 3975, 3983, 4002, 4010, 4016, 4024, 4034,
+    4042, 4048, 4056, 4078, 4086, 4092, 4100, 4110, 4118, 4124, 4132, 4151,
+    4159, 4165, 4173, 4183, 4191, 4197, 4205, 4071, 4079, 4085, 4093, 4103,
+    4111, 4117, 4125, 4144, 4152, 4158, 4166, 4176, 4184, 4190, 4198, 4220,
+    4228, 4234, 4242, 4252, 4260, 4266, 4274, 4293, 4301, 4307, 4315, 4325,
+    4333, 4339, 4347, 4373, 4381, 4387, 4395, 4405, 4413, 4419, 4427, 4446,
+    4454, 4460, 4468, 4478, 4486, 4492, 4500, 4522, 4530, 4536, 4544, 4554,
+    4562, 4568, 4576, 4595, 4603, 4609, 4617, 4627, 4635, 4641, 4649, 3742,
+    3750, 3756, 3764, 3774, 3782, 3788, 3796, 3815, 3823, 3829, 3837, 3847,
+    3855, 3861, 3869, 3891, 3899, 3905, 3913, 3923, 3931, 3937, 3945, 3964,
+    3972, 3978, 3986, 3996, 4004, 4010, 4018, 4044, 4052, 4058, 4066, 4076,
+    4084, 4090, 4098, 4117, 4125, 4131, 4139, 4149, 4157, 4163, 4171, 4193,
+    4201, 4207, 4215, 4225, 4233, 4239, 4247, 4266, 4274, 4280, 4288, 4298,
+    4306, 4312, 4320, 4186, 4194, 4200, 4208, 4218, 4226, 4232, 4240, 4259,
+    4267, 4273, 4281, 4291, 4299, 4305, 4313, 4335, 4343, 4349, 4357, 4367,
+    4375, 4381, 4389, 4408, 4416, 4422, 4430, 4440, 4448, 4454, 4462, 4488,
+    4496, 4502, 4510, 4520, 4528, 4534, 4542, 4561, 4569, 4575, 4583, 4593,
+    4601, 4607, 4615, 4637, 4645, 4651, 4659, 4669, 4677, 4683, 4691, 4710,
+    4718, 4724, 4732, 4742, 4750, 4756, 4764, 4562, 4570, 4576, 4584, 4594,
+    4602, 4608, 4616, 4635, 4643, 4649, 4657, 4667, 4675, 4681, 4689, 4711,
+    4719, 4725, 4733, 4743, 4751, 4757, 4765, 4784, 4792, 4798, 4806, 4816,
+    4824, 4830, 4838, 4864, 4872, 4878, 4886, 4896, 4904, 4910, 4918, 4937,
+    4945, 4951, 4959, 4969, 4977, 4983, 4991, 5013, 5021, 5027, 5035, 5045,
+    5053, 5059, 5067, 5086, 5094, 5100, 5108, 5118, 5126, 5132, 5140, 5006,
+    5014, 5020, 5028, 5038, 5046, 5052, 5060, 5079, 5087, 5093, 5101, 5111,
+    5119, 5125, 5133, 5155, 5163, 5169, 5177, 5187, 5195, 5201, 5209, 5228,
+    5236, 5242, 5250, 5260, 5268, 5274, 5282, 5308, 5316, 5322, 5330, 5340,
+    5348, 5354, 5362, 5381, 5389, 5395, 5403, 5413, 5421, 5427, 5435, 5457,
+    5465, 5471, 5479, 5489, 5497, 5503, 5511, 5530, 5538, 5544, 5552, 5562,
+    5570, 5576, 5584, 4853, 4861, 4867, 4875, 4885, 4893, 4899, 4907, 4926,
+    4934, 4940, 4948, 4958, 4966, 4972, 4980, 5002, 5010, 5016, 5024, 5034,
+    5042, 5048, 5056, 5075, 5083, 5089, 5097, 5107, 5115, 5121, 5129, 5155,
+    5163, 5169, 5177, 5187, 5195, 5201, 5209, 5228, 5236, 5242, 5250, 5260,
+    5268, 5274, 5282, 5304, 5312, 5318, 5326, 5336, 5344, 5350, 5358, 5377,
+    5385, 5391, 5399, 5409, 5417, 5423, 5431, 5297, 5305, 5311, 5319, 5329,
+    5337, 5343, 5351, 5370, 5378, 5384, 5392, 5402, 5410, 5416, 5424, 5446,
+    5454, 5460, 5468, 5478, 5486, 5492, 5500, 5519, 5527, 5533, 5541, 5551,
+    5559, 5565, 5573, 5599, 5607, 5613, 5621, 5631, 5639, 5645, 5653, 5672,
+    5680, 5686, 5694, 5704, 5712, 5718, 5726, 5748, 5756, 5762, 5770, 5780,
+    5788, 5794, 5802, 5821, 5829, 5835, 5843, 5853, 5861, 5867, 5875, 5673,
+    5681, 5687, 5695, 5705, 5713, 5719, 5727, 5746, 5754, 5760, 5768, 5778,
+    5786, 5792, 5800, 5822, 5830, 5836, 5844, 5854, 5862, 5868, 5876, 5895,
+    5903, 5909, 5917, 5927, 5935, 5941, 5949, 5975, 5983, 5989, 5997, 6007,
+    6015, 6021, 6029, 6048, 6056, 6062, 6070, 6080, 6088, 6094, 6102, 6124,
+    6132, 6138, 6146, 6156, 6164, 6170, 6178, 6197, 6205, 6211, 6219, 6229,
+    6237, 6243, 6251, 6117, 6125, 6131, 6139, 6149, 6157, 6163, 6171, 6190,
+    6198, 6204, 6212, 6222, 6230, 6236, 6244, 6266, 6274, 6280, 6288, 6298,
+    6306, 6312, 6320, 6339, 6347, 6353, 6361, 6371, 6379, 6385, 6393, 6419,
+    6427, 6433, 6441, 6451, 6459, 6465, 6473, 6492, 6500, 6506, 6514, 6524,
+    6532, 6538, 6546, 6568, 6576, 6582, 6590, 6600, 6608, 6614, 6622, 6641,
+    6649, 6655, 6663, 6673, 6681, 6687, 6695, 3742, 3750, 3756, 3764, 3774,
+    3782, 3788, 3796, 3815, 3823, 3829, 3837, 3847, 3855, 3861, 3869, 3891,
+    3899, 3905, 3913, 3923, 3931, 3937, 3945, 3964, 3972, 3978, 3986, 3996,
+    4004, 4010, 4018, 4044, 4052, 4058, 4066, 4076, 4084, 4090, 4098, 4117,
+    4125, 4131, 4139, 4149, 4157, 4163, 4171, 4193, 4201, 4207, 4215, 4225,
+    4233, 4239, 4247, 4266, 4274, 4280, 4288, 4298, 4306, 4312, 4320, 4186,
+    4194, 4200, 4208, 4218, 4226, 4232, 4240, 4259, 4267, 4273, 4281, 4291,
+    4299, 4305, 4313, 4335, 4343, 4349, 4357, 4367, 4375, 4381, 4389, 4408,
+    4416, 4422, 4430, 4440, 4448, 4454, 4462, 4488, 4496, 4502, 4510, 4520,
+    4528, 4534, 4542, 4561, 4569, 4575, 4583, 4593, 4601, 4607, 4615, 4637,
+    4645, 4651, 4659, 4669, 4677, 4683, 4691, 4710, 4718, 4724, 4732, 4742,
+    4750, 4756, 4764, 4562, 4570, 4576, 4584, 4594, 4602, 4608, 4616, 4635,
+    4643, 4649, 4657, 4667, 4675, 4681, 4689, 4711, 4719, 4725, 4733, 4743,
+    4751, 4757, 4765, 4784, 4792, 4798, 4806, 4816, 4824, 4830, 4838, 4864,
+    4872, 4878, 4886, 4896, 4904, 4910, 4918, 4937, 4945, 4951, 4959, 4969,
+    4977, 4983, 4991, 5013, 5021, 5027, 5035, 5045, 5053, 5059, 5067, 5086,
+    5094, 5100, 5108, 5118, 5126, 5132, 5140, 5006, 5014, 5020, 5028, 5038,
+    5046, 5052, 5060, 5079, 5087, 5093, 5101, 5111, 5119, 5125, 5133, 5155,
+    5163, 5169, 5177, 5187, 5195, 5201, 5209, 5228, 5236, 5242, 5250, 5260,
+    5268, 5274, 5282, 5308, 5316, 5322, 5330, 5340, 5348, 5354, 5362, 5381,
+    5389, 5395, 5403, 5413, 5421, 5427, 5435, 5457, 5465, 5471, 5479, 5489,
+    5497, 5503, 5511, 5530, 5538, 5544, 5552, 5562, 5570, 5576, 5584, 4853,
+    4861, 4867, 4875, 4885, 4893, 4899, 4907, 4926, 4934, 4940, 4948, 4958,
+    4966, 4972, 4980, 5002, 5010, 5016, 5024, 5034, 5042, 5048, 5056, 5075,
+    5083, 5089, 5097, 5107, 5115, 5121, 5129, 5155, 5163, 5169, 5177, 5187,
+    5195, 5201, 5209, 5228, 5236, 5242, 5250, 5260, 5268, 5274, 5282, 5304,
+    5312, 5318, 5326, 5336, 5344, 5350, 5358, 5377, 5385, 5391, 5399, 5409,
+    5417, 5423, 5431, 5297, 5305, 5311, 5319, 5329, 5337, 5343, 5351, 5370,
+    5378, 5384, 5392, 5402, 5410, 5416, 5424, 5446, 5454, 5460, 5468, 5478,
+    5486, 5492, 5500, 5519, 5527, 5533, 5541, 5551, 5559, 5565, 5573, 5599,
+    5607, 5613, 5621, 5631, 5639, 5645, 5653, 5672, 5680, 5686, 5694, 5704,
+    5712, 5718, 5726, 5748, 5756, 5762, 5770, 5780, 5788, 5794, 5802, 5821,
+    5829, 5835, 5843, 5853, 5861, 5867, 5875, 5673, 5681, 5687, 5695, 5705,
+    5713, 5719, 5727, 5746, 5754, 5760, 5768, 5778, 5786, 5792, 5800, 5822,
+    5830, 5836, 5844, 5854, 5862, 5868, 5876, 5895, 5903, 5909, 5917, 5927,
+    5935, 5941, 5949, 5975, 5983, 5989, 5997, 6007, 6015, 6021, 6029, 6048,
+    6056, 6062, 6070, 6080, 6088, 6094, 6102, 6124, 6132, 6138, 6146, 6156,
+    6164, 6170, 6178, 6197, 6205, 6211, 6219, 6229, 6237, 6243, 6251, 6117,
+    6125, 6131, 6139, 6149, 6157, 6163, 6171, 6190, 6198, 6204, 6212, 6222,
+    6230, 6236, 6244, 6266, 6274, 6280, 6288, 6298, 6306, 6312, 6320, 6339,
+    6347, 6353, 6361, 6371, 6379, 6385, 6393, 6419, 6427, 6433, 6441, 6451,
+    6459, 6465, 6473, 6492, 6500, 6506, 6514, 6524, 6532, 6538, 6546, 6568,
+    6576, 6582, 6590, 6600, 6608, 6614, 6622, 6641, 6649, 6655, 6663, 6673,
+    6681, 6687, 6695, 5788, 5796, 5802, 5810, 5820, 5828, 5834, 5842, 5861,
+    5869, 5875, 5883, 5893, 5901, 5907, 5915, 5937, 5945, 5951, 5959, 5969,
+    5977, 5983, 5991, 6010, 6018, 6024, 6032, 6042, 6050, 6056, 6064, 6090,
+    6098, 6104, 6112, 6122, 6130, 6136, 6144, 6163, 6171, 6177, 6185, 6195,
+    6203, 6209, 6217, 6239, 6247, 6253, 6261, 6271, 6279, 6285, 6293, 6312,
+    6320, 6326, 6334, 6344, 6352, 6358, 6366, 6232, 6240, 6246, 6254, 6264,
+    6272, 6278, 6286, 6305, 6313, 6319, 6327, 6337, 6345, 6351, 6359, 6381,
+    6389, 6395, 6403, 6413, 6421, 6427, 6435, 6454, 6462, 6468, 6476, 6486,
+    6494, 6500, 6508, 6534, 6542, 6548, 6556, 6566, 6574, 6580, 6588, 6607,
+    6615, 6621, 6629, 6639, 6647, 6653, 6661, 6683, 6691, 6697, 6705, 6715,
+    6723, 6729, 6737, 6756, 6764, 6770, 6778, 6788, 6796, 6802, 6810, 6608,
+    6616, 6622, 6630, 6640, 6648, 6654, 6662, 6681, 6689, 6695, 6703, 6713,
+    6721, 6727, 6735, 6757, 6765, 6771, 6779, 6789, 6797, 6803, 6811, 6830,
+    6838, 6844, 6852, 6862, 6870, 6876, 6884, 6910, 6918, 6924, 6932, 6942,
+    6950, 6956, 6964, 6983, 6991, 6997, 7005, 7015, 7023, 7029, 7037, 7059,
+    7067, 7073, 7081, 7091, 7099, 7105, 7113, 7132, 7140, 7146, 7154, 7164,
+    7172, 7178, 7186, 7052, 7060, 7066, 7074, 7084, 7092, 7098, 7106, 7125,
+    7133, 7139, 7147, 7157, 7165, 7171, 7179, 7201, 7209, 7215, 7223, 7233,
+    7241, 7247, 7255, 7274, 7282, 7288, 7296, 7306, 7314, 7320, 7328, 7354,
+    7362, 7368, 7376, 7386, 7394, 7400, 7408, 7427, 7435, 7441, 7449, 7459,
+    7467, 7473, 7481, 7503, 7511, 7517, 7525, 7535, 7543, 7549, 7557, 7576,
+    7584, 7590, 7598, 7608, 7616, 7622, 7630, 6899, 6907, 6913, 6921, 6931,
+    6939, 6945, 6953, 6972, 6980, 6986, 6994, 7004, 7012, 7018, 7026, 7048,
+    7056, 7062, 7070, 7080, 7088, 7094, 7102, 7121, 7129, 7135, 7143, 7153,
+    7161, 7167, 7175, 7201, 7209, 7215, 7223, 7233, 7241, 7247, 7255, 7274,
+    7282, 7288, 7296, 7306, 7314, 7320, 7328, 7350, 7358, 7364, 7372, 7382,
+    7390, 7396, 7404, 7423, 7431, 7437, 7445, 7455, 7463, 7469, 7477, 7343,
+    7351, 7357, 7365, 7375, 7383, 7389, 7397, 7416, 7424, 7430, 7438, 7448,
+    7456, 7462, 7470, 7492, 7500, 7506, 7514, 7524, 7532, 7538, 7546, 7565,
+    7573, 7579, 7587, 7597, 7605, 7611, 7619, 7645, 7653, 7659, 7667, 7677,
+    7685, 7691, 7699, 7718, 7726, 7732, 7740, 7750, 7758, 7764, 7772, 7794,
+    7802, 7808, 7816, 7826, 7834, 7840, 7848, 7867, 7875, 7881, 7889, 7899,
+    7907, 7913, 7921, 7719, 7727, 7733, 7741, 7751, 7759, 7765, 7773, 7792,
+    7800, 7806, 7814, 7824, 7832, 7838, 7846, 7868, 7876, 7882, 7890, 7900,
+    7908, 7914, 7922, 7941, 7949, 7955, 7963, 7973, 7981, 7987, 7995, 8021,
+    8029, 8035, 8043, 8053, 8061, 8067, 8075, 8094, 8102, 8108, 8116, 8126,
+    8134, 8140, 8148, 8170, 8178, 8184, 8192, 8202, 8210, 8216, 8224, 8243,
+    8251, 8257, 8265, 8275
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_DCT_VALUE_COST_H_
diff --git a/libvpx/libvpx/vp8/encoder/dct_value_tokens.h b/libvpx/libvpx/vp8/encoder/dct_value_tokens.h
new file mode 100644
index 0000000..c2aadef
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/dct_value_tokens.h
@@ -0,0 +1,712 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_ENCODER_DCT_VALUE_TOKENS_H_
+#define VP8_ENCODER_DCT_VALUE_TOKENS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Generated file, included by tokenize.c  */
+/* Values generated by fill_value_tokens() */
+
+static const TOKENVALUE dct_value_tokens[2048*2] =
+{
+    {10, 3963}, {10, 3961}, {10, 3959}, {10, 3957}, {10, 3955}, {10, 3953},
+    {10, 3951}, {10, 3949}, {10, 3947}, {10, 3945}, {10, 3943}, {10, 3941},
+    {10, 3939}, {10, 3937}, {10, 3935}, {10, 3933}, {10, 3931}, {10, 3929},
+    {10, 3927}, {10, 3925}, {10, 3923}, {10, 3921}, {10, 3919}, {10, 3917},
+    {10, 3915}, {10, 3913}, {10, 3911}, {10, 3909}, {10, 3907}, {10, 3905},
+    {10, 3903}, {10, 3901}, {10, 3899}, {10, 3897}, {10, 3895}, {10, 3893},
+    {10, 3891}, {10, 3889}, {10, 3887}, {10, 3885}, {10, 3883}, {10, 3881},
+    {10, 3879}, {10, 3877}, {10, 3875}, {10, 3873}, {10, 3871}, {10, 3869},
+    {10, 3867}, {10, 3865}, {10, 3863}, {10, 3861}, {10, 3859}, {10, 3857},
+    {10, 3855}, {10, 3853}, {10, 3851}, {10, 3849}, {10, 3847}, {10, 3845},
+    {10, 3843}, {10, 3841}, {10, 3839}, {10, 3837}, {10, 3835}, {10, 3833},
+    {10, 3831}, {10, 3829}, {10, 3827}, {10, 3825}, {10, 3823}, {10, 3821},
+    {10, 3819}, {10, 3817}, {10, 3815}, {10, 3813}, {10, 3811}, {10, 3809},
+    {10, 3807}, {10, 3805}, {10, 3803}, {10, 3801}, {10, 3799}, {10, 3797},
+    {10, 3795}, {10, 3793}, {10, 3791}, {10, 3789}, {10, 3787}, {10, 3785},
+    {10, 3783}, {10, 3781}, {10, 3779}, {10, 3777}, {10, 3775}, {10, 3773},
+    {10, 3771}, {10, 3769}, {10, 3767}, {10, 3765}, {10, 3763}, {10, 3761},
+    {10, 3759}, {10, 3757}, {10, 3755}, {10, 3753}, {10, 3751}, {10, 3749},
+    {10, 3747}, {10, 3745}, {10, 3743}, {10, 3741}, {10, 3739}, {10, 3737},
+    {10, 3735}, {10, 3733}, {10, 3731}, {10, 3729}, {10, 3727}, {10, 3725},
+    {10, 3723}, {10, 3721}, {10, 3719}, {10, 3717}, {10, 3715}, {10, 3713},
+    {10, 3711}, {10, 3709}, {10, 3707}, {10, 3705}, {10, 3703}, {10, 3701},
+    {10, 3699}, {10, 3697}, {10, 3695}, {10, 3693}, {10, 3691}, {10, 3689},
+    {10, 3687}, {10, 3685}, {10, 3683}, {10, 3681}, {10, 3679}, {10, 3677},
+    {10, 3675}, {10, 3673}, {10, 3671}, {10, 3669}, {10, 3667}, {10, 3665},
+    {10, 3663}, {10, 3661}, {10, 3659}, {10, 3657}, {10, 3655}, {10, 3653},
+    {10, 3651}, {10, 3649}, {10, 3647}, {10, 3645}, {10, 3643}, {10, 3641},
+    {10, 3639}, {10, 3637}, {10, 3635}, {10, 3633}, {10, 3631}, {10, 3629},
+    {10, 3627}, {10, 3625}, {10, 3623}, {10, 3621}, {10, 3619}, {10, 3617},
+    {10, 3615}, {10, 3613}, {10, 3611}, {10, 3609}, {10, 3607}, {10, 3605},
+    {10, 3603}, {10, 3601}, {10, 3599}, {10, 3597}, {10, 3595}, {10, 3593},
+    {10, 3591}, {10, 3589}, {10, 3587}, {10, 3585}, {10, 3583}, {10, 3581},
+    {10, 3579}, {10, 3577}, {10, 3575}, {10, 3573}, {10, 3571}, {10, 3569},
+    {10, 3567}, {10, 3565}, {10, 3563}, {10, 3561}, {10, 3559}, {10, 3557},
+    {10, 3555}, {10, 3553}, {10, 3551}, {10, 3549}, {10, 3547}, {10, 3545},
+    {10, 3543}, {10, 3541}, {10, 3539}, {10, 3537}, {10, 3535}, {10, 3533},
+    {10, 3531}, {10, 3529}, {10, 3527}, {10, 3525}, {10, 3523}, {10, 3521},
+    {10, 3519}, {10, 3517}, {10, 3515}, {10, 3513}, {10, 3511}, {10, 3509},
+    {10, 3507}, {10, 3505}, {10, 3503}, {10, 3501}, {10, 3499}, {10, 3497},
+    {10, 3495}, {10, 3493}, {10, 3491}, {10, 3489}, {10, 3487}, {10, 3485},
+    {10, 3483}, {10, 3481}, {10, 3479}, {10, 3477}, {10, 3475}, {10, 3473},
+    {10, 3471}, {10, 3469}, {10, 3467}, {10, 3465}, {10, 3463}, {10, 3461},
+    {10, 3459}, {10, 3457}, {10, 3455}, {10, 3453}, {10, 3451}, {10, 3449},
+    {10, 3447}, {10, 3445}, {10, 3443}, {10, 3441}, {10, 3439}, {10, 3437},
+    {10, 3435}, {10, 3433}, {10, 3431}, {10, 3429}, {10, 3427}, {10, 3425},
+    {10, 3423}, {10, 3421}, {10, 3419}, {10, 3417}, {10, 3415}, {10, 3413},
+    {10, 3411}, {10, 3409}, {10, 3407}, {10, 3405}, {10, 3403}, {10, 3401},
+    {10, 3399}, {10, 3397}, {10, 3395}, {10, 3393}, {10, 3391}, {10, 3389},
+    {10, 3387}, {10, 3385}, {10, 3383}, {10, 3381}, {10, 3379}, {10, 3377},
+    {10, 3375}, {10, 3373}, {10, 3371}, {10, 3369}, {10, 3367}, {10, 3365},
+    {10, 3363}, {10, 3361}, {10, 3359}, {10, 3357}, {10, 3355}, {10, 3353},
+    {10, 3351}, {10, 3349}, {10, 3347}, {10, 3345}, {10, 3343}, {10, 3341},
+    {10, 3339}, {10, 3337}, {10, 3335}, {10, 3333}, {10, 3331}, {10, 3329},
+    {10, 3327}, {10, 3325}, {10, 3323}, {10, 3321}, {10, 3319}, {10, 3317},
+    {10, 3315}, {10, 3313}, {10, 3311}, {10, 3309}, {10, 3307}, {10, 3305},
+    {10, 3303}, {10, 3301}, {10, 3299}, {10, 3297}, {10, 3295}, {10, 3293},
+    {10, 3291}, {10, 3289}, {10, 3287}, {10, 3285}, {10, 3283}, {10, 3281},
+    {10, 3279}, {10, 3277}, {10, 3275}, {10, 3273}, {10, 3271}, {10, 3269},
+    {10, 3267}, {10, 3265}, {10, 3263}, {10, 3261}, {10, 3259}, {10, 3257},
+    {10, 3255}, {10, 3253}, {10, 3251}, {10, 3249}, {10, 3247}, {10, 3245},
+    {10, 3243}, {10, 3241}, {10, 3239}, {10, 3237}, {10, 3235}, {10, 3233},
+    {10, 3231}, {10, 3229}, {10, 3227}, {10, 3225}, {10, 3223}, {10, 3221},
+    {10, 3219}, {10, 3217}, {10, 3215}, {10, 3213}, {10, 3211}, {10, 3209},
+    {10, 3207}, {10, 3205}, {10, 3203}, {10, 3201}, {10, 3199}, {10, 3197},
+    {10, 3195}, {10, 3193}, {10, 3191}, {10, 3189}, {10, 3187}, {10, 3185},
+    {10, 3183}, {10, 3181}, {10, 3179}, {10, 3177}, {10, 3175}, {10, 3173},
+    {10, 3171}, {10, 3169}, {10, 3167}, {10, 3165}, {10, 3163}, {10, 3161},
+    {10, 3159}, {10, 3157}, {10, 3155}, {10, 3153}, {10, 3151}, {10, 3149},
+    {10, 3147}, {10, 3145}, {10, 3143}, {10, 3141}, {10, 3139}, {10, 3137},
+    {10, 3135}, {10, 3133}, {10, 3131}, {10, 3129}, {10, 3127}, {10, 3125},
+    {10, 3123}, {10, 3121}, {10, 3119}, {10, 3117}, {10, 3115}, {10, 3113},
+    {10, 3111}, {10, 3109}, {10, 3107}, {10, 3105}, {10, 3103}, {10, 3101},
+    {10, 3099}, {10, 3097}, {10, 3095}, {10, 3093}, {10, 3091}, {10, 3089},
+    {10, 3087}, {10, 3085}, {10, 3083}, {10, 3081}, {10, 3079}, {10, 3077},
+    {10, 3075}, {10, 3073}, {10, 3071}, {10, 3069}, {10, 3067}, {10, 3065},
+    {10, 3063}, {10, 3061}, {10, 3059}, {10, 3057}, {10, 3055}, {10, 3053},
+    {10, 3051}, {10, 3049}, {10, 3047}, {10, 3045}, {10, 3043}, {10, 3041},
+    {10, 3039}, {10, 3037}, {10, 3035}, {10, 3033}, {10, 3031}, {10, 3029},
+    {10, 3027}, {10, 3025}, {10, 3023}, {10, 3021}, {10, 3019}, {10, 3017},
+    {10, 3015}, {10, 3013}, {10, 3011}, {10, 3009}, {10, 3007}, {10, 3005},
+    {10, 3003}, {10, 3001}, {10, 2999}, {10, 2997}, {10, 2995}, {10, 2993},
+    {10, 2991}, {10, 2989}, {10, 2987}, {10, 2985}, {10, 2983}, {10, 2981},
+    {10, 2979}, {10, 2977}, {10, 2975}, {10, 2973}, {10, 2971}, {10, 2969},
+    {10, 2967}, {10, 2965}, {10, 2963}, {10, 2961}, {10, 2959}, {10, 2957},
+    {10, 2955}, {10, 2953}, {10, 2951}, {10, 2949}, {10, 2947}, {10, 2945},
+    {10, 2943}, {10, 2941}, {10, 2939}, {10, 2937}, {10, 2935}, {10, 2933},
+    {10, 2931}, {10, 2929}, {10, 2927}, {10, 2925}, {10, 2923}, {10, 2921},
+    {10, 2919}, {10, 2917}, {10, 2915}, {10, 2913}, {10, 2911}, {10, 2909},
+    {10, 2907}, {10, 2905}, {10, 2903}, {10, 2901}, {10, 2899}, {10, 2897},
+    {10, 2895}, {10, 2893}, {10, 2891}, {10, 2889}, {10, 2887}, {10, 2885},
+    {10, 2883}, {10, 2881}, {10, 2879}, {10, 2877}, {10, 2875}, {10, 2873},
+    {10, 2871}, {10, 2869}, {10, 2867}, {10, 2865}, {10, 2863}, {10, 2861},
+    {10, 2859}, {10, 2857}, {10, 2855}, {10, 2853}, {10, 2851}, {10, 2849},
+    {10, 2847}, {10, 2845}, {10, 2843}, {10, 2841}, {10, 2839}, {10, 2837},
+    {10, 2835}, {10, 2833}, {10, 2831}, {10, 2829}, {10, 2827}, {10, 2825},
+    {10, 2823}, {10, 2821}, {10, 2819}, {10, 2817}, {10, 2815}, {10, 2813},
+    {10, 2811}, {10, 2809}, {10, 2807}, {10, 2805}, {10, 2803}, {10, 2801},
+    {10, 2799}, {10, 2797}, {10, 2795}, {10, 2793}, {10, 2791}, {10, 2789},
+    {10, 2787}, {10, 2785}, {10, 2783}, {10, 2781}, {10, 2779}, {10, 2777},
+    {10, 2775}, {10, 2773}, {10, 2771}, {10, 2769}, {10, 2767}, {10, 2765},
+    {10, 2763}, {10, 2761}, {10, 2759}, {10, 2757}, {10, 2755}, {10, 2753},
+    {10, 2751}, {10, 2749}, {10, 2747}, {10, 2745}, {10, 2743}, {10, 2741},
+    {10, 2739}, {10, 2737}, {10, 2735}, {10, 2733}, {10, 2731}, {10, 2729},
+    {10, 2727}, {10, 2725}, {10, 2723}, {10, 2721}, {10, 2719}, {10, 2717},
+    {10, 2715}, {10, 2713}, {10, 2711}, {10, 2709}, {10, 2707}, {10, 2705},
+    {10, 2703}, {10, 2701}, {10, 2699}, {10, 2697}, {10, 2695}, {10, 2693},
+    {10, 2691}, {10, 2689}, {10, 2687}, {10, 2685}, {10, 2683}, {10, 2681},
+    {10, 2679}, {10, 2677}, {10, 2675}, {10, 2673}, {10, 2671}, {10, 2669},
+    {10, 2667}, {10, 2665}, {10, 2663}, {10, 2661}, {10, 2659}, {10, 2657},
+    {10, 2655}, {10, 2653}, {10, 2651}, {10, 2649}, {10, 2647}, {10, 2645},
+    {10, 2643}, {10, 2641}, {10, 2639}, {10, 2637}, {10, 2635}, {10, 2633},
+    {10, 2631}, {10, 2629}, {10, 2627}, {10, 2625}, {10, 2623}, {10, 2621},
+    {10, 2619}, {10, 2617}, {10, 2615}, {10, 2613}, {10, 2611}, {10, 2609},
+    {10, 2607}, {10, 2605}, {10, 2603}, {10, 2601}, {10, 2599}, {10, 2597},
+    {10, 2595}, {10, 2593}, {10, 2591}, {10, 2589}, {10, 2587}, {10, 2585},
+    {10, 2583}, {10, 2581}, {10, 2579}, {10, 2577}, {10, 2575}, {10, 2573},
+    {10, 2571}, {10, 2569}, {10, 2567}, {10, 2565}, {10, 2563}, {10, 2561},
+    {10, 2559}, {10, 2557}, {10, 2555}, {10, 2553}, {10, 2551}, {10, 2549},
+    {10, 2547}, {10, 2545}, {10, 2543}, {10, 2541}, {10, 2539}, {10, 2537},
+    {10, 2535}, {10, 2533}, {10, 2531}, {10, 2529}, {10, 2527}, {10, 2525},
+    {10, 2523}, {10, 2521}, {10, 2519}, {10, 2517}, {10, 2515}, {10, 2513},
+    {10, 2511}, {10, 2509}, {10, 2507}, {10, 2505}, {10, 2503}, {10, 2501},
+    {10, 2499}, {10, 2497}, {10, 2495}, {10, 2493}, {10, 2491}, {10, 2489},
+    {10, 2487}, {10, 2485}, {10, 2483}, {10, 2481}, {10, 2479}, {10, 2477},
+    {10, 2475}, {10, 2473}, {10, 2471}, {10, 2469}, {10, 2467}, {10, 2465},
+    {10, 2463}, {10, 2461}, {10, 2459}, {10, 2457}, {10, 2455}, {10, 2453},
+    {10, 2451}, {10, 2449}, {10, 2447}, {10, 2445}, {10, 2443}, {10, 2441},
+    {10, 2439}, {10, 2437}, {10, 2435}, {10, 2433}, {10, 2431}, {10, 2429},
+    {10, 2427}, {10, 2425}, {10, 2423}, {10, 2421}, {10, 2419}, {10, 2417},
+    {10, 2415}, {10, 2413}, {10, 2411}, {10, 2409}, {10, 2407}, {10, 2405},
+    {10, 2403}, {10, 2401}, {10, 2399}, {10, 2397}, {10, 2395}, {10, 2393},
+    {10, 2391}, {10, 2389}, {10, 2387}, {10, 2385}, {10, 2383}, {10, 2381},
+    {10, 2379}, {10, 2377}, {10, 2375}, {10, 2373}, {10, 2371}, {10, 2369},
+    {10, 2367}, {10, 2365}, {10, 2363}, {10, 2361}, {10, 2359}, {10, 2357},
+    {10, 2355}, {10, 2353}, {10, 2351}, {10, 2349}, {10, 2347}, {10, 2345},
+    {10, 2343}, {10, 2341}, {10, 2339}, {10, 2337}, {10, 2335}, {10, 2333},
+    {10, 2331}, {10, 2329}, {10, 2327}, {10, 2325}, {10, 2323}, {10, 2321},
+    {10, 2319}, {10, 2317}, {10, 2315}, {10, 2313}, {10, 2311}, {10, 2309},
+    {10, 2307}, {10, 2305}, {10, 2303}, {10, 2301}, {10, 2299}, {10, 2297},
+    {10, 2295}, {10, 2293}, {10, 2291}, {10, 2289}, {10, 2287}, {10, 2285},
+    {10, 2283}, {10, 2281}, {10, 2279}, {10, 2277}, {10, 2275}, {10, 2273},
+    {10, 2271}, {10, 2269}, {10, 2267}, {10, 2265}, {10, 2263}, {10, 2261},
+    {10, 2259}, {10, 2257}, {10, 2255}, {10, 2253}, {10, 2251}, {10, 2249},
+    {10, 2247}, {10, 2245}, {10, 2243}, {10, 2241}, {10, 2239}, {10, 2237},
+    {10, 2235}, {10, 2233}, {10, 2231}, {10, 2229}, {10, 2227}, {10, 2225},
+    {10, 2223}, {10, 2221}, {10, 2219}, {10, 2217}, {10, 2215}, {10, 2213},
+    {10, 2211}, {10, 2209}, {10, 2207}, {10, 2205}, {10, 2203}, {10, 2201},
+    {10, 2199}, {10, 2197}, {10, 2195}, {10, 2193}, {10, 2191}, {10, 2189},
+    {10, 2187}, {10, 2185}, {10, 2183}, {10, 2181}, {10, 2179}, {10, 2177},
+    {10, 2175}, {10, 2173}, {10, 2171}, {10, 2169}, {10, 2167}, {10, 2165},
+    {10, 2163}, {10, 2161}, {10, 2159}, {10, 2157}, {10, 2155}, {10, 2153},
+    {10, 2151}, {10, 2149}, {10, 2147}, {10, 2145}, {10, 2143}, {10, 2141},
+    {10, 2139}, {10, 2137}, {10, 2135}, {10, 2133}, {10, 2131}, {10, 2129},
+    {10, 2127}, {10, 2125}, {10, 2123}, {10, 2121}, {10, 2119}, {10, 2117},
+    {10, 2115}, {10, 2113}, {10, 2111}, {10, 2109}, {10, 2107}, {10, 2105},
+    {10, 2103}, {10, 2101}, {10, 2099}, {10, 2097}, {10, 2095}, {10, 2093},
+    {10, 2091}, {10, 2089}, {10, 2087}, {10, 2085}, {10, 2083}, {10, 2081},
+    {10, 2079}, {10, 2077}, {10, 2075}, {10, 2073}, {10, 2071}, {10, 2069},
+    {10, 2067}, {10, 2065}, {10, 2063}, {10, 2061}, {10, 2059}, {10, 2057},
+    {10, 2055}, {10, 2053}, {10, 2051}, {10, 2049}, {10, 2047}, {10, 2045},
+    {10, 2043}, {10, 2041}, {10, 2039}, {10, 2037}, {10, 2035}, {10, 2033},
+    {10, 2031}, {10, 2029}, {10, 2027}, {10, 2025}, {10, 2023}, {10, 2021},
+    {10, 2019}, {10, 2017}, {10, 2015}, {10, 2013}, {10, 2011}, {10, 2009},
+    {10, 2007}, {10, 2005}, {10, 2003}, {10, 2001}, {10, 1999}, {10, 1997},
+    {10, 1995}, {10, 1993}, {10, 1991}, {10, 1989}, {10, 1987}, {10, 1985},
+    {10, 1983}, {10, 1981}, {10, 1979}, {10, 1977}, {10, 1975}, {10, 1973},
+    {10, 1971}, {10, 1969}, {10, 1967}, {10, 1965}, {10, 1963}, {10, 1961},
+    {10, 1959}, {10, 1957}, {10, 1955}, {10, 1953}, {10, 1951}, {10, 1949},
+    {10, 1947}, {10, 1945}, {10, 1943}, {10, 1941}, {10, 1939}, {10, 1937},
+    {10, 1935}, {10, 1933}, {10, 1931}, {10, 1929}, {10, 1927}, {10, 1925},
+    {10, 1923}, {10, 1921}, {10, 1919}, {10, 1917}, {10, 1915}, {10, 1913},
+    {10, 1911}, {10, 1909}, {10, 1907}, {10, 1905}, {10, 1903}, {10, 1901},
+    {10, 1899}, {10, 1897}, {10, 1895}, {10, 1893}, {10, 1891}, {10, 1889},
+    {10, 1887}, {10, 1885}, {10, 1883}, {10, 1881}, {10, 1879}, {10, 1877},
+    {10, 1875}, {10, 1873}, {10, 1871}, {10, 1869}, {10, 1867}, {10, 1865},
+    {10, 1863}, {10, 1861}, {10, 1859}, {10, 1857}, {10, 1855}, {10, 1853},
+    {10, 1851}, {10, 1849}, {10, 1847}, {10, 1845}, {10, 1843}, {10, 1841},
+    {10, 1839}, {10, 1837}, {10, 1835}, {10, 1833}, {10, 1831}, {10, 1829},
+    {10, 1827}, {10, 1825}, {10, 1823}, {10, 1821}, {10, 1819}, {10, 1817},
+    {10, 1815}, {10, 1813}, {10, 1811}, {10, 1809}, {10, 1807}, {10, 1805},
+    {10, 1803}, {10, 1801}, {10, 1799}, {10, 1797}, {10, 1795}, {10, 1793},
+    {10, 1791}, {10, 1789}, {10, 1787}, {10, 1785}, {10, 1783}, {10, 1781},
+    {10, 1779}, {10, 1777}, {10, 1775}, {10, 1773}, {10, 1771}, {10, 1769},
+    {10, 1767}, {10, 1765}, {10, 1763}, {10, 1761}, {10, 1759}, {10, 1757},
+    {10, 1755}, {10, 1753}, {10, 1751}, {10, 1749}, {10, 1747}, {10, 1745},
+    {10, 1743}, {10, 1741}, {10, 1739}, {10, 1737}, {10, 1735}, {10, 1733},
+    {10, 1731}, {10, 1729}, {10, 1727}, {10, 1725}, {10, 1723}, {10, 1721},
+    {10, 1719}, {10, 1717}, {10, 1715}, {10, 1713}, {10, 1711}, {10, 1709},
+    {10, 1707}, {10, 1705}, {10, 1703}, {10, 1701}, {10, 1699}, {10, 1697},
+    {10, 1695}, {10, 1693}, {10, 1691}, {10, 1689}, {10, 1687}, {10, 1685},
+    {10, 1683}, {10, 1681}, {10, 1679}, {10, 1677}, {10, 1675}, {10, 1673},
+    {10, 1671}, {10, 1669}, {10, 1667}, {10, 1665}, {10, 1663}, {10, 1661},
+    {10, 1659}, {10, 1657}, {10, 1655}, {10, 1653}, {10, 1651}, {10, 1649},
+    {10, 1647}, {10, 1645}, {10, 1643}, {10, 1641}, {10, 1639}, {10, 1637},
+    {10, 1635}, {10, 1633}, {10, 1631}, {10, 1629}, {10, 1627}, {10, 1625},
+    {10, 1623}, {10, 1621}, {10, 1619}, {10, 1617}, {10, 1615}, {10, 1613},
+    {10, 1611}, {10, 1609}, {10, 1607}, {10, 1605}, {10, 1603}, {10, 1601},
+    {10, 1599}, {10, 1597}, {10, 1595}, {10, 1593}, {10, 1591}, {10, 1589},
+    {10, 1587}, {10, 1585}, {10, 1583}, {10, 1581}, {10, 1579}, {10, 1577},
+    {10, 1575}, {10, 1573}, {10, 1571}, {10, 1569}, {10, 1567}, {10, 1565},
+    {10, 1563}, {10, 1561}, {10, 1559}, {10, 1557}, {10, 1555}, {10, 1553},
+    {10, 1551}, {10, 1549}, {10, 1547}, {10, 1545}, {10, 1543}, {10, 1541},
+    {10, 1539}, {10, 1537}, {10, 1535}, {10, 1533}, {10, 1531}, {10, 1529},
+    {10, 1527}, {10, 1525}, {10, 1523}, {10, 1521}, {10, 1519}, {10, 1517},
+    {10, 1515}, {10, 1513}, {10, 1511}, {10, 1509}, {10, 1507}, {10, 1505},
+    {10, 1503}, {10, 1501}, {10, 1499}, {10, 1497}, {10, 1495}, {10, 1493},
+    {10, 1491}, {10, 1489}, {10, 1487}, {10, 1485}, {10, 1483}, {10, 1481},
+    {10, 1479}, {10, 1477}, {10, 1475}, {10, 1473}, {10, 1471}, {10, 1469},
+    {10, 1467}, {10, 1465}, {10, 1463}, {10, 1461}, {10, 1459}, {10, 1457},
+    {10, 1455}, {10, 1453}, {10, 1451}, {10, 1449}, {10, 1447}, {10, 1445},
+    {10, 1443}, {10, 1441}, {10, 1439}, {10, 1437}, {10, 1435}, {10, 1433},
+    {10, 1431}, {10, 1429}, {10, 1427}, {10, 1425}, {10, 1423}, {10, 1421},
+    {10, 1419}, {10, 1417}, {10, 1415}, {10, 1413}, {10, 1411}, {10, 1409},
+    {10, 1407}, {10, 1405}, {10, 1403}, {10, 1401}, {10, 1399}, {10, 1397},
+    {10, 1395}, {10, 1393}, {10, 1391}, {10, 1389}, {10, 1387}, {10, 1385},
+    {10, 1383}, {10, 1381}, {10, 1379}, {10, 1377}, {10, 1375}, {10, 1373},
+    {10, 1371}, {10, 1369}, {10, 1367}, {10, 1365}, {10, 1363}, {10, 1361},
+    {10, 1359}, {10, 1357}, {10, 1355}, {10, 1353}, {10, 1351}, {10, 1349},
+    {10, 1347}, {10, 1345}, {10, 1343}, {10, 1341}, {10, 1339}, {10, 1337},
+    {10, 1335}, {10, 1333}, {10, 1331}, {10, 1329}, {10, 1327}, {10, 1325},
+    {10, 1323}, {10, 1321}, {10, 1319}, {10, 1317}, {10, 1315}, {10, 1313},
+    {10, 1311}, {10, 1309}, {10, 1307}, {10, 1305}, {10, 1303}, {10, 1301},
+    {10, 1299}, {10, 1297}, {10, 1295}, {10, 1293}, {10, 1291}, {10, 1289},
+    {10, 1287}, {10, 1285}, {10, 1283}, {10, 1281}, {10, 1279}, {10, 1277},
+    {10, 1275}, {10, 1273}, {10, 1271}, {10, 1269}, {10, 1267}, {10, 1265},
+    {10, 1263}, {10, 1261}, {10, 1259}, {10, 1257}, {10, 1255}, {10, 1253},
+    {10, 1251}, {10, 1249}, {10, 1247}, {10, 1245}, {10, 1243}, {10, 1241},
+    {10, 1239}, {10, 1237}, {10, 1235}, {10, 1233}, {10, 1231}, {10, 1229},
+    {10, 1227}, {10, 1225}, {10, 1223}, {10, 1221}, {10, 1219}, {10, 1217},
+    {10, 1215}, {10, 1213}, {10, 1211}, {10, 1209}, {10, 1207}, {10, 1205},
+    {10, 1203}, {10, 1201}, {10, 1199}, {10, 1197}, {10, 1195}, {10, 1193},
+    {10, 1191}, {10, 1189}, {10, 1187}, {10, 1185}, {10, 1183}, {10, 1181},
+    {10, 1179}, {10, 1177}, {10, 1175}, {10, 1173}, {10, 1171}, {10, 1169},
+    {10, 1167}, {10, 1165}, {10, 1163}, {10, 1161}, {10, 1159}, {10, 1157},
+    {10, 1155}, {10, 1153}, {10, 1151}, {10, 1149}, {10, 1147}, {10, 1145},
+    {10, 1143}, {10, 1141}, {10, 1139}, {10, 1137}, {10, 1135}, {10, 1133},
+    {10, 1131}, {10, 1129}, {10, 1127}, {10, 1125}, {10, 1123}, {10, 1121},
+    {10, 1119}, {10, 1117}, {10, 1115}, {10, 1113}, {10, 1111}, {10, 1109},
+    {10, 1107}, {10, 1105}, {10, 1103}, {10, 1101}, {10, 1099}, {10, 1097},
+    {10, 1095}, {10, 1093}, {10, 1091}, {10, 1089}, {10, 1087}, {10, 1085},
+    {10, 1083}, {10, 1081}, {10, 1079}, {10, 1077}, {10, 1075}, {10, 1073},
+    {10, 1071}, {10, 1069}, {10, 1067}, {10, 1065}, {10, 1063}, {10, 1061},
+    {10, 1059}, {10, 1057}, {10, 1055}, {10, 1053}, {10, 1051}, {10, 1049},
+    {10, 1047}, {10, 1045}, {10, 1043}, {10, 1041}, {10, 1039}, {10, 1037},
+    {10, 1035}, {10, 1033}, {10, 1031}, {10, 1029}, {10, 1027}, {10, 1025},
+    {10, 1023}, {10, 1021}, {10, 1019}, {10, 1017}, {10, 1015}, {10, 1013},
+    {10, 1011}, {10, 1009}, {10, 1007}, {10, 1005}, {10, 1003}, {10, 1001},
+    {10, 999}, {10, 997}, {10, 995}, {10, 993}, {10, 991}, {10, 989},
+    {10, 987}, {10, 985}, {10, 983}, {10, 981}, {10, 979}, {10, 977},
+    {10, 975}, {10, 973}, {10, 971}, {10, 969}, {10, 967}, {10, 965},
+    {10, 963}, {10, 961}, {10, 959}, {10, 957}, {10, 955}, {10, 953},
+    {10, 951}, {10, 949}, {10, 947}, {10, 945}, {10, 943}, {10, 941},
+    {10, 939}, {10, 937}, {10, 935}, {10, 933}, {10, 931}, {10, 929},
+    {10, 927}, {10, 925}, {10, 923}, {10, 921}, {10, 919}, {10, 917},
+    {10, 915}, {10, 913}, {10, 911}, {10, 909}, {10, 907}, {10, 905},
+    {10, 903}, {10, 901}, {10, 899}, {10, 897}, {10, 895}, {10, 893},
+    {10, 891}, {10, 889}, {10, 887}, {10, 885}, {10, 883}, {10, 881},
+    {10, 879}, {10, 877}, {10, 875}, {10, 873}, {10, 871}, {10, 869},
+    {10, 867}, {10, 865}, {10, 863}, {10, 861}, {10, 859}, {10, 857},
+    {10, 855}, {10, 853}, {10, 851}, {10, 849}, {10, 847}, {10, 845},
+    {10, 843}, {10, 841}, {10, 839}, {10, 837}, {10, 835}, {10, 833},
+    {10, 831}, {10, 829}, {10, 827}, {10, 825}, {10, 823}, {10, 821},
+    {10, 819}, {10, 817}, {10, 815}, {10, 813}, {10, 811}, {10, 809},
+    {10, 807}, {10, 805}, {10, 803}, {10, 801}, {10, 799}, {10, 797},
+    {10, 795}, {10, 793}, {10, 791}, {10, 789}, {10, 787}, {10, 785},
+    {10, 783}, {10, 781}, {10, 779}, {10, 777}, {10, 775}, {10, 773},
+    {10, 771}, {10, 769}, {10, 767}, {10, 765}, {10, 763}, {10, 761},
+    {10, 759}, {10, 757}, {10, 755}, {10, 753}, {10, 751}, {10, 749},
+    {10, 747}, {10, 745}, {10, 743}, {10, 741}, {10, 739}, {10, 737},
+    {10, 735}, {10, 733}, {10, 731}, {10, 729}, {10, 727}, {10, 725},
+    {10, 723}, {10, 721}, {10, 719}, {10, 717}, {10, 715}, {10, 713},
+    {10, 711}, {10, 709}, {10, 707}, {10, 705}, {10, 703}, {10, 701},
+    {10, 699}, {10, 697}, {10, 695}, {10, 693}, {10, 691}, {10, 689},
+    {10, 687}, {10, 685}, {10, 683}, {10, 681}, {10, 679}, {10, 677},
+    {10, 675}, {10, 673}, {10, 671}, {10, 669}, {10, 667}, {10, 665},
+    {10, 663}, {10, 661}, {10, 659}, {10, 657}, {10, 655}, {10, 653},
+    {10, 651}, {10, 649}, {10, 647}, {10, 645}, {10, 643}, {10, 641},
+    {10, 639}, {10, 637}, {10, 635}, {10, 633}, {10, 631}, {10, 629},
+    {10, 627}, {10, 625}, {10, 623}, {10, 621}, {10, 619}, {10, 617},
+    {10, 615}, {10, 613}, {10, 611}, {10, 609}, {10, 607}, {10, 605},
+    {10, 603}, {10, 601}, {10, 599}, {10, 597}, {10, 595}, {10, 593},
+    {10, 591}, {10, 589}, {10, 587}, {10, 585}, {10, 583}, {10, 581},
+    {10, 579}, {10, 577}, {10, 575}, {10, 573}, {10, 571}, {10, 569},
+    {10, 567}, {10, 565}, {10, 563}, {10, 561}, {10, 559}, {10, 557},
+    {10, 555}, {10, 553}, {10, 551}, {10, 549}, {10, 547}, {10, 545},
+    {10, 543}, {10, 541}, {10, 539}, {10, 537}, {10, 535}, {10, 533},
+    {10, 531}, {10, 529}, {10, 527}, {10, 525}, {10, 523}, {10, 521},
+    {10, 519}, {10, 517}, {10, 515}, {10, 513}, {10, 511}, {10, 509},
+    {10, 507}, {10, 505}, {10, 503}, {10, 501}, {10, 499}, {10, 497},
+    {10, 495}, {10, 493}, {10, 491}, {10, 489}, {10, 487}, {10, 485},
+    {10, 483}, {10, 481}, {10, 479}, {10, 477}, {10, 475}, {10, 473},
+    {10, 471}, {10, 469}, {10, 467}, {10, 465}, {10, 463}, {10, 461},
+    {10, 459}, {10, 457}, {10, 455}, {10, 453}, {10, 451}, {10, 449},
+    {10, 447}, {10, 445}, {10, 443}, {10, 441}, {10, 439}, {10, 437},
+    {10, 435}, {10, 433}, {10, 431}, {10, 429}, {10, 427}, {10, 425},
+    {10, 423}, {10, 421}, {10, 419}, {10, 417}, {10, 415}, {10, 413},
+    {10, 411}, {10, 409}, {10, 407}, {10, 405}, {10, 403}, {10, 401},
+    {10, 399}, {10, 397}, {10, 395}, {10, 393}, {10, 391}, {10, 389},
+    {10, 387}, {10, 385}, {10, 383}, {10, 381}, {10, 379}, {10, 377},
+    {10, 375}, {10, 373}, {10, 371}, {10, 369}, {10, 367}, {10, 365},
+    {10, 363}, {10, 361}, {10, 359}, {10, 357}, {10, 355}, {10, 353},
+    {10, 351}, {10, 349}, {10, 347}, {10, 345}, {10, 343}, {10, 341},
+    {10, 339}, {10, 337}, {10, 335}, {10, 333}, {10, 331}, {10, 329},
+    {10, 327}, {10, 325}, {10, 323}, {10, 321}, {10, 319}, {10, 317},
+    {10, 315}, {10, 313}, {10, 311}, {10, 309}, {10, 307}, {10, 305},
+    {10, 303}, {10, 301}, {10, 299}, {10, 297}, {10, 295}, {10, 293},
+    {10, 291}, {10, 289}, {10, 287}, {10, 285}, {10, 283}, {10, 281},
+    {10, 279}, {10, 277}, {10, 275}, {10, 273}, {10, 271}, {10, 269},
+    {10, 267}, {10, 265}, {10, 263}, {10, 261}, {10, 259}, {10, 257},
+    {10, 255}, {10, 253}, {10, 251}, {10, 249}, {10, 247}, {10, 245},
+    {10, 243}, {10, 241}, {10, 239}, {10, 237}, {10, 235}, {10, 233},
+    {10, 231}, {10, 229}, {10, 227}, {10, 225}, {10, 223}, {10, 221},
+    {10, 219}, {10, 217}, {10, 215}, {10, 213}, {10, 211}, {10, 209},
+    {10, 207}, {10, 205}, {10, 203}, {10, 201}, {10, 199}, {10, 197},
+    {10, 195}, {10, 193}, {10, 191}, {10, 189}, {10, 187}, {10, 185},
+    {10, 183}, {10, 181}, {10, 179}, {10, 177}, {10, 175}, {10, 173},
+    {10, 171}, {10, 169}, {10, 167}, {10, 165}, {10, 163}, {10, 161},
+    {10, 159}, {10, 157}, {10, 155}, {10, 153}, {10, 151}, {10, 149},
+    {10, 147}, {10, 145}, {10, 143}, {10, 141}, {10, 139}, {10, 137},
+    {10, 135}, {10, 133}, {10, 131}, {10, 129}, {10, 127}, {10, 125},
+    {10, 123}, {10, 121}, {10, 119}, {10, 117}, {10, 115}, {10, 113},
+    {10, 111}, {10, 109}, {10, 107}, {10, 105}, {10, 103}, {10, 101},
+    {10, 99}, {10, 97}, {10, 95}, {10, 93}, {10, 91}, {10, 89},
+    {10, 87}, {10, 85}, {10, 83}, {10, 81}, {10, 79}, {10, 77},
+    {10, 75}, {10, 73}, {10, 71}, {10, 69}, {10, 67}, {10, 65},
+    {10, 63}, {10, 61}, {10, 59}, {10, 57}, {10, 55}, {10, 53},
+    {10, 51}, {10, 49}, {10, 47}, {10, 45}, {10, 43}, {10, 41},
+    {10, 39}, {10, 37}, {10, 35}, {10, 33}, {10, 31}, {10, 29},
+    {10, 27}, {10, 25}, {10, 23}, {10, 21}, {10, 19}, {10, 17},
+    {10, 15}, {10, 13}, {10, 11}, {10, 9}, {10, 7}, {10, 5},
+    {10, 3}, {10, 1}, {9, 63}, {9, 61}, {9, 59}, {9, 57},
+    {9, 55}, {9, 53}, {9, 51}, {9, 49}, {9, 47}, {9, 45},
+    {9, 43}, {9, 41}, {9, 39}, {9, 37}, {9, 35}, {9, 33},
+    {9, 31}, {9, 29}, {9, 27}, {9, 25}, {9, 23}, {9, 21},
+    {9, 19}, {9, 17}, {9, 15}, {9, 13}, {9, 11}, {9, 9},
+    {9, 7}, {9, 5}, {9, 3}, {9, 1}, {8, 31}, {8, 29},
+    {8, 27}, {8, 25}, {8, 23}, {8, 21}, {8, 19}, {8, 17},
+    {8, 15}, {8, 13}, {8, 11}, {8, 9}, {8, 7}, {8, 5},
+    {8, 3}, {8, 1}, {7, 15}, {7, 13}, {7, 11}, {7, 9},
+    {7, 7}, {7, 5}, {7, 3}, {7, 1}, {6, 7}, {6, 5},
+    {6, 3}, {6, 1}, {5, 3}, {5, 1}, {4, 1}, {3, 1},
+    {2, 1}, {1, 1}, {0, 0}, {1, 0}, {2, 0}, {3, 0},
+    {4, 0}, {5, 0}, {5, 2}, {6, 0}, {6, 2}, {6, 4},
+    {6, 6}, {7, 0}, {7, 2}, {7, 4}, {7, 6}, {7, 8},
+    {7, 10}, {7, 12}, {7, 14}, {8, 0}, {8, 2}, {8, 4},
+    {8, 6}, {8, 8}, {8, 10}, {8, 12}, {8, 14}, {8, 16},
+    {8, 18}, {8, 20}, {8, 22}, {8, 24}, {8, 26}, {8, 28},
+    {8, 30}, {9, 0}, {9, 2}, {9, 4}, {9, 6}, {9, 8},
+    {9, 10}, {9, 12}, {9, 14}, {9, 16}, {9, 18}, {9, 20},
+    {9, 22}, {9, 24}, {9, 26}, {9, 28}, {9, 30}, {9, 32},
+    {9, 34}, {9, 36}, {9, 38}, {9, 40}, {9, 42}, {9, 44},
+    {9, 46}, {9, 48}, {9, 50}, {9, 52}, {9, 54}, {9, 56},
+    {9, 58}, {9, 60}, {9, 62}, {10, 0}, {10, 2}, {10, 4},
+    {10, 6}, {10, 8}, {10, 10}, {10, 12}, {10, 14}, {10, 16},
+    {10, 18}, {10, 20}, {10, 22}, {10, 24}, {10, 26}, {10, 28},
+    {10, 30}, {10, 32}, {10, 34}, {10, 36}, {10, 38}, {10, 40},
+    {10, 42}, {10, 44}, {10, 46}, {10, 48}, {10, 50}, {10, 52},
+    {10, 54}, {10, 56}, {10, 58}, {10, 60}, {10, 62}, {10, 64},
+    {10, 66}, {10, 68}, {10, 70}, {10, 72}, {10, 74}, {10, 76},
+    {10, 78}, {10, 80}, {10, 82}, {10, 84}, {10, 86}, {10, 88},
+    {10, 90}, {10, 92}, {10, 94}, {10, 96}, {10, 98}, {10, 100},
+    {10, 102}, {10, 104}, {10, 106}, {10, 108}, {10, 110}, {10, 112},
+    {10, 114}, {10, 116}, {10, 118}, {10, 120}, {10, 122}, {10, 124},
+    {10, 126}, {10, 128}, {10, 130}, {10, 132}, {10, 134}, {10, 136},
+    {10, 138}, {10, 140}, {10, 142}, {10, 144}, {10, 146}, {10, 148},
+    {10, 150}, {10, 152}, {10, 154}, {10, 156}, {10, 158}, {10, 160},
+    {10, 162}, {10, 164}, {10, 166}, {10, 168}, {10, 170}, {10, 172},
+    {10, 174}, {10, 176}, {10, 178}, {10, 180}, {10, 182}, {10, 184},
+    {10, 186}, {10, 188}, {10, 190}, {10, 192}, {10, 194}, {10, 196},
+    {10, 198}, {10, 200}, {10, 202}, {10, 204}, {10, 206}, {10, 208},
+    {10, 210}, {10, 212}, {10, 214}, {10, 216}, {10, 218}, {10, 220},
+    {10, 222}, {10, 224}, {10, 226}, {10, 228}, {10, 230}, {10, 232},
+    {10, 234}, {10, 236}, {10, 238}, {10, 240}, {10, 242}, {10, 244},
+    {10, 246}, {10, 248}, {10, 250}, {10, 252}, {10, 254}, {10, 256},
+    {10, 258}, {10, 260}, {10, 262}, {10, 264}, {10, 266}, {10, 268},
+    {10, 270}, {10, 272}, {10, 274}, {10, 276}, {10, 278}, {10, 280},
+    {10, 282}, {10, 284}, {10, 286}, {10, 288}, {10, 290}, {10, 292},
+    {10, 294}, {10, 296}, {10, 298}, {10, 300}, {10, 302}, {10, 304},
+    {10, 306}, {10, 308}, {10, 310}, {10, 312}, {10, 314}, {10, 316},
+    {10, 318}, {10, 320}, {10, 322}, {10, 324}, {10, 326}, {10, 328},
+    {10, 330}, {10, 332}, {10, 334}, {10, 336}, {10, 338}, {10, 340},
+    {10, 342}, {10, 344}, {10, 346}, {10, 348}, {10, 350}, {10, 352},
+    {10, 354}, {10, 356}, {10, 358}, {10, 360}, {10, 362}, {10, 364},
+    {10, 366}, {10, 368}, {10, 370}, {10, 372}, {10, 374}, {10, 376},
+    {10, 378}, {10, 380}, {10, 382}, {10, 384}, {10, 386}, {10, 388},
+    {10, 390}, {10, 392}, {10, 394}, {10, 396}, {10, 398}, {10, 400},
+    {10, 402}, {10, 404}, {10, 406}, {10, 408}, {10, 410}, {10, 412},
+    {10, 414}, {10, 416}, {10, 418}, {10, 420}, {10, 422}, {10, 424},
+    {10, 426}, {10, 428}, {10, 430}, {10, 432}, {10, 434}, {10, 436},
+    {10, 438}, {10, 440}, {10, 442}, {10, 444}, {10, 446}, {10, 448},
+    {10, 450}, {10, 452}, {10, 454}, {10, 456}, {10, 458}, {10, 460},
+    {10, 462}, {10, 464}, {10, 466}, {10, 468}, {10, 470}, {10, 472},
+    {10, 474}, {10, 476}, {10, 478}, {10, 480}, {10, 482}, {10, 484},
+    {10, 486}, {10, 488}, {10, 490}, {10, 492}, {10, 494}, {10, 496},
+    {10, 498}, {10, 500}, {10, 502}, {10, 504}, {10, 506}, {10, 508},
+    {10, 510}, {10, 512}, {10, 514}, {10, 516}, {10, 518}, {10, 520},
+    {10, 522}, {10, 524}, {10, 526}, {10, 528}, {10, 530}, {10, 532},
+    {10, 534}, {10, 536}, {10, 538}, {10, 540}, {10, 542}, {10, 544},
+    {10, 546}, {10, 548}, {10, 550}, {10, 552}, {10, 554}, {10, 556},
+    {10, 558}, {10, 560}, {10, 562}, {10, 564}, {10, 566}, {10, 568},
+    {10, 570}, {10, 572}, {10, 574}, {10, 576}, {10, 578}, {10, 580},
+    {10, 582}, {10, 584}, {10, 586}, {10, 588}, {10, 590}, {10, 592},
+    {10, 594}, {10, 596}, {10, 598}, {10, 600}, {10, 602}, {10, 604},
+    {10, 606}, {10, 608}, {10, 610}, {10, 612}, {10, 614}, {10, 616},
+    {10, 618}, {10, 620}, {10, 622}, {10, 624}, {10, 626}, {10, 628},
+    {10, 630}, {10, 632}, {10, 634}, {10, 636}, {10, 638}, {10, 640},
+    {10, 642}, {10, 644}, {10, 646}, {10, 648}, {10, 650}, {10, 652},
+    {10, 654}, {10, 656}, {10, 658}, {10, 660}, {10, 662}, {10, 664},
+    {10, 666}, {10, 668}, {10, 670}, {10, 672}, {10, 674}, {10, 676},
+    {10, 678}, {10, 680}, {10, 682}, {10, 684}, {10, 686}, {10, 688},
+    {10, 690}, {10, 692}, {10, 694}, {10, 696}, {10, 698}, {10, 700},
+    {10, 702}, {10, 704}, {10, 706}, {10, 708}, {10, 710}, {10, 712},
+    {10, 714}, {10, 716}, {10, 718}, {10, 720}, {10, 722}, {10, 724},
+    {10, 726}, {10, 728}, {10, 730}, {10, 732}, {10, 734}, {10, 736},
+    {10, 738}, {10, 740}, {10, 742}, {10, 744}, {10, 746}, {10, 748},
+    {10, 750}, {10, 752}, {10, 754}, {10, 756}, {10, 758}, {10, 760},
+    {10, 762}, {10, 764}, {10, 766}, {10, 768}, {10, 770}, {10, 772},
+    {10, 774}, {10, 776}, {10, 778}, {10, 780}, {10, 782}, {10, 784},
+    {10, 786}, {10, 788}, {10, 790}, {10, 792}, {10, 794}, {10, 796},
+    {10, 798}, {10, 800}, {10, 802}, {10, 804}, {10, 806}, {10, 808},
+    {10, 810}, {10, 812}, {10, 814}, {10, 816}, {10, 818}, {10, 820},
+    {10, 822}, {10, 824}, {10, 826}, {10, 828}, {10, 830}, {10, 832},
+    {10, 834}, {10, 836}, {10, 838}, {10, 840}, {10, 842}, {10, 844},
+    {10, 846}, {10, 848}, {10, 850}, {10, 852}, {10, 854}, {10, 856},
+    {10, 858}, {10, 860}, {10, 862}, {10, 864}, {10, 866}, {10, 868},
+    {10, 870}, {10, 872}, {10, 874}, {10, 876}, {10, 878}, {10, 880},
+    {10, 882}, {10, 884}, {10, 886}, {10, 888}, {10, 890}, {10, 892},
+    {10, 894}, {10, 896}, {10, 898}, {10, 900}, {10, 902}, {10, 904},
+    {10, 906}, {10, 908}, {10, 910}, {10, 912}, {10, 914}, {10, 916},
+    {10, 918}, {10, 920}, {10, 922}, {10, 924}, {10, 926}, {10, 928},
+    {10, 930}, {10, 932}, {10, 934}, {10, 936}, {10, 938}, {10, 940},
+    {10, 942}, {10, 944}, {10, 946}, {10, 948}, {10, 950}, {10, 952},
+    {10, 954}, {10, 956}, {10, 958}, {10, 960}, {10, 962}, {10, 964},
+    {10, 966}, {10, 968}, {10, 970}, {10, 972}, {10, 974}, {10, 976},
+    {10, 978}, {10, 980}, {10, 982}, {10, 984}, {10, 986}, {10, 988},
+    {10, 990}, {10, 992}, {10, 994}, {10, 996}, {10, 998}, {10, 1000},
+    {10, 1002}, {10, 1004}, {10, 1006}, {10, 1008}, {10, 1010}, {10, 1012},
+    {10, 1014}, {10, 1016}, {10, 1018}, {10, 1020}, {10, 1022}, {10, 1024},
+    {10, 1026}, {10, 1028}, {10, 1030}, {10, 1032}, {10, 1034}, {10, 1036},
+    {10, 1038}, {10, 1040}, {10, 1042}, {10, 1044}, {10, 1046}, {10, 1048},
+    {10, 1050}, {10, 1052}, {10, 1054}, {10, 1056}, {10, 1058}, {10, 1060},
+    {10, 1062}, {10, 1064}, {10, 1066}, {10, 1068}, {10, 1070}, {10, 1072},
+    {10, 1074}, {10, 1076}, {10, 1078}, {10, 1080}, {10, 1082}, {10, 1084},
+    {10, 1086}, {10, 1088}, {10, 1090}, {10, 1092}, {10, 1094}, {10, 1096},
+    {10, 1098}, {10, 1100}, {10, 1102}, {10, 1104}, {10, 1106}, {10, 1108},
+    {10, 1110}, {10, 1112}, {10, 1114}, {10, 1116}, {10, 1118}, {10, 1120},
+    {10, 1122}, {10, 1124}, {10, 1126}, {10, 1128}, {10, 1130}, {10, 1132},
+    {10, 1134}, {10, 1136}, {10, 1138}, {10, 1140}, {10, 1142}, {10, 1144},
+    {10, 1146}, {10, 1148}, {10, 1150}, {10, 1152}, {10, 1154}, {10, 1156},
+    {10, 1158}, {10, 1160}, {10, 1162}, {10, 1164}, {10, 1166}, {10, 1168},
+    {10, 1170}, {10, 1172}, {10, 1174}, {10, 1176}, {10, 1178}, {10, 1180},
+    {10, 1182}, {10, 1184}, {10, 1186}, {10, 1188}, {10, 1190}, {10, 1192},
+    {10, 1194}, {10, 1196}, {10, 1198}, {10, 1200}, {10, 1202}, {10, 1204},
+    {10, 1206}, {10, 1208}, {10, 1210}, {10, 1212}, {10, 1214}, {10, 1216},
+    {10, 1218}, {10, 1220}, {10, 1222}, {10, 1224}, {10, 1226}, {10, 1228},
+    {10, 1230}, {10, 1232}, {10, 1234}, {10, 1236}, {10, 1238}, {10, 1240},
+    {10, 1242}, {10, 1244}, {10, 1246}, {10, 1248}, {10, 1250}, {10, 1252},
+    {10, 1254}, {10, 1256}, {10, 1258}, {10, 1260}, {10, 1262}, {10, 1264},
+    {10, 1266}, {10, 1268}, {10, 1270}, {10, 1272}, {10, 1274}, {10, 1276},
+    {10, 1278}, {10, 1280}, {10, 1282}, {10, 1284}, {10, 1286}, {10, 1288},
+    {10, 1290}, {10, 1292}, {10, 1294}, {10, 1296}, {10, 1298}, {10, 1300},
+    {10, 1302}, {10, 1304}, {10, 1306}, {10, 1308}, {10, 1310}, {10, 1312},
+    {10, 1314}, {10, 1316}, {10, 1318}, {10, 1320}, {10, 1322}, {10, 1324},
+    {10, 1326}, {10, 1328}, {10, 1330}, {10, 1332}, {10, 1334}, {10, 1336},
+    {10, 1338}, {10, 1340}, {10, 1342}, {10, 1344}, {10, 1346}, {10, 1348},
+    {10, 1350}, {10, 1352}, {10, 1354}, {10, 1356}, {10, 1358}, {10, 1360},
+    {10, 1362}, {10, 1364}, {10, 1366}, {10, 1368}, {10, 1370}, {10, 1372},
+    {10, 1374}, {10, 1376}, {10, 1378}, {10, 1380}, {10, 1382}, {10, 1384},
+    {10, 1386}, {10, 1388}, {10, 1390}, {10, 1392}, {10, 1394}, {10, 1396},
+    {10, 1398}, {10, 1400}, {10, 1402}, {10, 1404}, {10, 1406}, {10, 1408},
+    {10, 1410}, {10, 1412}, {10, 1414}, {10, 1416}, {10, 1418}, {10, 1420},
+    {10, 1422}, {10, 1424}, {10, 1426}, {10, 1428}, {10, 1430}, {10, 1432},
+    {10, 1434}, {10, 1436}, {10, 1438}, {10, 1440}, {10, 1442}, {10, 1444},
+    {10, 1446}, {10, 1448}, {10, 1450}, {10, 1452}, {10, 1454}, {10, 1456},
+    {10, 1458}, {10, 1460}, {10, 1462}, {10, 1464}, {10, 1466}, {10, 1468},
+    {10, 1470}, {10, 1472}, {10, 1474}, {10, 1476}, {10, 1478}, {10, 1480},
+    {10, 1482}, {10, 1484}, {10, 1486}, {10, 1488}, {10, 1490}, {10, 1492},
+    {10, 1494}, {10, 1496}, {10, 1498}, {10, 1500}, {10, 1502}, {10, 1504},
+    {10, 1506}, {10, 1508}, {10, 1510}, {10, 1512}, {10, 1514}, {10, 1516},
+    {10, 1518}, {10, 1520}, {10, 1522}, {10, 1524}, {10, 1526}, {10, 1528},
+    {10, 1530}, {10, 1532}, {10, 1534}, {10, 1536}, {10, 1538}, {10, 1540},
+    {10, 1542}, {10, 1544}, {10, 1546}, {10, 1548}, {10, 1550}, {10, 1552},
+    {10, 1554}, {10, 1556}, {10, 1558}, {10, 1560}, {10, 1562}, {10, 1564},
+    {10, 1566}, {10, 1568}, {10, 1570}, {10, 1572}, {10, 1574}, {10, 1576},
+    {10, 1578}, {10, 1580}, {10, 1582}, {10, 1584}, {10, 1586}, {10, 1588},
+    {10, 1590}, {10, 1592}, {10, 1594}, {10, 1596}, {10, 1598}, {10, 1600},
+    {10, 1602}, {10, 1604}, {10, 1606}, {10, 1608}, {10, 1610}, {10, 1612},
+    {10, 1614}, {10, 1616}, {10, 1618}, {10, 1620}, {10, 1622}, {10, 1624},
+    {10, 1626}, {10, 1628}, {10, 1630}, {10, 1632}, {10, 1634}, {10, 1636},
+    {10, 1638}, {10, 1640}, {10, 1642}, {10, 1644}, {10, 1646}, {10, 1648},
+    {10, 1650}, {10, 1652}, {10, 1654}, {10, 1656}, {10, 1658}, {10, 1660},
+    {10, 1662}, {10, 1664}, {10, 1666}, {10, 1668}, {10, 1670}, {10, 1672},
+    {10, 1674}, {10, 1676}, {10, 1678}, {10, 1680}, {10, 1682}, {10, 1684},
+    {10, 1686}, {10, 1688}, {10, 1690}, {10, 1692}, {10, 1694}, {10, 1696},
+    {10, 1698}, {10, 1700}, {10, 1702}, {10, 1704}, {10, 1706}, {10, 1708},
+    {10, 1710}, {10, 1712}, {10, 1714}, {10, 1716}, {10, 1718}, {10, 1720},
+    {10, 1722}, {10, 1724}, {10, 1726}, {10, 1728}, {10, 1730}, {10, 1732},
+    {10, 1734}, {10, 1736}, {10, 1738}, {10, 1740}, {10, 1742}, {10, 1744},
+    {10, 1746}, {10, 1748}, {10, 1750}, {10, 1752}, {10, 1754}, {10, 1756},
+    {10, 1758}, {10, 1760}, {10, 1762}, {10, 1764}, {10, 1766}, {10, 1768},
+    {10, 1770}, {10, 1772}, {10, 1774}, {10, 1776}, {10, 1778}, {10, 1780},
+    {10, 1782}, {10, 1784}, {10, 1786}, {10, 1788}, {10, 1790}, {10, 1792},
+    {10, 1794}, {10, 1796}, {10, 1798}, {10, 1800}, {10, 1802}, {10, 1804},
+    {10, 1806}, {10, 1808}, {10, 1810}, {10, 1812}, {10, 1814}, {10, 1816},
+    {10, 1818}, {10, 1820}, {10, 1822}, {10, 1824}, {10, 1826}, {10, 1828},
+    {10, 1830}, {10, 1832}, {10, 1834}, {10, 1836}, {10, 1838}, {10, 1840},
+    {10, 1842}, {10, 1844}, {10, 1846}, {10, 1848}, {10, 1850}, {10, 1852},
+    {10, 1854}, {10, 1856}, {10, 1858}, {10, 1860}, {10, 1862}, {10, 1864},
+    {10, 1866}, {10, 1868}, {10, 1870}, {10, 1872}, {10, 1874}, {10, 1876},
+    {10, 1878}, {10, 1880}, {10, 1882}, {10, 1884}, {10, 1886}, {10, 1888},
+    {10, 1890}, {10, 1892}, {10, 1894}, {10, 1896}, {10, 1898}, {10, 1900},
+    {10, 1902}, {10, 1904}, {10, 1906}, {10, 1908}, {10, 1910}, {10, 1912},
+    {10, 1914}, {10, 1916}, {10, 1918}, {10, 1920}, {10, 1922}, {10, 1924},
+    {10, 1926}, {10, 1928}, {10, 1930}, {10, 1932}, {10, 1934}, {10, 1936},
+    {10, 1938}, {10, 1940}, {10, 1942}, {10, 1944}, {10, 1946}, {10, 1948},
+    {10, 1950}, {10, 1952}, {10, 1954}, {10, 1956}, {10, 1958}, {10, 1960},
+    {10, 1962}, {10, 1964}, {10, 1966}, {10, 1968}, {10, 1970}, {10, 1972},
+    {10, 1974}, {10, 1976}, {10, 1978}, {10, 1980}, {10, 1982}, {10, 1984},
+    {10, 1986}, {10, 1988}, {10, 1990}, {10, 1992}, {10, 1994}, {10, 1996},
+    {10, 1998}, {10, 2000}, {10, 2002}, {10, 2004}, {10, 2006}, {10, 2008},
+    {10, 2010}, {10, 2012}, {10, 2014}, {10, 2016}, {10, 2018}, {10, 2020},
+    {10, 2022}, {10, 2024}, {10, 2026}, {10, 2028}, {10, 2030}, {10, 2032},
+    {10, 2034}, {10, 2036}, {10, 2038}, {10, 2040}, {10, 2042}, {10, 2044},
+    {10, 2046}, {10, 2048}, {10, 2050}, {10, 2052}, {10, 2054}, {10, 2056},
+    {10, 2058}, {10, 2060}, {10, 2062}, {10, 2064}, {10, 2066}, {10, 2068},
+    {10, 2070}, {10, 2072}, {10, 2074}, {10, 2076}, {10, 2078}, {10, 2080},
+    {10, 2082}, {10, 2084}, {10, 2086}, {10, 2088}, {10, 2090}, {10, 2092},
+    {10, 2094}, {10, 2096}, {10, 2098}, {10, 2100}, {10, 2102}, {10, 2104},
+    {10, 2106}, {10, 2108}, {10, 2110}, {10, 2112}, {10, 2114}, {10, 2116},
+    {10, 2118}, {10, 2120}, {10, 2122}, {10, 2124}, {10, 2126}, {10, 2128},
+    {10, 2130}, {10, 2132}, {10, 2134}, {10, 2136}, {10, 2138}, {10, 2140},
+    {10, 2142}, {10, 2144}, {10, 2146}, {10, 2148}, {10, 2150}, {10, 2152},
+    {10, 2154}, {10, 2156}, {10, 2158}, {10, 2160}, {10, 2162}, {10, 2164},
+    {10, 2166}, {10, 2168}, {10, 2170}, {10, 2172}, {10, 2174}, {10, 2176},
+    {10, 2178}, {10, 2180}, {10, 2182}, {10, 2184}, {10, 2186}, {10, 2188},
+    {10, 2190}, {10, 2192}, {10, 2194}, {10, 2196}, {10, 2198}, {10, 2200},
+    {10, 2202}, {10, 2204}, {10, 2206}, {10, 2208}, {10, 2210}, {10, 2212},
+    {10, 2214}, {10, 2216}, {10, 2218}, {10, 2220}, {10, 2222}, {10, 2224},
+    {10, 2226}, {10, 2228}, {10, 2230}, {10, 2232}, {10, 2234}, {10, 2236},
+    {10, 2238}, {10, 2240}, {10, 2242}, {10, 2244}, {10, 2246}, {10, 2248},
+    {10, 2250}, {10, 2252}, {10, 2254}, {10, 2256}, {10, 2258}, {10, 2260},
+    {10, 2262}, {10, 2264}, {10, 2266}, {10, 2268}, {10, 2270}, {10, 2272},
+    {10, 2274}, {10, 2276}, {10, 2278}, {10, 2280}, {10, 2282}, {10, 2284},
+    {10, 2286}, {10, 2288}, {10, 2290}, {10, 2292}, {10, 2294}, {10, 2296},
+    {10, 2298}, {10, 2300}, {10, 2302}, {10, 2304}, {10, 2306}, {10, 2308},
+    {10, 2310}, {10, 2312}, {10, 2314}, {10, 2316}, {10, 2318}, {10, 2320},
+    {10, 2322}, {10, 2324}, {10, 2326}, {10, 2328}, {10, 2330}, {10, 2332},
+    {10, 2334}, {10, 2336}, {10, 2338}, {10, 2340}, {10, 2342}, {10, 2344},
+    {10, 2346}, {10, 2348}, {10, 2350}, {10, 2352}, {10, 2354}, {10, 2356},
+    {10, 2358}, {10, 2360}, {10, 2362}, {10, 2364}, {10, 2366}, {10, 2368},
+    {10, 2370}, {10, 2372}, {10, 2374}, {10, 2376}, {10, 2378}, {10, 2380},
+    {10, 2382}, {10, 2384}, {10, 2386}, {10, 2388}, {10, 2390}, {10, 2392},
+    {10, 2394}, {10, 2396}, {10, 2398}, {10, 2400}, {10, 2402}, {10, 2404},
+    {10, 2406}, {10, 2408}, {10, 2410}, {10, 2412}, {10, 2414}, {10, 2416},
+    {10, 2418}, {10, 2420}, {10, 2422}, {10, 2424}, {10, 2426}, {10, 2428},
+    {10, 2430}, {10, 2432}, {10, 2434}, {10, 2436}, {10, 2438}, {10, 2440},
+    {10, 2442}, {10, 2444}, {10, 2446}, {10, 2448}, {10, 2450}, {10, 2452},
+    {10, 2454}, {10, 2456}, {10, 2458}, {10, 2460}, {10, 2462}, {10, 2464},
+    {10, 2466}, {10, 2468}, {10, 2470}, {10, 2472}, {10, 2474}, {10, 2476},
+    {10, 2478}, {10, 2480}, {10, 2482}, {10, 2484}, {10, 2486}, {10, 2488},
+    {10, 2490}, {10, 2492}, {10, 2494}, {10, 2496}, {10, 2498}, {10, 2500},
+    {10, 2502}, {10, 2504}, {10, 2506}, {10, 2508}, {10, 2510}, {10, 2512},
+    {10, 2514}, {10, 2516}, {10, 2518}, {10, 2520}, {10, 2522}, {10, 2524},
+    {10, 2526}, {10, 2528}, {10, 2530}, {10, 2532}, {10, 2534}, {10, 2536},
+    {10, 2538}, {10, 2540}, {10, 2542}, {10, 2544}, {10, 2546}, {10, 2548},
+    {10, 2550}, {10, 2552}, {10, 2554}, {10, 2556}, {10, 2558}, {10, 2560},
+    {10, 2562}, {10, 2564}, {10, 2566}, {10, 2568}, {10, 2570}, {10, 2572},
+    {10, 2574}, {10, 2576}, {10, 2578}, {10, 2580}, {10, 2582}, {10, 2584},
+    {10, 2586}, {10, 2588}, {10, 2590}, {10, 2592}, {10, 2594}, {10, 2596},
+    {10, 2598}, {10, 2600}, {10, 2602}, {10, 2604}, {10, 2606}, {10, 2608},
+    {10, 2610}, {10, 2612}, {10, 2614}, {10, 2616}, {10, 2618}, {10, 2620},
+    {10, 2622}, {10, 2624}, {10, 2626}, {10, 2628}, {10, 2630}, {10, 2632},
+    {10, 2634}, {10, 2636}, {10, 2638}, {10, 2640}, {10, 2642}, {10, 2644},
+    {10, 2646}, {10, 2648}, {10, 2650}, {10, 2652}, {10, 2654}, {10, 2656},
+    {10, 2658}, {10, 2660}, {10, 2662}, {10, 2664}, {10, 2666}, {10, 2668},
+    {10, 2670}, {10, 2672}, {10, 2674}, {10, 2676}, {10, 2678}, {10, 2680},
+    {10, 2682}, {10, 2684}, {10, 2686}, {10, 2688}, {10, 2690}, {10, 2692},
+    {10, 2694}, {10, 2696}, {10, 2698}, {10, 2700}, {10, 2702}, {10, 2704},
+    {10, 2706}, {10, 2708}, {10, 2710}, {10, 2712}, {10, 2714}, {10, 2716},
+    {10, 2718}, {10, 2720}, {10, 2722}, {10, 2724}, {10, 2726}, {10, 2728},
+    {10, 2730}, {10, 2732}, {10, 2734}, {10, 2736}, {10, 2738}, {10, 2740},
+    {10, 2742}, {10, 2744}, {10, 2746}, {10, 2748}, {10, 2750}, {10, 2752},
+    {10, 2754}, {10, 2756}, {10, 2758}, {10, 2760}, {10, 2762}, {10, 2764},
+    {10, 2766}, {10, 2768}, {10, 2770}, {10, 2772}, {10, 2774}, {10, 2776},
+    {10, 2778}, {10, 2780}, {10, 2782}, {10, 2784}, {10, 2786}, {10, 2788},
+    {10, 2790}, {10, 2792}, {10, 2794}, {10, 2796}, {10, 2798}, {10, 2800},
+    {10, 2802}, {10, 2804}, {10, 2806}, {10, 2808}, {10, 2810}, {10, 2812},
+    {10, 2814}, {10, 2816}, {10, 2818}, {10, 2820}, {10, 2822}, {10, 2824},
+    {10, 2826}, {10, 2828}, {10, 2830}, {10, 2832}, {10, 2834}, {10, 2836},
+    {10, 2838}, {10, 2840}, {10, 2842}, {10, 2844}, {10, 2846}, {10, 2848},
+    {10, 2850}, {10, 2852}, {10, 2854}, {10, 2856}, {10, 2858}, {10, 2860},
+    {10, 2862}, {10, 2864}, {10, 2866}, {10, 2868}, {10, 2870}, {10, 2872},
+    {10, 2874}, {10, 2876}, {10, 2878}, {10, 2880}, {10, 2882}, {10, 2884},
+    {10, 2886}, {10, 2888}, {10, 2890}, {10, 2892}, {10, 2894}, {10, 2896},
+    {10, 2898}, {10, 2900}, {10, 2902}, {10, 2904}, {10, 2906}, {10, 2908},
+    {10, 2910}, {10, 2912}, {10, 2914}, {10, 2916}, {10, 2918}, {10, 2920},
+    {10, 2922}, {10, 2924}, {10, 2926}, {10, 2928}, {10, 2930}, {10, 2932},
+    {10, 2934}, {10, 2936}, {10, 2938}, {10, 2940}, {10, 2942}, {10, 2944},
+    {10, 2946}, {10, 2948}, {10, 2950}, {10, 2952}, {10, 2954}, {10, 2956},
+    {10, 2958}, {10, 2960}, {10, 2962}, {10, 2964}, {10, 2966}, {10, 2968},
+    {10, 2970}, {10, 2972}, {10, 2974}, {10, 2976}, {10, 2978}, {10, 2980},
+    {10, 2982}, {10, 2984}, {10, 2986}, {10, 2988}, {10, 2990}, {10, 2992},
+    {10, 2994}, {10, 2996}, {10, 2998}, {10, 3000}, {10, 3002}, {10, 3004},
+    {10, 3006}, {10, 3008}, {10, 3010}, {10, 3012}, {10, 3014}, {10, 3016},
+    {10, 3018}, {10, 3020}, {10, 3022}, {10, 3024}, {10, 3026}, {10, 3028},
+    {10, 3030}, {10, 3032}, {10, 3034}, {10, 3036}, {10, 3038}, {10, 3040},
+    {10, 3042}, {10, 3044}, {10, 3046}, {10, 3048}, {10, 3050}, {10, 3052},
+    {10, 3054}, {10, 3056}, {10, 3058}, {10, 3060}, {10, 3062}, {10, 3064},
+    {10, 3066}, {10, 3068}, {10, 3070}, {10, 3072}, {10, 3074}, {10, 3076},
+    {10, 3078}, {10, 3080}, {10, 3082}, {10, 3084}, {10, 3086}, {10, 3088},
+    {10, 3090}, {10, 3092}, {10, 3094}, {10, 3096}, {10, 3098}, {10, 3100},
+    {10, 3102}, {10, 3104}, {10, 3106}, {10, 3108}, {10, 3110}, {10, 3112},
+    {10, 3114}, {10, 3116}, {10, 3118}, {10, 3120}, {10, 3122}, {10, 3124},
+    {10, 3126}, {10, 3128}, {10, 3130}, {10, 3132}, {10, 3134}, {10, 3136},
+    {10, 3138}, {10, 3140}, {10, 3142}, {10, 3144}, {10, 3146}, {10, 3148},
+    {10, 3150}, {10, 3152}, {10, 3154}, {10, 3156}, {10, 3158}, {10, 3160},
+    {10, 3162}, {10, 3164}, {10, 3166}, {10, 3168}, {10, 3170}, {10, 3172},
+    {10, 3174}, {10, 3176}, {10, 3178}, {10, 3180}, {10, 3182}, {10, 3184},
+    {10, 3186}, {10, 3188}, {10, 3190}, {10, 3192}, {10, 3194}, {10, 3196},
+    {10, 3198}, {10, 3200}, {10, 3202}, {10, 3204}, {10, 3206}, {10, 3208},
+    {10, 3210}, {10, 3212}, {10, 3214}, {10, 3216}, {10, 3218}, {10, 3220},
+    {10, 3222}, {10, 3224}, {10, 3226}, {10, 3228}, {10, 3230}, {10, 3232},
+    {10, 3234}, {10, 3236}, {10, 3238}, {10, 3240}, {10, 3242}, {10, 3244},
+    {10, 3246}, {10, 3248}, {10, 3250}, {10, 3252}, {10, 3254}, {10, 3256},
+    {10, 3258}, {10, 3260}, {10, 3262}, {10, 3264}, {10, 3266}, {10, 3268},
+    {10, 3270}, {10, 3272}, {10, 3274}, {10, 3276}, {10, 3278}, {10, 3280},
+    {10, 3282}, {10, 3284}, {10, 3286}, {10, 3288}, {10, 3290}, {10, 3292},
+    {10, 3294}, {10, 3296}, {10, 3298}, {10, 3300}, {10, 3302}, {10, 3304},
+    {10, 3306}, {10, 3308}, {10, 3310}, {10, 3312}, {10, 3314}, {10, 3316},
+    {10, 3318}, {10, 3320}, {10, 3322}, {10, 3324}, {10, 3326}, {10, 3328},
+    {10, 3330}, {10, 3332}, {10, 3334}, {10, 3336}, {10, 3338}, {10, 3340},
+    {10, 3342}, {10, 3344}, {10, 3346}, {10, 3348}, {10, 3350}, {10, 3352},
+    {10, 3354}, {10, 3356}, {10, 3358}, {10, 3360}, {10, 3362}, {10, 3364},
+    {10, 3366}, {10, 3368}, {10, 3370}, {10, 3372}, {10, 3374}, {10, 3376},
+    {10, 3378}, {10, 3380}, {10, 3382}, {10, 3384}, {10, 3386}, {10, 3388},
+    {10, 3390}, {10, 3392}, {10, 3394}, {10, 3396}, {10, 3398}, {10, 3400},
+    {10, 3402}, {10, 3404}, {10, 3406}, {10, 3408}, {10, 3410}, {10, 3412},
+    {10, 3414}, {10, 3416}, {10, 3418}, {10, 3420}, {10, 3422}, {10, 3424},
+    {10, 3426}, {10, 3428}, {10, 3430}, {10, 3432}, {10, 3434}, {10, 3436},
+    {10, 3438}, {10, 3440}, {10, 3442}, {10, 3444}, {10, 3446}, {10, 3448},
+    {10, 3450}, {10, 3452}, {10, 3454}, {10, 3456}, {10, 3458}, {10, 3460},
+    {10, 3462}, {10, 3464}, {10, 3466}, {10, 3468}, {10, 3470}, {10, 3472},
+    {10, 3474}, {10, 3476}, {10, 3478}, {10, 3480}, {10, 3482}, {10, 3484},
+    {10, 3486}, {10, 3488}, {10, 3490}, {10, 3492}, {10, 3494}, {10, 3496},
+    {10, 3498}, {10, 3500}, {10, 3502}, {10, 3504}, {10, 3506}, {10, 3508},
+    {10, 3510}, {10, 3512}, {10, 3514}, {10, 3516}, {10, 3518}, {10, 3520},
+    {10, 3522}, {10, 3524}, {10, 3526}, {10, 3528}, {10, 3530}, {10, 3532},
+    {10, 3534}, {10, 3536}, {10, 3538}, {10, 3540}, {10, 3542}, {10, 3544},
+    {10, 3546}, {10, 3548}, {10, 3550}, {10, 3552}, {10, 3554}, {10, 3556},
+    {10, 3558}, {10, 3560}, {10, 3562}, {10, 3564}, {10, 3566}, {10, 3568},
+    {10, 3570}, {10, 3572}, {10, 3574}, {10, 3576}, {10, 3578}, {10, 3580},
+    {10, 3582}, {10, 3584}, {10, 3586}, {10, 3588}, {10, 3590}, {10, 3592},
+    {10, 3594}, {10, 3596}, {10, 3598}, {10, 3600}, {10, 3602}, {10, 3604},
+    {10, 3606}, {10, 3608}, {10, 3610}, {10, 3612}, {10, 3614}, {10, 3616},
+    {10, 3618}, {10, 3620}, {10, 3622}, {10, 3624}, {10, 3626}, {10, 3628},
+    {10, 3630}, {10, 3632}, {10, 3634}, {10, 3636}, {10, 3638}, {10, 3640},
+    {10, 3642}, {10, 3644}, {10, 3646}, {10, 3648}, {10, 3650}, {10, 3652},
+    {10, 3654}, {10, 3656}, {10, 3658}, {10, 3660}, {10, 3662}, {10, 3664},
+    {10, 3666}, {10, 3668}, {10, 3670}, {10, 3672}, {10, 3674}, {10, 3676},
+    {10, 3678}, {10, 3680}, {10, 3682}, {10, 3684}, {10, 3686}, {10, 3688},
+    {10, 3690}, {10, 3692}, {10, 3694}, {10, 3696}, {10, 3698}, {10, 3700},
+    {10, 3702}, {10, 3704}, {10, 3706}, {10, 3708}, {10, 3710}, {10, 3712},
+    {10, 3714}, {10, 3716}, {10, 3718}, {10, 3720}, {10, 3722}, {10, 3724},
+    {10, 3726}, {10, 3728}, {10, 3730}, {10, 3732}, {10, 3734}, {10, 3736},
+    {10, 3738}, {10, 3740}, {10, 3742}, {10, 3744}, {10, 3746}, {10, 3748},
+    {10, 3750}, {10, 3752}, {10, 3754}, {10, 3756}, {10, 3758}, {10, 3760},
+    {10, 3762}, {10, 3764}, {10, 3766}, {10, 3768}, {10, 3770}, {10, 3772},
+    {10, 3774}, {10, 3776}, {10, 3778}, {10, 3780}, {10, 3782}, {10, 3784},
+    {10, 3786}, {10, 3788}, {10, 3790}, {10, 3792}, {10, 3794}, {10, 3796},
+    {10, 3798}, {10, 3800}, {10, 3802}, {10, 3804}, {10, 3806}, {10, 3808},
+    {10, 3810}, {10, 3812}, {10, 3814}, {10, 3816}, {10, 3818}, {10, 3820},
+    {10, 3822}, {10, 3824}, {10, 3826}, {10, 3828}, {10, 3830}, {10, 3832},
+    {10, 3834}, {10, 3836}, {10, 3838}, {10, 3840}, {10, 3842}, {10, 3844},
+    {10, 3846}, {10, 3848}, {10, 3850}, {10, 3852}, {10, 3854}, {10, 3856},
+    {10, 3858}, {10, 3860}, {10, 3862}, {10, 3864}, {10, 3866}, {10, 3868},
+    {10, 3870}, {10, 3872}, {10, 3874}, {10, 3876}, {10, 3878}, {10, 3880},
+    {10, 3882}, {10, 3884}, {10, 3886}, {10, 3888}, {10, 3890}, {10, 3892},
+    {10, 3894}, {10, 3896}, {10, 3898}, {10, 3900}, {10, 3902}, {10, 3904},
+    {10, 3906}, {10, 3908}, {10, 3910}, {10, 3912}, {10, 3914}, {10, 3916},
+    {10, 3918}, {10, 3920}, {10, 3922}, {10, 3924}, {10, 3926}, {10, 3928},
+    {10, 3930}, {10, 3932}, {10, 3934}, {10, 3936}, {10, 3938}, {10, 3940},
+    {10, 3942}, {10, 3944}, {10, 3946}, {10, 3948}, {10, 3950}, {10, 3952},
+    {10, 3954}, {10, 3956}, {10, 3958}, {10, 3960}
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_DCT_VALUE_TOKENS_H_
diff --git a/libvpx/libvpx/vp8/encoder/defaultcoefcounts.h b/libvpx/libvpx/vp8/encoder/defaultcoefcounts.h
new file mode 100644
index 0000000..1e8e804
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/defaultcoefcounts.h
@@ -0,0 +1,236 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_ENCODER_DEFAULTCOEFCOUNTS_H_
+#define VP8_ENCODER_DEFAULTCOEFCOUNTS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Generated file, included by entropy.c */
+
+static const unsigned int default_coef_counts[BLOCK_TYPES]
+                                             [COEF_BANDS]
+                                             [PREV_COEF_CONTEXTS]
+                                             [MAX_ENTROPY_TOKENS] =
+{
+
+    {
+        /* Block Type ( 0 ) */
+        {
+            /* Coeff Band ( 0 ) */
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+        },
+        {
+            /* Coeff Band ( 1 ) */
+            {30190, 26544, 225,  24,   4,   0,   0,   0,   0,   0,   0, 4171593,},
+            {26846, 25157, 1241, 130,  26,   6,   1,   0,   0,   0,   0, 149987,},
+            {10484, 9538, 1006, 160,  36,  18,   0,   0,   0,   0,   0, 15104,},
+        },
+        {
+            /* Coeff Band ( 2 ) */
+            {25842, 40456, 1126,  83,  11,   2,   0,   0,   0,   0,   0,   0,},
+            {9338, 8010, 512,  73,   7,   3,   2,   0,   0,   0,   0, 43294,},
+            {1047, 751, 149,  31,  13,   6,   1,   0,   0,   0,   0, 879,},
+        },
+        {
+            /* Coeff Band ( 3 ) */
+            {26136, 9826, 252,  13,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {8134, 5574, 191,  14,   2,   0,   0,   0,   0,   0,   0, 35302,},
+            { 605, 677, 116,   9,   1,   0,   0,   0,   0,   0,   0, 611,},
+        },
+        {
+            /* Coeff Band ( 4 ) */
+            {10263, 15463, 283,  17,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {2773, 2191, 128,   9,   2,   2,   0,   0,   0,   0,   0, 10073,},
+            { 134, 125,  32,   4,   0,   2,   0,   0,   0,   0,   0,  50,},
+        },
+        {
+            /* Coeff Band ( 5 ) */
+            {10483, 2663,  23,   1,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {2137, 1251,  27,   1,   1,   0,   0,   0,   0,   0,   0, 14362,},
+            { 116, 156,  14,   2,   1,   0,   0,   0,   0,   0,   0, 190,},
+        },
+        {
+            /* Coeff Band ( 6 ) */
+            {40977, 27614, 412,  28,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {6113, 5213, 261,  22,   3,   0,   0,   0,   0,   0,   0, 26164,},
+            { 382, 312,  50,  14,   2,   0,   0,   0,   0,   0,   0, 345,},
+        },
+        {
+            /* Coeff Band ( 7 ) */
+            {   0,  26,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,  13,   0,   0,   0,   0,   0,   0,   0,   0,   0, 319,},
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   8,},
+        },
+    },
+    {
+        /* Block Type ( 1 ) */
+        {
+            /* Coeff Band ( 0 ) */
+            {3268, 19382, 1043, 250,  93,  82,  49,  26,  17,   8,  25, 82289,},
+            {8758, 32110, 5436, 1832, 827, 668, 420, 153,  24,   0,   3, 52914,},
+            {9337, 23725, 8487, 3954, 2107, 1836, 1069, 399,  59,   0,   0, 18620,},
+        },
+        {
+            /* Coeff Band ( 1 ) */
+            {12419, 8420, 452,  62,   9,   1,   0,   0,   0,   0,   0,   0,},
+            {11715, 8705, 693,  92,  15,   7,   2,   0,   0,   0,   0, 53988,},
+            {7603, 8585, 2306, 778, 270, 145,  39,   5,   0,   0,   0, 9136,},
+        },
+        {
+            /* Coeff Band ( 2 ) */
+            {15938, 14335, 1207, 184,  55,  13,   4,   1,   0,   0,   0,   0,},
+            {7415, 6829, 1138, 244,  71,  26,   7,   0,   0,   0,   0, 9980,},
+            {1580, 1824, 655, 241,  89,  46,  10,   2,   0,   0,   0, 429,},
+        },
+        {
+            /* Coeff Band ( 3 ) */
+            {19453, 5260, 201,  19,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {9173, 3758, 213,  22,   1,   1,   0,   0,   0,   0,   0, 9820,},
+            {1689, 1277, 276,  51,  17,   4,   0,   0,   0,   0,   0, 679,},
+        },
+        {
+            /* Coeff Band ( 4 ) */
+            {12076, 10667, 620,  85,  19,   9,   5,   0,   0,   0,   0,   0,},
+            {4665, 3625, 423,  55,  19,   9,   0,   0,   0,   0,   0, 5127,},
+            { 415, 440, 143,  34,  20,   7,   2,   0,   0,   0,   0, 101,},
+        },
+        {
+            /* Coeff Band ( 5 ) */
+            {12183, 4846, 115,  11,   1,   0,   0,   0,   0,   0,   0,   0,},
+            {4226, 3149, 177,  21,   2,   0,   0,   0,   0,   0,   0, 7157,},
+            { 375, 621, 189,  51,  11,   4,   1,   0,   0,   0,   0, 198,},
+        },
+        {
+            /* Coeff Band ( 6 ) */
+            {61658, 37743, 1203,  94,  10,   3,   0,   0,   0,   0,   0,   0,},
+            {15514, 11563, 903, 111,  14,   5,   0,   0,   0,   0,   0, 25195,},
+            { 929, 1077, 291,  78,  14,   7,   1,   0,   0,   0,   0, 507,},
+        },
+        {
+            /* Coeff Band ( 7 ) */
+            {   0, 990,  15,   3,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0, 412,  13,   0,   0,   0,   0,   0,   0,   0,   0, 1641,},
+            {   0,  18,   7,   1,   0,   0,   0,   0,   0,   0,   0,  30,},
+        },
+    },
+    {
+        /* Block Type ( 2 ) */
+        {
+            /* Coeff Band ( 0 ) */
+            { 953, 24519, 628, 120,  28,  12,   4,   0,   0,   0,   0, 2248798,},
+            {1525, 25654, 2647, 617, 239, 143,  42,   5,   0,   0,   0, 66837,},
+            {1180, 11011, 3001, 1237, 532, 448, 239,  54,   5,   0,   0, 7122,},
+        },
+        {
+            /* Coeff Band ( 1 ) */
+            {1356, 2220,  67,  10,   4,   1,   0,   0,   0,   0,   0,   0,},
+            {1450, 2544, 102,  18,   4,   3,   0,   0,   0,   0,   0, 57063,},
+            {1182, 2110, 470, 130,  41,  21,   0,   0,   0,   0,   0, 6047,},
+        },
+        {
+            /* Coeff Band ( 2 ) */
+            { 370, 3378, 200,  30,   5,   4,   1,   0,   0,   0,   0,   0,},
+            { 293, 1006, 131,  29,  11,   0,   0,   0,   0,   0,   0, 5404,},
+            { 114, 387,  98,  23,   4,   8,   1,   0,   0,   0,   0, 236,},
+        },
+        {
+            /* Coeff Band ( 3 ) */
+            { 579, 194,   4,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            { 395, 213,   5,   1,   0,   0,   0,   0,   0,   0,   0, 4157,},
+            { 119, 122,   4,   0,   0,   0,   0,   0,   0,   0,   0, 300,},
+        },
+        {
+            /* Coeff Band ( 4 ) */
+            {  38, 557,  19,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {  21, 114,  12,   1,   0,   0,   0,   0,   0,   0,   0, 427,},
+            {   0,   5,   0,   0,   0,   0,   0,   0,   0,   0,   0,   7,},
+        },
+        {
+            /* Coeff Band ( 5 ) */
+            {  52,   7,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {  18,   6,   0,   0,   0,   0,   0,   0,   0,   0,   0, 652,},
+            {   1,   1,   0,   0,   0,   0,   0,   0,   0,   0,   0,  30,},
+        },
+        {
+            /* Coeff Band ( 6 ) */
+            { 640, 569,  10,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {  25,  77,   2,   0,   0,   0,   0,   0,   0,   0,   0, 517,},
+            {   4,   7,   0,   0,   0,   0,   0,   0,   0,   0,   0,   3,},
+        },
+        {
+            /* Coeff Band ( 7 ) */
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+        },
+    },
+    {
+        /* Block Type ( 3 ) */
+        {
+            /* Coeff Band ( 0 ) */
+            {2506, 20161, 2707, 767, 261, 178, 107,  30,  14,   3,   0, 100694,},
+            {8806, 36478, 8817, 3268, 1280, 850, 401, 114,  42,   0,   0, 58572,},
+            {11003, 27214, 11798, 5716, 2482, 2072, 1048, 175,  32,   0,   0, 19284,},
+        },
+        {
+            /* Coeff Band ( 1 ) */
+            {9738, 11313, 959, 205,  70,  18,  11,   1,   0,   0,   0,   0,},
+            {12628, 15085, 1507, 273,  52,  19,   9,   0,   0,   0,   0, 54280,},
+            {10701, 15846, 5561, 1926, 813, 570, 249,  36,   0,   0,   0, 6460,},
+        },
+        {
+            /* Coeff Band ( 2 ) */
+            {6781, 22539, 2784, 634, 182, 123,  20,   4,   0,   0,   0,   0,},
+            {6263, 11544, 2649, 790, 259, 168,  27,   5,   0,   0,   0, 20539,},
+            {3109, 4075, 2031, 896, 457, 386, 158,  29,   0,   0,   0, 1138,},
+        },
+        {
+            /* Coeff Band ( 3 ) */
+            {11515, 4079, 465,  73,   5,  14,   2,   0,   0,   0,   0,   0,},
+            {9361, 5834, 650,  96,  24,   8,   4,   0,   0,   0,   0, 22181,},
+            {4343, 3974, 1360, 415, 132,  96,  14,   1,   0,   0,   0, 1267,},
+        },
+        {
+            /* Coeff Band ( 4 ) */
+            {4787, 9297, 823, 168,  44,  12,   4,   0,   0,   0,   0,   0,},
+            {3619, 4472, 719, 198,  60,  31,   3,   0,   0,   0,   0, 8401,},
+            {1157, 1175, 483, 182,  88,  31,   8,   0,   0,   0,   0, 268,},
+        },
+        {
+            /* Coeff Band ( 5 ) */
+            {8299, 1226,  32,   5,   1,   0,   0,   0,   0,   0,   0,   0,},
+            {3502, 1568,  57,   4,   1,   1,   0,   0,   0,   0,   0, 9811,},
+            {1055, 1070, 166,  29,   6,   1,   0,   0,   0,   0,   0, 527,},
+        },
+        {
+            /* Coeff Band ( 6 ) */
+            {27414, 27927, 1989, 347,  69,  26,   0,   0,   0,   0,   0,   0,},
+            {5876, 10074, 1574, 341,  91,  24,   4,   0,   0,   0,   0, 21954,},
+            {1571, 2171, 778, 324, 124,  65,  16,   0,   0,   0,   0, 979,},
+        },
+        {
+            /* Coeff Band ( 7 ) */
+            {   0,  29,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,},
+            {   0,  23,   0,   0,   0,   0,   0,   0,   0,   0,   0, 459,},
+            {   0,   1,   0,   0,   0,   0,   0,   0,   0,   0,   0,  13,},
+        },
+    },
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_DEFAULTCOEFCOUNTS_H_
diff --git a/libvpx/libvpx/vp8/encoder/denoising.c b/libvpx/libvpx/vp8/encoder/denoising.c
new file mode 100644
index 0000000..26ce120
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/denoising.c
@@ -0,0 +1,761 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+
+#include "denoising.h"
+
+#include "vp8/common/reconinter.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp8_rtcd.h"
+
+static const unsigned int NOISE_MOTION_THRESHOLD = 25 * 25;
+/* SSE_DIFF_THRESHOLD is selected as ~95% confidence assuming
+ * var(noise) ~= 100.
+ */
+static const unsigned int SSE_DIFF_THRESHOLD = 16 * 16 * 20;
+static const unsigned int SSE_THRESHOLD = 16 * 16 * 40;
+static const unsigned int SSE_THRESHOLD_HIGH = 16 * 16 * 80;
+
+/*
+ * The filter function was modified to reduce the computational complexity.
+ * Step 1:
+ * Instead of applying tap coefficients for each pixel, we calculated the
+ * pixel adjustments vs. pixel diff value ahead of time.
+ *     adjustment = filtered_value - current_raw
+ *                = (filter_coefficient * diff + 128) >> 8
+ * where
+ *     filter_coefficient = (255 << 8) / (256 + ((absdiff * 330) >> 3));
+ *     filter_coefficient += filter_coefficient /
+ *                           (3 + motion_magnitude_adjustment);
+ *     filter_coefficient is clamped to 0 ~ 255.
+ *
+ * Step 2:
+ * The adjustment vs. diff curve becomes flat very quick when diff increases.
+ * This allowed us to use only several levels to approximate the curve without
+ * changing the filtering algorithm too much.
+ * The adjustments were further corrected by checking the motion magnitude.
+ * The levels used are:
+ * diff       adjustment w/o motion correction   adjustment w/ motion correction
+ * [-255, -16]           -6                                   -7
+ * [-15, -8]             -4                                   -5
+ * [-7, -4]              -3                                   -4
+ * [-3, 3]               diff                                 diff
+ * [4, 7]                 3                                    4
+ * [8, 15]                4                                    5
+ * [16, 255]              6                                    7
+ */
+
+int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride,
+                          unsigned char *running_avg_y, int avg_y_stride,
+                          unsigned char *sig, int sig_stride,
+                          unsigned int motion_magnitude,
+                          int increase_denoising)
+{
+    unsigned char *running_avg_y_start = running_avg_y;
+    unsigned char *sig_start = sig;
+    int sum_diff_thresh;
+    int r, c;
+    int sum_diff = 0;
+    int adj_val[3] = {3, 4, 6};
+    int shift_inc1 = 0;
+    int shift_inc2 = 1;
+    int col_sum[16] = {0, 0, 0, 0,
+                       0, 0, 0, 0,
+                       0, 0, 0, 0,
+                       0, 0, 0, 0};
+    /* If motion_magnitude is small, making the denoiser more aggressive by
+     * increasing the adjustment for each level. Add another increment for
+     * blocks that are labeled for increase denoising. */
+    if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
+    {
+      if (increase_denoising) {
+        shift_inc1 = 1;
+        shift_inc2 = 2;
+      }
+      adj_val[0] += shift_inc2;
+      adj_val[1] += shift_inc2;
+      adj_val[2] += shift_inc2;
+    }
+
+    for (r = 0; r < 16; ++r)
+    {
+        for (c = 0; c < 16; ++c)
+        {
+            int diff = 0;
+            int adjustment = 0;
+            int absdiff = 0;
+
+            diff = mc_running_avg_y[c] - sig[c];
+            absdiff = abs(diff);
+
+            // When |diff| <= |3 + shift_inc1|, use pixel value from
+            // last denoised raw.
+            if (absdiff <= 3 + shift_inc1)
+            {
+                running_avg_y[c] = mc_running_avg_y[c];
+                col_sum[c] += diff;
+            }
+            else
+            {
+                if (absdiff >= 4 + shift_inc1 && absdiff <= 7)
+                    adjustment = adj_val[0];
+                else if (absdiff >= 8 && absdiff <= 15)
+                    adjustment = adj_val[1];
+                else
+                    adjustment = adj_val[2];
+
+                if (diff > 0)
+                {
+                    if ((sig[c] + adjustment) > 255)
+                        running_avg_y[c] = 255;
+                    else
+                        running_avg_y[c] = sig[c] + adjustment;
+
+                    col_sum[c] += adjustment;
+                }
+                else
+                {
+                    if ((sig[c] - adjustment) < 0)
+                        running_avg_y[c] = 0;
+                    else
+                        running_avg_y[c] = sig[c] - adjustment;
+
+                    col_sum[c] -= adjustment;
+                }
+            }
+        }
+
+        /* Update pointers for next iteration. */
+        sig += sig_stride;
+        mc_running_avg_y += mc_avg_y_stride;
+        running_avg_y += avg_y_stride;
+    }
+
+    for (c = 0; c < 16; ++c) {
+      // Below we clip the value in the same way which SSE code use.
+      // When adopting aggressive denoiser, the adj_val for each pixel
+      // could be at most 8 (this is current max adjustment of the map).
+      // In SSE code, we calculate the sum of adj_val for
+      // the columns, so the sum could be upto 128(16 rows). However,
+      // the range of the value is -128 ~ 127 in SSE code, that's why
+      // we do this change in C code.
+      // We don't do this for UV denoiser, since there are only 8 rows,
+      // and max adjustments <= 8, so the sum of the columns will not
+      // exceed 64.
+      if (col_sum[c] >= 128) {
+        col_sum[c] = 127;
+      }
+      sum_diff += col_sum[c];
+    }
+
+    sum_diff_thresh= SUM_DIFF_THRESHOLD;
+    if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
+    if (abs(sum_diff) > sum_diff_thresh) {
+      // Before returning to copy the block (i.e., apply no denoising), check
+      // if we can still apply some (weaker) temporal filtering to this block,
+      // that would otherwise not be denoised at all. Simplest is to apply
+      // an additional adjustment to running_avg_y to bring it closer to sig.
+      // The adjustment is capped by a maximum delta, and chosen such that
+      // in most cases the resulting sum_diff will be within the
+      // accceptable range given by sum_diff_thresh.
+
+      // The delta is set by the excess of absolute pixel diff over threshold.
+      int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        sig -= sig_stride * 16;
+        mc_running_avg_y -= mc_avg_y_stride * 16;
+        running_avg_y -= avg_y_stride * 16;
+        for (r = 0; r < 16; ++r) {
+          for (c = 0; c < 16; ++c) {
+            int diff = mc_running_avg_y[c] - sig[c];
+            int adjustment = abs(diff);
+            if (adjustment > delta)
+              adjustment = delta;
+            if (diff > 0) {
+              // Bring denoised signal down.
+              if (running_avg_y[c] - adjustment < 0)
+                running_avg_y[c] = 0;
+              else
+                running_avg_y[c] = running_avg_y[c] - adjustment;
+              col_sum[c] -= adjustment;
+            } else if (diff < 0) {
+              // Bring denoised signal up.
+              if (running_avg_y[c] + adjustment > 255)
+                running_avg_y[c] = 255;
+              else
+                running_avg_y[c] = running_avg_y[c] + adjustment;
+              col_sum[c] += adjustment;
+            }
+          }
+          // TODO(marpan): Check here if abs(sum_diff) has gone below the
+          // threshold sum_diff_thresh, and if so, we can exit the row loop.
+          sig += sig_stride;
+          mc_running_avg_y += mc_avg_y_stride;
+          running_avg_y += avg_y_stride;
+        }
+
+        sum_diff = 0;
+        for (c = 0; c < 16; ++c) {
+          if (col_sum[c] >= 128) {
+            col_sum[c] = 127;
+          }
+          sum_diff += col_sum[c];
+        }
+
+        if (abs(sum_diff) > sum_diff_thresh)
+          return COPY_BLOCK;
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+
+    vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
+    return FILTER_BLOCK;
+}
+
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg_uv,
+                             int mc_avg_uv_stride,
+                             unsigned char *running_avg_uv,
+                             int avg_uv_stride,
+                             unsigned char *sig,
+                             int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    unsigned char *running_avg_uv_start = running_avg_uv;
+    unsigned char *sig_start = sig;
+    int sum_diff_thresh;
+    int r, c;
+    int sum_diff = 0;
+    int sum_block = 0;
+    int adj_val[3] = {3, 4, 6};
+    int shift_inc1 = 0;
+    int shift_inc2 = 1;
+    /* If motion_magnitude is small, making the denoiser more aggressive by
+     * increasing the adjustment for each level. Add another increment for
+     * blocks that are labeled for increase denoising. */
+    if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) {
+      if (increase_denoising) {
+        shift_inc1 = 1;
+        shift_inc2 = 2;
+      }
+      adj_val[0] += shift_inc2;
+      adj_val[1] += shift_inc2;
+      adj_val[2] += shift_inc2;
+    }
+
+    // Avoid denoising color signal if its close to average level.
+    for (r = 0; r < 8; ++r) {
+      for (c = 0; c < 8; ++c) {
+        sum_block += sig[c];
+      }
+      sig += sig_stride;
+    }
+    if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
+      return COPY_BLOCK;
+    }
+
+    sig -= sig_stride * 8;
+    for (r = 0; r < 8; ++r) {
+      for (c = 0; c < 8; ++c) {
+        int diff = 0;
+        int adjustment = 0;
+        int absdiff = 0;
+
+        diff = mc_running_avg_uv[c] - sig[c];
+        absdiff = abs(diff);
+
+        // When |diff| <= |3 + shift_inc1|, use pixel value from
+        // last denoised raw.
+        if (absdiff <= 3 + shift_inc1) {
+          running_avg_uv[c] = mc_running_avg_uv[c];
+          sum_diff += diff;
+        } else {
+          if (absdiff >= 4 && absdiff <= 7)
+            adjustment = adj_val[0];
+          else if (absdiff >= 8 && absdiff <= 15)
+            adjustment = adj_val[1];
+          else
+            adjustment = adj_val[2];
+          if (diff > 0) {
+            if ((sig[c] + adjustment) > 255)
+              running_avg_uv[c] = 255;
+            else
+              running_avg_uv[c] = sig[c] + adjustment;
+            sum_diff += adjustment;
+          } else {
+            if ((sig[c] - adjustment) < 0)
+              running_avg_uv[c] = 0;
+            else
+              running_avg_uv[c] = sig[c] - adjustment;
+            sum_diff -= adjustment;
+          }
+        }
+      }
+      /* Update pointers for next iteration. */
+      sig += sig_stride;
+      mc_running_avg_uv += mc_avg_uv_stride;
+      running_avg_uv += avg_uv_stride;
+    }
+
+    sum_diff_thresh= SUM_DIFF_THRESHOLD_UV;
+    if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+    if (abs(sum_diff) > sum_diff_thresh) {
+      // Before returning to copy the block (i.e., apply no denoising), check
+      // if we can still apply some (weaker) temporal filtering to this block,
+      // that would otherwise not be denoised at all. Simplest is to apply
+      // an additional adjustment to running_avg_y to bring it closer to sig.
+      // The adjustment is capped by a maximum delta, and chosen such that
+      // in most cases the resulting sum_diff will be within the
+      // accceptable range given by sum_diff_thresh.
+
+      // The delta is set by the excess of absolute pixel diff over threshold.
+      int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        sig -= sig_stride * 8;
+        mc_running_avg_uv -= mc_avg_uv_stride * 8;
+        running_avg_uv -= avg_uv_stride * 8;
+        for (r = 0; r < 8; ++r) {
+          for (c = 0; c < 8; ++c) {
+            int diff = mc_running_avg_uv[c] - sig[c];
+            int adjustment = abs(diff);
+            if (adjustment > delta)
+              adjustment = delta;
+            if (diff > 0) {
+              // Bring denoised signal down.
+              if (running_avg_uv[c] - adjustment < 0)
+                running_avg_uv[c] = 0;
+              else
+                running_avg_uv[c] = running_avg_uv[c] - adjustment;
+              sum_diff -= adjustment;
+            } else if (diff < 0) {
+              // Bring denoised signal up.
+              if (running_avg_uv[c] + adjustment > 255)
+                running_avg_uv[c] = 255;
+              else
+                running_avg_uv[c] = running_avg_uv[c] + adjustment;
+              sum_diff += adjustment;
+            }
+          }
+          // TODO(marpan): Check here if abs(sum_diff) has gone below the
+          // threshold sum_diff_thresh, and if so, we can exit the row loop.
+          sig += sig_stride;
+          mc_running_avg_uv += mc_avg_uv_stride;
+          running_avg_uv += avg_uv_stride;
+        }
+        if (abs(sum_diff) > sum_diff_thresh)
+          return COPY_BLOCK;
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+
+    vp8_copy_mem8x8(running_avg_uv_start, avg_uv_stride, sig_start,
+                    sig_stride);
+    return FILTER_BLOCK;
+}
+
+void vp8_denoiser_set_parameters(VP8_DENOISER *denoiser, int mode) {
+  assert(mode > 0);  // Denoiser is allocated only if mode > 0.
+  if (mode == 1) {
+    denoiser->denoiser_mode = kDenoiserOnYOnly;
+  } else if (mode == 2) {
+    denoiser->denoiser_mode = kDenoiserOnYUV;
+  } else if (mode == 3) {
+    denoiser->denoiser_mode = kDenoiserOnYUVAggressive;
+  } else {
+    denoiser->denoiser_mode = kDenoiserOnYUV;
+  }
+  if (denoiser->denoiser_mode != kDenoiserOnYUVAggressive) {
+    denoiser->denoise_pars.scale_sse_thresh = 1;
+    denoiser->denoise_pars.scale_motion_thresh = 8;
+    denoiser->denoise_pars.scale_increase_filter = 0;
+    denoiser->denoise_pars.denoise_mv_bias = 95;
+    denoiser->denoise_pars.pickmode_mv_bias = 100;
+    denoiser->denoise_pars.qp_thresh = 0;
+    denoiser->denoise_pars.consec_zerolast = UINT_MAX;
+    denoiser->denoise_pars.spatial_blur = 0;
+  } else {
+    denoiser->denoise_pars.scale_sse_thresh = 2;
+    denoiser->denoise_pars.scale_motion_thresh = 16;
+    denoiser->denoise_pars.scale_increase_filter = 1;
+    denoiser->denoise_pars.denoise_mv_bias = 60;
+    denoiser->denoise_pars.pickmode_mv_bias = 75;
+    denoiser->denoise_pars.qp_thresh = 80;
+    denoiser->denoise_pars.consec_zerolast = 15;
+    denoiser->denoise_pars.spatial_blur = 0;
+  }
+}
+
+int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
+                          int num_mb_rows, int num_mb_cols, int mode)
+{
+    int i;
+    assert(denoiser);
+    denoiser->num_mb_cols = num_mb_cols;
+
+    for (i = 0; i < MAX_REF_FRAMES; i++)
+    {
+        denoiser->yv12_running_avg[i].flags = 0;
+
+        if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_running_avg[i]), width,
+                                        height, VP8BORDERINPIXELS)
+            < 0)
+        {
+            vp8_denoiser_free(denoiser);
+            return 1;
+        }
+        memset(denoiser->yv12_running_avg[i].buffer_alloc, 0,
+               denoiser->yv12_running_avg[i].frame_size);
+
+    }
+    denoiser->yv12_mc_running_avg.flags = 0;
+
+    if (vp8_yv12_alloc_frame_buffer(&(denoiser->yv12_mc_running_avg), width,
+                                   height, VP8BORDERINPIXELS) < 0)
+    {
+        vp8_denoiser_free(denoiser);
+        return 1;
+    }
+
+    memset(denoiser->yv12_mc_running_avg.buffer_alloc, 0,
+           denoiser->yv12_mc_running_avg.frame_size);
+
+    if (vp8_yv12_alloc_frame_buffer(&denoiser->yv12_last_source, width,
+                                    height, VP8BORDERINPIXELS) < 0) {
+      vp8_denoiser_free(denoiser);
+      return 1;
+    }
+    memset(denoiser->yv12_last_source.buffer_alloc, 0,
+           denoiser->yv12_last_source.frame_size);
+
+    denoiser->denoise_state = vpx_calloc((num_mb_rows * num_mb_cols), 1);
+    if (!denoiser->denoise_state)
+    {
+        vp8_denoiser_free(denoiser);
+        return 1;
+    }
+    memset(denoiser->denoise_state, 0, (num_mb_rows * num_mb_cols));
+    vp8_denoiser_set_parameters(denoiser, mode);
+    denoiser->nmse_source_diff = 0;
+    denoiser->nmse_source_diff_count = 0;
+    denoiser->qp_avg = 0;
+    // QP threshold below which we can go up to aggressive mode.
+    denoiser->qp_threshold_up = 80;
+    // QP threshold above which we can go back down to normal mode.
+    // For now keep this second threshold high, so not used currently.
+    denoiser->qp_threshold_down = 128;
+    // Bitrate thresholds and noise metric (nmse) thresholds for switching to
+    // aggressive mode.
+    // TODO(marpan): Adjust thresholds, including effect on resolution.
+    denoiser->bitrate_threshold = 400000;  // (bits/sec).
+    denoiser->threshold_aggressive_mode = 80;
+    if (width * height > 1280 * 720) {
+      denoiser->bitrate_threshold = 3000000;
+      denoiser->threshold_aggressive_mode = 200;
+    } else if (width * height > 960 * 540) {
+      denoiser->bitrate_threshold = 1200000;
+      denoiser->threshold_aggressive_mode = 120;
+    } else if (width * height > 640 * 480) {
+      denoiser->bitrate_threshold = 600000;
+      denoiser->threshold_aggressive_mode = 100;
+    }
+    return 0;
+}
+
+
+void vp8_denoiser_free(VP8_DENOISER *denoiser)
+{
+    int i;
+    assert(denoiser);
+
+    for (i = 0; i < MAX_REF_FRAMES ; i++)
+    {
+        vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_running_avg[i]);
+    }
+    vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_mc_running_avg);
+    vp8_yv12_de_alloc_frame_buffer(&denoiser->yv12_last_source);
+    vpx_free(denoiser->denoise_state);
+}
+
+void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
+                             MACROBLOCK *x,
+                             unsigned int best_sse,
+                             unsigned int zero_mv_sse,
+                             int recon_yoffset,
+                             int recon_uvoffset,
+                             loop_filter_info_n *lfi_n,
+                             int mb_row,
+                             int mb_col,
+                             int block_index,
+                             int consec_zero_last)
+
+{
+    int mv_row;
+    int mv_col;
+    unsigned int motion_threshold;
+    unsigned int motion_magnitude2;
+    unsigned int sse_thresh;
+    int sse_diff_thresh = 0;
+    // Spatial loop filter: only applied selectively based on
+    // temporal filter state of block relative to top/left neighbors.
+    int apply_spatial_loop_filter = 1;
+    MV_REFERENCE_FRAME frame = x->best_reference_frame;
+    MV_REFERENCE_FRAME zero_frame = x->best_zeromv_reference_frame;
+
+    enum vp8_denoiser_decision decision = FILTER_BLOCK;
+    enum vp8_denoiser_decision decision_u = COPY_BLOCK;
+    enum vp8_denoiser_decision decision_v = COPY_BLOCK;
+
+    if (zero_frame)
+    {
+        YV12_BUFFER_CONFIG *src = &denoiser->yv12_running_avg[frame];
+        YV12_BUFFER_CONFIG *dst = &denoiser->yv12_mc_running_avg;
+        YV12_BUFFER_CONFIG saved_pre,saved_dst;
+        MB_MODE_INFO saved_mbmi;
+        MACROBLOCKD *filter_xd = &x->e_mbd;
+        MB_MODE_INFO *mbmi = &filter_xd->mode_info_context->mbmi;
+        int sse_diff = 0;
+        // Bias on zero motion vector sse.
+        const int zero_bias = denoiser->denoise_pars.denoise_mv_bias;
+        zero_mv_sse = (unsigned int)((int64_t)zero_mv_sse * zero_bias / 100);
+        sse_diff = (int)zero_mv_sse - (int)best_sse;
+
+        saved_mbmi = *mbmi;
+
+        /* Use the best MV for the compensation. */
+        mbmi->ref_frame = x->best_reference_frame;
+        mbmi->mode = x->best_sse_inter_mode;
+        mbmi->mv = x->best_sse_mv;
+        mbmi->need_to_clamp_mvs = x->need_to_clamp_best_mvs;
+        mv_col = x->best_sse_mv.as_mv.col;
+        mv_row = x->best_sse_mv.as_mv.row;
+        // Bias to zero_mv if small amount of motion.
+        // Note sse_diff_thresh is intialized to zero, so this ensures
+        // we will always choose zero_mv for denoising if
+        // zero_mv_see <= best_sse (i.e., sse_diff <= 0).
+        if ((unsigned int)(mv_row * mv_row + mv_col * mv_col)
+            <= NOISE_MOTION_THRESHOLD)
+            sse_diff_thresh = (int)SSE_DIFF_THRESHOLD;
+
+        if (frame == INTRA_FRAME ||
+            sse_diff <= sse_diff_thresh)
+        {
+            /*
+             * Handle intra blocks as referring to last frame with zero motion
+             * and let the absolute pixel difference affect the filter factor.
+             * Also consider small amount of motion as being random walk due
+             * to noise, if it doesn't mean that we get a much bigger error.
+             * Note that any changes to the mode info only affects the
+             * denoising.
+             */
+            x->denoise_zeromv = 1;
+            mbmi->ref_frame =
+                    x->best_zeromv_reference_frame;
+
+            src = &denoiser->yv12_running_avg[zero_frame];
+
+            mbmi->mode = ZEROMV;
+            mbmi->mv.as_int = 0;
+            x->best_sse_inter_mode = ZEROMV;
+            x->best_sse_mv.as_int = 0;
+            best_sse = zero_mv_sse;
+        }
+
+        mv_row = x->best_sse_mv.as_mv.row;
+        mv_col = x->best_sse_mv.as_mv.col;
+        motion_magnitude2 = mv_row * mv_row + mv_col * mv_col;
+        motion_threshold = denoiser->denoise_pars.scale_motion_thresh *
+            NOISE_MOTION_THRESHOLD;
+
+        if (motion_magnitude2 <
+          denoiser->denoise_pars.scale_increase_filter * NOISE_MOTION_THRESHOLD)
+          x->increase_denoising = 1;
+
+        sse_thresh = denoiser->denoise_pars.scale_sse_thresh * SSE_THRESHOLD;
+        if (x->increase_denoising)
+          sse_thresh =
+              denoiser->denoise_pars.scale_sse_thresh * SSE_THRESHOLD_HIGH;
+
+        if (best_sse > sse_thresh || motion_magnitude2 > motion_threshold)
+          decision = COPY_BLOCK;
+
+        // If block is considered skin, don't denoise if the block
+        // (1) is selected as non-zero motion for current frame, or
+        // (2) has not been selected as ZERO_LAST mode at least x past frames
+        // in a row.
+        // TODO(marpan): Parameter "x" should be varied with framerate.
+        // In particualar, should be reduced for layers (base layer/LAST).
+        if (x->is_skin && (consec_zero_last < 2 || motion_magnitude2 > 0))
+          decision = COPY_BLOCK;
+
+        if (decision == FILTER_BLOCK) {
+          saved_pre = filter_xd->pre;
+          saved_dst = filter_xd->dst;
+
+          /* Compensate the running average. */
+          filter_xd->pre.y_buffer = src->y_buffer + recon_yoffset;
+          filter_xd->pre.u_buffer = src->u_buffer + recon_uvoffset;
+          filter_xd->pre.v_buffer = src->v_buffer + recon_uvoffset;
+          /* Write the compensated running average to the destination buffer. */
+          filter_xd->dst.y_buffer = dst->y_buffer + recon_yoffset;
+          filter_xd->dst.u_buffer = dst->u_buffer + recon_uvoffset;
+          filter_xd->dst.v_buffer = dst->v_buffer + recon_uvoffset;
+
+          if (!x->skip)
+          {
+              vp8_build_inter_predictors_mb(filter_xd);
+          }
+          else
+          {
+              vp8_build_inter16x16_predictors_mb(filter_xd,
+                                                 filter_xd->dst.y_buffer,
+                                                 filter_xd->dst.u_buffer,
+                                                 filter_xd->dst.v_buffer,
+                                                 filter_xd->dst.y_stride,
+                                                 filter_xd->dst.uv_stride);
+          }
+          filter_xd->pre = saved_pre;
+          filter_xd->dst = saved_dst;
+          *mbmi = saved_mbmi;
+        }
+    } else {
+      // zero_frame should always be 1 for real-time mode, as the
+      // ZEROMV mode is always checked, so we should never go into this branch.
+      // If case ZEROMV is not checked, then we will force no denoise (COPY).
+      decision = COPY_BLOCK;
+    }
+
+    if (decision == FILTER_BLOCK)
+    {
+        unsigned char *mc_running_avg_y =
+            denoiser->yv12_mc_running_avg.y_buffer + recon_yoffset;
+        int mc_avg_y_stride = denoiser->yv12_mc_running_avg.y_stride;
+        unsigned char *running_avg_y =
+            denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset;
+        int avg_y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride;
+
+        /* Filter. */
+        decision = vp8_denoiser_filter(mc_running_avg_y, mc_avg_y_stride,
+                                       running_avg_y, avg_y_stride,
+                                       x->thismb, 16, motion_magnitude2,
+                                       x->increase_denoising);
+        denoiser->denoise_state[block_index] = motion_magnitude2 > 0 ?
+            kFilterNonZeroMV : kFilterZeroMV;
+        // Only denoise UV for zero motion, and if y channel was denoised.
+        if (denoiser->denoiser_mode != kDenoiserOnYOnly &&
+            motion_magnitude2 == 0 &&
+            decision == FILTER_BLOCK) {
+          unsigned char *mc_running_avg_u =
+              denoiser->yv12_mc_running_avg.u_buffer + recon_uvoffset;
+          unsigned char *running_avg_u =
+              denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset;
+          unsigned char *mc_running_avg_v =
+              denoiser->yv12_mc_running_avg.v_buffer + recon_uvoffset;
+          unsigned char *running_avg_v =
+              denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset;
+          int mc_avg_uv_stride = denoiser->yv12_mc_running_avg.uv_stride;
+          int avg_uv_stride = denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
+          int signal_stride = x->block[16].src_stride;
+          decision_u =
+              vp8_denoiser_filter_uv(mc_running_avg_u, mc_avg_uv_stride,
+                                      running_avg_u, avg_uv_stride,
+                                      x->block[16].src + *x->block[16].base_src,
+                                      signal_stride, motion_magnitude2, 0);
+          decision_v =
+              vp8_denoiser_filter_uv(mc_running_avg_v, mc_avg_uv_stride,
+                                      running_avg_v, avg_uv_stride,
+                                      x->block[20].src + *x->block[20].base_src,
+                                      signal_stride, motion_magnitude2, 0);
+        }
+    }
+    if (decision == COPY_BLOCK)
+    {
+        /* No filtering of this block; it differs too much from the predictor,
+         * or the motion vector magnitude is considered too big.
+         */
+        x->denoise_zeromv = 0;
+        vp8_copy_mem16x16(
+                x->thismb, 16,
+                denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
+                denoiser->yv12_running_avg[INTRA_FRAME].y_stride);
+        denoiser->denoise_state[block_index] = kNoFilter;
+    }
+    if (denoiser->denoiser_mode != kDenoiserOnYOnly) {
+      if (decision_u == COPY_BLOCK) {
+        vp8_copy_mem8x8(
+            x->block[16].src + *x->block[16].base_src, x->block[16].src_stride,
+            denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset,
+            denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
+      }
+      if (decision_v == COPY_BLOCK) {
+        vp8_copy_mem8x8(
+            x->block[20].src + *x->block[20].base_src, x->block[16].src_stride,
+            denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset,
+            denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
+      }
+    }
+    // Option to selectively deblock the denoised signal, for y channel only.
+    if (apply_spatial_loop_filter) {
+      loop_filter_info lfi;
+      int apply_filter_col = 0;
+      int apply_filter_row = 0;
+      int apply_filter = 0;
+      int y_stride = denoiser->yv12_running_avg[INTRA_FRAME].y_stride;
+      int uv_stride =denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
+
+      // Fix filter level to some nominal value for now.
+      int filter_level = 48;
+
+      int hev_index = lfi_n->hev_thr_lut[INTER_FRAME][filter_level];
+      lfi.mblim = lfi_n->mblim[filter_level];
+      lfi.blim = lfi_n->blim[filter_level];
+      lfi.lim = lfi_n->lim[filter_level];
+      lfi.hev_thr = lfi_n->hev_thr[hev_index];
+
+      // Apply filter if there is a difference in the denoiser filter state
+      // between the current and left/top block, or if non-zero motion vector
+      // is used for the motion-compensated filtering.
+      if (mb_col > 0) {
+        apply_filter_col = !((denoiser->denoise_state[block_index] ==
+            denoiser->denoise_state[block_index - 1]) &&
+            denoiser->denoise_state[block_index] != kFilterNonZeroMV);
+        if (apply_filter_col) {
+          // Filter left vertical edge.
+          apply_filter = 1;
+          vp8_loop_filter_mbv(
+              denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
+              NULL, NULL, y_stride, uv_stride, &lfi);
+        }
+      }
+      if (mb_row > 0) {
+        apply_filter_row = !((denoiser->denoise_state[block_index] ==
+            denoiser->denoise_state[block_index - denoiser->num_mb_cols]) &&
+            denoiser->denoise_state[block_index] != kFilterNonZeroMV);
+        if (apply_filter_row) {
+          // Filter top horizontal edge.
+          apply_filter = 1;
+          vp8_loop_filter_mbh(
+              denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
+              NULL, NULL, y_stride, uv_stride, &lfi);
+        }
+      }
+      if (apply_filter) {
+        // Update the signal block |x|. Pixel changes are only to top and/or
+        // left boundary pixels: can we avoid full block copy here.
+        vp8_copy_mem16x16(
+            denoiser->yv12_running_avg[INTRA_FRAME].y_buffer + recon_yoffset,
+            y_stride, x->thismb, 16);
+      }
+    }
+}
diff --git a/libvpx/libvpx/vp8/encoder/denoising.h b/libvpx/libvpx/vp8/encoder/denoising.h
new file mode 100644
index 0000000..8c126c1
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/denoising.h
@@ -0,0 +1,118 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_ENCODER_DENOISING_H_
+#define VP8_ENCODER_DENOISING_H_
+
+#include "block.h"
+#include "vp8/common/loopfilter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define SUM_DIFF_THRESHOLD 512
+#define SUM_DIFF_THRESHOLD_HIGH 600
+#define MOTION_MAGNITUDE_THRESHOLD (8*3)
+
+#define SUM_DIFF_THRESHOLD_UV (96)   // (8 * 8 * 1.5)
+#define SUM_DIFF_THRESHOLD_HIGH_UV (8 * 8 * 2)
+#define SUM_DIFF_FROM_AVG_THRESH_UV (8 * 8 * 8)
+#define MOTION_MAGNITUDE_THRESHOLD_UV (8*3)
+
+#define MAX_GF_ARF_DENOISE_RANGE (8)
+
+enum vp8_denoiser_decision
+{
+  COPY_BLOCK,
+  FILTER_BLOCK
+};
+
+enum vp8_denoiser_filter_state {
+  kNoFilter,
+  kFilterZeroMV,
+  kFilterNonZeroMV
+};
+
+enum vp8_denoiser_mode {
+  kDenoiserOff,
+  kDenoiserOnYOnly,
+  kDenoiserOnYUV,
+  kDenoiserOnYUVAggressive,
+  kDenoiserOnAdaptive
+};
+
+typedef struct {
+  // Scale factor on sse threshold above which no denoising is done.
+  unsigned int scale_sse_thresh;
+  // Scale factor on motion magnitude threshold above which no
+  // denoising is done.
+  unsigned int scale_motion_thresh;
+  // Scale factor on motion magnitude below which we increase the strength of
+  // the temporal filter (in function vp8_denoiser_filter).
+  unsigned int scale_increase_filter;
+  // Scale factor to bias to ZEROMV for denoising.
+  unsigned int denoise_mv_bias;
+  // Scale factor to bias to ZEROMV for coding mode selection.
+  unsigned int pickmode_mv_bias;
+  // Quantizer threshold below which we use the segmentation map to switch off
+  // loop filter for blocks that have been coded as ZEROMV-LAST a certain number
+  // (consec_zerolast) of consecutive frames. Note that the delta-QP is set to
+  // 0 when segmentation map is used for shutting off loop filter.
+  unsigned int qp_thresh;
+  // Threshold for number of consecutive frames for blocks coded as ZEROMV-LAST.
+  unsigned int consec_zerolast;
+  // Threshold for amount of spatial blur on Y channel. 0 means no spatial blur.
+  unsigned int spatial_blur;
+} denoise_params;
+
+typedef struct vp8_denoiser
+{
+    YV12_BUFFER_CONFIG yv12_running_avg[MAX_REF_FRAMES];
+    YV12_BUFFER_CONFIG yv12_mc_running_avg;
+    // TODO(marpan): Should remove yv12_last_source and use vp8_lookahead_peak.
+    YV12_BUFFER_CONFIG yv12_last_source;
+    unsigned char* denoise_state;
+    int num_mb_cols;
+    int denoiser_mode;
+    int threshold_aggressive_mode;
+    int nmse_source_diff;
+    int nmse_source_diff_count;
+    int qp_avg;
+    int qp_threshold_up;
+    int qp_threshold_down;
+    int bitrate_threshold;
+    denoise_params denoise_pars;
+} VP8_DENOISER;
+
+int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
+                          int num_mb_rows, int num_mb_cols, int mode);
+
+void vp8_denoiser_free(VP8_DENOISER *denoiser);
+
+void vp8_denoiser_set_parameters(VP8_DENOISER *denoiser, int mode);
+
+void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
+                             MACROBLOCK *x,
+                             unsigned int best_sse,
+                             unsigned int zero_mv_sse,
+                             int recon_yoffset,
+                             int recon_uvoffset,
+                             loop_filter_info_n *lfi_n,
+                             int mb_row,
+                             int mb_col,
+                             int block_index,
+                             int consec_zero_last);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_DENOISING_H_
diff --git a/libvpx/libvpx/vp8/encoder/encodeframe.c b/libvpx/libvpx/vp8/encoder/encodeframe.c
new file mode 100644
index 0000000..9b05cd1
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/encodeframe.c
@@ -0,0 +1,1424 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "encodemb.h"
+#include "encodemv.h"
+#include "vp8/common/common.h"
+#include "onyx_int.h"
+#include "vp8/common/extend.h"
+#include "vp8/common/entropymode.h"
+#include "vp8/common/quant_common.h"
+#include "segmentation.h"
+#include "vp8/common/setupintrarecon.h"
+#include "encodeintra.h"
+#include "vp8/common/reconinter.h"
+#include "rdopt.h"
+#include "pickinter.h"
+#include "vp8/common/findnearmv.h"
+#include <stdio.h>
+#include <limits.h>
+#include "vp8/common/invtrans.h"
+#include "vpx_ports/vpx_timer.h"
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+#include "bitstream.h"
+#endif
+#include "encodeframe.h"
+
+extern void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t) ;
+extern void vp8_calc_ref_frame_costs(int *ref_frame_cost,
+                                     int prob_intra,
+                                     int prob_last,
+                                     int prob_garf
+                                    );
+extern void vp8_convert_rfct_to_prob(VP8_COMP *const cpi);
+extern void vp8cx_initialize_me_consts(VP8_COMP *cpi, int QIndex);
+extern void vp8_auto_select_speed(VP8_COMP *cpi);
+extern void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
+                                      MACROBLOCK *x,
+                                      MB_ROW_COMP *mbr_ei,
+                                      int count);
+static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x );
+
+#ifdef MODE_STATS
+unsigned int inter_y_modes[10] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+unsigned int inter_uv_modes[4] = {0, 0, 0, 0};
+unsigned int inter_b_modes[15]  = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+unsigned int y_modes[5]   = {0, 0, 0, 0, 0};
+unsigned int uv_modes[4]  = {0, 0, 0, 0};
+unsigned int b_modes[14]  = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+#endif
+
+
+/* activity_avg must be positive, or flat regions could get a zero weight
+ *  (infinite lambda), which confounds analysis.
+ * This also avoids the need for divide by zero checks in
+ *  vp8_activity_masking().
+ */
+#define VP8_ACTIVITY_AVG_MIN (64)
+
+/* This is used as a reference when computing the source variance for the
+ *  purposes of activity masking.
+ * Eventually this should be replaced by custom no-reference routines,
+ *  which will be faster.
+ */
+static const unsigned char VP8_VAR_OFFS[16]=
+{
+    128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128
+};
+
+
+/* Original activity measure from Tim T's code. */
+static unsigned int tt_activity_measure( VP8_COMP *cpi, MACROBLOCK *x )
+{
+    unsigned int act;
+    unsigned int sse;
+    (void)cpi;
+    /* TODO: This could also be done over smaller areas (8x8), but that would
+     *  require extensive changes elsewhere, as lambda is assumed to be fixed
+     *  over an entire MB in most of the code.
+     * Another option is to compute four 8x8 variances, and pick a single
+     *  lambda using a non-linear combination (e.g., the smallest, or second
+     *  smallest, etc.).
+     */
+    act =  vpx_variance16x16(x->src.y_buffer,
+                    x->src.y_stride, VP8_VAR_OFFS, 0, &sse);
+    act = act<<4;
+
+    /* If the region is flat, lower the activity some more. */
+    if (act < 8<<12)
+        act = act < 5<<12 ? act : 5<<12;
+
+    return act;
+}
+
+/* Stub for alternative experimental activity measures. */
+static unsigned int alt_activity_measure( VP8_COMP *cpi,
+                                          MACROBLOCK *x, int use_dc_pred )
+{
+    return vp8_encode_intra(cpi,x, use_dc_pred);
+}
+
+
+/* Measure the activity of the current macroblock
+ * What we measure here is TBD so abstracted to this function
+ */
+#define ALT_ACT_MEASURE 1
+static unsigned int mb_activity_measure( VP8_COMP *cpi, MACROBLOCK *x,
+                                  int mb_row, int mb_col)
+{
+    unsigned int mb_activity;
+
+    if  ( ALT_ACT_MEASURE )
+    {
+        int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
+
+        /* Or use and alternative. */
+        mb_activity = alt_activity_measure( cpi, x, use_dc_pred );
+    }
+    else
+    {
+        /* Original activity measure from Tim T's code. */
+        mb_activity = tt_activity_measure( cpi, x );
+    }
+
+    if ( mb_activity < VP8_ACTIVITY_AVG_MIN )
+        mb_activity = VP8_ACTIVITY_AVG_MIN;
+
+    return mb_activity;
+}
+
+/* Calculate an "average" mb activity value for the frame */
+#define ACT_MEDIAN 0
+static void calc_av_activity( VP8_COMP *cpi, int64_t activity_sum )
+{
+#if ACT_MEDIAN
+    /* Find median: Simple n^2 algorithm for experimentation */
+    {
+        unsigned int median;
+        unsigned int i,j;
+        unsigned int * sortlist;
+        unsigned int tmp;
+
+        /* Create a list to sort to */
+        CHECK_MEM_ERROR(sortlist,
+                        vpx_calloc(sizeof(unsigned int),
+                        cpi->common.MBs));
+
+        /* Copy map to sort list */
+        memcpy( sortlist, cpi->mb_activity_map,
+                sizeof(unsigned int) * cpi->common.MBs );
+
+
+        /* Ripple each value down to its correct position */
+        for ( i = 1; i < cpi->common.MBs; i ++ )
+        {
+            for ( j = i; j > 0; j -- )
+            {
+                if ( sortlist[j] < sortlist[j-1] )
+                {
+                    /* Swap values */
+                    tmp = sortlist[j-1];
+                    sortlist[j-1] = sortlist[j];
+                    sortlist[j] = tmp;
+                }
+                else
+                    break;
+            }
+        }
+
+        /* Even number MBs so estimate median as mean of two either side. */
+        median = ( 1 + sortlist[cpi->common.MBs >> 1] +
+                   sortlist[(cpi->common.MBs >> 1) + 1] ) >> 1;
+
+        cpi->activity_avg = median;
+
+        vpx_free(sortlist);
+    }
+#else
+    /* Simple mean for now */
+    cpi->activity_avg = (unsigned int)(activity_sum/cpi->common.MBs);
+#endif
+
+    if (cpi->activity_avg < VP8_ACTIVITY_AVG_MIN)
+        cpi->activity_avg = VP8_ACTIVITY_AVG_MIN;
+
+    /* Experimental code: return fixed value normalized for several clips */
+    if  ( ALT_ACT_MEASURE )
+        cpi->activity_avg = 100000;
+}
+
+#define USE_ACT_INDEX   0
+#define OUTPUT_NORM_ACT_STATS   0
+
+#if USE_ACT_INDEX
+/* Calculate and activity index for each mb */
+static void calc_activity_index( VP8_COMP *cpi, MACROBLOCK *x )
+{
+    VP8_COMMON *const cm = & cpi->common;
+    int mb_row, mb_col;
+
+    int64_t act;
+    int64_t a;
+    int64_t b;
+
+#if OUTPUT_NORM_ACT_STATS
+    FILE *f = fopen("norm_act.stt", "a");
+    fprintf(f, "\n%12d\n", cpi->activity_avg );
+#endif
+
+    /* Reset pointers to start of activity map */
+    x->mb_activity_ptr = cpi->mb_activity_map;
+
+    /* Calculate normalized mb activity number. */
+    for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
+    {
+        /* for each macroblock col in image */
+        for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
+        {
+            /* Read activity from the map */
+            act = *(x->mb_activity_ptr);
+
+            /* Calculate a normalized activity number */
+            a = act + 4*cpi->activity_avg;
+            b = 4*act + cpi->activity_avg;
+
+            if ( b >= a )
+                *(x->activity_ptr) = (int)((b + (a>>1))/a) - 1;
+            else
+                *(x->activity_ptr) = 1 - (int)((a + (b>>1))/b);
+
+#if OUTPUT_NORM_ACT_STATS
+            fprintf(f, " %6d", *(x->mb_activity_ptr));
+#endif
+            /* Increment activity map pointers */
+            x->mb_activity_ptr++;
+        }
+
+#if OUTPUT_NORM_ACT_STATS
+        fprintf(f, "\n");
+#endif
+
+    }
+
+#if OUTPUT_NORM_ACT_STATS
+    fclose(f);
+#endif
+
+}
+#endif
+
+/* Loop through all MBs. Note activity of each, average activity and
+ * calculate a normalized activity for each
+ */
+static void build_activity_map( VP8_COMP *cpi )
+{
+    MACROBLOCK *const x = & cpi->mb;
+    MACROBLOCKD *xd = &x->e_mbd;
+    VP8_COMMON *const cm = & cpi->common;
+
+#if ALT_ACT_MEASURE
+    YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
+    int recon_yoffset;
+    int recon_y_stride = new_yv12->y_stride;
+#endif
+
+    int mb_row, mb_col;
+    unsigned int mb_activity;
+    int64_t activity_sum = 0;
+
+    /* for each macroblock row in image */
+    for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
+    {
+#if ALT_ACT_MEASURE
+        /* reset above block coeffs */
+        xd->up_available = (mb_row != 0);
+        recon_yoffset = (mb_row * recon_y_stride * 16);
+#endif
+        /* for each macroblock col in image */
+        for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
+        {
+#if ALT_ACT_MEASURE
+            xd->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
+            xd->left_available = (mb_col != 0);
+            recon_yoffset += 16;
+#endif
+            /* Copy current mb to a buffer */
+            vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
+
+            /* measure activity */
+            mb_activity = mb_activity_measure( cpi, x, mb_row, mb_col );
+
+            /* Keep frame sum */
+            activity_sum += mb_activity;
+
+            /* Store MB level activity details. */
+            *x->mb_activity_ptr = mb_activity;
+
+            /* Increment activity map pointer */
+            x->mb_activity_ptr++;
+
+            /* adjust to the next column of source macroblocks */
+            x->src.y_buffer += 16;
+        }
+
+
+        /* adjust to the next row of mbs */
+        x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
+
+#if ALT_ACT_MEASURE
+        /* extend the recon for intra prediction */
+        vp8_extend_mb_row(new_yv12, xd->dst.y_buffer + 16,
+                          xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
+#endif
+
+    }
+
+    /* Calculate an "average" MB activity */
+    calc_av_activity(cpi, activity_sum);
+
+#if USE_ACT_INDEX
+    /* Calculate an activity index number of each mb */
+    calc_activity_index( cpi, x );
+#endif
+
+}
+
+/* Macroblock activity masking */
+void vp8_activity_masking(VP8_COMP *cpi, MACROBLOCK *x)
+{
+#if USE_ACT_INDEX
+    x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
+    x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
+    x->errorperbit += (x->errorperbit==0);
+#else
+    int64_t a;
+    int64_t b;
+    int64_t act = *(x->mb_activity_ptr);
+
+    /* Apply the masking to the RD multiplier. */
+    a = act + (2*cpi->activity_avg);
+    b = (2*act) + cpi->activity_avg;
+
+    x->rdmult = (unsigned int)(((int64_t)x->rdmult*b + (a>>1))/a);
+    x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
+    x->errorperbit += (x->errorperbit==0);
+#endif
+
+    /* Activity based Zbin adjustment */
+    adjust_act_zbin(cpi, x);
+}
+
+static
+void encode_mb_row(VP8_COMP *cpi,
+                   VP8_COMMON *cm,
+                   int mb_row,
+                   MACROBLOCK  *x,
+                   MACROBLOCKD *xd,
+                   TOKENEXTRA **tp,
+                   int *segment_counts,
+                   int *totalrate)
+{
+    int recon_yoffset, recon_uvoffset;
+    int mb_col;
+    int ref_fb_idx = cm->lst_fb_idx;
+    int dst_fb_idx = cm->new_fb_idx;
+    int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
+    int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
+    int map_index = (mb_row * cpi->common.mb_cols);
+
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+    const int num_part = (1 << cm->multi_token_partition);
+    TOKENEXTRA * tp_start = cpi->tok;
+    vp8_writer *w;
+#endif
+
+#if CONFIG_MULTITHREAD
+    const int nsync = cpi->mt_sync_range;
+    const int rightmost_col = cm->mb_cols + nsync;
+    const int *last_row_current_mb_col;
+    int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
+
+    if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
+        last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
+    else
+        last_row_current_mb_col = &rightmost_col;
+#endif
+
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+    if(num_part > 1)
+        w= &cpi->bc[1 + (mb_row % num_part)];
+    else
+        w = &cpi->bc[1];
+#endif
+
+    /* reset above block coeffs */
+    xd->above_context = cm->above_context;
+
+    xd->up_available = (mb_row != 0);
+    recon_yoffset = (mb_row * recon_y_stride * 16);
+    recon_uvoffset = (mb_row * recon_uv_stride * 8);
+
+    cpi->tplist[mb_row].start = *tp;
+    /* printf("Main mb_row = %d\n", mb_row); */
+
+    /* Distance of Mb to the top & bottom edges, specified in 1/8th pel
+     * units as they are always compared to values that are in 1/8th pel
+     */
+    xd->mb_to_top_edge = -((mb_row * 16) << 3);
+    xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
+
+    /* Set up limit values for vertical motion vector components
+     * to prevent them extending beyond the UMV borders
+     */
+    x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
+    x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
+                        + (VP8BORDERINPIXELS - 16);
+
+    /* Set the mb activity pointer to the start of the row. */
+    x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
+
+    /* for each macroblock col in image */
+    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
+    {
+
+#if  (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+        *tp = cpi->tok;
+#endif
+        /* Distance of Mb to the left & right edges, specified in
+         * 1/8th pel units as they are always compared to values
+         * that are in 1/8th pel units
+         */
+        xd->mb_to_left_edge = -((mb_col * 16) << 3);
+        xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
+
+        /* Set up limit values for horizontal motion vector components
+         * to prevent them extending beyond the UMV borders
+         */
+        x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
+        x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16)
+                            + (VP8BORDERINPIXELS - 16);
+
+        xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
+        xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
+        xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
+        xd->left_available = (mb_col != 0);
+
+        x->rddiv = cpi->RDDIV;
+        x->rdmult = cpi->RDMULT;
+
+        /* Copy current mb to a buffer */
+        vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
+
+#if CONFIG_MULTITHREAD
+        if (cpi->b_multi_threaded != 0) {
+            if (((mb_col - 1) % nsync) == 0) {
+                pthread_mutex_t *mutex = &cpi->pmutex[mb_row];
+                protected_write(mutex, current_mb_col, mb_col - 1);
+            }
+
+            if (mb_row && !(mb_col & (nsync - 1))) {
+                pthread_mutex_t *mutex = &cpi->pmutex[mb_row-1];
+                sync_read(mutex, mb_col, last_row_current_mb_col, nsync);
+            }
+        }
+#endif
+
+        if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
+            vp8_activity_masking(cpi, x);
+
+        /* Is segmentation enabled */
+        /* MB level adjustment to quantizer */
+        if (xd->segmentation_enabled)
+        {
+            /* Code to set segment id in xd->mbmi.segment_id for current MB
+             * (with range checking)
+             */
+            if (cpi->segmentation_map[map_index+mb_col] <= 3)
+                xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index+mb_col];
+            else
+                xd->mode_info_context->mbmi.segment_id = 0;
+
+            vp8cx_mb_init_quantizer(cpi, x, 1);
+        }
+        else
+            /* Set to Segment 0 by default */
+            xd->mode_info_context->mbmi.segment_id = 0;
+
+        x->active_ptr = cpi->active_map + map_index + mb_col;
+
+        if (cm->frame_type == KEY_FRAME)
+        {
+            *totalrate += vp8cx_encode_intra_macroblock(cpi, x, tp);
+#ifdef MODE_STATS
+            y_modes[xd->mbmi.mode] ++;
+#endif
+        }
+        else
+        {
+            *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
+
+#ifdef MODE_STATS
+            inter_y_modes[xd->mbmi.mode] ++;
+
+            if (xd->mbmi.mode == SPLITMV)
+            {
+                int b;
+
+                for (b = 0; b < xd->mbmi.partition_count; b++)
+                {
+                    inter_b_modes[x->partition->bmi[b].mode] ++;
+                }
+            }
+
+#endif
+
+            // Keep track of how many (consecutive) times a  block is coded
+            // as ZEROMV_LASTREF, for base layer frames.
+            // Reset to 0 if its coded as anything else.
+            if (cpi->current_layer == 0) {
+              if (xd->mode_info_context->mbmi.mode == ZEROMV &&
+                  xd->mode_info_context->mbmi.ref_frame == LAST_FRAME) {
+                // Increment, check for wrap-around.
+                if (cpi->consec_zero_last[map_index+mb_col] < 255)
+                  cpi->consec_zero_last[map_index+mb_col] += 1;
+                if (cpi->consec_zero_last_mvbias[map_index+mb_col] < 255)
+                  cpi->consec_zero_last_mvbias[map_index+mb_col] += 1;
+              } else {
+                cpi->consec_zero_last[map_index+mb_col] = 0;
+                cpi->consec_zero_last_mvbias[map_index+mb_col] = 0;
+              }
+              if (x->zero_last_dot_suppress)
+                cpi->consec_zero_last_mvbias[map_index+mb_col] = 0;
+            }
+
+            /* Special case code for cyclic refresh
+             * If cyclic update enabled then copy xd->mbmi.segment_id; (which
+             * may have been updated based on mode during
+             * vp8cx_encode_inter_macroblock()) back into the global
+             * segmentation map
+             */
+            if ((cpi->current_layer == 0) &&
+                (cpi->cyclic_refresh_mode_enabled &&
+                 xd->segmentation_enabled))
+            {
+                cpi->segmentation_map[map_index+mb_col] = xd->mode_info_context->mbmi.segment_id;
+
+                /* If the block has been refreshed mark it as clean (the
+                 * magnitude of the -ve influences how long it will be before
+                 * we consider another refresh):
+                 * Else if it was coded (last frame 0,0) and has not already
+                 * been refreshed then mark it as a candidate for cleanup
+                 * next time (marked 0) else mark it as dirty (1).
+                 */
+                if (xd->mode_info_context->mbmi.segment_id)
+                    cpi->cyclic_refresh_map[map_index+mb_col] = -1;
+                else if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
+                {
+                    if (cpi->cyclic_refresh_map[map_index+mb_col] == 1)
+                        cpi->cyclic_refresh_map[map_index+mb_col] = 0;
+                }
+                else
+                    cpi->cyclic_refresh_map[map_index+mb_col] = 1;
+
+            }
+        }
+
+        cpi->tplist[mb_row].stop = *tp;
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+        /* pack tokens for this MB */
+        {
+            int tok_count = *tp - tp_start;
+            vp8_pack_tokens(w, tp_start, tok_count);
+        }
+#endif
+        /* Increment pointer into gf usage flags structure. */
+        x->gf_active_ptr++;
+
+        /* Increment the activity mask pointers. */
+        x->mb_activity_ptr++;
+
+        /* adjust to the next column of macroblocks */
+        x->src.y_buffer += 16;
+        x->src.u_buffer += 8;
+        x->src.v_buffer += 8;
+
+        recon_yoffset += 16;
+        recon_uvoffset += 8;
+
+        /* Keep track of segment usage */
+        segment_counts[xd->mode_info_context->mbmi.segment_id] ++;
+
+        /* skip to next mb */
+        xd->mode_info_context++;
+        x->partition_info++;
+        xd->above_context++;
+    }
+
+    /* extend the recon for intra prediction */
+    vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
+                        xd->dst.y_buffer + 16,
+                        xd->dst.u_buffer + 8,
+                        xd->dst.v_buffer + 8);
+
+#if CONFIG_MULTITHREAD
+    if (cpi->b_multi_threaded != 0)
+        protected_write(&cpi->pmutex[mb_row], current_mb_col, rightmost_col);
+#endif
+
+    /* this is to account for the border */
+    xd->mode_info_context++;
+    x->partition_info++;
+}
+
+static void init_encode_frame_mb_context(VP8_COMP *cpi)
+{
+    MACROBLOCK *const x = & cpi->mb;
+    VP8_COMMON *const cm = & cpi->common;
+    MACROBLOCKD *const xd = & x->e_mbd;
+
+    /* GF active flags data structure */
+    x->gf_active_ptr = (signed char *)cpi->gf_active_flags;
+
+    /* Activity map pointer */
+    x->mb_activity_ptr = cpi->mb_activity_map;
+
+    x->act_zbin_adj = 0;
+
+    x->partition_info = x->pi;
+
+    xd->mode_info_context = cm->mi;
+    xd->mode_info_stride = cm->mode_info_stride;
+
+    xd->frame_type = cm->frame_type;
+
+    /* reset intra mode contexts */
+    if (cm->frame_type == KEY_FRAME)
+        vp8_init_mbmode_probs(cm);
+
+    /* Copy data over into macro block data structures. */
+    x->src = * cpi->Source;
+    xd->pre = cm->yv12_fb[cm->lst_fb_idx];
+    xd->dst = cm->yv12_fb[cm->new_fb_idx];
+
+    /* set up frame for intra coded blocks */
+    vp8_setup_intra_recon(&cm->yv12_fb[cm->new_fb_idx]);
+
+    vp8_build_block_offsets(x);
+
+    xd->mode_info_context->mbmi.mode = DC_PRED;
+    xd->mode_info_context->mbmi.uv_mode = DC_PRED;
+
+    xd->left_context = &cm->left_context;
+
+    x->mvc = cm->fc.mvc;
+
+    memset(cm->above_context, 0, sizeof(ENTROPY_CONTEXT_PLANES) * cm->mb_cols);
+
+    /* Special case treatment when GF and ARF are not sensible options
+     * for reference
+     */
+    if (cpi->ref_frame_flags == VP8_LAST_FRAME)
+        vp8_calc_ref_frame_costs(x->ref_frame_cost,
+                                 cpi->prob_intra_coded,255,128);
+    else if ((cpi->oxcf.number_of_layers > 1) &&
+               (cpi->ref_frame_flags == VP8_GOLD_FRAME))
+        vp8_calc_ref_frame_costs(x->ref_frame_cost,
+                                 cpi->prob_intra_coded,1,255);
+    else if ((cpi->oxcf.number_of_layers > 1) &&
+                (cpi->ref_frame_flags == VP8_ALTR_FRAME))
+        vp8_calc_ref_frame_costs(x->ref_frame_cost,
+                                 cpi->prob_intra_coded,1,1);
+    else
+        vp8_calc_ref_frame_costs(x->ref_frame_cost,
+                                 cpi->prob_intra_coded,
+                                 cpi->prob_last_coded,
+                                 cpi->prob_gf_coded);
+
+    xd->fullpixel_mask = 0xffffffff;
+    if(cm->full_pixel)
+        xd->fullpixel_mask = 0xfffffff8;
+
+    vp8_zero(x->coef_counts);
+    vp8_zero(x->ymode_count);
+    vp8_zero(x->uv_mode_count)
+    x->prediction_error = 0;
+    x->intra_error = 0;
+    vp8_zero(x->count_mb_ref_frame_usage);
+}
+
+#if CONFIG_MULTITHREAD
+static void sum_coef_counts(MACROBLOCK *x, MACROBLOCK *x_thread)
+{
+    int i = 0;
+    do
+    {
+        int j = 0;
+        do
+        {
+            int k = 0;
+            do
+            {
+                /* at every context */
+
+                /* calc probs and branch cts for this frame only */
+                int t = 0;      /* token/prob index */
+
+                do
+                {
+                    x->coef_counts [i][j][k][t] +=
+                        x_thread->coef_counts [i][j][k][t];
+                }
+                while (++t < ENTROPY_NODES);
+            }
+            while (++k < PREV_COEF_CONTEXTS);
+        }
+        while (++j < COEF_BANDS);
+    }
+    while (++i < BLOCK_TYPES);
+}
+#endif  // CONFIG_MULTITHREAD
+
+void vp8_encode_frame(VP8_COMP *cpi)
+{
+    int mb_row;
+    MACROBLOCK *const x = & cpi->mb;
+    VP8_COMMON *const cm = & cpi->common;
+    MACROBLOCKD *const xd = & x->e_mbd;
+    TOKENEXTRA *tp = cpi->tok;
+    int segment_counts[MAX_MB_SEGMENTS];
+    int totalrate;
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+    BOOL_CODER * bc = &cpi->bc[1]; /* bc[0] is for control partition */
+    const int num_part = (1 << cm->multi_token_partition);
+#endif
+
+    memset(segment_counts, 0, sizeof(segment_counts));
+    totalrate = 0;
+
+    if (cpi->compressor_speed == 2)
+    {
+        if (cpi->oxcf.cpu_used < 0)
+            cpi->Speed = -(cpi->oxcf.cpu_used);
+        else
+            vp8_auto_select_speed(cpi);
+    }
+
+    /* Functions setup for all frame types so we can use MC in AltRef */
+    if(!cm->use_bilinear_mc_filter)
+    {
+        xd->subpixel_predict        = vp8_sixtap_predict4x4;
+        xd->subpixel_predict8x4     = vp8_sixtap_predict8x4;
+        xd->subpixel_predict8x8     = vp8_sixtap_predict8x8;
+        xd->subpixel_predict16x16   = vp8_sixtap_predict16x16;
+    }
+    else
+    {
+        xd->subpixel_predict        = vp8_bilinear_predict4x4;
+        xd->subpixel_predict8x4     = vp8_bilinear_predict8x4;
+        xd->subpixel_predict8x8     = vp8_bilinear_predict8x8;
+        xd->subpixel_predict16x16   = vp8_bilinear_predict16x16;
+    }
+
+    cpi->mb.skip_true_count = 0;
+    cpi->tok_count = 0;
+
+#if 0
+    /* Experimental code */
+    cpi->frame_distortion = 0;
+    cpi->last_mb_distortion = 0;
+#endif
+
+    xd->mode_info_context = cm->mi;
+
+    vp8_zero(cpi->mb.MVcount);
+
+    vp8cx_frame_init_quantizer(cpi);
+
+    vp8_initialize_rd_consts(cpi, x,
+                             vp8_dc_quant(cm->base_qindex, cm->y1dc_delta_q));
+
+    vp8cx_initialize_me_consts(cpi, cm->base_qindex);
+
+    if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
+    {
+        /* Initialize encode frame context. */
+        init_encode_frame_mb_context(cpi);
+
+        /* Build a frame level activity map */
+        build_activity_map(cpi);
+    }
+
+    /* re-init encode frame context. */
+    init_encode_frame_mb_context(cpi);
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+    {
+        int i;
+        for(i = 0; i < num_part; i++)
+        {
+            vp8_start_encode(&bc[i], cpi->partition_d[i + 1],
+                    cpi->partition_d_end[i + 1]);
+            bc[i].error = &cm->error;
+        }
+    }
+
+#endif
+
+    {
+        struct vpx_usec_timer  emr_timer;
+        vpx_usec_timer_start(&emr_timer);
+
+#if CONFIG_MULTITHREAD
+        if (cpi->b_multi_threaded)
+        {
+            int i;
+
+            vp8cx_init_mbrthread_data(cpi, x, cpi->mb_row_ei,
+                                      cpi->encoding_thread_count);
+
+            for (i = 0; i < cm->mb_rows; i++)
+                cpi->mt_current_mb_col[i] = -1;
+
+            for (i = 0; i < cpi->encoding_thread_count; i++)
+            {
+                sem_post(&cpi->h_event_start_encoding[i]);
+            }
+
+            for (mb_row = 0; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
+            {
+                vp8_zero(cm->left_context)
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+                tp = cpi->tok;
+#else
+                tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
+#endif
+
+                encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
+
+                /* adjust to the next row of mbs */
+                x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
+                x->src.u_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
+                x->src.v_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
+
+                xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
+                x->partition_info  += xd->mode_info_stride * cpi->encoding_thread_count;
+                x->gf_active_ptr   += cm->mb_cols * cpi->encoding_thread_count;
+
+                if(mb_row == cm->mb_rows - 1)
+                {
+                    sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
+                }
+            }
+
+            sem_wait(&cpi->h_event_end_encoding); /* wait for other threads to finish */
+
+            for (mb_row = 0; mb_row < cm->mb_rows; mb_row ++)
+            {
+                cpi->tok_count += (unsigned int)
+                  (cpi->tplist[mb_row].stop - cpi->tplist[mb_row].start);
+            }
+
+            if (xd->segmentation_enabled)
+            {
+                int j;
+
+                if (xd->segmentation_enabled)
+                {
+                    for (i = 0; i < cpi->encoding_thread_count; i++)
+                    {
+                        for (j = 0; j < 4; j++)
+                            segment_counts[j] += cpi->mb_row_ei[i].segment_counts[j];
+                    }
+                }
+            }
+
+            for (i = 0; i < cpi->encoding_thread_count; i++)
+            {
+                int mode_count;
+                int c_idx;
+                totalrate += cpi->mb_row_ei[i].totalrate;
+
+                cpi->mb.skip_true_count += cpi->mb_row_ei[i].mb.skip_true_count;
+
+                for(mode_count = 0; mode_count < VP8_YMODES; mode_count++)
+                    cpi->mb.ymode_count[mode_count] +=
+                        cpi->mb_row_ei[i].mb.ymode_count[mode_count];
+
+                for(mode_count = 0; mode_count < VP8_UV_MODES; mode_count++)
+                    cpi->mb.uv_mode_count[mode_count] +=
+                        cpi->mb_row_ei[i].mb.uv_mode_count[mode_count];
+
+                for(c_idx = 0; c_idx < MVvals; c_idx++)
+                {
+                    cpi->mb.MVcount[0][c_idx] +=
+                        cpi->mb_row_ei[i].mb.MVcount[0][c_idx];
+                    cpi->mb.MVcount[1][c_idx] +=
+                        cpi->mb_row_ei[i].mb.MVcount[1][c_idx];
+                }
+
+                cpi->mb.prediction_error +=
+                    cpi->mb_row_ei[i].mb.prediction_error;
+                cpi->mb.intra_error += cpi->mb_row_ei[i].mb.intra_error;
+
+                for(c_idx = 0; c_idx < MAX_REF_FRAMES; c_idx++)
+                    cpi->mb.count_mb_ref_frame_usage[c_idx] +=
+                        cpi->mb_row_ei[i].mb.count_mb_ref_frame_usage[c_idx];
+
+                for(c_idx = 0; c_idx < MAX_ERROR_BINS; c_idx++)
+                    cpi->mb.error_bins[c_idx] +=
+                        cpi->mb_row_ei[i].mb.error_bins[c_idx];
+
+                /* add up counts for each thread */
+                sum_coef_counts(x, &cpi->mb_row_ei[i].mb);
+            }
+
+        }
+        else
+#endif  // CONFIG_MULTITHREAD
+        {
+
+            /* for each macroblock row in image */
+            for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
+            {
+                vp8_zero(cm->left_context)
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+                tp = cpi->tok;
+#endif
+
+                encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
+
+                /* adjust to the next row of mbs */
+                x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
+                x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
+                x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
+            }
+
+            cpi->tok_count = (unsigned int)(tp - cpi->tok);
+        }
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+        {
+            int i;
+            for(i = 0; i < num_part; i++)
+            {
+                vp8_stop_encode(&bc[i]);
+                cpi->partition_sz[i+1] = bc[i].pos;
+            }
+        }
+#endif
+
+        vpx_usec_timer_mark(&emr_timer);
+        cpi->time_encode_mb_row += vpx_usec_timer_elapsed(&emr_timer);
+    }
+
+
+    // Work out the segment probabilities if segmentation is enabled
+    // and needs to be updated
+    if (xd->segmentation_enabled && xd->update_mb_segmentation_map)
+    {
+        int tot_count;
+        int i;
+
+        /* Set to defaults */
+        memset(xd->mb_segment_tree_probs, 255 , sizeof(xd->mb_segment_tree_probs));
+
+        tot_count = segment_counts[0] + segment_counts[1] + segment_counts[2] + segment_counts[3];
+
+        if (tot_count)
+        {
+            xd->mb_segment_tree_probs[0] = ((segment_counts[0] + segment_counts[1]) * 255) / tot_count;
+
+            tot_count = segment_counts[0] + segment_counts[1];
+
+            if (tot_count > 0)
+            {
+                xd->mb_segment_tree_probs[1] = (segment_counts[0] * 255) / tot_count;
+            }
+
+            tot_count = segment_counts[2] + segment_counts[3];
+
+            if (tot_count > 0)
+                xd->mb_segment_tree_probs[2] = (segment_counts[2] * 255) / tot_count;
+
+            /* Zero probabilities not allowed */
+            for (i = 0; i < MB_FEATURE_TREE_PROBS; i ++)
+            {
+                if (xd->mb_segment_tree_probs[i] == 0)
+                    xd->mb_segment_tree_probs[i] = 1;
+            }
+        }
+    }
+
+    /* projected_frame_size in units of BYTES */
+    cpi->projected_frame_size = totalrate >> 8;
+
+    /* Make a note of the percentage MBs coded Intra. */
+    if (cm->frame_type == KEY_FRAME)
+    {
+        cpi->this_frame_percent_intra = 100;
+    }
+    else
+    {
+        int tot_modes;
+
+        tot_modes = cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME]
+                    + cpi->mb.count_mb_ref_frame_usage[LAST_FRAME]
+                    + cpi->mb.count_mb_ref_frame_usage[GOLDEN_FRAME]
+                    + cpi->mb.count_mb_ref_frame_usage[ALTREF_FRAME];
+
+        if (tot_modes)
+            cpi->this_frame_percent_intra =
+                cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME] * 100 / tot_modes;
+
+    }
+
+#if ! CONFIG_REALTIME_ONLY
+    /* Adjust the projected reference frame usage probability numbers to
+     * reflect what we have just seen. This may be useful when we make
+     * multiple iterations of the recode loop rather than continuing to use
+     * values from the previous frame.
+     */
+    if ((cm->frame_type != KEY_FRAME) && ((cpi->oxcf.number_of_layers > 1) ||
+        (!cm->refresh_alt_ref_frame && !cm->refresh_golden_frame)))
+    {
+      vp8_convert_rfct_to_prob(cpi);
+    }
+#endif
+}
+void vp8_setup_block_ptrs(MACROBLOCK *x)
+{
+    int r, c;
+    int i;
+
+    for (r = 0; r < 4; r++)
+    {
+        for (c = 0; c < 4; c++)
+        {
+            x->block[r*4+c].src_diff = x->src_diff + r * 4 * 16 + c * 4;
+        }
+    }
+
+    for (r = 0; r < 2; r++)
+    {
+        for (c = 0; c < 2; c++)
+        {
+            x->block[16 + r*2+c].src_diff = x->src_diff + 256 + r * 4 * 8 + c * 4;
+        }
+    }
+
+
+    for (r = 0; r < 2; r++)
+    {
+        for (c = 0; c < 2; c++)
+        {
+            x->block[20 + r*2+c].src_diff = x->src_diff + 320 + r * 4 * 8 + c * 4;
+        }
+    }
+
+    x->block[24].src_diff = x->src_diff + 384;
+
+
+    for (i = 0; i < 25; i++)
+    {
+        x->block[i].coeff = x->coeff + i * 16;
+    }
+}
+
+void vp8_build_block_offsets(MACROBLOCK *x)
+{
+    int block = 0;
+    int br, bc;
+
+    vp8_build_block_doffsets(&x->e_mbd);
+
+    /* y blocks */
+    x->thismb_ptr = &x->thismb[0];
+    for (br = 0; br < 4; br++)
+    {
+        for (bc = 0; bc < 4; bc++)
+        {
+            BLOCK *this_block = &x->block[block];
+            this_block->base_src = &x->thismb_ptr;
+            this_block->src_stride = 16;
+            this_block->src = 4 * br * 16 + 4 * bc;
+            ++block;
+        }
+    }
+
+    /* u blocks */
+    for (br = 0; br < 2; br++)
+    {
+        for (bc = 0; bc < 2; bc++)
+        {
+            BLOCK *this_block = &x->block[block];
+            this_block->base_src = &x->src.u_buffer;
+            this_block->src_stride = x->src.uv_stride;
+            this_block->src = 4 * br * this_block->src_stride + 4 * bc;
+            ++block;
+        }
+    }
+
+    /* v blocks */
+    for (br = 0; br < 2; br++)
+    {
+        for (bc = 0; bc < 2; bc++)
+        {
+            BLOCK *this_block = &x->block[block];
+            this_block->base_src = &x->src.v_buffer;
+            this_block->src_stride = x->src.uv_stride;
+            this_block->src = 4 * br * this_block->src_stride + 4 * bc;
+            ++block;
+        }
+    }
+}
+
+static void sum_intra_stats(VP8_COMP *cpi, MACROBLOCK *x)
+{
+    const MACROBLOCKD *xd = & x->e_mbd;
+    const MB_PREDICTION_MODE m = xd->mode_info_context->mbmi.mode;
+    const MB_PREDICTION_MODE uvm = xd->mode_info_context->mbmi.uv_mode;
+
+#ifdef MODE_STATS
+    const int is_key = cpi->common.frame_type == KEY_FRAME;
+
+    ++ (is_key ? uv_modes : inter_uv_modes)[uvm];
+
+    if (m == B_PRED)
+    {
+        unsigned int *const bct = is_key ? b_modes : inter_b_modes;
+
+        int b = 0;
+
+        do
+        {
+            ++ bct[xd->block[b].bmi.mode];
+        }
+        while (++b < 16);
+    }
+
+#else
+    (void)cpi;
+#endif
+
+    ++x->ymode_count[m];
+    ++x->uv_mode_count[uvm];
+
+}
+
+/* Experimental stub function to create a per MB zbin adjustment based on
+ * some previously calculated measure of MB activity.
+ */
+static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x )
+{
+#if USE_ACT_INDEX
+    x->act_zbin_adj = *(x->mb_activity_ptr);
+#else
+    int64_t a;
+    int64_t b;
+    int64_t act = *(x->mb_activity_ptr);
+
+    /* Apply the masking to the RD multiplier. */
+    a = act + 4*cpi->activity_avg;
+    b = 4*act + cpi->activity_avg;
+
+    if ( act > cpi->activity_avg )
+        x->act_zbin_adj = (int)(((int64_t)b + (a>>1))/a) - 1;
+    else
+        x->act_zbin_adj = 1 - (int)(((int64_t)a + (b>>1))/b);
+#endif
+}
+
+int vp8cx_encode_intra_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
+                                  TOKENEXTRA **t)
+{
+    MACROBLOCKD *xd = &x->e_mbd;
+    int rate;
+
+    if (cpi->sf.RD && cpi->compressor_speed != 2)
+        vp8_rd_pick_intra_mode(x, &rate);
+    else
+        vp8_pick_intra_mode(x, &rate);
+
+    if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
+    {
+        adjust_act_zbin( cpi, x );
+        vp8_update_zbin_extra(cpi, x);
+    }
+
+    if (x->e_mbd.mode_info_context->mbmi.mode == B_PRED)
+        vp8_encode_intra4x4mby(x);
+    else
+        vp8_encode_intra16x16mby(x);
+
+    vp8_encode_intra16x16mbuv(x);
+
+    sum_intra_stats(cpi, x);
+
+    vp8_tokenize_mb(cpi, x, t);
+
+    if (xd->mode_info_context->mbmi.mode != B_PRED)
+        vp8_inverse_transform_mby(xd);
+
+    vp8_dequant_idct_add_uv_block
+                    (xd->qcoeff+16*16, xd->dequant_uv,
+                     xd->dst.u_buffer, xd->dst.v_buffer,
+                     xd->dst.uv_stride, xd->eobs+16);
+    return rate;
+}
+#ifdef SPEEDSTATS
+extern int cnt_pm;
+#endif
+
+extern void vp8_fix_contexts(MACROBLOCKD *x);
+
+int vp8cx_encode_inter_macroblock
+(
+    VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t,
+    int recon_yoffset, int recon_uvoffset,
+    int mb_row, int mb_col
+)
+{
+    MACROBLOCKD *const xd = &x->e_mbd;
+    int intra_error = 0;
+    int rate;
+    int distortion;
+
+    x->skip = 0;
+
+    if (xd->segmentation_enabled)
+        x->encode_breakout = cpi->segment_encode_breakout[xd->mode_info_context->mbmi.segment_id];
+    else
+        x->encode_breakout = cpi->oxcf.encode_breakout;
+
+#if CONFIG_TEMPORAL_DENOISING
+    /* Reset the best sse mode/mv for each macroblock. */
+    x->best_reference_frame = INTRA_FRAME;
+    x->best_zeromv_reference_frame = INTRA_FRAME;
+    x->best_sse_inter_mode = 0;
+    x->best_sse_mv.as_int = 0;
+    x->need_to_clamp_best_mvs = 0;
+#endif
+
+    if (cpi->sf.RD)
+    {
+        int zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
+
+        /* Are we using the fast quantizer for the mode selection? */
+        if(cpi->sf.use_fastquant_for_pick)
+        {
+            x->quantize_b      = vp8_fast_quantize_b;
+
+            /* the fast quantizer does not use zbin_extra, so
+             * do not recalculate */
+            x->zbin_mode_boost_enabled = 0;
+        }
+        vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
+                               &distortion, &intra_error, mb_row, mb_col);
+
+        /* switch back to the regular quantizer for the encode */
+        if (cpi->sf.improved_quant)
+        {
+            x->quantize_b      = vp8_regular_quantize_b;
+        }
+
+        /* restore cpi->zbin_mode_boost_enabled */
+        x->zbin_mode_boost_enabled = zbin_mode_boost_enabled;
+
+    }
+    else
+    {
+        vp8_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
+                            &distortion, &intra_error, mb_row, mb_col);
+    }
+
+    x->prediction_error += distortion;
+    x->intra_error += intra_error;
+
+    if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
+    {
+        /* Adjust the zbin based on this MB rate. */
+        adjust_act_zbin( cpi, x );
+    }
+
+#if 0
+    /* Experimental RD code */
+    cpi->frame_distortion += distortion;
+    cpi->last_mb_distortion = distortion;
+#endif
+
+    /* MB level adjutment to quantizer setup */
+    if (xd->segmentation_enabled)
+    {
+        /* If cyclic update enabled */
+        if (cpi->current_layer == 0 && cpi->cyclic_refresh_mode_enabled)
+        {
+            /* Clear segment_id back to 0 if not coded (last frame 0,0) */
+            if ((xd->mode_info_context->mbmi.segment_id == 1) &&
+                ((xd->mode_info_context->mbmi.ref_frame != LAST_FRAME) || (xd->mode_info_context->mbmi.mode != ZEROMV)))
+            {
+                xd->mode_info_context->mbmi.segment_id = 0;
+
+                /* segment_id changed, so update */
+                vp8cx_mb_init_quantizer(cpi, x, 1);
+            }
+        }
+    }
+
+    {
+        /* Experimental code.
+         * Special case for gf and arf zeromv modes, for 1 temporal layer.
+         * Increase zbin size to supress noise.
+         */
+        x->zbin_mode_boost = 0;
+        if (x->zbin_mode_boost_enabled)
+        {
+            if ( xd->mode_info_context->mbmi.ref_frame != INTRA_FRAME )
+            {
+                if (xd->mode_info_context->mbmi.mode == ZEROMV)
+                {
+                    if (xd->mode_info_context->mbmi.ref_frame != LAST_FRAME &&
+                        cpi->oxcf.number_of_layers == 1)
+                        x->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
+                    else
+                        x->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST;
+                }
+                else if (xd->mode_info_context->mbmi.mode == SPLITMV)
+                    x->zbin_mode_boost = 0;
+                else
+                    x->zbin_mode_boost = MV_ZBIN_BOOST;
+            }
+        }
+
+        /* The fast quantizer doesn't use zbin_extra, only do so with
+         * the regular quantizer. */
+        if (cpi->sf.improved_quant)
+            vp8_update_zbin_extra(cpi, x);
+    }
+
+    x->count_mb_ref_frame_usage[xd->mode_info_context->mbmi.ref_frame] ++;
+
+    if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
+    {
+        vp8_encode_intra16x16mbuv(x);
+
+        if (xd->mode_info_context->mbmi.mode == B_PRED)
+        {
+            vp8_encode_intra4x4mby(x);
+        }
+        else
+        {
+            vp8_encode_intra16x16mby(x);
+        }
+
+        sum_intra_stats(cpi, x);
+    }
+    else
+    {
+        int ref_fb_idx;
+
+        if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
+            ref_fb_idx = cpi->common.lst_fb_idx;
+        else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
+            ref_fb_idx = cpi->common.gld_fb_idx;
+        else
+            ref_fb_idx = cpi->common.alt_fb_idx;
+
+        xd->pre.y_buffer = cpi->common.yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
+        xd->pre.u_buffer = cpi->common.yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
+        xd->pre.v_buffer = cpi->common.yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
+
+        if (!x->skip)
+        {
+            vp8_encode_inter16x16(x);
+        }
+        else
+            vp8_build_inter16x16_predictors_mb(xd, xd->dst.y_buffer,
+                                           xd->dst.u_buffer, xd->dst.v_buffer,
+                                           xd->dst.y_stride, xd->dst.uv_stride);
+
+    }
+
+    if (!x->skip)
+    {
+        vp8_tokenize_mb(cpi, x, t);
+
+        if (xd->mode_info_context->mbmi.mode != B_PRED)
+            vp8_inverse_transform_mby(xd);
+
+        vp8_dequant_idct_add_uv_block
+                        (xd->qcoeff+16*16, xd->dequant_uv,
+                         xd->dst.u_buffer, xd->dst.v_buffer,
+                         xd->dst.uv_stride, xd->eobs+16);
+    }
+    else
+    {
+        /* always set mb_skip_coeff as it is needed by the loopfilter */
+        xd->mode_info_context->mbmi.mb_skip_coeff = 1;
+
+        if (cpi->common.mb_no_coeff_skip)
+        {
+            x->skip_true_count ++;
+            vp8_fix_contexts(xd);
+        }
+        else
+        {
+            vp8_stuff_mb(cpi, x, t);
+        }
+    }
+
+    return rate;
+}
diff --git a/libvpx/libvpx/vp8/encoder/encodeframe.h b/libvpx/libvpx/vp8/encoder/encodeframe.h
new file mode 100644
index 0000000..e185c10
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/encodeframe.h
@@ -0,0 +1,35 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VP8_ENCODER_ENCODEFRAME_H_
+#define VP8_ENCODER_ENCODEFRAME_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+extern void vp8_activity_masking(VP8_COMP *cpi, MACROBLOCK *x);
+
+extern void vp8_build_block_offsets(MACROBLOCK *x);
+
+extern void vp8_setup_block_ptrs(MACROBLOCK *x);
+
+extern void vp8_encode_frame(VP8_COMP *cpi);
+
+extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
+        TOKENEXTRA **t,
+        int recon_yoffset, int recon_uvoffset,
+        int mb_row, int mb_col);
+
+extern int vp8cx_encode_intra_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
+        TOKENEXTRA **t);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_ENCODEFRAME_H_
diff --git a/libvpx/libvpx/vp8/encoder/encodeintra.c b/libvpx/libvpx/vp8/encoder/encodeintra.c
new file mode 100644
index 0000000..44be959
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/encodeintra.c
@@ -0,0 +1,140 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vp8/encoder/quantize.h"
+#include "vp8/common/reconintra.h"
+#include "vp8/common/reconintra4x4.h"
+#include "encodemb.h"
+#include "vp8/common/invtrans.h"
+#include "encodeintra.h"
+
+
+int vp8_encode_intra(VP8_COMP *cpi, MACROBLOCK *x, int use_dc_pred)
+{
+
+    int i;
+    int intra_pred_var = 0;
+    (void) cpi;
+
+    if (use_dc_pred)
+    {
+        x->e_mbd.mode_info_context->mbmi.mode = DC_PRED;
+        x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
+        x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME;
+
+        vp8_encode_intra16x16mby(x);
+
+        vp8_inverse_transform_mby(&x->e_mbd);
+    }
+    else
+    {
+        for (i = 0; i < 16; i++)
+        {
+            x->e_mbd.block[i].bmi.as_mode = B_DC_PRED;
+            vp8_encode_intra4x4block(x, i);
+        }
+    }
+
+    intra_pred_var = vpx_get_mb_ss(x->src_diff);
+
+    return intra_pred_var;
+}
+
+void vp8_encode_intra4x4block(MACROBLOCK *x, int ib)
+{
+    BLOCKD *b = &x->e_mbd.block[ib];
+    BLOCK *be = &x->block[ib];
+    int dst_stride = x->e_mbd.dst.y_stride;
+    unsigned char *dst = x->e_mbd.dst.y_buffer + b->offset;
+    unsigned char *Above = dst - dst_stride;
+    unsigned char *yleft = dst - 1;
+    unsigned char top_left = Above[-1];
+
+    vp8_intra4x4_predict(Above, yleft, dst_stride, b->bmi.as_mode,
+                         b->predictor, 16, top_left);
+
+    vp8_subtract_b(be, b, 16);
+
+    x->short_fdct4x4(be->src_diff, be->coeff, 32);
+
+    x->quantize_b(be, b);
+
+    if (*b->eob > 1)
+    {
+      vp8_short_idct4x4llm(b->dqcoeff, b->predictor, 16, dst, dst_stride);
+    }
+    else
+    {
+      vp8_dc_only_idct_add(b->dqcoeff[0], b->predictor, 16, dst, dst_stride);
+    }
+}
+
+void vp8_encode_intra4x4mby(MACROBLOCK *mb)
+{
+    int i;
+
+    MACROBLOCKD *xd = &mb->e_mbd;
+    intra_prediction_down_copy(xd, xd->dst.y_buffer - xd->dst.y_stride + 16);
+
+    for (i = 0; i < 16; i++)
+        vp8_encode_intra4x4block(mb, i);
+    return;
+}
+
+void vp8_encode_intra16x16mby(MACROBLOCK *x)
+{
+    BLOCK *b = &x->block[0];
+    MACROBLOCKD *xd = &x->e_mbd;
+
+    vp8_build_intra_predictors_mby_s(xd,
+                                         xd->dst.y_buffer - xd->dst.y_stride,
+                                         xd->dst.y_buffer - 1,
+                                         xd->dst.y_stride,
+                                         xd->dst.y_buffer,
+                                         xd->dst.y_stride);
+
+    vp8_subtract_mby(x->src_diff, *(b->base_src),
+        b->src_stride, xd->dst.y_buffer, xd->dst.y_stride);
+
+    vp8_transform_intra_mby(x);
+
+    vp8_quantize_mby(x);
+
+    if (x->optimize)
+        vp8_optimize_mby(x);
+}
+
+void vp8_encode_intra16x16mbuv(MACROBLOCK *x)
+{
+    MACROBLOCKD *xd = &x->e_mbd;
+
+    vp8_build_intra_predictors_mbuv_s(xd, xd->dst.u_buffer - xd->dst.uv_stride,
+                                      xd->dst.v_buffer - xd->dst.uv_stride,
+                                      xd->dst.u_buffer - 1,
+                                      xd->dst.v_buffer - 1,
+                                      xd->dst.uv_stride,
+                                      xd->dst.u_buffer, xd->dst.v_buffer,
+                                      xd->dst.uv_stride);
+
+    vp8_subtract_mbuv(x->src_diff, x->src.u_buffer,
+        x->src.v_buffer, x->src.uv_stride, xd->dst.u_buffer,
+        xd->dst.v_buffer, xd->dst.uv_stride);
+
+    vp8_transform_mbuv(x);
+
+    vp8_quantize_mbuv(x);
+
+    if (x->optimize)
+        vp8_optimize_mbuv(x);
+}
diff --git a/libvpx/libvpx/vp8/encoder/encodeintra.h b/libvpx/libvpx/vp8/encoder/encodeintra.h
new file mode 100644
index 0000000..a8d0284
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/encodeintra.h
@@ -0,0 +1,29 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_ENCODEINTRA_H_
+#define VP8_ENCODER_ENCODEINTRA_H_
+#include "onyx_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int vp8_encode_intra(VP8_COMP *cpi, MACROBLOCK *x, int use_dc_pred);
+void vp8_encode_intra16x16mby(MACROBLOCK *x);
+void vp8_encode_intra16x16mbuv(MACROBLOCK *x);
+void vp8_encode_intra4x4mby(MACROBLOCK *mb);
+void vp8_encode_intra4x4block(MACROBLOCK *x, int ib);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_ENCODEINTRA_H_
diff --git a/libvpx/libvpx/vp8/encoder/encodemb.c b/libvpx/libvpx/vp8/encoder/encodemb.c
new file mode 100644
index 0000000..932a157
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/encodemb.c
@@ -0,0 +1,593 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "encodemb.h"
+#include "vp8/common/reconinter.h"
+#include "vp8/encoder/quantize.h"
+#include "tokenize.h"
+#include "vp8/common/invtrans.h"
+#include "vpx_mem/vpx_mem.h"
+#include "rdopt.h"
+
+void vp8_subtract_b(BLOCK *be, BLOCKD *bd, int pitch) {
+  unsigned char *src_ptr = (*(be->base_src) + be->src);
+  short *diff_ptr = be->src_diff;
+  unsigned char *pred_ptr = bd->predictor;
+  int src_stride = be->src_stride;
+
+  vpx_subtract_block(4, 4, diff_ptr, pitch, src_ptr, src_stride,
+                     pred_ptr, pitch);
+}
+
+void vp8_subtract_mbuv(short *diff, unsigned char *usrc, unsigned char *vsrc,
+                         int src_stride, unsigned char *upred,
+                         unsigned char *vpred, int pred_stride) {
+  short *udiff = diff + 256;
+  short *vdiff = diff + 320;
+
+  vpx_subtract_block(8, 8, udiff, 8, usrc, src_stride, upred, pred_stride);
+  vpx_subtract_block(8, 8, vdiff, 8, vsrc, src_stride, vpred, pred_stride);
+}
+
+void vp8_subtract_mby(short *diff, unsigned char *src, int src_stride,
+                      unsigned char *pred, int pred_stride) {
+  vpx_subtract_block(16, 16, diff, 16, src, src_stride, pred, pred_stride);
+}
+
+static void vp8_subtract_mb(MACROBLOCK *x)
+{
+    BLOCK *b = &x->block[0];
+
+    vp8_subtract_mby(x->src_diff, *(b->base_src),
+        b->src_stride, x->e_mbd.dst.y_buffer, x->e_mbd.dst.y_stride);
+    vp8_subtract_mbuv(x->src_diff, x->src.u_buffer,
+        x->src.v_buffer, x->src.uv_stride, x->e_mbd.dst.u_buffer,
+        x->e_mbd.dst.v_buffer, x->e_mbd.dst.uv_stride);
+}
+
+static void build_dcblock(MACROBLOCK *x)
+{
+    short *src_diff_ptr = &x->src_diff[384];
+    int i;
+
+    for (i = 0; i < 16; i++)
+    {
+        src_diff_ptr[i] = x->coeff[i * 16];
+    }
+}
+
+void vp8_transform_mbuv(MACROBLOCK *x)
+{
+    int i;
+
+    for (i = 16; i < 24; i += 2)
+    {
+        x->short_fdct8x4(&x->block[i].src_diff[0],
+            &x->block[i].coeff[0], 16);
+    }
+}
+
+
+void vp8_transform_intra_mby(MACROBLOCK *x)
+{
+    int i;
+
+    for (i = 0; i < 16; i += 2)
+    {
+        x->short_fdct8x4(&x->block[i].src_diff[0],
+            &x->block[i].coeff[0], 32);
+    }
+
+    /* build dc block from 16 y dc values */
+    build_dcblock(x);
+
+    /* do 2nd order transform on the dc block */
+    x->short_walsh4x4(&x->block[24].src_diff[0],
+        &x->block[24].coeff[0], 8);
+
+}
+
+
+static void transform_mb(MACROBLOCK *x)
+{
+    int i;
+
+    for (i = 0; i < 16; i += 2)
+    {
+        x->short_fdct8x4(&x->block[i].src_diff[0],
+            &x->block[i].coeff[0], 32);
+    }
+
+    /* build dc block from 16 y dc values */
+    if (x->e_mbd.mode_info_context->mbmi.mode != SPLITMV)
+        build_dcblock(x);
+
+    for (i = 16; i < 24; i += 2)
+    {
+        x->short_fdct8x4(&x->block[i].src_diff[0],
+            &x->block[i].coeff[0], 16);
+    }
+
+    /* do 2nd order transform on the dc block */
+    if (x->e_mbd.mode_info_context->mbmi.mode != SPLITMV)
+        x->short_walsh4x4(&x->block[24].src_diff[0],
+        &x->block[24].coeff[0], 8);
+
+}
+
+
+static void transform_mby(MACROBLOCK *x)
+{
+    int i;
+
+    for (i = 0; i < 16; i += 2)
+    {
+        x->short_fdct8x4(&x->block[i].src_diff[0],
+            &x->block[i].coeff[0], 32);
+    }
+
+    /* build dc block from 16 y dc values */
+    if (x->e_mbd.mode_info_context->mbmi.mode != SPLITMV)
+    {
+        build_dcblock(x);
+        x->short_walsh4x4(&x->block[24].src_diff[0],
+            &x->block[24].coeff[0], 8);
+    }
+}
+
+
+
+#define RDTRUNC(RM,DM,R,D) ( (128+(R)*(RM)) & 0xFF )
+
+typedef struct vp8_token_state vp8_token_state;
+
+struct vp8_token_state{
+  int           rate;
+  int           error;
+  signed char   next;
+  signed char   token;
+  short         qc;
+};
+
+/* TODO: experiments to find optimal multiple numbers */
+#define Y1_RD_MULT 4
+#define UV_RD_MULT 2
+#define Y2_RD_MULT 16
+
+static const int plane_rd_mult[4]=
+{
+    Y1_RD_MULT,
+    Y2_RD_MULT,
+    UV_RD_MULT,
+    Y1_RD_MULT
+};
+
+static void optimize_b(MACROBLOCK *mb, int ib, int type,
+                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l)
+{
+    BLOCK *b;
+    BLOCKD *d;
+    vp8_token_state tokens[17][2];
+    unsigned best_mask[2];
+    const short *dequant_ptr;
+    const short *coeff_ptr;
+    short *qcoeff_ptr;
+    short *dqcoeff_ptr;
+    int eob;
+    int i0;
+    int rc;
+    int x;
+    int sz = 0;
+    int next;
+    int rdmult;
+    int rddiv;
+    int final_eob;
+    int rd_cost0;
+    int rd_cost1;
+    int rate0;
+    int rate1;
+    int error0;
+    int error1;
+    int t0;
+    int t1;
+    int best;
+    int band;
+    int pt;
+    int i;
+    int err_mult = plane_rd_mult[type];
+
+    b = &mb->block[ib];
+    d = &mb->e_mbd.block[ib];
+
+    dequant_ptr = d->dequant;
+    coeff_ptr = b->coeff;
+    qcoeff_ptr = d->qcoeff;
+    dqcoeff_ptr = d->dqcoeff;
+    i0 = !type;
+    eob = *d->eob;
+
+    /* Now set up a Viterbi trellis to evaluate alternative roundings. */
+    rdmult = mb->rdmult * err_mult;
+    if(mb->e_mbd.mode_info_context->mbmi.ref_frame==INTRA_FRAME)
+        rdmult = (rdmult * 9)>>4;
+
+    rddiv = mb->rddiv;
+    best_mask[0] = best_mask[1] = 0;
+    /* Initialize the sentinel node of the trellis. */
+    tokens[eob][0].rate = 0;
+    tokens[eob][0].error = 0;
+    tokens[eob][0].next = 16;
+    tokens[eob][0].token = DCT_EOB_TOKEN;
+    tokens[eob][0].qc = 0;
+    *(tokens[eob] + 1) = *(tokens[eob] + 0);
+    next = eob;
+    for (i = eob; i-- > i0;)
+    {
+        int base_bits;
+        int d2;
+        int dx;
+
+        rc = vp8_default_zig_zag1d[i];
+        x = qcoeff_ptr[rc];
+        /* Only add a trellis state for non-zero coefficients. */
+        if (x)
+        {
+            int shortcut=0;
+            error0 = tokens[next][0].error;
+            error1 = tokens[next][1].error;
+            /* Evaluate the first possibility for this state. */
+            rate0 = tokens[next][0].rate;
+            rate1 = tokens[next][1].rate;
+            t0 = (vp8_dct_value_tokens_ptr + x)->Token;
+            /* Consider both possible successor states. */
+            if (next < 16)
+            {
+                band = vp8_coef_bands[i + 1];
+                pt = vp8_prev_token_class[t0];
+                rate0 +=
+                    mb->token_costs[type][band][pt][tokens[next][0].token];
+                rate1 +=
+                    mb->token_costs[type][band][pt][tokens[next][1].token];
+            }
+            rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);
+            rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);
+            if (rd_cost0 == rd_cost1)
+            {
+                rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);
+                rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);
+            }
+            /* And pick the best. */
+            best = rd_cost1 < rd_cost0;
+            base_bits = *(vp8_dct_value_cost_ptr + x);
+            dx = dqcoeff_ptr[rc] - coeff_ptr[rc];
+            d2 = dx*dx;
+            tokens[i][0].rate = base_bits + (best ? rate1 : rate0);
+            tokens[i][0].error = d2 + (best ? error1 : error0);
+            tokens[i][0].next = next;
+            tokens[i][0].token = t0;
+            tokens[i][0].qc = x;
+            best_mask[0] |= best << i;
+            /* Evaluate the second possibility for this state. */
+            rate0 = tokens[next][0].rate;
+            rate1 = tokens[next][1].rate;
+
+            if((abs(x)*dequant_ptr[rc]>abs(coeff_ptr[rc])) &&
+               (abs(x)*dequant_ptr[rc]<abs(coeff_ptr[rc])+dequant_ptr[rc]))
+                shortcut = 1;
+            else
+                shortcut = 0;
+
+            if(shortcut)
+            {
+                sz = -(x < 0);
+                x -= 2*sz + 1;
+            }
+
+            /* Consider both possible successor states. */
+            if (!x)
+            {
+                /* If we reduced this coefficient to zero, check to see if
+                 *  we need to move the EOB back here.
+                 */
+                t0 = tokens[next][0].token == DCT_EOB_TOKEN ?
+                    DCT_EOB_TOKEN : ZERO_TOKEN;
+                t1 = tokens[next][1].token == DCT_EOB_TOKEN ?
+                    DCT_EOB_TOKEN : ZERO_TOKEN;
+            }
+            else
+            {
+                t0=t1 = (vp8_dct_value_tokens_ptr + x)->Token;
+            }
+            if (next < 16)
+            {
+                band = vp8_coef_bands[i + 1];
+                if(t0!=DCT_EOB_TOKEN)
+                {
+                    pt = vp8_prev_token_class[t0];
+                    rate0 += mb->token_costs[type][band][pt][
+                        tokens[next][0].token];
+                }
+                if(t1!=DCT_EOB_TOKEN)
+                {
+                    pt = vp8_prev_token_class[t1];
+                    rate1 += mb->token_costs[type][band][pt][
+                        tokens[next][1].token];
+                }
+            }
+
+            rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);
+            rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);
+            if (rd_cost0 == rd_cost1)
+            {
+                rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);
+                rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);
+            }
+            /* And pick the best. */
+            best = rd_cost1 < rd_cost0;
+            base_bits = *(vp8_dct_value_cost_ptr + x);
+
+            if(shortcut)
+            {
+                dx -= (dequant_ptr[rc] + sz) ^ sz;
+                d2 = dx*dx;
+            }
+            tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
+            tokens[i][1].error = d2 + (best ? error1 : error0);
+            tokens[i][1].next = next;
+            tokens[i][1].token =best?t1:t0;
+            tokens[i][1].qc = x;
+            best_mask[1] |= best << i;
+            /* Finally, make this the new head of the trellis. */
+            next = i;
+        }
+        /* There's no choice to make for a zero coefficient, so we don't
+         *  add a new trellis node, but we do need to update the costs.
+         */
+        else
+        {
+            band = vp8_coef_bands[i + 1];
+            t0 = tokens[next][0].token;
+            t1 = tokens[next][1].token;
+            /* Update the cost of each path if we're past the EOB token. */
+            if (t0 != DCT_EOB_TOKEN)
+            {
+                tokens[next][0].rate += mb->token_costs[type][band][0][t0];
+                tokens[next][0].token = ZERO_TOKEN;
+            }
+            if (t1 != DCT_EOB_TOKEN)
+            {
+                tokens[next][1].rate += mb->token_costs[type][band][0][t1];
+                tokens[next][1].token = ZERO_TOKEN;
+            }
+            /* Don't update next, because we didn't add a new node. */
+        }
+    }
+
+    /* Now pick the best path through the whole trellis. */
+    band = vp8_coef_bands[i + 1];
+    VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
+    rate0 = tokens[next][0].rate;
+    rate1 = tokens[next][1].rate;
+    error0 = tokens[next][0].error;
+    error1 = tokens[next][1].error;
+    t0 = tokens[next][0].token;
+    t1 = tokens[next][1].token;
+    rate0 += mb->token_costs[type][band][pt][t0];
+    rate1 += mb->token_costs[type][band][pt][t1];
+    rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);
+    rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);
+    if (rd_cost0 == rd_cost1)
+    {
+        rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);
+        rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);
+    }
+    best = rd_cost1 < rd_cost0;
+    final_eob = i0 - 1;
+    for (i = next; i < eob; i = next)
+    {
+        x = tokens[i][best].qc;
+        if (x)
+            final_eob = i;
+        rc = vp8_default_zig_zag1d[i];
+        qcoeff_ptr[rc] = x;
+        dqcoeff_ptr[rc] = x * dequant_ptr[rc];
+        next = tokens[i][best].next;
+        best = (best_mask[best] >> i) & 1;
+    }
+    final_eob++;
+
+    *a = *l = (final_eob != !type);
+    *d->eob = (char)final_eob;
+}
+static void check_reset_2nd_coeffs(MACROBLOCKD *x, int type,
+                                   ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l)
+{
+    int sum=0;
+    int i;
+    BLOCKD *bd = &x->block[24];
+
+    if(bd->dequant[0]>=35 && bd->dequant[1]>=35)
+        return;
+
+    for(i=0;i<(*bd->eob);i++)
+    {
+        int coef = bd->dqcoeff[vp8_default_zig_zag1d[i]];
+        sum+= (coef>=0)?coef:-coef;
+        if(sum>=35)
+            return;
+    }
+    /**************************************************************************
+    our inverse hadamard transform effectively is weighted sum of all 16 inputs
+    with weight either 1 or -1. It has a last stage scaling of (sum+3)>>3. And
+    dc only idct is (dc+4)>>3. So if all the sums are between -35 and 29, the
+    output after inverse wht and idct will be all zero. A sum of absolute value
+    smaller than 35 guarantees all 16 different (+1/-1) weighted sums in wht
+    fall between -35 and +35.
+    **************************************************************************/
+    if(sum < 35)
+    {
+        for(i=0;i<(*bd->eob);i++)
+        {
+            int rc = vp8_default_zig_zag1d[i];
+            bd->qcoeff[rc]=0;
+            bd->dqcoeff[rc]=0;
+        }
+        *bd->eob = 0;
+        *a = *l = (*bd->eob != !type);
+    }
+}
+
+static void optimize_mb(MACROBLOCK *x)
+{
+    int b;
+    int type;
+    int has_2nd_order;
+
+    ENTROPY_CONTEXT_PLANES t_above, t_left;
+    ENTROPY_CONTEXT *ta;
+    ENTROPY_CONTEXT *tl;
+
+    memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
+    memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
+
+    ta = (ENTROPY_CONTEXT *)&t_above;
+    tl = (ENTROPY_CONTEXT *)&t_left;
+
+    has_2nd_order = (x->e_mbd.mode_info_context->mbmi.mode != B_PRED
+        && x->e_mbd.mode_info_context->mbmi.mode != SPLITMV);
+    type = has_2nd_order ? PLANE_TYPE_Y_NO_DC : PLANE_TYPE_Y_WITH_DC;
+
+    for (b = 0; b < 16; b++)
+    {
+        optimize_b(x, b, type,
+            ta + vp8_block2above[b], tl + vp8_block2left[b]);
+    }
+
+    for (b = 16; b < 24; b++)
+    {
+        optimize_b(x, b, PLANE_TYPE_UV,
+            ta + vp8_block2above[b], tl + vp8_block2left[b]);
+    }
+
+    if (has_2nd_order)
+    {
+        b=24;
+        optimize_b(x, b, PLANE_TYPE_Y2,
+            ta + vp8_block2above[b], tl + vp8_block2left[b]);
+        check_reset_2nd_coeffs(&x->e_mbd, PLANE_TYPE_Y2,
+            ta + vp8_block2above[b], tl + vp8_block2left[b]);
+    }
+}
+
+
+void vp8_optimize_mby(MACROBLOCK *x)
+{
+    int b;
+    int type;
+    int has_2nd_order;
+
+    ENTROPY_CONTEXT_PLANES t_above, t_left;
+    ENTROPY_CONTEXT *ta;
+    ENTROPY_CONTEXT *tl;
+
+    if (!x->e_mbd.above_context)
+        return;
+
+    if (!x->e_mbd.left_context)
+        return;
+
+    memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
+    memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
+
+    ta = (ENTROPY_CONTEXT *)&t_above;
+    tl = (ENTROPY_CONTEXT *)&t_left;
+
+    has_2nd_order = (x->e_mbd.mode_info_context->mbmi.mode != B_PRED
+        && x->e_mbd.mode_info_context->mbmi.mode != SPLITMV);
+    type = has_2nd_order ? PLANE_TYPE_Y_NO_DC : PLANE_TYPE_Y_WITH_DC;
+
+    for (b = 0; b < 16; b++)
+    {
+        optimize_b(x, b, type,
+            ta + vp8_block2above[b], tl + vp8_block2left[b]);
+    }
+
+
+    if (has_2nd_order)
+    {
+        b=24;
+        optimize_b(x, b, PLANE_TYPE_Y2,
+            ta + vp8_block2above[b], tl + vp8_block2left[b]);
+        check_reset_2nd_coeffs(&x->e_mbd, PLANE_TYPE_Y2,
+            ta + vp8_block2above[b], tl + vp8_block2left[b]);
+    }
+}
+
+void vp8_optimize_mbuv(MACROBLOCK *x)
+{
+    int b;
+    ENTROPY_CONTEXT_PLANES t_above, t_left;
+    ENTROPY_CONTEXT *ta;
+    ENTROPY_CONTEXT *tl;
+
+    if (!x->e_mbd.above_context)
+        return;
+
+    if (!x->e_mbd.left_context)
+        return;
+
+    memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
+    memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
+
+    ta = (ENTROPY_CONTEXT *)&t_above;
+    tl = (ENTROPY_CONTEXT *)&t_left;
+
+    for (b = 16; b < 24; b++)
+    {
+        optimize_b(x, b, PLANE_TYPE_UV,
+            ta + vp8_block2above[b], tl + vp8_block2left[b]);
+    }
+}
+
+void vp8_encode_inter16x16(MACROBLOCK *x)
+{
+    vp8_build_inter_predictors_mb(&x->e_mbd);
+
+    vp8_subtract_mb(x);
+
+    transform_mb(x);
+
+    vp8_quantize_mb(x);
+
+    if (x->optimize)
+        optimize_mb(x);
+}
+
+/* this funciton is used by first pass only */
+void vp8_encode_inter16x16y(MACROBLOCK *x)
+{
+    BLOCK *b = &x->block[0];
+
+    vp8_build_inter16x16_predictors_mby(&x->e_mbd, x->e_mbd.dst.y_buffer,
+                                        x->e_mbd.dst.y_stride);
+
+    vp8_subtract_mby(x->src_diff, *(b->base_src),
+        b->src_stride, x->e_mbd.dst.y_buffer, x->e_mbd.dst.y_stride);
+
+    transform_mby(x);
+
+    vp8_quantize_mby(x);
+
+    vp8_inverse_transform_mby(&x->e_mbd);
+}
diff --git a/libvpx/libvpx/vp8/encoder/encodemb.h b/libvpx/libvpx/vp8/encoder/encodemb.h
new file mode 100644
index 0000000..10b3d86
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/encodemb.h
@@ -0,0 +1,41 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_ENCODEMB_H_
+#define VP8_ENCODER_ENCODEMB_H_
+
+#include "onyx_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+void vp8_encode_inter16x16(MACROBLOCK *x);
+
+void vp8_subtract_b(BLOCK *be, BLOCKD *bd, int pitch);
+void vp8_subtract_mbuv(short *diff, unsigned char *usrc, unsigned char *vsrc,
+                       int src_stride, unsigned char *upred,
+                       unsigned char *vpred, int pred_stride);
+void vp8_subtract_mby(short *diff, unsigned char *src, int src_stride,
+                      unsigned char *pred, int pred_stride);
+
+void vp8_build_dcblock(MACROBLOCK *b);
+void vp8_transform_mb(MACROBLOCK *mb);
+void vp8_transform_mbuv(MACROBLOCK *x);
+void vp8_transform_intra_mby(MACROBLOCK *x);
+
+void vp8_optimize_mby(MACROBLOCK *x);
+void vp8_optimize_mbuv(MACROBLOCK *x);
+void vp8_encode_inter16x16y(MACROBLOCK *x);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_ENCODEMB_H_
diff --git a/libvpx/libvpx/vp8/encoder/encodemv.c b/libvpx/libvpx/vp8/encoder/encodemv.c
new file mode 100644
index 0000000..2a74ff4
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/encodemv.c
@@ -0,0 +1,380 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vp8/common/common.h"
+#include "encodemv.h"
+#include "vp8/common/entropymode.h"
+#include "vp8/common/systemdependent.h"
+
+#include <math.h>
+
+#ifdef VP8_ENTROPY_STATS
+extern unsigned int active_section;
+#endif
+
+static void encode_mvcomponent(
+    vp8_writer *const w,
+    const int v,
+    const struct mv_context *mvc
+)
+{
+    const vp8_prob *p = mvc->prob;
+    const int x = v < 0 ? -v : v;
+
+    if (x < mvnum_short)     /* Small */
+    {
+        vp8_write(w, 0, p [mvpis_short]);
+        vp8_treed_write(w, vp8_small_mvtree, p + MVPshort, x, 3);
+
+        if (!x)
+            return;         /* no sign bit */
+    }
+    else                    /* Large */
+    {
+        int i = 0;
+
+        vp8_write(w, 1, p [mvpis_short]);
+
+        do
+            vp8_write(w, (x >> i) & 1, p [MVPbits + i]);
+
+        while (++i < 3);
+
+        i = mvlong_width - 1;  /* Skip bit 3, which is sometimes implicit */
+
+        do
+            vp8_write(w, (x >> i) & 1, p [MVPbits + i]);
+
+        while (--i > 3);
+
+        if (x & 0xFFF0)
+            vp8_write(w, (x >> 3) & 1, p [MVPbits + 3]);
+    }
+
+    vp8_write(w, v < 0, p [MVPsign]);
+}
+#if 0
+static int max_mv_r = 0;
+static int max_mv_c = 0;
+#endif
+void vp8_encode_motion_vector(vp8_writer *w, const MV *mv, const MV_CONTEXT *mvc)
+{
+
+#if 0
+    {
+        if (abs(mv->row >> 1) > max_mv_r)
+        {
+            FILE *f = fopen("maxmv.stt", "a");
+            max_mv_r = abs(mv->row >> 1);
+            fprintf(f, "New Mv Row Max %6d\n", (mv->row >> 1));
+
+            if ((abs(mv->row) / 2) != max_mv_r)
+                fprintf(f, "MV Row conversion error %6d\n", abs(mv->row) / 2);
+
+            fclose(f);
+        }
+
+        if (abs(mv->col >> 1) > max_mv_c)
+        {
+            FILE *f = fopen("maxmv.stt", "a");
+            fprintf(f, "New Mv Col Max %6d\n", (mv->col >> 1));
+            max_mv_c = abs(mv->col >> 1);
+            fclose(f);
+        }
+    }
+#endif
+
+    encode_mvcomponent(w, mv->row >> 1, &mvc[0]);
+    encode_mvcomponent(w, mv->col >> 1, &mvc[1]);
+}
+
+
+static unsigned int cost_mvcomponent(const int v, const struct mv_context *mvc)
+{
+    const vp8_prob *p = mvc->prob;
+    const int x = v;
+    unsigned int cost;
+
+    if (x < mvnum_short)
+    {
+        cost = vp8_cost_zero(p [mvpis_short])
+               + vp8_treed_cost(vp8_small_mvtree, p + MVPshort, x, 3);
+
+        if (!x)
+            return cost;
+    }
+    else
+    {
+        int i = 0;
+        cost = vp8_cost_one(p [mvpis_short]);
+
+        do
+            cost += vp8_cost_bit(p [MVPbits + i], (x >> i) & 1);
+
+        while (++i < 3);
+
+        i = mvlong_width - 1;  /* Skip bit 3, which is sometimes implicit */
+
+        do
+            cost += vp8_cost_bit(p [MVPbits + i], (x >> i) & 1);
+
+        while (--i > 3);
+
+        if (x & 0xFFF0)
+            cost += vp8_cost_bit(p [MVPbits + 3], (x >> 3) & 1);
+    }
+
+    return cost;   /* + vp8_cost_bit( p [MVPsign], v < 0); */
+}
+
+void vp8_build_component_cost_table(int *mvcost[2], const MV_CONTEXT *mvc, int mvc_flag[2])
+{
+    int i = 1;
+    unsigned int cost0 = 0;
+    unsigned int cost1 = 0;
+
+    vp8_clear_system_state();
+
+    i = 1;
+
+    if (mvc_flag[0])
+    {
+        mvcost [0] [0] = cost_mvcomponent(0, &mvc[0]);
+
+        do
+        {
+            cost0 = cost_mvcomponent(i, &mvc[0]);
+
+            mvcost [0] [i] = cost0 + vp8_cost_zero(mvc[0].prob[MVPsign]);
+            mvcost [0] [-i] = cost0 + vp8_cost_one(mvc[0].prob[MVPsign]);
+        }
+        while (++i <= mv_max);
+    }
+
+    i = 1;
+
+    if (mvc_flag[1])
+    {
+        mvcost [1] [0] = cost_mvcomponent(0, &mvc[1]);
+
+        do
+        {
+            cost1 = cost_mvcomponent(i, &mvc[1]);
+
+            mvcost [1] [i] = cost1 + vp8_cost_zero(mvc[1].prob[MVPsign]);
+            mvcost [1] [-i] = cost1 + vp8_cost_one(mvc[1].prob[MVPsign]);
+        }
+        while (++i <= mv_max);
+    }
+}
+
+
+/* Motion vector probability table update depends on benefit.
+ * Small correction allows for the fact that an update to an MV probability
+ * may have benefit in subsequent frames as well as the current one.
+ */
+#define MV_PROB_UPDATE_CORRECTION   -1
+
+
+static void calc_prob(vp8_prob *p, const unsigned int ct[2])
+{
+    const unsigned int tot = ct[0] + ct[1];
+
+    if (tot)
+    {
+        const vp8_prob x = ((ct[0] * 255) / tot) & -2;
+        *p = x ? x : 1;
+    }
+}
+
+static void update(
+    vp8_writer *const w,
+    const unsigned int ct[2],
+    vp8_prob *const cur_p,
+    const vp8_prob new_p,
+    const vp8_prob update_p,
+    int *updated
+)
+{
+    const int cur_b = vp8_cost_branch(ct, *cur_p);
+    const int new_b = vp8_cost_branch(ct, new_p);
+    const int cost = 7 + MV_PROB_UPDATE_CORRECTION + ((vp8_cost_one(update_p) - vp8_cost_zero(update_p) + 128) >> 8);
+
+    if (cur_b - new_b > cost)
+    {
+        *cur_p = new_p;
+        vp8_write(w, 1, update_p);
+        vp8_write_literal(w, new_p >> 1, 7);
+        *updated = 1;
+
+    }
+    else
+        vp8_write(w, 0, update_p);
+}
+
+static void write_component_probs(
+    vp8_writer *const w,
+    struct mv_context *cur_mvc,
+    const struct mv_context *default_mvc_,
+    const struct mv_context *update_mvc,
+    const unsigned int events [MVvals],
+    unsigned int rc,
+    int *updated
+)
+{
+    vp8_prob *Pcur = cur_mvc->prob;
+    const vp8_prob *default_mvc = default_mvc_->prob;
+    const vp8_prob *Pupdate = update_mvc->prob;
+    unsigned int is_short_ct[2], sign_ct[2];
+
+    unsigned int bit_ct [mvlong_width] [2];
+
+    unsigned int short_ct  [mvnum_short];
+    unsigned int short_bct [mvnum_short-1] [2];
+
+    vp8_prob Pnew [MVPcount];
+
+    (void) rc;
+    vp8_copy_array(Pnew, default_mvc, MVPcount);
+
+    vp8_zero(is_short_ct)
+    vp8_zero(sign_ct)
+    vp8_zero(bit_ct)
+    vp8_zero(short_ct)
+    vp8_zero(short_bct)
+
+
+    /* j=0 */
+    {
+        const int c = events [mv_max];
+
+        is_short_ct [0] += c;     /* Short vector */
+        short_ct [0] += c;       /* Magnitude distribution */
+    }
+
+    /* j: 1 ~ mv_max (1023) */
+    {
+        int j = 1;
+
+        do
+        {
+            const int c1 = events [mv_max + j];  /* positive */
+            const int c2 = events [mv_max - j];  /* negative */
+            const int c  = c1 + c2;
+            int a = j;
+
+            sign_ct [0] += c1;
+            sign_ct [1] += c2;
+
+            if (a < mvnum_short)
+            {
+                is_short_ct [0] += c;     /* Short vector */
+                short_ct [a] += c;       /* Magnitude distribution */
+            }
+            else
+            {
+                int k = mvlong_width - 1;
+                is_short_ct [1] += c;     /* Long vector */
+
+                /*  bit 3 not always encoded. */
+                do
+                    bit_ct [k] [(a >> k) & 1] += c;
+
+                while (--k >= 0);
+            }
+        }
+        while (++j <= mv_max);
+    }
+
+    calc_prob(Pnew + mvpis_short, is_short_ct);
+
+    calc_prob(Pnew + MVPsign, sign_ct);
+
+    {
+        vp8_prob p [mvnum_short - 1];    /* actually only need branch ct */
+        int j = 0;
+
+        vp8_tree_probs_from_distribution(
+            8, vp8_small_mvencodings, vp8_small_mvtree,
+            p, short_bct, short_ct,
+            256, 1
+        );
+
+        do
+            calc_prob(Pnew + MVPshort + j, short_bct[j]);
+
+        while (++j < mvnum_short - 1);
+    }
+
+    {
+        int j = 0;
+
+        do
+            calc_prob(Pnew + MVPbits + j, bit_ct[j]);
+
+        while (++j < mvlong_width);
+    }
+
+    update(w, is_short_ct, Pcur + mvpis_short, Pnew[mvpis_short], *Pupdate++, updated);
+
+    update(w, sign_ct, Pcur + MVPsign, Pnew[MVPsign], *Pupdate++, updated);
+
+    {
+        const vp8_prob *const new_p = Pnew + MVPshort;
+        vp8_prob *const cur_p = Pcur + MVPshort;
+
+        int j = 0;
+
+        do
+
+            update(w, short_bct[j], cur_p + j, new_p[j], *Pupdate++, updated);
+
+        while (++j < mvnum_short - 1);
+    }
+
+    {
+        const vp8_prob *const new_p = Pnew + MVPbits;
+        vp8_prob *const cur_p = Pcur + MVPbits;
+
+        int j = 0;
+
+        do
+
+            update(w, bit_ct[j], cur_p + j, new_p[j], *Pupdate++, updated);
+
+        while (++j < mvlong_width);
+    }
+}
+
+void vp8_write_mvprobs(VP8_COMP *cpi)
+{
+    vp8_writer *const w  = cpi->bc;
+    MV_CONTEXT *mvc = cpi->common.fc.mvc;
+    int flags[2] = {0, 0};
+#ifdef VP8_ENTROPY_STATS
+    active_section = 4;
+#endif
+    write_component_probs(
+        w, &mvc[0], &vp8_default_mv_context[0], &vp8_mv_update_probs[0],
+        cpi->mb.MVcount[0], 0, &flags[0]
+    );
+    write_component_probs(
+        w, &mvc[1], &vp8_default_mv_context[1], &vp8_mv_update_probs[1],
+        cpi->mb.MVcount[1], 1, &flags[1]
+    );
+
+    if (flags[0] || flags[1])
+        vp8_build_component_cost_table(cpi->mb.mvcost, (const MV_CONTEXT *) cpi->common.fc.mvc, flags);
+
+#ifdef VP8_ENTROPY_STATS
+    active_section = 5;
+#endif
+}
diff --git a/libvpx/libvpx/vp8/encoder/encodemv.h b/libvpx/libvpx/vp8/encoder/encodemv.h
new file mode 100644
index 0000000..722162b
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/encodemv.h
@@ -0,0 +1,29 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_ENCODEMV_H_
+#define VP8_ENCODER_ENCODEMV_H_
+
+#include "onyx_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_write_mvprobs(VP8_COMP *);
+void vp8_encode_motion_vector(vp8_writer *, const MV *, const MV_CONTEXT *);
+void vp8_build_component_cost_table(int *mvcost[2], const MV_CONTEXT *mvc, int mvc_flag[2]);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_ENCODEMV_H_
diff --git a/libvpx/libvpx/vp8/encoder/ethreading.c b/libvpx/libvpx/vp8/encoder/ethreading.c
new file mode 100644
index 0000000..2a0c298
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/ethreading.c
@@ -0,0 +1,679 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "onyx_int.h"
+#include "vp8/common/threading.h"
+#include "vp8/common/common.h"
+#include "vp8/common/extend.h"
+#include "bitstream.h"
+#include "encodeframe.h"
+
+#if CONFIG_MULTITHREAD
+
+extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x, int ok_to_skip);
+
+static THREAD_FUNCTION thread_loopfilter(void *p_data)
+{
+    VP8_COMP *cpi = (VP8_COMP *)(((LPFTHREAD_DATA *)p_data)->ptr1);
+    VP8_COMMON *cm = &cpi->common;
+
+    while (1)
+    {
+        if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0)
+            break;
+
+        if (sem_wait(&cpi->h_event_start_lpf) == 0)
+        {
+            /* we're shutting down */
+            if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0)
+                break;
+
+            vp8_loopfilter_frame(cpi, cm);
+
+            sem_post(&cpi->h_event_end_lpf);
+        }
+    }
+
+    return 0;
+}
+
+static
+THREAD_FUNCTION thread_encoding_proc(void *p_data)
+{
+    int ithread = ((ENCODETHREAD_DATA *)p_data)->ithread;
+    VP8_COMP *cpi = (VP8_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr1);
+    MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
+    ENTROPY_CONTEXT_PLANES mb_row_left_context;
+
+    while (1)
+    {
+        if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0)
+            break;
+
+        if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0)
+        {
+            const int nsync = cpi->mt_sync_range;
+            VP8_COMMON *cm = &cpi->common;
+            int mb_row;
+            MACROBLOCK *x = &mbri->mb;
+            MACROBLOCKD *xd = &x->e_mbd;
+            TOKENEXTRA *tp ;
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+            TOKENEXTRA *tp_start = cpi->tok + (1 + ithread) * (16 * 24);
+            const int num_part = (1 << cm->multi_token_partition);
+#endif
+
+            int *segment_counts = mbri->segment_counts;
+            int *totalrate = &mbri->totalrate;
+
+            /* we're shutting down */
+            if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0)
+                break;
+
+            xd->mode_info_context = cm->mi + cm->mode_info_stride *
+                (ithread + 1);
+            xd->mode_info_stride = cm->mode_info_stride;
+
+            for (mb_row = ithread + 1; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
+            {
+
+                int recon_yoffset, recon_uvoffset;
+                int mb_col;
+                int ref_fb_idx = cm->lst_fb_idx;
+                int dst_fb_idx = cm->new_fb_idx;
+                int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
+                int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
+                int map_index = (mb_row * cm->mb_cols);
+                const int *last_row_current_mb_col;
+                int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
+
+#if  (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+                vp8_writer *w = &cpi->bc[1 + (mb_row % num_part)];
+#else
+                tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
+                cpi->tplist[mb_row].start = tp;
+#endif
+
+                last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
+
+                /* reset above block coeffs */
+                xd->above_context = cm->above_context;
+                xd->left_context = &mb_row_left_context;
+
+                vp8_zero(mb_row_left_context);
+
+                xd->up_available = (mb_row != 0);
+                recon_yoffset = (mb_row * recon_y_stride * 16);
+                recon_uvoffset = (mb_row * recon_uv_stride * 8);
+
+                /* Set the mb activity pointer to the start of the row. */
+                x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
+
+                /* for each macroblock col in image */
+                for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
+                {
+                    if (((mb_col - 1) % nsync) == 0) {
+                        pthread_mutex_t *mutex = &cpi->pmutex[mb_row];
+                        protected_write(mutex, current_mb_col, mb_col - 1);
+                    }
+
+                    if (mb_row && !(mb_col & (nsync - 1))) {
+                      pthread_mutex_t *mutex = &cpi->pmutex[mb_row-1];
+                      sync_read(mutex, mb_col, last_row_current_mb_col, nsync);
+                    }
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+                    tp = tp_start;
+#endif
+
+                    /* Distance of Mb to the various image edges.
+                     * These specified to 8th pel as they are always compared
+                     * to values that are in 1/8th pel units
+                     */
+                    xd->mb_to_left_edge = -((mb_col * 16) << 3);
+                    xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
+                    xd->mb_to_top_edge = -((mb_row * 16) << 3);
+                    xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
+
+                    /* Set up limit values for motion vectors used to prevent
+                     * them extending outside the UMV borders
+                     */
+                    x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
+                    x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
+                    x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
+                    x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
+
+                    xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
+                    xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
+                    xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
+                    xd->left_available = (mb_col != 0);
+
+                    x->rddiv = cpi->RDDIV;
+                    x->rdmult = cpi->RDMULT;
+
+                    /* Copy current mb to a buffer */
+                    vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
+
+                    if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
+                        vp8_activity_masking(cpi, x);
+
+                    /* Is segmentation enabled */
+                    /* MB level adjustment to quantizer */
+                    if (xd->segmentation_enabled)
+                    {
+                        /* Code to set segment id in xd->mbmi.segment_id for
+                         * current MB (with range checking)
+                         */
+                        if (cpi->segmentation_map[map_index + mb_col] <= 3)
+                            xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index + mb_col];
+                        else
+                            xd->mode_info_context->mbmi.segment_id = 0;
+
+                        vp8cx_mb_init_quantizer(cpi, x, 1);
+                    }
+                    else
+                        /* Set to Segment 0 by default */
+                        xd->mode_info_context->mbmi.segment_id = 0;
+
+                    x->active_ptr = cpi->active_map + map_index + mb_col;
+
+                    if (cm->frame_type == KEY_FRAME)
+                    {
+                        *totalrate += vp8cx_encode_intra_macroblock(cpi, x, &tp);
+#ifdef MODE_STATS
+                        y_modes[xd->mbmi.mode] ++;
+#endif
+                    }
+                    else
+                    {
+                        *totalrate += vp8cx_encode_inter_macroblock(cpi, x, &tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
+
+#ifdef MODE_STATS
+                        inter_y_modes[xd->mbmi.mode] ++;
+
+                        if (xd->mbmi.mode == SPLITMV)
+                        {
+                            int b;
+
+                            for (b = 0; b < xd->mbmi.partition_count; b++)
+                            {
+                                inter_b_modes[x->partition->bmi[b].mode] ++;
+                            }
+                        }
+
+#endif
+                        // Keep track of how many (consecutive) times a  block
+                        // is coded as ZEROMV_LASTREF, for base layer frames.
+                        // Reset to 0 if its coded as anything else.
+                        if (cpi->current_layer == 0) {
+                          if (xd->mode_info_context->mbmi.mode == ZEROMV &&
+                              xd->mode_info_context->mbmi.ref_frame ==
+                                  LAST_FRAME) {
+                            // Increment, check for wrap-around.
+                            if (cpi->consec_zero_last[map_index+mb_col] < 255)
+                              cpi->consec_zero_last[map_index+mb_col] += 1;
+                            if (cpi->consec_zero_last_mvbias[map_index+mb_col] < 255)
+                              cpi->consec_zero_last_mvbias[map_index+mb_col] += 1;
+                          } else {
+                            cpi->consec_zero_last[map_index+mb_col] = 0;
+                            cpi->consec_zero_last_mvbias[map_index+mb_col] = 0;
+                          }
+                          if (x->zero_last_dot_suppress)
+                            cpi->consec_zero_last_mvbias[map_index+mb_col] = 0;
+                        }
+
+                        /* Special case code for cyclic refresh
+                         * If cyclic update enabled then copy
+                         * xd->mbmi.segment_id; (which may have been updated
+                         * based on mode during
+                         * vp8cx_encode_inter_macroblock()) back into the
+                         * global segmentation map
+                         */
+                        if ((cpi->current_layer == 0) &&
+                            (cpi->cyclic_refresh_mode_enabled &&
+                             xd->segmentation_enabled))
+                        {
+                            const MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
+                            cpi->segmentation_map[map_index + mb_col] = mbmi->segment_id;
+
+                            /* If the block has been refreshed mark it as clean
+                             * (the magnitude of the -ve influences how long it
+                             * will be before we consider another refresh):
+                             * Else if it was coded (last frame 0,0) and has
+                             * not already been refreshed then mark it as a
+                             * candidate for cleanup next time (marked 0) else
+                             * mark it as dirty (1).
+                             */
+                            if (mbmi->segment_id)
+                                cpi->cyclic_refresh_map[map_index + mb_col] = -1;
+                            else if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame == LAST_FRAME))
+                            {
+                                if (cpi->cyclic_refresh_map[map_index + mb_col] == 1)
+                                    cpi->cyclic_refresh_map[map_index + mb_col] = 0;
+                            }
+                            else
+                                cpi->cyclic_refresh_map[map_index + mb_col] = 1;
+
+                        }
+                    }
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+                    /* pack tokens for this MB */
+                    {
+                        int tok_count = tp - tp_start;
+                        vp8_pack_tokens(w, tp_start, tok_count);
+                    }
+#else
+                    cpi->tplist[mb_row].stop = tp;
+#endif
+                    /* Increment pointer into gf usage flags structure. */
+                    x->gf_active_ptr++;
+
+                    /* Increment the activity mask pointers. */
+                    x->mb_activity_ptr++;
+
+                    /* adjust to the next column of macroblocks */
+                    x->src.y_buffer += 16;
+                    x->src.u_buffer += 8;
+                    x->src.v_buffer += 8;
+
+                    recon_yoffset += 16;
+                    recon_uvoffset += 8;
+
+                    /* Keep track of segment usage */
+                    segment_counts[xd->mode_info_context->mbmi.segment_id]++;
+
+                    /* skip to next mb */
+                    xd->mode_info_context++;
+                    x->partition_info++;
+                    xd->above_context++;
+                }
+
+                vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
+                                    xd->dst.y_buffer + 16,
+                                    xd->dst.u_buffer + 8,
+                                    xd->dst.v_buffer + 8);
+
+                protected_write(&cpi->pmutex[mb_row], current_mb_col,
+                                mb_col + nsync);
+
+                /* this is to account for the border */
+                xd->mode_info_context++;
+                x->partition_info++;
+
+                x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
+                x->src.u_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
+                x->src.v_buffer += 8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
+
+                xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
+                x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
+                x->gf_active_ptr   += cm->mb_cols * cpi->encoding_thread_count;
+
+                if (mb_row == cm->mb_rows - 1)
+                {
+                    sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
+                }
+            }
+        }
+    }
+
+    /* printf("exit thread %d\n", ithread); */
+    return 0;
+}
+
+static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
+{
+
+    MACROBLOCK *x = mbsrc;
+    MACROBLOCK *z = mbdst;
+    int i;
+
+    z->ss               = x->ss;
+    z->ss_count          = x->ss_count;
+    z->searches_per_step  = x->searches_per_step;
+    z->errorperbit      = x->errorperbit;
+
+    z->sadperbit16      = x->sadperbit16;
+    z->sadperbit4       = x->sadperbit4;
+
+    /*
+    z->mv_col_min    = x->mv_col_min;
+    z->mv_col_max    = x->mv_col_max;
+    z->mv_row_min    = x->mv_row_min;
+    z->mv_row_max    = x->mv_row_max;
+    */
+
+    z->short_fdct4x4     = x->short_fdct4x4;
+    z->short_fdct8x4     = x->short_fdct8x4;
+    z->short_walsh4x4    = x->short_walsh4x4;
+    z->quantize_b        = x->quantize_b;
+    z->optimize          = x->optimize;
+
+    /*
+    z->mvc              = x->mvc;
+    z->src.y_buffer      = x->src.y_buffer;
+    z->src.u_buffer      = x->src.u_buffer;
+    z->src.v_buffer      = x->src.v_buffer;
+    */
+
+    z->mvcost[0] =  x->mvcost[0];
+    z->mvcost[1] =  x->mvcost[1];
+    z->mvsadcost[0] =  x->mvsadcost[0];
+    z->mvsadcost[1] =  x->mvsadcost[1];
+
+    z->token_costs = x->token_costs;
+    z->inter_bmode_costs = x->inter_bmode_costs;
+    z->mbmode_cost = x->mbmode_cost;
+    z->intra_uv_mode_cost = x->intra_uv_mode_cost;
+    z->bmode_costs = x->bmode_costs;
+
+    for (i = 0; i < 25; i++)
+    {
+        z->block[i].quant           = x->block[i].quant;
+        z->block[i].quant_fast      = x->block[i].quant_fast;
+        z->block[i].quant_shift     = x->block[i].quant_shift;
+        z->block[i].zbin            = x->block[i].zbin;
+        z->block[i].zrun_zbin_boost = x->block[i].zrun_zbin_boost;
+        z->block[i].round           = x->block[i].round;
+        z->block[i].src_stride      = x->block[i].src_stride;
+    }
+
+    z->q_index           = x->q_index;
+    z->act_zbin_adj      = x->act_zbin_adj;
+    z->last_act_zbin_adj = x->last_act_zbin_adj;
+
+    {
+        MACROBLOCKD *xd = &x->e_mbd;
+        MACROBLOCKD *zd = &z->e_mbd;
+
+        /*
+        zd->mode_info_context = xd->mode_info_context;
+        zd->mode_info        = xd->mode_info;
+
+        zd->mode_info_stride  = xd->mode_info_stride;
+        zd->frame_type       = xd->frame_type;
+        zd->up_available     = xd->up_available   ;
+        zd->left_available   = xd->left_available;
+        zd->left_context     = xd->left_context;
+        zd->last_frame_dc     = xd->last_frame_dc;
+        zd->last_frame_dccons = xd->last_frame_dccons;
+        zd->gold_frame_dc     = xd->gold_frame_dc;
+        zd->gold_frame_dccons = xd->gold_frame_dccons;
+        zd->mb_to_left_edge    = xd->mb_to_left_edge;
+        zd->mb_to_right_edge   = xd->mb_to_right_edge;
+        zd->mb_to_top_edge     = xd->mb_to_top_edge   ;
+        zd->mb_to_bottom_edge  = xd->mb_to_bottom_edge;
+        zd->gf_active_ptr     = xd->gf_active_ptr;
+        zd->frames_since_golden       = xd->frames_since_golden;
+        zd->frames_till_alt_ref_frame   = xd->frames_till_alt_ref_frame;
+        */
+        zd->subpixel_predict         = xd->subpixel_predict;
+        zd->subpixel_predict8x4      = xd->subpixel_predict8x4;
+        zd->subpixel_predict8x8      = xd->subpixel_predict8x8;
+        zd->subpixel_predict16x16    = xd->subpixel_predict16x16;
+        zd->segmentation_enabled     = xd->segmentation_enabled;
+        zd->mb_segement_abs_delta      = xd->mb_segement_abs_delta;
+        memcpy(zd->segment_feature_data, xd->segment_feature_data,
+               sizeof(xd->segment_feature_data));
+
+        memcpy(zd->dequant_y1_dc, xd->dequant_y1_dc, sizeof(xd->dequant_y1_dc));
+        memcpy(zd->dequant_y1, xd->dequant_y1, sizeof(xd->dequant_y1));
+        memcpy(zd->dequant_y2, xd->dequant_y2, sizeof(xd->dequant_y2));
+        memcpy(zd->dequant_uv, xd->dequant_uv, sizeof(xd->dequant_uv));
+
+#if 1
+        /*TODO:  Remove dequant from BLOCKD.  This is a temporary solution until
+         * the quantizer code uses a passed in pointer to the dequant constants.
+         * This will also require modifications to the x86 and neon assembly.
+         * */
+        for (i = 0; i < 16; i++)
+            zd->block[i].dequant = zd->dequant_y1;
+        for (i = 16; i < 24; i++)
+            zd->block[i].dequant = zd->dequant_uv;
+        zd->block[24].dequant = zd->dequant_y2;
+#endif
+
+
+        memcpy(z->rd_threshes, x->rd_threshes, sizeof(x->rd_threshes));
+        memcpy(z->rd_thresh_mult, x->rd_thresh_mult, sizeof(x->rd_thresh_mult));
+
+        z->zbin_over_quant = x->zbin_over_quant;
+        z->zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
+        z->zbin_mode_boost = x->zbin_mode_boost;
+
+        memset(z->error_bins, 0, sizeof(z->error_bins));
+    }
+}
+
+void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
+                               MACROBLOCK *x,
+                               MB_ROW_COMP *mbr_ei,
+                               int count
+                              )
+{
+
+    VP8_COMMON *const cm = & cpi->common;
+    MACROBLOCKD *const xd = & x->e_mbd;
+    int i;
+
+    for (i = 0; i < count; i++)
+    {
+        MACROBLOCK *mb = & mbr_ei[i].mb;
+        MACROBLOCKD *mbd = &mb->e_mbd;
+
+        mbd->subpixel_predict        = xd->subpixel_predict;
+        mbd->subpixel_predict8x4     = xd->subpixel_predict8x4;
+        mbd->subpixel_predict8x8     = xd->subpixel_predict8x8;
+        mbd->subpixel_predict16x16   = xd->subpixel_predict16x16;
+        mb->gf_active_ptr            = x->gf_active_ptr;
+
+        memset(mbr_ei[i].segment_counts, 0, sizeof(mbr_ei[i].segment_counts));
+        mbr_ei[i].totalrate = 0;
+
+        mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
+
+        mbd->frame_type = cm->frame_type;
+
+        mb->src = * cpi->Source;
+        mbd->pre = cm->yv12_fb[cm->lst_fb_idx];
+        mbd->dst = cm->yv12_fb[cm->new_fb_idx];
+
+        mb->src.y_buffer += 16 * x->src.y_stride * (i + 1);
+        mb->src.u_buffer +=  8 * x->src.uv_stride * (i + 1);
+        mb->src.v_buffer +=  8 * x->src.uv_stride * (i + 1);
+
+        vp8_build_block_offsets(mb);
+
+        mbd->left_context = &cm->left_context;
+        mb->mvc = cm->fc.mvc;
+
+        setup_mbby_copy(&mbr_ei[i].mb, x);
+
+        mbd->fullpixel_mask = 0xffffffff;
+        if(cm->full_pixel)
+            mbd->fullpixel_mask = 0xfffffff8;
+
+        vp8_zero(mb->coef_counts);
+        vp8_zero(x->ymode_count);
+        mb->skip_true_count = 0;
+        vp8_zero(mb->MVcount);
+        mb->prediction_error = 0;
+        mb->intra_error = 0;
+        vp8_zero(mb->count_mb_ref_frame_usage);
+        mb->mbs_tested_so_far = 0;
+        mb->mbs_zero_last_dot_suppress = 0;
+    }
+}
+
+int vp8cx_create_encoder_threads(VP8_COMP *cpi)
+{
+    const VP8_COMMON * cm = &cpi->common;
+
+    cpi->b_multi_threaded = 0;
+    cpi->encoding_thread_count = 0;
+
+    pthread_mutex_init(&cpi->mt_mutex, NULL);
+
+    if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
+    {
+        int ithread;
+        int th_count = cpi->oxcf.multi_threaded - 1;
+        int rc = 0;
+
+        /* don't allocate more threads than cores available */
+        if (cpi->oxcf.multi_threaded > cm->processor_core_count)
+            th_count = cm->processor_core_count - 1;
+
+        /* we have th_count + 1 (main) threads processing one row each */
+        /* no point to have more threads than the sync range allows */
+        if(th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1))
+        {
+            th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
+        }
+
+        if(th_count == 0)
+            return 0;
+
+        CHECK_MEM_ERROR(cpi->h_encoding_thread,
+                        vpx_malloc(sizeof(pthread_t) * th_count));
+        CHECK_MEM_ERROR(cpi->h_event_start_encoding,
+                        vpx_malloc(sizeof(sem_t) * th_count));
+        CHECK_MEM_ERROR(cpi->mb_row_ei,
+                        vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
+        memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
+        CHECK_MEM_ERROR(cpi->en_thread_data,
+                        vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
+
+        sem_init(&cpi->h_event_end_encoding, 0, 0);
+
+        cpi->b_multi_threaded = 1;
+        cpi->encoding_thread_count = th_count;
+
+        /*
+        printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
+               (cpi->encoding_thread_count +1));
+        */
+
+        for (ithread = 0; ithread < th_count; ithread++)
+        {
+            ENCODETHREAD_DATA *ethd = &cpi->en_thread_data[ithread];
+
+            /* Setup block ptrs and offsets */
+            vp8_setup_block_ptrs(&cpi->mb_row_ei[ithread].mb);
+            vp8_setup_block_dptrs(&cpi->mb_row_ei[ithread].mb.e_mbd);
+
+            sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
+
+            ethd->ithread = ithread;
+            ethd->ptr1 = (void *)cpi;
+            ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
+
+            rc = pthread_create(&cpi->h_encoding_thread[ithread], 0,
+                                thread_encoding_proc, ethd);
+            if(rc)
+                break;
+        }
+
+        if(rc)
+        {
+            /* shutdown other threads */
+            protected_write(&cpi->mt_mutex, &cpi->b_multi_threaded, 0);
+            for(--ithread; ithread >= 0; ithread--)
+            {
+                pthread_join(cpi->h_encoding_thread[ithread], 0);
+                sem_destroy(&cpi->h_event_start_encoding[ithread]);
+            }
+            sem_destroy(&cpi->h_event_end_encoding);
+
+            /* free thread related resources */
+            vpx_free(cpi->h_event_start_encoding);
+            vpx_free(cpi->h_encoding_thread);
+            vpx_free(cpi->mb_row_ei);
+            vpx_free(cpi->en_thread_data);
+
+            pthread_mutex_destroy(&cpi->mt_mutex);
+
+            return -1;
+        }
+
+
+        {
+            LPFTHREAD_DATA * lpfthd = &cpi->lpf_thread_data;
+
+            sem_init(&cpi->h_event_start_lpf, 0, 0);
+            sem_init(&cpi->h_event_end_lpf, 0, 0);
+
+            lpfthd->ptr1 = (void *)cpi;
+            rc = pthread_create(&cpi->h_filter_thread, 0, thread_loopfilter,
+                                lpfthd);
+
+            if(rc)
+            {
+                /* shutdown other threads */
+                protected_write(&cpi->mt_mutex, &cpi->b_multi_threaded, 0);
+                for(--ithread; ithread >= 0; ithread--)
+                {
+                    sem_post(&cpi->h_event_start_encoding[ithread]);
+                    pthread_join(cpi->h_encoding_thread[ithread], 0);
+                    sem_destroy(&cpi->h_event_start_encoding[ithread]);
+                }
+                sem_destroy(&cpi->h_event_end_encoding);
+                sem_destroy(&cpi->h_event_end_lpf);
+                sem_destroy(&cpi->h_event_start_lpf);
+
+                /* free thread related resources */
+                vpx_free(cpi->h_event_start_encoding);
+                vpx_free(cpi->h_encoding_thread);
+                vpx_free(cpi->mb_row_ei);
+                vpx_free(cpi->en_thread_data);
+
+                pthread_mutex_destroy(&cpi->mt_mutex);
+
+                return -2;
+            }
+        }
+    }
+    return 0;
+}
+
+void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
+{
+    if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded))
+    {
+        /* shutdown other threads */
+        protected_write(&cpi->mt_mutex, &cpi->b_multi_threaded, 0);
+        {
+            int i;
+
+            for (i = 0; i < cpi->encoding_thread_count; i++)
+            {
+                sem_post(&cpi->h_event_start_encoding[i]);
+                pthread_join(cpi->h_encoding_thread[i], 0);
+
+                sem_destroy(&cpi->h_event_start_encoding[i]);
+            }
+
+            sem_post(&cpi->h_event_start_lpf);
+            pthread_join(cpi->h_filter_thread, 0);
+        }
+
+        sem_destroy(&cpi->h_event_end_encoding);
+        sem_destroy(&cpi->h_event_end_lpf);
+        sem_destroy(&cpi->h_event_start_lpf);
+
+        /* free thread related resources */
+        vpx_free(cpi->h_event_start_encoding);
+        vpx_free(cpi->h_encoding_thread);
+        vpx_free(cpi->mb_row_ei);
+        vpx_free(cpi->en_thread_data);
+    }
+    pthread_mutex_destroy(&cpi->mt_mutex);
+}
+#endif
diff --git a/libvpx/libvpx/vp8/encoder/firstpass.c b/libvpx/libvpx/vp8/encoder/firstpass.c
new file mode 100644
index 0000000..c526a3e
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/firstpass.c
@@ -0,0 +1,3372 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <limits.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+#include "block.h"
+#include "onyx_int.h"
+#include "vpx_dsp/variance.h"
+#include "encodeintra.h"
+#include "vp8/common/common.h"
+#include "vp8/common/setupintrarecon.h"
+#include "vp8/common/systemdependent.h"
+#include "mcomp.h"
+#include "firstpass.h"
+#include "vpx_scale/vpx_scale.h"
+#include "encodemb.h"
+#include "vp8/common/extend.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp8/common/swapyv12buffer.h"
+#include "rdopt.h"
+#include "vp8/common/quant_common.h"
+#include "encodemv.h"
+#include "encodeframe.h"
+
+/* #define OUTPUT_FPF 1 */
+
+extern void vp8cx_frame_init_quantizer(VP8_COMP *cpi);
+
+#define GFQ_ADJUSTMENT vp8_gf_boost_qadjustment[Q]
+extern int vp8_kf_boost_qadjustment[QINDEX_RANGE];
+
+extern const int vp8_gf_boost_qadjustment[QINDEX_RANGE];
+
+#define IIFACTOR   1.5
+#define IIKFACTOR1 1.40
+#define IIKFACTOR2 1.5
+#define RMAX       14.0
+#define GF_RMAX    48.0
+
+#define KF_MB_INTRA_MIN 300
+#define GF_MB_INTRA_MIN 200
+
+#define DOUBLE_DIVIDE_CHECK(X) ((X)<0?(X)-.000001:(X)+.000001)
+
+#define POW1 (double)cpi->oxcf.two_pass_vbrbias/100.0
+#define POW2 (double)cpi->oxcf.two_pass_vbrbias/100.0
+
+#define NEW_BOOST 1
+
+static int vscale_lookup[7] = {0, 1, 1, 2, 2, 3, 3};
+static int hscale_lookup[7] = {0, 0, 1, 1, 2, 2, 3};
+
+
+static const int cq_level[QINDEX_RANGE] =
+{
+    0,0,1,1,2,3,3,4,4,5,6,6,7,8,8,9,
+    9,10,11,11,12,13,13,14,15,15,16,17,17,18,19,20,
+    20,21,22,22,23,24,24,25,26,27,27,28,29,30,30,31,
+    32,33,33,34,35,36,36,37,38,39,39,40,41,42,42,43,
+    44,45,46,46,47,48,49,50,50,51,52,53,54,55,55,56,
+    57,58,59,60,60,61,62,63,64,65,66,67,67,68,69,70,
+    71,72,73,74,75,75,76,77,78,79,80,81,82,83,84,85,
+    86,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100
+};
+
+static void find_next_key_frame(VP8_COMP *cpi, FIRSTPASS_STATS *this_frame);
+
+/* Resets the first pass file to the given position using a relative seek
+ * from the current position
+ */
+static void reset_fpf_position(VP8_COMP *cpi, FIRSTPASS_STATS *Position)
+{
+    cpi->twopass.stats_in = Position;
+}
+
+static int lookup_next_frame_stats(VP8_COMP *cpi, FIRSTPASS_STATS *next_frame)
+{
+    if (cpi->twopass.stats_in >= cpi->twopass.stats_in_end)
+        return EOF;
+
+    *next_frame = *cpi->twopass.stats_in;
+    return 1;
+}
+
+/* Read frame stats at an offset from the current position */
+static int read_frame_stats( VP8_COMP *cpi,
+                             FIRSTPASS_STATS *frame_stats,
+                             int offset )
+{
+    FIRSTPASS_STATS * fps_ptr = cpi->twopass.stats_in;
+
+    /* Check legality of offset */
+    if ( offset >= 0 )
+    {
+        if ( &fps_ptr[offset] >= cpi->twopass.stats_in_end )
+             return EOF;
+    }
+    else if ( offset < 0 )
+    {
+        if ( &fps_ptr[offset] < cpi->twopass.stats_in_start )
+             return EOF;
+    }
+
+    *frame_stats = fps_ptr[offset];
+    return 1;
+}
+
+static int input_stats(VP8_COMP *cpi, FIRSTPASS_STATS *fps)
+{
+    if (cpi->twopass.stats_in >= cpi->twopass.stats_in_end)
+        return EOF;
+
+    *fps = *cpi->twopass.stats_in;
+    cpi->twopass.stats_in =
+         (void*)((char *)cpi->twopass.stats_in + sizeof(FIRSTPASS_STATS));
+    return 1;
+}
+
+static void output_stats(const VP8_COMP            *cpi,
+                         struct vpx_codec_pkt_list *pktlist,
+                         FIRSTPASS_STATS            *stats)
+{
+    struct vpx_codec_cx_pkt pkt;
+    (void)cpi;
+    pkt.kind = VPX_CODEC_STATS_PKT;
+    pkt.data.twopass_stats.buf = stats;
+    pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
+    vpx_codec_pkt_list_add(pktlist, &pkt);
+
+/* TEMP debug code */
+#if OUTPUT_FPF
+
+    {
+        FILE *fpfile;
+        fpfile = fopen("firstpass.stt", "a");
+
+        fprintf(fpfile, "%12.0f %12.0f %12.0f %12.4f %12.4f %12.4f %12.4f"
+                " %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f"
+                " %12.0f %12.0f %12.4f\n",
+                stats->frame,
+                stats->intra_error,
+                stats->coded_error,
+                stats->ssim_weighted_pred_err,
+                stats->pcnt_inter,
+                stats->pcnt_motion,
+                stats->pcnt_second_ref,
+                stats->pcnt_neutral,
+                stats->MVr,
+                stats->mvr_abs,
+                stats->MVc,
+                stats->mvc_abs,
+                stats->MVrv,
+                stats->MVcv,
+                stats->mv_in_out_count,
+                stats->new_mv_count,
+                stats->count,
+                stats->duration);
+        fclose(fpfile);
+    }
+#endif
+}
+
+static void zero_stats(FIRSTPASS_STATS *section)
+{
+    section->frame      = 0.0;
+    section->intra_error = 0.0;
+    section->coded_error = 0.0;
+    section->ssim_weighted_pred_err = 0.0;
+    section->pcnt_inter  = 0.0;
+    section->pcnt_motion  = 0.0;
+    section->pcnt_second_ref = 0.0;
+    section->pcnt_neutral = 0.0;
+    section->MVr        = 0.0;
+    section->mvr_abs     = 0.0;
+    section->MVc        = 0.0;
+    section->mvc_abs     = 0.0;
+    section->MVrv       = 0.0;
+    section->MVcv       = 0.0;
+    section->mv_in_out_count  = 0.0;
+    section->new_mv_count = 0.0;
+    section->count      = 0.0;
+    section->duration   = 1.0;
+}
+
+static void accumulate_stats(FIRSTPASS_STATS *section, FIRSTPASS_STATS *frame)
+{
+    section->frame += frame->frame;
+    section->intra_error += frame->intra_error;
+    section->coded_error += frame->coded_error;
+    section->ssim_weighted_pred_err += frame->ssim_weighted_pred_err;
+    section->pcnt_inter  += frame->pcnt_inter;
+    section->pcnt_motion += frame->pcnt_motion;
+    section->pcnt_second_ref += frame->pcnt_second_ref;
+    section->pcnt_neutral += frame->pcnt_neutral;
+    section->MVr        += frame->MVr;
+    section->mvr_abs     += frame->mvr_abs;
+    section->MVc        += frame->MVc;
+    section->mvc_abs     += frame->mvc_abs;
+    section->MVrv       += frame->MVrv;
+    section->MVcv       += frame->MVcv;
+    section->mv_in_out_count  += frame->mv_in_out_count;
+    section->new_mv_count += frame->new_mv_count;
+    section->count      += frame->count;
+    section->duration   += frame->duration;
+}
+
+static void subtract_stats(FIRSTPASS_STATS *section, FIRSTPASS_STATS *frame)
+{
+    section->frame -= frame->frame;
+    section->intra_error -= frame->intra_error;
+    section->coded_error -= frame->coded_error;
+    section->ssim_weighted_pred_err -= frame->ssim_weighted_pred_err;
+    section->pcnt_inter  -= frame->pcnt_inter;
+    section->pcnt_motion -= frame->pcnt_motion;
+    section->pcnt_second_ref -= frame->pcnt_second_ref;
+    section->pcnt_neutral -= frame->pcnt_neutral;
+    section->MVr        -= frame->MVr;
+    section->mvr_abs     -= frame->mvr_abs;
+    section->MVc        -= frame->MVc;
+    section->mvc_abs     -= frame->mvc_abs;
+    section->MVrv       -= frame->MVrv;
+    section->MVcv       -= frame->MVcv;
+    section->mv_in_out_count  -= frame->mv_in_out_count;
+    section->new_mv_count -= frame->new_mv_count;
+    section->count      -= frame->count;
+    section->duration   -= frame->duration;
+}
+
+static void avg_stats(FIRSTPASS_STATS *section)
+{
+    if (section->count < 1.0)
+        return;
+
+    section->intra_error /= section->count;
+    section->coded_error /= section->count;
+    section->ssim_weighted_pred_err /= section->count;
+    section->pcnt_inter  /= section->count;
+    section->pcnt_second_ref /= section->count;
+    section->pcnt_neutral /= section->count;
+    section->pcnt_motion /= section->count;
+    section->MVr        /= section->count;
+    section->mvr_abs     /= section->count;
+    section->MVc        /= section->count;
+    section->mvc_abs     /= section->count;
+    section->MVrv       /= section->count;
+    section->MVcv       /= section->count;
+    section->mv_in_out_count   /= section->count;
+    section->duration   /= section->count;
+}
+
+/* Calculate a modified Error used in distributing bits between easier
+ * and harder frames
+ */
+static double calculate_modified_err(VP8_COMP *cpi, FIRSTPASS_STATS *this_frame)
+{
+    double av_err = ( cpi->twopass.total_stats.ssim_weighted_pred_err /
+                      cpi->twopass.total_stats.count );
+    double this_err = this_frame->ssim_weighted_pred_err;
+    double modified_err;
+
+    if (this_err > av_err)
+        modified_err = av_err * pow((this_err / DOUBLE_DIVIDE_CHECK(av_err)), POW1);
+    else
+        modified_err = av_err * pow((this_err / DOUBLE_DIVIDE_CHECK(av_err)), POW2);
+
+    return modified_err;
+}
+
+static const double weight_table[256] = {
+0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
+0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
+0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
+0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
+0.020000, 0.031250, 0.062500, 0.093750, 0.125000, 0.156250, 0.187500, 0.218750,
+0.250000, 0.281250, 0.312500, 0.343750, 0.375000, 0.406250, 0.437500, 0.468750,
+0.500000, 0.531250, 0.562500, 0.593750, 0.625000, 0.656250, 0.687500, 0.718750,
+0.750000, 0.781250, 0.812500, 0.843750, 0.875000, 0.906250, 0.937500, 0.968750,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000
+};
+
+static double simple_weight(YV12_BUFFER_CONFIG *source)
+{
+    int i, j;
+
+    unsigned char *src = source->y_buffer;
+    double sum_weights = 0.0;
+
+    /* Loop throught the Y plane raw examining levels and creating a weight
+     * for the image
+     */
+    i = source->y_height;
+    do
+    {
+        j = source->y_width;
+        do
+        {
+            sum_weights += weight_table[ *src];
+            src++;
+        }while(--j);
+        src -= source->y_width;
+        src += source->y_stride;
+    }while(--i);
+
+    sum_weights /= (source->y_height * source->y_width);
+
+    return sum_weights;
+}
+
+
+/* This function returns the current per frame maximum bitrate target */
+static int frame_max_bits(VP8_COMP *cpi)
+{
+    /* Max allocation for a single frame based on the max section guidelines
+     * passed in and how many bits are left
+     */
+    int max_bits;
+
+    /* For CBR we need to also consider buffer fullness.
+     * If we are running below the optimal level then we need to gradually
+     * tighten up on max_bits.
+     */
+    if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+    {
+        double buffer_fullness_ratio = (double)cpi->buffer_level / DOUBLE_DIVIDE_CHECK((double)cpi->oxcf.optimal_buffer_level);
+
+        /* For CBR base this on the target average bits per frame plus the
+         * maximum sedction rate passed in by the user
+         */
+        max_bits = (int)(cpi->av_per_frame_bandwidth * ((double)cpi->oxcf.two_pass_vbrmax_section / 100.0));
+
+        /* If our buffer is below the optimum level */
+        if (buffer_fullness_ratio < 1.0)
+        {
+            /* The lower of max_bits / 4 or cpi->av_per_frame_bandwidth / 4. */
+            int min_max_bits = ((cpi->av_per_frame_bandwidth >> 2) < (max_bits >> 2)) ? cpi->av_per_frame_bandwidth >> 2 : max_bits >> 2;
+
+            max_bits = (int)(max_bits * buffer_fullness_ratio);
+
+            /* Lowest value we will set ... which should allow the buffer to
+             * refill.
+             */
+            if (max_bits < min_max_bits)
+                max_bits = min_max_bits;
+        }
+    }
+    /* VBR */
+    else
+    {
+        /* For VBR base this on the bits and frames left plus the
+         * two_pass_vbrmax_section rate passed in by the user
+         */
+        max_bits = (int)(((double)cpi->twopass.bits_left / (cpi->twopass.total_stats.count - (double)cpi->common.current_video_frame)) * ((double)cpi->oxcf.two_pass_vbrmax_section / 100.0));
+    }
+
+    /* Trap case where we are out of bits */
+    if (max_bits < 0)
+        max_bits = 0;
+
+    return max_bits;
+}
+
+void vp8_init_first_pass(VP8_COMP *cpi)
+{
+    zero_stats(&cpi->twopass.total_stats);
+}
+
+void vp8_end_first_pass(VP8_COMP *cpi)
+{
+    output_stats(cpi, cpi->output_pkt_list, &cpi->twopass.total_stats);
+}
+
+static void zz_motion_search( VP8_COMP *cpi, MACROBLOCK * x,
+                              YV12_BUFFER_CONFIG * raw_buffer,
+                              int * raw_motion_err,
+                              YV12_BUFFER_CONFIG * recon_buffer,
+                              int * best_motion_err, int recon_yoffset)
+{
+    MACROBLOCKD * const xd = & x->e_mbd;
+    BLOCK *b = &x->block[0];
+    BLOCKD *d = &x->e_mbd.block[0];
+
+    unsigned char *src_ptr = (*(b->base_src) + b->src);
+    int src_stride = b->src_stride;
+    unsigned char *raw_ptr;
+    int raw_stride = raw_buffer->y_stride;
+    unsigned char *ref_ptr;
+    int ref_stride = x->e_mbd.pre.y_stride;
+    (void)cpi;
+
+    /* Set up pointers for this macro block raw buffer */
+    raw_ptr = (unsigned char *)(raw_buffer->y_buffer + recon_yoffset
+                                + d->offset);
+    vpx_mse16x16(src_ptr, src_stride, raw_ptr, raw_stride,
+                 (unsigned int *)(raw_motion_err));
+
+    /* Set up pointers for this macro block recon buffer */
+    xd->pre.y_buffer = recon_buffer->y_buffer + recon_yoffset;
+    ref_ptr = (unsigned char *)(xd->pre.y_buffer + d->offset );
+    vpx_mse16x16(src_ptr, src_stride, ref_ptr, ref_stride,
+                 (unsigned int *)(best_motion_err));
+}
+
+static void first_pass_motion_search(VP8_COMP *cpi, MACROBLOCK *x,
+                                     int_mv *ref_mv, MV *best_mv,
+                                     YV12_BUFFER_CONFIG *recon_buffer,
+                                     int *best_motion_err, int recon_yoffset )
+{
+    MACROBLOCKD *const xd = & x->e_mbd;
+    BLOCK *b = &x->block[0];
+    BLOCKD *d = &x->e_mbd.block[0];
+    int num00;
+
+    int_mv tmp_mv;
+    int_mv ref_mv_full;
+
+    int tmp_err;
+    int step_param = 3; /* Dont search over full range for first pass */
+    int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
+    int n;
+    vp8_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
+    int new_mv_mode_penalty = 256;
+
+    /* override the default variance function to use MSE */
+    v_fn_ptr.vf    = vpx_mse16x16;
+
+    /* Set up pointers for this macro block recon buffer */
+    xd->pre.y_buffer = recon_buffer->y_buffer + recon_yoffset;
+
+    /* Initial step/diamond search centred on best mv */
+    tmp_mv.as_int = 0;
+    ref_mv_full.as_mv.col = ref_mv->as_mv.col>>3;
+    ref_mv_full.as_mv.row = ref_mv->as_mv.row>>3;
+    tmp_err = cpi->diamond_search_sad(x, b, d, &ref_mv_full, &tmp_mv, step_param,
+                                      x->sadperbit16, &num00, &v_fn_ptr,
+                                      x->mvcost, ref_mv);
+    if ( tmp_err < INT_MAX-new_mv_mode_penalty )
+        tmp_err += new_mv_mode_penalty;
+
+    if (tmp_err < *best_motion_err)
+    {
+        *best_motion_err = tmp_err;
+        best_mv->row = tmp_mv.as_mv.row;
+        best_mv->col = tmp_mv.as_mv.col;
+    }
+
+    /* Further step/diamond searches as necessary */
+    n = num00;
+    num00 = 0;
+
+    while (n < further_steps)
+    {
+        n++;
+
+        if (num00)
+            num00--;
+        else
+        {
+            tmp_err = cpi->diamond_search_sad(x, b, d, &ref_mv_full, &tmp_mv,
+                                              step_param + n, x->sadperbit16,
+                                              &num00, &v_fn_ptr, x->mvcost,
+                                              ref_mv);
+            if ( tmp_err < INT_MAX-new_mv_mode_penalty )
+                tmp_err += new_mv_mode_penalty;
+
+            if (tmp_err < *best_motion_err)
+            {
+                *best_motion_err = tmp_err;
+                best_mv->row = tmp_mv.as_mv.row;
+                best_mv->col = tmp_mv.as_mv.col;
+            }
+        }
+    }
+}
+
+void vp8_first_pass(VP8_COMP *cpi)
+{
+    int mb_row, mb_col;
+    MACROBLOCK *const x = & cpi->mb;
+    VP8_COMMON *const cm = & cpi->common;
+    MACROBLOCKD *const xd = & x->e_mbd;
+
+    int recon_yoffset, recon_uvoffset;
+    YV12_BUFFER_CONFIG *lst_yv12 = &cm->yv12_fb[cm->lst_fb_idx];
+    YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
+    YV12_BUFFER_CONFIG *gld_yv12 = &cm->yv12_fb[cm->gld_fb_idx];
+    int recon_y_stride = lst_yv12->y_stride;
+    int recon_uv_stride = lst_yv12->uv_stride;
+    int64_t intra_error = 0;
+    int64_t coded_error = 0;
+
+    int sum_mvr = 0, sum_mvc = 0;
+    int sum_mvr_abs = 0, sum_mvc_abs = 0;
+    int sum_mvrs = 0, sum_mvcs = 0;
+    int mvcount = 0;
+    int intercount = 0;
+    int second_ref_count = 0;
+    int intrapenalty = 256;
+    int neutral_count = 0;
+    int new_mv_count = 0;
+    int sum_in_vectors = 0;
+    uint32_t lastmv_as_int = 0;
+
+    int_mv zero_ref_mv;
+
+    zero_ref_mv.as_int = 0;
+
+    vp8_clear_system_state();
+
+    x->src = * cpi->Source;
+    xd->pre = *lst_yv12;
+    xd->dst = *new_yv12;
+
+    x->partition_info = x->pi;
+
+    xd->mode_info_context = cm->mi;
+
+    if(!cm->use_bilinear_mc_filter)
+    {
+         xd->subpixel_predict        = vp8_sixtap_predict4x4;
+         xd->subpixel_predict8x4     = vp8_sixtap_predict8x4;
+         xd->subpixel_predict8x8     = vp8_sixtap_predict8x8;
+         xd->subpixel_predict16x16   = vp8_sixtap_predict16x16;
+     }
+     else
+     {
+         xd->subpixel_predict        = vp8_bilinear_predict4x4;
+         xd->subpixel_predict8x4     = vp8_bilinear_predict8x4;
+         xd->subpixel_predict8x8     = vp8_bilinear_predict8x8;
+         xd->subpixel_predict16x16   = vp8_bilinear_predict16x16;
+     }
+
+    vp8_build_block_offsets(x);
+
+    /* set up frame new frame for intra coded blocks */
+    vp8_setup_intra_recon(new_yv12);
+    vp8cx_frame_init_quantizer(cpi);
+
+    /* Initialise the MV cost table to the defaults */
+    {
+        int flag[2] = {1, 1};
+        vp8_initialize_rd_consts(cpi, x, vp8_dc_quant(cm->base_qindex, cm->y1dc_delta_q));
+        memcpy(cm->fc.mvc, vp8_default_mv_context, sizeof(vp8_default_mv_context));
+        vp8_build_component_cost_table(cpi->mb.mvcost, (const MV_CONTEXT *) cm->fc.mvc, flag);
+    }
+
+    /* for each macroblock row in image */
+    for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
+    {
+        int_mv best_ref_mv;
+
+        best_ref_mv.as_int = 0;
+
+        /* reset above block coeffs */
+        xd->up_available = (mb_row != 0);
+        recon_yoffset = (mb_row * recon_y_stride * 16);
+        recon_uvoffset = (mb_row * recon_uv_stride * 8);
+
+        /* Set up limit values for motion vectors to prevent them extending
+         * outside the UMV borders
+         */
+        x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
+        x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
+
+
+        /* for each macroblock col in image */
+        for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
+        {
+            int this_error;
+            int gf_motion_error = INT_MAX;
+            int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
+
+            xd->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
+            xd->dst.u_buffer = new_yv12->u_buffer + recon_uvoffset;
+            xd->dst.v_buffer = new_yv12->v_buffer + recon_uvoffset;
+            xd->left_available = (mb_col != 0);
+
+            /* Copy current mb to a buffer */
+            vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
+
+            /* do intra 16x16 prediction */
+            this_error = vp8_encode_intra(cpi, x, use_dc_pred);
+
+            /* "intrapenalty" below deals with situations where the intra
+             * and inter error scores are very low (eg a plain black frame)
+             * We do not have special cases in first pass for 0,0 and
+             * nearest etc so all inter modes carry an overhead cost
+             * estimate fot the mv. When the error score is very low this
+             * causes us to pick all or lots of INTRA modes and throw lots
+             * of key frames. This penalty adds a cost matching that of a
+             * 0,0 mv to the intra case.
+             */
+            this_error += intrapenalty;
+
+            /* Cumulative intra error total */
+            intra_error += (int64_t)this_error;
+
+            /* Set up limit values for motion vectors to prevent them
+             * extending outside the UMV borders
+             */
+            x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
+            x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
+
+            /* Other than for the first frame do a motion search */
+            if (cm->current_video_frame > 0)
+            {
+                BLOCKD *d = &x->e_mbd.block[0];
+                MV tmp_mv = {0, 0};
+                int tmp_err;
+                int motion_error = INT_MAX;
+                int raw_motion_error = INT_MAX;
+
+                /* Simple 0,0 motion with no mv overhead */
+                zz_motion_search( cpi, x, cpi->last_frame_unscaled_source,
+                                  &raw_motion_error, lst_yv12, &motion_error,
+                                  recon_yoffset );
+                d->bmi.mv.as_mv.row = 0;
+                d->bmi.mv.as_mv.col = 0;
+
+                if (raw_motion_error < cpi->oxcf.encode_breakout)
+                    goto skip_motion_search;
+
+                /* Test last reference frame using the previous best mv as the
+                 * starting point (best reference) for the search
+                 */
+                first_pass_motion_search(cpi, x, &best_ref_mv,
+                                        &d->bmi.mv.as_mv, lst_yv12,
+                                        &motion_error, recon_yoffset);
+
+                /* If the current best reference mv is not centred on 0,0
+                 * then do a 0,0 based search as well
+                 */
+                if (best_ref_mv.as_int)
+                {
+                   tmp_err = INT_MAX;
+                   first_pass_motion_search(cpi, x, &zero_ref_mv, &tmp_mv,
+                                     lst_yv12, &tmp_err, recon_yoffset);
+
+                   if ( tmp_err < motion_error )
+                   {
+                        motion_error = tmp_err;
+                        d->bmi.mv.as_mv.row = tmp_mv.row;
+                        d->bmi.mv.as_mv.col = tmp_mv.col;
+                   }
+                }
+
+                /* Experimental search in a second reference frame ((0,0)
+                 * based only)
+                 */
+                if (cm->current_video_frame > 1)
+                {
+                    first_pass_motion_search(cpi, x, &zero_ref_mv, &tmp_mv, gld_yv12, &gf_motion_error, recon_yoffset);
+
+                    if ((gf_motion_error < motion_error) && (gf_motion_error < this_error))
+                    {
+                        second_ref_count++;
+                    }
+
+                    /* Reset to last frame as reference buffer */
+                    xd->pre.y_buffer = lst_yv12->y_buffer + recon_yoffset;
+                    xd->pre.u_buffer = lst_yv12->u_buffer + recon_uvoffset;
+                    xd->pre.v_buffer = lst_yv12->v_buffer + recon_uvoffset;
+                }
+
+skip_motion_search:
+                /* Intra assumed best */
+                best_ref_mv.as_int = 0;
+
+                if (motion_error <= this_error)
+                {
+                    /* Keep a count of cases where the inter and intra were
+                     * very close and very low. This helps with scene cut
+                     * detection for example in cropped clips with black bars
+                     * at the sides or top and bottom.
+                     */
+                    if( (((this_error-intrapenalty) * 9) <=
+                         (motion_error*10)) &&
+                        (this_error < (2*intrapenalty)) )
+                    {
+                        neutral_count++;
+                    }
+
+                    d->bmi.mv.as_mv.row *= 8;
+                    d->bmi.mv.as_mv.col *= 8;
+                    this_error = motion_error;
+                    vp8_set_mbmode_and_mvs(x, NEWMV, &d->bmi.mv);
+                    vp8_encode_inter16x16y(x);
+                    sum_mvr += d->bmi.mv.as_mv.row;
+                    sum_mvr_abs += abs(d->bmi.mv.as_mv.row);
+                    sum_mvc += d->bmi.mv.as_mv.col;
+                    sum_mvc_abs += abs(d->bmi.mv.as_mv.col);
+                    sum_mvrs += d->bmi.mv.as_mv.row * d->bmi.mv.as_mv.row;
+                    sum_mvcs += d->bmi.mv.as_mv.col * d->bmi.mv.as_mv.col;
+                    intercount++;
+
+                    best_ref_mv.as_int = d->bmi.mv.as_int;
+
+                    /* Was the vector non-zero */
+                    if (d->bmi.mv.as_int)
+                    {
+                        mvcount++;
+
+                        /* Was it different from the last non zero vector */
+                        if ( d->bmi.mv.as_int != lastmv_as_int )
+                            new_mv_count++;
+                        lastmv_as_int = d->bmi.mv.as_int;
+
+                        /* Does the Row vector point inwards or outwards */
+                        if (mb_row < cm->mb_rows / 2)
+                        {
+                            if (d->bmi.mv.as_mv.row > 0)
+                                sum_in_vectors--;
+                            else if (d->bmi.mv.as_mv.row < 0)
+                                sum_in_vectors++;
+                        }
+                        else if (mb_row > cm->mb_rows / 2)
+                        {
+                            if (d->bmi.mv.as_mv.row > 0)
+                                sum_in_vectors++;
+                            else if (d->bmi.mv.as_mv.row < 0)
+                                sum_in_vectors--;
+                        }
+
+                        /* Does the Row vector point inwards or outwards */
+                        if (mb_col < cm->mb_cols / 2)
+                        {
+                            if (d->bmi.mv.as_mv.col > 0)
+                                sum_in_vectors--;
+                            else if (d->bmi.mv.as_mv.col < 0)
+                                sum_in_vectors++;
+                        }
+                        else if (mb_col > cm->mb_cols / 2)
+                        {
+                            if (d->bmi.mv.as_mv.col > 0)
+                                sum_in_vectors++;
+                            else if (d->bmi.mv.as_mv.col < 0)
+                                sum_in_vectors--;
+                        }
+                    }
+                }
+            }
+
+            coded_error += (int64_t)this_error;
+
+            /* adjust to the next column of macroblocks */
+            x->src.y_buffer += 16;
+            x->src.u_buffer += 8;
+            x->src.v_buffer += 8;
+
+            recon_yoffset += 16;
+            recon_uvoffset += 8;
+        }
+
+        /* adjust to the next row of mbs */
+        x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
+        x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
+        x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
+
+        /* extend the recon for intra prediction */
+        vp8_extend_mb_row(new_yv12, xd->dst.y_buffer + 16, xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
+        vp8_clear_system_state();
+    }
+
+    vp8_clear_system_state();
+    {
+        double weight = 0.0;
+
+        FIRSTPASS_STATS fps;
+
+        fps.frame      = cm->current_video_frame ;
+        fps.intra_error = (double)(intra_error >> 8);
+        fps.coded_error = (double)(coded_error >> 8);
+        weight = simple_weight(cpi->Source);
+
+
+        if (weight < 0.1)
+            weight = 0.1;
+
+        fps.ssim_weighted_pred_err = fps.coded_error * weight;
+
+        fps.pcnt_inter  = 0.0;
+        fps.pcnt_motion = 0.0;
+        fps.MVr        = 0.0;
+        fps.mvr_abs     = 0.0;
+        fps.MVc        = 0.0;
+        fps.mvc_abs     = 0.0;
+        fps.MVrv       = 0.0;
+        fps.MVcv       = 0.0;
+        fps.mv_in_out_count  = 0.0;
+        fps.new_mv_count = 0.0;
+        fps.count      = 1.0;
+
+        fps.pcnt_inter   = 1.0 * (double)intercount / cm->MBs;
+        fps.pcnt_second_ref = 1.0 * (double)second_ref_count / cm->MBs;
+        fps.pcnt_neutral = 1.0 * (double)neutral_count / cm->MBs;
+
+        if (mvcount > 0)
+        {
+            fps.MVr = (double)sum_mvr / (double)mvcount;
+            fps.mvr_abs = (double)sum_mvr_abs / (double)mvcount;
+            fps.MVc = (double)sum_mvc / (double)mvcount;
+            fps.mvc_abs = (double)sum_mvc_abs / (double)mvcount;
+            fps.MVrv = ((double)sum_mvrs - (fps.MVr * fps.MVr / (double)mvcount)) / (double)mvcount;
+            fps.MVcv = ((double)sum_mvcs - (fps.MVc * fps.MVc / (double)mvcount)) / (double)mvcount;
+            fps.mv_in_out_count = (double)sum_in_vectors / (double)(mvcount * 2);
+            fps.new_mv_count = new_mv_count;
+
+            fps.pcnt_motion = 1.0 * (double)mvcount / cpi->common.MBs;
+        }
+
+        /* TODO:  handle the case when duration is set to 0, or something less
+         * than the full time between subsequent cpi->source_time_stamps
+         */
+        fps.duration = (double)(cpi->source->ts_end
+                       - cpi->source->ts_start);
+
+        /* don't want to do output stats with a stack variable! */
+        memcpy(&cpi->twopass.this_frame_stats,
+               &fps,
+               sizeof(FIRSTPASS_STATS));
+        output_stats(cpi, cpi->output_pkt_list, &cpi->twopass.this_frame_stats);
+        accumulate_stats(&cpi->twopass.total_stats, &fps);
+    }
+
+    /* Copy the previous Last Frame into the GF buffer if specific
+     * conditions for doing so are met
+     */
+    if ((cm->current_video_frame > 0) &&
+        (cpi->twopass.this_frame_stats.pcnt_inter > 0.20) &&
+        ((cpi->twopass.this_frame_stats.intra_error /
+          DOUBLE_DIVIDE_CHECK(cpi->twopass.this_frame_stats.coded_error)) >
+         2.0))
+    {
+        vp8_yv12_copy_frame(lst_yv12, gld_yv12);
+    }
+
+    /* swap frame pointers so last frame refers to the frame we just
+     * compressed
+     */
+    vp8_swap_yv12_buffer(lst_yv12, new_yv12);
+    vp8_yv12_extend_frame_borders(lst_yv12);
+
+    /* Special case for the first frame. Copy into the GF buffer as a
+     * second reference.
+     */
+    if (cm->current_video_frame == 0)
+    {
+        vp8_yv12_copy_frame(lst_yv12, gld_yv12);
+    }
+
+
+    /* use this to see what the first pass reconstruction looks like */
+    if (0)
+    {
+        char filename[512];
+        FILE *recon_file;
+        sprintf(filename, "enc%04d.yuv", (int) cm->current_video_frame);
+
+        if (cm->current_video_frame == 0)
+            recon_file = fopen(filename, "wb");
+        else
+            recon_file = fopen(filename, "ab");
+
+        (void) fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1,
+                      recon_file);
+        fclose(recon_file);
+    }
+
+    cm->current_video_frame++;
+
+}
+extern const int vp8_bits_per_mb[2][QINDEX_RANGE];
+
+/* Estimate a cost per mb attributable to overheads such as the coding of
+ * modes and motion vectors.
+ * Currently simplistic in its assumptions for testing.
+ */
+
+static double bitcost( double prob )
+{
+  if (prob > 0.000122)
+    return -log(prob) / log(2.0);
+  else
+    return 13.0;
+}
+static int64_t estimate_modemvcost(VP8_COMP *cpi,
+                                     FIRSTPASS_STATS * fpstats)
+{
+    int mv_cost;
+    int64_t mode_cost;
+
+    double av_pct_inter = fpstats->pcnt_inter / fpstats->count;
+    double av_pct_motion = fpstats->pcnt_motion / fpstats->count;
+    double av_intra = (1.0 - av_pct_inter);
+
+    double zz_cost;
+    double motion_cost;
+    double intra_cost;
+
+    zz_cost = bitcost(av_pct_inter - av_pct_motion);
+    motion_cost = bitcost(av_pct_motion);
+    intra_cost = bitcost(av_intra);
+
+    /* Estimate of extra bits per mv overhead for mbs
+     * << 9 is the normalization to the (bits * 512) used in vp8_bits_per_mb
+     */
+    mv_cost = ((int)(fpstats->new_mv_count / fpstats->count) * 8) << 9;
+
+    /* Crude estimate of overhead cost from modes
+     * << 9 is the normalization to (bits * 512) used in vp8_bits_per_mb
+     */
+    mode_cost = (int64_t)((((av_pct_inter - av_pct_motion) * zz_cost) +
+                             (av_pct_motion * motion_cost) +
+                             (av_intra * intra_cost)) * cpi->common.MBs) * 512;
+
+    return mv_cost + mode_cost;
+}
+
+static double calc_correction_factor( double err_per_mb,
+                                      double err_devisor,
+                                      double pt_low,
+                                      double pt_high,
+                                      int Q )
+{
+    double power_term;
+    double error_term = err_per_mb / err_devisor;
+    double correction_factor;
+
+    /* Adjustment based on Q to power term. */
+    power_term = pt_low + (Q * 0.01);
+    power_term = (power_term > pt_high) ? pt_high : power_term;
+
+    /* Adjustments to error term */
+    /* TBD */
+
+    /* Calculate correction factor */
+    correction_factor = pow(error_term, power_term);
+
+    /* Clip range */
+    correction_factor =
+        (correction_factor < 0.05)
+            ? 0.05 : (correction_factor > 5.0) ? 5.0 : correction_factor;
+
+    return correction_factor;
+}
+
+static int estimate_max_q(VP8_COMP *cpi,
+                          FIRSTPASS_STATS * fpstats,
+                          int section_target_bandwitdh,
+                          int overhead_bits )
+{
+    int Q;
+    int num_mbs = cpi->common.MBs;
+    int target_norm_bits_per_mb;
+
+    double section_err = (fpstats->coded_error / fpstats->count);
+    double err_per_mb = section_err / num_mbs;
+    double err_correction_factor;
+    double speed_correction = 1.0;
+    int overhead_bits_per_mb;
+
+    if (section_target_bandwitdh <= 0)
+        return cpi->twopass.maxq_max_limit;       /* Highest value allowed */
+
+    target_norm_bits_per_mb =
+        (section_target_bandwitdh < (1 << 20))
+            ? (512 * section_target_bandwitdh) / num_mbs
+            : 512 * (section_target_bandwitdh / num_mbs);
+
+    /* Calculate a corrective factor based on a rolling ratio of bits spent
+     * vs target bits
+     */
+    if ((cpi->rolling_target_bits > 0) &&
+        (cpi->active_worst_quality < cpi->worst_quality))
+    {
+        double rolling_ratio;
+
+        rolling_ratio = (double)cpi->rolling_actual_bits /
+                        (double)cpi->rolling_target_bits;
+
+        if (rolling_ratio < 0.95)
+            cpi->twopass.est_max_qcorrection_factor -= 0.005;
+        else if (rolling_ratio > 1.05)
+            cpi->twopass.est_max_qcorrection_factor += 0.005;
+
+        cpi->twopass.est_max_qcorrection_factor =
+            (cpi->twopass.est_max_qcorrection_factor < 0.1)
+                ? 0.1
+                : (cpi->twopass.est_max_qcorrection_factor > 10.0)
+                    ? 10.0 : cpi->twopass.est_max_qcorrection_factor;
+    }
+
+    /* Corrections for higher compression speed settings
+     * (reduced compression expected)
+     */
+    if ((cpi->compressor_speed == 3) || (cpi->compressor_speed == 1))
+    {
+        if (cpi->oxcf.cpu_used <= 5)
+            speed_correction = 1.04 + (cpi->oxcf.cpu_used * 0.04);
+        else
+            speed_correction = 1.25;
+    }
+
+    /* Estimate of overhead bits per mb */
+    /* Correction to overhead bits for min allowed Q. */
+    overhead_bits_per_mb = overhead_bits / num_mbs;
+    overhead_bits_per_mb = (int)(overhead_bits_per_mb *
+                            pow( 0.98, (double)cpi->twopass.maxq_min_limit ));
+
+    /* Try and pick a max Q that will be high enough to encode the
+     * content at the given rate.
+     */
+    for (Q = cpi->twopass.maxq_min_limit; Q < cpi->twopass.maxq_max_limit; Q++)
+    {
+        int bits_per_mb_at_this_q;
+
+        /* Error per MB based correction factor */
+        err_correction_factor =
+            calc_correction_factor(err_per_mb, 150.0, 0.40, 0.90, Q);
+
+        bits_per_mb_at_this_q =
+            vp8_bits_per_mb[INTER_FRAME][Q] + overhead_bits_per_mb;
+
+        bits_per_mb_at_this_q = (int)(.5 + err_correction_factor
+            * speed_correction * cpi->twopass.est_max_qcorrection_factor
+            * cpi->twopass.section_max_qfactor
+            * (double)bits_per_mb_at_this_q);
+
+        /* Mode and motion overhead */
+        /* As Q rises in real encode loop rd code will force overhead down
+         * We make a crude adjustment for this here as *.98 per Q step.
+         */
+        overhead_bits_per_mb = (int)((double)overhead_bits_per_mb * 0.98);
+
+        if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
+            break;
+    }
+
+    /* Restriction on active max q for constrained quality mode. */
+    if ( (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY) &&
+         (Q < cpi->cq_target_quality) )
+    {
+        Q = cpi->cq_target_quality;
+    }
+
+    /* Adjust maxq_min_limit and maxq_max_limit limits based on
+     * average q observed in clip for non kf/gf.arf frames
+     * Give average a chance to settle though.
+     */
+    if ( (cpi->ni_frames >
+                  ((int)cpi->twopass.total_stats.count >> 8)) &&
+         (cpi->ni_frames > 150) )
+    {
+        cpi->twopass.maxq_max_limit = ((cpi->ni_av_qi + 32) < cpi->worst_quality)
+                                  ? (cpi->ni_av_qi + 32) : cpi->worst_quality;
+        cpi->twopass.maxq_min_limit = ((cpi->ni_av_qi - 32) > cpi->best_quality)
+                                  ? (cpi->ni_av_qi - 32) : cpi->best_quality;
+    }
+
+    return Q;
+}
+
+/* For cq mode estimate a cq level that matches the observed
+ * complexity and data rate.
+ */
+static int estimate_cq( VP8_COMP *cpi,
+                        FIRSTPASS_STATS * fpstats,
+                        int section_target_bandwitdh,
+                        int overhead_bits )
+{
+    int Q;
+    int num_mbs = cpi->common.MBs;
+    int target_norm_bits_per_mb;
+
+    double section_err = (fpstats->coded_error / fpstats->count);
+    double err_per_mb = section_err / num_mbs;
+    double err_correction_factor;
+    double speed_correction = 1.0;
+    double clip_iiratio;
+    double clip_iifactor;
+    int overhead_bits_per_mb;
+
+    if (0)
+    {
+        FILE *f = fopen("epmp.stt", "a");
+        fprintf(f, "%10.2f\n", err_per_mb );
+        fclose(f);
+    }
+
+    target_norm_bits_per_mb = (section_target_bandwitdh < (1 << 20))
+                              ? (512 * section_target_bandwitdh) / num_mbs
+                              : 512 * (section_target_bandwitdh / num_mbs);
+
+    /* Estimate of overhead bits per mb */
+    overhead_bits_per_mb = overhead_bits / num_mbs;
+
+    /* Corrections for higher compression speed settings
+     * (reduced compression expected)
+     */
+    if ((cpi->compressor_speed == 3) || (cpi->compressor_speed == 1))
+    {
+        if (cpi->oxcf.cpu_used <= 5)
+            speed_correction = 1.04 + (cpi->oxcf.cpu_used * 0.04);
+        else
+            speed_correction = 1.25;
+    }
+
+    /* II ratio correction factor for clip as a whole */
+    clip_iiratio = cpi->twopass.total_stats.intra_error /
+                   DOUBLE_DIVIDE_CHECK(cpi->twopass.total_stats.coded_error);
+    clip_iifactor = 1.0 - ((clip_iiratio - 10.0) * 0.025);
+    if (clip_iifactor < 0.80)
+        clip_iifactor = 0.80;
+
+    /* Try and pick a Q that can encode the content at the given rate. */
+    for (Q = 0; Q < MAXQ; Q++)
+    {
+        int bits_per_mb_at_this_q;
+
+        /* Error per MB based correction factor */
+        err_correction_factor =
+            calc_correction_factor(err_per_mb, 100.0, 0.40, 0.90, Q);
+
+        bits_per_mb_at_this_q =
+            vp8_bits_per_mb[INTER_FRAME][Q] + overhead_bits_per_mb;
+
+        bits_per_mb_at_this_q =
+            (int)( .5 + err_correction_factor *
+                        speed_correction *
+                        clip_iifactor *
+                        (double)bits_per_mb_at_this_q);
+
+        /* Mode and motion overhead */
+        /* As Q rises in real encode loop rd code will force overhead down
+         * We make a crude adjustment for this here as *.98 per Q step.
+         */
+        overhead_bits_per_mb = (int)((double)overhead_bits_per_mb * 0.98);
+
+        if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
+            break;
+    }
+
+    /* Clip value to range "best allowed to (worst allowed - 1)" */
+    Q = cq_level[Q];
+    if ( Q >= cpi->worst_quality )
+        Q = cpi->worst_quality - 1;
+    if ( Q < cpi->best_quality )
+        Q = cpi->best_quality;
+
+    return Q;
+}
+
+static int estimate_q(VP8_COMP *cpi, double section_err, int section_target_bandwitdh)
+{
+    int Q;
+    int num_mbs = cpi->common.MBs;
+    int target_norm_bits_per_mb;
+
+    double err_per_mb = section_err / num_mbs;
+    double err_correction_factor;
+    double speed_correction = 1.0;
+
+    target_norm_bits_per_mb = (section_target_bandwitdh < (1 << 20)) ? (512 * section_target_bandwitdh) / num_mbs : 512 * (section_target_bandwitdh / num_mbs);
+
+    /* Corrections for higher compression speed settings
+     * (reduced compression expected)
+     */
+    if ((cpi->compressor_speed == 3) || (cpi->compressor_speed == 1))
+    {
+        if (cpi->oxcf.cpu_used <= 5)
+            speed_correction = 1.04 + (cpi->oxcf.cpu_used * 0.04);
+        else
+            speed_correction = 1.25;
+    }
+
+    /* Try and pick a Q that can encode the content at the given rate. */
+    for (Q = 0; Q < MAXQ; Q++)
+    {
+        int bits_per_mb_at_this_q;
+
+        /* Error per MB based correction factor */
+        err_correction_factor =
+            calc_correction_factor(err_per_mb, 150.0, 0.40, 0.90, Q);
+
+        bits_per_mb_at_this_q =
+            (int)( .5 + ( err_correction_factor *
+                          speed_correction *
+                          cpi->twopass.est_max_qcorrection_factor *
+                          (double)vp8_bits_per_mb[INTER_FRAME][Q] / 1.0 ) );
+
+        if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
+            break;
+    }
+
+    return Q;
+}
+
+/* Estimate a worst case Q for a KF group */
+static int estimate_kf_group_q(VP8_COMP *cpi, double section_err, int section_target_bandwitdh, double group_iiratio)
+{
+    int Q;
+    int num_mbs = cpi->common.MBs;
+    int target_norm_bits_per_mb = (512 * section_target_bandwitdh) / num_mbs;
+    int bits_per_mb_at_this_q;
+
+    double err_per_mb = section_err / num_mbs;
+    double err_correction_factor;
+    double speed_correction = 1.0;
+    double current_spend_ratio = 1.0;
+
+    double pow_highq = (POW1 < 0.6) ? POW1 + 0.3 : 0.90;
+    double pow_lowq = (POW1 < 0.7) ? POW1 + 0.1 : 0.80;
+
+    double iiratio_correction_factor = 1.0;
+
+    double combined_correction_factor;
+
+    /* Trap special case where the target is <= 0 */
+    if (target_norm_bits_per_mb <= 0)
+        return MAXQ * 2;
+
+    /* Calculate a corrective factor based on a rolling ratio of bits spent
+     *  vs target bits
+     * This is clamped to the range 0.1 to 10.0
+     */
+    if (cpi->long_rolling_target_bits <= 0)
+        current_spend_ratio = 10.0;
+    else
+    {
+        current_spend_ratio = (double)cpi->long_rolling_actual_bits / (double)cpi->long_rolling_target_bits;
+        current_spend_ratio = (current_spend_ratio > 10.0) ? 10.0 : (current_spend_ratio < 0.1) ? 0.1 : current_spend_ratio;
+    }
+
+    /* Calculate a correction factor based on the quality of prediction in
+     * the sequence as indicated by intra_inter error score ratio (IIRatio)
+     * The idea here is to favour subsampling in the hardest sections vs
+     * the easyest.
+     */
+    iiratio_correction_factor = 1.0 - ((group_iiratio - 6.0) * 0.1);
+
+    if (iiratio_correction_factor < 0.5)
+        iiratio_correction_factor = 0.5;
+
+    /* Corrections for higher compression speed settings
+     * (reduced compression expected)
+     */
+    if ((cpi->compressor_speed == 3) || (cpi->compressor_speed == 1))
+    {
+        if (cpi->oxcf.cpu_used <= 5)
+            speed_correction = 1.04 + (cpi->oxcf.cpu_used * 0.04);
+        else
+            speed_correction = 1.25;
+    }
+
+    /* Combine the various factors calculated above */
+    combined_correction_factor = speed_correction * iiratio_correction_factor * current_spend_ratio;
+
+    /* Try and pick a Q that should be high enough to encode the content at
+     * the given rate.
+     */
+    for (Q = 0; Q < MAXQ; Q++)
+    {
+        /* Error per MB based correction factor */
+        err_correction_factor =
+            calc_correction_factor(err_per_mb, 150.0, pow_lowq, pow_highq, Q);
+
+        bits_per_mb_at_this_q =
+            (int)(.5 + ( err_correction_factor *
+                         combined_correction_factor *
+                         (double)vp8_bits_per_mb[INTER_FRAME][Q]) );
+
+        if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
+            break;
+    }
+
+    /* If we could not hit the target even at Max Q then estimate what Q
+     * would have been required
+     */
+    while ((bits_per_mb_at_this_q > target_norm_bits_per_mb)  && (Q < (MAXQ * 2)))
+    {
+
+        bits_per_mb_at_this_q = (int)(0.96 * bits_per_mb_at_this_q);
+        Q++;
+    }
+
+    if (0)
+    {
+        FILE *f = fopen("estkf_q.stt", "a");
+        fprintf(f, "%8d %8d %8d %8.2f %8.3f %8.2f %8.3f %8.3f %8.3f %8d\n", cpi->common.current_video_frame, bits_per_mb_at_this_q,
+                target_norm_bits_per_mb, err_per_mb, err_correction_factor,
+                current_spend_ratio, group_iiratio, iiratio_correction_factor,
+                (double)cpi->buffer_level / (double)cpi->oxcf.optimal_buffer_level, Q);
+        fclose(f);
+    }
+
+    return Q;
+}
+
+void vp8_init_second_pass(VP8_COMP *cpi)
+{
+    FIRSTPASS_STATS this_frame;
+    FIRSTPASS_STATS *start_pos;
+
+    double two_pass_min_rate = (double)(cpi->oxcf.target_bandwidth * cpi->oxcf.two_pass_vbrmin_section / 100);
+
+    zero_stats(&cpi->twopass.total_stats);
+    zero_stats(&cpi->twopass.total_left_stats);
+
+    if (!cpi->twopass.stats_in_end)
+        return;
+
+    cpi->twopass.total_stats = *cpi->twopass.stats_in_end;
+    cpi->twopass.total_left_stats = cpi->twopass.total_stats;
+
+    /* each frame can have a different duration, as the frame rate in the
+     * source isn't guaranteed to be constant.   The frame rate prior to
+     * the first frame encoded in the second pass is a guess.  However the
+     * sum duration is not. Its calculated based on the actual durations of
+     * all frames from the first pass.
+     */
+    vp8_new_framerate(cpi, 10000000.0 * cpi->twopass.total_stats.count / cpi->twopass.total_stats.duration);
+
+    cpi->output_framerate = cpi->framerate;
+    cpi->twopass.bits_left = (int64_t)(cpi->twopass.total_stats.duration * cpi->oxcf.target_bandwidth / 10000000.0) ;
+    cpi->twopass.bits_left -= (int64_t)(cpi->twopass.total_stats.duration * two_pass_min_rate / 10000000.0);
+
+    /* Calculate a minimum intra value to be used in determining the IIratio
+     * scores used in the second pass. We have this minimum to make sure
+     * that clips that are static but "low complexity" in the intra domain
+     * are still boosted appropriately for KF/GF/ARF
+     */
+    cpi->twopass.kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
+    cpi->twopass.gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
+
+    /* Scan the first pass file and calculate an average Intra / Inter error
+     * score ratio for the sequence
+     */
+    {
+        double sum_iiratio = 0.0;
+        double IIRatio;
+
+        start_pos = cpi->twopass.stats_in; /* Note starting "file" position */
+
+        while (input_stats(cpi, &this_frame) != EOF)
+        {
+            IIRatio = this_frame.intra_error / DOUBLE_DIVIDE_CHECK(this_frame.coded_error);
+            IIRatio = (IIRatio < 1.0) ? 1.0 : (IIRatio > 20.0) ? 20.0 : IIRatio;
+            sum_iiratio += IIRatio;
+        }
+
+        cpi->twopass.avg_iiratio = sum_iiratio / DOUBLE_DIVIDE_CHECK((double)cpi->twopass.total_stats.count);
+
+        /* Reset file position */
+        reset_fpf_position(cpi, start_pos);
+    }
+
+    /* Scan the first pass file and calculate a modified total error based
+     * upon the bias/power function used to allocate bits
+     */
+    {
+        start_pos = cpi->twopass.stats_in;  /* Note starting "file" position */
+
+        cpi->twopass.modified_error_total = 0.0;
+        cpi->twopass.modified_error_used = 0.0;
+
+        while (input_stats(cpi, &this_frame) != EOF)
+        {
+            cpi->twopass.modified_error_total += calculate_modified_err(cpi, &this_frame);
+        }
+        cpi->twopass.modified_error_left = cpi->twopass.modified_error_total;
+
+        reset_fpf_position(cpi, start_pos);  /* Reset file position */
+
+    }
+}
+
+void vp8_end_second_pass(VP8_COMP *cpi)
+{
+  (void)cpi;
+}
+
+/* This function gives and estimate of how badly we believe the prediction
+ * quality is decaying from frame to frame.
+ */
+static double get_prediction_decay_rate(VP8_COMP *cpi, FIRSTPASS_STATS *next_frame)
+{
+    double prediction_decay_rate;
+    double motion_decay;
+    double motion_pct = next_frame->pcnt_motion;
+    (void)cpi;
+
+    /* Initial basis is the % mbs inter coded */
+    prediction_decay_rate = next_frame->pcnt_inter;
+
+    /* High % motion -> somewhat higher decay rate */
+    motion_decay = (1.0 - (motion_pct / 20.0));
+    if (motion_decay < prediction_decay_rate)
+        prediction_decay_rate = motion_decay;
+
+    /* Adjustment to decay rate based on speed of motion */
+    {
+        double this_mv_rabs;
+        double this_mv_cabs;
+        double distance_factor;
+
+        this_mv_rabs = fabs(next_frame->mvr_abs * motion_pct);
+        this_mv_cabs = fabs(next_frame->mvc_abs * motion_pct);
+
+        distance_factor = sqrt((this_mv_rabs * this_mv_rabs) +
+                               (this_mv_cabs * this_mv_cabs)) / 250.0;
+        distance_factor = ((distance_factor > 1.0)
+                                ? 0.0 : (1.0 - distance_factor));
+        if (distance_factor < prediction_decay_rate)
+            prediction_decay_rate = distance_factor;
+    }
+
+    return prediction_decay_rate;
+}
+
+/* Function to test for a condition where a complex transition is followed
+ * by a static section. For example in slide shows where there is a fade
+ * between slides. This is to help with more optimal kf and gf positioning.
+ */
+static int detect_transition_to_still(
+    VP8_COMP *cpi,
+    int frame_interval,
+    int still_interval,
+    double loop_decay_rate,
+    double decay_accumulator )
+{
+    int trans_to_still = 0;
+
+    /* Break clause to detect very still sections after motion
+     * For example a static image after a fade or other transition
+     * instead of a clean scene cut.
+     */
+    if ( (frame_interval > MIN_GF_INTERVAL) &&
+         (loop_decay_rate >= 0.999) &&
+         (decay_accumulator < 0.9) )
+    {
+        int j;
+        FIRSTPASS_STATS * position = cpi->twopass.stats_in;
+        FIRSTPASS_STATS tmp_next_frame;
+        double decay_rate;
+
+        /* Look ahead a few frames to see if static condition persists... */
+        for ( j = 0; j < still_interval; j++ )
+        {
+            if (EOF == input_stats(cpi, &tmp_next_frame))
+                break;
+
+            decay_rate = get_prediction_decay_rate(cpi, &tmp_next_frame);
+            if ( decay_rate < 0.999 )
+                break;
+        }
+        /* Reset file position */
+        reset_fpf_position(cpi, position);
+
+        /* Only if it does do we signal a transition to still */
+        if ( j == still_interval )
+            trans_to_still = 1;
+    }
+
+    return trans_to_still;
+}
+
+/* This function detects a flash through the high relative pcnt_second_ref
+ * score in the frame following a flash frame. The offset passed in should
+ * reflect this
+ */
+static int detect_flash( VP8_COMP *cpi, int offset )
+{
+    FIRSTPASS_STATS next_frame;
+
+    int flash_detected = 0;
+
+    /* Read the frame data. */
+    /* The return is 0 (no flash detected) if not a valid frame */
+    if ( read_frame_stats(cpi, &next_frame, offset) != EOF )
+    {
+        /* What we are looking for here is a situation where there is a
+         * brief break in prediction (such as a flash) but subsequent frames
+         * are reasonably well predicted by an earlier (pre flash) frame.
+         * The recovery after a flash is indicated by a high pcnt_second_ref
+         * comapred to pcnt_inter.
+         */
+        if ( (next_frame.pcnt_second_ref > next_frame.pcnt_inter) &&
+             (next_frame.pcnt_second_ref >= 0.5 ) )
+        {
+            flash_detected = 1;
+
+            /*if (1)
+            {
+                FILE *f = fopen("flash.stt", "a");
+                fprintf(f, "%8.0f %6.2f %6.2f\n",
+                    next_frame.frame,
+                    next_frame.pcnt_inter,
+                    next_frame.pcnt_second_ref);
+                fclose(f);
+            }*/
+        }
+    }
+
+    return flash_detected;
+}
+
+/* Update the motion related elements to the GF arf boost calculation */
+static void accumulate_frame_motion_stats(
+    VP8_COMP *cpi,
+    FIRSTPASS_STATS * this_frame,
+    double * this_frame_mv_in_out,
+    double * mv_in_out_accumulator,
+    double * abs_mv_in_out_accumulator,
+    double * mv_ratio_accumulator )
+{
+    double this_frame_mvr_ratio;
+    double this_frame_mvc_ratio;
+    double motion_pct;
+    (void)cpi;
+
+    /* Accumulate motion stats. */
+    motion_pct = this_frame->pcnt_motion;
+
+    /* Accumulate Motion In/Out of frame stats */
+    *this_frame_mv_in_out = this_frame->mv_in_out_count * motion_pct;
+    *mv_in_out_accumulator += this_frame->mv_in_out_count * motion_pct;
+    *abs_mv_in_out_accumulator +=
+        fabs(this_frame->mv_in_out_count * motion_pct);
+
+    /* Accumulate a measure of how uniform (or conversely how random)
+     * the motion field is. (A ratio of absmv / mv)
+     */
+    if (motion_pct > 0.05)
+    {
+        this_frame_mvr_ratio = fabs(this_frame->mvr_abs) /
+                               DOUBLE_DIVIDE_CHECK(fabs(this_frame->MVr));
+
+        this_frame_mvc_ratio = fabs(this_frame->mvc_abs) /
+                               DOUBLE_DIVIDE_CHECK(fabs(this_frame->MVc));
+
+         *mv_ratio_accumulator +=
+            (this_frame_mvr_ratio < this_frame->mvr_abs)
+                ? (this_frame_mvr_ratio * motion_pct)
+                : this_frame->mvr_abs * motion_pct;
+
+        *mv_ratio_accumulator +=
+            (this_frame_mvc_ratio < this_frame->mvc_abs)
+                ? (this_frame_mvc_ratio * motion_pct)
+                : this_frame->mvc_abs * motion_pct;
+
+    }
+}
+
+/* Calculate a baseline boost number for the current frame. */
+static double calc_frame_boost(
+    VP8_COMP *cpi,
+    FIRSTPASS_STATS * this_frame,
+    double this_frame_mv_in_out )
+{
+    double frame_boost;
+
+    /* Underlying boost factor is based on inter intra error ratio */
+    if (this_frame->intra_error > cpi->twopass.gf_intra_err_min)
+        frame_boost = (IIFACTOR * this_frame->intra_error /
+                      DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
+    else
+        frame_boost = (IIFACTOR * cpi->twopass.gf_intra_err_min /
+                      DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
+
+    /* Increase boost for frames where new data coming into frame
+     * (eg zoom out). Slightly reduce boost if there is a net balance
+     * of motion out of the frame (zoom in).
+     * The range for this_frame_mv_in_out is -1.0 to +1.0
+     */
+    if (this_frame_mv_in_out > 0.0)
+        frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
+    /* In extreme case boost is halved */
+    else
+        frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
+
+    /* Clip to maximum */
+    if (frame_boost > GF_RMAX)
+        frame_boost = GF_RMAX;
+
+    return frame_boost;
+}
+
+#if NEW_BOOST
+static int calc_arf_boost(
+    VP8_COMP *cpi,
+    int offset,
+    int f_frames,
+    int b_frames,
+    int *f_boost,
+    int *b_boost )
+{
+    FIRSTPASS_STATS this_frame;
+
+    int i;
+    double boost_score = 0.0;
+    double mv_ratio_accumulator = 0.0;
+    double decay_accumulator = 1.0;
+    double this_frame_mv_in_out = 0.0;
+    double mv_in_out_accumulator = 0.0;
+    double abs_mv_in_out_accumulator = 0.0;
+    double r;
+    int flash_detected = 0;
+
+    /* Search forward from the proposed arf/next gf position */
+    for ( i = 0; i < f_frames; i++ )
+    {
+        if ( read_frame_stats(cpi, &this_frame, (i+offset)) == EOF )
+            break;
+
+        /* Update the motion related elements to the boost calculation */
+        accumulate_frame_motion_stats( cpi, &this_frame,
+            &this_frame_mv_in_out, &mv_in_out_accumulator,
+            &abs_mv_in_out_accumulator, &mv_ratio_accumulator );
+
+        /* Calculate the baseline boost number for this frame */
+        r = calc_frame_boost( cpi, &this_frame, this_frame_mv_in_out );
+
+        /* We want to discount the the flash frame itself and the recovery
+         * frame that follows as both will have poor scores.
+         */
+        flash_detected = detect_flash(cpi, (i+offset)) ||
+                         detect_flash(cpi, (i+offset+1));
+
+        /* Cumulative effect of prediction quality decay */
+        if ( !flash_detected )
+        {
+            decay_accumulator =
+                decay_accumulator *
+                get_prediction_decay_rate(cpi, &this_frame);
+            decay_accumulator =
+                decay_accumulator < 0.1 ? 0.1 : decay_accumulator;
+        }
+        boost_score += (decay_accumulator * r);
+
+        /* Break out conditions. */
+        if  ( (!flash_detected) &&
+              ((mv_ratio_accumulator > 100.0) ||
+               (abs_mv_in_out_accumulator > 3.0) ||
+               (mv_in_out_accumulator < -2.0) ) )
+        {
+            break;
+        }
+    }
+
+    *f_boost = (int)(boost_score * 100.0) >> 4;
+
+    /* Reset for backward looking loop */
+    boost_score = 0.0;
+    mv_ratio_accumulator = 0.0;
+    decay_accumulator = 1.0;
+    this_frame_mv_in_out = 0.0;
+    mv_in_out_accumulator = 0.0;
+    abs_mv_in_out_accumulator = 0.0;
+
+    /* Search forward from the proposed arf/next gf position */
+    for ( i = -1; i >= -b_frames; i-- )
+    {
+        if ( read_frame_stats(cpi, &this_frame, (i+offset)) == EOF )
+            break;
+
+        /* Update the motion related elements to the boost calculation */
+        accumulate_frame_motion_stats( cpi, &this_frame,
+            &this_frame_mv_in_out, &mv_in_out_accumulator,
+            &abs_mv_in_out_accumulator, &mv_ratio_accumulator );
+
+        /* Calculate the baseline boost number for this frame */
+        r = calc_frame_boost( cpi, &this_frame, this_frame_mv_in_out );
+
+        /* We want to discount the the flash frame itself and the recovery
+         * frame that follows as both will have poor scores.
+         */
+        flash_detected = detect_flash(cpi, (i+offset)) ||
+                         detect_flash(cpi, (i+offset+1));
+
+        /* Cumulative effect of prediction quality decay */
+        if ( !flash_detected )
+        {
+            decay_accumulator =
+                decay_accumulator *
+                get_prediction_decay_rate(cpi, &this_frame);
+            decay_accumulator =
+                decay_accumulator < 0.1 ? 0.1 : decay_accumulator;
+        }
+
+        boost_score += (decay_accumulator * r);
+
+        /* Break out conditions. */
+        if  ( (!flash_detected) &&
+              ((mv_ratio_accumulator > 100.0) ||
+               (abs_mv_in_out_accumulator > 3.0) ||
+               (mv_in_out_accumulator < -2.0) ) )
+        {
+            break;
+        }
+    }
+    *b_boost = (int)(boost_score * 100.0) >> 4;
+
+    return (*f_boost + *b_boost);
+}
+#endif
+
+/* Analyse and define a gf/arf group . */
+static void define_gf_group(VP8_COMP *cpi, FIRSTPASS_STATS *this_frame)
+{
+    FIRSTPASS_STATS next_frame;
+    FIRSTPASS_STATS *start_pos;
+    int i;
+    double r;
+    double boost_score = 0.0;
+    double old_boost_score = 0.0;
+    double gf_group_err = 0.0;
+    double gf_first_frame_err = 0.0;
+    double mod_frame_err = 0.0;
+
+    double mv_ratio_accumulator = 0.0;
+    double decay_accumulator = 1.0;
+
+    double loop_decay_rate = 1.00;          /* Starting decay rate */
+
+    double this_frame_mv_in_out = 0.0;
+    double mv_in_out_accumulator = 0.0;
+    double abs_mv_in_out_accumulator = 0.0;
+    double mod_err_per_mb_accumulator = 0.0;
+
+    int max_bits = frame_max_bits(cpi);     /* Max for a single frame */
+
+    unsigned int allow_alt_ref =
+                    cpi->oxcf.play_alternate && cpi->oxcf.lag_in_frames;
+
+    int alt_boost = 0;
+    int f_boost = 0;
+    int b_boost = 0;
+    int flash_detected;
+
+    cpi->twopass.gf_group_bits = 0;
+    cpi->twopass.gf_decay_rate = 0;
+
+    vp8_clear_system_state();
+
+    start_pos = cpi->twopass.stats_in;
+
+    memset(&next_frame, 0, sizeof(next_frame)); /* assure clean */
+
+    /* Load stats for the current frame. */
+    mod_frame_err = calculate_modified_err(cpi, this_frame);
+
+    /* Note the error of the frame at the start of the group (this will be
+     * the GF frame error if we code a normal gf
+     */
+    gf_first_frame_err = mod_frame_err;
+
+    /* Special treatment if the current frame is a key frame (which is also
+     * a gf). If it is then its error score (and hence bit allocation) need
+     * to be subtracted out from the calculation for the GF group
+     */
+    if (cpi->common.frame_type == KEY_FRAME)
+        gf_group_err -= gf_first_frame_err;
+
+    /* Scan forward to try and work out how many frames the next gf group
+     * should contain and what level of boost is appropriate for the GF
+     * or ARF that will be coded with the group
+     */
+    i = 0;
+
+    while (((i < cpi->twopass.static_scene_max_gf_interval) ||
+            ((cpi->twopass.frames_to_key - i) < MIN_GF_INTERVAL)) &&
+           (i < cpi->twopass.frames_to_key))
+    {
+        i++;
+
+        /* Accumulate error score of frames in this gf group */
+        mod_frame_err = calculate_modified_err(cpi, this_frame);
+
+        gf_group_err += mod_frame_err;
+
+        mod_err_per_mb_accumulator +=
+            mod_frame_err / DOUBLE_DIVIDE_CHECK((double)cpi->common.MBs);
+
+        if (EOF == input_stats(cpi, &next_frame))
+            break;
+
+        /* Test for the case where there is a brief flash but the prediction
+         * quality back to an earlier frame is then restored.
+         */
+        flash_detected = detect_flash(cpi, 0);
+
+        /* Update the motion related elements to the boost calculation */
+        accumulate_frame_motion_stats( cpi, &next_frame,
+            &this_frame_mv_in_out, &mv_in_out_accumulator,
+            &abs_mv_in_out_accumulator, &mv_ratio_accumulator );
+
+        /* Calculate a baseline boost number for this frame */
+        r = calc_frame_boost( cpi, &next_frame, this_frame_mv_in_out );
+
+        /* Cumulative effect of prediction quality decay */
+        if ( !flash_detected )
+        {
+            loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
+            decay_accumulator = decay_accumulator * loop_decay_rate;
+            decay_accumulator =
+                decay_accumulator < 0.1 ? 0.1 : decay_accumulator;
+        }
+        boost_score += (decay_accumulator * r);
+
+        /* Break clause to detect very still sections after motion
+         * For example a staic image after a fade or other transition.
+         */
+        if ( detect_transition_to_still( cpi, i, 5,
+                                         loop_decay_rate,
+                                         decay_accumulator ) )
+        {
+            allow_alt_ref = 0;
+            boost_score = old_boost_score;
+            break;
+        }
+
+        /* Break out conditions. */
+        if  (
+            /* Break at cpi->max_gf_interval unless almost totally static */
+            (i >= cpi->max_gf_interval && (decay_accumulator < 0.995)) ||
+            (
+                /* Dont break out with a very short interval */
+                (i > MIN_GF_INTERVAL) &&
+                /* Dont break out very close to a key frame */
+                ((cpi->twopass.frames_to_key - i) >= MIN_GF_INTERVAL) &&
+                ((boost_score > 20.0) || (next_frame.pcnt_inter < 0.75)) &&
+                (!flash_detected) &&
+                ((mv_ratio_accumulator > 100.0) ||
+                 (abs_mv_in_out_accumulator > 3.0) ||
+                 (mv_in_out_accumulator < -2.0) ||
+                 ((boost_score - old_boost_score) < 2.0))
+            ) )
+        {
+            boost_score = old_boost_score;
+            break;
+        }
+
+        memcpy(this_frame, &next_frame, sizeof(*this_frame));
+
+        old_boost_score = boost_score;
+    }
+
+    cpi->twopass.gf_decay_rate =
+        (i > 0) ? (int)(100.0 * (1.0 - decay_accumulator)) / i : 0;
+
+    /* When using CBR apply additional buffer related upper limits */
+    if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+    {
+        double max_boost;
+
+        /* For cbr apply buffer related limits */
+        if (cpi->drop_frames_allowed)
+        {
+            int64_t df_buffer_level = cpi->oxcf.drop_frames_water_mark *
+                                  (cpi->oxcf.optimal_buffer_level / 100);
+
+            if (cpi->buffer_level > df_buffer_level)
+                max_boost = ((double)((cpi->buffer_level - df_buffer_level) * 2 / 3) * 16.0) / DOUBLE_DIVIDE_CHECK((double)cpi->av_per_frame_bandwidth);
+            else
+                max_boost = 0.0;
+        }
+        else if (cpi->buffer_level > 0)
+        {
+            max_boost = ((double)(cpi->buffer_level * 2 / 3) * 16.0) / DOUBLE_DIVIDE_CHECK((double)cpi->av_per_frame_bandwidth);
+        }
+        else
+        {
+            max_boost = 0.0;
+        }
+
+        if (boost_score > max_boost)
+            boost_score = max_boost;
+    }
+
+    /* Dont allow conventional gf too near the next kf */
+    if ((cpi->twopass.frames_to_key - i) < MIN_GF_INTERVAL)
+    {
+        while (i < cpi->twopass.frames_to_key)
+        {
+            i++;
+
+            if (EOF == input_stats(cpi, this_frame))
+                break;
+
+            if (i < cpi->twopass.frames_to_key)
+            {
+                mod_frame_err = calculate_modified_err(cpi, this_frame);
+                gf_group_err += mod_frame_err;
+            }
+        }
+    }
+
+    cpi->gfu_boost = (int)(boost_score * 100.0) >> 4;
+
+#if NEW_BOOST
+    /* Alterrnative boost calculation for alt ref */
+    alt_boost = calc_arf_boost( cpi, 0, (i-1), (i-1), &f_boost, &b_boost );
+#endif
+
+    /* Should we use the alternate refernce frame */
+    if (allow_alt_ref &&
+        (i >= MIN_GF_INTERVAL) &&
+        /* dont use ARF very near next kf */
+        (i <= (cpi->twopass.frames_to_key - MIN_GF_INTERVAL)) &&
+#if NEW_BOOST
+        ((next_frame.pcnt_inter > 0.75) ||
+         (next_frame.pcnt_second_ref > 0.5)) &&
+        ((mv_in_out_accumulator / (double)i > -0.2) ||
+         (mv_in_out_accumulator > -2.0)) &&
+        (b_boost > 100) &&
+        (f_boost > 100) )
+#else
+        (next_frame.pcnt_inter > 0.75) &&
+        ((mv_in_out_accumulator / (double)i > -0.2) ||
+         (mv_in_out_accumulator > -2.0)) &&
+        (cpi->gfu_boost > 100) &&
+        (cpi->twopass.gf_decay_rate <=
+            (ARF_DECAY_THRESH + (cpi->gfu_boost / 200))) )
+#endif
+    {
+        int Boost;
+        int allocation_chunks;
+        int Q = (cpi->oxcf.fixed_q < 0)
+                ? cpi->last_q[INTER_FRAME] : cpi->oxcf.fixed_q;
+        int tmp_q;
+        int arf_frame_bits = 0;
+        int group_bits;
+
+#if NEW_BOOST
+        cpi->gfu_boost = alt_boost;
+#endif
+
+        /* Estimate the bits to be allocated to the group as a whole */
+        if ((cpi->twopass.kf_group_bits > 0) &&
+            (cpi->twopass.kf_group_error_left > 0))
+        {
+            group_bits = (int)((double)cpi->twopass.kf_group_bits *
+                (gf_group_err / (double)cpi->twopass.kf_group_error_left));
+        }
+        else
+            group_bits = 0;
+
+        /* Boost for arf frame */
+#if NEW_BOOST
+        Boost = (alt_boost * GFQ_ADJUSTMENT) / 100;
+#else
+        Boost = (cpi->gfu_boost * 3 * GFQ_ADJUSTMENT) / (2 * 100);
+#endif
+        Boost += (i * 50);
+
+        /* Set max and minimum boost and hence minimum allocation */
+        if (Boost > ((cpi->baseline_gf_interval + 1) * 200))
+            Boost = ((cpi->baseline_gf_interval + 1) * 200);
+        else if (Boost < 125)
+            Boost = 125;
+
+        allocation_chunks = (i * 100) + Boost;
+
+        /* Normalize Altboost and allocations chunck down to prevent overflow */
+        while (Boost > 1000)
+        {
+            Boost /= 2;
+            allocation_chunks /= 2;
+        }
+
+        /* Calculate the number of bits to be spent on the arf based on the
+         * boost number
+         */
+        arf_frame_bits = (int)((double)Boost * (group_bits /
+                               (double)allocation_chunks));
+
+        /* Estimate if there are enough bits available to make worthwhile use
+         * of an arf.
+         */
+        tmp_q = estimate_q(cpi, mod_frame_err, (int)arf_frame_bits);
+
+        /* Only use an arf if it is likely we will be able to code
+         * it at a lower Q than the surrounding frames.
+         */
+        if (tmp_q < cpi->worst_quality)
+        {
+            int half_gf_int;
+            int frames_after_arf;
+            int frames_bwd = cpi->oxcf.arnr_max_frames - 1;
+            int frames_fwd = cpi->oxcf.arnr_max_frames - 1;
+
+            cpi->source_alt_ref_pending = 1;
+
+            /*
+             * For alt ref frames the error score for the end frame of the
+             * group (the alt ref frame) should not contribute to the group
+             * total and hence the number of bit allocated to the group.
+             * Rather it forms part of the next group (it is the GF at the
+             * start of the next group)
+             * gf_group_err -= mod_frame_err;
+             *
+             * For alt ref frames alt ref frame is technically part of the
+             * GF frame for the next group but we always base the error
+             * calculation and bit allocation on the current group of frames.
+             *
+             * Set the interval till the next gf or arf.
+             * For ARFs this is the number of frames to be coded before the
+             * future frame that is coded as an ARF.
+             * The future frame itself is part of the next group
+             */
+            cpi->baseline_gf_interval = i;
+
+            /*
+             * Define the arnr filter width for this group of frames:
+             * We only filter frames that lie within a distance of half
+             * the GF interval from the ARF frame. We also have to trap
+             * cases where the filter extends beyond the end of clip.
+             * Note: this_frame->frame has been updated in the loop
+             * so it now points at the ARF frame.
+             */
+            half_gf_int = cpi->baseline_gf_interval >> 1;
+            frames_after_arf = (int)(cpi->twopass.total_stats.count -
+                               this_frame->frame - 1);
+
+            switch (cpi->oxcf.arnr_type)
+            {
+            case 1: /* Backward filter */
+                frames_fwd = 0;
+                if (frames_bwd > half_gf_int)
+                    frames_bwd = half_gf_int;
+                break;
+
+            case 2: /* Forward filter */
+                if (frames_fwd > half_gf_int)
+                    frames_fwd = half_gf_int;
+                if (frames_fwd > frames_after_arf)
+                    frames_fwd = frames_after_arf;
+                frames_bwd = 0;
+                break;
+
+            case 3: /* Centered filter */
+            default:
+                frames_fwd >>= 1;
+                if (frames_fwd > frames_after_arf)
+                    frames_fwd = frames_after_arf;
+                if (frames_fwd > half_gf_int)
+                    frames_fwd = half_gf_int;
+
+                frames_bwd = frames_fwd;
+
+                /* For even length filter there is one more frame backward
+                 * than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff.
+                 */
+                if (frames_bwd < half_gf_int)
+                    frames_bwd += (cpi->oxcf.arnr_max_frames+1) & 0x1;
+                break;
+            }
+
+            cpi->active_arnr_frames = frames_bwd + 1 + frames_fwd;
+        }
+        else
+        {
+            cpi->source_alt_ref_pending = 0;
+            cpi->baseline_gf_interval = i;
+        }
+    }
+    else
+    {
+        cpi->source_alt_ref_pending = 0;
+        cpi->baseline_gf_interval = i;
+    }
+
+    /*
+     * Now decide how many bits should be allocated to the GF group as  a
+     * proportion of those remaining in the kf group.
+     * The final key frame group in the clip is treated as a special case
+     * where cpi->twopass.kf_group_bits is tied to cpi->twopass.bits_left.
+     * This is also important for short clips where there may only be one
+     * key frame.
+     */
+    if (cpi->twopass.frames_to_key >= (int)(cpi->twopass.total_stats.count -
+                                            cpi->common.current_video_frame))
+    {
+        cpi->twopass.kf_group_bits =
+            (cpi->twopass.bits_left > 0) ? cpi->twopass.bits_left : 0;
+    }
+
+    /* Calculate the bits to be allocated to the group as a whole */
+    if ((cpi->twopass.kf_group_bits > 0) &&
+        (cpi->twopass.kf_group_error_left > 0))
+    {
+        cpi->twopass.gf_group_bits =
+            (int64_t)(cpi->twopass.kf_group_bits *
+                      (gf_group_err / cpi->twopass.kf_group_error_left));
+    }
+    else
+        cpi->twopass.gf_group_bits = 0;
+
+    cpi->twopass.gf_group_bits =
+        (cpi->twopass.gf_group_bits < 0)
+            ? 0
+            : (cpi->twopass.gf_group_bits > cpi->twopass.kf_group_bits)
+                ? cpi->twopass.kf_group_bits : cpi->twopass.gf_group_bits;
+
+    /* Clip cpi->twopass.gf_group_bits based on user supplied data rate
+     * variability limit (cpi->oxcf.two_pass_vbrmax_section)
+     */
+    if (cpi->twopass.gf_group_bits >
+        (int64_t)max_bits * cpi->baseline_gf_interval)
+        cpi->twopass.gf_group_bits =
+            (int64_t)max_bits * cpi->baseline_gf_interval;
+
+    /* Reset the file position */
+    reset_fpf_position(cpi, start_pos);
+
+    /* Update the record of error used so far (only done once per gf group) */
+    cpi->twopass.modified_error_used += gf_group_err;
+
+    /* Assign  bits to the arf or gf. */
+    for (i = 0; i <= (cpi->source_alt_ref_pending && cpi->common.frame_type != KEY_FRAME); i++) {
+        int Boost;
+        int allocation_chunks;
+        int Q = (cpi->oxcf.fixed_q < 0) ? cpi->last_q[INTER_FRAME] : cpi->oxcf.fixed_q;
+        int gf_bits;
+
+        /* For ARF frames */
+        if (cpi->source_alt_ref_pending && i == 0)
+        {
+#if NEW_BOOST
+            Boost = (alt_boost * GFQ_ADJUSTMENT) / 100;
+#else
+            Boost = (cpi->gfu_boost * 3 * GFQ_ADJUSTMENT) / (2 * 100);
+#endif
+            Boost += (cpi->baseline_gf_interval * 50);
+
+            /* Set max and minimum boost and hence minimum allocation */
+            if (Boost > ((cpi->baseline_gf_interval + 1) * 200))
+                Boost = ((cpi->baseline_gf_interval + 1) * 200);
+            else if (Boost < 125)
+                Boost = 125;
+
+            allocation_chunks =
+                ((cpi->baseline_gf_interval + 1) * 100) + Boost;
+        }
+        /* Else for standard golden frames */
+        else
+        {
+            /* boost based on inter / intra ratio of subsequent frames */
+            Boost = (cpi->gfu_boost * GFQ_ADJUSTMENT) / 100;
+
+            /* Set max and minimum boost and hence minimum allocation */
+            if (Boost > (cpi->baseline_gf_interval * 150))
+                Boost = (cpi->baseline_gf_interval * 150);
+            else if (Boost < 125)
+                Boost = 125;
+
+            allocation_chunks =
+                (cpi->baseline_gf_interval * 100) + (Boost - 100);
+        }
+
+        /* Normalize Altboost and allocations chunck down to prevent overflow */
+        while (Boost > 1000)
+        {
+            Boost /= 2;
+            allocation_chunks /= 2;
+        }
+
+        /* Calculate the number of bits to be spent on the gf or arf based on
+         * the boost number
+         */
+        gf_bits = (int)((double)Boost *
+                        (cpi->twopass.gf_group_bits /
+                         (double)allocation_chunks));
+
+        /* If the frame that is to be boosted is simpler than the average for
+         * the gf/arf group then use an alternative calculation
+         * based on the error score of the frame itself
+         */
+        if (mod_frame_err < gf_group_err / (double)cpi->baseline_gf_interval)
+        {
+            double  alt_gf_grp_bits;
+            int     alt_gf_bits;
+
+            alt_gf_grp_bits =
+                (double)cpi->twopass.kf_group_bits  *
+                (mod_frame_err * (double)cpi->baseline_gf_interval) /
+                DOUBLE_DIVIDE_CHECK((double)cpi->twopass.kf_group_error_left);
+
+            alt_gf_bits = (int)((double)Boost * (alt_gf_grp_bits /
+                                                 (double)allocation_chunks));
+
+            if (gf_bits > alt_gf_bits)
+            {
+                gf_bits = alt_gf_bits;
+            }
+        }
+        /* Else if it is harder than other frames in the group make sure it at
+         * least receives an allocation in keeping with its relative error
+         * score, otherwise it may be worse off than an "un-boosted" frame
+         */
+        else
+        {
+            int alt_gf_bits =
+                (int)((double)cpi->twopass.kf_group_bits *
+                      mod_frame_err /
+                      DOUBLE_DIVIDE_CHECK((double)cpi->twopass.kf_group_error_left));
+
+            if (alt_gf_bits > gf_bits)
+            {
+                gf_bits = alt_gf_bits;
+            }
+        }
+
+        /* Apply an additional limit for CBR */
+        if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+        {
+            if (cpi->twopass.gf_bits > (int)(cpi->buffer_level >> 1))
+                cpi->twopass.gf_bits = (int)(cpi->buffer_level >> 1);
+        }
+
+        /* Dont allow a negative value for gf_bits */
+        if (gf_bits < 0)
+            gf_bits = 0;
+
+        /* Add in minimum for a frame */
+        gf_bits += cpi->min_frame_bandwidth;
+
+        if (i == 0)
+        {
+            cpi->twopass.gf_bits = gf_bits;
+        }
+        if (i == 1 || (!cpi->source_alt_ref_pending && (cpi->common.frame_type != KEY_FRAME)))
+        {
+            /* Per frame bit target for this frame */
+            cpi->per_frame_bandwidth = gf_bits;
+        }
+    }
+
+    {
+        /* Adjust KF group bits and error remainin */
+        cpi->twopass.kf_group_error_left -= (int64_t)gf_group_err;
+        cpi->twopass.kf_group_bits -= cpi->twopass.gf_group_bits;
+
+        if (cpi->twopass.kf_group_bits < 0)
+            cpi->twopass.kf_group_bits = 0;
+
+        /* Note the error score left in the remaining frames of the group.
+         * For normal GFs we want to remove the error score for the first
+         * frame of the group (except in Key frame case where this has
+         * already happened)
+         */
+        if (!cpi->source_alt_ref_pending && cpi->common.frame_type != KEY_FRAME)
+            cpi->twopass.gf_group_error_left = (int)(gf_group_err -
+                                                     gf_first_frame_err);
+        else
+            cpi->twopass.gf_group_error_left = (int) gf_group_err;
+
+        cpi->twopass.gf_group_bits -= cpi->twopass.gf_bits - cpi->min_frame_bandwidth;
+
+        if (cpi->twopass.gf_group_bits < 0)
+            cpi->twopass.gf_group_bits = 0;
+
+        /* This condition could fail if there are two kfs very close together
+         * despite (MIN_GF_INTERVAL) and would cause a devide by 0 in the
+         * calculation of cpi->twopass.alt_extra_bits.
+         */
+        if ( cpi->baseline_gf_interval >= 3 )
+        {
+#if NEW_BOOST
+            int boost = (cpi->source_alt_ref_pending)
+                        ? b_boost : cpi->gfu_boost;
+#else
+            int boost = cpi->gfu_boost;
+#endif
+            if ( boost >= 150 )
+            {
+                int pct_extra;
+
+                pct_extra = (boost - 100) / 50;
+                pct_extra = (pct_extra > 20) ? 20 : pct_extra;
+
+                cpi->twopass.alt_extra_bits =
+                    (int)(cpi->twopass.gf_group_bits * pct_extra) / 100;
+                cpi->twopass.gf_group_bits -= cpi->twopass.alt_extra_bits;
+                cpi->twopass.alt_extra_bits /=
+                    ((cpi->baseline_gf_interval-1)>>1);
+            }
+            else
+                cpi->twopass.alt_extra_bits = 0;
+        }
+        else
+            cpi->twopass.alt_extra_bits = 0;
+    }
+
+    /* Adjustments based on a measure of complexity of the section */
+    if (cpi->common.frame_type != KEY_FRAME)
+    {
+        FIRSTPASS_STATS sectionstats;
+        double Ratio;
+
+        zero_stats(&sectionstats);
+        reset_fpf_position(cpi, start_pos);
+
+        for (i = 0 ; i < cpi->baseline_gf_interval ; i++)
+        {
+            input_stats(cpi, &next_frame);
+            accumulate_stats(&sectionstats, &next_frame);
+        }
+
+        avg_stats(&sectionstats);
+
+        cpi->twopass.section_intra_rating = (unsigned int)
+            (sectionstats.intra_error /
+            DOUBLE_DIVIDE_CHECK(sectionstats.coded_error));
+
+        Ratio = sectionstats.intra_error / DOUBLE_DIVIDE_CHECK(sectionstats.coded_error);
+        cpi->twopass.section_max_qfactor = 1.0 - ((Ratio - 10.0) * 0.025);
+
+        if (cpi->twopass.section_max_qfactor < 0.80)
+            cpi->twopass.section_max_qfactor = 0.80;
+
+        reset_fpf_position(cpi, start_pos);
+    }
+}
+
+/* Allocate bits to a normal frame that is neither a gf an arf or a key frame. */
+static void assign_std_frame_bits(VP8_COMP *cpi, FIRSTPASS_STATS *this_frame)
+{
+    int    target_frame_size;
+
+    double modified_err;
+    double err_fraction;
+
+    int max_bits = frame_max_bits(cpi);  /* Max for a single frame */
+
+    /* Calculate modified prediction error used in bit allocation */
+    modified_err = calculate_modified_err(cpi, this_frame);
+
+    /* What portion of the remaining GF group error is used by this frame */
+    if (cpi->twopass.gf_group_error_left > 0)
+        err_fraction = modified_err / cpi->twopass.gf_group_error_left;
+    else
+        err_fraction = 0.0;
+
+    /* How many of those bits available for allocation should we give it? */
+    target_frame_size = (int)((double)cpi->twopass.gf_group_bits * err_fraction);
+
+    /* Clip to target size to 0 - max_bits (or cpi->twopass.gf_group_bits)
+     * at the top end.
+     */
+    if (target_frame_size < 0)
+        target_frame_size = 0;
+    else
+    {
+        if (target_frame_size > max_bits)
+            target_frame_size = max_bits;
+
+        if (target_frame_size > cpi->twopass.gf_group_bits)
+            target_frame_size = (int)cpi->twopass.gf_group_bits;
+    }
+
+    /* Adjust error and bits remaining */
+    cpi->twopass.gf_group_error_left -= (int)modified_err;
+    cpi->twopass.gf_group_bits -= target_frame_size;
+
+    if (cpi->twopass.gf_group_bits < 0)
+        cpi->twopass.gf_group_bits = 0;
+
+    /* Add in the minimum number of bits that is set aside for every frame. */
+    target_frame_size += cpi->min_frame_bandwidth;
+
+    /* Every other frame gets a few extra bits */
+    if ( (cpi->frames_since_golden & 0x01) &&
+         (cpi->frames_till_gf_update_due > 0) )
+    {
+        target_frame_size += cpi->twopass.alt_extra_bits;
+    }
+
+    /* Per frame bit target for this frame */
+    cpi->per_frame_bandwidth = target_frame_size;
+}
+
+void vp8_second_pass(VP8_COMP *cpi)
+{
+    int tmp_q;
+    int frames_left = (int)(cpi->twopass.total_stats.count - cpi->common.current_video_frame);
+
+    FIRSTPASS_STATS this_frame;
+    FIRSTPASS_STATS this_frame_copy;
+
+    double this_frame_intra_error;
+    double this_frame_coded_error;
+
+    int overhead_bits;
+
+    vp8_zero(this_frame);
+
+    if (!cpi->twopass.stats_in)
+    {
+        return ;
+    }
+
+    vp8_clear_system_state();
+
+    if (EOF == input_stats(cpi, &this_frame))
+        return;
+
+    this_frame_intra_error = this_frame.intra_error;
+    this_frame_coded_error = this_frame.coded_error;
+
+    /* keyframe and section processing ! */
+    if (cpi->twopass.frames_to_key == 0)
+    {
+        /* Define next KF group and assign bits to it */
+        memcpy(&this_frame_copy, &this_frame, sizeof(this_frame));
+        find_next_key_frame(cpi, &this_frame_copy);
+
+        /* Special case: Error error_resilient_mode mode does not make much
+         * sense for two pass but with its current meaning this code is
+         * designed to stop outlandish behaviour if someone does set it when
+         * using two pass. It effectively disables GF groups. This is
+         * temporary code until we decide what should really happen in this
+         * case.
+         */
+        if (cpi->oxcf.error_resilient_mode)
+        {
+            cpi->twopass.gf_group_bits = cpi->twopass.kf_group_bits;
+            cpi->twopass.gf_group_error_left =
+                                  (int)cpi->twopass.kf_group_error_left;
+            cpi->baseline_gf_interval = cpi->twopass.frames_to_key;
+            cpi->frames_till_gf_update_due = cpi->baseline_gf_interval;
+            cpi->source_alt_ref_pending = 0;
+        }
+
+    }
+
+    /* Is this a GF / ARF (Note that a KF is always also a GF) */
+    if (cpi->frames_till_gf_update_due == 0)
+    {
+        /* Define next gf group and assign bits to it */
+        memcpy(&this_frame_copy, &this_frame, sizeof(this_frame));
+        define_gf_group(cpi, &this_frame_copy);
+
+        /* If we are going to code an altref frame at the end of the group
+         * and the current frame is not a key frame.... If the previous
+         * group used an arf this frame has already benefited from that arf
+         * boost and it should not be given extra bits If the previous
+         * group was NOT coded using arf we may want to apply some boost to
+         * this GF as well
+         */
+        if (cpi->source_alt_ref_pending && (cpi->common.frame_type != KEY_FRAME))
+        {
+            /* Assign a standard frames worth of bits from those allocated
+             * to the GF group
+             */
+            int bak = cpi->per_frame_bandwidth;
+            memcpy(&this_frame_copy, &this_frame, sizeof(this_frame));
+            assign_std_frame_bits(cpi, &this_frame_copy);
+            cpi->per_frame_bandwidth = bak;
+        }
+    }
+
+    /* Otherwise this is an ordinary frame */
+    else
+    {
+        /* Special case: Error error_resilient_mode mode does not make much
+         * sense for two pass but with its current meaning but this code is
+         * designed to stop outlandish behaviour if someone does set it
+         * when using two pass. It effectively disables GF groups. This is
+         * temporary code till we decide what should really happen in this
+         * case.
+         */
+        if (cpi->oxcf.error_resilient_mode)
+        {
+            cpi->frames_till_gf_update_due = cpi->twopass.frames_to_key;
+
+            if (cpi->common.frame_type != KEY_FRAME)
+            {
+                /* Assign bits from those allocated to the GF group */
+                memcpy(&this_frame_copy, &this_frame, sizeof(this_frame));
+                assign_std_frame_bits(cpi, &this_frame_copy);
+            }
+        }
+        else
+        {
+            /* Assign bits from those allocated to the GF group */
+            memcpy(&this_frame_copy, &this_frame, sizeof(this_frame));
+            assign_std_frame_bits(cpi, &this_frame_copy);
+        }
+    }
+
+    /* Keep a globally available copy of this and the next frame's iiratio. */
+    cpi->twopass.this_iiratio = (unsigned int)(this_frame_intra_error /
+                        DOUBLE_DIVIDE_CHECK(this_frame_coded_error));
+    {
+        FIRSTPASS_STATS next_frame;
+        if ( lookup_next_frame_stats(cpi, &next_frame) != EOF )
+        {
+            cpi->twopass.next_iiratio = (unsigned int)(next_frame.intra_error /
+                                DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
+        }
+    }
+
+    /* Set nominal per second bandwidth for this frame */
+    cpi->target_bandwidth = (int)
+    (cpi->per_frame_bandwidth * cpi->output_framerate);
+    if (cpi->target_bandwidth < 0)
+        cpi->target_bandwidth = 0;
+
+
+    /* Account for mv, mode and other overheads. */
+    overhead_bits = (int)estimate_modemvcost(
+                        cpi, &cpi->twopass.total_left_stats );
+
+    /* Special case code for first frame. */
+    if (cpi->common.current_video_frame == 0)
+    {
+        cpi->twopass.est_max_qcorrection_factor = 1.0;
+
+        /* Set a cq_level in constrained quality mode. */
+        if ( cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY )
+        {
+            int est_cq;
+
+            est_cq =
+                estimate_cq( cpi,
+                             &cpi->twopass.total_left_stats,
+                             (int)(cpi->twopass.bits_left / frames_left),
+                             overhead_bits );
+
+            cpi->cq_target_quality = cpi->oxcf.cq_level;
+            if ( est_cq > cpi->cq_target_quality )
+                cpi->cq_target_quality = est_cq;
+        }
+
+        /* guess at maxq needed in 2nd pass */
+        cpi->twopass.maxq_max_limit = cpi->worst_quality;
+        cpi->twopass.maxq_min_limit = cpi->best_quality;
+
+        tmp_q = estimate_max_q(
+                    cpi,
+                    &cpi->twopass.total_left_stats,
+                    (int)(cpi->twopass.bits_left / frames_left),
+                    overhead_bits );
+
+        /* Limit the maxq value returned subsequently.
+         * This increases the risk of overspend or underspend if the initial
+         * estimate for the clip is bad, but helps prevent excessive
+         * variation in Q, especially near the end of a clip
+         * where for example a small overspend may cause Q to crash
+         */
+        cpi->twopass.maxq_max_limit = ((tmp_q + 32) < cpi->worst_quality)
+                                  ? (tmp_q + 32) : cpi->worst_quality;
+        cpi->twopass.maxq_min_limit = ((tmp_q - 32) > cpi->best_quality)
+                                  ? (tmp_q - 32) : cpi->best_quality;
+
+        cpi->active_worst_quality         = tmp_q;
+        cpi->ni_av_qi                     = tmp_q;
+    }
+
+    /* The last few frames of a clip almost always have to few or too many
+     * bits and for the sake of over exact rate control we dont want to make
+     * radical adjustments to the allowed quantizer range just to use up a
+     * few surplus bits or get beneath the target rate.
+     */
+    else if ( (cpi->common.current_video_frame <
+                 (((unsigned int)cpi->twopass.total_stats.count * 255)>>8)) &&
+              ((cpi->common.current_video_frame + cpi->baseline_gf_interval) <
+                 (unsigned int)cpi->twopass.total_stats.count) )
+    {
+        if (frames_left < 1)
+            frames_left = 1;
+
+        tmp_q = estimate_max_q(
+                    cpi,
+                    &cpi->twopass.total_left_stats,
+                    (int)(cpi->twopass.bits_left / frames_left),
+                    overhead_bits );
+
+        /* Move active_worst_quality but in a damped way */
+        if (tmp_q > cpi->active_worst_quality)
+            cpi->active_worst_quality ++;
+        else if (tmp_q < cpi->active_worst_quality)
+            cpi->active_worst_quality --;
+
+        cpi->active_worst_quality =
+            ((cpi->active_worst_quality * 3) + tmp_q + 2) / 4;
+    }
+
+    cpi->twopass.frames_to_key --;
+
+    /* Update the total stats remaining sturcture */
+    subtract_stats(&cpi->twopass.total_left_stats, &this_frame );
+}
+
+
+static int test_candidate_kf(VP8_COMP *cpi,  FIRSTPASS_STATS *last_frame, FIRSTPASS_STATS *this_frame, FIRSTPASS_STATS *next_frame)
+{
+    int is_viable_kf = 0;
+
+    /* Does the frame satisfy the primary criteria of a key frame
+     *      If so, then examine how well it predicts subsequent frames
+     */
+    if ((this_frame->pcnt_second_ref < 0.10) &&
+        (next_frame->pcnt_second_ref < 0.10) &&
+        ((this_frame->pcnt_inter < 0.05) ||
+         (
+             ((this_frame->pcnt_inter - this_frame->pcnt_neutral) < .25) &&
+             ((this_frame->intra_error / DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) < 2.5) &&
+             ((fabs(last_frame->coded_error - this_frame->coded_error) / DOUBLE_DIVIDE_CHECK(this_frame->coded_error) > .40) ||
+              (fabs(last_frame->intra_error - this_frame->intra_error) / DOUBLE_DIVIDE_CHECK(this_frame->intra_error) > .40) ||
+              ((next_frame->intra_error / DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) > 3.5)
+             )
+         )
+        )
+       )
+    {
+        int i;
+        FIRSTPASS_STATS *start_pos;
+
+        FIRSTPASS_STATS local_next_frame;
+
+        double boost_score = 0.0;
+        double old_boost_score = 0.0;
+        double decay_accumulator = 1.0;
+        double next_iiratio;
+
+        memcpy(&local_next_frame, next_frame, sizeof(*next_frame));
+
+        /* Note the starting file position so we can reset to it */
+        start_pos = cpi->twopass.stats_in;
+
+        /* Examine how well the key frame predicts subsequent frames */
+        for (i = 0 ; i < 16; i++)
+        {
+            next_iiratio = (IIKFACTOR1 * local_next_frame.intra_error / DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error)) ;
+
+            if (next_iiratio > RMAX)
+                next_iiratio = RMAX;
+
+            /* Cumulative effect of decay in prediction quality */
+            if (local_next_frame.pcnt_inter > 0.85)
+                decay_accumulator = decay_accumulator * local_next_frame.pcnt_inter;
+            else
+                decay_accumulator = decay_accumulator * ((0.85 + local_next_frame.pcnt_inter) / 2.0);
+
+            /* Keep a running total */
+            boost_score += (decay_accumulator * next_iiratio);
+
+            /* Test various breakout clauses */
+            if ((local_next_frame.pcnt_inter < 0.05) ||
+                (next_iiratio < 1.5) ||
+                (((local_next_frame.pcnt_inter -
+                   local_next_frame.pcnt_neutral) < 0.20) &&
+                 (next_iiratio < 3.0)) ||
+                ((boost_score - old_boost_score) < 0.5) ||
+                (local_next_frame.intra_error < 200)
+               )
+            {
+                break;
+            }
+
+            old_boost_score = boost_score;
+
+            /* Get the next frame details */
+            if (EOF == input_stats(cpi, &local_next_frame))
+                break;
+        }
+
+        /* If there is tolerable prediction for at least the next 3 frames
+         * then break out else discard this pottential key frame and move on
+         */
+        if (boost_score > 5.0 && (i > 3))
+            is_viable_kf = 1;
+        else
+        {
+            /* Reset the file position */
+            reset_fpf_position(cpi, start_pos);
+
+            is_viable_kf = 0;
+        }
+    }
+
+    return is_viable_kf;
+}
+static void find_next_key_frame(VP8_COMP *cpi, FIRSTPASS_STATS *this_frame)
+{
+    int i,j;
+    FIRSTPASS_STATS last_frame;
+    FIRSTPASS_STATS first_frame;
+    FIRSTPASS_STATS next_frame;
+    FIRSTPASS_STATS *start_position;
+
+    double decay_accumulator = 1.0;
+    double boost_score = 0;
+    double old_boost_score = 0.0;
+    double loop_decay_rate;
+
+    double kf_mod_err = 0.0;
+    double kf_group_err = 0.0;
+    double kf_group_intra_err = 0.0;
+    double kf_group_coded_err = 0.0;
+    double recent_loop_decay[8] = {1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0};
+
+    memset(&next_frame, 0, sizeof(next_frame));
+
+    vp8_clear_system_state();
+    start_position = cpi->twopass.stats_in;
+
+    cpi->common.frame_type = KEY_FRAME;
+
+    /* is this a forced key frame by interval */
+    cpi->this_key_frame_forced = cpi->next_key_frame_forced;
+
+    /* Clear the alt ref active flag as this can never be active on a key
+     * frame
+     */
+    cpi->source_alt_ref_active = 0;
+
+    /* Kf is always a gf so clear frames till next gf counter */
+    cpi->frames_till_gf_update_due = 0;
+
+    cpi->twopass.frames_to_key = 1;
+
+    /* Take a copy of the initial frame details */
+    memcpy(&first_frame, this_frame, sizeof(*this_frame));
+
+    cpi->twopass.kf_group_bits = 0;
+    cpi->twopass.kf_group_error_left = 0;
+
+    kf_mod_err = calculate_modified_err(cpi, this_frame);
+
+    /* find the next keyframe */
+    i = 0;
+    while (cpi->twopass.stats_in < cpi->twopass.stats_in_end)
+    {
+        /* Accumulate kf group error */
+        kf_group_err += calculate_modified_err(cpi, this_frame);
+
+        /* These figures keep intra and coded error counts for all frames
+         * including key frames in the group. The effect of the key frame
+         * itself can be subtracted out using the first_frame data
+         * collected above
+         */
+        kf_group_intra_err += this_frame->intra_error;
+        kf_group_coded_err += this_frame->coded_error;
+
+        /* Load the next frame's stats. */
+        memcpy(&last_frame, this_frame, sizeof(*this_frame));
+        input_stats(cpi, this_frame);
+
+        /* Provided that we are not at the end of the file... */
+        if (cpi->oxcf.auto_key
+            && lookup_next_frame_stats(cpi, &next_frame) != EOF)
+        {
+            /* Normal scene cut check */
+            if ( ( i >= MIN_GF_INTERVAL ) &&
+                 test_candidate_kf(cpi, &last_frame, this_frame, &next_frame) )
+            {
+                break;
+            }
+
+            /* How fast is prediction quality decaying */
+            loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
+
+            /* We want to know something about the recent past... rather than
+             * as used elsewhere where we are concened with decay in prediction
+             * quality since the last GF or KF.
+             */
+            recent_loop_decay[i%8] = loop_decay_rate;
+            decay_accumulator = 1.0;
+            for (j = 0; j < 8; j++)
+            {
+                decay_accumulator = decay_accumulator * recent_loop_decay[j];
+            }
+
+            /* Special check for transition or high motion followed by a
+             * static scene.
+             */
+            if ( detect_transition_to_still( cpi, i,
+                                             ((int)(cpi->key_frame_frequency) -
+                                              (int)i),
+                                             loop_decay_rate,
+                                             decay_accumulator ) )
+            {
+                break;
+            }
+
+
+            /* Step on to the next frame */
+            cpi->twopass.frames_to_key ++;
+
+            /* If we don't have a real key frame within the next two
+             * forcekeyframeevery intervals then break out of the loop.
+             */
+            if (cpi->twopass.frames_to_key >= 2 *(int)cpi->key_frame_frequency)
+                break;
+        } else
+            cpi->twopass.frames_to_key ++;
+
+        i++;
+    }
+
+    /* If there is a max kf interval set by the user we must obey it.
+     * We already breakout of the loop above at 2x max.
+     * This code centers the extra kf if the actual natural
+     * interval is between 1x and 2x
+     */
+    if (cpi->oxcf.auto_key
+        && cpi->twopass.frames_to_key > (int)cpi->key_frame_frequency )
+    {
+        FIRSTPASS_STATS *current_pos = cpi->twopass.stats_in;
+        FIRSTPASS_STATS tmp_frame;
+
+        cpi->twopass.frames_to_key /= 2;
+
+        /* Copy first frame details */
+        memcpy(&tmp_frame, &first_frame, sizeof(first_frame));
+
+        /* Reset to the start of the group */
+        reset_fpf_position(cpi, start_position);
+
+        kf_group_err = 0;
+        kf_group_intra_err = 0;
+        kf_group_coded_err = 0;
+
+        /* Rescan to get the correct error data for the forced kf group */
+        for( i = 0; i < cpi->twopass.frames_to_key; i++ )
+        {
+            /* Accumulate kf group errors */
+            kf_group_err += calculate_modified_err(cpi, &tmp_frame);
+            kf_group_intra_err += tmp_frame.intra_error;
+            kf_group_coded_err += tmp_frame.coded_error;
+
+            /* Load a the next frame's stats */
+            input_stats(cpi, &tmp_frame);
+        }
+
+        /* Reset to the start of the group */
+        reset_fpf_position(cpi, current_pos);
+
+        cpi->next_key_frame_forced = 1;
+    }
+    else
+        cpi->next_key_frame_forced = 0;
+
+    /* Special case for the last frame of the file */
+    if (cpi->twopass.stats_in >= cpi->twopass.stats_in_end)
+    {
+        /* Accumulate kf group error */
+        kf_group_err += calculate_modified_err(cpi, this_frame);
+
+        /* These figures keep intra and coded error counts for all frames
+         * including key frames in the group. The effect of the key frame
+         * itself can be subtracted out using the first_frame data
+         * collected above
+         */
+        kf_group_intra_err += this_frame->intra_error;
+        kf_group_coded_err += this_frame->coded_error;
+    }
+
+    /* Calculate the number of bits that should be assigned to the kf group. */
+    if ((cpi->twopass.bits_left > 0) && (cpi->twopass.modified_error_left > 0.0))
+    {
+        /* Max for a single normal frame (not key frame) */
+        int max_bits = frame_max_bits(cpi);
+
+        /* Maximum bits for the kf group */
+        int64_t max_grp_bits;
+
+        /* Default allocation based on bits left and relative
+         * complexity of the section
+         */
+        cpi->twopass.kf_group_bits = (int64_t)( cpi->twopass.bits_left *
+                                          ( kf_group_err /
+                                            cpi->twopass.modified_error_left ));
+
+        /* Clip based on maximum per frame rate defined by the user. */
+        max_grp_bits = (int64_t)max_bits * (int64_t)cpi->twopass.frames_to_key;
+        if (cpi->twopass.kf_group_bits > max_grp_bits)
+            cpi->twopass.kf_group_bits = max_grp_bits;
+
+        /* Additional special case for CBR if buffer is getting full. */
+        if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+        {
+            int64_t opt_buffer_lvl = cpi->oxcf.optimal_buffer_level;
+            int64_t buffer_lvl = cpi->buffer_level;
+
+            /* If the buffer is near or above the optimal and this kf group is
+             * not being allocated much then increase the allocation a bit.
+             */
+            if (buffer_lvl >= opt_buffer_lvl)
+            {
+                int64_t high_water_mark = (opt_buffer_lvl +
+                                       cpi->oxcf.maximum_buffer_size) >> 1;
+
+                int64_t av_group_bits;
+
+                /* Av bits per frame * number of frames */
+                av_group_bits = (int64_t)cpi->av_per_frame_bandwidth *
+                                (int64_t)cpi->twopass.frames_to_key;
+
+                /* We are at or above the maximum. */
+                if (cpi->buffer_level >= high_water_mark)
+                {
+                    int64_t min_group_bits;
+
+                    min_group_bits = av_group_bits +
+                                     (int64_t)(buffer_lvl -
+                                                 high_water_mark);
+
+                    if (cpi->twopass.kf_group_bits < min_group_bits)
+                        cpi->twopass.kf_group_bits = min_group_bits;
+                }
+                /* We are above optimal but below the maximum */
+                else if (cpi->twopass.kf_group_bits < av_group_bits)
+                {
+                    int64_t bits_below_av = av_group_bits -
+                                              cpi->twopass.kf_group_bits;
+
+                    cpi->twopass.kf_group_bits +=
+                       (int64_t)((double)bits_below_av *
+                                   (double)(buffer_lvl - opt_buffer_lvl) /
+                                   (double)(high_water_mark - opt_buffer_lvl));
+                }
+            }
+        }
+    }
+    else
+        cpi->twopass.kf_group_bits = 0;
+
+    /* Reset the first pass file position */
+    reset_fpf_position(cpi, start_position);
+
+    /* determine how big to make this keyframe based on how well the
+     * subsequent frames use inter blocks
+     */
+    decay_accumulator = 1.0;
+    boost_score = 0.0;
+
+    for (i = 0 ; i < cpi->twopass.frames_to_key ; i++)
+    {
+        double r;
+
+        if (EOF == input_stats(cpi, &next_frame))
+            break;
+
+        if (next_frame.intra_error > cpi->twopass.kf_intra_err_min)
+            r = (IIKFACTOR2 * next_frame.intra_error /
+                     DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
+        else
+            r = (IIKFACTOR2 * cpi->twopass.kf_intra_err_min /
+                     DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
+
+        if (r > RMAX)
+            r = RMAX;
+
+        /* How fast is prediction quality decaying */
+        loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
+
+        decay_accumulator = decay_accumulator * loop_decay_rate;
+        decay_accumulator = decay_accumulator < 0.1 ? 0.1 : decay_accumulator;
+
+        boost_score += (decay_accumulator * r);
+
+        if ((i > MIN_GF_INTERVAL) &&
+            ((boost_score - old_boost_score) < 1.0))
+        {
+            break;
+        }
+
+        old_boost_score = boost_score;
+    }
+
+    if (1)
+    {
+        FIRSTPASS_STATS sectionstats;
+        double Ratio;
+
+        zero_stats(&sectionstats);
+        reset_fpf_position(cpi, start_position);
+
+        for (i = 0 ; i < cpi->twopass.frames_to_key ; i++)
+        {
+            input_stats(cpi, &next_frame);
+            accumulate_stats(&sectionstats, &next_frame);
+        }
+
+        avg_stats(&sectionstats);
+
+        cpi->twopass.section_intra_rating = (unsigned int)
+            (sectionstats.intra_error
+            / DOUBLE_DIVIDE_CHECK(sectionstats.coded_error));
+
+        Ratio = sectionstats.intra_error / DOUBLE_DIVIDE_CHECK(sectionstats.coded_error);
+        cpi->twopass.section_max_qfactor = 1.0 - ((Ratio - 10.0) * 0.025);
+
+        if (cpi->twopass.section_max_qfactor < 0.80)
+            cpi->twopass.section_max_qfactor = 0.80;
+    }
+
+    /* When using CBR apply additional buffer fullness related upper limits */
+    if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+    {
+        double max_boost;
+
+        if (cpi->drop_frames_allowed)
+        {
+            int df_buffer_level = (int)(cpi->oxcf.drop_frames_water_mark
+                                  * (cpi->oxcf.optimal_buffer_level / 100));
+
+            if (cpi->buffer_level > df_buffer_level)
+                max_boost = ((double)((cpi->buffer_level - df_buffer_level) * 2 / 3) * 16.0) / DOUBLE_DIVIDE_CHECK((double)cpi->av_per_frame_bandwidth);
+            else
+                max_boost = 0.0;
+        }
+        else if (cpi->buffer_level > 0)
+        {
+            max_boost = ((double)(cpi->buffer_level * 2 / 3) * 16.0) / DOUBLE_DIVIDE_CHECK((double)cpi->av_per_frame_bandwidth);
+        }
+        else
+        {
+            max_boost = 0.0;
+        }
+
+        if (boost_score > max_boost)
+            boost_score = max_boost;
+    }
+
+    /* Reset the first pass file position */
+    reset_fpf_position(cpi, start_position);
+
+    /* Work out how many bits to allocate for the key frame itself */
+    if (1)
+    {
+        int kf_boost = (int)boost_score;
+        int allocation_chunks;
+        int Counter = cpi->twopass.frames_to_key;
+        int alt_kf_bits;
+        YV12_BUFFER_CONFIG *lst_yv12 = &cpi->common.yv12_fb[cpi->common.lst_fb_idx];
+        /* Min boost based on kf interval */
+#if 0
+
+        while ((kf_boost < 48) && (Counter > 0))
+        {
+            Counter -= 2;
+            kf_boost ++;
+        }
+
+#endif
+
+        if (kf_boost < 48)
+        {
+            kf_boost += ((Counter + 1) >> 1);
+
+            if (kf_boost > 48) kf_boost = 48;
+        }
+
+        /* bigger frame sizes need larger kf boosts, smaller frames smaller
+         * boosts...
+         */
+        if ((lst_yv12->y_width * lst_yv12->y_height) > (320 * 240))
+            kf_boost += 2 * (lst_yv12->y_width * lst_yv12->y_height) / (320 * 240);
+        else if ((lst_yv12->y_width * lst_yv12->y_height) < (320 * 240))
+            kf_boost -= 4 * (320 * 240) / (lst_yv12->y_width * lst_yv12->y_height);
+
+        /* Min KF boost */
+        kf_boost = (int)((double)kf_boost * 100.0) >> 4; /* Scale 16 to 100 */
+        if (kf_boost < 250)
+            kf_boost = 250;
+
+        /*
+         * We do three calculations for kf size.
+         * The first is based on the error score for the whole kf group.
+         * The second (optionaly) on the key frames own error if this is
+         * smaller than the average for the group.
+         * The final one insures that the frame receives at least the
+         * allocation it would have received based on its own error score vs
+         * the error score remaining
+         * Special case if the sequence appears almost totaly static
+         * as measured by the decay accumulator. In this case we want to
+         * spend almost all of the bits on the key frame.
+         * cpi->twopass.frames_to_key-1 because key frame itself is taken
+         * care of by kf_boost.
+         */
+        if ( decay_accumulator >= 0.99 )
+        {
+            allocation_chunks =
+                ((cpi->twopass.frames_to_key - 1) * 10) + kf_boost;
+        }
+        else
+        {
+            allocation_chunks =
+                ((cpi->twopass.frames_to_key - 1) * 100) + kf_boost;
+        }
+
+        /* Normalize Altboost and allocations chunck down to prevent overflow */
+        while (kf_boost > 1000)
+        {
+            kf_boost /= 2;
+            allocation_chunks /= 2;
+        }
+
+        cpi->twopass.kf_group_bits = (cpi->twopass.kf_group_bits < 0) ? 0 : cpi->twopass.kf_group_bits;
+
+        /* Calculate the number of bits to be spent on the key frame */
+        cpi->twopass.kf_bits  = (int)((double)kf_boost * ((double)cpi->twopass.kf_group_bits / (double)allocation_chunks));
+
+        /* Apply an additional limit for CBR */
+        if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+        {
+            if (cpi->twopass.kf_bits > (int)((3 * cpi->buffer_level) >> 2))
+                cpi->twopass.kf_bits = (int)((3 * cpi->buffer_level) >> 2);
+        }
+
+        /* If the key frame is actually easier than the average for the
+         * kf group (which does sometimes happen... eg a blank intro frame)
+         * Then use an alternate calculation based on the kf error score
+         * which should give a smaller key frame.
+         */
+        if (kf_mod_err < kf_group_err / cpi->twopass.frames_to_key)
+        {
+            double  alt_kf_grp_bits =
+                        ((double)cpi->twopass.bits_left *
+                         (kf_mod_err * (double)cpi->twopass.frames_to_key) /
+                         DOUBLE_DIVIDE_CHECK(cpi->twopass.modified_error_left));
+
+            alt_kf_bits = (int)((double)kf_boost *
+                                (alt_kf_grp_bits / (double)allocation_chunks));
+
+            if (cpi->twopass.kf_bits > alt_kf_bits)
+            {
+                cpi->twopass.kf_bits = alt_kf_bits;
+            }
+        }
+        /* Else if it is much harder than other frames in the group make sure
+         * it at least receives an allocation in keeping with its relative
+         * error score
+         */
+        else
+        {
+            alt_kf_bits =
+                (int)((double)cpi->twopass.bits_left *
+                      (kf_mod_err /
+                       DOUBLE_DIVIDE_CHECK(cpi->twopass.modified_error_left)));
+
+            if (alt_kf_bits > cpi->twopass.kf_bits)
+            {
+                cpi->twopass.kf_bits = alt_kf_bits;
+            }
+        }
+
+        cpi->twopass.kf_group_bits -= cpi->twopass.kf_bits;
+        /* Add in the minimum frame allowance */
+        cpi->twopass.kf_bits += cpi->min_frame_bandwidth;
+
+        /* Peer frame bit target for this frame */
+        cpi->per_frame_bandwidth = cpi->twopass.kf_bits;
+
+        /* Convert to a per second bitrate */
+        cpi->target_bandwidth = (int)(cpi->twopass.kf_bits *
+                                      cpi->output_framerate);
+    }
+
+    /* Note the total error score of the kf group minus the key frame itself */
+    cpi->twopass.kf_group_error_left = (int)(kf_group_err - kf_mod_err);
+
+    /* Adjust the count of total modified error left. The count of bits left
+     * is adjusted elsewhere based on real coded frame sizes
+     */
+    cpi->twopass.modified_error_left -= kf_group_err;
+
+    if (cpi->oxcf.allow_spatial_resampling)
+    {
+        int resample_trigger = 0;
+        int last_kf_resampled = 0;
+        int kf_q;
+        int scale_val = 0;
+        int hr, hs, vr, vs;
+        int new_width = cpi->oxcf.Width;
+        int new_height = cpi->oxcf.Height;
+
+        int projected_buffer_level;
+        int tmp_q;
+
+        double projected_bits_perframe;
+        double group_iiratio = (kf_group_intra_err - first_frame.intra_error) / (kf_group_coded_err - first_frame.coded_error);
+        double err_per_frame = kf_group_err / cpi->twopass.frames_to_key;
+        double bits_per_frame;
+        double av_bits_per_frame;
+        double effective_size_ratio;
+
+        if ((cpi->common.Width != cpi->oxcf.Width) || (cpi->common.Height != cpi->oxcf.Height))
+            last_kf_resampled = 1;
+
+        /* Set back to unscaled by defaults */
+        cpi->common.horiz_scale = NORMAL;
+        cpi->common.vert_scale = NORMAL;
+
+        /* Calculate Average bits per frame. */
+        av_bits_per_frame = cpi->oxcf.target_bandwidth / DOUBLE_DIVIDE_CHECK((double)cpi->framerate);
+
+        /* CBR... Use the clip average as the target for deciding resample */
+        if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+        {
+            bits_per_frame = av_bits_per_frame;
+        }
+
+        /* In VBR we want to avoid downsampling in easy section unless we
+         * are under extreme pressure So use the larger of target bitrate
+         * for this section or average bitrate for sequence
+         */
+        else
+        {
+            /* This accounts for how hard the section is... */
+            bits_per_frame = (double)
+                (cpi->twopass.kf_group_bits / cpi->twopass.frames_to_key);
+
+            /* Dont turn to resampling in easy sections just because they
+             * have been assigned a small number of bits
+             */
+            if (bits_per_frame < av_bits_per_frame)
+                bits_per_frame = av_bits_per_frame;
+        }
+
+        /* bits_per_frame should comply with our minimum */
+        if (bits_per_frame < (cpi->oxcf.target_bandwidth * cpi->oxcf.two_pass_vbrmin_section / 100))
+            bits_per_frame = (cpi->oxcf.target_bandwidth * cpi->oxcf.two_pass_vbrmin_section / 100);
+
+        /* Work out if spatial resampling is necessary */
+        kf_q = estimate_kf_group_q(cpi, err_per_frame,
+                                  (int)bits_per_frame, group_iiratio);
+
+        /* If we project a required Q higher than the maximum allowed Q then
+         * make a guess at the actual size of frames in this section
+         */
+        projected_bits_perframe = bits_per_frame;
+        tmp_q = kf_q;
+
+        while (tmp_q > cpi->worst_quality)
+        {
+            projected_bits_perframe *= 1.04;
+            tmp_q--;
+        }
+
+        /* Guess at buffer level at the end of the section */
+        projected_buffer_level = (int)
+                    (cpi->buffer_level - (int)
+                    ((projected_bits_perframe - av_bits_per_frame) *
+                    cpi->twopass.frames_to_key));
+
+        if (0)
+        {
+            FILE *f = fopen("Subsamle.stt", "a");
+            fprintf(f, " %8d %8d %8d %8d %12.0f %8d %8d %8d\n",  cpi->common.current_video_frame, kf_q, cpi->common.horiz_scale, cpi->common.vert_scale,  kf_group_err / cpi->twopass.frames_to_key, (int)(cpi->twopass.kf_group_bits / cpi->twopass.frames_to_key), new_height, new_width);
+            fclose(f);
+        }
+
+        /* The trigger for spatial resampling depends on the various
+         * parameters such as whether we are streaming (CBR) or VBR.
+         */
+        if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+        {
+            /* Trigger resample if we are projected to fall below down
+             * sample level or resampled last time and are projected to
+             * remain below the up sample level
+             */
+            if ((projected_buffer_level < (cpi->oxcf.resample_down_water_mark * cpi->oxcf.optimal_buffer_level / 100)) ||
+                (last_kf_resampled && (projected_buffer_level < (cpi->oxcf.resample_up_water_mark * cpi->oxcf.optimal_buffer_level / 100))))
+                resample_trigger = 1;
+            else
+                resample_trigger = 0;
+        }
+        else
+        {
+            int64_t clip_bits = (int64_t)(cpi->twopass.total_stats.count * cpi->oxcf.target_bandwidth / DOUBLE_DIVIDE_CHECK((double)cpi->framerate));
+            int64_t over_spend = cpi->oxcf.starting_buffer_level - cpi->buffer_level;
+
+            /* If triggered last time the threshold for triggering again is
+             * reduced:
+             *
+             * Projected Q higher than allowed and Overspend > 5% of total
+             * bits
+             */
+            if ((last_kf_resampled && (kf_q > cpi->worst_quality)) ||
+                ((kf_q > cpi->worst_quality) &&
+                 (over_spend > clip_bits / 20)))
+                resample_trigger = 1;
+            else
+                resample_trigger = 0;
+
+        }
+
+        if (resample_trigger)
+        {
+            while ((kf_q >= cpi->worst_quality) && (scale_val < 6))
+            {
+                scale_val ++;
+
+                cpi->common.vert_scale   = vscale_lookup[scale_val];
+                cpi->common.horiz_scale  = hscale_lookup[scale_val];
+
+                Scale2Ratio(cpi->common.horiz_scale, &hr, &hs);
+                Scale2Ratio(cpi->common.vert_scale, &vr, &vs);
+
+                new_width = ((hs - 1) + (cpi->oxcf.Width * hr)) / hs;
+                new_height = ((vs - 1) + (cpi->oxcf.Height * vr)) / vs;
+
+                /* Reducing the area to 1/4 does not reduce the complexity
+                 * (err_per_frame) to 1/4... effective_sizeratio attempts
+                 * to provide a crude correction for this
+                 */
+                effective_size_ratio = (double)(new_width * new_height) / (double)(cpi->oxcf.Width * cpi->oxcf.Height);
+                effective_size_ratio = (1.0 + (3.0 * effective_size_ratio)) / 4.0;
+
+                /* Now try again and see what Q we get with the smaller
+                 * image size
+                 */
+                kf_q = estimate_kf_group_q(cpi,
+                                          err_per_frame * effective_size_ratio,
+                                          (int)bits_per_frame, group_iiratio);
+
+                if (0)
+                {
+                    FILE *f = fopen("Subsamle.stt", "a");
+                    fprintf(f, "******** %8d %8d %8d %12.0f %8d %8d %8d\n",  kf_q, cpi->common.horiz_scale, cpi->common.vert_scale,  kf_group_err / cpi->twopass.frames_to_key, (int)(cpi->twopass.kf_group_bits / cpi->twopass.frames_to_key), new_height, new_width);
+                    fclose(f);
+                }
+            }
+        }
+
+        if ((cpi->common.Width != new_width) || (cpi->common.Height != new_height))
+        {
+            cpi->common.Width = new_width;
+            cpi->common.Height = new_height;
+            vp8_alloc_compressor_data(cpi);
+        }
+    }
+}
diff --git a/libvpx/libvpx/vp8/encoder/firstpass.h b/libvpx/libvpx/vp8/encoder/firstpass.h
new file mode 100644
index 0000000..c409ebc
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/firstpass.h
@@ -0,0 +1,32 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_FIRSTPASS_H_
+#define VP8_ENCODER_FIRSTPASS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern void vp8_init_first_pass(VP8_COMP *cpi);
+extern void vp8_first_pass(VP8_COMP *cpi);
+extern void vp8_end_first_pass(VP8_COMP *cpi);
+
+extern void vp8_init_second_pass(VP8_COMP *cpi);
+extern void vp8_second_pass(VP8_COMP *cpi);
+extern void vp8_end_second_pass(VP8_COMP *cpi);
+
+extern size_t vp8_firstpass_stats_sz(unsigned int mb_count);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_FIRSTPASS_H_
diff --git a/libvpx/libvpx/vp8/encoder/lookahead.c b/libvpx/libvpx/vp8/encoder/lookahead.c
new file mode 100644
index 0000000..6623385
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/lookahead.c
@@ -0,0 +1,231 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <assert.h>
+#include <stdlib.h>
+#include "vpx_config.h"
+#include "lookahead.h"
+#include "vp8/common/extend.h"
+
+#define MAX_LAG_BUFFERS (CONFIG_REALTIME_ONLY? 1 : 25)
+
+struct lookahead_ctx
+{
+    unsigned int max_sz;         /* Absolute size of the queue */
+    unsigned int sz;             /* Number of buffers currently in the queue */
+    unsigned int read_idx;       /* Read index */
+    unsigned int write_idx;      /* Write index */
+    struct lookahead_entry *buf; /* Buffer list */
+};
+
+
+/* Return the buffer at the given absolute index and increment the index */
+static struct lookahead_entry *
+pop(struct lookahead_ctx *ctx,
+    unsigned int         *idx)
+{
+    unsigned int            index = *idx;
+    struct lookahead_entry *buf = ctx->buf + index;
+
+    assert(index < ctx->max_sz);
+    if(++index >= ctx->max_sz)
+        index -= ctx->max_sz;
+    *idx = index;
+    return buf;
+}
+
+
+void
+vp8_lookahead_destroy(struct lookahead_ctx *ctx)
+{
+    if(ctx)
+    {
+        if(ctx->buf)
+        {
+            unsigned int i;
+
+            for(i = 0; i < ctx->max_sz; i++)
+                vp8_yv12_de_alloc_frame_buffer(&ctx->buf[i].img);
+            free(ctx->buf);
+        }
+        free(ctx);
+    }
+}
+
+
+struct lookahead_ctx*
+vp8_lookahead_init(unsigned int width,
+                   unsigned int height,
+                   unsigned int depth)
+{
+    struct lookahead_ctx *ctx = NULL;
+    unsigned int i;
+
+    /* Clamp the lookahead queue depth */
+    if(depth < 1)
+        depth = 1;
+    else if(depth > MAX_LAG_BUFFERS)
+        depth = MAX_LAG_BUFFERS;
+
+    /* Keep last frame in lookahead buffer by increasing depth by 1.*/
+    depth += 1;
+
+    /* Align the buffer dimensions */
+    width = (width + 15) & ~15;
+    height = (height + 15) & ~15;
+
+    /* Allocate the lookahead structures */
+    ctx = calloc(1, sizeof(*ctx));
+    if(ctx)
+    {
+        ctx->max_sz = depth;
+        ctx->buf = calloc(depth, sizeof(*ctx->buf));
+        if(!ctx->buf)
+            goto bail;
+        for(i=0; i<depth; i++)
+            if (vp8_yv12_alloc_frame_buffer(&ctx->buf[i].img,
+                                            width, height, VP8BORDERINPIXELS))
+                goto bail;
+    }
+    return ctx;
+bail:
+    vp8_lookahead_destroy(ctx);
+    return NULL;
+}
+
+
+int
+vp8_lookahead_push(struct lookahead_ctx *ctx,
+                   YV12_BUFFER_CONFIG   *src,
+                   int64_t               ts_start,
+                   int64_t               ts_end,
+                   unsigned int          flags,
+                   unsigned char        *active_map)
+{
+    struct lookahead_entry* buf;
+    int row, col, active_end;
+    int mb_rows = (src->y_height + 15) >> 4;
+    int mb_cols = (src->y_width + 15) >> 4;
+
+    if(ctx->sz + 2 > ctx->max_sz)
+        return 1;
+    ctx->sz++;
+    buf = pop(ctx, &ctx->write_idx);
+
+    /* Only do this partial copy if the following conditions are all met:
+     * 1. Lookahead queue has has size of 1.
+     * 2. Active map is provided.
+     * 3. This is not a key frame, golden nor altref frame.
+     */
+    if (ctx->max_sz == 1 && active_map && !flags)
+    {
+        for (row = 0; row < mb_rows; ++row)
+        {
+            col = 0;
+
+            while (1)
+            {
+                /* Find the first active macroblock in this row. */
+                for (; col < mb_cols; ++col)
+                {
+                    if (active_map[col])
+                        break;
+                }
+
+                /* No more active macroblock in this row. */
+                if (col == mb_cols)
+                    break;
+
+                /* Find the end of active region in this row. */
+                active_end = col;
+
+                for (; active_end < mb_cols; ++active_end)
+                {
+                    if (!active_map[active_end])
+                        break;
+                }
+
+                /* Only copy this active region. */
+                vp8_copy_and_extend_frame_with_rect(src, &buf->img,
+                                                    row << 4,
+                                                    col << 4, 16,
+                                                    (active_end - col) << 4);
+
+                /* Start again from the end of this active region. */
+                col = active_end;
+            }
+
+            active_map += mb_cols;
+        }
+    }
+    else
+    {
+        vp8_copy_and_extend_frame(src, &buf->img);
+    }
+    buf->ts_start = ts_start;
+    buf->ts_end = ts_end;
+    buf->flags = flags;
+    return 0;
+}
+
+
+struct lookahead_entry*
+vp8_lookahead_pop(struct lookahead_ctx *ctx,
+                  int                   drain)
+{
+    struct lookahead_entry* buf = NULL;
+
+    assert(ctx != NULL);
+    if(ctx->sz && (drain || ctx->sz == ctx->max_sz - 1))
+    {
+        buf = pop(ctx, &ctx->read_idx);
+        ctx->sz--;
+    }
+    return buf;
+}
+
+
+struct lookahead_entry*
+vp8_lookahead_peek(struct lookahead_ctx *ctx,
+                   unsigned int          index,
+                   int                   direction)
+{
+    struct lookahead_entry* buf = NULL;
+
+    if (direction == PEEK_FORWARD)
+    {
+        assert(index < ctx->max_sz - 1);
+        if(index < ctx->sz)
+        {
+            index += ctx->read_idx;
+            if(index >= ctx->max_sz)
+                index -= ctx->max_sz;
+            buf = ctx->buf + index;
+        }
+    }
+    else if (direction == PEEK_BACKWARD)
+    {
+        assert(index == 1);
+
+        if(ctx->read_idx == 0)
+            index = ctx->max_sz - 1;
+        else
+            index = ctx->read_idx - index;
+        buf = ctx->buf + index;
+    }
+
+    return buf;
+}
+
+
+unsigned int
+vp8_lookahead_depth(struct lookahead_ctx *ctx)
+{
+    return ctx->sz;
+}
diff --git a/libvpx/libvpx/vp8/encoder/lookahead.h b/libvpx/libvpx/vp8/encoder/lookahead.h
new file mode 100644
index 0000000..cad68e6
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/lookahead.h
@@ -0,0 +1,117 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VP8_ENCODER_LOOKAHEAD_H_
+#define VP8_ENCODER_LOOKAHEAD_H_
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct lookahead_entry
+{
+    YV12_BUFFER_CONFIG  img;
+    int64_t             ts_start;
+    int64_t             ts_end;
+    unsigned int        flags;
+};
+
+
+struct lookahead_ctx;
+
+/**\brief Initializes the lookahead stage
+ *
+ * The lookahead stage is a queue of frame buffers on which some analysis
+ * may be done when buffers are enqueued.
+ *
+ *
+ */
+struct lookahead_ctx* vp8_lookahead_init(unsigned int width,
+                                         unsigned int height,
+                                         unsigned int depth
+                                         );
+
+
+/**\brief Destroys the lookahead stage
+ *
+ */
+void vp8_lookahead_destroy(struct lookahead_ctx *ctx);
+
+
+/**\brief Enqueue a source buffer
+ *
+ * This function will copy the source image into a new framebuffer with
+ * the expected stride/border.
+ *
+ * If active_map is non-NULL and there is only one frame in the queue, then copy
+ * only active macroblocks.
+ *
+ * \param[in] ctx         Pointer to the lookahead context
+ * \param[in] src         Pointer to the image to enqueue
+ * \param[in] ts_start    Timestamp for the start of this frame
+ * \param[in] ts_end      Timestamp for the end of this frame
+ * \param[in] flags       Flags set on this frame
+ * \param[in] active_map  Map that specifies which macroblock is active
+ */
+int
+vp8_lookahead_push(struct lookahead_ctx *ctx,
+                   YV12_BUFFER_CONFIG   *src,
+                   int64_t               ts_start,
+                   int64_t               ts_end,
+                   unsigned int          flags,
+                   unsigned char        *active_map);
+
+
+/**\brief Get the next source buffer to encode
+ *
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ * \param[in] drain     Flag indicating the buffer should be drained
+ *                      (return a buffer regardless of the current queue depth)
+ *
+ * \retval NULL, if drain set and queue is empty
+ * \retval NULL, if drain not set and queue not of the configured depth
+ *
+ */
+struct lookahead_entry*
+vp8_lookahead_pop(struct lookahead_ctx *ctx,
+                  int                   drain);
+
+
+#define PEEK_FORWARD   1
+#define PEEK_BACKWARD -1
+/**\brief Get a future source buffer to encode
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ * \param[in] index     Index of the frame to be returned, 0 == next frame
+ *
+ * \retval NULL, if no buffer exists at the specified index
+ *
+ */
+struct lookahead_entry*
+vp8_lookahead_peek(struct lookahead_ctx *ctx,
+                   unsigned int          index,
+                   int                   direction);
+
+
+/**\brief Get the number of frames currently in the lookahead queue
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ */
+unsigned int
+vp8_lookahead_depth(struct lookahead_ctx *ctx);
+
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_LOOKAHEAD_H_
diff --git a/libvpx/libvpx/vp8/encoder/mcomp.c b/libvpx/libvpx/vp8/encoder/mcomp.c
new file mode 100644
index 0000000..e20c1ea
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/mcomp.c
@@ -0,0 +1,2032 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "./vp8_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "onyx_int.h"
+#include "mcomp.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_config.h"
+#include <stdio.h>
+#include <limits.h>
+#include <math.h>
+#include "vp8/common/findnearmv.h"
+#include "vp8/common/common.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+#ifdef VP8_ENTROPY_STATS
+static int mv_ref_ct [31] [4] [2];
+static int mv_mode_cts [4] [2];
+#endif
+
+int vp8_mv_bit_cost(int_mv *mv, int_mv *ref, int *mvcost[2], int Weight)
+{
+    /* MV costing is based on the distribution of vectors in the previous
+     * frame and as such will tend to over state the cost of vectors. In
+     * addition coding a new vector can have a knock on effect on the cost
+     * of subsequent vectors and the quality of prediction from NEAR and
+     * NEAREST for subsequent blocks. The "Weight" parameter allows, to a
+     * limited extent, for some account to be taken of these factors.
+     */
+    return ((mvcost[0][(mv->as_mv.row - ref->as_mv.row) >> 1] + mvcost[1][(mv->as_mv.col - ref->as_mv.col) >> 1]) * Weight) >> 7;
+}
+
+static int mv_err_cost(int_mv *mv, int_mv *ref, int *mvcost[2], int error_per_bit)
+{
+    /* Ignore mv costing if mvcost is NULL */
+    if (mvcost)
+        return ((mvcost[0][(mv->as_mv.row - ref->as_mv.row) >> 1] +
+                 mvcost[1][(mv->as_mv.col - ref->as_mv.col) >> 1])
+                 * error_per_bit + 128) >> 8;
+    return 0;
+}
+
+static int mvsad_err_cost(int_mv *mv, int_mv *ref, int *mvsadcost[2], int error_per_bit)
+{
+    /* Calculate sad error cost on full pixel basis. */
+    /* Ignore mv costing if mvsadcost is NULL */
+    if (mvsadcost)
+        return ((mvsadcost[0][(mv->as_mv.row - ref->as_mv.row)] +
+                 mvsadcost[1][(mv->as_mv.col - ref->as_mv.col)])
+                * error_per_bit + 128) >> 8;
+    return 0;
+}
+
+void vp8_init_dsmotion_compensation(MACROBLOCK *x, int stride)
+{
+    int Len;
+    int search_site_count = 0;
+
+
+    /* Generate offsets for 4 search sites per step. */
+    Len = MAX_FIRST_STEP;
+    x->ss[search_site_count].mv.col = 0;
+    x->ss[search_site_count].mv.row = 0;
+    x->ss[search_site_count].offset = 0;
+    search_site_count++;
+
+    while (Len > 0)
+    {
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = 0;
+        x->ss[search_site_count].mv.row = -Len;
+        x->ss[search_site_count].offset = -Len * stride;
+        search_site_count++;
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = 0;
+        x->ss[search_site_count].mv.row = Len;
+        x->ss[search_site_count].offset = Len * stride;
+        search_site_count++;
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = -Len;
+        x->ss[search_site_count].mv.row = 0;
+        x->ss[search_site_count].offset = -Len;
+        search_site_count++;
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = Len;
+        x->ss[search_site_count].mv.row = 0;
+        x->ss[search_site_count].offset = Len;
+        search_site_count++;
+
+        /* Contract. */
+        Len /= 2;
+    }
+
+    x->ss_count = search_site_count;
+    x->searches_per_step = 4;
+}
+
+void vp8_init3smotion_compensation(MACROBLOCK *x, int stride)
+{
+    int Len;
+    int search_site_count = 0;
+
+    /* Generate offsets for 8 search sites per step. */
+    Len = MAX_FIRST_STEP;
+    x->ss[search_site_count].mv.col = 0;
+    x->ss[search_site_count].mv.row = 0;
+    x->ss[search_site_count].offset = 0;
+    search_site_count++;
+
+    while (Len > 0)
+    {
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = 0;
+        x->ss[search_site_count].mv.row = -Len;
+        x->ss[search_site_count].offset = -Len * stride;
+        search_site_count++;
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = 0;
+        x->ss[search_site_count].mv.row = Len;
+        x->ss[search_site_count].offset = Len * stride;
+        search_site_count++;
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = -Len;
+        x->ss[search_site_count].mv.row = 0;
+        x->ss[search_site_count].offset = -Len;
+        search_site_count++;
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = Len;
+        x->ss[search_site_count].mv.row = 0;
+        x->ss[search_site_count].offset = Len;
+        search_site_count++;
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = -Len;
+        x->ss[search_site_count].mv.row = -Len;
+        x->ss[search_site_count].offset = -Len * stride - Len;
+        search_site_count++;
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = Len;
+        x->ss[search_site_count].mv.row = -Len;
+        x->ss[search_site_count].offset = -Len * stride + Len;
+        search_site_count++;
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = -Len;
+        x->ss[search_site_count].mv.row = Len;
+        x->ss[search_site_count].offset = Len * stride - Len;
+        search_site_count++;
+
+        /* Compute offsets for search sites. */
+        x->ss[search_site_count].mv.col = Len;
+        x->ss[search_site_count].mv.row = Len;
+        x->ss[search_site_count].offset = Len * stride + Len;
+        search_site_count++;
+
+
+        /* Contract. */
+        Len /= 2;
+    }
+
+    x->ss_count = search_site_count;
+    x->searches_per_step = 8;
+}
+
+/*
+ * To avoid the penalty for crossing cache-line read, preload the reference
+ * area in a small buffer, which is aligned to make sure there won't be crossing
+ * cache-line read while reading from this buffer. This reduced the cpu
+ * cycles spent on reading ref data in sub-pixel filter functions.
+ * TODO: Currently, since sub-pixel search range here is -3 ~ 3, copy 22 rows x
+ * 32 cols area that is enough for 16x16 macroblock. Later, for SPLITMV, we
+ * could reduce the area.
+ */
+
+/* estimated cost of a motion vector (r,c) */
+#define MVC(r,c) (mvcost ? ((mvcost[0][(r)-rr] + mvcost[1][(c) - rc]) * error_per_bit + 128 )>>8 : 0)
+/* pointer to predictor base of a motionvector */
+#define PRE(r,c) (y + (((r)>>2) * y_stride + ((c)>>2) -(offset)))
+/* convert motion vector component to offset for svf calc */
+#define SP(x) (((x)&3)<<1)
+/* returns subpixel variance error function. */
+#define DIST(r,c) vfp->svf( PRE(r,c), y_stride, SP(c),SP(r), z,b->src_stride,&sse)
+#define IFMVCV(r,c,s,e) if ( c >= minc && c <= maxc && r >= minr && r <= maxr) s else e;
+/* returns distortion + motion vector cost */
+#define ERR(r,c) (MVC(r,c)+DIST(r,c))
+/* checks if (r,c) has better score than previous best */
+#define CHECK_BETTER(v,r,c) IFMVCV(r,c,{thismse = DIST(r,c); if((v = (MVC(r,c)+thismse)) < besterr) { besterr = v; br=r; bc=c; *distortion = thismse; *sse1 = sse; }}, v=UINT_MAX;)
+
+int vp8_find_best_sub_pixel_step_iteratively(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
+                                             int_mv *bestmv, int_mv *ref_mv,
+                                             int error_per_bit,
+                                             const vp8_variance_fn_ptr_t *vfp,
+                                             int *mvcost[2], int *distortion,
+                                             unsigned int *sse1)
+{
+    unsigned char *z = (*(b->base_src) + b->src);
+
+    int rr = ref_mv->as_mv.row >> 1, rc = ref_mv->as_mv.col >> 1;
+    int br = bestmv->as_mv.row * 4, bc = bestmv->as_mv.col * 4;
+    int tr = br, tc = bc;
+    unsigned int besterr;
+    unsigned int left, right, up, down, diag;
+    unsigned int sse;
+    unsigned int whichdir;
+    unsigned int halfiters = 4;
+    unsigned int quarteriters = 4;
+    int thismse;
+
+    int minc = VPXMAX(x->mv_col_min * 4,
+                      (ref_mv->as_mv.col >> 1) - ((1 << mvlong_width) - 1));
+    int maxc = VPXMIN(x->mv_col_max * 4,
+                      (ref_mv->as_mv.col >> 1) + ((1 << mvlong_width) - 1));
+    int minr = VPXMAX(x->mv_row_min * 4,
+                      (ref_mv->as_mv.row >> 1) - ((1 << mvlong_width) - 1));
+    int maxr = VPXMIN(x->mv_row_max * 4,
+                      (ref_mv->as_mv.row >> 1) + ((1 << mvlong_width) - 1));
+
+    int y_stride;
+    int offset;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+
+
+#if ARCH_X86 || ARCH_X86_64
+    MACROBLOCKD *xd = &x->e_mbd;
+    unsigned char *y_0 = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
+    unsigned char *y;
+    int buf_r1, buf_r2, buf_c1;
+
+    /* Clamping to avoid out-of-range data access */
+    buf_r1 = ((bestmv->as_mv.row - 3) < x->mv_row_min)?(bestmv->as_mv.row - x->mv_row_min):3;
+    buf_r2 = ((bestmv->as_mv.row + 3) > x->mv_row_max)?(x->mv_row_max - bestmv->as_mv.row):3;
+    buf_c1 = ((bestmv->as_mv.col - 3) < x->mv_col_min)?(bestmv->as_mv.col - x->mv_col_min):3;
+    y_stride = 32;
+
+    /* Copy to intermediate buffer before searching. */
+    vfp->copymem(y_0 - buf_c1 - pre_stride*buf_r1, pre_stride, xd->y_buf, y_stride, 16+buf_r1+buf_r2);
+    y = xd->y_buf + y_stride*buf_r1 +buf_c1;
+#else
+    unsigned char *y = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
+    y_stride = pre_stride;
+#endif
+
+    offset = (bestmv->as_mv.row) * y_stride + bestmv->as_mv.col;
+
+    /* central mv */
+    bestmv->as_mv.row *= 8;
+    bestmv->as_mv.col *= 8;
+
+    /* calculate central point error */
+    besterr = vfp->vf(y, y_stride, z, b->src_stride, sse1);
+    *distortion = besterr;
+    besterr += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
+
+    /* TODO: Each subsequent iteration checks at least one point in common
+     * with the last iteration could be 2 ( if diag selected)
+     */
+    while (--halfiters)
+    {
+        /* 1/2 pel */
+        CHECK_BETTER(left, tr, tc - 2);
+        CHECK_BETTER(right, tr, tc + 2);
+        CHECK_BETTER(up, tr - 2, tc);
+        CHECK_BETTER(down, tr + 2, tc);
+
+        whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
+
+        switch (whichdir)
+        {
+        case 0:
+            CHECK_BETTER(diag, tr - 2, tc - 2);
+            break;
+        case 1:
+            CHECK_BETTER(diag, tr - 2, tc + 2);
+            break;
+        case 2:
+            CHECK_BETTER(diag, tr + 2, tc - 2);
+            break;
+        case 3:
+            CHECK_BETTER(diag, tr + 2, tc + 2);
+            break;
+        }
+
+        /* no reason to check the same one again. */
+        if (tr == br && tc == bc)
+            break;
+
+        tr = br;
+        tc = bc;
+    }
+
+    /* TODO: Each subsequent iteration checks at least one point in common
+     * with the last iteration could be 2 ( if diag selected)
+     */
+
+    /* 1/4 pel */
+    while (--quarteriters)
+    {
+        CHECK_BETTER(left, tr, tc - 1);
+        CHECK_BETTER(right, tr, tc + 1);
+        CHECK_BETTER(up, tr - 1, tc);
+        CHECK_BETTER(down, tr + 1, tc);
+
+        whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
+
+        switch (whichdir)
+        {
+        case 0:
+            CHECK_BETTER(diag, tr - 1, tc - 1);
+            break;
+        case 1:
+            CHECK_BETTER(diag, tr - 1, tc + 1);
+            break;
+        case 2:
+            CHECK_BETTER(diag, tr + 1, tc - 1);
+            break;
+        case 3:
+            CHECK_BETTER(diag, tr + 1, tc + 1);
+            break;
+        }
+
+        /* no reason to check the same one again. */
+        if (tr == br && tc == bc)
+            break;
+
+        tr = br;
+        tc = bc;
+    }
+
+    bestmv->as_mv.row = br * 2;
+    bestmv->as_mv.col = bc * 2;
+
+    if ((abs(bestmv->as_mv.col - ref_mv->as_mv.col) > (MAX_FULL_PEL_VAL<<3)) ||
+        (abs(bestmv->as_mv.row - ref_mv->as_mv.row) > (MAX_FULL_PEL_VAL<<3)))
+        return INT_MAX;
+
+    return besterr;
+}
+#undef MVC
+#undef PRE
+#undef SP
+#undef DIST
+#undef IFMVCV
+#undef ERR
+#undef CHECK_BETTER
+
+int vp8_find_best_sub_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
+                                 int_mv *bestmv, int_mv *ref_mv,
+                                 int error_per_bit,
+                                 const vp8_variance_fn_ptr_t *vfp,
+                                 int *mvcost[2], int *distortion,
+                                 unsigned int *sse1)
+{
+    int bestmse = INT_MAX;
+    int_mv startmv;
+    int_mv this_mv;
+    unsigned char *z = (*(b->base_src) + b->src);
+    int left, right, up, down, diag;
+    unsigned int sse;
+    int whichdir ;
+    int thismse;
+    int y_stride;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+
+#if ARCH_X86 || ARCH_X86_64
+    MACROBLOCKD *xd = &x->e_mbd;
+    unsigned char *y_0 = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
+    unsigned char *y;
+
+    y_stride = 32;
+    /* Copy 18 rows x 32 cols area to intermediate buffer before searching. */
+     vfp->copymem(y_0 - 1 - pre_stride, pre_stride, xd->y_buf, y_stride, 18);
+     y = xd->y_buf + y_stride + 1;
+#else
+     unsigned char *y = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
+     y_stride = pre_stride;
+#endif
+
+    /* central mv */
+    bestmv->as_mv.row *= 8;
+    bestmv->as_mv.col *= 8;
+    startmv = *bestmv;
+
+    /* calculate central point error */
+    bestmse = vfp->vf(y, y_stride, z, b->src_stride, sse1);
+    *distortion = bestmse;
+    bestmse += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
+
+    /* go left then right and check error */
+    this_mv.as_mv.row = startmv.as_mv.row;
+    this_mv.as_mv.col = ((startmv.as_mv.col - 8) | 4);
+    thismse = vfp->svf_halfpix_h(y - 1, y_stride, z, b->src_stride, &sse);
+    left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (left < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = left;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    this_mv.as_mv.col += 8;
+    thismse = vfp->svf_halfpix_h(y, y_stride, z, b->src_stride, &sse);
+    right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (right < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = right;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    /* go up then down and check error */
+    this_mv.as_mv.col = startmv.as_mv.col;
+    this_mv.as_mv.row = ((startmv.as_mv.row - 8) | 4);
+    thismse =  vfp->svf_halfpix_v(y - y_stride, y_stride, z, b->src_stride, &sse);
+    up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (up < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = up;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    this_mv.as_mv.row += 8;
+    thismse = vfp->svf_halfpix_v(y, y_stride, z, b->src_stride, &sse);
+    down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (down < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = down;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+
+    /* now check 1 more diagonal */
+    whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
+    this_mv = startmv;
+
+    switch (whichdir)
+    {
+    case 0:
+        this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
+        this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
+        thismse = vfp->svf_halfpix_hv(y - 1 - y_stride, y_stride, z, b->src_stride, &sse);
+        break;
+    case 1:
+        this_mv.as_mv.col += 4;
+        this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
+        thismse = vfp->svf_halfpix_hv(y - y_stride, y_stride, z, b->src_stride, &sse);
+        break;
+    case 2:
+        this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
+        this_mv.as_mv.row += 4;
+        thismse = vfp->svf_halfpix_hv(y - 1, y_stride, z, b->src_stride, &sse);
+        break;
+    case 3:
+    default:
+        this_mv.as_mv.col += 4;
+        this_mv.as_mv.row += 4;
+        thismse = vfp->svf_halfpix_hv(y, y_stride, z, b->src_stride, &sse);
+        break;
+    }
+
+    diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (diag < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = diag;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+
+    /* time to check quarter pels. */
+    if (bestmv->as_mv.row < startmv.as_mv.row)
+        y -= y_stride;
+
+    if (bestmv->as_mv.col < startmv.as_mv.col)
+        y--;
+
+    startmv = *bestmv;
+
+
+
+    /* go left then right and check error */
+    this_mv.as_mv.row = startmv.as_mv.row;
+
+    if (startmv.as_mv.col & 7)
+    {
+        this_mv.as_mv.col = startmv.as_mv.col - 2;
+        thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
+    }
+    else
+    {
+        this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
+        thismse = vfp->svf(y - 1, y_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
+    }
+
+    left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (left < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = left;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    this_mv.as_mv.col += 4;
+    thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
+    right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (right < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = right;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    /* go up then down and check error */
+    this_mv.as_mv.col = startmv.as_mv.col;
+
+    if (startmv.as_mv.row & 7)
+    {
+        this_mv.as_mv.row = startmv.as_mv.row - 2;
+        thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
+    }
+    else
+    {
+        this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
+        thismse = vfp->svf(y - y_stride, y_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
+    }
+
+    up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (up < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = up;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    this_mv.as_mv.row += 4;
+    thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
+    down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (down < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = down;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+
+    /* now check 1 more diagonal */
+    whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
+
+    this_mv = startmv;
+
+    switch (whichdir)
+    {
+    case 0:
+
+        if (startmv.as_mv.row & 7)
+        {
+            this_mv.as_mv.row -= 2;
+
+            if (startmv.as_mv.col & 7)
+            {
+                this_mv.as_mv.col -= 2;
+                thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
+            }
+            else
+            {
+                this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
+                thismse = vfp->svf(y - 1, y_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);;
+            }
+        }
+        else
+        {
+            this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
+
+            if (startmv.as_mv.col & 7)
+            {
+                this_mv.as_mv.col -= 2;
+                thismse = vfp->svf(y - y_stride, y_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
+            }
+            else
+            {
+                this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
+                thismse = vfp->svf(y - y_stride - 1, y_stride, 6, 6, z, b->src_stride, &sse);
+            }
+        }
+
+        break;
+    case 1:
+        this_mv.as_mv.col += 2;
+
+        if (startmv.as_mv.row & 7)
+        {
+            this_mv.as_mv.row -= 2;
+            thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
+        }
+        else
+        {
+            this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
+            thismse = vfp->svf(y - y_stride, y_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
+        }
+
+        break;
+    case 2:
+        this_mv.as_mv.row += 2;
+
+        if (startmv.as_mv.col & 7)
+        {
+            this_mv.as_mv.col -= 2;
+            thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
+        }
+        else
+        {
+            this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
+            thismse = vfp->svf(y - 1, y_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
+        }
+
+        break;
+    case 3:
+        this_mv.as_mv.col += 2;
+        this_mv.as_mv.row += 2;
+        thismse = vfp->svf(y, y_stride,  this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
+        break;
+    }
+
+    diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (diag < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = diag;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    return bestmse;
+}
+
+int vp8_find_best_half_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
+                                  int_mv *bestmv, int_mv *ref_mv,
+                                  int error_per_bit,
+                                  const vp8_variance_fn_ptr_t *vfp,
+                                  int *mvcost[2], int *distortion,
+                                  unsigned int *sse1)
+{
+    int bestmse = INT_MAX;
+    int_mv startmv;
+    int_mv this_mv;
+    unsigned char *z = (*(b->base_src) + b->src);
+    int left, right, up, down, diag;
+    unsigned int sse;
+    int whichdir ;
+    int thismse;
+    int y_stride;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+
+#if ARCH_X86 || ARCH_X86_64
+    MACROBLOCKD *xd = &x->e_mbd;
+    unsigned char *y_0 = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
+    unsigned char *y;
+
+    y_stride = 32;
+    /* Copy 18 rows x 32 cols area to intermediate buffer before searching. */
+    vfp->copymem(y_0 - 1 - pre_stride, pre_stride, xd->y_buf, y_stride, 18);
+    y = xd->y_buf + y_stride + 1;
+#else
+    unsigned char *y = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
+    y_stride = pre_stride;
+#endif
+
+    /* central mv */
+    bestmv->as_mv.row *= 8;
+    bestmv->as_mv.col *= 8;
+    startmv = *bestmv;
+
+    /* calculate central point error */
+    bestmse = vfp->vf(y, y_stride, z, b->src_stride, sse1);
+    *distortion = bestmse;
+    bestmse += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
+
+    /* go left then right and check error */
+    this_mv.as_mv.row = startmv.as_mv.row;
+    this_mv.as_mv.col = ((startmv.as_mv.col - 8) | 4);
+    thismse = vfp->svf_halfpix_h(y - 1, y_stride, z, b->src_stride, &sse);
+    left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (left < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = left;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    this_mv.as_mv.col += 8;
+    thismse = vfp->svf_halfpix_h(y, y_stride, z, b->src_stride, &sse);
+    right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (right < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = right;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    /* go up then down and check error */
+    this_mv.as_mv.col = startmv.as_mv.col;
+    this_mv.as_mv.row = ((startmv.as_mv.row - 8) | 4);
+    thismse = vfp->svf_halfpix_v(y - y_stride, y_stride, z, b->src_stride, &sse);
+    up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (up < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = up;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    this_mv.as_mv.row += 8;
+    thismse = vfp->svf_halfpix_v(y, y_stride, z, b->src_stride, &sse);
+    down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (down < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = down;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    /* now check 1 more diagonal - */
+    whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
+    this_mv = startmv;
+
+    switch (whichdir)
+    {
+    case 0:
+        this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
+        this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
+        thismse = vfp->svf_halfpix_hv(y - 1 - y_stride, y_stride, z, b->src_stride, &sse);
+        break;
+    case 1:
+        this_mv.as_mv.col += 4;
+        this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
+        thismse = vfp->svf_halfpix_hv(y - y_stride, y_stride, z, b->src_stride, &sse);
+        break;
+    case 2:
+        this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
+        this_mv.as_mv.row += 4;
+        thismse = vfp->svf_halfpix_hv(y - 1, y_stride, z, b->src_stride, &sse);
+        break;
+    case 3:
+    default:
+        this_mv.as_mv.col += 4;
+        this_mv.as_mv.row += 4;
+        thismse = vfp->svf_halfpix_hv(y, y_stride, z, b->src_stride, &sse);
+        break;
+    }
+
+    diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+
+    if (diag < bestmse)
+    {
+        *bestmv = this_mv;
+        bestmse = diag;
+        *distortion = thismse;
+        *sse1 = sse;
+    }
+
+    return bestmse;
+}
+
+#define CHECK_BOUNDS(range) \
+{\
+    all_in = 1;\
+    all_in &= ((br-range) >= x->mv_row_min);\
+    all_in &= ((br+range) <= x->mv_row_max);\
+    all_in &= ((bc-range) >= x->mv_col_min);\
+    all_in &= ((bc+range) <= x->mv_col_max);\
+}
+
+#define CHECK_POINT \
+{\
+    if (this_mv.as_mv.col < x->mv_col_min) continue;\
+    if (this_mv.as_mv.col > x->mv_col_max) continue;\
+    if (this_mv.as_mv.row < x->mv_row_min) continue;\
+    if (this_mv.as_mv.row > x->mv_row_max) continue;\
+}
+
+#define CHECK_BETTER \
+{\
+    if (thissad < bestsad)\
+    {\
+        thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, sad_per_bit);\
+        if (thissad < bestsad)\
+        {\
+            bestsad = thissad;\
+            best_site = i;\
+        }\
+    }\
+}
+
+static const MV next_chkpts[6][3] =
+{
+    {{ -2, 0}, { -1, -2}, {1, -2}},
+    {{ -1, -2}, {1, -2}, {2, 0}},
+    {{1, -2}, {2, 0}, {1, 2}},
+    {{2, 0}, {1, 2}, { -1, 2}},
+    {{1, 2}, { -1, 2}, { -2, 0}},
+    {{ -1, 2}, { -2, 0}, { -1, -2}}
+};
+
+int vp8_hex_search
+(
+    MACROBLOCK *x,
+    BLOCK *b,
+    BLOCKD *d,
+    int_mv *ref_mv,
+    int_mv *best_mv,
+    int search_param,
+    int sad_per_bit,
+    const vp8_variance_fn_ptr_t *vfp,
+    int *mvsadcost[2],
+    int *mvcost[2],
+    int_mv *center_mv
+)
+{
+    MV hex[6] = { { -1, -2}, {1, -2}, {2, 0}, {1, 2}, { -1, 2}, { -2, 0} } ;
+    MV neighbors[4] = {{0, -1}, { -1, 0}, {1, 0}, {0, 1}} ;
+    int i, j;
+
+    unsigned char *what = (*(b->base_src) + b->src);
+    int what_stride = b->src_stride;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+
+    int in_what_stride = pre_stride;
+    int br, bc;
+    int_mv this_mv;
+    unsigned int bestsad;
+    unsigned int thissad;
+    unsigned char *base_offset;
+    unsigned char *this_offset;
+    int k = -1;
+    int all_in;
+    int best_site = -1;
+    int hex_range = 127;
+    int dia_range = 8;
+
+    int_mv fcenter_mv;
+    fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
+    fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
+
+    (void)mvcost;
+
+    /* adjust ref_mv to make sure it is within MV range */
+    vp8_clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
+    br = ref_mv->as_mv.row;
+    bc = ref_mv->as_mv.col;
+
+    /* Work out the start point for the search */
+    base_offset = (unsigned char *)(base_pre + d->offset);
+    this_offset = base_offset + (br * (pre_stride)) + bc;
+    this_mv.as_mv.row = br;
+    this_mv.as_mv.col = bc;
+    bestsad = vfp->sdf(what, what_stride, this_offset, in_what_stride)
+            + mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, sad_per_bit);
+
+#if CONFIG_MULTI_RES_ENCODING
+    /* Lower search range based on prediction info */
+    if (search_param >= 6) goto cal_neighbors;
+    else if (search_param >= 5) hex_range = 4;
+    else if (search_param >= 4) hex_range = 6;
+    else if (search_param >= 3) hex_range = 15;
+    else if (search_param >= 2) hex_range = 31;
+    else if (search_param >= 1) hex_range = 63;
+
+    dia_range = 8;
+#else
+    (void)search_param;
+#endif
+
+    /* hex search */
+    CHECK_BOUNDS(2)
+
+    if(all_in)
+    {
+        for (i = 0; i < 6; i++)
+        {
+            this_mv.as_mv.row = br + hex[i].row;
+            this_mv.as_mv.col = bc + hex[i].col;
+            this_offset = base_offset + (this_mv.as_mv.row * in_what_stride) + this_mv.as_mv.col;
+            thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride);
+            CHECK_BETTER
+        }
+    }else
+    {
+        for (i = 0; i < 6; i++)
+        {
+            this_mv.as_mv.row = br + hex[i].row;
+            this_mv.as_mv.col = bc + hex[i].col;
+            CHECK_POINT
+            this_offset = base_offset + (this_mv.as_mv.row * in_what_stride) + this_mv.as_mv.col;
+            thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride);
+            CHECK_BETTER
+        }
+    }
+
+    if (best_site == -1)
+        goto cal_neighbors;
+    else
+    {
+        br += hex[best_site].row;
+        bc += hex[best_site].col;
+        k = best_site;
+    }
+
+    for (j = 1; j < hex_range; j++)
+    {
+        best_site = -1;
+        CHECK_BOUNDS(2)
+
+        if(all_in)
+        {
+            for (i = 0; i < 3; i++)
+            {
+                this_mv.as_mv.row = br + next_chkpts[k][i].row;
+                this_mv.as_mv.col = bc + next_chkpts[k][i].col;
+                this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
+                thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride);
+                CHECK_BETTER
+            }
+        }else
+        {
+            for (i = 0; i < 3; i++)
+            {
+                this_mv.as_mv.row = br + next_chkpts[k][i].row;
+                this_mv.as_mv.col = bc + next_chkpts[k][i].col;
+                CHECK_POINT
+                this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
+                thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride);
+                CHECK_BETTER
+            }
+        }
+
+        if (best_site == -1)
+            break;
+        else
+        {
+            br += next_chkpts[k][best_site].row;
+            bc += next_chkpts[k][best_site].col;
+            k += 5 + best_site;
+            if (k >= 12) k -= 12;
+            else if (k >= 6) k -= 6;
+        }
+    }
+
+    /* check 4 1-away neighbors */
+cal_neighbors:
+    for (j = 0; j < dia_range; j++)
+    {
+        best_site = -1;
+        CHECK_BOUNDS(1)
+
+        if(all_in)
+        {
+            for (i = 0; i < 4; i++)
+            {
+                this_mv.as_mv.row = br + neighbors[i].row;
+                this_mv.as_mv.col = bc + neighbors[i].col;
+                this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
+                thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride);
+                CHECK_BETTER
+            }
+        }else
+        {
+            for (i = 0; i < 4; i++)
+            {
+                this_mv.as_mv.row = br + neighbors[i].row;
+                this_mv.as_mv.col = bc + neighbors[i].col;
+                CHECK_POINT
+                this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
+                thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride);
+                CHECK_BETTER
+            }
+        }
+
+        if (best_site == -1)
+            break;
+        else
+        {
+            br += neighbors[best_site].row;
+            bc += neighbors[best_site].col;
+        }
+    }
+
+    best_mv->as_mv.row = br;
+    best_mv->as_mv.col = bc;
+
+    return bestsad;
+}
+#undef CHECK_BOUNDS
+#undef CHECK_POINT
+#undef CHECK_BETTER
+
+int vp8_diamond_search_sad_c
+(
+    MACROBLOCK *x,
+    BLOCK *b,
+    BLOCKD *d,
+    int_mv *ref_mv,
+    int_mv *best_mv,
+    int search_param,
+    int sad_per_bit,
+    int *num00,
+    vp8_variance_fn_ptr_t *fn_ptr,
+    int *mvcost[2],
+    int_mv *center_mv
+)
+{
+    int i, j, step;
+
+    unsigned char *what = (*(b->base_src) + b->src);
+    int what_stride = b->src_stride;
+    unsigned char *in_what;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+    int in_what_stride = pre_stride;
+    unsigned char *best_address;
+
+    int tot_steps;
+    int_mv this_mv;
+
+    unsigned int bestsad;
+    unsigned int thissad;
+    int best_site = 0;
+    int last_site = 0;
+
+    int ref_row;
+    int ref_col;
+    int this_row_offset;
+    int this_col_offset;
+    search_site *ss;
+
+    unsigned char *check_here;
+
+    int *mvsadcost[2];
+    int_mv fcenter_mv;
+
+    mvsadcost[0] = x->mvsadcost[0];
+    mvsadcost[1] = x->mvsadcost[1];
+    fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
+    fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
+
+    vp8_clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
+    ref_row = ref_mv->as_mv.row;
+    ref_col = ref_mv->as_mv.col;
+    *num00 = 0;
+    best_mv->as_mv.row = ref_row;
+    best_mv->as_mv.col = ref_col;
+
+    /* Work out the start point for the search */
+    in_what = (unsigned char *)(base_pre + d->offset + (ref_row * pre_stride) + ref_col);
+    best_address = in_what;
+
+    /* Check the starting position */
+    bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride)
+            + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
+
+    /* search_param determines the length of the initial step and hence
+     * the number of iterations 0 = initial step (MAX_FIRST_STEP) pel :
+     * 1 = (MAX_FIRST_STEP/2) pel, 2 = (MAX_FIRST_STEP/4) pel... etc.
+     */
+    ss = &x->ss[search_param * x->searches_per_step];
+    tot_steps = (x->ss_count / x->searches_per_step) - search_param;
+
+    i = 1;
+
+    for (step = 0; step < tot_steps ; step++)
+    {
+        for (j = 0 ; j < x->searches_per_step ; j++)
+        {
+            /* Trap illegal vectors */
+            this_row_offset = best_mv->as_mv.row + ss[i].mv.row;
+            this_col_offset = best_mv->as_mv.col + ss[i].mv.col;
+
+            if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
+            (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
+
+            {
+                check_here = ss[i].offset + best_address;
+                thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
+
+                if (thissad < bestsad)
+                {
+                    this_mv.as_mv.row = this_row_offset;
+                    this_mv.as_mv.col = this_col_offset;
+                    thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
+                                              mvsadcost, sad_per_bit);
+
+                    if (thissad < bestsad)
+                    {
+                        bestsad = thissad;
+                        best_site = i;
+                    }
+                }
+            }
+
+            i++;
+        }
+
+        if (best_site != last_site)
+        {
+            best_mv->as_mv.row += ss[best_site].mv.row;
+            best_mv->as_mv.col += ss[best_site].mv.col;
+            best_address += ss[best_site].offset;
+            last_site = best_site;
+        }
+        else if (best_address == in_what)
+            (*num00)++;
+    }
+
+    this_mv.as_mv.row = best_mv->as_mv.row << 3;
+    this_mv.as_mv.col = best_mv->as_mv.col << 3;
+
+    return fn_ptr->vf(what, what_stride, best_address, in_what_stride, &thissad)
+           + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
+}
+
+int vp8_diamond_search_sadx4
+(
+    MACROBLOCK *x,
+    BLOCK *b,
+    BLOCKD *d,
+    int_mv *ref_mv,
+    int_mv *best_mv,
+    int search_param,
+    int sad_per_bit,
+    int *num00,
+    vp8_variance_fn_ptr_t *fn_ptr,
+    int *mvcost[2],
+    int_mv *center_mv
+)
+{
+    int i, j, step;
+
+    unsigned char *what = (*(b->base_src) + b->src);
+    int what_stride = b->src_stride;
+    unsigned char *in_what;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+    int in_what_stride = pre_stride;
+    unsigned char *best_address;
+
+    int tot_steps;
+    int_mv this_mv;
+
+    unsigned int bestsad;
+    unsigned int thissad;
+    int best_site = 0;
+    int last_site = 0;
+
+    int ref_row;
+    int ref_col;
+    int this_row_offset;
+    int this_col_offset;
+    search_site *ss;
+
+    unsigned char *check_here;
+
+    int *mvsadcost[2];
+    int_mv fcenter_mv;
+
+    mvsadcost[0] = x->mvsadcost[0];
+    mvsadcost[1] = x->mvsadcost[1];
+    fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
+    fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
+
+    vp8_clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
+    ref_row = ref_mv->as_mv.row;
+    ref_col = ref_mv->as_mv.col;
+    *num00 = 0;
+    best_mv->as_mv.row = ref_row;
+    best_mv->as_mv.col = ref_col;
+
+    /* Work out the start point for the search */
+    in_what = (unsigned char *)(base_pre + d->offset + (ref_row * pre_stride) + ref_col);
+    best_address = in_what;
+
+    /* Check the starting position */
+    bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride)
+            + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
+
+    /* search_param determines the length of the initial step and hence the
+     * number of iterations 0 = initial step (MAX_FIRST_STEP) pel : 1 =
+     * (MAX_FIRST_STEP/2) pel, 2 = (MAX_FIRST_STEP/4) pel... etc.
+     */
+    ss = &x->ss[search_param * x->searches_per_step];
+    tot_steps = (x->ss_count / x->searches_per_step) - search_param;
+
+    i = 1;
+
+    for (step = 0; step < tot_steps ; step++)
+    {
+        int all_in = 1, t;
+
+        /* To know if all neighbor points are within the bounds, 4 bounds
+         * checking are enough instead of checking 4 bounds for each
+         * points.
+         */
+        all_in &= ((best_mv->as_mv.row + ss[i].mv.row)> x->mv_row_min);
+        all_in &= ((best_mv->as_mv.row + ss[i+1].mv.row) < x->mv_row_max);
+        all_in &= ((best_mv->as_mv.col + ss[i+2].mv.col) > x->mv_col_min);
+        all_in &= ((best_mv->as_mv.col + ss[i+3].mv.col) < x->mv_col_max);
+
+        if (all_in)
+        {
+            unsigned int sad_array[4];
+
+            for (j = 0 ; j < x->searches_per_step ; j += 4)
+            {
+                const unsigned char *block_offset[4];
+
+                for (t = 0; t < 4; t++)
+                    block_offset[t] = ss[i+t].offset + best_address;
+
+                fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride, sad_array);
+
+                for (t = 0; t < 4; t++, i++)
+                {
+                    if (sad_array[t] < bestsad)
+                    {
+                        this_mv.as_mv.row = best_mv->as_mv.row + ss[i].mv.row;
+                        this_mv.as_mv.col = best_mv->as_mv.col + ss[i].mv.col;
+                        sad_array[t] += mvsad_err_cost(&this_mv, &fcenter_mv,
+                                                       mvsadcost, sad_per_bit);
+
+                        if (sad_array[t] < bestsad)
+                        {
+                            bestsad = sad_array[t];
+                            best_site = i;
+                        }
+                    }
+                }
+            }
+        }
+        else
+        {
+            for (j = 0 ; j < x->searches_per_step ; j++)
+            {
+                /* Trap illegal vectors */
+                this_row_offset = best_mv->as_mv.row + ss[i].mv.row;
+                this_col_offset = best_mv->as_mv.col + ss[i].mv.col;
+
+                if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
+                (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
+                {
+                    check_here = ss[i].offset + best_address;
+                    thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
+
+                    if (thissad < bestsad)
+                    {
+                        this_mv.as_mv.row = this_row_offset;
+                        this_mv.as_mv.col = this_col_offset;
+                        thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
+                                                  mvsadcost, sad_per_bit);
+
+                        if (thissad < bestsad)
+                        {
+                            bestsad = thissad;
+                            best_site = i;
+                        }
+                    }
+                }
+                i++;
+            }
+        }
+
+        if (best_site != last_site)
+        {
+            best_mv->as_mv.row += ss[best_site].mv.row;
+            best_mv->as_mv.col += ss[best_site].mv.col;
+            best_address += ss[best_site].offset;
+            last_site = best_site;
+        }
+        else if (best_address == in_what)
+            (*num00)++;
+    }
+
+    this_mv.as_mv.row = best_mv->as_mv.row * 8;
+    this_mv.as_mv.col = best_mv->as_mv.col * 8;
+
+    return fn_ptr->vf(what, what_stride, best_address, in_what_stride, &thissad)
+           + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
+}
+
+int vp8_full_search_sad_c(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
+                        int sad_per_bit, int distance,
+                        vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
+                        int_mv *center_mv)
+{
+    unsigned char *what = (*(b->base_src) + b->src);
+    int what_stride = b->src_stride;
+    unsigned char *in_what;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+    int in_what_stride = pre_stride;
+    int mv_stride = pre_stride;
+    unsigned char *bestaddress;
+    int_mv *best_mv = &d->bmi.mv;
+    int_mv this_mv;
+    unsigned int bestsad;
+    unsigned int thissad;
+    int r, c;
+
+    unsigned char *check_here;
+
+    int ref_row = ref_mv->as_mv.row;
+    int ref_col = ref_mv->as_mv.col;
+
+    int row_min = ref_row - distance;
+    int row_max = ref_row + distance;
+    int col_min = ref_col - distance;
+    int col_max = ref_col + distance;
+
+    int *mvsadcost[2];
+    int_mv fcenter_mv;
+
+    mvsadcost[0] = x->mvsadcost[0];
+    mvsadcost[1] = x->mvsadcost[1];
+    fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
+    fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
+
+    /* Work out the mid point for the search */
+    in_what = base_pre + d->offset;
+    bestaddress = in_what + (ref_row * pre_stride) + ref_col;
+
+    best_mv->as_mv.row = ref_row;
+    best_mv->as_mv.col = ref_col;
+
+    /* Baseline value at the centre */
+    bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride)
+            + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
+
+    /* Apply further limits to prevent us looking using vectors that
+     * stretch beyiond the UMV border
+     */
+    if (col_min < x->mv_col_min)
+        col_min = x->mv_col_min;
+
+    if (col_max > x->mv_col_max)
+        col_max = x->mv_col_max;
+
+    if (row_min < x->mv_row_min)
+        row_min = x->mv_row_min;
+
+    if (row_max > x->mv_row_max)
+        row_max = x->mv_row_max;
+
+    for (r = row_min; r < row_max ; r++)
+    {
+        this_mv.as_mv.row = r;
+        check_here = r * mv_stride + in_what + col_min;
+
+        for (c = col_min; c < col_max; c++)
+        {
+            thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
+
+            this_mv.as_mv.col = c;
+            thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
+                                      mvsadcost, sad_per_bit);
+
+            if (thissad < bestsad)
+            {
+                bestsad = thissad;
+                best_mv->as_mv.row = r;
+                best_mv->as_mv.col = c;
+                bestaddress = check_here;
+            }
+
+            check_here++;
+        }
+    }
+
+    this_mv.as_mv.row = best_mv->as_mv.row << 3;
+    this_mv.as_mv.col = best_mv->as_mv.col << 3;
+
+    return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, &thissad)
+           + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
+}
+
+int vp8_full_search_sadx3(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
+                          int sad_per_bit, int distance,
+                          vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
+                          int_mv *center_mv)
+{
+    unsigned char *what = (*(b->base_src) + b->src);
+    int what_stride = b->src_stride;
+    unsigned char *in_what;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+    int in_what_stride = pre_stride;
+    int mv_stride = pre_stride;
+    unsigned char *bestaddress;
+    int_mv *best_mv = &d->bmi.mv;
+    int_mv this_mv;
+    unsigned int bestsad;
+    unsigned int thissad;
+    int r, c;
+
+    unsigned char *check_here;
+
+    int ref_row = ref_mv->as_mv.row;
+    int ref_col = ref_mv->as_mv.col;
+
+    int row_min = ref_row - distance;
+    int row_max = ref_row + distance;
+    int col_min = ref_col - distance;
+    int col_max = ref_col + distance;
+
+    unsigned int sad_array[3];
+
+    int *mvsadcost[2];
+    int_mv fcenter_mv;
+
+    mvsadcost[0] = x->mvsadcost[0];
+    mvsadcost[1] = x->mvsadcost[1];
+    fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
+    fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
+
+    /* Work out the mid point for the search */
+    in_what = base_pre + d->offset;
+    bestaddress = in_what + (ref_row * pre_stride) + ref_col;
+
+    best_mv->as_mv.row = ref_row;
+    best_mv->as_mv.col = ref_col;
+
+    /* Baseline value at the centre */
+    bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride)
+            + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
+
+    /* Apply further limits to prevent us looking using vectors that stretch
+     * beyond the UMV border
+     */
+    if (col_min < x->mv_col_min)
+        col_min = x->mv_col_min;
+
+    if (col_max > x->mv_col_max)
+        col_max = x->mv_col_max;
+
+    if (row_min < x->mv_row_min)
+        row_min = x->mv_row_min;
+
+    if (row_max > x->mv_row_max)
+        row_max = x->mv_row_max;
+
+    for (r = row_min; r < row_max ; r++)
+    {
+        this_mv.as_mv.row = r;
+        check_here = r * mv_stride + in_what + col_min;
+        c = col_min;
+
+        while ((c + 2) < col_max)
+        {
+            int i;
+
+            fn_ptr->sdx3f(what, what_stride, check_here, in_what_stride, sad_array);
+
+            for (i = 0; i < 3; i++)
+            {
+                thissad = sad_array[i];
+
+                if (thissad < bestsad)
+                {
+                    this_mv.as_mv.col = c;
+                    thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
+                                              mvsadcost, sad_per_bit);
+
+                    if (thissad < bestsad)
+                    {
+                        bestsad = thissad;
+                        best_mv->as_mv.row = r;
+                        best_mv->as_mv.col = c;
+                        bestaddress = check_here;
+                    }
+                }
+
+                check_here++;
+                c++;
+            }
+        }
+
+        while (c < col_max)
+        {
+            thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
+
+            if (thissad < bestsad)
+            {
+                this_mv.as_mv.col = c;
+                thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
+                                          mvsadcost, sad_per_bit);
+
+                if (thissad < bestsad)
+                {
+                    bestsad = thissad;
+                    best_mv->as_mv.row = r;
+                    best_mv->as_mv.col = c;
+                    bestaddress = check_here;
+                }
+            }
+
+            check_here ++;
+            c ++;
+        }
+
+    }
+
+    this_mv.as_mv.row = best_mv->as_mv.row << 3;
+    this_mv.as_mv.col = best_mv->as_mv.col << 3;
+
+    return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, &thissad)
+           + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
+}
+
+int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
+                          int sad_per_bit, int distance,
+                          vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
+                          int_mv *center_mv)
+{
+    unsigned char *what = (*(b->base_src) + b->src);
+    int what_stride = b->src_stride;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+    unsigned char *in_what;
+    int in_what_stride = pre_stride;
+    int mv_stride = pre_stride;
+    unsigned char *bestaddress;
+    int_mv *best_mv = &d->bmi.mv;
+    int_mv this_mv;
+    unsigned int bestsad;
+    unsigned int thissad;
+    int r, c;
+
+    unsigned char *check_here;
+
+    int ref_row = ref_mv->as_mv.row;
+    int ref_col = ref_mv->as_mv.col;
+
+    int row_min = ref_row - distance;
+    int row_max = ref_row + distance;
+    int col_min = ref_col - distance;
+    int col_max = ref_col + distance;
+
+    DECLARE_ALIGNED(16, unsigned int, sad_array8[8]);
+    unsigned int sad_array[3];
+
+    int *mvsadcost[2];
+    int_mv fcenter_mv;
+
+    mvsadcost[0] = x->mvsadcost[0];
+    mvsadcost[1] = x->mvsadcost[1];
+    fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
+    fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
+
+    /* Work out the mid point for the search */
+    in_what = base_pre + d->offset;
+    bestaddress = in_what + (ref_row * pre_stride) + ref_col;
+
+    best_mv->as_mv.row = ref_row;
+    best_mv->as_mv.col = ref_col;
+
+    /* Baseline value at the centre */
+    bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride)
+            + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
+
+    /* Apply further limits to prevent us looking using vectors that stretch
+     * beyond the UMV border
+     */
+    if (col_min < x->mv_col_min)
+        col_min = x->mv_col_min;
+
+    if (col_max > x->mv_col_max)
+        col_max = x->mv_col_max;
+
+    if (row_min < x->mv_row_min)
+        row_min = x->mv_row_min;
+
+    if (row_max > x->mv_row_max)
+        row_max = x->mv_row_max;
+
+    for (r = row_min; r < row_max ; r++)
+    {
+        this_mv.as_mv.row = r;
+        check_here = r * mv_stride + in_what + col_min;
+        c = col_min;
+
+        while ((c + 7) < col_max)
+        {
+            int i;
+
+            fn_ptr->sdx8f(what, what_stride, check_here, in_what_stride, sad_array8);
+
+            for (i = 0; i < 8; i++)
+            {
+                thissad = sad_array8[i];
+
+                if (thissad < bestsad)
+                {
+                    this_mv.as_mv.col = c;
+                    thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
+                                              mvsadcost, sad_per_bit);
+
+                    if (thissad < bestsad)
+                    {
+                        bestsad = thissad;
+                        best_mv->as_mv.row = r;
+                        best_mv->as_mv.col = c;
+                        bestaddress = check_here;
+                    }
+                }
+
+                check_here++;
+                c++;
+            }
+        }
+
+        while ((c + 2) < col_max)
+        {
+            int i;
+
+            fn_ptr->sdx3f(what, what_stride, check_here , in_what_stride, sad_array);
+
+            for (i = 0; i < 3; i++)
+            {
+                thissad = sad_array[i];
+
+                if (thissad < bestsad)
+                {
+                    this_mv.as_mv.col = c;
+                    thissad  += mvsad_err_cost(&this_mv, &fcenter_mv,
+                        mvsadcost, sad_per_bit);
+
+                    if (thissad < bestsad)
+                    {
+                        bestsad = thissad;
+                        best_mv->as_mv.row = r;
+                        best_mv->as_mv.col = c;
+                        bestaddress = check_here;
+                    }
+                }
+
+                check_here++;
+                c++;
+            }
+        }
+
+        while (c < col_max)
+        {
+            thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride);
+
+            if (thissad < bestsad)
+            {
+                this_mv.as_mv.col = c;
+                thissad  += mvsad_err_cost(&this_mv, &fcenter_mv,
+                    mvsadcost, sad_per_bit);
+
+                if (thissad < bestsad)
+                {
+                    bestsad = thissad;
+                    best_mv->as_mv.row = r;
+                    best_mv->as_mv.col = c;
+                    bestaddress = check_here;
+                }
+            }
+
+            check_here ++;
+            c ++;
+        }
+    }
+
+    this_mv.as_mv.row = best_mv->as_mv.row * 8;
+    this_mv.as_mv.col = best_mv->as_mv.col * 8;
+
+    return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, &thissad)
+           + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
+}
+
+int vp8_refining_search_sad_c(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
+                            int error_per_bit, int search_range,
+                            vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
+                            int_mv *center_mv)
+{
+    MV neighbors[4] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0}};
+    int i, j;
+    short this_row_offset, this_col_offset;
+
+    int what_stride = b->src_stride;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+    int in_what_stride = pre_stride;
+    unsigned char *what = (*(b->base_src) + b->src);
+    unsigned char *best_address = (unsigned char *)(base_pre + d->offset +
+        (ref_mv->as_mv.row * pre_stride) + ref_mv->as_mv.col);
+    unsigned char *check_here;
+    int_mv this_mv;
+    unsigned int bestsad;
+    unsigned int thissad;
+
+    int *mvsadcost[2];
+    int_mv fcenter_mv;
+
+    mvsadcost[0] = x->mvsadcost[0];
+    mvsadcost[1] = x->mvsadcost[1];
+    fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
+    fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
+
+    bestsad = fn_ptr->sdf(what, what_stride, best_address, in_what_stride)
+            + mvsad_err_cost(ref_mv, &fcenter_mv, mvsadcost, error_per_bit);
+
+    for (i=0; i<search_range; i++)
+    {
+        int best_site = -1;
+
+        for (j = 0 ; j < 4 ; j++)
+        {
+            this_row_offset = ref_mv->as_mv.row + neighbors[j].row;
+            this_col_offset = ref_mv->as_mv.col + neighbors[j].col;
+
+            if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
+            (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
+            {
+                check_here = (neighbors[j].row)*in_what_stride + neighbors[j].col + best_address;
+                thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride);
+
+                if (thissad < bestsad)
+                {
+                    this_mv.as_mv.row = this_row_offset;
+                    this_mv.as_mv.col = this_col_offset;
+                    thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
+
+                    if (thissad < bestsad)
+                    {
+                        bestsad = thissad;
+                        best_site = j;
+                    }
+                }
+            }
+        }
+
+        if (best_site == -1)
+            break;
+        else
+        {
+            ref_mv->as_mv.row += neighbors[best_site].row;
+            ref_mv->as_mv.col += neighbors[best_site].col;
+            best_address += (neighbors[best_site].row)*in_what_stride + neighbors[best_site].col;
+        }
+    }
+
+    this_mv.as_mv.row = ref_mv->as_mv.row << 3;
+    this_mv.as_mv.col = ref_mv->as_mv.col << 3;
+
+    return fn_ptr->vf(what, what_stride, best_address, in_what_stride, &thissad)
+           + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
+}
+
+int vp8_refining_search_sadx4(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
+                              int_mv *ref_mv, int error_per_bit,
+                              int search_range, vp8_variance_fn_ptr_t *fn_ptr,
+                              int *mvcost[2], int_mv *center_mv)
+{
+    MV neighbors[4] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0}};
+    int i, j;
+    short this_row_offset, this_col_offset;
+
+    int what_stride = b->src_stride;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+    int in_what_stride = pre_stride;
+    unsigned char *what = (*(b->base_src) + b->src);
+    unsigned char *best_address = (unsigned char *)(base_pre + d->offset +
+        (ref_mv->as_mv.row * pre_stride) + ref_mv->as_mv.col);
+    unsigned char *check_here;
+    int_mv this_mv;
+    unsigned int bestsad;
+    unsigned int thissad;
+
+    int *mvsadcost[2];
+    int_mv fcenter_mv;
+
+    mvsadcost[0] = x->mvsadcost[0];
+    mvsadcost[1] = x->mvsadcost[1];
+    fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
+    fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
+
+    bestsad = fn_ptr->sdf(what, what_stride, best_address, in_what_stride)
+            + mvsad_err_cost(ref_mv, &fcenter_mv, mvsadcost, error_per_bit);
+
+    for (i=0; i<search_range; i++)
+    {
+        int best_site = -1;
+        int all_in = 1;
+
+        all_in &= ((ref_mv->as_mv.row - 1) > x->mv_row_min);
+        all_in &= ((ref_mv->as_mv.row + 1) < x->mv_row_max);
+        all_in &= ((ref_mv->as_mv.col - 1) > x->mv_col_min);
+        all_in &= ((ref_mv->as_mv.col + 1) < x->mv_col_max);
+
+        if(all_in)
+        {
+            unsigned int sad_array[4];
+            const unsigned char *block_offset[4];
+            block_offset[0] = best_address - in_what_stride;
+            block_offset[1] = best_address - 1;
+            block_offset[2] = best_address + 1;
+            block_offset[3] = best_address + in_what_stride;
+
+            fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride, sad_array);
+
+            for (j = 0; j < 4; j++)
+            {
+                if (sad_array[j] < bestsad)
+                {
+                    this_mv.as_mv.row = ref_mv->as_mv.row + neighbors[j].row;
+                    this_mv.as_mv.col = ref_mv->as_mv.col + neighbors[j].col;
+                    sad_array[j] += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
+
+                    if (sad_array[j] < bestsad)
+                    {
+                        bestsad = sad_array[j];
+                        best_site = j;
+                    }
+                }
+            }
+        }
+        else
+        {
+            for (j = 0 ; j < 4 ; j++)
+            {
+                this_row_offset = ref_mv->as_mv.row + neighbors[j].row;
+                this_col_offset = ref_mv->as_mv.col + neighbors[j].col;
+
+                if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
+                (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
+                {
+                    check_here = (neighbors[j].row)*in_what_stride + neighbors[j].col + best_address;
+                    thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride);
+
+                    if (thissad < bestsad)
+                    {
+                        this_mv.as_mv.row = this_row_offset;
+                        this_mv.as_mv.col = this_col_offset;
+                        thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
+
+                        if (thissad < bestsad)
+                        {
+                            bestsad = thissad;
+                            best_site = j;
+                        }
+                    }
+                }
+            }
+        }
+
+        if (best_site == -1)
+            break;
+        else
+        {
+            ref_mv->as_mv.row += neighbors[best_site].row;
+            ref_mv->as_mv.col += neighbors[best_site].col;
+            best_address += (neighbors[best_site].row)*in_what_stride + neighbors[best_site].col;
+        }
+    }
+
+    this_mv.as_mv.row = ref_mv->as_mv.row * 8;
+    this_mv.as_mv.col = ref_mv->as_mv.col * 8;
+
+    return fn_ptr->vf(what, what_stride, best_address, in_what_stride, &thissad)
+           + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
+}
+
+#ifdef VP8_ENTROPY_STATS
+void print_mode_context(void)
+{
+    FILE *f = fopen("modecont.c", "w");
+    int i, j;
+
+    fprintf(f, "#include \"entropy.h\"\n");
+    fprintf(f, "const int vp8_mode_contexts[6][4] =\n");
+    fprintf(f, "{\n");
+
+    for (j = 0; j < 6; j++)
+    {
+        fprintf(f, "  { /* %d */\n", j);
+        fprintf(f, "    ");
+
+        for (i = 0; i < 4; i++)
+        {
+            int overal_prob;
+            int this_prob;
+            int count;
+
+            /* Overall probs */
+            count = mv_mode_cts[i][0] + mv_mode_cts[i][1];
+
+            if (count)
+                overal_prob = 256 * mv_mode_cts[i][0] / count;
+            else
+                overal_prob = 128;
+
+            if (overal_prob == 0)
+                overal_prob = 1;
+
+            /* context probs */
+            count = mv_ref_ct[j][i][0] + mv_ref_ct[j][i][1];
+
+            if (count)
+                this_prob = 256 * mv_ref_ct[j][i][0] / count;
+            else
+                this_prob = 128;
+
+            if (this_prob == 0)
+                this_prob = 1;
+
+            fprintf(f, "%5d, ", this_prob);
+        }
+
+        fprintf(f, "  },\n");
+    }
+
+    fprintf(f, "};\n");
+    fclose(f);
+}
+
+/* MV ref count VP8_ENTROPY_STATS stats code */
+#ifdef VP8_ENTROPY_STATS
+void init_mv_ref_counts()
+{
+    memset(mv_ref_ct, 0, sizeof(mv_ref_ct));
+    memset(mv_mode_cts, 0, sizeof(mv_mode_cts));
+}
+
+void accum_mv_refs(MB_PREDICTION_MODE m, const int ct[4])
+{
+    if (m == ZEROMV)
+    {
+        ++mv_ref_ct [ct[0]] [0] [0];
+        ++mv_mode_cts[0][0];
+    }
+    else
+    {
+        ++mv_ref_ct [ct[0]] [0] [1];
+        ++mv_mode_cts[0][1];
+
+        if (m == NEARESTMV)
+        {
+            ++mv_ref_ct [ct[1]] [1] [0];
+            ++mv_mode_cts[1][0];
+        }
+        else
+        {
+            ++mv_ref_ct [ct[1]] [1] [1];
+            ++mv_mode_cts[1][1];
+
+            if (m == NEARMV)
+            {
+                ++mv_ref_ct [ct[2]] [2] [0];
+                ++mv_mode_cts[2][0];
+            }
+            else
+            {
+                ++mv_ref_ct [ct[2]] [2] [1];
+                ++mv_mode_cts[2][1];
+
+                if (m == NEWMV)
+                {
+                    ++mv_ref_ct [ct[3]] [3] [0];
+                    ++mv_mode_cts[3][0];
+                }
+                else
+                {
+                    ++mv_ref_ct [ct[3]] [3] [1];
+                    ++mv_mode_cts[3][1];
+                }
+            }
+        }
+    }
+}
+
+#endif/* END MV ref count VP8_ENTROPY_STATS stats code */
+
+#endif
diff --git a/libvpx/libvpx/vp8/encoder/mcomp.h b/libvpx/libvpx/vp8/encoder/mcomp.h
new file mode 100644
index 0000000..1694af8
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/mcomp.h
@@ -0,0 +1,115 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_MCOMP_H_
+#define VP8_ENCODER_MCOMP_H_
+
+#include "block.h"
+#include "vpx_dsp/variance.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef VP8_ENTROPY_STATS
+extern void init_mv_ref_counts();
+extern void accum_mv_refs(MB_PREDICTION_MODE, const int near_mv_ref_cts[4]);
+#endif
+
+
+/* The maximum number of steps in a step search given the largest allowed
+ * initial step
+ */
+#define MAX_MVSEARCH_STEPS 8
+
+/* Max full pel mv specified in 1 pel units */
+#define MAX_FULL_PEL_VAL ((1 << (MAX_MVSEARCH_STEPS)) - 1)
+
+/* Maximum size of the first step in full pel units */
+#define MAX_FIRST_STEP (1 << (MAX_MVSEARCH_STEPS-1))
+
+extern void print_mode_context(void);
+extern int vp8_mv_bit_cost(int_mv *mv, int_mv *ref, int *mvcost[2], int Weight);
+extern void vp8_init_dsmotion_compensation(MACROBLOCK *x, int stride);
+extern void vp8_init3smotion_compensation(MACROBLOCK *x,  int stride);
+
+
+extern int vp8_hex_search
+(
+    MACROBLOCK *x,
+    BLOCK *b,
+    BLOCKD *d,
+    int_mv *ref_mv,
+    int_mv *best_mv,
+    int search_param,
+    int error_per_bit,
+    const vp8_variance_fn_ptr_t *vf,
+    int *mvsadcost[2],
+    int *mvcost[2],
+    int_mv *center_mv
+);
+
+typedef int (fractional_mv_step_fp)
+    (MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *bestmv, int_mv *ref_mv,
+     int error_per_bit, const vp8_variance_fn_ptr_t *vfp, int *mvcost[2],
+     int *distortion, unsigned int *sse);
+
+extern fractional_mv_step_fp vp8_find_best_sub_pixel_step_iteratively;
+extern fractional_mv_step_fp vp8_find_best_sub_pixel_step;
+extern fractional_mv_step_fp vp8_find_best_half_pixel_step;
+extern fractional_mv_step_fp vp8_skip_fractional_mv_step;
+
+typedef int (*vp8_full_search_fn_t)
+    (
+     MACROBLOCK *x,
+     BLOCK *b,
+     BLOCKD *d,
+     int_mv *ref_mv,
+     int sad_per_bit,
+     int distance,
+     vp8_variance_fn_ptr_t *fn_ptr,
+     int *mvcost[2],
+     int_mv *center_mv
+    );
+
+typedef int (*vp8_refining_search_fn_t)
+    (
+     MACROBLOCK *x,
+     BLOCK *b,
+     BLOCKD *d,
+     int_mv *ref_mv,
+     int sad_per_bit,
+     int distance,
+     vp8_variance_fn_ptr_t *fn_ptr,
+     int *mvcost[2],
+     int_mv *center_mv
+    );
+
+typedef int (*vp8_diamond_search_fn_t)
+    (
+     MACROBLOCK *x,
+     BLOCK *b,
+     BLOCKD *d,
+     int_mv *ref_mv,
+     int_mv *best_mv,
+     int search_param,
+     int sad_per_bit,
+     int *num00,
+     vp8_variance_fn_ptr_t *fn_ptr,
+     int *mvcost[2],
+     int_mv *center_mv
+    );
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_MCOMP_H_
diff --git a/libvpx/libvpx/vp8/encoder/mips/msa/dct_msa.c b/libvpx/libvpx/vp8/encoder/mips/msa/dct_msa.c
new file mode 100644
index 0000000..be61ffa
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/mips/msa/dct_msa.c
@@ -0,0 +1,199 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+
+#define TRANSPOSE4x4_H(in0, in1, in2, in3, out0, out1, out2, out3)  \
+{                                                                   \
+    v8i16 s0_m, s1_m, tp0_m, tp1_m, tp2_m, tp3_m;                   \
+                                                                    \
+    ILVR_H2_SH(in2, in0, in3, in1, s0_m, s1_m);                     \
+    ILVRL_H2_SH(s1_m, s0_m, tp0_m, tp1_m);                          \
+    ILVL_H2_SH(in2, in0, in3, in1, s0_m, s1_m);                     \
+    ILVRL_H2_SH(s1_m, s0_m, tp2_m, tp3_m);                          \
+    PCKEV_D2_SH(tp2_m, tp0_m, tp3_m, tp1_m, out0, out2);            \
+    PCKOD_D2_SH(tp2_m, tp0_m, tp3_m, tp1_m, out1, out3);            \
+}
+
+#define SET_DOTP_VALUES(coeff, val0, val1, val2, const1, const2)    \
+{                                                                   \
+    v8i16 tmp0_m;                                                   \
+                                                                    \
+    SPLATI_H3_SH(coeff, val0, val1, val2, tmp0_m, const1, const2);  \
+    ILVEV_H2_SH(tmp0_m, const1, const2, tmp0_m, const1, const2);    \
+}
+
+#define RET_1_IF_NZERO_H(in0)       \
+({                                  \
+    v8i16 tmp0_m;                   \
+    v8i16 one_m = __msa_ldi_h(1);   \
+                                    \
+    tmp0_m = __msa_ceqi_h(in0, 0);  \
+    tmp0_m = tmp0_m ^ 255;          \
+    tmp0_m = one_m & tmp0_m;        \
+                                    \
+    tmp0_m;                         \
+})
+
+#define RET_1_IF_NZERO_W(in0)       \
+({                                  \
+    v4i32 tmp0_m;                   \
+    v4i32 one_m = __msa_ldi_w(1);   \
+                                    \
+    tmp0_m = __msa_ceqi_w(in0, 0);  \
+    tmp0_m = tmp0_m ^ 255;          \
+    tmp0_m = one_m & tmp0_m;        \
+                                    \
+    tmp0_m;                         \
+})
+
+#define RET_1_IF_NEG_W(in0)           \
+({                                    \
+    v4i32 tmp0_m;                     \
+                                      \
+    v4i32 one_m = __msa_ldi_w(1);     \
+    tmp0_m = __msa_clti_s_w(in0, 0);  \
+    tmp0_m = one_m & tmp0_m;          \
+                                      \
+    tmp0_m;                           \
+})
+
+void vp8_short_fdct4x4_msa(int16_t *input, int16_t *output, int32_t pitch)
+{
+    v8i16 in0, in1, in2, in3;
+    v8i16 temp0, temp1;
+    v8i16 const0, const1;
+    v8i16 coeff = { 2217, 5352, -5352, 14500, 7500, 12000, 25000, 26000 };
+    v4i32 out0, out1, out2, out3;
+    v8i16 zero = { 0 };
+
+    LD_SH4(input, pitch / 2, in0, in1, in2, in3);
+    TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+
+    BUTTERFLY_4(in0, in1, in2, in3, temp0, temp1, in1, in3);
+    SLLI_4V(temp0, temp1, in1, in3, 3);
+    in0 = temp0 + temp1;
+    in2 = temp0 - temp1;
+    SET_DOTP_VALUES(coeff, 0, 1, 2, const0, const1);
+    temp0 = __msa_ilvr_h(in3, in1);
+    in1 = __msa_splati_h(coeff, 3);
+    out0 = (v4i32)__msa_ilvev_h(zero, in1);
+    coeff = __msa_ilvl_h(zero, coeff);
+    out1 = __msa_splati_w((v4i32)coeff, 0);
+    DPADD_SH2_SW(temp0, temp0, const0, const1, out0, out1);
+    out0 >>= 12;
+    out1 >>= 12;
+    PCKEV_H2_SH(out0, out0, out1, out1, in1, in3);
+    TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+
+    BUTTERFLY_4(in0, in1, in2, in3, temp0, temp1, in1, in3);
+    in0 = temp0 + temp1 + 7;
+    in2 = temp0 - temp1 + 7;
+    in0 >>= 4;
+    in2 >>= 4;
+    ILVR_H2_SW(zero, in0, zero, in2, out0, out2);
+    temp1 = RET_1_IF_NZERO_H(in3);
+    ILVR_H2_SH(zero, temp1, in3, in1, temp1, temp0);
+    SPLATI_W2_SW(coeff, 2, out3, out1);
+    out3 += out1;
+    out1 = __msa_splati_w((v4i32)coeff, 1);
+    DPADD_SH2_SW(temp0, temp0, const0, const1, out1, out3);
+    out1 >>= 16;
+    out3 >>= 16;
+    out1 += (v4i32)temp1;
+    PCKEV_H2_SH(out1, out0, out3, out2, in0, in2);
+    ST_SH2(in0, in2, output, 8);
+}
+
+void vp8_short_fdct8x4_msa(int16_t *input, int16_t *output, int32_t pitch)
+{
+    v8i16 in0, in1, in2, in3;
+    v8i16 temp0, temp1, tmp0, tmp1;
+    v8i16 const0, const1, const2;
+    v8i16 coeff = { 2217, 5352, -5352, 14500, 7500, 12000, 25000, 26000 };
+    v8i16 zero = { 0 };
+    v4i32 vec0_w, vec1_w, vec2_w, vec3_w;
+
+    LD_SH4(input, pitch / 2, in0, in1, in2, in3);
+    TRANSPOSE4x4_H(in0, in1, in2, in3, in0, in1, in2, in3);
+
+    BUTTERFLY_4(in0, in1, in2, in3, temp0, temp1, in1, in3);
+    SLLI_4V(temp0, temp1, in1, in3, 3);
+    in0 = temp0 + temp1;
+    in2 = temp0 - temp1;
+    SET_DOTP_VALUES(coeff, 0, 1, 2, const1, const2);
+    temp0 = __msa_splati_h(coeff, 3);
+    vec1_w = (v4i32)__msa_ilvev_h(zero, temp0);
+    coeff = __msa_ilvl_h(zero, coeff);
+    vec3_w = __msa_splati_w((v4i32)coeff, 0);
+    ILVRL_H2_SH(in3, in1, tmp1, tmp0);
+    vec0_w = vec1_w;
+    vec2_w = vec3_w;
+    DPADD_SH4_SW(tmp1, tmp0, tmp1, tmp0, const1, const1, const2, const2,
+                 vec0_w, vec1_w, vec2_w, vec3_w);
+    SRA_4V(vec1_w, vec0_w, vec3_w, vec2_w, 12);
+    PCKEV_H2_SH(vec1_w, vec0_w, vec3_w, vec2_w, in1, in3);
+    TRANSPOSE4x4_H(in0, in1, in2, in3, in0, in1, in2, in3);
+
+    BUTTERFLY_4(in0, in1, in2, in3, temp0, temp1, in1, in3);
+    in0 = temp0 + temp1 + 7;
+    in2 = temp0 - temp1 + 7;
+    in0 >>= 4;
+    in2 >>= 4;
+    SPLATI_W2_SW(coeff, 2, vec3_w, vec1_w);
+    vec3_w += vec1_w;
+    vec1_w = __msa_splati_w((v4i32)coeff, 1);
+    const0 = RET_1_IF_NZERO_H(in3);
+    ILVRL_H2_SH(in3, in1, tmp1, tmp0);
+    vec0_w = vec1_w;
+    vec2_w = vec3_w;
+    DPADD_SH4_SW(tmp1, tmp0, tmp1, tmp0, const1, const1, const2, const2,
+                 vec0_w, vec1_w, vec2_w, vec3_w);
+    SRA_4V(vec1_w, vec0_w, vec3_w, vec2_w, 16);
+    PCKEV_H2_SH(vec1_w, vec0_w, vec3_w, vec2_w, in1, in3);
+    in1 += const0;
+    PCKEV_D2_SH(in1, in0, in3, in2, temp0, temp1);
+    ST_SH2(temp0, temp1, output, 8);
+
+    PCKOD_D2_SH(in1, in0, in3, in2, in0, in2);
+    ST_SH2(in0, in2, output + 16, 8);
+}
+
+void vp8_short_walsh4x4_msa(int16_t *input, int16_t *output, int32_t pitch)
+{
+    v8i16 in0_h, in1_h, in2_h, in3_h;
+    v4i32 in0_w, in1_w, in2_w, in3_w, temp0, temp1, temp2, temp3;
+
+    LD_SH4(input, pitch / 2, in0_h, in1_h, in2_h, in3_h);
+    TRANSPOSE4x4_SH_SH(in0_h, in1_h, in2_h, in3_h, in0_h, in1_h, in2_h, in3_h);
+
+    UNPCK_R_SH_SW(in0_h, in0_w);
+    UNPCK_R_SH_SW(in1_h, in1_w);
+    UNPCK_R_SH_SW(in2_h, in2_w);
+    UNPCK_R_SH_SW(in3_h, in3_w);
+    BUTTERFLY_4(in0_w, in1_w, in3_w, in2_w, temp0, temp3, temp2, temp1);
+    SLLI_4V(temp0, temp1, temp2, temp3, 2);
+    BUTTERFLY_4(temp0, temp1, temp2, temp3, in0_w, in1_w, in2_w, in3_w);
+    temp0 = RET_1_IF_NZERO_W(temp0);
+    in0_w += temp0;
+    TRANSPOSE4x4_SW_SW(in0_w, in1_w, in2_w, in3_w, in0_w, in1_w, in2_w, in3_w);
+
+    BUTTERFLY_4(in0_w, in1_w, in3_w, in2_w, temp0, temp3, temp2, temp1);
+    BUTTERFLY_4(temp0, temp1, temp2, temp3, in0_w, in1_w, in2_w, in3_w);
+    in0_w += RET_1_IF_NEG_W(in0_w);
+    in1_w += RET_1_IF_NEG_W(in1_w);
+    in2_w += RET_1_IF_NEG_W(in2_w);
+    in3_w += RET_1_IF_NEG_W(in3_w);
+    ADD4(in0_w, 3, in1_w, 3, in2_w, 3, in3_w, 3, in0_w, in1_w, in2_w, in3_w);
+    SRA_4V(in0_w, in1_w, in2_w, in3_w, 3);
+    PCKEV_H2_SH(in1_w, in0_w, in3_w, in2_w, in0_h, in1_h);
+    ST_SH2(in0_h, in1_h, output, 8);
+}
diff --git a/libvpx/libvpx/vp8/encoder/mips/msa/denoising_msa.c b/libvpx/libvpx/vp8/encoder/mips/msa/denoising_msa.c
new file mode 100644
index 0000000..66965c6
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/mips/msa/denoising_msa.c
@@ -0,0 +1,624 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include "./vp8_rtcd.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+#include "vp8/encoder/denoising.h"
+
+int32_t vp8_denoiser_filter_msa(uint8_t *mc_running_avg_y_ptr,
+                                int32_t mc_avg_y_stride,
+                                uint8_t *running_avg_y_ptr,
+                                int32_t avg_y_stride,
+                                uint8_t *sig_ptr, int32_t sig_stride,
+                                uint32_t motion_magnitude,
+                                int32_t increase_denoising)
+{
+    uint8_t *running_avg_y_start = running_avg_y_ptr;
+    uint8_t *sig_start = sig_ptr;
+    int32_t cnt = 0;
+    int32_t sum_diff = 0;
+    int32_t shift_inc1 = 3;
+    int32_t delta = 0;
+    int32_t sum_diff_thresh;
+    v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+    v16u8 src8, src9, src10, src11, src12, src13, src14, src15;
+    v16u8 mc_running_avg_y0, running_avg_y, sig0;
+    v16u8 mc_running_avg_y1, running_avg_y1, sig1;
+    v16u8 coeff0, coeff1;
+    v8i16 diff0, diff1, abs_diff0, abs_diff1, abs_diff_neg0, abs_diff_neg1;
+    v8i16 adjust0, adjust1, adjust2, adjust3;
+    v8i16 shift_inc1_vec = { 0 };
+    v8i16 col_sum0 = { 0 };
+    v8i16 col_sum1 = { 0 };
+    v8i16 col_sum2 = { 0 };
+    v8i16 col_sum3 = { 0 };
+    v8i16 temp0_h, temp1_h, temp2_h, temp3_h, cmp, delta_vec;
+    v4i32 temp0_w;
+    v2i64 temp0_d, temp1_d;
+    v8i16 zero = { 0 };
+    v8i16 one = __msa_ldi_h(1);
+    v8i16 four = __msa_ldi_h(4);
+    v8i16 val_127 = __msa_ldi_h(127);
+    v8i16 adj_val = { 6, 4, 3, 0, -6, -4, -3, 0 };
+
+    if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
+    {
+        adj_val = __msa_add_a_h(adj_val, one);
+        if (increase_denoising)
+        {
+            adj_val = __msa_add_a_h(adj_val, one);
+            shift_inc1 = 4;
+        }
+
+        temp0_h = zero - adj_val;
+        adj_val = (v8i16)__msa_ilvev_d((v2i64)temp0_h, (v2i64)adj_val);
+    }
+
+    adj_val = __msa_insert_h(adj_val, 3, cnt);
+    adj_val = __msa_insert_h(adj_val, 7, cnt);
+    shift_inc1_vec = __msa_fill_h(shift_inc1);
+
+    for (cnt = 8; cnt--;)
+    {
+        v8i16 mask0 = { 0 };
+        v8i16 mask1 = { 0 };
+
+        mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr);
+        sig0 = LD_UB(sig_ptr);
+        sig_ptr += sig_stride;
+        mc_running_avg_y_ptr += mc_avg_y_stride;
+
+        mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr);
+        sig1 = LD_UB(sig_ptr);
+
+        ILVRL_B2_UB(mc_running_avg_y0, sig0, coeff0, coeff1);
+        HSUB_UB2_SH(coeff0, coeff1, diff0, diff1);
+        abs_diff0 = __msa_add_a_h(diff0, zero);
+        abs_diff1 = __msa_add_a_h(diff1, zero);
+        cmp = __msa_clei_s_h(abs_diff0, 15);
+        cmp = cmp & one;
+        mask0 += cmp;
+        cmp = __msa_clei_s_h(abs_diff0, 7);
+        cmp = cmp & one;
+        mask0 += cmp;
+        cmp = abs_diff0 < shift_inc1_vec;
+        cmp = cmp & one;
+        mask0 += cmp;
+        cmp = __msa_clei_s_h(abs_diff1, 15);
+        cmp = cmp & one;
+        mask1 += cmp;
+        cmp = __msa_clei_s_h(abs_diff1, 7);
+        cmp = cmp & one;
+        mask1 += cmp;
+        cmp = abs_diff1 < shift_inc1_vec;
+        cmp = cmp & one;
+        mask1 += cmp;
+        temp0_h = __msa_clei_s_h(diff0, 0);
+        temp0_h = temp0_h & four;
+        mask0 += temp0_h;
+        temp1_h = __msa_clei_s_h(diff1, 0);
+        temp1_h = temp1_h & four;
+        mask1 += temp1_h;
+        VSHF_H2_SH(adj_val, adj_val, adj_val, adj_val, mask0, mask1, adjust0,
+                   adjust1);
+        temp2_h = __msa_ceqi_h(adjust0, 0);
+        temp3_h = __msa_ceqi_h(adjust1, 0);
+        adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0,
+                                     (v16u8)temp2_h);
+        adjust1 = (v8i16)__msa_bmnz_v((v16u8)adjust1, (v16u8)diff1,
+                                     (v16u8)temp3_h);
+        ADD2(col_sum0, adjust0, col_sum1, adjust1, col_sum0, col_sum1);
+        UNPCK_UB_SH(sig0, temp0_h, temp1_h);
+        ADD2(temp0_h, adjust0, temp1_h, adjust1, temp0_h, temp1_h);
+        MAXI_SH2_SH(temp0_h, temp1_h, 0);
+        SAT_UH2_SH(temp0_h, temp1_h, 7);
+        temp2_h = (v8i16)__msa_pckev_b((v16i8)temp3_h, (v16i8)temp2_h);
+        running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp1_h, (v16i8)temp0_h);
+        running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y0,
+                                     (v16u8)temp2_h);
+        ST_UB(running_avg_y, running_avg_y_ptr);
+        running_avg_y_ptr += avg_y_stride;
+
+        mask0 = zero;
+        mask1 = zero;
+        ILVRL_B2_UB(mc_running_avg_y1, sig1, coeff0, coeff1);
+        HSUB_UB2_SH(coeff0, coeff1, diff0, diff1);
+        abs_diff0 = __msa_add_a_h(diff0, zero);
+        abs_diff1 = __msa_add_a_h(diff1, zero);
+        cmp = __msa_clei_s_h(abs_diff0, 15);
+        cmp = cmp & one;
+        mask0 += cmp;
+        cmp = __msa_clei_s_h(abs_diff0, 7);
+        cmp = cmp & one;
+        mask0 += cmp;
+        cmp = abs_diff0 < shift_inc1_vec;
+        cmp = cmp & one;
+        mask0 += cmp;
+        cmp = __msa_clei_s_h(abs_diff1, 15);
+        cmp = cmp & one;
+        mask1 += cmp;
+        cmp = __msa_clei_s_h(abs_diff1, 7);
+        cmp = cmp & one;
+        mask1 += cmp;
+        cmp = abs_diff1 < shift_inc1_vec;
+        cmp = cmp & one;
+        mask1 += cmp;
+        temp0_h = __msa_clei_s_h(diff0, 0);
+        temp0_h = temp0_h & four;
+        mask0 += temp0_h;
+        temp1_h = __msa_clei_s_h(diff1, 0);
+        temp1_h = temp1_h & four;
+        mask1 += temp1_h;
+        VSHF_H2_SH(adj_val, adj_val, adj_val, adj_val, mask0, mask1, adjust0,
+                   adjust1);
+        temp2_h = __msa_ceqi_h(adjust0, 0);
+        temp3_h = __msa_ceqi_h(adjust1, 0);
+        adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0,
+                                      (v16u8)temp2_h);
+        adjust1 = (v8i16)__msa_bmnz_v((v16u8)adjust1, (v16u8)diff1,
+                                      (v16u8)temp3_h);
+        ADD2(col_sum0, adjust0, col_sum1, adjust1, col_sum0, col_sum1);
+        UNPCK_UB_SH(sig1, temp0_h, temp1_h);
+        ADD2(temp0_h, adjust0, temp1_h, adjust1, temp0_h, temp1_h);
+        MAXI_SH2_SH(temp0_h, temp1_h, 0);
+        SAT_UH2_SH(temp0_h, temp1_h, 7);
+        temp2_h = (v8i16)__msa_pckev_b((v16i8)temp3_h, (v16i8)temp2_h);
+        running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp1_h, (v16i8)temp0_h);
+        running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y1,
+                                     (v16u8)temp2_h);
+        ST_UB(running_avg_y, running_avg_y_ptr);
+        sig_ptr += sig_stride;
+        mc_running_avg_y_ptr += mc_avg_y_stride;
+        running_avg_y_ptr += avg_y_stride;
+    }
+
+    col_sum0 = __msa_min_s_h(col_sum0, val_127);
+    col_sum1 = __msa_min_s_h(col_sum1, val_127);
+    temp0_h = col_sum0 + col_sum1;
+    temp0_w = __msa_hadd_s_w(temp0_h, temp0_h);
+    temp0_d = __msa_hadd_s_d(temp0_w, temp0_w);
+    temp1_d = __msa_splati_d(temp0_d, 1);
+    temp0_d += temp1_d;
+    sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0);
+    sig_ptr -= sig_stride * 16;
+    mc_running_avg_y_ptr -= mc_avg_y_stride * 16;
+    running_avg_y_ptr -= avg_y_stride * 16;
+
+    if (increase_denoising)
+    {
+        sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
+    }
+
+    if (abs(sum_diff) > sum_diff_thresh)
+    {
+        delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
+        delta_vec = __msa_fill_h(delta);
+        if (delta < 4)
+        {
+            for (cnt = 8; cnt--;)
+            {
+                running_avg_y = LD_UB(running_avg_y_ptr);
+                mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr);
+                sig0 = LD_UB(sig_ptr);
+                sig_ptr += sig_stride;
+                mc_running_avg_y_ptr += mc_avg_y_stride;
+                running_avg_y_ptr += avg_y_stride;
+                mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr);
+                sig1 = LD_UB(sig_ptr);
+                running_avg_y1 = LD_UB(running_avg_y_ptr);
+                ILVRL_B2_UB(mc_running_avg_y0, sig0, coeff0, coeff1);
+                HSUB_UB2_SH(coeff0, coeff1, diff0, diff1);
+                abs_diff0 = __msa_add_a_h(diff0, zero);
+                abs_diff1 = __msa_add_a_h(diff1, zero);
+                temp0_h = abs_diff0 < delta_vec;
+                temp1_h = abs_diff1 < delta_vec;
+                abs_diff0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0,
+                                               (v16u8)delta_vec,
+                                               (v16u8)temp0_h);
+                abs_diff1 = (v8i16)__msa_bmz_v((v16u8)abs_diff1,
+                                               (v16u8)delta_vec,
+                                               (v16u8)temp1_h);
+                SUB2(zero, abs_diff0, zero, abs_diff1, abs_diff_neg0,
+                     abs_diff_neg1);
+                abs_diff_neg0 = zero - abs_diff0;
+                abs_diff_neg1 = zero - abs_diff1;
+                temp0_h = __msa_clei_s_h(diff0, 0);
+                temp1_h = __msa_clei_s_h(diff1, 0);
+                adjust0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0,
+                                              (v16u8)abs_diff_neg0,
+                                              (v16u8)temp0_h);
+                adjust1 = (v8i16)__msa_bmnz_v((v16u8)abs_diff1,
+                                              (v16u8)abs_diff_neg1,
+                                              (v16u8)temp1_h);
+                ILVRL_B2_SH(zero, running_avg_y, temp2_h, temp3_h);
+                ADD2(temp2_h, adjust0, temp3_h, adjust1, adjust2, adjust3);
+                MAXI_SH2_SH(adjust2, adjust3, 0);
+                SAT_UH2_SH(adjust2, adjust3, 7);
+                temp0_h = __msa_ceqi_h(diff0, 0);
+                temp1_h = __msa_ceqi_h(diff1, 0);
+                adjust2 = (v8i16)__msa_bmz_v((v16u8)adjust2, (v16u8)temp2_h,
+                                             (v16u8)temp0_h);
+                adjust3 = (v8i16)__msa_bmz_v((v16u8)adjust3, (v16u8)temp3_h,
+                                             (v16u8)temp1_h);
+                adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)zero,
+                                              (v16u8)temp0_h);
+                adjust1 = (v8i16)__msa_bmnz_v((v16u8)adjust1, (v16u8)zero,
+                                              (v16u8)temp1_h);
+                ADD2(col_sum2, adjust0, col_sum3, adjust1, col_sum2, col_sum3);
+                running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust3,
+                                                     (v16i8)adjust2);
+                ST_UB(running_avg_y, running_avg_y_ptr - avg_y_stride);
+                ILVRL_B2_UB(mc_running_avg_y1, sig1, coeff0, coeff1);
+                HSUB_UB2_SH(coeff0, coeff1, diff0, diff1);
+                abs_diff0 = __msa_add_a_h(diff0, zero);
+                abs_diff1 = __msa_add_a_h(diff1, zero);
+                temp0_h = abs_diff0 < delta_vec;
+                temp1_h = abs_diff1 < delta_vec;
+                abs_diff0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0,
+                                               (v16u8)delta_vec,
+                                               (v16u8)temp0_h);
+                abs_diff1 = (v8i16)__msa_bmz_v((v16u8)abs_diff1,
+                                               (v16u8)delta_vec,
+                                               (v16u8)temp1_h);
+                SUB2(zero, abs_diff0, zero, abs_diff1, abs_diff_neg0,
+                     abs_diff_neg1);
+                temp0_h = __msa_clei_s_h(diff0, 0);
+                temp1_h = __msa_clei_s_h(diff1, 0);
+                adjust0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0,
+                                              (v16u8)abs_diff_neg0,
+                                              (v16u8)temp0_h);
+                adjust1 = (v8i16)__msa_bmnz_v((v16u8)abs_diff1,
+                                              (v16u8)abs_diff_neg1,
+                                              (v16u8)temp1_h);
+                ILVRL_H2_SH(zero, running_avg_y1, temp2_h, temp3_h);
+                ADD2(temp2_h, adjust0, temp3_h, adjust1, adjust2, adjust3);
+                MAXI_SH2_SH(adjust2, adjust3, 0);
+                SAT_UH2_SH(adjust2, adjust3, 7);
+                temp0_h = __msa_ceqi_h(diff0, 0);
+                temp1_h = __msa_ceqi_h(diff1, 0);
+                adjust2 = (v8i16)__msa_bmz_v((v16u8)adjust2, (v16u8)temp2_h,
+                                             (v16u8)temp0_h);
+                adjust3 = (v8i16)__msa_bmz_v((v16u8)adjust3, (v16u8)temp3_h,
+                                             (v16u8)temp1_h);
+                adjust0 = (v8i16)__msa_bmz_v((v16u8)adjust0, (v16u8)zero,
+                                             (v16u8)temp0_h);
+                adjust1 = (v8i16)__msa_bmz_v((v16u8)adjust1, (v16u8)zero,
+                                             (v16u8)temp1_h);
+                ADD2(col_sum2, adjust0, col_sum3, adjust1, col_sum2, col_sum3);
+                running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust3,
+                                                     (v16i8)adjust2);
+                ST_UB(running_avg_y, running_avg_y_ptr);
+                running_avg_y_ptr += avg_y_stride;
+            }
+
+            col_sum2 = __msa_min_s_h(col_sum2, val_127);
+            col_sum3 = __msa_min_s_h(col_sum3, val_127);
+            temp0_h = col_sum2 + col_sum3;
+            temp0_w = __msa_hadd_s_w(temp0_h, temp0_h);
+            temp0_d = __msa_hadd_s_d(temp0_w, temp0_w);
+            temp1_d = __msa_splati_d(temp0_d, 1);
+            temp0_d += (v2i64)temp1_d;
+            sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0);
+            if (abs(sum_diff) > SUM_DIFF_THRESHOLD)
+            {
+                return COPY_BLOCK;
+            }
+        }
+        else
+        {
+            return COPY_BLOCK;
+        }
+    }
+
+    LD_UB8(sig_start, sig_stride, src0, src1, src2, src3, src4, src5, src6,
+           src7);
+    sig_start += (8 * sig_stride);
+    LD_UB8(sig_start, sig_stride, src8, src9, src10, src11, src12, src13,
+           src14, src15);
+
+    ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, running_avg_y_start,
+           avg_y_stride);
+    running_avg_y_start += (8 * avg_y_stride);
+    ST_UB8(src8, src9, src10, src11, src12, src13, src14, src15,
+           running_avg_y_start, avg_y_stride);
+
+    return FILTER_BLOCK;
+}
+
+int32_t vp8_denoiser_filter_uv_msa(uint8_t *mc_running_avg_y_ptr,
+                                   int32_t mc_avg_y_stride,
+                                   uint8_t *running_avg_y_ptr,
+                                   int32_t avg_y_stride,
+                                   uint8_t *sig_ptr,
+                                   int32_t sig_stride,
+                                   uint32_t motion_magnitude,
+                                   int32_t increase_denoising)
+{
+    uint8_t *running_avg_y_start = running_avg_y_ptr;
+    uint8_t *sig_start = sig_ptr;
+    int32_t cnt = 0;
+    int32_t sum_diff = 0;
+    int32_t shift_inc1 = 3;
+    int32_t delta = 0;
+    int32_t sum_block = 0;
+    int32_t sum_diff_thresh;
+    int64_t dst0, dst1, src0, src1, src2, src3;
+    v16u8 mc_running_avg_y0, running_avg_y, sig0;
+    v16u8 mc_running_avg_y1, running_avg_y1, sig1;
+    v16u8 sig2, sig3, sig4, sig5, sig6, sig7;
+    v16u8 coeff0;
+    v8i16 diff0, abs_diff0, abs_diff_neg0;
+    v8i16 adjust0, adjust2;
+    v8i16 shift_inc1_vec = { 0 };
+    v8i16 col_sum0 = { 0 };
+    v8i16 temp0_h, temp2_h, cmp, delta_vec;
+    v4i32 temp0_w;
+    v2i64 temp0_d, temp1_d;
+    v16i8 zero = { 0 };
+    v8i16 one = __msa_ldi_h(1);
+    v8i16 four = __msa_ldi_h(4);
+    v8i16 adj_val = { 6, 4, 3, 0, -6, -4, -3, 0 };
+
+
+    sig0 = LD_UB(sig_ptr);
+    sig_ptr += sig_stride;
+    temp0_h = (v8i16)__msa_ilvr_b(zero, (v16i8)sig0);
+    sig1 = LD_UB(sig_ptr);
+    sig_ptr += sig_stride;
+    temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig1);
+    sig2 = LD_UB(sig_ptr);
+    sig_ptr += sig_stride;
+    temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig2);
+    sig3 = LD_UB(sig_ptr);
+    sig_ptr += sig_stride;
+    temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig3);
+    sig4 = LD_UB(sig_ptr);
+    sig_ptr += sig_stride;
+    temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig4);
+    sig5 = LD_UB(sig_ptr);
+    sig_ptr += sig_stride;
+    temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig5);
+    sig6 = LD_UB(sig_ptr);
+    sig_ptr += sig_stride;
+    temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig6);
+    sig7 = LD_UB(sig_ptr);
+    sig_ptr += sig_stride;
+    temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig7);
+    temp0_w = __msa_hadd_s_w(temp0_h, temp0_h);
+    temp0_d = __msa_hadd_s_d(temp0_w, temp0_w);
+    temp1_d = __msa_splati_d(temp0_d, 1);
+    temp0_d += temp1_d;
+    sum_block = __msa_copy_s_w((v4i32)temp0_d, 0);
+    sig_ptr -= sig_stride * 8;
+
+    if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV)
+    {
+        return COPY_BLOCK;
+    }
+
+    if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
+    {
+        adj_val = __msa_add_a_h(adj_val, one);
+
+        if (increase_denoising)
+        {
+            adj_val = __msa_add_a_h(adj_val, one);
+            shift_inc1 = 4;
+        }
+
+        temp0_h = (v8i16)zero - adj_val;
+        adj_val = (v8i16)__msa_ilvev_d((v2i64)temp0_h, (v2i64)adj_val);
+    }
+
+    adj_val = __msa_insert_h(adj_val, 3, cnt);
+    adj_val = __msa_insert_h(adj_val, 7, cnt);
+    shift_inc1_vec = __msa_fill_h(shift_inc1);
+    for (cnt = 4; cnt--;)
+    {
+        v8i16 mask0 = { 0 };
+        mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr);
+        sig0 = LD_UB(sig_ptr);
+        sig_ptr += sig_stride;
+        mc_running_avg_y_ptr += mc_avg_y_stride;
+        mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr);
+        sig1 = LD_UB(sig_ptr);
+        coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y0, (v16i8)sig0);
+        diff0 = __msa_hsub_u_h(coeff0, coeff0);
+        abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero);
+        cmp = __msa_clei_s_h(abs_diff0, 15);
+        cmp = cmp & one;
+        mask0 += cmp;
+        cmp = __msa_clei_s_h(abs_diff0, 7);
+        cmp = cmp & one;
+        mask0 += cmp;
+        cmp = abs_diff0 < shift_inc1_vec;
+        cmp = cmp & one;
+        mask0 += cmp;
+        temp0_h = __msa_clei_s_h(diff0, 0);
+        temp0_h = temp0_h & four;
+        mask0 += temp0_h;
+        adjust0 = __msa_vshf_h(mask0, adj_val, adj_val);
+        temp2_h = __msa_ceqi_h(adjust0, 0);
+        adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0,
+                                      (v16u8)temp2_h);
+        col_sum0 += adjust0;
+        temp0_h = (v8i16)__msa_ilvr_b(zero, (v16i8)sig0);
+        temp0_h += adjust0;
+        temp0_h = __msa_maxi_s_h(temp0_h, 0);
+        temp0_h = (v8i16)__msa_sat_u_h((v8u16)temp0_h, 7);
+        temp2_h = (v8i16)__msa_pckev_b((v16i8)temp2_h, (v16i8)temp2_h);
+        running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp0_h, (v16i8)temp0_h);
+        running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y0,
+                                     (v16u8)temp2_h);
+        dst0 = __msa_copy_s_d((v2i64)running_avg_y,  0);
+        SD(dst0, running_avg_y_ptr);
+        running_avg_y_ptr += avg_y_stride;
+
+        mask0 = __msa_ldi_h(0);
+        coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y1, (v16i8)sig1);
+        diff0 = __msa_hsub_u_h(coeff0, coeff0);
+        abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero);
+        cmp = __msa_clei_s_h(abs_diff0, 15);
+        cmp = cmp & one;
+        mask0 += cmp;
+        cmp = __msa_clei_s_h(abs_diff0, 7);
+        cmp = cmp & one;
+        mask0 += cmp;
+        cmp = abs_diff0 < shift_inc1_vec;
+        cmp = cmp & one;
+        mask0 += cmp;
+        temp0_h = __msa_clei_s_h(diff0, 0);
+        temp0_h = temp0_h & four;
+        mask0 += temp0_h;
+        adjust0 = __msa_vshf_h(mask0, adj_val, adj_val);
+        temp2_h = __msa_ceqi_h(adjust0, 0);
+        adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0,
+                                      (v16u8)temp2_h);
+        col_sum0 += adjust0;
+        temp0_h = (v8i16)__msa_ilvr_b(zero, (v16i8)sig1);
+        temp0_h += adjust0;
+        temp0_h = __msa_maxi_s_h(temp0_h, 0);
+        temp0_h = (v8i16)__msa_sat_u_h((v8u16)temp0_h, 7);
+
+        temp2_h = (v8i16)__msa_pckev_b((v16i8)temp2_h, (v16i8)temp2_h);
+        running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp0_h, (v16i8)temp0_h);
+        running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y1,
+                                     (v16u8)temp2_h);
+        dst1 = __msa_copy_s_d((v2i64)running_avg_y, 0);
+        SD(dst1, running_avg_y_ptr);
+
+        sig_ptr += sig_stride;
+        mc_running_avg_y_ptr += mc_avg_y_stride;
+        running_avg_y_ptr += avg_y_stride;
+    }
+
+    temp0_h = col_sum0;
+    temp0_w = __msa_hadd_s_w(temp0_h, temp0_h);
+    temp0_d = __msa_hadd_s_d(temp0_w, temp0_w);
+    temp1_d = __msa_splati_d(temp0_d, 1);
+    temp0_d += temp1_d;
+    sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0);
+    sig_ptr -= sig_stride * 8;
+    mc_running_avg_y_ptr -= mc_avg_y_stride * 8;
+    running_avg_y_ptr -= avg_y_stride * 8;
+    sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
+
+    if (increase_denoising)
+    {
+        sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+    }
+
+    if (abs(sum_diff) > sum_diff_thresh)
+    {
+        delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
+        delta_vec = __msa_fill_h(delta);
+        if (delta < 4)
+        {
+            for (cnt = 4; cnt--;)
+            {
+                running_avg_y = LD_UB(running_avg_y_ptr);
+                mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr);
+                sig0 = LD_UB(sig_ptr);
+                /* Update pointers for next iteration. */
+                sig_ptr += sig_stride;
+                mc_running_avg_y_ptr += mc_avg_y_stride;
+                running_avg_y_ptr += avg_y_stride;
+
+                mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr);
+                sig1 = LD_UB(sig_ptr);
+                running_avg_y1 = LD_UB(running_avg_y_ptr);
+
+                coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y0,
+                                             (v16i8)sig0);
+                diff0 = __msa_hsub_u_h(coeff0, coeff0);
+                abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero);
+                temp0_h = delta_vec < abs_diff0;
+                abs_diff0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0,
+                                                (v16u8)delta_vec,
+                                                (v16u8)temp0_h);
+                abs_diff_neg0 = (v8i16)zero - abs_diff0;
+                temp0_h = __msa_clei_s_h(diff0, 0);
+                adjust0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0,
+                                             (v16u8)abs_diff_neg0,
+                                             (v16u8)temp0_h);
+                temp2_h = (v8i16)__msa_ilvr_b(zero, (v16i8)running_avg_y);
+                adjust2 = temp2_h + adjust0;
+                adjust2 = __msa_maxi_s_h(adjust2, 0);
+                adjust2 = (v8i16)__msa_sat_u_h((v8u16)adjust2, 7);
+                temp0_h = __msa_ceqi_h(diff0, 0);
+                adjust2 = (v8i16)__msa_bmnz_v((v16u8)adjust2, (v16u8)temp2_h,
+                                              (v16u8)temp0_h);
+                adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)zero,
+                                              (v16u8)temp0_h);
+                col_sum0 += adjust0;
+                running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust2,
+                                                     (v16i8)adjust2);
+                dst0 = __msa_copy_s_d((v2i64)running_avg_y, 0);
+                SD(dst0, running_avg_y_ptr - avg_y_stride);
+
+                coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y1,
+                                             (v16i8)sig1);
+                diff0 = __msa_hsub_u_h(coeff0, coeff0);
+                abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero);
+                temp0_h = delta_vec < abs_diff0;
+                abs_diff0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0,
+                                                (v16u8)delta_vec,
+                                                (v16u8)temp0_h);
+                abs_diff_neg0 = (v8i16)zero - abs_diff0;
+                temp0_h = __msa_clei_s_h(diff0, 0);
+                adjust0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0,
+                                             (v16u8)abs_diff_neg0,
+                                             (v16u8)temp0_h);
+                temp2_h = (v8i16)__msa_ilvr_b(zero, (v16i8)running_avg_y1);
+                adjust2 = temp2_h + adjust0;
+                adjust2 = __msa_maxi_s_h(adjust2, 0);
+                adjust2 = (v8i16)__msa_sat_u_h((v8u16)adjust2, 7);
+                temp0_h = __msa_ceqi_h(diff0, 0);
+                adjust2 = (v8i16)__msa_bmnz_v((v16u8)adjust2, (v16u8)temp2_h,
+                                              (v16u8)temp0_h);
+                adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)zero,
+                                              (v16u8)temp0_h);
+                col_sum0 += adjust0;
+                running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust2,
+                                                     (v16i8)adjust2);
+                dst1 = __msa_copy_s_d((v2i64)running_avg_y, 0);
+                SD(dst1, running_avg_y_ptr);
+                running_avg_y_ptr += avg_y_stride;
+            }
+
+            temp0_h = col_sum0;
+            temp0_w = __msa_hadd_s_w(temp0_h, temp0_h);
+            temp0_d = __msa_hadd_s_d(temp0_w, temp0_w);
+            temp1_d = __msa_splati_d(temp0_d, 1);
+            temp0_d += temp1_d;
+            sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0);
+
+            if (abs(sum_diff) > sum_diff_thresh)
+            {
+                return COPY_BLOCK;
+            }
+        }
+        else
+        {
+            return COPY_BLOCK;
+        }
+    }
+
+    LD4(sig_start, sig_stride, src0, src1, src2, src3);
+    sig_start += (4 * sig_stride);
+    SD4(src0, src1, src2, src3, running_avg_y_start, avg_y_stride);
+    running_avg_y_start += (4 * avg_y_stride);
+
+    LD4(sig_start, sig_stride, src0, src1, src2, src3);
+    SD4(src0, src1, src2, src3, running_avg_y_start, avg_y_stride);
+
+    return FILTER_BLOCK;
+}
diff --git a/libvpx/libvpx/vp8/encoder/mips/msa/encodeopt_msa.c b/libvpx/libvpx/vp8/encoder/mips/msa/encodeopt_msa.c
new file mode 100644
index 0000000..ea794a8
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/mips/msa/encodeopt_msa.c
@@ -0,0 +1,174 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+#include "vp8/encoder/block.h"
+
+int32_t vp8_block_error_msa(int16_t *coeff_ptr, int16_t *dq_coeff_ptr)
+{
+    int32_t err = 0;
+    uint32_t loop_cnt;
+    v8i16 coeff, dq_coeff, coeff0, coeff1;
+    v4i32 diff0, diff1;
+    v2i64 err0 = { 0 };
+    v2i64 err1 = { 0 };
+
+    for (loop_cnt = 2; loop_cnt--;)
+    {
+        coeff = LD_SH(coeff_ptr);
+        dq_coeff = LD_SH(dq_coeff_ptr);
+        ILVRL_H2_SH(coeff, dq_coeff, coeff0, coeff1);
+        HSUB_UH2_SW(coeff0, coeff1, diff0, diff1);
+        DPADD_SD2_SD(diff0, diff1, err0, err1);
+        coeff_ptr += 8;
+        dq_coeff_ptr += 8;
+    }
+
+    err0 += __msa_splati_d(err0, 1);
+    err1 += __msa_splati_d(err1, 1);
+    err = __msa_copy_s_d(err0, 0);
+    err += __msa_copy_s_d(err1, 0);
+
+    return err;
+}
+
+int32_t vp8_mbblock_error_msa(MACROBLOCK *mb, int32_t dc)
+{
+    BLOCK *be;
+    BLOCKD *bd;
+    int16_t *coeff_ptr, *dq_coeff_ptr;
+    int32_t err = 0;
+    uint32_t loop_cnt;
+    v8i16 coeff, coeff0, coeff1, coeff2, coeff3, coeff4;
+    v8i16 dq_coeff, dq_coeff2, dq_coeff3, dq_coeff4;
+    v4i32 diff0, diff1;
+    v2i64 err0, err1;
+    v16u8 zero  = { 0 };
+    v16u8 mask0 = (v16u8)__msa_ldi_b(255);
+
+    if (1 == dc)
+    {
+        mask0 = (v16u8)__msa_insve_w((v4i32)mask0, 0, (v4i32)zero);
+    }
+
+    for (loop_cnt = 0; loop_cnt < 8; loop_cnt++)
+    {
+        be = &mb->block[2 * loop_cnt];
+        bd = &mb->e_mbd.block[2 * loop_cnt];
+        coeff_ptr = be->coeff;
+        dq_coeff_ptr = bd->dqcoeff;
+        coeff = LD_SH(coeff_ptr);
+        dq_coeff = LD_SH(dq_coeff_ptr);
+        coeff_ptr += 8;
+        dq_coeff_ptr += 8;
+        coeff2 = LD_SH(coeff_ptr);
+        dq_coeff2 = LD_SH(dq_coeff_ptr);
+        be = &mb->block[2 * loop_cnt + 1];
+        bd = &mb->e_mbd.block[2 * loop_cnt + 1];
+        coeff_ptr = be->coeff;
+        dq_coeff_ptr = bd->dqcoeff;
+        coeff3 = LD_SH(coeff_ptr);
+        dq_coeff3 = LD_SH(dq_coeff_ptr);
+        coeff_ptr += 8;
+        dq_coeff_ptr += 8;
+        coeff4 = LD_SH(coeff_ptr);
+        dq_coeff4 = LD_SH(dq_coeff_ptr);
+        ILVRL_H2_SH(coeff, dq_coeff, coeff0, coeff1);
+        HSUB_UH2_SW(coeff0, coeff1, diff0, diff1);
+        diff0 = (v4i32)__msa_bmnz_v(zero, (v16u8)diff0, mask0);
+        DOTP_SW2_SD(diff0, diff1, diff0, diff1, err0, err1);
+        ILVRL_H2_SH(coeff2, dq_coeff2, coeff0, coeff1);
+        HSUB_UH2_SW(coeff0, coeff1, diff0, diff1);
+        DPADD_SD2_SD(diff0, diff1, err0, err1);
+        err0 += __msa_splati_d(err0, 1);
+        err1 += __msa_splati_d(err1, 1);
+        err += __msa_copy_s_d(err0, 0);
+        err += __msa_copy_s_d(err1, 0);
+
+        ILVRL_H2_SH(coeff3, dq_coeff3, coeff0, coeff1);
+        HSUB_UH2_SW(coeff0, coeff1, diff0, diff1);
+        diff0 = (v4i32)__msa_bmnz_v(zero, (v16u8)diff0, mask0);
+        DOTP_SW2_SD(diff0, diff1, diff0, diff1, err0, err1);
+        ILVRL_H2_SH(coeff4, dq_coeff4, coeff0, coeff1);
+        HSUB_UH2_SW(coeff0, coeff1, diff0, diff1);
+        DPADD_SD2_SD(diff0, diff1, err0, err1);
+        err0 += __msa_splati_d(err0, 1);
+        err1 += __msa_splati_d(err1, 1);
+        err += __msa_copy_s_d(err0, 0);
+        err += __msa_copy_s_d(err1, 0);
+    }
+
+    return err;
+}
+
+int32_t vp8_mbuverror_msa(MACROBLOCK *mb)
+{
+    BLOCK *be;
+    BLOCKD *bd;
+    int16_t *coeff_ptr, *dq_coeff_ptr;
+    int32_t err = 0;
+    uint32_t loop_cnt;
+    v8i16 coeff, coeff0, coeff1, coeff2, coeff3, coeff4;
+    v8i16 dq_coeff, dq_coeff2, dq_coeff3, dq_coeff4;
+    v4i32 diff0, diff1;
+    v2i64 err0, err1, err_dup0, err_dup1;
+
+    for (loop_cnt = 16; loop_cnt < 24; loop_cnt += 2)
+    {
+        be = &mb->block[loop_cnt];
+        bd = &mb->e_mbd.block[loop_cnt];
+        coeff_ptr = be->coeff;
+        dq_coeff_ptr = bd->dqcoeff;
+        coeff = LD_SH(coeff_ptr);
+        dq_coeff = LD_SH(dq_coeff_ptr);
+        coeff_ptr += 8;
+        dq_coeff_ptr += 8;
+        coeff2 = LD_SH(coeff_ptr);
+        dq_coeff2 = LD_SH(dq_coeff_ptr);
+        be = &mb->block[loop_cnt + 1];
+        bd = &mb->e_mbd.block[loop_cnt + 1];
+        coeff_ptr = be->coeff;
+        dq_coeff_ptr = bd->dqcoeff;
+        coeff3 = LD_SH(coeff_ptr);
+        dq_coeff3 = LD_SH(dq_coeff_ptr);
+        coeff_ptr += 8;
+        dq_coeff_ptr += 8;
+        coeff4 = LD_SH(coeff_ptr);
+        dq_coeff4 = LD_SH(dq_coeff_ptr);
+
+        ILVRL_H2_SH(coeff, dq_coeff, coeff0, coeff1);
+        HSUB_UH2_SW(coeff0, coeff1, diff0, diff1);
+        DOTP_SW2_SD(diff0, diff1, diff0, diff1, err0, err1);
+
+        ILVRL_H2_SH(coeff2, dq_coeff2, coeff0, coeff1);
+        HSUB_UH2_SW(coeff0, coeff1, diff0, diff1);
+        DPADD_SD2_SD(diff0, diff1, err0, err1);
+        err_dup0 = __msa_splati_d(err0, 1);
+        err_dup1 = __msa_splati_d(err1, 1);
+        ADD2(err0, err_dup0, err1, err_dup1, err0, err1);
+        err += __msa_copy_s_d(err0, 0);
+        err += __msa_copy_s_d(err1, 0);
+
+        ILVRL_H2_SH(coeff3, dq_coeff3, coeff0, coeff1);
+        HSUB_UH2_SW(coeff0, coeff1, diff0, diff1);
+        DOTP_SW2_SD(diff0, diff1, diff0, diff1, err0, err1);
+        ILVRL_H2_SH(coeff4, dq_coeff4, coeff0, coeff1);
+        HSUB_UH2_SW(coeff0, coeff1, diff0, diff1);
+        DPADD_SD2_SD(diff0, diff1, err0, err1);
+        err_dup0 = __msa_splati_d(err0, 1);
+        err_dup1 = __msa_splati_d(err1, 1);
+        ADD2(err0, err_dup0, err1, err_dup1, err0, err1);
+        err += __msa_copy_s_d(err0, 0);
+        err += __msa_copy_s_d(err1, 0);
+    }
+
+    return err;
+}
diff --git a/libvpx/libvpx/vp8/encoder/mips/msa/quantize_msa.c b/libvpx/libvpx/vp8/encoder/mips/msa/quantize_msa.c
new file mode 100644
index 0000000..0f97646
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/mips/msa/quantize_msa.c
@@ -0,0 +1,246 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+#include "vp8/encoder/block.h"
+
+static int8_t fast_quantize_b_msa(int16_t *coeff_ptr, int16_t *zbin,
+                                  int16_t *round, int16_t *quant,
+                                  int16_t *de_quant, int16_t *q_coeff,
+                                  int16_t *dq_coeff)
+{
+    int32_t cnt, eob;
+    v16i8 inv_zig_zag = { 0, 1, 5, 6, 2, 4, 7, 12,
+                          3, 8, 11, 13, 9, 10, 14, 15 };
+    v8i16 round0, round1;
+    v8i16 sign_z0, sign_z1;
+    v8i16 q_coeff0, q_coeff1;
+    v8i16 x0, x1, de_quant0, de_quant1;
+    v8i16 coeff0, coeff1, z0, z1;
+    v8i16 quant0, quant1, quant2, quant3;
+    v8i16 zero = { 0 };
+    v8i16 inv_zig_zag0, inv_zig_zag1;
+    v8i16 zigzag_mask0 = { 0, 1, 4, 8, 5, 2, 3, 6 };
+    v8i16 zigzag_mask1 = { 9, 12, 13, 10, 7, 11, 14, 15 };
+    v8i16 temp0_h, temp1_h, temp2_h, temp3_h;
+    v4i32 temp0_w, temp1_w, temp2_w, temp3_w;
+
+    ILVRL_B2_SH(zero, inv_zig_zag, inv_zig_zag0, inv_zig_zag1);
+    eob = -1;
+    LD_SH2(coeff_ptr, 8, coeff0, coeff1);
+    VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+               z0, z1);
+    LD_SH2(round, 8, coeff0, coeff1);
+    VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+               round0, round1);
+    LD_SH2(quant, 8, coeff0, coeff1);
+    VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+               quant0, quant2);
+    sign_z0 = z0 >> 15;
+    sign_z1 = z1 >> 15;
+    x0 = __msa_add_a_h(z0, zero);
+    x1 = __msa_add_a_h(z1, zero);
+    ILVL_H2_SH(quant0, quant0, quant2, quant2, quant1, quant3);
+    ILVR_H2_SH(quant0, quant0, quant2, quant2, quant0, quant2);
+    ILVL_H2_SH(round0, x0, round1, x1, temp1_h, temp3_h);
+    ILVR_H2_SH(round0, x0, round1, x1, temp0_h, temp2_h);
+    DOTP_SH4_SW(temp0_h, temp1_h, temp2_h, temp3_h, quant0, quant1, quant2,
+                quant3, temp0_w, temp1_w, temp2_w, temp3_w);
+    SRA_4V(temp0_w, temp1_w, temp2_w, temp3_w, 16);
+    PCKEV_H2_SH(temp1_w, temp0_w, temp3_w, temp2_w, x0, x1);
+    x0 = x0 ^ sign_z0;
+    x1 = x1 ^ sign_z1;
+    SUB2(x0, sign_z0, x1, sign_z1, x0, x1);
+    VSHF_H2_SH(x0, x1, x0, x1, inv_zig_zag0, inv_zig_zag1, q_coeff0, q_coeff1);
+    ST_SH2(q_coeff0, q_coeff1, q_coeff, 8);
+    LD_SH2(de_quant, 8, de_quant0, de_quant1);
+    q_coeff0 *= de_quant0;
+    q_coeff1 *= de_quant1;
+    ST_SH2(q_coeff0, q_coeff1, dq_coeff, 8);
+
+    for (cnt = 0; cnt < 16; ++cnt)
+    {
+        if ((cnt <= 7) && (x1[7 - cnt] != 0))
+        {
+            eob = (15 - cnt);
+            break;
+        }
+
+        if ((cnt > 7) && (x0[7 - (cnt - 8)] != 0))
+        {
+            eob = (7 - (cnt - 8));
+            break;
+        }
+    }
+
+    return (int8_t)(eob + 1);
+}
+
+static int8_t exact_regular_quantize_b_msa(int16_t *zbin_boost,
+                                           int16_t *coeff_ptr,
+                                           int16_t *zbin,
+                                           int16_t *round,
+                                           int16_t *quant,
+                                           int16_t *quant_shift,
+                                           int16_t *de_quant,
+                                           int16_t zbin_oq_in,
+                                           int16_t *q_coeff,
+                                           int16_t *dq_coeff)
+{
+    int32_t cnt, eob;
+    int16_t *boost_temp = zbin_boost;
+    v16i8 inv_zig_zag = { 0, 1, 5, 6, 2, 4, 7, 12,
+                          3, 8, 11, 13, 9, 10, 14, 15 };
+    v8i16 round0, round1;
+    v8i16 sign_z0, sign_z1;
+    v8i16 q_coeff0, q_coeff1;
+    v8i16 z_bin0, z_bin1, zbin_o_q;
+    v8i16 x0, x1, sign_x0, sign_x1, de_quant0, de_quant1;
+    v8i16 coeff0, coeff1, z0, z1;
+    v8i16 quant0, quant1, quant2, quant3;
+    v8i16 zero = { 0 };
+    v8i16 inv_zig_zag0, inv_zig_zag1;
+    v8i16 zigzag_mask0 = { 0, 1, 4, 8, 5, 2, 3, 6 };
+    v8i16 zigzag_mask1 = { 9, 12, 13, 10, 7, 11, 14, 15 };
+    v8i16 temp0_h, temp1_h, temp2_h, temp3_h;
+    v4i32 temp0_w, temp1_w, temp2_w, temp3_w;
+
+    ILVRL_B2_SH(zero, inv_zig_zag, inv_zig_zag0, inv_zig_zag1);
+    zbin_o_q = __msa_fill_h(zbin_oq_in);
+    eob = -1;
+    LD_SH2(coeff_ptr, 8, coeff0, coeff1);
+    VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+               z0, z1);
+    LD_SH2(round, 8, coeff0, coeff1);
+    VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+               round0, round1);
+    LD_SH2(quant, 8, coeff0, coeff1);
+    VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+               quant0, quant2);
+    LD_SH2(zbin, 8, coeff0, coeff1);
+    VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+               z_bin0, z_bin1);
+    sign_z0 = z0 >> 15;
+    sign_z1 = z1 >> 15;
+    x0 = __msa_add_a_h(z0, zero);
+    x1 = __msa_add_a_h(z1, zero);
+    SUB2(x0, z_bin0, x1, z_bin1, z_bin0, z_bin1);
+    SUB2(z_bin0, zbin_o_q, z_bin1, zbin_o_q, z_bin0, z_bin1);
+    ILVL_H2_SH(quant0, quant0, quant2, quant2, quant1, quant3);
+    ILVR_H2_SH(quant0, quant0, quant2, quant2, quant0, quant2);
+    ILVL_H2_SH(round0, x0, round1, x1, temp1_h, temp3_h);
+    ILVR_H2_SH(round0, x0, round1, x1, temp0_h, temp2_h);
+    DOTP_SH4_SW(temp0_h, temp1_h, temp2_h, temp3_h, quant0, quant1, quant2,
+                quant3, temp0_w, temp1_w, temp2_w, temp3_w);
+    SRA_4V(temp0_w, temp1_w, temp2_w, temp3_w, 16);
+    PCKEV_H2_SH(temp1_w, temp0_w, temp3_w, temp2_w, temp0_h, temp2_h);
+    LD_SH2(quant_shift, 8, coeff0, coeff1);
+    VSHF_H2_SH(coeff0, coeff1, coeff0, coeff1, zigzag_mask0, zigzag_mask1,
+               quant0, quant2);
+    ILVL_H2_SH(quant0, quant0, quant2, quant2, quant1, quant3);
+    ILVR_H2_SH(quant0, quant0, quant2, quant2, quant0, quant2);
+    ADD2(x0, round0, x1, round1, x0, x1);
+    ILVL_H2_SH(temp0_h, x0, temp2_h, x1, temp1_h, temp3_h);
+    ILVR_H2_SH(temp0_h, x0, temp2_h, x1, temp0_h, temp2_h);
+    DOTP_SH4_SW(temp0_h, temp1_h, temp2_h, temp3_h, quant0, quant1, quant2,
+                quant3, temp0_w, temp1_w, temp2_w, temp3_w);
+    SRA_4V(temp0_w, temp1_w, temp2_w, temp3_w, 16);
+    PCKEV_H2_SH(temp1_w, temp0_w, temp3_w, temp2_w, x0, x1);
+    sign_x0 = x0 ^ sign_z0;
+    sign_x1 = x1 ^ sign_z1;
+    SUB2(sign_x0, sign_z0, sign_x1, sign_z1, sign_x0, sign_x1);
+    for (cnt = 0; cnt < 16; ++cnt)
+    {
+        if (cnt <= 7)
+        {
+            if (boost_temp[0] <= z_bin0[cnt])
+            {
+                if (x0[cnt])
+                {
+                    eob = cnt;
+                    boost_temp = zbin_boost;
+                }
+                else
+                {
+                    boost_temp++;
+                }
+            }
+            else
+            {
+                sign_x0[cnt] = 0;
+                boost_temp++;
+            }
+        }
+        else
+        {
+            if (boost_temp[0] <= z_bin1[cnt - 8])
+            {
+                if (x1[cnt - 8])
+                {
+                    eob = cnt;
+                    boost_temp = zbin_boost;
+                }
+                else
+                {
+                    boost_temp++;
+                }
+            }
+            else
+            {
+                sign_x1[cnt - 8] = 0;
+                boost_temp++;
+            }
+        }
+    }
+
+    VSHF_H2_SH(sign_x0, sign_x1, sign_x0, sign_x1, inv_zig_zag0, inv_zig_zag1,
+               q_coeff0, q_coeff1);
+    ST_SH2(q_coeff0, q_coeff1, q_coeff, 8);
+    LD_SH2(de_quant, 8, de_quant0, de_quant1);
+    MUL2(de_quant0, q_coeff0, de_quant1, q_coeff1, de_quant0, de_quant1);
+    ST_SH2(de_quant0, de_quant1, dq_coeff, 8);
+
+    return (int8_t)(eob + 1);
+}
+
+void vp8_fast_quantize_b_msa(BLOCK *b, BLOCKD *d)
+{
+    int16_t *coeff_ptr = b->coeff;
+    int16_t *zbin_ptr = b->zbin;
+    int16_t *round_ptr = b->round;
+    int16_t *quant_ptr = b->quant_fast;
+    int16_t *qcoeff_ptr = d->qcoeff;
+    int16_t *dqcoeff_ptr = d->dqcoeff;
+    int16_t *dequant_ptr = d->dequant;
+
+    *d->eob = fast_quantize_b_msa(coeff_ptr, zbin_ptr, round_ptr, quant_ptr,
+                                  dequant_ptr, qcoeff_ptr, dqcoeff_ptr);
+}
+
+void vp8_regular_quantize_b_msa(BLOCK *b, BLOCKD *d)
+{
+    int16_t *zbin_boost_ptr = b->zrun_zbin_boost;
+    int16_t *coeff_ptr = b->coeff;
+    int16_t *zbin_ptr = b->zbin;
+    int16_t *round_ptr = b->round;
+    int16_t *quant_ptr = b->quant;
+    int16_t *quant_shift_ptr = b->quant_shift;
+    int16_t *qcoeff_ptr = d->qcoeff;
+    int16_t *dqcoeff_ptr = d->dqcoeff;
+    int16_t *dequant_ptr = d->dequant;
+    int16_t zbin_oq_value = b->zbin_extra;
+
+    *d->eob = exact_regular_quantize_b_msa(zbin_boost_ptr, coeff_ptr,
+                                           zbin_ptr, round_ptr,
+                                           quant_ptr, quant_shift_ptr,
+                                           dequant_ptr, zbin_oq_value,
+                                           qcoeff_ptr, dqcoeff_ptr);
+}
diff --git a/libvpx/libvpx/vp8/encoder/mips/msa/temporal_filter_msa.c b/libvpx/libvpx/vp8/encoder/mips/msa/temporal_filter_msa.c
new file mode 100644
index 0000000..5cca5e0
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/mips/msa/temporal_filter_msa.c
@@ -0,0 +1,303 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp8_rtcd.h"
+#include "vp8/common/mips/msa/vp8_macros_msa.h"
+
+static void temporal_filter_apply_16size_msa(uint8_t *frame1_ptr,
+                                             uint32_t stride,
+                                             uint8_t *frame2_ptr,
+                                             int32_t strength_in,
+                                             int32_t filter_wt_in,
+                                             uint32_t *acc, uint16_t *cnt)
+{
+    uint32_t row;
+    v16i8 frame1_0_b, frame1_1_b, frame2_0_b, frame2_1_b;
+    v16u8 frame_l, frame_h;
+    v16i8 zero = { 0 };
+    v8i16 frame2_0_h, frame2_1_h, mod0_h, mod1_h;
+    v8i16 diff0, diff1, cnt0, cnt1;
+    v4i32 const3, const16, filter_wt, strength;
+    v4i32 mod0_w, mod1_w, mod2_w, mod3_w;
+    v4i32 diff0_r, diff0_l, diff1_r, diff1_l;
+    v4i32 frame2_0, frame2_1, frame2_2, frame2_3;
+    v4i32 acc0, acc1, acc2, acc3;
+
+    filter_wt = __msa_fill_w(filter_wt_in);
+    strength = __msa_fill_w(strength_in);
+    const3 = __msa_ldi_w(3);
+    const16 = __msa_ldi_w(16);
+
+    for (row = 8; row--;)
+    {
+        frame1_0_b = LD_SB(frame1_ptr);
+        frame2_0_b = LD_SB(frame2_ptr);
+        frame1_ptr += stride;
+        frame2_ptr += 16;
+        frame1_1_b = LD_SB(frame1_ptr);
+        frame2_1_b = LD_SB(frame2_ptr);
+        LD_SW2(acc, 4, acc0, acc1);
+        LD_SW2(acc + 8, 4, acc2, acc3);
+        LD_SH2(cnt, 8, cnt0, cnt1);
+        ILVRL_B2_UB(frame1_0_b, frame2_0_b, frame_l, frame_h);
+        HSUB_UB2_SH(frame_l, frame_h, diff0, diff1);
+        UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+        UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+        MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l,
+             diff1_l, mod0_w, mod1_w, mod2_w, mod3_w);
+        MUL4(mod0_w, const3, mod1_w, const3, mod2_w, const3, mod3_w, const3,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+        diff0_r = (mod0_w < const16);
+        diff0_l = (mod1_w < const16);
+        diff1_r = (mod2_w < const16);
+        diff1_l = (mod3_w < const16);
+        SUB4(const16, mod0_w, const16, mod1_w, const16, mod2_w, const16, mod3_w,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        mod0_w = diff0_r & mod0_w;
+        mod1_w = diff0_l & mod1_w;
+        mod2_w = diff1_r & mod2_w;
+        mod3_w = diff1_l & mod3_w;
+        MUL4(mod0_w, filter_wt, mod1_w, filter_wt, mod2_w, filter_wt, mod3_w,
+             filter_wt, mod0_w, mod1_w, mod2_w, mod3_w);
+        PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h)
+        ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+        ST_SH2(mod0_h, mod1_h, cnt, 8);
+        cnt += 16;
+        ILVRL_B2_SH(zero, frame2_0_b, frame2_0_h, frame2_1_h);
+        UNPCK_SH_SW(frame2_0_h, frame2_0, frame2_1);
+        UNPCK_SH_SW(frame2_1_h, frame2_2, frame2_3);
+        MUL4(mod0_w, frame2_0, mod1_w, frame2_1, mod2_w, frame2_2, mod3_w,
+             frame2_3, mod0_w, mod1_w, mod2_w, mod3_w);
+        ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        ST_SW2(mod0_w, mod1_w, acc, 4);
+        ST_SW2(mod2_w, mod3_w, acc + 8, 4);
+        acc += 16;
+        LD_SW2(acc, 4, acc0, acc1);
+        LD_SW2(acc + 8, 4, acc2, acc3);
+        LD_SH2(cnt, 8, cnt0, cnt1);
+        ILVRL_B2_UB(frame1_1_b, frame2_1_b, frame_l, frame_h);
+        HSUB_UB2_SH(frame_l, frame_h, diff0, diff1);
+        UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+        UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+        MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l,
+             diff1_l, mod0_w, mod1_w, mod2_w, mod3_w);
+        MUL4(mod0_w, const3, mod1_w, const3, mod2_w, const3, mod3_w, const3,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+        diff0_r = (mod0_w < const16);
+        diff0_l = (mod1_w < const16);
+        diff1_r = (mod2_w < const16);
+        diff1_l = (mod3_w < const16);
+        SUB4(const16, mod0_w, const16, mod1_w, const16, mod2_w, const16, mod3_w,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        mod0_w = diff0_r & mod0_w;
+        mod1_w = diff0_l & mod1_w;
+        mod2_w = diff1_r & mod2_w;
+        mod3_w = diff1_l & mod3_w;
+        MUL4(mod0_w, filter_wt, mod1_w, filter_wt, mod2_w, filter_wt, mod3_w,
+             filter_wt, mod0_w, mod1_w, mod2_w, mod3_w);
+        PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+        ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+        ST_SH2(mod0_h, mod1_h, cnt, 8);
+        cnt += 16;
+
+        UNPCK_UB_SH(frame2_1_b, frame2_0_h, frame2_1_h);
+        UNPCK_SH_SW(frame2_0_h, frame2_0, frame2_1);
+        UNPCK_SH_SW(frame2_1_h, frame2_2, frame2_3);
+        MUL4(mod0_w, frame2_0, mod1_w, frame2_1, mod2_w, frame2_2, mod3_w,
+             frame2_3, mod0_w, mod1_w, mod2_w, mod3_w);
+        ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        ST_SW2(mod0_w, mod1_w, acc, 4);
+        ST_SW2(mod2_w, mod3_w, acc + 8, 4);
+        acc += 16;
+        frame1_ptr += stride;
+        frame2_ptr += 16;
+    }
+}
+
+static void temporal_filter_apply_8size_msa(uint8_t *frame1_ptr,
+                                            uint32_t stride,
+                                            uint8_t *frame2_ptr,
+                                            int32_t strength_in,
+                                            int32_t filter_wt_in,
+                                            uint32_t *acc, uint16_t *cnt)
+{
+    uint32_t row;
+    uint64_t f0, f1, f2, f3, f4, f5, f6, f7;
+    v16i8 frame1 = { 0 };
+    v16i8 frame2 = { 0 };
+    v16i8 frame3 = { 0 };
+    v16i8 frame4 = { 0 };
+    v16u8 frame_l, frame_h;
+    v8i16 frame2_0_h, frame2_1_h, mod0_h, mod1_h;
+    v8i16 diff0, diff1, cnt0, cnt1;
+    v4i32 const3, const16;
+    v4i32 filter_wt, strength;
+    v4i32 mod0_w, mod1_w, mod2_w, mod3_w;
+    v4i32 diff0_r, diff0_l, diff1_r, diff1_l;
+    v4i32 frame2_0, frame2_1, frame2_2, frame2_3;
+    v4i32 acc0, acc1, acc2, acc3;
+
+    filter_wt = __msa_fill_w(filter_wt_in);
+    strength = __msa_fill_w(strength_in);
+    const3 = __msa_ldi_w(3);
+    const16 = __msa_ldi_w(16);
+
+    for (row = 2; row--;)
+    {
+        LD2(frame1_ptr, stride, f0, f1);
+        frame1_ptr += (2 * stride);
+        LD2(frame2_ptr, 8, f2, f3);
+        frame2_ptr += 16;
+        LD2(frame1_ptr, stride, f4, f5);
+        frame1_ptr += (2 * stride);
+        LD2(frame2_ptr, 8, f6, f7);
+        frame2_ptr += 16;
+
+        LD_SW2(acc, 4, acc0, acc1);
+        LD_SW2(acc + 8, 4, acc2, acc3);
+        LD_SH2(cnt, 8, cnt0, cnt1);
+        INSERT_D2_SB(f0, f1, frame1);
+        INSERT_D2_SB(f2, f3, frame2);
+        INSERT_D2_SB(f4, f5, frame3);
+        INSERT_D2_SB(f6, f7, frame4);
+        ILVRL_B2_UB(frame1, frame2, frame_l, frame_h);
+        HSUB_UB2_SH(frame_l, frame_h, diff0, diff1);
+        UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+        UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+        MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l,
+             diff1_l, mod0_w, mod1_w, mod2_w, mod3_w);
+        MUL4(mod0_w, const3, mod1_w, const3, mod2_w, const3, mod3_w, const3,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+        diff0_r = (mod0_w < const16);
+        diff0_l = (mod1_w < const16);
+        diff1_r = (mod2_w < const16);
+        diff1_l = (mod3_w < const16);
+        SUB4(const16, mod0_w, const16, mod1_w, const16, mod2_w, const16, mod3_w,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        mod0_w = diff0_r & mod0_w;
+        mod1_w = diff0_l & mod1_w;
+        mod2_w = diff1_r & mod2_w;
+        mod3_w = diff1_l & mod3_w;
+        MUL4(mod0_w, filter_wt, mod1_w, filter_wt, mod2_w, filter_wt, mod3_w,
+             filter_wt, mod0_w, mod1_w, mod2_w, mod3_w);
+        PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+        ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+        ST_SH2(mod0_h, mod1_h, cnt, 8);
+        cnt += 16;
+
+        UNPCK_UB_SH(frame2, frame2_0_h, frame2_1_h);
+        UNPCK_SH_SW(frame2_0_h, frame2_0, frame2_1);
+        UNPCK_SH_SW(frame2_1_h, frame2_2, frame2_3);
+        MUL4(mod0_w, frame2_0, mod1_w, frame2_1, mod2_w, frame2_2, mod3_w,
+             frame2_3, mod0_w, mod1_w, mod2_w, mod3_w);
+        ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        ST_SW2(mod0_w, mod1_w, acc, 4);
+        ST_SW2(mod2_w, mod3_w, acc + 8, 4);
+        acc += 16;
+
+        LD_SW2(acc, 4, acc0, acc1);
+        LD_SW2(acc + 8, 4, acc2, acc3);
+        LD_SH2(cnt, 8, cnt0, cnt1);
+        ILVRL_B2_UB(frame3, frame4, frame_l, frame_h);
+        HSUB_UB2_SH(frame_l, frame_h, diff0, diff1);
+        UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+        UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+        MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l,
+             diff1_l, mod0_w, mod1_w, mod2_w, mod3_w);
+        MUL4(mod0_w, const3, mod1_w, const3, mod2_w, const3, mod3_w, const3,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+        diff0_r = (mod0_w < const16);
+        diff0_l = (mod1_w < const16);
+        diff1_r = (mod2_w < const16);
+        diff1_l = (mod3_w < const16);
+        SUB4(const16, mod0_w, const16, mod1_w, const16, mod2_w, const16, mod3_w,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        mod0_w = diff0_r & mod0_w;
+        mod1_w = diff0_l & mod1_w;
+        mod2_w = diff1_r & mod2_w;
+        mod3_w = diff1_l & mod3_w;
+        MUL4(mod0_w, filter_wt, mod1_w, filter_wt, mod2_w, filter_wt, mod3_w,
+             filter_wt, mod0_w, mod1_w, mod2_w, mod3_w);
+        PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+        ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+        ST_SH2(mod0_h, mod1_h, cnt, 8);
+        cnt += 16;
+
+        UNPCK_UB_SH(frame4, frame2_0_h, frame2_1_h);
+        UNPCK_SH_SW(frame2_0_h, frame2_0, frame2_1);
+        UNPCK_SH_SW(frame2_1_h, frame2_2, frame2_3);
+        MUL4(mod0_w, frame2_0, mod1_w, frame2_1, mod2_w, frame2_2, mod3_w,
+             frame2_3, mod0_w, mod1_w, mod2_w, mod3_w);
+        ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3,
+             mod0_w, mod1_w, mod2_w, mod3_w);
+        ST_SW2(mod0_w, mod1_w, acc, 4);
+        ST_SW2(mod2_w, mod3_w, acc + 8, 4);
+        acc += 16;
+    }
+}
+
+void vp8_temporal_filter_apply_msa(uint8_t *frame1, uint32_t stride,
+                                   uint8_t *frame2, uint32_t block_size,
+                                   int32_t strength,  int32_t filter_weight,
+                                   uint32_t *accumulator, uint16_t *count)
+{
+    if (8 == block_size)
+    {
+        temporal_filter_apply_8size_msa(frame1, stride, frame2, strength,
+                                        filter_weight, accumulator, count);
+    }
+    else if (16 == block_size)
+    {
+        temporal_filter_apply_16size_msa(frame1, stride, frame2, strength,
+                                         filter_weight, accumulator, count);
+    }
+    else
+    {
+        uint32_t i, j, k;
+        int32_t modifier;
+        int32_t byte = 0;
+        const int32_t rounding = strength > 0 ? 1 << (strength - 1) : 0;
+
+        for (i = 0, k = 0; i < block_size; ++i)
+        {
+            for (j = 0; j < block_size; ++j, ++k)
+            {
+                int src_byte = frame1[byte];
+                int pixel_value = *frame2++;
+
+                modifier = src_byte - pixel_value;
+                modifier *= modifier;
+                modifier *= 3;
+                modifier += rounding;
+                modifier >>= strength;
+
+                if (modifier > 16)
+                    modifier = 16;
+
+                modifier = 16 - modifier;
+                modifier *= filter_weight;
+
+                count[k] += modifier;
+                accumulator[k] += modifier * pixel_value;
+
+                byte++;
+            }
+
+            byte += stride - block_size;
+        }
+    }
+}
diff --git a/libvpx/libvpx/vp8/encoder/modecosts.c b/libvpx/libvpx/vp8/encoder/modecosts.c
new file mode 100644
index 0000000..ad0e930
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/modecosts.c
@@ -0,0 +1,55 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vp8/common/blockd.h"
+#include "modecosts.h"
+#include "onyx_int.h"
+#include "treewriter.h"
+#include "vp8/common/entropymode.h"
+
+
+void vp8_init_mode_costs(VP8_COMP *c)
+{
+    VP8_COMMON *x = &c->common;
+    struct rd_costs_struct *rd_costs = &c->rd_costs;
+
+    {
+        const vp8_tree_p T = vp8_bmode_tree;
+
+        int i = 0;
+
+        do
+        {
+            int j = 0;
+
+            do
+            {
+                vp8_cost_tokens(rd_costs->bmode_costs[i][j],
+                                vp8_kf_bmode_prob[i][j], T);
+            }
+            while (++j < VP8_BINTRAMODES);
+        }
+        while (++i < VP8_BINTRAMODES);
+
+        vp8_cost_tokens(rd_costs->inter_bmode_costs, x->fc.bmode_prob, T);
+    }
+    vp8_cost_tokens(rd_costs->inter_bmode_costs, x->fc.sub_mv_ref_prob,
+                    vp8_sub_mv_ref_tree);
+
+    vp8_cost_tokens(rd_costs->mbmode_cost[1], x->fc.ymode_prob, vp8_ymode_tree);
+    vp8_cost_tokens(rd_costs->mbmode_cost[0], vp8_kf_ymode_prob,
+                    vp8_kf_ymode_tree);
+
+    vp8_cost_tokens(rd_costs->intra_uv_mode_cost[1], x->fc.uv_mode_prob,
+                    vp8_uv_mode_tree);
+    vp8_cost_tokens(rd_costs->intra_uv_mode_cost[0], vp8_kf_uv_mode_prob,
+                    vp8_uv_mode_tree);
+}
diff --git a/libvpx/libvpx/vp8/encoder/modecosts.h b/libvpx/libvpx/vp8/encoder/modecosts.h
new file mode 100644
index 0000000..9871bff
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/modecosts.h
@@ -0,0 +1,27 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_MODECOSTS_H_
+#define VP8_ENCODER_MODECOSTS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP8_COMP;
+
+void vp8_init_mode_costs(struct VP8_COMP *x);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_MODECOSTS_H_
diff --git a/libvpx/libvpx/vp8/encoder/mr_dissim.c b/libvpx/libvpx/vp8/encoder/mr_dissim.c
new file mode 100644
index 0000000..886cba2
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/mr_dissim.c
@@ -0,0 +1,240 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <limits.h>
+#include "vpx_config.h"
+#include "onyx_int.h"
+#include "mr_dissim.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "rdopt.h"
+#include "vp8/common/common.h"
+
+void vp8_cal_low_res_mb_cols(VP8_COMP *cpi)
+{
+    int low_res_w;
+
+    /* Support arbitrary down-sampling factor */
+    unsigned int iw = cpi->oxcf.Width*cpi->oxcf.mr_down_sampling_factor.den
+                      + cpi->oxcf.mr_down_sampling_factor.num - 1;
+
+    low_res_w = iw/cpi->oxcf.mr_down_sampling_factor.num;
+    cpi->mr_low_res_mb_cols = ((low_res_w + 15) >> 4);
+}
+
+#define GET_MV(x)    \
+if(x->mbmi.ref_frame !=INTRA_FRAME)   \
+{   \
+    mvx[cnt] = x->mbmi.mv.as_mv.row;  \
+    mvy[cnt] = x->mbmi.mv.as_mv.col;  \
+    cnt++;    \
+}
+
+#define GET_MV_SIGN(x)    \
+if(x->mbmi.ref_frame !=INTRA_FRAME)   \
+{   \
+    mvx[cnt] = x->mbmi.mv.as_mv.row;  \
+    mvy[cnt] = x->mbmi.mv.as_mv.col;  \
+    if (cm->ref_frame_sign_bias[x->mbmi.ref_frame]  \
+        != cm->ref_frame_sign_bias[tmp->mbmi.ref_frame])  \
+    {  \
+        mvx[cnt] *= -1;   \
+        mvy[cnt] *= -1;   \
+    }  \
+    cnt++;  \
+}
+
+void vp8_cal_dissimilarity(VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = &cpi->common;
+    int i;
+
+    /* Note: The first row & first column in mip are outside the frame, which
+     * were initialized to all 0.(ref_frame, mode, mv...)
+     * Their ref_frame = 0 means they won't be counted in the following
+     * calculation.
+     */
+    if (cpi->oxcf.mr_total_resolutions >1
+        && cpi->oxcf.mr_encoder_id < (cpi->oxcf.mr_total_resolutions - 1))
+    {
+        /* Store info for show/no-show frames for supporting alt_ref.
+         * If parent frame is alt_ref, child has one too.
+         */
+        LOWER_RES_FRAME_INFO* store_info
+                      = (LOWER_RES_FRAME_INFO*)cpi->oxcf.mr_low_res_mode_info;
+
+        store_info->frame_type = cm->frame_type;
+
+        if(cm->frame_type != KEY_FRAME)
+        {
+            store_info->is_frame_dropped = 0;
+            for (i = 1; i < MAX_REF_FRAMES; i++)
+                store_info->low_res_ref_frames[i] = cpi->current_ref_frames[i];
+        }
+
+        if(cm->frame_type != KEY_FRAME)
+        {
+            int mb_row;
+            int mb_col;
+            /* Point to beginning of allocated MODE_INFO arrays. */
+            MODE_INFO *tmp = cm->mip + cm->mode_info_stride;
+            LOWER_RES_MB_INFO* store_mode_info = store_info->mb_info;
+
+            for (mb_row = 0; mb_row < cm->mb_rows; mb_row ++)
+            {
+                tmp++;
+                for (mb_col = 0; mb_col < cm->mb_cols; mb_col ++)
+                {
+                    int dissim = INT_MAX;
+
+                    if(tmp->mbmi.ref_frame !=INTRA_FRAME)
+                    {
+                        int              mvx[8];
+                        int              mvy[8];
+                        int              mmvx;
+                        int              mmvy;
+                        int              cnt=0;
+                        const MODE_INFO *here = tmp;
+                        const MODE_INFO *above = here - cm->mode_info_stride;
+                        const MODE_INFO *left = here - 1;
+                        const MODE_INFO *aboveleft = above - 1;
+                        const MODE_INFO *aboveright = NULL;
+                        const MODE_INFO *right = NULL;
+                        const MODE_INFO *belowleft = NULL;
+                        const MODE_INFO *below = NULL;
+                        const MODE_INFO *belowright = NULL;
+
+                        /* If alternate reference frame is used, we have to
+                         * check sign of MV. */
+                        if(cpi->oxcf.play_alternate)
+                        {
+                            /* Gather mv of neighboring MBs */
+                            GET_MV_SIGN(above)
+                            GET_MV_SIGN(left)
+                            GET_MV_SIGN(aboveleft)
+
+                            if(mb_col < (cm->mb_cols-1))
+                            {
+                                right = here + 1;
+                                aboveright = above + 1;
+                                GET_MV_SIGN(right)
+                                GET_MV_SIGN(aboveright)
+                            }
+
+                            if(mb_row < (cm->mb_rows-1))
+                            {
+                                below = here + cm->mode_info_stride;
+                                belowleft = below - 1;
+                                GET_MV_SIGN(below)
+                                GET_MV_SIGN(belowleft)
+                            }
+
+                            if(mb_col < (cm->mb_cols-1)
+                                && mb_row < (cm->mb_rows-1))
+                            {
+                                belowright = below + 1;
+                                GET_MV_SIGN(belowright)
+                            }
+                        }else
+                        {
+                            /* No alt_ref and gather mv of neighboring MBs */
+                            GET_MV(above)
+                            GET_MV(left)
+                            GET_MV(aboveleft)
+
+                            if(mb_col < (cm->mb_cols-1))
+                            {
+                                right = here + 1;
+                                aboveright = above + 1;
+                                GET_MV(right)
+                                GET_MV(aboveright)
+                            }
+
+                            if(mb_row < (cm->mb_rows-1))
+                            {
+                                below = here + cm->mode_info_stride;
+                                belowleft = below - 1;
+                                GET_MV(below)
+                                GET_MV(belowleft)
+                            }
+
+                            if(mb_col < (cm->mb_cols-1)
+                                && mb_row < (cm->mb_rows-1))
+                            {
+                                belowright = below + 1;
+                                GET_MV(belowright)
+                            }
+                        }
+
+                        if (cnt > 0)
+                        {
+                            int max_mvx = mvx[0];
+                            int min_mvx = mvx[0];
+                            int max_mvy = mvy[0];
+                            int min_mvy = mvy[0];
+                            int i;
+
+                            if (cnt > 1)
+                            {
+                                for (i=1; i< cnt; i++)
+                                {
+                                    if (mvx[i] > max_mvx) max_mvx = mvx[i];
+                                    else if (mvx[i] < min_mvx) min_mvx = mvx[i];
+                                    if (mvy[i] > max_mvy) max_mvy = mvy[i];
+                                    else if (mvy[i] < min_mvy) min_mvy = mvy[i];
+                                }
+                            }
+
+                            mmvx = VPXMAX(
+                                abs(min_mvx - here->mbmi.mv.as_mv.row),
+                                abs(max_mvx - here->mbmi.mv.as_mv.row));
+                            mmvy = VPXMAX(
+                                abs(min_mvy - here->mbmi.mv.as_mv.col),
+                                abs(max_mvy - here->mbmi.mv.as_mv.col));
+                            dissim = VPXMAX(mmvx, mmvy);
+                        }
+                    }
+
+                    /* Store mode info for next resolution encoding */
+                    store_mode_info->mode = tmp->mbmi.mode;
+                    store_mode_info->ref_frame = tmp->mbmi.ref_frame;
+                    store_mode_info->mv.as_int = tmp->mbmi.mv.as_int;
+                    store_mode_info->dissim = dissim;
+                    tmp++;
+                    store_mode_info++;
+                }
+            }
+        }
+    }
+}
+
+/* This function is called only when this frame is dropped at current
+   resolution level. */
+void vp8_store_drop_frame_info(VP8_COMP *cpi)
+{
+    /* If the frame is dropped in lower-resolution encoding, this information
+       is passed to higher resolution level so that the encoder knows there
+       is no mode & motion info available.
+     */
+    if (cpi->oxcf.mr_total_resolutions >1
+        && cpi->oxcf.mr_encoder_id < (cpi->oxcf.mr_total_resolutions - 1))
+    {
+        /* Store info for show/no-show frames for supporting alt_ref.
+         * If parent frame is alt_ref, child has one too.
+         */
+        LOWER_RES_FRAME_INFO* store_info
+                      = (LOWER_RES_FRAME_INFO*)cpi->oxcf.mr_low_res_mode_info;
+
+        /* Set frame_type to be INTER_FRAME since we won't drop key frame. */
+        store_info->frame_type = INTER_FRAME;
+        store_info->is_frame_dropped = 1;
+    }
+}
diff --git a/libvpx/libvpx/vp8/encoder/mr_dissim.h b/libvpx/libvpx/vp8/encoder/mr_dissim.h
new file mode 100644
index 0000000..5a59ce6
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/mr_dissim.h
@@ -0,0 +1,28 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_MR_DISSIM_H_
+#define VP8_ENCODER_MR_DISSIM_H_
+#include "vpx_config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern void vp8_cal_low_res_mb_cols(VP8_COMP *cpi);
+extern void vp8_cal_dissimilarity(VP8_COMP *cpi);
+extern void vp8_store_drop_frame_info(VP8_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_MR_DISSIM_H_
diff --git a/libvpx/libvpx/vp8/encoder/onyx_if.c b/libvpx/libvpx/vp8/encoder/onyx_if.c
new file mode 100644
index 0000000..d5a0fff
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/onyx_if.c
@@ -0,0 +1,5981 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "./vpx_scale_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vp8_rtcd.h"
+#include "vp8/common/onyxc_int.h"
+#include "vp8/common/blockd.h"
+#include "onyx_int.h"
+#include "vp8/common/systemdependent.h"
+#include "vp8/encoder/quantize.h"
+#include "vp8/common/alloccommon.h"
+#include "mcomp.h"
+#include "firstpass.h"
+#include "vpx/internal/vpx_psnr.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vp8/common/extend.h"
+#include "ratectrl.h"
+#include "vp8/common/quant_common.h"
+#include "segmentation.h"
+#if CONFIG_POSTPROC
+#include "vp8/common/postproc.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+#include "vp8/common/reconintra.h"
+#include "vp8/common/swapyv12buffer.h"
+#include "vp8/common/threading.h"
+#include "vpx_ports/vpx_timer.h"
+#if ARCH_ARM
+#include "vpx_ports/arm.h"
+#endif
+#if CONFIG_MULTI_RES_ENCODING
+#include "mr_dissim.h"
+#endif
+#include "encodeframe.h"
+
+#include <math.h>
+#include <stdio.h>
+#include <limits.h>
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+extern int vp8_update_coef_context(VP8_COMP *cpi);
+extern void vp8_update_coef_probs(VP8_COMP *cpi);
+#endif
+
+extern void vp8cx_pick_filter_level_fast(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi);
+extern void vp8cx_set_alt_lf_level(VP8_COMP *cpi, int filt_val);
+extern void vp8cx_pick_filter_level(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi);
+
+extern void vp8_deblock_frame(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *post, int filt_lvl, int low_var_thresh, int flag);
+extern void print_parms(VP8_CONFIG *ocf, char *filenam);
+extern unsigned int vp8_get_processor_freq();
+extern void print_tree_update_probs();
+extern int vp8cx_create_encoder_threads(VP8_COMP *cpi);
+extern void vp8cx_remove_encoder_threads(VP8_COMP *cpi);
+
+int vp8_estimate_entropy_savings(VP8_COMP *cpi);
+
+int vp8_calc_ss_err(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest);
+
+extern void vp8_temporal_filter_prepare_c(VP8_COMP *cpi, int distance);
+
+static void set_default_lf_deltas(VP8_COMP *cpi);
+
+extern const int vp8_gf_interval_table[101];
+
+#if CONFIG_INTERNAL_STATS
+#include "math.h"
+#include "vpx_dsp/ssim.h"
+#endif
+
+
+#ifdef OUTPUT_YUV_SRC
+FILE *yuv_file;
+#endif
+#ifdef OUTPUT_YUV_DENOISED
+FILE *yuv_denoised_file;
+#endif
+
+#if 0
+FILE *framepsnr;
+FILE *kf_list;
+FILE *keyfile;
+#endif
+
+#if 0
+extern int skip_true_count;
+extern int skip_false_count;
+#endif
+
+
+#ifdef VP8_ENTROPY_STATS
+extern int intra_mode_stats[10][10][10];
+#endif
+
+#ifdef SPEEDSTATS
+unsigned int frames_at_speed[16] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+unsigned int tot_pm = 0;
+unsigned int cnt_pm = 0;
+unsigned int tot_ef = 0;
+unsigned int cnt_ef = 0;
+#endif
+
+#ifdef MODE_STATS
+extern unsigned __int64 Sectionbits[50];
+extern int y_modes[5]  ;
+extern int uv_modes[4] ;
+extern int b_modes[10]  ;
+
+extern int inter_y_modes[10] ;
+extern int inter_uv_modes[4] ;
+extern unsigned int inter_b_modes[15];
+#endif
+
+extern const int vp8_bits_per_mb[2][QINDEX_RANGE];
+
+extern const int qrounding_factors[129];
+extern const int qzbin_factors[129];
+extern void vp8cx_init_quantizer(VP8_COMP *cpi);
+extern const int vp8cx_base_skip_false_prob[128];
+
+/* Tables relating active max Q to active min Q */
+static const unsigned char kf_low_motion_minq[QINDEX_RANGE] =
+{
+    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
+    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
+    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
+    0,0,0,0,1,1,1,1,1,1,1,1,2,2,2,2,
+    3,3,3,3,3,3,4,4,4,5,5,5,5,5,6,6,
+    6,6,7,7,8,8,8,8,9,9,10,10,10,10,11,11,
+    11,11,12,12,13,13,13,13,14,14,15,15,15,15,16,16,
+    16,16,17,17,18,18,18,18,19,20,20,21,21,22,23,23
+};
+static const unsigned char kf_high_motion_minq[QINDEX_RANGE] =
+{
+    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
+    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
+    1,1,1,1,1,1,1,1,2,2,2,2,3,3,3,3,
+    3,3,3,3,4,4,4,4,5,5,5,5,5,5,6,6,
+    6,6,7,7,8,8,8,8,9,9,10,10,10,10,11,11,
+    11,11,12,12,13,13,13,13,14,14,15,15,15,15,16,16,
+    16,16,17,17,18,18,18,18,19,19,20,20,20,20,21,21,
+    21,21,22,22,23,23,24,25,25,26,26,27,28,28,29,30
+};
+static const unsigned char gf_low_motion_minq[QINDEX_RANGE] =
+{
+    0,0,0,0,1,1,1,1,1,1,1,1,2,2,2,2,
+    3,3,3,3,4,4,4,4,5,5,5,5,6,6,6,6,
+    7,7,7,7,8,8,8,8,9,9,9,9,10,10,10,10,
+    11,11,12,12,13,13,14,14,15,15,16,16,17,17,18,18,
+    19,19,20,20,21,21,22,22,23,23,24,24,25,25,26,26,
+    27,27,28,28,29,29,30,30,31,31,32,32,33,33,34,34,
+    35,35,36,36,37,37,38,38,39,39,40,40,41,41,42,42,
+    43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58
+};
+static const unsigned char gf_mid_motion_minq[QINDEX_RANGE] =
+{
+    0,0,0,0,1,1,1,1,1,1,2,2,3,3,3,4,
+    4,4,5,5,5,6,6,6,7,7,7,8,8,8,9,9,
+    9,10,10,10,10,11,11,11,12,12,12,12,13,13,13,14,
+    14,14,15,15,16,16,17,17,18,18,19,19,20,20,21,21,
+    22,22,23,23,24,24,25,25,26,26,27,27,28,28,29,29,
+    30,30,31,31,32,32,33,33,34,34,35,35,36,36,37,37,
+    38,39,39,40,40,41,41,42,42,43,43,44,45,46,47,48,
+    49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64
+};
+static const unsigned char gf_high_motion_minq[QINDEX_RANGE] =
+{
+    0,0,0,0,1,1,1,1,1,2,2,2,3,3,3,4,
+    4,4,5,5,5,6,6,6,7,7,7,8,8,8,9,9,
+    9,10,10,10,11,11,12,12,13,13,14,14,15,15,16,16,
+    17,17,18,18,19,19,20,20,21,21,22,22,23,23,24,24,
+    25,25,26,26,27,27,28,28,29,29,30,30,31,31,32,32,
+    33,33,34,34,35,35,36,36,37,37,38,38,39,39,40,40,
+    41,41,42,42,43,44,45,46,47,48,49,50,51,52,53,54,
+    55,56,57,58,59,60,62,64,66,68,70,72,74,76,78,80
+};
+static const unsigned char inter_minq[QINDEX_RANGE] =
+{
+    0,0,1,1,2,3,3,4,4,5,6,6,7,8,8,9,
+    9,10,11,11,12,13,13,14,15,15,16,17,17,18,19,20,
+    20,21,22,22,23,24,24,25,26,27,27,28,29,30,30,31,
+    32,33,33,34,35,36,36,37,38,39,39,40,41,42,42,43,
+    44,45,46,46,47,48,49,50,50,51,52,53,54,55,55,56,
+    57,58,59,60,60,61,62,63,64,65,66,67,67,68,69,70,
+    71,72,73,74,75,75,76,77,78,79,80,81,82,83,84,85,
+    86,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100
+};
+
+#ifdef PACKET_TESTING
+extern FILE *vpxlogc;
+#endif
+
+static void save_layer_context(VP8_COMP *cpi)
+{
+    LAYER_CONTEXT *lc = &cpi->layer_context[cpi->current_layer];
+
+    /* Save layer dependent coding state */
+    lc->target_bandwidth                 = cpi->target_bandwidth;
+    lc->starting_buffer_level            = cpi->oxcf.starting_buffer_level;
+    lc->optimal_buffer_level             = cpi->oxcf.optimal_buffer_level;
+    lc->maximum_buffer_size              = cpi->oxcf.maximum_buffer_size;
+    lc->starting_buffer_level_in_ms      = cpi->oxcf.starting_buffer_level_in_ms;
+    lc->optimal_buffer_level_in_ms       = cpi->oxcf.optimal_buffer_level_in_ms;
+    lc->maximum_buffer_size_in_ms        = cpi->oxcf.maximum_buffer_size_in_ms;
+    lc->buffer_level                     = cpi->buffer_level;
+    lc->bits_off_target                  = cpi->bits_off_target;
+    lc->total_actual_bits                = cpi->total_actual_bits;
+    lc->worst_quality                    = cpi->worst_quality;
+    lc->active_worst_quality             = cpi->active_worst_quality;
+    lc->best_quality                     = cpi->best_quality;
+    lc->active_best_quality              = cpi->active_best_quality;
+    lc->ni_av_qi                         = cpi->ni_av_qi;
+    lc->ni_tot_qi                        = cpi->ni_tot_qi;
+    lc->ni_frames                        = cpi->ni_frames;
+    lc->avg_frame_qindex                 = cpi->avg_frame_qindex;
+    lc->rate_correction_factor           = cpi->rate_correction_factor;
+    lc->key_frame_rate_correction_factor = cpi->key_frame_rate_correction_factor;
+    lc->gf_rate_correction_factor        = cpi->gf_rate_correction_factor;
+    lc->zbin_over_quant                  = cpi->mb.zbin_over_quant;
+    lc->inter_frame_target               = cpi->inter_frame_target;
+    lc->total_byte_count                 = cpi->total_byte_count;
+    lc->filter_level                     = cpi->common.filter_level;
+
+    lc->last_frame_percent_intra         = cpi->last_frame_percent_intra;
+
+    memcpy (lc->count_mb_ref_frame_usage,
+            cpi->mb.count_mb_ref_frame_usage,
+            sizeof(cpi->mb.count_mb_ref_frame_usage));
+}
+
+static void restore_layer_context(VP8_COMP *cpi, const int layer)
+{
+    LAYER_CONTEXT *lc = &cpi->layer_context[layer];
+
+    /* Restore layer dependent coding state */
+    cpi->current_layer                    = layer;
+    cpi->target_bandwidth                 = lc->target_bandwidth;
+    cpi->oxcf.target_bandwidth            = lc->target_bandwidth;
+    cpi->oxcf.starting_buffer_level       = lc->starting_buffer_level;
+    cpi->oxcf.optimal_buffer_level        = lc->optimal_buffer_level;
+    cpi->oxcf.maximum_buffer_size         = lc->maximum_buffer_size;
+    cpi->oxcf.starting_buffer_level_in_ms = lc->starting_buffer_level_in_ms;
+    cpi->oxcf.optimal_buffer_level_in_ms  = lc->optimal_buffer_level_in_ms;
+    cpi->oxcf.maximum_buffer_size_in_ms   = lc->maximum_buffer_size_in_ms;
+    cpi->buffer_level                     = lc->buffer_level;
+    cpi->bits_off_target                  = lc->bits_off_target;
+    cpi->total_actual_bits                = lc->total_actual_bits;
+    cpi->active_worst_quality             = lc->active_worst_quality;
+    cpi->active_best_quality              = lc->active_best_quality;
+    cpi->ni_av_qi                         = lc->ni_av_qi;
+    cpi->ni_tot_qi                        = lc->ni_tot_qi;
+    cpi->ni_frames                        = lc->ni_frames;
+    cpi->avg_frame_qindex                 = lc->avg_frame_qindex;
+    cpi->rate_correction_factor           = lc->rate_correction_factor;
+    cpi->key_frame_rate_correction_factor = lc->key_frame_rate_correction_factor;
+    cpi->gf_rate_correction_factor        = lc->gf_rate_correction_factor;
+    cpi->mb.zbin_over_quant                  = lc->zbin_over_quant;
+    cpi->inter_frame_target               = lc->inter_frame_target;
+    cpi->total_byte_count                 = lc->total_byte_count;
+    cpi->common.filter_level              = lc->filter_level;
+
+    cpi->last_frame_percent_intra         = lc->last_frame_percent_intra;
+
+    memcpy (cpi->mb.count_mb_ref_frame_usage,
+            lc->count_mb_ref_frame_usage,
+            sizeof(cpi->mb.count_mb_ref_frame_usage));
+}
+
+static int rescale(int val, int num, int denom)
+{
+    int64_t llnum = num;
+    int64_t llden = denom;
+    int64_t llval = val;
+
+    return (int)(llval * llnum / llden);
+}
+
+static void init_temporal_layer_context(VP8_COMP *cpi,
+                                        VP8_CONFIG *oxcf,
+                                        const int layer,
+                                        double prev_layer_framerate)
+{
+    LAYER_CONTEXT *lc = &cpi->layer_context[layer];
+
+    lc->framerate = cpi->output_framerate / cpi->oxcf.rate_decimator[layer];
+    lc->target_bandwidth = cpi->oxcf.target_bitrate[layer] * 1000;
+
+    lc->starting_buffer_level_in_ms = oxcf->starting_buffer_level;
+    lc->optimal_buffer_level_in_ms  = oxcf->optimal_buffer_level;
+    lc->maximum_buffer_size_in_ms   = oxcf->maximum_buffer_size;
+
+    lc->starting_buffer_level =
+        rescale((int)(oxcf->starting_buffer_level),
+                lc->target_bandwidth, 1000);
+
+    if (oxcf->optimal_buffer_level == 0)
+      lc->optimal_buffer_level = lc->target_bandwidth / 8;
+    else
+      lc->optimal_buffer_level =
+          rescale((int)(oxcf->optimal_buffer_level),
+                  lc->target_bandwidth, 1000);
+
+    if (oxcf->maximum_buffer_size == 0)
+      lc->maximum_buffer_size = lc->target_bandwidth / 8;
+    else
+      lc->maximum_buffer_size =
+          rescale((int)(oxcf->maximum_buffer_size),
+                  lc->target_bandwidth, 1000);
+
+    /* Work out the average size of a frame within this layer */
+    if (layer > 0)
+      lc->avg_frame_size_for_layer =
+          (int)((cpi->oxcf.target_bitrate[layer] -
+                cpi->oxcf.target_bitrate[layer-1]) * 1000 /
+                (lc->framerate - prev_layer_framerate));
+
+     lc->active_worst_quality         = cpi->oxcf.worst_allowed_q;
+     lc->active_best_quality          = cpi->oxcf.best_allowed_q;
+     lc->avg_frame_qindex             = cpi->oxcf.worst_allowed_q;
+
+     lc->buffer_level                 = lc->starting_buffer_level;
+     lc->bits_off_target              = lc->starting_buffer_level;
+
+     lc->total_actual_bits                 = 0;
+     lc->ni_av_qi                          = 0;
+     lc->ni_tot_qi                         = 0;
+     lc->ni_frames                         = 0;
+     lc->rate_correction_factor            = 1.0;
+     lc->key_frame_rate_correction_factor  = 1.0;
+     lc->gf_rate_correction_factor         = 1.0;
+     lc->inter_frame_target                = 0;
+}
+
+// Upon a run-time change in temporal layers, reset the layer context parameters
+// for any "new" layers. For "existing" layers, let them inherit the parameters
+// from the previous layer state (at the same layer #). In future we may want
+// to better map the previous layer state(s) to the "new" ones.
+static void reset_temporal_layer_change(VP8_COMP *cpi,
+                                        VP8_CONFIG *oxcf,
+                                        const int prev_num_layers)
+{
+    int i;
+    double prev_layer_framerate = 0;
+    const int curr_num_layers = cpi->oxcf.number_of_layers;
+    // If the previous state was 1 layer, get current layer context from cpi.
+    // We need this to set the layer context for the new layers below.
+    if (prev_num_layers == 1)
+    {
+        cpi->current_layer = 0;
+        save_layer_context(cpi);
+    }
+    for (i = 0; i < curr_num_layers; i++)
+    {
+        LAYER_CONTEXT *lc = &cpi->layer_context[i];
+        if (i >= prev_num_layers)
+        {
+           init_temporal_layer_context(cpi, oxcf, i, prev_layer_framerate);
+        }
+        // The initial buffer levels are set based on their starting levels.
+        // We could set the buffer levels based on the previous state (normalized
+        // properly by the layer bandwidths) but we would need to keep track of
+        // the previous set of layer bandwidths (i.e., target_bitrate[i])
+        // before the layer change. For now, reset to the starting levels.
+        lc->buffer_level = cpi->oxcf.starting_buffer_level_in_ms *
+                           cpi->oxcf.target_bitrate[i];
+        lc->bits_off_target = lc->buffer_level;
+        // TDOD(marpan): Should we set the rate_correction_factor and
+        // active_worst/best_quality to values derived from the previous layer
+        // state (to smooth-out quality dips/rate fluctuation at transition)?
+
+        // We need to treat the 1 layer case separately: oxcf.target_bitrate[i]
+        // is not set for 1 layer, and the restore_layer_context/save_context()
+        // are not called in the encoding loop, so we need to call it here to
+        // pass the layer context state to |cpi|.
+        if (curr_num_layers == 1)
+        {
+            lc->target_bandwidth = cpi->oxcf.target_bandwidth;
+            lc->buffer_level = cpi->oxcf.starting_buffer_level_in_ms *
+                               lc->target_bandwidth  / 1000;
+            lc->bits_off_target = lc->buffer_level;
+            restore_layer_context(cpi, 0);
+        }
+        prev_layer_framerate = cpi->output_framerate /
+                               cpi->oxcf.rate_decimator[i];
+    }
+}
+
+static void setup_features(VP8_COMP *cpi)
+{
+    // If segmentation enabled set the update flags
+    if ( cpi->mb.e_mbd.segmentation_enabled )
+    {
+        cpi->mb.e_mbd.update_mb_segmentation_map = 1;
+        cpi->mb.e_mbd.update_mb_segmentation_data = 1;
+    }
+    else
+    {
+        cpi->mb.e_mbd.update_mb_segmentation_map = 0;
+        cpi->mb.e_mbd.update_mb_segmentation_data = 0;
+    }
+
+    cpi->mb.e_mbd.mode_ref_lf_delta_enabled = 0;
+    cpi->mb.e_mbd.mode_ref_lf_delta_update = 0;
+    memset(cpi->mb.e_mbd.ref_lf_deltas, 0, sizeof(cpi->mb.e_mbd.ref_lf_deltas));
+    memset(cpi->mb.e_mbd.mode_lf_deltas, 0, sizeof(cpi->mb.e_mbd.mode_lf_deltas));
+    memset(cpi->mb.e_mbd.last_ref_lf_deltas, 0, sizeof(cpi->mb.e_mbd.ref_lf_deltas));
+    memset(cpi->mb.e_mbd.last_mode_lf_deltas, 0, sizeof(cpi->mb.e_mbd.mode_lf_deltas));
+
+    set_default_lf_deltas(cpi);
+
+}
+
+
+static void dealloc_raw_frame_buffers(VP8_COMP *cpi);
+
+void vp8_initialize_enc(void)
+{
+    static volatile int init_done = 0;
+
+    if (!init_done) {
+        vpx_dsp_rtcd();
+        vp8_init_intra_predictors();
+        init_done = 1;
+    }
+}
+
+static void dealloc_compressor_data(VP8_COMP *cpi)
+{
+    vpx_free(cpi->tplist);
+    cpi->tplist = NULL;
+
+    /* Delete last frame MV storage buffers */
+    vpx_free(cpi->lfmv);
+    cpi->lfmv = 0;
+
+    vpx_free(cpi->lf_ref_frame_sign_bias);
+    cpi->lf_ref_frame_sign_bias = 0;
+
+    vpx_free(cpi->lf_ref_frame);
+    cpi->lf_ref_frame = 0;
+
+    /* Delete sementation map */
+    vpx_free(cpi->segmentation_map);
+    cpi->segmentation_map = 0;
+
+    vpx_free(cpi->active_map);
+    cpi->active_map = 0;
+
+    vp8_de_alloc_frame_buffers(&cpi->common);
+
+    vp8_yv12_de_alloc_frame_buffer(&cpi->pick_lf_lvl_frame);
+    vp8_yv12_de_alloc_frame_buffer(&cpi->scaled_source);
+    dealloc_raw_frame_buffers(cpi);
+
+    vpx_free(cpi->tok);
+    cpi->tok = 0;
+
+    /* Structure used to monitor GF usage */
+    vpx_free(cpi->gf_active_flags);
+    cpi->gf_active_flags = 0;
+
+    /* Activity mask based per mb zbin adjustments */
+    vpx_free(cpi->mb_activity_map);
+    cpi->mb_activity_map = 0;
+
+    vpx_free(cpi->mb.pip);
+    cpi->mb.pip = 0;
+
+#if CONFIG_MULTITHREAD
+    /* De-allocate mutex */
+    if (cpi->pmutex != NULL) {
+        VP8_COMMON *const pc = &cpi->common;
+        int i;
+
+        for (i = 0; i < pc->mb_rows; i++) {
+            pthread_mutex_destroy(&cpi->pmutex[i]);
+        }
+        vpx_free(cpi->pmutex);
+        cpi->pmutex = NULL;
+    }
+
+    vpx_free(cpi->mt_current_mb_col);
+    cpi->mt_current_mb_col = NULL;
+#endif
+}
+
+static void enable_segmentation(VP8_COMP *cpi)
+{
+    /* Set the appropriate feature bit */
+    cpi->mb.e_mbd.segmentation_enabled = 1;
+    cpi->mb.e_mbd.update_mb_segmentation_map = 1;
+    cpi->mb.e_mbd.update_mb_segmentation_data = 1;
+}
+static void disable_segmentation(VP8_COMP *cpi)
+{
+    /* Clear the appropriate feature bit */
+    cpi->mb.e_mbd.segmentation_enabled = 0;
+}
+
+/* Valid values for a segment are 0 to 3
+ * Segmentation map is arrange as [Rows][Columns]
+ */
+static void set_segmentation_map(VP8_COMP *cpi, unsigned char *segmentation_map)
+{
+    /* Copy in the new segmentation map */
+    memcpy(cpi->segmentation_map, segmentation_map, (cpi->common.mb_rows * cpi->common.mb_cols));
+
+    /* Signal that the map should be updated. */
+    cpi->mb.e_mbd.update_mb_segmentation_map = 1;
+    cpi->mb.e_mbd.update_mb_segmentation_data = 1;
+}
+
+/* The values given for each segment can be either deltas (from the default
+ * value chosen for the frame) or absolute values.
+ *
+ * Valid range for abs values is:
+ *    (0-127 for MB_LVL_ALT_Q), (0-63 for SEGMENT_ALT_LF)
+ * Valid range for delta values are:
+ *    (+/-127 for MB_LVL_ALT_Q), (+/-63 for SEGMENT_ALT_LF)
+ *
+ * abs_delta = SEGMENT_DELTADATA (deltas)
+ * abs_delta = SEGMENT_ABSDATA (use the absolute values given).
+ *
+ */
+static void set_segment_data(VP8_COMP *cpi, signed char *feature_data, unsigned char abs_delta)
+{
+    cpi->mb.e_mbd.mb_segement_abs_delta = abs_delta;
+    memcpy(cpi->segment_feature_data, feature_data, sizeof(cpi->segment_feature_data));
+}
+
+
+/* A simple function to cyclically refresh the background at a lower Q */
+static void cyclic_background_refresh(VP8_COMP *cpi, int Q, int lf_adjustment)
+{
+    unsigned char *seg_map = cpi->segmentation_map;
+    signed char feature_data[MB_LVL_MAX][MAX_MB_SEGMENTS];
+    int i;
+    int block_count = cpi->cyclic_refresh_mode_max_mbs_perframe;
+    int mbs_in_frame = cpi->common.mb_rows * cpi->common.mb_cols;
+
+    cpi->cyclic_refresh_q = Q / 2;
+
+    if (cpi->oxcf.screen_content_mode) {
+      // Modify quality ramp-up based on Q. Above some Q level, increase the
+      // number of blocks to be refreshed, and reduce it below the thredhold.
+      // Turn-off under certain conditions (i.e., away from key frame, and if
+      // we are at good quality (low Q) and most of the blocks were skipped-encoded
+      // in previous frame.
+      int qp_thresh = (cpi->oxcf.screen_content_mode == 2) ? 80 : 100;
+      if (Q >= qp_thresh) {
+        cpi->cyclic_refresh_mode_max_mbs_perframe =
+            (cpi->common.mb_rows * cpi->common.mb_cols) / 10;
+      } else if (cpi->frames_since_key > 250 &&
+                 Q < 20 &&
+                 cpi->mb.skip_true_count > (int)(0.95 * mbs_in_frame)) {
+        cpi->cyclic_refresh_mode_max_mbs_perframe = 0;
+      } else {
+        cpi->cyclic_refresh_mode_max_mbs_perframe =
+            (cpi->common.mb_rows * cpi->common.mb_cols) / 20;
+      }
+      block_count = cpi->cyclic_refresh_mode_max_mbs_perframe;
+    }
+
+    // Set every macroblock to be eligible for update.
+    // For key frame this will reset seg map to 0.
+    memset(cpi->segmentation_map, 0, mbs_in_frame);
+
+    if (cpi->common.frame_type != KEY_FRAME && block_count > 0)
+    {
+        /* Cycle through the macro_block rows */
+        /* MB loop to set local segmentation map */
+        i = cpi->cyclic_refresh_mode_index;
+        assert(i < mbs_in_frame);
+        do
+        {
+          /* If the MB is as a candidate for clean up then mark it for
+           * possible boost/refresh (segment 1) The segment id may get
+           * reset to 0 later if the MB gets coded anything other than
+           * last frame 0,0 as only (last frame 0,0) MBs are eligable for
+           * refresh : that is to say Mbs likely to be background blocks.
+           */
+          if (cpi->cyclic_refresh_map[i] == 0)
+          {
+              seg_map[i] = 1;
+              block_count --;
+          }
+          else if (cpi->cyclic_refresh_map[i] < 0)
+              cpi->cyclic_refresh_map[i]++;
+
+          i++;
+          if (i == mbs_in_frame)
+              i = 0;
+
+        }
+        while(block_count && i != cpi->cyclic_refresh_mode_index);
+
+        cpi->cyclic_refresh_mode_index = i;
+
+#if CONFIG_TEMPORAL_DENOISING
+        if (cpi->oxcf.noise_sensitivity > 0) {
+          if (cpi->denoiser.denoiser_mode == kDenoiserOnYUVAggressive &&
+              Q < (int)cpi->denoiser.denoise_pars.qp_thresh &&
+              (cpi->frames_since_key >
+               2 * cpi->denoiser.denoise_pars.consec_zerolast)) {
+            // Under aggressive denoising, use segmentation to turn off loop
+            // filter below some qp thresh. The filter is reduced for all
+            // blocks that have been encoded as ZEROMV LAST x frames in a row,
+            // where x is set by cpi->denoiser.denoise_pars.consec_zerolast.
+            // This is to avoid "dot" artifacts that can occur from repeated
+            // loop filtering on noisy input source.
+            cpi->cyclic_refresh_q = Q;
+            // lf_adjustment = -MAX_LOOP_FILTER;
+            lf_adjustment = -40;
+            for (i = 0; i < mbs_in_frame; ++i) {
+              seg_map[i] = (cpi->consec_zero_last[i] >
+                            cpi->denoiser.denoise_pars.consec_zerolast) ? 1 : 0;
+            }
+          }
+        }
+#endif
+    }
+
+    /* Activate segmentation. */
+    cpi->mb.e_mbd.update_mb_segmentation_map = 1;
+    cpi->mb.e_mbd.update_mb_segmentation_data = 1;
+    enable_segmentation(cpi);
+
+    /* Set up the quant segment data */
+    feature_data[MB_LVL_ALT_Q][0] = 0;
+    feature_data[MB_LVL_ALT_Q][1] = (cpi->cyclic_refresh_q - Q);
+    feature_data[MB_LVL_ALT_Q][2] = 0;
+    feature_data[MB_LVL_ALT_Q][3] = 0;
+
+    /* Set up the loop segment data */
+    feature_data[MB_LVL_ALT_LF][0] = 0;
+    feature_data[MB_LVL_ALT_LF][1] = lf_adjustment;
+    feature_data[MB_LVL_ALT_LF][2] = 0;
+    feature_data[MB_LVL_ALT_LF][3] = 0;
+
+    /* Initialise the feature data structure */
+    set_segment_data(cpi, &feature_data[0][0], SEGMENT_DELTADATA);
+
+}
+
+static void set_default_lf_deltas(VP8_COMP *cpi)
+{
+    cpi->mb.e_mbd.mode_ref_lf_delta_enabled = 1;
+    cpi->mb.e_mbd.mode_ref_lf_delta_update = 1;
+
+    memset(cpi->mb.e_mbd.ref_lf_deltas, 0, sizeof(cpi->mb.e_mbd.ref_lf_deltas));
+    memset(cpi->mb.e_mbd.mode_lf_deltas, 0, sizeof(cpi->mb.e_mbd.mode_lf_deltas));
+
+    /* Test of ref frame deltas */
+    cpi->mb.e_mbd.ref_lf_deltas[INTRA_FRAME] = 2;
+    cpi->mb.e_mbd.ref_lf_deltas[LAST_FRAME] = 0;
+    cpi->mb.e_mbd.ref_lf_deltas[GOLDEN_FRAME] = -2;
+    cpi->mb.e_mbd.ref_lf_deltas[ALTREF_FRAME] = -2;
+
+    cpi->mb.e_mbd.mode_lf_deltas[0] = 4;               /* BPRED */
+
+    if(cpi->oxcf.Mode == MODE_REALTIME)
+      cpi->mb.e_mbd.mode_lf_deltas[1] = -12;              /* Zero */
+    else
+      cpi->mb.e_mbd.mode_lf_deltas[1] = -2;              /* Zero */
+
+    cpi->mb.e_mbd.mode_lf_deltas[2] = 2;               /* New mv */
+    cpi->mb.e_mbd.mode_lf_deltas[3] = 4;               /* Split mv */
+}
+
+/* Convenience macros for mapping speed and mode into a continuous
+ * range
+ */
+#define GOOD(x) (x+1)
+#define RT(x) (x+7)
+
+static int speed_map(int speed, const int *map)
+{
+    int res;
+
+    do
+    {
+        res = *map++;
+    } while(speed >= *map++);
+    return res;
+}
+
+static const int thresh_mult_map_znn[] = {
+    /* map common to zero, nearest, and near */
+    0, GOOD(2), 1500, GOOD(3), 2000, RT(0), 1000, RT(2), 2000, INT_MAX
+};
+
+static const int thresh_mult_map_vhpred[] = {
+    1000, GOOD(2), 1500, GOOD(3), 2000, RT(0), 1000, RT(1), 2000,
+    RT(7), INT_MAX, INT_MAX
+};
+
+static const int thresh_mult_map_bpred[] = {
+    2000, GOOD(0), 2500, GOOD(2), 5000, GOOD(3), 7500, RT(0), 2500, RT(1), 5000,
+    RT(6), INT_MAX, INT_MAX
+};
+
+static const int thresh_mult_map_tm[] = {
+    1000, GOOD(2), 1500, GOOD(3), 2000, RT(0), 0, RT(1), 1000, RT(2), 2000,
+    RT(7), INT_MAX, INT_MAX
+};
+
+static const int thresh_mult_map_new1[] = {
+    1000, GOOD(2), 2000, RT(0), 2000, INT_MAX
+};
+
+static const int thresh_mult_map_new2[] = {
+    1000, GOOD(2), 2000, GOOD(3), 2500, GOOD(5), 4000, RT(0), 2000, RT(2), 2500,
+    RT(5), 4000, INT_MAX
+};
+
+static const int thresh_mult_map_split1[] = {
+    2500, GOOD(0), 1700, GOOD(2), 10000, GOOD(3), 25000, GOOD(4), INT_MAX,
+    RT(0), 5000, RT(1), 10000, RT(2), 25000, RT(3), INT_MAX, INT_MAX
+};
+
+static const int thresh_mult_map_split2[] = {
+    5000, GOOD(0), 4500, GOOD(2), 20000, GOOD(3), 50000, GOOD(4), INT_MAX,
+    RT(0), 10000, RT(1), 20000, RT(2), 50000, RT(3), INT_MAX, INT_MAX
+};
+
+static const int mode_check_freq_map_zn2[] = {
+    /* {zero,nearest}{2,3} */
+    0, RT(10), 1<<1, RT(11), 1<<2, RT(12), 1<<3, INT_MAX
+};
+
+static const int mode_check_freq_map_vhbpred[] = {
+    0, GOOD(5), 2, RT(0), 0, RT(3), 2, RT(5), 4, INT_MAX
+};
+
+static const int mode_check_freq_map_near2[] = {
+    0, GOOD(5), 2, RT(0), 0, RT(3), 2, RT(10), 1<<2, RT(11), 1<<3, RT(12), 1<<4,
+    INT_MAX
+};
+
+static const int mode_check_freq_map_new1[] = {
+    0, RT(10), 1<<1, RT(11), 1<<2, RT(12), 1<<3, INT_MAX
+};
+
+static const int mode_check_freq_map_new2[] = {
+    0, GOOD(5), 4, RT(0), 0, RT(3), 4, RT(10), 1<<3, RT(11), 1<<4, RT(12), 1<<5,
+    INT_MAX
+};
+
+static const int mode_check_freq_map_split1[] = {
+    0, GOOD(2), 2, GOOD(3), 7, RT(1), 2, RT(2), 7, INT_MAX
+};
+
+static const int mode_check_freq_map_split2[] = {
+    0, GOOD(1), 2, GOOD(2), 4, GOOD(3), 15, RT(1), 4, RT(2), 15, INT_MAX
+};
+
+void vp8_set_speed_features(VP8_COMP *cpi)
+{
+    SPEED_FEATURES *sf = &cpi->sf;
+    int Mode = cpi->compressor_speed;
+    int Speed = cpi->Speed;
+    int i;
+    VP8_COMMON *cm = &cpi->common;
+    int last_improved_quant = sf->improved_quant;
+    int ref_frames;
+
+    /* Initialise default mode frequency sampling variables */
+    for (i = 0; i < MAX_MODES; i ++)
+    {
+        cpi->mode_check_freq[i] = 0;
+    }
+
+    cpi->mb.mbs_tested_so_far = 0;
+    cpi->mb.mbs_zero_last_dot_suppress = 0;
+
+    /* best quality defaults */
+    sf->RD = 1;
+    sf->search_method = NSTEP;
+    sf->improved_quant = 1;
+    sf->improved_dct = 1;
+    sf->auto_filter = 1;
+    sf->recode_loop = 1;
+    sf->quarter_pixel_search = 1;
+    sf->half_pixel_search = 1;
+    sf->iterative_sub_pixel = 1;
+    sf->optimize_coefficients = 1;
+    sf->use_fastquant_for_pick = 0;
+    sf->no_skip_block4x4_search = 1;
+
+    sf->first_step = 0;
+    sf->max_step_search_steps = MAX_MVSEARCH_STEPS;
+    sf->improved_mv_pred = 1;
+
+    /* default thresholds to 0 */
+    for (i = 0; i < MAX_MODES; i++)
+        sf->thresh_mult[i] = 0;
+
+    /* Count enabled references */
+    ref_frames = 1;
+    if (cpi->ref_frame_flags & VP8_LAST_FRAME)
+        ref_frames++;
+    if (cpi->ref_frame_flags & VP8_GOLD_FRAME)
+        ref_frames++;
+    if (cpi->ref_frame_flags & VP8_ALTR_FRAME)
+        ref_frames++;
+
+    /* Convert speed to continuous range, with clamping */
+    if (Mode == 0)
+        Speed = 0;
+    else if (Mode == 2)
+        Speed = RT(Speed);
+    else
+    {
+        if (Speed > 5)
+            Speed = 5;
+        Speed = GOOD(Speed);
+    }
+
+    sf->thresh_mult[THR_ZERO1] =
+    sf->thresh_mult[THR_NEAREST1] =
+    sf->thresh_mult[THR_NEAR1] =
+    sf->thresh_mult[THR_DC] = 0; /* always */
+
+    sf->thresh_mult[THR_ZERO2] =
+    sf->thresh_mult[THR_ZERO3] =
+    sf->thresh_mult[THR_NEAREST2] =
+    sf->thresh_mult[THR_NEAREST3] =
+    sf->thresh_mult[THR_NEAR2]  =
+    sf->thresh_mult[THR_NEAR3]  = speed_map(Speed, thresh_mult_map_znn);
+
+    sf->thresh_mult[THR_V_PRED] =
+    sf->thresh_mult[THR_H_PRED] = speed_map(Speed, thresh_mult_map_vhpred);
+    sf->thresh_mult[THR_B_PRED] = speed_map(Speed, thresh_mult_map_bpred);
+    sf->thresh_mult[THR_TM]     = speed_map(Speed, thresh_mult_map_tm);
+    sf->thresh_mult[THR_NEW1]   = speed_map(Speed, thresh_mult_map_new1);
+    sf->thresh_mult[THR_NEW2]   =
+    sf->thresh_mult[THR_NEW3]   = speed_map(Speed, thresh_mult_map_new2);
+    sf->thresh_mult[THR_SPLIT1] = speed_map(Speed, thresh_mult_map_split1);
+    sf->thresh_mult[THR_SPLIT2] =
+    sf->thresh_mult[THR_SPLIT3] = speed_map(Speed, thresh_mult_map_split2);
+
+    // Special case for temporal layers.
+    // Reduce the thresholds for zero/nearest/near for GOLDEN, if GOLDEN is
+    // used as second reference. We don't modify thresholds for ALTREF case
+    // since ALTREF is usually used as long-term reference in temporal layers.
+    if ((cpi->Speed <= 6) &&
+        (cpi->oxcf.number_of_layers > 1) &&
+        (cpi->ref_frame_flags & VP8_LAST_FRAME) &&
+        (cpi->ref_frame_flags & VP8_GOLD_FRAME)) {
+      if (cpi->closest_reference_frame == GOLDEN_FRAME) {
+        sf->thresh_mult[THR_ZERO2] =  sf->thresh_mult[THR_ZERO2] >> 3;
+        sf->thresh_mult[THR_NEAREST2] = sf->thresh_mult[THR_NEAREST2] >> 3;
+        sf->thresh_mult[THR_NEAR2]  = sf->thresh_mult[THR_NEAR2] >> 3;
+      } else {
+        sf->thresh_mult[THR_ZERO2] =  sf->thresh_mult[THR_ZERO2] >> 1;
+        sf->thresh_mult[THR_NEAREST2] = sf->thresh_mult[THR_NEAREST2] >> 1;
+        sf->thresh_mult[THR_NEAR2]  = sf->thresh_mult[THR_NEAR2] >> 1;
+      }
+    }
+
+    cpi->mode_check_freq[THR_ZERO1] =
+    cpi->mode_check_freq[THR_NEAREST1] =
+    cpi->mode_check_freq[THR_NEAR1] =
+    cpi->mode_check_freq[THR_TM]     =
+    cpi->mode_check_freq[THR_DC] = 0; /* always */
+
+    cpi->mode_check_freq[THR_ZERO2] =
+    cpi->mode_check_freq[THR_ZERO3] =
+    cpi->mode_check_freq[THR_NEAREST2] =
+    cpi->mode_check_freq[THR_NEAREST3] = speed_map(Speed,
+                                                   mode_check_freq_map_zn2);
+
+    cpi->mode_check_freq[THR_NEAR2]  =
+    cpi->mode_check_freq[THR_NEAR3]  = speed_map(Speed,
+                                                 mode_check_freq_map_near2);
+
+    cpi->mode_check_freq[THR_V_PRED] =
+    cpi->mode_check_freq[THR_H_PRED] =
+    cpi->mode_check_freq[THR_B_PRED] = speed_map(Speed,
+                                                 mode_check_freq_map_vhbpred);
+    cpi->mode_check_freq[THR_NEW1]   = speed_map(Speed,
+                                                 mode_check_freq_map_new1);
+    cpi->mode_check_freq[THR_NEW2]   =
+    cpi->mode_check_freq[THR_NEW3]   = speed_map(Speed,
+                                                 mode_check_freq_map_new2);
+    cpi->mode_check_freq[THR_SPLIT1] = speed_map(Speed,
+                                                 mode_check_freq_map_split1);
+    cpi->mode_check_freq[THR_SPLIT2] =
+    cpi->mode_check_freq[THR_SPLIT3] = speed_map(Speed,
+                                                 mode_check_freq_map_split2);
+    Speed = cpi->Speed;
+    switch (Mode)
+    {
+#if !CONFIG_REALTIME_ONLY
+    case 0: /* best quality mode */
+        sf->first_step = 0;
+        sf->max_step_search_steps = MAX_MVSEARCH_STEPS;
+        break;
+    case 1:
+    case 3:
+        if (Speed > 0)
+        {
+            /* Disable coefficient optimization above speed 0 */
+            sf->optimize_coefficients = 0;
+            sf->use_fastquant_for_pick = 1;
+            sf->no_skip_block4x4_search = 0;
+
+            sf->first_step = 1;
+        }
+
+        if (Speed > 2)
+        {
+            sf->improved_quant = 0;
+            sf->improved_dct = 0;
+
+            /* Only do recode loop on key frames, golden frames and
+             * alt ref frames
+             */
+            sf->recode_loop = 2;
+
+        }
+
+        if (Speed > 3)
+        {
+            sf->auto_filter = 1;
+            sf->recode_loop = 0; /* recode loop off */
+            sf->RD = 0;         /* Turn rd off */
+
+        }
+
+        if (Speed > 4)
+        {
+            sf->auto_filter = 0;  /* Faster selection of loop filter */
+        }
+
+        break;
+#endif
+    case 2:
+        sf->optimize_coefficients = 0;
+        sf->recode_loop = 0;
+        sf->auto_filter = 1;
+        sf->iterative_sub_pixel = 1;
+        sf->search_method = NSTEP;
+
+        if (Speed > 0)
+        {
+            sf->improved_quant = 0;
+            sf->improved_dct = 0;
+
+            sf->use_fastquant_for_pick = 1;
+            sf->no_skip_block4x4_search = 0;
+            sf->first_step = 1;
+        }
+
+        if (Speed > 2)
+            sf->auto_filter = 0;  /* Faster selection of loop filter */
+
+        if (Speed > 3)
+        {
+            sf->RD = 0;
+            sf->auto_filter = 1;
+        }
+
+        if (Speed > 4)
+        {
+            sf->auto_filter = 0;  /* Faster selection of loop filter */
+            sf->search_method = HEX;
+            sf->iterative_sub_pixel = 0;
+        }
+
+        if (Speed > 6)
+        {
+            unsigned int sum = 0;
+            unsigned int total_mbs = cm->MBs;
+            int thresh;
+            unsigned int total_skip;
+
+            int min = 2000;
+
+            if (cpi->oxcf.encode_breakout > 2000)
+                min = cpi->oxcf.encode_breakout;
+
+            min >>= 7;
+
+            for (i = 0; i < min; i++)
+            {
+                sum += cpi->mb.error_bins[i];
+            }
+
+            total_skip = sum;
+            sum = 0;
+
+            /* i starts from 2 to make sure thresh started from 2048 */
+            for (; i < 1024; i++)
+            {
+                sum += cpi->mb.error_bins[i];
+
+                if (10 * sum >= (unsigned int)(cpi->Speed - 6)*(total_mbs - total_skip))
+                    break;
+            }
+
+            i--;
+            thresh = (i << 7);
+
+            if (thresh < 2000)
+                thresh = 2000;
+
+            if (ref_frames > 1)
+            {
+                sf->thresh_mult[THR_NEW1 ] = thresh;
+                sf->thresh_mult[THR_NEAREST1  ] = thresh >> 1;
+                sf->thresh_mult[THR_NEAR1     ] = thresh >> 1;
+            }
+
+            if (ref_frames > 2)
+            {
+                sf->thresh_mult[THR_NEW2] = thresh << 1;
+                sf->thresh_mult[THR_NEAREST2 ] = thresh;
+                sf->thresh_mult[THR_NEAR2    ] = thresh;
+            }
+
+            if (ref_frames > 3)
+            {
+                sf->thresh_mult[THR_NEW3] = thresh << 1;
+                sf->thresh_mult[THR_NEAREST3 ] = thresh;
+                sf->thresh_mult[THR_NEAR3    ] = thresh;
+            }
+
+            sf->improved_mv_pred = 0;
+        }
+
+        if (Speed > 8)
+            sf->quarter_pixel_search = 0;
+
+        if(cm->version == 0)
+        {
+            cm->filter_type = NORMAL_LOOPFILTER;
+
+            if (Speed >= 14)
+                cm->filter_type = SIMPLE_LOOPFILTER;
+        }
+        else
+        {
+            cm->filter_type = SIMPLE_LOOPFILTER;
+        }
+
+        /* This has a big hit on quality. Last resort */
+        if (Speed >= 15)
+            sf->half_pixel_search = 0;
+
+        memset(cpi->mb.error_bins, 0, sizeof(cpi->mb.error_bins));
+
+    }; /* switch */
+
+    /* Slow quant, dct and trellis not worthwhile for first pass
+     * so make sure they are always turned off.
+     */
+    if ( cpi->pass == 1 )
+    {
+        sf->improved_quant = 0;
+        sf->optimize_coefficients = 0;
+        sf->improved_dct = 0;
+    }
+
+    if (cpi->sf.search_method == NSTEP)
+    {
+        vp8_init3smotion_compensation(&cpi->mb, cm->yv12_fb[cm->lst_fb_idx].y_stride);
+    }
+    else if (cpi->sf.search_method == DIAMOND)
+    {
+        vp8_init_dsmotion_compensation(&cpi->mb, cm->yv12_fb[cm->lst_fb_idx].y_stride);
+    }
+
+    if (cpi->sf.improved_dct)
+    {
+        cpi->mb.short_fdct8x4 = vp8_short_fdct8x4;
+        cpi->mb.short_fdct4x4 = vp8_short_fdct4x4;
+    }
+    else
+    {
+        /* No fast FDCT defined for any platform at this time. */
+        cpi->mb.short_fdct8x4 = vp8_short_fdct8x4;
+        cpi->mb.short_fdct4x4 = vp8_short_fdct4x4;
+    }
+
+    cpi->mb.short_walsh4x4 = vp8_short_walsh4x4;
+
+    if (cpi->sf.improved_quant)
+    {
+        cpi->mb.quantize_b      = vp8_regular_quantize_b;
+    }
+    else
+    {
+        cpi->mb.quantize_b      = vp8_fast_quantize_b;
+    }
+    if (cpi->sf.improved_quant != last_improved_quant)
+        vp8cx_init_quantizer(cpi);
+
+    if (cpi->sf.iterative_sub_pixel == 1)
+    {
+        cpi->find_fractional_mv_step = vp8_find_best_sub_pixel_step_iteratively;
+    }
+    else if (cpi->sf.quarter_pixel_search)
+    {
+        cpi->find_fractional_mv_step = vp8_find_best_sub_pixel_step;
+    }
+    else if (cpi->sf.half_pixel_search)
+    {
+        cpi->find_fractional_mv_step = vp8_find_best_half_pixel_step;
+    }
+    else
+    {
+        cpi->find_fractional_mv_step = vp8_skip_fractional_mv_step;
+    }
+
+    if (cpi->sf.optimize_coefficients == 1 && cpi->pass!=1)
+        cpi->mb.optimize = 1;
+    else
+        cpi->mb.optimize = 0;
+
+    if (cpi->common.full_pixel)
+        cpi->find_fractional_mv_step = vp8_skip_fractional_mv_step;
+
+#ifdef SPEEDSTATS
+    frames_at_speed[cpi->Speed]++;
+#endif
+}
+#undef GOOD
+#undef RT
+
+static void alloc_raw_frame_buffers(VP8_COMP *cpi)
+{
+#if VP8_TEMPORAL_ALT_REF
+    int width = (cpi->oxcf.Width + 15) & ~15;
+    int height = (cpi->oxcf.Height + 15) & ~15;
+#endif
+
+    cpi->lookahead = vp8_lookahead_init(cpi->oxcf.Width, cpi->oxcf.Height,
+                                        cpi->oxcf.lag_in_frames);
+    if(!cpi->lookahead)
+        vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate lag buffers");
+
+#if VP8_TEMPORAL_ALT_REF
+
+    if (vp8_yv12_alloc_frame_buffer(&cpi->alt_ref_buffer,
+                                    width, height, VP8BORDERINPIXELS))
+        vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate altref buffer");
+
+#endif
+}
+
+
+static void dealloc_raw_frame_buffers(VP8_COMP *cpi)
+{
+#if VP8_TEMPORAL_ALT_REF
+    vp8_yv12_de_alloc_frame_buffer(&cpi->alt_ref_buffer);
+#endif
+    vp8_lookahead_destroy(cpi->lookahead);
+}
+
+
+static int vp8_alloc_partition_data(VP8_COMP *cpi)
+{
+        vpx_free(cpi->mb.pip);
+
+    cpi->mb.pip = vpx_calloc((cpi->common.mb_cols + 1) *
+                                (cpi->common.mb_rows + 1),
+                                sizeof(PARTITION_INFO));
+    if(!cpi->mb.pip)
+        return 1;
+
+    cpi->mb.pi = cpi->mb.pip + cpi->common.mode_info_stride + 1;
+
+    return 0;
+}
+
+void vp8_alloc_compressor_data(VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = & cpi->common;
+
+    int width = cm->Width;
+    int height = cm->Height;
+#if CONFIG_MULTITHREAD
+    int prev_mb_rows = cm->mb_rows;
+#endif
+
+    if (vp8_alloc_frame_buffers(cm, width, height))
+        vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate frame buffers");
+
+    if (vp8_alloc_partition_data(cpi))
+        vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate partition data");
+
+
+    if ((width & 0xf) != 0)
+        width += 16 - (width & 0xf);
+
+    if ((height & 0xf) != 0)
+        height += 16 - (height & 0xf);
+
+
+    if (vp8_yv12_alloc_frame_buffer(&cpi->pick_lf_lvl_frame,
+                                    width, height, VP8BORDERINPIXELS))
+        vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate last frame buffer");
+
+    if (vp8_yv12_alloc_frame_buffer(&cpi->scaled_source,
+                                    width, height, VP8BORDERINPIXELS))
+        vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate scaled source buffer");
+
+    vpx_free(cpi->tok);
+
+    {
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+        unsigned int tokens = 8 * 24 * 16; /* one MB for each thread */
+#else
+        unsigned int tokens = cm->mb_rows * cm->mb_cols * 24 * 16;
+#endif
+        CHECK_MEM_ERROR(cpi->tok, vpx_calloc(tokens, sizeof(*cpi->tok)));
+    }
+
+    /* Data used for real time vc mode to see if gf needs refreshing */
+    cpi->zeromv_count = 0;
+
+
+    /* Structures used to monitor GF usage */
+    vpx_free(cpi->gf_active_flags);
+    CHECK_MEM_ERROR(cpi->gf_active_flags,
+                    vpx_calloc(sizeof(*cpi->gf_active_flags),
+                    cm->mb_rows * cm->mb_cols));
+    cpi->gf_active_count = cm->mb_rows * cm->mb_cols;
+
+    vpx_free(cpi->mb_activity_map);
+    CHECK_MEM_ERROR(cpi->mb_activity_map,
+                    vpx_calloc(sizeof(*cpi->mb_activity_map),
+                    cm->mb_rows * cm->mb_cols));
+
+    /* allocate memory for storing last frame's MVs for MV prediction. */
+    vpx_free(cpi->lfmv);
+    CHECK_MEM_ERROR(cpi->lfmv, vpx_calloc((cm->mb_rows+2) * (cm->mb_cols+2),
+                    sizeof(*cpi->lfmv)));
+    vpx_free(cpi->lf_ref_frame_sign_bias);
+    CHECK_MEM_ERROR(cpi->lf_ref_frame_sign_bias,
+                    vpx_calloc((cm->mb_rows+2) * (cm->mb_cols+2),
+                    sizeof(*cpi->lf_ref_frame_sign_bias)));
+    vpx_free(cpi->lf_ref_frame);
+    CHECK_MEM_ERROR(cpi->lf_ref_frame,
+                    vpx_calloc((cm->mb_rows+2) * (cm->mb_cols+2),
+                    sizeof(*cpi->lf_ref_frame)));
+
+    /* Create the encoder segmentation map and set all entries to 0 */
+    vpx_free(cpi->segmentation_map);
+    CHECK_MEM_ERROR(cpi->segmentation_map,
+                    vpx_calloc(cm->mb_rows * cm->mb_cols,
+                    sizeof(*cpi->segmentation_map)));
+    cpi->cyclic_refresh_mode_index = 0;
+    vpx_free(cpi->active_map);
+    CHECK_MEM_ERROR(cpi->active_map,
+                    vpx_calloc(cm->mb_rows * cm->mb_cols,
+                    sizeof(*cpi->active_map)));
+    memset(cpi->active_map , 1, (cm->mb_rows * cm->mb_cols));
+
+#if CONFIG_MULTITHREAD
+    if (width < 640)
+        cpi->mt_sync_range = 1;
+    else if (width <= 1280)
+        cpi->mt_sync_range = 4;
+    else if (width <= 2560)
+        cpi->mt_sync_range = 8;
+    else
+        cpi->mt_sync_range = 16;
+
+    if (cpi->oxcf.multi_threaded > 1)
+    {
+        int i;
+
+        /* De-allocate and re-allocate mutex */
+        if (cpi->pmutex != NULL) {
+            for (i = 0; i < prev_mb_rows; i++) {
+                pthread_mutex_destroy(&cpi->pmutex[i]);
+            }
+            vpx_free(cpi->pmutex);
+            cpi->pmutex = NULL;
+        }
+
+        CHECK_MEM_ERROR(cpi->pmutex, vpx_malloc(sizeof(*cpi->pmutex) *
+                                                cm->mb_rows));
+        if (cpi->pmutex) {
+            for (i = 0; i < cm->mb_rows; i++) {
+                pthread_mutex_init(&cpi->pmutex[i], NULL);
+            }
+        }
+
+        vpx_free(cpi->mt_current_mb_col);
+        CHECK_MEM_ERROR(cpi->mt_current_mb_col,
+                    vpx_malloc(sizeof(*cpi->mt_current_mb_col) * cm->mb_rows));
+    }
+
+#endif
+
+    vpx_free(cpi->tplist);
+    CHECK_MEM_ERROR(cpi->tplist, vpx_malloc(sizeof(TOKENLIST) * cm->mb_rows));
+
+#if CONFIG_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity > 0) {
+      vp8_denoiser_free(&cpi->denoiser);
+      if (vp8_denoiser_allocate(&cpi->denoiser, width, height,
+                                cm->mb_rows, cm->mb_cols,
+                                cpi->oxcf.noise_sensitivity))
+          vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                             "Failed to allocate denoiser");
+    }
+#endif
+}
+
+
+/* Quant MOD */
+static const int q_trans[] =
+{
+    0,   1,  2,  3,  4,  5,  7,  8,
+    9,  10, 12, 13, 15, 17, 18, 19,
+    20,  21, 23, 24, 25, 26, 27, 28,
+    29,  30, 31, 33, 35, 37, 39, 41,
+    43,  45, 47, 49, 51, 53, 55, 57,
+    59,  61, 64, 67, 70, 73, 76, 79,
+    82,  85, 88, 91, 94, 97, 100, 103,
+    106, 109, 112, 115, 118, 121, 124, 127,
+};
+
+int vp8_reverse_trans(int x)
+{
+    int i;
+
+    for (i = 0; i < 64; i++)
+        if (q_trans[i] >= x)
+            return i;
+
+    return 63;
+}
+void vp8_new_framerate(VP8_COMP *cpi, double framerate)
+{
+    if(framerate < .1)
+        framerate = 30;
+
+    cpi->framerate              = framerate;
+    cpi->output_framerate       = framerate;
+    cpi->per_frame_bandwidth    = (int)(cpi->oxcf.target_bandwidth /
+                                  cpi->output_framerate);
+    cpi->av_per_frame_bandwidth = cpi->per_frame_bandwidth;
+    cpi->min_frame_bandwidth    = (int)(cpi->av_per_frame_bandwidth *
+                                  cpi->oxcf.two_pass_vbrmin_section / 100);
+
+    /* Set Maximum gf/arf interval */
+    cpi->max_gf_interval = ((int)(cpi->output_framerate / 2.0) + 2);
+
+    if(cpi->max_gf_interval < 12)
+        cpi->max_gf_interval = 12;
+
+    /* Extended interval for genuinely static scenes */
+    cpi->twopass.static_scene_max_gf_interval = cpi->key_frame_frequency >> 1;
+
+     /* Special conditions when altr ref frame enabled in lagged compress mode */
+    if (cpi->oxcf.play_alternate && cpi->oxcf.lag_in_frames)
+    {
+        if (cpi->max_gf_interval > cpi->oxcf.lag_in_frames - 1)
+            cpi->max_gf_interval = cpi->oxcf.lag_in_frames - 1;
+
+        if (cpi->twopass.static_scene_max_gf_interval > cpi->oxcf.lag_in_frames - 1)
+            cpi->twopass.static_scene_max_gf_interval = cpi->oxcf.lag_in_frames - 1;
+    }
+
+    if ( cpi->max_gf_interval > cpi->twopass.static_scene_max_gf_interval )
+        cpi->max_gf_interval = cpi->twopass.static_scene_max_gf_interval;
+}
+
+
+static void init_config(VP8_COMP *cpi, VP8_CONFIG *oxcf)
+{
+    VP8_COMMON *cm = &cpi->common;
+
+    cpi->oxcf = *oxcf;
+
+    cpi->auto_gold = 1;
+    cpi->auto_adjust_gold_quantizer = 1;
+
+    cm->version = oxcf->Version;
+    vp8_setup_version(cm);
+
+    /* Frame rate is not available on the first frame, as it's derived from
+     * the observed timestamps. The actual value used here doesn't matter
+     * too much, as it will adapt quickly.
+     */
+    if (oxcf->timebase.num > 0) {
+      cpi->framerate = (double)(oxcf->timebase.den) /
+                       (double)(oxcf->timebase.num);
+    } else {
+      cpi->framerate = 30;
+    }
+
+    /* If the reciprocal of the timebase seems like a reasonable framerate,
+     * then use that as a guess, otherwise use 30.
+     */
+    if (cpi->framerate > 180)
+        cpi->framerate = 30;
+
+    cpi->ref_framerate = cpi->framerate;
+
+    cpi->ref_frame_flags = VP8_ALTR_FRAME | VP8_GOLD_FRAME | VP8_LAST_FRAME;
+
+    cm->refresh_golden_frame = 0;
+    cm->refresh_last_frame = 1;
+    cm->refresh_entropy_probs = 1;
+
+    /* change includes all joint functionality */
+    vp8_change_config(cpi, oxcf);
+
+    /* Initialize active best and worst q and average q values. */
+    cpi->active_worst_quality         = cpi->oxcf.worst_allowed_q;
+    cpi->active_best_quality          = cpi->oxcf.best_allowed_q;
+    cpi->avg_frame_qindex             = cpi->oxcf.worst_allowed_q;
+
+    /* Initialise the starting buffer levels */
+    cpi->buffer_level                 = cpi->oxcf.starting_buffer_level;
+    cpi->bits_off_target              = cpi->oxcf.starting_buffer_level;
+
+    cpi->rolling_target_bits          = cpi->av_per_frame_bandwidth;
+    cpi->rolling_actual_bits          = cpi->av_per_frame_bandwidth;
+    cpi->long_rolling_target_bits     = cpi->av_per_frame_bandwidth;
+    cpi->long_rolling_actual_bits     = cpi->av_per_frame_bandwidth;
+
+    cpi->total_actual_bits            = 0;
+    cpi->total_target_vs_actual       = 0;
+
+    /* Temporal scalabilty */
+    if (cpi->oxcf.number_of_layers > 1)
+    {
+        unsigned int i;
+        double prev_layer_framerate=0;
+
+        for (i=0; i<cpi->oxcf.number_of_layers; i++)
+        {
+            init_temporal_layer_context(cpi, oxcf, i, prev_layer_framerate);
+            prev_layer_framerate = cpi->output_framerate /
+                                   cpi->oxcf.rate_decimator[i];
+        }
+    }
+
+#if VP8_TEMPORAL_ALT_REF
+    {
+        int i;
+
+        cpi->fixed_divide[0] = 0;
+
+        for (i = 1; i < 512; i++)
+            cpi->fixed_divide[i] = 0x80000 / i;
+    }
+#endif
+}
+
+static void update_layer_contexts (VP8_COMP *cpi)
+{
+    VP8_CONFIG *oxcf = &cpi->oxcf;
+
+    /* Update snapshots of the layer contexts to reflect new parameters */
+    if (oxcf->number_of_layers > 1)
+    {
+        unsigned int i;
+        double prev_layer_framerate=0;
+
+        assert(oxcf->number_of_layers <= VPX_TS_MAX_LAYERS);
+        for (i = 0; i < oxcf->number_of_layers && i < VPX_TS_MAX_LAYERS; ++i)
+        {
+            LAYER_CONTEXT *lc = &cpi->layer_context[i];
+
+            lc->framerate =
+                cpi->ref_framerate / oxcf->rate_decimator[i];
+            lc->target_bandwidth = oxcf->target_bitrate[i] * 1000;
+
+            lc->starting_buffer_level = rescale(
+                          (int)oxcf->starting_buffer_level_in_ms,
+                          lc->target_bandwidth, 1000);
+
+            if (oxcf->optimal_buffer_level == 0)
+                lc->optimal_buffer_level = lc->target_bandwidth / 8;
+            else
+                lc->optimal_buffer_level = rescale(
+                          (int)oxcf->optimal_buffer_level_in_ms,
+                          lc->target_bandwidth, 1000);
+
+            if (oxcf->maximum_buffer_size == 0)
+                lc->maximum_buffer_size = lc->target_bandwidth / 8;
+            else
+                lc->maximum_buffer_size = rescale(
+                          (int)oxcf->maximum_buffer_size_in_ms,
+                          lc->target_bandwidth, 1000);
+
+            /* Work out the average size of a frame within this layer */
+            if (i > 0)
+                lc->avg_frame_size_for_layer =
+                   (int)((oxcf->target_bitrate[i] -
+                          oxcf->target_bitrate[i-1]) * 1000 /
+                          (lc->framerate - prev_layer_framerate));
+
+            prev_layer_framerate = lc->framerate;
+        }
+    }
+}
+
+void vp8_change_config(VP8_COMP *cpi, VP8_CONFIG *oxcf)
+{
+    VP8_COMMON *cm = &cpi->common;
+    int last_w, last_h;
+    unsigned int prev_number_of_layers;
+
+    if (!cpi)
+        return;
+
+    if (!oxcf)
+        return;
+
+    if (cm->version != oxcf->Version)
+    {
+        cm->version = oxcf->Version;
+        vp8_setup_version(cm);
+    }
+
+    last_w = cpi->oxcf.Width;
+    last_h = cpi->oxcf.Height;
+    prev_number_of_layers = cpi->oxcf.number_of_layers;
+
+    cpi->oxcf = *oxcf;
+
+    switch (cpi->oxcf.Mode)
+    {
+
+    case MODE_REALTIME:
+        cpi->pass = 0;
+        cpi->compressor_speed = 2;
+
+        if (cpi->oxcf.cpu_used < -16)
+        {
+            cpi->oxcf.cpu_used = -16;
+        }
+
+        if (cpi->oxcf.cpu_used > 16)
+            cpi->oxcf.cpu_used = 16;
+
+        break;
+
+    case MODE_GOODQUALITY:
+        cpi->pass = 0;
+        cpi->compressor_speed = 1;
+
+        if (cpi->oxcf.cpu_used < -5)
+        {
+            cpi->oxcf.cpu_used = -5;
+        }
+
+        if (cpi->oxcf.cpu_used > 5)
+            cpi->oxcf.cpu_used = 5;
+
+        break;
+
+    case MODE_BESTQUALITY:
+        cpi->pass = 0;
+        cpi->compressor_speed = 0;
+        break;
+
+    case MODE_FIRSTPASS:
+        cpi->pass = 1;
+        cpi->compressor_speed = 1;
+        break;
+    case MODE_SECONDPASS:
+        cpi->pass = 2;
+        cpi->compressor_speed = 1;
+
+        if (cpi->oxcf.cpu_used < -5)
+        {
+            cpi->oxcf.cpu_used = -5;
+        }
+
+        if (cpi->oxcf.cpu_used > 5)
+            cpi->oxcf.cpu_used = 5;
+
+        break;
+    case MODE_SECONDPASS_BEST:
+        cpi->pass = 2;
+        cpi->compressor_speed = 0;
+        break;
+    }
+
+    if (cpi->pass == 0)
+        cpi->auto_worst_q = 1;
+
+    cpi->oxcf.worst_allowed_q = q_trans[oxcf->worst_allowed_q];
+    cpi->oxcf.best_allowed_q = q_trans[oxcf->best_allowed_q];
+    cpi->oxcf.cq_level = q_trans[cpi->oxcf.cq_level];
+
+    if (oxcf->fixed_q >= 0)
+    {
+        if (oxcf->worst_allowed_q < 0)
+            cpi->oxcf.fixed_q = q_trans[0];
+        else
+            cpi->oxcf.fixed_q = q_trans[oxcf->worst_allowed_q];
+
+        if (oxcf->alt_q < 0)
+            cpi->oxcf.alt_q = q_trans[0];
+        else
+            cpi->oxcf.alt_q = q_trans[oxcf->alt_q];
+
+        if (oxcf->key_q < 0)
+            cpi->oxcf.key_q = q_trans[0];
+        else
+            cpi->oxcf.key_q = q_trans[oxcf->key_q];
+
+        if (oxcf->gold_q < 0)
+            cpi->oxcf.gold_q = q_trans[0];
+        else
+            cpi->oxcf.gold_q = q_trans[oxcf->gold_q];
+
+    }
+
+    cpi->baseline_gf_interval =
+        cpi->oxcf.alt_freq ? cpi->oxcf.alt_freq : DEFAULT_GF_INTERVAL;
+
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+    cpi->oxcf.token_partitions = 3;
+#endif
+
+    if (cpi->oxcf.token_partitions >= 0 && cpi->oxcf.token_partitions <= 3)
+        cm->multi_token_partition =
+            (TOKEN_PARTITION) cpi->oxcf.token_partitions;
+
+    setup_features(cpi);
+
+    {
+        int i;
+
+        for (i = 0; i < MAX_MB_SEGMENTS; i++)
+            cpi->segment_encode_breakout[i] = cpi->oxcf.encode_breakout;
+    }
+
+    /* At the moment the first order values may not be > MAXQ */
+    if (cpi->oxcf.fixed_q > MAXQ)
+        cpi->oxcf.fixed_q = MAXQ;
+
+    /* local file playback mode == really big buffer */
+    if (cpi->oxcf.end_usage == USAGE_LOCAL_FILE_PLAYBACK)
+    {
+        cpi->oxcf.starting_buffer_level       = 60000;
+        cpi->oxcf.optimal_buffer_level        = 60000;
+        cpi->oxcf.maximum_buffer_size         = 240000;
+        cpi->oxcf.starting_buffer_level_in_ms = 60000;
+        cpi->oxcf.optimal_buffer_level_in_ms  = 60000;
+        cpi->oxcf.maximum_buffer_size_in_ms   = 240000;
+    }
+
+    /* Convert target bandwidth from Kbit/s to Bit/s */
+    cpi->oxcf.target_bandwidth       *= 1000;
+
+    cpi->oxcf.starting_buffer_level =
+        rescale((int)cpi->oxcf.starting_buffer_level,
+                cpi->oxcf.target_bandwidth, 1000);
+
+    /* Set or reset optimal and maximum buffer levels. */
+    if (cpi->oxcf.optimal_buffer_level == 0)
+        cpi->oxcf.optimal_buffer_level = cpi->oxcf.target_bandwidth / 8;
+    else
+        cpi->oxcf.optimal_buffer_level =
+            rescale((int)cpi->oxcf.optimal_buffer_level,
+                    cpi->oxcf.target_bandwidth, 1000);
+
+    if (cpi->oxcf.maximum_buffer_size == 0)
+        cpi->oxcf.maximum_buffer_size = cpi->oxcf.target_bandwidth / 8;
+    else
+        cpi->oxcf.maximum_buffer_size =
+            rescale((int)cpi->oxcf.maximum_buffer_size,
+                    cpi->oxcf.target_bandwidth, 1000);
+    // Under a configuration change, where maximum_buffer_size may change,
+    // keep buffer level clipped to the maximum allowed buffer size.
+    if (cpi->bits_off_target > cpi->oxcf.maximum_buffer_size) {
+      cpi->bits_off_target = cpi->oxcf.maximum_buffer_size;
+      cpi->buffer_level = cpi->bits_off_target;
+    }
+
+    /* Set up frame rate and related parameters rate control values. */
+    vp8_new_framerate(cpi, cpi->framerate);
+
+    /* Set absolute upper and lower quality limits */
+    cpi->worst_quality               = cpi->oxcf.worst_allowed_q;
+    cpi->best_quality                = cpi->oxcf.best_allowed_q;
+
+    /* active values should only be modified if out of new range */
+    if (cpi->active_worst_quality > cpi->oxcf.worst_allowed_q)
+    {
+      cpi->active_worst_quality = cpi->oxcf.worst_allowed_q;
+    }
+    /* less likely */
+    else if (cpi->active_worst_quality < cpi->oxcf.best_allowed_q)
+    {
+      cpi->active_worst_quality = cpi->oxcf.best_allowed_q;
+    }
+    if (cpi->active_best_quality < cpi->oxcf.best_allowed_q)
+    {
+      cpi->active_best_quality = cpi->oxcf.best_allowed_q;
+    }
+    /* less likely */
+    else if (cpi->active_best_quality > cpi->oxcf.worst_allowed_q)
+    {
+      cpi->active_best_quality = cpi->oxcf.worst_allowed_q;
+    }
+
+    cpi->buffered_mode = cpi->oxcf.optimal_buffer_level > 0;
+
+    cpi->cq_target_quality = cpi->oxcf.cq_level;
+
+    /* Only allow dropped frames in buffered mode */
+    cpi->drop_frames_allowed = cpi->oxcf.allow_df && cpi->buffered_mode;
+
+    cpi->target_bandwidth = cpi->oxcf.target_bandwidth;
+
+    // Check if the number of temporal layers has changed, and if so reset the
+    // pattern counter and set/initialize the temporal layer context for the
+    // new layer configuration.
+    if (cpi->oxcf.number_of_layers != prev_number_of_layers)
+    {
+        // If the number of temporal layers are changed we must start at the
+        // base of the pattern cycle, so set the layer id to 0 and reset
+        // the temporal pattern counter.
+        if (cpi->temporal_layer_id > 0) {
+          cpi->temporal_layer_id = 0;
+        }
+        cpi->temporal_pattern_counter = 0;
+        reset_temporal_layer_change(cpi, oxcf, prev_number_of_layers);
+    }
+
+    if (!cpi->initial_width)
+    {
+        cpi->initial_width = cpi->oxcf.Width;
+        cpi->initial_height = cpi->oxcf.Height;
+    }
+
+    cm->Width       = cpi->oxcf.Width;
+    cm->Height      = cpi->oxcf.Height;
+    assert(cm->Width <= cpi->initial_width);
+    assert(cm->Height <= cpi->initial_height);
+
+    /* TODO(jkoleszar): if an internal spatial resampling is active,
+     * and we downsize the input image, maybe we should clear the
+     * internal scale immediately rather than waiting for it to
+     * correct.
+     */
+
+    /* VP8 sharpness level mapping 0-7 (vs 0-10 in general VPx dialogs) */
+    if (cpi->oxcf.Sharpness > 7)
+        cpi->oxcf.Sharpness = 7;
+
+    cm->sharpness_level = cpi->oxcf.Sharpness;
+
+    if (cm->horiz_scale != NORMAL || cm->vert_scale != NORMAL)
+    {
+        int UNINITIALIZED_IS_SAFE(hr), UNINITIALIZED_IS_SAFE(hs);
+        int UNINITIALIZED_IS_SAFE(vr), UNINITIALIZED_IS_SAFE(vs);
+
+        Scale2Ratio(cm->horiz_scale, &hr, &hs);
+        Scale2Ratio(cm->vert_scale, &vr, &vs);
+
+        /* always go to the next whole number */
+        cm->Width = (hs - 1 + cpi->oxcf.Width * hr) / hs;
+        cm->Height = (vs - 1 + cpi->oxcf.Height * vr) / vs;
+    }
+
+    if (last_w != cpi->oxcf.Width || last_h != cpi->oxcf.Height)
+        cpi->force_next_frame_intra = 1;
+
+    if (((cm->Width + 15) & ~15) != cm->yv12_fb[cm->lst_fb_idx].y_width ||
+        ((cm->Height + 15) & ~15) != cm->yv12_fb[cm->lst_fb_idx].y_height ||
+        cm->yv12_fb[cm->lst_fb_idx].y_width == 0)
+    {
+        dealloc_raw_frame_buffers(cpi);
+        alloc_raw_frame_buffers(cpi);
+        vp8_alloc_compressor_data(cpi);
+    }
+
+    if (cpi->oxcf.fixed_q >= 0)
+    {
+        cpi->last_q[0] = cpi->oxcf.fixed_q;
+        cpi->last_q[1] = cpi->oxcf.fixed_q;
+    }
+
+    cpi->Speed = cpi->oxcf.cpu_used;
+
+    /* force to allowlag to 0 if lag_in_frames is 0; */
+    if (cpi->oxcf.lag_in_frames == 0)
+    {
+        cpi->oxcf.allow_lag = 0;
+    }
+    /* Limit on lag buffers as these are not currently dynamically allocated */
+    else if (cpi->oxcf.lag_in_frames > MAX_LAG_BUFFERS)
+        cpi->oxcf.lag_in_frames = MAX_LAG_BUFFERS;
+
+    /* YX Temp */
+    cpi->alt_ref_source = NULL;
+    cpi->is_src_frame_alt_ref = 0;
+
+#if CONFIG_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity)
+    {
+      if (!cpi->denoiser.yv12_mc_running_avg.buffer_alloc)
+      {
+        int width = (cpi->oxcf.Width + 15) & ~15;
+        int height = (cpi->oxcf.Height + 15) & ~15;
+        if (vp8_denoiser_allocate(&cpi->denoiser, width, height,
+                                  cm->mb_rows, cm->mb_cols,
+                                  cpi->oxcf.noise_sensitivity))
+            vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                               "Failed to allocate denoiser");
+      }
+    }
+#endif
+
+#if 0
+    /* Experimental RD Code */
+    cpi->frame_distortion = 0;
+    cpi->last_frame_distortion = 0;
+#endif
+
+}
+
+#ifndef M_LOG2_E
+#define M_LOG2_E 0.693147180559945309417
+#endif
+#define log2f(x) (log (x) / (float) M_LOG2_E)
+
+static void cal_mvsadcosts(int *mvsadcost[2])
+{
+    int i = 1;
+
+    mvsadcost [0] [0] = 300;
+    mvsadcost [1] [0] = 300;
+
+    do
+    {
+        double z = 256 * (2 * (log2f(8 * i) + .6));
+        mvsadcost [0][i] = (int) z;
+        mvsadcost [1][i] = (int) z;
+        mvsadcost [0][-i] = (int) z;
+        mvsadcost [1][-i] = (int) z;
+    }
+    while (++i <= mvfp_max);
+}
+
+struct VP8_COMP* vp8_create_compressor(VP8_CONFIG *oxcf)
+{
+    int i;
+
+    VP8_COMP *cpi;
+    VP8_COMMON *cm;
+
+    cpi = vpx_memalign(32, sizeof(VP8_COMP));
+    /* Check that the CPI instance is valid */
+    if (!cpi)
+        return 0;
+
+    cm = &cpi->common;
+
+    memset(cpi, 0, sizeof(VP8_COMP));
+
+    if (setjmp(cm->error.jmp))
+    {
+        cpi->common.error.setjmp = 0;
+        vp8_remove_compressor(&cpi);
+        return 0;
+    }
+
+    cpi->common.error.setjmp = 1;
+
+    CHECK_MEM_ERROR(cpi->mb.ss, vpx_calloc(sizeof(search_site), (MAX_MVSEARCH_STEPS * 8) + 1));
+
+    vp8_create_common(&cpi->common);
+
+    init_config(cpi, oxcf);
+
+    memcpy(cpi->base_skip_false_prob, vp8cx_base_skip_false_prob, sizeof(vp8cx_base_skip_false_prob));
+    cpi->common.current_video_frame   = 0;
+    cpi->temporal_pattern_counter     = 0;
+    cpi->temporal_layer_id            = -1;
+    cpi->kf_overspend_bits            = 0;
+    cpi->kf_bitrate_adjustment        = 0;
+    cpi->frames_till_gf_update_due      = 0;
+    cpi->gf_overspend_bits            = 0;
+    cpi->non_gf_bitrate_adjustment     = 0;
+    cpi->prob_last_coded              = 128;
+    cpi->prob_gf_coded                = 128;
+    cpi->prob_intra_coded             = 63;
+
+    /* Prime the recent reference frame usage counters.
+     * Hereafter they will be maintained as a sort of moving average
+     */
+    cpi->recent_ref_frame_usage[INTRA_FRAME]  = 1;
+    cpi->recent_ref_frame_usage[LAST_FRAME]   = 1;
+    cpi->recent_ref_frame_usage[GOLDEN_FRAME] = 1;
+    cpi->recent_ref_frame_usage[ALTREF_FRAME] = 1;
+
+    /* Set reference frame sign bias for ALTREF frame to 1 (for now) */
+    cpi->common.ref_frame_sign_bias[ALTREF_FRAME] = 1;
+
+    cpi->twopass.gf_decay_rate = 0;
+    cpi->baseline_gf_interval = DEFAULT_GF_INTERVAL;
+
+    cpi->gold_is_last = 0 ;
+    cpi->alt_is_last  = 0 ;
+    cpi->gold_is_alt  = 0 ;
+
+    cpi->active_map_enabled = 0;
+
+#if 0
+    /* Experimental code for lagged and one pass */
+    /* Initialise one_pass GF frames stats */
+    /* Update stats used for GF selection */
+    if (cpi->pass == 0)
+    {
+        cpi->one_pass_frame_index = 0;
+
+        for (i = 0; i < MAX_LAG_BUFFERS; i++)
+        {
+            cpi->one_pass_frame_stats[i].frames_so_far = 0;
+            cpi->one_pass_frame_stats[i].frame_intra_error = 0.0;
+            cpi->one_pass_frame_stats[i].frame_coded_error = 0.0;
+            cpi->one_pass_frame_stats[i].frame_pcnt_inter = 0.0;
+            cpi->one_pass_frame_stats[i].frame_pcnt_motion = 0.0;
+            cpi->one_pass_frame_stats[i].frame_mvr = 0.0;
+            cpi->one_pass_frame_stats[i].frame_mvr_abs = 0.0;
+            cpi->one_pass_frame_stats[i].frame_mvc = 0.0;
+            cpi->one_pass_frame_stats[i].frame_mvc_abs = 0.0;
+        }
+    }
+#endif
+
+    cpi->mse_source_denoised = 0;
+
+    /* Should we use the cyclic refresh method.
+     * Currently this is tied to error resilliant mode
+     */
+    cpi->cyclic_refresh_mode_enabled = cpi->oxcf.error_resilient_mode;
+    cpi->cyclic_refresh_mode_max_mbs_perframe = (cpi->common.mb_rows * cpi->common.mb_cols) / 7;
+    if (cpi->oxcf.number_of_layers == 1) {
+        cpi->cyclic_refresh_mode_max_mbs_perframe =
+            (cpi->common.mb_rows * cpi->common.mb_cols) / 20;
+    } else if (cpi->oxcf.number_of_layers == 2) {
+        cpi->cyclic_refresh_mode_max_mbs_perframe =
+            (cpi->common.mb_rows * cpi->common.mb_cols) / 10;
+    }
+    cpi->cyclic_refresh_mode_index = 0;
+    cpi->cyclic_refresh_q = 32;
+
+    if (cpi->cyclic_refresh_mode_enabled)
+    {
+        CHECK_MEM_ERROR(cpi->cyclic_refresh_map, vpx_calloc((cpi->common.mb_rows * cpi->common.mb_cols), 1));
+    }
+    else
+        cpi->cyclic_refresh_map = (signed char *) NULL;
+
+    CHECK_MEM_ERROR(cpi->consec_zero_last,
+                    vpx_calloc(cm->mb_rows * cm->mb_cols, 1));
+    CHECK_MEM_ERROR(cpi->consec_zero_last_mvbias,
+                    vpx_calloc((cpi->common.mb_rows * cpi->common.mb_cols), 1));
+
+#ifdef VP8_ENTROPY_STATS
+    init_context_counters();
+#endif
+
+    /*Initialize the feed-forward activity masking.*/
+    cpi->activity_avg = 90<<12;
+
+    /* Give a sensible default for the first frame. */
+    cpi->frames_since_key = 8;
+    cpi->key_frame_frequency = cpi->oxcf.key_freq;
+    cpi->this_key_frame_forced = 0;
+    cpi->next_key_frame_forced = 0;
+
+    cpi->source_alt_ref_pending = 0;
+    cpi->source_alt_ref_active = 0;
+    cpi->common.refresh_alt_ref_frame = 0;
+
+    cpi->force_maxqp = 0;
+
+    cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
+#if CONFIG_INTERNAL_STATS
+    cpi->b_calculate_ssimg = 0;
+
+    cpi->count = 0;
+    cpi->bytes = 0;
+
+    if (cpi->b_calculate_psnr)
+    {
+        cpi->total_sq_error = 0.0;
+        cpi->total_sq_error2 = 0.0;
+        cpi->total_y = 0.0;
+        cpi->total_u = 0.0;
+        cpi->total_v = 0.0;
+        cpi->total = 0.0;
+        cpi->totalp_y = 0.0;
+        cpi->totalp_u = 0.0;
+        cpi->totalp_v = 0.0;
+        cpi->totalp = 0.0;
+        cpi->tot_recode_hits = 0;
+        cpi->summed_quality = 0;
+        cpi->summed_weights = 0;
+    }
+
+    if (cpi->b_calculate_ssimg)
+    {
+        cpi->total_ssimg_y = 0;
+        cpi->total_ssimg_u = 0;
+        cpi->total_ssimg_v = 0;
+        cpi->total_ssimg_all = 0;
+    }
+
+#endif
+
+    cpi->first_time_stamp_ever = 0x7FFFFFFF;
+
+    cpi->frames_till_gf_update_due      = 0;
+    cpi->key_frame_count              = 1;
+
+    cpi->ni_av_qi                     = cpi->oxcf.worst_allowed_q;
+    cpi->ni_tot_qi                    = 0;
+    cpi->ni_frames                   = 0;
+    cpi->total_byte_count             = 0;
+
+    cpi->drop_frame                  = 0;
+
+    cpi->rate_correction_factor         = 1.0;
+    cpi->key_frame_rate_correction_factor = 1.0;
+    cpi->gf_rate_correction_factor  = 1.0;
+    cpi->twopass.est_max_qcorrection_factor  = 1.0;
+
+    for (i = 0; i < KEY_FRAME_CONTEXT; i++)
+    {
+        cpi->prior_key_frame_distance[i] = (int)cpi->output_framerate;
+    }
+
+#ifdef OUTPUT_YUV_SRC
+    yuv_file = fopen("bd.yuv", "ab");
+#endif
+#ifdef OUTPUT_YUV_DENOISED
+    yuv_denoised_file = fopen("denoised.yuv", "ab");
+#endif
+
+#if 0
+    framepsnr = fopen("framepsnr.stt", "a");
+    kf_list = fopen("kf_list.stt", "w");
+#endif
+
+    cpi->output_pkt_list = oxcf->output_pkt_list;
+
+#if !CONFIG_REALTIME_ONLY
+
+    if (cpi->pass == 1)
+    {
+        vp8_init_first_pass(cpi);
+    }
+    else if (cpi->pass == 2)
+    {
+        size_t packet_sz = sizeof(FIRSTPASS_STATS);
+        int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz);
+
+        cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf;
+        cpi->twopass.stats_in = cpi->twopass.stats_in_start;
+        cpi->twopass.stats_in_end = (void*)((char *)cpi->twopass.stats_in
+                            + (packets - 1) * packet_sz);
+        vp8_init_second_pass(cpi);
+    }
+
+#endif
+
+    if (cpi->compressor_speed == 2)
+    {
+        cpi->avg_encode_time      = 0;
+        cpi->avg_pick_mode_time    = 0;
+    }
+
+    vp8_set_speed_features(cpi);
+
+    /* Set starting values of RD threshold multipliers (128 = *1) */
+    for (i = 0; i < MAX_MODES; i++)
+    {
+        cpi->mb.rd_thresh_mult[i] = 128;
+    }
+
+#ifdef VP8_ENTROPY_STATS
+    init_mv_ref_counts();
+#endif
+
+#if CONFIG_MULTITHREAD
+    if(vp8cx_create_encoder_threads(cpi))
+    {
+        vp8_remove_compressor(&cpi);
+        return 0;
+    }
+#endif
+
+    cpi->fn_ptr[BLOCK_16X16].sdf            = vpx_sad16x16;
+    cpi->fn_ptr[BLOCK_16X16].vf             = vpx_variance16x16;
+    cpi->fn_ptr[BLOCK_16X16].svf            = vpx_sub_pixel_variance16x16;
+    cpi->fn_ptr[BLOCK_16X16].svf_halfpix_h  = vpx_variance_halfpixvar16x16_h;
+    cpi->fn_ptr[BLOCK_16X16].svf_halfpix_v  = vpx_variance_halfpixvar16x16_v;
+    cpi->fn_ptr[BLOCK_16X16].svf_halfpix_hv = vpx_variance_halfpixvar16x16_hv;
+    cpi->fn_ptr[BLOCK_16X16].sdx3f          = vpx_sad16x16x3;
+    cpi->fn_ptr[BLOCK_16X16].sdx8f          = vpx_sad16x16x8;
+    cpi->fn_ptr[BLOCK_16X16].sdx4df         = vpx_sad16x16x4d;
+
+    cpi->fn_ptr[BLOCK_16X8].sdf            = vpx_sad16x8;
+    cpi->fn_ptr[BLOCK_16X8].vf             = vpx_variance16x8;
+    cpi->fn_ptr[BLOCK_16X8].svf            = vpx_sub_pixel_variance16x8;
+    cpi->fn_ptr[BLOCK_16X8].svf_halfpix_h  = NULL;
+    cpi->fn_ptr[BLOCK_16X8].svf_halfpix_v  = NULL;
+    cpi->fn_ptr[BLOCK_16X8].svf_halfpix_hv = NULL;
+    cpi->fn_ptr[BLOCK_16X8].sdx3f          = vpx_sad16x8x3;
+    cpi->fn_ptr[BLOCK_16X8].sdx8f          = vpx_sad16x8x8;
+    cpi->fn_ptr[BLOCK_16X8].sdx4df         = vpx_sad16x8x4d;
+
+    cpi->fn_ptr[BLOCK_8X16].sdf            = vpx_sad8x16;
+    cpi->fn_ptr[BLOCK_8X16].vf             = vpx_variance8x16;
+    cpi->fn_ptr[BLOCK_8X16].svf            = vpx_sub_pixel_variance8x16;
+    cpi->fn_ptr[BLOCK_8X16].svf_halfpix_h  = NULL;
+    cpi->fn_ptr[BLOCK_8X16].svf_halfpix_v  = NULL;
+    cpi->fn_ptr[BLOCK_8X16].svf_halfpix_hv = NULL;
+    cpi->fn_ptr[BLOCK_8X16].sdx3f          = vpx_sad8x16x3;
+    cpi->fn_ptr[BLOCK_8X16].sdx8f          = vpx_sad8x16x8;
+    cpi->fn_ptr[BLOCK_8X16].sdx4df         = vpx_sad8x16x4d;
+
+    cpi->fn_ptr[BLOCK_8X8].sdf            = vpx_sad8x8;
+    cpi->fn_ptr[BLOCK_8X8].vf             = vpx_variance8x8;
+    cpi->fn_ptr[BLOCK_8X8].svf            = vpx_sub_pixel_variance8x8;
+    cpi->fn_ptr[BLOCK_8X8].svf_halfpix_h  = NULL;
+    cpi->fn_ptr[BLOCK_8X8].svf_halfpix_v  = NULL;
+    cpi->fn_ptr[BLOCK_8X8].svf_halfpix_hv = NULL;
+    cpi->fn_ptr[BLOCK_8X8].sdx3f          = vpx_sad8x8x3;
+    cpi->fn_ptr[BLOCK_8X8].sdx8f          = vpx_sad8x8x8;
+    cpi->fn_ptr[BLOCK_8X8].sdx4df         = vpx_sad8x8x4d;
+
+    cpi->fn_ptr[BLOCK_4X4].sdf            = vpx_sad4x4;
+    cpi->fn_ptr[BLOCK_4X4].vf             = vpx_variance4x4;
+    cpi->fn_ptr[BLOCK_4X4].svf            = vpx_sub_pixel_variance4x4;
+    cpi->fn_ptr[BLOCK_4X4].svf_halfpix_h  = NULL;
+    cpi->fn_ptr[BLOCK_4X4].svf_halfpix_v  = NULL;
+    cpi->fn_ptr[BLOCK_4X4].svf_halfpix_hv = NULL;
+    cpi->fn_ptr[BLOCK_4X4].sdx3f          = vpx_sad4x4x3;
+    cpi->fn_ptr[BLOCK_4X4].sdx8f          = vpx_sad4x4x8;
+    cpi->fn_ptr[BLOCK_4X4].sdx4df         = vpx_sad4x4x4d;
+
+#if ARCH_X86 || ARCH_X86_64
+    cpi->fn_ptr[BLOCK_16X16].copymem      = vp8_copy32xn;
+    cpi->fn_ptr[BLOCK_16X8].copymem       = vp8_copy32xn;
+    cpi->fn_ptr[BLOCK_8X16].copymem       = vp8_copy32xn;
+    cpi->fn_ptr[BLOCK_8X8].copymem        = vp8_copy32xn;
+    cpi->fn_ptr[BLOCK_4X4].copymem        = vp8_copy32xn;
+#endif
+
+    cpi->full_search_sad = vp8_full_search_sad;
+    cpi->diamond_search_sad = vp8_diamond_search_sad;
+    cpi->refining_search_sad = vp8_refining_search_sad;
+
+    /* make sure frame 1 is okay */
+    cpi->mb.error_bins[0] = cpi->common.MBs;
+
+    /* vp8cx_init_quantizer() is first called here. Add check in
+     * vp8cx_frame_init_quantizer() so that vp8cx_init_quantizer is only
+     * called later when needed. This will avoid unnecessary calls of
+     * vp8cx_init_quantizer() for every frame.
+     */
+    vp8cx_init_quantizer(cpi);
+
+    vp8_loop_filter_init(cm);
+
+    cpi->common.error.setjmp = 0;
+
+#if CONFIG_MULTI_RES_ENCODING
+
+    /* Calculate # of MBs in a row in lower-resolution level image. */
+    if (cpi->oxcf.mr_encoder_id > 0)
+        vp8_cal_low_res_mb_cols(cpi);
+
+#endif
+
+    /* setup RD costs to MACROBLOCK struct */
+
+    cpi->mb.mvcost[0] = &cpi->rd_costs.mvcosts[0][mv_max+1];
+    cpi->mb.mvcost[1] = &cpi->rd_costs.mvcosts[1][mv_max+1];
+    cpi->mb.mvsadcost[0] = &cpi->rd_costs.mvsadcosts[0][mvfp_max+1];
+    cpi->mb.mvsadcost[1] = &cpi->rd_costs.mvsadcosts[1][mvfp_max+1];
+
+    cal_mvsadcosts(cpi->mb.mvsadcost);
+
+    cpi->mb.mbmode_cost = cpi->rd_costs.mbmode_cost;
+    cpi->mb.intra_uv_mode_cost = cpi->rd_costs.intra_uv_mode_cost;
+    cpi->mb.bmode_costs = cpi->rd_costs.bmode_costs;
+    cpi->mb.inter_bmode_costs = cpi->rd_costs.inter_bmode_costs;
+    cpi->mb.token_costs = cpi->rd_costs.token_costs;
+
+    /* setup block ptrs & offsets */
+    vp8_setup_block_ptrs(&cpi->mb);
+    vp8_setup_block_dptrs(&cpi->mb.e_mbd);
+
+    return  cpi;
+}
+
+
+void vp8_remove_compressor(VP8_COMP **ptr)
+{
+    VP8_COMP *cpi = *ptr;
+
+    if (!cpi)
+        return;
+
+    if (cpi && (cpi->common.current_video_frame > 0))
+    {
+#if !CONFIG_REALTIME_ONLY
+
+        if (cpi->pass == 2)
+        {
+            vp8_end_second_pass(cpi);
+        }
+
+#endif
+
+#ifdef VP8_ENTROPY_STATS
+        print_context_counters();
+        print_tree_update_probs();
+        print_mode_context();
+#endif
+
+#if CONFIG_INTERNAL_STATS
+
+        if (cpi->pass != 1)
+        {
+            FILE *f = fopen("opsnr.stt", "a");
+            double time_encoded = (cpi->last_end_time_stamp_seen
+                                   - cpi->first_time_stamp_ever) / 10000000.000;
+            double total_encode_time = (cpi->time_receive_data +
+                                            cpi->time_compress_data) / 1000.000;
+            double dr = (double)cpi->bytes * 8.0 / 1000.0 / time_encoded;
+            const double target_rate = (double)cpi->oxcf.target_bandwidth / 1000;
+            const double rate_err = ((100.0 * (dr - target_rate)) / target_rate);
+
+            if (cpi->b_calculate_psnr)
+            {
+                if (cpi->oxcf.number_of_layers > 1)
+                {
+                    int i;
+
+                    fprintf(f, "Layer\tBitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\t"
+                               "GLPsnrP\tVPXSSIM\t\n");
+                    for (i=0; i<(int)cpi->oxcf.number_of_layers; i++)
+                    {
+                        double dr = (double)cpi->bytes_in_layer[i] *
+                                              8.0 / 1000.0  / time_encoded;
+                        double samples = 3.0 / 2 * cpi->frames_in_layer[i] *
+                                         cpi->common.Width * cpi->common.Height;
+                        double total_psnr =
+                            vpx_sse_to_psnr(samples, 255.0,
+                                            cpi->total_error2[i]);
+                        double total_psnr2 =
+                            vpx_sse_to_psnr(samples, 255.0,
+                                            cpi->total_error2_p[i]);
+                        double total_ssim = 100 * pow(cpi->sum_ssim[i] /
+                                                      cpi->sum_weights[i], 8.0);
+
+                        fprintf(f, "%5d\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
+                                   "%7.3f\t%7.3f\n",
+                                   i, dr,
+                                   cpi->sum_psnr[i] / cpi->frames_in_layer[i],
+                                   total_psnr,
+                                   cpi->sum_psnr_p[i] / cpi->frames_in_layer[i],
+                                   total_psnr2, total_ssim);
+                    }
+                }
+                else
+                {
+                    double samples = 3.0 / 2 * cpi->count *
+                                     cpi->common.Width * cpi->common.Height;
+                    double total_psnr = vpx_sse_to_psnr(samples, 255.0,
+                                                        cpi->total_sq_error);
+                    double total_psnr2 = vpx_sse_to_psnr(samples, 255.0,
+                                                         cpi->total_sq_error2);
+                    double total_ssim = 100 * pow(cpi->summed_quality /
+                                                      cpi->summed_weights, 8.0);
+
+                    fprintf(f, "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\t"
+                               "GLPsnrP\tVPXSSIM\t  Time(us)  Rc-Err "
+                               "Abs Err\n");
+                    fprintf(f, "%7.3f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
+                               "%7.3f\t%8.0f %7.2f %7.2f\n",
+                               dr, cpi->total / cpi->count, total_psnr,
+                               cpi->totalp / cpi->count, total_psnr2,
+                               total_ssim, total_encode_time,
+                               rate_err, fabs(rate_err));
+                }
+            }
+
+            if (cpi->b_calculate_ssimg)
+            {
+                if (cpi->oxcf.number_of_layers > 1)
+                {
+                    int i;
+
+                    fprintf(f, "Layer\tBitRate\tSSIM_Y\tSSIM_U\tSSIM_V\tSSIM_A\t"
+                               "Time(us)\n");
+                    for (i=0; i<(int)cpi->oxcf.number_of_layers; i++)
+                    {
+                        double dr = (double)cpi->bytes_in_layer[i] *
+                                    8.0 / 1000.0  / time_encoded;
+                        fprintf(f, "%5d\t%7.3f\t%6.4f\t"
+                                "%6.4f\t%6.4f\t%6.4f\t%8.0f\n",
+                                i, dr,
+                                cpi->total_ssimg_y_in_layer[i] /
+                                     cpi->frames_in_layer[i],
+                                cpi->total_ssimg_u_in_layer[i] /
+                                     cpi->frames_in_layer[i],
+                                cpi->total_ssimg_v_in_layer[i] /
+                                     cpi->frames_in_layer[i],
+                                cpi->total_ssimg_all_in_layer[i] /
+                                     cpi->frames_in_layer[i],
+                                total_encode_time);
+                    }
+                }
+                else
+                {
+                    fprintf(f, "BitRate\tSSIM_Y\tSSIM_U\tSSIM_V\tSSIM_A\t"
+                               "Time(us)\n");
+                    fprintf(f, "%7.3f\t%6.4f\t%6.4f\t%6.4f\t%6.4f\t%8.0f\n", dr,
+                            cpi->total_ssimg_y / cpi->count,
+                            cpi->total_ssimg_u / cpi->count,
+                            cpi->total_ssimg_v / cpi->count,
+                            cpi->total_ssimg_all / cpi->count, total_encode_time);
+                }
+            }
+
+            fclose(f);
+#if 0
+            f = fopen("qskip.stt", "a");
+            fprintf(f, "minq:%d -maxq:%d skiptrue:skipfalse = %d:%d\n", cpi->oxcf.best_allowed_q, cpi->oxcf.worst_allowed_q, skiptruecount, skipfalsecount);
+            fclose(f);
+#endif
+
+        }
+
+#endif
+
+
+#ifdef SPEEDSTATS
+
+        if (cpi->compressor_speed == 2)
+        {
+            int i;
+            FILE *f = fopen("cxspeed.stt", "a");
+            cnt_pm /= cpi->common.MBs;
+
+            for (i = 0; i < 16; i++)
+                fprintf(f, "%5d", frames_at_speed[i]);
+
+            fprintf(f, "\n");
+            fclose(f);
+        }
+
+#endif
+
+
+#ifdef MODE_STATS
+        {
+            extern int count_mb_seg[4];
+            FILE *f = fopen("modes.stt", "a");
+            double dr = (double)cpi->framerate * (double)bytes * (double)8 / (double)count / (double)1000 ;
+            fprintf(f, "intra_mode in Intra Frames:\n");
+            fprintf(f, "Y: %8d, %8d, %8d, %8d, %8d\n", y_modes[0], y_modes[1], y_modes[2], y_modes[3], y_modes[4]);
+            fprintf(f, "UV:%8d, %8d, %8d, %8d\n", uv_modes[0], uv_modes[1], uv_modes[2], uv_modes[3]);
+            fprintf(f, "B: ");
+            {
+                int i;
+
+                for (i = 0; i < 10; i++)
+                    fprintf(f, "%8d, ", b_modes[i]);
+
+                fprintf(f, "\n");
+
+            }
+
+            fprintf(f, "Modes in Inter Frames:\n");
+            fprintf(f, "Y: %8d, %8d, %8d, %8d, %8d, %8d, %8d, %8d, %8d, %8d\n",
+                    inter_y_modes[0], inter_y_modes[1], inter_y_modes[2], inter_y_modes[3], inter_y_modes[4],
+                    inter_y_modes[5], inter_y_modes[6], inter_y_modes[7], inter_y_modes[8], inter_y_modes[9]);
+            fprintf(f, "UV:%8d, %8d, %8d, %8d\n", inter_uv_modes[0], inter_uv_modes[1], inter_uv_modes[2], inter_uv_modes[3]);
+            fprintf(f, "B: ");
+            {
+                int i;
+
+                for (i = 0; i < 15; i++)
+                    fprintf(f, "%8d, ", inter_b_modes[i]);
+
+                fprintf(f, "\n");
+
+            }
+            fprintf(f, "P:%8d, %8d, %8d, %8d\n", count_mb_seg[0], count_mb_seg[1], count_mb_seg[2], count_mb_seg[3]);
+            fprintf(f, "PB:%8d, %8d, %8d, %8d\n", inter_b_modes[LEFT4X4], inter_b_modes[ABOVE4X4], inter_b_modes[ZERO4X4], inter_b_modes[NEW4X4]);
+
+
+
+            fclose(f);
+        }
+#endif
+
+#ifdef VP8_ENTROPY_STATS
+        {
+            int i, j, k;
+            FILE *fmode = fopen("modecontext.c", "w");
+
+            fprintf(fmode, "\n#include \"entropymode.h\"\n\n");
+            fprintf(fmode, "const unsigned int vp8_kf_default_bmode_counts ");
+            fprintf(fmode, "[VP8_BINTRAMODES] [VP8_BINTRAMODES] [VP8_BINTRAMODES] =\n{\n");
+
+            for (i = 0; i < 10; i++)
+            {
+
+                fprintf(fmode, "    { /* Above Mode :  %d */\n", i);
+
+                for (j = 0; j < 10; j++)
+                {
+
+                    fprintf(fmode, "        {");
+
+                    for (k = 0; k < 10; k++)
+                    {
+                        if (!intra_mode_stats[i][j][k])
+                            fprintf(fmode, " %5d, ", 1);
+                        else
+                            fprintf(fmode, " %5d, ", intra_mode_stats[i][j][k]);
+                    }
+
+                    fprintf(fmode, "}, /* left_mode %d */\n", j);
+
+                }
+
+                fprintf(fmode, "    },\n");
+
+            }
+
+            fprintf(fmode, "};\n");
+            fclose(fmode);
+        }
+#endif
+
+
+#if defined(SECTIONBITS_OUTPUT)
+
+        if (0)
+        {
+            int i;
+            FILE *f = fopen("tokenbits.stt", "a");
+
+            for (i = 0; i < 28; i++)
+                fprintf(f, "%8d", (int)(Sectionbits[i] / 256));
+
+            fprintf(f, "\n");
+            fclose(f);
+        }
+
+#endif
+
+#if 0
+        {
+            printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000);
+            printf("\n_frames recive_data encod_mb_row compress_frame  Total\n");
+            printf("%6d %10ld %10ld %10ld %10ld\n", cpi->common.current_video_frame, cpi->time_receive_data / 1000, cpi->time_encode_mb_row / 1000, cpi->time_compress_data / 1000, (cpi->time_receive_data + cpi->time_compress_data) / 1000);
+        }
+#endif
+
+    }
+
+#if CONFIG_MULTITHREAD
+    vp8cx_remove_encoder_threads(cpi);
+#endif
+
+#if CONFIG_TEMPORAL_DENOISING
+    vp8_denoiser_free(&cpi->denoiser);
+#endif
+    dealloc_compressor_data(cpi);
+    vpx_free(cpi->mb.ss);
+    vpx_free(cpi->tok);
+    vpx_free(cpi->cyclic_refresh_map);
+    vpx_free(cpi->consec_zero_last);
+    vpx_free(cpi->consec_zero_last_mvbias);
+
+    vp8_remove_common(&cpi->common);
+    vpx_free(cpi);
+    *ptr = 0;
+
+#ifdef OUTPUT_YUV_SRC
+    fclose(yuv_file);
+#endif
+#ifdef OUTPUT_YUV_DENOISED
+    fclose(yuv_denoised_file);
+#endif
+
+#if 0
+
+    if (keyfile)
+        fclose(keyfile);
+
+    if (framepsnr)
+        fclose(framepsnr);
+
+    if (kf_list)
+        fclose(kf_list);
+
+#endif
+
+}
+
+
+static uint64_t calc_plane_error(unsigned char *orig, int orig_stride,
+                                 unsigned char *recon, int recon_stride,
+                                 unsigned int cols, unsigned int rows)
+{
+    unsigned int row, col;
+    uint64_t total_sse = 0;
+    int diff;
+
+    for (row = 0; row + 16 <= rows; row += 16)
+    {
+        for (col = 0; col + 16 <= cols; col += 16)
+        {
+            unsigned int sse;
+
+            vpx_mse16x16(orig + col, orig_stride,
+                                            recon + col, recon_stride,
+                                            &sse);
+            total_sse += sse;
+        }
+
+        /* Handle odd-sized width */
+        if (col < cols)
+        {
+            unsigned int   border_row, border_col;
+            unsigned char *border_orig = orig;
+            unsigned char *border_recon = recon;
+
+            for (border_row = 0; border_row < 16; border_row++)
+            {
+                for (border_col = col; border_col < cols; border_col++)
+                {
+                    diff = border_orig[border_col] - border_recon[border_col];
+                    total_sse += diff * diff;
+                }
+
+                border_orig += orig_stride;
+                border_recon += recon_stride;
+            }
+        }
+
+        orig += orig_stride * 16;
+        recon += recon_stride * 16;
+    }
+
+    /* Handle odd-sized height */
+    for (; row < rows; row++)
+    {
+        for (col = 0; col < cols; col++)
+        {
+            diff = orig[col] - recon[col];
+            total_sse += diff * diff;
+        }
+
+        orig += orig_stride;
+        recon += recon_stride;
+    }
+
+    vp8_clear_system_state();
+    return total_sse;
+}
+
+
+static void generate_psnr_packet(VP8_COMP *cpi)
+{
+    YV12_BUFFER_CONFIG      *orig = cpi->Source;
+    YV12_BUFFER_CONFIG      *recon = cpi->common.frame_to_show;
+    struct vpx_codec_cx_pkt  pkt;
+    uint64_t                 sse;
+    int                      i;
+    unsigned int             width = cpi->common.Width;
+    unsigned int             height = cpi->common.Height;
+
+    pkt.kind = VPX_CODEC_PSNR_PKT;
+    sse = calc_plane_error(orig->y_buffer, orig->y_stride,
+                           recon->y_buffer, recon->y_stride,
+                           width, height);
+    pkt.data.psnr.sse[0] = sse;
+    pkt.data.psnr.sse[1] = sse;
+    pkt.data.psnr.samples[0] = width * height;
+    pkt.data.psnr.samples[1] = width * height;
+
+    width = (width + 1) / 2;
+    height = (height + 1) / 2;
+
+    sse = calc_plane_error(orig->u_buffer, orig->uv_stride,
+                           recon->u_buffer, recon->uv_stride,
+                           width, height);
+    pkt.data.psnr.sse[0] += sse;
+    pkt.data.psnr.sse[2] = sse;
+    pkt.data.psnr.samples[0] += width * height;
+    pkt.data.psnr.samples[2] = width * height;
+
+    sse = calc_plane_error(orig->v_buffer, orig->uv_stride,
+                           recon->v_buffer, recon->uv_stride,
+                           width, height);
+    pkt.data.psnr.sse[0] += sse;
+    pkt.data.psnr.sse[3] = sse;
+    pkt.data.psnr.samples[0] += width * height;
+    pkt.data.psnr.samples[3] = width * height;
+
+    for (i = 0; i < 4; i++)
+        pkt.data.psnr.psnr[i] = vpx_sse_to_psnr(pkt.data.psnr.samples[i], 255.0,
+                                                (double)(pkt.data.psnr.sse[i]));
+
+    vpx_codec_pkt_list_add(cpi->output_pkt_list, &pkt);
+}
+
+
+int vp8_use_as_reference(VP8_COMP *cpi, int ref_frame_flags)
+{
+    if (ref_frame_flags > 7)
+        return -1 ;
+
+    cpi->ref_frame_flags = ref_frame_flags;
+    return 0;
+}
+int vp8_update_reference(VP8_COMP *cpi, int ref_frame_flags)
+{
+    if (ref_frame_flags > 7)
+        return -1 ;
+
+    cpi->common.refresh_golden_frame = 0;
+    cpi->common.refresh_alt_ref_frame = 0;
+    cpi->common.refresh_last_frame   = 0;
+
+    if (ref_frame_flags & VP8_LAST_FRAME)
+        cpi->common.refresh_last_frame = 1;
+
+    if (ref_frame_flags & VP8_GOLD_FRAME)
+        cpi->common.refresh_golden_frame = 1;
+
+    if (ref_frame_flags & VP8_ALTR_FRAME)
+        cpi->common.refresh_alt_ref_frame = 1;
+
+    return 0;
+}
+
+int vp8_get_reference(VP8_COMP *cpi, enum vpx_ref_frame_type ref_frame_flag, YV12_BUFFER_CONFIG *sd)
+{
+    VP8_COMMON *cm = &cpi->common;
+    int ref_fb_idx;
+
+    if (ref_frame_flag == VP8_LAST_FRAME)
+        ref_fb_idx = cm->lst_fb_idx;
+    else if (ref_frame_flag == VP8_GOLD_FRAME)
+        ref_fb_idx = cm->gld_fb_idx;
+    else if (ref_frame_flag == VP8_ALTR_FRAME)
+        ref_fb_idx = cm->alt_fb_idx;
+    else
+        return -1;
+
+    vp8_yv12_copy_frame(&cm->yv12_fb[ref_fb_idx], sd);
+
+    return 0;
+}
+int vp8_set_reference(VP8_COMP *cpi, enum vpx_ref_frame_type ref_frame_flag, YV12_BUFFER_CONFIG *sd)
+{
+    VP8_COMMON *cm = &cpi->common;
+
+    int ref_fb_idx;
+
+    if (ref_frame_flag == VP8_LAST_FRAME)
+        ref_fb_idx = cm->lst_fb_idx;
+    else if (ref_frame_flag == VP8_GOLD_FRAME)
+        ref_fb_idx = cm->gld_fb_idx;
+    else if (ref_frame_flag == VP8_ALTR_FRAME)
+        ref_fb_idx = cm->alt_fb_idx;
+    else
+        return -1;
+
+    vp8_yv12_copy_frame(sd, &cm->yv12_fb[ref_fb_idx]);
+
+    return 0;
+}
+int vp8_update_entropy(VP8_COMP *cpi, int update)
+{
+    VP8_COMMON *cm = &cpi->common;
+    cm->refresh_entropy_probs = update;
+
+    return 0;
+}
+
+
+#if defined(OUTPUT_YUV_SRC) || defined(OUTPUT_YUV_DENOISED)
+void vp8_write_yuv_frame(FILE *yuv_file, YV12_BUFFER_CONFIG *s)
+{
+    unsigned char *src = s->y_buffer;
+    int h = s->y_height;
+
+    do
+    {
+        fwrite(src, s->y_width, 1,  yuv_file);
+        src += s->y_stride;
+    }
+    while (--h);
+
+    src = s->u_buffer;
+    h = s->uv_height;
+
+    do
+    {
+        fwrite(src, s->uv_width, 1,  yuv_file);
+        src += s->uv_stride;
+    }
+    while (--h);
+
+    src = s->v_buffer;
+    h = s->uv_height;
+
+    do
+    {
+        fwrite(src, s->uv_width, 1, yuv_file);
+        src += s->uv_stride;
+    }
+    while (--h);
+}
+#endif
+
+static void scale_and_extend_source(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = &cpi->common;
+
+    /* are we resizing the image */
+    if (cm->horiz_scale != 0 || cm->vert_scale != 0)
+    {
+#if CONFIG_SPATIAL_RESAMPLING
+        int UNINITIALIZED_IS_SAFE(hr), UNINITIALIZED_IS_SAFE(hs);
+        int UNINITIALIZED_IS_SAFE(vr), UNINITIALIZED_IS_SAFE(vs);
+        int tmp_height;
+
+        if (cm->vert_scale == 3)
+            tmp_height = 9;
+        else
+            tmp_height = 11;
+
+        Scale2Ratio(cm->horiz_scale, &hr, &hs);
+        Scale2Ratio(cm->vert_scale, &vr, &vs);
+
+        vpx_scale_frame(sd, &cpi->scaled_source, cm->temp_scale_frame.y_buffer,
+                        tmp_height, hs, hr, vs, vr, 0);
+
+        vp8_yv12_extend_frame_borders(&cpi->scaled_source);
+        cpi->Source = &cpi->scaled_source;
+#endif
+    }
+    else
+        cpi->Source = sd;
+}
+
+
+static int resize_key_frame(VP8_COMP *cpi)
+{
+#if CONFIG_SPATIAL_RESAMPLING
+    VP8_COMMON *cm = &cpi->common;
+
+    /* Do we need to apply resampling for one pass cbr.
+     * In one pass this is more limited than in two pass cbr.
+     * The test and any change is only made once per key frame sequence.
+     */
+    if (cpi->oxcf.allow_spatial_resampling && (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER))
+    {
+        int UNINITIALIZED_IS_SAFE(hr), UNINITIALIZED_IS_SAFE(hs);
+        int UNINITIALIZED_IS_SAFE(vr), UNINITIALIZED_IS_SAFE(vs);
+        int new_width, new_height;
+
+        /* If we are below the resample DOWN watermark then scale down a
+         * notch.
+         */
+        if (cpi->buffer_level < (cpi->oxcf.resample_down_water_mark * cpi->oxcf.optimal_buffer_level / 100))
+        {
+            cm->horiz_scale = (cm->horiz_scale < ONETWO) ? cm->horiz_scale + 1 : ONETWO;
+            cm->vert_scale = (cm->vert_scale < ONETWO) ? cm->vert_scale + 1 : ONETWO;
+        }
+        /* Should we now start scaling back up */
+        else if (cpi->buffer_level > (cpi->oxcf.resample_up_water_mark * cpi->oxcf.optimal_buffer_level / 100))
+        {
+            cm->horiz_scale = (cm->horiz_scale > NORMAL) ? cm->horiz_scale - 1 : NORMAL;
+            cm->vert_scale = (cm->vert_scale > NORMAL) ? cm->vert_scale - 1 : NORMAL;
+        }
+
+        /* Get the new height and width */
+        Scale2Ratio(cm->horiz_scale, &hr, &hs);
+        Scale2Ratio(cm->vert_scale, &vr, &vs);
+        new_width = ((hs - 1) + (cpi->oxcf.Width * hr)) / hs;
+        new_height = ((vs - 1) + (cpi->oxcf.Height * vr)) / vs;
+
+        /* If the image size has changed we need to reallocate the buffers
+         * and resample the source image
+         */
+        if ((cm->Width != new_width) || (cm->Height != new_height))
+        {
+            cm->Width = new_width;
+            cm->Height = new_height;
+            vp8_alloc_compressor_data(cpi);
+            scale_and_extend_source(cpi->un_scaled_source, cpi);
+            return 1;
+        }
+    }
+
+#endif
+    return 0;
+}
+
+
+static void update_alt_ref_frame_stats(VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = &cpi->common;
+
+    /* Select an interval before next GF or altref */
+    if (!cpi->auto_gold)
+        cpi->frames_till_gf_update_due = DEFAULT_GF_INTERVAL;
+
+    if ((cpi->pass != 2) && cpi->frames_till_gf_update_due)
+    {
+        cpi->current_gf_interval = cpi->frames_till_gf_update_due;
+
+        /* Set the bits per frame that we should try and recover in
+         * subsequent inter frames to account for the extra GF spend...
+         * note that his does not apply for GF updates that occur
+         * coincident with a key frame as the extra cost of key frames is
+         * dealt with elsewhere.
+         */
+        cpi->gf_overspend_bits += cpi->projected_frame_size;
+        cpi->non_gf_bitrate_adjustment = cpi->gf_overspend_bits / cpi->frames_till_gf_update_due;
+    }
+
+    /* Update data structure that monitors level of reference to last GF */
+    memset(cpi->gf_active_flags, 1, (cm->mb_rows * cm->mb_cols));
+    cpi->gf_active_count = cm->mb_rows * cm->mb_cols;
+
+    /* this frame refreshes means next frames don't unless specified by user */
+    cpi->frames_since_golden = 0;
+
+    /* Clear the alternate reference update pending flag. */
+    cpi->source_alt_ref_pending = 0;
+
+    /* Set the alternate reference frame active flag */
+    cpi->source_alt_ref_active = 1;
+
+
+}
+static void update_golden_frame_stats(VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = &cpi->common;
+
+    /* Update the Golden frame usage counts. */
+    if (cm->refresh_golden_frame)
+    {
+        /* Select an interval before next GF */
+        if (!cpi->auto_gold)
+            cpi->frames_till_gf_update_due = DEFAULT_GF_INTERVAL;
+
+        if ((cpi->pass != 2) && (cpi->frames_till_gf_update_due > 0))
+        {
+            cpi->current_gf_interval = cpi->frames_till_gf_update_due;
+
+            /* Set the bits per frame that we should try and recover in
+             * subsequent inter frames to account for the extra GF spend...
+             * note that his does not apply for GF updates that occur
+             * coincident with a key frame as the extra cost of key frames
+             * is dealt with elsewhere.
+             */
+            if ((cm->frame_type != KEY_FRAME) && !cpi->source_alt_ref_active)
+            {
+                /* Calcluate GF bits to be recovered
+                 * Projected size - av frame bits available for inter
+                 * frames for clip as a whole
+                 */
+                cpi->gf_overspend_bits += (cpi->projected_frame_size - cpi->inter_frame_target);
+            }
+
+            cpi->non_gf_bitrate_adjustment = cpi->gf_overspend_bits / cpi->frames_till_gf_update_due;
+
+        }
+
+        /* Update data structure that monitors level of reference to last GF */
+        memset(cpi->gf_active_flags, 1, (cm->mb_rows * cm->mb_cols));
+        cpi->gf_active_count = cm->mb_rows * cm->mb_cols;
+
+        /* this frame refreshes means next frames don't unless specified by
+         * user
+         */
+        cm->refresh_golden_frame = 0;
+        cpi->frames_since_golden = 0;
+
+        cpi->recent_ref_frame_usage[INTRA_FRAME] = 1;
+        cpi->recent_ref_frame_usage[LAST_FRAME] = 1;
+        cpi->recent_ref_frame_usage[GOLDEN_FRAME] = 1;
+        cpi->recent_ref_frame_usage[ALTREF_FRAME] = 1;
+
+        /* ******** Fixed Q test code only ************ */
+        /* If we are going to use the ALT reference for the next group of
+         * frames set a flag to say so.
+         */
+        if (cpi->oxcf.fixed_q >= 0 &&
+            cpi->oxcf.play_alternate && !cpi->common.refresh_alt_ref_frame)
+        {
+            cpi->source_alt_ref_pending = 1;
+            cpi->frames_till_gf_update_due = cpi->baseline_gf_interval;
+        }
+
+        if (!cpi->source_alt_ref_pending)
+            cpi->source_alt_ref_active = 0;
+
+        /* Decrement count down till next gf */
+        if (cpi->frames_till_gf_update_due > 0)
+            cpi->frames_till_gf_update_due--;
+
+    }
+    else if (!cpi->common.refresh_alt_ref_frame)
+    {
+        /* Decrement count down till next gf */
+        if (cpi->frames_till_gf_update_due > 0)
+            cpi->frames_till_gf_update_due--;
+
+        if (cpi->frames_till_alt_ref_frame)
+            cpi->frames_till_alt_ref_frame --;
+
+        cpi->frames_since_golden ++;
+
+        if (cpi->frames_since_golden > 1)
+        {
+            cpi->recent_ref_frame_usage[INTRA_FRAME] +=
+                cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME];
+            cpi->recent_ref_frame_usage[LAST_FRAME] +=
+                cpi->mb.count_mb_ref_frame_usage[LAST_FRAME];
+            cpi->recent_ref_frame_usage[GOLDEN_FRAME] +=
+                cpi->mb.count_mb_ref_frame_usage[GOLDEN_FRAME];
+            cpi->recent_ref_frame_usage[ALTREF_FRAME] +=
+                cpi->mb.count_mb_ref_frame_usage[ALTREF_FRAME];
+        }
+    }
+}
+
+/* This function updates the reference frame probability estimates that
+ * will be used during mode selection
+ */
+static void update_rd_ref_frame_probs(VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = &cpi->common;
+
+    const int *const rfct = cpi->mb.count_mb_ref_frame_usage;
+    const int rf_intra = rfct[INTRA_FRAME];
+    const int rf_inter = rfct[LAST_FRAME] + rfct[GOLDEN_FRAME] + rfct[ALTREF_FRAME];
+
+    if (cm->frame_type == KEY_FRAME)
+    {
+        cpi->prob_intra_coded = 255;
+        cpi->prob_last_coded  = 128;
+        cpi->prob_gf_coded  = 128;
+    }
+    else if (!(rf_intra + rf_inter))
+    {
+        cpi->prob_intra_coded = 63;
+        cpi->prob_last_coded  = 128;
+        cpi->prob_gf_coded    = 128;
+    }
+
+    /* update reference frame costs since we can do better than what we got
+     * last frame.
+     */
+    if (cpi->oxcf.number_of_layers == 1)
+    {
+        if (cpi->common.refresh_alt_ref_frame)
+        {
+            cpi->prob_intra_coded += 40;
+            if (cpi->prob_intra_coded > 255)
+                cpi->prob_intra_coded = 255;
+            cpi->prob_last_coded = 200;
+            cpi->prob_gf_coded = 1;
+        }
+        else if (cpi->frames_since_golden == 0)
+        {
+            cpi->prob_last_coded = 214;
+        }
+        else if (cpi->frames_since_golden == 1)
+        {
+            cpi->prob_last_coded = 192;
+            cpi->prob_gf_coded = 220;
+        }
+        else if (cpi->source_alt_ref_active)
+        {
+            cpi->prob_gf_coded -= 20;
+
+            if (cpi->prob_gf_coded < 10)
+                cpi->prob_gf_coded = 10;
+        }
+        if (!cpi->source_alt_ref_active)
+            cpi->prob_gf_coded = 255;
+    }
+}
+
+
+#if !CONFIG_REALTIME_ONLY
+/* 1 = key, 0 = inter */
+static int decide_key_frame(VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = &cpi->common;
+
+    int code_key_frame = 0;
+
+    cpi->kf_boost = 0;
+
+    if (cpi->Speed > 11)
+        return 0;
+
+    /* Clear down mmx registers */
+    vp8_clear_system_state();
+
+    if ((cpi->compressor_speed == 2) && (cpi->Speed >= 5) && (cpi->sf.RD == 0))
+    {
+        double change = 1.0 * abs((int)(cpi->mb.intra_error -
+            cpi->last_intra_error)) / (1 + cpi->last_intra_error);
+        double change2 = 1.0 * abs((int)(cpi->mb.prediction_error -
+            cpi->last_prediction_error)) / (1 + cpi->last_prediction_error);
+        double minerror = cm->MBs * 256;
+
+        cpi->last_intra_error = cpi->mb.intra_error;
+        cpi->last_prediction_error = cpi->mb.prediction_error;
+
+        if (10 * cpi->mb.intra_error / (1 + cpi->mb.prediction_error) < 15
+            && cpi->mb.prediction_error > minerror
+            && (change > .25 || change2 > .25))
+        {
+            /*(change > 1.4 || change < .75)&& cpi->this_frame_percent_intra > cpi->last_frame_percent_intra + 3*/
+            return 1;
+        }
+
+        return 0;
+
+    }
+
+    /* If the following are true we might as well code a key frame */
+    if (((cpi->this_frame_percent_intra == 100) &&
+         (cpi->this_frame_percent_intra > (cpi->last_frame_percent_intra + 2))) ||
+        ((cpi->this_frame_percent_intra > 95) &&
+         (cpi->this_frame_percent_intra >= (cpi->last_frame_percent_intra + 5))))
+    {
+        code_key_frame = 1;
+    }
+    /* in addition if the following are true and this is not a golden frame
+     * then code a key frame Note that on golden frames there often seems
+     * to be a pop in intra useage anyway hence this restriction is
+     * designed to prevent spurious key frames. The Intra pop needs to be
+     * investigated.
+     */
+    else if (((cpi->this_frame_percent_intra > 60) &&
+              (cpi->this_frame_percent_intra > (cpi->last_frame_percent_intra * 2))) ||
+             ((cpi->this_frame_percent_intra > 75) &&
+              (cpi->this_frame_percent_intra > (cpi->last_frame_percent_intra * 3 / 2))) ||
+             ((cpi->this_frame_percent_intra > 90) &&
+              (cpi->this_frame_percent_intra > (cpi->last_frame_percent_intra + 10))))
+    {
+        if (!cm->refresh_golden_frame)
+            code_key_frame = 1;
+    }
+
+    return code_key_frame;
+
+}
+
+static void Pass1Encode(VP8_COMP *cpi, unsigned long *size, unsigned char *dest, unsigned int *frame_flags)
+{
+    (void) size;
+    (void) dest;
+    (void) frame_flags;
+    vp8_set_quantizer(cpi, 26);
+
+    vp8_first_pass(cpi);
+}
+#endif
+
+#if 0
+void write_cx_frame_to_file(YV12_BUFFER_CONFIG *frame, int this_frame)
+{
+
+    /* write the frame */
+    FILE *yframe;
+    int i;
+    char filename[255];
+
+    sprintf(filename, "cx\\y%04d.raw", this_frame);
+    yframe = fopen(filename, "wb");
+
+    for (i = 0; i < frame->y_height; i++)
+        fwrite(frame->y_buffer + i * frame->y_stride, frame->y_width, 1, yframe);
+
+    fclose(yframe);
+    sprintf(filename, "cx\\u%04d.raw", this_frame);
+    yframe = fopen(filename, "wb");
+
+    for (i = 0; i < frame->uv_height; i++)
+        fwrite(frame->u_buffer + i * frame->uv_stride, frame->uv_width, 1, yframe);
+
+    fclose(yframe);
+    sprintf(filename, "cx\\v%04d.raw", this_frame);
+    yframe = fopen(filename, "wb");
+
+    for (i = 0; i < frame->uv_height; i++)
+        fwrite(frame->v_buffer + i * frame->uv_stride, frame->uv_width, 1, yframe);
+
+    fclose(yframe);
+}
+#endif
+/* return of 0 means drop frame */
+
+#if !CONFIG_REALTIME_ONLY
+/* Function to test for conditions that indeicate we should loop
+ * back and recode a frame.
+ */
+static int recode_loop_test( VP8_COMP *cpi,
+                              int high_limit, int low_limit,
+                              int q, int maxq, int minq )
+{
+    int force_recode = 0;
+    VP8_COMMON *cm = &cpi->common;
+
+    /* Is frame recode allowed at all
+     * Yes if either recode mode 1 is selected or mode two is selcted
+     * and the frame is a key frame. golden frame or alt_ref_frame
+     */
+    if ( (cpi->sf.recode_loop == 1) ||
+         ( (cpi->sf.recode_loop == 2) &&
+           ( (cm->frame_type == KEY_FRAME) ||
+             cm->refresh_golden_frame ||
+             cm->refresh_alt_ref_frame ) ) )
+    {
+        /* General over and under shoot tests */
+        if ( ((cpi->projected_frame_size > high_limit) && (q < maxq)) ||
+             ((cpi->projected_frame_size < low_limit) && (q > minq)) )
+        {
+            force_recode = 1;
+        }
+        /* Special Constrained quality tests */
+        else if (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY)
+        {
+            /* Undershoot and below auto cq level */
+            if ( (q > cpi->cq_target_quality) &&
+                 (cpi->projected_frame_size <
+                     ((cpi->this_frame_target * 7) >> 3)))
+            {
+                force_recode = 1;
+            }
+            /* Severe undershoot and between auto and user cq level */
+            else if ( (q > cpi->oxcf.cq_level) &&
+                      (cpi->projected_frame_size < cpi->min_frame_bandwidth) &&
+                      (cpi->active_best_quality > cpi->oxcf.cq_level))
+            {
+                force_recode = 1;
+                cpi->active_best_quality = cpi->oxcf.cq_level;
+            }
+        }
+    }
+
+    return force_recode;
+}
+#endif  // !CONFIG_REALTIME_ONLY
+
+static void update_reference_frames(VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = &cpi->common;
+    YV12_BUFFER_CONFIG *yv12_fb = cm->yv12_fb;
+
+    /* At this point the new frame has been encoded.
+     * If any buffer copy / swapping is signaled it should be done here.
+     */
+
+    if (cm->frame_type == KEY_FRAME)
+    {
+        yv12_fb[cm->new_fb_idx].flags |= VP8_GOLD_FRAME | VP8_ALTR_FRAME ;
+
+        yv12_fb[cm->gld_fb_idx].flags &= ~VP8_GOLD_FRAME;
+        yv12_fb[cm->alt_fb_idx].flags &= ~VP8_ALTR_FRAME;
+
+        cm->alt_fb_idx = cm->gld_fb_idx = cm->new_fb_idx;
+
+        cpi->current_ref_frames[GOLDEN_FRAME] = cm->current_video_frame;
+        cpi->current_ref_frames[ALTREF_FRAME] = cm->current_video_frame;
+    }
+    else    /* For non key frames */
+    {
+        if (cm->refresh_alt_ref_frame)
+        {
+            assert(!cm->copy_buffer_to_arf);
+
+            cm->yv12_fb[cm->new_fb_idx].flags |= VP8_ALTR_FRAME;
+            cm->yv12_fb[cm->alt_fb_idx].flags &= ~VP8_ALTR_FRAME;
+            cm->alt_fb_idx = cm->new_fb_idx;
+
+            cpi->current_ref_frames[ALTREF_FRAME] = cm->current_video_frame;
+        }
+        else if (cm->copy_buffer_to_arf)
+        {
+            assert(!(cm->copy_buffer_to_arf & ~0x3));
+
+            if (cm->copy_buffer_to_arf == 1)
+            {
+                if(cm->alt_fb_idx != cm->lst_fb_idx)
+                {
+                    yv12_fb[cm->lst_fb_idx].flags |= VP8_ALTR_FRAME;
+                    yv12_fb[cm->alt_fb_idx].flags &= ~VP8_ALTR_FRAME;
+                    cm->alt_fb_idx = cm->lst_fb_idx;
+
+                    cpi->current_ref_frames[ALTREF_FRAME] =
+                        cpi->current_ref_frames[LAST_FRAME];
+                }
+            }
+            else /* if (cm->copy_buffer_to_arf == 2) */
+            {
+                if(cm->alt_fb_idx != cm->gld_fb_idx)
+                {
+                    yv12_fb[cm->gld_fb_idx].flags |= VP8_ALTR_FRAME;
+                    yv12_fb[cm->alt_fb_idx].flags &= ~VP8_ALTR_FRAME;
+                    cm->alt_fb_idx = cm->gld_fb_idx;
+
+                    cpi->current_ref_frames[ALTREF_FRAME] =
+                        cpi->current_ref_frames[GOLDEN_FRAME];
+                }
+            }
+        }
+
+        if (cm->refresh_golden_frame)
+        {
+            assert(!cm->copy_buffer_to_gf);
+
+            cm->yv12_fb[cm->new_fb_idx].flags |= VP8_GOLD_FRAME;
+            cm->yv12_fb[cm->gld_fb_idx].flags &= ~VP8_GOLD_FRAME;
+            cm->gld_fb_idx = cm->new_fb_idx;
+
+            cpi->current_ref_frames[GOLDEN_FRAME] = cm->current_video_frame;
+        }
+        else if (cm->copy_buffer_to_gf)
+        {
+            assert(!(cm->copy_buffer_to_arf & ~0x3));
+
+            if (cm->copy_buffer_to_gf == 1)
+            {
+                if(cm->gld_fb_idx != cm->lst_fb_idx)
+                {
+                    yv12_fb[cm->lst_fb_idx].flags |= VP8_GOLD_FRAME;
+                    yv12_fb[cm->gld_fb_idx].flags &= ~VP8_GOLD_FRAME;
+                    cm->gld_fb_idx = cm->lst_fb_idx;
+
+                    cpi->current_ref_frames[GOLDEN_FRAME] =
+                        cpi->current_ref_frames[LAST_FRAME];
+                }
+            }
+            else /* if (cm->copy_buffer_to_gf == 2) */
+            {
+                if(cm->alt_fb_idx != cm->gld_fb_idx)
+                {
+                    yv12_fb[cm->alt_fb_idx].flags |= VP8_GOLD_FRAME;
+                    yv12_fb[cm->gld_fb_idx].flags &= ~VP8_GOLD_FRAME;
+                    cm->gld_fb_idx = cm->alt_fb_idx;
+
+                    cpi->current_ref_frames[GOLDEN_FRAME] =
+                        cpi->current_ref_frames[ALTREF_FRAME];
+                }
+            }
+        }
+    }
+
+    if (cm->refresh_last_frame)
+    {
+        cm->yv12_fb[cm->new_fb_idx].flags |= VP8_LAST_FRAME;
+        cm->yv12_fb[cm->lst_fb_idx].flags &= ~VP8_LAST_FRAME;
+        cm->lst_fb_idx = cm->new_fb_idx;
+
+        cpi->current_ref_frames[LAST_FRAME] = cm->current_video_frame;
+    }
+
+#if CONFIG_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity)
+    {
+        /* we shouldn't have to keep multiple copies as we know in advance which
+         * buffer we should start - for now to get something up and running
+         * I've chosen to copy the buffers
+         */
+        if (cm->frame_type == KEY_FRAME)
+        {
+            int i;
+            for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
+              vp8_yv12_copy_frame(cpi->Source,
+                                  &cpi->denoiser.yv12_running_avg[i]);
+        }
+        else /* For non key frames */
+        {
+            vp8_yv12_extend_frame_borders(
+                    &cpi->denoiser.yv12_running_avg[INTRA_FRAME]);
+
+            if (cm->refresh_alt_ref_frame || cm->copy_buffer_to_arf)
+            {
+                vp8_yv12_copy_frame(
+                        &cpi->denoiser.yv12_running_avg[INTRA_FRAME],
+                        &cpi->denoiser.yv12_running_avg[ALTREF_FRAME]);
+            }
+            if (cm->refresh_golden_frame || cm->copy_buffer_to_gf)
+            {
+                vp8_yv12_copy_frame(
+                        &cpi->denoiser.yv12_running_avg[INTRA_FRAME],
+                        &cpi->denoiser.yv12_running_avg[GOLDEN_FRAME]);
+            }
+            if(cm->refresh_last_frame)
+            {
+                vp8_yv12_copy_frame(
+                        &cpi->denoiser.yv12_running_avg[INTRA_FRAME],
+                        &cpi->denoiser.yv12_running_avg[LAST_FRAME]);
+            }
+        }
+        if (cpi->oxcf.noise_sensitivity == 4)
+          vp8_yv12_copy_frame(cpi->Source, &cpi->denoiser.yv12_last_source);
+
+    }
+#endif
+
+}
+
+static int measure_square_diff_partial(YV12_BUFFER_CONFIG *source,
+                                       YV12_BUFFER_CONFIG *dest,
+                                       VP8_COMP *cpi)
+    {
+        int i, j;
+        int Total = 0;
+        int num_blocks = 0;
+        int skip = 2;
+        int min_consec_zero_last = 10;
+        int tot_num_blocks = (source->y_height * source->y_width) >> 8;
+        unsigned char *src = source->y_buffer;
+        unsigned char *dst = dest->y_buffer;
+
+        /* Loop through the Y plane, every |skip| blocks along rows and colmumns,
+         * summing the square differences, and only for blocks that have been
+         * zero_last mode at least |x| frames in a row.
+         */
+        for (i = 0; i < source->y_height; i += 16 * skip)
+        {
+            int block_index_row = (i >> 4) * cpi->common.mb_cols;
+            for (j = 0; j < source->y_width; j += 16 * skip)
+            {
+                int index = block_index_row + (j >> 4);
+                if (cpi->consec_zero_last[index] >= min_consec_zero_last) {
+                  unsigned int sse;
+                  Total += vpx_mse16x16(src + j,
+                                        source->y_stride,
+                                        dst + j, dest->y_stride,
+                                        &sse);
+                  num_blocks++;
+                }
+            }
+            src += 16 * skip * source->y_stride;
+            dst += 16 * skip * dest->y_stride;
+        }
+        // Only return non-zero if we have at least ~1/16 samples for estimate.
+        if (num_blocks > (tot_num_blocks >> 4)) {
+        return (Total / num_blocks);
+        } else {
+          return 0;
+        }
+    }
+
+#if CONFIG_TEMPORAL_DENOISING
+static void process_denoiser_mode_change(VP8_COMP *cpi) {
+  const VP8_COMMON *const cm = &cpi->common;
+  int i, j;
+  int total = 0;
+  int num_blocks = 0;
+  // Number of blocks skipped along row/column in computing the
+  // nmse (normalized mean square error) of source.
+  int skip = 2;
+  // Only select blocks for computing nmse that have been encoded
+  // as ZERO LAST min_consec_zero_last frames in a row.
+  // Scale with number of temporal layers.
+  int min_consec_zero_last = 12 / cpi->oxcf.number_of_layers;
+  // Decision is tested for changing the denoising mode every
+  // num_mode_change times this function is called. Note that this
+  // function called every 8 frames, so (8 * num_mode_change) is number
+  // of frames where denoising mode change is tested for switch.
+  int num_mode_change = 20;
+  // Framerate factor, to compensate for larger mse at lower framerates.
+  // Use ref_framerate, which is full source framerate for temporal layers.
+  // TODO(marpan): Adjust this factor.
+  int fac_framerate = cpi->ref_framerate < 25.0f ? 80 : 100;
+  int tot_num_blocks = cm->mb_rows * cm->mb_cols;
+  int ystride = cpi->Source->y_stride;
+  unsigned char *src = cpi->Source->y_buffer;
+  unsigned char *dst = cpi->denoiser.yv12_last_source.y_buffer;
+  static const unsigned char const_source[16] = {
+      128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+      128, 128, 128};
+  int bandwidth = (int)(cpi->target_bandwidth);
+  // For temporal layers, use full bandwidth (top layer).
+  if (cpi->oxcf.number_of_layers > 1) {
+    LAYER_CONTEXT *lc = &cpi->layer_context[cpi->oxcf.number_of_layers - 1];
+    bandwidth = (int)(lc->target_bandwidth);
+  }
+  // Loop through the Y plane, every skip blocks along rows and columns,
+  // summing the normalized mean square error, only for blocks that have
+  // been encoded as ZEROMV LAST at least min_consec_zero_last least frames in
+  // a row and have small sum difference between current and previous frame.
+  // Normalization here is by the contrast of the current frame block.
+  for (i = 0; i < cm->Height; i += 16 * skip) {
+    int block_index_row = (i >> 4) * cm->mb_cols;
+    for (j = 0; j < cm->Width; j += 16 * skip) {
+      int index = block_index_row + (j >> 4);
+      if (cpi->consec_zero_last[index] >= min_consec_zero_last) {
+        unsigned int sse;
+        const unsigned int var = vpx_variance16x16(src + j,
+                                                   ystride,
+                                                   dst + j,
+                                                   ystride,
+                                                   &sse);
+        // Only consider this block as valid for noise measurement
+        // if the sum_diff average of the current and previous frame
+        // is small (to avoid effects from lighting change).
+        if ((sse - var) < 128) {
+          unsigned int sse2;
+          const unsigned int act = vpx_variance16x16(src + j,
+                                                     ystride,
+                                                     const_source,
+                                                     0,
+                                                     &sse2);
+          if (act > 0)
+            total += sse / act;
+          num_blocks++;
+        }
+      }
+    }
+    src += 16 * skip * ystride;
+    dst += 16 * skip * ystride;
+  }
+  total = total * fac_framerate / 100;
+
+  // Only consider this frame as valid sample if we have computed nmse over
+  // at least ~1/16 blocks, and Total > 0 (Total == 0 can happen if the
+  // application inputs duplicate frames, or contrast is all zero).
+  if (total > 0 &&
+      (num_blocks > (tot_num_blocks >> 4))) {
+    // Update the recursive mean square source_diff.
+    total = (total << 8) / num_blocks;
+    if (cpi->denoiser.nmse_source_diff_count == 0) {
+      // First sample in new interval.
+      cpi->denoiser.nmse_source_diff = total;
+      cpi->denoiser.qp_avg = cm->base_qindex;
+    } else {
+      // For subsequent samples, use average with weight ~1/4 for new sample.
+      cpi->denoiser.nmse_source_diff = (int)((total +
+          3 * cpi->denoiser.nmse_source_diff) >> 2);
+      cpi->denoiser.qp_avg = (int)((cm->base_qindex +
+          3 * cpi->denoiser.qp_avg) >> 2);
+    }
+    cpi->denoiser.nmse_source_diff_count++;
+  }
+  // Check for changing the denoiser mode, when we have obtained #samples =
+  // num_mode_change. Condition the change also on the bitrate and QP.
+  if (cpi->denoiser.nmse_source_diff_count == num_mode_change) {
+    // Check for going up: from normal to aggressive mode.
+    if ((cpi->denoiser.denoiser_mode == kDenoiserOnYUV) &&
+        (cpi->denoiser.nmse_source_diff >
+        cpi->denoiser.threshold_aggressive_mode) &&
+        (cpi->denoiser.qp_avg < cpi->denoiser.qp_threshold_up &&
+         bandwidth > cpi->denoiser.bitrate_threshold)) {
+      vp8_denoiser_set_parameters(&cpi->denoiser, kDenoiserOnYUVAggressive);
+    } else {
+      // Check for going down: from aggressive to normal mode.
+      if (((cpi->denoiser.denoiser_mode == kDenoiserOnYUVAggressive) &&
+          (cpi->denoiser.nmse_source_diff <
+          cpi->denoiser.threshold_aggressive_mode)) ||
+          ((cpi->denoiser.denoiser_mode == kDenoiserOnYUVAggressive) &&
+          (cpi->denoiser.qp_avg > cpi->denoiser.qp_threshold_down ||
+           bandwidth < cpi->denoiser.bitrate_threshold))) {
+        vp8_denoiser_set_parameters(&cpi->denoiser, kDenoiserOnYUV);
+      }
+    }
+    // Reset metric and counter for next interval.
+    cpi->denoiser.nmse_source_diff = 0;
+    cpi->denoiser.qp_avg = 0;
+    cpi->denoiser.nmse_source_diff_count = 0;
+  }
+}
+#endif
+
+void vp8_loopfilter_frame(VP8_COMP *cpi, VP8_COMMON *cm)
+{
+    const FRAME_TYPE frame_type = cm->frame_type;
+
+    int update_any_ref_buffers = 1;
+    if (cpi->common.refresh_last_frame == 0 &&
+        cpi->common.refresh_golden_frame == 0 &&
+        cpi->common.refresh_alt_ref_frame == 0) {
+        update_any_ref_buffers = 0;
+    }
+
+    if (cm->no_lpf)
+    {
+        cm->filter_level = 0;
+    }
+    else
+    {
+        struct vpx_usec_timer timer;
+
+        vp8_clear_system_state();
+
+        vpx_usec_timer_start(&timer);
+        if (cpi->sf.auto_filter == 0) {
+#if CONFIG_TEMPORAL_DENOISING
+            if (cpi->oxcf.noise_sensitivity && cm->frame_type != KEY_FRAME) {
+                // Use the denoised buffer for selecting base loop filter level.
+                // Denoised signal for current frame is stored in INTRA_FRAME.
+                // No denoising on key frames.
+                vp8cx_pick_filter_level_fast(
+                    &cpi->denoiser.yv12_running_avg[INTRA_FRAME], cpi);
+            } else {
+                vp8cx_pick_filter_level_fast(cpi->Source, cpi);
+            }
+#else
+            vp8cx_pick_filter_level_fast(cpi->Source, cpi);
+#endif
+        } else {
+#if CONFIG_TEMPORAL_DENOISING
+            if (cpi->oxcf.noise_sensitivity && cm->frame_type != KEY_FRAME) {
+                // Use the denoised buffer for selecting base loop filter level.
+                // Denoised signal for current frame is stored in INTRA_FRAME.
+                // No denoising on key frames.
+                vp8cx_pick_filter_level(
+                    &cpi->denoiser.yv12_running_avg[INTRA_FRAME], cpi);
+            } else {
+                vp8cx_pick_filter_level(cpi->Source, cpi);
+            }
+#else
+            vp8cx_pick_filter_level(cpi->Source, cpi);
+#endif
+        }
+
+
+        if (cm->filter_level > 0)
+        {
+            vp8cx_set_alt_lf_level(cpi, cm->filter_level);
+        }
+
+        vpx_usec_timer_mark(&timer);
+        cpi->time_pick_lpf += vpx_usec_timer_elapsed(&timer);
+    }
+
+#if CONFIG_MULTITHREAD
+    if (cpi->b_multi_threaded)
+        sem_post(&cpi->h_event_end_lpf); /* signal that we have set filter_level */
+#endif
+
+    // No need to apply loop-filter if the encoded frame does not update
+    // any reference buffers.
+    if (cm->filter_level > 0 && update_any_ref_buffers)
+    {
+        vp8_loop_filter_frame(cm, &cpi->mb.e_mbd, frame_type);
+    }
+
+    vp8_yv12_extend_frame_borders(cm->frame_to_show);
+
+}
+
+static void encode_frame_to_data_rate
+(
+    VP8_COMP *cpi,
+    unsigned long *size,
+    unsigned char *dest,
+    unsigned char* dest_end,
+    unsigned int *frame_flags
+)
+{
+    int Q;
+    int frame_over_shoot_limit;
+    int frame_under_shoot_limit;
+
+    int Loop = 0;
+    int loop_count;
+
+    VP8_COMMON *cm = &cpi->common;
+    int active_worst_qchanged = 0;
+
+#if !CONFIG_REALTIME_ONLY
+    int q_low;
+    int q_high;
+    int zbin_oq_high;
+    int zbin_oq_low = 0;
+    int top_index;
+    int bottom_index;
+    int overshoot_seen = 0;
+    int undershoot_seen = 0;
+#endif
+
+    int drop_mark = (int)(cpi->oxcf.drop_frames_water_mark *
+                          cpi->oxcf.optimal_buffer_level / 100);
+    int drop_mark75 = drop_mark * 2 / 3;
+    int drop_mark50 = drop_mark / 4;
+    int drop_mark25 = drop_mark / 8;
+
+
+    /* Clear down mmx registers to allow floating point in what follows */
+    vp8_clear_system_state();
+
+    if(cpi->force_next_frame_intra)
+    {
+        cm->frame_type = KEY_FRAME;  /* delayed intra frame */
+        cpi->force_next_frame_intra = 0;
+    }
+
+    /* For an alt ref frame in 2 pass we skip the call to the second pass
+     * function that sets the target bandwidth
+     */
+#if !CONFIG_REALTIME_ONLY
+
+    if (cpi->pass == 2)
+    {
+        if (cpi->common.refresh_alt_ref_frame)
+        {
+            /* Per frame bit target for the alt ref frame */
+            cpi->per_frame_bandwidth = cpi->twopass.gf_bits;
+            /* per second target bitrate */
+            cpi->target_bandwidth = (int)(cpi->twopass.gf_bits *
+                                          cpi->output_framerate);
+        }
+    }
+    else
+#endif
+        cpi->per_frame_bandwidth  = (int)(cpi->target_bandwidth / cpi->output_framerate);
+
+    /* Default turn off buffer to buffer copying */
+    cm->copy_buffer_to_gf = 0;
+    cm->copy_buffer_to_arf = 0;
+
+    /* Clear zbin over-quant value and mode boost values. */
+    cpi->mb.zbin_over_quant = 0;
+    cpi->mb.zbin_mode_boost = 0;
+
+    /* Enable or disable mode based tweaking of the zbin
+     * For 2 Pass Only used where GF/ARF prediction quality
+     * is above a threshold
+     */
+    cpi->mb.zbin_mode_boost_enabled = 1;
+    if (cpi->pass == 2)
+    {
+        if ( cpi->gfu_boost <= 400 )
+        {
+            cpi->mb.zbin_mode_boost_enabled = 0;
+        }
+    }
+
+    /* Current default encoder behaviour for the altref sign bias */
+    if (cpi->source_alt_ref_active)
+        cpi->common.ref_frame_sign_bias[ALTREF_FRAME] = 1;
+    else
+        cpi->common.ref_frame_sign_bias[ALTREF_FRAME] = 0;
+
+    /* Check to see if a key frame is signaled
+     * For two pass with auto key frame enabled cm->frame_type may already
+     * be set, but not for one pass.
+     */
+    if ((cm->current_video_frame == 0) ||
+        (cm->frame_flags & FRAMEFLAGS_KEY) ||
+        (cpi->oxcf.auto_key && (cpi->frames_since_key % cpi->key_frame_frequency == 0)))
+    {
+        /* Key frame from VFW/auto-keyframe/first frame */
+        cm->frame_type = KEY_FRAME;
+#if CONFIG_TEMPORAL_DENOISING
+        if (cpi->oxcf.noise_sensitivity == 4) {
+          // For adaptive mode, reset denoiser to normal mode on key frame.
+          vp8_denoiser_set_parameters(&cpi->denoiser, kDenoiserOnYUV);
+        }
+#endif
+    }
+
+#if CONFIG_MULTI_RES_ENCODING
+    if (cpi->oxcf.mr_total_resolutions > 1) {
+      LOWER_RES_FRAME_INFO* low_res_frame_info
+         = (LOWER_RES_FRAME_INFO*)cpi->oxcf.mr_low_res_mode_info;
+
+      if (cpi->oxcf.mr_encoder_id) {
+
+        // TODO(marpan): This constraint shouldn't be needed, as we would like
+        // to allow for key frame setting (forced or periodic) defined per
+        // spatial layer. For now, keep this in.
+        cm->frame_type = low_res_frame_info->frame_type;
+
+        // Check if lower resolution is available for motion vector reuse.
+        if(cm->frame_type != KEY_FRAME)
+        {
+          cpi->mr_low_res_mv_avail = 1;
+          cpi->mr_low_res_mv_avail &= !(low_res_frame_info->is_frame_dropped);
+
+          if (cpi->ref_frame_flags & VP8_LAST_FRAME)
+              cpi->mr_low_res_mv_avail &= (cpi->current_ref_frames[LAST_FRAME]
+                       == low_res_frame_info->low_res_ref_frames[LAST_FRAME]);
+
+          if (cpi->ref_frame_flags & VP8_GOLD_FRAME)
+              cpi->mr_low_res_mv_avail &= (cpi->current_ref_frames[GOLDEN_FRAME]
+                       == low_res_frame_info->low_res_ref_frames[GOLDEN_FRAME]);
+
+          // Don't use altref to determine whether low res is available.
+          // TODO (marpan): Should we make this type of condition on a
+          // per-reference frame basis?
+          /*
+          if (cpi->ref_frame_flags & VP8_ALTR_FRAME)
+              cpi->mr_low_res_mv_avail &= (cpi->current_ref_frames[ALTREF_FRAME]
+                       == low_res_frame_info->low_res_ref_frames[ALTREF_FRAME]);
+          */
+        }
+      }
+
+      // On a key frame: For the lowest resolution, keep track of the key frame
+      // counter value. For the higher resolutions, reset the current video
+      // frame counter to that of the lowest resolution.
+      // This is done to the handle the case where we may stop/start encoding
+      // higher layer(s). The restart-encoding of higher layer is only signaled
+      // by a key frame for now.
+      // TODO (marpan): Add flag to indicate restart-encoding of higher layer.
+      if (cm->frame_type == KEY_FRAME) {
+        if (cpi->oxcf.mr_encoder_id) {
+          // If the initial starting value of the buffer level is zero (this can
+          // happen because we may have not started encoding this higher stream),
+          // then reset it to non-zero value based on |starting_buffer_level|.
+          if (cpi->common.current_video_frame == 0 && cpi->buffer_level == 0) {
+            unsigned int i;
+            cpi->bits_off_target = cpi->oxcf.starting_buffer_level;
+            cpi->buffer_level = cpi->oxcf.starting_buffer_level;
+            for (i = 0; i < cpi->oxcf.number_of_layers; i++) {
+              LAYER_CONTEXT *lc = &cpi->layer_context[i];
+              lc->bits_off_target = lc->starting_buffer_level;
+              lc->buffer_level = lc->starting_buffer_level;
+            }
+          }
+          cpi->common.current_video_frame =
+              low_res_frame_info->key_frame_counter_value;
+        } else {
+          low_res_frame_info->key_frame_counter_value =
+              cpi->common.current_video_frame;
+        }
+      }
+
+    }
+#endif
+
+    // Find the reference frame closest to the current frame.
+    cpi->closest_reference_frame = LAST_FRAME;
+    if(cm->frame_type != KEY_FRAME) {
+      int i;
+      MV_REFERENCE_FRAME closest_ref = INTRA_FRAME;
+      if (cpi->ref_frame_flags & VP8_LAST_FRAME) {
+        closest_ref = LAST_FRAME;
+      } else if (cpi->ref_frame_flags & VP8_GOLD_FRAME) {
+        closest_ref = GOLDEN_FRAME;
+      } else if (cpi->ref_frame_flags & VP8_ALTR_FRAME) {
+        closest_ref = ALTREF_FRAME;
+      }
+      for(i = 1; i <= 3; i++) {
+        vpx_ref_frame_type_t ref_frame_type = (vpx_ref_frame_type_t)
+            ((i == 3) ? 4 : i);
+        if (cpi->ref_frame_flags & ref_frame_type) {
+          if ((cm->current_video_frame - cpi->current_ref_frames[i]) <
+              (cm->current_video_frame - cpi->current_ref_frames[closest_ref])) {
+            closest_ref = i;
+          }
+        }
+      }
+      cpi->closest_reference_frame = closest_ref;
+    }
+
+    /* Set various flags etc to special state if it is a key frame */
+    if (cm->frame_type == KEY_FRAME)
+    {
+        int i;
+
+        // Set the loop filter deltas and segmentation map update
+        setup_features(cpi);
+
+        /* The alternate reference frame cannot be active for a key frame */
+        cpi->source_alt_ref_active = 0;
+
+        /* Reset the RD threshold multipliers to default of * 1 (128) */
+        for (i = 0; i < MAX_MODES; i++)
+        {
+            cpi->mb.rd_thresh_mult[i] = 128;
+        }
+
+        // Reset the zero_last counter to 0 on key frame.
+        memset(cpi->consec_zero_last, 0, cm->mb_rows * cm->mb_cols);
+        memset(cpi->consec_zero_last_mvbias, 0,
+               (cpi->common.mb_rows * cpi->common.mb_cols));
+    }
+
+#if 0
+    /* Experimental code for lagged compress and one pass
+     * Initialise one_pass GF frames stats
+     * Update stats used for GF selection
+     */
+    {
+        cpi->one_pass_frame_index = cm->current_video_frame % MAX_LAG_BUFFERS;
+
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index ].frames_so_far = 0;
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index ].frame_intra_error = 0.0;
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index ].frame_coded_error = 0.0;
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index ].frame_pcnt_inter = 0.0;
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index ].frame_pcnt_motion = 0.0;
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index ].frame_mvr = 0.0;
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index ].frame_mvr_abs = 0.0;
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index ].frame_mvc = 0.0;
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index ].frame_mvc_abs = 0.0;
+    }
+#endif
+
+    update_rd_ref_frame_probs(cpi);
+
+    if (cpi->drop_frames_allowed)
+    {
+        /* The reset to decimation 0 is only done here for one pass.
+         * Once it is set two pass leaves decimation on till the next kf.
+         */
+        if ((cpi->buffer_level > drop_mark) && (cpi->decimation_factor > 0))
+            cpi->decimation_factor --;
+
+        if (cpi->buffer_level > drop_mark75 && cpi->decimation_factor > 0)
+            cpi->decimation_factor = 1;
+
+        else if (cpi->buffer_level < drop_mark25 && (cpi->decimation_factor == 2 || cpi->decimation_factor == 3))
+        {
+            cpi->decimation_factor = 3;
+        }
+        else if (cpi->buffer_level < drop_mark50 && (cpi->decimation_factor == 1 || cpi->decimation_factor == 2))
+        {
+            cpi->decimation_factor = 2;
+        }
+        else if (cpi->buffer_level < drop_mark75 && (cpi->decimation_factor == 0 || cpi->decimation_factor == 1))
+        {
+            cpi->decimation_factor = 1;
+        }
+    }
+
+    /* The following decimates the frame rate according to a regular
+     * pattern (i.e. to 1/2 or 2/3 frame rate) This can be used to help
+     * prevent buffer under-run in CBR mode. Alternatively it might be
+     * desirable in some situations to drop frame rate but throw more bits
+     * at each frame.
+     *
+     * Note that dropping a key frame can be problematic if spatial
+     * resampling is also active
+     */
+    if (cpi->decimation_factor > 0)
+    {
+        switch (cpi->decimation_factor)
+        {
+        case 1:
+            cpi->per_frame_bandwidth  = cpi->per_frame_bandwidth * 3 / 2;
+            break;
+        case 2:
+            cpi->per_frame_bandwidth  = cpi->per_frame_bandwidth * 5 / 4;
+            break;
+        case 3:
+            cpi->per_frame_bandwidth  = cpi->per_frame_bandwidth * 5 / 4;
+            break;
+        }
+
+        /* Note that we should not throw out a key frame (especially when
+         * spatial resampling is enabled).
+         */
+        if (cm->frame_type == KEY_FRAME)
+        {
+            cpi->decimation_count = cpi->decimation_factor;
+        }
+        else if (cpi->decimation_count > 0)
+        {
+            cpi->decimation_count --;
+
+            cpi->bits_off_target += cpi->av_per_frame_bandwidth;
+            if (cpi->bits_off_target > cpi->oxcf.maximum_buffer_size)
+                cpi->bits_off_target = cpi->oxcf.maximum_buffer_size;
+
+#if CONFIG_MULTI_RES_ENCODING
+            vp8_store_drop_frame_info(cpi);
+#endif
+
+            cm->current_video_frame++;
+            cpi->frames_since_key++;
+            // We advance the temporal pattern for dropped frames.
+            cpi->temporal_pattern_counter++;
+
+#if CONFIG_INTERNAL_STATS
+            cpi->count ++;
+#endif
+
+            cpi->buffer_level = cpi->bits_off_target;
+
+            if (cpi->oxcf.number_of_layers > 1)
+            {
+                unsigned int i;
+
+                /* Propagate bits saved by dropping the frame to higher
+                 * layers
+                 */
+                for (i=cpi->current_layer+1; i<cpi->oxcf.number_of_layers; i++)
+                {
+                    LAYER_CONTEXT *lc = &cpi->layer_context[i];
+                    lc->bits_off_target += (int)(lc->target_bandwidth /
+                                                 lc->framerate);
+                    if (lc->bits_off_target > lc->maximum_buffer_size)
+                        lc->bits_off_target = lc->maximum_buffer_size;
+                    lc->buffer_level = lc->bits_off_target;
+                }
+            }
+
+            return;
+        }
+        else
+            cpi->decimation_count = cpi->decimation_factor;
+    }
+    else
+        cpi->decimation_count = 0;
+
+    /* Decide how big to make the frame */
+    if (!vp8_pick_frame_size(cpi))
+    {
+        /*TODO: 2 drop_frame and return code could be put together. */
+#if CONFIG_MULTI_RES_ENCODING
+        vp8_store_drop_frame_info(cpi);
+#endif
+        cm->current_video_frame++;
+        cpi->frames_since_key++;
+        // We advance the temporal pattern for dropped frames.
+        cpi->temporal_pattern_counter++;
+        return;
+    }
+
+    /* Reduce active_worst_allowed_q for CBR if our buffer is getting too full.
+     * This has a knock on effect on active best quality as well.
+     * For CBR if the buffer reaches its maximum level then we can no longer
+     * save up bits for later frames so we might as well use them up
+     * on the current frame.
+     */
+    if ((cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) &&
+        (cpi->buffer_level >= cpi->oxcf.optimal_buffer_level) && cpi->buffered_mode)
+    {
+        /* Max adjustment is 1/4 */
+        int Adjustment = cpi->active_worst_quality / 4;
+
+        if (Adjustment)
+        {
+            int buff_lvl_step;
+
+            if (cpi->buffer_level < cpi->oxcf.maximum_buffer_size)
+            {
+                buff_lvl_step = (int)
+                                ((cpi->oxcf.maximum_buffer_size -
+                                  cpi->oxcf.optimal_buffer_level) /
+                                  Adjustment);
+
+                if (buff_lvl_step)
+                    Adjustment = (int)
+                                 ((cpi->buffer_level -
+                                 cpi->oxcf.optimal_buffer_level) /
+                                 buff_lvl_step);
+                else
+                    Adjustment = 0;
+            }
+
+            cpi->active_worst_quality -= Adjustment;
+
+            if(cpi->active_worst_quality < cpi->active_best_quality)
+                cpi->active_worst_quality = cpi->active_best_quality;
+        }
+    }
+
+    /* Set an active best quality and if necessary active worst quality
+     * There is some odd behavior for one pass here that needs attention.
+     */
+    if ( (cpi->pass == 2) || (cpi->ni_frames > 150))
+    {
+        vp8_clear_system_state();
+
+        Q = cpi->active_worst_quality;
+
+        if ( cm->frame_type == KEY_FRAME )
+        {
+            if ( cpi->pass == 2 )
+            {
+                if (cpi->gfu_boost > 600)
+                   cpi->active_best_quality = kf_low_motion_minq[Q];
+                else
+                   cpi->active_best_quality = kf_high_motion_minq[Q];
+
+                /* Special case for key frames forced because we have reached
+                 * the maximum key frame interval. Here force the Q to a range
+                 * based on the ambient Q to reduce the risk of popping
+                 */
+                if ( cpi->this_key_frame_forced )
+                {
+                    if ( cpi->active_best_quality > cpi->avg_frame_qindex * 7/8)
+                        cpi->active_best_quality = cpi->avg_frame_qindex * 7/8;
+                    else if ( cpi->active_best_quality < cpi->avg_frame_qindex >> 2 )
+                        cpi->active_best_quality = cpi->avg_frame_qindex >> 2;
+                }
+            }
+            /* One pass more conservative */
+            else
+               cpi->active_best_quality = kf_high_motion_minq[Q];
+        }
+
+        else if (cpi->oxcf.number_of_layers==1 &&
+                (cm->refresh_golden_frame || cpi->common.refresh_alt_ref_frame))
+        {
+            /* Use the lower of cpi->active_worst_quality and recent
+             * average Q as basis for GF/ARF Q limit unless last frame was
+             * a key frame.
+             */
+            if ( (cpi->frames_since_key > 1) &&
+               (cpi->avg_frame_qindex < cpi->active_worst_quality) )
+            {
+                Q = cpi->avg_frame_qindex;
+            }
+
+            /* For constrained quality dont allow Q less than the cq level */
+            if ( (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY) &&
+                 (Q < cpi->cq_target_quality) )
+            {
+                Q = cpi->cq_target_quality;
+            }
+
+            if ( cpi->pass == 2 )
+            {
+                if ( cpi->gfu_boost > 1000 )
+                    cpi->active_best_quality = gf_low_motion_minq[Q];
+                else if ( cpi->gfu_boost < 400 )
+                    cpi->active_best_quality = gf_high_motion_minq[Q];
+                else
+                    cpi->active_best_quality = gf_mid_motion_minq[Q];
+
+                /* Constrained quality use slightly lower active best. */
+                if ( cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY )
+                {
+                    cpi->active_best_quality =
+                        cpi->active_best_quality * 15/16;
+                }
+            }
+            /* One pass more conservative */
+            else
+                cpi->active_best_quality = gf_high_motion_minq[Q];
+        }
+        else
+        {
+            cpi->active_best_quality = inter_minq[Q];
+
+            /* For the constant/constrained quality mode we dont want
+             * q to fall below the cq level.
+             */
+            if ((cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY) &&
+                (cpi->active_best_quality < cpi->cq_target_quality) )
+            {
+                /* If we are strongly undershooting the target rate in the last
+                 * frames then use the user passed in cq value not the auto
+                 * cq value.
+                 */
+                if ( cpi->rolling_actual_bits < cpi->min_frame_bandwidth )
+                    cpi->active_best_quality = cpi->oxcf.cq_level;
+                else
+                    cpi->active_best_quality = cpi->cq_target_quality;
+            }
+        }
+
+        /* If CBR and the buffer is as full then it is reasonable to allow
+         * higher quality on the frames to prevent bits just going to waste.
+         */
+        if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+        {
+            /* Note that the use of >= here elliminates the risk of a devide
+             * by 0 error in the else if clause
+             */
+            if (cpi->buffer_level >= cpi->oxcf.maximum_buffer_size)
+                cpi->active_best_quality = cpi->best_quality;
+
+            else if (cpi->buffer_level > cpi->oxcf.optimal_buffer_level)
+            {
+                int Fraction = (int)
+                  (((cpi->buffer_level - cpi->oxcf.optimal_buffer_level) * 128)
+                  / (cpi->oxcf.maximum_buffer_size -
+                  cpi->oxcf.optimal_buffer_level));
+                int min_qadjustment = ((cpi->active_best_quality -
+                                        cpi->best_quality) * Fraction) / 128;
+
+                cpi->active_best_quality -= min_qadjustment;
+            }
+        }
+    }
+    /* Make sure constrained quality mode limits are adhered to for the first
+     * few frames of one pass encodes
+     */
+    else if (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY)
+    {
+        if ( (cm->frame_type == KEY_FRAME) ||
+             cm->refresh_golden_frame || cpi->common.refresh_alt_ref_frame )
+        {
+             cpi->active_best_quality = cpi->best_quality;
+        }
+        else if (cpi->active_best_quality < cpi->cq_target_quality)
+        {
+            cpi->active_best_quality = cpi->cq_target_quality;
+        }
+    }
+
+    /* Clip the active best and worst quality values to limits */
+    if (cpi->active_worst_quality > cpi->worst_quality)
+        cpi->active_worst_quality = cpi->worst_quality;
+
+    if (cpi->active_best_quality < cpi->best_quality)
+        cpi->active_best_quality = cpi->best_quality;
+
+    if ( cpi->active_worst_quality < cpi->active_best_quality )
+        cpi->active_worst_quality = cpi->active_best_quality;
+
+    /* Determine initial Q to try */
+    Q = vp8_regulate_q(cpi, cpi->this_frame_target);
+
+#if !CONFIG_REALTIME_ONLY
+
+    /* Set highest allowed value for Zbin over quant */
+    if (cm->frame_type == KEY_FRAME)
+        zbin_oq_high = 0;
+    else if ((cpi->oxcf.number_of_layers == 1) && ((cm->refresh_alt_ref_frame ||
+              (cm->refresh_golden_frame && !cpi->source_alt_ref_active))))
+    {
+          zbin_oq_high = 16;
+    }
+    else
+        zbin_oq_high = ZBIN_OQ_MAX;
+#endif
+
+    /* Setup background Q adjustment for error resilient mode.
+     * For multi-layer encodes only enable this for the base layer.
+    */
+    if (cpi->cyclic_refresh_mode_enabled)
+    {
+      // Special case for screen_content_mode with golden frame updates.
+      int disable_cr_gf = (cpi->oxcf.screen_content_mode == 2 &&
+                           cm->refresh_golden_frame);
+      if (cpi->current_layer == 0 && cpi->force_maxqp == 0 && !disable_cr_gf)
+        cyclic_background_refresh(cpi, Q, 0);
+      else
+        disable_segmentation(cpi);
+    }
+
+    vp8_compute_frame_size_bounds(cpi, &frame_under_shoot_limit, &frame_over_shoot_limit);
+
+#if !CONFIG_REALTIME_ONLY
+    /* Limit Q range for the adaptive loop. */
+    bottom_index = cpi->active_best_quality;
+    top_index    = cpi->active_worst_quality;
+    q_low  = cpi->active_best_quality;
+    q_high = cpi->active_worst_quality;
+#endif
+
+    vp8_save_coding_context(cpi);
+
+    loop_count = 0;
+
+    scale_and_extend_source(cpi->un_scaled_source, cpi);
+
+#if CONFIG_TEMPORAL_DENOISING && CONFIG_POSTPROC
+    // Option to apply spatial blur under the aggressive or adaptive
+    // (temporal denoising) mode.
+    if (cpi->oxcf.noise_sensitivity >= 3) {
+      if (cpi->denoiser.denoise_pars.spatial_blur != 0) {
+        vp8_de_noise(cm, cpi->Source, cpi->Source,
+            cpi->denoiser.denoise_pars.spatial_blur, 1, 0, 0);
+      }
+    }
+#endif
+
+#if !(CONFIG_REALTIME_ONLY) && CONFIG_POSTPROC && !(CONFIG_TEMPORAL_DENOISING)
+
+    if (cpi->oxcf.noise_sensitivity > 0)
+    {
+        unsigned char *src;
+        int l = 0;
+
+        switch (cpi->oxcf.noise_sensitivity)
+        {
+        case 1:
+            l = 20;
+            break;
+        case 2:
+            l = 40;
+            break;
+        case 3:
+            l = 60;
+            break;
+        case 4:
+            l = 80;
+            break;
+        case 5:
+            l = 100;
+            break;
+        case 6:
+            l = 150;
+            break;
+        }
+
+
+        if (cm->frame_type == KEY_FRAME)
+        {
+            vp8_de_noise(cm, cpi->Source, cpi->Source, l , 1,  0, 1);
+        }
+        else
+        {
+            vp8_de_noise(cm, cpi->Source, cpi->Source, l , 1,  0, 1);
+
+            src = cpi->Source->y_buffer;
+
+            if (cpi->Source->y_stride < 0)
+            {
+                src += cpi->Source->y_stride * (cpi->Source->y_height - 1);
+            }
+        }
+    }
+
+#endif
+
+
+#ifdef OUTPUT_YUV_SRC
+    vp8_write_yuv_frame(yuv_file, cpi->Source);
+#endif
+
+    do
+    {
+        vp8_clear_system_state();
+
+        vp8_set_quantizer(cpi, Q);
+
+        /* setup skip prob for costing in mode/mv decision */
+        if (cpi->common.mb_no_coeff_skip)
+        {
+            cpi->prob_skip_false = cpi->base_skip_false_prob[Q];
+
+            if (cm->frame_type != KEY_FRAME)
+            {
+                if (cpi->common.refresh_alt_ref_frame)
+                {
+                    if (cpi->last_skip_false_probs[2] != 0)
+                        cpi->prob_skip_false = cpi->last_skip_false_probs[2];
+
+                    /*
+                                        if(cpi->last_skip_false_probs[2]!=0 && abs(Q- cpi->last_skip_probs_q[2])<=16 )
+                       cpi->prob_skip_false = cpi->last_skip_false_probs[2];
+                                        else if (cpi->last_skip_false_probs[2]!=0)
+                       cpi->prob_skip_false = (cpi->last_skip_false_probs[2]  + cpi->prob_skip_false ) / 2;
+                       */
+                }
+                else if (cpi->common.refresh_golden_frame)
+                {
+                    if (cpi->last_skip_false_probs[1] != 0)
+                        cpi->prob_skip_false = cpi->last_skip_false_probs[1];
+
+                    /*
+                                        if(cpi->last_skip_false_probs[1]!=0 && abs(Q- cpi->last_skip_probs_q[1])<=16 )
+                       cpi->prob_skip_false = cpi->last_skip_false_probs[1];
+                                        else if (cpi->last_skip_false_probs[1]!=0)
+                       cpi->prob_skip_false = (cpi->last_skip_false_probs[1]  + cpi->prob_skip_false ) / 2;
+                       */
+                }
+                else
+                {
+                    if (cpi->last_skip_false_probs[0] != 0)
+                        cpi->prob_skip_false = cpi->last_skip_false_probs[0];
+
+                    /*
+                    if(cpi->last_skip_false_probs[0]!=0 && abs(Q- cpi->last_skip_probs_q[0])<=16 )
+                        cpi->prob_skip_false = cpi->last_skip_false_probs[0];
+                    else if(cpi->last_skip_false_probs[0]!=0)
+                        cpi->prob_skip_false = (cpi->last_skip_false_probs[0]  + cpi->prob_skip_false ) / 2;
+                        */
+                }
+
+                /* as this is for cost estimate, let's make sure it does not
+                 * go extreme eitehr way
+                 */
+                if (cpi->prob_skip_false < 5)
+                    cpi->prob_skip_false = 5;
+
+                if (cpi->prob_skip_false > 250)
+                    cpi->prob_skip_false = 250;
+
+                if (cpi->oxcf.number_of_layers == 1 && cpi->is_src_frame_alt_ref)
+                    cpi->prob_skip_false = 1;
+            }
+
+#if 0
+
+            if (cpi->pass != 1)
+            {
+                FILE *f = fopen("skip.stt", "a");
+                fprintf(f, "%d, %d, %4d ", cpi->common.refresh_golden_frame, cpi->common.refresh_alt_ref_frame, cpi->prob_skip_false);
+                fclose(f);
+            }
+
+#endif
+
+        }
+
+        if (cm->frame_type == KEY_FRAME)
+        {
+            if(resize_key_frame(cpi))
+            {
+              /* If the frame size has changed, need to reset Q, quantizer,
+               * and background refresh.
+               */
+              Q = vp8_regulate_q(cpi, cpi->this_frame_target);
+              if (cpi->cyclic_refresh_mode_enabled)
+              {
+                if (cpi->current_layer==0)
+                  cyclic_background_refresh(cpi, Q, 0);
+                else
+                  disable_segmentation(cpi);
+              }
+              // Reset the zero_last counter to 0 on key frame.
+              memset(cpi->consec_zero_last, 0, cm->mb_rows * cm->mb_cols);
+              memset(cpi->consec_zero_last_mvbias, 0,
+                     (cpi->common.mb_rows * cpi->common.mb_cols));
+              vp8_set_quantizer(cpi, Q);
+            }
+
+            vp8_setup_key_frame(cpi);
+        }
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+        {
+            if(cpi->oxcf.error_resilient_mode)
+                cm->refresh_entropy_probs = 0;
+
+            if (cpi->oxcf.error_resilient_mode & VPX_ERROR_RESILIENT_PARTITIONS)
+            {
+                if (cm->frame_type == KEY_FRAME)
+                    cm->refresh_entropy_probs = 1;
+            }
+
+            if (cm->refresh_entropy_probs == 0)
+            {
+                /* save a copy for later refresh */
+                memcpy(&cm->lfc, &cm->fc, sizeof(cm->fc));
+            }
+
+            vp8_update_coef_context(cpi);
+
+            vp8_update_coef_probs(cpi);
+
+            /* transform / motion compensation build reconstruction frame
+             * +pack coef partitions
+             */
+            vp8_encode_frame(cpi);
+
+            /* cpi->projected_frame_size is not needed for RT mode */
+        }
+#else
+        /* transform / motion compensation build reconstruction frame */
+        vp8_encode_frame(cpi);
+
+        if (cpi->oxcf.screen_content_mode == 2) {
+          if (vp8_drop_encodedframe_overshoot(cpi, Q))
+            return;
+        }
+
+        cpi->projected_frame_size -= vp8_estimate_entropy_savings(cpi);
+        cpi->projected_frame_size = (cpi->projected_frame_size > 0) ? cpi->projected_frame_size : 0;
+#endif
+        vp8_clear_system_state();
+
+        /* Test to see if the stats generated for this frame indicate that
+         * we should have coded a key frame (assuming that we didn't)!
+         */
+
+        if (cpi->pass != 2 && cpi->oxcf.auto_key && cm->frame_type != KEY_FRAME
+            && cpi->compressor_speed != 2)
+        {
+#if !CONFIG_REALTIME_ONLY
+            if (decide_key_frame(cpi))
+            {
+                /* Reset all our sizing numbers and recode */
+                cm->frame_type = KEY_FRAME;
+
+                vp8_pick_frame_size(cpi);
+
+                /* Clear the Alt reference frame active flag when we have
+                 * a key frame
+                 */
+                cpi->source_alt_ref_active = 0;
+
+                // Set the loop filter deltas and segmentation map update
+                setup_features(cpi);
+
+                vp8_restore_coding_context(cpi);
+
+                Q = vp8_regulate_q(cpi, cpi->this_frame_target);
+
+                vp8_compute_frame_size_bounds(cpi, &frame_under_shoot_limit, &frame_over_shoot_limit);
+
+                /* Limit Q range for the adaptive loop. */
+                bottom_index = cpi->active_best_quality;
+                top_index    = cpi->active_worst_quality;
+                q_low  = cpi->active_best_quality;
+                q_high = cpi->active_worst_quality;
+
+                loop_count++;
+                Loop = 1;
+
+                continue;
+            }
+#endif
+        }
+
+        vp8_clear_system_state();
+
+        if (frame_over_shoot_limit == 0)
+            frame_over_shoot_limit = 1;
+
+        /* Are we are overshooting and up against the limit of active max Q. */
+        if (((cpi->pass != 2) || (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)) &&
+            (Q == cpi->active_worst_quality)                     &&
+            (cpi->active_worst_quality < cpi->worst_quality)      &&
+            (cpi->projected_frame_size > frame_over_shoot_limit))
+        {
+            int over_size_percent = ((cpi->projected_frame_size - frame_over_shoot_limit) * 100) / frame_over_shoot_limit;
+
+            /* If so is there any scope for relaxing it */
+            while ((cpi->active_worst_quality < cpi->worst_quality) && (over_size_percent > 0))
+            {
+                cpi->active_worst_quality++;
+                /* Assume 1 qstep = about 4% on frame size. */
+                over_size_percent = (int)(over_size_percent * 0.96);
+            }
+#if !CONFIG_REALTIME_ONLY
+            top_index = cpi->active_worst_quality;
+#endif  // !CONFIG_REALTIME_ONLY
+            /* If we have updated the active max Q do not call
+             * vp8_update_rate_correction_factors() this loop.
+             */
+            active_worst_qchanged = 1;
+        }
+        else
+            active_worst_qchanged = 0;
+
+#if !CONFIG_REALTIME_ONLY
+        /* Special case handling for forced key frames */
+        if ( (cm->frame_type == KEY_FRAME) && cpi->this_key_frame_forced )
+        {
+            int last_q = Q;
+            int kf_err = vp8_calc_ss_err(cpi->Source,
+                                         &cm->yv12_fb[cm->new_fb_idx]);
+
+            /* The key frame is not good enough */
+            if ( kf_err > ((cpi->ambient_err * 7) >> 3) )
+            {
+                /* Lower q_high */
+                q_high = (Q > q_low) ? (Q - 1) : q_low;
+
+                /* Adjust Q */
+                Q = (q_high + q_low) >> 1;
+            }
+            /* The key frame is much better than the previous frame */
+            else if ( kf_err < (cpi->ambient_err >> 1) )
+            {
+                /* Raise q_low */
+                q_low = (Q < q_high) ? (Q + 1) : q_high;
+
+                /* Adjust Q */
+                Q = (q_high + q_low + 1) >> 1;
+            }
+
+            /* Clamp Q to upper and lower limits: */
+            if (Q > q_high)
+                Q = q_high;
+            else if (Q < q_low)
+                Q = q_low;
+
+            Loop = Q != last_q;
+        }
+
+        /* Is the projected frame size out of range and are we allowed
+         * to attempt to recode.
+         */
+        else if ( recode_loop_test( cpi,
+                               frame_over_shoot_limit, frame_under_shoot_limit,
+                               Q, top_index, bottom_index ) )
+        {
+            int last_q = Q;
+            int Retries = 0;
+
+            /* Frame size out of permitted range. Update correction factor
+             * & compute new Q to try...
+             */
+
+            /* Frame is too large */
+            if (cpi->projected_frame_size > cpi->this_frame_target)
+            {
+                /* Raise Qlow as to at least the current value */
+                q_low = (Q < q_high) ? (Q + 1) : q_high;
+
+                /* If we are using over quant do the same for zbin_oq_low */
+                if (cpi->mb.zbin_over_quant > 0)
+                    zbin_oq_low = (cpi->mb.zbin_over_quant < zbin_oq_high) ?
+                        (cpi->mb.zbin_over_quant + 1) : zbin_oq_high;
+
+                if (undershoot_seen)
+                {
+                    /* Update rate_correction_factor unless
+                     * cpi->active_worst_quality has changed.
+                     */
+                    if (!active_worst_qchanged)
+                        vp8_update_rate_correction_factors(cpi, 1);
+
+                    Q = (q_high + q_low + 1) / 2;
+
+                    /* Adjust cpi->zbin_over_quant (only allowed when Q
+                     * is max)
+                     */
+                    if (Q < MAXQ)
+                        cpi->mb.zbin_over_quant = 0;
+                    else
+                    {
+                        zbin_oq_low = (cpi->mb.zbin_over_quant < zbin_oq_high) ?
+                            (cpi->mb.zbin_over_quant + 1) : zbin_oq_high;
+                        cpi->mb.zbin_over_quant =
+                            (zbin_oq_high + zbin_oq_low) / 2;
+                    }
+                }
+                else
+                {
+                    /* Update rate_correction_factor unless
+                     * cpi->active_worst_quality has changed.
+                     */
+                    if (!active_worst_qchanged)
+                        vp8_update_rate_correction_factors(cpi, 0);
+
+                    Q = vp8_regulate_q(cpi, cpi->this_frame_target);
+
+                    while (((Q < q_low) ||
+                        (cpi->mb.zbin_over_quant < zbin_oq_low)) &&
+                        (Retries < 10))
+                    {
+                        vp8_update_rate_correction_factors(cpi, 0);
+                        Q = vp8_regulate_q(cpi, cpi->this_frame_target);
+                        Retries ++;
+                    }
+                }
+
+                overshoot_seen = 1;
+            }
+            /* Frame is too small */
+            else
+            {
+                if (cpi->mb.zbin_over_quant == 0)
+                    /* Lower q_high if not using over quant */
+                    q_high = (Q > q_low) ? (Q - 1) : q_low;
+                else
+                    /* else lower zbin_oq_high */
+                    zbin_oq_high = (cpi->mb.zbin_over_quant > zbin_oq_low) ?
+                        (cpi->mb.zbin_over_quant - 1) : zbin_oq_low;
+
+                if (overshoot_seen)
+                {
+                    /* Update rate_correction_factor unless
+                     * cpi->active_worst_quality has changed.
+                     */
+                    if (!active_worst_qchanged)
+                        vp8_update_rate_correction_factors(cpi, 1);
+
+                    Q = (q_high + q_low) / 2;
+
+                    /* Adjust cpi->zbin_over_quant (only allowed when Q
+                     * is max)
+                     */
+                    if (Q < MAXQ)
+                        cpi->mb.zbin_over_quant = 0;
+                    else
+                        cpi->mb.zbin_over_quant =
+                            (zbin_oq_high + zbin_oq_low) / 2;
+                }
+                else
+                {
+                    /* Update rate_correction_factor unless
+                     * cpi->active_worst_quality has changed.
+                     */
+                    if (!active_worst_qchanged)
+                        vp8_update_rate_correction_factors(cpi, 0);
+
+                    Q = vp8_regulate_q(cpi, cpi->this_frame_target);
+
+                    /* Special case reset for qlow for constrained quality.
+                     * This should only trigger where there is very substantial
+                     * undershoot on a frame and the auto cq level is above
+                     * the user passsed in value.
+                     */
+                    if ( (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY) &&
+                         (Q < q_low) )
+                    {
+                        q_low = Q;
+                    }
+
+                    while (((Q > q_high) ||
+                        (cpi->mb.zbin_over_quant > zbin_oq_high)) &&
+                        (Retries < 10))
+                    {
+                        vp8_update_rate_correction_factors(cpi, 0);
+                        Q = vp8_regulate_q(cpi, cpi->this_frame_target);
+                        Retries ++;
+                    }
+                }
+
+                undershoot_seen = 1;
+            }
+
+            /* Clamp Q to upper and lower limits: */
+            if (Q > q_high)
+                Q = q_high;
+            else if (Q < q_low)
+                Q = q_low;
+
+            /* Clamp cpi->zbin_over_quant */
+            cpi->mb.zbin_over_quant = (cpi->mb.zbin_over_quant < zbin_oq_low) ?
+                zbin_oq_low : (cpi->mb.zbin_over_quant > zbin_oq_high) ?
+                    zbin_oq_high : cpi->mb.zbin_over_quant;
+
+            Loop = Q != last_q;
+        }
+        else
+#endif
+            Loop = 0;
+
+        if (cpi->is_src_frame_alt_ref)
+            Loop = 0;
+
+        if (Loop == 1)
+        {
+            vp8_restore_coding_context(cpi);
+            loop_count++;
+#if CONFIG_INTERNAL_STATS
+            cpi->tot_recode_hits++;
+#endif
+        }
+    }
+    while (Loop == 1);
+
+#if 0
+    /* Experimental code for lagged and one pass
+     * Update stats used for one pass GF selection
+     */
+    {
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index].frame_coded_error = (double)cpi->prediction_error;
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index].frame_intra_error = (double)cpi->intra_error;
+        cpi->one_pass_frame_stats[cpi->one_pass_frame_index].frame_pcnt_inter = (double)(100 - cpi->this_frame_percent_intra) / 100.0;
+    }
+#endif
+
+    /* Special case code to reduce pulsing when key frames are forced at a
+     * fixed interval. Note the reconstruction error if it is the frame before
+     * the force key frame
+     */
+    if ( cpi->next_key_frame_forced && (cpi->twopass.frames_to_key == 0) )
+    {
+        cpi->ambient_err = vp8_calc_ss_err(cpi->Source,
+                                           &cm->yv12_fb[cm->new_fb_idx]);
+    }
+
+    /* This frame's MVs are saved and will be used in next frame's MV predictor.
+     * Last frame has one more line(add to bottom) and one more column(add to
+     * right) than cm->mip. The edge elements are initialized to 0.
+     */
+#if CONFIG_MULTI_RES_ENCODING
+    if(!cpi->oxcf.mr_encoder_id && cm->show_frame)
+#else
+    if(cm->show_frame)   /* do not save for altref frame */
+#endif
+    {
+        int mb_row;
+        int mb_col;
+        /* Point to beginning of allocated MODE_INFO arrays. */
+        MODE_INFO *tmp = cm->mip;
+
+        if(cm->frame_type != KEY_FRAME)
+        {
+            for (mb_row = 0; mb_row < cm->mb_rows+1; mb_row ++)
+            {
+                for (mb_col = 0; mb_col < cm->mb_cols+1; mb_col ++)
+                {
+                    if(tmp->mbmi.ref_frame != INTRA_FRAME)
+                        cpi->lfmv[mb_col + mb_row*(cm->mode_info_stride+1)].as_int = tmp->mbmi.mv.as_int;
+
+                    cpi->lf_ref_frame_sign_bias[mb_col + mb_row*(cm->mode_info_stride+1)] = cm->ref_frame_sign_bias[tmp->mbmi.ref_frame];
+                    cpi->lf_ref_frame[mb_col + mb_row*(cm->mode_info_stride+1)] = tmp->mbmi.ref_frame;
+                    tmp++;
+                }
+            }
+        }
+    }
+
+    /* Count last ref frame 0,0 usage on current encoded frame. */
+    {
+        int mb_row;
+        int mb_col;
+        /* Point to beginning of MODE_INFO arrays. */
+        MODE_INFO *tmp = cm->mi;
+
+        cpi->zeromv_count = 0;
+
+        if(cm->frame_type != KEY_FRAME)
+        {
+            for (mb_row = 0; mb_row < cm->mb_rows; mb_row ++)
+            {
+                for (mb_col = 0; mb_col < cm->mb_cols; mb_col ++)
+                {
+                    if (tmp->mbmi.mode == ZEROMV &&
+                       tmp->mbmi.ref_frame == LAST_FRAME)
+                        cpi->zeromv_count++;
+                    tmp++;
+                }
+                tmp++;
+            }
+        }
+    }
+
+#if CONFIG_MULTI_RES_ENCODING
+    vp8_cal_dissimilarity(cpi);
+#endif
+
+    /* Update the GF useage maps.
+     * This is done after completing the compression of a frame when all
+     * modes etc. are finalized but before loop filter
+     */
+    if (cpi->oxcf.number_of_layers == 1)
+        vp8_update_gf_useage_maps(cpi, cm, &cpi->mb);
+
+    if (cm->frame_type == KEY_FRAME)
+        cm->refresh_last_frame = 1;
+
+#if 0
+    {
+        FILE *f = fopen("gfactive.stt", "a");
+        fprintf(f, "%8d %8d %8d %8d %8d\n", cm->current_video_frame, (100 * cpi->gf_active_count) / (cpi->common.mb_rows * cpi->common.mb_cols), cpi->this_iiratio, cpi->next_iiratio, cm->refresh_golden_frame);
+        fclose(f);
+    }
+#endif
+
+    /* For inter frames the current default behavior is that when
+     * cm->refresh_golden_frame is set we copy the old GF over to the ARF buffer
+     * This is purely an encoder decision at present.
+     */
+    if (!cpi->oxcf.error_resilient_mode && cm->refresh_golden_frame)
+        cm->copy_buffer_to_arf  = 2;
+    else
+        cm->copy_buffer_to_arf  = 0;
+
+    cm->frame_to_show = &cm->yv12_fb[cm->new_fb_idx];
+
+#if CONFIG_TEMPORAL_DENOISING
+    // Get some measure of the amount of noise, by measuring the (partial) mse
+    // between source and denoised buffer, for y channel. Partial refers to
+    // computing the sse for a sub-sample of the frame (i.e., skip x blocks along row/column),
+    // and only for blocks in that set that are consecutive ZEROMV_LAST mode.
+    // Do this every ~8 frames, to further reduce complexity.
+    // TODO(marpan): Keep this for now for the case cpi->oxcf.noise_sensitivity < 4,
+    // should be removed in favor of the process_denoiser_mode_change() function below.
+    if (cpi->oxcf.noise_sensitivity > 0 &&
+       cpi->oxcf.noise_sensitivity < 4 &&
+       !cpi->oxcf.screen_content_mode &&
+       cpi->frames_since_key%8 == 0 &&
+       cm->frame_type != KEY_FRAME) {
+       cpi->mse_source_denoised = measure_square_diff_partial(
+           &cpi->denoiser.yv12_running_avg[INTRA_FRAME], cpi->Source, cpi);
+    }
+
+    // For the adaptive denoising mode (noise_sensitivity == 4), sample the mse
+    // of source diff (between current and previous frame), and determine if we
+    // should switch the denoiser mode. Sampling refers to computing the mse for
+    // a sub-sample of the frame (i.e., skip x blocks along row/column), and
+    // only for blocks in that set that have used ZEROMV LAST, along with some
+    // constraint on the sum diff between blocks. This process is called every
+    // ~8 frames, to further reduce complexity.
+    if (cpi->oxcf.noise_sensitivity == 4 &&
+        !cpi->oxcf.screen_content_mode &&
+        cpi->frames_since_key % 8 == 0 &&
+        cm->frame_type != KEY_FRAME) {
+      process_denoiser_mode_change(cpi);
+    }
+#endif
+
+#if CONFIG_MULTITHREAD
+    if (cpi->b_multi_threaded)
+    {
+        /* start loopfilter in separate thread */
+        sem_post(&cpi->h_event_start_lpf);
+    }
+    else
+#endif
+    {
+        vp8_loopfilter_frame(cpi, cm);
+    }
+
+    update_reference_frames(cpi);
+
+#ifdef OUTPUT_YUV_DENOISED
+    vp8_write_yuv_frame(yuv_denoised_file,
+                        &cpi->denoiser.yv12_running_avg[INTRA_FRAME]);
+#endif
+
+#if !(CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+    if (cpi->oxcf.error_resilient_mode)
+    {
+        cm->refresh_entropy_probs = 0;
+    }
+#endif
+
+#if CONFIG_MULTITHREAD
+    /* wait that filter_level is picked so that we can continue with stream packing */
+    if (cpi->b_multi_threaded)
+        sem_wait(&cpi->h_event_end_lpf);
+#endif
+
+    /* build the bitstream */
+    vp8_pack_bitstream(cpi, dest, dest_end, size);
+
+#if CONFIG_MULTITHREAD
+    /* wait for the lpf thread done */
+    if (cpi->b_multi_threaded)
+    {
+        sem_wait(&cpi->h_event_end_lpf);
+    }
+#endif
+
+    /* Move storing frame_type out of the above loop since it is also
+     * needed in motion search besides loopfilter */
+    cm->last_frame_type = cm->frame_type;
+
+    /* Update rate control heuristics */
+    cpi->total_byte_count += (*size);
+    cpi->projected_frame_size = (*size) << 3;
+
+    if (cpi->oxcf.number_of_layers > 1)
+    {
+        unsigned int i;
+        for (i=cpi->current_layer+1; i<cpi->oxcf.number_of_layers; i++)
+          cpi->layer_context[i].total_byte_count += (*size);
+    }
+
+    if (!active_worst_qchanged)
+        vp8_update_rate_correction_factors(cpi, 2);
+
+    cpi->last_q[cm->frame_type] = cm->base_qindex;
+
+    if (cm->frame_type == KEY_FRAME)
+    {
+        vp8_adjust_key_frame_context(cpi);
+    }
+
+    /* Keep a record of ambient average Q. */
+    if (cm->frame_type != KEY_FRAME)
+        cpi->avg_frame_qindex = (2 + 3 * cpi->avg_frame_qindex + cm->base_qindex) >> 2;
+
+    /* Keep a record from which we can calculate the average Q excluding
+     * GF updates and key frames
+     */
+    if ((cm->frame_type != KEY_FRAME) && ((cpi->oxcf.number_of_layers > 1) ||
+        (!cm->refresh_golden_frame && !cm->refresh_alt_ref_frame)))
+    {
+        cpi->ni_frames++;
+
+        /* Calculate the average Q for normal inter frames (not key or GFU
+         * frames).
+         */
+        if ( cpi->pass == 2 )
+        {
+            cpi->ni_tot_qi += Q;
+            cpi->ni_av_qi = (cpi->ni_tot_qi / cpi->ni_frames);
+        }
+        else
+        {
+            /* Damp value for first few frames */
+            if (cpi->ni_frames > 150 )
+            {
+                cpi->ni_tot_qi += Q;
+                cpi->ni_av_qi = (cpi->ni_tot_qi / cpi->ni_frames);
+            }
+            /* For one pass, early in the clip ... average the current frame Q
+             * value with the worstq entered by the user as a dampening measure
+             */
+            else
+            {
+                cpi->ni_tot_qi += Q;
+                cpi->ni_av_qi = ((cpi->ni_tot_qi / cpi->ni_frames) + cpi->worst_quality + 1) / 2;
+            }
+
+            /* If the average Q is higher than what was used in the last
+             * frame (after going through the recode loop to keep the frame
+             * size within range) then use the last frame value - 1. The -1
+             * is designed to stop Q and hence the data rate, from
+             * progressively falling away during difficult sections, but at
+             * the same time reduce the number of itterations around the
+             * recode loop.
+             */
+            if (Q > cpi->ni_av_qi)
+                cpi->ni_av_qi = Q - 1;
+        }
+    }
+
+    /* Update the buffer level variable. */
+    /* Non-viewable frames are a special case and are treated as pure overhead. */
+    if ( !cm->show_frame )
+        cpi->bits_off_target -= cpi->projected_frame_size;
+    else
+        cpi->bits_off_target += cpi->av_per_frame_bandwidth - cpi->projected_frame_size;
+
+    /* Clip the buffer level to the maximum specified buffer size */
+    if (cpi->bits_off_target > cpi->oxcf.maximum_buffer_size)
+        cpi->bits_off_target = cpi->oxcf.maximum_buffer_size;
+
+    // If the frame dropper is not enabled, don't let the buffer level go below
+    // some threshold, given here by -|maximum_buffer_size|. For now we only do
+    // this for screen content input.
+    if (cpi->drop_frames_allowed == 0 && cpi->oxcf.screen_content_mode &&
+        cpi->bits_off_target < -cpi->oxcf.maximum_buffer_size)
+        cpi->bits_off_target = -cpi->oxcf.maximum_buffer_size;
+
+    /* Rolling monitors of whether we are over or underspending used to
+     * help regulate min and Max Q in two pass.
+     */
+    cpi->rolling_target_bits = ((cpi->rolling_target_bits * 3) + cpi->this_frame_target + 2) / 4;
+    cpi->rolling_actual_bits = ((cpi->rolling_actual_bits * 3) + cpi->projected_frame_size + 2) / 4;
+    cpi->long_rolling_target_bits = ((cpi->long_rolling_target_bits * 31) + cpi->this_frame_target + 16) / 32;
+    cpi->long_rolling_actual_bits = ((cpi->long_rolling_actual_bits * 31) + cpi->projected_frame_size + 16) / 32;
+
+    /* Actual bits spent */
+    cpi->total_actual_bits += cpi->projected_frame_size;
+
+    /* Debug stats */
+    cpi->total_target_vs_actual += (cpi->this_frame_target - cpi->projected_frame_size);
+
+    cpi->buffer_level = cpi->bits_off_target;
+
+    /* Propagate values to higher temporal layers */
+    if (cpi->oxcf.number_of_layers > 1)
+    {
+        unsigned int i;
+
+        for (i=cpi->current_layer+1; i<cpi->oxcf.number_of_layers; i++)
+        {
+            LAYER_CONTEXT *lc = &cpi->layer_context[i];
+            int bits_off_for_this_layer =
+               (int)(lc->target_bandwidth / lc->framerate -
+                     cpi->projected_frame_size);
+
+            lc->bits_off_target += bits_off_for_this_layer;
+
+            /* Clip buffer level to maximum buffer size for the layer */
+            if (lc->bits_off_target > lc->maximum_buffer_size)
+                lc->bits_off_target = lc->maximum_buffer_size;
+
+            lc->total_actual_bits += cpi->projected_frame_size;
+            lc->total_target_vs_actual += bits_off_for_this_layer;
+            lc->buffer_level = lc->bits_off_target;
+        }
+    }
+
+    /* Update bits left to the kf and gf groups to account for overshoot
+     * or undershoot on these frames
+     */
+    if (cm->frame_type == KEY_FRAME)
+    {
+        cpi->twopass.kf_group_bits += cpi->this_frame_target - cpi->projected_frame_size;
+
+        if (cpi->twopass.kf_group_bits < 0)
+            cpi->twopass.kf_group_bits = 0 ;
+    }
+    else if (cm->refresh_golden_frame || cm->refresh_alt_ref_frame)
+    {
+        cpi->twopass.gf_group_bits += cpi->this_frame_target - cpi->projected_frame_size;
+
+        if (cpi->twopass.gf_group_bits < 0)
+            cpi->twopass.gf_group_bits = 0 ;
+    }
+
+    if (cm->frame_type != KEY_FRAME)
+    {
+        if (cpi->common.refresh_alt_ref_frame)
+        {
+            cpi->last_skip_false_probs[2] = cpi->prob_skip_false;
+            cpi->last_skip_probs_q[2] = cm->base_qindex;
+        }
+        else if (cpi->common.refresh_golden_frame)
+        {
+            cpi->last_skip_false_probs[1] = cpi->prob_skip_false;
+            cpi->last_skip_probs_q[1] = cm->base_qindex;
+        }
+        else
+        {
+            cpi->last_skip_false_probs[0] = cpi->prob_skip_false;
+            cpi->last_skip_probs_q[0] = cm->base_qindex;
+
+            /* update the baseline */
+            cpi->base_skip_false_prob[cm->base_qindex] = cpi->prob_skip_false;
+
+        }
+    }
+
+#if 0 && CONFIG_INTERNAL_STATS
+    {
+        FILE *f = fopen("tmp.stt", "a");
+
+        vp8_clear_system_state();
+
+        if (cpi->twopass.total_left_stats.coded_error != 0.0)
+            fprintf(f, "%10d %10d %10d %10d %10d %10"PRId64" %10"PRId64
+                       "%10"PRId64" %10d %6d %6d %6d %6d %5d %5d %5d %8d "
+                       "%8.2lf %"PRId64" %10.3lf %10"PRId64" %8d\n",
+                       cpi->common.current_video_frame, cpi->this_frame_target,
+                       cpi->projected_frame_size,
+                       (cpi->projected_frame_size - cpi->this_frame_target),
+                       cpi->total_target_vs_actual,
+                       cpi->buffer_level,
+                       (cpi->oxcf.starting_buffer_level-cpi->bits_off_target),
+                       cpi->total_actual_bits, cm->base_qindex,
+                       cpi->active_best_quality, cpi->active_worst_quality,
+                       cpi->ni_av_qi, cpi->cq_target_quality,
+                       cm->refresh_golden_frame, cm->refresh_alt_ref_frame,
+                       cm->frame_type, cpi->gfu_boost,
+                       cpi->twopass.est_max_qcorrection_factor,
+                       cpi->twopass.bits_left,
+                       cpi->twopass.total_left_stats.coded_error,
+                       (double)cpi->twopass.bits_left /
+                           cpi->twopass.total_left_stats.coded_error,
+                       cpi->tot_recode_hits);
+        else
+            fprintf(f, "%10d %10d %10d %10d %10d %10"PRId64" %10"PRId64
+                       "%10"PRId64" %10d %6d %6d %6d %6d %5d %5d %5d %8d "
+                       "%8.2lf %"PRId64" %10.3lf %8d\n",
+                       cpi->common.current_video_frame, cpi->this_frame_target,
+                       cpi->projected_frame_size,
+                       (cpi->projected_frame_size - cpi->this_frame_target),
+                       cpi->total_target_vs_actual,
+                       cpi->buffer_level,
+                       (cpi->oxcf.starting_buffer_level-cpi->bits_off_target),
+                       cpi->total_actual_bits, cm->base_qindex,
+                       cpi->active_best_quality, cpi->active_worst_quality,
+                       cpi->ni_av_qi, cpi->cq_target_quality,
+                       cm->refresh_golden_frame, cm->refresh_alt_ref_frame,
+                       cm->frame_type, cpi->gfu_boost,
+                       cpi->twopass.est_max_qcorrection_factor,
+                       cpi->twopass.bits_left,
+                       cpi->twopass.total_left_stats.coded_error,
+                       cpi->tot_recode_hits);
+
+        fclose(f);
+
+        {
+            FILE *fmodes = fopen("Modes.stt", "a");
+
+            fprintf(fmodes, "%6d:%1d:%1d:%1d ",
+                        cpi->common.current_video_frame,
+                        cm->frame_type, cm->refresh_golden_frame,
+                        cm->refresh_alt_ref_frame);
+
+            fprintf(fmodes, "\n");
+
+            fclose(fmodes);
+        }
+    }
+
+#endif
+
+    if (cm->refresh_golden_frame == 1)
+        cm->frame_flags = cm->frame_flags | FRAMEFLAGS_GOLDEN;
+    else
+        cm->frame_flags = cm->frame_flags&~FRAMEFLAGS_GOLDEN;
+
+    if (cm->refresh_alt_ref_frame == 1)
+        cm->frame_flags = cm->frame_flags | FRAMEFLAGS_ALTREF;
+    else
+        cm->frame_flags = cm->frame_flags&~FRAMEFLAGS_ALTREF;
+
+
+    if (cm->refresh_last_frame & cm->refresh_golden_frame)
+        /* both refreshed */
+        cpi->gold_is_last = 1;
+    else if (cm->refresh_last_frame ^ cm->refresh_golden_frame)
+        /* 1 refreshed but not the other */
+        cpi->gold_is_last = 0;
+
+    if (cm->refresh_last_frame & cm->refresh_alt_ref_frame)
+        /* both refreshed */
+        cpi->alt_is_last = 1;
+    else if (cm->refresh_last_frame ^ cm->refresh_alt_ref_frame)
+        /* 1 refreshed but not the other */
+        cpi->alt_is_last = 0;
+
+    if (cm->refresh_alt_ref_frame & cm->refresh_golden_frame)
+        /* both refreshed */
+        cpi->gold_is_alt = 1;
+    else if (cm->refresh_alt_ref_frame ^ cm->refresh_golden_frame)
+        /* 1 refreshed but not the other */
+        cpi->gold_is_alt = 0;
+
+    cpi->ref_frame_flags = VP8_ALTR_FRAME | VP8_GOLD_FRAME | VP8_LAST_FRAME;
+
+    if (cpi->gold_is_last)
+        cpi->ref_frame_flags &= ~VP8_GOLD_FRAME;
+
+    if (cpi->alt_is_last)
+        cpi->ref_frame_flags &= ~VP8_ALTR_FRAME;
+
+    if (cpi->gold_is_alt)
+        cpi->ref_frame_flags &= ~VP8_ALTR_FRAME;
+
+
+    if (!cpi->oxcf.error_resilient_mode)
+    {
+        if (cpi->oxcf.play_alternate && cm->refresh_alt_ref_frame && (cm->frame_type != KEY_FRAME))
+            /* Update the alternate reference frame stats as appropriate. */
+            update_alt_ref_frame_stats(cpi);
+        else
+            /* Update the Golden frame stats as appropriate. */
+            update_golden_frame_stats(cpi);
+    }
+
+    if (cm->frame_type == KEY_FRAME)
+    {
+        /* Tell the caller that the frame was coded as a key frame */
+        *frame_flags = cm->frame_flags | FRAMEFLAGS_KEY;
+
+        /* As this frame is a key frame  the next defaults to an inter frame. */
+        cm->frame_type = INTER_FRAME;
+
+        cpi->last_frame_percent_intra = 100;
+    }
+    else
+    {
+        *frame_flags = cm->frame_flags&~FRAMEFLAGS_KEY;
+
+        cpi->last_frame_percent_intra = cpi->this_frame_percent_intra;
+    }
+
+    /* Clear the one shot update flags for segmentation map and mode/ref
+     * loop filter deltas.
+     */
+    cpi->mb.e_mbd.update_mb_segmentation_map = 0;
+    cpi->mb.e_mbd.update_mb_segmentation_data = 0;
+    cpi->mb.e_mbd.mode_ref_lf_delta_update = 0;
+
+
+    /* Dont increment frame counters if this was an altref buffer update
+     * not a real frame
+     */
+    if (cm->show_frame)
+    {
+        cm->current_video_frame++;
+        cpi->frames_since_key++;
+        cpi->temporal_pattern_counter++;
+    }
+
+    /* reset to normal state now that we are done. */
+
+
+
+#if 0
+    {
+        char filename[512];
+        FILE *recon_file;
+        sprintf(filename, "enc%04d.yuv", (int) cm->current_video_frame);
+        recon_file = fopen(filename, "wb");
+        fwrite(cm->yv12_fb[cm->lst_fb_idx].buffer_alloc,
+               cm->yv12_fb[cm->lst_fb_idx].frame_size, 1, recon_file);
+        fclose(recon_file);
+    }
+#endif
+
+    /* DEBUG */
+    /* vp8_write_yuv_frame("encoder_recon.yuv", cm->frame_to_show); */
+
+
+}
+#if !CONFIG_REALTIME_ONLY
+static void Pass2Encode(VP8_COMP *cpi, unsigned long *size, unsigned char *dest, unsigned char * dest_end, unsigned int *frame_flags)
+{
+
+    if (!cpi->common.refresh_alt_ref_frame)
+        vp8_second_pass(cpi);
+
+    encode_frame_to_data_rate(cpi, size, dest, dest_end, frame_flags);
+    cpi->twopass.bits_left -= 8 * (int)(*size);
+
+    if (!cpi->common.refresh_alt_ref_frame)
+    {
+        double two_pass_min_rate = (double)(cpi->oxcf.target_bandwidth
+            *cpi->oxcf.two_pass_vbrmin_section / 100);
+        cpi->twopass.bits_left += (int64_t)(two_pass_min_rate / cpi->framerate);
+    }
+}
+#endif
+
+int vp8_receive_raw_frame(VP8_COMP *cpi, unsigned int frame_flags, YV12_BUFFER_CONFIG *sd, int64_t time_stamp, int64_t end_time)
+{
+    struct vpx_usec_timer  timer;
+    int                    res = 0;
+
+    vpx_usec_timer_start(&timer);
+
+    /* Reinit the lookahead buffer if the frame size changes */
+    if (sd->y_width != cpi->oxcf.Width || sd->y_height != cpi->oxcf.Height)
+    {
+        assert(cpi->oxcf.lag_in_frames < 2);
+        dealloc_raw_frame_buffers(cpi);
+        alloc_raw_frame_buffers(cpi);
+    }
+
+    if(vp8_lookahead_push(cpi->lookahead, sd, time_stamp, end_time,
+                          frame_flags, cpi->active_map_enabled ? cpi->active_map : NULL))
+        res = -1;
+    vpx_usec_timer_mark(&timer);
+    cpi->time_receive_data += vpx_usec_timer_elapsed(&timer);
+
+    return res;
+}
+
+
+static int frame_is_reference(const VP8_COMP *cpi)
+{
+    const VP8_COMMON *cm = &cpi->common;
+    const MACROBLOCKD *xd = &cpi->mb.e_mbd;
+
+    return cm->frame_type == KEY_FRAME || cm->refresh_last_frame
+           || cm->refresh_golden_frame || cm->refresh_alt_ref_frame
+           || cm->copy_buffer_to_gf || cm->copy_buffer_to_arf
+           || cm->refresh_entropy_probs
+           || xd->mode_ref_lf_delta_update
+           || xd->update_mb_segmentation_map || xd->update_mb_segmentation_data;
+}
+
+
+int vp8_get_compressed_data(VP8_COMP *cpi, unsigned int *frame_flags, unsigned long *size, unsigned char *dest, unsigned char *dest_end, int64_t *time_stamp, int64_t *time_end, int flush)
+{
+    VP8_COMMON *cm;
+    struct vpx_usec_timer  tsctimer;
+    struct vpx_usec_timer  ticktimer;
+    struct vpx_usec_timer  cmptimer;
+    YV12_BUFFER_CONFIG    *force_src_buffer = NULL;
+
+    if (!cpi)
+        return -1;
+
+    cm = &cpi->common;
+
+    if (setjmp(cpi->common.error.jmp))
+    {
+        cpi->common.error.setjmp = 0;
+        vp8_clear_system_state();
+        return VPX_CODEC_CORRUPT_FRAME;
+    }
+
+    cpi->common.error.setjmp = 1;
+
+    vpx_usec_timer_start(&cmptimer);
+
+    cpi->source = NULL;
+
+#if !CONFIG_REALTIME_ONLY
+    /* Should we code an alternate reference frame */
+    if (cpi->oxcf.error_resilient_mode == 0 &&
+        cpi->oxcf.play_alternate &&
+        cpi->source_alt_ref_pending)
+    {
+        if ((cpi->source = vp8_lookahead_peek(cpi->lookahead,
+                                              cpi->frames_till_gf_update_due,
+                                              PEEK_FORWARD)))
+        {
+            cpi->alt_ref_source = cpi->source;
+            if (cpi->oxcf.arnr_max_frames > 0)
+            {
+                vp8_temporal_filter_prepare_c(cpi,
+                                              cpi->frames_till_gf_update_due);
+                force_src_buffer = &cpi->alt_ref_buffer;
+            }
+            cpi->frames_till_alt_ref_frame = cpi->frames_till_gf_update_due;
+            cm->refresh_alt_ref_frame = 1;
+            cm->refresh_golden_frame = 0;
+            cm->refresh_last_frame = 0;
+            cm->show_frame = 0;
+            /* Clear Pending alt Ref flag. */
+            cpi->source_alt_ref_pending = 0;
+            cpi->is_src_frame_alt_ref = 0;
+        }
+    }
+#endif
+
+    if (!cpi->source)
+    {
+        /* Read last frame source if we are encoding first pass. */
+        if (cpi->pass == 1 && cm->current_video_frame > 0)
+        {
+            if((cpi->last_source = vp8_lookahead_peek(cpi->lookahead, 1,
+                                                      PEEK_BACKWARD)) == NULL)
+              return -1;
+        }
+
+
+        if ((cpi->source = vp8_lookahead_pop(cpi->lookahead, flush)))
+        {
+            cm->show_frame = 1;
+
+            cpi->is_src_frame_alt_ref = cpi->alt_ref_source
+                                        && (cpi->source == cpi->alt_ref_source);
+
+            if(cpi->is_src_frame_alt_ref)
+                cpi->alt_ref_source = NULL;
+        }
+    }
+
+    if (cpi->source)
+    {
+        cpi->Source = force_src_buffer ? force_src_buffer : &cpi->source->img;
+        cpi->un_scaled_source = cpi->Source;
+        *time_stamp = cpi->source->ts_start;
+        *time_end = cpi->source->ts_end;
+        *frame_flags = cpi->source->flags;
+
+        if (cpi->pass == 1 && cm->current_video_frame > 0)
+        {
+            cpi->last_frame_unscaled_source = &cpi->last_source->img;
+        }
+    }
+    else
+    {
+        *size = 0;
+#if !CONFIG_REALTIME_ONLY
+
+        if (flush && cpi->pass == 1 && !cpi->twopass.first_pass_done)
+        {
+            vp8_end_first_pass(cpi);    /* get last stats packet */
+            cpi->twopass.first_pass_done = 1;
+        }
+
+#endif
+
+        return -1;
+    }
+
+    if (cpi->source->ts_start < cpi->first_time_stamp_ever)
+    {
+        cpi->first_time_stamp_ever = cpi->source->ts_start;
+        cpi->last_end_time_stamp_seen = cpi->source->ts_start;
+    }
+
+    /* adjust frame rates based on timestamps given */
+    if (cm->show_frame)
+    {
+        int64_t this_duration;
+        int step = 0;
+
+        if (cpi->source->ts_start == cpi->first_time_stamp_ever)
+        {
+            this_duration = cpi->source->ts_end - cpi->source->ts_start;
+            step = 1;
+        }
+        else
+        {
+            int64_t last_duration;
+
+            this_duration = cpi->source->ts_end - cpi->last_end_time_stamp_seen;
+            last_duration = cpi->last_end_time_stamp_seen
+                            - cpi->last_time_stamp_seen;
+            /* do a step update if the duration changes by 10% */
+            if (last_duration)
+                step = (int)(((this_duration - last_duration) *
+                            10 / last_duration));
+        }
+
+        if (this_duration)
+        {
+            if (step)
+                cpi->ref_framerate = 10000000.0 / this_duration;
+            else
+            {
+                double avg_duration, interval;
+
+                /* Average this frame's rate into the last second's average
+                 * frame rate. If we haven't seen 1 second yet, then average
+                 * over the whole interval seen.
+                 */
+                interval = (double)(cpi->source->ts_end -
+                                    cpi->first_time_stamp_ever);
+                if(interval > 10000000.0)
+                    interval = 10000000;
+
+                avg_duration = 10000000.0 / cpi->ref_framerate;
+                avg_duration *= (interval - avg_duration + this_duration);
+                avg_duration /= interval;
+
+                cpi->ref_framerate = 10000000.0 / avg_duration;
+            }
+#if CONFIG_MULTI_RES_ENCODING
+            if (cpi->oxcf.mr_total_resolutions > 1) {
+              LOWER_RES_FRAME_INFO* low_res_frame_info = (LOWER_RES_FRAME_INFO*)
+                  cpi->oxcf.mr_low_res_mode_info;
+              // Frame rate should be the same for all spatial layers in
+              // multi-res-encoding (simulcast), so we constrain the frame for
+              // higher layers to be that of lowest resolution. This is needed
+              // as he application may decide to skip encoding a high layer and
+              // then start again, in which case a big jump in time-stamps will
+              // be received for that high layer, which will yield an incorrect
+              // frame rate (from time-stamp adjustment in above calculation).
+              if (cpi->oxcf.mr_encoder_id) {
+                 cpi->ref_framerate = low_res_frame_info->low_res_framerate;
+              }
+              else {
+                // Keep track of frame rate for lowest resolution.
+                low_res_frame_info->low_res_framerate = cpi->ref_framerate;
+              }
+            }
+#endif
+            if (cpi->oxcf.number_of_layers > 1)
+            {
+                unsigned int i;
+
+                /* Update frame rates for each layer */
+                assert(cpi->oxcf.number_of_layers <= VPX_TS_MAX_LAYERS);
+                for (i = 0; i < cpi->oxcf.number_of_layers &&
+                     i < VPX_TS_MAX_LAYERS; ++i)
+                {
+                    LAYER_CONTEXT *lc = &cpi->layer_context[i];
+                    lc->framerate = cpi->ref_framerate /
+                                    cpi->oxcf.rate_decimator[i];
+                }
+            }
+            else
+                vp8_new_framerate(cpi, cpi->ref_framerate);
+        }
+
+        cpi->last_time_stamp_seen = cpi->source->ts_start;
+        cpi->last_end_time_stamp_seen = cpi->source->ts_end;
+    }
+
+    if (cpi->oxcf.number_of_layers > 1)
+    {
+        int layer;
+
+        update_layer_contexts (cpi);
+
+        /* Restore layer specific context & set frame rate */
+        if (cpi->temporal_layer_id >= 0) {
+          layer = cpi->temporal_layer_id;
+        } else {
+          layer = cpi->oxcf.layer_id[
+                  cpi->temporal_pattern_counter % cpi->oxcf.periodicity];
+        }
+        restore_layer_context (cpi, layer);
+        vp8_new_framerate(cpi, cpi->layer_context[layer].framerate);
+    }
+
+    if (cpi->compressor_speed == 2)
+    {
+        vpx_usec_timer_start(&tsctimer);
+        vpx_usec_timer_start(&ticktimer);
+    }
+
+    cpi->lf_zeromv_pct = (cpi->zeromv_count * 100)/cm->MBs;
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+    {
+        int i;
+        const int num_part = (1 << cm->multi_token_partition);
+        /* the available bytes in dest */
+        const unsigned long dest_size = dest_end - dest;
+        const int tok_part_buff_size = (dest_size * 9) / (10 * num_part);
+
+        unsigned char *dp = dest;
+
+        cpi->partition_d[0] = dp;
+        dp += dest_size/10;         /* reserve 1/10 for control partition */
+        cpi->partition_d_end[0] = dp;
+
+        for(i = 0; i < num_part; i++)
+        {
+            cpi->partition_d[i + 1] = dp;
+            dp += tok_part_buff_size;
+            cpi->partition_d_end[i + 1] = dp;
+        }
+    }
+#endif
+
+    /* start with a 0 size frame */
+    *size = 0;
+
+    /* Clear down mmx registers */
+    vp8_clear_system_state();
+
+    cm->frame_type = INTER_FRAME;
+    cm->frame_flags = *frame_flags;
+
+#if 0
+
+    if (cm->refresh_alt_ref_frame)
+    {
+        cm->refresh_golden_frame = 0;
+        cm->refresh_last_frame = 0;
+    }
+    else
+    {
+        cm->refresh_golden_frame = 0;
+        cm->refresh_last_frame = 1;
+    }
+
+#endif
+    /* find a free buffer for the new frame */
+    {
+        int i = 0;
+        for(; i < NUM_YV12_BUFFERS; i++)
+        {
+            if(!cm->yv12_fb[i].flags)
+            {
+                cm->new_fb_idx = i;
+                break;
+            }
+        }
+
+        assert(i < NUM_YV12_BUFFERS );
+    }
+#if !CONFIG_REALTIME_ONLY
+
+    if (cpi->pass == 1)
+    {
+        Pass1Encode(cpi, size, dest, frame_flags);
+    }
+    else if (cpi->pass == 2)
+    {
+        Pass2Encode(cpi, size, dest, dest_end, frame_flags);
+    }
+    else
+#endif
+        encode_frame_to_data_rate(cpi, size, dest, dest_end, frame_flags);
+
+    if (cpi->compressor_speed == 2)
+    {
+        unsigned int duration, duration2;
+        vpx_usec_timer_mark(&tsctimer);
+        vpx_usec_timer_mark(&ticktimer);
+
+        duration = (int)(vpx_usec_timer_elapsed(&ticktimer));
+        duration2 = (unsigned int)((double)duration / 2);
+
+        if (cm->frame_type != KEY_FRAME)
+        {
+            if (cpi->avg_encode_time == 0)
+                cpi->avg_encode_time = duration;
+            else
+                cpi->avg_encode_time = (7 * cpi->avg_encode_time + duration) >> 3;
+        }
+
+        if (duration2)
+        {
+            {
+
+                if (cpi->avg_pick_mode_time == 0)
+                    cpi->avg_pick_mode_time = duration2;
+                else
+                    cpi->avg_pick_mode_time = (7 * cpi->avg_pick_mode_time + duration2) >> 3;
+            }
+        }
+
+    }
+
+    if (cm->refresh_entropy_probs == 0)
+    {
+        memcpy(&cm->fc, &cm->lfc, sizeof(cm->fc));
+    }
+
+    /* Save the contexts separately for alt ref, gold and last. */
+    /* (TODO jbb -> Optimize this with pointers to avoid extra copies. ) */
+    if(cm->refresh_alt_ref_frame)
+        memcpy(&cpi->lfc_a, &cm->fc, sizeof(cm->fc));
+
+    if(cm->refresh_golden_frame)
+        memcpy(&cpi->lfc_g, &cm->fc, sizeof(cm->fc));
+
+    if(cm->refresh_last_frame)
+        memcpy(&cpi->lfc_n, &cm->fc, sizeof(cm->fc));
+
+    /* if its a dropped frame honor the requests on subsequent frames */
+    if (*size > 0)
+    {
+        cpi->droppable = !frame_is_reference(cpi);
+
+        /* return to normal state */
+        cm->refresh_entropy_probs = 1;
+        cm->refresh_alt_ref_frame = 0;
+        cm->refresh_golden_frame = 0;
+        cm->refresh_last_frame = 1;
+        cm->frame_type = INTER_FRAME;
+
+    }
+
+    /* Save layer specific state */
+    if (cpi->oxcf.number_of_layers > 1)
+        save_layer_context (cpi);
+
+    vpx_usec_timer_mark(&cmptimer);
+    cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer);
+
+    if (cpi->b_calculate_psnr && cpi->pass != 1 && cm->show_frame)
+    {
+        generate_psnr_packet(cpi);
+    }
+
+#if CONFIG_INTERNAL_STATS
+
+    if (cpi->pass != 1)
+    {
+        cpi->bytes += *size;
+
+        if (cm->show_frame)
+        {
+            cpi->common.show_frame_mi = cpi->common.mi;
+            cpi->count ++;
+
+            if (cpi->b_calculate_psnr)
+            {
+                uint64_t ye,ue,ve;
+                double frame_psnr;
+                YV12_BUFFER_CONFIG      *orig = cpi->Source;
+                YV12_BUFFER_CONFIG      *recon = cpi->common.frame_to_show;
+                unsigned int y_width = cpi->common.Width;
+                unsigned int y_height = cpi->common.Height;
+                unsigned int uv_width = (y_width + 1) / 2;
+                unsigned int uv_height = (y_height + 1) / 2;
+                int y_samples = y_height * y_width;
+                int uv_samples = uv_height * uv_width;
+                int t_samples = y_samples + 2 * uv_samples;
+                double sq_error;
+
+                ye = calc_plane_error(orig->y_buffer, orig->y_stride,
+                  recon->y_buffer, recon->y_stride, y_width, y_height);
+
+                ue = calc_plane_error(orig->u_buffer, orig->uv_stride,
+                  recon->u_buffer, recon->uv_stride, uv_width, uv_height);
+
+                ve = calc_plane_error(orig->v_buffer, orig->uv_stride,
+                  recon->v_buffer, recon->uv_stride, uv_width, uv_height);
+
+                sq_error = (double)(ye + ue + ve);
+
+                frame_psnr = vpx_sse_to_psnr(t_samples, 255.0, sq_error);
+
+                cpi->total_y += vpx_sse_to_psnr(y_samples, 255.0, (double)ye);
+                cpi->total_u += vpx_sse_to_psnr(uv_samples, 255.0, (double)ue);
+                cpi->total_v += vpx_sse_to_psnr(uv_samples, 255.0, (double)ve);
+                cpi->total_sq_error += sq_error;
+                cpi->total  += frame_psnr;
+#if CONFIG_POSTPROC
+                {
+                    YV12_BUFFER_CONFIG      *pp = &cm->post_proc_buffer;
+                    double sq_error2;
+                    double frame_psnr2, frame_ssim2 = 0;
+                    double weight = 0;
+
+                    vp8_deblock(cm, cm->frame_to_show, &cm->post_proc_buffer, cm->filter_level * 10 / 6, 1, 0);
+                    vp8_clear_system_state();
+
+                    ye = calc_plane_error(orig->y_buffer, orig->y_stride,
+                      pp->y_buffer, pp->y_stride, y_width, y_height);
+
+                    ue = calc_plane_error(orig->u_buffer, orig->uv_stride,
+                      pp->u_buffer, pp->uv_stride, uv_width, uv_height);
+
+                    ve = calc_plane_error(orig->v_buffer, orig->uv_stride,
+                      pp->v_buffer, pp->uv_stride, uv_width, uv_height);
+
+                    sq_error2 = (double)(ye + ue + ve);
+
+                    frame_psnr2 = vpx_sse_to_psnr(t_samples, 255.0, sq_error2);
+
+                    cpi->totalp_y += vpx_sse_to_psnr(y_samples,
+                                                     255.0, (double)ye);
+                    cpi->totalp_u += vpx_sse_to_psnr(uv_samples,
+                                                     255.0, (double)ue);
+                    cpi->totalp_v += vpx_sse_to_psnr(uv_samples,
+                                                     255.0, (double)ve);
+                    cpi->total_sq_error2 += sq_error2;
+                    cpi->totalp  += frame_psnr2;
+
+                    frame_ssim2 = vpx_calc_ssim(cpi->Source,
+                      &cm->post_proc_buffer, &weight);
+
+                    cpi->summed_quality += frame_ssim2 * weight;
+                    cpi->summed_weights += weight;
+
+                    if (cpi->oxcf.number_of_layers > 1)
+                    {
+                         unsigned int i;
+
+                         for (i=cpi->current_layer;
+                                       i<cpi->oxcf.number_of_layers; i++)
+                         {
+                             cpi->frames_in_layer[i]++;
+
+                             cpi->bytes_in_layer[i] += *size;
+                             cpi->sum_psnr[i]       += frame_psnr;
+                             cpi->sum_psnr_p[i]     += frame_psnr2;
+                             cpi->total_error2[i]   += sq_error;
+                             cpi->total_error2_p[i] += sq_error2;
+                             cpi->sum_ssim[i]       += frame_ssim2 * weight;
+                             cpi->sum_weights[i]    += weight;
+                         }
+                    }
+                }
+#endif
+            }
+
+            if (cpi->b_calculate_ssimg)
+            {
+                double y, u, v, frame_all;
+                frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show,
+                    &y, &u, &v);
+
+                if (cpi->oxcf.number_of_layers > 1)
+                {
+                    unsigned int i;
+
+                    for (i=cpi->current_layer;
+                         i<cpi->oxcf.number_of_layers; i++)
+                    {
+                        if (!cpi->b_calculate_psnr)
+                            cpi->frames_in_layer[i]++;
+
+                        cpi->total_ssimg_y_in_layer[i] += y;
+                        cpi->total_ssimg_u_in_layer[i] += u;
+                        cpi->total_ssimg_v_in_layer[i] += v;
+                        cpi->total_ssimg_all_in_layer[i] += frame_all;
+                    }
+                }
+                else
+                {
+                    cpi->total_ssimg_y += y;
+                    cpi->total_ssimg_u += u;
+                    cpi->total_ssimg_v += v;
+                    cpi->total_ssimg_all += frame_all;
+                }
+            }
+
+        }
+    }
+
+#if 0
+
+    if (cpi->common.frame_type != 0 && cpi->common.base_qindex == cpi->oxcf.worst_allowed_q)
+    {
+        skiptruecount += cpi->skip_true_count;
+        skipfalsecount += cpi->skip_false_count;
+    }
+
+#endif
+#if 0
+
+    if (cpi->pass != 1)
+    {
+        FILE *f = fopen("skip.stt", "a");
+        fprintf(f, "frame:%4d flags:%4x Q:%4d P:%4d Size:%5d\n", cpi->common.current_video_frame, *frame_flags, cpi->common.base_qindex, cpi->prob_skip_false, *size);
+
+        if (cpi->is_src_frame_alt_ref == 1)
+            fprintf(f, "skipcount: %4d framesize: %d\n", cpi->skip_true_count , *size);
+
+        fclose(f);
+    }
+
+#endif
+#endif
+
+    cpi->common.error.setjmp = 0;
+
+    return 0;
+}
+
+int vp8_get_preview_raw_frame(VP8_COMP *cpi, YV12_BUFFER_CONFIG *dest, vp8_ppflags_t *flags)
+{
+    if (cpi->common.refresh_alt_ref_frame)
+        return -1;
+    else
+    {
+        int ret;
+
+#if CONFIG_POSTPROC
+        cpi->common.show_frame_mi = cpi->common.mi;
+        ret = vp8_post_proc_frame(&cpi->common, dest, flags);
+#else
+        (void)flags;
+
+        if (cpi->common.frame_to_show)
+        {
+            *dest = *cpi->common.frame_to_show;
+            dest->y_width = cpi->common.Width;
+            dest->y_height = cpi->common.Height;
+            dest->uv_height = cpi->common.Height / 2;
+            ret = 0;
+        }
+        else
+        {
+            ret = -1;
+        }
+
+#endif
+        vp8_clear_system_state();
+        return ret;
+    }
+}
+
+int vp8_set_roimap(VP8_COMP *cpi, unsigned char *map, unsigned int rows, unsigned int cols, int delta_q[4], int delta_lf[4], unsigned int threshold[4])
+{
+    signed char feature_data[MB_LVL_MAX][MAX_MB_SEGMENTS];
+    int internal_delta_q[MAX_MB_SEGMENTS];
+    const int range = 63;
+    int i;
+
+    // This method is currently incompatible with the cyclic refresh method
+    if ( cpi->cyclic_refresh_mode_enabled )
+        return -1;
+
+    // Check number of rows and columns match
+    if (cpi->common.mb_rows != (int)rows || cpi->common.mb_cols != (int)cols)
+        return -1;
+
+    // Range check the delta Q values and convert the external Q range values
+    // to internal ones.
+    if ( (abs(delta_q[0]) > range) || (abs(delta_q[1]) > range) ||
+         (abs(delta_q[2]) > range) || (abs(delta_q[3]) > range) )
+        return -1;
+
+    // Range check the delta lf values
+    if ( (abs(delta_lf[0]) > range) || (abs(delta_lf[1]) > range) ||
+         (abs(delta_lf[2]) > range) || (abs(delta_lf[3]) > range) )
+        return -1;
+
+    if (!map)
+    {
+        disable_segmentation(cpi);
+        return 0;
+    }
+
+    // Translate the external delta q values to internal values.
+    for ( i = 0; i < MAX_MB_SEGMENTS; i++ )
+        internal_delta_q[i] =
+            ( delta_q[i] >= 0 ) ? q_trans[delta_q[i]] : -q_trans[-delta_q[i]];
+
+    /* Set the segmentation Map */
+    set_segmentation_map(cpi, map);
+
+    /* Activate segmentation. */
+    enable_segmentation(cpi);
+
+    /* Set up the quant segment data */
+    feature_data[MB_LVL_ALT_Q][0] = internal_delta_q[0];
+    feature_data[MB_LVL_ALT_Q][1] = internal_delta_q[1];
+    feature_data[MB_LVL_ALT_Q][2] = internal_delta_q[2];
+    feature_data[MB_LVL_ALT_Q][3] = internal_delta_q[3];
+
+    /* Set up the loop segment data s */
+    feature_data[MB_LVL_ALT_LF][0] = delta_lf[0];
+    feature_data[MB_LVL_ALT_LF][1] = delta_lf[1];
+    feature_data[MB_LVL_ALT_LF][2] = delta_lf[2];
+    feature_data[MB_LVL_ALT_LF][3] = delta_lf[3];
+
+    cpi->segment_encode_breakout[0] = threshold[0];
+    cpi->segment_encode_breakout[1] = threshold[1];
+    cpi->segment_encode_breakout[2] = threshold[2];
+    cpi->segment_encode_breakout[3] = threshold[3];
+
+    /* Initialise the feature data structure */
+    set_segment_data(cpi, &feature_data[0][0], SEGMENT_DELTADATA);
+
+    return 0;
+}
+
+int vp8_set_active_map(VP8_COMP *cpi, unsigned char *map, unsigned int rows, unsigned int cols)
+{
+    if ((int)rows == cpi->common.mb_rows && (int)cols == cpi->common.mb_cols)
+    {
+        if (map)
+        {
+            memcpy(cpi->active_map, map, rows * cols);
+            cpi->active_map_enabled = 1;
+        }
+        else
+            cpi->active_map_enabled = 0;
+
+        return 0;
+    }
+    else
+    {
+        return -1 ;
+    }
+}
+
+int vp8_set_internal_size(VP8_COMP *cpi, VPX_SCALING horiz_mode, VPX_SCALING vert_mode)
+{
+    if (horiz_mode <= ONETWO)
+        cpi->common.horiz_scale = horiz_mode;
+    else
+        return -1;
+
+    if (vert_mode <= ONETWO)
+        cpi->common.vert_scale  = vert_mode;
+    else
+        return -1;
+
+    return 0;
+}
+
+
+
+int vp8_calc_ss_err(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest)
+{
+    int i, j;
+    int Total = 0;
+
+    unsigned char *src = source->y_buffer;
+    unsigned char *dst = dest->y_buffer;
+
+    /* Loop through the Y plane raw and reconstruction data summing
+     * (square differences)
+     */
+    for (i = 0; i < source->y_height; i += 16)
+    {
+        for (j = 0; j < source->y_width; j += 16)
+        {
+            unsigned int sse;
+            Total += vpx_mse16x16(src + j, source->y_stride,
+                                  dst + j, dest->y_stride, &sse);
+        }
+
+        src += 16 * source->y_stride;
+        dst += 16 * dest->y_stride;
+    }
+
+    return Total;
+}
+
+
+int vp8_get_quantizer(VP8_COMP *cpi)
+{
+    return cpi->common.base_qindex;
+}
diff --git a/libvpx/libvpx/vp8/encoder/onyx_int.h b/libvpx/libvpx/vp8/encoder/onyx_int.h
new file mode 100644
index 0000000..44fbbd4
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/onyx_int.h
@@ -0,0 +1,754 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_ONYX_INT_H_
+#define VP8_ENCODER_ONYX_INT_H_
+
+#include <stdio.h>
+#include "vpx_config.h"
+#include "vp8/common/onyx.h"
+#include "treewriter.h"
+#include "tokenize.h"
+#include "vp8/common/onyxc_int.h"
+#include "vpx_dsp/variance.h"
+#include "encodemb.h"
+#include "vp8/encoder/quantize.h"
+#include "vp8/common/entropy.h"
+#include "vp8/common/threading.h"
+#include "vpx_ports/mem.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "vpx/vp8.h"
+#include "mcomp.h"
+#include "vp8/common/findnearmv.h"
+#include "lookahead.h"
+#if CONFIG_TEMPORAL_DENOISING
+#include "vp8/encoder/denoising.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MIN_GF_INTERVAL             4
+#define DEFAULT_GF_INTERVAL         7
+
+#define KEY_FRAME_CONTEXT 5
+
+#define MAX_LAG_BUFFERS (CONFIG_REALTIME_ONLY? 1 : 25)
+
+#define AF_THRESH   25
+#define AF_THRESH2  100
+#define ARF_DECAY_THRESH 12
+
+
+#define MIN_THRESHMULT  32
+#define MAX_THRESHMULT  512
+
+#define GF_ZEROMV_ZBIN_BOOST 12
+#define LF_ZEROMV_ZBIN_BOOST 6
+#define MV_ZBIN_BOOST        4
+#define ZBIN_OQ_MAX 192
+
+#if !(CONFIG_REALTIME_ONLY)
+#define VP8_TEMPORAL_ALT_REF 1
+#endif
+
+typedef struct
+{
+    int kf_indicated;
+    unsigned int frames_since_key;
+    unsigned int frames_since_golden;
+    int filter_level;
+    int frames_till_gf_update_due;
+    int recent_ref_frame_usage[MAX_REF_FRAMES];
+
+    MV_CONTEXT mvc[2];
+    int mvcosts[2][MVvals+1];
+
+#ifdef MODE_STATS
+    int y_modes[5];
+    int uv_modes[4];
+    int b_modes[10];
+    int inter_y_modes[10];
+    int inter_uv_modes[4];
+    int inter_b_modes[10];
+#endif
+
+    vp8_prob ymode_prob[4], uv_mode_prob[3];   /* interframe intra mode probs */
+    vp8_prob kf_ymode_prob[4], kf_uv_mode_prob[3];   /* keyframe "" */
+
+    int ymode_count[5], uv_mode_count[4];  /* intra MB type cts this frame */
+
+    int count_mb_ref_frame_usage[MAX_REF_FRAMES];
+
+    int this_frame_percent_intra;
+    int last_frame_percent_intra;
+
+
+} CODING_CONTEXT;
+
+typedef struct
+{
+    double frame;
+    double intra_error;
+    double coded_error;
+    double ssim_weighted_pred_err;
+    double pcnt_inter;
+    double pcnt_motion;
+    double pcnt_second_ref;
+    double pcnt_neutral;
+    double MVr;
+    double mvr_abs;
+    double MVc;
+    double mvc_abs;
+    double MVrv;
+    double MVcv;
+    double mv_in_out_count;
+    double new_mv_count;
+    double duration;
+    double count;
+}
+FIRSTPASS_STATS;
+
+typedef struct
+{
+    int frames_so_far;
+    double frame_intra_error;
+    double frame_coded_error;
+    double frame_pcnt_inter;
+    double frame_pcnt_motion;
+    double frame_mvr;
+    double frame_mvr_abs;
+    double frame_mvc;
+    double frame_mvc_abs;
+
+} ONEPASS_FRAMESTATS;
+
+
+typedef enum
+{
+    THR_ZERO1          = 0,
+    THR_DC             = 1,
+
+    THR_NEAREST1       = 2,
+    THR_NEAR1          = 3,
+
+    THR_ZERO2          = 4,
+    THR_NEAREST2       = 5,
+
+    THR_ZERO3          = 6,
+    THR_NEAREST3       = 7,
+
+    THR_NEAR2          = 8,
+    THR_NEAR3          = 9,
+
+    THR_V_PRED         = 10,
+    THR_H_PRED         = 11,
+    THR_TM             = 12,
+
+    THR_NEW1           = 13,
+    THR_NEW2           = 14,
+    THR_NEW3           = 15,
+
+    THR_SPLIT1         = 16,
+    THR_SPLIT2         = 17,
+    THR_SPLIT3         = 18,
+
+    THR_B_PRED         = 19
+}
+THR_MODES;
+
+typedef enum
+{
+    DIAMOND = 0,
+    NSTEP = 1,
+    HEX = 2
+} SEARCH_METHODS;
+
+typedef struct
+{
+    int RD;
+    SEARCH_METHODS search_method;
+    int improved_quant;
+    int improved_dct;
+    int auto_filter;
+    int recode_loop;
+    int iterative_sub_pixel;
+    int half_pixel_search;
+    int quarter_pixel_search;
+    int thresh_mult[MAX_MODES];
+    int max_step_search_steps;
+    int first_step;
+    int optimize_coefficients;
+
+    int use_fastquant_for_pick;
+    int no_skip_block4x4_search;
+    int improved_mv_pred;
+
+} SPEED_FEATURES;
+
+typedef struct
+{
+    MACROBLOCK  mb;
+    int segment_counts[MAX_MB_SEGMENTS];
+    int totalrate;
+} MB_ROW_COMP;
+
+typedef struct
+{
+    TOKENEXTRA *start;
+    TOKENEXTRA *stop;
+} TOKENLIST;
+
+typedef struct
+{
+    int ithread;
+    void *ptr1;
+    void *ptr2;
+} ENCODETHREAD_DATA;
+typedef struct
+{
+    int ithread;
+    void *ptr1;
+} LPFTHREAD_DATA;
+
+enum
+{
+    BLOCK_16X8,
+    BLOCK_8X16,
+    BLOCK_8X8,
+    BLOCK_4X4,
+    BLOCK_16X16,
+    BLOCK_MAX_SEGMENTS
+};
+
+typedef struct
+{
+    /* Layer configuration */
+    double framerate;
+    int target_bandwidth;
+
+    /* Layer specific coding parameters */
+    int64_t starting_buffer_level;
+    int64_t optimal_buffer_level;
+    int64_t maximum_buffer_size;
+    int64_t starting_buffer_level_in_ms;
+    int64_t optimal_buffer_level_in_ms;
+    int64_t maximum_buffer_size_in_ms;
+
+    int avg_frame_size_for_layer;
+
+    int64_t buffer_level;
+    int64_t bits_off_target;
+
+    int64_t total_actual_bits;
+    int total_target_vs_actual;
+
+    int worst_quality;
+    int active_worst_quality;
+    int best_quality;
+    int active_best_quality;
+
+    int ni_av_qi;
+    int ni_tot_qi;
+    int ni_frames;
+    int avg_frame_qindex;
+
+    double rate_correction_factor;
+    double key_frame_rate_correction_factor;
+    double gf_rate_correction_factor;
+
+    int zbin_over_quant;
+
+    int inter_frame_target;
+    int64_t total_byte_count;
+
+    int filter_level;
+
+    int last_frame_percent_intra;
+
+    int count_mb_ref_frame_usage[MAX_REF_FRAMES];
+
+} LAYER_CONTEXT;
+
+typedef struct VP8_COMP
+{
+
+    DECLARE_ALIGNED(16, short, Y1quant[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, Y1quant_shift[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, Y1zbin[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, Y1round[QINDEX_RANGE][16]);
+
+    DECLARE_ALIGNED(16, short, Y2quant[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, Y2quant_shift[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, Y2zbin[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, Y2round[QINDEX_RANGE][16]);
+
+    DECLARE_ALIGNED(16, short, UVquant[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, UVquant_shift[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, UVzbin[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, UVround[QINDEX_RANGE][16]);
+
+    DECLARE_ALIGNED(16, short, zrun_zbin_boost_y1[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, zrun_zbin_boost_y2[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, zrun_zbin_boost_uv[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, Y1quant_fast[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, Y2quant_fast[QINDEX_RANGE][16]);
+    DECLARE_ALIGNED(16, short, UVquant_fast[QINDEX_RANGE][16]);
+
+
+    MACROBLOCK mb;
+    VP8_COMMON common;
+    vp8_writer bc[9]; /* one boolcoder for each partition */
+
+    VP8_CONFIG oxcf;
+
+    struct lookahead_ctx    *lookahead;
+    struct lookahead_entry  *source;
+    struct lookahead_entry  *alt_ref_source;
+    struct lookahead_entry  *last_source;
+
+    YV12_BUFFER_CONFIG *Source;
+    YV12_BUFFER_CONFIG *un_scaled_source;
+    YV12_BUFFER_CONFIG scaled_source;
+    YV12_BUFFER_CONFIG *last_frame_unscaled_source;
+
+    unsigned int frames_till_alt_ref_frame;
+    /* frame in src_buffers has been identified to be encoded as an alt ref */
+    int source_alt_ref_pending;
+    /* an alt ref frame has been encoded and is usable */
+    int source_alt_ref_active;
+    /* source of frame to encode is an exact copy of an alt ref frame */
+    int is_src_frame_alt_ref;
+
+    /* golden frame same as last frame ( short circuit gold searches) */
+    int gold_is_last;
+    /* Alt reference frame same as last ( short circuit altref search) */
+    int alt_is_last;
+    /* don't do both alt and gold search ( just do gold). */
+    int gold_is_alt;
+
+    YV12_BUFFER_CONFIG pick_lf_lvl_frame;
+
+    TOKENEXTRA *tok;
+    unsigned int tok_count;
+
+
+    unsigned int frames_since_key;
+    unsigned int key_frame_frequency;
+    unsigned int this_key_frame_forced;
+    unsigned int next_key_frame_forced;
+
+    /* Ambient reconstruction err target for force key frames */
+    int ambient_err;
+
+    unsigned int mode_check_freq[MAX_MODES];
+
+    int rd_baseline_thresh[MAX_MODES];
+
+    int RDMULT;
+    int RDDIV ;
+
+    CODING_CONTEXT coding_context;
+
+    /* Rate targetting variables */
+    int64_t last_prediction_error;
+    int64_t last_intra_error;
+
+    int this_frame_target;
+    int projected_frame_size;
+    int last_q[2];                   /* Separate values for Intra/Inter */
+
+    double rate_correction_factor;
+    double key_frame_rate_correction_factor;
+    double gf_rate_correction_factor;
+
+    int frames_since_golden;
+    /* Count down till next GF */
+    int frames_till_gf_update_due;
+
+    /* GF interval chosen when we coded the last GF */
+    int current_gf_interval;
+
+    /* Total bits overspent becasue of GF boost (cumulative) */
+    int gf_overspend_bits;
+
+    /* Used in the few frames following a GF to recover the extra bits
+     * spent in that GF
+     */
+    int non_gf_bitrate_adjustment;
+
+    /* Extra bits spent on key frames that need to be recovered */
+    int kf_overspend_bits;
+
+    /* Current number of bit s to try and recover on each inter frame. */
+    int kf_bitrate_adjustment;
+    int max_gf_interval;
+    int baseline_gf_interval;
+    int active_arnr_frames;
+
+    int64_t key_frame_count;
+    int prior_key_frame_distance[KEY_FRAME_CONTEXT];
+    /* Current section per frame bandwidth target */
+    int per_frame_bandwidth;
+    /* Average frame size target for clip */
+    int av_per_frame_bandwidth;
+    /* Minimum allocation that should be used for any frame */
+    int min_frame_bandwidth;
+    int inter_frame_target;
+    double output_framerate;
+    int64_t last_time_stamp_seen;
+    int64_t last_end_time_stamp_seen;
+    int64_t first_time_stamp_ever;
+
+    int ni_av_qi;
+    int ni_tot_qi;
+    int ni_frames;
+    int avg_frame_qindex;
+
+    int64_t total_byte_count;
+
+    int buffered_mode;
+
+    double framerate;
+    double ref_framerate;
+    int64_t buffer_level;
+    int64_t bits_off_target;
+
+    int rolling_target_bits;
+    int rolling_actual_bits;
+
+    int long_rolling_target_bits;
+    int long_rolling_actual_bits;
+
+    int64_t total_actual_bits;
+    int total_target_vs_actual; /* debug stats */
+
+    int worst_quality;
+    int active_worst_quality;
+    int best_quality;
+    int active_best_quality;
+
+    int cq_target_quality;
+
+    int drop_frames_allowed; /* Are we permitted to drop frames? */
+    int drop_frame;          /* Drop this frame? */
+
+    vp8_prob frame_coef_probs [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES];
+    char update_probs [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES];
+
+    unsigned int frame_branch_ct [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES][2];
+
+    int gfu_boost;
+    int kf_boost;
+    int last_boost;
+
+    int target_bandwidth;
+    struct vpx_codec_pkt_list  *output_pkt_list;
+
+#if 0
+    /* Experimental code for lagged and one pass */
+    ONEPASS_FRAMESTATS one_pass_frame_stats[MAX_LAG_BUFFERS];
+    int one_pass_frame_index;
+#endif
+
+    int decimation_factor;
+    int decimation_count;
+
+    /* for real time encoding */
+    int avg_encode_time;     /* microsecond */
+    int avg_pick_mode_time;  /* microsecond */
+    int Speed;
+    int compressor_speed;
+
+    int auto_gold;
+    int auto_adjust_gold_quantizer;
+    int auto_worst_q;
+    int cpu_used;
+    int pass;
+
+
+    int prob_intra_coded;
+    int prob_last_coded;
+    int prob_gf_coded;
+    int prob_skip_false;
+    int last_skip_false_probs[3];
+    int last_skip_probs_q[3];
+    int recent_ref_frame_usage[MAX_REF_FRAMES];
+
+    int this_frame_percent_intra;
+    int last_frame_percent_intra;
+
+    int ref_frame_flags;
+
+    SPEED_FEATURES sf;
+
+    /* Count ZEROMV on all reference frames. */
+    int zeromv_count;
+    int lf_zeromv_pct;
+
+    unsigned char *segmentation_map;
+    signed char segment_feature_data[MB_LVL_MAX][MAX_MB_SEGMENTS];
+    int  segment_encode_breakout[MAX_MB_SEGMENTS];
+
+    unsigned char *active_map;
+    unsigned int active_map_enabled;
+
+    /* Video conferencing cyclic refresh mode flags. This is a mode
+     * designed to clean up the background over time in live encoding
+     * scenarious. It uses segmentation.
+     */
+    int cyclic_refresh_mode_enabled;
+    int cyclic_refresh_mode_max_mbs_perframe;
+    int cyclic_refresh_mode_index;
+    int cyclic_refresh_q;
+    signed char *cyclic_refresh_map;
+    // Count on how many (consecutive) times a macroblock uses ZER0MV_LAST.
+    unsigned char *consec_zero_last;
+    // Counter that is reset when a block is checked for a mode-bias against
+    // ZEROMV_LASTREF.
+    unsigned char *consec_zero_last_mvbias;
+
+    // Frame counter for the temporal pattern. Counter is rest when the temporal
+    // layers are changed dynamically (run-time change).
+    unsigned int temporal_pattern_counter;
+    // Temporal layer id.
+    int temporal_layer_id;
+
+    // Measure of average squared difference between source and denoised signal.
+    int mse_source_denoised;
+
+    int force_maxqp;
+
+#if CONFIG_MULTITHREAD
+    /* multithread data */
+    pthread_mutex_t *pmutex;
+    pthread_mutex_t mt_mutex;           /* mutex for b_multi_threaded */
+    int * mt_current_mb_col;
+    int mt_sync_range;
+    int b_multi_threaded;
+    int encoding_thread_count;
+
+    pthread_t *h_encoding_thread;
+    pthread_t h_filter_thread;
+
+    MB_ROW_COMP *mb_row_ei;
+    ENCODETHREAD_DATA *en_thread_data;
+    LPFTHREAD_DATA lpf_thread_data;
+
+    /* events */
+    sem_t *h_event_start_encoding;
+    sem_t h_event_end_encoding;
+    sem_t h_event_start_lpf;
+    sem_t h_event_end_lpf;
+#endif
+
+    TOKENLIST *tplist;
+    unsigned int partition_sz[MAX_PARTITIONS];
+    unsigned char *partition_d[MAX_PARTITIONS];
+    unsigned char *partition_d_end[MAX_PARTITIONS];
+
+
+    fractional_mv_step_fp *find_fractional_mv_step;
+    vp8_full_search_fn_t full_search_sad;
+    vp8_refining_search_fn_t refining_search_sad;
+    vp8_diamond_search_fn_t diamond_search_sad;
+    vp8_variance_fn_ptr_t fn_ptr[BLOCK_MAX_SEGMENTS];
+    uint64_t time_receive_data;
+    uint64_t time_compress_data;
+    uint64_t time_pick_lpf;
+    uint64_t time_encode_mb_row;
+
+    int base_skip_false_prob[128];
+
+    FRAME_CONTEXT lfc_n; /* last frame entropy */
+    FRAME_CONTEXT lfc_a; /* last alt ref entropy */
+    FRAME_CONTEXT lfc_g; /* last gold ref entropy */
+
+
+    struct twopass_rc
+    {
+        unsigned int section_intra_rating;
+        double section_max_qfactor;
+        unsigned int next_iiratio;
+        unsigned int this_iiratio;
+        FIRSTPASS_STATS total_stats;
+        FIRSTPASS_STATS this_frame_stats;
+        FIRSTPASS_STATS *stats_in, *stats_in_end, *stats_in_start;
+        FIRSTPASS_STATS total_left_stats;
+        int first_pass_done;
+        int64_t bits_left;
+        int64_t clip_bits_total;
+        double avg_iiratio;
+        double modified_error_total;
+        double modified_error_used;
+        double modified_error_left;
+        double kf_intra_err_min;
+        double gf_intra_err_min;
+        int frames_to_key;
+        int maxq_max_limit;
+        int maxq_min_limit;
+        int gf_decay_rate;
+        int static_scene_max_gf_interval;
+        int kf_bits;
+        /* Remaining error from uncoded frames in a gf group. */
+        int gf_group_error_left;
+        /* Projected total bits available for a key frame group of frames */
+        int64_t kf_group_bits;
+        /* Error score of frames still to be coded in kf group */
+        int64_t kf_group_error_left;
+        /* Projected Bits available for a group including 1 GF or ARF */
+        int64_t gf_group_bits;
+        /* Bits for the golden frame or ARF */
+        int gf_bits;
+        int alt_extra_bits;
+        double est_max_qcorrection_factor;
+    } twopass;
+
+#if VP8_TEMPORAL_ALT_REF
+    YV12_BUFFER_CONFIG alt_ref_buffer;
+    YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS];
+    int fixed_divide[512];
+#endif
+
+#if CONFIG_INTERNAL_STATS
+    int    count;
+    double total_y;
+    double total_u;
+    double total_v;
+    double total ;
+    double total_sq_error;
+    double totalp_y;
+    double totalp_u;
+    double totalp_v;
+    double totalp;
+    double total_sq_error2;
+    int    bytes;
+    double summed_quality;
+    double summed_weights;
+    unsigned int tot_recode_hits;
+
+
+    double total_ssimg_y;
+    double total_ssimg_u;
+    double total_ssimg_v;
+    double total_ssimg_all;
+
+    int b_calculate_ssimg;
+#endif
+    int b_calculate_psnr;
+
+    /* Per MB activity measurement */
+    unsigned int activity_avg;
+    unsigned int * mb_activity_map;
+
+    /* Record of which MBs still refer to last golden frame either
+     * directly or through 0,0
+     */
+    unsigned char *gf_active_flags;
+    int gf_active_count;
+
+    int output_partition;
+
+    /* Store last frame's MV info for next frame MV prediction */
+    int_mv *lfmv;
+    int *lf_ref_frame_sign_bias;
+    int *lf_ref_frame;
+
+    /* force next frame to intra when kf_auto says so */
+    int force_next_frame_intra;
+
+    int droppable;
+
+    int initial_width;
+    int initial_height;
+
+#if CONFIG_TEMPORAL_DENOISING
+    VP8_DENOISER denoiser;
+#endif
+
+    /* Coding layer state variables */
+    unsigned int current_layer;
+    LAYER_CONTEXT layer_context[VPX_TS_MAX_LAYERS];
+
+    int64_t frames_in_layer[VPX_TS_MAX_LAYERS];
+    int64_t bytes_in_layer[VPX_TS_MAX_LAYERS];
+    double sum_psnr[VPX_TS_MAX_LAYERS];
+    double sum_psnr_p[VPX_TS_MAX_LAYERS];
+    double total_error2[VPX_TS_MAX_LAYERS];
+    double total_error2_p[VPX_TS_MAX_LAYERS];
+    double sum_ssim[VPX_TS_MAX_LAYERS];
+    double sum_weights[VPX_TS_MAX_LAYERS];
+
+    double total_ssimg_y_in_layer[VPX_TS_MAX_LAYERS];
+    double total_ssimg_u_in_layer[VPX_TS_MAX_LAYERS];
+    double total_ssimg_v_in_layer[VPX_TS_MAX_LAYERS];
+    double total_ssimg_all_in_layer[VPX_TS_MAX_LAYERS];
+
+#if CONFIG_MULTI_RES_ENCODING
+    /* Number of MBs per row at lower-resolution level */
+    int    mr_low_res_mb_cols;
+    /* Indicate if lower-res mv info is available */
+    unsigned char  mr_low_res_mv_avail;
+#endif
+    /* The frame number of each reference frames */
+    unsigned int current_ref_frames[MAX_REF_FRAMES];
+    // Closest reference frame to current frame.
+    MV_REFERENCE_FRAME closest_reference_frame;
+
+    struct rd_costs_struct
+    {
+        int mvcosts[2][MVvals+1];
+        int mvsadcosts[2][MVfpvals+1];
+        int mbmode_cost[2][MB_MODE_COUNT];
+        int intra_uv_mode_cost[2][MB_MODE_COUNT];
+        int bmode_costs[10][10][10];
+        int inter_bmode_costs[B_MODE_COUNT];
+        int token_costs[BLOCK_TYPES][COEF_BANDS]
+        [PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS];
+    } rd_costs;
+} VP8_COMP;
+
+void vp8_initialize_enc(void);
+
+void vp8_alloc_compressor_data(VP8_COMP *cpi);
+int vp8_reverse_trans(int x);
+void vp8_new_framerate(VP8_COMP *cpi, double framerate);
+void vp8_loopfilter_frame(VP8_COMP *cpi, VP8_COMMON *cm);
+
+void vp8_pack_bitstream(VP8_COMP *cpi, unsigned char *dest,
+                        unsigned char *dest_end, unsigned long *size);
+
+void vp8_tokenize_mb(VP8_COMP *, MACROBLOCK *, TOKENEXTRA **);
+
+void vp8_set_speed_features(VP8_COMP *cpi);
+
+#if CONFIG_DEBUG
+#define CHECK_MEM_ERROR(lval,expr) do {\
+        lval = (expr); \
+        if(!lval) \
+            vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,\
+                               "Failed to allocate "#lval" at %s:%d", \
+                               __FILE__,__LINE__);\
+    } while(0)
+#else
+#define CHECK_MEM_ERROR(lval,expr) do {\
+        lval = (expr); \
+        if(!lval) \
+            vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,\
+                               "Failed to allocate "#lval);\
+    } while(0)
+#endif
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_ONYX_INT_H_
diff --git a/libvpx/libvpx/vp8/encoder/pickinter.c b/libvpx/libvpx/vp8/encoder/pickinter.c
new file mode 100644
index 0000000..24b332d
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/pickinter.c
@@ -0,0 +1,1607 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <limits.h>
+#include "vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "onyx_int.h"
+#include "modecosts.h"
+#include "encodeintra.h"
+#include "vp8/common/common.h"
+#include "vp8/common/entropymode.h"
+#include "pickinter.h"
+#include "vp8/common/findnearmv.h"
+#include "encodemb.h"
+#include "vp8/common/reconinter.h"
+#include "vp8/common/reconintra.h"
+#include "vp8/common/reconintra4x4.h"
+#include "vpx_dsp/variance.h"
+#include "mcomp.h"
+#include "rdopt.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#if CONFIG_TEMPORAL_DENOISING
+#include "denoising.h"
+#endif
+
+#ifdef SPEEDSTATS
+extern unsigned int cnt_pm;
+#endif
+
+#define MODEL_MODE 1
+
+extern const int vp8_ref_frame_order[MAX_MODES];
+extern const MB_PREDICTION_MODE vp8_mode_order[MAX_MODES];
+
+// Fixed point implementation of a skin color classifier. Skin color
+// is model by a Gaussian distribution in the CbCr color space.
+// See ../../test/skin_color_detector_test.cc where the reference
+// skin color classifier is defined.
+
+// Fixed-point skin color model parameters.
+static const int skin_mean[5][2] =
+    {{7463, 9614}, {6400, 10240}, {7040, 10240}, {8320, 9280}, {6800, 9614}};
+static const int skin_inv_cov[4] = {4107, 1663, 1663, 2157};  // q16
+static const int skin_threshold[6] = {1570636, 1400000, 800000, 800000, 800000,
+    800000};  // q18
+
+// Evaluates the Mahalanobis distance measure for the input CbCr values.
+static int evaluate_skin_color_difference(int cb, int cr, int idx) {
+  const int cb_q6 = cb << 6;
+  const int cr_q6 = cr << 6;
+  const int cb_diff_q12 =
+      (cb_q6 - skin_mean[idx][0]) * (cb_q6 - skin_mean[idx][0]);
+  const int cbcr_diff_q12 =
+      (cb_q6 - skin_mean[idx][0]) * (cr_q6 - skin_mean[idx][1]);
+  const int cr_diff_q12 =
+      (cr_q6 - skin_mean[idx][1]) * (cr_q6 - skin_mean[idx][1]);
+  const int cb_diff_q2 = (cb_diff_q12 + (1 << 9)) >> 10;
+  const int cbcr_diff_q2 = (cbcr_diff_q12 + (1 << 9)) >> 10;
+  const int cr_diff_q2 = (cr_diff_q12 + (1 << 9)) >> 10;
+  const int skin_diff = skin_inv_cov[0] * cb_diff_q2 +
+      skin_inv_cov[1] * cbcr_diff_q2 +
+      skin_inv_cov[2] * cbcr_diff_q2 +
+      skin_inv_cov[3] * cr_diff_q2;
+  return skin_diff;
+}
+
+// Checks if the input yCbCr values corresponds to skin color.
+static int is_skin_color(int y, int cb, int cr, int consec_zeromv)
+{
+  if (y < 40 || y > 220)
+  {
+    return 0;
+  }
+  else
+  {
+    if (MODEL_MODE == 0)
+    {
+      return (evaluate_skin_color_difference(cb, cr, 0) < skin_threshold[0]);
+    }
+    else
+    {
+      int i = 0;
+      // No skin if block has been zero motion for long consecutive time.
+      if (consec_zeromv > 60)
+        return 0;
+      // Exit on grey.
+       if (cb == 128 && cr == 128)
+         return 0;
+       // Exit on very strong cb.
+       if (cb > 150 && cr < 110)
+         return 0;
+       for (; i < 5; i++) {
+         int skin_color_diff = evaluate_skin_color_difference(cb, cr, i);
+         if (skin_color_diff < skin_threshold[i + 1]) {
+            if (y < 60 && skin_color_diff > 3 * (skin_threshold[i + 1] >> 2))
+              return 0;
+            else if (consec_zeromv > 25 &&
+                     skin_color_diff > (skin_threshold[i + 1] >> 1))
+              return 0;
+            else
+             return 1;
+         }
+         // Exit if difference is much large than the threshold.
+         if (skin_color_diff > (skin_threshold[i + 1] << 3)) {
+           return 0;
+         }
+       }
+      return 0;
+    }
+  }
+}
+
+static int macroblock_corner_grad(unsigned char* signal, int stride,
+                                  int offsetx, int offsety, int sgnx, int sgny)
+{
+  int y1 = signal[offsetx * stride + offsety];
+  int y2 = signal[offsetx * stride + offsety + sgny];
+  int y3 = signal[(offsetx + sgnx) * stride + offsety];
+  int y4 = signal[(offsetx + sgnx) * stride + offsety + sgny];
+  return VPXMAX(VPXMAX(abs(y1 - y2), abs(y1 - y3)), abs(y1 - y4));
+}
+
+static int check_dot_artifact_candidate(VP8_COMP *cpi,
+                                        MACROBLOCK *x,
+                                        unsigned char *target_last,
+                                        int stride,
+                                        unsigned char* last_ref,
+                                        int mb_row,
+                                        int mb_col,
+                                        int channel)
+{
+  int threshold1 = 6;
+  int threshold2 = 3;
+  unsigned int max_num = (cpi->common.MBs) / 10;
+  int grad_last = 0;
+  int grad_source = 0;
+  int index = mb_row * cpi->common.mb_cols + mb_col;
+  // Threshold for #consecutive (base layer) frames using zero_last mode.
+  int num_frames = 30;
+  int shift = 15;
+  if (channel > 0) {
+    shift = 7;
+  }
+  if (cpi->oxcf.number_of_layers > 1)
+  {
+    num_frames = 20;
+  }
+  x->zero_last_dot_suppress = 0;
+  // Blocks on base layer frames that have been using ZEROMV_LAST repeatedly
+  // (i.e, at least |x| consecutive frames are candidates for increasing the
+  // rd adjustment for zero_last mode.
+  // Only allow this for at most |max_num| blocks per frame.
+  // Don't allow this for screen content input.
+  if (cpi->current_layer == 0 &&
+      cpi->consec_zero_last_mvbias[index] > num_frames &&
+      x->mbs_zero_last_dot_suppress < max_num &&
+      !cpi->oxcf.screen_content_mode)
+  {
+    // If this block is checked here, label it so we don't check it again until
+    // ~|x| framaes later.
+    x->zero_last_dot_suppress = 1;
+    // Dot artifact is noticeable as strong gradient at corners of macroblock,
+    // for flat areas. As a simple detector for now, we look for a high
+    // corner gradient on last ref, and a smaller gradient on source.
+    // Check 4 corners, return if any satisfy condition.
+    // Top-left:
+    grad_last = macroblock_corner_grad(last_ref, stride, 0, 0, 1, 1);
+    grad_source = macroblock_corner_grad(target_last, stride, 0, 0, 1, 1);
+    if (grad_last >= threshold1 && grad_source <= threshold2)
+    {
+       x->mbs_zero_last_dot_suppress++;
+       return 1;
+    }
+    // Top-right:
+    grad_last = macroblock_corner_grad(last_ref, stride, 0, shift, 1, -1);
+    grad_source = macroblock_corner_grad(target_last, stride, 0, shift, 1, -1);
+    if (grad_last >= threshold1 && grad_source <= threshold2)
+    {
+      x->mbs_zero_last_dot_suppress++;
+      return 1;
+    }
+    // Bottom-left:
+    grad_last = macroblock_corner_grad(last_ref, stride, shift, 0, -1, 1);
+    grad_source = macroblock_corner_grad(target_last, stride, shift, 0, -1, 1);
+    if (grad_last >= threshold1 && grad_source <= threshold2)
+    {
+      x->mbs_zero_last_dot_suppress++;
+      return 1;
+    }
+    // Bottom-right:
+    grad_last = macroblock_corner_grad(last_ref, stride, shift, shift, -1, -1);
+    grad_source = macroblock_corner_grad(target_last, stride, shift, shift, -1, -1);
+    if (grad_last >= threshold1 && grad_source <= threshold2)
+    {
+      x->mbs_zero_last_dot_suppress++;
+      return 1;
+    }
+    return 0;
+  }
+  return 0;
+}
+
+int vp8_skip_fractional_mv_step(MACROBLOCK *mb, BLOCK *b, BLOCKD *d,
+                                int_mv *bestmv, int_mv *ref_mv,
+                                int error_per_bit,
+                                const vp8_variance_fn_ptr_t *vfp,
+                                int *mvcost[2], int *distortion,
+                                unsigned int *sse)
+{
+    (void) b;
+    (void) d;
+    (void) ref_mv;
+    (void) error_per_bit;
+    (void) vfp;
+    (void) mb;
+    (void) mvcost;
+    (void) distortion;
+    (void) sse;
+    bestmv->as_mv.row <<= 3;
+    bestmv->as_mv.col <<= 3;
+    return 0;
+}
+
+
+int vp8_get_inter_mbpred_error(MACROBLOCK *mb,
+                                  const vp8_variance_fn_ptr_t *vfp,
+                                  unsigned int *sse,
+                                  int_mv this_mv)
+{
+
+    BLOCK *b = &mb->block[0];
+    BLOCKD *d = &mb->e_mbd.block[0];
+    unsigned char *what = (*(b->base_src) + b->src);
+    int what_stride = b->src_stride;
+    int pre_stride = mb->e_mbd.pre.y_stride;
+    unsigned char *in_what = mb->e_mbd.pre.y_buffer + d->offset ;
+    int in_what_stride = pre_stride;
+    int xoffset = this_mv.as_mv.col & 7;
+    int yoffset = this_mv.as_mv.row & 7;
+
+    in_what += (this_mv.as_mv.row >> 3) * pre_stride + (this_mv.as_mv.col >> 3);
+
+    if (xoffset | yoffset)
+    {
+        return vfp->svf(in_what, in_what_stride, xoffset, yoffset, what, what_stride, sse);
+    }
+    else
+    {
+        return vfp->vf(what, what_stride, in_what, in_what_stride, sse);
+    }
+
+}
+
+static int get_prediction_error(BLOCK *be, BLOCKD *b)
+{
+    unsigned char *sptr;
+    unsigned char *dptr;
+    sptr = (*(be->base_src) + be->src);
+    dptr = b->predictor;
+
+    return vpx_get4x4sse_cs(sptr, be->src_stride, dptr, 16);
+
+}
+
+static int pick_intra4x4block(
+    MACROBLOCK *x,
+    int ib,
+    B_PREDICTION_MODE *best_mode,
+    const int *mode_costs,
+
+    int *bestrate,
+    int *bestdistortion)
+{
+
+    BLOCKD *b = &x->e_mbd.block[ib];
+    BLOCK *be = &x->block[ib];
+    int dst_stride = x->e_mbd.dst.y_stride;
+    unsigned char *dst = x->e_mbd.dst.y_buffer + b->offset;
+    B_PREDICTION_MODE mode;
+    int best_rd = INT_MAX;
+    int rate;
+    int distortion;
+
+    unsigned char *Above = dst - dst_stride;
+    unsigned char *yleft = dst - 1;
+    unsigned char top_left = Above[-1];
+
+    for (mode = B_DC_PRED; mode <= B_HE_PRED; mode++)
+    {
+        int this_rd;
+
+        rate = mode_costs[mode];
+
+        vp8_intra4x4_predict(Above, yleft, dst_stride, mode,
+                             b->predictor, 16, top_left);
+        distortion = get_prediction_error(be, b);
+        this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
+
+        if (this_rd < best_rd)
+        {
+            *bestrate = rate;
+            *bestdistortion = distortion;
+            best_rd = this_rd;
+            *best_mode = mode;
+        }
+    }
+
+    b->bmi.as_mode = *best_mode;
+    vp8_encode_intra4x4block(x, ib);
+    return best_rd;
+}
+
+
+static int pick_intra4x4mby_modes
+(
+    MACROBLOCK *mb,
+    int *Rate,
+    int *best_dist
+)
+{
+    MACROBLOCKD *const xd = &mb->e_mbd;
+    int i;
+    int cost = mb->mbmode_cost [xd->frame_type] [B_PRED];
+    int error;
+    int distortion = 0;
+    const int *bmode_costs;
+
+    intra_prediction_down_copy(xd, xd->dst.y_buffer - xd->dst.y_stride + 16);
+
+    bmode_costs = mb->inter_bmode_costs;
+
+    for (i = 0; i < 16; i++)
+    {
+        MODE_INFO *const mic = xd->mode_info_context;
+        const int mis = xd->mode_info_stride;
+
+        B_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode);
+        int UNINITIALIZED_IS_SAFE(r), UNINITIALIZED_IS_SAFE(d);
+
+        if (mb->e_mbd.frame_type == KEY_FRAME)
+        {
+            const B_PREDICTION_MODE A = above_block_mode(mic, i, mis);
+            const B_PREDICTION_MODE L = left_block_mode(mic, i);
+
+            bmode_costs  = mb->bmode_costs[A][L];
+        }
+
+
+        pick_intra4x4block(mb, i, &best_mode, bmode_costs, &r, &d);
+
+        cost += r;
+        distortion += d;
+        mic->bmi[i].as_mode = best_mode;
+
+        /* Break out case where we have already exceeded best so far value
+         * that was passed in
+         */
+        if (distortion > *best_dist)
+            break;
+    }
+
+    *Rate = cost;
+
+    if (i == 16)
+    {
+        *best_dist = distortion;
+        error = RDCOST(mb->rdmult, mb->rddiv, cost, distortion);
+    }
+    else
+    {
+        *best_dist = INT_MAX;
+        error = INT_MAX;
+    }
+
+    return error;
+}
+
+static void pick_intra_mbuv_mode(MACROBLOCK *mb)
+{
+
+    MACROBLOCKD *x = &mb->e_mbd;
+    unsigned char *uabove_row = x->dst.u_buffer - x->dst.uv_stride;
+    unsigned char *vabove_row = x->dst.v_buffer - x->dst.uv_stride;
+    unsigned char *usrc_ptr = (mb->block[16].src + *mb->block[16].base_src);
+    unsigned char *vsrc_ptr = (mb->block[20].src + *mb->block[20].base_src);
+    int uvsrc_stride = mb->block[16].src_stride;
+    unsigned char uleft_col[8];
+    unsigned char vleft_col[8];
+    unsigned char utop_left = uabove_row[-1];
+    unsigned char vtop_left = vabove_row[-1];
+    int i, j;
+    int expected_udc;
+    int expected_vdc;
+    int shift;
+    int Uaverage = 0;
+    int Vaverage = 0;
+    int diff;
+    int pred_error[4] = {0, 0, 0, 0}, best_error = INT_MAX;
+    MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode);
+
+
+    for (i = 0; i < 8; i++)
+    {
+        uleft_col[i] = x->dst.u_buffer [i* x->dst.uv_stride -1];
+        vleft_col[i] = x->dst.v_buffer [i* x->dst.uv_stride -1];
+    }
+
+    if (!x->up_available && !x->left_available)
+    {
+        expected_udc = 128;
+        expected_vdc = 128;
+    }
+    else
+    {
+        shift = 2;
+
+        if (x->up_available)
+        {
+
+            for (i = 0; i < 8; i++)
+            {
+                Uaverage += uabove_row[i];
+                Vaverage += vabove_row[i];
+            }
+
+            shift ++;
+
+        }
+
+        if (x->left_available)
+        {
+            for (i = 0; i < 8; i++)
+            {
+                Uaverage += uleft_col[i];
+                Vaverage += vleft_col[i];
+            }
+
+            shift ++;
+
+        }
+
+        expected_udc = (Uaverage + (1 << (shift - 1))) >> shift;
+        expected_vdc = (Vaverage + (1 << (shift - 1))) >> shift;
+    }
+
+
+    for (i = 0; i < 8; i++)
+    {
+        for (j = 0; j < 8; j++)
+        {
+
+            int predu = uleft_col[i] + uabove_row[j] - utop_left;
+            int predv = vleft_col[i] + vabove_row[j] - vtop_left;
+            int u_p, v_p;
+
+            u_p = usrc_ptr[j];
+            v_p = vsrc_ptr[j];
+
+            if (predu < 0)
+                predu = 0;
+
+            if (predu > 255)
+                predu = 255;
+
+            if (predv < 0)
+                predv = 0;
+
+            if (predv > 255)
+                predv = 255;
+
+
+            diff = u_p - expected_udc;
+            pred_error[DC_PRED] += diff * diff;
+            diff = v_p - expected_vdc;
+            pred_error[DC_PRED] += diff * diff;
+
+
+            diff = u_p - uabove_row[j];
+            pred_error[V_PRED] += diff * diff;
+            diff = v_p - vabove_row[j];
+            pred_error[V_PRED] += diff * diff;
+
+
+            diff = u_p - uleft_col[i];
+            pred_error[H_PRED] += diff * diff;
+            diff = v_p - vleft_col[i];
+            pred_error[H_PRED] += diff * diff;
+
+
+            diff = u_p - predu;
+            pred_error[TM_PRED] += diff * diff;
+            diff = v_p - predv;
+            pred_error[TM_PRED] += diff * diff;
+
+
+        }
+
+        usrc_ptr += uvsrc_stride;
+        vsrc_ptr += uvsrc_stride;
+
+        if (i == 3)
+        {
+            usrc_ptr = (mb->block[18].src + *mb->block[18].base_src);
+            vsrc_ptr = (mb->block[22].src + *mb->block[22].base_src);
+        }
+
+
+
+    }
+
+
+    for (i = DC_PRED; i <= TM_PRED; i++)
+    {
+        if (best_error > pred_error[i])
+        {
+            best_error = pred_error[i];
+            best_mode = (MB_PREDICTION_MODE)i;
+        }
+    }
+
+
+    mb->e_mbd.mode_info_context->mbmi.uv_mode = best_mode;
+
+}
+
+static void update_mvcount(MACROBLOCK *x, int_mv *best_ref_mv)
+{
+    MACROBLOCKD *xd = &x->e_mbd;
+    /* Split MV modes currently not supported when RD is nopt enabled,
+     * therefore, only need to modify MVcount in NEWMV mode. */
+    if (xd->mode_info_context->mbmi.mode == NEWMV)
+    {
+        x->MVcount[0][mv_max+((xd->mode_info_context->mbmi.mv.as_mv.row -
+                                      best_ref_mv->as_mv.row) >> 1)]++;
+        x->MVcount[1][mv_max+((xd->mode_info_context->mbmi.mv.as_mv.col -
+                                      best_ref_mv->as_mv.col) >> 1)]++;
+    }
+}
+
+
+#if CONFIG_MULTI_RES_ENCODING
+static
+void get_lower_res_motion_info(VP8_COMP *cpi, MACROBLOCKD *xd, int *dissim,
+                               int *parent_ref_frame,
+                               MB_PREDICTION_MODE *parent_mode,
+                               int_mv *parent_ref_mv, int mb_row, int mb_col)
+{
+    LOWER_RES_MB_INFO* store_mode_info
+                          = ((LOWER_RES_FRAME_INFO*)cpi->oxcf.mr_low_res_mode_info)->mb_info;
+    unsigned int parent_mb_index;
+
+    /* Consider different down_sampling_factor.  */
+    {
+        /* TODO: Removed the loop that supports special down_sampling_factor
+         * such as 2, 4, 8. Will revisit it if needed.
+         * Should also try using a look-up table to see if it helps
+         * performance. */
+        int parent_mb_row, parent_mb_col;
+
+        parent_mb_row = mb_row*cpi->oxcf.mr_down_sampling_factor.den
+                    /cpi->oxcf.mr_down_sampling_factor.num;
+        parent_mb_col = mb_col*cpi->oxcf.mr_down_sampling_factor.den
+                    /cpi->oxcf.mr_down_sampling_factor.num;
+        parent_mb_index = parent_mb_row*cpi->mr_low_res_mb_cols + parent_mb_col;
+    }
+
+    /* Read lower-resolution mode & motion result from memory.*/
+    *parent_ref_frame = store_mode_info[parent_mb_index].ref_frame;
+    *parent_mode =  store_mode_info[parent_mb_index].mode;
+    *dissim = store_mode_info[parent_mb_index].dissim;
+
+    /* For highest-resolution encoder, adjust dissim value. Lower its quality
+     * for good performance. */
+    if (cpi->oxcf.mr_encoder_id == (cpi->oxcf.mr_total_resolutions - 1))
+        *dissim>>=1;
+
+    if(*parent_ref_frame != INTRA_FRAME)
+    {
+        /* Consider different down_sampling_factor.
+         * The result can be rounded to be more precise, but it takes more time.
+         */
+        (*parent_ref_mv).as_mv.row = store_mode_info[parent_mb_index].mv.as_mv.row
+                                  *cpi->oxcf.mr_down_sampling_factor.num
+                                  /cpi->oxcf.mr_down_sampling_factor.den;
+        (*parent_ref_mv).as_mv.col = store_mode_info[parent_mb_index].mv.as_mv.col
+                                  *cpi->oxcf.mr_down_sampling_factor.num
+                                  /cpi->oxcf.mr_down_sampling_factor.den;
+
+        vp8_clamp_mv2(parent_ref_mv, xd);
+    }
+}
+#endif
+
+static void check_for_encode_breakout(unsigned int sse, MACROBLOCK* x)
+{
+    MACROBLOCKD *xd = &x->e_mbd;
+
+    unsigned int threshold = (xd->block[0].dequant[1]
+        * xd->block[0].dequant[1] >>4);
+
+    if(threshold < x->encode_breakout)
+        threshold = x->encode_breakout;
+
+    if (sse < threshold )
+    {
+        /* Check u and v to make sure skip is ok */
+        unsigned int sse2 = 0;
+
+        sse2 = VP8_UVSSE(x);
+
+        if (sse2 * 2 < x->encode_breakout)
+            x->skip = 1;
+        else
+            x->skip = 0;
+    }
+}
+
+static int evaluate_inter_mode(unsigned int* sse, int rate2, int* distortion2,
+                               VP8_COMP *cpi, MACROBLOCK *x, int rd_adj)
+{
+    MB_PREDICTION_MODE this_mode = x->e_mbd.mode_info_context->mbmi.mode;
+    int_mv mv = x->e_mbd.mode_info_context->mbmi.mv;
+    int this_rd;
+    int denoise_aggressive = 0;
+    /* Exit early and don't compute the distortion if this macroblock
+     * is marked inactive. */
+    if (cpi->active_map_enabled && x->active_ptr[0] == 0)
+    {
+        *sse = 0;
+        *distortion2 = 0;
+        x->skip = 1;
+        return INT_MAX;
+    }
+
+    if((this_mode != NEWMV) ||
+        !(cpi->sf.half_pixel_search) || cpi->common.full_pixel==1)
+        *distortion2 = vp8_get_inter_mbpred_error(x,
+                                              &cpi->fn_ptr[BLOCK_16X16],
+                                              sse, mv);
+
+    this_rd = RDCOST(x->rdmult, x->rddiv, rate2, *distortion2);
+
+#if CONFIG_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity > 0) {
+      denoise_aggressive =
+        (cpi->denoiser.denoiser_mode == kDenoiserOnYUVAggressive) ? 1 : 0;
+    }
+#endif
+
+    // Adjust rd for ZEROMV and LAST, if LAST is the closest reference frame.
+    // TODO: We should also add condition on distance of closest to current.
+    if(!cpi->oxcf.screen_content_mode &&
+       this_mode == ZEROMV &&
+       x->e_mbd.mode_info_context->mbmi.ref_frame == LAST_FRAME &&
+       (denoise_aggressive || (cpi->closest_reference_frame == LAST_FRAME)))
+    {
+        // No adjustment if block is considered to be skin area.
+        if(x->is_skin)
+            rd_adj = 100;
+
+        this_rd = ((int64_t)this_rd) * rd_adj / 100;
+    }
+
+    check_for_encode_breakout(*sse, x);
+    return this_rd;
+}
+
+static void calculate_zeromv_rd_adjustment(VP8_COMP *cpi, MACROBLOCK *x,
+                                    int *rd_adjustment)
+{
+    MODE_INFO *mic = x->e_mbd.mode_info_context;
+    int_mv mv_l, mv_a, mv_al;
+    int local_motion_check = 0;
+
+    if (cpi->lf_zeromv_pct > 40)
+    {
+        /* left mb */
+        mic -= 1;
+        mv_l = mic->mbmi.mv;
+
+        if (mic->mbmi.ref_frame != INTRA_FRAME)
+            if( abs(mv_l.as_mv.row) < 8 && abs(mv_l.as_mv.col) < 8)
+                local_motion_check++;
+
+        /* above-left mb */
+        mic -= x->e_mbd.mode_info_stride;
+        mv_al = mic->mbmi.mv;
+
+        if (mic->mbmi.ref_frame != INTRA_FRAME)
+            if( abs(mv_al.as_mv.row) < 8 && abs(mv_al.as_mv.col) < 8)
+                local_motion_check++;
+
+        /* above mb */
+        mic += 1;
+        mv_a = mic->mbmi.mv;
+
+        if (mic->mbmi.ref_frame != INTRA_FRAME)
+            if( abs(mv_a.as_mv.row) < 8 && abs(mv_a.as_mv.col) < 8)
+                local_motion_check++;
+
+        if (((!x->e_mbd.mb_to_top_edge || !x->e_mbd.mb_to_left_edge)
+            && local_motion_check >0) ||  local_motion_check >2 )
+            *rd_adjustment = 80;
+        else if (local_motion_check > 0)
+            *rd_adjustment = 90;
+    }
+}
+
+void vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
+                         int recon_uvoffset, int *returnrate,
+                         int *returndistortion, int *returnintra, int mb_row,
+                         int mb_col)
+{
+    BLOCK *b = &x->block[0];
+    BLOCKD *d = &x->e_mbd.block[0];
+    MACROBLOCKD *xd = &x->e_mbd;
+    MB_MODE_INFO best_mbmode;
+
+    int_mv best_ref_mv_sb[2];
+    int_mv mode_mv_sb[2][MB_MODE_COUNT];
+    int_mv best_ref_mv;
+    int_mv *mode_mv;
+    MB_PREDICTION_MODE this_mode;
+    int num00;
+    int mdcounts[4];
+    int best_rd = INT_MAX;
+    int rd_adjustment = 100;
+    int best_intra_rd = INT_MAX;
+    int mode_index;
+    int rate;
+    int rate2;
+    int distortion2;
+    int bestsme = INT_MAX;
+    int best_mode_index = 0;
+    unsigned int sse = UINT_MAX, best_rd_sse = UINT_MAX;
+#if CONFIG_TEMPORAL_DENOISING
+    unsigned int zero_mv_sse = UINT_MAX, best_sse = UINT_MAX;
+#endif
+
+    int sf_improved_mv_pred = cpi->sf.improved_mv_pred;
+
+#if CONFIG_MULTI_RES_ENCODING
+    int dissim = INT_MAX;
+    int parent_ref_frame = 0;
+    int_mv parent_ref_mv;
+    MB_PREDICTION_MODE parent_mode = 0;
+    int parent_ref_valid = 0;
+#endif
+
+    int_mv mvp;
+
+    int near_sadidx[8] = {0, 1, 2, 3, 4, 5, 6, 7};
+    int saddone=0;
+    /* search range got from mv_pred(). It uses step_param levels. (0-7) */
+    int sr=0;
+
+    unsigned char *plane[4][3];
+    int ref_frame_map[4];
+    int sign_bias = 0;
+    int dot_artifact_candidate = 0;
+    get_predictor_pointers(cpi, plane, recon_yoffset, recon_uvoffset);
+
+    // If the current frame is using LAST as a reference, check for
+    // biasing the mode selection for dot artifacts.
+    if (cpi->ref_frame_flags & VP8_LAST_FRAME) {
+      unsigned char* target_y = x->src.y_buffer;
+      unsigned char* target_u = x->block[16].src + *x->block[16].base_src;
+      unsigned char* target_v = x->block[20].src + *x->block[20].base_src;
+      int stride = x->src.y_stride;
+      int stride_uv = x->block[16].src_stride;
+#if CONFIG_TEMPORAL_DENOISING
+      if (cpi->oxcf.noise_sensitivity) {
+        const int uv_denoise = (cpi->oxcf.noise_sensitivity >= 2) ? 1 : 0;
+        target_y =
+            cpi->denoiser.yv12_running_avg[LAST_FRAME].y_buffer + recon_yoffset;
+        stride = cpi->denoiser.yv12_running_avg[LAST_FRAME].y_stride;
+        if (uv_denoise) {
+          target_u =
+              cpi->denoiser.yv12_running_avg[LAST_FRAME].u_buffer +
+                  recon_uvoffset;
+          target_v =
+              cpi->denoiser.yv12_running_avg[LAST_FRAME].v_buffer +
+                  recon_uvoffset;
+          stride_uv = cpi->denoiser.yv12_running_avg[LAST_FRAME].uv_stride;
+        }
+      }
+#endif
+      dot_artifact_candidate =
+          check_dot_artifact_candidate(cpi, x, target_y, stride,
+              plane[LAST_FRAME][0], mb_row, mb_col, 0);
+      // If not found in Y channel, check UV channel.
+      if (!dot_artifact_candidate) {
+        dot_artifact_candidate =
+            check_dot_artifact_candidate(cpi, x, target_u, stride_uv,
+                plane[LAST_FRAME][1], mb_row, mb_col, 1);
+        if (!dot_artifact_candidate) {
+          dot_artifact_candidate =
+              check_dot_artifact_candidate(cpi, x, target_v, stride_uv,
+                  plane[LAST_FRAME][2], mb_row, mb_col, 2);
+        }
+      }
+    }
+
+#if CONFIG_MULTI_RES_ENCODING
+    // |parent_ref_valid| will be set here if potentially we can do mv resue for
+    // this higher resol (|cpi->oxcf.mr_encoder_id| > 0) frame.
+    // |parent_ref_valid| may be reset depending on |parent_ref_frame| for
+    // the current macroblock below.
+    parent_ref_valid = cpi->oxcf.mr_encoder_id && cpi->mr_low_res_mv_avail;
+    if (parent_ref_valid)
+    {
+        int parent_ref_flag;
+
+        get_lower_res_motion_info(cpi, xd, &dissim, &parent_ref_frame,
+                                  &parent_mode, &parent_ref_mv, mb_row, mb_col);
+
+        /* TODO(jkoleszar): The references available (ref_frame_flags) to the
+         * lower res encoder should match those available to this encoder, but
+         * there seems to be a situation where this mismatch can happen in the
+         * case of frame dropping and temporal layers. For example,
+         * GOLD being disallowed in ref_frame_flags, but being returned as
+         * parent_ref_frame.
+         *
+         * In this event, take the conservative approach of disabling the
+         * lower res info for this MB.
+         */
+
+        parent_ref_flag = 0;
+        // Note availability for mv reuse is only based on last and golden.
+        if (parent_ref_frame == LAST_FRAME)
+            parent_ref_flag = (cpi->ref_frame_flags & VP8_LAST_FRAME);
+        else if (parent_ref_frame == GOLDEN_FRAME)
+            parent_ref_flag = (cpi->ref_frame_flags & VP8_GOLD_FRAME);
+
+        //assert(!parent_ref_frame || parent_ref_flag);
+
+        // If |parent_ref_frame| did not match either last or golden then
+        // shut off mv reuse.
+        if (parent_ref_frame && !parent_ref_flag)
+            parent_ref_valid = 0;
+
+        // Don't do mv reuse since we want to allow for another mode besides
+        // ZEROMV_LAST to remove dot artifact.
+        if (dot_artifact_candidate)
+          parent_ref_valid = 0;
+    }
+#endif
+
+    // Check if current macroblock is in skin area.
+    {
+    const int y = (x->src.y_buffer[7 * x->src.y_stride + 7] +
+        x->src.y_buffer[7 * x->src.y_stride + 8] +
+        x->src.y_buffer[8 * x->src.y_stride + 7] +
+        x->src.y_buffer[8 * x->src.y_stride + 8]) >> 2;
+    const int cb = (x->src.u_buffer[3 * x->src.uv_stride + 3] +
+        x->src.u_buffer[3 * x->src.uv_stride + 4] +
+        x->src.u_buffer[4 * x->src.uv_stride + 3] +
+        x->src.u_buffer[4 * x->src.uv_stride + 4]) >> 2;
+    const int cr = (x->src.v_buffer[3 * x->src.uv_stride + 3] +
+        x->src.v_buffer[3 * x->src.uv_stride + 4] +
+        x->src.v_buffer[4 * x->src.uv_stride + 3] +
+        x->src.v_buffer[4 * x->src.uv_stride + 4]) >> 2;
+    x->is_skin = 0;
+    if (!cpi->oxcf.screen_content_mode) {
+      int block_index = mb_row * cpi->common.mb_cols + mb_col;
+      x->is_skin = is_skin_color(y, cb, cr, cpi->consec_zero_last[block_index]);
+    }
+    }
+#if CONFIG_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity) {
+      // Under aggressive denoising mode, should we use skin map to reduce denoiser
+      // and ZEROMV bias? Will need to revisit the accuracy of this detection for
+      // very noisy input. For now keep this as is (i.e., don't turn it off).
+      // if (cpi->denoiser.denoiser_mode == kDenoiserOnYUVAggressive)
+      //   x->is_skin = 0;
+    }
+#endif
+
+    mode_mv = mode_mv_sb[sign_bias];
+    best_ref_mv.as_int = 0;
+    memset(mode_mv_sb, 0, sizeof(mode_mv_sb));
+    memset(&best_mbmode, 0, sizeof(best_mbmode));
+
+    /* Setup search priorities */
+#if CONFIG_MULTI_RES_ENCODING
+    if (parent_ref_valid && parent_ref_frame && dissim < 8)
+    {
+        ref_frame_map[0] = -1;
+        ref_frame_map[1] = parent_ref_frame;
+        ref_frame_map[2] = -1;
+        ref_frame_map[3] = -1;
+    } else
+#endif
+    get_reference_search_order(cpi, ref_frame_map);
+
+    /* Check to see if there is at least 1 valid reference frame that we need
+     * to calculate near_mvs.
+     */
+    if (ref_frame_map[1] > 0)
+    {
+        sign_bias = vp8_find_near_mvs_bias(&x->e_mbd,
+                                           x->e_mbd.mode_info_context,
+                                           mode_mv_sb,
+                                           best_ref_mv_sb,
+                                           mdcounts,
+                                           ref_frame_map[1],
+                                           cpi->common.ref_frame_sign_bias);
+
+        mode_mv = mode_mv_sb[sign_bias];
+        best_ref_mv.as_int = best_ref_mv_sb[sign_bias].as_int;
+    }
+
+    /* Count of the number of MBs tested so far this frame */
+    x->mbs_tested_so_far++;
+
+    *returnintra = INT_MAX;
+    x->skip = 0;
+
+    x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME;
+
+    /* If the frame has big static background and current MB is in low
+    *  motion area, its mode decision is biased to ZEROMV mode.
+    *  No adjustment if cpu_used is <= -12 (i.e., cpi->Speed >= 12).
+    *  At such speed settings, ZEROMV is already heavily favored.
+    */
+    if (cpi->Speed < 12) {
+      calculate_zeromv_rd_adjustment(cpi, x, &rd_adjustment);
+    }
+
+#if CONFIG_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity) {
+      rd_adjustment = (int)(rd_adjustment *
+          cpi->denoiser.denoise_pars.pickmode_mv_bias / 100);
+    }
+#endif
+
+    if (dot_artifact_candidate)
+    {
+        // Bias against ZEROMV_LAST mode.
+        rd_adjustment = 150;
+    }
+
+
+    /* if we encode a new mv this is important
+     * find the best new motion vector
+     */
+    for (mode_index = 0; mode_index < MAX_MODES; mode_index++)
+    {
+        int frame_cost;
+        int this_rd = INT_MAX;
+        int this_ref_frame = ref_frame_map[vp8_ref_frame_order[mode_index]];
+
+        if (best_rd <= x->rd_threshes[mode_index])
+            continue;
+
+        if (this_ref_frame < 0)
+            continue;
+
+        x->e_mbd.mode_info_context->mbmi.ref_frame = this_ref_frame;
+
+        /* everything but intra */
+        if (x->e_mbd.mode_info_context->mbmi.ref_frame)
+        {
+            x->e_mbd.pre.y_buffer = plane[this_ref_frame][0];
+            x->e_mbd.pre.u_buffer = plane[this_ref_frame][1];
+            x->e_mbd.pre.v_buffer = plane[this_ref_frame][2];
+
+            if (sign_bias != cpi->common.ref_frame_sign_bias[this_ref_frame])
+            {
+                sign_bias = cpi->common.ref_frame_sign_bias[this_ref_frame];
+                mode_mv = mode_mv_sb[sign_bias];
+                best_ref_mv.as_int = best_ref_mv_sb[sign_bias].as_int;
+            }
+
+#if CONFIG_MULTI_RES_ENCODING
+            if (parent_ref_valid)
+            {
+                if (vp8_mode_order[mode_index] == NEARESTMV &&
+                    mode_mv[NEARESTMV].as_int ==0)
+                    continue;
+                if (vp8_mode_order[mode_index] == NEARMV &&
+                    mode_mv[NEARMV].as_int ==0)
+                    continue;
+
+                if (vp8_mode_order[mode_index] == NEWMV && parent_mode == ZEROMV
+                    && best_ref_mv.as_int==0)
+                    continue;
+                else if(vp8_mode_order[mode_index] == NEWMV && dissim==0
+                    && best_ref_mv.as_int==parent_ref_mv.as_int)
+                    continue;
+            }
+#endif
+        }
+
+        /* Check to see if the testing frequency for this mode is at its max
+         * If so then prevent it from being tested and increase the threshold
+         * for its testing */
+        if (x->mode_test_hit_counts[mode_index] &&
+                                         (cpi->mode_check_freq[mode_index] > 1))
+        {
+            if (x->mbs_tested_so_far <= (cpi->mode_check_freq[mode_index] *
+                                         x->mode_test_hit_counts[mode_index]))
+            {
+                /* Increase the threshold for coding this mode to make it less
+                 * likely to be chosen */
+                x->rd_thresh_mult[mode_index] += 4;
+
+                if (x->rd_thresh_mult[mode_index] > MAX_THRESHMULT)
+                    x->rd_thresh_mult[mode_index] = MAX_THRESHMULT;
+
+                x->rd_threshes[mode_index] =
+                                 (cpi->rd_baseline_thresh[mode_index] >> 7) *
+                                 x->rd_thresh_mult[mode_index];
+                continue;
+            }
+        }
+
+        /* We have now reached the point where we are going to test the current
+         * mode so increment the counter for the number of times it has been
+         * tested */
+        x->mode_test_hit_counts[mode_index] ++;
+
+        rate2 = 0;
+        distortion2 = 0;
+
+        this_mode = vp8_mode_order[mode_index];
+
+        x->e_mbd.mode_info_context->mbmi.mode = this_mode;
+        x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
+
+        /* Work out the cost assosciated with selecting the reference frame */
+        frame_cost =
+            x->ref_frame_cost[x->e_mbd.mode_info_context->mbmi.ref_frame];
+        rate2 += frame_cost;
+
+        /* Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
+         * unless ARNR filtering is enabled in which case we want
+         * an unfiltered alternative */
+        if (cpi->is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
+        {
+            if (this_mode != ZEROMV ||
+                x->e_mbd.mode_info_context->mbmi.ref_frame != ALTREF_FRAME)
+                continue;
+        }
+
+        switch (this_mode)
+        {
+        case B_PRED:
+            /* Pass best so far to pick_intra4x4mby_modes to use as breakout */
+            distortion2 = best_rd_sse;
+            pick_intra4x4mby_modes(x, &rate, &distortion2);
+
+            if (distortion2 == INT_MAX)
+            {
+                this_rd = INT_MAX;
+            }
+            else
+            {
+                rate2 += rate;
+                distortion2 = vpx_variance16x16(
+                                    *(b->base_src), b->src_stride,
+                                    x->e_mbd.predictor, 16, &sse);
+                this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+
+                if (this_rd < best_intra_rd)
+                {
+                    best_intra_rd = this_rd;
+                    *returnintra = distortion2;
+                }
+            }
+
+            break;
+
+        case SPLITMV:
+
+            /* Split MV modes currently not supported when RD is not enabled. */
+            break;
+
+        case DC_PRED:
+        case V_PRED:
+        case H_PRED:
+        case TM_PRED:
+            vp8_build_intra_predictors_mby_s(xd,
+                                             xd->dst.y_buffer - xd->dst.y_stride,
+                                             xd->dst.y_buffer - 1,
+                                             xd->dst.y_stride,
+                                             xd->predictor,
+                                             16);
+            distortion2 = vpx_variance16x16
+                                          (*(b->base_src), b->src_stride,
+                                          x->e_mbd.predictor, 16, &sse);
+            rate2 += x->mbmode_cost[x->e_mbd.frame_type][x->e_mbd.mode_info_context->mbmi.mode];
+            this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+
+            if (this_rd < best_intra_rd)
+            {
+                best_intra_rd = this_rd;
+                *returnintra = distortion2;
+            }
+            break;
+
+        case NEWMV:
+        {
+            int thissme;
+            int step_param;
+            int further_steps;
+            int n = 0;
+            int sadpb = x->sadperbit16;
+            int_mv mvp_full;
+
+            int col_min = ((best_ref_mv.as_mv.col+7)>>3) - MAX_FULL_PEL_VAL;
+            int row_min = ((best_ref_mv.as_mv.row+7)>>3) - MAX_FULL_PEL_VAL;
+            int col_max = (best_ref_mv.as_mv.col>>3)
+                         + MAX_FULL_PEL_VAL;
+            int row_max = (best_ref_mv.as_mv.row>>3)
+                         + MAX_FULL_PEL_VAL;
+
+            int tmp_col_min = x->mv_col_min;
+            int tmp_col_max = x->mv_col_max;
+            int tmp_row_min = x->mv_row_min;
+            int tmp_row_max = x->mv_row_max;
+
+            int speed_adjust = (cpi->Speed > 5) ? ((cpi->Speed >= 8)? 3 : 2) : 1;
+
+            /* Further step/diamond searches as necessary */
+            step_param = cpi->sf.first_step + speed_adjust;
+
+#if CONFIG_MULTI_RES_ENCODING
+            /* If lower-res frame is not available for mv reuse (because of
+               frame dropping or different temporal layer pattern), then higher
+               resol encoder does motion search without any previous knowledge.
+               Also, since last frame motion info is not stored, then we can not
+               use improved_mv_pred. */
+            if (cpi->oxcf.mr_encoder_id)
+                sf_improved_mv_pred = 0;
+
+            // Only use parent MV as predictor if this candidate reference frame
+            // (|this_ref_frame|) is equal to |parent_ref_frame|.
+            if (parent_ref_valid && (parent_ref_frame == this_ref_frame))
+            {
+                /* Use parent MV as predictor. Adjust search range
+                 * accordingly.
+                 */
+                mvp.as_int = parent_ref_mv.as_int;
+                mvp_full.as_mv.col = parent_ref_mv.as_mv.col>>3;
+                mvp_full.as_mv.row = parent_ref_mv.as_mv.row>>3;
+
+                if(dissim <=32) step_param += 3;
+                else if(dissim <=128) step_param += 2;
+                else step_param += 1;
+            }else
+#endif
+            {
+                if(sf_improved_mv_pred)
+                {
+                    if(!saddone)
+                    {
+                        vp8_cal_sad(cpi,xd,x, recon_yoffset ,&near_sadidx[0] );
+                        saddone = 1;
+                    }
+
+                    vp8_mv_pred(cpi, &x->e_mbd, x->e_mbd.mode_info_context,
+                                &mvp,x->e_mbd.mode_info_context->mbmi.ref_frame,
+                                cpi->common.ref_frame_sign_bias, &sr,
+                                &near_sadidx[0]);
+
+                    sr += speed_adjust;
+                    /* adjust search range according to sr from mv prediction */
+                    if(sr > step_param)
+                        step_param = sr;
+
+                    mvp_full.as_mv.col = mvp.as_mv.col>>3;
+                    mvp_full.as_mv.row = mvp.as_mv.row>>3;
+                }else
+                {
+                    mvp.as_int = best_ref_mv.as_int;
+                    mvp_full.as_mv.col = best_ref_mv.as_mv.col>>3;
+                    mvp_full.as_mv.row = best_ref_mv.as_mv.row>>3;
+                }
+            }
+
+#if CONFIG_MULTI_RES_ENCODING
+            if (parent_ref_valid && (parent_ref_frame == this_ref_frame) &&
+                dissim <= 2 &&
+                VPXMAX(abs(best_ref_mv.as_mv.row - parent_ref_mv.as_mv.row),
+                       abs(best_ref_mv.as_mv.col - parent_ref_mv.as_mv.col)) <=
+                    4)
+            {
+                d->bmi.mv.as_int = mvp_full.as_int;
+                mode_mv[NEWMV].as_int = mvp_full.as_int;
+
+                cpi->find_fractional_mv_step(x, b, d, &d->bmi.mv, &best_ref_mv,
+                                             x->errorperbit,
+                                             &cpi->fn_ptr[BLOCK_16X16],
+                                             cpi->mb.mvcost,
+                                             &distortion2,&sse);
+            }else
+#endif
+            {
+                /* Get intersection of UMV window and valid MV window to
+                 * reduce # of checks in diamond search. */
+                if (x->mv_col_min < col_min )
+                    x->mv_col_min = col_min;
+                if (x->mv_col_max > col_max )
+                    x->mv_col_max = col_max;
+                if (x->mv_row_min < row_min )
+                    x->mv_row_min = row_min;
+                if (x->mv_row_max > row_max )
+                    x->mv_row_max = row_max;
+
+                further_steps = (cpi->Speed >= 8)?
+                           0: (cpi->sf.max_step_search_steps - 1 - step_param);
+
+                if (cpi->sf.search_method == HEX)
+                {
+#if CONFIG_MULTI_RES_ENCODING
+                /* TODO: In higher-res pick_inter_mode, step_param is used to
+                 * modify hex search range. Here, set step_param to 0 not to
+                 * change the behavior in lowest-resolution encoder.
+                 * Will improve it later.
+                 */
+                /* Set step_param to 0 to ensure large-range motion search
+                 * when mv reuse if not valid (i.e. |parent_ref_valid| = 0),
+                 * or if this candidate reference frame (|this_ref_frame|) is
+                 * not equal to |parent_ref_frame|.
+                 */
+                if (!parent_ref_valid || (parent_ref_frame != this_ref_frame))
+                    step_param = 0;
+#endif
+                    bestsme = vp8_hex_search(x, b, d, &mvp_full, &d->bmi.mv,
+                                          step_param, sadpb,
+                                          &cpi->fn_ptr[BLOCK_16X16],
+                                          x->mvsadcost, x->mvcost, &best_ref_mv);
+                    mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
+                }
+                else
+                {
+                    bestsme = cpi->diamond_search_sad(x, b, d, &mvp_full,
+                                          &d->bmi.mv, step_param, sadpb, &num00,
+                                          &cpi->fn_ptr[BLOCK_16X16],
+                                          x->mvcost, &best_ref_mv);
+                    mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
+
+                    /* Further step/diamond searches as necessary */
+                    n = num00;
+                    num00 = 0;
+
+                    while (n < further_steps)
+                    {
+                        n++;
+
+                        if (num00)
+                            num00--;
+                        else
+                        {
+                            thissme =
+                            cpi->diamond_search_sad(x, b, d, &mvp_full,
+                                                    &d->bmi.mv,
+                                                    step_param + n,
+                                                    sadpb, &num00,
+                                                    &cpi->fn_ptr[BLOCK_16X16],
+                                                    x->mvcost, &best_ref_mv);
+                            if (thissme < bestsme)
+                            {
+                                bestsme = thissme;
+                                mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
+                            }
+                            else
+                            {
+                                d->bmi.mv.as_int = mode_mv[NEWMV].as_int;
+                            }
+                        }
+                    }
+                }
+
+                x->mv_col_min = tmp_col_min;
+                x->mv_col_max = tmp_col_max;
+                x->mv_row_min = tmp_row_min;
+                x->mv_row_max = tmp_row_max;
+
+                if (bestsme < INT_MAX)
+                    cpi->find_fractional_mv_step(x, b, d, &d->bmi.mv,
+                                             &best_ref_mv, x->errorperbit,
+                                             &cpi->fn_ptr[BLOCK_16X16],
+                                             cpi->mb.mvcost,
+                                             &distortion2,&sse);
+            }
+
+            mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
+            // The clamp below is not necessary from the perspective
+            // of VP8 bitstream, but is added to improve ChromeCast
+            // mirroring's robustness. Please do not remove.
+            vp8_clamp_mv2(&mode_mv[this_mode], xd);
+            /* mv cost; */
+            rate2 += vp8_mv_bit_cost(&mode_mv[NEWMV], &best_ref_mv,
+                                     cpi->mb.mvcost, 128);
+        }
+
+        case NEARESTMV:
+        case NEARMV:
+            if (mode_mv[this_mode].as_int == 0)
+                continue;
+
+        case ZEROMV:
+
+            /* Trap vectors that reach beyond the UMV borders
+             * Note that ALL New MV, Nearest MV Near MV and Zero MV code drops
+             * through to this point because of the lack of break statements
+             * in the previous two cases.
+             */
+            if (((mode_mv[this_mode].as_mv.row >> 3) < x->mv_row_min) ||
+                ((mode_mv[this_mode].as_mv.row >> 3) > x->mv_row_max) ||
+                ((mode_mv[this_mode].as_mv.col >> 3) < x->mv_col_min) ||
+                ((mode_mv[this_mode].as_mv.col >> 3) > x->mv_col_max))
+                continue;
+
+            rate2 += vp8_cost_mv_ref(this_mode, mdcounts);
+            x->e_mbd.mode_info_context->mbmi.mv.as_int =
+                                                    mode_mv[this_mode].as_int;
+            this_rd = evaluate_inter_mode(&sse, rate2, &distortion2, cpi, x,
+                                          rd_adjustment);
+
+            break;
+        default:
+            break;
+        }
+
+#if CONFIG_TEMPORAL_DENOISING
+        if (cpi->oxcf.noise_sensitivity)
+        {
+            /* Store for later use by denoiser. */
+            // Dont' denoise with GOLDEN OR ALTREF is they are old reference
+            // frames (greater than MAX_GF_ARF_DENOISE_RANGE frames in past).
+            int skip_old_reference = ((this_ref_frame != LAST_FRAME) &&
+                (cpi->common.current_video_frame -
+                 cpi->current_ref_frames[this_ref_frame] >
+                 MAX_GF_ARF_DENOISE_RANGE)) ? 1 : 0;
+            if (this_mode == ZEROMV && sse < zero_mv_sse &&
+                !skip_old_reference)
+            {
+                zero_mv_sse = sse;
+                x->best_zeromv_reference_frame =
+                        x->e_mbd.mode_info_context->mbmi.ref_frame;
+            }
+
+            // Store the best NEWMV in x for later use in the denoiser.
+            if (x->e_mbd.mode_info_context->mbmi.mode == NEWMV &&
+                sse < best_sse && !skip_old_reference)
+            {
+                best_sse = sse;
+                x->best_sse_inter_mode = NEWMV;
+                x->best_sse_mv = x->e_mbd.mode_info_context->mbmi.mv;
+                x->need_to_clamp_best_mvs =
+                    x->e_mbd.mode_info_context->mbmi.need_to_clamp_mvs;
+                x->best_reference_frame =
+                    x->e_mbd.mode_info_context->mbmi.ref_frame;
+            }
+        }
+#endif
+
+        if (this_rd < best_rd || x->skip)
+        {
+            /* Note index of best mode */
+            best_mode_index = mode_index;
+
+            *returnrate = rate2;
+            *returndistortion = distortion2;
+            best_rd_sse = sse;
+            best_rd = this_rd;
+            memcpy(&best_mbmode, &x->e_mbd.mode_info_context->mbmi,
+                   sizeof(MB_MODE_INFO));
+
+            /* Testing this mode gave rise to an improvement in best error
+             * score. Lower threshold a bit for next time
+             */
+            x->rd_thresh_mult[mode_index] =
+                     (x->rd_thresh_mult[mode_index] >= (MIN_THRESHMULT + 2)) ?
+                     x->rd_thresh_mult[mode_index] - 2 : MIN_THRESHMULT;
+            x->rd_threshes[mode_index] =
+                                   (cpi->rd_baseline_thresh[mode_index] >> 7) *
+                                   x->rd_thresh_mult[mode_index];
+        }
+
+        /* If the mode did not help improve the best error case then raise the
+         * threshold for testing that mode next time around.
+         */
+        else
+        {
+            x->rd_thresh_mult[mode_index] += 4;
+
+            if (x->rd_thresh_mult[mode_index] > MAX_THRESHMULT)
+                x->rd_thresh_mult[mode_index] = MAX_THRESHMULT;
+
+            x->rd_threshes[mode_index] =
+                         (cpi->rd_baseline_thresh[mode_index] >> 7) *
+                         x->rd_thresh_mult[mode_index];
+        }
+
+        if (x->skip)
+            break;
+    }
+
+    /* Reduce the activation RD thresholds for the best choice mode */
+    if ((cpi->rd_baseline_thresh[best_mode_index] > 0) && (cpi->rd_baseline_thresh[best_mode_index] < (INT_MAX >> 2)))
+    {
+        int best_adjustment = (x->rd_thresh_mult[best_mode_index] >> 3);
+
+        x->rd_thresh_mult[best_mode_index] =
+                        (x->rd_thresh_mult[best_mode_index]
+                        >= (MIN_THRESHMULT + best_adjustment)) ?
+                        x->rd_thresh_mult[best_mode_index] - best_adjustment :
+                        MIN_THRESHMULT;
+        x->rd_threshes[best_mode_index] =
+                        (cpi->rd_baseline_thresh[best_mode_index] >> 7) *
+                        x->rd_thresh_mult[best_mode_index];
+    }
+
+
+    {
+        int this_rdbin = (*returndistortion >> 7);
+
+        if (this_rdbin >= 1024)
+        {
+            this_rdbin = 1023;
+        }
+
+        x->error_bins[this_rdbin] ++;
+    }
+
+#if CONFIG_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity)
+    {
+        int block_index = mb_row * cpi->common.mb_cols + mb_col;
+        int reevaluate = 0;
+        int is_noisy = 0;
+        if (x->best_sse_inter_mode == DC_PRED)
+        {
+            /* No best MV found. */
+            x->best_sse_inter_mode = best_mbmode.mode;
+            x->best_sse_mv = best_mbmode.mv;
+            x->need_to_clamp_best_mvs = best_mbmode.need_to_clamp_mvs;
+            x->best_reference_frame = best_mbmode.ref_frame;
+            best_sse = best_rd_sse;
+        }
+        // For non-skin blocks that have selected ZEROMV for this current frame,
+        // and have been selecting ZEROMV_LAST (on the base layer frame) at
+        // least |x~20| consecutive past frames in a row, label the block for
+        // possible increase in denoising strength. We also condition this
+        // labeling on there being significant denoising in the scene
+        if  (cpi->oxcf.noise_sensitivity == 4) {
+          if (cpi->denoiser.nmse_source_diff >
+              70 * cpi->denoiser.threshold_aggressive_mode / 100)
+            is_noisy = 1;
+        } else {
+          if (cpi->mse_source_denoised > 1000)
+            is_noisy = 1;
+        }
+        x->increase_denoising = 0;
+        if (!x->is_skin &&
+            x->best_sse_inter_mode == ZEROMV &&
+            (x->best_reference_frame == LAST_FRAME ||
+            x->best_reference_frame == cpi->closest_reference_frame) &&
+            cpi->consec_zero_last[block_index] >= 20 &&
+            is_noisy) {
+            x->increase_denoising = 1;
+        }
+        x->denoise_zeromv = 0;
+        vp8_denoiser_denoise_mb(&cpi->denoiser, x, best_sse, zero_mv_sse,
+                                recon_yoffset, recon_uvoffset,
+                                &cpi->common.lf_info, mb_row, mb_col,
+                                block_index,
+                                cpi->consec_zero_last_mvbias[block_index]);
+
+        // Reevaluate ZEROMV after denoising: for large noise content
+        // (i.e., cpi->mse_source_denoised is above threshold), do this for all
+        // blocks that did not pick ZEROMV as best mode but are using ZEROMV
+        // for denoising. Otherwise, always re-evaluate for blocks that picked
+        // INTRA mode as best mode.
+        // Avoid blocks that have been biased against ZERO_LAST
+        // (i.e., dot artifact candidate blocks).
+        reevaluate = (best_mbmode.ref_frame == INTRA_FRAME) ||
+                     (best_mbmode.mode != ZEROMV &&
+                      x->denoise_zeromv &&
+                      cpi->mse_source_denoised > 2000);
+        if (!dot_artifact_candidate &&
+            reevaluate &&
+            x->best_zeromv_reference_frame != INTRA_FRAME)
+        {
+            int this_rd = 0;
+            int this_ref_frame = x->best_zeromv_reference_frame;
+            rd_adjustment = 100;
+            rate2 = x->ref_frame_cost[this_ref_frame] +
+                    vp8_cost_mv_ref(ZEROMV, mdcounts);
+            distortion2 = 0;
+
+            /* set up the proper prediction buffers for the frame */
+            x->e_mbd.mode_info_context->mbmi.ref_frame = this_ref_frame;
+            x->e_mbd.pre.y_buffer = plane[this_ref_frame][0];
+            x->e_mbd.pre.u_buffer = plane[this_ref_frame][1];
+            x->e_mbd.pre.v_buffer = plane[this_ref_frame][2];
+
+            x->e_mbd.mode_info_context->mbmi.mode = ZEROMV;
+            x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
+            x->e_mbd.mode_info_context->mbmi.mv.as_int = 0;
+            this_rd = evaluate_inter_mode(&sse, rate2, &distortion2, cpi, x,
+                                          rd_adjustment);
+
+            if (this_rd < best_rd)
+            {
+                memcpy(&best_mbmode, &x->e_mbd.mode_info_context->mbmi,
+                       sizeof(MB_MODE_INFO));
+            }
+        }
+
+    }
+#endif
+
+    if (cpi->is_src_frame_alt_ref &&
+        (best_mbmode.mode != ZEROMV || best_mbmode.ref_frame != ALTREF_FRAME))
+    {
+        x->e_mbd.mode_info_context->mbmi.mode = ZEROMV;
+        x->e_mbd.mode_info_context->mbmi.ref_frame = ALTREF_FRAME;
+        x->e_mbd.mode_info_context->mbmi.mv.as_int = 0;
+        x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
+        x->e_mbd.mode_info_context->mbmi.mb_skip_coeff =
+                                        (cpi->common.mb_no_coeff_skip);
+        x->e_mbd.mode_info_context->mbmi.partitioning = 0;
+
+        return;
+    }
+
+    /* set to the best mb mode, this copy can be skip if x->skip since it
+     * already has the right content */
+    if (!x->skip)
+        memcpy(&x->e_mbd.mode_info_context->mbmi, &best_mbmode,
+               sizeof(MB_MODE_INFO));
+
+    if (best_mbmode.mode <= B_PRED)
+    {
+        /* set mode_info_context->mbmi.uv_mode */
+        pick_intra_mbuv_mode(x);
+    }
+
+    if (sign_bias
+      != cpi->common.ref_frame_sign_bias[xd->mode_info_context->mbmi.ref_frame])
+        best_ref_mv.as_int = best_ref_mv_sb[!sign_bias].as_int;
+
+    update_mvcount(x, &best_ref_mv);
+}
+
+void vp8_pick_intra_mode(MACROBLOCK *x, int *rate_)
+{
+    int error4x4, error16x16 = INT_MAX;
+    int rate, best_rate = 0, distortion, best_sse;
+    MB_PREDICTION_MODE mode, best_mode = DC_PRED;
+    int this_rd;
+    unsigned int sse;
+    BLOCK *b = &x->block[0];
+    MACROBLOCKD *xd = &x->e_mbd;
+
+    xd->mode_info_context->mbmi.ref_frame = INTRA_FRAME;
+
+    pick_intra_mbuv_mode(x);
+
+    for (mode = DC_PRED; mode <= TM_PRED; mode ++)
+    {
+        xd->mode_info_context->mbmi.mode = mode;
+        vp8_build_intra_predictors_mby_s(xd,
+                                         xd->dst.y_buffer - xd->dst.y_stride,
+                                         xd->dst.y_buffer - 1,
+                                         xd->dst.y_stride,
+                                         xd->predictor,
+                                         16);
+        distortion = vpx_variance16x16
+            (*(b->base_src), b->src_stride, xd->predictor, 16, &sse);
+        rate = x->mbmode_cost[xd->frame_type][mode];
+        this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
+
+        if (error16x16 > this_rd)
+        {
+            error16x16 = this_rd;
+            best_mode = mode;
+            best_sse = sse;
+            best_rate = rate;
+        }
+    }
+    xd->mode_info_context->mbmi.mode = best_mode;
+
+    error4x4 = pick_intra4x4mby_modes(x, &rate,
+                                      &best_sse);
+    if (error4x4 < error16x16)
+    {
+        xd->mode_info_context->mbmi.mode = B_PRED;
+        best_rate = rate;
+    }
+
+    *rate_ = best_rate;
+}
diff --git a/libvpx/libvpx/vp8/encoder/pickinter.h b/libvpx/libvpx/vp8/encoder/pickinter.h
new file mode 100644
index 0000000..cf3b1f8
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/pickinter.h
@@ -0,0 +1,35 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_PICKINTER_H_
+#define VP8_ENCODER_PICKINTER_H_
+#include "vpx_config.h"
+#include "vp8/common/onyxc_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern void vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
+                                int recon_uvoffset, int *returnrate,
+                                int *returndistortion, int *returnintra,
+                                int mb_row, int mb_col);
+extern void vp8_pick_intra_mode(MACROBLOCK *x, int *rate);
+
+extern int vp8_get_inter_mbpred_error(MACROBLOCK *mb,
+                                      const vp8_variance_fn_ptr_t *vfp,
+                                      unsigned int *sse,
+                                      int_mv this_mv);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_PICKINTER_H_
diff --git a/libvpx/libvpx/vp8/encoder/picklpf.c b/libvpx/libvpx/vp8/encoder/picklpf.c
new file mode 100644
index 0000000..debd304
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/picklpf.c
@@ -0,0 +1,407 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+#include "vp8/common/onyxc_int.h"
+#include "onyx_int.h"
+#include "vp8/encoder/quantize.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vp8/common/alloccommon.h"
+#include "vp8/common/loopfilter.h"
+#if ARCH_ARM
+#include "vpx_ports/arm.h"
+#endif
+
+extern int vp8_calc_ss_err(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest);
+
+static void yv12_copy_partial_frame(YV12_BUFFER_CONFIG *src_ybc,
+                                    YV12_BUFFER_CONFIG *dst_ybc)
+{
+    unsigned char *src_y, *dst_y;
+    int yheight;
+    int ystride;
+    int yoffset;
+    int linestocopy;
+
+    yheight  = src_ybc->y_height;
+    ystride  = src_ybc->y_stride;
+
+    /* number of MB rows to use in partial filtering */
+    linestocopy = (yheight >> 4) / PARTIAL_FRAME_FRACTION;
+    linestocopy = linestocopy ? linestocopy << 4 : 16;     /* 16 lines per MB */
+
+    /* Copy extra 4 so that full filter context is available if filtering done
+     * on the copied partial frame and not original. Partial filter does mb
+     * filtering for top row also, which can modify3 pixels above.
+     */
+    linestocopy += 4;
+    /* partial image starts at ~middle of frame (macroblock border)*/
+    yoffset  = ystride * (((yheight >> 5) * 16) - 4);
+    src_y = src_ybc->y_buffer + yoffset;
+    dst_y = dst_ybc->y_buffer + yoffset;
+
+    memcpy(dst_y, src_y, ystride * linestocopy);
+}
+
+static int calc_partial_ssl_err(YV12_BUFFER_CONFIG *source,
+                                YV12_BUFFER_CONFIG *dest)
+{
+    int i, j;
+    int Total = 0;
+    int srcoffset, dstoffset;
+    unsigned char *src = source->y_buffer;
+    unsigned char *dst = dest->y_buffer;
+
+    int linestocopy;
+
+    /* number of MB rows to use in partial filtering */
+    linestocopy = (source->y_height >> 4) / PARTIAL_FRAME_FRACTION;
+    linestocopy = linestocopy ? linestocopy << 4 : 16;     /* 16 lines per MB */
+
+
+    /* partial image starts at ~middle of frame (macroblock border)*/
+    srcoffset = source->y_stride * ((dest->y_height >> 5) * 16);
+    dstoffset = dest->y_stride   * ((dest->y_height >> 5) * 16);
+
+    src += srcoffset;
+    dst += dstoffset;
+
+    /* Loop through the Y plane raw and reconstruction data summing
+     * (square differences)
+     */
+    for (i = 0; i < linestocopy; i += 16)
+    {
+        for (j = 0; j < source->y_width; j += 16)
+        {
+            unsigned int sse;
+            Total += vpx_mse16x16(src + j, source->y_stride,
+                                                     dst + j, dest->y_stride,
+                                                     &sse);
+        }
+
+        src += 16 * source->y_stride;
+        dst += 16 * dest->y_stride;
+    }
+
+    return Total;
+}
+
+/* Enforce a minimum filter level based upon baseline Q */
+static int get_min_filter_level(VP8_COMP *cpi, int base_qindex)
+{
+    int min_filter_level;
+
+    if (cpi->source_alt_ref_active && cpi->common.refresh_golden_frame &&
+        !cpi->common.refresh_alt_ref_frame)
+        min_filter_level = 0;
+    else
+    {
+        if (base_qindex <= 6)
+            min_filter_level = 0;
+        else if (base_qindex <= 16)
+            min_filter_level = 1;
+        else
+            min_filter_level = (base_qindex / 8);
+    }
+
+    return min_filter_level;
+}
+
+/* Enforce a maximum filter level based upon baseline Q */
+static int get_max_filter_level(VP8_COMP *cpi, int base_qindex)
+{
+    /* PGW August 2006: Highest filter values almost always a bad idea */
+
+    /* jbb chg: 20100118 - not so any more with this overquant stuff allow
+     * high values with lots of intra coming in.
+     */
+    int max_filter_level = MAX_LOOP_FILTER;
+    (void)base_qindex;
+
+    if (cpi->twopass.section_intra_rating > 8)
+        max_filter_level = MAX_LOOP_FILTER * 3 / 4;
+
+    return max_filter_level;
+}
+
+void vp8cx_pick_filter_level_fast(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = &cpi->common;
+
+    int best_err = 0;
+    int filt_err = 0;
+    int min_filter_level = get_min_filter_level(cpi, cm->base_qindex);
+    int max_filter_level = get_max_filter_level(cpi, cm->base_qindex);
+    int filt_val;
+    int best_filt_val;
+    YV12_BUFFER_CONFIG * saved_frame = cm->frame_to_show;
+
+    /* Replace unfiltered frame buffer with a new one */
+    cm->frame_to_show = &cpi->pick_lf_lvl_frame;
+
+    if (cm->frame_type == KEY_FRAME)
+        cm->sharpness_level = 0;
+    else
+        cm->sharpness_level = cpi->oxcf.Sharpness;
+
+    if (cm->sharpness_level != cm->last_sharpness_level)
+    {
+        vp8_loop_filter_update_sharpness(&cm->lf_info, cm->sharpness_level);
+        cm->last_sharpness_level = cm->sharpness_level;
+    }
+
+    /* Start the search at the previous frame filter level unless it is
+     * now out of range.
+     */
+    if (cm->filter_level < min_filter_level)
+        cm->filter_level = min_filter_level;
+    else if (cm->filter_level > max_filter_level)
+        cm->filter_level = max_filter_level;
+
+    filt_val = cm->filter_level;
+    best_filt_val = filt_val;
+
+    /* Get the err using the previous frame's filter value. */
+
+    /* Copy the unfiltered / processed recon buffer to the new buffer */
+    yv12_copy_partial_frame(saved_frame, cm->frame_to_show);
+    vp8_loop_filter_partial_frame(cm, &cpi->mb.e_mbd, filt_val);
+
+    best_err = calc_partial_ssl_err(sd, cm->frame_to_show);
+
+    filt_val -= 1 + (filt_val > 10);
+
+    /* Search lower filter levels */
+    while (filt_val >= min_filter_level)
+    {
+        /* Apply the loop filter */
+        yv12_copy_partial_frame(saved_frame, cm->frame_to_show);
+        vp8_loop_filter_partial_frame(cm, &cpi->mb.e_mbd, filt_val);
+
+        /* Get the err for filtered frame */
+        filt_err = calc_partial_ssl_err(sd, cm->frame_to_show);
+
+        /* Update the best case record or exit loop. */
+        if (filt_err < best_err)
+        {
+            best_err = filt_err;
+            best_filt_val = filt_val;
+        }
+        else
+            break;
+
+        /* Adjust filter level */
+        filt_val -= 1 + (filt_val > 10);
+    }
+
+    /* Search up (note that we have already done filt_val = cm->filter_level) */
+    filt_val = cm->filter_level + 1 + (filt_val > 10);
+
+    if (best_filt_val == cm->filter_level)
+    {
+        /* Resist raising filter level for very small gains */
+        best_err -= (best_err >> 10);
+
+        while (filt_val < max_filter_level)
+        {
+            /* Apply the loop filter */
+            yv12_copy_partial_frame(saved_frame, cm->frame_to_show);
+
+            vp8_loop_filter_partial_frame(cm, &cpi->mb.e_mbd, filt_val);
+
+            /* Get the err for filtered frame */
+            filt_err = calc_partial_ssl_err(sd, cm->frame_to_show);
+
+            /* Update the best case record or exit loop. */
+            if (filt_err < best_err)
+            {
+                /* Do not raise filter level if improvement is < 1 part
+                 * in 4096
+                 */
+                best_err = filt_err - (filt_err >> 10);
+
+                best_filt_val = filt_val;
+            }
+            else
+                break;
+
+            /* Adjust filter level */
+            filt_val += 1 + (filt_val > 10);
+        }
+    }
+
+    cm->filter_level = best_filt_val;
+
+    if (cm->filter_level < min_filter_level)
+        cm->filter_level = min_filter_level;
+
+    if (cm->filter_level > max_filter_level)
+        cm->filter_level = max_filter_level;
+
+    /* restore unfiltered frame pointer */
+    cm->frame_to_show = saved_frame;
+}
+
+/* Stub function for now Alt LF not used */
+void vp8cx_set_alt_lf_level(VP8_COMP *cpi, int filt_val)
+{
+    MACROBLOCKD *mbd = &cpi->mb.e_mbd;
+    (void) filt_val;
+
+    mbd->segment_feature_data[MB_LVL_ALT_LF][0] = cpi->segment_feature_data[MB_LVL_ALT_LF][0];
+    mbd->segment_feature_data[MB_LVL_ALT_LF][1] = cpi->segment_feature_data[MB_LVL_ALT_LF][1];
+    mbd->segment_feature_data[MB_LVL_ALT_LF][2] = cpi->segment_feature_data[MB_LVL_ALT_LF][2];
+    mbd->segment_feature_data[MB_LVL_ALT_LF][3] = cpi->segment_feature_data[MB_LVL_ALT_LF][3];
+}
+
+void vp8cx_pick_filter_level(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = &cpi->common;
+
+    int best_err = 0;
+    int filt_err = 0;
+    int min_filter_level = get_min_filter_level(cpi, cm->base_qindex);
+    int max_filter_level = get_max_filter_level(cpi, cm->base_qindex);
+
+    int filter_step;
+    int filt_high = 0;
+    int filt_mid;
+    int filt_low = 0;
+    int filt_best;
+    int filt_direction = 0;
+
+    /* Bias against raising loop filter and in favor of lowering it */
+    int Bias = 0;
+
+    int ss_err[MAX_LOOP_FILTER + 1];
+
+    YV12_BUFFER_CONFIG * saved_frame = cm->frame_to_show;
+
+    memset(ss_err, 0, sizeof(ss_err));
+
+    /* Replace unfiltered frame buffer with a new one */
+    cm->frame_to_show = &cpi->pick_lf_lvl_frame;
+
+    if (cm->frame_type == KEY_FRAME)
+        cm->sharpness_level = 0;
+    else
+        cm->sharpness_level = cpi->oxcf.Sharpness;
+
+    /* Start the search at the previous frame filter level unless it is
+     * now out of range.
+     */
+    filt_mid = cm->filter_level;
+
+    if (filt_mid < min_filter_level)
+        filt_mid = min_filter_level;
+    else if (filt_mid > max_filter_level)
+        filt_mid = max_filter_level;
+
+    /* Define the initial step size */
+    filter_step = (filt_mid < 16) ? 4 : filt_mid / 4;
+
+    /* Get baseline error score */
+
+    /* Copy the unfiltered / processed recon buffer to the new buffer */
+    vpx_yv12_copy_y(saved_frame, cm->frame_to_show);
+
+    vp8cx_set_alt_lf_level(cpi, filt_mid);
+    vp8_loop_filter_frame_yonly(cm, &cpi->mb.e_mbd, filt_mid);
+
+    best_err = vp8_calc_ss_err(sd, cm->frame_to_show);
+
+    ss_err[filt_mid] = best_err;
+
+    filt_best = filt_mid;
+
+    while (filter_step > 0)
+    {
+        Bias = (best_err >> (15 - (filt_mid / 8))) * filter_step;
+
+        if (cpi->twopass.section_intra_rating < 20)
+            Bias = Bias * cpi->twopass.section_intra_rating / 20;
+
+        filt_high = ((filt_mid + filter_step) > max_filter_level) ? max_filter_level : (filt_mid + filter_step);
+        filt_low = ((filt_mid - filter_step) < min_filter_level) ? min_filter_level : (filt_mid - filter_step);
+
+        if ((filt_direction <= 0) && (filt_low != filt_mid))
+        {
+            if(ss_err[filt_low] == 0)
+            {
+                /* Get Low filter error score */
+                vpx_yv12_copy_y(saved_frame, cm->frame_to_show);
+                vp8cx_set_alt_lf_level(cpi, filt_low);
+                vp8_loop_filter_frame_yonly(cm, &cpi->mb.e_mbd, filt_low);
+
+                filt_err = vp8_calc_ss_err(sd, cm->frame_to_show);
+                ss_err[filt_low] = filt_err;
+            }
+            else
+                filt_err = ss_err[filt_low];
+
+            /* If value is close to the best so far then bias towards a
+             * lower loop filter value.
+             */
+            if ((filt_err - Bias) < best_err)
+            {
+                /* Was it actually better than the previous best? */
+                if (filt_err < best_err)
+                    best_err = filt_err;
+
+                filt_best = filt_low;
+            }
+        }
+
+        /* Now look at filt_high */
+        if ((filt_direction >= 0) && (filt_high != filt_mid))
+        {
+            if(ss_err[filt_high] == 0)
+            {
+                vpx_yv12_copy_y(saved_frame, cm->frame_to_show);
+                vp8cx_set_alt_lf_level(cpi, filt_high);
+                vp8_loop_filter_frame_yonly(cm, &cpi->mb.e_mbd, filt_high);
+
+                filt_err = vp8_calc_ss_err(sd, cm->frame_to_show);
+                ss_err[filt_high] = filt_err;
+            }
+            else
+                filt_err = ss_err[filt_high];
+
+            /* Was it better than the previous best? */
+            if (filt_err < (best_err - Bias))
+            {
+                best_err = filt_err;
+                filt_best = filt_high;
+            }
+        }
+
+        /* Half the step distance if the best filter value was the same
+         * as last time
+         */
+        if (filt_best == filt_mid)
+        {
+            filter_step = filter_step / 2;
+            filt_direction = 0;
+        }
+        else
+        {
+            filt_direction = (filt_best < filt_mid) ? -1 : 1;
+            filt_mid = filt_best;
+        }
+    }
+
+    cm->filter_level = filt_best;
+
+    /* restore unfiltered frame pointer */
+    cm->frame_to_show = saved_frame;
+}
diff --git a/libvpx/libvpx/vp8/encoder/quantize.h b/libvpx/libvpx/vp8/encoder/quantize.h
new file mode 100644
index 0000000..7d36c2b
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/quantize.h
@@ -0,0 +1,34 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_QUANTIZE_H_
+#define VP8_ENCODER_QUANTIZE_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP8_COMP;
+struct macroblock;
+extern void vp8_quantize_mb(struct macroblock *x);
+extern void vp8_quantize_mby(struct macroblock *x);
+extern void vp8_quantize_mbuv(struct macroblock *x);
+extern void vp8_set_quantizer(struct VP8_COMP *cpi, int Q);
+extern void vp8cx_frame_init_quantizer(struct VP8_COMP *cpi);
+extern void vp8_update_zbin_extra(struct VP8_COMP *cpi, struct macroblock *x);
+extern void vp8cx_mb_init_quantizer(struct VP8_COMP *cpi, struct macroblock *x, int ok_to_skip);
+extern void vp8cx_init_quantizer(struct VP8_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_QUANTIZE_H_
diff --git a/libvpx/libvpx/vp8/encoder/ratectrl.c b/libvpx/libvpx/vp8/encoder/ratectrl.c
new file mode 100644
index 0000000..7da3d71
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/ratectrl.c
@@ -0,0 +1,1638 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <limits.h>
+#include <assert.h>
+
+#include "math.h"
+#include "vp8/common/common.h"
+#include "ratectrl.h"
+#include "vp8/common/entropymode.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp8/common/systemdependent.h"
+#include "encodemv.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+
+#define MIN_BPB_FACTOR          0.01
+#define MAX_BPB_FACTOR          50
+
+extern const MB_PREDICTION_MODE vp8_mode_order[MAX_MODES];
+
+
+
+#ifdef MODE_STATS
+extern int y_modes[5];
+extern int uv_modes[4];
+extern int b_modes[10];
+
+extern int inter_y_modes[10];
+extern int inter_uv_modes[4];
+extern int inter_b_modes[10];
+#endif
+
+/* Bits Per MB at different Q (Multiplied by 512) */
+#define BPER_MB_NORMBITS    9
+
+/* Work in progress recalibration of baseline rate tables based on
+ * the assumption that bits per mb is inversely proportional to the
+ * quantizer value.
+ */
+const int vp8_bits_per_mb[2][QINDEX_RANGE] =
+{
+    /* Intra case 450000/Qintra */
+    {
+        1125000,900000, 750000, 642857, 562500, 500000, 450000, 450000,
+        409090, 375000, 346153, 321428, 300000, 281250, 264705, 264705,
+        250000, 236842, 225000, 225000, 214285, 214285, 204545, 204545,
+        195652, 195652, 187500, 180000, 180000, 173076, 166666, 160714,
+        155172, 150000, 145161, 140625, 136363, 132352, 128571, 125000,
+        121621, 121621, 118421, 115384, 112500, 109756, 107142, 104651,
+        102272, 100000, 97826,  97826,  95744,  93750,  91836,  90000,
+        88235,  86538,  84905,  83333,  81818,  80357,  78947,  77586,
+        76271,  75000,  73770,  72580,  71428,  70312,  69230,  68181,
+        67164,  66176,  65217,  64285,  63380,  62500,  61643,  60810,
+        60000,  59210,  59210,  58441,  57692,  56962,  56250,  55555,
+        54878,  54216,  53571,  52941,  52325,  51724,  51136,  50561,
+        49450,  48387,  47368,  46875,  45918,  45000,  44554,  44117,
+        43269,  42452,  41666,  40909,  40178,  39473,  38793,  38135,
+        36885,  36290,  35714,  35156,  34615,  34090,  33582,  33088,
+        32608,  32142,  31468,  31034,  30405,  29801,  29220,  28662,
+    },
+    /* Inter case 285000/Qinter */
+    {
+        712500, 570000, 475000, 407142, 356250, 316666, 285000, 259090,
+        237500, 219230, 203571, 190000, 178125, 167647, 158333, 150000,
+        142500, 135714, 129545, 123913, 118750, 114000, 109615, 105555,
+        101785, 98275,  95000,  91935,  89062,  86363,  83823,  81428,
+        79166,  77027,  75000,  73076,  71250,  69512,  67857,  66279,
+        64772,  63333,  61956,  60638,  59375,  58163,  57000,  55882,
+        54807,  53773,  52777,  51818,  50892,  50000,  49137,  47500,
+        45967,  44531,  43181,  41911,  40714,  39583,  38513,  37500,
+        36538,  35625,  34756,  33928,  33139,  32386,  31666,  30978,
+        30319,  29687,  29081,  28500,  27941,  27403,  26886,  26388,
+        25909,  25446,  25000,  24568,  23949,  23360,  22800,  22265,
+        21755,  21268,  20802,  20357,  19930,  19520,  19127,  18750,
+        18387,  18037,  17701,  17378,  17065,  16764,  16473,  16101,
+        15745,  15405,  15079,  14766,  14467,  14179,  13902,  13636,
+        13380,  13133,  12895,  12666,  12445,  12179,  11924,  11632,
+        11445,  11220,  11003,  10795,  10594,  10401,  10215,  10035,
+    }
+};
+
+static const int kf_boost_qadjustment[QINDEX_RANGE] =
+{
+    128, 129, 130, 131, 132, 133, 134, 135,
+    136, 137, 138, 139, 140, 141, 142, 143,
+    144, 145, 146, 147, 148, 149, 150, 151,
+    152, 153, 154, 155, 156, 157, 158, 159,
+    160, 161, 162, 163, 164, 165, 166, 167,
+    168, 169, 170, 171, 172, 173, 174, 175,
+    176, 177, 178, 179, 180, 181, 182, 183,
+    184, 185, 186, 187, 188, 189, 190, 191,
+    192, 193, 194, 195, 196, 197, 198, 199,
+    200, 200, 201, 201, 202, 203, 203, 203,
+    204, 204, 205, 205, 206, 206, 207, 207,
+    208, 208, 209, 209, 210, 210, 211, 211,
+    212, 212, 213, 213, 214, 214, 215, 215,
+    216, 216, 217, 217, 218, 218, 219, 219,
+    220, 220, 220, 220, 220, 220, 220, 220,
+    220, 220, 220, 220, 220, 220, 220, 220,
+};
+
+/* #define GFQ_ADJUSTMENT (Q+100) */
+#define GFQ_ADJUSTMENT vp8_gf_boost_qadjustment[Q]
+const int vp8_gf_boost_qadjustment[QINDEX_RANGE] =
+{
+    80, 82, 84, 86, 88, 90, 92, 94,
+    96, 97, 98, 99, 100, 101, 102, 103,
+    104, 105, 106, 107, 108, 109, 110, 111,
+    112, 113, 114, 115, 116, 117, 118, 119,
+    120, 121, 122, 123, 124, 125, 126, 127,
+    128, 129, 130, 131, 132, 133, 134, 135,
+    136, 137, 138, 139, 140, 141, 142, 143,
+    144, 145, 146, 147, 148, 149, 150, 151,
+    152, 153, 154, 155, 156, 157, 158, 159,
+    160, 161, 162, 163, 164, 165, 166, 167,
+    168, 169, 170, 171, 172, 173, 174, 175,
+    176, 177, 178, 179, 180, 181, 182, 183,
+    184, 184, 185, 185, 186, 186, 187, 187,
+    188, 188, 189, 189, 190, 190, 191, 191,
+    192, 192, 193, 193, 194, 194, 194, 194,
+    195, 195, 196, 196, 197, 197, 198, 198
+};
+
+/*
+const int vp8_gf_boost_qadjustment[QINDEX_RANGE] =
+{
+    100,101,102,103,104,105,105,106,
+    106,107,107,108,109,109,110,111,
+    112,113,114,115,116,117,118,119,
+    120,121,122,123,124,125,126,127,
+    128,129,130,131,132,133,134,135,
+    136,137,138,139,140,141,142,143,
+    144,145,146,147,148,149,150,151,
+    152,153,154,155,156,157,158,159,
+    160,161,162,163,164,165,166,167,
+    168,169,170,170,171,171,172,172,
+    173,173,173,174,174,174,175,175,
+    175,176,176,176,177,177,177,177,
+    178,178,179,179,180,180,181,181,
+    182,182,183,183,184,184,185,185,
+    186,186,187,187,188,188,189,189,
+    190,190,191,191,192,192,193,193,
+};
+*/
+
+static const int kf_gf_boost_qlimits[QINDEX_RANGE] =
+{
+    150, 155, 160, 165, 170, 175, 180, 185,
+    190, 195, 200, 205, 210, 215, 220, 225,
+    230, 235, 240, 245, 250, 255, 260, 265,
+    270, 275, 280, 285, 290, 295, 300, 305,
+    310, 320, 330, 340, 350, 360, 370, 380,
+    390, 400, 410, 420, 430, 440, 450, 460,
+    470, 480, 490, 500, 510, 520, 530, 540,
+    550, 560, 570, 580, 590, 600, 600, 600,
+    600, 600, 600, 600, 600, 600, 600, 600,
+    600, 600, 600, 600, 600, 600, 600, 600,
+    600, 600, 600, 600, 600, 600, 600, 600,
+    600, 600, 600, 600, 600, 600, 600, 600,
+    600, 600, 600, 600, 600, 600, 600, 600,
+    600, 600, 600, 600, 600, 600, 600, 600,
+    600, 600, 600, 600, 600, 600, 600, 600,
+    600, 600, 600, 600, 600, 600, 600, 600,
+};
+
+static const int gf_adjust_table[101] =
+{
+    100,
+    115, 130, 145, 160, 175, 190, 200, 210, 220, 230,
+    240, 260, 270, 280, 290, 300, 310, 320, 330, 340,
+    350, 360, 370, 380, 390, 400, 400, 400, 400, 400,
+    400, 400, 400, 400, 400, 400, 400, 400, 400, 400,
+    400, 400, 400, 400, 400, 400, 400, 400, 400, 400,
+    400, 400, 400, 400, 400, 400, 400, 400, 400, 400,
+    400, 400, 400, 400, 400, 400, 400, 400, 400, 400,
+    400, 400, 400, 400, 400, 400, 400, 400, 400, 400,
+    400, 400, 400, 400, 400, 400, 400, 400, 400, 400,
+    400, 400, 400, 400, 400, 400, 400, 400, 400, 400,
+};
+
+static const int gf_intra_usage_adjustment[20] =
+{
+    125, 120, 115, 110, 105, 100,  95,  85,  80,  75,
+    70,  65,  60,  55,  50,  50,  50,  50,  50,  50,
+};
+
+static const int gf_interval_table[101] =
+{
+    7,
+    7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+    7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+    7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+    8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+    8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+    9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+    9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+    10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+    10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+    11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+};
+
+static const unsigned int prior_key_frame_weight[KEY_FRAME_CONTEXT] = { 1, 2, 3, 4, 5 };
+
+
+void vp8_save_coding_context(VP8_COMP *cpi)
+{
+    CODING_CONTEXT *const cc = & cpi->coding_context;
+
+    /* Stores a snapshot of key state variables which can subsequently be
+     * restored with a call to vp8_restore_coding_context. These functions are
+     * intended for use in a re-code loop in vp8_compress_frame where the
+     * quantizer value is adjusted between loop iterations.
+     */
+
+    cc->frames_since_key          = cpi->frames_since_key;
+    cc->filter_level             = cpi->common.filter_level;
+    cc->frames_till_gf_update_due   = cpi->frames_till_gf_update_due;
+    cc->frames_since_golden       = cpi->frames_since_golden;
+
+    vp8_copy(cc->mvc,      cpi->common.fc.mvc);
+    vp8_copy(cc->mvcosts,  cpi->rd_costs.mvcosts);
+
+    vp8_copy(cc->ymode_prob,   cpi->common.fc.ymode_prob);
+    vp8_copy(cc->uv_mode_prob,  cpi->common.fc.uv_mode_prob);
+
+    vp8_copy(cc->ymode_count, cpi->mb.ymode_count);
+    vp8_copy(cc->uv_mode_count, cpi->mb.uv_mode_count);
+
+
+    /* Stats */
+#ifdef MODE_STATS
+    vp8_copy(cc->y_modes,       y_modes);
+    vp8_copy(cc->uv_modes,      uv_modes);
+    vp8_copy(cc->b_modes,       b_modes);
+    vp8_copy(cc->inter_y_modes,  inter_y_modes);
+    vp8_copy(cc->inter_uv_modes, inter_uv_modes);
+    vp8_copy(cc->inter_b_modes,  inter_b_modes);
+#endif
+
+    cc->this_frame_percent_intra = cpi->this_frame_percent_intra;
+}
+
+
+void vp8_restore_coding_context(VP8_COMP *cpi)
+{
+    CODING_CONTEXT *const cc = & cpi->coding_context;
+
+    /* Restore key state variables to the snapshot state stored in the
+     * previous call to vp8_save_coding_context.
+     */
+
+    cpi->frames_since_key         =   cc->frames_since_key;
+    cpi->common.filter_level     =   cc->filter_level;
+    cpi->frames_till_gf_update_due  =   cc->frames_till_gf_update_due;
+    cpi->frames_since_golden       =   cc->frames_since_golden;
+
+    vp8_copy(cpi->common.fc.mvc, cc->mvc);
+
+    vp8_copy(cpi->rd_costs.mvcosts, cc->mvcosts);
+
+    vp8_copy(cpi->common.fc.ymode_prob,   cc->ymode_prob);
+    vp8_copy(cpi->common.fc.uv_mode_prob,  cc->uv_mode_prob);
+
+    vp8_copy(cpi->mb.ymode_count, cc->ymode_count);
+    vp8_copy(cpi->mb.uv_mode_count, cc->uv_mode_count);
+
+    /* Stats */
+#ifdef MODE_STATS
+    vp8_copy(y_modes, cc->y_modes);
+    vp8_copy(uv_modes, cc->uv_modes);
+    vp8_copy(b_modes, cc->b_modes);
+    vp8_copy(inter_y_modes, cc->inter_y_modes);
+    vp8_copy(inter_uv_modes, cc->inter_uv_modes);
+    vp8_copy(inter_b_modes, cc->inter_b_modes);
+#endif
+
+
+    cpi->this_frame_percent_intra = cc->this_frame_percent_intra;
+}
+
+
+void vp8_setup_key_frame(VP8_COMP *cpi)
+{
+    /* Setup for Key frame: */
+
+    vp8_default_coef_probs(& cpi->common);
+
+    memcpy(cpi->common.fc.mvc, vp8_default_mv_context, sizeof(vp8_default_mv_context));
+    {
+        int flag[2] = {1, 1};
+        vp8_build_component_cost_table(cpi->mb.mvcost, (const MV_CONTEXT *) cpi->common.fc.mvc, flag);
+    }
+
+    /* Make sure we initialize separate contexts for altref,gold, and normal.
+     * TODO shouldn't need 3 different copies of structure to do this!
+     */
+    memcpy(&cpi->lfc_a, &cpi->common.fc, sizeof(cpi->common.fc));
+    memcpy(&cpi->lfc_g, &cpi->common.fc, sizeof(cpi->common.fc));
+    memcpy(&cpi->lfc_n, &cpi->common.fc, sizeof(cpi->common.fc));
+
+    cpi->common.filter_level = cpi->common.base_qindex * 3 / 8 ;
+
+    /* Provisional interval before next GF */
+    if (cpi->auto_gold)
+        cpi->frames_till_gf_update_due = cpi->baseline_gf_interval;
+    else
+        cpi->frames_till_gf_update_due = DEFAULT_GF_INTERVAL;
+
+    cpi->common.refresh_golden_frame = 1;
+    cpi->common.refresh_alt_ref_frame = 1;
+}
+
+
+static int estimate_bits_at_q(int frame_kind, int Q, int MBs,
+                              double correction_factor)
+{
+    int Bpm = (int)(.5 + correction_factor * vp8_bits_per_mb[frame_kind][Q]);
+
+    /* Attempt to retain reasonable accuracy without overflow. The cutoff is
+     * chosen such that the maximum product of Bpm and MBs fits 31 bits. The
+     * largest Bpm takes 20 bits.
+     */
+    if (MBs > (1 << 11))
+        return (Bpm >> BPER_MB_NORMBITS) * MBs;
+    else
+        return (Bpm * MBs) >> BPER_MB_NORMBITS;
+}
+
+
+static void calc_iframe_target_size(VP8_COMP *cpi)
+{
+    /* boost defaults to half second */
+    int kf_boost;
+    uint64_t target;
+
+    /* Clear down mmx registers to allow floating point in what follows */
+    vp8_clear_system_state();
+
+    if (cpi->oxcf.fixed_q >= 0)
+    {
+        int Q = cpi->oxcf.key_q;
+
+        target = estimate_bits_at_q(INTRA_FRAME, Q, cpi->common.MBs,
+                                    cpi->key_frame_rate_correction_factor);
+    }
+    else if (cpi->pass == 2)
+    {
+        /* New Two pass RC */
+        target = cpi->per_frame_bandwidth;
+    }
+    /* First Frame is a special case */
+    else if (cpi->common.current_video_frame == 0)
+    {
+        /* 1 Pass there is no information on which to base size so use
+         * bandwidth per second * fraction of the initial buffer
+         * level
+         */
+        target = cpi->oxcf.starting_buffer_level / 2;
+
+        if(target > cpi->oxcf.target_bandwidth * 3 / 2)
+            target = cpi->oxcf.target_bandwidth * 3 / 2;
+    }
+    else
+    {
+        /* if this keyframe was forced, use a more recent Q estimate */
+        int Q = (cpi->common.frame_flags & FRAMEFLAGS_KEY)
+                ? cpi->avg_frame_qindex : cpi->ni_av_qi;
+
+        int initial_boost = 32; /* |3.0 * per_frame_bandwidth| */
+        /* Boost depends somewhat on frame rate: only used for 1 layer case. */
+        if (cpi->oxcf.number_of_layers == 1) {
+          kf_boost = VPXMAX(initial_boost,
+                            (int)(2 * cpi->output_framerate - 16));
+        }
+        else {
+          /* Initial factor: set target size to: |3.0 * per_frame_bandwidth|. */
+          kf_boost = initial_boost;
+        }
+
+        /* adjustment up based on q: this factor ranges from ~1.2 to 2.2. */
+        kf_boost = kf_boost * kf_boost_qadjustment[Q] / 100;
+
+        /* frame separation adjustment ( down) */
+        if (cpi->frames_since_key  < cpi->output_framerate / 2)
+            kf_boost = (int)(kf_boost
+                       * cpi->frames_since_key / (cpi->output_framerate / 2));
+
+        /* Minimal target size is |2* per_frame_bandwidth|. */
+        if (kf_boost < 16)
+            kf_boost = 16;
+
+        target = ((16 + kf_boost) * cpi->per_frame_bandwidth) >> 4;
+    }
+
+
+    if (cpi->oxcf.rc_max_intra_bitrate_pct)
+    {
+        unsigned int max_rate = cpi->per_frame_bandwidth
+                                * cpi->oxcf.rc_max_intra_bitrate_pct / 100;
+
+        if (target > max_rate)
+            target = max_rate;
+    }
+
+    cpi->this_frame_target = (int)target;
+
+    /* TODO: if we separate rate targeting from Q targetting, move this.
+     * Reset the active worst quality to the baseline value for key frames.
+     */
+    if (cpi->pass != 2)
+        cpi->active_worst_quality = cpi->worst_quality;
+
+#if 0
+    {
+        FILE *f;
+
+        f = fopen("kf_boost.stt", "a");
+        fprintf(f, " %8u %10d %10d %10d\n",
+                cpi->common.current_video_frame,  cpi->gfu_boost, cpi->baseline_gf_interval, cpi->source_alt_ref_pending);
+
+        fclose(f);
+    }
+#endif
+}
+
+
+/* Do the best we can to define the parameters for the next GF based on what
+ * information we have available.
+ */
+static void calc_gf_params(VP8_COMP *cpi)
+{
+    int Q = (cpi->oxcf.fixed_q < 0) ? cpi->last_q[INTER_FRAME] : cpi->oxcf.fixed_q;
+    int Boost = 0;
+
+    int gf_frame_useage = 0;      /* Golden frame useage since last GF */
+    int tot_mbs = cpi->recent_ref_frame_usage[INTRA_FRAME]  +
+                  cpi->recent_ref_frame_usage[LAST_FRAME]   +
+                  cpi->recent_ref_frame_usage[GOLDEN_FRAME] +
+                  cpi->recent_ref_frame_usage[ALTREF_FRAME];
+
+    int pct_gf_active = (100 * cpi->gf_active_count) / (cpi->common.mb_rows * cpi->common.mb_cols);
+
+    if (tot_mbs)
+        gf_frame_useage = (cpi->recent_ref_frame_usage[GOLDEN_FRAME] + cpi->recent_ref_frame_usage[ALTREF_FRAME]) * 100 / tot_mbs;
+
+    if (pct_gf_active > gf_frame_useage)
+        gf_frame_useage = pct_gf_active;
+
+    /* Not two pass */
+    if (cpi->pass != 2)
+    {
+        /* Single Pass lagged mode: TBD */
+        if (0)
+        {
+        }
+
+        /* Single Pass compression: Has to use current and historical data */
+        else
+        {
+#if 0
+            /* Experimental code */
+            int index = cpi->one_pass_frame_index;
+            int frames_to_scan = (cpi->max_gf_interval <= MAX_LAG_BUFFERS) ? cpi->max_gf_interval : MAX_LAG_BUFFERS;
+
+            /* ************** Experimental code - incomplete */
+            /*
+            double decay_val = 1.0;
+            double IIAccumulator = 0.0;
+            double last_iiaccumulator = 0.0;
+            double IIRatio;
+
+            cpi->one_pass_frame_index = cpi->common.current_video_frame%MAX_LAG_BUFFERS;
+
+            for ( i = 0; i < (frames_to_scan - 1); i++ )
+            {
+                if ( index < 0 )
+                    index = MAX_LAG_BUFFERS;
+                index --;
+
+                if ( cpi->one_pass_frame_stats[index].frame_coded_error > 0.0 )
+                {
+                    IIRatio = cpi->one_pass_frame_stats[index].frame_intra_error / cpi->one_pass_frame_stats[index].frame_coded_error;
+
+                    if ( IIRatio > 30.0 )
+                        IIRatio = 30.0;
+                }
+                else
+                    IIRatio = 30.0;
+
+                IIAccumulator += IIRatio * decay_val;
+
+                decay_val = decay_val * cpi->one_pass_frame_stats[index].frame_pcnt_inter;
+
+                if (    (i > MIN_GF_INTERVAL) &&
+                        ((IIAccumulator - last_iiaccumulator) < 2.0) )
+                {
+                    break;
+                }
+                last_iiaccumulator = IIAccumulator;
+            }
+
+            Boost = IIAccumulator*100.0/16.0;
+            cpi->baseline_gf_interval = i;
+
+            */
+#else
+
+            /*************************************************************/
+            /* OLD code */
+
+            /* Adjust boost based upon ambient Q */
+            Boost = GFQ_ADJUSTMENT;
+
+            /* Adjust based upon most recently measure intra useage */
+            Boost = Boost * gf_intra_usage_adjustment[(cpi->this_frame_percent_intra < 15) ? cpi->this_frame_percent_intra : 14] / 100;
+
+            /* Adjust gf boost based upon GF usage since last GF */
+            Boost = Boost * gf_adjust_table[gf_frame_useage] / 100;
+#endif
+        }
+
+        /* golden frame boost without recode loop often goes awry.  be
+         * safe by keeping numbers down.
+         */
+        if (!cpi->sf.recode_loop)
+        {
+            if (cpi->compressor_speed == 2)
+                Boost = Boost / 2;
+        }
+
+        /* Apply an upper limit based on Q for 1 pass encodes */
+        if (Boost > kf_gf_boost_qlimits[Q] && (cpi->pass == 0))
+            Boost = kf_gf_boost_qlimits[Q];
+
+        /* Apply lower limits to boost. */
+        else if (Boost < 110)
+            Boost = 110;
+
+        /* Note the boost used */
+        cpi->last_boost = Boost;
+
+    }
+
+    /* Estimate next interval
+     * This is updated once the real frame size/boost is known.
+     */
+    if (cpi->oxcf.fixed_q == -1)
+    {
+        if (cpi->pass == 2)         /* 2 Pass */
+        {
+            cpi->frames_till_gf_update_due = cpi->baseline_gf_interval;
+        }
+        else                            /* 1 Pass */
+        {
+            cpi->frames_till_gf_update_due = cpi->baseline_gf_interval;
+
+            if (cpi->last_boost > 750)
+                cpi->frames_till_gf_update_due++;
+
+            if (cpi->last_boost > 1000)
+                cpi->frames_till_gf_update_due++;
+
+            if (cpi->last_boost > 1250)
+                cpi->frames_till_gf_update_due++;
+
+            if (cpi->last_boost >= 1500)
+                cpi->frames_till_gf_update_due ++;
+
+            if (gf_interval_table[gf_frame_useage] > cpi->frames_till_gf_update_due)
+                cpi->frames_till_gf_update_due = gf_interval_table[gf_frame_useage];
+
+            if (cpi->frames_till_gf_update_due > cpi->max_gf_interval)
+                cpi->frames_till_gf_update_due = cpi->max_gf_interval;
+        }
+    }
+    else
+        cpi->frames_till_gf_update_due = cpi->baseline_gf_interval;
+
+    /* ARF on or off */
+    if (cpi->pass != 2)
+    {
+        /* For now Alt ref is not allowed except in 2 pass modes. */
+        cpi->source_alt_ref_pending = 0;
+
+        /*if ( cpi->oxcf.fixed_q == -1)
+        {
+            if ( cpi->oxcf.play_alternate && (cpi->last_boost > (100 + (AF_THRESH*cpi->frames_till_gf_update_due)) ) )
+                cpi->source_alt_ref_pending = 1;
+            else
+                cpi->source_alt_ref_pending = 0;
+        }*/
+    }
+}
+
+
+static void calc_pframe_target_size(VP8_COMP *cpi)
+{
+    int min_frame_target;
+    int old_per_frame_bandwidth = cpi->per_frame_bandwidth;
+
+    if ( cpi->current_layer > 0)
+        cpi->per_frame_bandwidth =
+            cpi->layer_context[cpi->current_layer].avg_frame_size_for_layer;
+
+    min_frame_target = 0;
+
+    if (cpi->pass == 2)
+    {
+        min_frame_target = cpi->min_frame_bandwidth;
+
+        if (min_frame_target < (cpi->av_per_frame_bandwidth >> 5))
+            min_frame_target = cpi->av_per_frame_bandwidth >> 5;
+    }
+    else if (min_frame_target < cpi->per_frame_bandwidth / 4)
+        min_frame_target = cpi->per_frame_bandwidth / 4;
+
+
+    /* Special alt reference frame case */
+    if((cpi->common.refresh_alt_ref_frame) && (cpi->oxcf.number_of_layers == 1))
+    {
+        if (cpi->pass == 2)
+        {
+            /* Per frame bit target for the alt ref frame */
+            cpi->per_frame_bandwidth = cpi->twopass.gf_bits;
+            cpi->this_frame_target = cpi->per_frame_bandwidth;
+        }
+
+        /* One Pass ??? TBD */
+    }
+
+    /* Normal frames (gf,and inter) */
+    else
+    {
+        /* 2 pass */
+        if (cpi->pass == 2)
+        {
+            cpi->this_frame_target = cpi->per_frame_bandwidth;
+        }
+        /* 1 pass */
+        else
+        {
+            int Adjustment;
+            /* Make rate adjustment to recover bits spent in key frame
+             * Test to see if the key frame inter data rate correction
+             * should still be in force
+             */
+            if (cpi->kf_overspend_bits > 0)
+            {
+                Adjustment = (cpi->kf_bitrate_adjustment <= cpi->kf_overspend_bits) ? cpi->kf_bitrate_adjustment : cpi->kf_overspend_bits;
+
+                if (Adjustment > (cpi->per_frame_bandwidth - min_frame_target))
+                    Adjustment = (cpi->per_frame_bandwidth - min_frame_target);
+
+                cpi->kf_overspend_bits -= Adjustment;
+
+                /* Calculate an inter frame bandwidth target for the next
+                 * few frames designed to recover any extra bits spent on
+                 * the key frame.
+                 */
+                cpi->this_frame_target = cpi->per_frame_bandwidth - Adjustment;
+
+                if (cpi->this_frame_target < min_frame_target)
+                    cpi->this_frame_target = min_frame_target;
+            }
+            else
+                cpi->this_frame_target = cpi->per_frame_bandwidth;
+
+            /* If appropriate make an adjustment to recover bits spent on a
+             * recent GF
+             */
+            if ((cpi->gf_overspend_bits > 0) && (cpi->this_frame_target > min_frame_target))
+            {
+                Adjustment = (cpi->non_gf_bitrate_adjustment <= cpi->gf_overspend_bits) ? cpi->non_gf_bitrate_adjustment : cpi->gf_overspend_bits;
+
+                if (Adjustment > (cpi->this_frame_target - min_frame_target))
+                    Adjustment = (cpi->this_frame_target - min_frame_target);
+
+                cpi->gf_overspend_bits -= Adjustment;
+                cpi->this_frame_target -= Adjustment;
+            }
+
+            /* Apply small + and - boosts for non gf frames */
+            if ((cpi->last_boost > 150) && (cpi->frames_till_gf_update_due > 0) &&
+                (cpi->current_gf_interval >= (MIN_GF_INTERVAL << 1)))
+            {
+                /* % Adjustment limited to the range 1% to 10% */
+                Adjustment = (cpi->last_boost - 100) >> 5;
+
+                if (Adjustment < 1)
+                    Adjustment = 1;
+                else if (Adjustment > 10)
+                    Adjustment = 10;
+
+                /* Convert to bits */
+                Adjustment = (cpi->this_frame_target * Adjustment) / 100;
+
+                if (Adjustment > (cpi->this_frame_target - min_frame_target))
+                    Adjustment = (cpi->this_frame_target - min_frame_target);
+
+                if (cpi->frames_since_golden == (cpi->current_gf_interval >> 1))
+                {
+                    Adjustment = (cpi->current_gf_interval - 1) * Adjustment;
+                    // Limit adjustment to 10% of current target.
+                    if (Adjustment > (10 * cpi->this_frame_target) / 100)
+                        Adjustment = (10 * cpi->this_frame_target) / 100;
+                    cpi->this_frame_target += Adjustment;
+                }
+                else
+                    cpi->this_frame_target -= Adjustment;
+            }
+        }
+    }
+
+    /* Sanity check that the total sum of adjustments is not above the
+     * maximum allowed That is that having allowed for KF and GF penalties
+     * we have not pushed the current interframe target to low. If the
+     * adjustment we apply here is not capable of recovering all the extra
+     * bits we have spent in the KF or GF then the remainder will have to
+     * be recovered over a longer time span via other buffer / rate control
+     * mechanisms.
+     */
+    if (cpi->this_frame_target < min_frame_target)
+        cpi->this_frame_target = min_frame_target;
+
+    if (!cpi->common.refresh_alt_ref_frame)
+        /* Note the baseline target data rate for this inter frame. */
+        cpi->inter_frame_target = cpi->this_frame_target;
+
+    /* One Pass specific code */
+    if (cpi->pass == 0)
+    {
+        /* Adapt target frame size with respect to any buffering constraints: */
+        if (cpi->buffered_mode)
+        {
+            int one_percent_bits = (int)
+                (1 + cpi->oxcf.optimal_buffer_level / 100);
+
+            if ((cpi->buffer_level < cpi->oxcf.optimal_buffer_level) ||
+                (cpi->bits_off_target < cpi->oxcf.optimal_buffer_level))
+            {
+                int percent_low = 0;
+
+                /* Decide whether or not we need to adjust the frame data
+                 * rate target.
+                 *
+                 * If we are are below the optimal buffer fullness level
+                 * and adherence to buffering constraints is important to
+                 * the end usage then adjust the per frame target.
+                 */
+                if ((cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) &&
+                    (cpi->buffer_level < cpi->oxcf.optimal_buffer_level))
+                {
+                    percent_low = (int)
+                        ((cpi->oxcf.optimal_buffer_level - cpi->buffer_level) /
+                        one_percent_bits);
+                }
+                /* Are we overshooting the long term clip data rate... */
+                else if (cpi->bits_off_target < 0)
+                {
+                    /* Adjust per frame data target downwards to compensate. */
+                    percent_low = (int)(100 * -cpi->bits_off_target /
+                                       (cpi->total_byte_count * 8));
+                }
+
+                if (percent_low > cpi->oxcf.under_shoot_pct)
+                    percent_low = cpi->oxcf.under_shoot_pct;
+                else if (percent_low < 0)
+                    percent_low = 0;
+
+                /* lower the target bandwidth for this frame. */
+                cpi->this_frame_target -=
+                        (cpi->this_frame_target * percent_low) / 200;
+
+                /* Are we using allowing control of active_worst_allowed_q
+                 * according to buffer level.
+                 */
+                if (cpi->auto_worst_q && cpi->ni_frames > 150)
+                {
+                    int64_t critical_buffer_level;
+
+                    /* For streaming applications the most important factor is
+                     * cpi->buffer_level as this takes into account the
+                     * specified short term buffering constraints. However,
+                     * hitting the long term clip data rate target is also
+                     * important.
+                     */
+                    if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+                    {
+                        /* Take the smaller of cpi->buffer_level and
+                         * cpi->bits_off_target
+                         */
+                        critical_buffer_level =
+                            (cpi->buffer_level < cpi->bits_off_target)
+                            ? cpi->buffer_level : cpi->bits_off_target;
+                    }
+                    /* For local file playback short term buffering constraints
+                     * are less of an issue
+                     */
+                    else
+                    {
+                        /* Consider only how we are doing for the clip as a
+                         * whole
+                         */
+                        critical_buffer_level = cpi->bits_off_target;
+                    }
+
+                    /* Set the active worst quality based upon the selected
+                     * buffer fullness number.
+                     */
+                    if (critical_buffer_level < cpi->oxcf.optimal_buffer_level)
+                    {
+                        if ( critical_buffer_level >
+                             (cpi->oxcf.optimal_buffer_level >> 2) )
+                        {
+                            int64_t qadjustment_range =
+                                      cpi->worst_quality - cpi->ni_av_qi;
+                            int64_t above_base =
+                                      (critical_buffer_level -
+                                       (cpi->oxcf.optimal_buffer_level >> 2));
+
+                            /* Step active worst quality down from
+                             * cpi->ni_av_qi when (critical_buffer_level ==
+                             * cpi->optimal_buffer_level) to
+                             * cpi->worst_quality when
+                             * (critical_buffer_level ==
+                             *     cpi->optimal_buffer_level >> 2)
+                             */
+                            cpi->active_worst_quality =
+                                cpi->worst_quality -
+                                (int)((qadjustment_range * above_base) /
+                                 (cpi->oxcf.optimal_buffer_level*3>>2));
+                        }
+                        else
+                        {
+                            cpi->active_worst_quality = cpi->worst_quality;
+                        }
+                    }
+                    else
+                    {
+                        cpi->active_worst_quality = cpi->ni_av_qi;
+                    }
+                }
+                else
+                {
+                    cpi->active_worst_quality = cpi->worst_quality;
+                }
+            }
+            else
+            {
+                int percent_high = 0;
+
+                if ((cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+                     && (cpi->buffer_level > cpi->oxcf.optimal_buffer_level))
+                {
+                    percent_high = (int)((cpi->buffer_level
+                                    - cpi->oxcf.optimal_buffer_level)
+                                   / one_percent_bits);
+                }
+                else if (cpi->bits_off_target > cpi->oxcf.optimal_buffer_level)
+                {
+                    percent_high = (int)((100 * cpi->bits_off_target)
+                                         / (cpi->total_byte_count * 8));
+                }
+
+                if (percent_high > cpi->oxcf.over_shoot_pct)
+                    percent_high = cpi->oxcf.over_shoot_pct;
+                else if (percent_high < 0)
+                    percent_high = 0;
+
+                cpi->this_frame_target += (cpi->this_frame_target *
+                                          percent_high) / 200;
+
+                /* Are we allowing control of active_worst_allowed_q according
+                 * to buffer level.
+                 */
+                if (cpi->auto_worst_q && cpi->ni_frames > 150)
+                {
+                    /* When using the relaxed buffer model stick to the
+                     * user specified value
+                     */
+                    cpi->active_worst_quality = cpi->ni_av_qi;
+                }
+                else
+                {
+                    cpi->active_worst_quality = cpi->worst_quality;
+                }
+            }
+
+            /* Set active_best_quality to prevent quality rising too high */
+            cpi->active_best_quality = cpi->best_quality;
+
+            /* Worst quality obviously must not be better than best quality */
+            if (cpi->active_worst_quality <= cpi->active_best_quality)
+                cpi->active_worst_quality = cpi->active_best_quality + 1;
+
+            if(cpi->active_worst_quality > 127)
+                cpi->active_worst_quality = 127;
+        }
+        /* Unbuffered mode (eg. video conferencing) */
+        else
+        {
+            /* Set the active worst quality */
+            cpi->active_worst_quality = cpi->worst_quality;
+        }
+
+        /* Special trap for constrained quality mode
+         * "active_worst_quality" may never drop below cq level
+         * for any frame type.
+         */
+        if ( cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY &&
+             cpi->active_worst_quality < cpi->cq_target_quality)
+        {
+            cpi->active_worst_quality = cpi->cq_target_quality;
+        }
+    }
+
+    /* Test to see if we have to drop a frame
+     * The auto-drop frame code is only used in buffered mode.
+     * In unbufferd mode (eg vide conferencing) the descision to
+     * code or drop a frame is made outside the codec in response to real
+     * world comms or buffer considerations.
+     */
+    if (cpi->drop_frames_allowed &&
+        (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER) &&
+        ((cpi->common.frame_type != KEY_FRAME)))
+    {
+        /* Check for a buffer underun-crisis in which case we have to drop
+         * a frame
+         */
+        if ((cpi->buffer_level < 0))
+        {
+#if 0
+            FILE *f = fopen("dec.stt", "a");
+            fprintf(f, "%10d %10d %10d %10d ***** BUFFER EMPTY\n",
+                    (int) cpi->common.current_video_frame,
+                    cpi->decimation_factor, cpi->common.horiz_scale,
+                    (cpi->buffer_level * 100) / cpi->oxcf.optimal_buffer_level);
+            fclose(f);
+#endif
+            cpi->drop_frame = 1;
+
+            /* Update the buffer level variable. */
+            cpi->bits_off_target += cpi->av_per_frame_bandwidth;
+            if (cpi->bits_off_target > cpi->oxcf.maximum_buffer_size)
+              cpi->bits_off_target = (int)cpi->oxcf.maximum_buffer_size;
+            cpi->buffer_level = cpi->bits_off_target;
+
+            if (cpi->oxcf.number_of_layers > 1) {
+              unsigned int i;
+
+              // Propagate bits saved by dropping the frame to higher layers.
+              for (i = cpi->current_layer + 1; i < cpi->oxcf.number_of_layers;
+                  i++) {
+                LAYER_CONTEXT *lc = &cpi->layer_context[i];
+                lc->bits_off_target += (int)(lc->target_bandwidth /
+                                             lc->framerate);
+                if (lc->bits_off_target > lc->maximum_buffer_size)
+                  lc->bits_off_target = lc->maximum_buffer_size;
+                lc->buffer_level = lc->bits_off_target;
+              }
+            }
+        }
+    }
+
+    /* Adjust target frame size for Golden Frames: */
+    if (cpi->oxcf.error_resilient_mode == 0 &&
+        (cpi->frames_till_gf_update_due == 0) && !cpi->drop_frame)
+    {
+        int Q = (cpi->oxcf.fixed_q < 0) ? cpi->last_q[INTER_FRAME] : cpi->oxcf.fixed_q;
+
+        int gf_frame_useage = 0;      /* Golden frame useage since last GF */
+        int tot_mbs = cpi->recent_ref_frame_usage[INTRA_FRAME]  +
+                      cpi->recent_ref_frame_usage[LAST_FRAME]   +
+                      cpi->recent_ref_frame_usage[GOLDEN_FRAME] +
+                      cpi->recent_ref_frame_usage[ALTREF_FRAME];
+
+        int pct_gf_active = (100 * cpi->gf_active_count) / (cpi->common.mb_rows * cpi->common.mb_cols);
+
+        if (tot_mbs)
+            gf_frame_useage = (cpi->recent_ref_frame_usage[GOLDEN_FRAME] + cpi->recent_ref_frame_usage[ALTREF_FRAME]) * 100 / tot_mbs;
+
+        if (pct_gf_active > gf_frame_useage)
+            gf_frame_useage = pct_gf_active;
+
+        /* Is a fixed manual GF frequency being used */
+        if (cpi->auto_gold)
+        {
+            /* For one pass throw a GF if recent frame intra useage is
+             * low or the GF useage is high
+             */
+            if ((cpi->pass == 0) && (cpi->this_frame_percent_intra < 15 || gf_frame_useage >= 5))
+                cpi->common.refresh_golden_frame = 1;
+
+            /* Two pass GF descision */
+            else if (cpi->pass == 2)
+                cpi->common.refresh_golden_frame = 1;
+        }
+
+#if 0
+
+        /* Debug stats */
+        if (0)
+        {
+            FILE *f;
+
+            f = fopen("gf_useaget.stt", "a");
+            fprintf(f, " %8ld %10ld %10ld %10ld %10ld\n",
+                    cpi->common.current_video_frame,  cpi->gfu_boost, GFQ_ADJUSTMENT, cpi->gfu_boost, gf_frame_useage);
+            fclose(f);
+        }
+
+#endif
+
+        if (cpi->common.refresh_golden_frame == 1)
+        {
+#if 0
+
+            if (0)
+            {
+                FILE *f;
+
+                f = fopen("GFexit.stt", "a");
+                fprintf(f, "%8ld GF coded\n", cpi->common.current_video_frame);
+                fclose(f);
+            }
+
+#endif
+
+            if (cpi->auto_adjust_gold_quantizer)
+            {
+                calc_gf_params(cpi);
+            }
+
+            /* If we are using alternate ref instead of gf then do not apply the
+             * boost It will instead be applied to the altref update Jims
+             * modified boost
+             */
+            if (!cpi->source_alt_ref_active)
+            {
+                if (cpi->oxcf.fixed_q < 0)
+                {
+                    if (cpi->pass == 2)
+                    {
+                        /* The spend on the GF is defined in the two pass
+                         * code for two pass encodes
+                         */
+                        cpi->this_frame_target = cpi->per_frame_bandwidth;
+                    }
+                    else
+                    {
+                        int Boost = cpi->last_boost;
+                        int frames_in_section = cpi->frames_till_gf_update_due + 1;
+                        int allocation_chunks = (frames_in_section * 100) + (Boost - 100);
+                        int bits_in_section = cpi->inter_frame_target * frames_in_section;
+
+                        /* Normalize Altboost and allocations chunck down to
+                         * prevent overflow
+                         */
+                        while (Boost > 1000)
+                        {
+                            Boost /= 2;
+                            allocation_chunks /= 2;
+                        }
+
+                        /* Avoid loss of precision but avoid overflow */
+                        if ((bits_in_section >> 7) > allocation_chunks)
+                            cpi->this_frame_target = Boost * (bits_in_section / allocation_chunks);
+                        else
+                            cpi->this_frame_target = (Boost * bits_in_section) / allocation_chunks;
+                    }
+                }
+                else
+                    cpi->this_frame_target =
+                        (estimate_bits_at_q(1, Q, cpi->common.MBs, 1.0)
+                         * cpi->last_boost) / 100;
+
+            }
+            /* If there is an active ARF at this location use the minimum
+             * bits on this frame even if it is a contructed arf.
+             * The active maximum quantizer insures that an appropriate
+             * number of bits will be spent if needed for contstructed ARFs.
+             */
+            else
+            {
+                cpi->this_frame_target = 0;
+            }
+
+            cpi->current_gf_interval = cpi->frames_till_gf_update_due;
+
+        }
+    }
+
+    cpi->per_frame_bandwidth = old_per_frame_bandwidth;
+}
+
+
+void vp8_update_rate_correction_factors(VP8_COMP *cpi, int damp_var)
+{
+    int    Q = cpi->common.base_qindex;
+    int    correction_factor = 100;
+    double rate_correction_factor;
+    double adjustment_limit;
+
+    int    projected_size_based_on_q = 0;
+
+    /* Clear down mmx registers to allow floating point in what follows */
+    vp8_clear_system_state();
+
+    if (cpi->common.frame_type == KEY_FRAME)
+    {
+        rate_correction_factor = cpi->key_frame_rate_correction_factor;
+    }
+    else
+    {
+        if (cpi->oxcf.number_of_layers == 1 &&
+           (cpi->common.refresh_alt_ref_frame ||
+            cpi->common.refresh_golden_frame))
+            rate_correction_factor = cpi->gf_rate_correction_factor;
+        else
+            rate_correction_factor = cpi->rate_correction_factor;
+    }
+
+    /* Work out how big we would have expected the frame to be at this Q
+     * given the current correction factor. Stay in double to avoid int
+     * overflow when values are large
+     */
+    projected_size_based_on_q = (int)(((.5 + rate_correction_factor * vp8_bits_per_mb[cpi->common.frame_type][Q]) * cpi->common.MBs) / (1 << BPER_MB_NORMBITS));
+
+    /* Make some allowance for cpi->zbin_over_quant */
+    if (cpi->mb.zbin_over_quant > 0)
+    {
+        int Z = cpi->mb.zbin_over_quant;
+        double Factor = 0.99;
+        double factor_adjustment = 0.01 / 256.0;
+
+        while (Z > 0)
+        {
+            Z --;
+            projected_size_based_on_q =
+                (int)(Factor * projected_size_based_on_q);
+            Factor += factor_adjustment;
+
+            if (Factor  >= 0.999)
+                Factor = 0.999;
+        }
+    }
+
+    /* Work out a size correction factor. */
+    if (projected_size_based_on_q > 0)
+        correction_factor = (100 * cpi->projected_frame_size) / projected_size_based_on_q;
+
+    /* More heavily damped adjustment used if we have been oscillating
+     * either side of target
+     */
+    switch (damp_var)
+    {
+    case 0:
+        adjustment_limit = 0.75;
+        break;
+    case 1:
+        adjustment_limit = 0.375;
+        break;
+    case 2:
+    default:
+        adjustment_limit = 0.25;
+        break;
+    }
+
+    if (correction_factor > 102)
+    {
+        /* We are not already at the worst allowable quality */
+        correction_factor = (int)(100.5 + ((correction_factor - 100) * adjustment_limit));
+        rate_correction_factor = ((rate_correction_factor * correction_factor) / 100);
+
+        /* Keep rate_correction_factor within limits */
+        if (rate_correction_factor > MAX_BPB_FACTOR)
+            rate_correction_factor = MAX_BPB_FACTOR;
+    }
+    else if (correction_factor < 99)
+    {
+        /* We are not already at the best allowable quality */
+        correction_factor = (int)(100.5 - ((100 - correction_factor) * adjustment_limit));
+        rate_correction_factor = ((rate_correction_factor * correction_factor) / 100);
+
+        /* Keep rate_correction_factor within limits */
+        if (rate_correction_factor < MIN_BPB_FACTOR)
+            rate_correction_factor = MIN_BPB_FACTOR;
+    }
+
+    if (cpi->common.frame_type == KEY_FRAME)
+        cpi->key_frame_rate_correction_factor = rate_correction_factor;
+    else
+    {
+        if (cpi->oxcf.number_of_layers == 1 &&
+           (cpi->common.refresh_alt_ref_frame ||
+            cpi->common.refresh_golden_frame))
+            cpi->gf_rate_correction_factor = rate_correction_factor;
+        else
+            cpi->rate_correction_factor = rate_correction_factor;
+    }
+}
+
+
+int vp8_regulate_q(VP8_COMP *cpi, int target_bits_per_frame)
+{
+    int Q = cpi->active_worst_quality;
+
+    if (cpi->force_maxqp == 1) {
+      cpi->active_worst_quality = cpi->worst_quality;
+      return cpi->worst_quality;
+    }
+
+    /* Reset Zbin OQ value */
+    cpi->mb.zbin_over_quant = 0;
+
+    if (cpi->oxcf.fixed_q >= 0)
+    {
+        Q = cpi->oxcf.fixed_q;
+
+        if (cpi->common.frame_type == KEY_FRAME)
+        {
+            Q = cpi->oxcf.key_q;
+        }
+        else if (cpi->oxcf.number_of_layers == 1 &&
+            cpi->common.refresh_alt_ref_frame)
+        {
+            Q = cpi->oxcf.alt_q;
+        }
+        else if (cpi->oxcf.number_of_layers == 1  &&
+            cpi->common.refresh_golden_frame)
+        {
+            Q = cpi->oxcf.gold_q;
+        }
+    }
+    else
+    {
+        int i;
+        int last_error = INT_MAX;
+        int target_bits_per_mb;
+        int bits_per_mb_at_this_q;
+        double correction_factor;
+
+        /* Select the appropriate correction factor based upon type of frame. */
+        if (cpi->common.frame_type == KEY_FRAME)
+            correction_factor = cpi->key_frame_rate_correction_factor;
+        else
+        {
+            if (cpi->oxcf.number_of_layers == 1 &&
+               (cpi->common.refresh_alt_ref_frame ||
+                cpi->common.refresh_golden_frame))
+                correction_factor = cpi->gf_rate_correction_factor;
+            else
+                correction_factor = cpi->rate_correction_factor;
+        }
+
+        /* Calculate required scaling factor based on target frame size and
+         * size of frame produced using previous Q
+         */
+        if (target_bits_per_frame >= (INT_MAX >> BPER_MB_NORMBITS))
+            /* Case where we would overflow int */
+            target_bits_per_mb = (target_bits_per_frame / cpi->common.MBs) << BPER_MB_NORMBITS;
+        else
+            target_bits_per_mb = (target_bits_per_frame << BPER_MB_NORMBITS) / cpi->common.MBs;
+
+        i = cpi->active_best_quality;
+
+        do
+        {
+            bits_per_mb_at_this_q = (int)(.5 + correction_factor * vp8_bits_per_mb[cpi->common.frame_type][i]);
+
+            if (bits_per_mb_at_this_q <= target_bits_per_mb)
+            {
+                if ((target_bits_per_mb - bits_per_mb_at_this_q) <= last_error)
+                    Q = i;
+                else
+                    Q = i - 1;
+
+                break;
+            }
+            else
+                last_error = bits_per_mb_at_this_q - target_bits_per_mb;
+        }
+        while (++i <= cpi->active_worst_quality);
+
+
+        /* If we are at MAXQ then enable Q over-run which seeks to claw
+         * back additional bits through things like the RD multiplier
+         * and zero bin size.
+         */
+        if (Q >= MAXQ)
+        {
+            int zbin_oqmax;
+
+            double Factor = 0.99;
+            double factor_adjustment = 0.01 / 256.0;
+
+            if (cpi->common.frame_type == KEY_FRAME)
+                zbin_oqmax = 0;
+            else if (cpi->oxcf.number_of_layers == 1 &&
+                (cpi->common.refresh_alt_ref_frame ||
+                (cpi->common.refresh_golden_frame &&
+                 !cpi->source_alt_ref_active)))
+                zbin_oqmax = 16;
+            else
+                zbin_oqmax = ZBIN_OQ_MAX;
+
+            /*{
+                double Factor = (double)target_bits_per_mb/(double)bits_per_mb_at_this_q;
+                double Oq;
+
+                Factor = Factor/1.2683;
+
+                Oq = pow( Factor, (1.0/-0.165) );
+
+                if ( Oq > zbin_oqmax )
+                    Oq = zbin_oqmax;
+
+                cpi->zbin_over_quant = (int)Oq;
+            }*/
+
+            /* Each incrment in the zbin is assumed to have a fixed effect
+             * on bitrate. This is not of course true. The effect will be
+             * highly clip dependent and may well have sudden steps. The
+             * idea here is to acheive higher effective quantizers than the
+             * normal maximum by expanding the zero bin and hence
+             * decreasing the number of low magnitude non zero coefficients.
+             */
+            while (cpi->mb.zbin_over_quant < zbin_oqmax)
+            {
+                cpi->mb.zbin_over_quant ++;
+
+                if (cpi->mb.zbin_over_quant > zbin_oqmax)
+                    cpi->mb.zbin_over_quant = zbin_oqmax;
+
+                /* Adjust bits_per_mb_at_this_q estimate */
+                bits_per_mb_at_this_q = (int)(Factor * bits_per_mb_at_this_q);
+                Factor += factor_adjustment;
+
+                if (Factor  >= 0.999)
+                    Factor = 0.999;
+
+                /* Break out if we get down to the target rate */
+                if (bits_per_mb_at_this_q <= target_bits_per_mb)
+                    break;
+            }
+
+        }
+    }
+
+    return Q;
+}
+
+
+static int estimate_keyframe_frequency(VP8_COMP *cpi)
+{
+    int i;
+
+    /* Average key frame frequency */
+    int av_key_frame_frequency = 0;
+
+    /* First key frame at start of sequence is a special case. We have no
+     * frequency data.
+     */
+    if (cpi->key_frame_count == 1)
+    {
+        /* Assume a default of 1 kf every 2 seconds, or the max kf interval,
+         * whichever is smaller.
+         */
+        int key_freq = cpi->oxcf.key_freq>0 ? cpi->oxcf.key_freq : 1;
+        av_key_frame_frequency = 1 + (int)cpi->output_framerate * 2;
+
+        if (cpi->oxcf.auto_key && av_key_frame_frequency > key_freq)
+            av_key_frame_frequency = key_freq;
+
+        cpi->prior_key_frame_distance[KEY_FRAME_CONTEXT - 1]
+            = av_key_frame_frequency;
+    }
+    else
+    {
+        unsigned int total_weight = 0;
+        int last_kf_interval =
+                (cpi->frames_since_key > 0) ? cpi->frames_since_key : 1;
+
+        /* reset keyframe context and calculate weighted average of last
+         * KEY_FRAME_CONTEXT keyframes
+         */
+        for (i = 0; i < KEY_FRAME_CONTEXT; i++)
+        {
+            if (i < KEY_FRAME_CONTEXT - 1)
+                cpi->prior_key_frame_distance[i]
+                    = cpi->prior_key_frame_distance[i+1];
+            else
+                cpi->prior_key_frame_distance[i] = last_kf_interval;
+
+            av_key_frame_frequency += prior_key_frame_weight[i]
+                                      * cpi->prior_key_frame_distance[i];
+            total_weight += prior_key_frame_weight[i];
+        }
+
+        av_key_frame_frequency  /= total_weight;
+
+    }
+    // TODO (marpan): Given the checks above, |av_key_frame_frequency|
+    // should always be above 0. But for now we keep the sanity check in.
+    if (av_key_frame_frequency == 0)
+        av_key_frame_frequency = 1;
+    return av_key_frame_frequency;
+}
+
+
+void vp8_adjust_key_frame_context(VP8_COMP *cpi)
+{
+    /* Clear down mmx registers to allow floating point in what follows */
+    vp8_clear_system_state();
+
+    /* Do we have any key frame overspend to recover? */
+    /* Two-pass overspend handled elsewhere. */
+    if ((cpi->pass != 2)
+         && (cpi->projected_frame_size > cpi->per_frame_bandwidth))
+    {
+        int overspend;
+
+        /* Update the count of key frame overspend to be recovered in
+         * subsequent frames. A portion of the KF overspend is treated as gf
+         * overspend (and hence recovered more quickly) as the kf is also a
+         * gf. Otherwise the few frames following each kf tend to get more
+         * bits allocated than those following other gfs.
+         */
+        overspend = (cpi->projected_frame_size - cpi->per_frame_bandwidth);
+
+        if (cpi->oxcf.number_of_layers > 1)
+            cpi->kf_overspend_bits += overspend;
+        else
+        {
+            cpi->kf_overspend_bits += overspend * 7 / 8;
+            cpi->gf_overspend_bits += overspend * 1 / 8;
+        }
+
+        /* Work out how much to try and recover per frame. */
+        cpi->kf_bitrate_adjustment = cpi->kf_overspend_bits
+                                     / estimate_keyframe_frequency(cpi);
+    }
+
+    cpi->frames_since_key = 0;
+    cpi->key_frame_count++;
+}
+
+
+void vp8_compute_frame_size_bounds(VP8_COMP *cpi, int *frame_under_shoot_limit, int *frame_over_shoot_limit)
+{
+    /* Set-up bounds on acceptable frame size: */
+    if (cpi->oxcf.fixed_q >= 0)
+    {
+        /* Fixed Q scenario: frame size never outranges target
+         * (there is no target!)
+         */
+        *frame_under_shoot_limit = 0;
+        *frame_over_shoot_limit  = INT_MAX;
+    }
+    else
+    {
+        if (cpi->common.frame_type == KEY_FRAME)
+        {
+            *frame_over_shoot_limit  = cpi->this_frame_target * 9 / 8;
+            *frame_under_shoot_limit = cpi->this_frame_target * 7 / 8;
+        }
+        else
+        {
+            if (cpi->oxcf.number_of_layers > 1 ||
+                cpi->common.refresh_alt_ref_frame ||
+                cpi->common.refresh_golden_frame)
+            {
+                *frame_over_shoot_limit  = cpi->this_frame_target * 9 / 8;
+                *frame_under_shoot_limit = cpi->this_frame_target * 7 / 8;
+            }
+            else
+            {
+                /* For CBR take buffer fullness into account */
+                if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
+                {
+                    if (cpi->buffer_level >= ((cpi->oxcf.optimal_buffer_level + cpi->oxcf.maximum_buffer_size) >> 1))
+                    {
+                        /* Buffer is too full so relax overshoot and tighten
+                         * undershoot
+                         */
+                        *frame_over_shoot_limit  = cpi->this_frame_target * 12 / 8;
+                        *frame_under_shoot_limit = cpi->this_frame_target * 6 / 8;
+                    }
+                    else if (cpi->buffer_level <= (cpi->oxcf.optimal_buffer_level >> 1))
+                    {
+                        /* Buffer is too low so relax undershoot and tighten
+                         * overshoot
+                         */
+                        *frame_over_shoot_limit  = cpi->this_frame_target * 10 / 8;
+                        *frame_under_shoot_limit = cpi->this_frame_target * 4 / 8;
+                    }
+                    else
+                    {
+                        *frame_over_shoot_limit  = cpi->this_frame_target * 11 / 8;
+                        *frame_under_shoot_limit = cpi->this_frame_target * 5 / 8;
+                    }
+                }
+                /* VBR and CQ mode */
+                /* Note that tighter restrictions here can help quality
+                 * but hurt encode speed
+                 */
+                else
+                {
+                    /* Stron overshoot limit for constrained quality */
+                    if (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY)
+                    {
+                        *frame_over_shoot_limit  = cpi->this_frame_target * 11 / 8;
+                        *frame_under_shoot_limit = cpi->this_frame_target * 2 / 8;
+                    }
+                    else
+                    {
+                        *frame_over_shoot_limit  = cpi->this_frame_target * 11 / 8;
+                        *frame_under_shoot_limit = cpi->this_frame_target * 5 / 8;
+                    }
+                }
+            }
+        }
+
+        /* For very small rate targets where the fractional adjustment
+         * (eg * 7/8) may be tiny make sure there is at least a minimum
+         * range.
+         */
+        *frame_over_shoot_limit += 200;
+        *frame_under_shoot_limit -= 200;
+        if ( *frame_under_shoot_limit < 0 )
+            *frame_under_shoot_limit = 0;
+
+    }
+}
+
+
+/* return of 0 means drop frame */
+int vp8_pick_frame_size(VP8_COMP *cpi)
+{
+    VP8_COMMON *cm = &cpi->common;
+
+    if (cm->frame_type == KEY_FRAME)
+        calc_iframe_target_size(cpi);
+    else
+    {
+        calc_pframe_target_size(cpi);
+
+        /* Check if we're dropping the frame: */
+        if (cpi->drop_frame)
+        {
+            cpi->drop_frame = 0;
+            return 0;
+        }
+    }
+    return 1;
+}
+// If this just encoded frame (mcomp/transform/quant, but before loopfilter and
+// pack_bitstream) has large overshoot, and was not being encoded close to the
+// max QP, then drop this frame and force next frame to be encoded at max QP.
+// Condition this on 1 pass CBR with screen content mode and frame dropper off.
+// TODO(marpan): Should do this exit condition during the encode_frame
+// (i.e., halfway during the encoding of the frame) to save cycles.
+int vp8_drop_encodedframe_overshoot(VP8_COMP *cpi, int Q) {
+  if (cpi->pass == 0 &&
+      cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER &&
+      cpi->drop_frames_allowed == 0 &&
+      cpi->common.frame_type != KEY_FRAME) {
+    // Note: the "projected_frame_size" from encode_frame() only gives estimate
+    // of mode/motion vector rate (in non-rd mode): so below we only require
+    // that projected_frame_size is somewhat greater than per-frame-bandwidth,
+    // but add additional condition with high threshold on prediction residual.
+
+    // QP threshold: only allow dropping if we are not close to qp_max.
+    int thresh_qp = 3 * cpi->worst_quality >> 2;
+    // Rate threshold, in bytes.
+    int thresh_rate = 2 * (cpi->av_per_frame_bandwidth >> 3);
+    // Threshold for the average (over all macroblocks) of the pixel-sum
+    // residual error over 16x16 block. Should add QP dependence on threshold?
+    int thresh_pred_err_mb = (256 << 4);
+    int pred_err_mb = (int)(cpi->mb.prediction_error / cpi->common.MBs);
+    if (Q < thresh_qp &&
+        cpi->projected_frame_size > thresh_rate &&
+        pred_err_mb > thresh_pred_err_mb) {
+      double new_correction_factor = cpi->rate_correction_factor;
+      const int target_size = cpi->av_per_frame_bandwidth;
+      int target_bits_per_mb;
+      // Drop this frame: advance frame counters, and set force_maxqp flag.
+      cpi->common.current_video_frame++;
+      cpi->frames_since_key++;
+      // Flag to indicate we will force next frame to be encoded at max QP.
+      cpi->force_maxqp = 1;
+      // Reset the buffer levels.
+      cpi->buffer_level = cpi->oxcf.optimal_buffer_level;
+      cpi->bits_off_target = cpi->oxcf.optimal_buffer_level;
+      // Compute a new rate correction factor, corresponding to the current
+      // target frame size and max_QP, and adjust the rate correction factor
+      // upwards, if needed.
+      // This is to prevent a bad state where the re-encoded frame at max_QP
+      // undershoots significantly, and then we end up dropping every other
+      // frame because the QP/rate_correction_factor may have been too low
+      // before the drop and then takes too long to come up.
+      if (target_size >= (INT_MAX >> BPER_MB_NORMBITS))
+        target_bits_per_mb =
+            (target_size / cpi->common.MBs) << BPER_MB_NORMBITS;
+      else
+        target_bits_per_mb =
+            (target_size << BPER_MB_NORMBITS) / cpi->common.MBs;
+      // Rate correction factor based on target_size_per_mb and max_QP.
+      new_correction_factor = (double)target_bits_per_mb /
+          (double)vp8_bits_per_mb[INTER_FRAME][cpi->worst_quality];
+      if (new_correction_factor > cpi->rate_correction_factor)
+        cpi->rate_correction_factor =
+            VPXMIN(2.0 * cpi->rate_correction_factor, new_correction_factor);
+      if (cpi->rate_correction_factor > MAX_BPB_FACTOR)
+        cpi->rate_correction_factor = MAX_BPB_FACTOR;
+      return 1;
+    } else {
+      cpi->force_maxqp = 0;
+      return 0;
+    }
+    cpi->force_maxqp = 0;
+    return 0;
+  }
+  cpi->force_maxqp = 0;
+  return 0;
+}
diff --git a/libvpx/libvpx/vp8/encoder/ratectrl.h b/libvpx/libvpx/vp8/encoder/ratectrl.h
new file mode 100644
index 0000000..703de9f
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/ratectrl.h
@@ -0,0 +1,39 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_RATECTRL_H_
+#define VP8_ENCODER_RATECTRL_H_
+
+#include "onyx_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern void vp8_save_coding_context(VP8_COMP *cpi);
+extern void vp8_restore_coding_context(VP8_COMP *cpi);
+
+extern void vp8_setup_key_frame(VP8_COMP *cpi);
+extern void vp8_update_rate_correction_factors(VP8_COMP *cpi, int damp_var);
+extern int vp8_regulate_q(VP8_COMP *cpi, int target_bits_per_frame);
+extern void vp8_adjust_key_frame_context(VP8_COMP *cpi);
+extern void vp8_compute_frame_size_bounds(VP8_COMP *cpi, int *frame_under_shoot_limit, int *frame_over_shoot_limit);
+
+/* return of 0 means drop frame */
+extern int vp8_pick_frame_size(VP8_COMP *cpi);
+
+extern int vp8_drop_encodedframe_overshoot(VP8_COMP *cpi, int Q);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_RATECTRL_H_
diff --git a/libvpx/libvpx/vp8/encoder/rdopt.c b/libvpx/libvpx/vp8/encoder/rdopt.c
new file mode 100644
index 0000000..6507ae9
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/rdopt.c
@@ -0,0 +1,2646 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <stdio.h>
+#include <math.h>
+#include <limits.h>
+#include <assert.h>
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "tokenize.h"
+#include "treewriter.h"
+#include "onyx_int.h"
+#include "modecosts.h"
+#include "encodeintra.h"
+#include "pickinter.h"
+#include "vp8/common/entropymode.h"
+#include "vp8/common/reconinter.h"
+#include "vp8/common/reconintra.h"
+#include "vp8/common/reconintra4x4.h"
+#include "vp8/common/findnearmv.h"
+#include "vp8/common/quant_common.h"
+#include "encodemb.h"
+#include "vp8/encoder/quantize.h"
+#include "vpx_dsp/variance.h"
+#include "mcomp.h"
+#include "rdopt.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp8/common/systemdependent.h"
+#if CONFIG_TEMPORAL_DENOISING
+#include "denoising.h"
+#endif
+extern void vp8_update_zbin_extra(VP8_COMP *cpi, MACROBLOCK *x);
+
+#define MAXF(a,b)            (((a) > (b)) ? (a) : (b))
+
+typedef struct rate_distortion_struct
+{
+    int rate2;
+    int rate_y;
+    int rate_uv;
+    int distortion2;
+    int distortion_uv;
+} RATE_DISTORTION;
+
+typedef struct best_mode_struct
+{
+  int yrd;
+  int rd;
+  int intra_rd;
+  MB_MODE_INFO mbmode;
+  union b_mode_info bmodes[16];
+  PARTITION_INFO partition;
+} BEST_MODE;
+
+static const int auto_speed_thresh[17] =
+{
+    1000,
+    200,
+    150,
+    130,
+    150,
+    125,
+    120,
+    115,
+    115,
+    115,
+    115,
+    115,
+    115,
+    115,
+    115,
+    115,
+    105
+};
+
+const MB_PREDICTION_MODE vp8_mode_order[MAX_MODES] =
+{
+    ZEROMV,
+    DC_PRED,
+
+    NEARESTMV,
+    NEARMV,
+
+    ZEROMV,
+    NEARESTMV,
+
+    ZEROMV,
+    NEARESTMV,
+
+    NEARMV,
+    NEARMV,
+
+    V_PRED,
+    H_PRED,
+    TM_PRED,
+
+    NEWMV,
+    NEWMV,
+    NEWMV,
+
+    SPLITMV,
+    SPLITMV,
+    SPLITMV,
+
+    B_PRED,
+};
+
+/* This table determines the search order in reference frame priority order,
+ * which may not necessarily match INTRA,LAST,GOLDEN,ARF
+ */
+const int vp8_ref_frame_order[MAX_MODES] =
+{
+    1,
+    0,
+
+    1,
+    1,
+
+    2,
+    2,
+
+    3,
+    3,
+
+    2,
+    3,
+
+    0,
+    0,
+    0,
+
+    1,
+    2,
+    3,
+
+    1,
+    2,
+    3,
+
+    0,
+};
+
+static void fill_token_costs(
+    int c[BLOCK_TYPES][COEF_BANDS][PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS],
+    const vp8_prob p[BLOCK_TYPES][COEF_BANDS][PREV_COEF_CONTEXTS][ENTROPY_NODES]
+)
+{
+    int i, j, k;
+
+
+    for (i = 0; i < BLOCK_TYPES; i++)
+        for (j = 0; j < COEF_BANDS; j++)
+            for (k = 0; k < PREV_COEF_CONTEXTS; k++)
+
+                /* check for pt=0 and band > 1 if block type 0
+                 * and 0 if blocktype 1
+                 */
+                if (k == 0 && j > (i == 0))
+                    vp8_cost_tokens2(c[i][j][k], p [i][j][k], vp8_coef_tree, 2);
+                else
+                    vp8_cost_tokens(c[i][j][k], p [i][j][k], vp8_coef_tree);
+}
+
+static const int rd_iifactor[32] =
+{
+    4, 4, 3, 2, 1, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0
+};
+
+/* values are now correlated to quantizer */
+static const int sad_per_bit16lut[QINDEX_RANGE] =
+{
+    2,  2,  2,  2,  2,  2,  2,  2,
+    2,  2,  2,  2,  2,  2,  2,  2,
+    3,  3,  3,  3,  3,  3,  3,  3,
+    3,  3,  3,  3,  3,  3,  4,  4,
+    4,  4,  4,  4,  4,  4,  4,  4,
+    4,  4,  5,  5,  5,  5,  5,  5,
+    5,  5,  5,  5,  5,  5,  6,  6,
+    6,  6,  6,  6,  6,  6,  6,  6,
+    6,  6,  7,  7,  7,  7,  7,  7,
+    7,  7,  7,  7,  7,  7,  8,  8,
+    8,  8,  8,  8,  8,  8,  8,  8,
+    8,  8,  9,  9,  9,  9,  9,  9,
+    9,  9,  9,  9,  9,  9,  10, 10,
+    10, 10, 10, 10, 10, 10, 11, 11,
+    11, 11, 11, 11, 12, 12, 12, 12,
+    12, 12, 13, 13, 13, 13, 14, 14
+};
+static const int sad_per_bit4lut[QINDEX_RANGE] =
+{
+    2,  2,  2,  2,  2,  2,  3,  3,
+    3,  3,  3,  3,  3,  3,  3,  3,
+    3,  3,  3,  3,  4,  4,  4,  4,
+    4,  4,  4,  4,  4,  4,  5,  5,
+    5,  5,  5,  5,  6,  6,  6,  6,
+    6,  6,  6,  6,  6,  6,  6,  6,
+    7,  7,  7,  7,  7,  7,  7,  7,
+    7,  7,  7,  7,  7,  8,  8,  8,
+    8,  8,  9,  9,  9,  9,  9,  9,
+    10, 10, 10, 10, 10, 10, 10, 10,
+    11, 11, 11, 11, 11, 11, 11, 11,
+    12, 12, 12, 12, 12, 12, 12, 12,
+    13, 13, 13, 13, 13, 13, 13, 14,
+    14, 14, 14, 14, 15, 15, 15, 15,
+    16, 16, 16, 16, 17, 17, 17, 18,
+    18, 18, 19, 19, 19, 20, 20, 20,
+};
+
+void vp8cx_initialize_me_consts(VP8_COMP *cpi, int QIndex)
+{
+    cpi->mb.sadperbit16 =  sad_per_bit16lut[QIndex];
+    cpi->mb.sadperbit4  =  sad_per_bit4lut[QIndex];
+}
+
+void vp8_initialize_rd_consts(VP8_COMP *cpi, MACROBLOCK *x, int Qvalue)
+{
+    int q;
+    int i;
+    double capped_q = (Qvalue < 160) ? (double)Qvalue : 160.0;
+    double rdconst = 2.80;
+
+    vp8_clear_system_state();
+
+    /* Further tests required to see if optimum is different
+     * for key frames, golden frames and arf frames.
+     */
+    cpi->RDMULT = (int)(rdconst * (capped_q * capped_q));
+
+    /* Extend rate multiplier along side quantizer zbin increases */
+    if (cpi->mb.zbin_over_quant  > 0)
+    {
+        double oq_factor;
+        double modq;
+
+        /* Experimental code using the same basic equation as used for Q above
+         * The units of cpi->mb.zbin_over_quant are 1/128 of Q bin size
+         */
+        oq_factor = 1.0 + ((double)0.0015625 * cpi->mb.zbin_over_quant);
+        modq = (int)((double)capped_q * oq_factor);
+        cpi->RDMULT = (int)(rdconst * (modq * modq));
+    }
+
+    if (cpi->pass == 2 && (cpi->common.frame_type != KEY_FRAME))
+    {
+        if (cpi->twopass.next_iiratio > 31)
+            cpi->RDMULT += (cpi->RDMULT * rd_iifactor[31]) >> 4;
+        else
+            cpi->RDMULT +=
+                (cpi->RDMULT * rd_iifactor[cpi->twopass.next_iiratio]) >> 4;
+    }
+
+    cpi->mb.errorperbit = (cpi->RDMULT / 110);
+    cpi->mb.errorperbit += (cpi->mb.errorperbit==0);
+
+    vp8_set_speed_features(cpi);
+
+    for (i = 0; i < MAX_MODES; i++)
+    {
+        x->mode_test_hit_counts[i] = 0;
+    }
+
+    q = (int)pow(Qvalue, 1.25);
+
+    if (q < 8)
+        q = 8;
+
+    if (cpi->RDMULT > 1000)
+    {
+        cpi->RDDIV = 1;
+        cpi->RDMULT /= 100;
+
+        for (i = 0; i < MAX_MODES; i++)
+        {
+            if (cpi->sf.thresh_mult[i] < INT_MAX)
+            {
+                x->rd_threshes[i] = cpi->sf.thresh_mult[i] * q / 100;
+            }
+            else
+            {
+                x->rd_threshes[i] = INT_MAX;
+            }
+
+            cpi->rd_baseline_thresh[i] = x->rd_threshes[i];
+        }
+    }
+    else
+    {
+        cpi->RDDIV = 100;
+
+        for (i = 0; i < MAX_MODES; i++)
+        {
+            if (cpi->sf.thresh_mult[i] < (INT_MAX / q))
+            {
+                x->rd_threshes[i] = cpi->sf.thresh_mult[i] * q;
+            }
+            else
+            {
+                x->rd_threshes[i] = INT_MAX;
+            }
+
+            cpi->rd_baseline_thresh[i] = x->rd_threshes[i];
+        }
+    }
+
+    {
+      /* build token cost array for the type of frame we have now */
+      FRAME_CONTEXT *l = &cpi->lfc_n;
+
+      if(cpi->common.refresh_alt_ref_frame)
+          l = &cpi->lfc_a;
+      else if(cpi->common.refresh_golden_frame)
+          l = &cpi->lfc_g;
+
+      fill_token_costs(
+          cpi->mb.token_costs,
+          (const vp8_prob( *)[8][3][11]) l->coef_probs
+      );
+      /*
+      fill_token_costs(
+          cpi->mb.token_costs,
+          (const vp8_prob( *)[8][3][11]) cpi->common.fc.coef_probs);
+      */
+
+
+      /* TODO make these mode costs depend on last,alt or gold too.  (jbb) */
+      vp8_init_mode_costs(cpi);
+    }
+
+}
+
+void vp8_auto_select_speed(VP8_COMP *cpi)
+{
+    int milliseconds_for_compress = (int)(1000000 / cpi->framerate);
+
+    milliseconds_for_compress = milliseconds_for_compress * (16 - cpi->oxcf.cpu_used) / 16;
+
+#if 0
+
+    if (0)
+    {
+        FILE *f;
+
+        f = fopen("speed.stt", "a");
+        fprintf(f, " %8ld %10ld %10ld %10ld\n",
+                cpi->common.current_video_frame, cpi->Speed, milliseconds_for_compress, cpi->avg_pick_mode_time);
+        fclose(f);
+    }
+
+#endif
+
+    if (cpi->avg_pick_mode_time < milliseconds_for_compress && (cpi->avg_encode_time - cpi->avg_pick_mode_time) < milliseconds_for_compress)
+    {
+        if (cpi->avg_pick_mode_time == 0)
+        {
+            cpi->Speed = 4;
+        }
+        else
+        {
+            if (milliseconds_for_compress * 100 < cpi->avg_encode_time * 95)
+            {
+                cpi->Speed          += 2;
+                cpi->avg_pick_mode_time = 0;
+                cpi->avg_encode_time = 0;
+
+                if (cpi->Speed > 16)
+                {
+                    cpi->Speed = 16;
+                }
+            }
+
+            if (milliseconds_for_compress * 100 > cpi->avg_encode_time * auto_speed_thresh[cpi->Speed])
+            {
+                cpi->Speed          -= 1;
+                cpi->avg_pick_mode_time = 0;
+                cpi->avg_encode_time = 0;
+
+                /* In real-time mode, cpi->speed is in [4, 16]. */
+                if (cpi->Speed < 4)
+                {
+                    cpi->Speed = 4;
+                }
+            }
+        }
+    }
+    else
+    {
+        cpi->Speed += 4;
+
+        if (cpi->Speed > 16)
+            cpi->Speed = 16;
+
+
+        cpi->avg_pick_mode_time = 0;
+        cpi->avg_encode_time = 0;
+    }
+}
+
+int vp8_block_error_c(short *coeff, short *dqcoeff)
+{
+    int i;
+    int error = 0;
+
+    for (i = 0; i < 16; i++)
+    {
+        int this_diff = coeff[i] - dqcoeff[i];
+        error += this_diff * this_diff;
+    }
+
+    return error;
+}
+
+int vp8_mbblock_error_c(MACROBLOCK *mb, int dc)
+{
+    BLOCK  *be;
+    BLOCKD *bd;
+    int i, j;
+    int berror, error = 0;
+
+    for (i = 0; i < 16; i++)
+    {
+        be = &mb->block[i];
+        bd = &mb->e_mbd.block[i];
+
+        berror = 0;
+
+        for (j = dc; j < 16; j++)
+        {
+            int this_diff = be->coeff[j] - bd->dqcoeff[j];
+            berror += this_diff * this_diff;
+        }
+
+        error += berror;
+    }
+
+    return error;
+}
+
+int vp8_mbuverror_c(MACROBLOCK *mb)
+{
+
+    BLOCK  *be;
+    BLOCKD *bd;
+
+
+    int i;
+    int error = 0;
+
+    for (i = 16; i < 24; i++)
+    {
+        be = &mb->block[i];
+        bd = &mb->e_mbd.block[i];
+
+        error += vp8_block_error_c(be->coeff, bd->dqcoeff);
+    }
+
+    return error;
+}
+
+int VP8_UVSSE(MACROBLOCK *x)
+{
+    unsigned char *uptr, *vptr;
+    unsigned char *upred_ptr = (*(x->block[16].base_src) + x->block[16].src);
+    unsigned char *vpred_ptr = (*(x->block[20].base_src) + x->block[20].src);
+    int uv_stride = x->block[16].src_stride;
+
+    unsigned int sse1 = 0;
+    unsigned int sse2 = 0;
+    int mv_row = x->e_mbd.mode_info_context->mbmi.mv.as_mv.row;
+    int mv_col = x->e_mbd.mode_info_context->mbmi.mv.as_mv.col;
+    int offset;
+    int pre_stride = x->e_mbd.pre.uv_stride;
+
+    if (mv_row < 0)
+        mv_row -= 1;
+    else
+        mv_row += 1;
+
+    if (mv_col < 0)
+        mv_col -= 1;
+    else
+        mv_col += 1;
+
+    mv_row /= 2;
+    mv_col /= 2;
+
+    offset = (mv_row >> 3) * pre_stride + (mv_col >> 3);
+    uptr = x->e_mbd.pre.u_buffer + offset;
+    vptr = x->e_mbd.pre.v_buffer + offset;
+
+    if ((mv_row | mv_col) & 7)
+    {
+        vpx_sub_pixel_variance8x8(uptr, pre_stride,
+            mv_col & 7, mv_row & 7, upred_ptr, uv_stride, &sse2);
+        vpx_sub_pixel_variance8x8(vptr, pre_stride,
+            mv_col & 7, mv_row & 7, vpred_ptr, uv_stride, &sse1);
+        sse2 += sse1;
+    }
+    else
+    {
+        vpx_variance8x8(uptr, pre_stride,
+            upred_ptr, uv_stride, &sse2);
+        vpx_variance8x8(vptr, pre_stride,
+            vpred_ptr, uv_stride, &sse1);
+        sse2 += sse1;
+    }
+    return sse2;
+
+}
+
+static int cost_coeffs(MACROBLOCK *mb, BLOCKD *b, int type, ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l)
+{
+    int c = !type;              /* start at coef 0, unless Y with Y2 */
+    int eob = (int)(*b->eob);
+    int pt ;    /* surrounding block/prev coef predictor */
+    int cost = 0;
+    short *qcoeff_ptr = b->qcoeff;
+
+    VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
+
+    assert(eob <= 16);
+    for (; c < eob; c++)
+    {
+        const int v = qcoeff_ptr[vp8_default_zig_zag1d[c]];
+        const int t = vp8_dct_value_tokens_ptr[v].Token;
+        cost += mb->token_costs [type] [vp8_coef_bands[c]] [pt] [t];
+        cost += vp8_dct_value_cost_ptr[v];
+        pt = vp8_prev_token_class[t];
+    }
+
+    if (c < 16)
+        cost += mb->token_costs [type] [vp8_coef_bands[c]] [pt] [DCT_EOB_TOKEN];
+
+    pt = (c != !type); /* is eob first coefficient; */
+    *a = *l = pt;
+
+    return cost;
+}
+
+static int vp8_rdcost_mby(MACROBLOCK *mb)
+{
+    int cost = 0;
+    int b;
+    MACROBLOCKD *x = &mb->e_mbd;
+    ENTROPY_CONTEXT_PLANES t_above, t_left;
+    ENTROPY_CONTEXT *ta;
+    ENTROPY_CONTEXT *tl;
+
+    memcpy(&t_above, mb->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
+    memcpy(&t_left, mb->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
+
+    ta = (ENTROPY_CONTEXT *)&t_above;
+    tl = (ENTROPY_CONTEXT *)&t_left;
+
+    for (b = 0; b < 16; b++)
+        cost += cost_coeffs(mb, x->block + b, PLANE_TYPE_Y_NO_DC,
+                    ta + vp8_block2above[b], tl + vp8_block2left[b]);
+
+    cost += cost_coeffs(mb, x->block + 24, PLANE_TYPE_Y2,
+                ta + vp8_block2above[24], tl + vp8_block2left[24]);
+
+    return cost;
+}
+
+static void macro_block_yrd( MACROBLOCK *mb,
+                             int *Rate,
+                             int *Distortion)
+{
+    int b;
+    MACROBLOCKD *const x = &mb->e_mbd;
+    BLOCK   *const mb_y2 = mb->block + 24;
+    BLOCKD *const x_y2  = x->block + 24;
+    short *Y2DCPtr = mb_y2->src_diff;
+    BLOCK *beptr;
+    int d;
+
+    vp8_subtract_mby( mb->src_diff, *(mb->block[0].base_src),
+        mb->block[0].src_stride,  mb->e_mbd.predictor, 16);
+
+    /* Fdct and building the 2nd order block */
+    for (beptr = mb->block; beptr < mb->block + 16; beptr += 2)
+    {
+        mb->short_fdct8x4(beptr->src_diff, beptr->coeff, 32);
+        *Y2DCPtr++ = beptr->coeff[0];
+        *Y2DCPtr++ = beptr->coeff[16];
+    }
+
+    /* 2nd order fdct */
+    mb->short_walsh4x4(mb_y2->src_diff, mb_y2->coeff, 8);
+
+    /* Quantization */
+    for (b = 0; b < 16; b++)
+    {
+        mb->quantize_b(&mb->block[b], &mb->e_mbd.block[b]);
+    }
+
+    /* DC predication and Quantization of 2nd Order block */
+    mb->quantize_b(mb_y2, x_y2);
+
+    /* Distortion */
+    d = vp8_mbblock_error(mb, 1) << 2;
+    d += vp8_block_error(mb_y2->coeff, x_y2->dqcoeff);
+
+    *Distortion = (d >> 4);
+
+    /* rate */
+    *Rate = vp8_rdcost_mby(mb);
+}
+
+static void copy_predictor(unsigned char *dst, const unsigned char *predictor)
+{
+    const unsigned int *p = (const unsigned int *)predictor;
+    unsigned int *d = (unsigned int *)dst;
+    d[0] = p[0];
+    d[4] = p[4];
+    d[8] = p[8];
+    d[12] = p[12];
+}
+static int rd_pick_intra4x4block(
+    MACROBLOCK *x,
+    BLOCK *be,
+    BLOCKD *b,
+    B_PREDICTION_MODE *best_mode,
+    const int *bmode_costs,
+    ENTROPY_CONTEXT *a,
+    ENTROPY_CONTEXT *l,
+
+    int *bestrate,
+    int *bestratey,
+    int *bestdistortion)
+{
+    B_PREDICTION_MODE mode;
+    int best_rd = INT_MAX;
+    int rate = 0;
+    int distortion;
+
+    ENTROPY_CONTEXT ta = *a, tempa = *a;
+    ENTROPY_CONTEXT tl = *l, templ = *l;
+    /*
+     * The predictor buffer is a 2d buffer with a stride of 16.  Create
+     * a temp buffer that meets the stride requirements, but we are only
+     * interested in the left 4x4 block
+     * */
+    DECLARE_ALIGNED(16, unsigned char,  best_predictor[16*4]);
+    DECLARE_ALIGNED(16, short, best_dqcoeff[16]);
+    int dst_stride = x->e_mbd.dst.y_stride;
+    unsigned char *dst = x->e_mbd.dst.y_buffer + b->offset;
+
+    unsigned char *Above = dst - dst_stride;
+    unsigned char *yleft = dst - 1;
+    unsigned char top_left = Above[-1];
+
+    for (mode = B_DC_PRED; mode <= B_HU_PRED; mode++)
+    {
+        int this_rd;
+        int ratey;
+
+        rate = bmode_costs[mode];
+
+        vp8_intra4x4_predict(Above, yleft, dst_stride, mode,
+                             b->predictor, 16, top_left);
+        vp8_subtract_b(be, b, 16);
+        x->short_fdct4x4(be->src_diff, be->coeff, 32);
+        x->quantize_b(be, b);
+
+        tempa = ta;
+        templ = tl;
+
+        ratey = cost_coeffs(x, b, PLANE_TYPE_Y_WITH_DC, &tempa, &templ);
+        rate += ratey;
+        distortion = vp8_block_error(be->coeff, b->dqcoeff) >> 2;
+
+        this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
+
+        if (this_rd < best_rd)
+        {
+            *bestrate = rate;
+            *bestratey = ratey;
+            *bestdistortion = distortion;
+            best_rd = this_rd;
+            *best_mode = mode;
+            *a = tempa;
+            *l = templ;
+            copy_predictor(best_predictor, b->predictor);
+            memcpy(best_dqcoeff, b->dqcoeff, 32);
+        }
+    }
+    b->bmi.as_mode = *best_mode;
+
+    vp8_short_idct4x4llm(best_dqcoeff, best_predictor, 16, dst, dst_stride);
+
+    return best_rd;
+}
+
+static int rd_pick_intra4x4mby_modes(MACROBLOCK *mb, int *Rate,
+                                     int *rate_y, int *Distortion, int best_rd)
+{
+    MACROBLOCKD *const xd = &mb->e_mbd;
+    int i;
+    int cost = mb->mbmode_cost [xd->frame_type] [B_PRED];
+    int distortion = 0;
+    int tot_rate_y = 0;
+    int64_t total_rd = 0;
+    ENTROPY_CONTEXT_PLANES t_above, t_left;
+    ENTROPY_CONTEXT *ta;
+    ENTROPY_CONTEXT *tl;
+    const int *bmode_costs;
+
+    memcpy(&t_above, mb->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
+    memcpy(&t_left, mb->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
+
+    ta = (ENTROPY_CONTEXT *)&t_above;
+    tl = (ENTROPY_CONTEXT *)&t_left;
+
+    intra_prediction_down_copy(xd, xd->dst.y_buffer - xd->dst.y_stride + 16);
+
+    bmode_costs = mb->inter_bmode_costs;
+
+    for (i = 0; i < 16; i++)
+    {
+        MODE_INFO *const mic = xd->mode_info_context;
+        const int mis = xd->mode_info_stride;
+        B_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode);
+        int UNINITIALIZED_IS_SAFE(r), UNINITIALIZED_IS_SAFE(ry), UNINITIALIZED_IS_SAFE(d);
+
+        if (mb->e_mbd.frame_type == KEY_FRAME)
+        {
+            const B_PREDICTION_MODE A = above_block_mode(mic, i, mis);
+            const B_PREDICTION_MODE L = left_block_mode(mic, i);
+
+            bmode_costs  = mb->bmode_costs[A][L];
+        }
+
+        total_rd += rd_pick_intra4x4block(
+            mb, mb->block + i, xd->block + i, &best_mode, bmode_costs,
+            ta + vp8_block2above[i],
+            tl + vp8_block2left[i], &r, &ry, &d);
+
+        cost += r;
+        distortion += d;
+        tot_rate_y += ry;
+
+        mic->bmi[i].as_mode = best_mode;
+
+        if(total_rd >= (int64_t)best_rd)
+            break;
+    }
+
+    if(total_rd >= (int64_t)best_rd)
+        return INT_MAX;
+
+    *Rate = cost;
+    *rate_y = tot_rate_y;
+    *Distortion = distortion;
+
+    return RDCOST(mb->rdmult, mb->rddiv, cost, distortion);
+}
+
+
+static int rd_pick_intra16x16mby_mode(MACROBLOCK *x,
+                                      int *Rate,
+                                      int *rate_y,
+                                      int *Distortion)
+{
+    MB_PREDICTION_MODE mode;
+    MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(mode_selected);
+    int rate, ratey;
+    int distortion;
+    int best_rd = INT_MAX;
+    int this_rd;
+    MACROBLOCKD *xd = &x->e_mbd;
+
+    /* Y Search for 16x16 intra prediction mode */
+    for (mode = DC_PRED; mode <= TM_PRED; mode++)
+    {
+        xd->mode_info_context->mbmi.mode = mode;
+
+        vp8_build_intra_predictors_mby_s(xd,
+                                         xd->dst.y_buffer - xd->dst.y_stride,
+                                         xd->dst.y_buffer - 1,
+                                         xd->dst.y_stride,
+                                         xd->predictor,
+                                         16);
+
+        macro_block_yrd(x, &ratey, &distortion);
+        rate = ratey + x->mbmode_cost[xd->frame_type]
+                                     [xd->mode_info_context->mbmi.mode];
+
+        this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
+
+        if (this_rd < best_rd)
+        {
+            mode_selected = mode;
+            best_rd = this_rd;
+            *Rate = rate;
+            *rate_y = ratey;
+            *Distortion = distortion;
+        }
+    }
+
+    xd->mode_info_context->mbmi.mode = mode_selected;
+    return best_rd;
+}
+
+static int rd_cost_mbuv(MACROBLOCK *mb)
+{
+    int b;
+    int cost = 0;
+    MACROBLOCKD *x = &mb->e_mbd;
+    ENTROPY_CONTEXT_PLANES t_above, t_left;
+    ENTROPY_CONTEXT *ta;
+    ENTROPY_CONTEXT *tl;
+
+    memcpy(&t_above, mb->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
+    memcpy(&t_left, mb->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
+
+    ta = (ENTROPY_CONTEXT *)&t_above;
+    tl = (ENTROPY_CONTEXT *)&t_left;
+
+    for (b = 16; b < 24; b++)
+        cost += cost_coeffs(mb, x->block + b, PLANE_TYPE_UV,
+                    ta + vp8_block2above[b], tl + vp8_block2left[b]);
+
+    return cost;
+}
+
+
+static int rd_inter16x16_uv(VP8_COMP *cpi, MACROBLOCK *x, int *rate,
+                            int *distortion, int fullpixel)
+{
+    (void)cpi;
+    (void)fullpixel;
+
+    vp8_build_inter16x16_predictors_mbuv(&x->e_mbd);
+    vp8_subtract_mbuv(x->src_diff,
+        x->src.u_buffer, x->src.v_buffer, x->src.uv_stride,
+        &x->e_mbd.predictor[256], &x->e_mbd.predictor[320], 8);
+
+    vp8_transform_mbuv(x);
+    vp8_quantize_mbuv(x);
+
+    *rate       = rd_cost_mbuv(x);
+    *distortion = vp8_mbuverror(x) / 4;
+
+    return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
+}
+
+static int rd_inter4x4_uv(VP8_COMP *cpi, MACROBLOCK *x, int *rate,
+                          int *distortion, int fullpixel)
+{
+    (void)cpi;
+    (void)fullpixel;
+
+    vp8_build_inter4x4_predictors_mbuv(&x->e_mbd);
+    vp8_subtract_mbuv(x->src_diff,
+        x->src.u_buffer, x->src.v_buffer, x->src.uv_stride,
+        &x->e_mbd.predictor[256], &x->e_mbd.predictor[320], 8);
+
+    vp8_transform_mbuv(x);
+    vp8_quantize_mbuv(x);
+
+    *rate       = rd_cost_mbuv(x);
+    *distortion = vp8_mbuverror(x) / 4;
+
+    return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
+}
+
+static void rd_pick_intra_mbuv_mode(MACROBLOCK *x, int *rate,
+                                    int *rate_tokenonly, int *distortion)
+{
+    MB_PREDICTION_MODE mode;
+    MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(mode_selected);
+    int best_rd = INT_MAX;
+    int UNINITIALIZED_IS_SAFE(d), UNINITIALIZED_IS_SAFE(r);
+    int rate_to;
+    MACROBLOCKD *xd = &x->e_mbd;
+
+    for (mode = DC_PRED; mode <= TM_PRED; mode++)
+    {
+        int this_rate;
+        int this_distortion;
+        int this_rd;
+
+        xd->mode_info_context->mbmi.uv_mode = mode;
+
+        vp8_build_intra_predictors_mbuv_s(xd,
+                                          xd->dst.u_buffer - xd->dst.uv_stride,
+                                          xd->dst.v_buffer - xd->dst.uv_stride,
+                                          xd->dst.u_buffer - 1,
+                                          xd->dst.v_buffer - 1,
+                                          xd->dst.uv_stride,
+                                          &xd->predictor[256], &xd->predictor[320],
+                                          8);
+
+
+        vp8_subtract_mbuv(x->src_diff,
+                      x->src.u_buffer, x->src.v_buffer, x->src.uv_stride,
+                      &xd->predictor[256], &xd->predictor[320], 8);
+        vp8_transform_mbuv(x);
+        vp8_quantize_mbuv(x);
+
+        rate_to = rd_cost_mbuv(x);
+        this_rate = rate_to + x->intra_uv_mode_cost[xd->frame_type][xd->mode_info_context->mbmi.uv_mode];
+
+        this_distortion = vp8_mbuverror(x) / 4;
+
+        this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
+
+        if (this_rd < best_rd)
+        {
+            best_rd = this_rd;
+            d = this_distortion;
+            r = this_rate;
+            *rate_tokenonly = rate_to;
+            mode_selected = mode;
+        }
+    }
+
+    *rate = r;
+    *distortion = d;
+
+    xd->mode_info_context->mbmi.uv_mode = mode_selected;
+}
+
+int vp8_cost_mv_ref(MB_PREDICTION_MODE m, const int near_mv_ref_ct[4])
+{
+    vp8_prob p [VP8_MVREFS-1];
+    assert(NEARESTMV <= m  &&  m <= SPLITMV);
+    vp8_mv_ref_probs(p, near_mv_ref_ct);
+    return vp8_cost_token(vp8_mv_ref_tree, p,
+                          vp8_mv_ref_encoding_array + (m - NEARESTMV));
+}
+
+void vp8_set_mbmode_and_mvs(MACROBLOCK *x, MB_PREDICTION_MODE mb, int_mv *mv)
+{
+    x->e_mbd.mode_info_context->mbmi.mode = mb;
+    x->e_mbd.mode_info_context->mbmi.mv.as_int = mv->as_int;
+}
+
+static int labels2mode(
+    MACROBLOCK *x,
+    int const *labelings, int which_label,
+    B_PREDICTION_MODE this_mode,
+    int_mv *this_mv, int_mv *best_ref_mv,
+    int *mvcost[2]
+)
+{
+    MACROBLOCKD *const xd = & x->e_mbd;
+    MODE_INFO *const mic = xd->mode_info_context;
+    const int mis = xd->mode_info_stride;
+
+    int cost = 0;
+    int thismvcost = 0;
+
+    /* We have to be careful retrieving previously-encoded motion vectors.
+       Ones from this macroblock have to be pulled from the BLOCKD array
+       as they have not yet made it to the bmi array in our MB_MODE_INFO. */
+
+    int i = 0;
+
+    do
+    {
+        BLOCKD *const d = xd->block + i;
+        const int row = i >> 2,  col = i & 3;
+
+        B_PREDICTION_MODE m;
+
+        if (labelings[i] != which_label)
+            continue;
+
+        if (col  &&  labelings[i] == labelings[i-1])
+            m = LEFT4X4;
+        else if (row  &&  labelings[i] == labelings[i-4])
+            m = ABOVE4X4;
+        else
+        {
+            /* the only time we should do costing for new motion vector
+             * or mode is when we are on a new label  (jbb May 08, 2007)
+             */
+            switch (m = this_mode)
+            {
+            case NEW4X4 :
+                thismvcost  = vp8_mv_bit_cost(this_mv, best_ref_mv, mvcost, 102);
+                break;
+            case LEFT4X4:
+                this_mv->as_int = col ? d[-1].bmi.mv.as_int : left_block_mv(mic, i);
+                break;
+            case ABOVE4X4:
+                this_mv->as_int = row ? d[-4].bmi.mv.as_int : above_block_mv(mic, i, mis);
+                break;
+            case ZERO4X4:
+                this_mv->as_int = 0;
+                break;
+            default:
+                break;
+            }
+
+            if (m == ABOVE4X4)  /* replace above with left if same */
+            {
+                int_mv left_mv;
+
+                left_mv.as_int = col ? d[-1].bmi.mv.as_int :
+                                        left_block_mv(mic, i);
+
+                if (left_mv.as_int == this_mv->as_int)
+                    m = LEFT4X4;
+            }
+
+            cost = x->inter_bmode_costs[ m];
+        }
+
+        d->bmi.mv.as_int = this_mv->as_int;
+
+        x->partition_info->bmi[i].mode = m;
+        x->partition_info->bmi[i].mv.as_int = this_mv->as_int;
+
+    }
+    while (++i < 16);
+
+    cost += thismvcost ;
+    return cost;
+}
+
+static int rdcost_mbsegment_y(MACROBLOCK *mb, const int *labels,
+                              int which_label, ENTROPY_CONTEXT *ta,
+                              ENTROPY_CONTEXT *tl)
+{
+    int cost = 0;
+    int b;
+    MACROBLOCKD *x = &mb->e_mbd;
+
+    for (b = 0; b < 16; b++)
+        if (labels[ b] == which_label)
+            cost += cost_coeffs(mb, x->block + b, PLANE_TYPE_Y_WITH_DC,
+                                ta + vp8_block2above[b],
+                                tl + vp8_block2left[b]);
+
+    return cost;
+
+}
+static unsigned int vp8_encode_inter_mb_segment(MACROBLOCK *x, int const *labels, int which_label)
+{
+    int i;
+    unsigned int distortion = 0;
+    int pre_stride = x->e_mbd.pre.y_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+
+
+    for (i = 0; i < 16; i++)
+    {
+        if (labels[i] == which_label)
+        {
+            BLOCKD *bd = &x->e_mbd.block[i];
+            BLOCK *be = &x->block[i];
+
+            vp8_build_inter_predictors_b(bd, 16, base_pre, pre_stride, x->e_mbd.subpixel_predict);
+            vp8_subtract_b(be, bd, 16);
+            x->short_fdct4x4(be->src_diff, be->coeff, 32);
+            x->quantize_b(be, bd);
+
+            distortion += vp8_block_error(be->coeff, bd->dqcoeff);
+        }
+    }
+
+    return distortion;
+}
+
+
+static const unsigned int segmentation_to_sseshift[4] = {3, 3, 2, 0};
+
+
+typedef struct
+{
+  int_mv *ref_mv;
+  int_mv mvp;
+
+  int segment_rd;
+  int segment_num;
+  int r;
+  int d;
+  int segment_yrate;
+  B_PREDICTION_MODE modes[16];
+  int_mv mvs[16];
+  unsigned char eobs[16];
+
+  int mvthresh;
+  int *mdcounts;
+
+  int_mv sv_mvp[4]; /* save 4 mvp from 8x8 */
+  int sv_istep[2];  /* save 2 initial step_param for 16x8/8x16 */
+
+} BEST_SEG_INFO;
+
+
+static void rd_check_segment(VP8_COMP *cpi, MACROBLOCK *x,
+                             BEST_SEG_INFO *bsi, unsigned int segmentation)
+{
+    int i;
+    int const *labels;
+    int br = 0;
+    int bd = 0;
+    B_PREDICTION_MODE this_mode;
+
+
+    int label_count;
+    int this_segment_rd = 0;
+    int label_mv_thresh;
+    int rate = 0;
+    int sbr = 0;
+    int sbd = 0;
+    int segmentyrate = 0;
+
+    vp8_variance_fn_ptr_t *v_fn_ptr;
+
+    ENTROPY_CONTEXT_PLANES t_above, t_left;
+    ENTROPY_CONTEXT *ta;
+    ENTROPY_CONTEXT *tl;
+    ENTROPY_CONTEXT_PLANES t_above_b, t_left_b;
+    ENTROPY_CONTEXT *ta_b;
+    ENTROPY_CONTEXT *tl_b;
+
+    memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
+    memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
+
+    ta = (ENTROPY_CONTEXT *)&t_above;
+    tl = (ENTROPY_CONTEXT *)&t_left;
+    ta_b = (ENTROPY_CONTEXT *)&t_above_b;
+    tl_b = (ENTROPY_CONTEXT *)&t_left_b;
+
+    br = 0;
+    bd = 0;
+
+    v_fn_ptr = &cpi->fn_ptr[segmentation];
+    labels = vp8_mbsplits[segmentation];
+    label_count = vp8_mbsplit_count[segmentation];
+
+    /* 64 makes this threshold really big effectively making it so that we
+     * very rarely check mvs on segments.   setting this to 1 would make mv
+     * thresh roughly equal to what it is for macroblocks
+     */
+    label_mv_thresh = 1 * bsi->mvthresh / label_count ;
+
+    /* Segmentation method overheads */
+    rate = vp8_cost_token(vp8_mbsplit_tree, vp8_mbsplit_probs, vp8_mbsplit_encodings + segmentation);
+    rate += vp8_cost_mv_ref(SPLITMV, bsi->mdcounts);
+    this_segment_rd += RDCOST(x->rdmult, x->rddiv, rate, 0);
+    br += rate;
+
+    for (i = 0; i < label_count; i++)
+    {
+        int_mv mode_mv[B_MODE_COUNT];
+        int best_label_rd = INT_MAX;
+        B_PREDICTION_MODE mode_selected = ZERO4X4;
+        int bestlabelyrate = 0;
+
+        /* search for the best motion vector on this segment */
+        for (this_mode = LEFT4X4; this_mode <= NEW4X4 ; this_mode ++)
+        {
+            int this_rd;
+            int distortion;
+            int labelyrate;
+            ENTROPY_CONTEXT_PLANES t_above_s, t_left_s;
+            ENTROPY_CONTEXT *ta_s;
+            ENTROPY_CONTEXT *tl_s;
+
+            memcpy(&t_above_s, &t_above, sizeof(ENTROPY_CONTEXT_PLANES));
+            memcpy(&t_left_s, &t_left, sizeof(ENTROPY_CONTEXT_PLANES));
+
+            ta_s = (ENTROPY_CONTEXT *)&t_above_s;
+            tl_s = (ENTROPY_CONTEXT *)&t_left_s;
+
+            if (this_mode == NEW4X4)
+            {
+                int sseshift;
+                int num00;
+                int step_param = 0;
+                int further_steps;
+                int n;
+                int thissme;
+                int bestsme = INT_MAX;
+                int_mv  temp_mv;
+                BLOCK *c;
+                BLOCKD *e;
+
+                /* Is the best so far sufficiently good that we cant justify
+                 * doing a new motion search.
+                 */
+                if (best_label_rd < label_mv_thresh)
+                    break;
+
+                if(cpi->compressor_speed)
+                {
+                    if (segmentation == BLOCK_8X16 || segmentation == BLOCK_16X8)
+                    {
+                        bsi->mvp.as_int = bsi->sv_mvp[i].as_int;
+                        if (i==1 && segmentation == BLOCK_16X8)
+                          bsi->mvp.as_int = bsi->sv_mvp[2].as_int;
+
+                        step_param = bsi->sv_istep[i];
+                    }
+
+                    /* use previous block's result as next block's MV
+                     * predictor.
+                     */
+                    if (segmentation == BLOCK_4X4 && i>0)
+                    {
+                        bsi->mvp.as_int = x->e_mbd.block[i-1].bmi.mv.as_int;
+                        if (i==4 || i==8 || i==12)
+                            bsi->mvp.as_int = x->e_mbd.block[i-4].bmi.mv.as_int;
+                        step_param = 2;
+                    }
+                }
+
+                further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
+
+                {
+                    int sadpb = x->sadperbit4;
+                    int_mv mvp_full;
+
+                    mvp_full.as_mv.row = bsi->mvp.as_mv.row >>3;
+                    mvp_full.as_mv.col = bsi->mvp.as_mv.col >>3;
+
+                    /* find first label */
+                    n = vp8_mbsplit_offset[segmentation][i];
+
+                    c = &x->block[n];
+                    e = &x->e_mbd.block[n];
+
+                    {
+                        bestsme = cpi->diamond_search_sad(x, c, e, &mvp_full,
+                                                &mode_mv[NEW4X4], step_param,
+                                                sadpb, &num00, v_fn_ptr,
+                                                x->mvcost, bsi->ref_mv);
+
+                        n = num00;
+                        num00 = 0;
+
+                        while (n < further_steps)
+                        {
+                            n++;
+
+                            if (num00)
+                                num00--;
+                            else
+                            {
+                                thissme = cpi->diamond_search_sad(x, c, e,
+                                                    &mvp_full, &temp_mv,
+                                                    step_param + n, sadpb,
+                                                    &num00, v_fn_ptr,
+                                                    x->mvcost, bsi->ref_mv);
+
+                                if (thissme < bestsme)
+                                {
+                                    bestsme = thissme;
+                                    mode_mv[NEW4X4].as_int = temp_mv.as_int;
+                                }
+                            }
+                        }
+                    }
+
+                    sseshift = segmentation_to_sseshift[segmentation];
+
+                    /* Should we do a full search (best quality only) */
+                    if ((cpi->compressor_speed == 0) && (bestsme >> sseshift) > 4000)
+                    {
+                        /* Check if mvp_full is within the range. */
+                        vp8_clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
+
+                        thissme = cpi->full_search_sad(x, c, e, &mvp_full,
+                                                       sadpb, 16, v_fn_ptr,
+                                                       x->mvcost, bsi->ref_mv);
+
+                        if (thissme < bestsme)
+                        {
+                            bestsme = thissme;
+                            mode_mv[NEW4X4].as_int = e->bmi.mv.as_int;
+                        }
+                        else
+                        {
+                            /* The full search result is actually worse so
+                             * re-instate the previous best vector
+                             */
+                            e->bmi.mv.as_int = mode_mv[NEW4X4].as_int;
+                        }
+                    }
+                }
+
+                if (bestsme < INT_MAX)
+                {
+                    int disto;
+                    unsigned int sse;
+                    cpi->find_fractional_mv_step(x, c, e, &mode_mv[NEW4X4],
+                        bsi->ref_mv, x->errorperbit, v_fn_ptr, x->mvcost,
+                        &disto, &sse);
+                }
+            } /* NEW4X4 */
+
+            rate = labels2mode(x, labels, i, this_mode, &mode_mv[this_mode],
+                               bsi->ref_mv, x->mvcost);
+
+            /* Trap vectors that reach beyond the UMV borders */
+            if (((mode_mv[this_mode].as_mv.row >> 3) < x->mv_row_min) || ((mode_mv[this_mode].as_mv.row >> 3) > x->mv_row_max) ||
+                ((mode_mv[this_mode].as_mv.col >> 3) < x->mv_col_min) || ((mode_mv[this_mode].as_mv.col >> 3) > x->mv_col_max))
+            {
+                continue;
+            }
+
+            distortion = vp8_encode_inter_mb_segment(x, labels, i) / 4;
+
+            labelyrate = rdcost_mbsegment_y(x, labels, i, ta_s, tl_s);
+            rate += labelyrate;
+
+            this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
+
+            if (this_rd < best_label_rd)
+            {
+                sbr = rate;
+                sbd = distortion;
+                bestlabelyrate = labelyrate;
+                mode_selected = this_mode;
+                best_label_rd = this_rd;
+
+                memcpy(ta_b, ta_s, sizeof(ENTROPY_CONTEXT_PLANES));
+                memcpy(tl_b, tl_s, sizeof(ENTROPY_CONTEXT_PLANES));
+
+            }
+        } /*for each 4x4 mode*/
+
+        memcpy(ta, ta_b, sizeof(ENTROPY_CONTEXT_PLANES));
+        memcpy(tl, tl_b, sizeof(ENTROPY_CONTEXT_PLANES));
+
+        labels2mode(x, labels, i, mode_selected, &mode_mv[mode_selected],
+                    bsi->ref_mv, x->mvcost);
+
+        br += sbr;
+        bd += sbd;
+        segmentyrate += bestlabelyrate;
+        this_segment_rd += best_label_rd;
+
+        if (this_segment_rd >= bsi->segment_rd)
+            break;
+
+    } /* for each label */
+
+    if (this_segment_rd < bsi->segment_rd)
+    {
+        bsi->r = br;
+        bsi->d = bd;
+        bsi->segment_yrate = segmentyrate;
+        bsi->segment_rd = this_segment_rd;
+        bsi->segment_num = segmentation;
+
+        /* store everything needed to come back to this!! */
+        for (i = 0; i < 16; i++)
+        {
+            bsi->mvs[i].as_mv = x->partition_info->bmi[i].mv.as_mv;
+            bsi->modes[i] = x->partition_info->bmi[i].mode;
+            bsi->eobs[i] = x->e_mbd.eobs[i];
+        }
+    }
+}
+
+static
+void vp8_cal_step_param(int sr, int *sp)
+{
+    int step = 0;
+
+    if (sr > MAX_FIRST_STEP) sr = MAX_FIRST_STEP;
+    else if (sr < 1) sr = 1;
+
+    while (sr>>=1)
+        step++;
+
+    *sp = MAX_MVSEARCH_STEPS - 1 - step;
+}
+
+static int vp8_rd_pick_best_mbsegmentation(VP8_COMP *cpi, MACROBLOCK *x,
+                                           int_mv *best_ref_mv, int best_rd,
+                                           int *mdcounts, int *returntotrate,
+                                           int *returnyrate, int *returndistortion,
+                                           int mvthresh)
+{
+    int i;
+    BEST_SEG_INFO bsi;
+
+    memset(&bsi, 0, sizeof(bsi));
+
+    bsi.segment_rd = best_rd;
+    bsi.ref_mv = best_ref_mv;
+    bsi.mvp.as_int = best_ref_mv->as_int;
+    bsi.mvthresh = mvthresh;
+    bsi.mdcounts = mdcounts;
+
+    for(i = 0; i < 16; i++)
+    {
+        bsi.modes[i] = ZERO4X4;
+    }
+
+    if(cpi->compressor_speed == 0)
+    {
+        /* for now, we will keep the original segmentation order
+           when in best quality mode */
+        rd_check_segment(cpi, x, &bsi, BLOCK_16X8);
+        rd_check_segment(cpi, x, &bsi, BLOCK_8X16);
+        rd_check_segment(cpi, x, &bsi, BLOCK_8X8);
+        rd_check_segment(cpi, x, &bsi, BLOCK_4X4);
+    }
+    else
+    {
+        int sr;
+
+        rd_check_segment(cpi, x, &bsi, BLOCK_8X8);
+
+        if (bsi.segment_rd < best_rd)
+        {
+            int col_min = ((best_ref_mv->as_mv.col+7)>>3) - MAX_FULL_PEL_VAL;
+            int row_min = ((best_ref_mv->as_mv.row+7)>>3) - MAX_FULL_PEL_VAL;
+            int col_max = (best_ref_mv->as_mv.col>>3) + MAX_FULL_PEL_VAL;
+            int row_max = (best_ref_mv->as_mv.row>>3) + MAX_FULL_PEL_VAL;
+
+            int tmp_col_min = x->mv_col_min;
+            int tmp_col_max = x->mv_col_max;
+            int tmp_row_min = x->mv_row_min;
+            int tmp_row_max = x->mv_row_max;
+
+            /* Get intersection of UMV window and valid MV window to reduce # of checks in diamond search. */
+            if (x->mv_col_min < col_min )
+                x->mv_col_min = col_min;
+            if (x->mv_col_max > col_max )
+                x->mv_col_max = col_max;
+            if (x->mv_row_min < row_min )
+                x->mv_row_min = row_min;
+            if (x->mv_row_max > row_max )
+                x->mv_row_max = row_max;
+
+            /* Get 8x8 result */
+            bsi.sv_mvp[0].as_int = bsi.mvs[0].as_int;
+            bsi.sv_mvp[1].as_int = bsi.mvs[2].as_int;
+            bsi.sv_mvp[2].as_int = bsi.mvs[8].as_int;
+            bsi.sv_mvp[3].as_int = bsi.mvs[10].as_int;
+
+            /* Use 8x8 result as 16x8/8x16's predictor MV. Adjust search range according to the closeness of 2 MV. */
+            /* block 8X16 */
+            {
+                sr = MAXF((abs(bsi.sv_mvp[0].as_mv.row - bsi.sv_mvp[2].as_mv.row))>>3, (abs(bsi.sv_mvp[0].as_mv.col - bsi.sv_mvp[2].as_mv.col))>>3);
+                vp8_cal_step_param(sr, &bsi.sv_istep[0]);
+
+                sr = MAXF((abs(bsi.sv_mvp[1].as_mv.row - bsi.sv_mvp[3].as_mv.row))>>3, (abs(bsi.sv_mvp[1].as_mv.col - bsi.sv_mvp[3].as_mv.col))>>3);
+                vp8_cal_step_param(sr, &bsi.sv_istep[1]);
+
+                rd_check_segment(cpi, x, &bsi, BLOCK_8X16);
+            }
+
+            /* block 16X8 */
+            {
+                sr = MAXF((abs(bsi.sv_mvp[0].as_mv.row - bsi.sv_mvp[1].as_mv.row))>>3, (abs(bsi.sv_mvp[0].as_mv.col - bsi.sv_mvp[1].as_mv.col))>>3);
+                vp8_cal_step_param(sr, &bsi.sv_istep[0]);
+
+                sr = MAXF((abs(bsi.sv_mvp[2].as_mv.row - bsi.sv_mvp[3].as_mv.row))>>3, (abs(bsi.sv_mvp[2].as_mv.col - bsi.sv_mvp[3].as_mv.col))>>3);
+                vp8_cal_step_param(sr, &bsi.sv_istep[1]);
+
+                rd_check_segment(cpi, x, &bsi, BLOCK_16X8);
+            }
+
+            /* If 8x8 is better than 16x8/8x16, then do 4x4 search */
+            /* Not skip 4x4 if speed=0 (good quality) */
+            if (cpi->sf.no_skip_block4x4_search || bsi.segment_num == BLOCK_8X8)  /* || (sv_segment_rd8x8-bsi.segment_rd) < sv_segment_rd8x8>>5) */
+            {
+                bsi.mvp.as_int = bsi.sv_mvp[0].as_int;
+                rd_check_segment(cpi, x, &bsi, BLOCK_4X4);
+            }
+
+            /* restore UMV window */
+            x->mv_col_min = tmp_col_min;
+            x->mv_col_max = tmp_col_max;
+            x->mv_row_min = tmp_row_min;
+            x->mv_row_max = tmp_row_max;
+        }
+    }
+
+    /* set it to the best */
+    for (i = 0; i < 16; i++)
+    {
+        BLOCKD *bd = &x->e_mbd.block[i];
+
+        bd->bmi.mv.as_int = bsi.mvs[i].as_int;
+        *bd->eob = bsi.eobs[i];
+    }
+
+    *returntotrate = bsi.r;
+    *returndistortion = bsi.d;
+    *returnyrate = bsi.segment_yrate;
+
+    /* save partitions */
+    x->e_mbd.mode_info_context->mbmi.partitioning = bsi.segment_num;
+    x->partition_info->count = vp8_mbsplit_count[bsi.segment_num];
+
+    for (i = 0; i < x->partition_info->count; i++)
+    {
+        int j;
+
+        j = vp8_mbsplit_offset[bsi.segment_num][i];
+
+        x->partition_info->bmi[i].mode = bsi.modes[j];
+        x->partition_info->bmi[i].mv.as_mv = bsi.mvs[j].as_mv;
+    }
+    /*
+     * used to set x->e_mbd.mode_info_context->mbmi.mv.as_int
+     */
+    x->partition_info->bmi[15].mv.as_int = bsi.mvs[15].as_int;
+
+    return bsi.segment_rd;
+}
+
+/* The improved MV prediction */
+void vp8_mv_pred
+(
+    VP8_COMP *cpi,
+    MACROBLOCKD *xd,
+    const MODE_INFO *here,
+    int_mv *mvp,
+    int refframe,
+    int *ref_frame_sign_bias,
+    int *sr,
+    int near_sadidx[]
+)
+{
+    const MODE_INFO *above = here - xd->mode_info_stride;
+    const MODE_INFO *left = here - 1;
+    const MODE_INFO *aboveleft = above - 1;
+    int_mv           near_mvs[8];
+    int              near_ref[8];
+    int_mv           mv;
+    int              vcnt=0;
+    int              find=0;
+    int              mb_offset;
+
+    int              mvx[8];
+    int              mvy[8];
+    int              i;
+
+    mv.as_int = 0;
+
+    if(here->mbmi.ref_frame != INTRA_FRAME)
+    {
+        near_mvs[0].as_int = near_mvs[1].as_int = near_mvs[2].as_int = near_mvs[3].as_int = near_mvs[4].as_int = near_mvs[5].as_int = near_mvs[6].as_int = near_mvs[7].as_int = 0;
+        near_ref[0] = near_ref[1] = near_ref[2] = near_ref[3] = near_ref[4] = near_ref[5] = near_ref[6] = near_ref[7] = 0;
+
+        /* read in 3 nearby block's MVs from current frame as prediction
+         * candidates.
+         */
+        if (above->mbmi.ref_frame != INTRA_FRAME)
+        {
+            near_mvs[vcnt].as_int = above->mbmi.mv.as_int;
+            mv_bias(ref_frame_sign_bias[above->mbmi.ref_frame], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+            near_ref[vcnt] =  above->mbmi.ref_frame;
+        }
+        vcnt++;
+        if (left->mbmi.ref_frame != INTRA_FRAME)
+        {
+            near_mvs[vcnt].as_int = left->mbmi.mv.as_int;
+            mv_bias(ref_frame_sign_bias[left->mbmi.ref_frame], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+            near_ref[vcnt] =  left->mbmi.ref_frame;
+        }
+        vcnt++;
+        if (aboveleft->mbmi.ref_frame != INTRA_FRAME)
+        {
+            near_mvs[vcnt].as_int = aboveleft->mbmi.mv.as_int;
+            mv_bias(ref_frame_sign_bias[aboveleft->mbmi.ref_frame], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+            near_ref[vcnt] =  aboveleft->mbmi.ref_frame;
+        }
+        vcnt++;
+
+        /* read in 5 nearby block's MVs from last frame. */
+        if(cpi->common.last_frame_type != KEY_FRAME)
+        {
+            mb_offset = (-xd->mb_to_top_edge/128 + 1) * (xd->mode_info_stride +1) + (-xd->mb_to_left_edge/128 +1) ;
+
+            /* current in last frame */
+            if (cpi->lf_ref_frame[mb_offset] != INTRA_FRAME)
+            {
+                near_mvs[vcnt].as_int = cpi->lfmv[mb_offset].as_int;
+                mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+                near_ref[vcnt] =  cpi->lf_ref_frame[mb_offset];
+            }
+            vcnt++;
+
+            /* above in last frame */
+            if (cpi->lf_ref_frame[mb_offset - xd->mode_info_stride-1] != INTRA_FRAME)
+            {
+                near_mvs[vcnt].as_int = cpi->lfmv[mb_offset - xd->mode_info_stride-1].as_int;
+                mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset - xd->mode_info_stride-1], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+                near_ref[vcnt] =  cpi->lf_ref_frame[mb_offset - xd->mode_info_stride-1];
+            }
+            vcnt++;
+
+            /* left in last frame */
+            if (cpi->lf_ref_frame[mb_offset-1] != INTRA_FRAME)
+            {
+                near_mvs[vcnt].as_int = cpi->lfmv[mb_offset -1].as_int;
+                mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset -1], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+                near_ref[vcnt] =  cpi->lf_ref_frame[mb_offset - 1];
+            }
+            vcnt++;
+
+            /* right in last frame */
+            if (cpi->lf_ref_frame[mb_offset +1] != INTRA_FRAME)
+            {
+                near_mvs[vcnt].as_int = cpi->lfmv[mb_offset +1].as_int;
+                mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset +1], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+                near_ref[vcnt] =  cpi->lf_ref_frame[mb_offset +1];
+            }
+            vcnt++;
+
+            /* below in last frame */
+            if (cpi->lf_ref_frame[mb_offset + xd->mode_info_stride +1] != INTRA_FRAME)
+            {
+                near_mvs[vcnt].as_int = cpi->lfmv[mb_offset + xd->mode_info_stride +1].as_int;
+                mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset + xd->mode_info_stride +1], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+                near_ref[vcnt] =  cpi->lf_ref_frame[mb_offset + xd->mode_info_stride +1];
+            }
+            vcnt++;
+        }
+
+        for(i=0; i< vcnt; i++)
+        {
+            if(near_ref[near_sadidx[i]] != INTRA_FRAME)
+            {
+                if(here->mbmi.ref_frame == near_ref[near_sadidx[i]])
+                {
+                    mv.as_int = near_mvs[near_sadidx[i]].as_int;
+                    find = 1;
+                    if (i < 3)
+                        *sr = 3;
+                    else
+                        *sr = 2;
+                    break;
+                }
+            }
+        }
+
+        if(!find)
+        {
+            for(i=0; i<vcnt; i++)
+            {
+                mvx[i] = near_mvs[i].as_mv.row;
+                mvy[i] = near_mvs[i].as_mv.col;
+            }
+
+            insertsortmv(mvx, vcnt);
+            insertsortmv(mvy, vcnt);
+            mv.as_mv.row = mvx[vcnt/2];
+            mv.as_mv.col = mvy[vcnt/2];
+
+            /* sr is set to 0 to allow calling function to decide the search
+             * range.
+             */
+            *sr = 0;
+        }
+    }
+
+    /* Set up return values */
+    mvp->as_int = mv.as_int;
+    vp8_clamp_mv2(mvp, xd);
+}
+
+void vp8_cal_sad(VP8_COMP *cpi, MACROBLOCKD *xd, MACROBLOCK *x, int recon_yoffset, int near_sadidx[])
+{
+    /* near_sad indexes:
+     *   0-cf above, 1-cf left, 2-cf aboveleft,
+     *   3-lf current, 4-lf above, 5-lf left, 6-lf right, 7-lf below
+     */
+    int near_sad[8] = {0};
+    BLOCK *b = &x->block[0];
+    unsigned char *src_y_ptr = *(b->base_src);
+
+    /* calculate sad for current frame 3 nearby MBs. */
+    if( xd->mb_to_top_edge==0 && xd->mb_to_left_edge ==0)
+    {
+        near_sad[0] = near_sad[1] = near_sad[2] = INT_MAX;
+    }else if(xd->mb_to_top_edge==0)
+    {   /* only has left MB for sad calculation. */
+        near_sad[0] = near_sad[2] = INT_MAX;
+        near_sad[1] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, xd->dst.y_buffer - 16,xd->dst.y_stride);
+    }else if(xd->mb_to_left_edge ==0)
+    {   /* only has left MB for sad calculation. */
+        near_sad[1] = near_sad[2] = INT_MAX;
+        near_sad[0] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, xd->dst.y_buffer - xd->dst.y_stride *16,xd->dst.y_stride);
+    }else
+    {
+        near_sad[0] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, xd->dst.y_buffer - xd->dst.y_stride *16,xd->dst.y_stride);
+        near_sad[1] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, xd->dst.y_buffer - 16,xd->dst.y_stride);
+        near_sad[2] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, xd->dst.y_buffer - xd->dst.y_stride *16 -16,xd->dst.y_stride);
+    }
+
+    if(cpi->common.last_frame_type != KEY_FRAME)
+    {
+        /* calculate sad for last frame 5 nearby MBs. */
+        unsigned char *pre_y_buffer = cpi->common.yv12_fb[cpi->common.lst_fb_idx].y_buffer + recon_yoffset;
+        int pre_y_stride = cpi->common.yv12_fb[cpi->common.lst_fb_idx].y_stride;
+
+        if(xd->mb_to_top_edge==0) near_sad[4] = INT_MAX;
+        if(xd->mb_to_left_edge ==0) near_sad[5] = INT_MAX;
+        if(xd->mb_to_right_edge ==0) near_sad[6] = INT_MAX;
+        if(xd->mb_to_bottom_edge==0) near_sad[7] = INT_MAX;
+
+        if(near_sad[4] != INT_MAX)
+            near_sad[4] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, pre_y_buffer - pre_y_stride *16, pre_y_stride);
+        if(near_sad[5] != INT_MAX)
+            near_sad[5] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, pre_y_buffer - 16, pre_y_stride);
+        near_sad[3] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, pre_y_buffer, pre_y_stride);
+        if(near_sad[6] != INT_MAX)
+            near_sad[6] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, pre_y_buffer + 16, pre_y_stride);
+        if(near_sad[7] != INT_MAX)
+            near_sad[7] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, pre_y_buffer + pre_y_stride *16, pre_y_stride);
+    }
+
+    if(cpi->common.last_frame_type != KEY_FRAME)
+    {
+        insertsortsad(near_sad, near_sadidx, 8);
+    }else
+    {
+        insertsortsad(near_sad, near_sadidx, 3);
+    }
+}
+
+static void rd_update_mvcount(MACROBLOCK *x, int_mv *best_ref_mv)
+{
+    if (x->e_mbd.mode_info_context->mbmi.mode == SPLITMV)
+    {
+        int i;
+
+        for (i = 0; i < x->partition_info->count; i++)
+        {
+            if (x->partition_info->bmi[i].mode == NEW4X4)
+            {
+                x->MVcount[0][mv_max+((x->partition_info->bmi[i].mv.as_mv.row
+                                          - best_ref_mv->as_mv.row) >> 1)]++;
+                x->MVcount[1][mv_max+((x->partition_info->bmi[i].mv.as_mv.col
+                                          - best_ref_mv->as_mv.col) >> 1)]++;
+            }
+        }
+    }
+    else if (x->e_mbd.mode_info_context->mbmi.mode == NEWMV)
+    {
+        x->MVcount[0][mv_max+((x->e_mbd.mode_info_context->mbmi.mv.as_mv.row
+                                          - best_ref_mv->as_mv.row) >> 1)]++;
+        x->MVcount[1][mv_max+((x->e_mbd.mode_info_context->mbmi.mv.as_mv.col
+                                          - best_ref_mv->as_mv.col) >> 1)]++;
+    }
+}
+
+static int evaluate_inter_mode_rd(int mdcounts[4],
+                                  RATE_DISTORTION* rd,
+                                  int* disable_skip,
+                                  VP8_COMP *cpi, MACROBLOCK *x)
+{
+    MB_PREDICTION_MODE this_mode = x->e_mbd.mode_info_context->mbmi.mode;
+    BLOCK *b = &x->block[0];
+    MACROBLOCKD *xd = &x->e_mbd;
+    int distortion;
+    vp8_build_inter16x16_predictors_mby(&x->e_mbd, x->e_mbd.predictor, 16);
+
+    if (cpi->active_map_enabled && x->active_ptr[0] == 0) {
+        x->skip = 1;
+    }
+    else if (x->encode_breakout)
+    {
+        unsigned int sse;
+        unsigned int var;
+        unsigned int threshold = (xd->block[0].dequant[1]
+                    * xd->block[0].dequant[1] >>4);
+
+        if(threshold < x->encode_breakout)
+            threshold = x->encode_breakout;
+
+        var = vpx_variance16x16
+                (*(b->base_src), b->src_stride,
+                x->e_mbd.predictor, 16, &sse);
+
+        if (sse < threshold)
+        {
+             unsigned int q2dc = xd->block[24].dequant[0];
+            /* If theres is no codeable 2nd order dc
+               or a very small uniform pixel change change */
+            if ((sse - var < q2dc * q2dc >>4) ||
+                (sse /2 > var && sse-var < 64))
+            {
+                /* Check u and v to make sure skip is ok */
+                unsigned int sse2 = VP8_UVSSE(x);
+                if (sse2 * 2 < threshold)
+                {
+                    x->skip = 1;
+                    rd->distortion2 = sse + sse2;
+                    rd->rate2 = 500;
+
+                    /* for best_yrd calculation */
+                    rd->rate_uv = 0;
+                    rd->distortion_uv = sse2;
+
+                    *disable_skip = 1;
+                    return RDCOST(x->rdmult, x->rddiv, rd->rate2,
+                                  rd->distortion2);
+                }
+            }
+        }
+    }
+
+
+    /* Add in the Mv/mode cost */
+    rd->rate2 += vp8_cost_mv_ref(this_mode, mdcounts);
+
+    /* Y cost and distortion */
+    macro_block_yrd(x, &rd->rate_y, &distortion);
+    rd->rate2 += rd->rate_y;
+    rd->distortion2 += distortion;
+
+    /* UV cost and distortion */
+    rd_inter16x16_uv(cpi, x, &rd->rate_uv, &rd->distortion_uv,
+                     cpi->common.full_pixel);
+    rd->rate2 += rd->rate_uv;
+    rd->distortion2 += rd->distortion_uv;
+    return INT_MAX;
+}
+
+static int calculate_final_rd_costs(int this_rd,
+                                    RATE_DISTORTION* rd,
+                                    int* other_cost,
+                                    int disable_skip,
+                                    int uv_intra_tteob,
+                                    int intra_rd_penalty,
+                                    VP8_COMP *cpi, MACROBLOCK *x)
+{
+    MB_PREDICTION_MODE this_mode = x->e_mbd.mode_info_context->mbmi.mode;
+
+    /* Where skip is allowable add in the default per mb cost for the no
+     * skip case. where we then decide to skip we have to delete this and
+     * replace it with the cost of signalling a skip
+     */
+    if (cpi->common.mb_no_coeff_skip)
+    {
+        *other_cost += vp8_cost_bit(cpi->prob_skip_false, 0);
+        rd->rate2 += *other_cost;
+    }
+
+    /* Estimate the reference frame signaling cost and add it
+     * to the rolling cost variable.
+     */
+    rd->rate2 +=
+        x->ref_frame_cost[x->e_mbd.mode_info_context->mbmi.ref_frame];
+
+    if (!disable_skip)
+    {
+        /* Test for the condition where skip block will be activated
+         * because there are no non zero coefficients and make any
+         * necessary adjustment for rate
+         */
+        if (cpi->common.mb_no_coeff_skip)
+        {
+            int i;
+            int tteob;
+            int has_y2_block = (this_mode!=SPLITMV && this_mode!=B_PRED);
+
+            tteob = 0;
+            if(has_y2_block)
+                tteob += x->e_mbd.eobs[24];
+
+            for (i = 0; i < 16; i++)
+                tteob += (x->e_mbd.eobs[i] > has_y2_block);
+
+            if (x->e_mbd.mode_info_context->mbmi.ref_frame)
+            {
+                for (i = 16; i < 24; i++)
+                    tteob += x->e_mbd.eobs[i];
+            }
+            else
+                tteob += uv_intra_tteob;
+
+            if (tteob == 0)
+            {
+                rd->rate2 -= (rd->rate_y + rd->rate_uv);
+                /* for best_yrd calculation */
+                rd->rate_uv = 0;
+
+                /* Back out no skip flag costing and add in skip flag costing */
+                if (cpi->prob_skip_false)
+                {
+                    int prob_skip_cost;
+
+                    prob_skip_cost = vp8_cost_bit(cpi->prob_skip_false, 1);
+                    prob_skip_cost -=
+                        (int)vp8_cost_bit(cpi->prob_skip_false, 0);
+                    rd->rate2 += prob_skip_cost;
+                    *other_cost += prob_skip_cost;
+                }
+            }
+        }
+        /* Calculate the final RD estimate for this mode */
+        this_rd = RDCOST(x->rdmult, x->rddiv, rd->rate2, rd->distortion2);
+        if (this_rd < INT_MAX && x->e_mbd.mode_info_context->mbmi.ref_frame
+                                 == INTRA_FRAME)
+            this_rd += intra_rd_penalty;
+    }
+    return this_rd;
+}
+
+static void update_best_mode(BEST_MODE* best_mode, int this_rd,
+                             RATE_DISTORTION* rd, int other_cost, MACROBLOCK *x)
+{
+    MB_PREDICTION_MODE this_mode = x->e_mbd.mode_info_context->mbmi.mode;
+
+    other_cost +=
+    x->ref_frame_cost[x->e_mbd.mode_info_context->mbmi.ref_frame];
+
+    /* Calculate the final y RD estimate for this mode */
+    best_mode->yrd = RDCOST(x->rdmult, x->rddiv, (rd->rate2-rd->rate_uv-other_cost),
+                      (rd->distortion2-rd->distortion_uv));
+
+    best_mode->rd = this_rd;
+    memcpy(&best_mode->mbmode, &x->e_mbd.mode_info_context->mbmi, sizeof(MB_MODE_INFO));
+    memcpy(&best_mode->partition, x->partition_info, sizeof(PARTITION_INFO));
+
+    if ((this_mode == B_PRED) || (this_mode == SPLITMV))
+    {
+        int i;
+        for (i = 0; i < 16; i++)
+        {
+            best_mode->bmodes[i] = x->e_mbd.block[i].bmi;
+        }
+    }
+}
+
+void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset,
+                            int recon_uvoffset, int *returnrate,
+                            int *returndistortion, int *returnintra,
+                            int mb_row, int mb_col)
+{
+    BLOCK *b = &x->block[0];
+    BLOCKD *d = &x->e_mbd.block[0];
+    MACROBLOCKD *xd = &x->e_mbd;
+    int_mv best_ref_mv_sb[2];
+    int_mv mode_mv_sb[2][MB_MODE_COUNT];
+    int_mv best_ref_mv;
+    int_mv *mode_mv;
+    MB_PREDICTION_MODE this_mode;
+    int num00;
+    int best_mode_index = 0;
+    BEST_MODE best_mode;
+
+    int i;
+    int mode_index;
+    int mdcounts[4];
+    int rate;
+    RATE_DISTORTION rd;
+    int uv_intra_rate, uv_intra_distortion, uv_intra_rate_tokenonly;
+    int uv_intra_tteob = 0;
+    int uv_intra_done = 0;
+
+    MB_PREDICTION_MODE uv_intra_mode = 0;
+    int_mv mvp;
+    int near_sadidx[8] = {0, 1, 2, 3, 4, 5, 6, 7};
+    int saddone=0;
+    /* search range got from mv_pred(). It uses step_param levels. (0-7) */
+    int sr=0;
+
+    unsigned char *plane[4][3];
+    int ref_frame_map[4];
+    int sign_bias = 0;
+
+    int intra_rd_penalty =  10* vp8_dc_quant(cpi->common.base_qindex,
+                                             cpi->common.y1dc_delta_q);
+
+#if CONFIG_TEMPORAL_DENOISING
+    unsigned int zero_mv_sse = UINT_MAX, best_sse = UINT_MAX,
+            best_rd_sse = UINT_MAX;
+#endif
+
+    mode_mv = mode_mv_sb[sign_bias];
+    best_ref_mv.as_int = 0;
+    best_mode.rd = INT_MAX;
+    best_mode.yrd = INT_MAX;
+    best_mode.intra_rd = INT_MAX;
+    memset(mode_mv_sb, 0, sizeof(mode_mv_sb));
+    memset(&best_mode.mbmode, 0, sizeof(best_mode.mbmode));
+    memset(&best_mode.bmodes, 0, sizeof(best_mode.bmodes));
+
+    /* Setup search priorities */
+    get_reference_search_order(cpi, ref_frame_map);
+
+    /* Check to see if there is at least 1 valid reference frame that we need
+     * to calculate near_mvs.
+     */
+    if (ref_frame_map[1] > 0)
+    {
+        sign_bias = vp8_find_near_mvs_bias(&x->e_mbd,
+                                           x->e_mbd.mode_info_context,
+                                           mode_mv_sb,
+                                           best_ref_mv_sb,
+                                           mdcounts,
+                                           ref_frame_map[1],
+                                           cpi->common.ref_frame_sign_bias);
+
+        mode_mv = mode_mv_sb[sign_bias];
+        best_ref_mv.as_int = best_ref_mv_sb[sign_bias].as_int;
+    }
+
+    get_predictor_pointers(cpi, plane, recon_yoffset, recon_uvoffset);
+
+    *returnintra = INT_MAX;
+    /* Count of the number of MBs tested so far this frame */
+    x->mbs_tested_so_far++;
+
+    x->skip = 0;
+
+    for (mode_index = 0; mode_index < MAX_MODES; mode_index++)
+    {
+        int this_rd = INT_MAX;
+        int disable_skip = 0;
+        int other_cost = 0;
+        int this_ref_frame = ref_frame_map[vp8_ref_frame_order[mode_index]];
+
+        /* Test best rd so far against threshold for trying this mode. */
+        if (best_mode.rd <= x->rd_threshes[mode_index])
+            continue;
+
+        if (this_ref_frame < 0)
+            continue;
+
+        /* These variables hold are rolling total cost and distortion for
+         * this mode
+         */
+        rd.rate2 = 0;
+        rd.distortion2 = 0;
+
+        this_mode = vp8_mode_order[mode_index];
+
+        x->e_mbd.mode_info_context->mbmi.mode = this_mode;
+        x->e_mbd.mode_info_context->mbmi.ref_frame = this_ref_frame;
+
+        /* Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
+         * unless ARNR filtering is enabled in which case we want
+         * an unfiltered alternative
+         */
+        if (cpi->is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
+        {
+            if (this_mode != ZEROMV || x->e_mbd.mode_info_context->mbmi.ref_frame != ALTREF_FRAME)
+                continue;
+        }
+
+        /* everything but intra */
+        if (x->e_mbd.mode_info_context->mbmi.ref_frame)
+        {
+            x->e_mbd.pre.y_buffer = plane[this_ref_frame][0];
+            x->e_mbd.pre.u_buffer = plane[this_ref_frame][1];
+            x->e_mbd.pre.v_buffer = plane[this_ref_frame][2];
+
+            if (sign_bias != cpi->common.ref_frame_sign_bias[this_ref_frame])
+            {
+                sign_bias = cpi->common.ref_frame_sign_bias[this_ref_frame];
+                mode_mv = mode_mv_sb[sign_bias];
+                best_ref_mv.as_int = best_ref_mv_sb[sign_bias].as_int;
+            }
+        }
+
+        /* Check to see if the testing frequency for this mode is at its
+         * max If so then prevent it from being tested and increase the
+         * threshold for its testing
+         */
+        if (x->mode_test_hit_counts[mode_index] && (cpi->mode_check_freq[mode_index] > 1))
+        {
+            if (x->mbs_tested_so_far  <= cpi->mode_check_freq[mode_index] * x->mode_test_hit_counts[mode_index])
+            {
+                /* Increase the threshold for coding this mode to make it
+                 * less likely to be chosen
+                 */
+                x->rd_thresh_mult[mode_index] += 4;
+
+                if (x->rd_thresh_mult[mode_index] > MAX_THRESHMULT)
+                    x->rd_thresh_mult[mode_index] = MAX_THRESHMULT;
+
+                x->rd_threshes[mode_index] =
+                    (cpi->rd_baseline_thresh[mode_index] >> 7) *
+                    x->rd_thresh_mult[mode_index];
+
+                continue;
+            }
+        }
+
+        /* We have now reached the point where we are going to test the
+         * current mode so increment the counter for the number of times
+         * it has been tested
+         */
+        x->mode_test_hit_counts[mode_index] ++;
+
+        /* Experimental code. Special case for gf and arf zeromv modes.
+         * Increase zbin size to supress noise
+         */
+        if (x->zbin_mode_boost_enabled)
+        {
+            if ( this_ref_frame == INTRA_FRAME )
+                x->zbin_mode_boost = 0;
+            else
+            {
+                if (vp8_mode_order[mode_index] == ZEROMV)
+                {
+                    if (this_ref_frame != LAST_FRAME)
+                        x->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
+                    else
+                        x->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST;
+                }
+                else if (vp8_mode_order[mode_index] == SPLITMV)
+                    x->zbin_mode_boost = 0;
+                else
+                    x->zbin_mode_boost = MV_ZBIN_BOOST;
+            }
+
+            vp8_update_zbin_extra(cpi, x);
+        }
+
+        if(!uv_intra_done && this_ref_frame == INTRA_FRAME)
+        {
+            rd_pick_intra_mbuv_mode(x, &uv_intra_rate,
+                                    &uv_intra_rate_tokenonly,
+                                    &uv_intra_distortion);
+            uv_intra_mode = x->e_mbd.mode_info_context->mbmi.uv_mode;
+
+            /*
+             * Total of the eobs is used later to further adjust rate2. Since uv
+             * block's intra eobs will be overwritten when we check inter modes,
+             * we need to save uv_intra_tteob here.
+             */
+            for (i = 16; i < 24; i++)
+                uv_intra_tteob += x->e_mbd.eobs[i];
+
+            uv_intra_done = 1;
+        }
+
+        switch (this_mode)
+        {
+        case B_PRED:
+        {
+            int tmp_rd;
+
+            /* Note the rate value returned here includes the cost of
+             * coding the BPRED mode: x->mbmode_cost[x->e_mbd.frame_type][BPRED]
+             */
+            int distortion;
+            tmp_rd = rd_pick_intra4x4mby_modes(x, &rate, &rd.rate_y, &distortion, best_mode.yrd);
+            rd.rate2 += rate;
+            rd.distortion2 += distortion;
+
+            if(tmp_rd < best_mode.yrd)
+            {
+                rd.rate2 += uv_intra_rate;
+                rd.rate_uv = uv_intra_rate_tokenonly;
+                rd.distortion2 += uv_intra_distortion;
+                rd.distortion_uv = uv_intra_distortion;
+            }
+            else
+            {
+                this_rd = INT_MAX;
+                disable_skip = 1;
+            }
+        }
+        break;
+
+        case SPLITMV:
+        {
+            int tmp_rd;
+            int this_rd_thresh;
+            int distortion;
+
+            this_rd_thresh = (vp8_ref_frame_order[mode_index] == 1) ?
+                x->rd_threshes[THR_NEW1] : x->rd_threshes[THR_NEW3];
+            this_rd_thresh = (vp8_ref_frame_order[mode_index] == 2) ?
+                x->rd_threshes[THR_NEW2] : this_rd_thresh;
+
+            tmp_rd = vp8_rd_pick_best_mbsegmentation(cpi, x, &best_ref_mv,
+                                                     best_mode.yrd, mdcounts,
+                                                     &rate, &rd.rate_y, &distortion, this_rd_thresh) ;
+
+            rd.rate2 += rate;
+            rd.distortion2 += distortion;
+
+            /* If even the 'Y' rd value of split is higher than best so far
+             * then dont bother looking at UV
+             */
+            if (tmp_rd < best_mode.yrd)
+            {
+                /* Now work out UV cost and add it in */
+                rd_inter4x4_uv(cpi, x, &rd.rate_uv, &rd.distortion_uv, cpi->common.full_pixel);
+                rd.rate2 += rd.rate_uv;
+                rd.distortion2 += rd.distortion_uv;
+            }
+            else
+            {
+                this_rd = INT_MAX;
+                disable_skip = 1;
+            }
+        }
+        break;
+        case DC_PRED:
+        case V_PRED:
+        case H_PRED:
+        case TM_PRED:
+        {
+            int distortion;
+            x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME;
+
+            vp8_build_intra_predictors_mby_s(xd,
+                                             xd->dst.y_buffer - xd->dst.y_stride,
+                                             xd->dst.y_buffer - 1,
+                                             xd->dst.y_stride,
+                                             xd->predictor,
+                                             16);
+            macro_block_yrd(x, &rd.rate_y, &distortion) ;
+            rd.rate2 += rd.rate_y;
+            rd.distortion2 += distortion;
+            rd.rate2 += x->mbmode_cost[x->e_mbd.frame_type][x->e_mbd.mode_info_context->mbmi.mode];
+            rd.rate2 += uv_intra_rate;
+            rd.rate_uv = uv_intra_rate_tokenonly;
+            rd.distortion2 += uv_intra_distortion;
+            rd.distortion_uv = uv_intra_distortion;
+        }
+        break;
+
+        case NEWMV:
+        {
+            int thissme;
+            int bestsme = INT_MAX;
+            int step_param = cpi->sf.first_step;
+            int further_steps;
+            int n;
+            int do_refine=1;   /* If last step (1-away) of n-step search doesn't pick the center point as the best match,
+                                  we will do a final 1-away diamond refining search  */
+
+            int sadpb = x->sadperbit16;
+            int_mv mvp_full;
+
+            int col_min = ((best_ref_mv.as_mv.col+7)>>3) - MAX_FULL_PEL_VAL;
+            int row_min = ((best_ref_mv.as_mv.row+7)>>3) - MAX_FULL_PEL_VAL;
+            int col_max = (best_ref_mv.as_mv.col>>3) + MAX_FULL_PEL_VAL;
+            int row_max = (best_ref_mv.as_mv.row>>3) + MAX_FULL_PEL_VAL;
+
+            int tmp_col_min = x->mv_col_min;
+            int tmp_col_max = x->mv_col_max;
+            int tmp_row_min = x->mv_row_min;
+            int tmp_row_max = x->mv_row_max;
+
+            if(!saddone)
+            {
+                vp8_cal_sad(cpi,xd,x, recon_yoffset ,&near_sadidx[0] );
+                saddone = 1;
+            }
+
+            vp8_mv_pred(cpi, &x->e_mbd, x->e_mbd.mode_info_context, &mvp,
+                        x->e_mbd.mode_info_context->mbmi.ref_frame, cpi->common.ref_frame_sign_bias, &sr, &near_sadidx[0]);
+
+            mvp_full.as_mv.col = mvp.as_mv.col>>3;
+            mvp_full.as_mv.row = mvp.as_mv.row>>3;
+
+            /* Get intersection of UMV window and valid MV window to
+             * reduce # of checks in diamond search.
+             */
+            if (x->mv_col_min < col_min )
+                x->mv_col_min = col_min;
+            if (x->mv_col_max > col_max )
+                x->mv_col_max = col_max;
+            if (x->mv_row_min < row_min )
+                x->mv_row_min = row_min;
+            if (x->mv_row_max > row_max )
+                x->mv_row_max = row_max;
+
+            /* adjust search range according to sr from mv prediction */
+            if(sr > step_param)
+                step_param = sr;
+
+            /* Initial step/diamond search */
+            {
+                bestsme = cpi->diamond_search_sad(x, b, d, &mvp_full, &d->bmi.mv,
+                                        step_param, sadpb, &num00,
+                                        &cpi->fn_ptr[BLOCK_16X16],
+                                        x->mvcost, &best_ref_mv);
+                mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
+
+                /* Further step/diamond searches as necessary */
+                further_steps = (cpi->sf.max_step_search_steps - 1) - step_param;
+
+                n = num00;
+                num00 = 0;
+
+                /* If there won't be more n-step search, check to see if refining search is needed. */
+                if (n > further_steps)
+                    do_refine = 0;
+
+                while (n < further_steps)
+                {
+                    n++;
+
+                    if (num00)
+                        num00--;
+                    else
+                    {
+                        thissme = cpi->diamond_search_sad(x, b, d, &mvp_full,
+                                    &d->bmi.mv, step_param + n, sadpb, &num00,
+                                    &cpi->fn_ptr[BLOCK_16X16], x->mvcost,
+                                    &best_ref_mv);
+
+                        /* check to see if refining search is needed. */
+                        if (num00 > (further_steps-n))
+                            do_refine = 0;
+
+                        if (thissme < bestsme)
+                        {
+                            bestsme = thissme;
+                            mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
+                        }
+                        else
+                        {
+                            d->bmi.mv.as_int = mode_mv[NEWMV].as_int;
+                        }
+                    }
+                }
+            }
+
+            /* final 1-away diamond refining search */
+            if (do_refine == 1)
+            {
+                int search_range;
+
+                search_range = 8;
+
+                thissme = cpi->refining_search_sad(x, b, d, &d->bmi.mv, sadpb,
+                                       search_range, &cpi->fn_ptr[BLOCK_16X16],
+                                       x->mvcost, &best_ref_mv);
+
+                if (thissme < bestsme)
+                {
+                    bestsme = thissme;
+                    mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
+                }
+                else
+                {
+                    d->bmi.mv.as_int = mode_mv[NEWMV].as_int;
+                }
+            }
+
+            x->mv_col_min = tmp_col_min;
+            x->mv_col_max = tmp_col_max;
+            x->mv_row_min = tmp_row_min;
+            x->mv_row_max = tmp_row_max;
+
+            if (bestsme < INT_MAX)
+            {
+                int dis; /* TODO: use dis in distortion calculation later. */
+                unsigned int sse;
+                cpi->find_fractional_mv_step(x, b, d, &d->bmi.mv, &best_ref_mv,
+                                             x->errorperbit,
+                                             &cpi->fn_ptr[BLOCK_16X16],
+                                             x->mvcost, &dis, &sse);
+            }
+
+            mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
+
+            /* Add the new motion vector cost to our rolling cost variable */
+            rd.rate2 += vp8_mv_bit_cost(&mode_mv[NEWMV], &best_ref_mv, x->mvcost, 96);
+        }
+
+        case NEARESTMV:
+        case NEARMV:
+            /* Clip "next_nearest" so that it does not extend to far out
+             * of image
+             */
+            vp8_clamp_mv2(&mode_mv[this_mode], xd);
+
+            /* Do not bother proceeding if the vector (from newmv, nearest
+             * or near) is 0,0 as this should then be coded using the zeromv
+             * mode.
+             */
+            if (((this_mode == NEARMV) || (this_mode == NEARESTMV)) && (mode_mv[this_mode].as_int == 0))
+                continue;
+
+        case ZEROMV:
+
+            /* Trap vectors that reach beyond the UMV borders
+             * Note that ALL New MV, Nearest MV Near MV and Zero MV code
+             * drops through to this point because of the lack of break
+             * statements in the previous two cases.
+             */
+            if (((mode_mv[this_mode].as_mv.row >> 3) < x->mv_row_min) || ((mode_mv[this_mode].as_mv.row >> 3) > x->mv_row_max) ||
+                ((mode_mv[this_mode].as_mv.col >> 3) < x->mv_col_min) || ((mode_mv[this_mode].as_mv.col >> 3) > x->mv_col_max))
+                continue;
+
+            vp8_set_mbmode_and_mvs(x, this_mode, &mode_mv[this_mode]);
+            this_rd = evaluate_inter_mode_rd(mdcounts, &rd,
+                                             &disable_skip, cpi, x);
+            break;
+
+        default:
+            break;
+        }
+
+        this_rd = calculate_final_rd_costs(this_rd, &rd, &other_cost,
+                                           disable_skip, uv_intra_tteob,
+                                           intra_rd_penalty, cpi, x);
+
+        /* Keep record of best intra distortion */
+        if ((x->e_mbd.mode_info_context->mbmi.ref_frame == INTRA_FRAME) &&
+            (this_rd < best_mode.intra_rd) )
+        {
+          best_mode.intra_rd = this_rd;
+            *returnintra = rd.distortion2 ;
+        }
+#if CONFIG_TEMPORAL_DENOISING
+        if (cpi->oxcf.noise_sensitivity)
+        {
+            unsigned int sse;
+            vp8_get_inter_mbpred_error(x,&cpi->fn_ptr[BLOCK_16X16],&sse,
+                                   mode_mv[this_mode]);
+
+            if (sse < best_rd_sse)
+                best_rd_sse = sse;
+
+            /* Store for later use by denoiser. */
+            if (this_mode == ZEROMV && sse < zero_mv_sse )
+            {
+                zero_mv_sse = sse;
+                x->best_zeromv_reference_frame =
+                        x->e_mbd.mode_info_context->mbmi.ref_frame;
+            }
+
+            /* Store the best NEWMV in x for later use in the denoiser. */
+            if (x->e_mbd.mode_info_context->mbmi.mode == NEWMV &&
+                    sse < best_sse)
+            {
+                best_sse = sse;
+                vp8_get_inter_mbpred_error(x,&cpi->fn_ptr[BLOCK_16X16],&best_sse,
+                                       mode_mv[this_mode]);
+                x->best_sse_inter_mode = NEWMV;
+                x->best_sse_mv = x->e_mbd.mode_info_context->mbmi.mv;
+                x->need_to_clamp_best_mvs =
+                    x->e_mbd.mode_info_context->mbmi.need_to_clamp_mvs;
+                x->best_reference_frame =
+                    x->e_mbd.mode_info_context->mbmi.ref_frame;
+            }
+        }
+#endif
+
+        /* Did this mode help.. i.i is it the new best mode */
+        if (this_rd < best_mode.rd || x->skip)
+        {
+            /* Note index of best mode so far */
+            best_mode_index = mode_index;
+            *returnrate = rd.rate2;
+            *returndistortion = rd.distortion2;
+            if (this_mode <= B_PRED)
+            {
+                x->e_mbd.mode_info_context->mbmi.uv_mode = uv_intra_mode;
+                /* required for left and above block mv */
+                x->e_mbd.mode_info_context->mbmi.mv.as_int = 0;
+            }
+            update_best_mode(&best_mode, this_rd, &rd, other_cost, x);
+
+
+            /* Testing this mode gave rise to an improvement in best error
+             * score. Lower threshold a bit for next time
+             */
+            x->rd_thresh_mult[mode_index] =
+                (x->rd_thresh_mult[mode_index] >= (MIN_THRESHMULT + 2)) ?
+                    x->rd_thresh_mult[mode_index] - 2 : MIN_THRESHMULT;
+        }
+
+        /* If the mode did not help improve the best error case then raise
+         * the threshold for testing that mode next time around.
+         */
+        else
+        {
+            x->rd_thresh_mult[mode_index] += 4;
+
+            if (x->rd_thresh_mult[mode_index] > MAX_THRESHMULT)
+                x->rd_thresh_mult[mode_index] = MAX_THRESHMULT;
+        }
+        x->rd_threshes[mode_index] =
+            (cpi->rd_baseline_thresh[mode_index] >> 7) *
+                x->rd_thresh_mult[mode_index];
+
+        if (x->skip)
+            break;
+
+    }
+
+    /* Reduce the activation RD thresholds for the best choice mode */
+    if ((cpi->rd_baseline_thresh[best_mode_index] > 0) && (cpi->rd_baseline_thresh[best_mode_index] < (INT_MAX >> 2)))
+    {
+        int best_adjustment = (x->rd_thresh_mult[best_mode_index] >> 2);
+
+        x->rd_thresh_mult[best_mode_index] =
+            (x->rd_thresh_mult[best_mode_index] >=
+                (MIN_THRESHMULT + best_adjustment)) ?
+                    x->rd_thresh_mult[best_mode_index] - best_adjustment :
+                    MIN_THRESHMULT;
+        x->rd_threshes[best_mode_index] =
+            (cpi->rd_baseline_thresh[best_mode_index] >> 7) *
+                x->rd_thresh_mult[best_mode_index];
+    }
+
+#if CONFIG_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity)
+    {
+        int block_index = mb_row * cpi->common.mb_cols + mb_col;
+        if (x->best_sse_inter_mode == DC_PRED)
+        {
+            /* No best MV found. */
+            x->best_sse_inter_mode = best_mode.mbmode.mode;
+            x->best_sse_mv = best_mode.mbmode.mv;
+            x->need_to_clamp_best_mvs = best_mode.mbmode.need_to_clamp_mvs;
+            x->best_reference_frame = best_mode.mbmode.ref_frame;
+            best_sse = best_rd_sse;
+        }
+        vp8_denoiser_denoise_mb(&cpi->denoiser, x, best_sse, zero_mv_sse,
+                                recon_yoffset, recon_uvoffset,
+                                &cpi->common.lf_info, mb_row, mb_col,
+                                block_index, 0);
+
+        /* Reevaluate ZEROMV after denoising. */
+        if (best_mode.mbmode.ref_frame == INTRA_FRAME &&
+            x->best_zeromv_reference_frame != INTRA_FRAME)
+        {
+            int this_rd = INT_MAX;
+            int disable_skip = 0;
+            int other_cost = 0;
+            int this_ref_frame = x->best_zeromv_reference_frame;
+            rd.rate2 = x->ref_frame_cost[this_ref_frame] +
+                    vp8_cost_mv_ref(ZEROMV, mdcounts);
+            rd.distortion2 = 0;
+
+            /* set up the proper prediction buffers for the frame */
+            x->e_mbd.mode_info_context->mbmi.ref_frame = this_ref_frame;
+            x->e_mbd.pre.y_buffer = plane[this_ref_frame][0];
+            x->e_mbd.pre.u_buffer = plane[this_ref_frame][1];
+            x->e_mbd.pre.v_buffer = plane[this_ref_frame][2];
+
+            x->e_mbd.mode_info_context->mbmi.mode = ZEROMV;
+            x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
+            x->e_mbd.mode_info_context->mbmi.mv.as_int = 0;
+
+            this_rd = evaluate_inter_mode_rd(mdcounts, &rd, &disable_skip, cpi, x);
+            this_rd = calculate_final_rd_costs(this_rd, &rd, &other_cost,
+                                               disable_skip, uv_intra_tteob,
+                                               intra_rd_penalty, cpi, x);
+            if (this_rd < best_mode.rd || x->skip)
+            {
+                *returnrate = rd.rate2;
+                *returndistortion = rd.distortion2;
+                update_best_mode(&best_mode, this_rd, &rd, other_cost, x);
+            }
+        }
+
+    }
+#endif
+
+    if (cpi->is_src_frame_alt_ref &&
+        (best_mode.mbmode.mode != ZEROMV || best_mode.mbmode.ref_frame != ALTREF_FRAME))
+    {
+        x->e_mbd.mode_info_context->mbmi.mode = ZEROMV;
+        x->e_mbd.mode_info_context->mbmi.ref_frame = ALTREF_FRAME;
+        x->e_mbd.mode_info_context->mbmi.mv.as_int = 0;
+        x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
+        x->e_mbd.mode_info_context->mbmi.mb_skip_coeff =
+                                        (cpi->common.mb_no_coeff_skip);
+        x->e_mbd.mode_info_context->mbmi.partitioning = 0;
+        return;
+    }
+
+
+    /* macroblock modes */
+    memcpy(&x->e_mbd.mode_info_context->mbmi, &best_mode.mbmode, sizeof(MB_MODE_INFO));
+
+    if (best_mode.mbmode.mode == B_PRED)
+    {
+        for (i = 0; i < 16; i++)
+            xd->mode_info_context->bmi[i].as_mode = best_mode.bmodes[i].as_mode;
+    }
+
+    if (best_mode.mbmode.mode == SPLITMV)
+    {
+        for (i = 0; i < 16; i++)
+            xd->mode_info_context->bmi[i].mv.as_int = best_mode.bmodes[i].mv.as_int;
+
+        memcpy(x->partition_info, &best_mode.partition, sizeof(PARTITION_INFO));
+
+        x->e_mbd.mode_info_context->mbmi.mv.as_int =
+                                      x->partition_info->bmi[15].mv.as_int;
+    }
+
+    if (sign_bias
+        != cpi->common.ref_frame_sign_bias[xd->mode_info_context->mbmi.ref_frame])
+        best_ref_mv.as_int = best_ref_mv_sb[!sign_bias].as_int;
+
+    rd_update_mvcount(x, &best_ref_mv);
+}
+
+void vp8_rd_pick_intra_mode(MACROBLOCK *x, int *rate_)
+{
+    int error4x4, error16x16;
+    int rate4x4, rate16x16 = 0, rateuv;
+    int dist4x4, dist16x16, distuv;
+    int rate;
+    int rate4x4_tokenonly = 0;
+    int rate16x16_tokenonly = 0;
+    int rateuv_tokenonly = 0;
+
+    x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME;
+
+    rd_pick_intra_mbuv_mode(x, &rateuv, &rateuv_tokenonly, &distuv);
+    rate = rateuv;
+
+    error16x16 = rd_pick_intra16x16mby_mode(x, &rate16x16, &rate16x16_tokenonly,
+                                            &dist16x16);
+
+    error4x4 = rd_pick_intra4x4mby_modes(x, &rate4x4, &rate4x4_tokenonly,
+                                         &dist4x4, error16x16);
+
+    if (error4x4 < error16x16)
+    {
+        x->e_mbd.mode_info_context->mbmi.mode = B_PRED;
+        rate += rate4x4;
+    }
+    else
+    {
+        rate += rate16x16;
+    }
+
+    *rate_ = rate;
+}
diff --git a/libvpx/libvpx/vp8/encoder/rdopt.h b/libvpx/libvpx/vp8/encoder/rdopt.h
new file mode 100644
index 0000000..1cb1a07
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/rdopt.h
@@ -0,0 +1,149 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_RDOPT_H_
+#define VP8_ENCODER_RDOPT_H_
+
+#include "./vpx_config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define RDCOST(RM,DM,R,D) ( ((128+(R)*(RM)) >> 8) + (DM)*(D) )
+
+static INLINE void insertsortmv(int arr[], int len)
+{
+    int i, j, k;
+
+    for ( i = 1 ; i <= len-1 ; i++ )
+    {
+        for ( j = 0 ; j < i ; j++ )
+        {
+            if ( arr[j] > arr[i] )
+            {
+                int temp;
+
+                temp = arr[i];
+
+                for ( k = i; k >j; k--)
+                    arr[k] = arr[k - 1] ;
+
+                arr[j] = temp ;
+            }
+        }
+    }
+}
+
+static INLINE void insertsortsad(int arr[],int idx[], int len)
+{
+    int i, j, k;
+
+    for ( i = 1 ; i <= len-1 ; i++ )
+    {
+        for ( j = 0 ; j < i ; j++ )
+        {
+            if ( arr[j] > arr[i] )
+            {
+                int temp, tempi;
+
+                temp = arr[i];
+                tempi = idx[i];
+
+                for ( k = i; k >j; k--)
+                {
+                    arr[k] = arr[k - 1] ;
+                    idx[k] = idx[k - 1];
+                }
+
+                arr[j] = temp ;
+                idx[j] = tempi;
+            }
+        }
+    }
+}
+
+extern void vp8_initialize_rd_consts(VP8_COMP *cpi, MACROBLOCK *x, int Qvalue);
+extern void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x,
+                                   int recon_yoffset, int recon_uvoffset,
+                                   int *returnrate, int *returndistortion,
+                                   int *returnintra, int mb_row, int mb_col);
+extern void vp8_rd_pick_intra_mode(MACROBLOCK *x, int *rate);
+
+
+static INLINE void get_plane_pointers(const YV12_BUFFER_CONFIG *fb,
+                                      unsigned char *plane[3],
+                                      unsigned int recon_yoffset,
+                                      unsigned int recon_uvoffset)
+{
+    plane[0] = fb->y_buffer + recon_yoffset;
+    plane[1] = fb->u_buffer + recon_uvoffset;
+    plane[2] = fb->v_buffer + recon_uvoffset;
+}
+
+
+static INLINE void get_predictor_pointers(const VP8_COMP *cpi,
+                                          unsigned char *plane[4][3],
+                                          unsigned int recon_yoffset,
+                                          unsigned int recon_uvoffset)
+{
+    if (cpi->ref_frame_flags & VP8_LAST_FRAME)
+        get_plane_pointers(&cpi->common.yv12_fb[cpi->common.lst_fb_idx],
+                           plane[LAST_FRAME], recon_yoffset, recon_uvoffset);
+
+    if (cpi->ref_frame_flags & VP8_GOLD_FRAME)
+        get_plane_pointers(&cpi->common.yv12_fb[cpi->common.gld_fb_idx],
+                           plane[GOLDEN_FRAME], recon_yoffset, recon_uvoffset);
+
+    if (cpi->ref_frame_flags & VP8_ALTR_FRAME)
+        get_plane_pointers(&cpi->common.yv12_fb[cpi->common.alt_fb_idx],
+                           plane[ALTREF_FRAME], recon_yoffset, recon_uvoffset);
+}
+
+
+static INLINE void get_reference_search_order(const VP8_COMP *cpi,
+                                              int ref_frame_map[4])
+{
+    int i=0;
+
+    ref_frame_map[i++] = INTRA_FRAME;
+    if (cpi->ref_frame_flags & VP8_LAST_FRAME)
+        ref_frame_map[i++] = LAST_FRAME;
+    if (cpi->ref_frame_flags & VP8_GOLD_FRAME)
+        ref_frame_map[i++] = GOLDEN_FRAME;
+    if (cpi->ref_frame_flags & VP8_ALTR_FRAME)
+        ref_frame_map[i++] = ALTREF_FRAME;
+    for(; i<4; i++)
+        ref_frame_map[i] = -1;
+}
+
+
+extern void vp8_mv_pred
+(
+    VP8_COMP *cpi,
+    MACROBLOCKD *xd,
+    const MODE_INFO *here,
+    int_mv *mvp,
+    int refframe,
+    int *ref_frame_sign_bias,
+    int *sr,
+    int near_sadidx[]
+);
+void vp8_cal_sad(VP8_COMP *cpi, MACROBLOCKD *xd, MACROBLOCK *x, int recon_yoffset, int near_sadidx[]);
+int VP8_UVSSE(MACROBLOCK *x);
+int vp8_cost_mv_ref(MB_PREDICTION_MODE m, const int near_mv_ref_ct[4]);
+void vp8_set_mbmode_and_mvs(MACROBLOCK *x, MB_PREDICTION_MODE mb, int_mv *mv);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_RDOPT_H_
diff --git a/libvpx/libvpx/vp8/encoder/segmentation.c b/libvpx/libvpx/vp8/encoder/segmentation.c
new file mode 100644
index 0000000..fdd22fc
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/segmentation.c
@@ -0,0 +1,66 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "segmentation.h"
+#include "vpx_mem/vpx_mem.h"
+
+void vp8_update_gf_useage_maps(VP8_COMP *cpi, VP8_COMMON *cm, MACROBLOCK *x)
+{
+    int mb_row, mb_col;
+
+    MODE_INFO *this_mb_mode_info = cm->mi;
+
+    x->gf_active_ptr = (signed char *)cpi->gf_active_flags;
+
+    if ((cm->frame_type == KEY_FRAME) || (cm->refresh_golden_frame))
+    {
+        /* Reset Gf useage monitors */
+        memset(cpi->gf_active_flags, 1, (cm->mb_rows * cm->mb_cols));
+        cpi->gf_active_count = cm->mb_rows * cm->mb_cols;
+    }
+    else
+    {
+        /* for each macroblock row in image */
+        for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
+        {
+            /* for each macroblock col in image */
+            for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
+            {
+
+                /* If using golden then set GF active flag if not already set.
+                 * If using last frame 0,0 mode then leave flag as it is
+                 * else if using non 0,0 motion or intra modes then clear
+                 * flag if it is currently set
+                 */
+                if ((this_mb_mode_info->mbmi.ref_frame == GOLDEN_FRAME) || (this_mb_mode_info->mbmi.ref_frame == ALTREF_FRAME))
+                {
+                    if (*(x->gf_active_ptr) == 0)
+                    {
+                        *(x->gf_active_ptr) = 1;
+                        cpi->gf_active_count ++;
+                    }
+                }
+                else if ((this_mb_mode_info->mbmi.mode != ZEROMV) && *(x->gf_active_ptr))
+                {
+                    *(x->gf_active_ptr) = 0;
+                    cpi->gf_active_count--;
+                }
+
+                x->gf_active_ptr++;          /* Step onto next entry */
+                this_mb_mode_info++;         /* skip to next mb */
+
+            }
+
+            /* this is to account for the border */
+            this_mb_mode_info++;
+        }
+    }
+}
diff --git a/libvpx/libvpx/vp8/encoder/segmentation.h b/libvpx/libvpx/vp8/encoder/segmentation.h
new file mode 100644
index 0000000..6b55005
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/segmentation.h
@@ -0,0 +1,28 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP8_ENCODER_SEGMENTATION_H_
+#define VP8_ENCODER_SEGMENTATION_H_
+
+#include "string.h"
+#include "vp8/common/blockd.h"
+#include "onyx_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern void vp8_update_gf_useage_maps(VP8_COMP *cpi, VP8_COMMON *cm, MACROBLOCK *x);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_SEGMENTATION_H_
diff --git a/libvpx/libvpx/vp8/encoder/temporal_filter.c b/libvpx/libvpx/vp8/encoder/temporal_filter.c
new file mode 100644
index 0000000..85d26c2
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/temporal_filter.c
@@ -0,0 +1,521 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vp8/common/onyxc_int.h"
+#include "onyx_int.h"
+#include "vp8/common/systemdependent.h"
+#include "vp8/encoder/quantize.h"
+#include "vp8/common/alloccommon.h"
+#include "mcomp.h"
+#include "firstpass.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vp8/common/extend.h"
+#include "ratectrl.h"
+#include "vp8/common/quant_common.h"
+#include "segmentation.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp8/common/swapyv12buffer.h"
+#include "vp8/common/threading.h"
+#include "vpx_ports/vpx_timer.h"
+
+#include <math.h>
+#include <limits.h>
+
+#define ALT_REF_MC_ENABLED 1    /* dis/enable MC in AltRef filtering */
+#define ALT_REF_SUBPEL_ENABLED 1 /* dis/enable subpel in MC AltRef filtering */
+
+#if VP8_TEMPORAL_ALT_REF
+
+static void vp8_temporal_filter_predictors_mb_c
+(
+    MACROBLOCKD *x,
+    unsigned char *y_mb_ptr,
+    unsigned char *u_mb_ptr,
+    unsigned char *v_mb_ptr,
+    int stride,
+    int mv_row,
+    int mv_col,
+    unsigned char *pred
+)
+{
+    int offset;
+    unsigned char *yptr, *uptr, *vptr;
+
+    /* Y */
+    yptr = y_mb_ptr + (mv_row >> 3) * stride + (mv_col >> 3);
+
+    if ((mv_row | mv_col) & 7)
+    {
+        x->subpixel_predict16x16(yptr, stride,
+                                    mv_col & 7, mv_row & 7, &pred[0], 16);
+    }
+    else
+    {
+        vp8_copy_mem16x16(yptr, stride, &pred[0], 16);
+    }
+
+    /* U & V */
+    mv_row >>= 1;
+    mv_col >>= 1;
+    stride = (stride + 1) >> 1;
+    offset = (mv_row >> 3) * stride + (mv_col >> 3);
+    uptr = u_mb_ptr + offset;
+    vptr = v_mb_ptr + offset;
+
+    if ((mv_row | mv_col) & 7)
+    {
+        x->subpixel_predict8x8(uptr, stride,
+                            mv_col & 7, mv_row & 7, &pred[256], 8);
+        x->subpixel_predict8x8(vptr, stride,
+                            mv_col & 7, mv_row & 7, &pred[320], 8);
+    }
+    else
+    {
+        vp8_copy_mem8x8(uptr, stride, &pred[256], 8);
+        vp8_copy_mem8x8(vptr, stride, &pred[320], 8);
+    }
+}
+void vp8_temporal_filter_apply_c
+(
+    unsigned char *frame1,
+    unsigned int stride,
+    unsigned char *frame2,
+    unsigned int block_size,
+    int strength,
+    int filter_weight,
+    unsigned int *accumulator,
+    unsigned short *count
+)
+{
+    unsigned int i, j, k;
+    int modifier;
+    int byte = 0;
+    const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
+
+    for (i = 0,k = 0; i < block_size; i++)
+    {
+        for (j = 0; j < block_size; j++, k++)
+        {
+
+            int src_byte = frame1[byte];
+            int pixel_value = *frame2++;
+
+            modifier   = src_byte - pixel_value;
+            /* This is an integer approximation of:
+             * float coeff = (3.0 * modifer * modifier) / pow(2, strength);
+             * modifier =  (int)roundf(coeff > 16 ? 0 : 16-coeff);
+             */
+            modifier  *= modifier;
+            modifier  *= 3;
+            modifier  += rounding;
+            modifier >>= strength;
+
+            if (modifier > 16)
+                modifier = 16;
+
+            modifier = 16 - modifier;
+            modifier *= filter_weight;
+
+            count[k] += modifier;
+            accumulator[k] += modifier * pixel_value;
+
+            byte++;
+        }
+
+        byte += stride - block_size;
+    }
+}
+
+#if ALT_REF_MC_ENABLED
+
+static int vp8_temporal_filter_find_matching_mb_c
+(
+    VP8_COMP *cpi,
+    YV12_BUFFER_CONFIG *arf_frame,
+    YV12_BUFFER_CONFIG *frame_ptr,
+    int mb_offset,
+    int error_thresh
+)
+{
+    MACROBLOCK *x = &cpi->mb;
+    int step_param;
+    int sadpb = x->sadperbit16;
+    int bestsme = INT_MAX;
+
+    BLOCK *b = &x->block[0];
+    BLOCKD *d = &x->e_mbd.block[0];
+    int_mv best_ref_mv1;
+    int_mv best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */
+
+    /* Save input state */
+    unsigned char **base_src = b->base_src;
+    int src = b->src;
+    int src_stride = b->src_stride;
+    unsigned char *base_pre = x->e_mbd.pre.y_buffer;
+    int pre = d->offset;
+    int pre_stride = x->e_mbd.pre.y_stride;
+
+    (void)error_thresh;
+
+    best_ref_mv1.as_int = 0;
+    best_ref_mv1_full.as_mv.col = best_ref_mv1.as_mv.col >>3;
+    best_ref_mv1_full.as_mv.row = best_ref_mv1.as_mv.row >>3;
+
+    /* Setup frame pointers */
+    b->base_src = &arf_frame->y_buffer;
+    b->src_stride = arf_frame->y_stride;
+    b->src = mb_offset;
+
+    x->e_mbd.pre.y_buffer = frame_ptr->y_buffer;
+    x->e_mbd.pre.y_stride = frame_ptr->y_stride;
+    d->offset = mb_offset;
+
+    /* Further step/diamond searches as necessary */
+    if (cpi->Speed < 8)
+    {
+        step_param = cpi->sf.first_step + (cpi->Speed > 5);
+    }
+    else
+    {
+        step_param = cpi->sf.first_step + 2;
+    }
+
+    /* TODO Check that the 16x16 vf & sdf are selected here */
+    /* Ignore mv costing by sending NULL cost arrays */
+    bestsme = vp8_hex_search(x, b, d, &best_ref_mv1_full, &d->bmi.mv,
+                             step_param, sadpb,
+                             &cpi->fn_ptr[BLOCK_16X16],
+                             NULL, NULL, &best_ref_mv1);
+
+#if ALT_REF_SUBPEL_ENABLED
+    /* Try sub-pixel MC? */
+    {
+        int distortion;
+        unsigned int sse;
+        /* Ignore mv costing by sending NULL cost array */
+        bestsme = cpi->find_fractional_mv_step(x, b, d,
+                                               &d->bmi.mv,
+                                               &best_ref_mv1,
+                                               x->errorperbit,
+                                               &cpi->fn_ptr[BLOCK_16X16],
+                                               NULL, &distortion, &sse);
+    }
+#endif
+
+    /* Save input state */
+    b->base_src = base_src;
+    b->src = src;
+    b->src_stride = src_stride;
+    x->e_mbd.pre.y_buffer = base_pre;
+    d->offset = pre;
+    x->e_mbd.pre.y_stride = pre_stride;
+
+    return bestsme;
+}
+#endif
+
+static void vp8_temporal_filter_iterate_c
+(
+    VP8_COMP *cpi,
+    int frame_count,
+    int alt_ref_index,
+    int strength
+)
+{
+    int byte;
+    int frame;
+    int mb_col, mb_row;
+    unsigned int filter_weight;
+    int mb_cols = cpi->common.mb_cols;
+    int mb_rows = cpi->common.mb_rows;
+    int mb_y_offset = 0;
+    int mb_uv_offset = 0;
+    DECLARE_ALIGNED(16, unsigned int, accumulator[16*16 + 8*8 + 8*8]);
+    DECLARE_ALIGNED(16, unsigned short, count[16*16 + 8*8 + 8*8]);
+    MACROBLOCKD *mbd = &cpi->mb.e_mbd;
+    YV12_BUFFER_CONFIG *f = cpi->frames[alt_ref_index];
+    unsigned char *dst1, *dst2;
+    DECLARE_ALIGNED(16, unsigned char,  predictor[16*16 + 8*8 + 8*8]);
+
+    /* Save input state */
+    unsigned char *y_buffer = mbd->pre.y_buffer;
+    unsigned char *u_buffer = mbd->pre.u_buffer;
+    unsigned char *v_buffer = mbd->pre.v_buffer;
+
+    for (mb_row = 0; mb_row < mb_rows; mb_row++)
+    {
+#if ALT_REF_MC_ENABLED
+        /* Source frames are extended to 16 pixels.  This is different than
+         *  L/A/G reference frames that have a border of 32 (VP8BORDERINPIXELS)
+         * A 6 tap filter is used for motion search.  This requires 2 pixels
+         *  before and 3 pixels after.  So the largest Y mv on a border would
+         *  then be 16 - 3.  The UV blocks are half the size of the Y and
+         *  therefore only extended by 8.  The largest mv that a UV block
+         *  can support is 8 - 3.  A UV mv is half of a Y mv.
+         *  (16 - 3) >> 1 == 6 which is greater than 8 - 3.
+         * To keep the mv in play for both Y and UV planes the max that it
+         *  can be on a border is therefore 16 - 5.
+         */
+        cpi->mb.mv_row_min = -((mb_row * 16) + (16 - 5));
+        cpi->mb.mv_row_max = ((cpi->common.mb_rows - 1 - mb_row) * 16)
+                                + (16 - 5);
+#endif
+
+        for (mb_col = 0; mb_col < mb_cols; mb_col++)
+        {
+            int i, j, k;
+            int stride;
+
+            memset(accumulator, 0, 384*sizeof(unsigned int));
+            memset(count, 0, 384*sizeof(unsigned short));
+
+#if ALT_REF_MC_ENABLED
+            cpi->mb.mv_col_min = -((mb_col * 16) + (16 - 5));
+            cpi->mb.mv_col_max = ((cpi->common.mb_cols - 1 - mb_col) * 16)
+                                    + (16 - 5);
+#endif
+
+            for (frame = 0; frame < frame_count; frame++)
+            {
+                if (cpi->frames[frame] == NULL)
+                    continue;
+
+                mbd->block[0].bmi.mv.as_mv.row = 0;
+                mbd->block[0].bmi.mv.as_mv.col = 0;
+
+                if (frame == alt_ref_index)
+                {
+                    filter_weight = 2;
+                }
+                else
+                {
+                    int err = 0;
+#if ALT_REF_MC_ENABLED
+#define THRESH_LOW   10000
+#define THRESH_HIGH  20000
+                    /* Find best match in this frame by MC */
+                    err = vp8_temporal_filter_find_matching_mb_c
+                              (cpi,
+                               cpi->frames[alt_ref_index],
+                               cpi->frames[frame],
+                               mb_y_offset,
+                               THRESH_LOW);
+#endif
+                    /* Assign higher weight to matching MB if it's error
+                     * score is lower. If not applying MC default behavior
+                     * is to weight all MBs equal.
+                     */
+                    filter_weight = err<THRESH_LOW
+                                       ? 2 : err<THRESH_HIGH ? 1 : 0;
+                }
+
+                if (filter_weight != 0)
+                {
+                    /* Construct the predictors */
+                    vp8_temporal_filter_predictors_mb_c
+                        (mbd,
+                         cpi->frames[frame]->y_buffer + mb_y_offset,
+                         cpi->frames[frame]->u_buffer + mb_uv_offset,
+                         cpi->frames[frame]->v_buffer + mb_uv_offset,
+                         cpi->frames[frame]->y_stride,
+                         mbd->block[0].bmi.mv.as_mv.row,
+                         mbd->block[0].bmi.mv.as_mv.col,
+                         predictor);
+
+                    /* Apply the filter (YUV) */
+                    vp8_temporal_filter_apply
+                        (f->y_buffer + mb_y_offset,
+                         f->y_stride,
+                         predictor,
+                         16,
+                         strength,
+                         filter_weight,
+                         accumulator,
+                         count);
+
+                    vp8_temporal_filter_apply
+                        (f->u_buffer + mb_uv_offset,
+                         f->uv_stride,
+                         predictor + 256,
+                         8,
+                         strength,
+                         filter_weight,
+                         accumulator + 256,
+                         count + 256);
+
+                    vp8_temporal_filter_apply
+                        (f->v_buffer + mb_uv_offset,
+                         f->uv_stride,
+                         predictor + 320,
+                         8,
+                         strength,
+                         filter_weight,
+                         accumulator + 320,
+                         count + 320);
+                }
+            }
+
+            /* Normalize filter output to produce AltRef frame */
+            dst1 = cpi->alt_ref_buffer.y_buffer;
+            stride = cpi->alt_ref_buffer.y_stride;
+            byte = mb_y_offset;
+            for (i = 0,k = 0; i < 16; i++)
+            {
+                for (j = 0; j < 16; j++, k++)
+                {
+                    unsigned int pval = accumulator[k] + (count[k] >> 1);
+                    pval *= cpi->fixed_divide[count[k]];
+                    pval >>= 19;
+
+                    dst1[byte] = (unsigned char)pval;
+
+                    /* move to next pixel */
+                    byte++;
+                }
+
+                byte += stride - 16;
+            }
+
+            dst1 = cpi->alt_ref_buffer.u_buffer;
+            dst2 = cpi->alt_ref_buffer.v_buffer;
+            stride = cpi->alt_ref_buffer.uv_stride;
+            byte = mb_uv_offset;
+            for (i = 0,k = 256; i < 8; i++)
+            {
+                for (j = 0; j < 8; j++, k++)
+                {
+                    int m=k+64;
+
+                    /* U */
+                    unsigned int pval = accumulator[k] + (count[k] >> 1);
+                    pval *= cpi->fixed_divide[count[k]];
+                    pval >>= 19;
+                    dst1[byte] = (unsigned char)pval;
+
+                    /* V */
+                    pval = accumulator[m] + (count[m] >> 1);
+                    pval *= cpi->fixed_divide[count[m]];
+                    pval >>= 19;
+                    dst2[byte] = (unsigned char)pval;
+
+                    /* move to next pixel */
+                    byte++;
+                }
+
+                byte += stride - 8;
+            }
+
+            mb_y_offset += 16;
+            mb_uv_offset += 8;
+        }
+
+        mb_y_offset += 16*(f->y_stride-mb_cols);
+        mb_uv_offset += 8*(f->uv_stride-mb_cols);
+    }
+
+    /* Restore input state */
+    mbd->pre.y_buffer = y_buffer;
+    mbd->pre.u_buffer = u_buffer;
+    mbd->pre.v_buffer = v_buffer;
+}
+
+void vp8_temporal_filter_prepare_c
+(
+    VP8_COMP *cpi,
+    int distance
+)
+{
+    int frame = 0;
+
+    int num_frames_backward = 0;
+    int num_frames_forward = 0;
+    int frames_to_blur_backward = 0;
+    int frames_to_blur_forward = 0;
+    int frames_to_blur = 0;
+    int start_frame = 0;
+
+    int strength = cpi->oxcf.arnr_strength;
+
+    int blur_type = cpi->oxcf.arnr_type;
+
+    int max_frames = cpi->active_arnr_frames;
+
+    num_frames_backward = distance;
+    num_frames_forward = vp8_lookahead_depth(cpi->lookahead)
+                         - (num_frames_backward + 1);
+
+    switch (blur_type)
+    {
+    case 1:
+        /* Backward Blur */
+
+        frames_to_blur_backward = num_frames_backward;
+
+        if (frames_to_blur_backward >= max_frames)
+            frames_to_blur_backward = max_frames - 1;
+
+        frames_to_blur = frames_to_blur_backward + 1;
+        break;
+
+    case 2:
+        /* Forward Blur */
+
+        frames_to_blur_forward = num_frames_forward;
+
+        if (frames_to_blur_forward >= max_frames)
+            frames_to_blur_forward = max_frames - 1;
+
+        frames_to_blur = frames_to_blur_forward + 1;
+        break;
+
+    case 3:
+    default:
+        /* Center Blur */
+        frames_to_blur_forward = num_frames_forward;
+        frames_to_blur_backward = num_frames_backward;
+
+        if (frames_to_blur_forward > frames_to_blur_backward)
+            frames_to_blur_forward = frames_to_blur_backward;
+
+        if (frames_to_blur_backward > frames_to_blur_forward)
+            frames_to_blur_backward = frames_to_blur_forward;
+
+        /* When max_frames is even we have 1 more frame backward than forward */
+        if (frames_to_blur_forward > (max_frames - 1) / 2)
+            frames_to_blur_forward = ((max_frames - 1) / 2);
+
+        if (frames_to_blur_backward > (max_frames / 2))
+            frames_to_blur_backward = (max_frames / 2);
+
+        frames_to_blur = frames_to_blur_backward + frames_to_blur_forward + 1;
+        break;
+    }
+
+    start_frame = distance + frames_to_blur_forward;
+
+    /* Setup frame pointers, NULL indicates frame not included in filter */
+    memset(cpi->frames, 0, max_frames*sizeof(YV12_BUFFER_CONFIG *));
+    for (frame = 0; frame < frames_to_blur; frame++)
+    {
+        int which_buffer =  start_frame - frame;
+        struct lookahead_entry* buf = vp8_lookahead_peek(cpi->lookahead,
+                                                         which_buffer,
+                                                         PEEK_FORWARD);
+        cpi->frames[frames_to_blur-1-frame] = &buf->img;
+    }
+
+    vp8_temporal_filter_iterate_c (
+        cpi,
+        frames_to_blur,
+        frames_to_blur_backward,
+        strength );
+}
+#endif
diff --git a/libvpx/libvpx/vp8/encoder/tokenize.c b/libvpx/libvpx/vp8/encoder/tokenize.c
new file mode 100644
index 0000000..afd46fb
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/tokenize.c
@@ -0,0 +1,608 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <math.h>
+#include <stdio.h>
+#include <string.h>
+#include <assert.h>
+#include "onyx_int.h"
+#include "tokenize.h"
+#include "vpx_mem/vpx_mem.h"
+
+/* Global event counters used for accumulating statistics across several
+   compressions, then generating context.c = initial stats. */
+
+#ifdef VP8_ENTROPY_STATS
+_int64 context_counters[BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [MAX_ENTROPY_TOKENS];
+#endif
+void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t) ;
+void vp8_fix_contexts(MACROBLOCKD *x);
+
+#include "dct_value_tokens.h"
+#include "dct_value_cost.h"
+
+const TOKENVALUE *const vp8_dct_value_tokens_ptr = dct_value_tokens +
+        DCT_MAX_VALUE;
+const short *const vp8_dct_value_cost_ptr = dct_value_cost + DCT_MAX_VALUE;
+
+#if 0
+int skip_true_count = 0;
+int skip_false_count = 0;
+#endif
+
+/* function used to generate dct_value_tokens and dct_value_cost tables */
+/*
+static void fill_value_tokens()
+{
+
+    TOKENVALUE *t = dct_value_tokens + DCT_MAX_VALUE;
+    const vp8_extra_bit_struct *e = vp8_extra_bits;
+
+    int i = -DCT_MAX_VALUE;
+    int sign = 1;
+
+    do
+    {
+        if (!i)
+            sign = 0;
+
+        {
+            const int a = sign ? -i : i;
+            int eb = sign;
+
+            if (a > 4)
+            {
+                int j = 4;
+
+                while (++j < 11  &&  e[j].base_val <= a) {}
+
+                t[i].Token = --j;
+                eb |= (a - e[j].base_val) << 1;
+            }
+            else
+                t[i].Token = a;
+
+            t[i].Extra = eb;
+        }
+
+        // initialize the cost for extra bits for all possible coefficient value.
+        {
+            int cost = 0;
+            const vp8_extra_bit_struct *p = vp8_extra_bits + t[i].Token;
+
+            if (p->base_val)
+            {
+                const int extra = t[i].Extra;
+                const int Length = p->Len;
+
+                if (Length)
+                    cost += vp8_treed_cost(p->tree, p->prob, extra >> 1, Length);
+
+                cost += vp8_cost_bit(vp8_prob_half, extra & 1); // sign
+                dct_value_cost[i + DCT_MAX_VALUE] = cost;
+            }
+
+        }
+
+    }
+    while (++i < DCT_MAX_VALUE);
+
+    vp8_dct_value_tokens_ptr = dct_value_tokens + DCT_MAX_VALUE;
+    vp8_dct_value_cost_ptr   = dct_value_cost + DCT_MAX_VALUE;
+}
+*/
+
+static void tokenize2nd_order_b
+(
+    MACROBLOCK *x,
+    TOKENEXTRA **tp,
+    VP8_COMP *cpi
+)
+{
+    MACROBLOCKD *xd = &x->e_mbd;
+    int pt;             /* near block/prev token context index */
+    int c;              /* start at DC */
+    TOKENEXTRA *t = *tp;/* store tokens starting here */
+    const BLOCKD *b;
+    const short *qcoeff_ptr;
+    ENTROPY_CONTEXT * a;
+    ENTROPY_CONTEXT * l;
+    int band, rc, v, token;
+    int eob;
+
+    b = xd->block + 24;
+    qcoeff_ptr = b->qcoeff;
+    a = (ENTROPY_CONTEXT *)xd->above_context + 8;
+    l = (ENTROPY_CONTEXT *)xd->left_context + 8;
+    eob = xd->eobs[24];
+    VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
+
+    if(!eob)
+    {
+        /* c = band for this case */
+        t->Token = DCT_EOB_TOKEN;
+        t->context_tree = cpi->common.fc.coef_probs [1] [0] [pt];
+        t->skip_eob_node = 0;
+
+        ++x->coef_counts       [1] [0] [pt] [DCT_EOB_TOKEN];
+        t++;
+        *tp = t;
+        *a = *l = 0;
+        return;
+    }
+
+    v = qcoeff_ptr[0];
+    t->Extra = vp8_dct_value_tokens_ptr[v].Extra;
+    token    = vp8_dct_value_tokens_ptr[v].Token;
+    t->Token = token;
+
+    t->context_tree = cpi->common.fc.coef_probs [1] [0] [pt];
+    t->skip_eob_node = 0;
+    ++x->coef_counts       [1] [0] [pt] [token];
+    pt = vp8_prev_token_class[token];
+    t++;
+    c = 1;
+
+    for (; c < eob; c++)
+    {
+        rc = vp8_default_zig_zag1d[c];
+        band = vp8_coef_bands[c];
+        v = qcoeff_ptr[rc];
+
+        t->Extra = vp8_dct_value_tokens_ptr[v].Extra;
+        token    = vp8_dct_value_tokens_ptr[v].Token;
+
+        t->Token = token;
+        t->context_tree = cpi->common.fc.coef_probs [1] [band] [pt];
+
+        t->skip_eob_node = ((pt == 0));
+
+        ++x->coef_counts       [1] [band] [pt] [token];
+
+        pt = vp8_prev_token_class[token];
+        t++;
+    }
+    if (c < 16)
+    {
+        band = vp8_coef_bands[c];
+        t->Token = DCT_EOB_TOKEN;
+        t->context_tree = cpi->common.fc.coef_probs [1] [band] [pt];
+
+        t->skip_eob_node = 0;
+
+        ++x->coef_counts       [1] [band] [pt] [DCT_EOB_TOKEN];
+
+        t++;
+    }
+
+    *tp = t;
+    *a = *l = 1;
+
+}
+
+static void tokenize1st_order_b
+(
+    MACROBLOCK *x,
+    TOKENEXTRA **tp,
+    int type,           /* which plane: 0=Y no DC, 1=Y2, 2=UV, 3=Y with DC */
+    VP8_COMP *cpi
+)
+{
+    MACROBLOCKD *xd = &x->e_mbd;
+    unsigned int block;
+    const BLOCKD *b;
+    int pt;             /* near block/prev token context index */
+    int c;
+    int token;
+    TOKENEXTRA *t = *tp;/* store tokens starting here */
+    const short *qcoeff_ptr;
+    ENTROPY_CONTEXT * a;
+    ENTROPY_CONTEXT * l;
+    int band, rc, v;
+    int tmp1, tmp2;
+
+    b = xd->block;
+    /* Luma */
+    for (block = 0; block < 16; block++, b++)
+    {
+        const int eob = *b->eob;
+        tmp1 = vp8_block2above[block];
+        tmp2 = vp8_block2left[block];
+        qcoeff_ptr = b->qcoeff;
+        a = (ENTROPY_CONTEXT *)xd->above_context + tmp1;
+        l = (ENTROPY_CONTEXT *)xd->left_context + tmp2;
+
+        VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
+
+        c = type ? 0 : 1;
+
+        if(c >= eob)
+        {
+            /* c = band for this case */
+            t->Token = DCT_EOB_TOKEN;
+            t->context_tree = cpi->common.fc.coef_probs [type] [c] [pt];
+            t->skip_eob_node = 0;
+
+            ++x->coef_counts       [type] [c] [pt] [DCT_EOB_TOKEN];
+            t++;
+            *tp = t;
+            *a = *l = 0;
+            continue;
+        }
+
+        v = qcoeff_ptr[c];
+
+        t->Extra = vp8_dct_value_tokens_ptr[v].Extra;
+        token    = vp8_dct_value_tokens_ptr[v].Token;
+        t->Token = token;
+
+        t->context_tree = cpi->common.fc.coef_probs [type] [c] [pt];
+        t->skip_eob_node = 0;
+        ++x->coef_counts       [type] [c] [pt] [token];
+        pt = vp8_prev_token_class[token];
+        t++;
+        c++;
+
+        assert(eob <= 16);
+        for (; c < eob; c++)
+        {
+            rc = vp8_default_zig_zag1d[c];
+            band = vp8_coef_bands[c];
+            v = qcoeff_ptr[rc];
+
+            t->Extra = vp8_dct_value_tokens_ptr[v].Extra;
+            token    = vp8_dct_value_tokens_ptr[v].Token;
+
+            t->Token = token;
+            t->context_tree = cpi->common.fc.coef_probs [type] [band] [pt];
+
+            t->skip_eob_node = (pt == 0);
+            ++x->coef_counts       [type] [band] [pt] [token];
+
+            pt = vp8_prev_token_class[token];
+            t++;
+        }
+        if (c < 16)
+        {
+            band = vp8_coef_bands[c];
+            t->Token = DCT_EOB_TOKEN;
+            t->context_tree = cpi->common.fc.coef_probs [type] [band] [pt];
+
+            t->skip_eob_node = 0;
+            ++x->coef_counts       [type] [band] [pt] [DCT_EOB_TOKEN];
+
+            t++;
+        }
+        *tp = t;
+        *a = *l = 1;
+    }
+
+    /* Chroma */
+    for (block = 16; block < 24; block++, b++)
+    {
+        const int eob = *b->eob;
+        tmp1 = vp8_block2above[block];
+        tmp2 = vp8_block2left[block];
+        qcoeff_ptr = b->qcoeff;
+        a = (ENTROPY_CONTEXT *)xd->above_context + tmp1;
+        l = (ENTROPY_CONTEXT *)xd->left_context + tmp2;
+
+        VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
+
+        if(!eob)
+        {
+            /* c = band for this case */
+            t->Token = DCT_EOB_TOKEN;
+            t->context_tree = cpi->common.fc.coef_probs [2] [0] [pt];
+            t->skip_eob_node = 0;
+
+            ++x->coef_counts       [2] [0] [pt] [DCT_EOB_TOKEN];
+            t++;
+            *tp = t;
+            *a = *l = 0;
+            continue;
+        }
+
+        v = qcoeff_ptr[0];
+
+        t->Extra = vp8_dct_value_tokens_ptr[v].Extra;
+        token    = vp8_dct_value_tokens_ptr[v].Token;
+        t->Token = token;
+
+        t->context_tree = cpi->common.fc.coef_probs [2] [0] [pt];
+        t->skip_eob_node = 0;
+        ++x->coef_counts       [2] [0] [pt] [token];
+        pt = vp8_prev_token_class[token];
+        t++;
+        c = 1;
+
+        assert(eob <= 16);
+        for (; c < eob; c++)
+        {
+            rc = vp8_default_zig_zag1d[c];
+            band = vp8_coef_bands[c];
+            v = qcoeff_ptr[rc];
+
+            t->Extra = vp8_dct_value_tokens_ptr[v].Extra;
+            token    = vp8_dct_value_tokens_ptr[v].Token;
+
+            t->Token = token;
+            t->context_tree = cpi->common.fc.coef_probs [2] [band] [pt];
+
+            t->skip_eob_node = (pt == 0);
+
+            ++x->coef_counts       [2] [band] [pt] [token];
+
+            pt = vp8_prev_token_class[token];
+            t++;
+        }
+        if (c < 16)
+        {
+            band = vp8_coef_bands[c];
+            t->Token = DCT_EOB_TOKEN;
+            t->context_tree = cpi->common.fc.coef_probs [2] [band] [pt];
+
+            t->skip_eob_node = 0;
+
+            ++x->coef_counts       [2] [band] [pt] [DCT_EOB_TOKEN];
+
+            t++;
+        }
+        *tp = t;
+        *a = *l = 1;
+    }
+}
+
+
+static int mb_is_skippable(MACROBLOCKD *x, int has_y2_block)
+{
+    int skip = 1;
+    int i = 0;
+
+    if (has_y2_block)
+    {
+        for (i = 0; i < 16; i++)
+            skip &= (x->eobs[i] < 2);
+    }
+
+    for (; i < 24 + has_y2_block; i++)
+        skip &= (!x->eobs[i]);
+
+    return skip;
+}
+
+
+void vp8_tokenize_mb(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t)
+{
+    MACROBLOCKD *xd = &x->e_mbd;
+    int plane_type;
+    int has_y2_block;
+
+    has_y2_block = (xd->mode_info_context->mbmi.mode != B_PRED
+                    && xd->mode_info_context->mbmi.mode != SPLITMV);
+
+    xd->mode_info_context->mbmi.mb_skip_coeff =
+        mb_is_skippable(xd, has_y2_block);
+    if (xd->mode_info_context->mbmi.mb_skip_coeff)
+    {
+        if (!cpi->common.mb_no_coeff_skip)
+        {
+            vp8_stuff_mb(cpi, x, t);
+        }
+        else
+        {
+            vp8_fix_contexts(xd);
+            x->skip_true_count++;
+        }
+
+        return;
+    }
+
+    plane_type = 3;
+    if(has_y2_block)
+    {
+        tokenize2nd_order_b(x, t, cpi);
+        plane_type = 0;
+    }
+
+    tokenize1st_order_b(x, t, plane_type, cpi);
+}
+
+
+#ifdef VP8_ENTROPY_STATS
+
+void init_context_counters(void)
+{
+    memset(context_counters, 0, sizeof(context_counters));
+}
+
+void print_context_counters()
+{
+
+    int type, band, pt, t;
+
+    FILE *const f = fopen("context.c", "w");
+
+    fprintf(f, "#include \"entropy.h\"\n");
+
+    fprintf(f, "\n/* *** GENERATED FILE: DO NOT EDIT *** */\n\n");
+
+    fprintf(f, "int Contexts[BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [MAX_ENTROPY_TOKENS];\n\n");
+
+    fprintf(f, "const int default_contexts[BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [MAX_ENTROPY_TOKENS] = {");
+
+# define Comma( X) (X? ",":"")
+
+    type = 0;
+
+    do
+    {
+        fprintf(f, "%s\n  { /* block Type %d */", Comma(type), type);
+
+        band = 0;
+
+        do
+        {
+            fprintf(f, "%s\n    { /* Coeff Band %d */", Comma(band), band);
+
+            pt = 0;
+
+            do
+            {
+                fprintf(f, "%s\n      {", Comma(pt));
+
+                t = 0;
+
+                do
+                {
+                    const _int64 x = context_counters [type] [band] [pt] [t];
+                    const int y = (int) x;
+
+                    assert(x == (_int64) y);  /* no overflow handling yet */
+                    fprintf(f, "%s %d", Comma(t), y);
+
+                }
+                while (++t < MAX_ENTROPY_TOKENS);
+
+                fprintf(f, "}");
+            }
+            while (++pt < PREV_COEF_CONTEXTS);
+
+            fprintf(f, "\n    }");
+
+        }
+        while (++band < COEF_BANDS);
+
+        fprintf(f, "\n  }");
+    }
+    while (++type < BLOCK_TYPES);
+
+    fprintf(f, "\n};\n");
+    fclose(f);
+}
+#endif
+
+
+static void stuff2nd_order_b
+(
+    TOKENEXTRA **tp,
+    ENTROPY_CONTEXT *a,
+    ENTROPY_CONTEXT *l,
+    VP8_COMP *cpi,
+    MACROBLOCK *x
+)
+{
+    int pt; /* near block/prev token context index */
+    TOKENEXTRA *t = *tp;        /* store tokens starting here */
+    VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
+
+    t->Token = DCT_EOB_TOKEN;
+    t->context_tree = cpi->common.fc.coef_probs [1] [0] [pt];
+    t->skip_eob_node = 0;
+    ++x->coef_counts       [1] [0] [pt] [DCT_EOB_TOKEN];
+    ++t;
+
+    *tp = t;
+    pt = 0;
+    *a = *l = pt;
+}
+
+static void stuff1st_order_b
+(
+    TOKENEXTRA **tp,
+    ENTROPY_CONTEXT *a,
+    ENTROPY_CONTEXT *l,
+    int type,
+    VP8_COMP *cpi,
+    MACROBLOCK *x
+)
+{
+    int pt; /* near block/prev token context index */
+    int band;
+    TOKENEXTRA *t = *tp;        /* store tokens starting here */
+    VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
+    band = type ? 0 : 1;
+    t->Token = DCT_EOB_TOKEN;
+    t->context_tree = cpi->common.fc.coef_probs [type] [band] [pt];
+    t->skip_eob_node = 0;
+    ++x->coef_counts       [type] [band] [pt] [DCT_EOB_TOKEN];
+    ++t;
+    *tp = t;
+    pt = 0; /* 0 <-> all coeff data is zero */
+    *a = *l = pt;
+}
+
+static
+void stuff1st_order_buv
+(
+    TOKENEXTRA **tp,
+    ENTROPY_CONTEXT *a,
+    ENTROPY_CONTEXT *l,
+    VP8_COMP *cpi,
+    MACROBLOCK *x
+)
+{
+    int pt; /* near block/prev token context index */
+    TOKENEXTRA *t = *tp;        /* store tokens starting here */
+    VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
+
+    t->Token = DCT_EOB_TOKEN;
+    t->context_tree = cpi->common.fc.coef_probs [2] [0] [pt];
+    t->skip_eob_node = 0;
+    ++x->coef_counts[2] [0] [pt] [DCT_EOB_TOKEN];
+    ++t;
+    *tp = t;
+    pt = 0; /* 0 <-> all coeff data is zero */
+    *a = *l = pt;
+}
+
+void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t)
+{
+    MACROBLOCKD *xd = &x->e_mbd;
+    ENTROPY_CONTEXT * A = (ENTROPY_CONTEXT *)xd->above_context;
+    ENTROPY_CONTEXT * L = (ENTROPY_CONTEXT *)xd->left_context;
+    int plane_type;
+    int b;
+    plane_type = 3;
+    if((xd->mode_info_context->mbmi.mode != B_PRED
+                        && xd->mode_info_context->mbmi.mode != SPLITMV))
+    {
+        stuff2nd_order_b(t,
+                     A + vp8_block2above[24], L + vp8_block2left[24], cpi, x);
+        plane_type = 0;
+    }
+
+    for (b = 0; b < 16; b++)
+        stuff1st_order_b(t,
+                         A + vp8_block2above[b],
+                         L + vp8_block2left[b], plane_type, cpi, x);
+
+    for (b = 16; b < 24; b++)
+        stuff1st_order_buv(t,
+                           A + vp8_block2above[b],
+                           L + vp8_block2left[b], cpi, x);
+
+}
+void vp8_fix_contexts(MACROBLOCKD *x)
+{
+    /* Clear entropy contexts for Y2 blocks */
+    if (x->mode_info_context->mbmi.mode != B_PRED && x->mode_info_context->mbmi.mode != SPLITMV)
+    {
+        memset(x->above_context, 0, sizeof(ENTROPY_CONTEXT_PLANES));
+        memset(x->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES));
+    }
+    else
+    {
+        memset(x->above_context, 0, sizeof(ENTROPY_CONTEXT_PLANES)-1);
+        memset(x->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES)-1);
+    }
+
+}
diff --git a/libvpx/libvpx/vp8/encoder/tokenize.h b/libvpx/libvpx/vp8/encoder/tokenize.h
new file mode 100644
index 0000000..b73a9ee
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/tokenize.h
@@ -0,0 +1,58 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_TOKENIZE_H_
+#define VP8_ENCODER_TOKENIZE_H_
+
+#include "vp8/common/entropy.h"
+#include "block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp8_tokenize_initialize();
+
+typedef struct
+{
+    short Token;
+    short Extra;
+} TOKENVALUE;
+
+typedef struct
+{
+    const vp8_prob *context_tree;
+    short           Extra;
+    unsigned char   Token;
+    unsigned char   skip_eob_node;
+} TOKENEXTRA;
+
+int rd_cost_mby(MACROBLOCKD *);
+
+#ifdef VP8_ENTROPY_STATS
+void init_context_counters();
+void print_context_counters();
+
+extern _int64 context_counters[BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [MAX_ENTROPY_TOKENS];
+#endif
+
+extern const short *const vp8_dct_value_cost_ptr;
+/* TODO: The Token field should be broken out into a separate char array to
+ *  improve cache locality, since it's needed for costing when the rest of the
+ *  fields are not.
+ */
+extern const TOKENVALUE *const vp8_dct_value_tokens_ptr;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_TOKENIZE_H_
diff --git a/libvpx/libvpx/vp8/encoder/treewriter.c b/libvpx/libvpx/vp8/encoder/treewriter.c
new file mode 100644
index 0000000..ef25f67
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/treewriter.c
@@ -0,0 +1,43 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "treewriter.h"
+
+static void cost(
+    int *const C,
+    vp8_tree T,
+    const vp8_prob *const P,
+    int i,
+    int c
+)
+{
+    const vp8_prob p = P [i>>1];
+
+    do
+    {
+        const vp8_tree_index j = T[i];
+        const int d = c + vp8_cost_bit(p, i & 1);
+
+        if (j <= 0)
+            C[-j] = d;
+        else
+            cost(C, T, P, j, d);
+    }
+    while (++i & 1);
+}
+void vp8_cost_tokens(int *c, const vp8_prob *p, vp8_tree t)
+{
+    cost(c, t, p, 0, 0);
+}
+void vp8_cost_tokens2(int *c, const vp8_prob *p, vp8_tree t,int start)
+{
+    cost(c, t, p, start, 0);
+}
diff --git a/libvpx/libvpx/vp8/encoder/treewriter.h b/libvpx/libvpx/vp8/encoder/treewriter.h
new file mode 100644
index 0000000..2debf92
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/treewriter.h
@@ -0,0 +1,135 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP8_ENCODER_TREEWRITER_H_
+#define VP8_ENCODER_TREEWRITER_H_
+
+/* Trees map alphabets into huffman-like codes suitable for an arithmetic
+   bit coder.  Timothy S Murphy  11 October 2004 */
+
+#include "./vpx_config.h"
+#include "vp8/common/treecoder.h"
+
+#include "boolhuff.h"       /* for now */
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef BOOL_CODER vp8_writer;
+
+#define vp8_write vp8_encode_bool
+#define vp8_write_literal vp8_encode_value
+#define vp8_write_bit( W, V) vp8_write( W, V, vp8_prob_half)
+
+#define vp8bc_write vp8bc_write_bool
+#define vp8bc_write_literal vp8bc_write_bits
+#define vp8bc_write_bit( W, V) vp8bc_write_bits( W, V, 1)
+
+
+/* Approximate length of an encoded bool in 256ths of a bit at given prob */
+
+#define vp8_cost_zero( x) ( vp8_prob_cost[x])
+#define vp8_cost_one( x)  vp8_cost_zero( vp8_complement(x))
+
+#define vp8_cost_bit( x, b) vp8_cost_zero( (b)?  vp8_complement(x) : (x) )
+
+/* VP8BC version is scaled by 2^20 rather than 2^8; see bool_coder.h */
+
+
+/* Both of these return bits, not scaled bits. */
+
+static INLINE unsigned int vp8_cost_branch(const unsigned int ct[2], vp8_prob p)
+{
+    /* Imitate existing calculation */
+
+    return ((ct[0] * vp8_cost_zero(p))
+            + (ct[1] * vp8_cost_one(p))) >> 8;
+}
+
+/* Small functions to write explicit values and tokens, as well as
+   estimate their lengths. */
+
+static void vp8_treed_write
+(
+    vp8_writer *const w,
+    vp8_tree t,
+    const vp8_prob *const p,
+    int v,
+    int n               /* number of bits in v, assumed nonzero */
+)
+{
+    vp8_tree_index i = 0;
+
+    do
+    {
+        const int b = (v >> --n) & 1;
+        vp8_write(w, b, p[i>>1]);
+        i = t[i+b];
+    }
+    while (n);
+}
+static INLINE void vp8_write_token
+(
+    vp8_writer *const w,
+    vp8_tree t,
+    const vp8_prob *const p,
+    vp8_token *const x
+)
+{
+    vp8_treed_write(w, t, p, x->value, x->Len);
+}
+
+static int vp8_treed_cost(
+    vp8_tree t,
+    const vp8_prob *const p,
+    int v,
+    int n               /* number of bits in v, assumed nonzero */
+)
+{
+    int c = 0;
+    vp8_tree_index i = 0;
+
+    do
+    {
+        const int b = (v >> --n) & 1;
+        c += vp8_cost_bit(p[i>>1], b);
+        i = t[i+b];
+    }
+    while (n);
+
+    return c;
+}
+static INLINE int vp8_cost_token
+(
+    vp8_tree t,
+    const vp8_prob *const p,
+    vp8_token *const x
+)
+{
+    return vp8_treed_cost(t, p, x->value, x->Len);
+}
+
+/* Fill array of costs for all possible token values. */
+
+void vp8_cost_tokens(
+    int *Costs, const vp8_prob *, vp8_tree
+);
+
+void vp8_cost_tokens2(
+    int *Costs, const vp8_prob *, vp8_tree, int
+);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP8_ENCODER_TREEWRITER_H_
diff --git a/libvpx/libvpx/vp8/encoder/vp8_quantize.c b/libvpx/libvpx/vp8/encoder/vp8_quantize.c
new file mode 100644
index 0000000..0d101ba
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/vp8_quantize.c
@@ -0,0 +1,583 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <math.h>
+#include "vpx_mem/vpx_mem.h"
+
+#include "onyx_int.h"
+#include "vp8/encoder/quantize.h"
+#include "vp8/common/quant_common.h"
+
+void vp8_fast_quantize_b_c(BLOCK *b, BLOCKD *d)
+{
+    int i, rc, eob;
+    int x, y, z, sz;
+    short *coeff_ptr   = b->coeff;
+    short *round_ptr   = b->round;
+    short *quant_ptr   = b->quant_fast;
+    short *qcoeff_ptr  = d->qcoeff;
+    short *dqcoeff_ptr = d->dqcoeff;
+    short *dequant_ptr = d->dequant;
+
+    eob = -1;
+    for (i = 0; i < 16; i++)
+    {
+        rc   = vp8_default_zig_zag1d[i];
+        z    = coeff_ptr[rc];
+
+        sz = (z >> 31);                              /* sign of z */
+        x  = (z ^ sz) - sz;                          /* x = abs(z) */
+
+        y  = ((x + round_ptr[rc]) * quant_ptr[rc]) >> 16; /* quantize (x) */
+        x  = (y ^ sz) - sz;                          /* get the sign back */
+        qcoeff_ptr[rc] = x;                          /* write to destination */
+        dqcoeff_ptr[rc] = x * dequant_ptr[rc];       /* dequantized value */
+
+        if (y)
+        {
+            eob = i;                                 /* last nonzero coeffs */
+        }
+    }
+    *d->eob = (char)(eob + 1);
+}
+
+void vp8_regular_quantize_b_c(BLOCK *b, BLOCKD *d)
+{
+    int i, rc, eob;
+    int zbin;
+    int x, y, z, sz;
+    short *zbin_boost_ptr  = b->zrun_zbin_boost;
+    short *coeff_ptr       = b->coeff;
+    short *zbin_ptr        = b->zbin;
+    short *round_ptr       = b->round;
+    short *quant_ptr       = b->quant;
+    short *quant_shift_ptr = b->quant_shift;
+    short *qcoeff_ptr      = d->qcoeff;
+    short *dqcoeff_ptr     = d->dqcoeff;
+    short *dequant_ptr     = d->dequant;
+    short zbin_oq_value    = b->zbin_extra;
+
+    memset(qcoeff_ptr, 0, 32);
+    memset(dqcoeff_ptr, 0, 32);
+
+    eob = -1;
+
+    for (i = 0; i < 16; i++)
+    {
+        rc   = vp8_default_zig_zag1d[i];
+        z    = coeff_ptr[rc];
+
+        zbin = zbin_ptr[rc] + *zbin_boost_ptr + zbin_oq_value;
+
+        zbin_boost_ptr ++;
+        sz = (z >> 31);                              /* sign of z */
+        x  = (z ^ sz) - sz;                          /* x = abs(z) */
+
+        if (x >= zbin)
+        {
+            x += round_ptr[rc];
+            y  = ((((x * quant_ptr[rc]) >> 16) + x)
+                 * quant_shift_ptr[rc]) >> 16;       /* quantize (x) */
+            x  = (y ^ sz) - sz;                      /* get the sign back */
+            qcoeff_ptr[rc]  = x;                     /* write to destination */
+            dqcoeff_ptr[rc] = x * dequant_ptr[rc];   /* dequantized value */
+
+            if (y)
+            {
+                eob = i;                             /* last nonzero coeffs */
+                zbin_boost_ptr = b->zrun_zbin_boost; /* reset zero runlength */
+            }
+        }
+    }
+
+    *d->eob = (char)(eob + 1);
+}
+
+void vp8_quantize_mby(MACROBLOCK *x)
+{
+    int i;
+    int has_2nd_order = (x->e_mbd.mode_info_context->mbmi.mode != B_PRED
+        && x->e_mbd.mode_info_context->mbmi.mode != SPLITMV);
+
+    for (i = 0; i < 16; i++)
+        x->quantize_b(&x->block[i], &x->e_mbd.block[i]);
+
+    if(has_2nd_order)
+        x->quantize_b(&x->block[24], &x->e_mbd.block[24]);
+}
+
+void vp8_quantize_mb(MACROBLOCK *x)
+{
+    int i;
+    int has_2nd_order=(x->e_mbd.mode_info_context->mbmi.mode != B_PRED
+        && x->e_mbd.mode_info_context->mbmi.mode != SPLITMV);
+
+    for (i = 0; i < 24+has_2nd_order; i++)
+        x->quantize_b(&x->block[i], &x->e_mbd.block[i]);
+}
+
+
+void vp8_quantize_mbuv(MACROBLOCK *x)
+{
+    int i;
+
+    for (i = 16; i < 24; i++)
+        x->quantize_b(&x->block[i], &x->e_mbd.block[i]);
+}
+
+static const int qrounding_factors[129] =
+{
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48
+};
+
+
+static const int qzbin_factors[129] =
+{
+    84, 84, 84, 84, 84, 84, 84, 84,
+    84, 84, 84, 84, 84, 84, 84, 84,
+    84, 84, 84, 84, 84, 84, 84, 84,
+    84, 84, 84, 84, 84, 84, 84, 84,
+    84, 84, 84, 84, 84, 84, 84, 84,
+    84, 84, 84, 84, 84, 84, 84, 84,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80
+};
+
+
+static const int qrounding_factors_y2[129] =
+{
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    48
+};
+
+
+static const int qzbin_factors_y2[129] =
+{
+    84, 84, 84, 84, 84, 84, 84, 84,
+    84, 84, 84, 84, 84, 84, 84, 84,
+    84, 84, 84, 84, 84, 84, 84, 84,
+    84, 84, 84, 84, 84, 84, 84, 84,
+    84, 84, 84, 84, 84, 84, 84, 84,
+    84, 84, 84, 84, 84, 84, 84, 84,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80, 80, 80, 80, 80, 80, 80, 80,
+    80
+};
+
+
+static void invert_quant(int improved_quant, short *quant,
+                         short *shift, short d)
+{
+    if(improved_quant)
+    {
+        unsigned t;
+        int l, m;
+        t = d;
+        for(l = 0; t > 1; l++)
+            t>>=1;
+        m = 1 + (1<<(16+l))/d;
+        *quant = (short)(m - (1<<16));
+        *shift = l;
+        /* use multiplication and constant shift by 16 */
+        *shift = 1 << (16 - *shift);
+    }
+    else
+    {
+        *quant = (1 << 16) / d;
+        *shift = 0;
+        /* use multiplication and constant shift by 16 */
+        *shift = 1 << (16 - *shift);
+    }
+}
+
+
+void vp8cx_init_quantizer(VP8_COMP *cpi)
+{
+    int i;
+    int quant_val;
+    int Q;
+
+    int zbin_boost[16] = {0, 0, 8, 10, 12, 14, 16, 20, 24, 28, 32, 36, 40, 44,
+                          44, 44};
+
+    for (Q = 0; Q < QINDEX_RANGE; Q++)
+    {
+        /* dc values */
+        quant_val = vp8_dc_quant(Q, cpi->common.y1dc_delta_q);
+        cpi->Y1quant_fast[Q][0] = (1 << 16) / quant_val;
+        invert_quant(cpi->sf.improved_quant, cpi->Y1quant[Q] + 0,
+                     cpi->Y1quant_shift[Q] + 0, quant_val);
+        cpi->Y1zbin[Q][0] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
+        cpi->Y1round[Q][0] = (qrounding_factors[Q] * quant_val) >> 7;
+        cpi->common.Y1dequant[Q][0] = quant_val;
+        cpi->zrun_zbin_boost_y1[Q][0] = (quant_val * zbin_boost[0]) >> 7;
+
+        quant_val = vp8_dc2quant(Q, cpi->common.y2dc_delta_q);
+        cpi->Y2quant_fast[Q][0] = (1 << 16) / quant_val;
+        invert_quant(cpi->sf.improved_quant, cpi->Y2quant[Q] + 0,
+                     cpi->Y2quant_shift[Q] + 0, quant_val);
+        cpi->Y2zbin[Q][0] = ((qzbin_factors_y2[Q] * quant_val) + 64) >> 7;
+        cpi->Y2round[Q][0] = (qrounding_factors_y2[Q] * quant_val) >> 7;
+        cpi->common.Y2dequant[Q][0] = quant_val;
+        cpi->zrun_zbin_boost_y2[Q][0] = (quant_val * zbin_boost[0]) >> 7;
+
+        quant_val = vp8_dc_uv_quant(Q, cpi->common.uvdc_delta_q);
+        cpi->UVquant_fast[Q][0] = (1 << 16) / quant_val;
+        invert_quant(cpi->sf.improved_quant, cpi->UVquant[Q] + 0,
+                     cpi->UVquant_shift[Q] + 0, quant_val);
+        cpi->UVzbin[Q][0] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;;
+        cpi->UVround[Q][0] = (qrounding_factors[Q] * quant_val) >> 7;
+        cpi->common.UVdequant[Q][0] = quant_val;
+        cpi->zrun_zbin_boost_uv[Q][0] = (quant_val * zbin_boost[0]) >> 7;
+
+        /* all the ac values = ; */
+        quant_val = vp8_ac_yquant(Q);
+        cpi->Y1quant_fast[Q][1] = (1 << 16) / quant_val;
+        invert_quant(cpi->sf.improved_quant, cpi->Y1quant[Q] + 1,
+                     cpi->Y1quant_shift[Q] + 1, quant_val);
+        cpi->Y1zbin[Q][1] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
+        cpi->Y1round[Q][1] = (qrounding_factors[Q] * quant_val) >> 7;
+        cpi->common.Y1dequant[Q][1] = quant_val;
+        cpi->zrun_zbin_boost_y1[Q][1] = (quant_val * zbin_boost[1]) >> 7;
+
+        quant_val = vp8_ac2quant(Q, cpi->common.y2ac_delta_q);
+        cpi->Y2quant_fast[Q][1] = (1 << 16) / quant_val;
+        invert_quant(cpi->sf.improved_quant, cpi->Y2quant[Q] + 1,
+                     cpi->Y2quant_shift[Q] + 1, quant_val);
+        cpi->Y2zbin[Q][1] = ((qzbin_factors_y2[Q] * quant_val) + 64) >> 7;
+        cpi->Y2round[Q][1] = (qrounding_factors_y2[Q] * quant_val) >> 7;
+        cpi->common.Y2dequant[Q][1] = quant_val;
+        cpi->zrun_zbin_boost_y2[Q][1] = (quant_val * zbin_boost[1]) >> 7;
+
+        quant_val = vp8_ac_uv_quant(Q, cpi->common.uvac_delta_q);
+        cpi->UVquant_fast[Q][1] = (1 << 16) / quant_val;
+        invert_quant(cpi->sf.improved_quant, cpi->UVquant[Q] + 1,
+                     cpi->UVquant_shift[Q] + 1, quant_val);
+        cpi->UVzbin[Q][1] = ((qzbin_factors[Q] * quant_val) + 64) >> 7;
+        cpi->UVround[Q][1] = (qrounding_factors[Q] * quant_val) >> 7;
+        cpi->common.UVdequant[Q][1] = quant_val;
+        cpi->zrun_zbin_boost_uv[Q][1] = (quant_val * zbin_boost[1]) >> 7;
+
+        for (i = 2; i < 16; i++)
+        {
+            cpi->Y1quant_fast[Q][i] = cpi->Y1quant_fast[Q][1];
+            cpi->Y1quant[Q][i] = cpi->Y1quant[Q][1];
+            cpi->Y1quant_shift[Q][i] = cpi->Y1quant_shift[Q][1];
+            cpi->Y1zbin[Q][i] = cpi->Y1zbin[Q][1];
+            cpi->Y1round[Q][i] = cpi->Y1round[Q][1];
+            cpi->zrun_zbin_boost_y1[Q][i] = (cpi->common.Y1dequant[Q][1] *
+                                             zbin_boost[i]) >> 7;
+
+            cpi->Y2quant_fast[Q][i] = cpi->Y2quant_fast[Q][1];
+            cpi->Y2quant[Q][i] = cpi->Y2quant[Q][1];
+            cpi->Y2quant_shift[Q][i] = cpi->Y2quant_shift[Q][1];
+            cpi->Y2zbin[Q][i] = cpi->Y2zbin[Q][1];
+            cpi->Y2round[Q][i] = cpi->Y2round[Q][1];
+            cpi->zrun_zbin_boost_y2[Q][i] = (cpi->common.Y2dequant[Q][1] *
+                                             zbin_boost[i]) >> 7;
+
+            cpi->UVquant_fast[Q][i] = cpi->UVquant_fast[Q][1];
+            cpi->UVquant[Q][i] = cpi->UVquant[Q][1];
+            cpi->UVquant_shift[Q][i] = cpi->UVquant_shift[Q][1];
+            cpi->UVzbin[Q][i] = cpi->UVzbin[Q][1];
+            cpi->UVround[Q][i] = cpi->UVround[Q][1];
+            cpi->zrun_zbin_boost_uv[Q][i] = (cpi->common.UVdequant[Q][1] *
+                                             zbin_boost[i]) >> 7;
+        }
+    }
+}
+
+#define ZBIN_EXTRA_Y \
+    (( cpi->common.Y1dequant[QIndex][1] *  \
+    ( x->zbin_over_quant +  \
+      x->zbin_mode_boost +  \
+      x->act_zbin_adj ) ) >> 7)
+
+#define ZBIN_EXTRA_UV \
+    (( cpi->common.UVdequant[QIndex][1] *  \
+    ( x->zbin_over_quant +  \
+      x->zbin_mode_boost +  \
+      x->act_zbin_adj ) ) >> 7)
+
+#define ZBIN_EXTRA_Y2 \
+    (( cpi->common.Y2dequant[QIndex][1] *  \
+    ( (x->zbin_over_quant / 2) +  \
+       x->zbin_mode_boost +  \
+       x->act_zbin_adj ) ) >> 7)
+
+void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x, int ok_to_skip)
+{
+    int i;
+    int QIndex;
+    MACROBLOCKD *xd = &x->e_mbd;
+    int zbin_extra;
+
+    /* Select the baseline MB Q index. */
+    if (xd->segmentation_enabled)
+    {
+        /* Abs Value */
+        if (xd->mb_segement_abs_delta == SEGMENT_ABSDATA)
+            QIndex = xd->segment_feature_data[MB_LVL_ALT_Q][xd->mode_info_context->mbmi.segment_id];
+        /* Delta Value */
+        else
+        {
+            QIndex = cpi->common.base_qindex + xd->segment_feature_data[MB_LVL_ALT_Q][xd->mode_info_context->mbmi.segment_id];
+            /* Clamp to valid range */
+            QIndex = (QIndex >= 0) ? ((QIndex <= MAXQ) ? QIndex : MAXQ) : 0;
+        }
+    }
+    else
+        QIndex = cpi->common.base_qindex;
+
+    /* This initialization should be called at least once. Use ok_to_skip to
+     * decide if it is ok to skip.
+     * Before encoding a frame, this function is always called with ok_to_skip
+     * =0, which means no skiping of calculations. The "last" values are
+     * initialized at that time.
+     */
+    if (!ok_to_skip || QIndex != x->q_index)
+    {
+
+        xd->dequant_y1_dc[0] = 1;
+        xd->dequant_y1[0] = cpi->common.Y1dequant[QIndex][0];
+        xd->dequant_y2[0] = cpi->common.Y2dequant[QIndex][0];
+        xd->dequant_uv[0] = cpi->common.UVdequant[QIndex][0];
+
+        for (i = 1; i < 16; i++)
+        {
+            xd->dequant_y1_dc[i] =
+            xd->dequant_y1[i] = cpi->common.Y1dequant[QIndex][1];
+            xd->dequant_y2[i] = cpi->common.Y2dequant[QIndex][1];
+            xd->dequant_uv[i] = cpi->common.UVdequant[QIndex][1];
+        }
+#if 1
+        /*TODO:  Remove dequant from BLOCKD.  This is a temporary solution until
+         * the quantizer code uses a passed in pointer to the dequant constants.
+         * This will also require modifications to the x86 and neon assembly.
+         * */
+        for (i = 0; i < 16; i++)
+            x->e_mbd.block[i].dequant = xd->dequant_y1;
+        for (i = 16; i < 24; i++)
+            x->e_mbd.block[i].dequant = xd->dequant_uv;
+        x->e_mbd.block[24].dequant = xd->dequant_y2;
+#endif
+
+        /* Y */
+        zbin_extra = ZBIN_EXTRA_Y;
+
+        for (i = 0; i < 16; i++)
+        {
+            x->block[i].quant = cpi->Y1quant[QIndex];
+            x->block[i].quant_fast = cpi->Y1quant_fast[QIndex];
+            x->block[i].quant_shift = cpi->Y1quant_shift[QIndex];
+            x->block[i].zbin = cpi->Y1zbin[QIndex];
+            x->block[i].round = cpi->Y1round[QIndex];
+            x->block[i].zrun_zbin_boost = cpi->zrun_zbin_boost_y1[QIndex];
+            x->block[i].zbin_extra = (short)zbin_extra;
+        }
+
+        /* UV */
+        zbin_extra = ZBIN_EXTRA_UV;
+
+        for (i = 16; i < 24; i++)
+        {
+            x->block[i].quant = cpi->UVquant[QIndex];
+            x->block[i].quant_fast = cpi->UVquant_fast[QIndex];
+            x->block[i].quant_shift = cpi->UVquant_shift[QIndex];
+            x->block[i].zbin = cpi->UVzbin[QIndex];
+            x->block[i].round = cpi->UVround[QIndex];
+            x->block[i].zrun_zbin_boost = cpi->zrun_zbin_boost_uv[QIndex];
+            x->block[i].zbin_extra = (short)zbin_extra;
+        }
+
+        /* Y2 */
+        zbin_extra = ZBIN_EXTRA_Y2;
+
+        x->block[24].quant_fast = cpi->Y2quant_fast[QIndex];
+        x->block[24].quant = cpi->Y2quant[QIndex];
+        x->block[24].quant_shift = cpi->Y2quant_shift[QIndex];
+        x->block[24].zbin = cpi->Y2zbin[QIndex];
+        x->block[24].round = cpi->Y2round[QIndex];
+        x->block[24].zrun_zbin_boost = cpi->zrun_zbin_boost_y2[QIndex];
+        x->block[24].zbin_extra = (short)zbin_extra;
+
+        /* save this macroblock QIndex for vp8_update_zbin_extra() */
+        x->q_index = QIndex;
+
+        x->last_zbin_over_quant = x->zbin_over_quant;
+        x->last_zbin_mode_boost = x->zbin_mode_boost;
+        x->last_act_zbin_adj = x->act_zbin_adj;
+
+
+
+    }
+    else if(x->last_zbin_over_quant != x->zbin_over_quant
+            || x->last_zbin_mode_boost != x->zbin_mode_boost
+            || x->last_act_zbin_adj != x->act_zbin_adj)
+    {
+        /* Y */
+        zbin_extra = ZBIN_EXTRA_Y;
+
+        for (i = 0; i < 16; i++)
+            x->block[i].zbin_extra = (short)zbin_extra;
+
+        /* UV */
+        zbin_extra = ZBIN_EXTRA_UV;
+
+        for (i = 16; i < 24; i++)
+            x->block[i].zbin_extra = (short)zbin_extra;
+
+        /* Y2 */
+        zbin_extra = ZBIN_EXTRA_Y2;
+        x->block[24].zbin_extra = (short)zbin_extra;
+
+        x->last_zbin_over_quant = x->zbin_over_quant;
+        x->last_zbin_mode_boost = x->zbin_mode_boost;
+        x->last_act_zbin_adj = x->act_zbin_adj;
+    }
+}
+
+void vp8_update_zbin_extra(VP8_COMP *cpi, MACROBLOCK *x)
+{
+    int i;
+    int QIndex = x->q_index;
+    int zbin_extra;
+
+    /* Y */
+    zbin_extra = ZBIN_EXTRA_Y;
+
+    for (i = 0; i < 16; i++)
+        x->block[i].zbin_extra = (short)zbin_extra;
+
+    /* UV */
+    zbin_extra = ZBIN_EXTRA_UV;
+
+    for (i = 16; i < 24; i++)
+        x->block[i].zbin_extra = (short)zbin_extra;
+
+    /* Y2 */
+    zbin_extra = ZBIN_EXTRA_Y2;
+    x->block[24].zbin_extra = (short)zbin_extra;
+}
+#undef ZBIN_EXTRA_Y
+#undef ZBIN_EXTRA_UV
+#undef ZBIN_EXTRA_Y2
+
+void vp8cx_frame_init_quantizer(VP8_COMP *cpi)
+{
+    /* Clear Zbin mode boost for default case */
+    cpi->mb.zbin_mode_boost = 0;
+
+    /* MB level quantizer setup */
+    vp8cx_mb_init_quantizer(cpi, &cpi->mb, 0);
+}
+
+
+void vp8_set_quantizer(struct VP8_COMP *cpi, int Q)
+{
+    VP8_COMMON *cm = &cpi->common;
+    MACROBLOCKD *mbd = &cpi->mb.e_mbd;
+    int update = 0;
+    int new_delta_q;
+    int new_uv_delta_q;
+    cm->base_qindex = Q;
+
+    /* if any of the delta_q values are changing update flag has to be set */
+    /* currently only y2dc_delta_q may change */
+
+    cm->y1dc_delta_q = 0;
+    cm->y2ac_delta_q = 0;
+
+    if (Q < 4)
+    {
+        new_delta_q = 4-Q;
+    }
+    else
+        new_delta_q = 0;
+
+    update |= cm->y2dc_delta_q != new_delta_q;
+    cm->y2dc_delta_q = new_delta_q;
+
+    new_uv_delta_q = 0;
+    // For screen content, lower the q value for UV channel. For now, select
+    // conservative delta; same delta for dc and ac, and decrease it with lower
+    // Q, and set to 0 below some threshold. May want to condition this in
+    // future on the variance/energy in UV channel.
+    if (cpi->oxcf.screen_content_mode && Q > 40) {
+      new_uv_delta_q = -(int)(0.15 * Q);
+      // Check range: magnitude of delta is 4 bits.
+      if (new_uv_delta_q < -15) {
+        new_uv_delta_q = -15;
+      }
+    }
+    update |= cm->uvdc_delta_q != new_uv_delta_q;
+    cm->uvdc_delta_q = new_uv_delta_q;
+    cm->uvac_delta_q = new_uv_delta_q;
+
+    /* Set Segment specific quatizers */
+    mbd->segment_feature_data[MB_LVL_ALT_Q][0] = cpi->segment_feature_data[MB_LVL_ALT_Q][0];
+    mbd->segment_feature_data[MB_LVL_ALT_Q][1] = cpi->segment_feature_data[MB_LVL_ALT_Q][1];
+    mbd->segment_feature_data[MB_LVL_ALT_Q][2] = cpi->segment_feature_data[MB_LVL_ALT_Q][2];
+    mbd->segment_feature_data[MB_LVL_ALT_Q][3] = cpi->segment_feature_data[MB_LVL_ALT_Q][3];
+
+    /* quantizer has to be reinitialized for any delta_q changes */
+    if(update)
+        vp8cx_init_quantizer(cpi);
+
+}
diff --git a/libvpx/libvpx/vp8/encoder/x86/dct_mmx.asm b/libvpx/libvpx/vp8/encoder/x86/dct_mmx.asm
new file mode 100644
index 0000000..6f188cb
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/dct_mmx.asm
@@ -0,0 +1,241 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vp8_short_fdct4x4_mmx(short *input, short *output, int pitch)
+global sym(vp8_short_fdct4x4_mmx) PRIVATE
+sym(vp8_short_fdct4x4_mmx):
+    push        rbp
+    mov         rbp,        rsp
+    SHADOW_ARGS_TO_STACK 3
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+        mov         rsi,        arg(0)      ; input
+        mov         rdi,        arg(1)      ; output
+
+        movsxd      rax,        dword ptr arg(2) ;pitch
+
+        lea         rcx,        [rsi + rax*2]
+        ; read the input data
+        movq        mm0,        [rsi]
+        movq        mm1,        [rsi + rax]
+
+        movq        mm2,        [rcx]
+        movq        mm4,        [rcx + rax]
+
+        ; transpose for the first stage
+        movq        mm3,        mm0         ; 00 01 02 03
+        movq        mm5,        mm2         ; 20 21 22 23
+
+        punpcklwd   mm0,        mm1         ; 00 10 01 11
+        punpckhwd   mm3,        mm1         ; 02 12 03 13
+
+        punpcklwd   mm2,        mm4         ; 20 30 21 31
+        punpckhwd   mm5,        mm4         ; 22 32 23 33
+
+        movq        mm1,        mm0         ; 00 10 01 11
+        punpckldq   mm0,        mm2         ; 00 10 20 30
+
+        punpckhdq   mm1,        mm2         ; 01 11 21 31
+
+        movq        mm2,        mm3         ; 02 12 03 13
+        punpckldq   mm2,        mm5         ; 02 12 22 32
+
+        punpckhdq   mm3,        mm5         ; 03 13 23 33
+
+        ; mm0 0
+        ; mm1 1
+        ; mm2 2
+        ; mm3 3
+
+        ; first stage
+        movq        mm5,        mm0
+        movq        mm4,        mm1
+
+        paddw       mm0,        mm3         ; a1 = 0 + 3
+        paddw       mm1,        mm2         ; b1 = 1 + 2
+
+        psubw       mm4,        mm2         ; c1 = 1 - 2
+        psubw       mm5,        mm3         ; d1 = 0 - 3
+
+        psllw       mm5,        3
+        psllw       mm4,        3
+
+        psllw       mm0,        3
+        psllw       mm1,        3
+
+        ; output 0 and 2
+        movq        mm2,        mm0         ; a1
+
+        paddw       mm0,        mm1         ; op[0] = a1 + b1
+        psubw       mm2,        mm1         ; op[2] = a1 - b1
+
+        ; output 1 and 3
+        ; interleave c1, d1
+        movq        mm1,        mm5         ; d1
+        punpcklwd   mm1,        mm4         ; c1 d1
+        punpckhwd   mm5,        mm4         ; c1 d1
+
+        movq        mm3,        mm1
+        movq        mm4,        mm5
+
+        pmaddwd     mm1,        MMWORD PTR[GLOBAL (_5352_2217)]    ; c1*2217 + d1*5352
+        pmaddwd     mm4,        MMWORD PTR[GLOBAL (_5352_2217)]    ; c1*2217 + d1*5352
+
+        pmaddwd     mm3,        MMWORD PTR[GLOBAL(_2217_neg5352)]  ; d1*2217 - c1*5352
+        pmaddwd     mm5,        MMWORD PTR[GLOBAL(_2217_neg5352)]  ; d1*2217 - c1*5352
+
+        paddd       mm1,        MMWORD PTR[GLOBAL(_14500)]
+        paddd       mm4,        MMWORD PTR[GLOBAL(_14500)]
+        paddd       mm3,        MMWORD PTR[GLOBAL(_7500)]
+        paddd       mm5,        MMWORD PTR[GLOBAL(_7500)]
+
+        psrad       mm1,        12          ; (c1 * 2217 + d1 * 5352 +  14500)>>12
+        psrad       mm4,        12          ; (c1 * 2217 + d1 * 5352 +  14500)>>12
+        psrad       mm3,        12          ; (d1 * 2217 - c1 * 5352 +   7500)>>12
+        psrad       mm5,        12          ; (d1 * 2217 - c1 * 5352 +   7500)>>12
+
+        packssdw    mm1,        mm4         ; op[1]
+        packssdw    mm3,        mm5         ; op[3]
+
+        ; done with vertical
+        ; transpose for the second stage
+        movq        mm4,        mm0         ; 00 10 20 30
+        movq        mm5,        mm2         ; 02 12 22 32
+
+        punpcklwd   mm0,        mm1         ; 00 01 10 11
+        punpckhwd   mm4,        mm1         ; 20 21 30 31
+
+        punpcklwd   mm2,        mm3         ; 02 03 12 13
+        punpckhwd   mm5,        mm3         ; 22 23 32 33
+
+        movq        mm1,        mm0         ; 00 01 10 11
+        punpckldq   mm0,        mm2         ; 00 01 02 03
+
+        punpckhdq   mm1,        mm2         ; 01 22 12 13
+
+        movq        mm2,        mm4         ; 20 31 30 31
+        punpckldq   mm2,        mm5         ; 20 21 22 23
+
+        punpckhdq   mm4,        mm5         ; 30 31 32 33
+
+        ; mm0 0
+        ; mm1 1
+        ; mm2 2
+        ; mm3 4
+
+        movq        mm5,        mm0
+        movq        mm3,        mm1
+
+        paddw       mm0,        mm4         ; a1 = 0 + 3
+        paddw       mm1,        mm2         ; b1 = 1 + 2
+
+        psubw       mm3,        mm2         ; c1 = 1 - 2
+        psubw       mm5,        mm4         ; d1 = 0 - 3
+
+        pxor        mm6,        mm6         ; zero out for compare
+
+        pcmpeqw     mm6,        mm5         ; d1 != 0
+
+        pandn       mm6,        MMWORD PTR[GLOBAL(_cmp_mask)]   ; clear upper,
+                                                                ; and keep bit 0 of lower
+
+        ; output 0 and 2
+        movq        mm2,        mm0         ; a1
+
+        paddw       mm0,        mm1         ; a1 + b1
+        psubw       mm2,        mm1         ; a1 - b1
+
+        paddw       mm0,        MMWORD PTR[GLOBAL(_7w)]
+        paddw       mm2,        MMWORD PTR[GLOBAL(_7w)]
+
+        psraw       mm0,        4           ; op[0] = (a1 + b1 + 7)>>4
+        psraw       mm2,        4           ; op[8] = (a1 - b1 + 7)>>4
+
+        movq        MMWORD PTR[rdi + 0 ],  mm0
+        movq        MMWORD PTR[rdi + 16],  mm2
+
+        ; output 1 and 3
+        ; interleave c1, d1
+        movq        mm1,        mm5         ; d1
+        punpcklwd   mm1,        mm3         ; c1 d1
+        punpckhwd   mm5,        mm3         ; c1 d1
+
+        movq        mm3,        mm1
+        movq        mm4,        mm5
+
+        pmaddwd     mm1,        MMWORD PTR[GLOBAL (_5352_2217)]    ; c1*2217 + d1*5352
+        pmaddwd     mm4,        MMWORD PTR[GLOBAL (_5352_2217)]    ; c1*2217 + d1*5352
+
+        pmaddwd     mm3,        MMWORD PTR[GLOBAL(_2217_neg5352)]  ; d1*2217 - c1*5352
+        pmaddwd     mm5,        MMWORD PTR[GLOBAL(_2217_neg5352)]  ; d1*2217 - c1*5352
+
+        paddd       mm1,        MMWORD PTR[GLOBAL(_12000)]
+        paddd       mm4,        MMWORD PTR[GLOBAL(_12000)]
+        paddd       mm3,        MMWORD PTR[GLOBAL(_51000)]
+        paddd       mm5,        MMWORD PTR[GLOBAL(_51000)]
+
+        psrad       mm1,        16          ; (c1 * 2217 + d1 * 5352 +  14500)>>16
+        psrad       mm4,        16          ; (c1 * 2217 + d1 * 5352 +  14500)>>16
+        psrad       mm3,        16          ; (d1 * 2217 - c1 * 5352 +   7500)>>16
+        psrad       mm5,        16          ; (d1 * 2217 - c1 * 5352 +   7500)>>16
+
+        packssdw    mm1,        mm4         ; op[4]
+        packssdw    mm3,        mm5         ; op[12]
+
+        paddw       mm1,        mm6         ; op[4] += (d1!=0)
+
+        movq        MMWORD PTR[rdi + 8 ],  mm1
+        movq        MMWORD PTR[rdi + 24],  mm3
+
+     ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 8
+_5352_2217:
+    dw 5352
+    dw 2217
+    dw 5352
+    dw 2217
+align 8
+_2217_neg5352:
+    dw 2217
+    dw -5352
+    dw 2217
+    dw -5352
+align 8
+_cmp_mask:
+    times 4 dw 1
+align 8
+_7w:
+    times 4 dw 7
+align 8
+_14500:
+    times 2 dd 14500
+align 8
+_7500:
+    times 2 dd 7500
+align 8
+_12000:
+    times 2 dd 12000
+align 8
+_51000:
+    times 2 dd 51000
diff --git a/libvpx/libvpx/vp8/encoder/x86/dct_sse2.asm b/libvpx/libvpx/vp8/encoder/x86/dct_sse2.asm
new file mode 100644
index 0000000..d06bca5
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/dct_sse2.asm
@@ -0,0 +1,432 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro STACK_FRAME_CREATE 0
+%if ABI_IS_32BIT
+  %define       input       rsi
+  %define       output      rdi
+  %define       pitch       rax
+    push        rbp
+    mov         rbp, rsp
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov         rsi, arg(0)
+    mov         rdi, arg(1)
+
+    movsxd      rax, dword ptr arg(2)
+    lea         rcx, [rsi + rax*2]
+%else
+  %if LIBVPX_YASM_WIN64
+    %define     input       rcx
+    %define     output      rdx
+    %define     pitch       r8
+    SAVE_XMM 7, u
+  %else
+    %define     input       rdi
+    %define     output      rsi
+    %define     pitch       rdx
+  %endif
+%endif
+%endmacro
+
+%macro STACK_FRAME_DESTROY 0
+  %define     input
+  %define     output
+  %define     pitch
+
+%if ABI_IS_32BIT
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    pop         rbp
+%else
+  %if LIBVPX_YASM_WIN64
+    RESTORE_XMM
+  %endif
+%endif
+    ret
+%endmacro
+
+;void vp8_short_fdct4x4_sse2(short *input, short *output, int pitch)
+global sym(vp8_short_fdct4x4_sse2) PRIVATE
+sym(vp8_short_fdct4x4_sse2):
+
+    STACK_FRAME_CREATE
+
+    movq        xmm0, MMWORD PTR[input        ] ;03 02 01 00
+    movq        xmm2, MMWORD PTR[input+  pitch] ;13 12 11 10
+    lea         input,          [input+2*pitch]
+    movq        xmm1, MMWORD PTR[input        ] ;23 22 21 20
+    movq        xmm3, MMWORD PTR[input+  pitch] ;33 32 31 30
+
+    punpcklqdq  xmm0, xmm2                      ;13 12 11 10 03 02 01 00
+    punpcklqdq  xmm1, xmm3                      ;33 32 31 30 23 22 21 20
+
+    movdqa      xmm2, xmm0
+    punpckldq   xmm0, xmm1                      ;23 22 03 02 21 20 01 00
+    punpckhdq   xmm2, xmm1                      ;33 32 13 12 31 30 11 10
+    movdqa      xmm1, xmm0
+    punpckldq   xmm0, xmm2                      ;31 21 30 20 11 10 01 00
+    pshufhw     xmm1, xmm1, 0b1h                ;22 23 02 03 xx xx xx xx
+    pshufhw     xmm2, xmm2, 0b1h                ;32 33 12 13 xx xx xx xx
+
+    punpckhdq   xmm1, xmm2                      ;32 33 22 23 12 13 02 03
+    movdqa      xmm3, xmm0
+    paddw       xmm0, xmm1                      ;b1 a1 b1 a1 b1 a1 b1 a1
+    psubw       xmm3, xmm1                      ;c1 d1 c1 d1 c1 d1 c1 d1
+    psllw       xmm0, 3                         ;b1 <<= 3 a1 <<= 3
+    psllw       xmm3, 3                         ;c1 <<= 3 d1 <<= 3
+
+    movdqa      xmm1, xmm0
+    pmaddwd     xmm0, XMMWORD PTR[GLOBAL(_mult_add)]    ;a1 + b1
+    pmaddwd     xmm1, XMMWORD PTR[GLOBAL(_mult_sub)]    ;a1 - b1
+    movdqa      xmm4, xmm3
+    pmaddwd     xmm3, XMMWORD PTR[GLOBAL(_5352_2217)]   ;c1*2217 + d1*5352
+    pmaddwd     xmm4, XMMWORD PTR[GLOBAL(_2217_neg5352)];d1*2217 - c1*5352
+
+    paddd       xmm3, XMMWORD PTR[GLOBAL(_14500)]
+    paddd       xmm4, XMMWORD PTR[GLOBAL(_7500)]
+    psrad       xmm3, 12            ;(c1 * 2217 + d1 * 5352 +  14500)>>12
+    psrad       xmm4, 12            ;(d1 * 2217 - c1 * 5352 +   7500)>>12
+
+    packssdw    xmm0, xmm1                      ;op[2] op[0]
+    packssdw    xmm3, xmm4                      ;op[3] op[1]
+    ; 23 22 21 20 03 02 01 00
+    ;
+    ; 33 32 31 30 13 12 11 10
+    ;
+    movdqa      xmm2, xmm0
+    punpcklqdq  xmm0, xmm3                      ;13 12 11 10 03 02 01 00
+    punpckhqdq  xmm2, xmm3                      ;23 22 21 20 33 32 31 30
+
+    movdqa      xmm3, xmm0
+    punpcklwd   xmm0, xmm2                      ;32 30 22 20 12 10 02 00
+    punpckhwd   xmm3, xmm2                      ;33 31 23 21 13 11 03 01
+    movdqa      xmm2, xmm0
+    punpcklwd   xmm0, xmm3                      ;13 12 11 10 03 02 01 00
+    punpckhwd   xmm2, xmm3                      ;33 32 31 30 23 22 21 20
+
+    movdqa      xmm5, XMMWORD PTR[GLOBAL(_7)]
+    pshufd      xmm2, xmm2, 04eh
+    movdqa      xmm3, xmm0
+    paddw       xmm0, xmm2                      ;b1 b1 b1 b1 a1 a1 a1 a1
+    psubw       xmm3, xmm2                      ;c1 c1 c1 c1 d1 d1 d1 d1
+
+    pshufd      xmm0, xmm0, 0d8h                ;b1 b1 a1 a1 b1 b1 a1 a1
+    movdqa      xmm2, xmm3                      ;save d1 for compare
+    pshufd      xmm3, xmm3, 0d8h                ;c1 c1 d1 d1 c1 c1 d1 d1
+    pshuflw     xmm0, xmm0, 0d8h                ;b1 b1 a1 a1 b1 a1 b1 a1
+    pshuflw     xmm3, xmm3, 0d8h                ;c1 c1 d1 d1 c1 d1 c1 d1
+    pshufhw     xmm0, xmm0, 0d8h                ;b1 a1 b1 a1 b1 a1 b1 a1
+    pshufhw     xmm3, xmm3, 0d8h                ;c1 d1 c1 d1 c1 d1 c1 d1
+    movdqa      xmm1, xmm0
+    pmaddwd     xmm0, XMMWORD PTR[GLOBAL(_mult_add)] ;a1 + b1
+    pmaddwd     xmm1, XMMWORD PTR[GLOBAL(_mult_sub)] ;a1 - b1
+
+    pxor        xmm4, xmm4                      ;zero out for compare
+    paddd       xmm0, xmm5
+    paddd       xmm1, xmm5
+    pcmpeqw     xmm2, xmm4
+    psrad       xmm0, 4                         ;(a1 + b1 + 7)>>4
+    psrad       xmm1, 4                         ;(a1 - b1 + 7)>>4
+    pandn       xmm2, XMMWORD PTR[GLOBAL(_cmp_mask)] ;clear upper,
+                                                     ;and keep bit 0 of lower
+
+    movdqa      xmm4, xmm3
+    pmaddwd     xmm3, XMMWORD PTR[GLOBAL(_5352_2217)]    ;c1*2217 + d1*5352
+    pmaddwd     xmm4, XMMWORD PTR[GLOBAL(_2217_neg5352)] ;d1*2217 - c1*5352
+    paddd       xmm3, XMMWORD PTR[GLOBAL(_12000)]
+    paddd       xmm4, XMMWORD PTR[GLOBAL(_51000)]
+    packssdw    xmm0, xmm1                      ;op[8] op[0]
+    psrad       xmm3, 16                ;(c1 * 2217 + d1 * 5352 +  12000)>>16
+    psrad       xmm4, 16                ;(d1 * 2217 - c1 * 5352 +  51000)>>16
+
+    packssdw    xmm3, xmm4                      ;op[12] op[4]
+    movdqa      xmm1, xmm0
+    paddw       xmm3, xmm2                      ;op[4] += (d1!=0)
+    punpcklqdq  xmm0, xmm3                      ;op[4] op[0]
+    punpckhqdq  xmm1, xmm3                      ;op[12] op[8]
+
+    movdqa      XMMWORD PTR[output +  0], xmm0
+    movdqa      XMMWORD PTR[output + 16], xmm1
+
+    STACK_FRAME_DESTROY
+
+;void vp8_short_fdct8x4_sse2(short *input, short *output, int pitch)
+global sym(vp8_short_fdct8x4_sse2) PRIVATE
+sym(vp8_short_fdct8x4_sse2):
+
+    STACK_FRAME_CREATE
+
+        ; read the input data
+        movdqa      xmm0,       [input        ]
+        movdqa      xmm2,       [input+  pitch]
+        lea         input,      [input+2*pitch]
+        movdqa      xmm4,       [input        ]
+        movdqa      xmm3,       [input+  pitch]
+
+        ; transpose for the first stage
+        movdqa      xmm1,       xmm0        ; 00 01 02 03 04 05 06 07
+        movdqa      xmm5,       xmm4        ; 20 21 22 23 24 25 26 27
+
+        punpcklwd   xmm0,       xmm2        ; 00 10 01 11 02 12 03 13
+        punpckhwd   xmm1,       xmm2        ; 04 14 05 15 06 16 07 17
+
+        punpcklwd   xmm4,       xmm3        ; 20 30 21 31 22 32 23 33
+        punpckhwd   xmm5,       xmm3        ; 24 34 25 35 26 36 27 37
+
+        movdqa      xmm2,       xmm0        ; 00 10 01 11 02 12 03 13
+        punpckldq   xmm0,       xmm4        ; 00 10 20 30 01 11 21 31
+
+        punpckhdq   xmm2,       xmm4        ; 02 12 22 32 03 13 23 33
+
+        movdqa      xmm4,       xmm1        ; 04 14 05 15 06 16 07 17
+        punpckldq   xmm4,       xmm5        ; 04 14 24 34 05 15 25 35
+
+        punpckhdq   xmm1,       xmm5        ; 06 16 26 36 07 17 27 37
+        movdqa      xmm3,       xmm2        ; 02 12 22 32 03 13 23 33
+
+        punpckhqdq  xmm3,       xmm1        ; 03 13 23 33 07 17 27 37
+        punpcklqdq  xmm2,       xmm1        ; 02 12 22 32 06 16 26 36
+
+        movdqa      xmm1,       xmm0        ; 00 10 20 30 01 11 21 31
+        punpcklqdq  xmm0,       xmm4        ; 00 10 20 30 04 14 24 34
+
+        punpckhqdq  xmm1,       xmm4        ; 01 11 21 32 05 15 25 35
+
+        ; xmm0 0
+        ; xmm1 1
+        ; xmm2 2
+        ; xmm3 3
+
+        ; first stage
+        movdqa      xmm5,       xmm0
+        movdqa      xmm4,       xmm1
+
+        paddw       xmm0,       xmm3        ; a1 = 0 + 3
+        paddw       xmm1,       xmm2        ; b1 = 1 + 2
+
+        psubw       xmm4,       xmm2        ; c1 = 1 - 2
+        psubw       xmm5,       xmm3        ; d1 = 0 - 3
+
+        psllw       xmm5,        3
+        psllw       xmm4,        3
+
+        psllw       xmm0,        3
+        psllw       xmm1,        3
+
+        ; output 0 and 2
+        movdqa      xmm2,       xmm0        ; a1
+
+        paddw       xmm0,       xmm1        ; op[0] = a1 + b1
+        psubw       xmm2,       xmm1        ; op[2] = a1 - b1
+
+        ; output 1 and 3
+        ; interleave c1, d1
+        movdqa      xmm1,       xmm5        ; d1
+        punpcklwd   xmm1,       xmm4        ; c1 d1
+        punpckhwd   xmm5,       xmm4        ; c1 d1
+
+        movdqa      xmm3,       xmm1
+        movdqa      xmm4,       xmm5
+
+        pmaddwd     xmm1,       XMMWORD PTR[GLOBAL (_5352_2217)]    ; c1*2217 + d1*5352
+        pmaddwd     xmm4,       XMMWORD PTR[GLOBAL (_5352_2217)]    ; c1*2217 + d1*5352
+
+        pmaddwd     xmm3,       XMMWORD PTR[GLOBAL(_2217_neg5352)]  ; d1*2217 - c1*5352
+        pmaddwd     xmm5,       XMMWORD PTR[GLOBAL(_2217_neg5352)]  ; d1*2217 - c1*5352
+
+        paddd       xmm1,       XMMWORD PTR[GLOBAL(_14500)]
+        paddd       xmm4,       XMMWORD PTR[GLOBAL(_14500)]
+        paddd       xmm3,       XMMWORD PTR[GLOBAL(_7500)]
+        paddd       xmm5,       XMMWORD PTR[GLOBAL(_7500)]
+
+        psrad       xmm1,       12          ; (c1 * 2217 + d1 * 5352 +  14500)>>12
+        psrad       xmm4,       12          ; (c1 * 2217 + d1 * 5352 +  14500)>>12
+        psrad       xmm3,       12          ; (d1 * 2217 - c1 * 5352 +   7500)>>12
+        psrad       xmm5,       12          ; (d1 * 2217 - c1 * 5352 +   7500)>>12
+
+        packssdw    xmm1,       xmm4        ; op[1]
+        packssdw    xmm3,       xmm5        ; op[3]
+
+        ; done with vertical
+        ; transpose for the second stage
+        movdqa      xmm4,       xmm0         ; 00 10 20 30 04 14 24 34
+        movdqa      xmm5,       xmm2         ; 02 12 22 32 06 16 26 36
+
+        punpcklwd   xmm0,       xmm1         ; 00 01 10 11 20 21 30 31
+        punpckhwd   xmm4,       xmm1         ; 04 05 14 15 24 25 34 35
+
+        punpcklwd   xmm2,       xmm3         ; 02 03 12 13 22 23 32 33
+        punpckhwd   xmm5,       xmm3         ; 06 07 16 17 26 27 36 37
+
+        movdqa      xmm1,       xmm0         ; 00 01 10 11 20 21 30 31
+        punpckldq   xmm0,       xmm2         ; 00 01 02 03 10 11 12 13
+
+        punpckhdq   xmm1,       xmm2         ; 20 21 22 23 30 31 32 33
+
+        movdqa      xmm2,       xmm4         ; 04 05 14 15 24 25 34 35
+        punpckldq   xmm2,       xmm5         ; 04 05 06 07 14 15 16 17
+
+        punpckhdq   xmm4,       xmm5         ; 24 25 26 27 34 35 36 37
+        movdqa      xmm3,       xmm1         ; 20 21 22 23 30 31 32 33
+
+        punpckhqdq  xmm3,       xmm4         ; 30 31 32 33 34 35 36 37
+        punpcklqdq  xmm1,       xmm4         ; 20 21 22 23 24 25 26 27
+
+        movdqa      xmm4,       xmm0         ; 00 01 02 03 10 11 12 13
+        punpcklqdq  xmm0,       xmm2         ; 00 01 02 03 04 05 06 07
+
+        punpckhqdq  xmm4,       xmm2         ; 10 11 12 13 14 15 16 17
+
+        ; xmm0 0
+        ; xmm1 4
+        ; xmm2 1
+        ; xmm3 3
+
+        movdqa      xmm5,       xmm0
+        movdqa      xmm2,       xmm1
+
+        paddw       xmm0,       xmm3        ; a1 = 0 + 3
+        paddw       xmm1,       xmm4        ; b1 = 1 + 2
+
+        psubw       xmm4,       xmm2        ; c1 = 1 - 2
+        psubw       xmm5,       xmm3        ; d1 = 0 - 3
+
+        pxor        xmm6,       xmm6        ; zero out for compare
+
+        pcmpeqw     xmm6,       xmm5        ; d1 != 0
+
+        pandn       xmm6,       XMMWORD PTR[GLOBAL(_cmp_mask8x4)]   ; clear upper,
+                                                                    ; and keep bit 0 of lower
+
+        ; output 0 and 2
+        movdqa      xmm2,       xmm0        ; a1
+
+        paddw       xmm0,       xmm1        ; a1 + b1
+        psubw       xmm2,       xmm1        ; a1 - b1
+
+        paddw       xmm0,       XMMWORD PTR[GLOBAL(_7w)]
+        paddw       xmm2,       XMMWORD PTR[GLOBAL(_7w)]
+
+        psraw       xmm0,       4           ; op[0] = (a1 + b1 + 7)>>4
+        psraw       xmm2,       4           ; op[8] = (a1 - b1 + 7)>>4
+
+        ; output 1 and 3
+        ; interleave c1, d1
+        movdqa      xmm1,       xmm5        ; d1
+        punpcklwd   xmm1,       xmm4        ; c1 d1
+        punpckhwd   xmm5,       xmm4        ; c1 d1
+
+        movdqa      xmm3,       xmm1
+        movdqa      xmm4,       xmm5
+
+        pmaddwd     xmm1,       XMMWORD PTR[GLOBAL (_5352_2217)]    ; c1*2217 + d1*5352
+        pmaddwd     xmm4,       XMMWORD PTR[GLOBAL (_5352_2217)]    ; c1*2217 + d1*5352
+
+        pmaddwd     xmm3,       XMMWORD PTR[GLOBAL(_2217_neg5352)]  ; d1*2217 - c1*5352
+        pmaddwd     xmm5,       XMMWORD PTR[GLOBAL(_2217_neg5352)]  ; d1*2217 - c1*5352
+
+        paddd       xmm1,       XMMWORD PTR[GLOBAL(_12000)]
+        paddd       xmm4,       XMMWORD PTR[GLOBAL(_12000)]
+        paddd       xmm3,       XMMWORD PTR[GLOBAL(_51000)]
+        paddd       xmm5,       XMMWORD PTR[GLOBAL(_51000)]
+
+        psrad       xmm1,       16          ; (c1 * 2217 + d1 * 5352 +  14500)>>16
+        psrad       xmm4,       16          ; (c1 * 2217 + d1 * 5352 +  14500)>>16
+        psrad       xmm3,       16          ; (d1 * 2217 - c1 * 5352 +   7500)>>16
+        psrad       xmm5,       16          ; (d1 * 2217 - c1 * 5352 +   7500)>>16
+
+        packssdw    xmm1,       xmm4        ; op[4]
+        packssdw    xmm3,       xmm5        ; op[12]
+
+        paddw       xmm1,       xmm6        ; op[4] += (d1!=0)
+
+        movdqa      xmm4,       xmm0
+        movdqa      xmm5,       xmm2
+
+        punpcklqdq  xmm0,       xmm1
+        punpckhqdq  xmm4,       xmm1
+
+        punpcklqdq  xmm2,       xmm3
+        punpckhqdq  xmm5,       xmm3
+
+        movdqa      XMMWORD PTR[output + 0 ],  xmm0
+        movdqa      XMMWORD PTR[output + 16],  xmm2
+        movdqa      XMMWORD PTR[output + 32],  xmm4
+        movdqa      XMMWORD PTR[output + 48],  xmm5
+
+    STACK_FRAME_DESTROY
+
+SECTION_RODATA
+align 16
+_5352_2217:
+    dw 5352
+    dw 2217
+    dw 5352
+    dw 2217
+    dw 5352
+    dw 2217
+    dw 5352
+    dw 2217
+align 16
+_2217_neg5352:
+    dw 2217
+    dw -5352
+    dw 2217
+    dw -5352
+    dw 2217
+    dw -5352
+    dw 2217
+    dw -5352
+align 16
+_mult_add:
+    times 8 dw 1
+align 16
+_cmp_mask:
+    times 4 dw 1
+    times 4 dw 0
+align 16
+_cmp_mask8x4:
+    times 8 dw 1
+align 16
+_mult_sub:
+    dw 1
+    dw -1
+    dw 1
+    dw -1
+    dw 1
+    dw -1
+    dw 1
+    dw -1
+align 16
+_7:
+    times 4 dd 7
+align 16
+_7w:
+    times 8 dw 7
+align 16
+_14500:
+    times 4 dd 14500
+align 16
+_7500:
+    times 4 dd 7500
+align 16
+_12000:
+    times 4 dd 12000
+align 16
+_51000:
+    times 4 dd 51000
diff --git a/libvpx/libvpx/vp8/encoder/x86/denoising_sse2.c b/libvpx/libvpx/vp8/encoder/x86/denoising_sse2.c
new file mode 100644
index 0000000..101d646
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/denoising_sse2.c
@@ -0,0 +1,379 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp8/encoder/denoising.h"
+#include "vp8/common/reconinter.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp8_rtcd.h"
+
+#include <emmintrin.h>
+#include "vpx_ports/emmintrin_compat.h"
+
+/* Compute the sum of all pixel differences of this MB. */
+static INLINE unsigned int abs_sum_diff_16x1(__m128i acc_diff) {
+  const __m128i k_1 = _mm_set1_epi16(1);
+  const __m128i acc_diff_lo = _mm_srai_epi16(
+      _mm_unpacklo_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_hi = _mm_srai_epi16(
+      _mm_unpackhi_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
+  const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
+  const __m128i hgfe_dcba = _mm_add_epi32(hg_fe_dc_ba,
+                                          _mm_srli_si128(hg_fe_dc_ba, 8));
+  const __m128i hgfedcba = _mm_add_epi32(hgfe_dcba,
+                                         _mm_srli_si128(hgfe_dcba, 4));
+  unsigned int sum_diff = abs(_mm_cvtsi128_si32(hgfedcba));
+
+  return sum_diff;
+}
+
+int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
+                             int mc_avg_y_stride,
+                             unsigned char *running_avg_y, int avg_y_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising)
+{
+    unsigned char *running_avg_y_start = running_avg_y;
+    unsigned char *sig_start = sig;
+    unsigned int sum_diff_thresh;
+    int r;
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 1 : 0;
+    __m128i acc_diff = _mm_setzero_si128();
+    const __m128i k_0 = _mm_setzero_si128();
+    const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+    const __m128i k_8 = _mm_set1_epi8(8);
+    const __m128i k_16 = _mm_set1_epi8(16);
+    /* Modify each level's adjustment according to motion_magnitude. */
+    const __m128i l3 = _mm_set1_epi8(
+                       (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ?
+                        7 + shift_inc : 6);
+    /* Difference between level 3 and level 2 is 2. */
+    const __m128i l32 = _mm_set1_epi8(2);
+    /* Difference between level 2 and level 1 is 1. */
+    const __m128i l21 = _mm_set1_epi8(1);
+
+    for (r = 0; r < 16; ++r)
+    {
+        /* Calculate differences */
+        const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
+        const __m128i v_mc_running_avg_y = _mm_loadu_si128(
+                                           (__m128i *)(&mc_running_avg_y[0]));
+        __m128i v_running_avg_y;
+        const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
+        const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
+        /* Obtain the sign. FF if diff is negative. */
+        const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+        /* Clamp absolute difference to 16 to be used to get mask. Doing this
+         * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */
+        const __m128i clamped_absdiff = _mm_min_epu8(
+                                        _mm_or_si128(pdiff, ndiff), k_16);
+        /* Get masks for l2 l1 and l0 adjustments */
+        const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
+        const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
+        const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
+        /* Get adjustments for l2, l1, and l0 */
+        __m128i adj2 = _mm_and_si128(mask2, l32);
+        const __m128i adj1 = _mm_and_si128(mask1, l21);
+        const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
+        __m128i adj,  padj, nadj;
+
+        /* Combine the adjustments and get absolute adjustments. */
+        adj2 = _mm_add_epi8(adj2, adj1);
+        adj = _mm_sub_epi8(l3, adj2);
+        adj = _mm_andnot_si128(mask0, adj);
+        adj = _mm_or_si128(adj, adj0);
+
+        /* Restore the sign and get positive and negative adjustments. */
+        padj = _mm_andnot_si128(diff_sign, adj);
+        nadj = _mm_and_si128(diff_sign, adj);
+
+        /* Calculate filtered value. */
+        v_running_avg_y = _mm_adds_epu8(v_sig, padj);
+        v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
+        _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
+
+        /* Adjustments <=7, and each element in acc_diff can fit in signed
+         * char.
+         */
+        acc_diff = _mm_adds_epi8(acc_diff, padj);
+        acc_diff = _mm_subs_epi8(acc_diff, nadj);
+
+        /* Update pointers for next iteration. */
+        sig += sig_stride;
+        mc_running_avg_y += mc_avg_y_stride;
+        running_avg_y += avg_y_stride;
+    }
+
+    {
+        /* Compute the sum of all pixel differences of this MB. */
+        unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+        sum_diff_thresh = SUM_DIFF_THRESHOLD;
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
+        if (abs_sum_diff > sum_diff_thresh) {
+          // Before returning to copy the block (i.e., apply no denoising),
+          // check if we can still apply some (weaker) temporal filtering to
+          // this block, that would otherwise not be denoised at all. Simplest
+          // is to apply an additional adjustment to running_avg_y to bring it
+          // closer to sig. The adjustment is capped by a maximum delta, and
+          // chosen such that in most cases the resulting sum_diff will be
+          // within the acceptable range given by sum_diff_thresh.
+
+          // The delta is set by the excess of absolute pixel diff over the
+          // threshold.
+          int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
+          // Only apply the adjustment for max delta up to 3.
+          if (delta < 4) {
+            const __m128i k_delta = _mm_set1_epi8(delta);
+            sig -= sig_stride * 16;
+            mc_running_avg_y -= mc_avg_y_stride * 16;
+            running_avg_y -= avg_y_stride * 16;
+            for (r = 0; r < 16; ++r) {
+              __m128i v_running_avg_y =
+                  _mm_loadu_si128((__m128i *)(&running_avg_y[0]));
+              // Calculate differences.
+              const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
+              const __m128i v_mc_running_avg_y =
+                  _mm_loadu_si128((__m128i *)(&mc_running_avg_y[0]));
+              const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
+              const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
+              // Obtain the sign. FF if diff is negative.
+              const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+              // Clamp absolute difference to delta to get the adjustment.
+              const __m128i adj =
+                  _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
+              // Restore the sign and get positive and negative adjustments.
+              __m128i padj, nadj;
+              padj = _mm_andnot_si128(diff_sign, adj);
+              nadj = _mm_and_si128(diff_sign, adj);
+              // Calculate filtered value.
+              v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj);
+              v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj);
+             _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
+
+             // Accumulate the adjustments.
+             acc_diff = _mm_subs_epi8(acc_diff, padj);
+             acc_diff = _mm_adds_epi8(acc_diff, nadj);
+
+             // Update pointers for next iteration.
+             sig += sig_stride;
+             mc_running_avg_y += mc_avg_y_stride;
+             running_avg_y += avg_y_stride;
+            }
+            abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+            if (abs_sum_diff > sum_diff_thresh) {
+              return COPY_BLOCK;
+            }
+          } else {
+            return COPY_BLOCK;
+          }
+        }
+    }
+
+    vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
+    return FILTER_BLOCK;
+}
+
+int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg,
+                             int mc_avg_stride,
+                             unsigned char *running_avg, int avg_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    unsigned char *running_avg_start = running_avg;
+    unsigned char *sig_start = sig;
+    unsigned int sum_diff_thresh;
+    int r;
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ? 1 : 0;
+    __m128i acc_diff = _mm_setzero_si128();
+    const __m128i k_0 = _mm_setzero_si128();
+    const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+    const __m128i k_8 = _mm_set1_epi8(8);
+    const __m128i k_16 = _mm_set1_epi8(16);
+    /* Modify each level's adjustment according to motion_magnitude. */
+    const __m128i l3 = _mm_set1_epi8(
+                       (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ?
+                        7 + shift_inc : 6);
+    /* Difference between level 3 and level 2 is 2. */
+    const __m128i l32 = _mm_set1_epi8(2);
+    /* Difference between level 2 and level 1 is 1. */
+    const __m128i l21 = _mm_set1_epi8(1);
+
+    {
+      const __m128i k_1 = _mm_set1_epi16(1);
+      __m128i vec_sum_block = _mm_setzero_si128();
+
+      // Avoid denoising color signal if its close to average level.
+      for (r = 0; r < 8; ++r) {
+        const __m128i v_sig = _mm_loadl_epi64((__m128i *)(&sig[0]));
+        const __m128i v_sig_unpack = _mm_unpacklo_epi8(v_sig, k_0);
+        vec_sum_block = _mm_add_epi16(vec_sum_block, v_sig_unpack);
+        sig += sig_stride;
+      }
+      sig -= sig_stride * 8;
+      {
+        const __m128i hg_fe_dc_ba = _mm_madd_epi16(vec_sum_block, k_1);
+        const __m128i hgfe_dcba = _mm_add_epi32(hg_fe_dc_ba,
+                                                _mm_srli_si128(hg_fe_dc_ba, 8));
+        const __m128i hgfedcba = _mm_add_epi32(hgfe_dcba,
+                                               _mm_srli_si128(hgfe_dcba, 4));
+        const int sum_block = _mm_cvtsi128_si32(hgfedcba);
+        if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
+          return COPY_BLOCK;
+        }
+      }
+    }
+
+    for (r = 0; r < 4; ++r) {
+        /* Calculate differences */
+        const __m128i v_sig_low = _mm_castpd_si128(
+            _mm_load_sd((double *)(&sig[0])));
+        const __m128i v_sig = _mm_castpd_si128(
+            _mm_loadh_pd(_mm_castsi128_pd(v_sig_low),
+                         (double *)(&sig[sig_stride])));
+        const __m128i v_mc_running_avg_low = _mm_castpd_si128(
+            _mm_load_sd((double *)(&mc_running_avg[0])));
+        const __m128i v_mc_running_avg = _mm_castpd_si128(
+            _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
+                         (double *)(&mc_running_avg[mc_avg_stride])));
+        const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
+        const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
+        /* Obtain the sign. FF if diff is negative. */
+        const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+        /* Clamp absolute difference to 16 to be used to get mask. Doing this
+         * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */
+        const __m128i clamped_absdiff = _mm_min_epu8(
+                                        _mm_or_si128(pdiff, ndiff), k_16);
+        /* Get masks for l2 l1 and l0 adjustments */
+        const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
+        const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
+        const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
+        /* Get adjustments for l2, l1, and l0 */
+        __m128i adj2 = _mm_and_si128(mask2, l32);
+        const __m128i adj1 = _mm_and_si128(mask1, l21);
+        const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
+        __m128i adj,  padj, nadj;
+        __m128i v_running_avg;
+
+        /* Combine the adjustments and get absolute adjustments. */
+        adj2 = _mm_add_epi8(adj2, adj1);
+        adj = _mm_sub_epi8(l3, adj2);
+        adj = _mm_andnot_si128(mask0, adj);
+        adj = _mm_or_si128(adj, adj0);
+
+        /* Restore the sign and get positive and negative adjustments. */
+        padj = _mm_andnot_si128(diff_sign, adj);
+        nadj = _mm_and_si128(diff_sign, adj);
+
+        /* Calculate filtered value. */
+        v_running_avg = _mm_adds_epu8(v_sig, padj);
+        v_running_avg = _mm_subs_epu8(v_running_avg, nadj);
+
+        _mm_storel_pd((double *)&running_avg[0],
+                      _mm_castsi128_pd(v_running_avg));
+        _mm_storeh_pd((double *)&running_avg[avg_stride],
+                      _mm_castsi128_pd(v_running_avg));
+
+        /* Adjustments <=7, and each element in acc_diff can fit in signed
+         * char.
+         */
+        acc_diff = _mm_adds_epi8(acc_diff, padj);
+        acc_diff = _mm_subs_epi8(acc_diff, nadj);
+
+        /* Update pointers for next iteration. */
+        sig += sig_stride * 2;
+        mc_running_avg += mc_avg_stride * 2;
+        running_avg += avg_stride * 2;
+    }
+
+    {
+        unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+        sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+        if (abs_sum_diff > sum_diff_thresh) {
+          // Before returning to copy the block (i.e., apply no denoising),
+          // check if we can still apply some (weaker) temporal filtering to
+          // this block, that would otherwise not be denoised at all. Simplest
+          // is to apply an additional adjustment to running_avg_y to bring it
+          // closer to sig. The adjustment is capped by a maximum delta, and
+          // chosen such that in most cases the resulting sum_diff will be
+          // within the acceptable range given by sum_diff_thresh.
+
+          // The delta is set by the excess of absolute pixel diff over the
+          // threshold.
+          int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
+          // Only apply the adjustment for max delta up to 3.
+          if (delta < 4) {
+            const __m128i k_delta = _mm_set1_epi8(delta);
+            sig -= sig_stride * 8;
+            mc_running_avg -= mc_avg_stride * 8;
+            running_avg -= avg_stride * 8;
+            for (r = 0; r < 4; ++r) {
+              // Calculate differences.
+              const __m128i v_sig_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&sig[0])));
+              const __m128i v_sig = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_sig_low),
+                               (double *)(&sig[sig_stride])));
+              const __m128i v_mc_running_avg_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&mc_running_avg[0])));
+              const __m128i v_mc_running_avg = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
+                               (double *)(&mc_running_avg[mc_avg_stride])));
+              const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
+              const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
+              // Obtain the sign. FF if diff is negative.
+              const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+              // Clamp absolute difference to delta to get the adjustment.
+              const __m128i adj =
+                  _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
+              // Restore the sign and get positive and negative adjustments.
+              __m128i padj, nadj;
+              const __m128i v_running_avg_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&running_avg[0])));
+              __m128i v_running_avg = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_running_avg_low),
+                               (double *)(&running_avg[avg_stride])));
+              padj = _mm_andnot_si128(diff_sign, adj);
+              nadj = _mm_and_si128(diff_sign, adj);
+              // Calculate filtered value.
+              v_running_avg = _mm_subs_epu8(v_running_avg, padj);
+              v_running_avg = _mm_adds_epu8(v_running_avg, nadj);
+
+              _mm_storel_pd((double *)&running_avg[0],
+                            _mm_castsi128_pd(v_running_avg));
+              _mm_storeh_pd((double *)&running_avg[avg_stride],
+                            _mm_castsi128_pd(v_running_avg));
+
+             // Accumulate the adjustments.
+             acc_diff = _mm_subs_epi8(acc_diff, padj);
+             acc_diff = _mm_adds_epi8(acc_diff, nadj);
+
+             // Update pointers for next iteration.
+             sig += sig_stride * 2;
+             mc_running_avg += mc_avg_stride * 2;
+             running_avg += avg_stride * 2;
+            }
+            abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+            if (abs_sum_diff > sum_diff_thresh) {
+              return COPY_BLOCK;
+            }
+          } else {
+            return COPY_BLOCK;
+          }
+        }
+    }
+
+    vp8_copy_mem8x8(running_avg_start, avg_stride, sig_start, sig_stride);
+    return FILTER_BLOCK;
+}
diff --git a/libvpx/libvpx/vp8/encoder/x86/encodeopt.asm b/libvpx/libvpx/vp8/encoder/x86/encodeopt.asm
new file mode 100644
index 0000000..fe26b18
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/encodeopt.asm
@@ -0,0 +1,386 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;int vp8_block_error_xmm(short *coeff_ptr,  short *dcoef_ptr)
+global sym(vp8_block_error_xmm) PRIVATE
+sym(vp8_block_error_xmm):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 2
+    push rsi
+    push rdi
+    ; end prologue
+
+        mov         rsi,        arg(0) ;coeff_ptr
+        mov         rdi,        arg(1) ;dcoef_ptr
+
+        movdqa      xmm0,       [rsi]
+        movdqa      xmm1,       [rdi]
+
+        movdqa      xmm2,       [rsi+16]
+        movdqa      xmm3,       [rdi+16]
+
+        psubw       xmm0,       xmm1
+        psubw       xmm2,       xmm3
+
+        pmaddwd     xmm0,       xmm0
+        pmaddwd     xmm2,       xmm2
+
+        paddd       xmm0,       xmm2
+
+        pxor        xmm5,       xmm5
+        movdqa      xmm1,       xmm0
+
+        punpckldq   xmm0,       xmm5
+        punpckhdq   xmm1,       xmm5
+
+        paddd       xmm0,       xmm1
+        movdqa      xmm1,       xmm0
+
+        psrldq      xmm0,       8
+        paddd       xmm0,       xmm1
+
+        movq        rax,        xmm0
+
+    pop rdi
+    pop rsi
+    ; begin epilog
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;int vp8_block_error_mmx(short *coeff_ptr,  short *dcoef_ptr)
+global sym(vp8_block_error_mmx) PRIVATE
+sym(vp8_block_error_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 2
+    push rsi
+    push rdi
+    ; end prolog
+
+
+        mov         rsi,        arg(0) ;coeff_ptr
+        pxor        mm7,        mm7
+
+        mov         rdi,        arg(1) ;dcoef_ptr
+        movq        mm3,        [rsi]
+
+        movq        mm4,        [rdi]
+        movq        mm5,        [rsi+8]
+
+        movq        mm6,        [rdi+8]
+        pxor        mm1,        mm1 ; from movd mm1, dc ; dc =0
+
+        movq        mm2,        mm7
+        psubw       mm5,        mm6
+
+        por         mm1,        mm2
+        pmaddwd     mm5,        mm5
+
+        pcmpeqw     mm1,        mm7
+        psubw       mm3,        mm4
+
+        pand        mm1,        mm3
+        pmaddwd     mm1,        mm1
+
+        paddd       mm1,        mm5
+        movq        mm3,        [rsi+16]
+
+        movq        mm4,        [rdi+16]
+        movq        mm5,        [rsi+24]
+
+        movq        mm6,        [rdi+24]
+        psubw       mm5,        mm6
+
+        pmaddwd     mm5,        mm5
+        psubw       mm3,        mm4
+
+        pmaddwd     mm3,        mm3
+        paddd       mm3,        mm5
+
+        paddd       mm1,        mm3
+        movq        mm0,        mm1
+
+        psrlq       mm1,        32
+        paddd       mm0,        mm1
+
+        movq        rax,        mm0
+
+    pop rdi
+    pop rsi
+    ; begin epilog
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;int vp8_mbblock_error_mmx_impl(short *coeff_ptr, short *dcoef_ptr, int dc);
+global sym(vp8_mbblock_error_mmx_impl) PRIVATE
+sym(vp8_mbblock_error_mmx_impl):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 3
+    push rsi
+    push rdi
+    ; end prolog
+
+
+        mov         rsi,        arg(0) ;coeff_ptr
+        pxor        mm7,        mm7
+
+        mov         rdi,        arg(1) ;dcoef_ptr
+        pxor        mm2,        mm2
+
+        movd        mm1,        dword ptr arg(2) ;dc
+        por         mm1,        mm2
+
+        pcmpeqw     mm1,        mm7
+        mov         rcx,        16
+
+.mberror_loop_mmx:
+        movq        mm3,       [rsi]
+        movq        mm4,       [rdi]
+
+        movq        mm5,       [rsi+8]
+        movq        mm6,       [rdi+8]
+
+
+        psubw       mm5,        mm6
+        pmaddwd     mm5,        mm5
+
+        psubw       mm3,        mm4
+        pand        mm3,        mm1
+
+        pmaddwd     mm3,        mm3
+        paddd       mm2,        mm5
+
+        paddd       mm2,        mm3
+        movq        mm3,       [rsi+16]
+
+        movq        mm4,       [rdi+16]
+        movq        mm5,       [rsi+24]
+
+        movq        mm6,       [rdi+24]
+        psubw       mm5,        mm6
+
+        pmaddwd     mm5,        mm5
+        psubw       mm3,        mm4
+
+        pmaddwd     mm3,        mm3
+        paddd       mm2,        mm5
+
+        paddd       mm2,        mm3
+        add         rsi,        32
+
+        add         rdi,        32
+        sub         rcx,        1
+
+        jnz         .mberror_loop_mmx
+
+        movq        mm0,        mm2
+        psrlq       mm2,        32
+
+        paddd       mm0,        mm2
+        movq        rax,        mm0
+
+    pop rdi
+    pop rsi
+    ; begin epilog
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;int vp8_mbblock_error_xmm_impl(short *coeff_ptr, short *dcoef_ptr, int dc);
+global sym(vp8_mbblock_error_xmm_impl) PRIVATE
+sym(vp8_mbblock_error_xmm_impl):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 3
+    SAVE_XMM 6
+    push rsi
+    push rdi
+    ; end prolog
+
+
+        mov         rsi,        arg(0) ;coeff_ptr
+        pxor        xmm6,       xmm6
+
+        mov         rdi,        arg(1) ;dcoef_ptr
+        pxor        xmm4,       xmm4
+
+        movd        xmm5,       dword ptr arg(2) ;dc
+        por         xmm5,       xmm4
+
+        pcmpeqw     xmm5,       xmm6
+        mov         rcx,        16
+
+.mberror_loop:
+        movdqa      xmm0,       [rsi]
+        movdqa      xmm1,       [rdi]
+
+        movdqa      xmm2,       [rsi+16]
+        movdqa      xmm3,       [rdi+16]
+
+
+        psubw       xmm2,       xmm3
+        pmaddwd     xmm2,       xmm2
+
+        psubw       xmm0,       xmm1
+        pand        xmm0,       xmm5
+
+        pmaddwd     xmm0,       xmm0
+        add         rsi,        32
+
+        add         rdi,        32
+
+        sub         rcx,        1
+        paddd       xmm4,       xmm2
+
+        paddd       xmm4,       xmm0
+        jnz         .mberror_loop
+
+        movdqa      xmm0,       xmm4
+        punpckldq   xmm0,       xmm6
+
+        punpckhdq   xmm4,       xmm6
+        paddd       xmm0,       xmm4
+
+        movdqa      xmm1,       xmm0
+        psrldq      xmm0,       8
+
+        paddd       xmm0,       xmm1
+        movq        rax,        xmm0
+
+    pop rdi
+    pop rsi
+    ; begin epilog
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;int vp8_mbuverror_mmx_impl(short *s_ptr, short *d_ptr);
+global sym(vp8_mbuverror_mmx_impl) PRIVATE
+sym(vp8_mbuverror_mmx_impl):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 2
+    push rsi
+    push rdi
+    ; end prolog
+
+
+        mov             rsi,        arg(0) ;s_ptr
+        mov             rdi,        arg(1) ;d_ptr
+
+        mov             rcx,        16
+        pxor            mm7,        mm7
+
+.mbuverror_loop_mmx:
+
+        movq            mm1,        [rsi]
+        movq            mm2,        [rdi]
+
+        psubw           mm1,        mm2
+        pmaddwd         mm1,        mm1
+
+
+        movq            mm3,        [rsi+8]
+        movq            mm4,        [rdi+8]
+
+        psubw           mm3,        mm4
+        pmaddwd         mm3,        mm3
+
+
+        paddd           mm7,        mm1
+        paddd           mm7,        mm3
+
+
+        add             rsi,        16
+        add             rdi,        16
+
+        dec             rcx
+        jnz             .mbuverror_loop_mmx
+
+        movq            mm0,        mm7
+        psrlq           mm7,        32
+
+        paddd           mm0,        mm7
+        movq            rax,        mm0
+
+    pop rdi
+    pop rsi
+    ; begin epilog
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;int vp8_mbuverror_xmm_impl(short *s_ptr, short *d_ptr);
+global sym(vp8_mbuverror_xmm_impl) PRIVATE
+sym(vp8_mbuverror_xmm_impl):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 2
+    push rsi
+    push rdi
+    ; end prolog
+
+
+        mov             rsi,        arg(0) ;s_ptr
+        mov             rdi,        arg(1) ;d_ptr
+
+        mov             rcx,        16
+        pxor            xmm3,       xmm3
+
+.mbuverror_loop:
+
+        movdqa          xmm1,       [rsi]
+        movdqa          xmm2,       [rdi]
+
+        psubw           xmm1,       xmm2
+        pmaddwd         xmm1,       xmm1
+
+        paddd           xmm3,       xmm1
+
+        add             rsi,        16
+        add             rdi,        16
+
+        dec             rcx
+        jnz             .mbuverror_loop
+
+        pxor        xmm0,           xmm0
+        movdqa      xmm1,           xmm3
+
+        movdqa      xmm2,           xmm1
+        punpckldq   xmm1,           xmm0
+
+        punpckhdq   xmm2,           xmm0
+        paddd       xmm1,           xmm2
+
+        movdqa      xmm2,           xmm1
+
+        psrldq      xmm1,           8
+        paddd       xmm1,           xmm2
+
+        movq            rax,            xmm1
+
+    pop rdi
+    pop rsi
+    ; begin epilog
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vp8/encoder/x86/fwalsh_sse2.asm b/libvpx/libvpx/vp8/encoder/x86/fwalsh_sse2.asm
new file mode 100644
index 0000000..f498927
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/fwalsh_sse2.asm
@@ -0,0 +1,164 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch)
+global sym(vp8_short_walsh4x4_sse2) PRIVATE
+sym(vp8_short_walsh4x4_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 3
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov     rsi, arg(0)           ; input
+    mov     rdi, arg(1)           ; output
+    movsxd  rdx, dword ptr arg(2) ; pitch
+
+    ; first for loop
+    movq    xmm0, MMWORD PTR [rsi]           ; load input
+    movq    xmm1, MMWORD PTR [rsi + rdx]
+    lea     rsi,  [rsi + rdx*2]
+    movq    xmm2, MMWORD PTR [rsi]
+    movq    xmm3, MMWORD PTR [rsi + rdx]
+
+    punpcklwd xmm0,  xmm1
+    punpcklwd xmm2,  xmm3
+
+    movdqa    xmm1, xmm0
+    punpckldq xmm0, xmm2           ; ip[1] ip[0]
+    punpckhdq xmm1, xmm2           ; ip[3] ip[2]
+
+    movdqa    xmm2, xmm0
+    paddw     xmm0, xmm1
+    psubw     xmm2, xmm1
+
+    psllw     xmm0, 2              ; d1  a1
+    psllw     xmm2, 2              ; c1  b1
+
+    movdqa    xmm1, xmm0
+    punpcklqdq xmm0, xmm2          ; b1  a1
+    punpckhqdq xmm1, xmm2          ; c1  d1
+
+    pxor      xmm6, xmm6
+    movq      xmm6, xmm0
+    pxor      xmm7, xmm7
+    pcmpeqw   xmm7, xmm6
+    paddw     xmm7, [GLOBAL(c1)]
+
+    movdqa    xmm2, xmm0
+    paddw     xmm0, xmm1           ; b1+c1  a1+d1
+    psubw     xmm2, xmm1           ; b1-c1  a1-d1
+    paddw     xmm0, xmm7           ; b1+c1  a1+d1+(a1!=0)
+
+    ; second for loop
+    ; input: 13  9  5  1 12  8  4  0 (xmm0)
+    ;        14 10  6  2 15 11  7  3 (xmm2)
+    ; after shuffle:
+    ;        13  5  9  1 12  4  8  0 (xmm0)
+    ;        14  6 10  2 15  7 11  3 (xmm1)
+    pshuflw   xmm3, xmm0, 0xd8
+    pshufhw   xmm0, xmm3, 0xd8
+    pshuflw   xmm3, xmm2, 0xd8
+    pshufhw   xmm1, xmm3, 0xd8
+
+    movdqa    xmm2, xmm0
+    pmaddwd   xmm0, [GLOBAL(c1)]    ; d11 a11 d10 a10
+    pmaddwd   xmm2, [GLOBAL(cn1)]   ; c11 b11 c10 b10
+    movdqa    xmm3, xmm1
+    pmaddwd   xmm1, [GLOBAL(c1)]    ; d12 a12 d13 a13
+    pmaddwd   xmm3, [GLOBAL(cn1)]   ; c12 b12 c13 b13
+
+    pshufd    xmm4, xmm0, 0xd8      ; d11 d10 a11 a10
+    pshufd    xmm5, xmm2, 0xd8      ; c11 c10 b11 b10
+    pshufd    xmm6, xmm1, 0x72      ; d13 d12 a13 a12
+    pshufd    xmm7, xmm3, 0x72      ; c13 c12 b13 b12
+
+    movdqa    xmm0, xmm4
+    punpcklqdq xmm0, xmm5           ; b11 b10 a11 a10
+    punpckhqdq xmm4, xmm5           ; c11 c10 d11 d10
+    movdqa    xmm1, xmm6
+    punpcklqdq xmm1, xmm7           ; b13 b12 a13 a12
+    punpckhqdq xmm6, xmm7           ; c13 c12 d13 d12
+
+    movdqa    xmm2, xmm0
+    paddd     xmm0, xmm4            ; b21 b20 a21 a20
+    psubd     xmm2, xmm4            ; c21 c20 d21 d20
+    movdqa    xmm3, xmm1
+    paddd     xmm1, xmm6            ; b23 b22 a23 a22
+    psubd     xmm3, xmm6            ; c23 c22 d23 d22
+
+    pxor      xmm4, xmm4
+    movdqa    xmm5, xmm4
+    pcmpgtd   xmm4, xmm0
+    pcmpgtd   xmm5, xmm2
+    pand      xmm4, [GLOBAL(cd1)]
+    pand      xmm5, [GLOBAL(cd1)]
+
+    pxor      xmm6, xmm6
+    movdqa    xmm7, xmm6
+    pcmpgtd   xmm6, xmm1
+    pcmpgtd   xmm7, xmm3
+    pand      xmm6, [GLOBAL(cd1)]
+    pand      xmm7, [GLOBAL(cd1)]
+
+    paddd     xmm0, xmm4
+    paddd     xmm2, xmm5
+    paddd     xmm0, [GLOBAL(cd3)]
+    paddd     xmm2, [GLOBAL(cd3)]
+    paddd     xmm1, xmm6
+    paddd     xmm3, xmm7
+    paddd     xmm1, [GLOBAL(cd3)]
+    paddd     xmm3, [GLOBAL(cd3)]
+
+    psrad     xmm0, 3
+    psrad     xmm1, 3
+    psrad     xmm2, 3
+    psrad     xmm3, 3
+    movdqa    xmm4, xmm0
+    punpcklqdq xmm0, xmm1           ; a23 a22 a21 a20
+    punpckhqdq xmm4, xmm1           ; b23 b22 b21 b20
+    movdqa    xmm5, xmm2
+    punpckhqdq xmm2, xmm3           ; c23 c22 c21 c20
+    punpcklqdq xmm5, xmm3           ; d23 d22 d21 d20
+
+    packssdw  xmm0, xmm4            ; b23 b22 b21 b20 a23 a22 a21 a20
+    packssdw  xmm2, xmm5            ; d23 d22 d21 d20 c23 c22 c21 c20
+
+    movdqa  XMMWORD PTR [rdi], xmm0
+    movdqa  XMMWORD PTR [rdi + 16], xmm2
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+c1:
+    dw 0x0001, 0x0001, 0x0001, 0x0001, 0x0001, 0x0001, 0x0001, 0x0001
+align 16
+cn1:
+    dw 0x0001, 0xffff, 0x0001, 0xffff, 0x0001, 0xffff, 0x0001, 0xffff
+align 16
+cd1:
+    dd 0x00000001, 0x00000001, 0x00000001, 0x00000001
+align 16
+cd3:
+    dd 0x00000003, 0x00000003, 0x00000003, 0x00000003
diff --git a/libvpx/libvpx/vp8/encoder/x86/quantize_mmx.asm b/libvpx/libvpx/vp8/encoder/x86/quantize_mmx.asm
new file mode 100644
index 0000000..2864ce1
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/quantize_mmx.asm
@@ -0,0 +1,286 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;int vp8_fast_quantize_b_impl_mmx(short *coeff_ptr, short *zbin_ptr,
+;                           short *qcoeff_ptr,short *dequant_ptr,
+;                           short *scan_mask, short *round_ptr,
+;                           short *quant_ptr, short *dqcoeff_ptr);
+global sym(vp8_fast_quantize_b_impl_mmx) PRIVATE
+sym(vp8_fast_quantize_b_impl_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 8
+    push rsi
+    push rdi
+    ; end prolog
+
+
+        mov             rsi,        arg(0) ;coeff_ptr
+        movq            mm0,        [rsi]
+
+        mov             rax,        arg(1) ;zbin_ptr
+        movq            mm1,        [rax]
+
+        movq            mm3,        mm0
+        psraw           mm0,        15
+
+        pxor            mm3,        mm0
+        psubw           mm3,        mm0         ; abs
+
+        movq            mm2,        mm3
+        pcmpgtw         mm1,        mm2
+
+        pandn           mm1,        mm2
+        movq            mm3,        mm1
+
+        mov             rdx,        arg(6) ;quant_ptr
+        movq            mm1,        [rdx]
+
+        mov             rcx,        arg(5) ;round_ptr
+        movq            mm2,        [rcx]
+
+        paddw           mm3,        mm2
+        pmulhuw         mm3,        mm1
+
+        pxor            mm3,        mm0
+        psubw           mm3,        mm0     ;gain the sign back
+
+        mov             rdi,        arg(2) ;qcoeff_ptr
+        movq            mm0,        mm3
+
+        movq            [rdi],      mm3
+
+        mov             rax,        arg(3) ;dequant_ptr
+        movq            mm2,        [rax]
+
+        pmullw          mm3,        mm2
+        mov             rax,        arg(7) ;dqcoeff_ptr
+
+        movq            [rax],      mm3
+
+        ; next 8
+        movq            mm4,        [rsi+8]
+
+        mov             rax,        arg(1) ;zbin_ptr
+        movq            mm5,        [rax+8]
+
+        movq            mm7,        mm4
+        psraw           mm4,        15
+
+        pxor            mm7,        mm4
+        psubw           mm7,        mm4         ; abs
+
+        movq            mm6,        mm7
+        pcmpgtw         mm5,        mm6
+
+        pandn           mm5,        mm6
+        movq            mm7,        mm5
+
+        movq            mm5,        [rdx+8]
+        movq            mm6,        [rcx+8]
+
+        paddw           mm7,        mm6
+        pmulhuw         mm7,        mm5
+
+        pxor            mm7,        mm4
+        psubw           mm7,        mm4;gain the sign back
+
+        mov             rdi,        arg(2) ;qcoeff_ptr
+
+        movq            mm1,        mm7
+        movq            [rdi+8],    mm7
+
+        mov             rax,        arg(3) ;dequant_ptr
+        movq            mm6,        [rax+8]
+
+        pmullw          mm7,        mm6
+        mov             rax,        arg(7) ;dqcoeff_ptr
+
+        movq            [rax+8],    mm7
+
+
+                ; next 8
+        movq            mm4,        [rsi+16]
+
+        mov             rax,        arg(1) ;zbin_ptr
+        movq            mm5,        [rax+16]
+
+        movq            mm7,        mm4
+        psraw           mm4,        15
+
+        pxor            mm7,        mm4
+        psubw           mm7,        mm4         ; abs
+
+        movq            mm6,        mm7
+        pcmpgtw         mm5,        mm6
+
+        pandn           mm5,        mm6
+        movq            mm7,        mm5
+
+        movq            mm5,        [rdx+16]
+        movq            mm6,        [rcx+16]
+
+        paddw           mm7,        mm6
+        pmulhuw         mm7,        mm5
+
+        pxor            mm7,        mm4
+        psubw           mm7,        mm4;gain the sign back
+
+        mov             rdi,        arg(2) ;qcoeff_ptr
+
+        movq            mm1,        mm7
+        movq            [rdi+16],   mm7
+
+        mov             rax,        arg(3) ;dequant_ptr
+        movq            mm6,        [rax+16]
+
+        pmullw          mm7,        mm6
+        mov             rax,        arg(7) ;dqcoeff_ptr
+
+        movq            [rax+16],   mm7
+
+
+                ; next 8
+        movq            mm4,        [rsi+24]
+
+        mov             rax,        arg(1) ;zbin_ptr
+        movq            mm5,        [rax+24]
+
+        movq            mm7,        mm4
+        psraw           mm4,        15
+
+        pxor            mm7,        mm4
+        psubw           mm7,        mm4         ; abs
+
+        movq            mm6,        mm7
+        pcmpgtw         mm5,        mm6
+
+        pandn           mm5,        mm6
+        movq            mm7,        mm5
+
+        movq            mm5,        [rdx+24]
+        movq            mm6,        [rcx+24]
+
+        paddw           mm7,        mm6
+        pmulhuw         mm7,        mm5
+
+        pxor            mm7,        mm4
+        psubw           mm7,        mm4;gain the sign back
+
+        mov             rdi,        arg(2) ;qcoeff_ptr
+
+        movq            mm1,        mm7
+        movq            [rdi+24],   mm7
+
+        mov             rax,        arg(3) ;dequant_ptr
+        movq            mm6,        [rax+24]
+
+        pmullw          mm7,        mm6
+        mov             rax,        arg(7) ;dqcoeff_ptr
+
+        movq            [rax+24],   mm7
+
+
+
+        mov             rdi,        arg(4) ;scan_mask
+        mov             rsi,        arg(2) ;qcoeff_ptr
+
+        pxor            mm5,        mm5
+        pxor            mm7,        mm7
+
+        movq            mm0,        [rsi]
+        movq            mm1,        [rsi+8]
+
+        movq            mm2,        [rdi]
+        movq            mm3,        [rdi+8];
+
+        pcmpeqw         mm0,        mm7
+        pcmpeqw         mm1,        mm7
+
+        pcmpeqw         mm6,        mm6
+        pxor            mm0,        mm6
+
+        pxor            mm1,        mm6
+        psrlw           mm0,        15
+
+        psrlw           mm1,        15
+        pmaddwd         mm0,        mm2
+
+        pmaddwd         mm1,        mm3
+        movq            mm5,        mm0
+
+        paddd           mm5,        mm1
+
+        movq            mm0,        [rsi+16]
+        movq            mm1,        [rsi+24]
+
+        movq            mm2,        [rdi+16]
+        movq            mm3,        [rdi+24];
+
+        pcmpeqw         mm0,        mm7
+        pcmpeqw         mm1,        mm7
+
+        pcmpeqw         mm6,        mm6
+        pxor            mm0,        mm6
+
+        pxor            mm1,        mm6
+        psrlw           mm0,        15
+
+        psrlw           mm1,        15
+        pmaddwd         mm0,        mm2
+
+        pmaddwd         mm1,        mm3
+        paddd           mm5,        mm0
+
+        paddd           mm5,        mm1
+        movq            mm0,        mm5
+
+        psrlq           mm5,        32
+        paddd           mm0,        mm5
+
+        ; eob adjustment begins here
+        movq            rcx,        mm0
+        and             rcx,        0xffff
+
+        xor             rdx,        rdx
+        sub             rdx,        rcx ; rdx=-rcx
+
+        bsr             rax,        rcx
+        inc             rax
+
+        sar             rdx,        31
+        and             rax,        rdx
+        ; Substitute the sse assembly for the old mmx mixed assembly/C. The
+        ; following is kept as reference
+        ;    movq            rcx,        mm0
+        ;    bsr             rax,        rcx
+        ;
+        ;    mov             eob,        rax
+        ;    mov             eee,        rcx
+        ;
+        ;if(eee==0)
+        ;{
+        ;    eob=-1;
+        ;}
+        ;else if(eee<0)
+        ;{
+        ;    eob=15;
+        ;}
+        ;d->eob = eob+1;
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vp8/encoder/x86/quantize_sse4.c b/libvpx/libvpx/vp8/encoder/x86/quantize_sse4.c
new file mode 100644
index 0000000..601dd23
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/quantize_sse4.c
@@ -0,0 +1,128 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <smmintrin.h> /* SSE4.1 */
+
+#include "./vp8_rtcd.h"
+#include "vp8/encoder/block.h"
+#include "vp8/common/entropy.h" /* vp8_default_inv_zig_zag */
+
+#define SELECT_EOB(i, z, x, y, q) \
+    do { \
+        short boost = *zbin_boost_ptr; \
+        short x_z = _mm_extract_epi16(x, z); \
+        short y_z = _mm_extract_epi16(y, z); \
+        int cmp = (x_z < boost) | (y_z == 0); \
+        zbin_boost_ptr++; \
+        if (cmp) \
+            break; \
+        q = _mm_insert_epi16(q, y_z, z); \
+        eob = i; \
+        zbin_boost_ptr = b->zrun_zbin_boost; \
+    } while (0)
+
+void vp8_regular_quantize_b_sse4_1(BLOCK *b, BLOCKD *d) {
+    char eob = 0;
+    short *zbin_boost_ptr  = b->zrun_zbin_boost;
+
+    __m128i sz0, x0, sz1, x1, y0, y1, x_minus_zbin0, x_minus_zbin1,
+            dqcoeff0, dqcoeff1;
+    __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift));
+    __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8));
+    __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
+    __m128i z1 = _mm_load_si128((__m128i *)(b->coeff+8));
+    __m128i zbin_extra = _mm_cvtsi32_si128(b->zbin_extra);
+    __m128i zbin0 = _mm_load_si128((__m128i *)(b->zbin));
+    __m128i zbin1 = _mm_load_si128((__m128i *)(b->zbin + 8));
+    __m128i round0 = _mm_load_si128((__m128i *)(b->round));
+    __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
+    __m128i quant0 = _mm_load_si128((__m128i *)(b->quant));
+    __m128i quant1 = _mm_load_si128((__m128i *)(b->quant + 8));
+    __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
+    __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
+    __m128i qcoeff0 = _mm_setzero_si128();
+    __m128i qcoeff1 = _mm_setzero_si128();
+
+    /* Duplicate to all lanes. */
+    zbin_extra = _mm_shufflelo_epi16(zbin_extra, 0);
+    zbin_extra = _mm_unpacklo_epi16(zbin_extra, zbin_extra);
+
+    /* Sign of z: z >> 15 */
+    sz0 = _mm_srai_epi16(z0, 15);
+    sz1 = _mm_srai_epi16(z1, 15);
+
+    /* x = abs(z): (z ^ sz) - sz */
+    x0 = _mm_xor_si128(z0, sz0);
+    x1 = _mm_xor_si128(z1, sz1);
+    x0 = _mm_sub_epi16(x0, sz0);
+    x1 = _mm_sub_epi16(x1, sz1);
+
+    /* zbin[] + zbin_extra */
+    zbin0 = _mm_add_epi16(zbin0, zbin_extra);
+    zbin1 = _mm_add_epi16(zbin1, zbin_extra);
+
+    /* In C x is compared to zbin where zbin = zbin[] + boost + extra. Rebalance
+     * the equation because boost is the only value which can change:
+     * x - (zbin[] + extra) >= boost */
+    x_minus_zbin0 = _mm_sub_epi16(x0, zbin0);
+    x_minus_zbin1 = _mm_sub_epi16(x1, zbin1);
+
+    /* All the remaining calculations are valid whether they are done now with
+     * simd or later inside the loop one at a time. */
+    x0 = _mm_add_epi16(x0, round0);
+    x1 = _mm_add_epi16(x1, round1);
+
+    y0 = _mm_mulhi_epi16(x0, quant0);
+    y1 = _mm_mulhi_epi16(x1, quant1);
+
+    y0 = _mm_add_epi16(y0, x0);
+    y1 = _mm_add_epi16(y1, x1);
+
+    /* Instead of shifting each value independently we convert the scaling
+     * factor with 1 << (16 - shift) so we can use multiply/return high half. */
+    y0 = _mm_mulhi_epi16(y0, quant_shift0);
+    y1 = _mm_mulhi_epi16(y1, quant_shift1);
+
+    /* Return the sign: (y ^ sz) - sz */
+    y0 = _mm_xor_si128(y0, sz0);
+    y1 = _mm_xor_si128(y1, sz1);
+    y0 = _mm_sub_epi16(y0, sz0);
+    y1 = _mm_sub_epi16(y1, sz1);
+
+    /* The loop gets unrolled anyway. Avoid the vp8_default_zig_zag1d lookup. */
+    SELECT_EOB(1, 0, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(2, 1, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(3, 4, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(4, 0, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(5, 5, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(6, 2, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(7, 3, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(8, 6, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(9, 1, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(10, 4, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(11, 5, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(12, 2, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(13, 7, x_minus_zbin0, y0, qcoeff0);
+    SELECT_EOB(14, 3, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(15, 6, x_minus_zbin1, y1, qcoeff1);
+    SELECT_EOB(16, 7, x_minus_zbin1, y1, qcoeff1);
+
+    _mm_store_si128((__m128i *)(d->qcoeff), qcoeff0);
+    _mm_store_si128((__m128i *)(d->qcoeff + 8), qcoeff1);
+
+    dqcoeff0 = _mm_mullo_epi16(qcoeff0, dequant0);
+    dqcoeff1 = _mm_mullo_epi16(qcoeff1, dequant1);
+
+    _mm_store_si128((__m128i *)(d->dqcoeff), dqcoeff0);
+    _mm_store_si128((__m128i *)(d->dqcoeff + 8), dqcoeff1);
+
+    *d->eob = eob;
+}
diff --git a/libvpx/libvpx/vp8/encoder/x86/quantize_ssse3.c b/libvpx/libvpx/vp8/encoder/x86/quantize_ssse3.c
new file mode 100644
index 0000000..14282db
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/quantize_ssse3.c
@@ -0,0 +1,114 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <tmmintrin.h> /* SSSE3 */
+
+#include "vp8/encoder/block.h"
+
+/* bitscan reverse (bsr) */
+#if defined(_MSC_VER)
+#include <intrin.h>
+#pragma intrinsic(_BitScanReverse)
+static int bsr(int mask) {
+  unsigned long eob;
+  _BitScanReverse(&eob, mask);
+  eob++;
+  if (mask == 0)
+    eob = 0;
+  return eob;
+}
+#else
+static int bsr(int mask) {
+  int eob;
+#if defined(__GNUC__) && __GNUC__
+  __asm__ __volatile__("bsr %1, %0" : "=r" (eob) : "r" (mask) : "flags");
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+  asm volatile("bsr %1, %0" : "=r" (eob) : "r" (mask) : "flags");
+#endif
+  eob++;
+  if (mask == 0)
+    eob = 0;
+  return eob;
+}
+#endif
+
+void vp8_fast_quantize_b_ssse3(BLOCK *b, BLOCKD *d) {
+  int eob, mask;
+
+  __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
+  __m128i z1 = _mm_load_si128((__m128i *)(b->coeff + 8));
+  __m128i round0 = _mm_load_si128((__m128i *)(b->round));
+  __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
+  __m128i quant_fast0 = _mm_load_si128((__m128i *)(b->quant_fast));
+  __m128i quant_fast1 = _mm_load_si128((__m128i *)(b->quant_fast + 8));
+  __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
+  __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
+
+  __m128i sz0, sz1, x, x0, x1, y0, y1, zeros, abs0, abs1;
+
+  DECLARE_ALIGNED(16, const uint8_t, pshufb_zig_zag_mask[16]) =
+    { 0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15 };
+  __m128i zig_zag = _mm_load_si128((const __m128i *)pshufb_zig_zag_mask);
+
+  /* sign of z: z >> 15 */
+  sz0 = _mm_srai_epi16(z0, 15);
+  sz1 = _mm_srai_epi16(z1, 15);
+
+  /* x = abs(z) */
+  x0 = _mm_abs_epi16(z0);
+  x1 = _mm_abs_epi16(z1);
+
+  /* x += round */
+  x0 = _mm_add_epi16(x0, round0);
+  x1 = _mm_add_epi16(x1, round1);
+
+  /* y = (x * quant) >> 16 */
+  y0 = _mm_mulhi_epi16(x0, quant_fast0);
+  y1 = _mm_mulhi_epi16(x1, quant_fast1);
+
+  /* ASM saves Y for EOB */
+  /* I think we can ignore that because adding the sign doesn't change anything
+   * and multiplying 0 by dequant is OK as well */
+  abs0 = y0;
+  abs1 = y1;
+
+  /* Restore the sign bit. */
+  y0 = _mm_xor_si128(y0, sz0);
+  y1 = _mm_xor_si128(y1, sz1);
+  x0 = _mm_sub_epi16(y0, sz0);
+  x1 = _mm_sub_epi16(y1, sz1);
+
+  /* qcoeff = x */
+  _mm_store_si128((__m128i *)(d->qcoeff), x0);
+  _mm_store_si128((__m128i *)(d->qcoeff + 8), x1);
+
+  /* x * dequant */
+  x0 = _mm_mullo_epi16(x0, dequant0);
+  x1 = _mm_mullo_epi16(x1, dequant1);
+
+  /* dqcoeff = x * dequant */
+  _mm_store_si128((__m128i *)(d->dqcoeff), x0);
+  _mm_store_si128((__m128i *)(d->dqcoeff + 8), x1);
+
+  zeros = _mm_setzero_si128();
+
+  x0 = _mm_cmpgt_epi16(abs0, zeros);
+  x1 = _mm_cmpgt_epi16(abs1, zeros);
+
+  x = _mm_packs_epi16(x0, x1);
+
+  x = _mm_shuffle_epi8(x, zig_zag);
+
+  mask = _mm_movemask_epi8(x);
+
+  eob = bsr(mask);
+
+  *d->eob = 0xFF & eob;
+}
diff --git a/libvpx/libvpx/vp8/encoder/x86/temporal_filter_apply_sse2.asm b/libvpx/libvpx/vp8/encoder/x86/temporal_filter_apply_sse2.asm
new file mode 100644
index 0000000..bd92b39
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/temporal_filter_apply_sse2.asm
@@ -0,0 +1,207 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+; void vp8_temporal_filter_apply_sse2 | arg
+;  (unsigned char  *frame1,           |  0
+;   unsigned int    stride,           |  1
+;   unsigned char  *frame2,           |  2
+;   unsigned int    block_size,       |  3
+;   int             strength,         |  4
+;   int             filter_weight,    |  5
+;   unsigned int   *accumulator,      |  6
+;   unsigned short *count)            |  7
+global sym(vp8_temporal_filter_apply_sse2) PRIVATE
+sym(vp8_temporal_filter_apply_sse2):
+
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 8
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ALIGN_STACK 16, rax
+    %define block_size    0
+    %define strength      16
+    %define filter_weight 32
+    %define rounding_bit  48
+    %define rbp_backup    64
+    %define stack_size    80
+    sub         rsp,           stack_size
+    mov         [rsp + rbp_backup], rbp
+    ; end prolog
+
+        mov         rdx,            arg(3)
+        mov         [rsp + block_size], rdx
+        movd        xmm6,            arg(4)
+        movdqa      [rsp + strength], xmm6 ; where strength is used, all 16 bytes are read
+
+        ; calculate the rounding bit outside the loop
+        ; 0x8000 >> (16 - strength)
+        mov         rdx,            16
+        sub         rdx,            arg(4) ; 16 - strength
+        movq        xmm4,           rdx    ; can't use rdx w/ shift
+        movdqa      xmm5,           [GLOBAL(_const_top_bit)]
+        psrlw       xmm5,           xmm4
+        movdqa      [rsp + rounding_bit], xmm5
+
+        mov         rsi,            arg(0) ; src/frame1
+        mov         rdx,            arg(2) ; predictor frame
+        mov         rdi,            arg(6) ; accumulator
+        mov         rax,            arg(7) ; count
+
+        ; dup the filter weight and store for later
+        movd        xmm0,           arg(5) ; filter_weight
+        pshuflw     xmm0,           xmm0, 0
+        punpcklwd   xmm0,           xmm0
+        movdqa      [rsp + filter_weight], xmm0
+
+        mov         rbp,            arg(1) ; stride
+        pxor        xmm7,           xmm7   ; zero for extraction
+
+        lea         rcx,            [rdx + 16*16*1]
+        cmp         dword ptr [rsp + block_size], 8
+        jne         .temporal_filter_apply_load_16
+        lea         rcx,            [rdx + 8*8*1]
+
+.temporal_filter_apply_load_8:
+        movq        xmm0,           [rsi]  ; first row
+        lea         rsi,            [rsi + rbp] ; += stride
+        punpcklbw   xmm0,           xmm7   ; src[ 0- 7]
+        movq        xmm1,           [rsi]  ; second row
+        lea         rsi,            [rsi + rbp] ; += stride
+        punpcklbw   xmm1,           xmm7   ; src[ 8-15]
+        jmp         .temporal_filter_apply_load_finished
+
+.temporal_filter_apply_load_16:
+        movdqa      xmm0,           [rsi]  ; src (frame1)
+        lea         rsi,            [rsi + rbp] ; += stride
+        movdqa      xmm1,           xmm0
+        punpcklbw   xmm0,           xmm7   ; src[ 0- 7]
+        punpckhbw   xmm1,           xmm7   ; src[ 8-15]
+
+.temporal_filter_apply_load_finished:
+        movdqa      xmm2,           [rdx]  ; predictor (frame2)
+        movdqa      xmm3,           xmm2
+        punpcklbw   xmm2,           xmm7   ; pred[ 0- 7]
+        punpckhbw   xmm3,           xmm7   ; pred[ 8-15]
+
+        ; modifier = src_byte - pixel_value
+        psubw       xmm0,           xmm2   ; src - pred[ 0- 7]
+        psubw       xmm1,           xmm3   ; src - pred[ 8-15]
+
+        ; modifier *= modifier
+        pmullw      xmm0,           xmm0   ; modifer[ 0- 7]^2
+        pmullw      xmm1,           xmm1   ; modifer[ 8-15]^2
+
+        ; modifier *= 3
+        pmullw      xmm0,           [GLOBAL(_const_3w)]
+        pmullw      xmm1,           [GLOBAL(_const_3w)]
+
+        ; modifer += 0x8000 >> (16 - strength)
+        paddw       xmm0,           [rsp + rounding_bit]
+        paddw       xmm1,           [rsp + rounding_bit]
+
+        ; modifier >>= strength
+        psrlw       xmm0,           [rsp + strength]
+        psrlw       xmm1,           [rsp + strength]
+
+        ; modifier = 16 - modifier
+        ; saturation takes care of modifier > 16
+        movdqa      xmm3,           [GLOBAL(_const_16w)]
+        movdqa      xmm2,           [GLOBAL(_const_16w)]
+        psubusw     xmm3,           xmm1
+        psubusw     xmm2,           xmm0
+
+        ; modifier *= filter_weight
+        pmullw      xmm2,           [rsp + filter_weight]
+        pmullw      xmm3,           [rsp + filter_weight]
+
+        ; count
+        movdqa      xmm4,           [rax]
+        movdqa      xmm5,           [rax+16]
+        ; += modifier
+        paddw       xmm4,           xmm2
+        paddw       xmm5,           xmm3
+        ; write back
+        movdqa      [rax],          xmm4
+        movdqa      [rax+16],       xmm5
+        lea         rax,            [rax + 16*2] ; count += 16*(sizeof(short))
+
+        ; load and extract the predictor up to shorts
+        pxor        xmm7,           xmm7
+        movdqa      xmm0,           [rdx]
+        lea         rdx,            [rdx + 16*1] ; pred += 16*(sizeof(char))
+        movdqa      xmm1,           xmm0
+        punpcklbw   xmm0,           xmm7   ; pred[ 0- 7]
+        punpckhbw   xmm1,           xmm7   ; pred[ 8-15]
+
+        ; modifier *= pixel_value
+        pmullw      xmm0,           xmm2
+        pmullw      xmm1,           xmm3
+
+        ; expand to double words
+        movdqa      xmm2,           xmm0
+        punpcklwd   xmm0,           xmm7   ; [ 0- 3]
+        punpckhwd   xmm2,           xmm7   ; [ 4- 7]
+        movdqa      xmm3,           xmm1
+        punpcklwd   xmm1,           xmm7   ; [ 8-11]
+        punpckhwd   xmm3,           xmm7   ; [12-15]
+
+        ; accumulator
+        movdqa      xmm4,           [rdi]
+        movdqa      xmm5,           [rdi+16]
+        movdqa      xmm6,           [rdi+32]
+        movdqa      xmm7,           [rdi+48]
+        ; += modifier
+        paddd       xmm4,           xmm0
+        paddd       xmm5,           xmm2
+        paddd       xmm6,           xmm1
+        paddd       xmm7,           xmm3
+        ; write back
+        movdqa      [rdi],          xmm4
+        movdqa      [rdi+16],       xmm5
+        movdqa      [rdi+32],       xmm6
+        movdqa      [rdi+48],       xmm7
+        lea         rdi,            [rdi + 16*4] ; accumulator += 16*(sizeof(int))
+
+        cmp         rdx,            rcx
+        je          .temporal_filter_apply_epilog
+        pxor        xmm7,           xmm7   ; zero for extraction
+        cmp         dword ptr [rsp + block_size], 16
+        je          .temporal_filter_apply_load_16
+        jmp         .temporal_filter_apply_load_8
+
+.temporal_filter_apply_epilog:
+    ; begin epilog
+    mov         rbp,            [rsp + rbp_backup]
+    add         rsp,            stack_size
+    pop         rsp
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+_const_3w:
+    times 8 dw 3
+align 16
+_const_top_bit:
+    times 8 dw 1<<15
+align 16
+_const_16w
+    times 8 dw 16
diff --git a/libvpx/libvpx/vp8/encoder/x86/vp8_enc_stubs_mmx.c b/libvpx/libvpx/vp8/encoder/x86/vp8_enc_stubs_mmx.c
new file mode 100644
index 0000000..7bf5155
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/vp8_enc_stubs_mmx.c
@@ -0,0 +1,67 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vpx_ports/x86.h"
+#include "vp8/encoder/block.h"
+
+void vp8_short_fdct4x4_mmx(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_mmx(short *input, short *output, int pitch)
+{
+    vp8_short_fdct4x4_mmx(input,   output,    pitch);
+    vp8_short_fdct4x4_mmx(input + 4, output + 16, pitch);
+}
+
+int vp8_fast_quantize_b_impl_mmx(short *coeff_ptr, short *zbin_ptr,
+                                 short *qcoeff_ptr, short *dequant_ptr,
+                                 const short *scan_mask, short *round_ptr,
+                                 short *quant_ptr, short *dqcoeff_ptr);
+void vp8_fast_quantize_b_mmx(BLOCK *b, BLOCKD *d)
+{
+    const short *scan_mask   = vp8_default_zig_zag_mask;
+    short *coeff_ptr   = b->coeff;
+    short *zbin_ptr    = b->zbin;
+    short *round_ptr   = b->round;
+    short *quant_ptr   = b->quant_fast;
+    short *qcoeff_ptr  = d->qcoeff;
+    short *dqcoeff_ptr = d->dqcoeff;
+    short *dequant_ptr = d->dequant;
+
+    *d->eob = (char)vp8_fast_quantize_b_impl_mmx(
+                                                 coeff_ptr,
+                                                 zbin_ptr,
+                                                 qcoeff_ptr,
+                                                 dequant_ptr,
+                                                 scan_mask,
+
+                                                 round_ptr,
+                                                 quant_ptr,
+                                                 dqcoeff_ptr
+                                                 );
+}
+
+int vp8_mbblock_error_mmx_impl(short *coeff_ptr, short *dcoef_ptr, int dc);
+int vp8_mbblock_error_mmx(MACROBLOCK *mb, int dc)
+{
+    short *coeff_ptr =  mb->block[0].coeff;
+    short *dcoef_ptr =  mb->e_mbd.block[0].dqcoeff;
+    return vp8_mbblock_error_mmx_impl(coeff_ptr, dcoef_ptr, dc);
+}
+
+int vp8_mbuverror_mmx_impl(short *s_ptr, short *d_ptr);
+int vp8_mbuverror_mmx(MACROBLOCK *mb)
+{
+    short *s_ptr = &mb->coeff[256];
+    short *d_ptr = &mb->e_mbd.dqcoeff[256];
+    return vp8_mbuverror_mmx_impl(s_ptr, d_ptr);
+}
+
diff --git a/libvpx/libvpx/vp8/encoder/x86/vp8_enc_stubs_sse2.c b/libvpx/libvpx/vp8/encoder/x86/vp8_enc_stubs_sse2.c
new file mode 100644
index 0000000..be9aaf3
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/vp8_enc_stubs_sse2.c
@@ -0,0 +1,32 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vpx_ports/x86.h"
+#include "vp8/encoder/block.h"
+
+int vp8_mbblock_error_xmm_impl(short *coeff_ptr, short *dcoef_ptr, int dc);
+int vp8_mbblock_error_xmm(MACROBLOCK *mb, int dc)
+{
+    short *coeff_ptr =  mb->block[0].coeff;
+    short *dcoef_ptr =  mb->e_mbd.block[0].dqcoeff;
+    return vp8_mbblock_error_xmm_impl(coeff_ptr, dcoef_ptr, dc);
+}
+
+int vp8_mbuverror_xmm_impl(short *s_ptr, short *d_ptr);
+int vp8_mbuverror_xmm(MACROBLOCK *mb)
+{
+    short *s_ptr = &mb->coeff[256];
+    short *d_ptr = &mb->e_mbd.dqcoeff[256];
+    return vp8_mbuverror_xmm_impl(s_ptr, d_ptr);
+}
+
diff --git a/libvpx/libvpx/vp8/encoder/x86/vp8_quantize_sse2.c b/libvpx/libvpx/vp8/encoder/x86/vp8_quantize_sse2.c
new file mode 100644
index 0000000..b4e92e0
--- /dev/null
+++ b/libvpx/libvpx/vp8/encoder/x86/vp8_quantize_sse2.c
@@ -0,0 +1,228 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_config.h"
+#include "vp8_rtcd.h"
+#include "vpx_ports/x86.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp8/encoder/block.h"
+#include "vp8/common/entropy.h" /* vp8_default_inv_zig_zag */
+
+#include <mmintrin.h> /* MMX */
+#include <xmmintrin.h> /* SSE */
+#include <emmintrin.h> /* SSE2 */
+
+#define SELECT_EOB(i, z) \
+    do { \
+        short boost = *zbin_boost_ptr; \
+        int cmp = (x[z] < boost) | (y[z] == 0); \
+        zbin_boost_ptr++; \
+        if (cmp) \
+            break; \
+        qcoeff_ptr[z] = y[z]; \
+        eob = i; \
+        zbin_boost_ptr = b->zrun_zbin_boost; \
+    } while (0)
+
+void vp8_regular_quantize_b_sse2(BLOCK *b, BLOCKD *d)
+{
+    char eob = 0;
+    short *zbin_boost_ptr;
+    short *qcoeff_ptr      = d->qcoeff;
+    DECLARE_ALIGNED(16, short, x[16]);
+    DECLARE_ALIGNED(16, short, y[16]);
+
+    __m128i sz0, x0, sz1, x1, y0, y1, x_minus_zbin0, x_minus_zbin1;
+    __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift));
+    __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8));
+    __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
+    __m128i z1 = _mm_load_si128((__m128i *)(b->coeff+8));
+    __m128i zbin_extra = _mm_cvtsi32_si128(b->zbin_extra);
+    __m128i zbin0 = _mm_load_si128((__m128i *)(b->zbin));
+    __m128i zbin1 = _mm_load_si128((__m128i *)(b->zbin + 8));
+    __m128i round0 = _mm_load_si128((__m128i *)(b->round));
+    __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
+    __m128i quant0 = _mm_load_si128((__m128i *)(b->quant));
+    __m128i quant1 = _mm_load_si128((__m128i *)(b->quant + 8));
+    __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
+    __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
+
+    memset(qcoeff_ptr, 0, 32);
+
+    /* Duplicate to all lanes. */
+    zbin_extra = _mm_shufflelo_epi16(zbin_extra, 0);
+    zbin_extra = _mm_unpacklo_epi16(zbin_extra, zbin_extra);
+
+    /* Sign of z: z >> 15 */
+    sz0 = _mm_srai_epi16(z0, 15);
+    sz1 = _mm_srai_epi16(z1, 15);
+
+    /* x = abs(z): (z ^ sz) - sz */
+    x0 = _mm_xor_si128(z0, sz0);
+    x1 = _mm_xor_si128(z1, sz1);
+    x0 = _mm_sub_epi16(x0, sz0);
+    x1 = _mm_sub_epi16(x1, sz1);
+
+    /* zbin[] + zbin_extra */
+    zbin0 = _mm_add_epi16(zbin0, zbin_extra);
+    zbin1 = _mm_add_epi16(zbin1, zbin_extra);
+
+    /* In C x is compared to zbin where zbin = zbin[] + boost + extra. Rebalance
+     * the equation because boost is the only value which can change:
+     * x - (zbin[] + extra) >= boost */
+    x_minus_zbin0 = _mm_sub_epi16(x0, zbin0);
+    x_minus_zbin1 = _mm_sub_epi16(x1, zbin1);
+
+    _mm_store_si128((__m128i *)(x), x_minus_zbin0);
+    _mm_store_si128((__m128i *)(x + 8), x_minus_zbin1);
+
+    /* All the remaining calculations are valid whether they are done now with
+     * simd or later inside the loop one at a time. */
+    x0 = _mm_add_epi16(x0, round0);
+    x1 = _mm_add_epi16(x1, round1);
+
+    y0 = _mm_mulhi_epi16(x0, quant0);
+    y1 = _mm_mulhi_epi16(x1, quant1);
+
+    y0 = _mm_add_epi16(y0, x0);
+    y1 = _mm_add_epi16(y1, x1);
+
+    /* Instead of shifting each value independently we convert the scaling
+     * factor with 1 << (16 - shift) so we can use multiply/return high half. */
+    y0 = _mm_mulhi_epi16(y0, quant_shift0);
+    y1 = _mm_mulhi_epi16(y1, quant_shift1);
+
+    /* Return the sign: (y ^ sz) - sz */
+    y0 = _mm_xor_si128(y0, sz0);
+    y1 = _mm_xor_si128(y1, sz1);
+    y0 = _mm_sub_epi16(y0, sz0);
+    y1 = _mm_sub_epi16(y1, sz1);
+
+    _mm_store_si128((__m128i *)(y), y0);
+    _mm_store_si128((__m128i *)(y + 8), y1);
+
+    zbin_boost_ptr = b->zrun_zbin_boost;
+
+    /* The loop gets unrolled anyway. Avoid the vp8_default_zig_zag1d lookup. */
+    SELECT_EOB(1, 0);
+    SELECT_EOB(2, 1);
+    SELECT_EOB(3, 4);
+    SELECT_EOB(4, 8);
+    SELECT_EOB(5, 5);
+    SELECT_EOB(6, 2);
+    SELECT_EOB(7, 3);
+    SELECT_EOB(8, 6);
+    SELECT_EOB(9, 9);
+    SELECT_EOB(10, 12);
+    SELECT_EOB(11, 13);
+    SELECT_EOB(12, 10);
+    SELECT_EOB(13, 7);
+    SELECT_EOB(14, 11);
+    SELECT_EOB(15, 14);
+    SELECT_EOB(16, 15);
+
+    y0 = _mm_load_si128((__m128i *)(d->qcoeff));
+    y1 = _mm_load_si128((__m128i *)(d->qcoeff + 8));
+
+    /* dqcoeff = qcoeff * dequant */
+    y0 = _mm_mullo_epi16(y0, dequant0);
+    y1 = _mm_mullo_epi16(y1, dequant1);
+
+    _mm_store_si128((__m128i *)(d->dqcoeff), y0);
+    _mm_store_si128((__m128i *)(d->dqcoeff + 8), y1);
+
+    *d->eob = eob;
+}
+
+void vp8_fast_quantize_b_sse2(BLOCK *b, BLOCKD *d)
+{
+  __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
+  __m128i z1 = _mm_load_si128((__m128i *)(b->coeff + 8));
+  __m128i round0 = _mm_load_si128((__m128i *)(b->round));
+  __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
+  __m128i quant_fast0 = _mm_load_si128((__m128i *)(b->quant_fast));
+  __m128i quant_fast1 = _mm_load_si128((__m128i *)(b->quant_fast + 8));
+  __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
+  __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
+  __m128i inv_zig_zag0 = _mm_load_si128((const __m128i *)(vp8_default_inv_zig_zag));
+  __m128i inv_zig_zag1 = _mm_load_si128((const __m128i *)(vp8_default_inv_zig_zag + 8));
+
+  __m128i sz0, sz1, x0, x1, y0, y1, xdq0, xdq1, zeros, ones;
+
+  /* sign of z: z >> 15 */
+  sz0 = _mm_srai_epi16(z0, 15);
+  sz1 = _mm_srai_epi16(z1, 15);
+
+  /* x = abs(z): (z ^ sz) - sz */
+  x0 = _mm_xor_si128(z0, sz0);
+  x1 = _mm_xor_si128(z1, sz1);
+  x0 = _mm_sub_epi16(x0, sz0);
+  x1 = _mm_sub_epi16(x1, sz1);
+
+  /* x += round */
+  x0 = _mm_add_epi16(x0, round0);
+  x1 = _mm_add_epi16(x1, round1);
+
+  /* y = (x * quant) >> 16 */
+  y0 = _mm_mulhi_epi16(x0, quant_fast0);
+  y1 = _mm_mulhi_epi16(x1, quant_fast1);
+
+  /* x = abs(y) = (y ^ sz) - sz */
+  y0 = _mm_xor_si128(y0, sz0);
+  y1 = _mm_xor_si128(y1, sz1);
+  x0 = _mm_sub_epi16(y0, sz0);
+  x1 = _mm_sub_epi16(y1, sz1);
+
+  /* qcoeff = x */
+  _mm_store_si128((__m128i *)(d->qcoeff), x0);
+  _mm_store_si128((__m128i *)(d->qcoeff + 8), x1);
+
+  /* x * dequant */
+  xdq0 = _mm_mullo_epi16(x0, dequant0);
+  xdq1 = _mm_mullo_epi16(x1, dequant1);
+
+  /* dqcoeff = x * dequant */
+  _mm_store_si128((__m128i *)(d->dqcoeff), xdq0);
+  _mm_store_si128((__m128i *)(d->dqcoeff + 8), xdq1);
+
+  /* build a mask for the zig zag */
+  zeros = _mm_setzero_si128();
+
+  x0 = _mm_cmpeq_epi16(x0, zeros);
+  x1 = _mm_cmpeq_epi16(x1, zeros);
+
+  ones = _mm_cmpeq_epi16(zeros, zeros);
+
+  x0 = _mm_xor_si128(x0, ones);
+  x1 = _mm_xor_si128(x1, ones);
+
+  x0 = _mm_and_si128(x0, inv_zig_zag0);
+  x1 = _mm_and_si128(x1, inv_zig_zag1);
+
+  x0 = _mm_max_epi16(x0, x1);
+
+  /* now down to 8 */
+  x1 = _mm_shuffle_epi32(x0, 0xE); // 0b00001110
+
+  x0 = _mm_max_epi16(x0, x1);
+
+  /* only 4 left */
+  x1 = _mm_shufflelo_epi16(x0, 0xE); // 0b00001110
+
+  x0 = _mm_max_epi16(x0, x1);
+
+  /* okay, just 2! */
+  x1 = _mm_shufflelo_epi16(x0, 0x1); // 0b00000001
+
+  x0 = _mm_max_epi16(x0, x1);
+
+  *d->eob = 0xFF & _mm_cvtsi128_si32(x0);
+}
diff --git a/libvpx/libvpx/vp8/exports_dec b/libvpx/libvpx/vp8/exports_dec
new file mode 100644
index 0000000..100ac5c
--- /dev/null
+++ b/libvpx/libvpx/vp8/exports_dec
@@ -0,0 +1,2 @@
+data vpx_codec_vp8_dx_algo
+text vpx_codec_vp8_dx
diff --git a/libvpx/libvpx/vp8/exports_enc b/libvpx/libvpx/vp8/exports_enc
new file mode 100644
index 0000000..29ff35e
--- /dev/null
+++ b/libvpx/libvpx/vp8/exports_enc
@@ -0,0 +1,2 @@
+data vpx_codec_vp8_cx_algo
+text vpx_codec_vp8_cx
diff --git a/libvpx/libvpx/vp8/vp8_common.mk b/libvpx/libvpx/vp8/vp8_common.mk
new file mode 100644
index 0000000..4c4e856
--- /dev/null
+++ b/libvpx/libvpx/vp8/vp8_common.mk
@@ -0,0 +1,169 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+VP8_COMMON_SRCS-yes += vp8_common.mk
+VP8_COMMON_SRCS-yes += common/ppflags.h
+VP8_COMMON_SRCS-yes += common/onyx.h
+VP8_COMMON_SRCS-yes += common/onyxd.h
+VP8_COMMON_SRCS-yes += common/alloccommon.c
+VP8_COMMON_SRCS-yes += common/blockd.c
+VP8_COMMON_SRCS-yes += common/coefupdateprobs.h
+VP8_COMMON_SRCS-yes += common/copy_c.c
+VP8_COMMON_SRCS-yes += common/debugmodes.c
+VP8_COMMON_SRCS-yes += common/default_coef_probs.h
+VP8_COMMON_SRCS-yes += common/dequantize.c
+VP8_COMMON_SRCS-yes += common/entropy.c
+VP8_COMMON_SRCS-yes += common/entropymode.c
+VP8_COMMON_SRCS-yes += common/entropymv.c
+VP8_COMMON_SRCS-yes += common/extend.c
+VP8_COMMON_SRCS-yes += common/filter.c
+VP8_COMMON_SRCS-yes += common/filter.h
+VP8_COMMON_SRCS-yes += common/findnearmv.c
+VP8_COMMON_SRCS-yes += common/generic/systemdependent.c
+VP8_COMMON_SRCS-yes += common/idct_blk.c
+VP8_COMMON_SRCS-yes += common/idctllm.c
+VP8_COMMON_SRCS-yes += common/alloccommon.h
+VP8_COMMON_SRCS-yes += common/blockd.h
+VP8_COMMON_SRCS-yes += common/common.h
+VP8_COMMON_SRCS-yes += common/entropy.h
+VP8_COMMON_SRCS-yes += common/entropymode.h
+VP8_COMMON_SRCS-yes += common/entropymv.h
+VP8_COMMON_SRCS-yes += common/extend.h
+VP8_COMMON_SRCS-yes += common/findnearmv.h
+VP8_COMMON_SRCS-yes += common/header.h
+VP8_COMMON_SRCS-yes += common/invtrans.h
+VP8_COMMON_SRCS-yes += common/loopfilter.h
+VP8_COMMON_SRCS-yes += common/modecont.h
+VP8_COMMON_SRCS-yes += common/mv.h
+VP8_COMMON_SRCS-yes += common/onyxc_int.h
+VP8_COMMON_SRCS-yes += common/quant_common.h
+VP8_COMMON_SRCS-yes += common/reconinter.h
+VP8_COMMON_SRCS-yes += common/reconintra.h
+VP8_COMMON_SRCS-yes += common/reconintra4x4.h
+VP8_COMMON_SRCS-yes += common/rtcd.c
+VP8_COMMON_SRCS-yes += common/rtcd_defs.pl
+VP8_COMMON_SRCS-yes += common/setupintrarecon.h
+VP8_COMMON_SRCS-yes += common/swapyv12buffer.h
+VP8_COMMON_SRCS-yes += common/systemdependent.h
+VP8_COMMON_SRCS-yes += common/threading.h
+VP8_COMMON_SRCS-yes += common/treecoder.h
+VP8_COMMON_SRCS-yes += common/vp8_loopfilter.c
+VP8_COMMON_SRCS-yes += common/loopfilter_filters.c
+VP8_COMMON_SRCS-yes += common/mbpitch.c
+VP8_COMMON_SRCS-yes += common/modecont.c
+VP8_COMMON_SRCS-yes += common/quant_common.c
+VP8_COMMON_SRCS-yes += common/reconinter.c
+VP8_COMMON_SRCS-yes += common/reconintra.c
+VP8_COMMON_SRCS-yes += common/reconintra4x4.c
+VP8_COMMON_SRCS-yes += common/setupintrarecon.c
+VP8_COMMON_SRCS-yes += common/swapyv12buffer.c
+VP8_COMMON_SRCS-yes += common/vp8_entropymodedata.h
+
+
+
+VP8_COMMON_SRCS-$(CONFIG_POSTPROC_VISUALIZER) += common/textblit.c
+VP8_COMMON_SRCS-yes += common/treecoder.c
+
+VP8_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/filter_x86.c
+VP8_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/filter_x86.h
+VP8_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/vp8_asm_stubs.c
+VP8_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/loopfilter_x86.c
+VP8_COMMON_SRCS-$(CONFIG_POSTPROC) += common/mfqe.c
+VP8_COMMON_SRCS-$(CONFIG_POSTPROC) += common/postproc.h
+VP8_COMMON_SRCS-$(CONFIG_POSTPROC) += common/postproc.c
+VP8_COMMON_SRCS-$(HAVE_MMX) += common/x86/dequantize_mmx.asm
+VP8_COMMON_SRCS-$(HAVE_MMX) += common/x86/idct_blk_mmx.c
+VP8_COMMON_SRCS-$(HAVE_MMX) += common/x86/idctllm_mmx.asm
+VP8_COMMON_SRCS-$(HAVE_MMX) += common/x86/iwalsh_mmx.asm
+VP8_COMMON_SRCS-$(HAVE_MMX) += common/x86/vp8_loopfilter_mmx.asm
+VP8_COMMON_SRCS-$(HAVE_MMX) += common/x86/recon_mmx.asm
+VP8_COMMON_SRCS-$(HAVE_MMX) += common/x86/subpixel_mmx.asm
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/copy_sse2.asm
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/idct_blk_sse2.c
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/idctllm_sse2.asm
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/recon_sse2.asm
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/subpixel_sse2.asm
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/loopfilter_sse2.asm
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/iwalsh_sse2.asm
+VP8_COMMON_SRCS-$(HAVE_SSE3) += common/x86/copy_sse3.asm
+VP8_COMMON_SRCS-$(HAVE_SSSE3) += common/x86/subpixel_ssse3.asm
+
+ifeq ($(CONFIG_POSTPROC),yes)
+VP8_COMMON_SRCS-$(HAVE_MMX) += common/x86/postproc_mmx.asm
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/mfqe_sse2.asm
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/postproc_sse2.asm
+endif
+
+ifeq ($(ARCH_X86_64),yes)
+VP8_COMMON_SRCS-$(HAVE_SSE2) += common/x86/loopfilter_block_sse2_x86_64.asm
+endif
+
+# common (c)
+VP8_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/idctllm_dspr2.c
+VP8_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/filter_dspr2.c
+VP8_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/vp8_loopfilter_filters_dspr2.c
+VP8_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/reconinter_dspr2.c
+VP8_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/idct_blk_dspr2.c
+VP8_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/dequantize_dspr2.c
+
+# common (c)
+VP8_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/bilinear_filter_msa.c
+VP8_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/copymem_msa.c
+VP8_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/idct_msa.c
+VP8_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/loopfilter_filters_msa.c
+VP8_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/sixtap_filter_msa.c
+VP8_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp8_macros_msa.h
+
+ifeq ($(CONFIG_POSTPROC),yes)
+VP8_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/mfqe_msa.c
+VP8_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/postproc_msa.c
+endif
+
+# common (c)
+VP8_COMMON_SRCS-$(ARCH_ARM)  += common/arm/filter_arm.c
+VP8_COMMON_SRCS-$(ARCH_ARM)  += common/arm/loopfilter_arm.c
+VP8_COMMON_SRCS-$(ARCH_ARM)  += common/arm/dequantize_arm.c
+
+# common (media)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/bilinearfilter_arm.c
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/bilinearfilter_arm.h
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/bilinearfilter_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/copymem8x4_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/copymem8x8_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/copymem16x16_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/dc_only_idct_add_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/iwalsh_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/filter_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/idct_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/loopfilter_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/simpleloopfilter_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/sixtappredict8x4_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/dequant_idct_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/dequantize_v6$(ASM)
+VP8_COMMON_SRCS-$(HAVE_MEDIA)  += common/arm/armv6/idct_blk_v6.c
+
+# common (neon intrinsics)
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/bilinearpredict_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/copymem_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/dc_only_idct_add_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/dequant_idct_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/dequantizeb_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/idct_blk_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/idct_dequant_0_2x_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/idct_dequant_full_2x_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/iwalsh_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/vp8_loopfilter_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/loopfiltersimplehorizontaledge_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/loopfiltersimpleverticaledge_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/mbloopfilter_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/shortidct4x4llm_neon.c
+VP8_COMMON_SRCS-$(HAVE_NEON)  += common/arm/neon/sixtappredict_neon.c
+
+$(eval $(call rtcd_h_template,vp8_rtcd,vp8/common/rtcd_defs.pl))
diff --git a/libvpx/libvpx/vp8/vp8_cx_iface.c b/libvpx/libvpx/vp8/vp8_cx_iface.c
new file mode 100644
index 0000000..22a82b7
--- /dev/null
+++ b/libvpx/libvpx/vp8/vp8_cx_iface.c
@@ -0,0 +1,1378 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "./vpx_config.h"
+#include "./vp8_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "vpx_version.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/vpx_once.h"
+#include "vp8/encoder/onyx_int.h"
+#include "vpx/vp8cx.h"
+#include "vp8/encoder/firstpass.h"
+#include "vp8/common/onyx.h"
+#include "vp8/common/common.h"
+#include <stdlib.h>
+#include <string.h>
+
+struct vp8_extracfg
+{
+    struct vpx_codec_pkt_list *pkt_list;
+    int                         cpu_used;                    /** available cpu percentage in 1/16*/
+    unsigned int                enable_auto_alt_ref;           /** if encoder decides to uses alternate reference frame */
+    unsigned int                noise_sensitivity;
+    unsigned int                Sharpness;
+    unsigned int                static_thresh;
+    unsigned int                token_partitions;
+    unsigned int                arnr_max_frames;    /* alt_ref Noise Reduction Max Frame Count */
+    unsigned int                arnr_strength;    /* alt_ref Noise Reduction Strength */
+    unsigned int                arnr_type;        /* alt_ref filter type */
+    vp8e_tuning                 tuning;
+    unsigned int                cq_level;         /* constrained quality level */
+    unsigned int                rc_max_intra_bitrate_pct;
+    unsigned int                screen_content_mode;
+
+};
+
+static struct vp8_extracfg default_extracfg = {
+  NULL,
+#if !(CONFIG_REALTIME_ONLY)
+  0,                          /* cpu_used      */
+#else
+  4,                          /* cpu_used      */
+#endif
+  0,                          /* enable_auto_alt_ref */
+  0,                          /* noise_sensitivity */
+  0,                          /* Sharpness */
+  0,                          /* static_thresh */
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+  VP8_EIGHT_TOKENPARTITION,
+#else
+  VP8_ONE_TOKENPARTITION,     /* token_partitions */
+#endif
+  0,                          /* arnr_max_frames */
+  3,                          /* arnr_strength */
+  3,                          /* arnr_type*/
+  0,                          /* tuning*/
+  10,                         /* cq_level */
+  0,                          /* rc_max_intra_bitrate_pct */
+  0,                          /* screen_content_mode */
+};
+
+struct vpx_codec_alg_priv
+{
+    vpx_codec_priv_t        base;
+    vpx_codec_enc_cfg_t     cfg;
+    struct vp8_extracfg     vp8_cfg;
+    VP8_CONFIG              oxcf;
+    struct VP8_COMP        *cpi;
+    unsigned char          *cx_data;
+    unsigned int            cx_data_sz;
+    vpx_image_t             preview_img;
+    unsigned int            next_frame_flag;
+    vp8_postproc_cfg_t      preview_ppcfg;
+    /* pkt_list size depends on the maximum number of lagged frames allowed. */
+    vpx_codec_pkt_list_decl(64) pkt_list;
+    unsigned int                fixed_kf_cntr;
+    vpx_enc_frame_flags_t   control_frame_flags;
+};
+
+
+static vpx_codec_err_t
+update_error_state(vpx_codec_alg_priv_t                 *ctx,
+                   const struct vpx_internal_error_info *error)
+{
+    vpx_codec_err_t res;
+
+    if ((res = error->error_code))
+        ctx->base.err_detail = error->has_detail
+                               ? error->detail
+                               : NULL;
+
+    return res;
+}
+
+
+#undef ERROR
+#define ERROR(str) do {\
+        ctx->base.err_detail = str;\
+        return VPX_CODEC_INVALID_PARAM;\
+    } while(0)
+
+#define RANGE_CHECK(p,memb,lo,hi) do {\
+        if(!(((p)->memb == lo || (p)->memb > (lo)) && (p)->memb <= hi)) \
+            ERROR(#memb " out of range ["#lo".."#hi"]");\
+    } while(0)
+
+#define RANGE_CHECK_HI(p,memb,hi) do {\
+        if(!((p)->memb <= (hi))) \
+            ERROR(#memb " out of range [.."#hi"]");\
+    } while(0)
+
+#define RANGE_CHECK_LO(p,memb,lo) do {\
+        if(!((p)->memb >= (lo))) \
+            ERROR(#memb " out of range ["#lo"..]");\
+    } while(0)
+
+#define RANGE_CHECK_BOOL(p,memb) do {\
+        if(!!((p)->memb) != (p)->memb) ERROR(#memb " expected boolean");\
+    } while(0)
+
+static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t      *ctx,
+                                       const vpx_codec_enc_cfg_t *cfg,
+                                       const struct vp8_extracfg *vp8_cfg,
+                                       int                        finalize)
+{
+    RANGE_CHECK(cfg, g_w,                   1, 16383); /* 14 bits available */
+    RANGE_CHECK(cfg, g_h,                   1, 16383); /* 14 bits available */
+    RANGE_CHECK(cfg, g_timebase.den,        1, 1000000000);
+    RANGE_CHECK(cfg, g_timebase.num,        1, 1000000000);
+    RANGE_CHECK_HI(cfg, g_profile,          3);
+    RANGE_CHECK_HI(cfg, rc_max_quantizer,   63);
+    RANGE_CHECK_HI(cfg, rc_min_quantizer,   cfg->rc_max_quantizer);
+    RANGE_CHECK_HI(cfg, g_threads,          64);
+#if CONFIG_REALTIME_ONLY
+    RANGE_CHECK_HI(cfg, g_lag_in_frames,    0);
+#elif CONFIG_MULTI_RES_ENCODING
+    if (ctx->base.enc.total_encoders > 1)
+        RANGE_CHECK_HI(cfg, g_lag_in_frames,    0);
+#else
+    RANGE_CHECK_HI(cfg, g_lag_in_frames,    25);
+#endif
+    RANGE_CHECK(cfg, rc_end_usage,          VPX_VBR, VPX_Q);
+    RANGE_CHECK_HI(cfg, rc_undershoot_pct,  1000);
+    RANGE_CHECK_HI(cfg, rc_overshoot_pct,   1000);
+    RANGE_CHECK_HI(cfg, rc_2pass_vbr_bias_pct, 100);
+    RANGE_CHECK(cfg, kf_mode,               VPX_KF_DISABLED, VPX_KF_AUTO);
+
+/* TODO: add spatial re-sampling support and frame dropping in
+ * multi-res-encoder.*/
+#if CONFIG_MULTI_RES_ENCODING
+    if (ctx->base.enc.total_encoders > 1)
+        RANGE_CHECK_HI(cfg, rc_resize_allowed,     0);
+#else
+    RANGE_CHECK_BOOL(cfg, rc_resize_allowed);
+#endif
+    RANGE_CHECK_HI(cfg, rc_dropframe_thresh,   100);
+    RANGE_CHECK_HI(cfg, rc_resize_up_thresh,   100);
+    RANGE_CHECK_HI(cfg, rc_resize_down_thresh, 100);
+
+#if CONFIG_REALTIME_ONLY
+    RANGE_CHECK(cfg,        g_pass,         VPX_RC_ONE_PASS, VPX_RC_ONE_PASS);
+#elif CONFIG_MULTI_RES_ENCODING
+    if (ctx->base.enc.total_encoders > 1)
+        RANGE_CHECK(cfg,    g_pass,         VPX_RC_ONE_PASS, VPX_RC_ONE_PASS);
+#else
+    RANGE_CHECK(cfg,        g_pass,         VPX_RC_ONE_PASS, VPX_RC_LAST_PASS);
+#endif
+
+    /* VP8 does not support a lower bound on the keyframe interval in
+     * automatic keyframe placement mode.
+     */
+    if (cfg->kf_mode != VPX_KF_DISABLED && cfg->kf_min_dist != cfg->kf_max_dist
+        && cfg->kf_min_dist > 0)
+        ERROR("kf_min_dist not supported in auto mode, use 0 "
+              "or kf_max_dist instead.");
+
+    RANGE_CHECK_BOOL(vp8_cfg,               enable_auto_alt_ref);
+    RANGE_CHECK(vp8_cfg, cpu_used,           -16, 16);
+
+#if CONFIG_REALTIME_ONLY && !CONFIG_TEMPORAL_DENOISING
+    RANGE_CHECK(vp8_cfg, noise_sensitivity,  0, 0);
+#else
+    RANGE_CHECK_HI(vp8_cfg, noise_sensitivity,  6);
+#endif
+
+    RANGE_CHECK(vp8_cfg, token_partitions,   VP8_ONE_TOKENPARTITION,
+                VP8_EIGHT_TOKENPARTITION);
+    RANGE_CHECK_HI(vp8_cfg, Sharpness,       7);
+    RANGE_CHECK(vp8_cfg, arnr_max_frames, 0, 15);
+    RANGE_CHECK_HI(vp8_cfg, arnr_strength,   6);
+    RANGE_CHECK(vp8_cfg, arnr_type,       1, 3);
+    RANGE_CHECK(vp8_cfg, cq_level, 0, 63);
+    RANGE_CHECK_HI(vp8_cfg, screen_content_mode, 2);
+    if (finalize && (cfg->rc_end_usage == VPX_CQ || cfg->rc_end_usage == VPX_Q))
+        RANGE_CHECK(vp8_cfg, cq_level,
+                    cfg->rc_min_quantizer, cfg->rc_max_quantizer);
+
+#if !(CONFIG_REALTIME_ONLY)
+    if (cfg->g_pass == VPX_RC_LAST_PASS)
+    {
+        size_t           packet_sz = sizeof(FIRSTPASS_STATS);
+        int              n_packets = (int)(cfg->rc_twopass_stats_in.sz /
+                                          packet_sz);
+        FIRSTPASS_STATS *stats;
+
+        if (!cfg->rc_twopass_stats_in.buf)
+            ERROR("rc_twopass_stats_in.buf not set.");
+
+        if (cfg->rc_twopass_stats_in.sz % packet_sz)
+            ERROR("rc_twopass_stats_in.sz indicates truncated packet.");
+
+        if (cfg->rc_twopass_stats_in.sz < 2 * packet_sz)
+            ERROR("rc_twopass_stats_in requires at least two packets.");
+
+        stats = (void*)((char *)cfg->rc_twopass_stats_in.buf
+                + (n_packets - 1) * packet_sz);
+
+        if ((int)(stats->count + 0.5) != n_packets - 1)
+            ERROR("rc_twopass_stats_in missing EOS stats packet");
+    }
+#endif
+
+    RANGE_CHECK(cfg, ts_number_layers, 1, 5);
+
+    if (cfg->ts_number_layers > 1)
+    {
+        unsigned int i;
+        RANGE_CHECK_HI(cfg, ts_periodicity, 16);
+
+        for (i=1; i<cfg->ts_number_layers; i++)
+            if (cfg->ts_target_bitrate[i] <= cfg->ts_target_bitrate[i-1] &&
+                cfg->rc_target_bitrate > 0)
+                ERROR("ts_target_bitrate entries are not strictly increasing");
+
+        RANGE_CHECK(cfg, ts_rate_decimator[cfg->ts_number_layers-1], 1, 1);
+        for (i=cfg->ts_number_layers-2; i>0; i--)
+            if (cfg->ts_rate_decimator[i-1] != 2*cfg->ts_rate_decimator[i])
+                ERROR("ts_rate_decimator factors are not powers of 2");
+
+        RANGE_CHECK_HI(cfg, ts_layer_id[i], cfg->ts_number_layers-1);
+    }
+
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+    if(cfg->g_threads > (1 << vp8_cfg->token_partitions))
+        ERROR("g_threads cannot be bigger than number of token partitions");
+#endif
+
+    return VPX_CODEC_OK;
+}
+
+
+static vpx_codec_err_t validate_img(vpx_codec_alg_priv_t *ctx,
+                                    const vpx_image_t    *img)
+{
+    switch (img->fmt)
+    {
+    case VPX_IMG_FMT_YV12:
+    case VPX_IMG_FMT_I420:
+    case VPX_IMG_FMT_VPXI420:
+    case VPX_IMG_FMT_VPXYV12:
+        break;
+    default:
+        ERROR("Invalid image format. Only YV12 and I420 images are supported");
+    }
+
+    if ((img->d_w != ctx->cfg.g_w) || (img->d_h != ctx->cfg.g_h))
+        ERROR("Image size must match encoder init configuration size");
+
+    return VPX_CODEC_OK;
+}
+
+
+static vpx_codec_err_t set_vp8e_config(VP8_CONFIG *oxcf,
+                                       vpx_codec_enc_cfg_t cfg,
+                                       struct vp8_extracfg vp8_cfg,
+                                       vpx_codec_priv_enc_mr_cfg_t *mr_cfg)
+{
+    oxcf->multi_threaded         = cfg.g_threads;
+    oxcf->Version               = cfg.g_profile;
+
+    oxcf->Width                 = cfg.g_w;
+    oxcf->Height                = cfg.g_h;
+    oxcf->timebase              = cfg.g_timebase;
+
+    oxcf->error_resilient_mode = cfg.g_error_resilient;
+
+    switch (cfg.g_pass)
+    {
+    case VPX_RC_ONE_PASS:
+        oxcf->Mode = MODE_BESTQUALITY;
+        break;
+    case VPX_RC_FIRST_PASS:
+        oxcf->Mode = MODE_FIRSTPASS;
+        break;
+    case VPX_RC_LAST_PASS:
+        oxcf->Mode = MODE_SECONDPASS_BEST;
+        break;
+    }
+
+    if (cfg.g_pass == VPX_RC_FIRST_PASS || cfg.g_pass == VPX_RC_ONE_PASS)
+    {
+        oxcf->allow_lag     = 0;
+        oxcf->lag_in_frames = 0;
+    }
+    else
+    {
+        oxcf->allow_lag     = (cfg.g_lag_in_frames) > 0;
+        oxcf->lag_in_frames = cfg.g_lag_in_frames;
+    }
+
+    oxcf->allow_df               = (cfg.rc_dropframe_thresh > 0);
+    oxcf->drop_frames_water_mark   = cfg.rc_dropframe_thresh;
+
+    oxcf->allow_spatial_resampling = cfg.rc_resize_allowed;
+    oxcf->resample_up_water_mark   = cfg.rc_resize_up_thresh;
+    oxcf->resample_down_water_mark = cfg.rc_resize_down_thresh;
+
+    if (cfg.rc_end_usage == VPX_VBR) {
+      oxcf->end_usage = USAGE_LOCAL_FILE_PLAYBACK;
+    } else if (cfg.rc_end_usage == VPX_CBR) {
+      oxcf->end_usage = USAGE_STREAM_FROM_SERVER;
+    } else if (cfg.rc_end_usage == VPX_CQ) {
+      oxcf->end_usage = USAGE_CONSTRAINED_QUALITY;
+    } else if (cfg.rc_end_usage == VPX_Q) {
+      oxcf->end_usage = USAGE_CONSTANT_QUALITY;
+    }
+
+    oxcf->target_bandwidth         = cfg.rc_target_bitrate;
+    oxcf->rc_max_intra_bitrate_pct = vp8_cfg.rc_max_intra_bitrate_pct;
+
+    oxcf->best_allowed_q           = cfg.rc_min_quantizer;
+    oxcf->worst_allowed_q          = cfg.rc_max_quantizer;
+    oxcf->cq_level                 = vp8_cfg.cq_level;
+    oxcf->fixed_q = -1;
+
+    oxcf->under_shoot_pct          = cfg.rc_undershoot_pct;
+    oxcf->over_shoot_pct           = cfg.rc_overshoot_pct;
+
+    oxcf->maximum_buffer_size_in_ms   = cfg.rc_buf_sz;
+    oxcf->starting_buffer_level_in_ms = cfg.rc_buf_initial_sz;
+    oxcf->optimal_buffer_level_in_ms  = cfg.rc_buf_optimal_sz;
+
+    oxcf->maximum_buffer_size      = cfg.rc_buf_sz;
+    oxcf->starting_buffer_level    = cfg.rc_buf_initial_sz;
+    oxcf->optimal_buffer_level     = cfg.rc_buf_optimal_sz;
+
+    oxcf->two_pass_vbrbias         = cfg.rc_2pass_vbr_bias_pct;
+    oxcf->two_pass_vbrmin_section  = cfg.rc_2pass_vbr_minsection_pct;
+    oxcf->two_pass_vbrmax_section  = cfg.rc_2pass_vbr_maxsection_pct;
+
+    oxcf->auto_key                 = cfg.kf_mode == VPX_KF_AUTO
+                                       && cfg.kf_min_dist != cfg.kf_max_dist;
+    oxcf->key_freq                 = cfg.kf_max_dist;
+
+    oxcf->number_of_layers         = cfg.ts_number_layers;
+    oxcf->periodicity              = cfg.ts_periodicity;
+
+    if (oxcf->number_of_layers > 1)
+    {
+        memcpy (oxcf->target_bitrate, cfg.ts_target_bitrate,
+                sizeof(cfg.ts_target_bitrate));
+        memcpy (oxcf->rate_decimator, cfg.ts_rate_decimator,
+                sizeof(cfg.ts_rate_decimator));
+        memcpy (oxcf->layer_id, cfg.ts_layer_id, sizeof(cfg.ts_layer_id));
+    }
+
+#if CONFIG_MULTI_RES_ENCODING
+    /* When mr_cfg is NULL, oxcf->mr_total_resolutions and oxcf->mr_encoder_id
+     * are both memset to 0, which ensures the correct logic under this
+     * situation.
+     */
+    if(mr_cfg)
+    {
+        oxcf->mr_total_resolutions        = mr_cfg->mr_total_resolutions;
+        oxcf->mr_encoder_id               = mr_cfg->mr_encoder_id;
+        oxcf->mr_down_sampling_factor.num = mr_cfg->mr_down_sampling_factor.num;
+        oxcf->mr_down_sampling_factor.den = mr_cfg->mr_down_sampling_factor.den;
+        oxcf->mr_low_res_mode_info        = mr_cfg->mr_low_res_mode_info;
+    }
+#else
+    (void)mr_cfg;
+#endif
+
+    oxcf->cpu_used               = vp8_cfg.cpu_used;
+    oxcf->encode_breakout        = vp8_cfg.static_thresh;
+    oxcf->play_alternate         = vp8_cfg.enable_auto_alt_ref;
+    oxcf->noise_sensitivity      = vp8_cfg.noise_sensitivity;
+    oxcf->Sharpness              = vp8_cfg.Sharpness;
+    oxcf->token_partitions       = vp8_cfg.token_partitions;
+
+    oxcf->two_pass_stats_in      = cfg.rc_twopass_stats_in;
+    oxcf->output_pkt_list        = vp8_cfg.pkt_list;
+
+    oxcf->arnr_max_frames        = vp8_cfg.arnr_max_frames;
+    oxcf->arnr_strength          = vp8_cfg.arnr_strength;
+    oxcf->arnr_type              = vp8_cfg.arnr_type;
+
+    oxcf->tuning                 = vp8_cfg.tuning;
+
+    oxcf->screen_content_mode    = vp8_cfg.screen_content_mode;
+
+    /*
+        printf("Current VP8 Settings: \n");
+        printf("target_bandwidth: %d\n", oxcf->target_bandwidth);
+        printf("noise_sensitivity: %d\n", oxcf->noise_sensitivity);
+        printf("Sharpness: %d\n",    oxcf->Sharpness);
+        printf("cpu_used: %d\n",  oxcf->cpu_used);
+        printf("Mode: %d\n",     oxcf->Mode);
+        printf("auto_key: %d\n",  oxcf->auto_key);
+        printf("key_freq: %d\n", oxcf->key_freq);
+        printf("end_usage: %d\n", oxcf->end_usage);
+        printf("under_shoot_pct: %d\n", oxcf->under_shoot_pct);
+        printf("over_shoot_pct: %d\n", oxcf->over_shoot_pct);
+        printf("starting_buffer_level: %d\n", oxcf->starting_buffer_level);
+        printf("optimal_buffer_level: %d\n",  oxcf->optimal_buffer_level);
+        printf("maximum_buffer_size: %d\n", oxcf->maximum_buffer_size);
+        printf("fixed_q: %d\n",  oxcf->fixed_q);
+        printf("worst_allowed_q: %d\n", oxcf->worst_allowed_q);
+        printf("best_allowed_q: %d\n", oxcf->best_allowed_q);
+        printf("allow_spatial_resampling: %d\n",  oxcf->allow_spatial_resampling);
+        printf("resample_down_water_mark: %d\n", oxcf->resample_down_water_mark);
+        printf("resample_up_water_mark: %d\n", oxcf->resample_up_water_mark);
+        printf("allow_df: %d\n", oxcf->allow_df);
+        printf("drop_frames_water_mark: %d\n", oxcf->drop_frames_water_mark);
+        printf("two_pass_vbrbias: %d\n",  oxcf->two_pass_vbrbias);
+        printf("two_pass_vbrmin_section: %d\n", oxcf->two_pass_vbrmin_section);
+        printf("two_pass_vbrmax_section: %d\n", oxcf->two_pass_vbrmax_section);
+        printf("allow_lag: %d\n", oxcf->allow_lag);
+        printf("lag_in_frames: %d\n", oxcf->lag_in_frames);
+        printf("play_alternate: %d\n", oxcf->play_alternate);
+        printf("Version: %d\n", oxcf->Version);
+        printf("multi_threaded: %d\n",   oxcf->multi_threaded);
+        printf("encode_breakout: %d\n", oxcf->encode_breakout);
+    */
+    return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t vp8e_set_config(vpx_codec_alg_priv_t       *ctx,
+                                       const vpx_codec_enc_cfg_t  *cfg)
+{
+    vpx_codec_err_t res;
+
+    if (cfg->g_w != ctx->cfg.g_w || cfg->g_h != ctx->cfg.g_h)
+    {
+        if (cfg->g_lag_in_frames > 1 || cfg->g_pass != VPX_RC_ONE_PASS)
+            ERROR("Cannot change width or height after initialization");
+        if ((ctx->cpi->initial_width && (int)cfg->g_w > ctx->cpi->initial_width) ||
+            (ctx->cpi->initial_height && (int)cfg->g_h > ctx->cpi->initial_height))
+            ERROR("Cannot increase width or height larger than their initial values");
+    }
+
+    /* Prevent increasing lag_in_frames. This check is stricter than it needs
+     * to be -- the limit is not increasing past the first lag_in_frames
+     * value, but we don't track the initial config, only the last successful
+     * config.
+     */
+    if ((cfg->g_lag_in_frames > ctx->cfg.g_lag_in_frames))
+        ERROR("Cannot increase lag_in_frames");
+
+    res = validate_config(ctx, cfg, &ctx->vp8_cfg, 0);
+
+    if (!res)
+    {
+        ctx->cfg = *cfg;
+        set_vp8e_config(&ctx->oxcf, ctx->cfg, ctx->vp8_cfg, NULL);
+        vp8_change_config(ctx->cpi, &ctx->oxcf);
+    }
+
+    return res;
+}
+
+static vpx_codec_err_t get_quantizer(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL)
+    return VPX_CODEC_INVALID_PARAM;
+  *arg = vp8_get_quantizer(ctx->cpi);
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t get_quantizer64(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL)
+    return VPX_CODEC_INVALID_PARAM;
+  *arg = vp8_reverse_trans(vp8_get_quantizer(ctx->cpi));
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t update_extracfg(vpx_codec_alg_priv_t *ctx,
+                                       const struct vp8_extracfg *extra_cfg)
+{
+  const vpx_codec_err_t res = validate_config(ctx, &ctx->cfg, extra_cfg, 0);
+  if (res == VPX_CODEC_OK) {
+    ctx->vp8_cfg = *extra_cfg;
+    set_vp8e_config(&ctx->oxcf, ctx->cfg, ctx->vp8_cfg, NULL);
+    vp8_change_config(ctx->cpi, &ctx->oxcf);
+  }
+  return res;
+}
+
+static vpx_codec_err_t set_cpu_used(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.cpu_used = CAST(VP8E_SET_CPUUSED, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_enable_auto_alt_ref(vpx_codec_alg_priv_t *ctx,
+                                               va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.enable_auto_alt_ref = CAST(VP8E_SET_ENABLEAUTOALTREF, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_noise_sensitivity(vpx_codec_alg_priv_t *ctx,
+                                             va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.noise_sensitivity = CAST(VP8E_SET_NOISE_SENSITIVITY, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_sharpness(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.Sharpness = CAST(VP8E_SET_SHARPNESS, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_static_thresh(vpx_codec_alg_priv_t *ctx,
+                                         va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.static_thresh = CAST(VP8E_SET_STATIC_THRESHOLD, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_token_partitions(vpx_codec_alg_priv_t *ctx,
+                                            va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.token_partitions = CAST(VP8E_SET_TOKEN_PARTITIONS, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_arnr_max_frames(vpx_codec_alg_priv_t *ctx,
+                                           va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.arnr_max_frames = CAST(VP8E_SET_ARNR_MAXFRAMES, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_arnr_strength(vpx_codec_alg_priv_t *ctx,
+                                         va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.arnr_strength = CAST(VP8E_SET_ARNR_STRENGTH, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_arnr_type(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.arnr_type = CAST(VP8E_SET_ARNR_TYPE, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_tuning(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.tuning = CAST(VP8E_SET_TUNING, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_cq_level(vpx_codec_alg_priv_t *ctx, va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.cq_level = CAST(VP8E_SET_CQ_LEVEL, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_rc_max_intra_bitrate_pct(vpx_codec_alg_priv_t *ctx,
+                                                    va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.rc_max_intra_bitrate_pct =
+      CAST(VP8E_SET_MAX_INTRA_BITRATE_PCT, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t set_screen_content_mode(vpx_codec_alg_priv_t *ctx,
+                                               va_list args)
+{
+  struct vp8_extracfg extra_cfg = ctx->vp8_cfg;
+  extra_cfg.screen_content_mode =
+      CAST(VP8E_SET_SCREEN_CONTENT_MODE, args);
+  return update_extracfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t vp8e_mr_alloc_mem(const vpx_codec_enc_cfg_t *cfg,
+                                        void **mem_loc)
+{
+    vpx_codec_err_t res = 0;
+
+#if CONFIG_MULTI_RES_ENCODING
+    LOWER_RES_FRAME_INFO *shared_mem_loc;
+    int mb_rows = ((cfg->g_w + 15) >>4);
+    int mb_cols = ((cfg->g_h + 15) >>4);
+
+    shared_mem_loc = calloc(1, sizeof(LOWER_RES_FRAME_INFO));
+    if(!shared_mem_loc)
+    {
+        res = VPX_CODEC_MEM_ERROR;
+    }
+
+    shared_mem_loc->mb_info = calloc(mb_rows*mb_cols, sizeof(LOWER_RES_MB_INFO));
+    if(!(shared_mem_loc->mb_info))
+    {
+        res = VPX_CODEC_MEM_ERROR;
+    }
+    else
+    {
+        *mem_loc = (void *)shared_mem_loc;
+        res = VPX_CODEC_OK;
+    }
+#else
+    (void)cfg;
+    (void)mem_loc;
+#endif
+    return res;
+}
+
+static vpx_codec_err_t vp8e_init(vpx_codec_ctx_t *ctx,
+                                 vpx_codec_priv_enc_mr_cfg_t *mr_cfg)
+{
+    vpx_codec_err_t        res = VPX_CODEC_OK;
+
+
+    vp8_rtcd();
+    vpx_dsp_rtcd();
+    vpx_scale_rtcd();
+
+    if (!ctx->priv)
+    {
+        struct vpx_codec_alg_priv *priv =
+            (struct vpx_codec_alg_priv *)vpx_calloc(1, sizeof(*priv));
+
+        if (!priv)
+        {
+            return VPX_CODEC_MEM_ERROR;
+        }
+
+        ctx->priv = (vpx_codec_priv_t *)priv;
+        ctx->priv->init_flags = ctx->init_flags;
+
+        if (ctx->config.enc)
+        {
+            /* Update the reference to the config structure to an
+             * internal copy.
+             */
+            priv->cfg = *ctx->config.enc;
+            ctx->config.enc = &priv->cfg;
+        }
+
+        priv->vp8_cfg = default_extracfg;
+        priv->vp8_cfg.pkt_list = &priv->pkt_list.head;
+
+        priv->cx_data_sz = priv->cfg.g_w * priv->cfg.g_h * 3 / 2 * 2;
+
+        if (priv->cx_data_sz < 32768) priv->cx_data_sz = 32768;
+
+        priv->cx_data = malloc(priv->cx_data_sz);
+
+        if (!priv->cx_data)
+        {
+            return VPX_CODEC_MEM_ERROR;
+        }
+
+        if(mr_cfg)
+            ctx->priv->enc.total_encoders   = mr_cfg->mr_total_resolutions;
+        else
+            ctx->priv->enc.total_encoders   = 1;
+
+        once(vp8_initialize_enc);
+
+        res = validate_config(priv, &priv->cfg, &priv->vp8_cfg, 0);
+
+        if (!res)
+        {
+            set_vp8e_config(&priv->oxcf, priv->cfg, priv->vp8_cfg, mr_cfg);
+            priv->cpi = vp8_create_compressor(&priv->oxcf);
+            if (!priv->cpi)
+                res = VPX_CODEC_MEM_ERROR;
+        }
+    }
+
+    return res;
+}
+
+static vpx_codec_err_t vp8e_destroy(vpx_codec_alg_priv_t *ctx)
+{
+#if CONFIG_MULTI_RES_ENCODING
+    /* Free multi-encoder shared memory */
+    if (ctx->oxcf.mr_total_resolutions > 0 && (ctx->oxcf.mr_encoder_id == ctx->oxcf.mr_total_resolutions-1))
+    {
+        LOWER_RES_FRAME_INFO *shared_mem_loc = (LOWER_RES_FRAME_INFO *)ctx->oxcf.mr_low_res_mode_info;
+        free(shared_mem_loc->mb_info);
+        free(ctx->oxcf.mr_low_res_mode_info);
+    }
+#endif
+
+    free(ctx->cx_data);
+    vp8_remove_compressor(&ctx->cpi);
+    vpx_free(ctx);
+    return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t image2yuvconfig(const vpx_image_t   *img,
+                                       YV12_BUFFER_CONFIG  *yv12)
+{
+    const int y_w = img->d_w;
+    const int y_h = img->d_h;
+    const int uv_w = (img->d_w + 1) / 2;
+    const int uv_h = (img->d_h + 1) / 2;
+    vpx_codec_err_t        res = VPX_CODEC_OK;
+    yv12->y_buffer = img->planes[VPX_PLANE_Y];
+    yv12->u_buffer = img->planes[VPX_PLANE_U];
+    yv12->v_buffer = img->planes[VPX_PLANE_V];
+
+    yv12->y_crop_width  = y_w;
+    yv12->y_crop_height = y_h;
+    yv12->y_width  = y_w;
+    yv12->y_height = y_h;
+    yv12->uv_crop_width = uv_w;
+    yv12->uv_crop_height = uv_h;
+    yv12->uv_width = uv_w;
+    yv12->uv_height = uv_h;
+
+    yv12->y_stride = img->stride[VPX_PLANE_Y];
+    yv12->uv_stride = img->stride[VPX_PLANE_U];
+
+    yv12->border  = (img->stride[VPX_PLANE_Y] - img->w) / 2;
+    return res;
+}
+
+static void pick_quickcompress_mode(vpx_codec_alg_priv_t  *ctx,
+                                    unsigned long          duration,
+                                    unsigned long          deadline)
+{
+    int new_qc;
+
+#if !(CONFIG_REALTIME_ONLY)
+    /* Use best quality mode if no deadline is given. */
+    new_qc = MODE_BESTQUALITY;
+
+    if (deadline)
+    {
+        uint64_t     duration_us;
+
+        /* Convert duration parameter from stream timebase to microseconds */
+        duration_us = (uint64_t)duration * 1000000
+                      * (uint64_t)ctx->cfg.g_timebase.num
+                      / (uint64_t)ctx->cfg.g_timebase.den;
+
+        /* If the deadline is more that the duration this frame is to be shown,
+         * use good quality mode. Otherwise use realtime mode.
+         */
+        new_qc = (deadline > duration_us) ? MODE_GOODQUALITY : MODE_REALTIME;
+    }
+
+#else
+    (void)duration;
+    new_qc = MODE_REALTIME;
+#endif
+
+    if (deadline == VPX_DL_REALTIME)
+        new_qc = MODE_REALTIME;
+    else if (ctx->cfg.g_pass == VPX_RC_FIRST_PASS)
+        new_qc = MODE_FIRSTPASS;
+    else if (ctx->cfg.g_pass == VPX_RC_LAST_PASS)
+        new_qc = (new_qc == MODE_BESTQUALITY)
+                 ? MODE_SECONDPASS_BEST
+                 : MODE_SECONDPASS;
+
+    if (ctx->oxcf.Mode != new_qc)
+    {
+        ctx->oxcf.Mode = new_qc;
+        vp8_change_config(ctx->cpi, &ctx->oxcf);
+    }
+}
+
+static vpx_codec_err_t set_reference_and_update(vpx_codec_alg_priv_t *ctx,
+                                                int flags)
+{
+
+    /* Handle Flags */
+    if (((flags & VP8_EFLAG_NO_UPD_GF) && (flags & VP8_EFLAG_FORCE_GF))
+        || ((flags & VP8_EFLAG_NO_UPD_ARF) && (flags & VP8_EFLAG_FORCE_ARF)))
+    {
+        ctx->base.err_detail = "Conflicting flags.";
+        return VPX_CODEC_INVALID_PARAM;
+    }
+
+    if (flags & (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF
+                 | VP8_EFLAG_NO_REF_ARF))
+    {
+        int ref = 7;
+
+        if (flags & VP8_EFLAG_NO_REF_LAST)
+            ref ^= VP8_LAST_FRAME;
+
+        if (flags & VP8_EFLAG_NO_REF_GF)
+            ref ^= VP8_GOLD_FRAME;
+
+        if (flags & VP8_EFLAG_NO_REF_ARF)
+            ref ^= VP8_ALTR_FRAME;
+
+        vp8_use_as_reference(ctx->cpi, ref);
+    }
+
+    if (flags & (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF
+                 | VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_FORCE_GF
+                 | VP8_EFLAG_FORCE_ARF))
+    {
+        int upd = 7;
+
+        if (flags & VP8_EFLAG_NO_UPD_LAST)
+            upd ^= VP8_LAST_FRAME;
+
+        if (flags & VP8_EFLAG_NO_UPD_GF)
+            upd ^= VP8_GOLD_FRAME;
+
+        if (flags & VP8_EFLAG_NO_UPD_ARF)
+            upd ^= VP8_ALTR_FRAME;
+
+        vp8_update_reference(ctx->cpi, upd);
+    }
+
+    if (flags & VP8_EFLAG_NO_UPD_ENTROPY)
+    {
+        vp8_update_entropy(ctx->cpi, 0);
+    }
+
+    return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t vp8e_encode(vpx_codec_alg_priv_t  *ctx,
+                                   const vpx_image_t     *img,
+                                   vpx_codec_pts_t        pts,
+                                   unsigned long          duration,
+                                   vpx_enc_frame_flags_t  flags,
+                                   unsigned long          deadline)
+{
+    vpx_codec_err_t res = VPX_CODEC_OK;
+
+    if (!ctx->cfg.rc_target_bitrate)
+        return res;
+
+    if (img)
+        res = validate_img(ctx, img);
+
+    if (!res)
+        res = validate_config(ctx, &ctx->cfg, &ctx->vp8_cfg, 1);
+
+    pick_quickcompress_mode(ctx, duration, deadline);
+    vpx_codec_pkt_list_init(&ctx->pkt_list);
+
+    // If no flags are set in the encode call, then use the frame flags as
+    // defined via the control function: vp8e_set_frame_flags.
+    if (!flags) {
+        flags = ctx->control_frame_flags;
+    }
+    ctx->control_frame_flags = 0;
+
+    if (!res)
+        res = set_reference_and_update(ctx, flags);
+
+    /* Handle fixed keyframe intervals */
+    if (ctx->cfg.kf_mode == VPX_KF_AUTO
+        && ctx->cfg.kf_min_dist == ctx->cfg.kf_max_dist)
+    {
+        if (++ctx->fixed_kf_cntr > ctx->cfg.kf_min_dist)
+        {
+            flags |= VPX_EFLAG_FORCE_KF;
+            ctx->fixed_kf_cntr = 1;
+        }
+    }
+
+    /* Initialize the encoder instance on the first frame*/
+    if (!res && ctx->cpi)
+    {
+        unsigned int lib_flags;
+        YV12_BUFFER_CONFIG sd;
+        int64_t dst_time_stamp, dst_end_time_stamp;
+        unsigned long size, cx_data_sz;
+        unsigned char *cx_data;
+        unsigned char *cx_data_end;
+        int comp_data_state = 0;
+
+        /* Set up internal flags */
+        if (ctx->base.init_flags & VPX_CODEC_USE_PSNR)
+            ((VP8_COMP *)ctx->cpi)->b_calculate_psnr = 1;
+
+        if (ctx->base.init_flags & VPX_CODEC_USE_OUTPUT_PARTITION)
+            ((VP8_COMP *)ctx->cpi)->output_partition = 1;
+
+        /* Convert API flags to internal codec lib flags */
+        lib_flags = (flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
+
+        /* vp8 use 10,000,000 ticks/second as time stamp */
+        dst_time_stamp    = pts * 10000000 * ctx->cfg.g_timebase.num / ctx->cfg.g_timebase.den;
+        dst_end_time_stamp = (pts + duration) * 10000000 * ctx->cfg.g_timebase.num / ctx->cfg.g_timebase.den;
+
+        if (img != NULL)
+        {
+            res = image2yuvconfig(img, &sd);
+
+            if (vp8_receive_raw_frame(ctx->cpi, ctx->next_frame_flag | lib_flags,
+                                      &sd, dst_time_stamp, dst_end_time_stamp))
+            {
+                VP8_COMP *cpi = (VP8_COMP *)ctx->cpi;
+                res = update_error_state(ctx, &cpi->common.error);
+            }
+
+            /* reset for next frame */
+            ctx->next_frame_flag = 0;
+        }
+
+        cx_data = ctx->cx_data;
+        cx_data_sz = ctx->cx_data_sz;
+        cx_data_end = ctx->cx_data + cx_data_sz;
+        lib_flags = 0;
+
+        while (cx_data_sz >= ctx->cx_data_sz / 2)
+        {
+            comp_data_state = vp8_get_compressed_data(ctx->cpi,
+                                                  &lib_flags,
+                                                  &size,
+                                                  cx_data,
+                                                  cx_data_end,
+                                                  &dst_time_stamp,
+                                                  &dst_end_time_stamp,
+                                                  !img);
+
+            if(comp_data_state == VPX_CODEC_CORRUPT_FRAME)
+                return VPX_CODEC_CORRUPT_FRAME;
+            else if(comp_data_state == -1)
+                break;
+
+            if (size)
+            {
+                vpx_codec_pts_t    round, delta;
+                vpx_codec_cx_pkt_t pkt;
+                VP8_COMP *cpi = (VP8_COMP *)ctx->cpi;
+
+                /* Add the frame packet to the list of returned packets. */
+                round = (vpx_codec_pts_t)10000000
+                        * ctx->cfg.g_timebase.num / 2 - 1;
+                delta = (dst_end_time_stamp - dst_time_stamp);
+                pkt.kind = VPX_CODEC_CX_FRAME_PKT;
+                pkt.data.frame.pts =
+                    (dst_time_stamp * ctx->cfg.g_timebase.den + round)
+                    / ctx->cfg.g_timebase.num / 10000000;
+                pkt.data.frame.duration = (unsigned long)
+                    ((delta * ctx->cfg.g_timebase.den + round)
+                    / ctx->cfg.g_timebase.num / 10000000);
+                pkt.data.frame.flags = lib_flags << 16;
+
+                if (lib_flags & FRAMEFLAGS_KEY)
+                    pkt.data.frame.flags |= VPX_FRAME_IS_KEY;
+
+                if (!cpi->common.show_frame)
+                {
+                    pkt.data.frame.flags |= VPX_FRAME_IS_INVISIBLE;
+
+                    /* This timestamp should be as close as possible to the
+                     * prior PTS so that if a decoder uses pts to schedule when
+                     * to do this, we start right after last frame was decoded.
+                     * Invisible frames have no duration.
+                     */
+                    pkt.data.frame.pts = ((cpi->last_time_stamp_seen
+                        * ctx->cfg.g_timebase.den + round)
+                        / ctx->cfg.g_timebase.num / 10000000) + 1;
+                    pkt.data.frame.duration = 0;
+                }
+
+                if (cpi->droppable)
+                    pkt.data.frame.flags |= VPX_FRAME_IS_DROPPABLE;
+
+                if (cpi->output_partition)
+                {
+                    int i;
+                    const int num_partitions =
+                            (1 << cpi->common.multi_token_partition) + 1;
+
+                    pkt.data.frame.flags |= VPX_FRAME_IS_FRAGMENT;
+
+                    for (i = 0; i < num_partitions; ++i)
+                    {
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+                        pkt.data.frame.buf = cpi->partition_d[i];
+#else
+                        pkt.data.frame.buf = cx_data;
+                        cx_data += cpi->partition_sz[i];
+                        cx_data_sz -= cpi->partition_sz[i];
+#endif
+                        pkt.data.frame.sz = cpi->partition_sz[i];
+                        pkt.data.frame.partition_id = i;
+                        /* don't set the fragment bit for the last partition */
+                        if (i == (num_partitions - 1))
+                            pkt.data.frame.flags &= ~VPX_FRAME_IS_FRAGMENT;
+                        vpx_codec_pkt_list_add(&ctx->pkt_list.head, &pkt);
+                    }
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+                    /* In lagged mode the encoder can buffer multiple frames.
+                     * We don't want this in partitioned output because
+                     * partitions are spread all over the output buffer.
+                     * So, force an exit!
+                     */
+                    cx_data_sz -= ctx->cx_data_sz / 2;
+#endif
+                }
+                else
+                {
+                    pkt.data.frame.buf = cx_data;
+                    pkt.data.frame.sz  = size;
+                    pkt.data.frame.partition_id = -1;
+                    vpx_codec_pkt_list_add(&ctx->pkt_list.head, &pkt);
+                    cx_data += size;
+                    cx_data_sz -= size;
+                }
+            }
+        }
+    }
+
+    return res;
+}
+
+
+static const vpx_codec_cx_pkt_t *vp8e_get_cxdata(vpx_codec_alg_priv_t  *ctx,
+        vpx_codec_iter_t      *iter)
+{
+    return vpx_codec_pkt_list_get(&ctx->pkt_list.head, iter);
+}
+
+static vpx_codec_err_t vp8e_set_reference(vpx_codec_alg_priv_t *ctx,
+                                          va_list args)
+{
+    vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
+
+    if (data)
+    {
+        vpx_ref_frame_t *frame = (vpx_ref_frame_t *)data;
+        YV12_BUFFER_CONFIG sd;
+
+        image2yuvconfig(&frame->img, &sd);
+        vp8_set_reference(ctx->cpi, frame->frame_type, &sd);
+        return VPX_CODEC_OK;
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+
+}
+
+static vpx_codec_err_t vp8e_get_reference(vpx_codec_alg_priv_t *ctx,
+                                          va_list args)
+{
+
+    vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
+
+    if (data)
+    {
+        vpx_ref_frame_t *frame = (vpx_ref_frame_t *)data;
+        YV12_BUFFER_CONFIG sd;
+
+        image2yuvconfig(&frame->img, &sd);
+        vp8_get_reference(ctx->cpi, frame->frame_type, &sd);
+        return VPX_CODEC_OK;
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+}
+
+static vpx_codec_err_t vp8e_set_previewpp(vpx_codec_alg_priv_t *ctx,
+                                          va_list args)
+{
+#if CONFIG_POSTPROC
+    vp8_postproc_cfg_t *data = va_arg(args, vp8_postproc_cfg_t *);
+
+    if (data)
+    {
+        ctx->preview_ppcfg = *((vp8_postproc_cfg_t *)data);
+        return VPX_CODEC_OK;
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+#else
+    (void)ctx;
+    (void)args;
+    return VPX_CODEC_INCAPABLE;
+#endif
+}
+
+
+static vpx_image_t *vp8e_get_preview(vpx_codec_alg_priv_t *ctx)
+{
+
+    YV12_BUFFER_CONFIG sd;
+    vp8_ppflags_t flags;
+    vp8_zero(flags);
+
+    if (ctx->preview_ppcfg.post_proc_flag)
+    {
+        flags.post_proc_flag        = ctx->preview_ppcfg.post_proc_flag;
+        flags.deblocking_level      = ctx->preview_ppcfg.deblocking_level;
+        flags.noise_level           = ctx->preview_ppcfg.noise_level;
+    }
+
+    if (0 == vp8_get_preview_raw_frame(ctx->cpi, &sd, &flags))
+    {
+
+        /*
+        vpx_img_wrap(&ctx->preview_img, VPX_IMG_FMT_YV12,
+            sd.y_width + 2*VP8BORDERINPIXELS,
+            sd.y_height + 2*VP8BORDERINPIXELS,
+            1,
+            sd.buffer_alloc);
+        vpx_img_set_rect(&ctx->preview_img,
+            VP8BORDERINPIXELS, VP8BORDERINPIXELS,
+            sd.y_width, sd.y_height);
+            */
+
+        ctx->preview_img.bps = 12;
+        ctx->preview_img.planes[VPX_PLANE_Y] = sd.y_buffer;
+        ctx->preview_img.planes[VPX_PLANE_U] = sd.u_buffer;
+        ctx->preview_img.planes[VPX_PLANE_V] = sd.v_buffer;
+
+        ctx->preview_img.fmt = VPX_IMG_FMT_I420;
+        ctx->preview_img.x_chroma_shift = 1;
+        ctx->preview_img.y_chroma_shift = 1;
+
+        ctx->preview_img.d_w = sd.y_width;
+        ctx->preview_img.d_h = sd.y_height;
+        ctx->preview_img.stride[VPX_PLANE_Y] = sd.y_stride;
+        ctx->preview_img.stride[VPX_PLANE_U] = sd.uv_stride;
+        ctx->preview_img.stride[VPX_PLANE_V] = sd.uv_stride;
+        ctx->preview_img.w   = sd.y_width;
+        ctx->preview_img.h   = sd.y_height;
+
+        return &ctx->preview_img;
+    }
+    else
+        return NULL;
+}
+
+static vpx_codec_err_t vp8e_set_frame_flags(vpx_codec_alg_priv_t *ctx,
+                                            va_list args)
+{
+    int frame_flags = va_arg(args, int);
+    ctx->control_frame_flags = frame_flags;
+    return set_reference_and_update(ctx, frame_flags);
+}
+
+static vpx_codec_err_t vp8e_set_temporal_layer_id(vpx_codec_alg_priv_t *ctx,
+                                                  va_list args)
+{
+    int layer_id = va_arg(args, int);
+    if (layer_id < 0 || layer_id >= (int)ctx->cfg.ts_number_layers) {
+      return VPX_CODEC_INVALID_PARAM;
+    }
+    ctx->cpi->temporal_layer_id = layer_id;
+    return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t vp8e_set_roi_map(vpx_codec_alg_priv_t *ctx,
+                                        va_list args)
+{
+    vpx_roi_map_t *data = va_arg(args, vpx_roi_map_t *);
+
+    if (data)
+    {
+        vpx_roi_map_t *roi = (vpx_roi_map_t *)data;
+
+        if (!vp8_set_roimap(ctx->cpi, roi->roi_map, roi->rows, roi->cols, roi->delta_q, roi->delta_lf, roi->static_threshold))
+            return VPX_CODEC_OK;
+        else
+            return VPX_CODEC_INVALID_PARAM;
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+}
+
+
+static vpx_codec_err_t vp8e_set_activemap(vpx_codec_alg_priv_t *ctx,
+                                          va_list args)
+{
+    vpx_active_map_t *data = va_arg(args, vpx_active_map_t *);
+
+    if (data)
+    {
+
+        vpx_active_map_t *map = (vpx_active_map_t *)data;
+
+        if (!vp8_set_active_map(ctx->cpi, map->active_map, map->rows, map->cols))
+            return VPX_CODEC_OK;
+        else
+            return VPX_CODEC_INVALID_PARAM;
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+}
+
+static vpx_codec_err_t vp8e_set_scalemode(vpx_codec_alg_priv_t *ctx,
+                                          va_list args)
+{
+
+    vpx_scaling_mode_t *data =  va_arg(args, vpx_scaling_mode_t *);
+
+    if (data)
+    {
+        int res;
+        vpx_scaling_mode_t scalemode = *(vpx_scaling_mode_t *)data ;
+        res = vp8_set_internal_size(ctx->cpi,
+                                    (VPX_SCALING)scalemode.h_scaling_mode,
+                                    (VPX_SCALING)scalemode.v_scaling_mode);
+
+        if (!res)
+        {
+            /*force next frame a key frame to effect scaling mode */
+            ctx->next_frame_flag |= FRAMEFLAGS_KEY;
+            return VPX_CODEC_OK;
+        }
+        else
+            return VPX_CODEC_INVALID_PARAM;
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+}
+
+
+static vpx_codec_ctrl_fn_map_t vp8e_ctf_maps[] =
+{
+    {VP8_SET_REFERENCE,                 vp8e_set_reference},
+    {VP8_COPY_REFERENCE,                vp8e_get_reference},
+    {VP8_SET_POSTPROC,                  vp8e_set_previewpp},
+    {VP8E_SET_FRAME_FLAGS,              vp8e_set_frame_flags},
+    {VP8E_SET_TEMPORAL_LAYER_ID,        vp8e_set_temporal_layer_id},
+    {VP8E_SET_ROI_MAP,                  vp8e_set_roi_map},
+    {VP8E_SET_ACTIVEMAP,                vp8e_set_activemap},
+    {VP8E_SET_SCALEMODE,                vp8e_set_scalemode},
+    {VP8E_SET_CPUUSED,                  set_cpu_used},
+    {VP8E_SET_NOISE_SENSITIVITY,        set_noise_sensitivity},
+    {VP8E_SET_ENABLEAUTOALTREF,         set_enable_auto_alt_ref},
+    {VP8E_SET_SHARPNESS,                set_sharpness},
+    {VP8E_SET_STATIC_THRESHOLD,         set_static_thresh},
+    {VP8E_SET_TOKEN_PARTITIONS,         set_token_partitions},
+    {VP8E_GET_LAST_QUANTIZER,           get_quantizer},
+    {VP8E_GET_LAST_QUANTIZER_64,        get_quantizer64},
+    {VP8E_SET_ARNR_MAXFRAMES,           set_arnr_max_frames},
+    {VP8E_SET_ARNR_STRENGTH ,           set_arnr_strength},
+    {VP8E_SET_ARNR_TYPE     ,           set_arnr_type},
+    {VP8E_SET_TUNING,                   set_tuning},
+    {VP8E_SET_CQ_LEVEL,                 set_cq_level},
+    {VP8E_SET_MAX_INTRA_BITRATE_PCT,    set_rc_max_intra_bitrate_pct},
+    {VP8E_SET_SCREEN_CONTENT_MODE,      set_screen_content_mode},
+    { -1, NULL},
+};
+
+static vpx_codec_enc_cfg_map_t vp8e_usage_cfg_map[] =
+{
+    {
+    0,
+    {
+        0,                  /* g_usage */
+        0,                  /* g_threads */
+        0,                  /* g_profile */
+
+        320,                /* g_width */
+        240,                /* g_height */
+        VPX_BITS_8,         /* g_bit_depth */
+        8,                  /* g_input_bit_depth */
+
+        {1, 30},            /* g_timebase */
+
+        0,                  /* g_error_resilient */
+
+        VPX_RC_ONE_PASS,    /* g_pass */
+
+        0,                  /* g_lag_in_frames */
+
+        0,                  /* rc_dropframe_thresh */
+        0,                  /* rc_resize_allowed */
+        1,                  /* rc_scaled_width */
+        1,                  /* rc_scaled_height */
+        60,                 /* rc_resize_down_thresold */
+        30,                 /* rc_resize_up_thresold */
+
+        VPX_VBR,            /* rc_end_usage */
+        {NULL, 0},          /* rc_twopass_stats_in */
+        {NULL, 0},          /* rc_firstpass_mb_stats_in */
+        256,                /* rc_target_bandwidth */
+        4,                  /* rc_min_quantizer */
+        63,                 /* rc_max_quantizer */
+        100,                /* rc_undershoot_pct */
+        100,                /* rc_overshoot_pct */
+
+        6000,               /* rc_max_buffer_size */
+        4000,               /* rc_buffer_initial_size; */
+        5000,               /* rc_buffer_optimal_size; */
+
+        50,                 /* rc_two_pass_vbrbias  */
+        0,                  /* rc_two_pass_vbrmin_section */
+        400,                /* rc_two_pass_vbrmax_section */
+
+        /* keyframing settings (kf) */
+        VPX_KF_AUTO,        /* g_kfmode*/
+        0,                  /* kf_min_dist */
+        128,                /* kf_max_dist */
+
+        VPX_SS_DEFAULT_LAYERS, /* ss_number_layers */
+        {0},
+        {0},                /* ss_target_bitrate */
+        1,                  /* ts_number_layers */
+        {0},                /* ts_target_bitrate */
+        {0},                /* ts_rate_decimator */
+        0,                  /* ts_periodicity */
+        {0},                /* ts_layer_id */
+        {0},                /* layer_target_bitrate */
+        0                   /* temporal_layering_mode */
+    }},
+};
+
+
+#ifndef VERSION_STRING
+#define VERSION_STRING
+#endif
+CODEC_INTERFACE(vpx_codec_vp8_cx) =
+{
+    "WebM Project VP8 Encoder" VERSION_STRING,
+    VPX_CODEC_INTERNAL_ABI_VERSION,
+    VPX_CODEC_CAP_ENCODER | VPX_CODEC_CAP_PSNR |
+    VPX_CODEC_CAP_OUTPUT_PARTITION,
+    /* vpx_codec_caps_t          caps; */
+    vp8e_init,          /* vpx_codec_init_fn_t       init; */
+    vp8e_destroy,       /* vpx_codec_destroy_fn_t    destroy; */
+    vp8e_ctf_maps,      /* vpx_codec_ctrl_fn_map_t  *ctrl_maps; */
+    {
+        NULL,    /* vpx_codec_peek_si_fn_t    peek_si; */
+        NULL,    /* vpx_codec_get_si_fn_t     get_si; */
+        NULL,    /* vpx_codec_decode_fn_t     decode; */
+        NULL,    /* vpx_codec_frame_get_fn_t  frame_get; */
+        NULL,    /* vpx_codec_set_fb_fn_t     set_fb_fn; */
+    },
+    {
+        1,                  /* 1 cfg map */
+        vp8e_usage_cfg_map, /* vpx_codec_enc_cfg_map_t    cfg_maps; */
+        vp8e_encode,        /* vpx_codec_encode_fn_t      encode; */
+        vp8e_get_cxdata,    /* vpx_codec_get_cx_data_fn_t   get_cx_data; */
+        vp8e_set_config,
+        NULL,
+        vp8e_get_preview,
+        vp8e_mr_alloc_mem,
+    } /* encoder functions */
+};
diff --git a/libvpx/libvpx/vp8/vp8_dx_iface.c b/libvpx/libvpx/vp8/vp8_dx_iface.c
new file mode 100644
index 0000000..fc9288d
--- /dev/null
+++ b/libvpx/libvpx/vp8/vp8_dx_iface.c
@@ -0,0 +1,828 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+#include "./vp8_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+#include "vpx/vpx_decoder.h"
+#include "vpx/vp8dx.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "vpx_version.h"
+#include "common/alloccommon.h"
+#include "common/common.h"
+#include "common/onyxd.h"
+#include "decoder/onyxd_int.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#if CONFIG_ERROR_CONCEALMENT
+#include "decoder/error_concealment.h"
+#endif
+#include "decoder/decoderthreading.h"
+
+#define VP8_CAP_POSTPROC (CONFIG_POSTPROC ? VPX_CODEC_CAP_POSTPROC : 0)
+#define VP8_CAP_ERROR_CONCEALMENT (CONFIG_ERROR_CONCEALMENT ? \
+                                    VPX_CODEC_CAP_ERROR_CONCEALMENT : 0)
+
+typedef vpx_codec_stream_info_t  vp8_stream_info_t;
+
+/* Structures for handling memory allocations */
+typedef enum
+{
+    VP8_SEG_ALG_PRIV     = 256,
+    VP8_SEG_MAX
+} mem_seg_id_t;
+#define NELEMENTS(x) ((int)(sizeof(x)/sizeof(x[0])))
+
+struct vpx_codec_alg_priv
+{
+    vpx_codec_priv_t        base;
+    vpx_codec_dec_cfg_t     cfg;
+    vp8_stream_info_t       si;
+    int                     decoder_init;
+    int                     postproc_cfg_set;
+    vp8_postproc_cfg_t      postproc_cfg;
+#if CONFIG_POSTPROC_VISUALIZER
+    unsigned int            dbg_postproc_flag;
+    int                     dbg_color_ref_frame_flag;
+    int                     dbg_color_mb_modes_flag;
+    int                     dbg_color_b_modes_flag;
+    int                     dbg_display_mv_flag;
+#endif
+    vpx_decrypt_cb          decrypt_cb;
+    void                    *decrypt_state;
+    vpx_image_t             img;
+    int                     img_setup;
+    struct frame_buffers    yv12_frame_buffers;
+    void                    *user_priv;
+    FRAGMENT_DATA           fragments;
+};
+
+static int vp8_init_ctx(vpx_codec_ctx_t *ctx)
+{
+    vpx_codec_alg_priv_t *priv =
+        (vpx_codec_alg_priv_t *)vpx_calloc(1, sizeof(*priv));
+    if (!priv) return 1;
+
+    ctx->priv = (vpx_codec_priv_t *)priv;
+    ctx->priv->init_flags = ctx->init_flags;
+
+    priv->si.sz = sizeof(priv->si);
+    priv->decrypt_cb = NULL;
+    priv->decrypt_state = NULL;
+
+    if (ctx->config.dec)
+    {
+        /* Update the reference to the config structure to an internal copy. */
+        priv->cfg = *ctx->config.dec;
+        ctx->config.dec = &priv->cfg;
+    }
+
+    return 0;
+}
+
+static vpx_codec_err_t vp8_init(vpx_codec_ctx_t *ctx,
+                                vpx_codec_priv_enc_mr_cfg_t *data)
+{
+    vpx_codec_err_t res = VPX_CODEC_OK;
+    vpx_codec_alg_priv_t *priv = NULL;
+    (void) data;
+
+    vp8_rtcd();
+    vpx_dsp_rtcd();
+    vpx_scale_rtcd();
+
+    /* This function only allocates space for the vpx_codec_alg_priv_t
+     * structure. More memory may be required at the time the stream
+     * information becomes known.
+     */
+    if (!ctx->priv) {
+      if (vp8_init_ctx(ctx)) return VPX_CODEC_MEM_ERROR;
+      priv = (vpx_codec_alg_priv_t *)ctx->priv;
+
+      /* initialize number of fragments to zero */
+      priv->fragments.count = 0;
+      /* is input fragments enabled? */
+      priv->fragments.enabled =
+          (priv->base.init_flags & VPX_CODEC_USE_INPUT_FRAGMENTS);
+
+      /*post processing level initialized to do nothing */
+    } else {
+      priv = (vpx_codec_alg_priv_t *)ctx->priv;
+    }
+
+    priv->yv12_frame_buffers.use_frame_threads =
+        (ctx->priv->init_flags & VPX_CODEC_USE_FRAME_THREADING);
+
+    /* for now, disable frame threading */
+    priv->yv12_frame_buffers.use_frame_threads = 0;
+
+    if (priv->yv12_frame_buffers.use_frame_threads &&
+        ((ctx->priv->init_flags & VPX_CODEC_USE_ERROR_CONCEALMENT) ||
+         (ctx->priv->init_flags & VPX_CODEC_USE_INPUT_FRAGMENTS))) {
+      /* row-based threading, error concealment, and input fragments will
+       * not be supported when using frame-based threading */
+      res = VPX_CODEC_INVALID_PARAM;
+    }
+
+    return res;
+}
+
+static vpx_codec_err_t vp8_destroy(vpx_codec_alg_priv_t *ctx)
+{
+    vp8_remove_decoder_instances(&ctx->yv12_frame_buffers);
+
+    vpx_free(ctx);
+
+    return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t vp8_peek_si_internal(const uint8_t *data,
+                                            unsigned int data_sz,
+                                            vpx_codec_stream_info_t *si,
+                                            vpx_decrypt_cb decrypt_cb,
+                                            void *decrypt_state)
+{
+    vpx_codec_err_t res = VPX_CODEC_OK;
+
+    assert(data != NULL);
+
+    if(data + data_sz <= data)
+    {
+        res = VPX_CODEC_INVALID_PARAM;
+    }
+    else
+    {
+        /* Parse uncompresssed part of key frame header.
+         * 3 bytes:- including version, frame type and an offset
+         * 3 bytes:- sync code (0x9d, 0x01, 0x2a)
+         * 4 bytes:- including image width and height in the lowest 14 bits
+         *           of each 2-byte value.
+         */
+        uint8_t clear_buffer[10];
+        const uint8_t *clear = data;
+        if (decrypt_cb)
+        {
+            int n = VPXMIN(sizeof(clear_buffer), data_sz);
+            decrypt_cb(decrypt_state, data, clear_buffer, n);
+            clear = clear_buffer;
+        }
+        si->is_kf = 0;
+
+        if (data_sz >= 10 && !(clear[0] & 0x01))  /* I-Frame */
+        {
+            si->is_kf = 1;
+
+            /* vet via sync code */
+            if (clear[3] != 0x9d || clear[4] != 0x01 || clear[5] != 0x2a)
+                return VPX_CODEC_UNSUP_BITSTREAM;
+
+            si->w = (clear[6] | (clear[7] << 8)) & 0x3fff;
+            si->h = (clear[8] | (clear[9] << 8)) & 0x3fff;
+
+            /*printf("w=%d, h=%d\n", si->w, si->h);*/
+            if (!(si->h | si->w))
+                res = VPX_CODEC_UNSUP_BITSTREAM;
+        }
+        else
+        {
+            res = VPX_CODEC_UNSUP_BITSTREAM;
+        }
+    }
+
+    return res;
+}
+
+static vpx_codec_err_t vp8_peek_si(const uint8_t *data,
+                                   unsigned int data_sz,
+                                   vpx_codec_stream_info_t *si) {
+    return vp8_peek_si_internal(data, data_sz, si, NULL, NULL);
+}
+
+static vpx_codec_err_t vp8_get_si(vpx_codec_alg_priv_t    *ctx,
+                                  vpx_codec_stream_info_t *si)
+{
+
+    unsigned int sz;
+
+    if (si->sz >= sizeof(vp8_stream_info_t))
+        sz = sizeof(vp8_stream_info_t);
+    else
+        sz = sizeof(vpx_codec_stream_info_t);
+
+    memcpy(si, &ctx->si, sz);
+    si->sz = sz;
+
+    return VPX_CODEC_OK;
+}
+
+
+static vpx_codec_err_t
+update_error_state(vpx_codec_alg_priv_t                 *ctx,
+                   const struct vpx_internal_error_info *error)
+{
+    vpx_codec_err_t res;
+
+    if ((res = error->error_code))
+        ctx->base.err_detail = error->has_detail
+                               ? error->detail
+                               : NULL;
+
+    return res;
+}
+
+static void yuvconfig2image(vpx_image_t               *img,
+                            const YV12_BUFFER_CONFIG  *yv12,
+                            void                      *user_priv)
+{
+    /** vpx_img_wrap() doesn't allow specifying independent strides for
+      * the Y, U, and V planes, nor other alignment adjustments that
+      * might be representable by a YV12_BUFFER_CONFIG, so we just
+      * initialize all the fields.*/
+    img->fmt = VPX_IMG_FMT_I420;
+    img->w = yv12->y_stride;
+    img->h = (yv12->y_height + 2 * VP8BORDERINPIXELS + 15) & ~15;
+    img->d_w = img->r_w = yv12->y_width;
+    img->d_h = img->r_h = yv12->y_height;
+    img->x_chroma_shift = 1;
+    img->y_chroma_shift = 1;
+    img->planes[VPX_PLANE_Y] = yv12->y_buffer;
+    img->planes[VPX_PLANE_U] = yv12->u_buffer;
+    img->planes[VPX_PLANE_V] = yv12->v_buffer;
+    img->planes[VPX_PLANE_ALPHA] = NULL;
+    img->stride[VPX_PLANE_Y] = yv12->y_stride;
+    img->stride[VPX_PLANE_U] = yv12->uv_stride;
+    img->stride[VPX_PLANE_V] = yv12->uv_stride;
+    img->stride[VPX_PLANE_ALPHA] = yv12->y_stride;
+    img->bit_depth = 8;
+    img->bps = 12;
+    img->user_priv = user_priv;
+    img->img_data = yv12->buffer_alloc;
+    img->img_data_owner = 0;
+    img->self_allocd = 0;
+}
+
+static int
+update_fragments(vpx_codec_alg_priv_t  *ctx,
+                 const uint8_t         *data,
+                 unsigned int           data_sz,
+                 vpx_codec_err_t       *res)
+{
+    *res = VPX_CODEC_OK;
+
+    if (ctx->fragments.count == 0)
+    {
+        /* New frame, reset fragment pointers and sizes */
+        memset((void*)ctx->fragments.ptrs, 0, sizeof(ctx->fragments.ptrs));
+        memset(ctx->fragments.sizes, 0, sizeof(ctx->fragments.sizes));
+    }
+    if (ctx->fragments.enabled && !(data == NULL && data_sz == 0))
+    {
+        /* Store a pointer to this fragment and return. We haven't
+         * received the complete frame yet, so we will wait with decoding.
+         */
+        ctx->fragments.ptrs[ctx->fragments.count] = data;
+        ctx->fragments.sizes[ctx->fragments.count] = data_sz;
+        ctx->fragments.count++;
+        if (ctx->fragments.count > (1 << EIGHT_PARTITION) + 1)
+        {
+            ctx->fragments.count = 0;
+            *res = VPX_CODEC_INVALID_PARAM;
+            return -1;
+        }
+        return 0;
+    }
+
+    if (!ctx->fragments.enabled && (data == NULL && data_sz == 0))
+    {
+        return 0;
+    }
+
+    if (!ctx->fragments.enabled)
+    {
+        ctx->fragments.ptrs[0] = data;
+        ctx->fragments.sizes[0] = data_sz;
+        ctx->fragments.count = 1;
+    }
+
+    return 1;
+}
+
+static vpx_codec_err_t vp8_decode(vpx_codec_alg_priv_t  *ctx,
+                                  const uint8_t         *data,
+                                  unsigned int            data_sz,
+                                  void                    *user_priv,
+                                  long                    deadline)
+{
+    vpx_codec_err_t res = VPX_CODEC_OK;
+    unsigned int resolution_change = 0;
+    unsigned int w, h;
+
+    if (!ctx->fragments.enabled && (data == NULL && data_sz == 0))
+    {
+        return 0;
+    }
+
+    /* Update the input fragment data */
+    if(update_fragments(ctx, data, data_sz, &res) <= 0)
+        return res;
+
+    /* Determine the stream parameters. Note that we rely on peek_si to
+     * validate that we have a buffer that does not wrap around the top
+     * of the heap.
+     */
+    w = ctx->si.w;
+    h = ctx->si.h;
+
+    res = vp8_peek_si_internal(ctx->fragments.ptrs[0], ctx->fragments.sizes[0],
+                               &ctx->si, ctx->decrypt_cb, ctx->decrypt_state);
+
+    if((res == VPX_CODEC_UNSUP_BITSTREAM) && !ctx->si.is_kf)
+    {
+        /* the peek function returns an error for non keyframes, however for
+         * this case, it is not an error */
+        res = VPX_CODEC_OK;
+    }
+
+    if(!ctx->decoder_init && !ctx->si.is_kf)
+        res = VPX_CODEC_UNSUP_BITSTREAM;
+
+    if ((ctx->si.h != h) || (ctx->si.w != w))
+        resolution_change = 1;
+
+    /* Initialize the decoder instance on the first frame*/
+    if (!res && !ctx->decoder_init)
+    {
+      VP8D_CONFIG oxcf;
+
+      oxcf.Width = ctx->si.w;
+      oxcf.Height = ctx->si.h;
+      oxcf.Version = 9;
+      oxcf.postprocess = 0;
+      oxcf.max_threads = ctx->cfg.threads;
+      oxcf.error_concealment =
+          (ctx->base.init_flags & VPX_CODEC_USE_ERROR_CONCEALMENT);
+
+      /* If postprocessing was enabled by the application and a
+       * configuration has not been provided, default it.
+       */
+       if (!ctx->postproc_cfg_set
+           && (ctx->base.init_flags & VPX_CODEC_USE_POSTPROC)) {
+         ctx->postproc_cfg.post_proc_flag =
+             VP8_DEBLOCK | VP8_DEMACROBLOCK | VP8_MFQE;
+         ctx->postproc_cfg.deblocking_level = 4;
+         ctx->postproc_cfg.noise_level = 0;
+       }
+
+       res = vp8_create_decoder_instances(&ctx->yv12_frame_buffers, &oxcf);
+       ctx->decoder_init = 1;
+    }
+
+    /* Set these even if already initialized.  The caller may have changed the
+     * decrypt config between frames.
+     */
+    if (ctx->decoder_init) {
+      ctx->yv12_frame_buffers.pbi[0]->decrypt_cb = ctx->decrypt_cb;
+      ctx->yv12_frame_buffers.pbi[0]->decrypt_state = ctx->decrypt_state;
+    }
+
+    if (!res)
+    {
+        VP8D_COMP *pbi = ctx->yv12_frame_buffers.pbi[0];
+        if (resolution_change)
+        {
+            VP8_COMMON *const pc = & pbi->common;
+            MACROBLOCKD *const xd  = & pbi->mb;
+#if CONFIG_MULTITHREAD
+            int i;
+#endif
+            pc->Width = ctx->si.w;
+            pc->Height = ctx->si.h;
+            {
+                int prev_mb_rows = pc->mb_rows;
+
+                if (setjmp(pbi->common.error.jmp))
+                {
+                    pbi->common.error.setjmp = 0;
+                    vp8_clear_system_state();
+                    /* same return value as used in vp8dx_receive_compressed_data */
+                    return -1;
+                }
+
+                pbi->common.error.setjmp = 1;
+
+                if (pc->Width <= 0)
+                {
+                    pc->Width = w;
+                    vpx_internal_error(&pc->error, VPX_CODEC_CORRUPT_FRAME,
+                                       "Invalid frame width");
+                }
+
+                if (pc->Height <= 0)
+                {
+                    pc->Height = h;
+                    vpx_internal_error(&pc->error, VPX_CODEC_CORRUPT_FRAME,
+                                       "Invalid frame height");
+                }
+
+                if (vp8_alloc_frame_buffers(pc, pc->Width, pc->Height))
+                    vpx_internal_error(&pc->error, VPX_CODEC_MEM_ERROR,
+                                       "Failed to allocate frame buffers");
+
+                xd->pre = pc->yv12_fb[pc->lst_fb_idx];
+                xd->dst = pc->yv12_fb[pc->new_fb_idx];
+
+#if CONFIG_MULTITHREAD
+                for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
+                {
+                    pbi->mb_row_di[i].mbd.dst = pc->yv12_fb[pc->new_fb_idx];
+                    vp8_build_block_doffsets(&pbi->mb_row_di[i].mbd);
+                }
+#endif
+                vp8_build_block_doffsets(&pbi->mb);
+
+                /* allocate memory for last frame MODE_INFO array */
+#if CONFIG_ERROR_CONCEALMENT
+
+                if (pbi->ec_enabled)
+                {
+                    /* old prev_mip was released by vp8_de_alloc_frame_buffers()
+                     * called in vp8_alloc_frame_buffers() */
+                    pc->prev_mip = vpx_calloc(
+                                       (pc->mb_cols + 1) * (pc->mb_rows + 1),
+                                       sizeof(MODE_INFO));
+
+                    if (!pc->prev_mip)
+                    {
+                        vp8_de_alloc_frame_buffers(pc);
+                        vpx_internal_error(&pc->error, VPX_CODEC_MEM_ERROR,
+                                           "Failed to allocate"
+                                           "last frame MODE_INFO array");
+                    }
+
+                    pc->prev_mi = pc->prev_mip + pc->mode_info_stride + 1;
+
+                    if (vp8_alloc_overlap_lists(pbi))
+                        vpx_internal_error(&pc->error, VPX_CODEC_MEM_ERROR,
+                                           "Failed to allocate overlap lists "
+                                           "for error concealment");
+                }
+
+#endif
+
+#if CONFIG_MULTITHREAD
+                if (pbi->b_multithreaded_rd)
+                    vp8mt_alloc_temp_buffers(pbi, pc->Width, prev_mb_rows);
+#else
+                (void)prev_mb_rows;
+#endif
+            }
+
+            pbi->common.error.setjmp = 0;
+
+            /* required to get past the first get_free_fb() call */
+            pbi->common.fb_idx_ref_cnt[0] = 0;
+        }
+
+        /* update the pbi fragment data */
+        pbi->fragments = ctx->fragments;
+
+        ctx->user_priv = user_priv;
+        if (vp8dx_receive_compressed_data(pbi, data_sz, data, deadline))
+        {
+            res = update_error_state(ctx, &pbi->common.error);
+        }
+
+        /* get ready for the next series of fragments */
+        ctx->fragments.count = 0;
+    }
+
+    return res;
+}
+
+static vpx_image_t *vp8_get_frame(vpx_codec_alg_priv_t  *ctx,
+                                  vpx_codec_iter_t      *iter)
+{
+    vpx_image_t *img = NULL;
+
+    /* iter acts as a flip flop, so an image is only returned on the first
+     * call to get_frame.
+     */
+    if (!(*iter) && ctx->yv12_frame_buffers.pbi[0])
+    {
+        YV12_BUFFER_CONFIG sd;
+        int64_t time_stamp = 0, time_end_stamp = 0;
+        vp8_ppflags_t flags;
+        vp8_zero(flags);
+
+        if (ctx->base.init_flags & VPX_CODEC_USE_POSTPROC)
+        {
+            flags.post_proc_flag= ctx->postproc_cfg.post_proc_flag
+#if CONFIG_POSTPROC_VISUALIZER
+
+                                | ((ctx->dbg_color_ref_frame_flag != 0) ? VP8D_DEBUG_CLR_FRM_REF_BLKS : 0)
+                                | ((ctx->dbg_color_mb_modes_flag != 0) ? VP8D_DEBUG_CLR_BLK_MODES : 0)
+                                | ((ctx->dbg_color_b_modes_flag != 0) ? VP8D_DEBUG_CLR_BLK_MODES : 0)
+                                | ((ctx->dbg_display_mv_flag != 0) ? VP8D_DEBUG_DRAW_MV : 0)
+#endif
+                                ;
+            flags.deblocking_level      = ctx->postproc_cfg.deblocking_level;
+            flags.noise_level           = ctx->postproc_cfg.noise_level;
+#if CONFIG_POSTPROC_VISUALIZER
+            flags.display_ref_frame_flag= ctx->dbg_color_ref_frame_flag;
+            flags.display_mb_modes_flag = ctx->dbg_color_mb_modes_flag;
+            flags.display_b_modes_flag  = ctx->dbg_color_b_modes_flag;
+            flags.display_mv_flag       = ctx->dbg_display_mv_flag;
+#endif
+        }
+
+        if (0 == vp8dx_get_raw_frame(ctx->yv12_frame_buffers.pbi[0], &sd,
+                                     &time_stamp, &time_end_stamp, &flags))
+        {
+            yuvconfig2image(&ctx->img, &sd, ctx->user_priv);
+
+            img = &ctx->img;
+            *iter = img;
+        }
+    }
+
+    return img;
+}
+
+static vpx_codec_err_t image2yuvconfig(const vpx_image_t   *img,
+                                       YV12_BUFFER_CONFIG  *yv12)
+{
+    const int y_w = img->d_w;
+    const int y_h = img->d_h;
+    const int uv_w = (img->d_w + 1) / 2;
+    const int uv_h = (img->d_h + 1) / 2;
+    vpx_codec_err_t        res = VPX_CODEC_OK;
+    yv12->y_buffer = img->planes[VPX_PLANE_Y];
+    yv12->u_buffer = img->planes[VPX_PLANE_U];
+    yv12->v_buffer = img->planes[VPX_PLANE_V];
+
+    yv12->y_crop_width  = y_w;
+    yv12->y_crop_height = y_h;
+    yv12->y_width  = y_w;
+    yv12->y_height = y_h;
+    yv12->uv_crop_width = uv_w;
+    yv12->uv_crop_height = uv_h;
+    yv12->uv_width = uv_w;
+    yv12->uv_height = uv_h;
+
+    yv12->y_stride = img->stride[VPX_PLANE_Y];
+    yv12->uv_stride = img->stride[VPX_PLANE_U];
+
+    yv12->border  = (img->stride[VPX_PLANE_Y] - img->d_w) / 2;
+    return res;
+}
+
+
+static vpx_codec_err_t vp8_set_reference(vpx_codec_alg_priv_t *ctx,
+                                         va_list args)
+{
+
+    vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
+
+    if (data && !ctx->yv12_frame_buffers.use_frame_threads)
+    {
+        vpx_ref_frame_t *frame = (vpx_ref_frame_t *)data;
+        YV12_BUFFER_CONFIG sd;
+
+        image2yuvconfig(&frame->img, &sd);
+
+        return vp8dx_set_reference(ctx->yv12_frame_buffers.pbi[0],
+                                   frame->frame_type, &sd);
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+
+}
+
+static vpx_codec_err_t vp8_get_reference(vpx_codec_alg_priv_t *ctx,
+                                         va_list args)
+{
+
+    vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
+
+    if (data && !ctx->yv12_frame_buffers.use_frame_threads)
+    {
+        vpx_ref_frame_t *frame = (vpx_ref_frame_t *)data;
+        YV12_BUFFER_CONFIG sd;
+
+        image2yuvconfig(&frame->img, &sd);
+
+        return vp8dx_get_reference(ctx->yv12_frame_buffers.pbi[0],
+                                   frame->frame_type, &sd);
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+
+}
+
+static vpx_codec_err_t vp8_set_postproc(vpx_codec_alg_priv_t *ctx,
+                                        va_list args)
+{
+#if CONFIG_POSTPROC
+    vp8_postproc_cfg_t *data = va_arg(args, vp8_postproc_cfg_t *);
+
+    if (data)
+    {
+        ctx->postproc_cfg_set = 1;
+        ctx->postproc_cfg = *((vp8_postproc_cfg_t *)data);
+        return VPX_CODEC_OK;
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+
+#else
+    (void)ctx;
+    (void)args;
+    return VPX_CODEC_INCAPABLE;
+#endif
+}
+
+
+static vpx_codec_err_t vp8_set_dbg_color_ref_frame(vpx_codec_alg_priv_t *ctx,
+                                                   va_list args) {
+#if CONFIG_POSTPROC_VISUALIZER && CONFIG_POSTPROC
+  ctx->dbg_color_ref_frame_flag = va_arg(args, int);
+  return VPX_CODEC_OK;
+#else
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
+#endif
+}
+
+static vpx_codec_err_t vp8_set_dbg_color_mb_modes(vpx_codec_alg_priv_t *ctx,
+                                                  va_list args) {
+#if CONFIG_POSTPROC_VISUALIZER && CONFIG_POSTPROC
+  ctx->dbg_color_mb_modes_flag = va_arg(args, int);
+  return VPX_CODEC_OK;
+#else
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
+#endif
+}
+
+static vpx_codec_err_t vp8_set_dbg_color_b_modes(vpx_codec_alg_priv_t *ctx,
+                                                 va_list args) {
+#if CONFIG_POSTPROC_VISUALIZER && CONFIG_POSTPROC
+  ctx->dbg_color_b_modes_flag = va_arg(args, int);
+  return VPX_CODEC_OK;
+#else
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
+#endif
+}
+
+static vpx_codec_err_t vp8_set_dbg_display_mv(vpx_codec_alg_priv_t *ctx,
+                                              va_list args) {
+#if CONFIG_POSTPROC_VISUALIZER && CONFIG_POSTPROC
+  ctx->dbg_display_mv_flag = va_arg(args, int);
+  return VPX_CODEC_OK;
+#else
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
+#endif
+}
+
+static vpx_codec_err_t vp8_get_last_ref_updates(vpx_codec_alg_priv_t *ctx,
+                                                va_list args)
+{
+    int *update_info = va_arg(args, int *);
+
+    if (update_info && !ctx->yv12_frame_buffers.use_frame_threads)
+    {
+        VP8D_COMP *pbi = (VP8D_COMP *)ctx->yv12_frame_buffers.pbi[0];
+
+        *update_info = pbi->common.refresh_alt_ref_frame * (int) VP8_ALTR_FRAME
+            + pbi->common.refresh_golden_frame * (int) VP8_GOLD_FRAME
+            + pbi->common.refresh_last_frame * (int) VP8_LAST_FRAME;
+
+        return VPX_CODEC_OK;
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+}
+
+extern int vp8dx_references_buffer( VP8_COMMON *oci, int ref_frame );
+static vpx_codec_err_t vp8_get_last_ref_frame(vpx_codec_alg_priv_t *ctx,
+                                              va_list args)
+{
+    int *ref_info = va_arg(args, int *);
+
+    if (ref_info && !ctx->yv12_frame_buffers.use_frame_threads)
+    {
+        VP8D_COMP *pbi = (VP8D_COMP *)ctx->yv12_frame_buffers.pbi[0];
+        VP8_COMMON *oci = &pbi->common;
+        *ref_info =
+            (vp8dx_references_buffer( oci, ALTREF_FRAME )?VP8_ALTR_FRAME:0) |
+            (vp8dx_references_buffer( oci, GOLDEN_FRAME )?VP8_GOLD_FRAME:0) |
+            (vp8dx_references_buffer( oci, LAST_FRAME )?VP8_LAST_FRAME:0);
+
+        return VPX_CODEC_OK;
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+}
+
+static vpx_codec_err_t vp8_get_frame_corrupted(vpx_codec_alg_priv_t *ctx,
+                                               va_list args)
+{
+
+    int *corrupted = va_arg(args, int *);
+    VP8D_COMP *pbi = (VP8D_COMP *)ctx->yv12_frame_buffers.pbi[0];
+
+    if (corrupted && pbi)
+    {
+        const YV12_BUFFER_CONFIG *const frame = pbi->common.frame_to_show;
+        if (frame == NULL) return VPX_CODEC_ERROR;
+        *corrupted = frame->corrupted;
+        return VPX_CODEC_OK;
+    }
+    else
+        return VPX_CODEC_INVALID_PARAM;
+
+}
+
+static vpx_codec_err_t vp8_set_decryptor(vpx_codec_alg_priv_t *ctx,
+                                         va_list args)
+{
+    vpx_decrypt_init *init = va_arg(args, vpx_decrypt_init *);
+
+    if (init)
+    {
+        ctx->decrypt_cb = init->decrypt_cb;
+        ctx->decrypt_state = init->decrypt_state;
+    }
+    else
+    {
+        ctx->decrypt_cb = NULL;
+        ctx->decrypt_state = NULL;
+    }
+    return VPX_CODEC_OK;
+}
+
+vpx_codec_ctrl_fn_map_t vp8_ctf_maps[] =
+{
+    {VP8_SET_REFERENCE,             vp8_set_reference},
+    {VP8_COPY_REFERENCE,            vp8_get_reference},
+    {VP8_SET_POSTPROC,              vp8_set_postproc},
+    {VP8_SET_DBG_COLOR_REF_FRAME,   vp8_set_dbg_color_ref_frame},
+    {VP8_SET_DBG_COLOR_MB_MODES,    vp8_set_dbg_color_mb_modes},
+    {VP8_SET_DBG_COLOR_B_MODES,     vp8_set_dbg_color_b_modes},
+    {VP8_SET_DBG_DISPLAY_MV,        vp8_set_dbg_display_mv},
+    {VP8D_GET_LAST_REF_UPDATES,     vp8_get_last_ref_updates},
+    {VP8D_GET_FRAME_CORRUPTED,      vp8_get_frame_corrupted},
+    {VP8D_GET_LAST_REF_USED,        vp8_get_last_ref_frame},
+    {VPXD_SET_DECRYPTOR,            vp8_set_decryptor},
+    { -1, NULL},
+};
+
+
+#ifndef VERSION_STRING
+#define VERSION_STRING
+#endif
+CODEC_INTERFACE(vpx_codec_vp8_dx) =
+{
+    "WebM Project VP8 Decoder" VERSION_STRING,
+    VPX_CODEC_INTERNAL_ABI_VERSION,
+    VPX_CODEC_CAP_DECODER | VP8_CAP_POSTPROC | VP8_CAP_ERROR_CONCEALMENT |
+    VPX_CODEC_CAP_INPUT_FRAGMENTS,
+    /* vpx_codec_caps_t          caps; */
+    vp8_init,         /* vpx_codec_init_fn_t       init; */
+    vp8_destroy,      /* vpx_codec_destroy_fn_t    destroy; */
+    vp8_ctf_maps,     /* vpx_codec_ctrl_fn_map_t  *ctrl_maps; */
+    {
+        vp8_peek_si,      /* vpx_codec_peek_si_fn_t    peek_si; */
+        vp8_get_si,       /* vpx_codec_get_si_fn_t     get_si; */
+        vp8_decode,       /* vpx_codec_decode_fn_t     decode; */
+        vp8_get_frame,    /* vpx_codec_frame_get_fn_t  frame_get; */
+        NULL,
+    },
+    { /* encoder functions */
+        0,
+        NULL,  /* vpx_codec_enc_cfg_map_t */
+        NULL,  /* vpx_codec_encode_fn_t */
+        NULL,  /* vpx_codec_get_cx_data_fn_t */
+        NULL,  /* vpx_codec_enc_config_set_fn_t */
+        NULL,  /* vpx_codec_get_global_headers_fn_t */
+        NULL,  /* vpx_codec_get_preview_frame_fn_t */
+        NULL   /* vpx_codec_enc_mr_get_mem_loc_fn_t */
+    }
+};
diff --git a/libvpx/libvpx/vp8/vp8cx.mk b/libvpx/libvpx/vp8/vp8cx.mk
new file mode 100644
index 0000000..857a631
--- /dev/null
+++ b/libvpx/libvpx/vp8/vp8cx.mk
@@ -0,0 +1,117 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+VP8_CX_EXPORTS += exports_enc
+
+VP8_CX_SRCS-yes += $(VP8_COMMON_SRCS-yes)
+VP8_CX_SRCS-no  += $(VP8_COMMON_SRCS-no)
+VP8_CX_SRCS_REMOVE-yes += $(VP8_COMMON_SRCS_REMOVE-yes)
+VP8_CX_SRCS_REMOVE-no  += $(VP8_COMMON_SRCS_REMOVE-no)
+
+ifeq ($(ARCH_ARM),yes)
+  include $(SRC_PATH_BARE)/$(VP8_PREFIX)vp8cx_arm.mk
+endif
+
+VP8_CX_SRCS-yes += vp8cx.mk
+
+VP8_CX_SRCS-yes += vp8_cx_iface.c
+
+VP8_CX_SRCS-yes += encoder/defaultcoefcounts.h
+VP8_CX_SRCS-yes += encoder/bitstream.c
+VP8_CX_SRCS-yes += encoder/boolhuff.c
+VP8_CX_SRCS-yes += encoder/dct.c
+VP8_CX_SRCS-yes += encoder/encodeframe.c
+VP8_CX_SRCS-yes += encoder/encodeframe.h
+VP8_CX_SRCS-yes += encoder/encodeintra.c
+VP8_CX_SRCS-yes += encoder/encodemb.c
+VP8_CX_SRCS-yes += encoder/encodemv.c
+VP8_CX_SRCS-$(CONFIG_MULTITHREAD) += encoder/ethreading.c
+VP8_CX_SRCS-yes += encoder/firstpass.c
+VP8_CX_SRCS-yes += encoder/block.h
+VP8_CX_SRCS-yes += encoder/boolhuff.h
+VP8_CX_SRCS-yes += encoder/bitstream.h
+VP8_CX_SRCS-$(CONFIG_TEMPORAL_DENOISING) += encoder/denoising.h
+VP8_CX_SRCS-$(CONFIG_TEMPORAL_DENOISING) += encoder/denoising.c
+VP8_CX_SRCS-yes += encoder/encodeintra.h
+VP8_CX_SRCS-yes += encoder/encodemb.h
+VP8_CX_SRCS-yes += encoder/encodemv.h
+VP8_CX_SRCS-yes += encoder/firstpass.h
+VP8_CX_SRCS-yes += encoder/lookahead.c
+VP8_CX_SRCS-yes += encoder/lookahead.h
+VP8_CX_SRCS-yes += encoder/mcomp.h
+VP8_CX_SRCS-yes += encoder/modecosts.h
+VP8_CX_SRCS-yes += encoder/onyx_int.h
+VP8_CX_SRCS-yes += encoder/pickinter.h
+VP8_CX_SRCS-yes += encoder/quantize.h
+VP8_CX_SRCS-yes += encoder/ratectrl.h
+VP8_CX_SRCS-yes += encoder/rdopt.h
+VP8_CX_SRCS-yes += encoder/tokenize.h
+VP8_CX_SRCS-yes += encoder/treewriter.h
+VP8_CX_SRCS-yes += encoder/mcomp.c
+VP8_CX_SRCS-yes += encoder/modecosts.c
+VP8_CX_SRCS-yes += encoder/onyx_if.c
+VP8_CX_SRCS-yes += encoder/pickinter.c
+VP8_CX_SRCS-yes += encoder/picklpf.c
+VP8_CX_SRCS-yes += encoder/vp8_quantize.c
+VP8_CX_SRCS-yes += encoder/ratectrl.c
+VP8_CX_SRCS-yes += encoder/rdopt.c
+VP8_CX_SRCS-yes += encoder/segmentation.c
+VP8_CX_SRCS-yes += encoder/segmentation.h
+VP8_CX_SRCS-yes += encoder/tokenize.c
+VP8_CX_SRCS-yes += encoder/dct_value_cost.h
+VP8_CX_SRCS-yes += encoder/dct_value_tokens.h
+VP8_CX_SRCS-yes += encoder/treewriter.c
+VP8_CX_SRCS-$(CONFIG_INTERNAL_STATS) += common/postproc.h
+VP8_CX_SRCS-$(CONFIG_INTERNAL_STATS) += common/postproc.c
+VP8_CX_SRCS-yes += encoder/temporal_filter.c
+VP8_CX_SRCS-$(CONFIG_MULTI_RES_ENCODING) += encoder/mr_dissim.c
+VP8_CX_SRCS-$(CONFIG_MULTI_RES_ENCODING) += encoder/mr_dissim.h
+
+ifeq ($(CONFIG_REALTIME_ONLY),yes)
+VP8_CX_SRCS_REMOVE-yes += encoder/firstpass.c
+VP8_CX_SRCS_REMOVE-yes += encoder/temporal_filter.c
+endif
+
+VP8_CX_SRCS-$(HAVE_MMX) += encoder/x86/dct_mmx.asm
+VP8_CX_SRCS-$(HAVE_MMX) += encoder/x86/vp8_enc_stubs_mmx.c
+VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_sse2.asm
+VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/fwalsh_sse2.asm
+VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp8_quantize_sse2.c
+VP8_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/quantize_ssse3.c
+VP8_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/quantize_sse4.c
+
+ifeq ($(CONFIG_TEMPORAL_DENOISING),yes)
+VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/denoising_sse2.c
+endif
+
+VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/temporal_filter_apply_sse2.asm
+VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp8_enc_stubs_sse2.c
+VP8_CX_SRCS-$(ARCH_X86)$(ARCH_X86_64) += encoder/x86/quantize_mmx.asm
+VP8_CX_SRCS-$(ARCH_X86)$(ARCH_X86_64) += encoder/x86/encodeopt.asm
+
+ifeq ($(CONFIG_REALTIME_ONLY),yes)
+VP8_CX_SRCS_REMOVE-$(HAVE_SSE2) += encoder/x86/temporal_filter_apply_sse2.asm
+endif
+
+VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/dct_msa.c
+VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/encodeopt_msa.c
+VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/quantize_msa.c
+VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/temporal_filter_msa.c
+
+ifeq ($(CONFIG_TEMPORAL_DENOISING),yes)
+VP8_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/denoising_msa.c
+endif
+
+ifeq ($(CONFIG_REALTIME_ONLY),yes)
+VP8_CX_SRCS_REMOVE-$(HAVE_MSA) += encoder/mips/msa/temporal_filter_msa.c
+endif
+
+VP8_CX_SRCS-yes := $(filter-out $(VP8_CX_SRCS_REMOVE-yes),$(VP8_CX_SRCS-yes))
diff --git a/libvpx/libvpx/vp8/vp8cx_arm.mk b/libvpx/libvpx/vp8/vp8cx_arm.mk
new file mode 100644
index 0000000..838b53d
--- /dev/null
+++ b/libvpx/libvpx/vp8/vp8cx_arm.mk
@@ -0,0 +1,28 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+VP8_CX_SRCS-$(ARCH_ARM)  += vp8cx_arm.mk
+
+#File list for arm
+# encoder
+VP8_CX_SRCS-$(ARCH_ARM)  += encoder/arm/dct_arm.c
+
+#File list for media
+# encoder
+VP8_CX_SRCS-$(HAVE_MEDIA)  += encoder/arm/armv6/vp8_short_fdct4x4_armv6$(ASM)
+VP8_CX_SRCS-$(HAVE_MEDIA)  += encoder/arm/armv6/walsh_v6$(ASM)
+
+#File list for neon
+# encoder
+VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/denoising_neon.c
+VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/fastquantizeb_neon.c
+VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/shortfdct_neon.c
+VP8_CX_SRCS-$(HAVE_NEON)  += encoder/arm/neon/vp8_shortwalsh4x4_neon.c
diff --git a/libvpx/libvpx/vp8/vp8dx.mk b/libvpx/libvpx/vp8/vp8dx.mk
new file mode 100644
index 0000000..892ed70
--- /dev/null
+++ b/libvpx/libvpx/vp8/vp8dx.mk
@@ -0,0 +1,39 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+VP8_DX_EXPORTS += exports_dec
+
+VP8_DX_SRCS-yes += $(VP8_COMMON_SRCS-yes)
+VP8_DX_SRCS-no  += $(VP8_COMMON_SRCS-no)
+VP8_DX_SRCS_REMOVE-yes += $(VP8_COMMON_SRCS_REMOVE-yes)
+VP8_DX_SRCS_REMOVE-no  += $(VP8_COMMON_SRCS_REMOVE-no)
+
+VP8_DX_SRCS-yes += vp8dx.mk
+
+VP8_DX_SRCS-yes += vp8_dx_iface.c
+
+VP8_DX_SRCS-yes += decoder/dboolhuff.c
+VP8_DX_SRCS-yes += decoder/decodemv.c
+VP8_DX_SRCS-yes += decoder/decodeframe.c
+VP8_DX_SRCS-yes += decoder/detokenize.c
+VP8_DX_SRCS-$(CONFIG_ERROR_CONCEALMENT) += decoder/ec_types.h
+VP8_DX_SRCS-$(CONFIG_ERROR_CONCEALMENT) += decoder/error_concealment.h
+VP8_DX_SRCS-$(CONFIG_ERROR_CONCEALMENT) += decoder/error_concealment.c
+VP8_DX_SRCS-yes += decoder/dboolhuff.h
+VP8_DX_SRCS-yes += decoder/decodemv.h
+VP8_DX_SRCS-yes += decoder/decoderthreading.h
+VP8_DX_SRCS-yes += decoder/detokenize.h
+VP8_DX_SRCS-yes += decoder/onyxd_int.h
+VP8_DX_SRCS-yes += decoder/treereader.h
+VP8_DX_SRCS-yes += decoder/onyxd_if.c
+VP8_DX_SRCS-$(CONFIG_MULTITHREAD) += decoder/threading.c
+
+VP8_DX_SRCS-yes := $(filter-out $(VP8_DX_SRCS_REMOVE-yes),$(VP8_DX_SRCS-yes))
diff --git a/libvpx/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.c b/libvpx/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.c
new file mode 100644
index 0000000..1761fad
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.c
@@ -0,0 +1,248 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "vp9/common/vp9_common.h"
+
+static int16_t sinpi_1_9 = 0x14a3;
+static int16_t sinpi_2_9 = 0x26c9;
+static int16_t sinpi_3_9 = 0x3441;
+static int16_t sinpi_4_9 = 0x3b6c;
+static int16_t cospi_8_64 = 0x3b21;
+static int16_t cospi_16_64 = 0x2d41;
+static int16_t cospi_24_64 = 0x187e;
+
+static INLINE void TRANSPOSE4X4(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16) {
+    int32x4_t q8s32, q9s32;
+    int16x4x2_t d0x2s16, d1x2s16;
+    int32x4x2_t q0x2s32;
+
+    d0x2s16 = vtrn_s16(vget_low_s16(*q8s16), vget_high_s16(*q8s16));
+    d1x2s16 = vtrn_s16(vget_low_s16(*q9s16), vget_high_s16(*q9s16));
+
+    q8s32 = vreinterpretq_s32_s16(vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]));
+    q9s32 = vreinterpretq_s32_s16(vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]));
+    q0x2s32 = vtrnq_s32(q8s32, q9s32);
+
+    *q8s16 = vreinterpretq_s16_s32(q0x2s32.val[0]);
+    *q9s16 = vreinterpretq_s16_s32(q0x2s32.val[1]);
+    return;
+}
+
+static INLINE void GENERATE_COSINE_CONSTANTS(
+        int16x4_t *d0s16,
+        int16x4_t *d1s16,
+        int16x4_t *d2s16) {
+    *d0s16 = vdup_n_s16(cospi_8_64);
+    *d1s16 = vdup_n_s16(cospi_16_64);
+    *d2s16 = vdup_n_s16(cospi_24_64);
+    return;
+}
+
+static INLINE void GENERATE_SINE_CONSTANTS(
+        int16x4_t *d3s16,
+        int16x4_t *d4s16,
+        int16x4_t *d5s16,
+        int16x8_t *q3s16) {
+    *d3s16 = vdup_n_s16(sinpi_1_9);
+    *d4s16 = vdup_n_s16(sinpi_2_9);
+    *q3s16 = vdupq_n_s16(sinpi_3_9);
+    *d5s16 = vdup_n_s16(sinpi_4_9);
+    return;
+}
+
+static INLINE void IDCT4x4_1D(
+        int16x4_t *d0s16,
+        int16x4_t *d1s16,
+        int16x4_t *d2s16,
+        int16x8_t *q8s16,
+        int16x8_t *q9s16) {
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d23s16, d24s16;
+    int16x4_t d26s16, d27s16, d28s16, d29s16;
+    int32x4_t q10s32, q13s32, q14s32, q15s32;
+    int16x8_t q13s16, q14s16;
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+
+    d23s16 = vadd_s16(d16s16, d18s16);
+    d24s16 = vsub_s16(d16s16, d18s16);
+
+    q15s32 = vmull_s16(d17s16, *d2s16);
+    q10s32 = vmull_s16(d17s16, *d0s16);
+    q13s32 = vmull_s16(d23s16, *d1s16);
+    q14s32 = vmull_s16(d24s16, *d1s16);
+    q15s32 = vmlsl_s16(q15s32, d19s16, *d0s16);
+    q10s32 = vmlal_s16(q10s32, d19s16, *d2s16);
+
+    d26s16 = vqrshrn_n_s32(q13s32, 14);
+    d27s16 = vqrshrn_n_s32(q14s32, 14);
+    d29s16 = vqrshrn_n_s32(q15s32, 14);
+    d28s16 = vqrshrn_n_s32(q10s32, 14);
+
+    q13s16 = vcombine_s16(d26s16, d27s16);
+    q14s16 = vcombine_s16(d28s16, d29s16);
+    *q8s16 = vaddq_s16(q13s16, q14s16);
+    *q9s16 = vsubq_s16(q13s16, q14s16);
+    *q9s16 = vcombine_s16(vget_high_s16(*q9s16),
+                          vget_low_s16(*q9s16));  // vswp
+    return;
+}
+
+static INLINE void IADST4x4_1D(
+        int16x4_t *d3s16,
+        int16x4_t *d4s16,
+        int16x4_t *d5s16,
+        int16x8_t *q3s16,
+        int16x8_t *q8s16,
+        int16x8_t *q9s16) {
+    int16x4_t d6s16, d16s16, d17s16, d18s16, d19s16;
+    int32x4_t q8s32, q9s32, q10s32, q11s32, q12s32, q13s32, q14s32, q15s32;
+
+    d6s16 = vget_low_s16(*q3s16);
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+
+    q10s32 = vmull_s16(*d3s16, d16s16);
+    q11s32 = vmull_s16(*d4s16, d16s16);
+    q12s32 = vmull_s16(d6s16, d17s16);
+    q13s32 = vmull_s16(*d5s16, d18s16);
+    q14s32 = vmull_s16(*d3s16, d18s16);
+    q15s32 = vmovl_s16(d16s16);
+    q15s32 = vaddw_s16(q15s32, d19s16);
+    q8s32  = vmull_s16(*d4s16, d19s16);
+    q15s32 = vsubw_s16(q15s32, d18s16);
+    q9s32  = vmull_s16(*d5s16, d19s16);
+
+    q10s32 = vaddq_s32(q10s32, q13s32);
+    q10s32 = vaddq_s32(q10s32, q8s32);
+    q11s32 = vsubq_s32(q11s32, q14s32);
+    q8s32  = vdupq_n_s32(sinpi_3_9);
+    q11s32 = vsubq_s32(q11s32, q9s32);
+    q15s32 = vmulq_s32(q15s32, q8s32);
+
+    q13s32 = vaddq_s32(q10s32, q12s32);
+    q10s32 = vaddq_s32(q10s32, q11s32);
+    q14s32 = vaddq_s32(q11s32, q12s32);
+    q10s32 = vsubq_s32(q10s32, q12s32);
+
+    d16s16 = vqrshrn_n_s32(q13s32, 14);
+    d17s16 = vqrshrn_n_s32(q14s32, 14);
+    d18s16 = vqrshrn_n_s32(q15s32, 14);
+    d19s16 = vqrshrn_n_s32(q10s32, 14);
+
+    *q8s16 = vcombine_s16(d16s16, d17s16);
+    *q9s16 = vcombine_s16(d18s16, d19s16);
+    return;
+}
+
+void vp9_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest,
+                            int dest_stride, int tx_type) {
+    uint8x8_t d26u8, d27u8;
+    int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16;
+    uint32x2_t d26u32, d27u32;
+    int16x8_t q3s16, q8s16, q9s16;
+    uint16x8_t q8u16, q9u16;
+
+    d26u32 = d27u32 = vdup_n_u32(0);
+
+    q8s16 = vld1q_s16(input);
+    q9s16 = vld1q_s16(input + 8);
+
+    TRANSPOSE4X4(&q8s16, &q9s16);
+
+    switch (tx_type) {
+      case 0:  // idct_idct is not supported. Fall back to C
+        vp9_iht4x4_16_add_c(input, dest, dest_stride, tx_type);
+        return;
+        break;
+      case 1:  // iadst_idct
+        // generate constants
+        GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
+        GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
+
+        // first transform rows
+        IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
+
+        // transpose the matrix
+        TRANSPOSE4X4(&q8s16, &q9s16);
+
+        // then transform columns
+        IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+        break;
+      case 2:  // idct_iadst
+        // generate constantsyy
+        GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
+        GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
+
+        // first transform rows
+        IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+
+        // transpose the matrix
+        TRANSPOSE4X4(&q8s16, &q9s16);
+
+        // then transform columns
+        IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
+        break;
+      case 3:  // iadst_iadst
+        // generate constants
+        GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
+
+        // first transform rows
+        IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+
+        // transpose the matrix
+        TRANSPOSE4X4(&q8s16, &q9s16);
+
+        // then transform columns
+        IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+        break;
+      default:  // iadst_idct
+        assert(0);
+        break;
+    }
+
+    q8s16 = vrshrq_n_s16(q8s16, 4);
+    q9s16 = vrshrq_n_s16(q9s16, 4);
+
+    d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 0);
+    dest += dest_stride;
+    d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 1);
+    dest += dest_stride;
+    d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 0);
+    dest += dest_stride;
+    d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 1);
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
+
+    d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+
+    vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 1);
+    dest -= dest_stride;
+    vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 0);
+    dest -= dest_stride;
+    vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 1);
+    dest -= dest_stride;
+    vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 0);
+    return;
+}
diff --git a/libvpx/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.c b/libvpx/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.c
new file mode 100644
index 0000000..04b342c
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.c
@@ -0,0 +1,624 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "vp9/common/vp9_common.h"
+
+static int16_t cospi_2_64 = 16305;
+static int16_t cospi_4_64 = 16069;
+static int16_t cospi_6_64 = 15679;
+static int16_t cospi_8_64 = 15137;
+static int16_t cospi_10_64 = 14449;
+static int16_t cospi_12_64 = 13623;
+static int16_t cospi_14_64 = 12665;
+static int16_t cospi_16_64 = 11585;
+static int16_t cospi_18_64 = 10394;
+static int16_t cospi_20_64 = 9102;
+static int16_t cospi_22_64 = 7723;
+static int16_t cospi_24_64 = 6270;
+static int16_t cospi_26_64 = 4756;
+static int16_t cospi_28_64 = 3196;
+static int16_t cospi_30_64 = 1606;
+
+static INLINE void TRANSPOSE8X8(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+    int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    *q8s16  = vcombine_s16(d16s16, d24s16);  // vswp d17, d24
+    *q9s16  = vcombine_s16(d18s16, d26s16);  // vswp d19, d26
+    *q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+    *q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+    *q12s16 = vcombine_s16(d17s16, d25s16);
+    *q13s16 = vcombine_s16(d19s16, d27s16);
+    *q14s16 = vcombine_s16(d21s16, d29s16);
+    *q15s16 = vcombine_s16(d23s16, d31s16);
+
+    q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q8s16),
+                        vreinterpretq_s32_s16(*q10s16));
+    q1x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q9s16),
+                        vreinterpretq_s32_s16(*q11s16));
+    q2x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q12s16),
+                        vreinterpretq_s32_s16(*q14s16));
+    q3x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q13s16),
+                        vreinterpretq_s32_s16(*q15s16));
+
+    q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                        vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+    q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                        vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+    q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                        vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+    q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                        vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+    *q8s16  = q0x2s16.val[0];
+    *q9s16  = q0x2s16.val[1];
+    *q10s16 = q1x2s16.val[0];
+    *q11s16 = q1x2s16.val[1];
+    *q12s16 = q2x2s16.val[0];
+    *q13s16 = q2x2s16.val[1];
+    *q14s16 = q3x2s16.val[0];
+    *q15s16 = q3x2s16.val[1];
+    return;
+}
+
+static INLINE void IDCT8x8_1D(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
+
+    d0s16 = vdup_n_s16(cospi_28_64);
+    d1s16 = vdup_n_s16(cospi_4_64);
+    d2s16 = vdup_n_s16(cospi_12_64);
+    d3s16 = vdup_n_s16(cospi_20_64);
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    q2s32 = vmull_s16(d18s16, d0s16);
+    q3s32 = vmull_s16(d19s16, d0s16);
+    q5s32 = vmull_s16(d26s16, d2s16);
+    q6s32 = vmull_s16(d27s16, d2s16);
+
+    q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
+    q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
+    q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
+    q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
+
+    d8s16  = vqrshrn_n_s32(q2s32, 14);
+    d9s16  = vqrshrn_n_s32(q3s32, 14);
+    d10s16 = vqrshrn_n_s32(q5s32, 14);
+    d11s16 = vqrshrn_n_s32(q6s32, 14);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    q2s32 = vmull_s16(d18s16, d1s16);
+    q3s32 = vmull_s16(d19s16, d1s16);
+    q9s32 = vmull_s16(d26s16, d3s16);
+    q13s32 = vmull_s16(d27s16, d3s16);
+
+    q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
+    q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
+    q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
+    q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
+
+    d14s16 = vqrshrn_n_s32(q2s32, 14);
+    d15s16 = vqrshrn_n_s32(q3s32, 14);
+    d12s16 = vqrshrn_n_s32(q9s32, 14);
+    d13s16 = vqrshrn_n_s32(q13s32, 14);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+    q7s16 = vcombine_s16(d14s16, d15s16);
+
+    d0s16 = vdup_n_s16(cospi_16_64);
+
+    q2s32 = vmull_s16(d16s16, d0s16);
+    q3s32 = vmull_s16(d17s16, d0s16);
+    q13s32 = vmull_s16(d16s16, d0s16);
+    q15s32 = vmull_s16(d17s16, d0s16);
+
+    q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
+    q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
+    q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
+    q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
+
+    d0s16 = vdup_n_s16(cospi_24_64);
+    d1s16 = vdup_n_s16(cospi_8_64);
+
+    d18s16 = vqrshrn_n_s32(q2s32, 14);
+    d19s16 = vqrshrn_n_s32(q3s32, 14);
+    d22s16 = vqrshrn_n_s32(q13s32, 14);
+    d23s16 = vqrshrn_n_s32(q15s32, 14);
+    *q9s16  = vcombine_s16(d18s16, d19s16);
+    *q11s16 = vcombine_s16(d22s16, d23s16);
+
+    q2s32 = vmull_s16(d20s16, d0s16);
+    q3s32 = vmull_s16(d21s16, d0s16);
+    q8s32 = vmull_s16(d20s16, d1s16);
+    q12s32 = vmull_s16(d21s16, d1s16);
+
+    q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
+    q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
+    q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
+    q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
+
+    d26s16 = vqrshrn_n_s32(q2s32, 14);
+    d27s16 = vqrshrn_n_s32(q3s32, 14);
+    d30s16 = vqrshrn_n_s32(q8s32, 14);
+    d31s16 = vqrshrn_n_s32(q12s32, 14);
+    *q13s16 = vcombine_s16(d26s16, d27s16);
+    *q15s16 = vcombine_s16(d30s16, d31s16);
+
+    q0s16 = vaddq_s16(*q9s16, *q15s16);
+    q1s16 = vaddq_s16(*q11s16, *q13s16);
+    q2s16 = vsubq_s16(*q11s16, *q13s16);
+    q3s16 = vsubq_s16(*q9s16, *q15s16);
+
+    *q13s16 = vsubq_s16(q4s16, q5s16);
+    q4s16   = vaddq_s16(q4s16, q5s16);
+    *q14s16 = vsubq_s16(q7s16, q6s16);
+    q7s16   = vaddq_s16(q7s16, q6s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+
+    d16s16 = vdup_n_s16(cospi_16_64);
+
+    q9s32  = vmull_s16(d28s16, d16s16);
+    q10s32 = vmull_s16(d29s16, d16s16);
+    q11s32 = vmull_s16(d28s16, d16s16);
+    q12s32 = vmull_s16(d29s16, d16s16);
+
+    q9s32  = vmlsl_s16(q9s32,  d26s16, d16s16);
+    q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+    q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+    q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+    d10s16 = vqrshrn_n_s32(q9s32, 14);
+    d11s16 = vqrshrn_n_s32(q10s32, 14);
+    d12s16 = vqrshrn_n_s32(q11s32, 14);
+    d13s16 = vqrshrn_n_s32(q12s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    *q8s16  = vaddq_s16(q0s16, q7s16);
+    *q9s16  = vaddq_s16(q1s16, q6s16);
+    *q10s16 = vaddq_s16(q2s16, q5s16);
+    *q11s16 = vaddq_s16(q3s16, q4s16);
+    *q12s16 = vsubq_s16(q3s16, q4s16);
+    *q13s16 = vsubq_s16(q2s16, q5s16);
+    *q14s16 = vsubq_s16(q1s16, q6s16);
+    *q15s16 = vsubq_s16(q0s16, q7s16);
+    return;
+}
+
+static INLINE void IADST8X8_1D(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int16x8_t q2s16, q4s16, q5s16, q6s16;
+    int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q7s32, q8s32;
+    int32x4_t q9s32, q10s32, q11s32, q12s32, q13s32, q14s32, q15s32;
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    d14s16 = vdup_n_s16(cospi_2_64);
+    d15s16 = vdup_n_s16(cospi_30_64);
+
+    q1s32 = vmull_s16(d30s16, d14s16);
+    q2s32 = vmull_s16(d31s16, d14s16);
+    q3s32 = vmull_s16(d30s16, d15s16);
+    q4s32 = vmull_s16(d31s16, d15s16);
+
+    d30s16 = vdup_n_s16(cospi_18_64);
+    d31s16 = vdup_n_s16(cospi_14_64);
+
+    q1s32 = vmlal_s16(q1s32, d16s16, d15s16);
+    q2s32 = vmlal_s16(q2s32, d17s16, d15s16);
+    q3s32 = vmlsl_s16(q3s32, d16s16, d14s16);
+    q4s32 = vmlsl_s16(q4s32, d17s16, d14s16);
+
+    q5s32 = vmull_s16(d22s16, d30s16);
+    q6s32 = vmull_s16(d23s16, d30s16);
+    q7s32 = vmull_s16(d22s16, d31s16);
+    q8s32 = vmull_s16(d23s16, d31s16);
+
+    q5s32 = vmlal_s16(q5s32, d24s16, d31s16);
+    q6s32 = vmlal_s16(q6s32, d25s16, d31s16);
+    q7s32 = vmlsl_s16(q7s32, d24s16, d30s16);
+    q8s32 = vmlsl_s16(q8s32, d25s16, d30s16);
+
+    q11s32 = vaddq_s32(q1s32, q5s32);
+    q12s32 = vaddq_s32(q2s32, q6s32);
+    q1s32 = vsubq_s32(q1s32, q5s32);
+    q2s32 = vsubq_s32(q2s32, q6s32);
+
+    d22s16 = vqrshrn_n_s32(q11s32, 14);
+    d23s16 = vqrshrn_n_s32(q12s32, 14);
+    *q11s16 = vcombine_s16(d22s16, d23s16);
+
+    q12s32 = vaddq_s32(q3s32, q7s32);
+    q15s32 = vaddq_s32(q4s32, q8s32);
+    q3s32 = vsubq_s32(q3s32, q7s32);
+    q4s32 = vsubq_s32(q4s32, q8s32);
+
+    d2s16  = vqrshrn_n_s32(q1s32, 14);
+    d3s16  = vqrshrn_n_s32(q2s32, 14);
+    d24s16 = vqrshrn_n_s32(q12s32, 14);
+    d25s16 = vqrshrn_n_s32(q15s32, 14);
+    d6s16  = vqrshrn_n_s32(q3s32, 14);
+    d7s16  = vqrshrn_n_s32(q4s32, 14);
+    *q12s16 = vcombine_s16(d24s16, d25s16);
+
+    d0s16 = vdup_n_s16(cospi_10_64);
+    d1s16 = vdup_n_s16(cospi_22_64);
+    q4s32 = vmull_s16(d26s16, d0s16);
+    q5s32 = vmull_s16(d27s16, d0s16);
+    q2s32 = vmull_s16(d26s16, d1s16);
+    q6s32 = vmull_s16(d27s16, d1s16);
+
+    d30s16 = vdup_n_s16(cospi_26_64);
+    d31s16 = vdup_n_s16(cospi_6_64);
+
+    q4s32 = vmlal_s16(q4s32, d20s16, d1s16);
+    q5s32 = vmlal_s16(q5s32, d21s16, d1s16);
+    q2s32 = vmlsl_s16(q2s32, d20s16, d0s16);
+    q6s32 = vmlsl_s16(q6s32, d21s16, d0s16);
+
+    q0s32 = vmull_s16(d18s16, d30s16);
+    q13s32 = vmull_s16(d19s16, d30s16);
+
+    q0s32 = vmlal_s16(q0s32, d28s16, d31s16);
+    q13s32 = vmlal_s16(q13s32, d29s16, d31s16);
+
+    q10s32 = vmull_s16(d18s16, d31s16);
+    q9s32 = vmull_s16(d19s16, d31s16);
+
+    q10s32 = vmlsl_s16(q10s32, d28s16, d30s16);
+    q9s32 = vmlsl_s16(q9s32, d29s16, d30s16);
+
+    q14s32 = vaddq_s32(q2s32, q10s32);
+    q15s32 = vaddq_s32(q6s32, q9s32);
+    q2s32 = vsubq_s32(q2s32, q10s32);
+    q6s32 = vsubq_s32(q6s32, q9s32);
+
+    d28s16 = vqrshrn_n_s32(q14s32, 14);
+    d29s16 = vqrshrn_n_s32(q15s32, 14);
+    d4s16 = vqrshrn_n_s32(q2s32, 14);
+    d5s16 = vqrshrn_n_s32(q6s32, 14);
+    *q14s16 = vcombine_s16(d28s16, d29s16);
+
+    q9s32 = vaddq_s32(q4s32, q0s32);
+    q10s32 = vaddq_s32(q5s32, q13s32);
+    q4s32 = vsubq_s32(q4s32, q0s32);
+    q5s32 = vsubq_s32(q5s32, q13s32);
+
+    d30s16 = vdup_n_s16(cospi_8_64);
+    d31s16 = vdup_n_s16(cospi_24_64);
+
+    d18s16 = vqrshrn_n_s32(q9s32, 14);
+    d19s16 = vqrshrn_n_s32(q10s32, 14);
+    d8s16 = vqrshrn_n_s32(q4s32, 14);
+    d9s16 = vqrshrn_n_s32(q5s32, 14);
+    *q9s16 = vcombine_s16(d18s16, d19s16);
+
+    q5s32 = vmull_s16(d2s16, d30s16);
+    q6s32 = vmull_s16(d3s16, d30s16);
+    q7s32 = vmull_s16(d2s16, d31s16);
+    q0s32 = vmull_s16(d3s16, d31s16);
+
+    q5s32 = vmlal_s16(q5s32, d6s16, d31s16);
+    q6s32 = vmlal_s16(q6s32, d7s16, d31s16);
+    q7s32 = vmlsl_s16(q7s32, d6s16, d30s16);
+    q0s32 = vmlsl_s16(q0s32, d7s16, d30s16);
+
+    q1s32 = vmull_s16(d4s16, d30s16);
+    q3s32 = vmull_s16(d5s16, d30s16);
+    q10s32 = vmull_s16(d4s16, d31s16);
+    q2s32 = vmull_s16(d5s16, d31s16);
+
+    q1s32 = vmlsl_s16(q1s32, d8s16, d31s16);
+    q3s32 = vmlsl_s16(q3s32, d9s16, d31s16);
+    q10s32 = vmlal_s16(q10s32, d8s16, d30s16);
+    q2s32 = vmlal_s16(q2s32, d9s16, d30s16);
+
+    *q8s16 = vaddq_s16(*q11s16, *q9s16);
+    *q11s16 = vsubq_s16(*q11s16, *q9s16);
+    q4s16 = vaddq_s16(*q12s16, *q14s16);
+    *q12s16 = vsubq_s16(*q12s16, *q14s16);
+
+    q14s32 = vaddq_s32(q5s32, q1s32);
+    q15s32 = vaddq_s32(q6s32, q3s32);
+    q5s32 = vsubq_s32(q5s32, q1s32);
+    q6s32 = vsubq_s32(q6s32, q3s32);
+
+    d18s16 = vqrshrn_n_s32(q14s32, 14);
+    d19s16 = vqrshrn_n_s32(q15s32, 14);
+    d10s16 = vqrshrn_n_s32(q5s32, 14);
+    d11s16 = vqrshrn_n_s32(q6s32, 14);
+    *q9s16 = vcombine_s16(d18s16, d19s16);
+
+    q1s32 = vaddq_s32(q7s32, q10s32);
+    q3s32 = vaddq_s32(q0s32, q2s32);
+    q7s32 = vsubq_s32(q7s32, q10s32);
+    q0s32 = vsubq_s32(q0s32, q2s32);
+
+    d28s16 = vqrshrn_n_s32(q1s32, 14);
+    d29s16 = vqrshrn_n_s32(q3s32, 14);
+    d14s16 = vqrshrn_n_s32(q7s32, 14);
+    d15s16 = vqrshrn_n_s32(q0s32, 14);
+    *q14s16 = vcombine_s16(d28s16, d29s16);
+
+    d30s16 = vdup_n_s16(cospi_16_64);
+
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    q2s32 = vmull_s16(d22s16, d30s16);
+    q3s32 = vmull_s16(d23s16, d30s16);
+    q13s32 = vmull_s16(d22s16, d30s16);
+    q1s32 = vmull_s16(d23s16, d30s16);
+
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    q2s32 = vmlal_s16(q2s32, d24s16, d30s16);
+    q3s32 = vmlal_s16(q3s32, d25s16, d30s16);
+    q13s32 = vmlsl_s16(q13s32, d24s16, d30s16);
+    q1s32 = vmlsl_s16(q1s32, d25s16, d30s16);
+
+    d4s16 = vqrshrn_n_s32(q2s32, 14);
+    d5s16 = vqrshrn_n_s32(q3s32, 14);
+    d24s16 = vqrshrn_n_s32(q13s32, 14);
+    d25s16 = vqrshrn_n_s32(q1s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    *q12s16 = vcombine_s16(d24s16, d25s16);
+
+    q13s32 = vmull_s16(d10s16, d30s16);
+    q1s32 = vmull_s16(d11s16, d30s16);
+    q11s32 = vmull_s16(d10s16, d30s16);
+    q0s32 = vmull_s16(d11s16, d30s16);
+
+    q13s32 = vmlal_s16(q13s32, d14s16, d30s16);
+    q1s32 = vmlal_s16(q1s32, d15s16, d30s16);
+    q11s32 = vmlsl_s16(q11s32, d14s16, d30s16);
+    q0s32 = vmlsl_s16(q0s32, d15s16, d30s16);
+
+    d20s16 = vqrshrn_n_s32(q13s32, 14);
+    d21s16 = vqrshrn_n_s32(q1s32, 14);
+    d12s16 = vqrshrn_n_s32(q11s32, 14);
+    d13s16 = vqrshrn_n_s32(q0s32, 14);
+    *q10s16 = vcombine_s16(d20s16, d21s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    q5s16 = vdupq_n_s16(0);
+
+    *q9s16  = vsubq_s16(q5s16, *q9s16);
+    *q11s16 = vsubq_s16(q5s16, q2s16);
+    *q13s16 = vsubq_s16(q5s16, q6s16);
+    *q15s16 = vsubq_s16(q5s16, q4s16);
+    return;
+}
+
+void vp9_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest,
+                            int dest_stride, int tx_type) {
+    int i;
+    uint8_t *d1, *d2;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8;
+    uint64x1_t d0u64, d1u64, d2u64, d3u64;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16;
+
+    q8s16  = vld1q_s16(input);
+    q9s16  = vld1q_s16(input + 8);
+    q10s16 = vld1q_s16(input + 8 * 2);
+    q11s16 = vld1q_s16(input + 8 * 3);
+    q12s16 = vld1q_s16(input + 8 * 4);
+    q13s16 = vld1q_s16(input + 8 * 5);
+    q14s16 = vld1q_s16(input + 8 * 6);
+    q15s16 = vld1q_s16(input + 8 * 7);
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    switch (tx_type) {
+      case 0:  // idct_idct is not supported. Fall back to C
+        vp9_iht8x8_64_add_c(input, dest, dest_stride, tx_type);
+        return;
+        break;
+      case 1:  // iadst_idct
+        // generate IDCT constants
+        // GENERATE_IDCT_CONSTANTS
+
+        // first transform rows
+        IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                   &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // transpose the matrix
+        TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                     &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // generate IADST constants
+        // GENERATE_IADST_CONSTANTS
+
+        // then transform columns
+        IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                    &q12s16, &q13s16, &q14s16, &q15s16);
+        break;
+      case 2:  // idct_iadst
+        // generate IADST constants
+        // GENERATE_IADST_CONSTANTS
+
+        // first transform rows
+        IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                    &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // transpose the matrix
+        TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                     &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // generate IDCT constants
+        // GENERATE_IDCT_CONSTANTS
+
+        // then transform columns
+        IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                   &q12s16, &q13s16, &q14s16, &q15s16);
+        break;
+      case 3:  // iadst_iadst
+        // generate IADST constants
+        // GENERATE_IADST_CONSTANTS
+
+        // first transform rows
+        IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                    &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // transpose the matrix
+        TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                     &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // then transform columns
+        IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                    &q12s16, &q13s16, &q14s16, &q15s16);
+        break;
+      default:  // iadst_idct
+        assert(0);
+        break;
+    }
+
+    q8s16 = vrshrq_n_s16(q8s16, 5);
+    q9s16 = vrshrq_n_s16(q9s16, 5);
+    q10s16 = vrshrq_n_s16(q10s16, 5);
+    q11s16 = vrshrq_n_s16(q11s16, 5);
+    q12s16 = vrshrq_n_s16(q12s16, 5);
+    q13s16 = vrshrq_n_s16(q13s16, 5);
+    q14s16 = vrshrq_n_s16(q14s16, 5);
+    q15s16 = vrshrq_n_s16(q15s16, 5);
+
+    for (d1 = d2 = dest, i = 0; i < 2; i++) {
+        if (i != 0) {
+            q8s16 = q12s16;
+            q9s16 = q13s16;
+            q10s16 = q14s16;
+            q11s16 = q15s16;
+        }
+
+        d0u64 = vld1_u64((uint64_t *)d1);
+        d1 += dest_stride;
+        d1u64 = vld1_u64((uint64_t *)d1);
+        d1 += dest_stride;
+        d2u64 = vld1_u64((uint64_t *)d1);
+        d1 += dest_stride;
+        d3u64 = vld1_u64((uint64_t *)d1);
+        d1 += dest_stride;
+
+        q8u16  = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                          vreinterpret_u8_u64(d0u64));
+        q9u16  = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                          vreinterpret_u8_u64(d1u64));
+        q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
+                          vreinterpret_u8_u64(d2u64));
+        q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
+                          vreinterpret_u8_u64(d3u64));
+
+        d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+        d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+        d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+        d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+        d2 += dest_stride;
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vp9/common/mips/dspr2/vp9_itrans16_dspr2.c b/libvpx/libvpx/vp9/common/mips/dspr2/vp9_itrans16_dspr2.c
new file mode 100644
index 0000000..6ca83a0
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/mips/dspr2/vp9_itrans16_dspr2.c
@@ -0,0 +1,108 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_idct.h"
+#include "vpx_dsp/mips/inv_txfm_dspr2.h"
+#include "vpx_dsp/txfm_common.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+void vp9_iht16x16_256_add_dspr2(const int16_t *input, uint8_t *dest,
+                                int pitch, int tx_type) {
+  int i, j;
+  DECLARE_ALIGNED(32, int16_t,  out[16 * 16]);
+  int16_t *outptr = out;
+  int16_t temp_out[16];
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp    %[pos],    1    \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  switch (tx_type) {
+    case DCT_DCT:     // DCT in both horizontal and vertical
+      idct16_rows_dspr2(input, outptr, 16);
+      idct16_cols_add_blk_dspr2(out, dest, pitch);
+      break;
+    case ADST_DCT:    // ADST in vertical, DCT in horizontal
+      idct16_rows_dspr2(input, outptr, 16);
+
+      outptr = out;
+
+      for (i = 0; i < 16; ++i) {
+        iadst16_dspr2(outptr, temp_out);
+
+        for (j = 0; j < 16; ++j)
+          dest[j * pitch + i] =
+                    clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 6)
+                                      + dest[j * pitch + i]);
+        outptr += 16;
+      }
+      break;
+    case DCT_ADST:    // DCT in vertical, ADST in horizontal
+    {
+      int16_t temp_in[16 * 16];
+
+      for (i = 0; i < 16; ++i) {
+        /* prefetch row */
+        prefetch_load((const uint8_t *)(input + 16));
+
+        iadst16_dspr2(input, outptr);
+        input += 16;
+        outptr += 16;
+      }
+
+      for (i = 0; i < 16; ++i)
+        for (j = 0; j < 16; ++j)
+            temp_in[j * 16 + i] = out[i * 16 + j];
+
+      idct16_cols_add_blk_dspr2(temp_in, dest, pitch);
+    }
+    break;
+    case ADST_ADST:   // ADST in both directions
+    {
+      int16_t temp_in[16];
+
+      for (i = 0; i < 16; ++i) {
+        /* prefetch row */
+        prefetch_load((const uint8_t *)(input + 16));
+
+        iadst16_dspr2(input, outptr);
+        input += 16;
+        outptr += 16;
+      }
+
+      for (i = 0; i < 16; ++i) {
+        for (j = 0; j < 16; ++j)
+          temp_in[j] = out[j * 16 + i];
+        iadst16_dspr2(temp_in, temp_out);
+        for (j = 0; j < 16; ++j)
+          dest[j * pitch + i] =
+                    clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 6)
+                                      + dest[j * pitch + i]);
+      }
+    }
+    break;
+    default:
+      printf("vp9_short_iht16x16_add_dspr2 : Invalid tx_type\n");
+      break;
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vp9/common/mips/dspr2/vp9_itrans4_dspr2.c b/libvpx/libvpx/vp9/common/mips/dspr2/vp9_itrans4_dspr2.c
new file mode 100644
index 0000000..c10979b
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/mips/dspr2/vp9_itrans4_dspr2.c
@@ -0,0 +1,97 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_idct.h"
+#include "vpx_dsp/mips/inv_txfm_dspr2.h"
+#include "vpx_dsp/txfm_common.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+void vp9_iht4x4_16_add_dspr2(const int16_t *input, uint8_t *dest,
+                             int dest_stride, int tx_type) {
+  int i, j;
+  DECLARE_ALIGNED(32, int16_t, out[4 * 4]);
+  int16_t *outptr = out;
+  int16_t temp_in[4 * 4], temp_out[4];
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  switch (tx_type) {
+    case DCT_DCT:   // DCT in both horizontal and vertical
+      vpx_idct4_rows_dspr2(input, outptr);
+      vpx_idct4_columns_add_blk_dspr2(&out[0], dest, dest_stride);
+      break;
+    case ADST_DCT:  // ADST in vertical, DCT in horizontal
+      vpx_idct4_rows_dspr2(input, outptr);
+
+      outptr = out;
+
+      for (i = 0; i < 4; ++i) {
+        iadst4_dspr2(outptr, temp_out);
+
+        for (j = 0; j < 4; ++j)
+          dest[j * dest_stride + i] =
+                    clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 4)
+                                      + dest[j * dest_stride + i]);
+
+        outptr += 4;
+      }
+      break;
+    case DCT_ADST:  // DCT in vertical, ADST in horizontal
+      for (i = 0; i < 4; ++i) {
+        iadst4_dspr2(input, outptr);
+        input  += 4;
+        outptr += 4;
+      }
+
+      for (i = 0; i < 4; ++i) {
+        for (j = 0; j < 4; ++j) {
+          temp_in[i * 4 + j] = out[j * 4 + i];
+        }
+      }
+      vpx_idct4_columns_add_blk_dspr2(&temp_in[0], dest, dest_stride);
+      break;
+    case ADST_ADST:  // ADST in both directions
+      for (i = 0; i < 4; ++i) {
+        iadst4_dspr2(input, outptr);
+        input  += 4;
+        outptr += 4;
+      }
+
+      for (i = 0; i < 4; ++i) {
+        for (j = 0; j < 4; ++j)
+          temp_in[j] = out[j * 4 + i];
+        iadst4_dspr2(temp_in, temp_out);
+
+        for (j = 0; j < 4; ++j)
+          dest[j * dest_stride + i] =
+                  clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 4)
+                                      + dest[j * dest_stride + i]);
+      }
+      break;
+    default:
+      printf("vp9_short_iht4x4_add_dspr2 : Invalid tx_type\n");
+      break;
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vp9/common/mips/dspr2/vp9_itrans8_dspr2.c b/libvpx/libvpx/vp9/common/mips/dspr2/vp9_itrans8_dspr2.c
new file mode 100644
index 0000000..37f3ca9
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/mips/dspr2/vp9_itrans8_dspr2.c
@@ -0,0 +1,93 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vpx_dsp/mips/inv_txfm_dspr2.h"
+#include "vpx_dsp/txfm_common.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+void vp9_iht8x8_64_add_dspr2(const int16_t *input, uint8_t *dest,
+                             int dest_stride, int tx_type) {
+  int i, j;
+  DECLARE_ALIGNED(32, int16_t, out[8 * 8]);
+  int16_t *outptr = out;
+  int16_t temp_in[8 * 8], temp_out[8];
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp    %[pos],    1    \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  switch (tx_type) {
+    case DCT_DCT:     // DCT in both horizontal and vertical
+      idct8_rows_dspr2(input, outptr, 8);
+      idct8_columns_add_blk_dspr2(&out[0], dest, dest_stride);
+      break;
+    case ADST_DCT:    // ADST in vertical, DCT in horizontal
+      idct8_rows_dspr2(input, outptr, 8);
+
+      for (i = 0; i < 8; ++i) {
+        iadst8_dspr2(&out[i * 8], temp_out);
+
+        for (j = 0; j < 8; ++j)
+          dest[j * dest_stride + i] =
+                    clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 5)
+                                      + dest[j * dest_stride + i]);
+      }
+      break;
+    case DCT_ADST:    // DCT in vertical, ADST in horizontal
+      for (i = 0; i < 8; ++i) {
+        iadst8_dspr2(input, outptr);
+        input += 8;
+        outptr += 8;
+      }
+
+      for (i = 0; i < 8; ++i) {
+        for (j = 0; j < 8; ++j) {
+          temp_in[i * 8 + j] = out[j * 8 + i];
+        }
+      }
+      idct8_columns_add_blk_dspr2(&temp_in[0], dest, dest_stride);
+      break;
+    case ADST_ADST:   // ADST in both directions
+      for (i = 0; i < 8; ++i) {
+        iadst8_dspr2(input, outptr);
+        input += 8;
+        outptr += 8;
+      }
+
+      for (i = 0; i < 8; ++i) {
+        for (j = 0; j < 8; ++j)
+          temp_in[j] = out[j * 8 + i];
+
+        iadst8_dspr2(temp_in, temp_out);
+
+        for (j = 0; j < 8; ++j)
+          dest[j * dest_stride + i] =
+                clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 5)
+                                      + dest[j * dest_stride + i]);
+      }
+      break;
+    default:
+      printf("vp9_short_iht8x8_add_dspr2 : Invalid tx_type\n");
+      break;
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vp9/common/mips/msa/vp9_idct16x16_msa.c b/libvpx/libvpx/vp9/common/mips/msa/vp9_idct16x16_msa.c
new file mode 100644
index 0000000..5adf0aa
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/mips/msa/vp9_idct16x16_msa.c
@@ -0,0 +1,81 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_enums.h"
+#include "vpx_dsp/mips/inv_txfm_msa.h"
+
+void vp9_iht16x16_256_add_msa(const int16_t *input, uint8_t *dst,
+                              int32_t dst_stride, int32_t tx_type) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, out[16 * 16]);
+  int16_t *out_ptr = &out[0];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      /* transform rows */
+      for (i = 0; i < 2; ++i) {
+        /* process 16 * 8 block */
+        vpx_idct16_1d_rows_msa((input + (i << 7)), (out_ptr + (i << 7)));
+      }
+
+      /* transform columns */
+      for (i = 0; i < 2; ++i) {
+        /* process 8 * 16 block */
+        vpx_idct16_1d_columns_addblk_msa((out_ptr + (i << 3)), (dst + (i << 3)),
+                                         dst_stride);
+      }
+      break;
+    case ADST_DCT:
+      /* transform rows */
+      for (i = 0; i < 2; ++i) {
+        /* process 16 * 8 block */
+        vpx_idct16_1d_rows_msa((input + (i << 7)), (out_ptr + (i << 7)));
+      }
+
+      /* transform columns */
+      for (i = 0; i < 2; ++i) {
+        vpx_iadst16_1d_columns_addblk_msa((out_ptr + (i << 3)),
+                                          (dst + (i << 3)), dst_stride);
+      }
+      break;
+    case DCT_ADST:
+      /* transform rows */
+      for (i = 0; i < 2; ++i) {
+        /* process 16 * 8 block */
+        vpx_iadst16_1d_rows_msa((input + (i << 7)), (out_ptr + (i << 7)));
+      }
+
+      /* transform columns */
+      for (i = 0; i < 2; ++i) {
+        /* process 8 * 16 block */
+        vpx_idct16_1d_columns_addblk_msa((out_ptr + (i << 3)), (dst + (i << 3)),
+                                         dst_stride);
+      }
+      break;
+    case ADST_ADST:
+      /* transform rows */
+      for (i = 0; i < 2; ++i) {
+        /* process 16 * 8 block */
+        vpx_iadst16_1d_rows_msa((input + (i << 7)), (out_ptr + (i << 7)));
+      }
+
+      /* transform columns */
+      for (i = 0; i < 2; ++i) {
+        vpx_iadst16_1d_columns_addblk_msa((out_ptr + (i << 3)),
+                                          (dst + (i << 3)), dst_stride);
+      }
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
diff --git a/libvpx/libvpx/vp9/common/mips/msa/vp9_idct4x4_msa.c b/libvpx/libvpx/vp9/common/mips/msa/vp9_idct4x4_msa.c
new file mode 100644
index 0000000..75977b1
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/mips/msa/vp9_idct4x4_msa.c
@@ -0,0 +1,62 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_enums.h"
+#include "vpx_dsp/mips/inv_txfm_msa.h"
+
+void vp9_iht4x4_16_add_msa(const int16_t *input, uint8_t *dst,
+                           int32_t dst_stride, int32_t tx_type) {
+  v8i16 in0, in1, in2, in3;
+
+  /* load vector elements of 4x4 block */
+  LD4x4_SH(input, in0, in1, in2, in3);
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      /* DCT in horizontal */
+      VP9_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      /* DCT in vertical */
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      VP9_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case ADST_DCT:
+      /* DCT in horizontal */
+      VP9_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      /* ADST in vertical */
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      VP9_IADST4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case DCT_ADST:
+      /* ADST in horizontal */
+      VP9_IADST4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      /* DCT in vertical */
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      VP9_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case ADST_ADST:
+      /* ADST in horizontal */
+      VP9_IADST4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      /* ADST in vertical */
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      VP9_IADST4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  /* final rounding (add 2^3, divide by 2^4) and shift */
+  SRARI_H4_SH(in0, in1, in2, in3, 4);
+  /* add block and store 4x4 */
+  ADDBLK_ST4x4_UB(in0, in1, in2, in3, dst, dst_stride);
+}
diff --git a/libvpx/libvpx/vp9/common/mips/msa/vp9_idct8x8_msa.c b/libvpx/libvpx/vp9/common/mips/msa/vp9_idct8x8_msa.c
new file mode 100644
index 0000000..65d2993
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/mips/msa/vp9_idct8x8_msa.c
@@ -0,0 +1,80 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_enums.h"
+#include "vpx_dsp/mips/inv_txfm_msa.h"
+
+void vp9_iht8x8_64_add_msa(const int16_t *input, uint8_t *dst,
+                           int32_t dst_stride, int32_t tx_type) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  /* load vector elements of 8x8 block */
+  LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      /* DCT in horizontal */
+      VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+      /* DCT in vertical */
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    case ADST_DCT:
+      /* DCT in horizontal */
+      VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+      /* ADST in vertical */
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    case DCT_ADST:
+      /* ADST in horizontal */
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      /* DCT in vertical */
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    case ADST_ADST:
+      /* ADST in horizontal */
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      /* ADST in vertical */
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  /* final rounding (add 2^4, divide by 2^5) and shift */
+  SRARI_H4_SH(in0, in1, in2, in3, 5);
+  SRARI_H4_SH(in4, in5, in6, in7, 5);
+
+  /* add block and store 8x8 */
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, in0, in1, in2, in3);
+  dst += (4 * dst_stride);
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, in4, in5, in6, in7);
+}
diff --git a/libvpx/libvpx/vp9/common/mips/msa/vp9_mfqe_msa.c b/libvpx/libvpx/vp9/common/mips/msa/vp9_mfqe_msa.c
new file mode 100644
index 0000000..7257cd6
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/mips/msa/vp9_mfqe_msa.c
@@ -0,0 +1,137 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vpx_dsp/mips/macros_msa.h"
+
+static void filter_by_weight8x8_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                    uint8_t *dst_ptr, int32_t dst_stride,
+                                    int32_t src_weight) {
+  int32_t dst_weight = (1 << MFQE_PRECISION) - src_weight;
+  int32_t row;
+  uint64_t src0_d, src1_d, dst0_d, dst1_d;
+  v16i8 src0 = { 0 };
+  v16i8 src1 = { 0 };
+  v16i8 dst0 = { 0 };
+  v16i8 dst1 = { 0 };
+  v8i16 src_wt, dst_wt, res_h_r, res_h_l, src_r, src_l, dst_r, dst_l;
+
+  src_wt = __msa_fill_h(src_weight);
+  dst_wt = __msa_fill_h(dst_weight);
+
+  for (row = 2; row--;) {
+    LD2(src_ptr, src_stride, src0_d, src1_d);
+    src_ptr += (2 * src_stride);
+    LD2(dst_ptr, dst_stride, dst0_d, dst1_d);
+    INSERT_D2_SB(src0_d, src1_d, src0);
+    INSERT_D2_SB(dst0_d, dst1_d, dst0);
+
+    LD2(src_ptr, src_stride, src0_d, src1_d);
+    src_ptr += (2 * src_stride);
+    LD2((dst_ptr + 2 * dst_stride), dst_stride, dst0_d, dst1_d);
+    INSERT_D2_SB(src0_d, src1_d, src1);
+    INSERT_D2_SB(dst0_d, dst1_d, dst1);
+
+    UNPCK_UB_SH(src0, src_r, src_l);
+    UNPCK_UB_SH(dst0, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    dst0 = (v16i8)__msa_pckev_b((v16i8)res_h_l, (v16i8)res_h_r);
+    ST8x2_UB(dst0, dst_ptr, dst_stride);
+    dst_ptr += (2 * dst_stride);
+
+    UNPCK_UB_SH(src1, src_r, src_l);
+    UNPCK_UB_SH(dst1, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    dst1 = (v16i8)__msa_pckev_b((v16i8)res_h_l, (v16i8)res_h_r);
+    ST8x2_UB(dst1, dst_ptr, dst_stride);
+    dst_ptr += (2 * dst_stride);
+  }
+}
+
+static void filter_by_weight16x16_msa(const uint8_t *src_ptr,
+                                      int32_t src_stride,
+                                      uint8_t *dst_ptr,
+                                      int32_t dst_stride,
+                                      int32_t src_weight) {
+  int32_t dst_weight = (1 << MFQE_PRECISION) - src_weight;
+  int32_t row;
+  v16i8 src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  v8i16 src_wt, dst_wt, res_h_r, res_h_l, src_r, src_l, dst_r, dst_l;
+
+  src_wt = __msa_fill_h(src_weight);
+  dst_wt = __msa_fill_h(dst_weight);
+
+  for (row = 4; row--;) {
+    LD_SB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_SB4(dst_ptr, dst_stride, dst0, dst1, dst2, dst3);
+
+    UNPCK_UB_SH(src0, src_r, src_l);
+    UNPCK_UB_SH(dst0, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+    dst_ptr += dst_stride;
+
+    UNPCK_UB_SH(src1, src_r, src_l);
+    UNPCK_UB_SH(dst1, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+    dst_ptr += dst_stride;
+
+    UNPCK_UB_SH(src2, src_r, src_l);
+    UNPCK_UB_SH(dst2, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+    dst_ptr += dst_stride;
+
+    UNPCK_UB_SH(src3, src_r, src_l);
+    UNPCK_UB_SH(dst3, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+    dst_ptr += dst_stride;
+  }
+}
+
+void vp9_filter_by_weight8x8_msa(const uint8_t *src, int src_stride,
+                                 uint8_t *dst, int dst_stride,
+                                 int src_weight) {
+  filter_by_weight8x8_msa(src, src_stride, dst, dst_stride, src_weight);
+}
+
+void vp9_filter_by_weight16x16_msa(const uint8_t *src, int src_stride,
+                                   uint8_t *dst, int dst_stride,
+                                   int src_weight) {
+  filter_by_weight16x16_msa(src, src_stride, dst, dst_stride, src_weight);
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_alloccommon.c b/libvpx/libvpx/vp9/common/vp9_alloccommon.c
new file mode 100644
index 0000000..7dd1005
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_alloccommon.c
@@ -0,0 +1,201 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_onyxc_int.h"
+
+// TODO(hkuang): Don't need to lock the whole pool after implementing atomic
+// frame reference count.
+void lock_buffer_pool(BufferPool *const pool) {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_lock(&pool->pool_mutex);
+#else
+  (void)pool;
+#endif
+}
+
+void unlock_buffer_pool(BufferPool *const pool) {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_unlock(&pool->pool_mutex);
+#else
+  (void)pool;
+#endif
+}
+
+void vp9_set_mb_mi(VP9_COMMON *cm, int width, int height) {
+  const int aligned_width = ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2);
+  const int aligned_height = ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2);
+
+  cm->mi_cols = aligned_width >> MI_SIZE_LOG2;
+  cm->mi_rows = aligned_height >> MI_SIZE_LOG2;
+  cm->mi_stride = calc_mi_size(cm->mi_cols);
+
+  cm->mb_cols = (cm->mi_cols + 1) >> 1;
+  cm->mb_rows = (cm->mi_rows + 1) >> 1;
+  cm->MBs = cm->mb_rows * cm->mb_cols;
+}
+
+static int alloc_seg_map(VP9_COMMON *cm, int seg_map_size) {
+  int i;
+
+  for (i = 0; i < NUM_PING_PONG_BUFFERS; ++i) {
+    cm->seg_map_array[i] = (uint8_t *)vpx_calloc(seg_map_size, 1);
+    if (cm->seg_map_array[i] == NULL)
+      return 1;
+  }
+  cm->seg_map_alloc_size = seg_map_size;
+
+  // Init the index.
+  cm->seg_map_idx = 0;
+  cm->prev_seg_map_idx = 1;
+
+  cm->current_frame_seg_map = cm->seg_map_array[cm->seg_map_idx];
+  if (!cm->frame_parallel_decode)
+    cm->last_frame_seg_map = cm->seg_map_array[cm->prev_seg_map_idx];
+
+  return 0;
+}
+
+static void free_seg_map(VP9_COMMON *cm) {
+  int i;
+
+  for (i = 0; i < NUM_PING_PONG_BUFFERS; ++i) {
+    vpx_free(cm->seg_map_array[i]);
+    cm->seg_map_array[i] = NULL;
+  }
+
+  cm->current_frame_seg_map = NULL;
+
+  if (!cm->frame_parallel_decode) {
+    cm->last_frame_seg_map = NULL;
+  }
+}
+
+void vp9_free_ref_frame_buffers(BufferPool *pool) {
+  int i;
+
+  for (i = 0; i < FRAME_BUFFERS; ++i) {
+    if (pool->frame_bufs[i].ref_count > 0 &&
+        pool->frame_bufs[i].raw_frame_buffer.data != NULL) {
+      pool->release_fb_cb(pool->cb_priv, &pool->frame_bufs[i].raw_frame_buffer);
+      pool->frame_bufs[i].ref_count = 0;
+    }
+    vpx_free(pool->frame_bufs[i].mvs);
+    pool->frame_bufs[i].mvs = NULL;
+    vpx_free_frame_buffer(&pool->frame_bufs[i].buf);
+  }
+}
+
+void vp9_free_postproc_buffers(VP9_COMMON *cm) {
+#if CONFIG_VP9_POSTPROC
+  vpx_free_frame_buffer(&cm->post_proc_buffer);
+  vpx_free_frame_buffer(&cm->post_proc_buffer_int);
+#else
+  (void)cm;
+#endif
+}
+
+void vp9_free_context_buffers(VP9_COMMON *cm) {
+  cm->free_mi(cm);
+  free_seg_map(cm);
+  vpx_free(cm->above_context);
+  cm->above_context = NULL;
+  vpx_free(cm->above_seg_context);
+  cm->above_seg_context = NULL;
+  vpx_free(cm->lf.lfm);
+  cm->lf.lfm = NULL;
+}
+
+
+int vp9_alloc_loop_filter(VP9_COMMON *cm) {
+  vpx_free(cm->lf.lfm);
+  // Each lfm holds bit masks for all the 8x8 blocks in a 64x64 region.  The
+  // stride and rows are rounded up / truncated to a multiple of 8.
+  cm->lf.lfm_stride = (cm->mi_cols + (MI_BLOCK_SIZE - 1)) >> 3;
+  cm->lf.lfm = (LOOP_FILTER_MASK *)vpx_calloc(
+      ((cm->mi_rows + (MI_BLOCK_SIZE - 1)) >> 3) * cm->lf.lfm_stride,
+      sizeof(*cm->lf.lfm));
+  if (!cm->lf.lfm)
+    return 1;
+  return 0;
+}
+
+int vp9_alloc_context_buffers(VP9_COMMON *cm, int width, int height) {
+  int new_mi_size;
+
+  vp9_set_mb_mi(cm, width, height);
+  new_mi_size = cm->mi_stride * calc_mi_size(cm->mi_rows);
+  if (cm->mi_alloc_size < new_mi_size) {
+    cm->free_mi(cm);
+    if (cm->alloc_mi(cm, new_mi_size))
+      goto fail;
+  }
+
+  if (cm->seg_map_alloc_size < cm->mi_rows * cm->mi_cols) {
+    // Create the segmentation map structure and set to 0.
+    free_seg_map(cm);
+    if (alloc_seg_map(cm, cm->mi_rows * cm->mi_cols))
+      goto fail;
+  }
+
+  if (cm->above_context_alloc_cols < cm->mi_cols) {
+    vpx_free(cm->above_context);
+    cm->above_context = (ENTROPY_CONTEXT *)vpx_calloc(
+        2 * mi_cols_aligned_to_sb(cm->mi_cols) * MAX_MB_PLANE,
+        sizeof(*cm->above_context));
+    if (!cm->above_context) goto fail;
+
+    vpx_free(cm->above_seg_context);
+    cm->above_seg_context = (PARTITION_CONTEXT *)vpx_calloc(
+        mi_cols_aligned_to_sb(cm->mi_cols), sizeof(*cm->above_seg_context));
+    if (!cm->above_seg_context) goto fail;
+    cm->above_context_alloc_cols = cm->mi_cols;
+  }
+
+  if (vp9_alloc_loop_filter(cm))
+    goto fail;
+
+  return 0;
+
+ fail:
+  vp9_free_context_buffers(cm);
+  return 1;
+}
+
+void vp9_remove_common(VP9_COMMON *cm) {
+  vp9_free_context_buffers(cm);
+
+  vpx_free(cm->fc);
+  cm->fc = NULL;
+  vpx_free(cm->frame_contexts);
+  cm->frame_contexts = NULL;
+}
+
+void vp9_init_context_buffers(VP9_COMMON *cm) {
+  cm->setup_mi(cm);
+  if (cm->last_frame_seg_map && !cm->frame_parallel_decode)
+    memset(cm->last_frame_seg_map, 0, cm->mi_rows * cm->mi_cols);
+}
+
+void vp9_swap_current_and_last_seg_map(VP9_COMMON *cm) {
+  // Swap indices.
+  const int tmp = cm->seg_map_idx;
+  cm->seg_map_idx = cm->prev_seg_map_idx;
+  cm->prev_seg_map_idx = tmp;
+
+  cm->current_frame_seg_map = cm->seg_map_array[cm->seg_map_idx];
+  cm->last_frame_seg_map = cm->seg_map_array[cm->prev_seg_map_idx];
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_alloccommon.h b/libvpx/libvpx/vp9/common/vp9_alloccommon.h
new file mode 100644
index 0000000..e53955b
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_alloccommon.h
@@ -0,0 +1,45 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_COMMON_VP9_ALLOCCOMMON_H_
+#define VP9_COMMON_VP9_ALLOCCOMMON_H_
+
+#define INVALID_IDX -1  // Invalid buffer index.
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9Common;
+struct BufferPool;
+
+void vp9_remove_common(struct VP9Common *cm);
+
+int vp9_alloc_loop_filter(struct VP9Common *cm);
+int vp9_alloc_context_buffers(struct VP9Common *cm, int width, int height);
+void vp9_init_context_buffers(struct VP9Common *cm);
+void vp9_free_context_buffers(struct VP9Common *cm);
+
+void vp9_free_ref_frame_buffers(struct BufferPool *pool);
+void vp9_free_postproc_buffers(struct VP9Common *cm);
+
+int vp9_alloc_state_buffers(struct VP9Common *cm, int width, int height);
+void vp9_free_state_buffers(struct VP9Common *cm);
+
+void vp9_set_mb_mi(struct VP9Common *cm, int width, int height);
+
+void vp9_swap_current_and_last_seg_map(struct VP9Common *cm);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ALLOCCOMMON_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_blockd.c b/libvpx/libvpx/vp9/common/vp9_blockd.c
new file mode 100644
index 0000000..7bab27d
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_blockd.c
@@ -0,0 +1,135 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_blockd.h"
+
+PREDICTION_MODE vp9_left_block_mode(const MODE_INFO *cur_mi,
+                                    const MODE_INFO *left_mi, int b) {
+  if (b == 0 || b == 2) {
+    if (!left_mi || is_inter_block(left_mi))
+      return DC_PRED;
+
+    return get_y_mode(left_mi, b + 1);
+  } else {
+    assert(b == 1 || b == 3);
+    return cur_mi->bmi[b - 1].as_mode;
+  }
+}
+
+PREDICTION_MODE vp9_above_block_mode(const MODE_INFO *cur_mi,
+                                     const MODE_INFO *above_mi, int b) {
+  if (b == 0 || b == 1) {
+    if (!above_mi || is_inter_block(above_mi))
+      return DC_PRED;
+
+    return get_y_mode(above_mi, b + 2);
+  } else {
+    assert(b == 2 || b == 3);
+    return cur_mi->bmi[b - 2].as_mode;
+  }
+}
+
+void vp9_foreach_transformed_block_in_plane(
+    const MACROBLOCKD *const xd, BLOCK_SIZE bsize, int plane,
+    foreach_transformed_block_visitor visit, void *arg) {
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const MODE_INFO* mi = xd->mi[0];
+  // block and transform sizes, in number of 4x4 blocks log 2 ("*_b")
+  // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8
+  // transform size varies per plane, look it up in a common way.
+  const TX_SIZE tx_size = plane ? get_uv_tx_size(mi, pd)
+                                : mi->tx_size;
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+  const int step = 1 << (tx_size << 1);
+  int i = 0, r, c;
+
+  // If mb_to_right_edge is < 0 we are in a situation in which
+  // the current block size extends into the UMV and we won't
+  // visit the sub blocks that are wholly within the UMV.
+  const int max_blocks_wide = num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 :
+      xd->mb_to_right_edge >> (5 + pd->subsampling_x));
+  const int max_blocks_high = num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 :
+      xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
+  const int extra_step = ((num_4x4_w - max_blocks_wide) >> tx_size) * step;
+
+  // Keep track of the row and column of the blocks we use so that we know
+  // if we are in the unrestricted motion border.
+  for (r = 0; r < max_blocks_high; r += (1 << tx_size)) {
+    // Skip visiting the sub blocks that are wholly within the UMV.
+    for (c = 0; c < max_blocks_wide; c += (1 << tx_size)) {
+      visit(plane, i, plane_bsize, tx_size, arg);
+      i += step;
+    }
+    i += extra_step;
+  }
+}
+
+void vp9_foreach_transformed_block(const MACROBLOCKD* const xd,
+                                   BLOCK_SIZE bsize,
+                                   foreach_transformed_block_visitor visit,
+                                   void *arg) {
+  int plane;
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane)
+    vp9_foreach_transformed_block_in_plane(xd, bsize, plane, visit, arg);
+}
+
+void vp9_set_contexts(const MACROBLOCKD *xd, struct macroblockd_plane *pd,
+                      BLOCK_SIZE plane_bsize, TX_SIZE tx_size, int has_eob,
+                      int aoff, int loff) {
+  ENTROPY_CONTEXT *const a = pd->above_context + aoff;
+  ENTROPY_CONTEXT *const l = pd->left_context + loff;
+  const int tx_size_in_blocks = 1 << tx_size;
+
+  // above
+  if (has_eob && xd->mb_to_right_edge < 0) {
+    int i;
+    const int blocks_wide = num_4x4_blocks_wide_lookup[plane_bsize] +
+                            (xd->mb_to_right_edge >> (5 + pd->subsampling_x));
+    int above_contexts = tx_size_in_blocks;
+    if (above_contexts + aoff > blocks_wide)
+      above_contexts = blocks_wide - aoff;
+
+    for (i = 0; i < above_contexts; ++i)
+      a[i] = has_eob;
+    for (i = above_contexts; i < tx_size_in_blocks; ++i)
+      a[i] = 0;
+  } else {
+    memset(a, has_eob, sizeof(ENTROPY_CONTEXT) * tx_size_in_blocks);
+  }
+
+  // left
+  if (has_eob && xd->mb_to_bottom_edge < 0) {
+    int i;
+    const int blocks_high = num_4x4_blocks_high_lookup[plane_bsize] +
+                            (xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
+    int left_contexts = tx_size_in_blocks;
+    if (left_contexts + loff > blocks_high)
+      left_contexts = blocks_high - loff;
+
+    for (i = 0; i < left_contexts; ++i)
+      l[i] = has_eob;
+    for (i = left_contexts; i < tx_size_in_blocks; ++i)
+      l[i] = 0;
+  } else {
+    memset(l, has_eob, sizeof(ENTROPY_CONTEXT) * tx_size_in_blocks);
+  }
+}
+
+void vp9_setup_block_planes(MACROBLOCKD *xd, int ss_x, int ss_y) {
+  int i;
+
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    xd->plane[i].subsampling_x = i ? ss_x : 0;
+    xd->plane[i].subsampling_y = i ? ss_y : 0;
+  }
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_blockd.h b/libvpx/libvpx/vp9/common/vp9_blockd.h
new file mode 100644
index 0000000..3d26fb2
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_blockd.h
@@ -0,0 +1,305 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_COMMON_VP9_BLOCKD_H_
+#define VP9_COMMON_VP9_BLOCKD_H_
+
+#include "./vpx_config.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+#include "vpx_scale/yv12config.h"
+
+#include "vp9/common/vp9_common_data.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_mv.h"
+#include "vp9/common/vp9_scale.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_tile_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_MB_PLANE 3
+
+typedef enum {
+  KEY_FRAME = 0,
+  INTER_FRAME = 1,
+  FRAME_TYPES,
+} FRAME_TYPE;
+
+static INLINE int is_inter_mode(PREDICTION_MODE mode) {
+  return mode >= NEARESTMV && mode <= NEWMV;
+}
+
+/* For keyframes, intra block modes are predicted by the (already decoded)
+   modes for the Y blocks to the left and above us; for interframes, there
+   is a single probability table. */
+
+typedef struct {
+  PREDICTION_MODE as_mode;
+  int_mv as_mv[2];  // first, second inter predictor motion vectors
+} b_mode_info;
+
+// Note that the rate-distortion optimization loop, bit-stream writer, and
+// decoder implementation modules critically rely on the defined entry values
+// specified herein. They should be refactored concurrently.
+
+#define NONE           -1
+#define INTRA_FRAME     0
+#define LAST_FRAME      1
+#define GOLDEN_FRAME    2
+#define ALTREF_FRAME    3
+#define MAX_REF_FRAMES  4
+typedef int8_t MV_REFERENCE_FRAME;
+
+// This structure now relates to 8x8 block regions.
+typedef struct MODE_INFO {
+  // Common for both INTER and INTRA blocks
+  BLOCK_SIZE sb_type;
+  PREDICTION_MODE mode;
+  TX_SIZE tx_size;
+  int8_t skip;
+  int8_t segment_id;
+  int8_t seg_id_predicted;  // valid only when temporal_update is enabled
+
+  // Only for INTRA blocks
+  PREDICTION_MODE uv_mode;
+
+  // Only for INTER blocks
+  INTERP_FILTER interp_filter;
+  MV_REFERENCE_FRAME ref_frame[2];
+
+  // TODO(slavarnway): Delete and use bmi[3].as_mv[] instead.
+  int_mv mv[2];
+
+  b_mode_info bmi[4];
+} MODE_INFO;
+
+static INLINE PREDICTION_MODE get_y_mode(const MODE_INFO *mi, int block) {
+  return mi->sb_type < BLOCK_8X8 ? mi->bmi[block].as_mode
+                                 : mi->mode;
+}
+
+static INLINE int is_inter_block(const MODE_INFO *mi) {
+  return mi->ref_frame[0] > INTRA_FRAME;
+}
+
+static INLINE int has_second_ref(const MODE_INFO *mi) {
+  return mi->ref_frame[1] > INTRA_FRAME;
+}
+
+PREDICTION_MODE vp9_left_block_mode(const MODE_INFO *cur_mi,
+                                    const MODE_INFO *left_mi, int b);
+
+PREDICTION_MODE vp9_above_block_mode(const MODE_INFO *cur_mi,
+                                     const MODE_INFO *above_mi, int b);
+
+enum mv_precision {
+  MV_PRECISION_Q3,
+  MV_PRECISION_Q4
+};
+
+struct buf_2d {
+  uint8_t *buf;
+  int stride;
+};
+
+struct macroblockd_plane {
+  tran_low_t *dqcoeff;
+  int subsampling_x;
+  int subsampling_y;
+  struct buf_2d dst;
+  struct buf_2d pre[2];
+  ENTROPY_CONTEXT *above_context;
+  ENTROPY_CONTEXT *left_context;
+  int16_t seg_dequant[MAX_SEGMENTS][2];
+
+  // number of 4x4s in current block
+  uint16_t n4_w, n4_h;
+  // log2 of n4_w, n4_h
+  uint8_t n4_wl, n4_hl;
+
+  // encoder
+  const int16_t *dequant;
+};
+
+#define BLOCK_OFFSET(x, i) ((x) + (i) * 16)
+
+typedef struct RefBuffer {
+  // TODO(dkovalev): idx is not really required and should be removed, now it
+  // is used in vp9_onyxd_if.c
+  int idx;
+  YV12_BUFFER_CONFIG *buf;
+  struct scale_factors sf;
+} RefBuffer;
+
+typedef struct macroblockd {
+  struct macroblockd_plane plane[MAX_MB_PLANE];
+  uint8_t bmode_blocks_wl;
+  uint8_t bmode_blocks_hl;
+
+  FRAME_COUNTS *counts;
+  TileInfo tile;
+
+  int mi_stride;
+
+  MODE_INFO **mi;
+  MODE_INFO *left_mi;
+  MODE_INFO *above_mi;
+
+  unsigned int max_blocks_wide;
+  unsigned int max_blocks_high;
+
+  const vpx_prob (*partition_probs)[PARTITION_TYPES - 1];
+
+  /* Distance of MB away from frame edges */
+  int mb_to_left_edge;
+  int mb_to_right_edge;
+  int mb_to_top_edge;
+  int mb_to_bottom_edge;
+
+  FRAME_CONTEXT *fc;
+
+  /* pointers to reference frames */
+  RefBuffer *block_refs[2];
+
+  /* pointer to current frame */
+  const YV12_BUFFER_CONFIG *cur_buf;
+
+  ENTROPY_CONTEXT *above_context[MAX_MB_PLANE];
+  ENTROPY_CONTEXT left_context[MAX_MB_PLANE][16];
+
+  PARTITION_CONTEXT *above_seg_context;
+  PARTITION_CONTEXT left_seg_context[8];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  /* Bit depth: 8, 10, 12 */
+  int bd;
+#endif
+
+  int lossless;
+  int corrupted;
+
+  struct vpx_internal_error_info *error_info;
+} MACROBLOCKD;
+
+static INLINE PLANE_TYPE get_plane_type(int plane) {
+  return (PLANE_TYPE)(plane > 0);
+}
+
+static INLINE BLOCK_SIZE get_subsize(BLOCK_SIZE bsize,
+                                     PARTITION_TYPE partition) {
+  return subsize_lookup[partition][bsize];
+}
+
+extern const TX_TYPE intra_mode_to_tx_type_lookup[INTRA_MODES];
+
+static INLINE TX_TYPE get_tx_type(PLANE_TYPE plane_type,
+                                  const MACROBLOCKD *xd) {
+  const MODE_INFO *const mi = xd->mi[0];
+
+  if (plane_type != PLANE_TYPE_Y || xd->lossless || is_inter_block(mi))
+    return DCT_DCT;
+
+  return intra_mode_to_tx_type_lookup[mi->mode];
+}
+
+static INLINE TX_TYPE get_tx_type_4x4(PLANE_TYPE plane_type,
+                                      const MACROBLOCKD *xd, int ib) {
+  const MODE_INFO *const mi = xd->mi[0];
+
+  if (plane_type != PLANE_TYPE_Y || xd->lossless || is_inter_block(mi))
+    return DCT_DCT;
+
+  return intra_mode_to_tx_type_lookup[get_y_mode(mi, ib)];
+}
+
+void vp9_setup_block_planes(MACROBLOCKD *xd, int ss_x, int ss_y);
+
+static INLINE TX_SIZE get_uv_tx_size_impl(TX_SIZE y_tx_size, BLOCK_SIZE bsize,
+                                          int xss, int yss) {
+  if (bsize < BLOCK_8X8) {
+    return TX_4X4;
+  } else {
+    const BLOCK_SIZE plane_bsize = ss_size_lookup[bsize][xss][yss];
+    return VPXMIN(y_tx_size, max_txsize_lookup[plane_bsize]);
+  }
+}
+
+static INLINE TX_SIZE get_uv_tx_size(const MODE_INFO *mi,
+                                     const struct macroblockd_plane *pd) {
+  return get_uv_tx_size_impl(mi->tx_size, mi->sb_type, pd->subsampling_x,
+                             pd->subsampling_y);
+}
+
+static INLINE BLOCK_SIZE get_plane_block_size(BLOCK_SIZE bsize,
+    const struct macroblockd_plane *pd) {
+  return ss_size_lookup[bsize][pd->subsampling_x][pd->subsampling_y];
+}
+
+static INLINE void reset_skip_context(MACROBLOCKD *xd, BLOCK_SIZE bsize) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    struct macroblockd_plane *const pd = &xd->plane[i];
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+    memset(pd->above_context, 0,
+           sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide_lookup[plane_bsize]);
+    memset(pd->left_context, 0,
+           sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high_lookup[plane_bsize]);
+  }
+}
+
+static INLINE const vpx_prob *get_y_mode_probs(const MODE_INFO *mi,
+                                               const MODE_INFO *above_mi,
+                                               const MODE_INFO *left_mi,
+                                               int block) {
+  const PREDICTION_MODE above = vp9_above_block_mode(mi, above_mi, block);
+  const PREDICTION_MODE left = vp9_left_block_mode(mi, left_mi, block);
+  return vp9_kf_y_mode_prob[above][left];
+}
+
+typedef void (*foreach_transformed_block_visitor)(int plane, int block,
+                                                  BLOCK_SIZE plane_bsize,
+                                                  TX_SIZE tx_size,
+                                                  void *arg);
+
+void vp9_foreach_transformed_block_in_plane(
+    const MACROBLOCKD *const xd, BLOCK_SIZE bsize, int plane,
+    foreach_transformed_block_visitor visit, void *arg);
+
+
+void vp9_foreach_transformed_block(
+    const MACROBLOCKD* const xd, BLOCK_SIZE bsize,
+    foreach_transformed_block_visitor visit, void *arg);
+
+static INLINE void txfrm_block_to_raster_xy(BLOCK_SIZE plane_bsize,
+                                            TX_SIZE tx_size, int block,
+                                            int *x, int *y) {
+  const int bwl = b_width_log2_lookup[plane_bsize];
+  const int tx_cols_log2 = bwl - tx_size;
+  const int tx_cols = 1 << tx_cols_log2;
+  const int raster_mb = block >> (tx_size << 1);
+  *x = (raster_mb & (tx_cols - 1)) << tx_size;
+  *y = (raster_mb >> tx_cols_log2) << tx_size;
+}
+
+void vp9_set_contexts(const MACROBLOCKD *xd, struct macroblockd_plane *pd,
+                      BLOCK_SIZE plane_bsize, TX_SIZE tx_size, int has_eob,
+                      int aoff, int loff);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_BLOCKD_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_common.h b/libvpx/libvpx/vp9/common/vp9_common.h
new file mode 100644
index 0000000..908fa80
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_common.h
@@ -0,0 +1,74 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_COMMON_H_
+#define VP9_COMMON_VP9_COMMON_H_
+
+/* Interface header for common constant data structures and lookup tables */
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/bitops.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Only need this for fixed-size arrays, for structs just assign.
+#define vp9_copy(dest, src) {            \
+    assert(sizeof(dest) == sizeof(src)); \
+    memcpy(dest, src, sizeof(src));  \
+  }
+
+// Use this for variably-sized arrays.
+#define vp9_copy_array(dest, src, n) {       \
+    assert(sizeof(*dest) == sizeof(*src));   \
+    memcpy(dest, src, n * sizeof(*src)); \
+  }
+
+#define vp9_zero(dest) memset(&(dest), 0, sizeof(dest))
+#define vp9_zero_array(dest, n) memset(dest, 0, n * sizeof(*dest))
+
+static INLINE int get_unsigned_bits(unsigned int num_values) {
+  return num_values > 0 ? get_msb(num_values) + 1 : 0;
+}
+
+#if CONFIG_DEBUG
+#define CHECK_MEM_ERROR(cm, lval, expr) do { \
+  lval = (expr); \
+  if (!lval) \
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, \
+                       "Failed to allocate "#lval" at %s:%d", \
+                       __FILE__, __LINE__); \
+  } while (0)
+#else
+#define CHECK_MEM_ERROR(cm, lval, expr) do { \
+  lval = (expr); \
+  if (!lval) \
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, \
+                       "Failed to allocate "#lval); \
+  } while (0)
+#endif
+
+#define VP9_SYNC_CODE_0 0x49
+#define VP9_SYNC_CODE_1 0x83
+#define VP9_SYNC_CODE_2 0x42
+
+#define VP9_FRAME_MARKER 0x2
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_COMMON_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_common_data.c b/libvpx/libvpx/vp9/common/vp9_common_data.c
new file mode 100644
index 0000000..3409d04
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_common_data.c
@@ -0,0 +1,176 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_common_data.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+// Log 2 conversion lookup tables for block width and height
+const uint8_t b_width_log2_lookup[BLOCK_SIZES] =
+  {0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4};
+const uint8_t b_height_log2_lookup[BLOCK_SIZES] =
+  {0, 1, 0, 1, 2, 1, 2, 3, 2, 3, 4, 3, 4};
+const uint8_t num_4x4_blocks_wide_lookup[BLOCK_SIZES] =
+  {1, 1, 2, 2, 2, 4, 4, 4, 8, 8, 8, 16, 16};
+const uint8_t num_4x4_blocks_high_lookup[BLOCK_SIZES] =
+  {1, 2, 1, 2, 4, 2, 4, 8, 4, 8, 16, 8, 16};
+// Log 2 conversion lookup tables for modeinfo width and height
+const uint8_t mi_width_log2_lookup[BLOCK_SIZES] =
+  {0, 0, 0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3};
+const uint8_t num_8x8_blocks_wide_lookup[BLOCK_SIZES] =
+  {1, 1, 1, 1, 1, 2, 2, 2, 4, 4, 4, 8, 8};
+const uint8_t num_8x8_blocks_high_lookup[BLOCK_SIZES] =
+  {1, 1, 1, 1, 2, 1, 2, 4, 2, 4, 8, 4, 8};
+
+// VPXMIN(3, VPXMIN(b_width_log2(bsize), b_height_log2(bsize)))
+const uint8_t size_group_lookup[BLOCK_SIZES] =
+  {0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 3};
+
+const uint8_t num_pels_log2_lookup[BLOCK_SIZES] =
+  {4, 5, 5, 6, 7, 7, 8, 9, 9, 10, 11, 11, 12};
+
+const PARTITION_TYPE partition_lookup[][BLOCK_SIZES] = {
+  {  // 4X4
+    // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
+    PARTITION_NONE, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID
+  }, {  // 8X8
+    // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
+    PARTITION_SPLIT, PARTITION_VERT, PARTITION_HORZ, PARTITION_NONE,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID
+  }, {  // 16X16
+    // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT,
+    PARTITION_VERT, PARTITION_HORZ, PARTITION_NONE, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID
+  }, {  // 32X32
+    // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT,
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_VERT,
+    PARTITION_HORZ, PARTITION_NONE, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID
+  }, {  // 64X64
+    // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT,
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT,
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_VERT, PARTITION_HORZ,
+    PARTITION_NONE
+  }
+};
+
+const BLOCK_SIZE subsize_lookup[PARTITION_TYPES][BLOCK_SIZES] = {
+  {     // PARTITION_NONE
+    BLOCK_4X4,   BLOCK_4X8,   BLOCK_8X4,
+    BLOCK_8X8,   BLOCK_8X16,  BLOCK_16X8,
+    BLOCK_16X16, BLOCK_16X32, BLOCK_32X16,
+    BLOCK_32X32, BLOCK_32X64, BLOCK_64X32,
+    BLOCK_64X64,
+  }, {  // PARTITION_HORZ
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_8X4,     BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_16X8,    BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_32X16,   BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_64X32,
+  }, {  // PARTITION_VERT
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_4X8,     BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_8X16,    BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_16X32,   BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_32X64,
+  }, {  // PARTITION_SPLIT
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_4X4,     BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_8X8,     BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_16X16,   BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_32X32,
+  }
+};
+
+const TX_SIZE max_txsize_lookup[BLOCK_SIZES] = {
+  TX_4X4,   TX_4X4,   TX_4X4,
+  TX_8X8,   TX_8X8,   TX_8X8,
+  TX_16X16, TX_16X16, TX_16X16,
+  TX_32X32, TX_32X32, TX_32X32, TX_32X32
+};
+
+const BLOCK_SIZE txsize_to_bsize[TX_SIZES] = {
+    BLOCK_4X4,  // TX_4X4
+    BLOCK_8X8,  // TX_8X8
+    BLOCK_16X16,  // TX_16X16
+    BLOCK_32X32,  // TX_32X32
+};
+
+const TX_SIZE tx_mode_to_biggest_tx_size[TX_MODES] = {
+  TX_4X4,  // ONLY_4X4
+  TX_8X8,  // ALLOW_8X8
+  TX_16X16,  // ALLOW_16X16
+  TX_32X32,  // ALLOW_32X32
+  TX_32X32,  // TX_MODE_SELECT
+};
+
+const BLOCK_SIZE ss_size_lookup[BLOCK_SIZES][2][2] = {
+//  ss_x == 0    ss_x == 0        ss_x == 1      ss_x == 1
+//  ss_y == 0    ss_y == 1        ss_y == 0      ss_y == 1
+  {{BLOCK_4X4,   BLOCK_INVALID}, {BLOCK_INVALID, BLOCK_INVALID}},
+  {{BLOCK_4X8,   BLOCK_4X4},     {BLOCK_INVALID, BLOCK_INVALID}},
+  {{BLOCK_8X4,   BLOCK_INVALID}, {BLOCK_4X4,     BLOCK_INVALID}},
+  {{BLOCK_8X8,   BLOCK_8X4},     {BLOCK_4X8,     BLOCK_4X4}},
+  {{BLOCK_8X16,  BLOCK_8X8},     {BLOCK_INVALID, BLOCK_4X8}},
+  {{BLOCK_16X8,  BLOCK_INVALID}, {BLOCK_8X8,     BLOCK_8X4}},
+  {{BLOCK_16X16, BLOCK_16X8},    {BLOCK_8X16,    BLOCK_8X8}},
+  {{BLOCK_16X32, BLOCK_16X16},   {BLOCK_INVALID, BLOCK_8X16}},
+  {{BLOCK_32X16, BLOCK_INVALID}, {BLOCK_16X16,   BLOCK_16X8}},
+  {{BLOCK_32X32, BLOCK_32X16},   {BLOCK_16X32,   BLOCK_16X16}},
+  {{BLOCK_32X64, BLOCK_32X32},   {BLOCK_INVALID, BLOCK_16X32}},
+  {{BLOCK_64X32, BLOCK_INVALID}, {BLOCK_32X32,   BLOCK_32X16}},
+  {{BLOCK_64X64, BLOCK_64X32},   {BLOCK_32X64,   BLOCK_32X32}},
+};
+
+// Generates 4 bit field in which each bit set to 1 represents
+// a blocksize partition  1111 means we split 64x64, 32x32, 16x16
+// and 8x8.  1000 means we just split the 64x64 to 32x32
+const struct {
+  PARTITION_CONTEXT above;
+  PARTITION_CONTEXT left;
+} partition_context_lookup[BLOCK_SIZES]= {
+  {15, 15},  // 4X4   - {0b1111, 0b1111}
+  {15, 14},  // 4X8   - {0b1111, 0b1110}
+  {14, 15},  // 8X4   - {0b1110, 0b1111}
+  {14, 14},  // 8X8   - {0b1110, 0b1110}
+  {14, 12},  // 8X16  - {0b1110, 0b1100}
+  {12, 14},  // 16X8  - {0b1100, 0b1110}
+  {12, 12},  // 16X16 - {0b1100, 0b1100}
+  {12, 8 },  // 16X32 - {0b1100, 0b1000}
+  {8,  12},  // 32X16 - {0b1000, 0b1100}
+  {8,  8 },  // 32X32 - {0b1000, 0b1000}
+  {8,  0 },  // 32X64 - {0b1000, 0b0000}
+  {0,  8 },  // 64X32 - {0b0000, 0b1000}
+  {0,  0 },  // 64X64 - {0b0000, 0b0000}
+};
+
+#if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
+const uint8_t need_top_left[INTRA_MODES] = {
+    0,  // DC_PRED
+    0,  // V_PRED
+    0,  // H_PRED
+    0,  // D45_PRED
+    1,  // D135_PRED
+    1,  // D117_PRED
+    1,  // D153_PRED
+    0,  // D207_PRED
+    0,  // D63_PRED
+    1,  // TM_PRED
+};
+#endif  // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vp9/common/vp9_common_data.h b/libvpx/libvpx/vp9/common/vp9_common_data.h
new file mode 100644
index 0000000..0ae24da
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_common_data.h
@@ -0,0 +1,44 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_COMMON_DATA_H_
+#define VP9_COMMON_VP9_COMMON_DATA_H_
+
+#include "vp9/common/vp9_enums.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern const uint8_t b_width_log2_lookup[BLOCK_SIZES];
+extern const uint8_t b_height_log2_lookup[BLOCK_SIZES];
+extern const uint8_t mi_width_log2_lookup[BLOCK_SIZES];
+extern const uint8_t num_8x8_blocks_wide_lookup[BLOCK_SIZES];
+extern const uint8_t num_8x8_blocks_high_lookup[BLOCK_SIZES];
+extern const uint8_t num_4x4_blocks_high_lookup[BLOCK_SIZES];
+extern const uint8_t num_4x4_blocks_wide_lookup[BLOCK_SIZES];
+extern const uint8_t size_group_lookup[BLOCK_SIZES];
+extern const uint8_t num_pels_log2_lookup[BLOCK_SIZES];
+extern const PARTITION_TYPE partition_lookup[][BLOCK_SIZES];
+extern const BLOCK_SIZE subsize_lookup[PARTITION_TYPES][BLOCK_SIZES];
+extern const TX_SIZE max_txsize_lookup[BLOCK_SIZES];
+extern const BLOCK_SIZE txsize_to_bsize[TX_SIZES];
+extern const TX_SIZE tx_mode_to_biggest_tx_size[TX_MODES];
+extern const BLOCK_SIZE ss_size_lookup[BLOCK_SIZES][2][2];
+#if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
+extern const uint8_t need_top_left[INTRA_MODES];
+#endif  // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_COMMON_DATA_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_debugmodes.c b/libvpx/libvpx/vp9/common/vp9_debugmodes.c
new file mode 100644
index 0000000..d9c1fd9
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_debugmodes.c
@@ -0,0 +1,91 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdio.h>
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_onyxc_int.h"
+
+static void log_frame_info(VP9_COMMON *cm, const char *str, FILE *f) {
+  fprintf(f, "%s", str);
+  fprintf(f, "(Frame %d, Show:%d, Q:%d): \n", cm->current_video_frame,
+          cm->show_frame, cm->base_qindex);
+}
+/* This function dereferences a pointer to the mbmi structure
+ * and uses the passed in member offset to print out the value of an integer
+ * for each mbmi member value in the mi structure.
+ */
+static void print_mi_data(VP9_COMMON *cm, FILE *file, const char *descriptor,
+                          size_t member_offset) {
+  int mi_row, mi_col;
+  MODE_INFO **mi = cm->mi_grid_visible;
+  int rows = cm->mi_rows;
+  int cols = cm->mi_cols;
+  char prefix = descriptor[0];
+
+  log_frame_info(cm, descriptor, file);
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    fprintf(file, "%c ", prefix);
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      fprintf(file, "%2d ",
+              *((int*) ((char *) (mi[0]) +
+                                  member_offset)));
+      mi++;
+    }
+    fprintf(file, "\n");
+    mi += 8;
+  }
+  fprintf(file, "\n");
+}
+
+void vp9_print_modes_and_motion_vectors(VP9_COMMON *cm, const char *file) {
+  int mi_row;
+  int mi_col;
+  FILE *mvs = fopen(file, "a");
+  MODE_INFO **mi = cm->mi_grid_visible;
+  int rows = cm->mi_rows;
+  int cols = cm->mi_cols;
+
+  print_mi_data(cm, mvs, "Partitions:", offsetof(MODE_INFO, sb_type));
+  print_mi_data(cm, mvs, "Modes:", offsetof(MODE_INFO, mode));
+  print_mi_data(cm, mvs, "Ref frame:", offsetof(MODE_INFO, ref_frame[0]));
+  print_mi_data(cm, mvs, "Transform:", offsetof(MODE_INFO, tx_size));
+  print_mi_data(cm, mvs, "UV Modes:", offsetof(MODE_INFO, uv_mode));
+
+  // output skip infomation.
+  log_frame_info(cm, "Skips:", mvs);
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    fprintf(mvs, "S ");
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      fprintf(mvs, "%2d ", mi[0]->skip);
+      mi++;
+    }
+    fprintf(mvs, "\n");
+    mi += 8;
+  }
+  fprintf(mvs, "\n");
+
+  // output motion vectors.
+  log_frame_info(cm, "Vectors ", mvs);
+  mi = cm->mi_grid_visible;
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    fprintf(mvs, "V ");
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      fprintf(mvs, "%4d:%4d ", mi[0]->mv[0].as_mv.row,
+                               mi[0]->mv[0].as_mv.col);
+      mi++;
+    }
+    fprintf(mvs, "\n");
+    mi += 8;
+  }
+  fprintf(mvs, "\n");
+
+  fclose(mvs);
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_entropy.c b/libvpx/libvpx/vp9/common/vp9_entropy.c
new file mode 100644
index 0000000..7b490af
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_entropy.c
@@ -0,0 +1,802 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx/vpx_integer.h"
+
+// Unconstrained Node Tree
+const vpx_tree_index vp9_coef_con_tree[TREE_SIZE(ENTROPY_TOKENS)] = {
+  2, 6,                                // 0 = LOW_VAL
+  -TWO_TOKEN, 4,                       // 1 = TWO
+  -THREE_TOKEN, -FOUR_TOKEN,           // 2 = THREE
+  8, 10,                               // 3 = HIGH_LOW
+  -CATEGORY1_TOKEN, -CATEGORY2_TOKEN,  // 4 = CAT_ONE
+  12, 14,                              // 5 = CAT_THREEFOUR
+  -CATEGORY3_TOKEN, -CATEGORY4_TOKEN,  // 6 = CAT_THREE
+  -CATEGORY5_TOKEN, -CATEGORY6_TOKEN   // 7 = CAT_FIVE
+};
+
+const vpx_prob vp9_cat1_prob[] = { 159 };
+const vpx_prob vp9_cat2_prob[] = { 165, 145 };
+const vpx_prob vp9_cat3_prob[] = { 173, 148, 140 };
+const vpx_prob vp9_cat4_prob[] = { 176, 155, 140, 135 };
+const vpx_prob vp9_cat5_prob[] = { 180, 157, 141, 134, 130 };
+const vpx_prob vp9_cat6_prob[] = {
+    254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129
+};
+#if CONFIG_VP9_HIGHBITDEPTH
+const vpx_prob vp9_cat6_prob_high12[] = {
+    255, 255, 255, 255, 254, 254, 254, 252, 249,
+    243, 230, 196, 177, 153, 140, 133, 130, 129
+};
+#endif
+
+const uint8_t vp9_coefband_trans_8x8plus[1024] = {
+  0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4,
+  4, 4, 4, 4, 4, 5,
+  // beyond MAXBAND_INDEX+1 all values are filled as 5
+                    5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+};
+
+const uint8_t vp9_coefband_trans_4x4[16] = {
+  0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 5, 5, 5,
+};
+
+const uint8_t vp9_pt_energy_class[ENTROPY_TOKENS] = {
+  0, 1, 2, 3, 3, 4, 4, 5, 5, 5, 5, 5
+};
+
+// Model obtained from a 2-sided zero-centerd distribuition derived
+// from a Pareto distribution. The cdf of the distribution is:
+// cdf(x) = 0.5 + 0.5 * sgn(x) * [1 - {alpha/(alpha + |x|)} ^ beta]
+//
+// For a given beta and a given probablity of the 1-node, the alpha
+// is first solved, and then the {alpha, beta} pair is used to generate
+// the probabilities for the rest of the nodes.
+
+// beta = 8
+
+// Every odd line in this table can be generated from the even lines
+// by averaging :
+// vp9_pareto8_full[l][node] = (vp9_pareto8_full[l-1][node] +
+//                              vp9_pareto8_full[l+1][node] ) >> 1;
+const vpx_prob vp9_pareto8_full[COEFF_PROB_MODELS][MODEL_NODES] = {
+  {  3,  86, 128,   6,  86,  23,  88,  29},
+  {  6,  86, 128,  11,  87,  42,  91,  52},
+  {  9,  86, 129,  17,  88,  61,  94,  76},
+  { 12,  86, 129,  22,  88,  77,  97,  93},
+  { 15,  87, 129,  28,  89,  93, 100, 110},
+  { 17,  87, 129,  33,  90, 105, 103, 123},
+  { 20,  88, 130,  38,  91, 118, 106, 136},
+  { 23,  88, 130,  43,  91, 128, 108, 146},
+  { 26,  89, 131,  48,  92, 139, 111, 156},
+  { 28,  89, 131,  53,  93, 147, 114, 163},
+  { 31,  90, 131,  58,  94, 156, 117, 171},
+  { 34,  90, 131,  62,  94, 163, 119, 177},
+  { 37,  90, 132,  66,  95, 171, 122, 184},
+  { 39,  90, 132,  70,  96, 177, 124, 189},
+  { 42,  91, 132,  75,  97, 183, 127, 194},
+  { 44,  91, 132,  79,  97, 188, 129, 198},
+  { 47,  92, 133,  83,  98, 193, 132, 202},
+  { 49,  92, 133,  86,  99, 197, 134, 205},
+  { 52,  93, 133,  90, 100, 201, 137, 208},
+  { 54,  93, 133,  94, 100, 204, 139, 211},
+  { 57,  94, 134,  98, 101, 208, 142, 214},
+  { 59,  94, 134, 101, 102, 211, 144, 216},
+  { 62,  94, 135, 105, 103, 214, 146, 218},
+  { 64,  94, 135, 108, 103, 216, 148, 220},
+  { 66,  95, 135, 111, 104, 219, 151, 222},
+  { 68,  95, 135, 114, 105, 221, 153, 223},
+  { 71,  96, 136, 117, 106, 224, 155, 225},
+  { 73,  96, 136, 120, 106, 225, 157, 226},
+  { 76,  97, 136, 123, 107, 227, 159, 228},
+  { 78,  97, 136, 126, 108, 229, 160, 229},
+  { 80,  98, 137, 129, 109, 231, 162, 231},
+  { 82,  98, 137, 131, 109, 232, 164, 232},
+  { 84,  98, 138, 134, 110, 234, 166, 233},
+  { 86,  98, 138, 137, 111, 235, 168, 234},
+  { 89,  99, 138, 140, 112, 236, 170, 235},
+  { 91,  99, 138, 142, 112, 237, 171, 235},
+  { 93, 100, 139, 145, 113, 238, 173, 236},
+  { 95, 100, 139, 147, 114, 239, 174, 237},
+  { 97, 101, 140, 149, 115, 240, 176, 238},
+  { 99, 101, 140, 151, 115, 241, 177, 238},
+  {101, 102, 140, 154, 116, 242, 179, 239},
+  {103, 102, 140, 156, 117, 242, 180, 239},
+  {105, 103, 141, 158, 118, 243, 182, 240},
+  {107, 103, 141, 160, 118, 243, 183, 240},
+  {109, 104, 141, 162, 119, 244, 185, 241},
+  {111, 104, 141, 164, 119, 244, 186, 241},
+  {113, 104, 142, 166, 120, 245, 187, 242},
+  {114, 104, 142, 168, 121, 245, 188, 242},
+  {116, 105, 143, 170, 122, 246, 190, 243},
+  {118, 105, 143, 171, 122, 246, 191, 243},
+  {120, 106, 143, 173, 123, 247, 192, 244},
+  {121, 106, 143, 175, 124, 247, 193, 244},
+  {123, 107, 144, 177, 125, 248, 195, 244},
+  {125, 107, 144, 178, 125, 248, 196, 244},
+  {127, 108, 145, 180, 126, 249, 197, 245},
+  {128, 108, 145, 181, 127, 249, 198, 245},
+  {130, 109, 145, 183, 128, 249, 199, 245},
+  {132, 109, 145, 184, 128, 249, 200, 245},
+  {134, 110, 146, 186, 129, 250, 201, 246},
+  {135, 110, 146, 187, 130, 250, 202, 246},
+  {137, 111, 147, 189, 131, 251, 203, 246},
+  {138, 111, 147, 190, 131, 251, 204, 246},
+  {140, 112, 147, 192, 132, 251, 205, 247},
+  {141, 112, 147, 193, 132, 251, 206, 247},
+  {143, 113, 148, 194, 133, 251, 207, 247},
+  {144, 113, 148, 195, 134, 251, 207, 247},
+  {146, 114, 149, 197, 135, 252, 208, 248},
+  {147, 114, 149, 198, 135, 252, 209, 248},
+  {149, 115, 149, 199, 136, 252, 210, 248},
+  {150, 115, 149, 200, 137, 252, 210, 248},
+  {152, 115, 150, 201, 138, 252, 211, 248},
+  {153, 115, 150, 202, 138, 252, 212, 248},
+  {155, 116, 151, 204, 139, 253, 213, 249},
+  {156, 116, 151, 205, 139, 253, 213, 249},
+  {158, 117, 151, 206, 140, 253, 214, 249},
+  {159, 117, 151, 207, 141, 253, 215, 249},
+  {161, 118, 152, 208, 142, 253, 216, 249},
+  {162, 118, 152, 209, 142, 253, 216, 249},
+  {163, 119, 153, 210, 143, 253, 217, 249},
+  {164, 119, 153, 211, 143, 253, 217, 249},
+  {166, 120, 153, 212, 144, 254, 218, 250},
+  {167, 120, 153, 212, 145, 254, 219, 250},
+  {168, 121, 154, 213, 146, 254, 220, 250},
+  {169, 121, 154, 214, 146, 254, 220, 250},
+  {171, 122, 155, 215, 147, 254, 221, 250},
+  {172, 122, 155, 216, 147, 254, 221, 250},
+  {173, 123, 155, 217, 148, 254, 222, 250},
+  {174, 123, 155, 217, 149, 254, 222, 250},
+  {176, 124, 156, 218, 150, 254, 223, 250},
+  {177, 124, 156, 219, 150, 254, 223, 250},
+  {178, 125, 157, 220, 151, 254, 224, 251},
+  {179, 125, 157, 220, 151, 254, 224, 251},
+  {180, 126, 157, 221, 152, 254, 225, 251},
+  {181, 126, 157, 221, 152, 254, 225, 251},
+  {183, 127, 158, 222, 153, 254, 226, 251},
+  {184, 127, 158, 223, 154, 254, 226, 251},
+  {185, 128, 159, 224, 155, 255, 227, 251},
+  {186, 128, 159, 224, 155, 255, 227, 251},
+  {187, 129, 160, 225, 156, 255, 228, 251},
+  {188, 130, 160, 225, 156, 255, 228, 251},
+  {189, 131, 160, 226, 157, 255, 228, 251},
+  {190, 131, 160, 226, 158, 255, 228, 251},
+  {191, 132, 161, 227, 159, 255, 229, 251},
+  {192, 132, 161, 227, 159, 255, 229, 251},
+  {193, 133, 162, 228, 160, 255, 230, 252},
+  {194, 133, 162, 229, 160, 255, 230, 252},
+  {195, 134, 163, 230, 161, 255, 231, 252},
+  {196, 134, 163, 230, 161, 255, 231, 252},
+  {197, 135, 163, 231, 162, 255, 231, 252},
+  {198, 135, 163, 231, 162, 255, 231, 252},
+  {199, 136, 164, 232, 163, 255, 232, 252},
+  {200, 136, 164, 232, 164, 255, 232, 252},
+  {201, 137, 165, 233, 165, 255, 233, 252},
+  {201, 137, 165, 233, 165, 255, 233, 252},
+  {202, 138, 166, 233, 166, 255, 233, 252},
+  {203, 138, 166, 233, 166, 255, 233, 252},
+  {204, 139, 166, 234, 167, 255, 234, 252},
+  {205, 139, 166, 234, 167, 255, 234, 252},
+  {206, 140, 167, 235, 168, 255, 235, 252},
+  {206, 140, 167, 235, 168, 255, 235, 252},
+  {207, 141, 168, 236, 169, 255, 235, 252},
+  {208, 141, 168, 236, 170, 255, 235, 252},
+  {209, 142, 169, 237, 171, 255, 236, 252},
+  {209, 143, 169, 237, 171, 255, 236, 252},
+  {210, 144, 169, 237, 172, 255, 236, 252},
+  {211, 144, 169, 237, 172, 255, 236, 252},
+  {212, 145, 170, 238, 173, 255, 237, 252},
+  {213, 145, 170, 238, 173, 255, 237, 252},
+  {214, 146, 171, 239, 174, 255, 237, 253},
+  {214, 146, 171, 239, 174, 255, 237, 253},
+  {215, 147, 172, 240, 175, 255, 238, 253},
+  {215, 147, 172, 240, 175, 255, 238, 253},
+  {216, 148, 173, 240, 176, 255, 238, 253},
+  {217, 148, 173, 240, 176, 255, 238, 253},
+  {218, 149, 173, 241, 177, 255, 239, 253},
+  {218, 149, 173, 241, 178, 255, 239, 253},
+  {219, 150, 174, 241, 179, 255, 239, 253},
+  {219, 151, 174, 241, 179, 255, 239, 253},
+  {220, 152, 175, 242, 180, 255, 240, 253},
+  {221, 152, 175, 242, 180, 255, 240, 253},
+  {222, 153, 176, 242, 181, 255, 240, 253},
+  {222, 153, 176, 242, 181, 255, 240, 253},
+  {223, 154, 177, 243, 182, 255, 240, 253},
+  {223, 154, 177, 243, 182, 255, 240, 253},
+  {224, 155, 178, 244, 183, 255, 241, 253},
+  {224, 155, 178, 244, 183, 255, 241, 253},
+  {225, 156, 178, 244, 184, 255, 241, 253},
+  {225, 157, 178, 244, 184, 255, 241, 253},
+  {226, 158, 179, 244, 185, 255, 242, 253},
+  {227, 158, 179, 244, 185, 255, 242, 253},
+  {228, 159, 180, 245, 186, 255, 242, 253},
+  {228, 159, 180, 245, 186, 255, 242, 253},
+  {229, 160, 181, 245, 187, 255, 242, 253},
+  {229, 160, 181, 245, 187, 255, 242, 253},
+  {230, 161, 182, 246, 188, 255, 243, 253},
+  {230, 162, 182, 246, 188, 255, 243, 253},
+  {231, 163, 183, 246, 189, 255, 243, 253},
+  {231, 163, 183, 246, 189, 255, 243, 253},
+  {232, 164, 184, 247, 190, 255, 243, 253},
+  {232, 164, 184, 247, 190, 255, 243, 253},
+  {233, 165, 185, 247, 191, 255, 244, 253},
+  {233, 165, 185, 247, 191, 255, 244, 253},
+  {234, 166, 185, 247, 192, 255, 244, 253},
+  {234, 167, 185, 247, 192, 255, 244, 253},
+  {235, 168, 186, 248, 193, 255, 244, 253},
+  {235, 168, 186, 248, 193, 255, 244, 253},
+  {236, 169, 187, 248, 194, 255, 244, 253},
+  {236, 169, 187, 248, 194, 255, 244, 253},
+  {236, 170, 188, 248, 195, 255, 245, 253},
+  {236, 170, 188, 248, 195, 255, 245, 253},
+  {237, 171, 189, 249, 196, 255, 245, 254},
+  {237, 172, 189, 249, 196, 255, 245, 254},
+  {238, 173, 190, 249, 197, 255, 245, 254},
+  {238, 173, 190, 249, 197, 255, 245, 254},
+  {239, 174, 191, 249, 198, 255, 245, 254},
+  {239, 174, 191, 249, 198, 255, 245, 254},
+  {240, 175, 192, 249, 199, 255, 246, 254},
+  {240, 176, 192, 249, 199, 255, 246, 254},
+  {240, 177, 193, 250, 200, 255, 246, 254},
+  {240, 177, 193, 250, 200, 255, 246, 254},
+  {241, 178, 194, 250, 201, 255, 246, 254},
+  {241, 178, 194, 250, 201, 255, 246, 254},
+  {242, 179, 195, 250, 202, 255, 246, 254},
+  {242, 180, 195, 250, 202, 255, 246, 254},
+  {242, 181, 196, 250, 203, 255, 247, 254},
+  {242, 181, 196, 250, 203, 255, 247, 254},
+  {243, 182, 197, 251, 204, 255, 247, 254},
+  {243, 183, 197, 251, 204, 255, 247, 254},
+  {244, 184, 198, 251, 205, 255, 247, 254},
+  {244, 184, 198, 251, 205, 255, 247, 254},
+  {244, 185, 199, 251, 206, 255, 247, 254},
+  {244, 185, 199, 251, 206, 255, 247, 254},
+  {245, 186, 200, 251, 207, 255, 247, 254},
+  {245, 187, 200, 251, 207, 255, 247, 254},
+  {246, 188, 201, 252, 207, 255, 248, 254},
+  {246, 188, 201, 252, 207, 255, 248, 254},
+  {246, 189, 202, 252, 208, 255, 248, 254},
+  {246, 190, 202, 252, 208, 255, 248, 254},
+  {247, 191, 203, 252, 209, 255, 248, 254},
+  {247, 191, 203, 252, 209, 255, 248, 254},
+  {247, 192, 204, 252, 210, 255, 248, 254},
+  {247, 193, 204, 252, 210, 255, 248, 254},
+  {248, 194, 205, 252, 211, 255, 248, 254},
+  {248, 194, 205, 252, 211, 255, 248, 254},
+  {248, 195, 206, 252, 212, 255, 249, 254},
+  {248, 196, 206, 252, 212, 255, 249, 254},
+  {249, 197, 207, 253, 213, 255, 249, 254},
+  {249, 197, 207, 253, 213, 255, 249, 254},
+  {249, 198, 208, 253, 214, 255, 249, 254},
+  {249, 199, 209, 253, 214, 255, 249, 254},
+  {250, 200, 210, 253, 215, 255, 249, 254},
+  {250, 200, 210, 253, 215, 255, 249, 254},
+  {250, 201, 211, 253, 215, 255, 249, 254},
+  {250, 202, 211, 253, 215, 255, 249, 254},
+  {250, 203, 212, 253, 216, 255, 249, 254},
+  {250, 203, 212, 253, 216, 255, 249, 254},
+  {251, 204, 213, 253, 217, 255, 250, 254},
+  {251, 205, 213, 253, 217, 255, 250, 254},
+  {251, 206, 214, 254, 218, 255, 250, 254},
+  {251, 206, 215, 254, 218, 255, 250, 254},
+  {252, 207, 216, 254, 219, 255, 250, 254},
+  {252, 208, 216, 254, 219, 255, 250, 254},
+  {252, 209, 217, 254, 220, 255, 250, 254},
+  {252, 210, 217, 254, 220, 255, 250, 254},
+  {252, 211, 218, 254, 221, 255, 250, 254},
+  {252, 212, 218, 254, 221, 255, 250, 254},
+  {253, 213, 219, 254, 222, 255, 250, 254},
+  {253, 213, 220, 254, 222, 255, 250, 254},
+  {253, 214, 221, 254, 223, 255, 250, 254},
+  {253, 215, 221, 254, 223, 255, 250, 254},
+  {253, 216, 222, 254, 224, 255, 251, 254},
+  {253, 217, 223, 254, 224, 255, 251, 254},
+  {253, 218, 224, 254, 225, 255, 251, 254},
+  {253, 219, 224, 254, 225, 255, 251, 254},
+  {254, 220, 225, 254, 225, 255, 251, 254},
+  {254, 221, 226, 254, 225, 255, 251, 254},
+  {254, 222, 227, 255, 226, 255, 251, 254},
+  {254, 223, 227, 255, 226, 255, 251, 254},
+  {254, 224, 228, 255, 227, 255, 251, 254},
+  {254, 225, 229, 255, 227, 255, 251, 254},
+  {254, 226, 230, 255, 228, 255, 251, 254},
+  {254, 227, 230, 255, 229, 255, 251, 254},
+  {255, 228, 231, 255, 230, 255, 251, 254},
+  {255, 229, 232, 255, 230, 255, 251, 254},
+  {255, 230, 233, 255, 231, 255, 252, 254},
+  {255, 231, 234, 255, 231, 255, 252, 254},
+  {255, 232, 235, 255, 232, 255, 252, 254},
+  {255, 233, 236, 255, 232, 255, 252, 254},
+  {255, 235, 237, 255, 233, 255, 252, 254},
+  {255, 236, 238, 255, 234, 255, 252, 254},
+  {255, 238, 240, 255, 235, 255, 252, 255},
+  {255, 239, 241, 255, 235, 255, 252, 254},
+  {255, 241, 243, 255, 236, 255, 252, 254},
+  {255, 243, 245, 255, 237, 255, 252, 254},
+  {255, 246, 247, 255, 239, 255, 253, 255},
+};
+
+static const vp9_coeff_probs_model default_coef_probs_4x4[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        { 195,  29, 183 }, {  84,  49, 136 }, {   8,  42,  71 }
+      }, {  // Band 1
+        {  31, 107, 169 }, {  35,  99, 159 }, {  17,  82, 140 },
+        {   8,  66, 114 }, {   2,  44,  76 }, {   1,  19,  32 }
+      }, {  // Band 2
+        {  40, 132, 201 }, {  29, 114, 187 }, {  13,  91, 157 },
+        {   7,  75, 127 }, {   3,  58,  95 }, {   1,  28,  47 }
+      }, {  // Band 3
+        {  69, 142, 221 }, {  42, 122, 201 }, {  15,  91, 159 },
+        {   6,  67, 121 }, {   1,  42,  77 }, {   1,  17,  31 }
+      }, {  // Band 4
+        { 102, 148, 228 }, {  67, 117, 204 }, {  17,  82, 154 },
+        {   6,  59, 114 }, {   2,  39,  75 }, {   1,  15,  29 }
+      }, {  // Band 5
+        { 156,  57, 233 }, { 119,  57, 212 }, {  58,  48, 163 },
+        {  29,  40, 124 }, {  12,  30,  81 }, {   3,  12,  31 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 191, 107, 226 }, { 124, 117, 204 }, {  25,  99, 155 }
+      }, {  // Band 1
+        {  29, 148, 210 }, {  37, 126, 194 }, {   8,  93, 157 },
+        {   2,  68, 118 }, {   1,  39,  69 }, {   1,  17,  33 }
+      }, {  // Band 2
+        {  41, 151, 213 }, {  27, 123, 193 }, {   3,  82, 144 },
+        {   1,  58, 105 }, {   1,  32,  60 }, {   1,  13,  26 }
+      }, {  // Band 3
+        {  59, 159, 220 }, {  23, 126, 198 }, {   4,  88, 151 },
+        {   1,  66, 114 }, {   1,  38,  71 }, {   1,  18,  34 }
+      }, {  // Band 4
+        { 114, 136, 232 }, {  51, 114, 207 }, {  11,  83, 155 },
+        {   3,  56, 105 }, {   1,  33,  65 }, {   1,  17,  34 }
+      }, {  // Band 5
+        { 149,  65, 234 }, { 121,  57, 215 }, {  61,  49, 166 },
+        {  28,  36, 114 }, {  12,  25,  76 }, {   3,  16,  42 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 214,  49, 220 }, { 132,  63, 188 }, {  42,  65, 137 }
+      }, {  // Band 1
+        {  85, 137, 221 }, { 104, 131, 216 }, {  49, 111, 192 },
+        {  21,  87, 155 }, {   2,  49,  87 }, {   1,  16,  28 }
+      }, {  // Band 2
+        {  89, 163, 230 }, {  90, 137, 220 }, {  29, 100, 183 },
+        {  10,  70, 135 }, {   2,  42,  81 }, {   1,  17,  33 }
+      }, {  // Band 3
+        { 108, 167, 237 }, {  55, 133, 222 }, {  15,  97, 179 },
+        {   4,  72, 135 }, {   1,  45,  85 }, {   1,  19,  38 }
+      }, {  // Band 4
+        { 124, 146, 240 }, {  66, 124, 224 }, {  17,  88, 175 },
+        {   4,  58, 122 }, {   1,  36,  75 }, {   1,  18,  37 }
+      }, {  //  Band 5
+        { 141,  79, 241 }, { 126,  70, 227 }, {  66,  58, 182 },
+        {  30,  44, 136 }, {  12,  34,  96 }, {   2,  20,  47 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 229,  99, 249 }, { 143, 111, 235 }, {  46, 109, 192 }
+      }, {  // Band 1
+        {  82, 158, 236 }, {  94, 146, 224 }, {  25, 117, 191 },
+        {   9,  87, 149 }, {   3,  56,  99 }, {   1,  33,  57 }
+      }, {  // Band 2
+        {  83, 167, 237 }, {  68, 145, 222 }, {  10, 103, 177 },
+        {   2,  72, 131 }, {   1,  41,  79 }, {   1,  20,  39 }
+      }, {  // Band 3
+        {  99, 167, 239 }, {  47, 141, 224 }, {  10, 104, 178 },
+        {   2,  73, 133 }, {   1,  44,  85 }, {   1,  22,  47 }
+      }, {  // Band 4
+        { 127, 145, 243 }, {  71, 129, 228 }, {  17,  93, 177 },
+        {   3,  61, 124 }, {   1,  41,  84 }, {   1,  21,  52 }
+      }, {  // Band 5
+        { 157,  78, 244 }, { 140,  72, 231 }, {  69,  58, 184 },
+        {  31,  44, 137 }, {  14,  38, 105 }, {   8,  23,  61 }
+      }
+    }
+  }
+};
+
+static const vp9_coeff_probs_model default_coef_probs_8x8[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        { 125,  34, 187 }, {  52,  41, 133 }, {   6,  31,  56 }
+      }, {  // Band 1
+        {  37, 109, 153 }, {  51, 102, 147 }, {  23,  87, 128 },
+        {   8,  67, 101 }, {   1,  41,  63 }, {   1,  19,  29 }
+      }, {  // Band 2
+        {  31, 154, 185 }, {  17, 127, 175 }, {   6,  96, 145 },
+        {   2,  73, 114 }, {   1,  51,  82 }, {   1,  28,  45 }
+      }, {  // Band 3
+        {  23, 163, 200 }, {  10, 131, 185 }, {   2,  93, 148 },
+        {   1,  67, 111 }, {   1,  41,  69 }, {   1,  14,  24 }
+      }, {  // Band 4
+        {  29, 176, 217 }, {  12, 145, 201 }, {   3, 101, 156 },
+        {   1,  69, 111 }, {   1,  39,  63 }, {   1,  14,  23 }
+      }, {  // Band 5
+        {  57, 192, 233 }, {  25, 154, 215 }, {   6, 109, 167 },
+        {   3,  78, 118 }, {   1,  48,  69 }, {   1,  21,  29 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 202, 105, 245 }, { 108, 106, 216 }, {  18,  90, 144 }
+      }, {  // Band 1
+        {  33, 172, 219 }, {  64, 149, 206 }, {  14, 117, 177 },
+        {   5,  90, 141 }, {   2,  61,  95 }, {   1,  37,  57 }
+      }, {  // Band 2
+        {  33, 179, 220 }, {  11, 140, 198 }, {   1,  89, 148 },
+        {   1,  60, 104 }, {   1,  33,  57 }, {   1,  12,  21 }
+      }, {  // Band 3
+        {  30, 181, 221 }, {   8, 141, 198 }, {   1,  87, 145 },
+        {   1,  58, 100 }, {   1,  31,  55 }, {   1,  12,  20 }
+      }, {  // Band 4
+        {  32, 186, 224 }, {   7, 142, 198 }, {   1,  86, 143 },
+        {   1,  58, 100 }, {   1,  31,  55 }, {   1,  12,  22 }
+      }, {  // Band 5
+        {  57, 192, 227 }, {  20, 143, 204 }, {   3,  96, 154 },
+        {   1,  68, 112 }, {   1,  42,  69 }, {   1,  19,  32 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 212,  35, 215 }, { 113,  47, 169 }, {  29,  48, 105 }
+      }, {  // Band 1
+        {  74, 129, 203 }, { 106, 120, 203 }, {  49, 107, 178 },
+        {  19,  84, 144 }, {   4,  50,  84 }, {   1,  15,  25 }
+      }, {  // Band 2
+        {  71, 172, 217 }, {  44, 141, 209 }, {  15, 102, 173 },
+        {   6,  76, 133 }, {   2,  51,  89 }, {   1,  24,  42 }
+      }, {  // Band 3
+        {  64, 185, 231 }, {  31, 148, 216 }, {   8, 103, 175 },
+        {   3,  74, 131 }, {   1,  46,  81 }, {   1,  18,  30 }
+      }, {  // Band 4
+        {  65, 196, 235 }, {  25, 157, 221 }, {   5, 105, 174 },
+        {   1,  67, 120 }, {   1,  38,  69 }, {   1,  15,  30 }
+      }, {  // Band 5
+        {  65, 204, 238 }, {  30, 156, 224 }, {   7, 107, 177 },
+        {   2,  70, 124 }, {   1,  42,  73 }, {   1,  18,  34 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 225,  86, 251 }, { 144, 104, 235 }, {  42,  99, 181 }
+      }, {  // Band 1
+        {  85, 175, 239 }, { 112, 165, 229 }, {  29, 136, 200 },
+        {  12, 103, 162 }, {   6,  77, 123 }, {   2,  53,  84 }
+      }, {  // Band 2
+        {  75, 183, 239 }, {  30, 155, 221 }, {   3, 106, 171 },
+        {   1,  74, 128 }, {   1,  44,  76 }, {   1,  17,  28 }
+      }, {  // Band 3
+        {  73, 185, 240 }, {  27, 159, 222 }, {   2, 107, 172 },
+        {   1,  75, 127 }, {   1,  42,  73 }, {   1,  17,  29 }
+      }, {  // Band 4
+        {  62, 190, 238 }, {  21, 159, 222 }, {   2, 107, 172 },
+        {   1,  72, 122 }, {   1,  40,  71 }, {   1,  18,  32 }
+      }, {  // Band 5
+        {  61, 199, 240 }, {  27, 161, 226 }, {   4, 113, 180 },
+        {   1,  76, 129 }, {   1,  46,  80 }, {   1,  23,  41 }
+      }
+    }
+  }
+};
+
+static const vp9_coeff_probs_model default_coef_probs_16x16[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        {   7,  27, 153 }, {   5,  30,  95 }, {   1,  16,  30 }
+      }, {  // Band 1
+        {  50,  75, 127 }, {  57,  75, 124 }, {  27,  67, 108 },
+        {  10,  54,  86 }, {   1,  33,  52 }, {   1,  12,  18 }
+      }, {  // Band 2
+        {  43, 125, 151 }, {  26, 108, 148 }, {   7,  83, 122 },
+        {   2,  59,  89 }, {   1,  38,  60 }, {   1,  17,  27 }
+      }, {  // Band 3
+        {  23, 144, 163 }, {  13, 112, 154 }, {   2,  75, 117 },
+        {   1,  50,  81 }, {   1,  31,  51 }, {   1,  14,  23 }
+      }, {  // Band 4
+        {  18, 162, 185 }, {   6, 123, 171 }, {   1,  78, 125 },
+        {   1,  51,  86 }, {   1,  31,  54 }, {   1,  14,  23 }
+      }, {  // Band 5
+        {  15, 199, 227 }, {   3, 150, 204 }, {   1,  91, 146 },
+        {   1,  55,  95 }, {   1,  30,  53 }, {   1,  11,  20 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        {  19,  55, 240 }, {  19,  59, 196 }, {   3,  52, 105 }
+      }, {  // Band 1
+        {  41, 166, 207 }, { 104, 153, 199 }, {  31, 123, 181 },
+        {  14, 101, 152 }, {   5,  72, 106 }, {   1,  36,  52 }
+      }, {  // Band 2
+        {  35, 176, 211 }, {  12, 131, 190 }, {   2,  88, 144 },
+        {   1,  60, 101 }, {   1,  36,  60 }, {   1,  16,  28 }
+      }, {  // Band 3
+        {  28, 183, 213 }, {   8, 134, 191 }, {   1,  86, 142 },
+        {   1,  56,  96 }, {   1,  30,  53 }, {   1,  12,  20 }
+      }, {  // Band 4
+        {  20, 190, 215 }, {   4, 135, 192 }, {   1,  84, 139 },
+        {   1,  53,  91 }, {   1,  28,  49 }, {   1,  11,  20 }
+      }, {  // Band 5
+        {  13, 196, 216 }, {   2, 137, 192 }, {   1,  86, 143 },
+        {   1,  57,  99 }, {   1,  32,  56 }, {   1,  13,  24 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 211,  29, 217 }, {  96,  47, 156 }, {  22,  43,  87 }
+      }, {  // Band 1
+        {  78, 120, 193 }, { 111, 116, 186 }, {  46, 102, 164 },
+        {  15,  80, 128 }, {   2,  49,  76 }, {   1,  18,  28 }
+      }, {  // Band 2
+        {  71, 161, 203 }, {  42, 132, 192 }, {  10,  98, 150 },
+        {   3,  69, 109 }, {   1,  44,  70 }, {   1,  18,  29 }
+      }, {  // Band 3
+        {  57, 186, 211 }, {  30, 140, 196 }, {   4,  93, 146 },
+        {   1,  62, 102 }, {   1,  38,  65 }, {   1,  16,  27 }
+      }, {  // Band 4
+        {  47, 199, 217 }, {  14, 145, 196 }, {   1,  88, 142 },
+        {   1,  57,  98 }, {   1,  36,  62 }, {   1,  15,  26 }
+      }, {  // Band 5
+        {  26, 219, 229 }, {   5, 155, 207 }, {   1,  94, 151 },
+        {   1,  60, 104 }, {   1,  36,  62 }, {   1,  16,  28 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 233,  29, 248 }, { 146,  47, 220 }, {  43,  52, 140 }
+      }, {  // Band 1
+        { 100, 163, 232 }, { 179, 161, 222 }, {  63, 142, 204 },
+        {  37, 113, 174 }, {  26,  89, 137 }, {  18,  68,  97 }
+      }, {  // Band 2
+        {  85, 181, 230 }, {  32, 146, 209 }, {   7, 100, 164 },
+        {   3,  71, 121 }, {   1,  45,  77 }, {   1,  18,  30 }
+      }, {  // Band 3
+        {  65, 187, 230 }, {  20, 148, 207 }, {   2,  97, 159 },
+        {   1,  68, 116 }, {   1,  40,  70 }, {   1,  14,  29 }
+      }, {  // Band 4
+        {  40, 194, 227 }, {   8, 147, 204 }, {   1,  94, 155 },
+        {   1,  65, 112 }, {   1,  39,  66 }, {   1,  14,  26 }
+      }, {  // Band 5
+        {  16, 208, 228 }, {   3, 151, 207 }, {   1,  98, 160 },
+        {   1,  67, 117 }, {   1,  41,  74 }, {   1,  17,  31 }
+      }
+    }
+  }
+};
+
+static const vp9_coeff_probs_model default_coef_probs_32x32[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        {  17,  38, 140 }, {   7,  34,  80 }, {   1,  17,  29 }
+      }, {  // Band 1
+        {  37,  75, 128 }, {  41,  76, 128 }, {  26,  66, 116 },
+        {  12,  52,  94 }, {   2,  32,  55 }, {   1,  10,  16 }
+      }, {  // Band 2
+        {  50, 127, 154 }, {  37, 109, 152 }, {  16,  82, 121 },
+        {   5,  59,  85 }, {   1,  35,  54 }, {   1,  13,  20 }
+      }, {  // Band 3
+        {  40, 142, 167 }, {  17, 110, 157 }, {   2,  71, 112 },
+        {   1,  44,  72 }, {   1,  27,  45 }, {   1,  11,  17 }
+      }, {  // Band 4
+        {  30, 175, 188 }, {   9, 124, 169 }, {   1,  74, 116 },
+        {   1,  48,  78 }, {   1,  30,  49 }, {   1,  11,  18 }
+      }, {  // Band 5
+        {  10, 222, 223 }, {   2, 150, 194 }, {   1,  83, 128 },
+        {   1,  48,  79 }, {   1,  27,  45 }, {   1,  11,  17 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        {  36,  41, 235 }, {  29,  36, 193 }, {  10,  27, 111 }
+      }, {  // Band 1
+        {  85, 165, 222 }, { 177, 162, 215 }, { 110, 135, 195 },
+        {  57, 113, 168 }, {  23,  83, 120 }, {  10,  49,  61 }
+      }, {  // Band 2
+        {  85, 190, 223 }, {  36, 139, 200 }, {   5,  90, 146 },
+        {   1,  60, 103 }, {   1,  38,  65 }, {   1,  18,  30 }
+      }, {  // Band 3
+        {  72, 202, 223 }, {  23, 141, 199 }, {   2,  86, 140 },
+        {   1,  56,  97 }, {   1,  36,  61 }, {   1,  16,  27 }
+      }, {  // Band 4
+        {  55, 218, 225 }, {  13, 145, 200 }, {   1,  86, 141 },
+        {   1,  57,  99 }, {   1,  35,  61 }, {   1,  13,  22 }
+      }, {  // Band 5
+        {  15, 235, 212 }, {   1, 132, 184 }, {   1,  84, 139 },
+        {   1,  57,  97 }, {   1,  34,  56 }, {   1,  14,  23 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 181,  21, 201 }, {  61,  37, 123 }, {  10,  38,  71 }
+      }, {  // Band 1
+        {  47, 106, 172 }, {  95, 104, 173 }, {  42,  93, 159 },
+        {  18,  77, 131 }, {   4,  50,  81 }, {   1,  17,  23 }
+      }, {  // Band 2
+        {  62, 147, 199 }, {  44, 130, 189 }, {  28, 102, 154 },
+        {  18,  75, 115 }, {   2,  44,  65 }, {   1,  12,  19 }
+      }, {  // Band 3
+        {  55, 153, 210 }, {  24, 130, 194 }, {   3,  93, 146 },
+        {   1,  61,  97 }, {   1,  31,  50 }, {   1,  10,  16 }
+      }, {  // Band 4
+        {  49, 186, 223 }, {  17, 148, 204 }, {   1,  96, 142 },
+        {   1,  53,  83 }, {   1,  26,  44 }, {   1,  11,  17 }
+      }, {  // Band 5
+        {  13, 217, 212 }, {   2, 136, 180 }, {   1,  78, 124 },
+        {   1,  50,  83 }, {   1,  29,  49 }, {   1,  14,  23 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 197,  13, 247 }, {  82,  17, 222 }, {  25,  17, 162 }
+      }, {  // Band 1
+        { 126, 186, 247 }, { 234, 191, 243 }, { 176, 177, 234 },
+        { 104, 158, 220 }, {  66, 128, 186 }, {  55,  90, 137 }
+      }, {  // Band 2
+        { 111, 197, 242 }, {  46, 158, 219 }, {   9, 104, 171 },
+        {   2,  65, 125 }, {   1,  44,  80 }, {   1,  17,  91 }
+      }, {  // Band 3
+        { 104, 208, 245 }, {  39, 168, 224 }, {   3, 109, 162 },
+        {   1,  79, 124 }, {   1,  50, 102 }, {   1,  43, 102 }
+      }, {  // Band 4
+        {  84, 220, 246 }, {  31, 177, 231 }, {   2, 115, 180 },
+        {   1,  79, 134 }, {   1,  55,  77 }, {   1,  60,  79 }
+      }, {  // Band 5
+        {  43, 243, 240 }, {   8, 180, 217 }, {   1, 115, 166 },
+        {   1,  84, 121 }, {   1,  51,  67 }, {   1,  16,   6 }
+      }
+    }
+  }
+};
+
+static void extend_to_full_distribution(vpx_prob *probs, vpx_prob p) {
+  assert(p != 0);
+  memcpy(probs, vp9_pareto8_full[p - 1], MODEL_NODES * sizeof(vpx_prob));
+}
+
+void vp9_model_to_full_probs(const vpx_prob *model, vpx_prob *full) {
+  if (full != model)
+    memcpy(full, model, sizeof(vpx_prob) * UNCONSTRAINED_NODES);
+  extend_to_full_distribution(&full[UNCONSTRAINED_NODES], model[PIVOT_NODE]);
+}
+
+void vp9_default_coef_probs(VP9_COMMON *cm) {
+  vp9_copy(cm->fc->coef_probs[TX_4X4], default_coef_probs_4x4);
+  vp9_copy(cm->fc->coef_probs[TX_8X8], default_coef_probs_8x8);
+  vp9_copy(cm->fc->coef_probs[TX_16X16], default_coef_probs_16x16);
+  vp9_copy(cm->fc->coef_probs[TX_32X32], default_coef_probs_32x32);
+}
+
+#define COEF_COUNT_SAT 24
+#define COEF_MAX_UPDATE_FACTOR 112
+#define COEF_COUNT_SAT_KEY 24
+#define COEF_MAX_UPDATE_FACTOR_KEY 112
+#define COEF_COUNT_SAT_AFTER_KEY 24
+#define COEF_MAX_UPDATE_FACTOR_AFTER_KEY 128
+
+static void adapt_coef_probs(VP9_COMMON *cm, TX_SIZE tx_size,
+                             unsigned int count_sat,
+                             unsigned int update_factor) {
+  const FRAME_CONTEXT *pre_fc = &cm->frame_contexts[cm->frame_context_idx];
+  vp9_coeff_probs_model *const probs = cm->fc->coef_probs[tx_size];
+  const vp9_coeff_probs_model *const pre_probs = pre_fc->coef_probs[tx_size];
+  vp9_coeff_count_model *counts = cm->counts.coef[tx_size];
+  unsigned int (*eob_counts)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
+      cm->counts.eob_branch[tx_size];
+  int i, j, k, l, m;
+
+  for (i = 0; i < PLANE_TYPES; ++i)
+    for (j = 0; j < REF_TYPES; ++j)
+      for (k = 0; k < COEF_BANDS; ++k)
+        for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+          const int n0 = counts[i][j][k][l][ZERO_TOKEN];
+          const int n1 = counts[i][j][k][l][ONE_TOKEN];
+          const int n2 = counts[i][j][k][l][TWO_TOKEN];
+          const int neob = counts[i][j][k][l][EOB_MODEL_TOKEN];
+          const unsigned int branch_ct[UNCONSTRAINED_NODES][2] = {
+            { neob, eob_counts[i][j][k][l] - neob },
+            { n0, n1 + n2 },
+            { n1, n2 }
+          };
+          for (m = 0; m < UNCONSTRAINED_NODES; ++m)
+            probs[i][j][k][l][m] = merge_probs(pre_probs[i][j][k][l][m],
+                                               branch_ct[m],
+                                               count_sat, update_factor);
+        }
+}
+
+void vp9_adapt_coef_probs(VP9_COMMON *cm) {
+  TX_SIZE t;
+  unsigned int count_sat, update_factor;
+
+  if (frame_is_intra_only(cm)) {
+    update_factor = COEF_MAX_UPDATE_FACTOR_KEY;
+    count_sat = COEF_COUNT_SAT_KEY;
+  } else if (cm->last_frame_type == KEY_FRAME) {
+    update_factor = COEF_MAX_UPDATE_FACTOR_AFTER_KEY;  /* adapt quickly */
+    count_sat = COEF_COUNT_SAT_AFTER_KEY;
+  } else {
+    update_factor = COEF_MAX_UPDATE_FACTOR;
+    count_sat = COEF_COUNT_SAT;
+  }
+  for (t = TX_4X4; t <= TX_32X32; t++)
+    adapt_coef_probs(cm, t, count_sat, update_factor);
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_entropy.h b/libvpx/libvpx/vp9/common/vp9_entropy.h
new file mode 100644
index 0000000..63b3bff
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_entropy.h
@@ -0,0 +1,200 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_ENTROPY_H_
+#define VP9_COMMON_VP9_ENTROPY_H_
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/prob.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define DIFF_UPDATE_PROB 252
+
+// Coefficient token alphabet
+#define ZERO_TOKEN      0   // 0     Extra Bits 0+0
+#define ONE_TOKEN       1   // 1     Extra Bits 0+1
+#define TWO_TOKEN       2   // 2     Extra Bits 0+1
+#define THREE_TOKEN     3   // 3     Extra Bits 0+1
+#define FOUR_TOKEN      4   // 4     Extra Bits 0+1
+#define CATEGORY1_TOKEN 5   // 5-6   Extra Bits 1+1
+#define CATEGORY2_TOKEN 6   // 7-10  Extra Bits 2+1
+#define CATEGORY3_TOKEN 7   // 11-18 Extra Bits 3+1
+#define CATEGORY4_TOKEN 8   // 19-34 Extra Bits 4+1
+#define CATEGORY5_TOKEN 9   // 35-66 Extra Bits 5+1
+#define CATEGORY6_TOKEN 10  // 67+   Extra Bits 14+1
+#define EOB_TOKEN       11  // EOB   Extra Bits 0+0
+
+#define ENTROPY_TOKENS 12
+
+#define ENTROPY_NODES 11
+
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_pt_energy_class[ENTROPY_TOKENS]);
+
+#define CAT1_MIN_VAL    5
+#define CAT2_MIN_VAL    7
+#define CAT3_MIN_VAL   11
+#define CAT4_MIN_VAL   19
+#define CAT5_MIN_VAL   35
+#define CAT6_MIN_VAL   67
+
+// Extra bit probabilities.
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat1_prob[1]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat2_prob[2]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat3_prob[3]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat4_prob[4]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat5_prob[5]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat6_prob[14]);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat1_prob_high10[1]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat2_prob_high10[2]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat3_prob_high10[3]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat4_prob_high10[4]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat5_prob_high10[5]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat6_prob_high10[16]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat1_prob_high12[1]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat2_prob_high12[2]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat3_prob_high12[3]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat4_prob_high12[4]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat5_prob_high12[5]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat6_prob_high12[18]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#define EOB_MODEL_TOKEN 3
+
+#define DCT_MAX_VALUE           16384
+#if CONFIG_VP9_HIGHBITDEPTH
+#define DCT_MAX_VALUE_HIGH10    65536
+#define DCT_MAX_VALUE_HIGH12   262144
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+/* Coefficients are predicted via a 3-dimensional probability table. */
+
+#define REF_TYPES 2  // intra=0, inter=1
+
+/* Middle dimension reflects the coefficient position within the transform. */
+#define COEF_BANDS 6
+
+/* Inside dimension is measure of nearby complexity, that reflects the energy
+   of nearby coefficients are nonzero.  For the first coefficient (DC, unless
+   block type is 0), we look at the (already encoded) blocks above and to the
+   left of the current block.  The context index is then the number (0,1,or 2)
+   of these blocks having nonzero coefficients.
+   After decoding a coefficient, the measure is determined by the size of the
+   most recently decoded coefficient.
+   Note that the intuitive meaning of this measure changes as coefficients
+   are decoded, e.g., prior to the first token, a zero means that my neighbors
+   are empty while, after the first token, because of the use of end-of-block,
+   a zero means we just decoded a zero and hence guarantees that a non-zero
+   coefficient will appear later in this block.  However, this shift
+   in meaning is perfectly OK because our context depends also on the
+   coefficient band (and since zigzag positions 0, 1, and 2 are in
+   distinct bands). */
+
+#define COEFF_CONTEXTS 6
+#define BAND_COEFF_CONTEXTS(band) ((band) == 0 ? 3 : COEFF_CONTEXTS)
+
+// #define ENTROPY_STATS
+
+typedef unsigned int vp9_coeff_count[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS]
+                                    [ENTROPY_TOKENS];
+typedef unsigned int vp9_coeff_stats[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS]
+                                    [ENTROPY_NODES][2];
+
+#define SUBEXP_PARAM                4   /* Subexponential code parameter */
+#define MODULUS_PARAM               13  /* Modulus parameter */
+
+struct VP9Common;
+void vp9_default_coef_probs(struct VP9Common *cm);
+void vp9_adapt_coef_probs(struct VP9Common *cm);
+
+// This is the index in the scan order beyond which all coefficients for
+// 8x8 transform and above are in the top band.
+// This macro is currently unused but may be used by certain implementations
+#define MAXBAND_INDEX 21
+
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_coefband_trans_8x8plus[1024]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_coefband_trans_4x4[16]);
+
+static INLINE const uint8_t *get_band_translate(TX_SIZE tx_size) {
+  return tx_size == TX_4X4 ? vp9_coefband_trans_4x4
+                           : vp9_coefband_trans_8x8plus;
+}
+
+// 128 lists of probabilities are stored for the following ONE node probs:
+// 1, 3, 5, 7, ..., 253, 255
+// In between probabilities are interpolated linearly
+
+#define COEFF_PROB_MODELS 255
+
+#define UNCONSTRAINED_NODES         3
+
+#define PIVOT_NODE                  2   // which node is pivot
+
+#define MODEL_NODES (ENTROPY_NODES - UNCONSTRAINED_NODES)
+extern const vpx_tree_index vp9_coef_con_tree[TREE_SIZE(ENTROPY_TOKENS)];
+extern const vpx_prob vp9_pareto8_full[COEFF_PROB_MODELS][MODEL_NODES];
+
+typedef vpx_prob vp9_coeff_probs_model[REF_TYPES][COEF_BANDS]
+                                      [COEFF_CONTEXTS][UNCONSTRAINED_NODES];
+
+typedef unsigned int vp9_coeff_count_model[REF_TYPES][COEF_BANDS]
+                                          [COEFF_CONTEXTS]
+                                          [UNCONSTRAINED_NODES + 1];
+
+void vp9_model_to_full_probs(const vpx_prob *model, vpx_prob *full);
+
+typedef char ENTROPY_CONTEXT;
+
+static INLINE int combine_entropy_contexts(ENTROPY_CONTEXT a,
+                                           ENTROPY_CONTEXT b) {
+  return (a != 0) + (b != 0);
+}
+
+static INLINE int get_entropy_context(TX_SIZE tx_size, const ENTROPY_CONTEXT *a,
+                                      const ENTROPY_CONTEXT *l) {
+  ENTROPY_CONTEXT above_ec = 0, left_ec = 0;
+
+  switch (tx_size) {
+    case TX_4X4:
+      above_ec = a[0] != 0;
+      left_ec = l[0] != 0;
+      break;
+    case TX_8X8:
+      above_ec = !!*(const uint16_t *)a;
+      left_ec  = !!*(const uint16_t *)l;
+      break;
+    case TX_16X16:
+      above_ec = !!*(const uint32_t *)a;
+      left_ec  = !!*(const uint32_t *)l;
+      break;
+    case TX_32X32:
+      above_ec = !!*(const uint64_t *)a;
+      left_ec  = !!*(const uint64_t *)l;
+      break;
+    default:
+      assert(0 && "Invalid transform size.");
+      break;
+  }
+
+  return combine_entropy_contexts(above_ec, left_ec);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ENTROPY_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_entropymode.c b/libvpx/libvpx/vp9/common/vp9_entropymode.c
new file mode 100644
index 0000000..670348b
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_entropymode.c
@@ -0,0 +1,469 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_seg_common.h"
+
+const vpx_prob vp9_kf_y_mode_prob[INTRA_MODES][INTRA_MODES][INTRA_MODES - 1] = {
+  {  // above = dc
+    { 137,  30,  42, 148, 151, 207,  70,  52,  91 },  // left = dc
+    {  92,  45, 102, 136, 116, 180,  74,  90, 100 },  // left = v
+    {  73,  32,  19, 187, 222, 215,  46,  34, 100 },  // left = h
+    {  91,  30,  32, 116, 121, 186,  93,  86,  94 },  // left = d45
+    {  72,  35,  36, 149,  68, 206,  68,  63, 105 },  // left = d135
+    {  73,  31,  28, 138,  57, 124,  55, 122, 151 },  // left = d117
+    {  67,  23,  21, 140, 126, 197,  40,  37, 171 },  // left = d153
+    {  86,  27,  28, 128, 154, 212,  45,  43,  53 },  // left = d207
+    {  74,  32,  27, 107,  86, 160,  63, 134, 102 },  // left = d63
+    {  59,  67,  44, 140, 161, 202,  78,  67, 119 }   // left = tm
+  }, {  // above = v
+    {  63,  36, 126, 146, 123, 158,  60,  90,  96 },  // left = dc
+    {  43,  46, 168, 134, 107, 128,  69, 142,  92 },  // left = v
+    {  44,  29,  68, 159, 201, 177,  50,  57,  77 },  // left = h
+    {  58,  38,  76, 114,  97, 172,  78, 133,  92 },  // left = d45
+    {  46,  41,  76, 140,  63, 184,  69, 112,  57 },  // left = d135
+    {  38,  32,  85, 140,  46, 112,  54, 151, 133 },  // left = d117
+    {  39,  27,  61, 131, 110, 175,  44,  75, 136 },  // left = d153
+    {  52,  30,  74, 113, 130, 175,  51,  64,  58 },  // left = d207
+    {  47,  35,  80, 100,  74, 143,  64, 163,  74 },  // left = d63
+    {  36,  61, 116, 114, 128, 162,  80, 125,  82 }   // left = tm
+  }, {  // above = h
+    {  82,  26,  26, 171, 208, 204,  44,  32, 105 },  // left = dc
+    {  55,  44,  68, 166, 179, 192,  57,  57, 108 },  // left = v
+    {  42,  26,  11, 199, 241, 228,  23,  15,  85 },  // left = h
+    {  68,  42,  19, 131, 160, 199,  55,  52,  83 },  // left = d45
+    {  58,  50,  25, 139, 115, 232,  39,  52, 118 },  // left = d135
+    {  50,  35,  33, 153, 104, 162,  64,  59, 131 },  // left = d117
+    {  44,  24,  16, 150, 177, 202,  33,  19, 156 },  // left = d153
+    {  55,  27,  12, 153, 203, 218,  26,  27,  49 },  // left = d207
+    {  53,  49,  21, 110, 116, 168,  59,  80,  76 },  // left = d63
+    {  38,  72,  19, 168, 203, 212,  50,  50, 107 }   // left = tm
+  }, {  // above = d45
+    { 103,  26,  36, 129, 132, 201,  83,  80,  93 },  // left = dc
+    {  59,  38,  83, 112, 103, 162,  98, 136,  90 },  // left = v
+    {  62,  30,  23, 158, 200, 207,  59,  57,  50 },  // left = h
+    {  67,  30,  29,  84,  86, 191, 102,  91,  59 },  // left = d45
+    {  60,  32,  33, 112,  71, 220,  64,  89, 104 },  // left = d135
+    {  53,  26,  34, 130,  56, 149,  84, 120, 103 },  // left = d117
+    {  53,  21,  23, 133, 109, 210,  56,  77, 172 },  // left = d153
+    {  77,  19,  29, 112, 142, 228,  55,  66,  36 },  // left = d207
+    {  61,  29,  29,  93,  97, 165,  83, 175, 162 },  // left = d63
+    {  47,  47,  43, 114, 137, 181, 100,  99,  95 }   // left = tm
+  }, {  // above = d135
+    {  69,  23,  29, 128,  83, 199,  46,  44, 101 },  // left = dc
+    {  53,  40,  55, 139,  69, 183,  61,  80, 110 },  // left = v
+    {  40,  29,  19, 161, 180, 207,  43,  24,  91 },  // left = h
+    {  60,  34,  19, 105,  61, 198,  53,  64,  89 },  // left = d45
+    {  52,  31,  22, 158,  40, 209,  58,  62,  89 },  // left = d135
+    {  44,  31,  29, 147,  46, 158,  56, 102, 198 },  // left = d117
+    {  35,  19,  12, 135,  87, 209,  41,  45, 167 },  // left = d153
+    {  55,  25,  21, 118,  95, 215,  38,  39,  66 },  // left = d207
+    {  51,  38,  25, 113,  58, 164,  70,  93,  97 },  // left = d63
+    {  47,  54,  34, 146, 108, 203,  72, 103, 151 }   // left = tm
+  }, {  // above = d117
+    {  64,  19,  37, 156,  66, 138,  49,  95, 133 },  // left = dc
+    {  46,  27,  80, 150,  55, 124,  55, 121, 135 },  // left = v
+    {  36,  23,  27, 165, 149, 166,  54,  64, 118 },  // left = h
+    {  53,  21,  36, 131,  63, 163,  60, 109,  81 },  // left = d45
+    {  40,  26,  35, 154,  40, 185,  51,  97, 123 },  // left = d135
+    {  35,  19,  34, 179,  19,  97,  48, 129, 124 },  // left = d117
+    {  36,  20,  26, 136,  62, 164,  33,  77, 154 },  // left = d153
+    {  45,  18,  32, 130,  90, 157,  40,  79,  91 },  // left = d207
+    {  45,  26,  28, 129,  45, 129,  49, 147, 123 },  // left = d63
+    {  38,  44,  51, 136,  74, 162,  57,  97, 121 }   // left = tm
+  }, {  // above = d153
+    {  75,  17,  22, 136, 138, 185,  32,  34, 166 },  // left = dc
+    {  56,  39,  58, 133, 117, 173,  48,  53, 187 },  // left = v
+    {  35,  21,  12, 161, 212, 207,  20,  23, 145 },  // left = h
+    {  56,  29,  19, 117, 109, 181,  55,  68, 112 },  // left = d45
+    {  47,  29,  17, 153,  64, 220,  59,  51, 114 },  // left = d135
+    {  46,  16,  24, 136,  76, 147,  41,  64, 172 },  // left = d117
+    {  34,  17,  11, 108, 152, 187,  13,  15, 209 },  // left = d153
+    {  51,  24,  14, 115, 133, 209,  32,  26, 104 },  // left = d207
+    {  55,  30,  18, 122,  79, 179,  44,  88, 116 },  // left = d63
+    {  37,  49,  25, 129, 168, 164,  41,  54, 148 }   // left = tm
+  }, {  // above = d207
+    {  82,  22,  32, 127, 143, 213,  39,  41,  70 },  // left = dc
+    {  62,  44,  61, 123, 105, 189,  48,  57,  64 },  // left = v
+    {  47,  25,  17, 175, 222, 220,  24,  30,  86 },  // left = h
+    {  68,  36,  17, 106, 102, 206,  59,  74,  74 },  // left = d45
+    {  57,  39,  23, 151,  68, 216,  55,  63,  58 },  // left = d135
+    {  49,  30,  35, 141,  70, 168,  82,  40, 115 },  // left = d117
+    {  51,  25,  15, 136, 129, 202,  38,  35, 139 },  // left = d153
+    {  68,  26,  16, 111, 141, 215,  29,  28,  28 },  // left = d207
+    {  59,  39,  19, 114,  75, 180,  77, 104,  42 },  // left = d63
+    {  40,  61,  26, 126, 152, 206,  61,  59,  93 }   // left = tm
+  }, {  // above = d63
+    {  78,  23,  39, 111, 117, 170,  74, 124,  94 },  // left = dc
+    {  48,  34,  86, 101,  92, 146,  78, 179, 134 },  // left = v
+    {  47,  22,  24, 138, 187, 178,  68,  69,  59 },  // left = h
+    {  56,  25,  33, 105, 112, 187,  95, 177, 129 },  // left = d45
+    {  48,  31,  27, 114,  63, 183,  82, 116,  56 },  // left = d135
+    {  43,  28,  37, 121,  63, 123,  61, 192, 169 },  // left = d117
+    {  42,  17,  24, 109,  97, 177,  56,  76, 122 },  // left = d153
+    {  58,  18,  28, 105, 139, 182,  70,  92,  63 },  // left = d207
+    {  46,  23,  32,  74,  86, 150,  67, 183,  88 },  // left = d63
+    {  36,  38,  48,  92, 122, 165,  88, 137,  91 }   // left = tm
+  }, {  // above = tm
+    {  65,  70,  60, 155, 159, 199,  61,  60,  81 },  // left = dc
+    {  44,  78, 115, 132, 119, 173,  71, 112,  93 },  // left = v
+    {  39,  38,  21, 184, 227, 206,  42,  32,  64 },  // left = h
+    {  58,  47,  36, 124, 137, 193,  80,  82,  78 },  // left = d45
+    {  49,  50,  35, 144,  95, 205,  63,  78,  59 },  // left = d135
+    {  41,  53,  52, 148,  71, 142,  65, 128,  51 },  // left = d117
+    {  40,  36,  28, 143, 143, 202,  40,  55, 137 },  // left = d153
+    {  52,  34,  29, 129, 183, 227,  42,  35,  43 },  // left = d207
+    {  42,  44,  44, 104, 105, 164,  64, 130,  80 },  // left = d63
+    {  43,  81,  53, 140, 169, 204,  68,  84,  72 }   // left = tm
+  }
+};
+
+const vpx_prob vp9_kf_uv_mode_prob[INTRA_MODES][INTRA_MODES - 1] = {
+  { 144,  11,  54, 157, 195, 130,  46,  58, 108 },  // y = dc
+  { 118,  15, 123, 148, 131, 101,  44,  93, 131 },  // y = v
+  { 113,  12,  23, 188, 226, 142,  26,  32, 125 },  // y = h
+  { 120,  11,  50, 123, 163, 135,  64,  77, 103 },  // y = d45
+  { 113,   9,  36, 155, 111, 157,  32,  44, 161 },  // y = d135
+  { 116,   9,  55, 176,  76,  96,  37,  61, 149 },  // y = d117
+  { 115,   9,  28, 141, 161, 167,  21,  25, 193 },  // y = d153
+  { 120,  12,  32, 145, 195, 142,  32,  38,  86 },  // y = d207
+  { 116,  12,  64, 120, 140, 125,  49, 115, 121 },  // y = d63
+  { 102,  19,  66, 162, 182, 122,  35,  59, 128 }   // y = tm
+};
+
+static const vpx_prob default_if_y_probs[BLOCK_SIZE_GROUPS][INTRA_MODES - 1] = {
+  {  65,  32,  18, 144, 162, 194,  41,  51,  98 },  // block_size < 8x8
+  { 132,  68,  18, 165, 217, 196,  45,  40,  78 },  // block_size < 16x16
+  { 173,  80,  19, 176, 240, 193,  64,  35,  46 },  // block_size < 32x32
+  { 221, 135,  38, 194, 248, 121,  96,  85,  29 }   // block_size >= 32x32
+};
+
+static const vpx_prob default_if_uv_probs[INTRA_MODES][INTRA_MODES - 1] = {
+  { 120,   7,  76, 176, 208, 126,  28,  54, 103 },  // y = dc
+  {  48,  12, 154, 155, 139,  90,  34, 117, 119 },  // y = v
+  {  67,   6,  25, 204, 243, 158,  13,  21,  96 },  // y = h
+  {  97,   5,  44, 131, 176, 139,  48,  68,  97 },  // y = d45
+  {  83,   5,  42, 156, 111, 152,  26,  49, 152 },  // y = d135
+  {  80,   5,  58, 178,  74,  83,  33,  62, 145 },  // y = d117
+  {  86,   5,  32, 154, 192, 168,  14,  22, 163 },  // y = d153
+  {  85,   5,  32, 156, 216, 148,  19,  29,  73 },  // y = d207
+  {  77,   7,  64, 116, 132, 122,  37, 126, 120 },  // y = d63
+  { 101,  21, 107, 181, 192, 103,  19,  67, 125 }   // y = tm
+};
+
+const vpx_prob vp9_kf_partition_probs[PARTITION_CONTEXTS]
+                                     [PARTITION_TYPES - 1] = {
+  // 8x8 -> 4x4
+  { 158,  97,  94 },  // a/l both not split
+  {  93,  24,  99 },  // a split, l not split
+  {  85, 119,  44 },  // l split, a not split
+  {  62,  59,  67 },  // a/l both split
+  // 16x16 -> 8x8
+  { 149,  53,  53 },  // a/l both not split
+  {  94,  20,  48 },  // a split, l not split
+  {  83,  53,  24 },  // l split, a not split
+  {  52,  18,  18 },  // a/l both split
+  // 32x32 -> 16x16
+  { 150,  40,  39 },  // a/l both not split
+  {  78,  12,  26 },  // a split, l not split
+  {  67,  33,  11 },  // l split, a not split
+  {  24,   7,   5 },  // a/l both split
+  // 64x64 -> 32x32
+  { 174,  35,  49 },  // a/l both not split
+  {  68,  11,  27 },  // a split, l not split
+  {  57,  15,   9 },  // l split, a not split
+  {  12,   3,   3 },  // a/l both split
+};
+
+static const vpx_prob default_partition_probs[PARTITION_CONTEXTS]
+                                             [PARTITION_TYPES - 1] = {
+  // 8x8 -> 4x4
+  { 199, 122, 141 },  // a/l both not split
+  { 147,  63, 159 },  // a split, l not split
+  { 148, 133, 118 },  // l split, a not split
+  { 121, 104, 114 },  // a/l both split
+  // 16x16 -> 8x8
+  { 174,  73,  87 },  // a/l both not split
+  {  92,  41,  83 },  // a split, l not split
+  {  82,  99,  50 },  // l split, a not split
+  {  53,  39,  39 },  // a/l both split
+  // 32x32 -> 16x16
+  { 177,  58,  59 },  // a/l both not split
+  {  68,  26,  63 },  // a split, l not split
+  {  52,  79,  25 },  // l split, a not split
+  {  17,  14,  12 },  // a/l both split
+  // 64x64 -> 32x32
+  { 222,  34,  30 },  // a/l both not split
+  {  72,  16,  44 },  // a split, l not split
+  {  58,  32,  12 },  // l split, a not split
+  {  10,   7,   6 },  // a/l both split
+};
+
+static const vpx_prob default_inter_mode_probs[INTER_MODE_CONTEXTS]
+                                              [INTER_MODES - 1] = {
+  {2,       173,   34},  // 0 = both zero mv
+  {7,       145,   85},  // 1 = one zero mv + one a predicted mv
+  {7,       166,   63},  // 2 = two predicted mvs
+  {7,       94,    66},  // 3 = one predicted/zero and one new mv
+  {8,       64,    46},  // 4 = two new mvs
+  {17,      81,    31},  // 5 = one intra neighbour + x
+  {25,      29,    30},  // 6 = two intra neighbours
+};
+
+/* Array indices are identical to previously-existing INTRAMODECONTEXTNODES. */
+const vpx_tree_index vp9_intra_mode_tree[TREE_SIZE(INTRA_MODES)] = {
+  -DC_PRED, 2,                      /* 0 = DC_NODE */
+  -TM_PRED, 4,                      /* 1 = TM_NODE */
+  -V_PRED, 6,                       /* 2 = V_NODE */
+  8, 12,                            /* 3 = COM_NODE */
+  -H_PRED, 10,                      /* 4 = H_NODE */
+  -D135_PRED, -D117_PRED,           /* 5 = D135_NODE */
+  -D45_PRED, 14,                    /* 6 = D45_NODE */
+  -D63_PRED, 16,                    /* 7 = D63_NODE */
+  -D153_PRED, -D207_PRED             /* 8 = D153_NODE */
+};
+
+const vpx_tree_index vp9_inter_mode_tree[TREE_SIZE(INTER_MODES)] = {
+  -INTER_OFFSET(ZEROMV), 2,
+  -INTER_OFFSET(NEARESTMV), 4,
+  -INTER_OFFSET(NEARMV), -INTER_OFFSET(NEWMV)
+};
+
+const vpx_tree_index vp9_partition_tree[TREE_SIZE(PARTITION_TYPES)] = {
+  -PARTITION_NONE, 2,
+  -PARTITION_HORZ, 4,
+  -PARTITION_VERT, -PARTITION_SPLIT
+};
+
+static const vpx_prob default_intra_inter_p[INTRA_INTER_CONTEXTS] = {
+  9, 102, 187, 225
+};
+
+static const vpx_prob default_comp_inter_p[COMP_INTER_CONTEXTS] = {
+  239, 183, 119,  96,  41
+};
+
+static const vpx_prob default_comp_ref_p[REF_CONTEXTS] = {
+  50, 126, 123, 221, 226
+};
+
+static const vpx_prob default_single_ref_p[REF_CONTEXTS][2] = {
+  {  33,  16 },
+  {  77,  74 },
+  { 142, 142 },
+  { 172, 170 },
+  { 238, 247 }
+};
+
+static const struct tx_probs default_tx_probs = {
+  { { 3, 136, 37 },
+    { 5, 52,  13 } },
+
+  { { 20, 152 },
+    { 15, 101 } },
+
+  { { 100 },
+    { 66  } }
+};
+
+void tx_counts_to_branch_counts_32x32(const unsigned int *tx_count_32x32p,
+                                      unsigned int (*ct_32x32p)[2]) {
+  ct_32x32p[0][0] = tx_count_32x32p[TX_4X4];
+  ct_32x32p[0][1] = tx_count_32x32p[TX_8X8] +
+                    tx_count_32x32p[TX_16X16] +
+                    tx_count_32x32p[TX_32X32];
+  ct_32x32p[1][0] = tx_count_32x32p[TX_8X8];
+  ct_32x32p[1][1] = tx_count_32x32p[TX_16X16] +
+                    tx_count_32x32p[TX_32X32];
+  ct_32x32p[2][0] = tx_count_32x32p[TX_16X16];
+  ct_32x32p[2][1] = tx_count_32x32p[TX_32X32];
+}
+
+void tx_counts_to_branch_counts_16x16(const unsigned int *tx_count_16x16p,
+                                      unsigned int (*ct_16x16p)[2]) {
+  ct_16x16p[0][0] = tx_count_16x16p[TX_4X4];
+  ct_16x16p[0][1] = tx_count_16x16p[TX_8X8] + tx_count_16x16p[TX_16X16];
+  ct_16x16p[1][0] = tx_count_16x16p[TX_8X8];
+  ct_16x16p[1][1] = tx_count_16x16p[TX_16X16];
+}
+
+void tx_counts_to_branch_counts_8x8(const unsigned int *tx_count_8x8p,
+                                    unsigned int (*ct_8x8p)[2]) {
+  ct_8x8p[0][0] = tx_count_8x8p[TX_4X4];
+  ct_8x8p[0][1] = tx_count_8x8p[TX_8X8];
+}
+
+static const vpx_prob default_skip_probs[SKIP_CONTEXTS] = {
+  192, 128, 64
+};
+
+static const vpx_prob default_switchable_interp_prob[SWITCHABLE_FILTER_CONTEXTS]
+                                                    [SWITCHABLE_FILTERS - 1] = {
+  { 235, 162, },
+  { 36, 255, },
+  { 34, 3, },
+  { 149, 144, },
+};
+
+static void init_mode_probs(FRAME_CONTEXT *fc) {
+  vp9_copy(fc->uv_mode_prob, default_if_uv_probs);
+  vp9_copy(fc->y_mode_prob, default_if_y_probs);
+  vp9_copy(fc->switchable_interp_prob, default_switchable_interp_prob);
+  vp9_copy(fc->partition_prob, default_partition_probs);
+  vp9_copy(fc->intra_inter_prob, default_intra_inter_p);
+  vp9_copy(fc->comp_inter_prob, default_comp_inter_p);
+  vp9_copy(fc->comp_ref_prob, default_comp_ref_p);
+  vp9_copy(fc->single_ref_prob, default_single_ref_p);
+  fc->tx_probs = default_tx_probs;
+  vp9_copy(fc->skip_probs, default_skip_probs);
+  vp9_copy(fc->inter_mode_probs, default_inter_mode_probs);
+}
+
+const vpx_tree_index vp9_switchable_interp_tree
+                         [TREE_SIZE(SWITCHABLE_FILTERS)] = {
+  -EIGHTTAP, 2,
+  -EIGHTTAP_SMOOTH, -EIGHTTAP_SHARP
+};
+
+void vp9_adapt_mode_probs(VP9_COMMON *cm) {
+  int i, j;
+  FRAME_CONTEXT *fc = cm->fc;
+  const FRAME_CONTEXT *pre_fc = &cm->frame_contexts[cm->frame_context_idx];
+  const FRAME_COUNTS *counts = &cm->counts;
+
+  for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+    fc->intra_inter_prob[i] = mode_mv_merge_probs(pre_fc->intra_inter_prob[i],
+                                                  counts->intra_inter[i]);
+  for (i = 0; i < COMP_INTER_CONTEXTS; i++)
+    fc->comp_inter_prob[i] = mode_mv_merge_probs(pre_fc->comp_inter_prob[i],
+                                                 counts->comp_inter[i]);
+  for (i = 0; i < REF_CONTEXTS; i++)
+    fc->comp_ref_prob[i] = mode_mv_merge_probs(pre_fc->comp_ref_prob[i],
+                                               counts->comp_ref[i]);
+  for (i = 0; i < REF_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      fc->single_ref_prob[i][j] = mode_mv_merge_probs(
+          pre_fc->single_ref_prob[i][j], counts->single_ref[i][j]);
+
+  for (i = 0; i < INTER_MODE_CONTEXTS; i++)
+    vpx_tree_merge_probs(vp9_inter_mode_tree, pre_fc->inter_mode_probs[i],
+                counts->inter_mode[i], fc->inter_mode_probs[i]);
+
+  for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
+    vpx_tree_merge_probs(vp9_intra_mode_tree, pre_fc->y_mode_prob[i],
+                counts->y_mode[i], fc->y_mode_prob[i]);
+
+  for (i = 0; i < INTRA_MODES; ++i)
+    vpx_tree_merge_probs(vp9_intra_mode_tree, pre_fc->uv_mode_prob[i],
+                         counts->uv_mode[i], fc->uv_mode_prob[i]);
+
+  for (i = 0; i < PARTITION_CONTEXTS; i++)
+    vpx_tree_merge_probs(vp9_partition_tree, pre_fc->partition_prob[i],
+                         counts->partition[i], fc->partition_prob[i]);
+
+  if (cm->interp_filter == SWITCHABLE) {
+    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+      vpx_tree_merge_probs(vp9_switchable_interp_tree,
+                           pre_fc->switchable_interp_prob[i],
+                           counts->switchable_interp[i],
+                           fc->switchable_interp_prob[i]);
+  }
+
+  if (cm->tx_mode == TX_MODE_SELECT) {
+    int j;
+    unsigned int branch_ct_8x8p[TX_SIZES - 3][2];
+    unsigned int branch_ct_16x16p[TX_SIZES - 2][2];
+    unsigned int branch_ct_32x32p[TX_SIZES - 1][2];
+
+    for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
+      tx_counts_to_branch_counts_8x8(counts->tx.p8x8[i], branch_ct_8x8p);
+      for (j = 0; j < TX_SIZES - 3; ++j)
+        fc->tx_probs.p8x8[i][j] = mode_mv_merge_probs(
+            pre_fc->tx_probs.p8x8[i][j], branch_ct_8x8p[j]);
+
+      tx_counts_to_branch_counts_16x16(counts->tx.p16x16[i], branch_ct_16x16p);
+      for (j = 0; j < TX_SIZES - 2; ++j)
+        fc->tx_probs.p16x16[i][j] = mode_mv_merge_probs(
+            pre_fc->tx_probs.p16x16[i][j], branch_ct_16x16p[j]);
+
+      tx_counts_to_branch_counts_32x32(counts->tx.p32x32[i], branch_ct_32x32p);
+      for (j = 0; j < TX_SIZES - 1; ++j)
+        fc->tx_probs.p32x32[i][j] = mode_mv_merge_probs(
+            pre_fc->tx_probs.p32x32[i][j], branch_ct_32x32p[j]);
+    }
+  }
+
+  for (i = 0; i < SKIP_CONTEXTS; ++i)
+    fc->skip_probs[i] = mode_mv_merge_probs(
+        pre_fc->skip_probs[i], counts->skip[i]);
+}
+
+static void set_default_lf_deltas(struct loopfilter *lf) {
+  lf->mode_ref_delta_enabled = 1;
+  lf->mode_ref_delta_update = 1;
+
+  lf->ref_deltas[INTRA_FRAME] = 1;
+  lf->ref_deltas[LAST_FRAME] = 0;
+  lf->ref_deltas[GOLDEN_FRAME] = -1;
+  lf->ref_deltas[ALTREF_FRAME] = -1;
+
+  lf->mode_deltas[0] = 0;
+  lf->mode_deltas[1] = 0;
+}
+
+void vp9_setup_past_independence(VP9_COMMON *cm) {
+  // Reset the segment feature data to the default stats:
+  // Features disabled, 0, with delta coding (Default state).
+  struct loopfilter *const lf = &cm->lf;
+
+  int i;
+  vp9_clearall_segfeatures(&cm->seg);
+  cm->seg.abs_delta = SEGMENT_DELTADATA;
+
+  if (cm->last_frame_seg_map && !cm->frame_parallel_decode)
+    memset(cm->last_frame_seg_map, 0, (cm->mi_rows * cm->mi_cols));
+
+  if (cm->current_frame_seg_map)
+    memset(cm->current_frame_seg_map, 0, (cm->mi_rows * cm->mi_cols));
+
+  // Reset the mode ref deltas for loop filter
+  vp9_zero(lf->last_ref_deltas);
+  vp9_zero(lf->last_mode_deltas);
+  set_default_lf_deltas(lf);
+
+  // To force update of the sharpness
+  lf->last_sharpness_level = -1;
+
+  vp9_default_coef_probs(cm);
+  init_mode_probs(cm->fc);
+  vp9_init_mv_probs(cm);
+  cm->fc->initialized = 1;
+
+  if (cm->frame_type == KEY_FRAME ||
+      cm->error_resilient_mode || cm->reset_frame_context == 3) {
+    // Reset all frame contexts.
+    for (i = 0; i < FRAME_CONTEXTS; ++i)
+      cm->frame_contexts[i] = *cm->fc;
+  } else if (cm->reset_frame_context == 2) {
+    // Reset only the frame context specified in the frame header.
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+  }
+
+  // prev_mip will only be allocated in encoder.
+  if (frame_is_intra_only(cm) && cm->prev_mip && !cm->frame_parallel_decode)
+    memset(cm->prev_mip, 0,
+           cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->prev_mip));
+
+  vp9_zero(cm->ref_frame_sign_bias);
+
+  cm->frame_context_idx = 0;
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_entropymode.h b/libvpx/libvpx/vp9/common/vp9_entropymode.h
new file mode 100644
index 0000000..0285be1
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_entropymode.h
@@ -0,0 +1,107 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_ENTROPYMODE_H_
+#define VP9_COMMON_VP9_ENTROPYMODE_H_
+
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_filter.h"
+#include "vpx_dsp/vpx_filter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define BLOCK_SIZE_GROUPS 4
+
+#define TX_SIZE_CONTEXTS 2
+
+#define INTER_OFFSET(mode) ((mode) - NEARESTMV)
+
+struct VP9Common;
+
+struct tx_probs {
+  vpx_prob p32x32[TX_SIZE_CONTEXTS][TX_SIZES - 1];
+  vpx_prob p16x16[TX_SIZE_CONTEXTS][TX_SIZES - 2];
+  vpx_prob p8x8[TX_SIZE_CONTEXTS][TX_SIZES - 3];
+};
+
+struct tx_counts {
+  unsigned int p32x32[TX_SIZE_CONTEXTS][TX_SIZES];
+  unsigned int p16x16[TX_SIZE_CONTEXTS][TX_SIZES - 1];
+  unsigned int p8x8[TX_SIZE_CONTEXTS][TX_SIZES - 2];
+  unsigned int tx_totals[TX_SIZES];
+};
+
+typedef struct frame_contexts {
+  vpx_prob y_mode_prob[BLOCK_SIZE_GROUPS][INTRA_MODES - 1];
+  vpx_prob uv_mode_prob[INTRA_MODES][INTRA_MODES - 1];
+  vpx_prob partition_prob[PARTITION_CONTEXTS][PARTITION_TYPES - 1];
+  vp9_coeff_probs_model coef_probs[TX_SIZES][PLANE_TYPES];
+  vpx_prob switchable_interp_prob[SWITCHABLE_FILTER_CONTEXTS]
+                                 [SWITCHABLE_FILTERS - 1];
+  vpx_prob inter_mode_probs[INTER_MODE_CONTEXTS][INTER_MODES - 1];
+  vpx_prob intra_inter_prob[INTRA_INTER_CONTEXTS];
+  vpx_prob comp_inter_prob[COMP_INTER_CONTEXTS];
+  vpx_prob single_ref_prob[REF_CONTEXTS][2];
+  vpx_prob comp_ref_prob[REF_CONTEXTS];
+  struct tx_probs tx_probs;
+  vpx_prob skip_probs[SKIP_CONTEXTS];
+  nmv_context nmvc;
+  int initialized;
+} FRAME_CONTEXT;
+
+typedef struct FRAME_COUNTS {
+  unsigned int y_mode[BLOCK_SIZE_GROUPS][INTRA_MODES];
+  unsigned int uv_mode[INTRA_MODES][INTRA_MODES];
+  unsigned int partition[PARTITION_CONTEXTS][PARTITION_TYPES];
+  vp9_coeff_count_model coef[TX_SIZES][PLANE_TYPES];
+  unsigned int eob_branch[TX_SIZES][PLANE_TYPES][REF_TYPES]
+                         [COEF_BANDS][COEFF_CONTEXTS];
+  unsigned int switchable_interp[SWITCHABLE_FILTER_CONTEXTS]
+                                [SWITCHABLE_FILTERS];
+  unsigned int inter_mode[INTER_MODE_CONTEXTS][INTER_MODES];
+  unsigned int intra_inter[INTRA_INTER_CONTEXTS][2];
+  unsigned int comp_inter[COMP_INTER_CONTEXTS][2];
+  unsigned int single_ref[REF_CONTEXTS][2][2];
+  unsigned int comp_ref[REF_CONTEXTS][2];
+  struct tx_counts tx;
+  unsigned int skip[SKIP_CONTEXTS][2];
+  nmv_context_counts mv;
+} FRAME_COUNTS;
+
+extern const vpx_prob vp9_kf_uv_mode_prob[INTRA_MODES][INTRA_MODES - 1];
+extern const vpx_prob vp9_kf_y_mode_prob[INTRA_MODES][INTRA_MODES]
+                                        [INTRA_MODES - 1];
+extern const vpx_prob vp9_kf_partition_probs[PARTITION_CONTEXTS]
+                                            [PARTITION_TYPES - 1];
+extern const vpx_tree_index vp9_intra_mode_tree[TREE_SIZE(INTRA_MODES)];
+extern const vpx_tree_index vp9_inter_mode_tree[TREE_SIZE(INTER_MODES)];
+extern const vpx_tree_index vp9_partition_tree[TREE_SIZE(PARTITION_TYPES)];
+extern const vpx_tree_index vp9_switchable_interp_tree
+                                [TREE_SIZE(SWITCHABLE_FILTERS)];
+
+void vp9_setup_past_independence(struct VP9Common *cm);
+
+void vp9_adapt_mode_probs(struct VP9Common *cm);
+
+void tx_counts_to_branch_counts_32x32(const unsigned int *tx_count_32x32p,
+                                      unsigned int (*ct_32x32p)[2]);
+void tx_counts_to_branch_counts_16x16(const unsigned int *tx_count_16x16p,
+                                      unsigned int (*ct_16x16p)[2]);
+void tx_counts_to_branch_counts_8x8(const unsigned int *tx_count_8x8p,
+                                    unsigned int (*ct_8x8p)[2]);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ENTROPYMODE_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_entropymv.c b/libvpx/libvpx/vp9/common/vp9_entropymv.c
new file mode 100644
index 0000000..566ae91
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_entropymv.c
@@ -0,0 +1,210 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_entropymv.h"
+
+const vpx_tree_index vp9_mv_joint_tree[TREE_SIZE(MV_JOINTS)] = {
+  -MV_JOINT_ZERO, 2,
+  -MV_JOINT_HNZVZ, 4,
+  -MV_JOINT_HZVNZ, -MV_JOINT_HNZVNZ
+};
+
+const vpx_tree_index vp9_mv_class_tree[TREE_SIZE(MV_CLASSES)] = {
+  -MV_CLASS_0, 2,
+  -MV_CLASS_1, 4,
+  6, 8,
+  -MV_CLASS_2, -MV_CLASS_3,
+  10, 12,
+  -MV_CLASS_4, -MV_CLASS_5,
+  -MV_CLASS_6, 14,
+  16, 18,
+  -MV_CLASS_7, -MV_CLASS_8,
+  -MV_CLASS_9, -MV_CLASS_10,
+};
+
+const vpx_tree_index vp9_mv_class0_tree[TREE_SIZE(CLASS0_SIZE)] = {
+  -0, -1,
+};
+
+const vpx_tree_index vp9_mv_fp_tree[TREE_SIZE(MV_FP_SIZE)] = {
+  -0, 2,
+  -1, 4,
+  -2, -3
+};
+
+static const nmv_context default_nmv_context = {
+  {32, 64, 96},
+  {
+    { // Vertical component
+      128,                                                  // sign
+      {224, 144, 192, 168, 192, 176, 192, 198, 198, 245},   // class
+      {216},                                                // class0
+      {136, 140, 148, 160, 176, 192, 224, 234, 234, 240},   // bits
+      {{128, 128, 64}, {96, 112, 64}},                      // class0_fp
+      {64, 96, 64},                                         // fp
+      160,                                                  // class0_hp bit
+      128,                                                  // hp
+    },
+    { // Horizontal component
+      128,                                                  // sign
+      {216, 128, 176, 160, 176, 176, 192, 198, 198, 208},   // class
+      {208},                                                // class0
+      {136, 140, 148, 160, 176, 192, 224, 234, 234, 240},   // bits
+      {{128, 128, 64}, {96, 112, 64}},                      // class0_fp
+      {64, 96, 64},                                         // fp
+      160,                                                  // class0_hp bit
+      128,                                                  // hp
+    }
+  },
+};
+
+static const uint8_t log_in_base_2[] = {
+  0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4,
+  4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10
+};
+
+static INLINE int mv_class_base(MV_CLASS_TYPE c) {
+  return c ? CLASS0_SIZE << (c + 2) : 0;
+}
+
+MV_CLASS_TYPE vp9_get_mv_class(int z, int *offset) {
+  const MV_CLASS_TYPE c = (z >= CLASS0_SIZE * 4096) ?
+      MV_CLASS_10 : (MV_CLASS_TYPE)log_in_base_2[z >> 3];
+  if (offset)
+    *offset = z - mv_class_base(c);
+  return c;
+}
+
+static void inc_mv_component(int v, nmv_component_counts *comp_counts,
+                             int incr, int usehp) {
+  int s, z, c, o, d, e, f;
+  assert(v != 0);            /* should not be zero */
+  s = v < 0;
+  comp_counts->sign[s] += incr;
+  z = (s ? -v : v) - 1;       /* magnitude - 1 */
+
+  c = vp9_get_mv_class(z, &o);
+  comp_counts->classes[c] += incr;
+
+  d = (o >> 3);               /* int mv data */
+  f = (o >> 1) & 3;           /* fractional pel mv data */
+  e = (o & 1);                /* high precision mv data */
+
+  if (c == MV_CLASS_0) {
+    comp_counts->class0[d] += incr;
+    comp_counts->class0_fp[d][f] += incr;
+    comp_counts->class0_hp[e] += usehp * incr;
+  } else {
+    int i;
+    int b = c + CLASS0_BITS - 1;  // number of bits
+    for (i = 0; i < b; ++i)
+      comp_counts->bits[i][((d >> i) & 1)] += incr;
+    comp_counts->fp[f] += incr;
+    comp_counts->hp[e] += usehp * incr;
+  }
+}
+
+void vp9_inc_mv(const MV *mv, nmv_context_counts *counts) {
+  if (counts != NULL) {
+    const MV_JOINT_TYPE j = vp9_get_mv_joint(mv);
+    ++counts->joints[j];
+
+    if (mv_joint_vertical(j)) {
+      inc_mv_component(mv->row, &counts->comps[0], 1, 1);
+    }
+
+    if (mv_joint_horizontal(j)) {
+      inc_mv_component(mv->col, &counts->comps[1], 1, 1);
+    }
+  }
+}
+
+void vp9_adapt_mv_probs(VP9_COMMON *cm, int allow_hp) {
+  int i, j;
+
+  nmv_context *fc = &cm->fc->nmvc;
+  const nmv_context *pre_fc = &cm->frame_contexts[cm->frame_context_idx].nmvc;
+  const nmv_context_counts *counts = &cm->counts.mv;
+
+  vpx_tree_merge_probs(vp9_mv_joint_tree, pre_fc->joints, counts->joints,
+                       fc->joints);
+
+  for (i = 0; i < 2; ++i) {
+    nmv_component *comp = &fc->comps[i];
+    const nmv_component *pre_comp = &pre_fc->comps[i];
+    const nmv_component_counts *c = &counts->comps[i];
+
+    comp->sign = mode_mv_merge_probs(pre_comp->sign, c->sign);
+    vpx_tree_merge_probs(vp9_mv_class_tree, pre_comp->classes, c->classes,
+                         comp->classes);
+    vpx_tree_merge_probs(vp9_mv_class0_tree, pre_comp->class0, c->class0,
+                         comp->class0);
+
+    for (j = 0; j < MV_OFFSET_BITS; ++j)
+      comp->bits[j] = mode_mv_merge_probs(pre_comp->bits[j], c->bits[j]);
+
+    for (j = 0; j < CLASS0_SIZE; ++j)
+      vpx_tree_merge_probs(vp9_mv_fp_tree, pre_comp->class0_fp[j],
+                           c->class0_fp[j], comp->class0_fp[j]);
+
+    vpx_tree_merge_probs(vp9_mv_fp_tree, pre_comp->fp, c->fp, comp->fp);
+
+    if (allow_hp) {
+      comp->class0_hp = mode_mv_merge_probs(pre_comp->class0_hp, c->class0_hp);
+      comp->hp = mode_mv_merge_probs(pre_comp->hp, c->hp);
+    }
+  }
+}
+
+void vp9_init_mv_probs(VP9_COMMON *cm) {
+  cm->fc->nmvc = default_nmv_context;
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_entropymv.h b/libvpx/libvpx/vp9/common/vp9_entropymv.h
new file mode 100644
index 0000000..2f05ad4
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_entropymv.h
@@ -0,0 +1,140 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_COMMON_VP9_ENTROPYMV_H_
+#define VP9_COMMON_VP9_ENTROPYMV_H_
+
+#include "./vpx_config.h"
+
+#include "vpx_dsp/prob.h"
+
+#include "vp9/common/vp9_mv.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9Common;
+
+void vp9_init_mv_probs(struct VP9Common *cm);
+
+void vp9_adapt_mv_probs(struct VP9Common *cm, int usehp);
+
+// Integer pel reference mv threshold for use of high-precision 1/8 mv
+#define COMPANDED_MVREF_THRESH 8
+
+static INLINE int use_mv_hp(const MV *ref) {
+  return (abs(ref->row) >> 3) < COMPANDED_MVREF_THRESH &&
+         (abs(ref->col) >> 3) < COMPANDED_MVREF_THRESH;
+}
+
+#define MV_UPDATE_PROB 252
+
+/* Symbols for coding which components are zero jointly */
+#define MV_JOINTS     4
+typedef enum {
+  MV_JOINT_ZERO = 0,             /* Zero vector */
+  MV_JOINT_HNZVZ = 1,            /* Vert zero, hor nonzero */
+  MV_JOINT_HZVNZ = 2,            /* Hor zero, vert nonzero */
+  MV_JOINT_HNZVNZ = 3,           /* Both components nonzero */
+} MV_JOINT_TYPE;
+
+static INLINE int mv_joint_vertical(MV_JOINT_TYPE type) {
+  return type == MV_JOINT_HZVNZ || type == MV_JOINT_HNZVNZ;
+}
+
+static INLINE int mv_joint_horizontal(MV_JOINT_TYPE type) {
+  return type == MV_JOINT_HNZVZ || type == MV_JOINT_HNZVNZ;
+}
+
+/* Symbols for coding magnitude class of nonzero components */
+#define MV_CLASSES     11
+typedef enum {
+  MV_CLASS_0 = 0,      /* (0, 2]     integer pel */
+  MV_CLASS_1 = 1,      /* (2, 4]     integer pel */
+  MV_CLASS_2 = 2,      /* (4, 8]     integer pel */
+  MV_CLASS_3 = 3,      /* (8, 16]    integer pel */
+  MV_CLASS_4 = 4,      /* (16, 32]   integer pel */
+  MV_CLASS_5 = 5,      /* (32, 64]   integer pel */
+  MV_CLASS_6 = 6,      /* (64, 128]  integer pel */
+  MV_CLASS_7 = 7,      /* (128, 256] integer pel */
+  MV_CLASS_8 = 8,      /* (256, 512] integer pel */
+  MV_CLASS_9 = 9,      /* (512, 1024] integer pel */
+  MV_CLASS_10 = 10,    /* (1024,2048] integer pel */
+} MV_CLASS_TYPE;
+
+#define CLASS0_BITS    1  /* bits at integer precision for class 0 */
+#define CLASS0_SIZE    (1 << CLASS0_BITS)
+#define MV_OFFSET_BITS (MV_CLASSES + CLASS0_BITS - 2)
+#define MV_FP_SIZE 4
+
+#define MV_MAX_BITS    (MV_CLASSES + CLASS0_BITS + 2)
+#define MV_MAX         ((1 << MV_MAX_BITS) - 1)
+#define MV_VALS        ((MV_MAX << 1) + 1)
+
+#define MV_IN_USE_BITS 14
+#define MV_UPP   ((1 << MV_IN_USE_BITS) - 1)
+#define MV_LOW   (-(1 << MV_IN_USE_BITS))
+
+extern const vpx_tree_index vp9_mv_joint_tree[];
+extern const vpx_tree_index vp9_mv_class_tree[];
+extern const vpx_tree_index vp9_mv_class0_tree[];
+extern const vpx_tree_index vp9_mv_fp_tree[];
+
+typedef struct {
+  vpx_prob sign;
+  vpx_prob classes[MV_CLASSES - 1];
+  vpx_prob class0[CLASS0_SIZE - 1];
+  vpx_prob bits[MV_OFFSET_BITS];
+  vpx_prob class0_fp[CLASS0_SIZE][MV_FP_SIZE - 1];
+  vpx_prob fp[MV_FP_SIZE - 1];
+  vpx_prob class0_hp;
+  vpx_prob hp;
+} nmv_component;
+
+typedef struct {
+  vpx_prob joints[MV_JOINTS - 1];
+  nmv_component comps[2];
+} nmv_context;
+
+static INLINE MV_JOINT_TYPE vp9_get_mv_joint(const MV *mv) {
+  if (mv->row == 0) {
+    return mv->col == 0 ? MV_JOINT_ZERO : MV_JOINT_HNZVZ;
+  } else {
+    return mv->col == 0 ? MV_JOINT_HZVNZ : MV_JOINT_HNZVNZ;
+  }
+}
+
+MV_CLASS_TYPE vp9_get_mv_class(int z, int *offset);
+
+typedef struct {
+  unsigned int sign[2];
+  unsigned int classes[MV_CLASSES];
+  unsigned int class0[CLASS0_SIZE];
+  unsigned int bits[MV_OFFSET_BITS][2];
+  unsigned int class0_fp[CLASS0_SIZE][MV_FP_SIZE];
+  unsigned int fp[MV_FP_SIZE];
+  unsigned int class0_hp[2];
+  unsigned int hp[2];
+} nmv_component_counts;
+
+typedef struct {
+  unsigned int joints[MV_JOINTS];
+  nmv_component_counts comps[2];
+} nmv_context_counts;
+
+void vp9_inc_mv(const MV *mv, nmv_context_counts *mvctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ENTROPYMV_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_enums.h b/libvpx/libvpx/vp9/common/vp9_enums.h
new file mode 100644
index 0000000..d089f23
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_enums.h
@@ -0,0 +1,147 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_ENUMS_H_
+#define VP9_COMMON_VP9_ENUMS_H_
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MI_SIZE_LOG2 3
+#define MI_BLOCK_SIZE_LOG2 (6 - MI_SIZE_LOG2)  // 64 = 2^6
+
+#define MI_SIZE (1 << MI_SIZE_LOG2)  // pixels per mi-unit
+#define MI_BLOCK_SIZE (1 << MI_BLOCK_SIZE_LOG2)  // mi-units per max block
+
+#define MI_MASK (MI_BLOCK_SIZE - 1)
+
+// Bitstream profiles indicated by 2-3 bits in the uncompressed header.
+// 00: Profile 0.  8-bit 4:2:0 only.
+// 10: Profile 1.  8-bit 4:4:4, 4:2:2, and 4:4:0.
+// 01: Profile 2.  10-bit and 12-bit color only, with 4:2:0 sampling.
+// 110: Profile 3. 10-bit and 12-bit color only, with 4:2:2/4:4:4/4:4:0
+//                 sampling.
+// 111: Undefined profile.
+typedef enum BITSTREAM_PROFILE {
+  PROFILE_0,
+  PROFILE_1,
+  PROFILE_2,
+  PROFILE_3,
+  MAX_PROFILES
+} BITSTREAM_PROFILE;
+
+#define BLOCK_4X4     0
+#define BLOCK_4X8     1
+#define BLOCK_8X4     2
+#define BLOCK_8X8     3
+#define BLOCK_8X16    4
+#define BLOCK_16X8    5
+#define BLOCK_16X16   6
+#define BLOCK_16X32   7
+#define BLOCK_32X16   8
+#define BLOCK_32X32   9
+#define BLOCK_32X64  10
+#define BLOCK_64X32  11
+#define BLOCK_64X64  12
+#define BLOCK_SIZES  13
+#define BLOCK_INVALID BLOCK_SIZES
+typedef uint8_t BLOCK_SIZE;
+
+typedef enum PARTITION_TYPE {
+  PARTITION_NONE,
+  PARTITION_HORZ,
+  PARTITION_VERT,
+  PARTITION_SPLIT,
+  PARTITION_TYPES,
+  PARTITION_INVALID = PARTITION_TYPES
+} PARTITION_TYPE;
+
+typedef char PARTITION_CONTEXT;
+#define PARTITION_PLOFFSET   4  // number of probability models per block size
+#define PARTITION_CONTEXTS (4 * PARTITION_PLOFFSET)
+
+// block transform size
+typedef uint8_t TX_SIZE;
+#define TX_4X4   ((TX_SIZE)0)   // 4x4 transform
+#define TX_8X8   ((TX_SIZE)1)   // 8x8 transform
+#define TX_16X16 ((TX_SIZE)2)   // 16x16 transform
+#define TX_32X32 ((TX_SIZE)3)   // 32x32 transform
+#define TX_SIZES ((TX_SIZE)4)
+
+// frame transform mode
+typedef enum {
+  ONLY_4X4            = 0,        // only 4x4 transform used
+  ALLOW_8X8           = 1,        // allow block transform size up to 8x8
+  ALLOW_16X16         = 2,        // allow block transform size up to 16x16
+  ALLOW_32X32         = 3,        // allow block transform size up to 32x32
+  TX_MODE_SELECT      = 4,        // transform specified for each block
+  TX_MODES            = 5,
+} TX_MODE;
+
+typedef enum {
+  DCT_DCT   = 0,                      // DCT  in both horizontal and vertical
+  ADST_DCT  = 1,                      // ADST in vertical, DCT in horizontal
+  DCT_ADST  = 2,                      // DCT  in vertical, ADST in horizontal
+  ADST_ADST = 3,                      // ADST in both directions
+  TX_TYPES = 4
+} TX_TYPE;
+
+typedef enum {
+  VP9_LAST_FLAG = 1 << 0,
+  VP9_GOLD_FLAG = 1 << 1,
+  VP9_ALT_FLAG = 1 << 2,
+} VP9_REFFRAME;
+
+typedef enum {
+  PLANE_TYPE_Y  = 0,
+  PLANE_TYPE_UV = 1,
+  PLANE_TYPES
+} PLANE_TYPE;
+
+#define DC_PRED    0       // Average of above and left pixels
+#define V_PRED     1       // Vertical
+#define H_PRED     2       // Horizontal
+#define D45_PRED   3       // Directional 45  deg = round(arctan(1/1) * 180/pi)
+#define D135_PRED  4       // Directional 135 deg = 180 - 45
+#define D117_PRED  5       // Directional 117 deg = 180 - 63
+#define D153_PRED  6       // Directional 153 deg = 180 - 27
+#define D207_PRED  7       // Directional 207 deg = 180 + 27
+#define D63_PRED   8       // Directional 63  deg = round(arctan(2/1) * 180/pi)
+#define TM_PRED    9       // True-motion
+#define NEARESTMV 10
+#define NEARMV    11
+#define ZEROMV    12
+#define NEWMV     13
+#define MB_MODE_COUNT 14
+typedef uint8_t PREDICTION_MODE;
+
+#define INTRA_MODES (TM_PRED + 1)
+
+#define INTER_MODES (1 + NEWMV - NEARESTMV)
+
+#define SKIP_CONTEXTS 3
+#define INTER_MODE_CONTEXTS 7
+
+/* Segment Feature Masks */
+#define MAX_MV_REF_CANDIDATES 2
+
+#define INTRA_INTER_CONTEXTS 4
+#define COMP_INTER_CONTEXTS 5
+#define REF_CONTEXTS 5
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ENUMS_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_filter.c b/libvpx/libvpx/vp9/common/vp9_filter.c
new file mode 100644
index 0000000..4b2198f
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_filter.c
@@ -0,0 +1,104 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_filter.h"
+
+DECLARE_ALIGNED(256, static const InterpKernel,
+                bilinear_filters[SUBPEL_SHIFTS]) = {
+  { 0, 0, 0, 128,   0, 0, 0, 0 },
+  { 0, 0, 0, 120,   8, 0, 0, 0 },
+  { 0, 0, 0, 112,  16, 0, 0, 0 },
+  { 0, 0, 0, 104,  24, 0, 0, 0 },
+  { 0, 0, 0,  96,  32, 0, 0, 0 },
+  { 0, 0, 0,  88,  40, 0, 0, 0 },
+  { 0, 0, 0,  80,  48, 0, 0, 0 },
+  { 0, 0, 0,  72,  56, 0, 0, 0 },
+  { 0, 0, 0,  64,  64, 0, 0, 0 },
+  { 0, 0, 0,  56,  72, 0, 0, 0 },
+  { 0, 0, 0,  48,  80, 0, 0, 0 },
+  { 0, 0, 0,  40,  88, 0, 0, 0 },
+  { 0, 0, 0,  32,  96, 0, 0, 0 },
+  { 0, 0, 0,  24, 104, 0, 0, 0 },
+  { 0, 0, 0,  16, 112, 0, 0, 0 },
+  { 0, 0, 0,   8, 120, 0, 0, 0 }
+};
+
+// Lagrangian interpolation filter
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8[SUBPEL_SHIFTS]) = {
+  { 0,   0,   0, 128,   0,   0,   0,  0},
+  { 0,   1,  -5, 126,   8,  -3,   1,  0},
+  { -1,   3, -10, 122,  18,  -6,   2,  0},
+  { -1,   4, -13, 118,  27,  -9,   3, -1},
+  { -1,   4, -16, 112,  37, -11,   4, -1},
+  { -1,   5, -18, 105,  48, -14,   4, -1},
+  { -1,   5, -19,  97,  58, -16,   5, -1},
+  { -1,   6, -19,  88,  68, -18,   5, -1},
+  { -1,   6, -19,  78,  78, -19,   6, -1},
+  { -1,   5, -18,  68,  88, -19,   6, -1},
+  { -1,   5, -16,  58,  97, -19,   5, -1},
+  { -1,   4, -14,  48, 105, -18,   5, -1},
+  { -1,   4, -11,  37, 112, -16,   4, -1},
+  { -1,   3,  -9,  27, 118, -13,   4, -1},
+  { 0,   2,  -6,  18, 122, -10,   3, -1},
+  { 0,   1,  -3,   8, 126,  -5,   1,  0}
+};
+
+// DCT based filter
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8s[SUBPEL_SHIFTS]) = {
+  {0,   0,   0, 128,   0,   0,   0, 0},
+  {-1,   3,  -7, 127,   8,  -3,   1, 0},
+  {-2,   5, -13, 125,  17,  -6,   3, -1},
+  {-3,   7, -17, 121,  27, -10,   5, -2},
+  {-4,   9, -20, 115,  37, -13,   6, -2},
+  {-4,  10, -23, 108,  48, -16,   8, -3},
+  {-4,  10, -24, 100,  59, -19,   9, -3},
+  {-4,  11, -24,  90,  70, -21,  10, -4},
+  {-4,  11, -23,  80,  80, -23,  11, -4},
+  {-4,  10, -21,  70,  90, -24,  11, -4},
+  {-3,   9, -19,  59, 100, -24,  10, -4},
+  {-3,   8, -16,  48, 108, -23,  10, -4},
+  {-2,   6, -13,  37, 115, -20,   9, -4},
+  {-2,   5, -10,  27, 121, -17,   7, -3},
+  {-1,   3,  -6,  17, 125, -13,   5, -2},
+  {0,   1,  -3,   8, 127,  -7,   3, -1}
+};
+
+// freqmultiplier = 0.5
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8lp[SUBPEL_SHIFTS]) = {
+  { 0,  0,  0, 128,  0,  0,  0,  0},
+  {-3, -1, 32,  64, 38,  1, -3,  0},
+  {-2, -2, 29,  63, 41,  2, -3,  0},
+  {-2, -2, 26,  63, 43,  4, -4,  0},
+  {-2, -3, 24,  62, 46,  5, -4,  0},
+  {-2, -3, 21,  60, 49,  7, -4,  0},
+  {-1, -4, 18,  59, 51,  9, -4,  0},
+  {-1, -4, 16,  57, 53, 12, -4, -1},
+  {-1, -4, 14,  55, 55, 14, -4, -1},
+  {-1, -4, 12,  53, 57, 16, -4, -1},
+  { 0, -4,  9,  51, 59, 18, -4, -1},
+  { 0, -4,  7,  49, 60, 21, -3, -2},
+  { 0, -4,  5,  46, 62, 24, -3, -2},
+  { 0, -4,  4,  43, 63, 26, -2, -2},
+  { 0, -3,  2,  41, 63, 29, -2, -2},
+  { 0, -3,  1,  38, 64, 32, -1, -3}
+};
+
+
+const InterpKernel *vp9_filter_kernels[4] = {
+  sub_pel_filters_8,
+  sub_pel_filters_8lp,
+  sub_pel_filters_8s,
+  bilinear_filters
+};
diff --git a/libvpx/libvpx/vp9/common/vp9_filter.h b/libvpx/libvpx/vp9/common/vp9_filter.h
new file mode 100644
index 0000000..efa24bc
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_filter.h
@@ -0,0 +1,42 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_FILTER_H_
+#define VP9_COMMON_VP9_FILTER_H_
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_filter.h"
+#include "vpx_ports/mem.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define EIGHTTAP            0
+#define EIGHTTAP_SMOOTH     1
+#define EIGHTTAP_SHARP      2
+#define SWITCHABLE_FILTERS  3 /* Number of switchable filters */
+#define BILINEAR            3
+// The codec can operate in four possible inter prediction filter mode:
+// 8-tap, 8-tap-smooth, 8-tap-sharp, and switching between the three.
+#define SWITCHABLE_FILTER_CONTEXTS (SWITCHABLE_FILTERS + 1)
+#define SWITCHABLE 4 /* should be the last one */
+
+typedef uint8_t INTERP_FILTER;
+
+extern const InterpKernel *vp9_filter_kernels[4];
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_FILTER_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_frame_buffers.c b/libvpx/libvpx/vp9/common/vp9_frame_buffers.c
new file mode 100644
index 0000000..0f41d66
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_frame_buffers.c
@@ -0,0 +1,86 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_frame_buffers.h"
+#include "vpx_mem/vpx_mem.h"
+
+int vp9_alloc_internal_frame_buffers(InternalFrameBufferList *list) {
+  assert(list != NULL);
+  vp9_free_internal_frame_buffers(list);
+
+  list->num_internal_frame_buffers =
+      VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS;
+  list->int_fb =
+      (InternalFrameBuffer *)vpx_calloc(list->num_internal_frame_buffers,
+                                        sizeof(*list->int_fb));
+  return (list->int_fb == NULL);
+}
+
+void vp9_free_internal_frame_buffers(InternalFrameBufferList *list) {
+  int i;
+
+  assert(list != NULL);
+
+  for (i = 0; i < list->num_internal_frame_buffers; ++i) {
+    vpx_free(list->int_fb[i].data);
+    list->int_fb[i].data = NULL;
+  }
+  vpx_free(list->int_fb);
+  list->int_fb = NULL;
+}
+
+int vp9_get_frame_buffer(void *cb_priv, size_t min_size,
+                         vpx_codec_frame_buffer_t *fb) {
+  int i;
+  InternalFrameBufferList *const int_fb_list =
+      (InternalFrameBufferList *)cb_priv;
+  if (int_fb_list == NULL)
+    return -1;
+
+  // Find a free frame buffer.
+  for (i = 0; i < int_fb_list->num_internal_frame_buffers; ++i) {
+    if (!int_fb_list->int_fb[i].in_use)
+      break;
+  }
+
+  if (i == int_fb_list->num_internal_frame_buffers)
+    return -1;
+
+  if (int_fb_list->int_fb[i].size < min_size) {
+    int_fb_list->int_fb[i].data =
+        (uint8_t *)vpx_realloc(int_fb_list->int_fb[i].data, min_size);
+    if (!int_fb_list->int_fb[i].data)
+      return -1;
+
+    // This memset is needed for fixing valgrind error from C loop filter
+    // due to access uninitialized memory in frame border. It could be
+    // removed if border is totally removed.
+    memset(int_fb_list->int_fb[i].data, 0, min_size);
+    int_fb_list->int_fb[i].size = min_size;
+  }
+
+  fb->data = int_fb_list->int_fb[i].data;
+  fb->size = int_fb_list->int_fb[i].size;
+  int_fb_list->int_fb[i].in_use = 1;
+
+  // Set the frame buffer's private data to point at the internal frame buffer.
+  fb->priv = &int_fb_list->int_fb[i];
+  return 0;
+}
+
+int vp9_release_frame_buffer(void *cb_priv, vpx_codec_frame_buffer_t *fb) {
+  InternalFrameBuffer *const int_fb = (InternalFrameBuffer *)fb->priv;
+  (void)cb_priv;
+  if (int_fb)
+    int_fb->in_use = 0;
+  return 0;
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_frame_buffers.h b/libvpx/libvpx/vp9/common/vp9_frame_buffers.h
new file mode 100644
index 0000000..e2cfe61
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_frame_buffers.h
@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_FRAME_BUFFERS_H_
+#define VP9_COMMON_VP9_FRAME_BUFFERS_H_
+
+#include "vpx/vpx_frame_buffer.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct InternalFrameBuffer {
+  uint8_t *data;
+  size_t size;
+  int in_use;
+} InternalFrameBuffer;
+
+typedef struct InternalFrameBufferList {
+  int num_internal_frame_buffers;
+  InternalFrameBuffer *int_fb;
+} InternalFrameBufferList;
+
+// Initializes |list|. Returns 0 on success.
+int vp9_alloc_internal_frame_buffers(InternalFrameBufferList *list);
+
+// Free any data allocated to the frame buffers.
+void vp9_free_internal_frame_buffers(InternalFrameBufferList *list);
+
+// Callback used by libvpx to request an external frame buffer. |cb_priv|
+// Callback private data, which points to an InternalFrameBufferList.
+// |min_size| is the minimum size in bytes needed to decode the next frame.
+// |fb| pointer to the frame buffer.
+int vp9_get_frame_buffer(void *cb_priv, size_t min_size,
+                         vpx_codec_frame_buffer_t *fb);
+
+// Callback used by libvpx when there are no references to the frame buffer.
+// |cb_priv| is not used. |fb| pointer to the frame buffer.
+int vp9_release_frame_buffer(void *cb_priv, vpx_codec_frame_buffer_t *fb);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_FRAME_BUFFERS_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_idct.c b/libvpx/libvpx/vp9/common/vp9_idct.c
new file mode 100644
index 0000000..1b42014
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_idct.c
@@ -0,0 +1,405 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_idct.h"
+#include "vpx_dsp/inv_txfm.h"
+#include "vpx_ports/mem.h"
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                         int tx_type) {
+  const transform_2d IHT_4[] = {
+    { idct4_c, idct4_c  },  // DCT_DCT  = 0
+    { iadst4_c, idct4_c  },   // ADST_DCT = 1
+    { idct4_c, iadst4_c },    // DCT_ADST = 2
+    { iadst4_c, iadst4_c }      // ADST_ADST = 3
+  };
+
+  int i, j;
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[4], temp_out[4];
+
+  // inverse transform row vectors
+  for (i = 0; i < 4; ++i) {
+    IHT_4[tx_type].rows(input, outptr);
+    input  += 4;
+    outptr += 4;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j)
+      temp_in[j] = out[j * 4 + i];
+    IHT_4[tx_type].cols(temp_in, temp_out);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 4));
+    }
+  }
+}
+
+static const transform_2d IHT_8[] = {
+  { idct8_c,  idct8_c  },  // DCT_DCT  = 0
+  { iadst8_c, idct8_c  },  // ADST_DCT = 1
+  { idct8_c,  iadst8_c },  // DCT_ADST = 2
+  { iadst8_c, iadst8_c }   // ADST_ADST = 3
+};
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                         int tx_type) {
+  int i, j;
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[8], temp_out[8];
+  const transform_2d ht = IHT_8[tx_type];
+
+  // inverse transform row vectors
+  for (i = 0; i < 8; ++i) {
+    ht.rows(input, outptr);
+    input += 8;
+    outptr += 8;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 5));
+    }
+  }
+}
+
+static const transform_2d IHT_16[] = {
+  { idct16_c,  idct16_c  },  // DCT_DCT  = 0
+  { iadst16_c, idct16_c  },  // ADST_DCT = 1
+  { idct16_c,  iadst16_c },  // DCT_ADST = 2
+  { iadst16_c, iadst16_c }   // ADST_ADST = 3
+};
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  int i, j;
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[16], temp_out[16];
+  const transform_2d ht = IHT_16[tx_type];
+
+  // Rows
+  for (i = 0; i < 16; ++i) {
+    ht.rows(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j * 16 + i];
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+// idct
+void vp9_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob) {
+  if (eob > 1)
+    vpx_idct4x4_16_add(input, dest, stride);
+  else
+    vpx_idct4x4_1_add(input, dest, stride);
+}
+
+
+void vp9_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob) {
+  if (eob > 1)
+    vpx_iwht4x4_16_add(input, dest, stride);
+  else
+    vpx_iwht4x4_1_add(input, dest, stride);
+}
+
+void vp9_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob) {
+  // If dc is 1, then input[0] is the reconstructed value, do not need
+  // dequantization. Also, when dc is 1, dc is counted in eobs, namely eobs >=1.
+
+  // The calculation can be simplified if there are not many non-zero dct
+  // coefficients. Use eobs to decide what to do.
+  // TODO(yunqingwang): "eobs = 1" case is also handled in vp9_short_idct8x8_c.
+  // Combine that with code here.
+  if (eob == 1)
+    // DC only DCT coefficient
+    vpx_idct8x8_1_add(input, dest, stride);
+  else if (eob <= 12)
+    vpx_idct8x8_12_add(input, dest, stride);
+  else
+    vpx_idct8x8_64_add(input, dest, stride);
+}
+
+void vp9_idct16x16_add(const tran_low_t *input, uint8_t *dest, int stride,
+                       int eob) {
+  /* The calculation can be simplified if there are not many non-zero dct
+   * coefficients. Use eobs to separate different cases. */
+  if (eob == 1)
+    /* DC only DCT coefficient. */
+    vpx_idct16x16_1_add(input, dest, stride);
+  else if (eob <= 10)
+    vpx_idct16x16_10_add(input, dest, stride);
+  else
+    vpx_idct16x16_256_add(input, dest, stride);
+}
+
+void vp9_idct32x32_add(const tran_low_t *input, uint8_t *dest, int stride,
+                       int eob) {
+  if (eob == 1)
+    vpx_idct32x32_1_add(input, dest, stride);
+  else if (eob <= 34)
+    // non-zero coeff only in upper-left 8x8
+    vpx_idct32x32_34_add(input, dest, stride);
+  else if (eob <= 135)
+    // non-zero coeff only in upper-left 16x16
+    vpx_idct32x32_135_add(input, dest, stride);
+  else
+    vpx_idct32x32_1024_add(input, dest, stride);
+}
+
+// iht
+void vp9_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                    int stride, int eob) {
+  if (tx_type == DCT_DCT)
+    vp9_idct4x4_add(input, dest, stride, eob);
+  else
+    vp9_iht4x4_16_add(input, dest, stride, tx_type);
+}
+
+void vp9_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                    int stride, int eob) {
+  if (tx_type == DCT_DCT) {
+    vp9_idct8x8_add(input, dest, stride, eob);
+  } else {
+    vp9_iht8x8_64_add(input, dest, stride, tx_type);
+  }
+}
+
+void vp9_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                      int stride, int eob) {
+  if (tx_type == DCT_DCT) {
+    vp9_idct16x16_add(input, dest, stride, eob);
+  } else {
+    vp9_iht16x16_256_add(input, dest, stride, tx_type);
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int tx_type, int bd) {
+  const highbd_transform_2d IHT_4[] = {
+    { vpx_highbd_idct4_c, vpx_highbd_idct4_c  },    // DCT_DCT  = 0
+    { vpx_highbd_iadst4_c, vpx_highbd_idct4_c },    // ADST_DCT = 1
+    { vpx_highbd_idct4_c, vpx_highbd_iadst4_c },    // DCT_ADST = 2
+    { vpx_highbd_iadst4_c, vpx_highbd_iadst4_c }    // ADST_ADST = 3
+  };
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[4], temp_out[4];
+
+  // Inverse transform row vectors.
+  for (i = 0; i < 4; ++i) {
+    IHT_4[tx_type].rows(input, outptr, bd);
+    input  += 4;
+    outptr += 4;
+  }
+
+  // Inverse transform column vectors.
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j)
+      temp_in[j] = out[j * 4 + i];
+    IHT_4[tx_type].cols(temp_in, temp_out, bd);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
+    }
+  }
+}
+
+static const highbd_transform_2d HIGH_IHT_8[] = {
+  { vpx_highbd_idct8_c,  vpx_highbd_idct8_c  },  // DCT_DCT  = 0
+  { vpx_highbd_iadst8_c, vpx_highbd_idct8_c  },  // ADST_DCT = 1
+  { vpx_highbd_idct8_c,  vpx_highbd_iadst8_c },  // DCT_ADST = 2
+  { vpx_highbd_iadst8_c, vpx_highbd_iadst8_c }   // ADST_ADST = 3
+};
+
+void vp9_highbd_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int tx_type, int bd) {
+  int i, j;
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[8], temp_out[8];
+  const highbd_transform_2d ht = HIGH_IHT_8[tx_type];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Inverse transform row vectors.
+  for (i = 0; i < 8; ++i) {
+    ht.rows(input, outptr, bd);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Inverse transform column vectors.
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    ht.cols(temp_in, temp_out, bd);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+    }
+  }
+}
+
+static const highbd_transform_2d HIGH_IHT_16[] = {
+  { vpx_highbd_idct16_c,  vpx_highbd_idct16_c  },  // DCT_DCT  = 0
+  { vpx_highbd_iadst16_c, vpx_highbd_idct16_c  },  // ADST_DCT = 1
+  { vpx_highbd_idct16_c,  vpx_highbd_iadst16_c },  // DCT_ADST = 2
+  { vpx_highbd_iadst16_c, vpx_highbd_iadst16_c }   // ADST_ADST = 3
+};
+
+void vp9_highbd_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int tx_type, int bd) {
+  int i, j;
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[16], temp_out[16];
+  const highbd_transform_2d ht = HIGH_IHT_16[tx_type];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  for (i = 0; i < 16; ++i) {
+    ht.rows(input, outptr, bd);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j * 16 + i];
+    ht.cols(temp_in, temp_out, bd);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+// idct
+void vp9_highbd_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd) {
+  if (eob > 1)
+    vpx_highbd_idct4x4_16_add(input, dest, stride, bd);
+  else
+    vpx_highbd_idct4x4_1_add(input, dest, stride, bd);
+}
+
+
+void vp9_highbd_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd) {
+  if (eob > 1)
+    vpx_highbd_iwht4x4_16_add(input, dest, stride, bd);
+  else
+    vpx_highbd_iwht4x4_1_add(input, dest, stride, bd);
+}
+
+void vp9_highbd_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd) {
+  // If dc is 1, then input[0] is the reconstructed value, do not need
+  // dequantization. Also, when dc is 1, dc is counted in eobs, namely eobs >=1.
+
+  // The calculation can be simplified if there are not many non-zero dct
+  // coefficients. Use eobs to decide what to do.
+  // TODO(yunqingwang): "eobs = 1" case is also handled in vp9_short_idct8x8_c.
+  // Combine that with code here.
+  // DC only DCT coefficient
+  if (eob == 1) {
+    vpx_highbd_idct8x8_1_add(input, dest, stride, bd);
+  } else if (eob <= 10) {
+    vpx_highbd_idct8x8_10_add(input, dest, stride, bd);
+  } else {
+    vpx_highbd_idct8x8_64_add(input, dest, stride, bd);
+  }
+}
+
+void vp9_highbd_idct16x16_add(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, int bd) {
+  // The calculation can be simplified if there are not many non-zero dct
+  // coefficients. Use eobs to separate different cases.
+  // DC only DCT coefficient.
+  if (eob == 1) {
+    vpx_highbd_idct16x16_1_add(input, dest, stride, bd);
+  } else if (eob <= 10) {
+    vpx_highbd_idct16x16_10_add(input, dest, stride, bd);
+  } else {
+    vpx_highbd_idct16x16_256_add(input, dest, stride, bd);
+  }
+}
+
+void vp9_highbd_idct32x32_add(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, int bd) {
+  // Non-zero coeff only in upper-left 8x8
+  if (eob == 1) {
+    vpx_highbd_idct32x32_1_add(input, dest, stride, bd);
+  } else if (eob <= 34) {
+    vpx_highbd_idct32x32_34_add(input, dest, stride, bd);
+  } else {
+    vpx_highbd_idct32x32_1024_add(input, dest, stride, bd);
+  }
+}
+
+// iht
+void vp9_highbd_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd) {
+  if (tx_type == DCT_DCT)
+    vp9_highbd_idct4x4_add(input, dest, stride, eob, bd);
+  else
+    vp9_highbd_iht4x4_16_add(input, dest, stride, tx_type, bd);
+}
+
+void vp9_highbd_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd) {
+  if (tx_type == DCT_DCT) {
+    vp9_highbd_idct8x8_add(input, dest, stride, eob, bd);
+  } else {
+    vp9_highbd_iht8x8_64_add(input, dest, stride, tx_type, bd);
+  }
+}
+
+void vp9_highbd_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd) {
+  if (tx_type == DCT_DCT) {
+    vp9_highbd_idct16x16_add(input, dest, stride, eob, bd);
+  } else {
+    vp9_highbd_iht16x16_256_add(input, dest, stride, tx_type, bd);
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vp9/common/vp9_idct.h b/libvpx/libvpx/vp9/common/vp9_idct.h
new file mode 100644
index 0000000..b5a3fbf
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_idct.h
@@ -0,0 +1,81 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_IDCT_H_
+#define VP9_COMMON_VP9_IDCT_H_
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+#include "vpx_dsp/inv_txfm.h"
+#include "vpx_dsp/txfm_common.h"
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef void (*transform_1d)(const tran_low_t*, tran_low_t*);
+
+typedef struct {
+  transform_1d cols, rows;  // vertical and horizontal
+} transform_2d;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*highbd_transform_1d)(const tran_low_t*, tran_low_t*, int bd);
+
+typedef struct {
+  highbd_transform_1d cols, rows;  // vertical and horizontal
+} highbd_transform_2d;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+void vp9_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+void vp9_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+void vp9_idct16x16_add(const tran_low_t *input, uint8_t *dest, int stride,
+                       int eob);
+void vp9_idct32x32_add(const tran_low_t *input, uint8_t *dest, int stride,
+                       int eob);
+
+void vp9_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                    int stride, int eob);
+void vp9_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                    int stride, int eob);
+void vp9_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                      int stride, int eob);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd);
+void vp9_highbd_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd);
+void vp9_highbd_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd);
+void vp9_highbd_idct16x16_add(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, int bd);
+void vp9_highbd_idct32x32_add(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, int bd);
+void vp9_highbd_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd);
+void vp9_highbd_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd);
+void vp9_highbd_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input,
+                             uint8_t *dest, int stride, int eob, int bd);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_IDCT_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_loopfilter.c b/libvpx/libvpx/vp9/common/vp9_loopfilter.c
new file mode 100644
index 0000000..183dec4
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_loopfilter.c
@@ -0,0 +1,1697 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_seg_common.h"
+
+// 64 bit masks for left transform size. Each 1 represents a position where
+// we should apply a loop filter across the left border of an 8x8 block
+// boundary.
+//
+// In the case of TX_16X16->  ( in low order byte first we end up with
+// a mask that looks like this
+//
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//
+// A loopfilter should be applied to every other 8x8 horizontally.
+static const uint64_t left_64x64_txform_mask[TX_SIZES]= {
+  0xffffffffffffffffULL,  // TX_4X4
+  0xffffffffffffffffULL,  // TX_8x8
+  0x5555555555555555ULL,  // TX_16x16
+  0x1111111111111111ULL,  // TX_32x32
+};
+
+// 64 bit masks for above transform size. Each 1 represents a position where
+// we should apply a loop filter across the top border of an 8x8 block
+// boundary.
+//
+// In the case of TX_32x32 ->  ( in low order byte first we end up with
+// a mask that looks like this
+//
+//    11111111
+//    00000000
+//    00000000
+//    00000000
+//    11111111
+//    00000000
+//    00000000
+//    00000000
+//
+// A loopfilter should be applied to every other 4 the row vertically.
+static const uint64_t above_64x64_txform_mask[TX_SIZES]= {
+  0xffffffffffffffffULL,  // TX_4X4
+  0xffffffffffffffffULL,  // TX_8x8
+  0x00ff00ff00ff00ffULL,  // TX_16x16
+  0x000000ff000000ffULL,  // TX_32x32
+};
+
+// 64 bit masks for prediction sizes (left). Each 1 represents a position
+// where left border of an 8x8 block. These are aligned to the right most
+// appropriate bit, and then shifted into place.
+//
+// In the case of TX_16x32 ->  ( low order byte first ) we end up with
+// a mask that looks like this :
+//
+//  10000000
+//  10000000
+//  10000000
+//  10000000
+//  00000000
+//  00000000
+//  00000000
+//  00000000
+static const uint64_t left_prediction_mask[BLOCK_SIZES] = {
+  0x0000000000000001ULL,  // BLOCK_4X4,
+  0x0000000000000001ULL,  // BLOCK_4X8,
+  0x0000000000000001ULL,  // BLOCK_8X4,
+  0x0000000000000001ULL,  // BLOCK_8X8,
+  0x0000000000000101ULL,  // BLOCK_8X16,
+  0x0000000000000001ULL,  // BLOCK_16X8,
+  0x0000000000000101ULL,  // BLOCK_16X16,
+  0x0000000001010101ULL,  // BLOCK_16X32,
+  0x0000000000000101ULL,  // BLOCK_32X16,
+  0x0000000001010101ULL,  // BLOCK_32X32,
+  0x0101010101010101ULL,  // BLOCK_32X64,
+  0x0000000001010101ULL,  // BLOCK_64X32,
+  0x0101010101010101ULL,  // BLOCK_64X64
+};
+
+// 64 bit mask to shift and set for each prediction size.
+static const uint64_t above_prediction_mask[BLOCK_SIZES] = {
+  0x0000000000000001ULL,  // BLOCK_4X4
+  0x0000000000000001ULL,  // BLOCK_4X8
+  0x0000000000000001ULL,  // BLOCK_8X4
+  0x0000000000000001ULL,  // BLOCK_8X8
+  0x0000000000000001ULL,  // BLOCK_8X16,
+  0x0000000000000003ULL,  // BLOCK_16X8
+  0x0000000000000003ULL,  // BLOCK_16X16
+  0x0000000000000003ULL,  // BLOCK_16X32,
+  0x000000000000000fULL,  // BLOCK_32X16,
+  0x000000000000000fULL,  // BLOCK_32X32,
+  0x000000000000000fULL,  // BLOCK_32X64,
+  0x00000000000000ffULL,  // BLOCK_64X32,
+  0x00000000000000ffULL,  // BLOCK_64X64
+};
+// 64 bit mask to shift and set for each prediction size. A bit is set for
+// each 8x8 block that would be in the left most block of the given block
+// size in the 64x64 block.
+static const uint64_t size_mask[BLOCK_SIZES] = {
+  0x0000000000000001ULL,  // BLOCK_4X4
+  0x0000000000000001ULL,  // BLOCK_4X8
+  0x0000000000000001ULL,  // BLOCK_8X4
+  0x0000000000000001ULL,  // BLOCK_8X8
+  0x0000000000000101ULL,  // BLOCK_8X16,
+  0x0000000000000003ULL,  // BLOCK_16X8
+  0x0000000000000303ULL,  // BLOCK_16X16
+  0x0000000003030303ULL,  // BLOCK_16X32,
+  0x0000000000000f0fULL,  // BLOCK_32X16,
+  0x000000000f0f0f0fULL,  // BLOCK_32X32,
+  0x0f0f0f0f0f0f0f0fULL,  // BLOCK_32X64,
+  0x00000000ffffffffULL,  // BLOCK_64X32,
+  0xffffffffffffffffULL,  // BLOCK_64X64
+};
+
+// These are used for masking the left and above borders.
+static const uint64_t left_border =  0x1111111111111111ULL;
+static const uint64_t above_border = 0x000000ff000000ffULL;
+
+// 16 bit masks for uv transform sizes.
+static const uint16_t left_64x64_txform_mask_uv[TX_SIZES]= {
+  0xffff,  // TX_4X4
+  0xffff,  // TX_8x8
+  0x5555,  // TX_16x16
+  0x1111,  // TX_32x32
+};
+
+static const uint16_t above_64x64_txform_mask_uv[TX_SIZES]= {
+  0xffff,  // TX_4X4
+  0xffff,  // TX_8x8
+  0x0f0f,  // TX_16x16
+  0x000f,  // TX_32x32
+};
+
+// 16 bit left mask to shift and set for each uv prediction size.
+static const uint16_t left_prediction_mask_uv[BLOCK_SIZES] = {
+  0x0001,  // BLOCK_4X4,
+  0x0001,  // BLOCK_4X8,
+  0x0001,  // BLOCK_8X4,
+  0x0001,  // BLOCK_8X8,
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8,
+  0x0001,  // BLOCK_16X16,
+  0x0011,  // BLOCK_16X32,
+  0x0001,  // BLOCK_32X16,
+  0x0011,  // BLOCK_32X32,
+  0x1111,  // BLOCK_32X64
+  0x0011,  // BLOCK_64X32,
+  0x1111,  // BLOCK_64X64
+};
+// 16 bit above mask to shift and set for uv each prediction size.
+static const uint16_t above_prediction_mask_uv[BLOCK_SIZES] = {
+  0x0001,  // BLOCK_4X4
+  0x0001,  // BLOCK_4X8
+  0x0001,  // BLOCK_8X4
+  0x0001,  // BLOCK_8X8
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8
+  0x0001,  // BLOCK_16X16
+  0x0001,  // BLOCK_16X32,
+  0x0003,  // BLOCK_32X16,
+  0x0003,  // BLOCK_32X32,
+  0x0003,  // BLOCK_32X64,
+  0x000f,  // BLOCK_64X32,
+  0x000f,  // BLOCK_64X64
+};
+
+// 64 bit mask to shift and set for each uv prediction size
+static const uint16_t size_mask_uv[BLOCK_SIZES] = {
+  0x0001,  // BLOCK_4X4
+  0x0001,  // BLOCK_4X8
+  0x0001,  // BLOCK_8X4
+  0x0001,  // BLOCK_8X8
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8
+  0x0001,  // BLOCK_16X16
+  0x0011,  // BLOCK_16X32,
+  0x0003,  // BLOCK_32X16,
+  0x0033,  // BLOCK_32X32,
+  0x3333,  // BLOCK_32X64,
+  0x00ff,  // BLOCK_64X32,
+  0xffff,  // BLOCK_64X64
+};
+static const uint16_t left_border_uv =  0x1111;
+static const uint16_t above_border_uv = 0x000f;
+
+static const int mode_lf_lut[MB_MODE_COUNT] = {
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0,  // INTRA_MODES
+  1, 1, 0, 1                     // INTER_MODES (ZEROMV == 0)
+};
+
+static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
+  int lvl;
+
+  // For each possible value for the loop filter fill out limits
+  for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++) {
+    // Set loop filter parameters that control sharpness.
+    int block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));
+
+    if (sharpness_lvl > 0) {
+      if (block_inside_limit > (9 - sharpness_lvl))
+        block_inside_limit = (9 - sharpness_lvl);
+    }
+
+    if (block_inside_limit < 1)
+      block_inside_limit = 1;
+
+    memset(lfi->lfthr[lvl].lim, block_inside_limit, SIMD_WIDTH);
+    memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
+           SIMD_WIDTH);
+  }
+}
+
+static uint8_t get_filter_level(const loop_filter_info_n *lfi_n,
+                                const MODE_INFO *mi) {
+  return lfi_n->lvl[mi->segment_id][mi->ref_frame[0]]
+                   [mode_lf_lut[mi->mode]];
+}
+
+void vp9_loop_filter_init(VP9_COMMON *cm) {
+  loop_filter_info_n *lfi = &cm->lf_info;
+  struct loopfilter *lf = &cm->lf;
+  int lvl;
+
+  // init limits for given sharpness
+  update_sharpness(lfi, lf->sharpness_level);
+  lf->last_sharpness_level = lf->sharpness_level;
+
+  // init hev threshold const vectors
+  for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++)
+    memset(lfi->lfthr[lvl].hev_thr, (lvl >> 4), SIMD_WIDTH);
+}
+
+void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) {
+  int seg_id;
+  // n_shift is the multiplier for lf_deltas
+  // the multiplier is 1 for when filter_lvl is between 0 and 31;
+  // 2 when filter_lvl is between 32 and 63
+  const int scale = 1 << (default_filt_lvl >> 5);
+  loop_filter_info_n *const lfi = &cm->lf_info;
+  struct loopfilter *const lf = &cm->lf;
+  const struct segmentation *const seg = &cm->seg;
+
+  // update limits if sharpness has changed
+  if (lf->last_sharpness_level != lf->sharpness_level) {
+    update_sharpness(lfi, lf->sharpness_level);
+    lf->last_sharpness_level = lf->sharpness_level;
+  }
+
+  for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
+    int lvl_seg = default_filt_lvl;
+    if (segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
+      const int data = get_segdata(seg, seg_id, SEG_LVL_ALT_LF);
+      lvl_seg = clamp(seg->abs_delta == SEGMENT_ABSDATA ?
+                      data : default_filt_lvl + data,
+                      0, MAX_LOOP_FILTER);
+    }
+
+    if (!lf->mode_ref_delta_enabled) {
+      // we could get rid of this if we assume that deltas are set to
+      // zero when not in use; encoder always uses deltas
+      memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id]));
+    } else {
+      int ref, mode;
+      const int intra_lvl = lvl_seg + lf->ref_deltas[INTRA_FRAME] * scale;
+      lfi->lvl[seg_id][INTRA_FRAME][0] = clamp(intra_lvl, 0, MAX_LOOP_FILTER);
+
+      for (ref = LAST_FRAME; ref < MAX_REF_FRAMES; ++ref) {
+        for (mode = 0; mode < MAX_MODE_LF_DELTAS; ++mode) {
+          const int inter_lvl = lvl_seg + lf->ref_deltas[ref] * scale
+                                        + lf->mode_deltas[mode] * scale;
+          lfi->lvl[seg_id][ref][mode] = clamp(inter_lvl, 0, MAX_LOOP_FILTER);
+        }
+      }
+    }
+  }
+}
+
+static void filter_selectively_vert_row2(int subsampling_factor,
+                                         uint8_t *s, int pitch,
+                                         unsigned int mask_16x16,
+                                         unsigned int mask_8x8,
+                                         unsigned int mask_4x4,
+                                         unsigned int mask_4x4_int,
+                                         const loop_filter_thresh *lfthr,
+                                         const uint8_t *lfl) {
+  const int dual_mask_cutoff = subsampling_factor ? 0xff : 0xffff;
+  const int lfl_forward = subsampling_factor ? 4 : 8;
+  const unsigned int dual_one = 1 | (1 << lfl_forward);
+  unsigned int mask;
+  uint8_t *ss[2];
+  ss[0] = s;
+
+  for (mask =
+           (mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int) & dual_mask_cutoff;
+       mask; mask = (mask & ~dual_one) >> 1) {
+    if (mask & dual_one) {
+      const loop_filter_thresh *lfis[2];
+      lfis[0] = lfthr + *lfl;
+      lfis[1] = lfthr + *(lfl + lfl_forward);
+      ss[1] = ss[0] + 8 * pitch;
+
+      if (mask_16x16 & dual_one) {
+        if ((mask_16x16 & dual_one) == dual_one) {
+          vpx_lpf_vertical_16_dual(ss[0], pitch, lfis[0]->mblim, lfis[0]->lim,
+                                   lfis[0]->hev_thr);
+        } else {
+          const loop_filter_thresh *lfi = lfis[!(mask_16x16 & 1)];
+          vpx_lpf_vertical_16(ss[!(mask_16x16 & 1)], pitch, lfi->mblim,
+                              lfi->lim, lfi->hev_thr);
+        }
+      }
+
+      if (mask_8x8 & dual_one) {
+        if ((mask_8x8 & dual_one) == dual_one) {
+          vpx_lpf_vertical_8_dual(ss[0], pitch, lfis[0]->mblim, lfis[0]->lim,
+                                  lfis[0]->hev_thr, lfis[1]->mblim,
+                                  lfis[1]->lim, lfis[1]->hev_thr);
+        } else {
+          const loop_filter_thresh *lfi = lfis[!(mask_8x8 & 1)];
+          vpx_lpf_vertical_8(ss[!(mask_8x8 & 1)], pitch, lfi->mblim, lfi->lim,
+                             lfi->hev_thr);
+        }
+      }
+
+      if (mask_4x4 & dual_one) {
+        if ((mask_4x4 & dual_one) == dual_one) {
+          vpx_lpf_vertical_4_dual(ss[0], pitch, lfis[0]->mblim, lfis[0]->lim,
+                                  lfis[0]->hev_thr, lfis[1]->mblim,
+                                  lfis[1]->lim, lfis[1]->hev_thr);
+        } else {
+          const loop_filter_thresh *lfi = lfis[!(mask_4x4 & 1)];
+          vpx_lpf_vertical_4(ss[!(mask_4x4 & 1)], pitch, lfi->mblim, lfi->lim,
+                             lfi->hev_thr);
+        }
+      }
+
+      if (mask_4x4_int & dual_one) {
+        if ((mask_4x4_int & dual_one) == dual_one) {
+          vpx_lpf_vertical_4_dual(ss[0] + 4, pitch, lfis[0]->mblim,
+                                  lfis[0]->lim, lfis[0]->hev_thr,
+                                  lfis[1]->mblim, lfis[1]->lim,
+                                  lfis[1]->hev_thr);
+        } else {
+          const loop_filter_thresh *lfi = lfis[!(mask_4x4_int & 1)];
+          vpx_lpf_vertical_4(ss[!(mask_4x4_int & 1)] + 4, pitch, lfi->mblim,
+                             lfi->lim, lfi->hev_thr);
+        }
+      }
+    }
+
+    ss[0] += 8;
+    lfl += 1;
+    mask_16x16 >>= 1;
+    mask_8x8 >>= 1;
+    mask_4x4 >>= 1;
+    mask_4x4_int >>= 1;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_filter_selectively_vert_row2(int subsampling_factor,
+                                                uint16_t *s, int pitch,
+                                                unsigned int mask_16x16,
+                                                unsigned int mask_8x8,
+                                                unsigned int mask_4x4,
+                                                unsigned int mask_4x4_int,
+                                                const loop_filter_thresh *lfthr,
+                                                const uint8_t *lfl, int bd) {
+  const int dual_mask_cutoff = subsampling_factor ? 0xff : 0xffff;
+  const int lfl_forward = subsampling_factor ? 4 : 8;
+  const unsigned int dual_one = 1 | (1 << lfl_forward);
+  unsigned int mask;
+  uint16_t *ss[2];
+  ss[0] = s;
+
+  for (mask =
+           (mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int) & dual_mask_cutoff;
+       mask; mask = (mask & ~dual_one) >> 1) {
+    if (mask & dual_one) {
+      const loop_filter_thresh *lfis[2];
+      lfis[0] = lfthr + *lfl;
+      lfis[1] = lfthr + *(lfl + lfl_forward);
+      ss[1] = ss[0] + 8 * pitch;
+
+      if (mask_16x16 & dual_one) {
+        if ((mask_16x16 & dual_one) == dual_one) {
+          vpx_highbd_lpf_vertical_16_dual(ss[0], pitch, lfis[0]->mblim,
+                                          lfis[0]->lim, lfis[0]->hev_thr, bd);
+        } else {
+          const loop_filter_thresh *lfi = lfis[!(mask_16x16 & 1)];
+          vpx_highbd_lpf_vertical_16(ss[!(mask_16x16 & 1)], pitch, lfi->mblim,
+                                     lfi->lim, lfi->hev_thr, bd);
+        }
+      }
+
+      if (mask_8x8 & dual_one) {
+        if ((mask_8x8 & dual_one) == dual_one) {
+          vpx_highbd_lpf_vertical_8_dual(ss[0], pitch, lfis[0]->mblim,
+                                         lfis[0]->lim, lfis[0]->hev_thr,
+                                         lfis[1]->mblim, lfis[1]->lim,
+                                         lfis[1]->hev_thr, bd);
+        } else {
+          const loop_filter_thresh *lfi = lfis[!(mask_8x8 & 1)];
+          vpx_highbd_lpf_vertical_8(ss[!(mask_8x8 & 1)], pitch, lfi->mblim,
+                                    lfi->lim, lfi->hev_thr, bd);
+        }
+      }
+
+      if (mask_4x4 & dual_one) {
+        if ((mask_4x4 & dual_one) == dual_one) {
+          vpx_highbd_lpf_vertical_4_dual(ss[0], pitch, lfis[0]->mblim,
+                                         lfis[0]->lim, lfis[0]->hev_thr,
+                                         lfis[1]->mblim, lfis[1]->lim,
+                                         lfis[1]->hev_thr, bd);
+        } else {
+          const loop_filter_thresh *lfi = lfis[!(mask_4x4 & 1)];
+          vpx_highbd_lpf_vertical_4(ss[!(mask_4x4 & 1)], pitch, lfi->mblim,
+                                    lfi->lim, lfi->hev_thr, bd);
+        }
+      }
+
+      if (mask_4x4_int & dual_one) {
+        if ((mask_4x4_int & dual_one) == dual_one) {
+          vpx_highbd_lpf_vertical_4_dual(ss[0] + 4, pitch, lfis[0]->mblim,
+                                         lfis[0]->lim, lfis[0]->hev_thr,
+                                         lfis[1]->mblim, lfis[1]->lim,
+                                         lfis[1]->hev_thr, bd);
+        } else {
+          const loop_filter_thresh *lfi = lfis[!(mask_4x4_int & 1)];
+          vpx_highbd_lpf_vertical_4(ss[!(mask_4x4_int & 1)] + 4, pitch,
+                                    lfi->mblim, lfi->lim, lfi->hev_thr, bd);
+        }
+      }
+    }
+
+    ss[0] += 8;
+    lfl += 1;
+    mask_16x16 >>= 1;
+    mask_8x8 >>= 1;
+    mask_4x4 >>= 1;
+    mask_4x4_int >>= 1;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void filter_selectively_horiz(uint8_t *s, int pitch,
+                                     unsigned int mask_16x16,
+                                     unsigned int mask_8x8,
+                                     unsigned int mask_4x4,
+                                     unsigned int mask_4x4_int,
+                                     const loop_filter_thresh *lfthr,
+                                     const uint8_t *lfl) {
+  unsigned int mask;
+  int count;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
+       mask; mask >>= count) {
+    count = 1;
+    if (mask & 1) {
+      const loop_filter_thresh *lfi = lfthr + *lfl;
+
+      if (mask_16x16 & 1) {
+        if ((mask_16x16 & 3) == 3) {
+          vpx_lpf_horizontal_edge_16(s, pitch, lfi->mblim, lfi->lim,
+                                     lfi->hev_thr);
+          count = 2;
+        } else {
+          vpx_lpf_horizontal_edge_8(s, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr);
+        }
+      } else if (mask_8x8 & 1) {
+        if ((mask_8x8 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfthr + *(lfl + 1);
+
+          vpx_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr, lfin->mblim, lfin->lim,
+                                    lfin->hev_thr);
+
+          if ((mask_4x4_int & 3) == 3) {
+            vpx_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
+                                      lfi->lim, lfi->hev_thr, lfin->mblim,
+                                      lfin->lim, lfin->hev_thr);
+          } else {
+            if (mask_4x4_int & 1)
+              vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                   lfi->hev_thr);
+            else if (mask_4x4_int & 2)
+              vpx_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                   lfin->lim, lfin->hev_thr);
+          }
+          count = 2;
+        } else {
+          vpx_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+
+          if (mask_4x4_int & 1)
+            vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                 lfi->hev_thr);
+        }
+      } else if (mask_4x4 & 1) {
+        if ((mask_4x4 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfthr + *(lfl + 1);
+
+          vpx_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr, lfin->mblim, lfin->lim,
+                                    lfin->hev_thr);
+          if ((mask_4x4_int & 3) == 3) {
+            vpx_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
+                                      lfi->lim, lfi->hev_thr, lfin->mblim,
+                                      lfin->lim, lfin->hev_thr);
+          } else {
+            if (mask_4x4_int & 1)
+              vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                   lfi->hev_thr);
+            else if (mask_4x4_int & 2)
+              vpx_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                   lfin->lim, lfin->hev_thr);
+          }
+          count = 2;
+        } else {
+          vpx_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+
+          if (mask_4x4_int & 1)
+            vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                 lfi->hev_thr);
+        }
+      } else {
+        vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                             lfi->hev_thr);
+      }
+    }
+    s += 8 * count;
+    lfl += count;
+    mask_16x16 >>= count;
+    mask_8x8 >>= count;
+    mask_4x4 >>= count;
+    mask_4x4_int >>= count;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_filter_selectively_horiz(uint16_t *s, int pitch,
+                                            unsigned int mask_16x16,
+                                            unsigned int mask_8x8,
+                                            unsigned int mask_4x4,
+                                            unsigned int mask_4x4_int,
+                                            const loop_filter_thresh *lfthr,
+                                            const uint8_t *lfl, int bd) {
+  unsigned int mask;
+  int count;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
+       mask; mask >>= count) {
+    count = 1;
+    if (mask & 1) {
+      const loop_filter_thresh *lfi = lfthr + *lfl;
+
+      if (mask_16x16 & 1) {
+        if ((mask_16x16 & 3) == 3) {
+          vpx_highbd_lpf_horizontal_edge_16(s, pitch, lfi->mblim, lfi->lim,
+                                            lfi->hev_thr, bd);
+          count = 2;
+        } else {
+          vpx_highbd_lpf_horizontal_edge_8(s, pitch, lfi->mblim, lfi->lim,
+                                           lfi->hev_thr, bd);
+        }
+      } else if (mask_8x8 & 1) {
+        if ((mask_8x8 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfthr + *(lfl + 1);
+
+          vpx_highbd_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
+                                           lfi->hev_thr, lfin->mblim, lfin->lim,
+                                           lfin->hev_thr, bd);
+
+          if ((mask_4x4_int & 3) == 3) {
+            vpx_highbd_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
+                                             lfi->lim, lfi->hev_thr,
+                                             lfin->mblim, lfin->lim,
+                                             lfin->hev_thr, bd);
+          } else {
+            if (mask_4x4_int & 1) {
+              vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                          lfi->lim, lfi->hev_thr, bd);
+            } else if (mask_4x4_int & 2) {
+              vpx_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                          lfin->lim, lfin->hev_thr, bd);
+            }
+          }
+          count = 2;
+        } else {
+          vpx_highbd_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim,
+                                      lfi->hev_thr, bd);
+
+          if (mask_4x4_int & 1) {
+            vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                        lfi->lim, lfi->hev_thr, bd);
+          }
+        }
+      } else if (mask_4x4 & 1) {
+        if ((mask_4x4 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfthr + *(lfl + 1);
+
+          vpx_highbd_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
+                                           lfi->hev_thr, lfin->mblim, lfin->lim,
+                                           lfin->hev_thr, bd);
+          if ((mask_4x4_int & 3) == 3) {
+            vpx_highbd_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
+                                             lfi->lim, lfi->hev_thr,
+                                             lfin->mblim, lfin->lim,
+                                             lfin->hev_thr, bd);
+          } else {
+            if (mask_4x4_int & 1) {
+              vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                          lfi->lim, lfi->hev_thr, bd);
+            } else if (mask_4x4_int & 2) {
+              vpx_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                          lfin->lim, lfin->hev_thr, bd);
+            }
+          }
+          count = 2;
+        } else {
+          vpx_highbd_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim,
+                                      lfi->hev_thr, bd);
+
+          if (mask_4x4_int & 1) {
+            vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                        lfi->lim, lfi->hev_thr, bd);
+          }
+        }
+      } else {
+        vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr, bd);
+      }
+    }
+    s += 8 * count;
+    lfl += count;
+    mask_16x16 >>= count;
+    mask_8x8 >>= count;
+    mask_4x4 >>= count;
+    mask_4x4_int >>= count;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+// This function ors into the current lfm structure, where to do loop
+// filters for the specific mi we are looking at. It uses information
+// including the block_size_type (32x16, 32x32, etc.), the transform size,
+// whether there were any coefficients encoded, and the loop filter strength
+// block we are currently looking at. Shift is used to position the
+// 1's we produce.
+static void build_masks(const loop_filter_info_n *const lfi_n,
+                        const MODE_INFO *mi, const int shift_y,
+                        const int shift_uv,
+                        LOOP_FILTER_MASK *lfm) {
+  const BLOCK_SIZE block_size = mi->sb_type;
+  const TX_SIZE tx_size_y = mi->tx_size;
+  const TX_SIZE tx_size_uv = get_uv_tx_size_impl(tx_size_y, block_size, 1, 1);
+  const int filter_level = get_filter_level(lfi_n, mi);
+  uint64_t *const left_y = &lfm->left_y[tx_size_y];
+  uint64_t *const above_y = &lfm->above_y[tx_size_y];
+  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
+  uint16_t *const left_uv = &lfm->left_uv[tx_size_uv];
+  uint16_t *const above_uv = &lfm->above_uv[tx_size_uv];
+  uint16_t *const int_4x4_uv = &lfm->int_4x4_uv;
+  int i;
+
+  // If filter level is 0 we don't loop filter.
+  if (!filter_level) {
+    return;
+  } else {
+    const int w = num_8x8_blocks_wide_lookup[block_size];
+    const int h = num_8x8_blocks_high_lookup[block_size];
+    int index = shift_y;
+    for (i = 0; i < h; i++) {
+      memset(&lfm->lfl_y[index], filter_level, w);
+      index += 8;
+    }
+  }
+
+  // These set 1 in the current block size for the block size edges.
+  // For instance if the block size is 32x16, we'll set:
+  //    above =   1111
+  //              0000
+  //    and
+  //    left  =   1000
+  //          =   1000
+  // NOTE : In this example the low bit is left most ( 1000 ) is stored as
+  //        1,  not 8...
+  //
+  // U and V set things on a 16 bit scale.
+  //
+  *above_y |= above_prediction_mask[block_size] << shift_y;
+  *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
+  *left_y |= left_prediction_mask[block_size] << shift_y;
+  *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;
+
+  // If the block has no coefficients and is not intra we skip applying
+  // the loop filter on block edges.
+  if (mi->skip && is_inter_block(mi))
+    return;
+
+  // Here we are adding a mask for the transform size. The transform
+  // size mask is set to be correct for a 64x64 prediction block size. We
+  // mask to match the size of the block we are working on and then shift it
+  // into place..
+  *above_y |= (size_mask[block_size] &
+               above_64x64_txform_mask[tx_size_y]) << shift_y;
+  *above_uv |= (size_mask_uv[block_size] &
+                above_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
+
+  *left_y |= (size_mask[block_size] &
+              left_64x64_txform_mask[tx_size_y]) << shift_y;
+  *left_uv |= (size_mask_uv[block_size] &
+               left_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
+
+  // Here we are trying to determine what to do with the internal 4x4 block
+  // boundaries.  These differ from the 4x4 boundaries on the outside edge of
+  // an 8x8 in that the internal ones can be skipped and don't depend on
+  // the prediction block size.
+  if (tx_size_y == TX_4X4)
+    *int_4x4_y |= size_mask[block_size] << shift_y;
+
+  if (tx_size_uv == TX_4X4)
+    *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
+}
+
+// This function does the same thing as the one above with the exception that
+// it only affects the y masks. It exists because for blocks < 16x16 in size,
+// we only update u and v masks on the first block.
+static void build_y_mask(const loop_filter_info_n *const lfi_n,
+                         const MODE_INFO *mi, const int shift_y,
+                         LOOP_FILTER_MASK *lfm) {
+  const BLOCK_SIZE block_size = mi->sb_type;
+  const TX_SIZE tx_size_y = mi->tx_size;
+  const int filter_level = get_filter_level(lfi_n, mi);
+  uint64_t *const left_y = &lfm->left_y[tx_size_y];
+  uint64_t *const above_y = &lfm->above_y[tx_size_y];
+  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
+  int i;
+
+  if (!filter_level) {
+    return;
+  } else {
+    const int w = num_8x8_blocks_wide_lookup[block_size];
+    const int h = num_8x8_blocks_high_lookup[block_size];
+    int index = shift_y;
+    for (i = 0; i < h; i++) {
+      memset(&lfm->lfl_y[index], filter_level, w);
+      index += 8;
+    }
+  }
+
+  *above_y |= above_prediction_mask[block_size] << shift_y;
+  *left_y |= left_prediction_mask[block_size] << shift_y;
+
+  if (mi->skip && is_inter_block(mi))
+    return;
+
+  *above_y |= (size_mask[block_size] &
+               above_64x64_txform_mask[tx_size_y]) << shift_y;
+
+  *left_y |= (size_mask[block_size] &
+              left_64x64_txform_mask[tx_size_y]) << shift_y;
+
+  if (tx_size_y == TX_4X4)
+    *int_4x4_y |= size_mask[block_size] << shift_y;
+}
+
+void vp9_adjust_mask(VP9_COMMON *const cm, const int mi_row,
+                     const int mi_col, LOOP_FILTER_MASK *lfm) {
+  int i;
+
+  // The largest loopfilter we have is 16x16 so we use the 16x16 mask
+  // for 32x32 transforms also.
+  lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32];
+  lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32];
+  lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32];
+  lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32];
+
+  // We do at least 8 tap filter on every 32x32 even if the transform size
+  // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and
+  // remove it from the 4x4.
+  lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border;
+  lfm->left_y[TX_4X4] &= ~left_border;
+  lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border;
+  lfm->above_y[TX_4X4] &= ~above_border;
+  lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv;
+  lfm->left_uv[TX_4X4] &= ~left_border_uv;
+  lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv;
+  lfm->above_uv[TX_4X4] &= ~above_border_uv;
+
+  // We do some special edge handling.
+  if (mi_row + MI_BLOCK_SIZE > cm->mi_rows) {
+    const uint64_t rows = cm->mi_rows - mi_row;
+
+    // Each pixel inside the border gets a 1,
+    const uint64_t mask_y = (((uint64_t) 1 << (rows << 3)) - 1);
+    const uint16_t mask_uv = (((uint16_t) 1 << (((rows + 1) >> 1) << 2)) - 1);
+
+    // Remove values completely outside our border.
+    for (i = 0; i < TX_32X32; i++) {
+      lfm->left_y[i] &= mask_y;
+      lfm->above_y[i] &= mask_y;
+      lfm->left_uv[i] &= mask_uv;
+      lfm->above_uv[i] &= mask_uv;
+    }
+    lfm->int_4x4_y &= mask_y;
+    lfm->int_4x4_uv &= mask_uv;
+
+    // We don't apply a wide loop filter on the last uv block row. If set
+    // apply the shorter one instead.
+    if (rows == 1) {
+      lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16];
+      lfm->above_uv[TX_16X16] = 0;
+    }
+    if (rows == 5) {
+      lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00;
+      lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00);
+    }
+  }
+
+  if (mi_col + MI_BLOCK_SIZE > cm->mi_cols) {
+    const uint64_t columns = cm->mi_cols - mi_col;
+
+    // Each pixel inside the border gets a 1, the multiply copies the border
+    // to where we need it.
+    const uint64_t mask_y  = (((1 << columns) - 1)) * 0x0101010101010101ULL;
+    const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111;
+
+    // Internal edges are not applied on the last column of the image so
+    // we mask 1 more for the internal edges
+    const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111;
+
+    // Remove the bits outside the image edge.
+    for (i = 0; i < TX_32X32; i++) {
+      lfm->left_y[i] &= mask_y;
+      lfm->above_y[i] &= mask_y;
+      lfm->left_uv[i] &= mask_uv;
+      lfm->above_uv[i] &= mask_uv;
+    }
+    lfm->int_4x4_y &= mask_y;
+    lfm->int_4x4_uv &= mask_uv_int;
+
+    // We don't apply a wide loop filter on the last uv column. If set
+    // apply the shorter one instead.
+    if (columns == 1) {
+      lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16];
+      lfm->left_uv[TX_16X16] = 0;
+    }
+    if (columns == 5) {
+      lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc);
+      lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc);
+    }
+  }
+  // We don't apply a loop filter on the first column in the image, mask that
+  // out.
+  if (mi_col == 0) {
+    for (i = 0; i < TX_32X32; i++) {
+      lfm->left_y[i] &= 0xfefefefefefefefeULL;
+      lfm->left_uv[i] &= 0xeeee;
+    }
+  }
+
+  // Assert if we try to apply 2 different loop filters at the same position.
+  assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_8X8]));
+  assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_4X4]));
+  assert(!(lfm->left_y[TX_8X8] & lfm->left_y[TX_4X4]));
+  assert(!(lfm->int_4x4_y & lfm->left_y[TX_16X16]));
+  assert(!(lfm->left_uv[TX_16X16]&lfm->left_uv[TX_8X8]));
+  assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_4X4]));
+  assert(!(lfm->left_uv[TX_8X8] & lfm->left_uv[TX_4X4]));
+  assert(!(lfm->int_4x4_uv & lfm->left_uv[TX_16X16]));
+  assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_8X8]));
+  assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_4X4]));
+  assert(!(lfm->above_y[TX_8X8] & lfm->above_y[TX_4X4]));
+  assert(!(lfm->int_4x4_y & lfm->above_y[TX_16X16]));
+  assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_8X8]));
+  assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_4X4]));
+  assert(!(lfm->above_uv[TX_8X8] & lfm->above_uv[TX_4X4]));
+  assert(!(lfm->int_4x4_uv & lfm->above_uv[TX_16X16]));
+}
+
+// This function sets up the bit masks for the entire 64x64 region represented
+// by mi_row, mi_col.
+void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
+                    MODE_INFO **mi, const int mode_info_stride,
+                    LOOP_FILTER_MASK *lfm) {
+  int idx_32, idx_16, idx_8;
+  const loop_filter_info_n *const lfi_n = &cm->lf_info;
+  MODE_INFO **mip = mi;
+  MODE_INFO **mip2 = mi;
+
+  // These are offsets to the next mi in the 64x64 block. It is what gets
+  // added to the mi ptr as we go through each loop. It helps us to avoid
+  // setting up special row and column counters for each index. The last step
+  // brings us out back to the starting position.
+  const int offset_32[] = {4, (mode_info_stride << 2) - 4, 4,
+                           -(mode_info_stride << 2) - 4};
+  const int offset_16[] = {2, (mode_info_stride << 1) - 2, 2,
+                           -(mode_info_stride << 1) - 2};
+  const int offset[] = {1, mode_info_stride - 1, 1, -mode_info_stride - 1};
+
+  // Following variables represent shifts to position the current block
+  // mask over the appropriate block. A shift of 36 to the left will move
+  // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left
+  // 4 rows to the appropriate spot.
+  const int shift_32_y[] = {0, 4, 32, 36};
+  const int shift_16_y[] = {0, 2, 16, 18};
+  const int shift_8_y[] = {0, 1, 8, 9};
+  const int shift_32_uv[] = {0, 2, 8, 10};
+  const int shift_16_uv[] = {0, 1, 4, 5};
+  const int max_rows = (mi_row + MI_BLOCK_SIZE > cm->mi_rows ?
+                        cm->mi_rows - mi_row : MI_BLOCK_SIZE);
+  const int max_cols = (mi_col + MI_BLOCK_SIZE > cm->mi_cols ?
+                        cm->mi_cols - mi_col : MI_BLOCK_SIZE);
+
+  vp9_zero(*lfm);
+  assert(mip[0] != NULL);
+
+  switch (mip[0]->sb_type) {
+    case BLOCK_64X64:
+      build_masks(lfi_n, mip[0] , 0, 0, lfm);
+      break;
+    case BLOCK_64X32:
+      build_masks(lfi_n, mip[0], 0, 0, lfm);
+      mip2 = mip + mode_info_stride * 4;
+      if (4 >= max_rows)
+        break;
+      build_masks(lfi_n, mip2[0], 32, 8, lfm);
+      break;
+    case BLOCK_32X64:
+      build_masks(lfi_n, mip[0], 0, 0, lfm);
+      mip2 = mip + 4;
+      if (4 >= max_cols)
+        break;
+      build_masks(lfi_n, mip2[0], 4, 2, lfm);
+      break;
+    default:
+      for (idx_32 = 0; idx_32 < 4; mip += offset_32[idx_32], ++idx_32) {
+        const int shift_y = shift_32_y[idx_32];
+        const int shift_uv = shift_32_uv[idx_32];
+        const int mi_32_col_offset = ((idx_32 & 1) << 2);
+        const int mi_32_row_offset = ((idx_32 >> 1) << 2);
+        if (mi_32_col_offset >= max_cols || mi_32_row_offset >= max_rows)
+          continue;
+        switch (mip[0]->sb_type) {
+          case BLOCK_32X32:
+            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+            break;
+          case BLOCK_32X16:
+            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+            if (mi_32_row_offset + 2 >= max_rows)
+              continue;
+            mip2 = mip + mode_info_stride * 2;
+            build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm);
+            break;
+          case BLOCK_16X32:
+            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+            if (mi_32_col_offset + 2 >= max_cols)
+              continue;
+            mip2 = mip + 2;
+            build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm);
+            break;
+          default:
+            for (idx_16 = 0; idx_16 < 4; mip += offset_16[idx_16], ++idx_16) {
+              const int shift_y = shift_32_y[idx_32] + shift_16_y[idx_16];
+              const int shift_uv = shift_32_uv[idx_32] + shift_16_uv[idx_16];
+              const int mi_16_col_offset = mi_32_col_offset +
+                  ((idx_16 & 1) << 1);
+              const int mi_16_row_offset = mi_32_row_offset +
+                  ((idx_16 >> 1) << 1);
+
+              if (mi_16_col_offset >= max_cols || mi_16_row_offset >= max_rows)
+                continue;
+
+              switch (mip[0]->sb_type) {
+                case BLOCK_16X16:
+                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+                  break;
+                case BLOCK_16X8:
+                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+                  if (mi_16_row_offset + 1 >= max_rows)
+                    continue;
+                  mip2 = mip + mode_info_stride;
+                  build_y_mask(lfi_n, mip2[0], shift_y+8, lfm);
+                  break;
+                case BLOCK_8X16:
+                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+                  if (mi_16_col_offset +1 >= max_cols)
+                    continue;
+                  mip2 = mip + 1;
+                  build_y_mask(lfi_n, mip2[0], shift_y+1, lfm);
+                  break;
+                default: {
+                  const int shift_y = shift_32_y[idx_32] +
+                                      shift_16_y[idx_16] +
+                                      shift_8_y[0];
+                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+                  mip += offset[0];
+                  for (idx_8 = 1; idx_8 < 4; mip += offset[idx_8], ++idx_8) {
+                    const int shift_y = shift_32_y[idx_32] +
+                                        shift_16_y[idx_16] +
+                                        shift_8_y[idx_8];
+                    const int mi_8_col_offset = mi_16_col_offset +
+                        ((idx_8 & 1));
+                    const int mi_8_row_offset = mi_16_row_offset +
+                        ((idx_8 >> 1));
+
+                    if (mi_8_col_offset >= max_cols ||
+                        mi_8_row_offset >= max_rows)
+                      continue;
+                    build_y_mask(lfi_n, mip[0], shift_y, lfm);
+                  }
+                  break;
+                }
+              }
+            }
+            break;
+        }
+      }
+      break;
+  }
+}
+
+static void filter_selectively_vert(uint8_t *s, int pitch,
+                                    unsigned int mask_16x16,
+                                    unsigned int mask_8x8,
+                                    unsigned int mask_4x4,
+                                    unsigned int mask_4x4_int,
+                                    const loop_filter_thresh *lfthr,
+                                    const uint8_t *lfl) {
+  unsigned int mask;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
+       mask; mask >>= 1) {
+    const loop_filter_thresh *lfi = lfthr + *lfl;
+
+    if (mask & 1) {
+      if (mask_16x16 & 1) {
+        vpx_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+      } else if (mask_8x8 & 1) {
+        vpx_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+      } else if (mask_4x4 & 1) {
+        vpx_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+      }
+    }
+    if (mask_4x4_int & 1)
+      vpx_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+    s += 8;
+    lfl += 1;
+    mask_16x16 >>= 1;
+    mask_8x8 >>= 1;
+    mask_4x4 >>= 1;
+    mask_4x4_int >>= 1;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_filter_selectively_vert(uint16_t *s, int pitch,
+                                           unsigned int mask_16x16,
+                                           unsigned int mask_8x8,
+                                           unsigned int mask_4x4,
+                                           unsigned int mask_4x4_int,
+                                           const loop_filter_thresh *lfthr,
+                                           const uint8_t *lfl, int bd) {
+  unsigned int mask;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
+       mask; mask >>= 1) {
+    const loop_filter_thresh *lfi = lfthr + *lfl;
+
+    if (mask & 1) {
+      if (mask_16x16 & 1) {
+        vpx_highbd_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim,
+                                   lfi->hev_thr, bd);
+      } else if (mask_8x8 & 1) {
+        vpx_highbd_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim,
+                                  lfi->hev_thr, bd);
+      } else if (mask_4x4 & 1) {
+        vpx_highbd_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim,
+                                  lfi->hev_thr, bd);
+      }
+    }
+    if (mask_4x4_int & 1)
+      vpx_highbd_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim,
+                                lfi->hev_thr, bd);
+    s += 8;
+    lfl += 1;
+    mask_16x16 >>= 1;
+    mask_8x8 >>= 1;
+    mask_4x4 >>= 1;
+    mask_4x4_int >>= 1;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_filter_block_plane_non420(VP9_COMMON *cm,
+                                   struct macroblockd_plane *plane,
+                                   MODE_INFO **mi_8x8,
+                                   int mi_row, int mi_col) {
+  const int ss_x = plane->subsampling_x;
+  const int ss_y = plane->subsampling_y;
+  const int row_step = 1 << ss_y;
+  const int col_step = 1 << ss_x;
+  const int row_step_stride = cm->mi_stride * row_step;
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t* const dst0 = dst->buf;
+  unsigned int mask_16x16[MI_BLOCK_SIZE] = {0};
+  unsigned int mask_8x8[MI_BLOCK_SIZE] = {0};
+  unsigned int mask_4x4[MI_BLOCK_SIZE] = {0};
+  unsigned int mask_4x4_int[MI_BLOCK_SIZE] = {0};
+  uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE];
+  int r, c;
+
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
+    unsigned int mask_16x16_c = 0;
+    unsigned int mask_8x8_c = 0;
+    unsigned int mask_4x4_c = 0;
+    unsigned int border_mask;
+
+    // Determine the vertical edges that need filtering
+    for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
+      const MODE_INFO *mi = mi_8x8[c];
+      const BLOCK_SIZE sb_type = mi[0].sb_type;
+      const int skip_this = mi[0].skip && is_inter_block(mi);
+      // left edge of current unit is block/partition edge -> no skip
+      const int block_edge_left = (num_4x4_blocks_wide_lookup[sb_type] > 1) ?
+          !(c & (num_8x8_blocks_wide_lookup[sb_type] - 1)) : 1;
+      const int skip_this_c = skip_this && !block_edge_left;
+      // top edge of current unit is block/partition edge -> no skip
+      const int block_edge_above = (num_4x4_blocks_high_lookup[sb_type] > 1) ?
+          !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
+      const int skip_this_r = skip_this && !block_edge_above;
+      const TX_SIZE tx_size = get_uv_tx_size(mi, plane);
+      const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
+      const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
+
+      // Filter level can vary per MI
+      if (!(lfl[(r << 3) + (c >> ss_x)] =
+            get_filter_level(&cm->lf_info, mi)))
+        continue;
+
+      // Build masks based on the transform size of each block
+      if (tx_size == TX_32X32) {
+        if (!skip_this_c && ((c >> ss_x) & 3) == 0) {
+          if (!skip_border_4x4_c)
+            mask_16x16_c |= 1 << (c >> ss_x);
+          else
+            mask_8x8_c |= 1 << (c >> ss_x);
+        }
+        if (!skip_this_r && ((r >> ss_y) & 3) == 0) {
+          if (!skip_border_4x4_r)
+            mask_16x16[r] |= 1 << (c >> ss_x);
+          else
+            mask_8x8[r] |= 1 << (c >> ss_x);
+        }
+      } else if (tx_size == TX_16X16) {
+        if (!skip_this_c && ((c >> ss_x) & 1) == 0) {
+          if (!skip_border_4x4_c)
+            mask_16x16_c |= 1 << (c >> ss_x);
+          else
+            mask_8x8_c |= 1 << (c >> ss_x);
+        }
+        if (!skip_this_r && ((r >> ss_y) & 1) == 0) {
+          if (!skip_border_4x4_r)
+            mask_16x16[r] |= 1 << (c >> ss_x);
+          else
+            mask_8x8[r] |= 1 << (c >> ss_x);
+        }
+      } else {
+        // force 8x8 filtering on 32x32 boundaries
+        if (!skip_this_c) {
+          if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
+            mask_8x8_c |= 1 << (c >> ss_x);
+          else
+            mask_4x4_c |= 1 << (c >> ss_x);
+        }
+
+        if (!skip_this_r) {
+          if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
+            mask_8x8[r] |= 1 << (c >> ss_x);
+          else
+            mask_4x4[r] |= 1 << (c >> ss_x);
+        }
+
+        if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
+          mask_4x4_int[r] |= 1 << (c >> ss_x);
+      }
+    }
+
+    // Disable filtering on the leftmost column
+    border_mask = ~(mi_col == 0);
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_vert(CONVERT_TO_SHORTPTR(dst->buf),
+                                     dst->stride,
+                                     mask_16x16_c & border_mask,
+                                     mask_8x8_c & border_mask,
+                                     mask_4x4_c & border_mask,
+                                     mask_4x4_int[r],
+                                     cm->lf_info.lfthr, &lfl[r << 3],
+                                     (int)cm->bit_depth);
+    } else {
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      filter_selectively_vert(dst->buf, dst->stride,
+                              mask_16x16_c & border_mask,
+                              mask_8x8_c & border_mask,
+                              mask_4x4_c & border_mask,
+                              mask_4x4_int[r],
+                              cm->lf_info.lfthr, &lfl[r << 3]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    dst->buf += 8 * dst->stride;
+    mi_8x8 += row_step_stride;
+  }
+
+  // Now do horizontal pass
+  dst->buf = dst0;
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
+    const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
+    const unsigned int mask_4x4_int_r = skip_border_4x4_r ? 0 : mask_4x4_int[r];
+
+    unsigned int mask_16x16_r;
+    unsigned int mask_8x8_r;
+    unsigned int mask_4x4_r;
+
+    if (mi_row + r == 0) {
+      mask_16x16_r = 0;
+      mask_8x8_r = 0;
+      mask_4x4_r = 0;
+    } else {
+      mask_16x16_r = mask_16x16[r];
+      mask_8x8_r = mask_8x8[r];
+      mask_4x4_r = mask_4x4[r];
+    }
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
+                                      dst->stride,
+                                      mask_16x16_r,
+                                      mask_8x8_r,
+                                      mask_4x4_r,
+                                      mask_4x4_int_r,
+                                      cm->lf_info.lfthr, &lfl[r << 3],
+                                      (int)cm->bit_depth);
+    } else {
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      filter_selectively_horiz(dst->buf, dst->stride,
+                               mask_16x16_r,
+                               mask_8x8_r,
+                               mask_4x4_r,
+                               mask_4x4_int_r,
+                               cm->lf_info.lfthr, &lfl[r << 3]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    dst->buf += 8 * dst->stride;
+  }
+}
+
+void vp9_filter_block_plane_ss00(VP9_COMMON *const cm,
+                                 struct macroblockd_plane *const plane,
+                                 int mi_row,
+                                 LOOP_FILTER_MASK *lfm) {
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t *const dst0 = dst->buf;
+  int r;
+  uint64_t mask_16x16 = lfm->left_y[TX_16X16];
+  uint64_t mask_8x8 = lfm->left_y[TX_8X8];
+  uint64_t mask_4x4 = lfm->left_y[TX_4X4];
+  uint64_t mask_4x4_int = lfm->int_4x4_y;
+
+  assert(plane->subsampling_x == 0 && plane->subsampling_y == 0);
+
+  // Vertical pass: do 2 rows at one time
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
+    // Disable filtering on the leftmost column.
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_vert_row2(plane->subsampling_x,
+                                          CONVERT_TO_SHORTPTR(dst->buf),
+                                          dst->stride,
+                                          (unsigned int)mask_16x16,
+                                          (unsigned int)mask_8x8,
+                                          (unsigned int)mask_4x4,
+                                          (unsigned int)mask_4x4_int,
+                                          cm->lf_info.lfthr,
+                                          &lfm->lfl_y[r << 3],
+                                          (int)cm->bit_depth);
+    } else {
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      filter_selectively_vert_row2(plane->subsampling_x, dst->buf, dst->stride,
+                                   (unsigned int)mask_16x16,
+                                   (unsigned int)mask_8x8,
+                                   (unsigned int)mask_4x4,
+                                   (unsigned int)mask_4x4_int,
+                                   cm->lf_info.lfthr, &lfm->lfl_y[r << 3]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    dst->buf += 16 * dst->stride;
+    mask_16x16 >>= 16;
+    mask_8x8 >>= 16;
+    mask_4x4 >>= 16;
+    mask_4x4_int >>= 16;
+  }
+
+  // Horizontal pass
+  dst->buf = dst0;
+  mask_16x16 = lfm->above_y[TX_16X16];
+  mask_8x8 = lfm->above_y[TX_8X8];
+  mask_4x4 = lfm->above_y[TX_4X4];
+  mask_4x4_int = lfm->int_4x4_y;
+
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r++) {
+    unsigned int mask_16x16_r;
+    unsigned int mask_8x8_r;
+    unsigned int mask_4x4_r;
+
+    if (mi_row + r == 0) {
+      mask_16x16_r = 0;
+      mask_8x8_r = 0;
+      mask_4x4_r = 0;
+    } else {
+      mask_16x16_r = mask_16x16 & 0xff;
+      mask_8x8_r = mask_8x8 & 0xff;
+      mask_4x4_r = mask_4x4 & 0xff;
+    }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
+                                      dst->stride, mask_16x16_r, mask_8x8_r,
+                                      mask_4x4_r, mask_4x4_int & 0xff,
+                                      cm->lf_info.lfthr, &lfm->lfl_y[r << 3],
+                                      (int)cm->bit_depth);
+    } else {
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
+                               mask_4x4_r, mask_4x4_int & 0xff,
+                               cm->lf_info.lfthr, &lfm->lfl_y[r << 3]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    dst->buf += 8 * dst->stride;
+    mask_16x16 >>= 8;
+    mask_8x8 >>= 8;
+    mask_4x4 >>= 8;
+    mask_4x4_int >>= 8;
+  }
+}
+
+void vp9_filter_block_plane_ss11(VP9_COMMON *const cm,
+                                 struct macroblockd_plane *const plane,
+                                 int mi_row,
+                                 LOOP_FILTER_MASK *lfm) {
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t *const dst0 = dst->buf;
+  int r, c;
+  uint8_t lfl_uv[16];
+
+  uint16_t mask_16x16 = lfm->left_uv[TX_16X16];
+  uint16_t mask_8x8 = lfm->left_uv[TX_8X8];
+  uint16_t mask_4x4 = lfm->left_uv[TX_4X4];
+  uint16_t mask_4x4_int = lfm->int_4x4_uv;
+
+  assert(plane->subsampling_x == 1 && plane->subsampling_y == 1);
+
+  // Vertical pass: do 2 rows at one time
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 4) {
+    for (c = 0; c < (MI_BLOCK_SIZE >> 1); c++) {
+      lfl_uv[(r << 1) + c] = lfm->lfl_y[(r << 3) + (c << 1)];
+      lfl_uv[((r + 2) << 1) + c] = lfm->lfl_y[((r + 2) << 3) + (c << 1)];
+    }
+
+    // Disable filtering on the leftmost column.
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_vert_row2(plane->subsampling_x,
+                                          CONVERT_TO_SHORTPTR(dst->buf),
+                                          dst->stride,
+                                          (unsigned int)mask_16x16,
+                                          (unsigned int)mask_8x8,
+                                          (unsigned int)mask_4x4,
+                                          (unsigned int)mask_4x4_int,
+                                          cm->lf_info.lfthr, &lfl_uv[r << 1],
+                                          (int)cm->bit_depth);
+    } else {
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      filter_selectively_vert_row2(plane->subsampling_x, dst->buf, dst->stride,
+                                   (unsigned int)mask_16x16,
+                                   (unsigned int)mask_8x8,
+                                   (unsigned int)mask_4x4,
+                                   (unsigned int)mask_4x4_int,
+                                   cm->lf_info.lfthr, &lfl_uv[r << 1]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    dst->buf += 16 * dst->stride;
+    mask_16x16 >>= 8;
+    mask_8x8 >>= 8;
+    mask_4x4 >>= 8;
+    mask_4x4_int >>= 8;
+  }
+
+  // Horizontal pass
+  dst->buf = dst0;
+  mask_16x16 = lfm->above_uv[TX_16X16];
+  mask_8x8 = lfm->above_uv[TX_8X8];
+  mask_4x4 = lfm->above_uv[TX_4X4];
+  mask_4x4_int = lfm->int_4x4_uv;
+
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
+    const int skip_border_4x4_r = mi_row + r == cm->mi_rows - 1;
+    const unsigned int mask_4x4_int_r =
+        skip_border_4x4_r ? 0 : (mask_4x4_int & 0xf);
+    unsigned int mask_16x16_r;
+    unsigned int mask_8x8_r;
+    unsigned int mask_4x4_r;
+
+    if (mi_row + r == 0) {
+      mask_16x16_r = 0;
+      mask_8x8_r = 0;
+      mask_4x4_r = 0;
+    } else {
+      mask_16x16_r = mask_16x16 & 0xf;
+      mask_8x8_r = mask_8x8 & 0xf;
+      mask_4x4_r = mask_4x4 & 0xf;
+    }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
+                                      dst->stride, mask_16x16_r, mask_8x8_r,
+                                      mask_4x4_r, mask_4x4_int_r,
+                                      cm->lf_info.lfthr, &lfl_uv[r << 1],
+                                      (int)cm->bit_depth);
+    } else {
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
+                               mask_4x4_r, mask_4x4_int_r, cm->lf_info.lfthr,
+                               &lfl_uv[r << 1]);
+#if CONFIG_VP9_HIGHBITDEPTH
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    dst->buf += 8 * dst->stride;
+    mask_16x16 >>= 4;
+    mask_8x8 >>= 4;
+    mask_4x4 >>= 4;
+    mask_4x4_int >>= 4;
+  }
+}
+
+static void loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer, VP9_COMMON *cm,
+                             struct macroblockd_plane planes[MAX_MB_PLANE],
+                             int start, int stop, int y_only) {
+  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
+  enum lf_path path;
+  int mi_row, mi_col;
+
+  if (y_only)
+    path = LF_PATH_444;
+  else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
+    path = LF_PATH_420;
+  else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
+    path = LF_PATH_444;
+  else
+    path = LF_PATH_SLOW;
+
+  for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) {
+    MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+    LOOP_FILTER_MASK *lfm = get_lfm(&cm->lf, mi_row, 0);
+
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE, ++lfm) {
+      int plane;
+
+      vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
+
+      // TODO(jimbankoski): For 444 only need to do y mask.
+      vp9_adjust_mask(cm, mi_row, mi_col, lfm);
+
+      vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, lfm);
+      for (plane = 1; plane < num_planes; ++plane) {
+        switch (path) {
+          case LF_PATH_420:
+            vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, lfm);
+            break;
+          case LF_PATH_444:
+            vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, lfm);
+            break;
+          case LF_PATH_SLOW:
+            vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
+                                          mi_row, mi_col);
+            break;
+        }
+      }
+    }
+  }
+}
+
+void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
+                           VP9_COMMON *cm, MACROBLOCKD *xd,
+                           int frame_filter_level,
+                           int y_only, int partial_frame) {
+  int start_mi_row, end_mi_row, mi_rows_to_filter;
+  if (!frame_filter_level) return;
+  start_mi_row = 0;
+  mi_rows_to_filter = cm->mi_rows;
+  if (partial_frame && cm->mi_rows > 8) {
+    start_mi_row = cm->mi_rows >> 1;
+    start_mi_row &= 0xfffffff8;
+    mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
+  }
+  end_mi_row = start_mi_row + mi_rows_to_filter;
+  loop_filter_rows(frame, cm, xd->plane, start_mi_row, end_mi_row, y_only);
+}
+
+// Used by the encoder to build the loopfilter masks.
+// TODO(slavarnway): Do the encoder the same way the decoder does it and
+//                   build the masks in line as part of the encode process.
+void vp9_build_mask_frame(VP9_COMMON *cm, int frame_filter_level,
+                          int partial_frame) {
+  int start_mi_row, end_mi_row, mi_rows_to_filter;
+  int mi_col, mi_row;
+  if (!frame_filter_level) return;
+  start_mi_row = 0;
+  mi_rows_to_filter = cm->mi_rows;
+  if (partial_frame && cm->mi_rows > 8) {
+    start_mi_row = cm->mi_rows >> 1;
+    start_mi_row &= 0xfffffff8;
+    mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
+  }
+  end_mi_row = start_mi_row + mi_rows_to_filter;
+
+  vp9_loop_filter_frame_init(cm, frame_filter_level);
+
+  for (mi_row = start_mi_row; mi_row < end_mi_row; mi_row += MI_BLOCK_SIZE) {
+    MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
+      // vp9_setup_mask() zeros lfm
+      vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
+                     get_lfm(&cm->lf, mi_row, mi_col));
+    }
+  }
+}
+
+// 8x8 blocks in a superblock.  A "1" represents the first block in a 16x16
+// or greater area.
+static const uint8_t first_block_in_16x16[8][8] = {
+  {1, 0, 1, 0, 1, 0, 1, 0},
+  {0, 0, 0, 0, 0, 0, 0, 0},
+  {1, 0, 1, 0, 1, 0, 1, 0},
+  {0, 0, 0, 0, 0, 0, 0, 0},
+  {1, 0, 1, 0, 1, 0, 1, 0},
+  {0, 0, 0, 0, 0, 0, 0, 0},
+  {1, 0, 1, 0, 1, 0, 1, 0},
+  {0, 0, 0, 0, 0, 0, 0, 0}
+};
+
+// This function sets up the bit masks for a block represented
+// by mi_row, mi_col in a 64x64 region.
+// TODO(SJL): This function only works for yv12.
+void vp9_build_mask(VP9_COMMON *cm, const MODE_INFO *mi, int mi_row,
+                    int mi_col, int bw, int bh) {
+  const BLOCK_SIZE block_size = mi->sb_type;
+  const TX_SIZE tx_size_y = mi->tx_size;
+  const loop_filter_info_n *const lfi_n = &cm->lf_info;
+  const int filter_level = get_filter_level(lfi_n, mi);
+  const TX_SIZE tx_size_uv = get_uv_tx_size_impl(tx_size_y, block_size, 1, 1);
+  LOOP_FILTER_MASK *const lfm = get_lfm(&cm->lf, mi_row, mi_col);
+  uint64_t *const left_y = &lfm->left_y[tx_size_y];
+  uint64_t *const above_y = &lfm->above_y[tx_size_y];
+  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
+  uint16_t *const left_uv = &lfm->left_uv[tx_size_uv];
+  uint16_t *const above_uv = &lfm->above_uv[tx_size_uv];
+  uint16_t *const int_4x4_uv = &lfm->int_4x4_uv;
+  const int row_in_sb = (mi_row & 7);
+  const int col_in_sb = (mi_col & 7);
+  const int shift_y = col_in_sb + (row_in_sb << 3);
+  const int shift_uv = (col_in_sb >> 1) + ((row_in_sb >> 1) << 2);
+  const int build_uv = first_block_in_16x16[row_in_sb][col_in_sb];
+
+  if (!filter_level) {
+    return;
+  } else {
+    int index = shift_y;
+    int i;
+    for (i = 0; i < bh; i++) {
+      memset(&lfm->lfl_y[index], filter_level, bw);
+      index += 8;
+    }
+  }
+
+  // These set 1 in the current block size for the block size edges.
+  // For instance if the block size is 32x16, we'll set:
+  //    above =   1111
+  //              0000
+  //    and
+  //    left  =   1000
+  //          =   1000
+  // NOTE : In this example the low bit is left most ( 1000 ) is stored as
+  //        1,  not 8...
+  //
+  // U and V set things on a 16 bit scale.
+  //
+  *above_y |= above_prediction_mask[block_size] << shift_y;
+  *left_y |= left_prediction_mask[block_size] << shift_y;
+
+  if (build_uv) {
+    *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
+    *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;
+  }
+
+  // If the block has no coefficients and is not intra we skip applying
+  // the loop filter on block edges.
+  if (mi->skip && is_inter_block(mi))
+    return;
+
+  // Add a mask for the transform size. The transform size mask is set to
+  // be correct for a 64x64 prediction block size. Mask to match the size of
+  // the block we are working on and then shift it into place.
+  *above_y |= (size_mask[block_size] &
+               above_64x64_txform_mask[tx_size_y]) << shift_y;
+  *left_y |= (size_mask[block_size] &
+              left_64x64_txform_mask[tx_size_y]) << shift_y;
+
+  if (build_uv) {
+    *above_uv |= (size_mask_uv[block_size] &
+                  above_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
+
+    *left_uv |= (size_mask_uv[block_size] &
+                 left_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
+  }
+
+  // Try to determine what to do with the internal 4x4 block boundaries.  These
+  // differ from the 4x4 boundaries on the outside edge of an 8x8 in that the
+  // internal ones can be skipped and don't depend on the prediction block size.
+  if (tx_size_y == TX_4X4)
+    *int_4x4_y |= size_mask[block_size] << shift_y;
+
+  if (build_uv && tx_size_uv == TX_4X4)
+    *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
+}
+
+void vp9_loop_filter_data_reset(
+    LFWorkerData *lf_data, YV12_BUFFER_CONFIG *frame_buffer,
+    struct VP9Common *cm, const struct macroblockd_plane planes[MAX_MB_PLANE]) {
+  lf_data->frame_buffer = frame_buffer;
+  lf_data->cm = cm;
+  lf_data->start = 0;
+  lf_data->stop = 0;
+  lf_data->y_only = 0;
+  memcpy(lf_data->planes, planes, sizeof(lf_data->planes));
+}
+
+void vp9_reset_lfm(VP9_COMMON *const cm) {
+  if (cm->lf.filter_level) {
+    memset(cm->lf.lfm, 0,
+           ((cm->mi_rows + (MI_BLOCK_SIZE - 1)) >> 3) * cm->lf.lfm_stride *
+            sizeof(*cm->lf.lfm));
+  }
+}
+
+int vp9_loop_filter_worker(LFWorkerData *const lf_data, void *unused) {
+  (void)unused;
+  loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
+                   lf_data->start, lf_data->stop, lf_data->y_only);
+  return 1;
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_loopfilter.h b/libvpx/libvpx/vp9/common/vp9_loopfilter.h
new file mode 100644
index 0000000..fca8830
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_loopfilter.h
@@ -0,0 +1,166 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_LOOPFILTER_H_
+#define VP9_COMMON_VP9_LOOPFILTER_H_
+
+#include "vpx_ports/mem.h"
+#include "./vpx_config.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_LOOP_FILTER 63
+#define MAX_SHARPNESS 7
+
+#define SIMD_WIDTH 16
+
+#define MAX_REF_LF_DELTAS       4
+#define MAX_MODE_LF_DELTAS      2
+
+enum lf_path {
+  LF_PATH_420,
+  LF_PATH_444,
+  LF_PATH_SLOW,
+};
+
+// Need to align this structure so when it is declared and
+// passed it can be loaded into vector registers.
+typedef struct {
+  DECLARE_ALIGNED(SIMD_WIDTH, uint8_t, mblim[SIMD_WIDTH]);
+  DECLARE_ALIGNED(SIMD_WIDTH, uint8_t, lim[SIMD_WIDTH]);
+  DECLARE_ALIGNED(SIMD_WIDTH, uint8_t, hev_thr[SIMD_WIDTH]);
+} loop_filter_thresh;
+
+typedef struct {
+  loop_filter_thresh lfthr[MAX_LOOP_FILTER + 1];
+  uint8_t lvl[MAX_SEGMENTS][MAX_REF_FRAMES][MAX_MODE_LF_DELTAS];
+} loop_filter_info_n;
+
+// This structure holds bit masks for all 8x8 blocks in a 64x64 region.
+// Each 1 bit represents a position in which we want to apply the loop filter.
+// Left_ entries refer to whether we apply a filter on the border to the
+// left of the block.   Above_ entries refer to whether or not to apply a
+// filter on the above border.   Int_ entries refer to whether or not to
+// apply borders on the 4x4 edges within the 8x8 block that each bit
+// represents.
+// Since each transform is accompanied by a potentially different type of
+// loop filter there is a different entry in the array for each transform size.
+typedef struct {
+  uint64_t left_y[TX_SIZES];
+  uint64_t above_y[TX_SIZES];
+  uint64_t int_4x4_y;
+  uint16_t left_uv[TX_SIZES];
+  uint16_t above_uv[TX_SIZES];
+  uint16_t int_4x4_uv;
+  uint8_t lfl_y[64];
+} LOOP_FILTER_MASK;
+
+struct loopfilter {
+  int filter_level;
+  int last_filt_level;
+
+  int sharpness_level;
+  int last_sharpness_level;
+
+  uint8_t mode_ref_delta_enabled;
+  uint8_t mode_ref_delta_update;
+
+  // 0 = Intra, Last, GF, ARF
+  signed char ref_deltas[MAX_REF_LF_DELTAS];
+  signed char last_ref_deltas[MAX_REF_LF_DELTAS];
+
+  // 0 = ZERO_MV, MV
+  signed char mode_deltas[MAX_MODE_LF_DELTAS];
+  signed char last_mode_deltas[MAX_MODE_LF_DELTAS];
+
+  LOOP_FILTER_MASK *lfm;
+  int lfm_stride;
+};
+
+/* assorted loopfilter functions which get used elsewhere */
+struct VP9Common;
+struct macroblockd;
+struct VP9LfSyncData;
+
+// This function sets up the bit masks for the entire 64x64 region represented
+// by mi_row, mi_col.
+void vp9_setup_mask(struct VP9Common *const cm,
+                    const int mi_row, const int mi_col,
+                    MODE_INFO **mi_8x8, const int mode_info_stride,
+                    LOOP_FILTER_MASK *lfm);
+
+void vp9_filter_block_plane_ss00(struct VP9Common *const cm,
+                                 struct macroblockd_plane *const plane,
+                                 int mi_row,
+                                 LOOP_FILTER_MASK *lfm);
+
+void vp9_filter_block_plane_ss11(struct VP9Common *const cm,
+                                 struct macroblockd_plane *const plane,
+                                 int mi_row,
+                                 LOOP_FILTER_MASK *lfm);
+
+void vp9_filter_block_plane_non420(struct VP9Common *cm,
+                                   struct macroblockd_plane *plane,
+                                   MODE_INFO **mi_8x8,
+                                   int mi_row, int mi_col);
+
+void vp9_loop_filter_init(struct VP9Common *cm);
+
+// Update the loop filter for the current frame.
+// This should be called before vp9_loop_filter_frame(), vp9_build_mask_frame()
+// calls this function directly.
+void vp9_loop_filter_frame_init(struct VP9Common *cm, int default_filt_lvl);
+
+void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
+                           struct VP9Common *cm,
+                           struct macroblockd *mbd,
+                           int filter_level,
+                           int y_only, int partial_frame);
+
+// Get the superblock lfm for a given mi_row, mi_col.
+static INLINE LOOP_FILTER_MASK *get_lfm(const struct loopfilter *lf,
+                                        const int mi_row, const int mi_col) {
+  return &lf->lfm[(mi_col >> 3) + ((mi_row >> 3) * lf->lfm_stride)];
+}
+
+void vp9_build_mask(struct VP9Common *cm, const MODE_INFO *mi, int mi_row,
+                    int mi_col, int bw, int bh);
+void vp9_adjust_mask(struct VP9Common *const cm, const int mi_row,
+                     const int mi_col, LOOP_FILTER_MASK *lfm);
+void vp9_build_mask_frame(struct VP9Common *cm, int frame_filter_level,
+                          int partial_frame);
+void vp9_reset_lfm(struct VP9Common *const cm);
+
+typedef struct LoopFilterWorkerData {
+  YV12_BUFFER_CONFIG *frame_buffer;
+  struct VP9Common *cm;
+  struct macroblockd_plane planes[MAX_MB_PLANE];
+
+  int start;
+  int stop;
+  int y_only;
+} LFWorkerData;
+
+void vp9_loop_filter_data_reset(
+    LFWorkerData *lf_data, YV12_BUFFER_CONFIG *frame_buffer,
+    struct VP9Common *cm, const struct macroblockd_plane planes[MAX_MB_PLANE]);
+
+// Operates on the rows described by 'lf_data'.
+int vp9_loop_filter_worker(LFWorkerData *const lf_data, void *unused);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_LOOPFILTER_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_mfqe.c b/libvpx/libvpx/vp9/common/vp9_mfqe.c
new file mode 100644
index 0000000..f526466
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_mfqe.c
@@ -0,0 +1,394 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_postproc.h"
+
+// TODO(jackychen): Replace this function with SSE2 code. There is
+// one SSE2 implementation in vp8, so will consider how to share it
+// between vp8 and vp9.
+static void filter_by_weight(const uint8_t *src, int src_stride,
+                             uint8_t *dst, int dst_stride,
+                             int block_size, int src_weight) {
+  const int dst_weight = (1 << MFQE_PRECISION) - src_weight;
+  const int rounding_bit = 1 << (MFQE_PRECISION - 1);
+  int r, c;
+
+  for (r = 0; r < block_size; r++) {
+    for (c = 0; c < block_size; c++) {
+      dst[c] = (src[c] * src_weight + dst[c] * dst_weight + rounding_bit)
+               >> MFQE_PRECISION;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vp9_filter_by_weight8x8_c(const uint8_t *src, int src_stride,
+                               uint8_t *dst, int dst_stride, int src_weight) {
+  filter_by_weight(src, src_stride, dst, dst_stride, 8, src_weight);
+}
+
+void vp9_filter_by_weight16x16_c(const uint8_t *src, int src_stride,
+                                 uint8_t *dst, int dst_stride,
+                                 int src_weight) {
+  filter_by_weight(src, src_stride, dst, dst_stride, 16, src_weight);
+}
+
+static void filter_by_weight32x32(const uint8_t *src, int src_stride,
+                                  uint8_t *dst, int dst_stride, int weight) {
+  vp9_filter_by_weight16x16(src, src_stride, dst, dst_stride, weight);
+  vp9_filter_by_weight16x16(src + 16, src_stride, dst + 16, dst_stride,
+                            weight);
+  vp9_filter_by_weight16x16(src + src_stride * 16, src_stride,
+                            dst + dst_stride * 16, dst_stride, weight);
+  vp9_filter_by_weight16x16(src + src_stride * 16 + 16, src_stride,
+                            dst + dst_stride * 16 + 16, dst_stride, weight);
+}
+
+static void filter_by_weight64x64(const uint8_t *src, int src_stride,
+                                  uint8_t *dst, int dst_stride, int weight) {
+  filter_by_weight32x32(src, src_stride, dst, dst_stride, weight);
+  filter_by_weight32x32(src + 32, src_stride, dst + 32,
+                        dst_stride, weight);
+  filter_by_weight32x32(src + src_stride * 32, src_stride,
+                        dst + dst_stride * 32, dst_stride, weight);
+  filter_by_weight32x32(src + src_stride * 32 + 32, src_stride,
+                        dst + dst_stride * 32 + 32, dst_stride, weight);
+}
+
+static void apply_ifactor(const uint8_t *y, int y_stride, uint8_t *yd,
+                          int yd_stride, const uint8_t *u, const uint8_t *v,
+                          int uv_stride, uint8_t *ud, uint8_t *vd,
+                          int uvd_stride, BLOCK_SIZE block_size,
+                          int weight) {
+  if (block_size == BLOCK_16X16) {
+    vp9_filter_by_weight16x16(y, y_stride, yd, yd_stride, weight);
+    vp9_filter_by_weight8x8(u, uv_stride, ud, uvd_stride, weight);
+    vp9_filter_by_weight8x8(v, uv_stride, vd, uvd_stride, weight);
+  } else if (block_size == BLOCK_32X32) {
+    filter_by_weight32x32(y, y_stride, yd, yd_stride, weight);
+    vp9_filter_by_weight16x16(u, uv_stride, ud, uvd_stride, weight);
+    vp9_filter_by_weight16x16(v, uv_stride, vd, uvd_stride, weight);
+  } else if (block_size == BLOCK_64X64) {
+    filter_by_weight64x64(y, y_stride, yd, yd_stride, weight);
+    filter_by_weight32x32(u, uv_stride, ud, uvd_stride, weight);
+    filter_by_weight32x32(v, uv_stride, vd, uvd_stride, weight);
+  }
+}
+
+// TODO(jackychen): Determine whether replace it with assembly code.
+static void copy_mem8x8(const uint8_t *src, int src_stride,
+                        uint8_t *dst, int dst_stride) {
+  int r;
+  for (r = 0; r < 8; r++) {
+    memcpy(dst, src, 8);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void copy_mem16x16(const uint8_t *src, int src_stride,
+                          uint8_t *dst, int dst_stride) {
+  int r;
+  for (r = 0; r < 16; r++) {
+    memcpy(dst, src, 16);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void copy_mem32x32(const uint8_t *src, int src_stride,
+                          uint8_t *dst, int dst_stride) {
+  copy_mem16x16(src, src_stride, dst, dst_stride);
+  copy_mem16x16(src + 16, src_stride, dst + 16, dst_stride);
+  copy_mem16x16(src + src_stride * 16, src_stride,
+                dst + dst_stride * 16, dst_stride);
+  copy_mem16x16(src + src_stride * 16 + 16, src_stride,
+                dst + dst_stride * 16 + 16, dst_stride);
+}
+
+static void copy_mem64x64(const uint8_t *src, int src_stride,
+                          uint8_t *dst, int dst_stride) {
+  copy_mem32x32(src, src_stride, dst, dst_stride);
+  copy_mem32x32(src + 32, src_stride, dst + 32, dst_stride);
+  copy_mem32x32(src + src_stride * 32, src_stride,
+                dst + src_stride * 32, dst_stride);
+  copy_mem32x32(src + src_stride * 32 + 32, src_stride,
+                dst + src_stride * 32 + 32, dst_stride);
+}
+
+static void copy_block(const uint8_t *y, const uint8_t *u, const uint8_t *v,
+                       int y_stride, int uv_stride, uint8_t *yd, uint8_t *ud,
+                       uint8_t *vd, int yd_stride, int uvd_stride,
+                       BLOCK_SIZE bs) {
+  if (bs == BLOCK_16X16) {
+    copy_mem16x16(y, y_stride, yd, yd_stride);
+    copy_mem8x8(u, uv_stride, ud, uvd_stride);
+    copy_mem8x8(v, uv_stride, vd, uvd_stride);
+  } else if (bs == BLOCK_32X32) {
+    copy_mem32x32(y, y_stride, yd, yd_stride);
+    copy_mem16x16(u, uv_stride, ud, uvd_stride);
+    copy_mem16x16(v, uv_stride, vd, uvd_stride);
+  } else {
+    copy_mem64x64(y, y_stride, yd, yd_stride);
+    copy_mem32x32(u, uv_stride, ud, uvd_stride);
+    copy_mem32x32(v, uv_stride, vd, uvd_stride);
+  }
+}
+
+static void get_thr(BLOCK_SIZE bs, int qdiff, int *sad_thr, int *vdiff_thr) {
+  const int adj = qdiff >> MFQE_PRECISION;
+  if (bs == BLOCK_16X16) {
+    *sad_thr = 7 + adj;
+  } else if (bs == BLOCK_32X32) {
+    *sad_thr = 6 + adj;
+  } else {  // BLOCK_64X64
+    *sad_thr = 5 + adj;
+  }
+  *vdiff_thr = 125 + qdiff;
+}
+
+static void mfqe_block(BLOCK_SIZE bs, const uint8_t *y, const uint8_t *u,
+                       const uint8_t *v, int y_stride, int uv_stride,
+                       uint8_t *yd, uint8_t *ud, uint8_t *vd, int yd_stride,
+                       int uvd_stride, int qdiff) {
+  int sad, sad_thr, vdiff, vdiff_thr;
+  uint32_t sse;
+
+  get_thr(bs, qdiff, &sad_thr, &vdiff_thr);
+
+  if (bs == BLOCK_16X16) {
+    vdiff = (vpx_variance16x16(y, y_stride, yd, yd_stride, &sse) + 128) >> 8;
+    sad = (vpx_sad16x16(y, y_stride, yd, yd_stride) + 128) >> 8;
+  } else if (bs == BLOCK_32X32) {
+    vdiff = (vpx_variance32x32(y, y_stride, yd, yd_stride, &sse) + 512) >> 10;
+    sad = (vpx_sad32x32(y, y_stride, yd, yd_stride) + 512) >> 10;
+  } else /* if (bs == BLOCK_64X64) */ {
+    vdiff = (vpx_variance64x64(y, y_stride, yd, yd_stride, &sse) + 2048) >> 12;
+    sad = (vpx_sad64x64(y, y_stride, yd, yd_stride) + 2048) >> 12;
+  }
+
+  // vdiff > sad * 3 means vdiff should not be too small, otherwise,
+  // it might be a lighting change in smooth area. When there is a
+  // lighting change in smooth area, it is dangerous to do MFQE.
+  if (sad > 1 && vdiff > sad * 3) {
+    const int weight = 1 << MFQE_PRECISION;
+    int ifactor = weight * sad * vdiff / (sad_thr * vdiff_thr);
+    // When ifactor equals weight, no MFQE is done.
+    if (ifactor > weight) {
+      ifactor = weight;
+    }
+    apply_ifactor(y, y_stride, yd, yd_stride, u, v, uv_stride, ud, vd,
+                  uvd_stride, bs, ifactor);
+  } else {
+    // Copy the block from current frame (i.e., no mfqe is done).
+    copy_block(y, u, v, y_stride, uv_stride, yd, ud, vd,
+               yd_stride, uvd_stride, bs);
+  }
+}
+
+static int mfqe_decision(MODE_INFO *mi, BLOCK_SIZE cur_bs) {
+  // Check the motion in current block(for inter frame),
+  // or check the motion in the correlated block in last frame (for keyframe).
+  const int mv_len_square = mi->mv[0].as_mv.row *
+                            mi->mv[0].as_mv.row +
+                            mi->mv[0].as_mv.col *
+                            mi->mv[0].as_mv.col;
+  const int mv_threshold = 100;
+  return mi->mode >= NEARESTMV &&  // Not an intra block
+         cur_bs >= BLOCK_16X16 &&
+         mv_len_square <= mv_threshold;
+}
+
+// Process each partiton in a super block, recursively.
+static void mfqe_partition(VP9_COMMON *cm, MODE_INFO *mi, BLOCK_SIZE bs,
+                           const uint8_t *y, const uint8_t *u,
+                           const uint8_t *v, int y_stride, int uv_stride,
+                           uint8_t *yd, uint8_t *ud, uint8_t *vd,
+                           int yd_stride, int uvd_stride) {
+  int mi_offset, y_offset, uv_offset;
+  const BLOCK_SIZE cur_bs = mi->sb_type;
+  const int qdiff = cm->base_qindex - cm->postproc_state.last_base_qindex;
+  const int bsl = b_width_log2_lookup[bs];
+  PARTITION_TYPE partition = partition_lookup[bsl][cur_bs];
+  const BLOCK_SIZE subsize = get_subsize(bs, partition);
+
+  if (cur_bs < BLOCK_8X8) {
+    // If there are blocks smaller than 8x8, it must be on the boundary.
+    return;
+  }
+  // No MFQE on blocks smaller than 16x16
+  if (bs == BLOCK_16X16) {
+    partition = PARTITION_NONE;
+  }
+  if (bs == BLOCK_64X64) {
+    mi_offset = 4;
+    y_offset = 32;
+    uv_offset = 16;
+  } else {
+    mi_offset = 2;
+    y_offset = 16;
+    uv_offset = 8;
+  }
+  switch (partition) {
+    BLOCK_SIZE mfqe_bs, bs_tmp;
+    case PARTITION_HORZ:
+      if (bs == BLOCK_64X64) {
+        mfqe_bs = BLOCK_64X32;
+        bs_tmp = BLOCK_32X32;
+      } else {
+        mfqe_bs = BLOCK_32X16;
+        bs_tmp = BLOCK_16X16;
+      }
+      if (mfqe_decision(mi, mfqe_bs)) {
+        // Do mfqe on the first square partition.
+        mfqe_block(bs_tmp, y, u, v, y_stride, uv_stride,
+                   yd, ud, vd, yd_stride, uvd_stride, qdiff);
+        // Do mfqe on the second square partition.
+        mfqe_block(bs_tmp, y + y_offset, u + uv_offset, v + uv_offset,
+                   y_stride, uv_stride, yd + y_offset, ud + uv_offset,
+                   vd + uv_offset, yd_stride, uvd_stride, qdiff);
+      }
+      if (mfqe_decision(mi + mi_offset * cm->mi_stride, mfqe_bs)) {
+        // Do mfqe on the first square partition.
+        mfqe_block(bs_tmp, y + y_offset * y_stride, u + uv_offset * uv_stride,
+                   v + uv_offset * uv_stride, y_stride, uv_stride,
+                   yd + y_offset * yd_stride, ud + uv_offset * uvd_stride,
+                   vd + uv_offset * uvd_stride, yd_stride, uvd_stride, qdiff);
+        // Do mfqe on the second square partition.
+        mfqe_block(bs_tmp, y + y_offset * y_stride + y_offset,
+                   u + uv_offset * uv_stride + uv_offset,
+                   v + uv_offset * uv_stride + uv_offset, y_stride,
+                   uv_stride, yd + y_offset * yd_stride + y_offset,
+                   ud + uv_offset * uvd_stride + uv_offset,
+                   vd + uv_offset * uvd_stride + uv_offset,
+                   yd_stride, uvd_stride, qdiff);
+      }
+      break;
+    case PARTITION_VERT:
+      if (bs == BLOCK_64X64) {
+        mfqe_bs = BLOCK_32X64;
+        bs_tmp = BLOCK_32X32;
+      } else {
+        mfqe_bs = BLOCK_16X32;
+        bs_tmp = BLOCK_16X16;
+      }
+      if (mfqe_decision(mi, mfqe_bs)) {
+        // Do mfqe on the first square partition.
+        mfqe_block(bs_tmp, y, u, v, y_stride, uv_stride,
+                   yd, ud, vd, yd_stride, uvd_stride, qdiff);
+        // Do mfqe on the second square partition.
+        mfqe_block(bs_tmp, y + y_offset * y_stride, u + uv_offset * uv_stride,
+                   v + uv_offset * uv_stride, y_stride, uv_stride,
+                   yd + y_offset * yd_stride, ud + uv_offset * uvd_stride,
+                   vd + uv_offset * uvd_stride, yd_stride, uvd_stride, qdiff);
+      }
+      if (mfqe_decision(mi + mi_offset, mfqe_bs)) {
+        // Do mfqe on the first square partition.
+        mfqe_block(bs_tmp, y + y_offset, u + uv_offset, v + uv_offset,
+                   y_stride, uv_stride, yd + y_offset, ud + uv_offset,
+                   vd + uv_offset, yd_stride, uvd_stride, qdiff);
+        // Do mfqe on the second square partition.
+        mfqe_block(bs_tmp, y + y_offset * y_stride + y_offset,
+                   u + uv_offset * uv_stride + uv_offset,
+                   v + uv_offset * uv_stride + uv_offset, y_stride,
+                   uv_stride, yd + y_offset * yd_stride + y_offset,
+                   ud + uv_offset * uvd_stride + uv_offset,
+                   vd + uv_offset * uvd_stride + uv_offset,
+                   yd_stride, uvd_stride, qdiff);
+      }
+      break;
+    case PARTITION_NONE:
+      if (mfqe_decision(mi, cur_bs)) {
+        // Do mfqe on this partition.
+        mfqe_block(cur_bs, y, u, v, y_stride, uv_stride,
+                   yd, ud, vd, yd_stride, uvd_stride, qdiff);
+      } else {
+        // Copy the block from current frame(i.e., no mfqe is done).
+        copy_block(y, u, v, y_stride, uv_stride, yd, ud, vd,
+                   yd_stride, uvd_stride, bs);
+      }
+      break;
+    case PARTITION_SPLIT:
+      // Recursion on four square partitions, e.g. if bs is 64X64,
+      // then look into four 32X32 blocks in it.
+      mfqe_partition(cm, mi, subsize, y, u, v, y_stride, uv_stride, yd, ud, vd,
+                     yd_stride, uvd_stride);
+      mfqe_partition(cm, mi + mi_offset, subsize, y + y_offset, u + uv_offset,
+                     v + uv_offset, y_stride, uv_stride, yd + y_offset,
+                     ud + uv_offset, vd + uv_offset, yd_stride, uvd_stride);
+      mfqe_partition(cm, mi + mi_offset * cm->mi_stride, subsize,
+                     y + y_offset * y_stride, u + uv_offset * uv_stride,
+                     v + uv_offset * uv_stride, y_stride, uv_stride,
+                     yd + y_offset * yd_stride, ud + uv_offset * uvd_stride,
+                     vd + uv_offset * uvd_stride, yd_stride, uvd_stride);
+      mfqe_partition(cm, mi + mi_offset * cm->mi_stride + mi_offset,
+                     subsize, y + y_offset * y_stride + y_offset,
+                     u + uv_offset * uv_stride + uv_offset,
+                     v + uv_offset * uv_stride + uv_offset, y_stride,
+                     uv_stride, yd + y_offset * yd_stride + y_offset,
+                     ud + uv_offset * uvd_stride + uv_offset,
+                     vd + uv_offset * uvd_stride + uv_offset,
+                     yd_stride, uvd_stride);
+      break;
+    default:
+      assert(0);
+  }
+}
+
+void vp9_mfqe(VP9_COMMON *cm) {
+  int mi_row, mi_col;
+  // Current decoded frame.
+  const YV12_BUFFER_CONFIG *show = cm->frame_to_show;
+  // Last decoded frame and will store the MFQE result.
+  YV12_BUFFER_CONFIG *dest = &cm->post_proc_buffer;
+  // Loop through each super block.
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row += MI_BLOCK_SIZE) {
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
+      MODE_INFO *mi;
+      MODE_INFO *mi_local = cm->mi + (mi_row * cm->mi_stride + mi_col);
+      // Motion Info in last frame.
+      MODE_INFO *mi_prev = cm->postproc_state.prev_mi +
+                           (mi_row * cm->mi_stride + mi_col);
+      const uint32_t y_stride = show->y_stride;
+      const uint32_t uv_stride = show->uv_stride;
+      const uint32_t yd_stride = dest->y_stride;
+      const uint32_t uvd_stride = dest->uv_stride;
+      const uint32_t row_offset_y = mi_row << 3;
+      const uint32_t row_offset_uv = mi_row << 2;
+      const uint32_t col_offset_y = mi_col << 3;
+      const uint32_t col_offset_uv = mi_col << 2;
+      const uint8_t *y = show->y_buffer + row_offset_y * y_stride +
+                         col_offset_y;
+      const uint8_t *u = show->u_buffer + row_offset_uv * uv_stride +
+                         col_offset_uv;
+      const uint8_t *v = show->v_buffer + row_offset_uv * uv_stride +
+                         col_offset_uv;
+      uint8_t *yd = dest->y_buffer + row_offset_y * yd_stride + col_offset_y;
+      uint8_t *ud = dest->u_buffer + row_offset_uv * uvd_stride +
+                    col_offset_uv;
+      uint8_t *vd = dest->v_buffer + row_offset_uv * uvd_stride +
+                    col_offset_uv;
+      if (frame_is_intra_only(cm)) {
+        mi = mi_prev;
+      } else {
+        mi = mi_local;
+      }
+      mfqe_partition(cm, mi, BLOCK_64X64, y, u, v, y_stride, uv_stride, yd, ud,
+                     vd, yd_stride, uvd_stride);
+    }
+  }
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_mfqe.h b/libvpx/libvpx/vp9/common/vp9_mfqe.h
new file mode 100644
index 0000000..dfff8c2
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_mfqe.h
@@ -0,0 +1,31 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_MFQE_H_
+#define VP9_COMMON_VP9_MFQE_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Multiframe Quality Enhancement.
+// The aim for MFQE is to replace pixel blocks in the current frame with
+// the correlated pixel blocks (with higher quality) in the last frame.
+// The replacement can only be taken in stationary blocks by checking
+// the motion of the blocks and other conditions such as the SAD of
+// the current block and correlated block, the variance of the block
+// difference, etc.
+void vp9_mfqe(struct VP9Common *cm);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_MFQE_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_mv.h b/libvpx/libvpx/vp9/common/vp9_mv.h
new file mode 100644
index 0000000..5d89da8
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_mv.h
@@ -0,0 +1,55 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_MV_H_
+#define VP9_COMMON_VP9_MV_H_
+
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct mv {
+  int16_t row;
+  int16_t col;
+} MV;
+
+typedef union int_mv {
+  uint32_t as_int;
+  MV as_mv;
+} int_mv; /* facilitates faster equality tests and copies */
+
+typedef struct mv32 {
+  int32_t row;
+  int32_t col;
+} MV32;
+
+static INLINE int is_zero_mv(const MV *mv) {
+  return *((const uint32_t *)mv) == 0;
+}
+
+static INLINE int is_equal_mv(const MV *a, const MV *b) {
+  return  *((const uint32_t *)a) == *((const uint32_t *)b);
+}
+
+static INLINE void clamp_mv(MV *mv, int min_col, int max_col,
+                            int min_row, int max_row) {
+  mv->col = clamp(mv->col, min_col, max_col);
+  mv->row = clamp(mv->row, min_row, max_row);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_MV_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_mvref_common.c b/libvpx/libvpx/vp9/common/vp9_mvref_common.c
new file mode 100644
index 0000000..0eb01a5
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_mvref_common.c
@@ -0,0 +1,201 @@
+
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_mvref_common.h"
+
+// This function searches the neighborhood of a given MB/SB
+// to try and find candidate reference vectors.
+static void find_mv_refs_idx(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                             MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                             int_mv *mv_ref_list,
+                             int block, int mi_row, int mi_col,
+                             uint8_t *mode_context) {
+  const int *ref_sign_bias = cm->ref_frame_sign_bias;
+  int i, refmv_count = 0;
+  const POSITION *const mv_ref_search = mv_ref_blocks[mi->sb_type];
+  int different_ref_found = 0;
+  int context_counter = 0;
+  const MV_REF *const prev_frame_mvs = cm->use_prev_frame_mvs ?
+      cm->prev_frame->mvs + mi_row * cm->mi_cols + mi_col : NULL;
+  const TileInfo *const tile = &xd->tile;
+
+  // Blank the reference vector list
+  memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
+
+  // The nearest 2 blocks are treated differently
+  // if the size < 8x8 we get the mv from the bmi substructure,
+  // and we also need to keep a mode count.
+  for (i = 0; i < 2; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row *
+                                                   xd->mi_stride];
+      // Keep counts for entropy encoding.
+      context_counter += mode_2_counter[candidate_mi->mode];
+      different_ref_found = 1;
+
+      if (candidate_mi->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, block),
+                        refmv_count, mv_ref_list, Done);
+      else if (candidate_mi->ref_frame[1] == ref_frame)
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 1, mv_ref->col, block),
+                        refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  // Check the rest of the neighbors in much the same way
+  // as before except we don't need to keep track of sub blocks or
+  // mode counts.
+  for (; i < MVREF_NEIGHBOURS; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MODE_INFO *const candidate_mi =
+          xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
+      different_ref_found = 1;
+
+      if (candidate_mi->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(candidate_mi->mv[0], refmv_count, mv_ref_list, Done);
+      else if (candidate_mi->ref_frame[1] == ref_frame)
+        ADD_MV_REF_LIST(candidate_mi->mv[1], refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  // Check the last frame's mode and mv info.
+  if (cm->use_prev_frame_mvs) {
+    if (prev_frame_mvs->ref_frame[0] == ref_frame) {
+      ADD_MV_REF_LIST(prev_frame_mvs->mv[0], refmv_count, mv_ref_list, Done);
+    } else if (prev_frame_mvs->ref_frame[1] == ref_frame) {
+      ADD_MV_REF_LIST(prev_frame_mvs->mv[1], refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  // Since we couldn't find 2 mvs from the same reference frame
+  // go back through the neighbors and find motion vectors from
+  // different reference frames.
+  if (different_ref_found) {
+    for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
+      const POSITION *mv_ref = &mv_ref_search[i];
+      if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+        const MODE_INFO *const candidate_mi =
+            xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
+
+        // If the candidate is INTRA we don't want to consider its mv.
+        IF_DIFF_REF_FRAME_ADD_MV(candidate_mi, ref_frame, ref_sign_bias,
+                                 refmv_count, mv_ref_list, Done);
+      }
+    }
+  }
+
+  // Since we still don't have a candidate we'll try the last frame.
+  if (cm->use_prev_frame_mvs) {
+    if (prev_frame_mvs->ref_frame[0] != ref_frame &&
+        prev_frame_mvs->ref_frame[0] > INTRA_FRAME) {
+      int_mv mv = prev_frame_mvs->mv[0];
+      if (ref_sign_bias[prev_frame_mvs->ref_frame[0]] !=
+          ref_sign_bias[ref_frame]) {
+        mv.as_mv.row *= -1;
+        mv.as_mv.col *= -1;
+      }
+      ADD_MV_REF_LIST(mv, refmv_count, mv_ref_list, Done);
+    }
+
+    if (prev_frame_mvs->ref_frame[1] > INTRA_FRAME &&
+        prev_frame_mvs->ref_frame[1] != ref_frame &&
+        prev_frame_mvs->mv[1].as_int != prev_frame_mvs->mv[0].as_int) {
+      int_mv mv = prev_frame_mvs->mv[1];
+      if (ref_sign_bias[prev_frame_mvs->ref_frame[1]] !=
+          ref_sign_bias[ref_frame]) {
+        mv.as_mv.row *= -1;
+        mv.as_mv.col *= -1;
+      }
+      ADD_MV_REF_LIST(mv, refmv_count, mv_ref_list, Done);
+    }
+  }
+
+ Done:
+
+  mode_context[ref_frame] = counter_to_context[context_counter];
+
+  // Clamp vectors
+  for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
+    clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
+}
+
+void vp9_find_mv_refs(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                      MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                      int_mv *mv_ref_list,
+                      int mi_row, int mi_col,
+                      uint8_t *mode_context) {
+  find_mv_refs_idx(cm, xd, mi, ref_frame, mv_ref_list, -1,
+                   mi_row, mi_col, mode_context);
+}
+
+void vp9_find_best_ref_mvs(MACROBLOCKD *xd, int allow_hp,
+                           int_mv *mvlist, int_mv *nearest_mv,
+                           int_mv *near_mv) {
+  int i;
+  // Make sure all the candidates are properly clamped etc
+  for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i) {
+    lower_mv_precision(&mvlist[i].as_mv, allow_hp);
+    clamp_mv2(&mvlist[i].as_mv, xd);
+  }
+  *nearest_mv = mvlist[0];
+  *near_mv = mvlist[1];
+}
+
+void vp9_append_sub8x8_mvs_for_idx(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                   int block, int ref, int mi_row, int mi_col,
+                                   int_mv *nearest_mv, int_mv *near_mv,
+                                   uint8_t *mode_context) {
+  int_mv mv_list[MAX_MV_REF_CANDIDATES];
+  MODE_INFO *const mi = xd->mi[0];
+  b_mode_info *bmi = mi->bmi;
+  int n;
+
+  assert(MAX_MV_REF_CANDIDATES == 2);
+
+  find_mv_refs_idx(cm, xd, mi, mi->ref_frame[ref], mv_list, block,
+                   mi_row, mi_col, mode_context);
+
+  near_mv->as_int = 0;
+  switch (block) {
+    case 0:
+      nearest_mv->as_int = mv_list[0].as_int;
+      near_mv->as_int = mv_list[1].as_int;
+      break;
+    case 1:
+    case 2:
+      nearest_mv->as_int = bmi[0].as_mv[ref].as_int;
+      for (n = 0; n < MAX_MV_REF_CANDIDATES; ++n)
+        if (nearest_mv->as_int != mv_list[n].as_int) {
+          near_mv->as_int = mv_list[n].as_int;
+          break;
+        }
+      break;
+    case 3: {
+      int_mv candidates[2 + MAX_MV_REF_CANDIDATES];
+      candidates[0] = bmi[1].as_mv[ref];
+      candidates[1] = bmi[0].as_mv[ref];
+      candidates[2] = mv_list[0];
+      candidates[3] = mv_list[1];
+
+      nearest_mv->as_int = bmi[2].as_mv[ref].as_int;
+      for (n = 0; n < 2 + MAX_MV_REF_CANDIDATES; ++n)
+        if (nearest_mv->as_int != candidates[n].as_int) {
+          near_mv->as_int = candidates[n].as_int;
+          break;
+        }
+      break;
+    }
+    default:
+      assert(0 && "Invalid block index.");
+  }
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_mvref_common.h b/libvpx/libvpx/vp9/common/vp9_mvref_common.h
new file mode 100644
index 0000000..4380843
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_mvref_common.h
@@ -0,0 +1,241 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VP9_COMMON_VP9_MVREF_COMMON_H_
+#define VP9_COMMON_VP9_MVREF_COMMON_H_
+
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define LEFT_TOP_MARGIN ((VP9_ENC_BORDER_IN_PIXELS - VP9_INTERP_EXTEND) << 3)
+#define RIGHT_BOTTOM_MARGIN ((VP9_ENC_BORDER_IN_PIXELS -\
+                                VP9_INTERP_EXTEND) << 3)
+
+#define MVREF_NEIGHBOURS 8
+
+typedef struct position {
+  int row;
+  int col;
+} POSITION;
+
+typedef enum {
+  BOTH_ZERO = 0,
+  ZERO_PLUS_PREDICTED = 1,
+  BOTH_PREDICTED = 2,
+  NEW_PLUS_NON_INTRA = 3,
+  BOTH_NEW = 4,
+  INTRA_PLUS_NON_INTRA = 5,
+  BOTH_INTRA = 6,
+  INVALID_CASE = 9
+} motion_vector_context;
+
+// This is used to figure out a context for the ref blocks. The code flattens
+// an array that would have 3 possible counts (0, 1 & 2) for 3 choices by
+// adding 9 for each intra block, 3 for each zero mv and 1 for each new
+// motion vector. This single number is then converted into a context
+// with a single lookup ( counter_to_context ).
+static const int mode_2_counter[MB_MODE_COUNT] = {
+  9,  // DC_PRED
+  9,  // V_PRED
+  9,  // H_PRED
+  9,  // D45_PRED
+  9,  // D135_PRED
+  9,  // D117_PRED
+  9,  // D153_PRED
+  9,  // D207_PRED
+  9,  // D63_PRED
+  9,  // TM_PRED
+  0,  // NEARESTMV
+  0,  // NEARMV
+  3,  // ZEROMV
+  1,  // NEWMV
+};
+
+// There are 3^3 different combinations of 3 counts that can be either 0,1 or
+// 2. However the actual count can never be greater than 2 so the highest
+// counter we need is 18. 9 is an invalid counter that's never used.
+static const int counter_to_context[19] = {
+  BOTH_PREDICTED,  // 0
+  NEW_PLUS_NON_INTRA,  // 1
+  BOTH_NEW,  // 2
+  ZERO_PLUS_PREDICTED,  // 3
+  NEW_PLUS_NON_INTRA,  // 4
+  INVALID_CASE,  // 5
+  BOTH_ZERO,  // 6
+  INVALID_CASE,  // 7
+  INVALID_CASE,  // 8
+  INTRA_PLUS_NON_INTRA,  // 9
+  INTRA_PLUS_NON_INTRA,  // 10
+  INVALID_CASE,  // 11
+  INTRA_PLUS_NON_INTRA,  // 12
+  INVALID_CASE,  // 13
+  INVALID_CASE,  // 14
+  INVALID_CASE,  // 15
+  INVALID_CASE,  // 16
+  INVALID_CASE,  // 17
+  BOTH_INTRA  // 18
+};
+
+static const POSITION mv_ref_blocks[BLOCK_SIZES][MVREF_NEIGHBOURS] = {
+  // 4X4
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 4X8
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X4
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X8
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X16
+  {{0, -1}, {-1, 0}, {1, -1}, {-1, -1}, {0, -2}, {-2, 0}, {-2, -1}, {-1, -2}},
+  // 16X8
+  {{-1, 0}, {0, -1}, {-1, 1}, {-1, -1}, {-2, 0}, {0, -2}, {-1, -2}, {-2, -1}},
+  // 16X16
+  {{-1, 0}, {0, -1}, {-1, 1}, {1, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 16X32
+  {{0, -1}, {-1, 0}, {2, -1}, {-1, -1}, {-1, 1}, {0, -3}, {-3, 0}, {-3, -3}},
+  // 32X16
+  {{-1, 0}, {0, -1}, {-1, 2}, {-1, -1}, {1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 32X32
+  {{-1, 1}, {1, -1}, {-1, 2}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 32X64
+  {{0, -1}, {-1, 0}, {4, -1}, {-1, 2}, {-1, -1}, {0, -3}, {-3, 0}, {2, -1}},
+  // 64X32
+  {{-1, 0}, {0, -1}, {-1, 4}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-1, 2}},
+  // 64X64
+  {{-1, 3}, {3, -1}, {-1, 4}, {4, -1}, {-1, -1}, {-1, 0}, {0, -1}, {-1, 6}}
+};
+
+static const int idx_n_column_to_subblock[4][2] = {
+  {1, 2},
+  {1, 3},
+  {3, 2},
+  {3, 3}
+};
+
+// clamp_mv_ref
+#define MV_BORDER (16 << 3)  // Allow 16 pels in 1/8th pel units
+
+static INLINE void clamp_mv_ref(MV *mv, const MACROBLOCKD *xd) {
+  clamp_mv(mv, xd->mb_to_left_edge - MV_BORDER,
+               xd->mb_to_right_edge + MV_BORDER,
+               xd->mb_to_top_edge - MV_BORDER,
+               xd->mb_to_bottom_edge + MV_BORDER);
+}
+
+// This function returns either the appropriate sub block or block's mv
+// on whether the block_size < 8x8 and we have check_sub_blocks set.
+static INLINE int_mv get_sub_block_mv(const MODE_INFO *candidate, int which_mv,
+                                      int search_col, int block_idx) {
+  return block_idx >= 0 && candidate->sb_type < BLOCK_8X8
+          ? candidate->bmi[idx_n_column_to_subblock[block_idx][search_col == 0]]
+              .as_mv[which_mv]
+          : candidate->mv[which_mv];
+}
+
+
+// Performs mv sign inversion if indicated by the reference frame combination.
+static INLINE int_mv scale_mv(const MODE_INFO *mi, int ref,
+                              const MV_REFERENCE_FRAME this_ref_frame,
+                              const int *ref_sign_bias) {
+  int_mv mv = mi->mv[ref];
+  if (ref_sign_bias[mi->ref_frame[ref]] != ref_sign_bias[this_ref_frame]) {
+    mv.as_mv.row *= -1;
+    mv.as_mv.col *= -1;
+  }
+  return mv;
+}
+
+// This macro is used to add a motion vector mv_ref list if it isn't
+// already in the list.  If it's the second motion vector it will also
+// skip all additional processing and jump to Done!
+#define ADD_MV_REF_LIST(mv, refmv_count, mv_ref_list, Done) \
+  do { \
+    if (refmv_count) { \
+      if ((mv).as_int != (mv_ref_list)[0].as_int) { \
+        (mv_ref_list)[(refmv_count)] = (mv); \
+        goto Done; \
+      } \
+    } else { \
+      (mv_ref_list)[(refmv_count)++] = (mv); \
+    } \
+  } while (0)
+
+// If either reference frame is different, not INTRA, and they
+// are different from each other scale and add the mv to our list.
+#define IF_DIFF_REF_FRAME_ADD_MV(mbmi, ref_frame, ref_sign_bias, refmv_count, \
+                                 mv_ref_list, Done) \
+  do { \
+    if (is_inter_block(mbmi)) { \
+      if ((mbmi)->ref_frame[0] != ref_frame) \
+        ADD_MV_REF_LIST(scale_mv((mbmi), 0, ref_frame, ref_sign_bias), \
+                        refmv_count, mv_ref_list, Done); \
+      if (has_second_ref(mbmi) && \
+          (mbmi)->ref_frame[1] != ref_frame && \
+          (mbmi)->mv[1].as_int != (mbmi)->mv[0].as_int) \
+        ADD_MV_REF_LIST(scale_mv((mbmi), 1, ref_frame, ref_sign_bias), \
+                        refmv_count, mv_ref_list, Done); \
+    } \
+  } while (0)
+
+
+// Checks that the given mi_row, mi_col and search point
+// are inside the borders of the tile.
+static INLINE int is_inside(const TileInfo *const tile,
+                            int mi_col, int mi_row, int mi_rows,
+                            const POSITION *mi_pos) {
+  return !(mi_row + mi_pos->row < 0 ||
+           mi_col + mi_pos->col < tile->mi_col_start ||
+           mi_row + mi_pos->row >= mi_rows ||
+           mi_col + mi_pos->col >= tile->mi_col_end);
+}
+
+// TODO(jingning): this mv clamping function should be block size dependent.
+static INLINE void clamp_mv2(MV *mv, const MACROBLOCKD *xd) {
+  clamp_mv(mv, xd->mb_to_left_edge - LEFT_TOP_MARGIN,
+               xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN,
+               xd->mb_to_top_edge - LEFT_TOP_MARGIN,
+               xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN);
+}
+
+static INLINE void lower_mv_precision(MV *mv, int allow_hp) {
+  const int use_hp = allow_hp && use_mv_hp(mv);
+  if (!use_hp) {
+    if (mv->row & 1)
+      mv->row += (mv->row > 0 ? -1 : 1);
+    if (mv->col & 1)
+      mv->col += (mv->col > 0 ? -1 : 1);
+  }
+}
+
+typedef void (*find_mv_refs_sync)(void *const data, int mi_row);
+void vp9_find_mv_refs(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                      MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                      int_mv *mv_ref_list, int mi_row, int mi_col,
+                      uint8_t *mode_context);
+
+// check a list of motion vectors by sad score using a number rows of pixels
+// above and a number cols of pixels in the left to select the one with best
+// score to use as ref motion vector
+void vp9_find_best_ref_mvs(MACROBLOCKD *xd, int allow_hp,
+                           int_mv *mvlist, int_mv *nearest_mv, int_mv *near_mv);
+
+void vp9_append_sub8x8_mvs_for_idx(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                   int block, int ref, int mi_row, int mi_col,
+                                   int_mv *nearest_mv, int_mv *near_mv,
+                                   uint8_t *mode_context);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_MVREF_COMMON_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_onyxc_int.h b/libvpx/libvpx/vp9/common/vp9_onyxc_int.h
new file mode 100644
index 0000000..3fd935e
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_onyxc_int.h
@@ -0,0 +1,446 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_ONYXC_INT_H_
+#define VP9_COMMON_VP9_ONYXC_INT_H_
+
+#include "./vpx_config.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "vpx_util/vpx_thread.h"
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_frame_buffers.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_tile_common.h"
+
+#if CONFIG_VP9_POSTPROC
+#include "vp9/common/vp9_postproc.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define REFS_PER_FRAME 3
+
+#define REF_FRAMES_LOG2 3
+#define REF_FRAMES (1 << REF_FRAMES_LOG2)
+
+// 4 scratch frames for the new frames to support a maximum of 4 cores decoding
+// in parallel, 3 for scaled references on the encoder.
+// TODO(hkuang): Add ondemand frame buffers instead of hardcoding the number
+// of framebuffers.
+// TODO(jkoleszar): These 3 extra references could probably come from the
+// normal reference pool.
+#define FRAME_BUFFERS (REF_FRAMES + 7)
+
+#define FRAME_CONTEXTS_LOG2 2
+#define FRAME_CONTEXTS (1 << FRAME_CONTEXTS_LOG2)
+
+#define NUM_PING_PONG_BUFFERS 2
+
+extern const struct {
+  PARTITION_CONTEXT above;
+  PARTITION_CONTEXT left;
+} partition_context_lookup[BLOCK_SIZES];
+
+
+typedef enum {
+  SINGLE_REFERENCE      = 0,
+  COMPOUND_REFERENCE    = 1,
+  REFERENCE_MODE_SELECT = 2,
+  REFERENCE_MODES       = 3,
+} REFERENCE_MODE;
+
+typedef struct {
+  int_mv mv[2];
+  MV_REFERENCE_FRAME ref_frame[2];
+} MV_REF;
+
+typedef struct {
+  int ref_count;
+  MV_REF *mvs;
+  int mi_rows;
+  int mi_cols;
+  vpx_codec_frame_buffer_t raw_frame_buffer;
+  YV12_BUFFER_CONFIG buf;
+
+  // The Following variables will only be used in frame parallel decode.
+
+  // frame_worker_owner indicates which FrameWorker owns this buffer. NULL means
+  // that no FrameWorker owns, or is decoding, this buffer.
+  VPxWorker *frame_worker_owner;
+
+  // row and col indicate which position frame has been decoded to in real
+  // pixel unit. They are reset to -1 when decoding begins and set to INT_MAX
+  // when the frame is fully decoded.
+  int row;
+  int col;
+} RefCntBuffer;
+
+typedef struct BufferPool {
+  // Protect BufferPool from being accessed by several FrameWorkers at
+  // the same time during frame parallel decode.
+  // TODO(hkuang): Try to use atomic variable instead of locking the whole pool.
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t pool_mutex;
+#endif
+
+  // Private data associated with the frame buffer callbacks.
+  void *cb_priv;
+
+  vpx_get_frame_buffer_cb_fn_t get_fb_cb;
+  vpx_release_frame_buffer_cb_fn_t release_fb_cb;
+
+  RefCntBuffer frame_bufs[FRAME_BUFFERS];
+
+  // Frame buffers allocated internally by the codec.
+  InternalFrameBufferList int_frame_buffers;
+} BufferPool;
+
+typedef struct VP9Common {
+  struct vpx_internal_error_info  error;
+  vpx_color_space_t color_space;
+  vpx_color_range_t color_range;
+  int width;
+  int height;
+  int render_width;
+  int render_height;
+  int last_width;
+  int last_height;
+
+  // TODO(jkoleszar): this implies chroma ss right now, but could vary per
+  // plane. Revisit as part of the future change to YV12_BUFFER_CONFIG to
+  // support additional planes.
+  int subsampling_x;
+  int subsampling_y;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  int use_highbitdepth;  // Marks if we need to use 16bit frame buffers.
+#endif
+
+  YV12_BUFFER_CONFIG *frame_to_show;
+  RefCntBuffer *prev_frame;
+
+  // TODO(hkuang): Combine this with cur_buf in macroblockd.
+  RefCntBuffer *cur_frame;
+
+  int ref_frame_map[REF_FRAMES]; /* maps fb_idx to reference slot */
+
+  // Prepare ref_frame_map for the next frame.
+  // Only used in frame parallel decode.
+  int next_ref_frame_map[REF_FRAMES];
+
+  // TODO(jkoleszar): could expand active_ref_idx to 4, with 0 as intra, and
+  // roll new_fb_idx into it.
+
+  // Each frame can reference REFS_PER_FRAME buffers
+  RefBuffer frame_refs[REFS_PER_FRAME];
+
+  int new_fb_idx;
+
+#if CONFIG_VP9_POSTPROC
+  YV12_BUFFER_CONFIG post_proc_buffer;
+  YV12_BUFFER_CONFIG post_proc_buffer_int;
+#endif
+
+  FRAME_TYPE last_frame_type;  /* last frame's frame type for motion search.*/
+  FRAME_TYPE frame_type;
+
+  int show_frame;
+  int last_show_frame;
+  int show_existing_frame;
+
+  // Flag signaling that the frame is encoded using only INTRA modes.
+  uint8_t intra_only;
+  uint8_t last_intra_only;
+
+  int allow_high_precision_mv;
+
+  // Flag signaling that the frame context should be reset to default values.
+  // 0 or 1 implies don't reset, 2 reset just the context specified in the
+  // frame header, 3 reset all contexts.
+  int reset_frame_context;
+
+  // MBs, mb_rows/cols is in 16-pixel units; mi_rows/cols is in
+  // MODE_INFO (8-pixel) units.
+  int MBs;
+  int mb_rows, mi_rows;
+  int mb_cols, mi_cols;
+  int mi_stride;
+
+  /* profile settings */
+  TX_MODE tx_mode;
+
+  int base_qindex;
+  int y_dc_delta_q;
+  int uv_dc_delta_q;
+  int uv_ac_delta_q;
+  int16_t y_dequant[MAX_SEGMENTS][2];
+  int16_t uv_dequant[MAX_SEGMENTS][2];
+
+  /* We allocate a MODE_INFO struct for each macroblock, together with
+     an extra row on top and column on the left to simplify prediction. */
+  int mi_alloc_size;
+  MODE_INFO *mip; /* Base of allocated array */
+  MODE_INFO *mi;  /* Corresponds to upper left visible macroblock */
+
+  // TODO(agrange): Move prev_mi into encoder structure.
+  // prev_mip and prev_mi will only be allocated in VP9 encoder.
+  MODE_INFO *prev_mip; /* MODE_INFO array 'mip' from last decoded frame */
+  MODE_INFO *prev_mi;  /* 'mi' from last frame (points into prev_mip) */
+
+  // Separate mi functions between encoder and decoder.
+  int (*alloc_mi)(struct VP9Common *cm, int mi_size);
+  void (*free_mi)(struct VP9Common *cm);
+  void (*setup_mi)(struct VP9Common *cm);
+
+  // Grid of pointers to 8x8 MODE_INFO structs.  Any 8x8 not in the visible
+  // area will be NULL.
+  MODE_INFO **mi_grid_base;
+  MODE_INFO **mi_grid_visible;
+  MODE_INFO **prev_mi_grid_base;
+  MODE_INFO **prev_mi_grid_visible;
+
+  // Whether to use previous frame's motion vectors for prediction.
+  int use_prev_frame_mvs;
+
+  // Persistent mb segment id map used in prediction.
+  int seg_map_idx;
+  int prev_seg_map_idx;
+
+  uint8_t *seg_map_array[NUM_PING_PONG_BUFFERS];
+  uint8_t *last_frame_seg_map;
+  uint8_t *current_frame_seg_map;
+  int seg_map_alloc_size;
+
+  INTERP_FILTER interp_filter;
+
+  loop_filter_info_n lf_info;
+
+  int refresh_frame_context;    /* Two state 0 = NO, 1 = YES */
+
+  int ref_frame_sign_bias[MAX_REF_FRAMES];    /* Two state 0, 1 */
+
+  struct loopfilter lf;
+  struct segmentation seg;
+
+  // TODO(hkuang): Remove this as it is the same as frame_parallel_decode
+  // in pbi.
+  int frame_parallel_decode;  // frame-based threading.
+
+  // Context probabilities for reference frame prediction
+  MV_REFERENCE_FRAME comp_fixed_ref;
+  MV_REFERENCE_FRAME comp_var_ref[2];
+  REFERENCE_MODE reference_mode;
+
+  FRAME_CONTEXT *fc;  /* this frame entropy */
+  FRAME_CONTEXT *frame_contexts;   // FRAME_CONTEXTS
+  unsigned int  frame_context_idx; /* Context to use/update */
+  FRAME_COUNTS counts;
+
+  unsigned int current_video_frame;
+  BITSTREAM_PROFILE profile;
+
+  // VPX_BITS_8 in profile 0 or 1, VPX_BITS_10 or VPX_BITS_12 in profile 2 or 3.
+  vpx_bit_depth_t bit_depth;
+  vpx_bit_depth_t dequant_bit_depth;  // bit_depth of current dequantizer
+
+#if CONFIG_VP9_POSTPROC
+  struct postproc_state  postproc_state;
+#endif
+
+  int error_resilient_mode;
+  int frame_parallel_decoding_mode;
+
+  int log2_tile_cols, log2_tile_rows;
+  int byte_alignment;
+  int skip_loop_filter;
+
+  // Private data associated with the frame buffer callbacks.
+  void *cb_priv;
+  vpx_get_frame_buffer_cb_fn_t get_fb_cb;
+  vpx_release_frame_buffer_cb_fn_t release_fb_cb;
+
+  // Handles memory for the codec.
+  InternalFrameBufferList int_frame_buffers;
+
+  // External BufferPool passed from outside.
+  BufferPool *buffer_pool;
+
+  PARTITION_CONTEXT *above_seg_context;
+  ENTROPY_CONTEXT *above_context;
+  int above_context_alloc_cols;
+} VP9_COMMON;
+
+// TODO(hkuang): Don't need to lock the whole pool after implementing atomic
+// frame reference count.
+void lock_buffer_pool(BufferPool *const pool);
+void unlock_buffer_pool(BufferPool *const pool);
+
+static INLINE YV12_BUFFER_CONFIG *get_ref_frame(VP9_COMMON *cm, int index) {
+  if (index < 0 || index >= REF_FRAMES)
+    return NULL;
+  if (cm->ref_frame_map[index] < 0)
+    return NULL;
+  assert(cm->ref_frame_map[index] < FRAME_BUFFERS);
+  return &cm->buffer_pool->frame_bufs[cm->ref_frame_map[index]].buf;
+}
+
+static INLINE YV12_BUFFER_CONFIG *get_frame_new_buffer(VP9_COMMON *cm) {
+  return &cm->buffer_pool->frame_bufs[cm->new_fb_idx].buf;
+}
+
+static INLINE int get_free_fb(VP9_COMMON *cm) {
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+  int i;
+
+  lock_buffer_pool(cm->buffer_pool);
+  for (i = 0; i < FRAME_BUFFERS; ++i)
+    if (frame_bufs[i].ref_count == 0)
+      break;
+
+  if (i != FRAME_BUFFERS) {
+    frame_bufs[i].ref_count = 1;
+  } else {
+    // Reset i to be INVALID_IDX to indicate no free buffer found.
+    i = INVALID_IDX;
+  }
+
+  unlock_buffer_pool(cm->buffer_pool);
+  return i;
+}
+
+static INLINE void ref_cnt_fb(RefCntBuffer *bufs, int *idx, int new_idx) {
+  const int ref_index = *idx;
+
+  if (ref_index >= 0 && bufs[ref_index].ref_count > 0)
+    bufs[ref_index].ref_count--;
+
+  *idx = new_idx;
+
+  bufs[new_idx].ref_count++;
+}
+
+static INLINE int mi_cols_aligned_to_sb(int n_mis) {
+  return ALIGN_POWER_OF_TWO(n_mis, MI_BLOCK_SIZE_LOG2);
+}
+
+static INLINE int frame_is_intra_only(const VP9_COMMON *const cm) {
+  return cm->frame_type == KEY_FRAME || cm->intra_only;
+}
+
+static INLINE void set_partition_probs(const VP9_COMMON *const cm,
+                                       MACROBLOCKD *const xd) {
+  xd->partition_probs =
+      frame_is_intra_only(cm) ?
+          &vp9_kf_partition_probs[0] :
+          (const vpx_prob (*)[PARTITION_TYPES - 1])cm->fc->partition_prob;
+}
+
+static INLINE void vp9_init_macroblockd(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                        tran_low_t *dqcoeff) {
+  int i;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    xd->plane[i].dqcoeff = dqcoeff;
+    xd->above_context[i] = cm->above_context +
+        i * sizeof(*cm->above_context) * 2 * mi_cols_aligned_to_sb(cm->mi_cols);
+
+    if (get_plane_type(i) == PLANE_TYPE_Y) {
+      memcpy(xd->plane[i].seg_dequant, cm->y_dequant, sizeof(cm->y_dequant));
+    } else {
+      memcpy(xd->plane[i].seg_dequant, cm->uv_dequant, sizeof(cm->uv_dequant));
+    }
+    xd->fc = cm->fc;
+  }
+
+  xd->above_seg_context = cm->above_seg_context;
+  xd->mi_stride = cm->mi_stride;
+  xd->error_info = &cm->error;
+
+  set_partition_probs(cm, xd);
+}
+
+static INLINE const vpx_prob* get_partition_probs(const MACROBLOCKD *xd,
+                                                  int ctx) {
+  return xd->partition_probs[ctx];
+}
+
+static INLINE void set_skip_context(MACROBLOCKD *xd, int mi_row, int mi_col) {
+  const int above_idx = mi_col * 2;
+  const int left_idx = (mi_row * 2) & 15;
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    struct macroblockd_plane *const pd = &xd->plane[i];
+    pd->above_context = &xd->above_context[i][above_idx >> pd->subsampling_x];
+    pd->left_context = &xd->left_context[i][left_idx >> pd->subsampling_y];
+  }
+}
+
+static INLINE int calc_mi_size(int len) {
+  // len is in mi units.
+  return len + MI_BLOCK_SIZE;
+}
+
+static INLINE void set_mi_row_col(MACROBLOCKD *xd, const TileInfo *const tile,
+                                  int mi_row, int bh,
+                                  int mi_col, int bw,
+                                  int mi_rows, int mi_cols) {
+  xd->mb_to_top_edge    = -((mi_row * MI_SIZE) * 8);
+  xd->mb_to_bottom_edge = ((mi_rows - bh - mi_row) * MI_SIZE) * 8;
+  xd->mb_to_left_edge   = -((mi_col * MI_SIZE) * 8);
+  xd->mb_to_right_edge  = ((mi_cols - bw - mi_col) * MI_SIZE) * 8;
+
+  // Are edges available for intra prediction?
+  xd->above_mi = (mi_row != 0) ? xd->mi[-xd->mi_stride] : NULL;
+  xd->left_mi  = (mi_col > tile->mi_col_start) ? xd->mi[-1] : NULL;
+}
+
+static INLINE void update_partition_context(MACROBLOCKD *xd,
+                                            int mi_row, int mi_col,
+                                            BLOCK_SIZE subsize,
+                                            BLOCK_SIZE bsize) {
+  PARTITION_CONTEXT *const above_ctx = xd->above_seg_context + mi_col;
+  PARTITION_CONTEXT *const left_ctx = xd->left_seg_context + (mi_row & MI_MASK);
+
+  // num_4x4_blocks_wide_lookup[bsize] / 2
+  const int bs = num_8x8_blocks_wide_lookup[bsize];
+
+  // update the partition context at the end notes. set partition bits
+  // of block sizes larger than the current one to be one, and partition
+  // bits of smaller block sizes to be zero.
+  memset(above_ctx, partition_context_lookup[subsize].above, bs);
+  memset(left_ctx, partition_context_lookup[subsize].left, bs);
+}
+
+static INLINE int partition_plane_context(const MACROBLOCKD *xd,
+                                          int mi_row, int mi_col,
+                                          BLOCK_SIZE bsize) {
+  const PARTITION_CONTEXT *above_ctx = xd->above_seg_context + mi_col;
+  const PARTITION_CONTEXT *left_ctx = xd->left_seg_context + (mi_row & MI_MASK);
+  const int bsl = mi_width_log2_lookup[bsize];
+  int above = (*above_ctx >> bsl) & 1 , left = (*left_ctx >> bsl) & 1;
+
+  assert(b_width_log2_lookup[bsize] == b_height_log2_lookup[bsize]);
+  assert(bsl >= 0);
+
+  return (left * 2 + above) + bsl * PARTITION_PLOFFSET;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ONYXC_INT_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_postproc.c b/libvpx/libvpx/vp9/common/vp9_postproc.c
new file mode 100644
index 0000000..c04cc8f
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_postproc.c
@@ -0,0 +1,720 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+#include "./vpx_scale_rtcd.h"
+#include "./vp9_rtcd.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/system_state.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vpx_scale/yv12config.h"
+
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_postproc.h"
+#include "vp9/common/vp9_textblit.h"
+
+#if CONFIG_VP9_POSTPROC
+static const int16_t kernel5[] = {
+  1, 1, 4, 1, 1
+};
+
+const int16_t vp9_rv[] = {
+  8, 5, 2, 2, 8, 12, 4, 9, 8, 3,
+  0, 3, 9, 0, 0, 0, 8, 3, 14, 4,
+  10, 1, 11, 14, 1, 14, 9, 6, 12, 11,
+  8, 6, 10, 0, 0, 8, 9, 0, 3, 14,
+  8, 11, 13, 4, 2, 9, 0, 3, 9, 6,
+  1, 2, 3, 14, 13, 1, 8, 2, 9, 7,
+  3, 3, 1, 13, 13, 6, 6, 5, 2, 7,
+  11, 9, 11, 8, 7, 3, 2, 0, 13, 13,
+  14, 4, 12, 5, 12, 10, 8, 10, 13, 10,
+  4, 14, 4, 10, 0, 8, 11, 1, 13, 7,
+  7, 14, 6, 14, 13, 2, 13, 5, 4, 4,
+  0, 10, 0, 5, 13, 2, 12, 7, 11, 13,
+  8, 0, 4, 10, 7, 2, 7, 2, 2, 5,
+  3, 4, 7, 3, 3, 14, 14, 5, 9, 13,
+  3, 14, 3, 6, 3, 0, 11, 8, 13, 1,
+  13, 1, 12, 0, 10, 9, 7, 6, 2, 8,
+  5, 2, 13, 7, 1, 13, 14, 7, 6, 7,
+  9, 6, 10, 11, 7, 8, 7, 5, 14, 8,
+  4, 4, 0, 8, 7, 10, 0, 8, 14, 11,
+  3, 12, 5, 7, 14, 3, 14, 5, 2, 6,
+  11, 12, 12, 8, 0, 11, 13, 1, 2, 0,
+  5, 10, 14, 7, 8, 0, 4, 11, 0, 8,
+  0, 3, 10, 5, 8, 0, 11, 6, 7, 8,
+  10, 7, 13, 9, 2, 5, 1, 5, 10, 2,
+  4, 3, 5, 6, 10, 8, 9, 4, 11, 14,
+  0, 10, 0, 5, 13, 2, 12, 7, 11, 13,
+  8, 0, 4, 10, 7, 2, 7, 2, 2, 5,
+  3, 4, 7, 3, 3, 14, 14, 5, 9, 13,
+  3, 14, 3, 6, 3, 0, 11, 8, 13, 1,
+  13, 1, 12, 0, 10, 9, 7, 6, 2, 8,
+  5, 2, 13, 7, 1, 13, 14, 7, 6, 7,
+  9, 6, 10, 11, 7, 8, 7, 5, 14, 8,
+  4, 4, 0, 8, 7, 10, 0, 8, 14, 11,
+  3, 12, 5, 7, 14, 3, 14, 5, 2, 6,
+  11, 12, 12, 8, 0, 11, 13, 1, 2, 0,
+  5, 10, 14, 7, 8, 0, 4, 11, 0, 8,
+  0, 3, 10, 5, 8, 0, 11, 6, 7, 8,
+  10, 7, 13, 9, 2, 5, 1, 5, 10, 2,
+  4, 3, 5, 6, 10, 8, 9, 4, 11, 14,
+  3, 8, 3, 7, 8, 5, 11, 4, 12, 3,
+  11, 9, 14, 8, 14, 13, 4, 3, 1, 2,
+  14, 6, 5, 4, 4, 11, 4, 6, 2, 1,
+  5, 8, 8, 12, 13, 5, 14, 10, 12, 13,
+  0, 9, 5, 5, 11, 10, 13, 9, 10, 13,
+};
+
+static const uint8_t q_diff_thresh = 20;
+static const uint8_t last_q_thresh = 170;
+
+void vp9_post_proc_down_and_across_c(const uint8_t *src_ptr,
+                                     uint8_t *dst_ptr,
+                                     int src_pixels_per_line,
+                                     int dst_pixels_per_line,
+                                     int rows,
+                                     int cols,
+                                     int flimit) {
+  uint8_t const *p_src;
+  uint8_t *p_dst;
+  int row, col, i, v, kernel;
+  int pitch = src_pixels_per_line;
+  uint8_t d[8];
+  (void)dst_pixels_per_line;
+
+  for (row = 0; row < rows; row++) {
+    /* post_proc_down for one row */
+    p_src = src_ptr;
+    p_dst = dst_ptr;
+
+    for (col = 0; col < cols; col++) {
+      kernel = 4;
+      v = p_src[col];
+
+      for (i = -2; i <= 2; i++) {
+        if (abs(v - p_src[col + i * pitch]) > flimit)
+          goto down_skip_convolve;
+
+        kernel += kernel5[2 + i] * p_src[col + i * pitch];
+      }
+
+      v = (kernel >> 3);
+    down_skip_convolve:
+      p_dst[col] = v;
+    }
+
+    /* now post_proc_across */
+    p_src = dst_ptr;
+    p_dst = dst_ptr;
+
+    for (i = 0; i < 8; i++)
+      d[i] = p_src[i];
+
+    for (col = 0; col < cols; col++) {
+      kernel = 4;
+      v = p_src[col];
+
+      d[col & 7] = v;
+
+      for (i = -2; i <= 2; i++) {
+        if (abs(v - p_src[col + i]) > flimit)
+          goto across_skip_convolve;
+
+        kernel += kernel5[2 + i] * p_src[col + i];
+      }
+
+      d[col & 7] = (kernel >> 3);
+    across_skip_convolve:
+
+      if (col >= 2)
+        p_dst[col - 2] = d[(col - 2) & 7];
+    }
+
+    /* handle the last two pixels */
+    p_dst[col - 2] = d[(col - 2) & 7];
+    p_dst[col - 1] = d[(col - 1) & 7];
+
+
+    /* next row */
+    src_ptr += pitch;
+    dst_ptr += pitch;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_post_proc_down_and_across_c(const uint16_t *src_ptr,
+                                            uint16_t *dst_ptr,
+                                            int src_pixels_per_line,
+                                            int dst_pixels_per_line,
+                                            int rows,
+                                            int cols,
+                                            int flimit) {
+  uint16_t const *p_src;
+  uint16_t *p_dst;
+  int row, col, i, v, kernel;
+  int pitch = src_pixels_per_line;
+  uint16_t d[8];
+
+  for (row = 0; row < rows; row++) {
+    // post_proc_down for one row.
+    p_src = src_ptr;
+    p_dst = dst_ptr;
+
+    for (col = 0; col < cols; col++) {
+      kernel = 4;
+      v = p_src[col];
+
+      for (i = -2; i <= 2; i++) {
+        if (abs(v - p_src[col + i * pitch]) > flimit)
+          goto down_skip_convolve;
+
+        kernel += kernel5[2 + i] * p_src[col + i * pitch];
+      }
+
+      v = (kernel >> 3);
+
+    down_skip_convolve:
+      p_dst[col] = v;
+    }
+
+    /* now post_proc_across */
+    p_src = dst_ptr;
+    p_dst = dst_ptr;
+
+    for (i = 0; i < 8; i++)
+      d[i] = p_src[i];
+
+    for (col = 0; col < cols; col++) {
+      kernel = 4;
+      v = p_src[col];
+
+      d[col & 7] = v;
+
+      for (i = -2; i <= 2; i++) {
+        if (abs(v - p_src[col + i]) > flimit)
+          goto across_skip_convolve;
+
+        kernel += kernel5[2 + i] * p_src[col + i];
+      }
+
+      d[col & 7] = (kernel >> 3);
+
+    across_skip_convolve:
+      if (col >= 2)
+        p_dst[col - 2] = d[(col - 2) & 7];
+    }
+
+    /* handle the last two pixels */
+    p_dst[col - 2] = d[(col - 2) & 7];
+    p_dst[col - 1] = d[(col - 1) & 7];
+
+
+    /* next row */
+    src_ptr += pitch;
+    dst_ptr += dst_pixels_per_line;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static int q2mbl(int x) {
+  if (x < 20) x = 20;
+
+  x = 50 + (x - 50) * 10 / 8;
+  return x * x / 3;
+}
+
+void vp9_mbpost_proc_across_ip_c(uint8_t *src, int pitch,
+                                 int rows, int cols, int flimit) {
+  int r, c, i;
+  uint8_t *s = src;
+  uint8_t d[16];
+
+  for (r = 0; r < rows; r++) {
+    int sumsq = 0;
+    int sum = 0;
+
+    for (i = -8; i <= 6; i++) {
+      sumsq += s[i] * s[i];
+      sum += s[i];
+      d[i + 8] = 0;
+    }
+
+    for (c = 0; c < cols + 8; c++) {
+      int x = s[c + 7] - s[c - 8];
+      int y = s[c + 7] + s[c - 8];
+
+      sum += x;
+      sumsq += x * y;
+
+      d[c & 15] = s[c];
+
+      if (sumsq * 15 - sum * sum < flimit) {
+        d[c & 15] = (8 + sum + s[c]) >> 4;
+      }
+
+      s[c - 8] = d[(c - 8) & 15];
+    }
+    s += pitch;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_mbpost_proc_across_ip_c(uint16_t *src, int pitch,
+                                        int rows, int cols, int flimit) {
+  int r, c, i;
+
+  uint16_t *s = src;
+  uint16_t d[16];
+
+
+  for (r = 0; r < rows; r++) {
+    int sumsq = 0;
+    int sum   = 0;
+
+    for (i = -8; i <= 6; i++) {
+      sumsq += s[i] * s[i];
+      sum   += s[i];
+      d[i + 8] = 0;
+    }
+
+    for (c = 0; c < cols + 8; c++) {
+      int x = s[c + 7] - s[c - 8];
+      int y = s[c + 7] + s[c - 8];
+
+      sum  += x;
+      sumsq += x * y;
+
+      d[c & 15] = s[c];
+
+      if (sumsq * 15 - sum * sum < flimit) {
+        d[c & 15] = (8 + sum + s[c]) >> 4;
+      }
+
+      s[c - 8] = d[(c - 8) & 15];
+    }
+
+    s += pitch;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_mbpost_proc_down_c(uint8_t *dst, int pitch,
+                            int rows, int cols, int flimit) {
+  int r, c, i;
+  const short *rv3 = &vp9_rv[63 & rand()]; // NOLINT
+
+  for (c = 0; c < cols; c++) {
+    uint8_t *s = &dst[c];
+    int sumsq = 0;
+    int sum   = 0;
+    uint8_t d[16];
+    const int16_t *rv2 = rv3 + ((c * 17) & 127);
+
+    for (i = -8; i <= 6; i++) {
+      sumsq += s[i * pitch] * s[i * pitch];
+      sum   += s[i * pitch];
+    }
+
+    for (r = 0; r < rows + 8; r++) {
+      sumsq += s[7 * pitch] * s[ 7 * pitch] - s[-8 * pitch] * s[-8 * pitch];
+      sum  += s[7 * pitch] - s[-8 * pitch];
+      d[r & 15] = s[0];
+
+      if (sumsq * 15 - sum * sum < flimit) {
+        d[r & 15] = (rv2[r & 127] + sum + s[0]) >> 4;
+      }
+
+      s[-8 * pitch] = d[(r - 8) & 15];
+      s += pitch;
+    }
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_mbpost_proc_down_c(uint16_t *dst, int pitch,
+                                   int rows, int cols, int flimit) {
+  int r, c, i;
+  const int16_t *rv3 = &vp9_rv[63 & rand()];  // NOLINT
+
+  for (c = 0; c < cols; c++) {
+    uint16_t *s = &dst[c];
+    int sumsq = 0;
+    int sum = 0;
+    uint16_t d[16];
+    const int16_t *rv2 = rv3 + ((c * 17) & 127);
+
+    for (i = -8; i <= 6; i++) {
+      sumsq += s[i * pitch] * s[i * pitch];
+      sum += s[i * pitch];
+    }
+
+    for (r = 0; r < rows + 8; r++) {
+      sumsq += s[7 * pitch] * s[ 7 * pitch] - s[-8 * pitch] * s[-8 * pitch];
+      sum += s[7 * pitch] - s[-8 * pitch];
+      d[r & 15] = s[0];
+
+      if (sumsq * 15 - sum * sum < flimit) {
+        d[r & 15] = (rv2[r & 127] + sum + s[0]) >> 4;
+      }
+
+      s[-8 * pitch] = d[(r - 8) & 15];
+      s += pitch;
+    }
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void deblock_and_de_macro_block(YV12_BUFFER_CONFIG   *source,
+                                       YV12_BUFFER_CONFIG   *post,
+                                       int                   q,
+                                       int                   low_var_thresh,
+                                       int                   flag) {
+  double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
+  int ppl = (int)(level + .5);
+  (void) low_var_thresh;
+  (void) flag;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (source->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vp9_highbd_post_proc_down_and_across(CONVERT_TO_SHORTPTR(source->y_buffer),
+                                         CONVERT_TO_SHORTPTR(post->y_buffer),
+                                         source->y_stride, post->y_stride,
+                                         source->y_height, source->y_width,
+                                         ppl);
+
+    vp9_highbd_mbpost_proc_across_ip(CONVERT_TO_SHORTPTR(post->y_buffer),
+                                     post->y_stride, post->y_height,
+                                     post->y_width, q2mbl(q));
+
+    vp9_highbd_mbpost_proc_down(CONVERT_TO_SHORTPTR(post->y_buffer),
+                                post->y_stride, post->y_height,
+                                post->y_width, q2mbl(q));
+
+    vp9_highbd_post_proc_down_and_across(CONVERT_TO_SHORTPTR(source->u_buffer),
+                                         CONVERT_TO_SHORTPTR(post->u_buffer),
+                                         source->uv_stride, post->uv_stride,
+                                         source->uv_height, source->uv_width,
+                                         ppl);
+    vp9_highbd_post_proc_down_and_across(CONVERT_TO_SHORTPTR(source->v_buffer),
+                                         CONVERT_TO_SHORTPTR(post->v_buffer),
+                                         source->uv_stride, post->uv_stride,
+                                         source->uv_height, source->uv_width,
+                                         ppl);
+  } else {
+    vp9_post_proc_down_and_across(source->y_buffer, post->y_buffer,
+                                  source->y_stride, post->y_stride,
+                                  source->y_height, source->y_width, ppl);
+
+    vp9_mbpost_proc_across_ip(post->y_buffer, post->y_stride, post->y_height,
+                              post->y_width, q2mbl(q));
+
+    vp9_mbpost_proc_down(post->y_buffer, post->y_stride, post->y_height,
+                         post->y_width, q2mbl(q));
+
+    vp9_post_proc_down_and_across(source->u_buffer, post->u_buffer,
+                                  source->uv_stride, post->uv_stride,
+                                  source->uv_height, source->uv_width, ppl);
+    vp9_post_proc_down_and_across(source->v_buffer, post->v_buffer,
+                                  source->uv_stride, post->uv_stride,
+                                  source->uv_height, source->uv_width, ppl);
+  }
+#else
+  vp9_post_proc_down_and_across(source->y_buffer, post->y_buffer,
+                                source->y_stride, post->y_stride,
+                                source->y_height, source->y_width, ppl);
+
+  vp9_mbpost_proc_across_ip(post->y_buffer, post->y_stride, post->y_height,
+                            post->y_width, q2mbl(q));
+
+  vp9_mbpost_proc_down(post->y_buffer, post->y_stride, post->y_height,
+                       post->y_width, q2mbl(q));
+
+  vp9_post_proc_down_and_across(source->u_buffer, post->u_buffer,
+                                source->uv_stride, post->uv_stride,
+                                source->uv_height, source->uv_width, ppl);
+  vp9_post_proc_down_and_across(source->v_buffer, post->v_buffer,
+                                source->uv_stride, post->uv_stride,
+                                source->uv_height, source->uv_width, ppl);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+void vp9_deblock(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst,
+                 int q) {
+  const int ppl = (int)(6.0e-05 * q * q * q - 0.0067 * q * q + 0.306 * q
+                        + 0.0065 + 0.5);
+  int i;
+
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  const int src_widths[3] = {src->y_width, src->uv_width, src->uv_width};
+  const int src_heights[3] = {src->y_height, src->uv_height, src->uv_height};
+
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    assert((src->flags & YV12_FLAG_HIGHBITDEPTH) ==
+           (dst->flags & YV12_FLAG_HIGHBITDEPTH));
+    if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+      vp9_highbd_post_proc_down_and_across(CONVERT_TO_SHORTPTR(srcs[i]),
+                                           CONVERT_TO_SHORTPTR(dsts[i]),
+                                           src_strides[i], dst_strides[i],
+                                           src_heights[i], src_widths[i], ppl);
+    } else {
+      vp9_post_proc_down_and_across(srcs[i], dsts[i],
+                                    src_strides[i], dst_strides[i],
+                                    src_heights[i], src_widths[i], ppl);
+    }
+#else
+    vp9_post_proc_down_and_across(srcs[i], dsts[i],
+                                  src_strides[i], dst_strides[i],
+                                  src_heights[i], src_widths[i], ppl);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+}
+
+void vp9_denoise(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst,
+                 int q) {
+  const int ppl = (int)(6.0e-05 * q * q * q - 0.0067 * q * q + 0.306 * q
+                        + 0.0065 + 0.5);
+  int i;
+
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  const int src_widths[3] = {src->y_width, src->uv_width, src->uv_width};
+  const int src_heights[3] = {src->y_height, src->uv_height, src->uv_height};
+
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    const int src_stride = src_strides[i];
+    const int src_width = src_widths[i] - 4;
+    const int src_height = src_heights[i] - 4;
+    const int dst_stride = dst_strides[i];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+    assert((src->flags & YV12_FLAG_HIGHBITDEPTH) ==
+           (dst->flags & YV12_FLAG_HIGHBITDEPTH));
+    if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+      const uint16_t *const src_plane = CONVERT_TO_SHORTPTR(
+          srcs[i] + 2 * src_stride + 2);
+      uint16_t *const dst_plane = CONVERT_TO_SHORTPTR(
+          dsts[i] + 2 * dst_stride + 2);
+      vp9_highbd_post_proc_down_and_across(src_plane, dst_plane, src_stride,
+                                           dst_stride, src_height, src_width,
+                                           ppl);
+    } else {
+      const uint8_t *const src_plane = srcs[i] + 2 * src_stride + 2;
+      uint8_t *const dst_plane = dsts[i] + 2 * dst_stride + 2;
+
+      vp9_post_proc_down_and_across(src_plane, dst_plane, src_stride,
+                                    dst_stride, src_height, src_width, ppl);
+    }
+#else
+    const uint8_t *const src_plane = srcs[i] + 2 * src_stride + 2;
+    uint8_t *const dst_plane = dsts[i] + 2 * dst_stride + 2;
+    vp9_post_proc_down_and_across(src_plane, dst_plane, src_stride, dst_stride,
+                                  src_height, src_width, ppl);
+#endif
+  }
+}
+
+static double gaussian(double sigma, double mu, double x) {
+  return 1 / (sigma * sqrt(2.0 * 3.14159265)) *
+         (exp(-(x - mu) * (x - mu) / (2 * sigma * sigma)));
+}
+
+static void fillrd(struct postproc_state *state, int q, int a) {
+  char char_dist[300];
+
+  double sigma;
+  int ai = a, qi = q, i;
+
+  vpx_clear_system_state();
+
+  sigma = ai + .5 + .6 * (63 - qi) / 63.0;
+
+  /* set up a lookup table of 256 entries that matches
+   * a gaussian distribution with sigma determined by q.
+   */
+  {
+    int next, j;
+
+    next = 0;
+
+    for (i = -32; i < 32; i++) {
+      int a_i = (int)(0.5 + 256 * gaussian(sigma, 0, i));
+
+      if (a_i) {
+        for (j = 0; j < a_i; j++) {
+          char_dist[next + j] = (char) i;
+        }
+
+        next = next + j;
+      }
+    }
+
+    for (; next < 256; next++)
+      char_dist[next] = 0;
+  }
+
+  for (i = 0; i < 3072; i++) {
+    state->noise[i] = char_dist[rand() & 0xff];  // NOLINT
+  }
+
+  for (i = 0; i < 16; i++) {
+    state->blackclamp[i] = -char_dist[0];
+    state->whiteclamp[i] = -char_dist[0];
+    state->bothclamp[i] = -2 * char_dist[0];
+  }
+
+  state->last_q = q;
+  state->last_noise = a;
+}
+
+static void swap_mi_and_prev_mi(VP9_COMMON *cm) {
+  // Current mip will be the prev_mip for the next frame.
+  MODE_INFO *temp = cm->postproc_state.prev_mip;
+  cm->postproc_state.prev_mip = cm->mip;
+  cm->mip = temp;
+
+  // Update the upper left visible macroblock ptrs.
+  cm->mi = cm->mip + cm->mi_stride + 1;
+  cm->postproc_state.prev_mi = cm->postproc_state.prev_mip + cm->mi_stride + 1;
+}
+
+int vp9_post_proc_frame(struct VP9Common *cm,
+                        YV12_BUFFER_CONFIG *dest, vp9_ppflags_t *ppflags) {
+  const int q = VPXMIN(105, cm->lf.filter_level * 2);
+  const int flags = ppflags->post_proc_flag;
+  YV12_BUFFER_CONFIG *const ppbuf = &cm->post_proc_buffer;
+  struct postproc_state *const ppstate = &cm->postproc_state;
+
+  if (!cm->frame_to_show)
+    return -1;
+
+  if (!flags) {
+    *dest = *cm->frame_to_show;
+    return 0;
+  }
+
+  vpx_clear_system_state();
+
+  // Alloc memory for prev_mip in the first frame.
+  if (cm->current_video_frame == 1) {
+    cm->postproc_state.last_base_qindex = cm->base_qindex;
+    cm->postproc_state.last_frame_valid = 1;
+    ppstate->prev_mip = vpx_calloc(cm->mi_alloc_size, sizeof(*cm->mip));
+    if (!ppstate->prev_mip) {
+      return 1;
+    }
+    ppstate->prev_mi = ppstate->prev_mip + cm->mi_stride + 1;
+    memset(ppstate->prev_mip, 0,
+           cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mip));
+  }
+
+  // Allocate post_proc_buffer_int if needed.
+  if ((flags & VP9D_MFQE) && !cm->post_proc_buffer_int.buffer_alloc) {
+    if ((flags & VP9D_DEMACROBLOCK) || (flags & VP9D_DEBLOCK)) {
+      const int width = ALIGN_POWER_OF_TWO(cm->width, 4);
+      const int height = ALIGN_POWER_OF_TWO(cm->height, 4);
+
+      if (vpx_alloc_frame_buffer(&cm->post_proc_buffer_int, width, height,
+                                 cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 cm->use_highbitdepth,
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+                                 VP9_ENC_BORDER_IN_PIXELS,
+                                 cm->byte_alignment) < 0) {
+        vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate MFQE framebuffer");
+      }
+
+      // Ensure that postproc is set to all 0s so that post proc
+      // doesn't pull random data in from edge.
+      memset(cm->post_proc_buffer_int.buffer_alloc, 128,
+             cm->post_proc_buffer.frame_size);
+    }
+  }
+
+  if (vpx_realloc_frame_buffer(&cm->post_proc_buffer, cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               VP9_DEC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL) < 0)
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate post-processing buffer");
+
+  if ((flags & VP9D_MFQE) && cm->current_video_frame >= 2 &&
+      cm->postproc_state.last_frame_valid && cm->bit_depth == 8 &&
+      cm->postproc_state.last_base_qindex <= last_q_thresh &&
+      cm->base_qindex - cm->postproc_state.last_base_qindex >= q_diff_thresh) {
+    vp9_mfqe(cm);
+    // TODO(jackychen): Consider whether enable deblocking by default
+    // if mfqe is enabled. Need to take both the quality and the speed
+    // into consideration.
+    if ((flags & VP9D_DEMACROBLOCK) || (flags & VP9D_DEBLOCK)) {
+      vp8_yv12_copy_frame(ppbuf, &cm->post_proc_buffer_int);
+    }
+    if ((flags & VP9D_DEMACROBLOCK) && cm->post_proc_buffer_int.buffer_alloc) {
+      deblock_and_de_macro_block(&cm->post_proc_buffer_int, ppbuf,
+                                 q + (ppflags->deblocking_level - 5) * 10,
+                                 1, 0);
+    } else if (flags & VP9D_DEBLOCK) {
+      vp9_deblock(&cm->post_proc_buffer_int, ppbuf, q);
+    } else {
+      vp8_yv12_copy_frame(&cm->post_proc_buffer_int, ppbuf);
+    }
+  } else if (flags & VP9D_DEMACROBLOCK) {
+    deblock_and_de_macro_block(cm->frame_to_show, ppbuf,
+                               q + (ppflags->deblocking_level - 5) * 10, 1, 0);
+  } else if (flags & VP9D_DEBLOCK) {
+    vp9_deblock(cm->frame_to_show, ppbuf, q);
+  } else {
+    vp8_yv12_copy_frame(cm->frame_to_show, ppbuf);
+  }
+
+  cm->postproc_state.last_base_qindex = cm->base_qindex;
+  cm->postproc_state.last_frame_valid = 1;
+
+  if (flags & VP9D_ADDNOISE) {
+    const int noise_level = ppflags->noise_level;
+    if (ppstate->last_q != q ||
+        ppstate->last_noise != noise_level) {
+      fillrd(ppstate, 63 - q, noise_level);
+    }
+    vpx_plane_add_noise(ppbuf->y_buffer, ppstate->noise, ppstate->blackclamp,
+                        ppstate->whiteclamp, ppstate->bothclamp,
+                        ppbuf->y_width, ppbuf->y_height, ppbuf->y_stride);
+  }
+
+  *dest = *ppbuf;
+
+  /* handle problem with extending borders */
+  dest->y_width = cm->width;
+  dest->y_height = cm->height;
+  dest->uv_width = dest->y_width >> cm->subsampling_x;
+  dest->uv_height = dest->y_height >> cm->subsampling_y;
+
+  swap_mi_and_prev_mi(cm);
+  return 0;
+}
+#endif  // CONFIG_VP9_POSTPROC
diff --git a/libvpx/libvpx/vp9/common/vp9_postproc.h b/libvpx/libvpx/vp9/common/vp9_postproc.h
new file mode 100644
index 0000000..035c9cd
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_postproc.h
@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_COMMON_VP9_POSTPROC_H_
+#define VP9_COMMON_VP9_POSTPROC_H_
+
+#include "vpx_ports/mem.h"
+#include "vpx_scale/yv12config.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_mfqe.h"
+#include "vp9/common/vp9_ppflags.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct postproc_state {
+  int last_q;
+  int last_noise;
+  char noise[3072];
+  int last_base_qindex;
+  int last_frame_valid;
+  MODE_INFO *prev_mip;
+  MODE_INFO *prev_mi;
+  DECLARE_ALIGNED(16, char, blackclamp[16]);
+  DECLARE_ALIGNED(16, char, whiteclamp[16]);
+  DECLARE_ALIGNED(16, char, bothclamp[16]);
+};
+
+struct VP9Common;
+
+#define MFQE_PRECISION 4
+
+int vp9_post_proc_frame(struct VP9Common *cm,
+                        YV12_BUFFER_CONFIG *dest, vp9_ppflags_t *flags);
+
+void vp9_denoise(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, int q);
+
+void vp9_deblock(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, int q);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_POSTPROC_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_ppflags.h b/libvpx/libvpx/vp9/common/vp9_ppflags.h
new file mode 100644
index 0000000..12b989f
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_ppflags.h
@@ -0,0 +1,43 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_PPFLAGS_H_
+#define VP9_COMMON_VP9_PPFLAGS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum {
+  VP9D_NOFILTERING            = 0,
+  VP9D_DEBLOCK                = 1 << 0,
+  VP9D_DEMACROBLOCK           = 1 << 1,
+  VP9D_ADDNOISE               = 1 << 2,
+  VP9D_DEBUG_TXT_FRAME_INFO   = 1 << 3,
+  VP9D_DEBUG_TXT_MBLK_MODES   = 1 << 4,
+  VP9D_DEBUG_TXT_DC_DIFF      = 1 << 5,
+  VP9D_DEBUG_TXT_RATE_INFO    = 1 << 6,
+  VP9D_DEBUG_DRAW_MV          = 1 << 7,
+  VP9D_DEBUG_CLR_BLK_MODES    = 1 << 8,
+  VP9D_DEBUG_CLR_FRM_REF_BLKS = 1 << 9,
+  VP9D_MFQE                   = 1 << 10
+};
+
+typedef struct {
+  int post_proc_flag;
+  int deblocking_level;
+  int noise_level;
+} vp9_ppflags_t;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_PPFLAGS_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_pred_common.c b/libvpx/libvpx/vp9/common/vp9_pred_common.c
new file mode 100644
index 0000000..8f90e70
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_pred_common.c
@@ -0,0 +1,314 @@
+
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+// Returns a context number for the given MB prediction signal
+int vp9_get_pred_context_switchable_interp(const MACROBLOCKD *xd) {
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int left_type = left_mi && is_inter_block(left_mi) ?
+                            left_mi->interp_filter : SWITCHABLE_FILTERS;
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const int above_type = above_mi && is_inter_block(above_mi) ?
+                             above_mi->interp_filter : SWITCHABLE_FILTERS;
+
+  if (left_type == above_type)
+    return left_type;
+  else if (left_type == SWITCHABLE_FILTERS)
+    return above_type;
+  else if (above_type == SWITCHABLE_FILTERS)
+    return left_type;
+  else
+    return SWITCHABLE_FILTERS;
+}
+
+int vp9_get_reference_mode_context(const VP9_COMMON *cm,
+                                   const MACROBLOCKD *xd) {
+  int ctx;
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int has_above = !!above_mi;
+  const int has_left = !!left_mi;
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  if (has_above && has_left) {  // both edges available
+    if (!has_second_ref(above_mi) && !has_second_ref(left_mi))
+      // neither edge uses comp pred (0/1)
+      ctx = (above_mi->ref_frame[0] == cm->comp_fixed_ref) ^
+            (left_mi->ref_frame[0] == cm->comp_fixed_ref);
+    else if (!has_second_ref(above_mi))
+      // one of two edges uses comp pred (2/3)
+      ctx = 2 + (above_mi->ref_frame[0] == cm->comp_fixed_ref ||
+                 !is_inter_block(above_mi));
+    else if (!has_second_ref(left_mi))
+      // one of two edges uses comp pred (2/3)
+      ctx = 2 + (left_mi->ref_frame[0] == cm->comp_fixed_ref ||
+                 !is_inter_block(left_mi));
+    else  // both edges use comp pred (4)
+      ctx = 4;
+  } else if (has_above || has_left) {  // one edge available
+    const MODE_INFO *edge_mi = has_above ? above_mi : left_mi;
+
+    if (!has_second_ref(edge_mi))
+      // edge does not use comp pred (0/1)
+      ctx = edge_mi->ref_frame[0] == cm->comp_fixed_ref;
+    else
+      // edge uses comp pred (3)
+      ctx = 3;
+  } else {  // no edges available (1)
+    ctx = 1;
+  }
+  assert(ctx >= 0 && ctx < COMP_INTER_CONTEXTS);
+  return ctx;
+}
+
+// Returns a context number for the given MB prediction signal
+int vp9_get_pred_context_comp_ref_p(const VP9_COMMON *cm,
+                                    const MACROBLOCKD *xd) {
+  int pred_context;
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int above_in_image = !!above_mi;
+  const int left_in_image = !!left_mi;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  const int fix_ref_idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
+  const int var_ref_idx = !fix_ref_idx;
+
+  if (above_in_image && left_in_image) {  // both edges available
+    const int above_intra = !is_inter_block(above_mi);
+    const int left_intra = !is_inter_block(left_mi);
+
+    if (above_intra && left_intra) {  // intra/intra (2)
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter
+      const MODE_INFO *edge_mi = above_intra ? left_mi : above_mi;
+
+      if (!has_second_ref(edge_mi))  // single pred (1/3)
+        pred_context = 1 + 2 * (edge_mi->ref_frame[0] != cm->comp_var_ref[1]);
+      else  // comp pred (1/3)
+        pred_context = 1 + 2 * (edge_mi->ref_frame[var_ref_idx]
+                                    != cm->comp_var_ref[1]);
+    } else {  // inter/inter
+      const int l_sg = !has_second_ref(left_mi);
+      const int a_sg = !has_second_ref(above_mi);
+      const MV_REFERENCE_FRAME vrfa = a_sg ? above_mi->ref_frame[0]
+                                           : above_mi->ref_frame[var_ref_idx];
+      const MV_REFERENCE_FRAME vrfl = l_sg ? left_mi->ref_frame[0]
+                                           : left_mi->ref_frame[var_ref_idx];
+
+      if (vrfa == vrfl && cm->comp_var_ref[1] == vrfa) {
+        pred_context = 0;
+      } else if (l_sg && a_sg) {  // single/single
+        if ((vrfa == cm->comp_fixed_ref && vrfl == cm->comp_var_ref[0]) ||
+            (vrfl == cm->comp_fixed_ref && vrfa == cm->comp_var_ref[0]))
+          pred_context = 4;
+        else if (vrfa == vrfl)
+          pred_context = 3;
+        else
+          pred_context = 1;
+      } else if (l_sg || a_sg) {  // single/comp
+        const MV_REFERENCE_FRAME vrfc = l_sg ? vrfa : vrfl;
+        const MV_REFERENCE_FRAME rfs = a_sg ? vrfa : vrfl;
+        if (vrfc == cm->comp_var_ref[1] && rfs != cm->comp_var_ref[1])
+          pred_context = 1;
+        else if (rfs == cm->comp_var_ref[1] && vrfc != cm->comp_var_ref[1])
+          pred_context = 2;
+        else
+          pred_context = 4;
+      } else if (vrfa == vrfl) {  // comp/comp
+        pred_context = 4;
+      } else {
+        pred_context = 2;
+      }
+    }
+  } else if (above_in_image || left_in_image) {  // one edge available
+    const MODE_INFO *edge_mi = above_in_image ? above_mi : left_mi;
+
+    if (!is_inter_block(edge_mi)) {
+      pred_context = 2;
+    } else {
+      if (has_second_ref(edge_mi))
+        pred_context = 4 * (edge_mi->ref_frame[var_ref_idx]
+                              != cm->comp_var_ref[1]);
+      else
+        pred_context = 3 * (edge_mi->ref_frame[0] != cm->comp_var_ref[1]);
+    }
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+
+  return pred_context;
+}
+
+int vp9_get_pred_context_single_ref_p1(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int has_above = !!above_mi;
+  const int has_left = !!left_mi;
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mi);
+    const int left_intra = !is_inter_block(left_mi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MODE_INFO *edge_mi = above_intra ? left_mi : above_mi;
+      if (!has_second_ref(edge_mi))
+        pred_context = 4 * (edge_mi->ref_frame[0] == LAST_FRAME);
+      else
+        pred_context = 1 + (edge_mi->ref_frame[0] == LAST_FRAME ||
+                            edge_mi->ref_frame[1] == LAST_FRAME);
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mi);
+      const int left_has_second = has_second_ref(left_mi);
+      const MV_REFERENCE_FRAME above0 = above_mi->ref_frame[0];
+      const MV_REFERENCE_FRAME above1 = above_mi->ref_frame[1];
+      const MV_REFERENCE_FRAME left0 = left_mi->ref_frame[0];
+      const MV_REFERENCE_FRAME left1 = left_mi->ref_frame[1];
+
+      if (above_has_second && left_has_second) {
+        pred_context = 1 + (above0 == LAST_FRAME || above1 == LAST_FRAME ||
+                            left0 == LAST_FRAME || left1 == LAST_FRAME);
+      } else if (above_has_second || left_has_second) {
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf1 = above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf2 = above_has_second ? above1 : left1;
+
+        if (rfs == LAST_FRAME)
+          pred_context = 3 + (crf1 == LAST_FRAME || crf2 == LAST_FRAME);
+        else
+          pred_context = (crf1 == LAST_FRAME || crf2 == LAST_FRAME);
+      } else {
+        pred_context = 2 * (above0 == LAST_FRAME) + 2 * (left0 == LAST_FRAME);
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MODE_INFO *edge_mi = has_above ? above_mi : left_mi;
+    if (!is_inter_block(edge_mi)) {  // intra
+      pred_context = 2;
+    } else {  // inter
+      if (!has_second_ref(edge_mi))
+        pred_context = 4 * (edge_mi->ref_frame[0] == LAST_FRAME);
+      else
+        pred_context = 1 + (edge_mi->ref_frame[0] == LAST_FRAME ||
+                            edge_mi->ref_frame[1] == LAST_FRAME);
+    }
+  } else {  // no edges available
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
+
+int vp9_get_pred_context_single_ref_p2(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int has_above = !!above_mi;
+  const int has_left = !!left_mi;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mi);
+    const int left_intra = !is_inter_block(left_mi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MODE_INFO *edge_mi = above_intra ? left_mi : above_mi;
+      if (!has_second_ref(edge_mi)) {
+        if (edge_mi->ref_frame[0] == LAST_FRAME)
+          pred_context = 3;
+        else
+          pred_context = 4 * (edge_mi->ref_frame[0] == GOLDEN_FRAME);
+      } else {
+        pred_context = 1 + 2 * (edge_mi->ref_frame[0] == GOLDEN_FRAME ||
+                                edge_mi->ref_frame[1] == GOLDEN_FRAME);
+      }
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mi);
+      const int left_has_second = has_second_ref(left_mi);
+      const MV_REFERENCE_FRAME above0 = above_mi->ref_frame[0];
+      const MV_REFERENCE_FRAME above1 = above_mi->ref_frame[1];
+      const MV_REFERENCE_FRAME left0 = left_mi->ref_frame[0];
+      const MV_REFERENCE_FRAME left1 = left_mi->ref_frame[1];
+
+      if (above_has_second && left_has_second) {
+        if (above0 == left0 && above1 == left1)
+          pred_context = 3 * (above0 == GOLDEN_FRAME ||
+                              above1 == GOLDEN_FRAME ||
+                              left0 == GOLDEN_FRAME ||
+                              left1 == GOLDEN_FRAME);
+        else
+          pred_context = 2;
+      } else if (above_has_second || left_has_second) {
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf1 = above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf2 = above_has_second ? above1 : left1;
+
+        if (rfs == GOLDEN_FRAME)
+          pred_context = 3 + (crf1 == GOLDEN_FRAME || crf2 == GOLDEN_FRAME);
+        else if (rfs == ALTREF_FRAME)
+          pred_context = crf1 == GOLDEN_FRAME || crf2 == GOLDEN_FRAME;
+        else
+          pred_context = 1 + 2 * (crf1 == GOLDEN_FRAME || crf2 == GOLDEN_FRAME);
+      } else {
+        if (above0 == LAST_FRAME && left0 == LAST_FRAME) {
+          pred_context = 3;
+        } else if (above0 == LAST_FRAME || left0 == LAST_FRAME) {
+          const MV_REFERENCE_FRAME edge0 = (above0 == LAST_FRAME) ? left0
+                                                                  : above0;
+          pred_context = 4 * (edge0 == GOLDEN_FRAME);
+        } else {
+          pred_context = 2 * (above0 == GOLDEN_FRAME) +
+                             2 * (left0 == GOLDEN_FRAME);
+        }
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MODE_INFO *edge_mi = has_above ? above_mi : left_mi;
+
+    if (!is_inter_block(edge_mi) ||
+        (edge_mi->ref_frame[0] == LAST_FRAME && !has_second_ref(edge_mi)))
+      pred_context = 2;
+    else if (!has_second_ref(edge_mi))
+      pred_context = 4 * (edge_mi->ref_frame[0] == GOLDEN_FRAME);
+    else
+      pred_context = 3 * (edge_mi->ref_frame[0] == GOLDEN_FRAME ||
+                          edge_mi->ref_frame[1] == GOLDEN_FRAME);
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_pred_common.h b/libvpx/libvpx/vp9/common/vp9_pred_common.h
new file mode 100644
index 0000000..f3c676e
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_pred_common.h
@@ -0,0 +1,192 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_PRED_COMMON_H_
+#define VP9_COMMON_VP9_PRED_COMMON_H_
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE int get_segment_id(const VP9_COMMON *cm,
+                                 const uint8_t *segment_ids,
+                                 BLOCK_SIZE bsize, int mi_row, int mi_col) {
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = num_8x8_blocks_wide_lookup[bsize];
+  const int bh = num_8x8_blocks_high_lookup[bsize];
+  const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
+  const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
+  int x, y, segment_id = MAX_SEGMENTS;
+
+  for (y = 0; y < ymis; ++y)
+    for (x = 0; x < xmis; ++x)
+      segment_id =
+          VPXMIN(segment_id, segment_ids[mi_offset + y * cm->mi_cols + x]);
+
+  assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
+  return segment_id;
+}
+
+static INLINE int vp9_get_pred_context_seg_id(const MACROBLOCKD *xd) {
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int above_sip = (above_mi != NULL) ?
+                        above_mi->seg_id_predicted : 0;
+  const int left_sip = (left_mi != NULL) ? left_mi->seg_id_predicted : 0;
+
+  return above_sip + left_sip;
+}
+
+static INLINE vpx_prob vp9_get_pred_prob_seg_id(const struct segmentation *seg,
+                                                const MACROBLOCKD *xd) {
+  return seg->pred_probs[vp9_get_pred_context_seg_id(xd)];
+}
+
+static INLINE int vp9_get_skip_context(const MACROBLOCKD *xd) {
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int above_skip = (above_mi != NULL) ? above_mi->skip : 0;
+  const int left_skip = (left_mi != NULL) ? left_mi->skip : 0;
+  return above_skip + left_skip;
+}
+
+static INLINE vpx_prob vp9_get_skip_prob(const VP9_COMMON *cm,
+                                         const MACROBLOCKD *xd) {
+  return cm->fc->skip_probs[vp9_get_skip_context(xd)];
+}
+
+int vp9_get_pred_context_switchable_interp(const MACROBLOCKD *xd);
+
+// The mode info data structure has a one element border above and to the
+// left of the entries corresponding to real macroblocks.
+// The prediction flags in these dummy entries are initialized to 0.
+// 0 - inter/inter, inter/--, --/inter, --/--
+// 1 - intra/inter, inter/intra
+// 2 - intra/--, --/intra
+// 3 - intra/intra
+static INLINE int get_intra_inter_context(const MACROBLOCKD *xd) {
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int has_above = !!above_mi;
+  const int has_left = !!left_mi;
+
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mi);
+    const int left_intra = !is_inter_block(left_mi);
+    return left_intra && above_intra ? 3 : left_intra || above_intra;
+  } else if (has_above || has_left) {  // one edge available
+    return 2 * !is_inter_block(has_above ? above_mi : left_mi);
+  }
+  return 0;
+}
+
+static INLINE vpx_prob vp9_get_intra_inter_prob(const VP9_COMMON *cm,
+                                                const MACROBLOCKD *xd) {
+  return cm->fc->intra_inter_prob[get_intra_inter_context(xd)];
+}
+
+int vp9_get_reference_mode_context(const VP9_COMMON *cm, const MACROBLOCKD *xd);
+
+static INLINE vpx_prob vp9_get_reference_mode_prob(const VP9_COMMON *cm,
+                                                   const MACROBLOCKD *xd) {
+  return cm->fc->comp_inter_prob[vp9_get_reference_mode_context(cm, xd)];
+}
+
+int vp9_get_pred_context_comp_ref_p(const VP9_COMMON *cm,
+                                    const MACROBLOCKD *xd);
+
+static INLINE vpx_prob vp9_get_pred_prob_comp_ref_p(const VP9_COMMON *cm,
+                                                    const MACROBLOCKD *xd) {
+  const int pred_context = vp9_get_pred_context_comp_ref_p(cm, xd);
+  return cm->fc->comp_ref_prob[pred_context];
+}
+
+int vp9_get_pred_context_single_ref_p1(const MACROBLOCKD *xd);
+
+static INLINE vpx_prob vp9_get_pred_prob_single_ref_p1(const VP9_COMMON *cm,
+                                                       const MACROBLOCKD *xd) {
+  return cm->fc->single_ref_prob[vp9_get_pred_context_single_ref_p1(xd)][0];
+}
+
+int vp9_get_pred_context_single_ref_p2(const MACROBLOCKD *xd);
+
+static INLINE vpx_prob vp9_get_pred_prob_single_ref_p2(const VP9_COMMON *cm,
+                                                       const MACROBLOCKD *xd) {
+  return cm->fc->single_ref_prob[vp9_get_pred_context_single_ref_p2(xd)][1];
+}
+
+// Returns a context number for the given MB prediction signal
+// The mode info data structure has a one element border above and to the
+// left of the entries corresponding to real blocks.
+// The prediction flags in these dummy entries are initialized to 0.
+static INLINE int get_tx_size_context(const MACROBLOCKD *xd) {
+  const int max_tx_size = max_txsize_lookup[xd->mi[0]->sb_type];
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int has_above = !!above_mi;
+  const int has_left = !!left_mi;
+  int above_ctx = (has_above && !above_mi->skip) ? (int)above_mi->tx_size
+                                                 : max_tx_size;
+  int left_ctx = (has_left && !left_mi->skip) ? (int)left_mi->tx_size
+                                              : max_tx_size;
+  if (!has_left)
+    left_ctx = above_ctx;
+
+  if (!has_above)
+    above_ctx = left_ctx;
+
+  return (above_ctx + left_ctx) > max_tx_size;
+}
+
+static INLINE const vpx_prob *get_tx_probs(TX_SIZE max_tx_size, int ctx,
+                                           const struct tx_probs *tx_probs) {
+  switch (max_tx_size) {
+    case TX_8X8:
+      return tx_probs->p8x8[ctx];
+    case TX_16X16:
+      return tx_probs->p16x16[ctx];
+    case TX_32X32:
+      return tx_probs->p32x32[ctx];
+    default:
+      assert(0 && "Invalid max_tx_size.");
+      return NULL;
+  }
+}
+
+static INLINE const vpx_prob *get_tx_probs2(TX_SIZE max_tx_size,
+                                            const MACROBLOCKD *xd,
+                                            const struct tx_probs *tx_probs) {
+  return get_tx_probs(max_tx_size, get_tx_size_context(xd), tx_probs);
+}
+
+static INLINE unsigned int *get_tx_counts(TX_SIZE max_tx_size, int ctx,
+                                          struct tx_counts *tx_counts) {
+  switch (max_tx_size) {
+    case TX_8X8:
+      return tx_counts->p8x8[ctx];
+    case TX_16X16:
+      return tx_counts->p16x16[ctx];
+    case TX_32X32:
+      return tx_counts->p32x32[ctx];
+    default:
+      assert(0 && "Invalid max_tx_size.");
+      return NULL;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_PRED_COMMON_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_quant_common.c b/libvpx/libvpx/vp9/common/vp9_quant_common.c
new file mode 100644
index 0000000..d83f3c1
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_quant_common.c
@@ -0,0 +1,278 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+static const int16_t dc_qlookup[QINDEX_RANGE] = {
+  4,       8,    8,    9,   10,   11,   12,   12,
+  13,     14,   15,   16,   17,   18,   19,   19,
+  20,     21,   22,   23,   24,   25,   26,   26,
+  27,     28,   29,   30,   31,   32,   32,   33,
+  34,     35,   36,   37,   38,   38,   39,   40,
+  41,     42,   43,   43,   44,   45,   46,   47,
+  48,     48,   49,   50,   51,   52,   53,   53,
+  54,     55,   56,   57,   57,   58,   59,   60,
+  61,     62,   62,   63,   64,   65,   66,   66,
+  67,     68,   69,   70,   70,   71,   72,   73,
+  74,     74,   75,   76,   77,   78,   78,   79,
+  80,     81,   81,   82,   83,   84,   85,   85,
+  87,     88,   90,   92,   93,   95,   96,   98,
+  99,    101,  102,  104,  105,  107,  108,  110,
+  111,   113,  114,  116,  117,  118,  120,  121,
+  123,   125,  127,  129,  131,  134,  136,  138,
+  140,   142,  144,  146,  148,  150,  152,  154,
+  156,   158,  161,  164,  166,  169,  172,  174,
+  177,   180,  182,  185,  187,  190,  192,  195,
+  199,   202,  205,  208,  211,  214,  217,  220,
+  223,   226,  230,  233,  237,  240,  243,  247,
+  250,   253,  257,  261,  265,  269,  272,  276,
+  280,   284,  288,  292,  296,  300,  304,  309,
+  313,   317,  322,  326,  330,  335,  340,  344,
+  349,   354,  359,  364,  369,  374,  379,  384,
+  389,   395,  400,  406,  411,  417,  423,  429,
+  435,   441,  447,  454,  461,  467,  475,  482,
+  489,   497,  505,  513,  522,  530,  539,  549,
+  559,   569,  579,  590,  602,  614,  626,  640,
+  654,   668,  684,  700,  717,  736,  755,  775,
+  796,   819,  843,  869,  896,  925,  955,  988,
+  1022, 1058, 1098, 1139, 1184, 1232, 1282, 1336,
+};
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static const int16_t dc_qlookup_10[QINDEX_RANGE] = {
+  4,     9,    10,    13,    15,    17,    20,    22,
+  25,    28,    31,    34,    37,    40,    43,    47,
+  50,    53,    57,    60,    64,    68,    71,    75,
+  78,    82,    86,    90,    93,    97,   101,   105,
+  109,   113,   116,   120,   124,   128,   132,   136,
+  140,   143,   147,   151,   155,   159,   163,   166,
+  170,   174,   178,   182,   185,   189,   193,   197,
+  200,   204,   208,   212,   215,   219,   223,   226,
+  230,   233,   237,   241,   244,   248,   251,   255,
+  259,   262,   266,   269,   273,   276,   280,   283,
+  287,   290,   293,   297,   300,   304,   307,   310,
+  314,   317,   321,   324,   327,   331,   334,   337,
+  343,   350,   356,   362,   369,   375,   381,   387,
+  394,   400,   406,   412,   418,   424,   430,   436,
+  442,   448,   454,   460,   466,   472,   478,   484,
+  490,   499,   507,   516,   525,   533,   542,   550,
+  559,   567,   576,   584,   592,   601,   609,   617,
+  625,   634,   644,   655,   666,   676,   687,   698,
+  708,   718,   729,   739,   749,   759,   770,   782,
+  795,   807,   819,   831,   844,   856,   868,   880,
+  891,   906,   920,   933,   947,   961,   975,   988,
+  1001,  1015,  1030,  1045,  1061,  1076,  1090,  1105,
+  1120,  1137,  1153,  1170,  1186,  1202,  1218,  1236,
+  1253,  1271,  1288,  1306,  1323,  1342,  1361,  1379,
+  1398,  1416,  1436,  1456,  1476,  1496,  1516,  1537,
+  1559,  1580,  1601,  1624,  1647,  1670,  1692,  1717,
+  1741,  1766,  1791,  1817,  1844,  1871,  1900,  1929,
+  1958,  1990,  2021,  2054,  2088,  2123,  2159,  2197,
+  2236,  2276,  2319,  2363,  2410,  2458,  2508,  2561,
+  2616,  2675,  2737,  2802,  2871,  2944,  3020,  3102,
+  3188,  3280,  3375,  3478,  3586,  3702,  3823,  3953,
+  4089,  4236,  4394,  4559,  4737,  4929,  5130,  5347,
+};
+
+static const int16_t dc_qlookup_12[QINDEX_RANGE] = {
+  4,    12,    18,    25,    33,    41,    50,    60,
+  70,    80,    91,   103,   115,   127,   140,   153,
+  166,   180,   194,   208,   222,   237,   251,   266,
+  281,   296,   312,   327,   343,   358,   374,   390,
+  405,   421,   437,   453,   469,   484,   500,   516,
+  532,   548,   564,   580,   596,   611,   627,   643,
+  659,   674,   690,   706,   721,   737,   752,   768,
+  783,   798,   814,   829,   844,   859,   874,   889,
+  904,   919,   934,   949,   964,   978,   993,  1008,
+  1022,  1037,  1051,  1065,  1080,  1094,  1108,  1122,
+  1136,  1151,  1165,  1179,  1192,  1206,  1220,  1234,
+  1248,  1261,  1275,  1288,  1302,  1315,  1329,  1342,
+  1368,  1393,  1419,  1444,  1469,  1494,  1519,  1544,
+  1569,  1594,  1618,  1643,  1668,  1692,  1717,  1741,
+  1765,  1789,  1814,  1838,  1862,  1885,  1909,  1933,
+  1957,  1992,  2027,  2061,  2096,  2130,  2165,  2199,
+  2233,  2267,  2300,  2334,  2367,  2400,  2434,  2467,
+  2499,  2532,  2575,  2618,  2661,  2704,  2746,  2788,
+  2830,  2872,  2913,  2954,  2995,  3036,  3076,  3127,
+  3177,  3226,  3275,  3324,  3373,  3421,  3469,  3517,
+  3565,  3621,  3677,  3733,  3788,  3843,  3897,  3951,
+  4005,  4058,  4119,  4181,  4241,  4301,  4361,  4420,
+  4479,  4546,  4612,  4677,  4742,  4807,  4871,  4942,
+  5013,  5083,  5153,  5222,  5291,  5367,  5442,  5517,
+  5591,  5665,  5745,  5825,  5905,  5984,  6063,  6149,
+  6234,  6319,  6404,  6495,  6587,  6678,  6769,  6867,
+  6966,  7064,  7163,  7269,  7376,  7483,  7599,  7715,
+  7832,  7958,  8085,  8214,  8352,  8492,  8635,  8788,
+  8945,  9104,  9275,  9450,  9639,  9832, 10031, 10245,
+  10465, 10702, 10946, 11210, 11482, 11776, 12081, 12409,
+  12750, 13118, 13501, 13913, 14343, 14807, 15290, 15812,
+  16356, 16943, 17575, 18237, 18949, 19718, 20521, 21387,
+};
+#endif
+
+static const int16_t ac_qlookup[QINDEX_RANGE] = {
+  4,       8,    9,   10,   11,   12,   13,   14,
+  15,     16,   17,   18,   19,   20,   21,   22,
+  23,     24,   25,   26,   27,   28,   29,   30,
+  31,     32,   33,   34,   35,   36,   37,   38,
+  39,     40,   41,   42,   43,   44,   45,   46,
+  47,     48,   49,   50,   51,   52,   53,   54,
+  55,     56,   57,   58,   59,   60,   61,   62,
+  63,     64,   65,   66,   67,   68,   69,   70,
+  71,     72,   73,   74,   75,   76,   77,   78,
+  79,     80,   81,   82,   83,   84,   85,   86,
+  87,     88,   89,   90,   91,   92,   93,   94,
+  95,     96,   97,   98,   99,  100,  101,  102,
+  104,   106,  108,  110,  112,  114,  116,  118,
+  120,   122,  124,  126,  128,  130,  132,  134,
+  136,   138,  140,  142,  144,  146,  148,  150,
+  152,   155,  158,  161,  164,  167,  170,  173,
+  176,   179,  182,  185,  188,  191,  194,  197,
+  200,   203,  207,  211,  215,  219,  223,  227,
+  231,   235,  239,  243,  247,  251,  255,  260,
+  265,   270,  275,  280,  285,  290,  295,  300,
+  305,   311,  317,  323,  329,  335,  341,  347,
+  353,   359,  366,  373,  380,  387,  394,  401,
+  408,   416,  424,  432,  440,  448,  456,  465,
+  474,   483,  492,  501,  510,  520,  530,  540,
+  550,   560,  571,  582,  593,  604,  615,  627,
+  639,   651,  663,  676,  689,  702,  715,  729,
+  743,   757,  771,  786,  801,  816,  832,  848,
+  864,   881,  898,  915,  933,  951,  969,  988,
+  1007, 1026, 1046, 1066, 1087, 1108, 1129, 1151,
+  1173, 1196, 1219, 1243, 1267, 1292, 1317, 1343,
+  1369, 1396, 1423, 1451, 1479, 1508, 1537, 1567,
+  1597, 1628, 1660, 1692, 1725, 1759, 1793, 1828,
+};
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static const int16_t ac_qlookup_10[QINDEX_RANGE] = {
+  4,     9,    11,    13,    16,    18,    21,    24,
+  27,    30,    33,    37,    40,    44,    48,    51,
+  55,    59,    63,    67,    71,    75,    79,    83,
+  88,    92,    96,   100,   105,   109,   114,   118,
+  122,   127,   131,   136,   140,   145,   149,   154,
+  158,   163,   168,   172,   177,   181,   186,   190,
+  195,   199,   204,   208,   213,   217,   222,   226,
+  231,   235,   240,   244,   249,   253,   258,   262,
+  267,   271,   275,   280,   284,   289,   293,   297,
+  302,   306,   311,   315,   319,   324,   328,   332,
+  337,   341,   345,   349,   354,   358,   362,   367,
+  371,   375,   379,   384,   388,   392,   396,   401,
+  409,   417,   425,   433,   441,   449,   458,   466,
+  474,   482,   490,   498,   506,   514,   523,   531,
+  539,   547,   555,   563,   571,   579,   588,   596,
+  604,   616,   628,   640,   652,   664,   676,   688,
+  700,   713,   725,   737,   749,   761,   773,   785,
+  797,   809,   825,   841,   857,   873,   889,   905,
+  922,   938,   954,   970,   986,  1002,  1018,  1038,
+  1058,  1078,  1098,  1118,  1138,  1158,  1178,  1198,
+  1218,  1242,  1266,  1290,  1314,  1338,  1362,  1386,
+  1411,  1435,  1463,  1491,  1519,  1547,  1575,  1603,
+  1631,  1663,  1695,  1727,  1759,  1791,  1823,  1859,
+  1895,  1931,  1967,  2003,  2039,  2079,  2119,  2159,
+  2199,  2239,  2283,  2327,  2371,  2415,  2459,  2507,
+  2555,  2603,  2651,  2703,  2755,  2807,  2859,  2915,
+  2971,  3027,  3083,  3143,  3203,  3263,  3327,  3391,
+  3455,  3523,  3591,  3659,  3731,  3803,  3876,  3952,
+  4028,  4104,  4184,  4264,  4348,  4432,  4516,  4604,
+  4692,  4784,  4876,  4972,  5068,  5168,  5268,  5372,
+  5476,  5584,  5692,  5804,  5916,  6032,  6148,  6268,
+  6388,  6512,  6640,  6768,  6900,  7036,  7172,  7312,
+};
+
+static const int16_t ac_qlookup_12[QINDEX_RANGE] = {
+  4,    13,    19,    27,    35,    44,    54,    64,
+  75,    87,    99,   112,   126,   139,   154,   168,
+  183,   199,   214,   230,   247,   263,   280,   297,
+  314,   331,   349,   366,   384,   402,   420,   438,
+  456,   475,   493,   511,   530,   548,   567,   586,
+  604,   623,   642,   660,   679,   698,   716,   735,
+  753,   772,   791,   809,   828,   846,   865,   884,
+  902,   920,   939,   957,   976,   994,  1012,  1030,
+  1049,  1067,  1085,  1103,  1121,  1139,  1157,  1175,
+  1193,  1211,  1229,  1246,  1264,  1282,  1299,  1317,
+  1335,  1352,  1370,  1387,  1405,  1422,  1440,  1457,
+  1474,  1491,  1509,  1526,  1543,  1560,  1577,  1595,
+  1627,  1660,  1693,  1725,  1758,  1791,  1824,  1856,
+  1889,  1922,  1954,  1987,  2020,  2052,  2085,  2118,
+  2150,  2183,  2216,  2248,  2281,  2313,  2346,  2378,
+  2411,  2459,  2508,  2556,  2605,  2653,  2701,  2750,
+  2798,  2847,  2895,  2943,  2992,  3040,  3088,  3137,
+  3185,  3234,  3298,  3362,  3426,  3491,  3555,  3619,
+  3684,  3748,  3812,  3876,  3941,  4005,  4069,  4149,
+  4230,  4310,  4390,  4470,  4550,  4631,  4711,  4791,
+  4871,  4967,  5064,  5160,  5256,  5352,  5448,  5544,
+  5641,  5737,  5849,  5961,  6073,  6185,  6297,  6410,
+  6522,  6650,  6778,  6906,  7034,  7162,  7290,  7435,
+  7579,  7723,  7867,  8011,  8155,  8315,  8475,  8635,
+  8795,  8956,  9132,  9308,  9484,  9660,  9836, 10028,
+  10220, 10412, 10604, 10812, 11020, 11228, 11437, 11661,
+  11885, 12109, 12333, 12573, 12813, 13053, 13309, 13565,
+  13821, 14093, 14365, 14637, 14925, 15213, 15502, 15806,
+  16110, 16414, 16734, 17054, 17390, 17726, 18062, 18414,
+  18766, 19134, 19502, 19886, 20270, 20670, 21070, 21486,
+  21902, 22334, 22766, 23214, 23662, 24126, 24590, 25070,
+  25551, 26047, 26559, 27071, 27599, 28143, 28687, 29247,
+};
+#endif
+
+int16_t vp9_dc_quant(int qindex, int delta, vpx_bit_depth_t bit_depth) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      return dc_qlookup[clamp(qindex + delta, 0, MAXQ)];
+    case VPX_BITS_10:
+      return dc_qlookup_10[clamp(qindex + delta, 0, MAXQ)];
+    case VPX_BITS_12:
+      return dc_qlookup_12[clamp(qindex + delta, 0, MAXQ)];
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  (void) bit_depth;
+  return dc_qlookup[clamp(qindex + delta, 0, MAXQ)];
+#endif
+}
+
+int16_t vp9_ac_quant(int qindex, int delta, vpx_bit_depth_t bit_depth) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      return ac_qlookup[clamp(qindex + delta, 0, MAXQ)];
+    case VPX_BITS_10:
+      return ac_qlookup_10[clamp(qindex + delta, 0, MAXQ)];
+    case VPX_BITS_12:
+      return ac_qlookup_12[clamp(qindex + delta, 0, MAXQ)];
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  (void) bit_depth;
+  return ac_qlookup[clamp(qindex + delta, 0, MAXQ)];
+#endif
+}
+
+int vp9_get_qindex(const struct segmentation *seg, int segment_id,
+                   int base_qindex) {
+  if (segfeature_active(seg, segment_id, SEG_LVL_ALT_Q)) {
+    const int data = get_segdata(seg, segment_id, SEG_LVL_ALT_Q);
+    const int seg_qindex = seg->abs_delta == SEGMENT_ABSDATA ?
+        data : base_qindex + data;
+    return clamp(seg_qindex, 0, MAXQ);
+  } else {
+    return base_qindex;
+  }
+}
+
diff --git a/libvpx/libvpx/vp9/common/vp9_quant_common.h b/libvpx/libvpx/vp9/common/vp9_quant_common.h
new file mode 100644
index 0000000..4bae4a8
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_quant_common.h
@@ -0,0 +1,36 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_QUANT_COMMON_H_
+#define VP9_COMMON_VP9_QUANT_COMMON_H_
+
+#include "vpx/vpx_codec.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MINQ 0
+#define MAXQ 255
+#define QINDEX_RANGE (MAXQ - MINQ + 1)
+#define QINDEX_BITS 8
+
+int16_t vp9_dc_quant(int qindex, int delta, vpx_bit_depth_t bit_depth);
+int16_t vp9_ac_quant(int qindex, int delta, vpx_bit_depth_t bit_depth);
+
+int vp9_get_qindex(const struct segmentation *seg, int segment_id,
+                   int base_qindex);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_QUANT_COMMON_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_reconinter.c b/libvpx/libvpx/vp9/common/vp9_reconinter.c
new file mode 100644
index 0000000..84718e9
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_reconinter.c
@@ -0,0 +1,309 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "./vpx_scale_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_build_inter_predictor(const uint8_t *src, int src_stride,
+                                      uint8_t *dst, int dst_stride,
+                                      const MV *src_mv,
+                                      const struct scale_factors *sf,
+                                      int w, int h, int ref,
+                                      const InterpKernel *kernel,
+                                      enum mv_precision precision,
+                                      int x, int y, int bd) {
+  const int is_q4 = precision == MV_PRECISION_Q4;
+  const MV mv_q4 = { is_q4 ? src_mv->row : src_mv->row * 2,
+                     is_q4 ? src_mv->col : src_mv->col * 2 };
+  MV32 mv = vp9_scale_mv(&mv_q4, x, y, sf);
+  const int subpel_x = mv.col & SUBPEL_MASK;
+  const int subpel_y = mv.row & SUBPEL_MASK;
+
+  src += (mv.row >> SUBPEL_BITS) * src_stride + (mv.col >> SUBPEL_BITS);
+
+  highbd_inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
+                         sf, w, h, ref, kernel, sf->x_step_q4, sf->y_step_q4,
+                         bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_build_inter_predictor(const uint8_t *src, int src_stride,
+                               uint8_t *dst, int dst_stride,
+                               const MV *src_mv,
+                               const struct scale_factors *sf,
+                               int w, int h, int ref,
+                               const InterpKernel *kernel,
+                               enum mv_precision precision,
+                               int x, int y) {
+  const int is_q4 = precision == MV_PRECISION_Q4;
+  const MV mv_q4 = { is_q4 ? src_mv->row : src_mv->row * 2,
+                     is_q4 ? src_mv->col : src_mv->col * 2 };
+  MV32 mv = vp9_scale_mv(&mv_q4, x, y, sf);
+  const int subpel_x = mv.col & SUBPEL_MASK;
+  const int subpel_y = mv.row & SUBPEL_MASK;
+
+  src += (mv.row >> SUBPEL_BITS) * src_stride + (mv.col >> SUBPEL_BITS);
+
+  inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
+                  sf, w, h, ref, kernel, sf->x_step_q4, sf->y_step_q4);
+}
+
+static INLINE int round_mv_comp_q4(int value) {
+  return (value < 0 ? value - 2 : value + 2) / 4;
+}
+
+static MV mi_mv_pred_q4(const MODE_INFO *mi, int idx) {
+  MV res = { round_mv_comp_q4(mi->bmi[0].as_mv[idx].as_mv.row +
+                              mi->bmi[1].as_mv[idx].as_mv.row +
+                              mi->bmi[2].as_mv[idx].as_mv.row +
+                              mi->bmi[3].as_mv[idx].as_mv.row),
+             round_mv_comp_q4(mi->bmi[0].as_mv[idx].as_mv.col +
+                              mi->bmi[1].as_mv[idx].as_mv.col +
+                              mi->bmi[2].as_mv[idx].as_mv.col +
+                              mi->bmi[3].as_mv[idx].as_mv.col) };
+  return res;
+}
+
+static INLINE int round_mv_comp_q2(int value) {
+  return (value < 0 ? value - 1 : value + 1) / 2;
+}
+
+static MV mi_mv_pred_q2(const MODE_INFO *mi, int idx, int block0, int block1) {
+  MV res = { round_mv_comp_q2(mi->bmi[block0].as_mv[idx].as_mv.row +
+                              mi->bmi[block1].as_mv[idx].as_mv.row),
+             round_mv_comp_q2(mi->bmi[block0].as_mv[idx].as_mv.col +
+                              mi->bmi[block1].as_mv[idx].as_mv.col) };
+  return res;
+}
+
+// TODO(jkoleszar): yet another mv clamping function :-(
+MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd, const MV *src_mv,
+                             int bw, int bh, int ss_x, int ss_y) {
+  // If the MV points so far into the UMV border that no visible pixels
+  // are used for reconstruction, the subpel part of the MV can be
+  // discarded and the MV limited to 16 pixels with equivalent results.
+  const int spel_left = (VP9_INTERP_EXTEND + bw) << SUBPEL_BITS;
+  const int spel_right = spel_left - SUBPEL_SHIFTS;
+  const int spel_top = (VP9_INTERP_EXTEND + bh) << SUBPEL_BITS;
+  const int spel_bottom = spel_top - SUBPEL_SHIFTS;
+  MV clamped_mv = {
+    src_mv->row * (1 << (1 - ss_y)),
+    src_mv->col * (1 << (1 - ss_x))
+  };
+  assert(ss_x <= 1);
+  assert(ss_y <= 1);
+
+  clamp_mv(&clamped_mv,
+           xd->mb_to_left_edge * (1 << (1 - ss_x)) - spel_left,
+           xd->mb_to_right_edge * (1 << (1 - ss_x)) + spel_right,
+           xd->mb_to_top_edge * (1 << (1 - ss_y)) - spel_top,
+           xd->mb_to_bottom_edge * (1 << (1 - ss_y)) + spel_bottom);
+
+  return clamped_mv;
+}
+
+MV average_split_mvs(const struct macroblockd_plane *pd,
+                     const MODE_INFO *mi, int ref, int block) {
+  const int ss_idx = ((pd->subsampling_x > 0) << 1) | (pd->subsampling_y > 0);
+  MV res = {0, 0};
+  switch (ss_idx) {
+    case 0:
+      res = mi->bmi[block].as_mv[ref].as_mv;
+      break;
+    case 1:
+      res = mi_mv_pred_q2(mi, ref, block, block + 2);
+      break;
+    case 2:
+      res = mi_mv_pred_q2(mi, ref, block, block + 1);
+      break;
+    case 3:
+      res = mi_mv_pred_q4(mi, ref);
+      break;
+    default:
+      assert(ss_idx <= 3 && ss_idx >= 0);
+  }
+  return res;
+}
+
+static void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
+                                   int bw, int bh,
+                                   int x, int y, int w, int h,
+                                   int mi_x, int mi_y) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const MODE_INFO *mi = xd->mi[0];
+  const int is_compound = has_second_ref(mi);
+  const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
+  int ref;
+
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
+    struct buf_2d *const pre_buf = &pd->pre[ref];
+    struct buf_2d *const dst_buf = &pd->dst;
+    uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
+    const MV mv = mi->sb_type < BLOCK_8X8
+               ? average_split_mvs(pd, mi, ref, block)
+               : mi->mv[ref].as_mv;
+
+    // TODO(jkoleszar): This clamping is done in the incorrect place for the
+    // scaling case. It needs to be done on the scaled MV, not the pre-scaling
+    // MV. Note however that it performs the subsampling aware scaling so
+    // that the result is always q4.
+    // mv_precision precision is MV_PRECISION_Q4.
+    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, &mv, bw, bh,
+                                               pd->subsampling_x,
+                                               pd->subsampling_y);
+
+    uint8_t *pre;
+    MV32 scaled_mv;
+    int xs, ys, subpel_x, subpel_y;
+    const int is_scaled = vp9_is_scaled(sf);
+
+    if (is_scaled) {
+      // Co-ordinate of containing block to pixel precision.
+      const int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
+      const int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
+#if CONFIG_BETTER_HW_COMPATIBILITY
+      assert(xd->mi[0]->sb_type != BLOCK_4X8 &&
+             xd->mi[0]->sb_type != BLOCK_8X4);
+      assert(mv_q4.row == mv.row * (1 << (1 - pd->subsampling_y)) &&
+             mv_q4.col == mv.col * (1 << (1 - pd->subsampling_x)));
+#endif
+      if (plane == 0)
+        pre_buf->buf = xd->block_refs[ref]->buf->y_buffer;
+      else if (plane == 1)
+        pre_buf->buf = xd->block_refs[ref]->buf->u_buffer;
+      else
+        pre_buf->buf = xd->block_refs[ref]->buf->v_buffer;
+
+      pre_buf->buf += scaled_buffer_offset(x_start + x, y_start + y,
+                                           pre_buf->stride, sf);
+      pre = pre_buf->buf;
+      scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
+      xs = sf->x_step_q4;
+      ys = sf->y_step_q4;
+    } else {
+      pre = pre_buf->buf + (y * pre_buf->stride + x);
+      scaled_mv.row = mv_q4.row;
+      scaled_mv.col = mv_q4.col;
+      xs = ys = 16;
+    }
+    subpel_x = scaled_mv.col & SUBPEL_MASK;
+    subpel_y = scaled_mv.row & SUBPEL_MASK;
+    pre += (scaled_mv.row >> SUBPEL_BITS) * pre_buf->stride
+           + (scaled_mv.col >> SUBPEL_BITS);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      highbd_inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
+                             subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys,
+                             xd->bd);
+    } else {
+      inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
+                      subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys);
+    }
+#else
+    inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
+                    subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+}
+
+static void build_inter_predictors_for_planes(MACROBLOCKD *xd, BLOCK_SIZE bsize,
+                                              int mi_row, int mi_col,
+                                              int plane_from, int plane_to) {
+  int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
+  for (plane = plane_from; plane <= plane_to; ++plane) {
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
+                                                        &xd->plane[plane]);
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+    const int bw = 4 * num_4x4_w;
+    const int bh = 4 * num_4x4_h;
+
+    if (xd->mi[0]->sb_type < BLOCK_8X8) {
+      int i = 0, x, y;
+      assert(bsize == BLOCK_8X8);
+      for (y = 0; y < num_4x4_h; ++y)
+        for (x = 0; x < num_4x4_w; ++x)
+           build_inter_predictors(xd, plane, i++, bw, bh,
+                                  4 * x, 4 * y, 4, 4, mi_x, mi_y);
+    } else {
+      build_inter_predictors(xd, plane, 0, bw, bh,
+                             0, 0, bw, bh, mi_x, mi_y);
+    }
+  }
+}
+
+void vp9_build_inter_predictors_sby(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                    BLOCK_SIZE bsize) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, 0, 0);
+}
+
+void vp9_build_inter_predictors_sbp(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                    BLOCK_SIZE bsize, int plane) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, plane, plane);
+}
+
+void vp9_build_inter_predictors_sbuv(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                     BLOCK_SIZE bsize) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, 1,
+                                    MAX_MB_PLANE - 1);
+}
+
+void vp9_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                   BLOCK_SIZE bsize) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, 0,
+                                    MAX_MB_PLANE - 1);
+}
+
+void vp9_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col) {
+  uint8_t *const buffers[MAX_MB_PLANE] = { src->y_buffer, src->u_buffer,
+      src->v_buffer};
+  const int strides[MAX_MB_PLANE] = { src->y_stride, src->uv_stride,
+      src->uv_stride};
+  int i;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    struct macroblockd_plane *const pd = &planes[i];
+    setup_pred_plane(&pd->dst, buffers[i], strides[i], mi_row, mi_col, NULL,
+                     pd->subsampling_x, pd->subsampling_y);
+  }
+}
+
+void vp9_setup_pre_planes(MACROBLOCKD *xd, int idx,
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col,
+                          const struct scale_factors *sf) {
+  if (src != NULL) {
+    int i;
+    uint8_t *const buffers[MAX_MB_PLANE] = { src->y_buffer, src->u_buffer,
+        src->v_buffer};
+    const int strides[MAX_MB_PLANE] = { src->y_stride, src->uv_stride,
+        src->uv_stride};
+    for (i = 0; i < MAX_MB_PLANE; ++i) {
+      struct macroblockd_plane *const pd = &xd->plane[i];
+      setup_pred_plane(&pd->pre[idx], buffers[i], strides[i], mi_row, mi_col,
+                       sf, pd->subsampling_x, pd->subsampling_y);
+    }
+  }
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_reconinter.h b/libvpx/libvpx/vp9/common/vp9_reconinter.h
new file mode 100644
index 0000000..07745e3
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_reconinter.h
@@ -0,0 +1,119 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_RECONINTER_H_
+#define VP9_COMMON_VP9_RECONINTER_H_
+
+#include "vp9/common/vp9_filter.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_filter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE void inter_predictor(const uint8_t *src, int src_stride,
+                                   uint8_t *dst, int dst_stride,
+                                   const int subpel_x,
+                                   const int subpel_y,
+                                   const struct scale_factors *sf,
+                                   int w, int h, int ref,
+                                   const InterpKernel *kernel,
+                                   int xs, int ys) {
+  sf->predict[subpel_x != 0][subpel_y != 0][ref](
+      src, src_stride, dst, dst_stride,
+      kernel[subpel_x], xs, kernel[subpel_y], ys, w, h);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE void highbd_inter_predictor(const uint8_t *src, int src_stride,
+                                          uint8_t *dst, int dst_stride,
+                                          const int subpel_x,
+                                          const int subpel_y,
+                                          const struct scale_factors *sf,
+                                          int w, int h, int ref,
+                                          const InterpKernel *kernel,
+                                          int xs, int ys, int bd) {
+  sf->highbd_predict[subpel_x != 0][subpel_y != 0][ref](
+      src, src_stride, dst, dst_stride,
+      kernel[subpel_x], xs, kernel[subpel_y], ys, w, h, bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+MV average_split_mvs(const struct macroblockd_plane *pd, const MODE_INFO *mi,
+                     int ref, int block);
+
+MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd, const MV *src_mv,
+                             int bw, int bh, int ss_x, int ss_y);
+
+void vp9_build_inter_predictors_sby(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                    BLOCK_SIZE bsize);
+
+void vp9_build_inter_predictors_sbp(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                    BLOCK_SIZE bsize, int plane);
+
+void vp9_build_inter_predictors_sbuv(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                     BLOCK_SIZE bsize);
+
+void vp9_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                   BLOCK_SIZE bsize);
+
+void vp9_build_inter_predictor(const uint8_t *src, int src_stride,
+                               uint8_t *dst, int dst_stride,
+                               const MV *mv_q3,
+                               const struct scale_factors *sf,
+                               int w, int h, int do_avg,
+                               const InterpKernel *kernel,
+                               enum mv_precision precision,
+                               int x, int y);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_build_inter_predictor(const uint8_t *src, int src_stride,
+                                      uint8_t *dst, int dst_stride,
+                                      const MV *mv_q3,
+                                      const struct scale_factors *sf,
+                                      int w, int h, int do_avg,
+                                      const InterpKernel *kernel,
+                                      enum mv_precision precision,
+                                      int x, int y, int bd);
+#endif
+
+static INLINE int scaled_buffer_offset(int x_offset, int y_offset, int stride,
+                                       const struct scale_factors *sf) {
+  const int x = sf ? sf->scale_value_x(x_offset, sf) : x_offset;
+  const int y = sf ? sf->scale_value_y(y_offset, sf) : y_offset;
+  return y * stride + x;
+}
+
+static INLINE void setup_pred_plane(struct buf_2d *dst,
+                                    uint8_t *src, int stride,
+                                    int mi_row, int mi_col,
+                                    const struct scale_factors *scale,
+                                    int subsampling_x, int subsampling_y) {
+  const int x = (MI_SIZE * mi_col) >> subsampling_x;
+  const int y = (MI_SIZE * mi_row) >> subsampling_y;
+  dst->buf = src + scaled_buffer_offset(x, y, stride, scale);
+  dst->stride = stride;
+}
+
+void vp9_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col);
+
+void vp9_setup_pre_planes(MACROBLOCKD *xd, int idx,
+                          const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
+                          const struct scale_factors *sf);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_RECONINTER_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_reconintra.c b/libvpx/libvpx/vp9/common/vp9_reconintra.c
new file mode 100644
index 0000000..4457858
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_reconintra.c
@@ -0,0 +1,445 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#if CONFIG_VP9_HIGHBITDEPTH
+#include "vpx_dsp/vpx_dsp_common.h"
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/vpx_once.h"
+
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_onyxc_int.h"
+
+const TX_TYPE intra_mode_to_tx_type_lookup[INTRA_MODES] = {
+  DCT_DCT,    // DC
+  ADST_DCT,   // V
+  DCT_ADST,   // H
+  DCT_DCT,    // D45
+  ADST_ADST,  // D135
+  ADST_DCT,   // D117
+  DCT_ADST,   // D153
+  DCT_ADST,   // D207
+  ADST_DCT,   // D63
+  ADST_ADST,  // TM
+};
+
+enum {
+  NEED_LEFT = 1 << 1,
+  NEED_ABOVE = 1 << 2,
+  NEED_ABOVERIGHT = 1 << 3,
+};
+
+static const uint8_t extend_modes[INTRA_MODES] = {
+  NEED_ABOVE | NEED_LEFT,       // DC
+  NEED_ABOVE,                   // V
+  NEED_LEFT,                    // H
+  NEED_ABOVERIGHT,              // D45
+  NEED_LEFT | NEED_ABOVE,       // D135
+  NEED_LEFT | NEED_ABOVE,       // D117
+  NEED_LEFT | NEED_ABOVE,       // D153
+  NEED_LEFT,                    // D207
+  NEED_ABOVERIGHT,              // D63
+  NEED_LEFT | NEED_ABOVE,       // TM
+};
+
+typedef void (*intra_pred_fn)(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left);
+
+static intra_pred_fn pred[INTRA_MODES][TX_SIZES];
+static intra_pred_fn dc_pred[2][2][TX_SIZES];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*intra_high_pred_fn)(uint16_t *dst, ptrdiff_t stride,
+                                   const uint16_t *above, const uint16_t *left,
+                                   int bd);
+static intra_high_pred_fn pred_high[INTRA_MODES][4];
+static intra_high_pred_fn dc_pred_high[2][2][4];
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void vp9_init_intra_predictors_internal(void) {
+#define INIT_ALL_SIZES(p, type) \
+  p[TX_4X4] = vpx_##type##_predictor_4x4; \
+  p[TX_8X8] = vpx_##type##_predictor_8x8; \
+  p[TX_16X16] = vpx_##type##_predictor_16x16; \
+  p[TX_32X32] = vpx_##type##_predictor_32x32
+
+  INIT_ALL_SIZES(pred[V_PRED], v);
+  INIT_ALL_SIZES(pred[H_PRED], h);
+  INIT_ALL_SIZES(pred[D207_PRED], d207);
+  INIT_ALL_SIZES(pred[D45_PRED], d45);
+  INIT_ALL_SIZES(pred[D63_PRED], d63);
+  INIT_ALL_SIZES(pred[D117_PRED], d117);
+  INIT_ALL_SIZES(pred[D135_PRED], d135);
+  INIT_ALL_SIZES(pred[D153_PRED], d153);
+  INIT_ALL_SIZES(pred[TM_PRED], tm);
+
+  INIT_ALL_SIZES(dc_pred[0][0], dc_128);
+  INIT_ALL_SIZES(dc_pred[0][1], dc_top);
+  INIT_ALL_SIZES(dc_pred[1][0], dc_left);
+  INIT_ALL_SIZES(dc_pred[1][1], dc);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  INIT_ALL_SIZES(pred_high[V_PRED], highbd_v);
+  INIT_ALL_SIZES(pred_high[H_PRED], highbd_h);
+  INIT_ALL_SIZES(pred_high[D207_PRED], highbd_d207);
+  INIT_ALL_SIZES(pred_high[D45_PRED], highbd_d45);
+  INIT_ALL_SIZES(pred_high[D63_PRED], highbd_d63);
+  INIT_ALL_SIZES(pred_high[D117_PRED], highbd_d117);
+  INIT_ALL_SIZES(pred_high[D135_PRED], highbd_d135);
+  INIT_ALL_SIZES(pred_high[D153_PRED], highbd_d153);
+  INIT_ALL_SIZES(pred_high[TM_PRED], highbd_tm);
+
+  INIT_ALL_SIZES(dc_pred_high[0][0], highbd_dc_128);
+  INIT_ALL_SIZES(dc_pred_high[0][1], highbd_dc_top);
+  INIT_ALL_SIZES(dc_pred_high[1][0], highbd_dc_left);
+  INIT_ALL_SIZES(dc_pred_high[1][1], highbd_dc);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#undef intra_pred_allsizes
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void build_intra_predictors_high(const MACROBLOCKD *xd,
+                                        const uint8_t *ref8,
+                                        int ref_stride,
+                                        uint8_t *dst8,
+                                        int dst_stride,
+                                        PREDICTION_MODE mode,
+                                        TX_SIZE tx_size,
+                                        int up_available,
+                                        int left_available,
+                                        int right_available,
+                                        int x, int y,
+                                        int plane, int bd) {
+  int i;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  DECLARE_ALIGNED(16, uint16_t, left_col[32]);
+  DECLARE_ALIGNED(16, uint16_t, above_data[64 + 16]);
+  uint16_t *above_row = above_data + 16;
+  const uint16_t *const_above_row = above_row;
+  const int bs = 4 << tx_size;
+  int frame_width, frame_height;
+  int x0, y0;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int need_left = extend_modes[mode] & NEED_LEFT;
+  const int need_above = extend_modes[mode] & NEED_ABOVE;
+  const int need_aboveright = extend_modes[mode] & NEED_ABOVERIGHT;
+  int base = 128 << (bd - 8);
+  // 127 127 127 .. 127 127 127 127 127 127
+  // 129  A   B  ..  Y   Z
+  // 129  C   D  ..  W   X
+  // 129  E   F  ..  U   V
+  // 129  G   H  ..  S   T   T   T   T   T
+  // For 10 bit and 12 bit, 127 and 129 are replaced by base -1 and base + 1.
+
+  // Get current frame pointer, width and height.
+  if (plane == 0) {
+    frame_width = xd->cur_buf->y_width;
+    frame_height = xd->cur_buf->y_height;
+  } else {
+    frame_width = xd->cur_buf->uv_width;
+    frame_height = xd->cur_buf->uv_height;
+  }
+
+  // Get block position in current frame.
+  x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
+  y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
+
+  // NEED_LEFT
+  if (need_left) {
+    if (left_available) {
+      if (xd->mb_to_bottom_edge < 0) {
+        /* slower path if the block needs border extension */
+        if (y0 + bs <= frame_height) {
+          for (i = 0; i < bs; ++i)
+            left_col[i] = ref[i * ref_stride - 1];
+        } else {
+          const int extend_bottom = frame_height - y0;
+          for (i = 0; i < extend_bottom; ++i)
+            left_col[i] = ref[i * ref_stride - 1];
+          for (; i < bs; ++i)
+            left_col[i] = ref[(extend_bottom - 1) * ref_stride - 1];
+        }
+      } else {
+        /* faster path if the block does not need extension */
+        for (i = 0; i < bs; ++i)
+          left_col[i] = ref[i * ref_stride - 1];
+      }
+    } else {
+      vpx_memset16(left_col, base + 1, bs);
+    }
+  }
+
+  // NEED_ABOVE
+  if (need_above) {
+    if (up_available) {
+      const uint16_t *above_ref = ref - ref_stride;
+      if (xd->mb_to_right_edge < 0) {
+        /* slower path if the block needs border extension */
+        if (x0 + bs <= frame_width) {
+          memcpy(above_row, above_ref, bs * sizeof(above_row[0]));
+        } else if (x0 <= frame_width) {
+          const int r = frame_width - x0;
+          memcpy(above_row, above_ref, r * sizeof(above_row[0]));
+          vpx_memset16(above_row + r, above_row[r - 1], x0 + bs - frame_width);
+        }
+      } else {
+        /* faster path if the block does not need extension */
+        if (bs == 4 && right_available && left_available) {
+          const_above_row = above_ref;
+        } else {
+          memcpy(above_row, above_ref, bs * sizeof(above_row[0]));
+        }
+      }
+      above_row[-1] = left_available ? above_ref[-1] : (base + 1);
+    } else {
+      vpx_memset16(above_row, base - 1, bs);
+      above_row[-1] = base - 1;
+    }
+  }
+
+  // NEED_ABOVERIGHT
+  if (need_aboveright) {
+    if (up_available) {
+      const uint16_t *above_ref = ref - ref_stride;
+      if (xd->mb_to_right_edge < 0) {
+        /* slower path if the block needs border extension */
+        if (x0 + 2 * bs <= frame_width) {
+          if (right_available && bs == 4) {
+            memcpy(above_row, above_ref, 2 * bs * sizeof(above_row[0]));
+          } else {
+            memcpy(above_row, above_ref, bs * sizeof(above_row[0]));
+            vpx_memset16(above_row + bs, above_row[bs - 1], bs);
+          }
+        } else if (x0 + bs <= frame_width) {
+          const int r = frame_width - x0;
+          if (right_available && bs == 4) {
+            memcpy(above_row, above_ref, r * sizeof(above_row[0]));
+            vpx_memset16(above_row + r, above_row[r - 1],
+                         x0 + 2 * bs - frame_width);
+          } else {
+            memcpy(above_row, above_ref, bs * sizeof(above_row[0]));
+            vpx_memset16(above_row + bs, above_row[bs - 1], bs);
+          }
+        } else if (x0 <= frame_width) {
+          const int r = frame_width - x0;
+          memcpy(above_row, above_ref, r * sizeof(above_row[0]));
+          vpx_memset16(above_row + r, above_row[r - 1],
+                       x0 + 2 * bs - frame_width);
+        }
+        above_row[-1] = left_available ? above_ref[-1] : (base + 1);
+      } else {
+        /* faster path if the block does not need extension */
+        if (bs == 4 && right_available && left_available) {
+          const_above_row = above_ref;
+        } else {
+          memcpy(above_row, above_ref, bs * sizeof(above_row[0]));
+          if (bs == 4 && right_available)
+            memcpy(above_row + bs, above_ref + bs, bs * sizeof(above_row[0]));
+          else
+            vpx_memset16(above_row + bs, above_row[bs - 1], bs);
+          above_row[-1] = left_available ? above_ref[-1] : (base + 1);
+        }
+      }
+    } else {
+      vpx_memset16(above_row, base - 1, bs * 2);
+      above_row[-1] = base - 1;
+    }
+  }
+
+  // predict
+  if (mode == DC_PRED) {
+    dc_pred_high[left_available][up_available][tx_size](dst, dst_stride,
+                                                        const_above_row,
+                                                        left_col, xd->bd);
+  } else {
+    pred_high[mode][tx_size](dst, dst_stride, const_above_row, left_col,
+                             xd->bd);
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void build_intra_predictors(const MACROBLOCKD *xd, const uint8_t *ref,
+                                   int ref_stride, uint8_t *dst, int dst_stride,
+                                   PREDICTION_MODE mode, TX_SIZE tx_size,
+                                   int up_available, int left_available,
+                                   int right_available, int x, int y,
+                                   int plane) {
+  int i;
+  DECLARE_ALIGNED(16, uint8_t, left_col[32]);
+  DECLARE_ALIGNED(16, uint8_t, above_data[64 + 16]);
+  uint8_t *above_row = above_data + 16;
+  const uint8_t *const_above_row = above_row;
+  const int bs = 4 << tx_size;
+  int frame_width, frame_height;
+  int x0, y0;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+
+  // 127 127 127 .. 127 127 127 127 127 127
+  // 129  A   B  ..  Y   Z
+  // 129  C   D  ..  W   X
+  // 129  E   F  ..  U   V
+  // 129  G   H  ..  S   T   T   T   T   T
+  // ..
+
+  // Get current frame pointer, width and height.
+  if (plane == 0) {
+    frame_width = xd->cur_buf->y_width;
+    frame_height = xd->cur_buf->y_height;
+  } else {
+    frame_width = xd->cur_buf->uv_width;
+    frame_height = xd->cur_buf->uv_height;
+  }
+
+  // Get block position in current frame.
+  x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
+  y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
+
+  // NEED_LEFT
+  if (extend_modes[mode] & NEED_LEFT) {
+    if (left_available) {
+      if (xd->mb_to_bottom_edge < 0) {
+        /* slower path if the block needs border extension */
+        if (y0 + bs <= frame_height) {
+          for (i = 0; i < bs; ++i)
+            left_col[i] = ref[i * ref_stride - 1];
+        } else {
+          const int extend_bottom = frame_height - y0;
+          for (i = 0; i < extend_bottom; ++i)
+            left_col[i] = ref[i * ref_stride - 1];
+          for (; i < bs; ++i)
+            left_col[i] = ref[(extend_bottom - 1) * ref_stride - 1];
+        }
+      } else {
+        /* faster path if the block does not need extension */
+        for (i = 0; i < bs; ++i)
+          left_col[i] = ref[i * ref_stride - 1];
+      }
+    } else {
+      memset(left_col, 129, bs);
+    }
+  }
+
+  // NEED_ABOVE
+  if (extend_modes[mode] & NEED_ABOVE) {
+    if (up_available) {
+      const uint8_t *above_ref = ref - ref_stride;
+      if (xd->mb_to_right_edge < 0) {
+        /* slower path if the block needs border extension */
+        if (x0 + bs <= frame_width) {
+          memcpy(above_row, above_ref, bs);
+        } else if (x0 <= frame_width) {
+          const int r = frame_width - x0;
+          memcpy(above_row, above_ref, r);
+          memset(above_row + r, above_row[r - 1], x0 + bs - frame_width);
+        }
+      } else {
+        /* faster path if the block does not need extension */
+        if (bs == 4 && right_available && left_available) {
+          const_above_row = above_ref;
+        } else {
+          memcpy(above_row, above_ref, bs);
+        }
+      }
+      above_row[-1] = left_available ? above_ref[-1] : 129;
+    } else {
+      memset(above_row, 127, bs);
+      above_row[-1] = 127;
+    }
+  }
+
+  // NEED_ABOVERIGHT
+  if (extend_modes[mode] & NEED_ABOVERIGHT) {
+    if (up_available) {
+      const uint8_t *above_ref = ref - ref_stride;
+      if (xd->mb_to_right_edge < 0) {
+        /* slower path if the block needs border extension */
+        if (x0 + 2 * bs <= frame_width) {
+          if (right_available && bs == 4) {
+            memcpy(above_row, above_ref, 2 * bs);
+          } else {
+            memcpy(above_row, above_ref, bs);
+            memset(above_row + bs, above_row[bs - 1], bs);
+          }
+        } else if (x0 + bs <= frame_width) {
+          const int r = frame_width - x0;
+          if (right_available && bs == 4) {
+            memcpy(above_row, above_ref, r);
+            memset(above_row + r, above_row[r - 1], x0 + 2 * bs - frame_width);
+          } else {
+            memcpy(above_row, above_ref, bs);
+            memset(above_row + bs, above_row[bs - 1], bs);
+          }
+        } else if (x0 <= frame_width) {
+          const int r = frame_width - x0;
+          memcpy(above_row, above_ref, r);
+          memset(above_row + r, above_row[r - 1], x0 + 2 * bs - frame_width);
+        }
+      } else {
+        /* faster path if the block does not need extension */
+        if (bs == 4 && right_available && left_available) {
+          const_above_row = above_ref;
+        } else {
+          memcpy(above_row, above_ref, bs);
+          if (bs == 4 && right_available)
+            memcpy(above_row + bs, above_ref + bs, bs);
+          else
+            memset(above_row + bs, above_row[bs - 1], bs);
+        }
+      }
+      above_row[-1] = left_available ? above_ref[-1] : 129;
+    } else {
+      memset(above_row, 127, bs * 2);
+      above_row[-1] = 127;
+    }
+  }
+
+  // predict
+  if (mode == DC_PRED) {
+    dc_pred[left_available][up_available][tx_size](dst, dst_stride,
+                                                   const_above_row, left_col);
+  } else {
+    pred[mode][tx_size](dst, dst_stride, const_above_row, left_col);
+  }
+}
+
+void vp9_predict_intra_block(const MACROBLOCKD *xd, int bwl_in,
+                             TX_SIZE tx_size, PREDICTION_MODE mode,
+                             const uint8_t *ref, int ref_stride,
+                             uint8_t *dst, int dst_stride,
+                             int aoff, int loff, int plane) {
+  const int bw = (1 << bwl_in);
+  const int txw = (1 << tx_size);
+  const int have_top = loff || (xd->above_mi != NULL);
+  const int have_left = aoff || (xd->left_mi != NULL);
+  const int have_right = (aoff + txw) < bw;
+  const int x = aoff * 4;
+  const int y = loff * 4;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    build_intra_predictors_high(xd, ref, ref_stride, dst, dst_stride, mode,
+                                tx_size, have_top, have_left, have_right,
+                                x, y, plane, xd->bd);
+    return;
+  }
+#endif
+  build_intra_predictors(xd, ref, ref_stride, dst, dst_stride, mode, tx_size,
+                         have_top, have_left, have_right, x, y, plane);
+}
+
+void vp9_init_intra_predictors(void) {
+  once(vp9_init_intra_predictors_internal);
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_reconintra.h b/libvpx/libvpx/vp9/common/vp9_reconintra.h
new file mode 100644
index 0000000..de45380
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_reconintra.h
@@ -0,0 +1,32 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_RECONINTRA_H_
+#define VP9_COMMON_VP9_RECONINTRA_H_
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_init_intra_predictors(void);
+
+void vp9_predict_intra_block(const MACROBLOCKD *xd, int bwl_in,
+                             TX_SIZE tx_size, PREDICTION_MODE mode,
+                             const uint8_t *ref, int ref_stride,
+                             uint8_t *dst, int dst_stride,
+                             int aoff, int loff, int plane);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_RECONINTRA_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_rtcd.c b/libvpx/libvpx/vp9/common/vp9_rtcd.c
new file mode 100644
index 0000000..2dfa09f
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_rtcd.c
@@ -0,0 +1,19 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_config.h"
+#define RTCD_C
+#include "./vp9_rtcd.h"
+#include "vpx_ports/vpx_once.h"
+
+void vp9_rtcd() {
+    // TODO(JBB): Remove this once, by insuring that both the encoder and
+    // decoder setup functions are protected by once();
+    once(setup_rtcd_internal);
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_rtcd_defs.pl b/libvpx/libvpx/vp9/common/vp9_rtcd_defs.pl
new file mode 100644
index 0000000..8461336
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_rtcd_defs.pl
@@ -0,0 +1,317 @@
+sub vp9_common_forward_decls() {
+print <<EOF
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+EOF
+}
+forward_decls qw/vp9_common_forward_decls/;
+
+# x86inc.asm had specific constraints. break it out so it's easy to disable.
+# zero all the variables to avoid tricky else conditions.
+$mmx_x86inc = $sse_x86inc = $sse2_x86inc = $ssse3_x86inc = $avx_x86inc =
+  $avx2_x86inc = '';
+$mmx_x86_64_x86inc = $sse_x86_64_x86inc = $sse2_x86_64_x86inc =
+  $ssse3_x86_64_x86inc = $avx_x86_64_x86inc = $avx2_x86_64_x86inc = '';
+if (vpx_config("CONFIG_USE_X86INC") eq "yes") {
+  $mmx_x86inc = 'mmx';
+  $sse_x86inc = 'sse';
+  $sse2_x86inc = 'sse2';
+  $ssse3_x86inc = 'ssse3';
+  $avx_x86inc = 'avx';
+  $avx2_x86inc = 'avx2';
+  if ($opts{arch} eq "x86_64") {
+    $mmx_x86_64_x86inc = 'mmx';
+    $sse_x86_64_x86inc = 'sse';
+    $sse2_x86_64_x86inc = 'sse2';
+    $ssse3_x86_64_x86inc = 'ssse3';
+    $avx_x86_64_x86inc = 'avx';
+    $avx2_x86_64_x86inc = 'avx2';
+  }
+}
+
+# functions that are 64 bit only.
+$mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = '';
+if ($opts{arch} eq "x86_64") {
+  $mmx_x86_64 = 'mmx';
+  $sse2_x86_64 = 'sse2';
+  $ssse3_x86_64 = 'ssse3';
+  $avx_x86_64 = 'avx';
+  $avx2_x86_64 = 'avx2';
+}
+
+#
+# post proc
+#
+if (vpx_config("CONFIG_VP9_POSTPROC") eq "yes") {
+add_proto qw/void vp9_mbpost_proc_down/, "uint8_t *dst, int pitch, int rows, int cols, int flimit";
+specialize qw/vp9_mbpost_proc_down sse2/;
+$vp9_mbpost_proc_down_sse2=vp9_mbpost_proc_down_xmm;
+
+add_proto qw/void vp9_mbpost_proc_across_ip/, "uint8_t *src, int pitch, int rows, int cols, int flimit";
+specialize qw/vp9_mbpost_proc_across_ip sse2/;
+$vp9_mbpost_proc_across_ip_sse2=vp9_mbpost_proc_across_ip_xmm;
+
+add_proto qw/void vp9_post_proc_down_and_across/, "const uint8_t *src_ptr, uint8_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit";
+specialize qw/vp9_post_proc_down_and_across sse2/;
+$vp9_post_proc_down_and_across_sse2=vp9_post_proc_down_and_across_xmm;
+
+add_proto qw/void vp9_filter_by_weight16x16/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight";
+specialize qw/vp9_filter_by_weight16x16 sse2 msa/;
+
+add_proto qw/void vp9_filter_by_weight8x8/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight";
+specialize qw/vp9_filter_by_weight8x8 sse2 msa/;
+}
+
+#
+# dct
+#
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
+  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
+    add_proto qw/void vp9_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp9_iht4x4_16_add/;
+
+    add_proto qw/void vp9_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp9_iht8x8_64_add/;
+
+    add_proto qw/void vp9_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+    specialize qw/vp9_iht16x16_256_add/;
+  } else {
+    add_proto qw/void vp9_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp9_iht4x4_16_add sse2/;
+
+    add_proto qw/void vp9_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp9_iht8x8_64_add sse2/;
+
+    add_proto qw/void vp9_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+    specialize qw/vp9_iht16x16_256_add sse2/;
+  }
+} else {
+  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
+  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
+    add_proto qw/void vp9_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp9_iht4x4_16_add/;
+
+    add_proto qw/void vp9_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp9_iht8x8_64_add/;
+
+    add_proto qw/void vp9_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+    specialize qw/vp9_iht16x16_256_add/;
+  } else {
+    add_proto qw/void vp9_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp9_iht4x4_16_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void vp9_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp9_iht8x8_64_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void vp9_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+    specialize qw/vp9_iht16x16_256_add sse2 dspr2 msa/;
+  }
+}
+
+# High bitdepth functions
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  #
+  # Sub Pixel Filters
+  #
+  add_proto qw/void vp9_highbd_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_highbd_convolve_copy/;
+
+  add_proto qw/void vp9_highbd_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_highbd_convolve_avg/;
+
+  add_proto qw/void vp9_highbd_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_highbd_convolve8/, "$sse2_x86_64";
+
+  add_proto qw/void vp9_highbd_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_highbd_convolve8_horiz/, "$sse2_x86_64";
+
+  add_proto qw/void vp9_highbd_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_highbd_convolve8_vert/, "$sse2_x86_64";
+
+  add_proto qw/void vp9_highbd_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_highbd_convolve8_avg/, "$sse2_x86_64";
+
+  add_proto qw/void vp9_highbd_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_highbd_convolve8_avg_horiz/, "$sse2_x86_64";
+
+  add_proto qw/void vp9_highbd_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_highbd_convolve8_avg_vert/, "$sse2_x86_64";
+
+  #
+  # post proc
+  #
+  if (vpx_config("CONFIG_VP9_POSTPROC") eq "yes") {
+    add_proto qw/void vp9_highbd_mbpost_proc_down/, "uint16_t *dst, int pitch, int rows, int cols, int flimit";
+    specialize qw/vp9_highbd_mbpost_proc_down/;
+
+    add_proto qw/void vp9_highbd_mbpost_proc_across_ip/, "uint16_t *src, int pitch, int rows, int cols, int flimit";
+    specialize qw/vp9_highbd_mbpost_proc_across_ip/;
+
+    add_proto qw/void vp9_highbd_post_proc_down_and_across/, "const uint16_t *src_ptr, uint16_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit";
+    specialize qw/vp9_highbd_post_proc_down_and_across/;
+  }
+
+  #
+  # dct
+  #
+  # Note as optimized versions of these functions are added we need to add a check to ensure
+  # that when CONFIG_EMULATE_HARDWARE is on, it defaults to the C versions only.
+  add_proto qw/void vp9_highbd_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+  specialize qw/vp9_highbd_iht4x4_16_add/;
+
+  add_proto qw/void vp9_highbd_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+  specialize qw/vp9_highbd_iht8x8_64_add/;
+
+  add_proto qw/void vp9_highbd_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type, int bd";
+  specialize qw/vp9_highbd_iht16x16_256_add/;
+}
+
+#
+# Encoder functions below this point.
+#
+if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
+
+# ENCODEMB INVOKE
+
+#
+# Denoiser
+#
+if (vpx_config("CONFIG_VP9_TEMPORAL_DENOISING") eq "yes") {
+  add_proto qw/int vp9_denoiser_filter/, "const uint8_t *sig, int sig_stride, const uint8_t *mc_avg, int mc_avg_stride, uint8_t *avg, int avg_stride, int increase_denoising, BLOCK_SIZE bs, int motion_magnitude";
+  specialize qw/vp9_denoiser_filter sse2/;
+}
+
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/int64_t vp9_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
+  specialize qw/vp9_block_error/;
+
+  add_proto qw/int64_t vp9_highbd_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd";
+  specialize qw/vp9_highbd_block_error/, "$sse2_x86inc";
+
+  add_proto qw/int64_t vp9_highbd_block_error_8bit/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
+  specialize qw/vp9_highbd_block_error_8bit/, "$sse2_x86inc", "$avx_x86inc";
+
+  add_proto qw/void vp9_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  specialize qw/vp9_quantize_fp/;
+
+  add_proto qw/void vp9_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  specialize qw/vp9_quantize_fp_32x32/;
+
+  add_proto qw/void vp9_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  specialize qw/vp9_fdct8x8_quant/;
+} else {
+  add_proto qw/int64_t vp9_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
+  specialize qw/vp9_block_error avx2 msa/, "$sse2_x86inc";
+
+  add_proto qw/int64_t vp9_block_error_fp/, "const int16_t *coeff, const int16_t *dqcoeff, int block_size";
+  specialize qw/vp9_block_error_fp neon/, "$sse2_x86inc";
+
+  add_proto qw/void vp9_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  specialize qw/vp9_quantize_fp neon sse2/, "$ssse3_x86_64_x86inc";
+
+  add_proto qw/void vp9_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  specialize qw/vp9_quantize_fp_32x32/, "$ssse3_x86_64_x86inc";
+
+  add_proto qw/void vp9_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  specialize qw/vp9_fdct8x8_quant sse2 ssse3 neon/;
+}
+
+# fdct functions
+
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void vp9_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/vp9_fht4x4 sse2/;
+
+  add_proto qw/void vp9_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/vp9_fht8x8 sse2/;
+
+  add_proto qw/void vp9_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/vp9_fht16x16 sse2/;
+
+  add_proto qw/void vp9_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vp9_fwht4x4/, "$sse2_x86inc";
+} else {
+  add_proto qw/void vp9_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/vp9_fht4x4 sse2 msa/;
+
+  add_proto qw/void vp9_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/vp9_fht8x8 sse2 msa/;
+
+  add_proto qw/void vp9_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/vp9_fht16x16 sse2 msa/;
+
+  add_proto qw/void vp9_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vp9_fwht4x4 msa/, "$sse2_x86inc";
+}
+
+#
+# Motion search
+#
+add_proto qw/int vp9_full_search_sad/, "const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv";
+specialize qw/vp9_full_search_sad sse3 sse4_1/;
+$vp9_full_search_sad_sse3=vp9_full_search_sadx3;
+$vp9_full_search_sad_sse4_1=vp9_full_search_sadx8;
+
+add_proto qw/int vp9_diamond_search_sad/, "const struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct vp9_variance_vtable *fn_ptr, const struct mv *center_mv";
+specialize qw/vp9_diamond_search_sad avx/;
+
+add_proto qw/void vp9_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
+specialize qw/vp9_temporal_filter_apply sse2 msa/;
+
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+
+  # ENCODEMB INVOKE
+
+  add_proto qw/void vp9_highbd_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  specialize qw/vp9_highbd_quantize_fp/;
+
+  add_proto qw/void vp9_highbd_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  specialize qw/vp9_highbd_quantize_fp_32x32/;
+
+  # fdct functions
+  add_proto qw/void vp9_highbd_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/vp9_highbd_fht4x4/;
+
+  add_proto qw/void vp9_highbd_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/vp9_highbd_fht8x8/;
+
+  add_proto qw/void vp9_highbd_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/vp9_highbd_fht16x16/;
+
+  add_proto qw/void vp9_highbd_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vp9_highbd_fwht4x4/;
+
+  add_proto qw/void vp9_highbd_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
+  specialize qw/vp9_highbd_temporal_filter_apply/;
+
+}
+# End vp9_high encoder functions
+
+#
+# frame based scale
+#
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+} else {
+  add_proto qw/void vp9_scale_and_extend_frame/, "const struct yv12_buffer_config *src, struct yv12_buffer_config *dst";
+  specialize qw/vp9_scale_and_extend_frame ssse3/;
+}
+
+}
+# end encoder functions
+1;
diff --git a/libvpx/libvpx/vp9/common/vp9_scale.c b/libvpx/libvpx/vp9/common/vp9_scale.c
new file mode 100644
index 0000000..b763b92
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_scale.c
@@ -0,0 +1,175 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vp9/common/vp9_filter.h"
+#include "vp9/common/vp9_scale.h"
+#include "vpx_dsp/vpx_filter.h"
+
+static INLINE int scaled_x(int val, const struct scale_factors *sf) {
+  return (int)((int64_t)val * sf->x_scale_fp >> REF_SCALE_SHIFT);
+}
+
+static INLINE int scaled_y(int val, const struct scale_factors *sf) {
+  return (int)((int64_t)val * sf->y_scale_fp >> REF_SCALE_SHIFT);
+}
+
+static int unscaled_value(int val, const struct scale_factors *sf) {
+  (void) sf;
+  return val;
+}
+
+static int get_fixed_point_scale_factor(int other_size, int this_size) {
+  // Calculate scaling factor once for each reference frame
+  // and use fixed point scaling factors in decoding and encoding routines.
+  // Hardware implementations can calculate scale factor in device driver
+  // and use multiplication and shifting on hardware instead of division.
+  return (other_size << REF_SCALE_SHIFT) / this_size;
+}
+
+MV32 vp9_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf) {
+  const int x_off_q4 = scaled_x(x << SUBPEL_BITS, sf) & SUBPEL_MASK;
+  const int y_off_q4 = scaled_y(y << SUBPEL_BITS, sf) & SUBPEL_MASK;
+  const MV32 res = {
+    scaled_y(mv->row, sf) + y_off_q4,
+    scaled_x(mv->col, sf) + x_off_q4
+  };
+  return res;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+                                       int other_w, int other_h,
+                                       int this_w, int this_h,
+                                       int use_highbd) {
+#else
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+                                       int other_w, int other_h,
+                                       int this_w, int this_h) {
+#endif
+  if (!valid_ref_frame_size(other_w, other_h, this_w, this_h)) {
+    sf->x_scale_fp = REF_INVALID_SCALE;
+    sf->y_scale_fp = REF_INVALID_SCALE;
+    return;
+  }
+
+  sf->x_scale_fp = get_fixed_point_scale_factor(other_w, this_w);
+  sf->y_scale_fp = get_fixed_point_scale_factor(other_h, this_h);
+  sf->x_step_q4 = scaled_x(16, sf);
+  sf->y_step_q4 = scaled_y(16, sf);
+
+  if (vp9_is_scaled(sf)) {
+    sf->scale_value_x = scaled_x;
+    sf->scale_value_y = scaled_y;
+  } else {
+    sf->scale_value_x = unscaled_value;
+    sf->scale_value_y = unscaled_value;
+  }
+
+  // TODO(agrange): Investigate the best choice of functions to use here
+  // for EIGHTTAP_SMOOTH. Since it is not interpolating, need to choose what
+  // to do at full-pel offsets. The current selection, where the filter is
+  // applied in one direction only, and not at all for 0,0, seems to give the
+  // best quality, but it may be worth trying an additional mode that does
+  // do the filtering on full-pel.
+
+  if (sf->x_step_q4 == 16) {
+    if (sf->y_step_q4 == 16) {
+      // No scaling in either direction.
+      sf->predict[0][0][0] = vpx_convolve_copy;
+      sf->predict[0][0][1] = vpx_convolve_avg;
+      sf->predict[0][1][0] = vpx_convolve8_vert;
+      sf->predict[0][1][1] = vpx_convolve8_avg_vert;
+      sf->predict[1][0][0] = vpx_convolve8_horiz;
+      sf->predict[1][0][1] = vpx_convolve8_avg_horiz;
+    } else {
+      // No scaling in x direction. Must always scale in the y direction.
+      sf->predict[0][0][0] = vpx_scaled_vert;
+      sf->predict[0][0][1] = vpx_scaled_avg_vert;
+      sf->predict[0][1][0] = vpx_scaled_vert;
+      sf->predict[0][1][1] = vpx_scaled_avg_vert;
+      sf->predict[1][0][0] = vpx_scaled_2d;
+      sf->predict[1][0][1] = vpx_scaled_avg_2d;
+    }
+  } else {
+    if (sf->y_step_q4 == 16) {
+      // No scaling in the y direction. Must always scale in the x direction.
+      sf->predict[0][0][0] = vpx_scaled_horiz;
+      sf->predict[0][0][1] = vpx_scaled_avg_horiz;
+      sf->predict[0][1][0] = vpx_scaled_2d;
+      sf->predict[0][1][1] = vpx_scaled_avg_2d;
+      sf->predict[1][0][0] = vpx_scaled_horiz;
+      sf->predict[1][0][1] = vpx_scaled_avg_horiz;
+    } else {
+      // Must always scale in both directions.
+      sf->predict[0][0][0] = vpx_scaled_2d;
+      sf->predict[0][0][1] = vpx_scaled_avg_2d;
+      sf->predict[0][1][0] = vpx_scaled_2d;
+      sf->predict[0][1][1] = vpx_scaled_avg_2d;
+      sf->predict[1][0][0] = vpx_scaled_2d;
+      sf->predict[1][0][1] = vpx_scaled_avg_2d;
+    }
+  }
+
+  // 2D subpel motion always gets filtered in both directions
+
+  if ((sf->x_step_q4 != 16) || (sf->y_step_q4 != 16)) {
+    sf->predict[1][1][0] = vpx_scaled_2d;
+    sf->predict[1][1][1] = vpx_scaled_avg_2d;
+  } else {
+    sf->predict[1][1][0] = vpx_convolve8;
+    sf->predict[1][1][1] = vpx_convolve8_avg;
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (use_highbd) {
+    if (sf->x_step_q4 == 16) {
+      if (sf->y_step_q4 == 16) {
+        // No scaling in either direction.
+        sf->highbd_predict[0][0][0] = vpx_highbd_convolve_copy;
+        sf->highbd_predict[0][0][1] = vpx_highbd_convolve_avg;
+        sf->highbd_predict[0][1][0] = vpx_highbd_convolve8_vert;
+        sf->highbd_predict[0][1][1] = vpx_highbd_convolve8_avg_vert;
+        sf->highbd_predict[1][0][0] = vpx_highbd_convolve8_horiz;
+        sf->highbd_predict[1][0][1] = vpx_highbd_convolve8_avg_horiz;
+      } else {
+        // No scaling in x direction. Must always scale in the y direction.
+        sf->highbd_predict[0][0][0] = vpx_highbd_convolve8_vert;
+        sf->highbd_predict[0][0][1] = vpx_highbd_convolve8_avg_vert;
+        sf->highbd_predict[0][1][0] = vpx_highbd_convolve8_vert;
+        sf->highbd_predict[0][1][1] = vpx_highbd_convolve8_avg_vert;
+        sf->highbd_predict[1][0][0] = vpx_highbd_convolve8;
+        sf->highbd_predict[1][0][1] = vpx_highbd_convolve8_avg;
+      }
+    } else {
+      if (sf->y_step_q4 == 16) {
+        // No scaling in the y direction. Must always scale in the x direction.
+        sf->highbd_predict[0][0][0] = vpx_highbd_convolve8_horiz;
+        sf->highbd_predict[0][0][1] = vpx_highbd_convolve8_avg_horiz;
+        sf->highbd_predict[0][1][0] = vpx_highbd_convolve8;
+        sf->highbd_predict[0][1][1] = vpx_highbd_convolve8_avg;
+        sf->highbd_predict[1][0][0] = vpx_highbd_convolve8_horiz;
+        sf->highbd_predict[1][0][1] = vpx_highbd_convolve8_avg_horiz;
+      } else {
+        // Must always scale in both directions.
+        sf->highbd_predict[0][0][0] = vpx_highbd_convolve8;
+        sf->highbd_predict[0][0][1] = vpx_highbd_convolve8_avg;
+        sf->highbd_predict[0][1][0] = vpx_highbd_convolve8;
+        sf->highbd_predict[0][1][1] = vpx_highbd_convolve8_avg;
+        sf->highbd_predict[1][0][0] = vpx_highbd_convolve8;
+        sf->highbd_predict[1][0][1] = vpx_highbd_convolve8_avg;
+      }
+    }
+    // 2D subpel motion always gets filtered in both directions.
+    sf->highbd_predict[1][1][0] = vpx_highbd_convolve8;
+    sf->highbd_predict[1][1][1] = vpx_highbd_convolve8_avg;
+  }
+#endif
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_scale.h b/libvpx/libvpx/vp9/common/vp9_scale.h
new file mode 100644
index 0000000..5e91041
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_scale.h
@@ -0,0 +1,75 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_SCALE_H_
+#define VP9_COMMON_VP9_SCALE_H_
+
+#include "vp9/common/vp9_mv.h"
+#include "vpx_dsp/vpx_convolve.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define REF_SCALE_SHIFT 14
+#define REF_NO_SCALE (1 << REF_SCALE_SHIFT)
+#define REF_INVALID_SCALE -1
+
+struct scale_factors {
+  int x_scale_fp;   // horizontal fixed point scale factor
+  int y_scale_fp;   // vertical fixed point scale factor
+  int x_step_q4;
+  int y_step_q4;
+
+  int (*scale_value_x)(int val, const struct scale_factors *sf);
+  int (*scale_value_y)(int val, const struct scale_factors *sf);
+
+  convolve_fn_t predict[2][2][2];  // horiz, vert, avg
+#if CONFIG_VP9_HIGHBITDEPTH
+  highbd_convolve_fn_t highbd_predict[2][2][2];  // horiz, vert, avg
+#endif
+};
+
+MV32 vp9_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+                                       int other_w, int other_h,
+                                       int this_w, int this_h,
+                                       int use_high);
+#else
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+                                       int other_w, int other_h,
+                                       int this_w, int this_h);
+#endif
+
+static INLINE int vp9_is_valid_scale(const struct scale_factors *sf) {
+  return sf->x_scale_fp != REF_INVALID_SCALE &&
+         sf->y_scale_fp != REF_INVALID_SCALE;
+}
+
+static INLINE int vp9_is_scaled(const struct scale_factors *sf) {
+  return vp9_is_valid_scale(sf) &&
+         (sf->x_scale_fp != REF_NO_SCALE || sf->y_scale_fp != REF_NO_SCALE);
+}
+
+static INLINE int valid_ref_frame_size(int ref_width, int ref_height,
+                                      int this_width, int this_height) {
+  return 2 * this_width >= ref_width &&
+         2 * this_height >= ref_height &&
+         this_width <= 16 * ref_width &&
+         this_height <= 16 * ref_height;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_SCALE_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_scan.c b/libvpx/libvpx/vp9/common/vp9_scan.c
new file mode 100644
index 0000000..8b8b09f
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_scan.c
@@ -0,0 +1,725 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_scan.h"
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_4x4[16]) = {
+  0,  4,  1,  5,
+  8,  2, 12,  9,
+  3,  6, 13, 10,
+  7, 14, 11, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, col_scan_4x4[16]) = {
+  0,  4,  8,  1,
+  12,  5,  9,  2,
+  13,  6, 10,  3,
+  7, 14, 11, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, row_scan_4x4[16]) = {
+  0,  1,  4,  2,
+  5,  3,  6,  8,
+  9,  7, 12, 10,
+  13, 11, 14, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_8x8[64]) = {
+  0,  8,  1, 16,  9,  2, 17, 24,
+  10,  3, 18, 25, 32, 11,  4, 26,
+  33, 19, 40, 12, 34, 27,  5, 41,
+  20, 48, 13, 35, 42, 28, 21,  6,
+  49, 56, 36, 43, 29,  7, 14, 50,
+  57, 44, 22, 37, 15, 51, 58, 30,
+  45, 23, 52, 59, 38, 31, 60, 53,
+  46, 39, 61, 54, 47, 62, 55, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, col_scan_8x8[64]) = {
+  0,  8, 16,  1, 24,  9, 32, 17,
+  2, 40, 25, 10, 33, 18, 48,  3,
+  26, 41, 11, 56, 19, 34,  4, 49,
+  27, 42, 12, 35, 20, 57, 50, 28,
+  5, 43, 13, 36, 58, 51, 21, 44,
+  6, 29, 59, 37, 14, 52, 22,  7,
+  45, 60, 30, 15, 38, 53, 23, 46,
+  31, 61, 39, 54, 47, 62, 55, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, row_scan_8x8[64]) = {
+  0,  1,  2,  8,  9,  3, 16, 10,
+  4, 17, 11, 24,  5, 18, 25, 12,
+  19, 26, 32,  6, 13, 20, 33, 27,
+  7, 34, 40, 21, 28, 41, 14, 35,
+  48, 42, 29, 36, 49, 22, 43, 15,
+  56, 37, 50, 44, 30, 57, 23, 51,
+  58, 45, 38, 52, 31, 59, 53, 46,
+  60, 39, 61, 47, 54, 55, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_16x16[256]) = {
+  0, 16, 1, 32, 17, 2, 48, 33, 18, 3, 64, 34, 49, 19, 65, 80,
+  50, 4, 35, 66, 20, 81, 96, 51, 5, 36, 82, 97, 67, 112, 21, 52,
+  98, 37, 83, 113, 6, 68, 128, 53, 22, 99, 114, 84, 7, 129, 38, 69,
+  100, 115, 144, 130, 85, 54, 23, 8, 145, 39, 70, 116, 101, 131, 160, 146,
+  55, 86, 24, 71, 132, 117, 161, 40, 9, 102, 147, 176, 162, 87, 56, 25,
+  133, 118, 177, 148, 72, 103, 41, 163, 10, 192, 178, 88, 57, 134, 149, 119,
+  26, 164, 73, 104, 193, 42, 179, 208, 11, 135, 89, 165, 120, 150, 58, 194,
+  180, 27, 74, 209, 105, 151, 136, 43, 90, 224, 166, 195, 181, 121, 210, 59,
+  12, 152, 106, 167, 196, 75, 137, 225, 211, 240, 182, 122, 91, 28, 197, 13,
+  226, 168, 183, 153, 44, 212, 138, 107, 241, 60, 29, 123, 198, 184, 227, 169,
+  242, 76, 213, 154, 45, 92, 14, 199, 139, 61, 228, 214, 170, 185, 243, 108,
+  77, 155, 30, 15, 200, 229, 124, 215, 244, 93, 46, 186, 171, 201, 109, 140,
+  230, 62, 216, 245, 31, 125, 78, 156, 231, 47, 187, 202, 217, 94, 246, 141,
+  63, 232, 172, 110, 247, 157, 79, 218, 203, 126, 233, 188, 248, 95, 173, 142,
+  219, 111, 249, 234, 158, 127, 189, 204, 250, 235, 143, 174, 220, 205, 159,
+  251,
+  190, 221, 175, 236, 237, 191, 206, 252, 222, 253, 207, 238, 223, 254, 239,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, col_scan_16x16[256]) = {
+  0, 16, 32, 48, 1, 64, 17, 80, 33, 96, 49, 2, 65, 112, 18, 81,
+  34, 128, 50, 97, 3, 66, 144, 19, 113, 35, 82, 160, 98, 51, 129, 4,
+  67, 176, 20, 114, 145, 83, 36, 99, 130, 52, 192, 5, 161, 68, 115, 21,
+  146, 84, 208, 177, 37, 131, 100, 53, 162, 224, 69, 6, 116, 193, 147, 85,
+  22, 240, 132, 38, 178, 101, 163, 54, 209, 117, 70, 7, 148, 194, 86, 179,
+  225, 23, 133, 39, 164, 8, 102, 210, 241, 55, 195, 118, 149, 71, 180, 24,
+  87, 226, 134, 165, 211, 40, 103, 56, 72, 150, 196, 242, 119, 9, 181, 227,
+  88, 166, 25, 135, 41, 104, 212, 57, 151, 197, 120, 73, 243, 182, 136, 167,
+  213, 89, 10, 228, 105, 152, 198, 26, 42, 121, 183, 244, 168, 58, 137, 229,
+  74, 214, 90, 153, 199, 184, 11, 106, 245, 27, 122, 230, 169, 43, 215, 59,
+  200, 138, 185, 246, 75, 12, 91, 154, 216, 231, 107, 28, 44, 201, 123, 170,
+  60, 247, 232, 76, 139, 13, 92, 217, 186, 248, 155, 108, 29, 124, 45, 202,
+  233, 171, 61, 14, 77, 140, 15, 249, 93, 30, 187, 156, 218, 46, 109, 125,
+  62, 172, 78, 203, 31, 141, 234, 94, 47, 188, 63, 157, 110, 250, 219, 79,
+  126, 204, 173, 142, 95, 189, 111, 235, 158, 220, 251, 127, 174, 143, 205,
+  236,
+  159, 190, 221, 252, 175, 206, 237, 191, 253, 222, 238, 207, 254, 223, 239,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, row_scan_16x16[256]) = {
+  0, 1, 2, 16, 3, 17, 4, 18, 32, 5, 33, 19, 6, 34, 48, 20,
+  49, 7, 35, 21, 50, 64, 8, 36, 65, 22, 51, 37, 80, 9, 66, 52,
+  23, 38, 81, 67, 10, 53, 24, 82, 68, 96, 39, 11, 54, 83, 97, 69,
+  25, 98, 84, 40, 112, 55, 12, 70, 99, 113, 85, 26, 41, 56, 114, 100,
+  13, 71, 128, 86, 27, 115, 101, 129, 42, 57, 72, 116, 14, 87, 130, 102,
+  144, 73, 131, 117, 28, 58, 15, 88, 43, 145, 103, 132, 146, 118, 74, 160,
+  89, 133, 104, 29, 59, 147, 119, 44, 161, 148, 90, 105, 134, 162, 120, 176,
+  75, 135, 149, 30, 60, 163, 177, 45, 121, 91, 106, 164, 178, 150, 192, 136,
+  165, 179, 31, 151, 193, 76, 122, 61, 137, 194, 107, 152, 180, 208, 46, 166,
+  167, 195, 92, 181, 138, 209, 123, 153, 224, 196, 77, 168, 210, 182, 240, 108,
+  197, 62, 154, 225, 183, 169, 211, 47, 139, 93, 184, 226, 212, 241, 198, 170,
+  124, 155, 199, 78, 213, 185, 109, 227, 200, 63, 228, 242, 140, 214, 171, 186,
+  156, 229, 243, 125, 94, 201, 244, 215, 216, 230, 141, 187, 202, 79, 172, 110,
+  157, 245, 217, 231, 95, 246, 232, 126, 203, 247, 233, 173, 218, 142, 111,
+  158,
+  188, 248, 127, 234, 219, 249, 189, 204, 143, 174, 159, 250, 235, 205, 220,
+  175,
+  190, 251, 221, 191, 206, 236, 207, 237, 252, 222, 253, 223, 238, 239, 254,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_32x32[1024]) = {
+  0, 32, 1, 64, 33, 2, 96, 65, 34, 128, 3, 97, 66, 160,
+  129, 35, 98, 4, 67, 130, 161, 192, 36, 99, 224, 5, 162, 193,
+  68, 131, 37, 100,
+  225, 194, 256, 163, 69, 132, 6, 226, 257, 288, 195, 101, 164, 38,
+  258, 7, 227, 289, 133, 320, 70, 196, 165, 290, 259, 228, 39, 321,
+  102, 352, 8, 197,
+  71, 134, 322, 291, 260, 353, 384, 229, 166, 103, 40, 354, 323, 292,
+  135, 385, 198, 261, 72, 9, 416, 167, 386, 355, 230, 324, 104, 293,
+  41, 417, 199, 136,
+  262, 387, 448, 325, 356, 10, 73, 418, 231, 168, 449, 294, 388, 105,
+  419, 263, 42, 200, 357, 450, 137, 480, 74, 326, 232, 11, 389, 169,
+  295, 420, 106, 451,
+  481, 358, 264, 327, 201, 43, 138, 512, 482, 390, 296, 233, 170, 421,
+  75, 452, 359, 12, 513, 265, 483, 328, 107, 202, 514, 544, 422, 391,
+  453, 139, 44, 234,
+  484, 297, 360, 171, 76, 515, 545, 266, 329, 454, 13, 423, 203, 108,
+  546, 485, 576, 298, 235, 140, 361, 330, 172, 547, 45, 455, 267, 577,
+  486, 77, 204, 362,
+  608, 14, 299, 578, 109, 236, 487, 609, 331, 141, 579, 46, 15, 173,
+  610, 363, 78, 205, 16, 110, 237, 611, 142, 47, 174, 79, 206, 17,
+  111, 238, 48, 143,
+  80, 175, 112, 207, 49, 18, 239, 81, 113, 19, 50, 82, 114, 51,
+  83, 115, 640, 516, 392, 268, 144, 20, 672, 641, 548, 517, 424,
+  393, 300, 269, 176, 145,
+  52, 21, 704, 673, 642, 580, 549, 518, 456, 425, 394, 332, 301,
+  270, 208, 177, 146, 84, 53, 22, 736, 705, 674, 643, 612, 581,
+  550, 519, 488, 457, 426, 395,
+  364, 333, 302, 271, 240, 209, 178, 147, 116, 85, 54, 23, 737,
+  706, 675, 613, 582, 551, 489, 458, 427, 365, 334, 303, 241,
+  210, 179, 117, 86, 55, 738, 707,
+  614, 583, 490, 459, 366, 335, 242, 211, 118, 87, 739, 615, 491,
+  367, 243, 119, 768, 644, 520, 396, 272, 148, 24, 800, 769, 676,
+  645, 552, 521, 428, 397, 304,
+  273, 180, 149, 56, 25, 832, 801, 770, 708, 677, 646, 584, 553,
+  522, 460, 429, 398, 336, 305, 274, 212, 181, 150, 88, 57, 26,
+  864, 833, 802, 771, 740, 709,
+  678, 647, 616, 585, 554, 523, 492, 461, 430, 399, 368, 337, 306,
+  275, 244, 213, 182, 151, 120, 89, 58, 27, 865, 834, 803, 741,
+  710, 679, 617, 586, 555, 493,
+  462, 431, 369, 338, 307, 245, 214, 183, 121, 90, 59, 866, 835,
+  742, 711, 618, 587, 494, 463, 370, 339, 246, 215, 122, 91, 867,
+  743, 619, 495, 371, 247, 123,
+  896, 772, 648, 524, 400, 276, 152, 28, 928, 897, 804, 773, 680,
+  649, 556, 525, 432, 401, 308, 277, 184, 153, 60, 29, 960, 929,
+  898, 836, 805, 774, 712, 681,
+  650, 588, 557, 526, 464, 433, 402, 340, 309, 278, 216, 185, 154,
+  92, 61, 30, 992, 961, 930, 899, 868, 837, 806, 775, 744, 713, 682,
+  651, 620, 589, 558, 527,
+  496, 465, 434, 403, 372, 341, 310, 279, 248, 217, 186, 155, 124,
+  93, 62, 31, 993, 962, 931, 869, 838, 807, 745, 714, 683, 621, 590,
+  559, 497, 466, 435, 373,
+  342, 311, 249, 218, 187, 125, 94, 63, 994, 963, 870, 839, 746, 715,
+  622, 591, 498, 467, 374, 343, 250, 219, 126, 95, 995, 871, 747, 623,
+  499, 375, 251, 127,
+  900, 776, 652, 528, 404, 280, 156, 932, 901, 808, 777, 684, 653, 560,
+  529, 436, 405, 312, 281, 188, 157, 964, 933, 902, 840, 809, 778, 716,
+  685, 654, 592, 561,
+  530, 468, 437, 406, 344, 313, 282, 220, 189, 158, 996, 965, 934, 903,
+  872, 841, 810, 779, 748, 717, 686, 655, 624, 593, 562, 531, 500, 469,
+  438, 407, 376, 345,
+  314, 283, 252, 221, 190, 159, 997, 966, 935, 873, 842, 811, 749, 718,
+  687, 625, 594, 563, 501, 470, 439, 377, 346, 315, 253, 222, 191, 998,
+  967, 874, 843, 750,
+  719, 626, 595, 502, 471, 378, 347, 254, 223, 999, 875, 751, 627, 503,
+  379, 255, 904, 780, 656, 532, 408, 284, 936, 905, 812, 781, 688, 657,
+  564, 533, 440, 409,
+  316, 285, 968, 937, 906, 844, 813, 782, 720, 689, 658, 596, 565, 534,
+  472, 441, 410, 348, 317, 286, 1000, 969, 938, 907, 876, 845, 814, 783,
+  752, 721, 690, 659,
+  628, 597, 566, 535, 504, 473, 442, 411, 380, 349, 318, 287, 1001, 970,
+  939, 877, 846, 815, 753, 722, 691, 629, 598, 567, 505, 474, 443, 381,
+  350, 319, 1002, 971,
+  878, 847, 754, 723, 630, 599, 506, 475, 382, 351, 1003, 879, 755, 631,
+  507, 383, 908, 784, 660, 536, 412, 940, 909, 816, 785, 692, 661, 568,
+  537, 444, 413, 972,
+  941, 910, 848, 817, 786, 724, 693, 662, 600, 569, 538, 476, 445, 414,
+  1004, 973, 942, 911, 880, 849, 818, 787, 756, 725, 694, 663, 632, 601,
+  570, 539, 508, 477,
+  446, 415, 1005, 974, 943, 881, 850, 819, 757, 726, 695, 633, 602, 571,
+  509, 478, 447, 1006, 975, 882, 851, 758, 727, 634, 603, 510, 479,
+  1007, 883, 759, 635, 511,
+  912, 788, 664, 540, 944, 913, 820, 789, 696, 665, 572, 541, 976, 945,
+  914, 852, 821, 790, 728, 697, 666, 604, 573, 542, 1008, 977, 946, 915,
+  884, 853, 822, 791,
+  760, 729, 698, 667, 636, 605, 574, 543, 1009, 978, 947, 885, 854, 823,
+  761, 730, 699, 637, 606, 575, 1010, 979, 886, 855, 762, 731, 638, 607,
+  1011, 887, 763, 639,
+  916, 792, 668, 948, 917, 824, 793, 700, 669, 980, 949, 918, 856, 825,
+  794, 732, 701, 670, 1012, 981, 950, 919, 888, 857, 826, 795, 764, 733,
+  702, 671, 1013, 982,
+  951, 889, 858, 827, 765, 734, 703, 1014, 983, 890, 859, 766, 735, 1015,
+  891, 767, 920, 796, 952, 921, 828, 797, 984, 953, 922, 860, 829, 798,
+  1016, 985, 954, 923,
+  892, 861, 830, 799, 1017, 986, 955, 893, 862, 831, 1018, 987, 894, 863,
+  1019, 895, 924, 956, 925, 988, 957, 926, 1020, 989, 958, 927, 1021,
+  990, 959, 1022, 991, 1023,
+};
+
+// Neighborhood 2-tuples for various scans and blocksizes,
+// in {top, left} order for each position in corresponding scan order.
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 0, 1, 4, 4, 4, 1, 1, 8, 8, 5, 8, 2, 2, 2, 5, 9, 12, 6, 9,
+  3, 6, 10, 13, 7, 10, 11, 14, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                col_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 4, 4, 0, 0, 8, 8, 1, 1, 5, 5, 1, 1, 9, 9, 2, 2, 6, 6, 2, 2, 3,
+  3, 10, 10, 7, 7, 11, 11, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                row_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 0, 1, 1, 4, 4, 2, 2, 5, 5, 4, 4, 8, 8, 6, 6, 8, 8, 9, 9, 12,
+  12, 10, 10, 13, 13, 14, 14, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                col_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 8, 8, 0, 0, 16, 16, 1, 1, 24, 24, 9, 9, 1, 1, 32, 32, 17, 17, 2,
+  2, 25, 25, 10, 10, 40, 40, 2, 2, 18, 18, 33, 33, 3, 3, 48, 48, 11, 11, 26,
+  26, 3, 3, 41, 41, 19, 19, 34, 34, 4, 4, 27, 27, 12, 12, 49, 49, 42, 42, 20,
+  20, 4, 4, 35, 35, 5, 5, 28, 28, 50, 50, 43, 43, 13, 13, 36, 36, 5, 5, 21, 21,
+  51, 51, 29, 29, 6, 6, 44, 44, 14, 14, 6, 6, 37, 37, 52, 52, 22, 22, 7, 7, 30,
+  30, 45, 45, 15, 15, 38, 38, 23, 23, 53, 53, 31, 31, 46, 46, 39, 39, 54, 54,
+  47, 47, 55, 55, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                row_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 1, 1, 0, 0, 8, 8, 2, 2, 8, 8, 9, 9, 3, 3, 16, 16, 10, 10, 16, 16,
+  4, 4, 17, 17, 24, 24, 11, 11, 18, 18, 25, 25, 24, 24, 5, 5, 12, 12, 19, 19,
+  32, 32, 26, 26, 6, 6, 33, 33, 32, 32, 20, 20, 27, 27, 40, 40, 13, 13, 34, 34,
+  40, 40, 41, 41, 28, 28, 35, 35, 48, 48, 21, 21, 42, 42, 14, 14, 48, 48, 36,
+  36, 49, 49, 43, 43, 29, 29, 56, 56, 22, 22, 50, 50, 57, 57, 44, 44, 37, 37,
+  51, 51, 30, 30, 58, 58, 52, 52, 45, 45, 59, 59, 38, 38, 60, 60, 46, 46, 53,
+  53, 54, 54, 61, 61, 62, 62, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 0, 8, 8, 1, 8, 1, 1, 9, 16, 16, 16, 2, 9, 2, 2, 10, 17, 17,
+  24, 24, 24, 3, 10, 3, 3, 18, 25, 25, 32, 11, 18, 32, 32, 4, 11, 26, 33, 19,
+  26, 4, 4, 33, 40, 12, 19, 40, 40, 5, 12, 27, 34, 34, 41, 20, 27, 13, 20, 5,
+  5, 41, 48, 48, 48, 28, 35, 35, 42, 21, 28, 6, 6, 6, 13, 42, 49, 49, 56, 36,
+  43, 14, 21, 29, 36, 7, 14, 43, 50, 50, 57, 22, 29, 37, 44, 15, 22, 44, 51,
+  51, 58, 30, 37, 23, 30, 52, 59, 45, 52, 38, 45, 31, 38, 53, 60, 46, 53, 39,
+  46, 54, 61, 47, 54, 55, 62, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                col_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 16, 16, 32, 32, 0, 0, 48, 48, 1, 1, 64, 64,
+  17, 17, 80, 80, 33, 33, 1, 1, 49, 49, 96, 96, 2, 2, 65, 65,
+  18, 18, 112, 112, 34, 34, 81, 81, 2, 2, 50, 50, 128, 128, 3, 3,
+  97, 97, 19, 19, 66, 66, 144, 144, 82, 82, 35, 35, 113, 113, 3, 3,
+  51, 51, 160, 160, 4, 4, 98, 98, 129, 129, 67, 67, 20, 20, 83, 83,
+  114, 114, 36, 36, 176, 176, 4, 4, 145, 145, 52, 52, 99, 99, 5, 5,
+  130, 130, 68, 68, 192, 192, 161, 161, 21, 21, 115, 115, 84, 84, 37, 37,
+  146, 146, 208, 208, 53, 53, 5, 5, 100, 100, 177, 177, 131, 131, 69, 69,
+  6, 6, 224, 224, 116, 116, 22, 22, 162, 162, 85, 85, 147, 147, 38, 38,
+  193, 193, 101, 101, 54, 54, 6, 6, 132, 132, 178, 178, 70, 70, 163, 163,
+  209, 209, 7, 7, 117, 117, 23, 23, 148, 148, 7, 7, 86, 86, 194, 194,
+  225, 225, 39, 39, 179, 179, 102, 102, 133, 133, 55, 55, 164, 164, 8, 8,
+  71, 71, 210, 210, 118, 118, 149, 149, 195, 195, 24, 24, 87, 87, 40, 40,
+  56, 56, 134, 134, 180, 180, 226, 226, 103, 103, 8, 8, 165, 165, 211, 211,
+  72, 72, 150, 150, 9, 9, 119, 119, 25, 25, 88, 88, 196, 196, 41, 41,
+  135, 135, 181, 181, 104, 104, 57, 57, 227, 227, 166, 166, 120, 120, 151, 151,
+  197, 197, 73, 73, 9, 9, 212, 212, 89, 89, 136, 136, 182, 182, 10, 10,
+  26, 26, 105, 105, 167, 167, 228, 228, 152, 152, 42, 42, 121, 121, 213, 213,
+  58, 58, 198, 198, 74, 74, 137, 137, 183, 183, 168, 168, 10, 10, 90, 90,
+  229, 229, 11, 11, 106, 106, 214, 214, 153, 153, 27, 27, 199, 199, 43, 43,
+  184, 184, 122, 122, 169, 169, 230, 230, 59, 59, 11, 11, 75, 75, 138, 138,
+  200, 200, 215, 215, 91, 91, 12, 12, 28, 28, 185, 185, 107, 107, 154, 154,
+  44, 44, 231, 231, 216, 216, 60, 60, 123, 123, 12, 12, 76, 76, 201, 201,
+  170, 170, 232, 232, 139, 139, 92, 92, 13, 13, 108, 108, 29, 29, 186, 186,
+  217, 217, 155, 155, 45, 45, 13, 13, 61, 61, 124, 124, 14, 14, 233, 233,
+  77, 77, 14, 14, 171, 171, 140, 140, 202, 202, 30, 30, 93, 93, 109, 109,
+  46, 46, 156, 156, 62, 62, 187, 187, 15, 15, 125, 125, 218, 218, 78, 78,
+  31, 31, 172, 172, 47, 47, 141, 141, 94, 94, 234, 234, 203, 203, 63, 63,
+  110, 110, 188, 188, 157, 157, 126, 126, 79, 79, 173, 173, 95, 95, 219, 219,
+  142, 142, 204, 204, 235, 235, 111, 111, 158, 158, 127, 127, 189, 189, 220,
+  220, 143, 143, 174, 174, 205, 205, 236, 236, 159, 159, 190, 190, 221, 221,
+  175, 175, 237, 237, 206, 206, 222, 222, 191, 191, 238, 238, 207, 207, 223,
+  223, 239, 239, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                row_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 1, 1, 0, 0, 2, 2, 16, 16, 3, 3, 17, 17,
+  16, 16, 4, 4, 32, 32, 18, 18, 5, 5, 33, 33, 32, 32, 19, 19,
+  48, 48, 6, 6, 34, 34, 20, 20, 49, 49, 48, 48, 7, 7, 35, 35,
+  64, 64, 21, 21, 50, 50, 36, 36, 64, 64, 8, 8, 65, 65, 51, 51,
+  22, 22, 37, 37, 80, 80, 66, 66, 9, 9, 52, 52, 23, 23, 81, 81,
+  67, 67, 80, 80, 38, 38, 10, 10, 53, 53, 82, 82, 96, 96, 68, 68,
+  24, 24, 97, 97, 83, 83, 39, 39, 96, 96, 54, 54, 11, 11, 69, 69,
+  98, 98, 112, 112, 84, 84, 25, 25, 40, 40, 55, 55, 113, 113, 99, 99,
+  12, 12, 70, 70, 112, 112, 85, 85, 26, 26, 114, 114, 100, 100, 128, 128,
+  41, 41, 56, 56, 71, 71, 115, 115, 13, 13, 86, 86, 129, 129, 101, 101,
+  128, 128, 72, 72, 130, 130, 116, 116, 27, 27, 57, 57, 14, 14, 87, 87,
+  42, 42, 144, 144, 102, 102, 131, 131, 145, 145, 117, 117, 73, 73, 144, 144,
+  88, 88, 132, 132, 103, 103, 28, 28, 58, 58, 146, 146, 118, 118, 43, 43,
+  160, 160, 147, 147, 89, 89, 104, 104, 133, 133, 161, 161, 119, 119, 160, 160,
+  74, 74, 134, 134, 148, 148, 29, 29, 59, 59, 162, 162, 176, 176, 44, 44,
+  120, 120, 90, 90, 105, 105, 163, 163, 177, 177, 149, 149, 176, 176, 135, 135,
+  164, 164, 178, 178, 30, 30, 150, 150, 192, 192, 75, 75, 121, 121, 60, 60,
+  136, 136, 193, 193, 106, 106, 151, 151, 179, 179, 192, 192, 45, 45, 165, 165,
+  166, 166, 194, 194, 91, 91, 180, 180, 137, 137, 208, 208, 122, 122, 152, 152,
+  208, 208, 195, 195, 76, 76, 167, 167, 209, 209, 181, 181, 224, 224, 107, 107,
+  196, 196, 61, 61, 153, 153, 224, 224, 182, 182, 168, 168, 210, 210, 46, 46,
+  138, 138, 92, 92, 183, 183, 225, 225, 211, 211, 240, 240, 197, 197, 169, 169,
+  123, 123, 154, 154, 198, 198, 77, 77, 212, 212, 184, 184, 108, 108, 226, 226,
+  199, 199, 62, 62, 227, 227, 241, 241, 139, 139, 213, 213, 170, 170, 185, 185,
+  155, 155, 228, 228, 242, 242, 124, 124, 93, 93, 200, 200, 243, 243, 214, 214,
+  215, 215, 229, 229, 140, 140, 186, 186, 201, 201, 78, 78, 171, 171, 109, 109,
+  156, 156, 244, 244, 216, 216, 230, 230, 94, 94, 245, 245, 231, 231, 125, 125,
+  202, 202, 246, 246, 232, 232, 172, 172, 217, 217, 141, 141, 110, 110, 157,
+  157, 187, 187, 247, 247, 126, 126, 233, 233, 218, 218, 248, 248, 188, 188,
+  203, 203, 142, 142, 173, 173, 158, 158, 249, 249, 234, 234, 204, 204, 219,
+  219, 174, 174, 189, 189, 250, 250, 220, 220, 190, 190, 205, 205, 235, 235,
+  206, 206, 236, 236, 251, 251, 221, 221, 252, 252, 222, 222, 237, 237, 238,
+  238, 253, 253, 254, 254, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 0, 16, 16, 1, 16, 1, 1, 32, 32, 17, 32,
+  2, 17, 2, 2, 48, 48, 18, 33, 33, 48, 3, 18, 49, 64, 64, 64,
+  34, 49, 3, 3, 19, 34, 50, 65, 4, 19, 65, 80, 80, 80, 35, 50,
+  4, 4, 20, 35, 66, 81, 81, 96, 51, 66, 96, 96, 5, 20, 36, 51,
+  82, 97, 21, 36, 67, 82, 97, 112, 5, 5, 52, 67, 112, 112, 37, 52,
+  6, 21, 83, 98, 98, 113, 68, 83, 6, 6, 113, 128, 22, 37, 53, 68,
+  84, 99, 99, 114, 128, 128, 114, 129, 69, 84, 38, 53, 7, 22, 7, 7,
+  129, 144, 23, 38, 54, 69, 100, 115, 85, 100, 115, 130, 144, 144, 130, 145,
+  39, 54, 70, 85, 8, 23, 55, 70, 116, 131, 101, 116, 145, 160, 24, 39,
+  8, 8, 86, 101, 131, 146, 160, 160, 146, 161, 71, 86, 40, 55, 9, 24,
+  117, 132, 102, 117, 161, 176, 132, 147, 56, 71, 87, 102, 25, 40, 147, 162,
+  9, 9, 176, 176, 162, 177, 72, 87, 41, 56, 118, 133, 133, 148, 103, 118,
+  10, 25, 148, 163, 57, 72, 88, 103, 177, 192, 26, 41, 163, 178, 192, 192,
+  10, 10, 119, 134, 73, 88, 149, 164, 104, 119, 134, 149, 42, 57, 178, 193,
+  164, 179, 11, 26, 58, 73, 193, 208, 89, 104, 135, 150, 120, 135, 27, 42,
+  74, 89, 208, 208, 150, 165, 179, 194, 165, 180, 105, 120, 194, 209, 43, 58,
+  11, 11, 136, 151, 90, 105, 151, 166, 180, 195, 59, 74, 121, 136, 209, 224,
+  195, 210, 224, 224, 166, 181, 106, 121, 75, 90, 12, 27, 181, 196, 12, 12,
+  210, 225, 152, 167, 167, 182, 137, 152, 28, 43, 196, 211, 122, 137, 91, 106,
+  225, 240, 44, 59, 13, 28, 107, 122, 182, 197, 168, 183, 211, 226, 153, 168,
+  226, 241, 60, 75, 197, 212, 138, 153, 29, 44, 76, 91, 13, 13, 183, 198,
+  123, 138, 45, 60, 212, 227, 198, 213, 154, 169, 169, 184, 227, 242, 92, 107,
+  61, 76, 139, 154, 14, 29, 14, 14, 184, 199, 213, 228, 108, 123, 199, 214,
+  228, 243, 77, 92, 30, 45, 170, 185, 155, 170, 185, 200, 93, 108, 124, 139,
+  214, 229, 46, 61, 200, 215, 229, 244, 15, 30, 109, 124, 62, 77, 140, 155,
+  215, 230, 31, 46, 171, 186, 186, 201, 201, 216, 78, 93, 230, 245, 125, 140,
+  47, 62, 216, 231, 156, 171, 94, 109, 231, 246, 141, 156, 63, 78, 202, 217,
+  187, 202, 110, 125, 217, 232, 172, 187, 232, 247, 79, 94, 157, 172, 126, 141,
+  203, 218, 95, 110, 233, 248, 218, 233, 142, 157, 111, 126, 173, 188, 188, 203,
+  234, 249, 219, 234, 127, 142, 158, 173, 204, 219, 189, 204, 143, 158, 235,
+  250, 174, 189, 205, 220, 159, 174, 220, 235, 221, 236, 175, 190, 190, 205,
+  236, 251, 206, 221, 237, 252, 191, 206, 222, 237, 207, 222, 238, 253, 223,
+  238, 239, 254, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_32x32_neighbors[1025 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 0, 32, 32, 1, 32, 1, 1, 64, 64, 33, 64,
+  2, 33, 96, 96, 2, 2, 65, 96, 34, 65, 128, 128, 97, 128, 3, 34,
+  66, 97, 3, 3, 35, 66, 98, 129, 129, 160, 160, 160, 4, 35, 67, 98,
+  192, 192, 4, 4, 130, 161, 161, 192, 36, 67, 99, 130, 5, 36, 68, 99,
+  193, 224, 162, 193, 224, 224, 131, 162, 37, 68, 100, 131, 5, 5, 194, 225,
+  225, 256, 256, 256, 163, 194, 69, 100, 132, 163, 6, 37, 226, 257, 6, 6,
+  195, 226, 257, 288, 101, 132, 288, 288, 38, 69, 164, 195, 133, 164, 258, 289,
+  227, 258, 196, 227, 7, 38, 289, 320, 70, 101, 320, 320, 7, 7, 165, 196,
+  39, 70, 102, 133, 290, 321, 259, 290, 228, 259, 321, 352, 352, 352, 197, 228,
+  134, 165, 71, 102, 8, 39, 322, 353, 291, 322, 260, 291, 103, 134, 353, 384,
+  166, 197, 229, 260, 40, 71, 8, 8, 384, 384, 135, 166, 354, 385, 323, 354,
+  198, 229, 292, 323, 72, 103, 261, 292, 9, 40, 385, 416, 167, 198, 104, 135,
+  230, 261, 355, 386, 416, 416, 293, 324, 324, 355, 9, 9, 41, 72, 386, 417,
+  199, 230, 136, 167, 417, 448, 262, 293, 356, 387, 73, 104, 387, 418, 231, 262,
+  10, 41, 168, 199, 325, 356, 418, 449, 105, 136, 448, 448, 42, 73, 294, 325,
+  200, 231, 10, 10, 357, 388, 137, 168, 263, 294, 388, 419, 74, 105, 419, 450,
+  449, 480, 326, 357, 232, 263, 295, 326, 169, 200, 11, 42, 106, 137, 480, 480,
+  450, 481, 358, 389, 264, 295, 201, 232, 138, 169, 389, 420, 43, 74, 420, 451,
+  327, 358, 11, 11, 481, 512, 233, 264, 451, 482, 296, 327, 75, 106, 170, 201,
+  482, 513, 512, 512, 390, 421, 359, 390, 421, 452, 107, 138, 12, 43, 202, 233,
+  452, 483, 265, 296, 328, 359, 139, 170, 44, 75, 483, 514, 513, 544, 234, 265,
+  297, 328, 422, 453, 12, 12, 391, 422, 171, 202, 76, 107, 514, 545, 453, 484,
+  544, 544, 266, 297, 203, 234, 108, 139, 329, 360, 298, 329, 140, 171, 515,
+  546, 13, 44, 423, 454, 235, 266, 545, 576, 454, 485, 45, 76, 172, 203, 330,
+  361, 576, 576, 13, 13, 267, 298, 546, 577, 77, 108, 204, 235, 455, 486, 577,
+  608, 299, 330, 109, 140, 547, 578, 14, 45, 14, 14, 141, 172, 578, 609, 331,
+  362, 46, 77, 173, 204, 15, 15, 78, 109, 205, 236, 579, 610, 110, 141, 15, 46,
+  142, 173, 47, 78, 174, 205, 16, 16, 79, 110, 206, 237, 16, 47, 111, 142,
+  48, 79, 143, 174, 80, 111, 175, 206, 17, 48, 17, 17, 207, 238, 49, 80,
+  81, 112, 18, 18, 18, 49, 50, 81, 82, 113, 19, 50, 51, 82, 83, 114, 608, 608,
+  484, 515, 360, 391, 236, 267, 112, 143, 19, 19, 640, 640, 609, 640, 516, 547,
+  485, 516, 392, 423, 361, 392, 268, 299, 237, 268, 144, 175, 113, 144, 20, 51,
+  20, 20, 672, 672, 641, 672, 610, 641, 548, 579, 517, 548, 486, 517, 424, 455,
+  393, 424, 362, 393, 300, 331, 269, 300, 238, 269, 176, 207, 145, 176, 114,
+  145, 52, 83, 21, 52, 21, 21, 704, 704, 673, 704, 642, 673, 611, 642, 580,
+  611, 549, 580, 518, 549, 487, 518, 456, 487, 425, 456, 394, 425, 363, 394,
+  332, 363, 301, 332, 270, 301, 239, 270, 208, 239, 177, 208, 146, 177, 115,
+  146, 84, 115, 53, 84, 22, 53, 22, 22, 705, 736, 674, 705, 643, 674, 581, 612,
+  550, 581, 519, 550, 457, 488, 426, 457, 395, 426, 333, 364, 302, 333, 271,
+  302, 209, 240, 178, 209, 147, 178, 85, 116, 54, 85, 23, 54, 706, 737, 675,
+  706, 582, 613, 551, 582, 458, 489, 427, 458, 334, 365, 303, 334, 210, 241,
+  179, 210, 86, 117, 55, 86, 707, 738, 583, 614, 459, 490, 335, 366, 211, 242,
+  87, 118, 736, 736, 612, 643, 488, 519, 364, 395, 240, 271, 116, 147, 23, 23,
+  768, 768, 737, 768, 644, 675, 613, 644, 520, 551, 489, 520, 396, 427, 365,
+  396, 272, 303, 241, 272, 148, 179, 117, 148, 24, 55, 24, 24, 800, 800, 769,
+  800, 738, 769, 676, 707, 645, 676, 614, 645, 552, 583, 521, 552, 490, 521,
+  428, 459, 397, 428, 366, 397, 304, 335, 273, 304, 242, 273, 180, 211, 149,
+  180, 118, 149, 56, 87, 25, 56, 25, 25, 832, 832, 801, 832, 770, 801, 739,
+  770, 708, 739, 677, 708, 646, 677, 615, 646, 584, 615, 553, 584, 522, 553,
+  491, 522, 460, 491, 429, 460, 398, 429, 367, 398, 336, 367, 305, 336, 274,
+  305, 243, 274, 212, 243, 181, 212, 150, 181, 119, 150, 88, 119, 57, 88, 26,
+  57, 26, 26, 833, 864, 802, 833, 771, 802, 709, 740, 678, 709, 647, 678, 585,
+  616, 554, 585, 523, 554, 461, 492, 430, 461, 399, 430, 337, 368, 306, 337,
+  275, 306, 213, 244, 182, 213, 151, 182, 89, 120, 58, 89, 27, 58, 834, 865,
+  803, 834, 710, 741, 679, 710, 586, 617, 555, 586, 462, 493, 431, 462, 338,
+  369, 307, 338, 214, 245, 183, 214, 90, 121, 59, 90, 835, 866, 711, 742, 587,
+  618, 463, 494, 339, 370, 215, 246, 91, 122, 864, 864, 740, 771, 616, 647,
+  492, 523, 368, 399, 244, 275, 120, 151, 27, 27, 896, 896, 865, 896, 772, 803,
+  741, 772, 648, 679, 617, 648, 524, 555, 493, 524, 400, 431, 369, 400, 276,
+  307, 245, 276, 152, 183, 121, 152, 28, 59, 28, 28, 928, 928, 897, 928, 866,
+  897, 804, 835, 773, 804, 742, 773, 680, 711, 649, 680, 618, 649, 556, 587,
+  525, 556, 494, 525, 432, 463, 401, 432, 370, 401, 308, 339, 277, 308, 246,
+  277, 184, 215, 153, 184, 122, 153, 60, 91, 29, 60, 29, 29, 960, 960, 929,
+  960, 898, 929, 867, 898, 836, 867, 805, 836, 774, 805, 743, 774, 712, 743,
+  681, 712, 650, 681, 619, 650, 588, 619, 557, 588, 526, 557, 495, 526, 464,
+  495, 433, 464, 402, 433, 371, 402, 340, 371, 309, 340, 278, 309, 247, 278,
+  216, 247, 185, 216, 154, 185, 123, 154, 92, 123, 61, 92, 30, 61, 30, 30,
+  961, 992, 930, 961, 899, 930, 837, 868, 806, 837, 775, 806, 713, 744, 682,
+  713, 651, 682, 589, 620, 558, 589, 527, 558, 465, 496, 434, 465, 403, 434,
+  341, 372, 310, 341, 279, 310, 217, 248, 186, 217, 155, 186, 93, 124, 62, 93,
+  31, 62, 962, 993, 931, 962, 838, 869, 807, 838, 714, 745, 683, 714, 590, 621,
+  559, 590, 466, 497, 435, 466, 342, 373, 311, 342, 218, 249, 187, 218, 94,
+  125, 63, 94, 963, 994, 839, 870, 715, 746, 591, 622, 467, 498, 343, 374, 219,
+  250, 95, 126, 868, 899, 744, 775, 620, 651, 496, 527, 372, 403, 248, 279,
+  124, 155, 900, 931, 869, 900, 776, 807, 745, 776, 652, 683, 621, 652, 528,
+  559, 497, 528, 404, 435, 373, 404, 280, 311, 249, 280, 156, 187, 125, 156,
+  932, 963, 901, 932, 870, 901, 808, 839, 777, 808, 746, 777, 684, 715, 653,
+  684, 622, 653, 560, 591, 529, 560, 498, 529, 436, 467, 405, 436, 374, 405,
+  312, 343, 281, 312, 250, 281, 188, 219, 157, 188, 126, 157, 964, 995, 933,
+  964, 902, 933, 871, 902, 840, 871, 809, 840, 778, 809, 747, 778, 716, 747,
+  685, 716, 654, 685, 623, 654, 592, 623, 561, 592, 530, 561, 499, 530, 468,
+  499, 437, 468, 406, 437, 375, 406, 344, 375, 313, 344, 282, 313, 251, 282,
+  220, 251, 189, 220, 158, 189, 127, 158, 965, 996, 934, 965, 903, 934, 841,
+  872, 810, 841, 779, 810, 717, 748, 686, 717, 655, 686, 593, 624, 562, 593,
+  531, 562, 469, 500, 438, 469, 407, 438, 345, 376, 314, 345, 283, 314, 221,
+  252, 190, 221, 159, 190, 966, 997, 935, 966, 842, 873, 811, 842, 718, 749,
+  687, 718, 594, 625, 563, 594, 470, 501, 439, 470, 346, 377, 315, 346, 222,
+  253, 191, 222, 967, 998, 843, 874, 719, 750, 595, 626, 471, 502, 347, 378,
+  223, 254, 872, 903, 748, 779, 624, 655, 500, 531, 376, 407, 252, 283, 904,
+  935, 873, 904, 780, 811, 749, 780, 656, 687, 625, 656, 532, 563, 501, 532,
+  408, 439, 377, 408, 284, 315, 253, 284, 936, 967, 905, 936, 874, 905, 812,
+  843, 781, 812, 750, 781, 688, 719, 657, 688, 626, 657, 564, 595, 533, 564,
+  502, 533, 440, 471, 409, 440, 378, 409, 316, 347, 285, 316, 254, 285, 968,
+  999, 937, 968, 906, 937, 875, 906, 844, 875, 813, 844, 782, 813, 751, 782,
+  720, 751, 689, 720, 658, 689, 627, 658, 596, 627, 565, 596, 534, 565, 503,
+  534, 472, 503, 441, 472, 410, 441, 379, 410, 348, 379, 317, 348, 286, 317,
+  255, 286, 969, 1000, 938, 969, 907, 938, 845, 876, 814, 845, 783, 814, 721,
+  752, 690, 721, 659, 690, 597, 628, 566, 597, 535, 566, 473, 504, 442, 473,
+  411, 442, 349, 380, 318, 349, 287, 318, 970, 1001, 939, 970, 846, 877, 815,
+  846, 722, 753, 691, 722, 598, 629, 567, 598, 474, 505, 443, 474, 350, 381,
+  319, 350, 971, 1002, 847, 878, 723, 754, 599, 630, 475, 506, 351, 382, 876,
+  907, 752, 783, 628, 659, 504, 535, 380, 411, 908, 939, 877, 908, 784, 815,
+  753, 784, 660, 691, 629, 660, 536, 567, 505, 536, 412, 443, 381, 412, 940,
+  971, 909, 940, 878, 909, 816, 847, 785, 816, 754, 785, 692, 723, 661, 692,
+  630, 661, 568, 599, 537, 568, 506, 537, 444, 475, 413, 444, 382, 413, 972,
+  1003, 941, 972, 910, 941, 879, 910, 848, 879, 817, 848, 786, 817, 755, 786,
+  724, 755, 693, 724, 662, 693, 631, 662, 600, 631, 569, 600, 538, 569, 507,
+  538, 476, 507, 445, 476, 414, 445, 383, 414, 973, 1004, 942, 973, 911, 942,
+  849, 880, 818, 849, 787, 818, 725, 756, 694, 725, 663, 694, 601, 632, 570,
+  601, 539, 570, 477, 508, 446, 477, 415, 446, 974, 1005, 943, 974, 850, 881,
+  819, 850, 726, 757, 695, 726, 602, 633, 571, 602, 478, 509, 447, 478, 975,
+  1006, 851, 882, 727, 758, 603, 634, 479, 510, 880, 911, 756, 787, 632, 663,
+  508, 539, 912, 943, 881, 912, 788, 819, 757, 788, 664, 695, 633, 664, 540,
+  571, 509, 540, 944, 975, 913, 944, 882, 913, 820, 851, 789, 820, 758, 789,
+  696, 727, 665, 696, 634, 665, 572, 603, 541, 572, 510, 541, 976, 1007, 945,
+  976, 914, 945, 883, 914, 852, 883, 821, 852, 790, 821, 759, 790, 728, 759,
+  697, 728, 666, 697, 635, 666, 604, 635, 573, 604, 542, 573, 511, 542, 977,
+  1008, 946, 977, 915, 946, 853, 884, 822, 853, 791, 822, 729, 760, 698, 729,
+  667, 698, 605, 636, 574, 605, 543, 574, 978, 1009, 947, 978, 854, 885, 823,
+  854, 730, 761, 699, 730, 606, 637, 575, 606, 979, 1010, 855, 886, 731, 762,
+  607, 638, 884, 915, 760, 791, 636, 667, 916, 947, 885, 916, 792, 823, 761,
+  792, 668, 699, 637, 668, 948, 979, 917, 948, 886, 917, 824, 855, 793, 824,
+  762, 793, 700, 731, 669, 700, 638, 669, 980, 1011, 949, 980, 918, 949, 887,
+  918, 856, 887, 825, 856, 794, 825, 763, 794, 732, 763, 701, 732, 670, 701,
+  639, 670, 981, 1012, 950, 981, 919, 950, 857, 888, 826, 857, 795, 826, 733,
+  764, 702, 733, 671, 702, 982, 1013, 951, 982, 858, 889, 827, 858, 734, 765,
+  703, 734, 983, 1014, 859, 890, 735, 766, 888, 919, 764, 795, 920, 951, 889,
+  920, 796, 827, 765, 796, 952, 983, 921, 952, 890, 921, 828, 859, 797, 828,
+  766, 797, 984, 1015, 953, 984, 922, 953, 891, 922, 860, 891, 829, 860, 798,
+  829, 767, 798, 985, 1016, 954, 985, 923, 954, 861, 892, 830, 861, 799, 830,
+  986, 1017, 955, 986, 862, 893, 831, 862, 987, 1018, 863, 894, 892, 923, 924,
+  955, 893, 924, 956, 987, 925, 956, 894, 925, 988, 1019, 957, 988, 926, 957,
+  895, 926, 989, 1020, 958, 989, 927, 958, 990, 1021, 959, 990, 991, 1022, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_default_iscan_4x4[16]) = {
+  0, 2, 5, 8, 1, 3, 9, 12, 4, 7, 11, 14, 6, 10, 13, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_col_iscan_4x4[16]) = {
+  0, 3, 7, 11, 1, 5, 9, 12, 2, 6, 10, 14, 4, 8, 13, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_row_iscan_4x4[16]) = {
+  0, 1, 3, 5, 2, 4, 6, 9, 7, 8, 11, 13, 10, 12, 14, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_col_iscan_8x8[64]) = {
+  0, 3, 8, 15, 22, 32, 40, 47, 1, 5, 11, 18, 26, 34, 44, 51,
+  2, 7, 13, 20, 28, 38, 46, 54, 4, 10, 16, 24, 31, 41, 50, 56,
+  6, 12, 21, 27, 35, 43, 52, 58, 9, 17, 25, 33, 39, 48, 55, 60,
+  14, 23, 30, 37, 45, 53, 59, 62, 19, 29, 36, 42, 49, 57, 61, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_row_iscan_8x8[64]) = {
+  0, 1, 2, 5, 8, 12, 19, 24, 3, 4, 7, 10, 15, 20, 30, 39,
+  6, 9, 13, 16, 21, 27, 37, 46, 11, 14, 17, 23, 28, 34, 44, 52,
+  18, 22, 25, 31, 35, 41, 50, 57, 26, 29, 33, 38, 43, 49, 55, 59,
+  32, 36, 42, 47, 51, 54, 60, 61, 40, 45, 48, 53, 56, 58, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_default_iscan_8x8[64]) = {
+  0, 2, 5, 9, 14, 22, 31, 37, 1, 4, 8, 13, 19, 26, 38, 44,
+  3, 6, 10, 17, 24, 30, 42, 49, 7, 11, 15, 21, 29, 36, 47, 53,
+  12, 16, 20, 27, 34, 43, 52, 57, 18, 23, 28, 35, 41, 48, 56, 60,
+  25, 32, 39, 45, 50, 55, 59, 62, 33, 40, 46, 51, 54, 58, 61, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_col_iscan_16x16[256]) = {
+  0, 4, 11, 20, 31, 43, 59, 75, 85, 109, 130, 150, 165, 181, 195, 198,
+  1, 6, 14, 23, 34, 47, 64, 81, 95, 114, 135, 153, 171, 188, 201, 212,
+  2, 8, 16, 25, 38, 52, 67, 83, 101, 116, 136, 157, 172, 190, 205, 216,
+  3, 10, 18, 29, 41, 55, 71, 89, 103, 119, 141, 159, 176, 194, 208, 218,
+  5, 12, 21, 32, 45, 58, 74, 93, 104, 123, 144, 164, 179, 196, 210, 223,
+  7, 15, 26, 37, 49, 63, 78, 96, 112, 129, 146, 166, 182, 200, 215, 228,
+  9, 19, 28, 39, 54, 69, 86, 102, 117, 132, 151, 170, 187, 206, 220, 230,
+  13, 24, 35, 46, 60, 73, 91, 108, 122, 137, 154, 174, 189, 207, 224, 235,
+  17, 30, 40, 53, 66, 82, 98, 115, 126, 142, 161, 180, 197, 213, 227, 237,
+  22, 36, 48, 62, 76, 92, 105, 120, 133, 147, 167, 186, 203, 219, 232, 240,
+  27, 44, 56, 70, 84, 99, 113, 127, 140, 156, 175, 193, 209, 226, 236, 244,
+  33, 51, 68, 79, 94, 110, 125, 138, 149, 162, 184, 202, 217, 229, 241, 247,
+  42, 61, 77, 90, 106, 121, 134, 148, 160, 173, 191, 211, 225, 238, 245, 251,
+  50, 72, 87, 100, 118, 128, 145, 158, 168, 183, 204, 222, 233, 242, 249, 253,
+  57, 80, 97, 111, 131, 143, 155, 169, 178, 192, 214, 231, 239, 246, 250, 254,
+  65, 88, 107, 124, 139, 152, 163, 177, 185, 199, 221, 234, 243, 248, 252, 255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_row_iscan_16x16[256]) = {
+  0, 1, 2, 4, 6, 9, 12, 17, 22, 29, 36, 43, 54, 64, 76, 86,
+  3, 5, 7, 11, 15, 19, 25, 32, 38, 48, 59, 68, 84, 99, 115, 130,
+  8, 10, 13, 18, 23, 27, 33, 42, 51, 60, 72, 88, 103, 119, 142, 167,
+  14, 16, 20, 26, 31, 37, 44, 53, 61, 73, 85, 100, 116, 135, 161, 185,
+  21, 24, 30, 35, 40, 47, 55, 65, 74, 81, 94, 112, 133, 154, 179, 205,
+  28, 34, 39, 45, 50, 58, 67, 77, 87, 96, 106, 121, 146, 169, 196, 212,
+  41, 46, 49, 56, 63, 70, 79, 90, 98, 107, 122, 138, 159, 182, 207, 222,
+  52, 57, 62, 69, 75, 83, 93, 102, 110, 120, 134, 150, 176, 195, 215, 226,
+  66, 71, 78, 82, 91, 97, 108, 113, 127, 136, 148, 168, 188, 202, 221, 232,
+  80, 89, 92, 101, 105, 114, 125, 131, 139, 151, 162, 177, 192, 208, 223, 234,
+  95, 104, 109, 117, 123, 128, 143, 144, 155, 165, 175, 190, 206, 219, 233, 239,
+  111, 118, 124, 129, 140, 147, 157, 164, 170, 181, 191, 203, 224, 230, 240,
+  243, 126, 132, 137, 145, 153, 160, 174, 178, 184, 197, 204, 216, 231, 237,
+  244, 246, 141, 149, 156, 166, 172, 180, 189, 199, 200, 210, 220, 228, 238,
+  242, 249, 251, 152, 163, 171, 183, 186, 193, 201, 211, 214, 218, 227, 236,
+  245, 247, 252, 253, 158, 173, 187, 194, 198, 209, 213, 217, 225, 229, 235,
+  241, 248, 250, 254, 255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_default_iscan_16x16[256]) = {
+  0, 2, 5, 9, 17, 24, 36, 44, 55, 72, 88, 104, 128, 143, 166, 179,
+  1, 4, 8, 13, 20, 30, 40, 54, 66, 79, 96, 113, 141, 154, 178, 196,
+  3, 7, 11, 18, 25, 33, 46, 57, 71, 86, 101, 119, 148, 164, 186, 201,
+  6, 12, 16, 23, 31, 39, 53, 64, 78, 92, 110, 127, 153, 169, 193, 208,
+  10, 14, 19, 28, 37, 47, 58, 67, 84, 98, 114, 133, 161, 176, 198, 214,
+  15, 21, 26, 34, 43, 52, 65, 77, 91, 106, 120, 140, 165, 185, 205, 221,
+  22, 27, 32, 41, 48, 60, 73, 85, 99, 116, 130, 151, 175, 190, 211, 225,
+  29, 35, 42, 49, 59, 69, 81, 95, 108, 125, 139, 155, 182, 197, 217, 229,
+  38, 45, 51, 61, 68, 80, 93, 105, 118, 134, 150, 168, 191, 207, 223, 234,
+  50, 56, 63, 74, 83, 94, 109, 117, 129, 147, 163, 177, 199, 213, 228, 238,
+  62, 70, 76, 87, 97, 107, 122, 131, 145, 159, 172, 188, 210, 222, 235, 242,
+  75, 82, 90, 102, 112, 124, 138, 146, 157, 173, 187, 202, 219, 230, 240, 245,
+  89, 100, 111, 123, 132, 142, 156, 167, 180, 189, 203, 216, 231, 237, 246, 250,
+  103, 115, 126, 136, 149, 162, 171, 183, 194, 204, 215, 224, 236, 241, 248,
+  252, 121, 135, 144, 158, 170, 181, 192, 200, 209, 218, 227, 233, 243, 244,
+  251, 254, 137, 152, 160, 174, 184, 195, 206, 212, 220, 226, 232, 239, 247,
+  249, 253, 255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_default_iscan_32x32[1024]) = {
+  0, 2, 5, 10, 17, 25, 38, 47, 62, 83, 101, 121, 145, 170, 193, 204,
+  210, 219, 229, 233, 245, 257, 275, 299, 342, 356, 377, 405, 455, 471, 495,
+  527, 1, 4, 8, 15, 22, 30, 45, 58, 74, 92, 112, 133, 158, 184, 203, 215, 222,
+  228, 234, 237, 256, 274, 298, 317, 355, 376, 404, 426, 470, 494, 526, 551,
+  3, 7, 12, 18, 28, 36, 52, 64, 82, 102, 118, 142, 164, 189, 208, 217, 224,
+  231, 235, 238, 273, 297, 316, 329, 375, 403, 425, 440, 493, 525, 550, 567,
+  6, 11, 16, 23, 31, 43, 60, 73, 90, 109, 126, 150, 173, 196, 211, 220, 226,
+  232, 236, 239, 296, 315, 328, 335, 402, 424, 439, 447, 524, 549, 566, 575,
+  9, 14, 19, 29, 37, 50, 65, 78, 95, 116, 134, 157, 179, 201, 214, 223, 244,
+  255, 272, 295, 341, 354, 374, 401, 454, 469, 492, 523, 582, 596, 617, 645,
+  13, 20, 26, 35, 44, 54, 72, 85, 105, 123, 140, 163, 182, 205, 216, 225,
+  254, 271, 294, 314, 353, 373, 400, 423, 468, 491, 522, 548, 595, 616, 644,
+  666, 21, 27, 33, 42, 53, 63, 80, 94, 113, 132, 151, 172, 190, 209, 218, 227,
+  270, 293, 313, 327, 372, 399, 422, 438, 490, 521, 547, 565, 615, 643, 665,
+  680, 24, 32, 39, 48, 57, 71, 88, 104, 120, 139, 159, 178, 197, 212, 221, 230,
+  292, 312, 326, 334, 398, 421, 437, 446, 520, 546, 564, 574, 642, 664, 679,
+  687, 34, 40, 46, 56, 68, 81, 96, 111, 130, 147, 167, 186, 243, 253, 269, 291,
+  340, 352, 371, 397, 453, 467, 489, 519, 581, 594, 614, 641, 693, 705, 723,
+  747, 41, 49, 55, 67, 77, 91, 107, 124, 138, 161, 177, 194, 252, 268, 290,
+  311, 351, 370, 396, 420, 466, 488, 518, 545, 593, 613, 640, 663, 704, 722,
+  746, 765, 51, 59, 66, 76, 89, 99, 119, 131, 149, 168, 181, 200, 267, 289,
+  310, 325, 369, 395, 419, 436, 487, 517, 544, 563, 612, 639, 662, 678, 721,
+  745, 764, 777, 61, 69, 75, 87, 100, 114, 129, 144, 162, 180, 191, 207, 288,
+  309, 324, 333, 394, 418, 435, 445, 516, 543, 562, 573, 638, 661, 677, 686,
+  744, 763, 776, 783, 70, 79, 86, 97, 108, 122, 137, 155, 242, 251, 266, 287,
+  339, 350, 368, 393, 452, 465, 486, 515, 580, 592, 611, 637, 692, 703, 720,
+  743, 788, 798, 813, 833, 84, 93, 103, 110, 125, 141, 154, 171, 250, 265, 286,
+  308, 349, 367, 392, 417, 464, 485, 514, 542, 591, 610, 636, 660, 702, 719,
+  742, 762, 797, 812, 832, 848, 98, 106, 115, 127, 143, 156, 169, 185, 264,
+  285, 307, 323, 366, 391, 416, 434, 484, 513, 541, 561, 609, 635, 659, 676,
+  718, 741, 761, 775, 811, 831, 847, 858, 117, 128, 136, 148, 160, 175, 188,
+  198, 284, 306, 322, 332, 390, 415, 433, 444, 512, 540, 560, 572, 634, 658,
+  675, 685, 740, 760, 774, 782, 830, 846, 857, 863, 135, 146, 152, 165, 241,
+  249, 263, 283, 338, 348, 365, 389, 451, 463, 483, 511, 579, 590, 608, 633,
+  691, 701, 717, 739, 787, 796, 810, 829, 867, 875, 887, 903, 153, 166, 174,
+  183, 248, 262, 282, 305, 347, 364, 388, 414, 462, 482, 510, 539, 589, 607,
+  632, 657, 700, 716, 738, 759, 795, 809, 828, 845, 874, 886, 902, 915, 176,
+  187, 195, 202, 261, 281, 304, 321, 363, 387, 413, 432, 481, 509, 538, 559,
+  606, 631, 656, 674, 715, 737, 758, 773, 808, 827, 844, 856, 885, 901, 914,
+  923, 192, 199, 206, 213, 280, 303, 320, 331, 386, 412, 431, 443, 508, 537,
+  558, 571, 630, 655, 673, 684, 736, 757, 772, 781, 826, 843, 855, 862, 900,
+  913, 922, 927, 240, 247, 260, 279, 337, 346, 362, 385, 450, 461, 480, 507,
+  578, 588, 605, 629, 690, 699, 714, 735, 786, 794, 807, 825, 866, 873, 884,
+  899, 930, 936, 945, 957, 246, 259, 278, 302, 345, 361, 384, 411, 460, 479,
+  506, 536, 587, 604, 628, 654, 698, 713, 734, 756, 793, 806, 824, 842, 872,
+  883, 898, 912, 935, 944, 956, 966, 258, 277, 301, 319, 360, 383, 410, 430,
+  478, 505, 535, 557, 603, 627, 653, 672, 712, 733, 755, 771, 805, 823, 841,
+  854, 882, 897, 911, 921, 943, 955, 965, 972, 276, 300, 318, 330, 382, 409,
+  429, 442, 504, 534, 556, 570, 626, 652, 671, 683, 732, 754, 770, 780, 822,
+  840, 853, 861, 896, 910, 920, 926, 954, 964, 971, 975, 336, 344, 359, 381,
+  449, 459, 477, 503, 577, 586, 602, 625, 689, 697, 711, 731, 785, 792, 804,
+  821, 865, 871, 881, 895, 929, 934, 942, 953, 977, 981, 987, 995, 343, 358,
+  380, 408, 458, 476, 502, 533, 585, 601, 624, 651, 696, 710, 730, 753, 791,
+  803, 820, 839, 870, 880, 894, 909, 933, 941, 952, 963, 980, 986, 994, 1001,
+  357, 379, 407, 428, 475, 501, 532, 555, 600, 623, 650, 670, 709, 729, 752,
+  769, 802, 819, 838, 852, 879, 893, 908, 919, 940, 951, 962, 970, 985, 993,
+  1000, 1005, 378, 406, 427, 441, 500, 531, 554, 569, 622, 649, 669, 682, 728,
+  751, 768, 779, 818, 837, 851, 860, 892, 907, 918, 925, 950, 961, 969, 974,
+  992, 999, 1004, 1007, 448, 457, 474, 499, 576, 584, 599, 621, 688, 695, 708,
+  727, 784, 790, 801, 817, 864, 869, 878, 891, 928, 932, 939, 949, 976, 979,
+  984, 991, 1008, 1010, 1013, 1017, 456, 473, 498, 530, 583, 598, 620, 648,
+  694, 707, 726, 750, 789, 800, 816, 836, 868, 877, 890, 906, 931, 938, 948,
+  960, 978, 983, 990, 998, 1009, 1012, 1016, 1020, 472, 497, 529, 553, 597,
+  619, 647, 668, 706, 725, 749, 767, 799, 815, 835, 850, 876, 889, 905, 917,
+  937, 947, 959, 968, 982, 989, 997, 1003, 1011, 1015, 1019, 1022, 496, 528,
+  552, 568, 618, 646, 667, 681, 724, 748, 766, 778, 814, 834, 849, 859, 888,
+  904, 916, 924, 946, 958, 967, 973, 988, 996, 1002, 1006, 1014, 1018, 1021,
+  1023,
+};
+
+const scan_order vp9_default_scan_orders[TX_SIZES] = {
+  {default_scan_4x4,   vp9_default_iscan_4x4,   default_scan_4x4_neighbors},
+  {default_scan_8x8,   vp9_default_iscan_8x8,   default_scan_8x8_neighbors},
+  {default_scan_16x16, vp9_default_iscan_16x16, default_scan_16x16_neighbors},
+  {default_scan_32x32, vp9_default_iscan_32x32, default_scan_32x32_neighbors},
+};
+
+const scan_order vp9_scan_orders[TX_SIZES][TX_TYPES] = {
+  {  // TX_4X4
+    {default_scan_4x4, vp9_default_iscan_4x4, default_scan_4x4_neighbors},
+    {row_scan_4x4,     vp9_row_iscan_4x4,     row_scan_4x4_neighbors},
+    {col_scan_4x4,     vp9_col_iscan_4x4,     col_scan_4x4_neighbors},
+    {default_scan_4x4, vp9_default_iscan_4x4, default_scan_4x4_neighbors}
+  }, {  // TX_8X8
+    {default_scan_8x8, vp9_default_iscan_8x8, default_scan_8x8_neighbors},
+    {row_scan_8x8,     vp9_row_iscan_8x8,     row_scan_8x8_neighbors},
+    {col_scan_8x8,     vp9_col_iscan_8x8,     col_scan_8x8_neighbors},
+    {default_scan_8x8, vp9_default_iscan_8x8, default_scan_8x8_neighbors}
+  }, {  // TX_16X16
+    {default_scan_16x16, vp9_default_iscan_16x16, default_scan_16x16_neighbors},
+    {row_scan_16x16,     vp9_row_iscan_16x16,     row_scan_16x16_neighbors},
+    {col_scan_16x16,     vp9_col_iscan_16x16,     col_scan_16x16_neighbors},
+    {default_scan_16x16, vp9_default_iscan_16x16, default_scan_16x16_neighbors}
+  }, {  // TX_32X32
+    {default_scan_32x32, vp9_default_iscan_32x32, default_scan_32x32_neighbors},
+    {default_scan_32x32, vp9_default_iscan_32x32, default_scan_32x32_neighbors},
+    {default_scan_32x32, vp9_default_iscan_32x32, default_scan_32x32_neighbors},
+    {default_scan_32x32, vp9_default_iscan_32x32, default_scan_32x32_neighbors},
+  }
+};
diff --git a/libvpx/libvpx/vp9/common/vp9_scan.h b/libvpx/libvpx/vp9/common/vp9_scan.h
new file mode 100644
index 0000000..4c1ee81
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_scan.h
@@ -0,0 +1,57 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_SCAN_H_
+#define VP9_COMMON_VP9_SCAN_H_
+
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_enums.h"
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_NEIGHBORS 2
+
+typedef struct {
+  const int16_t *scan;
+  const int16_t *iscan;
+  const int16_t *neighbors;
+} scan_order;
+
+extern const scan_order vp9_default_scan_orders[TX_SIZES];
+extern const scan_order vp9_scan_orders[TX_SIZES][TX_TYPES];
+
+static INLINE int get_coef_context(const int16_t *neighbors,
+                                   const uint8_t *token_cache, int c) {
+  return (1 + token_cache[neighbors[MAX_NEIGHBORS * c + 0]] +
+          token_cache[neighbors[MAX_NEIGHBORS * c + 1]]) >> 1;
+}
+
+static INLINE const scan_order *get_scan(const MACROBLOCKD *xd, TX_SIZE tx_size,
+                                         PLANE_TYPE type, int block_idx) {
+  const MODE_INFO *const mi = xd->mi[0];
+
+  if (is_inter_block(mi) || type != PLANE_TYPE_Y || xd->lossless) {
+    return &vp9_default_scan_orders[tx_size];
+  } else {
+    const PREDICTION_MODE mode = get_y_mode(mi, block_idx);
+    return &vp9_scan_orders[tx_size][intra_mode_to_tx_type_lookup[mode]];
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_SCAN_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_seg_common.c b/libvpx/libvpx/vp9/common/vp9_seg_common.c
new file mode 100644
index 0000000..7af6162
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_seg_common.c
@@ -0,0 +1,64 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_quant_common.h"
+
+static const int seg_feature_data_signed[SEG_LVL_MAX] = { 1, 1, 0, 0 };
+
+static const int seg_feature_data_max[SEG_LVL_MAX] = {
+  MAXQ, MAX_LOOP_FILTER, 3, 0 };
+
+// These functions provide access to new segment level features.
+// Eventually these function may be "optimized out" but for the moment,
+// the coding mechanism is still subject to change so these provide a
+// convenient single point of change.
+
+void vp9_clearall_segfeatures(struct segmentation *seg) {
+  vp9_zero(seg->feature_data);
+  vp9_zero(seg->feature_mask);
+  seg->aq_av_offset = 0;
+}
+
+void vp9_enable_segfeature(struct segmentation *seg, int segment_id,
+                           SEG_LVL_FEATURES feature_id) {
+  seg->feature_mask[segment_id] |= 1 << feature_id;
+}
+
+int vp9_seg_feature_data_max(SEG_LVL_FEATURES feature_id) {
+  return seg_feature_data_max[feature_id];
+}
+
+int vp9_is_segfeature_signed(SEG_LVL_FEATURES feature_id) {
+  return seg_feature_data_signed[feature_id];
+}
+
+void vp9_set_segdata(struct segmentation *seg, int segment_id,
+                     SEG_LVL_FEATURES feature_id, int seg_data) {
+  assert(seg_data <= seg_feature_data_max[feature_id]);
+  if (seg_data < 0) {
+    assert(seg_feature_data_signed[feature_id]);
+    assert(-seg_data <= seg_feature_data_max[feature_id]);
+  }
+
+  seg->feature_data[segment_id][feature_id] = seg_data;
+}
+
+const vpx_tree_index vp9_segment_tree[TREE_SIZE(MAX_SEGMENTS)] = {
+  2,  4,  6,  8, 10, 12,
+  0, -1, -2, -3, -4, -5, -6, -7
+};
+
+
+// TBD? Functions to read and write segment data with range / validity checking
diff --git a/libvpx/libvpx/vp9/common/vp9_seg_common.h b/libvpx/libvpx/vp9/common/vp9_seg_common.h
new file mode 100644
index 0000000..99a9440
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_seg_common.h
@@ -0,0 +1,87 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_SEG_COMMON_H_
+#define VP9_COMMON_VP9_SEG_COMMON_H_
+
+#include "vpx_dsp/prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define SEGMENT_DELTADATA   0
+#define SEGMENT_ABSDATA     1
+
+#define MAX_SEGMENTS     8
+#define SEG_TREE_PROBS   (MAX_SEGMENTS-1)
+
+#define PREDICTION_PROBS 3
+
+// Segment level features.
+typedef enum {
+  SEG_LVL_ALT_Q = 0,               // Use alternate Quantizer ....
+  SEG_LVL_ALT_LF = 1,              // Use alternate loop filter value...
+  SEG_LVL_REF_FRAME = 2,           // Optional Segment reference frame
+  SEG_LVL_SKIP = 3,                // Optional Segment (0,0) + skip mode
+  SEG_LVL_MAX = 4                  // Number of features supported
+} SEG_LVL_FEATURES;
+
+
+struct segmentation {
+  uint8_t enabled;
+  uint8_t update_map;
+  uint8_t update_data;
+  uint8_t abs_delta;
+  uint8_t temporal_update;
+
+  vpx_prob tree_probs[SEG_TREE_PROBS];
+  vpx_prob pred_probs[PREDICTION_PROBS];
+
+  int16_t feature_data[MAX_SEGMENTS][SEG_LVL_MAX];
+  uint32_t feature_mask[MAX_SEGMENTS];
+  int aq_av_offset;
+};
+
+static INLINE int segfeature_active(const struct segmentation *seg,
+                                    int segment_id,
+                                    SEG_LVL_FEATURES feature_id) {
+  return seg->enabled &&
+         (seg->feature_mask[segment_id] & (1 << feature_id));
+}
+
+void vp9_clearall_segfeatures(struct segmentation *seg);
+
+void vp9_enable_segfeature(struct segmentation *seg,
+                           int segment_id,
+                           SEG_LVL_FEATURES feature_id);
+
+int vp9_seg_feature_data_max(SEG_LVL_FEATURES feature_id);
+
+int vp9_is_segfeature_signed(SEG_LVL_FEATURES feature_id);
+
+void vp9_set_segdata(struct segmentation *seg,
+                     int segment_id,
+                     SEG_LVL_FEATURES feature_id,
+                     int seg_data);
+
+static INLINE int get_segdata(const struct segmentation *seg, int segment_id,
+                              SEG_LVL_FEATURES feature_id) {
+  return seg->feature_data[segment_id][feature_id];
+}
+
+extern const vpx_tree_index vp9_segment_tree[TREE_SIZE(MAX_SEGMENTS)];
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_SEG_COMMON_H_
+
diff --git a/libvpx/libvpx/vp9/common/vp9_textblit.c b/libvpx/libvpx/vp9/common/vp9_textblit.c
new file mode 100644
index 0000000..60e95e0
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_textblit.c
@@ -0,0 +1,120 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "vp9/common/vp9_textblit.h"
+
+static const int font[] = {
+  0x0, 0x5C00, 0x8020, 0xAFABEA, 0xD7EC0, 0x1111111, 0x1855740, 0x18000,
+  0x45C0, 0x74400, 0x51140, 0x23880, 0xC4000, 0x21080, 0x80000, 0x111110,
+  0xE9D72E, 0x87E40, 0x12AD732, 0xAAD62A, 0x4F94C4, 0x4D6B7, 0x456AA,
+  0x3E8423, 0xAAD6AA, 0xAAD6A2, 0x2800, 0x2A00, 0x8A880, 0x52940, 0x22A20,
+  0x15422, 0x6AD62E, 0x1E4A53E, 0xAAD6BF, 0x8C62E, 0xE8C63F, 0x118D6BF,
+  0x1094BF, 0xCAC62E, 0x1F2109F, 0x118FE31, 0xF8C628, 0x8A89F, 0x108421F,
+  0x1F1105F, 0x1F4105F, 0xE8C62E, 0x2294BF, 0x164C62E, 0x12694BF, 0x8AD6A2,
+  0x10FC21, 0x1F8421F, 0x744107, 0xF8220F, 0x1151151, 0x117041, 0x119D731,
+  0x47E0, 0x1041041, 0xFC400, 0x10440, 0x1084210, 0x820
+};
+
+static void plot(int x, int y, unsigned char *image, int pitch) {
+  image[x + y * pitch] ^= 255;
+}
+
+void vp9_blit_text(const char *msg, unsigned char *address, const int pitch) {
+  int letter_bitmap;
+  unsigned char *output_pos = address;
+  int colpos = 0;
+
+  while (msg[colpos] != 0) {
+    char letter = msg[colpos];
+    int fontcol, fontrow;
+
+    if (letter <= 'Z' && letter >= ' ')
+      letter_bitmap = font[letter - ' '];
+    else if (letter <= 'z' && letter >= 'a')
+      letter_bitmap = font[letter - 'a' + 'A' - ' '];
+    else
+      letter_bitmap = font[0];
+
+    for (fontcol = 6; fontcol >= 0; fontcol--)
+      for (fontrow = 0; fontrow < 5; fontrow++)
+        output_pos[fontrow * pitch + fontcol] =
+          ((letter_bitmap >> (fontcol * 5)) & (1 << fontrow) ? 255 : 0);
+
+    output_pos += 7;
+    colpos++;
+  }
+}
+
+
+
+/* Bresenham line algorithm */
+void vp9_blit_line(int x0, int x1, int y0, int y1, unsigned char *image,
+                   int pitch) {
+  int steep = abs(y1 - y0) > abs(x1 - x0);
+  int deltax, deltay;
+  int error, ystep, y, x;
+
+  if (steep) {
+    int t;
+    t = x0;
+    x0 = y0;
+    y0 = t;
+
+    t = x1;
+    x1 = y1;
+    y1 = t;
+  }
+
+  if (x0 > x1) {
+    int t;
+    t = x0;
+    x0 = x1;
+    x1 = t;
+
+    t = y0;
+    y0 = y1;
+    y1 = t;
+  }
+
+  deltax = x1 - x0;
+  deltay = abs(y1 - y0);
+  error  = deltax / 2;
+
+  y = y0;
+
+  if (y0 < y1)
+    ystep = 1;
+  else
+    ystep = -1;
+
+  if (steep) {
+    for (x = x0; x <= x1; x++) {
+      plot(y, x, image, pitch);
+
+      error = error - deltay;
+      if (error < 0) {
+        y = y + ystep;
+        error = error + deltax;
+      }
+    }
+  } else {
+    for (x = x0; x <= x1; x++) {
+      plot(x, y, image, pitch);
+
+      error = error - deltay;
+      if (error < 0) {
+        y = y + ystep;
+        error = error + deltax;
+      }
+    }
+  }
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_textblit.h b/libvpx/libvpx/vp9/common/vp9_textblit.h
new file mode 100644
index 0000000..158ec1b
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_textblit.h
@@ -0,0 +1,27 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_TEXTBLIT_H_
+#define VP9_COMMON_VP9_TEXTBLIT_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_blit_text(const char *msg, unsigned char *address, int pitch);
+
+void vp9_blit_line(int x0, int x1, int y0, int y1, unsigned char *image,
+                   int pitch);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_TEXTBLIT_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_thread_common.c b/libvpx/libvpx/vp9/common/vp9_thread_common.c
new file mode 100644
index 0000000..db78d6b
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_thread_common.c
@@ -0,0 +1,435 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_thread_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_loopfilter.h"
+
+#if CONFIG_MULTITHREAD
+static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
+  const int kMaxTryLocks = 4000;
+  int locked = 0;
+  int i;
+
+  for (i = 0; i < kMaxTryLocks; ++i) {
+    if (!pthread_mutex_trylock(mutex)) {
+      locked = 1;
+      break;
+    }
+  }
+
+  if (!locked)
+    pthread_mutex_lock(mutex);
+}
+#endif  // CONFIG_MULTITHREAD
+
+static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {
+#if CONFIG_MULTITHREAD
+  const int nsync = lf_sync->sync_range;
+
+  if (r && !(c & (nsync - 1))) {
+    pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
+    mutex_lock(mutex);
+
+    while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
+      pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
+    }
+    pthread_mutex_unlock(mutex);
+  }
+#else
+  (void)lf_sync;
+  (void)r;
+  (void)c;
+#endif  // CONFIG_MULTITHREAD
+}
+
+static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
+                              const int sb_cols) {
+#if CONFIG_MULTITHREAD
+  const int nsync = lf_sync->sync_range;
+  int cur;
+  // Only signal when there are enough filtered SB for next row to run.
+  int sig = 1;
+
+  if (c < sb_cols - 1) {
+    cur = c;
+    if (c % nsync)
+      sig = 0;
+  } else {
+    cur = sb_cols + nsync;
+  }
+
+  if (sig) {
+    mutex_lock(&lf_sync->mutex_[r]);
+
+    lf_sync->cur_sb_col[r] = cur;
+
+    pthread_cond_signal(&lf_sync->cond_[r]);
+    pthread_mutex_unlock(&lf_sync->mutex_[r]);
+  }
+#else
+  (void)lf_sync;
+  (void)r;
+  (void)c;
+  (void)sb_cols;
+#endif  // CONFIG_MULTITHREAD
+}
+
+// Implement row loopfiltering for each thread.
+static INLINE
+void thread_loop_filter_rows(const YV12_BUFFER_CONFIG *const frame_buffer,
+                             VP9_COMMON *const cm,
+                             struct macroblockd_plane planes[MAX_MB_PLANE],
+                             int start, int stop, int y_only,
+                             VP9LfSync *const lf_sync) {
+  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
+  const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
+  int mi_row, mi_col;
+  enum lf_path path;
+  if (y_only)
+    path = LF_PATH_444;
+  else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
+    path = LF_PATH_420;
+  else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
+    path = LF_PATH_444;
+  else
+    path = LF_PATH_SLOW;
+
+  for (mi_row = start; mi_row < stop;
+       mi_row += lf_sync->num_workers * MI_BLOCK_SIZE) {
+    MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+    LOOP_FILTER_MASK *lfm = get_lfm(&cm->lf, mi_row, 0);
+
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE, ++lfm) {
+      const int r = mi_row >> MI_BLOCK_SIZE_LOG2;
+      const int c = mi_col >> MI_BLOCK_SIZE_LOG2;
+      int plane;
+
+      sync_read(lf_sync, r, c);
+
+      vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
+
+      vp9_adjust_mask(cm, mi_row, mi_col, lfm);
+
+      vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, lfm);
+      for (plane = 1; plane < num_planes; ++plane) {
+        switch (path) {
+          case LF_PATH_420:
+            vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, lfm);
+            break;
+          case LF_PATH_444:
+            vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, lfm);
+            break;
+          case LF_PATH_SLOW:
+            vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
+                                          mi_row, mi_col);
+            break;
+        }
+      }
+
+      sync_write(lf_sync, r, c, sb_cols);
+    }
+  }
+}
+
+// Row-based multi-threaded loopfilter hook
+static int loop_filter_row_worker(VP9LfSync *const lf_sync,
+                                  LFWorkerData *const lf_data) {
+  thread_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
+                          lf_data->start, lf_data->stop, lf_data->y_only,
+                          lf_sync);
+  return 1;
+}
+
+static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame,
+                                VP9_COMMON *cm,
+                                struct macroblockd_plane planes[MAX_MB_PLANE],
+                                int start, int stop, int y_only,
+                                VPxWorker *workers, int nworkers,
+                                VP9LfSync *lf_sync) {
+  const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
+  // Number of superblock rows and cols
+  const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
+  // Decoder may allocate more threads than number of tiles based on user's
+  // input.
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int num_workers = VPXMIN(nworkers, tile_cols);
+  int i;
+
+  if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
+      num_workers > lf_sync->num_workers) {
+    vp9_loop_filter_dealloc(lf_sync);
+    vp9_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
+  }
+
+  // Initialize cur_sb_col to -1 for all SB rows.
+  memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
+
+  // Set up loopfilter thread data.
+  // The decoder is capping num_workers because it has been observed that using
+  // more threads on the loopfilter than there are cores will hurt performance
+  // on Android. This is because the system will only schedule the tile decode
+  // workers on cores equal to the number of tile columns. Then if the decoder
+  // tries to use more threads for the loopfilter, it will hurt performance
+  // because of contention. If the multithreading code changes in the future
+  // then the number of workers used by the loopfilter should be revisited.
+  for (i = 0; i < num_workers; ++i) {
+    VPxWorker *const worker = &workers[i];
+    LFWorkerData *const lf_data = &lf_sync->lfdata[i];
+
+    worker->hook = (VPxWorkerHook)loop_filter_row_worker;
+    worker->data1 = lf_sync;
+    worker->data2 = lf_data;
+
+    // Loopfilter data
+    vp9_loop_filter_data_reset(lf_data, frame, cm, planes);
+    lf_data->start = start + i * MI_BLOCK_SIZE;
+    lf_data->stop = stop;
+    lf_data->y_only = y_only;
+
+    // Start loopfiltering
+    if (i == num_workers - 1) {
+      winterface->execute(worker);
+    } else {
+      winterface->launch(worker);
+    }
+  }
+
+  // Wait till all rows are finished
+  for (i = 0; i < num_workers; ++i) {
+    winterface->sync(&workers[i]);
+  }
+}
+
+void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+                              VP9_COMMON *cm,
+                              struct macroblockd_plane planes[MAX_MB_PLANE],
+                              int frame_filter_level,
+                              int y_only, int partial_frame,
+                              VPxWorker *workers, int num_workers,
+                              VP9LfSync *lf_sync) {
+  int start_mi_row, end_mi_row, mi_rows_to_filter;
+
+  if (!frame_filter_level) return;
+
+  start_mi_row = 0;
+  mi_rows_to_filter = cm->mi_rows;
+  if (partial_frame && cm->mi_rows > 8) {
+    start_mi_row = cm->mi_rows >> 1;
+    start_mi_row &= 0xfffffff8;
+    mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
+  }
+  end_mi_row = start_mi_row + mi_rows_to_filter;
+  vp9_loop_filter_frame_init(cm, frame_filter_level);
+
+  loop_filter_rows_mt(frame, cm, planes, start_mi_row, end_mi_row,
+                      y_only, workers, num_workers, lf_sync);
+}
+
+// Set up nsync by width.
+static INLINE int get_sync_range(int width) {
+  // nsync numbers are picked by testing. For example, for 4k
+  // video, using 4 gives best performance.
+  if (width < 640)
+    return 1;
+  else if (width <= 1280)
+    return 2;
+  else if (width <= 4096)
+    return 4;
+  else
+    return 8;
+}
+
+// Allocate memory for lf row synchronization
+void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows,
+                           int width, int num_workers) {
+  lf_sync->rows = rows;
+#if CONFIG_MULTITHREAD
+  {
+    int i;
+
+    CHECK_MEM_ERROR(cm, lf_sync->mutex_,
+                    vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
+    if (lf_sync->mutex_) {
+      for (i = 0; i < rows; ++i) {
+        pthread_mutex_init(&lf_sync->mutex_[i], NULL);
+      }
+    }
+
+    CHECK_MEM_ERROR(cm, lf_sync->cond_,
+                    vpx_malloc(sizeof(*lf_sync->cond_) * rows));
+    if (lf_sync->cond_) {
+      for (i = 0; i < rows; ++i) {
+        pthread_cond_init(&lf_sync->cond_[i], NULL);
+      }
+    }
+  }
+#endif  // CONFIG_MULTITHREAD
+
+  CHECK_MEM_ERROR(cm, lf_sync->lfdata,
+                  vpx_malloc(num_workers * sizeof(*lf_sync->lfdata)));
+  lf_sync->num_workers = num_workers;
+
+  CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
+                  vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
+
+  // Set up nsync.
+  lf_sync->sync_range = get_sync_range(width);
+}
+
+// Deallocate lf synchronization related mutex and data
+void vp9_loop_filter_dealloc(VP9LfSync *lf_sync) {
+  if (lf_sync != NULL) {
+#if CONFIG_MULTITHREAD
+    int i;
+
+    if (lf_sync->mutex_ != NULL) {
+      for (i = 0; i < lf_sync->rows; ++i) {
+        pthread_mutex_destroy(&lf_sync->mutex_[i]);
+      }
+      vpx_free(lf_sync->mutex_);
+    }
+    if (lf_sync->cond_ != NULL) {
+      for (i = 0; i < lf_sync->rows; ++i) {
+        pthread_cond_destroy(&lf_sync->cond_[i]);
+      }
+      vpx_free(lf_sync->cond_);
+    }
+#endif  // CONFIG_MULTITHREAD
+    vpx_free(lf_sync->lfdata);
+    vpx_free(lf_sync->cur_sb_col);
+    // clear the structure as the source of this call may be a resize in which
+    // case this call will be followed by an _alloc() which may fail.
+    vp9_zero(*lf_sync);
+  }
+}
+
+// Accumulate frame counts.
+void vp9_accumulate_frame_counts(FRAME_COUNTS *accum,
+                                 const FRAME_COUNTS *counts, int is_dec) {
+  int i, j, k, l, m;
+
+  for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
+    for (j = 0; j < INTRA_MODES; j++)
+      accum->y_mode[i][j] += counts->y_mode[i][j];
+
+  for (i = 0; i < INTRA_MODES; i++)
+    for (j = 0; j < INTRA_MODES; j++)
+      accum->uv_mode[i][j] += counts->uv_mode[i][j];
+
+  for (i = 0; i < PARTITION_CONTEXTS; i++)
+    for (j = 0; j < PARTITION_TYPES; j++)
+      accum->partition[i][j] += counts->partition[i][j];
+
+  if (is_dec) {
+    int n;
+    for (i = 0; i < TX_SIZES; i++)
+      for (j = 0; j < PLANE_TYPES; j++)
+        for (k = 0; k < REF_TYPES; k++)
+          for (l = 0; l < COEF_BANDS; l++)
+            for (m = 0; m < COEFF_CONTEXTS; m++) {
+              accum->eob_branch[i][j][k][l][m] +=
+                  counts->eob_branch[i][j][k][l][m];
+              for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
+                accum->coef[i][j][k][l][m][n] +=
+                    counts->coef[i][j][k][l][m][n];
+            }
+  } else {
+    for (i = 0; i < TX_SIZES; i++)
+      for (j = 0; j < PLANE_TYPES; j++)
+        for (k = 0; k < REF_TYPES; k++)
+          for (l = 0; l < COEF_BANDS; l++)
+            for (m = 0; m < COEFF_CONTEXTS; m++)
+              accum->eob_branch[i][j][k][l][m] +=
+                  counts->eob_branch[i][j][k][l][m];
+                // In the encoder, coef is only updated at frame
+                // level, so not need to accumulate it here.
+                // for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
+                //   accum->coef[i][j][k][l][m][n] +=
+                //       counts->coef[i][j][k][l][m][n];
+  }
+
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+    for (j = 0; j < SWITCHABLE_FILTERS; j++)
+      accum->switchable_interp[i][j] += counts->switchable_interp[i][j];
+
+  for (i = 0; i < INTER_MODE_CONTEXTS; i++)
+    for (j = 0; j < INTER_MODES; j++)
+      accum->inter_mode[i][j] += counts->inter_mode[i][j];
+
+  for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      accum->intra_inter[i][j] += counts->intra_inter[i][j];
+
+  for (i = 0; i < COMP_INTER_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      accum->comp_inter[i][j] += counts->comp_inter[i][j];
+
+  for (i = 0; i < REF_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      for (k = 0; k < 2; k++)
+      accum->single_ref[i][j][k] += counts->single_ref[i][j][k];
+
+  for (i = 0; i < REF_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      accum->comp_ref[i][j] += counts->comp_ref[i][j];
+
+  for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
+    for (j = 0; j < TX_SIZES; j++)
+      accum->tx.p32x32[i][j] += counts->tx.p32x32[i][j];
+
+    for (j = 0; j < TX_SIZES - 1; j++)
+      accum->tx.p16x16[i][j] += counts->tx.p16x16[i][j];
+
+    for (j = 0; j < TX_SIZES - 2; j++)
+      accum->tx.p8x8[i][j] += counts->tx.p8x8[i][j];
+  }
+
+  for (i = 0; i < TX_SIZES; i++)
+    accum->tx.tx_totals[i] += counts->tx.tx_totals[i];
+
+  for (i = 0; i < SKIP_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      accum->skip[i][j] += counts->skip[i][j];
+
+  for (i = 0; i < MV_JOINTS; i++)
+    accum->mv.joints[i] += counts->mv.joints[i];
+
+  for (k = 0; k < 2; k++) {
+    nmv_component_counts *const comps = &accum->mv.comps[k];
+    const nmv_component_counts *const comps_t = &counts->mv.comps[k];
+
+    for (i = 0; i < 2; i++) {
+      comps->sign[i] += comps_t->sign[i];
+      comps->class0_hp[i] += comps_t->class0_hp[i];
+      comps->hp[i] += comps_t->hp[i];
+    }
+
+    for (i = 0; i < MV_CLASSES; i++)
+      comps->classes[i] += comps_t->classes[i];
+
+    for (i = 0; i < CLASS0_SIZE; i++) {
+      comps->class0[i] += comps_t->class0[i];
+      for (j = 0; j < MV_FP_SIZE; j++)
+        comps->class0_fp[i][j] += comps_t->class0_fp[i][j];
+    }
+
+    for (i = 0; i < MV_OFFSET_BITS; i++)
+      for (j = 0; j < 2; j++)
+        comps->bits[i][j] += comps_t->bits[i][j];
+
+    for (i = 0; i < MV_FP_SIZE; i++)
+      comps->fp[i] += comps_t->fp[i];
+  }
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_thread_common.h b/libvpx/libvpx/vp9/common/vp9_thread_common.h
new file mode 100644
index 0000000..b3b60c2
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_thread_common.h
@@ -0,0 +1,65 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_THREAD_COMMON_H_
+#define VP9_COMMON_VP9_THREAD_COMMON_H_
+#include "./vpx_config.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vpx_util/vpx_thread.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9Common;
+struct FRAME_COUNTS;
+
+// Loopfilter row synchronization
+typedef struct VP9LfSyncData {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t *mutex_;
+  pthread_cond_t *cond_;
+#endif
+  // Allocate memory to store the loop-filtered superblock index in each row.
+  int *cur_sb_col;
+  // The optimal sync_range for different resolution and platform should be
+  // determined by testing. Currently, it is chosen to be a power-of-2 number.
+  int sync_range;
+  int rows;
+
+  // Row-based parallel loopfilter data
+  LFWorkerData *lfdata;
+  int num_workers;
+} VP9LfSync;
+
+// Allocate memory for loopfilter row synchronization.
+void vp9_loop_filter_alloc(VP9LfSync *lf_sync, struct VP9Common *cm, int rows,
+                           int width, int num_workers);
+
+// Deallocate loopfilter synchronization related mutex and data.
+void vp9_loop_filter_dealloc(VP9LfSync *lf_sync);
+
+// Multi-threaded loopfilter that uses the tile threads.
+void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+                              struct VP9Common *cm,
+                              struct macroblockd_plane planes[MAX_MB_PLANE],
+                              int frame_filter_level,
+                              int y_only, int partial_frame,
+                              VPxWorker *workers, int num_workers,
+                              VP9LfSync *lf_sync);
+
+void vp9_accumulate_frame_counts(struct FRAME_COUNTS *accum,
+                                 const struct FRAME_COUNTS *counts, int is_dec);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_THREAD_COMMON_H_
diff --git a/libvpx/libvpx/vp9/common/vp9_tile_common.c b/libvpx/libvpx/vp9/common/vp9_tile_common.c
new file mode 100644
index 0000000..9fcb97c
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_tile_common.c
@@ -0,0 +1,59 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_tile_common.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+#define MIN_TILE_WIDTH_B64 4
+#define MAX_TILE_WIDTH_B64 64
+
+static int get_tile_offset(int idx, int mis, int log2) {
+  const int sb_cols = mi_cols_aligned_to_sb(mis) >> MI_BLOCK_SIZE_LOG2;
+  const int offset = ((idx * sb_cols) >> log2) << MI_BLOCK_SIZE_LOG2;
+  return VPXMIN(offset, mis);
+}
+
+void vp9_tile_set_row(TileInfo *tile, const VP9_COMMON *cm, int row) {
+  tile->mi_row_start = get_tile_offset(row, cm->mi_rows, cm->log2_tile_rows);
+  tile->mi_row_end = get_tile_offset(row + 1, cm->mi_rows, cm->log2_tile_rows);
+}
+
+void vp9_tile_set_col(TileInfo *tile, const VP9_COMMON *cm, int col) {
+  tile->mi_col_start = get_tile_offset(col, cm->mi_cols, cm->log2_tile_cols);
+  tile->mi_col_end = get_tile_offset(col + 1, cm->mi_cols, cm->log2_tile_cols);
+}
+
+void vp9_tile_init(TileInfo *tile, const VP9_COMMON *cm, int row, int col) {
+  vp9_tile_set_row(tile, cm, row);
+  vp9_tile_set_col(tile, cm, col);
+}
+
+static int get_min_log2_tile_cols(const int sb64_cols) {
+  int min_log2 = 0;
+  while ((MAX_TILE_WIDTH_B64 << min_log2) < sb64_cols)
+    ++min_log2;
+  return min_log2;
+}
+
+static int get_max_log2_tile_cols(const int sb64_cols) {
+  int max_log2 = 1;
+  while ((sb64_cols >> max_log2) >= MIN_TILE_WIDTH_B64)
+    ++max_log2;
+  return max_log2 - 1;
+}
+
+void vp9_get_tile_n_bits(int mi_cols,
+                         int *min_log2_tile_cols, int *max_log2_tile_cols) {
+  const int sb64_cols = mi_cols_aligned_to_sb(mi_cols) >> MI_BLOCK_SIZE_LOG2;
+  *min_log2_tile_cols = get_min_log2_tile_cols(sb64_cols);
+  *max_log2_tile_cols = get_max_log2_tile_cols(sb64_cols);
+  assert(*min_log2_tile_cols <= *max_log2_tile_cols);
+}
diff --git a/libvpx/libvpx/vp9/common/vp9_tile_common.h b/libvpx/libvpx/vp9/common/vp9_tile_common.h
new file mode 100644
index 0000000..ae58805
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/vp9_tile_common.h
@@ -0,0 +1,40 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_TILE_COMMON_H_
+#define VP9_COMMON_VP9_TILE_COMMON_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9Common;
+
+typedef struct TileInfo {
+  int mi_row_start, mi_row_end;
+  int mi_col_start, mi_col_end;
+} TileInfo;
+
+// initializes 'tile->mi_(row|col)_(start|end)' for (row, col) based on
+// 'cm->log2_tile_(rows|cols)' & 'cm->mi_(rows|cols)'
+void vp9_tile_init(TileInfo *tile, const struct VP9Common *cm,
+                   int row, int col);
+
+void vp9_tile_set_row(TileInfo *tile, const struct VP9Common *cm, int row);
+void vp9_tile_set_col(TileInfo *tile, const struct VP9Common *cm, int col);
+
+void vp9_get_tile_n_bits(int mi_cols,
+                         int *min_log2_tile_cols, int *max_log2_tile_cols);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_TILE_COMMON_H_
diff --git a/libvpx/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.c b/libvpx/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.c
new file mode 100644
index 0000000..1c77b57
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.c
@@ -0,0 +1,181 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "vpx_dsp/x86/inv_txfm_sse2.h"
+#include "vpx_dsp/x86/txfm_common_sse2.h"
+#include "vpx_ports/mem.h"
+
+void vp9_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  __m128i in[2];
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i eight = _mm_set1_epi16(8);
+
+  in[0] = load_input_data(input);
+  in[1] = load_input_data(input + 8);
+
+  switch (tx_type) {
+    case 0:  // DCT_DCT
+      idct4_sse2(in);
+      idct4_sse2(in);
+      break;
+    case 1:  // ADST_DCT
+      idct4_sse2(in);
+      iadst4_sse2(in);
+      break;
+    case 2:  // DCT_ADST
+      iadst4_sse2(in);
+      idct4_sse2(in);
+      break;
+    case 3:  // ADST_ADST
+      iadst4_sse2(in);
+      iadst4_sse2(in);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  // Final round and shift
+  in[0] = _mm_add_epi16(in[0], eight);
+  in[1] = _mm_add_epi16(in[1], eight);
+
+  in[0] = _mm_srai_epi16(in[0], 4);
+  in[1] = _mm_srai_epi16(in[1], 4);
+
+  // Reconstruction and Store
+  {
+    __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest));
+    __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2));
+    d0 = _mm_unpacklo_epi32(d0,
+                            _mm_cvtsi32_si128(*(const int *)(dest + stride)));
+    d2 = _mm_unpacklo_epi32(
+        d2, _mm_cvtsi32_si128(*(const int *)(dest + stride * 3)));
+    d0 = _mm_unpacklo_epi8(d0, zero);
+    d2 = _mm_unpacklo_epi8(d2, zero);
+    d0 = _mm_add_epi16(d0, in[0]);
+    d2 = _mm_add_epi16(d2, in[1]);
+    d0 = _mm_packus_epi16(d0, d2);
+    // store result[0]
+    *(int *)dest = _mm_cvtsi128_si32(d0);
+    // store result[1]
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride) = _mm_cvtsi128_si32(d0);
+    // store result[2]
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0);
+    // store result[3]
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0);
+  }
+}
+
+void vp9_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  __m128i in[8];
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+
+  // load input data
+  in[0] = load_input_data(input);
+  in[1] = load_input_data(input + 8 * 1);
+  in[2] = load_input_data(input + 8 * 2);
+  in[3] = load_input_data(input + 8 * 3);
+  in[4] = load_input_data(input + 8 * 4);
+  in[5] = load_input_data(input + 8 * 5);
+  in[6] = load_input_data(input + 8 * 6);
+  in[7] = load_input_data(input + 8 * 7);
+
+  switch (tx_type) {
+    case 0:  // DCT_DCT
+      idct8_sse2(in);
+      idct8_sse2(in);
+      break;
+    case 1:  // ADST_DCT
+      idct8_sse2(in);
+      iadst8_sse2(in);
+      break;
+    case 2:  // DCT_ADST
+      iadst8_sse2(in);
+      idct8_sse2(in);
+      break;
+    case 3:  // ADST_ADST
+      iadst8_sse2(in);
+      iadst8_sse2(in);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+  in[4] = _mm_adds_epi16(in[4], final_rounding);
+  in[5] = _mm_adds_epi16(in[5], final_rounding);
+  in[6] = _mm_adds_epi16(in[6], final_rounding);
+  in[7] = _mm_adds_epi16(in[7], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 5);
+  in[1] = _mm_srai_epi16(in[1], 5);
+  in[2] = _mm_srai_epi16(in[2], 5);
+  in[3] = _mm_srai_epi16(in[3], 5);
+  in[4] = _mm_srai_epi16(in[4], 5);
+  in[5] = _mm_srai_epi16(in[5], 5);
+  in[6] = _mm_srai_epi16(in[6], 5);
+  in[7] = _mm_srai_epi16(in[7], 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in[0]);
+  RECON_AND_STORE(dest + 1 * stride, in[1]);
+  RECON_AND_STORE(dest + 2 * stride, in[2]);
+  RECON_AND_STORE(dest + 3 * stride, in[3]);
+  RECON_AND_STORE(dest + 4 * stride, in[4]);
+  RECON_AND_STORE(dest + 5 * stride, in[5]);
+  RECON_AND_STORE(dest + 6 * stride, in[6]);
+  RECON_AND_STORE(dest + 7 * stride, in[7]);
+}
+
+void vp9_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest,
+                               int stride, int tx_type) {
+  __m128i in0[16], in1[16];
+
+  load_buffer_8x16(input, in0);
+  input += 8;
+  load_buffer_8x16(input, in1);
+
+  switch (tx_type) {
+    case 0:  // DCT_DCT
+      idct16_sse2(in0, in1);
+      idct16_sse2(in0, in1);
+      break;
+    case 1:  // ADST_DCT
+      idct16_sse2(in0, in1);
+      iadst16_sse2(in0, in1);
+      break;
+    case 2:  // DCT_ADST
+      iadst16_sse2(in0, in1);
+      idct16_sse2(in0, in1);
+      break;
+    case 3:  // ADST_ADST
+      iadst16_sse2(in0, in1);
+      iadst16_sse2(in0, in1);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  write_buffer_8x16(dest, in0, stride);
+  dest += 8;
+  write_buffer_8x16(dest, in1, stride);
+}
diff --git a/libvpx/libvpx/vp9/common/x86/vp9_mfqe_sse2.asm b/libvpx/libvpx/vp9/common/x86/vp9_mfqe_sse2.asm
new file mode 100644
index 0000000..6029420
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/x86/vp9_mfqe_sse2.asm
@@ -0,0 +1,287 @@
+;
+;  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+;  This file is a duplicate of mfqe_sse2.asm in VP8.
+;  TODO(jackychen): Find a way to fix the duplicate.
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vp9_filter_by_weight16x16_sse2
+;(
+;    unsigned char *src,
+;    int            src_stride,
+;    unsigned char *dst,
+;    int            dst_stride,
+;    int            src_weight
+;)
+global sym(vp9_filter_by_weight16x16_sse2) PRIVATE
+sym(vp9_filter_by_weight16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movd        xmm0, arg(4)                ; src_weight
+    pshuflw     xmm0, xmm0, 0x0             ; replicate to all low words
+    punpcklqdq  xmm0, xmm0                  ; replicate to all hi words
+
+    movdqa      xmm1, [GLOBAL(tMFQE)]
+    psubw       xmm1, xmm0                  ; dst_weight
+
+    mov         rax, arg(0)                 ; src
+    mov         rsi, arg(1)                 ; src_stride
+    mov         rdx, arg(2)                 ; dst
+    mov         rdi, arg(3)                 ; dst_stride
+
+    mov         rcx, 16                     ; loop count
+    pxor        xmm6, xmm6
+
+.combine
+    movdqa      xmm2, [rax]
+    movdqa      xmm4, [rdx]
+    add         rax, rsi
+
+    ; src * src_weight
+    movdqa      xmm3, xmm2
+    punpcklbw   xmm2, xmm6
+    punpckhbw   xmm3, xmm6
+    pmullw      xmm2, xmm0
+    pmullw      xmm3, xmm0
+
+    ; dst * dst_weight
+    movdqa      xmm5, xmm4
+    punpcklbw   xmm4, xmm6
+    punpckhbw   xmm5, xmm6
+    pmullw      xmm4, xmm1
+    pmullw      xmm5, xmm1
+
+    ; sum, round and shift
+    paddw       xmm2, xmm4
+    paddw       xmm3, xmm5
+    paddw       xmm2, [GLOBAL(tMFQE_round)]
+    paddw       xmm3, [GLOBAL(tMFQE_round)]
+    psrlw       xmm2, 4
+    psrlw       xmm3, 4
+
+    packuswb    xmm2, xmm3
+    movdqa      [rdx], xmm2
+    add         rdx, rdi
+
+    dec         rcx
+    jnz         .combine
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+
+    ret
+
+;void vp9_filter_by_weight8x8_sse2
+;(
+;    unsigned char *src,
+;    int            src_stride,
+;    unsigned char *dst,
+;    int            dst_stride,
+;    int            src_weight
+;)
+global sym(vp9_filter_by_weight8x8_sse2) PRIVATE
+sym(vp9_filter_by_weight8x8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movd        xmm0, arg(4)                ; src_weight
+    pshuflw     xmm0, xmm0, 0x0             ; replicate to all low words
+    punpcklqdq  xmm0, xmm0                  ; replicate to all hi words
+
+    movdqa      xmm1, [GLOBAL(tMFQE)]
+    psubw       xmm1, xmm0                  ; dst_weight
+
+    mov         rax, arg(0)                 ; src
+    mov         rsi, arg(1)                 ; src_stride
+    mov         rdx, arg(2)                 ; dst
+    mov         rdi, arg(3)                 ; dst_stride
+
+    mov         rcx, 8                      ; loop count
+    pxor        xmm4, xmm4
+
+.combine
+    movq        xmm2, [rax]
+    movq        xmm3, [rdx]
+    add         rax, rsi
+
+    ; src * src_weight
+    punpcklbw   xmm2, xmm4
+    pmullw      xmm2, xmm0
+
+    ; dst * dst_weight
+    punpcklbw   xmm3, xmm4
+    pmullw      xmm3, xmm1
+
+    ; sum, round and shift
+    paddw       xmm2, xmm3
+    paddw       xmm2, [GLOBAL(tMFQE_round)]
+    psrlw       xmm2, 4
+
+    packuswb    xmm2, xmm4
+    movq        [rdx], xmm2
+    add         rdx, rdi
+
+    dec         rcx
+    jnz         .combine
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+
+    ret
+
+;void vp9_variance_and_sad_16x16_sse2 | arg
+;(
+;    unsigned char *src1,          0
+;    int            stride1,       1
+;    unsigned char *src2,          2
+;    int            stride2,       3
+;    unsigned int  *variance,      4
+;    unsigned int  *sad,           5
+;)
+global sym(vp9_variance_and_sad_16x16_sse2) PRIVATE
+sym(vp9_variance_and_sad_16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov         rax,        arg(0)          ; src1
+    mov         rcx,        arg(1)          ; stride1
+    mov         rdx,        arg(2)          ; src2
+    mov         rdi,        arg(3)          ; stride2
+
+    mov         rsi,        16              ; block height
+
+    ; Prep accumulator registers
+    pxor        xmm3, xmm3                  ; SAD
+    pxor        xmm4, xmm4                  ; sum of src2
+    pxor        xmm5, xmm5                  ; sum of src2^2
+
+    ; Because we're working with the actual output frames
+    ; we can't depend on any kind of data alignment.
+.accumulate
+    movdqa      xmm0, [rax]                 ; src1
+    movdqa      xmm1, [rdx]                 ; src2
+    add         rax, rcx                    ; src1 + stride1
+    add         rdx, rdi                    ; src2 + stride2
+
+    ; SAD(src1, src2)
+    psadbw      xmm0, xmm1
+    paddusw     xmm3, xmm0
+
+    ; SUM(src2)
+    pxor        xmm2, xmm2
+    psadbw      xmm2, xmm1                  ; sum src2 by misusing SAD against 0
+    paddusw     xmm4, xmm2
+
+    ; pmaddubsw would be ideal if it took two unsigned values. instead,
+    ; it expects a signed and an unsigned value. so instead we zero extend
+    ; and operate on words.
+    pxor        xmm2, xmm2
+    movdqa      xmm0, xmm1
+    punpcklbw   xmm0, xmm2
+    punpckhbw   xmm1, xmm2
+    pmaddwd     xmm0, xmm0
+    pmaddwd     xmm1, xmm1
+    paddd       xmm5, xmm0
+    paddd       xmm5, xmm1
+
+    sub         rsi,        1
+    jnz         .accumulate
+
+    ; phaddd only operates on adjacent double words.
+    ; Finalize SAD and store
+    movdqa      xmm0, xmm3
+    psrldq      xmm0, 8
+    paddusw     xmm0, xmm3
+    paddd       xmm0, [GLOBAL(t128)]
+    psrld       xmm0, 8
+
+    mov         rax,  arg(5)
+    movd        [rax], xmm0
+
+    ; Accumulate sum of src2
+    movdqa      xmm0, xmm4
+    psrldq      xmm0, 8
+    paddusw     xmm0, xmm4
+    ; Square src2. Ignore high value
+    pmuludq     xmm0, xmm0
+    psrld       xmm0, 8
+
+    ; phaddw could be used to sum adjacent values but we want
+    ; all the values summed. promote to doubles, accumulate,
+    ; shift and sum
+    pxor        xmm2, xmm2
+    movdqa      xmm1, xmm5
+    punpckldq   xmm1, xmm2
+    punpckhdq   xmm5, xmm2
+    paddd       xmm1, xmm5
+    movdqa      xmm2, xmm1
+    psrldq      xmm1, 8
+    paddd       xmm1, xmm2
+
+    psubd       xmm1, xmm0
+
+    ; (variance + 128) >> 8
+    paddd       xmm1, [GLOBAL(t128)]
+    psrld       xmm1, 8
+    mov         rax,  arg(4)
+
+    movd        [rax], xmm1
+
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+t128:
+%ifndef __NASM_VER__
+    ddq 128
+%elif CONFIG_BIG_ENDIAN
+    dq  0, 128
+%else
+    dq  128, 0
+%endif
+align 16
+tMFQE: ; 1 << MFQE_PRECISION
+    times 8 dw 0x10
+align 16
+tMFQE_round: ; 1 << (MFQE_PRECISION - 1)
+    times 8 dw 0x08
diff --git a/libvpx/libvpx/vp9/common/x86/vp9_postproc_sse2.asm b/libvpx/libvpx/vp9/common/x86/vp9_postproc_sse2.asm
new file mode 100644
index 0000000..4307628
--- /dev/null
+++ b/libvpx/libvpx/vp9/common/x86/vp9_postproc_sse2.asm
@@ -0,0 +1,632 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vp9_post_proc_down_and_across_xmm
+;(
+;    unsigned char *src_ptr,
+;    unsigned char *dst_ptr,
+;    int src_pixels_per_line,
+;    int dst_pixels_per_line,
+;    int rows,
+;    int cols,
+;    int flimit
+;)
+global sym(vp9_post_proc_down_and_across_xmm) PRIVATE
+sym(vp9_post_proc_down_and_across_xmm):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+%if ABI_IS_32BIT=1 && CONFIG_PIC=1
+    ALIGN_STACK 16, rax
+    ; move the global rd onto the stack, since we don't have enough registers
+    ; to do PIC addressing
+    movdqa      xmm0, [GLOBAL(rd42)]
+    sub         rsp, 16
+    movdqa      [rsp], xmm0
+%define RD42 [rsp]
+%else
+%define RD42 [GLOBAL(rd42)]
+%endif
+
+
+        movd        xmm2,       dword ptr arg(6) ;flimit
+        punpcklwd   xmm2,       xmm2
+        punpckldq   xmm2,       xmm2
+        punpcklqdq  xmm2,       xmm2
+
+        mov         rsi,        arg(0) ;src_ptr
+        mov         rdi,        arg(1) ;dst_ptr
+
+        movsxd      rcx,        DWORD PTR arg(4) ;rows
+        movsxd      rax,        DWORD PTR arg(2) ;src_pixels_per_line ; destination pitch?
+        pxor        xmm0,       xmm0              ; mm0 = 00000000
+
+.nextrow:
+
+        xor         rdx,        rdx       ; clear out rdx for use as loop counter
+.nextcol:
+        movq        xmm3,       QWORD PTR [rsi]         ; mm4 = r0 p0..p7
+        punpcklbw   xmm3,       xmm0                    ; mm3 = p0..p3
+        movdqa      xmm1,       xmm3                    ; mm1 = p0..p3
+        psllw       xmm3,       2                       ;
+
+        movq        xmm5,       QWORD PTR [rsi + rax]   ; mm4 = r1 p0..p7
+        punpcklbw   xmm5,       xmm0                    ; mm5 = r1 p0..p3
+        paddusw     xmm3,       xmm5                    ; mm3 += mm6
+
+        ; thresholding
+        movdqa      xmm7,       xmm1                    ; mm7 = r0 p0..p3
+        psubusw     xmm7,       xmm5                    ; mm7 = r0 p0..p3 - r1 p0..p3
+        psubusw     xmm5,       xmm1                    ; mm5 = r1 p0..p3 - r0 p0..p3
+        paddusw     xmm7,       xmm5                    ; mm7 = abs(r0 p0..p3 - r1 p0..p3)
+        pcmpgtw     xmm7,       xmm2
+
+        movq        xmm5,       QWORD PTR [rsi + 2*rax] ; mm4 = r2 p0..p7
+        punpcklbw   xmm5,       xmm0                    ; mm5 = r2 p0..p3
+        paddusw     xmm3,       xmm5                    ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1                    ; mm6 = r0 p0..p3
+        psubusw     xmm6,       xmm5                    ; mm6 = r0 p0..p3 - r2 p0..p3
+        psubusw     xmm5,       xmm1                    ; mm5 = r2 p0..p3 - r2 p0..p3
+        paddusw     xmm6,       xmm5                    ; mm6 = abs(r0 p0..p3 - r2 p0..p3)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6                    ; accumulate thresholds
+
+
+        neg         rax
+        movq        xmm5,       QWORD PTR [rsi+2*rax]   ; mm4 = r-2 p0..p7
+        punpcklbw   xmm5,       xmm0                    ; mm5 = r-2 p0..p3
+        paddusw     xmm3,       xmm5                    ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1                    ; mm6 = r0 p0..p3
+        psubusw     xmm6,       xmm5                    ; mm6 = p0..p3 - r-2 p0..p3
+        psubusw     xmm5,       xmm1                    ; mm5 = r-2 p0..p3 - p0..p3
+        paddusw     xmm6,       xmm5                    ; mm6 = abs(r0 p0..p3 - r-2 p0..p3)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6                    ; accumulate thresholds
+
+        movq        xmm4,       QWORD PTR [rsi+rax]     ; mm4 = r-1 p0..p7
+        punpcklbw   xmm4,       xmm0                    ; mm4 = r-1 p0..p3
+        paddusw     xmm3,       xmm4                    ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1                    ; mm6 = r0 p0..p3
+        psubusw     xmm6,       xmm4                    ; mm6 = p0..p3 - r-2 p0..p3
+        psubusw     xmm4,       xmm1                    ; mm5 = r-1 p0..p3 - p0..p3
+        paddusw     xmm6,       xmm4                    ; mm6 = abs(r0 p0..p3 - r-1 p0..p3)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6                    ; accumulate thresholds
+
+
+        paddusw     xmm3,       RD42                    ; mm3 += round value
+        psraw       xmm3,       3                       ; mm3 /= 8
+
+        pand        xmm1,       xmm7                    ; mm1 select vals > thresh from source
+        pandn       xmm7,       xmm3                    ; mm7 select vals < thresh from blurred result
+        paddusw     xmm1,       xmm7                    ; combination
+
+        packuswb    xmm1,       xmm0                    ; pack to bytes
+        movq        QWORD PTR [rdi], xmm1             ;
+
+        neg         rax                   ; pitch is positive
+        add         rsi,        8
+        add         rdi,        8
+
+        add         rdx,        8
+        cmp         edx,        dword arg(5) ;cols
+
+        jl          .nextcol
+
+        ; done with the all cols, start the across filtering in place
+        sub         rsi,        rdx
+        sub         rdi,        rdx
+
+        xor         rdx,        rdx
+        movq        mm0,        QWORD PTR [rdi-8];
+
+.acrossnextcol:
+        movq        xmm7,       QWORD PTR [rdi +rdx -2]
+        movd        xmm4,       DWORD PTR [rdi +rdx +6]
+
+        pslldq      xmm4,       8
+        por         xmm4,       xmm7
+
+        movdqa      xmm3,       xmm4
+        psrldq      xmm3,       2
+        punpcklbw   xmm3,       xmm0              ; mm3 = p0..p3
+        movdqa      xmm1,       xmm3              ; mm1 = p0..p3
+        psllw       xmm3,       2
+
+
+        movdqa      xmm5,       xmm4
+        psrldq      xmm5,       3
+        punpcklbw   xmm5,       xmm0              ; mm5 = p1..p4
+        paddusw     xmm3,       xmm5              ; mm3 += mm6
+
+        ; thresholding
+        movdqa      xmm7,       xmm1              ; mm7 = p0..p3
+        psubusw     xmm7,       xmm5              ; mm7 = p0..p3 - p1..p4
+        psubusw     xmm5,       xmm1              ; mm5 = p1..p4 - p0..p3
+        paddusw     xmm7,       xmm5              ; mm7 = abs(p0..p3 - p1..p4)
+        pcmpgtw     xmm7,       xmm2
+
+        movdqa      xmm5,       xmm4
+        psrldq      xmm5,       4
+        punpcklbw   xmm5,       xmm0              ; mm5 = p2..p5
+        paddusw     xmm3,       xmm5              ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1              ; mm6 = p0..p3
+        psubusw     xmm6,       xmm5              ; mm6 = p0..p3 - p1..p4
+        psubusw     xmm5,       xmm1              ; mm5 = p1..p4 - p0..p3
+        paddusw     xmm6,       xmm5              ; mm6 = abs(p0..p3 - p1..p4)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6              ; accumulate thresholds
+
+
+        movdqa      xmm5,       xmm4              ; mm5 = p-2..p5
+        punpcklbw   xmm5,       xmm0              ; mm5 = p-2..p1
+        paddusw     xmm3,       xmm5              ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1              ; mm6 = p0..p3
+        psubusw     xmm6,       xmm5              ; mm6 = p0..p3 - p1..p4
+        psubusw     xmm5,       xmm1              ; mm5 = p1..p4 - p0..p3
+        paddusw     xmm6,       xmm5              ; mm6 = abs(p0..p3 - p1..p4)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6              ; accumulate thresholds
+
+        psrldq      xmm4,       1                   ; mm4 = p-1..p5
+        punpcklbw   xmm4,       xmm0              ; mm4 = p-1..p2
+        paddusw     xmm3,       xmm4              ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1              ; mm6 = p0..p3
+        psubusw     xmm6,       xmm4              ; mm6 = p0..p3 - p1..p4
+        psubusw     xmm4,       xmm1              ; mm5 = p1..p4 - p0..p3
+        paddusw     xmm6,       xmm4              ; mm6 = abs(p0..p3 - p1..p4)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6              ; accumulate thresholds
+
+        paddusw     xmm3,       RD42              ; mm3 += round value
+        psraw       xmm3,       3                 ; mm3 /= 8
+
+        pand        xmm1,       xmm7              ; mm1 select vals > thresh from source
+        pandn       xmm7,       xmm3              ; mm7 select vals < thresh from blurred result
+        paddusw     xmm1,       xmm7              ; combination
+
+        packuswb    xmm1,       xmm0              ; pack to bytes
+        movq        QWORD PTR [rdi+rdx-8],  mm0   ; store previous four bytes
+        movdq2q     mm0,        xmm1
+
+        add         rdx,        8
+        cmp         edx,        dword arg(5) ;cols
+        jl          .acrossnextcol;
+
+        ; last 8 pixels
+        movq        QWORD PTR [rdi+rdx-8],  mm0
+
+        ; done with this rwo
+        add         rsi,rax               ; next line
+        mov         eax, dword arg(3) ;dst_pixels_per_line ; destination pitch?
+        add         rdi,rax               ; next destination
+        mov         eax, dword arg(2) ;src_pixels_per_line ; destination pitch?
+
+        dec         rcx                   ; decrement count
+        jnz         .nextrow              ; next row
+
+%if ABI_IS_32BIT=1 && CONFIG_PIC=1
+    add rsp,16
+    pop rsp
+%endif
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%undef RD42
+
+
+;void vp9_mbpost_proc_down_xmm(unsigned char *dst,
+;                            int pitch, int rows, int cols,int flimit)
+extern sym(vp9_rv)
+global sym(vp9_mbpost_proc_down_xmm) PRIVATE
+sym(vp9_mbpost_proc_down_xmm):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 128+16
+
+    ; unsigned char d[16][8] at [rsp]
+    ; create flimit2 at [rsp+128]
+    mov         eax, dword ptr arg(4) ;flimit
+    mov         [rsp+128], eax
+    mov         [rsp+128+4], eax
+    mov         [rsp+128+8], eax
+    mov         [rsp+128+12], eax
+%define flimit4 [rsp+128]
+
+%if ABI_IS_32BIT=0
+    lea         r8,       [GLOBAL(sym(vp9_rv))]
+%endif
+
+    ;rows +=8;
+    add         dword arg(2), 8
+
+    ;for(c=0; c<cols; c+=8)
+.loop_col:
+            mov         rsi,        arg(0) ; s
+            pxor        xmm0,       xmm0        ;
+
+            movsxd      rax,        dword ptr arg(1) ;pitch       ;
+            neg         rax                                     ; rax = -pitch
+
+            lea         rsi,        [rsi + rax*8];              ; rdi = s[-pitch*8]
+            neg         rax
+
+
+            pxor        xmm5,       xmm5
+            pxor        xmm6,       xmm6        ;
+
+            pxor        xmm7,       xmm7        ;
+            mov         rdi,        rsi
+
+            mov         rcx,        15          ;
+
+.loop_initvar:
+            movq        xmm1,       QWORD PTR [rdi];
+            punpcklbw   xmm1,       xmm0        ;
+
+            paddw       xmm5,       xmm1        ;
+            pmullw      xmm1,       xmm1        ;
+
+            movdqa      xmm2,       xmm1        ;
+            punpcklwd   xmm1,       xmm0        ;
+
+            punpckhwd   xmm2,       xmm0        ;
+            paddd       xmm6,       xmm1        ;
+
+            paddd       xmm7,       xmm2        ;
+            lea         rdi,        [rdi+rax]   ;
+
+            dec         rcx
+            jne         .loop_initvar
+            ;save the var and sum
+            xor         rdx,        rdx
+.loop_row:
+            movq        xmm1,       QWORD PTR [rsi]     ; [s-pitch*8]
+            movq        xmm2,       QWORD PTR [rdi]     ; [s+pitch*7]
+
+            punpcklbw   xmm1,       xmm0
+            punpcklbw   xmm2,       xmm0
+
+            paddw       xmm5,       xmm2
+            psubw       xmm5,       xmm1
+
+            pmullw      xmm2,       xmm2
+            movdqa      xmm4,       xmm2
+
+            punpcklwd   xmm2,       xmm0
+            punpckhwd   xmm4,       xmm0
+
+            paddd       xmm6,       xmm2
+            paddd       xmm7,       xmm4
+
+            pmullw      xmm1,       xmm1
+            movdqa      xmm2,       xmm1
+
+            punpcklwd   xmm1,       xmm0
+            psubd       xmm6,       xmm1
+
+            punpckhwd   xmm2,       xmm0
+            psubd       xmm7,       xmm2
+
+
+            movdqa      xmm3,       xmm6
+            pslld       xmm3,       4
+
+            psubd       xmm3,       xmm6
+            movdqa      xmm1,       xmm5
+
+            movdqa      xmm4,       xmm5
+            pmullw      xmm1,       xmm1
+
+            pmulhw      xmm4,       xmm4
+            movdqa      xmm2,       xmm1
+
+            punpcklwd   xmm1,       xmm4
+            punpckhwd   xmm2,       xmm4
+
+            movdqa      xmm4,       xmm7
+            pslld       xmm4,       4
+
+            psubd       xmm4,       xmm7
+
+            psubd       xmm3,       xmm1
+            psubd       xmm4,       xmm2
+
+            psubd       xmm3,       flimit4
+            psubd       xmm4,       flimit4
+
+            psrad       xmm3,       31
+            psrad       xmm4,       31
+
+            packssdw    xmm3,       xmm4
+            packsswb    xmm3,       xmm0
+
+            movq        xmm1,       QWORD PTR [rsi+rax*8]
+
+            movq        xmm2,       xmm1
+            punpcklbw   xmm1,       xmm0
+
+            paddw       xmm1,       xmm5
+            mov         rcx,        rdx
+
+            and         rcx,        127
+%if ABI_IS_32BIT=1 && CONFIG_PIC=1
+            push        rax
+            lea         rax,        [GLOBAL(sym(vp9_rv))]
+            movdqu      xmm4,       [rax + rcx*2] ;vp9_rv[rcx*2]
+            pop         rax
+%elif ABI_IS_32BIT=0
+            movdqu      xmm4,       [r8 + rcx*2] ;vp9_rv[rcx*2]
+%else
+            movdqu      xmm4,       [sym(vp9_rv) + rcx*2]
+%endif
+
+            paddw       xmm1,       xmm4
+            ;paddw     xmm1,       eight8s
+            psraw       xmm1,       4
+
+            packuswb    xmm1,       xmm0
+            pand        xmm1,       xmm3
+
+            pandn       xmm3,       xmm2
+            por         xmm1,       xmm3
+
+            and         rcx,        15
+            movq        QWORD PTR   [rsp + rcx*8], xmm1 ;d[rcx*8]
+
+            mov         rcx,        rdx
+            sub         rcx,        8
+
+            and         rcx,        15
+            movq        mm0,        [rsp + rcx*8] ;d[rcx*8]
+
+            movq        [rsi],      mm0
+            lea         rsi,        [rsi+rax]
+
+            lea         rdi,        [rdi+rax]
+            add         rdx,        1
+
+            cmp         edx,        dword arg(2) ;rows
+            jl          .loop_row
+
+        add         dword arg(0), 8 ; s += 8
+        sub         dword arg(3), 8 ; cols -= 8
+        cmp         dword arg(3), 0
+        jg          .loop_col
+
+    add         rsp, 128+16
+    pop         rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%undef flimit4
+
+
+;void vp9_mbpost_proc_across_ip_xmm(unsigned char *src,
+;                                int pitch, int rows, int cols,int flimit)
+global sym(vp9_mbpost_proc_across_ip_xmm) PRIVATE
+sym(vp9_mbpost_proc_across_ip_xmm):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16
+
+    ; create flimit4 at [rsp]
+    mov         eax, dword ptr arg(4) ;flimit
+    mov         [rsp], eax
+    mov         [rsp+4], eax
+    mov         [rsp+8], eax
+    mov         [rsp+12], eax
+%define flimit4 [rsp]
+
+
+    ;for(r=0;r<rows;r++)
+.ip_row_loop:
+
+        xor         rdx,    rdx ;sumsq=0;
+        xor         rcx,    rcx ;sum=0;
+        mov         rsi,    arg(0); s
+        mov         rdi,    -8
+.ip_var_loop:
+        ;for(i=-8;i<=6;i++)
+        ;{
+        ;    sumsq += s[i]*s[i];
+        ;    sum   += s[i];
+        ;}
+        movzx       eax, byte [rsi+rdi]
+        add         ecx, eax
+        mul         al
+        add         edx, eax
+        add         rdi, 1
+        cmp         rdi, 6
+        jle         .ip_var_loop
+
+
+            ;mov         rax,    sumsq
+            ;movd        xmm7,   rax
+            movd        xmm7,   edx
+
+            ;mov         rax,    sum
+            ;movd        xmm6,   rax
+            movd        xmm6,   ecx
+
+            mov         rsi,    arg(0) ;s
+            xor         rcx,    rcx
+
+            movsxd      rdx,    dword arg(3) ;cols
+            add         rdx,    8
+            pxor        mm0,    mm0
+            pxor        mm1,    mm1
+
+            pxor        xmm0,   xmm0
+.nextcol4:
+
+            movd        xmm1,   DWORD PTR [rsi+rcx-8]   ; -8 -7 -6 -5
+            movd        xmm2,   DWORD PTR [rsi+rcx+7]   ; +7 +8 +9 +10
+
+            punpcklbw   xmm1,   xmm0                    ; expanding
+            punpcklbw   xmm2,   xmm0                    ; expanding
+
+            punpcklwd   xmm1,   xmm0                    ; expanding to dwords
+            punpcklwd   xmm2,   xmm0                    ; expanding to dwords
+
+            psubd       xmm2,   xmm1                    ; 7--8   8--7   9--6 10--5
+            paddd       xmm1,   xmm1                    ; -8*2   -7*2   -6*2 -5*2
+
+            paddd       xmm1,   xmm2                    ; 7+-8   8+-7   9+-6 10+-5
+            pmaddwd     xmm1,   xmm2                    ; squared of 7+-8   8+-7   9+-6 10+-5
+
+            paddd       xmm6,   xmm2
+            paddd       xmm7,   xmm1
+
+            pshufd      xmm6,   xmm6,   0               ; duplicate the last ones
+            pshufd      xmm7,   xmm7,   0               ; duplicate the last ones
+
+            psrldq      xmm1,       4                   ; 8--7   9--6 10--5  0000
+            psrldq      xmm2,       4                   ; 8--7   9--6 10--5  0000
+
+            pshufd      xmm3,   xmm1,   3               ; 0000  8--7   8--7   8--7 squared
+            pshufd      xmm4,   xmm2,   3               ; 0000  8--7   8--7   8--7 squared
+
+            paddd       xmm6,   xmm4
+            paddd       xmm7,   xmm3
+
+            pshufd      xmm3,   xmm1,   01011111b       ; 0000  0000   9--6   9--6 squared
+            pshufd      xmm4,   xmm2,   01011111b       ; 0000  0000   9--6   9--6 squared
+
+            paddd       xmm7,   xmm3
+            paddd       xmm6,   xmm4
+
+            pshufd      xmm3,   xmm1,   10111111b       ; 0000  0000   8--7   8--7 squared
+            pshufd      xmm4,   xmm2,   10111111b       ; 0000  0000   8--7   8--7 squared
+
+            paddd       xmm7,   xmm3
+            paddd       xmm6,   xmm4
+
+            movdqa      xmm3,   xmm6
+            pmaddwd     xmm3,   xmm3
+
+            movdqa      xmm5,   xmm7
+            pslld       xmm5,   4
+
+            psubd       xmm5,   xmm7
+            psubd       xmm5,   xmm3
+
+            psubd       xmm5,   flimit4
+            psrad       xmm5,   31
+
+            packssdw    xmm5,   xmm0
+            packsswb    xmm5,   xmm0
+
+            movd        xmm1,   DWORD PTR [rsi+rcx]
+            movq        xmm2,   xmm1
+
+            punpcklbw   xmm1,   xmm0
+            punpcklwd   xmm1,   xmm0
+
+            paddd       xmm1,   xmm6
+            paddd       xmm1,   [GLOBAL(four8s)]
+
+            psrad       xmm1,   4
+            packssdw    xmm1,   xmm0
+
+            packuswb    xmm1,   xmm0
+            pand        xmm1,   xmm5
+
+            pandn       xmm5,   xmm2
+            por         xmm5,   xmm1
+
+            movd        [rsi+rcx-8],  mm0
+            movq        mm0,    mm1
+
+            movdq2q     mm1,    xmm5
+            psrldq      xmm7,   12
+
+            psrldq      xmm6,   12
+            add         rcx,    4
+
+            cmp         rcx,    rdx
+            jl          .nextcol4
+
+        ;s+=pitch;
+        movsxd rax, dword arg(1)
+        add    arg(0), rax
+
+        sub dword arg(2), 1 ;rows-=1
+        cmp dword arg(2), 0
+        jg .ip_row_loop
+
+    add         rsp, 16
+    pop         rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%undef flimit4
+
+
+SECTION_RODATA
+align 16
+rd42:
+    times 8 dw 0x04
+four8s:
+    times 4 dd 8
diff --git a/libvpx/libvpx/vp9/decoder/vp9_decodeframe.c b/libvpx/libvpx/vp9/decoder/vp9_decodeframe.c
new file mode 100644
index 0000000..d639129
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_decodeframe.c
@@ -0,0 +1,2271 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdlib.h>  // qsort()
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+
+#include "vpx_dsp/bitreader_buffer.h"
+#include "vpx_dsp/bitreader.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/mem_ops.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vpx_util/vpx_thread.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_idct.h"
+#include "vp9/common/vp9_thread_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_tile_common.h"
+
+#include "vp9/decoder/vp9_decodeframe.h"
+#include "vp9/decoder/vp9_detokenize.h"
+#include "vp9/decoder/vp9_decodemv.h"
+#include "vp9/decoder/vp9_decoder.h"
+#include "vp9/decoder/vp9_dsubexp.h"
+
+#define MAX_VP9_HEADER_SIZE 80
+
+static int is_compound_reference_allowed(const VP9_COMMON *cm) {
+  int i;
+  for (i = 1; i < REFS_PER_FRAME; ++i)
+    if (cm->ref_frame_sign_bias[i + 1] != cm->ref_frame_sign_bias[1])
+      return 1;
+
+  return 0;
+}
+
+static void setup_compound_reference_mode(VP9_COMMON *cm) {
+  if (cm->ref_frame_sign_bias[LAST_FRAME] ==
+          cm->ref_frame_sign_bias[GOLDEN_FRAME]) {
+    cm->comp_fixed_ref = ALTREF_FRAME;
+    cm->comp_var_ref[0] = LAST_FRAME;
+    cm->comp_var_ref[1] = GOLDEN_FRAME;
+  } else if (cm->ref_frame_sign_bias[LAST_FRAME] ==
+                 cm->ref_frame_sign_bias[ALTREF_FRAME]) {
+    cm->comp_fixed_ref = GOLDEN_FRAME;
+    cm->comp_var_ref[0] = LAST_FRAME;
+    cm->comp_var_ref[1] = ALTREF_FRAME;
+  } else {
+    cm->comp_fixed_ref = LAST_FRAME;
+    cm->comp_var_ref[0] = GOLDEN_FRAME;
+    cm->comp_var_ref[1] = ALTREF_FRAME;
+  }
+}
+
+static int read_is_valid(const uint8_t *start, size_t len, const uint8_t *end) {
+  return len != 0 && len <= (size_t)(end - start);
+}
+
+static int decode_unsigned_max(struct vpx_read_bit_buffer *rb, int max) {
+  const int data = vpx_rb_read_literal(rb, get_unsigned_bits(max));
+  return data > max ? max : data;
+}
+
+static TX_MODE read_tx_mode(vpx_reader *r) {
+  TX_MODE tx_mode = vpx_read_literal(r, 2);
+  if (tx_mode == ALLOW_32X32)
+    tx_mode += vpx_read_bit(r);
+  return tx_mode;
+}
+
+static void read_tx_mode_probs(struct tx_probs *tx_probs, vpx_reader *r) {
+  int i, j;
+
+  for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
+    for (j = 0; j < TX_SIZES - 3; ++j)
+      vp9_diff_update_prob(r, &tx_probs->p8x8[i][j]);
+
+  for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
+    for (j = 0; j < TX_SIZES - 2; ++j)
+      vp9_diff_update_prob(r, &tx_probs->p16x16[i][j]);
+
+  for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
+    for (j = 0; j < TX_SIZES - 1; ++j)
+      vp9_diff_update_prob(r, &tx_probs->p32x32[i][j]);
+}
+
+static void read_switchable_interp_probs(FRAME_CONTEXT *fc, vpx_reader *r) {
+  int i, j;
+  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
+    for (i = 0; i < SWITCHABLE_FILTERS - 1; ++i)
+      vp9_diff_update_prob(r, &fc->switchable_interp_prob[j][i]);
+}
+
+static void read_inter_mode_probs(FRAME_CONTEXT *fc, vpx_reader *r) {
+  int i, j;
+  for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
+    for (j = 0; j < INTER_MODES - 1; ++j)
+      vp9_diff_update_prob(r, &fc->inter_mode_probs[i][j]);
+}
+
+static REFERENCE_MODE read_frame_reference_mode(const VP9_COMMON *cm,
+                                                vpx_reader *r) {
+  if (is_compound_reference_allowed(cm)) {
+    return vpx_read_bit(r) ? (vpx_read_bit(r) ? REFERENCE_MODE_SELECT
+                                              : COMPOUND_REFERENCE)
+                           : SINGLE_REFERENCE;
+  } else {
+    return SINGLE_REFERENCE;
+  }
+}
+
+static void read_frame_reference_mode_probs(VP9_COMMON *cm, vpx_reader *r) {
+  FRAME_CONTEXT *const fc = cm->fc;
+  int i;
+
+  if (cm->reference_mode == REFERENCE_MODE_SELECT)
+    for (i = 0; i < COMP_INTER_CONTEXTS; ++i)
+      vp9_diff_update_prob(r, &fc->comp_inter_prob[i]);
+
+  if (cm->reference_mode != COMPOUND_REFERENCE)
+    for (i = 0; i < REF_CONTEXTS; ++i) {
+      vp9_diff_update_prob(r, &fc->single_ref_prob[i][0]);
+      vp9_diff_update_prob(r, &fc->single_ref_prob[i][1]);
+    }
+
+  if (cm->reference_mode != SINGLE_REFERENCE)
+    for (i = 0; i < REF_CONTEXTS; ++i)
+      vp9_diff_update_prob(r, &fc->comp_ref_prob[i]);
+}
+
+static void update_mv_probs(vpx_prob *p, int n, vpx_reader *r) {
+  int i;
+  for (i = 0; i < n; ++i)
+    if (vpx_read(r, MV_UPDATE_PROB))
+      p[i] = (vpx_read_literal(r, 7) << 1) | 1;
+}
+
+static void read_mv_probs(nmv_context *ctx, int allow_hp, vpx_reader *r) {
+  int i, j;
+
+  update_mv_probs(ctx->joints, MV_JOINTS - 1, r);
+
+  for (i = 0; i < 2; ++i) {
+    nmv_component *const comp_ctx = &ctx->comps[i];
+    update_mv_probs(&comp_ctx->sign, 1, r);
+    update_mv_probs(comp_ctx->classes, MV_CLASSES - 1, r);
+    update_mv_probs(comp_ctx->class0, CLASS0_SIZE - 1, r);
+    update_mv_probs(comp_ctx->bits, MV_OFFSET_BITS, r);
+  }
+
+  for (i = 0; i < 2; ++i) {
+    nmv_component *const comp_ctx = &ctx->comps[i];
+    for (j = 0; j < CLASS0_SIZE; ++j)
+      update_mv_probs(comp_ctx->class0_fp[j], MV_FP_SIZE - 1, r);
+    update_mv_probs(comp_ctx->fp, 3, r);
+  }
+
+  if (allow_hp) {
+    for (i = 0; i < 2; ++i) {
+      nmv_component *const comp_ctx = &ctx->comps[i];
+      update_mv_probs(&comp_ctx->class0_hp, 1, r);
+      update_mv_probs(&comp_ctx->hp, 1, r);
+    }
+  }
+}
+
+static void inverse_transform_block_inter(MACROBLOCKD* xd, int plane,
+                                          const TX_SIZE tx_size,
+                                          uint8_t *dst, int stride,
+                                          int eob) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *const dqcoeff = pd->dqcoeff;
+  assert(eob > 0);
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    if (xd->lossless) {
+      vp9_highbd_iwht4x4_add(dqcoeff, dst, stride, eob, xd->bd);
+    } else {
+      switch (tx_size) {
+        case TX_4X4:
+          vp9_highbd_idct4x4_add(dqcoeff, dst, stride, eob, xd->bd);
+          break;
+        case TX_8X8:
+          vp9_highbd_idct8x8_add(dqcoeff, dst, stride, eob, xd->bd);
+          break;
+        case TX_16X16:
+          vp9_highbd_idct16x16_add(dqcoeff, dst, stride, eob, xd->bd);
+          break;
+        case TX_32X32:
+          vp9_highbd_idct32x32_add(dqcoeff, dst, stride, eob, xd->bd);
+          break;
+        default:
+          assert(0 && "Invalid transform size");
+      }
+    }
+  } else {
+    if (xd->lossless) {
+      vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
+    } else {
+      switch (tx_size) {
+        case TX_4X4:
+          vp9_idct4x4_add(dqcoeff, dst, stride, eob);
+          break;
+        case TX_8X8:
+          vp9_idct8x8_add(dqcoeff, dst, stride, eob);
+          break;
+        case TX_16X16:
+          vp9_idct16x16_add(dqcoeff, dst, stride, eob);
+          break;
+        case TX_32X32:
+          vp9_idct32x32_add(dqcoeff, dst, stride, eob);
+          break;
+        default:
+          assert(0 && "Invalid transform size");
+          return;
+      }
+    }
+  }
+#else
+  if (xd->lossless) {
+    vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
+  } else {
+    switch (tx_size) {
+      case TX_4X4:
+        vp9_idct4x4_add(dqcoeff, dst, stride, eob);
+        break;
+      case TX_8X8:
+        vp9_idct8x8_add(dqcoeff, dst, stride, eob);
+        break;
+      case TX_16X16:
+        vp9_idct16x16_add(dqcoeff, dst, stride, eob);
+        break;
+      case TX_32X32:
+        vp9_idct32x32_add(dqcoeff, dst, stride, eob);
+        break;
+      default:
+        assert(0 && "Invalid transform size");
+        return;
+    }
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  if (eob == 1) {
+    dqcoeff[0] = 0;
+  } else {
+    if (tx_size <= TX_16X16 && eob <= 10)
+      memset(dqcoeff, 0, 4 * (4 << tx_size) * sizeof(dqcoeff[0]));
+    else if (tx_size == TX_32X32 && eob <= 34)
+      memset(dqcoeff, 0, 256 * sizeof(dqcoeff[0]));
+    else
+      memset(dqcoeff, 0, (16 << (tx_size << 1)) * sizeof(dqcoeff[0]));
+  }
+}
+
+static void inverse_transform_block_intra(MACROBLOCKD* xd, int plane,
+                                          const TX_TYPE tx_type,
+                                          const TX_SIZE tx_size,
+                                          uint8_t *dst, int stride,
+                                          int eob) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *const dqcoeff = pd->dqcoeff;
+  assert(eob > 0);
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    if (xd->lossless) {
+      vp9_highbd_iwht4x4_add(dqcoeff, dst, stride, eob, xd->bd);
+    } else {
+      switch (tx_size) {
+        case TX_4X4:
+          vp9_highbd_iht4x4_add(tx_type, dqcoeff, dst, stride, eob, xd->bd);
+          break;
+        case TX_8X8:
+          vp9_highbd_iht8x8_add(tx_type, dqcoeff, dst, stride, eob, xd->bd);
+          break;
+        case TX_16X16:
+          vp9_highbd_iht16x16_add(tx_type, dqcoeff, dst, stride, eob, xd->bd);
+          break;
+        case TX_32X32:
+          vp9_highbd_idct32x32_add(dqcoeff, dst, stride, eob, xd->bd);
+          break;
+        default:
+          assert(0 && "Invalid transform size");
+      }
+    }
+  } else {
+    if (xd->lossless) {
+      vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
+    } else {
+      switch (tx_size) {
+        case TX_4X4:
+          vp9_iht4x4_add(tx_type, dqcoeff, dst, stride, eob);
+          break;
+        case TX_8X8:
+          vp9_iht8x8_add(tx_type, dqcoeff, dst, stride, eob);
+          break;
+        case TX_16X16:
+          vp9_iht16x16_add(tx_type, dqcoeff, dst, stride, eob);
+          break;
+        case TX_32X32:
+          vp9_idct32x32_add(dqcoeff, dst, stride, eob);
+          break;
+        default:
+          assert(0 && "Invalid transform size");
+          return;
+      }
+    }
+  }
+#else
+  if (xd->lossless) {
+    vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
+  } else {
+    switch (tx_size) {
+      case TX_4X4:
+        vp9_iht4x4_add(tx_type, dqcoeff, dst, stride, eob);
+        break;
+      case TX_8X8:
+        vp9_iht8x8_add(tx_type, dqcoeff, dst, stride, eob);
+        break;
+      case TX_16X16:
+        vp9_iht16x16_add(tx_type, dqcoeff, dst, stride, eob);
+        break;
+      case TX_32X32:
+        vp9_idct32x32_add(dqcoeff, dst, stride, eob);
+        break;
+      default:
+        assert(0 && "Invalid transform size");
+        return;
+    }
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  if (eob == 1) {
+    dqcoeff[0] = 0;
+  } else {
+    if (tx_type == DCT_DCT && tx_size <= TX_16X16 && eob <= 10)
+      memset(dqcoeff, 0, 4 * (4 << tx_size) * sizeof(dqcoeff[0]));
+    else if (tx_size == TX_32X32 && eob <= 34)
+      memset(dqcoeff, 0, 256 * sizeof(dqcoeff[0]));
+    else
+      memset(dqcoeff, 0, (16 << (tx_size << 1)) * sizeof(dqcoeff[0]));
+  }
+}
+
+static void predict_and_reconstruct_intra_block(MACROBLOCKD *const xd,
+                                                vpx_reader *r,
+                                                MODE_INFO *const mi,
+                                                int plane,
+                                                int row, int col,
+                                                TX_SIZE tx_size) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  PREDICTION_MODE mode = (plane == 0) ? mi->mode : mi->uv_mode;
+  uint8_t *dst;
+  dst = &pd->dst.buf[4 * row * pd->dst.stride + 4 * col];
+
+  if (mi->sb_type < BLOCK_8X8)
+    if (plane == 0)
+      mode = xd->mi[0]->bmi[(row << 1) + col].as_mode;
+
+  vp9_predict_intra_block(xd, pd->n4_wl, tx_size, mode,
+                          dst, pd->dst.stride, dst, pd->dst.stride,
+                          col, row, plane);
+
+  if (!mi->skip) {
+    const TX_TYPE tx_type = (plane || xd->lossless) ?
+        DCT_DCT : intra_mode_to_tx_type_lookup[mode];
+    const scan_order *sc = (plane || xd->lossless) ?
+        &vp9_default_scan_orders[tx_size] : &vp9_scan_orders[tx_size][tx_type];
+    const int eob = vp9_decode_block_tokens(xd, plane, sc, col, row, tx_size,
+                                            r, mi->segment_id);
+    if (eob > 0) {
+      inverse_transform_block_intra(xd, plane, tx_type, tx_size,
+                                    dst, pd->dst.stride, eob);
+    }
+  }
+}
+
+static int reconstruct_inter_block(MACROBLOCKD *const xd, vpx_reader *r,
+                                   MODE_INFO *const mi, int plane,
+                                   int row, int col, TX_SIZE tx_size) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const scan_order *sc = &vp9_default_scan_orders[tx_size];
+  const int eob = vp9_decode_block_tokens(xd, plane, sc, col, row, tx_size, r,
+                                          mi->segment_id);
+
+  if (eob > 0) {
+    inverse_transform_block_inter(
+        xd, plane, tx_size, &pd->dst.buf[4 * row * pd->dst.stride + 4 * col],
+        pd->dst.stride, eob);
+  }
+  return eob;
+}
+
+static void build_mc_border(const uint8_t *src, int src_stride,
+                            uint8_t *dst, int dst_stride,
+                            int x, int y, int b_w, int b_h, int w, int h) {
+  // Get a pointer to the start of the real data for this row.
+  const uint8_t *ref_row = src - x - y * src_stride;
+
+  if (y >= h)
+    ref_row += (h - 1) * src_stride;
+  else if (y > 0)
+    ref_row += y * src_stride;
+
+  do {
+    int right = 0, copy;
+    int left = x < 0 ? -x : 0;
+
+    if (left > b_w)
+      left = b_w;
+
+    if (x + b_w > w)
+      right = x + b_w - w;
+
+    if (right > b_w)
+      right = b_w;
+
+    copy = b_w - left - right;
+
+    if (left)
+      memset(dst, ref_row[0], left);
+
+    if (copy)
+      memcpy(dst + left, ref_row + x + left, copy);
+
+    if (right)
+      memset(dst + left + copy, ref_row[w - 1], right);
+
+    dst += dst_stride;
+    ++y;
+
+    if (y > 0 && y < h)
+      ref_row += src_stride;
+  } while (--b_h);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void high_build_mc_border(const uint8_t *src8, int src_stride,
+                                 uint16_t *dst, int dst_stride,
+                                 int x, int y, int b_w, int b_h,
+                                 int w, int h) {
+  // Get a pointer to the start of the real data for this row.
+  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  const uint16_t *ref_row = src - x - y * src_stride;
+
+  if (y >= h)
+    ref_row += (h - 1) * src_stride;
+  else if (y > 0)
+    ref_row += y * src_stride;
+
+  do {
+    int right = 0, copy;
+    int left = x < 0 ? -x : 0;
+
+    if (left > b_w)
+      left = b_w;
+
+    if (x + b_w > w)
+      right = x + b_w - w;
+
+    if (right > b_w)
+      right = b_w;
+
+    copy = b_w - left - right;
+
+    if (left)
+      vpx_memset16(dst, ref_row[0], left);
+
+    if (copy)
+      memcpy(dst + left, ref_row + x + left, copy * sizeof(uint16_t));
+
+    if (right)
+      vpx_memset16(dst + left + copy, ref_row[w - 1], right);
+
+    dst += dst_stride;
+    ++y;
+
+    if (y > 0 && y < h)
+      ref_row += src_stride;
+  } while (--b_h);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void extend_and_predict(const uint8_t *buf_ptr1, int pre_buf_stride,
+                               int x0, int y0, int b_w, int b_h,
+                               int frame_width, int frame_height,
+                               int border_offset,
+                               uint8_t *const dst, int dst_buf_stride,
+                               int subpel_x, int subpel_y,
+                               const InterpKernel *kernel,
+                               const struct scale_factors *sf,
+                               MACROBLOCKD *xd,
+                               int w, int h, int ref, int xs, int ys) {
+  DECLARE_ALIGNED(16, uint16_t, mc_buf_high[80 * 2 * 80 * 2]);
+  const uint8_t *buf_ptr;
+
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    high_build_mc_border(buf_ptr1, pre_buf_stride, mc_buf_high, b_w,
+                         x0, y0, b_w, b_h, frame_width, frame_height);
+    buf_ptr = CONVERT_TO_BYTEPTR(mc_buf_high) + border_offset;
+  } else {
+    build_mc_border(buf_ptr1, pre_buf_stride, (uint8_t *)mc_buf_high, b_w,
+                    x0, y0, b_w, b_h, frame_width, frame_height);
+    buf_ptr = ((uint8_t *)mc_buf_high) + border_offset;
+  }
+
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    highbd_inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
+                           subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
+  } else {
+    inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
+                    subpel_y, sf, w, h, ref, kernel, xs, ys);
+  }
+}
+#else
+static void extend_and_predict(const uint8_t *buf_ptr1, int pre_buf_stride,
+                               int x0, int y0, int b_w, int b_h,
+                               int frame_width, int frame_height,
+                               int border_offset,
+                               uint8_t *const dst, int dst_buf_stride,
+                               int subpel_x, int subpel_y,
+                               const InterpKernel *kernel,
+                               const struct scale_factors *sf,
+                               int w, int h, int ref, int xs, int ys) {
+  DECLARE_ALIGNED(16, uint8_t, mc_buf[80 * 2 * 80 * 2]);
+  const uint8_t *buf_ptr;
+
+  build_mc_border(buf_ptr1, pre_buf_stride, mc_buf, b_w,
+                  x0, y0, b_w, b_h, frame_width, frame_height);
+  buf_ptr = mc_buf + border_offset;
+
+  inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
+                  subpel_y, sf, w, h, ref, kernel, xs, ys);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void dec_build_inter_predictors(VPxWorker *const worker, MACROBLOCKD *xd,
+                                       int plane, int bw, int bh, int x,
+                                       int y, int w, int h, int mi_x, int mi_y,
+                                       const InterpKernel *kernel,
+                                       const struct scale_factors *sf,
+                                       struct buf_2d *pre_buf,
+                                       struct buf_2d *dst_buf, const MV* mv,
+                                       RefCntBuffer *ref_frame_buf,
+                                       int is_scaled, int ref) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
+  MV32 scaled_mv;
+  int xs, ys, x0, y0, x0_16, y0_16, frame_width, frame_height,
+      buf_stride, subpel_x, subpel_y;
+  uint8_t *ref_frame, *buf_ptr;
+
+  // Get reference frame pointer, width and height.
+  if (plane == 0) {
+    frame_width = ref_frame_buf->buf.y_crop_width;
+    frame_height = ref_frame_buf->buf.y_crop_height;
+    ref_frame = ref_frame_buf->buf.y_buffer;
+  } else {
+    frame_width = ref_frame_buf->buf.uv_crop_width;
+    frame_height = ref_frame_buf->buf.uv_crop_height;
+    ref_frame = plane == 1 ? ref_frame_buf->buf.u_buffer
+                         : ref_frame_buf->buf.v_buffer;
+  }
+
+  if (is_scaled) {
+    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, mv, bw, bh,
+                                               pd->subsampling_x,
+                                               pd->subsampling_y);
+    // Co-ordinate of containing block to pixel precision.
+    int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
+    int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
+#if CONFIG_BETTER_HW_COMPATIBILITY
+    assert(xd->mi[0]->sb_type != BLOCK_4X8 &&
+           xd->mi[0]->sb_type != BLOCK_8X4);
+    assert(mv_q4.row == mv->row * (1 << (1 - pd->subsampling_y)) &&
+           mv_q4.col == mv->col * (1 << (1 - pd->subsampling_x)));
+#endif
+    // Co-ordinate of the block to 1/16th pixel precision.
+    x0_16 = (x_start + x) << SUBPEL_BITS;
+    y0_16 = (y_start + y) << SUBPEL_BITS;
+
+    // Co-ordinate of current block in reference frame
+    // to 1/16th pixel precision.
+    x0_16 = sf->scale_value_x(x0_16, sf);
+    y0_16 = sf->scale_value_y(y0_16, sf);
+
+    // Map the top left corner of the block into the reference frame.
+    x0 = sf->scale_value_x(x_start + x, sf);
+    y0 = sf->scale_value_y(y_start + y, sf);
+
+    // Scale the MV and incorporate the sub-pixel offset of the block
+    // in the reference frame.
+    scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
+    xs = sf->x_step_q4;
+    ys = sf->y_step_q4;
+  } else {
+    // Co-ordinate of containing block to pixel precision.
+    x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
+    y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
+
+    // Co-ordinate of the block to 1/16th pixel precision.
+    x0_16 = x0 << SUBPEL_BITS;
+    y0_16 = y0 << SUBPEL_BITS;
+
+    scaled_mv.row = mv->row * (1 << (1 - pd->subsampling_y));
+    scaled_mv.col = mv->col * (1 << (1 - pd->subsampling_x));
+    xs = ys = 16;
+  }
+  subpel_x = scaled_mv.col & SUBPEL_MASK;
+  subpel_y = scaled_mv.row & SUBPEL_MASK;
+
+  // Calculate the top left corner of the best matching block in the
+  // reference frame.
+  x0 += scaled_mv.col >> SUBPEL_BITS;
+  y0 += scaled_mv.row >> SUBPEL_BITS;
+  x0_16 += scaled_mv.col;
+  y0_16 += scaled_mv.row;
+
+  // Get reference block pointer.
+  buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
+  buf_stride = pre_buf->stride;
+
+  // Do border extension if there is motion or the
+  // width/height is not a multiple of 8 pixels.
+  if (is_scaled || scaled_mv.col || scaled_mv.row ||
+      (frame_width & 0x7) || (frame_height & 0x7)) {
+    int y1 = ((y0_16 + (h - 1) * ys) >> SUBPEL_BITS) + 1;
+
+    // Get reference block bottom right horizontal coordinate.
+    int x1 = ((x0_16 + (w - 1) * xs) >> SUBPEL_BITS) + 1;
+    int x_pad = 0, y_pad = 0;
+
+    if (subpel_x || (sf->x_step_q4 != SUBPEL_SHIFTS)) {
+      x0 -= VP9_INTERP_EXTEND - 1;
+      x1 += VP9_INTERP_EXTEND;
+      x_pad = 1;
+    }
+
+    if (subpel_y || (sf->y_step_q4 != SUBPEL_SHIFTS)) {
+      y0 -= VP9_INTERP_EXTEND - 1;
+      y1 += VP9_INTERP_EXTEND;
+      y_pad = 1;
+    }
+
+    // Wait until reference block is ready. Pad 7 more pixels as last 7
+    // pixels of each superblock row can be changed by next superblock row.
+    if (worker != NULL)
+      vp9_frameworker_wait(worker, ref_frame_buf,
+                           VPXMAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
+
+    // Skip border extension if block is inside the frame.
+    if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width - 1 ||
+        y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
+      // Extend the border.
+      const uint8_t *const buf_ptr1 = ref_frame + y0 * buf_stride + x0;
+      const int b_w = x1 - x0 + 1;
+      const int b_h = y1 - y0 + 1;
+      const int border_offset = y_pad * 3 * b_w + x_pad * 3;
+
+      extend_and_predict(buf_ptr1, buf_stride, x0, y0, b_w, b_h,
+                         frame_width, frame_height, border_offset,
+                         dst, dst_buf->stride,
+                         subpel_x, subpel_y,
+                         kernel, sf,
+#if CONFIG_VP9_HIGHBITDEPTH
+                         xd,
+#endif
+                         w, h, ref, xs, ys);
+      return;
+    }
+  } else {
+    // Wait until reference block is ready. Pad 7 more pixels as last 7
+    // pixels of each superblock row can be changed by next superblock row.
+    if (worker != NULL) {
+      const int y1 = (y0_16 + (h - 1) * ys) >> SUBPEL_BITS;
+      vp9_frameworker_wait(worker, ref_frame_buf,
+                           VPXMAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
+    }
+  }
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    highbd_inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                           subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
+  } else {
+    inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                    subpel_y, sf, w, h, ref, kernel, xs, ys);
+  }
+#else
+  inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                  subpel_y, sf, w, h, ref, kernel, xs, ys);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+static void dec_build_inter_predictors_sb(VP9Decoder *const pbi,
+                                          MACROBLOCKD *xd,
+                                          int mi_row, int mi_col) {
+  int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
+  const MODE_INFO *mi = xd->mi[0];
+  const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
+  const BLOCK_SIZE sb_type = mi->sb_type;
+  const int is_compound = has_second_ref(mi);
+  int ref;
+  int is_scaled;
+  VPxWorker *const fwo = pbi->frame_parallel_decode ?
+      pbi->frame_worker_owner : NULL;
+
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    const MV_REFERENCE_FRAME frame = mi->ref_frame[ref];
+    RefBuffer *ref_buf = &pbi->common.frame_refs[frame - LAST_FRAME];
+    const struct scale_factors *const sf = &ref_buf->sf;
+    const int idx = ref_buf->idx;
+    BufferPool *const pool = pbi->common.buffer_pool;
+    RefCntBuffer *const ref_frame_buf = &pool->frame_bufs[idx];
+
+    if (!vp9_is_valid_scale(sf))
+      vpx_internal_error(xd->error_info, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Reference frame has invalid dimensions");
+
+    is_scaled = vp9_is_scaled(sf);
+    vp9_setup_pre_planes(xd, ref, ref_buf->buf, mi_row, mi_col,
+                         is_scaled ? sf : NULL);
+    xd->block_refs[ref] = ref_buf;
+
+    if (sb_type < BLOCK_8X8) {
+      for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+        struct macroblockd_plane *const pd = &xd->plane[plane];
+        struct buf_2d *const dst_buf = &pd->dst;
+        const int num_4x4_w = pd->n4_w;
+        const int num_4x4_h = pd->n4_h;
+        const int n4w_x4 = 4 * num_4x4_w;
+        const int n4h_x4 = 4 * num_4x4_h;
+        struct buf_2d *const pre_buf = &pd->pre[ref];
+        int i = 0, x, y;
+        for (y = 0; y < num_4x4_h; ++y) {
+          for (x = 0; x < num_4x4_w; ++x) {
+            const MV mv = average_split_mvs(pd, mi, ref, i++);
+            dec_build_inter_predictors(fwo, xd, plane, n4w_x4, n4h_x4,
+                                       4 * x, 4 * y, 4, 4, mi_x, mi_y, kernel,
+                                       sf, pre_buf, dst_buf, &mv,
+                                       ref_frame_buf, is_scaled, ref);
+          }
+        }
+      }
+    } else {
+      const MV mv = mi->mv[ref].as_mv;
+      for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+        struct macroblockd_plane *const pd = &xd->plane[plane];
+        struct buf_2d *const dst_buf = &pd->dst;
+        const int num_4x4_w = pd->n4_w;
+        const int num_4x4_h = pd->n4_h;
+        const int n4w_x4 = 4 * num_4x4_w;
+        const int n4h_x4 = 4 * num_4x4_h;
+        struct buf_2d *const pre_buf = &pd->pre[ref];
+        dec_build_inter_predictors(fwo, xd, plane, n4w_x4, n4h_x4,
+                                   0, 0, n4w_x4, n4h_x4, mi_x, mi_y, kernel,
+                                   sf, pre_buf, dst_buf, &mv,
+                                   ref_frame_buf, is_scaled, ref);
+      }
+    }
+  }
+}
+
+static INLINE TX_SIZE dec_get_uv_tx_size(const MODE_INFO *mi,
+                                         int n4_wl, int n4_hl) {
+  // get minimum log2 num4x4s dimension
+  const int x = VPXMIN(n4_wl, n4_hl);
+  return VPXMIN(mi->tx_size,  x);
+}
+
+static INLINE void dec_reset_skip_context(MACROBLOCKD *xd) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    struct macroblockd_plane *const pd = &xd->plane[i];
+    memset(pd->above_context, 0, sizeof(ENTROPY_CONTEXT) * pd->n4_w);
+    memset(pd->left_context, 0, sizeof(ENTROPY_CONTEXT) * pd->n4_h);
+  }
+}
+
+static void set_plane_n4(MACROBLOCKD *const xd, int bw, int bh, int bwl,
+                         int bhl) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    xd->plane[i].n4_w = (bw << 1) >> xd->plane[i].subsampling_x;
+    xd->plane[i].n4_h = (bh << 1) >> xd->plane[i].subsampling_y;
+    xd->plane[i].n4_wl = bwl - xd->plane[i].subsampling_x;
+    xd->plane[i].n4_hl = bhl - xd->plane[i].subsampling_y;
+  }
+}
+
+static MODE_INFO *set_offsets(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                              BLOCK_SIZE bsize, int mi_row, int mi_col,
+                              int bw, int bh, int x_mis, int y_mis,
+                              int bwl, int bhl) {
+  const int offset = mi_row * cm->mi_stride + mi_col;
+  int x, y;
+  const TileInfo *const tile = &xd->tile;
+
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = &cm->mi[offset];
+  // TODO(slavarnway): Generate sb_type based on bwl and bhl, instead of
+  // passing bsize from decode_partition().
+  xd->mi[0]->sb_type = bsize;
+  for (y = 0; y < y_mis; ++y)
+    for (x = !y; x < x_mis; ++x) {
+      xd->mi[y * cm->mi_stride + x] = xd->mi[0];
+    }
+
+  set_plane_n4(xd, bw, bh, bwl, bhl);
+
+  set_skip_context(xd, mi_row, mi_col);
+
+  // Distance of Mb to the various image edges. These are specified to 8th pel
+  // as they are always compared to values that are in 1/8th pel units
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
+
+  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
+  return xd->mi[0];
+}
+
+static void decode_block(VP9Decoder *const pbi, MACROBLOCKD *const xd,
+                         int mi_row, int mi_col,
+                         vpx_reader *r, BLOCK_SIZE bsize,
+                         int bwl, int bhl) {
+  VP9_COMMON *const cm = &pbi->common;
+  const int less8x8 = bsize < BLOCK_8X8;
+  const int bw = 1 << (bwl - 1);
+  const int bh = 1 << (bhl - 1);
+  const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
+
+  MODE_INFO *mi = set_offsets(cm, xd, bsize, mi_row, mi_col,
+                              bw, bh, x_mis, y_mis, bwl, bhl);
+
+  if (bsize >= BLOCK_8X8 && (cm->subsampling_x || cm->subsampling_y)) {
+    const BLOCK_SIZE uv_subsize =
+        ss_size_lookup[bsize][cm->subsampling_x][cm->subsampling_y];
+    if (uv_subsize == BLOCK_INVALID)
+      vpx_internal_error(xd->error_info,
+                         VPX_CODEC_CORRUPT_FRAME, "Invalid block size.");
+  }
+
+  vp9_read_mode_info(pbi, xd, mi_row, mi_col, r, x_mis, y_mis);
+
+  if (mi->skip) {
+    dec_reset_skip_context(xd);
+  }
+
+  if (!is_inter_block(mi)) {
+    int plane;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+      const struct macroblockd_plane *const pd = &xd->plane[plane];
+      const TX_SIZE tx_size =
+          plane ? dec_get_uv_tx_size(mi, pd->n4_wl, pd->n4_hl)
+                  : mi->tx_size;
+      const int num_4x4_w = pd->n4_w;
+      const int num_4x4_h = pd->n4_h;
+      const int step = (1 << tx_size);
+      int row, col;
+      const int max_blocks_wide = num_4x4_w + (xd->mb_to_right_edge >= 0 ?
+          0 : xd->mb_to_right_edge >> (5 + pd->subsampling_x));
+      const int max_blocks_high = num_4x4_h + (xd->mb_to_bottom_edge >= 0 ?
+          0 : xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
+
+      xd->max_blocks_wide = xd->mb_to_right_edge >= 0 ? 0 : max_blocks_wide;
+      xd->max_blocks_high = xd->mb_to_bottom_edge >= 0 ? 0 : max_blocks_high;
+
+      for (row = 0; row < max_blocks_high; row += step)
+        for (col = 0; col < max_blocks_wide; col += step)
+          predict_and_reconstruct_intra_block(xd, r, mi, plane,
+                                              row, col, tx_size);
+    }
+  } else {
+    // Prediction
+    dec_build_inter_predictors_sb(pbi, xd, mi_row, mi_col);
+
+    // Reconstruction
+    if (!mi->skip) {
+      int eobtotal = 0;
+      int plane;
+
+      for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+        const struct macroblockd_plane *const pd = &xd->plane[plane];
+        const TX_SIZE tx_size =
+            plane ? dec_get_uv_tx_size(mi, pd->n4_wl, pd->n4_hl)
+                    : mi->tx_size;
+        const int num_4x4_w = pd->n4_w;
+        const int num_4x4_h = pd->n4_h;
+        const int step = (1 << tx_size);
+        int row, col;
+        const int max_blocks_wide = num_4x4_w + (xd->mb_to_right_edge >= 0 ?
+            0 : xd->mb_to_right_edge >> (5 + pd->subsampling_x));
+        const int max_blocks_high = num_4x4_h + (xd->mb_to_bottom_edge >= 0 ?
+            0 : xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
+
+        xd->max_blocks_wide = xd->mb_to_right_edge >= 0 ? 0 : max_blocks_wide;
+        xd->max_blocks_high = xd->mb_to_bottom_edge >= 0 ? 0 : max_blocks_high;
+
+        for (row = 0; row < max_blocks_high; row += step)
+          for (col = 0; col < max_blocks_wide; col += step)
+            eobtotal += reconstruct_inter_block(xd, r, mi, plane, row, col,
+                                                tx_size);
+      }
+
+      if (!less8x8 && eobtotal == 0)
+        mi->skip = 1;  // skip loopfilter
+    }
+  }
+
+  xd->corrupted |= vpx_reader_has_error(r);
+
+  if (cm->lf.filter_level) {
+    vp9_build_mask(cm, mi, mi_row, mi_col, bw, bh);
+  }
+}
+
+static INLINE int dec_partition_plane_context(const MACROBLOCKD *xd,
+                                              int mi_row, int mi_col,
+                                              int bsl) {
+  const PARTITION_CONTEXT *above_ctx = xd->above_seg_context + mi_col;
+  const PARTITION_CONTEXT *left_ctx = xd->left_seg_context + (mi_row & MI_MASK);
+  int above = (*above_ctx >> bsl) & 1 , left = (*left_ctx >> bsl) & 1;
+
+//  assert(bsl >= 0);
+
+  return (left * 2 + above) + bsl * PARTITION_PLOFFSET;
+}
+
+static INLINE void dec_update_partition_context(MACROBLOCKD *xd,
+                                                int mi_row, int mi_col,
+                                                BLOCK_SIZE subsize,
+                                                int bw) {
+  PARTITION_CONTEXT *const above_ctx = xd->above_seg_context + mi_col;
+  PARTITION_CONTEXT *const left_ctx = xd->left_seg_context + (mi_row & MI_MASK);
+
+  // update the partition context at the end notes. set partition bits
+  // of block sizes larger than the current one to be one, and partition
+  // bits of smaller block sizes to be zero.
+  memset(above_ctx, partition_context_lookup[subsize].above, bw);
+  memset(left_ctx, partition_context_lookup[subsize].left, bw);
+}
+
+static PARTITION_TYPE read_partition(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                     vpx_reader *r,
+                                     int has_rows, int has_cols, int bsl) {
+  const int ctx = dec_partition_plane_context(xd, mi_row, mi_col, bsl);
+  const vpx_prob *const probs = get_partition_probs(xd, ctx);
+  FRAME_COUNTS *counts = xd->counts;
+  PARTITION_TYPE p;
+
+  if (has_rows && has_cols)
+    p = (PARTITION_TYPE)vpx_read_tree(r, vp9_partition_tree, probs);
+  else if (!has_rows && has_cols)
+    p = vpx_read(r, probs[1]) ? PARTITION_SPLIT : PARTITION_HORZ;
+  else if (has_rows && !has_cols)
+    p = vpx_read(r, probs[2]) ? PARTITION_SPLIT : PARTITION_VERT;
+  else
+    p = PARTITION_SPLIT;
+
+  if (counts)
+    ++counts->partition[ctx][p];
+
+  return p;
+}
+
+// TODO(slavarnway): eliminate bsize and subsize in future commits
+static void decode_partition(VP9Decoder *const pbi, MACROBLOCKD *const xd,
+                             int mi_row, int mi_col,
+                             vpx_reader* r, BLOCK_SIZE bsize, int n4x4_l2) {
+  VP9_COMMON *const cm = &pbi->common;
+  const int n8x8_l2 = n4x4_l2 - 1;
+  const int num_8x8_wh = 1 << n8x8_l2;
+  const int hbs = num_8x8_wh >> 1;
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+  const int has_rows = (mi_row + hbs) < cm->mi_rows;
+  const int has_cols = (mi_col + hbs) < cm->mi_cols;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  partition = read_partition(xd, mi_row, mi_col, r, has_rows, has_cols,
+                             n8x8_l2);
+  subsize = subsize_lookup[partition][bsize];  // get_subsize(bsize, partition);
+  if (!hbs) {
+    // calculate bmode block dimensions (log 2)
+    xd->bmode_blocks_wl = 1 >> !!(partition & PARTITION_VERT);
+    xd->bmode_blocks_hl = 1 >> !!(partition & PARTITION_HORZ);
+    decode_block(pbi, xd, mi_row, mi_col, r, subsize, 1, 1);
+  } else {
+    switch (partition) {
+      case PARTITION_NONE:
+        decode_block(pbi, xd, mi_row, mi_col, r, subsize, n4x4_l2, n4x4_l2);
+        break;
+      case PARTITION_HORZ:
+        decode_block(pbi, xd, mi_row, mi_col, r, subsize, n4x4_l2, n8x8_l2);
+        if (has_rows)
+          decode_block(pbi, xd, mi_row + hbs, mi_col, r, subsize, n4x4_l2,
+                       n8x8_l2);
+        break;
+      case PARTITION_VERT:
+        decode_block(pbi, xd, mi_row, mi_col, r, subsize, n8x8_l2, n4x4_l2);
+        if (has_cols)
+          decode_block(pbi, xd, mi_row, mi_col + hbs, r, subsize, n8x8_l2,
+                       n4x4_l2);
+        break;
+      case PARTITION_SPLIT:
+        decode_partition(pbi, xd, mi_row, mi_col, r, subsize, n8x8_l2);
+        decode_partition(pbi, xd, mi_row, mi_col + hbs, r, subsize, n8x8_l2);
+        decode_partition(pbi, xd, mi_row + hbs, mi_col, r, subsize, n8x8_l2);
+        decode_partition(pbi, xd, mi_row + hbs, mi_col + hbs, r, subsize,
+                         n8x8_l2);
+        break;
+      default:
+        assert(0 && "Invalid partition type");
+    }
+  }
+
+  // update partition context
+  if (bsize >= BLOCK_8X8 &&
+      (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
+    dec_update_partition_context(xd, mi_row, mi_col, subsize, num_8x8_wh);
+}
+
+static void setup_token_decoder(const uint8_t *data,
+                                const uint8_t *data_end,
+                                size_t read_size,
+                                struct vpx_internal_error_info *error_info,
+                                vpx_reader *r,
+                                vpx_decrypt_cb decrypt_cb,
+                                void *decrypt_state) {
+  // Validate the calculated partition length. If the buffer
+  // described by the partition can't be fully read, then restrict
+  // it to the portion that can be (for EC mode) or throw an error.
+  if (!read_is_valid(data, read_size, data_end))
+    vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt tile length");
+
+  if (vpx_reader_init(r, data, read_size, decrypt_cb, decrypt_state))
+    vpx_internal_error(error_info, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate bool decoder %d", 1);
+}
+
+static void read_coef_probs_common(vp9_coeff_probs_model *coef_probs,
+                                   vpx_reader *r) {
+  int i, j, k, l, m;
+
+  if (vpx_read_bit(r))
+    for (i = 0; i < PLANE_TYPES; ++i)
+      for (j = 0; j < REF_TYPES; ++j)
+        for (k = 0; k < COEF_BANDS; ++k)
+          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
+            for (m = 0; m < UNCONSTRAINED_NODES; ++m)
+              vp9_diff_update_prob(r, &coef_probs[i][j][k][l][m]);
+}
+
+static void read_coef_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode,
+                            vpx_reader *r) {
+    const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+    TX_SIZE tx_size;
+    for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
+      read_coef_probs_common(fc->coef_probs[tx_size], r);
+}
+
+static void setup_segmentation(struct segmentation *seg,
+                               struct vpx_read_bit_buffer *rb) {
+  int i, j;
+
+  seg->update_map = 0;
+  seg->update_data = 0;
+
+  seg->enabled = vpx_rb_read_bit(rb);
+  if (!seg->enabled)
+    return;
+
+  // Segmentation map update
+  seg->update_map = vpx_rb_read_bit(rb);
+  if (seg->update_map) {
+    for (i = 0; i < SEG_TREE_PROBS; i++)
+      seg->tree_probs[i] = vpx_rb_read_bit(rb) ? vpx_rb_read_literal(rb, 8)
+                                               : MAX_PROB;
+
+    seg->temporal_update = vpx_rb_read_bit(rb);
+    if (seg->temporal_update) {
+      for (i = 0; i < PREDICTION_PROBS; i++)
+        seg->pred_probs[i] = vpx_rb_read_bit(rb) ? vpx_rb_read_literal(rb, 8)
+                                                 : MAX_PROB;
+    } else {
+      for (i = 0; i < PREDICTION_PROBS; i++)
+        seg->pred_probs[i] = MAX_PROB;
+    }
+  }
+
+  // Segmentation data update
+  seg->update_data = vpx_rb_read_bit(rb);
+  if (seg->update_data) {
+    seg->abs_delta = vpx_rb_read_bit(rb);
+
+    vp9_clearall_segfeatures(seg);
+
+    for (i = 0; i < MAX_SEGMENTS; i++) {
+      for (j = 0; j < SEG_LVL_MAX; j++) {
+        int data = 0;
+        const int feature_enabled = vpx_rb_read_bit(rb);
+        if (feature_enabled) {
+          vp9_enable_segfeature(seg, i, j);
+          data = decode_unsigned_max(rb, vp9_seg_feature_data_max(j));
+          if (vp9_is_segfeature_signed(j))
+            data = vpx_rb_read_bit(rb) ? -data : data;
+        }
+        vp9_set_segdata(seg, i, j, data);
+      }
+    }
+  }
+}
+
+static void setup_loopfilter(struct loopfilter *lf,
+                             struct vpx_read_bit_buffer *rb) {
+  lf->filter_level = vpx_rb_read_literal(rb, 6);
+  lf->sharpness_level = vpx_rb_read_literal(rb, 3);
+
+  // Read in loop filter deltas applied at the MB level based on mode or ref
+  // frame.
+  lf->mode_ref_delta_update = 0;
+
+  lf->mode_ref_delta_enabled = vpx_rb_read_bit(rb);
+  if (lf->mode_ref_delta_enabled) {
+    lf->mode_ref_delta_update = vpx_rb_read_bit(rb);
+    if (lf->mode_ref_delta_update) {
+      int i;
+
+      for (i = 0; i < MAX_REF_LF_DELTAS; i++)
+        if (vpx_rb_read_bit(rb))
+          lf->ref_deltas[i] = vpx_rb_read_signed_literal(rb, 6);
+
+      for (i = 0; i < MAX_MODE_LF_DELTAS; i++)
+        if (vpx_rb_read_bit(rb))
+          lf->mode_deltas[i] = vpx_rb_read_signed_literal(rb, 6);
+    }
+  }
+}
+
+static INLINE int read_delta_q(struct vpx_read_bit_buffer *rb) {
+  return vpx_rb_read_bit(rb) ? vpx_rb_read_signed_literal(rb, 4) : 0;
+}
+
+static void setup_quantization(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                               struct vpx_read_bit_buffer *rb) {
+  cm->base_qindex = vpx_rb_read_literal(rb, QINDEX_BITS);
+  cm->y_dc_delta_q = read_delta_q(rb);
+  cm->uv_dc_delta_q = read_delta_q(rb);
+  cm->uv_ac_delta_q = read_delta_q(rb);
+  cm->dequant_bit_depth = cm->bit_depth;
+  xd->lossless = cm->base_qindex == 0 &&
+                 cm->y_dc_delta_q == 0 &&
+                 cm->uv_dc_delta_q == 0 &&
+                 cm->uv_ac_delta_q == 0;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  xd->bd = (int)cm->bit_depth;
+#endif
+}
+
+static void setup_segmentation_dequant(VP9_COMMON *const cm) {
+  // Build y/uv dequant values based on segmentation.
+  if (cm->seg.enabled) {
+    int i;
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      const int qindex = vp9_get_qindex(&cm->seg, i, cm->base_qindex);
+      cm->y_dequant[i][0] = vp9_dc_quant(qindex, cm->y_dc_delta_q,
+                                         cm->bit_depth);
+      cm->y_dequant[i][1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
+      cm->uv_dequant[i][0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q,
+                                          cm->bit_depth);
+      cm->uv_dequant[i][1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q,
+                                          cm->bit_depth);
+    }
+  } else {
+    const int qindex = cm->base_qindex;
+    // When segmentation is disabled, only the first value is used.  The
+    // remaining are don't cares.
+    cm->y_dequant[0][0] = vp9_dc_quant(qindex, cm->y_dc_delta_q, cm->bit_depth);
+    cm->y_dequant[0][1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
+    cm->uv_dequant[0][0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q,
+                                        cm->bit_depth);
+    cm->uv_dequant[0][1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q,
+                                        cm->bit_depth);
+  }
+}
+
+static INTERP_FILTER read_interp_filter(struct vpx_read_bit_buffer *rb) {
+  const INTERP_FILTER literal_to_filter[] = { EIGHTTAP_SMOOTH,
+                                              EIGHTTAP,
+                                              EIGHTTAP_SHARP,
+                                              BILINEAR };
+  return vpx_rb_read_bit(rb) ? SWITCHABLE
+                             : literal_to_filter[vpx_rb_read_literal(rb, 2)];
+}
+
+static void setup_render_size(VP9_COMMON *cm, struct vpx_read_bit_buffer *rb) {
+  cm->render_width = cm->width;
+  cm->render_height = cm->height;
+  if (vpx_rb_read_bit(rb))
+    vp9_read_frame_size(rb, &cm->render_width, &cm->render_height);
+}
+
+static void resize_mv_buffer(VP9_COMMON *cm) {
+  vpx_free(cm->cur_frame->mvs);
+  cm->cur_frame->mi_rows = cm->mi_rows;
+  cm->cur_frame->mi_cols = cm->mi_cols;
+  CHECK_MEM_ERROR(cm, cm->cur_frame->mvs,
+                  (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
+                                       sizeof(*cm->cur_frame->mvs)));
+}
+
+static void resize_context_buffers(VP9_COMMON *cm, int width, int height) {
+#if CONFIG_SIZE_LIMIT
+  if (width > DECODE_WIDTH_LIMIT || height > DECODE_HEIGHT_LIMIT)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Dimensions of %dx%d beyond allowed size of %dx%d.",
+                       width, height, DECODE_WIDTH_LIMIT, DECODE_HEIGHT_LIMIT);
+#endif
+  if (cm->width != width || cm->height != height) {
+    const int new_mi_rows =
+        ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2) >> MI_SIZE_LOG2;
+    const int new_mi_cols =
+        ALIGN_POWER_OF_TWO(width,  MI_SIZE_LOG2) >> MI_SIZE_LOG2;
+
+    // Allocations in vp9_alloc_context_buffers() depend on individual
+    // dimensions as well as the overall size.
+    if (new_mi_cols > cm->mi_cols || new_mi_rows > cm->mi_rows) {
+      if (vp9_alloc_context_buffers(cm, width, height))
+        vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate context buffers");
+    } else {
+      vp9_set_mb_mi(cm, width, height);
+    }
+    vp9_init_context_buffers(cm);
+    cm->width = width;
+    cm->height = height;
+  }
+  if (cm->cur_frame->mvs == NULL || cm->mi_rows > cm->cur_frame->mi_rows ||
+      cm->mi_cols > cm->cur_frame->mi_cols) {
+    resize_mv_buffer(cm);
+  }
+}
+
+static void setup_frame_size(VP9_COMMON *cm, struct vpx_read_bit_buffer *rb) {
+  int width, height;
+  BufferPool *const pool = cm->buffer_pool;
+  vp9_read_frame_size(rb, &width, &height);
+  resize_context_buffers(cm, width, height);
+  setup_render_size(cm, rb);
+
+  lock_buffer_pool(pool);
+  if (vpx_realloc_frame_buffer(
+          get_frame_new_buffer(cm), cm->width, cm->height,
+          cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+          cm->use_highbitdepth,
+#endif
+          VP9_DEC_BORDER_IN_PIXELS,
+          cm->byte_alignment,
+          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
+          pool->cb_priv)) {
+    unlock_buffer_pool(pool);
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate frame buffer");
+  }
+  unlock_buffer_pool(pool);
+
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
+  pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_space = cm->color_space;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_range = cm->color_range;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_width  = cm->render_width;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_height = cm->render_height;
+}
+
+static INLINE int valid_ref_frame_img_fmt(vpx_bit_depth_t ref_bit_depth,
+                                          int ref_xss, int ref_yss,
+                                          vpx_bit_depth_t this_bit_depth,
+                                          int this_xss, int this_yss) {
+  return ref_bit_depth == this_bit_depth && ref_xss == this_xss &&
+         ref_yss == this_yss;
+}
+
+static void setup_frame_size_with_refs(VP9_COMMON *cm,
+                                       struct vpx_read_bit_buffer *rb) {
+  int width, height;
+  int found = 0, i;
+  int has_valid_ref_frame = 0;
+  BufferPool *const pool = cm->buffer_pool;
+  for (i = 0; i < REFS_PER_FRAME; ++i) {
+    if (vpx_rb_read_bit(rb)) {
+      if (cm->frame_refs[i].idx != INVALID_IDX) {
+        YV12_BUFFER_CONFIG *const buf = cm->frame_refs[i].buf;
+        width = buf->y_crop_width;
+        height = buf->y_crop_height;
+        found = 1;
+        break;
+      } else {
+        vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                           "Failed to decode frame size");
+      }
+    }
+  }
+
+  if (!found)
+    vp9_read_frame_size(rb, &width, &height);
+
+  if (width <= 0 || height <= 0)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid frame size");
+
+  // Check to make sure at least one of frames that this frame references
+  // has valid dimensions.
+  for (i = 0; i < REFS_PER_FRAME; ++i) {
+    RefBuffer *const ref_frame = &cm->frame_refs[i];
+    has_valid_ref_frame |= (ref_frame->idx != INVALID_IDX &&
+                            valid_ref_frame_size(ref_frame->buf->y_crop_width,
+                                                 ref_frame->buf->y_crop_height,
+                                                 width, height));
+  }
+  if (!has_valid_ref_frame)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Referenced frame has invalid size");
+  for (i = 0; i < REFS_PER_FRAME; ++i) {
+    RefBuffer *const ref_frame = &cm->frame_refs[i];
+    if (ref_frame->idx == INVALID_IDX ||
+        !valid_ref_frame_img_fmt(ref_frame->buf->bit_depth,
+                                 ref_frame->buf->subsampling_x,
+                                 ref_frame->buf->subsampling_y,
+                                 cm->bit_depth,
+                                 cm->subsampling_x,
+                                 cm->subsampling_y))
+      vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                         "Referenced frame has incompatible color format");
+  }
+
+  resize_context_buffers(cm, width, height);
+  setup_render_size(cm, rb);
+
+  lock_buffer_pool(pool);
+  if (vpx_realloc_frame_buffer(
+          get_frame_new_buffer(cm), cm->width, cm->height,
+          cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+          cm->use_highbitdepth,
+#endif
+          VP9_DEC_BORDER_IN_PIXELS,
+          cm->byte_alignment,
+          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
+          pool->cb_priv)) {
+    unlock_buffer_pool(pool);
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate frame buffer");
+  }
+  unlock_buffer_pool(pool);
+
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
+  pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_space = cm->color_space;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_range = cm->color_range;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_width  = cm->render_width;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_height = cm->render_height;
+}
+
+static void setup_tile_info(VP9_COMMON *cm, struct vpx_read_bit_buffer *rb) {
+  int min_log2_tile_cols, max_log2_tile_cols, max_ones;
+  vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
+
+  // columns
+  max_ones = max_log2_tile_cols - min_log2_tile_cols;
+  cm->log2_tile_cols = min_log2_tile_cols;
+  while (max_ones-- && vpx_rb_read_bit(rb))
+    cm->log2_tile_cols++;
+
+  if (cm->log2_tile_cols > 6)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid number of tile columns");
+
+  // rows
+  cm->log2_tile_rows = vpx_rb_read_bit(rb);
+  if (cm->log2_tile_rows)
+    cm->log2_tile_rows += vpx_rb_read_bit(rb);
+}
+
+// Reads the next tile returning its size and adjusting '*data' accordingly
+// based on 'is_last'.
+static void get_tile_buffer(const uint8_t *const data_end,
+                            int is_last,
+                            struct vpx_internal_error_info *error_info,
+                            const uint8_t **data,
+                            vpx_decrypt_cb decrypt_cb, void *decrypt_state,
+                            TileBuffer *buf) {
+  size_t size;
+
+  if (!is_last) {
+    if (!read_is_valid(*data, 4, data_end))
+      vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
+                         "Truncated packet or corrupt tile length");
+
+    if (decrypt_cb) {
+      uint8_t be_data[4];
+      decrypt_cb(decrypt_state, *data, be_data, 4);
+      size = mem_get_be32(be_data);
+    } else {
+      size = mem_get_be32(*data);
+    }
+    *data += 4;
+
+    if (size > (size_t)(data_end - *data))
+      vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
+                         "Truncated packet or corrupt tile size");
+  } else {
+    size = data_end - *data;
+  }
+
+  buf->data = *data;
+  buf->size = size;
+
+  *data += size;
+}
+
+static void get_tile_buffers(VP9Decoder *pbi,
+                             const uint8_t *data, const uint8_t *data_end,
+                             int tile_cols, int tile_rows,
+                             TileBuffer (*tile_buffers)[1 << 6]) {
+  int r, c;
+
+  for (r = 0; r < tile_rows; ++r) {
+    for (c = 0; c < tile_cols; ++c) {
+      const int is_last = (r == tile_rows - 1) && (c == tile_cols - 1);
+      TileBuffer *const buf = &tile_buffers[r][c];
+      buf->col = c;
+      get_tile_buffer(data_end, is_last, &pbi->common.error, &data,
+                      pbi->decrypt_cb, pbi->decrypt_state, buf);
+    }
+  }
+}
+
+static const uint8_t *decode_tiles(VP9Decoder *pbi,
+                                   const uint8_t *data,
+                                   const uint8_t *data_end) {
+  VP9_COMMON *const cm = &pbi->common;
+  const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
+  const int aligned_cols = mi_cols_aligned_to_sb(cm->mi_cols);
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  TileBuffer tile_buffers[4][1 << 6];
+  int tile_row, tile_col;
+  int mi_row, mi_col;
+  TileWorkerData *tile_data = NULL;
+
+  if (cm->lf.filter_level && !cm->skip_loop_filter &&
+      pbi->lf_worker.data1 == NULL) {
+    CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
+                    vpx_memalign(32, sizeof(LFWorkerData)));
+    pbi->lf_worker.hook = (VPxWorkerHook)vp9_loop_filter_worker;
+    if (pbi->max_threads > 1 && !winterface->reset(&pbi->lf_worker)) {
+      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                         "Loop filter thread creation failed");
+    }
+  }
+
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+    // Be sure to sync as we might be resuming after a failed frame decode.
+    winterface->sync(&pbi->lf_worker);
+    vp9_loop_filter_data_reset(lf_data, get_frame_new_buffer(cm), cm,
+                               pbi->mb.plane);
+  }
+
+  assert(tile_rows <= 4);
+  assert(tile_cols <= (1 << 6));
+
+  // Note: this memset assumes above_context[0], [1] and [2]
+  // are allocated as part of the same buffer.
+  memset(cm->above_context, 0,
+         sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_cols);
+
+  memset(cm->above_seg_context, 0,
+         sizeof(*cm->above_seg_context) * aligned_cols);
+
+  vp9_reset_lfm(cm);
+
+  get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
+
+  // Load all tile information into tile_data.
+  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
+    for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+      const TileBuffer *const buf = &tile_buffers[tile_row][tile_col];
+      tile_data = pbi->tile_worker_data + tile_cols * tile_row + tile_col;
+      tile_data->xd = pbi->mb;
+      tile_data->xd.corrupted = 0;
+      tile_data->xd.counts =
+          cm->frame_parallel_decoding_mode ? NULL : &cm->counts;
+      vp9_zero(tile_data->dqcoeff);
+      vp9_tile_init(&tile_data->xd.tile, cm, tile_row, tile_col);
+      setup_token_decoder(buf->data, data_end, buf->size, &cm->error,
+                          &tile_data->bit_reader, pbi->decrypt_cb,
+                          pbi->decrypt_state);
+      vp9_init_macroblockd(cm, &tile_data->xd, tile_data->dqcoeff);
+    }
+  }
+
+  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
+    TileInfo tile;
+    vp9_tile_set_row(&tile, cm, tile_row);
+    for (mi_row = tile.mi_row_start; mi_row < tile.mi_row_end;
+         mi_row += MI_BLOCK_SIZE) {
+      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+        const int col = pbi->inv_tile_order ?
+                        tile_cols - tile_col - 1 : tile_col;
+        tile_data = pbi->tile_worker_data + tile_cols * tile_row + col;
+        vp9_tile_set_col(&tile, cm, col);
+        vp9_zero(tile_data->xd.left_context);
+        vp9_zero(tile_data->xd.left_seg_context);
+        for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
+             mi_col += MI_BLOCK_SIZE) {
+          decode_partition(pbi, &tile_data->xd, mi_row,
+                           mi_col, &tile_data->bit_reader, BLOCK_64X64, 4);
+        }
+        pbi->mb.corrupted |= tile_data->xd.corrupted;
+        if (pbi->mb.corrupted)
+            vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                               "Failed to decode tile data");
+      }
+      // Loopfilter one row.
+      if (cm->lf.filter_level && !cm->skip_loop_filter) {
+        const int lf_start = mi_row - MI_BLOCK_SIZE;
+        LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+
+        // delay the loopfilter by 1 macroblock row.
+        if (lf_start < 0) continue;
+
+        // decoding has completed: finish up the loop filter in this thread.
+        if (mi_row + MI_BLOCK_SIZE >= cm->mi_rows) continue;
+
+        winterface->sync(&pbi->lf_worker);
+        lf_data->start = lf_start;
+        lf_data->stop = mi_row;
+        if (pbi->max_threads > 1) {
+          winterface->launch(&pbi->lf_worker);
+        } else {
+          winterface->execute(&pbi->lf_worker);
+        }
+      }
+      // After loopfiltering, the last 7 row pixels in each superblock row may
+      // still be changed by the longest loopfilter of the next superblock
+      // row.
+      if (pbi->frame_parallel_decode)
+        vp9_frameworker_broadcast(pbi->cur_buf,
+                                  mi_row << MI_BLOCK_SIZE_LOG2);
+    }
+  }
+
+  // Loopfilter remaining rows in the frame.
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+    winterface->sync(&pbi->lf_worker);
+    lf_data->start = lf_data->stop;
+    lf_data->stop = cm->mi_rows;
+    winterface->execute(&pbi->lf_worker);
+  }
+
+  // Get last tile data.
+  tile_data = pbi->tile_worker_data + tile_cols * tile_rows - 1;
+
+  if (pbi->frame_parallel_decode)
+    vp9_frameworker_broadcast(pbi->cur_buf, INT_MAX);
+  return vpx_reader_find_end(&tile_data->bit_reader);
+}
+
+// On entry 'tile_data->data_end' points to the end of the input frame, on exit
+// it is updated to reflect the bitreader position of the final tile column if
+// present in the tile buffer group or NULL otherwise.
+static int tile_worker_hook(TileWorkerData *const tile_data,
+                            VP9Decoder *const pbi) {
+  TileInfo *volatile tile = &tile_data->xd.tile;
+  const int final_col = (1 << pbi->common.log2_tile_cols) - 1;
+  const uint8_t *volatile bit_reader_end = NULL;
+  volatile int n = tile_data->buf_start;
+  tile_data->error_info.setjmp = 1;
+
+  if (setjmp(tile_data->error_info.jmp)) {
+    tile_data->error_info.setjmp = 0;
+    tile_data->xd.corrupted = 1;
+    tile_data->data_end = NULL;
+    return 0;
+  }
+
+  tile_data->xd.error_info = &tile_data->error_info;
+  tile_data->xd.corrupted = 0;
+
+  do {
+    int mi_row, mi_col;
+    const TileBuffer *const buf = pbi->tile_buffers + n;
+    vp9_zero(tile_data->dqcoeff);
+    vp9_tile_init(tile, &pbi->common, 0, buf->col);
+    setup_token_decoder(buf->data, tile_data->data_end, buf->size,
+                        &tile_data->error_info, &tile_data->bit_reader,
+                        pbi->decrypt_cb, pbi->decrypt_state);
+    vp9_init_macroblockd(&pbi->common, &tile_data->xd, tile_data->dqcoeff);
+
+    for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
+         mi_row += MI_BLOCK_SIZE) {
+      vp9_zero(tile_data->xd.left_context);
+      vp9_zero(tile_data->xd.left_seg_context);
+      for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
+           mi_col += MI_BLOCK_SIZE) {
+        decode_partition(pbi, &tile_data->xd, mi_row, mi_col,
+                         &tile_data->bit_reader, BLOCK_64X64, 4);
+      }
+    }
+
+    if (buf->col == final_col) {
+      bit_reader_end = vpx_reader_find_end(&tile_data->bit_reader);
+    }
+  } while (!tile_data->xd.corrupted && ++n <= tile_data->buf_end);
+
+  tile_data->data_end = bit_reader_end;
+  return !tile_data->xd.corrupted;
+}
+
+// sorts in descending order
+static int compare_tile_buffers(const void *a, const void *b) {
+  const TileBuffer *const buf1 = (const TileBuffer*)a;
+  const TileBuffer *const buf2 = (const TileBuffer*)b;
+  return (int)(buf2->size - buf1->size);
+}
+
+static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
+                                      const uint8_t *data,
+                                      const uint8_t *data_end) {
+  VP9_COMMON *const cm = &pbi->common;
+  const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
+  const uint8_t *bit_reader_end = NULL;
+  const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  const int num_workers = VPXMIN(pbi->max_threads, tile_cols);
+  int n;
+
+  assert(tile_cols <= (1 << 6));
+  assert(tile_rows == 1);
+  (void)tile_rows;
+
+  if (pbi->num_tile_workers == 0) {
+    const int num_threads = pbi->max_threads;
+    CHECK_MEM_ERROR(cm, pbi->tile_workers,
+                    vpx_malloc(num_threads * sizeof(*pbi->tile_workers)));
+    for (n = 0; n < num_threads; ++n) {
+      VPxWorker *const worker = &pbi->tile_workers[n];
+      ++pbi->num_tile_workers;
+
+      winterface->init(worker);
+      if (n < num_threads - 1 && !winterface->reset(worker)) {
+        vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                           "Tile decoder thread creation failed");
+      }
+    }
+  }
+
+  // Reset tile decoding hook
+  for (n = 0; n < num_workers; ++n) {
+    VPxWorker *const worker = &pbi->tile_workers[n];
+    TileWorkerData *const tile_data =
+        &pbi->tile_worker_data[n + pbi->total_tiles];
+    winterface->sync(worker);
+    tile_data->xd = pbi->mb;
+    tile_data->xd.counts =
+        cm->frame_parallel_decoding_mode ? NULL : &tile_data->counts;
+    worker->hook = (VPxWorkerHook)tile_worker_hook;
+    worker->data1 = tile_data;
+    worker->data2 = pbi;
+  }
+
+  // Note: this memset assumes above_context[0], [1] and [2]
+  // are allocated as part of the same buffer.
+  memset(cm->above_context, 0,
+         sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_mi_cols);
+  memset(cm->above_seg_context, 0,
+         sizeof(*cm->above_seg_context) * aligned_mi_cols);
+
+  vp9_reset_lfm(cm);
+
+  // Load tile data into tile_buffers
+  get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows,
+                   &pbi->tile_buffers);
+
+  // Sort the buffers based on size in descending order.
+  qsort(pbi->tile_buffers, tile_cols, sizeof(pbi->tile_buffers[0]),
+        compare_tile_buffers);
+
+  if (num_workers == tile_cols) {
+    // Rearrange the tile buffers such that the largest, and
+    // presumably the most difficult, tile will be decoded in the main thread.
+    // This should help minimize the number of instances where the main thread
+    // is waiting for a worker to complete.
+    const TileBuffer largest = pbi->tile_buffers[0];
+    memmove(pbi->tile_buffers, pbi->tile_buffers + 1,
+            (tile_cols - 1) * sizeof(pbi->tile_buffers[0]));
+    pbi->tile_buffers[tile_cols - 1] = largest;
+  } else {
+    int start = 0, end = tile_cols - 2;
+    TileBuffer tmp;
+
+    // Interleave the tiles to distribute the load between threads, assuming a
+    // larger tile implies it is more difficult to decode.
+    while (start < end) {
+      tmp = pbi->tile_buffers[start];
+      pbi->tile_buffers[start] = pbi->tile_buffers[end];
+      pbi->tile_buffers[end] = tmp;
+      start += 2;
+      end -= 2;
+    }
+  }
+
+  // Initialize thread frame counts.
+  if (!cm->frame_parallel_decoding_mode) {
+    for (n = 0; n < num_workers; ++n) {
+      TileWorkerData *const tile_data =
+          (TileWorkerData*)pbi->tile_workers[n].data1;
+      vp9_zero(tile_data->counts);
+    }
+  }
+
+  {
+    const int base = tile_cols / num_workers;
+    const int remain = tile_cols % num_workers;
+    int buf_start = 0;
+
+    for (n = 0; n < num_workers; ++n) {
+      const int count = base + (remain + n) / num_workers;
+      VPxWorker *const worker = &pbi->tile_workers[n];
+      TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
+
+      tile_data->buf_start = buf_start;
+      tile_data->buf_end = buf_start + count - 1;
+      tile_data->data_end = data_end;
+      buf_start += count;
+
+      worker->had_error = 0;
+      if (n == num_workers - 1) {
+        assert(tile_data->buf_end == tile_cols - 1);
+        winterface->execute(worker);
+      } else {
+        winterface->launch(worker);
+      }
+    }
+
+    for (; n > 0; --n) {
+      VPxWorker *const worker = &pbi->tile_workers[n - 1];
+      TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
+      // TODO(jzern): The tile may have specific error data associated with
+      // its vpx_internal_error_info which could be propagated to the main info
+      // in cm. Additionally once the threads have been synced and an error is
+      // detected, there's no point in continuing to decode tiles.
+      pbi->mb.corrupted |= !winterface->sync(worker);
+      if (!bit_reader_end) bit_reader_end = tile_data->data_end;
+    }
+  }
+
+  // Accumulate thread frame counts.
+  if (!cm->frame_parallel_decoding_mode) {
+    for (n = 0; n < num_workers; ++n) {
+      TileWorkerData *const tile_data =
+          (TileWorkerData*)pbi->tile_workers[n].data1;
+      vp9_accumulate_frame_counts(&cm->counts, &tile_data->counts, 1);
+    }
+  }
+
+  assert(bit_reader_end || pbi->mb.corrupted);
+  return bit_reader_end;
+}
+
+static void error_handler(void *data) {
+  VP9_COMMON *const cm = (VP9_COMMON *)data;
+  vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME, "Truncated packet");
+}
+
+static void read_bitdepth_colorspace_sampling(
+    VP9_COMMON *cm, struct vpx_read_bit_buffer *rb) {
+  if (cm->profile >= PROFILE_2) {
+    cm->bit_depth = vpx_rb_read_bit(rb) ? VPX_BITS_12 : VPX_BITS_10;
+#if CONFIG_VP9_HIGHBITDEPTH
+    cm->use_highbitdepth = 1;
+#endif
+  } else {
+    cm->bit_depth = VPX_BITS_8;
+#if CONFIG_VP9_HIGHBITDEPTH
+    cm->use_highbitdepth = 0;
+#endif
+  }
+  cm->color_space = vpx_rb_read_literal(rb, 3);
+  if (cm->color_space != VPX_CS_SRGB) {
+    cm->color_range = (vpx_color_range_t)vpx_rb_read_bit(rb);
+    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
+      cm->subsampling_x = vpx_rb_read_bit(rb);
+      cm->subsampling_y = vpx_rb_read_bit(rb);
+      if (cm->subsampling_x == 1 && cm->subsampling_y == 1)
+        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                           "4:2:0 color not supported in profile 1 or 3");
+      if (vpx_rb_read_bit(rb))
+        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                           "Reserved bit set");
+    } else {
+      cm->subsampling_y = cm->subsampling_x = 1;
+    }
+  } else {
+    cm->color_range = VPX_CR_FULL_RANGE;
+    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
+      // Note if colorspace is SRGB then 4:4:4 chroma sampling is assumed.
+      // 4:2:2 or 4:4:0 chroma sampling is not allowed.
+      cm->subsampling_y = cm->subsampling_x = 0;
+      if (vpx_rb_read_bit(rb))
+        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                           "Reserved bit set");
+    } else {
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "4:4:4 color not supported in profile 0 or 2");
+    }
+  }
+}
+
+static size_t read_uncompressed_header(VP9Decoder *pbi,
+                                       struct vpx_read_bit_buffer *rb) {
+  VP9_COMMON *const cm = &pbi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = pool->frame_bufs;
+  int i, mask, ref_index = 0;
+  size_t sz;
+
+  cm->last_frame_type = cm->frame_type;
+  cm->last_intra_only = cm->intra_only;
+
+  if (vpx_rb_read_literal(rb, 2) != VP9_FRAME_MARKER)
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Invalid frame marker");
+
+  cm->profile = vp9_read_profile(rb);
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (cm->profile >= MAX_PROFILES)
+    vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                       "Unsupported bitstream profile");
+#else
+  if (cm->profile >= PROFILE_2)
+    vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                       "Unsupported bitstream profile");
+#endif
+
+  cm->show_existing_frame = vpx_rb_read_bit(rb);
+  if (cm->show_existing_frame) {
+    // Show an existing frame directly.
+    const int frame_to_show = cm->ref_frame_map[vpx_rb_read_literal(rb, 3)];
+    lock_buffer_pool(pool);
+    if (frame_to_show < 0 || frame_bufs[frame_to_show].ref_count < 1) {
+      unlock_buffer_pool(pool);
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Buffer %d does not contain a decoded frame",
+                         frame_to_show);
+    }
+
+    ref_cnt_fb(frame_bufs, &cm->new_fb_idx, frame_to_show);
+    unlock_buffer_pool(pool);
+    pbi->refresh_frame_flags = 0;
+    cm->lf.filter_level = 0;
+    cm->show_frame = 1;
+
+    if (pbi->frame_parallel_decode) {
+      for (i = 0; i < REF_FRAMES; ++i)
+        cm->next_ref_frame_map[i] = cm->ref_frame_map[i];
+    }
+    return 0;
+  }
+
+  cm->frame_type = (FRAME_TYPE) vpx_rb_read_bit(rb);
+  cm->show_frame = vpx_rb_read_bit(rb);
+  cm->error_resilient_mode = vpx_rb_read_bit(rb);
+
+  if (cm->frame_type == KEY_FRAME) {
+    if (!vp9_read_sync_code(rb))
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Invalid frame sync code");
+
+    read_bitdepth_colorspace_sampling(cm, rb);
+    pbi->refresh_frame_flags = (1 << REF_FRAMES) - 1;
+
+    for (i = 0; i < REFS_PER_FRAME; ++i) {
+      cm->frame_refs[i].idx = INVALID_IDX;
+      cm->frame_refs[i].buf = NULL;
+    }
+
+    setup_frame_size(cm, rb);
+    if (pbi->need_resync) {
+      memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+      pbi->need_resync = 0;
+    }
+  } else {
+    cm->intra_only = cm->show_frame ? 0 : vpx_rb_read_bit(rb);
+
+    cm->reset_frame_context = cm->error_resilient_mode ?
+        0 : vpx_rb_read_literal(rb, 2);
+
+    if (cm->intra_only) {
+      if (!vp9_read_sync_code(rb))
+        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                           "Invalid frame sync code");
+      if (cm->profile > PROFILE_0) {
+        read_bitdepth_colorspace_sampling(cm, rb);
+      } else {
+        // NOTE: The intra-only frame header does not include the specification
+        // of either the color format or color sub-sampling in profile 0. VP9
+        // specifies that the default color format should be YUV 4:2:0 in this
+        // case (normative).
+        cm->color_space = VPX_CS_BT_601;
+        cm->color_range = VPX_CR_STUDIO_RANGE;
+        cm->subsampling_y = cm->subsampling_x = 1;
+        cm->bit_depth = VPX_BITS_8;
+#if CONFIG_VP9_HIGHBITDEPTH
+        cm->use_highbitdepth = 0;
+#endif
+      }
+
+      pbi->refresh_frame_flags = vpx_rb_read_literal(rb, REF_FRAMES);
+      setup_frame_size(cm, rb);
+      if (pbi->need_resync) {
+        memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+        pbi->need_resync = 0;
+      }
+    } else if (pbi->need_resync != 1) {  /* Skip if need resync */
+      pbi->refresh_frame_flags = vpx_rb_read_literal(rb, REF_FRAMES);
+      for (i = 0; i < REFS_PER_FRAME; ++i) {
+        const int ref = vpx_rb_read_literal(rb, REF_FRAMES_LOG2);
+        const int idx = cm->ref_frame_map[ref];
+        RefBuffer *const ref_frame = &cm->frame_refs[i];
+        ref_frame->idx = idx;
+        ref_frame->buf = &frame_bufs[idx].buf;
+        cm->ref_frame_sign_bias[LAST_FRAME + i] = vpx_rb_read_bit(rb);
+      }
+
+      setup_frame_size_with_refs(cm, rb);
+
+      cm->allow_high_precision_mv = vpx_rb_read_bit(rb);
+      cm->interp_filter = read_interp_filter(rb);
+
+      for (i = 0; i < REFS_PER_FRAME; ++i) {
+        RefBuffer *const ref_buf = &cm->frame_refs[i];
+#if CONFIG_VP9_HIGHBITDEPTH
+        vp9_setup_scale_factors_for_frame(&ref_buf->sf,
+                                          ref_buf->buf->y_crop_width,
+                                          ref_buf->buf->y_crop_height,
+                                          cm->width, cm->height,
+                                          cm->use_highbitdepth);
+#else
+        vp9_setup_scale_factors_for_frame(&ref_buf->sf,
+                                          ref_buf->buf->y_crop_width,
+                                          ref_buf->buf->y_crop_height,
+                                          cm->width, cm->height);
+#endif
+      }
+    }
+  }
+#if CONFIG_VP9_HIGHBITDEPTH
+  get_frame_new_buffer(cm)->bit_depth = cm->bit_depth;
+#endif
+  get_frame_new_buffer(cm)->color_space = cm->color_space;
+  get_frame_new_buffer(cm)->color_range = cm->color_range;
+  get_frame_new_buffer(cm)->render_width  = cm->render_width;
+  get_frame_new_buffer(cm)->render_height = cm->render_height;
+
+  if (pbi->need_resync) {
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Keyframe / intra-only frame required to reset decoder"
+                       " state");
+  }
+
+  if (!cm->error_resilient_mode) {
+    cm->refresh_frame_context = vpx_rb_read_bit(rb);
+    cm->frame_parallel_decoding_mode = vpx_rb_read_bit(rb);
+    if (!cm->frame_parallel_decoding_mode)
+      vp9_zero(cm->counts);
+  } else {
+    cm->refresh_frame_context = 0;
+    cm->frame_parallel_decoding_mode = 1;
+  }
+
+  // This flag will be overridden by the call to vp9_setup_past_independence
+  // below, forcing the use of context 0 for those frame types.
+  cm->frame_context_idx = vpx_rb_read_literal(rb, FRAME_CONTEXTS_LOG2);
+
+  // Generate next_ref_frame_map.
+  lock_buffer_pool(pool);
+  for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+    if (mask & 1) {
+      cm->next_ref_frame_map[ref_index] = cm->new_fb_idx;
+      ++frame_bufs[cm->new_fb_idx].ref_count;
+    } else {
+      cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+    }
+    // Current thread holds the reference frame.
+    if (cm->ref_frame_map[ref_index] >= 0)
+      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+    ++ref_index;
+  }
+
+  for (; ref_index < REF_FRAMES; ++ref_index) {
+    cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+    // Current thread holds the reference frame.
+    if (cm->ref_frame_map[ref_index] >= 0)
+      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+  }
+  unlock_buffer_pool(pool);
+  pbi->hold_ref_buf = 1;
+
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode)
+    vp9_setup_past_independence(cm);
+
+  setup_loopfilter(&cm->lf, rb);
+  setup_quantization(cm, &pbi->mb, rb);
+  setup_segmentation(&cm->seg, rb);
+  setup_segmentation_dequant(cm);
+
+  setup_tile_info(cm, rb);
+  sz = vpx_rb_read_literal(rb, 16);
+
+  if (sz == 0)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid header size");
+
+  return sz;
+}
+
+static int read_compressed_header(VP9Decoder *pbi, const uint8_t *data,
+                                  size_t partition_size) {
+  VP9_COMMON *const cm = &pbi->common;
+  MACROBLOCKD *const xd = &pbi->mb;
+  FRAME_CONTEXT *const fc = cm->fc;
+  vpx_reader r;
+  int k;
+
+  if (vpx_reader_init(&r, data, partition_size, pbi->decrypt_cb,
+                      pbi->decrypt_state))
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate bool decoder 0");
+
+  cm->tx_mode = xd->lossless ? ONLY_4X4 : read_tx_mode(&r);
+  if (cm->tx_mode == TX_MODE_SELECT)
+    read_tx_mode_probs(&fc->tx_probs, &r);
+  read_coef_probs(fc, cm->tx_mode, &r);
+
+  for (k = 0; k < SKIP_CONTEXTS; ++k)
+    vp9_diff_update_prob(&r, &fc->skip_probs[k]);
+
+  if (!frame_is_intra_only(cm)) {
+    nmv_context *const nmvc = &fc->nmvc;
+    int i, j;
+
+    read_inter_mode_probs(fc, &r);
+
+    if (cm->interp_filter == SWITCHABLE)
+      read_switchable_interp_probs(fc, &r);
+
+    for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+      vp9_diff_update_prob(&r, &fc->intra_inter_prob[i]);
+
+    cm->reference_mode = read_frame_reference_mode(cm, &r);
+    if (cm->reference_mode != SINGLE_REFERENCE)
+      setup_compound_reference_mode(cm);
+    read_frame_reference_mode_probs(cm, &r);
+
+    for (j = 0; j < BLOCK_SIZE_GROUPS; j++)
+      for (i = 0; i < INTRA_MODES - 1; ++i)
+        vp9_diff_update_prob(&r, &fc->y_mode_prob[j][i]);
+
+    for (j = 0; j < PARTITION_CONTEXTS; ++j)
+      for (i = 0; i < PARTITION_TYPES - 1; ++i)
+        vp9_diff_update_prob(&r, &fc->partition_prob[j][i]);
+
+    read_mv_probs(nmvc, cm->allow_high_precision_mv, &r);
+  }
+
+  return vpx_reader_has_error(&r);
+}
+
+static struct vpx_read_bit_buffer *init_read_bit_buffer(
+    VP9Decoder *pbi,
+    struct vpx_read_bit_buffer *rb,
+    const uint8_t *data,
+    const uint8_t *data_end,
+    uint8_t clear_data[MAX_VP9_HEADER_SIZE]) {
+  rb->bit_offset = 0;
+  rb->error_handler = error_handler;
+  rb->error_handler_data = &pbi->common;
+  if (pbi->decrypt_cb) {
+    const int n = (int)VPXMIN(MAX_VP9_HEADER_SIZE, data_end - data);
+    pbi->decrypt_cb(pbi->decrypt_state, data, clear_data, n);
+    rb->bit_buffer = clear_data;
+    rb->bit_buffer_end = clear_data + n;
+  } else {
+    rb->bit_buffer = data;
+    rb->bit_buffer_end = data_end;
+  }
+  return rb;
+}
+
+//------------------------------------------------------------------------------
+
+int vp9_read_sync_code(struct vpx_read_bit_buffer *const rb) {
+  return vpx_rb_read_literal(rb, 8) == VP9_SYNC_CODE_0 &&
+         vpx_rb_read_literal(rb, 8) == VP9_SYNC_CODE_1 &&
+         vpx_rb_read_literal(rb, 8) == VP9_SYNC_CODE_2;
+}
+
+void vp9_read_frame_size(struct vpx_read_bit_buffer *rb,
+                         int *width, int *height) {
+  *width = vpx_rb_read_literal(rb, 16) + 1;
+  *height = vpx_rb_read_literal(rb, 16) + 1;
+}
+
+BITSTREAM_PROFILE vp9_read_profile(struct vpx_read_bit_buffer *rb) {
+  int profile = vpx_rb_read_bit(rb);
+  profile |= vpx_rb_read_bit(rb) << 1;
+  if (profile > 2)
+    profile += vpx_rb_read_bit(rb);
+  return (BITSTREAM_PROFILE) profile;
+}
+
+void vp9_decode_frame(VP9Decoder *pbi,
+                      const uint8_t *data, const uint8_t *data_end,
+                      const uint8_t **p_data_end) {
+  VP9_COMMON *const cm = &pbi->common;
+  MACROBLOCKD *const xd = &pbi->mb;
+  struct vpx_read_bit_buffer rb;
+  int context_updated = 0;
+  uint8_t clear_data[MAX_VP9_HEADER_SIZE];
+  const size_t first_partition_size = read_uncompressed_header(pbi,
+      init_read_bit_buffer(pbi, &rb, data, data_end, clear_data));
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  YV12_BUFFER_CONFIG *const new_fb = get_frame_new_buffer(cm);
+  xd->cur_buf = new_fb;
+
+  if (!first_partition_size) {
+    // showing a frame directly
+    *p_data_end = data + (cm->profile <= PROFILE_2 ? 1 : 2);
+    return;
+  }
+
+  data += vpx_rb_bytes_read(&rb);
+  if (!read_is_valid(data, first_partition_size, data_end))
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt header length");
+
+  cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
+                           cm->width == cm->last_width &&
+                           cm->height == cm->last_height &&
+                           !cm->last_intra_only &&
+                           cm->last_show_frame &&
+                           (cm->last_frame_type != KEY_FRAME);
+
+  vp9_setup_block_planes(xd, cm->subsampling_x, cm->subsampling_y);
+
+  *cm->fc = cm->frame_contexts[cm->frame_context_idx];
+  if (!cm->fc->initialized)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Uninitialized entropy context.");
+
+  xd->corrupted = 0;
+  new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);
+  if (new_fb->corrupted)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Decode failed. Frame data header is corrupted.");
+
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
+  }
+
+  // If encoded in frame parallel mode, frame context is ready after decoding
+  // the frame header.
+  if (pbi->frame_parallel_decode && cm->frame_parallel_decoding_mode) {
+    VPxWorker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    if (cm->refresh_frame_context) {
+      context_updated = 1;
+      cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+    }
+    vp9_frameworker_lock_stats(worker);
+    pbi->cur_buf->row = -1;
+    pbi->cur_buf->col = -1;
+    frame_worker_data->frame_context_ready = 1;
+    // Signal the main thread that context is ready.
+    vp9_frameworker_signal_stats(worker);
+    vp9_frameworker_unlock_stats(worker);
+  }
+
+  if (pbi->tile_worker_data == NULL ||
+      (tile_cols * tile_rows) != pbi->total_tiles) {
+    const int num_tile_workers = tile_cols * tile_rows +
+        ((pbi->max_threads > 1) ? pbi->max_threads : 0);
+    const size_t twd_size = num_tile_workers * sizeof(*pbi->tile_worker_data);
+    // Ensure tile data offsets will be properly aligned. This may fail on
+    // platforms without DECLARE_ALIGNED().
+    assert((sizeof(*pbi->tile_worker_data) % 16) == 0);
+    vpx_free(pbi->tile_worker_data);
+    CHECK_MEM_ERROR(cm, pbi->tile_worker_data, vpx_memalign(32, twd_size));
+    pbi->total_tiles = tile_rows * tile_cols;
+  }
+
+  if (pbi->max_threads > 1 && tile_rows == 1 && tile_cols > 1) {
+    // Multi-threaded tile decoder
+    *p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
+    if (!xd->corrupted) {
+      if (!cm->skip_loop_filter) {
+        // If multiple threads are used to decode tiles, then we use those
+        // threads to do parallel loopfiltering.
+        vp9_loop_filter_frame_mt(new_fb, cm, pbi->mb.plane,
+                                 cm->lf.filter_level, 0, 0, pbi->tile_workers,
+                                 pbi->num_tile_workers, &pbi->lf_row_sync);
+      }
+    } else {
+      vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                         "Decode failed. Frame data is corrupted.");
+    }
+  } else {
+    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
+  }
+
+  if (!xd->corrupted) {
+    if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
+      vp9_adapt_coef_probs(cm);
+
+      if (!frame_is_intra_only(cm)) {
+        vp9_adapt_mode_probs(cm);
+        vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
+      }
+    }
+  } else {
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Decode failed. Frame data is corrupted.");
+  }
+
+  // Non frame parallel update frame context here.
+  if (cm->refresh_frame_context && !context_updated)
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+}
diff --git a/libvpx/libvpx/vp9/decoder/vp9_decodeframe.h b/libvpx/libvpx/vp9/decoder/vp9_decodeframe.h
new file mode 100644
index 0000000..ce33cbd
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_decodeframe.h
@@ -0,0 +1,37 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_DECODER_VP9_DECODEFRAME_H_
+#define VP9_DECODER_VP9_DECODEFRAME_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "vp9/common/vp9_enums.h"
+
+struct VP9Decoder;
+struct vpx_read_bit_buffer;
+
+int vp9_read_sync_code(struct vpx_read_bit_buffer *const rb);
+void vp9_read_frame_size(struct vpx_read_bit_buffer *rb,
+                         int *width, int *height);
+BITSTREAM_PROFILE vp9_read_profile(struct vpx_read_bit_buffer *rb);
+
+void vp9_decode_frame(struct VP9Decoder *pbi,
+                      const uint8_t *data, const uint8_t *data_end,
+                      const uint8_t **p_data_end);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DECODEFRAME_H_
diff --git a/libvpx/libvpx/vp9/decoder/vp9_decodemv.c b/libvpx/libvpx/vp9/decoder/vp9_decodemv.c
new file mode 100644
index 0000000..ffc6839
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_decodemv.c
@@ -0,0 +1,911 @@
+/*
+  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/decoder/vp9_decodemv.h"
+#include "vp9/decoder/vp9_decodeframe.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+
+static PREDICTION_MODE read_intra_mode(vpx_reader *r, const vpx_prob *p) {
+  return (PREDICTION_MODE)vpx_read_tree(r, vp9_intra_mode_tree, p);
+}
+
+static PREDICTION_MODE read_intra_mode_y(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                         vpx_reader *r, int size_group) {
+  const PREDICTION_MODE y_mode =
+      read_intra_mode(r, cm->fc->y_mode_prob[size_group]);
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts)
+    ++counts->y_mode[size_group][y_mode];
+  return y_mode;
+}
+
+static PREDICTION_MODE read_intra_mode_uv(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                          vpx_reader *r,
+                                          PREDICTION_MODE y_mode) {
+  const PREDICTION_MODE uv_mode = read_intra_mode(r,
+                                         cm->fc->uv_mode_prob[y_mode]);
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts)
+    ++counts->uv_mode[y_mode][uv_mode];
+  return uv_mode;
+}
+
+static PREDICTION_MODE read_inter_mode(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                       vpx_reader *r, int ctx) {
+  const int mode = vpx_read_tree(r, vp9_inter_mode_tree,
+                                 cm->fc->inter_mode_probs[ctx]);
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts)
+    ++counts->inter_mode[ctx][mode];
+
+  return NEARESTMV + mode;
+}
+
+static int read_segment_id(vpx_reader *r, const struct segmentation *seg) {
+  return vpx_read_tree(r, vp9_segment_tree, seg->tree_probs);
+}
+
+static TX_SIZE read_selected_tx_size(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                     TX_SIZE max_tx_size, vpx_reader *r) {
+  FRAME_COUNTS *counts = xd->counts;
+  const int ctx = get_tx_size_context(xd);
+  const vpx_prob *tx_probs = get_tx_probs(max_tx_size, ctx, &cm->fc->tx_probs);
+  int tx_size = vpx_read(r, tx_probs[0]);
+  if (tx_size != TX_4X4 && max_tx_size >= TX_16X16) {
+    tx_size += vpx_read(r, tx_probs[1]);
+    if (tx_size != TX_8X8 && max_tx_size >= TX_32X32)
+      tx_size += vpx_read(r, tx_probs[2]);
+  }
+
+  if (counts)
+    ++get_tx_counts(max_tx_size, ctx, &counts->tx)[tx_size];
+  return (TX_SIZE)tx_size;
+}
+
+static INLINE TX_SIZE read_tx_size(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                   int allow_select, vpx_reader *r) {
+  TX_MODE tx_mode = cm->tx_mode;
+  BLOCK_SIZE bsize = xd->mi[0]->sb_type;
+  const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
+  if (allow_select && tx_mode == TX_MODE_SELECT && bsize >= BLOCK_8X8)
+    return read_selected_tx_size(cm, xd, max_tx_size, r);
+  else
+    return VPXMIN(max_tx_size, tx_mode_to_biggest_tx_size[tx_mode]);
+}
+
+static int dec_get_segment_id(const VP9_COMMON *cm, const uint8_t *segment_ids,
+                              int mi_offset, int x_mis, int y_mis) {
+  int x, y, segment_id = INT_MAX;
+
+  for (y = 0; y < y_mis; y++)
+    for (x = 0; x < x_mis; x++)
+      segment_id =
+          VPXMIN(segment_id, segment_ids[mi_offset + y * cm->mi_cols + x]);
+
+  assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
+  return segment_id;
+}
+
+static void set_segment_id(VP9_COMMON *cm, int mi_offset,
+                           int x_mis, int y_mis, int segment_id) {
+  int x, y;
+
+  assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
+
+  for (y = 0; y < y_mis; y++)
+    for (x = 0; x < x_mis; x++)
+      cm->current_frame_seg_map[mi_offset + y * cm->mi_cols + x] = segment_id;
+}
+
+static void copy_segment_id(const VP9_COMMON *cm,
+                           const uint8_t *last_segment_ids,
+                           uint8_t *current_segment_ids,
+                           int mi_offset, int x_mis, int y_mis) {
+  int x, y;
+
+  for (y = 0; y < y_mis; y++)
+    for (x = 0; x < x_mis; x++)
+      current_segment_ids[mi_offset + y * cm->mi_cols + x] =  last_segment_ids ?
+          last_segment_ids[mi_offset + y * cm->mi_cols + x] : 0;
+}
+
+static int read_intra_segment_id(VP9_COMMON *const cm, int mi_offset,
+                                 int x_mis, int y_mis,
+                                 vpx_reader *r) {
+  struct segmentation *const seg = &cm->seg;
+  int segment_id;
+
+  if (!seg->enabled)
+    return 0;  // Default for disabled segmentation
+
+  if (!seg->update_map) {
+    copy_segment_id(cm, cm->last_frame_seg_map, cm->current_frame_seg_map,
+                    mi_offset, x_mis, y_mis);
+    return 0;
+  }
+
+  segment_id = read_segment_id(r, seg);
+  set_segment_id(cm, mi_offset, x_mis, y_mis, segment_id);
+  return segment_id;
+}
+
+static int read_inter_segment_id(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                                 int mi_row, int mi_col, vpx_reader *r,
+                                 int x_mis, int y_mis) {
+  struct segmentation *const seg = &cm->seg;
+  MODE_INFO *const mi = xd->mi[0];
+  int predicted_segment_id, segment_id;
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+
+  if (!seg->enabled)
+    return 0;  // Default for disabled segmentation
+
+  predicted_segment_id = cm->last_frame_seg_map ?
+      dec_get_segment_id(cm, cm->last_frame_seg_map, mi_offset, x_mis, y_mis) :
+      0;
+
+  if (!seg->update_map) {
+    copy_segment_id(cm, cm->last_frame_seg_map, cm->current_frame_seg_map,
+                    mi_offset, x_mis, y_mis);
+    return predicted_segment_id;
+  }
+
+  if (seg->temporal_update) {
+    const vpx_prob pred_prob = vp9_get_pred_prob_seg_id(seg, xd);
+    mi->seg_id_predicted = vpx_read(r, pred_prob);
+    segment_id = mi->seg_id_predicted ? predicted_segment_id
+                                      : read_segment_id(r, seg);
+  } else {
+    segment_id = read_segment_id(r, seg);
+  }
+  set_segment_id(cm, mi_offset, x_mis, y_mis, segment_id);
+  return segment_id;
+}
+
+static int read_skip(VP9_COMMON *cm, const MACROBLOCKD *xd,
+                     int segment_id, vpx_reader *r) {
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
+    return 1;
+  } else {
+    const int ctx = vp9_get_skip_context(xd);
+    const int skip = vpx_read(r, cm->fc->skip_probs[ctx]);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts)
+      ++counts->skip[ctx][skip];
+    return skip;
+  }
+}
+
+static void read_intra_frame_mode_info(VP9_COMMON *const cm,
+                                       MACROBLOCKD *const xd,
+                                       int mi_row, int mi_col, vpx_reader *r,
+                                       int x_mis, int y_mis) {
+  MODE_INFO *const mi = xd->mi[0];
+  const MODE_INFO *above_mi = xd->above_mi;
+  const MODE_INFO *left_mi  = xd->left_mi;
+  const BLOCK_SIZE bsize = mi->sb_type;
+  int i;
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+
+  mi->segment_id = read_intra_segment_id(cm, mi_offset, x_mis, y_mis, r);
+  mi->skip = read_skip(cm, xd, mi->segment_id, r);
+  mi->tx_size = read_tx_size(cm, xd, 1, r);
+  mi->ref_frame[0] = INTRA_FRAME;
+  mi->ref_frame[1] = NONE;
+
+  switch (bsize) {
+    case BLOCK_4X4:
+      for (i = 0; i < 4; ++i)
+        mi->bmi[i].as_mode =
+            read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, i));
+      mi->mode = mi->bmi[3].as_mode;
+      break;
+    case BLOCK_4X8:
+      mi->bmi[0].as_mode = mi->bmi[2].as_mode =
+          read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 0));
+      mi->bmi[1].as_mode = mi->bmi[3].as_mode = mi->mode =
+          read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 1));
+      break;
+    case BLOCK_8X4:
+      mi->bmi[0].as_mode = mi->bmi[1].as_mode =
+          read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 0));
+      mi->bmi[2].as_mode = mi->bmi[3].as_mode = mi->mode =
+          read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 2));
+      break;
+    default:
+      mi->mode = read_intra_mode(r,
+                                 get_y_mode_probs(mi, above_mi, left_mi, 0));
+  }
+
+  mi->uv_mode = read_intra_mode(r, vp9_kf_uv_mode_prob[mi->mode]);
+}
+
+static int read_mv_component(vpx_reader *r,
+                             const nmv_component *mvcomp, int usehp) {
+  int mag, d, fr, hp;
+  const int sign = vpx_read(r, mvcomp->sign);
+  const int mv_class = vpx_read_tree(r, vp9_mv_class_tree, mvcomp->classes);
+  const int class0 = mv_class == MV_CLASS_0;
+
+  // Integer part
+  if (class0) {
+    d = vpx_read_tree(r, vp9_mv_class0_tree, mvcomp->class0);
+    mag = 0;
+  } else {
+    int i;
+    const int n = mv_class + CLASS0_BITS - 1;  // number of bits
+
+    d = 0;
+    for (i = 0; i < n; ++i)
+      d |= vpx_read(r, mvcomp->bits[i]) << i;
+    mag = CLASS0_SIZE << (mv_class + 2);
+  }
+
+  // Fractional part
+  fr = vpx_read_tree(r, vp9_mv_fp_tree, class0 ? mvcomp->class0_fp[d]
+                                               : mvcomp->fp);
+
+  // High precision part (if hp is not used, the default value of the hp is 1)
+  hp = usehp ? vpx_read(r, class0 ? mvcomp->class0_hp : mvcomp->hp)
+             : 1;
+
+  // Result
+  mag += ((d << 3) | (fr << 1) | hp) + 1;
+  return sign ? -mag : mag;
+}
+
+static INLINE void read_mv(vpx_reader *r, MV *mv, const MV *ref,
+                           const nmv_context *ctx,
+                           nmv_context_counts *counts, int allow_hp) {
+  const MV_JOINT_TYPE joint_type =
+      (MV_JOINT_TYPE)vpx_read_tree(r, vp9_mv_joint_tree, ctx->joints);
+  const int use_hp = allow_hp && use_mv_hp(ref);
+  MV diff = {0, 0};
+
+  if (mv_joint_vertical(joint_type))
+    diff.row = read_mv_component(r, &ctx->comps[0], use_hp);
+
+  if (mv_joint_horizontal(joint_type))
+    diff.col = read_mv_component(r, &ctx->comps[1], use_hp);
+
+  vp9_inc_mv(&diff, counts);
+
+  mv->row = ref->row + diff.row;
+  mv->col = ref->col + diff.col;
+}
+
+static REFERENCE_MODE read_block_reference_mode(VP9_COMMON *cm,
+                                                const MACROBLOCKD *xd,
+                                                vpx_reader *r) {
+  if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+    const int ctx = vp9_get_reference_mode_context(cm, xd);
+    const REFERENCE_MODE mode =
+        (REFERENCE_MODE)vpx_read(r, cm->fc->comp_inter_prob[ctx]);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts)
+      ++counts->comp_inter[ctx][mode];
+    return mode;  // SINGLE_REFERENCE or COMPOUND_REFERENCE
+  } else {
+    return cm->reference_mode;
+  }
+}
+
+// Read the referncence frame
+static void read_ref_frames(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                            vpx_reader *r,
+                            int segment_id, MV_REFERENCE_FRAME ref_frame[2]) {
+  FRAME_CONTEXT *const fc = cm->fc;
+  FRAME_COUNTS *counts = xd->counts;
+
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
+    ref_frame[0] = (MV_REFERENCE_FRAME)get_segdata(&cm->seg, segment_id,
+                                                   SEG_LVL_REF_FRAME);
+    ref_frame[1] = NONE;
+  } else {
+    const REFERENCE_MODE mode = read_block_reference_mode(cm, xd, r);
+    // FIXME(rbultje) I'm pretty sure this breaks segmentation ref frame coding
+    if (mode == COMPOUND_REFERENCE) {
+      const int idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
+      const int ctx = vp9_get_pred_context_comp_ref_p(cm, xd);
+      const int bit = vpx_read(r, fc->comp_ref_prob[ctx]);
+      if (counts)
+        ++counts->comp_ref[ctx][bit];
+      ref_frame[idx] = cm->comp_fixed_ref;
+      ref_frame[!idx] = cm->comp_var_ref[bit];
+    } else if (mode == SINGLE_REFERENCE) {
+      const int ctx0 = vp9_get_pred_context_single_ref_p1(xd);
+      const int bit0 = vpx_read(r, fc->single_ref_prob[ctx0][0]);
+      if (counts)
+        ++counts->single_ref[ctx0][0][bit0];
+      if (bit0) {
+        const int ctx1 = vp9_get_pred_context_single_ref_p2(xd);
+        const int bit1 = vpx_read(r, fc->single_ref_prob[ctx1][1]);
+        if (counts)
+          ++counts->single_ref[ctx1][1][bit1];
+        ref_frame[0] = bit1 ? ALTREF_FRAME : GOLDEN_FRAME;
+      } else {
+        ref_frame[0] = LAST_FRAME;
+      }
+
+      ref_frame[1] = NONE;
+    } else {
+      assert(0 && "Invalid prediction mode.");
+    }
+  }
+}
+
+// TODO(slavarnway): Move this decoder version of
+// vp9_get_pred_context_switchable_interp() to vp9_pred_common.h and update the
+// encoder.
+//
+// Returns a context number for the given MB prediction signal
+static int dec_get_pred_context_switchable_interp(const MACROBLOCKD *xd) {
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int left_type = left_mi ? left_mi->interp_filter : SWITCHABLE_FILTERS;
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const int above_type = above_mi ? above_mi->interp_filter
+                             : SWITCHABLE_FILTERS;
+
+  if (left_type == above_type)
+    return left_type;
+  else if (left_type == SWITCHABLE_FILTERS)
+    return above_type;
+  else if (above_type == SWITCHABLE_FILTERS)
+    return left_type;
+  else
+    return SWITCHABLE_FILTERS;
+}
+
+static INLINE INTERP_FILTER read_switchable_interp_filter(
+    VP9_COMMON *const cm, MACROBLOCKD *const xd,
+    vpx_reader *r) {
+  const int ctx = dec_get_pred_context_switchable_interp(xd);
+  const INTERP_FILTER type =
+      (INTERP_FILTER)vpx_read_tree(r, vp9_switchable_interp_tree,
+                                   cm->fc->switchable_interp_prob[ctx]);
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts)
+    ++counts->switchable_interp[ctx][type];
+  return type;
+}
+
+static void read_intra_block_mode_info(VP9_COMMON *const cm,
+                                       MACROBLOCKD *const xd, MODE_INFO *mi,
+                                       vpx_reader *r) {
+  const BLOCK_SIZE bsize = mi->sb_type;
+  int i;
+
+  switch (bsize) {
+    case BLOCK_4X4:
+      for (i = 0; i < 4; ++i)
+        mi->bmi[i].as_mode = read_intra_mode_y(cm, xd, r, 0);
+      mi->mode = mi->bmi[3].as_mode;
+      break;
+    case BLOCK_4X8:
+      mi->bmi[0].as_mode = mi->bmi[2].as_mode = read_intra_mode_y(cm, xd,
+                                                                  r, 0);
+      mi->bmi[1].as_mode = mi->bmi[3].as_mode = mi->mode =
+          read_intra_mode_y(cm, xd, r, 0);
+      break;
+    case BLOCK_8X4:
+      mi->bmi[0].as_mode = mi->bmi[1].as_mode = read_intra_mode_y(cm, xd,
+                                                                  r, 0);
+      mi->bmi[2].as_mode = mi->bmi[3].as_mode = mi->mode =
+          read_intra_mode_y(cm, xd, r, 0);
+      break;
+    default:
+      mi->mode = read_intra_mode_y(cm, xd, r, size_group_lookup[bsize]);
+  }
+
+  mi->uv_mode = read_intra_mode_uv(cm, xd, r, mi->mode);
+
+  // Initialize interp_filter here so we do not have to check for inter block
+  // modes in dec_get_pred_context_switchable_interp()
+  mi->interp_filter = SWITCHABLE_FILTERS;
+
+  mi->ref_frame[0] = INTRA_FRAME;
+  mi->ref_frame[1] = NONE;
+}
+
+static INLINE int is_mv_valid(const MV *mv) {
+  return mv->row > MV_LOW && mv->row < MV_UPP &&
+         mv->col > MV_LOW && mv->col < MV_UPP;
+}
+
+static INLINE void copy_mv_pair(int_mv *dst, const int_mv *src) {
+  memcpy(dst, src, sizeof(*dst) * 2);
+}
+
+static INLINE void zero_mv_pair(int_mv *dst) {
+  memset(dst, 0, sizeof(*dst) * 2);
+}
+
+static INLINE int assign_mv(VP9_COMMON *cm, MACROBLOCKD *xd,
+                            PREDICTION_MODE mode,
+                            int_mv mv[2], int_mv ref_mv[2],
+                            int_mv near_nearest_mv[2],
+                            int is_compound, int allow_hp, vpx_reader *r) {
+  int i;
+  int ret = 1;
+
+  switch (mode) {
+    case NEWMV: {
+      FRAME_COUNTS *counts = xd->counts;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+      for (i = 0; i < 1 + is_compound; ++i) {
+        read_mv(r, &mv[i].as_mv, &ref_mv[i].as_mv, &cm->fc->nmvc, mv_counts,
+                allow_hp);
+        ret = ret && is_mv_valid(&mv[i].as_mv);
+      }
+      break;
+    }
+    case NEARMV:
+    case NEARESTMV: {
+      copy_mv_pair(mv, near_nearest_mv);
+      break;
+    }
+    case ZEROMV: {
+      zero_mv_pair(mv);
+      break;
+    }
+    default: {
+      return 0;
+    }
+  }
+  return ret;
+}
+
+static int read_is_inter_block(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                               int segment_id, vpx_reader *r) {
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
+    return get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME) != INTRA_FRAME;
+  } else {
+    const int ctx = get_intra_inter_context(xd);
+    const int is_inter = vpx_read(r, cm->fc->intra_inter_prob[ctx]);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts)
+      ++counts->intra_inter[ctx][is_inter];
+    return is_inter;
+  }
+}
+
+static void dec_find_best_ref_mvs(int allow_hp, int_mv *mvlist, int_mv *best_mv,
+                                  int refmv_count) {
+  int i;
+
+  // Make sure all the candidates are properly clamped etc
+  for (i = 0; i < refmv_count; ++i) {
+    lower_mv_precision(&mvlist[i].as_mv, allow_hp);
+    *best_mv = mvlist[i];
+  }
+}
+
+static void fpm_sync(void *const data, int mi_row) {
+  VP9Decoder *const pbi = (VP9Decoder *)data;
+  vp9_frameworker_wait(pbi->frame_worker_owner, pbi->common.prev_frame,
+                       mi_row << MI_BLOCK_SIZE_LOG2);
+}
+
+// This macro is used to add a motion vector mv_ref list if it isn't
+// already in the list.  If it's the second motion vector or early_break
+// it will also skip all additional processing and jump to Done!
+#define ADD_MV_REF_LIST_EB(mv, refmv_count, mv_ref_list, Done) \
+  do { \
+    if (refmv_count) { \
+      if ((mv).as_int != (mv_ref_list)[0].as_int) { \
+        (mv_ref_list)[(refmv_count)] = (mv); \
+        refmv_count++; \
+        goto Done; \
+      } \
+    } else { \
+      (mv_ref_list)[(refmv_count)++] = (mv); \
+      if (early_break) \
+        goto Done; \
+    } \
+  } while (0)
+
+// If either reference frame is different, not INTRA, and they
+// are different from each other scale and add the mv to our list.
+#define IF_DIFF_REF_FRAME_ADD_MV_EB(mbmi, ref_frame, ref_sign_bias, \
+                                    refmv_count, mv_ref_list, Done) \
+  do { \
+    if (is_inter_block(mbmi)) { \
+      if ((mbmi)->ref_frame[0] != ref_frame) \
+        ADD_MV_REF_LIST_EB(scale_mv((mbmi), 0, ref_frame, ref_sign_bias), \
+                           refmv_count, mv_ref_list, Done); \
+      if (has_second_ref(mbmi) && \
+          (mbmi)->ref_frame[1] != ref_frame && \
+          (mbmi)->mv[1].as_int != (mbmi)->mv[0].as_int) \
+        ADD_MV_REF_LIST_EB(scale_mv((mbmi), 1, ref_frame, ref_sign_bias), \
+                           refmv_count, mv_ref_list, Done); \
+    } \
+  } while (0)
+
+// This function searches the neighborhood of a given MB/SB
+// to try and find candidate reference vectors.
+static int dec_find_mv_refs(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                            PREDICTION_MODE mode, MV_REFERENCE_FRAME ref_frame,
+                            const POSITION *const mv_ref_search,
+                            int_mv *mv_ref_list,
+                            int mi_row, int mi_col, int block, int is_sub8x8,
+                            find_mv_refs_sync sync, void *const data) {
+  const int *ref_sign_bias = cm->ref_frame_sign_bias;
+  int i, refmv_count = 0;
+  int different_ref_found = 0;
+  const MV_REF *const prev_frame_mvs = cm->use_prev_frame_mvs ?
+      cm->prev_frame->mvs + mi_row * cm->mi_cols + mi_col : NULL;
+  const TileInfo *const tile = &xd->tile;
+  // If mode is nearestmv or newmv (uses nearestmv as a reference) then stop
+  // searching after the first mv is found.
+  const int early_break = (mode != NEARMV);
+
+  // Blank the reference vector list
+  memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
+
+  i = 0;
+  if (is_sub8x8) {
+    // If the size < 8x8 we get the mv from the bmi substructure for the
+    // nearest two blocks.
+    for (i = 0; i < 2; ++i) {
+      const POSITION *const mv_ref = &mv_ref_search[i];
+      if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+        const MODE_INFO *const candidate_mi =
+            xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
+        different_ref_found = 1;
+
+        if (candidate_mi->ref_frame[0] == ref_frame)
+          ADD_MV_REF_LIST_EB(
+              get_sub_block_mv(candidate_mi, 0, mv_ref->col, block),
+                  refmv_count, mv_ref_list, Done);
+        else if (candidate_mi->ref_frame[1] == ref_frame)
+          ADD_MV_REF_LIST_EB(
+              get_sub_block_mv(candidate_mi, 1, mv_ref->col, block),
+                  refmv_count, mv_ref_list, Done);
+      }
+    }
+  }
+
+  // Check the rest of the neighbors in much the same way
+  // as before except we don't need to keep track of sub blocks or
+  // mode counts.
+  for (; i < MVREF_NEIGHBOURS; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MODE_INFO *const candidate =
+          xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
+      different_ref_found = 1;
+
+      if (candidate->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST_EB(candidate->mv[0], refmv_count, mv_ref_list, Done);
+      else if (candidate->ref_frame[1] == ref_frame)
+        ADD_MV_REF_LIST_EB(candidate->mv[1], refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  // TODO(hkuang): Remove this sync after fixing pthread_cond_broadcast
+  // on windows platform. The sync here is unnecessary if use_prev_frame_mvs
+  // is 0. But after removing it, there will be hang in the unit test on windows
+  // due to several threads waiting for a thread's signal.
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+    if (cm->frame_parallel_decode && sync != NULL) {
+      sync(data, mi_row);
+    }
+#endif
+
+  // Check the last frame's mode and mv info.
+  if (prev_frame_mvs) {
+    // Synchronize here for frame parallel decode if sync function is provided.
+    if (cm->frame_parallel_decode && sync != NULL) {
+      sync(data, mi_row);
+    }
+
+    if (prev_frame_mvs->ref_frame[0] == ref_frame) {
+      ADD_MV_REF_LIST_EB(prev_frame_mvs->mv[0], refmv_count, mv_ref_list, Done);
+    } else if (prev_frame_mvs->ref_frame[1] == ref_frame) {
+      ADD_MV_REF_LIST_EB(prev_frame_mvs->mv[1], refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  // Since we couldn't find 2 mvs from the same reference frame
+  // go back through the neighbors and find motion vectors from
+  // different reference frames.
+  if (different_ref_found) {
+    for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
+      const POSITION *mv_ref = &mv_ref_search[i];
+      if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+        const MODE_INFO *const candidate =
+            xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
+
+        // If the candidate is INTRA we don't want to consider its mv.
+        IF_DIFF_REF_FRAME_ADD_MV_EB(candidate, ref_frame, ref_sign_bias,
+                                    refmv_count, mv_ref_list, Done);
+      }
+    }
+  }
+
+  // Since we still don't have a candidate we'll try the last frame.
+  if (prev_frame_mvs) {
+    if (prev_frame_mvs->ref_frame[0] != ref_frame &&
+        prev_frame_mvs->ref_frame[0] > INTRA_FRAME) {
+      int_mv mv = prev_frame_mvs->mv[0];
+      if (ref_sign_bias[prev_frame_mvs->ref_frame[0]] !=
+          ref_sign_bias[ref_frame]) {
+        mv.as_mv.row *= -1;
+        mv.as_mv.col *= -1;
+      }
+      ADD_MV_REF_LIST_EB(mv, refmv_count, mv_ref_list, Done);
+    }
+
+    if (prev_frame_mvs->ref_frame[1] > INTRA_FRAME &&
+        prev_frame_mvs->ref_frame[1] != ref_frame &&
+        prev_frame_mvs->mv[1].as_int != prev_frame_mvs->mv[0].as_int) {
+      int_mv mv = prev_frame_mvs->mv[1];
+      if (ref_sign_bias[prev_frame_mvs->ref_frame[1]] !=
+          ref_sign_bias[ref_frame]) {
+        mv.as_mv.row *= -1;
+        mv.as_mv.col *= -1;
+      }
+      ADD_MV_REF_LIST_EB(mv, refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  if (mode == NEARMV)
+    refmv_count = MAX_MV_REF_CANDIDATES;
+  else
+    // we only care about the nearestmv for the remaining modes
+    refmv_count = 1;
+
+ Done:
+  // Clamp vectors
+  for (i = 0; i < refmv_count; ++i)
+    clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
+
+  return refmv_count;
+}
+
+static void append_sub8x8_mvs_for_idx(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                      const POSITION *const mv_ref_search,
+                                      PREDICTION_MODE b_mode, int block,
+                                      int ref, int mi_row, int mi_col,
+                                      int_mv *best_sub8x8) {
+  int_mv mv_list[MAX_MV_REF_CANDIDATES];
+  MODE_INFO *const mi = xd->mi[0];
+  b_mode_info *bmi = mi->bmi;
+  int n;
+  int refmv_count;
+
+  assert(MAX_MV_REF_CANDIDATES == 2);
+
+  refmv_count = dec_find_mv_refs(cm, xd, b_mode, mi->ref_frame[ref],
+                                 mv_ref_search, mv_list, mi_row, mi_col, block,
+                                 1, NULL, NULL);
+
+  switch (block) {
+    case 0:
+      best_sub8x8->as_int = mv_list[refmv_count - 1].as_int;
+      break;
+    case 1:
+    case 2:
+      if (b_mode == NEARESTMV) {
+        best_sub8x8->as_int = bmi[0].as_mv[ref].as_int;
+      } else {
+        best_sub8x8->as_int = 0;
+        for (n = 0; n < refmv_count; ++n)
+          if (bmi[0].as_mv[ref].as_int != mv_list[n].as_int) {
+            best_sub8x8->as_int = mv_list[n].as_int;
+            break;
+          }
+      }
+      break;
+    case 3:
+      if (b_mode == NEARESTMV) {
+        best_sub8x8->as_int = bmi[2].as_mv[ref].as_int;
+      } else {
+        int_mv candidates[2 + MAX_MV_REF_CANDIDATES];
+        candidates[0] = bmi[1].as_mv[ref];
+        candidates[1] = bmi[0].as_mv[ref];
+        candidates[2] = mv_list[0];
+        candidates[3] = mv_list[1];
+        best_sub8x8->as_int = 0;
+        for (n = 0; n < 2 + MAX_MV_REF_CANDIDATES; ++n)
+          if (bmi[2].as_mv[ref].as_int != candidates[n].as_int) {
+            best_sub8x8->as_int = candidates[n].as_int;
+            break;
+          }
+      }
+      break;
+    default:
+      assert(0 && "Invalid block index.");
+  }
+}
+
+static uint8_t get_mode_context(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                                const POSITION *const mv_ref_search,
+                                int mi_row, int mi_col) {
+  int i;
+  int context_counter = 0;
+  const TileInfo *const tile = &xd->tile;
+
+  // Get mode count from nearest 2 blocks
+  for (i = 0; i < 2; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MODE_INFO *const candidate =
+          xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
+      // Keep counts for entropy encoding.
+      context_counter += mode_2_counter[candidate->mode];
+    }
+  }
+
+  return counter_to_context[context_counter];
+}
+
+static void read_inter_block_mode_info(VP9Decoder *const pbi,
+                                       MACROBLOCKD *const xd,
+                                       MODE_INFO *const mi,
+                                       int mi_row, int mi_col, vpx_reader *r) {
+  VP9_COMMON *const cm = &pbi->common;
+  const BLOCK_SIZE bsize = mi->sb_type;
+  const int allow_hp = cm->allow_high_precision_mv;
+  int_mv best_ref_mvs[2];
+  int ref, is_compound;
+  uint8_t inter_mode_ctx;
+  const POSITION *const mv_ref_search = mv_ref_blocks[bsize];
+
+  read_ref_frames(cm, xd, r, mi->segment_id, mi->ref_frame);
+  is_compound = has_second_ref(mi);
+  inter_mode_ctx = get_mode_context(cm, xd, mv_ref_search, mi_row, mi_col);
+
+  if (segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP)) {
+    mi->mode = ZEROMV;
+    if (bsize < BLOCK_8X8) {
+        vpx_internal_error(xd->error_info, VPX_CODEC_UNSUP_BITSTREAM,
+                           "Invalid usage of segement feature on small blocks");
+        return;
+    }
+  } else {
+    if (bsize >= BLOCK_8X8)
+      mi->mode = read_inter_mode(cm, xd, r, inter_mode_ctx);
+    else
+      // Sub 8x8 blocks use the nearestmv as a ref_mv if the b_mode is NEWMV.
+      // Setting mode to NEARESTMV forces the search to stop after the nearestmv
+      // has been found. After b_modes have been read, mode will be overwritten
+      // by the last b_mode.
+      mi->mode = NEARESTMV;
+
+    if (mi->mode != ZEROMV) {
+      for (ref = 0; ref < 1 + is_compound; ++ref) {
+        int_mv tmp_mvs[MAX_MV_REF_CANDIDATES];
+        const MV_REFERENCE_FRAME frame = mi->ref_frame[ref];
+        int refmv_count;
+
+        refmv_count = dec_find_mv_refs(cm, xd, mi->mode, frame, mv_ref_search,
+                                       tmp_mvs, mi_row, mi_col, -1, 0,
+                                       fpm_sync, (void *)pbi);
+
+        dec_find_best_ref_mvs(allow_hp, tmp_mvs, &best_ref_mvs[ref],
+                              refmv_count);
+      }
+    }
+  }
+
+  mi->interp_filter = (cm->interp_filter == SWITCHABLE)
+                      ? read_switchable_interp_filter(cm, xd, r)
+                      : cm->interp_filter;
+
+  if (bsize < BLOCK_8X8) {
+    const int num_4x4_w = 1 << xd->bmode_blocks_wl;
+    const int num_4x4_h = 1 << xd->bmode_blocks_hl;
+    int idx, idy;
+    PREDICTION_MODE b_mode;
+    int_mv best_sub8x8[2];
+    for (idy = 0; idy < 2; idy += num_4x4_h) {
+      for (idx = 0; idx < 2; idx += num_4x4_w) {
+        const int j = idy * 2 + idx;
+        b_mode = read_inter_mode(cm, xd, r, inter_mode_ctx);
+
+        if (b_mode == NEARESTMV || b_mode == NEARMV) {
+          for (ref = 0; ref < 1 + is_compound; ++ref)
+            append_sub8x8_mvs_for_idx(cm, xd, mv_ref_search, b_mode, j, ref,
+                                      mi_row, mi_col, &best_sub8x8[ref]);
+        }
+
+        if (!assign_mv(cm, xd, b_mode, mi->bmi[j].as_mv, best_ref_mvs,
+                       best_sub8x8, is_compound, allow_hp, r)) {
+          xd->corrupted |= 1;
+          break;
+        }
+
+        if (num_4x4_h == 2)
+          mi->bmi[j + 2] = mi->bmi[j];
+        if (num_4x4_w == 2)
+          mi->bmi[j + 1] = mi->bmi[j];
+      }
+    }
+
+    mi->mode = b_mode;
+
+    copy_mv_pair(mi->mv, mi->bmi[3].as_mv);
+  } else {
+    xd->corrupted |= !assign_mv(cm, xd, mi->mode, mi->mv, best_ref_mvs,
+                                best_ref_mvs, is_compound, allow_hp, r);
+  }
+}
+
+static void read_inter_frame_mode_info(VP9Decoder *const pbi,
+                                       MACROBLOCKD *const xd,
+                                       int mi_row, int mi_col, vpx_reader *r,
+                                       int x_mis, int y_mis) {
+  VP9_COMMON *const cm = &pbi->common;
+  MODE_INFO *const mi = xd->mi[0];
+  int inter_block;
+
+  mi->segment_id = read_inter_segment_id(cm, xd, mi_row, mi_col, r, x_mis,
+                                         y_mis);
+  mi->skip = read_skip(cm, xd, mi->segment_id, r);
+  inter_block = read_is_inter_block(cm, xd, mi->segment_id, r);
+  mi->tx_size = read_tx_size(cm, xd, !mi->skip || !inter_block, r);
+
+  if (inter_block)
+    read_inter_block_mode_info(pbi, xd, mi, mi_row, mi_col, r);
+  else
+    read_intra_block_mode_info(cm, xd, mi, r);
+}
+
+static INLINE void copy_ref_frame_pair(MV_REFERENCE_FRAME *dst,
+                                       const MV_REFERENCE_FRAME *src) {
+  memcpy(dst, src, sizeof(*dst) * 2);
+}
+
+void vp9_read_mode_info(VP9Decoder *const pbi, MACROBLOCKD *xd,
+                        int mi_row, int mi_col, vpx_reader *r,
+                        int x_mis, int y_mis) {
+  VP9_COMMON *const cm = &pbi->common;
+  MODE_INFO *const mi = xd->mi[0];
+  MV_REF* frame_mvs = cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
+  int w, h;
+
+  if (frame_is_intra_only(cm)) {
+    read_intra_frame_mode_info(cm, xd, mi_row, mi_col, r, x_mis, y_mis);
+  } else {
+    read_inter_frame_mode_info(pbi, xd, mi_row, mi_col, r, x_mis, y_mis);
+
+    for (h = 0; h < y_mis; ++h) {
+      for (w = 0; w < x_mis; ++w) {
+        MV_REF *const mv = frame_mvs + w;
+        copy_ref_frame_pair(mv->ref_frame, mi->ref_frame);
+        copy_mv_pair(mv->mv, mi->mv);
+      }
+      frame_mvs += cm->mi_cols;
+    }
+  }
+#if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
+    if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
+        (xd->above_mi == NULL || xd->left_mi == NULL) &&
+        !is_inter_block(mi) && need_top_left[mi->uv_mode])
+      assert(0);
+#endif  // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
+}
diff --git a/libvpx/libvpx/vp9/decoder/vp9_decodemv.h b/libvpx/libvpx/vp9/decoder/vp9_decodemv.h
new file mode 100644
index 0000000..45569ec
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_decodemv.h
@@ -0,0 +1,30 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_DECODER_VP9_DECODEMV_H_
+#define VP9_DECODER_VP9_DECODEMV_H_
+
+#include "vpx_dsp/bitreader.h"
+
+#include "vp9/decoder/vp9_decoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_read_mode_info(VP9Decoder *const pbi, MACROBLOCKD *xd,
+                        int mi_row, int mi_col, vpx_reader *r,
+                        int x_mis, int y_mis);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DECODEMV_H_
diff --git a/libvpx/libvpx/vp9/decoder/vp9_decoder.c b/libvpx/libvpx/vp9/decoder/vp9_decoder.c
new file mode 100644
index 0000000..935c04f
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_decoder.c
@@ -0,0 +1,518 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <stdio.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/system_state.h"
+#include "vpx_ports/vpx_once.h"
+#include "vpx_ports/vpx_timer.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vpx_util/vpx_thread.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#if CONFIG_VP9_POSTPROC
+#include "vp9/common/vp9_postproc.h"
+#endif
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconintra.h"
+
+#include "vp9/decoder/vp9_decodeframe.h"
+#include "vp9/decoder/vp9_decoder.h"
+#include "vp9/decoder/vp9_detokenize.h"
+
+static void initialize_dec(void) {
+  static volatile int init_done = 0;
+
+  if (!init_done) {
+    vp9_rtcd();
+    vpx_dsp_rtcd();
+    vpx_scale_rtcd();
+    vp9_init_intra_predictors();
+    init_done = 1;
+  }
+}
+
+static void vp9_dec_setup_mi(VP9_COMMON *cm) {
+  cm->mi = cm->mip + cm->mi_stride + 1;
+  cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
+  memset(cm->mi_grid_base, 0,
+         cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
+}
+
+static int vp9_dec_alloc_mi(VP9_COMMON *cm, int mi_size) {
+  cm->mip = vpx_calloc(mi_size, sizeof(*cm->mip));
+  if (!cm->mip)
+    return 1;
+  cm->mi_alloc_size = mi_size;
+  cm->mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
+  if (!cm->mi_grid_base)
+    return 1;
+  return 0;
+}
+
+static void vp9_dec_free_mi(VP9_COMMON *cm) {
+  vpx_free(cm->mip);
+  cm->mip = NULL;
+  vpx_free(cm->mi_grid_base);
+  cm->mi_grid_base = NULL;
+}
+
+VP9Decoder *vp9_decoder_create(BufferPool *const pool) {
+  VP9Decoder *volatile const pbi = vpx_memalign(32, sizeof(*pbi));
+  VP9_COMMON *volatile const cm = pbi ? &pbi->common : NULL;
+
+  if (!cm)
+    return NULL;
+
+  vp9_zero(*pbi);
+
+  if (setjmp(cm->error.jmp)) {
+    cm->error.setjmp = 0;
+    vp9_decoder_remove(pbi);
+    return NULL;
+  }
+
+  cm->error.setjmp = 1;
+
+  CHECK_MEM_ERROR(cm, cm->fc,
+                  (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc)));
+  CHECK_MEM_ERROR(cm, cm->frame_contexts,
+                  (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS,
+                  sizeof(*cm->frame_contexts)));
+
+  pbi->need_resync = 1;
+  once(initialize_dec);
+
+  // Initialize the references to not point to any frame buffers.
+  memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+  memset(&cm->next_ref_frame_map, -1, sizeof(cm->next_ref_frame_map));
+
+  cm->current_video_frame = 0;
+  pbi->ready_for_new_data = 1;
+  pbi->common.buffer_pool = pool;
+
+  cm->bit_depth = VPX_BITS_8;
+  cm->dequant_bit_depth = VPX_BITS_8;
+
+  cm->alloc_mi = vp9_dec_alloc_mi;
+  cm->free_mi = vp9_dec_free_mi;
+  cm->setup_mi = vp9_dec_setup_mi;
+
+  vp9_loop_filter_init(cm);
+
+  cm->error.setjmp = 0;
+
+  vpx_get_worker_interface()->init(&pbi->lf_worker);
+
+  return pbi;
+}
+
+void vp9_decoder_remove(VP9Decoder *pbi) {
+  int i;
+
+  if (!pbi)
+    return;
+
+  vpx_get_worker_interface()->end(&pbi->lf_worker);
+  vpx_free(pbi->lf_worker.data1);
+
+  for (i = 0; i < pbi->num_tile_workers; ++i) {
+    VPxWorker *const worker = &pbi->tile_workers[i];
+    vpx_get_worker_interface()->end(worker);
+  }
+
+  vpx_free(pbi->tile_worker_data);
+  vpx_free(pbi->tile_workers);
+
+  if (pbi->num_tile_workers > 0) {
+    vp9_loop_filter_dealloc(&pbi->lf_row_sync);
+  }
+
+  vpx_free(pbi);
+}
+
+static int equal_dimensions(const YV12_BUFFER_CONFIG *a,
+                            const YV12_BUFFER_CONFIG *b) {
+    return a->y_height == b->y_height && a->y_width == b->y_width &&
+           a->uv_height == b->uv_height && a->uv_width == b->uv_width;
+}
+
+vpx_codec_err_t vp9_copy_reference_dec(VP9Decoder *pbi,
+                                       VP9_REFFRAME ref_frame_flag,
+                                       YV12_BUFFER_CONFIG *sd) {
+  VP9_COMMON *cm = &pbi->common;
+
+  /* TODO(jkoleszar): The decoder doesn't have any real knowledge of what the
+   * encoder is using the frame buffers for. This is just a stub to keep the
+   * vpxenc --test-decode functionality working, and will be replaced in a
+   * later commit that adds VP9-specific controls for this functionality.
+   */
+  if (ref_frame_flag == VP9_LAST_FLAG) {
+    const YV12_BUFFER_CONFIG *const cfg = get_ref_frame(cm, 0);
+    if (cfg == NULL) {
+      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                         "No 'last' reference frame");
+      return VPX_CODEC_ERROR;
+    }
+    if (!equal_dimensions(cfg, sd))
+      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                         "Incorrect buffer dimensions");
+    else
+      vp8_yv12_copy_frame(cfg, sd);
+  } else {
+    vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                       "Invalid reference frame");
+  }
+
+  return cm->error.error_code;
+}
+
+
+vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
+                                      VP9_REFFRAME ref_frame_flag,
+                                      YV12_BUFFER_CONFIG *sd) {
+  RefBuffer *ref_buf = NULL;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+
+  // TODO(jkoleszar): The decoder doesn't have any real knowledge of what the
+  // encoder is using the frame buffers for. This is just a stub to keep the
+  // vpxenc --test-decode functionality working, and will be replaced in a
+  // later commit that adds VP9-specific controls for this functionality.
+  if (ref_frame_flag == VP9_LAST_FLAG) {
+    ref_buf = &cm->frame_refs[0];
+  } else if (ref_frame_flag == VP9_GOLD_FLAG) {
+    ref_buf = &cm->frame_refs[1];
+  } else if (ref_frame_flag == VP9_ALT_FLAG) {
+    ref_buf = &cm->frame_refs[2];
+  } else {
+    vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                       "Invalid reference frame");
+    return cm->error.error_code;
+  }
+
+  if (!equal_dimensions(ref_buf->buf, sd)) {
+    vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                       "Incorrect buffer dimensions");
+  } else {
+    int *ref_fb_ptr = &ref_buf->idx;
+
+    // Find an empty frame buffer.
+    const int free_fb = get_free_fb(cm);
+    if (cm->new_fb_idx == INVALID_IDX) {
+      vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                         "Unable to find free frame buffer");
+      return cm->error.error_code;
+    }
+
+    // Decrease ref_count since it will be increased again in
+    // ref_cnt_fb() below.
+    --frame_bufs[free_fb].ref_count;
+
+    // Manage the reference counters and copy image.
+    ref_cnt_fb(frame_bufs, ref_fb_ptr, free_fb);
+    ref_buf->buf = &frame_bufs[*ref_fb_ptr].buf;
+    vp8_yv12_copy_frame(sd, ref_buf->buf);
+  }
+
+  return cm->error.error_code;
+}
+
+/* If any buffer updating is signaled it should be done here. */
+static void swap_frame_buffers(VP9Decoder *pbi) {
+  int ref_index = 0, mask;
+  VP9_COMMON *const cm = &pbi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+
+  lock_buffer_pool(pool);
+  for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+    const int old_idx = cm->ref_frame_map[ref_index];
+    // Current thread releases the holding of reference frame.
+    decrease_ref_count(old_idx, frame_bufs, pool);
+
+    // Release the reference frame in reference map.
+    if (mask & 1) {
+      decrease_ref_count(old_idx, frame_bufs, pool);
+    }
+    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
+    ++ref_index;
+  }
+
+  // Current thread releases the holding of reference frame.
+  for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
+    const int old_idx = cm->ref_frame_map[ref_index];
+    decrease_ref_count(old_idx, frame_bufs, pool);
+    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
+  }
+  unlock_buffer_pool(pool);
+  pbi->hold_ref_buf = 0;
+  cm->frame_to_show = get_frame_new_buffer(cm);
+
+  if (!pbi->frame_parallel_decode || !cm->show_frame) {
+    lock_buffer_pool(pool);
+    --frame_bufs[cm->new_fb_idx].ref_count;
+    unlock_buffer_pool(pool);
+  }
+
+  // Invalidate these references until the next frame starts.
+  for (ref_index = 0; ref_index < 3; ref_index++)
+    cm->frame_refs[ref_index].idx = -1;
+}
+
+int vp9_receive_compressed_data(VP9Decoder *pbi,
+                                size_t size, const uint8_t **psource) {
+  VP9_COMMON *volatile const cm = &pbi->common;
+  BufferPool *volatile const pool = cm->buffer_pool;
+  RefCntBuffer *volatile const frame_bufs = cm->buffer_pool->frame_bufs;
+  const uint8_t *source = *psource;
+  int retcode = 0;
+  cm->error.error_code = VPX_CODEC_OK;
+
+  if (size == 0) {
+    // This is used to signal that we are missing frames.
+    // We do not know if the missing frame(s) was supposed to update
+    // any of the reference buffers, but we act conservative and
+    // mark only the last buffer as corrupted.
+    //
+    // TODO(jkoleszar): Error concealment is undefined and non-normative
+    // at this point, but if it becomes so, [0] may not always be the correct
+    // thing to do here.
+    if (cm->frame_refs[0].idx > 0) {
+      assert(cm->frame_refs[0].buf != NULL);
+      cm->frame_refs[0].buf->corrupted = 1;
+    }
+  }
+
+  pbi->ready_for_new_data = 0;
+
+  // Check if the previous frame was a frame without any references to it.
+  // Release frame buffer if not decoding in frame parallel mode.
+  if (!pbi->frame_parallel_decode && cm->new_fb_idx >= 0
+      && frame_bufs[cm->new_fb_idx].ref_count == 0)
+    pool->release_fb_cb(pool->cb_priv,
+                        &frame_bufs[cm->new_fb_idx].raw_frame_buffer);
+  // Find a free frame buffer. Return error if can not find any.
+  cm->new_fb_idx = get_free_fb(cm);
+  if (cm->new_fb_idx == INVALID_IDX) {
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Unable to find free frame buffer");
+    return cm->error.error_code;
+  }
+
+  // Assign a MV array to the frame buffer.
+  cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
+
+  pbi->hold_ref_buf = 0;
+  if (pbi->frame_parallel_decode) {
+    VPxWorker *const worker = pbi->frame_worker_owner;
+    vp9_frameworker_lock_stats(worker);
+    frame_bufs[cm->new_fb_idx].frame_worker_owner = worker;
+    // Reset decoding progress.
+    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
+    pbi->cur_buf->row = -1;
+    pbi->cur_buf->col = -1;
+    vp9_frameworker_unlock_stats(worker);
+  } else {
+    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
+  }
+
+
+  if (setjmp(cm->error.jmp)) {
+    const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
+    int i;
+
+    cm->error.setjmp = 0;
+    pbi->ready_for_new_data = 1;
+
+    // Synchronize all threads immediately as a subsequent decode call may
+    // cause a resize invalidating some allocations.
+    winterface->sync(&pbi->lf_worker);
+    for (i = 0; i < pbi->num_tile_workers; ++i) {
+      winterface->sync(&pbi->tile_workers[i]);
+    }
+
+    lock_buffer_pool(pool);
+    // Release all the reference buffers if worker thread is holding them.
+    if (pbi->hold_ref_buf == 1) {
+      int ref_index = 0, mask;
+      for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+        const int old_idx = cm->ref_frame_map[ref_index];
+        // Current thread releases the holding of reference frame.
+        decrease_ref_count(old_idx, frame_bufs, pool);
+
+        // Release the reference frame in reference map.
+        if (mask & 1) {
+          decrease_ref_count(old_idx, frame_bufs, pool);
+        }
+        ++ref_index;
+      }
+
+      // Current thread releases the holding of reference frame.
+      for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
+        const int old_idx = cm->ref_frame_map[ref_index];
+        decrease_ref_count(old_idx, frame_bufs, pool);
+      }
+      pbi->hold_ref_buf = 0;
+    }
+    // Release current frame.
+    decrease_ref_count(cm->new_fb_idx, frame_bufs, pool);
+    unlock_buffer_pool(pool);
+
+    vpx_clear_system_state();
+    return -1;
+  }
+
+  cm->error.setjmp = 1;
+  vp9_decode_frame(pbi, source, source + size, psource);
+
+  swap_frame_buffers(pbi);
+
+  vpx_clear_system_state();
+
+  if (!cm->show_existing_frame) {
+    cm->last_show_frame = cm->show_frame;
+    cm->prev_frame = cm->cur_frame;
+    if (cm->seg.enabled && !pbi->frame_parallel_decode)
+      vp9_swap_current_and_last_seg_map(cm);
+  }
+
+  // Update progress in frame parallel decode.
+  if (pbi->frame_parallel_decode) {
+    // Need to lock the mutex here as another thread may
+    // be accessing this buffer.
+    VPxWorker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    vp9_frameworker_lock_stats(worker);
+
+    if (cm->show_frame) {
+      cm->current_video_frame++;
+    }
+    frame_worker_data->frame_decoded = 1;
+    frame_worker_data->frame_context_ready = 1;
+    vp9_frameworker_signal_stats(worker);
+    vp9_frameworker_unlock_stats(worker);
+  } else {
+    cm->last_width = cm->width;
+    cm->last_height = cm->height;
+    if (cm->show_frame) {
+      cm->current_video_frame++;
+    }
+  }
+
+  cm->error.setjmp = 0;
+  return retcode;
+}
+
+int vp9_get_raw_frame(VP9Decoder *pbi, YV12_BUFFER_CONFIG *sd,
+                      vp9_ppflags_t *flags) {
+  VP9_COMMON *const cm = &pbi->common;
+  int ret = -1;
+#if !CONFIG_VP9_POSTPROC
+  (void)*flags;
+#endif
+
+  if (pbi->ready_for_new_data == 1)
+    return ret;
+
+  pbi->ready_for_new_data = 1;
+
+  /* no raw frame to show!!! */
+  if (!cm->show_frame)
+    return ret;
+
+  pbi->ready_for_new_data = 1;
+
+#if CONFIG_VP9_POSTPROC
+  if (!cm->show_existing_frame) {
+    ret = vp9_post_proc_frame(cm, sd, flags);
+  } else {
+    *sd = *cm->frame_to_show;
+    ret = 0;
+  }
+#else
+  *sd = *cm->frame_to_show;
+  ret = 0;
+#endif /*!CONFIG_POSTPROC*/
+  vpx_clear_system_state();
+  return ret;
+}
+
+vpx_codec_err_t vp9_parse_superframe_index(const uint8_t *data,
+                                           size_t data_sz,
+                                           uint32_t sizes[8], int *count,
+                                           vpx_decrypt_cb decrypt_cb,
+                                           void *decrypt_state) {
+  // A chunk ending with a byte matching 0xc0 is an invalid chunk unless
+  // it is a super frame index. If the last byte of real video compression
+  // data is 0xc0 the encoder must add a 0 byte. If we have the marker but
+  // not the associated matching marker byte at the front of the index we have
+  // an invalid bitstream and need to return an error.
+
+  uint8_t marker;
+
+  assert(data_sz);
+  marker = read_marker(decrypt_cb, decrypt_state, data + data_sz - 1);
+  *count = 0;
+
+  if ((marker & 0xe0) == 0xc0) {
+    const uint32_t frames = (marker & 0x7) + 1;
+    const uint32_t mag = ((marker >> 3) & 0x3) + 1;
+    const size_t index_sz = 2 + mag * frames;
+
+    // This chunk is marked as having a superframe index but doesn't have
+    // enough data for it, thus it's an invalid superframe index.
+    if (data_sz < index_sz)
+      return VPX_CODEC_CORRUPT_FRAME;
+
+    {
+      const uint8_t marker2 = read_marker(decrypt_cb, decrypt_state,
+                                          data + data_sz - index_sz);
+
+      // This chunk is marked as having a superframe index but doesn't have
+      // the matching marker byte at the front of the index therefore it's an
+      // invalid chunk.
+      if (marker != marker2)
+        return VPX_CODEC_CORRUPT_FRAME;
+    }
+
+    {
+      // Found a valid superframe index.
+      uint32_t i, j;
+      const uint8_t *x = &data[data_sz - index_sz + 1];
+
+      // Frames has a maximum of 8 and mag has a maximum of 4.
+      uint8_t clear_buffer[32];
+      assert(sizeof(clear_buffer) >= frames * mag);
+      if (decrypt_cb) {
+        decrypt_cb(decrypt_state, x, clear_buffer, frames * mag);
+        x = clear_buffer;
+      }
+
+      for (i = 0; i < frames; ++i) {
+        uint32_t this_sz = 0;
+
+        for (j = 0; j < mag; ++j)
+          this_sz |= ((uint32_t)(*x++)) << (j * 8);
+        sizes[i] = this_sz;
+      }
+      *count = frames;
+    }
+  }
+  return VPX_CODEC_OK;
+}
diff --git a/libvpx/libvpx/vp9/decoder/vp9_decoder.h b/libvpx/libvpx/vp9/decoder/vp9_decoder.h
new file mode 100644
index 0000000..7111a36
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_decoder.h
@@ -0,0 +1,137 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_DECODER_VP9_DECODER_H_
+#define VP9_DECODER_VP9_DECODER_H_
+
+#include "./vpx_config.h"
+
+#include "vpx/vpx_codec.h"
+#include "vpx_dsp/bitreader.h"
+#include "vpx_scale/yv12config.h"
+#include "vpx_util/vpx_thread.h"
+
+#include "vp9/common/vp9_thread_common.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_ppflags.h"
+#include "vp9/decoder/vp9_dthread.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct TileBuffer {
+  const uint8_t *data;
+  size_t size;
+  int col;  // only used with multi-threaded decoding
+} TileBuffer;
+
+typedef struct TileWorkerData {
+  const uint8_t *data_end;
+  int buf_start, buf_end;  // pbi->tile_buffers to decode, inclusive
+  vpx_reader bit_reader;
+  FRAME_COUNTS counts;
+  DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+  /* dqcoeff are shared by all the planes. So planes must be decoded serially */
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[32 * 32]);
+  struct vpx_internal_error_info error_info;
+} TileWorkerData;
+
+typedef struct VP9Decoder {
+  DECLARE_ALIGNED(16, MACROBLOCKD, mb);
+
+  DECLARE_ALIGNED(16, VP9_COMMON, common);
+
+  int ready_for_new_data;
+
+  int refresh_frame_flags;
+
+  int frame_parallel_decode;  // frame-based threading.
+
+  // TODO(hkuang): Combine this with cur_buf in macroblockd as they are
+  // the same.
+  RefCntBuffer *cur_buf;   //  Current decoding frame buffer.
+
+  VPxWorker *frame_worker_owner;   // frame_worker that owns this pbi.
+  VPxWorker lf_worker;
+  VPxWorker *tile_workers;
+  TileWorkerData *tile_worker_data;
+  TileBuffer tile_buffers[64];
+  int num_tile_workers;
+  int total_tiles;
+
+  VP9LfSync lf_row_sync;
+
+  vpx_decrypt_cb decrypt_cb;
+  void *decrypt_state;
+
+  int max_threads;
+  int inv_tile_order;
+  int need_resync;  // wait for key/intra-only frame.
+  int hold_ref_buf;  // hold the reference buffer.
+} VP9Decoder;
+
+int vp9_receive_compressed_data(struct VP9Decoder *pbi,
+                                size_t size, const uint8_t **dest);
+
+int vp9_get_raw_frame(struct VP9Decoder *pbi, YV12_BUFFER_CONFIG *sd,
+                      vp9_ppflags_t *flags);
+
+vpx_codec_err_t vp9_copy_reference_dec(struct VP9Decoder *pbi,
+                                       VP9_REFFRAME ref_frame_flag,
+                                       YV12_BUFFER_CONFIG *sd);
+
+vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
+                                      VP9_REFFRAME ref_frame_flag,
+                                      YV12_BUFFER_CONFIG *sd);
+
+static INLINE uint8_t read_marker(vpx_decrypt_cb decrypt_cb,
+                                  void *decrypt_state,
+                                  const uint8_t *data) {
+  if (decrypt_cb) {
+    uint8_t marker;
+    decrypt_cb(decrypt_state, data, &marker, 1);
+    return marker;
+  }
+  return *data;
+}
+
+// This function is exposed for use in tests, as well as the inlined function
+// "read_marker".
+vpx_codec_err_t vp9_parse_superframe_index(const uint8_t *data,
+                                           size_t data_sz,
+                                           uint32_t sizes[8], int *count,
+                                           vpx_decrypt_cb decrypt_cb,
+                                           void *decrypt_state);
+
+struct VP9Decoder *vp9_decoder_create(BufferPool *const pool);
+
+void vp9_decoder_remove(struct VP9Decoder *pbi);
+
+static INLINE void decrease_ref_count(int idx, RefCntBuffer *const frame_bufs,
+                                      BufferPool *const pool) {
+  if (idx >= 0 && frame_bufs[idx].ref_count > 0) {
+    --frame_bufs[idx].ref_count;
+    // A worker may only get a free framebuffer index when calling get_free_fb.
+    // But the private buffer is not set up until finish decoding header.
+    // So any error happens during decoding header, the frame_bufs will not
+    // have valid priv buffer.
+    if (frame_bufs[idx].ref_count == 0 &&
+        frame_bufs[idx].raw_frame_buffer.priv) {
+      pool->release_fb_cb(pool->cb_priv, &frame_bufs[idx].raw_frame_buffer);
+    }
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DECODER_H_
diff --git a/libvpx/libvpx/vp9/decoder/vp9_detokenize.c b/libvpx/libvpx/vp9/decoder/vp9_detokenize.c
new file mode 100644
index 0000000..47dc107
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_detokenize.c
@@ -0,0 +1,224 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+#include "vp9/common/vp9_idct.h"
+#endif
+
+#include "vp9/decoder/vp9_detokenize.h"
+
+#define EOB_CONTEXT_NODE            0
+#define ZERO_CONTEXT_NODE           1
+#define ONE_CONTEXT_NODE            2
+
+#define INCREMENT_COUNT(token)                              \
+  do {                                                      \
+     if (counts)                                            \
+       ++coef_counts[band][ctx][token];                     \
+  } while (0)
+
+static INLINE int read_coeff(const vpx_prob *probs, int n, vpx_reader *r) {
+  int i, val = 0;
+  for (i = 0; i < n; ++i)
+    val = (val << 1) | vpx_read(r, probs[i]);
+  return val;
+}
+
+static int decode_coefs(const MACROBLOCKD *xd,
+                        PLANE_TYPE type,
+                        tran_low_t *dqcoeff, TX_SIZE tx_size, const int16_t *dq,
+                        int ctx, const int16_t *scan, const int16_t *nb,
+                        vpx_reader *r) {
+  FRAME_COUNTS *counts = xd->counts;
+  const int max_eob = 16 << (tx_size << 1);
+  const FRAME_CONTEXT *const fc = xd->fc;
+  const int ref = is_inter_block(xd->mi[0]);
+  int band, c = 0;
+  const vpx_prob (*coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
+      fc->coef_probs[tx_size][type][ref];
+  const vpx_prob *prob;
+  unsigned int (*coef_counts)[COEFF_CONTEXTS][UNCONSTRAINED_NODES + 1];
+  unsigned int (*eob_branch_count)[COEFF_CONTEXTS];
+  uint8_t token_cache[32 * 32];
+  const uint8_t *band_translate = get_band_translate(tx_size);
+  const int dq_shift = (tx_size == TX_32X32);
+  int v, token;
+  int16_t dqv = dq[0];
+  const uint8_t *const cat6_prob =
+#if CONFIG_VP9_HIGHBITDEPTH
+      (xd->bd == VPX_BITS_12) ? vp9_cat6_prob_high12 :
+      (xd->bd == VPX_BITS_10) ? vp9_cat6_prob_high12 + 2 :
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      vp9_cat6_prob;
+  const int cat6_bits =
+#if CONFIG_VP9_HIGHBITDEPTH
+      (xd->bd == VPX_BITS_12) ? 18 :
+      (xd->bd == VPX_BITS_10) ? 16 :
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      14;
+
+  if (counts) {
+    coef_counts = counts->coef[tx_size][type][ref];
+    eob_branch_count = counts->eob_branch[tx_size][type][ref];
+  }
+
+  while (c < max_eob) {
+    int val = -1;
+    band = *band_translate++;
+    prob = coef_probs[band][ctx];
+    if (counts)
+      ++eob_branch_count[band][ctx];
+    if (!vpx_read(r, prob[EOB_CONTEXT_NODE])) {
+      INCREMENT_COUNT(EOB_MODEL_TOKEN);
+      break;
+    }
+
+    while (!vpx_read(r, prob[ZERO_CONTEXT_NODE])) {
+      INCREMENT_COUNT(ZERO_TOKEN);
+      dqv = dq[1];
+      token_cache[scan[c]] = 0;
+      ++c;
+      if (c >= max_eob)
+        return c;  // zero tokens at the end (no eob token)
+      ctx = get_coef_context(nb, token_cache, c);
+      band = *band_translate++;
+      prob = coef_probs[band][ctx];
+    }
+
+    if (!vpx_read(r, prob[ONE_CONTEXT_NODE])) {
+      INCREMENT_COUNT(ONE_TOKEN);
+      token = ONE_TOKEN;
+      val = 1;
+    } else {
+      INCREMENT_COUNT(TWO_TOKEN);
+      token = vpx_read_tree(r, vp9_coef_con_tree,
+                            vp9_pareto8_full[prob[PIVOT_NODE] - 1]);
+      switch (token) {
+        case TWO_TOKEN:
+        case THREE_TOKEN:
+        case FOUR_TOKEN:
+          val = token;
+          break;
+        case CATEGORY1_TOKEN:
+          val = CAT1_MIN_VAL + read_coeff(vp9_cat1_prob, 1, r);
+          break;
+        case CATEGORY2_TOKEN:
+          val = CAT2_MIN_VAL + read_coeff(vp9_cat2_prob, 2, r);
+          break;
+        case CATEGORY3_TOKEN:
+          val = CAT3_MIN_VAL + read_coeff(vp9_cat3_prob, 3, r);
+          break;
+        case CATEGORY4_TOKEN:
+          val = CAT4_MIN_VAL + read_coeff(vp9_cat4_prob, 4, r);
+          break;
+        case CATEGORY5_TOKEN:
+          val = CAT5_MIN_VAL + read_coeff(vp9_cat5_prob, 5, r);
+          break;
+        case CATEGORY6_TOKEN:
+          val = CAT6_MIN_VAL + read_coeff(cat6_prob, cat6_bits, r);
+          break;
+      }
+    }
+    v = (val * dqv) >> dq_shift;
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+#if CONFIG_VP9_HIGHBITDEPTH
+    dqcoeff[scan[c]] = highbd_check_range((vpx_read_bit(r) ? -v : v),
+                                          xd->bd);
+#else
+    dqcoeff[scan[c]] = check_range(vpx_read_bit(r) ? -v : v);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#else
+    dqcoeff[scan[c]] = vpx_read_bit(r) ? -v : v;
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+    token_cache[scan[c]] = vp9_pt_energy_class[token];
+    ++c;
+    ctx = get_coef_context(nb, token_cache, c);
+    dqv = dq[1];
+  }
+
+  return c;
+}
+
+static void get_ctx_shift(MACROBLOCKD *xd, int *ctx_shift_a, int *ctx_shift_l,
+                          int x, int y, unsigned int tx_size_in_blocks) {
+  if (xd->max_blocks_wide) {
+    if (tx_size_in_blocks + x > xd->max_blocks_wide)
+      *ctx_shift_a = (tx_size_in_blocks - (xd->max_blocks_wide - x)) * 8;
+  }
+  if (xd->max_blocks_high) {
+    if (tx_size_in_blocks + y > xd->max_blocks_high)
+      *ctx_shift_l = (tx_size_in_blocks - (xd->max_blocks_high - y)) * 8;
+  }
+}
+
+int vp9_decode_block_tokens(MACROBLOCKD *xd, int plane, const scan_order *sc,
+                            int x, int y, TX_SIZE tx_size, vpx_reader *r,
+                            int seg_id) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int16_t *const dequant = pd->seg_dequant[seg_id];
+  int eob;
+  ENTROPY_CONTEXT *a = pd->above_context + x;
+  ENTROPY_CONTEXT *l = pd->left_context + y;
+  int ctx;
+  int ctx_shift_a = 0;
+  int ctx_shift_l = 0;
+
+  switch (tx_size) {
+    case TX_4X4:
+      ctx  = a[0] != 0;
+      ctx += l[0] != 0;
+      eob = decode_coefs(xd, get_plane_type(plane), pd->dqcoeff, tx_size,
+                         dequant, ctx, sc->scan, sc->neighbors, r);
+      a[0] = l[0] = (eob > 0);
+      break;
+    case TX_8X8:
+      get_ctx_shift(xd, &ctx_shift_a, &ctx_shift_l, x, y, 1 << TX_8X8);
+      ctx  = !!*(const uint16_t *)a;
+      ctx += !!*(const uint16_t *)l;
+      eob = decode_coefs(xd, get_plane_type(plane), pd->dqcoeff, tx_size,
+                         dequant, ctx, sc->scan, sc->neighbors, r);
+      *(uint16_t *)a = ((eob > 0) * 0x0101) >> ctx_shift_a;
+      *(uint16_t *)l = ((eob > 0) * 0x0101) >> ctx_shift_l;
+      break;
+    case TX_16X16:
+      get_ctx_shift(xd, &ctx_shift_a, &ctx_shift_l, x, y, 1 << TX_16X16);
+      ctx  = !!*(const uint32_t *)a;
+      ctx += !!*(const uint32_t *)l;
+      eob = decode_coefs(xd, get_plane_type(plane), pd->dqcoeff, tx_size,
+                         dequant, ctx, sc->scan, sc->neighbors, r);
+      *(uint32_t *)a = ((eob > 0) * 0x01010101) >> ctx_shift_a;
+      *(uint32_t *)l = ((eob > 0) * 0x01010101) >> ctx_shift_l;
+      break;
+    case TX_32X32:
+      get_ctx_shift(xd, &ctx_shift_a, &ctx_shift_l, x, y, 1 << TX_32X32);
+      // NOTE: casting to uint64_t here is safe because the default memory
+      // alignment is at least 8 bytes and the TX_32X32 is aligned on 8 byte
+      // boundaries.
+      ctx  = !!*(const uint64_t *)a;
+      ctx += !!*(const uint64_t *)l;
+      eob = decode_coefs(xd, get_plane_type(plane), pd->dqcoeff, tx_size,
+                         dequant, ctx, sc->scan, sc->neighbors, r);
+      *(uint64_t *)a = ((eob > 0) * 0x0101010101010101ULL) >> ctx_shift_a;
+      *(uint64_t *)l = ((eob > 0) * 0x0101010101010101ULL) >> ctx_shift_l;
+      break;
+    default:
+      assert(0 && "Invalid transform size.");
+      eob = 0;
+      break;
+  }
+
+  return eob;
+}
diff --git a/libvpx/libvpx/vp9/decoder/vp9_detokenize.h b/libvpx/libvpx/vp9/decoder/vp9_detokenize.h
new file mode 100644
index 0000000..d242d44
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_detokenize.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_DECODER_VP9_DETOKENIZE_H_
+#define VP9_DECODER_VP9_DETOKENIZE_H_
+
+#include "vpx_dsp/bitreader.h"
+#include "vp9/decoder/vp9_decoder.h"
+#include "vp9/common/vp9_scan.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int vp9_decode_block_tokens(MACROBLOCKD *xd,
+                            int plane, const scan_order *sc,
+                            int x, int y,
+                            TX_SIZE tx_size, vpx_reader *r,
+                            int seg_id);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DETOKENIZE_H_
diff --git a/libvpx/libvpx/vp9/decoder/vp9_dsubexp.c b/libvpx/libvpx/vp9/decoder/vp9_dsubexp.c
new file mode 100644
index 0000000..05b3853
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_dsubexp.c
@@ -0,0 +1,76 @@
+/*
+  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_entropy.h"
+
+#include "vp9/decoder/vp9_dsubexp.h"
+
+static int inv_recenter_nonneg(int v, int m) {
+  if (v > 2 * m)
+    return v;
+
+  return (v & 1) ? m - ((v + 1) >> 1) : m + (v >> 1);
+}
+
+static int decode_uniform(vpx_reader *r) {
+  const int l = 8;
+  const int m = (1 << l) - 191;
+  const int v = vpx_read_literal(r, l - 1);
+  return v < m ?  v : (v << 1) - m + vpx_read_bit(r);
+}
+
+static int inv_remap_prob(int v, int m) {
+  static uint8_t inv_map_table[MAX_PROB] = {
+      7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176, 189,
+    202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,  10,  11,
+     12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,  25,  26,  27,
+     28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,
+     44,  45,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  60,
+     61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  73,  74,  75,  76,
+     77,  78,  79,  80,  81,  82,  83,  84,  86,  87,  88,  89,  90,  91,  92,
+     93,  94,  95,  96,  97,  99, 100, 101, 102, 103, 104, 105, 106, 107, 108,
+    109, 110, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 125,
+    126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 138, 139, 140, 141,
+    142, 143, 144, 145, 146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157,
+    158, 159, 160, 161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173,
+    174, 175, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190,
+    191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,
+    207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221, 222,
+    223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238,
+    239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 253
+  };
+  assert(v < (int)(sizeof(inv_map_table) / sizeof(inv_map_table[0])));
+  v = inv_map_table[v];
+  m--;
+  if ((m << 1) <= MAX_PROB) {
+    return 1 + inv_recenter_nonneg(v, m);
+  } else {
+    return MAX_PROB - inv_recenter_nonneg(v, MAX_PROB - 1 - m);
+  }
+}
+
+static int decode_term_subexp(vpx_reader *r) {
+  if (!vpx_read_bit(r))
+    return vpx_read_literal(r, 4);
+  if (!vpx_read_bit(r))
+    return vpx_read_literal(r, 4) + 16;
+  if (!vpx_read_bit(r))
+    return vpx_read_literal(r, 5) + 32;
+  return decode_uniform(r) + 64;
+}
+
+void vp9_diff_update_prob(vpx_reader *r, vpx_prob* p) {
+  if (vpx_read(r, DIFF_UPDATE_PROB)) {
+    const int delp = decode_term_subexp(r);
+    *p = (vpx_prob)inv_remap_prob(delp, *p);
+  }
+}
diff --git a/libvpx/libvpx/vp9/decoder/vp9_dsubexp.h b/libvpx/libvpx/vp9/decoder/vp9_dsubexp.h
new file mode 100644
index 0000000..a8bcc70
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_dsubexp.h
@@ -0,0 +1,27 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_DECODER_VP9_DSUBEXP_H_
+#define VP9_DECODER_VP9_DSUBEXP_H_
+
+#include "vpx_dsp/bitreader.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_diff_update_prob(vpx_reader *r, vpx_prob* p);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DSUBEXP_H_
diff --git a/libvpx/libvpx/vp9/decoder/vp9_dthread.c b/libvpx/libvpx/vp9/decoder/vp9_dthread.c
new file mode 100644
index 0000000..14a7144
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_dthread.c
@@ -0,0 +1,189 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/decoder/vp9_dthread.h"
+#include "vp9/decoder/vp9_decoder.h"
+
+// #define DEBUG_THREAD
+
+// TODO(hkuang): Clean up all the #ifdef in this file.
+void vp9_frameworker_lock_stats(VPxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  pthread_mutex_lock(&worker_data->stats_mutex);
+#else
+  (void)worker;
+#endif
+}
+
+void vp9_frameworker_unlock_stats(VPxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  pthread_mutex_unlock(&worker_data->stats_mutex);
+#else
+  (void)worker;
+#endif
+}
+
+void vp9_frameworker_signal_stats(VPxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+
+// TODO(hkuang): Fix the pthread_cond_broadcast in windows wrapper.
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+  pthread_cond_signal(&worker_data->stats_cond);
+#else
+  pthread_cond_broadcast(&worker_data->stats_cond);
+#endif
+
+#else
+  (void)worker;
+#endif
+}
+
+// This macro prevents thread_sanitizer from reporting known concurrent writes.
+#if defined(__has_feature)
+#if __has_feature(thread_sanitizer)
+#define BUILDING_WITH_TSAN
+#endif
+#endif
+
+// TODO(hkuang): Remove worker parameter as it is only used in debug code.
+void vp9_frameworker_wait(VPxWorker *const worker, RefCntBuffer *const ref_buf,
+                          int row) {
+#if CONFIG_MULTITHREAD
+  if (!ref_buf)
+    return;
+
+#ifndef BUILDING_WITH_TSAN
+  // The following line of code will get harmless tsan error but it is the key
+  // to get best performance.
+  if (ref_buf->row >= row && ref_buf->buf.corrupted != 1) return;
+#endif
+
+  {
+    // Find the worker thread that owns the reference frame. If the reference
+    // frame has been fully decoded, it may not have owner.
+    VPxWorker *const ref_worker = ref_buf->frame_worker_owner;
+    FrameWorkerData *const ref_worker_data =
+        (FrameWorkerData *)ref_worker->data1;
+    const VP9Decoder *const pbi = ref_worker_data->pbi;
+
+#ifdef DEBUG_THREAD
+    {
+      FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+      printf("%d %p worker is waiting for %d %p worker (%d)  ref %d \r\n",
+             worker_data->worker_id, worker, ref_worker_data->worker_id,
+             ref_buf->frame_worker_owner, row, ref_buf->row);
+    }
+#endif
+
+    vp9_frameworker_lock_stats(ref_worker);
+    while (ref_buf->row < row && pbi->cur_buf == ref_buf &&
+           ref_buf->buf.corrupted != 1) {
+      pthread_cond_wait(&ref_worker_data->stats_cond,
+                        &ref_worker_data->stats_mutex);
+    }
+
+    if (ref_buf->buf.corrupted == 1) {
+      FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+      vp9_frameworker_unlock_stats(ref_worker);
+      vpx_internal_error(&worker_data->pbi->common.error,
+                         VPX_CODEC_CORRUPT_FRAME,
+                         "Worker %p failed to decode frame", worker);
+    }
+    vp9_frameworker_unlock_stats(ref_worker);
+  }
+#else
+  (void)worker;
+  (void)ref_buf;
+  (void)row;
+  (void)ref_buf;
+#endif  // CONFIG_MULTITHREAD
+}
+
+void vp9_frameworker_broadcast(RefCntBuffer *const buf, int row) {
+#if CONFIG_MULTITHREAD
+  VPxWorker *worker = buf->frame_worker_owner;
+
+#ifdef DEBUG_THREAD
+  {
+    FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+    printf("%d %p worker decode to (%d) \r\n", worker_data->worker_id,
+           buf->frame_worker_owner, row);
+  }
+#endif
+
+  vp9_frameworker_lock_stats(worker);
+  buf->row = row;
+  vp9_frameworker_signal_stats(worker);
+  vp9_frameworker_unlock_stats(worker);
+#else
+  (void)buf;
+  (void)row;
+#endif  // CONFIG_MULTITHREAD
+}
+
+void vp9_frameworker_copy_context(VPxWorker *const dst_worker,
+                                  VPxWorker *const src_worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const src_worker_data = (FrameWorkerData *)src_worker->data1;
+  FrameWorkerData *const dst_worker_data = (FrameWorkerData *)dst_worker->data1;
+  VP9_COMMON *const src_cm = &src_worker_data->pbi->common;
+  VP9_COMMON *const dst_cm = &dst_worker_data->pbi->common;
+  int i;
+
+  // Wait until source frame's context is ready.
+  vp9_frameworker_lock_stats(src_worker);
+  while (!src_worker_data->frame_context_ready) {
+    pthread_cond_wait(&src_worker_data->stats_cond,
+        &src_worker_data->stats_mutex);
+  }
+
+  dst_cm->last_frame_seg_map = src_cm->seg.enabled ?
+      src_cm->current_frame_seg_map : src_cm->last_frame_seg_map;
+  dst_worker_data->pbi->need_resync = src_worker_data->pbi->need_resync;
+  vp9_frameworker_unlock_stats(src_worker);
+
+  dst_cm->bit_depth = src_cm->bit_depth;
+#if CONFIG_VP9_HIGHBITDEPTH
+  dst_cm->use_highbitdepth = src_cm->use_highbitdepth;
+#endif
+  dst_cm->prev_frame = src_cm->show_existing_frame ?
+                       src_cm->prev_frame : src_cm->cur_frame;
+  dst_cm->last_width = !src_cm->show_existing_frame ?
+                       src_cm->width : src_cm->last_width;
+  dst_cm->last_height = !src_cm->show_existing_frame ?
+                        src_cm->height : src_cm->last_height;
+  dst_cm->subsampling_x = src_cm->subsampling_x;
+  dst_cm->subsampling_y = src_cm->subsampling_y;
+  dst_cm->frame_type = src_cm->frame_type;
+  dst_cm->last_show_frame = !src_cm->show_existing_frame ?
+                            src_cm->show_frame : src_cm->last_show_frame;
+  for (i = 0; i < REF_FRAMES; ++i)
+    dst_cm->ref_frame_map[i] = src_cm->next_ref_frame_map[i];
+
+  memcpy(dst_cm->lf_info.lfthr, src_cm->lf_info.lfthr,
+         (MAX_LOOP_FILTER + 1) * sizeof(loop_filter_thresh));
+  dst_cm->lf.last_sharpness_level = src_cm->lf.sharpness_level;
+  dst_cm->lf.filter_level = src_cm->lf.filter_level;
+  memcpy(dst_cm->lf.ref_deltas, src_cm->lf.ref_deltas, MAX_REF_LF_DELTAS);
+  memcpy(dst_cm->lf.mode_deltas, src_cm->lf.mode_deltas, MAX_MODE_LF_DELTAS);
+  dst_cm->seg = src_cm->seg;
+  memcpy(dst_cm->frame_contexts, src_cm->frame_contexts,
+         FRAME_CONTEXTS * sizeof(dst_cm->frame_contexts[0]));
+#else
+  (void) dst_worker;
+  (void) src_worker;
+#endif  // CONFIG_MULTITHREAD
+}
diff --git a/libvpx/libvpx/vp9/decoder/vp9_dthread.h b/libvpx/libvpx/vp9/decoder/vp9_dthread.h
new file mode 100644
index 0000000..ba7c38a
--- /dev/null
+++ b/libvpx/libvpx/vp9/decoder/vp9_dthread.h
@@ -0,0 +1,74 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_DECODER_VP9_DTHREAD_H_
+#define VP9_DECODER_VP9_DTHREAD_H_
+
+#include "./vpx_config.h"
+#include "vpx_util/vpx_thread.h"
+#include "vpx/internal/vpx_codec_internal.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9Common;
+struct VP9Decoder;
+
+// WorkerData for the FrameWorker thread. It contains all the information of
+// the worker and decode structures for decoding a frame.
+typedef struct FrameWorkerData {
+  struct VP9Decoder *pbi;
+  const uint8_t *data;
+  const uint8_t *data_end;
+  size_t data_size;
+  void *user_priv;
+  int result;
+  int worker_id;
+  int received_frame;
+
+  // scratch_buffer is used in frame parallel mode only.
+  // It is used to make a copy of the compressed data.
+  uint8_t *scratch_buffer;
+  size_t scratch_buffer_size;
+
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t stats_mutex;
+  pthread_cond_t stats_cond;
+#endif
+
+  int frame_context_ready;  // Current frame's context is ready to read.
+  int frame_decoded;        // Finished decoding current frame.
+} FrameWorkerData;
+
+void vp9_frameworker_lock_stats(VPxWorker *const worker);
+void vp9_frameworker_unlock_stats(VPxWorker *const worker);
+void vp9_frameworker_signal_stats(VPxWorker *const worker);
+
+// Wait until ref_buf has been decoded to row in real pixel unit.
+// Note: worker may already finish decoding ref_buf and release it in order to
+// start decoding next frame. So need to check whether worker is still decoding
+// ref_buf.
+void vp9_frameworker_wait(VPxWorker *const worker, RefCntBuffer *const ref_buf,
+                          int row);
+
+// FrameWorker broadcasts its decoding progress so other workers that are
+// waiting on it can resume decoding.
+void vp9_frameworker_broadcast(RefCntBuffer *const buf, int row);
+
+// Copy necessary decoding context from src worker to dst worker.
+void vp9_frameworker_copy_context(VPxWorker *const dst_worker,
+                                  VPxWorker *const src_worker);
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DTHREAD_H_
diff --git a/libvpx/libvpx/vp9/encoder/arm/neon/vp9_dct_neon.c b/libvpx/libvpx/vp9/encoder/arm/neon/vp9_dct_neon.c
new file mode 100644
index 0000000..11e8773
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/arm/neon/vp9_dct_neon.c
@@ -0,0 +1,36 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vpx_dsp/txfm_common.h"
+
+void vp9_fdct8x8_quant_neon(const int16_t *input, int stride,
+                            int16_t* coeff_ptr, intptr_t n_coeffs,
+                            int skip_block, const int16_t* zbin_ptr,
+                            const int16_t* round_ptr, const int16_t* quant_ptr,
+                            const int16_t* quant_shift_ptr,
+                            int16_t* qcoeff_ptr, int16_t* dqcoeff_ptr,
+                            const int16_t* dequant_ptr, uint16_t* eob_ptr,
+                            const int16_t* scan_ptr,
+                            const int16_t* iscan_ptr) {
+  int16_t temp_buffer[64];
+  (void)coeff_ptr;
+
+  vpx_fdct8x8_neon(input, temp_buffer, stride);
+  vp9_quantize_fp_neon(temp_buffer, n_coeffs, skip_block, zbin_ptr, round_ptr,
+                       quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
+                       dequant_ptr, eob_ptr, scan_ptr, iscan_ptr);
+}
diff --git a/libvpx/libvpx/vp9/encoder/arm/neon/vp9_error_neon.c b/libvpx/libvpx/vp9/encoder/arm/neon/vp9_error_neon.c
new file mode 100644
index 0000000..1c75031
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/arm/neon/vp9_error_neon.c
@@ -0,0 +1,41 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./vp9_rtcd.h"
+
+int64_t vp9_block_error_fp_neon(const int16_t *coeff, const int16_t *dqcoeff,
+                                int block_size) {
+  int64x2_t error = vdupq_n_s64(0);
+
+  assert(block_size >= 8);
+  assert((block_size % 8) == 0);
+
+  do {
+    const int16x8_t c = vld1q_s16(coeff);
+    const int16x8_t d = vld1q_s16(dqcoeff);
+    const int16x8_t diff = vsubq_s16(c, d);
+    const int16x4_t diff_lo = vget_low_s16(diff);
+    const int16x4_t diff_hi = vget_high_s16(diff);
+    // diff is 15-bits, the squares 30, so we can store 2 in 31-bits before
+    // accumulating them in 64-bits.
+    const int32x4_t err0 = vmull_s16(diff_lo, diff_lo);
+    const int32x4_t err1 = vmlal_s16(err0, diff_hi, diff_hi);
+    const int64x2_t err2 = vaddl_s32(vget_low_s32(err1), vget_high_s32(err1));
+    error = vaddq_s64(error, err2);
+    coeff += 8;
+    dqcoeff += 8;
+    block_size -= 8;
+  } while (block_size != 0);
+
+  return vgetq_lane_s64(error, 0) + vgetq_lane_s64(error, 1);
+}
diff --git a/libvpx/libvpx/vp9/encoder/arm/neon/vp9_quantize_neon.c b/libvpx/libvpx/vp9/encoder/arm/neon/vp9_quantize_neon.c
new file mode 100644
index 0000000..47363c7
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/arm/neon/vp9_quantize_neon.c
@@ -0,0 +1,118 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include <math.h>
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_rd.h"
+
+void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
+                          int skip_block, const int16_t *zbin_ptr,
+                          const int16_t *round_ptr, const int16_t *quant_ptr,
+                          const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                          int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                          uint16_t *eob_ptr,
+                          const int16_t *scan, const int16_t *iscan) {
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)scan;
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    int i;
+    const int16x8_t v_zero = vdupq_n_s16(0);
+    const int16x8_t v_one = vdupq_n_s16(1);
+    int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
+    int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
+    int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
+    int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    // adjust for dc
+    v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
+    v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
+    v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
+    // process dc and the first seven ac coeffs
+    {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[0]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+      const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
+                                           vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
+                                           vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 = vcombine_s16(vshrn_n_s32(v_tmp_lo, 16),
+                                            vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+      vst1q_s16(&qcoeff_ptr[0], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[0], v_dqcoeff);
+      v_round = vmovq_n_s16(round_ptr[1]);
+      v_quant = vmovq_n_s16(quant_ptr[1]);
+      v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    }
+    // now process the rest of the ac coeffs
+    for (i = 8; i < count; i += 8) {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[i]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+      const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
+                                           vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
+                                           vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 = vcombine_s16(vshrn_n_s32(v_tmp_lo, 16),
+                                            vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+      vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[i], v_dqcoeff);
+    }
+    {
+      const int16x4_t v_eobmax_3210 =
+          vmax_s16(vget_low_s16(v_eobmax_76543210),
+                   vget_high_s16(v_eobmax_76543210));
+      const int64x1_t v_eobmax_xx32 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
+      const int16x4_t v_eobmax_tmp =
+          vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
+      const int64x1_t v_eobmax_xxx3 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
+      const int16x4_t v_eobmax_final =
+          vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
+
+      *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
+    }
+  } else {
+    memset(qcoeff_ptr, 0, count * sizeof(int16_t));
+    memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
+    *eob_ptr = 0;
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/mips/msa/vp9_error_msa.c b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_error_msa.c
new file mode 100644
index 0000000..1dc70bd
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_error_msa.c
@@ -0,0 +1,114 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "vpx_dsp/mips/macros_msa.h"
+
+#define BLOCK_ERROR_BLOCKSIZE_MSA(BSize)                                   \
+static int64_t block_error_##BSize##size_msa(const int16_t *coeff_ptr,     \
+                                             const int16_t *dq_coeff_ptr,  \
+                                             int64_t *ssz) {               \
+  int64_t err = 0;                                                         \
+  uint32_t loop_cnt;                                                       \
+  v8i16 coeff, dq_coeff, coeff_r_h, coeff_l_h;                             \
+  v4i32 diff_r, diff_l, coeff_r_w, coeff_l_w;                              \
+  v2i64 sq_coeff_r, sq_coeff_l;                                            \
+  v2i64 err0, err_dup0, err1, err_dup1;                                    \
+                                                                           \
+  coeff = LD_SH(coeff_ptr);                                                \
+  dq_coeff = LD_SH(dq_coeff_ptr);                                          \
+  UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                                \
+  ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                      \
+  HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                       \
+  DOTP_SW2_SD(coeff_r_w, coeff_l_w, coeff_r_w, coeff_l_w,                  \
+              sq_coeff_r, sq_coeff_l);                                     \
+  DOTP_SW2_SD(diff_r, diff_l, diff_r, diff_l, err0, err1);                 \
+                                                                           \
+  coeff = LD_SH(coeff_ptr + 8);                                            \
+  dq_coeff = LD_SH(dq_coeff_ptr + 8);                                      \
+  UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                                \
+  ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                      \
+  HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                       \
+  DPADD_SD2_SD(coeff_r_w, coeff_l_w, sq_coeff_r, sq_coeff_l);              \
+  DPADD_SD2_SD(diff_r, diff_l, err0, err1);                                \
+                                                                           \
+  coeff_ptr += 16;                                                         \
+  dq_coeff_ptr += 16;                                                      \
+                                                                           \
+  for (loop_cnt = ((BSize >> 4) - 1); loop_cnt--;) {                       \
+    coeff = LD_SH(coeff_ptr);                                              \
+    dq_coeff = LD_SH(dq_coeff_ptr);                                        \
+    UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                              \
+    ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                    \
+    HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                     \
+    DPADD_SD2_SD(coeff_r_w, coeff_l_w, sq_coeff_r, sq_coeff_l);            \
+    DPADD_SD2_SD(diff_r, diff_l, err0, err1);                              \
+                                                                           \
+    coeff = LD_SH(coeff_ptr + 8);                                          \
+    dq_coeff = LD_SH(dq_coeff_ptr + 8);                                    \
+    UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                              \
+    ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                    \
+    HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                     \
+    DPADD_SD2_SD(coeff_r_w, coeff_l_w, sq_coeff_r, sq_coeff_l);            \
+    DPADD_SD2_SD(diff_r, diff_l, err0, err1);                              \
+                                                                           \
+    coeff_ptr += 16;                                                       \
+    dq_coeff_ptr += 16;                                                    \
+  }                                                                        \
+                                                                           \
+  err_dup0 = __msa_splati_d(sq_coeff_r, 1);                                \
+  err_dup1 = __msa_splati_d(sq_coeff_l, 1);                                \
+  sq_coeff_r += err_dup0;                                                  \
+  sq_coeff_l += err_dup1;                                                  \
+  *ssz = __msa_copy_s_d(sq_coeff_r, 0);                                    \
+  *ssz += __msa_copy_s_d(sq_coeff_l, 0);                                   \
+                                                                           \
+  err_dup0 = __msa_splati_d(err0, 1);                                      \
+  err_dup1 = __msa_splati_d(err1, 1);                                      \
+  err0 += err_dup0;                                                        \
+  err1 += err_dup1;                                                        \
+  err = __msa_copy_s_d(err0, 0);                                           \
+  err += __msa_copy_s_d(err1, 0);                                          \
+                                                                           \
+  return err;                                                              \
+}
+
+BLOCK_ERROR_BLOCKSIZE_MSA(16);
+BLOCK_ERROR_BLOCKSIZE_MSA(64);
+BLOCK_ERROR_BLOCKSIZE_MSA(256);
+BLOCK_ERROR_BLOCKSIZE_MSA(1024);
+
+int64_t vp9_block_error_msa(const tran_low_t *coeff_ptr,
+                            const tran_low_t *dq_coeff_ptr,
+                            intptr_t blk_size, int64_t *ssz) {
+  int64_t err;
+  const int16_t *coeff = (const int16_t *)coeff_ptr;
+  const int16_t *dq_coeff = (const int16_t *)dq_coeff_ptr;
+
+  switch (blk_size) {
+    case 16:
+      err = block_error_16size_msa(coeff, dq_coeff, ssz);
+      break;
+    case 64:
+      err = block_error_64size_msa(coeff, dq_coeff, ssz);
+      break;
+    case 256:
+      err = block_error_256size_msa(coeff, dq_coeff, ssz);
+      break;
+    case 1024:
+      err = block_error_1024size_msa(coeff, dq_coeff, ssz);
+      break;
+    default:
+      err = vp9_block_error_c(coeff_ptr, dq_coeff_ptr, blk_size, ssz);
+      break;
+  }
+
+  return err;
+}
diff --git a/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct16x16_msa.c b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct16x16_msa.c
new file mode 100644
index 0000000..6dabb58
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct16x16_msa.c
@@ -0,0 +1,507 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_enums.h"
+#include "vp9/encoder/mips/msa/vp9_fdct_msa.h"
+#include "vpx_dsp/mips/fwd_txfm_msa.h"
+
+static void fadst16_cols_step1_msa(const int16_t *input, int32_t stride,
+                                   const int32_t *const0, int16_t *int_buf) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 tp0, tp1, tp2, tp3, g0, g1, g2, g3, g8, g9, g10, g11, h0, h1, h2, h3;
+  v4i32 k0, k1, k2, k3;
+
+  /* load input data */
+  r0 = LD_SH(input);
+  r15 = LD_SH(input + 15 * stride);
+  r7 = LD_SH(input + 7 * stride);
+  r8 = LD_SH(input + 8 * stride);
+  SLLI_4V(r0, r15, r7, r8, 2);
+
+  /* stage 1 */
+  LD_SW2(const0, 4, k0, k1);
+  LD_SW2(const0 + 8, 4, k2, k3);
+  MADD_BF(r15, r0, r7, r8, k0, k1, k2, k3, g0, g1, g2, g3);
+
+  r3 = LD_SH(input + 3 * stride);
+  r4 = LD_SH(input + 4 * stride);
+  r11 = LD_SH(input + 11 * stride);
+  r12 = LD_SH(input + 12 * stride);
+  SLLI_4V(r3, r4, r11, r12, 2);
+
+  LD_SW2(const0 + 4 * 4, 4, k0, k1);
+  LD_SW2(const0 + 4 * 6, 4, k2, k3);
+  MADD_BF(r11, r4, r3, r12, k0, k1, k2, k3, g8, g9, g10, g11);
+
+  /* stage 2 */
+  BUTTERFLY_4(g0, g2, g10, g8, tp0, tp2, tp3, tp1);
+  ST_SH2(tp0, tp2, int_buf, 8);
+  ST_SH2(tp1, tp3, int_buf + 4 * 8, 8);
+
+  LD_SW2(const0 + 4 * 8, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 10);
+  MADD_BF(g1, g3, g9, g11, k0, k1, k2, k0, h0, h1, h2, h3);
+
+  ST_SH2(h0, h1, int_buf + 8 * 8, 8);
+  ST_SH2(h3, h2, int_buf + 12 * 8, 8);
+
+  r9 = LD_SH(input + 9 * stride);
+  r6 = LD_SH(input + 6 * stride);
+  r1 = LD_SH(input + stride);
+  r14 = LD_SH(input + 14 * stride);
+  SLLI_4V(r9, r6, r1, r14, 2);
+
+  LD_SW2(const0 + 4 * 11, 4, k0, k1);
+  LD_SW2(const0 + 4 * 13, 4, k2, k3);
+  MADD_BF(r9, r6, r1, r14, k0, k1, k2, k3, g0, g1, g2, g3);
+
+  ST_SH2(g1, g3, int_buf + 3 * 8, 4 * 8);
+
+  r13 = LD_SH(input + 13 * stride);
+  r2 = LD_SH(input + 2 * stride);
+  r5 = LD_SH(input + 5 * stride);
+  r10 = LD_SH(input + 10 * stride);
+  SLLI_4V(r13, r2, r5, r10, 2);
+
+  LD_SW2(const0 + 4 * 15, 4, k0, k1);
+  LD_SW2(const0 + 4 * 17, 4, k2, k3);
+  MADD_BF(r13, r2, r5, r10, k0, k1, k2, k3, h0, h1, h2, h3);
+
+  ST_SH2(h1, h3, int_buf + 11 * 8, 4 * 8);
+
+  BUTTERFLY_4(h0, h2, g2, g0, tp0, tp1, tp2, tp3);
+  ST_SH4(tp0, tp1, tp2, tp3, int_buf + 2 * 8, 4 * 8);
+}
+
+static void fadst16_cols_step2_msa(int16_t *int_buf, const int32_t *const0,
+                                   int16_t *out) {
+  int16_t *out_ptr = out + 128;
+  v8i16 tp0, tp1, tp2, tp3, g5, g7, g13, g15;
+  v8i16 h0, h1, h2, h3, h4, h5, h6, h7, h10, h11;
+  v8i16 out0, out1, out2, out3, out4, out5, out6, out7;
+  v8i16 out8, out9, out10, out11, out12, out13, out14, out15;
+  v4i32 k0, k1, k2, k3;
+
+  LD_SH2(int_buf + 3 * 8, 4 * 8, g13, g15);
+  LD_SH2(int_buf + 11 * 8, 4 * 8, g5, g7);
+  LD_SW2(const0 + 4 * 19, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 21);
+  MADD_BF(g7, g5, g15, g13, k0, k1, k2, k0, h4, h5, h6, h7);
+
+  tp0 = LD_SH(int_buf + 4 * 8);
+  tp1 = LD_SH(int_buf + 5 * 8);
+  tp3 = LD_SH(int_buf + 10 * 8);
+  tp2 = LD_SH(int_buf + 14 * 8);
+  LD_SW2(const0 + 4 * 22, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 24);
+  MADD_BF(tp0, tp1, tp2, tp3, k0, k1, k2, k0, out4, out6, out5, out7);
+  out4 = -out4;
+  ST_SH(out4, (out + 3 * 16));
+  ST_SH(out5, (out_ptr + 4 * 16));
+
+  h1 = LD_SH(int_buf + 9 * 8);
+  h3 = LD_SH(int_buf + 12 * 8);
+  MADD_BF(h1, h3, h5, h7, k0, k1, k2, k0, out12, out14, out13, out15);
+  out13 = -out13;
+  ST_SH(out12, (out + 2 * 16));
+  ST_SH(out13, (out_ptr + 5 * 16));
+
+  tp0 = LD_SH(int_buf);
+  tp1 = LD_SH(int_buf + 8);
+  tp2 = LD_SH(int_buf + 2 * 8);
+  tp3 = LD_SH(int_buf + 6 * 8);
+
+  BUTTERFLY_4(tp0, tp1, tp3, tp2, out0, out1, h11, h10);
+  out1 = -out1;
+  ST_SH(out0, (out));
+  ST_SH(out1, (out_ptr + 7 * 16));
+
+  h0 = LD_SH(int_buf + 8 * 8);
+  h2 = LD_SH(int_buf + 13 * 8);
+
+  BUTTERFLY_4(h0, h2, h6, h4, out8, out9, out11, out10);
+  out8 = -out8;
+  ST_SH(out8, (out + 16));
+  ST_SH(out9, (out_ptr + 6 * 16));
+
+  /* stage 4 */
+  LD_SW2(const0 + 4 * 25, 4, k0, k1);
+  LD_SW2(const0 + 4 * 27, 4, k2, k3);
+  MADD_SHORT(h10, h11, k1, k2, out2, out3);
+  ST_SH(out2, (out + 7 * 16));
+  ST_SH(out3, (out_ptr));
+
+  MADD_SHORT(out6, out7, k0, k3, out6, out7);
+  ST_SH(out6, (out + 4 * 16));
+  ST_SH(out7, (out_ptr + 3 * 16));
+
+  MADD_SHORT(out10, out11, k0, k3, out10, out11);
+  ST_SH(out10, (out + 6 * 16));
+  ST_SH(out11, (out_ptr + 16));
+
+  MADD_SHORT(out14, out15, k1, k2, out14, out15);
+  ST_SH(out14, (out + 5 * 16));
+  ST_SH(out15, (out_ptr + 2 * 16));
+}
+
+static void fadst16_transpose_postproc_msa(int16_t *input, int16_t *out) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 l0, l1, l2, l3, l4, l5, l6, l7, l8, l9, l10, l11, l12, l13, l14, l15;
+
+  /* load input data */
+  LD_SH8(input, 16, l0, l1, l2, l3, l4, l5, l6, l7);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7,
+                     r0, r1, r2, r3, r4, r5, r6, r7);
+  FDCT_POSTPROC_2V_NEG_H(r0, r1);
+  FDCT_POSTPROC_2V_NEG_H(r2, r3);
+  FDCT_POSTPROC_2V_NEG_H(r4, r5);
+  FDCT_POSTPROC_2V_NEG_H(r6, r7);
+  ST_SH8(r0, r1, r2, r3, r4, r5, r6, r7, out, 8);
+  out += 64;
+
+  LD_SH8(input + 8, 16, l8, l9, l10, l11, l12, l13, l14, l15);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15,
+                     r8, r9, r10, r11, r12, r13, r14, r15);
+  FDCT_POSTPROC_2V_NEG_H(r8, r9);
+  FDCT_POSTPROC_2V_NEG_H(r10, r11);
+  FDCT_POSTPROC_2V_NEG_H(r12, r13);
+  FDCT_POSTPROC_2V_NEG_H(r14, r15);
+  ST_SH8(r8, r9, r10, r11, r12, r13, r14, r15, out, 8);
+  out += 64;
+
+  /* load input data */
+  input += 128;
+  LD_SH8(input, 16, l0, l1, l2, l3, l4, l5, l6, l7);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7,
+                     r0, r1, r2, r3, r4, r5, r6, r7);
+  FDCT_POSTPROC_2V_NEG_H(r0, r1);
+  FDCT_POSTPROC_2V_NEG_H(r2, r3);
+  FDCT_POSTPROC_2V_NEG_H(r4, r5);
+  FDCT_POSTPROC_2V_NEG_H(r6, r7);
+  ST_SH8(r0, r1, r2, r3, r4, r5, r6, r7, out, 8);
+  out += 64;
+
+  LD_SH8(input + 8, 16, l8, l9, l10, l11, l12, l13, l14, l15);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15,
+                     r8, r9, r10, r11, r12, r13, r14, r15);
+  FDCT_POSTPROC_2V_NEG_H(r8, r9);
+  FDCT_POSTPROC_2V_NEG_H(r10, r11);
+  FDCT_POSTPROC_2V_NEG_H(r12, r13);
+  FDCT_POSTPROC_2V_NEG_H(r14, r15);
+  ST_SH8(r8, r9, r10, r11, r12, r13, r14, r15, out, 8);
+}
+
+static void fadst16_rows_step1_msa(int16_t *input, const int32_t *const0,
+                                   int16_t *int_buf) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 tp0, tp1, tp2, tp3, g0, g1, g2, g3, g8, g9, g10, g11, h0, h1, h2, h3;
+  v4i32 k0, k1, k2, k3;
+
+  /* load input data */
+  r0 = LD_SH(input);
+  r7 = LD_SH(input + 7 * 8);
+  r8 = LD_SH(input + 8 * 8);
+  r15 = LD_SH(input + 15 * 8);
+
+  /* stage 1 */
+  LD_SW2(const0, 4, k0, k1);
+  LD_SW2(const0 + 4 * 2, 4, k2, k3);
+  MADD_BF(r15, r0, r7, r8, k0, k1, k2, k3, g0, g1, g2, g3);
+
+  r3 = LD_SH(input + 3 * 8);
+  r4 = LD_SH(input + 4 * 8);
+  r11 = LD_SH(input + 11 * 8);
+  r12 = LD_SH(input + 12 * 8);
+
+  LD_SW2(const0 + 4 * 4, 4, k0, k1);
+  LD_SW2(const0 + 4 * 6, 4, k2, k3);
+  MADD_BF(r11, r4, r3, r12, k0, k1, k2, k3, g8, g9, g10, g11);
+
+  /* stage 2 */
+  BUTTERFLY_4(g0, g2, g10, g8, tp0, tp2, tp3, tp1);
+  ST_SH2(tp0, tp1, int_buf, 4 * 8);
+  ST_SH2(tp2, tp3, int_buf + 8, 4 * 8);
+
+  LD_SW2(const0 + 4 * 8, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 10);
+  MADD_BF(g1, g3, g9, g11, k0, k1, k2, k0, h0, h1, h2, h3);
+  ST_SH2(h0, h3, int_buf + 8 * 8, 4 * 8);
+  ST_SH2(h1, h2, int_buf + 9 * 8, 4 * 8);
+
+  r1 = LD_SH(input + 8);
+  r6 = LD_SH(input + 6 * 8);
+  r9 = LD_SH(input + 9 * 8);
+  r14 = LD_SH(input + 14 * 8);
+
+  LD_SW2(const0 + 4 * 11, 4, k0, k1);
+  LD_SW2(const0 + 4 * 13, 4, k2, k3);
+  MADD_BF(r9, r6, r1, r14, k0, k1, k2, k3, g0, g1, g2, g3);
+  ST_SH2(g1, g3, int_buf + 3 * 8, 4 * 8);
+
+  r2 = LD_SH(input + 2 * 8);
+  r5 = LD_SH(input + 5 * 8);
+  r10 = LD_SH(input + 10 * 8);
+  r13 = LD_SH(input + 13 * 8);
+
+  LD_SW2(const0 + 4 * 15, 4, k0, k1);
+  LD_SW2(const0 + 4 * 17, 4, k2, k3);
+  MADD_BF(r13, r2, r5, r10, k0, k1, k2, k3, h0, h1, h2, h3);
+  ST_SH2(h1, h3, int_buf + 11 * 8, 4 * 8);
+  BUTTERFLY_4(h0, h2, g2, g0, tp0, tp1, tp2, tp3);
+  ST_SH4(tp0, tp1, tp2, tp3, int_buf + 2 * 8, 4 * 8);
+}
+
+static void fadst16_rows_step2_msa(int16_t *int_buf, const int32_t *const0,
+                                   int16_t *out) {
+  int16_t *out_ptr = out + 8;
+  v8i16 tp0, tp1, tp2, tp3, g5, g7, g13, g15;
+  v8i16 h0, h1, h2, h3, h4, h5, h6, h7, h10, h11;
+  v8i16 out0, out1, out2, out3, out4, out5, out6, out7;
+  v8i16 out8, out9, out10, out11, out12, out13, out14, out15;
+  v4i32 k0, k1, k2, k3;
+
+  g13 = LD_SH(int_buf + 3 * 8);
+  g15 = LD_SH(int_buf + 7 * 8);
+  g5 = LD_SH(int_buf + 11 * 8);
+  g7 = LD_SH(int_buf + 15 * 8);
+
+  LD_SW2(const0 + 4 * 19, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 21);
+  MADD_BF(g7, g5, g15, g13, k0, k1, k2, k0, h4, h5, h6, h7);
+
+  tp0 = LD_SH(int_buf + 4 * 8);
+  tp1 = LD_SH(int_buf + 5 * 8);
+  tp3 = LD_SH(int_buf + 10 * 8);
+  tp2 = LD_SH(int_buf + 14 * 8);
+
+  LD_SW2(const0 + 4 * 22, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 24);
+  MADD_BF(tp0, tp1, tp2, tp3, k0, k1, k2, k0, out4, out6, out5, out7);
+  out4 = -out4;
+  ST_SH(out4, (out + 3 * 16));
+  ST_SH(out5, (out_ptr + 4 * 16));
+
+  h1 = LD_SH(int_buf + 9 * 8);
+  h3 = LD_SH(int_buf + 12 * 8);
+  MADD_BF(h1, h3, h5, h7, k0, k1, k2, k0, out12, out14, out13, out15);
+  out13 = -out13;
+  ST_SH(out12, (out + 2 * 16));
+  ST_SH(out13, (out_ptr + 5 * 16));
+
+  tp0 = LD_SH(int_buf);
+  tp1 = LD_SH(int_buf + 8);
+  tp2 = LD_SH(int_buf + 2 * 8);
+  tp3 = LD_SH(int_buf + 6 * 8);
+
+  BUTTERFLY_4(tp0, tp1, tp3, tp2, out0, out1, h11, h10);
+  out1 = -out1;
+  ST_SH(out0, (out));
+  ST_SH(out1, (out_ptr + 7 * 16));
+
+  h0 = LD_SH(int_buf + 8 * 8);
+  h2 = LD_SH(int_buf + 13 * 8);
+  BUTTERFLY_4(h0, h2, h6, h4, out8, out9, out11, out10);
+  out8 = -out8;
+  ST_SH(out8, (out + 16));
+  ST_SH(out9, (out_ptr + 6 * 16));
+
+  /* stage 4 */
+  LD_SW2(const0 + 4 * 25, 4, k0, k1);
+  LD_SW2(const0 + 4 * 27, 4, k2, k3);
+  MADD_SHORT(h10, h11, k1, k2, out2, out3);
+  ST_SH(out2, (out + 7 * 16));
+  ST_SH(out3, (out_ptr));
+
+  MADD_SHORT(out6, out7, k0, k3, out6, out7);
+  ST_SH(out6, (out + 4 * 16));
+  ST_SH(out7, (out_ptr + 3 * 16));
+
+  MADD_SHORT(out10, out11, k0, k3, out10, out11);
+  ST_SH(out10, (out + 6 * 16));
+  ST_SH(out11, (out_ptr + 16));
+
+  MADD_SHORT(out14, out15, k1, k2, out14, out15);
+  ST_SH(out14, (out + 5 * 16));
+  ST_SH(out15, (out_ptr + 2 * 16));
+}
+
+static void fadst16_transpose_msa(int16_t *input, int16_t *out) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 l0, l1, l2, l3, l4, l5, l6, l7, l8, l9, l10, l11, l12, l13, l14, l15;
+
+  /* load input data */
+  LD_SH16(input, 8, l0, l8, l1, l9, l2, l10, l3, l11,
+          l4, l12, l5, l13, l6, l14, l7, l15);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7,
+                     r0, r1, r2, r3, r4, r5, r6, r7);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15,
+                     r8, r9, r10, r11, r12, r13, r14, r15);
+  ST_SH8(r0, r8, r1, r9, r2, r10, r3, r11, out, 8);
+  ST_SH8(r4, r12, r5, r13, r6, r14, r7, r15, (out + 64), 8);
+  out += 16 * 8;
+
+  /* load input data */
+  input += 128;
+  LD_SH16(input, 8, l0, l8, l1, l9, l2, l10, l3, l11,
+          l4, l12, l5, l13, l6, l14, l7, l15);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7,
+                     r0, r1, r2, r3, r4, r5, r6, r7);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15,
+                     r8, r9, r10, r11, r12, r13, r14, r15);
+  ST_SH8(r0, r8, r1, r9, r2, r10, r3, r11, out, 8);
+  ST_SH8(r4, r12, r5, r13, r6, r14, r7, r15, (out + 64), 8);
+}
+
+static void postproc_fdct16x8_1d_row(int16_t *intermediate, int16_t *output) {
+  int16_t *temp = intermediate;
+  int16_t *out = output;
+  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11;
+  v8i16 in12, in13, in14, in15;
+
+  LD_SH8(temp, 16, in0, in1, in2, in3, in4, in5, in6, in7);
+  temp = intermediate + 8;
+  LD_SH8(temp, 16, in8, in9, in10, in11, in12, in13, in14, in15);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15,
+                     in8, in9, in10, in11, in12, in13, in14, in15);
+  FDCT_POSTPROC_2V_NEG_H(in0, in1);
+  FDCT_POSTPROC_2V_NEG_H(in2, in3);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  FDCT_POSTPROC_2V_NEG_H(in6, in7);
+  FDCT_POSTPROC_2V_NEG_H(in8, in9);
+  FDCT_POSTPROC_2V_NEG_H(in10, in11);
+  FDCT_POSTPROC_2V_NEG_H(in12, in13);
+  FDCT_POSTPROC_2V_NEG_H(in14, in15);
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+               in8, in9, in10, in11, in12, in13, in14, in15,
+               tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7,
+               in8, in9, in10, in11, in12, in13, in14, in15);
+  temp = intermediate;
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, temp, 16);
+  FDCT8x16_EVEN(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7,
+                tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7);
+  temp = intermediate;
+  LD_SH8(temp, 16, in8, in9, in10, in11, in12, in13, in14, in15);
+  FDCT8x16_ODD(in8, in9, in10, in11, in12, in13, in14, in15,
+               in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3,
+                     tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3);
+  ST_SH8(tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3, out, 16);
+  TRANSPOSE8x8_SH_SH(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7,
+                     tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7);
+  out = output + 8;
+  ST_SH8(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7, out, 16);
+}
+
+void vp9_fht16x16_msa(const int16_t *input, int16_t *output,
+                      int32_t stride, int32_t tx_type) {
+  DECLARE_ALIGNED(32, int16_t, tmp[256]);
+  DECLARE_ALIGNED(32, int16_t, trans_buf[256]);
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[128]);
+  int32_t i;
+  int16_t *ptmpbuf = &tmp_buf[0];
+  int16_t *trans = &trans_buf[0];
+  const int32_t const_arr[29 * 4] = {
+    52707308, 52707308, 52707308, 52707308,
+    -1072430300, -1072430300, -1072430300, -1072430300,
+    795618043, 795618043, 795618043, 795618043,
+    -721080468, -721080468, -721080468, -721080468,
+    459094491, 459094491, 459094491, 459094491,
+    -970646691, -970646691, -970646691, -970646691,
+    1010963856, 1010963856, 1010963856, 1010963856,
+    -361743294, -361743294, -361743294, -361743294,
+    209469125, 209469125, 209469125, 209469125,
+    -1053094788, -1053094788, -1053094788, -1053094788,
+    1053160324, 1053160324, 1053160324, 1053160324,
+    639644520, 639644520, 639644520, 639644520,
+    -862444000, -862444000, -862444000, -862444000,
+    1062144356, 1062144356, 1062144356, 1062144356,
+    -157532337, -157532337, -157532337, -157532337,
+    260914709, 260914709, 260914709, 260914709,
+    -1041559667, -1041559667, -1041559667, -1041559667,
+    920985831, 920985831, 920985831, 920985831,
+    -551995675, -551995675, -551995675, -551995675,
+    596522295, 596522295, 596522295, 596522295,
+    892853362, 892853362, 892853362, 892853362,
+    -892787826, -892787826, -892787826, -892787826,
+    410925857, 410925857, 410925857, 410925857,
+    -992012162, -992012162, -992012162, -992012162,
+    992077698, 992077698, 992077698, 992077698,
+    759246145, 759246145, 759246145, 759246145,
+    -759180609, -759180609, -759180609, -759180609,
+    -759222975, -759222975, -759222975, -759222975,
+    759288511, 759288511, 759288511, 759288511 };
+
+  switch (tx_type) {
+    case DCT_DCT:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fdct8x16_1d_column(input + 8 * i, tmp + 8 * i, stride);
+      }
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        fdct16x8_1d_row(tmp + (128 * i), output + (128 * i));
+      }
+      break;
+    case ADST_DCT:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_cols_step1_msa(input + (i << 3), stride, const_arr, ptmpbuf);
+        fadst16_cols_step2_msa(ptmpbuf, const_arr, tmp + (i << 3));
+      }
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        postproc_fdct16x8_1d_row(tmp + (128 * i), output + (128 * i));
+      }
+      break;
+    case DCT_ADST:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fdct8x16_1d_column(input + 8 * i, tmp + 8 * i, stride);
+      }
+
+      fadst16_transpose_postproc_msa(tmp, trans);
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_rows_step1_msa(trans + (i << 7), const_arr, ptmpbuf);
+        fadst16_rows_step2_msa(ptmpbuf, const_arr, tmp + (i << 7));
+      }
+
+      fadst16_transpose_msa(tmp, output);
+      break;
+    case ADST_ADST:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_cols_step1_msa(input + (i << 3), stride, const_arr, ptmpbuf);
+        fadst16_cols_step2_msa(ptmpbuf, const_arr, tmp + (i << 3));
+      }
+
+      fadst16_transpose_postproc_msa(tmp, trans);
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_rows_step1_msa(trans + (i << 7), const_arr, ptmpbuf);
+        fadst16_rows_step2_msa(ptmpbuf, const_arr, tmp + (i << 7));
+      }
+
+      fadst16_transpose_msa(tmp, output);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct4x4_msa.c b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct4x4_msa.c
new file mode 100644
index 0000000..574016f
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct4x4_msa.c
@@ -0,0 +1,99 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_enums.h"
+#include "vp9/encoder/mips/msa/vp9_fdct_msa.h"
+
+void vp9_fwht4x4_msa(const int16_t *input, int16_t *output,
+                     int32_t src_stride) {
+  v8i16 in0, in1, in2, in3, in4;
+
+  LD_SH4(input, src_stride, in0, in1, in2, in3);
+
+  in0 += in1;
+  in3 -= in2;
+  in4 = (in0 - in3) >> 1;
+  SUB2(in4, in1, in4, in2, in1, in2);
+  in0 -= in2;
+  in3 += in1;
+
+  TRANSPOSE4x4_SH_SH(in0, in2, in3, in1, in0, in2, in3, in1);
+
+  in0 += in2;
+  in1 -= in3;
+  in4 = (in0 - in1) >> 1;
+  SUB2(in4, in2, in4, in3, in2, in3);
+  in0 -= in3;
+  in1 += in2;
+
+  SLLI_4V(in0, in1, in2, in3, 2);
+
+  TRANSPOSE4x4_SH_SH(in0, in3, in1, in2, in0, in3, in1, in2);
+
+  ST4x2_UB(in0, output, 4);
+  ST4x2_UB(in3, output + 4, 4);
+  ST4x2_UB(in1, output + 8, 4);
+  ST4x2_UB(in2, output + 12, 4);
+}
+
+void vp9_fht4x4_msa(const int16_t *input, int16_t *output, int32_t stride,
+                    int32_t tx_type) {
+  v8i16 in0, in1, in2, in3;
+
+  LD_SH4(input, stride, in0, in1, in2, in3);
+
+  /* fdct4 pre-process */
+  {
+    v8i16 temp, mask;
+    v16i8 zero = { 0 };
+    v16i8 one = __msa_ldi_b(1);
+
+    mask = (v8i16)__msa_sldi_b(zero, one, 15);
+    SLLI_4V(in0, in1, in2, in3, 4);
+    temp = __msa_ceqi_h(in0, 0);
+    temp = (v8i16)__msa_xori_b((v16u8)temp, 255);
+    temp = mask & temp;
+    in0 += temp;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+      VP9_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      VP9_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case ADST_DCT:
+      VP9_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      VP9_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case DCT_ADST:
+      VP9_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      VP9_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case ADST_ADST:
+      VP9_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      VP9_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  ADD4(in0, 1, in1, 1, in2, 1, in3, 1, in0, in1, in2, in3);
+  SRA_4V(in0, in1, in2, in3, 2);
+  PCKEV_D2_SH(in1, in0, in3, in2, in0, in2);
+  ST_SH2(in0, in2, output, 8);
+}
diff --git a/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct8x8_msa.c b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct8x8_msa.c
new file mode 100644
index 0000000..7c3c635
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct8x8_msa.c
@@ -0,0 +1,66 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_enums.h"
+#include "vp9/encoder/mips/msa/vp9_fdct_msa.h"
+
+void vp9_fht8x8_msa(const int16_t *input, int16_t *output, int32_t stride,
+                    int32_t tx_type) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  LD_SH8(input, stride, in0, in1, in2, in3, in4, in5, in6, in7);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    case ADST_DCT:
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    case DCT_ADST:
+      VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    case ADST_ADST:
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7);
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output, 8);
+}
diff --git a/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct_msa.h b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct_msa.h
new file mode 100644
index 0000000..d7d40cb
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_fdct_msa.h
@@ -0,0 +1,117 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_MIPS_MSA_VP9_FDCT_MSA_H_
+#define VP9_ENCODER_MIPS_MSA_VP9_FDCT_MSA_H_
+
+#include "vpx_dsp/mips/fwd_txfm_msa.h"
+#include "vpx_dsp/mips/txfm_macros_msa.h"
+#include "vpx_ports/mem.h"
+
+#define VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,                   \
+                  out0, out1, out2, out3, out4, out5, out6, out7) {         \
+  v8i16 cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst4_m;                        \
+  v8i16 vec0_m, vec1_m, vec2_m, vec3_m, s0_m, s1_m;                         \
+  v8i16 coeff0_m = { cospi_2_64, cospi_6_64, cospi_10_64, cospi_14_64,      \
+                     cospi_18_64, cospi_22_64, cospi_26_64, cospi_30_64 };  \
+  v8i16 coeff1_m = { cospi_8_64, -cospi_8_64, cospi_16_64, -cospi_16_64,    \
+                     cospi_24_64, -cospi_24_64, 0, 0 };                     \
+                                                                            \
+  SPLATI_H2_SH(coeff0_m, 0, 7, cnst0_m, cnst1_m);                           \
+  cnst2_m = -cnst0_m;                                                       \
+  ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);        \
+  SPLATI_H2_SH(coeff0_m, 4, 3, cnst2_m, cnst3_m);                           \
+  cnst4_m = -cnst2_m;                                                       \
+  ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);        \
+                                                                            \
+  ILVRL_H2_SH(in0, in7, vec1_m, vec0_m);                                    \
+  ILVRL_H2_SH(in4, in3, vec3_m, vec2_m);                                    \
+  DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m,            \
+                        cnst1_m, cnst2_m, cnst3_m, in7, in0,                \
+                        in4, in3);                                          \
+                                                                            \
+  SPLATI_H2_SH(coeff0_m, 2, 5, cnst0_m, cnst1_m);                           \
+  cnst2_m = -cnst0_m;                                                       \
+  ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);        \
+  SPLATI_H2_SH(coeff0_m, 6, 1, cnst2_m, cnst3_m);                           \
+  cnst4_m = -cnst2_m;                                                       \
+  ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);        \
+                                                                            \
+  ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                    \
+  ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                    \
+                                                                            \
+  DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m,            \
+                        cnst1_m, cnst2_m, cnst3_m, in5, in2,                \
+                        in6, in1);                                          \
+  BUTTERFLY_4(in7, in0, in2, in5, s1_m, s0_m, in2, in5);                    \
+  out7 = -s0_m;                                                             \
+  out0 = s1_m;                                                              \
+                                                                            \
+  SPLATI_H4_SH(coeff1_m, 0, 4, 1, 5, cnst0_m, cnst1_m, cnst2_m, cnst3_m);   \
+                                                                            \
+  ILVEV_H2_SH(cnst3_m, cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst2_m);        \
+  cnst0_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                \
+  cnst1_m = cnst0_m;                                                        \
+                                                                            \
+  ILVRL_H2_SH(in4, in3, vec1_m, vec0_m);                                    \
+  ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                    \
+  DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m,            \
+                        cnst2_m, cnst3_m, cnst1_m, out1, out6,              \
+                        s0_m, s1_m);                                        \
+                                                                            \
+  SPLATI_H2_SH(coeff1_m, 2, 3, cnst0_m, cnst1_m);                           \
+  cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                \
+                                                                            \
+  ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                    \
+  ILVRL_H2_SH(s0_m, s1_m, vec3_m, vec2_m);                                  \
+  out3 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                    \
+  out4 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst1_m);                    \
+  out2 = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst0_m);                    \
+  out5 = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst1_m);                    \
+                                                                            \
+  out1 = -out1;                                                             \
+  out3 = -out3;                                                             \
+  out5 = -out5;                                                             \
+}
+
+#define VP9_FADST4(in0, in1, in2, in3, out0, out1, out2, out3) {  \
+  v4i32 s0_m, s1_m, s2_m, s3_m, constant_m;                       \
+  v4i32 in0_r_m, in1_r_m, in2_r_m, in3_r_m;                       \
+                                                                  \
+  UNPCK_R_SH_SW(in0, in0_r_m);                                    \
+  UNPCK_R_SH_SW(in1, in1_r_m);                                    \
+  UNPCK_R_SH_SW(in2, in2_r_m);                                    \
+  UNPCK_R_SH_SW(in3, in3_r_m);                                    \
+                                                                  \
+  constant_m = __msa_fill_w(sinpi_4_9);                           \
+  MUL2(in0_r_m, constant_m, in3_r_m, constant_m, s1_m, s0_m);     \
+                                                                  \
+  constant_m = __msa_fill_w(sinpi_1_9);                           \
+  s0_m += in0_r_m * constant_m;                                   \
+  s1_m -= in1_r_m * constant_m;                                   \
+                                                                  \
+  constant_m = __msa_fill_w(sinpi_2_9);                           \
+  s0_m += in1_r_m * constant_m;                                   \
+  s1_m += in3_r_m * constant_m;                                   \
+                                                                  \
+  s2_m = in0_r_m + in1_r_m - in3_r_m;                             \
+                                                                  \
+  constant_m = __msa_fill_w(sinpi_3_9);                           \
+  MUL2(in2_r_m, constant_m, s2_m, constant_m, s3_m, in1_r_m);     \
+                                                                  \
+  in0_r_m = s0_m + s3_m;                                          \
+  s2_m = s1_m - s3_m;                                             \
+  s3_m = s1_m - s0_m + s3_m;                                      \
+                                                                  \
+  SRARI_W4_SW(in0_r_m, in1_r_m, s2_m, s3_m, DCT_CONST_BITS);      \
+  PCKEV_H4_SH(in0_r_m, in0_r_m, in1_r_m, in1_r_m, s2_m, s2_m,     \
+              s3_m, s3_m, out0, out1, out2, out3);                \
+}
+#endif  /* VP9_ENCODER_MIPS_MSA_VP9_FDCT_MSA_H_ */
diff --git a/libvpx/libvpx/vp9/encoder/mips/msa/vp9_temporal_filter_msa.c b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_temporal_filter_msa.c
new file mode 100644
index 0000000..363aabb
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/mips/msa/vp9_temporal_filter_msa.c
@@ -0,0 +1,289 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "vpx_dsp/mips/macros_msa.h"
+
+static void temporal_filter_apply_8size_msa(uint8_t *frm1_ptr,
+                                            uint32_t stride,
+                                            uint8_t *frm2_ptr,
+                                            int32_t filt_sth,
+                                            int32_t filt_wgt,
+                                            uint32_t *acc,
+                                            uint16_t *cnt) {
+  uint32_t row;
+  uint64_t f0, f1, f2, f3;
+  v16i8 frm2, frm1 = { 0 };
+  v16i8 frm4, frm3 = { 0 };
+  v16u8 frm_r, frm_l;
+  v8i16 frm2_r, frm2_l;
+  v8i16 diff0, diff1, mod0_h, mod1_h;
+  v4i32 cnst3, cnst16, filt_wt, strength;
+  v4i32 mod0_w, mod1_w, mod2_w, mod3_w;
+  v4i32 diff0_r, diff0_l, diff1_r, diff1_l;
+  v4i32 frm2_rr, frm2_rl, frm2_lr, frm2_ll;
+  v4i32 acc0, acc1, acc2, acc3;
+  v8i16 cnt0, cnt1;
+
+  filt_wt = __msa_fill_w(filt_wgt);
+  strength = __msa_fill_w(filt_sth);
+  cnst3 = __msa_ldi_w(3);
+  cnst16 = __msa_ldi_w(16);
+
+  for (row = 2; row--;) {
+    LD4(frm1_ptr, stride, f0, f1, f2, f3);
+    frm1_ptr += (4 * stride);
+
+    LD_SB2(frm2_ptr, 16, frm2, frm4);
+    frm2_ptr += 32;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc + 8, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    INSERT_D2_SB(f0, f1, frm1);
+    INSERT_D2_SB(f2, f3, frm3);
+    ILVRL_B2_UB(frm1, frm2, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l,
+         diff1_l, mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+
+    UNPCK_UB_SH(frm2, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc + 8, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    ILVRL_B2_UB(frm3, frm4, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l,
+         diff1_l, mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+    UNPCK_UB_SH(frm4, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+  }
+}
+
+static void temporal_filter_apply_16size_msa(uint8_t *frm1_ptr,
+                                             uint32_t stride,
+                                             uint8_t *frm2_ptr,
+                                             int32_t filt_sth,
+                                             int32_t filt_wgt,
+                                             uint32_t *acc,
+                                             uint16_t *cnt) {
+  uint32_t row;
+  v16i8 frm1, frm2, frm3, frm4;
+  v16u8 frm_r, frm_l;
+  v16i8 zero = { 0 };
+  v8u16 frm2_r, frm2_l;
+  v8i16 diff0, diff1, mod0_h, mod1_h;
+  v4i32 cnst3, cnst16, filt_wt, strength;
+  v4i32 mod0_w, mod1_w, mod2_w, mod3_w;
+  v4i32 diff0_r, diff0_l, diff1_r, diff1_l;
+  v4i32 frm2_rr, frm2_rl, frm2_lr, frm2_ll;
+  v4i32 acc0, acc1, acc2, acc3;
+  v8i16 cnt0, cnt1;
+
+  filt_wt = __msa_fill_w(filt_wgt);
+  strength = __msa_fill_w(filt_sth);
+  cnst3 = __msa_ldi_w(3);
+  cnst16 = __msa_ldi_w(16);
+
+  for (row = 8; row--;) {
+    LD_SB2(frm1_ptr, stride, frm1, frm3);
+    frm1_ptr += stride;
+
+    LD_SB2(frm2_ptr, 16, frm2, frm4);
+    frm2_ptr += 16;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    ILVRL_B2_UB(frm1, frm2, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l, diff1_l,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+
+    ILVRL_B2_UH(zero, frm2, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc + 8, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    ILVRL_B2_UB(frm3, frm4, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l, diff1_l,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+
+    ILVRL_B2_UH(zero, frm4, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+
+    frm1_ptr += stride;
+    frm2_ptr += 16;
+  }
+}
+
+void vp9_temporal_filter_apply_msa(uint8_t *frame1_ptr, uint32_t stride,
+                                   uint8_t *frame2_ptr, uint32_t blk_w,
+                                   uint32_t blk_h, int32_t strength,
+                                   int32_t filt_wgt, uint32_t *accu,
+                                   uint16_t *cnt) {
+  if (8 == (blk_w * blk_h)) {
+    temporal_filter_apply_8size_msa(frame1_ptr, stride, frame2_ptr,
+                                    strength, filt_wgt, accu, cnt);
+  } else if (16 == (blk_w * blk_h)) {
+    temporal_filter_apply_16size_msa(frame1_ptr, stride, frame2_ptr,
+                                     strength, filt_wgt, accu, cnt);
+  } else {
+    vp9_temporal_filter_apply_c(frame1_ptr, stride, frame2_ptr, blk_w, blk_h,
+                                strength, filt_wgt, accu, cnt);
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_aq_360.c b/libvpx/libvpx/vp9/encoder/vp9_aq_360.c
new file mode 100644
index 0000000..7d411f6
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_aq_360.c
@@ -0,0 +1,74 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "vpx_ports/mem.h"
+#include "vpx_ports/system_state.h"
+
+#include "vp9/encoder/vp9_aq_360.h"
+#include "vp9/encoder/vp9_aq_variance.h"
+
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_segmentation.h"
+
+static const double rate_ratio[MAX_SEGMENTS] =
+  {1.0, 0.75, 0.6, 0.5, 0.4, 0.3, 0.25};
+
+// Sets segment id 0 for the equatorial region, 1 for temperate region
+// and 2 for the polar regions
+unsigned int vp9_360aq_segment_id(int mi_row, int mi_rows) {
+  if (mi_row < mi_rows / 8 || mi_row > mi_rows - mi_rows / 8)
+    return 2;
+  else if (mi_row < mi_rows / 4 || mi_row > mi_rows - mi_rows / 4)
+    return 1;
+  else
+    return 0;
+}
+
+void vp9_360aq_frame_setup(VP9_COMP *cpi) {
+  VP9_COMMON *cm = &cpi->common;
+  struct segmentation *seg = &cm->seg;
+  int i;
+
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
+    vp9_enable_segmentation(seg);
+    vp9_clearall_segfeatures(seg);
+
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    vpx_clear_system_state();
+
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      int qindex_delta =
+          vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type, cm->base_qindex,
+                                     rate_ratio[i], cm->bit_depth);
+
+      // We don't allow qindex 0 in a segment if the base value is not 0.
+      // Q index 0 (lossless) implies 4x4 encoding only and in AQ mode a segment
+      // Q delta is sometimes applied without going back around the rd loop.
+      // This could lead to an illegal combination of partition size and q.
+      if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
+        qindex_delta = -cm->base_qindex + 1;
+      }
+
+      // No need to enable SEG_LVL_ALT_Q for this segment.
+      if (rate_ratio[i] == 1.0) {
+        continue;
+      }
+
+      vp9_set_segdata(seg, i, SEG_LVL_ALT_Q, qindex_delta);
+      vp9_enable_segfeature(seg, i, SEG_LVL_ALT_Q);
+    }
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_aq_360.h b/libvpx/libvpx/vp9/encoder/vp9_aq_360.h
new file mode 100644
index 0000000..fb861cb
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_aq_360.h
@@ -0,0 +1,28 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_AQ_360_H_
+#define VP9_ENCODER_VP9_AQ_360_H_
+
+#include "vp9/encoder/vp9_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vp9_360aq_segment_id(int mi_row, int mi_rows);
+void vp9_360aq_frame_setup(VP9_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_AQ_VARIANCE_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_aq_complexity.c b/libvpx/libvpx/vp9/encoder/vp9_aq_complexity.c
new file mode 100644
index 0000000..2d979ec
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_aq_complexity.c
@@ -0,0 +1,161 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <math.h>
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/system_state.h"
+
+#include "vp9/encoder/vp9_aq_complexity.h"
+#include "vp9/encoder/vp9_aq_variance.h"
+#include "vp9/encoder/vp9_encodeframe.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/encoder/vp9_segmentation.h"
+
+#define AQ_C_SEGMENTS  5
+#define DEFAULT_AQ2_SEG 3   // Neutral Q segment
+#define AQ_C_STRENGTHS 3
+static const double aq_c_q_adj_factor[AQ_C_STRENGTHS][AQ_C_SEGMENTS] =
+  { {1.75, 1.25, 1.05, 1.00, 0.90},
+    {2.00, 1.50, 1.15, 1.00, 0.85},
+    {2.50, 1.75, 1.25, 1.00, 0.80} };
+static const double aq_c_transitions[AQ_C_STRENGTHS][AQ_C_SEGMENTS] =
+  { {0.15, 0.30, 0.55, 2.00, 100.0},
+    {0.20, 0.40, 0.65, 2.00, 100.0},
+    {0.25, 0.50, 0.75, 2.00, 100.0} };
+static const double aq_c_var_thresholds[AQ_C_STRENGTHS][AQ_C_SEGMENTS] =
+  { {-4.0, -3.0, -2.0, 100.00, 100.0},
+    {-3.5, -2.5, -1.5, 100.00, 100.0},
+    {-3.0, -2.0, -1.0, 100.00, 100.0} };
+
+static int get_aq_c_strength(int q_index, vpx_bit_depth_t bit_depth) {
+  // Approximate base quatizer (truncated to int)
+  const int base_quant = vp9_ac_quant(q_index, 0, bit_depth) / 4;
+  return (base_quant > 10) + (base_quant > 25);
+}
+
+void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  struct segmentation *const seg = &cm->seg;
+
+  // Make SURE use of floating point in this function is safe.
+  vpx_clear_system_state();
+
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode ||
+      cpi->refresh_alt_ref_frame ||
+      (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+    int segment;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex, cm->bit_depth);
+
+    // Clear down the segment map.
+    memset(cpi->segmentation_map, DEFAULT_AQ2_SEG, cm->mi_rows * cm->mi_cols);
+
+    vp9_clearall_segfeatures(seg);
+
+    // Segmentation only makes sense if the target bits per SB is above a
+    // threshold. Below this the overheads will usually outweigh any benefit.
+    if (cpi->rc.sb64_target_rate < 256) {
+      vp9_disable_segmentation(seg);
+      return;
+    }
+
+    vp9_enable_segmentation(seg);
+
+    // Select delta coding method.
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    // Default segment "Q" feature is disabled so it defaults to the baseline Q.
+    vp9_disable_segfeature(seg, DEFAULT_AQ2_SEG, SEG_LVL_ALT_Q);
+
+    // Use some of the segments for in frame Q adjustment.
+    for (segment = 0; segment < AQ_C_SEGMENTS; ++segment) {
+      int qindex_delta;
+
+      if (segment == DEFAULT_AQ2_SEG)
+        continue;
+
+      qindex_delta =
+        vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type, cm->base_qindex,
+                                   aq_c_q_adj_factor[aq_strength][segment],
+                                   cm->bit_depth);
+
+
+      // For AQ complexity mode, we dont allow Q0 in a segment if the base
+      // Q is not 0. Q0 (lossless) implies 4x4 only and in AQ mode 2 a segment
+      // Q delta is sometimes applied without going back around the rd loop.
+      // This could lead to an illegal combination of partition size and q.
+      if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
+        qindex_delta = -cm->base_qindex + 1;
+      }
+      if ((cm->base_qindex + qindex_delta) > 0) {
+        vp9_enable_segfeature(seg, segment, SEG_LVL_ALT_Q);
+        vp9_set_segdata(seg, segment, SEG_LVL_ALT_Q, qindex_delta);
+      }
+    }
+  }
+}
+
+#define DEFAULT_LV_THRESH 10.0
+#define MIN_DEFAULT_LV_THRESH 8.0
+// Select a segment for the current block.
+// The choice of segment for a block depends on the ratio of the projected
+// bits for the block vs a target average and its spatial complexity.
+void vp9_caq_select_segment(VP9_COMP *cpi, MACROBLOCK *mb, BLOCK_SIZE bs,
+                            int mi_row, int mi_col, int projected_rate) {
+  VP9_COMMON *const cm = &cpi->common;
+
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
+  const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
+  const int xmis = VPXMIN(cm->mi_cols - mi_col, num_8x8_blocks_wide_lookup[bs]);
+  const int ymis = VPXMIN(cm->mi_rows - mi_row, num_8x8_blocks_high_lookup[bs]);
+  int x, y;
+  int i;
+  unsigned char segment;
+
+  if (0) {
+    segment = DEFAULT_AQ2_SEG;
+  } else {
+    // Rate depends on fraction of a SB64 in frame (xmis * ymis / bw * bh).
+    // It is converted to bits * 256 units.
+    const int target_rate = (cpi->rc.sb64_target_rate * xmis * ymis * 256) /
+                            (bw * bh);
+    double logvar;
+    double low_var_thresh;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex, cm->bit_depth);
+
+    vpx_clear_system_state();
+    low_var_thresh = (cpi->oxcf.pass == 2)
+      ? VPXMAX(cpi->twopass.mb_av_energy, MIN_DEFAULT_LV_THRESH)
+      : DEFAULT_LV_THRESH;
+
+    vp9_setup_src_planes(mb, cpi->Source, mi_row, mi_col);
+    logvar = vp9_log_block_var(cpi, mb, bs);
+
+    segment = AQ_C_SEGMENTS - 1;    // Just in case no break out below.
+    for (i = 0; i < AQ_C_SEGMENTS; ++i) {
+      // Test rate against a threshold value and variance against a threshold.
+      // Increasing segment number (higher variance and complexity) = higher Q.
+      if ((projected_rate <
+           target_rate * aq_c_transitions[aq_strength][i]) &&
+          (logvar < (low_var_thresh + aq_c_var_thresholds[aq_strength][i]))) {
+        segment = i;
+        break;
+      }
+    }
+  }
+
+  // Fill in the entires in the segment map corresponding to this SB64.
+  for (y = 0; y < ymis; y++) {
+    for (x = 0; x < xmis; x++) {
+      cpi->segmentation_map[mi_offset + y * cm->mi_cols + x] = segment;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_aq_complexity.h b/libvpx/libvpx/vp9/encoder/vp9_aq_complexity.h
new file mode 100644
index 0000000..e9acb1c
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_aq_complexity.h
@@ -0,0 +1,37 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_AQ_COMPLEXITY_H_
+#define VP9_ENCODER_VP9_AQ_COMPLEXITY_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "vp9/common/vp9_enums.h"
+
+struct VP9_COMP;
+struct macroblock;
+
+// Select a segment for the current Block.
+void vp9_caq_select_segment(struct VP9_COMP *cpi, struct macroblock *,
+                            BLOCK_SIZE bs,
+                            int mi_row, int mi_col, int projected_rate);
+
+// This function sets up a set of segments with delta Q values around
+// the baseline frame quantizer.
+void vp9_setup_in_frame_q_adj(struct VP9_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_AQ_COMPLEXITY_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.c b/libvpx/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.c
new file mode 100644
index 0000000..3e1a0a5
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.c
@@ -0,0 +1,613 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <math.h>
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/system_state.h"
+
+#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
+
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_segmentation.h"
+
+CYCLIC_REFRESH *vp9_cyclic_refresh_alloc(int mi_rows, int mi_cols) {
+  size_t last_coded_q_map_size;
+  CYCLIC_REFRESH *const cr = vpx_calloc(1, sizeof(*cr));
+  if (cr == NULL)
+    return NULL;
+
+  cr->map = vpx_calloc(mi_rows * mi_cols, sizeof(*cr->map));
+  if (cr->map == NULL) {
+    vp9_cyclic_refresh_free(cr);
+    return NULL;
+  }
+  last_coded_q_map_size = mi_rows * mi_cols * sizeof(*cr->last_coded_q_map);
+  cr->last_coded_q_map = vpx_malloc(last_coded_q_map_size);
+  if (cr->last_coded_q_map == NULL) {
+    vp9_cyclic_refresh_free(cr);
+    return NULL;
+  }
+  assert(MAXQ <= 255);
+  memset(cr->last_coded_q_map, MAXQ, last_coded_q_map_size);
+  return cr;
+}
+
+void vp9_cyclic_refresh_free(CYCLIC_REFRESH *cr) {
+  vpx_free(cr->map);
+  vpx_free(cr->last_coded_q_map);
+  vpx_free(cr);
+}
+
+// Check if this coding block, of size bsize, should be considered for refresh
+// (lower-qp coding). Decision can be based on various factors, such as
+// size of the coding block (i.e., below min_block size rejected), coding
+// mode, and rate/distortion.
+static int candidate_refresh_aq(const CYCLIC_REFRESH *cr,
+                                const MODE_INFO *mi,
+                                int64_t rate,
+                                int64_t dist,
+                                int bsize) {
+  MV mv = mi->mv[0].as_mv;
+  // Reject the block for lower-qp coding if projected distortion
+  // is above the threshold, and any of the following is true:
+  // 1) mode uses large mv
+  // 2) mode is an intra-mode
+  // Otherwise accept for refresh.
+  if (dist > cr->thresh_dist_sb &&
+      (mv.row > cr->motion_thresh || mv.row < -cr->motion_thresh ||
+       mv.col > cr->motion_thresh || mv.col < -cr->motion_thresh ||
+       !is_inter_block(mi)))
+    return CR_SEGMENT_ID_BASE;
+  else  if (bsize >= BLOCK_16X16 &&
+            rate < cr->thresh_rate_sb &&
+            is_inter_block(mi) &&
+            mi->mv[0].as_int == 0 &&
+            cr->rate_boost_fac > 10)
+    // More aggressive delta-q for bigger blocks with zero motion.
+    return CR_SEGMENT_ID_BOOST2;
+  else
+    return CR_SEGMENT_ID_BOOST1;
+}
+
+// Compute delta-q for the segment.
+static int compute_deltaq(const VP9_COMP *cpi, int q, double rate_factor) {
+  const CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  int deltaq = vp9_compute_qdelta_by_rate(rc, cpi->common.frame_type,
+                                          q, rate_factor,
+                                          cpi->common.bit_depth);
+  if ((-deltaq) > cr->max_qdelta_perc * q / 100) {
+    deltaq = -cr->max_qdelta_perc * q / 100;
+  }
+  return deltaq;
+}
+
+// For the just encoded frame, estimate the bits, incorporating the delta-q
+// from non-base segment. For now ignore effect of multiple segments
+// (with different delta-q). Note this function is called in the postencode
+// (called from rc_update_rate_correction_factors()).
+int vp9_cyclic_refresh_estimate_bits_at_q(const VP9_COMP *cpi,
+                                          double correction_factor) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int estimated_bits;
+  int mbs = cm->MBs;
+  int num8x8bl = mbs << 2;
+  // Weight for non-base segments: use actual number of blocks refreshed in
+  // previous/just encoded frame. Note number of blocks here is in 8x8 units.
+  double weight_segment1 = (double)cr->actual_num_seg1_blocks / num8x8bl;
+  double weight_segment2 = (double)cr->actual_num_seg2_blocks / num8x8bl;
+  // Take segment weighted average for estimated bits.
+  estimated_bits = (int)((1.0 - weight_segment1 - weight_segment2) *
+      vp9_estimate_bits_at_q(cm->frame_type, cm->base_qindex, mbs,
+                             correction_factor, cm->bit_depth) +
+                             weight_segment1 *
+      vp9_estimate_bits_at_q(cm->frame_type,
+                             cm->base_qindex + cr->qindex_delta[1], mbs,
+                             correction_factor, cm->bit_depth) +
+                             weight_segment2 *
+      vp9_estimate_bits_at_q(cm->frame_type,
+                             cm->base_qindex + cr->qindex_delta[2], mbs,
+                             correction_factor, cm->bit_depth));
+  return estimated_bits;
+}
+
+// Prior to encoding the frame, estimate the bits per mb, for a given q = i and
+// a corresponding delta-q (for segment 1). This function is called in the
+// rc_regulate_q() to set the base qp index.
+// Note: the segment map is set to either 0/CR_SEGMENT_ID_BASE (no refresh) or
+// to 1/CR_SEGMENT_ID_BOOST1 (refresh) for each superblock, prior to encoding.
+int vp9_cyclic_refresh_rc_bits_per_mb(const VP9_COMP *cpi, int i,
+                                      double correction_factor) {
+  const VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int bits_per_mb;
+  int num8x8bl = cm->MBs << 2;
+  // Weight for segment prior to encoding: take the average of the target
+  // number for the frame to be encoded and the actual from the previous frame.
+  int target_refresh = cr->percent_refresh * cm->mi_rows * cm->mi_cols / 100;
+  double weight_segment = (double)((target_refresh +
+      cr->actual_num_seg1_blocks + cr->actual_num_seg2_blocks) >> 1) /
+      num8x8bl;
+  // Compute delta-q corresponding to qindex i.
+  int deltaq = compute_deltaq(cpi, i, cr->rate_ratio_qdelta);
+  // Take segment weighted average for bits per mb.
+  bits_per_mb = (int)((1.0 - weight_segment) *
+      vp9_rc_bits_per_mb(cm->frame_type, i, correction_factor, cm->bit_depth) +
+      weight_segment *
+      vp9_rc_bits_per_mb(cm->frame_type, i + deltaq, correction_factor,
+                         cm->bit_depth));
+  return bits_per_mb;
+}
+
+// Prior to coding a given prediction block, of size bsize at (mi_row, mi_col),
+// check if we should reset the segment_id, and update the cyclic_refresh map
+// and segmentation map.
+void vp9_cyclic_refresh_update_segment(VP9_COMP *const cpi,
+                                       MODE_INFO *const mi,
+                                       int mi_row, int mi_col,
+                                       BLOCK_SIZE bsize,
+                                       int64_t rate,
+                                       int64_t dist,
+                                       int skip,
+                                       struct macroblock_plane *const p) {
+  const VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  const int bw = num_8x8_blocks_wide_lookup[bsize];
+  const int bh = num_8x8_blocks_high_lookup[bsize];
+  const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
+  const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
+  const int block_index = mi_row * cm->mi_cols + mi_col;
+  int refresh_this_block = candidate_refresh_aq(cr, mi, rate, dist, bsize);
+  // Default is to not update the refresh map.
+  int new_map_value = cr->map[block_index];
+  int x = 0; int y = 0;
+
+  int is_skin = 0;
+  if (refresh_this_block == 0 &&
+      bsize <= BLOCK_16X16 &&
+      cpi->use_skin_detection) {
+    is_skin = vp9_compute_skin_block(p[0].src.buf,
+                                     p[1].src.buf,
+                                     p[2].src.buf,
+                                     p[0].src.stride,
+                                     p[1].src.stride,
+                                     bsize,
+                                     0,
+                                     0);
+    if (is_skin)
+      refresh_this_block = 1;
+  }
+
+  if (cpi->oxcf.rc_mode == VPX_VBR && mi->ref_frame[0] == GOLDEN_FRAME)
+    refresh_this_block = 0;
+
+  // If this block is labeled for refresh, check if we should reset the
+  // segment_id.
+  if (cyclic_refresh_segment_id_boosted(mi->segment_id)) {
+    mi->segment_id = refresh_this_block;
+    // Reset segment_id if it will be skipped.
+    if (skip)
+      mi->segment_id = CR_SEGMENT_ID_BASE;
+  }
+
+  // Update the cyclic refresh map, to be used for setting segmentation map
+  // for the next frame. If the block  will be refreshed this frame, mark it
+  // as clean. The magnitude of the -ve influences how long before we consider
+  // it for refresh again.
+  if (cyclic_refresh_segment_id_boosted(mi->segment_id)) {
+    new_map_value = -cr->time_for_refresh;
+  } else if (refresh_this_block) {
+    // Else if it is accepted as candidate for refresh, and has not already
+    // been refreshed (marked as 1) then mark it as a candidate for cleanup
+    // for future time (marked as 0), otherwise don't update it.
+    if (cr->map[block_index] == 1)
+      new_map_value = 0;
+  } else {
+    // Leave it marked as block that is not candidate for refresh.
+    new_map_value = 1;
+  }
+
+  // Update entries in the cyclic refresh map with new_map_value, and
+  // copy mbmi->segment_id into global segmentation map.
+  for (y = 0; y < ymis; y++)
+    for (x = 0; x < xmis; x++) {
+      int map_offset = block_index + y * cm->mi_cols + x;
+      cr->map[map_offset] = new_map_value;
+      cpi->segmentation_map[map_offset] = mi->segment_id;
+    }
+}
+
+void vp9_cyclic_refresh_update_sb_postencode(VP9_COMP *const cpi,
+                                             const MODE_INFO *const mi,
+                                             int mi_row, int mi_col,
+                                             BLOCK_SIZE bsize) {
+  const VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  const int bw = num_8x8_blocks_wide_lookup[bsize];
+  const int bh = num_8x8_blocks_high_lookup[bsize];
+  const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
+  const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
+  const int block_index = mi_row * cm->mi_cols + mi_col;
+  int x, y;
+  for (y = 0; y < ymis; y++)
+    for (x = 0; x < xmis; x++) {
+      int map_offset = block_index + y * cm->mi_cols + x;
+      // Inter skip blocks were clearly not coded at the current qindex, so
+      // don't update the map for them. For cases where motion is non-zero or
+      // the reference frame isn't the previous frame, the previous value in
+      // the map for this spatial location is not entirely correct.
+      if ((!is_inter_block(mi) || !mi->skip) &&
+          mi->segment_id <= CR_SEGMENT_ID_BOOST2) {
+        cr->last_coded_q_map[map_offset] = clamp(
+            cm->base_qindex + cr->qindex_delta[mi->segment_id], 0, MAXQ);
+      } else if (is_inter_block(mi) && mi->skip &&
+                 mi->segment_id <= CR_SEGMENT_ID_BOOST2) {
+        cr->last_coded_q_map[map_offset] = VPXMIN(
+            clamp(cm->base_qindex + cr->qindex_delta[mi->segment_id],
+                  0, MAXQ),
+            cr->last_coded_q_map[map_offset]);
+      }
+    }
+}
+
+// Update the actual number of blocks that were applied the segment delta q.
+void vp9_cyclic_refresh_postencode(VP9_COMP *const cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  unsigned char *const seg_map = cpi->segmentation_map;
+  int mi_row, mi_col;
+  cr->actual_num_seg1_blocks = 0;
+  cr->actual_num_seg2_blocks = 0;
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++)
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
+      if (cyclic_refresh_segment_id(
+          seg_map[mi_row * cm->mi_cols + mi_col]) == CR_SEGMENT_ID_BOOST1)
+        cr->actual_num_seg1_blocks++;
+      else if (cyclic_refresh_segment_id(
+          seg_map[mi_row * cm->mi_cols + mi_col]) == CR_SEGMENT_ID_BOOST2)
+        cr->actual_num_seg2_blocks++;
+    }
+}
+
+// Set golden frame update interval, for non-svc 1 pass CBR mode.
+void vp9_cyclic_refresh_set_golden_update(VP9_COMP *const cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  // Set minimum gf_interval for GF update to a multiple of the refresh period,
+  // with some max limit. Depending on past encoding stats, GF flag may be
+  // reset and update may not occur until next baseline_gf_interval.
+  if (cr->percent_refresh > 0)
+    rc->baseline_gf_interval = VPXMIN(4 * (100 / cr->percent_refresh), 40);
+  else
+    rc->baseline_gf_interval = 40;
+  if (cpi->oxcf.rc_mode == VPX_VBR)
+    rc->baseline_gf_interval = 20;
+}
+
+// Update some encoding stats (from the just encoded frame). If this frame's
+// background has high motion, refresh the golden frame. Otherwise, if the
+// golden reference is to be updated check if we should NOT update the golden
+// ref.
+void vp9_cyclic_refresh_check_golden_update(VP9_COMP *const cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int mi_row, mi_col;
+  double fraction_low = 0.0;
+  int low_content_frame = 0;
+  MODE_INFO **mi = cm->mi_grid_visible;
+  RATE_CONTROL *const rc = &cpi->rc;
+  const int rows = cm->mi_rows, cols = cm->mi_cols;
+  int cnt1 = 0, cnt2 = 0;
+  int force_gf_refresh = 0;
+  int flag_force_gf_high_motion = 0;
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      if (flag_force_gf_high_motion == 1) {
+        int16_t abs_mvr = mi[0]->mv[0].as_mv.row >= 0 ?
+            mi[0]->mv[0].as_mv.row : -1 * mi[0]->mv[0].as_mv.row;
+        int16_t abs_mvc = mi[0]->mv[0].as_mv.col >= 0 ?
+            mi[0]->mv[0].as_mv.col : -1 * mi[0]->mv[0].as_mv.col;
+        // Calculate the motion of the background.
+        if (abs_mvr <= 16 && abs_mvc <= 16) {
+          cnt1++;
+          if (abs_mvr == 0 && abs_mvc == 0)
+            cnt2++;
+        }
+      }
+      mi++;
+      // Accumulate low_content_frame.
+      if (cr->map[mi_row * cols + mi_col] < 1)
+        low_content_frame++;
+    }
+    mi += 8;
+  }
+  // For video conference clips, if the background has high motion in current
+  // frame because of the camera movement, set this frame as the golden frame.
+  // Use 70% and 5% as the thresholds for golden frame refreshing.
+  // Also, force this frame as a golden update frame if this frame will change
+  // the resolution (resize_pending != 0).
+  if (cpi->resize_pending != 0 ||
+     (cnt1 * 100 > (70 * rows * cols) && cnt2 * 20 < cnt1)) {
+    vp9_cyclic_refresh_set_golden_update(cpi);
+    rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+
+    if (rc->frames_till_gf_update_due > rc->frames_to_key)
+      rc->frames_till_gf_update_due = rc->frames_to_key;
+    cpi->refresh_golden_frame = 1;
+    force_gf_refresh = 1;
+  }
+  fraction_low =
+      (double)low_content_frame / (rows * cols);
+  // Update average.
+  cr->low_content_avg = (fraction_low + 3 * cr->low_content_avg) / 4;
+  if (!force_gf_refresh && cpi->refresh_golden_frame == 1) {
+    // Don't update golden reference if the amount of low_content for the
+    // current encoded frame is small, or if the recursive average of the
+    // low_content over the update interval window falls below threshold.
+    if (fraction_low < 0.8 || cr->low_content_avg < 0.7)
+      cpi->refresh_golden_frame = 0;
+    // Reset for next internal.
+    cr->low_content_avg = fraction_low;
+  }
+}
+
+// Update the segmentation map, and related quantities: cyclic refresh map,
+// refresh sb_index, and target number of blocks to be refreshed.
+// The map is set to either 0/CR_SEGMENT_ID_BASE (no refresh) or to
+// 1/CR_SEGMENT_ID_BOOST1 (refresh) for each superblock.
+// Blocks labeled as BOOST1 may later get set to BOOST2 (during the
+// encoding of the superblock).
+static void cyclic_refresh_update_map(VP9_COMP *const cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  unsigned char *const seg_map = cpi->segmentation_map;
+  int i, block_count, bl_index, sb_rows, sb_cols, sbs_in_frame;
+  int xmis, ymis, x, y;
+  int consec_zero_mv_thresh = 0;
+  int qindex_thresh = 0;
+  int count_sel = 0;
+  int count_tot = 0;
+  memset(seg_map, CR_SEGMENT_ID_BASE, cm->mi_rows * cm->mi_cols);
+  sb_cols = (cm->mi_cols + MI_BLOCK_SIZE - 1) / MI_BLOCK_SIZE;
+  sb_rows = (cm->mi_rows + MI_BLOCK_SIZE - 1) / MI_BLOCK_SIZE;
+  sbs_in_frame = sb_cols * sb_rows;
+  // Number of target blocks to get the q delta (segment 1).
+  block_count = cr->percent_refresh * cm->mi_rows * cm->mi_cols / 100;
+  // Set the segmentation map: cycle through the superblocks, starting at
+  // cr->mb_index, and stopping when either block_count blocks have been found
+  // to be refreshed, or we have passed through whole frame.
+  assert(cr->sb_index < sbs_in_frame);
+  i = cr->sb_index;
+  cr->target_num_seg_blocks = 0;
+  if (cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
+    consec_zero_mv_thresh = 100;
+  }
+  qindex_thresh =
+      cpi->oxcf.content == VP9E_CONTENT_SCREEN
+      ? vp9_get_qindex(&cm->seg, CR_SEGMENT_ID_BOOST2, cm->base_qindex)
+      : vp9_get_qindex(&cm->seg, CR_SEGMENT_ID_BOOST1, cm->base_qindex);
+  // More aggressive settings for noisy content.
+  if (cpi->noise_estimate.enabled && cpi->noise_estimate.level >= kMedium) {
+    consec_zero_mv_thresh = 80;
+    qindex_thresh =
+        VPXMAX(vp9_get_qindex(&cm->seg, CR_SEGMENT_ID_BOOST1, cm->base_qindex),
+                              7 * cm->base_qindex >> 3);
+  }
+  do {
+    int sum_map = 0;
+    // Get the mi_row/mi_col corresponding to superblock index i.
+    int sb_row_index = (i / sb_cols);
+    int sb_col_index = i - sb_row_index * sb_cols;
+    int mi_row = sb_row_index * MI_BLOCK_SIZE;
+    int mi_col = sb_col_index * MI_BLOCK_SIZE;
+    assert(mi_row >= 0 && mi_row < cm->mi_rows);
+    assert(mi_col >= 0 && mi_col < cm->mi_cols);
+    bl_index = mi_row * cm->mi_cols + mi_col;
+    // Loop through all 8x8 blocks in superblock and update map.
+    xmis =
+        VPXMIN(cm->mi_cols - mi_col, num_8x8_blocks_wide_lookup[BLOCK_64X64]);
+    ymis =
+        VPXMIN(cm->mi_rows - mi_row, num_8x8_blocks_high_lookup[BLOCK_64X64]);
+    for (y = 0; y < ymis; y++) {
+      for (x = 0; x < xmis; x++) {
+        const int bl_index2 = bl_index + y * cm->mi_cols + x;
+        // If the block is as a candidate for clean up then mark it
+        // for possible boost/refresh (segment 1). The segment id may get
+        // reset to 0 later if block gets coded anything other than ZEROMV.
+        if (cr->map[bl_index2] == 0) {
+          count_tot++;
+          if (cr->last_coded_q_map[bl_index2] > qindex_thresh ||
+              cpi->consec_zero_mv[bl_index2] < consec_zero_mv_thresh) {
+            sum_map++;
+            count_sel++;
+          }
+        } else if (cr->map[bl_index2] < 0) {
+          cr->map[bl_index2]++;
+        }
+      }
+    }
+    // Enforce constant segment over superblock.
+    // If segment is at least half of superblock, set to 1.
+    if (sum_map >= xmis * ymis / 2) {
+      for (y = 0; y < ymis; y++)
+        for (x = 0; x < xmis; x++) {
+          seg_map[bl_index + y * cm->mi_cols + x] = CR_SEGMENT_ID_BOOST1;
+        }
+      cr->target_num_seg_blocks += xmis * ymis;
+    }
+    i++;
+    if (i == sbs_in_frame) {
+      i = 0;
+    }
+  } while (cr->target_num_seg_blocks < block_count && i != cr->sb_index);
+  cr->sb_index = i;
+  cr->reduce_refresh = 0;
+  if (count_sel < (3 * count_tot) >> 2)
+    cr->reduce_refresh = 1;
+}
+
+// Set cyclic refresh parameters.
+void vp9_cyclic_refresh_update_parameters(VP9_COMP *const cpi) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  cr->percent_refresh = 10;
+  if (cr->reduce_refresh)
+    cr->percent_refresh = 5;
+  cr->max_qdelta_perc = 50;
+  cr->time_for_refresh = 0;
+  cr->motion_thresh = 32;
+  cr->rate_boost_fac = 15;
+  // Use larger delta-qp (increase rate_ratio_qdelta) for first few (~4)
+  // periods of the refresh cycle, after a key frame.
+  // Account for larger interval on base layer for temporal layers.
+  if (cr->percent_refresh > 0 &&
+      rc->frames_since_key <  (4 * cpi->svc.number_temporal_layers) *
+      (100 / cr->percent_refresh)) {
+    cr->rate_ratio_qdelta = 3.0;
+  } else {
+    cr->rate_ratio_qdelta = 2.0;
+    if (cpi->noise_estimate.enabled && cpi->noise_estimate.level >= kMedium) {
+      // Reduce the delta-qp if the estimated source noise is above threshold.
+      cr->rate_ratio_qdelta = 1.7;
+      cr->rate_boost_fac = 13;
+    }
+  }
+  // Adjust some parameters for low resolutions at low bitrates.
+  if (cm->width <= 352 &&
+      cm->height <= 288 &&
+      rc->avg_frame_bandwidth < 3400) {
+    cr->motion_thresh = 4;
+    cr->rate_boost_fac = 10;
+  }
+  if (cpi->svc.spatial_layer_id > 0) {
+    cr->motion_thresh = 4;
+    cr->rate_boost_fac = 12;
+  }
+  if (cpi->oxcf.rc_mode == VPX_VBR) {
+    // To be adjusted for VBR mode, e.g., based on gf period and boost.
+    // For now use smaller qp-delta (than CBR), no second boosted seg, and
+    // turn-off (no refresh) on golden refresh (since it's already boosted).
+    cr->percent_refresh = 10;
+    cr->rate_ratio_qdelta = 1.5;
+    cr->rate_boost_fac = 10;
+    if (cpi->refresh_golden_frame == 1) {
+      cr->percent_refresh = 0;
+      cr->rate_ratio_qdelta = 1.0;
+    }
+  }
+}
+
+// Setup cyclic background refresh: set delta q and segmentation map.
+void vp9_cyclic_refresh_setup(VP9_COMP *const cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  struct segmentation *const seg = &cm->seg;
+  // TODO(marpan): Look into whether we should reduce the amount/delta-qp
+  // instead of completely shutting off at low bitrates. For now keep it on.
+  // const int apply_cyclic_refresh = apply_cyclic_refresh_bitrate(cm, rc);
+  const int apply_cyclic_refresh = 1;
+  if (cm->current_video_frame == 0)
+    cr->low_content_avg = 0.0;
+  // Don't apply refresh on key frame or temporal enhancement layer frames.
+  if (!apply_cyclic_refresh ||
+      (cm->frame_type == KEY_FRAME) ||
+      (cpi->svc.temporal_layer_id > 0)) {
+    // Set segmentation map to 0 and disable.
+    unsigned char *const seg_map = cpi->segmentation_map;
+    memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
+    vp9_disable_segmentation(&cm->seg);
+    if (cm->frame_type == KEY_FRAME) {
+      memset(cr->last_coded_q_map, MAXQ,
+             cm->mi_rows * cm->mi_cols * sizeof(*cr->last_coded_q_map));
+      cr->sb_index = 0;
+    }
+    return;
+  } else {
+    int qindex_delta = 0;
+    int qindex2;
+    const double q = vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth);
+    vpx_clear_system_state();
+    // Set rate threshold to some multiple (set to 2 for now) of the target
+    // rate (target is given by sb64_target_rate and scaled by 256).
+    cr->thresh_rate_sb = ((int64_t)(rc->sb64_target_rate) << 8) << 2;
+    // Distortion threshold, quadratic in Q, scale factor to be adjusted.
+    // q will not exceed 457, so (q * q) is within 32bit; see:
+    // vp9_convert_qindex_to_q(), vp9_ac_quant(), ac_qlookup*[].
+    cr->thresh_dist_sb = ((int64_t)(q * q)) << 2;
+
+    // Set up segmentation.
+    // Clear down the segment map.
+    vp9_enable_segmentation(&cm->seg);
+    vp9_clearall_segfeatures(seg);
+    // Select delta coding method.
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    // Note: setting temporal_update has no effect, as the seg-map coding method
+    // (temporal or spatial) is determined in vp9_choose_segmap_coding_method(),
+    // based on the coding cost of each method. For error_resilient mode on the
+    // last_frame_seg_map is set to 0, so if temporal coding is used, it is
+    // relative to 0 previous map.
+    // seg->temporal_update = 0;
+
+    // Segment BASE "Q" feature is disabled so it defaults to the baseline Q.
+    vp9_disable_segfeature(seg, CR_SEGMENT_ID_BASE, SEG_LVL_ALT_Q);
+    // Use segment BOOST1 for in-frame Q adjustment.
+    vp9_enable_segfeature(seg, CR_SEGMENT_ID_BOOST1, SEG_LVL_ALT_Q);
+    // Use segment BOOST2 for more aggressive in-frame Q adjustment.
+    vp9_enable_segfeature(seg, CR_SEGMENT_ID_BOOST2, SEG_LVL_ALT_Q);
+
+    // Set the q delta for segment BOOST1.
+    qindex_delta = compute_deltaq(cpi, cm->base_qindex, cr->rate_ratio_qdelta);
+    cr->qindex_delta[1] = qindex_delta;
+
+    // Compute rd-mult for segment BOOST1.
+    qindex2 = clamp(cm->base_qindex + cm->y_dc_delta_q + qindex_delta, 0, MAXQ);
+
+    cr->rdmult = vp9_compute_rd_mult(cpi, qindex2);
+
+    vp9_set_segdata(seg, CR_SEGMENT_ID_BOOST1, SEG_LVL_ALT_Q, qindex_delta);
+
+    // Set a more aggressive (higher) q delta for segment BOOST2.
+    qindex_delta = compute_deltaq(
+        cpi, cm->base_qindex,
+        VPXMIN(CR_MAX_RATE_TARGET_RATIO,
+               0.1 * cr->rate_boost_fac * cr->rate_ratio_qdelta));
+    cr->qindex_delta[2] = qindex_delta;
+    vp9_set_segdata(seg, CR_SEGMENT_ID_BOOST2, SEG_LVL_ALT_Q, qindex_delta);
+
+    // Reset if resoluton change has occurred.
+    if (cpi->resize_pending != 0)
+      vp9_cyclic_refresh_reset_resize(cpi);
+
+    // Update the segmentation and refresh map.
+    cyclic_refresh_update_map(cpi);
+  }
+}
+
+int vp9_cyclic_refresh_get_rdmult(const CYCLIC_REFRESH *cr) {
+  return cr->rdmult;
+}
+
+void vp9_cyclic_refresh_reset_resize(VP9_COMP *const cpi) {
+  const VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  memset(cr->map, 0, cm->mi_rows * cm->mi_cols);
+  memset(cr->last_coded_q_map, MAXQ, cm->mi_rows * cm->mi_cols);
+  cr->sb_index = 0;
+  cpi->refresh_golden_frame = 1;
+  cpi->refresh_alt_ref_frame = 1;
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.h b/libvpx/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.h
new file mode 100644
index 0000000..35eea18
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.h
@@ -0,0 +1,145 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_AQ_CYCLICREFRESH_H_
+#define VP9_ENCODER_VP9_AQ_CYCLICREFRESH_H_
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_skin_detection.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The segment ids used in cyclic refresh: from base (no boost) to increasing
+// boost (higher delta-qp).
+#define CR_SEGMENT_ID_BASE    0
+#define CR_SEGMENT_ID_BOOST1  1
+#define CR_SEGMENT_ID_BOOST2  2
+
+// Maximum rate target ratio for setting segment delta-qp.
+#define CR_MAX_RATE_TARGET_RATIO 4.0
+
+struct CYCLIC_REFRESH {
+  // Percentage of blocks per frame that are targeted as candidates
+  // for cyclic refresh.
+  int percent_refresh;
+  // Maximum q-delta as percentage of base q.
+  int max_qdelta_perc;
+  // Superblock starting index for cycling through the frame.
+  int sb_index;
+  // Controls how long block will need to wait to be refreshed again, in
+  // excess of the cycle time, i.e., in the case of all zero motion, block
+  // will be refreshed every (100/percent_refresh + time_for_refresh) frames.
+  int time_for_refresh;
+  // Target number of (8x8) blocks that are set for delta-q.
+  int target_num_seg_blocks;
+  // Actual number of (8x8) blocks that were applied delta-q.
+  int actual_num_seg1_blocks;
+  int actual_num_seg2_blocks;
+  // RD mult. parameters for segment 1.
+  int rdmult;
+  // Cyclic refresh map.
+  signed char *map;
+  // Map of the last q a block was coded at.
+  uint8_t *last_coded_q_map;
+  // Thresholds applied to the projected rate/distortion of the coding block,
+  // when deciding whether block should be refreshed.
+  int64_t thresh_rate_sb;
+  int64_t thresh_dist_sb;
+  // Threshold applied to the motion vector (in units of 1/8 pel) of the
+  // coding block, when deciding whether block should be refreshed.
+  int16_t motion_thresh;
+  // Rate target ratio to set q delta.
+  double rate_ratio_qdelta;
+  // Boost factor for rate target ratio, for segment CR_SEGMENT_ID_BOOST2.
+  int rate_boost_fac;
+  double low_content_avg;
+  int qindex_delta[3];
+  int reduce_refresh;
+};
+
+struct VP9_COMP;
+
+typedef struct CYCLIC_REFRESH CYCLIC_REFRESH;
+
+CYCLIC_REFRESH *vp9_cyclic_refresh_alloc(int mi_rows, int mi_cols);
+
+void vp9_cyclic_refresh_free(CYCLIC_REFRESH *cr);
+
+// Estimate the bits, incorporating the delta-q from segment 1, after encoding
+// the frame.
+int vp9_cyclic_refresh_estimate_bits_at_q(const struct VP9_COMP *cpi,
+                                          double correction_factor);
+
+// Estimate the bits per mb, for a given q = i and a corresponding delta-q
+// (for segment 1), prior to encoding the frame.
+int vp9_cyclic_refresh_rc_bits_per_mb(const struct VP9_COMP *cpi, int i,
+                                      double correction_factor);
+
+// Prior to coding a given prediction block, of size bsize at (mi_row, mi_col),
+// check if we should reset the segment_id, and update the cyclic_refresh map
+// and segmentation map.
+void vp9_cyclic_refresh_update_segment(struct VP9_COMP *const cpi,
+                                       MODE_INFO *const mi,
+                                       int mi_row, int mi_col, BLOCK_SIZE bsize,
+                                       int64_t rate, int64_t dist, int skip,
+                                       struct macroblock_plane *const p);
+
+void vp9_cyclic_refresh_update_sb_postencode(struct VP9_COMP *const cpi,
+                                             const MODE_INFO *const mi,
+                                             int mi_row, int mi_col,
+                                             BLOCK_SIZE bsize);
+
+// Update the segmentation map, and related quantities: cyclic refresh map,
+// refresh sb_index, and target number of blocks to be refreshed.
+void vp9_cyclic_refresh_update__map(struct VP9_COMP *const cpi);
+
+// Update the actual number of blocks that were applied the segment delta q.
+void vp9_cyclic_refresh_postencode(struct VP9_COMP *const cpi);
+
+// Set golden frame update interval, for non-svc 1 pass CBR mode.
+void vp9_cyclic_refresh_set_golden_update(struct VP9_COMP *const cpi);
+
+// Check if we should not update golden reference, based on past refresh stats.
+void vp9_cyclic_refresh_check_golden_update(struct VP9_COMP *const cpi);
+
+// Set/update global/frame level refresh parameters.
+void vp9_cyclic_refresh_update_parameters(struct VP9_COMP *const cpi);
+
+// Setup cyclic background refresh: set delta q and segmentation map.
+void vp9_cyclic_refresh_setup(struct VP9_COMP *const cpi);
+
+int vp9_cyclic_refresh_get_rdmult(const CYCLIC_REFRESH *cr);
+
+void vp9_cyclic_refresh_reset_resize(struct VP9_COMP *const cpi);
+
+static INLINE int cyclic_refresh_segment_id_boosted(int segment_id) {
+  return segment_id == CR_SEGMENT_ID_BOOST1 ||
+         segment_id == CR_SEGMENT_ID_BOOST2;
+}
+
+static INLINE int cyclic_refresh_segment_id(int segment_id) {
+  if (segment_id == CR_SEGMENT_ID_BOOST1)
+    return CR_SEGMENT_ID_BOOST1;
+  else if (segment_id == CR_SEGMENT_ID_BOOST2)
+    return CR_SEGMENT_ID_BOOST2;
+  else
+    return CR_SEGMENT_ID_BASE;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_AQ_CYCLICREFRESH_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_aq_variance.c b/libvpx/libvpx/vp9/encoder/vp9_aq_variance.c
new file mode 100644
index 0000000..59ef5fa
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_aq_variance.c
@@ -0,0 +1,207 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "vpx_ports/mem.h"
+#include "vpx_ports/system_state.h"
+
+#include "vp9/encoder/vp9_aq_variance.h"
+
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_segmentation.h"
+
+#define ENERGY_MIN (-4)
+#define ENERGY_MAX (1)
+#define ENERGY_SPAN (ENERGY_MAX - ENERGY_MIN +  1)
+#define ENERGY_IN_BOUNDS(energy)\
+  assert((energy) >= ENERGY_MIN && (energy) <= ENERGY_MAX)
+
+static const double rate_ratio[MAX_SEGMENTS] =
+  {2.5, 2.0, 1.5, 1.0, 0.75, 1.0, 1.0, 1.0};
+static const int segment_id[ENERGY_SPAN] = {0, 1, 1, 2, 3, 4};
+
+#define SEGMENT_ID(i) segment_id[(i) - ENERGY_MIN]
+
+DECLARE_ALIGNED(16, static const uint8_t, vp9_64_zeros[64]) = {0};
+#if CONFIG_VP9_HIGHBITDEPTH
+DECLARE_ALIGNED(16, static const uint16_t, vp9_highbd_64_zeros[64]) = {0};
+#endif
+
+unsigned int vp9_vaq_segment_id(int energy) {
+  ENERGY_IN_BOUNDS(energy);
+  return SEGMENT_ID(energy);
+}
+
+void vp9_vaq_frame_setup(VP9_COMP *cpi) {
+  VP9_COMMON *cm = &cpi->common;
+  struct segmentation *seg = &cm->seg;
+  int i;
+
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode ||
+      cpi->refresh_alt_ref_frame ||
+      (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+    vp9_enable_segmentation(seg);
+    vp9_clearall_segfeatures(seg);
+
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    vpx_clear_system_state();
+
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      int qindex_delta =
+          vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type, cm->base_qindex,
+                                     rate_ratio[i], cm->bit_depth);
+
+      // We don't allow qindex 0 in a segment if the base value is not 0.
+      // Q index 0 (lossless) implies 4x4 encoding only and in AQ mode a segment
+      // Q delta is sometimes applied without going back around the rd loop.
+      // This could lead to an illegal combination of partition size and q.
+      if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
+        qindex_delta = -cm->base_qindex + 1;
+      }
+
+      // No need to enable SEG_LVL_ALT_Q for this segment.
+      if (rate_ratio[i] == 1.0) {
+        continue;
+      }
+
+      vp9_set_segdata(seg, i, SEG_LVL_ALT_Q, qindex_delta);
+      vp9_enable_segfeature(seg, i, SEG_LVL_ALT_Q);
+    }
+  }
+}
+
+/* TODO(agrange, paulwilkins): The block_variance calls the unoptimized versions
+ * of variance() and highbd_8_variance(). It should not.
+ */
+static void aq_variance(const uint8_t *a, int  a_stride,
+                        const uint8_t *b, int  b_stride,
+                        int  w, int  h, unsigned int *sse, int *sum) {
+  int i, j;
+
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void aq_highbd_variance64(const uint8_t *a8, int  a_stride,
+                                 const uint8_t *b8, int  b_stride,
+                                 int w, int h, uint64_t *sse, uint64_t *sum) {
+  int i, j;
+
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+static void aq_highbd_8_variance(const uint8_t *a8, int  a_stride,
+                                 const uint8_t *b8, int  b_stride,
+                                 int w, int h, unsigned int *sse, int *sum) {
+  uint64_t sse_long = 0;
+  uint64_t sum_long = 0;
+  aq_highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
+  *sse = (unsigned int)sse_long;
+  *sum = (int)sum_long;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static unsigned int block_variance(VP9_COMP *cpi, MACROBLOCK *x,
+                                   BLOCK_SIZE bs) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  unsigned int var, sse;
+  int right_overflow = (xd->mb_to_right_edge < 0) ?
+      ((-xd->mb_to_right_edge) >> 3) : 0;
+  int bottom_overflow = (xd->mb_to_bottom_edge < 0) ?
+      ((-xd->mb_to_bottom_edge) >> 3) : 0;
+
+  if (right_overflow || bottom_overflow) {
+    const int bw = 8 * num_8x8_blocks_wide_lookup[bs] - right_overflow;
+    const int bh = 8 * num_8x8_blocks_high_lookup[bs] - bottom_overflow;
+    int avg;
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      aq_highbd_8_variance(x->plane[0].src.buf, x->plane[0].src.stride,
+                           CONVERT_TO_BYTEPTR(vp9_highbd_64_zeros), 0, bw, bh,
+                           &sse, &avg);
+      sse >>= 2 * (xd->bd - 8);
+      avg >>= (xd->bd - 8);
+    } else {
+      aq_variance(x->plane[0].src.buf, x->plane[0].src.stride,
+                  vp9_64_zeros, 0, bw, bh, &sse, &avg);
+    }
+#else
+    aq_variance(x->plane[0].src.buf, x->plane[0].src.stride,
+                vp9_64_zeros, 0, bw, bh, &sse, &avg);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    var = sse - (((int64_t)avg * avg) / (bw * bh));
+    return (unsigned int)(((uint64_t)256 * var) / (bw * bh));
+  } else {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf,
+                               x->plane[0].src.stride,
+                               CONVERT_TO_BYTEPTR(vp9_highbd_64_zeros),
+                               0, &sse);
+    } else {
+      var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf,
+                               x->plane[0].src.stride,
+                               vp9_64_zeros, 0, &sse);
+    }
+#else
+    var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf,
+                             x->plane[0].src.stride,
+                             vp9_64_zeros, 0, &sse);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    return (unsigned int)(((uint64_t)256 * var) >> num_pels_log2_lookup[bs]);
+  }
+}
+
+double vp9_log_block_var(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs) {
+  unsigned int var = block_variance(cpi, x, bs);
+  vpx_clear_system_state();
+  return log(var + 1.0);
+}
+
+#define DEFAULT_E_MIDPOINT 10.0
+int vp9_block_energy(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs) {
+  double energy;
+  double energy_midpoint;
+  vpx_clear_system_state();
+  energy_midpoint =
+    (cpi->oxcf.pass == 2) ? cpi->twopass.mb_av_energy : DEFAULT_E_MIDPOINT;
+  energy = vp9_log_block_var(cpi, x, bs) - energy_midpoint;
+  return clamp((int)round(energy), ENERGY_MIN, ENERGY_MAX);
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_aq_variance.h b/libvpx/libvpx/vp9/encoder/vp9_aq_variance.h
new file mode 100644
index 0000000..a0effa3
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_aq_variance.h
@@ -0,0 +1,31 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_AQ_VARIANCE_H_
+#define VP9_ENCODER_VP9_AQ_VARIANCE_H_
+
+#include "vp9/encoder/vp9_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vp9_vaq_segment_id(int energy);
+void vp9_vaq_frame_setup(VP9_COMP *cpi);
+
+int vp9_block_energy(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs);
+double vp9_log_block_var(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_AQ_VARIANCE_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_bitstream.c b/libvpx/libvpx/vp9/encoder/vp9_bitstream.c
new file mode 100644
index 0000000..73a2db0
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_bitstream.c
@@ -0,0 +1,1245 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+#include <limits.h>
+
+#include "vpx/vpx_encoder.h"
+#include "vpx_dsp/bitwriter_buffer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem_ops.h"
+#include "vpx_ports/system_state.h"
+
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_tile_common.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_bitstream.h"
+#include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_segmentation.h"
+#include "vp9/encoder/vp9_subexp.h"
+#include "vp9/encoder/vp9_tokenize.h"
+
+static const struct vp9_token intra_mode_encodings[INTRA_MODES] = {
+  {0, 1}, {6, 3}, {28, 5}, {30, 5}, {58, 6}, {59, 6}, {126, 7}, {127, 7},
+  {62, 6}, {2, 2}};
+static const struct vp9_token switchable_interp_encodings[SWITCHABLE_FILTERS] =
+  {{0, 1}, {2, 2}, {3, 2}};
+static const struct vp9_token partition_encodings[PARTITION_TYPES] =
+  {{0, 1}, {2, 2}, {6, 3}, {7, 3}};
+static const struct vp9_token inter_mode_encodings[INTER_MODES] =
+  {{2, 2}, {6, 3}, {0, 1}, {7, 3}};
+
+static void write_intra_mode(vpx_writer *w, PREDICTION_MODE mode,
+                             const vpx_prob *probs) {
+  vp9_write_token(w, vp9_intra_mode_tree, probs, &intra_mode_encodings[mode]);
+}
+
+static void write_inter_mode(vpx_writer *w, PREDICTION_MODE mode,
+                             const vpx_prob *probs) {
+  assert(is_inter_mode(mode));
+  vp9_write_token(w, vp9_inter_mode_tree, probs,
+                  &inter_mode_encodings[INTER_OFFSET(mode)]);
+}
+
+static void encode_unsigned_max(struct vpx_write_bit_buffer *wb,
+                                int data, int max) {
+  vpx_wb_write_literal(wb, data, get_unsigned_bits(max));
+}
+
+static void prob_diff_update(const vpx_tree_index *tree,
+                             vpx_prob probs[/*n - 1*/],
+                             const unsigned int counts[/*n - 1*/],
+                             int n, vpx_writer *w) {
+  int i;
+  unsigned int branch_ct[32][2];
+
+  // Assuming max number of probabilities <= 32
+  assert(n <= 32);
+
+  vp9_tree_probs_from_distribution(tree, branch_ct, counts);
+  for (i = 0; i < n - 1; ++i)
+    vp9_cond_prob_diff_update(w, &probs[i], branch_ct[i]);
+}
+
+static void write_selected_tx_size(const VP9_COMMON *cm,
+                                   const MACROBLOCKD *xd, vpx_writer *w) {
+  TX_SIZE tx_size = xd->mi[0]->tx_size;
+  BLOCK_SIZE bsize = xd->mi[0]->sb_type;
+  const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
+  const vpx_prob *const tx_probs = get_tx_probs2(max_tx_size, xd,
+                                                 &cm->fc->tx_probs);
+  vpx_write(w, tx_size != TX_4X4, tx_probs[0]);
+  if (tx_size != TX_4X4 && max_tx_size >= TX_16X16) {
+    vpx_write(w, tx_size != TX_8X8, tx_probs[1]);
+    if (tx_size != TX_8X8 && max_tx_size >= TX_32X32)
+      vpx_write(w, tx_size != TX_16X16, tx_probs[2]);
+  }
+}
+
+static int write_skip(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                      int segment_id, const MODE_INFO *mi, vpx_writer *w) {
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
+    return 1;
+  } else {
+    const int skip = mi->skip;
+    vpx_write(w, skip, vp9_get_skip_prob(cm, xd));
+    return skip;
+  }
+}
+
+static void update_skip_probs(VP9_COMMON *cm, vpx_writer *w,
+                              FRAME_COUNTS *counts) {
+  int k;
+
+  for (k = 0; k < SKIP_CONTEXTS; ++k)
+    vp9_cond_prob_diff_update(w, &cm->fc->skip_probs[k], counts->skip[k]);
+}
+
+static void update_switchable_interp_probs(VP9_COMMON *cm, vpx_writer *w,
+                                           FRAME_COUNTS *counts) {
+  int j;
+  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
+    prob_diff_update(vp9_switchable_interp_tree,
+                     cm->fc->switchable_interp_prob[j],
+                     counts->switchable_interp[j], SWITCHABLE_FILTERS, w);
+}
+
+static void pack_mb_tokens(vpx_writer *w,
+                           TOKENEXTRA **tp, const TOKENEXTRA *const stop,
+                           vpx_bit_depth_t bit_depth) {
+  const TOKENEXTRA *p;
+  const vp9_extra_bit *const extra_bits =
+#if CONFIG_VP9_HIGHBITDEPTH
+    (bit_depth == VPX_BITS_12) ? vp9_extra_bits_high12 :
+    (bit_depth == VPX_BITS_10) ? vp9_extra_bits_high10 :
+    vp9_extra_bits;
+#else
+    vp9_extra_bits;
+    (void) bit_depth;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  for (p = *tp; p < stop && p->token != EOSB_TOKEN; ++p) {
+    if (p->token == EOB_TOKEN) {
+      vpx_write(w, 0, p->context_tree[0]);
+      continue;
+    }
+    vpx_write(w, 1, p->context_tree[0]);
+    while (p->token == ZERO_TOKEN) {
+      vpx_write(w, 0, p->context_tree[1]);
+      ++p;
+      if (p == stop || p->token == EOSB_TOKEN) {
+        *tp = (TOKENEXTRA*)(uintptr_t)p + (p->token == EOSB_TOKEN);
+        return;
+      }
+    }
+
+    {
+      const int t = p->token;
+      const vpx_prob *const context_tree = p->context_tree;
+      assert(t != ZERO_TOKEN);
+      assert(t != EOB_TOKEN);
+      assert(t != EOSB_TOKEN);
+      vpx_write(w, 1, context_tree[1]);
+      if (t == ONE_TOKEN) {
+        vpx_write(w, 0, context_tree[2]);
+        vpx_write_bit(w, p->extra & 1);
+      } else {  // t >= TWO_TOKEN && t < EOB_TOKEN
+        const struct vp9_token *const a = &vp9_coef_encodings[t];
+        const int v = a->value;
+        const int n = a->len;
+        const int e = p->extra;
+        vpx_write(w, 1, context_tree[2]);
+        vp9_write_tree(w, vp9_coef_con_tree,
+                       vp9_pareto8_full[context_tree[PIVOT_NODE] - 1], v,
+                       n - UNCONSTRAINED_NODES, 0);
+        if (t >= CATEGORY1_TOKEN) {
+          const vp9_extra_bit *const b = &extra_bits[t];
+          const unsigned char *pb = b->prob;
+          int v = e >> 1;
+          int n = b->len;  // number of bits in v, assumed nonzero
+          do {
+            const int bb = (v >> --n) & 1;
+            vpx_write(w, bb, *pb++);
+          } while (n);
+        }
+        vpx_write_bit(w, e & 1);
+      }
+    }
+  }
+  *tp = (TOKENEXTRA*)(uintptr_t)p + (p->token == EOSB_TOKEN);
+}
+
+static void write_segment_id(vpx_writer *w, const struct segmentation *seg,
+                             int segment_id) {
+  if (seg->enabled && seg->update_map)
+    vp9_write_tree(w, vp9_segment_tree, seg->tree_probs, segment_id, 3, 0);
+}
+
+// This function encodes the reference frame
+static void write_ref_frames(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                             vpx_writer *w) {
+  const MODE_INFO *const mi = xd->mi[0];
+  const int is_compound = has_second_ref(mi);
+  const int segment_id = mi->segment_id;
+
+  // If segment level coding of this signal is disabled...
+  // or the segment allows multiple reference frame options
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
+    assert(!is_compound);
+    assert(mi->ref_frame[0] ==
+               get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME));
+  } else {
+    // does the feature use compound prediction or not
+    // (if not specified at the frame/segment level)
+    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+      vpx_write(w, is_compound, vp9_get_reference_mode_prob(cm, xd));
+    } else {
+      assert(!is_compound == (cm->reference_mode == SINGLE_REFERENCE));
+    }
+
+    if (is_compound) {
+      vpx_write(w, mi->ref_frame[0] == GOLDEN_FRAME,
+                vp9_get_pred_prob_comp_ref_p(cm, xd));
+    } else {
+      const int bit0 = mi->ref_frame[0] != LAST_FRAME;
+      vpx_write(w, bit0, vp9_get_pred_prob_single_ref_p1(cm, xd));
+      if (bit0) {
+        const int bit1 = mi->ref_frame[0] != GOLDEN_FRAME;
+        vpx_write(w, bit1, vp9_get_pred_prob_single_ref_p2(cm, xd));
+      }
+    }
+  }
+}
+
+static void pack_inter_mode_mvs(VP9_COMP *cpi, const MODE_INFO *mi,
+                                vpx_writer *w) {
+  VP9_COMMON *const cm = &cpi->common;
+  const nmv_context *nmvc = &cm->fc->nmvc;
+  const MACROBLOCK *const x = &cpi->td.mb;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct segmentation *const seg = &cm->seg;
+  const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  const PREDICTION_MODE mode = mi->mode;
+  const int segment_id = mi->segment_id;
+  const BLOCK_SIZE bsize = mi->sb_type;
+  const int allow_hp = cm->allow_high_precision_mv;
+  const int is_inter = is_inter_block(mi);
+  const int is_compound = has_second_ref(mi);
+  int skip, ref;
+
+  if (seg->update_map) {
+    if (seg->temporal_update) {
+      const int pred_flag = mi->seg_id_predicted;
+      vpx_prob pred_prob = vp9_get_pred_prob_seg_id(seg, xd);
+      vpx_write(w, pred_flag, pred_prob);
+      if (!pred_flag)
+        write_segment_id(w, seg, segment_id);
+    } else {
+      write_segment_id(w, seg, segment_id);
+    }
+  }
+
+  skip = write_skip(cm, xd, segment_id, mi, w);
+
+  if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
+    vpx_write(w, is_inter, vp9_get_intra_inter_prob(cm, xd));
+
+  if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT &&
+      !(is_inter && skip)) {
+    write_selected_tx_size(cm, xd, w);
+  }
+
+  if (!is_inter) {
+    if (bsize >= BLOCK_8X8) {
+      write_intra_mode(w, mode, cm->fc->y_mode_prob[size_group_lookup[bsize]]);
+    } else {
+      int idx, idy;
+      const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+      const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+      for (idy = 0; idy < 2; idy += num_4x4_h) {
+        for (idx = 0; idx < 2; idx += num_4x4_w) {
+          const PREDICTION_MODE b_mode = mi->bmi[idy * 2 + idx].as_mode;
+          write_intra_mode(w, b_mode, cm->fc->y_mode_prob[0]);
+        }
+      }
+    }
+    write_intra_mode(w, mi->uv_mode, cm->fc->uv_mode_prob[mode]);
+  } else {
+    const int mode_ctx = mbmi_ext->mode_context[mi->ref_frame[0]];
+    const vpx_prob *const inter_probs = cm->fc->inter_mode_probs[mode_ctx];
+    write_ref_frames(cm, xd, w);
+
+    // If segment skip is not enabled code the mode.
+    if (!segfeature_active(seg, segment_id, SEG_LVL_SKIP)) {
+      if (bsize >= BLOCK_8X8) {
+        write_inter_mode(w, mode, inter_probs);
+      }
+    }
+
+    if (cm->interp_filter == SWITCHABLE) {
+      const int ctx = vp9_get_pred_context_switchable_interp(xd);
+      vp9_write_token(w, vp9_switchable_interp_tree,
+                      cm->fc->switchable_interp_prob[ctx],
+                      &switchable_interp_encodings[mi->interp_filter]);
+      ++cpi->interp_filter_selected[0][mi->interp_filter];
+    } else {
+      assert(mi->interp_filter == cm->interp_filter);
+    }
+
+    if (bsize < BLOCK_8X8) {
+      const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+      const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+      int idx, idy;
+      for (idy = 0; idy < 2; idy += num_4x4_h) {
+        for (idx = 0; idx < 2; idx += num_4x4_w) {
+          const int j = idy * 2 + idx;
+          const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
+          write_inter_mode(w, b_mode, inter_probs);
+          if (b_mode == NEWMV) {
+            for (ref = 0; ref < 1 + is_compound; ++ref)
+              vp9_encode_mv(cpi, w, &mi->bmi[j].as_mv[ref].as_mv,
+                            &mbmi_ext->ref_mvs[mi->ref_frame[ref]][0].as_mv,
+                            nmvc, allow_hp);
+          }
+        }
+      }
+    } else {
+      if (mode == NEWMV) {
+        for (ref = 0; ref < 1 + is_compound; ++ref)
+          vp9_encode_mv(cpi, w, &mi->mv[ref].as_mv,
+                        &mbmi_ext->ref_mvs[mi->ref_frame[ref]][0].as_mv, nmvc,
+                        allow_hp);
+      }
+    }
+  }
+}
+
+static void write_mb_modes_kf(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                              MODE_INFO **mi_8x8, vpx_writer *w) {
+  const struct segmentation *const seg = &cm->seg;
+  const MODE_INFO *const mi = mi_8x8[0];
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const BLOCK_SIZE bsize = mi->sb_type;
+
+  if (seg->update_map)
+    write_segment_id(w, seg, mi->segment_id);
+
+  write_skip(cm, xd, mi->segment_id, mi, w);
+
+  if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT)
+    write_selected_tx_size(cm, xd, w);
+
+  if (bsize >= BLOCK_8X8) {
+    write_intra_mode(w, mi->mode, get_y_mode_probs(mi, above_mi, left_mi, 0));
+  } else {
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+    int idx, idy;
+
+    for (idy = 0; idy < 2; idy += num_4x4_h) {
+      for (idx = 0; idx < 2; idx += num_4x4_w) {
+        const int block = idy * 2 + idx;
+        write_intra_mode(w, mi->bmi[block].as_mode,
+                         get_y_mode_probs(mi, above_mi, left_mi, block));
+      }
+    }
+  }
+
+  write_intra_mode(w, mi->uv_mode, vp9_kf_uv_mode_prob[mi->mode]);
+}
+
+static void write_modes_b(VP9_COMP *cpi, const TileInfo *const tile,
+                          vpx_writer *w, TOKENEXTRA **tok,
+                          const TOKENEXTRA *const tok_end,
+                          int mi_row, int mi_col) {
+  const VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+  MODE_INFO *m;
+
+  xd->mi = cm->mi_grid_visible + (mi_row * cm->mi_stride + mi_col);
+  m = xd->mi[0];
+
+  cpi->td.mb.mbmi_ext = cpi->td.mb.mbmi_ext_base +
+      (mi_row * cm->mi_cols + mi_col);
+
+  set_mi_row_col(xd, tile,
+                 mi_row, num_8x8_blocks_high_lookup[m->sb_type],
+                 mi_col, num_8x8_blocks_wide_lookup[m->sb_type],
+                 cm->mi_rows, cm->mi_cols);
+  if (frame_is_intra_only(cm)) {
+    write_mb_modes_kf(cm, xd, xd->mi, w);
+  } else {
+    pack_inter_mode_mvs(cpi, m, w);
+  }
+
+  assert(*tok < tok_end);
+  pack_mb_tokens(w, tok, tok_end, cm->bit_depth);
+}
+
+static void write_partition(const VP9_COMMON *const cm,
+                            const MACROBLOCKD *const xd,
+                            int hbs, int mi_row, int mi_col,
+                            PARTITION_TYPE p, BLOCK_SIZE bsize, vpx_writer *w) {
+  const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
+  const vpx_prob *const probs = xd->partition_probs[ctx];
+  const int has_rows = (mi_row + hbs) < cm->mi_rows;
+  const int has_cols = (mi_col + hbs) < cm->mi_cols;
+
+  if (has_rows && has_cols) {
+    vp9_write_token(w, vp9_partition_tree, probs, &partition_encodings[p]);
+  } else if (!has_rows && has_cols) {
+    assert(p == PARTITION_SPLIT || p == PARTITION_HORZ);
+    vpx_write(w, p == PARTITION_SPLIT, probs[1]);
+  } else if (has_rows && !has_cols) {
+    assert(p == PARTITION_SPLIT || p == PARTITION_VERT);
+    vpx_write(w, p == PARTITION_SPLIT, probs[2]);
+  } else {
+    assert(p == PARTITION_SPLIT);
+  }
+}
+
+static void write_modes_sb(VP9_COMP *cpi,
+                           const TileInfo *const tile, vpx_writer *w,
+                           TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
+                           int mi_row, int mi_col, BLOCK_SIZE bsize) {
+  const VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+
+  const int bsl = b_width_log2_lookup[bsize];
+  const int bs = (1 << bsl) / 4;
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+  const MODE_INFO *m = NULL;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  m = cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col];
+
+  partition = partition_lookup[bsl][m->sb_type];
+  write_partition(cm, xd, bs, mi_row, mi_col, partition, bsize, w);
+  subsize = get_subsize(bsize, partition);
+  if (subsize < BLOCK_8X8) {
+    write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+  } else {
+    switch (partition) {
+      case PARTITION_NONE:
+        write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+        break;
+      case PARTITION_HORZ:
+        write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+        if (mi_row + bs < cm->mi_rows)
+          write_modes_b(cpi, tile, w, tok, tok_end, mi_row + bs, mi_col);
+        break;
+      case PARTITION_VERT:
+        write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+        if (mi_col + bs < cm->mi_cols)
+          write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col + bs);
+        break;
+      case PARTITION_SPLIT:
+        write_modes_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col, subsize);
+        write_modes_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col + bs,
+                       subsize);
+        write_modes_sb(cpi, tile, w, tok, tok_end, mi_row + bs, mi_col,
+                       subsize);
+        write_modes_sb(cpi, tile, w, tok, tok_end, mi_row + bs, mi_col + bs,
+                       subsize);
+        break;
+      default:
+        assert(0);
+    }
+  }
+
+  // update partition context
+  if (bsize >= BLOCK_8X8 &&
+      (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
+    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+}
+
+static void write_modes(VP9_COMP *cpi,
+                        const TileInfo *const tile, vpx_writer *w,
+                        TOKENEXTRA **tok, const TOKENEXTRA *const tok_end) {
+  const VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+  int mi_row, mi_col;
+
+  set_partition_probs(cm, xd);
+
+  for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
+       mi_row += MI_BLOCK_SIZE) {
+    vp9_zero(xd->left_seg_context);
+    for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
+         mi_col += MI_BLOCK_SIZE)
+      write_modes_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col,
+                     BLOCK_64X64);
+  }
+}
+
+static void build_tree_distribution(VP9_COMP *cpi, TX_SIZE tx_size,
+                                    vp9_coeff_stats *coef_branch_ct,
+                                    vp9_coeff_probs_model *coef_probs) {
+  vp9_coeff_count *coef_counts = cpi->td.rd_counts.coef_counts[tx_size];
+  unsigned int (*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
+      cpi->common.counts.eob_branch[tx_size];
+  int i, j, k, l, m;
+
+  for (i = 0; i < PLANE_TYPES; ++i) {
+    for (j = 0; j < REF_TYPES; ++j) {
+      for (k = 0; k < COEF_BANDS; ++k) {
+        for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+          vp9_tree_probs_from_distribution(vp9_coef_tree,
+                                           coef_branch_ct[i][j][k][l],
+                                           coef_counts[i][j][k][l]);
+          coef_branch_ct[i][j][k][l][0][1] = eob_branch_ct[i][j][k][l] -
+                                             coef_branch_ct[i][j][k][l][0][0];
+          for (m = 0; m < UNCONSTRAINED_NODES; ++m)
+            coef_probs[i][j][k][l][m] = get_binary_prob(
+                                            coef_branch_ct[i][j][k][l][m][0],
+                                            coef_branch_ct[i][j][k][l][m][1]);
+        }
+      }
+    }
+  }
+}
+
+static void update_coef_probs_common(vpx_writer* const bc, VP9_COMP *cpi,
+                                     TX_SIZE tx_size,
+                                     vp9_coeff_stats *frame_branch_ct,
+                                     vp9_coeff_probs_model *new_coef_probs) {
+  vp9_coeff_probs_model *old_coef_probs = cpi->common.fc->coef_probs[tx_size];
+  const vpx_prob upd = DIFF_UPDATE_PROB;
+  const int entropy_nodes_update = UNCONSTRAINED_NODES;
+  int i, j, k, l, t;
+  int stepsize = cpi->sf.coeff_prob_appx_step;
+
+  switch (cpi->sf.use_fast_coef_updates) {
+    case TWO_LOOP: {
+      /* dry run to see if there is any update at all needed */
+      int savings = 0;
+      int update[2] = {0, 0};
+      for (i = 0; i < PLANE_TYPES; ++i) {
+        for (j = 0; j < REF_TYPES; ++j) {
+          for (k = 0; k < COEF_BANDS; ++k) {
+            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+              for (t = 0; t < entropy_nodes_update; ++t) {
+                vpx_prob newp = new_coef_probs[i][j][k][l][t];
+                const vpx_prob oldp = old_coef_probs[i][j][k][l][t];
+                int s;
+                int u = 0;
+                if (t == PIVOT_NODE)
+                  s = vp9_prob_diff_update_savings_search_model(
+                      frame_branch_ct[i][j][k][l][0], oldp, &newp, upd,
+                      stepsize);
+                else
+                  s = vp9_prob_diff_update_savings_search(
+                      frame_branch_ct[i][j][k][l][t], oldp, &newp, upd);
+                if (s > 0 && newp != oldp)
+                  u = 1;
+                if (u)
+                  savings += s - (int)(vp9_cost_zero(upd));
+                else
+                  savings -= (int)(vp9_cost_zero(upd));
+                update[u]++;
+              }
+            }
+          }
+        }
+      }
+
+      // printf("Update %d %d, savings %d\n", update[0], update[1], savings);
+      /* Is coef updated at all */
+      if (update[1] == 0 || savings < 0) {
+        vpx_write_bit(bc, 0);
+        return;
+      }
+      vpx_write_bit(bc, 1);
+      for (i = 0; i < PLANE_TYPES; ++i) {
+        for (j = 0; j < REF_TYPES; ++j) {
+          for (k = 0; k < COEF_BANDS; ++k) {
+            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+              // calc probs and branch cts for this frame only
+              for (t = 0; t < entropy_nodes_update; ++t) {
+                vpx_prob newp = new_coef_probs[i][j][k][l][t];
+                vpx_prob *oldp = old_coef_probs[i][j][k][l] + t;
+                const vpx_prob upd = DIFF_UPDATE_PROB;
+                int s;
+                int u = 0;
+                if (t == PIVOT_NODE)
+                  s = vp9_prob_diff_update_savings_search_model(
+                      frame_branch_ct[i][j][k][l][0],
+                      *oldp, &newp, upd, stepsize);
+                else
+                  s = vp9_prob_diff_update_savings_search(
+                      frame_branch_ct[i][j][k][l][t],
+                      *oldp, &newp, upd);
+                if (s > 0 && newp != *oldp)
+                  u = 1;
+                vpx_write(bc, u, upd);
+                if (u) {
+                  /* send/use new probability */
+                  vp9_write_prob_diff_update(bc, newp, *oldp);
+                  *oldp = newp;
+                }
+              }
+            }
+          }
+        }
+      }
+      return;
+    }
+
+    case ONE_LOOP_REDUCED: {
+      int updates = 0;
+      int noupdates_before_first = 0;
+      for (i = 0; i < PLANE_TYPES; ++i) {
+        for (j = 0; j < REF_TYPES; ++j) {
+          for (k = 0; k < COEF_BANDS; ++k) {
+            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+              // calc probs and branch cts for this frame only
+              for (t = 0; t < entropy_nodes_update; ++t) {
+                vpx_prob newp = new_coef_probs[i][j][k][l][t];
+                vpx_prob *oldp = old_coef_probs[i][j][k][l] + t;
+                int s;
+                int u = 0;
+
+                if (t == PIVOT_NODE) {
+                  s = vp9_prob_diff_update_savings_search_model(
+                      frame_branch_ct[i][j][k][l][0],
+                      *oldp, &newp, upd, stepsize);
+                } else {
+                  s = vp9_prob_diff_update_savings_search(
+                      frame_branch_ct[i][j][k][l][t],
+                      *oldp, &newp, upd);
+                }
+
+                if (s > 0 && newp != *oldp)
+                  u = 1;
+                updates += u;
+                if (u == 0 && updates == 0) {
+                  noupdates_before_first++;
+                  continue;
+                }
+                if (u == 1 && updates == 1) {
+                  int v;
+                  // first update
+                  vpx_write_bit(bc, 1);
+                  for (v = 0; v < noupdates_before_first; ++v)
+                    vpx_write(bc, 0, upd);
+                }
+                vpx_write(bc, u, upd);
+                if (u) {
+                  /* send/use new probability */
+                  vp9_write_prob_diff_update(bc, newp, *oldp);
+                  *oldp = newp;
+                }
+              }
+            }
+          }
+        }
+      }
+      if (updates == 0) {
+        vpx_write_bit(bc, 0);  // no updates
+      }
+      return;
+    }
+    default:
+      assert(0);
+  }
+}
+
+static void update_coef_probs(VP9_COMP *cpi, vpx_writer* w) {
+  const TX_MODE tx_mode = cpi->common.tx_mode;
+  const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+  TX_SIZE tx_size;
+  for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size) {
+    vp9_coeff_stats frame_branch_ct[PLANE_TYPES];
+    vp9_coeff_probs_model frame_coef_probs[PLANE_TYPES];
+    if (cpi->td.counts->tx.tx_totals[tx_size] <= 20 ||
+        (tx_size >= TX_16X16 && cpi->sf.tx_size_search_method == USE_TX_8X8)) {
+      vpx_write_bit(w, 0);
+    } else {
+      build_tree_distribution(cpi, tx_size, frame_branch_ct,
+                              frame_coef_probs);
+      update_coef_probs_common(w, cpi, tx_size, frame_branch_ct,
+                               frame_coef_probs);
+    }
+  }
+}
+
+static void encode_loopfilter(struct loopfilter *lf,
+                              struct vpx_write_bit_buffer *wb) {
+  int i;
+
+  // Encode the loop filter level and type
+  vpx_wb_write_literal(wb, lf->filter_level, 6);
+  vpx_wb_write_literal(wb, lf->sharpness_level, 3);
+
+  // Write out loop filter deltas applied at the MB level based on mode or
+  // ref frame (if they are enabled).
+  vpx_wb_write_bit(wb, lf->mode_ref_delta_enabled);
+
+  if (lf->mode_ref_delta_enabled) {
+    vpx_wb_write_bit(wb, lf->mode_ref_delta_update);
+    if (lf->mode_ref_delta_update) {
+      for (i = 0; i < MAX_REF_LF_DELTAS; i++) {
+        const int delta = lf->ref_deltas[i];
+        const int changed = delta != lf->last_ref_deltas[i];
+        vpx_wb_write_bit(wb, changed);
+        if (changed) {
+          lf->last_ref_deltas[i] = delta;
+          vpx_wb_write_literal(wb, abs(delta) & 0x3F, 6);
+          vpx_wb_write_bit(wb, delta < 0);
+        }
+      }
+
+      for (i = 0; i < MAX_MODE_LF_DELTAS; i++) {
+        const int delta = lf->mode_deltas[i];
+        const int changed = delta != lf->last_mode_deltas[i];
+        vpx_wb_write_bit(wb, changed);
+        if (changed) {
+          lf->last_mode_deltas[i] = delta;
+          vpx_wb_write_literal(wb, abs(delta) & 0x3F, 6);
+          vpx_wb_write_bit(wb, delta < 0);
+        }
+      }
+    }
+  }
+}
+
+static void write_delta_q(struct vpx_write_bit_buffer *wb, int delta_q) {
+  if (delta_q != 0) {
+    vpx_wb_write_bit(wb, 1);
+    vpx_wb_write_literal(wb, abs(delta_q), 4);
+    vpx_wb_write_bit(wb, delta_q < 0);
+  } else {
+    vpx_wb_write_bit(wb, 0);
+  }
+}
+
+static void encode_quantization(const VP9_COMMON *const cm,
+                                struct vpx_write_bit_buffer *wb) {
+  vpx_wb_write_literal(wb, cm->base_qindex, QINDEX_BITS);
+  write_delta_q(wb, cm->y_dc_delta_q);
+  write_delta_q(wb, cm->uv_dc_delta_q);
+  write_delta_q(wb, cm->uv_ac_delta_q);
+}
+
+static void encode_segmentation(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                struct vpx_write_bit_buffer *wb) {
+  int i, j;
+
+  const struct segmentation *seg = &cm->seg;
+
+  vpx_wb_write_bit(wb, seg->enabled);
+  if (!seg->enabled)
+    return;
+
+  // Segmentation map
+  vpx_wb_write_bit(wb, seg->update_map);
+  if (seg->update_map) {
+    // Select the coding strategy (temporal or spatial)
+    vp9_choose_segmap_coding_method(cm, xd);
+    // Write out probabilities used to decode unpredicted  macro-block segments
+    for (i = 0; i < SEG_TREE_PROBS; i++) {
+      const int prob = seg->tree_probs[i];
+      const int update = prob != MAX_PROB;
+      vpx_wb_write_bit(wb, update);
+      if (update)
+        vpx_wb_write_literal(wb, prob, 8);
+    }
+
+    // Write out the chosen coding method.
+    vpx_wb_write_bit(wb, seg->temporal_update);
+    if (seg->temporal_update) {
+      for (i = 0; i < PREDICTION_PROBS; i++) {
+        const int prob = seg->pred_probs[i];
+        const int update = prob != MAX_PROB;
+        vpx_wb_write_bit(wb, update);
+        if (update)
+          vpx_wb_write_literal(wb, prob, 8);
+      }
+    }
+  }
+
+  // Segmentation data
+  vpx_wb_write_bit(wb, seg->update_data);
+  if (seg->update_data) {
+    vpx_wb_write_bit(wb, seg->abs_delta);
+
+    for (i = 0; i < MAX_SEGMENTS; i++) {
+      for (j = 0; j < SEG_LVL_MAX; j++) {
+        const int active = segfeature_active(seg, i, j);
+        vpx_wb_write_bit(wb, active);
+        if (active) {
+          const int data = get_segdata(seg, i, j);
+          const int data_max = vp9_seg_feature_data_max(j);
+
+          if (vp9_is_segfeature_signed(j)) {
+            encode_unsigned_max(wb, abs(data), data_max);
+            vpx_wb_write_bit(wb, data < 0);
+          } else {
+            encode_unsigned_max(wb, data, data_max);
+          }
+        }
+      }
+    }
+  }
+}
+
+static void encode_txfm_probs(VP9_COMMON *cm, vpx_writer *w,
+                              FRAME_COUNTS *counts) {
+  // Mode
+  vpx_write_literal(w, VPXMIN(cm->tx_mode, ALLOW_32X32), 2);
+  if (cm->tx_mode >= ALLOW_32X32)
+    vpx_write_bit(w, cm->tx_mode == TX_MODE_SELECT);
+
+  // Probabilities
+  if (cm->tx_mode == TX_MODE_SELECT) {
+    int i, j;
+    unsigned int ct_8x8p[TX_SIZES - 3][2];
+    unsigned int ct_16x16p[TX_SIZES - 2][2];
+    unsigned int ct_32x32p[TX_SIZES - 1][2];
+
+
+    for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
+      tx_counts_to_branch_counts_8x8(counts->tx.p8x8[i], ct_8x8p);
+      for (j = 0; j < TX_SIZES - 3; j++)
+        vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p8x8[i][j], ct_8x8p[j]);
+    }
+
+    for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
+      tx_counts_to_branch_counts_16x16(counts->tx.p16x16[i], ct_16x16p);
+      for (j = 0; j < TX_SIZES - 2; j++)
+        vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p16x16[i][j],
+                                  ct_16x16p[j]);
+    }
+
+    for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
+      tx_counts_to_branch_counts_32x32(counts->tx.p32x32[i], ct_32x32p);
+      for (j = 0; j < TX_SIZES - 1; j++)
+        vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p32x32[i][j],
+                                  ct_32x32p[j]);
+    }
+  }
+}
+
+static void write_interp_filter(INTERP_FILTER filter,
+                                struct vpx_write_bit_buffer *wb) {
+  const int filter_to_literal[] = { 1, 0, 2, 3 };
+
+  vpx_wb_write_bit(wb, filter == SWITCHABLE);
+  if (filter != SWITCHABLE)
+    vpx_wb_write_literal(wb, filter_to_literal[filter], 2);
+}
+
+static void fix_interp_filter(VP9_COMMON *cm, FRAME_COUNTS *counts) {
+  if (cm->interp_filter == SWITCHABLE) {
+    // Check to see if only one of the filters is actually used
+    int count[SWITCHABLE_FILTERS];
+    int i, j, c = 0;
+    for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
+      count[i] = 0;
+      for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
+        count[i] += counts->switchable_interp[j][i];
+      c += (count[i] > 0);
+    }
+    if (c == 1) {
+      // Only one filter is used. So set the filter at frame level
+      for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
+        if (count[i]) {
+          cm->interp_filter = i;
+          break;
+        }
+      }
+    }
+  }
+}
+
+static void write_tile_info(const VP9_COMMON *const cm,
+                            struct vpx_write_bit_buffer *wb) {
+  int min_log2_tile_cols, max_log2_tile_cols, ones;
+  vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
+
+  // columns
+  ones = cm->log2_tile_cols - min_log2_tile_cols;
+  while (ones--)
+    vpx_wb_write_bit(wb, 1);
+
+  if (cm->log2_tile_cols < max_log2_tile_cols)
+    vpx_wb_write_bit(wb, 0);
+
+  // rows
+  vpx_wb_write_bit(wb, cm->log2_tile_rows != 0);
+  if (cm->log2_tile_rows != 0)
+    vpx_wb_write_bit(wb, cm->log2_tile_rows != 1);
+}
+
+int vp9_get_refresh_mask(VP9_COMP *cpi) {
+  if (vp9_preserve_existing_gf(cpi)) {
+    // We have decided to preserve the previously existing golden frame as our
+    // new ARF frame. However, in the short term we leave it in the GF slot and,
+    // if we're updating the GF with the current decoded frame, we save it
+    // instead to the ARF slot.
+    // Later, in the function vp9_encoder.c:vp9_update_reference_frames() we
+    // will swap gld_fb_idx and alt_fb_idx to achieve our objective. We do it
+    // there so that it can be done outside of the recode loop.
+    // Note: This is highly specific to the use of ARF as a forward reference,
+    // and this needs to be generalized as other uses are implemented
+    // (like RTC/temporal scalability).
+    return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
+           (cpi->refresh_golden_frame << cpi->alt_fb_idx);
+  } else {
+    int arf_idx = cpi->alt_fb_idx;
+    if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
+      const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+      arf_idx = gf_group->arf_update_idx[gf_group->index];
+    }
+    return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
+           (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
+           (cpi->refresh_alt_ref_frame << arf_idx);
+  }
+}
+
+static size_t encode_tiles(VP9_COMP *cpi, uint8_t *data_ptr) {
+  VP9_COMMON *const cm = &cpi->common;
+  vpx_writer residual_bc;
+  int tile_row, tile_col;
+  TOKENEXTRA *tok_end;
+  size_t total_size = 0;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+
+  memset(cm->above_seg_context, 0,
+         sizeof(*cm->above_seg_context) * mi_cols_aligned_to_sb(cm->mi_cols));
+
+  for (tile_row = 0; tile_row < tile_rows; tile_row++) {
+    for (tile_col = 0; tile_col < tile_cols; tile_col++) {
+      int tile_idx = tile_row * tile_cols + tile_col;
+      TOKENEXTRA *tok = cpi->tile_tok[tile_row][tile_col];
+
+      tok_end = cpi->tile_tok[tile_row][tile_col] +
+          cpi->tok_count[tile_row][tile_col];
+
+      if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1)
+        vpx_start_encode(&residual_bc, data_ptr + total_size + 4);
+      else
+        vpx_start_encode(&residual_bc, data_ptr + total_size);
+
+      write_modes(cpi, &cpi->tile_data[tile_idx].tile_info,
+                  &residual_bc, &tok, tok_end);
+      assert(tok == tok_end);
+      vpx_stop_encode(&residual_bc);
+      if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1) {
+        // size of this tile
+        mem_put_be32(data_ptr + total_size, residual_bc.pos);
+        total_size += 4;
+      }
+
+      total_size += residual_bc.pos;
+    }
+  }
+
+  return total_size;
+}
+
+static void write_render_size(const VP9_COMMON *cm,
+                              struct vpx_write_bit_buffer *wb) {
+  const int scaling_active = cm->width != cm->render_width ||
+                             cm->height != cm->render_height;
+  vpx_wb_write_bit(wb, scaling_active);
+  if (scaling_active) {
+    vpx_wb_write_literal(wb, cm->render_width - 1, 16);
+    vpx_wb_write_literal(wb, cm->render_height - 1, 16);
+  }
+}
+
+static void write_frame_size(const VP9_COMMON *cm,
+                             struct vpx_write_bit_buffer *wb) {
+  vpx_wb_write_literal(wb, cm->width - 1, 16);
+  vpx_wb_write_literal(wb, cm->height - 1, 16);
+
+  write_render_size(cm, wb);
+}
+
+static void write_frame_size_with_refs(VP9_COMP *cpi,
+                                       struct vpx_write_bit_buffer *wb) {
+  VP9_COMMON *const cm = &cpi->common;
+  int found = 0;
+
+  MV_REFERENCE_FRAME ref_frame;
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, ref_frame);
+
+    // Set "found" to 0 for temporal svc and for spatial svc key frame
+    if (cpi->use_svc &&
+        ((cpi->svc.number_temporal_layers > 1 &&
+         cpi->oxcf.rc_mode == VPX_CBR) ||
+        (cpi->svc.number_spatial_layers > 1 &&
+         cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame) ||
+        (is_two_pass_svc(cpi) &&
+         cpi->svc.encode_empty_frame_state == ENCODING &&
+         cpi->svc.layer_context[0].frames_from_key_frame <
+         cpi->svc.number_temporal_layers + 1))) {
+      found = 0;
+    } else if (cfg != NULL) {
+      found = cm->width == cfg->y_crop_width &&
+              cm->height == cfg->y_crop_height;
+    }
+    vpx_wb_write_bit(wb, found);
+    if (found) {
+      break;
+    }
+  }
+
+  if (!found) {
+    vpx_wb_write_literal(wb, cm->width - 1, 16);
+    vpx_wb_write_literal(wb, cm->height - 1, 16);
+  }
+
+  write_render_size(cm, wb);
+}
+
+static void write_sync_code(struct vpx_write_bit_buffer *wb) {
+  vpx_wb_write_literal(wb, VP9_SYNC_CODE_0, 8);
+  vpx_wb_write_literal(wb, VP9_SYNC_CODE_1, 8);
+  vpx_wb_write_literal(wb, VP9_SYNC_CODE_2, 8);
+}
+
+static void write_profile(BITSTREAM_PROFILE profile,
+                          struct vpx_write_bit_buffer *wb) {
+  switch (profile) {
+    case PROFILE_0:
+      vpx_wb_write_literal(wb, 0, 2);
+      break;
+    case PROFILE_1:
+      vpx_wb_write_literal(wb, 2, 2);
+      break;
+    case PROFILE_2:
+      vpx_wb_write_literal(wb, 1, 2);
+      break;
+    case PROFILE_3:
+      vpx_wb_write_literal(wb, 6, 3);
+      break;
+    default:
+      assert(0);
+  }
+}
+
+static void write_bitdepth_colorspace_sampling(
+    VP9_COMMON *const cm, struct vpx_write_bit_buffer *wb) {
+  if (cm->profile >= PROFILE_2) {
+    assert(cm->bit_depth > VPX_BITS_8);
+    vpx_wb_write_bit(wb, cm->bit_depth == VPX_BITS_10 ? 0 : 1);
+  }
+  vpx_wb_write_literal(wb, cm->color_space, 3);
+  if (cm->color_space != VPX_CS_SRGB) {
+    // 0: [16, 235] (i.e. xvYCC), 1: [0, 255]
+    vpx_wb_write_bit(wb, cm->color_range);
+    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
+      assert(cm->subsampling_x != 1 || cm->subsampling_y != 1);
+      vpx_wb_write_bit(wb, cm->subsampling_x);
+      vpx_wb_write_bit(wb, cm->subsampling_y);
+      vpx_wb_write_bit(wb, 0);  // unused
+    } else {
+      assert(cm->subsampling_x == 1 && cm->subsampling_y == 1);
+    }
+  } else {
+    assert(cm->profile == PROFILE_1 || cm->profile == PROFILE_3);
+    vpx_wb_write_bit(wb, 0);  // unused
+  }
+}
+
+static void write_uncompressed_header(VP9_COMP *cpi,
+                                      struct vpx_write_bit_buffer *wb) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+
+  vpx_wb_write_literal(wb, VP9_FRAME_MARKER, 2);
+
+  write_profile(cm->profile, wb);
+
+  vpx_wb_write_bit(wb, 0);  // show_existing_frame
+  vpx_wb_write_bit(wb, cm->frame_type);
+  vpx_wb_write_bit(wb, cm->show_frame);
+  vpx_wb_write_bit(wb, cm->error_resilient_mode);
+
+  if (cm->frame_type == KEY_FRAME) {
+    write_sync_code(wb);
+    write_bitdepth_colorspace_sampling(cm, wb);
+    write_frame_size(cm, wb);
+  } else {
+    // In spatial svc if it's not error_resilient_mode then we need to code all
+    // visible frames as invisible. But we need to keep the show_frame flag so
+    // that the publisher could know whether it is supposed to be visible.
+    // So we will code the show_frame flag as it is. Then code the intra_only
+    // bit here. This will make the bitstream incompatible. In the player we
+    // will change to show_frame flag to 0, then add an one byte frame with
+    // show_existing_frame flag which tells the decoder which frame we want to
+    // show.
+    if (!cm->show_frame)
+      vpx_wb_write_bit(wb, cm->intra_only);
+
+    if (!cm->error_resilient_mode)
+      vpx_wb_write_literal(wb, cm->reset_frame_context, 2);
+
+    if (cm->intra_only) {
+      write_sync_code(wb);
+
+      // Note for profile 0, 420 8bpp is assumed.
+      if (cm->profile > PROFILE_0) {
+        write_bitdepth_colorspace_sampling(cm, wb);
+      }
+
+      vpx_wb_write_literal(wb, vp9_get_refresh_mask(cpi), REF_FRAMES);
+      write_frame_size(cm, wb);
+    } else {
+      MV_REFERENCE_FRAME ref_frame;
+      vpx_wb_write_literal(wb, vp9_get_refresh_mask(cpi), REF_FRAMES);
+      for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+        assert(get_ref_frame_map_idx(cpi, ref_frame) != INVALID_IDX);
+        vpx_wb_write_literal(wb, get_ref_frame_map_idx(cpi, ref_frame),
+                             REF_FRAMES_LOG2);
+        vpx_wb_write_bit(wb, cm->ref_frame_sign_bias[ref_frame]);
+      }
+
+      write_frame_size_with_refs(cpi, wb);
+
+      vpx_wb_write_bit(wb, cm->allow_high_precision_mv);
+
+      fix_interp_filter(cm, cpi->td.counts);
+      write_interp_filter(cm->interp_filter, wb);
+    }
+  }
+
+  if (!cm->error_resilient_mode) {
+    vpx_wb_write_bit(wb, cm->refresh_frame_context);
+    vpx_wb_write_bit(wb, cm->frame_parallel_decoding_mode);
+  }
+
+  vpx_wb_write_literal(wb, cm->frame_context_idx, FRAME_CONTEXTS_LOG2);
+
+  encode_loopfilter(&cm->lf, wb);
+  encode_quantization(cm, wb);
+  encode_segmentation(cm, xd, wb);
+
+  write_tile_info(cm, wb);
+}
+
+static size_t write_compressed_header(VP9_COMP *cpi, uint8_t *data) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+  FRAME_CONTEXT *const fc = cm->fc;
+  FRAME_COUNTS *counts = cpi->td.counts;
+  vpx_writer header_bc;
+
+  vpx_start_encode(&header_bc, data);
+
+  if (xd->lossless)
+    cm->tx_mode = ONLY_4X4;
+  else
+    encode_txfm_probs(cm, &header_bc, counts);
+
+  update_coef_probs(cpi, &header_bc);
+  update_skip_probs(cm, &header_bc, counts);
+
+  if (!frame_is_intra_only(cm)) {
+    int i;
+
+    for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
+      prob_diff_update(vp9_inter_mode_tree, cm->fc->inter_mode_probs[i],
+                       counts->inter_mode[i], INTER_MODES, &header_bc);
+
+    if (cm->interp_filter == SWITCHABLE)
+      update_switchable_interp_probs(cm, &header_bc, counts);
+
+    for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+      vp9_cond_prob_diff_update(&header_bc, &fc->intra_inter_prob[i],
+                                counts->intra_inter[i]);
+
+    if (cpi->allow_comp_inter_inter) {
+      const int use_compound_pred = cm->reference_mode != SINGLE_REFERENCE;
+      const int use_hybrid_pred = cm->reference_mode == REFERENCE_MODE_SELECT;
+
+      vpx_write_bit(&header_bc, use_compound_pred);
+      if (use_compound_pred) {
+        vpx_write_bit(&header_bc, use_hybrid_pred);
+        if (use_hybrid_pred)
+          for (i = 0; i < COMP_INTER_CONTEXTS; i++)
+            vp9_cond_prob_diff_update(&header_bc, &fc->comp_inter_prob[i],
+                                      counts->comp_inter[i]);
+      }
+    }
+
+    if (cm->reference_mode != COMPOUND_REFERENCE) {
+      for (i = 0; i < REF_CONTEXTS; i++) {
+        vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][0],
+                                  counts->single_ref[i][0]);
+        vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][1],
+                                  counts->single_ref[i][1]);
+      }
+    }
+
+    if (cm->reference_mode != SINGLE_REFERENCE)
+      for (i = 0; i < REF_CONTEXTS; i++)
+        vp9_cond_prob_diff_update(&header_bc, &fc->comp_ref_prob[i],
+                                  counts->comp_ref[i]);
+
+    for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
+      prob_diff_update(vp9_intra_mode_tree, cm->fc->y_mode_prob[i],
+                       counts->y_mode[i], INTRA_MODES, &header_bc);
+
+    for (i = 0; i < PARTITION_CONTEXTS; ++i)
+      prob_diff_update(vp9_partition_tree, fc->partition_prob[i],
+                       counts->partition[i], PARTITION_TYPES, &header_bc);
+
+    vp9_write_nmv_probs(cm, cm->allow_high_precision_mv, &header_bc,
+                        &counts->mv);
+  }
+
+  vpx_stop_encode(&header_bc);
+  assert(header_bc.pos <= 0xffff);
+
+  return header_bc.pos;
+}
+
+void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t *size) {
+  uint8_t *data = dest;
+  size_t first_part_size, uncompressed_hdr_size;
+  struct vpx_write_bit_buffer wb = {data, 0};
+  struct vpx_write_bit_buffer saved_wb;
+
+  write_uncompressed_header(cpi, &wb);
+  saved_wb = wb;
+  vpx_wb_write_literal(&wb, 0, 16);  // don't know in advance first part. size
+
+  uncompressed_hdr_size = vpx_wb_bytes_written(&wb);
+  data += uncompressed_hdr_size;
+
+  vpx_clear_system_state();
+
+  first_part_size = write_compressed_header(cpi, data);
+  data += first_part_size;
+  // TODO(jbb): Figure out what to do if first_part_size > 16 bits.
+  vpx_wb_write_literal(&saved_wb, (int)first_part_size, 16);
+
+  data += encode_tiles(cpi, data);
+
+  *size = data - dest;
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_bitstream.h b/libvpx/libvpx/vp9/encoder/vp9_bitstream.h
new file mode 100644
index 0000000..f24d20f
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_bitstream.h
@@ -0,0 +1,39 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_BITSTREAM_H_
+#define VP9_ENCODER_VP9_BITSTREAM_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "vp9/encoder/vp9_encoder.h"
+
+int vp9_get_refresh_mask(VP9_COMP *cpi);
+
+void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t *size);
+
+static INLINE int vp9_preserve_existing_gf(VP9_COMP *cpi) {
+  return !cpi->multi_arf_allowed && cpi->refresh_golden_frame &&
+         cpi->rc.is_src_frame_alt_ref &&
+         (!cpi->use_svc ||      // Add spatial svc base layer case here
+          (is_two_pass_svc(cpi) &&
+           cpi->svc.spatial_layer_id == 0 &&
+           cpi->svc.layer_context[0].gold_ref_idx >=0 &&
+           cpi->oxcf.ss_enable_auto_arf[0]));
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_BITSTREAM_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_block.h b/libvpx/libvpx/vp9/encoder/vp9_block.h
new file mode 100644
index 0000000..069c335
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_block.h
@@ -0,0 +1,165 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_BLOCK_H_
+#define VP9_ENCODER_VP9_BLOCK_H_
+
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_entropy.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  unsigned int sse;
+  int sum;
+  unsigned int var;
+} diff;
+
+struct macroblock_plane {
+  DECLARE_ALIGNED(16, int16_t, src_diff[64 * 64]);
+  tran_low_t *qcoeff;
+  tran_low_t *coeff;
+  uint16_t *eobs;
+  struct buf_2d src;
+
+  // Quantizer setings
+  int16_t *quant_fp;
+  int16_t *round_fp;
+  int16_t *quant;
+  int16_t *quant_shift;
+  int16_t *zbin;
+  int16_t *round;
+
+  int64_t quant_thred[2];
+};
+
+/* The [2] dimension is for whether we skip the EOB node (i.e. if previous
+ * coefficient in this block was zero) or not. */
+typedef unsigned int vp9_coeff_cost[PLANE_TYPES][REF_TYPES][COEF_BANDS][2]
+                                   [COEFF_CONTEXTS][ENTROPY_TOKENS];
+
+typedef struct {
+  int_mv ref_mvs[MAX_REF_FRAMES][MAX_MV_REF_CANDIDATES];
+  uint8_t mode_context[MAX_REF_FRAMES];
+} MB_MODE_INFO_EXT;
+
+typedef struct macroblock MACROBLOCK;
+struct macroblock {
+  struct macroblock_plane plane[MAX_MB_PLANE];
+
+  MACROBLOCKD e_mbd;
+  MB_MODE_INFO_EXT *mbmi_ext;
+  MB_MODE_INFO_EXT *mbmi_ext_base;
+  int skip_block;
+  int select_tx_size;
+  int skip_recode;
+  int skip_optimize;
+  int q_index;
+
+  // The equivalent error at the current rdmult of one whole bit (not one
+  // bitcost unit).
+  int errorperbit;
+  // The equivalend SAD error of one (whole) bit at the current quantizer
+  // for large blocks.
+  int sadperbit16;
+  // The equivalend SAD error of one (whole) bit at the current quantizer
+  // for sub-8x8 blocks.
+  int sadperbit4;
+  int rddiv;
+  int rdmult;
+  int mb_energy;
+  int * m_search_count_ptr;
+  int * ex_search_count_ptr;
+
+  // These are set to their default values at the beginning, and then adjusted
+  // further in the encoding process.
+  BLOCK_SIZE min_partition_size;
+  BLOCK_SIZE max_partition_size;
+
+  int mv_best_ref_index[MAX_REF_FRAMES];
+  unsigned int max_mv_context[MAX_REF_FRAMES];
+  unsigned int source_variance;
+  unsigned int pred_sse[MAX_REF_FRAMES];
+  int pred_mv_sad[MAX_REF_FRAMES];
+
+  int nmvjointcost[MV_JOINTS];
+  int *nmvcost[2];
+  int *nmvcost_hp[2];
+  int **mvcost;
+
+  int nmvjointsadcost[MV_JOINTS];
+  int *nmvsadcost[2];
+  int *nmvsadcost_hp[2];
+  int **mvsadcost;
+
+  // These define limits to motion vector components to prevent them
+  // from extending outside the UMV borders
+  int mv_col_min;
+  int mv_col_max;
+  int mv_row_min;
+  int mv_row_max;
+
+  // Notes transform blocks where no coefficents are coded.
+  // Set during mode selection. Read during block encoding.
+  uint8_t zcoeff_blk[TX_SIZES][256];
+
+  int skip;
+
+  int encode_breakout;
+
+  // note that token_costs is the cost when eob node is skipped
+  vp9_coeff_cost token_costs[TX_SIZES];
+
+  int optimize;
+
+  // indicate if it is in the rd search loop or encoding process
+  int use_lp32x32fdct;
+  int skip_encode;
+
+  // use fast quantization process
+  int quant_fp;
+
+  // skip forward transform and quantization
+  uint8_t skip_txfm[MAX_MB_PLANE << 2];
+  #define SKIP_TXFM_NONE 0
+  #define SKIP_TXFM_AC_DC 1
+  #define SKIP_TXFM_AC_ONLY 2
+
+  int64_t bsse[MAX_MB_PLANE << 2];
+
+  // Used to store sub partition's choices.
+  MV pred_mv[MAX_REF_FRAMES];
+
+  // Strong color activity detection. Used in RTC coding mode to enhance
+  // the visual quality at the boundary of moving color objects.
+  uint8_t color_sensitivity[2];
+
+  uint8_t sb_is_skin;
+
+  // Used to save the status of whether a block has a low variance in
+  // choose_partitioning. 0 for 64x64, 1~2 for 64x32, 3~4 for 32x64, 5~8 for
+  // 32x32, 9~24 for 16x16.
+  uint8_t variance_low[25];
+
+  void (*fwd_txm4x4)(const int16_t *input, tran_low_t *output, int stride);
+  void (*itxm_add)(const tran_low_t *input, uint8_t *dest, int stride, int eob);
+#if CONFIG_VP9_HIGHBITDEPTH
+  void (*highbd_itxm_add)(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, int bd);
+#endif
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_BLOCK_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_blockiness.c b/libvpx/libvpx/vp9/encoder/vp9_blockiness.c
new file mode 100644
index 0000000..1a89ce4
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_blockiness.c
@@ -0,0 +1,134 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <stdlib.h>
+
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/system_state.h"
+
+static int horizontal_filter(const uint8_t *s) {
+  return (s[1] - s[-2]) * 2 + (s[-1] - s[0]) * 6;
+}
+
+static int vertical_filter(const uint8_t *s, int p) {
+  return (s[p] - s[-2 * p]) * 2 + (s[-p] - s[0]) * 6;
+}
+
+static int variance(int sum, int sum_squared, int size) {
+  return sum_squared / size - (sum / size) * (sum / size);
+}
+// Calculate a blockiness level for a vertical block edge.
+// This function returns a new blockiness metric that's defined as
+
+//              p0 p1 p2 p3
+//              q0 q1 q2 q3
+// block edge ->
+//              r0 r1 r2 r3
+//              s0 s1 s2 s3
+
+// blockiness =  p0*-2+q0*6+r0*-6+s0*2 +
+//               p1*-2+q1*6+r1*-6+s1*2 +
+//               p2*-2+q2*6+r2*-6+s2*2 +
+//               p3*-2+q3*6+r3*-6+s3*2 ;
+
+// reconstructed_blockiness = abs(blockiness from reconstructed buffer -
+//                                blockiness from source buffer,0)
+//
+// I make the assumption that flat blocks are much more visible than high
+// contrast blocks. As such, I scale the result of the blockiness calc
+// by dividing the blockiness by the variance of the pixels on either side
+// of the edge as follows:
+// var_0 = (q0^2+q1^2+q2^2+q3^2) - ((q0 + q1 + q2 + q3) / 4 )^2
+// var_1 = (r0^2+r1^2+r2^2+r3^2) - ((r0 + r1 + r2 + r3) / 4 )^2
+// The returned blockiness is the scaled value
+// Reconstructed blockiness / ( 1 + var_0 + var_1 ) ;
+static int blockiness_vertical(const uint8_t *s, int sp, const uint8_t *r,
+                               int rp, int size) {
+  int s_blockiness = 0;
+  int r_blockiness = 0;
+  int sum_0 = 0;
+  int sum_sq_0 = 0;
+  int sum_1 = 0;
+  int sum_sq_1 = 0;
+  int i;
+  int var_0;
+  int var_1;
+  for (i = 0; i < size; ++i, s += sp, r += rp) {
+    s_blockiness += horizontal_filter(s);
+    r_blockiness += horizontal_filter(r);
+    sum_0 += s[0];
+    sum_sq_0 += s[0]*s[0];
+    sum_1 += s[-1];
+    sum_sq_1 += s[-1]*s[-1];
+  }
+  var_0 = variance(sum_0, sum_sq_0, size);
+  var_1 = variance(sum_1, sum_sq_1, size);
+  r_blockiness = abs(r_blockiness);
+  s_blockiness = abs(s_blockiness);
+
+  if (r_blockiness > s_blockiness)
+    return (r_blockiness - s_blockiness) / (1 + var_0 + var_1);
+  else
+    return 0;
+}
+
+// Calculate a blockiness level for a horizontal block edge
+// same as above.
+static int blockiness_horizontal(const uint8_t *s, int sp, const uint8_t *r,
+                                 int rp, int size) {
+  int s_blockiness = 0;
+  int r_blockiness = 0;
+  int sum_0 = 0;
+  int sum_sq_0 = 0;
+  int sum_1 = 0;
+  int sum_sq_1 = 0;
+  int i;
+  int var_0;
+  int var_1;
+  for (i = 0; i < size; ++i, ++s, ++r) {
+    s_blockiness += vertical_filter(s, sp);
+    r_blockiness += vertical_filter(r, rp);
+    sum_0 += s[0];
+    sum_sq_0 += s[0] * s[0];
+    sum_1 += s[-sp];
+    sum_sq_1 += s[-sp] * s[-sp];
+  }
+  var_0 = variance(sum_0, sum_sq_0, size);
+  var_1 = variance(sum_1, sum_sq_1, size);
+  r_blockiness = abs(r_blockiness);
+  s_blockiness = abs(s_blockiness);
+
+  if (r_blockiness > s_blockiness)
+    return (r_blockiness - s_blockiness) / (1 + var_0 + var_1);
+  else
+    return 0;
+}
+
+// This function returns the blockiness for the entire frame currently by
+// looking at all borders in steps of 4.
+double vp9_get_blockiness(const uint8_t *img1, int img1_pitch,
+                          const uint8_t *img2, int img2_pitch,
+                          int width, int height) {
+  double blockiness = 0;
+  int i, j;
+  vpx_clear_system_state();
+  for (i = 0; i < height; i += 4, img1 += img1_pitch * 4,
+       img2 += img2_pitch * 4) {
+    for (j = 0; j < width; j += 4) {
+      if (i > 0 && i < height && j > 0 && j < width) {
+        blockiness += blockiness_vertical(img1 + j, img1_pitch,
+                                          img2 + j, img2_pitch, 4);
+        blockiness += blockiness_horizontal(img1 + j, img1_pitch,
+                                            img2 + j, img2_pitch, 4);
+      }
+    }
+  }
+  blockiness /= width * height / 16;
+  return blockiness;
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_context_tree.c b/libvpx/libvpx/vp9/encoder/vp9_context_tree.c
new file mode 100644
index 0000000..396ed3f
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_context_tree.c
@@ -0,0 +1,161 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_encoder.h"
+
+static const BLOCK_SIZE square[] = {
+  BLOCK_8X8,
+  BLOCK_16X16,
+  BLOCK_32X32,
+  BLOCK_64X64,
+};
+
+static void alloc_mode_context(VP9_COMMON *cm, int num_4x4_blk,
+                               PICK_MODE_CONTEXT *ctx) {
+  const int num_blk = (num_4x4_blk < 4 ? 4 : num_4x4_blk);
+  const int num_pix = num_blk << 4;
+  int i, k;
+  ctx->num_4x4_blk = num_blk;
+
+  CHECK_MEM_ERROR(cm, ctx->zcoeff_blk,
+                  vpx_calloc(num_blk, sizeof(uint8_t)));
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (k = 0; k < 3; ++k) {
+      CHECK_MEM_ERROR(cm, ctx->coeff[i][k],
+                      vpx_memalign(32, num_pix * sizeof(*ctx->coeff[i][k])));
+      CHECK_MEM_ERROR(cm, ctx->qcoeff[i][k],
+                      vpx_memalign(32, num_pix * sizeof(*ctx->qcoeff[i][k])));
+      CHECK_MEM_ERROR(cm, ctx->dqcoeff[i][k],
+                      vpx_memalign(32, num_pix * sizeof(*ctx->dqcoeff[i][k])));
+      CHECK_MEM_ERROR(cm, ctx->eobs[i][k],
+                      vpx_memalign(32, num_blk * sizeof(*ctx->eobs[i][k])));
+      ctx->coeff_pbuf[i][k]   = ctx->coeff[i][k];
+      ctx->qcoeff_pbuf[i][k]  = ctx->qcoeff[i][k];
+      ctx->dqcoeff_pbuf[i][k] = ctx->dqcoeff[i][k];
+      ctx->eobs_pbuf[i][k]    = ctx->eobs[i][k];
+    }
+  }
+}
+
+static void free_mode_context(PICK_MODE_CONTEXT *ctx) {
+  int i, k;
+  vpx_free(ctx->zcoeff_blk);
+  ctx->zcoeff_blk = 0;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (k = 0; k < 3; ++k) {
+      vpx_free(ctx->coeff[i][k]);
+      ctx->coeff[i][k] = 0;
+      vpx_free(ctx->qcoeff[i][k]);
+      ctx->qcoeff[i][k] = 0;
+      vpx_free(ctx->dqcoeff[i][k]);
+      ctx->dqcoeff[i][k] = 0;
+      vpx_free(ctx->eobs[i][k]);
+      ctx->eobs[i][k] = 0;
+    }
+  }
+}
+
+static void alloc_tree_contexts(VP9_COMMON *cm, PC_TREE *tree,
+                                int num_4x4_blk) {
+  alloc_mode_context(cm, num_4x4_blk, &tree->none);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->horizontal[0]);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->vertical[0]);
+
+  if (num_4x4_blk > 4) {
+    alloc_mode_context(cm, num_4x4_blk/2, &tree->horizontal[1]);
+    alloc_mode_context(cm, num_4x4_blk/2, &tree->vertical[1]);
+  } else {
+    memset(&tree->horizontal[1], 0, sizeof(tree->horizontal[1]));
+    memset(&tree->vertical[1], 0, sizeof(tree->vertical[1]));
+  }
+}
+
+static void free_tree_contexts(PC_TREE *tree) {
+  free_mode_context(&tree->none);
+  free_mode_context(&tree->horizontal[0]);
+  free_mode_context(&tree->horizontal[1]);
+  free_mode_context(&tree->vertical[0]);
+  free_mode_context(&tree->vertical[1]);
+}
+
+// This function sets up a tree of contexts such that at each square
+// partition level. There are contexts for none, horizontal, vertical, and
+// split.  Along with a block_size value and a selected block_size which
+// represents the state of our search.
+void vp9_setup_pc_tree(VP9_COMMON *cm, ThreadData *td) {
+  int i, j;
+  const int leaf_nodes = 64;
+  const int tree_nodes = 64 + 16 + 4 + 1;
+  int pc_tree_index = 0;
+  PC_TREE *this_pc;
+  PICK_MODE_CONTEXT *this_leaf;
+  int square_index = 1;
+  int nodes;
+
+  vpx_free(td->leaf_tree);
+  CHECK_MEM_ERROR(cm, td->leaf_tree, vpx_calloc(leaf_nodes,
+                                                sizeof(*td->leaf_tree)));
+  vpx_free(td->pc_tree);
+  CHECK_MEM_ERROR(cm, td->pc_tree, vpx_calloc(tree_nodes,
+                                              sizeof(*td->pc_tree)));
+
+  this_pc = &td->pc_tree[0];
+  this_leaf = &td->leaf_tree[0];
+
+  // 4x4 blocks smaller than 8x8 but in the same 8x8 block share the same
+  // context so we only need to allocate 1 for each 8x8 block.
+  for (i = 0; i < leaf_nodes; ++i)
+    alloc_mode_context(cm, 1, &td->leaf_tree[i]);
+
+  // Sets up all the leaf nodes in the tree.
+  for (pc_tree_index = 0; pc_tree_index < leaf_nodes; ++pc_tree_index) {
+    PC_TREE *const tree = &td->pc_tree[pc_tree_index];
+    tree->block_size = square[0];
+    alloc_tree_contexts(cm, tree, 4);
+    tree->leaf_split[0] = this_leaf++;
+    for (j = 1; j < 4; j++)
+      tree->leaf_split[j] = tree->leaf_split[0];
+  }
+
+  // Each node has 4 leaf nodes, fill each block_size level of the tree
+  // from leafs to the root.
+  for (nodes = 16; nodes > 0; nodes >>= 2) {
+    for (i = 0; i < nodes; ++i) {
+      PC_TREE *const tree = &td->pc_tree[pc_tree_index];
+      alloc_tree_contexts(cm, tree, 4 << (2 * square_index));
+      tree->block_size = square[square_index];
+      for (j = 0; j < 4; j++)
+        tree->split[j] = this_pc++;
+      ++pc_tree_index;
+    }
+    ++square_index;
+  }
+  td->pc_root = &td->pc_tree[tree_nodes - 1];
+  td->pc_root[0].none.best_mode_index = 2;
+}
+
+void vp9_free_pc_tree(ThreadData *td) {
+  const int tree_nodes = 64 + 16 + 4 + 1;
+  int i;
+
+  // Set up all 4x4 mode contexts
+  for (i = 0; i < 64; ++i)
+    free_mode_context(&td->leaf_tree[i]);
+
+  // Sets up all the leaf nodes in the tree.
+  for (i = 0; i < tree_nodes; ++i)
+    free_tree_contexts(&td->pc_tree[i]);
+
+  vpx_free(td->pc_tree);
+  td->pc_tree = NULL;
+  vpx_free(td->leaf_tree);
+  td->leaf_tree = NULL;
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_context_tree.h b/libvpx/libvpx/vp9/encoder/vp9_context_tree.h
new file mode 100644
index 0000000..86ba03d
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_context_tree.h
@@ -0,0 +1,96 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_CONTEXT_TREE_H_
+#define VP9_ENCODER_VP9_CONTEXT_TREE_H_
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/encoder/vp9_block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9_COMP;
+struct VP9Common;
+struct ThreadData;
+
+// Structure to hold snapshot of coding context during the mode picking process
+typedef struct {
+  MODE_INFO mic;
+  MB_MODE_INFO_EXT mbmi_ext;
+  uint8_t *zcoeff_blk;
+  tran_low_t *coeff[MAX_MB_PLANE][3];
+  tran_low_t *qcoeff[MAX_MB_PLANE][3];
+  tran_low_t *dqcoeff[MAX_MB_PLANE][3];
+  uint16_t *eobs[MAX_MB_PLANE][3];
+
+  // dual buffer pointers, 0: in use, 1: best in store
+  tran_low_t *coeff_pbuf[MAX_MB_PLANE][3];
+  tran_low_t *qcoeff_pbuf[MAX_MB_PLANE][3];
+  tran_low_t *dqcoeff_pbuf[MAX_MB_PLANE][3];
+  uint16_t *eobs_pbuf[MAX_MB_PLANE][3];
+
+  int is_coded;
+  int num_4x4_blk;
+  int skip;
+  int pred_pixel_ready;
+  // For current partition, only if all Y, U, and V transform blocks'
+  // coefficients are quantized to 0, skippable is set to 0.
+  int skippable;
+  uint8_t skip_txfm[MAX_MB_PLANE << 2];
+  int best_mode_index;
+  int hybrid_pred_diff;
+  int comp_pred_diff;
+  int single_pred_diff;
+  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+
+  // TODO(jingning) Use RD_COST struct here instead. This involves a boarder
+  // scope of refactoring.
+  int rate;
+  int64_t dist;
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  unsigned int newmv_sse;
+  unsigned int zeromv_sse;
+  unsigned int zeromv_lastref_sse;
+  PREDICTION_MODE best_sse_inter_mode;
+  int_mv best_sse_mv;
+  MV_REFERENCE_FRAME best_reference_frame;
+  MV_REFERENCE_FRAME best_zeromv_reference_frame;
+#endif
+
+  // motion vector cache for adaptive motion search control in partition
+  // search loop
+  MV pred_mv[MAX_REF_FRAMES];
+  INTERP_FILTER pred_interp_filter;
+} PICK_MODE_CONTEXT;
+
+typedef struct PC_TREE {
+  int index;
+  PARTITION_TYPE partitioning;
+  BLOCK_SIZE block_size;
+  PICK_MODE_CONTEXT none;
+  PICK_MODE_CONTEXT horizontal[2];
+  PICK_MODE_CONTEXT vertical[2];
+  union {
+    struct PC_TREE *split[4];
+    PICK_MODE_CONTEXT *leaf_split[4];
+  };
+} PC_TREE;
+
+void vp9_setup_pc_tree(struct VP9Common *cm, struct ThreadData *td);
+void vp9_free_pc_tree(struct ThreadData *td);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif /* VP9_ENCODER_VP9_CONTEXT_TREE_H_ */
diff --git a/libvpx/libvpx/vp9/encoder/vp9_cost.c b/libvpx/libvpx/vp9/encoder/vp9_cost.c
new file mode 100644
index 0000000..5d14742
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_cost.c
@@ -0,0 +1,66 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <assert.h>
+
+#include "vp9/encoder/vp9_cost.h"
+
+/* round(-log2(i/256.) * (1 << VP9_PROB_COST_SHIFT))
+   Begins with a bogus entry for simpler addressing. */
+const uint16_t vp9_prob_cost[256] = {
+    4096, 4096, 3584, 3284, 3072, 2907, 2772, 2659, 2560, 2473, 2395, 2325,
+    2260, 2201, 2147, 2096, 2048, 2003, 1961, 1921, 1883, 1847, 1813, 1780,
+    1748, 1718, 1689, 1661, 1635, 1609, 1584, 1559, 1536, 1513, 1491, 1470,
+    1449, 1429, 1409, 1390, 1371, 1353, 1335, 1318, 1301, 1284, 1268, 1252,
+    1236, 1221, 1206, 1192, 1177, 1163, 1149, 1136, 1123, 1110, 1097, 1084,
+    1072, 1059, 1047, 1036, 1024, 1013, 1001, 990,  979,  968,  958,  947,
+    937,  927,  917,  907,  897,  887,  878,  868,  859,  850,  841,  832,
+    823,  814,  806,  797,  789,  780,  772,  764,  756,  748,  740,  732,
+    724,  717,  709,  702,  694,  687,  680,  673,  665,  658,  651,  644,
+    637,  631,  624,  617,  611,  604,  598,  591,  585,  578,  572,  566,
+    560,  554,  547,  541,  535,  530,  524,  518,  512,  506,  501,  495,
+    489,  484,  478,  473,  467,  462,  456,  451,  446,  441,  435,  430,
+    425,  420,  415,  410,  405,  400,  395,  390,  385,  380,  375,  371,
+    366,  361,  356,  352,  347,  343,  338,  333,  329,  324,  320,  316,
+    311,  307,  302,  298,  294,  289,  285,  281,  277,  273,  268,  264,
+    260,  256,  252,  248,  244,  240,  236,  232,  228,  224,  220,  216,
+    212,  209,  205,  201,  197,  194,  190,  186,  182,  179,  175,  171,
+    168,  164,  161,  157,  153,  150,  146,  143,  139,  136,  132,  129,
+    125,  122,  119,  115,  112,  109,  105,  102,  99,   95,   92,   89,
+    86,   82,   79,   76,   73,   70,   66,   63,   60,   57,   54,   51,
+    48,   45,   42,   38,   35,   32,   29,   26,   23,   20,   18,   15,
+    12,   9,    6,    3};
+
+static void cost(int *costs, vpx_tree tree, const vpx_prob *probs,
+                 int i, int c) {
+  const vpx_prob prob = probs[i / 2];
+  int b;
+
+  assert(prob != 0);
+  for (b = 0; b <= 1; ++b) {
+    const int cc = c + vp9_cost_bit(prob, b);
+    const vpx_tree_index ii = tree[i + b];
+
+    if (ii <= 0)
+      costs[-ii] = cc;
+    else
+      cost(costs, tree, probs, ii, cc);
+  }
+}
+
+void vp9_cost_tokens(int *costs, const vpx_prob *probs, vpx_tree tree) {
+  cost(costs, tree, probs, 0, 0);
+}
+
+void vp9_cost_tokens_skip(int *costs, const vpx_prob *probs, vpx_tree tree) {
+  assert(tree[0] <= 0 && tree[1] > 0);
+
+  costs[-tree[0]] = vp9_cost_bit(probs[0], 0);
+  cost(costs, tree, probs, 2, 0);
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_cost.h b/libvpx/libvpx/vp9/encoder/vp9_cost.h
new file mode 100644
index 0000000..0c70b78
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_cost.h
@@ -0,0 +1,59 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_COST_H_
+#define VP9_ENCODER_VP9_COST_H_
+
+#include "vpx_dsp/prob.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern const uint16_t vp9_prob_cost[256];
+
+// The factor to scale from cost in bits to cost in vp9_prob_cost units.
+#define VP9_PROB_COST_SHIFT 9
+
+#define vp9_cost_zero(prob) (vp9_prob_cost[prob])
+
+#define vp9_cost_one(prob) vp9_cost_zero(256 - (prob))
+
+#define vp9_cost_bit(prob, bit) vp9_cost_zero((bit) ? 256 - (prob) \
+                                                    : (prob))
+
+static INLINE unsigned int cost_branch256(const unsigned int ct[2],
+                                          vpx_prob p) {
+  return ct[0] * vp9_cost_zero(p) + ct[1] * vp9_cost_one(p);
+}
+
+static INLINE int treed_cost(vpx_tree tree, const vpx_prob *probs,
+                             int bits, int len) {
+  int cost = 0;
+  vpx_tree_index i = 0;
+
+  do {
+    const int bit = (bits >> --len) & 1;
+    cost += vp9_cost_bit(probs[i >> 1], bit);
+    i = tree[i + bit];
+  } while (len);
+
+  return cost;
+}
+
+void vp9_cost_tokens(int *costs, const vpx_prob *probs, vpx_tree tree);
+void vp9_cost_tokens_skip(int *costs, const vpx_prob *probs, vpx_tree tree);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_COST_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_dct.c b/libvpx/libvpx/vp9/encoder/vp9_dct.c
new file mode 100644
index 0000000..f94540b
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_dct.c
@@ -0,0 +1,810 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <math.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_idct.h"
+#include "vpx_dsp/fwd_txfm.h"
+#include "vpx_ports/mem.h"
+
+static void fdct4(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t step[4];
+  tran_high_t temp1, temp2;
+
+  step[0] = input[0] + input[3];
+  step[1] = input[1] + input[2];
+  step[2] = input[1] - input[2];
+  step[3] = input[0] - input[3];
+
+  temp1 = (step[0] + step[1]) * cospi_16_64;
+  temp2 = (step[0] - step[1]) * cospi_16_64;
+  output[0] = (tran_low_t)fdct_round_shift(temp1);
+  output[2] = (tran_low_t)fdct_round_shift(temp2);
+  temp1 = step[2] * cospi_24_64 + step[3] * cospi_8_64;
+  temp2 = -step[2] * cospi_8_64 + step[3] * cospi_24_64;
+  output[1] = (tran_low_t)fdct_round_shift(temp1);
+  output[3] = (tran_low_t)fdct_round_shift(temp2);
+}
+
+static void fdct8(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
+  tran_high_t t0, t1, t2, t3;                  // needs32
+  tran_high_t x0, x1, x2, x3;                  // canbe16
+
+  // stage 1
+  s0 = input[0] + input[7];
+  s1 = input[1] + input[6];
+  s2 = input[2] + input[5];
+  s3 = input[3] + input[4];
+  s4 = input[3] - input[4];
+  s5 = input[2] - input[5];
+  s6 = input[1] - input[6];
+  s7 = input[0] - input[7];
+
+  // fdct4(step, step);
+  x0 = s0 + s3;
+  x1 = s1 + s2;
+  x2 = s1 - s2;
+  x3 = s0 - s3;
+  t0 = (x0 + x1) * cospi_16_64;
+  t1 = (x0 - x1) * cospi_16_64;
+  t2 =  x2 * cospi_24_64 + x3 *  cospi_8_64;
+  t3 = -x2 * cospi_8_64  + x3 * cospi_24_64;
+  output[0] = (tran_low_t)fdct_round_shift(t0);
+  output[2] = (tran_low_t)fdct_round_shift(t2);
+  output[4] = (tran_low_t)fdct_round_shift(t1);
+  output[6] = (tran_low_t)fdct_round_shift(t3);
+
+  // Stage 2
+  t0 = (s6 - s5) * cospi_16_64;
+  t1 = (s6 + s5) * cospi_16_64;
+  t2 = (tran_low_t)fdct_round_shift(t0);
+  t3 = (tran_low_t)fdct_round_shift(t1);
+
+  // Stage 3
+  x0 = s4 + t2;
+  x1 = s4 - t2;
+  x2 = s7 - t3;
+  x3 = s7 + t3;
+
+  // Stage 4
+  t0 = x0 * cospi_28_64 + x3 *   cospi_4_64;
+  t1 = x1 * cospi_12_64 + x2 *  cospi_20_64;
+  t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
+  t3 = x3 * cospi_28_64 + x0 *  -cospi_4_64;
+  output[1] = (tran_low_t)fdct_round_shift(t0);
+  output[3] = (tran_low_t)fdct_round_shift(t2);
+  output[5] = (tran_low_t)fdct_round_shift(t1);
+  output[7] = (tran_low_t)fdct_round_shift(t3);
+}
+
+static void fdct16(const tran_low_t in[16], tran_low_t out[16]) {
+  tran_high_t step1[8];      // canbe16
+  tran_high_t step2[8];      // canbe16
+  tran_high_t step3[8];      // canbe16
+  tran_high_t input[8];      // canbe16
+  tran_high_t temp1, temp2;  // needs32
+
+  // step 1
+  input[0] = in[0] + in[15];
+  input[1] = in[1] + in[14];
+  input[2] = in[2] + in[13];
+  input[3] = in[3] + in[12];
+  input[4] = in[4] + in[11];
+  input[5] = in[5] + in[10];
+  input[6] = in[6] + in[ 9];
+  input[7] = in[7] + in[ 8];
+
+  step1[0] = in[7] - in[ 8];
+  step1[1] = in[6] - in[ 9];
+  step1[2] = in[5] - in[10];
+  step1[3] = in[4] - in[11];
+  step1[4] = in[3] - in[12];
+  step1[5] = in[2] - in[13];
+  step1[6] = in[1] - in[14];
+  step1[7] = in[0] - in[15];
+
+  // fdct8(step, step);
+  {
+    tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
+    tran_high_t t0, t1, t2, t3;                  // needs32
+    tran_high_t x0, x1, x2, x3;                  // canbe16
+
+    // stage 1
+    s0 = input[0] + input[7];
+    s1 = input[1] + input[6];
+    s2 = input[2] + input[5];
+    s3 = input[3] + input[4];
+    s4 = input[3] - input[4];
+    s5 = input[2] - input[5];
+    s6 = input[1] - input[6];
+    s7 = input[0] - input[7];
+
+    // fdct4(step, step);
+    x0 = s0 + s3;
+    x1 = s1 + s2;
+    x2 = s1 - s2;
+    x3 = s0 - s3;
+    t0 = (x0 + x1) * cospi_16_64;
+    t1 = (x0 - x1) * cospi_16_64;
+    t2 = x3 * cospi_8_64  + x2 * cospi_24_64;
+    t3 = x3 * cospi_24_64 - x2 * cospi_8_64;
+    out[0] = (tran_low_t)fdct_round_shift(t0);
+    out[4] = (tran_low_t)fdct_round_shift(t2);
+    out[8] = (tran_low_t)fdct_round_shift(t1);
+    out[12] = (tran_low_t)fdct_round_shift(t3);
+
+    // Stage 2
+    t0 = (s6 - s5) * cospi_16_64;
+    t1 = (s6 + s5) * cospi_16_64;
+    t2 = fdct_round_shift(t0);
+    t3 = fdct_round_shift(t1);
+
+    // Stage 3
+    x0 = s4 + t2;
+    x1 = s4 - t2;
+    x2 = s7 - t3;
+    x3 = s7 + t3;
+
+    // Stage 4
+    t0 = x0 * cospi_28_64 + x3 *   cospi_4_64;
+    t1 = x1 * cospi_12_64 + x2 *  cospi_20_64;
+    t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
+    t3 = x3 * cospi_28_64 + x0 *  -cospi_4_64;
+    out[2] = (tran_low_t)fdct_round_shift(t0);
+    out[6] = (tran_low_t)fdct_round_shift(t2);
+    out[10] = (tran_low_t)fdct_round_shift(t1);
+    out[14] = (tran_low_t)fdct_round_shift(t3);
+  }
+
+  // step 2
+  temp1 = (step1[5] - step1[2]) * cospi_16_64;
+  temp2 = (step1[4] - step1[3]) * cospi_16_64;
+  step2[2] = fdct_round_shift(temp1);
+  step2[3] = fdct_round_shift(temp2);
+  temp1 = (step1[4] + step1[3]) * cospi_16_64;
+  temp2 = (step1[5] + step1[2]) * cospi_16_64;
+  step2[4] = fdct_round_shift(temp1);
+  step2[5] = fdct_round_shift(temp2);
+
+  // step 3
+  step3[0] = step1[0] + step2[3];
+  step3[1] = step1[1] + step2[2];
+  step3[2] = step1[1] - step2[2];
+  step3[3] = step1[0] - step2[3];
+  step3[4] = step1[7] - step2[4];
+  step3[5] = step1[6] - step2[5];
+  step3[6] = step1[6] + step2[5];
+  step3[7] = step1[7] + step2[4];
+
+  // step 4
+  temp1 = step3[1] *  -cospi_8_64 + step3[6] * cospi_24_64;
+  temp2 = step3[2] * cospi_24_64 + step3[5] *  cospi_8_64;
+  step2[1] = fdct_round_shift(temp1);
+  step2[2] = fdct_round_shift(temp2);
+  temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
+  temp2 = step3[1] * cospi_24_64 + step3[6] *  cospi_8_64;
+  step2[5] = fdct_round_shift(temp1);
+  step2[6] = fdct_round_shift(temp2);
+
+  // step 5
+  step1[0] = step3[0] + step2[1];
+  step1[1] = step3[0] - step2[1];
+  step1[2] = step3[3] + step2[2];
+  step1[3] = step3[3] - step2[2];
+  step1[4] = step3[4] - step2[5];
+  step1[5] = step3[4] + step2[5];
+  step1[6] = step3[7] - step2[6];
+  step1[7] = step3[7] + step2[6];
+
+  // step 6
+  temp1 = step1[0] * cospi_30_64 + step1[7] *  cospi_2_64;
+  temp2 = step1[1] * cospi_14_64 + step1[6] * cospi_18_64;
+  out[1] = (tran_low_t)fdct_round_shift(temp1);
+  out[9] = (tran_low_t)fdct_round_shift(temp2);
+
+  temp1 = step1[2] * cospi_22_64 + step1[5] * cospi_10_64;
+  temp2 = step1[3] *  cospi_6_64 + step1[4] * cospi_26_64;
+  out[5] = (tran_low_t)fdct_round_shift(temp1);
+  out[13] = (tran_low_t)fdct_round_shift(temp2);
+
+  temp1 = step1[3] * -cospi_26_64 + step1[4] *  cospi_6_64;
+  temp2 = step1[2] * -cospi_10_64 + step1[5] * cospi_22_64;
+  out[3] = (tran_low_t)fdct_round_shift(temp1);
+  out[11] = (tran_low_t)fdct_round_shift(temp2);
+
+  temp1 = step1[1] * -cospi_18_64 + step1[6] * cospi_14_64;
+  temp2 = step1[0] *  -cospi_2_64 + step1[7] * cospi_30_64;
+  out[7] = (tran_low_t)fdct_round_shift(temp1);
+  out[15] = (tran_low_t)fdct_round_shift(temp2);
+}
+
+static void fadst4(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t x0, x1, x2, x3;
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  x0 = input[0];
+  x1 = input[1];
+  x2 = input[2];
+  x3 = input[3];
+
+  if (!(x0 | x1 | x2 | x3)) {
+    output[0] = output[1] = output[2] = output[3] = 0;
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_4_9 * x0;
+  s2 = sinpi_2_9 * x1;
+  s3 = sinpi_1_9 * x1;
+  s4 = sinpi_3_9 * x2;
+  s5 = sinpi_4_9 * x3;
+  s6 = sinpi_2_9 * x3;
+  s7 = x0 + x1 - x3;
+
+  x0 = s0 + s2 + s5;
+  x1 = sinpi_3_9 * s7;
+  x2 = s1 - s3 + s6;
+  x3 = s4;
+
+  s0 = x0 + x3;
+  s1 = x1;
+  s2 = x2 - x3;
+  s3 = x2 - x0 + x3;
+
+  // 1-D transform scaling factor is sqrt(2).
+  output[0] = (tran_low_t)fdct_round_shift(s0);
+  output[1] = (tran_low_t)fdct_round_shift(s1);
+  output[2] = (tran_low_t)fdct_round_shift(s2);
+  output[3] = (tran_low_t)fdct_round_shift(s3);
+}
+
+static void fadst8(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_high_t x0 = input[7];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[5];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[3];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[1];
+  tran_high_t x7 = input[6];
+
+  // stage 1
+  s0 = cospi_2_64  * x0 + cospi_30_64 * x1;
+  s1 = cospi_30_64 * x0 - cospi_2_64  * x1;
+  s2 = cospi_10_64 * x2 + cospi_22_64 * x3;
+  s3 = cospi_22_64 * x2 - cospi_10_64 * x3;
+  s4 = cospi_18_64 * x4 + cospi_14_64 * x5;
+  s5 = cospi_14_64 * x4 - cospi_18_64 * x5;
+  s6 = cospi_26_64 * x6 + cospi_6_64  * x7;
+  s7 = cospi_6_64  * x6 - cospi_26_64 * x7;
+
+  x0 = fdct_round_shift(s0 + s4);
+  x1 = fdct_round_shift(s1 + s5);
+  x2 = fdct_round_shift(s2 + s6);
+  x3 = fdct_round_shift(s3 + s7);
+  x4 = fdct_round_shift(s0 - s4);
+  x5 = fdct_round_shift(s1 - s5);
+  x6 = fdct_round_shift(s2 - s6);
+  x7 = fdct_round_shift(s3 - s7);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = cospi_8_64  * x4 + cospi_24_64 * x5;
+  s5 = cospi_24_64 * x4 - cospi_8_64  * x5;
+  s6 = - cospi_24_64 * x6 + cospi_8_64  * x7;
+  s7 =   cospi_8_64  * x6 + cospi_24_64 * x7;
+
+  x0 = s0 + s2;
+  x1 = s1 + s3;
+  x2 = s0 - s2;
+  x3 = s1 - s3;
+  x4 = fdct_round_shift(s4 + s6);
+  x5 = fdct_round_shift(s5 + s7);
+  x6 = fdct_round_shift(s4 - s6);
+  x7 = fdct_round_shift(s5 - s7);
+
+  // stage 3
+  s2 = cospi_16_64 * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (x6 - x7);
+
+  x2 = fdct_round_shift(s2);
+  x3 = fdct_round_shift(s3);
+  x6 = fdct_round_shift(s6);
+  x7 = fdct_round_shift(s7);
+
+  output[0] = (tran_low_t)x0;
+  output[1] = (tran_low_t)-x4;
+  output[2] = (tran_low_t)x6;
+  output[3] = (tran_low_t)-x2;
+  output[4] = (tran_low_t)x3;
+  output[5] = (tran_low_t)-x7;
+  output[6] = (tran_low_t)x5;
+  output[7] = (tran_low_t)-x1;
+}
+
+static void fadst16(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7, s8;
+  tran_high_t s9, s10, s11, s12, s13, s14, s15;
+
+  tran_high_t x0 = input[15];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[13];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[11];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[9];
+  tran_high_t x7 = input[6];
+  tran_high_t x8 = input[7];
+  tran_high_t x9 = input[8];
+  tran_high_t x10 = input[5];
+  tran_high_t x11 = input[10];
+  tran_high_t x12 = input[3];
+  tran_high_t x13 = input[12];
+  tran_high_t x14 = input[1];
+  tran_high_t x15 = input[14];
+
+  // stage 1
+  s0 = x0 * cospi_1_64  + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64  + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64  + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64  - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64  - x15 * cospi_29_64;
+
+  x0 = fdct_round_shift(s0 + s8);
+  x1 = fdct_round_shift(s1 + s9);
+  x2 = fdct_round_shift(s2 + s10);
+  x3 = fdct_round_shift(s3 + s11);
+  x4 = fdct_round_shift(s4 + s12);
+  x5 = fdct_round_shift(s5 + s13);
+  x6 = fdct_round_shift(s6 + s14);
+  x7 = fdct_round_shift(s7 + s15);
+  x8  = fdct_round_shift(s0 - s8);
+  x9  = fdct_round_shift(s1 - s9);
+  x10 = fdct_round_shift(s2 - s10);
+  x11 = fdct_round_shift(s3 - s11);
+  x12 = fdct_round_shift(s4 - s12);
+  x13 = fdct_round_shift(s5 - s13);
+  x14 = fdct_round_shift(s6 - s14);
+  x15 = fdct_round_shift(s7 - s15);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 =    x8 * cospi_4_64   + x9 * cospi_28_64;
+  s9 =    x8 * cospi_28_64  - x9 * cospi_4_64;
+  s10 =   x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 =   x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = - x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 =   x12 * cospi_4_64  + x13 * cospi_28_64;
+  s14 = - x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 =   x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = s0 + s4;
+  x1 = s1 + s5;
+  x2 = s2 + s6;
+  x3 = s3 + s7;
+  x4 = s0 - s4;
+  x5 = s1 - s5;
+  x6 = s2 - s6;
+  x7 = s3 - s7;
+  x8 = fdct_round_shift(s8 + s12);
+  x9 = fdct_round_shift(s9 + s13);
+  x10 = fdct_round_shift(s10 + s14);
+  x11 = fdct_round_shift(s11 + s15);
+  x12 = fdct_round_shift(s8 - s12);
+  x13 = fdct_round_shift(s9 - s13);
+  x14 = fdct_round_shift(s10 - s14);
+  x15 = fdct_round_shift(s11 - s15);
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64  + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = - x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 =   x6 * cospi_8_64  + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64  + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = - x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 =   x14 * cospi_8_64  + x15 * cospi_24_64;
+
+  x0 = s0 + s2;
+  x1 = s1 + s3;
+  x2 = s0 - s2;
+  x3 = s1 - s3;
+  x4 = fdct_round_shift(s4 + s6);
+  x5 = fdct_round_shift(s5 + s7);
+  x6 = fdct_round_shift(s4 - s6);
+  x7 = fdct_round_shift(s5 - s7);
+  x8 = s8 + s10;
+  x9 = s9 + s11;
+  x10 = s8 - s10;
+  x11 = s9 - s11;
+  x12 = fdct_round_shift(s12 + s14);
+  x13 = fdct_round_shift(s13 + s15);
+  x14 = fdct_round_shift(s12 - s14);
+  x15 = fdct_round_shift(s13 - s15);
+
+  // stage 4
+  s2 = (- cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (- x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (- x10 + x11);
+  s14 = (- cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = fdct_round_shift(s2);
+  x3 = fdct_round_shift(s3);
+  x6 = fdct_round_shift(s6);
+  x7 = fdct_round_shift(s7);
+  x10 = fdct_round_shift(s10);
+  x11 = fdct_round_shift(s11);
+  x14 = fdct_round_shift(s14);
+  x15 = fdct_round_shift(s15);
+
+  output[0] = (tran_low_t)x0;
+  output[1] = (tran_low_t)-x8;
+  output[2] = (tran_low_t)x12;
+  output[3] = (tran_low_t)-x4;
+  output[4] = (tran_low_t)x6;
+  output[5] = (tran_low_t)x14;
+  output[6] = (tran_low_t)x10;
+  output[7] = (tran_low_t)x2;
+  output[8] = (tran_low_t)x3;
+  output[9] = (tran_low_t)x11;
+  output[10] = (tran_low_t)x15;
+  output[11] = (tran_low_t)x7;
+  output[12] = (tran_low_t)x5;
+  output[13] = (tran_low_t)-x13;
+  output[14] = (tran_low_t)x9;
+  output[15] = (tran_low_t)-x1;
+}
+
+static const transform_2d FHT_4[] = {
+  { fdct4,  fdct4  },  // DCT_DCT  = 0
+  { fadst4, fdct4  },  // ADST_DCT = 1
+  { fdct4,  fadst4 },  // DCT_ADST = 2
+  { fadst4, fadst4 }   // ADST_ADST = 3
+};
+
+static const transform_2d FHT_8[] = {
+  { fdct8,  fdct8  },  // DCT_DCT  = 0
+  { fadst8, fdct8  },  // ADST_DCT = 1
+  { fdct8,  fadst8 },  // DCT_ADST = 2
+  { fadst8, fadst8 }   // ADST_ADST = 3
+};
+
+static const transform_2d FHT_16[] = {
+  { fdct16,  fdct16  },  // DCT_DCT  = 0
+  { fadst16, fdct16  },  // ADST_DCT = 1
+  { fdct16,  fadst16 },  // DCT_ADST = 2
+  { fadst16, fadst16 }   // ADST_ADST = 3
+};
+
+void vp9_fht4x4_c(const int16_t *input, tran_low_t *output,
+                  int stride, int tx_type) {
+  if (tx_type == DCT_DCT) {
+    vpx_fdct4x4_c(input, output, stride);
+  } else {
+    tran_low_t out[4 * 4];
+    int i, j;
+    tran_low_t temp_in[4], temp_out[4];
+    const transform_2d ht = FHT_4[tx_type];
+
+    // Columns
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j)
+        temp_in[j] = input[j * stride + i] * 16;
+      if (i == 0 && temp_in[0])
+        temp_in[0] += 1;
+      ht.cols(temp_in, temp_out);
+      for (j = 0; j < 4; ++j)
+        out[j * 4 + i] = temp_out[j];
+    }
+
+    // Rows
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j)
+        temp_in[j] = out[j + i * 4];
+      ht.rows(temp_in, temp_out);
+      for (j = 0; j < 4; ++j)
+        output[j + i * 4] = (temp_out[j] + 1) >> 2;
+    }
+  }
+}
+
+void vp9_fdct8x8_quant_c(const int16_t *input, int stride,
+                         tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                         int skip_block,
+                         const int16_t *zbin_ptr, const int16_t *round_ptr,
+                         const int16_t *quant_ptr,
+                         const int16_t *quant_shift_ptr,
+                         tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                         const int16_t *dequant_ptr,
+                         uint16_t *eob_ptr,
+                         const int16_t *scan, const int16_t *iscan) {
+  int eob = -1;
+
+  int i, j;
+  tran_low_t intermediate[64];
+
+  // Transform columns
+  {
+    tran_low_t *output = intermediate;
+    tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
+    tran_high_t t0, t1, t2, t3;                  // needs32
+    tran_high_t x0, x1, x2, x3;                  // canbe16
+
+    int i;
+    for (i = 0; i < 8; i++) {
+      // stage 1
+      s0 = (input[0 * stride] + input[7 * stride]) * 4;
+      s1 = (input[1 * stride] + input[6 * stride]) * 4;
+      s2 = (input[2 * stride] + input[5 * stride]) * 4;
+      s3 = (input[3 * stride] + input[4 * stride]) * 4;
+      s4 = (input[3 * stride] - input[4 * stride]) * 4;
+      s5 = (input[2 * stride] - input[5 * stride]) * 4;
+      s6 = (input[1 * stride] - input[6 * stride]) * 4;
+      s7 = (input[0 * stride] - input[7 * stride]) * 4;
+
+      // fdct4(step, step);
+      x0 = s0 + s3;
+      x1 = s1 + s2;
+      x2 = s1 - s2;
+      x3 = s0 - s3;
+      t0 = (x0 + x1) * cospi_16_64;
+      t1 = (x0 - x1) * cospi_16_64;
+      t2 =  x2 * cospi_24_64 + x3 *  cospi_8_64;
+      t3 = -x2 * cospi_8_64  + x3 * cospi_24_64;
+      output[0 * 8] = (tran_low_t)fdct_round_shift(t0);
+      output[2 * 8] = (tran_low_t)fdct_round_shift(t2);
+      output[4 * 8] = (tran_low_t)fdct_round_shift(t1);
+      output[6 * 8] = (tran_low_t)fdct_round_shift(t3);
+
+      // Stage 2
+      t0 = (s6 - s5) * cospi_16_64;
+      t1 = (s6 + s5) * cospi_16_64;
+      t2 = fdct_round_shift(t0);
+      t3 = fdct_round_shift(t1);
+
+      // Stage 3
+      x0 = s4 + t2;
+      x1 = s4 - t2;
+      x2 = s7 - t3;
+      x3 = s7 + t3;
+
+      // Stage 4
+      t0 = x0 * cospi_28_64 + x3 *   cospi_4_64;
+      t1 = x1 * cospi_12_64 + x2 *  cospi_20_64;
+      t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
+      t3 = x3 * cospi_28_64 + x0 *  -cospi_4_64;
+      output[1 * 8] = (tran_low_t)fdct_round_shift(t0);
+      output[3 * 8] = (tran_low_t)fdct_round_shift(t2);
+      output[5 * 8] = (tran_low_t)fdct_round_shift(t1);
+      output[7 * 8] = (tran_low_t)fdct_round_shift(t3);
+      input++;
+      output++;
+    }
+  }
+
+  // Rows
+  for (i = 0; i < 8; ++i) {
+    fdct8(&intermediate[i * 8], &coeff_ptr[i * 8]);
+    for (j = 0; j < 8; ++j)
+      coeff_ptr[j + i * 8] /= 2;
+  }
+
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+      tmp = (tmp * quant_ptr[rc != 0]) >> 16;
+
+      qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+
+      if (tmp)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void vp9_fht8x8_c(const int16_t *input, tran_low_t *output,
+                  int stride, int tx_type) {
+  if (tx_type == DCT_DCT) {
+    vpx_fdct8x8_c(input, output, stride);
+  } else {
+    tran_low_t out[64];
+    int i, j;
+    tran_low_t temp_in[8], temp_out[8];
+    const transform_2d ht = FHT_8[tx_type];
+
+    // Columns
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j)
+        temp_in[j] = input[j * stride + i] * 4;
+      ht.cols(temp_in, temp_out);
+      for (j = 0; j < 8; ++j)
+        out[j * 8 + i] = temp_out[j];
+    }
+
+    // Rows
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j)
+        temp_in[j] = out[j + i * 8];
+      ht.rows(temp_in, temp_out);
+      for (j = 0; j < 8; ++j)
+        output[j + i * 8] = (temp_out[j] + (temp_out[j] < 0)) >> 1;
+    }
+  }
+}
+
+/* 4-point reversible, orthonormal Walsh-Hadamard in 3.5 adds, 0.5 shifts per
+   pixel. */
+void vp9_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride) {
+  int i;
+  tran_high_t a1, b1, c1, d1, e1;
+  const int16_t *ip_pass0 = input;
+  const tran_low_t *ip = NULL;
+  tran_low_t *op = output;
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip_pass0[0 * stride];
+    b1 = ip_pass0[1 * stride];
+    c1 = ip_pass0[2 * stride];
+    d1 = ip_pass0[3 * stride];
+
+    a1 += b1;
+    d1 = d1 - c1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= c1;
+    d1 += b1;
+    op[0] = (tran_low_t)a1;
+    op[4] = (tran_low_t)c1;
+    op[8] = (tran_low_t)d1;
+    op[12] = (tran_low_t)b1;
+
+    ip_pass0++;
+    op++;
+  }
+  ip = output;
+  op = output;
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip[0];
+    b1 = ip[1];
+    c1 = ip[2];
+    d1 = ip[3];
+
+    a1 += b1;
+    d1 -= c1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= c1;
+    d1 += b1;
+    op[0] = (tran_low_t)(a1 * UNIT_QUANT_FACTOR);
+    op[1] = (tran_low_t)(c1 * UNIT_QUANT_FACTOR);
+    op[2] = (tran_low_t)(d1 * UNIT_QUANT_FACTOR);
+    op[3] = (tran_low_t)(b1 * UNIT_QUANT_FACTOR);
+
+    ip += 4;
+    op += 4;
+  }
+}
+
+void vp9_fht16x16_c(const int16_t *input, tran_low_t *output,
+                    int stride, int tx_type) {
+  if (tx_type == DCT_DCT) {
+    vpx_fdct16x16_c(input, output, stride);
+  } else {
+    tran_low_t out[256];
+    int i, j;
+    tran_low_t temp_in[16], temp_out[16];
+    const transform_2d ht = FHT_16[tx_type];
+
+    // Columns
+    for (i = 0; i < 16; ++i) {
+      for (j = 0; j < 16; ++j)
+        temp_in[j] = input[j * stride + i] * 4;
+      ht.cols(temp_in, temp_out);
+      for (j = 0; j < 16; ++j)
+        out[j * 16 + i] = (temp_out[j] + 1 + (temp_out[j] < 0)) >> 2;
+    }
+
+    // Rows
+    for (i = 0; i < 16; ++i) {
+      for (j = 0; j < 16; ++j)
+        temp_in[j] = out[j + i * 16];
+      ht.rows(temp_in, temp_out);
+      for (j = 0; j < 16; ++j)
+        output[j + i * 16] = temp_out[j];
+    }
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_fht4x4_c(const int16_t *input, tran_low_t *output,
+                         int stride, int tx_type) {
+  vp9_fht4x4_c(input, output, stride, tx_type);
+}
+
+void vp9_highbd_fht8x8_c(const int16_t *input, tran_low_t *output,
+                         int stride, int tx_type) {
+  vp9_fht8x8_c(input, output, stride, tx_type);
+}
+
+void vp9_highbd_fwht4x4_c(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  vp9_fwht4x4_c(input, output, stride);
+}
+
+void vp9_highbd_fht16x16_c(const int16_t *input, tran_low_t *output,
+                           int stride, int tx_type) {
+  vp9_fht16x16_c(input, output, stride, tx_type);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vp9/encoder/vp9_denoiser.c b/libvpx/libvpx/vp9/encoder/vp9_denoiser.c
new file mode 100644
index 0000000..42d456e
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_denoiser.c
@@ -0,0 +1,615 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_denoiser.h"
+#include "vp9/encoder/vp9_encoder.h"
+
+#ifdef OUTPUT_YUV_DENOISED
+static void make_grayscale(YV12_BUFFER_CONFIG *yuv);
+#endif
+
+static int absdiff_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  (void)bs;
+  return 3 + (increase_denoising ? 1 : 0);
+}
+
+static int delta_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  (void)bs;
+  (void)increase_denoising;
+  return 4;
+}
+
+static int noise_motion_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  (void)bs;
+  (void)increase_denoising;
+  return 625;
+}
+
+static unsigned int sse_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 80 : 40);
+}
+
+static int sse_diff_thresh(BLOCK_SIZE bs, int increase_denoising,
+                           int motion_magnitude) {
+  if (motion_magnitude >
+      noise_motion_thresh(bs, increase_denoising)) {
+    if (increase_denoising)
+      return (1 << num_pels_log2_lookup[bs]) << 2;
+    else
+      return 0;
+  } else {
+    return (1 << num_pels_log2_lookup[bs]) << 4;
+  }
+}
+
+static int total_adj_weak_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 3 : 2);
+}
+
+// TODO(jackychen): If increase_denoising is enabled in the future,
+// we might need to update the code for calculating 'total_adj' in
+// case the C code is not bit-exact with corresponding sse2 code.
+int vp9_denoiser_filter_c(const uint8_t *sig, int sig_stride,
+                          const uint8_t *mc_avg,
+                          int mc_avg_stride,
+                          uint8_t *avg, int avg_stride,
+                          int increase_denoising,
+                          BLOCK_SIZE bs,
+                          int motion_magnitude) {
+  int r, c;
+  const uint8_t *sig_start = sig;
+  const uint8_t *mc_avg_start = mc_avg;
+  uint8_t *avg_start = avg;
+  int diff, adj, absdiff, delta;
+  int adj_val[] = {3, 4, 6};
+  int total_adj = 0;
+  int shift_inc = 1;
+
+  // If motion_magnitude is small, making the denoiser more aggressive by
+  // increasing the adjustment for each level. Add another increment for
+  // blocks that are labeled for increase denoising.
+  if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) {
+    if (increase_denoising) {
+      shift_inc = 2;
+    }
+    adj_val[0] += shift_inc;
+    adj_val[1] += shift_inc;
+    adj_val[2] += shift_inc;
+  }
+
+  // First attempt to apply a strong temporal denoising filter.
+  for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+    for (c = 0; c < (4 << b_width_log2_lookup[bs]); ++c) {
+      diff = mc_avg[c] - sig[c];
+      absdiff = abs(diff);
+
+      if (absdiff <= absdiff_thresh(bs, increase_denoising)) {
+        avg[c] = mc_avg[c];
+        total_adj += diff;
+      } else {
+        switch (absdiff) {
+          case 4: case 5: case 6: case 7:
+            adj = adj_val[0];
+            break;
+          case 8: case 9: case 10: case 11:
+          case 12: case 13: case 14: case 15:
+            adj = adj_val[1];
+            break;
+          default:
+            adj = adj_val[2];
+        }
+        if (diff > 0) {
+          avg[c] = VPXMIN(UINT8_MAX, sig[c] + adj);
+          total_adj += adj;
+        } else {
+          avg[c] = VPXMAX(0, sig[c] - adj);
+          total_adj -= adj;
+        }
+      }
+    }
+    sig += sig_stride;
+    avg += avg_stride;
+    mc_avg += mc_avg_stride;
+  }
+
+  // If the strong filter did not modify the signal too much, we're all set.
+  if (abs(total_adj) <= total_adj_strong_thresh(bs, increase_denoising)) {
+    return FILTER_BLOCK;
+  }
+
+  // Otherwise, we try to dampen the filter if the delta is not too high.
+  delta = ((abs(total_adj) - total_adj_strong_thresh(bs, increase_denoising))
+           >> num_pels_log2_lookup[bs]) + 1;
+
+  if (delta >= delta_thresh(bs, increase_denoising)) {
+    return COPY_BLOCK;
+  }
+
+  mc_avg =  mc_avg_start;
+  avg = avg_start;
+  sig = sig_start;
+  for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+    for (c = 0; c < (4 << b_width_log2_lookup[bs]); ++c) {
+      diff = mc_avg[c] - sig[c];
+      adj = abs(diff);
+      if (adj > delta) {
+        adj = delta;
+      }
+      if (diff > 0) {
+        // Diff positive means we made positive adjustment above
+        // (in first try/attempt), so now make negative adjustment to bring
+        // denoised signal down.
+        avg[c] = VPXMAX(0, avg[c] - adj);
+        total_adj -= adj;
+      } else {
+        // Diff negative means we made negative adjustment above
+        // (in first try/attempt), so now make positive adjustment to bring
+        // denoised signal up.
+        avg[c] = VPXMIN(UINT8_MAX, avg[c] + adj);
+        total_adj += adj;
+      }
+    }
+    sig += sig_stride;
+    avg += avg_stride;
+    mc_avg += mc_avg_stride;
+  }
+
+  // We can use the filter if it has been sufficiently dampened
+  if (abs(total_adj) <= total_adj_weak_thresh(bs, increase_denoising)) {
+    return FILTER_BLOCK;
+  }
+  return COPY_BLOCK;
+}
+
+static uint8_t *block_start(uint8_t *framebuf, int stride,
+                            int mi_row, int mi_col) {
+  return framebuf + (stride * mi_row  << 3) + (mi_col << 3);
+}
+
+static VP9_DENOISER_DECISION perform_motion_compensation(VP9_DENOISER *denoiser,
+                                                         MACROBLOCK *mb,
+                                                         BLOCK_SIZE bs,
+                                                         int increase_denoising,
+                                                         int mi_row,
+                                                         int mi_col,
+                                                         PICK_MODE_CONTEXT *ctx,
+                                                         int motion_magnitude,
+                                                         int is_skin,
+                                                         int *zeromv_filter,
+                                                         int consec_zeromv) {
+  int sse_diff = ctx->zeromv_sse - ctx->newmv_sse;
+  MV_REFERENCE_FRAME frame;
+  MACROBLOCKD *filter_mbd = &mb->e_mbd;
+  MODE_INFO *mi = filter_mbd->mi[0];
+  MODE_INFO saved_mi;
+  int i;
+  struct buf_2d saved_dst[MAX_MB_PLANE];
+  struct buf_2d saved_pre[MAX_MB_PLANE];
+
+  frame = ctx->best_reference_frame;
+  saved_mi = *mi;
+
+  if (is_skin && (motion_magnitude > 0 || consec_zeromv < 4))
+    return COPY_BLOCK;
+
+  // Avoid denoising for small block (unless motion is small).
+  // Small blocks are selected in variance partition (before encoding) and
+  // will typically lie on moving areas.
+  if (denoiser->denoising_level < kDenHigh &&
+      motion_magnitude > 16 && bs <= BLOCK_8X8)
+    return COPY_BLOCK;
+
+  // If the best reference frame uses inter-prediction and there is enough of a
+  // difference in sum-squared-error, use it.
+  if (frame != INTRA_FRAME &&
+      ctx->newmv_sse != UINT_MAX &&
+      sse_diff > sse_diff_thresh(bs, increase_denoising, motion_magnitude)) {
+    mi->ref_frame[0] = ctx->best_reference_frame;
+    mi->mode = ctx->best_sse_inter_mode;
+    mi->mv[0] = ctx->best_sse_mv;
+  } else {
+    // Otherwise, use the zero reference frame.
+    frame = ctx->best_zeromv_reference_frame;
+    ctx->newmv_sse = ctx->zeromv_sse;
+    // Bias to last reference.
+    if (frame != LAST_FRAME &&
+        ((ctx->zeromv_lastref_sse < (5 * ctx->zeromv_sse) >> 2) ||
+         denoiser->denoising_level >= kDenHigh)) {
+      frame = LAST_FRAME;
+      ctx->newmv_sse = ctx->zeromv_lastref_sse;
+    }
+    mi->ref_frame[0] = frame;
+    mi->mode = ZEROMV;
+    mi->mv[0].as_int = 0;
+    ctx->best_sse_inter_mode = ZEROMV;
+    ctx->best_sse_mv.as_int = 0;
+    *zeromv_filter = 1;
+    if (denoiser->denoising_level > kDenMedium) {
+      motion_magnitude = 0;
+    }
+  }
+
+  if (ctx->newmv_sse > sse_thresh(bs, increase_denoising)) {
+    // Restore everything to its original state
+    *mi = saved_mi;
+    return COPY_BLOCK;
+  }
+  if (motion_magnitude >
+     (noise_motion_thresh(bs, increase_denoising) << 3)) {
+    // Restore everything to its original state
+    *mi = saved_mi;
+    return COPY_BLOCK;
+  }
+
+  // We will restore these after motion compensation.
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    saved_pre[i] = filter_mbd->plane[i].pre[0];
+    saved_dst[i] = filter_mbd->plane[i].dst;
+  }
+
+  // Set the pointers in the MACROBLOCKD to point to the buffers in the denoiser
+  // struct.
+  filter_mbd->plane[0].pre[0].buf =
+      block_start(denoiser->running_avg_y[frame].y_buffer,
+                  denoiser->running_avg_y[frame].y_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[0].pre[0].stride =
+      denoiser->running_avg_y[frame].y_stride;
+  filter_mbd->plane[1].pre[0].buf =
+       block_start(denoiser->running_avg_y[frame].u_buffer,
+                  denoiser->running_avg_y[frame].uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[1].pre[0].stride =
+      denoiser->running_avg_y[frame].uv_stride;
+  filter_mbd->plane[2].pre[0].buf =
+      block_start(denoiser->running_avg_y[frame].v_buffer,
+                  denoiser->running_avg_y[frame].uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[2].pre[0].stride =
+      denoiser->running_avg_y[frame].uv_stride;
+
+  filter_mbd->plane[0].dst.buf =
+      block_start(denoiser->mc_running_avg_y.y_buffer,
+                  denoiser->mc_running_avg_y.y_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[0].dst.stride = denoiser->mc_running_avg_y.y_stride;
+  filter_mbd->plane[1].dst.buf =
+      block_start(denoiser->mc_running_avg_y.u_buffer,
+                  denoiser->mc_running_avg_y.uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[1].dst.stride = denoiser->mc_running_avg_y.uv_stride;
+  filter_mbd->plane[2].dst.buf =
+      block_start(denoiser->mc_running_avg_y.v_buffer,
+                  denoiser->mc_running_avg_y.uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[2].dst.stride = denoiser->mc_running_avg_y.uv_stride;
+
+  vp9_build_inter_predictors_sby(filter_mbd, mi_row, mi_col, bs);
+
+  // Restore everything to its original state
+  *mi = saved_mi;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    filter_mbd->plane[i].pre[0] = saved_pre[i];
+    filter_mbd->plane[i].dst = saved_dst[i];
+  }
+
+  return FILTER_BLOCK;
+}
+
+void vp9_denoiser_denoise(VP9_COMP *cpi, MACROBLOCK *mb,
+                          int mi_row, int mi_col, BLOCK_SIZE bs,
+                          PICK_MODE_CONTEXT *ctx,
+                          VP9_DENOISER_DECISION *denoiser_decision) {
+  int mv_col, mv_row;
+  int motion_magnitude = 0;
+  int zeromv_filter = 0;
+  VP9_DENOISER *denoiser = &cpi->denoiser;
+  VP9_DENOISER_DECISION decision = COPY_BLOCK;
+  YV12_BUFFER_CONFIG avg = denoiser->running_avg_y[INTRA_FRAME];
+  YV12_BUFFER_CONFIG mc_avg = denoiser->mc_running_avg_y;
+  uint8_t *avg_start = block_start(avg.y_buffer, avg.y_stride, mi_row, mi_col);
+  uint8_t *mc_avg_start = block_start(mc_avg.y_buffer, mc_avg.y_stride,
+                                          mi_row, mi_col);
+  struct buf_2d src = mb->plane[0].src;
+  int is_skin = 0;
+  int consec_zeromv = 0;
+  mv_col = ctx->best_sse_mv.as_mv.col;
+  mv_row = ctx->best_sse_mv.as_mv.row;
+  motion_magnitude = mv_row * mv_row + mv_col * mv_col;
+
+  if (cpi->use_skin_detection &&
+      bs <= BLOCK_32X32 &&
+      denoiser->denoising_level < kDenHigh) {
+    int motion_level = (motion_magnitude < 16) ? 0 : 1;
+    // If motion for current block is small/zero, compute consec_zeromv for
+    // skin detection (early exit in skin detection is done for large
+    // consec_zeromv when current block has small/zero motion).
+    consec_zeromv = 0;
+    if (motion_level == 0) {
+      VP9_COMMON * const cm = &cpi->common;
+      int j, i;
+      // Loop through the 8x8 sub-blocks.
+      const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
+      const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
+      const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
+      const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
+      const int block_index = mi_row * cm->mi_cols + mi_col;
+      consec_zeromv = 100;
+      for (i = 0; i < ymis; i++) {
+        for (j = 0; j < xmis; j++) {
+          int bl_index = block_index + i * cm->mi_cols + j;
+          consec_zeromv = VPXMIN(cpi->consec_zero_mv[bl_index], consec_zeromv);
+          // No need to keep checking 8x8 blocks if any of the sub-blocks
+          // has small consec_zeromv (since threshold for no_skin based on
+          // zero/small motion in skin detection is high, i.e, > 4).
+          if (consec_zeromv < 4) {
+            i = ymis;
+            j = xmis;
+          }
+        }
+      }
+    }
+    // TODO(marpan): Compute skin detection over sub-blocks.
+    is_skin = vp9_compute_skin_block(mb->plane[0].src.buf,
+                                     mb->plane[1].src.buf,
+                                     mb->plane[2].src.buf,
+                                     mb->plane[0].src.stride,
+                                     mb->plane[1].src.stride,
+                                     bs,
+                                     consec_zeromv,
+                                     motion_level);
+  }
+  if (!is_skin &&
+      denoiser->denoising_level == kDenHigh) {
+    denoiser->increase_denoising = 1;
+  } else {
+    denoiser->increase_denoising = 0;
+  }
+
+  if (denoiser->denoising_level >= kDenLow)
+    decision = perform_motion_compensation(denoiser, mb, bs,
+                                           denoiser->increase_denoising,
+                                           mi_row, mi_col, ctx,
+                                           motion_magnitude,
+                                           is_skin,
+                                           &zeromv_filter,
+                                           consec_zeromv);
+
+  if (decision == FILTER_BLOCK) {
+    decision = vp9_denoiser_filter(src.buf, src.stride,
+                                 mc_avg_start, mc_avg.y_stride,
+                                 avg_start, avg.y_stride,
+                                 denoiser->increase_denoising,
+                                 bs, motion_magnitude);
+  }
+
+  if (decision == FILTER_BLOCK) {
+    vpx_convolve_copy(avg_start, avg.y_stride, src.buf, src.stride,
+                      NULL, 0, NULL, 0,
+                      num_4x4_blocks_wide_lookup[bs] << 2,
+                      num_4x4_blocks_high_lookup[bs] << 2);
+  } else {  // COPY_BLOCK
+    vpx_convolve_copy(src.buf, src.stride, avg_start, avg.y_stride,
+                      NULL, 0, NULL, 0,
+                      num_4x4_blocks_wide_lookup[bs] << 2,
+                      num_4x4_blocks_high_lookup[bs] << 2);
+  }
+  *denoiser_decision = decision;
+  if (decision == FILTER_BLOCK && zeromv_filter == 1)
+    *denoiser_decision = FILTER_ZEROMV_BLOCK;
+}
+
+static void copy_frame(YV12_BUFFER_CONFIG * const dest,
+                       const YV12_BUFFER_CONFIG * const src) {
+  int r;
+  const uint8_t *srcbuf = src->y_buffer;
+  uint8_t *destbuf = dest->y_buffer;
+
+  assert(dest->y_width == src->y_width);
+  assert(dest->y_height == src->y_height);
+
+  for (r = 0; r < dest->y_height; ++r) {
+    memcpy(destbuf, srcbuf, dest->y_width);
+    destbuf += dest->y_stride;
+    srcbuf += src->y_stride;
+  }
+}
+
+static void swap_frame_buffer(YV12_BUFFER_CONFIG * const dest,
+                              YV12_BUFFER_CONFIG * const src) {
+  uint8_t *tmp_buf = dest->y_buffer;
+  assert(dest->y_width == src->y_width);
+  assert(dest->y_height == src->y_height);
+  dest->y_buffer = src->y_buffer;
+  src->y_buffer = tmp_buf;
+}
+
+void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
+                                    YV12_BUFFER_CONFIG src,
+                                    FRAME_TYPE frame_type,
+                                    int refresh_alt_ref_frame,
+                                    int refresh_golden_frame,
+                                    int refresh_last_frame,
+                                    int resized) {
+  // Copy source into denoised reference buffers on KEY_FRAME or
+  // if the just encoded frame was resized.
+  if (frame_type == KEY_FRAME || resized != 0 || denoiser->reset) {
+    int i;
+    // Start at 1 so as not to overwrite the INTRA_FRAME
+    for (i = 1; i < MAX_REF_FRAMES; ++i)
+      copy_frame(&denoiser->running_avg_y[i], &src);
+    denoiser->reset = 0;
+    return;
+  }
+
+  // If more than one refresh occurs, must copy frame buffer.
+  if ((refresh_alt_ref_frame + refresh_golden_frame + refresh_last_frame)
+      > 1) {
+    if (refresh_alt_ref_frame) {
+      copy_frame(&denoiser->running_avg_y[ALTREF_FRAME],
+                 &denoiser->running_avg_y[INTRA_FRAME]);
+    }
+    if (refresh_golden_frame) {
+      copy_frame(&denoiser->running_avg_y[GOLDEN_FRAME],
+                 &denoiser->running_avg_y[INTRA_FRAME]);
+    }
+    if (refresh_last_frame) {
+      copy_frame(&denoiser->running_avg_y[LAST_FRAME],
+                 &denoiser->running_avg_y[INTRA_FRAME]);
+    }
+  } else {
+    if (refresh_alt_ref_frame) {
+      swap_frame_buffer(&denoiser->running_avg_y[ALTREF_FRAME],
+                        &denoiser->running_avg_y[INTRA_FRAME]);
+    }
+    if (refresh_golden_frame) {
+      swap_frame_buffer(&denoiser->running_avg_y[GOLDEN_FRAME],
+                        &denoiser->running_avg_y[INTRA_FRAME]);
+    }
+    if (refresh_last_frame) {
+      swap_frame_buffer(&denoiser->running_avg_y[LAST_FRAME],
+                        &denoiser->running_avg_y[INTRA_FRAME]);
+    }
+  }
+}
+
+void vp9_denoiser_reset_frame_stats(PICK_MODE_CONTEXT *ctx) {
+  ctx->zeromv_sse = UINT_MAX;
+  ctx->newmv_sse = UINT_MAX;
+  ctx->zeromv_lastref_sse = UINT_MAX;
+  ctx->best_sse_mv.as_int = 0;
+}
+
+void vp9_denoiser_update_frame_stats(MODE_INFO *mi, unsigned int sse,
+                                     PREDICTION_MODE mode,
+                                     PICK_MODE_CONTEXT *ctx) {
+  if (mi->mv[0].as_int == 0 && sse < ctx->zeromv_sse) {
+    ctx->zeromv_sse = sse;
+    ctx->best_zeromv_reference_frame = mi->ref_frame[0];
+    if (mi->ref_frame[0] == LAST_FRAME)
+      ctx->zeromv_lastref_sse = sse;
+  }
+
+  if (mi->mv[0].as_int != 0 && sse < ctx->newmv_sse) {
+    ctx->newmv_sse = sse;
+    ctx->best_sse_inter_mode = mode;
+    ctx->best_sse_mv = mi->mv[0];
+    ctx->best_reference_frame = mi->ref_frame[0];
+  }
+}
+
+int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
+                       int ssx, int ssy,
+#if CONFIG_VP9_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       int border) {
+  int i, fail;
+  const int legacy_byte_alignment = 0;
+  assert(denoiser != NULL);
+
+  for (i = 0; i < MAX_REF_FRAMES; ++i) {
+    fail = vpx_alloc_frame_buffer(&denoiser->running_avg_y[i], width, height,
+                                  ssx, ssy,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                  use_highbitdepth,
+#endif
+                                  border, legacy_byte_alignment);
+    if (fail) {
+      vp9_denoiser_free(denoiser);
+      return 1;
+    }
+#ifdef OUTPUT_YUV_DENOISED
+    make_grayscale(&denoiser->running_avg_y[i]);
+#endif
+  }
+
+  fail = vpx_alloc_frame_buffer(&denoiser->mc_running_avg_y, width, height,
+                                ssx, ssy,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                use_highbitdepth,
+#endif
+                                border, legacy_byte_alignment);
+  if (fail) {
+    vp9_denoiser_free(denoiser);
+    return 1;
+  }
+
+  fail = vpx_alloc_frame_buffer(&denoiser->last_source, width, height,
+                                ssx, ssy,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                use_highbitdepth,
+#endif
+                                border, legacy_byte_alignment);
+  if (fail) {
+    vp9_denoiser_free(denoiser);
+    return 1;
+  }
+#ifdef OUTPUT_YUV_DENOISED
+  make_grayscale(&denoiser->running_avg_y[i]);
+#endif
+  denoiser->increase_denoising = 0;
+  denoiser->frame_buffer_initialized = 1;
+  denoiser->denoising_level = kDenLow;
+  denoiser->prev_denoising_level = kDenLow;
+  denoiser->reset = 0;
+  return 0;
+}
+
+void vp9_denoiser_free(VP9_DENOISER *denoiser) {
+  int i;
+  denoiser->frame_buffer_initialized = 0;
+  if (denoiser == NULL) {
+    return;
+  }
+  for (i = 0; i < MAX_REF_FRAMES; ++i) {
+    vpx_free_frame_buffer(&denoiser->running_avg_y[i]);
+  }
+  vpx_free_frame_buffer(&denoiser->mc_running_avg_y);
+  vpx_free_frame_buffer(&denoiser->last_source);
+}
+
+void vp9_denoiser_set_noise_level(VP9_DENOISER *denoiser,
+                                  int noise_level) {
+  denoiser->denoising_level = noise_level;
+  if (denoiser->denoising_level > kDenLowLow &&
+      denoiser->prev_denoising_level == kDenLowLow)
+    denoiser->reset = 1;
+  else
+    denoiser->reset = 0;
+  denoiser->prev_denoising_level = denoiser->denoising_level;
+}
+
+#ifdef OUTPUT_YUV_DENOISED
+static void make_grayscale(YV12_BUFFER_CONFIG *yuv) {
+  int r, c;
+  uint8_t *u = yuv->u_buffer;
+  uint8_t *v = yuv->v_buffer;
+
+  for (r = 0; r < yuv->uv_height; ++r) {
+    for (c = 0; c < yuv->uv_width; ++c) {
+      u[c] = UINT8_MAX / 2;
+      v[c] = UINT8_MAX / 2;
+    }
+    u += yuv->uv_stride;
+    v += yuv->uv_stride;
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vp9/encoder/vp9_denoiser.h b/libvpx/libvpx/vp9/encoder/vp9_denoiser.h
new file mode 100644
index 0000000..9c86e5a
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_denoiser.h
@@ -0,0 +1,107 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_DENOISER_H_
+#define VP9_ENCODER_DENOISER_H_
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_skin_detection.h"
+#include "vpx_scale/yv12config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MOTION_MAGNITUDE_THRESHOLD (8 * 3)
+
+typedef enum vp9_denoiser_decision {
+  COPY_BLOCK,
+  FILTER_BLOCK,
+  FILTER_ZEROMV_BLOCK
+} VP9_DENOISER_DECISION;
+
+typedef enum vp9_denoiser_level {
+  kDenLowLow,
+  kDenLow,
+  kDenMedium,
+  kDenHigh
+} VP9_DENOISER_LEVEL;
+
+typedef struct vp9_denoiser {
+  YV12_BUFFER_CONFIG running_avg_y[MAX_REF_FRAMES];
+  YV12_BUFFER_CONFIG mc_running_avg_y;
+  YV12_BUFFER_CONFIG last_source;
+  int increase_denoising;
+  int frame_buffer_initialized;
+  int reset;
+  VP9_DENOISER_LEVEL denoising_level;
+  VP9_DENOISER_LEVEL prev_denoising_level;
+} VP9_DENOISER;
+
+typedef struct {
+  int64_t zero_last_cost_orig;
+  int *ref_frame_cost;
+  int_mv (*frame_mv)[MAX_REF_FRAMES];
+  int reuse_inter_pred;
+  TX_SIZE best_tx_size;
+  PREDICTION_MODE best_mode;
+  MV_REFERENCE_FRAME best_ref_frame;
+  INTERP_FILTER best_pred_filter;
+  uint8_t best_mode_skip_txfm;
+} VP9_PICKMODE_CTX_DEN;
+
+struct VP9_COMP;
+
+void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
+                                    YV12_BUFFER_CONFIG src,
+                                    FRAME_TYPE frame_type,
+                                    int refresh_alt_ref_frame,
+                                    int refresh_golden_frame,
+                                    int refresh_last_frame,
+                                    int resized);
+
+void vp9_denoiser_denoise(struct VP9_COMP *cpi, MACROBLOCK *mb,
+                          int mi_row, int mi_col, BLOCK_SIZE bs,
+                          PICK_MODE_CONTEXT *ctx ,
+                          VP9_DENOISER_DECISION *denoiser_decision);
+
+void vp9_denoiser_reset_frame_stats(PICK_MODE_CONTEXT *ctx);
+
+void vp9_denoiser_update_frame_stats(MODE_INFO *mi,
+                                     unsigned int sse, PREDICTION_MODE mode,
+                                     PICK_MODE_CONTEXT *ctx);
+
+int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
+                       int ssx, int ssy,
+#if CONFIG_VP9_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       int border);
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+// This function is used by both c and sse2 denoiser implementations.
+// Define it as a static function within the scope where vp9_denoiser.h
+// is referenced.
+static INLINE int total_adj_strong_thresh(BLOCK_SIZE bs,
+                                          int increase_denoising) {
+  return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 3 : 2);
+}
+#endif
+
+void vp9_denoiser_free(VP9_DENOISER *denoiser);
+
+void vp9_denoiser_set_noise_level(VP9_DENOISER *denoiser,
+                                  int noise_level);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_DENOISER_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_encodeframe.c b/libvpx/libvpx/vp9/encoder/vp9_encodeframe.c
new file mode 100644
index 0000000..f66ed9e
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_encodeframe.c
@@ -0,0 +1,4515 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/vpx_timer.h"
+#include "vpx_ports/system_state.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_idct.h"
+#include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_tile_common.h"
+
+#include "vp9/encoder/vp9_aq_360.h"
+#include "vp9/encoder/vp9_aq_complexity.h"
+#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
+#include "vp9/encoder/vp9_aq_variance.h"
+#include "vp9/encoder/vp9_encodeframe.h"
+#include "vp9/encoder/vp9_encodemb.h"
+#include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_ethread.h"
+#include "vp9/encoder/vp9_extend.h"
+#include "vp9/encoder/vp9_pickmode.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_rdopt.h"
+#include "vp9/encoder/vp9_segmentation.h"
+#include "vp9/encoder/vp9_tokenize.h"
+
+static void encode_superblock(VP9_COMP *cpi, ThreadData * td,
+                              TOKENEXTRA **t, int output_enabled,
+                              int mi_row, int mi_col, BLOCK_SIZE bsize,
+                              PICK_MODE_CONTEXT *ctx);
+
+// This is used as a reference when computing the source variance for the
+//  purposes of activity masking.
+// Eventually this should be replaced by custom no-reference routines,
+//  which will be faster.
+static const uint8_t VP9_VAR_OFFS[64] = {
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128
+};
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128
+};
+
+static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4
+};
+
+static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16
+};
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
+                                           const struct buf_2d *ref,
+                                           BLOCK_SIZE bs) {
+  unsigned int sse;
+  const unsigned int var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
+                                              VP9_VAR_OFFS, 0, &sse);
+  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+unsigned int vp9_high_get_sby_perpixel_variance(
+    VP9_COMP *cpi, const struct buf_2d *ref, BLOCK_SIZE bs, int bd) {
+  unsigned int var, sse;
+  switch (bd) {
+    case 10:
+      var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
+                               CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10),
+                               0, &sse);
+      break;
+    case 12:
+      var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
+                               CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12),
+                               0, &sse);
+      break;
+    case 8:
+    default:
+      var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
+                               CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8),
+                               0, &sse);
+      break;
+  }
+  return ROUND_POWER_OF_TWO((int64_t)var, num_pels_log2_lookup[bs]);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
+                                                   const struct buf_2d *ref,
+                                                   int mi_row, int mi_col,
+                                                   BLOCK_SIZE bs) {
+  unsigned int sse, var;
+  uint8_t *last_y;
+  const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
+
+  assert(last != NULL);
+  last_y =
+      &last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
+  var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
+  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
+}
+
+static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi, MACROBLOCK *x,
+                                                   int mi_row,
+                                                   int mi_col) {
+  unsigned int var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
+                                                    mi_row, mi_col,
+                                                    BLOCK_64X64);
+  if (var < 8)
+    return BLOCK_64X64;
+  else if (var < 128)
+    return BLOCK_32X32;
+  else if (var < 2048)
+    return BLOCK_16X16;
+  else
+    return BLOCK_8X8;
+}
+
+// Lighter version of set_offsets that only sets the mode info
+// pointers.
+static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
+                                         MACROBLOCK *const x,
+                                         MACROBLOCKD *const xd,
+                                         int mi_row,
+                                         int mi_col) {
+  const int idx_str = xd->mi_stride * mi_row + mi_col;
+  xd->mi = cm->mi_grid_visible + idx_str;
+  xd->mi[0] = cm->mi + idx_str;
+  x->mbmi_ext = x->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
+}
+
+static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
+                        MACROBLOCK *const x, int mi_row, int mi_col,
+                        BLOCK_SIZE bsize) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *mi;
+  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
+  const int mi_height = num_8x8_blocks_high_lookup[bsize];
+  const struct segmentation *const seg = &cm->seg;
+
+  set_skip_context(xd, mi_row, mi_col);
+
+  set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
+
+  mi = xd->mi[0];
+
+  // Set up destination pointers.
+  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
+
+  // Set up limit values for MV components.
+  // Mv beyond the range do not produce new/different prediction block.
+  x->mv_row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
+  x->mv_col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
+  x->mv_row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
+  x->mv_col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
+
+  // Set up distance of MB to edge of frame in 1/8th pel units.
+  assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
+  set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
+                 cm->mi_rows, cm->mi_cols);
+
+  // Set up source buffers.
+  vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
+
+  // R/D setup.
+  x->rddiv = cpi->rd.RDDIV;
+  x->rdmult = cpi->rd.RDMULT;
+
+  // Setup segment ID.
+  if (seg->enabled) {
+    if (cpi->oxcf.aq_mode != VARIANCE_AQ &&
+        cpi->oxcf.aq_mode != EQUATOR360_AQ) {
+      const uint8_t *const map = seg->update_map ? cpi->segmentation_map
+                                                 : cm->last_frame_seg_map;
+      mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
+    }
+    vp9_init_plane_quantizers(cpi, x);
+
+    x->encode_breakout = cpi->segment_encode_breakout[mi->segment_id];
+  } else {
+    mi->segment_id = 0;
+    x->encode_breakout = cpi->encode_breakout;
+  }
+
+  // required by vp9_append_sub8x8_mvs_for_idx() and vp9_find_best_ref_mvs()
+  xd->tile = *tile;
+}
+
+static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                      int mi_row, int mi_col,
+                                      BLOCK_SIZE bsize) {
+  const int block_width = VPXMIN(num_8x8_blocks_wide_lookup[bsize],
+                                 cm->mi_cols - mi_col);
+  const int block_height = VPXMIN(num_8x8_blocks_high_lookup[bsize],
+                                  cm->mi_rows - mi_row);
+  const int mi_stride = xd->mi_stride;
+  MODE_INFO *const src_mi = xd->mi[0];
+  int i, j;
+  for (j = 0; j < block_height; ++j)
+    for (i = 0; i < block_width; ++i)
+      xd->mi[j * mi_stride + i] = src_mi;
+}
+
+static void set_block_size(VP9_COMP * const cpi,
+                           MACROBLOCK *const x,
+                           MACROBLOCKD *const xd,
+                           int mi_row, int mi_col,
+                           BLOCK_SIZE bsize) {
+  if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
+    set_mode_info_offsets(&cpi->common, x, xd, mi_row, mi_col);
+    xd->mi[0]->sb_type = bsize;
+  }
+}
+
+typedef struct {
+  int64_t sum_square_error;
+  int64_t sum_error;
+  int log2_count;
+  int variance;
+} var;
+
+typedef struct {
+  var none;
+  var horz[2];
+  var vert[2];
+} partition_variance;
+
+typedef struct {
+  partition_variance part_variances;
+  var split[4];
+} v4x4;
+
+typedef struct {
+  partition_variance part_variances;
+  v4x4 split[4];
+} v8x8;
+
+typedef struct {
+  partition_variance part_variances;
+  v8x8 split[4];
+} v16x16;
+
+typedef struct {
+  partition_variance part_variances;
+  v16x16 split[4];
+} v32x32;
+
+typedef struct {
+  partition_variance part_variances;
+  v32x32 split[4];
+} v64x64;
+
+typedef struct {
+  partition_variance *part_variances;
+  var *split[4];
+} variance_node;
+
+typedef enum {
+  V16X16,
+  V32X32,
+  V64X64,
+} TREE_LEVEL;
+
+static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
+  int i;
+  node->part_variances = NULL;
+  switch (bsize) {
+    case BLOCK_64X64: {
+      v64x64 *vt = (v64x64 *) data;
+      node->part_variances = &vt->part_variances;
+      for (i = 0; i < 4; i++)
+        node->split[i] = &vt->split[i].part_variances.none;
+      break;
+    }
+    case BLOCK_32X32: {
+      v32x32 *vt = (v32x32 *) data;
+      node->part_variances = &vt->part_variances;
+      for (i = 0; i < 4; i++)
+        node->split[i] = &vt->split[i].part_variances.none;
+      break;
+    }
+    case BLOCK_16X16: {
+      v16x16 *vt = (v16x16 *) data;
+      node->part_variances = &vt->part_variances;
+      for (i = 0; i < 4; i++)
+        node->split[i] = &vt->split[i].part_variances.none;
+      break;
+    }
+    case BLOCK_8X8: {
+      v8x8 *vt = (v8x8 *) data;
+      node->part_variances = &vt->part_variances;
+      for (i = 0; i < 4; i++)
+        node->split[i] = &vt->split[i].part_variances.none;
+      break;
+    }
+    case BLOCK_4X4: {
+      v4x4 *vt = (v4x4 *) data;
+      node->part_variances = &vt->part_variances;
+      for (i = 0; i < 4; i++)
+        node->split[i] = &vt->split[i];
+      break;
+    }
+    default: {
+      assert(0);
+      break;
+    }
+  }
+}
+
+// Set variance values given sum square error, sum error, count.
+static void fill_variance(int64_t s2, int64_t s, int c, var *v) {
+  v->sum_square_error = s2;
+  v->sum_error = s;
+  v->log2_count = c;
+}
+
+static void get_variance(var *v) {
+  v->variance = (int)(256 * (v->sum_square_error -
+      ((v->sum_error * v->sum_error) >> v->log2_count)) >> v->log2_count);
+}
+
+static void sum_2_variances(const var *a, const var *b, var *r) {
+  assert(a->log2_count == b->log2_count);
+  fill_variance(a->sum_square_error + b->sum_square_error,
+                a->sum_error + b->sum_error, a->log2_count + 1, r);
+}
+
+static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
+  variance_node node;
+  memset(&node, 0, sizeof(node));
+  tree_to_node(data, bsize, &node);
+  sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
+  sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
+  sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
+  sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
+  sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
+                  &node.part_variances->none);
+}
+
+static int set_vt_partitioning(VP9_COMP *cpi,
+                               MACROBLOCK *const x,
+                               MACROBLOCKD *const xd,
+                               void *data,
+                               BLOCK_SIZE bsize,
+                               int mi_row,
+                               int mi_col,
+                               int64_t threshold,
+                               BLOCK_SIZE bsize_min,
+                               int force_split) {
+  VP9_COMMON * const cm = &cpi->common;
+  variance_node vt;
+  const int block_width = num_8x8_blocks_wide_lookup[bsize];
+  const int block_height = num_8x8_blocks_high_lookup[bsize];
+
+  assert(block_height == block_width);
+  tree_to_node(data, bsize, &vt);
+
+  if (force_split == 1)
+    return 0;
+
+  // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
+  // variance is below threshold, otherwise split will be selected.
+  // No check for vert/horiz split as too few samples for variance.
+  if (bsize == bsize_min) {
+    // Variance already computed to set the force_split.
+    if (cm->frame_type == KEY_FRAME)
+      get_variance(&vt.part_variances->none);
+    if (mi_col + block_width / 2 < cm->mi_cols &&
+        mi_row + block_height / 2 < cm->mi_rows &&
+        vt.part_variances->none.variance < threshold) {
+      set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
+      return 1;
+    }
+    return 0;
+  } else if (bsize > bsize_min) {
+    // Variance already computed to set the force_split.
+    if (cm->frame_type == KEY_FRAME)
+      get_variance(&vt.part_variances->none);
+    // For key frame: take split for bsize above 32X32 or very high variance.
+    if (cm->frame_type == KEY_FRAME &&
+        (bsize > BLOCK_32X32 ||
+        vt.part_variances->none.variance > (threshold << 4))) {
+      return 0;
+    }
+    // If variance is low, take the bsize (no split).
+    if (mi_col + block_width / 2 < cm->mi_cols &&
+        mi_row + block_height / 2 < cm->mi_rows &&
+        vt.part_variances->none.variance < threshold) {
+      set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
+      return 1;
+    }
+
+    // Check vertical split.
+    if (mi_row + block_height / 2 < cm->mi_rows) {
+      BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
+      get_variance(&vt.part_variances->vert[0]);
+      get_variance(&vt.part_variances->vert[1]);
+      if (vt.part_variances->vert[0].variance < threshold &&
+          vt.part_variances->vert[1].variance < threshold &&
+          get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
+        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
+        set_block_size(cpi, x, xd, mi_row, mi_col + block_width / 2, subsize);
+        return 1;
+      }
+    }
+    // Check horizontal split.
+    if (mi_col + block_width / 2 < cm->mi_cols) {
+      BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
+      get_variance(&vt.part_variances->horz[0]);
+      get_variance(&vt.part_variances->horz[1]);
+      if (vt.part_variances->horz[0].variance < threshold &&
+          vt.part_variances->horz[1].variance < threshold &&
+          get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
+        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
+        set_block_size(cpi, x, xd, mi_row + block_height / 2, mi_col, subsize);
+        return 1;
+      }
+    }
+
+    return 0;
+  }
+  return 0;
+}
+
+// Set the variance split thresholds for following the block sizes:
+// 0 - threshold_64x64, 1 - threshold_32x32, 2 - threshold_16x16,
+// 3 - vbp_threshold_8x8. vbp_threshold_8x8 (to split to 4x4 partition) is
+// currently only used on key frame.
+static void set_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int is_key_frame = (cm->frame_type == KEY_FRAME);
+  const int threshold_multiplier = is_key_frame ? 20 : 1;
+  int64_t threshold_base = (int64_t)(threshold_multiplier *
+      cpi->y_dequant[q][1]);
+  if (is_key_frame) {
+    thresholds[0] = threshold_base;
+    thresholds[1] = threshold_base >> 2;
+    thresholds[2] = threshold_base >> 2;
+    thresholds[3] = threshold_base << 2;
+  } else {
+    // Increase base variance threshold based on  estimated noise level.
+    if (cpi->noise_estimate.enabled) {
+      NOISE_LEVEL noise_level = vp9_noise_estimate_extract_level(
+          &cpi->noise_estimate);
+      if (noise_level == kHigh)
+        threshold_base = 3 * threshold_base;
+      else if (noise_level == kMedium)
+        threshold_base = threshold_base << 1;
+      else if (noise_level < kLow)
+        threshold_base = (7 * threshold_base) >> 3;
+    }
+    if (cm->width <= 352 && cm->height <= 288) {
+      thresholds[0] = threshold_base >> 3;
+      thresholds[1] = threshold_base >> 1;
+      thresholds[2] = threshold_base << 3;
+    } else {
+      thresholds[0] = threshold_base;
+      thresholds[1] = (5 * threshold_base) >> 2;
+      if (cm->width >= 1920 && cm->height >= 1080)
+        thresholds[1] = (7 * threshold_base) >> 2;
+      thresholds[2] = threshold_base << cpi->oxcf.speed;
+    }
+  }
+}
+
+void vp9_set_variance_partition_thresholds(VP9_COMP *cpi, int q) {
+  VP9_COMMON *const cm = &cpi->common;
+  SPEED_FEATURES *const sf = &cpi->sf;
+  const int is_key_frame = (cm->frame_type == KEY_FRAME);
+  if (sf->partition_search_type != VAR_BASED_PARTITION &&
+      sf->partition_search_type != REFERENCE_PARTITION) {
+    return;
+  } else {
+    set_vbp_thresholds(cpi, cpi->vbp_thresholds, q);
+    // The thresholds below are not changed locally.
+    if (is_key_frame) {
+      cpi->vbp_threshold_sad = 0;
+      cpi->vbp_bsize_min = BLOCK_8X8;
+    } else {
+      if (cm->width <= 352 && cm->height <= 288)
+        cpi->vbp_threshold_sad = 10;
+      else
+        cpi->vbp_threshold_sad = (cpi->y_dequant[q][1] << 1) > 1000 ?
+            (cpi->y_dequant[q][1] << 1) : 1000;
+      cpi->vbp_bsize_min = BLOCK_16X16;
+    }
+    cpi->vbp_threshold_minmax = 15 + (q >> 3);
+  }
+}
+
+// Compute the minmax over the 8x8 subblocks.
+static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
+                              int dp, int x16_idx, int y16_idx,
+#if CONFIG_VP9_HIGHBITDEPTH
+                              int highbd_flag,
+#endif
+                              int pixels_wide,
+                              int pixels_high) {
+  int k;
+  int minmax_max = 0;
+  int minmax_min = 255;
+  // Loop over the 4 8x8 subblocks.
+  for (k = 0; k < 4; k++) {
+    int x8_idx = x16_idx + ((k & 1) << 3);
+    int y8_idx = y16_idx + ((k >> 1) << 3);
+    int min = 0;
+    int max = 0;
+    if (x8_idx < pixels_wide && y8_idx < pixels_high) {
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
+        vpx_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
+                              d + y8_idx * dp + x8_idx, dp,
+                              &min, &max);
+      } else {
+        vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
+                       d + y8_idx * dp + x8_idx, dp,
+                       &min, &max);
+      }
+#else
+      vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
+                     d + y8_idx * dp + x8_idx, dp,
+                     &min, &max);
+#endif
+      if ((max - min) > minmax_max)
+        minmax_max = (max - min);
+      if ((max - min) < minmax_min)
+        minmax_min = (max - min);
+    }
+  }
+  return (minmax_max - minmax_min);
+}
+
+static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
+                                 int dp, int x8_idx, int y8_idx, v8x8 *vst,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 int highbd_flag,
+#endif
+                                 int pixels_wide,
+                                 int pixels_high,
+                                 int is_key_frame) {
+  int k;
+  for (k = 0; k < 4; k++) {
+    int x4_idx = x8_idx + ((k & 1) << 2);
+    int y4_idx = y8_idx + ((k >> 1) << 2);
+    unsigned int sse = 0;
+    int sum = 0;
+    if (x4_idx < pixels_wide && y4_idx < pixels_high) {
+      int s_avg;
+      int d_avg = 128;
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
+        s_avg = vpx_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
+        if (!is_key_frame)
+          d_avg = vpx_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
+      } else {
+        s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
+        if (!is_key_frame)
+          d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
+      }
+#else
+      s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
+      if (!is_key_frame)
+        d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
+#endif
+      sum = s_avg - d_avg;
+      sse = sum * sum;
+    }
+    fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
+  }
+}
+
+static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
+                                 int dp, int x16_idx, int y16_idx, v16x16 *vst,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 int highbd_flag,
+#endif
+                                 int pixels_wide,
+                                 int pixels_high,
+                                 int is_key_frame) {
+  int k;
+  for (k = 0; k < 4; k++) {
+    int x8_idx = x16_idx + ((k & 1) << 3);
+    int y8_idx = y16_idx + ((k >> 1) << 3);
+    unsigned int sse = 0;
+    int sum = 0;
+    if (x8_idx < pixels_wide && y8_idx < pixels_high) {
+      int s_avg;
+      int d_avg = 128;
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
+        s_avg = vpx_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
+        if (!is_key_frame)
+          d_avg = vpx_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
+      } else {
+        s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
+        if (!is_key_frame)
+          d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
+      }
+#else
+      s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
+      if (!is_key_frame)
+        d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
+#endif
+      sum = s_avg - d_avg;
+      sse = sum * sum;
+    }
+    fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
+  }
+}
+
+#if !CONFIG_VP9_HIGHBITDEPTH
+// Check if most of the superblock is skin content, and if so, force split to
+// 32x32, and set x->sb_is_skin for use in mode selection.
+static int skin_sb_split(VP9_COMP *cpi, MACROBLOCK *x, const int low_res,
+                         int mi_row, int mi_col, int *force_split) {
+  VP9_COMMON * const cm = &cpi->common;
+  // Avoid checking superblocks on/near boundary and avoid low resolutions.
+  // Note superblock may still pick 64X64 if y_sad is very small
+  // (i.e., y_sad < cpi->vbp_threshold_sad) below. For now leave this as is.
+  if (!low_res && (mi_col >= 8 && mi_col + 8 < cm->mi_cols && mi_row >= 8 &&
+      mi_row + 8 < cm->mi_rows)) {
+    int num_16x16_skin = 0;
+    int num_16x16_nonskin = 0;
+    uint8_t *ysignal = x->plane[0].src.buf;
+    uint8_t *usignal = x->plane[1].src.buf;
+    uint8_t *vsignal = x->plane[2].src.buf;
+    int sp = x->plane[0].src.stride;
+    int spuv = x->plane[1].src.stride;
+    const int block_index = mi_row * cm->mi_cols + mi_col;
+    const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
+    const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
+    const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
+    const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
+    // Loop through the 16x16 sub-blocks.
+    int i, j;
+    for (i = 0; i < ymis; i+=2) {
+      for (j = 0; j < xmis; j+=2) {
+        int bl_index = block_index + i * cm->mi_cols + j;
+        int bl_index1 = bl_index + 1;
+        int bl_index2 = bl_index + cm->mi_cols;
+        int bl_index3 = bl_index2 + 1;
+        int consec_zeromv = VPXMIN(cpi->consec_zero_mv[bl_index],
+                                   VPXMIN(cpi->consec_zero_mv[bl_index1],
+                                   VPXMIN(cpi->consec_zero_mv[bl_index2],
+                                   cpi->consec_zero_mv[bl_index3])));
+        int is_skin = vp9_compute_skin_block(ysignal,
+                                             usignal,
+                                             vsignal,
+                                             sp,
+                                             spuv,
+                                             BLOCK_16X16,
+                                             consec_zeromv,
+                                             0);
+        num_16x16_skin += is_skin;
+        num_16x16_nonskin += (1 - is_skin);
+        if (num_16x16_nonskin > 3) {
+          // Exit loop if at least 4 of the 16x16 blocks are not skin.
+          i = ymis;
+          break;
+        }
+        ysignal += 16;
+        usignal += 8;
+        vsignal += 8;
+      }
+      ysignal += (sp << 4) - 64;
+      usignal += (spuv << 3) - 32;
+      vsignal += (spuv << 3) - 32;
+    }
+    if (num_16x16_skin > 12) {
+      *force_split = 1;
+      return 1;
+    }
+  }
+  return 0;
+}
+#endif
+
+static void set_low_temp_var_flag(VP9_COMP *cpi, MACROBLOCK *x,
+                                  MACROBLOCKD *xd, v64x64 *vt,
+                                  int force_split[], int64_t thresholds[],
+                                  MV_REFERENCE_FRAME ref_frame_partition,
+                                  int mi_col, int mi_row) {
+  int i, j;
+  VP9_COMMON * const cm = &cpi->common;
+  const int mv_thr = cm->width > 640 ? 8 : 4;
+  // Check temporal variance for bsize >= 16x16, if LAST_FRAME was selected and
+  // int_pro mv is small. If the temporal variance is small set the flag
+  // variance_low for the block. The variance threshold can be adjusted, the
+  // higher the more aggressive.
+  if (ref_frame_partition == LAST_FRAME &&
+      (cpi->sf.short_circuit_low_temp_var == 1 ||
+       (xd->mi[0]->mv[0].as_mv.col < mv_thr &&
+        xd->mi[0]->mv[0].as_mv.col > -mv_thr &&
+        xd->mi[0]->mv[0].as_mv.row < mv_thr &&
+        xd->mi[0]->mv[0].as_mv.row > -mv_thr))) {
+    if (xd->mi[0]->sb_type == BLOCK_64X64 &&
+        (vt->part_variances).none.variance < (thresholds[0] >> 1)) {
+      x->variance_low[0] = 1;
+    } else if (xd->mi[0]->sb_type == BLOCK_64X32) {
+      for (i = 0; i < 2; i++) {
+        if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
+          x->variance_low[i + 1] = 1;
+      }
+    } else if (xd->mi[0]->sb_type == BLOCK_32X64) {
+      for (i = 0; i < 2; i++) {
+        if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
+          x->variance_low[i + 3] = 1;
+      }
+    } else {
+      for (i = 0; i < 4; i++) {
+        if (!force_split[i + 1]) {
+          // 32x32
+          if (vt->split[i].part_variances.none.variance <
+              (thresholds[1] >> 1))
+            x->variance_low[i + 5] = 1;
+        } else if (cpi->sf.short_circuit_low_temp_var == 2) {
+          int idx[4] = {0, 4, xd->mi_stride << 2, (xd->mi_stride << 2) + 4};
+          const int idx_str = cm->mi_stride * mi_row + mi_col + idx[i];
+          MODE_INFO **this_mi = cm->mi_grid_visible + idx_str;
+          // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
+          // inside.
+          if ((*this_mi)->sb_type == BLOCK_16X16 ||
+              (*this_mi)->sb_type == BLOCK_32X16 ||
+              (*this_mi)->sb_type == BLOCK_16X32) {
+            for (j = 0; j < 4; j++) {
+              if (vt->split[i].split[j].part_variances.none.variance <
+                  (thresholds[2] >> 8))
+                x->variance_low[(i << 2) + j + 9] = 1;
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+// This function chooses partitioning based on the variance between source and
+// reconstructed last, where variance is computed for down-sampled inputs.
+static int choose_partitioning(VP9_COMP *cpi,
+                               const TileInfo *const tile,
+                               MACROBLOCK *x,
+                               int mi_row, int mi_col) {
+  VP9_COMMON * const cm = &cpi->common;
+  MACROBLOCKD *xd = &x->e_mbd;
+  int i, j, k, m;
+  v64x64 vt;
+  v16x16 vt2[16];
+  int force_split[21];
+  int avg_32x32;
+  int avg_16x16[4];
+  uint8_t *s;
+  const uint8_t *d;
+  int sp;
+  int dp;
+  // Ref frame used in partitioning.
+  MV_REFERENCE_FRAME ref_frame_partition = LAST_FRAME;
+  int pixels_wide = 64, pixels_high = 64;
+  int64_t thresholds[4] = {cpi->vbp_thresholds[0], cpi->vbp_thresholds[1],
+      cpi->vbp_thresholds[2], cpi->vbp_thresholds[3]};
+
+  // For the variance computation under SVC mode, we treat the frame as key if
+  // the reference (base layer frame) is key frame (i.e., is_key_frame == 1).
+  const int is_key_frame = (cm->frame_type == KEY_FRAME ||
+      (is_one_pass_cbr_svc(cpi) &&
+      cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame));
+  // Always use 4x4 partition for key frame.
+  const int use_4x4_partition = cm->frame_type == KEY_FRAME;
+  const int low_res = (cm->width <= 352 && cm->height <= 288);
+  int variance4x4downsample[16];
+  int segment_id;
+
+  set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
+  segment_id = xd->mi[0]->segment_id;
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
+    if (cyclic_refresh_segment_id_boosted(segment_id)) {
+      int q = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+      set_vbp_thresholds(cpi, thresholds, q);
+    }
+  }
+
+  memset(x->variance_low, 0, sizeof(x->variance_low));
+
+  if (xd->mb_to_right_edge < 0)
+    pixels_wide += (xd->mb_to_right_edge >> 3);
+  if (xd->mb_to_bottom_edge < 0)
+    pixels_high += (xd->mb_to_bottom_edge >> 3);
+
+  s = x->plane[0].src.buf;
+  sp = x->plane[0].src.stride;
+
+  // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
+  // 5-20 for the 16x16 blocks.
+  force_split[0] = 0;
+
+  if (!is_key_frame) {
+    // In the case of spatial/temporal scalable coding, the assumption here is
+    // that the temporal reference frame will always be of type LAST_FRAME.
+    // TODO(marpan): If that assumption is broken, we need to revisit this code.
+    MODE_INFO *mi = xd->mi[0];
+    unsigned int uv_sad;
+    const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
+
+    const YV12_BUFFER_CONFIG *yv12_g = NULL;
+    unsigned int y_sad, y_sad_g, y_sad_thr;
+    const BLOCK_SIZE bsize = BLOCK_32X32
+        + (mi_col + 4 < cm->mi_cols) * 2 + (mi_row + 4 < cm->mi_rows);
+
+    assert(yv12 != NULL);
+
+    if (!(is_one_pass_cbr_svc(cpi) && cpi->svc.spatial_layer_id)) {
+      // For now, GOLDEN will not be used for non-zero spatial layers, since
+      // it may not be a temporal reference.
+      yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
+    }
+
+    if (yv12_g && yv12_g != yv12 &&
+       (cpi->ref_frame_flags & VP9_GOLD_FLAG)) {
+      vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
+                           &cm->frame_refs[GOLDEN_FRAME - 1].sf);
+      y_sad_g = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
+                                       x->plane[0].src.stride,
+                                       xd->plane[0].pre[0].buf,
+                                       xd->plane[0].pre[0].stride);
+    } else {
+      y_sad_g = UINT_MAX;
+    }
+
+    vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
+                         &cm->frame_refs[LAST_FRAME - 1].sf);
+    mi->ref_frame[0] = LAST_FRAME;
+    mi->ref_frame[1] = NONE;
+    mi->sb_type = BLOCK_64X64;
+    mi->mv[0].as_int = 0;
+    mi->interp_filter = BILINEAR;
+
+    y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
+    // Pick ref frame for partitioning, bias last frame when y_sad_g and y_sad
+    // are close if short_circuit_low_temp_var is on.
+    y_sad_thr = cpi->sf.short_circuit_low_temp_var ? (y_sad * 7) >> 3 : y_sad;
+    if (y_sad_g < y_sad_thr) {
+      vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
+                           &cm->frame_refs[GOLDEN_FRAME - 1].sf);
+      mi->ref_frame[0] = GOLDEN_FRAME;
+      mi->mv[0].as_int = 0;
+      y_sad = y_sad_g;
+      ref_frame_partition = GOLDEN_FRAME;
+    } else {
+      x->pred_mv[LAST_FRAME] = mi->mv[0].as_mv;
+      ref_frame_partition = LAST_FRAME;
+    }
+
+    set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
+    vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
+
+    x->sb_is_skin = 0;
+#if !CONFIG_VP9_HIGHBITDEPTH
+    if (cpi->use_skin_detection)
+      x->sb_is_skin = skin_sb_split(cpi, x, low_res, mi_row, mi_col,
+                                    &force_split[0]);
+#endif
+
+    for (i = 1; i <= 2; ++i) {
+      struct macroblock_plane  *p = &x->plane[i];
+      struct macroblockd_plane *pd = &xd->plane[i];
+      const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
+
+      if (bs == BLOCK_INVALID)
+        uv_sad = UINT_MAX;
+      else
+        uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride,
+                                     pd->dst.buf, pd->dst.stride);
+
+        // TODO(marpan): Investigate if we should lower this threshold if
+        // superblock is detected as skin.
+        x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
+    }
+
+    d = xd->plane[0].dst.buf;
+    dp = xd->plane[0].dst.stride;
+
+    // If the y_sad is very small, take 64x64 as partition and exit.
+    // Don't check on boosted segment for now, as 64x64 is suppressed there.
+    if (segment_id == CR_SEGMENT_ID_BASE &&
+        y_sad < cpi->vbp_threshold_sad) {
+      const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
+      const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
+      if (mi_col + block_width / 2 < cm->mi_cols &&
+          mi_row + block_height / 2 < cm->mi_rows) {
+        set_block_size(cpi, x, xd, mi_row, mi_col, BLOCK_64X64);
+        return 0;
+      }
+    }
+  } else {
+    d = VP9_VAR_OFFS;
+    dp = 0;
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      switch (xd->bd) {
+        case 10:
+          d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10);
+          break;
+        case 12:
+          d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12);
+          break;
+        case 8:
+        default:
+          d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8);
+          break;
+      }
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+
+  // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
+  // for splits.
+  for (i = 0; i < 4; i++) {
+    const int x32_idx = ((i & 1) << 5);
+    const int y32_idx = ((i >> 1) << 5);
+    const int i2 = i << 2;
+    force_split[i + 1] = 0;
+    avg_16x16[i] = 0;
+    for (j = 0; j < 4; j++) {
+      const int x16_idx = x32_idx + ((j & 1) << 4);
+      const int y16_idx = y32_idx + ((j >> 1) << 4);
+      const int split_index = 5 + i2 + j;
+      v16x16 *vst = &vt.split[i].split[j];
+      force_split[split_index] = 0;
+      variance4x4downsample[i2 + j] = 0;
+      if (!is_key_frame) {
+        fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
+#if CONFIG_VP9_HIGHBITDEPTH
+                            xd->cur_buf->flags,
+#endif
+                            pixels_wide,
+                            pixels_high,
+                            is_key_frame);
+        fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
+        get_variance(&vt.split[i].split[j].part_variances.none);
+        avg_16x16[i] += vt.split[i].split[j].part_variances.none.variance;
+        if (vt.split[i].split[j].part_variances.none.variance >
+            thresholds[2]) {
+          // 16X16 variance is above threshold for split, so force split to 8x8
+          // for this 16x16 block (this also forces splits for upper levels).
+          force_split[split_index] = 1;
+          force_split[i + 1] = 1;
+          force_split[0] = 1;
+        } else if (cpi->oxcf.speed < 8 &&
+                   vt.split[i].split[j].part_variances.none.variance >
+                   thresholds[1] &&
+                   !cyclic_refresh_segment_id_boosted(segment_id)) {
+          // We have some nominal amount of 16x16 variance (based on average),
+          // compute the minmax over the 8x8 sub-blocks, and if above threshold,
+          // force split to 8x8 block for this 16x16 block.
+          int minmax = compute_minmax_8x8(s, sp, d, dp, x16_idx, y16_idx,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                          xd->cur_buf->flags,
+#endif
+                                          pixels_wide, pixels_high);
+          if (minmax > cpi->vbp_threshold_minmax) {
+            force_split[split_index] = 1;
+            force_split[i + 1] = 1;
+            force_split[0] = 1;
+          }
+        }
+      }
+      if (is_key_frame || (low_res &&
+          vt.split[i].split[j].part_variances.none.variance >
+          (thresholds[1] << 1))) {
+        force_split[split_index] = 0;
+        // Go down to 4x4 down-sampling for variance.
+        variance4x4downsample[i2 + j] = 1;
+        for (k = 0; k < 4; k++) {
+          int x8_idx = x16_idx + ((k & 1) << 3);
+          int y8_idx = y16_idx + ((k >> 1) << 3);
+          v8x8 *vst2 = is_key_frame ? &vst->split[k] :
+              &vt2[i2 + j].split[k];
+          fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
+#if CONFIG_VP9_HIGHBITDEPTH
+                               xd->cur_buf->flags,
+#endif
+                               pixels_wide,
+                               pixels_high,
+                               is_key_frame);
+        }
+      }
+    }
+  }
+  // Fill the rest of the variance tree by summing split partition values.
+  avg_32x32 = 0;
+  for (i = 0; i < 4; i++) {
+    const int i2 = i << 2;
+    for (j = 0; j < 4; j++) {
+      if (variance4x4downsample[i2 + j] == 1) {
+        v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] :
+            &vt.split[i].split[j];
+        for (m = 0; m < 4; m++)
+          fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
+        fill_variance_tree(vtemp, BLOCK_16X16);
+        // If variance of this 16x16 block is above the threshold, force block
+        // to split. This also forces a split on the upper levels.
+        get_variance(&vtemp->part_variances.none);
+        if (vtemp->part_variances.none.variance > thresholds[2]) {
+          force_split[5 + i2 + j] = 1;
+          force_split[i + 1] = 1;
+          force_split[0] = 1;
+        }
+      }
+    }
+    fill_variance_tree(&vt.split[i], BLOCK_32X32);
+    // If variance of this 32x32 block is above the threshold, or if its above
+    // (some threshold of) the average variance over the sub-16x16 blocks, then
+    // force this block to split. This also forces a split on the upper
+    // (64x64) level.
+    if (!force_split[i + 1]) {
+      get_variance(&vt.split[i].part_variances.none);
+      if (vt.split[i].part_variances.none.variance > thresholds[1] ||
+          (!is_key_frame &&
+          vt.split[i].part_variances.none.variance > (thresholds[1] >> 1) &&
+          vt.split[i].part_variances.none.variance > (avg_16x16[i] >> 1))) {
+        force_split[i + 1] = 1;
+        force_split[0] = 1;
+      }
+      avg_32x32 += vt.split[i].part_variances.none.variance;
+    }
+  }
+  if (!force_split[0]) {
+    fill_variance_tree(&vt, BLOCK_64X64);
+    get_variance(&vt.part_variances.none);
+    // If variance of this 64x64 block is above (some threshold of) the average
+    // variance over the sub-32x32 blocks, then force this block to split.
+    if (!is_key_frame &&
+        vt.part_variances.none.variance > (5 * avg_32x32) >> 4)
+      force_split[0] = 1;
+  }
+
+  // Now go through the entire structure, splitting every block size until
+  // we get to one that's got a variance lower than our threshold.
+  if ( mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
+      !set_vt_partitioning(cpi, x, xd, &vt, BLOCK_64X64, mi_row, mi_col,
+                           thresholds[0], BLOCK_16X16, force_split[0])) {
+    for (i = 0; i < 4; ++i) {
+      const int x32_idx = ((i & 1) << 2);
+      const int y32_idx = ((i >> 1) << 2);
+      const int i2 = i << 2;
+      if (!set_vt_partitioning(cpi, x, xd, &vt.split[i], BLOCK_32X32,
+                               (mi_row + y32_idx), (mi_col + x32_idx),
+                               thresholds[1], BLOCK_16X16,
+                               force_split[i + 1])) {
+        for (j = 0; j < 4; ++j) {
+          const int x16_idx = ((j & 1) << 1);
+          const int y16_idx = ((j >> 1) << 1);
+          // For inter frames: if variance4x4downsample[] == 1 for this 16x16
+          // block, then the variance is based on 4x4 down-sampling, so use vt2
+          // in set_vt_partioning(), otherwise use vt.
+          v16x16 *vtemp = (!is_key_frame &&
+                           variance4x4downsample[i2 + j] == 1) ?
+                           &vt2[i2 + j] : &vt.split[i].split[j];
+          if (!set_vt_partitioning(cpi, x, xd, vtemp, BLOCK_16X16,
+                                   mi_row + y32_idx + y16_idx,
+                                   mi_col + x32_idx + x16_idx,
+                                   thresholds[2],
+                                   cpi->vbp_bsize_min,
+                                   force_split[5 + i2  + j])) {
+            for (k = 0; k < 4; ++k) {
+              const int x8_idx = (k & 1);
+              const int y8_idx = (k >> 1);
+              if (use_4x4_partition) {
+                if (!set_vt_partitioning(cpi, x, xd, &vtemp->split[k],
+                                         BLOCK_8X8,
+                                         mi_row + y32_idx + y16_idx + y8_idx,
+                                         mi_col + x32_idx + x16_idx + x8_idx,
+                                         thresholds[3], BLOCK_8X8, 0)) {
+                  set_block_size(cpi, x, xd,
+                                 (mi_row + y32_idx + y16_idx + y8_idx),
+                                 (mi_col + x32_idx + x16_idx + x8_idx),
+                                 BLOCK_4X4);
+                }
+              } else {
+                set_block_size(cpi, x, xd,
+                               (mi_row + y32_idx + y16_idx + y8_idx),
+                               (mi_col + x32_idx + x16_idx + x8_idx),
+                               BLOCK_8X8);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  if (cpi->sf.short_circuit_low_temp_var) {
+    set_low_temp_var_flag(cpi, x, xd, &vt, force_split, thresholds,
+                          ref_frame_partition, mi_col, mi_row);
+  }
+  return 0;
+}
+
+static void update_state(VP9_COMP *cpi, ThreadData *td,
+                         PICK_MODE_CONTEXT *ctx,
+                         int mi_row, int mi_col, BLOCK_SIZE bsize,
+                         int output_enabled) {
+  int i, x_idx, y;
+  VP9_COMMON *const cm = &cpi->common;
+  RD_COUNTS *const rdc = &td->rd_counts;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *const p = x->plane;
+  struct macroblockd_plane *const pd = xd->plane;
+  MODE_INFO *mi = &ctx->mic;
+  MODE_INFO *const xdmi = xd->mi[0];
+  MODE_INFO *mi_addr = xd->mi[0];
+  const struct segmentation *const seg = &cm->seg;
+  const int bw = num_8x8_blocks_wide_lookup[mi->sb_type];
+  const int bh = num_8x8_blocks_high_lookup[mi->sb_type];
+  const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
+  MV_REF *const frame_mvs =
+      cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
+  int w, h;
+
+  const int mis = cm->mi_stride;
+  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
+  const int mi_height = num_8x8_blocks_high_lookup[bsize];
+  int max_plane;
+
+  assert(mi->sb_type == bsize);
+
+  *mi_addr = *mi;
+  *x->mbmi_ext = ctx->mbmi_ext;
+
+  // If segmentation in use
+  if (seg->enabled) {
+    // For in frame complexity AQ copy the segment id from the segment map.
+    if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
+      const uint8_t *const map = seg->update_map ? cpi->segmentation_map
+                                                 : cm->last_frame_seg_map;
+      mi_addr->segment_id =
+        get_segment_id(cm, map, bsize, mi_row, mi_col);
+    }
+    // Else for cyclic refresh mode update the segment map, set the segment id
+    // and then update the quantizer.
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
+      vp9_cyclic_refresh_update_segment(cpi, xd->mi[0], mi_row,
+                                        mi_col, bsize, ctx->rate, ctx->dist,
+                                        x->skip, p);
+    }
+  }
+
+  max_plane = is_inter_block(xdmi) ? MAX_MB_PLANE : 1;
+  for (i = 0; i < max_plane; ++i) {
+    p[i].coeff = ctx->coeff_pbuf[i][1];
+    p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
+    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
+    p[i].eobs = ctx->eobs_pbuf[i][1];
+  }
+
+  for (i = max_plane; i < MAX_MB_PLANE; ++i) {
+    p[i].coeff = ctx->coeff_pbuf[i][2];
+    p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
+    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
+    p[i].eobs = ctx->eobs_pbuf[i][2];
+  }
+
+  // Restore the coding context of the MB to that that was in place
+  // when the mode was picked for it
+  for (y = 0; y < mi_height; y++)
+    for (x_idx = 0; x_idx < mi_width; x_idx++)
+      if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
+        && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
+        xd->mi[x_idx + y * mis] = mi_addr;
+      }
+
+  if (cpi->oxcf.aq_mode != NO_AQ)
+    vp9_init_plane_quantizers(cpi, x);
+
+  if (is_inter_block(xdmi) && xdmi->sb_type < BLOCK_8X8) {
+    xdmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
+    xdmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
+  }
+
+  x->skip = ctx->skip;
+  memcpy(x->zcoeff_blk[xdmi->tx_size], ctx->zcoeff_blk,
+         sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
+
+  if (!output_enabled)
+    return;
+
+#if CONFIG_INTERNAL_STATS
+  if (frame_is_intra_only(cm)) {
+    static const int kf_mode_index[] = {
+      THR_DC        /*DC_PRED*/,
+      THR_V_PRED    /*V_PRED*/,
+      THR_H_PRED    /*H_PRED*/,
+      THR_D45_PRED  /*D45_PRED*/,
+      THR_D135_PRED /*D135_PRED*/,
+      THR_D117_PRED /*D117_PRED*/,
+      THR_D153_PRED /*D153_PRED*/,
+      THR_D207_PRED /*D207_PRED*/,
+      THR_D63_PRED  /*D63_PRED*/,
+      THR_TM        /*TM_PRED*/,
+    };
+    ++cpi->mode_chosen_counts[kf_mode_index[xdmi->mode]];
+  } else {
+    // Note how often each mode chosen as best
+    ++cpi->mode_chosen_counts[ctx->best_mode_index];
+  }
+#endif
+  if (!frame_is_intra_only(cm)) {
+    if (is_inter_block(xdmi)) {
+      vp9_update_mv_count(td);
+
+      if (cm->interp_filter == SWITCHABLE) {
+        const int ctx = vp9_get_pred_context_switchable_interp(xd);
+        ++td->counts->switchable_interp[ctx][xdmi->interp_filter];
+      }
+    }
+
+    rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
+    rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
+    rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
+
+    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+      rdc->filter_diff[i] += ctx->best_filter_diff[i];
+  }
+
+  for (h = 0; h < y_mis; ++h) {
+    MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
+    for (w = 0; w < x_mis; ++w) {
+      MV_REF *const mv = frame_mv + w;
+      mv->ref_frame[0] = mi->ref_frame[0];
+      mv->ref_frame[1] = mi->ref_frame[1];
+      mv->mv[0].as_int = mi->mv[0].as_int;
+      mv->mv[1].as_int = mi->mv[1].as_int;
+    }
+  }
+}
+
+void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col) {
+  uint8_t *const buffers[3] = {src->y_buffer, src->u_buffer, src->v_buffer };
+  const int strides[3] = {src->y_stride, src->uv_stride, src->uv_stride };
+  int i;
+
+  // Set current frame pointer.
+  x->e_mbd.cur_buf = src;
+
+  for (i = 0; i < MAX_MB_PLANE; i++)
+    setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
+                     NULL, x->e_mbd.plane[i].subsampling_x,
+                     x->e_mbd.plane[i].subsampling_y);
+}
+
+static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
+                                   RD_COST *rd_cost, BLOCK_SIZE bsize) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  INTERP_FILTER filter_ref;
+
+  filter_ref = vp9_get_pred_context_switchable_interp(xd);
+  if (filter_ref == SWITCHABLE_FILTERS)
+    filter_ref = EIGHTTAP;
+
+  mi->sb_type = bsize;
+  mi->mode = ZEROMV;
+  mi->tx_size =
+      VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[tx_mode]);
+  mi->skip = 1;
+  mi->uv_mode = DC_PRED;
+  mi->ref_frame[0] = LAST_FRAME;
+  mi->ref_frame[1] = NONE;
+  mi->mv[0].as_int = 0;
+  mi->interp_filter = filter_ref;
+
+  xd->mi[0]->bmi[0].as_mv[0].as_int = 0;
+  x->skip = 1;
+
+  vp9_rd_cost_init(rd_cost);
+}
+
+static int set_segment_rdmult(VP9_COMP *const cpi,
+                               MACROBLOCK *const x,
+                               int8_t segment_id) {
+  int segment_qindex;
+  VP9_COMMON *const cm = &cpi->common;
+  vp9_init_plane_quantizers(cpi, x);
+  vpx_clear_system_state();
+  segment_qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+  return vp9_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
+}
+
+static void rd_pick_sb_modes(VP9_COMP *cpi,
+                             TileDataEnc *tile_data,
+                             MACROBLOCK *const x,
+                             int mi_row, int mi_col, RD_COST *rd_cost,
+                             BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                             int64_t best_rd) {
+  VP9_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *mi;
+  struct macroblock_plane *const p = x->plane;
+  struct macroblockd_plane *const pd = xd->plane;
+  const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
+  int i, orig_rdmult;
+
+  vpx_clear_system_state();
+
+  // Use the lower precision, but faster, 32x32 fdct for mode selection.
+  x->use_lp32x32fdct = 1;
+
+  set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+  mi = xd->mi[0];
+  mi->sb_type = bsize;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    p[i].coeff = ctx->coeff_pbuf[i][0];
+    p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
+    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
+    p[i].eobs = ctx->eobs_pbuf[i][0];
+  }
+  ctx->is_coded = 0;
+  ctx->skippable = 0;
+  ctx->pred_pixel_ready = 0;
+  x->skip_recode = 0;
+
+  // Set to zero to make sure we do not use the previous encoded frame stats
+  mi->skip = 0;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    x->source_variance =
+        vp9_high_get_sby_perpixel_variance(cpi, &x->plane[0].src,
+                                           bsize, xd->bd);
+  } else {
+    x->source_variance =
+      vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
+  }
+#else
+  x->source_variance =
+    vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  // Save rdmult before it might be changed, so it can be restored later.
+  orig_rdmult = x->rdmult;
+
+  if (aq_mode == VARIANCE_AQ) {
+    const int energy = bsize <= BLOCK_16X16 ? x->mb_energy
+                                            : vp9_block_energy(cpi, x, bsize);
+    if (cm->frame_type == KEY_FRAME ||
+        cpi->refresh_alt_ref_frame ||
+        (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+      mi->segment_id = vp9_vaq_segment_id(energy);
+    } else {
+      const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
+                                                    : cm->last_frame_seg_map;
+      mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
+    }
+    x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
+  } else if (aq_mode == EQUATOR360_AQ) {
+    if (cm->frame_type == KEY_FRAME) {
+      mi->segment_id = vp9_360aq_segment_id(mi_row, cm->mi_rows);
+    } else {
+      const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
+                                                    : cm->last_frame_seg_map;
+      mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
+    }
+    x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
+  } else if (aq_mode == COMPLEXITY_AQ) {
+    x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
+  } else if (aq_mode == CYCLIC_REFRESH_AQ) {
+    const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
+                                                  : cm->last_frame_seg_map;
+    // If segment is boosted, use rdmult for that segment.
+    if (cyclic_refresh_segment_id_boosted(
+            get_segment_id(cm, map, bsize, mi_row, mi_col)))
+      x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
+  }
+
+  // Find best coding mode & reconstruct the MB so it is available
+  // as a predictor for MBs that follow in the SB
+  if (frame_is_intra_only(cm)) {
+    vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
+  } else {
+    if (bsize >= BLOCK_8X8) {
+      if (segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP))
+        vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
+                                           ctx, best_rd);
+      else
+        vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col,
+                                  rd_cost, bsize, ctx, best_rd);
+    } else {
+      vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col,
+                                    rd_cost, bsize, ctx, best_rd);
+    }
+  }
+
+
+  // Examine the resulting rate and for AQ mode 2 make a segment choice.
+  if ((rd_cost->rate != INT_MAX) &&
+      (aq_mode == COMPLEXITY_AQ) && (bsize >= BLOCK_16X16) &&
+      (cm->frame_type == KEY_FRAME ||
+       cpi->refresh_alt_ref_frame ||
+       (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
+    vp9_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
+  }
+
+  x->rdmult = orig_rdmult;
+
+  // TODO(jingning) The rate-distortion optimization flow needs to be
+  // refactored to provide proper exit/return handle.
+  if (rd_cost->rate == INT_MAX)
+    rd_cost->rdcost = INT64_MAX;
+
+  ctx->rate = rd_cost->rate;
+  ctx->dist = rd_cost->dist;
+}
+
+static void update_stats(VP9_COMMON *cm, ThreadData *td) {
+  const MACROBLOCK *x = &td->mb;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const MODE_INFO *const mi = xd->mi[0];
+  const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  const BLOCK_SIZE bsize = mi->sb_type;
+
+  if (!frame_is_intra_only(cm)) {
+    FRAME_COUNTS *const counts = td->counts;
+    const int inter_block = is_inter_block(mi);
+    const int seg_ref_active = segfeature_active(&cm->seg, mi->segment_id,
+                                                 SEG_LVL_REF_FRAME);
+    if (!seg_ref_active) {
+      counts->intra_inter[get_intra_inter_context(xd)][inter_block]++;
+      // If the segment reference feature is enabled we have only a single
+      // reference frame allowed for the segment so exclude it from
+      // the reference frame counts used to work out probabilities.
+      if (inter_block) {
+        const MV_REFERENCE_FRAME ref0 = mi->ref_frame[0];
+        if (cm->reference_mode == REFERENCE_MODE_SELECT)
+          counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
+                            [has_second_ref(mi)]++;
+
+        if (has_second_ref(mi)) {
+          counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
+                          [ref0 == GOLDEN_FRAME]++;
+        } else {
+          counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
+                            [ref0 != LAST_FRAME]++;
+          if (ref0 != LAST_FRAME)
+            counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
+                              [ref0 != GOLDEN_FRAME]++;
+        }
+      }
+    }
+    if (inter_block &&
+        !segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP)) {
+      const int mode_ctx = mbmi_ext->mode_context[mi->ref_frame[0]];
+      if (bsize >= BLOCK_8X8) {
+        const PREDICTION_MODE mode = mi->mode;
+        ++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
+      } else {
+        const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+        const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+        int idx, idy;
+        for (idy = 0; idy < 2; idy += num_4x4_h) {
+          for (idx = 0; idx < 2; idx += num_4x4_w) {
+            const int j = idy * 2 + idx;
+            const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
+            ++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
+          }
+        }
+      }
+    }
+  }
+}
+
+static void restore_context(MACROBLOCK *const x, int mi_row, int mi_col,
+                            ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
+                            ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
+                            PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
+                            BLOCK_SIZE bsize) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int p;
+  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
+  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
+  int mi_width = num_8x8_blocks_wide_lookup[bsize];
+  int mi_height = num_8x8_blocks_high_lookup[bsize];
+  for (p = 0; p < MAX_MB_PLANE; p++) {
+    memcpy(
+        xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
+        a + num_4x4_blocks_wide * p,
+        (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
+        xd->plane[p].subsampling_x);
+    memcpy(
+        xd->left_context[p]
+            + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
+        l + num_4x4_blocks_high * p,
+        (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
+        xd->plane[p].subsampling_y);
+  }
+  memcpy(xd->above_seg_context + mi_col, sa,
+         sizeof(*xd->above_seg_context) * mi_width);
+  memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
+         sizeof(xd->left_seg_context[0]) * mi_height);
+}
+
+static void save_context(MACROBLOCK *const x, int mi_row, int mi_col,
+                         ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
+                         ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
+                         PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
+                         BLOCK_SIZE bsize) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  int p;
+  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
+  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
+  int mi_width = num_8x8_blocks_wide_lookup[bsize];
+  int mi_height = num_8x8_blocks_high_lookup[bsize];
+
+  // buffer the above/left context information of the block in search.
+  for (p = 0; p < MAX_MB_PLANE; ++p) {
+    memcpy(
+        a + num_4x4_blocks_wide * p,
+        xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
+        (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
+        xd->plane[p].subsampling_x);
+    memcpy(
+        l + num_4x4_blocks_high * p,
+        xd->left_context[p]
+            + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
+        (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
+        xd->plane[p].subsampling_y);
+  }
+  memcpy(sa, xd->above_seg_context + mi_col,
+         sizeof(*xd->above_seg_context) * mi_width);
+  memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
+         sizeof(xd->left_seg_context[0]) * mi_height);
+}
+
+static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
+                     ThreadData *td,
+                     TOKENEXTRA **tp, int mi_row, int mi_col,
+                     int output_enabled, BLOCK_SIZE bsize,
+                     PICK_MODE_CONTEXT *ctx) {
+  MACROBLOCK *const x = &td->mb;
+  set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
+  update_state(cpi, td, ctx, mi_row, mi_col, bsize, output_enabled);
+  encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
+
+  if (output_enabled) {
+    update_stats(&cpi->common, td);
+
+    (*tp)->token = EOSB_TOKEN;
+    (*tp)++;
+  }
+}
+
+static void encode_sb(VP9_COMP *cpi, ThreadData *td,
+                      const TileInfo *const tile,
+                      TOKENEXTRA **tp, int mi_row, int mi_col,
+                      int output_enabled, BLOCK_SIZE bsize,
+                      PC_TREE *pc_tree) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+
+  const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
+  int ctx;
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize = bsize;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  if (bsize >= BLOCK_8X8) {
+    ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
+    subsize = get_subsize(bsize, pc_tree->partitioning);
+  } else {
+    ctx = 0;
+    subsize = BLOCK_4X4;
+  }
+
+  partition = partition_lookup[bsl][subsize];
+  if (output_enabled && bsize != BLOCK_4X4)
+    td->counts->partition[ctx][partition]++;
+
+  switch (partition) {
+    case PARTITION_NONE:
+      encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
+               &pc_tree->none);
+      break;
+    case PARTITION_VERT:
+      encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
+               &pc_tree->vertical[0]);
+      if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
+        encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, output_enabled,
+                 subsize, &pc_tree->vertical[1]);
+      }
+      break;
+    case PARTITION_HORZ:
+      encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
+               &pc_tree->horizontal[0]);
+      if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
+        encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, output_enabled,
+                 subsize, &pc_tree->horizontal[1]);
+      }
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8) {
+        encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
+                 pc_tree->leaf_split[0]);
+      } else {
+        encode_sb(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                  pc_tree->split[0]);
+        encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
+                  subsize, pc_tree->split[1]);
+        encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
+                  subsize, pc_tree->split[2]);
+        encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
+                  subsize, pc_tree->split[3]);
+      }
+      break;
+    default:
+      assert(0 && "Invalid partition type.");
+      break;
+  }
+
+  if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
+    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+}
+
+// Check to see if the given partition size is allowed for a specified number
+// of 8x8 block rows and columns remaining in the image.
+// If not then return the largest allowed partition size
+static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize,
+                                      int rows_left, int cols_left,
+                                      int *bh, int *bw) {
+  if (rows_left <= 0 || cols_left <= 0) {
+    return VPXMIN(bsize, BLOCK_8X8);
+  } else {
+    for (; bsize > 0; bsize -= 3) {
+      *bh = num_8x8_blocks_high_lookup[bsize];
+      *bw = num_8x8_blocks_wide_lookup[bsize];
+      if ((*bh <= rows_left) && (*bw <= cols_left)) {
+        break;
+      }
+    }
+  }
+  return bsize;
+}
+
+static void set_partial_b64x64_partition(MODE_INFO *mi, int mis,
+    int bh_in, int bw_in, int row8x8_remaining, int col8x8_remaining,
+    BLOCK_SIZE bsize, MODE_INFO **mi_8x8) {
+  int bh = bh_in;
+  int r, c;
+  for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
+    int bw = bw_in;
+    for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
+      const int index = r * mis + c;
+      mi_8x8[index] = mi + index;
+      mi_8x8[index]->sb_type = find_partition_size(bsize,
+          row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
+    }
+  }
+}
+
+// This function attempts to set all mode info entries in a given SB64
+// to the same block partition size.
+// However, at the bottom and right borders of the image the requested size
+// may not be allowed in which case this code attempts to choose the largest
+// allowable partition.
+static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
+                                   MODE_INFO **mi_8x8, int mi_row, int mi_col,
+                                   BLOCK_SIZE bsize) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int mis = cm->mi_stride;
+  const int row8x8_remaining = tile->mi_row_end - mi_row;
+  const int col8x8_remaining = tile->mi_col_end - mi_col;
+  int block_row, block_col;
+  MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
+  int bh = num_8x8_blocks_high_lookup[bsize];
+  int bw = num_8x8_blocks_wide_lookup[bsize];
+
+  assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
+
+  // Apply the requested partition size to the SB64 if it is all "in image"
+  if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
+      (row8x8_remaining >= MI_BLOCK_SIZE)) {
+    for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
+      for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
+        int index = block_row * mis + block_col;
+        mi_8x8[index] = mi_upper_left + index;
+        mi_8x8[index]->sb_type = bsize;
+      }
+    }
+  } else {
+    // Else this is a partial SB64.
+    set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
+        col8x8_remaining, bsize, mi_8x8);
+  }
+}
+
+static const struct {
+  int row;
+  int col;
+} coord_lookup[16] = {
+    // 32x32 index = 0
+    {0, 0}, {0, 2}, {2, 0}, {2, 2},
+    // 32x32 index = 1
+    {0, 4}, {0, 6}, {2, 4}, {2, 6},
+    // 32x32 index = 2
+    {4, 0}, {4, 2}, {6, 0}, {6, 2},
+    // 32x32 index = 3
+    {4, 4}, {4, 6}, {6, 4}, {6, 6},
+};
+
+static void set_source_var_based_partition(VP9_COMP *cpi,
+                                           const TileInfo *const tile,
+                                           MACROBLOCK *const x,
+                                           MODE_INFO **mi_8x8,
+                                           int mi_row, int mi_col) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int mis = cm->mi_stride;
+  const int row8x8_remaining = tile->mi_row_end - mi_row;
+  const int col8x8_remaining = tile->mi_col_end - mi_col;
+  MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
+
+  vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
+
+  assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
+
+  // In-image SB64
+  if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
+      (row8x8_remaining >= MI_BLOCK_SIZE)) {
+    int i, j;
+    int index;
+    diff d32[4];
+    const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
+    int is_larger_better = 0;
+    int use32x32 = 0;
+    unsigned int thr = cpi->source_var_thresh;
+
+    memset(d32, 0, 4 * sizeof(diff));
+
+    for (i = 0; i < 4; i++) {
+      diff *d16[4];
+
+      for (j = 0; j < 4; j++) {
+        int b_mi_row = coord_lookup[i * 4 + j].row;
+        int b_mi_col = coord_lookup[i * 4 + j].col;
+        int boffset = b_mi_row / 2 * cm->mb_cols +
+                      b_mi_col / 2;
+
+        d16[j] = cpi->source_diff_var + offset + boffset;
+
+        index = b_mi_row * mis + b_mi_col;
+        mi_8x8[index] = mi_upper_left + index;
+        mi_8x8[index]->sb_type = BLOCK_16X16;
+
+        // TODO(yunqingwang): If d16[j].var is very large, use 8x8 partition
+        // size to further improve quality.
+      }
+
+      is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
+          (d16[2]->var < thr) && (d16[3]->var < thr);
+
+      // Use 32x32 partition
+      if (is_larger_better) {
+        use32x32 += 1;
+
+        for (j = 0; j < 4; j++) {
+          d32[i].sse += d16[j]->sse;
+          d32[i].sum += d16[j]->sum;
+        }
+
+        d32[i].var = d32[i].sse - (((int64_t)d32[i].sum * d32[i].sum) >> 10);
+
+        index = coord_lookup[i*4].row * mis + coord_lookup[i*4].col;
+        mi_8x8[index] = mi_upper_left + index;
+        mi_8x8[index]->sb_type = BLOCK_32X32;
+      }
+    }
+
+    if (use32x32 == 4) {
+      thr <<= 1;
+      is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
+          (d32[2].var < thr) && (d32[3].var < thr);
+
+      // Use 64x64 partition
+      if (is_larger_better) {
+        mi_8x8[0] = mi_upper_left;
+        mi_8x8[0]->sb_type = BLOCK_64X64;
+      }
+    }
+  } else {   // partial in-image SB64
+    int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
+    int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
+    set_partial_b64x64_partition(mi_upper_left, mis, bh, bw,
+        row8x8_remaining, col8x8_remaining, BLOCK_16X16, mi_8x8);
+  }
+}
+
+static void update_state_rt(VP9_COMP *cpi, ThreadData *td,
+                            PICK_MODE_CONTEXT *ctx,
+                            int mi_row, int mi_col, int bsize) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  struct macroblock_plane *const p = x->plane;
+  const struct segmentation *const seg = &cm->seg;
+  const int bw = num_8x8_blocks_wide_lookup[mi->sb_type];
+  const int bh = num_8x8_blocks_high_lookup[mi->sb_type];
+  const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
+
+  *(xd->mi[0]) = ctx->mic;
+  *(x->mbmi_ext) = ctx->mbmi_ext;
+
+  if (seg->enabled && cpi->oxcf.aq_mode != NO_AQ) {
+    // For in frame complexity AQ or variance AQ, copy segment_id from
+    // segmentation_map.
+    if (cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ) {
+      const uint8_t *const map = seg->update_map ? cpi->segmentation_map
+                                                 : cm->last_frame_seg_map;
+      mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
+    } else {
+    // Setting segmentation map for cyclic_refresh.
+      vp9_cyclic_refresh_update_segment(cpi, mi, mi_row, mi_col, bsize,
+                                        ctx->rate, ctx->dist, x->skip, p);
+    }
+    vp9_init_plane_quantizers(cpi, x);
+  }
+
+  if (is_inter_block(mi)) {
+    vp9_update_mv_count(td);
+    if (cm->interp_filter == SWITCHABLE) {
+      const int pred_ctx = vp9_get_pred_context_switchable_interp(xd);
+      ++td->counts->switchable_interp[pred_ctx][mi->interp_filter];
+    }
+
+    if (mi->sb_type < BLOCK_8X8) {
+      mi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
+      mi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
+    }
+  }
+
+  if (cm->use_prev_frame_mvs ||
+      (cpi->svc.use_base_mv && cpi->svc.number_spatial_layers > 1
+        && cpi->svc.spatial_layer_id != cpi->svc.number_spatial_layers - 1)) {
+    MV_REF *const frame_mvs =
+        cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
+    int w, h;
+
+    for (h = 0; h < y_mis; ++h) {
+      MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
+      for (w = 0; w < x_mis; ++w) {
+        MV_REF *const mv = frame_mv + w;
+        mv->ref_frame[0] = mi->ref_frame[0];
+        mv->ref_frame[1] = mi->ref_frame[1];
+        mv->mv[0].as_int = mi->mv[0].as_int;
+        mv->mv[1].as_int = mi->mv[1].as_int;
+      }
+    }
+  }
+
+  x->skip = ctx->skip;
+  x->skip_txfm[0] = mi->segment_id ? 0 : ctx->skip_txfm[0];
+}
+
+static void encode_b_rt(VP9_COMP *cpi, ThreadData *td,
+                        const TileInfo *const tile,
+                        TOKENEXTRA **tp, int mi_row, int mi_col,
+                        int output_enabled, BLOCK_SIZE bsize,
+                        PICK_MODE_CONTEXT *ctx) {
+  MACROBLOCK *const x = &td->mb;
+  set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
+  update_state_rt(cpi, td, ctx, mi_row, mi_col, bsize);
+
+  encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
+  update_stats(&cpi->common, td);
+
+  (*tp)->token = EOSB_TOKEN;
+  (*tp)++;
+}
+
+static void encode_sb_rt(VP9_COMP *cpi, ThreadData *td,
+                         const TileInfo *const tile,
+                         TOKENEXTRA **tp, int mi_row, int mi_col,
+                         int output_enabled, BLOCK_SIZE bsize,
+                         PC_TREE *pc_tree) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+
+  const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
+  int ctx;
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  if (bsize >= BLOCK_8X8) {
+    const int idx_str = xd->mi_stride * mi_row + mi_col;
+    MODE_INFO ** mi_8x8 = cm->mi_grid_visible + idx_str;
+    ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
+    subsize = mi_8x8[0]->sb_type;
+  } else {
+    ctx = 0;
+    subsize = BLOCK_4X4;
+  }
+
+  partition = partition_lookup[bsl][subsize];
+  if (output_enabled && bsize != BLOCK_4X4)
+    td->counts->partition[ctx][partition]++;
+
+  switch (partition) {
+    case PARTITION_NONE:
+      encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                  &pc_tree->none);
+      break;
+    case PARTITION_VERT:
+      encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                  &pc_tree->vertical[0]);
+      if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
+        encode_b_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
+                    subsize, &pc_tree->vertical[1]);
+      }
+      break;
+    case PARTITION_HORZ:
+      encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                  &pc_tree->horizontal[0]);
+      if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
+        encode_b_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
+                    subsize, &pc_tree->horizontal[1]);
+      }
+      break;
+    case PARTITION_SPLIT:
+      subsize = get_subsize(bsize, PARTITION_SPLIT);
+      encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
+                   pc_tree->split[0]);
+      encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
+                   subsize, pc_tree->split[1]);
+      encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
+                   subsize, pc_tree->split[2]);
+      encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs,
+                   output_enabled, subsize, pc_tree->split[3]);
+      break;
+    default:
+      assert(0 && "Invalid partition type.");
+      break;
+  }
+
+  if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
+    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+}
+
+static void rd_use_partition(VP9_COMP *cpi,
+                             ThreadData *td,
+                             TileDataEnc *tile_data,
+                             MODE_INFO **mi_8x8, TOKENEXTRA **tp,
+                             int mi_row, int mi_col,
+                             BLOCK_SIZE bsize,
+                             int *rate, int64_t *dist,
+                             int do_recon, PC_TREE *pc_tree) {
+  VP9_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int mis = cm->mi_stride;
+  const int bsl = b_width_log2_lookup[bsize];
+  const int mi_step = num_4x4_blocks_wide_lookup[bsize] / 2;
+  const int bss = (1 << bsl) / 4;
+  int i, pl;
+  PARTITION_TYPE partition = PARTITION_NONE;
+  BLOCK_SIZE subsize;
+  ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
+  PARTITION_CONTEXT sl[8], sa[8];
+  RD_COST last_part_rdc, none_rdc, chosen_rdc;
+  BLOCK_SIZE sub_subsize = BLOCK_4X4;
+  int splits_below = 0;
+  BLOCK_SIZE bs_type = mi_8x8[0]->sb_type;
+  int do_partition_search = 1;
+  PICK_MODE_CONTEXT *ctx = &pc_tree->none;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  assert(num_4x4_blocks_wide_lookup[bsize] ==
+         num_4x4_blocks_high_lookup[bsize]);
+
+  vp9_rd_cost_reset(&last_part_rdc);
+  vp9_rd_cost_reset(&none_rdc);
+  vp9_rd_cost_reset(&chosen_rdc);
+
+  partition = partition_lookup[bsl][bs_type];
+  subsize = get_subsize(bsize, partition);
+
+  pc_tree->partitioning = partition;
+  save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+
+  if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode != NO_AQ) {
+    set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+    x->mb_energy = vp9_block_energy(cpi, x, bsize);
+  }
+
+  if (do_partition_search &&
+      cpi->sf.partition_search_type == SEARCH_PARTITION &&
+      cpi->sf.adjust_partitioning_from_last_frame) {
+    // Check if any of the sub blocks are further split.
+    if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
+      sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
+      splits_below = 1;
+      for (i = 0; i < 4; i++) {
+        int jj = i >> 1, ii = i & 0x01;
+        MODE_INFO *this_mi = mi_8x8[jj * bss * mis + ii * bss];
+        if (this_mi && this_mi->sb_type >= sub_subsize) {
+          splits_below = 0;
+        }
+      }
+    }
+
+    // If partition is not none try none unless each of the 4 splits are split
+    // even further..
+    if (partition != PARTITION_NONE && !splits_below &&
+        mi_row + (mi_step >> 1) < cm->mi_rows &&
+        mi_col + (mi_step >> 1) < cm->mi_cols) {
+      pc_tree->partitioning = PARTITION_NONE;
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, bsize,
+                       ctx, INT64_MAX);
+
+      pl = partition_plane_context(xd, mi_row, mi_col, bsize);
+
+      if (none_rdc.rate < INT_MAX) {
+        none_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
+        none_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, none_rdc.rate,
+                                 none_rdc.dist);
+      }
+
+      restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+      mi_8x8[0]->sb_type = bs_type;
+      pc_tree->partitioning = partition;
+    }
+  }
+
+  switch (partition) {
+    case PARTITION_NONE:
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
+                       bsize, ctx, INT64_MAX);
+      break;
+    case PARTITION_HORZ:
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
+                       subsize, &pc_tree->horizontal[0],
+                       INT64_MAX);
+      if (last_part_rdc.rate != INT_MAX &&
+          bsize >= BLOCK_8X8 && mi_row + (mi_step >> 1) < cm->mi_rows) {
+        RD_COST tmp_rdc;
+        PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
+        vp9_rd_cost_init(&tmp_rdc);
+        update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
+        encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
+        rd_pick_sb_modes(cpi, tile_data, x,
+                         mi_row + (mi_step >> 1), mi_col, &tmp_rdc,
+                         subsize, &pc_tree->horizontal[1], INT64_MAX);
+        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
+          vp9_rd_cost_reset(&last_part_rdc);
+          break;
+        }
+        last_part_rdc.rate += tmp_rdc.rate;
+        last_part_rdc.dist += tmp_rdc.dist;
+        last_part_rdc.rdcost += tmp_rdc.rdcost;
+      }
+      break;
+    case PARTITION_VERT:
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
+                       subsize, &pc_tree->vertical[0], INT64_MAX);
+      if (last_part_rdc.rate != INT_MAX &&
+          bsize >= BLOCK_8X8 && mi_col + (mi_step >> 1) < cm->mi_cols) {
+        RD_COST tmp_rdc;
+        PICK_MODE_CONTEXT *ctx = &pc_tree->vertical[0];
+        vp9_rd_cost_init(&tmp_rdc);
+        update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
+        encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
+        rd_pick_sb_modes(cpi, tile_data, x,
+                         mi_row, mi_col + (mi_step >> 1), &tmp_rdc,
+                         subsize, &pc_tree->vertical[bsize > BLOCK_8X8],
+                         INT64_MAX);
+        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
+          vp9_rd_cost_reset(&last_part_rdc);
+          break;
+        }
+        last_part_rdc.rate += tmp_rdc.rate;
+        last_part_rdc.dist += tmp_rdc.dist;
+        last_part_rdc.rdcost += tmp_rdc.rdcost;
+      }
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8) {
+        rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
+                         subsize, pc_tree->leaf_split[0], INT64_MAX);
+        break;
+      }
+      last_part_rdc.rate = 0;
+      last_part_rdc.dist = 0;
+      last_part_rdc.rdcost = 0;
+      for (i = 0; i < 4; i++) {
+        int x_idx = (i & 1) * (mi_step >> 1);
+        int y_idx = (i >> 1) * (mi_step >> 1);
+        int jj = i >> 1, ii = i & 0x01;
+        RD_COST tmp_rdc;
+        if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
+          continue;
+
+        vp9_rd_cost_init(&tmp_rdc);
+        rd_use_partition(cpi, td, tile_data,
+                         mi_8x8 + jj * bss * mis + ii * bss, tp,
+                         mi_row + y_idx, mi_col + x_idx, subsize,
+                         &tmp_rdc.rate, &tmp_rdc.dist,
+                         i != 3, pc_tree->split[i]);
+        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
+          vp9_rd_cost_reset(&last_part_rdc);
+          break;
+        }
+        last_part_rdc.rate += tmp_rdc.rate;
+        last_part_rdc.dist += tmp_rdc.dist;
+      }
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  pl = partition_plane_context(xd, mi_row, mi_col, bsize);
+  if (last_part_rdc.rate < INT_MAX) {
+    last_part_rdc.rate += cpi->partition_cost[pl][partition];
+    last_part_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                                  last_part_rdc.rate, last_part_rdc.dist);
+  }
+
+  if (do_partition_search
+      && cpi->sf.adjust_partitioning_from_last_frame
+      && cpi->sf.partition_search_type == SEARCH_PARTITION
+      && partition != PARTITION_SPLIT && bsize > BLOCK_8X8
+      && (mi_row + mi_step < cm->mi_rows ||
+          mi_row + (mi_step >> 1) == cm->mi_rows)
+      && (mi_col + mi_step < cm->mi_cols ||
+          mi_col + (mi_step >> 1) == cm->mi_cols)) {
+    BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
+    chosen_rdc.rate = 0;
+    chosen_rdc.dist = 0;
+    restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+    pc_tree->partitioning = PARTITION_SPLIT;
+
+    // Split partition.
+    for (i = 0; i < 4; i++) {
+      int x_idx = (i & 1) * (mi_step >> 1);
+      int y_idx = (i >> 1) * (mi_step >> 1);
+      RD_COST tmp_rdc;
+      ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
+      PARTITION_CONTEXT sl[8], sa[8];
+
+      if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
+        continue;
+
+      save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+      pc_tree->split[i]->partitioning = PARTITION_NONE;
+      rd_pick_sb_modes(cpi, tile_data, x,
+                       mi_row + y_idx, mi_col + x_idx, &tmp_rdc,
+                       split_subsize, &pc_tree->split[i]->none, INT64_MAX);
+
+      restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+
+      if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
+        vp9_rd_cost_reset(&chosen_rdc);
+        break;
+      }
+
+      chosen_rdc.rate += tmp_rdc.rate;
+      chosen_rdc.dist += tmp_rdc.dist;
+
+      if (i != 3)
+        encode_sb(cpi, td, tile_info, tp,  mi_row + y_idx, mi_col + x_idx, 0,
+                  split_subsize, pc_tree->split[i]);
+
+      pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
+                                   split_subsize);
+      chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
+    }
+    pl = partition_plane_context(xd, mi_row, mi_col, bsize);
+    if (chosen_rdc.rate < INT_MAX) {
+      chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
+      chosen_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                                 chosen_rdc.rate, chosen_rdc.dist);
+    }
+  }
+
+  // If last_part is better set the partitioning to that.
+  if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
+    mi_8x8[0]->sb_type = bsize;
+    if (bsize >= BLOCK_8X8)
+      pc_tree->partitioning = partition;
+    chosen_rdc = last_part_rdc;
+  }
+  // If none was better set the partitioning to that.
+  if (none_rdc.rdcost < chosen_rdc.rdcost) {
+    if (bsize >= BLOCK_8X8)
+      pc_tree->partitioning = PARTITION_NONE;
+    chosen_rdc = none_rdc;
+  }
+
+  restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+
+  // We must have chosen a partitioning and encoding or we'll fail later on.
+  // No other opportunities for success.
+  if (bsize == BLOCK_64X64)
+    assert(chosen_rdc.rate < INT_MAX && chosen_rdc.dist < INT64_MAX);
+
+  if (do_recon) {
+    int output_enabled = (bsize == BLOCK_64X64);
+    encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
+              pc_tree);
+  }
+
+  *rate = chosen_rdc.rate;
+  *dist = chosen_rdc.dist;
+}
+
+static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
+  BLOCK_4X4,   BLOCK_4X4,   BLOCK_4X4,
+  BLOCK_4X4,   BLOCK_4X4,   BLOCK_4X4,
+  BLOCK_8X8,   BLOCK_8X8,   BLOCK_8X8,
+  BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
+  BLOCK_16X16
+};
+
+static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
+  BLOCK_8X8,   BLOCK_16X16, BLOCK_16X16,
+  BLOCK_16X16, BLOCK_32X32, BLOCK_32X32,
+  BLOCK_32X32, BLOCK_64X64, BLOCK_64X64,
+  BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
+  BLOCK_64X64
+};
+
+
+// Look at all the mode_info entries for blocks that are part of this
+// partition and find the min and max values for sb_type.
+// At the moment this is designed to work on a 64x64 SB but could be
+// adjusted to use a size parameter.
+//
+// The min and max are assumed to have been initialized prior to calling this
+// function so repeat calls can accumulate a min and max of more than one sb64.
+static void get_sb_partition_size_range(MACROBLOCKD *xd, MODE_INFO **mi_8x8,
+                                        BLOCK_SIZE *min_block_size,
+                                        BLOCK_SIZE *max_block_size,
+                                        int bs_hist[BLOCK_SIZES]) {
+  int sb_width_in_blocks = MI_BLOCK_SIZE;
+  int sb_height_in_blocks  = MI_BLOCK_SIZE;
+  int i, j;
+  int index = 0;
+
+  // Check the sb_type for each block that belongs to this region.
+  for (i = 0; i < sb_height_in_blocks; ++i) {
+    for (j = 0; j < sb_width_in_blocks; ++j) {
+      MODE_INFO *mi = mi_8x8[index+j];
+      BLOCK_SIZE sb_type = mi ? mi->sb_type : 0;
+      bs_hist[sb_type]++;
+      *min_block_size = VPXMIN(*min_block_size, sb_type);
+      *max_block_size = VPXMAX(*max_block_size, sb_type);
+    }
+    index += xd->mi_stride;
+  }
+}
+
+// Next square block size less or equal than current block size.
+static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
+  BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
+  BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
+  BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
+  BLOCK_32X32, BLOCK_32X32, BLOCK_32X32,
+  BLOCK_64X64
+};
+
+// Look at neighboring blocks and set a min and max partition size based on
+// what they chose.
+static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
+                                    MACROBLOCKD *const xd,
+                                    int mi_row, int mi_col,
+                                    BLOCK_SIZE *min_block_size,
+                                    BLOCK_SIZE *max_block_size) {
+  VP9_COMMON *const cm = &cpi->common;
+  MODE_INFO **mi = xd->mi;
+  const int left_in_image = !!xd->left_mi;
+  const int above_in_image = !!xd->above_mi;
+  const int row8x8_remaining = tile->mi_row_end - mi_row;
+  const int col8x8_remaining = tile->mi_col_end - mi_col;
+  int bh, bw;
+  BLOCK_SIZE min_size = BLOCK_4X4;
+  BLOCK_SIZE max_size = BLOCK_64X64;
+  int bs_hist[BLOCK_SIZES] = {0};
+
+  // Trap case where we do not have a prediction.
+  if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
+    // Default "min to max" and "max to min"
+    min_size = BLOCK_64X64;
+    max_size = BLOCK_4X4;
+
+    // NOTE: each call to get_sb_partition_size_range() uses the previous
+    // passed in values for min and max as a starting point.
+    // Find the min and max partition used in previous frame at this location
+    if (cm->frame_type != KEY_FRAME) {
+      MODE_INFO **prev_mi =
+          &cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col];
+      get_sb_partition_size_range(xd, prev_mi, &min_size, &max_size, bs_hist);
+    }
+    // Find the min and max partition sizes used in the left SB64
+    if (left_in_image) {
+      MODE_INFO **left_sb64_mi = &mi[-MI_BLOCK_SIZE];
+      get_sb_partition_size_range(xd, left_sb64_mi, &min_size, &max_size,
+                                  bs_hist);
+    }
+    // Find the min and max partition sizes used in the above SB64.
+    if (above_in_image) {
+      MODE_INFO **above_sb64_mi = &mi[-xd->mi_stride * MI_BLOCK_SIZE];
+      get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size,
+                                  bs_hist);
+    }
+
+    // Adjust observed min and max for "relaxed" auto partition case.
+    if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
+      min_size = min_partition_size[min_size];
+      max_size = max_partition_size[max_size];
+    }
+  }
+
+  // Check border cases where max and min from neighbors may not be legal.
+  max_size = find_partition_size(max_size,
+                                 row8x8_remaining, col8x8_remaining,
+                                 &bh, &bw);
+  // Test for blocks at the edge of the active image.
+  // This may be the actual edge of the image or where there are formatting
+  // bars.
+  if (vp9_active_edge_sb(cpi, mi_row, mi_col)) {
+    min_size = BLOCK_4X4;
+  } else {
+    min_size =
+        VPXMIN(cpi->sf.rd_auto_partition_min_limit, VPXMIN(min_size, max_size));
+  }
+
+  // When use_square_partition_only is true, make sure at least one square
+  // partition is allowed by selecting the next smaller square size as
+  // *min_block_size.
+  if (cpi->sf.use_square_partition_only &&
+      next_square_size[max_size] < min_size) {
+     min_size = next_square_size[max_size];
+  }
+
+  *min_block_size = min_size;
+  *max_block_size = max_size;
+}
+
+// TODO(jingning) refactor functions setting partition search range
+static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                int mi_row, int mi_col, BLOCK_SIZE bsize,
+                                BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
+  int mi_width  = num_8x8_blocks_wide_lookup[bsize];
+  int mi_height = num_8x8_blocks_high_lookup[bsize];
+  int idx, idy;
+
+  MODE_INFO *mi;
+  const int idx_str = cm->mi_stride * mi_row + mi_col;
+  MODE_INFO **prev_mi = &cm->prev_mi_grid_visible[idx_str];
+  BLOCK_SIZE bs, min_size, max_size;
+
+  min_size = BLOCK_64X64;
+  max_size = BLOCK_4X4;
+
+  if (prev_mi) {
+    for (idy = 0; idy < mi_height; ++idy) {
+      for (idx = 0; idx < mi_width; ++idx) {
+        mi = prev_mi[idy * cm->mi_stride + idx];
+        bs = mi ? mi->sb_type : bsize;
+        min_size = VPXMIN(min_size, bs);
+        max_size = VPXMAX(max_size, bs);
+      }
+    }
+  }
+
+  if (xd->left_mi) {
+    for (idy = 0; idy < mi_height; ++idy) {
+      mi = xd->mi[idy * cm->mi_stride - 1];
+      bs = mi ? mi->sb_type : bsize;
+      min_size = VPXMIN(min_size, bs);
+      max_size = VPXMAX(max_size, bs);
+    }
+  }
+
+  if (xd->above_mi) {
+    for (idx = 0; idx < mi_width; ++idx) {
+      mi = xd->mi[idx - cm->mi_stride];
+      bs = mi ? mi->sb_type : bsize;
+      min_size = VPXMIN(min_size, bs);
+      max_size = VPXMAX(max_size, bs);
+    }
+  }
+
+  if (min_size == max_size) {
+    min_size = min_partition_size[min_size];
+    max_size = max_partition_size[max_size];
+  }
+
+  *min_bs = min_size;
+  *max_bs = max_size;
+}
+
+static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
+  memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
+}
+
+static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
+  memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
+}
+
+#if CONFIG_FP_MB_STATS
+const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] =
+  {1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4};
+const int num_16x16_blocks_high_lookup[BLOCK_SIZES] =
+  {1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4};
+const int qindex_skip_threshold_lookup[BLOCK_SIZES] =
+  {0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120};
+const int qindex_split_threshold_lookup[BLOCK_SIZES] =
+  {0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120};
+const int complexity_16x16_blocks_threshold[BLOCK_SIZES] =
+  {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6};
+
+typedef enum {
+  MV_ZERO = 0,
+  MV_LEFT = 1,
+  MV_UP = 2,
+  MV_RIGHT = 3,
+  MV_DOWN = 4,
+  MV_INVALID
+} MOTION_DIRECTION;
+
+static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
+  if (fp_byte & FPMB_MOTION_ZERO_MASK) {
+    return MV_ZERO;
+  } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
+    return MV_LEFT;
+  } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
+    return MV_RIGHT;
+  } else if (fp_byte & FPMB_MOTION_UP_MASK) {
+    return MV_UP;
+  } else {
+    return MV_DOWN;
+  }
+}
+
+static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
+                                           MOTION_DIRECTION that_mv) {
+  if (this_mv == that_mv) {
+    return 0;
+  } else {
+    return abs(this_mv - that_mv) == 2 ? 2 : 1;
+  }
+}
+#endif
+
+// TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
+// unlikely to be selected depending on previous rate-distortion optimization
+// results, for encoding speed-up.
+static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
+                              TileDataEnc *tile_data,
+                              TOKENEXTRA **tp, int mi_row, int mi_col,
+                              BLOCK_SIZE bsize, RD_COST *rd_cost,
+                              int64_t best_rd, PC_TREE *pc_tree) {
+  VP9_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
+  ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
+  PARTITION_CONTEXT sl[8], sa[8];
+  TOKENEXTRA *tp_orig = *tp;
+  PICK_MODE_CONTEXT *ctx = &pc_tree->none;
+  int i;
+  const int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
+  BLOCK_SIZE subsize;
+  RD_COST this_rdc, sum_rdc, best_rdc;
+  int do_split = bsize >= BLOCK_8X8;
+  int do_rect = 1;
+
+  // Override skipping rectangular partition operations for edge blocks
+  const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
+  const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
+  const int xss = x->e_mbd.plane[1].subsampling_x;
+  const int yss = x->e_mbd.plane[1].subsampling_y;
+
+  BLOCK_SIZE min_size = x->min_partition_size;
+  BLOCK_SIZE max_size = x->max_partition_size;
+
+#if CONFIG_FP_MB_STATS
+  unsigned int src_diff_var = UINT_MAX;
+  int none_complexity = 0;
+#endif
+
+  int partition_none_allowed = !force_horz_split && !force_vert_split;
+  int partition_horz_allowed = !force_vert_split && yss <= xss &&
+                               bsize >= BLOCK_8X8;
+  int partition_vert_allowed = !force_horz_split && xss <= yss &&
+                               bsize >= BLOCK_8X8;
+
+  int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_dist_thr;
+  int rate_breakout_thr = cpi->sf.partition_search_breakout_rate_thr;
+
+  (void)*tp_orig;
+
+  assert(num_8x8_blocks_wide_lookup[bsize] ==
+             num_8x8_blocks_high_lookup[bsize]);
+
+  // Adjust dist breakout threshold according to the partition size.
+  dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
+      b_height_log2_lookup[bsize]);
+  rate_breakout_thr *= num_pels_log2_lookup[bsize];
+
+  vp9_rd_cost_init(&this_rdc);
+  vp9_rd_cost_init(&sum_rdc);
+  vp9_rd_cost_reset(&best_rdc);
+  best_rdc.rdcost = best_rd;
+
+  set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+
+  if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode != NO_AQ)
+    x->mb_energy = vp9_block_energy(cpi, x, bsize);
+
+  if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
+    int cb_partition_search_ctrl = ((pc_tree->index == 0 || pc_tree->index == 3)
+        + get_chessboard_index(cm->current_video_frame)) & 0x1;
+
+    if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
+      set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
+  }
+
+  // Determine partition types in search according to the speed features.
+  // The threshold set here has to be of square block size.
+  if (cpi->sf.auto_min_max_partition_size) {
+    partition_none_allowed &= (bsize <= max_size && bsize >= min_size);
+    partition_horz_allowed &= ((bsize <= max_size && bsize > min_size) ||
+                                force_horz_split);
+    partition_vert_allowed &= ((bsize <= max_size && bsize > min_size) ||
+                                force_vert_split);
+    do_split &= bsize > min_size;
+  }
+
+  if (cpi->sf.use_square_partition_only &&
+      bsize > cpi->sf.use_square_only_threshold) {
+    if (cpi->use_svc) {
+      if (!vp9_active_h_edge(cpi, mi_row, mi_step) || x->e_mbd.lossless)
+        partition_horz_allowed &= force_horz_split;
+      if (!vp9_active_v_edge(cpi, mi_row, mi_step) || x->e_mbd.lossless)
+        partition_vert_allowed &= force_vert_split;
+    } else {
+      partition_horz_allowed &= force_horz_split;
+      partition_vert_allowed &= force_vert_split;
+    }
+  }
+
+  save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+
+#if CONFIG_FP_MB_STATS
+  if (cpi->use_fp_mb_stats) {
+    set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+    src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
+                                                  mi_row, mi_col, bsize);
+  }
+#endif
+
+#if CONFIG_FP_MB_STATS
+  // Decide whether we shall split directly and skip searching NONE by using
+  // the first pass block statistics
+  if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
+      partition_none_allowed && src_diff_var > 4 &&
+      cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
+    int mb_row = mi_row >> 1;
+    int mb_col = mi_col >> 1;
+    int mb_row_end =
+        VPXMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
+    int mb_col_end =
+        VPXMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
+    int r, c;
+
+    // compute a complexity measure, basically measure inconsistency of motion
+    // vectors obtained from the first pass in the current block
+    for (r = mb_row; r < mb_row_end ; r++) {
+      for (c = mb_col; c < mb_col_end; c++) {
+        const int mb_index = r * cm->mb_cols + c;
+
+        MOTION_DIRECTION this_mv;
+        MOTION_DIRECTION right_mv;
+        MOTION_DIRECTION bottom_mv;
+
+        this_mv =
+            get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
+
+        // to its right
+        if (c != mb_col_end - 1) {
+          right_mv = get_motion_direction_fp(
+              cpi->twopass.this_frame_mb_stats[mb_index + 1]);
+          none_complexity += get_motion_inconsistency(this_mv, right_mv);
+        }
+
+        // to its bottom
+        if (r != mb_row_end - 1) {
+          bottom_mv = get_motion_direction_fp(
+              cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
+          none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
+        }
+
+        // do not count its left and top neighbors to avoid double counting
+      }
+    }
+
+    if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
+      partition_none_allowed = 0;
+    }
+  }
+#endif
+
+  // PARTITION_NONE
+  if (partition_none_allowed) {
+    rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
+                     &this_rdc, bsize, ctx, best_rdc.rdcost);
+    if (this_rdc.rate != INT_MAX) {
+      if (bsize >= BLOCK_8X8) {
+        this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
+        this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                                 this_rdc.rate, this_rdc.dist);
+      }
+
+      if (this_rdc.rdcost < best_rdc.rdcost) {
+        best_rdc = this_rdc;
+        if (bsize >= BLOCK_8X8)
+          pc_tree->partitioning = PARTITION_NONE;
+
+        // If all y, u, v transform blocks in this partition are skippable, and
+        // the dist & rate are within the thresholds, the partition search is
+        // terminated for current branch of the partition search tree.
+        if (!x->e_mbd.lossless && ctx->skippable  &&
+            ((best_rdc.dist < (dist_breakout_thr >> 2)) ||
+             (best_rdc.dist < dist_breakout_thr &&
+              best_rdc.rate < rate_breakout_thr))) {
+          do_split = 0;
+          do_rect = 0;
+        }
+
+#if CONFIG_FP_MB_STATS
+        // Check if every 16x16 first pass block statistics has zero
+        // motion and the corresponding first pass residue is small enough.
+        // If that is the case, check the difference variance between the
+        // current frame and the last frame. If the variance is small enough,
+        // stop further splitting in RD optimization
+        if (cpi->use_fp_mb_stats && do_split != 0 &&
+            cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
+          int mb_row = mi_row >> 1;
+          int mb_col = mi_col >> 1;
+          int mb_row_end =
+              VPXMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
+          int mb_col_end =
+              VPXMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
+          int r, c;
+
+          int skip = 1;
+          for (r = mb_row; r < mb_row_end; r++) {
+            for (c = mb_col; c < mb_col_end; c++) {
+              const int mb_index = r * cm->mb_cols + c;
+              if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
+                    FPMB_MOTION_ZERO_MASK) ||
+                  !(cpi->twopass.this_frame_mb_stats[mb_index] &
+                    FPMB_ERROR_SMALL_MASK)) {
+                skip = 0;
+                break;
+              }
+            }
+            if (skip == 0) {
+              break;
+            }
+          }
+
+          if (skip) {
+            if (src_diff_var == UINT_MAX) {
+              set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+              src_diff_var = get_sby_perpixel_diff_variance(
+                  cpi, &x->plane[0].src, mi_row, mi_col, bsize);
+            }
+            if (src_diff_var < 8) {
+              do_split = 0;
+              do_rect = 0;
+            }
+          }
+        }
+#endif
+      }
+    }
+    restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+  }
+
+  // store estimated motion vector
+  if (cpi->sf.adaptive_motion_search)
+    store_pred_mv(x, ctx);
+
+  // PARTITION_SPLIT
+  // TODO(jingning): use the motion vectors given by the above search as
+  // the starting point of motion search in the following partition type check.
+  if (do_split) {
+    subsize = get_subsize(bsize, PARTITION_SPLIT);
+    if (bsize == BLOCK_8X8) {
+      i = 4;
+      if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
+        pc_tree->leaf_split[0]->pred_interp_filter =
+            ctx->mic.interp_filter;
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
+                       pc_tree->leaf_split[0], best_rdc.rdcost);
+
+      if (sum_rdc.rate == INT_MAX)
+        sum_rdc.rdcost = INT64_MAX;
+    } else {
+      for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
+        const int x_idx =  (i & 1) * mi_step;
+        const int y_idx = (i >> 1) * mi_step;
+
+        if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
+          continue;
+
+        if (cpi->sf.adaptive_motion_search)
+          load_pred_mv(x, ctx);
+
+        pc_tree->split[i]->index = i;
+        rd_pick_partition(cpi, td, tile_data, tp,
+                          mi_row + y_idx, mi_col + x_idx,
+                          subsize, &this_rdc,
+                          best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
+
+        if (this_rdc.rate == INT_MAX) {
+          sum_rdc.rdcost = INT64_MAX;
+          break;
+        } else {
+          sum_rdc.rate += this_rdc.rate;
+          sum_rdc.dist += this_rdc.dist;
+          sum_rdc.rdcost += this_rdc.rdcost;
+        }
+      }
+    }
+
+    if (sum_rdc.rdcost < best_rdc.rdcost && i == 4) {
+      sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
+      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                              sum_rdc.rate, sum_rdc.dist);
+
+      if (sum_rdc.rdcost < best_rdc.rdcost) {
+        best_rdc = sum_rdc;
+        pc_tree->partitioning = PARTITION_SPLIT;
+
+        // Rate and distortion based partition search termination clause.
+        if (!x->e_mbd.lossless &&
+            ((best_rdc.dist < (dist_breakout_thr >> 2)) ||
+             (best_rdc.dist < dist_breakout_thr &&
+              best_rdc.rate < rate_breakout_thr))) {
+          do_rect = 0;
+        }
+      }
+    } else {
+      // skip rectangular partition test when larger block size
+      // gives better rd cost
+      if ((cpi->sf.less_rectangular_check) &&
+          ((bsize > cpi->sf.use_square_only_threshold) ||
+           (best_rdc.dist < dist_breakout_thr)))
+        do_rect &= !partition_none_allowed;
+    }
+    restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+  }
+
+  // PARTITION_HORZ
+  if (partition_horz_allowed &&
+      (do_rect || vp9_active_h_edge(cpi, mi_row, mi_step))) {
+    subsize = get_subsize(bsize, PARTITION_HORZ);
+    if (cpi->sf.adaptive_motion_search)
+      load_pred_mv(x, ctx);
+    if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+        partition_none_allowed)
+      pc_tree->horizontal[0].pred_interp_filter =
+          ctx->mic.interp_filter;
+    rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
+                     &pc_tree->horizontal[0], best_rdc.rdcost);
+
+    if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + mi_step < cm->mi_rows &&
+        bsize > BLOCK_8X8) {
+      PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
+      update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
+      encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
+
+      if (cpi->sf.adaptive_motion_search)
+        load_pred_mv(x, ctx);
+      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+          partition_none_allowed)
+        pc_tree->horizontal[1].pred_interp_filter =
+            ctx->mic.interp_filter;
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col,
+                       &this_rdc, subsize, &pc_tree->horizontal[1],
+                       best_rdc.rdcost - sum_rdc.rdcost);
+      if (this_rdc.rate == INT_MAX) {
+        sum_rdc.rdcost = INT64_MAX;
+      } else {
+        sum_rdc.rate += this_rdc.rate;
+        sum_rdc.dist += this_rdc.dist;
+        sum_rdc.rdcost += this_rdc.rdcost;
+      }
+    }
+
+    if (sum_rdc.rdcost < best_rdc.rdcost) {
+      sum_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
+      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+      if (sum_rdc.rdcost < best_rdc.rdcost) {
+        best_rdc = sum_rdc;
+        pc_tree->partitioning = PARTITION_HORZ;
+
+        if ((cpi->sf.less_rectangular_check) &&
+            (bsize > cpi->sf.use_square_only_threshold))
+          do_rect = 0;
+      }
+    }
+    restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+  }
+
+  // PARTITION_VERT
+  if (partition_vert_allowed &&
+      (do_rect || vp9_active_v_edge(cpi, mi_col, mi_step))) {
+    subsize = get_subsize(bsize, PARTITION_VERT);
+
+    if (cpi->sf.adaptive_motion_search)
+      load_pred_mv(x, ctx);
+    if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+        partition_none_allowed)
+      pc_tree->vertical[0].pred_interp_filter =
+          ctx->mic.interp_filter;
+    rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
+                     &pc_tree->vertical[0], best_rdc.rdcost);
+    if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + mi_step < cm->mi_cols &&
+        bsize > BLOCK_8X8) {
+      update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
+      encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize,
+                        &pc_tree->vertical[0]);
+
+      if (cpi->sf.adaptive_motion_search)
+        load_pred_mv(x, ctx);
+      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+          partition_none_allowed)
+        pc_tree->vertical[1].pred_interp_filter =
+            ctx->mic.interp_filter;
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step,
+                       &this_rdc, subsize,
+                       &pc_tree->vertical[1], best_rdc.rdcost - sum_rdc.rdcost);
+      if (this_rdc.rate == INT_MAX) {
+        sum_rdc.rdcost = INT64_MAX;
+      } else {
+        sum_rdc.rate += this_rdc.rate;
+        sum_rdc.dist += this_rdc.dist;
+        sum_rdc.rdcost += this_rdc.rdcost;
+      }
+    }
+
+    if (sum_rdc.rdcost < best_rdc.rdcost) {
+      sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
+      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                              sum_rdc.rate, sum_rdc.dist);
+      if (sum_rdc.rdcost < best_rdc.rdcost) {
+        best_rdc = sum_rdc;
+        pc_tree->partitioning = PARTITION_VERT;
+      }
+    }
+    restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
+  }
+
+  // TODO(jbb): This code added so that we avoid static analysis
+  // warning related to the fact that best_rd isn't used after this
+  // point.  This code should be refactored so that the duplicate
+  // checks occur in some sub function and thus are used...
+  (void) best_rd;
+  *rd_cost = best_rdc;
+
+  if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
+      pc_tree->index != 3) {
+    int output_enabled = (bsize == BLOCK_64X64);
+    encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
+              bsize, pc_tree);
+  }
+
+  if (bsize == BLOCK_64X64) {
+    assert(tp_orig < *tp);
+    assert(best_rdc.rate < INT_MAX);
+    assert(best_rdc.dist < INT64_MAX);
+  } else {
+    assert(tp_orig == *tp);
+  }
+}
+
+static void encode_rd_sb_row(VP9_COMP *cpi,
+                             ThreadData *td,
+                             TileDataEnc *tile_data,
+                             int mi_row,
+                             TOKENEXTRA **tp) {
+  VP9_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  SPEED_FEATURES *const sf = &cpi->sf;
+  const int mi_col_start = tile_info->mi_col_start;
+  const int mi_col_end = tile_info->mi_col_end;
+  int mi_col;
+
+  // Initialize the left context for the new SB row
+  memset(&xd->left_context, 0, sizeof(xd->left_context));
+  memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
+
+  // Code each SB in the row
+  for (mi_col = mi_col_start; mi_col < mi_col_end; mi_col += MI_BLOCK_SIZE) {
+    const struct segmentation *const seg = &cm->seg;
+    int dummy_rate;
+    int64_t dummy_dist;
+    RD_COST dummy_rdc;
+    int i;
+    int seg_skip = 0;
+
+    const int idx_str = cm->mi_stride * mi_row + mi_col;
+    MODE_INFO **mi = cm->mi_grid_visible + idx_str;
+
+    if (sf->adaptive_pred_interp_filter) {
+      for (i = 0; i < 64; ++i)
+        td->leaf_tree[i].pred_interp_filter = SWITCHABLE;
+
+      for (i = 0; i < 64; ++i) {
+        td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
+        td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
+        td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
+        td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
+      }
+    }
+
+    vp9_zero(x->pred_mv);
+    td->pc_root->index = 0;
+
+    if (seg->enabled) {
+      const uint8_t *const map = seg->update_map ? cpi->segmentation_map
+                                                 : cm->last_frame_seg_map;
+      int segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
+      seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
+    }
+
+    x->source_variance = UINT_MAX;
+    if (sf->partition_search_type == FIXED_PARTITION || seg_skip) {
+      const BLOCK_SIZE bsize =
+          seg_skip ? BLOCK_64X64 : sf->always_this_block_size;
+      set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
+      set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
+      rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
+                       BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
+    } else if (cpi->partition_search_skippable_frame) {
+      BLOCK_SIZE bsize;
+      set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
+      bsize = get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col);
+      set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
+      rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
+                       BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
+    } else if (sf->partition_search_type == VAR_BASED_PARTITION &&
+               cm->frame_type != KEY_FRAME) {
+      choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
+      rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
+                       BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
+    } else {
+      // If required set upper and lower partition size limits
+      if (sf->auto_min_max_partition_size) {
+        set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
+        rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col,
+                                &x->min_partition_size,
+                                &x->max_partition_size);
+      }
+      rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, BLOCK_64X64,
+                        &dummy_rdc, INT64_MAX, td->pc_root);
+    }
+  }
+}
+
+static void init_encode_frame_mb_context(VP9_COMP *cpi) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
+
+  // Copy data over into macro block data structures.
+  vp9_setup_src_planes(x, cpi->Source, 0, 0);
+
+  vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
+
+  // Note: this memset assumes above_context[0], [1] and [2]
+  // are allocated as part of the same buffer.
+  memset(xd->above_context[0], 0,
+         sizeof(*xd->above_context[0]) *
+         2 * aligned_mi_cols * MAX_MB_PLANE);
+  memset(xd->above_seg_context, 0,
+         sizeof(*xd->above_seg_context) * aligned_mi_cols);
+}
+
+static int check_dual_ref_flags(VP9_COMP *cpi) {
+  const int ref_flags = cpi->ref_frame_flags;
+
+  if (segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
+    return 0;
+  } else {
+    return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG)
+        + !!(ref_flags & VP9_ALT_FLAG)) >= 2;
+  }
+}
+
+static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
+  int mi_row, mi_col;
+  const int mis = cm->mi_stride;
+  MODE_INFO **mi_ptr = cm->mi_grid_visible;
+
+  for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
+    for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
+      if (mi_ptr[mi_col]->tx_size > max_tx_size)
+        mi_ptr[mi_col]->tx_size = max_tx_size;
+    }
+  }
+}
+
+static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
+  if (frame_is_intra_only(&cpi->common))
+    return INTRA_FRAME;
+  else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
+    return ALTREF_FRAME;
+  else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
+    return GOLDEN_FRAME;
+  else
+    return LAST_FRAME;
+}
+
+static TX_MODE select_tx_mode(const VP9_COMP *cpi, MACROBLOCKD *const xd) {
+  if (xd->lossless)
+    return ONLY_4X4;
+  if (cpi->common.frame_type == KEY_FRAME &&
+      cpi->sf.use_nonrd_pick_mode)
+    return ALLOW_16X16;
+  if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
+    return ALLOW_32X32;
+  else if (cpi->sf.tx_size_search_method == USE_FULL_RD||
+           cpi->sf.tx_size_search_method == USE_TX_8X8)
+    return TX_MODE_SELECT;
+  else
+    return cpi->common.tx_mode;
+}
+
+static void hybrid_intra_mode_search(VP9_COMP *cpi, MACROBLOCK *const x,
+                                     RD_COST *rd_cost, BLOCK_SIZE bsize,
+                                     PICK_MODE_CONTEXT *ctx) {
+  if (bsize < BLOCK_16X16)
+    vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
+  else
+    vp9_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
+}
+
+static void nonrd_pick_sb_modes(VP9_COMP *cpi,
+                                TileDataEnc *tile_data, MACROBLOCK *const x,
+                                int mi_row, int mi_col, RD_COST *rd_cost,
+                                BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
+  VP9_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *mi;
+  ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
+  BLOCK_SIZE bs = VPXMAX(bsize, BLOCK_8X8);  // processing unit block size
+  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bs];
+  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bs];
+  int plane;
+
+  set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+  mi = xd->mi[0];
+  mi->sb_type = bsize;
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    struct macroblockd_plane *pd = &xd->plane[plane];
+    memcpy(a + num_4x4_blocks_wide * plane, pd->above_context,
+           (sizeof(a[0]) * num_4x4_blocks_wide) >> pd->subsampling_x);
+    memcpy(l + num_4x4_blocks_high * plane, pd->left_context,
+           (sizeof(l[0]) * num_4x4_blocks_high) >> pd->subsampling_y);
+  }
+
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
+    if (cyclic_refresh_segment_id_boosted(mi->segment_id))
+      x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
+
+  if (cm->frame_type == KEY_FRAME)
+    hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx);
+  else if (segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP))
+    set_mode_info_seg_skip(x, cm->tx_mode, rd_cost, bsize);
+  else if (bsize >= BLOCK_8X8)
+    vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col,
+                        rd_cost, bsize, ctx);
+  else
+    vp9_pick_inter_mode_sub8x8(cpi, x, mi_row, mi_col,
+                               rd_cost, bsize, ctx);
+
+  duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    struct macroblockd_plane *pd = &xd->plane[plane];
+    memcpy(pd->above_context, a + num_4x4_blocks_wide * plane,
+           (sizeof(a[0]) * num_4x4_blocks_wide) >> pd->subsampling_x);
+    memcpy(pd->left_context, l + num_4x4_blocks_high * plane,
+           (sizeof(l[0]) * num_4x4_blocks_high) >> pd->subsampling_y);
+  }
+
+  if (rd_cost->rate == INT_MAX)
+    vp9_rd_cost_reset(rd_cost);
+
+  ctx->rate = rd_cost->rate;
+  ctx->dist = rd_cost->dist;
+}
+
+static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x,
+                              int mi_row, int mi_col,
+                              BLOCK_SIZE bsize,
+                              PC_TREE *pc_tree) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
+  PARTITION_TYPE partition = pc_tree->partitioning;
+  BLOCK_SIZE subsize = get_subsize(bsize, partition);
+
+  assert(bsize >= BLOCK_8X8);
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  switch (partition) {
+    case PARTITION_NONE:
+      set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
+      *(xd->mi[0]) = pc_tree->none.mic;
+      *(x->mbmi_ext) = pc_tree->none.mbmi_ext;
+      duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
+      break;
+    case PARTITION_VERT:
+      set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
+      *(xd->mi[0]) = pc_tree->vertical[0].mic;
+      *(x->mbmi_ext) = pc_tree->vertical[0].mbmi_ext;
+      duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
+
+      if (mi_col + hbs < cm->mi_cols) {
+        set_mode_info_offsets(cm, x, xd, mi_row, mi_col + hbs);
+        *(xd->mi[0]) = pc_tree->vertical[1].mic;
+        *(x->mbmi_ext) = pc_tree->vertical[1].mbmi_ext;
+        duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, subsize);
+      }
+      break;
+    case PARTITION_HORZ:
+      set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
+      *(xd->mi[0]) = pc_tree->horizontal[0].mic;
+      *(x->mbmi_ext) = pc_tree->horizontal[0].mbmi_ext;
+      duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
+      if (mi_row + hbs < cm->mi_rows) {
+        set_mode_info_offsets(cm, x, xd, mi_row + hbs, mi_col);
+        *(xd->mi[0]) = pc_tree->horizontal[1].mic;
+        *(x->mbmi_ext) = pc_tree->horizontal[1].mbmi_ext;
+        duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, subsize);
+      }
+      break;
+    case PARTITION_SPLIT: {
+      fill_mode_info_sb(cm, x, mi_row, mi_col, subsize, pc_tree->split[0]);
+      fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
+                        pc_tree->split[1]);
+      fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
+                        pc_tree->split[2]);
+      fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
+                        pc_tree->split[3]);
+      break;
+    }
+    default:
+      break;
+  }
+}
+
+// Reset the prediction pixel ready flag recursively.
+static void pred_pixel_ready_reset(PC_TREE *pc_tree, BLOCK_SIZE bsize) {
+  pc_tree->none.pred_pixel_ready = 0;
+  pc_tree->horizontal[0].pred_pixel_ready = 0;
+  pc_tree->horizontal[1].pred_pixel_ready = 0;
+  pc_tree->vertical[0].pred_pixel_ready = 0;
+  pc_tree->vertical[1].pred_pixel_ready = 0;
+
+  if (bsize > BLOCK_8X8) {
+    BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
+    int i;
+    for (i = 0; i < 4; ++i)
+      pred_pixel_ready_reset(pc_tree->split[i], subsize);
+  }
+}
+
+static void nonrd_pick_partition(VP9_COMP *cpi, ThreadData *td,
+                                 TileDataEnc *tile_data,
+                                 TOKENEXTRA **tp, int mi_row,
+                                 int mi_col, BLOCK_SIZE bsize, RD_COST *rd_cost,
+                                 int do_recon, int64_t best_rd,
+                                 PC_TREE *pc_tree) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  VP9_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
+  TOKENEXTRA *tp_orig = *tp;
+  PICK_MODE_CONTEXT *ctx = &pc_tree->none;
+  int i;
+  BLOCK_SIZE subsize = bsize;
+  RD_COST this_rdc, sum_rdc, best_rdc;
+  int do_split = bsize >= BLOCK_8X8;
+  int do_rect = 1;
+  // Override skipping rectangular partition operations for edge blocks
+  const int force_horz_split = (mi_row + ms >= cm->mi_rows);
+  const int force_vert_split = (mi_col + ms >= cm->mi_cols);
+  const int xss = x->e_mbd.plane[1].subsampling_x;
+  const int yss = x->e_mbd.plane[1].subsampling_y;
+
+  int partition_none_allowed = !force_horz_split && !force_vert_split;
+  int partition_horz_allowed = !force_vert_split && yss <= xss &&
+                               bsize >= BLOCK_8X8;
+  int partition_vert_allowed = !force_horz_split && xss <= yss &&
+                               bsize >= BLOCK_8X8;
+  (void) *tp_orig;
+
+  assert(num_8x8_blocks_wide_lookup[bsize] ==
+             num_8x8_blocks_high_lookup[bsize]);
+
+  vp9_rd_cost_init(&sum_rdc);
+  vp9_rd_cost_reset(&best_rdc);
+  best_rdc.rdcost = best_rd;
+
+  // Determine partition types in search according to the speed features.
+  // The threshold set here has to be of square block size.
+  if (sf->auto_min_max_partition_size) {
+    partition_none_allowed &= (bsize <= x->max_partition_size &&
+                               bsize >= x->min_partition_size);
+    partition_horz_allowed &= ((bsize <= x->max_partition_size &&
+                                bsize > x->min_partition_size) ||
+                                force_horz_split);
+    partition_vert_allowed &= ((bsize <= x->max_partition_size &&
+                                bsize > x->min_partition_size) ||
+                                force_vert_split);
+    do_split &= bsize > x->min_partition_size;
+  }
+  if (sf->use_square_partition_only) {
+    partition_horz_allowed &= force_horz_split;
+    partition_vert_allowed &= force_vert_split;
+  }
+
+  ctx->pred_pixel_ready = !(partition_vert_allowed ||
+                            partition_horz_allowed ||
+                            do_split);
+
+  // PARTITION_NONE
+  if (partition_none_allowed) {
+    nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
+                        &this_rdc, bsize, ctx);
+    ctx->mic = *xd->mi[0];
+    ctx->mbmi_ext = *x->mbmi_ext;
+    ctx->skip_txfm[0] = x->skip_txfm[0];
+    ctx->skip = x->skip;
+
+    if (this_rdc.rate != INT_MAX) {
+      int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
+      this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
+      this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                              this_rdc.rate, this_rdc.dist);
+      if (this_rdc.rdcost < best_rdc.rdcost) {
+        int64_t dist_breakout_thr = sf->partition_search_breakout_dist_thr;
+        int64_t rate_breakout_thr = sf->partition_search_breakout_rate_thr;
+
+        dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
+            b_height_log2_lookup[bsize]);
+
+        rate_breakout_thr *= num_pels_log2_lookup[bsize];
+
+        best_rdc = this_rdc;
+        if (bsize >= BLOCK_8X8)
+          pc_tree->partitioning = PARTITION_NONE;
+
+        if (!x->e_mbd.lossless &&
+            this_rdc.rate < rate_breakout_thr &&
+            this_rdc.dist < dist_breakout_thr) {
+          do_split = 0;
+          do_rect = 0;
+        }
+      }
+    }
+  }
+
+  // store estimated motion vector
+  store_pred_mv(x, ctx);
+
+  // PARTITION_SPLIT
+  if (do_split) {
+    int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
+    sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
+    sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+    subsize = get_subsize(bsize, PARTITION_SPLIT);
+    for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
+      const int x_idx = (i & 1) * ms;
+      const int y_idx = (i >> 1) * ms;
+
+      if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
+        continue;
+      load_pred_mv(x, ctx);
+      nonrd_pick_partition(cpi, td, tile_data, tp,
+                           mi_row + y_idx, mi_col + x_idx,
+                           subsize, &this_rdc, 0,
+                           best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
+
+      if (this_rdc.rate == INT_MAX) {
+        vp9_rd_cost_reset(&sum_rdc);
+      } else {
+        sum_rdc.rate += this_rdc.rate;
+        sum_rdc.dist += this_rdc.dist;
+        sum_rdc.rdcost += this_rdc.rdcost;
+      }
+    }
+
+    if (sum_rdc.rdcost < best_rdc.rdcost) {
+      best_rdc = sum_rdc;
+      pc_tree->partitioning = PARTITION_SPLIT;
+    } else {
+      // skip rectangular partition test when larger block size
+      // gives better rd cost
+      if (sf->less_rectangular_check)
+        do_rect &= !partition_none_allowed;
+    }
+  }
+
+  // PARTITION_HORZ
+  if (partition_horz_allowed && do_rect) {
+    subsize = get_subsize(bsize, PARTITION_HORZ);
+    if (sf->adaptive_motion_search)
+      load_pred_mv(x, ctx);
+    pc_tree->horizontal[0].pred_pixel_ready = 1;
+    nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
+                        &pc_tree->horizontal[0]);
+
+    pc_tree->horizontal[0].mic = *xd->mi[0];
+    pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
+    pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
+    pc_tree->horizontal[0].skip = x->skip;
+
+    if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + ms < cm->mi_rows) {
+      load_pred_mv(x, ctx);
+      pc_tree->horizontal[1].pred_pixel_ready = 1;
+      nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + ms, mi_col,
+                          &this_rdc, subsize,
+                          &pc_tree->horizontal[1]);
+
+      pc_tree->horizontal[1].mic = *xd->mi[0];
+      pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
+      pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
+      pc_tree->horizontal[1].skip = x->skip;
+
+      if (this_rdc.rate == INT_MAX) {
+        vp9_rd_cost_reset(&sum_rdc);
+      } else {
+        int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
+        this_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
+        sum_rdc.rate += this_rdc.rate;
+        sum_rdc.dist += this_rdc.dist;
+        sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                                sum_rdc.rate, sum_rdc.dist);
+      }
+    }
+
+    if (sum_rdc.rdcost < best_rdc.rdcost) {
+      best_rdc = sum_rdc;
+      pc_tree->partitioning = PARTITION_HORZ;
+    } else {
+      pred_pixel_ready_reset(pc_tree, bsize);
+    }
+  }
+
+  // PARTITION_VERT
+  if (partition_vert_allowed && do_rect) {
+    subsize = get_subsize(bsize, PARTITION_VERT);
+    if (sf->adaptive_motion_search)
+      load_pred_mv(x, ctx);
+    pc_tree->vertical[0].pred_pixel_ready = 1;
+    nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
+                        &pc_tree->vertical[0]);
+    pc_tree->vertical[0].mic = *xd->mi[0];
+    pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
+    pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
+    pc_tree->vertical[0].skip = x->skip;
+
+    if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + ms < cm->mi_cols) {
+      load_pred_mv(x, ctx);
+      pc_tree->vertical[1].pred_pixel_ready = 1;
+      nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + ms,
+                          &this_rdc, subsize,
+                          &pc_tree->vertical[1]);
+      pc_tree->vertical[1].mic = *xd->mi[0];
+      pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
+      pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
+      pc_tree->vertical[1].skip = x->skip;
+
+      if (this_rdc.rate == INT_MAX) {
+        vp9_rd_cost_reset(&sum_rdc);
+      } else {
+        int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
+        sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
+        sum_rdc.rate += this_rdc.rate;
+        sum_rdc.dist += this_rdc.dist;
+        sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                                sum_rdc.rate, sum_rdc.dist);
+      }
+    }
+
+    if (sum_rdc.rdcost < best_rdc.rdcost) {
+      best_rdc = sum_rdc;
+      pc_tree->partitioning = PARTITION_VERT;
+    } else {
+      pred_pixel_ready_reset(pc_tree, bsize);
+    }
+  }
+
+  *rd_cost = best_rdc;
+
+  if (best_rdc.rate == INT_MAX) {
+    vp9_rd_cost_reset(rd_cost);
+    return;
+  }
+
+  // update mode info array
+  fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, pc_tree);
+
+  if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && do_recon) {
+    int output_enabled = (bsize == BLOCK_64X64);
+    encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
+                 bsize, pc_tree);
+  }
+
+  if (bsize == BLOCK_64X64 && do_recon) {
+    assert(tp_orig < *tp);
+    assert(best_rdc.rate < INT_MAX);
+    assert(best_rdc.dist < INT64_MAX);
+  } else {
+    assert(tp_orig == *tp);
+  }
+}
+
+static void nonrd_select_partition(VP9_COMP *cpi,
+                                   ThreadData *td,
+                                   TileDataEnc *tile_data,
+                                   MODE_INFO **mi,
+                                   TOKENEXTRA **tp,
+                                   int mi_row, int mi_col,
+                                   BLOCK_SIZE bsize, int output_enabled,
+                                   RD_COST *rd_cost, PC_TREE *pc_tree) {
+  VP9_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
+  const int mis = cm->mi_stride;
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+  RD_COST this_rdc;
+
+  vp9_rd_cost_reset(&this_rdc);
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  subsize = (bsize >= BLOCK_8X8) ? mi[0]->sb_type : BLOCK_4X4;
+  partition = partition_lookup[bsl][subsize];
+
+  if (bsize == BLOCK_32X32 && subsize == BLOCK_32X32) {
+    x->max_partition_size = BLOCK_32X32;
+    x->min_partition_size = BLOCK_16X16;
+    nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
+                         rd_cost, 0, INT64_MAX, pc_tree);
+  } else if (bsize == BLOCK_32X32 && partition != PARTITION_NONE &&
+             subsize >= BLOCK_16X16) {
+    x->max_partition_size = BLOCK_32X32;
+    x->min_partition_size = BLOCK_8X8;
+    nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
+                         rd_cost, 0, INT64_MAX, pc_tree);
+  } else if (bsize == BLOCK_16X16 && partition != PARTITION_NONE) {
+    x->max_partition_size = BLOCK_16X16;
+    x->min_partition_size = BLOCK_8X8;
+    nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
+                         rd_cost, 0, INT64_MAX, pc_tree);
+  } else {
+    switch (partition) {
+      case PARTITION_NONE:
+        pc_tree->none.pred_pixel_ready = 1;
+        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
+                            subsize, &pc_tree->none);
+        pc_tree->none.mic = *xd->mi[0];
+        pc_tree->none.mbmi_ext = *x->mbmi_ext;
+        pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
+        pc_tree->none.skip = x->skip;
+        break;
+      case PARTITION_VERT:
+        pc_tree->vertical[0].pred_pixel_ready = 1;
+        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
+                            subsize, &pc_tree->vertical[0]);
+        pc_tree->vertical[0].mic = *xd->mi[0];
+        pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
+        pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
+        pc_tree->vertical[0].skip = x->skip;
+        if (mi_col + hbs < cm->mi_cols) {
+          pc_tree->vertical[1].pred_pixel_ready = 1;
+          nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
+                              &this_rdc, subsize, &pc_tree->vertical[1]);
+          pc_tree->vertical[1].mic = *xd->mi[0];
+          pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
+          pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
+          pc_tree->vertical[1].skip = x->skip;
+          if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
+              rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
+            rd_cost->rate += this_rdc.rate;
+            rd_cost->dist += this_rdc.dist;
+          }
+        }
+        break;
+      case PARTITION_HORZ:
+        pc_tree->horizontal[0].pred_pixel_ready = 1;
+        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
+                            subsize, &pc_tree->horizontal[0]);
+        pc_tree->horizontal[0].mic = *xd->mi[0];
+        pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
+        pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
+        pc_tree->horizontal[0].skip = x->skip;
+        if (mi_row + hbs < cm->mi_rows) {
+          pc_tree->horizontal[1].pred_pixel_ready = 1;
+          nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
+                              &this_rdc, subsize, &pc_tree->horizontal[1]);
+          pc_tree->horizontal[1].mic = *xd->mi[0];
+          pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
+          pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
+          pc_tree->horizontal[1].skip = x->skip;
+          if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
+              rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
+            rd_cost->rate += this_rdc.rate;
+            rd_cost->dist += this_rdc.dist;
+          }
+        }
+        break;
+      case PARTITION_SPLIT:
+        subsize = get_subsize(bsize, PARTITION_SPLIT);
+        nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
+                               subsize, output_enabled, rd_cost,
+                               pc_tree->split[0]);
+        nonrd_select_partition(cpi, td, tile_data, mi + hbs, tp,
+                               mi_row, mi_col + hbs, subsize, output_enabled,
+                               &this_rdc, pc_tree->split[1]);
+        if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
+            rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
+          rd_cost->rate += this_rdc.rate;
+          rd_cost->dist += this_rdc.dist;
+        }
+        nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis, tp,
+                               mi_row + hbs, mi_col, subsize, output_enabled,
+                               &this_rdc, pc_tree->split[2]);
+        if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
+            rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
+          rd_cost->rate += this_rdc.rate;
+          rd_cost->dist += this_rdc.dist;
+        }
+        nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
+                               mi_row + hbs, mi_col + hbs, subsize,
+                               output_enabled, &this_rdc, pc_tree->split[3]);
+        if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
+            rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
+          rd_cost->rate += this_rdc.rate;
+          rd_cost->dist += this_rdc.dist;
+        }
+        break;
+      default:
+        assert(0 && "Invalid partition type.");
+        break;
+    }
+  }
+
+  if (bsize == BLOCK_64X64 && output_enabled)
+    encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, 1, bsize, pc_tree);
+}
+
+
+static void nonrd_use_partition(VP9_COMP *cpi,
+                                ThreadData *td,
+                                TileDataEnc *tile_data,
+                                MODE_INFO **mi,
+                                TOKENEXTRA **tp,
+                                int mi_row, int mi_col,
+                                BLOCK_SIZE bsize, int output_enabled,
+                                RD_COST *dummy_cost, PC_TREE *pc_tree) {
+  VP9_COMMON *const cm = &cpi->common;
+  TileInfo *tile_info = &tile_data->tile_info;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
+  const int mis = cm->mi_stride;
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  subsize = (bsize >= BLOCK_8X8) ? mi[0]->sb_type : BLOCK_4X4;
+  partition = partition_lookup[bsl][subsize];
+
+  if (output_enabled && bsize != BLOCK_4X4) {
+    int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
+    td->counts->partition[ctx][partition]++;
+  }
+
+  switch (partition) {
+    case PARTITION_NONE:
+      pc_tree->none.pred_pixel_ready = 1;
+      nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
+                          subsize, &pc_tree->none);
+      pc_tree->none.mic = *xd->mi[0];
+      pc_tree->none.mbmi_ext = *x->mbmi_ext;
+      pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
+      pc_tree->none.skip = x->skip;
+      encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
+                  subsize, &pc_tree->none);
+      break;
+    case PARTITION_VERT:
+      pc_tree->vertical[0].pred_pixel_ready = 1;
+      nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
+                          subsize, &pc_tree->vertical[0]);
+      pc_tree->vertical[0].mic = *xd->mi[0];
+      pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
+      pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
+      pc_tree->vertical[0].skip = x->skip;
+      encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
+                  subsize, &pc_tree->vertical[0]);
+      if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
+        pc_tree->vertical[1].pred_pixel_ready = 1;
+        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
+                            dummy_cost, subsize, &pc_tree->vertical[1]);
+        pc_tree->vertical[1].mic = *xd->mi[0];
+        pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
+        pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
+        pc_tree->vertical[1].skip = x->skip;
+        encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col + hbs,
+                    output_enabled, subsize, &pc_tree->vertical[1]);
+      }
+      break;
+    case PARTITION_HORZ:
+      pc_tree->horizontal[0].pred_pixel_ready = 1;
+      nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
+                          subsize, &pc_tree->horizontal[0]);
+      pc_tree->horizontal[0].mic = *xd->mi[0];
+      pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
+      pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
+      pc_tree->horizontal[0].skip = x->skip;
+      encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
+                  subsize, &pc_tree->horizontal[0]);
+
+      if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
+        pc_tree->horizontal[1].pred_pixel_ready = 1;
+        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
+                            dummy_cost, subsize, &pc_tree->horizontal[1]);
+        pc_tree->horizontal[1].mic = *xd->mi[0];
+        pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
+        pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
+        pc_tree->horizontal[1].skip = x->skip;
+        encode_b_rt(cpi, td, tile_info, tp, mi_row + hbs, mi_col,
+                    output_enabled, subsize, &pc_tree->horizontal[1]);
+      }
+      break;
+    case PARTITION_SPLIT:
+      subsize = get_subsize(bsize, PARTITION_SPLIT);
+      if (bsize == BLOCK_8X8) {
+        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
+                            subsize, pc_tree->leaf_split[0]);
+        encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col,
+                    output_enabled, subsize, pc_tree->leaf_split[0]);
+      } else {
+        nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
+                            subsize, output_enabled, dummy_cost,
+                            pc_tree->split[0]);
+        nonrd_use_partition(cpi, td, tile_data, mi + hbs, tp,
+                            mi_row, mi_col + hbs, subsize, output_enabled,
+                            dummy_cost, pc_tree->split[1]);
+        nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis, tp,
+                            mi_row + hbs, mi_col, subsize, output_enabled,
+                            dummy_cost, pc_tree->split[2]);
+        nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
+                            mi_row + hbs, mi_col + hbs, subsize, output_enabled,
+                            dummy_cost, pc_tree->split[3]);
+      }
+      break;
+    default:
+      assert(0 && "Invalid partition type.");
+      break;
+  }
+
+  if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
+    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+}
+
+static void encode_nonrd_sb_row(VP9_COMP *cpi,
+                                ThreadData *td,
+                                TileDataEnc *tile_data,
+                                int mi_row,
+                                TOKENEXTRA **tp) {
+  SPEED_FEATURES *const sf = &cpi->sf;
+  VP9_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int mi_col_start = tile_info->mi_col_start;
+  const int mi_col_end = tile_info->mi_col_end;
+  int mi_col;
+
+  // Initialize the left context for the new SB row
+  memset(&xd->left_context, 0, sizeof(xd->left_context));
+  memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
+
+  // Code each SB in the row
+  for (mi_col = mi_col_start; mi_col < mi_col_end; mi_col += MI_BLOCK_SIZE) {
+    const struct segmentation *const seg = &cm->seg;
+    RD_COST dummy_rdc;
+    const int idx_str = cm->mi_stride * mi_row + mi_col;
+    MODE_INFO **mi = cm->mi_grid_visible + idx_str;
+    PARTITION_SEARCH_TYPE partition_search_type = sf->partition_search_type;
+    BLOCK_SIZE bsize = BLOCK_64X64;
+    int seg_skip = 0;
+    x->source_variance = UINT_MAX;
+    vp9_zero(x->pred_mv);
+    vp9_rd_cost_init(&dummy_rdc);
+    x->color_sensitivity[0] = 0;
+    x->color_sensitivity[1] = 0;
+    x->sb_is_skin = 0;
+
+    if (seg->enabled) {
+      const uint8_t *const map = seg->update_map ? cpi->segmentation_map
+                                                 : cm->last_frame_seg_map;
+      int segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
+      seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
+      if (seg_skip) {
+        partition_search_type = FIXED_PARTITION;
+      }
+    }
+
+    // Set the partition type of the 64X64 block
+    switch (partition_search_type) {
+      case VAR_BASED_PARTITION:
+        // TODO(jingning, marpan): The mode decision and encoding process
+        // support both intra and inter sub8x8 block coding for RTC mode.
+        // Tune the thresholds accordingly to use sub8x8 block coding for
+        // coding performance improvement.
+        choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
+        nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
+                            BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
+        break;
+      case SOURCE_VAR_BASED_PARTITION:
+        set_source_var_based_partition(cpi, tile_info, x, mi, mi_row, mi_col);
+        nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
+                            BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
+        break;
+      case FIXED_PARTITION:
+        if (!seg_skip)
+          bsize = sf->always_this_block_size;
+        set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
+        nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
+                            BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
+        break;
+      case REFERENCE_PARTITION:
+        set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
+        // Use nonrd_pick_partition on scene-cut for VBR, or on qp-segment
+        // if cyclic_refresh is enabled.
+        // nonrd_pick_partition does not support 4x4 partition, so avoid it
+        // on key frame for now.
+        if ((cpi->oxcf.rc_mode == VPX_VBR && cpi->rc.high_source_sad &&
+            cm->frame_type != KEY_FRAME) ||
+            (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
+            xd->mi[0]->segment_id)) {
+          // Use lower max_partition_size for low resoultions.
+          if (cm->width <= 352 && cm->height <= 288)
+            x->max_partition_size = BLOCK_32X32;
+          else
+            x->max_partition_size = BLOCK_64X64;
+          x->min_partition_size = BLOCK_8X8;
+          nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col,
+                               BLOCK_64X64, &dummy_rdc, 1,
+                               INT64_MAX, td->pc_root);
+        } else {
+          choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
+          // TODO(marpan): Seems like nonrd_select_partition does not support
+          // 4x4 partition. Since 4x4 is used on key frame, use this switch
+          // for now.
+          if (cm->frame_type == KEY_FRAME)
+            nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
+                                BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
+          else
+            nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
+                                   BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
+        }
+
+        break;
+      default:
+        assert(0);
+        break;
+    }
+  }
+}
+// end RTC play code
+
+static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  const VP9_COMMON *const cm = &cpi->common;
+
+  const uint8_t *src = cpi->Source->y_buffer;
+  const uint8_t *last_src = cpi->Last_Source->y_buffer;
+  const int src_stride = cpi->Source->y_stride;
+  const int last_stride = cpi->Last_Source->y_stride;
+
+  // Pick cutoff threshold
+  const int cutoff = (VPXMIN(cm->width, cm->height) >= 720) ?
+      (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100) :
+      (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
+  DECLARE_ALIGNED(16, int, hist[VAR_HIST_BINS]);
+  diff *var16 = cpi->source_diff_var;
+
+  int sum = 0;
+  int i, j;
+
+  memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
+
+  for (i = 0; i < cm->mb_rows; i++) {
+    for (j = 0; j < cm->mb_cols; j++) {
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (cm->use_highbitdepth) {
+        switch (cm->bit_depth) {
+          case VPX_BITS_8:
+            vpx_highbd_8_get16x16var(src, src_stride, last_src, last_stride,
+                                   &var16->sse, &var16->sum);
+            break;
+          case VPX_BITS_10:
+            vpx_highbd_10_get16x16var(src, src_stride, last_src, last_stride,
+                                    &var16->sse, &var16->sum);
+            break;
+          case VPX_BITS_12:
+            vpx_highbd_12_get16x16var(src, src_stride, last_src, last_stride,
+                                      &var16->sse, &var16->sum);
+            break;
+          default:
+            assert(0 && "cm->bit_depth should be VPX_BITS_8, VPX_BITS_10"
+                   " or VPX_BITS_12");
+            return -1;
+        }
+      } else {
+        vpx_get16x16var(src, src_stride, last_src, last_stride,
+                        &var16->sse, &var16->sum);
+      }
+#else
+      vpx_get16x16var(src, src_stride, last_src, last_stride,
+                      &var16->sse, &var16->sum);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      var16->var = var16->sse -
+          (((uint32_t)var16->sum * var16->sum) >> 8);
+
+      if (var16->var >= VAR_HIST_MAX_BG_VAR)
+        hist[VAR_HIST_BINS - 1]++;
+      else
+        hist[var16->var / VAR_HIST_FACTOR]++;
+
+      src += 16;
+      last_src += 16;
+      var16++;
+    }
+
+    src = src - cm->mb_cols * 16 + 16 * src_stride;
+    last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
+  }
+
+  cpi->source_var_thresh = 0;
+
+  if (hist[VAR_HIST_BINS - 1] < cutoff) {
+    for (i = 0; i < VAR_HIST_BINS - 1; i++) {
+      sum += hist[i];
+
+      if (sum > cutoff) {
+        cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
+        return 0;
+      }
+    }
+  }
+
+  return sf->search_type_check_frequency;
+}
+
+static void source_var_based_partition_search_method(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  SPEED_FEATURES *const sf = &cpi->sf;
+
+  if (cm->frame_type == KEY_FRAME) {
+    // For key frame, use SEARCH_PARTITION.
+    sf->partition_search_type = SEARCH_PARTITION;
+  } else if (cm->intra_only) {
+    sf->partition_search_type = FIXED_PARTITION;
+  } else {
+    if (cm->last_width != cm->width || cm->last_height != cm->height) {
+      if (cpi->source_diff_var)
+        vpx_free(cpi->source_diff_var);
+
+      CHECK_MEM_ERROR(cm, cpi->source_diff_var,
+                      vpx_calloc(cm->MBs, sizeof(diff)));
+    }
+
+    if (!cpi->frames_till_next_var_check)
+      cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
+
+    if (cpi->frames_till_next_var_check > 0) {
+      sf->partition_search_type = FIXED_PARTITION;
+      cpi->frames_till_next_var_check--;
+    }
+  }
+}
+
+static int get_skip_encode_frame(const VP9_COMMON *cm, ThreadData *const td) {
+  unsigned int intra_count = 0, inter_count = 0;
+  int j;
+
+  for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
+    intra_count += td->counts->intra_inter[j][0];
+    inter_count += td->counts->intra_inter[j][1];
+  }
+
+  return (intra_count << 2) < inter_count &&
+         cm->frame_type != KEY_FRAME && cm->show_frame;
+}
+
+void vp9_init_tile_data(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  int tile_col, tile_row;
+  TOKENEXTRA *pre_tok = cpi->tile_tok[0][0];
+  int tile_tok = 0;
+
+  if (cpi->tile_data == NULL || cpi->allocated_tiles < tile_cols * tile_rows) {
+    if (cpi->tile_data != NULL)
+      vpx_free(cpi->tile_data);
+    CHECK_MEM_ERROR(cm, cpi->tile_data,
+        vpx_malloc(tile_cols * tile_rows * sizeof(*cpi->tile_data)));
+    cpi->allocated_tiles = tile_cols * tile_rows;
+
+    for (tile_row = 0; tile_row < tile_rows; ++tile_row)
+      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+        TileDataEnc *tile_data =
+            &cpi->tile_data[tile_row * tile_cols + tile_col];
+        int i, j;
+        for (i = 0; i < BLOCK_SIZES; ++i) {
+          for (j = 0; j < MAX_MODES; ++j) {
+            tile_data->thresh_freq_fact[i][j] = 32;
+            tile_data->mode_map[i][j] = j;
+          }
+        }
+      }
+  }
+
+  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
+    for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+      TileInfo *tile_info =
+          &cpi->tile_data[tile_row * tile_cols + tile_col].tile_info;
+      vp9_tile_init(tile_info, cm, tile_row, tile_col);
+
+      cpi->tile_tok[tile_row][tile_col] = pre_tok + tile_tok;
+      pre_tok = cpi->tile_tok[tile_row][tile_col];
+      tile_tok = allocated_tokens(*tile_info);
+    }
+  }
+}
+
+void vp9_encode_tile(VP9_COMP *cpi, ThreadData *td,
+                     int tile_row, int tile_col) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  TileDataEnc *this_tile =
+      &cpi->tile_data[tile_row * tile_cols + tile_col];
+  const TileInfo * const tile_info = &this_tile->tile_info;
+  TOKENEXTRA *tok = cpi->tile_tok[tile_row][tile_col];
+  const int mi_row_start = tile_info->mi_row_start;
+  const int mi_row_end = tile_info->mi_row_end;
+  int mi_row;
+
+  // Set up pointers to per thread motion search counters.
+  td->mb.m_search_count_ptr = &td->rd_counts.m_search_count;
+  td->mb.ex_search_count_ptr = &td->rd_counts.ex_search_count;
+
+  for (mi_row = mi_row_start; mi_row < mi_row_end; mi_row += MI_BLOCK_SIZE) {
+    if (cpi->sf.use_nonrd_pick_mode)
+      encode_nonrd_sb_row(cpi, td, this_tile, mi_row, &tok);
+    else
+      encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok);
+  }
+  cpi->tok_count[tile_row][tile_col] =
+      (unsigned int)(tok - cpi->tile_tok[tile_row][tile_col]);
+  assert(tok - cpi->tile_tok[tile_row][tile_col] <=
+      allocated_tokens(*tile_info));
+}
+
+static void encode_tiles(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  int tile_col, tile_row;
+
+  vp9_init_tile_data(cpi);
+
+  for (tile_row = 0; tile_row < tile_rows; ++tile_row)
+    for (tile_col = 0; tile_col < tile_cols; ++tile_col)
+      vp9_encode_tile(cpi, &cpi->td, tile_row, tile_col);
+}
+
+#if CONFIG_FP_MB_STATS
+static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
+                            VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
+  uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
+      cm->current_video_frame * cm->MBs * sizeof(uint8_t);
+
+  if (mb_stats_in > firstpass_mb_stats->mb_stats_end)
+    return EOF;
+
+  *this_frame_mb_stats = mb_stats_in;
+
+  return 1;
+}
+#endif
+
+static void encode_frame_internal(VP9_COMP *cpi) {
+  SPEED_FEATURES *const sf = &cpi->sf;
+  ThreadData *const td = &cpi->td;
+  MACROBLOCK *const x = &td->mb;
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  RD_COUNTS *const rdc = &cpi->td.rd_counts;
+
+  xd->mi = cm->mi_grid_visible;
+  xd->mi[0] = cm->mi;
+
+  vp9_zero(*td->counts);
+  vp9_zero(rdc->coef_counts);
+  vp9_zero(rdc->comp_pred_diff);
+  vp9_zero(rdc->filter_diff);
+  rdc->m_search_count = 0;   // Count of motion search hits.
+  rdc->ex_search_count = 0;  // Exhaustive mesh search hits.
+
+  xd->lossless = cm->base_qindex == 0 &&
+                 cm->y_dc_delta_q == 0 &&
+                 cm->uv_dc_delta_q == 0 &&
+                 cm->uv_ac_delta_q == 0;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (cm->use_highbitdepth)
+    x->fwd_txm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
+  else
+    x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
+  x->highbd_itxm_add = xd->lossless ? vp9_highbd_iwht4x4_add :
+                                      vp9_highbd_idct4x4_add;
+#else
+  x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  x->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
+
+  if (xd->lossless)
+    x->optimize = 0;
+
+  cm->tx_mode = select_tx_mode(cpi, xd);
+
+  vp9_frame_init_quantizer(cpi);
+
+  vp9_initialize_rd_consts(cpi);
+  vp9_initialize_me_consts(cpi, x, cm->base_qindex);
+  init_encode_frame_mb_context(cpi);
+  cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
+                           cm->width == cm->last_width &&
+                           cm->height == cm->last_height &&
+                           !cm->intra_only &&
+                           cm->last_show_frame;
+  // Special case: set prev_mi to NULL when the previous mode info
+  // context cannot be used.
+  cm->prev_mi = cm->use_prev_frame_mvs ?
+                cm->prev_mip + cm->mi_stride + 1 : NULL;
+
+  x->quant_fp = cpi->sf.use_quant_fp;
+  vp9_zero(x->skip_txfm);
+  if (sf->use_nonrd_pick_mode) {
+    // Initialize internal buffer pointers for rtc coding, where non-RD
+    // mode decision is used and hence no buffer pointer swap needed.
+    int i;
+    struct macroblock_plane *const p = x->plane;
+    struct macroblockd_plane *const pd = xd->plane;
+    PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
+
+    for (i = 0; i < MAX_MB_PLANE; ++i) {
+      p[i].coeff = ctx->coeff_pbuf[i][0];
+      p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
+      pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
+      p[i].eobs = ctx->eobs_pbuf[i][0];
+    }
+    vp9_zero(x->zcoeff_blk);
+
+    if (cm->frame_type != KEY_FRAME &&
+        cpi->rc.frames_since_golden == 0 &&
+        !cpi->use_svc)
+      cpi->ref_frame_flags &= (~VP9_GOLD_FLAG);
+
+    if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
+      source_var_based_partition_search_method(cpi);
+  }
+
+  {
+    struct vpx_usec_timer emr_timer;
+    vpx_usec_timer_start(&emr_timer);
+
+#if CONFIG_FP_MB_STATS
+    if (cpi->use_fp_mb_stats) {
+      input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
+                       &cpi->twopass.this_frame_mb_stats);
+    }
+#endif
+
+    // If allowed, encoding tiles in parallel with one thread handling one tile.
+    if (VPXMIN(cpi->oxcf.max_threads, 1 << cm->log2_tile_cols) > 1)
+      vp9_encode_tiles_mt(cpi);
+    else
+      encode_tiles(cpi);
+
+    vpx_usec_timer_mark(&emr_timer);
+    cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
+  }
+
+  sf->skip_encode_frame = sf->skip_encode_sb ?
+      get_skip_encode_frame(cm, td) : 0;
+
+#if 0
+  // Keep record of the total distortion this time around for future use
+  cpi->last_frame_distortion = cpi->frame_distortion;
+#endif
+}
+
+static INTERP_FILTER get_interp_filter(
+    const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
+  if (!is_alt_ref &&
+      threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
+      threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
+      threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
+    return EIGHTTAP_SMOOTH;
+  } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
+             threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
+    return EIGHTTAP_SHARP;
+  } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
+    return EIGHTTAP;
+  } else {
+    return SWITCHABLE;
+  }
+}
+
+static int compute_frame_aq_offset(struct VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
+  struct segmentation *const seg = &cm->seg;
+
+  int mi_row, mi_col;
+  int sum_delta = 0;
+  int map_index = 0;
+  int qdelta_index;
+  int segment_id;
+
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
+    MODE_INFO **mi_8x8 = mi_8x8_ptr;
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++, mi_8x8++) {
+      segment_id = mi_8x8[0]->segment_id;
+      qdelta_index = get_segdata(seg, segment_id, SEG_LVL_ALT_Q);
+      sum_delta += qdelta_index;
+      map_index++;
+    }
+    mi_8x8_ptr += cm->mi_stride;
+  }
+
+  return sum_delta / (cm->mi_rows * cm->mi_cols);
+}
+
+void vp9_encode_frame(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+
+  // In the longer term the encoder should be generalized to match the
+  // decoder such that we allow compound where one of the 3 buffers has a
+  // different sign bias and that buffer is then the fixed ref. However, this
+  // requires further work in the rd loop. For now the only supported encoder
+  // side behavior is where the ALT ref buffer has opposite sign bias to
+  // the other two.
+  if (!frame_is_intra_only(cm)) {
+    if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
+             cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
+        (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
+             cm->ref_frame_sign_bias[LAST_FRAME])) {
+      cpi->allow_comp_inter_inter = 0;
+    } else {
+      cpi->allow_comp_inter_inter = 1;
+      cm->comp_fixed_ref = ALTREF_FRAME;
+      cm->comp_var_ref[0] = LAST_FRAME;
+      cm->comp_var_ref[1] = GOLDEN_FRAME;
+    }
+  }
+
+  if (cpi->sf.frame_parameter_update) {
+    int i;
+    RD_OPT *const rd_opt = &cpi->rd;
+    FRAME_COUNTS *counts = cpi->td.counts;
+    RD_COUNTS *const rdc = &cpi->td.rd_counts;
+
+    // This code does a single RD pass over the whole frame assuming
+    // either compound, single or hybrid prediction as per whatever has
+    // worked best for that type of frame in the past.
+    // It also predicts whether another coding mode would have worked
+    // better that this coding mode. If that is the case, it remembers
+    // that for subsequent frames.
+    // It does the same analysis for transform size selection also.
+    const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
+    int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
+    int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
+    const int is_alt_ref = frame_type == ALTREF_FRAME;
+
+    /* prediction (compound, single or hybrid) mode selection */
+    if (is_alt_ref || !cpi->allow_comp_inter_inter)
+      cm->reference_mode = SINGLE_REFERENCE;
+    else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
+             mode_thrs[COMPOUND_REFERENCE] >
+                 mode_thrs[REFERENCE_MODE_SELECT] &&
+             check_dual_ref_flags(cpi) &&
+             cpi->static_mb_pct == 100)
+      cm->reference_mode = COMPOUND_REFERENCE;
+    else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
+      cm->reference_mode = SINGLE_REFERENCE;
+    else
+      cm->reference_mode = REFERENCE_MODE_SELECT;
+
+    if (cm->interp_filter == SWITCHABLE)
+      cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
+
+    encode_frame_internal(cpi);
+
+    for (i = 0; i < REFERENCE_MODES; ++i)
+      mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2;
+
+    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+      filter_thrs[i] = (filter_thrs[i] + rdc->filter_diff[i] / cm->MBs) / 2;
+
+    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+      int single_count_zero = 0;
+      int comp_count_zero = 0;
+
+      for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
+        single_count_zero += counts->comp_inter[i][0];
+        comp_count_zero += counts->comp_inter[i][1];
+      }
+
+      if (comp_count_zero == 0) {
+        cm->reference_mode = SINGLE_REFERENCE;
+        vp9_zero(counts->comp_inter);
+      } else if (single_count_zero == 0) {
+        cm->reference_mode = COMPOUND_REFERENCE;
+        vp9_zero(counts->comp_inter);
+      }
+    }
+
+    if (cm->tx_mode == TX_MODE_SELECT) {
+      int count4x4 = 0;
+      int count8x8_lp = 0, count8x8_8x8p = 0;
+      int count16x16_16x16p = 0, count16x16_lp = 0;
+      int count32x32 = 0;
+
+      for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
+        count4x4 += counts->tx.p32x32[i][TX_4X4];
+        count4x4 += counts->tx.p16x16[i][TX_4X4];
+        count4x4 += counts->tx.p8x8[i][TX_4X4];
+
+        count8x8_lp += counts->tx.p32x32[i][TX_8X8];
+        count8x8_lp += counts->tx.p16x16[i][TX_8X8];
+        count8x8_8x8p += counts->tx.p8x8[i][TX_8X8];
+
+        count16x16_16x16p += counts->tx.p16x16[i][TX_16X16];
+        count16x16_lp += counts->tx.p32x32[i][TX_16X16];
+        count32x32 += counts->tx.p32x32[i][TX_32X32];
+      }
+      if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
+          count32x32 == 0) {
+        cm->tx_mode = ALLOW_8X8;
+        reset_skip_tx_size(cm, TX_8X8);
+      } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
+                 count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
+        cm->tx_mode = ONLY_4X4;
+        reset_skip_tx_size(cm, TX_4X4);
+      } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
+        cm->tx_mode = ALLOW_32X32;
+      } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
+        cm->tx_mode = ALLOW_16X16;
+        reset_skip_tx_size(cm, TX_16X16);
+      }
+    }
+  } else {
+    cm->reference_mode = SINGLE_REFERENCE;
+    encode_frame_internal(cpi);
+  }
+
+  // If segmentated AQ is enabled compute the average AQ weighting.
+  if (cm->seg.enabled && (cpi->oxcf.aq_mode != NO_AQ) &&
+      (cm->seg.update_map || cm->seg.update_data)) {
+    cm->seg.aq_av_offset = compute_frame_aq_offset(cpi);
+  }
+}
+static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
+  const PREDICTION_MODE y_mode = mi->mode;
+  const PREDICTION_MODE uv_mode = mi->uv_mode;
+  const BLOCK_SIZE bsize = mi->sb_type;
+
+  if (bsize < BLOCK_8X8) {
+    int idx, idy;
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+    for (idy = 0; idy < 2; idy += num_4x4_h)
+      for (idx = 0; idx < 2; idx += num_4x4_w)
+        ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
+  } else {
+    ++counts->y_mode[size_group_lookup[bsize]][y_mode];
+  }
+
+  ++counts->uv_mode[y_mode][uv_mode];
+}
+
+static void update_zeromv_cnt(VP9_COMP *const cpi,
+                              const MODE_INFO *const mi,
+                              int mi_row, int mi_col,
+                              BLOCK_SIZE bsize) {
+  const VP9_COMMON *const cm = &cpi->common;
+  MV mv = mi->mv[0].as_mv;
+  const int bw = num_8x8_blocks_wide_lookup[bsize];
+  const int bh = num_8x8_blocks_high_lookup[bsize];
+  const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
+  const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
+  const int block_index = mi_row * cm->mi_cols + mi_col;
+  int x, y;
+  for (y = 0; y < ymis; y++)
+    for (x = 0; x < xmis; x++) {
+      int map_offset = block_index + y * cm->mi_cols + x;
+      if (is_inter_block(mi) && mi->segment_id <= CR_SEGMENT_ID_BOOST2) {
+        if (abs(mv.row) < 8 && abs(mv.col) < 8) {
+          if (cpi->consec_zero_mv[map_offset] < 255)
+           cpi->consec_zero_mv[map_offset]++;
+        } else {
+          cpi->consec_zero_mv[map_offset] = 0;
+        }
+      }
+    }
+}
+
+static void encode_superblock(VP9_COMP *cpi, ThreadData *td,
+                              TOKENEXTRA **t, int output_enabled,
+                              int mi_row, int mi_col, BLOCK_SIZE bsize,
+                              PICK_MODE_CONTEXT *ctx) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  const int seg_skip = segfeature_active(&cm->seg, mi->segment_id,
+                                         SEG_LVL_SKIP);
+
+  x->skip_recode = !x->select_tx_size && mi->sb_type >= BLOCK_8X8 &&
+                   cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
+                   cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
+                   cpi->sf.allow_skip_recode;
+
+  if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
+    memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
+
+  x->skip_optimize = ctx->is_coded;
+  ctx->is_coded = 1;
+  x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
+  x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
+                    x->q_index < QIDX_SKIP_THRESH);
+
+  if (x->skip_encode)
+    return;
+
+  if (!is_inter_block(mi)) {
+    int plane;
+#if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
+    if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
+        (xd->above_mi == NULL || xd->left_mi == NULL) &&
+        need_top_left[mi->uv_mode])
+      assert(0);
+#endif  // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
+    mi->skip = 1;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane)
+      vp9_encode_intra_block_plane(x, VPXMAX(bsize, BLOCK_8X8), plane, 1);
+    if (output_enabled)
+      sum_intra_stats(td->counts, mi);
+    vp9_tokenize_sb(cpi, td, t, !output_enabled, seg_skip,
+                    VPXMAX(bsize, BLOCK_8X8));
+  } else {
+    int ref;
+    const int is_compound = has_second_ref(mi);
+    set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
+    for (ref = 0; ref < 1 + is_compound; ++ref) {
+      YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi,
+                                                     mi->ref_frame[ref]);
+      assert(cfg != NULL);
+      vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
+                           &xd->block_refs[ref]->sf);
+    }
+    if (!(cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready) || seg_skip)
+      vp9_build_inter_predictors_sby(xd, mi_row, mi_col,
+                                     VPXMAX(bsize, BLOCK_8X8));
+
+    vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col,
+                                    VPXMAX(bsize, BLOCK_8X8));
+
+    vp9_encode_sb(x, VPXMAX(bsize, BLOCK_8X8));
+    vp9_tokenize_sb(cpi, td, t, !output_enabled, seg_skip,
+                    VPXMAX(bsize, BLOCK_8X8));
+  }
+
+  if (output_enabled) {
+    if (cm->tx_mode == TX_MODE_SELECT &&
+        mi->sb_type >= BLOCK_8X8  &&
+        !(is_inter_block(mi) && (mi->skip || seg_skip))) {
+      ++get_tx_counts(max_txsize_lookup[bsize], get_tx_size_context(xd),
+                      &td->counts->tx)[mi->tx_size];
+    } else {
+      // The new intra coding scheme requires no change of transform size
+      if (is_inter_block(mi)) {
+        mi->tx_size = VPXMIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
+                             max_txsize_lookup[bsize]);
+      } else {
+        mi->tx_size = (bsize >= BLOCK_8X8) ? mi->tx_size : TX_4X4;
+      }
+    }
+
+    ++td->counts->tx.tx_totals[mi->tx_size];
+    ++td->counts->tx.tx_totals[get_uv_tx_size(mi, &xd->plane[1])];
+    if (cm->seg.enabled && cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
+      vp9_cyclic_refresh_update_sb_postencode(cpi, mi, mi_row, mi_col, bsize);
+    if (cpi->oxcf.pass == 0 && cpi->svc.temporal_layer_id == 0)
+      update_zeromv_cnt(cpi, mi, mi_row, mi_col, bsize);
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_encodeframe.h b/libvpx/libvpx/vp9/encoder/vp9_encodeframe.h
new file mode 100644
index 0000000..6aaa564
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_encodeframe.h
@@ -0,0 +1,49 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_ENCODEFRAME_H_
+#define VP9_ENCODER_VP9_ENCODEFRAME_H_
+
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct macroblock;
+struct yv12_buffer_config;
+struct VP9_COMP;
+struct ThreadData;
+
+// Constants used in SOURCE_VAR_BASED_PARTITION
+#define VAR_HIST_MAX_BG_VAR 1000
+#define VAR_HIST_FACTOR 10
+#define VAR_HIST_BINS (VAR_HIST_MAX_BG_VAR / VAR_HIST_FACTOR + 1)
+#define VAR_HIST_LARGE_CUT_OFF 75
+#define VAR_HIST_SMALL_CUT_OFF 45
+
+void vp9_setup_src_planes(struct macroblock *x,
+                          const struct yv12_buffer_config *src,
+                          int mi_row, int mi_col);
+
+void vp9_encode_frame(struct VP9_COMP *cpi);
+
+void vp9_init_tile_data(struct VP9_COMP *cpi);
+void vp9_encode_tile(struct VP9_COMP *cpi, struct ThreadData *td,
+                     int tile_row, int tile_col);
+
+void vp9_set_variance_partition_thresholds(struct VP9_COMP *cpi, int q);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_ENCODEFRAME_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_encodemb.c b/libvpx/libvpx/vp9/encoder/vp9_encodemb.c
new file mode 100644
index 0000000..169943c
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_encodemb.c
@@ -0,0 +1,1019 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_dsp/quantize.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_idct.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_scan.h"
+
+#include "vp9/encoder/vp9_encodemb.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_tokenize.h"
+
+struct optimize_ctx {
+  ENTROPY_CONTEXT ta[MAX_MB_PLANE][16];
+  ENTROPY_CONTEXT tl[MAX_MB_PLANE][16];
+};
+
+void vp9_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
+  struct macroblock_plane *const p = &x->plane[plane];
+  const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+  const int bw = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
+  const int bh = 4 * num_4x4_blocks_high_lookup[plane_bsize];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (x->e_mbd.cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vpx_highbd_subtract_block(bh, bw, p->src_diff, bw, p->src.buf,
+                              p->src.stride, pd->dst.buf, pd->dst.stride,
+                              x->e_mbd.bd);
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
+                     pd->dst.buf, pd->dst.stride);
+}
+
+typedef struct vp9_token_state {
+  int64_t       error;
+  int           rate;
+  int16_t       next;
+  int16_t       token;
+  tran_low_t    qc;
+  tran_low_t    dqc;
+} vp9_token_state;
+
+static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] ={ {10, 6}, {8, 7}, };
+
+#define UPDATE_RD_COST()\
+{\
+  rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);\
+  rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);\
+}
+
+// This function is a place holder for now but may ultimately need
+// to scan previous tokens to work out the correct context.
+static int trellis_get_coeff_context(const int16_t *scan,
+                                     const int16_t *nb,
+                                     int idx, int token,
+                                     uint8_t *token_cache) {
+  int bak = token_cache[scan[idx]], pt;
+  token_cache[scan[idx]] = vp9_pt_energy_class[token];
+  pt = get_coef_context(nb, token_cache, idx + 1);
+  token_cache[scan[idx]] = bak;
+  return pt;
+}
+
+static int optimize_b(MACROBLOCK *mb, int plane, int block,
+                      TX_SIZE tx_size, int ctx) {
+  MACROBLOCKD *const xd = &mb->e_mbd;
+  struct macroblock_plane *const p = &mb->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int ref = is_inter_block(xd->mi[0]);
+  vp9_token_state tokens[1025][2];
+  unsigned best_index[1025][2];
+  uint8_t token_cache[1024];
+  const tran_low_t *const coeff = BLOCK_OFFSET(mb->plane[plane].coeff, block);
+  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  const int eob = p->eobs[block];
+  const PLANE_TYPE type = get_plane_type(plane);
+  const int default_eob = 16 << (tx_size << 1);
+  const int shift = (tx_size == TX_32X32);
+  const int16_t* const dequant_ptr = pd->dequant;
+  const uint8_t* const band_translate = get_band_translate(tx_size);
+  const scan_order *const so = get_scan(xd, tx_size, type, block);
+  const int16_t *const scan = so->scan;
+  const int16_t *const nb = so->neighbors;
+  const int dq_step[2] = { dequant_ptr[0] >> shift, dequant_ptr[1] >> shift };
+  int next = eob, sz = 0;
+  const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][type]) >> 1;
+  const int64_t rddiv = mb->rddiv;
+  int64_t rd_cost0, rd_cost1;
+  int rate0, rate1;
+  int64_t error0, error1;
+  int16_t t0, t1;
+  EXTRABIT e0;
+  int best, band, pt, i, final_eob;
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
+#else
+  const int *cat6_high_cost = vp9_get_high_cost_table(8);
+#endif
+
+  assert((!type && !plane) || (type && plane));
+  assert(eob <= default_eob);
+
+  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
+  /* Initialize the sentinel node of the trellis. */
+  tokens[eob][0].rate = 0;
+  tokens[eob][0].error = 0;
+  tokens[eob][0].next = default_eob;
+  tokens[eob][0].token = EOB_TOKEN;
+  tokens[eob][0].qc = 0;
+  tokens[eob][1] = tokens[eob][0];
+
+  for (i = 0; i < eob; i++)
+    token_cache[scan[i]] =
+        vp9_pt_energy_class[vp9_get_token(qcoeff[scan[i]])];
+
+  for (i = eob; i-- > 0;) {
+    int base_bits, d2, dx;
+    const int rc = scan[i];
+    int x = qcoeff[rc];
+    /* Only add a trellis state for non-zero coefficients. */
+    if (x) {
+      int shortcut = 0;
+      error0 = tokens[next][0].error;
+      error1 = tokens[next][1].error;
+      /* Evaluate the first possibility for this state. */
+      rate0 = tokens[next][0].rate;
+      rate1 = tokens[next][1].rate;
+      vp9_get_token_extra(x, &t0, &e0);
+      /* Consider both possible successor states. */
+      if (next < default_eob) {
+        band = band_translate[i + 1];
+        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
+        rate0 += mb->token_costs[tx_size][type][ref][band][0][pt]
+                                [tokens[next][0].token];
+        rate1 += mb->token_costs[tx_size][type][ref][band][0][pt]
+                                [tokens[next][1].token];
+      }
+      UPDATE_RD_COST();
+      /* And pick the best. */
+      best = rd_cost1 < rd_cost0;
+      base_bits = vp9_get_cost(t0, e0, cat6_high_cost);
+      dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        dx >>= xd->bd - 8;
+      }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      d2 = dx * dx;
+      tokens[i][0].rate = base_bits + (best ? rate1 : rate0);
+      tokens[i][0].error = d2 + (best ? error1 : error0);
+      tokens[i][0].next = next;
+      tokens[i][0].token = t0;
+      tokens[i][0].qc = x;
+      tokens[i][0].dqc = dqcoeff[rc];
+      best_index[i][0] = best;
+
+      /* Evaluate the second possibility for this state. */
+      rate0 = tokens[next][0].rate;
+      rate1 = tokens[next][1].rate;
+
+      if ((abs(x) * dequant_ptr[rc != 0] > (abs(coeff[rc]) << shift)) &&
+          (abs(x) * dequant_ptr[rc != 0] < (abs(coeff[rc]) << shift) +
+                                               dequant_ptr[rc != 0]))
+        shortcut = 1;
+      else
+        shortcut = 0;
+
+      if (shortcut) {
+        sz = -(x < 0);
+        x -= 2 * sz + 1;
+      } else {
+        tokens[i][1] = tokens[i][0];
+        best_index[i][1] = best_index[i][0];
+        next = i;
+        continue;
+      }
+
+      /* Consider both possible successor states. */
+      if (!x) {
+        /* If we reduced this coefficient to zero, check to see if
+         *  we need to move the EOB back here.
+         */
+        t0 = tokens[next][0].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
+        t1 = tokens[next][1].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
+        e0 = 0;
+      } else {
+        vp9_get_token_extra(x, &t0, &e0);
+        t1 = t0;
+      }
+      if (next < default_eob) {
+        band = band_translate[i + 1];
+        if (t0 != EOB_TOKEN) {
+          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
+          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
+                                  [tokens[next][0].token];
+        }
+        if (t1 != EOB_TOKEN) {
+          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
+          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
+                                  [tokens[next][1].token];
+        }
+      }
+
+      UPDATE_RD_COST();
+      /* And pick the best. */
+      best = rd_cost1 < rd_cost0;
+      base_bits = vp9_get_cost(t0, e0, cat6_high_cost);
+
+      if (shortcut) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          dx -= ((dequant_ptr[rc != 0] >> (xd->bd - 8)) + sz) ^ sz;
+        } else {
+          dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
+        }
+#else
+        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        d2 = dx * dx;
+      }
+      tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
+      tokens[i][1].error = d2 + (best ? error1 : error0);
+      tokens[i][1].next = next;
+      tokens[i][1].token = best ? t1 : t0;
+      tokens[i][1].qc = x;
+
+      if (x) {
+        tran_low_t offset = dq_step[rc != 0];
+        // The 32x32 transform coefficient uses half quantization step size.
+        // Account for the rounding difference in the dequantized coefficeint
+        // value when the quantization index is dropped from an even number
+        // to an odd number.
+        if (shift & x)
+          offset += (dequant_ptr[rc != 0] & 0x01);
+
+        if (sz == 0)
+          tokens[i][1].dqc = dqcoeff[rc] - offset;
+        else
+          tokens[i][1].dqc = dqcoeff[rc] + offset;
+      } else {
+        tokens[i][1].dqc = 0;
+      }
+
+      best_index[i][1] = best;
+      /* Finally, make this the new head of the trellis. */
+      next = i;
+    } else {
+      /* There's no choice to make for a zero coefficient, so we don't
+       *  add a new trellis node, but we do need to update the costs.
+       */
+      band = band_translate[i + 1];
+      t0 = tokens[next][0].token;
+      t1 = tokens[next][1].token;
+      /* Update the cost of each path if we're past the EOB token. */
+      if (t0 != EOB_TOKEN) {
+        tokens[next][0].rate +=
+            mb->token_costs[tx_size][type][ref][band][1][0][t0];
+        tokens[next][0].token = ZERO_TOKEN;
+      }
+      if (t1 != EOB_TOKEN) {
+        tokens[next][1].rate +=
+            mb->token_costs[tx_size][type][ref][band][1][0][t1];
+        tokens[next][1].token = ZERO_TOKEN;
+      }
+      best_index[i][0] = best_index[i][1] = 0;
+      /* Don't update next, because we didn't add a new node. */
+    }
+  }
+
+  /* Now pick the best path through the whole trellis. */
+  band = band_translate[i + 1];
+  rate0 = tokens[next][0].rate;
+  rate1 = tokens[next][1].rate;
+  error0 = tokens[next][0].error;
+  error1 = tokens[next][1].error;
+  t0 = tokens[next][0].token;
+  t1 = tokens[next][1].token;
+  rate0 += mb->token_costs[tx_size][type][ref][band][0][ctx][t0];
+  rate1 += mb->token_costs[tx_size][type][ref][band][0][ctx][t1];
+  UPDATE_RD_COST();
+  best = rd_cost1 < rd_cost0;
+  final_eob = -1;
+
+  for (i = next; i < eob; i = next) {
+    const int x = tokens[i][best].qc;
+    const int rc = scan[i];
+    if (x) final_eob = i;
+    qcoeff[rc] = x;
+    dqcoeff[rc] = tokens[i][best].dqc;
+    next = tokens[i][best].next;
+    best = best_index[i][best];
+  }
+  final_eob++;
+
+  mb->plane[plane].eobs[block] = final_eob;
+  return final_eob;
+}
+
+static INLINE void fdct32x32(int rd_transform,
+                             const int16_t *src, tran_low_t *dst,
+                             int src_stride) {
+  if (rd_transform)
+    vpx_fdct32x32_rd(src, dst, src_stride);
+  else
+    vpx_fdct32x32(src, dst, src_stride);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE void highbd_fdct32x32(int rd_transform, const int16_t *src,
+                                    tran_low_t *dst, int src_stride) {
+  if (rd_transform)
+    vpx_highbd_fdct32x32_rd(src, dst, src_stride);
+  else
+    vpx_highbd_fdct32x32(src, dst, src_stride);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_xform_quant_fp(MACROBLOCK *x, int plane, int block,
+                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const struct macroblock_plane *const p = &x->plane[plane];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
+  tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  uint16_t *const eob = &p->eobs[block];
+  const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
+  int i, j;
+  const int16_t *src_diff;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
+  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    switch (tx_size) {
+      case TX_32X32:
+        highbd_fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
+        vp9_highbd_quantize_fp_32x32(coeff, 1024, x->skip_block, p->zbin,
+                                     p->round_fp, p->quant_fp, p->quant_shift,
+                                     qcoeff, dqcoeff, pd->dequant,
+                                     eob, scan_order->scan,
+                                     scan_order->iscan);
+        break;
+      case TX_16X16:
+        vpx_highbd_fdct16x16(src_diff, coeff, diff_stride);
+        vp9_highbd_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
+                               p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                               pd->dequant, eob,
+                               scan_order->scan, scan_order->iscan);
+        break;
+      case TX_8X8:
+        vpx_highbd_fdct8x8(src_diff, coeff, diff_stride);
+        vp9_highbd_quantize_fp(coeff, 64, x->skip_block, p->zbin, p->round_fp,
+                               p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                               pd->dequant, eob,
+                               scan_order->scan, scan_order->iscan);
+        break;
+      case TX_4X4:
+        x->fwd_txm4x4(src_diff, coeff, diff_stride);
+        vp9_highbd_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
+                               p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                               pd->dequant, eob,
+                               scan_order->scan, scan_order->iscan);
+        break;
+      default:
+        assert(0);
+    }
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  switch (tx_size) {
+    case TX_32X32:
+      fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
+      vp9_quantize_fp_32x32(coeff, 1024, x->skip_block, p->zbin, p->round_fp,
+                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                            pd->dequant, eob, scan_order->scan,
+                            scan_order->iscan);
+      break;
+    case TX_16X16:
+      vpx_fdct16x16(src_diff, coeff, diff_stride);
+      vp9_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
+                      p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                      pd->dequant, eob,
+                      scan_order->scan, scan_order->iscan);
+      break;
+    case TX_8X8:
+      vp9_fdct8x8_quant(src_diff, diff_stride, coeff, 64,
+                        x->skip_block, p->zbin, p->round_fp,
+                        p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                        pd->dequant, eob,
+                        scan_order->scan, scan_order->iscan);
+      break;
+    case TX_4X4:
+      x->fwd_txm4x4(src_diff, coeff, diff_stride);
+      vp9_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
+                      p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                      pd->dequant, eob,
+                      scan_order->scan, scan_order->iscan);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
+
+void vp9_xform_quant_dc(MACROBLOCK *x, int plane, int block,
+                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const struct macroblock_plane *const p = &x->plane[plane];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  uint16_t *const eob = &p->eobs[block];
+  const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
+  int i, j;
+  const int16_t *src_diff;
+
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
+  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    switch (tx_size) {
+      case TX_32X32:
+        vpx_highbd_fdct32x32_1(src_diff, coeff, diff_stride);
+        vpx_highbd_quantize_dc_32x32(coeff, x->skip_block, p->round,
+                                     p->quant_fp[0], qcoeff, dqcoeff,
+                                     pd->dequant[0], eob);
+        break;
+      case TX_16X16:
+        vpx_highbd_fdct16x16_1(src_diff, coeff, diff_stride);
+        vpx_highbd_quantize_dc(coeff, 256, x->skip_block, p->round,
+                               p->quant_fp[0], qcoeff, dqcoeff,
+                               pd->dequant[0], eob);
+        break;
+      case TX_8X8:
+        vpx_highbd_fdct8x8_1(src_diff, coeff, diff_stride);
+        vpx_highbd_quantize_dc(coeff, 64, x->skip_block, p->round,
+                               p->quant_fp[0], qcoeff, dqcoeff,
+                               pd->dequant[0], eob);
+        break;
+      case TX_4X4:
+        x->fwd_txm4x4(src_diff, coeff, diff_stride);
+        vpx_highbd_quantize_dc(coeff, 16, x->skip_block, p->round,
+                               p->quant_fp[0], qcoeff, dqcoeff,
+                               pd->dequant[0], eob);
+        break;
+      default:
+        assert(0);
+    }
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  switch (tx_size) {
+    case TX_32X32:
+      vpx_fdct32x32_1(src_diff, coeff, diff_stride);
+      vpx_quantize_dc_32x32(coeff, x->skip_block, p->round,
+                            p->quant_fp[0], qcoeff, dqcoeff,
+                            pd->dequant[0], eob);
+      break;
+    case TX_16X16:
+      vpx_fdct16x16_1(src_diff, coeff, diff_stride);
+      vpx_quantize_dc(coeff, 256, x->skip_block, p->round,
+                     p->quant_fp[0], qcoeff, dqcoeff,
+                     pd->dequant[0], eob);
+      break;
+    case TX_8X8:
+      vpx_fdct8x8_1(src_diff, coeff, diff_stride);
+      vpx_quantize_dc(coeff, 64, x->skip_block, p->round,
+                      p->quant_fp[0], qcoeff, dqcoeff,
+                      pd->dequant[0], eob);
+      break;
+    case TX_4X4:
+      x->fwd_txm4x4(src_diff, coeff, diff_stride);
+      vpx_quantize_dc(coeff, 16, x->skip_block, p->round,
+                      p->quant_fp[0], qcoeff, dqcoeff,
+                      pd->dequant[0], eob);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
+
+void vp9_xform_quant(MACROBLOCK *x, int plane, int block,
+                     BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const struct macroblock_plane *const p = &x->plane[plane];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
+  tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  uint16_t *const eob = &p->eobs[block];
+  const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
+  int i, j;
+  const int16_t *src_diff;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
+  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+     switch (tx_size) {
+      case TX_32X32:
+        highbd_fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
+        vpx_highbd_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin,
+                                    p->round, p->quant, p->quant_shift, qcoeff,
+                                    dqcoeff, pd->dequant, eob,
+                                    scan_order->scan, scan_order->iscan);
+        break;
+      case TX_16X16:
+        vpx_highbd_fdct16x16(src_diff, coeff, diff_stride);
+        vpx_highbd_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
+                              p->quant, p->quant_shift, qcoeff, dqcoeff,
+                              pd->dequant, eob,
+                              scan_order->scan, scan_order->iscan);
+        break;
+      case TX_8X8:
+        vpx_highbd_fdct8x8(src_diff, coeff, diff_stride);
+        vpx_highbd_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
+                              p->quant, p->quant_shift, qcoeff, dqcoeff,
+                              pd->dequant, eob,
+                              scan_order->scan, scan_order->iscan);
+        break;
+      case TX_4X4:
+        x->fwd_txm4x4(src_diff, coeff, diff_stride);
+        vpx_highbd_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
+                              p->quant, p->quant_shift, qcoeff, dqcoeff,
+                              pd->dequant, eob,
+                              scan_order->scan, scan_order->iscan);
+        break;
+      default:
+        assert(0);
+    }
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  switch (tx_size) {
+    case TX_32X32:
+      fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
+      vpx_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
+                           p->quant, p->quant_shift, qcoeff, dqcoeff,
+                           pd->dequant, eob, scan_order->scan,
+                           scan_order->iscan);
+      break;
+    case TX_16X16:
+      vpx_fdct16x16(src_diff, coeff, diff_stride);
+      vpx_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
+                     p->quant, p->quant_shift, qcoeff, dqcoeff,
+                     pd->dequant, eob,
+                     scan_order->scan, scan_order->iscan);
+      break;
+    case TX_8X8:
+      vpx_fdct8x8(src_diff, coeff, diff_stride);
+      vpx_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
+                     p->quant, p->quant_shift, qcoeff, dqcoeff,
+                     pd->dequant, eob,
+                     scan_order->scan, scan_order->iscan);
+      break;
+    case TX_4X4:
+      x->fwd_txm4x4(src_diff, coeff, diff_stride);
+      vpx_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
+                     p->quant, p->quant_shift, qcoeff, dqcoeff,
+                     pd->dequant, eob,
+                     scan_order->scan, scan_order->iscan);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
+
+static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
+                         TX_SIZE tx_size, void *arg) {
+  struct encode_b_args *const args = arg;
+  MACROBLOCK *const x = args->x;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct optimize_ctx *const ctx = args->ctx;
+  struct macroblock_plane *const p = &x->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  int i, j;
+  uint8_t *dst;
+  ENTROPY_CONTEXT *a, *l;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
+  dst = &pd->dst.buf[4 * j * pd->dst.stride + 4 * i];
+  a = &ctx->ta[plane][i];
+  l = &ctx->tl[plane][j];
+
+  // TODO(jingning): per transformed block zero forcing only enabled for
+  // luma component. will integrate chroma components as well.
+  if (x->zcoeff_blk[tx_size][block] && plane == 0) {
+    p->eobs[block] = 0;
+    *a = *l = 0;
+    return;
+  }
+
+  if (!x->skip_recode) {
+    if (x->quant_fp) {
+      // Encoding process for rtc mode
+      if (x->skip_txfm[0] == SKIP_TXFM_AC_DC && plane == 0) {
+        // skip forward transform
+        p->eobs[block] = 0;
+        *a = *l = 0;
+        return;
+      } else {
+        vp9_xform_quant_fp(x, plane, block, plane_bsize, tx_size);
+      }
+    } else {
+      if (max_txsize_lookup[plane_bsize] == tx_size) {
+        int txfm_blk_index = (plane << 2) + (block >> (tx_size << 1));
+        if (x->skip_txfm[txfm_blk_index] == SKIP_TXFM_NONE) {
+          // full forward transform and quantization
+          vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
+        } else if (x->skip_txfm[txfm_blk_index] == SKIP_TXFM_AC_ONLY) {
+          // fast path forward transform and quantization
+          vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
+        } else {
+          // skip forward transform
+          p->eobs[block] = 0;
+          *a = *l = 0;
+          return;
+        }
+      } else {
+        vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
+      }
+    }
+  }
+
+  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
+    const int ctx = combine_entropy_contexts(*a, *l);
+    *a = *l = optimize_b(x, plane, block, tx_size, ctx) > 0;
+  } else {
+    *a = *l = p->eobs[block] > 0;
+  }
+
+  if (p->eobs[block])
+    *(args->skip) = 0;
+
+  if (x->skip_encode || p->eobs[block] == 0)
+    return;
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    switch (tx_size) {
+      case TX_32X32:
+        vp9_highbd_idct32x32_add(dqcoeff, dst, pd->dst.stride,
+                                 p->eobs[block], xd->bd);
+        break;
+      case TX_16X16:
+        vp9_highbd_idct16x16_add(dqcoeff, dst, pd->dst.stride,
+                                 p->eobs[block], xd->bd);
+        break;
+      case TX_8X8:
+        vp9_highbd_idct8x8_add(dqcoeff, dst, pd->dst.stride,
+                               p->eobs[block], xd->bd);
+        break;
+      case TX_4X4:
+        // this is like vp9_short_idct4x4 but has a special case around eob<=1
+        // which is significant (not just an optimization) for the lossless
+        // case.
+        x->highbd_itxm_add(dqcoeff, dst, pd->dst.stride,
+                           p->eobs[block], xd->bd);
+        break;
+      default:
+        assert(0 && "Invalid transform size");
+    }
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  switch (tx_size) {
+    case TX_32X32:
+      vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
+      break;
+    case TX_16X16:
+      vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
+      break;
+    case TX_8X8:
+      vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
+      break;
+    case TX_4X4:
+      // this is like vp9_short_idct4x4 but has a special case around eob<=1
+      // which is significant (not just an optimization) for the lossless
+      // case.
+      x->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
+      break;
+    default:
+      assert(0 && "Invalid transform size");
+      break;
+  }
+}
+
+static void encode_block_pass1(int plane, int block, BLOCK_SIZE plane_bsize,
+                               TX_SIZE tx_size, void *arg) {
+  MACROBLOCK *const x = (MACROBLOCK *)arg;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *const p = &x->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  int i, j;
+  uint8_t *dst;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
+  dst = &pd->dst.buf[4 * j * pd->dst.stride + 4 * i];
+
+  vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
+
+  if (p->eobs[block] > 0) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+       x->highbd_itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block], xd->bd);
+       return;
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    x->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
+  }
+}
+
+void vp9_encode_sby_pass1(MACROBLOCK *x, BLOCK_SIZE bsize) {
+  vp9_subtract_plane(x, bsize, 0);
+  vp9_foreach_transformed_block_in_plane(&x->e_mbd, bsize, 0,
+                                         encode_block_pass1, x);
+}
+
+void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct optimize_ctx ctx;
+  MODE_INFO *mi = xd->mi[0];
+  struct encode_b_args arg = {x, &ctx, &mi->skip};
+  int plane;
+
+  mi->skip = 1;
+
+  if (x->skip)
+    return;
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    if (!x->skip_recode)
+      vp9_subtract_plane(x, bsize, plane);
+
+    if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
+      const struct macroblockd_plane* const pd = &xd->plane[plane];
+      const TX_SIZE tx_size = plane ? get_uv_tx_size(mi, pd) : mi->tx_size;
+      vp9_get_entropy_contexts(bsize, tx_size, pd,
+                               ctx.ta[plane], ctx.tl[plane]);
+    }
+
+    vp9_foreach_transformed_block_in_plane(xd, bsize, plane, encode_block,
+                                           &arg);
+  }
+}
+
+void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
+                            TX_SIZE tx_size, void *arg) {
+  struct encode_b_args* const args = arg;
+  MACROBLOCK *const x = args->x;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  struct macroblock_plane *const p = &x->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *coeff = BLOCK_OFFSET(p->coeff, block);
+  tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  const scan_order *scan_order;
+  TX_TYPE tx_type = DCT_DCT;
+  PREDICTION_MODE mode;
+  const int bwl = b_width_log2_lookup[plane_bsize];
+  const int diff_stride = 4 * (1 << bwl);
+  uint8_t *src, *dst;
+  int16_t *src_diff;
+  uint16_t *eob = &p->eobs[block];
+  const int src_stride = p->src.stride;
+  const int dst_stride = pd->dst.stride;
+  int i, j;
+  struct optimize_ctx *const ctx = args->ctx;
+  ENTROPY_CONTEXT *a = NULL;
+  ENTROPY_CONTEXT *l = NULL;
+  int entropy_ctx = 0;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
+  dst = &pd->dst.buf[4 * (j * dst_stride + i)];
+  src = &p->src.buf[4 * (j * src_stride + i)];
+  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
+  if (args->ctx != NULL) {
+    a = &ctx->ta[plane][i];
+    l = &ctx->tl[plane][j];
+    entropy_ctx = combine_entropy_contexts(*a, *l);
+  }
+
+  if (tx_size == TX_4X4) {
+    tx_type = get_tx_type_4x4(get_plane_type(plane), xd, block);
+    scan_order = &vp9_scan_orders[TX_4X4][tx_type];
+    mode = plane == 0 ? get_y_mode(xd->mi[0], block) : mi->uv_mode;
+  } else {
+    mode = plane == 0 ? mi->mode : mi->uv_mode;
+    if (tx_size == TX_32X32) {
+      scan_order = &vp9_default_scan_orders[TX_32X32];
+    } else {
+      tx_type = get_tx_type(get_plane_type(plane), xd);
+      scan_order = &vp9_scan_orders[tx_size][tx_type];
+    }
+  }
+
+  vp9_predict_intra_block(xd, bwl, tx_size, mode, x->skip_encode ? src : dst,
+                          x->skip_encode ? src_stride : dst_stride,
+                          dst, dst_stride, i, j, plane);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    switch (tx_size) {
+      case TX_32X32:
+        if (!x->skip_recode) {
+          vpx_highbd_subtract_block(32, 32, src_diff, diff_stride,
+                                    src, src_stride, dst, dst_stride, xd->bd);
+          highbd_fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
+          vpx_highbd_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin,
+                                      p->round, p->quant, p->quant_shift,
+                                      qcoeff, dqcoeff, pd->dequant, eob,
+                                      scan_order->scan, scan_order->iscan);
+        }
+        if (!x->skip_encode && *eob) {
+          vp9_highbd_idct32x32_add(dqcoeff, dst, dst_stride, *eob, xd->bd);
+        }
+        break;
+      case TX_16X16:
+        if (!x->skip_recode) {
+          vpx_highbd_subtract_block(16, 16, src_diff, diff_stride,
+                                    src, src_stride, dst, dst_stride, xd->bd);
+          if (tx_type == DCT_DCT)
+            vpx_highbd_fdct16x16(src_diff, coeff, diff_stride);
+          else
+            vp9_highbd_fht16x16(src_diff, coeff, diff_stride, tx_type);
+          vpx_highbd_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
+                                p->quant, p->quant_shift, qcoeff, dqcoeff,
+                                pd->dequant, eob,
+                                scan_order->scan, scan_order->iscan);
+        }
+        if (!x->skip_encode && *eob) {
+          vp9_highbd_iht16x16_add(tx_type, dqcoeff, dst, dst_stride,
+                                  *eob, xd->bd);
+        }
+        break;
+      case TX_8X8:
+        if (!x->skip_recode) {
+          vpx_highbd_subtract_block(8, 8, src_diff, diff_stride,
+                                    src, src_stride, dst, dst_stride, xd->bd);
+          if (tx_type == DCT_DCT)
+            vpx_highbd_fdct8x8(src_diff, coeff, diff_stride);
+          else
+            vp9_highbd_fht8x8(src_diff, coeff, diff_stride, tx_type);
+          vpx_highbd_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
+                                p->quant, p->quant_shift, qcoeff, dqcoeff,
+                                pd->dequant, eob,
+                                scan_order->scan, scan_order->iscan);
+        }
+        if (!x->skip_encode && *eob) {
+          vp9_highbd_iht8x8_add(tx_type, dqcoeff, dst, dst_stride, *eob,
+                                xd->bd);
+        }
+        break;
+      case TX_4X4:
+        if (!x->skip_recode) {
+          vpx_highbd_subtract_block(4, 4, src_diff, diff_stride,
+                                    src, src_stride, dst, dst_stride, xd->bd);
+          if (tx_type != DCT_DCT)
+            vp9_highbd_fht4x4(src_diff, coeff, diff_stride, tx_type);
+          else
+            x->fwd_txm4x4(src_diff, coeff, diff_stride);
+          vpx_highbd_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
+                                p->quant, p->quant_shift, qcoeff, dqcoeff,
+                                pd->dequant, eob,
+                                scan_order->scan, scan_order->iscan);
+        }
+
+        if (!x->skip_encode && *eob) {
+          if (tx_type == DCT_DCT) {
+            // this is like vp9_short_idct4x4 but has a special case around
+            // eob<=1 which is significant (not just an optimization) for the
+            // lossless case.
+            x->highbd_itxm_add(dqcoeff, dst, dst_stride, *eob, xd->bd);
+          } else {
+            vp9_highbd_iht4x4_16_add(dqcoeff, dst, dst_stride, tx_type, xd->bd);
+          }
+        }
+        break;
+      default:
+        assert(0);
+        return;
+    }
+    if (*eob)
+      *(args->skip) = 0;
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  switch (tx_size) {
+    case TX_32X32:
+      if (!x->skip_recode) {
+        vpx_subtract_block(32, 32, src_diff, diff_stride,
+                           src, src_stride, dst, dst_stride);
+        fdct32x32(x->use_lp32x32fdct, src_diff, coeff, diff_stride);
+        vpx_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
+                             p->quant, p->quant_shift, qcoeff, dqcoeff,
+                             pd->dequant, eob, scan_order->scan,
+                             scan_order->iscan);
+      }
+      if (args->ctx != NULL && !x->skip_recode) {
+       *a = *l = optimize_b(x, plane, block, tx_size, entropy_ctx) > 0;
+      }
+      if (!x->skip_encode && *eob)
+        vp9_idct32x32_add(dqcoeff, dst, dst_stride, *eob);
+      break;
+    case TX_16X16:
+      if (!x->skip_recode) {
+        vpx_subtract_block(16, 16, src_diff, diff_stride,
+                           src, src_stride, dst, dst_stride);
+        vp9_fht16x16(src_diff, coeff, diff_stride, tx_type);
+        vpx_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
+                       p->quant, p->quant_shift, qcoeff, dqcoeff,
+                       pd->dequant, eob, scan_order->scan,
+                       scan_order->iscan);
+      }
+      if (args->ctx != NULL && !x->skip_recode) {
+        *a = *l = optimize_b(x, plane, block, tx_size, entropy_ctx) > 0;
+      }
+      if (!x->skip_encode && *eob)
+        vp9_iht16x16_add(tx_type, dqcoeff, dst, dst_stride, *eob);
+      break;
+    case TX_8X8:
+      if (!x->skip_recode) {
+        vpx_subtract_block(8, 8, src_diff, diff_stride,
+                           src, src_stride, dst, dst_stride);
+        vp9_fht8x8(src_diff, coeff, diff_stride, tx_type);
+        vpx_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant,
+                       p->quant_shift, qcoeff, dqcoeff,
+                       pd->dequant, eob, scan_order->scan,
+                       scan_order->iscan);
+      }
+      if (args->ctx != NULL && !x->skip_recode) {
+        *a = *l = optimize_b(x, plane, block, tx_size, entropy_ctx) > 0;
+      }
+      if (!x->skip_encode && *eob)
+        vp9_iht8x8_add(tx_type, dqcoeff, dst, dst_stride, *eob);
+      break;
+    case TX_4X4:
+      if (!x->skip_recode) {
+        vpx_subtract_block(4, 4, src_diff, diff_stride,
+                           src, src_stride, dst, dst_stride);
+        if (tx_type != DCT_DCT)
+          vp9_fht4x4(src_diff, coeff, diff_stride, tx_type);
+        else
+          x->fwd_txm4x4(src_diff, coeff, diff_stride);
+        vpx_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant,
+                       p->quant_shift, qcoeff, dqcoeff,
+                       pd->dequant, eob, scan_order->scan,
+                       scan_order->iscan);
+      }
+      if (args->ctx != NULL && !x->skip_recode) {
+        *a = *l = optimize_b(x, plane, block, tx_size, entropy_ctx) > 0;
+      }
+      if (!x->skip_encode && *eob) {
+        if (tx_type == DCT_DCT)
+          // this is like vp9_short_idct4x4 but has a special case around eob<=1
+          // which is significant (not just an optimization) for the lossless
+          // case.
+          x->itxm_add(dqcoeff, dst, dst_stride, *eob);
+        else
+          vp9_iht4x4_16_add(dqcoeff, dst, dst_stride, tx_type);
+      }
+      break;
+    default:
+      assert(0);
+      break;
+  }
+  if (*eob)
+    *(args->skip) = 0;
+}
+
+void vp9_encode_intra_block_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane,
+                                  int enable_optimize_b) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  struct optimize_ctx ctx;
+  struct encode_b_args arg = {x, NULL, &xd->mi[0]->skip};
+
+  if (enable_optimize_b && x->optimize &&
+      (!x->skip_recode || !x->skip_optimize)) {
+    const struct macroblockd_plane* const pd = &xd->plane[plane];
+    const TX_SIZE tx_size = plane ? get_uv_tx_size(xd->mi[0], pd) :
+        xd->mi[0]->tx_size;
+    vp9_get_entropy_contexts(bsize, tx_size, pd, ctx.ta[plane], ctx.tl[plane]);
+    arg.ctx = &ctx;
+  }
+
+  vp9_foreach_transformed_block_in_plane(xd, bsize, plane,
+                                         vp9_encode_block_intra, &arg);
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_encodemb.h b/libvpx/libvpx/vp9/encoder/vp9_encodemb.h
new file mode 100644
index 0000000..25b0b23
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_encodemb.h
@@ -0,0 +1,47 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_ENCODEMB_H_
+#define VP9_ENCODER_VP9_ENCODEMB_H_
+
+#include "./vpx_config.h"
+#include "vp9/encoder/vp9_block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct encode_b_args {
+  MACROBLOCK *x;
+  struct optimize_ctx *ctx;
+  int8_t *skip;
+};
+void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize);
+void vp9_encode_sby_pass1(MACROBLOCK *x, BLOCK_SIZE bsize);
+void vp9_xform_quant_fp(MACROBLOCK *x, int plane, int block,
+                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+void vp9_xform_quant_dc(MACROBLOCK *x, int plane, int block,
+                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+void vp9_xform_quant(MACROBLOCK *x, int plane, int block,
+                     BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+
+void vp9_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+
+void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
+                            TX_SIZE tx_size, void *arg);
+
+void vp9_encode_intra_block_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane,
+                                  int enable_optimize_b);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_ENCODEMB_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_encodemv.c b/libvpx/libvpx/vp9/encoder/vp9_encodemv.c
new file mode 100644
index 0000000..71f27cc
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_encodemv.c
@@ -0,0 +1,282 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropymode.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_encodemv.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+
+static struct vp9_token mv_joint_encodings[MV_JOINTS];
+static struct vp9_token mv_class_encodings[MV_CLASSES];
+static struct vp9_token mv_fp_encodings[MV_FP_SIZE];
+static struct vp9_token mv_class0_encodings[CLASS0_SIZE];
+
+void vp9_entropy_mv_init(void) {
+  vp9_tokens_from_tree(mv_joint_encodings, vp9_mv_joint_tree);
+  vp9_tokens_from_tree(mv_class_encodings, vp9_mv_class_tree);
+  vp9_tokens_from_tree(mv_class0_encodings, vp9_mv_class0_tree);
+  vp9_tokens_from_tree(mv_fp_encodings, vp9_mv_fp_tree);
+}
+
+static void encode_mv_component(vpx_writer* w, int comp,
+                                const nmv_component* mvcomp, int usehp) {
+  int offset;
+  const int sign = comp < 0;
+  const int mag = sign ? -comp : comp;
+  const int mv_class = vp9_get_mv_class(mag - 1, &offset);
+  const int d = offset >> 3;                // int mv data
+  const int fr = (offset >> 1) & 3;         // fractional mv data
+  const int hp = offset & 1;                // high precision mv data
+
+  assert(comp != 0);
+
+  // Sign
+  vpx_write(w, sign, mvcomp->sign);
+
+  // Class
+  vp9_write_token(w, vp9_mv_class_tree, mvcomp->classes,
+                  &mv_class_encodings[mv_class]);
+
+  // Integer bits
+  if (mv_class == MV_CLASS_0) {
+    vp9_write_token(w, vp9_mv_class0_tree, mvcomp->class0,
+                    &mv_class0_encodings[d]);
+  } else {
+    int i;
+    const int n = mv_class + CLASS0_BITS - 1;  // number of bits
+    for (i = 0; i < n; ++i)
+      vpx_write(w, (d >> i) & 1, mvcomp->bits[i]);
+  }
+
+  // Fractional bits
+  vp9_write_token(w, vp9_mv_fp_tree,
+                  mv_class == MV_CLASS_0 ?  mvcomp->class0_fp[d] : mvcomp->fp,
+                  &mv_fp_encodings[fr]);
+
+  // High precision bit
+  if (usehp)
+    vpx_write(w, hp,
+              mv_class == MV_CLASS_0 ? mvcomp->class0_hp : mvcomp->hp);
+}
+
+
+static void build_nmv_component_cost_table(int *mvcost,
+                                           const nmv_component* const mvcomp,
+                                           int usehp) {
+  int sign_cost[2], class_cost[MV_CLASSES], class0_cost[CLASS0_SIZE];
+  int bits_cost[MV_OFFSET_BITS][2];
+  int class0_fp_cost[CLASS0_SIZE][MV_FP_SIZE], fp_cost[MV_FP_SIZE];
+  int class0_hp_cost[2], hp_cost[2];
+  int i;
+  int c, o;
+
+  sign_cost[0] = vp9_cost_zero(mvcomp->sign);
+  sign_cost[1] = vp9_cost_one(mvcomp->sign);
+  vp9_cost_tokens(class_cost, mvcomp->classes, vp9_mv_class_tree);
+  vp9_cost_tokens(class0_cost, mvcomp->class0, vp9_mv_class0_tree);
+  for (i = 0; i < MV_OFFSET_BITS; ++i) {
+    bits_cost[i][0] = vp9_cost_zero(mvcomp->bits[i]);
+    bits_cost[i][1] = vp9_cost_one(mvcomp->bits[i]);
+  }
+
+  for (i = 0; i < CLASS0_SIZE; ++i)
+    vp9_cost_tokens(class0_fp_cost[i], mvcomp->class0_fp[i], vp9_mv_fp_tree);
+  vp9_cost_tokens(fp_cost, mvcomp->fp, vp9_mv_fp_tree);
+
+  // Always build the hp costs to avoid an uninitialized warning from gcc
+  class0_hp_cost[0] = vp9_cost_zero(mvcomp->class0_hp);
+  class0_hp_cost[1] = vp9_cost_one(mvcomp->class0_hp);
+  hp_cost[0] = vp9_cost_zero(mvcomp->hp);
+  hp_cost[1] = vp9_cost_one(mvcomp->hp);
+
+  mvcost[0] = 0;
+  // MV_CLASS_0
+  for (o = 0; o < (CLASS0_SIZE << 3); ++o) {
+    int d, e, f;
+    int cost = class_cost[MV_CLASS_0];
+    int v = o + 1;
+    d = (o >> 3);               /* int mv data */
+    f = (o >> 1) & 3;           /* fractional pel mv data */
+    cost += class0_cost[d];
+    cost += class0_fp_cost[d][f];
+    if (usehp) {
+      e = (o & 1);                /* high precision mv data */
+      cost += class0_hp_cost[e];
+    }
+    mvcost[v] = cost + sign_cost[0];
+    mvcost[-v] = cost + sign_cost[1];
+  }
+  for (c = MV_CLASS_1; c < MV_CLASSES; ++c) {
+    int d;
+    for (d = 0; d < (1 << c); ++d) {
+      int f;
+      int whole_cost = class_cost[c];
+      int b = c + CLASS0_BITS - 1;  /* number of bits */
+      for (i = 0; i < b; ++i)
+        whole_cost += bits_cost[i][((d >> i) & 1)];
+      for (f = 0; f < 4; ++f) {
+        int cost = whole_cost + fp_cost[f];
+        int v = (CLASS0_SIZE << (c + 2)) + d * 8 + f * 2 /* + e */ + 1;
+        if (usehp) {
+          mvcost[v] = cost + hp_cost[0] + sign_cost[0];
+          mvcost[-v] = cost + hp_cost[0] + sign_cost[1];
+          if (v + 1 > MV_MAX) break;
+          mvcost[v + 1] = cost + hp_cost[1] + sign_cost[0];
+          mvcost[-v - 1] = cost + hp_cost[1] + sign_cost[1];
+        } else {
+          mvcost[v] = cost + sign_cost[0];
+          mvcost[-v] = cost + sign_cost[1];
+          if (v + 1 > MV_MAX) break;
+          mvcost[v + 1] = cost + sign_cost[0];
+          mvcost[-v - 1] = cost + sign_cost[1];
+        }
+      }
+    }
+  }
+}
+
+static int update_mv(vpx_writer *w, const unsigned int ct[2], vpx_prob *cur_p,
+                     vpx_prob upd_p) {
+  const vpx_prob new_p = get_binary_prob(ct[0], ct[1]) | 1;
+  const int update = cost_branch256(ct, *cur_p) + vp9_cost_zero(upd_p) >
+                     cost_branch256(ct, new_p) + vp9_cost_one(upd_p) +
+                         (7 << VP9_PROB_COST_SHIFT);
+  vpx_write(w, update, upd_p);
+  if (update) {
+    *cur_p = new_p;
+    vpx_write_literal(w, new_p >> 1, 7);
+  }
+  return update;
+}
+
+static void write_mv_update(const vpx_tree_index *tree,
+                            vpx_prob probs[/*n - 1*/],
+                            const unsigned int counts[/*n - 1*/],
+                            int n, vpx_writer *w) {
+  int i;
+  unsigned int branch_ct[32][2];
+
+  // Assuming max number of probabilities <= 32
+  assert(n <= 32);
+
+  vp9_tree_probs_from_distribution(tree, branch_ct, counts);
+  for (i = 0; i < n - 1; ++i)
+    update_mv(w, branch_ct[i], &probs[i], MV_UPDATE_PROB);
+}
+
+void vp9_write_nmv_probs(VP9_COMMON *cm, int usehp, vpx_writer *w,
+                         nmv_context_counts *const counts) {
+  int i, j;
+  nmv_context *const mvc = &cm->fc->nmvc;
+
+  write_mv_update(vp9_mv_joint_tree, mvc->joints, counts->joints, MV_JOINTS, w);
+
+  for (i = 0; i < 2; ++i) {
+    nmv_component *comp = &mvc->comps[i];
+    nmv_component_counts *comp_counts = &counts->comps[i];
+
+    update_mv(w, comp_counts->sign, &comp->sign, MV_UPDATE_PROB);
+    write_mv_update(vp9_mv_class_tree, comp->classes, comp_counts->classes,
+                    MV_CLASSES, w);
+    write_mv_update(vp9_mv_class0_tree, comp->class0, comp_counts->class0,
+                    CLASS0_SIZE, w);
+    for (j = 0; j < MV_OFFSET_BITS; ++j)
+      update_mv(w, comp_counts->bits[j], &comp->bits[j], MV_UPDATE_PROB);
+  }
+
+  for (i = 0; i < 2; ++i) {
+    for (j = 0; j < CLASS0_SIZE; ++j)
+      write_mv_update(vp9_mv_fp_tree, mvc->comps[i].class0_fp[j],
+                      counts->comps[i].class0_fp[j], MV_FP_SIZE, w);
+
+    write_mv_update(vp9_mv_fp_tree, mvc->comps[i].fp, counts->comps[i].fp,
+                    MV_FP_SIZE, w);
+  }
+
+  if (usehp) {
+    for (i = 0; i < 2; ++i) {
+      update_mv(w, counts->comps[i].class0_hp, &mvc->comps[i].class0_hp,
+                MV_UPDATE_PROB);
+      update_mv(w, counts->comps[i].hp, &mvc->comps[i].hp, MV_UPDATE_PROB);
+    }
+  }
+}
+
+void vp9_encode_mv(VP9_COMP* cpi, vpx_writer* w,
+                   const MV* mv, const MV* ref,
+                   const nmv_context* mvctx, int usehp) {
+  const MV diff = {mv->row - ref->row,
+                   mv->col - ref->col};
+  const MV_JOINT_TYPE j = vp9_get_mv_joint(&diff);
+  usehp = usehp && use_mv_hp(ref);
+
+  vp9_write_token(w, vp9_mv_joint_tree, mvctx->joints, &mv_joint_encodings[j]);
+  if (mv_joint_vertical(j))
+    encode_mv_component(w, diff.row, &mvctx->comps[0], usehp);
+
+  if (mv_joint_horizontal(j))
+    encode_mv_component(w, diff.col, &mvctx->comps[1], usehp);
+
+  // If auto_mv_step_size is enabled then keep track of the largest
+  // motion vector component used.
+  if (cpi->sf.mv.auto_mv_step_size) {
+    unsigned int maxv = VPXMAX(abs(mv->row), abs(mv->col)) >> 3;
+    cpi->max_mv_magnitude = VPXMAX(maxv, cpi->max_mv_magnitude);
+  }
+}
+
+void vp9_build_nmv_cost_table(int *mvjoint, int *mvcost[2],
+                              const nmv_context* ctx, int usehp) {
+  vp9_cost_tokens(mvjoint, ctx->joints, vp9_mv_joint_tree);
+  build_nmv_component_cost_table(mvcost[0], &ctx->comps[0], usehp);
+  build_nmv_component_cost_table(mvcost[1], &ctx->comps[1], usehp);
+}
+
+static void inc_mvs(const MODE_INFO *mi, const MB_MODE_INFO_EXT *mbmi_ext,
+                    const int_mv mvs[2],
+                    nmv_context_counts *counts) {
+  int i;
+
+  for (i = 0; i < 1 + has_second_ref(mi); ++i) {
+    const MV *ref = &mbmi_ext->ref_mvs[mi->ref_frame[i]][0].as_mv;
+    const MV diff = {mvs[i].as_mv.row - ref->row,
+                     mvs[i].as_mv.col - ref->col};
+    vp9_inc_mv(&diff, counts);
+  }
+}
+
+void vp9_update_mv_count(ThreadData *td) {
+  const MACROBLOCKD *xd = &td->mb.e_mbd;
+  const MODE_INFO *mi = xd->mi[0];
+  const MB_MODE_INFO_EXT *mbmi_ext = td->mb.mbmi_ext;
+
+  if (mi->sb_type < BLOCK_8X8) {
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[mi->sb_type];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[mi->sb_type];
+    int idx, idy;
+
+    for (idy = 0; idy < 2; idy += num_4x4_h) {
+      for (idx = 0; idx < 2; idx += num_4x4_w) {
+        const int i = idy * 2 + idx;
+        if (mi->bmi[i].as_mode == NEWMV)
+          inc_mvs(mi, mbmi_ext, mi->bmi[i].as_mv, &td->counts->mv);
+      }
+    }
+  } else {
+    if (mi->mode == NEWMV)
+      inc_mvs(mi, mbmi_ext, mi->mv, &td->counts->mv);
+  }
+}
+
diff --git a/libvpx/libvpx/vp9/encoder/vp9_encodemv.h b/libvpx/libvpx/vp9/encoder/vp9_encodemv.h
new file mode 100644
index 0000000..5fb114c
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_encodemv.h
@@ -0,0 +1,38 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_ENCODEMV_H_
+#define VP9_ENCODER_VP9_ENCODEMV_H_
+
+#include "vp9/encoder/vp9_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_entropy_mv_init(void);
+
+void vp9_write_nmv_probs(VP9_COMMON *cm, int usehp, vpx_writer *w,
+                         nmv_context_counts *const counts);
+
+void vp9_encode_mv(VP9_COMP *cpi, vpx_writer* w, const MV* mv, const MV* ref,
+                   const nmv_context* mvctx, int usehp);
+
+void vp9_build_nmv_cost_table(int *mvjoint, int *mvcost[2],
+                              const nmv_context* mvctx, int usehp);
+
+void vp9_update_mv_count(ThreadData *td);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_ENCODEMV_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_encoder.c b/libvpx/libvpx/vp9/encoder/vp9_encoder.c
new file mode 100644
index 0000000..147f970
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_encoder.c
@@ -0,0 +1,5229 @@
+/*
+ * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdio.h>
+#include <limits.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+#include "vpx/internal/vpx_psnr.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/vpx_filter.h"
+#if CONFIG_INTERNAL_STATS
+#include "vpx_dsp/ssim.h"
+#endif
+#include "vpx_ports/mem.h"
+#include "vpx_ports/system_state.h"
+#include "vpx_ports/vpx_timer.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_filter.h"
+#include "vp9/common/vp9_idct.h"
+#if CONFIG_VP9_POSTPROC
+#include "vp9/common/vp9_postproc.h"
+#endif
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_tile_common.h"
+
+#include "vp9/encoder/vp9_aq_360.h"
+#include "vp9/encoder/vp9_aq_complexity.h"
+#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
+#include "vp9/encoder/vp9_aq_variance.h"
+#include "vp9/encoder/vp9_bitstream.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_encodeframe.h"
+#include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_ethread.h"
+#include "vp9/encoder/vp9_firstpass.h"
+#include "vp9/encoder/vp9_mbgraph.h"
+#include "vp9/encoder/vp9_noise_estimate.h"
+#include "vp9/encoder/vp9_picklpf.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_resize.h"
+#include "vp9/encoder/vp9_segmentation.h"
+#include "vp9/encoder/vp9_skin_detection.h"
+#include "vp9/encoder/vp9_speed_features.h"
+#include "vp9/encoder/vp9_svc_layercontext.h"
+#include "vp9/encoder/vp9_temporal_filter.h"
+
+#define AM_SEGMENT_ID_INACTIVE 7
+#define AM_SEGMENT_ID_ACTIVE 0
+
+#define ALTREF_HIGH_PRECISION_MV 1      // Whether to use high precision mv
+                                         //  for altref computation.
+#define HIGH_PRECISION_MV_QTHRESH 200   // Q threshold for high precision
+                                         // mv. Choose a very high value for
+                                         // now so that HIGH_PRECISION is always
+                                         // chosen.
+// #define OUTPUT_YUV_REC
+
+#ifdef OUTPUT_YUV_DENOISED
+FILE *yuv_denoised_file = NULL;
+#endif
+#ifdef OUTPUT_YUV_SKINMAP
+FILE *yuv_skinmap_file = NULL;
+#endif
+#ifdef OUTPUT_YUV_REC
+FILE *yuv_rec_file;
+#endif
+
+#if 0
+FILE *framepsnr;
+FILE *kf_list;
+FILE *keyfile;
+#endif
+
+static const Vp9LevelSpec vp9_level_defs[VP9_LEVELS] = {
+  {LEVEL_1,   829440,      36864,    200,    400,   2, 1,  4,  8},
+  {LEVEL_1_1, 2764800,     73728,    800,    1000,  2, 1,  4,  8},
+  {LEVEL_2,   4608000,     122880,   1800,   1500,  2, 1,  4,  8},
+  {LEVEL_2_1, 9216000,     245760,   3600,   2800,  2, 2,  4,  8},
+  {LEVEL_3,   20736000,    552960,   7200,   6000,  2, 4,  4,  8},
+  {LEVEL_3_1, 36864000,    983040,   12000,  10000, 2, 4,  4,  8},
+  {LEVEL_4,   83558400,    2228224,  18000,  16000, 4, 4,  4,  8},
+  {LEVEL_4_1, 160432128,   2228224,  30000,  18000, 4, 4,  5,  6},
+  {LEVEL_5,   311951360,   8912896,  60000,  36000, 6, 8,  6,  4},
+  {LEVEL_5_1, 588251136,   8912896,  120000, 46000, 8, 8,  10, 4},
+  // TODO(huisu): update max_cpb_size for level 5_2 ~ 6_2 when
+  // they are finalized (currently TBD).
+  {LEVEL_5_2, 1176502272,  8912896,  180000, 0,     8, 8,  10, 4},
+  {LEVEL_6,   1176502272,  35651584, 180000, 0,     8, 16, 10, 4},
+  {LEVEL_6_1, 2353004544u, 35651584, 240000, 0,     8, 16, 10, 4},
+  {LEVEL_6_2, 4706009088u, 35651584, 480000, 0,     8, 16, 10, 4},
+};
+
+static INLINE void Scale2Ratio(VPX_SCALING mode, int *hr, int *hs) {
+  switch (mode) {
+    case NORMAL:
+      *hr = 1;
+      *hs = 1;
+      break;
+    case FOURFIVE:
+      *hr = 4;
+      *hs = 5;
+      break;
+    case THREEFIVE:
+      *hr = 3;
+      *hs = 5;
+    break;
+    case ONETWO:
+      *hr = 1;
+      *hs = 2;
+    break;
+    default:
+      *hr = 1;
+      *hs = 1;
+       assert(0);
+      break;
+  }
+}
+
+// Mark all inactive blocks as active. Other segmentation features may be set
+// so memset cannot be used, instead only inactive blocks should be reset.
+static void suppress_active_map(VP9_COMP *cpi) {
+  unsigned char *const seg_map = cpi->segmentation_map;
+
+  if (cpi->active_map.enabled || cpi->active_map.update) {
+    const int rows = cpi->common.mi_rows;
+    const int cols = cpi->common.mi_cols;
+    int i;
+
+    for (i = 0; i < rows * cols; ++i)
+      if (seg_map[i] == AM_SEGMENT_ID_INACTIVE)
+        seg_map[i] = AM_SEGMENT_ID_ACTIVE;
+  }
+}
+
+static void apply_active_map(VP9_COMP *cpi) {
+  struct segmentation *const seg = &cpi->common.seg;
+  unsigned char *const seg_map = cpi->segmentation_map;
+  const unsigned char *const active_map = cpi->active_map.map;
+  int i;
+
+  assert(AM_SEGMENT_ID_ACTIVE == CR_SEGMENT_ID_BASE);
+
+  if (frame_is_intra_only(&cpi->common)) {
+    cpi->active_map.enabled = 0;
+    cpi->active_map.update = 1;
+  }
+
+  if (cpi->active_map.update) {
+    if (cpi->active_map.enabled) {
+      for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i)
+        if (seg_map[i] == AM_SEGMENT_ID_ACTIVE) seg_map[i] = active_map[i];
+      vp9_enable_segmentation(seg);
+      vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
+      vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
+      // Setting the data to -MAX_LOOP_FILTER will result in the computed loop
+      // filter level being zero regardless of the value of seg->abs_delta.
+      vp9_set_segdata(seg, AM_SEGMENT_ID_INACTIVE,
+                      SEG_LVL_ALT_LF, -MAX_LOOP_FILTER);
+    } else {
+      vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
+      vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
+      if (seg->enabled) {
+        seg->update_data = 1;
+        seg->update_map = 1;
+      }
+    }
+    cpi->active_map.update = 0;
+  }
+}
+
+static void init_level_info(Vp9LevelInfo *level_info) {
+  Vp9LevelStats *const level_stats = &level_info->level_stats;
+  Vp9LevelSpec *const level_spec = &level_info->level_spec;
+
+  memset(level_stats, 0, sizeof(*level_stats));
+  memset(level_spec, 0, sizeof(*level_spec));
+  level_spec->level = LEVEL_UNKNOWN;
+  level_spec->min_altref_distance = INT_MAX;
+}
+
+VP9_LEVEL vp9_get_level(const Vp9LevelSpec * const level_spec) {
+  int i;
+  const Vp9LevelSpec *this_level;
+
+  vpx_clear_system_state();
+
+  for (i = 0; i < VP9_LEVELS; ++i) {
+    this_level = &vp9_level_defs[i];
+    if ((double)level_spec->max_luma_sample_rate * (1 + SAMPLE_RATE_GRACE_P) >
+        (double)this_level->max_luma_sample_rate ||
+        level_spec->max_luma_picture_size > this_level->max_luma_picture_size ||
+        level_spec->average_bitrate > this_level->average_bitrate ||
+        level_spec->max_cpb_size > this_level->max_cpb_size ||
+        level_spec->compression_ratio < this_level->compression_ratio ||
+        level_spec->max_col_tiles > this_level->max_col_tiles ||
+        level_spec->min_altref_distance < this_level->min_altref_distance ||
+        level_spec->max_ref_frame_buffers > this_level->max_ref_frame_buffers)
+      continue;
+    break;
+  }
+  return (i == VP9_LEVELS) ? LEVEL_UNKNOWN : vp9_level_defs[i].level;
+}
+
+int vp9_set_active_map(VP9_COMP* cpi,
+                       unsigned char* new_map_16x16,
+                       int rows,
+                       int cols) {
+  if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols) {
+    unsigned char *const active_map_8x8 = cpi->active_map.map;
+    const int mi_rows = cpi->common.mi_rows;
+    const int mi_cols = cpi->common.mi_cols;
+    cpi->active_map.update = 1;
+    if (new_map_16x16) {
+      int r, c;
+      for (r = 0; r < mi_rows; ++r) {
+        for (c = 0; c < mi_cols; ++c) {
+          active_map_8x8[r * mi_cols + c] =
+              new_map_16x16[(r >> 1) * cols + (c >> 1)]
+                  ? AM_SEGMENT_ID_ACTIVE
+                  : AM_SEGMENT_ID_INACTIVE;
+        }
+      }
+      cpi->active_map.enabled = 1;
+    } else {
+      cpi->active_map.enabled = 0;
+    }
+    return 0;
+  } else {
+    return -1;
+  }
+}
+
+int vp9_get_active_map(VP9_COMP* cpi,
+                       unsigned char* new_map_16x16,
+                       int rows,
+                       int cols) {
+  if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols &&
+      new_map_16x16) {
+    unsigned char* const seg_map_8x8 = cpi->segmentation_map;
+    const int mi_rows = cpi->common.mi_rows;
+    const int mi_cols = cpi->common.mi_cols;
+    memset(new_map_16x16, !cpi->active_map.enabled, rows * cols);
+    if (cpi->active_map.enabled) {
+      int r, c;
+      for (r = 0; r < mi_rows; ++r) {
+        for (c = 0; c < mi_cols; ++c) {
+          // Cyclic refresh segments are considered active despite not having
+          // AM_SEGMENT_ID_ACTIVE
+          new_map_16x16[(r >> 1) * cols + (c >> 1)] |=
+              seg_map_8x8[r * mi_cols + c] != AM_SEGMENT_ID_INACTIVE;
+        }
+      }
+    }
+    return 0;
+  } else {
+    return -1;
+  }
+}
+
+void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv) {
+  MACROBLOCK *const mb = &cpi->td.mb;
+  cpi->common.allow_high_precision_mv = allow_high_precision_mv;
+  if (cpi->common.allow_high_precision_mv) {
+    mb->mvcost = mb->nmvcost_hp;
+    mb->mvsadcost = mb->nmvsadcost_hp;
+  } else {
+    mb->mvcost = mb->nmvcost;
+    mb->mvsadcost = mb->nmvsadcost;
+  }
+}
+
+static void setup_frame(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  // Set up entropy context depending on frame type. The decoder mandates
+  // the use of the default context, index 0, for keyframes and inter
+  // frames where the error_resilient_mode or intra_only flag is set. For
+  // other inter-frames the encoder currently uses only two contexts;
+  // context 1 for ALTREF frames and context 0 for the others.
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
+    vp9_setup_past_independence(cm);
+  } else {
+    if (!cpi->use_svc)
+      cm->frame_context_idx = cpi->refresh_alt_ref_frame;
+  }
+
+  if (cm->frame_type == KEY_FRAME) {
+    if (!is_two_pass_svc(cpi))
+      cpi->refresh_golden_frame = 1;
+    cpi->refresh_alt_ref_frame = 1;
+    vp9_zero(cpi->interp_filter_selected);
+  } else {
+    *cm->fc = cm->frame_contexts[cm->frame_context_idx];
+    vp9_zero(cpi->interp_filter_selected[0]);
+  }
+}
+
+static void vp9_enc_setup_mi(VP9_COMMON *cm) {
+  int i;
+  cm->mi = cm->mip + cm->mi_stride + 1;
+  memset(cm->mip, 0, cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mip));
+  cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
+  // Clear top border row
+  memset(cm->prev_mip, 0, sizeof(*cm->prev_mip) * cm->mi_stride);
+  // Clear left border column
+  for (i = 1; i < cm->mi_rows + 1; ++i)
+    memset(&cm->prev_mip[i * cm->mi_stride], 0, sizeof(*cm->prev_mip));
+
+  cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
+  cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
+
+  memset(cm->mi_grid_base, 0,
+         cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
+}
+
+static int vp9_enc_alloc_mi(VP9_COMMON *cm, int mi_size) {
+  cm->mip = vpx_calloc(mi_size, sizeof(*cm->mip));
+  if (!cm->mip)
+    return 1;
+  cm->prev_mip = vpx_calloc(mi_size, sizeof(*cm->prev_mip));
+  if (!cm->prev_mip)
+    return 1;
+  cm->mi_alloc_size = mi_size;
+
+  cm->mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
+  if (!cm->mi_grid_base)
+    return 1;
+  cm->prev_mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
+  if (!cm->prev_mi_grid_base)
+    return 1;
+
+  return 0;
+}
+
+static void vp9_enc_free_mi(VP9_COMMON *cm) {
+  vpx_free(cm->mip);
+  cm->mip = NULL;
+  vpx_free(cm->prev_mip);
+  cm->prev_mip = NULL;
+  vpx_free(cm->mi_grid_base);
+  cm->mi_grid_base = NULL;
+  vpx_free(cm->prev_mi_grid_base);
+  cm->prev_mi_grid_base = NULL;
+}
+
+static void vp9_swap_mi_and_prev_mi(VP9_COMMON *cm) {
+  // Current mip will be the prev_mip for the next frame.
+  MODE_INFO **temp_base = cm->prev_mi_grid_base;
+  MODE_INFO *temp = cm->prev_mip;
+  cm->prev_mip = cm->mip;
+  cm->mip = temp;
+
+  // Update the upper left visible macroblock ptrs.
+  cm->mi = cm->mip + cm->mi_stride + 1;
+  cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
+
+  cm->prev_mi_grid_base = cm->mi_grid_base;
+  cm->mi_grid_base = temp_base;
+  cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
+  cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
+}
+
+void vp9_initialize_enc(void) {
+  static volatile int init_done = 0;
+
+  if (!init_done) {
+    vp9_rtcd();
+    vpx_dsp_rtcd();
+    vpx_scale_rtcd();
+    vp9_init_intra_predictors();
+    vp9_init_me_luts();
+    vp9_rc_init_minq_luts();
+    vp9_entropy_mv_init();
+    vp9_temporal_filter_init();
+    init_done = 1;
+  }
+}
+
+static void dealloc_compressor_data(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  int i;
+
+  vpx_free(cpi->mbmi_ext_base);
+  cpi->mbmi_ext_base = NULL;
+
+  vpx_free(cpi->tile_data);
+  cpi->tile_data = NULL;
+
+  // Delete sementation map
+  vpx_free(cpi->segmentation_map);
+  cpi->segmentation_map = NULL;
+  vpx_free(cpi->coding_context.last_frame_seg_map_copy);
+  cpi->coding_context.last_frame_seg_map_copy = NULL;
+
+  vpx_free(cpi->nmvcosts[0]);
+  vpx_free(cpi->nmvcosts[1]);
+  cpi->nmvcosts[0] = NULL;
+  cpi->nmvcosts[1] = NULL;
+
+  vpx_free(cpi->nmvcosts_hp[0]);
+  vpx_free(cpi->nmvcosts_hp[1]);
+  cpi->nmvcosts_hp[0] = NULL;
+  cpi->nmvcosts_hp[1] = NULL;
+
+  vpx_free(cpi->nmvsadcosts[0]);
+  vpx_free(cpi->nmvsadcosts[1]);
+  cpi->nmvsadcosts[0] = NULL;
+  cpi->nmvsadcosts[1] = NULL;
+
+  vpx_free(cpi->nmvsadcosts_hp[0]);
+  vpx_free(cpi->nmvsadcosts_hp[1]);
+  cpi->nmvsadcosts_hp[0] = NULL;
+  cpi->nmvsadcosts_hp[1] = NULL;
+
+  vp9_cyclic_refresh_free(cpi->cyclic_refresh);
+  cpi->cyclic_refresh = NULL;
+
+  vpx_free(cpi->active_map.map);
+  cpi->active_map.map = NULL;
+
+  vpx_free(cpi->consec_zero_mv);
+  cpi->consec_zero_mv = NULL;
+
+  vp9_free_ref_frame_buffers(cm->buffer_pool);
+#if CONFIG_VP9_POSTPROC
+  vp9_free_postproc_buffers(cm);
+#endif
+  vp9_free_context_buffers(cm);
+
+  vpx_free_frame_buffer(&cpi->last_frame_uf);
+  vpx_free_frame_buffer(&cpi->scaled_source);
+  vpx_free_frame_buffer(&cpi->scaled_last_source);
+  vpx_free_frame_buffer(&cpi->alt_ref_buffer);
+  vp9_lookahead_destroy(cpi->lookahead);
+
+  vpx_free(cpi->tile_tok[0][0]);
+  cpi->tile_tok[0][0] = 0;
+
+  vp9_free_pc_tree(&cpi->td);
+
+  for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
+    LAYER_CONTEXT *const lc = &cpi->svc.layer_context[i];
+    vpx_free(lc->rc_twopass_stats_in.buf);
+    lc->rc_twopass_stats_in.buf = NULL;
+    lc->rc_twopass_stats_in.sz = 0;
+  }
+
+  if (cpi->source_diff_var != NULL) {
+    vpx_free(cpi->source_diff_var);
+    cpi->source_diff_var = NULL;
+  }
+
+  for (i = 0; i < MAX_LAG_BUFFERS; ++i) {
+    vpx_free_frame_buffer(&cpi->svc.scaled_frames[i]);
+  }
+  memset(&cpi->svc.scaled_frames[0], 0,
+         MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0]));
+
+  vpx_free_frame_buffer(&cpi->svc.scaled_temp);
+  memset(&cpi->svc.scaled_temp, 0, sizeof(cpi->svc.scaled_temp));
+
+  vpx_free_frame_buffer(&cpi->svc.empty_frame.img);
+  memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame));
+
+  vp9_free_svc_cyclic_refresh(cpi);
+}
+
+static void save_coding_context(VP9_COMP *cpi) {
+  CODING_CONTEXT *const cc = &cpi->coding_context;
+  VP9_COMMON *cm = &cpi->common;
+
+  // Stores a snapshot of key state variables which can subsequently be
+  // restored with a call to vp9_restore_coding_context. These functions are
+  // intended for use in a re-code loop in vp9_compress_frame where the
+  // quantizer value is adjusted between loop iterations.
+  vp9_copy(cc->nmvjointcost,  cpi->td.mb.nmvjointcost);
+
+  memcpy(cc->nmvcosts[0], cpi->nmvcosts[0],
+         MV_VALS * sizeof(*cpi->nmvcosts[0]));
+  memcpy(cc->nmvcosts[1], cpi->nmvcosts[1],
+         MV_VALS * sizeof(*cpi->nmvcosts[1]));
+  memcpy(cc->nmvcosts_hp[0], cpi->nmvcosts_hp[0],
+         MV_VALS * sizeof(*cpi->nmvcosts_hp[0]));
+  memcpy(cc->nmvcosts_hp[1], cpi->nmvcosts_hp[1],
+         MV_VALS * sizeof(*cpi->nmvcosts_hp[1]));
+
+  vp9_copy(cc->segment_pred_probs, cm->seg.pred_probs);
+
+  memcpy(cpi->coding_context.last_frame_seg_map_copy,
+         cm->last_frame_seg_map, (cm->mi_rows * cm->mi_cols));
+
+  vp9_copy(cc->last_ref_lf_deltas, cm->lf.last_ref_deltas);
+  vp9_copy(cc->last_mode_lf_deltas, cm->lf.last_mode_deltas);
+
+  cc->fc = *cm->fc;
+}
+
+static void restore_coding_context(VP9_COMP *cpi) {
+  CODING_CONTEXT *const cc = &cpi->coding_context;
+  VP9_COMMON *cm = &cpi->common;
+
+  // Restore key state variables to the snapshot state stored in the
+  // previous call to vp9_save_coding_context.
+  vp9_copy(cpi->td.mb.nmvjointcost, cc->nmvjointcost);
+
+  memcpy(cpi->nmvcosts[0], cc->nmvcosts[0], MV_VALS * sizeof(*cc->nmvcosts[0]));
+  memcpy(cpi->nmvcosts[1], cc->nmvcosts[1], MV_VALS * sizeof(*cc->nmvcosts[1]));
+  memcpy(cpi->nmvcosts_hp[0], cc->nmvcosts_hp[0],
+         MV_VALS * sizeof(*cc->nmvcosts_hp[0]));
+  memcpy(cpi->nmvcosts_hp[1], cc->nmvcosts_hp[1],
+         MV_VALS * sizeof(*cc->nmvcosts_hp[1]));
+
+  vp9_copy(cm->seg.pred_probs, cc->segment_pred_probs);
+
+  memcpy(cm->last_frame_seg_map,
+         cpi->coding_context.last_frame_seg_map_copy,
+         (cm->mi_rows * cm->mi_cols));
+
+  vp9_copy(cm->lf.last_ref_deltas, cc->last_ref_lf_deltas);
+  vp9_copy(cm->lf.last_mode_deltas, cc->last_mode_lf_deltas);
+
+  *cm->fc = cc->fc;
+}
+
+static void configure_static_seg_features(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  struct segmentation *const seg = &cm->seg;
+
+  int high_q = (int)(rc->avg_q > 48.0);
+  int qi_delta;
+
+  // Disable and clear down for KF
+  if (cm->frame_type == KEY_FRAME) {
+    // Clear down the global segmentation map
+    memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
+    seg->update_map = 0;
+    seg->update_data = 0;
+    cpi->static_mb_pct = 0;
+
+    // Disable segmentation
+    vp9_disable_segmentation(seg);
+
+    // Clear down the segment features.
+    vp9_clearall_segfeatures(seg);
+  } else if (cpi->refresh_alt_ref_frame) {
+    // If this is an alt ref frame
+    // Clear down the global segmentation map
+    memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
+    seg->update_map = 0;
+    seg->update_data = 0;
+    cpi->static_mb_pct = 0;
+
+    // Disable segmentation and individual segment features by default
+    vp9_disable_segmentation(seg);
+    vp9_clearall_segfeatures(seg);
+
+    // Scan frames from current to arf frame.
+    // This function re-enables segmentation if appropriate.
+    vp9_update_mbgraph_stats(cpi);
+
+    // If segmentation was enabled set those features needed for the
+    // arf itself.
+    if (seg->enabled) {
+      seg->update_map = 1;
+      seg->update_data = 1;
+
+      qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 0.875,
+                                    cm->bit_depth);
+      vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta - 2);
+      vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
+
+      vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
+      vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
+
+      // Where relevant assume segment data is delta data
+      seg->abs_delta = SEGMENT_DELTADATA;
+    }
+  } else if (seg->enabled) {
+    // All other frames if segmentation has been enabled
+
+    // First normal frame in a valid gf or alt ref group
+    if (rc->frames_since_golden == 0) {
+      // Set up segment features for normal frames in an arf group
+      if (rc->source_alt_ref_active) {
+        seg->update_map = 0;
+        seg->update_data = 1;
+        seg->abs_delta = SEGMENT_DELTADATA;
+
+        qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 1.125,
+                                      cm->bit_depth);
+        vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta + 2);
+        vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
+
+        vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
+        vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
+
+        // Segment coding disabled for compred testing
+        if (high_q || (cpi->static_mb_pct == 100)) {
+          vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
+          vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
+          vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
+        }
+      } else {
+        // Disable segmentation and clear down features if alt ref
+        // is not active for this group
+
+        vp9_disable_segmentation(seg);
+
+        memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
+
+        seg->update_map = 0;
+        seg->update_data = 0;
+
+        vp9_clearall_segfeatures(seg);
+      }
+    } else if (rc->is_src_frame_alt_ref) {
+      // Special case where we are coding over the top of a previous
+      // alt ref frame.
+      // Segment coding disabled for compred testing
+
+      // Enable ref frame features for segment 0 as well
+      vp9_enable_segfeature(seg, 0, SEG_LVL_REF_FRAME);
+      vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
+
+      // All mbs should use ALTREF_FRAME
+      vp9_clear_segdata(seg, 0, SEG_LVL_REF_FRAME);
+      vp9_set_segdata(seg, 0, SEG_LVL_REF_FRAME, ALTREF_FRAME);
+      vp9_clear_segdata(seg, 1, SEG_LVL_REF_FRAME);
+      vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
+
+      // Skip all MBs if high Q (0,0 mv and skip coeffs)
+      if (high_q) {
+        vp9_enable_segfeature(seg, 0, SEG_LVL_SKIP);
+        vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
+      }
+      // Enable data update
+      seg->update_data = 1;
+    } else {
+      // All other frames.
+
+      // No updates.. leave things as they are.
+      seg->update_map = 0;
+      seg->update_data = 0;
+    }
+  }
+}
+
+static void update_reference_segmentation_map(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
+  uint8_t *cache_ptr = cm->last_frame_seg_map;
+  int row, col;
+
+  for (row = 0; row < cm->mi_rows; row++) {
+    MODE_INFO **mi_8x8 = mi_8x8_ptr;
+    uint8_t *cache = cache_ptr;
+    for (col = 0; col < cm->mi_cols; col++, mi_8x8++, cache++)
+      cache[0] = mi_8x8[0]->segment_id;
+    mi_8x8_ptr += cm->mi_stride;
+    cache_ptr += cm->mi_cols;
+  }
+}
+
+static void alloc_raw_frame_buffers(VP9_COMP *cpi) {
+  VP9_COMMON *cm = &cpi->common;
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
+
+  if (!cpi->lookahead)
+    cpi->lookahead = vp9_lookahead_init(oxcf->width, oxcf->height,
+                                        cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                      cm->use_highbitdepth,
+#endif
+                                      oxcf->lag_in_frames);
+  if (!cpi->lookahead)
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate lag buffers");
+
+  // TODO(agrange) Check if ARF is enabled and skip allocation if not.
+  if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer,
+                               oxcf->width, oxcf->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL))
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate altref buffer");
+}
+
+static void alloc_util_frame_buffers(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  if (vpx_realloc_frame_buffer(&cpi->last_frame_uf,
+                               cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL))
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate last frame buffer");
+
+  if (vpx_realloc_frame_buffer(&cpi->scaled_source,
+                               cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL))
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate scaled source buffer");
+
+  if (vpx_realloc_frame_buffer(&cpi->scaled_last_source,
+                               cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL))
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate scaled last source buffer");
+}
+
+
+static int alloc_context_buffers_ext(VP9_COMP *cpi) {
+  VP9_COMMON *cm = &cpi->common;
+  int mi_size = cm->mi_cols * cm->mi_rows;
+
+  cpi->mbmi_ext_base = vpx_calloc(mi_size, sizeof(*cpi->mbmi_ext_base));
+  if (!cpi->mbmi_ext_base)
+    return 1;
+
+  return 0;
+}
+
+static void alloc_compressor_data(VP9_COMP *cpi) {
+  VP9_COMMON *cm = &cpi->common;
+
+  vp9_alloc_context_buffers(cm, cm->width, cm->height);
+
+  alloc_context_buffers_ext(cpi);
+
+  vpx_free(cpi->tile_tok[0][0]);
+
+  {
+    unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols);
+    CHECK_MEM_ERROR(cm, cpi->tile_tok[0][0],
+        vpx_calloc(tokens, sizeof(*cpi->tile_tok[0][0])));
+  }
+
+  vp9_setup_pc_tree(&cpi->common, &cpi->td);
+}
+
+void vp9_new_framerate(VP9_COMP *cpi, double framerate) {
+  cpi->framerate = framerate < 0.1 ? 30 : framerate;
+  vp9_rc_update_framerate(cpi);
+}
+
+static void set_tile_limits(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+
+  int min_log2_tile_cols, max_log2_tile_cols;
+  vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
+
+  if (is_two_pass_svc(cpi) &&
+      (cpi->svc.encode_empty_frame_state == ENCODING ||
+      cpi->svc.number_spatial_layers > 1)) {
+    cm->log2_tile_cols = 0;
+    cm->log2_tile_rows = 0;
+  } else {
+    cm->log2_tile_cols = clamp(cpi->oxcf.tile_columns,
+                               min_log2_tile_cols, max_log2_tile_cols);
+    cm->log2_tile_rows = cpi->oxcf.tile_rows;
+  }
+}
+
+static void update_frame_size(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+
+  vp9_set_mb_mi(cm, cm->width, cm->height);
+  vp9_init_context_buffers(cm);
+  vp9_init_macroblockd(cm, xd, NULL);
+  cpi->td.mb.mbmi_ext_base = cpi->mbmi_ext_base;
+  memset(cpi->mbmi_ext_base, 0,
+         cm->mi_rows * cm->mi_cols * sizeof(*cpi->mbmi_ext_base));
+
+  set_tile_limits(cpi);
+
+  if (is_two_pass_svc(cpi)) {
+    if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer,
+                                 cm->width, cm->height,
+                                 cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 cm->use_highbitdepth,
+#endif
+                                 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                                 NULL, NULL, NULL))
+      vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                         "Failed to reallocate alt_ref_buffer");
+  }
+}
+
+static void init_buffer_indices(VP9_COMP *cpi) {
+  cpi->lst_fb_idx = 0;
+  cpi->gld_fb_idx = 1;
+  cpi->alt_fb_idx = 2;
+}
+
+static void init_config(struct VP9_COMP *cpi, VP9EncoderConfig *oxcf) {
+  VP9_COMMON *const cm = &cpi->common;
+
+  cpi->oxcf = *oxcf;
+  cpi->framerate = oxcf->init_framerate;
+  cm->profile = oxcf->profile;
+  cm->bit_depth = oxcf->bit_depth;
+#if CONFIG_VP9_HIGHBITDEPTH
+  cm->use_highbitdepth = oxcf->use_highbitdepth;
+#endif
+  cm->color_space = oxcf->color_space;
+  cm->color_range = oxcf->color_range;
+
+  cpi->target_level = oxcf->target_level;
+  cpi->keep_level_stats = oxcf->target_level != LEVEL_MAX;
+
+  cm->width = oxcf->width;
+  cm->height = oxcf->height;
+  alloc_compressor_data(cpi);
+
+  cpi->svc.temporal_layering_mode = oxcf->temporal_layering_mode;
+
+  // Single thread case: use counts in common.
+  cpi->td.counts = &cm->counts;
+
+  // Spatial scalability.
+  cpi->svc.number_spatial_layers = oxcf->ss_number_layers;
+  // Temporal scalability.
+  cpi->svc.number_temporal_layers = oxcf->ts_number_layers;
+
+  if ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) ||
+      ((cpi->svc.number_temporal_layers > 1 ||
+        cpi->svc.number_spatial_layers > 1) &&
+       cpi->oxcf.pass != 1)) {
+    vp9_init_layer_context(cpi);
+  }
+
+  // change includes all joint functionality
+  vp9_change_config(cpi, oxcf);
+
+  cpi->static_mb_pct = 0;
+  cpi->ref_frame_flags = 0;
+
+  init_buffer_indices(cpi);
+
+  vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
+}
+
+static void set_rc_buffer_sizes(RATE_CONTROL *rc,
+                                const VP9EncoderConfig *oxcf) {
+  const int64_t bandwidth = oxcf->target_bandwidth;
+  const int64_t starting = oxcf->starting_buffer_level_ms;
+  const int64_t optimal = oxcf->optimal_buffer_level_ms;
+  const int64_t maximum = oxcf->maximum_buffer_size_ms;
+
+  rc->starting_buffer_level = starting * bandwidth / 1000;
+  rc->optimal_buffer_level = (optimal == 0) ? bandwidth / 8
+                                            : optimal * bandwidth / 1000;
+  rc->maximum_buffer_size = (maximum == 0) ? bandwidth / 8
+                                           : maximum * bandwidth / 1000;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+#define HIGHBD_BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF) \
+    cpi->fn_ptr[BT].sdf = SDF; \
+    cpi->fn_ptr[BT].sdaf = SDAF; \
+    cpi->fn_ptr[BT].vf = VF; \
+    cpi->fn_ptr[BT].svf = SVF; \
+    cpi->fn_ptr[BT].svaf = SVAF; \
+    cpi->fn_ptr[BT].sdx3f = SDX3F; \
+    cpi->fn_ptr[BT].sdx8f = SDX8F; \
+    cpi->fn_ptr[BT].sdx4df = SDX4DF;
+
+#define MAKE_BFP_SAD_WRAPPER(fnname) \
+static unsigned int fnname##_bits8(const uint8_t *src_ptr, \
+                                   int source_stride, \
+                                   const uint8_t *ref_ptr, \
+                                   int ref_stride) {  \
+  return fnname(src_ptr, source_stride, ref_ptr, ref_stride); \
+} \
+static unsigned int fnname##_bits10(const uint8_t *src_ptr, \
+                                    int source_stride, \
+                                    const uint8_t *ref_ptr, \
+                                    int ref_stride) {  \
+  return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 2; \
+} \
+static unsigned int fnname##_bits12(const uint8_t *src_ptr, \
+                                    int source_stride, \
+                                    const uint8_t *ref_ptr, \
+                                    int ref_stride) {  \
+  return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 4; \
+}
+
+#define MAKE_BFP_SADAVG_WRAPPER(fnname) static unsigned int \
+fnname##_bits8(const uint8_t *src_ptr, \
+               int source_stride, \
+               const uint8_t *ref_ptr, \
+               int ref_stride, \
+               const uint8_t *second_pred) {  \
+  return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred); \
+} \
+static unsigned int fnname##_bits10(const uint8_t *src_ptr, \
+                                    int source_stride, \
+                                    const uint8_t *ref_ptr, \
+                                    int ref_stride, \
+                                    const uint8_t *second_pred) {  \
+  return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \
+                second_pred) >> 2; \
+} \
+static unsigned int fnname##_bits12(const uint8_t *src_ptr, \
+                                    int source_stride, \
+                                    const uint8_t *ref_ptr, \
+                                    int ref_stride, \
+                                    const uint8_t *second_pred) {  \
+  return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \
+                second_pred) >> 4; \
+}
+
+#define MAKE_BFP_SAD3_WRAPPER(fnname) \
+static void fnname##_bits8(const uint8_t *src_ptr, \
+                           int source_stride, \
+                           const uint8_t *ref_ptr, \
+                           int  ref_stride, \
+                           unsigned int *sad_array) {  \
+  fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
+} \
+static void fnname##_bits10(const uint8_t *src_ptr, \
+                            int source_stride, \
+                            const uint8_t *ref_ptr, \
+                            int  ref_stride, \
+                            unsigned int *sad_array) {  \
+  int i; \
+  fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
+  for (i = 0; i < 3; i++) \
+    sad_array[i] >>= 2; \
+} \
+static void fnname##_bits12(const uint8_t *src_ptr, \
+                            int source_stride, \
+                            const uint8_t *ref_ptr, \
+                            int  ref_stride, \
+                            unsigned int *sad_array) {  \
+  int i; \
+  fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
+  for (i = 0; i < 3; i++) \
+    sad_array[i] >>= 4; \
+}
+
+#define MAKE_BFP_SAD8_WRAPPER(fnname) \
+static void fnname##_bits8(const uint8_t *src_ptr, \
+                           int source_stride, \
+                           const uint8_t *ref_ptr, \
+                           int  ref_stride, \
+                           unsigned int *sad_array) {  \
+  fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
+} \
+static void fnname##_bits10(const uint8_t *src_ptr, \
+                            int source_stride, \
+                            const uint8_t *ref_ptr, \
+                            int  ref_stride, \
+                            unsigned int *sad_array) {  \
+  int i; \
+  fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
+  for (i = 0; i < 8; i++) \
+    sad_array[i] >>= 2; \
+} \
+static void fnname##_bits12(const uint8_t *src_ptr, \
+                            int source_stride, \
+                            const uint8_t *ref_ptr, \
+                            int  ref_stride, \
+                            unsigned int *sad_array) {  \
+  int i; \
+  fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
+  for (i = 0; i < 8; i++) \
+    sad_array[i] >>= 4; \
+}
+#define MAKE_BFP_SAD4D_WRAPPER(fnname) \
+static void fnname##_bits8(const uint8_t *src_ptr, \
+                           int source_stride, \
+                           const uint8_t* const ref_ptr[], \
+                           int  ref_stride, \
+                           unsigned int *sad_array) {  \
+  fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
+} \
+static void fnname##_bits10(const uint8_t *src_ptr, \
+                            int source_stride, \
+                            const uint8_t* const ref_ptr[], \
+                            int  ref_stride, \
+                            unsigned int *sad_array) {  \
+  int i; \
+  fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
+  for (i = 0; i < 4; i++) \
+  sad_array[i] >>= 2; \
+} \
+static void fnname##_bits12(const uint8_t *src_ptr, \
+                            int source_stride, \
+                            const uint8_t* const ref_ptr[], \
+                            int  ref_stride, \
+                            unsigned int *sad_array) {  \
+  int i; \
+  fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
+  for (i = 0; i < 4; i++) \
+  sad_array[i] >>= 4; \
+}
+
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x16)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x16_avg)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x16x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x32)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x32_avg)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x32x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x32)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x32_avg)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x32x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x64)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x64_avg)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x64x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x32)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x32_avg)
+MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad32x32x3)
+MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad32x32x8)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x32x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x64)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x64_avg)
+MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad64x64x3)
+MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad64x64x8)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x64x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x16)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x16_avg)
+MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x16x3)
+MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x16x8)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x16x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x8)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x8_avg)
+MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x8x3)
+MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x8x8)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x8x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x16)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x16_avg)
+MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x16x3)
+MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x16x8)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x16x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x8)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x8_avg)
+MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x8x3)
+MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x8x8)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x8x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x4)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x4_avg)
+MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x4x8)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x4x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x8)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x8_avg)
+MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x8x8)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x8x4d)
+MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x4)
+MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x4_avg)
+MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad4x4x3)
+MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x4x8)
+MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x4x4d)
+
+static void  highbd_set_var_fns(VP9_COMP *const cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  if (cm->use_highbitdepth) {
+    switch (cm->bit_depth) {
+      case VPX_BITS_8:
+        HIGHBD_BFP(BLOCK_32X16,
+                   vpx_highbd_sad32x16_bits8,
+                   vpx_highbd_sad32x16_avg_bits8,
+                   vpx_highbd_8_variance32x16,
+                   vpx_highbd_8_sub_pixel_variance32x16,
+                   vpx_highbd_8_sub_pixel_avg_variance32x16,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad32x16x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_16X32,
+                   vpx_highbd_sad16x32_bits8,
+                   vpx_highbd_sad16x32_avg_bits8,
+                   vpx_highbd_8_variance16x32,
+                   vpx_highbd_8_sub_pixel_variance16x32,
+                   vpx_highbd_8_sub_pixel_avg_variance16x32,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad16x32x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_64X32,
+                   vpx_highbd_sad64x32_bits8,
+                   vpx_highbd_sad64x32_avg_bits8,
+                   vpx_highbd_8_variance64x32,
+                   vpx_highbd_8_sub_pixel_variance64x32,
+                   vpx_highbd_8_sub_pixel_avg_variance64x32,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad64x32x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_32X64,
+                   vpx_highbd_sad32x64_bits8,
+                   vpx_highbd_sad32x64_avg_bits8,
+                   vpx_highbd_8_variance32x64,
+                   vpx_highbd_8_sub_pixel_variance32x64,
+                   vpx_highbd_8_sub_pixel_avg_variance32x64,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad32x64x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_32X32,
+                   vpx_highbd_sad32x32_bits8,
+                   vpx_highbd_sad32x32_avg_bits8,
+                   vpx_highbd_8_variance32x32,
+                   vpx_highbd_8_sub_pixel_variance32x32,
+                   vpx_highbd_8_sub_pixel_avg_variance32x32,
+                   vpx_highbd_sad32x32x3_bits8,
+                   vpx_highbd_sad32x32x8_bits8,
+                   vpx_highbd_sad32x32x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_64X64,
+                   vpx_highbd_sad64x64_bits8,
+                   vpx_highbd_sad64x64_avg_bits8,
+                   vpx_highbd_8_variance64x64,
+                   vpx_highbd_8_sub_pixel_variance64x64,
+                   vpx_highbd_8_sub_pixel_avg_variance64x64,
+                   vpx_highbd_sad64x64x3_bits8,
+                   vpx_highbd_sad64x64x8_bits8,
+                   vpx_highbd_sad64x64x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_16X16,
+                   vpx_highbd_sad16x16_bits8,
+                   vpx_highbd_sad16x16_avg_bits8,
+                   vpx_highbd_8_variance16x16,
+                   vpx_highbd_8_sub_pixel_variance16x16,
+                   vpx_highbd_8_sub_pixel_avg_variance16x16,
+                   vpx_highbd_sad16x16x3_bits8,
+                   vpx_highbd_sad16x16x8_bits8,
+                   vpx_highbd_sad16x16x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_16X8,
+                   vpx_highbd_sad16x8_bits8,
+                   vpx_highbd_sad16x8_avg_bits8,
+                   vpx_highbd_8_variance16x8,
+                   vpx_highbd_8_sub_pixel_variance16x8,
+                   vpx_highbd_8_sub_pixel_avg_variance16x8,
+                   vpx_highbd_sad16x8x3_bits8,
+                   vpx_highbd_sad16x8x8_bits8,
+                   vpx_highbd_sad16x8x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_8X16,
+                   vpx_highbd_sad8x16_bits8,
+                   vpx_highbd_sad8x16_avg_bits8,
+                   vpx_highbd_8_variance8x16,
+                   vpx_highbd_8_sub_pixel_variance8x16,
+                   vpx_highbd_8_sub_pixel_avg_variance8x16,
+                   vpx_highbd_sad8x16x3_bits8,
+                   vpx_highbd_sad8x16x8_bits8,
+                   vpx_highbd_sad8x16x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_8X8,
+                   vpx_highbd_sad8x8_bits8,
+                   vpx_highbd_sad8x8_avg_bits8,
+                   vpx_highbd_8_variance8x8,
+                   vpx_highbd_8_sub_pixel_variance8x8,
+                   vpx_highbd_8_sub_pixel_avg_variance8x8,
+                   vpx_highbd_sad8x8x3_bits8,
+                   vpx_highbd_sad8x8x8_bits8,
+                   vpx_highbd_sad8x8x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_8X4,
+                   vpx_highbd_sad8x4_bits8,
+                   vpx_highbd_sad8x4_avg_bits8,
+                   vpx_highbd_8_variance8x4,
+                   vpx_highbd_8_sub_pixel_variance8x4,
+                   vpx_highbd_8_sub_pixel_avg_variance8x4,
+                   NULL,
+                   vpx_highbd_sad8x4x8_bits8,
+                   vpx_highbd_sad8x4x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_4X8,
+                   vpx_highbd_sad4x8_bits8,
+                   vpx_highbd_sad4x8_avg_bits8,
+                   vpx_highbd_8_variance4x8,
+                   vpx_highbd_8_sub_pixel_variance4x8,
+                   vpx_highbd_8_sub_pixel_avg_variance4x8,
+                   NULL,
+                   vpx_highbd_sad4x8x8_bits8,
+                   vpx_highbd_sad4x8x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_4X4,
+                   vpx_highbd_sad4x4_bits8,
+                   vpx_highbd_sad4x4_avg_bits8,
+                   vpx_highbd_8_variance4x4,
+                   vpx_highbd_8_sub_pixel_variance4x4,
+                   vpx_highbd_8_sub_pixel_avg_variance4x4,
+                   vpx_highbd_sad4x4x3_bits8,
+                   vpx_highbd_sad4x4x8_bits8,
+                   vpx_highbd_sad4x4x4d_bits8)
+        break;
+
+      case VPX_BITS_10:
+        HIGHBD_BFP(BLOCK_32X16,
+                   vpx_highbd_sad32x16_bits10,
+                   vpx_highbd_sad32x16_avg_bits10,
+                   vpx_highbd_10_variance32x16,
+                   vpx_highbd_10_sub_pixel_variance32x16,
+                   vpx_highbd_10_sub_pixel_avg_variance32x16,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad32x16x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_16X32,
+                   vpx_highbd_sad16x32_bits10,
+                   vpx_highbd_sad16x32_avg_bits10,
+                   vpx_highbd_10_variance16x32,
+                   vpx_highbd_10_sub_pixel_variance16x32,
+                   vpx_highbd_10_sub_pixel_avg_variance16x32,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad16x32x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_64X32,
+                   vpx_highbd_sad64x32_bits10,
+                   vpx_highbd_sad64x32_avg_bits10,
+                   vpx_highbd_10_variance64x32,
+                   vpx_highbd_10_sub_pixel_variance64x32,
+                   vpx_highbd_10_sub_pixel_avg_variance64x32,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad64x32x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_32X64,
+                   vpx_highbd_sad32x64_bits10,
+                   vpx_highbd_sad32x64_avg_bits10,
+                   vpx_highbd_10_variance32x64,
+                   vpx_highbd_10_sub_pixel_variance32x64,
+                   vpx_highbd_10_sub_pixel_avg_variance32x64,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad32x64x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_32X32,
+                   vpx_highbd_sad32x32_bits10,
+                   vpx_highbd_sad32x32_avg_bits10,
+                   vpx_highbd_10_variance32x32,
+                   vpx_highbd_10_sub_pixel_variance32x32,
+                   vpx_highbd_10_sub_pixel_avg_variance32x32,
+                   vpx_highbd_sad32x32x3_bits10,
+                   vpx_highbd_sad32x32x8_bits10,
+                   vpx_highbd_sad32x32x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_64X64,
+                   vpx_highbd_sad64x64_bits10,
+                   vpx_highbd_sad64x64_avg_bits10,
+                   vpx_highbd_10_variance64x64,
+                   vpx_highbd_10_sub_pixel_variance64x64,
+                   vpx_highbd_10_sub_pixel_avg_variance64x64,
+                   vpx_highbd_sad64x64x3_bits10,
+                   vpx_highbd_sad64x64x8_bits10,
+                   vpx_highbd_sad64x64x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_16X16,
+                   vpx_highbd_sad16x16_bits10,
+                   vpx_highbd_sad16x16_avg_bits10,
+                   vpx_highbd_10_variance16x16,
+                   vpx_highbd_10_sub_pixel_variance16x16,
+                   vpx_highbd_10_sub_pixel_avg_variance16x16,
+                   vpx_highbd_sad16x16x3_bits10,
+                   vpx_highbd_sad16x16x8_bits10,
+                   vpx_highbd_sad16x16x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_16X8,
+                   vpx_highbd_sad16x8_bits10,
+                   vpx_highbd_sad16x8_avg_bits10,
+                   vpx_highbd_10_variance16x8,
+                   vpx_highbd_10_sub_pixel_variance16x8,
+                   vpx_highbd_10_sub_pixel_avg_variance16x8,
+                   vpx_highbd_sad16x8x3_bits10,
+                   vpx_highbd_sad16x8x8_bits10,
+                   vpx_highbd_sad16x8x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_8X16,
+                   vpx_highbd_sad8x16_bits10,
+                   vpx_highbd_sad8x16_avg_bits10,
+                   vpx_highbd_10_variance8x16,
+                   vpx_highbd_10_sub_pixel_variance8x16,
+                   vpx_highbd_10_sub_pixel_avg_variance8x16,
+                   vpx_highbd_sad8x16x3_bits10,
+                   vpx_highbd_sad8x16x8_bits10,
+                   vpx_highbd_sad8x16x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_8X8,
+                   vpx_highbd_sad8x8_bits10,
+                   vpx_highbd_sad8x8_avg_bits10,
+                   vpx_highbd_10_variance8x8,
+                   vpx_highbd_10_sub_pixel_variance8x8,
+                   vpx_highbd_10_sub_pixel_avg_variance8x8,
+                   vpx_highbd_sad8x8x3_bits10,
+                   vpx_highbd_sad8x8x8_bits10,
+                   vpx_highbd_sad8x8x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_8X4,
+                   vpx_highbd_sad8x4_bits10,
+                   vpx_highbd_sad8x4_avg_bits10,
+                   vpx_highbd_10_variance8x4,
+                   vpx_highbd_10_sub_pixel_variance8x4,
+                   vpx_highbd_10_sub_pixel_avg_variance8x4,
+                   NULL,
+                   vpx_highbd_sad8x4x8_bits10,
+                   vpx_highbd_sad8x4x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_4X8,
+                   vpx_highbd_sad4x8_bits10,
+                   vpx_highbd_sad4x8_avg_bits10,
+                   vpx_highbd_10_variance4x8,
+                   vpx_highbd_10_sub_pixel_variance4x8,
+                   vpx_highbd_10_sub_pixel_avg_variance4x8,
+                   NULL,
+                   vpx_highbd_sad4x8x8_bits10,
+                   vpx_highbd_sad4x8x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_4X4,
+                   vpx_highbd_sad4x4_bits10,
+                   vpx_highbd_sad4x4_avg_bits10,
+                   vpx_highbd_10_variance4x4,
+                   vpx_highbd_10_sub_pixel_variance4x4,
+                   vpx_highbd_10_sub_pixel_avg_variance4x4,
+                   vpx_highbd_sad4x4x3_bits10,
+                   vpx_highbd_sad4x4x8_bits10,
+                   vpx_highbd_sad4x4x4d_bits10)
+        break;
+
+      case VPX_BITS_12:
+        HIGHBD_BFP(BLOCK_32X16,
+                   vpx_highbd_sad32x16_bits12,
+                   vpx_highbd_sad32x16_avg_bits12,
+                   vpx_highbd_12_variance32x16,
+                   vpx_highbd_12_sub_pixel_variance32x16,
+                   vpx_highbd_12_sub_pixel_avg_variance32x16,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad32x16x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_16X32,
+                   vpx_highbd_sad16x32_bits12,
+                   vpx_highbd_sad16x32_avg_bits12,
+                   vpx_highbd_12_variance16x32,
+                   vpx_highbd_12_sub_pixel_variance16x32,
+                   vpx_highbd_12_sub_pixel_avg_variance16x32,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad16x32x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_64X32,
+                   vpx_highbd_sad64x32_bits12,
+                   vpx_highbd_sad64x32_avg_bits12,
+                   vpx_highbd_12_variance64x32,
+                   vpx_highbd_12_sub_pixel_variance64x32,
+                   vpx_highbd_12_sub_pixel_avg_variance64x32,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad64x32x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_32X64,
+                   vpx_highbd_sad32x64_bits12,
+                   vpx_highbd_sad32x64_avg_bits12,
+                   vpx_highbd_12_variance32x64,
+                   vpx_highbd_12_sub_pixel_variance32x64,
+                   vpx_highbd_12_sub_pixel_avg_variance32x64,
+                   NULL,
+                   NULL,
+                   vpx_highbd_sad32x64x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_32X32,
+                   vpx_highbd_sad32x32_bits12,
+                   vpx_highbd_sad32x32_avg_bits12,
+                   vpx_highbd_12_variance32x32,
+                   vpx_highbd_12_sub_pixel_variance32x32,
+                   vpx_highbd_12_sub_pixel_avg_variance32x32,
+                   vpx_highbd_sad32x32x3_bits12,
+                   vpx_highbd_sad32x32x8_bits12,
+                   vpx_highbd_sad32x32x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_64X64,
+                   vpx_highbd_sad64x64_bits12,
+                   vpx_highbd_sad64x64_avg_bits12,
+                   vpx_highbd_12_variance64x64,
+                   vpx_highbd_12_sub_pixel_variance64x64,
+                   vpx_highbd_12_sub_pixel_avg_variance64x64,
+                   vpx_highbd_sad64x64x3_bits12,
+                   vpx_highbd_sad64x64x8_bits12,
+                   vpx_highbd_sad64x64x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_16X16,
+                   vpx_highbd_sad16x16_bits12,
+                   vpx_highbd_sad16x16_avg_bits12,
+                   vpx_highbd_12_variance16x16,
+                   vpx_highbd_12_sub_pixel_variance16x16,
+                   vpx_highbd_12_sub_pixel_avg_variance16x16,
+                   vpx_highbd_sad16x16x3_bits12,
+                   vpx_highbd_sad16x16x8_bits12,
+                   vpx_highbd_sad16x16x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_16X8,
+                   vpx_highbd_sad16x8_bits12,
+                   vpx_highbd_sad16x8_avg_bits12,
+                   vpx_highbd_12_variance16x8,
+                   vpx_highbd_12_sub_pixel_variance16x8,
+                   vpx_highbd_12_sub_pixel_avg_variance16x8,
+                   vpx_highbd_sad16x8x3_bits12,
+                   vpx_highbd_sad16x8x8_bits12,
+                   vpx_highbd_sad16x8x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_8X16,
+                   vpx_highbd_sad8x16_bits12,
+                   vpx_highbd_sad8x16_avg_bits12,
+                   vpx_highbd_12_variance8x16,
+                   vpx_highbd_12_sub_pixel_variance8x16,
+                   vpx_highbd_12_sub_pixel_avg_variance8x16,
+                   vpx_highbd_sad8x16x3_bits12,
+                   vpx_highbd_sad8x16x8_bits12,
+                   vpx_highbd_sad8x16x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_8X8,
+                   vpx_highbd_sad8x8_bits12,
+                   vpx_highbd_sad8x8_avg_bits12,
+                   vpx_highbd_12_variance8x8,
+                   vpx_highbd_12_sub_pixel_variance8x8,
+                   vpx_highbd_12_sub_pixel_avg_variance8x8,
+                   vpx_highbd_sad8x8x3_bits12,
+                   vpx_highbd_sad8x8x8_bits12,
+                   vpx_highbd_sad8x8x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_8X4,
+                   vpx_highbd_sad8x4_bits12,
+                   vpx_highbd_sad8x4_avg_bits12,
+                   vpx_highbd_12_variance8x4,
+                   vpx_highbd_12_sub_pixel_variance8x4,
+                   vpx_highbd_12_sub_pixel_avg_variance8x4,
+                   NULL,
+                   vpx_highbd_sad8x4x8_bits12,
+                   vpx_highbd_sad8x4x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_4X8,
+                   vpx_highbd_sad4x8_bits12,
+                   vpx_highbd_sad4x8_avg_bits12,
+                   vpx_highbd_12_variance4x8,
+                   vpx_highbd_12_sub_pixel_variance4x8,
+                   vpx_highbd_12_sub_pixel_avg_variance4x8,
+                   NULL,
+                   vpx_highbd_sad4x8x8_bits12,
+                   vpx_highbd_sad4x8x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_4X4,
+                   vpx_highbd_sad4x4_bits12,
+                   vpx_highbd_sad4x4_avg_bits12,
+                   vpx_highbd_12_variance4x4,
+                   vpx_highbd_12_sub_pixel_variance4x4,
+                   vpx_highbd_12_sub_pixel_avg_variance4x4,
+                   vpx_highbd_sad4x4x3_bits12,
+                   vpx_highbd_sad4x4x8_bits12,
+                   vpx_highbd_sad4x4x4d_bits12)
+        break;
+
+      default:
+        assert(0 && "cm->bit_depth should be VPX_BITS_8, "
+                    "VPX_BITS_10 or VPX_BITS_12");
+    }
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void realloc_segmentation_maps(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+
+  // Create the encoder segmentation map and set all entries to 0
+  vpx_free(cpi->segmentation_map);
+  CHECK_MEM_ERROR(cm, cpi->segmentation_map,
+                  vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
+
+  // Create a map used for cyclic background refresh.
+  if (cpi->cyclic_refresh)
+    vp9_cyclic_refresh_free(cpi->cyclic_refresh);
+  CHECK_MEM_ERROR(cm, cpi->cyclic_refresh,
+                  vp9_cyclic_refresh_alloc(cm->mi_rows, cm->mi_cols));
+
+  // Create a map used to mark inactive areas.
+  vpx_free(cpi->active_map.map);
+  CHECK_MEM_ERROR(cm, cpi->active_map.map,
+                  vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
+
+  // And a place holder structure is the coding context
+  // for use if we want to save and restore it
+  vpx_free(cpi->coding_context.last_frame_seg_map_copy);
+  CHECK_MEM_ERROR(cm, cpi->coding_context.last_frame_seg_map_copy,
+                  vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
+}
+
+void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
+  VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int last_w = cpi->oxcf.width;
+  int last_h = cpi->oxcf.height;
+
+  if (cm->profile != oxcf->profile)
+    cm->profile = oxcf->profile;
+  cm->bit_depth = oxcf->bit_depth;
+  cm->color_space = oxcf->color_space;
+  cm->color_range = oxcf->color_range;
+
+  cpi->target_level = oxcf->target_level;
+  cpi->keep_level_stats = oxcf->target_level != LEVEL_MAX;
+
+  if (cm->profile <= PROFILE_1)
+    assert(cm->bit_depth == VPX_BITS_8);
+  else
+    assert(cm->bit_depth > VPX_BITS_8);
+
+  cpi->oxcf = *oxcf;
+#if CONFIG_VP9_HIGHBITDEPTH
+  cpi->td.mb.e_mbd.bd = (int)cm->bit_depth;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  if ((oxcf->pass == 0) && (oxcf->rc_mode == VPX_Q)) {
+    rc->baseline_gf_interval = FIXED_GF_INTERVAL;
+  } else {
+    rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2;
+  }
+
+  cpi->refresh_golden_frame = 0;
+  cpi->refresh_last_frame = 1;
+  cm->refresh_frame_context = 1;
+  cm->reset_frame_context = 0;
+
+  vp9_reset_segment_features(&cm->seg);
+  vp9_set_high_precision_mv(cpi, 0);
+
+  {
+    int i;
+
+    for (i = 0; i < MAX_SEGMENTS; i++)
+      cpi->segment_encode_breakout[i] = cpi->oxcf.encode_breakout;
+  }
+  cpi->encode_breakout = cpi->oxcf.encode_breakout;
+
+  set_rc_buffer_sizes(rc, &cpi->oxcf);
+
+  // Under a configuration change, where maximum_buffer_size may change,
+  // keep buffer level clipped to the maximum allowed buffer size.
+  rc->bits_off_target = VPXMIN(rc->bits_off_target, rc->maximum_buffer_size);
+  rc->buffer_level = VPXMIN(rc->buffer_level, rc->maximum_buffer_size);
+
+  // Set up frame rate and related parameters rate control values.
+  vp9_new_framerate(cpi, cpi->framerate);
+
+  // Set absolute upper and lower quality limits
+  rc->worst_quality = cpi->oxcf.worst_allowed_q;
+  rc->best_quality = cpi->oxcf.best_allowed_q;
+
+  cm->interp_filter = cpi->sf.default_interp_filter;
+
+  if (cpi->oxcf.render_width > 0 && cpi->oxcf.render_height > 0) {
+    cm->render_width = cpi->oxcf.render_width;
+    cm->render_height = cpi->oxcf.render_height;
+  } else {
+    cm->render_width = cpi->oxcf.width;
+    cm->render_height = cpi->oxcf.height;
+  }
+  if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
+    cm->width = cpi->oxcf.width;
+    cm->height = cpi->oxcf.height;
+    cpi->external_resize = 1;
+  }
+
+  if (cpi->initial_width) {
+    int new_mi_size = 0;
+    vp9_set_mb_mi(cm, cm->width, cm->height);
+    new_mi_size = cm->mi_stride * calc_mi_size(cm->mi_rows);
+    if (cm->mi_alloc_size < new_mi_size) {
+      vp9_free_context_buffers(cm);
+      alloc_compressor_data(cpi);
+      realloc_segmentation_maps(cpi);
+      cpi->initial_width = cpi->initial_height = 0;
+      cpi->external_resize = 0;
+    } else if (cm->mi_alloc_size == new_mi_size &&
+             (cpi->oxcf.width > last_w || cpi->oxcf.height > last_h)) {
+        vp9_alloc_loop_filter(cm);
+    }
+  }
+
+  update_frame_size(cpi);
+
+  if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
+    memset(cpi->consec_zero_mv, 0,
+               cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
+      vp9_cyclic_refresh_reset_resize(cpi);
+  }
+
+  if ((cpi->svc.number_temporal_layers > 1 &&
+      cpi->oxcf.rc_mode == VPX_CBR) ||
+      ((cpi->svc.number_temporal_layers > 1 ||
+        cpi->svc.number_spatial_layers > 1) &&
+       cpi->oxcf.pass != 1)) {
+    vp9_update_layer_context_change_config(cpi,
+                                           (int)cpi->oxcf.target_bandwidth);
+  }
+
+  cpi->alt_ref_source = NULL;
+  rc->is_src_frame_alt_ref = 0;
+
+#if 0
+  // Experimental RD Code
+  cpi->frame_distortion = 0;
+  cpi->last_frame_distortion = 0;
+#endif
+
+  set_tile_limits(cpi);
+
+  cpi->ext_refresh_frame_flags_pending = 0;
+  cpi->ext_refresh_frame_context_pending = 0;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  highbd_set_var_fns(cpi);
+#endif
+}
+
+#ifndef M_LOG2_E
+#define M_LOG2_E 0.693147180559945309417
+#endif
+#define log2f(x) (log (x) / (float) M_LOG2_E)
+
+/***********************************************************************
+ * Read before modifying 'cal_nmvjointsadcost' or 'cal_nmvsadcosts'    *
+ ***********************************************************************
+ * The following 2 functions ('cal_nmvjointsadcost' and                *
+ * 'cal_nmvsadcosts') are used to calculate cost lookup tables         *
+ * used by 'vp9_diamond_search_sad'. The C implementation of the       *
+ * function is generic, but the AVX intrinsics optimised version       *
+ * relies on the following properties of the computed tables:          *
+ * For cal_nmvjointsadcost:                                            *
+ *   - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3]     *
+ * For cal_nmvsadcosts:                                                *
+ *   - For all i: mvsadcost[0][i] == mvsadcost[1][i]                   *
+ *         (Equal costs for both components)                           *
+ *   - For all i: mvsadcost[0][i] == mvsadcost[0][-i]                  *
+ *         (Cost function is even)                                     *
+ * If these do not hold, then the AVX optimised version of the         *
+ * 'vp9_diamond_search_sad' function cannot be used as it is, in which *
+ * case you can revert to using the C function instead.                *
+ ***********************************************************************/
+
+static void cal_nmvjointsadcost(int *mvjointsadcost) {
+  /*********************************************************************
+   * Warning: Read the comments above before modifying this function   *
+   *********************************************************************/
+  mvjointsadcost[0] = 600;
+  mvjointsadcost[1] = 300;
+  mvjointsadcost[2] = 300;
+  mvjointsadcost[3] = 300;
+}
+
+static void cal_nmvsadcosts(int *mvsadcost[2]) {
+  /*********************************************************************
+   * Warning: Read the comments above before modifying this function   *
+   *********************************************************************/
+  int i = 1;
+
+  mvsadcost[0][0] = 0;
+  mvsadcost[1][0] = 0;
+
+  do {
+    double z = 256 * (2 * (log2f(8 * i) + .6));
+    mvsadcost[0][i] = (int)z;
+    mvsadcost[1][i] = (int)z;
+    mvsadcost[0][-i] = (int)z;
+    mvsadcost[1][-i] = (int)z;
+  } while (++i <= MV_MAX);
+}
+
+static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
+  int i = 1;
+
+  mvsadcost[0][0] = 0;
+  mvsadcost[1][0] = 0;
+
+  do {
+    double z = 256 * (2 * (log2f(8 * i) + .6));
+    mvsadcost[0][i] = (int)z;
+    mvsadcost[1][i] = (int)z;
+    mvsadcost[0][-i] = (int)z;
+    mvsadcost[1][-i] = (int)z;
+  } while (++i <= MV_MAX);
+}
+
+VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
+                                BufferPool *const pool) {
+  unsigned int i;
+  VP9_COMP *volatile const cpi = vpx_memalign(32, sizeof(VP9_COMP));
+  VP9_COMMON *volatile const cm = cpi != NULL ? &cpi->common : NULL;
+
+  if (!cm)
+    return NULL;
+
+  vp9_zero(*cpi);
+
+  if (setjmp(cm->error.jmp)) {
+    cm->error.setjmp = 0;
+    vp9_remove_compressor(cpi);
+    return 0;
+  }
+
+  cm->error.setjmp = 1;
+  cm->alloc_mi = vp9_enc_alloc_mi;
+  cm->free_mi = vp9_enc_free_mi;
+  cm->setup_mi = vp9_enc_setup_mi;
+
+  CHECK_MEM_ERROR(cm, cm->fc,
+                  (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc)));
+  CHECK_MEM_ERROR(cm, cm->frame_contexts,
+                  (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS,
+                  sizeof(*cm->frame_contexts)));
+
+  cpi->use_svc = 0;
+  cpi->resize_state = 0;
+  cpi->external_resize = 0;
+  cpi->resize_avg_qp = 0;
+  cpi->resize_buffer_underflow = 0;
+  cpi->use_skin_detection = 0;
+  cpi->common.buffer_pool = pool;
+
+  init_config(cpi, oxcf);
+  vp9_rc_init(&cpi->oxcf, oxcf->pass, &cpi->rc);
+
+  cm->current_video_frame = 0;
+  cpi->partition_search_skippable_frame = 0;
+  cpi->tile_data = NULL;
+
+  realloc_segmentation_maps(cpi);
+
+  CHECK_MEM_ERROR(cm, cpi->consec_zero_mv,
+                  vpx_calloc(cm->mi_rows * cm->mi_cols,
+                             sizeof(*cpi->consec_zero_mv)));
+
+  CHECK_MEM_ERROR(cm, cpi->nmvcosts[0],
+                  vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[0])));
+  CHECK_MEM_ERROR(cm, cpi->nmvcosts[1],
+                  vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[1])));
+  CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[0],
+                  vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[0])));
+  CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[1],
+                  vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[1])));
+  CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[0],
+                  vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[0])));
+  CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[1],
+                  vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[1])));
+  CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[0],
+                  vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[0])));
+  CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[1],
+                  vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[1])));
+
+  for (i = 0; i < (sizeof(cpi->mbgraph_stats) /
+                   sizeof(cpi->mbgraph_stats[0])); i++) {
+    CHECK_MEM_ERROR(cm, cpi->mbgraph_stats[i].mb_stats,
+                    vpx_calloc(cm->MBs *
+                               sizeof(*cpi->mbgraph_stats[i].mb_stats), 1));
+  }
+
+#if CONFIG_FP_MB_STATS
+  cpi->use_fp_mb_stats = 0;
+  if (cpi->use_fp_mb_stats) {
+    // a place holder used to store the first pass mb stats in the first pass
+    CHECK_MEM_ERROR(cm, cpi->twopass.frame_mb_stats_buf,
+                    vpx_calloc(cm->MBs * sizeof(uint8_t), 1));
+  } else {
+    cpi->twopass.frame_mb_stats_buf = NULL;
+  }
+#endif
+
+  cpi->refresh_alt_ref_frame = 0;
+  cpi->multi_arf_last_grp_enabled = 0;
+
+  cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
+
+  init_level_info(&cpi->level_info);
+
+#if CONFIG_INTERNAL_STATS
+  cpi->b_calculate_ssimg = 0;
+  cpi->b_calculate_blockiness = 1;
+  cpi->b_calculate_consistency = 1;
+  cpi->total_inconsistency = 0;
+  cpi->psnr.worst = 100.0;
+  cpi->worst_ssim = 100.0;
+
+  cpi->count = 0;
+  cpi->bytes = 0;
+
+  if (cpi->b_calculate_psnr) {
+    cpi->total_sq_error = 0;
+    cpi->total_samples = 0;
+
+    cpi->totalp_sq_error = 0;
+    cpi->totalp_samples = 0;
+
+    cpi->tot_recode_hits = 0;
+    cpi->summed_quality = 0;
+    cpi->summed_weights = 0;
+    cpi->summedp_quality = 0;
+    cpi->summedp_weights = 0;
+  }
+
+  if (cpi->b_calculate_ssimg) {
+    cpi->ssimg.worst= 100.0;
+  }
+  cpi->fastssim.worst = 100.0;
+
+  cpi->psnrhvs.worst = 100.0;
+
+  if (cpi->b_calculate_blockiness) {
+    cpi->total_blockiness = 0;
+    cpi->worst_blockiness = 0.0;
+  }
+
+  if (cpi->b_calculate_consistency) {
+    CHECK_MEM_ERROR(cm, cpi->ssim_vars,
+                    vpx_malloc(sizeof(*cpi->ssim_vars) * 4 *
+                               cpi->common.mi_rows * cpi->common.mi_cols));
+    cpi->worst_consistency = 100.0;
+  }
+
+#endif
+
+  cpi->first_time_stamp_ever = INT64_MAX;
+
+  /*********************************************************************
+   * Warning: Read the comments around 'cal_nmvjointsadcost' and       *
+   * 'cal_nmvsadcosts' before modifying how these tables are computed. *
+   *********************************************************************/
+  cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
+  cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX];
+  cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX];
+  cpi->td.mb.nmvsadcost[0] = &cpi->nmvsadcosts[0][MV_MAX];
+  cpi->td.mb.nmvsadcost[1] = &cpi->nmvsadcosts[1][MV_MAX];
+  cal_nmvsadcosts(cpi->td.mb.nmvsadcost);
+
+  cpi->td.mb.nmvcost_hp[0] = &cpi->nmvcosts_hp[0][MV_MAX];
+  cpi->td.mb.nmvcost_hp[1] = &cpi->nmvcosts_hp[1][MV_MAX];
+  cpi->td.mb.nmvsadcost_hp[0] = &cpi->nmvsadcosts_hp[0][MV_MAX];
+  cpi->td.mb.nmvsadcost_hp[1] = &cpi->nmvsadcosts_hp[1][MV_MAX];
+  cal_nmvsadcosts_hp(cpi->td.mb.nmvsadcost_hp);
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#ifdef OUTPUT_YUV_DENOISED
+  yuv_denoised_file = fopen("denoised.yuv", "ab");
+#endif
+#endif
+#ifdef OUTPUT_YUV_SKINMAP
+  yuv_skinmap_file = fopen("skinmap.yuv", "ab");
+#endif
+#ifdef OUTPUT_YUV_REC
+  yuv_rec_file = fopen("rec.yuv", "wb");
+#endif
+
+#if 0
+  framepsnr = fopen("framepsnr.stt", "a");
+  kf_list = fopen("kf_list.stt", "w");
+#endif
+
+  cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
+
+  if (oxcf->pass == 1) {
+    vp9_init_first_pass(cpi);
+  } else if (oxcf->pass == 2) {
+    const size_t packet_sz = sizeof(FIRSTPASS_STATS);
+    const int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz);
+
+    if (cpi->svc.number_spatial_layers > 1
+        || cpi->svc.number_temporal_layers > 1) {
+      FIRSTPASS_STATS *const stats = oxcf->two_pass_stats_in.buf;
+      FIRSTPASS_STATS *stats_copy[VPX_SS_MAX_LAYERS] = {0};
+      int i;
+
+      for (i = 0; i < oxcf->ss_number_layers; ++i) {
+        FIRSTPASS_STATS *const last_packet_for_layer =
+            &stats[packets - oxcf->ss_number_layers + i];
+        const int layer_id = (int)last_packet_for_layer->spatial_layer_id;
+        const int packets_in_layer = (int)last_packet_for_layer->count + 1;
+        if (layer_id >= 0 && layer_id < oxcf->ss_number_layers) {
+          LAYER_CONTEXT *const lc = &cpi->svc.layer_context[layer_id];
+
+          vpx_free(lc->rc_twopass_stats_in.buf);
+
+          lc->rc_twopass_stats_in.sz = packets_in_layer * packet_sz;
+          CHECK_MEM_ERROR(cm, lc->rc_twopass_stats_in.buf,
+                          vpx_malloc(lc->rc_twopass_stats_in.sz));
+          lc->twopass.stats_in_start = lc->rc_twopass_stats_in.buf;
+          lc->twopass.stats_in = lc->twopass.stats_in_start;
+          lc->twopass.stats_in_end = lc->twopass.stats_in_start
+                                     + packets_in_layer - 1;
+          stats_copy[layer_id] = lc->rc_twopass_stats_in.buf;
+        }
+      }
+
+      for (i = 0; i < packets; ++i) {
+        const int layer_id = (int)stats[i].spatial_layer_id;
+        if (layer_id >= 0 && layer_id < oxcf->ss_number_layers
+            && stats_copy[layer_id] != NULL) {
+          *stats_copy[layer_id] = stats[i];
+          ++stats_copy[layer_id];
+        }
+      }
+
+      vp9_init_second_pass_spatial_svc(cpi);
+    } else {
+#if CONFIG_FP_MB_STATS
+      if (cpi->use_fp_mb_stats) {
+        const size_t psz = cpi->common.MBs * sizeof(uint8_t);
+        const int ps = (int)(oxcf->firstpass_mb_stats_in.sz / psz);
+
+        cpi->twopass.firstpass_mb_stats.mb_stats_start =
+            oxcf->firstpass_mb_stats_in.buf;
+        cpi->twopass.firstpass_mb_stats.mb_stats_end =
+            cpi->twopass.firstpass_mb_stats.mb_stats_start +
+            (ps - 1) * cpi->common.MBs * sizeof(uint8_t);
+      }
+#endif
+
+      cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf;
+      cpi->twopass.stats_in = cpi->twopass.stats_in_start;
+      cpi->twopass.stats_in_end = &cpi->twopass.stats_in[packets - 1];
+
+      vp9_init_second_pass(cpi);
+    }
+  }
+
+  vp9_set_speed_features_framesize_independent(cpi);
+  vp9_set_speed_features_framesize_dependent(cpi);
+
+  // Allocate memory to store variances for a frame.
+  CHECK_MEM_ERROR(cm, cpi->source_diff_var,
+                  vpx_calloc(cm->MBs, sizeof(diff)));
+  cpi->source_var_thresh = 0;
+  cpi->frames_till_next_var_check = 0;
+
+#define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF)\
+    cpi->fn_ptr[BT].sdf            = SDF; \
+    cpi->fn_ptr[BT].sdaf           = SDAF; \
+    cpi->fn_ptr[BT].vf             = VF; \
+    cpi->fn_ptr[BT].svf            = SVF; \
+    cpi->fn_ptr[BT].svaf           = SVAF; \
+    cpi->fn_ptr[BT].sdx3f          = SDX3F; \
+    cpi->fn_ptr[BT].sdx8f          = SDX8F; \
+    cpi->fn_ptr[BT].sdx4df         = SDX4DF;
+
+  BFP(BLOCK_32X16, vpx_sad32x16, vpx_sad32x16_avg,
+      vpx_variance32x16, vpx_sub_pixel_variance32x16,
+      vpx_sub_pixel_avg_variance32x16, NULL, NULL, vpx_sad32x16x4d)
+
+  BFP(BLOCK_16X32, vpx_sad16x32, vpx_sad16x32_avg,
+      vpx_variance16x32, vpx_sub_pixel_variance16x32,
+      vpx_sub_pixel_avg_variance16x32, NULL, NULL, vpx_sad16x32x4d)
+
+  BFP(BLOCK_64X32, vpx_sad64x32, vpx_sad64x32_avg,
+      vpx_variance64x32, vpx_sub_pixel_variance64x32,
+      vpx_sub_pixel_avg_variance64x32, NULL, NULL, vpx_sad64x32x4d)
+
+  BFP(BLOCK_32X64, vpx_sad32x64, vpx_sad32x64_avg,
+      vpx_variance32x64, vpx_sub_pixel_variance32x64,
+      vpx_sub_pixel_avg_variance32x64, NULL, NULL, vpx_sad32x64x4d)
+
+  BFP(BLOCK_32X32, vpx_sad32x32, vpx_sad32x32_avg,
+      vpx_variance32x32, vpx_sub_pixel_variance32x32,
+      vpx_sub_pixel_avg_variance32x32, vpx_sad32x32x3, vpx_sad32x32x8,
+      vpx_sad32x32x4d)
+
+  BFP(BLOCK_64X64, vpx_sad64x64, vpx_sad64x64_avg,
+      vpx_variance64x64, vpx_sub_pixel_variance64x64,
+      vpx_sub_pixel_avg_variance64x64, vpx_sad64x64x3, vpx_sad64x64x8,
+      vpx_sad64x64x4d)
+
+  BFP(BLOCK_16X16, vpx_sad16x16, vpx_sad16x16_avg,
+      vpx_variance16x16, vpx_sub_pixel_variance16x16,
+      vpx_sub_pixel_avg_variance16x16, vpx_sad16x16x3, vpx_sad16x16x8,
+      vpx_sad16x16x4d)
+
+  BFP(BLOCK_16X8, vpx_sad16x8, vpx_sad16x8_avg,
+      vpx_variance16x8, vpx_sub_pixel_variance16x8,
+      vpx_sub_pixel_avg_variance16x8,
+      vpx_sad16x8x3, vpx_sad16x8x8, vpx_sad16x8x4d)
+
+  BFP(BLOCK_8X16, vpx_sad8x16, vpx_sad8x16_avg,
+      vpx_variance8x16, vpx_sub_pixel_variance8x16,
+      vpx_sub_pixel_avg_variance8x16,
+      vpx_sad8x16x3, vpx_sad8x16x8, vpx_sad8x16x4d)
+
+  BFP(BLOCK_8X8, vpx_sad8x8, vpx_sad8x8_avg,
+      vpx_variance8x8, vpx_sub_pixel_variance8x8,
+      vpx_sub_pixel_avg_variance8x8,
+      vpx_sad8x8x3, vpx_sad8x8x8, vpx_sad8x8x4d)
+
+  BFP(BLOCK_8X4, vpx_sad8x4, vpx_sad8x4_avg,
+      vpx_variance8x4, vpx_sub_pixel_variance8x4,
+      vpx_sub_pixel_avg_variance8x4, NULL, vpx_sad8x4x8, vpx_sad8x4x4d)
+
+  BFP(BLOCK_4X8, vpx_sad4x8, vpx_sad4x8_avg,
+      vpx_variance4x8, vpx_sub_pixel_variance4x8,
+      vpx_sub_pixel_avg_variance4x8, NULL, vpx_sad4x8x8, vpx_sad4x8x4d)
+
+  BFP(BLOCK_4X4, vpx_sad4x4, vpx_sad4x4_avg,
+      vpx_variance4x4, vpx_sub_pixel_variance4x4,
+      vpx_sub_pixel_avg_variance4x4,
+      vpx_sad4x4x3, vpx_sad4x4x8, vpx_sad4x4x4d)
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  highbd_set_var_fns(cpi);
+#endif
+
+  /* vp9_init_quantizer() is first called here. Add check in
+   * vp9_frame_init_quantizer() so that vp9_init_quantizer is only
+   * called later when needed. This will avoid unnecessary calls of
+   * vp9_init_quantizer() for every frame.
+   */
+  vp9_init_quantizer(cpi);
+
+  vp9_loop_filter_init(cm);
+
+  cm->error.setjmp = 0;
+
+  return cpi;
+}
+
+#if CONFIG_INTERNAL_STATS
+#define SNPRINT(H, T) \
+  snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T))
+
+#define SNPRINT2(H, T, V) \
+  snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T), (V))
+#endif  // CONFIG_INTERNAL_STATS
+
+void vp9_remove_compressor(VP9_COMP *cpi) {
+  VP9_COMMON *cm;
+  unsigned int i;
+  int t;
+
+  if (!cpi)
+    return;
+
+  cm = &cpi->common;
+  if (cm->current_video_frame > 0) {
+#if CONFIG_INTERNAL_STATS
+    vpx_clear_system_state();
+
+    if (cpi->oxcf.pass != 1) {
+      char headings[512] = {0};
+      char results[512] = {0};
+      FILE *f = fopen("opsnr.stt", "a");
+      double time_encoded = (cpi->last_end_time_stamp_seen
+                             - cpi->first_time_stamp_ever) / 10000000.000;
+      double total_encode_time = (cpi->time_receive_data +
+                                  cpi->time_compress_data)   / 1000.000;
+      const double dr =
+          (double)cpi->bytes * (double) 8 / (double)1000 / time_encoded;
+      const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
+      const double target_rate = (double)cpi->oxcf.target_bandwidth / 1000;
+      const double rate_err = ((100.0 * (dr - target_rate)) / target_rate);
+
+      if (cpi->b_calculate_psnr) {
+        const double total_psnr =
+            vpx_sse_to_psnr((double)cpi->total_samples, peak,
+                            (double)cpi->total_sq_error);
+        const double totalp_psnr =
+            vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
+                            (double)cpi->totalp_sq_error);
+        const double total_ssim = 100 * pow(cpi->summed_quality /
+                                            cpi->summed_weights, 8.0);
+        const double totalp_ssim = 100 * pow(cpi->summedp_quality /
+                                             cpi->summedp_weights, 8.0);
+
+        snprintf(headings, sizeof(headings),
+                 "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
+                 "VPXSSIM\tVPSSIMP\tFASTSIM\tPSNRHVS\t"
+                 "WstPsnr\tWstSsim\tWstFast\tWstHVS");
+        snprintf(results, sizeof(results),
+                 "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
+                 "%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
+                 "%7.3f\t%7.3f\t%7.3f\t%7.3f",
+                 dr, cpi->psnr.stat[ALL] / cpi->count, total_psnr,
+                 cpi->psnrp.stat[ALL] / cpi->count, totalp_psnr,
+                 total_ssim, totalp_ssim,
+                 cpi->fastssim.stat[ALL] / cpi->count,
+                 cpi->psnrhvs.stat[ALL] / cpi->count,
+                 cpi->psnr.worst, cpi->worst_ssim, cpi->fastssim.worst,
+                 cpi->psnrhvs.worst);
+
+        if (cpi->b_calculate_blockiness) {
+          SNPRINT(headings, "\t  Block\tWstBlck");
+          SNPRINT2(results, "\t%7.3f", cpi->total_blockiness / cpi->count);
+          SNPRINT2(results, "\t%7.3f", cpi->worst_blockiness);
+        }
+
+        if (cpi->b_calculate_consistency) {
+          double consistency =
+              vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
+                              (double)cpi->total_inconsistency);
+
+          SNPRINT(headings, "\tConsist\tWstCons");
+          SNPRINT2(results, "\t%7.3f", consistency);
+          SNPRINT2(results, "\t%7.3f", cpi->worst_consistency);
+        }
+
+        if (cpi->b_calculate_ssimg) {
+          SNPRINT(headings, "\t  SSIMG\tWtSSIMG");
+          SNPRINT2(results, "\t%7.3f", cpi->ssimg.stat[ALL] / cpi->count);
+          SNPRINT2(results, "\t%7.3f", cpi->ssimg.worst);
+        }
+
+        fprintf(f, "%s\t    Time  Rc-Err Abs Err\n", headings);
+        fprintf(f, "%s\t%8.0f %7.2f %7.2f\n", results,
+                total_encode_time, rate_err, fabs(rate_err));
+      }
+
+      fclose(f);
+    }
+
+#endif
+
+#if 0
+    {
+      printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000);
+      printf("\n_frames recive_data encod_mb_row compress_frame  Total\n");
+      printf("%6d %10ld %10ld %10ld %10ld\n", cpi->common.current_video_frame,
+             cpi->time_receive_data / 1000, cpi->time_encode_sb_row / 1000,
+             cpi->time_compress_data / 1000,
+             (cpi->time_receive_data + cpi->time_compress_data) / 1000);
+    }
+#endif
+  }
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  vp9_denoiser_free(&(cpi->denoiser));
+#endif
+
+  for (t = 0; t < cpi->num_workers; ++t) {
+    VPxWorker *const worker = &cpi->workers[t];
+    EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
+
+    // Deallocate allocated threads.
+    vpx_get_worker_interface()->end(worker);
+
+    // Deallocate allocated thread data.
+    if (t < cpi->num_workers - 1) {
+      vpx_free(thread_data->td->counts);
+      vp9_free_pc_tree(thread_data->td);
+      vpx_free(thread_data->td);
+    }
+  }
+  vpx_free(cpi->tile_thr_data);
+  vpx_free(cpi->workers);
+
+  if (cpi->num_workers > 1)
+    vp9_loop_filter_dealloc(&cpi->lf_row_sync);
+
+  dealloc_compressor_data(cpi);
+
+  for (i = 0; i < sizeof(cpi->mbgraph_stats) /
+                  sizeof(cpi->mbgraph_stats[0]); ++i) {
+    vpx_free(cpi->mbgraph_stats[i].mb_stats);
+  }
+
+#if CONFIG_FP_MB_STATS
+  if (cpi->use_fp_mb_stats) {
+    vpx_free(cpi->twopass.frame_mb_stats_buf);
+    cpi->twopass.frame_mb_stats_buf = NULL;
+  }
+#endif
+
+  vp9_remove_common(cm);
+  vp9_free_ref_frame_buffers(cm->buffer_pool);
+#if CONFIG_VP9_POSTPROC
+  vp9_free_postproc_buffers(cm);
+#endif
+  vpx_free(cpi);
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#ifdef OUTPUT_YUV_DENOISED
+  fclose(yuv_denoised_file);
+#endif
+#endif
+#ifdef OUTPUT_YUV_SKINMAP
+  fclose(yuv_skinmap_file);
+#endif
+#ifdef OUTPUT_YUV_REC
+  fclose(yuv_rec_file);
+#endif
+
+#if 0
+
+  if (keyfile)
+    fclose(keyfile);
+
+  if (framepsnr)
+    fclose(framepsnr);
+
+  if (kf_list)
+    fclose(kf_list);
+
+#endif
+}
+
+/* TODO(yaowu): The block_variance calls the unoptimized versions of variance()
+ * and highbd_8_variance(). It should not.
+ */
+static void encoder_variance(const uint8_t *a, int  a_stride,
+                             const uint8_t *b, int  b_stride,
+                             int  w, int  h, unsigned int *sse, int *sum) {
+  int i, j;
+
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void encoder_highbd_variance64(const uint8_t *a8, int  a_stride,
+                                      const uint8_t *b8, int  b_stride,
+                                      int w, int h, uint64_t *sse,
+                                      int64_t *sum) {
+  int i, j;
+
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+static void encoder_highbd_8_variance(const uint8_t *a8, int  a_stride,
+                                      const uint8_t *b8, int  b_stride,
+                                      int w, int h,
+                                      unsigned int *sse, int *sum) {
+  uint64_t sse_long = 0;
+  int64_t sum_long = 0;
+  encoder_highbd_variance64(a8, a_stride, b8, b_stride, w, h,
+                            &sse_long, &sum_long);
+  *sse = (unsigned int)sse_long;
+  *sum = (int)sum_long;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static int64_t get_sse(const uint8_t *a, int a_stride,
+                       const uint8_t *b, int b_stride,
+                       int width, int height) {
+  const int dw = width % 16;
+  const int dh = height % 16;
+  int64_t total_sse = 0;
+  unsigned int sse = 0;
+  int sum = 0;
+  int x, y;
+
+  if (dw > 0) {
+    encoder_variance(&a[width - dw], a_stride, &b[width - dw], b_stride,
+                     dw, height, &sse, &sum);
+    total_sse += sse;
+  }
+
+  if (dh > 0) {
+    encoder_variance(&a[(height - dh) * a_stride], a_stride,
+                     &b[(height - dh) * b_stride], b_stride,
+                     width - dw, dh, &sse, &sum);
+    total_sse += sse;
+  }
+
+  for (y = 0; y < height / 16; ++y) {
+    const uint8_t *pa = a;
+    const uint8_t *pb = b;
+    for (x = 0; x < width / 16; ++x) {
+      vpx_mse16x16(pa, a_stride, pb, b_stride, &sse);
+      total_sse += sse;
+
+      pa += 16;
+      pb += 16;
+    }
+
+    a += 16 * a_stride;
+    b += 16 * b_stride;
+  }
+
+  return total_sse;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static int64_t highbd_get_sse_shift(const uint8_t *a8, int a_stride,
+                                    const uint8_t *b8, int b_stride,
+                                    int width, int height,
+                                    unsigned int input_shift) {
+  const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  const uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  int64_t total_sse = 0;
+  int x, y;
+  for (y = 0; y < height; ++y) {
+    for (x = 0; x < width; ++x) {
+      int64_t diff;
+      diff = (a[x] >> input_shift) - (b[x] >> input_shift);
+      total_sse += diff * diff;
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+  return total_sse;
+}
+
+static int64_t highbd_get_sse(const uint8_t *a, int a_stride,
+                              const uint8_t *b, int b_stride,
+                              int width, int height) {
+  int64_t total_sse = 0;
+  int x, y;
+  const int dw = width % 16;
+  const int dh = height % 16;
+  unsigned int sse = 0;
+  int sum = 0;
+  if (dw > 0) {
+    encoder_highbd_8_variance(&a[width - dw], a_stride,
+                              &b[width - dw], b_stride,
+                              dw, height, &sse, &sum);
+    total_sse += sse;
+  }
+  if (dh > 0) {
+    encoder_highbd_8_variance(&a[(height - dh) * a_stride], a_stride,
+                              &b[(height - dh) * b_stride], b_stride,
+                              width - dw, dh, &sse, &sum);
+    total_sse += sse;
+  }
+  for (y = 0; y < height / 16; ++y) {
+    const uint8_t *pa = a;
+    const uint8_t *pb = b;
+    for (x = 0; x < width / 16; ++x) {
+      vpx_highbd_8_mse16x16(pa, a_stride, pb, b_stride, &sse);
+      total_sse += sse;
+      pa += 16;
+      pb += 16;
+    }
+    a += 16 * a_stride;
+    b += 16 * b_stride;
+  }
+  return total_sse;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+typedef struct {
+  double psnr[4];       // total/y/u/v
+  uint64_t sse[4];      // total/y/u/v
+  uint32_t samples[4];  // total/y/u/v
+} PSNR_STATS;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void calc_highbd_psnr(const YV12_BUFFER_CONFIG *a,
+                             const YV12_BUFFER_CONFIG *b,
+                             PSNR_STATS *psnr,
+                             unsigned int bit_depth,
+                             unsigned int in_bit_depth) {
+  const int widths[3] =
+      {a->y_crop_width,  a->uv_crop_width,  a->uv_crop_width };
+  const int heights[3] =
+      {a->y_crop_height, a->uv_crop_height, a->uv_crop_height};
+  const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer,  a->v_buffer };
+  const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride};
+  const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer,  b->v_buffer };
+  const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride};
+  int i;
+  uint64_t total_sse = 0;
+  uint32_t total_samples = 0;
+  const double peak = (double)((1 << in_bit_depth) - 1);
+  const unsigned int input_shift = bit_depth - in_bit_depth;
+
+  for (i = 0; i < 3; ++i) {
+    const int w = widths[i];
+    const int h = heights[i];
+    const uint32_t samples = w * h;
+    uint64_t sse;
+    if (a->flags & YV12_FLAG_HIGHBITDEPTH) {
+      if (input_shift) {
+        sse = highbd_get_sse_shift(a_planes[i], a_strides[i],
+                                   b_planes[i], b_strides[i], w, h,
+                                   input_shift);
+      } else {
+        sse = highbd_get_sse(a_planes[i], a_strides[i],
+                             b_planes[i], b_strides[i], w, h);
+      }
+    } else {
+      sse = get_sse(a_planes[i], a_strides[i],
+                    b_planes[i], b_strides[i],
+                    w, h);
+    }
+    psnr->sse[1 + i] = sse;
+    psnr->samples[1 + i] = samples;
+    psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse);
+
+    total_sse += sse;
+    total_samples += samples;
+  }
+
+  psnr->sse[0] = total_sse;
+  psnr->samples[0] = total_samples;
+  psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak,
+                                  (double)total_sse);
+}
+
+#else  // !CONFIG_VP9_HIGHBITDEPTH
+
+static void calc_psnr(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b,
+                      PSNR_STATS *psnr) {
+  static const double peak = 255.0;
+  const int widths[3]        = {
+      a->y_crop_width, a->uv_crop_width, a->uv_crop_width};
+  const int heights[3]       = {
+      a->y_crop_height, a->uv_crop_height, a->uv_crop_height};
+  const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer};
+  const int a_strides[3]     = {a->y_stride, a->uv_stride, a->uv_stride};
+  const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer};
+  const int b_strides[3]     = {b->y_stride, b->uv_stride, b->uv_stride};
+  int i;
+  uint64_t total_sse = 0;
+  uint32_t total_samples = 0;
+
+  for (i = 0; i < 3; ++i) {
+    const int w = widths[i];
+    const int h = heights[i];
+    const uint32_t samples = w * h;
+    const uint64_t sse = get_sse(a_planes[i], a_strides[i],
+                                 b_planes[i], b_strides[i],
+                                 w, h);
+    psnr->sse[1 + i] = sse;
+    psnr->samples[1 + i] = samples;
+    psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse);
+
+    total_sse += sse;
+    total_samples += samples;
+  }
+
+  psnr->sse[0] = total_sse;
+  psnr->samples[0] = total_samples;
+  psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak,
+                                  (double)total_sse);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void generate_psnr_packet(VP9_COMP *cpi) {
+  struct vpx_codec_cx_pkt pkt;
+  int i;
+  PSNR_STATS psnr;
+#if CONFIG_VP9_HIGHBITDEPTH
+  calc_highbd_psnr(cpi->Source, cpi->common.frame_to_show, &psnr,
+                   cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth);
+#else
+  calc_psnr(cpi->Source, cpi->common.frame_to_show, &psnr);
+#endif
+
+  for (i = 0; i < 4; ++i) {
+    pkt.data.psnr.samples[i] = psnr.samples[i];
+    pkt.data.psnr.sse[i] = psnr.sse[i];
+    pkt.data.psnr.psnr[i] = psnr.psnr[i];
+  }
+  pkt.kind = VPX_CODEC_PSNR_PKT;
+  if (cpi->use_svc)
+    cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+        cpi->svc.number_temporal_layers].psnr_pkt = pkt.data.psnr;
+  else
+    vpx_codec_pkt_list_add(cpi->output_pkt_list, &pkt);
+}
+
+int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags) {
+  if (ref_frame_flags > 7)
+    return -1;
+
+  cpi->ref_frame_flags = ref_frame_flags;
+  return 0;
+}
+
+void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags) {
+  cpi->ext_refresh_golden_frame = (ref_frame_flags & VP9_GOLD_FLAG) != 0;
+  cpi->ext_refresh_alt_ref_frame = (ref_frame_flags & VP9_ALT_FLAG) != 0;
+  cpi->ext_refresh_last_frame = (ref_frame_flags & VP9_LAST_FLAG) != 0;
+  cpi->ext_refresh_frame_flags_pending = 1;
+}
+
+static YV12_BUFFER_CONFIG *get_vp9_ref_frame_buffer(VP9_COMP *cpi,
+                                VP9_REFFRAME ref_frame_flag) {
+  MV_REFERENCE_FRAME ref_frame = NONE;
+  if (ref_frame_flag == VP9_LAST_FLAG)
+    ref_frame = LAST_FRAME;
+  else if (ref_frame_flag == VP9_GOLD_FLAG)
+    ref_frame = GOLDEN_FRAME;
+  else if (ref_frame_flag == VP9_ALT_FLAG)
+    ref_frame = ALTREF_FRAME;
+
+  return ref_frame == NONE ? NULL : get_ref_frame_buffer(cpi, ref_frame);
+}
+
+int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
+                           YV12_BUFFER_CONFIG *sd) {
+  YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
+  if (cfg) {
+    vp8_yv12_copy_frame(cfg, sd);
+    return 0;
+  } else {
+    return -1;
+  }
+}
+
+int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
+                          YV12_BUFFER_CONFIG *sd) {
+  YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
+  if (cfg) {
+    vp8_yv12_copy_frame(sd, cfg);
+    return 0;
+  } else {
+    return -1;
+  }
+}
+
+int vp9_update_entropy(VP9_COMP * cpi, int update) {
+  cpi->ext_refresh_frame_context = update;
+  cpi->ext_refresh_frame_context_pending = 1;
+  return 0;
+}
+
+#if defined(OUTPUT_YUV_DENOISED) || defined(OUTPUT_YUV_SKINMAP)
+// The denoiser buffer is allocated as a YUV 440 buffer. This function writes it
+// as YUV 420. We simply use the top-left pixels of the UV buffers, since we do
+// not denoise the UV channels at this time. If ever we implement UV channel
+// denoising we will have to modify this.
+void vp9_write_yuv_frame_420(YV12_BUFFER_CONFIG *s, FILE *f) {
+  uint8_t *src = s->y_buffer;
+  int h = s->y_height;
+
+  do {
+    fwrite(src, s->y_width, 1, f);
+    src += s->y_stride;
+  } while (--h);
+
+  src = s->u_buffer;
+  h = s->uv_height;
+
+  do {
+    fwrite(src, s->uv_width, 1, f);
+    src += s->uv_stride;
+  } while (--h);
+
+  src = s->v_buffer;
+  h = s->uv_height;
+
+  do {
+    fwrite(src, s->uv_width, 1, f);
+    src += s->uv_stride;
+  } while (--h);
+}
+#endif
+
+#ifdef OUTPUT_YUV_REC
+void vp9_write_yuv_rec_frame(VP9_COMMON *cm) {
+  YV12_BUFFER_CONFIG *s = cm->frame_to_show;
+  uint8_t *src = s->y_buffer;
+  int h = cm->height;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (s->flags & YV12_FLAG_HIGHBITDEPTH) {
+    uint16_t *src16 = CONVERT_TO_SHORTPTR(s->y_buffer);
+
+    do {
+      fwrite(src16, s->y_width, 2,  yuv_rec_file);
+      src16 += s->y_stride;
+    } while (--h);
+
+    src16 = CONVERT_TO_SHORTPTR(s->u_buffer);
+    h = s->uv_height;
+
+    do {
+      fwrite(src16, s->uv_width, 2,  yuv_rec_file);
+      src16 += s->uv_stride;
+    } while (--h);
+
+    src16 = CONVERT_TO_SHORTPTR(s->v_buffer);
+    h = s->uv_height;
+
+    do {
+      fwrite(src16, s->uv_width, 2, yuv_rec_file);
+      src16 += s->uv_stride;
+    } while (--h);
+
+    fflush(yuv_rec_file);
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  do {
+    fwrite(src, s->y_width, 1,  yuv_rec_file);
+    src += s->y_stride;
+  } while (--h);
+
+  src = s->u_buffer;
+  h = s->uv_height;
+
+  do {
+    fwrite(src, s->uv_width, 1,  yuv_rec_file);
+    src += s->uv_stride;
+  } while (--h);
+
+  src = s->v_buffer;
+  h = s->uv_height;
+
+  do {
+    fwrite(src, s->uv_width, 1, yuv_rec_file);
+    src += s->uv_stride;
+  } while (--h);
+
+  fflush(yuv_rec_file);
+}
+#endif
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
+                                                YV12_BUFFER_CONFIG *dst,
+                                                int bd) {
+#else
+static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
+                                                YV12_BUFFER_CONFIG *dst) {
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  // TODO(dkovalev): replace YV12_BUFFER_CONFIG with vpx_image_t
+  int i;
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  const int src_widths[3] = {src->y_crop_width, src->uv_crop_width,
+                             src->uv_crop_width };
+  const int src_heights[3] = {src->y_crop_height, src->uv_crop_height,
+                              src->uv_crop_height};
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
+  const int dst_widths[3] = {dst->y_crop_width, dst->uv_crop_width,
+                             dst->uv_crop_width};
+  const int dst_heights[3] = {dst->y_crop_height, dst->uv_crop_height,
+                              dst->uv_crop_height};
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+      vp9_highbd_resize_plane(srcs[i], src_heights[i], src_widths[i],
+                              src_strides[i], dsts[i], dst_heights[i],
+                              dst_widths[i], dst_strides[i], bd);
+    } else {
+      vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
+                       dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
+    }
+#else
+    vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
+                     dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+  vpx_extend_frame_borders(dst);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                                   YV12_BUFFER_CONFIG *dst, int bd) {
+  const int src_w = src->y_crop_width;
+  const int src_h = src->y_crop_height;
+  const int dst_w = dst->y_crop_width;
+  const int dst_h = dst->y_crop_height;
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
+  const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
+  int x, y, i;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    const int factor = (i == 0 || i == 3 ? 1 : 2);
+    const int src_stride = src_strides[i];
+    const int dst_stride = dst_strides[i];
+    for (y = 0; y < dst_h; y += 16) {
+      const int y_q4 = y * (16 / factor) * src_h / dst_h;
+      for (x = 0; x < dst_w; x += 16) {
+        const int x_q4 = x * (16 / factor) * src_w / dst_w;
+        const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h *
+                                   src_stride + (x / factor) * src_w / dst_w;
+        uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
+
+        if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+          vpx_highbd_convolve8(src_ptr, src_stride, dst_ptr, dst_stride,
+                               kernel[x_q4 & 0xf], 16 * src_w / dst_w,
+                               kernel[y_q4 & 0xf], 16 * src_h / dst_h,
+                               16 / factor, 16 / factor, bd);
+        } else {
+          vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
+                        kernel[x_q4 & 0xf], 16 * src_w / dst_w,
+                        kernel[y_q4 & 0xf], 16 * src_h / dst_h,
+                        16 / factor, 16 / factor);
+        }
+      }
+    }
+  }
+
+  vpx_extend_frame_borders(dst);
+}
+#else
+void vp9_scale_and_extend_frame_c(const YV12_BUFFER_CONFIG *src,
+                                  YV12_BUFFER_CONFIG *dst) {
+  const int src_w = src->y_crop_width;
+  const int src_h = src->y_crop_height;
+  const int dst_w = dst->y_crop_width;
+  const int dst_h = dst->y_crop_height;
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
+  const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
+  int x, y, i;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    const int factor = (i == 0 || i == 3 ? 1 : 2);
+    const int src_stride = src_strides[i];
+    const int dst_stride = dst_strides[i];
+    for (y = 0; y < dst_h; y += 16) {
+      const int y_q4 = y * (16 / factor) * src_h / dst_h;
+      for (x = 0; x < dst_w; x += 16) {
+        const int x_q4 = x * (16 / factor) * src_w / dst_w;
+        const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h *
+                                   src_stride + (x / factor) * src_w / dst_w;
+        uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
+
+        vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
+                      kernel[x_q4 & 0xf], 16 * src_w / dst_w,
+                      kernel[y_q4 & 0xf], 16 * src_h / dst_h,
+                      16 / factor, 16 / factor);
+      }
+    }
+  }
+
+  vpx_extend_frame_borders(dst);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static int scale_down(VP9_COMP *cpi, int q) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+  int scale = 0;
+  assert(frame_is_kf_gf_arf(cpi));
+
+  if (rc->frame_size_selector == UNSCALED &&
+      q >= rc->rf_level_maxq[gf_group->rf_level[gf_group->index]]) {
+    const int max_size_thresh = (int)(rate_thresh_mult[SCALE_STEP1]
+        * VPXMAX(rc->this_frame_target, rc->avg_frame_bandwidth));
+    scale = rc->projected_frame_size > max_size_thresh ? 1 : 0;
+  }
+  return scale;
+}
+
+static int big_rate_miss(VP9_COMP *cpi, int high_limit, int low_limit) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+
+  return (rc->projected_frame_size > ((high_limit * 3) / 2)) ||
+         (rc->projected_frame_size < (low_limit / 2));
+}
+
+// Function to test for conditions that indicate we should loop
+// back and recode a frame.
+static int recode_loop_test(VP9_COMP *cpi,
+                            int high_limit, int low_limit,
+                            int q, int maxq, int minq) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const int frame_is_kfgfarf = frame_is_kf_gf_arf(cpi);
+  int force_recode = 0;
+
+  if ((rc->projected_frame_size >= rc->max_frame_bandwidth) ||
+      big_rate_miss(cpi, high_limit, low_limit) ||
+      (cpi->sf.recode_loop == ALLOW_RECODE) ||
+      (frame_is_kfgfarf &&
+       (cpi->sf.recode_loop == ALLOW_RECODE_KFARFGF))) {
+    if (frame_is_kfgfarf &&
+        (oxcf->resize_mode == RESIZE_DYNAMIC) &&
+        scale_down(cpi, q)) {
+        // Code this group at a lower resolution.
+        cpi->resize_pending = 1;
+        return 1;
+    }
+
+    // TODO(agrange) high_limit could be greater than the scale-down threshold.
+    if ((rc->projected_frame_size > high_limit && q < maxq) ||
+        (rc->projected_frame_size < low_limit && q > minq)) {
+      force_recode = 1;
+    } else if (cpi->oxcf.rc_mode == VPX_CQ) {
+      // Deal with frame undershoot and whether or not we are
+      // below the automatically set cq level.
+      if (q > oxcf->cq_level &&
+          rc->projected_frame_size < ((rc->this_frame_target * 7) >> 3)) {
+        force_recode = 1;
+      }
+    }
+  }
+  return force_recode;
+}
+
+void vp9_update_reference_frames(VP9_COMP *cpi) {
+  VP9_COMMON * const cm = &cpi->common;
+  BufferPool *const pool = cm->buffer_pool;
+
+  // At this point the new frame has been encoded.
+  // If any buffer copy / swapping is signaled it should be done here.
+  if (cm->frame_type == KEY_FRAME) {
+    ref_cnt_fb(pool->frame_bufs,
+               &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
+    ref_cnt_fb(pool->frame_bufs,
+               &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
+  } else if (vp9_preserve_existing_gf(cpi)) {
+    // We have decided to preserve the previously existing golden frame as our
+    // new ARF frame. However, in the short term in function
+    // vp9_get_refresh_mask() we left it in the GF slot and, if
+    // we're updating the GF with the current decoded frame, we save it to the
+    // ARF slot instead.
+    // We now have to update the ARF with the current frame and swap gld_fb_idx
+    // and alt_fb_idx so that, overall, we've stored the old GF in the new ARF
+    // slot and, if we're updating the GF, the current frame becomes the new GF.
+    int tmp;
+
+    ref_cnt_fb(pool->frame_bufs,
+               &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
+
+    tmp = cpi->alt_fb_idx;
+    cpi->alt_fb_idx = cpi->gld_fb_idx;
+    cpi->gld_fb_idx = tmp;
+
+    if (is_two_pass_svc(cpi)) {
+      cpi->svc.layer_context[0].gold_ref_idx = cpi->gld_fb_idx;
+      cpi->svc.layer_context[0].alt_ref_idx = cpi->alt_fb_idx;
+    }
+  } else { /* For non key/golden frames */
+    if (cpi->refresh_alt_ref_frame) {
+      int arf_idx = cpi->alt_fb_idx;
+      if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
+        const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+        arf_idx = gf_group->arf_update_idx[gf_group->index];
+      }
+
+      ref_cnt_fb(pool->frame_bufs,
+                 &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
+      memcpy(cpi->interp_filter_selected[ALTREF_FRAME],
+             cpi->interp_filter_selected[0],
+             sizeof(cpi->interp_filter_selected[0]));
+    }
+
+    if (cpi->refresh_golden_frame) {
+      ref_cnt_fb(pool->frame_bufs,
+                 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
+      if (!cpi->rc.is_src_frame_alt_ref)
+        memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
+               cpi->interp_filter_selected[0],
+               sizeof(cpi->interp_filter_selected[0]));
+      else
+        memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
+               cpi->interp_filter_selected[ALTREF_FRAME],
+               sizeof(cpi->interp_filter_selected[ALTREF_FRAME]));
+    }
+  }
+
+  if (cpi->refresh_last_frame) {
+    ref_cnt_fb(pool->frame_bufs,
+               &cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx);
+    if (!cpi->rc.is_src_frame_alt_ref)
+      memcpy(cpi->interp_filter_selected[LAST_FRAME],
+             cpi->interp_filter_selected[0],
+             sizeof(cpi->interp_filter_selected[0]));
+  }
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0 &&
+      cpi->denoiser.denoising_level > kDenLowLow) {
+    vp9_denoiser_update_frame_info(&cpi->denoiser,
+                                   *cpi->Source,
+                                   cpi->common.frame_type,
+                                   cpi->refresh_alt_ref_frame,
+                                   cpi->refresh_golden_frame,
+                                   cpi->refresh_last_frame,
+                                   cpi->resize_pending);
+  }
+#endif
+  if (is_one_pass_cbr_svc(cpi)) {
+    // Keep track of frame index for each reference frame.
+    SVC *const svc = &cpi->svc;
+    if (cm->frame_type == KEY_FRAME) {
+      svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;
+      svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;
+      svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;
+    } else {
+      if (cpi->refresh_last_frame)
+        svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;
+      if (cpi->refresh_golden_frame)
+        svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;
+      if (cpi->refresh_alt_ref_frame)
+        svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;
+    }
+  }
+}
+
+static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
+  MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
+  struct loopfilter *lf = &cm->lf;
+
+  if (xd->lossless) {
+      lf->filter_level = 0;
+      lf->last_filt_level = 0;
+  } else {
+    struct vpx_usec_timer timer;
+
+    vpx_clear_system_state();
+
+    vpx_usec_timer_start(&timer);
+
+    if (!cpi->rc.is_src_frame_alt_ref) {
+      if ((cpi->common.frame_type == KEY_FRAME) &&
+          (!cpi->rc.this_key_frame_forced)) {
+        lf->last_filt_level = 0;
+      }
+      vp9_pick_filter_level(cpi->Source, cpi, cpi->sf.lpf_pick);
+      lf->last_filt_level = lf->filter_level;
+    } else {
+      lf->filter_level = 0;
+    }
+
+    vpx_usec_timer_mark(&timer);
+    cpi->time_pick_lpf += vpx_usec_timer_elapsed(&timer);
+  }
+
+  if (lf->filter_level > 0) {
+    vp9_build_mask_frame(cm, lf->filter_level, 0);
+
+    if (cpi->num_workers > 1)
+      vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane,
+                               lf->filter_level, 0, 0,
+                               cpi->workers, cpi->num_workers,
+                               &cpi->lf_row_sync);
+    else
+      vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
+  }
+
+  vpx_extend_frame_inner_borders(cm->frame_to_show);
+}
+
+static INLINE void alloc_frame_mvs(VP9_COMMON *const cm,
+                                   int buffer_idx) {
+  RefCntBuffer *const new_fb_ptr = &cm->buffer_pool->frame_bufs[buffer_idx];
+  if (new_fb_ptr->mvs == NULL ||
+      new_fb_ptr->mi_rows < cm->mi_rows ||
+      new_fb_ptr->mi_cols < cm->mi_cols) {
+    vpx_free(new_fb_ptr->mvs);
+    CHECK_MEM_ERROR(cm, new_fb_ptr->mvs,
+                    (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
+                                         sizeof(*new_fb_ptr->mvs)));
+    new_fb_ptr->mi_rows = cm->mi_rows;
+    new_fb_ptr->mi_cols = cm->mi_cols;
+  }
+}
+
+void vp9_scale_references(VP9_COMP *cpi) {
+  VP9_COMMON *cm = &cpi->common;
+  MV_REFERENCE_FRAME ref_frame;
+  const VP9_REFFRAME ref_mask[3] = {VP9_LAST_FLAG, VP9_GOLD_FLAG, VP9_ALT_FLAG};
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    // Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1).
+    if (cpi->ref_frame_flags & ref_mask[ref_frame - 1]) {
+      BufferPool *const pool = cm->buffer_pool;
+      const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi,
+                                                                 ref_frame);
+
+      if (ref == NULL) {
+        cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
+        continue;
+      }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
+        RefCntBuffer *new_fb_ptr = NULL;
+        int force_scaling = 0;
+        int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
+        if (new_fb == INVALID_IDX) {
+          new_fb = get_free_fb(cm);
+          force_scaling = 1;
+        }
+        if (new_fb == INVALID_IDX)
+          return;
+        new_fb_ptr = &pool->frame_bufs[new_fb];
+        if (force_scaling ||
+            new_fb_ptr->buf.y_crop_width != cm->width ||
+            new_fb_ptr->buf.y_crop_height != cm->height) {
+          if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
+                                       cm->subsampling_x, cm->subsampling_y,
+                                       cm->use_highbitdepth,
+                                       VP9_ENC_BORDER_IN_PIXELS,
+                                       cm->byte_alignment, NULL, NULL, NULL))
+            vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                               "Failed to allocate frame buffer");
+          scale_and_extend_frame(ref, &new_fb_ptr->buf, (int)cm->bit_depth);
+          cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
+          alloc_frame_mvs(cm, new_fb);
+        }
+#else
+      if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
+        RefCntBuffer *new_fb_ptr = NULL;
+        int force_scaling = 0;
+        int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
+        if (new_fb == INVALID_IDX) {
+          new_fb = get_free_fb(cm);
+          force_scaling = 1;
+        }
+        if (new_fb == INVALID_IDX)
+          return;
+        new_fb_ptr = &pool->frame_bufs[new_fb];
+        if (force_scaling ||
+            new_fb_ptr->buf.y_crop_width != cm->width ||
+            new_fb_ptr->buf.y_crop_height != cm->height) {
+          if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
+                                       cm->subsampling_x, cm->subsampling_y,
+                                       VP9_ENC_BORDER_IN_PIXELS,
+                                       cm->byte_alignment, NULL, NULL, NULL))
+            vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                               "Failed to allocate frame buffer");
+          vp9_scale_and_extend_frame(ref, &new_fb_ptr->buf);
+          cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
+          alloc_frame_mvs(cm, new_fb);
+        }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      } else {
+        int buf_idx;
+        RefCntBuffer *buf = NULL;
+        if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
+          // Check for release of scaled reference.
+          buf_idx = cpi->scaled_ref_idx[ref_frame - 1];
+          buf = (buf_idx != INVALID_IDX) ? &pool->frame_bufs[buf_idx] : NULL;
+          if (buf != NULL) {
+            --buf->ref_count;
+            cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
+          }
+        }
+        buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+        buf = &pool->frame_bufs[buf_idx];
+        buf->buf.y_crop_width = ref->y_crop_width;
+        buf->buf.y_crop_height = ref->y_crop_height;
+        cpi->scaled_ref_idx[ref_frame - 1] = buf_idx;
+        ++buf->ref_count;
+      }
+    } else {
+      if (cpi->oxcf.pass != 0 || cpi->use_svc)
+        cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
+    }
+  }
+}
+
+static void release_scaled_references(VP9_COMP *cpi) {
+  VP9_COMMON *cm = &cpi->common;
+  int i;
+  if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
+    // Only release scaled references under certain conditions:
+    // if reference will be updated, or if scaled reference has same resolution.
+    int refresh[3];
+    refresh[0] = (cpi->refresh_last_frame) ? 1 : 0;
+    refresh[1] = (cpi->refresh_golden_frame) ? 1 : 0;
+    refresh[2] = (cpi->refresh_alt_ref_frame) ? 1 : 0;
+    for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
+      const int idx = cpi->scaled_ref_idx[i - 1];
+      RefCntBuffer *const buf = idx != INVALID_IDX ?
+          &cm->buffer_pool->frame_bufs[idx] : NULL;
+      const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, i);
+      if (buf != NULL &&
+          (refresh[i - 1] ||
+          (buf->buf.y_crop_width == ref->y_crop_width &&
+           buf->buf.y_crop_height == ref->y_crop_height))) {
+        --buf->ref_count;
+        cpi->scaled_ref_idx[i -1] = INVALID_IDX;
+      }
+    }
+  } else {
+    for (i = 0; i < MAX_REF_FRAMES; ++i) {
+      const int idx = cpi->scaled_ref_idx[i];
+      RefCntBuffer *const buf = idx != INVALID_IDX ?
+          &cm->buffer_pool->frame_bufs[idx] : NULL;
+      if (buf != NULL) {
+        --buf->ref_count;
+        cpi->scaled_ref_idx[i] = INVALID_IDX;
+      }
+    }
+  }
+}
+
+static void full_to_model_count(unsigned int *model_count,
+                                unsigned int *full_count) {
+  int n;
+  model_count[ZERO_TOKEN] = full_count[ZERO_TOKEN];
+  model_count[ONE_TOKEN] = full_count[ONE_TOKEN];
+  model_count[TWO_TOKEN] = full_count[TWO_TOKEN];
+  for (n = THREE_TOKEN; n < EOB_TOKEN; ++n)
+    model_count[TWO_TOKEN] += full_count[n];
+  model_count[EOB_MODEL_TOKEN] = full_count[EOB_TOKEN];
+}
+
+static void full_to_model_counts(vp9_coeff_count_model *model_count,
+                                 vp9_coeff_count *full_count) {
+  int i, j, k, l;
+
+  for (i = 0; i < PLANE_TYPES; ++i)
+    for (j = 0; j < REF_TYPES; ++j)
+      for (k = 0; k < COEF_BANDS; ++k)
+        for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
+          full_to_model_count(model_count[i][j][k][l], full_count[i][j][k][l]);
+}
+
+#if 0 && CONFIG_INTERNAL_STATS
+static void output_frame_level_debug_stats(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  FILE *const f = fopen("tmp.stt", cm->current_video_frame ? "a" : "w");
+  int64_t recon_err;
+
+  vpx_clear_system_state();
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    recon_err = vp9_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
+  } else {
+    recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
+  }
+#else
+  recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+
+  if (cpi->twopass.total_left_stats.coded_error != 0.0) {
+    double dc_quant_devisor;
+#if CONFIG_VP9_HIGHBITDEPTH
+    switch (cm->bit_depth) {
+      case VPX_BITS_8:
+        dc_quant_devisor = 4.0;
+        break;
+      case VPX_BITS_10:
+        dc_quant_devisor = 16.0;
+        break;
+      case VPX_BITS_12:
+        dc_quant_devisor = 64.0;
+        break;
+      default:
+        assert(0 && "bit_depth must be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+        break;
+    }
+#else
+    dc_quant_devisor = 4.0;
+#endif
+
+    fprintf(f, "%10u %dx%d %10d %10d %d %d %10d %10d %10d %10d"
+       "%10"PRId64" %10"PRId64" %5d %5d %10"PRId64" "
+       "%10"PRId64" %10"PRId64" %10d "
+       "%7.2lf %7.2lf %7.2lf %7.2lf %7.2lf"
+        "%6d %6d %5d %5d %5d "
+        "%10"PRId64" %10.3lf"
+        "%10lf %8u %10"PRId64" %10d %10d %10d %10d %10d\n",
+        cpi->common.current_video_frame,
+        cm->width, cm->height,
+        cpi->td.rd_counts.m_search_count,
+        cpi->td.rd_counts.ex_search_count,
+        cpi->rc.source_alt_ref_pending,
+        cpi->rc.source_alt_ref_active,
+        cpi->rc.this_frame_target,
+        cpi->rc.projected_frame_size,
+        cpi->rc.projected_frame_size / cpi->common.MBs,
+        (cpi->rc.projected_frame_size - cpi->rc.this_frame_target),
+        cpi->rc.vbr_bits_off_target,
+        cpi->rc.vbr_bits_off_target_fast,
+        cpi->twopass.extend_minq,
+        cpi->twopass.extend_minq_fast,
+        cpi->rc.total_target_vs_actual,
+        (cpi->rc.starting_buffer_level - cpi->rc.bits_off_target),
+        cpi->rc.total_actual_bits, cm->base_qindex,
+        vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth),
+        (double)vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth) /
+            dc_quant_devisor,
+        vp9_convert_qindex_to_q(cpi->twopass.active_worst_quality,
+                                cm->bit_depth),
+        cpi->rc.avg_q,
+        vp9_convert_qindex_to_q(cpi->oxcf.cq_level, cm->bit_depth),
+        cpi->refresh_last_frame, cpi->refresh_golden_frame,
+        cpi->refresh_alt_ref_frame, cm->frame_type, cpi->rc.gfu_boost,
+        cpi->twopass.bits_left,
+        cpi->twopass.total_left_stats.coded_error,
+        cpi->twopass.bits_left /
+            (1 + cpi->twopass.total_left_stats.coded_error),
+        cpi->tot_recode_hits, recon_err, cpi->rc.kf_boost,
+        cpi->twopass.kf_zeromotion_pct,
+        cpi->twopass.fr_content_type,
+        cm->lf.filter_level,
+        cm->seg.aq_av_offset);
+  }
+  fclose(f);
+
+  if (0) {
+    FILE *const fmodes = fopen("Modes.stt", "a");
+    int i;
+
+    fprintf(fmodes, "%6d:%1d:%1d:%1d ", cpi->common.current_video_frame,
+            cm->frame_type, cpi->refresh_golden_frame,
+            cpi->refresh_alt_ref_frame);
+
+    for (i = 0; i < MAX_MODES; ++i)
+      fprintf(fmodes, "%5d ", cpi->mode_chosen_counts[i]);
+
+    fprintf(fmodes, "\n");
+
+    fclose(fmodes);
+  }
+}
+#endif
+
+static void set_mv_search_params(VP9_COMP *cpi) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const unsigned int max_mv_def = VPXMIN(cm->width, cm->height);
+
+  // Default based on max resolution.
+  cpi->mv_step_param = vp9_init_search_range(max_mv_def);
+
+  if (cpi->sf.mv.auto_mv_step_size) {
+    if (frame_is_intra_only(cm)) {
+      // Initialize max_mv_magnitude for use in the first INTER frame
+      // after a key/intra-only frame.
+      cpi->max_mv_magnitude = max_mv_def;
+    } else {
+      if (cm->show_frame) {
+        // Allow mv_steps to correspond to twice the max mv magnitude found
+        // in the previous frame, capped by the default max_mv_magnitude based
+        // on resolution.
+        cpi->mv_step_param = vp9_init_search_range(
+            VPXMIN(max_mv_def, 2 * cpi->max_mv_magnitude));
+      }
+      cpi->max_mv_magnitude = 0;
+    }
+  }
+}
+
+static void set_size_independent_vars(VP9_COMP *cpi) {
+  vp9_set_speed_features_framesize_independent(cpi);
+  vp9_set_rd_speed_thresholds(cpi);
+  vp9_set_rd_speed_thresholds_sub8x8(cpi);
+  cpi->common.interp_filter = cpi->sf.default_interp_filter;
+}
+
+static void set_size_dependent_vars(VP9_COMP *cpi, int *q,
+                                    int *bottom_index, int *top_index) {
+  VP9_COMMON *const cm = &cpi->common;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+
+  // Setup variables that depend on the dimensions of the frame.
+  vp9_set_speed_features_framesize_dependent(cpi);
+
+  // Decide q and q bounds.
+  *q = vp9_rc_pick_q_and_bounds(cpi, bottom_index, top_index);
+
+  if (!frame_is_intra_only(cm)) {
+    vp9_set_high_precision_mv(cpi, (*q) < HIGH_PRECISION_MV_QTHRESH);
+  }
+
+  // Configure experimental use of segmentation for enhanced coding of
+  // static regions if indicated.
+  // Only allowed in the second pass of a two pass encode, as it requires
+  // lagged coding, and if the relevant speed feature flag is set.
+  if (oxcf->pass == 2 && cpi->sf.static_segmentation)
+    configure_static_seg_features(cpi);
+
+#if CONFIG_VP9_POSTPROC && !(CONFIG_VP9_TEMPORAL_DENOISING)
+  if (oxcf->noise_sensitivity > 0) {
+    int l = 0;
+    switch (oxcf->noise_sensitivity) {
+      case 1:
+        l = 20;
+        break;
+      case 2:
+        l = 40;
+        break;
+      case 3:
+        l = 60;
+        break;
+      case 4:
+      case 5:
+        l = 100;
+        break;
+      case 6:
+        l = 150;
+        break;
+    }
+    vp9_denoise(cpi->Source, cpi->Source, l);
+  }
+#endif  // CONFIG_VP9_POSTPROC
+}
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+static void setup_denoiser_buffer(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  if (cpi->oxcf.noise_sensitivity > 0 &&
+      !cpi->denoiser.frame_buffer_initialized) {
+    if (vp9_denoiser_alloc(&cpi->denoiser, cm->width, cm->height,
+                           cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                           cm->use_highbitdepth,
+#endif
+                           VP9_ENC_BORDER_IN_PIXELS))
+      vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                         "Failed to allocate denoiser");
+  }
+}
+#endif
+
+static void init_motion_estimation(VP9_COMP *cpi) {
+  int y_stride = cpi->scaled_source.y_stride;
+
+  if (cpi->sf.mv.search_method == NSTEP) {
+    vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride);
+  } else if (cpi->sf.mv.search_method == DIAMOND) {
+    vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
+  }
+}
+
+static void set_frame_size(VP9_COMP *cpi) {
+  int ref_frame;
+  VP9_COMMON *const cm = &cpi->common;
+  VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+
+  if (oxcf->pass == 2 &&
+      oxcf->rc_mode == VPX_VBR &&
+      ((oxcf->resize_mode == RESIZE_FIXED && cm->current_video_frame == 0) ||
+        (oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending))) {
+    calculate_coded_size(
+        cpi, &oxcf->scaled_frame_width, &oxcf->scaled_frame_height);
+
+    // There has been a change in frame size.
+    vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
+                         oxcf->scaled_frame_height);
+  }
+
+  if (oxcf->pass == 0 &&
+      oxcf->rc_mode == VPX_CBR &&
+      !cpi->use_svc &&
+      oxcf->resize_mode == RESIZE_DYNAMIC &&
+      cpi->resize_pending != 0) {
+    oxcf->scaled_frame_width =
+        (oxcf->width * cpi->resize_scale_num) / cpi->resize_scale_den;
+    oxcf->scaled_frame_height =
+        (oxcf->height * cpi->resize_scale_num) /cpi->resize_scale_den;
+    // There has been a change in frame size.
+    vp9_set_size_literal(cpi,
+                         oxcf->scaled_frame_width,
+                         oxcf->scaled_frame_height);
+
+    // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
+    set_mv_search_params(cpi);
+
+    vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
+#if CONFIG_VP9_TEMPORAL_DENOISING
+    // Reset the denoiser on the resized frame.
+    if (cpi->oxcf.noise_sensitivity > 0) {
+      vp9_denoiser_free(&(cpi->denoiser));
+      setup_denoiser_buffer(cpi);
+      // Dynamic resize is only triggered for non-SVC, so we can force
+      // golden frame update here as temporary fix to denoiser.
+      cpi->refresh_golden_frame = 1;
+    }
+#endif
+  }
+
+  if ((oxcf->pass == 2) &&
+      (!cpi->use_svc ||
+          (is_two_pass_svc(cpi) &&
+              cpi->svc.encode_empty_frame_state != ENCODING))) {
+    vp9_set_target_rate(cpi);
+  }
+
+  alloc_frame_mvs(cm, cm->new_fb_idx);
+
+  // Reset the frame pointers to the current frame size.
+  if (vpx_realloc_frame_buffer(get_frame_new_buffer(cm), cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL))
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate frame buffer");
+
+  alloc_util_frame_buffers(cpi);
+  init_motion_estimation(cpi);
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
+    const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+
+    ref_buf->idx = buf_idx;
+
+    if (buf_idx != INVALID_IDX) {
+      YV12_BUFFER_CONFIG *const buf = &cm->buffer_pool->frame_bufs[buf_idx].buf;
+      ref_buf->buf = buf;
+#if CONFIG_VP9_HIGHBITDEPTH
+      vp9_setup_scale_factors_for_frame(&ref_buf->sf,
+                                        buf->y_crop_width, buf->y_crop_height,
+                                        cm->width, cm->height,
+                                        (buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
+                                            1 : 0);
+#else
+      vp9_setup_scale_factors_for_frame(&ref_buf->sf,
+                                        buf->y_crop_width, buf->y_crop_height,
+                                        cm->width, cm->height);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      if (vp9_is_scaled(&ref_buf->sf))
+        vpx_extend_frame_borders(buf);
+    } else {
+      ref_buf->buf = NULL;
+    }
+  }
+
+  set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
+}
+
+static void encode_without_recode_loop(VP9_COMP *cpi,
+                                       size_t *size,
+                                       uint8_t *dest) {
+  VP9_COMMON *const cm = &cpi->common;
+  int q = 0, bottom_index = 0, top_index = 0;  // Dummy variables.
+
+  vpx_clear_system_state();
+
+  set_frame_size(cpi);
+
+  if (is_one_pass_cbr_svc(cpi) &&
+      cpi->un_scaled_source->y_width == cm->width << 2 &&
+      cpi->un_scaled_source->y_height == cm->height << 2 &&
+      cpi->svc.scaled_temp.y_width == cm->width << 1 &&
+      cpi->svc.scaled_temp.y_height == cm->height << 1) {
+    cpi->Source = vp9_svc_twostage_scale(cm,
+                                         cpi->un_scaled_source,
+                                         &cpi->scaled_source,
+                                         &cpi->svc.scaled_temp);
+  } else {
+    cpi->Source = vp9_scale_if_required(cm,
+                                        cpi->un_scaled_source,
+                                        &cpi->scaled_source,
+                                        (cpi->oxcf.pass == 0));
+  }
+  // Avoid scaling last_source unless its needed.
+  // Last source is needed if vp9_avg_source_sad() is used, or if
+  // partition_search_type == SOURCE_VAR_BASED_PARTITION, or if noise
+  // estimation is enabled.
+  if (cpi->unscaled_last_source != NULL &&
+      (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
+      (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_VBR &&
+      cpi->oxcf.mode == REALTIME && cpi->oxcf.speed >= 5) ||
+      cpi->sf.partition_search_type == SOURCE_VAR_BASED_PARTITION ||
+      cpi->noise_estimate.enabled))
+    cpi->Last_Source = vp9_scale_if_required(cm,
+                                             cpi->unscaled_last_source,
+                                             &cpi->scaled_last_source,
+                                             (cpi->oxcf.pass == 0));
+
+  if (cm->frame_type == KEY_FRAME || cpi->resize_pending != 0) {
+    memset(cpi->consec_zero_mv, 0,
+           cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
+  }
+
+  vp9_update_noise_estimate(cpi);
+
+  if (cpi->oxcf.pass == 0 &&
+      cpi->oxcf.mode == REALTIME &&
+      cpi->oxcf.speed >= 5 &&
+      cpi->resize_state == 0 &&
+      cm->frame_type != KEY_FRAME &&
+      (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
+       cpi->oxcf.rc_mode == VPX_VBR))
+    vp9_avg_source_sad(cpi);
+
+  // For 1 pass SVC, since only ZEROMV is allowed for upsampled reference
+  // frame (i.e, svc->force_zero_mode_spatial_ref = 0), we can avoid this
+  // frame-level upsampling.
+  if (frame_is_intra_only(cm) == 0 && !is_one_pass_cbr_svc(cpi)) {
+    vp9_scale_references(cpi);
+  }
+
+  set_size_independent_vars(cpi);
+  set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
+
+  if (cpi->oxcf.speed >= 5 &&
+      cpi->oxcf.pass == 0 &&
+      cpi->oxcf.rc_mode == VPX_CBR &&
+      cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
+      cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
+    cpi->use_skin_detection = 1;
+  }
+
+  vp9_set_quantizer(cm, q);
+  vp9_set_variance_partition_thresholds(cpi, q);
+
+  setup_frame(cpi);
+
+  suppress_active_map(cpi);
+  // Variance adaptive and in frame q adjustment experiments are mutually
+  // exclusive.
+  if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
+    vp9_vaq_frame_setup(cpi);
+  } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
+    vp9_360aq_frame_setup(cpi);
+  } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
+    vp9_setup_in_frame_q_adj(cpi);
+  } else if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
+    vp9_cyclic_refresh_setup(cpi);
+  }
+  apply_active_map(cpi);
+
+  // transform / motion compensation build reconstruction frame
+  vp9_encode_frame(cpi);
+
+  // Check if we should drop this frame because of high overshoot.
+  // Only for frames where high temporal-source sad is detected.
+  if (cpi->oxcf.pass == 0 &&
+      cpi->oxcf.rc_mode == VPX_CBR &&
+      cpi->resize_state == 0 &&
+      cm->frame_type != KEY_FRAME &&
+      cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
+      cpi->rc.high_source_sad == 1) {
+    int frame_size = 0;
+    // Get an estimate of the encoded frame size.
+    save_coding_context(cpi);
+    vp9_pack_bitstream(cpi, dest, size);
+    restore_coding_context(cpi);
+    frame_size = (int)(*size) << 3;
+    // Check if encoded frame will overshoot too much, and if so, set the q and
+    // adjust some rate control parameters, and return to re-encode the frame.
+    if (vp9_encodedframe_overshoot(cpi, frame_size, &q)) {
+      vpx_clear_system_state();
+      vp9_set_quantizer(cm, q);
+      vp9_set_variance_partition_thresholds(cpi, q);
+      suppress_active_map(cpi);
+      // Turn-off cyclic refresh for re-encoded frame.
+      if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
+        unsigned char *const seg_map = cpi->segmentation_map;
+        memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
+        vp9_disable_segmentation(&cm->seg);
+      }
+      apply_active_map(cpi);
+      vp9_encode_frame(cpi);
+    }
+  }
+
+  // Update some stats from cyclic refresh, and check if we should not update
+  // golden reference, for non-SVC 1 pass CBR.
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
+      cm->frame_type != KEY_FRAME &&
+      !cpi->use_svc &&
+      cpi->ext_refresh_frame_flags_pending == 0 &&
+      (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR))
+    vp9_cyclic_refresh_check_golden_update(cpi);
+
+  // Update the skip mb flag probabilities based on the distribution
+  // seen in the last encoder iteration.
+  // update_base_skip_probs(cpi);
+  vpx_clear_system_state();
+}
+
+static void encode_with_recode_loop(VP9_COMP *cpi,
+                                    size_t *size,
+                                    uint8_t *dest) {
+  VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int bottom_index, top_index;
+  int loop_count = 0;
+  int loop_at_this_size = 0;
+  int loop = 0;
+  int overshoot_seen = 0;
+  int undershoot_seen = 0;
+  int frame_over_shoot_limit;
+  int frame_under_shoot_limit;
+  int q = 0, q_low = 0, q_high = 0;
+
+  set_size_independent_vars(cpi);
+
+  do {
+    vpx_clear_system_state();
+
+    set_frame_size(cpi);
+
+    if (loop_count == 0 || cpi->resize_pending != 0) {
+      set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
+
+      // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
+      set_mv_search_params(cpi);
+
+      // Reset the loop state for new frame size.
+      overshoot_seen = 0;
+      undershoot_seen = 0;
+
+      // Reconfiguration for change in frame size has concluded.
+      cpi->resize_pending = 0;
+
+      q_low = bottom_index;
+      q_high = top_index;
+
+      loop_at_this_size = 0;
+    }
+
+    // Decide frame size bounds first time through.
+    if (loop_count == 0) {
+      vp9_rc_compute_frame_size_bounds(cpi, rc->this_frame_target,
+                                       &frame_under_shoot_limit,
+                                       &frame_over_shoot_limit);
+    }
+
+    cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
+                                      &cpi->scaled_source,
+                                      (cpi->oxcf.pass == 0));
+
+    if (cpi->unscaled_last_source != NULL)
+      cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
+                                               &cpi->scaled_last_source,
+                                               (cpi->oxcf.pass == 0));
+
+    if (frame_is_intra_only(cm) == 0) {
+      if (loop_count > 0) {
+        release_scaled_references(cpi);
+      }
+      vp9_scale_references(cpi);
+    }
+
+    vp9_set_quantizer(cm, q);
+
+    if (loop_count == 0)
+      setup_frame(cpi);
+
+    // Variance adaptive and in frame q adjustment experiments are mutually
+    // exclusive.
+    if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
+      vp9_vaq_frame_setup(cpi);
+    } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
+      vp9_360aq_frame_setup(cpi);
+    } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
+      vp9_setup_in_frame_q_adj(cpi);
+    }
+
+    // transform / motion compensation build reconstruction frame
+    vp9_encode_frame(cpi);
+
+    // Update the skip mb flag probabilities based on the distribution
+    // seen in the last encoder iteration.
+    // update_base_skip_probs(cpi);
+
+    vpx_clear_system_state();
+
+    // Dummy pack of the bitstream using up to date stats to get an
+    // accurate estimate of output frame size to determine if we need
+    // to recode.
+    if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF) {
+      save_coding_context(cpi);
+      if (!cpi->sf.use_nonrd_pick_mode)
+        vp9_pack_bitstream(cpi, dest, size);
+
+      rc->projected_frame_size = (int)(*size) << 3;
+      restore_coding_context(cpi);
+
+      if (frame_over_shoot_limit == 0)
+        frame_over_shoot_limit = 1;
+    }
+
+    if (cpi->oxcf.rc_mode == VPX_Q) {
+      loop = 0;
+    } else {
+      if ((cm->frame_type == KEY_FRAME) &&
+           rc->this_key_frame_forced &&
+           (rc->projected_frame_size < rc->max_frame_bandwidth)) {
+        int last_q = q;
+        int64_t kf_err;
+
+        int64_t high_err_target = cpi->ambient_err;
+        int64_t low_err_target = cpi->ambient_err >> 1;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+        if (cm->use_highbitdepth) {
+          kf_err = vp9_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
+        } else {
+          kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
+        }
+#else
+        kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+        // Prevent possible divide by zero error below for perfect KF
+        kf_err += !kf_err;
+
+        // The key frame is not good enough or we can afford
+        // to make it better without undue risk of popping.
+        if ((kf_err > high_err_target &&
+             rc->projected_frame_size <= frame_over_shoot_limit) ||
+            (kf_err > low_err_target &&
+             rc->projected_frame_size <= frame_under_shoot_limit)) {
+          // Lower q_high
+          q_high = q > q_low ? q - 1 : q_low;
+
+          // Adjust Q
+          q = (int)((q * high_err_target) / kf_err);
+          q = VPXMIN(q, (q_high + q_low) >> 1);
+        } else if (kf_err < low_err_target &&
+                   rc->projected_frame_size >= frame_under_shoot_limit) {
+          // The key frame is much better than the previous frame
+          // Raise q_low
+          q_low = q < q_high ? q + 1 : q_high;
+
+          // Adjust Q
+          q = (int)((q * low_err_target) / kf_err);
+          q = VPXMIN(q, (q_high + q_low + 1) >> 1);
+        }
+
+        // Clamp Q to upper and lower limits:
+        q = clamp(q, q_low, q_high);
+
+        loop = q != last_q;
+      } else if (recode_loop_test(
+          cpi, frame_over_shoot_limit, frame_under_shoot_limit,
+          q, VPXMAX(q_high, top_index), bottom_index)) {
+        // Is the projected frame size out of range and are we allowed
+        // to attempt to recode.
+        int last_q = q;
+        int retries = 0;
+
+        if (cpi->resize_pending == 1) {
+          // Change in frame size so go back around the recode loop.
+          cpi->rc.frame_size_selector =
+              SCALE_STEP1 - cpi->rc.frame_size_selector;
+          cpi->rc.next_frame_size_selector = cpi->rc.frame_size_selector;
+
+#if CONFIG_INTERNAL_STATS
+          ++cpi->tot_recode_hits;
+#endif
+          ++loop_count;
+          loop = 1;
+          continue;
+        }
+
+        // Frame size out of permitted range:
+        // Update correction factor & compute new Q to try...
+
+        // Frame is too large
+        if (rc->projected_frame_size > rc->this_frame_target) {
+          // Special case if the projected size is > the max allowed.
+          if (rc->projected_frame_size >= rc->max_frame_bandwidth)
+            q_high = rc->worst_quality;
+
+          // Raise Qlow as to at least the current value
+          q_low = q < q_high ? q + 1 : q_high;
+
+          if (undershoot_seen || loop_at_this_size > 1) {
+            // Update rate_correction_factor unless
+            vp9_rc_update_rate_correction_factors(cpi);
+
+            q = (q_high + q_low + 1) / 2;
+          } else {
+            // Update rate_correction_factor unless
+            vp9_rc_update_rate_correction_factors(cpi);
+
+            q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
+                                  bottom_index, VPXMAX(q_high, top_index));
+
+            while (q < q_low && retries < 10) {
+              vp9_rc_update_rate_correction_factors(cpi);
+              q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
+                                    bottom_index, VPXMAX(q_high, top_index));
+              retries++;
+            }
+          }
+
+          overshoot_seen = 1;
+        } else {
+          // Frame is too small
+          q_high = q > q_low ? q - 1 : q_low;
+
+          if (overshoot_seen || loop_at_this_size > 1) {
+            vp9_rc_update_rate_correction_factors(cpi);
+            q = (q_high + q_low) / 2;
+          } else {
+            vp9_rc_update_rate_correction_factors(cpi);
+            q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
+                                   bottom_index, top_index);
+            // Special case reset for qlow for constrained quality.
+            // This should only trigger where there is very substantial
+            // undershoot on a frame and the auto cq level is above
+            // the user passsed in value.
+            if (cpi->oxcf.rc_mode == VPX_CQ &&
+                q < q_low) {
+              q_low = q;
+            }
+
+            while (q > q_high && retries < 10) {
+              vp9_rc_update_rate_correction_factors(cpi);
+              q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
+                                     bottom_index, top_index);
+              retries++;
+            }
+          }
+
+          undershoot_seen = 1;
+        }
+
+        // Clamp Q to upper and lower limits:
+        q = clamp(q, q_low, q_high);
+
+        loop = (q != last_q);
+      } else {
+        loop = 0;
+      }
+    }
+
+    // Special case for overlay frame.
+    if (rc->is_src_frame_alt_ref &&
+        rc->projected_frame_size < rc->max_frame_bandwidth)
+      loop = 0;
+
+    if (loop) {
+      ++loop_count;
+      ++loop_at_this_size;
+
+#if CONFIG_INTERNAL_STATS
+      ++cpi->tot_recode_hits;
+#endif
+    }
+  } while (loop);
+}
+
+static int get_ref_frame_flags(const VP9_COMP *cpi) {
+  const int *const map = cpi->common.ref_frame_map;
+  const int gold_is_last = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idx];
+  const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idx];
+  const int gold_is_alt = map[cpi->gld_fb_idx] == map[cpi->alt_fb_idx];
+  int flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
+
+  if (gold_is_last)
+    flags &= ~VP9_GOLD_FLAG;
+
+  if (cpi->rc.frames_till_gf_update_due == INT_MAX &&
+      (cpi->svc.number_temporal_layers == 1 &&
+       cpi->svc.number_spatial_layers == 1))
+    flags &= ~VP9_GOLD_FLAG;
+
+  if (alt_is_last)
+    flags &= ~VP9_ALT_FLAG;
+
+  if (gold_is_alt)
+    flags &= ~VP9_ALT_FLAG;
+
+  return flags;
+}
+
+static void set_ext_overrides(VP9_COMP *cpi) {
+  // Overrides the defaults with the externally supplied values with
+  // vp9_update_reference() and vp9_update_entropy() calls
+  // Note: The overrides are valid only for the next frame passed
+  // to encode_frame_to_data_rate() function
+  if (cpi->ext_refresh_frame_context_pending) {
+    cpi->common.refresh_frame_context = cpi->ext_refresh_frame_context;
+    cpi->ext_refresh_frame_context_pending = 0;
+  }
+  if (cpi->ext_refresh_frame_flags_pending) {
+    cpi->refresh_last_frame = cpi->ext_refresh_last_frame;
+    cpi->refresh_golden_frame = cpi->ext_refresh_golden_frame;
+    cpi->refresh_alt_ref_frame = cpi->ext_refresh_alt_ref_frame;
+  }
+}
+
+YV12_BUFFER_CONFIG *vp9_svc_twostage_scale(VP9_COMMON *cm,
+                                           YV12_BUFFER_CONFIG *unscaled,
+                                           YV12_BUFFER_CONFIG *scaled,
+                                           YV12_BUFFER_CONFIG *scaled_temp) {
+  if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
+      cm->mi_rows * MI_SIZE != unscaled->y_height) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    scale_and_extend_frame(unscaled, scaled_temp, (int)cm->bit_depth);
+    scale_and_extend_frame(scaled_temp, scaled, (int)cm->bit_depth);
+#else
+    vp9_scale_and_extend_frame(unscaled, scaled_temp);
+    vp9_scale_and_extend_frame(scaled_temp, scaled);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    return scaled;
+  } else {
+    return unscaled;
+  }
+}
+
+YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
+                                          YV12_BUFFER_CONFIG *unscaled,
+                                          YV12_BUFFER_CONFIG *scaled,
+                                          int use_normative_scaler) {
+  if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
+      cm->mi_rows * MI_SIZE != unscaled->y_height) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (use_normative_scaler &&
+        unscaled->y_width <= (scaled->y_width << 1) &&
+        unscaled->y_height <= (scaled->y_height << 1))
+      scale_and_extend_frame(unscaled, scaled, (int)cm->bit_depth);
+    else
+      scale_and_extend_frame_nonnormative(unscaled, scaled, (int)cm->bit_depth);
+#else
+    if (use_normative_scaler &&
+        unscaled->y_width <= (scaled->y_width << 1) &&
+        unscaled->y_height <= (scaled->y_height << 1))
+      vp9_scale_and_extend_frame(unscaled, scaled);
+    else
+      scale_and_extend_frame_nonnormative(unscaled, scaled);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    return scaled;
+  } else {
+    return unscaled;
+  }
+}
+
+static void set_arf_sign_bias(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  int arf_sign_bias;
+
+  if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    arf_sign_bias = cpi->rc.source_alt_ref_active &&
+                    (!cpi->refresh_alt_ref_frame ||
+                     (gf_group->rf_level[gf_group->index] == GF_ARF_LOW));
+  } else {
+    arf_sign_bias =
+      (cpi->rc.source_alt_ref_active && !cpi->refresh_alt_ref_frame);
+  }
+  cm->ref_frame_sign_bias[ALTREF_FRAME] = arf_sign_bias;
+}
+
+static int setup_interp_filter_search_mask(VP9_COMP *cpi) {
+  INTERP_FILTER ifilter;
+  int ref_total[MAX_REF_FRAMES] = {0};
+  MV_REFERENCE_FRAME ref;
+  int mask = 0;
+  if (cpi->common.last_frame_type == KEY_FRAME ||
+      cpi->refresh_alt_ref_frame)
+    return mask;
+  for (ref = LAST_FRAME; ref <= ALTREF_FRAME; ++ref)
+    for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter)
+      ref_total[ref] += cpi->interp_filter_selected[ref][ifilter];
+
+  for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter) {
+    if ((ref_total[LAST_FRAME] &&
+        cpi->interp_filter_selected[LAST_FRAME][ifilter] == 0) &&
+        (ref_total[GOLDEN_FRAME] == 0 ||
+         cpi->interp_filter_selected[GOLDEN_FRAME][ifilter] * 50
+           < ref_total[GOLDEN_FRAME]) &&
+        (ref_total[ALTREF_FRAME] == 0 ||
+         cpi->interp_filter_selected[ALTREF_FRAME][ifilter] * 50
+           < ref_total[ALTREF_FRAME]))
+      mask |= 1 << ifilter;
+  }
+  return mask;
+}
+
+static void encode_frame_to_data_rate(VP9_COMP *cpi,
+                                      size_t *size,
+                                      uint8_t *dest,
+                                      unsigned int *frame_flags) {
+  VP9_COMMON *const cm = &cpi->common;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  struct segmentation *const seg = &cm->seg;
+  TX_SIZE t;
+
+  set_ext_overrides(cpi);
+  vpx_clear_system_state();
+
+  // Set the arf sign bias for this frame.
+  set_arf_sign_bias(cpi);
+
+  // Set default state for segment based loop filter update flags.
+  cm->lf.mode_ref_delta_update = 0;
+
+  if (cpi->oxcf.pass == 2 &&
+      cpi->sf.adaptive_interp_filter_search)
+    cpi->sf.interp_filter_search_mask =
+        setup_interp_filter_search_mask(cpi);
+
+  // Set various flags etc to special state if it is a key frame.
+  if (frame_is_intra_only(cm)) {
+    // Reset the loop filter deltas and segmentation map.
+    vp9_reset_segment_features(&cm->seg);
+
+    // If segmentation is enabled force a map update for key frames.
+    if (seg->enabled) {
+      seg->update_map = 1;
+      seg->update_data = 1;
+    }
+
+    // The alternate reference frame cannot be active for a key frame.
+    cpi->rc.source_alt_ref_active = 0;
+
+    cm->error_resilient_mode = oxcf->error_resilient_mode;
+    cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
+
+    // By default, encoder assumes decoder can use prev_mi.
+    if (cm->error_resilient_mode) {
+      cm->frame_parallel_decoding_mode = 1;
+      cm->reset_frame_context = 0;
+      cm->refresh_frame_context = 0;
+    } else if (cm->intra_only) {
+      // Only reset the current context.
+      cm->reset_frame_context = 2;
+    }
+  }
+  if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0) {
+    // Use context 0 for intra only empty frame, but the last frame context
+    // for other empty frames.
+    if (cpi->svc.encode_empty_frame_state == ENCODING) {
+      if (cpi->svc.encode_intra_empty_frame != 0)
+        cm->frame_context_idx = 0;
+      else
+        cm->frame_context_idx = FRAME_CONTEXTS - 1;
+    } else {
+    cm->frame_context_idx =
+        cpi->svc.spatial_layer_id * cpi->svc.number_temporal_layers +
+        cpi->svc.temporal_layer_id;
+    }
+
+    cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
+
+    // The probs will be updated based on the frame type of its previous
+    // frame if frame_parallel_decoding_mode is 0. The type may vary for
+    // the frame after a key frame in base layer since we may drop enhancement
+    // layers. So set frame_parallel_decoding_mode to 1 in this case.
+    if (cm->frame_parallel_decoding_mode == 0) {
+      if (cpi->svc.number_temporal_layers == 1) {
+        if (cpi->svc.spatial_layer_id == 0 &&
+            cpi->svc.layer_context[0].last_frame_type == KEY_FRAME)
+          cm->frame_parallel_decoding_mode = 1;
+      } else if (cpi->svc.spatial_layer_id == 0) {
+        // Find the 2nd frame in temporal base layer and 1st frame in temporal
+        // enhancement layers from the key frame.
+        int i;
+        for (i = 0; i < cpi->svc.number_temporal_layers; ++i) {
+          if (cpi->svc.layer_context[0].frames_from_key_frame == 1 << i) {
+            cm->frame_parallel_decoding_mode = 1;
+            break;
+          }
+        }
+      }
+    }
+  }
+
+  // For 1 pass CBR, check if we are dropping this frame.
+  // For spatial layers, for now only check for frame-dropping on first spatial
+  // layer, and if decision is to drop, we drop whole super-frame.
+  if (oxcf->pass == 0 &&
+      oxcf->rc_mode == VPX_CBR &&
+      cm->frame_type != KEY_FRAME) {
+    if (vp9_rc_drop_frame(cpi) ||
+        (is_one_pass_cbr_svc(cpi) && cpi->svc.rc_drop_superframe == 1)) {
+      vp9_rc_postencode_update_drop_frame(cpi);
+      ++cm->current_video_frame;
+      cpi->ext_refresh_frame_flags_pending = 0;
+      cpi->svc.rc_drop_superframe = 1;
+      // TODO(marpan): Advancing the svc counters on dropped frames can break
+      // the referencing scheme for the fixed svc patterns defined in
+      // vp9_one_pass_cbr_svc_start_layer(). Look into fixing this issue, but
+      // for now, don't advance the svc frame counters on dropped frame.
+      // if (cpi->use_svc)
+      //   vp9_inc_frame_in_layer(cpi);
+      return;
+    }
+  }
+
+  vpx_clear_system_state();
+
+#if CONFIG_INTERNAL_STATS
+  memset(cpi->mode_chosen_counts, 0,
+         MAX_MODES * sizeof(*cpi->mode_chosen_counts));
+#endif
+
+  if (cpi->sf.recode_loop == DISALLOW_RECODE) {
+    encode_without_recode_loop(cpi, size, dest);
+  } else {
+    encode_with_recode_loop(cpi, size, dest);
+  }
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#ifdef OUTPUT_YUV_DENOISED
+  if (oxcf->noise_sensitivity > 0) {
+    vp9_write_yuv_frame_420(&cpi->denoiser.running_avg_y[INTRA_FRAME],
+                            yuv_denoised_file);
+  }
+#endif
+#endif
+#ifdef OUTPUT_YUV_SKINMAP
+  if (cpi->common.current_video_frame > 1) {
+    vp9_compute_skin_map(cpi, yuv_skinmap_file);
+  }
+#endif
+
+  // Special case code to reduce pulsing when key frames are forced at a
+  // fixed interval. Note the reconstruction error if it is the frame before
+  // the force key frame
+  if (cpi->rc.next_key_frame_forced && cpi->rc.frames_to_key == 1) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      cpi->ambient_err = vp9_highbd_get_y_sse(cpi->Source,
+                                              get_frame_new_buffer(cm));
+    } else {
+      cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
+    }
+#else
+    cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+
+  // If the encoder forced a KEY_FRAME decision
+  if (cm->frame_type == KEY_FRAME)
+    cpi->refresh_last_frame = 1;
+
+  cm->frame_to_show = get_frame_new_buffer(cm);
+  cm->frame_to_show->color_space = cm->color_space;
+  cm->frame_to_show->color_range = cm->color_range;
+  cm->frame_to_show->render_width  = cm->render_width;
+  cm->frame_to_show->render_height = cm->render_height;
+
+  // Pick the loop filter level for the frame.
+  loopfilter_frame(cpi, cm);
+
+  // build the bitstream
+  vp9_pack_bitstream(cpi, dest, size);
+
+  if (cm->seg.update_map)
+    update_reference_segmentation_map(cpi);
+
+  if (frame_is_intra_only(cm) == 0) {
+    release_scaled_references(cpi);
+  }
+  vp9_update_reference_frames(cpi);
+
+  for (t = TX_4X4; t <= TX_32X32; t++)
+    full_to_model_counts(cpi->td.counts->coef[t],
+                         cpi->td.rd_counts.coef_counts[t]);
+
+  if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode)
+    vp9_adapt_coef_probs(cm);
+
+  if (!frame_is_intra_only(cm)) {
+    if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
+      vp9_adapt_mode_probs(cm);
+      vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
+    }
+  }
+
+  cpi->ext_refresh_frame_flags_pending = 0;
+
+  if (cpi->refresh_golden_frame == 1)
+    cpi->frame_flags |= FRAMEFLAGS_GOLDEN;
+  else
+    cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN;
+
+  if (cpi->refresh_alt_ref_frame == 1)
+    cpi->frame_flags |= FRAMEFLAGS_ALTREF;
+  else
+    cpi->frame_flags &= ~FRAMEFLAGS_ALTREF;
+
+  cpi->ref_frame_flags = get_ref_frame_flags(cpi);
+
+  cm->last_frame_type = cm->frame_type;
+
+  if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
+    vp9_rc_postencode_update(cpi, *size);
+
+#if 0
+  output_frame_level_debug_stats(cpi);
+#endif
+
+  if (cm->frame_type == KEY_FRAME) {
+    // Tell the caller that the frame was coded as a key frame
+    *frame_flags = cpi->frame_flags | FRAMEFLAGS_KEY;
+  } else {
+    *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY;
+  }
+
+  // Clear the one shot update flags for segmentation map and mode/ref loop
+  // filter deltas.
+  cm->seg.update_map = 0;
+  cm->seg.update_data = 0;
+  cm->lf.mode_ref_delta_update = 0;
+
+  // keep track of the last coded dimensions
+  cm->last_width = cm->width;
+  cm->last_height = cm->height;
+
+  // reset to normal state now that we are done.
+  if (!cm->show_existing_frame)
+    cm->last_show_frame = cm->show_frame;
+
+  if (cm->show_frame) {
+    vp9_swap_mi_and_prev_mi(cm);
+    // Don't increment frame counters if this was an altref buffer
+    // update not a real frame
+    ++cm->current_video_frame;
+    if (cpi->use_svc)
+      vp9_inc_frame_in_layer(cpi);
+  }
+  cm->prev_frame = cm->cur_frame;
+
+  if (cpi->use_svc)
+    cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+                           cpi->svc.number_temporal_layers +
+                           cpi->svc.temporal_layer_id].last_frame_type =
+                               cm->frame_type;
+}
+
+static void SvcEncode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
+                      unsigned int *frame_flags) {
+  vp9_rc_get_svc_params(cpi);
+  encode_frame_to_data_rate(cpi, size, dest, frame_flags);
+}
+
+static void Pass0Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
+                        unsigned int *frame_flags) {
+  if (cpi->oxcf.rc_mode == VPX_CBR) {
+    vp9_rc_get_one_pass_cbr_params(cpi);
+  } else {
+    vp9_rc_get_one_pass_vbr_params(cpi);
+  }
+  encode_frame_to_data_rate(cpi, size, dest, frame_flags);
+}
+
+static void Pass2Encode(VP9_COMP *cpi, size_t *size,
+                        uint8_t *dest, unsigned int *frame_flags) {
+  cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
+  encode_frame_to_data_rate(cpi, size, dest, frame_flags);
+
+  if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
+    vp9_twopass_postencode_update(cpi);
+}
+
+static void init_ref_frame_bufs(VP9_COMMON *cm) {
+  int i;
+  BufferPool *const pool = cm->buffer_pool;
+  cm->new_fb_idx = INVALID_IDX;
+  for (i = 0; i < REF_FRAMES; ++i) {
+    cm->ref_frame_map[i] = INVALID_IDX;
+    pool->frame_bufs[i].ref_count = 0;
+  }
+}
+
+static void check_initial_width(VP9_COMP *cpi,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                int use_highbitdepth,
+#endif
+                                int subsampling_x, int subsampling_y) {
+  VP9_COMMON *const cm = &cpi->common;
+
+  if (!cpi->initial_width ||
+#if CONFIG_VP9_HIGHBITDEPTH
+      cm->use_highbitdepth != use_highbitdepth ||
+#endif
+      cm->subsampling_x != subsampling_x ||
+      cm->subsampling_y != subsampling_y) {
+    cm->subsampling_x = subsampling_x;
+    cm->subsampling_y = subsampling_y;
+#if CONFIG_VP9_HIGHBITDEPTH
+    cm->use_highbitdepth = use_highbitdepth;
+#endif
+
+    alloc_raw_frame_buffers(cpi);
+    init_ref_frame_bufs(cm);
+    alloc_util_frame_buffers(cpi);
+
+    init_motion_estimation(cpi);  // TODO(agrange) This can be removed.
+
+    cpi->initial_width = cm->width;
+    cpi->initial_height = cm->height;
+    cpi->initial_mbs = cm->MBs;
+  }
+}
+
+int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
+                          YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
+                          int64_t end_time) {
+  VP9_COMMON *const cm = &cpi->common;
+  struct vpx_usec_timer timer;
+  int res = 0;
+  const int subsampling_x = sd->subsampling_x;
+  const int subsampling_y = sd->subsampling_y;
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int use_highbitdepth = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0;
+#endif
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  check_initial_width(cpi, use_highbitdepth, subsampling_x, subsampling_y);
+#else
+  check_initial_width(cpi, subsampling_x, subsampling_y);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  setup_denoiser_buffer(cpi);
+#endif
+  vpx_usec_timer_start(&timer);
+
+  if (vp9_lookahead_push(cpi->lookahead, sd, time_stamp, end_time,
+#if CONFIG_VP9_HIGHBITDEPTH
+                         use_highbitdepth,
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+                         frame_flags))
+    res = -1;
+  vpx_usec_timer_mark(&timer);
+  cpi->time_receive_data += vpx_usec_timer_elapsed(&timer);
+
+  if ((cm->profile == PROFILE_0 || cm->profile == PROFILE_2) &&
+      (subsampling_x != 1 || subsampling_y != 1)) {
+    vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
+                       "Non-4:2:0 color format requires profile 1 or 3");
+    res = -1;
+  }
+  if ((cm->profile == PROFILE_1 || cm->profile == PROFILE_3) &&
+      (subsampling_x == 1 && subsampling_y == 1)) {
+    vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
+                       "4:2:0 color format requires profile 0 or 2");
+    res = -1;
+  }
+
+  return res;
+}
+
+
+static int frame_is_reference(const VP9_COMP *cpi) {
+  const VP9_COMMON *cm = &cpi->common;
+
+  return cm->frame_type == KEY_FRAME ||
+         cpi->refresh_last_frame ||
+         cpi->refresh_golden_frame ||
+         cpi->refresh_alt_ref_frame ||
+         cm->refresh_frame_context ||
+         cm->lf.mode_ref_delta_update ||
+         cm->seg.update_map ||
+         cm->seg.update_data;
+}
+
+static void adjust_frame_rate(VP9_COMP *cpi,
+                              const struct lookahead_entry *source) {
+  int64_t this_duration;
+  int step = 0;
+
+  if (source->ts_start == cpi->first_time_stamp_ever) {
+    this_duration = source->ts_end - source->ts_start;
+    step = 1;
+  } else {
+    int64_t last_duration = cpi->last_end_time_stamp_seen
+        - cpi->last_time_stamp_seen;
+
+    this_duration = source->ts_end - cpi->last_end_time_stamp_seen;
+
+    // do a step update if the duration changes by 10%
+    if (last_duration)
+      step = (int)((this_duration - last_duration) * 10 / last_duration);
+  }
+
+  if (this_duration) {
+    if (step) {
+      vp9_new_framerate(cpi, 10000000.0 / this_duration);
+    } else {
+      // Average this frame's rate into the last second's average
+      // frame rate. If we haven't seen 1 second yet, then average
+      // over the whole interval seen.
+      const double interval = VPXMIN(
+          (double)(source->ts_end - cpi->first_time_stamp_ever), 10000000.0);
+      double avg_duration = 10000000.0 / cpi->framerate;
+      avg_duration *= (interval - avg_duration + this_duration);
+      avg_duration /= interval;
+
+      vp9_new_framerate(cpi, 10000000.0 / avg_duration);
+    }
+  }
+  cpi->last_time_stamp_seen = source->ts_start;
+  cpi->last_end_time_stamp_seen = source->ts_end;
+}
+
+// Returns 0 if this is not an alt ref else the offset of the source frame
+// used as the arf midpoint.
+static int get_arf_src_index(VP9_COMP *cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  int arf_src_index = 0;
+  if (is_altref_enabled(cpi)) {
+    if (cpi->oxcf.pass == 2) {
+      const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+      if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
+        arf_src_index = gf_group->arf_src_offset[gf_group->index];
+      }
+    } else if (rc->source_alt_ref_pending) {
+      arf_src_index = rc->frames_till_gf_update_due;
+    }
+  }
+  return arf_src_index;
+}
+
+static void check_src_altref(VP9_COMP *cpi,
+                             const struct lookahead_entry *source) {
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  if (cpi->oxcf.pass == 2) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    rc->is_src_frame_alt_ref =
+      (gf_group->update_type[gf_group->index] == OVERLAY_UPDATE);
+  } else {
+    rc->is_src_frame_alt_ref = cpi->alt_ref_source &&
+                               (source == cpi->alt_ref_source);
+  }
+
+  if (rc->is_src_frame_alt_ref) {
+    // Current frame is an ARF overlay frame.
+    cpi->alt_ref_source = NULL;
+
+    // Don't refresh the last buffer for an ARF overlay frame. It will
+    // become the GF so preserve last as an alternative prediction option.
+    cpi->refresh_last_frame = 0;
+  }
+}
+
+#if CONFIG_INTERNAL_STATS
+extern double vp9_get_blockiness(const uint8_t *img1, int img1_pitch,
+                                 const uint8_t *img2, int img2_pitch,
+                                 int width, int height);
+
+static void adjust_image_stat(double y, double u, double v, double all,
+                              ImageStat *s) {
+  s->stat[Y] += y;
+  s->stat[U] += u;
+  s->stat[V] += v;
+  s->stat[ALL] += all;
+  s->worst = VPXMIN(s->worst, all);
+}
+#endif  // CONFIG_INTERNAL_STATS
+
+static void update_level_info(VP9_COMP *cpi, size_t *size, int arf_src_index) {
+  VP9_COMMON *const cm = &cpi->common;
+  Vp9LevelInfo *const level_info = &cpi->level_info;
+  Vp9LevelSpec *const level_spec = &level_info->level_spec;
+  Vp9LevelStats *const level_stats = &level_info->level_stats;
+  int i, idx;
+  uint64_t luma_samples, dur_end;
+  const uint32_t luma_pic_size = cm->width * cm->height;
+  double cpb_data_size;
+
+  vpx_clear_system_state();
+
+  // update level_stats
+  level_stats->total_compressed_size += *size;
+  if (cm->show_frame) {
+    level_stats->total_uncompressed_size +=
+        luma_pic_size +
+        2 * (luma_pic_size >> (cm->subsampling_x + cm->subsampling_y));
+    level_stats->time_encoded =
+        (cpi->last_end_time_stamp_seen - cpi->first_time_stamp_ever) /
+        (double)TICKS_PER_SEC;
+  }
+
+  if (arf_src_index > 0) {
+    if (!level_stats->seen_first_altref) {
+      level_stats->seen_first_altref = 1;
+    } else if (level_stats->frames_since_last_altref <
+             level_spec->min_altref_distance) {
+      level_spec->min_altref_distance = level_stats->frames_since_last_altref;
+    }
+    level_stats->frames_since_last_altref = 0;
+  } else {
+    ++level_stats->frames_since_last_altref;
+  }
+
+  if (level_stats->frame_window_buffer.len < FRAME_WINDOW_SIZE - 1) {
+    idx = (level_stats->frame_window_buffer.start +
+           level_stats->frame_window_buffer.len++) % FRAME_WINDOW_SIZE;
+  } else {
+    idx = level_stats->frame_window_buffer.start;
+    level_stats->frame_window_buffer.start = (idx + 1) % FRAME_WINDOW_SIZE;
+  }
+  level_stats->frame_window_buffer.buf[idx].ts = cpi->last_time_stamp_seen;
+  level_stats->frame_window_buffer.buf[idx].size = (uint32_t)(*size);
+  level_stats->frame_window_buffer.buf[idx].luma_samples = luma_pic_size;
+
+  if (cm->frame_type == KEY_FRAME) {
+    level_stats->ref_refresh_map = 0;
+  } else {
+    int count = 0;
+    level_stats->ref_refresh_map |= vp9_get_refresh_mask(cpi);
+    // Also need to consider the case where the encoder refers to a buffer
+    // that has been implicitly refreshed after encoding a keyframe.
+    if (!cm->intra_only) {
+      level_stats->ref_refresh_map |= (1 << cpi->lst_fb_idx);
+      level_stats->ref_refresh_map |= (1 << cpi->gld_fb_idx);
+      level_stats->ref_refresh_map |= (1 << cpi->alt_fb_idx);
+    }
+    for (i = 0; i < REF_FRAMES; ++i) {
+      count += (level_stats->ref_refresh_map >> i) & 1;
+    }
+    if (count > level_spec->max_ref_frame_buffers) {
+      level_spec->max_ref_frame_buffers = count;
+    }
+  }
+
+  // update average_bitrate
+  level_spec->average_bitrate =
+      (double)level_stats->total_compressed_size / 125.0 /
+      level_stats->time_encoded;
+
+  // update max_luma_sample_rate
+  luma_samples = 0;
+  for (i = 0; i < level_stats->frame_window_buffer.len; ++i) {
+    idx = (level_stats->frame_window_buffer.start +
+           level_stats->frame_window_buffer.len - 1 - i) % FRAME_WINDOW_SIZE;
+    if (i == 0) {
+      dur_end = level_stats->frame_window_buffer.buf[idx].ts;
+    }
+    if (dur_end - level_stats->frame_window_buffer.buf[idx].ts >=
+        TICKS_PER_SEC) {
+      break;
+    }
+    luma_samples += level_stats->frame_window_buffer.buf[idx].luma_samples;
+  }
+  if (luma_samples > level_spec->max_luma_sample_rate) {
+    level_spec->max_luma_sample_rate = luma_samples;
+  }
+
+  // update max_cpb_size
+  cpb_data_size = 0;
+  for (i = 0; i < CPB_WINDOW_SIZE; ++i) {
+    if (i >= level_stats->frame_window_buffer.len) break;
+    idx = (level_stats->frame_window_buffer.start +
+           level_stats->frame_window_buffer.len - 1 - i) % FRAME_WINDOW_SIZE;
+    cpb_data_size += level_stats->frame_window_buffer.buf[idx].size;
+  }
+  cpb_data_size = cpb_data_size / 125.0;
+  if (cpb_data_size > level_spec->max_cpb_size) {
+    level_spec->max_cpb_size = cpb_data_size;
+  }
+
+  // update max_luma_picture_size
+  if (luma_pic_size > level_spec->max_luma_picture_size) {
+    level_spec->max_luma_picture_size = luma_pic_size;
+  }
+
+  // update compression_ratio
+  level_spec->compression_ratio =
+      (double)level_stats->total_uncompressed_size * cm->bit_depth /
+      level_stats->total_compressed_size / 8.0;
+
+  // update max_col_tiles
+  if (level_spec->max_col_tiles < (1 << cm->log2_tile_cols)) {
+    level_spec->max_col_tiles = (1 << cm->log2_tile_cols);
+  }
+}
+
+int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
+                            size_t *size, uint8_t *dest,
+                            int64_t *time_stamp, int64_t *time_end, int flush) {
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  VP9_COMMON *const cm = &cpi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  RATE_CONTROL *const rc = &cpi->rc;
+  struct vpx_usec_timer  cmptimer;
+  YV12_BUFFER_CONFIG *force_src_buffer = NULL;
+  struct lookahead_entry *last_source = NULL;
+  struct lookahead_entry *source = NULL;
+  int arf_src_index;
+  int i;
+
+  if (is_two_pass_svc(cpi)) {
+#if CONFIG_SPATIAL_SVC
+    vp9_svc_start_frame(cpi);
+    // Use a small empty frame instead of a real frame
+    if (cpi->svc.encode_empty_frame_state == ENCODING)
+      source = &cpi->svc.empty_frame;
+#endif
+    if (oxcf->pass == 2)
+      vp9_restore_layer_context(cpi);
+  } else if (is_one_pass_cbr_svc(cpi)) {
+    vp9_one_pass_cbr_svc_start_layer(cpi);
+  }
+
+  vpx_usec_timer_start(&cmptimer);
+
+  vp9_set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV);
+
+  // Is multi-arf enabled.
+  // Note that at the moment multi_arf is only configured for 2 pass VBR and
+  // will not work properly with svc.
+  if ((oxcf->pass == 2) && !cpi->use_svc &&
+      (cpi->oxcf.enable_auto_arf > 1))
+    cpi->multi_arf_allowed = 1;
+  else
+    cpi->multi_arf_allowed = 0;
+
+  // Normal defaults
+  cm->reset_frame_context = 0;
+  cm->refresh_frame_context = 1;
+  if (!is_one_pass_cbr_svc(cpi)) {
+    cpi->refresh_last_frame = 1;
+    cpi->refresh_golden_frame = 0;
+    cpi->refresh_alt_ref_frame = 0;
+  }
+
+  // Should we encode an arf frame.
+  arf_src_index = get_arf_src_index(cpi);
+
+  // Skip alt frame if we encode the empty frame
+  if (is_two_pass_svc(cpi) && source != NULL)
+    arf_src_index = 0;
+
+  if (arf_src_index) {
+    for (i = 0; i <= arf_src_index; ++i) {
+      struct lookahead_entry *e = vp9_lookahead_peek(cpi->lookahead, i);
+      // Avoid creating an alt-ref if there's a forced keyframe pending.
+      if (e == NULL) {
+        break;
+      } else if (e->flags == VPX_EFLAG_FORCE_KF) {
+        arf_src_index = 0;
+        flush = 1;
+        break;
+      }
+    }
+  }
+
+  if (arf_src_index) {
+    assert(arf_src_index <= rc->frames_to_key);
+
+    if ((source = vp9_lookahead_peek(cpi->lookahead, arf_src_index)) != NULL) {
+      cpi->alt_ref_source = source;
+
+#if CONFIG_SPATIAL_SVC
+      if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0) {
+        int i;
+        // Reference a hidden frame from a lower layer
+        for (i = cpi->svc.spatial_layer_id - 1; i >= 0; --i) {
+          if (oxcf->ss_enable_auto_arf[i]) {
+            cpi->gld_fb_idx = cpi->svc.layer_context[i].alt_ref_idx;
+            break;
+          }
+        }
+      }
+      cpi->svc.layer_context[cpi->svc.spatial_layer_id].has_alt_frame = 1;
+#endif
+
+      if ((oxcf->arnr_max_frames > 0) && (oxcf->arnr_strength > 0)) {
+        // Produce the filtered ARF frame.
+        vp9_temporal_filter(cpi, arf_src_index);
+        vpx_extend_frame_borders(&cpi->alt_ref_buffer);
+        force_src_buffer = &cpi->alt_ref_buffer;
+      }
+
+      cm->show_frame = 0;
+      cm->intra_only = 0;
+      cpi->refresh_alt_ref_frame = 1;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_last_frame = 0;
+      rc->is_src_frame_alt_ref = 0;
+      rc->source_alt_ref_pending = 0;
+    } else {
+      rc->source_alt_ref_pending = 0;
+    }
+  }
+
+  if (!source) {
+    // Get last frame source.
+    if (cm->current_video_frame > 0) {
+      if ((last_source = vp9_lookahead_peek(cpi->lookahead, -1)) == NULL)
+        return -1;
+    }
+
+    // Read in the source frame.
+    if (cpi->use_svc)
+      source = vp9_svc_lookahead_pop(cpi, cpi->lookahead, flush);
+    else
+      source = vp9_lookahead_pop(cpi->lookahead, flush);
+
+    if (source != NULL) {
+      cm->show_frame = 1;
+      cm->intra_only = 0;
+      // if the flags indicate intra frame, but if the current picture is for
+      // non-zero spatial layer, it should not be an intra picture.
+      // TODO(Won Kap): this needs to change if per-layer intra frame is
+      // allowed.
+      if ((source->flags & VPX_EFLAG_FORCE_KF) &&
+          cpi->svc.spatial_layer_id > cpi->svc.first_spatial_layer_to_encode) {
+        source->flags &= ~(unsigned int)(VPX_EFLAG_FORCE_KF);
+      }
+
+      // Check to see if the frame should be encoded as an arf overlay.
+      check_src_altref(cpi, source);
+    }
+  }
+
+  if (source) {
+    cpi->un_scaled_source = cpi->Source = force_src_buffer ? force_src_buffer
+                                                           : &source->img;
+
+    cpi->unscaled_last_source = last_source != NULL ? &last_source->img : NULL;
+
+    *time_stamp = source->ts_start;
+    *time_end = source->ts_end;
+    *frame_flags = (source->flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
+
+  } else {
+    *size = 0;
+    if (flush && oxcf->pass == 1 && !cpi->twopass.first_pass_done) {
+      vp9_end_first_pass(cpi);    /* get last stats packet */
+      cpi->twopass.first_pass_done = 1;
+    }
+    return -1;
+  }
+
+  if (source->ts_start < cpi->first_time_stamp_ever) {
+    cpi->first_time_stamp_ever = source->ts_start;
+    cpi->last_end_time_stamp_seen = source->ts_start;
+  }
+
+  // Clear down mmx registers
+  vpx_clear_system_state();
+
+  // adjust frame rates based on timestamps given
+  if (cm->show_frame) {
+    adjust_frame_rate(cpi, source);
+  }
+
+  if (is_one_pass_cbr_svc(cpi)) {
+    vp9_update_temporal_layer_framerate(cpi);
+    vp9_restore_layer_context(cpi);
+  }
+
+  // Find a free buffer for the new frame, releasing the reference previously
+  // held.
+  if (cm->new_fb_idx != INVALID_IDX) {
+    --pool->frame_bufs[cm->new_fb_idx].ref_count;
+  }
+  cm->new_fb_idx = get_free_fb(cm);
+
+  if (cm->new_fb_idx == INVALID_IDX)
+    return -1;
+
+  cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
+
+  if (!cpi->use_svc && cpi->multi_arf_allowed) {
+    if (cm->frame_type == KEY_FRAME) {
+      init_buffer_indices(cpi);
+    } else if (oxcf->pass == 2) {
+      const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+      cpi->alt_fb_idx = gf_group->arf_ref_idx[gf_group->index];
+    }
+  }
+
+  // Start with a 0 size frame.
+  *size = 0;
+
+  cpi->frame_flags = *frame_flags;
+
+  if ((oxcf->pass == 2) &&
+      (!cpi->use_svc ||
+          (is_two_pass_svc(cpi) &&
+              cpi->svc.encode_empty_frame_state != ENCODING))) {
+    vp9_rc_get_second_pass_params(cpi);
+  } else if (oxcf->pass == 1) {
+    set_frame_size(cpi);
+  }
+
+  if (cpi->oxcf.pass != 0 ||
+      cpi->use_svc ||
+      frame_is_intra_only(cm) == 1) {
+    for (i = 0; i < MAX_REF_FRAMES; ++i)
+      cpi->scaled_ref_idx[i] = INVALID_IDX;
+  }
+
+  if (oxcf->pass == 1 &&
+      (!cpi->use_svc || is_two_pass_svc(cpi))) {
+    const int lossless = is_lossless_requested(oxcf);
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cpi->oxcf.use_highbitdepth)
+      cpi->td.mb.fwd_txm4x4 = lossless ?
+          vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
+    else
+      cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
+    cpi->td.mb.highbd_itxm_add = lossless ? vp9_highbd_iwht4x4_add :
+                                         vp9_highbd_idct4x4_add;
+#else
+    cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    cpi->td.mb.itxm_add = lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
+    vp9_first_pass(cpi, source);
+  } else if (oxcf->pass == 2 &&
+      (!cpi->use_svc || is_two_pass_svc(cpi))) {
+    Pass2Encode(cpi, size, dest, frame_flags);
+  } else if (cpi->use_svc) {
+    SvcEncode(cpi, size, dest, frame_flags);
+  } else {
+    // One pass encode
+    Pass0Encode(cpi, size, dest, frame_flags);
+  }
+
+  if (cm->refresh_frame_context)
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+
+  // No frame encoded, or frame was dropped, release scaled references.
+  if ((*size == 0) && (frame_is_intra_only(cm) == 0)) {
+    release_scaled_references(cpi);
+  }
+
+  if (*size > 0) {
+    cpi->droppable = !frame_is_reference(cpi);
+  }
+
+  // Save layer specific state.
+  if (is_one_pass_cbr_svc(cpi) ||
+        ((cpi->svc.number_temporal_layers > 1 ||
+          cpi->svc.number_spatial_layers > 1) &&
+         oxcf->pass == 2)) {
+    vp9_save_layer_context(cpi);
+  }
+
+  vpx_usec_timer_mark(&cmptimer);
+  cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer);
+
+  if (cpi->b_calculate_psnr && oxcf->pass != 1 && cm->show_frame)
+    generate_psnr_packet(cpi);
+
+  if (cpi->keep_level_stats && oxcf->pass != 1)
+    update_level_info(cpi, size, arf_src_index);
+
+#if CONFIG_INTERNAL_STATS
+
+  if (oxcf->pass != 1) {
+    double samples = 0.0;
+    cpi->bytes += (int)(*size);
+
+    if (cm->show_frame) {
+      cpi->count++;
+
+      if (cpi->b_calculate_psnr) {
+        YV12_BUFFER_CONFIG *orig = cpi->Source;
+        YV12_BUFFER_CONFIG *recon = cpi->common.frame_to_show;
+        YV12_BUFFER_CONFIG *pp = &cm->post_proc_buffer;
+        PSNR_STATS psnr;
+#if CONFIG_VP9_HIGHBITDEPTH
+        calc_highbd_psnr(orig, recon, &psnr, cpi->td.mb.e_mbd.bd,
+                         cpi->oxcf.input_bit_depth);
+#else
+        calc_psnr(orig, recon, &psnr);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+        adjust_image_stat(psnr.psnr[1], psnr.psnr[2], psnr.psnr[3],
+                          psnr.psnr[0], &cpi->psnr);
+        cpi->total_sq_error += psnr.sse[0];
+        cpi->total_samples += psnr.samples[0];
+        samples = psnr.samples[0];
+
+        {
+          PSNR_STATS psnr2;
+          double frame_ssim2 = 0, weight = 0;
+#if CONFIG_VP9_POSTPROC
+          if (vpx_alloc_frame_buffer(&cm->post_proc_buffer,
+                                     recon->y_crop_width, recon->y_crop_height,
+                                     cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                     cm->use_highbitdepth,
+#endif
+                                     VP9_ENC_BORDER_IN_PIXELS,
+                                     cm->byte_alignment) < 0) {
+            vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                               "Failed to allocate post processing buffer");
+          }
+
+          vp9_deblock(cm->frame_to_show, &cm->post_proc_buffer,
+                      cm->lf.filter_level * 10 / 6);
+#endif
+          vpx_clear_system_state();
+
+#if CONFIG_VP9_HIGHBITDEPTH
+          calc_highbd_psnr(orig, pp, &psnr2, cpi->td.mb.e_mbd.bd,
+                           cpi->oxcf.input_bit_depth);
+#else
+          calc_psnr(orig, pp, &psnr2);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+          cpi->totalp_sq_error += psnr2.sse[0];
+          cpi->totalp_samples += psnr2.samples[0];
+          adjust_image_stat(psnr2.psnr[1], psnr2.psnr[2], psnr2.psnr[3],
+                            psnr2.psnr[0], &cpi->psnrp);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+          if (cm->use_highbitdepth) {
+            frame_ssim2 = vpx_highbd_calc_ssim(orig, recon, &weight,
+                                               (int)cm->bit_depth);
+          } else {
+            frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
+          }
+#else
+          frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+          cpi->worst_ssim = VPXMIN(cpi->worst_ssim, frame_ssim2);
+          cpi->summed_quality += frame_ssim2 * weight;
+          cpi->summed_weights += weight;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+          if (cm->use_highbitdepth) {
+            frame_ssim2 = vpx_highbd_calc_ssim(
+                orig, &cm->post_proc_buffer, &weight, (int)cm->bit_depth);
+          } else {
+            frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
+          }
+#else
+          frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+          cpi->summedp_quality += frame_ssim2 * weight;
+          cpi->summedp_weights += weight;
+#if 0
+          {
+            FILE *f = fopen("q_used.stt", "a");
+            fprintf(f, "%5d : Y%f7.3:U%f7.3:V%f7.3:F%f7.3:S%7.3f\n",
+                    cpi->common.current_video_frame, y2, u2, v2,
+                    frame_psnr2, frame_ssim2);
+            fclose(f);
+          }
+#endif
+        }
+      }
+      if (cpi->b_calculate_blockiness) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        if (!cm->use_highbitdepth)
+#endif
+        {
+          double frame_blockiness = vp9_get_blockiness(
+              cpi->Source->y_buffer, cpi->Source->y_stride,
+              cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
+              cpi->Source->y_width, cpi->Source->y_height);
+          cpi->worst_blockiness =
+              VPXMAX(cpi->worst_blockiness, frame_blockiness);
+          cpi->total_blockiness += frame_blockiness;
+        }
+      }
+
+      if (cpi->b_calculate_consistency) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        if (!cm->use_highbitdepth)
+#endif
+        {
+          double this_inconsistency = vpx_get_ssim_metrics(
+              cpi->Source->y_buffer, cpi->Source->y_stride,
+              cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
+              cpi->Source->y_width, cpi->Source->y_height, cpi->ssim_vars,
+              &cpi->metrics, 1);
+
+          const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
+          double consistency = vpx_sse_to_psnr(samples, peak,
+                                             (double)cpi->total_inconsistency);
+          if (consistency > 0.0)
+            cpi->worst_consistency =
+                VPXMIN(cpi->worst_consistency, consistency);
+          cpi->total_inconsistency += this_inconsistency;
+        }
+      }
+
+      if (cpi->b_calculate_ssimg) {
+        double y, u, v, frame_all;
+#if CONFIG_VP9_HIGHBITDEPTH
+        if (cm->use_highbitdepth) {
+          frame_all = vpx_highbd_calc_ssimg(cpi->Source, cm->frame_to_show, &y,
+                                            &u, &v, (int)cm->bit_depth);
+        } else {
+          frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u,
+                                     &v);
+        }
+#else
+        frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u, &v);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        adjust_image_stat(y, u, v, frame_all, &cpi->ssimg);
+      }
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (!cm->use_highbitdepth)
+#endif
+      {
+        double y, u, v, frame_all;
+        frame_all = vpx_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u,
+                                      &v);
+        adjust_image_stat(y, u, v, frame_all, &cpi->fastssim);
+        /* TODO(JBB): add 10/12 bit support */
+      }
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (!cm->use_highbitdepth)
+#endif
+      {
+        double y, u, v, frame_all;
+        frame_all = vpx_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v);
+        adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs);
+      }
+    }
+  }
+
+#endif
+
+  if (is_two_pass_svc(cpi)) {
+    if (cpi->svc.encode_empty_frame_state == ENCODING) {
+      cpi->svc.encode_empty_frame_state = ENCODED;
+      cpi->svc.encode_intra_empty_frame = 0;
+    }
+
+    if (cm->show_frame) {
+      ++cpi->svc.spatial_layer_to_encode;
+      if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
+        cpi->svc.spatial_layer_to_encode = 0;
+
+      // May need the empty frame after an visible frame.
+      cpi->svc.encode_empty_frame_state = NEED_TO_ENCODE;
+    }
+  } else if (is_one_pass_cbr_svc(cpi)) {
+    if (cm->show_frame) {
+      ++cpi->svc.spatial_layer_to_encode;
+      if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
+        cpi->svc.spatial_layer_to_encode = 0;
+    }
+  }
+  vpx_clear_system_state();
+  return 0;
+}
+
+int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
+                              vp9_ppflags_t *flags) {
+  VP9_COMMON *cm = &cpi->common;
+#if !CONFIG_VP9_POSTPROC
+  (void)flags;
+#endif
+
+  if (!cm->show_frame) {
+    return -1;
+  } else {
+    int ret;
+#if CONFIG_VP9_POSTPROC
+    ret = vp9_post_proc_frame(cm, dest, flags);
+#else
+    if (cm->frame_to_show) {
+      *dest = *cm->frame_to_show;
+      dest->y_width = cm->width;
+      dest->y_height = cm->height;
+      dest->uv_width = cm->width >> cm->subsampling_x;
+      dest->uv_height = cm->height >> cm->subsampling_y;
+      ret = 0;
+    } else {
+      ret = -1;
+    }
+#endif  // !CONFIG_VP9_POSTPROC
+    vpx_clear_system_state();
+    return ret;
+  }
+}
+
+int vp9_set_internal_size(VP9_COMP *cpi,
+                          VPX_SCALING horiz_mode, VPX_SCALING vert_mode) {
+  VP9_COMMON *cm = &cpi->common;
+  int hr = 0, hs = 0, vr = 0, vs = 0;
+
+  if (horiz_mode > ONETWO || vert_mode > ONETWO)
+    return -1;
+
+  Scale2Ratio(horiz_mode, &hr, &hs);
+  Scale2Ratio(vert_mode, &vr, &vs);
+
+  // always go to the next whole number
+  cm->width = (hs - 1 + cpi->oxcf.width * hr) / hs;
+  cm->height = (vs - 1 + cpi->oxcf.height * vr) / vs;
+  if (cm->current_video_frame) {
+    assert(cm->width <= cpi->initial_width);
+    assert(cm->height <= cpi->initial_height);
+  }
+
+  update_frame_size(cpi);
+
+  return 0;
+}
+
+int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
+                         unsigned int height) {
+  VP9_COMMON *cm = &cpi->common;
+#if CONFIG_VP9_HIGHBITDEPTH
+  check_initial_width(cpi, cm->use_highbitdepth, 1, 1);
+#else
+  check_initial_width(cpi, 1, 1);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  setup_denoiser_buffer(cpi);
+#endif
+
+  if (width) {
+    cm->width = width;
+    if (cm->width > cpi->initial_width) {
+      cm->width = cpi->initial_width;
+      printf("Warning: Desired width too large, changed to %d\n", cm->width);
+    }
+  }
+
+  if (height) {
+    cm->height = height;
+    if (cm->height > cpi->initial_height) {
+      cm->height = cpi->initial_height;
+      printf("Warning: Desired height too large, changed to %d\n", cm->height);
+    }
+  }
+  assert(cm->width <= cpi->initial_width);
+  assert(cm->height <= cpi->initial_height);
+
+  update_frame_size(cpi);
+
+  return 0;
+}
+
+void vp9_set_svc(VP9_COMP *cpi, int use_svc) {
+  cpi->use_svc = use_svc;
+  return;
+}
+
+int64_t vp9_get_y_sse(const YV12_BUFFER_CONFIG *a,
+                      const YV12_BUFFER_CONFIG *b) {
+  assert(a->y_crop_width == b->y_crop_width);
+  assert(a->y_crop_height == b->y_crop_height);
+
+  return get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
+                 a->y_crop_width, a->y_crop_height);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+int64_t vp9_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
+                             const YV12_BUFFER_CONFIG *b) {
+  assert(a->y_crop_width == b->y_crop_width);
+  assert(a->y_crop_height == b->y_crop_height);
+  assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
+  assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
+
+  return highbd_get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
+                        a->y_crop_width, a->y_crop_height);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+int vp9_get_quantizer(VP9_COMP *cpi) {
+  return cpi->common.base_qindex;
+}
+
+void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags) {
+  if (flags & (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF |
+               VP8_EFLAG_NO_REF_ARF)) {
+    int ref = 7;
+
+    if (flags & VP8_EFLAG_NO_REF_LAST)
+      ref ^= VP9_LAST_FLAG;
+
+    if (flags & VP8_EFLAG_NO_REF_GF)
+      ref ^= VP9_GOLD_FLAG;
+
+    if (flags & VP8_EFLAG_NO_REF_ARF)
+      ref ^= VP9_ALT_FLAG;
+
+    vp9_use_as_reference(cpi, ref);
+  }
+
+  if (flags & (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF |
+               VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_FORCE_GF |
+               VP8_EFLAG_FORCE_ARF)) {
+    int upd = 7;
+
+    if (flags & VP8_EFLAG_NO_UPD_LAST)
+      upd ^= VP9_LAST_FLAG;
+
+    if (flags & VP8_EFLAG_NO_UPD_GF)
+      upd ^= VP9_GOLD_FLAG;
+
+    if (flags & VP8_EFLAG_NO_UPD_ARF)
+      upd ^= VP9_ALT_FLAG;
+
+    vp9_update_reference(cpi, upd);
+  }
+
+  if (flags & VP8_EFLAG_NO_UPD_ENTROPY) {
+    vp9_update_entropy(cpi, 0);
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_encoder.h b/libvpx/libvpx/vp9/encoder/vp9_encoder.h
new file mode 100644
index 0000000..128b623
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_encoder.h
@@ -0,0 +1,756 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_ENCODER_H_
+#define VP9_ENCODER_VP9_ENCODER_H_
+
+#include <stdio.h>
+
+#include "./vpx_config.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "vpx/vp8cx.h"
+#if CONFIG_INTERNAL_STATS
+#include "vpx_dsp/ssim.h"
+#endif
+#include "vpx_dsp/variance.h"
+#include "vpx_ports/system_state.h"
+#include "vpx_util/vpx_thread.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_ppflags.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_thread_common.h"
+#include "vp9/common/vp9_onyxc_int.h"
+
+#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_encodemb.h"
+#include "vp9/encoder/vp9_firstpass.h"
+#include "vp9/encoder/vp9_lookahead.h"
+#include "vp9/encoder/vp9_mbgraph.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_noise_estimate.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_speed_features.h"
+#include "vp9/encoder/vp9_svc_layercontext.h"
+#include "vp9/encoder/vp9_tokenize.h"
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#include "vp9/encoder/vp9_denoiser.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// vp9 uses 10,000,000 ticks/second as time stamp
+#define TICKS_PER_SEC 10000000
+
+typedef struct {
+  int nmvjointcost[MV_JOINTS];
+  int nmvcosts[2][MV_VALS];
+  int nmvcosts_hp[2][MV_VALS];
+
+  vpx_prob segment_pred_probs[PREDICTION_PROBS];
+
+  unsigned char *last_frame_seg_map_copy;
+
+  // 0 = Intra, Last, GF, ARF
+  signed char last_ref_lf_deltas[MAX_REF_LF_DELTAS];
+  // 0 = ZERO_MV, MV
+  signed char last_mode_lf_deltas[MAX_MODE_LF_DELTAS];
+
+  FRAME_CONTEXT fc;
+} CODING_CONTEXT;
+
+
+typedef enum {
+  // encode_breakout is disabled.
+  ENCODE_BREAKOUT_DISABLED = 0,
+  // encode_breakout is enabled.
+  ENCODE_BREAKOUT_ENABLED = 1,
+  // encode_breakout is enabled with small max_thresh limit.
+  ENCODE_BREAKOUT_LIMITED = 2
+} ENCODE_BREAKOUT_TYPE;
+
+typedef enum {
+  NORMAL      = 0,
+  FOURFIVE    = 1,
+  THREEFIVE   = 2,
+  ONETWO      = 3
+} VPX_SCALING;
+
+typedef enum {
+  // Good Quality Fast Encoding. The encoder balances quality with the amount of
+  // time it takes to encode the output. Speed setting controls how fast.
+  GOOD,
+
+  // The encoder places priority on the quality of the output over encoding
+  // speed. The output is compressed at the highest possible quality. This
+  // option takes the longest amount of time to encode. Speed setting ignored.
+  BEST,
+
+  // Realtime/Live Encoding. This mode is optimized for realtime encoding (for
+  // example, capturing a television signal or feed from a live camera). Speed
+  // setting controls how fast.
+  REALTIME
+} MODE;
+
+typedef enum {
+  FRAMEFLAGS_KEY    = 1 << 0,
+  FRAMEFLAGS_GOLDEN = 1 << 1,
+  FRAMEFLAGS_ALTREF = 1 << 2,
+} FRAMETYPE_FLAGS;
+
+typedef enum {
+  NO_AQ = 0,
+  VARIANCE_AQ = 1,
+  COMPLEXITY_AQ = 2,
+  CYCLIC_REFRESH_AQ = 3,
+  EQUATOR360_AQ = 4,
+  AQ_MODE_COUNT  // This should always be the last member of the enum
+} AQ_MODE;
+
+typedef enum {
+  RESIZE_NONE = 0,    // No frame resizing allowed (except for SVC).
+  RESIZE_FIXED = 1,   // All frames are coded at the specified dimension.
+  RESIZE_DYNAMIC = 2  // Coded size of each frame is determined by the codec.
+} RESIZE_TYPE;
+
+typedef struct VP9EncoderConfig {
+  BITSTREAM_PROFILE profile;
+  vpx_bit_depth_t bit_depth;     // Codec bit-depth.
+  int width;  // width of data passed to the compressor
+  int height;  // height of data passed to the compressor
+  unsigned int input_bit_depth;  // Input bit depth.
+  double init_framerate;  // set to passed in framerate
+  int64_t target_bandwidth;  // bandwidth to be used in bits per second
+
+  int noise_sensitivity;  // pre processing blur: recommendation 0
+  int sharpness;  // sharpening output: recommendation 0:
+  int speed;
+  // maximum allowed bitrate for any intra frame in % of bitrate target.
+  unsigned int rc_max_intra_bitrate_pct;
+  // maximum allowed bitrate for any inter frame in % of bitrate target.
+  unsigned int rc_max_inter_bitrate_pct;
+  // percent of rate boost for golden frame in CBR mode.
+  unsigned int gf_cbr_boost_pct;
+
+  MODE mode;
+  int pass;
+
+  // Key Framing Operations
+  int auto_key;  // autodetect cut scenes and set the keyframes
+  int key_freq;  // maximum distance to key frame.
+
+  int lag_in_frames;  // how many frames lag before we start encoding
+
+  // ----------------------------------------------------------------
+  // DATARATE CONTROL OPTIONS
+
+  // vbr, cbr, constrained quality or constant quality
+  enum vpx_rc_mode rc_mode;
+
+  // buffer targeting aggressiveness
+  int under_shoot_pct;
+  int over_shoot_pct;
+
+  // buffering parameters
+  int64_t starting_buffer_level_ms;
+  int64_t optimal_buffer_level_ms;
+  int64_t maximum_buffer_size_ms;
+
+  // Frame drop threshold.
+  int drop_frames_water_mark;
+
+  // controlling quality
+  int fixed_q;
+  int worst_allowed_q;
+  int best_allowed_q;
+  int cq_level;
+  AQ_MODE aq_mode;  // Adaptive Quantization mode
+
+  // Internal frame size scaling.
+  RESIZE_TYPE resize_mode;
+  int scaled_frame_width;
+  int scaled_frame_height;
+
+  // Enable feature to reduce the frame quantization every x frames.
+  int frame_periodic_boost;
+
+  // two pass datarate control
+  int two_pass_vbrbias;        // two pass datarate control tweaks
+  int two_pass_vbrmin_section;
+  int two_pass_vbrmax_section;
+  // END DATARATE CONTROL OPTIONS
+  // ----------------------------------------------------------------
+
+  // Spatial and temporal scalability.
+  int ss_number_layers;  // Number of spatial layers.
+  int ts_number_layers;  // Number of temporal layers.
+  // Bitrate allocation for spatial layers.
+  int layer_target_bitrate[VPX_MAX_LAYERS];
+  int ss_target_bitrate[VPX_SS_MAX_LAYERS];
+  int ss_enable_auto_arf[VPX_SS_MAX_LAYERS];
+  // Bitrate allocation (CBR mode) and framerate factor, for temporal layers.
+  int ts_rate_decimator[VPX_TS_MAX_LAYERS];
+
+  int enable_auto_arf;
+
+  int encode_breakout;  // early breakout : for video conf recommend 800
+
+  /* Bitfield defining the error resiliency features to enable.
+   * Can provide decodable frames after losses in previous
+   * frames and decodable partitions after losses in the same frame.
+   */
+  unsigned int error_resilient_mode;
+
+  /* Bitfield defining the parallel decoding mode where the
+   * decoding in successive frames may be conducted in parallel
+   * just by decoding the frame headers.
+   */
+  unsigned int frame_parallel_decoding_mode;
+
+  int arnr_max_frames;
+  int arnr_strength;
+
+  int min_gf_interval;
+  int max_gf_interval;
+
+  int tile_columns;
+  int tile_rows;
+
+  int max_threads;
+
+  int target_level;
+
+  vpx_fixed_buf_t two_pass_stats_in;
+  struct vpx_codec_pkt_list *output_pkt_list;
+
+#if CONFIG_FP_MB_STATS
+  vpx_fixed_buf_t firstpass_mb_stats_in;
+#endif
+
+  vp8e_tuning tuning;
+  vp9e_tune_content content;
+#if CONFIG_VP9_HIGHBITDEPTH
+  int use_highbitdepth;
+#endif
+  vpx_color_space_t color_space;
+  vpx_color_range_t color_range;
+  int render_width;
+  int render_height;
+  VP9E_TEMPORAL_LAYERING_MODE temporal_layering_mode;
+} VP9EncoderConfig;
+
+static INLINE int is_lossless_requested(const VP9EncoderConfig *cfg) {
+  return cfg->best_allowed_q == 0 && cfg->worst_allowed_q == 0;
+}
+
+// TODO(jingning) All spatially adaptive variables should go to TileDataEnc.
+typedef struct TileDataEnc {
+  TileInfo tile_info;
+  int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
+  int mode_map[BLOCK_SIZES][MAX_MODES];
+} TileDataEnc;
+
+typedef struct RD_COUNTS {
+  vp9_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
+  int64_t comp_pred_diff[REFERENCE_MODES];
+  int64_t filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+  int m_search_count;
+  int ex_search_count;
+} RD_COUNTS;
+
+typedef struct ThreadData {
+  MACROBLOCK mb;
+  RD_COUNTS rd_counts;
+  FRAME_COUNTS *counts;
+
+  PICK_MODE_CONTEXT *leaf_tree;
+  PC_TREE *pc_tree;
+  PC_TREE *pc_root;
+} ThreadData;
+
+struct EncWorkerData;
+
+typedef struct ActiveMap {
+  int enabled;
+  int update;
+  unsigned char *map;
+} ActiveMap;
+
+typedef enum {
+  Y,
+  U,
+  V,
+  ALL
+} STAT_TYPE;
+
+typedef struct IMAGE_STAT {
+  double stat[ALL+1];
+  double worst;
+} ImageStat;
+
+#define CPB_WINDOW_SIZE 4
+#define FRAME_WINDOW_SIZE 128
+#define SAMPLE_RATE_GRACE_P 0.015
+#define VP9_LEVELS 14
+
+typedef enum {
+  LEVEL_UNKNOWN = 0,
+  LEVEL_1 = 10,
+  LEVEL_1_1 = 11,
+  LEVEL_2 = 20,
+  LEVEL_2_1 = 21,
+  LEVEL_3 = 30,
+  LEVEL_3_1 = 31,
+  LEVEL_4 = 40,
+  LEVEL_4_1 = 41,
+  LEVEL_5 = 50,
+  LEVEL_5_1 = 51,
+  LEVEL_5_2 = 52,
+  LEVEL_6 = 60,
+  LEVEL_6_1 = 61,
+  LEVEL_6_2 = 62,
+  LEVEL_MAX = 255
+} VP9_LEVEL;
+
+typedef struct {
+  VP9_LEVEL level;
+  uint64_t max_luma_sample_rate;
+  uint32_t max_luma_picture_size;
+  double average_bitrate;  // in kilobits per second
+  double max_cpb_size;  // in kilobits
+  double compression_ratio;
+  uint8_t max_col_tiles;
+  uint32_t min_altref_distance;
+  uint8_t max_ref_frame_buffers;
+} Vp9LevelSpec;
+
+typedef struct {
+  int64_t ts;  // timestamp
+  uint32_t luma_samples;
+  uint32_t size;  // in bytes
+} FrameRecord;
+
+typedef struct {
+  FrameRecord buf[FRAME_WINDOW_SIZE];
+  uint8_t start;
+  uint8_t len;
+} FrameWindowBuffer;
+
+typedef struct {
+  uint8_t seen_first_altref;
+  uint32_t frames_since_last_altref;
+  uint64_t total_compressed_size;
+  uint64_t total_uncompressed_size;
+  double time_encoded;  // in seconds
+  FrameWindowBuffer frame_window_buffer;
+  int ref_refresh_map;
+} Vp9LevelStats;
+
+typedef struct {
+  Vp9LevelStats level_stats;
+  Vp9LevelSpec level_spec;
+} Vp9LevelInfo;
+
+typedef struct VP9_COMP {
+  QUANTS quants;
+  ThreadData td;
+  MB_MODE_INFO_EXT *mbmi_ext_base;
+  DECLARE_ALIGNED(16, int16_t, y_dequant[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_dequant[QINDEX_RANGE][8]);
+  VP9_COMMON common;
+  VP9EncoderConfig oxcf;
+  struct lookahead_ctx    *lookahead;
+  struct lookahead_entry  *alt_ref_source;
+
+  YV12_BUFFER_CONFIG *Source;
+  YV12_BUFFER_CONFIG *Last_Source;  // NULL for first frame and alt_ref frames
+  YV12_BUFFER_CONFIG *un_scaled_source;
+  YV12_BUFFER_CONFIG scaled_source;
+  YV12_BUFFER_CONFIG *unscaled_last_source;
+  YV12_BUFFER_CONFIG scaled_last_source;
+
+  TileDataEnc *tile_data;
+  int allocated_tiles;  // Keep track of memory allocated for tiles.
+
+  // For a still frame, this flag is set to 1 to skip partition search.
+  int partition_search_skippable_frame;
+
+  int scaled_ref_idx[MAX_REF_FRAMES];
+  int lst_fb_idx;
+  int gld_fb_idx;
+  int alt_fb_idx;
+
+  int refresh_last_frame;
+  int refresh_golden_frame;
+  int refresh_alt_ref_frame;
+
+  int ext_refresh_frame_flags_pending;
+  int ext_refresh_last_frame;
+  int ext_refresh_golden_frame;
+  int ext_refresh_alt_ref_frame;
+
+  int ext_refresh_frame_context_pending;
+  int ext_refresh_frame_context;
+
+  YV12_BUFFER_CONFIG last_frame_uf;
+
+  TOKENEXTRA *tile_tok[4][1 << 6];
+  uint32_t tok_count[4][1 << 6];
+
+  // Ambient reconstruction err target for force key frames
+  int64_t ambient_err;
+
+  RD_OPT rd;
+
+  CODING_CONTEXT coding_context;
+
+  int *nmvcosts[2];
+  int *nmvcosts_hp[2];
+  int *nmvsadcosts[2];
+  int *nmvsadcosts_hp[2];
+
+  int64_t last_time_stamp_seen;
+  int64_t last_end_time_stamp_seen;
+  int64_t first_time_stamp_ever;
+
+  RATE_CONTROL rc;
+  double framerate;
+
+  int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE];
+
+  struct vpx_codec_pkt_list  *output_pkt_list;
+
+  MBGRAPH_FRAME_STATS mbgraph_stats[MAX_LAG_BUFFERS];
+  int mbgraph_n_frames;             // number of frames filled in the above
+  int static_mb_pct;                // % forced skip mbs by segmentation
+  int ref_frame_flags;
+
+  SPEED_FEATURES sf;
+
+  uint32_t max_mv_magnitude;
+  int mv_step_param;
+
+  int allow_comp_inter_inter;
+
+  // Default value is 1. From first pass stats, encode_breakout may be disabled.
+  ENCODE_BREAKOUT_TYPE allow_encode_breakout;
+
+  // Get threshold from external input. A suggested threshold is 800 for HD
+  // clips, and 300 for < HD clips.
+  int encode_breakout;
+
+  uint8_t *segmentation_map;
+
+  // segment threashold for encode breakout
+  int segment_encode_breakout[MAX_SEGMENTS];
+
+  CYCLIC_REFRESH *cyclic_refresh;
+  ActiveMap active_map;
+
+  fractional_mv_step_fp *find_fractional_mv_step;
+  vp9_full_search_fn_t full_search_sad;
+  vp9_diamond_search_fn_t diamond_search_sad;
+  vp9_variance_fn_ptr_t fn_ptr[BLOCK_SIZES];
+  uint64_t time_receive_data;
+  uint64_t time_compress_data;
+  uint64_t time_pick_lpf;
+  uint64_t time_encode_sb_row;
+
+#if CONFIG_FP_MB_STATS
+  int use_fp_mb_stats;
+#endif
+
+  TWO_PASS twopass;
+
+  YV12_BUFFER_CONFIG alt_ref_buffer;
+
+#if CONFIG_INTERNAL_STATS
+  unsigned int mode_chosen_counts[MAX_MODES];
+
+  int count;
+  uint64_t total_sq_error;
+  uint64_t total_samples;
+  ImageStat psnr;
+
+  uint64_t totalp_sq_error;
+  uint64_t totalp_samples;
+  ImageStat psnrp;
+
+  double total_blockiness;
+  double worst_blockiness;
+
+  int    bytes;
+  double summed_quality;
+  double summed_weights;
+  double summedp_quality;
+  double summedp_weights;
+  unsigned int tot_recode_hits;
+  double worst_ssim;
+
+  ImageStat ssimg;
+  ImageStat fastssim;
+  ImageStat psnrhvs;
+
+  int b_calculate_ssimg;
+  int b_calculate_blockiness;
+
+  int b_calculate_consistency;
+
+  double total_inconsistency;
+  double worst_consistency;
+  Ssimv *ssim_vars;
+  Metrics metrics;
+#endif
+  int b_calculate_psnr;
+
+  int droppable;
+
+  int initial_width;
+  int initial_height;
+  int initial_mbs;  // Number of MBs in the full-size frame; to be used to
+                    // normalize the firstpass stats. This will differ from the
+                    // number of MBs in the current frame when the frame is
+                    // scaled.
+
+  int use_svc;
+
+  SVC svc;
+
+  // Store frame variance info in SOURCE_VAR_BASED_PARTITION search type.
+  diff *source_diff_var;
+  // The threshold used in SOURCE_VAR_BASED_PARTITION search type.
+  unsigned int source_var_thresh;
+  int frames_till_next_var_check;
+
+  int frame_flags;
+
+  search_site_config ss_cfg;
+
+  int mbmode_cost[INTRA_MODES];
+  unsigned int inter_mode_cost[INTER_MODE_CONTEXTS][INTER_MODES];
+  int intra_uv_mode_cost[FRAME_TYPES][INTRA_MODES][INTRA_MODES];
+  int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
+  int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
+  int partition_cost[PARTITION_CONTEXTS][PARTITION_TYPES];
+
+  int multi_arf_allowed;
+  int multi_arf_enabled;
+  int multi_arf_last_grp_enabled;
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  VP9_DENOISER denoiser;
+#endif
+
+  int resize_pending;
+  int resize_state;
+  int external_resize;
+  int resize_scale_num;
+  int resize_scale_den;
+  int resize_avg_qp;
+  int resize_buffer_underflow;
+  int resize_count;
+
+  int use_skin_detection;
+
+  int target_level;
+
+  NOISE_ESTIMATE noise_estimate;
+
+  // Count on how many consecutive times a block uses small/zeromv for encoding.
+  uint8_t *consec_zero_mv;
+
+  // VAR_BASED_PARTITION thresholds
+  // 0 - threshold_64x64; 1 - threshold_32x32;
+  // 2 - threshold_16x16; 3 - vbp_threshold_8x8;
+  int64_t vbp_thresholds[4];
+  int64_t vbp_threshold_minmax;
+  int64_t vbp_threshold_sad;
+  BLOCK_SIZE vbp_bsize_min;
+
+  // Multi-threading
+  int num_workers;
+  VPxWorker *workers;
+  struct EncWorkerData *tile_thr_data;
+  VP9LfSync lf_row_sync;
+
+  int keep_level_stats;
+  Vp9LevelInfo level_info;
+} VP9_COMP;
+
+void vp9_initialize_enc(void);
+
+struct VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
+                                       BufferPool *const pool);
+void vp9_remove_compressor(VP9_COMP *cpi);
+
+void vp9_change_config(VP9_COMP *cpi, const VP9EncoderConfig *oxcf);
+
+  // receive a frames worth of data. caller can assume that a copy of this
+  // frame is made and not just a copy of the pointer..
+int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
+                          YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
+                          int64_t end_time_stamp);
+
+int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
+                            size_t *size, uint8_t *dest,
+                            int64_t *time_stamp, int64_t *time_end, int flush);
+
+int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
+                              vp9_ppflags_t *flags);
+
+int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags);
+
+void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags);
+
+int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
+                           YV12_BUFFER_CONFIG *sd);
+
+int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
+                          YV12_BUFFER_CONFIG *sd);
+
+int vp9_update_entropy(VP9_COMP *cpi, int update);
+
+int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols);
+
+int vp9_get_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols);
+
+int vp9_set_internal_size(VP9_COMP *cpi,
+                          VPX_SCALING horiz_mode, VPX_SCALING vert_mode);
+
+int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
+                         unsigned int height);
+
+void vp9_set_svc(VP9_COMP *cpi, int use_svc);
+
+int vp9_get_quantizer(struct VP9_COMP *cpi);
+
+static INLINE int frame_is_kf_gf_arf(const VP9_COMP *cpi) {
+  return frame_is_intra_only(&cpi->common) ||
+         cpi->refresh_alt_ref_frame ||
+         (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref);
+}
+
+static INLINE int get_ref_frame_map_idx(const VP9_COMP *cpi,
+                                        MV_REFERENCE_FRAME ref_frame) {
+  if (ref_frame == LAST_FRAME) {
+    return cpi->lst_fb_idx;
+  } else if (ref_frame == GOLDEN_FRAME) {
+    return cpi->gld_fb_idx;
+  } else {
+    return cpi->alt_fb_idx;
+  }
+}
+
+static INLINE int get_ref_frame_buf_idx(const VP9_COMP *const cpi,
+                                        int ref_frame) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const int map_idx = get_ref_frame_map_idx(cpi, ref_frame);
+  return (map_idx != INVALID_IDX) ? cm->ref_frame_map[map_idx] : INVALID_IDX;
+}
+
+static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer(
+    VP9_COMP *cpi, MV_REFERENCE_FRAME ref_frame) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+  return
+      buf_idx != INVALID_IDX ? &cm->buffer_pool->frame_bufs[buf_idx].buf : NULL;
+}
+
+static INLINE int get_token_alloc(int mb_rows, int mb_cols) {
+  // TODO(JBB): double check we can't exceed this token count if we have a
+  // 32x32 transform crossing a boundary at a multiple of 16.
+  // mb_rows, cols are in units of 16 pixels. We assume 3 planes all at full
+  // resolution. We assume up to 1 token per pixel, and then allow
+  // a head room of 4.
+  return mb_rows * mb_cols * (16 * 16 * 3 + 4);
+}
+
+// Get the allocated token size for a tile. It does the same calculation as in
+// the frame token allocation.
+static INLINE int allocated_tokens(TileInfo tile) {
+  int tile_mb_rows = (tile.mi_row_end - tile.mi_row_start + 1) >> 1;
+  int tile_mb_cols = (tile.mi_col_end - tile.mi_col_start + 1) >> 1;
+
+  return get_token_alloc(tile_mb_rows, tile_mb_cols);
+}
+
+int64_t vp9_get_y_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b);
+#if CONFIG_VP9_HIGHBITDEPTH
+int64_t vp9_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
+                             const YV12_BUFFER_CONFIG *b);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_scale_references(VP9_COMP *cpi);
+
+void vp9_update_reference_frames(VP9_COMP *cpi);
+
+void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv);
+
+YV12_BUFFER_CONFIG *vp9_svc_twostage_scale(VP9_COMMON *cm,
+                                           YV12_BUFFER_CONFIG *unscaled,
+                                           YV12_BUFFER_CONFIG *scaled,
+                                           YV12_BUFFER_CONFIG *scaled_temp);
+
+YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
+                                          YV12_BUFFER_CONFIG *unscaled,
+                                          YV12_BUFFER_CONFIG *scaled,
+                                          int use_normative_scaler);
+
+void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags);
+
+static INLINE int is_two_pass_svc(const struct VP9_COMP *const cpi) {
+  return cpi->use_svc && cpi->oxcf.pass != 0;
+}
+
+static INLINE int is_one_pass_cbr_svc(const struct VP9_COMP *const cpi) {
+  return (cpi->use_svc && cpi->oxcf.pass == 0);
+}
+
+static INLINE int is_altref_enabled(const VP9_COMP *const cpi) {
+  return cpi->oxcf.mode != REALTIME && cpi->oxcf.lag_in_frames > 0 &&
+         (cpi->oxcf.enable_auto_arf &&
+          (!is_two_pass_svc(cpi) ||
+           cpi->oxcf.ss_enable_auto_arf[cpi->svc.spatial_layer_id]));
+}
+
+static INLINE void set_ref_ptrs(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                MV_REFERENCE_FRAME ref0,
+                                MV_REFERENCE_FRAME ref1) {
+  xd->block_refs[0] = &cm->frame_refs[ref0 >= LAST_FRAME ? ref0 - LAST_FRAME
+                                                         : 0];
+  xd->block_refs[1] = &cm->frame_refs[ref1 >= LAST_FRAME ? ref1 - LAST_FRAME
+                                                         : 0];
+}
+
+static INLINE int get_chessboard_index(const int frame_index) {
+  return frame_index & 0x1;
+}
+
+static INLINE int *cond_cost_list(const struct VP9_COMP *cpi, int *cost_list) {
+  return cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL;
+}
+
+VP9_LEVEL vp9_get_level(const Vp9LevelSpec *const level_spec);
+
+void vp9_new_framerate(VP9_COMP *cpi, double framerate);
+
+#define LAYER_IDS_TO_IDX(sl, tl, num_tl) ((sl) * (num_tl) + (tl))
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_ENCODER_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_ethread.c b/libvpx/libvpx/vp9/encoder/vp9_ethread.c
new file mode 100644
index 0000000..1d1926c
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_ethread.c
@@ -0,0 +1,203 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/encoder/vp9_encodeframe.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_ethread.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+static void accumulate_rd_opt(ThreadData *td, ThreadData *td_t) {
+  int i, j, k, l, m, n;
+
+  for (i = 0; i < REFERENCE_MODES; i++)
+    td->rd_counts.comp_pred_diff[i] += td_t->rd_counts.comp_pred_diff[i];
+
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+    td->rd_counts.filter_diff[i] += td_t->rd_counts.filter_diff[i];
+
+  for (i = 0; i < TX_SIZES; i++)
+    for (j = 0; j < PLANE_TYPES; j++)
+      for (k = 0; k < REF_TYPES; k++)
+        for (l = 0; l < COEF_BANDS; l++)
+          for (m = 0; m < COEFF_CONTEXTS; m++)
+            for (n = 0; n < ENTROPY_TOKENS; n++)
+              td->rd_counts.coef_counts[i][j][k][l][m][n] +=
+                  td_t->rd_counts.coef_counts[i][j][k][l][m][n];
+
+  // Counts of all motion searches and exhuastive mesh searches.
+  td->rd_counts.m_search_count += td_t->rd_counts.m_search_count;
+  td->rd_counts.ex_search_count += td_t->rd_counts.ex_search_count;
+}
+
+static int enc_worker_hook(EncWorkerData *const thread_data, void *unused) {
+  VP9_COMP *const cpi = thread_data->cpi;
+  const VP9_COMMON *const cm = &cpi->common;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  int t;
+
+  (void) unused;
+
+  for (t = thread_data->start; t < tile_rows * tile_cols;
+      t += cpi->num_workers) {
+    int tile_row = t / tile_cols;
+    int tile_col = t % tile_cols;
+
+    vp9_encode_tile(cpi, thread_data->td, tile_row, tile_col);
+  }
+
+  return 0;
+}
+
+static int get_max_tile_cols(VP9_COMP *cpi) {
+  const int aligned_width = ALIGN_POWER_OF_TWO(cpi->oxcf.width, MI_SIZE_LOG2);
+  int mi_cols = aligned_width >> MI_SIZE_LOG2;
+  int min_log2_tile_cols, max_log2_tile_cols;
+  int log2_tile_cols;
+
+  vp9_get_tile_n_bits(mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
+  log2_tile_cols = clamp(cpi->oxcf.tile_columns,
+                   min_log2_tile_cols, max_log2_tile_cols);
+  return (1 << log2_tile_cols);
+}
+
+void vp9_encode_tiles_mt(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
+  const int num_workers = VPXMIN(cpi->oxcf.max_threads, tile_cols);
+  int i;
+
+  vp9_init_tile_data(cpi);
+
+  // Only run once to create threads and allocate thread data.
+  if (cpi->num_workers == 0) {
+    int allocated_workers = num_workers;
+
+    // While using SVC, we need to allocate threads according to the highest
+    // resolution.
+    if (cpi->use_svc) {
+      int max_tile_cols = get_max_tile_cols(cpi);
+      allocated_workers = VPXMIN(cpi->oxcf.max_threads, max_tile_cols);
+    }
+
+    CHECK_MEM_ERROR(cm, cpi->workers,
+                    vpx_malloc(allocated_workers * sizeof(*cpi->workers)));
+
+    CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
+                    vpx_calloc(allocated_workers,
+                    sizeof(*cpi->tile_thr_data)));
+
+    for (i = 0; i < allocated_workers; i++) {
+      VPxWorker *const worker = &cpi->workers[i];
+      EncWorkerData *thread_data = &cpi->tile_thr_data[i];
+
+      ++cpi->num_workers;
+      winterface->init(worker);
+
+      if (i < allocated_workers - 1) {
+        thread_data->cpi = cpi;
+
+        // Allocate thread data.
+        CHECK_MEM_ERROR(cm, thread_data->td,
+                        vpx_memalign(32, sizeof(*thread_data->td)));
+        vp9_zero(*thread_data->td);
+
+        // Set up pc_tree.
+        thread_data->td->leaf_tree = NULL;
+        thread_data->td->pc_tree = NULL;
+        vp9_setup_pc_tree(cm, thread_data->td);
+
+        // Allocate frame counters in thread data.
+        CHECK_MEM_ERROR(cm, thread_data->td->counts,
+                        vpx_calloc(1, sizeof(*thread_data->td->counts)));
+
+        // Create threads
+        if (!winterface->reset(worker))
+          vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                             "Tile encoder thread creation failed");
+      } else {
+        // Main thread acts as a worker and uses the thread data in cpi.
+        thread_data->cpi = cpi;
+        thread_data->td = &cpi->td;
+      }
+
+      winterface->sync(worker);
+    }
+  }
+
+  for (i = 0; i < num_workers; i++) {
+    VPxWorker *const worker = &cpi->workers[i];
+    EncWorkerData *thread_data;
+
+    worker->hook = (VPxWorkerHook)enc_worker_hook;
+    worker->data1 = &cpi->tile_thr_data[i];
+    worker->data2 = NULL;
+    thread_data = (EncWorkerData*)worker->data1;
+
+    // Before encoding a frame, copy the thread data from cpi.
+    if (thread_data->td != &cpi->td) {
+      thread_data->td->mb = cpi->td.mb;
+      thread_data->td->rd_counts = cpi->td.rd_counts;
+    }
+    if (thread_data->td->counts != &cpi->common.counts) {
+      memcpy(thread_data->td->counts, &cpi->common.counts,
+             sizeof(cpi->common.counts));
+    }
+
+    // Handle use_nonrd_pick_mode case.
+    if (cpi->sf.use_nonrd_pick_mode) {
+      MACROBLOCK *const x = &thread_data->td->mb;
+      MACROBLOCKD *const xd = &x->e_mbd;
+      struct macroblock_plane *const p = x->plane;
+      struct macroblockd_plane *const pd = xd->plane;
+      PICK_MODE_CONTEXT *ctx = &thread_data->td->pc_root->none;
+      int j;
+
+      for (j = 0; j < MAX_MB_PLANE; ++j) {
+        p[j].coeff = ctx->coeff_pbuf[j][0];
+        p[j].qcoeff = ctx->qcoeff_pbuf[j][0];
+        pd[j].dqcoeff = ctx->dqcoeff_pbuf[j][0];
+        p[j].eobs = ctx->eobs_pbuf[j][0];
+      }
+    }
+  }
+
+  // Encode a frame
+  for (i = 0; i < num_workers; i++) {
+    VPxWorker *const worker = &cpi->workers[i];
+    EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
+
+    // Set the starting tile for each thread.
+    thread_data->start = i;
+
+    if (i == cpi->num_workers - 1)
+      winterface->execute(worker);
+    else
+      winterface->launch(worker);
+  }
+
+  // Encoding ends.
+  for (i = 0; i < num_workers; i++) {
+    VPxWorker *const worker = &cpi->workers[i];
+    winterface->sync(worker);
+  }
+
+  for (i = 0; i < num_workers; i++) {
+    VPxWorker *const worker = &cpi->workers[i];
+    EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
+
+    // Accumulate counters.
+    if (i < cpi->num_workers - 1) {
+      vp9_accumulate_frame_counts(&cm->counts, thread_data->td->counts, 0);
+      accumulate_rd_opt(&cpi->td, thread_data->td);
+    }
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_ethread.h b/libvpx/libvpx/vp9/encoder/vp9_ethread.h
new file mode 100644
index 0000000..1efa4dc
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_ethread.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_ETHREAD_H_
+#define VP9_ENCODER_VP9_ETHREAD_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9_COMP;
+struct ThreadData;
+
+typedef struct EncWorkerData {
+  struct VP9_COMP *cpi;
+  struct ThreadData *td;
+  int start;
+} EncWorkerData;
+
+void vp9_encode_tiles_mt(struct VP9_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_ETHREAD_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_extend.c b/libvpx/libvpx/vp9/encoder/vp9_extend.c
new file mode 100644
index 0000000..92585b8
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_extend.c
@@ -0,0 +1,201 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/encoder/vp9_extend.h"
+
+static void copy_and_extend_plane(const uint8_t *src, int src_pitch,
+                                  uint8_t *dst, int dst_pitch,
+                                  int w, int h,
+                                  int extend_top, int extend_left,
+                                  int extend_bottom, int extend_right) {
+  int i, linesize;
+
+  // copy the left and right most columns out
+  const uint8_t *src_ptr1 = src;
+  const uint8_t *src_ptr2 = src + w - 1;
+  uint8_t *dst_ptr1 = dst - extend_left;
+  uint8_t *dst_ptr2 = dst + w;
+
+  for (i = 0; i < h; i++) {
+    memset(dst_ptr1, src_ptr1[0], extend_left);
+    memcpy(dst_ptr1 + extend_left, src_ptr1, w);
+    memset(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_pitch;
+    src_ptr2 += src_pitch;
+    dst_ptr1 += dst_pitch;
+    dst_ptr2 += dst_pitch;
+  }
+
+  // Now copy the top and bottom lines into each line of the respective
+  // borders
+  src_ptr1 = dst - extend_left;
+  src_ptr2 = dst + dst_pitch * (h - 1) - extend_left;
+  dst_ptr1 = dst + dst_pitch * (-extend_top) - extend_left;
+  dst_ptr2 = dst + dst_pitch * (h) - extend_left;
+  linesize = extend_left + extend_right + w;
+
+  for (i = 0; i < extend_top; i++) {
+    memcpy(dst_ptr1, src_ptr1, linesize);
+    dst_ptr1 += dst_pitch;
+  }
+
+  for (i = 0; i < extend_bottom; i++) {
+    memcpy(dst_ptr2, src_ptr2, linesize);
+    dst_ptr2 += dst_pitch;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_copy_and_extend_plane(const uint8_t *src8, int src_pitch,
+                                         uint8_t *dst8, int dst_pitch,
+                                         int w, int h,
+                                         int extend_top, int extend_left,
+                                         int extend_bottom, int extend_right) {
+  int i, linesize;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+
+  // copy the left and right most columns out
+  const uint16_t *src_ptr1 = src;
+  const uint16_t *src_ptr2 = src + w - 1;
+  uint16_t *dst_ptr1 = dst - extend_left;
+  uint16_t *dst_ptr2 = dst + w;
+
+  for (i = 0; i < h; i++) {
+    vpx_memset16(dst_ptr1, src_ptr1[0], extend_left);
+    memcpy(dst_ptr1 + extend_left, src_ptr1, w * sizeof(src_ptr1[0]));
+    vpx_memset16(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_pitch;
+    src_ptr2 += src_pitch;
+    dst_ptr1 += dst_pitch;
+    dst_ptr2 += dst_pitch;
+  }
+
+  // Now copy the top and bottom lines into each line of the respective
+  // borders
+  src_ptr1 = dst - extend_left;
+  src_ptr2 = dst + dst_pitch * (h - 1) - extend_left;
+  dst_ptr1 = dst + dst_pitch * (-extend_top) - extend_left;
+  dst_ptr2 = dst + dst_pitch * (h) - extend_left;
+  linesize = extend_left + extend_right + w;
+
+  for (i = 0; i < extend_top; i++) {
+    memcpy(dst_ptr1, src_ptr1, linesize * sizeof(src_ptr1[0]));
+    dst_ptr1 += dst_pitch;
+  }
+
+  for (i = 0; i < extend_bottom; i++) {
+    memcpy(dst_ptr2, src_ptr2, linesize * sizeof(src_ptr2[0]));
+    dst_ptr2 += dst_pitch;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_copy_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                               YV12_BUFFER_CONFIG *dst) {
+  // Extend src frame in buffer
+  // Altref filtering assumes 16 pixel extension
+  const int et_y = 16;
+  const int el_y = 16;
+  // Motion estimation may use src block variance with the block size up
+  // to 64x64, so the right and bottom need to be extended to 64 multiple
+  // or up to 16, whichever is greater.
+  const int er_y =
+      VPXMAX(src->y_width + 16, ALIGN_POWER_OF_TWO(src->y_width, 6)) -
+      src->y_crop_width;
+  const int eb_y =
+      VPXMAX(src->y_height + 16, ALIGN_POWER_OF_TWO(src->y_height, 6)) -
+      src->y_crop_height;
+  const int uv_width_subsampling = (src->uv_width != src->y_width);
+  const int uv_height_subsampling = (src->uv_height != src->y_height);
+  const int et_uv = et_y >> uv_height_subsampling;
+  const int el_uv = el_y >> uv_width_subsampling;
+  const int eb_uv = eb_y >> uv_height_subsampling;
+  const int er_uv = er_y >> uv_width_subsampling;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+    highbd_copy_and_extend_plane(src->y_buffer, src->y_stride,
+                                 dst->y_buffer, dst->y_stride,
+                                 src->y_crop_width, src->y_crop_height,
+                                 et_y, el_y, eb_y, er_y);
+
+    highbd_copy_and_extend_plane(src->u_buffer, src->uv_stride,
+                                 dst->u_buffer, dst->uv_stride,
+                                 src->uv_crop_width, src->uv_crop_height,
+                                 et_uv, el_uv, eb_uv, er_uv);
+
+    highbd_copy_and_extend_plane(src->v_buffer, src->uv_stride,
+                                 dst->v_buffer, dst->uv_stride,
+                                 src->uv_crop_width, src->uv_crop_height,
+                                 et_uv, el_uv, eb_uv, er_uv);
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  copy_and_extend_plane(src->y_buffer, src->y_stride,
+                        dst->y_buffer, dst->y_stride,
+                        src->y_crop_width, src->y_crop_height,
+                        et_y, el_y, eb_y, er_y);
+
+  copy_and_extend_plane(src->u_buffer, src->uv_stride,
+                        dst->u_buffer, dst->uv_stride,
+                        src->uv_crop_width, src->uv_crop_height,
+                        et_uv, el_uv, eb_uv, er_uv);
+
+  copy_and_extend_plane(src->v_buffer, src->uv_stride,
+                        dst->v_buffer, dst->uv_stride,
+                        src->uv_crop_width, src->uv_crop_height,
+                        et_uv, el_uv, eb_uv, er_uv);
+}
+
+void vp9_copy_and_extend_frame_with_rect(const YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst,
+                                         int srcy, int srcx,
+                                         int srch, int srcw) {
+  // If the side is not touching the bounder then don't extend.
+  const int et_y = srcy ? 0 : dst->border;
+  const int el_y = srcx ? 0 : dst->border;
+  const int eb_y = srcy + srch != src->y_height ? 0 :
+                      dst->border + dst->y_height - src->y_height;
+  const int er_y = srcx + srcw != src->y_width ? 0 :
+                      dst->border + dst->y_width - src->y_width;
+  const int src_y_offset = srcy * src->y_stride + srcx;
+  const int dst_y_offset = srcy * dst->y_stride + srcx;
+
+  const int et_uv = ROUND_POWER_OF_TWO(et_y, 1);
+  const int el_uv = ROUND_POWER_OF_TWO(el_y, 1);
+  const int eb_uv = ROUND_POWER_OF_TWO(eb_y, 1);
+  const int er_uv = ROUND_POWER_OF_TWO(er_y, 1);
+  const int src_uv_offset = ((srcy * src->uv_stride) >> 1) + (srcx >> 1);
+  const int dst_uv_offset = ((srcy * dst->uv_stride) >> 1) + (srcx >> 1);
+  const int srch_uv = ROUND_POWER_OF_TWO(srch, 1);
+  const int srcw_uv = ROUND_POWER_OF_TWO(srcw, 1);
+
+  copy_and_extend_plane(src->y_buffer + src_y_offset, src->y_stride,
+                        dst->y_buffer + dst_y_offset, dst->y_stride,
+                        srcw, srch,
+                        et_y, el_y, eb_y, er_y);
+
+  copy_and_extend_plane(src->u_buffer + src_uv_offset, src->uv_stride,
+                        dst->u_buffer + dst_uv_offset, dst->uv_stride,
+                        srcw_uv, srch_uv,
+                        et_uv, el_uv, eb_uv, er_uv);
+
+  copy_and_extend_plane(src->v_buffer + src_uv_offset, src->uv_stride,
+                        dst->v_buffer + dst_uv_offset, dst->uv_stride,
+                        srcw_uv, srch_uv,
+                        et_uv, el_uv, eb_uv, er_uv);
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_extend.h b/libvpx/libvpx/vp9/encoder/vp9_extend.h
new file mode 100644
index 0000000..058fe09
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_extend.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_EXTEND_H_
+#define VP9_ENCODER_VP9_EXTEND_H_
+
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+void vp9_copy_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                               YV12_BUFFER_CONFIG *dst);
+
+void vp9_copy_and_extend_frame_with_rect(const YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst,
+                                         int srcy, int srcx,
+                                         int srch, int srcw);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_EXTEND_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_firstpass.c b/libvpx/libvpx/vp9/encoder/vp9_firstpass.c
new file mode 100644
index 0000000..53a3ec7
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_firstpass.c
@@ -0,0 +1,2937 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/system_state.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vpx_scale/yv12config.h"
+
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconinter.h"  // vp9_setup_dst_planes()
+#include "vp9/encoder/vp9_aq_variance.h"
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_encodeframe.h"
+#include "vp9/encoder/vp9_encodemb.h"
+#include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_extend.h"
+#include "vp9/encoder/vp9_firstpass.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vpx_dsp/variance.h"
+
+#define OUTPUT_FPF          0
+#define ARF_STATS_OUTPUT    0
+
+#define BOOST_BREAKOUT      12.5
+#define BOOST_FACTOR        12.5
+#define FACTOR_PT_LOW       0.70
+#define FACTOR_PT_HIGH      0.90
+#define FIRST_PASS_Q        10.0
+#define GF_MAX_BOOST        96.0
+#define INTRA_MODE_PENALTY  1024
+#define KF_MAX_BOOST        128.0
+#define MIN_ARF_GF_BOOST    240
+#define MIN_DECAY_FACTOR    0.01
+#define MIN_KF_BOOST        300
+#define NEW_MV_MODE_PENALTY 32
+#define SVC_FACTOR_PT_LOW   0.45
+#define DARK_THRESH         64
+#define DEFAULT_GRP_WEIGHT  1.0
+#define RC_FACTOR_MIN       0.75
+#define RC_FACTOR_MAX       1.75
+
+
+#define NCOUNT_INTRA_THRESH 8192
+#define NCOUNT_INTRA_FACTOR 3
+
+
+#define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x) - 0.000001 : (x) + 0.000001)
+
+#if ARF_STATS_OUTPUT
+unsigned int arf_count = 0;
+#endif
+
+// Resets the first pass file to the given position using a relative seek from
+// the current position.
+static void reset_fpf_position(TWO_PASS *p,
+                               const FIRSTPASS_STATS *position) {
+  p->stats_in = position;
+}
+
+// Read frame stats at an offset from the current position.
+static const FIRSTPASS_STATS *read_frame_stats(const TWO_PASS *p, int offset) {
+  if ((offset >= 0 && p->stats_in + offset >= p->stats_in_end) ||
+      (offset < 0 && p->stats_in + offset < p->stats_in_start)) {
+    return NULL;
+  }
+
+  return &p->stats_in[offset];
+}
+
+static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
+  if (p->stats_in >= p->stats_in_end)
+    return EOF;
+
+  *fps = *p->stats_in;
+  ++p->stats_in;
+  return 1;
+}
+
+static void output_stats(FIRSTPASS_STATS *stats,
+                         struct vpx_codec_pkt_list *pktlist) {
+  struct vpx_codec_cx_pkt pkt;
+  pkt.kind = VPX_CODEC_STATS_PKT;
+  pkt.data.twopass_stats.buf = stats;
+  pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
+  vpx_codec_pkt_list_add(pktlist, &pkt);
+
+// TEMP debug code
+#if OUTPUT_FPF
+  {
+    FILE *fpfile;
+    fpfile = fopen("firstpass.stt", "a");
+
+    fprintf(fpfile, "%12.0lf %12.4lf %12.0lf %12.0lf %12.0lf %12.4lf %12.4lf"
+            "%12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf"
+            "%12.4lf %12.4lf %12.4lf %12.0lf %12.0lf %12.0lf %12.4lf\n",
+            stats->frame,
+            stats->weight,
+            stats->intra_error,
+            stats->coded_error,
+            stats->sr_coded_error,
+            stats->pcnt_inter,
+            stats->pcnt_motion,
+            stats->pcnt_second_ref,
+            stats->pcnt_neutral,
+            stats->intra_skip_pct,
+            stats->intra_smooth_pct,
+            stats->inactive_zone_rows,
+            stats->inactive_zone_cols,
+            stats->MVr,
+            stats->mvr_abs,
+            stats->MVc,
+            stats->mvc_abs,
+            stats->MVrv,
+            stats->MVcv,
+            stats->mv_in_out_count,
+            stats->new_mv_count,
+            stats->count,
+            stats->duration);
+    fclose(fpfile);
+  }
+#endif
+}
+
+#if CONFIG_FP_MB_STATS
+static void output_fpmb_stats(uint8_t *this_frame_mb_stats, VP9_COMMON *cm,
+                         struct vpx_codec_pkt_list *pktlist) {
+  struct vpx_codec_cx_pkt pkt;
+  pkt.kind = VPX_CODEC_FPMB_STATS_PKT;
+  pkt.data.firstpass_mb_stats.buf = this_frame_mb_stats;
+  pkt.data.firstpass_mb_stats.sz = cm->initial_mbs * sizeof(uint8_t);
+  vpx_codec_pkt_list_add(pktlist, &pkt);
+}
+#endif
+
+static void zero_stats(FIRSTPASS_STATS *section) {
+  section->frame              = 0.0;
+  section->weight             = 0.0;
+  section->intra_error        = 0.0;
+  section->coded_error        = 0.0;
+  section->sr_coded_error     = 0.0;
+  section->pcnt_inter         = 0.0;
+  section->pcnt_motion        = 0.0;
+  section->pcnt_second_ref    = 0.0;
+  section->pcnt_neutral       = 0.0;
+  section->intra_skip_pct     = 0.0;
+  section->intra_smooth_pct   = 0.0;
+  section->inactive_zone_rows = 0.0;
+  section->inactive_zone_cols = 0.0;
+  section->MVr                = 0.0;
+  section->mvr_abs            = 0.0;
+  section->MVc                = 0.0;
+  section->mvc_abs            = 0.0;
+  section->MVrv               = 0.0;
+  section->MVcv               = 0.0;
+  section->mv_in_out_count    = 0.0;
+  section->new_mv_count       = 0.0;
+  section->count              = 0.0;
+  section->duration           = 1.0;
+  section->spatial_layer_id   = 0;
+}
+
+static void accumulate_stats(FIRSTPASS_STATS *section,
+                             const FIRSTPASS_STATS *frame) {
+  section->frame              += frame->frame;
+  section->weight             += frame->weight;
+  section->spatial_layer_id    = frame->spatial_layer_id;
+  section->intra_error        += frame->intra_error;
+  section->coded_error        += frame->coded_error;
+  section->sr_coded_error     += frame->sr_coded_error;
+  section->pcnt_inter         += frame->pcnt_inter;
+  section->pcnt_motion        += frame->pcnt_motion;
+  section->pcnt_second_ref    += frame->pcnt_second_ref;
+  section->pcnt_neutral       += frame->pcnt_neutral;
+  section->intra_skip_pct     += frame->intra_skip_pct;
+  section->intra_smooth_pct   += frame->intra_smooth_pct;
+  section->inactive_zone_rows += frame->inactive_zone_rows;
+  section->inactive_zone_cols += frame->inactive_zone_cols;
+  section->MVr                += frame->MVr;
+  section->mvr_abs            += frame->mvr_abs;
+  section->MVc                += frame->MVc;
+  section->mvc_abs            += frame->mvc_abs;
+  section->MVrv               += frame->MVrv;
+  section->MVcv               += frame->MVcv;
+  section->mv_in_out_count    += frame->mv_in_out_count;
+  section->new_mv_count       += frame->new_mv_count;
+  section->count              += frame->count;
+  section->duration           += frame->duration;
+}
+
+static void subtract_stats(FIRSTPASS_STATS *section,
+                           const FIRSTPASS_STATS *frame) {
+  section->frame              -= frame->frame;
+  section->weight             -= frame->weight;
+  section->intra_error        -= frame->intra_error;
+  section->coded_error        -= frame->coded_error;
+  section->sr_coded_error     -= frame->sr_coded_error;
+  section->pcnt_inter         -= frame->pcnt_inter;
+  section->pcnt_motion        -= frame->pcnt_motion;
+  section->pcnt_second_ref    -= frame->pcnt_second_ref;
+  section->pcnt_neutral       -= frame->pcnt_neutral;
+  section->intra_skip_pct     -= frame->intra_skip_pct;
+  section->intra_smooth_pct   -= frame->intra_smooth_pct;
+  section->inactive_zone_rows -= frame->inactive_zone_rows;
+  section->inactive_zone_cols -= frame->inactive_zone_cols;
+  section->MVr                -= frame->MVr;
+  section->mvr_abs            -= frame->mvr_abs;
+  section->MVc                -= frame->MVc;
+  section->mvc_abs            -= frame->mvc_abs;
+  section->MVrv               -= frame->MVrv;
+  section->MVcv               -= frame->MVcv;
+  section->mv_in_out_count    -= frame->mv_in_out_count;
+  section->new_mv_count       -= frame->new_mv_count;
+  section->count              -= frame->count;
+  section->duration           -= frame->duration;
+}
+
+// Calculate an active area of the image that discounts formatting
+// bars and partially discounts other 0 energy areas.
+#define MIN_ACTIVE_AREA 0.5
+#define MAX_ACTIVE_AREA 1.0
+static double calculate_active_area(const VP9_COMP *cpi,
+                                    const FIRSTPASS_STATS *this_frame) {
+  double active_pct;
+
+  active_pct = 1.0 -
+    ((this_frame->intra_skip_pct / 2) +
+     ((this_frame->inactive_zone_rows * 2) / (double)cpi->common.mb_rows));
+  return fclamp(active_pct, MIN_ACTIVE_AREA, MAX_ACTIVE_AREA);
+}
+
+// Calculate a modified Error used in distributing bits between easier and
+// harder frames.
+#define ACT_AREA_CORRECTION 0.5
+static double calculate_modified_err(const VP9_COMP *cpi,
+                                     const TWO_PASS *twopass,
+                                     const VP9EncoderConfig *oxcf,
+                                     const FIRSTPASS_STATS *this_frame) {
+  const FIRSTPASS_STATS *const stats = &twopass->total_stats;
+  const double av_weight = stats->weight / stats->count;
+  const double av_err = (stats->coded_error * av_weight) / stats->count;
+  double modified_error =
+    av_err * pow(this_frame->coded_error * this_frame->weight /
+                 DOUBLE_DIVIDE_CHECK(av_err), oxcf->two_pass_vbrbias / 100.0);
+
+  // Correction for active area. Frames with a reduced active area
+  // (eg due to formatting bars) have a higher error per mb for the
+  // remaining active MBs. The correction here assumes that coding
+  // 0.5N blocks of complexity 2X is a little easier than coding N
+  // blocks of complexity X.
+  modified_error *=
+    pow(calculate_active_area(cpi, this_frame), ACT_AREA_CORRECTION);
+
+  return fclamp(modified_error,
+                twopass->modified_error_min, twopass->modified_error_max);
+}
+
+// This function returns the maximum target rate per frame.
+static int frame_max_bits(const RATE_CONTROL *rc,
+                          const VP9EncoderConfig *oxcf) {
+  int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
+                          (int64_t)oxcf->two_pass_vbrmax_section) / 100;
+  if (max_bits < 0)
+    max_bits = 0;
+  else if (max_bits > rc->max_frame_bandwidth)
+    max_bits = rc->max_frame_bandwidth;
+
+  return (int)max_bits;
+}
+
+void vp9_init_first_pass(VP9_COMP *cpi) {
+  zero_stats(&cpi->twopass.total_stats);
+}
+
+void vp9_end_first_pass(VP9_COMP *cpi) {
+  if (is_two_pass_svc(cpi)) {
+    int i;
+    for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
+      output_stats(&cpi->svc.layer_context[i].twopass.total_stats,
+                   cpi->output_pkt_list);
+    }
+  } else {
+    output_stats(&cpi->twopass.total_stats, cpi->output_pkt_list);
+  }
+}
+
+static vpx_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
+  switch (bsize) {
+    case BLOCK_8X8:
+      return vpx_mse8x8;
+    case BLOCK_16X8:
+      return vpx_mse16x8;
+    case BLOCK_8X16:
+      return vpx_mse8x16;
+    default:
+      return vpx_mse16x16;
+  }
+}
+
+static unsigned int get_prediction_error(BLOCK_SIZE bsize,
+                                         const struct buf_2d *src,
+                                         const struct buf_2d *ref) {
+  unsigned int sse;
+  const vpx_variance_fn_t fn = get_block_variance_fn(bsize);
+  fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
+  return sse;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static vpx_variance_fn_t highbd_get_block_variance_fn(BLOCK_SIZE bsize,
+                                                      int bd) {
+  switch (bd) {
+    default:
+      switch (bsize) {
+        case BLOCK_8X8:
+          return vpx_highbd_8_mse8x8;
+        case BLOCK_16X8:
+          return vpx_highbd_8_mse16x8;
+        case BLOCK_8X16:
+          return vpx_highbd_8_mse8x16;
+        default:
+          return vpx_highbd_8_mse16x16;
+      }
+      break;
+    case 10:
+      switch (bsize) {
+        case BLOCK_8X8:
+          return vpx_highbd_10_mse8x8;
+        case BLOCK_16X8:
+          return vpx_highbd_10_mse16x8;
+        case BLOCK_8X16:
+          return vpx_highbd_10_mse8x16;
+        default:
+          return vpx_highbd_10_mse16x16;
+      }
+      break;
+    case 12:
+      switch (bsize) {
+        case BLOCK_8X8:
+          return vpx_highbd_12_mse8x8;
+        case BLOCK_16X8:
+          return vpx_highbd_12_mse16x8;
+        case BLOCK_8X16:
+          return vpx_highbd_12_mse8x16;
+        default:
+          return vpx_highbd_12_mse16x16;
+      }
+      break;
+  }
+}
+
+static unsigned int highbd_get_prediction_error(BLOCK_SIZE bsize,
+                                                const struct buf_2d *src,
+                                                const struct buf_2d *ref,
+                                                int bd) {
+  unsigned int sse;
+  const vpx_variance_fn_t fn = highbd_get_block_variance_fn(bsize, bd);
+  fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
+  return sse;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+// Refine the motion search range according to the frame dimension
+// for first pass test.
+static int get_search_range(const VP9_COMP *cpi) {
+  int sr = 0;
+  const int dim = VPXMIN(cpi->initial_width, cpi->initial_height);
+
+  while ((dim << sr) < MAX_FULL_PEL_VAL)
+    ++sr;
+  return sr;
+}
+
+static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
+                                     const MV *ref_mv, MV *best_mv,
+                                     int *best_motion_err) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MV tmp_mv = {0, 0};
+  MV ref_mv_full = {ref_mv->row >> 3, ref_mv->col >> 3};
+  int num00, tmp_err, n;
+  const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
+  vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
+  const int new_mv_mode_penalty = NEW_MV_MODE_PENALTY;
+
+  int step_param = 3;
+  int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
+  const int sr = get_search_range(cpi);
+  step_param += sr;
+  further_steps -= sr;
+
+  // Override the default variance function to use MSE.
+  v_fn_ptr.vf = get_block_variance_fn(bsize);
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    v_fn_ptr.vf = highbd_get_block_variance_fn(bsize, xd->bd);
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  // Center the initial step/diamond search on best mv.
+  tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
+                                    step_param,
+                                    x->sadperbit16, &num00, &v_fn_ptr, ref_mv);
+  if (tmp_err < INT_MAX)
+    tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
+  if (tmp_err < INT_MAX - new_mv_mode_penalty)
+    tmp_err += new_mv_mode_penalty;
+
+  if (tmp_err < *best_motion_err) {
+    *best_motion_err = tmp_err;
+    *best_mv = tmp_mv;
+  }
+
+  // Carry out further step/diamond searches as necessary.
+  n = num00;
+  num00 = 0;
+
+  while (n < further_steps) {
+    ++n;
+
+    if (num00) {
+      --num00;
+    } else {
+      tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
+                                        step_param + n, x->sadperbit16,
+                                        &num00, &v_fn_ptr, ref_mv);
+      if (tmp_err < INT_MAX)
+        tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
+      if (tmp_err < INT_MAX - new_mv_mode_penalty)
+        tmp_err += new_mv_mode_penalty;
+
+      if (tmp_err < *best_motion_err) {
+        *best_motion_err = tmp_err;
+        *best_mv = tmp_mv;
+      }
+    }
+  }
+}
+
+static BLOCK_SIZE get_bsize(const VP9_COMMON *cm, int mb_row, int mb_col) {
+  if (2 * mb_col + 1 < cm->mi_cols) {
+    return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_16X16
+                                        : BLOCK_16X8;
+  } else {
+    return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_8X16
+                                        : BLOCK_8X8;
+  }
+}
+
+static int find_fp_qindex(vpx_bit_depth_t bit_depth) {
+  int i;
+
+  for (i = 0; i < QINDEX_RANGE; ++i)
+    if (vp9_convert_qindex_to_q(i, bit_depth) >= FIRST_PASS_Q)
+      break;
+
+  if (i == QINDEX_RANGE)
+    i--;
+
+  return i;
+}
+
+static void set_first_pass_params(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  if (!cpi->refresh_alt_ref_frame &&
+      (cm->current_video_frame == 0 ||
+       (cpi->frame_flags & FRAMEFLAGS_KEY))) {
+    cm->frame_type = KEY_FRAME;
+  } else {
+    cm->frame_type = INTER_FRAME;
+  }
+  // Do not use periodic key frames.
+  cpi->rc.frames_to_key = INT_MAX;
+}
+
+// This threshold is used to track blocks where to all intents and purposes
+// the intra prediction error 0. Though the metric we test against
+// is technically a sse we are mainly interested in blocks where all the pixels
+// in the 8 bit domain have an error of <= 1 (where error = sse) so a
+// linear scaling for 10 and 12 bit gives similar results.
+#define UL_INTRA_THRESH 50
+static int get_ul_intra_threshold(VP9_COMMON *cm) {
+  int ret_val = UL_INTRA_THRESH;
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    switch (cm->bit_depth) {
+      case VPX_BITS_8:
+        ret_val = UL_INTRA_THRESH;
+        break;
+      case VPX_BITS_10:
+        ret_val = UL_INTRA_THRESH >> 2;
+        break;
+      case VPX_BITS_12:
+        ret_val = UL_INTRA_THRESH >> 4;
+        break;
+      default:
+        assert(0 && "cm->bit_depth should be VPX_BITS_8, "
+                    "VPX_BITS_10 or VPX_BITS_12");
+    }
+  }
+#else
+  (void) cm;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  return ret_val;
+}
+
+#define SMOOTH_INTRA_THRESH 4000
+static int get_smooth_intra_threshold(VP9_COMMON *cm) {
+  int ret_val = SMOOTH_INTRA_THRESH;
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    switch (cm->bit_depth) {
+      case VPX_BITS_8:
+        ret_val = SMOOTH_INTRA_THRESH;
+        break;
+      case VPX_BITS_10:
+        ret_val = SMOOTH_INTRA_THRESH >> 2;
+        break;
+      case VPX_BITS_12:
+        ret_val = SMOOTH_INTRA_THRESH >> 4;
+        break;
+      default:
+        assert(0 && "cm->bit_depth should be VPX_BITS_8, "
+                    "VPX_BITS_10 or VPX_BITS_12");
+    }
+  }
+#else
+  (void) cm;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  return ret_val;
+}
+
+#define INVALID_ROW -1
+void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
+  int mb_row, mb_col;
+  MACROBLOCK *const x = &cpi->td.mb;
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  TileInfo tile;
+  struct macroblock_plane *const p = x->plane;
+  struct macroblockd_plane *const pd = xd->plane;
+  const PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
+  int i;
+
+  int recon_yoffset, recon_uvoffset;
+  int64_t intra_error = 0;
+  int64_t coded_error = 0;
+  int64_t sr_coded_error = 0;
+
+  int sum_mvr = 0, sum_mvc = 0;
+  int sum_mvr_abs = 0, sum_mvc_abs = 0;
+  int64_t sum_mvrs = 0, sum_mvcs = 0;
+  int mvcount = 0;
+  int intercount = 0;
+  int second_ref_count = 0;
+  const int intrapenalty = INTRA_MODE_PENALTY;
+  double neutral_count;
+  int intra_skip_count = 0;
+  int intra_smooth_count = 0;
+  int image_data_start_row = INVALID_ROW;
+  int new_mv_count = 0;
+  int sum_in_vectors = 0;
+  MV lastmv = {0, 0};
+  TWO_PASS *twopass = &cpi->twopass;
+  const MV zero_mv = {0, 0};
+  int recon_y_stride, recon_uv_stride, uv_mb_height;
+
+  YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
+  YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
+  YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
+  const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
+
+  LAYER_CONTEXT *const lc = is_two_pass_svc(cpi) ?
+        &cpi->svc.layer_context[cpi->svc.spatial_layer_id] : NULL;
+  double intra_factor;
+  double brightness_factor;
+  BufferPool *const pool = cm->buffer_pool;
+  MODE_INFO mi_above, mi_left;
+
+  // First pass code requires valid last and new frame buffers.
+  assert(new_yv12 != NULL);
+  assert((lc != NULL) || frame_is_intra_only(cm) || (lst_yv12 != NULL));
+
+#if CONFIG_FP_MB_STATS
+  if (cpi->use_fp_mb_stats) {
+    vp9_zero_array(cpi->twopass.frame_mb_stats_buf, cm->initial_mbs);
+  }
+#endif
+
+  vpx_clear_system_state();
+
+  intra_factor = 0.0;
+  brightness_factor = 0.0;
+  neutral_count = 0.0;
+
+  set_first_pass_params(cpi);
+  vp9_set_quantizer(cm, find_fp_qindex(cm->bit_depth));
+
+  if (lc != NULL) {
+    twopass = &lc->twopass;
+
+    cpi->lst_fb_idx = cpi->svc.spatial_layer_id;
+    cpi->ref_frame_flags = VP9_LAST_FLAG;
+
+    if (cpi->svc.number_spatial_layers + cpi->svc.spatial_layer_id <
+        REF_FRAMES) {
+      cpi->gld_fb_idx =
+          cpi->svc.number_spatial_layers + cpi->svc.spatial_layer_id;
+      cpi->ref_frame_flags |= VP9_GOLD_FLAG;
+      cpi->refresh_golden_frame = (lc->current_video_frame_in_layer == 0);
+    } else {
+      cpi->refresh_golden_frame = 0;
+    }
+
+    if (lc->current_video_frame_in_layer == 0)
+      cpi->ref_frame_flags = 0;
+
+    vp9_scale_references(cpi);
+
+    // Use either last frame or alt frame for motion search.
+    if (cpi->ref_frame_flags & VP9_LAST_FLAG) {
+      first_ref_buf = vp9_get_scaled_ref_frame(cpi, LAST_FRAME);
+      if (first_ref_buf == NULL)
+        first_ref_buf = get_ref_frame_buffer(cpi, LAST_FRAME);
+    }
+
+    if (cpi->ref_frame_flags & VP9_GOLD_FLAG) {
+      gld_yv12 = vp9_get_scaled_ref_frame(cpi, GOLDEN_FRAME);
+      if (gld_yv12 == NULL) {
+        gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
+      }
+    } else {
+      gld_yv12 = NULL;
+    }
+
+    set_ref_ptrs(cm, xd,
+                 (cpi->ref_frame_flags & VP9_LAST_FLAG) ? LAST_FRAME: NONE,
+                 (cpi->ref_frame_flags & VP9_GOLD_FLAG) ? GOLDEN_FRAME : NONE);
+
+    cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
+                                        &cpi->scaled_source, 0);
+  }
+
+  vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
+
+  vp9_setup_src_planes(x, cpi->Source, 0, 0);
+  vp9_setup_dst_planes(xd->plane, new_yv12, 0, 0);
+
+  if (!frame_is_intra_only(cm)) {
+    vp9_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
+  }
+
+  xd->mi = cm->mi_grid_visible;
+  xd->mi[0] = cm->mi;
+
+  vp9_frame_init_quantizer(cpi);
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    p[i].coeff = ctx->coeff_pbuf[i][1];
+    p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
+    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
+    p[i].eobs = ctx->eobs_pbuf[i][1];
+  }
+  x->skip_recode = 0;
+
+  vp9_init_mv_probs(cm);
+  vp9_initialize_rd_consts(cpi);
+
+  // Tiling is ignored in the first pass.
+  vp9_tile_init(&tile, cm, 0, 0);
+
+  recon_y_stride = new_yv12->y_stride;
+  recon_uv_stride = new_yv12->uv_stride;
+  uv_mb_height = 16 >> (new_yv12->y_height > new_yv12->uv_height);
+
+  for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
+    MV best_ref_mv = {0, 0};
+
+    // Reset above block coeffs.
+    recon_yoffset = (mb_row * recon_y_stride * 16);
+    recon_uvoffset = (mb_row * recon_uv_stride * uv_mb_height);
+
+    // Set up limit values for motion vectors to prevent them extending
+    // outside the UMV borders.
+    x->mv_row_min = -((mb_row * 16) + BORDER_MV_PIXELS_B16);
+    x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
+                    + BORDER_MV_PIXELS_B16;
+
+    for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
+      int this_error;
+      const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
+      const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);
+      double log_intra;
+      int level_sample;
+
+#if CONFIG_FP_MB_STATS
+      const int mb_index = mb_row * cm->mb_cols + mb_col;
+#endif
+
+      vpx_clear_system_state();
+
+      xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
+      xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
+      xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
+      xd->mi[0]->sb_type = bsize;
+      xd->mi[0]->ref_frame[0] = INTRA_FRAME;
+      set_mi_row_col(xd, &tile,
+                     mb_row << 1, num_8x8_blocks_high_lookup[bsize],
+                     mb_col << 1, num_8x8_blocks_wide_lookup[bsize],
+                     cm->mi_rows, cm->mi_cols);
+      // Are edges available for intra prediction?
+      // Since the firstpass does not populate the mi_grid_visible,
+      // above_mi/left_mi must be overwritten with a nonzero value when edges
+      // are available.  Required by vp9_predict_intra_block().
+      xd->above_mi = (mb_row != 0) ? &mi_above : NULL;
+      xd->left_mi  = (mb_col > tile.mi_col_start) ? &mi_left : NULL;
+
+      // Do intra 16x16 prediction.
+      x->skip_encode = 0;
+      xd->mi[0]->mode = DC_PRED;
+      xd->mi[0]->tx_size = use_dc_pred ?
+         (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
+      vp9_encode_intra_block_plane(x, bsize, 0, 0);
+      this_error = vpx_get_mb_ss(x->plane[0].src_diff);
+
+      // Keep a record of blocks that have almost no intra error residual
+      // (i.e. are in effect completely flat and untextured in the intra
+      // domain). In natural videos this is uncommon, but it is much more
+      // common in animations, graphics and screen content, so may be used
+      // as a signal to detect these types of content.
+      if (this_error < get_ul_intra_threshold(cm)) {
+        ++intra_skip_count;
+      } else if ((mb_col > 0) && (image_data_start_row == INVALID_ROW)) {
+        image_data_start_row = mb_row;
+      }
+      if (this_error < get_smooth_intra_threshold(cm)) {
+        ++intra_smooth_count;
+      }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (cm->use_highbitdepth) {
+        switch (cm->bit_depth) {
+          case VPX_BITS_8:
+            break;
+          case VPX_BITS_10:
+            this_error >>= 4;
+            break;
+          case VPX_BITS_12:
+            this_error >>= 8;
+            break;
+          default:
+            assert(0 && "cm->bit_depth should be VPX_BITS_8, "
+                        "VPX_BITS_10 or VPX_BITS_12");
+            return;
+        }
+      }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+      vpx_clear_system_state();
+      log_intra = log(this_error + 1.0);
+      if (log_intra < 10.0)
+        intra_factor += 1.0 + ((10.0 - log_intra) * 0.05);
+      else
+        intra_factor += 1.0;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (cm->use_highbitdepth)
+        level_sample = CONVERT_TO_SHORTPTR(x->plane[0].src.buf)[0];
+      else
+        level_sample = x->plane[0].src.buf[0];
+#else
+      level_sample = x->plane[0].src.buf[0];
+#endif
+      if ((level_sample < DARK_THRESH) && (log_intra < 9.0))
+        brightness_factor += 1.0 + (0.01 * (DARK_THRESH - level_sample));
+      else
+        brightness_factor += 1.0;
+
+      // Intrapenalty below deals with situations where the intra and inter
+      // error scores are very low (e.g. a plain black frame).
+      // We do not have special cases in first pass for 0,0 and nearest etc so
+      // all inter modes carry an overhead cost estimate for the mv.
+      // When the error score is very low this causes us to pick all or lots of
+      // INTRA modes and throw lots of key frames.
+      // This penalty adds a cost matching that of a 0,0 mv to the intra case.
+      this_error += intrapenalty;
+
+      // Accumulate the intra error.
+      intra_error += (int64_t)this_error;
+
+#if CONFIG_FP_MB_STATS
+      if (cpi->use_fp_mb_stats) {
+        // initialization
+        cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
+      }
+#endif
+
+      // Set up limit values for motion vectors to prevent them extending
+      // outside the UMV borders.
+      x->mv_col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
+      x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;
+
+      // Other than for the first frame do a motion search.
+      if ((lc == NULL && cm->current_video_frame > 0) ||
+          (lc != NULL && lc->current_video_frame_in_layer > 0)) {
+        int tmp_err, motion_error, raw_motion_error;
+        // Assume 0,0 motion with no mv overhead.
+        MV mv = {0, 0} , tmp_mv = {0, 0};
+        struct buf_2d unscaled_last_source_buf_2d;
+
+        xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
+#if CONFIG_VP9_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          motion_error = highbd_get_prediction_error(
+              bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
+        } else {
+          motion_error = get_prediction_error(
+              bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
+        }
+#else
+        motion_error = get_prediction_error(
+            bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+        // Compute the motion error of the 0,0 motion using the last source
+        // frame as the reference. Skip the further motion search on
+        // reconstructed frame if this error is small.
+        unscaled_last_source_buf_2d.buf =
+            cpi->unscaled_last_source->y_buffer + recon_yoffset;
+        unscaled_last_source_buf_2d.stride =
+            cpi->unscaled_last_source->y_stride;
+#if CONFIG_VP9_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          raw_motion_error = highbd_get_prediction_error(
+              bsize, &x->plane[0].src, &unscaled_last_source_buf_2d, xd->bd);
+        } else {
+          raw_motion_error = get_prediction_error(
+              bsize, &x->plane[0].src, &unscaled_last_source_buf_2d);
+        }
+#else
+        raw_motion_error = get_prediction_error(
+            bsize, &x->plane[0].src, &unscaled_last_source_buf_2d);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+        // TODO(pengchong): Replace the hard-coded threshold
+        if (raw_motion_error > 25 || lc != NULL) {
+          // Test last reference frame using the previous best mv as the
+          // starting point (best reference) for the search.
+          first_pass_motion_search(cpi, x, &best_ref_mv, &mv, &motion_error);
+
+          // If the current best reference mv is not centered on 0,0 then do a
+          // 0,0 based search as well.
+          if (!is_zero_mv(&best_ref_mv)) {
+            tmp_err = INT_MAX;
+            first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv, &tmp_err);
+
+            if (tmp_err < motion_error) {
+              motion_error = tmp_err;
+              mv = tmp_mv;
+            }
+          }
+
+          // Search in an older reference frame.
+          if (((lc == NULL && cm->current_video_frame > 1) ||
+               (lc != NULL && lc->current_video_frame_in_layer > 1))
+              && gld_yv12 != NULL) {
+            // Assume 0,0 motion with no mv overhead.
+            int gf_motion_error;
+
+            xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
+#if CONFIG_VP9_HIGHBITDEPTH
+            if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+              gf_motion_error = highbd_get_prediction_error(
+                  bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
+            } else {
+              gf_motion_error = get_prediction_error(
+                  bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
+            }
+#else
+            gf_motion_error = get_prediction_error(
+                bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+            first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv,
+                                     &gf_motion_error);
+
+            if (gf_motion_error < motion_error && gf_motion_error < this_error)
+              ++second_ref_count;
+
+            // Reset to last frame as reference buffer.
+            xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
+            xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
+            xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
+
+            // In accumulating a score for the older reference frame take the
+            // best of the motion predicted score and the intra coded error
+            // (just as will be done for) accumulation of "coded_error" for
+            // the last frame.
+            if (gf_motion_error < this_error)
+              sr_coded_error += gf_motion_error;
+            else
+              sr_coded_error += this_error;
+          } else {
+            sr_coded_error += motion_error;
+          }
+        } else {
+          sr_coded_error += motion_error;
+        }
+
+        // Start by assuming that intra mode is best.
+        best_ref_mv.row = 0;
+        best_ref_mv.col = 0;
+
+#if CONFIG_FP_MB_STATS
+        if (cpi->use_fp_mb_stats) {
+          // intra predication statistics
+          cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
+          cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_DCINTRA_MASK;
+          cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
+          if (this_error > FPMB_ERROR_LARGE_TH) {
+            cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LARGE_MASK;
+          } else if (this_error < FPMB_ERROR_SMALL_TH) {
+            cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_SMALL_MASK;
+          }
+        }
+#endif
+
+        if (motion_error <= this_error) {
+          vpx_clear_system_state();
+
+          // Keep a count of cases where the inter and intra were very close
+          // and very low. This helps with scene cut detection for example in
+          // cropped clips with black bars at the sides or top and bottom.
+          if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
+              (this_error < (2 * intrapenalty))) {
+            neutral_count += 1.0;
+          // Also track cases where the intra is not much worse than the inter
+          // and use this in limiting the GF/arf group length.
+          } else if ((this_error > NCOUNT_INTRA_THRESH) &&
+                     (this_error < (NCOUNT_INTRA_FACTOR * motion_error))) {
+            neutral_count += (double)motion_error /
+                             DOUBLE_DIVIDE_CHECK((double)this_error);
+          }
+
+          mv.row *= 8;
+          mv.col *= 8;
+          this_error = motion_error;
+          xd->mi[0]->mode = NEWMV;
+          xd->mi[0]->mv[0].as_mv = mv;
+          xd->mi[0]->tx_size = TX_4X4;
+          xd->mi[0]->ref_frame[0] = LAST_FRAME;
+          xd->mi[0]->ref_frame[1] = NONE;
+          vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, bsize);
+          vp9_encode_sby_pass1(x, bsize);
+          sum_mvr += mv.row;
+          sum_mvr_abs += abs(mv.row);
+          sum_mvc += mv.col;
+          sum_mvc_abs += abs(mv.col);
+          sum_mvrs += mv.row * mv.row;
+          sum_mvcs += mv.col * mv.col;
+          ++intercount;
+
+          best_ref_mv = mv;
+
+#if CONFIG_FP_MB_STATS
+          if (cpi->use_fp_mb_stats) {
+            // inter predication statistics
+            cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
+            cpi->twopass.frame_mb_stats_buf[mb_index] &= ~FPMB_DCINTRA_MASK;
+            cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
+            if (this_error > FPMB_ERROR_LARGE_TH) {
+              cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                  FPMB_ERROR_LARGE_MASK;
+            } else if (this_error < FPMB_ERROR_SMALL_TH) {
+              cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                  FPMB_ERROR_SMALL_MASK;
+            }
+          }
+#endif
+
+          if (!is_zero_mv(&mv)) {
+            ++mvcount;
+
+#if CONFIG_FP_MB_STATS
+            if (cpi->use_fp_mb_stats) {
+              cpi->twopass.frame_mb_stats_buf[mb_index] &=
+                  ~FPMB_MOTION_ZERO_MASK;
+              // check estimated motion direction
+              if (mv.as_mv.col > 0 && mv.as_mv.col >= abs(mv.as_mv.row)) {
+                // right direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_RIGHT_MASK;
+              } else if (mv.as_mv.row < 0 &&
+                         abs(mv.as_mv.row) >= abs(mv.as_mv.col)) {
+                // up direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_UP_MASK;
+              } else if (mv.as_mv.col < 0 &&
+                         abs(mv.as_mv.col) >= abs(mv.as_mv.row)) {
+                // left direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_LEFT_MASK;
+              } else {
+                // down direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_DOWN_MASK;
+              }
+            }
+#endif
+
+            // Non-zero vector, was it different from the last non zero vector?
+            if (!is_equal_mv(&mv, &lastmv))
+              ++new_mv_count;
+            lastmv = mv;
+
+            // Does the row vector point inwards or outwards?
+            if (mb_row < cm->mb_rows / 2) {
+              if (mv.row > 0)
+                --sum_in_vectors;
+              else if (mv.row < 0)
+                ++sum_in_vectors;
+            } else if (mb_row > cm->mb_rows / 2) {
+              if (mv.row > 0)
+                ++sum_in_vectors;
+              else if (mv.row < 0)
+                --sum_in_vectors;
+            }
+
+            // Does the col vector point inwards or outwards?
+            if (mb_col < cm->mb_cols / 2) {
+              if (mv.col > 0)
+                --sum_in_vectors;
+              else if (mv.col < 0)
+                ++sum_in_vectors;
+            } else if (mb_col > cm->mb_cols / 2) {
+              if (mv.col > 0)
+                ++sum_in_vectors;
+              else if (mv.col < 0)
+                --sum_in_vectors;
+            }
+          }
+        }
+      } else {
+        sr_coded_error += (int64_t)this_error;
+      }
+      coded_error += (int64_t)this_error;
+
+      // Adjust to the next column of MBs.
+      x->plane[0].src.buf += 16;
+      x->plane[1].src.buf += uv_mb_height;
+      x->plane[2].src.buf += uv_mb_height;
+
+      recon_yoffset += 16;
+      recon_uvoffset += uv_mb_height;
+    }
+
+    // Adjust to the next row of MBs.
+    x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
+    x->plane[1].src.buf += uv_mb_height * x->plane[1].src.stride -
+                           uv_mb_height * cm->mb_cols;
+    x->plane[2].src.buf += uv_mb_height * x->plane[1].src.stride -
+                           uv_mb_height * cm->mb_cols;
+
+    vpx_clear_system_state();
+  }
+
+  // Clamp the image start to rows/2. This number of rows is discarded top
+  // and bottom as dead data so rows / 2 means the frame is blank.
+  if ((image_data_start_row > cm->mb_rows / 2) ||
+      (image_data_start_row == INVALID_ROW)) {
+    image_data_start_row = cm->mb_rows / 2;
+  }
+  // Exclude any image dead zone
+  if (image_data_start_row > 0) {
+    intra_skip_count =
+        VPXMAX(0, intra_skip_count - (image_data_start_row * cm->mb_cols * 2));
+  }
+
+  {
+    FIRSTPASS_STATS fps;
+    // The minimum error here insures some bit allocation to frames even
+    // in static regions. The allocation per MB declines for larger formats
+    // where the typical "real" energy per MB also falls.
+    // Initial estimate here uses sqrt(mbs) to define the min_err, where the
+    // number of mbs is proportional to the image area.
+    const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
+                        ? cpi->initial_mbs : cpi->common.MBs;
+    const double min_err = 200 * sqrt(num_mbs);
+
+    intra_factor = intra_factor / (double)num_mbs;
+    brightness_factor = brightness_factor / (double)num_mbs;
+    fps.weight = intra_factor * brightness_factor;
+
+    fps.frame = cm->current_video_frame;
+    fps.spatial_layer_id = cpi->svc.spatial_layer_id;
+    fps.coded_error = (double)(coded_error >> 8) + min_err;
+    fps.sr_coded_error = (double)(sr_coded_error >> 8) + min_err;
+    fps.intra_error = (double)(intra_error >> 8) + min_err;
+    fps.count = 1.0;
+    fps.pcnt_inter = (double)intercount / num_mbs;
+    fps.pcnt_second_ref = (double)second_ref_count / num_mbs;
+    fps.pcnt_neutral = (double)neutral_count / num_mbs;
+    fps.intra_skip_pct = (double)intra_skip_count / num_mbs;
+    fps.intra_smooth_pct = (double)intra_smooth_count / num_mbs;
+    fps.inactive_zone_rows = (double)image_data_start_row;
+    // Currently set to 0 as most issues relate to letter boxing.
+    fps.inactive_zone_cols = (double)0;
+
+    if (mvcount > 0) {
+      fps.MVr = (double)sum_mvr / mvcount;
+      fps.mvr_abs = (double)sum_mvr_abs / mvcount;
+      fps.MVc = (double)sum_mvc / mvcount;
+      fps.mvc_abs = (double)sum_mvc_abs / mvcount;
+      fps.MVrv = ((double)sum_mvrs -
+                  ((double)sum_mvr * sum_mvr / mvcount)) / mvcount;
+      fps.MVcv = ((double)sum_mvcs -
+                  ((double)sum_mvc * sum_mvc / mvcount)) / mvcount;
+      fps.mv_in_out_count = (double)sum_in_vectors / (mvcount * 2);
+      fps.new_mv_count = new_mv_count;
+      fps.pcnt_motion = (double)mvcount / num_mbs;
+    } else {
+      fps.MVr = 0.0;
+      fps.mvr_abs = 0.0;
+      fps.MVc = 0.0;
+      fps.mvc_abs = 0.0;
+      fps.MVrv = 0.0;
+      fps.MVcv = 0.0;
+      fps.mv_in_out_count = 0.0;
+      fps.new_mv_count = 0.0;
+      fps.pcnt_motion = 0.0;
+    }
+
+    // Dont allow a value of 0 for duration.
+    // (Section duration is also defaulted to minimum of 1.0).
+    fps.duration = VPXMAX(1.0, (double)(source->ts_end - source->ts_start));
+
+    // Don't want to do output stats with a stack variable!
+    twopass->this_frame_stats = fps;
+    output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
+    accumulate_stats(&twopass->total_stats, &fps);
+
+#if CONFIG_FP_MB_STATS
+    if (cpi->use_fp_mb_stats) {
+      output_fpmb_stats(twopass->frame_mb_stats_buf, cm, cpi->output_pkt_list);
+    }
+#endif
+  }
+
+  // Copy the previous Last Frame back into gf and and arf buffers if
+  // the prediction is good enough... but also don't allow it to lag too far.
+  if ((twopass->sr_update_lag > 3) ||
+      ((cm->current_video_frame > 0) &&
+       (twopass->this_frame_stats.pcnt_inter > 0.20) &&
+       ((twopass->this_frame_stats.intra_error /
+         DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
+    if (gld_yv12 != NULL) {
+      ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
+                 cm->ref_frame_map[cpi->lst_fb_idx]);
+    }
+    twopass->sr_update_lag = 1;
+  } else {
+    ++twopass->sr_update_lag;
+  }
+
+  vpx_extend_frame_borders(new_yv12);
+
+  if (lc != NULL) {
+    vp9_update_reference_frames(cpi);
+  } else {
+    // The frame we just compressed now becomes the last frame.
+    ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->lst_fb_idx],
+               cm->new_fb_idx);
+  }
+
+  // Special case for the first frame. Copy into the GF buffer as a second
+  // reference.
+  if (cm->current_video_frame == 0 && cpi->gld_fb_idx != INVALID_IDX &&
+      lc == NULL) {
+    ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
+               cm->ref_frame_map[cpi->lst_fb_idx]);
+  }
+
+  // Use this to see what the first pass reconstruction looks like.
+  if (0) {
+    char filename[512];
+    FILE *recon_file;
+    snprintf(filename, sizeof(filename), "enc%04d.yuv",
+             (int)cm->current_video_frame);
+
+    if (cm->current_video_frame == 0)
+      recon_file = fopen(filename, "wb");
+    else
+      recon_file = fopen(filename, "ab");
+
+    (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
+    fclose(recon_file);
+  }
+
+  ++cm->current_video_frame;
+  if (cpi->use_svc)
+    vp9_inc_frame_in_layer(cpi);
+}
+
+static double calc_correction_factor(double err_per_mb,
+                                     double err_divisor,
+                                     double pt_low,
+                                     double pt_high,
+                                     int q,
+                                     vpx_bit_depth_t bit_depth) {
+  const double error_term = err_per_mb / err_divisor;
+
+  // Adjustment based on actual quantizer to power term.
+  const double power_term =
+      VPXMIN(vp9_convert_qindex_to_q(q, bit_depth) * 0.01 + pt_low, pt_high);
+
+  // Calculate correction factor.
+  if (power_term < 1.0)
+    assert(error_term >= 0.0);
+
+  return fclamp(pow(error_term, power_term), 0.05, 5.0);
+}
+
+#define ERR_DIVISOR         115.0
+static int get_twopass_worst_quality(VP9_COMP *cpi,
+                                     const double section_err,
+                                     double inactive_zone,
+                                     int section_target_bandwidth) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  TWO_PASS *const twopass = &cpi->twopass;
+
+  // Clamp the target rate to VBR min / max limts.
+  const int target_rate =
+      vp9_rc_clamp_pframe_target_size(cpi, section_target_bandwidth);
+
+  inactive_zone = fclamp(inactive_zone, 0.0, 1.0);
+
+  if (target_rate <= 0) {
+    return rc->worst_quality;  // Highest value allowed
+  } else {
+    const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
+                        ? cpi->initial_mbs : cpi->common.MBs;
+    const int active_mbs = VPXMAX(1, num_mbs - (int)(num_mbs * inactive_zone));
+    const double av_err_per_mb = section_err / active_mbs;
+    const double speed_term = 1.0 + 0.04 * oxcf->speed;
+    double last_group_rate_err;
+    const int target_norm_bits_per_mb = ((uint64_t)target_rate <<
+                                         BPER_MB_NORMBITS) / active_mbs;
+    int q;
+    int is_svc_upper_layer = 0;
+
+    if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0)
+      is_svc_upper_layer = 1;
+
+    // based on recent history adjust expectations of bits per macroblock.
+    last_group_rate_err = (double)twopass->rolling_arf_group_actual_bits /
+        DOUBLE_DIVIDE_CHECK((double)twopass->rolling_arf_group_target_bits);
+    last_group_rate_err =
+        VPXMAX(0.25, VPXMIN(4.0, last_group_rate_err));
+    twopass->bpm_factor *= (3.0 + last_group_rate_err) / 4.0;
+    twopass->bpm_factor =
+        VPXMAX(0.25, VPXMIN(4.0, twopass->bpm_factor));
+
+    // Try and pick a max Q that will be high enough to encode the
+    // content at the given rate.
+    for (q = rc->best_quality; q < rc->worst_quality; ++q) {
+      const double factor =
+          calc_correction_factor(av_err_per_mb,
+                                 ERR_DIVISOR,
+                                 is_svc_upper_layer ? SVC_FACTOR_PT_LOW :
+                                 FACTOR_PT_LOW, FACTOR_PT_HIGH, q,
+                                 cpi->common.bit_depth);
+      const int bits_per_mb =
+        vp9_rc_bits_per_mb(INTER_FRAME, q,
+                           factor * speed_term * cpi->twopass.bpm_factor,
+                           cpi->common.bit_depth);
+      if (bits_per_mb <= target_norm_bits_per_mb)
+        break;
+    }
+
+    // Restriction on active max q for constrained quality mode.
+    if (cpi->oxcf.rc_mode == VPX_CQ)
+      q = VPXMAX(q, oxcf->cq_level);
+    return q;
+  }
+}
+
+static void setup_rf_level_maxq(VP9_COMP *cpi) {
+  int i;
+  RATE_CONTROL *const rc = &cpi->rc;
+  for (i = INTER_NORMAL; i < RATE_FACTOR_LEVELS; ++i) {
+    int qdelta = vp9_frame_type_qdelta(cpi, i, rc->worst_quality);
+    rc->rf_level_maxq[i] = VPXMAX(rc->worst_quality + qdelta, rc->best_quality);
+  }
+}
+
+static void init_subsampling(VP9_COMP *cpi) {
+  const VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  const int w = cm->width;
+  const int h = cm->height;
+  int i;
+
+  for (i = 0; i < FRAME_SCALE_STEPS; ++i) {
+    // Note: Frames with odd-sized dimensions may result from this scaling.
+    rc->frame_width[i] = (w * 16) / frame_scale_factor[i];
+    rc->frame_height[i] = (h * 16) / frame_scale_factor[i];
+  }
+
+  setup_rf_level_maxq(cpi);
+}
+
+void calculate_coded_size(VP9_COMP *cpi,
+                          int *scaled_frame_width,
+                          int *scaled_frame_height) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  *scaled_frame_width = rc->frame_width[rc->frame_size_selector];
+  *scaled_frame_height = rc->frame_height[rc->frame_size_selector];
+}
+
+void vp9_init_second_pass(VP9_COMP *cpi) {
+  SVC *const svc = &cpi->svc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const int is_two_pass_svc = (svc->number_spatial_layers > 1) ||
+                              (svc->number_temporal_layers > 1);
+  RATE_CONTROL *const rc = &cpi->rc;
+  TWO_PASS *const twopass = is_two_pass_svc ?
+      &svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
+  double frame_rate;
+  FIRSTPASS_STATS *stats;
+
+  zero_stats(&twopass->total_stats);
+  zero_stats(&twopass->total_left_stats);
+
+  if (!twopass->stats_in_end)
+    return;
+
+  stats = &twopass->total_stats;
+
+  *stats = *twopass->stats_in_end;
+  twopass->total_left_stats = *stats;
+
+  frame_rate = 10000000.0 * stats->count / stats->duration;
+  // Each frame can have a different duration, as the frame rate in the source
+  // isn't guaranteed to be constant. The frame rate prior to the first frame
+  // encoded in the second pass is a guess. However, the sum duration is not.
+  // It is calculated based on the actual durations of all frames from the
+  // first pass.
+
+  if (is_two_pass_svc) {
+    vp9_update_spatial_layer_framerate(cpi, frame_rate);
+    twopass->bits_left = (int64_t)(stats->duration *
+        svc->layer_context[svc->spatial_layer_id].target_bandwidth /
+        10000000.0);
+  } else {
+    vp9_new_framerate(cpi, frame_rate);
+    twopass->bits_left = (int64_t)(stats->duration * oxcf->target_bandwidth /
+                             10000000.0);
+  }
+
+  // This variable monitors how far behind the second ref update is lagging.
+  twopass->sr_update_lag = 1;
+
+  // Scan the first pass file and calculate a modified total error based upon
+  // the bias/power function used to allocate bits.
+  {
+    const double avg_error = stats->coded_error /
+                             DOUBLE_DIVIDE_CHECK(stats->count);
+    const FIRSTPASS_STATS *s = twopass->stats_in;
+    double modified_error_total = 0.0;
+    twopass->modified_error_min = (avg_error *
+                                      oxcf->two_pass_vbrmin_section) / 100;
+    twopass->modified_error_max = (avg_error *
+                                      oxcf->two_pass_vbrmax_section) / 100;
+    while (s < twopass->stats_in_end) {
+      modified_error_total += calculate_modified_err(cpi, twopass, oxcf, s);
+      ++s;
+    }
+    twopass->modified_error_left = modified_error_total;
+  }
+
+  // Reset the vbr bits off target counters
+  rc->vbr_bits_off_target = 0;
+  rc->vbr_bits_off_target_fast = 0;
+  rc->rate_error_estimate = 0;
+
+  // Static sequence monitor variables.
+  twopass->kf_zeromotion_pct = 100;
+  twopass->last_kfgroup_zeromotion_pct = 100;
+
+  // Initialize bits per macro_block estimate correction factor.
+  twopass->bpm_factor = 1.0;
+  // Initiallize actual and target bits counters for ARF groups so that
+  // at the start we have a neutral bpm adjustment.
+  twopass->rolling_arf_group_target_bits = 1;
+  twopass->rolling_arf_group_actual_bits = 1;
+
+  if (oxcf->resize_mode != RESIZE_NONE) {
+    init_subsampling(cpi);
+  }
+}
+
+#define SR_DIFF_PART 0.0015
+#define INTRA_PART 0.005
+#define DEFAULT_DECAY_LIMIT 0.75
+#define LOW_SR_DIFF_TRHESH 0.1
+#define SR_DIFF_MAX 128.0
+#define LOW_CODED_ERR_PER_MB 10.0
+#define NCOUNT_FRAME_II_THRESH 6.0
+
+static double get_sr_decay_rate(const VP9_COMP *cpi,
+                                const FIRSTPASS_STATS *frame) {
+  const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
+                      ? cpi->initial_mbs : cpi->common.MBs;
+  double sr_diff =
+      (frame->sr_coded_error - frame->coded_error) / num_mbs;
+  double sr_decay = 1.0;
+  double modified_pct_inter;
+  double modified_pcnt_intra;
+  const double motion_amplitude_part =
+      frame->pcnt_motion *
+      ((frame->mvc_abs + frame->mvr_abs) /
+       (cpi->initial_height + cpi->initial_width));
+
+  modified_pct_inter = frame->pcnt_inter;
+  if (((frame->coded_error / num_mbs) > LOW_CODED_ERR_PER_MB) &&
+      ((frame->intra_error / DOUBLE_DIVIDE_CHECK(frame->coded_error)) <
+       (double)NCOUNT_FRAME_II_THRESH)) {
+    modified_pct_inter = frame->pcnt_inter - frame->pcnt_neutral;
+  }
+  modified_pcnt_intra = 100 * (1.0 - modified_pct_inter);
+
+
+  if ((sr_diff > LOW_SR_DIFF_TRHESH)) {
+    sr_diff = VPXMIN(sr_diff, SR_DIFF_MAX);
+    sr_decay = 1.0 - (SR_DIFF_PART * sr_diff) -
+               motion_amplitude_part -
+               (INTRA_PART * modified_pcnt_intra);
+  }
+  return VPXMAX(sr_decay, VPXMIN(DEFAULT_DECAY_LIMIT, modified_pct_inter));
+}
+
+// This function gives an estimate of how badly we believe the prediction
+// quality is decaying from frame to frame.
+static double get_zero_motion_factor(const VP9_COMP *cpi,
+                                     const FIRSTPASS_STATS *frame) {
+  const double zero_motion_pct = frame->pcnt_inter -
+                                 frame->pcnt_motion;
+  double sr_decay = get_sr_decay_rate(cpi, frame);
+  return VPXMIN(sr_decay, zero_motion_pct);
+}
+
+#define ZM_POWER_FACTOR 0.75
+
+static double get_prediction_decay_rate(const VP9_COMP *cpi,
+                                        const FIRSTPASS_STATS *next_frame) {
+  const double sr_decay_rate = get_sr_decay_rate(cpi, next_frame);
+  const double zero_motion_factor =
+    (0.95 * pow((next_frame->pcnt_inter - next_frame->pcnt_motion),
+                ZM_POWER_FACTOR));
+
+  return VPXMAX(zero_motion_factor,
+                (sr_decay_rate + ((1.0 - sr_decay_rate) * zero_motion_factor)));
+}
+
+// Function to test for a condition where a complex transition is followed
+// by a static section. For example in slide shows where there is a fade
+// between slides. This is to help with more optimal kf and gf positioning.
+static int detect_transition_to_still(VP9_COMP *cpi,
+                                      int frame_interval, int still_interval,
+                                      double loop_decay_rate,
+                                      double last_decay_rate) {
+  TWO_PASS *const twopass = &cpi->twopass;
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  // Break clause to detect very still sections after motion
+  // For example a static image after a fade or other transition
+  // instead of a clean scene cut.
+  if (frame_interval > rc->min_gf_interval &&
+      loop_decay_rate >= 0.999 &&
+      last_decay_rate < 0.9) {
+    int j;
+
+    // Look ahead a few frames to see if static condition persists...
+    for (j = 0; j < still_interval; ++j) {
+      const FIRSTPASS_STATS *stats = &twopass->stats_in[j];
+      if (stats >= twopass->stats_in_end)
+        break;
+
+      if (stats->pcnt_inter - stats->pcnt_motion < 0.999)
+        break;
+    }
+
+    // Only if it does do we signal a transition to still.
+    return j == still_interval;
+  }
+
+  return 0;
+}
+
+// This function detects a flash through the high relative pcnt_second_ref
+// score in the frame following a flash frame. The offset passed in should
+// reflect this.
+static int detect_flash(const TWO_PASS *twopass, int offset) {
+  const FIRSTPASS_STATS *const next_frame = read_frame_stats(twopass, offset);
+
+  // What we are looking for here is a situation where there is a
+  // brief break in prediction (such as a flash) but subsequent frames
+  // are reasonably well predicted by an earlier (pre flash) frame.
+  // The recovery after a flash is indicated by a high pcnt_second_ref
+  // compared to pcnt_inter.
+  return next_frame != NULL &&
+         next_frame->pcnt_second_ref > next_frame->pcnt_inter &&
+         next_frame->pcnt_second_ref >= 0.5;
+}
+
+// Update the motion related elements to the GF arf boost calculation.
+static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
+                                          double *mv_in_out,
+                                          double *mv_in_out_accumulator,
+                                          double *abs_mv_in_out_accumulator,
+                                          double *mv_ratio_accumulator) {
+  const double pct = stats->pcnt_motion;
+
+  // Accumulate Motion In/Out of frame stats.
+  *mv_in_out = stats->mv_in_out_count * pct;
+  *mv_in_out_accumulator += *mv_in_out;
+  *abs_mv_in_out_accumulator += fabs(*mv_in_out);
+
+  // Accumulate a measure of how uniform (or conversely how random) the motion
+  // field is (a ratio of abs(mv) / mv).
+  if (pct > 0.05) {
+    const double mvr_ratio = fabs(stats->mvr_abs) /
+                                 DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
+    const double mvc_ratio = fabs(stats->mvc_abs) /
+                                 DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
+
+    *mv_ratio_accumulator += pct * (mvr_ratio < stats->mvr_abs ?
+                                       mvr_ratio : stats->mvr_abs);
+    *mv_ratio_accumulator += pct * (mvc_ratio < stats->mvc_abs ?
+                                       mvc_ratio : stats->mvc_abs);
+  }
+}
+
+#define BASELINE_ERR_PER_MB 1000.0
+static double calc_frame_boost(VP9_COMP *cpi,
+                               const FIRSTPASS_STATS *this_frame,
+                               double this_frame_mv_in_out,
+                               double max_boost) {
+  double frame_boost;
+  const double lq =
+    vp9_convert_qindex_to_q(cpi->rc.avg_frame_qindex[INTER_FRAME],
+                            cpi->common.bit_depth);
+  const double boost_q_correction = VPXMIN((0.5 + (lq * 0.015)), 1.5);
+  int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
+                ? cpi->initial_mbs : cpi->common.MBs;
+
+  // Correct for any inactive region in the image
+  num_mbs = (int)VPXMAX(1, num_mbs * calculate_active_area(cpi, this_frame));
+
+  // Underlying boost factor is based on inter error ratio.
+  frame_boost = (BASELINE_ERR_PER_MB * num_mbs) /
+                DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
+  frame_boost = frame_boost * BOOST_FACTOR * boost_q_correction;
+
+  // Increase boost for frames where new data coming into frame (e.g. zoom out).
+  // Slightly reduce boost if there is a net balance of motion out of the frame
+  // (zoom in). The range for this_frame_mv_in_out is -1.0 to +1.0.
+  if (this_frame_mv_in_out > 0.0)
+    frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
+  // In the extreme case the boost is halved.
+  else
+    frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
+
+  return VPXMIN(frame_boost, max_boost * boost_q_correction);
+}
+
+static int calc_arf_boost(VP9_COMP *cpi, int offset,
+                          int f_frames, int b_frames,
+                          int *f_boost, int *b_boost) {
+  TWO_PASS *const twopass = &cpi->twopass;
+  int i;
+  double boost_score = 0.0;
+  double mv_ratio_accumulator = 0.0;
+  double decay_accumulator = 1.0;
+  double this_frame_mv_in_out = 0.0;
+  double mv_in_out_accumulator = 0.0;
+  double abs_mv_in_out_accumulator = 0.0;
+  int arf_boost;
+  int flash_detected = 0;
+
+  // Search forward from the proposed arf/next gf position.
+  for (i = 0; i < f_frames; ++i) {
+    const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
+    if (this_frame == NULL)
+      break;
+
+    // Update the motion related elements to the boost calculation.
+    accumulate_frame_motion_stats(this_frame,
+                                  &this_frame_mv_in_out, &mv_in_out_accumulator,
+                                  &abs_mv_in_out_accumulator,
+                                  &mv_ratio_accumulator);
+
+    // We want to discount the flash frame itself and the recovery
+    // frame that follows as both will have poor scores.
+    flash_detected = detect_flash(twopass, i + offset) ||
+                     detect_flash(twopass, i + offset + 1);
+
+    // Accumulate the effect of prediction quality decay.
+    if (!flash_detected) {
+      decay_accumulator *= get_prediction_decay_rate(cpi, this_frame);
+      decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
+                          ? MIN_DECAY_FACTOR : decay_accumulator;
+    }
+
+    boost_score += decay_accumulator * calc_frame_boost(cpi, this_frame,
+                                                        this_frame_mv_in_out,
+                                                        GF_MAX_BOOST);
+  }
+
+  *f_boost = (int)boost_score;
+
+  // Reset for backward looking loop.
+  boost_score = 0.0;
+  mv_ratio_accumulator = 0.0;
+  decay_accumulator = 1.0;
+  this_frame_mv_in_out = 0.0;
+  mv_in_out_accumulator = 0.0;
+  abs_mv_in_out_accumulator = 0.0;
+
+  // Search backward towards last gf position.
+  for (i = -1; i >= -b_frames; --i) {
+    const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
+    if (this_frame == NULL)
+      break;
+
+    // Update the motion related elements to the boost calculation.
+    accumulate_frame_motion_stats(this_frame,
+                                  &this_frame_mv_in_out, &mv_in_out_accumulator,
+                                  &abs_mv_in_out_accumulator,
+                                  &mv_ratio_accumulator);
+
+    // We want to discount the the flash frame itself and the recovery
+    // frame that follows as both will have poor scores.
+    flash_detected = detect_flash(twopass, i + offset) ||
+                     detect_flash(twopass, i + offset + 1);
+
+    // Cumulative effect of prediction quality decay.
+    if (!flash_detected) {
+      decay_accumulator *= get_prediction_decay_rate(cpi, this_frame);
+      decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
+                              ? MIN_DECAY_FACTOR : decay_accumulator;
+    }
+
+    boost_score += decay_accumulator * calc_frame_boost(cpi, this_frame,
+                                                        this_frame_mv_in_out,
+                                                        GF_MAX_BOOST);
+  }
+  *b_boost = (int)boost_score;
+
+  arf_boost = (*f_boost + *b_boost);
+  if (arf_boost < ((b_frames + f_frames) * 20))
+    arf_boost = ((b_frames + f_frames) * 20);
+  arf_boost = VPXMAX(arf_boost, MIN_ARF_GF_BOOST);
+
+  return arf_boost;
+}
+
+// Calculate a section intra ratio used in setting max loop filter.
+static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
+                                         const FIRSTPASS_STATS *end,
+                                         int section_length) {
+  const FIRSTPASS_STATS *s = begin;
+  double intra_error = 0.0;
+  double coded_error = 0.0;
+  int i = 0;
+
+  while (s < end && i < section_length) {
+    intra_error += s->intra_error;
+    coded_error += s->coded_error;
+    ++s;
+    ++i;
+  }
+
+  return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
+}
+
+// Calculate the total bits to allocate in this GF/ARF group.
+static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
+                                             double gf_group_err) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const TWO_PASS *const twopass = &cpi->twopass;
+  const int max_bits = frame_max_bits(rc, &cpi->oxcf);
+  int64_t total_group_bits;
+
+  // Calculate the bits to be allocated to the group as a whole.
+  if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
+    total_group_bits = (int64_t)(twopass->kf_group_bits *
+                                 (gf_group_err / twopass->kf_group_error_left));
+  } else {
+    total_group_bits = 0;
+  }
+
+  // Clamp odd edge cases.
+  total_group_bits = (total_group_bits < 0) ?
+     0 : (total_group_bits > twopass->kf_group_bits) ?
+     twopass->kf_group_bits : total_group_bits;
+
+  // Clip based on user supplied data rate variability limit.
+  if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
+    total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
+
+  return total_group_bits;
+}
+
+// Calculate the number bits extra to assign to boosted frames in a group.
+static int calculate_boost_bits(int frame_count,
+                                int boost, int64_t total_group_bits) {
+  int allocation_chunks;
+
+  // return 0 for invalid inputs (could arise e.g. through rounding errors)
+  if (!boost || (total_group_bits <= 0) || (frame_count <= 0) )
+    return 0;
+
+  allocation_chunks = (frame_count * 100) + boost;
+
+  // Prevent overflow.
+  if (boost > 1023) {
+    int divisor = boost >> 10;
+    boost /= divisor;
+    allocation_chunks /= divisor;
+  }
+
+  // Calculate the number of extra bits for use in the boosted frame or frames.
+  return VPXMAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks),
+                0);
+}
+
+// Current limit on maximum number of active arfs in a GF/ARF group.
+#define MAX_ACTIVE_ARFS 2
+#define ARF_SLOT1 2
+#define ARF_SLOT2 3
+// This function indirects the choice of buffers for arfs.
+// At the moment the values are fixed but this may change as part of
+// the integration process with other codec features that swap buffers around.
+static void get_arf_buffer_indices(unsigned char *arf_buffer_indices) {
+  arf_buffer_indices[0] = ARF_SLOT1;
+  arf_buffer_indices[1] = ARF_SLOT2;
+}
+
+static void allocate_gf_group_bits(VP9_COMP *cpi, int64_t gf_group_bits,
+                                   double group_error, int gf_arf_bits) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  TWO_PASS *const twopass = &cpi->twopass;
+  GF_GROUP *const gf_group = &twopass->gf_group;
+  FIRSTPASS_STATS frame_stats;
+  int i;
+  int frame_index = 1;
+  int target_frame_size;
+  int key_frame;
+  const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
+  int64_t total_group_bits = gf_group_bits;
+  double modified_err = 0.0;
+  double err_fraction;
+  int mid_boost_bits = 0;
+  int mid_frame_idx;
+  unsigned char arf_buffer_indices[MAX_ACTIVE_ARFS];
+  int alt_frame_index = frame_index;
+  int has_temporal_layers = is_two_pass_svc(cpi) &&
+                            cpi->svc.number_temporal_layers > 1;
+
+  // Only encode alt reference frame in temporal base layer.
+  if (has_temporal_layers)
+    alt_frame_index = cpi->svc.number_temporal_layers;
+
+  key_frame = cpi->common.frame_type == KEY_FRAME ||
+              vp9_is_upper_layer_key_frame(cpi);
+
+  get_arf_buffer_indices(arf_buffer_indices);
+
+  // For key frames the frame target rate is already set and it
+  // is also the golden frame.
+  if (!key_frame) {
+    if (rc->source_alt_ref_active) {
+      gf_group->update_type[0] = OVERLAY_UPDATE;
+      gf_group->rf_level[0] = INTER_NORMAL;
+      gf_group->bit_allocation[0] = 0;
+    } else {
+      gf_group->update_type[0] = GF_UPDATE;
+      gf_group->rf_level[0] = GF_ARF_STD;
+      gf_group->bit_allocation[0] = gf_arf_bits;
+    }
+    gf_group->arf_update_idx[0] = arf_buffer_indices[0];
+    gf_group->arf_ref_idx[0] = arf_buffer_indices[0];
+
+    // Step over the golden frame / overlay frame
+    if (EOF == input_stats(twopass, &frame_stats))
+      return;
+  }
+
+  // Deduct the boost bits for arf (or gf if it is not a key frame)
+  // from the group total.
+  if (rc->source_alt_ref_pending || !key_frame)
+    total_group_bits -= gf_arf_bits;
+
+  // Store the bits to spend on the ARF if there is one.
+  if (rc->source_alt_ref_pending) {
+    gf_group->update_type[alt_frame_index] = ARF_UPDATE;
+    gf_group->rf_level[alt_frame_index] = GF_ARF_STD;
+    gf_group->bit_allocation[alt_frame_index] = gf_arf_bits;
+
+    if (has_temporal_layers)
+      gf_group->arf_src_offset[alt_frame_index] =
+          (unsigned char)(rc->baseline_gf_interval -
+                          cpi->svc.number_temporal_layers);
+    else
+      gf_group->arf_src_offset[alt_frame_index] =
+          (unsigned char)(rc->baseline_gf_interval - 1);
+
+    gf_group->arf_update_idx[alt_frame_index] = arf_buffer_indices[0];
+    gf_group->arf_ref_idx[alt_frame_index] =
+      arf_buffer_indices[cpi->multi_arf_last_grp_enabled &&
+                         rc->source_alt_ref_active];
+    if (!has_temporal_layers)
+      ++frame_index;
+
+    if (cpi->multi_arf_enabled) {
+      // Set aside a slot for a level 1 arf.
+      gf_group->update_type[frame_index] = ARF_UPDATE;
+      gf_group->rf_level[frame_index] = GF_ARF_LOW;
+      gf_group->arf_src_offset[frame_index] =
+        (unsigned char)((rc->baseline_gf_interval >> 1) - 1);
+      gf_group->arf_update_idx[frame_index] = arf_buffer_indices[1];
+      gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[0];
+      ++frame_index;
+    }
+  }
+
+  // Define middle frame
+  mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
+
+  // Allocate bits to the other frames in the group.
+  for (i = 0; i < rc->baseline_gf_interval - rc->source_alt_ref_pending; ++i) {
+    int arf_idx = 0;
+    if (EOF == input_stats(twopass, &frame_stats))
+      break;
+
+    if (has_temporal_layers && frame_index == alt_frame_index) {
+      ++frame_index;
+    }
+
+    modified_err = calculate_modified_err(cpi, twopass, oxcf, &frame_stats);
+
+    if (group_error > 0)
+      err_fraction = modified_err / DOUBLE_DIVIDE_CHECK(group_error);
+    else
+      err_fraction = 0.0;
+
+    target_frame_size = (int)((double)total_group_bits * err_fraction);
+
+    if (rc->source_alt_ref_pending && cpi->multi_arf_enabled) {
+      mid_boost_bits += (target_frame_size >> 4);
+      target_frame_size -= (target_frame_size >> 4);
+
+      if (frame_index <= mid_frame_idx)
+        arf_idx = 1;
+    }
+    gf_group->arf_update_idx[frame_index] = arf_buffer_indices[arf_idx];
+    gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[arf_idx];
+
+    target_frame_size = clamp(target_frame_size, 0,
+                              VPXMIN(max_bits, (int)total_group_bits));
+
+    gf_group->update_type[frame_index] = LF_UPDATE;
+    gf_group->rf_level[frame_index] = INTER_NORMAL;
+
+    gf_group->bit_allocation[frame_index] = target_frame_size;
+    ++frame_index;
+  }
+
+  // Note:
+  // We need to configure the frame at the end of the sequence + 1 that will be
+  // the start frame for the next group. Otherwise prior to the call to
+  // vp9_rc_get_second_pass_params() the data will be undefined.
+  gf_group->arf_update_idx[frame_index] = arf_buffer_indices[0];
+  gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[0];
+
+  if (rc->source_alt_ref_pending) {
+    gf_group->update_type[frame_index] = OVERLAY_UPDATE;
+    gf_group->rf_level[frame_index] = INTER_NORMAL;
+
+    // Final setup for second arf and its overlay.
+    if (cpi->multi_arf_enabled) {
+      gf_group->bit_allocation[2] =
+          gf_group->bit_allocation[mid_frame_idx] + mid_boost_bits;
+      gf_group->update_type[mid_frame_idx] = OVERLAY_UPDATE;
+      gf_group->bit_allocation[mid_frame_idx] = 0;
+    }
+  } else {
+    gf_group->update_type[frame_index] = GF_UPDATE;
+    gf_group->rf_level[frame_index] = GF_ARF_STD;
+  }
+
+  // Note whether multi-arf was enabled this group for next time.
+  cpi->multi_arf_last_grp_enabled = cpi->multi_arf_enabled;
+}
+
+// Analyse and define a gf/arf group.
+static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
+  VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  TWO_PASS *const twopass = &cpi->twopass;
+  FIRSTPASS_STATS next_frame;
+  const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
+  int i;
+
+  double boost_score = 0.0;
+  double old_boost_score = 0.0;
+  double gf_group_err = 0.0;
+  double gf_group_raw_error = 0.0;
+  double gf_group_skip_pct = 0.0;
+  double gf_group_inactive_zone_rows = 0.0;
+  double gf_first_frame_err = 0.0;
+  double mod_frame_err = 0.0;
+
+  double mv_ratio_accumulator = 0.0;
+  double decay_accumulator = 1.0;
+  double zero_motion_accumulator = 1.0;
+
+  double loop_decay_rate = 1.00;
+  double last_loop_decay_rate = 1.00;
+
+  double this_frame_mv_in_out = 0.0;
+  double mv_in_out_accumulator = 0.0;
+  double abs_mv_in_out_accumulator = 0.0;
+  double mv_ratio_accumulator_thresh;
+  unsigned int allow_alt_ref = is_altref_enabled(cpi);
+
+  int f_boost = 0;
+  int b_boost = 0;
+  int flash_detected;
+  int active_max_gf_interval;
+  int active_min_gf_interval;
+  int64_t gf_group_bits;
+  double gf_group_error_left;
+  int gf_arf_bits;
+  const int is_key_frame = frame_is_intra_only(cm);
+  const int arf_active_or_kf = is_key_frame || rc->source_alt_ref_active;
+
+  // Reset the GF group data structures unless this is a key
+  // frame in which case it will already have been done.
+  if (is_key_frame == 0) {
+    vp9_zero(twopass->gf_group);
+  }
+
+  vpx_clear_system_state();
+  vp9_zero(next_frame);
+
+  // Load stats for the current frame.
+  mod_frame_err = calculate_modified_err(cpi, twopass, oxcf, this_frame);
+
+  // Note the error of the frame at the start of the group. This will be
+  // the GF frame error if we code a normal gf.
+  gf_first_frame_err = mod_frame_err;
+
+  // If this is a key frame or the overlay from a previous arf then
+  // the error score / cost of this frame has already been accounted for.
+  if (arf_active_or_kf) {
+    gf_group_err -= gf_first_frame_err;
+    gf_group_raw_error -= this_frame->coded_error;
+    gf_group_skip_pct -= this_frame->intra_skip_pct;
+    gf_group_inactive_zone_rows -= this_frame->inactive_zone_rows;
+  }
+
+  // Motion breakout threshold for loop below depends on image size.
+  mv_ratio_accumulator_thresh =
+      (cpi->initial_height + cpi->initial_width) / 4.0;
+
+  // Set a maximum and minimum interval for the GF group.
+  // If the image appears almost completely static we can extend beyond this.
+  {
+    int int_max_q =
+      (int)(vp9_convert_qindex_to_q(twopass->active_worst_quality,
+                                   cpi->common.bit_depth));
+    int int_lbq =
+      (int)(vp9_convert_qindex_to_q(rc->last_boosted_qindex,
+                                   cpi->common.bit_depth));
+    active_min_gf_interval =
+      rc->min_gf_interval + arf_active_or_kf + VPXMIN(2, int_max_q / 200);
+    if (active_min_gf_interval > rc->max_gf_interval)
+      active_min_gf_interval = rc->max_gf_interval;
+
+    if (cpi->multi_arf_allowed) {
+      active_max_gf_interval = rc->max_gf_interval;
+    } else {
+      // The value chosen depends on the active Q range. At low Q we have
+      // bits to spare and are better with a smaller interval and smaller boost.
+      // At high Q when there are few bits to spare we are better with a longer
+      // interval to spread the cost of the GF.
+      active_max_gf_interval =
+        12 + arf_active_or_kf + VPXMIN(4, (int_lbq / 6));
+
+      // We have: active_min_gf_interval <= rc->max_gf_interval
+      if (active_max_gf_interval < active_min_gf_interval)
+        active_max_gf_interval = active_min_gf_interval;
+      else if (active_max_gf_interval > rc->max_gf_interval)
+        active_max_gf_interval = rc->max_gf_interval;
+
+      // Would the active max drop us out just before the near the next kf?
+      if ((active_max_gf_interval <= rc->frames_to_key) &&
+          (active_max_gf_interval >=
+              (rc->frames_to_key - rc->min_gf_interval)))
+        active_max_gf_interval = rc->frames_to_key / 2;
+    }
+  }
+
+  i = 0;
+  while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
+    ++i;
+
+    // Accumulate error score of frames in this gf group.
+    mod_frame_err = calculate_modified_err(cpi, twopass, oxcf, this_frame);
+    gf_group_err += mod_frame_err;
+    gf_group_raw_error += this_frame->coded_error;
+    gf_group_skip_pct += this_frame->intra_skip_pct;
+    gf_group_inactive_zone_rows += this_frame->inactive_zone_rows;
+
+    if (EOF == input_stats(twopass, &next_frame))
+      break;
+
+    // Test for the case where there is a brief flash but the prediction
+    // quality back to an earlier frame is then restored.
+    flash_detected = detect_flash(twopass, 0);
+
+    // Update the motion related elements to the boost calculation.
+    accumulate_frame_motion_stats(&next_frame,
+                                  &this_frame_mv_in_out, &mv_in_out_accumulator,
+                                  &abs_mv_in_out_accumulator,
+                                  &mv_ratio_accumulator);
+
+    // Accumulate the effect of prediction quality decay.
+    if (!flash_detected) {
+      last_loop_decay_rate = loop_decay_rate;
+      loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
+
+      decay_accumulator = decay_accumulator * loop_decay_rate;
+
+      // Monitor for static sections.
+      zero_motion_accumulator = VPXMIN(
+          zero_motion_accumulator, get_zero_motion_factor(cpi, &next_frame));
+
+      // Break clause to detect very still sections after motion. For example,
+      // a static image after a fade or other transition.
+      if (detect_transition_to_still(cpi, i, 5, loop_decay_rate,
+                                     last_loop_decay_rate)) {
+        allow_alt_ref = 0;
+        break;
+      }
+    }
+
+    // Calculate a boost number for this frame.
+    boost_score += decay_accumulator * calc_frame_boost(cpi, &next_frame,
+                                                        this_frame_mv_in_out,
+                                                        GF_MAX_BOOST);
+
+    // Break out conditions.
+    if (
+      // Break at active_max_gf_interval unless almost totally static.
+      ((i >= active_max_gf_interval) &&
+       (zero_motion_accumulator < 0.995)) ||
+      (
+        // Don't break out with a very short interval.
+        (i >= active_min_gf_interval) &&
+        // If possible dont break very close to a kf
+        ((rc->frames_to_key - i) >= rc->min_gf_interval) &&
+        (!flash_detected) &&
+        ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
+         (abs_mv_in_out_accumulator > 3.0) ||
+         (mv_in_out_accumulator < -2.0) ||
+         ((boost_score - old_boost_score) < BOOST_BREAKOUT)))) {
+      boost_score = old_boost_score;
+      break;
+    }
+
+    *this_frame = next_frame;
+    old_boost_score = boost_score;
+  }
+
+  // Was the group length constrained by the requirement for a new KF?
+  rc->constrained_gf_group = (i >= rc->frames_to_key) ? 1 : 0;
+
+  // Should we use the alternate reference frame.
+  if (allow_alt_ref &&
+    (i < cpi->oxcf.lag_in_frames) &&
+    (i >= rc->min_gf_interval)) {
+    // Calculate the boost for alt ref.
+    rc->gfu_boost = calc_arf_boost(cpi, 0, (i - 1), (i - 1), &f_boost,
+      &b_boost);
+    rc->source_alt_ref_pending = 1;
+
+    // Test to see if multi arf is appropriate.
+    cpi->multi_arf_enabled =
+      (cpi->multi_arf_allowed && (rc->baseline_gf_interval >= 6) &&
+      (zero_motion_accumulator < 0.995)) ? 1 : 0;
+  } else {
+    rc->gfu_boost = VPXMAX((int)boost_score, MIN_ARF_GF_BOOST);
+    rc->source_alt_ref_pending = 0;
+  }
+
+  // Set the interval until the next gf.
+  rc->baseline_gf_interval = i - (is_key_frame || rc->source_alt_ref_pending);
+
+  // Only encode alt reference frame in temporal base layer. So
+  // baseline_gf_interval should be multiple of a temporal layer group
+  // (typically the frame distance between two base layer frames)
+  if (is_two_pass_svc(cpi) && cpi->svc.number_temporal_layers > 1) {
+    int count = (1 << (cpi->svc.number_temporal_layers - 1)) - 1;
+    int new_gf_interval = (rc->baseline_gf_interval + count) & (~count);
+    int j;
+    for (j = 0; j < new_gf_interval - rc->baseline_gf_interval; ++j) {
+      if (EOF == input_stats(twopass, this_frame))
+        break;
+      gf_group_err += calculate_modified_err(cpi, twopass, oxcf, this_frame);
+      gf_group_raw_error += this_frame->coded_error;
+      gf_group_skip_pct += this_frame->intra_skip_pct;
+      gf_group_inactive_zone_rows += this_frame->inactive_zone_rows;
+    }
+    rc->baseline_gf_interval = new_gf_interval;
+  }
+
+  rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+
+  // Reset the file position.
+  reset_fpf_position(twopass, start_pos);
+
+  // Calculate the bits to be allocated to the gf/arf group as a whole
+  gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
+
+  // Calculate an estimate of the maxq needed for the group.
+  // We are more agressive about correcting for sections
+  // where there could be significant overshoot than for easier
+  // sections where we do not wish to risk creating an overshoot
+  // of the allocated bit budget.
+  if ((cpi->oxcf.rc_mode != VPX_Q) && (rc->baseline_gf_interval > 1)) {
+    const int vbr_group_bits_per_frame =
+      (int)(gf_group_bits / rc->baseline_gf_interval);
+    const double group_av_err = gf_group_raw_error  / rc->baseline_gf_interval;
+    const double group_av_skip_pct =
+      gf_group_skip_pct / rc->baseline_gf_interval;
+    const double group_av_inactive_zone =
+      ((gf_group_inactive_zone_rows * 2) /
+       (rc->baseline_gf_interval * (double)cm->mb_rows));
+    int tmp_q =
+        get_twopass_worst_quality(cpi, group_av_err,
+                                  (group_av_skip_pct + group_av_inactive_zone),
+                                  vbr_group_bits_per_frame);
+    twopass->active_worst_quality =
+        (tmp_q + (twopass->active_worst_quality * 3)) >> 2;
+  }
+
+  // Calculate the extra bits to be used for boosted frame(s)
+  gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval,
+                                     rc->gfu_boost, gf_group_bits);
+
+  // Adjust KF group bits and error remaining.
+  twopass->kf_group_error_left -= (int64_t)gf_group_err;
+
+  // If this is an arf update we want to remove the score for the overlay
+  // frame at the end which will usually be very cheap to code.
+  // The overlay frame has already, in effect, been coded so we want to spread
+  // the remaining bits among the other frames.
+  // For normal GFs remove the score for the GF itself unless this is
+  // also a key frame in which case it has already been accounted for.
+  if (rc->source_alt_ref_pending) {
+    gf_group_error_left = gf_group_err - mod_frame_err;
+  } else if (is_key_frame == 0) {
+    gf_group_error_left = gf_group_err - gf_first_frame_err;
+  } else {
+    gf_group_error_left = gf_group_err;
+  }
+
+  // Allocate bits to each of the frames in the GF group.
+  allocate_gf_group_bits(cpi, gf_group_bits, gf_group_error_left, gf_arf_bits);
+
+  // Reset the file position.
+  reset_fpf_position(twopass, start_pos);
+
+  // Calculate a section intra ratio used in setting max loop filter.
+  if (cpi->common.frame_type != KEY_FRAME) {
+    twopass->section_intra_rating =
+        calculate_section_intra_ratio(start_pos, twopass->stats_in_end,
+                                      rc->baseline_gf_interval);
+  }
+
+  if (oxcf->resize_mode == RESIZE_DYNAMIC) {
+    // Default to starting GF groups at normal frame size.
+    cpi->rc.next_frame_size_selector = UNSCALED;
+  }
+
+  // Reset rolling actual and target bits counters for ARF groups.
+  twopass->rolling_arf_group_target_bits = 0;
+  twopass->rolling_arf_group_actual_bits = 0;
+}
+
+// Threshold for use of the lagging second reference frame. High second ref
+// usage may point to a transient event like a flash or occlusion rather than
+// a real scene cut.
+#define SECOND_REF_USEAGE_THRESH 0.1
+// Minimum % intra coding observed in first pass (1.0 = 100%)
+#define MIN_INTRA_LEVEL 0.25
+// Minimum ratio between the % of intra coding and inter coding in the first
+// pass after discounting neutral blocks (discounting neutral blocks in this
+// way helps catch scene cuts in clips with very flat areas or letter box
+// format clips with image padding.
+#define INTRA_VS_INTER_THRESH 2.0
+// Hard threshold where the first pass chooses intra for almost all blocks.
+// In such a case even if the frame is not a scene cut coding a key frame
+// may be a good option.
+#define VERY_LOW_INTER_THRESH 0.05
+// Maximum threshold for the relative ratio of intra error score vs best
+// inter error score.
+#define KF_II_ERR_THRESHOLD 2.5
+// In real scene cuts there is almost always a sharp change in the intra
+// or inter error score.
+#define ERR_CHANGE_THRESHOLD 0.4
+// For real scene cuts we expect an improvment in the intra inter error
+// ratio in the next frame.
+#define II_IMPROVEMENT_THRESHOLD 3.5
+#define KF_II_MAX 128.0
+
+static int test_candidate_kf(TWO_PASS *twopass,
+                             const FIRSTPASS_STATS *last_frame,
+                             const FIRSTPASS_STATS *this_frame,
+                             const FIRSTPASS_STATS *next_frame) {
+  int is_viable_kf = 0;
+  double pcnt_intra = 1.0 - this_frame->pcnt_inter;
+  double modified_pcnt_inter =
+    this_frame->pcnt_inter - this_frame->pcnt_neutral;
+
+  // Does the frame satisfy the primary criteria of a key frame?
+  // See above for an explanation of the test criteria.
+  // If so, then examine how well it predicts subsequent frames.
+  if ((this_frame->pcnt_second_ref < SECOND_REF_USEAGE_THRESH) &&
+      (next_frame->pcnt_second_ref < SECOND_REF_USEAGE_THRESH) &&
+      ((this_frame->pcnt_inter < VERY_LOW_INTER_THRESH) ||
+       ((pcnt_intra > MIN_INTRA_LEVEL) &&
+        (pcnt_intra > (INTRA_VS_INTER_THRESH * modified_pcnt_inter)) &&
+        ((this_frame->intra_error /
+          DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) <
+          KF_II_ERR_THRESHOLD) &&
+        ((fabs(last_frame->coded_error - this_frame->coded_error) /
+          DOUBLE_DIVIDE_CHECK(this_frame->coded_error) >
+          ERR_CHANGE_THRESHOLD) ||
+         (fabs(last_frame->intra_error - this_frame->intra_error) /
+          DOUBLE_DIVIDE_CHECK(this_frame->intra_error) >
+          ERR_CHANGE_THRESHOLD) ||
+         ((next_frame->intra_error /
+          DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) >
+          II_IMPROVEMENT_THRESHOLD))))) {
+    int i;
+    const FIRSTPASS_STATS *start_pos = twopass->stats_in;
+    FIRSTPASS_STATS local_next_frame = *next_frame;
+    double boost_score = 0.0;
+    double old_boost_score = 0.0;
+    double decay_accumulator = 1.0;
+
+    // Examine how well the key frame predicts subsequent frames.
+    for (i = 0; i < 16; ++i) {
+      double next_iiratio = (BOOST_FACTOR * local_next_frame.intra_error /
+                             DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
+
+      if (next_iiratio > KF_II_MAX)
+        next_iiratio = KF_II_MAX;
+
+      // Cumulative effect of decay in prediction quality.
+      if (local_next_frame.pcnt_inter > 0.85)
+        decay_accumulator *= local_next_frame.pcnt_inter;
+      else
+        decay_accumulator *= (0.85 + local_next_frame.pcnt_inter) / 2.0;
+
+      // Keep a running total.
+      boost_score += (decay_accumulator * next_iiratio);
+
+      // Test various breakout clauses.
+      if ((local_next_frame.pcnt_inter < 0.05) ||
+          (next_iiratio < 1.5) ||
+          (((local_next_frame.pcnt_inter -
+             local_next_frame.pcnt_neutral) < 0.20) &&
+           (next_iiratio < 3.0)) ||
+          ((boost_score - old_boost_score) < 3.0) ||
+          (local_next_frame.intra_error < 200)) {
+        break;
+      }
+
+      old_boost_score = boost_score;
+
+      // Get the next frame details
+      if (EOF == input_stats(twopass, &local_next_frame))
+        break;
+    }
+
+    // If there is tolerable prediction for at least the next 3 frames then
+    // break out else discard this potential key frame and move on
+    if (boost_score > 30.0 && (i > 3)) {
+      is_viable_kf = 1;
+    } else {
+      // Reset the file position
+      reset_fpf_position(twopass, start_pos);
+
+      is_viable_kf = 0;
+    }
+  }
+
+  return is_viable_kf;
+}
+
+#define FRAMES_TO_CHECK_DECAY 8
+
+static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
+  int i, j;
+  RATE_CONTROL *const rc = &cpi->rc;
+  TWO_PASS *const twopass = &cpi->twopass;
+  GF_GROUP *const gf_group = &twopass->gf_group;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const FIRSTPASS_STATS first_frame = *this_frame;
+  const FIRSTPASS_STATS *const start_position = twopass->stats_in;
+  FIRSTPASS_STATS next_frame;
+  FIRSTPASS_STATS last_frame;
+  int kf_bits = 0;
+  int loop_decay_counter = 0;
+  double decay_accumulator = 1.0;
+  double av_decay_accumulator = 0.0;
+  double zero_motion_accumulator = 1.0;
+  double boost_score = 0.0;
+  double kf_mod_err = 0.0;
+  double kf_group_err = 0.0;
+  double recent_loop_decay[FRAMES_TO_CHECK_DECAY];
+
+  vp9_zero(next_frame);
+
+  cpi->common.frame_type = KEY_FRAME;
+
+  // Reset the GF group data structures.
+  vp9_zero(*gf_group);
+
+  // Is this a forced key frame by interval.
+  rc->this_key_frame_forced = rc->next_key_frame_forced;
+
+  // Clear the alt ref active flag and last group multi arf flags as they
+  // can never be set for a key frame.
+  rc->source_alt_ref_active = 0;
+  cpi->multi_arf_last_grp_enabled = 0;
+
+  // KF is always a GF so clear frames till next gf counter.
+  rc->frames_till_gf_update_due = 0;
+
+  rc->frames_to_key = 1;
+
+  twopass->kf_group_bits = 0;        // Total bits available to kf group
+  twopass->kf_group_error_left = 0;  // Group modified error score.
+
+  kf_mod_err = calculate_modified_err(cpi, twopass, oxcf, this_frame);
+
+  // Initialize the decay rates for the recent frames to check
+  for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j)
+    recent_loop_decay[j] = 1.0;
+
+  // Find the next keyframe.
+  i = 0;
+  while (twopass->stats_in < twopass->stats_in_end &&
+         rc->frames_to_key < cpi->oxcf.key_freq) {
+    // Accumulate kf group error.
+    kf_group_err += calculate_modified_err(cpi, twopass, oxcf, this_frame);
+
+    // Load the next frame's stats.
+    last_frame = *this_frame;
+    input_stats(twopass, this_frame);
+
+    // Provided that we are not at the end of the file...
+    if (cpi->oxcf.auto_key && twopass->stats_in < twopass->stats_in_end) {
+      double loop_decay_rate;
+
+      // Check for a scene cut.
+      if (test_candidate_kf(twopass, &last_frame, this_frame,
+                            twopass->stats_in))
+        break;
+
+      // How fast is the prediction quality decaying?
+      loop_decay_rate = get_prediction_decay_rate(cpi, twopass->stats_in);
+
+      // We want to know something about the recent past... rather than
+      // as used elsewhere where we are concerned with decay in prediction
+      // quality since the last GF or KF.
+      recent_loop_decay[i % FRAMES_TO_CHECK_DECAY] = loop_decay_rate;
+      decay_accumulator = 1.0;
+      for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j)
+        decay_accumulator *= recent_loop_decay[j];
+
+      // Special check for transition or high motion followed by a
+      // static scene.
+      if (detect_transition_to_still(cpi, i, cpi->oxcf.key_freq - i,
+                                     loop_decay_rate, decay_accumulator))
+        break;
+
+      // Step on to the next frame.
+      ++rc->frames_to_key;
+
+      // If we don't have a real key frame within the next two
+      // key_freq intervals then break out of the loop.
+      if (rc->frames_to_key >= 2 * cpi->oxcf.key_freq)
+        break;
+    } else {
+      ++rc->frames_to_key;
+    }
+    ++i;
+  }
+
+  // If there is a max kf interval set by the user we must obey it.
+  // We already breakout of the loop above at 2x max.
+  // This code centers the extra kf if the actual natural interval
+  // is between 1x and 2x.
+  if (cpi->oxcf.auto_key &&
+      rc->frames_to_key > cpi->oxcf.key_freq) {
+    FIRSTPASS_STATS tmp_frame = first_frame;
+
+    rc->frames_to_key /= 2;
+
+    // Reset to the start of the group.
+    reset_fpf_position(twopass, start_position);
+
+    kf_group_err = 0.0;
+
+    // Rescan to get the correct error data for the forced kf group.
+    for (i = 0; i < rc->frames_to_key; ++i) {
+      kf_group_err += calculate_modified_err(cpi, twopass, oxcf, &tmp_frame);
+      input_stats(twopass, &tmp_frame);
+    }
+    rc->next_key_frame_forced = 1;
+  } else if (twopass->stats_in == twopass->stats_in_end ||
+             rc->frames_to_key >= cpi->oxcf.key_freq) {
+    rc->next_key_frame_forced = 1;
+  } else {
+    rc->next_key_frame_forced = 0;
+  }
+
+  if (is_two_pass_svc(cpi) && cpi->svc.number_temporal_layers > 1) {
+    int count = (1 << (cpi->svc.number_temporal_layers - 1)) - 1;
+    int new_frame_to_key = (rc->frames_to_key + count) & (~count);
+    int j;
+    for (j = 0; j < new_frame_to_key - rc->frames_to_key; ++j) {
+      if (EOF == input_stats(twopass, this_frame))
+        break;
+      kf_group_err += calculate_modified_err(cpi, twopass, oxcf, this_frame);
+    }
+    rc->frames_to_key = new_frame_to_key;
+  }
+
+  // Special case for the last key frame of the file.
+  if (twopass->stats_in >= twopass->stats_in_end) {
+    // Accumulate kf group error.
+    kf_group_err += calculate_modified_err(cpi, twopass, oxcf, this_frame);
+  }
+
+  // Calculate the number of bits that should be assigned to the kf group.
+  if (twopass->bits_left > 0 && twopass->modified_error_left > 0.0) {
+    // Maximum number of bits for a single normal frame (not key frame).
+    const int max_bits = frame_max_bits(rc, &cpi->oxcf);
+
+    // Maximum number of bits allocated to the key frame group.
+    int64_t max_grp_bits;
+
+    // Default allocation based on bits left and relative
+    // complexity of the section.
+    twopass->kf_group_bits = (int64_t)(twopass->bits_left *
+       (kf_group_err / twopass->modified_error_left));
+
+    // Clip based on maximum per frame rate defined by the user.
+    max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
+    if (twopass->kf_group_bits > max_grp_bits)
+      twopass->kf_group_bits = max_grp_bits;
+  } else {
+    twopass->kf_group_bits = 0;
+  }
+  twopass->kf_group_bits = VPXMAX(0, twopass->kf_group_bits);
+
+  // Reset the first pass file position.
+  reset_fpf_position(twopass, start_position);
+
+  // Scan through the kf group collating various stats used to determine
+  // how many bits to spend on it.
+  decay_accumulator = 1.0;
+  boost_score = 0.0;
+  for (i = 0; i < (rc->frames_to_key - 1); ++i) {
+    if (EOF == input_stats(twopass, &next_frame))
+      break;
+
+    // Monitor for static sections.
+    zero_motion_accumulator = VPXMIN(
+        zero_motion_accumulator, get_zero_motion_factor(cpi, &next_frame));
+
+    // Not all frames in the group are necessarily used in calculating boost.
+    if ((i <= rc->max_gf_interval) ||
+        ((i <= (rc->max_gf_interval * 4)) && (decay_accumulator > 0.5))) {
+      const double frame_boost =
+        calc_frame_boost(cpi, &next_frame, 0, KF_MAX_BOOST);
+
+      // How fast is prediction quality decaying.
+      if (!detect_flash(twopass, 0)) {
+        const double loop_decay_rate =
+          get_prediction_decay_rate(cpi, &next_frame);
+        decay_accumulator *= loop_decay_rate;
+        decay_accumulator = VPXMAX(decay_accumulator, MIN_DECAY_FACTOR);
+        av_decay_accumulator += decay_accumulator;
+        ++loop_decay_counter;
+      }
+      boost_score += (decay_accumulator * frame_boost);
+    }
+  }
+  av_decay_accumulator /= (double)loop_decay_counter;
+
+  reset_fpf_position(twopass, start_position);
+
+  // Store the zero motion percentage
+  twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
+
+  // Calculate a section intra ratio used in setting max loop filter.
+  twopass->section_intra_rating =
+      calculate_section_intra_ratio(start_position, twopass->stats_in_end,
+                                    rc->frames_to_key);
+
+  // Apply various clamps for min and max boost
+  rc->kf_boost = (int)(av_decay_accumulator * boost_score);
+  rc->kf_boost = VPXMAX(rc->kf_boost, (rc->frames_to_key * 3));
+  rc->kf_boost = VPXMAX(rc->kf_boost, MIN_KF_BOOST);
+
+  // Work out how many bits to allocate for the key frame itself.
+  kf_bits = calculate_boost_bits((rc->frames_to_key - 1),
+                                  rc->kf_boost, twopass->kf_group_bits);
+
+  twopass->kf_group_bits -= kf_bits;
+
+  // Save the bits to spend on the key frame.
+  gf_group->bit_allocation[0] = kf_bits;
+  gf_group->update_type[0] = KF_UPDATE;
+  gf_group->rf_level[0] = KF_STD;
+
+  // Note the total error score of the kf group minus the key frame itself.
+  twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
+
+  // Adjust the count of total modified error left.
+  // The count of bits left is adjusted elsewhere based on real coded frame
+  // sizes.
+  twopass->modified_error_left -= kf_group_err;
+
+  if (oxcf->resize_mode == RESIZE_DYNAMIC) {
+    // Default to normal-sized frame on keyframes.
+    cpi->rc.next_frame_size_selector = UNSCALED;
+  }
+}
+
+// Define the reference buffers that will be updated post encode.
+static void configure_buffer_updates(VP9_COMP *cpi) {
+  TWO_PASS *const twopass = &cpi->twopass;
+
+  cpi->rc.is_src_frame_alt_ref = 0;
+  switch (twopass->gf_group.update_type[twopass->gf_group.index]) {
+    case KF_UPDATE:
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 1;
+      cpi->refresh_alt_ref_frame = 1;
+      break;
+    case LF_UPDATE:
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_alt_ref_frame = 0;
+      break;
+    case GF_UPDATE:
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 1;
+      cpi->refresh_alt_ref_frame = 0;
+      break;
+    case OVERLAY_UPDATE:
+      cpi->refresh_last_frame = 0;
+      cpi->refresh_golden_frame = 1;
+      cpi->refresh_alt_ref_frame = 0;
+      cpi->rc.is_src_frame_alt_ref = 1;
+      break;
+    case ARF_UPDATE:
+      cpi->refresh_last_frame = 0;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_alt_ref_frame = 1;
+      break;
+    default:
+      assert(0);
+      break;
+  }
+  if (is_two_pass_svc(cpi)) {
+    if (cpi->svc.temporal_layer_id > 0) {
+      cpi->refresh_last_frame = 0;
+      cpi->refresh_golden_frame = 0;
+    }
+    if (cpi->svc.layer_context[cpi->svc.spatial_layer_id].gold_ref_idx < 0)
+      cpi->refresh_golden_frame = 0;
+    if (cpi->alt_ref_source == NULL)
+      cpi->refresh_alt_ref_frame = 0;
+  }
+}
+
+static int is_skippable_frame(const VP9_COMP *cpi) {
+  // If the current frame does not have non-zero motion vector detected in the
+  // first  pass, and so do its previous and forward frames, then this frame
+  // can be skipped for partition check, and the partition size is assigned
+  // according to the variance
+  const SVC *const svc = &cpi->svc;
+  const TWO_PASS *const twopass = is_two_pass_svc(cpi) ?
+      &svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
+
+  return (!frame_is_intra_only(&cpi->common) &&
+    twopass->stats_in - 2 > twopass->stats_in_start &&
+    twopass->stats_in < twopass->stats_in_end &&
+    (twopass->stats_in - 1)->pcnt_inter - (twopass->stats_in - 1)->pcnt_motion
+    == 1 &&
+    (twopass->stats_in - 2)->pcnt_inter - (twopass->stats_in - 2)->pcnt_motion
+    == 1 &&
+    twopass->stats_in->pcnt_inter - twopass->stats_in->pcnt_motion == 1);
+}
+
+void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  TWO_PASS *const twopass = &cpi->twopass;
+  GF_GROUP *const gf_group = &twopass->gf_group;
+  FIRSTPASS_STATS this_frame;
+
+  int target_rate;
+  LAYER_CONTEXT *const lc = is_two_pass_svc(cpi) ?
+        &cpi->svc.layer_context[cpi->svc.spatial_layer_id] : 0;
+
+  if (!twopass->stats_in)
+    return;
+
+  // If this is an arf frame then we dont want to read the stats file or
+  // advance the input pointer as we already have what we need.
+  if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
+    int target_rate;
+    configure_buffer_updates(cpi);
+    target_rate = gf_group->bit_allocation[gf_group->index];
+    target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
+    rc->base_frame_target = target_rate;
+
+    cm->frame_type = INTER_FRAME;
+
+    if (lc != NULL) {
+      if (cpi->svc.spatial_layer_id == 0) {
+        lc->is_key_frame = 0;
+      } else {
+        lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
+
+        if (lc->is_key_frame)
+          cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
+      }
+    }
+
+    // Do the firstpass stats indicate that this frame is skippable for the
+    // partition search?
+    if (cpi->sf.allow_partition_search_skip &&
+        cpi->oxcf.pass == 2 && (!cpi->use_svc || is_two_pass_svc(cpi))) {
+      cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
+    }
+
+    return;
+  }
+
+  vpx_clear_system_state();
+
+  if (cpi->oxcf.rc_mode == VPX_Q) {
+    twopass->active_worst_quality = cpi->oxcf.cq_level;
+  } else if (cm->current_video_frame == 0 ||
+             (lc != NULL && lc->current_video_frame_in_layer == 0)) {
+    const int frames_left = (int)(twopass->total_stats.count -
+        ((lc != NULL) ? lc->current_video_frame_in_layer
+                      : cm->current_video_frame));
+    // Special case code for first frame.
+    const int section_target_bandwidth = (int)(twopass->bits_left /
+                                               frames_left);
+    const double section_length = twopass->total_left_stats.count;
+    const double section_error =
+      twopass->total_left_stats.coded_error / section_length;
+    const double section_intra_skip =
+      twopass->total_left_stats.intra_skip_pct / section_length;
+    const double section_inactive_zone =
+      (twopass->total_left_stats.inactive_zone_rows * 2) /
+      ((double)cm->mb_rows * section_length);
+    int tmp_q;
+
+    tmp_q = get_twopass_worst_quality(cpi, section_error,
+        section_intra_skip + section_inactive_zone, section_target_bandwidth);
+
+    twopass->active_worst_quality = tmp_q;
+    twopass->baseline_active_worst_quality = tmp_q;
+    rc->ni_av_qi = tmp_q;
+    rc->last_q[INTER_FRAME] = tmp_q;
+    rc->avg_q = vp9_convert_qindex_to_q(tmp_q, cm->bit_depth);
+    rc->avg_frame_qindex[INTER_FRAME] = tmp_q;
+    rc->last_q[KEY_FRAME] = (tmp_q + cpi->oxcf.best_allowed_q) / 2;
+    rc->avg_frame_qindex[KEY_FRAME] = rc->last_q[KEY_FRAME];
+  }
+  vp9_zero(this_frame);
+  if (EOF == input_stats(twopass, &this_frame))
+    return;
+
+  // Set the frame content type flag.
+  if (this_frame.intra_skip_pct >= FC_ANIMATION_THRESH)
+    twopass->fr_content_type = FC_GRAPHICS_ANIMATION;
+  else
+    twopass->fr_content_type = FC_NORMAL;
+
+  // Keyframe and section processing.
+  if (rc->frames_to_key == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY)) {
+    FIRSTPASS_STATS this_frame_copy;
+    this_frame_copy = this_frame;
+    // Define next KF group and assign bits to it.
+    find_next_key_frame(cpi, &this_frame);
+    this_frame = this_frame_copy;
+  } else {
+    cm->frame_type = INTER_FRAME;
+  }
+
+  if (lc != NULL) {
+    if (cpi->svc.spatial_layer_id == 0) {
+      lc->is_key_frame = (cm->frame_type == KEY_FRAME);
+      if (lc->is_key_frame) {
+        cpi->ref_frame_flags &=
+            (~VP9_LAST_FLAG & ~VP9_GOLD_FLAG & ~VP9_ALT_FLAG);
+        lc->frames_from_key_frame = 0;
+        // Encode an intra only empty frame since we have a key frame.
+        cpi->svc.encode_intra_empty_frame = 1;
+      }
+    } else {
+      cm->frame_type = INTER_FRAME;
+      lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
+
+      if (lc->is_key_frame) {
+        cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
+        lc->frames_from_key_frame = 0;
+      }
+    }
+  }
+
+  // Define a new GF/ARF group. (Should always enter here for key frames).
+  if (rc->frames_till_gf_update_due == 0) {
+    define_gf_group(cpi, &this_frame);
+
+    rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+    if (lc != NULL)
+      cpi->refresh_golden_frame = 1;
+
+#if ARF_STATS_OUTPUT
+    {
+      FILE *fpfile;
+      fpfile = fopen("arf.stt", "a");
+      ++arf_count;
+      fprintf(fpfile, "%10d %10ld %10d %10d %10ld\n",
+              cm->current_video_frame, rc->frames_till_gf_update_due,
+              rc->kf_boost, arf_count, rc->gfu_boost);
+
+      fclose(fpfile);
+    }
+#endif
+  }
+
+  configure_buffer_updates(cpi);
+
+  // Do the firstpass stats indicate that this frame is skippable for the
+  // partition search?
+  if (cpi->sf.allow_partition_search_skip && cpi->oxcf.pass == 2 &&
+      (!cpi->use_svc || is_two_pass_svc(cpi))) {
+    cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
+  }
+
+  target_rate = gf_group->bit_allocation[gf_group->index];
+  rc->base_frame_target = target_rate;
+
+  {
+    const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
+                        ? cpi->initial_mbs : cpi->common.MBs;
+    // The multiplication by 256 reverses a scaling factor of (>> 8)
+    // applied when combining MB error values for the frame.
+    twopass->mb_av_energy =
+      log(((this_frame.intra_error * 256.0) / num_mbs) + 1.0);
+    twopass->mb_smooth_pct = this_frame.intra_smooth_pct;
+  }
+
+  // Update the total stats remaining structure.
+  subtract_stats(&twopass->total_left_stats, &this_frame);
+}
+
+#define MINQ_ADJ_LIMIT 48
+#define MINQ_ADJ_LIMIT_CQ 20
+#define HIGH_UNDERSHOOT_RATIO 2
+void vp9_twopass_postencode_update(VP9_COMP *cpi) {
+  TWO_PASS *const twopass = &cpi->twopass;
+  RATE_CONTROL *const rc = &cpi->rc;
+  VP9_COMMON *const cm = &cpi->common;
+  const int bits_used = rc->base_frame_target;
+
+  // VBR correction is done through rc->vbr_bits_off_target. Based on the
+  // sign of this value, a limited % adjustment is made to the target rate
+  // of subsequent frames, to try and push it back towards 0. This method
+  // is designed to prevent extreme behaviour at the end of a clip
+  // or group of frames.
+  rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
+  twopass->bits_left = VPXMAX(twopass->bits_left - bits_used, 0);
+
+  // Target vs actual bits for this arf group.
+  twopass->rolling_arf_group_target_bits += rc->this_frame_target;
+  twopass->rolling_arf_group_actual_bits += rc->projected_frame_size;
+
+  // Calculate the pct rc error.
+  if (rc->total_actual_bits) {
+    rc->rate_error_estimate =
+      (int)((rc->vbr_bits_off_target * 100) / rc->total_actual_bits);
+    rc->rate_error_estimate = clamp(rc->rate_error_estimate, -100, 100);
+  } else {
+    rc->rate_error_estimate = 0;
+  }
+
+  if (cpi->common.frame_type != KEY_FRAME &&
+      !vp9_is_upper_layer_key_frame(cpi)) {
+    twopass->kf_group_bits -= bits_used;
+    twopass->last_kfgroup_zeromotion_pct = twopass->kf_zeromotion_pct;
+  }
+  twopass->kf_group_bits = VPXMAX(twopass->kf_group_bits, 0);
+
+  // Increment the gf group index ready for the next frame.
+  ++twopass->gf_group.index;
+
+  // If the rate control is drifting consider adjustment to min or maxq.
+  if ((cpi->oxcf.rc_mode != VPX_Q) &&
+      !cpi->rc.is_src_frame_alt_ref) {
+    const int maxq_adj_limit =
+      rc->worst_quality - twopass->active_worst_quality;
+    const int minq_adj_limit =
+        (cpi->oxcf.rc_mode == VPX_CQ ? MINQ_ADJ_LIMIT_CQ : MINQ_ADJ_LIMIT);
+    int aq_extend_min = 0;
+    int aq_extend_max = 0;
+
+    // Extend min or Max Q range to account for imbalance from the base
+    // value when using AQ.
+    if (cpi->oxcf.aq_mode != NO_AQ) {
+      if (cm->seg.aq_av_offset < 0) {
+        // The balance of the AQ map tends towarda lowering the average Q.
+        aq_extend_min = 0;
+        aq_extend_max = VPXMIN(maxq_adj_limit, -cm->seg.aq_av_offset);
+      } else {
+        // The balance of the AQ map tends towards raising the average Q.
+        aq_extend_min = VPXMIN(minq_adj_limit, cm->seg.aq_av_offset);
+        aq_extend_max = 0;
+      }
+    }
+
+    // Undershoot.
+    if (rc->rate_error_estimate > cpi->oxcf.under_shoot_pct) {
+      --twopass->extend_maxq;
+      if (rc->rolling_target_bits >= rc->rolling_actual_bits)
+        ++twopass->extend_minq;
+    // Overshoot.
+    } else if (rc->rate_error_estimate < -cpi->oxcf.over_shoot_pct) {
+      --twopass->extend_minq;
+      if (rc->rolling_target_bits < rc->rolling_actual_bits)
+        ++twopass->extend_maxq;
+    } else {
+      // Adjustment for extreme local overshoot.
+      if (rc->projected_frame_size > (2 * rc->base_frame_target) &&
+          rc->projected_frame_size > (2 * rc->avg_frame_bandwidth))
+        ++twopass->extend_maxq;
+
+      // Unwind undershoot or overshoot adjustment.
+      if (rc->rolling_target_bits < rc->rolling_actual_bits)
+        --twopass->extend_minq;
+      else if (rc->rolling_target_bits > rc->rolling_actual_bits)
+        --twopass->extend_maxq;
+    }
+
+    twopass->extend_minq =
+        clamp(twopass->extend_minq, aq_extend_min, minq_adj_limit);
+    twopass->extend_maxq =
+        clamp(twopass->extend_maxq, aq_extend_max, maxq_adj_limit);
+
+    // If there is a big and undexpected undershoot then feed the extra
+    // bits back in quickly. One situation where this may happen is if a
+    // frame is unexpectedly almost perfectly predicted by the ARF or GF
+    // but not very well predcited by the previous frame.
+    if (!frame_is_kf_gf_arf(cpi) && !cpi->rc.is_src_frame_alt_ref) {
+      int fast_extra_thresh = rc->base_frame_target / HIGH_UNDERSHOOT_RATIO;
+      if (rc->projected_frame_size < fast_extra_thresh) {
+        rc->vbr_bits_off_target_fast +=
+          fast_extra_thresh - rc->projected_frame_size;
+        rc->vbr_bits_off_target_fast =
+          VPXMIN(rc->vbr_bits_off_target_fast, (4 * rc->avg_frame_bandwidth));
+
+        // Fast adaptation of minQ if necessary to use up the extra bits.
+        if (rc->avg_frame_bandwidth) {
+          twopass->extend_minq_fast =
+            (int)(rc->vbr_bits_off_target_fast * 8 / rc->avg_frame_bandwidth);
+        }
+        twopass->extend_minq_fast = VPXMIN(
+            twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
+      } else if (rc->vbr_bits_off_target_fast) {
+        twopass->extend_minq_fast = VPXMIN(
+            twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
+      } else {
+        twopass->extend_minq_fast = 0;
+      }
+    }
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_firstpass.h b/libvpx/libvpx/vp9/encoder/vp9_firstpass.h
new file mode 100644
index 0000000..7607288
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_firstpass.h
@@ -0,0 +1,163 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_FIRSTPASS_H_
+#define VP9_ENCODER_VP9_FIRSTPASS_H_
+
+#include "vp9/encoder/vp9_lookahead.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_FP_MB_STATS
+
+#define FPMB_DCINTRA_MASK 0x01
+
+#define FPMB_MOTION_ZERO_MASK 0x02
+#define FPMB_MOTION_LEFT_MASK 0x04
+#define FPMB_MOTION_RIGHT_MASK 0x08
+#define FPMB_MOTION_UP_MASK 0x10
+#define FPMB_MOTION_DOWN_MASK 0x20
+
+#define FPMB_ERROR_SMALL_MASK 0x40
+#define FPMB_ERROR_LARGE_MASK 0x80
+#define FPMB_ERROR_SMALL_TH 2000
+#define FPMB_ERROR_LARGE_TH 48000
+
+typedef struct {
+  uint8_t *mb_stats_start;
+  uint8_t *mb_stats_end;
+} FIRSTPASS_MB_STATS;
+#endif
+
+typedef struct {
+  double frame;
+  double weight;
+  double intra_error;
+  double coded_error;
+  double sr_coded_error;
+  double pcnt_inter;
+  double pcnt_motion;
+  double pcnt_second_ref;
+  double pcnt_neutral;
+  double intra_skip_pct;
+  double intra_smooth_pct;    // % of blocks that are smooth
+  double inactive_zone_rows;  // Image mask rows top and bottom.
+  double inactive_zone_cols;  // Image mask columns at left and right edges.
+  double MVr;
+  double mvr_abs;
+  double MVc;
+  double mvc_abs;
+  double MVrv;
+  double MVcv;
+  double mv_in_out_count;
+  double new_mv_count;
+  double duration;
+  double count;
+  int64_t spatial_layer_id;
+} FIRSTPASS_STATS;
+
+typedef enum {
+  KF_UPDATE = 0,
+  LF_UPDATE = 1,
+  GF_UPDATE = 2,
+  ARF_UPDATE = 3,
+  OVERLAY_UPDATE = 4,
+  FRAME_UPDATE_TYPES = 5
+} FRAME_UPDATE_TYPE;
+
+#define FC_ANIMATION_THRESH 0.15
+typedef enum {
+  FC_NORMAL = 0,
+  FC_GRAPHICS_ANIMATION = 1,
+  FRAME_CONTENT_TYPES = 2
+} FRAME_CONTENT_TYPE;
+
+typedef struct {
+  unsigned char index;
+  RATE_FACTOR_LEVEL rf_level[(MAX_LAG_BUFFERS * 2) + 1];
+  FRAME_UPDATE_TYPE update_type[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_src_offset[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_update_idx[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_ref_idx[(MAX_LAG_BUFFERS * 2) + 1];
+  int bit_allocation[(MAX_LAG_BUFFERS * 2) + 1];
+} GF_GROUP;
+
+typedef struct {
+  unsigned int section_intra_rating;
+  FIRSTPASS_STATS total_stats;
+  FIRSTPASS_STATS this_frame_stats;
+  const FIRSTPASS_STATS *stats_in;
+  const FIRSTPASS_STATS *stats_in_start;
+  const FIRSTPASS_STATS *stats_in_end;
+  FIRSTPASS_STATS total_left_stats;
+  int first_pass_done;
+  int64_t bits_left;
+  double modified_error_min;
+  double modified_error_max;
+  double modified_error_left;
+  double mb_av_energy;
+  double mb_smooth_pct;
+
+#if CONFIG_FP_MB_STATS
+  uint8_t *frame_mb_stats_buf;
+  uint8_t *this_frame_mb_stats;
+  FIRSTPASS_MB_STATS firstpass_mb_stats;
+#endif
+  // An indication of the content type of the current frame
+  FRAME_CONTENT_TYPE fr_content_type;
+
+  // Projected total bits available for a key frame group of frames
+  int64_t kf_group_bits;
+
+  // Error score of frames still to be coded in kf group
+  int64_t kf_group_error_left;
+
+  double bpm_factor;
+  int rolling_arf_group_target_bits;
+  int rolling_arf_group_actual_bits;
+
+  int sr_update_lag;
+  int kf_zeromotion_pct;
+  int last_kfgroup_zeromotion_pct;
+  int active_worst_quality;
+  int baseline_active_worst_quality;
+  int extend_minq;
+  int extend_maxq;
+  int extend_minq_fast;
+
+  GF_GROUP gf_group;
+} TWO_PASS;
+
+struct VP9_COMP;
+
+void vp9_init_first_pass(struct VP9_COMP *cpi);
+void vp9_rc_get_first_pass_params(struct VP9_COMP *cpi);
+void vp9_first_pass(struct VP9_COMP *cpi, const struct lookahead_entry *source);
+void vp9_end_first_pass(struct VP9_COMP *cpi);
+
+void vp9_init_second_pass(struct VP9_COMP *cpi);
+void vp9_rc_get_second_pass_params(struct VP9_COMP *cpi);
+void vp9_twopass_postencode_update(struct VP9_COMP *cpi);
+
+// Post encode update of the rate control parameters for 2-pass
+void vp9_twopass_postencode_update(struct VP9_COMP *cpi);
+
+void calculate_coded_size(struct VP9_COMP *cpi,
+                          int *scaled_frame_width,
+                          int *scaled_frame_height);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_FIRSTPASS_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_lookahead.c b/libvpx/libvpx/vp9/encoder/vp9_lookahead.c
new file mode 100644
index 0000000..787bcf4
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_lookahead.c
@@ -0,0 +1,245 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <assert.h>
+#include <stdlib.h>
+
+#include "./vpx_config.h"
+
+#include "vp9/common/vp9_common.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_extend.h"
+#include "vp9/encoder/vp9_lookahead.h"
+
+/* Return the buffer at the given absolute index and increment the index */
+static struct lookahead_entry *pop(struct lookahead_ctx *ctx,
+                                   int *idx) {
+  int index = *idx;
+  struct lookahead_entry *buf = ctx->buf + index;
+
+  assert(index < ctx->max_sz);
+  if (++index >= ctx->max_sz)
+    index -= ctx->max_sz;
+  *idx = index;
+  return buf;
+}
+
+
+void vp9_lookahead_destroy(struct lookahead_ctx *ctx) {
+  if (ctx) {
+    if (ctx->buf) {
+      int i;
+
+      for (i = 0; i < ctx->max_sz; i++)
+        vpx_free_frame_buffer(&ctx->buf[i].img);
+      free(ctx->buf);
+    }
+    free(ctx);
+  }
+}
+
+
+struct lookahead_ctx *vp9_lookahead_init(unsigned int width,
+                                         unsigned int height,
+                                         unsigned int subsampling_x,
+                                         unsigned int subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                         int use_highbitdepth,
+#endif
+                                         unsigned int depth) {
+  struct lookahead_ctx *ctx = NULL;
+
+  // Clamp the lookahead queue depth
+  depth = clamp(depth, 1, MAX_LAG_BUFFERS);
+
+  // Allocate memory to keep previous source frames available.
+  depth += MAX_PRE_FRAMES;
+
+  // Allocate the lookahead structures
+  ctx = calloc(1, sizeof(*ctx));
+  if (ctx) {
+    const int legacy_byte_alignment = 0;
+    unsigned int i;
+    ctx->max_sz = depth;
+    ctx->buf = calloc(depth, sizeof(*ctx->buf));
+    if (!ctx->buf)
+      goto bail;
+    for (i = 0; i < depth; i++)
+      if (vpx_alloc_frame_buffer(&ctx->buf[i].img,
+                                 width, height, subsampling_x, subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 use_highbitdepth,
+#endif
+                                 VP9_ENC_BORDER_IN_PIXELS,
+                                 legacy_byte_alignment))
+        goto bail;
+  }
+  return ctx;
+ bail:
+  vp9_lookahead_destroy(ctx);
+  return NULL;
+}
+
+#define USE_PARTIAL_COPY 0
+
+int vp9_lookahead_push(struct lookahead_ctx *ctx, YV12_BUFFER_CONFIG *src,
+                       int64_t ts_start, int64_t ts_end,
+#if CONFIG_VP9_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       unsigned int flags) {
+  struct lookahead_entry *buf;
+#if USE_PARTIAL_COPY
+  int row, col, active_end;
+  int mb_rows = (src->y_height + 15) >> 4;
+  int mb_cols = (src->y_width + 15) >> 4;
+#endif
+  int width = src->y_crop_width;
+  int height = src->y_crop_height;
+  int uv_width = src->uv_crop_width;
+  int uv_height = src->uv_crop_height;
+  int subsampling_x = src->subsampling_x;
+  int subsampling_y = src->subsampling_y;
+  int larger_dimensions, new_dimensions;
+
+  if (ctx->sz + 1  + MAX_PRE_FRAMES > ctx->max_sz)
+    return 1;
+  ctx->sz++;
+  buf = pop(ctx, &ctx->write_idx);
+
+  new_dimensions = width != buf->img.y_crop_width ||
+                   height != buf->img.y_crop_height ||
+                   uv_width != buf->img.uv_crop_width ||
+                   uv_height != buf->img.uv_crop_height;
+  larger_dimensions = width > buf->img.y_width ||
+                      height > buf->img.y_height ||
+                      uv_width > buf->img.uv_width ||
+                      uv_height > buf->img.uv_height;
+  assert(!larger_dimensions || new_dimensions);
+
+#if USE_PARTIAL_COPY
+  // TODO(jkoleszar): This is disabled for now, as
+  // vp9_copy_and_extend_frame_with_rect is not subsampling/alpha aware.
+
+  // Only do this partial copy if the following conditions are all met:
+  // 1. Lookahead queue has has size of 1.
+  // 2. Active map is provided.
+  // 3. This is not a key frame, golden nor altref frame.
+  if (!new_dimensions && ctx->max_sz == 1 && active_map && !flags) {
+    for (row = 0; row < mb_rows; ++row) {
+      col = 0;
+
+      while (1) {
+        // Find the first active macroblock in this row.
+        for (; col < mb_cols; ++col) {
+          if (active_map[col])
+            break;
+        }
+
+        // No more active macroblock in this row.
+        if (col == mb_cols)
+          break;
+
+        // Find the end of active region in this row.
+        active_end = col;
+
+        for (; active_end < mb_cols; ++active_end) {
+          if (!active_map[active_end])
+            break;
+        }
+
+        // Only copy this active region.
+        vp9_copy_and_extend_frame_with_rect(src, &buf->img,
+                                            row << 4,
+                                            col << 4, 16,
+                                            (active_end - col) << 4);
+
+        // Start again from the end of this active region.
+        col = active_end;
+      }
+
+      active_map += mb_cols;
+    }
+  } else {
+#endif
+    if (larger_dimensions) {
+      YV12_BUFFER_CONFIG new_img;
+      memset(&new_img, 0, sizeof(new_img));
+      if (vpx_alloc_frame_buffer(&new_img,
+                                 width, height, subsampling_x, subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 use_highbitdepth,
+#endif
+                                 VP9_ENC_BORDER_IN_PIXELS,
+                                 0))
+          return 1;
+      vpx_free_frame_buffer(&buf->img);
+      buf->img = new_img;
+    } else if (new_dimensions) {
+      buf->img.y_crop_width = src->y_crop_width;
+      buf->img.y_crop_height = src->y_crop_height;
+      buf->img.uv_crop_width = src->uv_crop_width;
+      buf->img.uv_crop_height = src->uv_crop_height;
+      buf->img.subsampling_x = src->subsampling_x;
+      buf->img.subsampling_y = src->subsampling_y;
+    }
+    // Partial copy not implemented yet
+    vp9_copy_and_extend_frame(src, &buf->img);
+#if USE_PARTIAL_COPY
+  }
+#endif
+
+  buf->ts_start = ts_start;
+  buf->ts_end = ts_end;
+  buf->flags = flags;
+  return 0;
+}
+
+
+struct lookahead_entry *vp9_lookahead_pop(struct lookahead_ctx *ctx,
+                                          int drain) {
+  struct lookahead_entry *buf = NULL;
+
+  if (ctx && ctx->sz && (drain || ctx->sz == ctx->max_sz - MAX_PRE_FRAMES)) {
+    buf = pop(ctx, &ctx->read_idx);
+    ctx->sz--;
+  }
+  return buf;
+}
+
+
+struct lookahead_entry *vp9_lookahead_peek(struct lookahead_ctx *ctx,
+                                           int index) {
+  struct lookahead_entry *buf = NULL;
+
+  if (index >= 0) {
+    // Forward peek
+    if (index < ctx->sz) {
+      index += ctx->read_idx;
+      if (index >= ctx->max_sz)
+        index -= ctx->max_sz;
+      buf = ctx->buf + index;
+    }
+  } else if (index < 0) {
+    // Backward peek
+    if (-index <= MAX_PRE_FRAMES) {
+      index += ctx->read_idx;
+      if (index < 0)
+        index += ctx->max_sz;
+      buf = ctx->buf + index;
+    }
+  }
+
+  return buf;
+}
+
+unsigned int vp9_lookahead_depth(struct lookahead_ctx *ctx) {
+  return ctx->sz;
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_lookahead.h b/libvpx/libvpx/vp9/encoder/vp9_lookahead.h
new file mode 100644
index 0000000..db0fd1c
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_lookahead.h
@@ -0,0 +1,124 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_LOOKAHEAD_H_
+#define VP9_ENCODER_VP9_LOOKAHEAD_H_
+
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+
+#if CONFIG_SPATIAL_SVC
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_LAG_BUFFERS 25
+
+struct lookahead_entry {
+  YV12_BUFFER_CONFIG  img;
+  int64_t             ts_start;
+  int64_t             ts_end;
+  unsigned int        flags;
+};
+
+// The max of past frames we want to keep in the queue.
+#define MAX_PRE_FRAMES 1
+
+struct lookahead_ctx {
+  int max_sz;                  /* Absolute size of the queue */
+  int sz;                      /* Number of buffers currently in the queue */
+  int read_idx;                /* Read index */
+  int write_idx;               /* Write index */
+  struct lookahead_entry *buf; /* Buffer list */
+};
+
+/**\brief Initializes the lookahead stage
+ *
+ * The lookahead stage is a queue of frame buffers on which some analysis
+ * may be done when buffers are enqueued.
+ */
+struct lookahead_ctx *vp9_lookahead_init(unsigned int width,
+                                         unsigned int height,
+                                         unsigned int subsampling_x,
+                                         unsigned int subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                         int use_highbitdepth,
+#endif
+                                         unsigned int depth);
+
+
+/**\brief Destroys the lookahead stage
+ */
+void vp9_lookahead_destroy(struct lookahead_ctx *ctx);
+
+
+/**\brief Enqueue a source buffer
+ *
+ * This function will copy the source image into a new framebuffer with
+ * the expected stride/border.
+ *
+ * If active_map is non-NULL and there is only one frame in the queue, then copy
+ * only active macroblocks.
+ *
+ * \param[in] ctx         Pointer to the lookahead context
+ * \param[in] src         Pointer to the image to enqueue
+ * \param[in] ts_start    Timestamp for the start of this frame
+ * \param[in] ts_end      Timestamp for the end of this frame
+ * \param[in] flags       Flags set on this frame
+ * \param[in] active_map  Map that specifies which macroblock is active
+ */
+int vp9_lookahead_push(struct lookahead_ctx *ctx, YV12_BUFFER_CONFIG *src,
+                       int64_t ts_start, int64_t ts_end,
+#if CONFIG_VP9_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       unsigned int flags);
+
+
+/**\brief Get the next source buffer to encode
+ *
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ * \param[in] drain     Flag indicating the buffer should be drained
+ *                      (return a buffer regardless of the current queue depth)
+ *
+ * \retval NULL, if drain set and queue is empty
+ * \retval NULL, if drain not set and queue not of the configured depth
+ */
+struct lookahead_entry *vp9_lookahead_pop(struct lookahead_ctx *ctx,
+                                          int drain);
+
+
+/**\brief Get a future source buffer to encode
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ * \param[in] index     Index of the frame to be returned, 0 == next frame
+ *
+ * \retval NULL, if no buffer exists at the specified index
+ */
+struct lookahead_entry *vp9_lookahead_peek(struct lookahead_ctx *ctx,
+                                           int index);
+
+
+/**\brief Get the number of frames currently in the lookahead queue
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ */
+unsigned int vp9_lookahead_depth(struct lookahead_ctx *ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_LOOKAHEAD_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_mbgraph.c b/libvpx/libvpx/vp9/encoder/vp9_mbgraph.c
new file mode 100644
index 0000000..14a0b16
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_mbgraph.c
@@ -0,0 +1,428 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/system_state.h"
+#include "vp9/encoder/vp9_segmentation.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+
+
+static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
+                                              const MV *ref_mv,
+                                              MV *dst_mv,
+                                              int mb_row,
+                                              int mb_col) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
+  const SEARCH_METHODS old_search_method = mv_sf->search_method;
+  const vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
+
+  const int tmp_col_min = x->mv_col_min;
+  const int tmp_col_max = x->mv_col_max;
+  const int tmp_row_min = x->mv_row_min;
+  const int tmp_row_max = x->mv_row_max;
+  MV ref_full;
+  int cost_list[5];
+
+  // Further step/diamond searches as necessary
+  int step_param = mv_sf->reduce_first_step_size;
+  step_param = VPXMIN(step_param, MAX_MVSEARCH_STEPS - 2);
+
+  vp9_set_mv_search_range(x, ref_mv);
+
+  ref_full.col = ref_mv->col >> 3;
+  ref_full.row = ref_mv->row >> 3;
+
+  mv_sf->search_method = HEX;
+  vp9_full_pixel_search(cpi, x, BLOCK_16X16, &ref_full, step_param,
+                        x->errorperbit, cond_cost_list(cpi, cost_list), ref_mv,
+                        dst_mv, 0, 0);
+  mv_sf->search_method = old_search_method;
+
+  // Try sub-pixel MC
+  // if (bestsme > error_thresh && bestsme < INT_MAX)
+  {
+    uint32_t distortion;
+    uint32_t sse;
+    cpi->find_fractional_mv_step(
+        x, dst_mv, ref_mv, cpi->common.allow_high_precision_mv, x->errorperbit,
+        &v_fn_ptr, 0, mv_sf->subpel_iters_per_step,
+        cond_cost_list(cpi, cost_list),
+        NULL, NULL,
+        &distortion, &sse, NULL, 0, 0);
+  }
+
+  xd->mi[0]->mode = NEWMV;
+  xd->mi[0]->mv[0].as_mv = *dst_mv;
+
+  vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
+
+  /* restore UMV window */
+  x->mv_col_min = tmp_col_min;
+  x->mv_col_max = tmp_col_max;
+  x->mv_row_min = tmp_row_min;
+  x->mv_row_max = tmp_row_max;
+
+  return vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                      xd->plane[0].dst.buf, xd->plane[0].dst.stride);
+}
+
+static int do_16x16_motion_search(VP9_COMP *cpi, const MV *ref_mv,
+                                  int_mv *dst_mv, int mb_row, int mb_col) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  unsigned int err, tmp_err;
+  MV tmp_mv;
+
+  // Try zero MV first
+  // FIXME should really use something like near/nearest MV and/or MV prediction
+  err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
+  dst_mv->as_int = 0;
+
+  // Test last reference frame using the previous best mv as the
+  // starting point (best reference) for the search
+  tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv, mb_row, mb_col);
+  if (tmp_err < err) {
+    err = tmp_err;
+    dst_mv->as_mv = tmp_mv;
+  }
+
+  // If the current best reference mv is not centered on 0,0 then do a 0,0
+  // based search as well.
+  if (ref_mv->row != 0 || ref_mv->col != 0) {
+    unsigned int tmp_err;
+    MV zero_ref_mv = {0, 0}, tmp_mv;
+
+    tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, &tmp_mv,
+                                        mb_row, mb_col);
+    if (tmp_err < err) {
+      dst_mv->as_mv = tmp_mv;
+      err = tmp_err;
+    }
+  }
+
+  return err;
+}
+
+static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  unsigned int err;
+
+  // Try zero MV first
+  // FIXME should really use something like near/nearest MV and/or MV prediction
+  err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
+
+  dst_mv->as_int = 0;
+
+  return err;
+}
+static int find_best_16x16_intra(VP9_COMP *cpi, PREDICTION_MODE *pbest_mode) {
+  MACROBLOCK   *const x  = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  PREDICTION_MODE best_mode = -1, mode;
+  unsigned int best_err = INT_MAX;
+
+  // calculate SATD for each intra prediction mode;
+  // we're intentionally not doing 4x4, we just want a rough estimate
+  for (mode = DC_PRED; mode <= TM_PRED; mode++) {
+    unsigned int err;
+
+    xd->mi[0]->mode = mode;
+    vp9_predict_intra_block(xd, 2, TX_16X16, mode,
+                            x->plane[0].src.buf, x->plane[0].src.stride,
+                            xd->plane[0].dst.buf, xd->plane[0].dst.stride,
+                            0, 0, 0);
+    err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                       xd->plane[0].dst.buf, xd->plane[0].dst.stride);
+
+    // find best
+    if (err < best_err) {
+      best_err  = err;
+      best_mode = mode;
+    }
+  }
+
+  if (pbest_mode)
+    *pbest_mode = best_mode;
+
+  return best_err;
+}
+
+static void update_mbgraph_mb_stats
+(
+  VP9_COMP *cpi,
+  MBGRAPH_MB_STATS *stats,
+  YV12_BUFFER_CONFIG *buf,
+  int mb_y_offset,
+  YV12_BUFFER_CONFIG *golden_ref,
+  const MV *prev_golden_ref_mv,
+  YV12_BUFFER_CONFIG *alt_ref,
+  int mb_row,
+  int mb_col
+) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int intra_error;
+  VP9_COMMON *cm = &cpi->common;
+
+  // FIXME in practice we're completely ignoring chroma here
+  x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
+  x->plane[0].src.stride = buf->y_stride;
+
+  xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
+  xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
+
+  // do intra 16x16 prediction
+  intra_error = find_best_16x16_intra(cpi,
+                                      &stats->ref[INTRA_FRAME].m.mode);
+  if (intra_error <= 0)
+    intra_error = 1;
+  stats->ref[INTRA_FRAME].err = intra_error;
+
+  // Golden frame MV search, if it exists and is different than last frame
+  if (golden_ref) {
+    int g_motion_error;
+    xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
+    xd->plane[0].pre[0].stride = golden_ref->y_stride;
+    g_motion_error = do_16x16_motion_search(cpi,
+                                            prev_golden_ref_mv,
+                                            &stats->ref[GOLDEN_FRAME].m.mv,
+                                            mb_row, mb_col);
+    stats->ref[GOLDEN_FRAME].err = g_motion_error;
+  } else {
+    stats->ref[GOLDEN_FRAME].err = INT_MAX;
+    stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
+  }
+
+  // Do an Alt-ref frame MV search, if it exists and is different than
+  // last/golden frame.
+  if (alt_ref) {
+    int a_motion_error;
+    xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
+    xd->plane[0].pre[0].stride = alt_ref->y_stride;
+    a_motion_error = do_16x16_zerozero_search(cpi,
+                                              &stats->ref[ALTREF_FRAME].m.mv);
+
+    stats->ref[ALTREF_FRAME].err = a_motion_error;
+  } else {
+    stats->ref[ALTREF_FRAME].err = INT_MAX;
+    stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
+  }
+}
+
+static void update_mbgraph_frame_stats(VP9_COMP *cpi,
+                                       MBGRAPH_FRAME_STATS *stats,
+                                       YV12_BUFFER_CONFIG *buf,
+                                       YV12_BUFFER_CONFIG *golden_ref,
+                                       YV12_BUFFER_CONFIG *alt_ref) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  VP9_COMMON *const cm = &cpi->common;
+
+  int mb_col, mb_row, offset = 0;
+  int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
+  MV gld_top_mv = {0, 0};
+  MODE_INFO mi_local;
+  MODE_INFO mi_above, mi_left;
+
+  vp9_zero(mi_local);
+  // Set up limit values for motion vectors to prevent them extending outside
+  // the UMV borders.
+  x->mv_row_min     = -BORDER_MV_PIXELS_B16;
+  x->mv_row_max     = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
+  // Signal to vp9_predict_intra_block() that above is not available
+  xd->above_mi = NULL;
+
+  xd->plane[0].dst.stride  = buf->y_stride;
+  xd->plane[0].pre[0].stride  = buf->y_stride;
+  xd->plane[1].dst.stride = buf->uv_stride;
+  xd->mi[0] = &mi_local;
+  mi_local.sb_type = BLOCK_16X16;
+  mi_local.ref_frame[0] = LAST_FRAME;
+  mi_local.ref_frame[1] = NONE;
+
+  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
+    MV gld_left_mv = gld_top_mv;
+    int mb_y_in_offset  = mb_y_offset;
+    int arf_y_in_offset = arf_y_offset;
+    int gld_y_in_offset = gld_y_offset;
+
+    // Set up limit values for motion vectors to prevent them extending outside
+    // the UMV borders.
+    x->mv_col_min      = -BORDER_MV_PIXELS_B16;
+    x->mv_col_max      = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
+    // Signal to vp9_predict_intra_block() that left is not available
+    xd->left_mi = NULL;
+
+    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
+      MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];
+
+      update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset,
+                              golden_ref, &gld_left_mv, alt_ref,
+                              mb_row, mb_col);
+      gld_left_mv = mb_stats->ref[GOLDEN_FRAME].m.mv.as_mv;
+      if (mb_col == 0) {
+        gld_top_mv = gld_left_mv;
+      }
+      // Signal to vp9_predict_intra_block() that left is available
+      xd->left_mi = &mi_left;
+
+      mb_y_in_offset    += 16;
+      gld_y_in_offset   += 16;
+      arf_y_in_offset   += 16;
+      x->mv_col_min     -= 16;
+      x->mv_col_max     -= 16;
+    }
+
+    // Signal to vp9_predict_intra_block() that above is available
+    xd->above_mi = &mi_above;
+
+    mb_y_offset     += buf->y_stride * 16;
+    gld_y_offset    += golden_ref->y_stride * 16;
+    if (alt_ref)
+      arf_y_offset    += alt_ref->y_stride * 16;
+    x->mv_row_min   -= 16;
+    x->mv_row_max   -= 16;
+    offset          += cm->mb_cols;
+  }
+}
+
+// void separate_arf_mbs_byzz
+static void separate_arf_mbs(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  int mb_col, mb_row, offset, i;
+  int mi_row, mi_col;
+  int ncnt[4] = { 0 };
+  int n_frames = cpi->mbgraph_n_frames;
+
+  int *arf_not_zz;
+
+  CHECK_MEM_ERROR(cm, arf_not_zz,
+                  vpx_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz),
+                             1));
+
+  // We are not interested in results beyond the alt ref itself.
+  if (n_frames > cpi->rc.frames_till_gf_update_due)
+    n_frames = cpi->rc.frames_till_gf_update_due;
+
+  // defer cost to reference frames
+  for (i = n_frames - 1; i >= 0; i--) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+
+    for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
+         offset += cm->mb_cols, mb_row++) {
+      for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
+        MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];
+
+        int altref_err = mb_stats->ref[ALTREF_FRAME].err;
+        int intra_err  = mb_stats->ref[INTRA_FRAME ].err;
+        int golden_err = mb_stats->ref[GOLDEN_FRAME].err;
+
+        // Test for altref vs intra and gf and that its mv was 0,0.
+        if (altref_err > 1000 ||
+            altref_err > intra_err ||
+            altref_err > golden_err) {
+          arf_not_zz[offset + mb_col]++;
+        }
+      }
+    }
+  }
+
+  // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
+  // of bound access in segmentation_map
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
+      // If any of the blocks in the sequence failed then the MB
+      // goes in segment 0
+      if (arf_not_zz[mi_row / 2 * cm->mb_cols + mi_col / 2]) {
+        ncnt[0]++;
+        cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
+      } else {
+        cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
+        ncnt[1]++;
+      }
+    }
+  }
+
+  // Only bother with segmentation if over 10% of the MBs in static segment
+  // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
+  if (1) {
+    // Note % of blocks that are marked as static
+    if (cm->MBs)
+      cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);
+
+    // This error case should not be reachable as this function should
+    // never be called with the common data structure uninitialized.
+    else
+      cpi->static_mb_pct = 0;
+
+    vp9_enable_segmentation(&cm->seg);
+  } else {
+    cpi->static_mb_pct = 0;
+    vp9_disable_segmentation(&cm->seg);
+  }
+
+  // Free localy allocated storage
+  vpx_free(arf_not_zz);
+}
+
+void vp9_update_mbgraph_stats(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  int i, n_frames = vp9_lookahead_depth(cpi->lookahead);
+  YV12_BUFFER_CONFIG *golden_ref = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
+
+  assert(golden_ref != NULL);
+
+  // we need to look ahead beyond where the ARF transitions into
+  // being a GF - so exit if we don't look ahead beyond that
+  if (n_frames <= cpi->rc.frames_till_gf_update_due)
+    return;
+
+  if (n_frames > MAX_LAG_BUFFERS)
+    n_frames = MAX_LAG_BUFFERS;
+
+  cpi->mbgraph_n_frames = n_frames;
+  for (i = 0; i < n_frames; i++) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+    memset(frame_stats->mb_stats, 0,
+           cm->mb_rows * cm->mb_cols * sizeof(*cpi->mbgraph_stats[i].mb_stats));
+  }
+
+  // do motion search to find contribution of each reference to data
+  // later on in this GF group
+  // FIXME really, the GF/last MC search should be done forward, and
+  // the ARF MC search backwards, to get optimal results for MV caching
+  for (i = 0; i < n_frames; i++) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+    struct lookahead_entry *q_cur = vp9_lookahead_peek(cpi->lookahead, i);
+
+    assert(q_cur != NULL);
+
+    update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img,
+                               golden_ref, cpi->Source);
+  }
+
+  vpx_clear_system_state();
+
+  separate_arf_mbs(cpi);
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_mbgraph.h b/libvpx/libvpx/vp9/encoder/vp9_mbgraph.h
new file mode 100644
index 0000000..c3af972
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_mbgraph.h
@@ -0,0 +1,40 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_MBGRAPH_H_
+#define VP9_ENCODER_VP9_MBGRAPH_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  struct {
+    int err;
+    union {
+      int_mv mv;
+      PREDICTION_MODE mode;
+    } m;
+  } ref[MAX_REF_FRAMES];
+} MBGRAPH_MB_STATS;
+
+typedef struct {
+  MBGRAPH_MB_STATS *mb_stats;
+} MBGRAPH_FRAME_STATS;
+
+struct VP9_COMP;
+
+void vp9_update_mbgraph_stats(struct VP9_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_MBGRAPH_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_mcomp.c b/libvpx/libvpx/vp9/encoder/vp9_mcomp.c
new file mode 100644
index 0000000..2ebacc0
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_mcomp.c
@@ -0,0 +1,2578 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_reconinter.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_mcomp.h"
+
+// #define NEW_DIAMOND_SEARCH
+
+static INLINE const uint8_t *get_buf_from_mv(const struct buf_2d *buf,
+                                             const MV *mv) {
+  return &buf->buf[mv->row * buf->stride + mv->col];
+}
+
+void vp9_set_mv_search_range(MACROBLOCK *x, const MV *mv) {
+  int col_min = (mv->col >> 3) - MAX_FULL_PEL_VAL + (mv->col & 7 ? 1 : 0);
+  int row_min = (mv->row >> 3) - MAX_FULL_PEL_VAL + (mv->row & 7 ? 1 : 0);
+  int col_max = (mv->col >> 3) + MAX_FULL_PEL_VAL;
+  int row_max = (mv->row >> 3) + MAX_FULL_PEL_VAL;
+
+  col_min = VPXMAX(col_min, (MV_LOW >> 3) + 1);
+  row_min = VPXMAX(row_min, (MV_LOW >> 3) + 1);
+  col_max = VPXMIN(col_max, (MV_UPP >> 3) - 1);
+  row_max = VPXMIN(row_max, (MV_UPP >> 3) - 1);
+
+  // Get intersection of UMV window and valid MV window to reduce # of checks
+  // in diamond search.
+  if (x->mv_col_min < col_min)
+    x->mv_col_min = col_min;
+  if (x->mv_col_max > col_max)
+    x->mv_col_max = col_max;
+  if (x->mv_row_min < row_min)
+    x->mv_row_min = row_min;
+  if (x->mv_row_max > row_max)
+    x->mv_row_max = row_max;
+}
+
+int vp9_init_search_range(int size) {
+  int sr = 0;
+  // Minimum search size no matter what the passed in value.
+  size = VPXMAX(16, size);
+
+  while ((size << sr) < MAX_FULL_PEL_VAL)
+    sr++;
+
+  sr = VPXMIN(sr, MAX_MVSEARCH_STEPS - 2);
+  return sr;
+}
+
+static INLINE int mv_cost(const MV *mv,
+                          const int *joint_cost, int *const comp_cost[2]) {
+  assert(mv->row >= -MV_MAX && mv->row < MV_MAX);
+  assert(mv->col >= -MV_MAX && mv->col < MV_MAX);
+  return joint_cost[vp9_get_mv_joint(mv)] +
+             comp_cost[0][mv->row] + comp_cost[1][mv->col];
+}
+
+int vp9_mv_bit_cost(const MV *mv, const MV *ref,
+                    const int *mvjcost, int *mvcost[2], int weight) {
+  const MV diff = { mv->row - ref->row,
+                    mv->col - ref->col };
+  return ROUND_POWER_OF_TWO(mv_cost(&diff, mvjcost, mvcost) * weight, 7);
+}
+
+#define PIXEL_TRANSFORM_ERROR_SCALE 4
+static int mv_err_cost(const MV *mv, const MV *ref, const int *mvjcost,
+                       int *mvcost[2], int error_per_bit) {
+  if (mvcost) {
+    const MV diff = {mv->row - ref->row, mv->col - ref->col};
+    // This product sits at a 32-bit ceiling right now and any additional
+    // accuracy in either bit cost or error cost will cause it to overflow.
+    return ROUND_POWER_OF_TWO(
+        (unsigned)mv_cost(&diff, mvjcost, mvcost) * error_per_bit,
+        RDDIV_BITS + VP9_PROB_COST_SHIFT - RD_EPB_SHIFT +
+            PIXEL_TRANSFORM_ERROR_SCALE);
+  }
+  return 0;
+}
+
+static int mvsad_err_cost(const MACROBLOCK *x, const MV *mv, const MV *ref,
+                          int sad_per_bit) {
+  const MV diff = { mv->row - ref->row,
+                    mv->col - ref->col };
+  return ROUND_POWER_OF_TWO(
+      (unsigned)mv_cost(&diff, x->nmvjointsadcost, x->nmvsadcost) *
+          sad_per_bit,
+      VP9_PROB_COST_SHIFT);
+}
+
+void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride) {
+  int len;
+  int ss_count = 0;
+
+  for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
+    // Generate offsets for 4 search sites per step.
+    const MV ss_mvs[] = {{-len, 0}, {len, 0}, {0, -len}, {0, len}};
+    int i;
+    for (i = 0; i < 4; ++i, ++ss_count) {
+      cfg->ss_mv[ss_count] = ss_mvs[i];
+      cfg->ss_os[ss_count] = ss_mvs[i].row * stride + ss_mvs[i].col;
+    }
+  }
+
+  cfg->searches_per_step = 4;
+  cfg->total_steps = ss_count / cfg->searches_per_step;
+}
+
+void vp9_init3smotion_compensation(search_site_config *cfg, int stride) {
+  int len;
+  int ss_count = 0;
+
+  for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
+    // Generate offsets for 8 search sites per step.
+    const MV ss_mvs[8] = {
+      {-len,  0  }, {len,  0  }, { 0,   -len}, {0,    len},
+      {-len, -len}, {-len, len}, {len,  -len}, {len,  len}
+    };
+    int i;
+    for (i = 0; i < 8; ++i, ++ss_count) {
+      cfg->ss_mv[ss_count] = ss_mvs[i];
+      cfg->ss_os[ss_count] = ss_mvs[i].row * stride + ss_mvs[i].col;
+    }
+  }
+
+  cfg->searches_per_step = 8;
+  cfg->total_steps = ss_count / cfg->searches_per_step;
+}
+
+/* Estimated (square) error cost of a motion vector (r,c). The 14 scale comes
+ * from the same math as in mv_err_cost(). */
+#define MVC(r, c)                                              \
+    (mvcost ?                                                  \
+     ((unsigned)(mvjcost[((r) != rr) * 2 + ((c) != rc)] +      \
+       mvcost[0][((r) - rr)] + mvcost[1][((c) - rc)]) *        \
+      error_per_bit + 8192) >> 14 : 0)
+
+
+// convert motion vector component to offset for sv[a]f calc
+static INLINE int sp(int x) {
+  return x & 7;
+}
+
+static INLINE const uint8_t *pre(const uint8_t *buf, int stride, int r, int c) {
+  return &buf[(r >> 3) * stride + (c >> 3)];
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+/* checks if (r, c) has better score than previous best */
+#define CHECK_BETTER(v, r, c) \
+  if (c >= minc && c <= maxc && r >= minr && r <= maxr) {              \
+    int64_t tmpmse;                                                    \
+    if (second_pred == NULL) {                                         \
+      thismse = vfp->svf(pre(y, y_stride, r, c), y_stride, sp(c),      \
+                         sp(r), z, src_stride, &sse);                  \
+    } else {                                                           \
+      thismse = vfp->svaf(pre(y, y_stride, r, c), y_stride, sp(c),     \
+                          sp(r), z, src_stride, &sse, second_pred);    \
+    }                                                                  \
+    tmpmse = thismse;                                                  \
+    tmpmse += MVC(r, c);                                               \
+    if (tmpmse >= INT_MAX) {                                           \
+      v = INT_MAX;                                                     \
+    } else if ((v = (uint32_t)tmpmse) < besterr) {                     \
+      besterr = v;                                                     \
+      br = r;                                                          \
+      bc = c;                                                          \
+      *distortion = thismse;                                           \
+      *sse1 = sse;                                                     \
+    }                                                                  \
+  } else {                                                             \
+    v = INT_MAX;                                                       \
+  }
+#else
+/* checks if (r, c) has better score than previous best */
+#define CHECK_BETTER(v, r, c) \
+  if (c >= minc && c <= maxc && r >= minr && r <= maxr) {              \
+    if (second_pred == NULL)                                           \
+      thismse = vfp->svf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z, \
+                             src_stride, &sse);                        \
+    else                                                               \
+      thismse = vfp->svaf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), \
+                              z, src_stride, &sse, second_pred);       \
+    if ((v = MVC(r, c) + thismse) < besterr) {                         \
+      besterr = v;                                                     \
+      br = r;                                                          \
+      bc = c;                                                          \
+      *distortion = thismse;                                           \
+      *sse1 = sse;                                                     \
+    }                                                                  \
+  } else {                                                             \
+    v = INT_MAX;                                                       \
+  }
+
+#endif
+#define FIRST_LEVEL_CHECKS                              \
+  {                                                     \
+    unsigned int left, right, up, down, diag;           \
+    CHECK_BETTER(left, tr, tc - hstep);                 \
+    CHECK_BETTER(right, tr, tc + hstep);                \
+    CHECK_BETTER(up, tr - hstep, tc);                   \
+    CHECK_BETTER(down, tr + hstep, tc);                 \
+    whichdir = (left < right ? 0 : 1) +                 \
+               (up < down ? 0 : 2);                     \
+    switch (whichdir) {                                 \
+      case 0:                                           \
+        CHECK_BETTER(diag, tr - hstep, tc - hstep);     \
+        break;                                          \
+      case 1:                                           \
+        CHECK_BETTER(diag, tr - hstep, tc + hstep);     \
+        break;                                          \
+      case 2:                                           \
+        CHECK_BETTER(diag, tr + hstep, tc - hstep);     \
+        break;                                          \
+      case 3:                                           \
+        CHECK_BETTER(diag, tr + hstep, tc + hstep);     \
+        break;                                          \
+    }                                                   \
+  }
+
+#define SECOND_LEVEL_CHECKS                             \
+  {                                                     \
+    int kr, kc;                                         \
+    unsigned int second;                                \
+    if (tr != br && tc != bc) {                         \
+      kr = br - tr;                                     \
+      kc = bc - tc;                                     \
+      CHECK_BETTER(second, tr + kr, tc + 2 * kc);       \
+      CHECK_BETTER(second, tr + 2 * kr, tc + kc);       \
+    } else if (tr == br && tc != bc) {                  \
+      kc = bc - tc;                                     \
+      CHECK_BETTER(second, tr + hstep, tc + 2 * kc);    \
+      CHECK_BETTER(second, tr - hstep, tc + 2 * kc);    \
+      switch (whichdir) {                               \
+        case 0:                                         \
+        case 1:                                         \
+          CHECK_BETTER(second, tr + hstep, tc + kc);    \
+          break;                                        \
+        case 2:                                         \
+        case 3:                                         \
+          CHECK_BETTER(second, tr - hstep, tc + kc);    \
+          break;                                        \
+      }                                                 \
+    } else if (tr != br && tc == bc) {                  \
+      kr = br - tr;                                     \
+      CHECK_BETTER(second, tr + 2 * kr, tc + hstep);    \
+      CHECK_BETTER(second, tr + 2 * kr, tc - hstep);    \
+      switch (whichdir) {                               \
+        case 0:                                         \
+        case 2:                                         \
+          CHECK_BETTER(second, tr + kr, tc + hstep);    \
+          break;                                        \
+        case 1:                                         \
+        case 3:                                         \
+          CHECK_BETTER(second, tr + kr, tc - hstep);    \
+          break;                                        \
+      }                                                 \
+    }                                                   \
+  }
+
+// TODO(yunqingwang): SECOND_LEVEL_CHECKS_BEST was a rewrote of
+// SECOND_LEVEL_CHECKS, and SECOND_LEVEL_CHECKS should be rewritten
+// later in the same way.
+#define SECOND_LEVEL_CHECKS_BEST                        \
+  {                                                     \
+    unsigned int second;                                \
+    int br0 = br;                                       \
+    int bc0 = bc;                                       \
+    assert(tr == br || tc == bc);                       \
+    if (tr == br && tc != bc) {                         \
+      kc = bc - tc;                                     \
+    } else if (tr != br && tc == bc) {                  \
+      kr = br - tr;                                     \
+    }                                                   \
+    CHECK_BETTER(second, br0 + kr, bc0);                \
+    CHECK_BETTER(second, br0, bc0 + kc);                \
+    if (br0 != br || bc0 != bc) {                       \
+      CHECK_BETTER(second, br0 + kr, bc0 + kc);         \
+    }                                                   \
+  }
+
+#define SETUP_SUBPEL_SEARCH                                                \
+  const uint8_t *const z = x->plane[0].src.buf;                            \
+  const int src_stride = x->plane[0].src.stride;                           \
+  const MACROBLOCKD *xd = &x->e_mbd;                                       \
+  unsigned int besterr = INT_MAX;                                          \
+  unsigned int sse;                                                        \
+  unsigned int whichdir;                                                   \
+  int thismse;                                                             \
+  const unsigned int halfiters = iters_per_step;                           \
+  const unsigned int quarteriters = iters_per_step;                        \
+  const unsigned int eighthiters = iters_per_step;                         \
+  const int y_stride = xd->plane[0].pre[0].stride;                         \
+  const int offset = bestmv->row * y_stride + bestmv->col;                 \
+  const uint8_t *const y = xd->plane[0].pre[0].buf;                        \
+                                                                           \
+  int rr = ref_mv->row;                                                    \
+  int rc = ref_mv->col;                                                    \
+  int br = bestmv->row * 8;                                                \
+  int bc = bestmv->col * 8;                                                \
+  int hstep = 4;                                                           \
+  const int minc = VPXMAX(x->mv_col_min * 8, ref_mv->col - MV_MAX);        \
+  const int maxc = VPXMIN(x->mv_col_max * 8, ref_mv->col + MV_MAX);        \
+  const int minr = VPXMAX(x->mv_row_min * 8, ref_mv->row - MV_MAX);        \
+  const int maxr = VPXMIN(x->mv_row_max * 8, ref_mv->row + MV_MAX);        \
+  int tr = br;                                                             \
+  int tc = bc;                                                             \
+                                                                           \
+  bestmv->row *= 8;                                                        \
+  bestmv->col *= 8;
+
+static unsigned int setup_center_error(const MACROBLOCKD *xd,
+                                       const MV *bestmv,
+                                       const MV *ref_mv,
+                                       int error_per_bit,
+                                       const vp9_variance_fn_ptr_t *vfp,
+                                       const uint8_t *const src,
+                                       const int src_stride,
+                                       const uint8_t *const y,
+                                       int y_stride,
+                                       const uint8_t *second_pred,
+                                       int w, int h, int offset,
+                                       int *mvjcost, int *mvcost[2],
+                                       uint32_t *sse1,
+                                       uint32_t *distortion) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  uint64_t besterr;
+  if (second_pred != NULL) {
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      DECLARE_ALIGNED(16, uint16_t, comp_pred16[64 * 64]);
+      vpx_highbd_comp_avg_pred(comp_pred16, second_pred, w, h, y + offset,
+                               y_stride);
+      besterr = vfp->vf(CONVERT_TO_BYTEPTR(comp_pred16), w, src, src_stride,
+                        sse1);
+    } else {
+      DECLARE_ALIGNED(16, uint8_t, comp_pred[64 * 64]);
+      vpx_comp_avg_pred(comp_pred, second_pred, w, h, y + offset, y_stride);
+      besterr = vfp->vf(comp_pred, w, src, src_stride, sse1);
+    }
+  } else {
+    besterr = vfp->vf(y + offset, y_stride, src, src_stride, sse1);
+  }
+  *distortion = (uint32_t)besterr;
+  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
+  if (besterr >= UINT32_MAX)
+    return UINT32_MAX;
+  return (uint32_t)besterr;
+#else
+  uint32_t besterr;
+  (void) xd;
+  if (second_pred != NULL) {
+    DECLARE_ALIGNED(16, uint8_t, comp_pred[64 * 64]);
+    vpx_comp_avg_pred(comp_pred, second_pred, w, h, y + offset, y_stride);
+    besterr = vfp->vf(comp_pred, w, src, src_stride, sse1);
+  } else {
+    besterr = vfp->vf(y + offset, y_stride, src, src_stride, sse1);
+  }
+  *distortion = besterr;
+  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
+  return besterr;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+static INLINE int divide_and_round(const int n, const int d) {
+  return ((n < 0) ^ (d < 0)) ? ((n - d / 2) / d) : ((n + d / 2) / d);
+}
+
+static INLINE int is_cost_list_wellbehaved(int *cost_list) {
+  return cost_list[0] < cost_list[1] &&
+         cost_list[0] < cost_list[2] &&
+         cost_list[0] < cost_list[3] &&
+         cost_list[0] < cost_list[4];
+}
+
+// Returns surface minima estimate at given precision in 1/2^n bits.
+// Assume a model for the cost surface: S = A(x - x0)^2 + B(y - y0)^2 + C
+// For a given set of costs S0, S1, S2, S3, S4 at points
+// (y, x) = (0, 0), (0, -1), (1, 0), (0, 1) and (-1, 0) respectively,
+// the solution for the location of the minima (x0, y0) is given by:
+// x0 = 1/2 (S1 - S3)/(S1 + S3 - 2*S0),
+// y0 = 1/2 (S4 - S2)/(S4 + S2 - 2*S0).
+// The code below is an integerized version of that.
+static void get_cost_surf_min(int *cost_list, int *ir, int *ic,
+                              int bits) {
+  *ic = divide_and_round((cost_list[1] - cost_list[3]) * (1 << (bits - 1)),
+                         (cost_list[1] - 2 * cost_list[0] + cost_list[3]));
+  *ir = divide_and_round((cost_list[4] - cost_list[2]) * (1 << (bits - 1)),
+                         (cost_list[4] - 2 * cost_list[0] + cost_list[2]));
+}
+
+uint32_t vp9_skip_sub_pixel_tree(
+    const MACROBLOCK *x,
+    MV *bestmv, const MV *ref_mv,
+    int allow_hp,
+    int error_per_bit,
+    const vp9_variance_fn_ptr_t *vfp,
+    int forced_stop,
+    int iters_per_step,
+    int *cost_list,
+    int *mvjcost, int *mvcost[2],
+    uint32_t *distortion,
+    uint32_t *sse1,
+    const uint8_t *second_pred,
+    int w, int h) {
+  SETUP_SUBPEL_SEARCH;
+  besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
+                               z, src_stride, y, y_stride, second_pred,
+                               w, h, offset, mvjcost, mvcost,
+                               sse1, distortion);
+  (void) halfiters;
+  (void) quarteriters;
+  (void) eighthiters;
+  (void) whichdir;
+  (void) allow_hp;
+  (void) forced_stop;
+  (void) hstep;
+  (void) rr;
+  (void) rc;
+  (void) minr;
+  (void) minc;
+  (void) maxr;
+  (void) maxc;
+  (void) tr;
+  (void) tc;
+  (void) sse;
+  (void) thismse;
+  (void) cost_list;
+
+  if ((abs(bestmv->col - ref_mv->col) > (MAX_FULL_PEL_VAL << 3)) ||
+      (abs(bestmv->row - ref_mv->row) > (MAX_FULL_PEL_VAL << 3)))
+    return UINT32_MAX;
+
+  return besterr;
+}
+
+uint32_t vp9_find_best_sub_pixel_tree_pruned_evenmore(
+    const MACROBLOCK *x,
+    MV *bestmv, const MV *ref_mv,
+    int allow_hp,
+    int error_per_bit,
+    const vp9_variance_fn_ptr_t *vfp,
+    int forced_stop,
+    int iters_per_step,
+    int *cost_list,
+    int *mvjcost, int *mvcost[2],
+    uint32_t *distortion,
+    uint32_t *sse1,
+    const uint8_t *second_pred,
+    int w, int h) {
+  SETUP_SUBPEL_SEARCH;
+  besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
+                               z, src_stride, y, y_stride, second_pred,
+                               w, h, offset, mvjcost, mvcost,
+                               sse1, distortion);
+  (void) halfiters;
+  (void) quarteriters;
+  (void) eighthiters;
+  (void) whichdir;
+  (void) allow_hp;
+  (void) forced_stop;
+  (void) hstep;
+
+  if (cost_list &&
+      cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
+      cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
+      cost_list[4] != INT_MAX &&
+      is_cost_list_wellbehaved(cost_list)) {
+    int ir, ic;
+    unsigned int minpt;
+    get_cost_surf_min(cost_list, &ir, &ic, 2);
+    if (ir != 0 || ic != 0) {
+      CHECK_BETTER(minpt, tr + 2 * ir, tc + 2 * ic);
+    }
+  } else {
+    FIRST_LEVEL_CHECKS;
+    if (halfiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+
+    tr = br;
+    tc = bc;
+
+    // Each subsequent iteration checks at least one point in common with
+    // the last iteration could be 2 ( if diag selected) 1/4 pel
+    // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
+    if (forced_stop != 2) {
+      hstep >>= 1;
+      FIRST_LEVEL_CHECKS;
+      if (quarteriters > 1) {
+        SECOND_LEVEL_CHECKS;
+      }
+    }
+  }
+
+  tr = br;
+  tc = bc;
+
+  if (allow_hp && use_mv_hp(ref_mv) && forced_stop == 0) {
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (eighthiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+  }
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  if ((abs(bestmv->col - ref_mv->col) > (MAX_FULL_PEL_VAL << 3)) ||
+      (abs(bestmv->row - ref_mv->row) > (MAX_FULL_PEL_VAL << 3)))
+    return INT_MAX;
+
+  return besterr;
+}
+
+uint32_t vp9_find_best_sub_pixel_tree_pruned_more(const MACROBLOCK *x,
+                                             MV *bestmv, const MV *ref_mv,
+                                             int allow_hp,
+                                             int error_per_bit,
+                                             const vp9_variance_fn_ptr_t *vfp,
+                                             int forced_stop,
+                                             int iters_per_step,
+                                             int *cost_list,
+                                             int *mvjcost, int *mvcost[2],
+                                             uint32_t *distortion,
+                                             uint32_t *sse1,
+                                             const uint8_t *second_pred,
+                                             int w, int h) {
+  SETUP_SUBPEL_SEARCH;
+  besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
+                               z, src_stride, y, y_stride, second_pred,
+                               w, h, offset, mvjcost, mvcost,
+                               sse1, distortion);
+  if (cost_list &&
+      cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
+      cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
+      cost_list[4] != INT_MAX &&
+      is_cost_list_wellbehaved(cost_list)) {
+    unsigned int minpt;
+    int ir, ic;
+    get_cost_surf_min(cost_list, &ir, &ic, 1);
+    if (ir != 0 || ic != 0) {
+      CHECK_BETTER(minpt, tr + ir * hstep, tc + ic * hstep);
+    }
+  } else {
+    FIRST_LEVEL_CHECKS;
+    if (halfiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+  }
+
+  // Each subsequent iteration checks at least one point in common with
+  // the last iteration could be 2 ( if diag selected) 1/4 pel
+
+  // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
+  if (forced_stop != 2) {
+    tr = br;
+    tc = bc;
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (quarteriters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+  }
+
+  if (allow_hp && use_mv_hp(ref_mv) && forced_stop == 0) {
+    tr = br;
+    tc = bc;
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (eighthiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+  }
+  // These lines insure static analysis doesn't warn that
+  // tr and tc aren't used after the above point.
+  (void) tr;
+  (void) tc;
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  if ((abs(bestmv->col - ref_mv->col) > (MAX_FULL_PEL_VAL << 3)) ||
+      (abs(bestmv->row - ref_mv->row) > (MAX_FULL_PEL_VAL << 3)))
+    return UINT32_MAX;
+
+  return besterr;
+}
+
+uint32_t vp9_find_best_sub_pixel_tree_pruned(const MACROBLOCK *x,
+                                        MV *bestmv, const MV *ref_mv,
+                                        int allow_hp,
+                                        int error_per_bit,
+                                        const vp9_variance_fn_ptr_t *vfp,
+                                        int forced_stop,
+                                        int iters_per_step,
+                                        int *cost_list,
+                                        int *mvjcost, int *mvcost[2],
+                                        uint32_t *distortion,
+                                        uint32_t *sse1,
+                                        const uint8_t *second_pred,
+                                        int w, int h) {
+  SETUP_SUBPEL_SEARCH;
+  besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
+                               z, src_stride, y, y_stride, second_pred,
+                               w, h, offset, mvjcost, mvcost,
+                               sse1, distortion);
+  if (cost_list &&
+      cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
+      cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
+      cost_list[4] != INT_MAX) {
+    unsigned int left, right, up, down, diag;
+    whichdir = (cost_list[1] < cost_list[3] ? 0 : 1) +
+               (cost_list[2] < cost_list[4] ? 0 : 2);
+    switch (whichdir) {
+      case 0:
+        CHECK_BETTER(left, tr, tc - hstep);
+        CHECK_BETTER(down, tr + hstep, tc);
+        CHECK_BETTER(diag, tr + hstep, tc - hstep);
+        break;
+      case 1:
+        CHECK_BETTER(right, tr, tc + hstep);
+        CHECK_BETTER(down, tr + hstep, tc);
+        CHECK_BETTER(diag, tr + hstep, tc + hstep);
+        break;
+      case 2:
+        CHECK_BETTER(left, tr, tc - hstep);
+        CHECK_BETTER(up, tr - hstep, tc);
+        CHECK_BETTER(diag, tr - hstep, tc - hstep);
+        break;
+      case 3:
+        CHECK_BETTER(right, tr, tc + hstep);
+        CHECK_BETTER(up, tr - hstep, tc);
+        CHECK_BETTER(diag, tr - hstep, tc + hstep);
+        break;
+    }
+  } else {
+    FIRST_LEVEL_CHECKS;
+    if (halfiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+  }
+
+  tr = br;
+  tc = bc;
+
+  // Each subsequent iteration checks at least one point in common with
+  // the last iteration could be 2 ( if diag selected) 1/4 pel
+
+  // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
+  if (forced_stop != 2) {
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (quarteriters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+    tr = br;
+    tc = bc;
+  }
+
+  if (allow_hp && use_mv_hp(ref_mv) && forced_stop == 0) {
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (eighthiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+    tr = br;
+    tc = bc;
+  }
+  // These lines insure static analysis doesn't warn that
+  // tr and tc aren't used after the above point.
+  (void) tr;
+  (void) tc;
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  if ((abs(bestmv->col - ref_mv->col) > (MAX_FULL_PEL_VAL << 3)) ||
+      (abs(bestmv->row - ref_mv->row) > (MAX_FULL_PEL_VAL << 3)))
+    return INT_MAX;
+
+  return besterr;
+}
+
+static const MV search_step_table[12] = {
+    // left, right, up, down
+    {0, -4}, {0, 4}, {-4, 0}, {4, 0},
+    {0, -2}, {0, 2}, {-2, 0}, {2, 0},
+    {0, -1}, {0, 1}, {-1, 0}, {1, 0}
+};
+
+uint32_t vp9_find_best_sub_pixel_tree(const MACROBLOCK *x,
+                                      MV *bestmv, const MV *ref_mv,
+                                      int allow_hp,
+                                      int error_per_bit,
+                                      const vp9_variance_fn_ptr_t *vfp,
+                                      int forced_stop,
+                                      int iters_per_step,
+                                      int *cost_list,
+                                      int *mvjcost, int *mvcost[2],
+                                      uint32_t *distortion,
+                                      uint32_t *sse1,
+                                      const uint8_t *second_pred,
+                                      int w, int h) {
+  const uint8_t *const z = x->plane[0].src.buf;
+  const uint8_t *const src_address = z;
+  const int src_stride = x->plane[0].src.stride;
+  const MACROBLOCKD *xd = &x->e_mbd;
+  unsigned int besterr = INT_MAX;
+  unsigned int sse;
+  int thismse;
+  const int y_stride = xd->plane[0].pre[0].stride;
+  const int offset = bestmv->row * y_stride + bestmv->col;
+  const uint8_t *const y = xd->plane[0].pre[0].buf;
+
+  int rr = ref_mv->row;
+  int rc = ref_mv->col;
+  int br = bestmv->row * 8;
+  int bc = bestmv->col * 8;
+  int hstep = 4;
+  int iter, round = 3 - forced_stop;
+  const int minc = VPXMAX(x->mv_col_min * 8, ref_mv->col - MV_MAX);
+  const int maxc = VPXMIN(x->mv_col_max * 8, ref_mv->col + MV_MAX);
+  const int minr = VPXMAX(x->mv_row_min * 8, ref_mv->row - MV_MAX);
+  const int maxr = VPXMIN(x->mv_row_max * 8, ref_mv->row + MV_MAX);
+  int tr = br;
+  int tc = bc;
+  const MV *search_step = search_step_table;
+  int idx, best_idx = -1;
+  unsigned int cost_array[5];
+  int kr, kc;
+
+  if (!(allow_hp && use_mv_hp(ref_mv)))
+    if (round == 3)
+      round = 2;
+
+  bestmv->row *= 8;
+  bestmv->col *= 8;
+
+  besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
+                               z, src_stride, y, y_stride, second_pred,
+                               w, h, offset, mvjcost, mvcost,
+                               sse1, distortion);
+
+  (void) cost_list;  // to silence compiler warning
+
+  for (iter = 0; iter < round; ++iter) {
+    // Check vertical and horizontal sub-pixel positions.
+    for (idx = 0; idx < 4; ++idx) {
+      tr = br + search_step[idx].row;
+      tc = bc + search_step[idx].col;
+      if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
+        const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
+        MV this_mv;
+        this_mv.row = tr;
+        this_mv.col = tc;
+        if (second_pred == NULL)
+          thismse = vfp->svf(pre_address, y_stride, sp(tc), sp(tr),
+                             src_address, src_stride, &sse);
+        else
+          thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
+                              src_address, src_stride, &sse, second_pred);
+        cost_array[idx] = thismse +
+            mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit);
+
+        if (cost_array[idx] < besterr) {
+          best_idx = idx;
+          besterr = cost_array[idx];
+          *distortion = thismse;
+          *sse1 = sse;
+        }
+      } else {
+        cost_array[idx] = INT_MAX;
+      }
+    }
+
+    // Check diagonal sub-pixel position
+    kc = (cost_array[0] <= cost_array[1] ? -hstep : hstep);
+    kr = (cost_array[2] <= cost_array[3] ? -hstep : hstep);
+
+    tc = bc + kc;
+    tr = br + kr;
+    if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
+      const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
+      MV this_mv = {tr, tc};
+      if (second_pred == NULL)
+        thismse = vfp->svf(pre_address, y_stride, sp(tc), sp(tr),
+                           src_address, src_stride, &sse);
+      else
+        thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
+                            src_address, src_stride, &sse, second_pred);
+      cost_array[4] = thismse +
+          mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit);
+
+      if (cost_array[4] < besterr) {
+        best_idx = 4;
+        besterr = cost_array[4];
+        *distortion = thismse;
+        *sse1 = sse;
+      }
+    } else {
+      cost_array[idx] = INT_MAX;
+    }
+
+    if (best_idx < 4 && best_idx >= 0) {
+      br += search_step[best_idx].row;
+      bc += search_step[best_idx].col;
+    } else if (best_idx == 4) {
+      br = tr;
+      bc = tc;
+    }
+
+    if (iters_per_step > 1 && best_idx != -1)
+      SECOND_LEVEL_CHECKS_BEST;
+
+    tr = br;
+    tc = bc;
+
+    search_step += 4;
+    hstep >>= 1;
+    best_idx = -1;
+  }
+
+  // Each subsequent iteration checks at least one point in common with
+  // the last iteration could be 2 ( if diag selected) 1/4 pel
+
+  // These lines insure static analysis doesn't warn that
+  // tr and tc aren't used after the above point.
+  (void) tr;
+  (void) tc;
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  if ((abs(bestmv->col - ref_mv->col) > (MAX_FULL_PEL_VAL << 3)) ||
+      (abs(bestmv->row - ref_mv->row) > (MAX_FULL_PEL_VAL << 3)))
+    return INT_MAX;
+
+  return besterr;
+}
+
+#undef MVC
+#undef CHECK_BETTER
+
+static INLINE int check_bounds(const MACROBLOCK *x, int row, int col,
+                               int range) {
+  return ((row - range) >= x->mv_row_min) &
+         ((row + range) <= x->mv_row_max) &
+         ((col - range) >= x->mv_col_min) &
+         ((col + range) <= x->mv_col_max);
+}
+
+static INLINE int is_mv_in(const MACROBLOCK *x, const MV *mv) {
+  return (mv->col >= x->mv_col_min) && (mv->col <= x->mv_col_max) &&
+         (mv->row >= x->mv_row_min) && (mv->row <= x->mv_row_max);
+}
+
+#define CHECK_BETTER \
+  {\
+    if (thissad < bestsad) {\
+      if (use_mvcost) \
+        thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);\
+      if (thissad < bestsad) {\
+        bestsad = thissad;\
+        best_site = i;\
+      }\
+    }\
+  }
+
+#define MAX_PATTERN_SCALES         11
+#define MAX_PATTERN_CANDIDATES      8  // max number of canddiates per scale
+#define PATTERN_CANDIDATES_REF      3  // number of refinement candidates
+
+// Calculate and return a sad+mvcost list around an integer best pel.
+static INLINE void calc_int_cost_list(const MACROBLOCK *x,
+                                      const MV *ref_mv,
+                                      int sadpb,
+                                      const vp9_variance_fn_ptr_t *fn_ptr,
+                                      const MV *best_mv,
+                                      int *cost_list) {
+  static const MV neighbors[4] = {{0, -1}, {1, 0}, {0, 1}, {-1, 0}};
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &x->e_mbd.plane[0].pre[0];
+  const MV fcenter_mv = {ref_mv->row >> 3, ref_mv->col >> 3};
+  int br = best_mv->row;
+  int bc = best_mv->col;
+  MV this_mv;
+  int i;
+  unsigned int sse;
+
+  this_mv.row = br;
+  this_mv.col = bc;
+  cost_list[0] = fn_ptr->vf(what->buf, what->stride,
+                            get_buf_from_mv(in_what, &this_mv),
+                            in_what->stride, &sse) +
+      mvsad_err_cost(x, &this_mv, &fcenter_mv, sadpb);
+  if (check_bounds(x, br, bc, 1)) {
+    for (i = 0; i < 4; i++) {
+      const MV this_mv = {br + neighbors[i].row,
+        bc + neighbors[i].col};
+      cost_list[i + 1] = fn_ptr->vf(what->buf, what->stride,
+                                    get_buf_from_mv(in_what, &this_mv),
+                                    in_what->stride, &sse) +
+                                    mv_err_cost(&this_mv, &fcenter_mv,
+                                                x->nmvjointcost, x->mvcost,
+                                                x->errorperbit);
+    }
+  } else {
+    for (i = 0; i < 4; i++) {
+      const MV this_mv = {br + neighbors[i].row,
+        bc + neighbors[i].col};
+      if (!is_mv_in(x, &this_mv))
+        cost_list[i + 1] = INT_MAX;
+      else
+        cost_list[i + 1] = fn_ptr->vf(what->buf, what->stride,
+                                      get_buf_from_mv(in_what, &this_mv),
+                                      in_what->stride, &sse) +
+                                      mv_err_cost(&this_mv, &fcenter_mv,
+                                                  x->nmvjointcost, x->mvcost,
+                                                  x->errorperbit);
+    }
+  }
+}
+
+// Generic pattern search function that searches over multiple scales.
+// Each scale can have a different number of candidates and shape of
+// candidates as indicated in the num_candidates and candidates arrays
+// passed into this function
+//
+static int vp9_pattern_search(const MACROBLOCK *x,
+                              MV *ref_mv,
+                              int search_param,
+                              int sad_per_bit,
+                              int do_init_search,
+                              int *cost_list,
+                              const vp9_variance_fn_ptr_t *vfp,
+                              int use_mvcost,
+                              const MV *center_mv,
+                              MV *best_mv,
+                              const int num_candidates[MAX_PATTERN_SCALES],
+                              const MV candidates[MAX_PATTERN_SCALES]
+                                                 [MAX_PATTERN_CANDIDATES]) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  static const int search_param_to_steps[MAX_MVSEARCH_STEPS] = {
+    10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
+  };
+  int i, s, t;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  int br, bc;
+  int bestsad = INT_MAX;
+  int thissad;
+  int k = -1;
+  const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+  int best_init_s = search_param_to_steps[search_param];
+  // adjust ref_mv to make sure it is within MV range
+  clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
+  br = ref_mv->row;
+  bc = ref_mv->col;
+
+  // Work out the start point for the search
+  bestsad = vfp->sdf(what->buf, what->stride,
+                     get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+
+  // Search all possible scales upto the search param around the center point
+  // pick the scale of the point that is best as the starting scale of
+  // further steps around it.
+  if (do_init_search) {
+    s = best_init_s;
+    best_init_s = -1;
+    for (t = 0; t <= s; ++t) {
+      int best_site = -1;
+      if (check_bounds(x, br, bc, 1 << t)) {
+        for (i = 0; i < num_candidates[t]; i++) {
+          const MV this_mv = {br + candidates[t][i].row,
+                              bc + candidates[t][i].col};
+          thissad = vfp->sdf(what->buf, what->stride,
+                             get_buf_from_mv(in_what, &this_mv),
+                             in_what->stride);
+          CHECK_BETTER
+        }
+      } else {
+        for (i = 0; i < num_candidates[t]; i++) {
+          const MV this_mv = {br + candidates[t][i].row,
+                              bc + candidates[t][i].col};
+          if (!is_mv_in(x, &this_mv))
+            continue;
+          thissad = vfp->sdf(what->buf, what->stride,
+                             get_buf_from_mv(in_what, &this_mv),
+                             in_what->stride);
+          CHECK_BETTER
+        }
+      }
+      if (best_site == -1) {
+        continue;
+      } else {
+        best_init_s = t;
+        k = best_site;
+      }
+    }
+    if (best_init_s != -1) {
+      br += candidates[best_init_s][k].row;
+      bc += candidates[best_init_s][k].col;
+    }
+  }
+
+  // If the center point is still the best, just skip this and move to
+  // the refinement step.
+  if (best_init_s != -1) {
+    int best_site = -1;
+    s = best_init_s;
+
+    do {
+      // No need to search all 6 points the 1st time if initial search was used
+      if (!do_init_search || s != best_init_s) {
+        if (check_bounds(x, br, bc, 1 << s)) {
+          for (i = 0; i < num_candidates[s]; i++) {
+            const MV this_mv = {br + candidates[s][i].row,
+                                bc + candidates[s][i].col};
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        } else {
+          for (i = 0; i < num_candidates[s]; i++) {
+            const MV this_mv = {br + candidates[s][i].row,
+                                bc + candidates[s][i].col};
+            if (!is_mv_in(x, &this_mv))
+              continue;
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        }
+
+        if (best_site == -1) {
+          continue;
+        } else {
+          br += candidates[s][best_site].row;
+          bc += candidates[s][best_site].col;
+          k = best_site;
+        }
+      }
+
+      do {
+        int next_chkpts_indices[PATTERN_CANDIDATES_REF];
+        best_site = -1;
+        next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
+        next_chkpts_indices[1] = k;
+        next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
+
+        if (check_bounds(x, br, bc, 1 << s)) {
+          for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
+            const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,
+                                bc + candidates[s][next_chkpts_indices[i]].col};
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        } else {
+          for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
+            const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,
+                                bc + candidates[s][next_chkpts_indices[i]].col};
+            if (!is_mv_in(x, &this_mv))
+              continue;
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        }
+
+        if (best_site != -1) {
+          k = next_chkpts_indices[best_site];
+          br += candidates[s][k].row;
+          bc += candidates[s][k].col;
+        }
+      } while (best_site != -1);
+    } while (s--);
+  }
+
+  // Returns the one-away integer pel sad values around the best as follows:
+  // cost_list[0]: cost at the best integer pel
+  // cost_list[1]: cost at delta {0, -1} (left)   from the best integer pel
+  // cost_list[2]: cost at delta { 1, 0} (bottom) from the best integer pel
+  // cost_list[3]: cost at delta { 0, 1} (right)  from the best integer pel
+  // cost_list[4]: cost at delta {-1, 0} (top)    from the best integer pel
+  if (cost_list) {
+    const MV best_mv = { br, bc };
+    calc_int_cost_list(x, &fcenter_mv, sad_per_bit, vfp, &best_mv, cost_list);
+  }
+  best_mv->row = br;
+  best_mv->col = bc;
+  return bestsad;
+}
+
+// A specialized function where the smallest scale search candidates
+// are 4 1-away neighbors, and cost_list is non-null
+// TODO(debargha): Merge this function with the one above. Also remove
+// use_mvcost option since it is always 1, to save unnecessary branches.
+static int vp9_pattern_search_sad(const MACROBLOCK *x,
+                                  MV *ref_mv,
+                                  int search_param,
+                                  int sad_per_bit,
+                                  int do_init_search,
+                                  int *cost_list,
+                                  const vp9_variance_fn_ptr_t *vfp,
+                                  int use_mvcost,
+                                  const MV *center_mv,
+                                  MV *best_mv,
+                                  const int num_candidates[MAX_PATTERN_SCALES],
+                                  const MV candidates[MAX_PATTERN_SCALES]
+                                                     [MAX_PATTERN_CANDIDATES]) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  static const int search_param_to_steps[MAX_MVSEARCH_STEPS] = {
+    10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
+  };
+  int i, s, t;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  int br, bc;
+  int bestsad = INT_MAX;
+  int thissad;
+  int k = -1;
+  const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+  int best_init_s = search_param_to_steps[search_param];
+  // adjust ref_mv to make sure it is within MV range
+  clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
+  br = ref_mv->row;
+  bc = ref_mv->col;
+  if (cost_list != NULL) {
+    cost_list[0] = cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] =
+        INT_MAX;
+  }
+
+  // Work out the start point for the search
+  bestsad = vfp->sdf(what->buf, what->stride,
+                     get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+
+  // Search all possible scales upto the search param around the center point
+  // pick the scale of the point that is best as the starting scale of
+  // further steps around it.
+  if (do_init_search) {
+    s = best_init_s;
+    best_init_s = -1;
+    for (t = 0; t <= s; ++t) {
+      int best_site = -1;
+      if (check_bounds(x, br, bc, 1 << t)) {
+        for (i = 0; i < num_candidates[t]; i++) {
+          const MV this_mv = {br + candidates[t][i].row,
+                              bc + candidates[t][i].col};
+          thissad = vfp->sdf(what->buf, what->stride,
+                             get_buf_from_mv(in_what, &this_mv),
+                             in_what->stride);
+          CHECK_BETTER
+        }
+      } else {
+        for (i = 0; i < num_candidates[t]; i++) {
+          const MV this_mv = {br + candidates[t][i].row,
+                              bc + candidates[t][i].col};
+          if (!is_mv_in(x, &this_mv))
+            continue;
+          thissad = vfp->sdf(what->buf, what->stride,
+                             get_buf_from_mv(in_what, &this_mv),
+                             in_what->stride);
+          CHECK_BETTER
+        }
+      }
+      if (best_site == -1) {
+        continue;
+      } else {
+        best_init_s = t;
+        k = best_site;
+      }
+    }
+    if (best_init_s != -1) {
+      br += candidates[best_init_s][k].row;
+      bc += candidates[best_init_s][k].col;
+    }
+  }
+
+  // If the center point is still the best, just skip this and move to
+  // the refinement step.
+  if (best_init_s != -1) {
+    int do_sad = (num_candidates[0] == 4 && cost_list != NULL);
+    int best_site = -1;
+    s = best_init_s;
+
+    for (; s >= do_sad; s--) {
+      if (!do_init_search || s != best_init_s) {
+        if (check_bounds(x, br, bc, 1 << s)) {
+          for (i = 0; i < num_candidates[s]; i++) {
+            const MV this_mv = {br + candidates[s][i].row,
+                                bc + candidates[s][i].col};
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        } else {
+          for (i = 0; i < num_candidates[s]; i++) {
+            const MV this_mv = {br + candidates[s][i].row,
+                                bc + candidates[s][i].col};
+            if (!is_mv_in(x, &this_mv))
+              continue;
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        }
+
+        if (best_site == -1) {
+          continue;
+        } else {
+          br += candidates[s][best_site].row;
+          bc += candidates[s][best_site].col;
+          k = best_site;
+        }
+      }
+
+      do {
+        int next_chkpts_indices[PATTERN_CANDIDATES_REF];
+        best_site = -1;
+        next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
+        next_chkpts_indices[1] = k;
+        next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
+
+        if (check_bounds(x, br, bc, 1 << s)) {
+          for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
+            const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,
+                                bc + candidates[s][next_chkpts_indices[i]].col};
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        } else {
+          for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
+            const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,
+                                bc + candidates[s][next_chkpts_indices[i]].col};
+            if (!is_mv_in(x, &this_mv))
+              continue;
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        }
+
+        if (best_site != -1) {
+          k = next_chkpts_indices[best_site];
+          br += candidates[s][k].row;
+          bc += candidates[s][k].col;
+        }
+      } while (best_site != -1);
+    }
+
+    // Note: If we enter the if below, then cost_list must be non-NULL.
+    if (s == 0) {
+      cost_list[0] = bestsad;
+      if (!do_init_search || s != best_init_s) {
+        if (check_bounds(x, br, bc, 1 << s)) {
+          for (i = 0; i < num_candidates[s]; i++) {
+            const MV this_mv = {br + candidates[s][i].row,
+                                bc + candidates[s][i].col};
+            cost_list[i + 1] =
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        } else {
+          for (i = 0; i < num_candidates[s]; i++) {
+            const MV this_mv = {br + candidates[s][i].row,
+                                bc + candidates[s][i].col};
+            if (!is_mv_in(x, &this_mv))
+              continue;
+            cost_list[i + 1] =
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        }
+
+        if (best_site != -1) {
+          br += candidates[s][best_site].row;
+          bc += candidates[s][best_site].col;
+          k = best_site;
+        }
+      }
+      while (best_site != -1) {
+        int next_chkpts_indices[PATTERN_CANDIDATES_REF];
+        best_site = -1;
+        next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
+        next_chkpts_indices[1] = k;
+        next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
+        cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] = INT_MAX;
+        cost_list[((k + 2) % 4) + 1] = cost_list[0];
+        cost_list[0] = bestsad;
+
+        if (check_bounds(x, br, bc, 1 << s)) {
+          for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
+            const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,
+                                bc + candidates[s][next_chkpts_indices[i]].col};
+            cost_list[next_chkpts_indices[i] + 1] =
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        } else {
+          for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
+            const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,
+                                bc + candidates[s][next_chkpts_indices[i]].col};
+            if (!is_mv_in(x, &this_mv)) {
+              cost_list[next_chkpts_indices[i] + 1] = INT_MAX;
+              continue;
+            }
+            cost_list[next_chkpts_indices[i] + 1] =
+            thissad = vfp->sdf(what->buf, what->stride,
+                               get_buf_from_mv(in_what, &this_mv),
+                               in_what->stride);
+            CHECK_BETTER
+          }
+        }
+
+        if (best_site != -1) {
+          k = next_chkpts_indices[best_site];
+          br += candidates[s][k].row;
+          bc += candidates[s][k].col;
+        }
+      }
+    }
+  }
+
+  // Returns the one-away integer pel sad values around the best as follows:
+  // cost_list[0]: sad at the best integer pel
+  // cost_list[1]: sad at delta {0, -1} (left)   from the best integer pel
+  // cost_list[2]: sad at delta { 1, 0} (bottom) from the best integer pel
+  // cost_list[3]: sad at delta { 0, 1} (right)  from the best integer pel
+  // cost_list[4]: sad at delta {-1, 0} (top)    from the best integer pel
+  if (cost_list) {
+    static const MV neighbors[4] = {{0, -1}, {1, 0}, {0, 1}, {-1, 0}};
+    if (cost_list[0] == INT_MAX) {
+      cost_list[0] = bestsad;
+      if (check_bounds(x, br, bc, 1)) {
+        for (i = 0; i < 4; i++) {
+          const MV this_mv = { br + neighbors[i].row,
+                               bc + neighbors[i].col };
+          cost_list[i + 1] = vfp->sdf(what->buf, what->stride,
+                                     get_buf_from_mv(in_what, &this_mv),
+                                     in_what->stride);
+        }
+      } else {
+        for (i = 0; i < 4; i++) {
+          const MV this_mv = {br + neighbors[i].row,
+            bc + neighbors[i].col};
+          if (!is_mv_in(x, &this_mv))
+            cost_list[i + 1] = INT_MAX;
+          else
+            cost_list[i + 1] = vfp->sdf(what->buf, what->stride,
+                                       get_buf_from_mv(in_what, &this_mv),
+                                       in_what->stride);
+        }
+      }
+    } else {
+      if (use_mvcost) {
+        for (i = 0; i < 4; i++) {
+          const MV this_mv = {br + neighbors[i].row,
+            bc + neighbors[i].col};
+          if (cost_list[i + 1] != INT_MAX) {
+            cost_list[i + 1] +=
+                mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
+          }
+        }
+      }
+    }
+  }
+  best_mv->row = br;
+  best_mv->col = bc;
+  return bestsad;
+}
+
+int vp9_get_mvpred_var(const MACROBLOCK *x,
+                       const MV *best_mv, const MV *center_mv,
+                       const vp9_variance_fn_ptr_t *vfp,
+                       int use_mvcost) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const MV mv = {best_mv->row * 8, best_mv->col * 8};
+  uint32_t unused;
+#if CONFIG_VP9_HIGHBITDEPTH
+  uint64_t err= vfp->vf(what->buf, what->stride,
+                        get_buf_from_mv(in_what, best_mv),
+                        in_what->stride, &unused);
+  err += (use_mvcost ?  mv_err_cost(&mv, center_mv, x->nmvjointcost,
+                                    x->mvcost, x->errorperbit) : 0);
+  if (err >= INT_MAX)
+    return INT_MAX;
+  return (int)err;
+#else
+  return vfp->vf(what->buf, what->stride,
+                 get_buf_from_mv(in_what, best_mv), in_what->stride, &unused) +
+      (use_mvcost ?  mv_err_cost(&mv, center_mv, x->nmvjointcost,
+                                 x->mvcost, x->errorperbit) : 0);
+#endif
+}
+
+int vp9_get_mvpred_av_var(const MACROBLOCK *x,
+                          const MV *best_mv, const MV *center_mv,
+                          const uint8_t *second_pred,
+                          const vp9_variance_fn_ptr_t *vfp,
+                          int use_mvcost) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const MV mv = {best_mv->row * 8, best_mv->col * 8};
+  unsigned int unused;
+
+  return vfp->svaf(get_buf_from_mv(in_what, best_mv), in_what->stride, 0, 0,
+                   what->buf, what->stride, &unused, second_pred) +
+      (use_mvcost ?  mv_err_cost(&mv, center_mv, x->nmvjointcost,
+                                 x->mvcost, x->errorperbit) : 0);
+}
+
+static int hex_search(const MACROBLOCK *x,
+                      MV *ref_mv,
+                      int search_param,
+                      int sad_per_bit,
+                      int do_init_search,
+                      int *cost_list,
+                      const vp9_variance_fn_ptr_t *vfp,
+                      int use_mvcost,
+                      const MV *center_mv, MV *best_mv) {
+  // First scale has 8-closest points, the rest have 6 points in hex shape
+  // at increasing scales
+  static const int hex_num_candidates[MAX_PATTERN_SCALES] = {
+    8, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6
+  };
+  // Note that the largest candidate step at each scale is 2^scale
+  static const MV hex_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = {
+    {{-1, -1}, {0, -1}, {1, -1}, {1, 0}, {1, 1}, { 0, 1}, { -1, 1}, {-1, 0}},
+    {{-1, -2}, {1, -2}, {2, 0}, {1, 2}, { -1, 2}, { -2, 0}},
+    {{-2, -4}, {2, -4}, {4, 0}, {2, 4}, { -2, 4}, { -4, 0}},
+    {{-4, -8}, {4, -8}, {8, 0}, {4, 8}, { -4, 8}, { -8, 0}},
+    {{-8, -16}, {8, -16}, {16, 0}, {8, 16}, { -8, 16}, { -16, 0}},
+    {{-16, -32}, {16, -32}, {32, 0}, {16, 32}, { -16, 32}, { -32, 0}},
+    {{-32, -64}, {32, -64}, {64, 0}, {32, 64}, { -32, 64}, { -64, 0}},
+    {{-64, -128}, {64, -128}, {128, 0}, {64, 128}, { -64, 128}, { -128, 0}},
+    {{-128, -256}, {128, -256}, {256, 0}, {128, 256}, { -128, 256}, { -256, 0}},
+    {{-256, -512}, {256, -512}, {512, 0}, {256, 512}, { -256, 512}, { -512, 0}},
+    {{-512, -1024}, {512, -1024}, {1024, 0}, {512, 1024}, { -512, 1024},
+      { -1024, 0}},
+  };
+  return vp9_pattern_search(x, ref_mv, search_param, sad_per_bit,
+                            do_init_search, cost_list, vfp, use_mvcost,
+                            center_mv, best_mv,
+                            hex_num_candidates, hex_candidates);
+}
+
+static int bigdia_search(const MACROBLOCK *x,
+                         MV *ref_mv,
+                         int search_param,
+                         int sad_per_bit,
+                         int do_init_search,
+                         int *cost_list,
+                         const vp9_variance_fn_ptr_t *vfp,
+                         int use_mvcost,
+                         const MV *center_mv,
+                         MV *best_mv) {
+  // First scale has 4-closest points, the rest have 8 points in diamond
+  // shape at increasing scales
+  static const int bigdia_num_candidates[MAX_PATTERN_SCALES] = {
+    4, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  };
+  // Note that the largest candidate step at each scale is 2^scale
+  static const MV bigdia_candidates[MAX_PATTERN_SCALES]
+                                   [MAX_PATTERN_CANDIDATES] = {
+    {{0, -1}, {1, 0}, { 0, 1}, {-1, 0}},
+    {{-1, -1}, {0, -2}, {1, -1}, {2, 0}, {1, 1}, {0, 2}, {-1, 1}, {-2, 0}},
+    {{-2, -2}, {0, -4}, {2, -2}, {4, 0}, {2, 2}, {0, 4}, {-2, 2}, {-4, 0}},
+    {{-4, -4}, {0, -8}, {4, -4}, {8, 0}, {4, 4}, {0, 8}, {-4, 4}, {-8, 0}},
+    {{-8, -8}, {0, -16}, {8, -8}, {16, 0}, {8, 8}, {0, 16}, {-8, 8}, {-16, 0}},
+    {{-16, -16}, {0, -32}, {16, -16}, {32, 0}, {16, 16}, {0, 32},
+      {-16, 16}, {-32, 0}},
+    {{-32, -32}, {0, -64}, {32, -32}, {64, 0}, {32, 32}, {0, 64},
+      {-32, 32}, {-64, 0}},
+    {{-64, -64}, {0, -128}, {64, -64}, {128, 0}, {64, 64}, {0, 128},
+      {-64, 64}, {-128, 0}},
+    {{-128, -128}, {0, -256}, {128, -128}, {256, 0}, {128, 128}, {0, 256},
+      {-128, 128}, {-256, 0}},
+    {{-256, -256}, {0, -512}, {256, -256}, {512, 0}, {256, 256}, {0, 512},
+      {-256, 256}, {-512, 0}},
+    {{-512, -512}, {0, -1024}, {512, -512}, {1024, 0}, {512, 512}, {0, 1024},
+      {-512, 512}, {-1024, 0}},
+  };
+  return vp9_pattern_search_sad(x, ref_mv, search_param, sad_per_bit,
+                                do_init_search, cost_list, vfp, use_mvcost,
+                                center_mv, best_mv,
+                                bigdia_num_candidates, bigdia_candidates);
+}
+
+static int square_search(const MACROBLOCK *x,
+                         MV *ref_mv,
+                         int search_param,
+                         int sad_per_bit,
+                         int do_init_search,
+                         int *cost_list,
+                         const vp9_variance_fn_ptr_t *vfp,
+                         int use_mvcost,
+                         const MV *center_mv,
+                         MV *best_mv) {
+  // All scales have 8 closest points in square shape
+  static const int square_num_candidates[MAX_PATTERN_SCALES] = {
+    8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  };
+  // Note that the largest candidate step at each scale is 2^scale
+  static const MV square_candidates[MAX_PATTERN_SCALES]
+                                   [MAX_PATTERN_CANDIDATES] = {
+    {{-1, -1}, {0, -1}, {1, -1}, {1, 0}, {1, 1}, {0, 1}, {-1, 1}, {-1, 0}},
+    {{-2, -2}, {0, -2}, {2, -2}, {2, 0}, {2, 2}, {0, 2}, {-2, 2}, {-2, 0}},
+    {{-4, -4}, {0, -4}, {4, -4}, {4, 0}, {4, 4}, {0, 4}, {-4, 4}, {-4, 0}},
+    {{-8, -8}, {0, -8}, {8, -8}, {8, 0}, {8, 8}, {0, 8}, {-8, 8}, {-8, 0}},
+    {{-16, -16}, {0, -16}, {16, -16}, {16, 0}, {16, 16}, {0, 16},
+      {-16, 16}, {-16, 0}},
+    {{-32, -32}, {0, -32}, {32, -32}, {32, 0}, {32, 32}, {0, 32},
+      {-32, 32}, {-32, 0}},
+    {{-64, -64}, {0, -64}, {64, -64}, {64, 0}, {64, 64}, {0, 64},
+      {-64, 64}, {-64, 0}},
+    {{-128, -128}, {0, -128}, {128, -128}, {128, 0}, {128, 128}, {0, 128},
+      {-128, 128}, {-128, 0}},
+    {{-256, -256}, {0, -256}, {256, -256}, {256, 0}, {256, 256}, {0, 256},
+      {-256, 256}, {-256, 0}},
+    {{-512, -512}, {0, -512}, {512, -512}, {512, 0}, {512, 512}, {0, 512},
+      {-512, 512}, {-512, 0}},
+    {{-1024, -1024}, {0, -1024}, {1024, -1024}, {1024, 0}, {1024, 1024},
+      {0, 1024}, {-1024, 1024}, {-1024, 0}},
+  };
+  return vp9_pattern_search(x, ref_mv, search_param, sad_per_bit,
+                            do_init_search, cost_list, vfp, use_mvcost,
+                            center_mv, best_mv,
+                            square_num_candidates, square_candidates);
+}
+
+static int fast_hex_search(const MACROBLOCK *x,
+                           MV *ref_mv,
+                           int search_param,
+                           int sad_per_bit,
+                           int do_init_search,  // must be zero for fast_hex
+                           int *cost_list,
+                           const vp9_variance_fn_ptr_t *vfp,
+                           int use_mvcost,
+                           const MV *center_mv,
+                           MV *best_mv) {
+  return hex_search(x, ref_mv, VPXMAX(MAX_MVSEARCH_STEPS - 2, search_param),
+                    sad_per_bit, do_init_search, cost_list, vfp, use_mvcost,
+                    center_mv, best_mv);
+}
+
+static int fast_dia_search(const MACROBLOCK *x,
+                           MV *ref_mv,
+                           int search_param,
+                           int sad_per_bit,
+                           int do_init_search,
+                           int *cost_list,
+                           const vp9_variance_fn_ptr_t *vfp,
+                           int use_mvcost,
+                           const MV *center_mv,
+                           MV *best_mv) {
+  return bigdia_search(
+      x, ref_mv, VPXMAX(MAX_MVSEARCH_STEPS - 2, search_param), sad_per_bit,
+      do_init_search, cost_list, vfp, use_mvcost, center_mv, best_mv);
+}
+
+#undef CHECK_BETTER
+
+// Exhuastive motion search around a given centre position with a given
+// step size.
+static int exhuastive_mesh_search(const MACROBLOCK *x,
+                                  MV *ref_mv, MV *best_mv,
+                                  int range, int step, int sad_per_bit,
+                                  const vp9_variance_fn_ptr_t *fn_ptr,
+                                  const MV *center_mv) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  MV fcenter_mv = {center_mv->row, center_mv->col};
+  unsigned int best_sad = INT_MAX;
+  int r, c, i;
+  int start_col, end_col, start_row, end_row;
+  int col_step = (step > 1) ? step : 4;
+
+  assert(step >= 1);
+
+  clamp_mv(&fcenter_mv, x->mv_col_min, x->mv_col_max,
+           x->mv_row_min, x->mv_row_max);
+  *best_mv = fcenter_mv;
+  best_sad = fn_ptr->sdf(what->buf, what->stride,
+             get_buf_from_mv(in_what, &fcenter_mv), in_what->stride) +
+             mvsad_err_cost(x, &fcenter_mv, ref_mv, sad_per_bit);
+  start_row = VPXMAX(-range, x->mv_row_min - fcenter_mv.row);
+  start_col = VPXMAX(-range, x->mv_col_min - fcenter_mv.col);
+  end_row = VPXMIN(range, x->mv_row_max - fcenter_mv.row);
+  end_col = VPXMIN(range, x->mv_col_max - fcenter_mv.col);
+
+  for (r = start_row; r <= end_row; r += step) {
+    for (c = start_col; c <= end_col; c += col_step) {
+      // Step > 1 means we are not checking every location in this pass.
+      if (step > 1) {
+        const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c};
+        unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
+                           get_buf_from_mv(in_what, &mv), in_what->stride);
+        if (sad < best_sad) {
+          sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
+          if (sad < best_sad) {
+            best_sad = sad;
+            *best_mv = mv;
+          }
+        }
+      } else {
+        // 4 sads in a single call if we are checking every location
+        if (c + 3 <= end_col) {
+          unsigned int sads[4];
+          const uint8_t *addrs[4];
+          for (i = 0; i < 4; ++i) {
+            const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c + i};
+            addrs[i] = get_buf_from_mv(in_what, &mv);
+          }
+          fn_ptr->sdx4df(what->buf, what->stride, addrs,
+                         in_what->stride, sads);
+
+          for (i = 0; i < 4; ++i) {
+            if (sads[i] < best_sad) {
+              const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c + i};
+              const unsigned int sad = sads[i] +
+                  mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
+              if (sad < best_sad) {
+                best_sad = sad;
+                *best_mv = mv;
+              }
+            }
+          }
+        } else {
+          for (i = 0; i < end_col - c; ++i) {
+            const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c + i};
+            unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
+                get_buf_from_mv(in_what, &mv), in_what->stride);
+            if (sad < best_sad) {
+              sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
+              if (sad < best_sad) {
+                best_sad = sad;
+                *best_mv = mv;
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  return best_sad;
+}
+
+int vp9_diamond_search_sad_c(const MACROBLOCK *x,
+                             const search_site_config *cfg,
+                             MV *ref_mv, MV *best_mv, int search_param,
+                             int sad_per_bit, int *num00,
+                             const vp9_variance_fn_ptr_t *fn_ptr,
+                             const MV *center_mv) {
+  int i, j, step;
+
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  uint8_t *what = x->plane[0].src.buf;
+  const int what_stride = x->plane[0].src.stride;
+  const uint8_t *in_what;
+  const int in_what_stride = xd->plane[0].pre[0].stride;
+  const uint8_t *best_address;
+
+  unsigned int bestsad = INT_MAX;
+  int best_site = -1;
+  int last_site = -1;
+
+  int ref_row;
+  int ref_col;
+
+  // search_param determines the length of the initial step and hence the number
+  // of iterations.
+  // 0 = initial step (MAX_FIRST_STEP) pel
+  // 1 = (MAX_FIRST_STEP/2) pel,
+  // 2 = (MAX_FIRST_STEP/4) pel...
+//  const search_site *ss = &cfg->ss[search_param * cfg->searches_per_step];
+  const MV *ss_mv = &cfg->ss_mv[search_param * cfg->searches_per_step];
+  const intptr_t *ss_os = &cfg->ss_os[search_param * cfg->searches_per_step];
+  const int tot_steps = cfg->total_steps - search_param;
+
+  const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+  clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
+  ref_row = ref_mv->row;
+  ref_col = ref_mv->col;
+  *num00 = 0;
+  best_mv->row = ref_row;
+  best_mv->col = ref_col;
+
+  // Work out the start point for the search
+  in_what = xd->plane[0].pre[0].buf + ref_row * in_what_stride + ref_col;
+  best_address = in_what;
+
+  // Check the starting position
+  bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride)
+                + mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit);
+
+  i = 0;
+
+  for (step = 0; step < tot_steps; step++) {
+    int all_in = 1, t;
+
+    // All_in is true if every one of the points we are checking are within
+    // the bounds of the image.
+    all_in &= ((best_mv->row + ss_mv[i].row) > x->mv_row_min);
+    all_in &= ((best_mv->row + ss_mv[i + 1].row) < x->mv_row_max);
+    all_in &= ((best_mv->col + ss_mv[i + 2].col) > x->mv_col_min);
+    all_in &= ((best_mv->col + ss_mv[i + 3].col) < x->mv_col_max);
+
+    // If all the pixels are within the bounds we don't check whether the
+    // search point is valid in this loop,  otherwise we check each point
+    // for validity..
+    if (all_in) {
+      unsigned int sad_array[4];
+
+      for (j = 0; j < cfg->searches_per_step; j += 4) {
+        unsigned char const *block_offset[4];
+
+        for (t = 0; t < 4; t++)
+          block_offset[t] = ss_os[i + t] + best_address;
+
+        fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride,
+                       sad_array);
+
+        for (t = 0; t < 4; t++, i++) {
+          if (sad_array[t] < bestsad) {
+            const MV this_mv = {best_mv->row + ss_mv[i].row,
+                                best_mv->col + ss_mv[i].col};
+            sad_array[t] += mvsad_err_cost(x, &this_mv, &fcenter_mv,
+                                           sad_per_bit);
+            if (sad_array[t] < bestsad) {
+              bestsad = sad_array[t];
+              best_site = i;
+            }
+          }
+        }
+      }
+    } else {
+      for (j = 0; j < cfg->searches_per_step; j++) {
+        // Trap illegal vectors
+        const MV this_mv = {best_mv->row + ss_mv[i].row,
+                            best_mv->col + ss_mv[i].col};
+
+        if (is_mv_in(x, &this_mv)) {
+          const uint8_t *const check_here = ss_os[i] + best_address;
+          unsigned int thissad = fn_ptr->sdf(what, what_stride, check_here,
+                                             in_what_stride);
+
+          if (thissad < bestsad) {
+            thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
+            if (thissad < bestsad) {
+              bestsad = thissad;
+              best_site = i;
+            }
+          }
+        }
+        i++;
+      }
+    }
+    if (best_site != last_site) {
+      best_mv->row += ss_mv[best_site].row;
+      best_mv->col += ss_mv[best_site].col;
+      best_address += ss_os[best_site];
+      last_site = best_site;
+#if defined(NEW_DIAMOND_SEARCH)
+      while (1) {
+        const MV this_mv = {best_mv->row + ss_mv[best_site].row,
+                            best_mv->col + ss_mv[best_site].col};
+        if (is_mv_in(x, &this_mv)) {
+          const uint8_t *const check_here = ss_os[best_site] + best_address;
+          unsigned int thissad = fn_ptr->sdf(what, what_stride, check_here,
+                                             in_what_stride);
+          if (thissad < bestsad) {
+            thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
+            if (thissad < bestsad) {
+              bestsad = thissad;
+              best_mv->row += ss_mv[best_site].row;
+              best_mv->col += ss_mv[best_site].col;
+              best_address += ss_os[best_site];
+              continue;
+            }
+          }
+        }
+        break;
+      }
+#endif
+    } else if (best_address == in_what) {
+      (*num00)++;
+    }
+  }
+  return bestsad;
+}
+
+static int vector_match(int16_t *ref, int16_t *src, int bwl) {
+  int best_sad = INT_MAX;
+  int this_sad;
+  int d;
+  int center, offset = 0;
+  int bw = 4 << bwl;  // redundant variable, to be changed in the experiments.
+  for (d = 0; d <= bw; d += 16) {
+    this_sad = vpx_vector_var(&ref[d], src, bwl);
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      offset = d;
+    }
+  }
+  center = offset;
+
+  for (d = -8; d <= 8; d += 16) {
+    int this_pos = offset + d;
+    // check limit
+    if (this_pos < 0 || this_pos > bw)
+      continue;
+    this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      center = this_pos;
+    }
+  }
+  offset = center;
+
+  for (d = -4; d <= 4; d += 8) {
+    int this_pos = offset + d;
+    // check limit
+    if (this_pos < 0 || this_pos > bw)
+      continue;
+    this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      center = this_pos;
+    }
+  }
+  offset = center;
+
+  for (d = -2; d <= 2; d += 4) {
+    int this_pos = offset + d;
+    // check limit
+    if (this_pos < 0 || this_pos > bw)
+      continue;
+    this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      center = this_pos;
+    }
+  }
+  offset = center;
+
+  for (d = -1; d <= 1; d += 2) {
+    int this_pos = offset + d;
+    // check limit
+    if (this_pos < 0 || this_pos > bw)
+      continue;
+    this_sad = vpx_vector_var(&ref[this_pos], src, bwl);
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      center = this_pos;
+    }
+  }
+
+  return (center - (bw >> 1));
+}
+
+static const MV search_pos[4] = {
+    {-1, 0}, {0, -1}, {0, 1}, {1, 0},
+};
+
+unsigned int vp9_int_pro_motion_estimation(const VP9_COMP *cpi, MACROBLOCK *x,
+                                           BLOCK_SIZE bsize,
+                                           int mi_row, int mi_col) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
+  DECLARE_ALIGNED(16, int16_t, hbuf[128]);
+  DECLARE_ALIGNED(16, int16_t, vbuf[128]);
+  DECLARE_ALIGNED(16, int16_t, src_hbuf[64]);
+  DECLARE_ALIGNED(16, int16_t, src_vbuf[64]);
+  int idx;
+  const int bw = 4 << b_width_log2_lookup[bsize];
+  const int bh = 4 << b_height_log2_lookup[bsize];
+  const int search_width = bw << 1;
+  const int search_height = bh << 1;
+  const int src_stride = x->plane[0].src.stride;
+  const int ref_stride = xd->plane[0].pre[0].stride;
+  uint8_t const *ref_buf, *src_buf;
+  MV *tmp_mv = &xd->mi[0]->mv[0].as_mv;
+  unsigned int best_sad, tmp_sad, this_sad[4];
+  MV this_mv;
+  const int norm_factor = 3 + (bw >> 5);
+  const YV12_BUFFER_CONFIG *scaled_ref_frame =
+      vp9_get_scaled_ref_frame(cpi, mi->ref_frame[0]);
+
+  if (scaled_ref_frame) {
+    int i;
+    // Swap out the reference frame for a version that's been scaled to
+    // match the resolution of the current frame, allowing the existing
+    // motion search code to be used without additional modifications.
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      backup_yv12[i] = xd->plane[i].pre[0];
+    vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  {
+    unsigned int this_sad;
+    tmp_mv->row = 0;
+    tmp_mv->col = 0;
+    this_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf, src_stride,
+                                      xd->plane[0].pre[0].buf, ref_stride);
+
+    if (scaled_ref_frame) {
+      int i;
+      for (i = 0; i < MAX_MB_PLANE; i++)
+        xd->plane[i].pre[0] = backup_yv12[i];
+    }
+    return this_sad;
+  }
+#endif
+
+  // Set up prediction 1-D reference set
+  ref_buf = xd->plane[0].pre[0].buf - (bw >> 1);
+  for (idx = 0; idx < search_width; idx += 16) {
+    vpx_int_pro_row(&hbuf[idx], ref_buf, ref_stride, bh);
+    ref_buf += 16;
+  }
+
+  ref_buf = xd->plane[0].pre[0].buf - (bh >> 1) * ref_stride;
+  for (idx = 0; idx < search_height; ++idx) {
+    vbuf[idx] = vpx_int_pro_col(ref_buf, bw) >> norm_factor;
+    ref_buf += ref_stride;
+  }
+
+  // Set up src 1-D reference set
+  for (idx = 0; idx < bw; idx += 16) {
+    src_buf = x->plane[0].src.buf + idx;
+    vpx_int_pro_row(&src_hbuf[idx], src_buf, src_stride, bh);
+  }
+
+  src_buf = x->plane[0].src.buf;
+  for (idx = 0; idx < bh; ++idx) {
+    src_vbuf[idx] = vpx_int_pro_col(src_buf, bw) >> norm_factor;
+    src_buf += src_stride;
+  }
+
+  // Find the best match per 1-D search
+  tmp_mv->col = vector_match(hbuf, src_hbuf, b_width_log2_lookup[bsize]);
+  tmp_mv->row = vector_match(vbuf, src_vbuf, b_height_log2_lookup[bsize]);
+
+  this_mv = *tmp_mv;
+  src_buf = x->plane[0].src.buf;
+  ref_buf = xd->plane[0].pre[0].buf + this_mv.row * ref_stride + this_mv.col;
+  best_sad = cpi->fn_ptr[bsize].sdf(src_buf, src_stride, ref_buf, ref_stride);
+
+  {
+    const uint8_t * const pos[4] = {
+        ref_buf - ref_stride,
+        ref_buf - 1,
+        ref_buf + 1,
+        ref_buf + ref_stride,
+    };
+
+    cpi->fn_ptr[bsize].sdx4df(src_buf, src_stride, pos, ref_stride, this_sad);
+  }
+
+  for (idx = 0; idx < 4; ++idx) {
+    if (this_sad[idx] < best_sad) {
+      best_sad = this_sad[idx];
+      tmp_mv->row = search_pos[idx].row + this_mv.row;
+      tmp_mv->col = search_pos[idx].col + this_mv.col;
+    }
+  }
+
+  if (this_sad[0] < this_sad[3])
+    this_mv.row -= 1;
+  else
+    this_mv.row += 1;
+
+  if (this_sad[1] < this_sad[2])
+    this_mv.col -= 1;
+  else
+    this_mv.col += 1;
+
+  ref_buf = xd->plane[0].pre[0].buf + this_mv.row * ref_stride + this_mv.col;
+
+  tmp_sad = cpi->fn_ptr[bsize].sdf(src_buf, src_stride,
+                                   ref_buf, ref_stride);
+  if (best_sad > tmp_sad) {
+    *tmp_mv = this_mv;
+    best_sad = tmp_sad;
+  }
+
+  tmp_mv->row *= 8;
+  tmp_mv->col *= 8;
+
+  if (scaled_ref_frame) {
+    int i;
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      xd->plane[i].pre[0] = backup_yv12[i];
+  }
+
+  return best_sad;
+}
+
+// Runs sequence of diamond searches in smaller steps for RD.
+/* do_refine: If last step (1-away) of n-step search doesn't pick the center
+              point as the best match, we will do a final 1-away diamond
+              refining search  */
+static int full_pixel_diamond(const VP9_COMP *cpi, MACROBLOCK *x,
+                              MV *mvp_full, int step_param,
+                              int sadpb, int further_steps, int do_refine,
+                              int *cost_list,
+                              const vp9_variance_fn_ptr_t *fn_ptr,
+                              const MV *ref_mv, MV *dst_mv) {
+  MV temp_mv;
+  int thissme, n, num00 = 0;
+  int bestsme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
+                                        step_param, sadpb, &n,
+                                        fn_ptr, ref_mv);
+  if (bestsme < INT_MAX)
+    bestsme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
+  *dst_mv = temp_mv;
+
+  // If there won't be more n-step search, check to see if refining search is
+  // needed.
+  if (n > further_steps)
+    do_refine = 0;
+
+  while (n < further_steps) {
+    ++n;
+
+    if (num00) {
+      num00--;
+    } else {
+      thissme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
+                                        step_param + n, sadpb, &num00,
+                                        fn_ptr, ref_mv);
+      if (thissme < INT_MAX)
+        thissme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
+
+      // check to see if refining search is needed.
+      if (num00 > further_steps - n)
+        do_refine = 0;
+
+      if (thissme < bestsme) {
+        bestsme = thissme;
+        *dst_mv = temp_mv;
+      }
+    }
+  }
+
+  // final 1-away diamond refining search
+  if (do_refine) {
+    const int search_range = 8;
+    MV best_mv = *dst_mv;
+    thissme = vp9_refining_search_sad(x, &best_mv, sadpb, search_range,
+                                       fn_ptr, ref_mv);
+    if (thissme < INT_MAX)
+      thissme = vp9_get_mvpred_var(x, &best_mv, ref_mv, fn_ptr, 1);
+    if (thissme < bestsme) {
+      bestsme = thissme;
+      *dst_mv = best_mv;
+    }
+  }
+
+  // Return cost list.
+  if (cost_list) {
+    calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, dst_mv, cost_list);
+  }
+  return bestsme;
+}
+
+#define MIN_RANGE 7
+#define MAX_RANGE 256
+#define MIN_INTERVAL 1
+// Runs an limited range exhaustive mesh search using a pattern set
+// according to the encode speed profile.
+static int full_pixel_exhaustive(VP9_COMP *cpi, MACROBLOCK *x,
+                                 MV *centre_mv_full, int sadpb,  int *cost_list,
+                                 const vp9_variance_fn_ptr_t *fn_ptr,
+                                 const MV *ref_mv, MV *dst_mv) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  MV temp_mv = {centre_mv_full->row, centre_mv_full->col};
+  MV f_ref_mv = {ref_mv->row >> 3, ref_mv->col >> 3};
+  int bestsme;
+  int i;
+  int interval = sf->mesh_patterns[0].interval;
+  int range = sf->mesh_patterns[0].range;
+  int baseline_interval_divisor;
+
+  // Keep track of number of exhaustive calls (this frame in this thread).
+  ++(*x->ex_search_count_ptr);
+
+  // Trap illegal values for interval and range for this function.
+  if ((range < MIN_RANGE) || (range > MAX_RANGE) ||
+      (interval < MIN_INTERVAL) || (interval > range))
+    return INT_MAX;
+
+  baseline_interval_divisor = range / interval;
+
+  // Check size of proposed first range against magnitude of the centre
+  // value used as a starting point.
+  range = VPXMAX(range, (5 * VPXMAX(abs(temp_mv.row), abs(temp_mv.col))) / 4);
+  range = VPXMIN(range, MAX_RANGE);
+  interval = VPXMAX(interval, range / baseline_interval_divisor);
+
+  // initial search
+  bestsme = exhuastive_mesh_search(x, &f_ref_mv, &temp_mv, range,
+                                  interval, sadpb, fn_ptr, &temp_mv);
+
+  if ((interval > MIN_INTERVAL) && (range > MIN_RANGE)) {
+    // Progressive searches with range and step size decreasing each time
+    // till we reach a step size of 1. Then break out.
+    for (i = 1; i < MAX_MESH_STEP; ++i) {
+      // First pass with coarser step and longer range
+      bestsme = exhuastive_mesh_search(x, &f_ref_mv, &temp_mv,
+                                       sf->mesh_patterns[i].range,
+                                       sf->mesh_patterns[i].interval,
+                                       sadpb, fn_ptr, &temp_mv);
+
+      if (sf->mesh_patterns[i].interval == 1)
+        break;
+    }
+  }
+
+  if (bestsme < INT_MAX)
+    bestsme = vp9_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
+  *dst_mv = temp_mv;
+
+  // Return cost list.
+  if (cost_list) {
+    calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, dst_mv, cost_list);
+  }
+  return bestsme;
+}
+
+int vp9_full_search_sad_c(const MACROBLOCK *x, const MV *ref_mv,
+                          int sad_per_bit, int distance,
+                          const vp9_variance_fn_ptr_t *fn_ptr,
+                          const MV *center_mv, MV *best_mv) {
+  int r, c;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const int row_min = VPXMAX(ref_mv->row - distance, x->mv_row_min);
+  const int row_max = VPXMIN(ref_mv->row + distance, x->mv_row_max);
+  const int col_min = VPXMAX(ref_mv->col - distance, x->mv_col_min);
+  const int col_max = VPXMIN(ref_mv->col + distance, x->mv_col_max);
+  const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+  int best_sad = fn_ptr->sdf(what->buf, what->stride,
+      get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+  *best_mv = *ref_mv;
+
+  for (r = row_min; r < row_max; ++r) {
+    for (c = col_min; c < col_max; ++c) {
+      const MV mv = {r, c};
+      const int sad = fn_ptr->sdf(what->buf, what->stride,
+          get_buf_from_mv(in_what, &mv), in_what->stride) +
+              mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+      if (sad < best_sad) {
+        best_sad = sad;
+        *best_mv = mv;
+      }
+    }
+  }
+  return best_sad;
+}
+
+int vp9_full_search_sadx3(const MACROBLOCK *x, const MV *ref_mv,
+                          int sad_per_bit, int distance,
+                          const vp9_variance_fn_ptr_t *fn_ptr,
+                          const MV *center_mv, MV *best_mv) {
+  int r;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const int row_min = VPXMAX(ref_mv->row - distance, x->mv_row_min);
+  const int row_max = VPXMIN(ref_mv->row + distance, x->mv_row_max);
+  const int col_min = VPXMAX(ref_mv->col - distance, x->mv_col_min);
+  const int col_max = VPXMIN(ref_mv->col + distance, x->mv_col_max);
+  const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+  unsigned int best_sad = fn_ptr->sdf(what->buf, what->stride,
+      get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+  *best_mv = *ref_mv;
+
+  for (r = row_min; r < row_max; ++r) {
+    int c = col_min;
+    const uint8_t *check_here = &in_what->buf[r * in_what->stride + c];
+
+    if (fn_ptr->sdx3f != NULL) {
+      while ((c + 2) < col_max) {
+        int i;
+        DECLARE_ALIGNED(16, uint32_t, sads[3]);
+
+        fn_ptr->sdx3f(what->buf, what->stride, check_here, in_what->stride,
+                      sads);
+
+        for (i = 0; i < 3; ++i) {
+          unsigned int sad = sads[i];
+          if (sad < best_sad) {
+            const MV mv = {r, c};
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              *best_mv = mv;
+            }
+          }
+          ++check_here;
+          ++c;
+        }
+      }
+    }
+
+    while (c < col_max) {
+      unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
+                                     check_here, in_what->stride);
+      if (sad < best_sad) {
+        const MV mv = {r, c};
+        sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+        if (sad < best_sad) {
+          best_sad = sad;
+          *best_mv = mv;
+        }
+      }
+      ++check_here;
+      ++c;
+    }
+  }
+
+  return best_sad;
+}
+
+int vp9_full_search_sadx8(const MACROBLOCK *x, const MV *ref_mv,
+                          int sad_per_bit, int distance,
+                          const vp9_variance_fn_ptr_t *fn_ptr,
+                          const MV *center_mv, MV *best_mv) {
+  int r;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const int row_min = VPXMAX(ref_mv->row - distance, x->mv_row_min);
+  const int row_max = VPXMIN(ref_mv->row + distance, x->mv_row_max);
+  const int col_min = VPXMAX(ref_mv->col - distance, x->mv_col_min);
+  const int col_max = VPXMIN(ref_mv->col + distance, x->mv_col_max);
+  const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+  unsigned int best_sad = fn_ptr->sdf(what->buf, what->stride,
+      get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+  *best_mv = *ref_mv;
+
+  for (r = row_min; r < row_max; ++r) {
+    int c = col_min;
+    const uint8_t *check_here = &in_what->buf[r * in_what->stride + c];
+
+    if (fn_ptr->sdx8f != NULL) {
+      while ((c + 7) < col_max) {
+        int i;
+        DECLARE_ALIGNED(16, uint32_t, sads[8]);
+
+        fn_ptr->sdx8f(what->buf, what->stride, check_here, in_what->stride,
+                      sads);
+
+        for (i = 0; i < 8; ++i) {
+          unsigned int sad = sads[i];
+          if (sad < best_sad) {
+            const MV mv = {r, c};
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              *best_mv = mv;
+            }
+          }
+          ++check_here;
+          ++c;
+        }
+      }
+    }
+
+    if (fn_ptr->sdx3f != NULL) {
+      while ((c + 2) < col_max) {
+        int i;
+        DECLARE_ALIGNED(16, uint32_t, sads[3]);
+
+        fn_ptr->sdx3f(what->buf, what->stride, check_here, in_what->stride,
+                      sads);
+
+        for (i = 0; i < 3; ++i) {
+          unsigned int sad = sads[i];
+          if (sad < best_sad) {
+            const MV mv = {r, c};
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              *best_mv = mv;
+            }
+          }
+          ++check_here;
+          ++c;
+        }
+      }
+    }
+
+    while (c < col_max) {
+      unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
+                                     check_here, in_what->stride);
+      if (sad < best_sad) {
+        const MV mv = {r, c};
+        sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+        if (sad < best_sad) {
+          best_sad = sad;
+          *best_mv = mv;
+        }
+      }
+      ++check_here;
+      ++c;
+    }
+  }
+
+  return best_sad;
+}
+
+int vp9_refining_search_sad(const MACROBLOCK *x,
+                            MV *ref_mv, int error_per_bit,
+                            int search_range,
+                            const vp9_variance_fn_ptr_t *fn_ptr,
+                            const MV *center_mv) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const MV neighbors[4] = {{ -1, 0}, {0, -1}, {0, 1}, {1, 0}};
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+  const uint8_t *best_address = get_buf_from_mv(in_what, ref_mv);
+  unsigned int best_sad = fn_ptr->sdf(what->buf, what->stride, best_address,
+                                    in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
+  int i, j;
+
+  for (i = 0; i < search_range; i++) {
+    int best_site = -1;
+    const int all_in = ((ref_mv->row - 1) > x->mv_row_min) &
+                       ((ref_mv->row + 1) < x->mv_row_max) &
+                       ((ref_mv->col - 1) > x->mv_col_min) &
+                       ((ref_mv->col + 1) < x->mv_col_max);
+
+    if (all_in) {
+      unsigned int sads[4];
+      const uint8_t *const positions[4] = {
+        best_address - in_what->stride,
+        best_address - 1,
+        best_address + 1,
+        best_address + in_what->stride
+      };
+
+      fn_ptr->sdx4df(what->buf, what->stride, positions, in_what->stride, sads);
+
+      for (j = 0; j < 4; ++j) {
+        if (sads[j] < best_sad) {
+          const MV mv = {ref_mv->row + neighbors[j].row,
+                         ref_mv->col + neighbors[j].col};
+          sads[j] += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
+          if (sads[j] < best_sad) {
+            best_sad = sads[j];
+            best_site = j;
+          }
+        }
+      }
+    } else {
+      for (j = 0; j < 4; ++j) {
+        const MV mv = {ref_mv->row + neighbors[j].row,
+                       ref_mv->col + neighbors[j].col};
+
+        if (is_mv_in(x, &mv)) {
+          unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
+                                         get_buf_from_mv(in_what, &mv),
+                                         in_what->stride);
+          if (sad < best_sad) {
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              best_site = j;
+            }
+          }
+        }
+      }
+    }
+
+    if (best_site == -1) {
+      break;
+    } else {
+      ref_mv->row += neighbors[best_site].row;
+      ref_mv->col += neighbors[best_site].col;
+      best_address = get_buf_from_mv(in_what, ref_mv);
+    }
+  }
+
+  return best_sad;
+}
+
+// This function is called when we do joint motion search in comp_inter_inter
+// mode.
+int vp9_refining_search_8p_c(const MACROBLOCK *x,
+                             MV *ref_mv, int error_per_bit,
+                             int search_range,
+                             const vp9_variance_fn_ptr_t *fn_ptr,
+                             const MV *center_mv,
+                             const uint8_t *second_pred) {
+  const MV neighbors[8] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0},
+                           {-1, -1}, {1, -1}, {-1, 1}, {1, 1}};
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+  unsigned int best_sad = fn_ptr->sdaf(what->buf, what->stride,
+      get_buf_from_mv(in_what, ref_mv), in_what->stride, second_pred) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
+  int i, j;
+
+  for (i = 0; i < search_range; ++i) {
+    int best_site = -1;
+
+    for (j = 0; j < 8; ++j) {
+      const MV mv = {ref_mv->row + neighbors[j].row,
+                     ref_mv->col + neighbors[j].col};
+
+      if (is_mv_in(x, &mv)) {
+        unsigned int sad = fn_ptr->sdaf(what->buf, what->stride,
+            get_buf_from_mv(in_what, &mv), in_what->stride, second_pred);
+        if (sad < best_sad) {
+          sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
+          if (sad < best_sad) {
+            best_sad = sad;
+            best_site = j;
+          }
+        }
+      }
+    }
+
+    if (best_site == -1) {
+      break;
+    } else {
+      ref_mv->row += neighbors[best_site].row;
+      ref_mv->col += neighbors[best_site].col;
+    }
+  }
+  return best_sad;
+}
+
+#define MIN_EX_SEARCH_LIMIT 128
+static int is_exhaustive_allowed(VP9_COMP *cpi, MACROBLOCK *x) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  const int max_ex = VPXMAX(MIN_EX_SEARCH_LIMIT,
+      (*x->m_search_count_ptr * sf->max_exaustive_pct) / 100);
+
+  return sf->allow_exhaustive_searches &&
+      (sf->exhaustive_searches_thresh < INT_MAX) &&
+      (*x->ex_search_count_ptr <= max_ex) &&
+      !cpi->rc.is_src_frame_alt_ref;
+}
+
+int vp9_full_pixel_search(VP9_COMP *cpi, MACROBLOCK *x,
+                          BLOCK_SIZE bsize, MV *mvp_full,
+                          int step_param, int error_per_bit,
+                          int *cost_list,
+                          const MV *ref_mv, MV *tmp_mv,
+                          int var_max, int rd) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  const SEARCH_METHODS method = sf->mv.search_method;
+  vp9_variance_fn_ptr_t *fn_ptr = &cpi->fn_ptr[bsize];
+  int var = 0;
+  if (cost_list) {
+    cost_list[0] = INT_MAX;
+    cost_list[1] = INT_MAX;
+    cost_list[2] = INT_MAX;
+    cost_list[3] = INT_MAX;
+    cost_list[4] = INT_MAX;
+  }
+
+  // Keep track of number of searches (this frame in this thread).
+  ++(*x->m_search_count_ptr);
+
+  switch (method) {
+    case FAST_DIAMOND:
+      var = fast_dia_search(x, mvp_full, step_param, error_per_bit, 0,
+                            cost_list, fn_ptr, 1, ref_mv, tmp_mv);
+      break;
+    case FAST_HEX:
+      var = fast_hex_search(x, mvp_full, step_param, error_per_bit, 0,
+                            cost_list, fn_ptr, 1, ref_mv, tmp_mv);
+      break;
+    case HEX:
+      var = hex_search(x, mvp_full, step_param, error_per_bit, 1,
+                       cost_list, fn_ptr, 1, ref_mv, tmp_mv);
+      break;
+    case SQUARE:
+      var = square_search(x, mvp_full, step_param, error_per_bit, 1,
+                          cost_list, fn_ptr, 1, ref_mv, tmp_mv);
+      break;
+    case BIGDIA:
+      var = bigdia_search(x, mvp_full, step_param, error_per_bit, 1,
+                          cost_list, fn_ptr, 1, ref_mv, tmp_mv);
+      break;
+    case NSTEP:
+      var = full_pixel_diamond(cpi, x, mvp_full, step_param, error_per_bit,
+                               MAX_MVSEARCH_STEPS - 1 - step_param,
+                               1, cost_list, fn_ptr, ref_mv, tmp_mv);
+
+      // Should we allow a follow on exhaustive search?
+      if (is_exhaustive_allowed(cpi, x)) {
+        int64_t exhuastive_thr = sf->exhaustive_searches_thresh;
+        exhuastive_thr >>= 8 - (b_width_log2_lookup[bsize] +
+                                b_height_log2_lookup[bsize]);
+
+        // Threshold variance for an exhaustive full search.
+        if (var > exhuastive_thr) {
+            int var_ex;
+          MV tmp_mv_ex;
+          var_ex = full_pixel_exhaustive(cpi, x, tmp_mv,
+                                         error_per_bit, cost_list, fn_ptr,
+                                         ref_mv, &tmp_mv_ex);
+
+          if (var_ex < var) {
+            var = var_ex;
+            *tmp_mv = tmp_mv_ex;
+          }
+        }
+      }
+      break;
+    default:
+      assert(0 && "Invalid search method.");
+  }
+
+  if (method != NSTEP && rd && var < var_max)
+    var = vp9_get_mvpred_var(x, tmp_mv, ref_mv, fn_ptr, 1);
+
+  return var;
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_mcomp.h b/libvpx/libvpx/vp9/encoder/vp9_mcomp.h
new file mode 100644
index 0000000..86cd267
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_mcomp.h
@@ -0,0 +1,136 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_MCOMP_H_
+#define VP9_ENCODER_VP9_MCOMP_H_
+
+#include "vp9/encoder/vp9_block.h"
+#include "vpx_dsp/variance.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The maximum number of steps in a step search given the largest
+// allowed initial step
+#define MAX_MVSEARCH_STEPS 11
+// Max full pel mv specified in the unit of full pixel
+// Enable the use of motion vector in range [-1023, 1023].
+#define MAX_FULL_PEL_VAL ((1 << (MAX_MVSEARCH_STEPS - 1)) - 1)
+// Maximum size of the first step in full pel units
+#define MAX_FIRST_STEP (1 << (MAX_MVSEARCH_STEPS-1))
+// Allowed motion vector pixel distance outside image border
+// for Block_16x16
+#define BORDER_MV_PIXELS_B16 (16 + VP9_INTERP_EXTEND)
+
+typedef struct search_site_config {
+  // motion search sites
+  MV  ss_mv[8 * MAX_MVSEARCH_STEPS];        // Motion vector
+  intptr_t ss_os[8 * MAX_MVSEARCH_STEPS];   // Offset
+  int searches_per_step;
+  int total_steps;
+} search_site_config;
+
+void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride);
+void vp9_init3smotion_compensation(search_site_config *cfg,  int stride);
+
+void vp9_set_mv_search_range(MACROBLOCK *x, const MV *mv);
+int vp9_mv_bit_cost(const MV *mv, const MV *ref,
+                    const int *mvjcost, int *mvcost[2], int weight);
+
+// Utility to compute variance + MV rate cost for a given MV
+int vp9_get_mvpred_var(const MACROBLOCK *x,
+                       const MV *best_mv, const MV *center_mv,
+                       const vp9_variance_fn_ptr_t *vfp,
+                       int use_mvcost);
+int vp9_get_mvpred_av_var(const MACROBLOCK *x,
+                          const MV *best_mv, const MV *center_mv,
+                          const uint8_t *second_pred,
+                          const vp9_variance_fn_ptr_t *vfp,
+                          int use_mvcost);
+
+struct VP9_COMP;
+struct SPEED_FEATURES;
+
+int vp9_init_search_range(int size);
+
+int vp9_refining_search_sad(const struct macroblock *x,
+                            struct mv *ref_mv,
+                            int sad_per_bit, int distance,
+                            const struct vp9_variance_vtable *fn_ptr,
+                            const struct mv *center_mv);
+
+// Perform integral projection based motion estimation.
+unsigned int vp9_int_pro_motion_estimation(const struct VP9_COMP *cpi,
+                                           MACROBLOCK *x,
+                                           BLOCK_SIZE bsize,
+                                           int mi_row, int mi_col);
+
+typedef uint32_t (fractional_mv_step_fp) (
+    const MACROBLOCK *x,
+    MV *bestmv, const MV *ref_mv,
+    int allow_hp,
+    int error_per_bit,
+    const vp9_variance_fn_ptr_t *vfp,
+    int forced_stop,  // 0 - full, 1 - qtr only, 2 - half only
+    int iters_per_step,
+    int *cost_list,
+    int *mvjcost, int *mvcost[2],
+    uint32_t *distortion, uint32_t *sse1,
+    const uint8_t *second_pred,
+    int w, int h);
+
+extern fractional_mv_step_fp vp9_find_best_sub_pixel_tree;
+extern fractional_mv_step_fp vp9_find_best_sub_pixel_tree_pruned;
+extern fractional_mv_step_fp vp9_find_best_sub_pixel_tree_pruned_more;
+extern fractional_mv_step_fp vp9_find_best_sub_pixel_tree_pruned_evenmore;
+extern fractional_mv_step_fp vp9_skip_sub_pixel_tree;
+
+typedef int (*vp9_full_search_fn_t)(const MACROBLOCK *x,
+                                    const MV *ref_mv, int sad_per_bit,
+                                    int distance,
+                                    const vp9_variance_fn_ptr_t *fn_ptr,
+                                    const MV *center_mv, MV *best_mv);
+
+typedef int (*vp9_refining_search_fn_t)(const MACROBLOCK *x,
+                                        MV *ref_mv, int sad_per_bit,
+                                        int distance,
+                                        const vp9_variance_fn_ptr_t *fn_ptr,
+                                        const MV *center_mv);
+
+typedef int (*vp9_diamond_search_fn_t)(const MACROBLOCK *x,
+                                       const search_site_config *cfg,
+                                       MV *ref_mv, MV *best_mv,
+                                       int search_param, int sad_per_bit,
+                                       int *num00,
+                                       const vp9_variance_fn_ptr_t *fn_ptr,
+                                       const MV *center_mv);
+
+int vp9_refining_search_8p_c(const MACROBLOCK *x,
+                             MV *ref_mv, int error_per_bit,
+                             int search_range,
+                             const vp9_variance_fn_ptr_t *fn_ptr,
+                             const MV *center_mv, const uint8_t *second_pred);
+
+struct VP9_COMP;
+
+int vp9_full_pixel_search(struct VP9_COMP *cpi, MACROBLOCK *x,
+                          BLOCK_SIZE bsize, MV *mvp_full,
+                          int step_param, int error_per_bit,
+                          int *cost_list,
+                          const MV *ref_mv, MV *tmp_mv,
+                          int var_max, int rd);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_MCOMP_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_noise_estimate.c b/libvpx/libvpx/vp9/encoder/vp9_noise_estimate.c
new file mode 100644
index 0000000..4b43b38
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_noise_estimate.c
@@ -0,0 +1,263 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_noise_estimate.h"
+#include "vp9/encoder/vp9_encoder.h"
+
+void vp9_noise_estimate_init(NOISE_ESTIMATE *const ne,
+                             int width,
+                             int height) {
+  ne->enabled = 0;
+  ne->level = kLowLow;
+  ne->value = 0;
+  ne->count = 0;
+  ne->thresh = 90;
+  ne->last_w = 0;
+  ne->last_h = 0;
+  if (width * height >= 1920 * 1080) {
+    ne->thresh = 200;
+  } else if (width * height >= 1280 * 720) {
+    ne->thresh = 130;
+  }
+  ne->num_frames_estimate = 20;
+}
+
+static int enable_noise_estimation(VP9_COMP *const cpi) {
+  // Enable noise estimation if denoising is on.
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0)
+    return 1;
+#endif
+  // Only allow noise estimate under certain encoding mode.
+  // Enabled for 1 pass CBR, speed >=5, and if resolution is same as original.
+  // Not enabled for SVC mode and screen_content_mode.
+  // Not enabled for low resolutions.
+  if (cpi->oxcf.pass == 0 &&
+      cpi->oxcf.rc_mode == VPX_CBR &&
+      cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
+      cpi->oxcf.speed >= 5 &&
+      cpi->resize_state == ORIG &&
+      cpi->resize_pending == 0 &&
+      !cpi->use_svc &&
+      cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
+      cpi->common.width >= 640 &&
+      cpi->common.height >= 480)
+    return 1;
+  else
+    return 0;
+}
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+static void copy_frame(YV12_BUFFER_CONFIG * const dest,
+                       const YV12_BUFFER_CONFIG * const src) {
+  int r;
+  const uint8_t *srcbuf = src->y_buffer;
+  uint8_t *destbuf = dest->y_buffer;
+
+  assert(dest->y_width == src->y_width);
+  assert(dest->y_height == src->y_height);
+
+  for (r = 0; r < dest->y_height; ++r) {
+    memcpy(destbuf, srcbuf, dest->y_width);
+    destbuf += dest->y_stride;
+    srcbuf += src->y_stride;
+  }
+}
+#endif  // CONFIG_VP9_TEMPORAL_DENOISING
+
+NOISE_LEVEL vp9_noise_estimate_extract_level(NOISE_ESTIMATE *const ne) {
+  int noise_level = kLowLow;
+  if (ne->value > (ne->thresh << 1)) {
+    noise_level = kHigh;
+  } else {
+    if (ne->value > ne->thresh)
+      noise_level = kMedium;
+    else if (ne->value > ((9 * ne->thresh) >> 4))
+      noise_level = kLow;
+    else
+      noise_level = kLowLow;
+  }
+  return noise_level;
+}
+
+void vp9_update_noise_estimate(VP9_COMP *const cpi) {
+  const VP9_COMMON *const cm = &cpi->common;
+  NOISE_ESTIMATE *const ne = &cpi->noise_estimate;
+  // Estimate of noise level every frame_period frames.
+  int frame_period = 8;
+  int thresh_consec_zeromv = 6;
+  unsigned int thresh_sum_diff = 100;
+  unsigned int thresh_sum_spatial = (200 * 200) << 8;
+  unsigned int thresh_spatial_var = (32 * 32) << 8;
+  int min_blocks_estimate = cm->mi_rows * cm->mi_cols >> 7;
+  // Estimate is between current source and last source.
+  YV12_BUFFER_CONFIG *last_source = cpi->Last_Source;
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0)
+    last_source = &cpi->denoiser.last_source;
+#endif
+  ne->enabled = enable_noise_estimation(cpi);
+  if (!ne->enabled ||
+      cm->current_video_frame % frame_period != 0 ||
+      last_source == NULL ||
+      ne->last_w != cm->width ||
+      ne->last_h != cm->height) {
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0)
+    copy_frame(&cpi->denoiser.last_source, cpi->Source);
+#endif
+    if (last_source != NULL) {
+      ne->last_w = cm->width;
+      ne->last_h = cm->height;
+    }
+    return;
+  } else if (cpi->rc.avg_frame_low_motion < 50) {
+    // Force noise estimation to 0 and denoiser off if content has high motion.
+    ne->level = kLowLow;
+#if CONFIG_VP9_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity > 0)
+      vp9_denoiser_set_noise_level(&cpi->denoiser, ne->level);
+#endif
+    return;
+  } else {
+    int num_samples = 0;
+    uint64_t avg_est = 0;
+    int bsize = BLOCK_16X16;
+    static const unsigned char const_source[16] = {
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+    // Loop over sub-sample of 16x16 blocks of frame, and for blocks that have
+    // been encoded as zero/small mv at least x consecutive frames, compute
+    // the variance to update estimate of noise in the source.
+    const uint8_t *src_y = cpi->Source->y_buffer;
+    const int src_ystride = cpi->Source->y_stride;
+    const uint8_t *last_src_y = last_source->y_buffer;
+    const int last_src_ystride = last_source->y_stride;
+    const uint8_t *src_u = cpi->Source->u_buffer;
+    const uint8_t *src_v = cpi->Source->v_buffer;
+    const int src_uvstride = cpi->Source->uv_stride;
+    int mi_row, mi_col;
+    int num_low_motion = 0;
+    int frame_low_motion = 1;
+    for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
+      for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
+        int bl_index = mi_row * cm->mi_cols + mi_col;
+        if (cpi->consec_zero_mv[bl_index] > thresh_consec_zeromv)
+          num_low_motion++;
+      }
+    }
+    if (num_low_motion < ((3 * cm->mi_rows * cm->mi_cols) >> 3))
+      frame_low_motion = 0;
+    for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
+      for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
+        // 16x16 blocks, 1/4 sample of frame.
+        if (mi_row % 4 == 0 && mi_col % 4 == 0 &&
+            mi_row < cm->mi_rows - 1 &&
+            mi_col < cm->mi_cols - 1) {
+          int bl_index = mi_row * cm->mi_cols + mi_col;
+          int bl_index1 = bl_index + 1;
+          int bl_index2 = bl_index + cm->mi_cols;
+          int bl_index3 = bl_index2 + 1;
+          // Only consider blocks that are likely steady background. i.e, have
+          // been encoded as zero/low motion x (= thresh_consec_zeromv) frames
+          // in a row. consec_zero_mv[] defined for 8x8 blocks, so consider all
+          // 4 sub-blocks for 16x16 block. Also, avoid skin blocks.
+          int consec_zeromv = VPXMIN(cpi->consec_zero_mv[bl_index],
+                                     VPXMIN(cpi->consec_zero_mv[bl_index1],
+                                     VPXMIN(cpi->consec_zero_mv[bl_index2],
+                                     cpi->consec_zero_mv[bl_index3])));
+          int is_skin = 0;
+          if (cpi->use_skin_detection) {
+            is_skin = vp9_compute_skin_block(src_y,
+                                             src_u,
+                                             src_v,
+                                             src_ystride,
+                                             src_uvstride,
+                                             bsize,
+                                             consec_zeromv,
+                                             0);
+          }
+          if (frame_low_motion &&
+              cpi->consec_zero_mv[bl_index] > thresh_consec_zeromv &&
+              cpi->consec_zero_mv[bl_index1] > thresh_consec_zeromv &&
+              cpi->consec_zero_mv[bl_index2] > thresh_consec_zeromv &&
+              cpi->consec_zero_mv[bl_index3] > thresh_consec_zeromv &&
+              !is_skin) {
+            // Compute variance.
+            unsigned int sse;
+            unsigned int variance = cpi->fn_ptr[bsize].vf(src_y,
+                                                          src_ystride,
+                                                          last_src_y,
+                                                          last_src_ystride,
+                                                          &sse);
+            // Only consider this block as valid for noise measurement if the
+            // average term (sse - variance = N * avg^{2}, N = 16X16) of the
+            // temporal residual is small (avoid effects from lighting change).
+            if ((sse - variance) < thresh_sum_diff) {
+              unsigned int sse2;
+              const unsigned int spatial_variance =
+                  cpi->fn_ptr[bsize].vf(src_y, src_ystride, const_source,
+                                        0, &sse2);
+              // Avoid blocks with high brightness and high spatial variance.
+              if ((sse2 - spatial_variance) < thresh_sum_spatial &&
+                  spatial_variance < thresh_spatial_var) {
+                avg_est += variance / ((spatial_variance >> 9) + 1);
+                num_samples++;
+              }
+            }
+          }
+        }
+        src_y += 8;
+        last_src_y += 8;
+        src_u += 4;
+        src_v += 4;
+      }
+      src_y += (src_ystride << 3) - (cm->mi_cols << 3);
+      last_src_y += (last_src_ystride << 3) - (cm->mi_cols << 3);
+      src_u += (src_uvstride << 2) - (cm->mi_cols << 2);
+      src_v += (src_uvstride << 2) - (cm->mi_cols << 2);
+    }
+    ne->last_w = cm->width;
+    ne->last_h = cm->height;
+    // Update noise estimate if we have at a minimum number of block samples,
+    // and avg_est > 0 (avg_est == 0 can happen if the application inputs
+    // duplicate frames).
+    if (num_samples > min_blocks_estimate && avg_est > 0) {
+      // Normalize.
+      avg_est = avg_est / num_samples;
+      // Update noise estimate.
+      ne->value = (int)((15 * ne->value + avg_est) >> 4);
+      ne->count++;
+      if (ne->count == ne->num_frames_estimate) {
+        // Reset counter and check noise level condition.
+        ne->num_frames_estimate = 30;
+        ne->count = 0;
+        ne->level = vp9_noise_estimate_extract_level(ne);
+#if CONFIG_VP9_TEMPORAL_DENOISING
+        if (cpi->oxcf.noise_sensitivity > 0)
+          vp9_denoiser_set_noise_level(&cpi->denoiser, ne->level);
+#endif
+      }
+    }
+  }
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0)
+    copy_frame(&cpi->denoiser.last_source, cpi->Source);
+#endif
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_noise_estimate.h b/libvpx/libvpx/vp9/encoder/vp9_noise_estimate.h
new file mode 100644
index 0000000..826d125
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_noise_estimate.h
@@ -0,0 +1,58 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_NOISE_ESTIMATE_H_
+#define VP9_ENCODER_NOISE_ESTIMATE_H_
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_skin_detection.h"
+#include "vpx_scale/yv12config.h"
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#include "vp9/encoder/vp9_denoiser.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef enum noise_level {
+  kLowLow,
+  kLow,
+  kMedium,
+  kHigh
+} NOISE_LEVEL;
+
+typedef struct noise_estimate {
+  int enabled;
+  NOISE_LEVEL level;
+  int value;
+  int thresh;
+  int count;
+  int last_w;
+  int last_h;
+  int num_frames_estimate;
+} NOISE_ESTIMATE;
+
+struct VP9_COMP;
+
+void vp9_noise_estimate_init(NOISE_ESTIMATE *const ne,
+                             int width,
+                             int height);
+
+NOISE_LEVEL vp9_noise_estimate_extract_level(NOISE_ESTIMATE *const ne);
+
+void vp9_update_noise_estimate(struct VP9_COMP *const cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_NOISE_ESTIMATE_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_picklpf.c b/libvpx/libvpx/vp9/encoder/vp9_picklpf.c
new file mode 100644
index 0000000..f6b1dfc
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_picklpf.c
@@ -0,0 +1,195 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+
+#include "./vpx_scale_rtcd.h"
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_quant_common.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_picklpf.h"
+#include "vp9/encoder/vp9_quantize.h"
+
+static int get_max_filter_level(const VP9_COMP *cpi) {
+  if (cpi->oxcf.pass == 2) {
+    return cpi->twopass.section_intra_rating > 8 ? MAX_LOOP_FILTER * 3 / 4
+                                                 : MAX_LOOP_FILTER;
+  } else {
+    return MAX_LOOP_FILTER;
+  }
+}
+
+
+static int64_t try_filter_frame(const YV12_BUFFER_CONFIG *sd,
+                                VP9_COMP *const cpi,
+                                int filt_level, int partial_frame) {
+  VP9_COMMON *const cm = &cpi->common;
+  int64_t filt_err;
+
+  vp9_build_mask_frame(cm, filt_level, partial_frame);
+
+  if (cpi->num_workers > 1)
+    vp9_loop_filter_frame_mt(cm->frame_to_show, cm, cpi->td.mb.e_mbd.plane,
+                             filt_level, 1, partial_frame,
+                             cpi->workers, cpi->num_workers, &cpi->lf_row_sync);
+  else
+    vp9_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filt_level,
+                          1, partial_frame);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    filt_err = vp9_highbd_get_y_sse(sd, cm->frame_to_show);
+  } else {
+    filt_err = vp9_get_y_sse(sd, cm->frame_to_show);
+  }
+#else
+  filt_err = vp9_get_y_sse(sd, cm->frame_to_show);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  // Re-instate the unfiltered frame
+  vpx_yv12_copy_y(&cpi->last_frame_uf, cm->frame_to_show);
+
+  return filt_err;
+}
+
+static int search_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
+                               int partial_frame) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const struct loopfilter *const lf = &cm->lf;
+  const int min_filter_level = 0;
+  const int max_filter_level = get_max_filter_level(cpi);
+  int filt_direction = 0;
+  int64_t best_err;
+  int filt_best;
+
+  // Start the search at the previous frame filter level unless it is now out of
+  // range.
+  int filt_mid =
+      clamp(lf->last_filt_level, min_filter_level, max_filter_level);
+  int filter_step = filt_mid < 16 ? 4 : filt_mid / 4;
+  // Sum squared error at each filter level
+  int64_t ss_err[MAX_LOOP_FILTER + 1];
+
+  // Set each entry to -1
+  memset(ss_err, 0xFF, sizeof(ss_err));
+
+  //  Make a copy of the unfiltered / processed recon buffer
+  vpx_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_uf);
+
+  best_err = try_filter_frame(sd, cpi, filt_mid, partial_frame);
+  filt_best = filt_mid;
+  ss_err[filt_mid] = best_err;
+
+  while (filter_step > 0) {
+    const int filt_high = VPXMIN(filt_mid + filter_step, max_filter_level);
+    const int filt_low = VPXMAX(filt_mid - filter_step, min_filter_level);
+
+    // Bias against raising loop filter in favor of lowering it.
+    int64_t bias = (best_err >> (15 - (filt_mid / 8))) * filter_step;
+
+    if ((cpi->oxcf.pass == 2) && (cpi->twopass.section_intra_rating < 20))
+      bias = (bias * cpi->twopass.section_intra_rating) / 20;
+
+    // yx, bias less for large block size
+    if (cm->tx_mode != ONLY_4X4)
+      bias >>= 1;
+
+    if (filt_direction <= 0 && filt_low != filt_mid) {
+      // Get Low filter error score
+      if (ss_err[filt_low] < 0) {
+        ss_err[filt_low] = try_filter_frame(sd, cpi, filt_low, partial_frame);
+      }
+      // If value is close to the best so far then bias towards a lower loop
+      // filter value.
+      if ((ss_err[filt_low] - bias) < best_err) {
+        // Was it actually better than the previous best?
+        if (ss_err[filt_low] < best_err)
+          best_err = ss_err[filt_low];
+
+        filt_best = filt_low;
+      }
+    }
+
+    // Now look at filt_high
+    if (filt_direction >= 0 && filt_high != filt_mid) {
+      if (ss_err[filt_high] < 0) {
+        ss_err[filt_high] = try_filter_frame(sd, cpi, filt_high, partial_frame);
+      }
+      // Was it better than the previous best?
+      if (ss_err[filt_high] < (best_err - bias)) {
+        best_err = ss_err[filt_high];
+        filt_best = filt_high;
+      }
+    }
+
+    // Half the step distance if the best filter value was the same as last time
+    if (filt_best == filt_mid) {
+      filter_step /= 2;
+      filt_direction = 0;
+    } else {
+      filt_direction = (filt_best < filt_mid) ? -1 : 1;
+      filt_mid = filt_best;
+    }
+  }
+
+  return filt_best;
+}
+
+void vp9_pick_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
+                           LPF_PICK_METHOD method) {
+  VP9_COMMON *const cm = &cpi->common;
+  struct loopfilter *const lf = &cm->lf;
+
+  lf->sharpness_level = cm->frame_type == KEY_FRAME ? 0
+                                                    : cpi->oxcf.sharpness;
+
+  if (method == LPF_PICK_MINIMAL_LPF && lf->filter_level) {
+      lf->filter_level = 0;
+  } else if (method >= LPF_PICK_FROM_Q) {
+    const int min_filter_level = 0;
+    const int max_filter_level = get_max_filter_level(cpi);
+    const int q = vp9_ac_quant(cm->base_qindex, 0, cm->bit_depth);
+    // These values were determined by linear fitting the result of the
+    // searched level, filt_guess = q * 0.316206 + 3.87252
+#if CONFIG_VP9_HIGHBITDEPTH
+    int filt_guess;
+    switch (cm->bit_depth) {
+      case VPX_BITS_8:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 1015158, 18);
+        break;
+      case VPX_BITS_10:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 4060632, 20);
+        break;
+      case VPX_BITS_12:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 16242526, 22);
+        break;
+      default:
+        assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 "
+                    "or VPX_BITS_12");
+        return;
+    }
+#else
+    int filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 1015158, 18);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    if (cm->frame_type == KEY_FRAME)
+      filt_guess -= 4;
+    lf->filter_level = clamp(filt_guess, min_filter_level, max_filter_level);
+  } else {
+    lf->filter_level = search_filter_level(sd, cpi,
+                                           method == LPF_PICK_FROM_SUBIMAGE);
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_picklpf.h b/libvpx/libvpx/vp9/encoder/vp9_picklpf.h
new file mode 100644
index 0000000..33c490f
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_picklpf.h
@@ -0,0 +1,30 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_PICKLPF_H_
+#define VP9_ENCODER_VP9_PICKLPF_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "vp9/encoder/vp9_encoder.h"
+
+struct yv12_buffer_config;
+struct VP9_COMP;
+
+void vp9_pick_filter_level(const struct yv12_buffer_config *sd,
+                           struct VP9_COMP *cpi, LPF_PICK_METHOD method);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_PICKLPF_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_pickmode.c b/libvpx/libvpx/vp9/encoder/vp9_pickmode.c
new file mode 100644
index 0000000..ba6a0c6
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_pickmode.c
@@ -0,0 +1,2331 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_scan.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_pickmode.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+
+typedef struct {
+  uint8_t *data;
+  int stride;
+  int in_use;
+} PRED_BUFFER;
+
+
+static const int pos_shift_16x16[4][4] = {
+  {9, 10, 13, 14},
+  {11, 12, 15, 16},
+  {17, 18, 21, 22},
+  {19, 20, 23, 24}
+};
+
+static int mv_refs_rt(VP9_COMP *cpi, const VP9_COMMON *cm,
+                      const MACROBLOCK *x,
+                      const MACROBLOCKD *xd,
+                      const TileInfo *const tile,
+                      MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                      int_mv *mv_ref_list, int_mv *base_mv,
+                      int mi_row, int mi_col, int use_base_mv) {
+  const int *ref_sign_bias = cm->ref_frame_sign_bias;
+  int i, refmv_count = 0;
+
+  const POSITION *const mv_ref_search = mv_ref_blocks[mi->sb_type];
+
+  int different_ref_found = 0;
+  int context_counter = 0;
+  int const_motion = 0;
+
+  // Blank the reference vector list
+  memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
+
+  // The nearest 2 blocks are treated differently
+  // if the size < 8x8 we get the mv from the bmi substructure,
+  // and we also need to keep a mode count.
+  for (i = 0; i < 2; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row *
+                                                   xd->mi_stride];
+      // Keep counts for entropy encoding.
+      context_counter += mode_2_counter[candidate_mi->mode];
+      different_ref_found = 1;
+
+      if (candidate_mi->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
+                        refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  const_motion = 1;
+
+  // Check the rest of the neighbors in much the same way
+  // as before except we don't need to keep track of sub blocks or
+  // mode counts.
+  for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row *
+                                                   xd->mi_stride];
+      different_ref_found = 1;
+
+      if (candidate_mi->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(candidate_mi->mv[0], refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  // Since we couldn't find 2 mvs from the same reference frame
+  // go back through the neighbors and find motion vectors from
+  // different reference frames.
+  if (different_ref_found && !refmv_count) {
+    for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
+      const POSITION *mv_ref = &mv_ref_search[i];
+      if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+        const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row
+                                              * xd->mi_stride];
+
+        // If the candidate is INTRA we don't want to consider its mv.
+        IF_DIFF_REF_FRAME_ADD_MV(candidate_mi, ref_frame, ref_sign_bias,
+                                 refmv_count, mv_ref_list, Done);
+      }
+    }
+  }
+  if (use_base_mv &&
+      !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
+      ref_frame == LAST_FRAME) {
+    // Get base layer mv.
+    MV_REF *candidate =
+        &cm->prev_frame->mvs[(mi_col>>1) + (mi_row>>1) * (cm->mi_cols>>1)];
+    if (candidate->mv[0].as_int != INVALID_MV) {
+        base_mv->as_mv.row = (candidate->mv[0].as_mv.row * 2);
+        base_mv->as_mv.col = (candidate->mv[0].as_mv.col * 2);
+      clamp_mv_ref(&base_mv->as_mv, xd);
+    } else {
+      base_mv->as_int = INVALID_MV;
+    }
+  }
+
+ Done:
+
+  x->mbmi_ext->mode_context[ref_frame] = counter_to_context[context_counter];
+
+  // Clamp vectors
+  for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
+    clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
+
+  return const_motion;
+}
+
+static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
+                                  BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                  int_mv *tmp_mv, int *rate_mv,
+                                  int64_t best_rd_sofar, int use_base_mv) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
+  const int step_param = cpi->sf.mv.fullpel_search_step_param;
+  const int sadpb = x->sadperbit16;
+  MV mvp_full;
+  const int ref = mi->ref_frame[0];
+  const MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
+  MV center_mv;
+  uint32_t dis;
+  int rate_mode;
+  const int tmp_col_min = x->mv_col_min;
+  const int tmp_col_max = x->mv_col_max;
+  const int tmp_row_min = x->mv_row_min;
+  const int tmp_row_max = x->mv_row_max;
+  int rv = 0;
+  int cost_list[5];
+  const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
+                                                                        ref);
+  if (scaled_ref_frame) {
+    int i;
+    // Swap out the reference frame for a version that's been scaled to
+    // match the resolution of the current frame, allowing the existing
+    // motion search code to be used without additional modifications.
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      backup_yv12[i] = xd->plane[i].pre[0];
+    vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
+  }
+  vp9_set_mv_search_range(x, &ref_mv);
+
+  assert(x->mv_best_ref_index[ref] <= 2);
+  if (x->mv_best_ref_index[ref] < 2)
+    mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
+  else
+    mvp_full = x->pred_mv[ref];
+
+  mvp_full.col >>= 3;
+  mvp_full.row >>= 3;
+
+  if (!use_base_mv)
+    center_mv = ref_mv;
+  else
+    center_mv = tmp_mv->as_mv;
+
+  vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
+                        cond_cost_list(cpi, cost_list),
+                        &center_mv, &tmp_mv->as_mv, INT_MAX, 0);
+
+  x->mv_col_min = tmp_col_min;
+  x->mv_col_max = tmp_col_max;
+  x->mv_row_min = tmp_row_min;
+  x->mv_row_max = tmp_row_max;
+
+  // calculate the bit cost on motion vector
+  mvp_full.row = tmp_mv->as_mv.row * 8;
+  mvp_full.col = tmp_mv->as_mv.col * 8;
+
+  *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv,
+                             x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+
+  rate_mode = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]]
+                                  [INTER_OFFSET(NEWMV)];
+  rv = !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) >
+         best_rd_sofar);
+
+  if (rv) {
+    cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
+                                 cpi->common.allow_high_precision_mv,
+                                 x->errorperbit,
+                                 &cpi->fn_ptr[bsize],
+                                 cpi->sf.mv.subpel_force_stop,
+                                 cpi->sf.mv.subpel_iters_per_step,
+                                 cond_cost_list(cpi, cost_list),
+                                 x->nmvjointcost, x->mvcost,
+                                 &dis, &x->pred_sse[ref], NULL, 0, 0);
+    *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
+                               x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+  }
+
+  if (scaled_ref_frame) {
+    int i;
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      xd->plane[i].pre[0] = backup_yv12[i];
+  }
+  return rv;
+}
+
+static void block_variance(const uint8_t *src, int src_stride,
+                           const uint8_t *ref, int ref_stride,
+                           int w, int h, unsigned int *sse, int *sum,
+                           int block_size, unsigned int *sse8x8,
+                           int *sum8x8, unsigned int *var8x8) {
+  int i, j, k = 0;
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      vpx_get8x8var(src + src_stride * i + j, src_stride,
+                    ref + ref_stride * i + j, ref_stride,
+                    &sse8x8[k], &sum8x8[k]);
+      *sse += sse8x8[k];
+      *sum += sum8x8[k];
+      var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6);
+      k++;
+    }
+  }
+}
+
+static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
+                               unsigned int *sse_i, int *sum_i,
+                               unsigned int *var_o, unsigned int *sse_o,
+                               int *sum_o) {
+  const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
+  const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
+  const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
+  int i, j, k = 0;
+
+  for (i = 0; i < nh; i += 2) {
+    for (j = 0; j < nw; j += 2) {
+      sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
+          sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
+      sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
+          sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
+      var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
+          (b_width_log2_lookup[unit_size] +
+              b_height_log2_lookup[unit_size] + 6));
+      k++;
+    }
+  }
+}
+
+static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
+                                    MACROBLOCK *x, MACROBLOCKD *xd,
+                                    int *out_rate_sum, int64_t *out_dist_sum,
+                                    unsigned int *var_y, unsigned int *sse_y,
+                                    int mi_row, int mi_col, int *early_term) {
+  // Note our transform coeffs are 8 times an orthogonal transform.
+  // Hence quantizer step is also 8 times. To get effective quantizer
+  // we need to divide by 8 before sending to modeling function.
+  unsigned int sse;
+  int rate;
+  int64_t dist;
+  struct macroblock_plane *const p = &x->plane[0];
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  const uint32_t dc_quant = pd->dequant[0];
+  const uint32_t ac_quant = pd->dequant[1];
+  const int64_t dc_thr = dc_quant * dc_quant >> 6;
+  const int64_t ac_thr = ac_quant * ac_quant >> 6;
+  unsigned int var;
+  int sum;
+  int skip_dc = 0;
+
+  const int bw = b_width_log2_lookup[bsize];
+  const int bh = b_height_log2_lookup[bsize];
+  const int num8x8 = 1 << (bw + bh - 2);
+  unsigned int sse8x8[64] = {0};
+  int sum8x8[64] = {0};
+  unsigned int var8x8[64] = {0};
+  TX_SIZE tx_size;
+  int i, k;
+
+  // Calculate variance for whole partition, and also save 8x8 blocks' variance
+  // to be used in following transform skipping test.
+  block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
+                 4 << bw, 4 << bh, &sse, &sum, 8, sse8x8, sum8x8, var8x8);
+  var = sse - (((int64_t)sum * sum) >> (bw + bh + 4));
+
+  *var_y = var;
+  *sse_y = sse;
+
+  if (cpi->common.tx_mode == TX_MODE_SELECT) {
+    if (sse > (var << 2))
+      tx_size = VPXMIN(max_txsize_lookup[bsize],
+                       tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+    else
+      tx_size = TX_8X8;
+
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
+        cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
+      tx_size = TX_8X8;
+    else if (tx_size > TX_16X16)
+      tx_size = TX_16X16;
+  } else {
+    tx_size = VPXMIN(max_txsize_lookup[bsize],
+                     tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+  }
+
+  assert(tx_size >= TX_8X8);
+  xd->mi[0]->tx_size = tx_size;
+
+  // Evaluate if the partition block is a skippable block in Y plane.
+  {
+    unsigned int sse16x16[16] = {0};
+    int sum16x16[16] = {0};
+    unsigned int var16x16[16] = {0};
+    const int num16x16 = num8x8 >> 2;
+
+    unsigned int sse32x32[4] = {0};
+    int sum32x32[4] = {0};
+    unsigned int var32x32[4] = {0};
+    const int num32x32 = num8x8 >> 4;
+
+    int ac_test = 1;
+    int dc_test = 1;
+    const int num = (tx_size == TX_8X8) ? num8x8 :
+        ((tx_size == TX_16X16) ? num16x16 : num32x32);
+    const unsigned int *sse_tx = (tx_size == TX_8X8) ? sse8x8 :
+        ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
+    const unsigned int *var_tx = (tx_size == TX_8X8) ? var8x8 :
+        ((tx_size == TX_16X16) ? var16x16 : var32x32);
+
+    // Calculate variance if tx_size > TX_8X8
+    if (tx_size >= TX_16X16)
+      calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
+                         sum16x16);
+    if (tx_size == TX_32X32)
+      calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
+                         sse32x32, sum32x32);
+
+    // Skipping test
+    x->skip_txfm[0] = SKIP_TXFM_NONE;
+    for (k = 0; k < num; k++)
+      // Check if all ac coefficients can be quantized to zero.
+      if (!(var_tx[k] < ac_thr || var == 0)) {
+        ac_test = 0;
+        break;
+      }
+
+    for (k = 0; k < num; k++)
+      // Check if dc coefficient can be quantized to zero.
+      if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
+        dc_test = 0;
+        break;
+      }
+
+    if (ac_test) {
+      x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
+
+      if (dc_test)
+        x->skip_txfm[0] = SKIP_TXFM_AC_DC;
+    } else if (dc_test) {
+      skip_dc = 1;
+    }
+  }
+
+  if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
+    int skip_uv[2] = {0};
+    unsigned int var_uv[2];
+    unsigned int sse_uv[2];
+
+    *out_rate_sum = 0;
+    *out_dist_sum = sse << 4;
+
+    // Transform skipping test in UV planes.
+    for (i = 1; i <= 2; i++) {
+      struct macroblock_plane *const p = &x->plane[i];
+      struct macroblockd_plane *const pd = &xd->plane[i];
+      const TX_SIZE uv_tx_size = get_uv_tx_size(xd->mi[0], pd);
+      const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
+      const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
+      const int uv_bw = b_width_log2_lookup[uv_bsize];
+      const int uv_bh = b_height_log2_lookup[uv_bsize];
+      const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
+          (uv_bh - b_height_log2_lookup[unit_size]);
+      const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
+      const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
+      int j = i - 1;
+
+      vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
+      var_uv[j] = cpi->fn_ptr[uv_bsize].vf(p->src.buf, p->src.stride,
+          pd->dst.buf, pd->dst.stride, &sse_uv[j]);
+
+      if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
+          (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
+        skip_uv[j] = 1;
+      else
+        break;
+    }
+
+    // If the transform in YUV planes are skippable, the mode search checks
+    // fewer inter modes and doesn't check intra modes.
+    if (skip_uv[0] & skip_uv[1]) {
+      *early_term = 1;
+    }
+
+    return;
+  }
+
+  if (!skip_dc) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
+                                   dc_quant >> (xd->bd - 5), &rate, &dist);
+    } else {
+      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
+                                   dc_quant >> 3, &rate, &dist);
+    }
+#else
+    vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
+                                 dc_quant >> 3, &rate, &dist);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+
+  if (!skip_dc) {
+    *out_rate_sum = rate >> 1;
+    *out_dist_sum = dist << 3;
+  } else {
+    *out_rate_sum = 0;
+    *out_dist_sum = (sse - var) << 4;
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
+                                 ac_quant >> (xd->bd - 5), &rate, &dist);
+  } else {
+    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
+                                 ac_quant >> 3, &rate, &dist);
+  }
+#else
+  vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
+                               ac_quant >> 3, &rate, &dist);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  *out_rate_sum += rate;
+  *out_dist_sum += dist << 4;
+}
+
+static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
+                              MACROBLOCK *x, MACROBLOCKD *xd,
+                              int *out_rate_sum, int64_t *out_dist_sum,
+                              unsigned int *var_y, unsigned int *sse_y) {
+  // Note our transform coeffs are 8 times an orthogonal transform.
+  // Hence quantizer step is also 8 times. To get effective quantizer
+  // we need to divide by 8 before sending to modeling function.
+  unsigned int sse;
+  int rate;
+  int64_t dist;
+  struct macroblock_plane *const p = &x->plane[0];
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  const int64_t dc_thr = p->quant_thred[0] >> 6;
+  const int64_t ac_thr = p->quant_thred[1] >> 6;
+  const uint32_t dc_quant = pd->dequant[0];
+  const uint32_t ac_quant = pd->dequant[1];
+  unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
+                                           pd->dst.buf, pd->dst.stride, &sse);
+  int skip_dc = 0;
+
+  *var_y = var;
+  *sse_y = sse;
+
+  if (cpi->common.tx_mode == TX_MODE_SELECT) {
+    if (sse > (var << 2))
+      xd->mi[0]->tx_size =
+          VPXMIN(max_txsize_lookup[bsize],
+                 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+    else
+      xd->mi[0]->tx_size = TX_8X8;
+
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
+        cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
+      xd->mi[0]->tx_size = TX_8X8;
+    else if (xd->mi[0]->tx_size > TX_16X16)
+      xd->mi[0]->tx_size = TX_16X16;
+  } else {
+    xd->mi[0]->tx_size =
+        VPXMIN(max_txsize_lookup[bsize],
+               tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+  }
+
+  // Evaluate if the partition block is a skippable block in Y plane.
+  {
+    const BLOCK_SIZE unit_size =
+        txsize_to_bsize[xd->mi[0]->tx_size];
+    const unsigned int num_blk_log2 =
+        (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
+        (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
+    const unsigned int sse_tx = sse >> num_blk_log2;
+    const unsigned int var_tx = var >> num_blk_log2;
+
+    x->skip_txfm[0] = SKIP_TXFM_NONE;
+    // Check if all ac coefficients can be quantized to zero.
+    if (var_tx < ac_thr || var == 0) {
+      x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
+      // Check if dc coefficient can be quantized to zero.
+      if (sse_tx - var_tx < dc_thr || sse == var)
+        x->skip_txfm[0] = SKIP_TXFM_AC_DC;
+    } else {
+      if (sse_tx - var_tx < dc_thr || sse == var)
+        skip_dc = 1;
+    }
+  }
+
+  if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
+    *out_rate_sum = 0;
+    *out_dist_sum = sse << 4;
+    return;
+  }
+
+  if (!skip_dc) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
+                                   dc_quant >> (xd->bd - 5), &rate, &dist);
+    } else {
+      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
+                                   dc_quant >> 3, &rate, &dist);
+    }
+#else
+    vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
+                                 dc_quant >> 3, &rate, &dist);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+
+  if (!skip_dc) {
+    *out_rate_sum = rate >> 1;
+    *out_dist_sum = dist << 3;
+  } else {
+    *out_rate_sum = 0;
+    *out_dist_sum = (sse - var) << 4;
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
+                                 ac_quant >> (xd->bd - 5), &rate, &dist);
+  } else {
+    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
+                                 ac_quant >> 3, &rate, &dist);
+  }
+#else
+  vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
+                               ac_quant >> 3, &rate, &dist);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  *out_rate_sum += rate;
+  *out_dist_sum += dist << 4;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
+                      int *skippable, int64_t *sse, BLOCK_SIZE bsize,
+                      TX_SIZE tx_size) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  unsigned int var_y, sse_y;
+
+  (void)tx_size;
+  model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist, &var_y,
+                    &sse_y);
+  *sse = INT_MAX;
+  *skippable = 0;
+  return;
+}
+#else
+static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
+                      int *skippable, int64_t *sse, BLOCK_SIZE bsize,
+                      TX_SIZE tx_size) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  const struct macroblockd_plane *pd = &xd->plane[0];
+  struct macroblock_plane *const p = &x->plane[0];
+  const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+  const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+  const int step = 1 << (tx_size << 1);
+  const int block_step = (1 << tx_size);
+  int block = 0, r, c;
+  const int max_blocks_wide = num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 :
+      xd->mb_to_right_edge >> 5);
+  const int max_blocks_high = num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 :
+      xd->mb_to_bottom_edge >> 5);
+  int eob_cost = 0;
+  const int bw = 4 * num_4x4_w;
+  const int bh = 4 * num_4x4_h;
+
+  (void)cpi;
+
+  // The max tx_size passed in is TX_16X16.
+  assert(tx_size != TX_32X32);
+
+  vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
+                     pd->dst.buf, pd->dst.stride);
+  *skippable = 1;
+  // Keep track of the row and column of the blocks we use so that we know
+  // if we are in the unrestricted motion border.
+  for (r = 0; r < max_blocks_high; r += block_step) {
+    for (c = 0; c < num_4x4_w; c += block_step) {
+      if (c < max_blocks_wide) {
+        const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
+        tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+        tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+        tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+        uint16_t *const eob = &p->eobs[block];
+        const int diff_stride = bw;
+        const int16_t *src_diff;
+        src_diff = &p->src_diff[(r * diff_stride + c) << 2];
+
+        switch (tx_size) {
+          case TX_16X16:
+            vpx_hadamard_16x16(src_diff, diff_stride, (int16_t *)coeff);
+            vp9_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
+                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                            pd->dequant, eob,
+                            scan_order->scan, scan_order->iscan);
+            break;
+          case TX_8X8:
+            vpx_hadamard_8x8(src_diff, diff_stride, (int16_t *)coeff);
+            vp9_quantize_fp(coeff, 64, x->skip_block, p->zbin, p->round_fp,
+                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                            pd->dequant, eob,
+                            scan_order->scan, scan_order->iscan);
+            break;
+          case TX_4X4:
+            x->fwd_txm4x4(src_diff, coeff, diff_stride);
+            vp9_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
+                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                            pd->dequant, eob,
+                            scan_order->scan, scan_order->iscan);
+            break;
+          default:
+            assert(0);
+            break;
+        }
+        *skippable &= (*eob == 0);
+        eob_cost += 1;
+      }
+      block += step;
+    }
+  }
+
+  this_rdc->rate = 0;
+  if (*sse < INT64_MAX) {
+    *sse = (*sse << 6) >> 2;
+    if (*skippable) {
+      this_rdc->dist = *sse;
+      return;
+    }
+  }
+
+  block = 0;
+  this_rdc->dist = 0;
+  for (r = 0; r < max_blocks_high; r += block_step) {
+    for (c = 0; c < num_4x4_w; c += block_step) {
+      if (c < max_blocks_wide) {
+        tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+        tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+        tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+        uint16_t *const eob = &p->eobs[block];
+
+        if (*eob == 1)
+          this_rdc->rate += (int)abs(qcoeff[0]);
+        else if (*eob > 1)
+          this_rdc->rate += vpx_satd((const int16_t *)qcoeff, step << 4);
+
+        this_rdc->dist +=
+            vp9_block_error_fp(coeff, dqcoeff, step << 4) >> 2;
+      }
+      block += step;
+    }
+  }
+
+  // If skippable is set, rate gets clobbered later.
+  this_rdc->rate <<= (2 + VP9_PROB_COST_SHIFT);
+  this_rdc->rate += (eob_cost << VP9_PROB_COST_SHIFT);
+}
+#endif
+
+static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize,
+                               MACROBLOCK *x, MACROBLOCKD *xd,
+                               RD_COST *this_rdc,
+                               unsigned int *var_y, unsigned int *sse_y,
+                               int start_plane, int stop_plane) {
+  // Note our transform coeffs are 8 times an orthogonal transform.
+  // Hence quantizer step is also 8 times. To get effective quantizer
+  // we need to divide by 8 before sending to modeling function.
+  unsigned int sse;
+  int rate;
+  int64_t dist;
+  int i;
+
+  this_rdc->rate = 0;
+  this_rdc->dist = 0;
+
+  for (i = start_plane; i <= stop_plane; ++i) {
+    struct macroblock_plane *const p = &x->plane[i];
+    struct macroblockd_plane *const pd = &xd->plane[i];
+    const uint32_t dc_quant = pd->dequant[0];
+    const uint32_t ac_quant = pd->dequant[1];
+    const BLOCK_SIZE bs = plane_bsize;
+    unsigned int var;
+
+    if (!x->color_sensitivity[i - 1])
+      continue;
+
+    var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
+                             pd->dst.buf, pd->dst.stride, &sse);
+    *var_y += var;
+    *sse_y += sse;
+
+  #if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
+                                   dc_quant >> (xd->bd - 5), &rate, &dist);
+    } else {
+      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
+                                   dc_quant >> 3, &rate, &dist);
+    }
+  #else
+    vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
+                                 dc_quant >> 3, &rate, &dist);
+  #endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    this_rdc->rate += rate >> 1;
+    this_rdc->dist += dist << 3;
+
+  #if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
+                                   ac_quant >> (xd->bd - 5), &rate, &dist);
+    } else {
+      vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
+                                   ac_quant >> 3, &rate, &dist);
+    }
+  #else
+    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
+                                 ac_quant >> 3, &rate, &dist);
+  #endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    this_rdc->rate += rate;
+    this_rdc->dist += dist << 4;
+  }
+}
+
+static int get_pred_buffer(PRED_BUFFER *p, int len) {
+  int i;
+
+  for (i = 0; i < len; i++) {
+    if (!p[i].in_use) {
+      p[i].in_use = 1;
+      return i;
+    }
+  }
+  return -1;
+}
+
+static void free_pred_buffer(PRED_BUFFER *p) {
+  if (p != NULL)
+    p->in_use = 0;
+}
+
+static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x,
+                                 BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                 MV_REFERENCE_FRAME ref_frame,
+                                 PREDICTION_MODE this_mode,
+                                 unsigned int var_y, unsigned int sse_y,
+                                 struct buf_2d yv12_mb[][MAX_MB_PLANE],
+                                 int *rate, int64_t *dist) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
+  unsigned int var = var_y, sse = sse_y;
+  // Skipping threshold for ac.
+  unsigned int thresh_ac;
+  // Skipping threshold for dc.
+  unsigned int thresh_dc;
+  int motion_low = 1;
+  if (mi->mv[0].as_mv.row > 64 ||
+      mi->mv[0].as_mv.row < -64 ||
+      mi->mv[0].as_mv.col > 64 ||
+      mi->mv[0].as_mv.col < -64)
+    motion_low = 0;
+  if (x->encode_breakout > 0 && motion_low == 1) {
+    // Set a maximum for threshold to avoid big PSNR loss in low bit rate
+    // case. Use extreme low threshold for static frames to limit
+    // skipping.
+    const unsigned int max_thresh = 36000;
+    // The encode_breakout input
+    const unsigned int min_thresh =
+        VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh);
+#if CONFIG_VP9_HIGHBITDEPTH
+    const int shift = (xd->bd << 1) - 16;
+#endif
+
+    // Calculate threshold according to dequant value.
+    thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
+#if CONFIG_VP9_HIGHBITDEPTH
+    if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
+      thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
+
+    // Adjust ac threshold according to partition size.
+    thresh_ac >>=
+        8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
+
+    thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
+#if CONFIG_VP9_HIGHBITDEPTH
+    if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
+      thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  } else {
+    thresh_ac = 0;
+    thresh_dc = 0;
+  }
+
+  // Y skipping condition checking for ac and dc.
+  if (var <= thresh_ac && (sse - var) <= thresh_dc) {
+    unsigned int sse_u, sse_v;
+    unsigned int var_u, var_v;
+    unsigned int thresh_ac_uv = thresh_ac;
+    unsigned int thresh_dc_uv = thresh_dc;
+    if (x->sb_is_skin) {
+      thresh_ac_uv = 0;
+      thresh_dc_uv = 0;
+    }
+
+    // Skip UV prediction unless breakout is zero (lossless) to save
+    // computation with low impact on the result
+    if (x->encode_breakout == 0) {
+      xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
+      xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
+      vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
+    }
+
+    var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
+                                    x->plane[1].src.stride,
+                                    xd->plane[1].dst.buf,
+                                    xd->plane[1].dst.stride, &sse_u);
+
+    // U skipping condition checking
+    if (((var_u << 2) <= thresh_ac_uv) && (sse_u - var_u <= thresh_dc_uv)) {
+      var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
+                                      x->plane[2].src.stride,
+                                      xd->plane[2].dst.buf,
+                                      xd->plane[2].dst.stride, &sse_v);
+
+      // V skipping condition checking
+      if (((var_v << 2) <= thresh_ac_uv) && (sse_v - var_v <= thresh_dc_uv)) {
+        x->skip = 1;
+
+        // The cost of skip bit needs to be added.
+        *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
+                                    [INTER_OFFSET(this_mode)];
+
+        // More on this part of rate
+        // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+
+        // Scaling factor for SSE from spatial domain to frequency
+        // domain is 16. Adjust distortion accordingly.
+        // TODO(yunqingwang): In this function, only y-plane dist is
+        // calculated.
+        *dist = (sse << 4);  // + ((sse_u + sse_v) << 4);
+
+        // *disable_skip = 1;
+      }
+    }
+  }
+}
+
+struct estimate_block_intra_args {
+  VP9_COMP *cpi;
+  MACROBLOCK *x;
+  PREDICTION_MODE mode;
+  int skippable;
+  RD_COST *rdc;
+};
+
+static void estimate_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
+                                 TX_SIZE tx_size, void *arg) {
+  struct estimate_block_intra_args* const args = arg;
+  VP9_COMP *const cpi = args->cpi;
+  MACROBLOCK *const x = args->x;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *const p = &x->plane[0];
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
+  uint8_t *const src_buf_base = p->src.buf;
+  uint8_t *const dst_buf_base = pd->dst.buf;
+  const int src_stride = p->src.stride;
+  const int dst_stride = pd->dst.stride;
+  int i, j;
+  RD_COST this_rdc;
+
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
+
+  p->src.buf = &src_buf_base[4 * (j * src_stride + i)];
+  pd->dst.buf = &dst_buf_base[4 * (j * dst_stride + i)];
+  // Use source buffer as an approximation for the fully reconstructed buffer.
+  vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize],
+                          tx_size, args->mode,
+                          x->skip_encode ? p->src.buf : pd->dst.buf,
+                          x->skip_encode ? src_stride : dst_stride,
+                          pd->dst.buf, dst_stride,
+                          i, j, plane);
+
+  if (plane == 0) {
+    int64_t this_sse = INT64_MAX;
+    // TODO(jingning): This needs further refactoring.
+    block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx,
+              VPXMIN(tx_size, TX_16X16));
+  } else {
+    unsigned int var = 0;
+    unsigned int sse = 0;
+    model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane,
+                       plane);
+  }
+
+  p->src.buf = src_buf_base;
+  pd->dst.buf = dst_buf_base;
+  args->rdc->rate += this_rdc.rate;
+  args->rdc->dist += this_rdc.dist;
+}
+
+static const THR_MODES mode_idx[MAX_REF_FRAMES - 1][4] = {
+  {THR_DC, THR_V_PRED, THR_H_PRED, THR_TM},
+  {THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV},
+  {THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG},
+};
+
+static const PREDICTION_MODE intra_mode_list[] = {
+  DC_PRED, V_PRED, H_PRED, TM_PRED
+};
+
+static int mode_offset(const PREDICTION_MODE mode) {
+  if (mode >= NEARESTMV) {
+    return INTER_OFFSET(mode);
+  } else {
+    switch (mode) {
+      case DC_PRED:
+        return 0;
+      case V_PRED:
+        return 1;
+      case H_PRED:
+        return 2;
+      case TM_PRED:
+        return 3;
+      default:
+        return -1;
+    }
+  }
+}
+
+static INLINE void update_thresh_freq_fact(VP9_COMP *cpi,
+                                           TileDataEnc *tile_data,
+                                           BLOCK_SIZE bsize,
+                                           MV_REFERENCE_FRAME ref_frame,
+                                           THR_MODES best_mode_idx,
+                                           PREDICTION_MODE mode) {
+  THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
+  int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
+  if (thr_mode_idx == best_mode_idx)
+    *freq_fact -= (*freq_fact >> 4);
+  else
+    *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
+                        cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
+}
+
+void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
+                         BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  RD_COST this_rdc, best_rdc;
+  PREDICTION_MODE this_mode;
+  struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
+  const TX_SIZE intra_tx_size =
+      VPXMIN(max_txsize_lookup[bsize],
+             tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+  MODE_INFO *const mic = xd->mi[0];
+  int *bmode_costs;
+  const MODE_INFO *above_mi = xd->above_mi;
+  const MODE_INFO *left_mi = xd->left_mi;
+  const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
+  const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
+  bmode_costs = cpi->y_mode_costs[A][L];
+
+  (void) ctx;
+  vp9_rd_cost_reset(&best_rdc);
+  vp9_rd_cost_reset(&this_rdc);
+
+  mi->ref_frame[0] = INTRA_FRAME;
+  mi->mv[0].as_int = INVALID_MV;
+  mi->uv_mode = DC_PRED;
+  memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
+
+  // Change the limit of this loop to add other intra prediction
+  // mode tests.
+  for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
+    this_rdc.dist = this_rdc.rate = 0;
+    args.mode = this_mode;
+    args.skippable = 1;
+    args.rdc = &this_rdc;
+    mi->tx_size = intra_tx_size;
+    vp9_foreach_transformed_block_in_plane(xd, bsize, 0,
+                                           estimate_block_intra, &args);
+    if (args.skippable) {
+      x->skip_txfm[0] = SKIP_TXFM_AC_DC;
+      this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
+    } else {
+      x->skip_txfm[0] = SKIP_TXFM_NONE;
+      this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
+    }
+    this_rdc.rate += bmode_costs[this_mode];
+    this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                             this_rdc.rate, this_rdc.dist);
+
+    if (this_rdc.rdcost < best_rdc.rdcost) {
+      best_rdc = this_rdc;
+      mi->mode = this_mode;
+    }
+  }
+
+  *rd_cost = best_rdc;
+}
+
+static void init_ref_frame_cost(VP9_COMMON *const cm,
+                                MACROBLOCKD *const xd,
+                                int ref_frame_cost[MAX_REF_FRAMES]) {
+  vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
+  vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
+  vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
+
+  ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
+  ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
+    ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
+
+  ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
+  ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
+  ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
+  ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
+  ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
+}
+
+typedef struct {
+  MV_REFERENCE_FRAME ref_frame;
+  PREDICTION_MODE pred_mode;
+} REF_MODE;
+
+#define RT_INTER_MODES 8
+static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
+    {LAST_FRAME, ZEROMV},
+    {LAST_FRAME, NEARESTMV},
+    {GOLDEN_FRAME, ZEROMV},
+    {LAST_FRAME, NEARMV},
+    {LAST_FRAME, NEWMV},
+    {GOLDEN_FRAME, NEARESTMV},
+    {GOLDEN_FRAME, NEARMV},
+    {GOLDEN_FRAME, NEWMV}
+};
+static const REF_MODE ref_mode_set_svc[RT_INTER_MODES] = {
+    {LAST_FRAME, ZEROMV},
+    {GOLDEN_FRAME, ZEROMV},
+    {LAST_FRAME, NEARESTMV},
+    {LAST_FRAME, NEARMV},
+    {GOLDEN_FRAME, NEARESTMV},
+    {GOLDEN_FRAME, NEARMV},
+    {LAST_FRAME, NEWMV},
+    {GOLDEN_FRAME, NEWMV}
+};
+
+static int set_intra_cost_penalty(const VP9_COMP *const cpi, BLOCK_SIZE bsize) {
+  const VP9_COMMON *const cm = &cpi->common;
+  // Reduce the intra cost penalty for small blocks (<=16x16).
+  int reduction_fac =
+      (bsize <= BLOCK_16X16) ? ((bsize <= BLOCK_8X8) ? 4 : 2) : 0;
+  if (cpi->noise_estimate.enabled && cpi->noise_estimate.level == kHigh)
+     // Don't reduce intra cost penalty if estimated noise level is high.
+     reduction_fac = 0;
+  return vp9_get_intra_cost_penalty(
+      cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth) >> reduction_fac;
+}
+
+static INLINE void find_predictors(VP9_COMP *cpi, MACROBLOCK *x,
+                                 MV_REFERENCE_FRAME ref_frame,
+                                 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
+                                 int const_motion[MAX_REF_FRAMES],
+                                 int *ref_frame_skip_mask,
+                                 const int flag_list[4],
+                                 TileDataEnc *tile_data,
+                                 int mi_row, int mi_col,
+                                 struct buf_2d yv12_mb[4][MAX_MB_PLANE],
+                                 BLOCK_SIZE bsize,
+                                 int force_skip_low_temp_var) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
+  TileInfo *const tile_info = &tile_data->tile_info;
+  // TODO(jingning) placeholder for inter-frame non-RD mode decision.
+  x->pred_mv_sad[ref_frame] = INT_MAX;
+  frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
+  frame_mv[ZEROMV][ref_frame].as_int = 0;
+  // this needs various further optimizations. to be continued..
+  if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
+    int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
+    const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
+    vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
+                         sf, sf);
+    if (cm->use_prev_frame_mvs) {
+      vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame,
+                       candidates, mi_row, mi_col,
+                       x->mbmi_ext->mode_context);
+    } else {
+      const_motion[ref_frame] =
+          mv_refs_rt(cpi, cm, x, xd, tile_info, xd->mi[0], ref_frame,
+                     candidates, &frame_mv[NEWMV][ref_frame], mi_row, mi_col,
+                     (int)(cpi->svc.use_base_mv && cpi->svc.spatial_layer_id));
+    }
+    vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
+                          &frame_mv[NEARESTMV][ref_frame],
+                          &frame_mv[NEARMV][ref_frame]);
+    // Early exit for golden frame if force_skip_low_temp_var is set.
+    if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8 &&
+        !(force_skip_low_temp_var && ref_frame == GOLDEN_FRAME)) {
+      vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
+                  ref_frame, bsize);
+    }
+  } else {
+    *ref_frame_skip_mask |= (1 << ref_frame);
+  }
+}
+
+static void vp9_large_block_mv_bias(const NOISE_ESTIMATE *ne, RD_COST *this_rdc,
+                                    BLOCK_SIZE bsize, int mv_row, int mv_col,
+                                    int is_last_frame) {
+  // Bias against non-zero (above some threshold) motion for large blocks.
+  // This is temporary fix to avoid selection of large mv for big blocks.
+  if (mv_row > 64 || mv_row < -64 || mv_col > 64 || mv_col < -64) {
+    if (bsize == BLOCK_64X64)
+      this_rdc->rdcost = this_rdc->rdcost << 1;
+    else if (bsize >= BLOCK_32X32)
+      this_rdc->rdcost = 3 * this_rdc->rdcost >> 1;
+  }
+  // If noise estimation is enabled, and estimated level is above threshold,
+  // add a bias to LAST reference with small motion, for large blocks.
+  if (ne->enabled && ne->level >= kMedium &&
+      bsize >= BLOCK_32X32 && is_last_frame &&
+      mv_row < 8 && mv_row > -8 && mv_col < 8 && mv_col > -8) {
+    this_rdc->rdcost = 7 * this_rdc->rdcost >> 3;
+  }
+}
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+static void vp9_pickmode_ctx_den_update(
+    VP9_PICKMODE_CTX_DEN *ctx_den,
+    int64_t zero_last_cost_orig,
+    int ref_frame_cost[MAX_REF_FRAMES],
+    int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
+    int reuse_inter_pred,
+    TX_SIZE best_tx_size,
+    PREDICTION_MODE best_mode,
+    MV_REFERENCE_FRAME best_ref_frame,
+    INTERP_FILTER best_pred_filter,
+    uint8_t best_mode_skip_txfm) {
+  ctx_den->zero_last_cost_orig = zero_last_cost_orig;
+  ctx_den->ref_frame_cost = ref_frame_cost;
+  ctx_den->frame_mv = frame_mv;
+  ctx_den->reuse_inter_pred = reuse_inter_pred;
+  ctx_den->best_tx_size = best_tx_size;
+  ctx_den->best_mode = best_mode;
+  ctx_den->best_ref_frame = best_ref_frame;
+  ctx_den->best_pred_filter = best_pred_filter;
+  ctx_den->best_mode_skip_txfm = best_mode_skip_txfm;
+}
+
+static void recheck_zeromv_after_denoising(
+    VP9_COMP *cpi, MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
+    VP9_DENOISER_DECISION decision, VP9_PICKMODE_CTX_DEN *ctx_den,
+    struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_COST *best_rdc, BLOCK_SIZE bsize,
+    int mi_row, int mi_col) {
+  // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
+  // denoised result. Only do this under noise conditions, and if rdcost of
+  // ZEROMV onoriginal source is not significantly higher than rdcost of best
+  // mode.
+  if (cpi->noise_estimate.enabled &&
+      cpi->noise_estimate.level > kLow &&
+      ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
+      ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
+       (ctx_den->best_ref_frame == GOLDEN_FRAME &&
+        decision == FILTER_ZEROMV_BLOCK))) {
+    // Check if we should pick ZEROMV on denoised signal.
+    int rate = 0;
+    int64_t dist = 0;
+    uint32_t var_y = UINT_MAX;
+    uint32_t sse_y = UINT_MAX;
+    RD_COST this_rdc;
+    mi->mode = ZEROMV;
+    mi->ref_frame[0] = LAST_FRAME;
+    mi->ref_frame[1] = NONE;
+    mi->mv[0].as_int = 0;
+    mi->interp_filter = EIGHTTAP;
+    xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0];
+    vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
+    model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
+    this_rdc.rate = rate + ctx_den->ref_frame_cost[LAST_FRAME] +
+        cpi->inter_mode_cost[x->mbmi_ext->mode_context[LAST_FRAME]]
+                            [INTER_OFFSET(ZEROMV)];
+    this_rdc.dist = dist;
+    this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, rate, dist);
+    // Switch to ZEROMV if the rdcost for ZEROMV on denoised source
+    // is lower than best_ref mode (on original source).
+    if (this_rdc.rdcost > best_rdc->rdcost) {
+      this_rdc = *best_rdc;
+      mi->mode = ctx_den->best_mode;
+      mi->ref_frame[0] = ctx_den->best_ref_frame;
+      mi->interp_filter = ctx_den->best_pred_filter;
+      if (ctx_den->best_ref_frame == INTRA_FRAME)
+        mi->mv[0].as_int = INVALID_MV;
+      else if (ctx_den->best_ref_frame == GOLDEN_FRAME) {
+        mi->mv[0].as_int = ctx_den->frame_mv[ctx_den->best_mode]
+                                            [ctx_den->best_ref_frame].as_int;
+        if (ctx_den->reuse_inter_pred) {
+          xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0];
+          vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
+        }
+      }
+      mi->tx_size = ctx_den->best_tx_size;
+      x->skip_txfm[0] = ctx_den->best_mode_skip_txfm;
+    } else {
+      ctx_den->best_ref_frame = LAST_FRAME;
+      *best_rdc = this_rdc;
+    }
+  }
+}
+#endif  // CONFIG_VP9_TEMPORAL_DENOISING
+
+static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low,
+                                              int mi_row, int mi_col,
+                                              BLOCK_SIZE bsize) {
+  const int i = (mi_row & 0x7) >> 1;
+  const int j = (mi_col & 0x7) >> 1;
+  int force_skip_low_temp_var = 0;
+  // Set force_skip_low_temp_var based on the block size and block offset.
+  if (bsize == BLOCK_64X64) {
+    force_skip_low_temp_var = variance_low[0];
+  } else if (bsize == BLOCK_64X32) {
+    if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
+      force_skip_low_temp_var = variance_low[1];
+    } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
+      force_skip_low_temp_var = variance_low[2];
+    }
+  } else if (bsize == BLOCK_32X64) {
+    if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
+      force_skip_low_temp_var = variance_low[3];
+    } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
+      force_skip_low_temp_var = variance_low[4];
+    }
+  } else if (bsize == BLOCK_32X32) {
+    if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
+      force_skip_low_temp_var = variance_low[5];
+    } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
+      force_skip_low_temp_var = variance_low[6];
+    } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
+      force_skip_low_temp_var = variance_low[7];
+    } else if ((mi_col & 0x7) && (mi_row & 0x7)) {
+      force_skip_low_temp_var = variance_low[8];
+    }
+  } else if (bsize == BLOCK_16X16) {
+    force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
+  } else if (bsize == BLOCK_32X16) {
+    // The col shift index for the second 16x16 block.
+    const int j2 = ((mi_col + 2) & 0x7) >> 1;
+    // Only if each 16x16 block inside has low temporal variance.
+    force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
+                              variance_low[pos_shift_16x16[i][j2]];
+  } else if (bsize == BLOCK_16X32) {
+    // The row shift index for the second 16x16 block.
+    const int i2 = ((mi_row + 2) & 0x7) >> 1;
+    force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
+                              variance_low[pos_shift_16x16[i2][j]];
+  }
+  return force_skip_low_temp_var;
+}
+
+void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
+                         TileDataEnc *tile_data,
+                         int mi_row, int mi_col, RD_COST *rd_cost,
+                         BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
+  VP9_COMMON *const cm = &cpi->common;
+  SPEED_FEATURES *const sf = &cpi->sf;
+  const SVC *const svc = &cpi->svc;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  PREDICTION_MODE best_mode = ZEROMV;
+  MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
+  MV_REFERENCE_FRAME usable_ref_frame;
+  TX_SIZE best_tx_size = TX_SIZES;
+  INTERP_FILTER best_pred_filter = EIGHTTAP;
+  int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
+  struct buf_2d yv12_mb[4][MAX_MB_PLANE];
+  static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
+                                    VP9_ALT_FLAG };
+  RD_COST this_rdc, best_rdc;
+  uint8_t skip_txfm = SKIP_TXFM_NONE, best_mode_skip_txfm = SKIP_TXFM_NONE;
+  // var_y and sse_y are saved to be used in skipping checking
+  unsigned int var_y = UINT_MAX;
+  unsigned int sse_y = UINT_MAX;
+  const int intra_cost_penalty = set_intra_cost_penalty(cpi, bsize);
+  int64_t inter_mode_thresh = RDCOST(x->rdmult, x->rddiv,
+                                           intra_cost_penalty, 0);
+  const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize];
+  const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
+  INTERP_FILTER filter_ref;
+  const int bsl = mi_width_log2_lookup[bsize];
+  const int pred_filter_search = cm->interp_filter == SWITCHABLE ?
+      (((mi_row + mi_col) >> bsl) +
+       get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
+  int const_motion[MAX_REF_FRAMES] = { 0 };
+  const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
+  const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
+  // For speed 6, the result of interp filter is reused later in actual encoding
+  // process.
+  // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
+  PRED_BUFFER tmp[4];
+  DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
+#if CONFIG_VP9_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
+#endif
+  struct buf_2d orig_dst = pd->dst;
+  PRED_BUFFER *best_pred = NULL;
+  PRED_BUFFER *this_mode_pred = NULL;
+  const int pixels_in_block = bh * bw;
+  int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
+  int ref_frame_skip_mask = 0;
+  int idx;
+  int best_pred_sad = INT_MAX;
+  int best_early_term = 0;
+  int ref_frame_cost[MAX_REF_FRAMES];
+  int svc_force_zero_mode[3] = {0};
+  int perform_intra_pred = 1;
+  int use_golden_nonzeromv = 1;
+  int force_skip_low_temp_var = 0;
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  VP9_PICKMODE_CTX_DEN ctx_den;
+  int64_t zero_last_cost_orig = INT64_MAX;
+#endif
+
+  init_ref_frame_cost(cm, xd, ref_frame_cost);
+
+  if (reuse_inter_pred) {
+    int i;
+    for (i = 0; i < 3; i++) {
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (cm->use_highbitdepth)
+        tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
+      else
+        tmp[i].data = &pred_buf[pixels_in_block * i];
+#else
+      tmp[i].data = &pred_buf[pixels_in_block * i];
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      tmp[i].stride = bw;
+      tmp[i].in_use = 0;
+    }
+    tmp[3].data = pd->dst.buf;
+    tmp[3].stride = pd->dst.stride;
+    tmp[3].in_use = 0;
+  }
+
+  x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
+  x->skip = 0;
+
+  // Instead of using vp9_get_pred_context_switchable_interp(xd) to assign
+  // filter_ref, we use a less strict condition on assigning filter_ref.
+  // This is to reduce the probabily of entering the flow of not assigning
+  // filter_ref and then skip filter search.
+  if (xd->above_mi && is_inter_block(xd->above_mi))
+    filter_ref = xd->above_mi->interp_filter;
+  else if (xd->left_mi && is_inter_block(xd->left_mi))
+    filter_ref = xd->left_mi->interp_filter;
+  else
+    filter_ref = cm->interp_filter;
+
+  // initialize mode decisions
+  vp9_rd_cost_reset(&best_rdc);
+  vp9_rd_cost_reset(rd_cost);
+  mi->sb_type = bsize;
+  mi->ref_frame[0] = NONE;
+  mi->ref_frame[1] = NONE;
+  mi->tx_size = VPXMIN(max_txsize_lookup[bsize],
+                       tx_mode_to_biggest_tx_size[cm->tx_mode]);
+
+  if (sf->short_circuit_flat_blocks) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+      x->source_variance = vp9_high_get_sby_perpixel_variance(
+          cpi, &x->plane[0].src, bsize, xd->bd);
+    else
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      x->source_variance =
+          vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
+  }
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0 &&
+      cpi->denoiser.denoising_level > kDenLowLow) {
+    vp9_denoiser_reset_frame_stats(ctx);
+  }
+#endif
+
+  if (cpi->rc.frames_since_golden == 0 && !cpi->use_svc) {
+    usable_ref_frame = LAST_FRAME;
+  } else {
+    usable_ref_frame = GOLDEN_FRAME;
+  }
+
+  // For svc mode, on spatial_layer_id > 0: if the reference has different scale
+  // constrain the inter mode to only test zero motion.
+  if (cpi->use_svc &&
+      svc ->force_zero_mode_spatial_ref &&
+      cpi->svc.spatial_layer_id > 0) {
+    if (cpi->ref_frame_flags & flag_list[LAST_FRAME]) {
+      struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
+      if (vp9_is_scaled(sf))
+        svc_force_zero_mode[LAST_FRAME - 1] = 1;
+    }
+    if (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) {
+      struct scale_factors *const sf = &cm->frame_refs[GOLDEN_FRAME - 1].sf;
+      if (vp9_is_scaled(sf))
+        svc_force_zero_mode[GOLDEN_FRAME - 1] = 1;
+    }
+  }
+
+  if (cpi->sf.short_circuit_low_temp_var) {
+    force_skip_low_temp_var =
+        get_force_skip_low_temp_var(&x->variance_low[0], mi_row, mi_col, bsize);
+  }
+
+  if (!((cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) &&
+      !svc_force_zero_mode[GOLDEN_FRAME - 1] && !force_skip_low_temp_var))
+    use_golden_nonzeromv = 0;
+
+  for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
+    find_predictors(cpi, x, ref_frame, frame_mv, const_motion,
+                    &ref_frame_skip_mask, flag_list, tile_data, mi_row, mi_col,
+                    yv12_mb, bsize, force_skip_low_temp_var);
+  }
+
+  for (idx = 0; idx < RT_INTER_MODES; ++idx) {
+    int rate_mv = 0;
+    int mode_rd_thresh;
+    int mode_index;
+    int i;
+    int64_t this_sse;
+    int is_skippable;
+    int this_early_term = 0;
+    PREDICTION_MODE this_mode = ref_mode_set[idx].pred_mode;
+
+    if (cpi->use_svc)
+      this_mode = ref_mode_set_svc[idx].pred_mode;
+
+    if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
+        this_mode != NEARESTMV) {
+      continue;
+    }
+
+    if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode)))
+      continue;
+
+    ref_frame = ref_mode_set[idx].ref_frame;
+    if (cpi->use_svc) {
+      ref_frame = ref_mode_set_svc[idx].ref_frame;
+    }
+
+    if (!(cpi->ref_frame_flags & flag_list[ref_frame]))
+      continue;
+
+    if (const_motion[ref_frame] && this_mode == NEARMV)
+      continue;
+
+    // Skip non-zeromv mode search for golden frame if force_skip_low_temp_var
+    // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
+    // later.
+    if (force_skip_low_temp_var && ref_frame == GOLDEN_FRAME &&
+        frame_mv[this_mode][ref_frame].as_int != 0) {
+      continue;
+    }
+
+    if (cpi->sf.short_circuit_low_temp_var == 2 &&
+        force_skip_low_temp_var && ref_frame == LAST_FRAME &&
+        this_mode == NEWMV) {
+      continue;
+    }
+
+    if (cpi->use_svc) {
+      if (svc_force_zero_mode[ref_frame - 1] &&
+          frame_mv[this_mode][ref_frame].as_int != 0)
+        continue;
+    }
+
+    if (!force_skip_low_temp_var &&
+        !(frame_mv[this_mode][ref_frame].as_int == 0 &&
+          ref_frame == LAST_FRAME)) {
+      i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
+      if ((cpi->ref_frame_flags & flag_list[i]) && sf->reference_masking)
+        if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
+          ref_frame_skip_mask |= (1 << ref_frame);
+    }
+    if (ref_frame_skip_mask & (1 << ref_frame))
+      continue;
+
+    // Select prediction reference frames.
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
+
+    mi->ref_frame[0] = ref_frame;
+    set_ref_ptrs(cm, xd, ref_frame, NONE);
+
+    mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
+    mode_rd_thresh = best_mode_skip_txfm ?
+            rd_threshes[mode_index] << 1 : rd_threshes[mode_index];
+    if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
+                            rd_thresh_freq_fact[mode_index]))
+      continue;
+
+    if (this_mode == NEWMV) {
+      if (ref_frame > LAST_FRAME &&
+          !cpi->use_svc &&
+          cpi->oxcf.rc_mode == VPX_CBR) {
+        int tmp_sad;
+        uint32_t dis;
+        int cost_list[5];
+
+        if (bsize < BLOCK_16X16)
+          continue;
+
+        tmp_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
+
+        if (tmp_sad > x->pred_mv_sad[LAST_FRAME])
+          continue;
+        if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad)
+          continue;
+
+        frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
+        rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
+          &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
+          x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+        frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
+        frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
+
+        cpi->find_fractional_mv_step(x, &frame_mv[NEWMV][ref_frame].as_mv,
+          &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
+          cpi->common.allow_high_precision_mv,
+          x->errorperbit,
+          &cpi->fn_ptr[bsize],
+          cpi->sf.mv.subpel_force_stop,
+          cpi->sf.mv.subpel_iters_per_step,
+          cond_cost_list(cpi, cost_list),
+          x->nmvjointcost, x->mvcost, &dis,
+          &x->pred_sse[ref_frame], NULL, 0, 0);
+      } else if (svc->use_base_mv && svc->spatial_layer_id) {
+        if (frame_mv[NEWMV][ref_frame].as_int != INVALID_MV &&
+            frame_mv[NEWMV][ref_frame].as_int != 0) {
+          const int pre_stride = xd->plane[0].pre[0].stride;
+          int base_mv_sad = INT_MAX;
+          const uint8_t * const pre_buf = xd->plane[0].pre[0].buf +
+              (frame_mv[NEWMV][ref_frame].as_mv.row >> 3) * pre_stride +
+              (frame_mv[NEWMV][ref_frame].as_mv.col >> 3);
+          base_mv_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
+                                       x->plane[0].src.stride,
+                                       pre_buf, pre_stride);
+
+          // TODO(wonkap): make the decision to use base layer mv on RD;
+          // not just SAD.
+          if (base_mv_sad < x->pred_mv_sad[ref_frame]) {
+            // Base layer mv is good.
+            if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
+                &frame_mv[NEWMV][ref_frame], &rate_mv, best_rdc.rdcost, 1)) {
+                continue;
+            }
+          } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
+            &frame_mv[NEWMV][ref_frame], &rate_mv, best_rdc.rdcost, 0)) {
+            continue;
+          }
+        } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
+          &frame_mv[NEWMV][ref_frame], &rate_mv, best_rdc.rdcost, 0)) {
+          continue;
+        }
+      } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
+        &frame_mv[NEWMV][ref_frame], &rate_mv, best_rdc.rdcost, 0)) {
+        continue;
+      }
+    }
+
+    // If use_golden_nonzeromv is false, NEWMV mode is skipped for golden, no
+    // need to compute best_pred_sad which is only used to skip golden NEWMV.
+    if (use_golden_nonzeromv && this_mode == NEWMV &&
+        ref_frame == LAST_FRAME &&
+        frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
+      const int pre_stride = xd->plane[0].pre[0].stride;
+      const uint8_t * const pre_buf = xd->plane[0].pre[0].buf +
+          (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
+          (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
+      best_pred_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
+                                   x->plane[0].src.stride,
+                                   pre_buf, pre_stride);
+      x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
+    }
+
+    if (this_mode != NEARESTMV &&
+        frame_mv[this_mode][ref_frame].as_int ==
+            frame_mv[NEARESTMV][ref_frame].as_int)
+      continue;
+
+    mi->mode = this_mode;
+    mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
+
+    // Search for the best prediction filter type, when the resulting
+    // motion vector is at sub-pixel accuracy level for luma component, i.e.,
+    // the last three bits are all zeros.
+    if (reuse_inter_pred) {
+      if (!this_mode_pred) {
+        this_mode_pred = &tmp[3];
+      } else {
+        this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
+        pd->dst.buf = this_mode_pred->data;
+        pd->dst.stride = bw;
+      }
+    }
+
+    if ((this_mode == NEWMV || filter_ref == SWITCHABLE) && pred_filter_search
+        && (ref_frame == LAST_FRAME ||
+           (ref_frame == GOLDEN_FRAME &&
+           (cpi->use_svc || cpi->oxcf.rc_mode == VPX_VBR))) &&
+           (((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) != 0)) {
+      int pf_rate[3];
+      int64_t pf_dist[3];
+      unsigned int pf_var[3];
+      unsigned int pf_sse[3];
+      TX_SIZE pf_tx_size[3];
+      int64_t best_cost = INT64_MAX;
+      INTERP_FILTER best_filter = SWITCHABLE, filter;
+      PRED_BUFFER *current_pred = this_mode_pred;
+
+      for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
+        int64_t cost;
+        mi->interp_filter = filter;
+        vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
+        model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
+                          &pf_var[filter], &pf_sse[filter]);
+        pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
+        cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
+        pf_tx_size[filter] = mi->tx_size;
+        if (cost < best_cost) {
+          best_filter = filter;
+          best_cost = cost;
+          skip_txfm = x->skip_txfm[0];
+
+          if (reuse_inter_pred) {
+            if (this_mode_pred != current_pred) {
+              free_pred_buffer(this_mode_pred);
+              this_mode_pred = current_pred;
+            }
+            current_pred = &tmp[get_pred_buffer(tmp, 3)];
+            pd->dst.buf = current_pred->data;
+            pd->dst.stride = bw;
+          }
+        }
+      }
+
+      if (reuse_inter_pred && this_mode_pred != current_pred)
+        free_pred_buffer(current_pred);
+
+      mi->interp_filter = best_filter;
+      mi->tx_size = pf_tx_size[best_filter];
+      this_rdc.rate = pf_rate[best_filter];
+      this_rdc.dist = pf_dist[best_filter];
+      var_y = pf_var[best_filter];
+      sse_y = pf_sse[best_filter];
+      x->skip_txfm[0] = skip_txfm;
+      if (reuse_inter_pred) {
+        pd->dst.buf = this_mode_pred->data;
+        pd->dst.stride = this_mode_pred->stride;
+      }
+    } else {
+      // TODO(jackychen): the low-bitdepth condition causes a segfault in
+      // high-bitdepth builds.
+      // https://bugs.chromium.org/p/webm/issues/detail?id=1250
+#if CONFIG_VP9_HIGHBITDEPTH
+      const int large_block = bsize > BLOCK_32X32;
+#else
+      const int large_block = bsize >= BLOCK_32X32;
+#endif
+      mi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
+      vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
+
+      // For large partition blocks, extra testing is done.
+      if (cpi->oxcf.rc_mode == VPX_CBR && large_block &&
+          !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) &&
+          cm->base_qindex) {
+        model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
+                                &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
+                                &this_early_term);
+      } else {
+        model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
+                          &var_y, &sse_y);
+      }
+    }
+
+    if (!this_early_term) {
+      this_sse = (int64_t)sse_y;
+      block_yrd(cpi, x, &this_rdc, &is_skippable, &this_sse, bsize,
+                VPXMIN(mi->tx_size, TX_16X16));
+      x->skip_txfm[0] = is_skippable;
+      if (is_skippable) {
+        this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+      } else {
+        if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
+            RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
+          this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
+        } else {
+          this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+          this_rdc.dist = this_sse;
+          x->skip_txfm[0] = SKIP_TXFM_AC_DC;
+        }
+      }
+
+      if (cm->interp_filter == SWITCHABLE) {
+        if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07)
+          this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
+      }
+    } else {
+      this_rdc.rate += cm->interp_filter == SWITCHABLE ?
+          vp9_get_switchable_rate(cpi, xd) : 0;
+      this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+    }
+
+    if (x->color_sensitivity[0] || x->color_sensitivity[1]) {
+      RD_COST rdc_uv;
+      const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]);
+      if (x->color_sensitivity[0])
+        vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
+      if (x->color_sensitivity[1])
+        vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
+      model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2);
+      this_rdc.rate += rdc_uv.rate;
+      this_rdc.dist += rdc_uv.dist;
+    }
+
+    this_rdc.rate += rate_mv;
+    this_rdc.rate +=
+        cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]][INTER_OFFSET(
+            this_mode)];
+    this_rdc.rate += ref_frame_cost[ref_frame];
+    this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
+
+    // Bias against non-zero motion
+    if (cpi->oxcf.rc_mode == VPX_CBR &&
+        cpi->oxcf.speed >= 5 &&
+        cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
+        !x->sb_is_skin) {
+      vp9_large_block_mv_bias(&cpi->noise_estimate, &this_rdc, bsize,
+                              frame_mv[this_mode][ref_frame].as_mv.row,
+                              frame_mv[this_mode][ref_frame].as_mv.col,
+                              ref_frame == LAST_FRAME);
+    }
+
+    // Skipping checking: test to see if this block can be reconstructed by
+    // prediction only.
+    if (cpi->allow_encode_breakout) {
+      encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
+                           var_y, sse_y, yv12_mb, &this_rdc.rate,
+                           &this_rdc.dist);
+      if (x->skip) {
+        this_rdc.rate += rate_mv;
+        this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate,
+                                 this_rdc.dist);
+      }
+    }
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+    if (cpi->oxcf.noise_sensitivity > 0 &&
+        cpi->denoiser.denoising_level > kDenLowLow) {
+      vp9_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
+      // Keep track of zero_last cost.
+      if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0)
+        zero_last_cost_orig = this_rdc.rdcost;
+    }
+#else
+    (void)ctx;
+#endif
+
+    if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
+      best_rdc = this_rdc;
+      best_mode = this_mode;
+      best_pred_filter = mi->interp_filter;
+      best_tx_size = mi->tx_size;
+      best_ref_frame = ref_frame;
+      best_mode_skip_txfm = x->skip_txfm[0];
+      best_early_term = this_early_term;
+
+      if (reuse_inter_pred) {
+        free_pred_buffer(best_pred);
+        best_pred = this_mode_pred;
+      }
+    } else {
+      if (reuse_inter_pred)
+        free_pred_buffer(this_mode_pred);
+    }
+
+    if (x->skip)
+      break;
+
+    // If early termination flag is 1 and at least 2 modes are checked,
+    // the mode search is terminated.
+    if (best_early_term && idx > 0) {
+      x->skip = 1;
+      break;
+    }
+  }
+
+  mi->mode          = best_mode;
+  mi->interp_filter = best_pred_filter;
+  mi->tx_size       = best_tx_size;
+  mi->ref_frame[0]  = best_ref_frame;
+  mi->mv[0].as_int  = frame_mv[best_mode][best_ref_frame].as_int;
+  xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int;
+  x->skip_txfm[0] = best_mode_skip_txfm;
+
+  // For spatial enhancemanent layer: perform intra prediction only if base
+  // layer is chosen as the reference. Always perform intra prediction if
+  // LAST is the only reference or is_key_frame is set.
+  if (cpi->svc.spatial_layer_id) {
+    perform_intra_pred =
+        cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame ||
+        !(cpi->ref_frame_flags & flag_list[GOLDEN_FRAME])  ||
+        (!cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame
+            && svc_force_zero_mode[best_ref_frame - 1]);
+    inter_mode_thresh = (inter_mode_thresh << 1) + inter_mode_thresh;
+  }
+  // Perform intra prediction search, if the best SAD is above a certain
+  // threshold.
+  if ((!force_skip_low_temp_var || bsize < BLOCK_32X32) &&
+      perform_intra_pred &&
+      (best_rdc.rdcost == INT64_MAX ||
+       (!x->skip && best_rdc.rdcost > inter_mode_thresh &&
+        bsize <= cpi->sf.max_intra_bsize))) {
+    struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
+    int i;
+    TX_SIZE best_intra_tx_size = TX_SIZES;
+    TX_SIZE intra_tx_size =
+        VPXMIN(max_txsize_lookup[bsize],
+               tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+    if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16)
+      intra_tx_size = TX_16X16;
+
+    if (reuse_inter_pred && best_pred != NULL) {
+      if (best_pred->data == orig_dst.buf) {
+        this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
+#if CONFIG_VP9_HIGHBITDEPTH
+        if (cm->use_highbitdepth)
+          vpx_highbd_convolve_copy(best_pred->data, best_pred->stride,
+                                   this_mode_pred->data, this_mode_pred->stride,
+                                   NULL, 0, NULL, 0, bw, bh, xd->bd);
+        else
+          vpx_convolve_copy(best_pred->data, best_pred->stride,
+                          this_mode_pred->data, this_mode_pred->stride,
+                          NULL, 0, NULL, 0, bw, bh);
+#else
+        vpx_convolve_copy(best_pred->data, best_pred->stride,
+                          this_mode_pred->data, this_mode_pred->stride,
+                          NULL, 0, NULL, 0, bw, bh);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        best_pred = this_mode_pred;
+      }
+    }
+    pd->dst = orig_dst;
+
+    for (i = 0; i < 4; ++i) {
+      const PREDICTION_MODE this_mode = intra_mode_list[i];
+      THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
+      int mode_rd_thresh = rd_threshes[mode_index];
+      if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
+          this_mode != DC_PRED) {
+        continue;
+      }
+
+      if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
+        continue;
+
+      if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
+                              rd_thresh_freq_fact[mode_index]))
+        continue;
+
+      mi->mode = this_mode;
+      mi->ref_frame[0] = INTRA_FRAME;
+      this_rdc.dist = this_rdc.rate = 0;
+      args.mode = this_mode;
+      args.skippable = 1;
+      args.rdc = &this_rdc;
+      mi->tx_size = intra_tx_size;
+      vp9_foreach_transformed_block_in_plane(xd, bsize, 0,
+                                             estimate_block_intra, &args);
+      // Check skip cost here since skippable is not set for for uv, this
+      // mirrors the behavior used by inter
+      if (args.skippable) {
+        x->skip_txfm[0] = SKIP_TXFM_AC_DC;
+        this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
+      } else {
+        x->skip_txfm[0] = SKIP_TXFM_NONE;
+        this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
+      }
+      // Inter and intra RD will mismatch in scale for non-screen content.
+      if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
+        if (x->color_sensitivity[0])
+          vp9_foreach_transformed_block_in_plane(xd, bsize, 1,
+                                                 estimate_block_intra, &args);
+        if (x->color_sensitivity[1])
+          vp9_foreach_transformed_block_in_plane(xd, bsize, 2,
+                                                 estimate_block_intra, &args);
+      }
+      this_rdc.rate += cpi->mbmode_cost[this_mode];
+      this_rdc.rate += ref_frame_cost[INTRA_FRAME];
+      this_rdc.rate += intra_cost_penalty;
+      this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                               this_rdc.rate, this_rdc.dist);
+
+      if (this_rdc.rdcost < best_rdc.rdcost) {
+        best_rdc = this_rdc;
+        best_mode = this_mode;
+        best_intra_tx_size = mi->tx_size;
+        best_ref_frame = INTRA_FRAME;
+        mi->uv_mode = this_mode;
+        mi->mv[0].as_int = INVALID_MV;
+        best_mode_skip_txfm = x->skip_txfm[0];
+      }
+    }
+
+    // Reset mb_mode_info to the best inter mode.
+    if (best_ref_frame != INTRA_FRAME) {
+      mi->tx_size = best_tx_size;
+    } else {
+      mi->tx_size = best_intra_tx_size;
+    }
+  }
+
+  pd->dst = orig_dst;
+  mi->mode = best_mode;
+  mi->ref_frame[0] = best_ref_frame;
+  x->skip_txfm[0] = best_mode_skip_txfm;
+
+  if (!is_inter_block(mi)) {
+    mi->interp_filter = SWITCHABLE_FILTERS;
+  }
+
+  if (reuse_inter_pred && best_pred != NULL) {
+    if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (cm->use_highbitdepth)
+        vpx_highbd_convolve_copy(best_pred->data, best_pred->stride,
+                                 pd->dst.buf, pd->dst.stride, NULL, 0,
+                                 NULL, 0, bw, bh, xd->bd);
+      else
+        vpx_convolve_copy(best_pred->data, best_pred->stride,
+                          pd->dst.buf, pd->dst.stride, NULL, 0,
+                          NULL, 0, bw, bh);
+#else
+      vpx_convolve_copy(best_pred->data, best_pred->stride,
+                        pd->dst.buf, pd->dst.stride, NULL, 0,
+                        NULL, 0, bw, bh);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    }
+  }
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0 &&
+      cpi->resize_pending == 0 &&
+      cpi->denoiser.denoising_level > kDenLowLow &&
+      cpi->denoiser.reset == 0) {
+    VP9_DENOISER_DECISION decision = COPY_BLOCK;
+    vp9_pickmode_ctx_den_update(&ctx_den, zero_last_cost_orig, ref_frame_cost,
+                                frame_mv, reuse_inter_pred, best_tx_size,
+                                best_mode, best_ref_frame, best_pred_filter,
+                                best_mode_skip_txfm);
+    vp9_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision);
+    recheck_zeromv_after_denoising(cpi, mi, x, xd, decision, &ctx_den, yv12_mb,
+                                   &best_rdc, bsize, mi_row, mi_col);
+    best_ref_frame = ctx_den.best_ref_frame;
+  }
+#endif
+
+  if (cpi->sf.adaptive_rd_thresh) {
+    THR_MODES best_mode_idx = mode_idx[best_ref_frame][mode_offset(mi->mode)];
+
+    if (best_ref_frame == INTRA_FRAME) {
+      // Only consider the modes that are included in the intra_mode_list.
+      int intra_modes = sizeof(intra_mode_list)/sizeof(PREDICTION_MODE);
+      int i;
+
+      // TODO(yunqingwang): Check intra mode mask and only update freq_fact
+      // for those valid modes.
+      for (i = 0; i < intra_modes; i++) {
+        update_thresh_freq_fact(cpi, tile_data, bsize, INTRA_FRAME,
+                                best_mode_idx, intra_mode_list[i]);
+      }
+    } else {
+      for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
+        PREDICTION_MODE this_mode;
+        if (best_ref_frame != ref_frame) continue;
+        for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
+          update_thresh_freq_fact(cpi, tile_data, bsize, ref_frame,
+                                  best_mode_idx, this_mode);
+        }
+      }
+    }
+  }
+
+  *rd_cost = best_rdc;
+}
+
+void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
+                                int mi_row, int mi_col, RD_COST *rd_cost,
+                                BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
+  VP9_COMMON *const cm = &cpi->common;
+  SPEED_FEATURES *const sf = &cpi->sf;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  const struct segmentation *const seg = &cm->seg;
+  MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
+  MV_REFERENCE_FRAME best_ref_frame = NONE;
+  unsigned char segment_id = mi->segment_id;
+  struct buf_2d yv12_mb[4][MAX_MB_PLANE];
+  static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
+                                    VP9_ALT_FLAG };
+  int64_t best_rd = INT64_MAX;
+  b_mode_info bsi[MAX_REF_FRAMES][4];
+  int ref_frame_skip_mask = 0;
+  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
+  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
+  int idx, idy;
+
+  x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
+  ctx->pred_pixel_ready = 0;
+
+  for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
+    const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
+    int_mv dummy_mv[2];
+    x->pred_mv_sad[ref_frame] = INT_MAX;
+
+    if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
+      int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame];
+      const struct scale_factors *const sf =
+                             &cm->frame_refs[ref_frame - 1].sf;
+      vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
+                           sf, sf);
+      vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame,
+                       candidates, mi_row, mi_col, mbmi_ext->mode_context);
+
+      vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
+                            &dummy_mv[0], &dummy_mv[1]);
+    } else {
+      ref_frame_skip_mask |= (1 << ref_frame);
+    }
+  }
+
+  mi->sb_type = bsize;
+  mi->tx_size = TX_4X4;
+  mi->uv_mode = DC_PRED;
+  mi->ref_frame[0] = LAST_FRAME;
+  mi->ref_frame[1] = NONE;
+  mi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
+                                                      : cm->interp_filter;
+
+  for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
+    int64_t this_rd = 0;
+    int plane;
+
+    if (ref_frame_skip_mask & (1 << ref_frame))
+      continue;
+
+#if CONFIG_BETTER_HW_COMPATIBILITY
+    if ((bsize == BLOCK_8X4 || bsize == BLOCK_4X8) &&
+        ref_frame > INTRA_FRAME &&
+        vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
+      continue;
+#endif
+
+    // TODO(jingning, agrange): Scaling reference frame not supported for
+    // sub8x8 blocks. Is this supported now?
+    if (ref_frame > INTRA_FRAME &&
+        vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
+      continue;
+
+    // If the segment reference frame feature is enabled....
+    // then do nothing if the current ref frame is not allowed..
+    if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
+        get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
+      continue;
+
+    mi->ref_frame[0] = ref_frame;
+    x->skip = 0;
+    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
+
+    // Select prediction reference frames.
+    for (plane = 0; plane < MAX_MB_PLANE; plane++)
+      xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
+
+    for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
+      for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
+        int_mv b_mv[MB_MODE_COUNT];
+        int64_t b_best_rd = INT64_MAX;
+        const int i = idy * 2 + idx;
+        PREDICTION_MODE this_mode;
+        RD_COST this_rdc;
+        unsigned int var_y, sse_y;
+
+        struct macroblock_plane *p = &x->plane[0];
+        struct macroblockd_plane *pd = &xd->plane[0];
+
+        const struct buf_2d orig_src = p->src;
+        const struct buf_2d orig_dst = pd->dst;
+        struct buf_2d orig_pre[2];
+        memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
+
+        // set buffer pointers for sub8x8 motion search.
+        p->src.buf =
+            &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
+        pd->dst.buf =
+            &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
+        pd->pre[0].buf =
+            &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8,
+                                                    i, pd->pre[0].stride)];
+
+        b_mv[ZEROMV].as_int = 0;
+        b_mv[NEWMV].as_int = INVALID_MV;
+        vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col,
+                                      &b_mv[NEARESTMV],
+                                      &b_mv[NEARMV],
+                                      mbmi_ext->mode_context);
+
+        for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
+          int b_rate = 0;
+          xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
+
+          if (this_mode == NEWMV) {
+            const int step_param = cpi->sf.mv.fullpel_search_step_param;
+            MV mvp_full;
+            MV tmp_mv;
+            int cost_list[5];
+            const int tmp_col_min = x->mv_col_min;
+            const int tmp_col_max = x->mv_col_max;
+            const int tmp_row_min = x->mv_row_min;
+            const int tmp_row_max = x->mv_row_max;
+            uint32_t dummy_dist;
+
+            if (i == 0) {
+              mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
+              mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
+            } else {
+              mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
+              mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
+            }
+
+            vp9_set_mv_search_range(x, &mbmi_ext->ref_mvs[0]->as_mv);
+
+            vp9_full_pixel_search(
+                cpi, x, bsize, &mvp_full, step_param, x->sadperbit4,
+                cond_cost_list(cpi, cost_list),
+                &mbmi_ext->ref_mvs[ref_frame][0].as_mv, &tmp_mv,
+                INT_MAX, 0);
+
+            x->mv_col_min = tmp_col_min;
+            x->mv_col_max = tmp_col_max;
+            x->mv_row_min = tmp_row_min;
+            x->mv_row_max = tmp_row_max;
+
+            // calculate the bit cost on motion vector
+            mvp_full.row = tmp_mv.row * 8;
+            mvp_full.col = tmp_mv.col * 8;
+
+            b_rate += vp9_mv_bit_cost(&mvp_full,
+                                      &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
+                                      x->nmvjointcost, x->mvcost,
+                                      MV_COST_WEIGHT);
+
+            b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
+                                          [INTER_OFFSET(NEWMV)];
+            if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd)
+              continue;
+
+            cpi->find_fractional_mv_step(x, &tmp_mv,
+                                         &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
+                                         cpi->common.allow_high_precision_mv,
+                                         x->errorperbit,
+                                         &cpi->fn_ptr[bsize],
+                                         cpi->sf.mv.subpel_force_stop,
+                                         cpi->sf.mv.subpel_iters_per_step,
+                                         cond_cost_list(cpi, cost_list),
+                                         x->nmvjointcost, x->mvcost,
+                                         &dummy_dist,
+                                         &x->pred_sse[ref_frame], NULL, 0, 0);
+
+            xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
+          } else {
+            b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
+                                          [INTER_OFFSET(this_mode)];
+          }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            vp9_highbd_build_inter_predictor(pd->pre[0].buf, pd->pre[0].stride,
+                                    pd->dst.buf, pd->dst.stride,
+                                    &xd->mi[0]->bmi[i].as_mv[0].as_mv,
+                                    &xd->block_refs[0]->sf,
+                                    4 * num_4x4_blocks_wide,
+                                    4 * num_4x4_blocks_high, 0,
+                                    vp9_filter_kernels[mi->interp_filter],
+                                    MV_PRECISION_Q3,
+                                    mi_col * MI_SIZE + 4 * (i & 0x01),
+                                    mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
+          } else {
+#endif
+            vp9_build_inter_predictor(pd->pre[0].buf, pd->pre[0].stride,
+                                     pd->dst.buf, pd->dst.stride,
+                                     &xd->mi[0]->bmi[i].as_mv[0].as_mv,
+                                     &xd->block_refs[0]->sf,
+                                     4 * num_4x4_blocks_wide,
+                                     4 * num_4x4_blocks_high, 0,
+                                     vp9_filter_kernels[mi->interp_filter],
+                                     MV_PRECISION_Q3,
+                                     mi_col * MI_SIZE + 4 * (i & 0x01),
+                                     mi_row * MI_SIZE + 4 * (i >> 1));
+
+#if CONFIG_VP9_HIGHBITDEPTH
+          }
+#endif
+
+          model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
+                            &var_y, &sse_y);
+
+          this_rdc.rate += b_rate;
+          this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
+                                   this_rdc.rate, this_rdc.dist);
+          if (this_rdc.rdcost < b_best_rd) {
+            b_best_rd = this_rdc.rdcost;
+            bsi[ref_frame][i].as_mode = this_mode;
+            bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
+          }
+        }  // mode search
+
+        // restore source and prediction buffer pointers.
+        p->src = orig_src;
+        pd->pre[0] = orig_pre[0];
+        pd->dst = orig_dst;
+        this_rd += b_best_rd;
+
+        xd->mi[0]->bmi[i] = bsi[ref_frame][i];
+        if (num_4x4_blocks_wide > 1)
+          xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
+        if (num_4x4_blocks_high > 1)
+          xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
+      }
+    }  // loop through sub8x8 blocks
+
+    if (this_rd < best_rd) {
+      best_rd = this_rd;
+      best_ref_frame = ref_frame;
+    }
+  }  // reference frames
+
+  mi->tx_size = TX_4X4;
+  mi->ref_frame[0] = best_ref_frame;
+  for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
+    for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
+      const int block = idy * 2 + idx;
+      xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
+      if (num_4x4_blocks_wide > 1)
+        xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
+      if (num_4x4_blocks_high > 1)
+        xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
+    }
+  }
+  mi->mode = xd->mi[0]->bmi[3].as_mode;
+  ctx->mic = *(xd->mi[0]);
+  ctx->mbmi_ext = *x->mbmi_ext;
+  ctx->skip_txfm[0] = SKIP_TXFM_NONE;
+  ctx->skip = 0;
+  // Dummy assignment for speed -5. No effect in speed -6.
+  rd_cost->rdcost = best_rd;
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_pickmode.h b/libvpx/libvpx/vp9/encoder/vp9_pickmode.h
new file mode 100644
index 0000000..a43bb81
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_pickmode.h
@@ -0,0 +1,38 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_PICKMODE_H_
+#define VP9_ENCODER_VP9_PICKMODE_H_
+
+#include "vp9/encoder/vp9_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
+                         BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx);
+
+void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
+                         TileDataEnc *tile_data,
+                         int mi_row, int mi_col, RD_COST *rd_cost,
+                         BLOCK_SIZE bsize,
+                         PICK_MODE_CONTEXT *ctx);
+
+void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
+                                int mi_row, int mi_col, RD_COST *rd_cost,
+                                BLOCK_SIZE bsize,
+                                PICK_MODE_CONTEXT *ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_PICKMODE_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_quantize.c b/libvpx/libvpx/vp9/encoder/vp9_quantize.c
new file mode 100644
index 0000000..d68b684
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_quantize.c
@@ -0,0 +1,388 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_rd.h"
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                       int skip_block,
+                       const int16_t *zbin_ptr, const int16_t *round_ptr,
+                       const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
+                       tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                       const int16_t *dequant_ptr,
+                       uint16_t *eob_ptr,
+                       const int16_t *scan, const int16_t *iscan) {
+  int i, eob = -1;
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+      tmp = (tmp * quant_ptr[rc != 0]) >> 16;
+
+      qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+
+      if (tmp)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_fp_c(const tran_low_t *coeff_ptr,
+                              intptr_t count,
+                              int skip_block,
+                              const int16_t *zbin_ptr,
+                              const int16_t *round_ptr,
+                              const int16_t *quant_ptr,
+                              const int16_t *quant_shift_ptr,
+                              tran_low_t *qcoeff_ptr,
+                              tran_low_t *dqcoeff_ptr,
+                              const int16_t *dequant_ptr,
+                              uint16_t *eob_ptr,
+                              const int16_t *scan,
+                              const int16_t *iscan) {
+  int i;
+  int eob = -1;
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int64_t tmp = abs_coeff + round_ptr[rc != 0];
+      const int abs_qcoeff = (int)((tmp * quant_ptr[rc != 0]) >> 16);
+      qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+      if (abs_qcoeff)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+// TODO(jingning) Refactor this file and combine functions with similar
+// operations.
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block,
+                             const int16_t *zbin_ptr, const int16_t *round_ptr,
+                             const int16_t *quant_ptr,
+                             const int16_t *quant_shift_ptr,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             const int16_t *dequant_ptr,
+                             uint16_t *eob_ptr,
+                             const int16_t *scan, const int16_t *iscan) {
+  int i, eob = -1;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      int tmp = 0;
+      int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff >= (dequant_ptr[rc != 0] >> 2)) {
+        abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+        abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+        tmp = (abs_coeff * quant_ptr[rc != 0]) >> 15;
+        qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+      }
+
+      if (tmp)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_fp_32x32_c(const tran_low_t *coeff_ptr,
+                                    intptr_t n_coeffs, int skip_block,
+                                    const int16_t *zbin_ptr,
+                                    const int16_t *round_ptr,
+                                    const int16_t *quant_ptr,
+                                    const int16_t *quant_shift_ptr,
+                                    tran_low_t *qcoeff_ptr,
+                                    tran_low_t *dqcoeff_ptr,
+                                    const int16_t *dequant_ptr,
+                                    uint16_t *eob_ptr,
+                                    const int16_t *scan, const int16_t *iscan) {
+  int i, eob = -1;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    for (i = 0; i < n_coeffs; i++) {
+      uint32_t abs_qcoeff = 0;
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff >= (dequant_ptr[rc != 0] >> 2)) {
+        const int64_t tmp = abs_coeff
+                           + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+        abs_qcoeff = (uint32_t) ((tmp * quant_ptr[rc != 0]) >> 15);
+        qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+      }
+
+      if (abs_qcoeff)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+void vp9_regular_quantize_b_4x4(MACROBLOCK *x, int plane, int block,
+                                const int16_t *scan, const int16_t *iscan) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vpx_highbd_quantize_b(BLOCK_OFFSET(p->coeff, block),
+                          16, x->skip_block,
+                          p->zbin, p->round, p->quant, p->quant_shift,
+                          BLOCK_OFFSET(p->qcoeff, block),
+                          BLOCK_OFFSET(pd->dqcoeff, block),
+                          pd->dequant, &p->eobs[block],
+                          scan, iscan);
+    return;
+  }
+#endif
+  vpx_quantize_b(BLOCK_OFFSET(p->coeff, block),
+                 16, x->skip_block,
+                 p->zbin, p->round, p->quant, p->quant_shift,
+                 BLOCK_OFFSET(p->qcoeff, block),
+                 BLOCK_OFFSET(pd->dqcoeff, block),
+                 pd->dequant, &p->eobs[block], scan, iscan);
+}
+
+static void invert_quant(int16_t *quant, int16_t *shift, int d) {
+  unsigned t;
+  int l, m;
+  t = d;
+  for (l = 0; t > 1; l++)
+    t >>= 1;
+  m = 1 + (1 << (16 + l)) / d;
+  *quant = (int16_t)(m - (1 << 16));
+  *shift = 1 << (16 - l);
+}
+
+static int get_qzbin_factor(int q, vpx_bit_depth_t bit_depth) {
+  const int quant = vp9_dc_quant(q, 0, bit_depth);
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      return q == 0 ? 64 : (quant < 148 ? 84 : 80);
+    case VPX_BITS_10:
+      return q == 0 ? 64 : (quant < 592 ? 84 : 80);
+    case VPX_BITS_12:
+      return q == 0 ? 64 : (quant < 2368 ? 84 : 80);
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  (void) bit_depth;
+  return q == 0 ? 64 : (quant < 148 ? 84 : 80);
+#endif
+}
+
+void vp9_init_quantizer(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  QUANTS *const quants = &cpi->quants;
+  int i, q, quant;
+
+  for (q = 0; q < QINDEX_RANGE; q++) {
+    const int qzbin_factor = get_qzbin_factor(q, cm->bit_depth);
+    const int qrounding_factor = q == 0 ? 64 : 48;
+
+    for (i = 0; i < 2; ++i) {
+      int qrounding_factor_fp = i == 0 ? 48 : 42;
+      if (q == 0)
+        qrounding_factor_fp = 64;
+
+      // y
+      quant = i == 0 ? vp9_dc_quant(q, cm->y_dc_delta_q, cm->bit_depth)
+                     : vp9_ac_quant(q, 0, cm->bit_depth);
+      invert_quant(&quants->y_quant[q][i], &quants->y_quant_shift[q][i], quant);
+      quants->y_quant_fp[q][i] = (1 << 16) / quant;
+      quants->y_round_fp[q][i] = (qrounding_factor_fp * quant) >> 7;
+      quants->y_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
+      quants->y_round[q][i] = (qrounding_factor * quant) >> 7;
+      cpi->y_dequant[q][i] = quant;
+
+      // uv
+      quant = i == 0 ? vp9_dc_quant(q, cm->uv_dc_delta_q, cm->bit_depth)
+                     : vp9_ac_quant(q, cm->uv_ac_delta_q, cm->bit_depth);
+      invert_quant(&quants->uv_quant[q][i],
+                   &quants->uv_quant_shift[q][i], quant);
+      quants->uv_quant_fp[q][i] = (1 << 16) / quant;
+      quants->uv_round_fp[q][i] = (qrounding_factor_fp * quant) >> 7;
+      quants->uv_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
+      quants->uv_round[q][i] = (qrounding_factor * quant) >> 7;
+      cpi->uv_dequant[q][i] = quant;
+    }
+
+    for (i = 2; i < 8; i++) {
+      quants->y_quant[q][i] = quants->y_quant[q][1];
+      quants->y_quant_fp[q][i] = quants->y_quant_fp[q][1];
+      quants->y_round_fp[q][i] = quants->y_round_fp[q][1];
+      quants->y_quant_shift[q][i] = quants->y_quant_shift[q][1];
+      quants->y_zbin[q][i] = quants->y_zbin[q][1];
+      quants->y_round[q][i] = quants->y_round[q][1];
+      cpi->y_dequant[q][i] = cpi->y_dequant[q][1];
+
+      quants->uv_quant[q][i] = quants->uv_quant[q][1];
+      quants->uv_quant_fp[q][i] = quants->uv_quant_fp[q][1];
+      quants->uv_round_fp[q][i] = quants->uv_round_fp[q][1];
+      quants->uv_quant_shift[q][i] = quants->uv_quant_shift[q][1];
+      quants->uv_zbin[q][i] = quants->uv_zbin[q][1];
+      quants->uv_round[q][i] = quants->uv_round[q][1];
+      cpi->uv_dequant[q][i] = cpi->uv_dequant[q][1];
+    }
+  }
+}
+
+void vp9_init_plane_quantizers(VP9_COMP *cpi, MACROBLOCK *x) {
+  const VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  QUANTS *const quants = &cpi->quants;
+  const int segment_id = xd->mi[0]->segment_id;
+  const int qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+  const int rdmult = vp9_compute_rd_mult(cpi, qindex + cm->y_dc_delta_q);
+  int i;
+
+  // Y
+  x->plane[0].quant = quants->y_quant[qindex];
+  x->plane[0].quant_fp = quants->y_quant_fp[qindex];
+  x->plane[0].round_fp = quants->y_round_fp[qindex];
+  x->plane[0].quant_shift = quants->y_quant_shift[qindex];
+  x->plane[0].zbin = quants->y_zbin[qindex];
+  x->plane[0].round = quants->y_round[qindex];
+  xd->plane[0].dequant = cpi->y_dequant[qindex];
+
+  x->plane[0].quant_thred[0] = x->plane[0].zbin[0] * x->plane[0].zbin[0];
+  x->plane[0].quant_thred[1] = x->plane[0].zbin[1] * x->plane[0].zbin[1];
+
+  // UV
+  for (i = 1; i < 3; i++) {
+    x->plane[i].quant = quants->uv_quant[qindex];
+    x->plane[i].quant_fp = quants->uv_quant_fp[qindex];
+    x->plane[i].round_fp = quants->uv_round_fp[qindex];
+    x->plane[i].quant_shift = quants->uv_quant_shift[qindex];
+    x->plane[i].zbin = quants->uv_zbin[qindex];
+    x->plane[i].round = quants->uv_round[qindex];
+    xd->plane[i].dequant = cpi->uv_dequant[qindex];
+
+    x->plane[i].quant_thred[0] = x->plane[i].zbin[0] * x->plane[i].zbin[0];
+    x->plane[i].quant_thred[1] = x->plane[i].zbin[1] * x->plane[i].zbin[1];
+  }
+
+  x->skip_block = segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP);
+  x->q_index = qindex;
+
+  set_error_per_bit(x, rdmult);
+
+  vp9_initialize_me_consts(cpi, x, x->q_index);
+}
+
+void vp9_frame_init_quantizer(VP9_COMP *cpi) {
+  vp9_init_plane_quantizers(cpi, &cpi->td.mb);
+}
+
+void vp9_set_quantizer(VP9_COMMON *cm, int q) {
+  // quantizer has to be reinitialized with vp9_init_quantizer() if any
+  // delta_q changes.
+  cm->base_qindex = q;
+  cm->y_dc_delta_q = 0;
+  cm->uv_dc_delta_q = 0;
+  cm->uv_ac_delta_q = 0;
+}
+
+// Table that converts 0-63 Q-range values passed in outside to the Qindex
+// range used internally.
+static const int quantizer_to_qindex[] = {
+  0,    4,   8,  12,  16,  20,  24,  28,
+  32,   36,  40,  44,  48,  52,  56,  60,
+  64,   68,  72,  76,  80,  84,  88,  92,
+  96,  100, 104, 108, 112, 116, 120, 124,
+  128, 132, 136, 140, 144, 148, 152, 156,
+  160, 164, 168, 172, 176, 180, 184, 188,
+  192, 196, 200, 204, 208, 212, 216, 220,
+  224, 228, 232, 236, 240, 244, 249, 255,
+};
+
+int vp9_quantizer_to_qindex(int quantizer) {
+  return quantizer_to_qindex[quantizer];
+}
+
+int vp9_qindex_to_quantizer(int qindex) {
+  int quantizer;
+
+  for (quantizer = 0; quantizer < 64; ++quantizer)
+    if (quantizer_to_qindex[quantizer] >= qindex)
+      return quantizer;
+
+  return 63;
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_quantize.h b/libvpx/libvpx/vp9/encoder/vp9_quantize.h
new file mode 100644
index 0000000..6132036
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_quantize.h
@@ -0,0 +1,62 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_QUANTIZE_H_
+#define VP9_ENCODER_VP9_QUANTIZE_H_
+
+#include "./vpx_config.h"
+#include "vp9/encoder/vp9_block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  DECLARE_ALIGNED(16, int16_t, y_quant[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_quant_shift[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_zbin[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_round[QINDEX_RANGE][8]);
+
+  // TODO(jingning): in progress of re-working the quantization. will decide
+  // if we want to deprecate the current use of y_quant.
+  DECLARE_ALIGNED(16, int16_t, y_quant_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_quant_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_round_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_round_fp[QINDEX_RANGE][8]);
+
+  DECLARE_ALIGNED(16, int16_t, uv_quant[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_quant_shift[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_zbin[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_round[QINDEX_RANGE][8]);
+} QUANTS;
+
+void vp9_regular_quantize_b_4x4(MACROBLOCK *x, int plane, int block,
+                                const int16_t *scan, const int16_t *iscan);
+
+struct VP9_COMP;
+struct VP9Common;
+
+void vp9_frame_init_quantizer(struct VP9_COMP *cpi);
+
+void vp9_init_plane_quantizers(struct VP9_COMP *cpi, MACROBLOCK *x);
+
+void vp9_init_quantizer(struct VP9_COMP *cpi);
+
+void vp9_set_quantizer(struct VP9Common *cm, int q);
+
+int vp9_quantizer_to_qindex(int quantizer);
+
+int vp9_qindex_to_quantizer(int qindex);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_QUANTIZE_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_ratectrl.c b/libvpx/libvpx/vp9/encoder/vp9_ratectrl.c
new file mode 100644
index 0000000..b45f8d0
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_ratectrl.c
@@ -0,0 +1,2238 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/system_state.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+
+// Max rate target for 1080P and below encodes under normal circumstances
+// (1920 * 1080 / (16 * 16)) * MAX_MB_RATE bits per MB
+#define MAX_MB_RATE 250
+#define MAXRATE_1080P 2025000
+
+#define DEFAULT_KF_BOOST 2000
+#define DEFAULT_GF_BOOST 2000
+
+#define LIMIT_QRANGE_FOR_ALTREF_AND_KEY 1
+
+#define MIN_BPB_FACTOR 0.005
+#define MAX_BPB_FACTOR 50
+
+#define FRAME_OVERHEAD_BITS 200
+
+#if CONFIG_VP9_HIGHBITDEPTH
+#define ASSIGN_MINQ_TABLE(bit_depth, name) \
+  do { \
+    switch (bit_depth) { \
+      case VPX_BITS_8: \
+        name = name##_8; \
+        break; \
+      case VPX_BITS_10: \
+        name = name##_10; \
+        break; \
+      case VPX_BITS_12: \
+        name = name##_12; \
+        break; \
+      default: \
+        assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10" \
+                    " or VPX_BITS_12"); \
+        name = NULL; \
+    } \
+  } while (0)
+#else
+#define ASSIGN_MINQ_TABLE(bit_depth, name) \
+  do { \
+    (void) bit_depth; \
+    name = name##_8; \
+  } while (0)
+#endif
+
+// Tables relating active max Q to active min Q
+static int kf_low_motion_minq_8[QINDEX_RANGE];
+static int kf_high_motion_minq_8[QINDEX_RANGE];
+static int arfgf_low_motion_minq_8[QINDEX_RANGE];
+static int arfgf_high_motion_minq_8[QINDEX_RANGE];
+static int inter_minq_8[QINDEX_RANGE];
+static int rtc_minq_8[QINDEX_RANGE];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static int kf_low_motion_minq_10[QINDEX_RANGE];
+static int kf_high_motion_minq_10[QINDEX_RANGE];
+static int arfgf_low_motion_minq_10[QINDEX_RANGE];
+static int arfgf_high_motion_minq_10[QINDEX_RANGE];
+static int inter_minq_10[QINDEX_RANGE];
+static int rtc_minq_10[QINDEX_RANGE];
+static int kf_low_motion_minq_12[QINDEX_RANGE];
+static int kf_high_motion_minq_12[QINDEX_RANGE];
+static int arfgf_low_motion_minq_12[QINDEX_RANGE];
+static int arfgf_high_motion_minq_12[QINDEX_RANGE];
+static int inter_minq_12[QINDEX_RANGE];
+static int rtc_minq_12[QINDEX_RANGE];
+#endif
+
+static int gf_high = 2000;
+static int gf_low = 400;
+static int kf_high = 5000;
+static int kf_low = 400;
+
+// Functions to compute the active minq lookup table entries based on a
+// formulaic approach to facilitate easier adjustment of the Q tables.
+// The formulae were derived from computing a 3rd order polynomial best
+// fit to the original data (after plotting real maxq vs minq (not q index))
+static int get_minq_index(double maxq, double x3, double x2, double x1,
+                          vpx_bit_depth_t bit_depth) {
+  int i;
+  const double minqtarget = VPXMIN(((x3 * maxq + x2) * maxq + x1) * maxq,
+                                   maxq);
+
+  // Special case handling to deal with the step from q2.0
+  // down to lossless mode represented by q 1.0.
+  if (minqtarget <= 2.0)
+    return 0;
+
+  for (i = 0; i < QINDEX_RANGE; i++) {
+    if (minqtarget <= vp9_convert_qindex_to_q(i, bit_depth))
+      return i;
+  }
+
+  return QINDEX_RANGE - 1;
+}
+
+static void init_minq_luts(int *kf_low_m, int *kf_high_m,
+                           int *arfgf_low, int *arfgf_high,
+                           int *inter, int *rtc, vpx_bit_depth_t bit_depth) {
+  int i;
+  for (i = 0; i < QINDEX_RANGE; i++) {
+    const double maxq = vp9_convert_qindex_to_q(i, bit_depth);
+    kf_low_m[i] = get_minq_index(maxq, 0.000001, -0.0004, 0.150, bit_depth);
+    kf_high_m[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.55, bit_depth);
+    arfgf_low[i] = get_minq_index(maxq, 0.0000015, -0.0009, 0.30, bit_depth);
+    arfgf_high[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.55, bit_depth);
+    inter[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.70, bit_depth);
+    rtc[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.70, bit_depth);
+  }
+}
+
+void vp9_rc_init_minq_luts(void) {
+  init_minq_luts(kf_low_motion_minq_8, kf_high_motion_minq_8,
+                 arfgf_low_motion_minq_8, arfgf_high_motion_minq_8,
+                 inter_minq_8, rtc_minq_8, VPX_BITS_8);
+#if CONFIG_VP9_HIGHBITDEPTH
+  init_minq_luts(kf_low_motion_minq_10, kf_high_motion_minq_10,
+                 arfgf_low_motion_minq_10, arfgf_high_motion_minq_10,
+                 inter_minq_10, rtc_minq_10, VPX_BITS_10);
+  init_minq_luts(kf_low_motion_minq_12, kf_high_motion_minq_12,
+                 arfgf_low_motion_minq_12, arfgf_high_motion_minq_12,
+                 inter_minq_12, rtc_minq_12, VPX_BITS_12);
+#endif
+}
+
+// These functions use formulaic calculations to make playing with the
+// quantizer tables easier. If necessary they can be replaced by lookup
+// tables if and when things settle down in the experimental bitstream
+double vp9_convert_qindex_to_q(int qindex, vpx_bit_depth_t bit_depth) {
+  // Convert the index to a real Q value (scaled down to match old Q values)
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      return vp9_ac_quant(qindex, 0, bit_depth) / 4.0;
+    case VPX_BITS_10:
+      return vp9_ac_quant(qindex, 0, bit_depth) / 16.0;
+    case VPX_BITS_12:
+      return vp9_ac_quant(qindex, 0, bit_depth) / 64.0;
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1.0;
+  }
+#else
+  return vp9_ac_quant(qindex, 0, bit_depth) / 4.0;
+#endif
+}
+
+int vp9_rc_bits_per_mb(FRAME_TYPE frame_type, int qindex,
+                       double correction_factor,
+                       vpx_bit_depth_t bit_depth) {
+  const double q = vp9_convert_qindex_to_q(qindex, bit_depth);
+  int enumerator = frame_type == KEY_FRAME ? 2700000 : 1800000;
+
+  assert(correction_factor <= MAX_BPB_FACTOR &&
+         correction_factor >= MIN_BPB_FACTOR);
+
+  // q based adjustment to baseline enumerator
+  enumerator += (int)(enumerator * q) >> 12;
+  return (int)(enumerator * correction_factor / q);
+}
+
+int vp9_estimate_bits_at_q(FRAME_TYPE frame_type, int q, int mbs,
+                           double correction_factor,
+                           vpx_bit_depth_t bit_depth) {
+  const int bpm = (int)(vp9_rc_bits_per_mb(frame_type, q, correction_factor,
+                                           bit_depth));
+  return VPXMAX(FRAME_OVERHEAD_BITS,
+                (int)((uint64_t)bpm * mbs) >> BPER_MB_NORMBITS);
+}
+
+int vp9_rc_clamp_pframe_target_size(const VP9_COMP *const cpi, int target) {
+  const RATE_CONTROL *rc = &cpi->rc;
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
+  const int min_frame_target = VPXMAX(rc->min_frame_bandwidth,
+                                      rc->avg_frame_bandwidth >> 5);
+  if (target < min_frame_target)
+    target = min_frame_target;
+  if (cpi->refresh_golden_frame && rc->is_src_frame_alt_ref) {
+    // If there is an active ARF at this location use the minimum
+    // bits on this frame even if it is a constructed arf.
+    // The active maximum quantizer insures that an appropriate
+    // number of bits will be spent if needed for constructed ARFs.
+    target = min_frame_target;
+  }
+  // Clip the frame target to the maximum allowed value.
+  if (target > rc->max_frame_bandwidth)
+    target = rc->max_frame_bandwidth;
+  if (oxcf->rc_max_inter_bitrate_pct) {
+    const int max_rate = rc->avg_frame_bandwidth *
+                         oxcf->rc_max_inter_bitrate_pct / 100;
+    target = VPXMIN(target, max_rate);
+  }
+  return target;
+}
+
+int vp9_rc_clamp_iframe_target_size(const VP9_COMP *const cpi, int target) {
+  const RATE_CONTROL *rc = &cpi->rc;
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
+  if (oxcf->rc_max_intra_bitrate_pct) {
+    const int max_rate = rc->avg_frame_bandwidth *
+                             oxcf->rc_max_intra_bitrate_pct / 100;
+    target = VPXMIN(target, max_rate);
+  }
+  if (target > rc->max_frame_bandwidth)
+    target = rc->max_frame_bandwidth;
+  return target;
+}
+
+// Update the buffer level for higher temporal layers, given the encoded current
+// temporal layer.
+static void update_layer_buffer_level(SVC *svc, int encoded_frame_size) {
+  int i = 0;
+  int current_temporal_layer = svc->temporal_layer_id;
+  for (i = current_temporal_layer + 1;
+      i < svc->number_temporal_layers; ++i) {
+    const int layer = LAYER_IDS_TO_IDX(svc->spatial_layer_id, i,
+                                       svc->number_temporal_layers);
+    LAYER_CONTEXT *lc = &svc->layer_context[layer];
+    RATE_CONTROL *lrc = &lc->rc;
+    int bits_off_for_this_layer = (int)(lc->target_bandwidth / lc->framerate -
+        encoded_frame_size);
+    lrc->bits_off_target += bits_off_for_this_layer;
+
+    // Clip buffer level to maximum buffer size for the layer.
+    lrc->bits_off_target =
+        VPXMIN(lrc->bits_off_target, lrc->maximum_buffer_size);
+    lrc->buffer_level = lrc->bits_off_target;
+  }
+}
+
+// Update the buffer level: leaky bucket model.
+static void update_buffer_level(VP9_COMP *cpi, int encoded_frame_size) {
+  const VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  // Non-viewable frames are a special case and are treated as pure overhead.
+  if (!cm->show_frame) {
+    rc->bits_off_target -= encoded_frame_size;
+  } else {
+    rc->bits_off_target += rc->avg_frame_bandwidth - encoded_frame_size;
+  }
+
+  // Clip the buffer level to the maximum specified buffer size.
+  rc->bits_off_target = VPXMIN(rc->bits_off_target, rc->maximum_buffer_size);
+
+  // For screen-content mode, and if frame-dropper is off, don't let buffer
+  // level go below threshold, given here as -rc->maximum_ buffer_size.
+  if (cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
+      cpi->oxcf.drop_frames_water_mark == 0)
+    rc->bits_off_target = VPXMAX(rc->bits_off_target, -rc->maximum_buffer_size);
+
+  rc->buffer_level = rc->bits_off_target;
+
+  if (is_one_pass_cbr_svc(cpi)) {
+    update_layer_buffer_level(&cpi->svc, encoded_frame_size);
+  }
+}
+
+int vp9_rc_get_default_min_gf_interval(
+    int width, int height, double framerate) {
+  // Assume we do not need any constraint lower than 4K 20 fps
+  static const double factor_safe = 3840 * 2160 * 20.0;
+  const double factor = width * height * framerate;
+  const int default_interval =
+      clamp((int)(framerate * 0.125), MIN_GF_INTERVAL, MAX_GF_INTERVAL);
+
+  if (factor <= factor_safe)
+    return default_interval;
+  else
+    return VPXMAX(default_interval,
+                  (int)(MIN_GF_INTERVAL * factor / factor_safe + 0.5));
+  // Note this logic makes:
+  // 4K24: 5
+  // 4K30: 6
+  // 4K60: 12
+}
+
+int vp9_rc_get_default_max_gf_interval(double framerate, int min_gf_interval) {
+  int interval = VPXMIN(MAX_GF_INTERVAL, (int)(framerate * 0.75));
+  interval += (interval & 0x01);  // Round to even value
+  return VPXMAX(interval, min_gf_interval);
+}
+
+void vp9_rc_init(const VP9EncoderConfig *oxcf, int pass, RATE_CONTROL *rc) {
+  int i;
+
+  if (pass == 0 && oxcf->rc_mode == VPX_CBR) {
+    rc->avg_frame_qindex[KEY_FRAME] = oxcf->worst_allowed_q;
+    rc->avg_frame_qindex[INTER_FRAME] = oxcf->worst_allowed_q;
+  } else {
+    rc->avg_frame_qindex[KEY_FRAME] = (oxcf->worst_allowed_q +
+                                       oxcf->best_allowed_q) / 2;
+    rc->avg_frame_qindex[INTER_FRAME] = (oxcf->worst_allowed_q +
+                                         oxcf->best_allowed_q) / 2;
+  }
+
+  rc->last_q[KEY_FRAME] = oxcf->best_allowed_q;
+  rc->last_q[INTER_FRAME] = oxcf->worst_allowed_q;
+
+  rc->buffer_level =    rc->starting_buffer_level;
+  rc->bits_off_target = rc->starting_buffer_level;
+
+  rc->rolling_target_bits      = rc->avg_frame_bandwidth;
+  rc->rolling_actual_bits      = rc->avg_frame_bandwidth;
+  rc->long_rolling_target_bits = rc->avg_frame_bandwidth;
+  rc->long_rolling_actual_bits = rc->avg_frame_bandwidth;
+
+  rc->total_actual_bits = 0;
+  rc->total_target_bits = 0;
+  rc->total_target_vs_actual = 0;
+  rc->avg_frame_low_motion = 0;
+  rc->high_source_sad = 0;
+  rc->count_last_scene_change = 0;
+  rc->avg_source_sad = 0;
+
+  rc->frames_since_key = 8;  // Sensible default for first frame.
+  rc->this_key_frame_forced = 0;
+  rc->next_key_frame_forced = 0;
+  rc->source_alt_ref_pending = 0;
+  rc->source_alt_ref_active = 0;
+
+  rc->frames_till_gf_update_due = 0;
+  rc->ni_av_qi = oxcf->worst_allowed_q;
+  rc->ni_tot_qi = 0;
+  rc->ni_frames = 0;
+
+  rc->tot_q = 0.0;
+  rc->avg_q = vp9_convert_qindex_to_q(oxcf->worst_allowed_q, oxcf->bit_depth);
+
+  for (i = 0; i < RATE_FACTOR_LEVELS; ++i) {
+    rc->rate_correction_factors[i] = 1.0;
+  }
+
+  rc->min_gf_interval = oxcf->min_gf_interval;
+  rc->max_gf_interval = oxcf->max_gf_interval;
+  if (rc->min_gf_interval == 0)
+    rc->min_gf_interval = vp9_rc_get_default_min_gf_interval(
+        oxcf->width, oxcf->height, oxcf->init_framerate);
+  if (rc->max_gf_interval == 0)
+    rc->max_gf_interval = vp9_rc_get_default_max_gf_interval(
+        oxcf->init_framerate, rc->min_gf_interval);
+  rc->baseline_gf_interval = (rc->min_gf_interval + rc->max_gf_interval) / 2;
+}
+
+int vp9_rc_drop_frame(VP9_COMP *cpi) {
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
+  RATE_CONTROL *const rc = &cpi->rc;
+  if (!oxcf->drop_frames_water_mark ||
+      (is_one_pass_cbr_svc(cpi) &&
+       cpi->svc.spatial_layer_id > cpi->svc.first_spatial_layer_to_encode)) {
+    return 0;
+  } else {
+    if (rc->buffer_level < 0) {
+      // Always drop if buffer is below 0.
+      return 1;
+    } else {
+      // If buffer is below drop_mark, for now just drop every other frame
+      // (starting with the next frame) until it increases back over drop_mark.
+      int drop_mark = (int)(oxcf->drop_frames_water_mark *
+          rc->optimal_buffer_level / 100);
+      if ((rc->buffer_level > drop_mark) &&
+          (rc->decimation_factor > 0)) {
+        --rc->decimation_factor;
+      } else if (rc->buffer_level <= drop_mark &&
+          rc->decimation_factor == 0) {
+        rc->decimation_factor = 1;
+      }
+      if (rc->decimation_factor > 0) {
+        if (rc->decimation_count > 0) {
+          --rc->decimation_count;
+          return 1;
+        } else {
+          rc->decimation_count = rc->decimation_factor;
+          return 0;
+        }
+      } else {
+        rc->decimation_count = 0;
+        return 0;
+      }
+    }
+  }
+}
+
+static double get_rate_correction_factor(const VP9_COMP *cpi) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  double rcf;
+
+  if (cpi->common.frame_type == KEY_FRAME) {
+    rcf = rc->rate_correction_factors[KF_STD];
+  } else if (cpi->oxcf.pass == 2) {
+    RATE_FACTOR_LEVEL rf_lvl =
+      cpi->twopass.gf_group.rf_level[cpi->twopass.gf_group.index];
+    rcf = rc->rate_correction_factors[rf_lvl];
+  } else {
+    if ((cpi->refresh_alt_ref_frame || cpi->refresh_golden_frame) &&
+        !rc->is_src_frame_alt_ref && !cpi->use_svc &&
+        (cpi->oxcf.rc_mode != VPX_CBR || cpi->oxcf.gf_cbr_boost_pct > 20))
+      rcf = rc->rate_correction_factors[GF_ARF_STD];
+    else
+      rcf = rc->rate_correction_factors[INTER_NORMAL];
+  }
+  rcf *= rcf_mult[rc->frame_size_selector];
+  return fclamp(rcf, MIN_BPB_FACTOR, MAX_BPB_FACTOR);
+}
+
+static void set_rate_correction_factor(VP9_COMP *cpi, double factor) {
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  // Normalize RCF to account for the size-dependent scaling factor.
+  factor /= rcf_mult[cpi->rc.frame_size_selector];
+
+  factor = fclamp(factor, MIN_BPB_FACTOR, MAX_BPB_FACTOR);
+
+  if (cpi->common.frame_type == KEY_FRAME) {
+    rc->rate_correction_factors[KF_STD] = factor;
+  } else if (cpi->oxcf.pass == 2) {
+    RATE_FACTOR_LEVEL rf_lvl =
+      cpi->twopass.gf_group.rf_level[cpi->twopass.gf_group.index];
+    rc->rate_correction_factors[rf_lvl] = factor;
+  } else {
+    if ((cpi->refresh_alt_ref_frame || cpi->refresh_golden_frame) &&
+        !rc->is_src_frame_alt_ref && !cpi->use_svc &&
+        (cpi->oxcf.rc_mode != VPX_CBR || cpi->oxcf.gf_cbr_boost_pct > 20))
+      rc->rate_correction_factors[GF_ARF_STD] = factor;
+    else
+      rc->rate_correction_factors[INTER_NORMAL] = factor;
+  }
+}
+
+void vp9_rc_update_rate_correction_factors(VP9_COMP *cpi) {
+  const VP9_COMMON *const cm = &cpi->common;
+  int correction_factor = 100;
+  double rate_correction_factor = get_rate_correction_factor(cpi);
+  double adjustment_limit;
+
+  int projected_size_based_on_q = 0;
+
+  // Do not update the rate factors for arf overlay frames.
+  if (cpi->rc.is_src_frame_alt_ref)
+    return;
+
+  // Clear down mmx registers to allow floating point in what follows
+  vpx_clear_system_state();
+
+  // Work out how big we would have expected the frame to be at this Q given
+  // the current correction factor.
+  // Stay in double to avoid int overflow when values are large
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cpi->common.seg.enabled) {
+    projected_size_based_on_q =
+        vp9_cyclic_refresh_estimate_bits_at_q(cpi, rate_correction_factor);
+  } else {
+    projected_size_based_on_q = vp9_estimate_bits_at_q(cpi->common.frame_type,
+                                                       cm->base_qindex,
+                                                       cm->MBs,
+                                                       rate_correction_factor,
+                                                       cm->bit_depth);
+  }
+  // Work out a size correction factor.
+  if (projected_size_based_on_q > FRAME_OVERHEAD_BITS)
+    correction_factor = (int)((100 * (int64_t)cpi->rc.projected_frame_size) /
+                        projected_size_based_on_q);
+
+  // More heavily damped adjustment used if we have been oscillating either side
+  // of target.
+  adjustment_limit = 0.25 +
+      0.5 * VPXMIN(1, fabs(log10(0.01 * correction_factor)));
+
+  cpi->rc.q_2_frame = cpi->rc.q_1_frame;
+  cpi->rc.q_1_frame = cm->base_qindex;
+  cpi->rc.rc_2_frame = cpi->rc.rc_1_frame;
+  if (correction_factor > 110)
+    cpi->rc.rc_1_frame = -1;
+  else if (correction_factor < 90)
+    cpi->rc.rc_1_frame = 1;
+  else
+    cpi->rc.rc_1_frame = 0;
+
+  // Turn off oscilation detection in the case of massive overshoot.
+  if (cpi->rc.rc_1_frame == -1 && cpi->rc.rc_2_frame == 1 &&
+      correction_factor > 1000) {
+    cpi->rc.rc_2_frame = 0;
+  }
+
+  if (correction_factor > 102) {
+    // We are not already at the worst allowable quality
+    correction_factor = (int)(100 + ((correction_factor - 100) *
+                                  adjustment_limit));
+    rate_correction_factor = (rate_correction_factor * correction_factor) / 100;
+    // Keep rate_correction_factor within limits
+    if (rate_correction_factor > MAX_BPB_FACTOR)
+      rate_correction_factor = MAX_BPB_FACTOR;
+  } else if (correction_factor < 99) {
+    // We are not already at the best allowable quality
+    correction_factor = (int)(100 - ((100 - correction_factor) *
+                                  adjustment_limit));
+    rate_correction_factor = (rate_correction_factor * correction_factor) / 100;
+
+    // Keep rate_correction_factor within limits
+    if (rate_correction_factor < MIN_BPB_FACTOR)
+      rate_correction_factor = MIN_BPB_FACTOR;
+  }
+
+  set_rate_correction_factor(cpi, rate_correction_factor);
+}
+
+
+int vp9_rc_regulate_q(const VP9_COMP *cpi, int target_bits_per_frame,
+                      int active_best_quality, int active_worst_quality) {
+  const VP9_COMMON *const cm = &cpi->common;
+  int q = active_worst_quality;
+  int last_error = INT_MAX;
+  int i, target_bits_per_mb, bits_per_mb_at_this_q;
+  const double correction_factor = get_rate_correction_factor(cpi);
+
+  // Calculate required scaling factor based on target frame size and size of
+  // frame produced using previous Q.
+  target_bits_per_mb =
+      ((uint64_t)target_bits_per_frame << BPER_MB_NORMBITS) / cm->MBs;
+
+  i = active_best_quality;
+
+  do {
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
+        cm->seg.enabled &&
+        cpi->svc.temporal_layer_id == 0) {
+      bits_per_mb_at_this_q =
+          (int)vp9_cyclic_refresh_rc_bits_per_mb(cpi, i, correction_factor);
+    } else {
+      bits_per_mb_at_this_q = (int)vp9_rc_bits_per_mb(cm->frame_type, i,
+                                                      correction_factor,
+                                                      cm->bit_depth);
+    }
+
+    if (bits_per_mb_at_this_q <= target_bits_per_mb) {
+      if ((target_bits_per_mb - bits_per_mb_at_this_q) <= last_error)
+        q = i;
+      else
+        q = i - 1;
+
+      break;
+    } else {
+      last_error = bits_per_mb_at_this_q - target_bits_per_mb;
+    }
+  } while (++i <= active_worst_quality);
+
+  // In CBR mode, this makes sure q is between oscillating Qs to prevent
+  // resonance.
+  if (cpi->oxcf.rc_mode == VPX_CBR &&
+      (cpi->rc.rc_1_frame * cpi->rc.rc_2_frame == -1) &&
+      cpi->rc.q_1_frame != cpi->rc.q_2_frame) {
+    q = clamp(q, VPXMIN(cpi->rc.q_1_frame, cpi->rc.q_2_frame),
+              VPXMAX(cpi->rc.q_1_frame, cpi->rc.q_2_frame));
+  }
+  return q;
+}
+
+static int get_active_quality(int q, int gfu_boost, int low, int high,
+                              int *low_motion_minq, int *high_motion_minq) {
+  if (gfu_boost > high) {
+    return low_motion_minq[q];
+  } else if (gfu_boost < low) {
+    return high_motion_minq[q];
+  } else {
+    const int gap = high - low;
+    const int offset = high - gfu_boost;
+    const int qdiff = high_motion_minq[q] - low_motion_minq[q];
+    const int adjustment = ((offset * qdiff) + (gap >> 1)) / gap;
+    return low_motion_minq[q] + adjustment;
+  }
+}
+
+static int get_kf_active_quality(const RATE_CONTROL *const rc, int q,
+                                 vpx_bit_depth_t bit_depth) {
+  int *kf_low_motion_minq;
+  int *kf_high_motion_minq;
+  ASSIGN_MINQ_TABLE(bit_depth, kf_low_motion_minq);
+  ASSIGN_MINQ_TABLE(bit_depth, kf_high_motion_minq);
+  return get_active_quality(q, rc->kf_boost, kf_low, kf_high,
+                            kf_low_motion_minq, kf_high_motion_minq);
+}
+
+static int get_gf_active_quality(const RATE_CONTROL *const rc, int q,
+                                 vpx_bit_depth_t bit_depth) {
+  int *arfgf_low_motion_minq;
+  int *arfgf_high_motion_minq;
+  ASSIGN_MINQ_TABLE(bit_depth, arfgf_low_motion_minq);
+  ASSIGN_MINQ_TABLE(bit_depth, arfgf_high_motion_minq);
+  return get_active_quality(q, rc->gfu_boost, gf_low, gf_high,
+                            arfgf_low_motion_minq, arfgf_high_motion_minq);
+}
+
+static int calc_active_worst_quality_one_pass_vbr(const VP9_COMP *cpi) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const unsigned int curr_frame = cpi->common.current_video_frame;
+  int active_worst_quality;
+
+  if (cpi->common.frame_type == KEY_FRAME) {
+    active_worst_quality = curr_frame == 0 ? rc->worst_quality
+                                           : rc->last_q[KEY_FRAME] << 1;
+  } else {
+    if (!rc->is_src_frame_alt_ref &&
+        (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+      active_worst_quality =  curr_frame == 1 ? rc->last_q[KEY_FRAME] * 5 >> 2
+                                              : rc->last_q[INTER_FRAME];
+    } else {
+      active_worst_quality = curr_frame == 1 ? rc->last_q[KEY_FRAME] << 1 :
+          VPXMIN(rc->last_q[INTER_FRAME] << 1,
+                (rc->avg_frame_qindex[INTER_FRAME] * 3 >> 1));
+    }
+  }
+  return VPXMIN(active_worst_quality, rc->worst_quality);
+}
+
+// Adjust active_worst_quality level based on buffer level.
+static int calc_active_worst_quality_one_pass_cbr(const VP9_COMP *cpi) {
+  // Adjust active_worst_quality: If buffer is above the optimal/target level,
+  // bring active_worst_quality down depending on fullness of buffer.
+  // If buffer is below the optimal level, let the active_worst_quality go from
+  // ambient Q (at buffer = optimal level) to worst_quality level
+  // (at buffer = critical level).
+  const VP9_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *rc = &cpi->rc;
+  // Buffer level below which we push active_worst to worst_quality.
+  int64_t critical_level = rc->optimal_buffer_level >> 3;
+  int64_t buff_lvl_step = 0;
+  int adjustment = 0;
+  int active_worst_quality;
+  int ambient_qp;
+  unsigned int num_frames_weight_key = 5 * cpi->svc.number_temporal_layers;
+  if (cm->frame_type == KEY_FRAME)
+    return rc->worst_quality;
+  // For ambient_qp we use minimum of avg_frame_qindex[KEY_FRAME/INTER_FRAME]
+  // for the first few frames following key frame. These are both initialized
+  // to worst_quality and updated with (3/4, 1/4) average in postencode_update.
+  // So for first few frames following key, the qp of that key frame is weighted
+  // into the active_worst_quality setting.
+  ambient_qp = (cm->current_video_frame < num_frames_weight_key) ?
+                   VPXMIN(rc->avg_frame_qindex[INTER_FRAME],
+                          rc->avg_frame_qindex[KEY_FRAME]) :
+                   rc->avg_frame_qindex[INTER_FRAME];
+  active_worst_quality = VPXMIN(rc->worst_quality, ambient_qp * 5 >> 2);
+  if (rc->buffer_level > rc->optimal_buffer_level) {
+    // Adjust down.
+    // Maximum limit for down adjustment, ~30%.
+    int max_adjustment_down = active_worst_quality / 3;
+    if (max_adjustment_down) {
+      buff_lvl_step = ((rc->maximum_buffer_size -
+                        rc->optimal_buffer_level) / max_adjustment_down);
+      if (buff_lvl_step)
+        adjustment = (int)((rc->buffer_level - rc->optimal_buffer_level) /
+                            buff_lvl_step);
+      active_worst_quality -= adjustment;
+    }
+  } else if (rc->buffer_level > critical_level) {
+    // Adjust up from ambient Q.
+    if (critical_level) {
+      buff_lvl_step = (rc->optimal_buffer_level - critical_level);
+      if (buff_lvl_step) {
+        adjustment = (int)((rc->worst_quality - ambient_qp) *
+                           (rc->optimal_buffer_level - rc->buffer_level) /
+                           buff_lvl_step);
+      }
+      active_worst_quality = ambient_qp + adjustment;
+    }
+  } else {
+    // Set to worst_quality if buffer is below critical level.
+    active_worst_quality = rc->worst_quality;
+  }
+  return active_worst_quality;
+}
+
+static int rc_pick_q_and_bounds_one_pass_cbr(const VP9_COMP *cpi,
+                                             int *bottom_index,
+                                             int *top_index) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  int active_best_quality;
+  int active_worst_quality = calc_active_worst_quality_one_pass_cbr(cpi);
+  int q;
+  int *rtc_minq;
+  ASSIGN_MINQ_TABLE(cm->bit_depth, rtc_minq);
+
+  if (frame_is_intra_only(cm)) {
+    active_best_quality = rc->best_quality;
+    // Handle the special case for key frames forced when we have reached
+    // the maximum key frame interval. Here force the Q to a range
+    // based on the ambient Q to reduce the risk of popping.
+    if (rc->this_key_frame_forced) {
+      int qindex = rc->last_boosted_qindex;
+      double last_boosted_q = vp9_convert_qindex_to_q(qindex, cm->bit_depth);
+      int delta_qindex = vp9_compute_qdelta(rc, last_boosted_q,
+                                            (last_boosted_q * 0.75),
+                                            cm->bit_depth);
+      active_best_quality = VPXMAX(qindex + delta_qindex, rc->best_quality);
+    } else if (cm->current_video_frame > 0) {
+      // not first frame of one pass and kf_boost is set
+      double q_adj_factor = 1.0;
+      double q_val;
+
+      active_best_quality =
+          get_kf_active_quality(rc, rc->avg_frame_qindex[KEY_FRAME],
+                                cm->bit_depth);
+
+      // Allow somewhat lower kf minq with small image formats.
+      if ((cm->width * cm->height) <= (352 * 288)) {
+        q_adj_factor -= 0.25;
+      }
+
+      // Convert the adjustment factor to a qindex delta
+      // on active_best_quality.
+      q_val = vp9_convert_qindex_to_q(active_best_quality, cm->bit_depth);
+      active_best_quality += vp9_compute_qdelta(rc, q_val,
+                                                q_val * q_adj_factor,
+                                                cm->bit_depth);
+    }
+  } else if (!rc->is_src_frame_alt_ref &&
+             !cpi->use_svc &&
+             (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+    // Use the lower of active_worst_quality and recent
+    // average Q as basis for GF/ARF best Q limit unless last frame was
+    // a key frame.
+    if (rc->frames_since_key > 1 &&
+        rc->avg_frame_qindex[INTER_FRAME] < active_worst_quality) {
+      q = rc->avg_frame_qindex[INTER_FRAME];
+    } else {
+      q = active_worst_quality;
+    }
+    active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+  } else {
+    // Use the lower of active_worst_quality and recent/average Q.
+    if (cm->current_video_frame > 1) {
+      if (rc->avg_frame_qindex[INTER_FRAME] < active_worst_quality)
+        active_best_quality = rtc_minq[rc->avg_frame_qindex[INTER_FRAME]];
+      else
+        active_best_quality = rtc_minq[active_worst_quality];
+    } else {
+      if (rc->avg_frame_qindex[KEY_FRAME] < active_worst_quality)
+        active_best_quality = rtc_minq[rc->avg_frame_qindex[KEY_FRAME]];
+      else
+        active_best_quality = rtc_minq[active_worst_quality];
+    }
+  }
+
+  // Clip the active best and worst quality values to limits
+  active_best_quality = clamp(active_best_quality,
+                              rc->best_quality, rc->worst_quality);
+  active_worst_quality = clamp(active_worst_quality,
+                               active_best_quality, rc->worst_quality);
+
+  *top_index = active_worst_quality;
+  *bottom_index = active_best_quality;
+
+#if LIMIT_QRANGE_FOR_ALTREF_AND_KEY
+  // Limit Q range for the adaptive loop.
+  if (cm->frame_type == KEY_FRAME &&
+      !rc->this_key_frame_forced  &&
+      !(cm->current_video_frame == 0)) {
+    int qdelta = 0;
+    vpx_clear_system_state();
+    qdelta = vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type,
+                                        active_worst_quality, 2.0,
+                                        cm->bit_depth);
+    *top_index = active_worst_quality + qdelta;
+    *top_index = (*top_index > *bottom_index) ? *top_index : *bottom_index;
+  }
+#endif
+
+  // Special case code to try and match quality with forced key frames
+  if (cm->frame_type == KEY_FRAME && rc->this_key_frame_forced) {
+    q = rc->last_boosted_qindex;
+  } else {
+    q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
+                          active_best_quality, active_worst_quality);
+    if (q > *top_index) {
+      // Special case when we are targeting the max allowed rate
+      if (rc->this_frame_target >= rc->max_frame_bandwidth)
+        *top_index = q;
+      else
+        q = *top_index;
+    }
+  }
+  assert(*top_index <= rc->worst_quality &&
+         *top_index >= rc->best_quality);
+  assert(*bottom_index <= rc->worst_quality &&
+         *bottom_index >= rc->best_quality);
+  assert(q <= rc->worst_quality && q >= rc->best_quality);
+  return q;
+}
+
+static int get_active_cq_level_one_pass(
+    const RATE_CONTROL *rc, const VP9EncoderConfig *const oxcf) {
+  static const double cq_adjust_threshold = 0.1;
+  int active_cq_level = oxcf->cq_level;
+  if (oxcf->rc_mode == VPX_CQ &&
+      rc->total_target_bits > 0) {
+    const double x = (double)rc->total_actual_bits / rc->total_target_bits;
+    if (x < cq_adjust_threshold) {
+      active_cq_level = (int)(active_cq_level * x / cq_adjust_threshold);
+    }
+  }
+  return active_cq_level;
+}
+
+#define SMOOTH_PCT_MIN  0.1
+#define SMOOTH_PCT_DIV  0.05
+static int get_active_cq_level_two_pass(
+    const TWO_PASS *twopass, const RATE_CONTROL *rc,
+    const VP9EncoderConfig *const oxcf) {
+  static const double cq_adjust_threshold = 0.1;
+  int active_cq_level = oxcf->cq_level;
+  if (oxcf->rc_mode == VPX_CQ) {
+    if (twopass->mb_smooth_pct > SMOOTH_PCT_MIN) {
+      active_cq_level -= (int)((twopass->mb_smooth_pct - SMOOTH_PCT_MIN) /
+          SMOOTH_PCT_DIV);
+      active_cq_level = VPXMAX(active_cq_level, 0);
+    }
+    if (rc->total_target_bits > 0) {
+      const double x = (double)rc->total_actual_bits / rc->total_target_bits;
+      if (x < cq_adjust_threshold) {
+        active_cq_level = (int)(active_cq_level * x / cq_adjust_threshold);
+      }
+    }
+  }
+  return active_cq_level;
+}
+
+static int rc_pick_q_and_bounds_one_pass_vbr(const VP9_COMP *cpi,
+                                             int *bottom_index,
+                                             int *top_index) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const int cq_level = get_active_cq_level_one_pass(rc, oxcf);
+  int active_best_quality;
+  int active_worst_quality = calc_active_worst_quality_one_pass_vbr(cpi);
+  int q;
+  int *inter_minq;
+  ASSIGN_MINQ_TABLE(cm->bit_depth, inter_minq);
+
+  if (frame_is_intra_only(cm)) {
+    if (oxcf->rc_mode == VPX_Q) {
+      int qindex = cq_level;
+      double q = vp9_convert_qindex_to_q(qindex, cm->bit_depth);
+      int delta_qindex = vp9_compute_qdelta(rc, q, q * 0.25,
+                                            cm->bit_depth);
+      active_best_quality = VPXMAX(qindex + delta_qindex, rc->best_quality);
+    } else if (rc->this_key_frame_forced) {
+      // Handle the special case for key frames forced when we have reached
+      // the maximum key frame interval. Here force the Q to a range
+      // based on the ambient Q to reduce the risk of popping.
+      int qindex = rc->last_boosted_qindex;
+      double last_boosted_q = vp9_convert_qindex_to_q(qindex, cm->bit_depth);
+      int delta_qindex = vp9_compute_qdelta(rc, last_boosted_q,
+                                            last_boosted_q * 0.75,
+                                            cm->bit_depth);
+      active_best_quality = VPXMAX(qindex + delta_qindex, rc->best_quality);
+    } else {
+      // not first frame of one pass and kf_boost is set
+      double q_adj_factor = 1.0;
+      double q_val;
+
+      active_best_quality =
+          get_kf_active_quality(rc, rc->avg_frame_qindex[KEY_FRAME],
+                                cm->bit_depth);
+
+      // Allow somewhat lower kf minq with small image formats.
+      if ((cm->width * cm->height) <= (352 * 288)) {
+        q_adj_factor -= 0.25;
+      }
+
+      // Convert the adjustment factor to a qindex delta
+      // on active_best_quality.
+      q_val = vp9_convert_qindex_to_q(active_best_quality, cm->bit_depth);
+      active_best_quality += vp9_compute_qdelta(rc, q_val,
+                                                q_val * q_adj_factor,
+                                                cm->bit_depth);
+    }
+  } else if (!rc->is_src_frame_alt_ref &&
+             (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+    // Use the lower of active_worst_quality and recent
+    // average Q as basis for GF/ARF best Q limit unless last frame was
+    // a key frame.
+    if (rc->frames_since_key > 1) {
+      if (rc->avg_frame_qindex[INTER_FRAME] < active_worst_quality) {
+        q = rc->avg_frame_qindex[INTER_FRAME];
+      } else {
+        q = active_worst_quality;
+      }
+    } else {
+      q = rc->avg_frame_qindex[KEY_FRAME];
+    }
+    // For constrained quality dont allow Q less than the cq level
+    if (oxcf->rc_mode == VPX_CQ) {
+      if (q < cq_level)
+        q = cq_level;
+
+      active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+
+      // Constrained quality use slightly lower active best.
+      active_best_quality = active_best_quality * 15 / 16;
+
+    } else if (oxcf->rc_mode == VPX_Q) {
+      int qindex = cq_level;
+      double q = vp9_convert_qindex_to_q(qindex, cm->bit_depth);
+      int delta_qindex;
+      if (cpi->refresh_alt_ref_frame)
+        delta_qindex = vp9_compute_qdelta(rc, q, q * 0.40, cm->bit_depth);
+      else
+        delta_qindex = vp9_compute_qdelta(rc, q, q * 0.50, cm->bit_depth);
+      active_best_quality = VPXMAX(qindex + delta_qindex, rc->best_quality);
+    } else {
+      active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+    }
+  } else {
+    if (oxcf->rc_mode == VPX_Q) {
+      int qindex = cq_level;
+      double q = vp9_convert_qindex_to_q(qindex, cm->bit_depth);
+      double delta_rate[FIXED_GF_INTERVAL] =
+          {0.50, 1.0, 0.85, 1.0, 0.70, 1.0, 0.85, 1.0};
+      int delta_qindex =
+          vp9_compute_qdelta(rc, q,
+                             q * delta_rate[cm->current_video_frame %
+                             FIXED_GF_INTERVAL], cm->bit_depth);
+      active_best_quality = VPXMAX(qindex + delta_qindex, rc->best_quality);
+    } else {
+      // Use the min of the average Q and active_worst_quality as basis for
+      // active_best.
+      if (cm->current_video_frame > 1) {
+        q = VPXMIN(rc->avg_frame_qindex[INTER_FRAME], active_worst_quality);
+        active_best_quality = inter_minq[q];
+      } else {
+        active_best_quality = inter_minq[rc->avg_frame_qindex[KEY_FRAME]];
+      }
+      // For the constrained quality mode we don't want
+      // q to fall below the cq level.
+      if ((oxcf->rc_mode == VPX_CQ) &&
+          (active_best_quality < cq_level)) {
+        active_best_quality = cq_level;
+      }
+    }
+  }
+
+  // Clip the active best and worst quality values to limits
+  active_best_quality = clamp(active_best_quality,
+                              rc->best_quality, rc->worst_quality);
+  active_worst_quality = clamp(active_worst_quality,
+                               active_best_quality, rc->worst_quality);
+
+  *top_index = active_worst_quality;
+  *bottom_index = active_best_quality;
+
+#if LIMIT_QRANGE_FOR_ALTREF_AND_KEY
+  {
+    int qdelta = 0;
+    vpx_clear_system_state();
+
+    // Limit Q range for the adaptive loop.
+    if (cm->frame_type == KEY_FRAME &&
+        !rc->this_key_frame_forced &&
+        !(cm->current_video_frame == 0)) {
+      qdelta = vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type,
+                                          active_worst_quality, 2.0,
+                                          cm->bit_depth);
+    } else if (!rc->is_src_frame_alt_ref &&
+               (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+      qdelta = vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type,
+                                          active_worst_quality, 1.75,
+                                          cm->bit_depth);
+    }
+    *top_index = active_worst_quality + qdelta;
+    *top_index = (*top_index > *bottom_index) ? *top_index : *bottom_index;
+  }
+#endif
+
+  if (oxcf->rc_mode == VPX_Q) {
+    q = active_best_quality;
+  // Special case code to try and match quality with forced key frames
+  } else if ((cm->frame_type == KEY_FRAME) && rc->this_key_frame_forced) {
+    q = rc->last_boosted_qindex;
+  } else {
+    q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
+                          active_best_quality, active_worst_quality);
+    if (q > *top_index) {
+      // Special case when we are targeting the max allowed rate
+      if (rc->this_frame_target >= rc->max_frame_bandwidth)
+        *top_index = q;
+      else
+        q = *top_index;
+    }
+  }
+
+  assert(*top_index <= rc->worst_quality &&
+         *top_index >= rc->best_quality);
+  assert(*bottom_index <= rc->worst_quality &&
+         *bottom_index >= rc->best_quality);
+  assert(q <= rc->worst_quality && q >= rc->best_quality);
+  return q;
+}
+
+int vp9_frame_type_qdelta(const VP9_COMP *cpi, int rf_level, int q) {
+  static const double rate_factor_deltas[RATE_FACTOR_LEVELS] = {
+    1.00,  // INTER_NORMAL
+    1.00,  // INTER_HIGH
+    1.50,  // GF_ARF_LOW
+    1.75,  // GF_ARF_STD
+    2.00,  // KF_STD
+  };
+  static const FRAME_TYPE frame_type[RATE_FACTOR_LEVELS] =
+      {INTER_FRAME, INTER_FRAME, INTER_FRAME, INTER_FRAME, KEY_FRAME};
+  const VP9_COMMON *const cm = &cpi->common;
+  int qdelta = vp9_compute_qdelta_by_rate(&cpi->rc, frame_type[rf_level],
+                                          q, rate_factor_deltas[rf_level],
+                                          cm->bit_depth);
+  return qdelta;
+}
+
+#define STATIC_MOTION_THRESH 95
+static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi,
+                                         int *bottom_index,
+                                         int *top_index) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const GF_GROUP *gf_group = &cpi->twopass.gf_group;
+  const int cq_level = get_active_cq_level_two_pass(&cpi->twopass, rc, oxcf);
+  int active_best_quality;
+  int active_worst_quality = cpi->twopass.active_worst_quality;
+  int q;
+  int *inter_minq;
+  ASSIGN_MINQ_TABLE(cm->bit_depth, inter_minq);
+
+  if (frame_is_intra_only(cm) || vp9_is_upper_layer_key_frame(cpi)) {
+    // Handle the special case for key frames forced when we have reached
+    // the maximum key frame interval. Here force the Q to a range
+    // based on the ambient Q to reduce the risk of popping.
+    if (rc->this_key_frame_forced) {
+      double last_boosted_q;
+      int delta_qindex;
+      int qindex;
+
+      if (cpi->twopass.last_kfgroup_zeromotion_pct >= STATIC_MOTION_THRESH) {
+        qindex = VPXMIN(rc->last_kf_qindex, rc->last_boosted_qindex);
+        active_best_quality = qindex;
+        last_boosted_q = vp9_convert_qindex_to_q(qindex, cm->bit_depth);
+        delta_qindex = vp9_compute_qdelta(rc, last_boosted_q,
+                                              last_boosted_q * 1.25,
+                                              cm->bit_depth);
+        active_worst_quality =
+            VPXMIN(qindex + delta_qindex, active_worst_quality);
+      } else {
+        qindex = rc->last_boosted_qindex;
+        last_boosted_q = vp9_convert_qindex_to_q(qindex, cm->bit_depth);
+        delta_qindex = vp9_compute_qdelta(rc, last_boosted_q,
+                                              last_boosted_q * 0.75,
+                                              cm->bit_depth);
+        active_best_quality = VPXMAX(qindex + delta_qindex, rc->best_quality);
+      }
+    } else {
+      // Not forced keyframe.
+      double q_adj_factor = 1.0;
+      double q_val;
+      // Baseline value derived from cpi->active_worst_quality and kf boost.
+      active_best_quality = get_kf_active_quality(rc, active_worst_quality,
+                                                  cm->bit_depth);
+
+      // Allow somewhat lower kf minq with small image formats.
+      if ((cm->width * cm->height) <= (352 * 288)) {
+        q_adj_factor -= 0.25;
+      }
+
+      // Make a further adjustment based on the kf zero motion measure.
+      q_adj_factor += 0.05 - (0.001 * (double)cpi->twopass.kf_zeromotion_pct);
+
+      // Convert the adjustment factor to a qindex delta
+      // on active_best_quality.
+      q_val = vp9_convert_qindex_to_q(active_best_quality, cm->bit_depth);
+      active_best_quality += vp9_compute_qdelta(rc, q_val,
+                                                q_val * q_adj_factor,
+                                                cm->bit_depth);
+    }
+  } else if (!rc->is_src_frame_alt_ref &&
+             (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+    // Use the lower of active_worst_quality and recent
+    // average Q as basis for GF/ARF best Q limit unless last frame was
+    // a key frame.
+    if (rc->frames_since_key > 1 &&
+        rc->avg_frame_qindex[INTER_FRAME] < active_worst_quality) {
+      q = rc->avg_frame_qindex[INTER_FRAME];
+    } else {
+      q = active_worst_quality;
+    }
+    // For constrained quality dont allow Q less than the cq level
+    if (oxcf->rc_mode == VPX_CQ) {
+      if (q < cq_level)
+        q = cq_level;
+
+      active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+
+      // Constrained quality use slightly lower active best.
+      active_best_quality = active_best_quality * 15 / 16;
+
+    } else if (oxcf->rc_mode == VPX_Q) {
+      if (!cpi->refresh_alt_ref_frame) {
+        active_best_quality = cq_level;
+      } else {
+        active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+
+        // Modify best quality for second level arfs. For mode VPX_Q this
+        // becomes the baseline frame q.
+        if (gf_group->rf_level[gf_group->index] == GF_ARF_LOW)
+          active_best_quality = (active_best_quality + cq_level + 1) / 2;
+      }
+    } else {
+      active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+    }
+  } else {
+    if (oxcf->rc_mode == VPX_Q) {
+      active_best_quality = cq_level;
+    } else {
+      active_best_quality = inter_minq[active_worst_quality];
+
+      // For the constrained quality mode we don't want
+      // q to fall below the cq level.
+      if ((oxcf->rc_mode == VPX_CQ) &&
+          (active_best_quality < cq_level)) {
+        active_best_quality = cq_level;
+      }
+    }
+  }
+
+  // Extension to max or min Q if undershoot or overshoot is outside
+  // the permitted range.
+  if (cpi->oxcf.rc_mode != VPX_Q) {
+    if (frame_is_intra_only(cm) ||
+        (!rc->is_src_frame_alt_ref &&
+         (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame))) {
+      active_best_quality -=
+        (cpi->twopass.extend_minq + cpi->twopass.extend_minq_fast);
+      active_worst_quality += (cpi->twopass.extend_maxq / 2);
+    } else {
+      active_best_quality -=
+        (cpi->twopass.extend_minq + cpi->twopass.extend_minq_fast) / 2;
+      active_worst_quality += cpi->twopass.extend_maxq;
+    }
+  }
+
+#if LIMIT_QRANGE_FOR_ALTREF_AND_KEY
+  vpx_clear_system_state();
+  // Static forced key frames Q restrictions dealt with elsewhere.
+  if (!((frame_is_intra_only(cm) || vp9_is_upper_layer_key_frame(cpi))) ||
+      !rc->this_key_frame_forced ||
+      (cpi->twopass.last_kfgroup_zeromotion_pct < STATIC_MOTION_THRESH)) {
+    int qdelta = vp9_frame_type_qdelta(cpi, gf_group->rf_level[gf_group->index],
+                                       active_worst_quality);
+    active_worst_quality = VPXMAX(active_worst_quality + qdelta,
+                                  active_best_quality);
+  }
+#endif
+
+  // Modify active_best_quality for downscaled normal frames.
+  if (rc->frame_size_selector != UNSCALED && !frame_is_kf_gf_arf(cpi)) {
+    int qdelta = vp9_compute_qdelta_by_rate(rc, cm->frame_type,
+                                            active_best_quality, 2.0,
+                                            cm->bit_depth);
+    active_best_quality =
+        VPXMAX(active_best_quality + qdelta, rc->best_quality);
+  }
+
+  active_best_quality = clamp(active_best_quality,
+                              rc->best_quality, rc->worst_quality);
+  active_worst_quality = clamp(active_worst_quality,
+                               active_best_quality, rc->worst_quality);
+
+  if (oxcf->rc_mode == VPX_Q) {
+    q = active_best_quality;
+  // Special case code to try and match quality with forced key frames.
+  } else if ((frame_is_intra_only(cm) || vp9_is_upper_layer_key_frame(cpi)) &&
+             rc->this_key_frame_forced) {
+    // If static since last kf use better of last boosted and last kf q.
+    if (cpi->twopass.last_kfgroup_zeromotion_pct >= STATIC_MOTION_THRESH) {
+      q = VPXMIN(rc->last_kf_qindex, rc->last_boosted_qindex);
+    } else {
+      q = rc->last_boosted_qindex;
+    }
+  } else {
+    q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
+                          active_best_quality, active_worst_quality);
+    if (q > active_worst_quality) {
+      // Special case when we are targeting the max allowed rate.
+      if (rc->this_frame_target >= rc->max_frame_bandwidth)
+        active_worst_quality = q;
+      else
+        q = active_worst_quality;
+    }
+  }
+  clamp(q, active_best_quality, active_worst_quality);
+
+  *top_index = active_worst_quality;
+  *bottom_index = active_best_quality;
+
+  assert(*top_index <= rc->worst_quality &&
+         *top_index >= rc->best_quality);
+  assert(*bottom_index <= rc->worst_quality &&
+         *bottom_index >= rc->best_quality);
+  assert(q <= rc->worst_quality && q >= rc->best_quality);
+  return q;
+}
+
+int vp9_rc_pick_q_and_bounds(const VP9_COMP *cpi,
+                             int *bottom_index, int *top_index) {
+  int q;
+  if (cpi->oxcf.pass == 0) {
+    if (cpi->oxcf.rc_mode == VPX_CBR)
+      q = rc_pick_q_and_bounds_one_pass_cbr(cpi, bottom_index, top_index);
+    else
+      q = rc_pick_q_and_bounds_one_pass_vbr(cpi, bottom_index, top_index);
+  } else {
+    q = rc_pick_q_and_bounds_two_pass(cpi, bottom_index, top_index);
+  }
+  if (cpi->sf.use_nonrd_pick_mode) {
+    if (cpi->sf.force_frame_boost == 1)
+      q -= cpi->sf.max_delta_qindex;
+
+    if (q < *bottom_index)
+      *bottom_index = q;
+    else if (q > *top_index)
+      *top_index = q;
+  }
+  return q;
+}
+
+void vp9_rc_compute_frame_size_bounds(const VP9_COMP *cpi,
+                                      int frame_target,
+                                      int *frame_under_shoot_limit,
+                                      int *frame_over_shoot_limit) {
+  if (cpi->oxcf.rc_mode == VPX_Q) {
+    *frame_under_shoot_limit = 0;
+    *frame_over_shoot_limit  = INT_MAX;
+  } else {
+    // For very small rate targets where the fractional adjustment
+    // may be tiny make sure there is at least a minimum range.
+    const int tolerance = (cpi->sf.recode_tolerance * frame_target) / 100;
+    *frame_under_shoot_limit = VPXMAX(frame_target - tolerance - 200, 0);
+    *frame_over_shoot_limit = VPXMIN(frame_target + tolerance + 200,
+                                     cpi->rc.max_frame_bandwidth);
+  }
+}
+
+void vp9_rc_set_frame_target(VP9_COMP *cpi, int target) {
+  const VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  rc->this_frame_target = target;
+
+  // Modify frame size target when down-scaling.
+  if (cpi->oxcf.resize_mode == RESIZE_DYNAMIC &&
+      rc->frame_size_selector != UNSCALED)
+    rc->this_frame_target = (int)(rc->this_frame_target
+        * rate_thresh_mult[rc->frame_size_selector]);
+
+  // Target rate per SB64 (including partial SB64s.
+  rc->sb64_target_rate = ((int64_t)rc->this_frame_target * 64 * 64) /
+                             (cm->width * cm->height);
+}
+
+static void update_alt_ref_frame_stats(VP9_COMP *cpi) {
+  // this frame refreshes means next frames don't unless specified by user
+  RATE_CONTROL *const rc = &cpi->rc;
+  rc->frames_since_golden = 0;
+
+  // Mark the alt ref as done (setting to 0 means no further alt refs pending).
+  rc->source_alt_ref_pending = 0;
+
+  // Set the alternate reference frame active flag
+  rc->source_alt_ref_active = 1;
+}
+
+static void update_golden_frame_stats(VP9_COMP *cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  // Update the Golden frame usage counts.
+  if (cpi->refresh_golden_frame) {
+    // this frame refreshes means next frames don't unless specified by user
+    rc->frames_since_golden = 0;
+
+    // If we are not using alt ref in the up and coming group clear the arf
+    // active flag. In multi arf group case, if the index is not 0 then
+    // we are overlaying a mid group arf so should not reset the flag.
+    if (cpi->oxcf.pass == 2) {
+      if (!rc->source_alt_ref_pending && (cpi->twopass.gf_group.index == 0))
+        rc->source_alt_ref_active = 0;
+    } else if (!rc->source_alt_ref_pending) {
+      rc->source_alt_ref_active = 0;
+    }
+
+    // Decrement count down till next gf
+    if (rc->frames_till_gf_update_due > 0)
+      rc->frames_till_gf_update_due--;
+
+  } else if (!cpi->refresh_alt_ref_frame) {
+    // Decrement count down till next gf
+    if (rc->frames_till_gf_update_due > 0)
+      rc->frames_till_gf_update_due--;
+
+    rc->frames_since_golden++;
+  }
+}
+
+static void compute_frame_low_motion(VP9_COMP *const cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  int mi_row, mi_col;
+  MODE_INFO **mi = cm->mi_grid_visible;
+  RATE_CONTROL *const rc = &cpi->rc;
+  const int rows = cm->mi_rows, cols = cm->mi_cols;
+  int cnt_zeromv = 0;
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      if (abs(mi[0]->mv[0].as_mv.row) < 16 &&
+          abs(mi[0]->mv[0].as_mv.col) < 16)
+        cnt_zeromv++;
+      mi++;
+    }
+    mi += 8;
+  }
+  cnt_zeromv = 100 * cnt_zeromv / (rows * cols);
+  rc->avg_frame_low_motion = (3 * rc->avg_frame_low_motion + cnt_zeromv) >> 2;
+}
+
+void vp9_rc_postencode_update(VP9_COMP *cpi, uint64_t bytes_used) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  RATE_CONTROL *const rc = &cpi->rc;
+  const int qindex = cm->base_qindex;
+
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
+    vp9_cyclic_refresh_postencode(cpi);
+  }
+
+  // Update rate control heuristics
+  rc->projected_frame_size = (int)(bytes_used << 3);
+
+  // Post encode loop adjustment of Q prediction.
+  vp9_rc_update_rate_correction_factors(cpi);
+
+  // Keep a record of last Q and ambient average Q.
+  if (cm->frame_type == KEY_FRAME) {
+    rc->last_q[KEY_FRAME] = qindex;
+    rc->avg_frame_qindex[KEY_FRAME] =
+        ROUND_POWER_OF_TWO(3 * rc->avg_frame_qindex[KEY_FRAME] + qindex, 2);
+    if (cpi->use_svc) {
+      int i = 0;
+      SVC *svc = &cpi->svc;
+      for (i = 0; i < svc->number_temporal_layers; ++i) {
+        const int layer = LAYER_IDS_TO_IDX(svc->spatial_layer_id, i,
+                                           svc->number_temporal_layers);
+        LAYER_CONTEXT *lc = &svc->layer_context[layer];
+        RATE_CONTROL *lrc = &lc->rc;
+        lrc->last_q[KEY_FRAME] = rc->last_q[KEY_FRAME];
+        lrc->avg_frame_qindex[KEY_FRAME] = rc->avg_frame_qindex[KEY_FRAME];
+      }
+    }
+  } else {
+    if ((cpi->use_svc && oxcf->rc_mode == VPX_CBR) ||
+        (!rc->is_src_frame_alt_ref &&
+         !(cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame))) {
+      rc->last_q[INTER_FRAME] = qindex;
+      rc->avg_frame_qindex[INTER_FRAME] =
+        ROUND_POWER_OF_TWO(3 * rc->avg_frame_qindex[INTER_FRAME] + qindex, 2);
+      rc->ni_frames++;
+      rc->tot_q += vp9_convert_qindex_to_q(qindex, cm->bit_depth);
+      rc->avg_q = rc->tot_q / rc->ni_frames;
+      // Calculate the average Q for normal inter frames (not key or GFU
+      // frames).
+      rc->ni_tot_qi += qindex;
+      rc->ni_av_qi = rc->ni_tot_qi / rc->ni_frames;
+    }
+  }
+
+  // Keep record of last boosted (KF/KF/ARF) Q value.
+  // If the current frame is coded at a lower Q then we also update it.
+  // If all mbs in this group are skipped only update if the Q value is
+  // better than that already stored.
+  // This is used to help set quality in forced key frames to reduce popping
+  if ((qindex < rc->last_boosted_qindex) ||
+      (cm->frame_type == KEY_FRAME) ||
+      (!rc->constrained_gf_group &&
+       (cpi->refresh_alt_ref_frame ||
+        (cpi->refresh_golden_frame && !rc->is_src_frame_alt_ref)))) {
+    rc->last_boosted_qindex = qindex;
+  }
+  if (cm->frame_type == KEY_FRAME)
+    rc->last_kf_qindex = qindex;
+
+  update_buffer_level(cpi, rc->projected_frame_size);
+
+  // Rolling monitors of whether we are over or underspending used to help
+  // regulate min and Max Q in two pass.
+  if (cm->frame_type != KEY_FRAME) {
+    rc->rolling_target_bits = ROUND_POWER_OF_TWO(
+        rc->rolling_target_bits * 3 + rc->this_frame_target, 2);
+    rc->rolling_actual_bits = ROUND_POWER_OF_TWO(
+        rc->rolling_actual_bits * 3 + rc->projected_frame_size, 2);
+    rc->long_rolling_target_bits = ROUND_POWER_OF_TWO(
+        rc->long_rolling_target_bits * 31 + rc->this_frame_target, 5);
+    rc->long_rolling_actual_bits = ROUND_POWER_OF_TWO(
+        rc->long_rolling_actual_bits * 31 + rc->projected_frame_size, 5);
+  }
+
+  // Actual bits spent
+  rc->total_actual_bits += rc->projected_frame_size;
+  rc->total_target_bits += cm->show_frame ? rc->avg_frame_bandwidth : 0;
+
+  rc->total_target_vs_actual = rc->total_actual_bits - rc->total_target_bits;
+
+  if (!cpi->use_svc || is_two_pass_svc(cpi)) {
+    if (is_altref_enabled(cpi) && cpi->refresh_alt_ref_frame &&
+        (cm->frame_type != KEY_FRAME))
+      // Update the alternate reference frame stats as appropriate.
+      update_alt_ref_frame_stats(cpi);
+    else
+      // Update the Golden frame stats as appropriate.
+      update_golden_frame_stats(cpi);
+  }
+
+  if (cm->frame_type == KEY_FRAME)
+    rc->frames_since_key = 0;
+  if (cm->show_frame) {
+    rc->frames_since_key++;
+    rc->frames_to_key--;
+  }
+
+  // Trigger the resizing of the next frame if it is scaled.
+  if (oxcf->pass != 0) {
+    cpi->resize_pending =
+        rc->next_frame_size_selector != rc->frame_size_selector;
+    rc->frame_size_selector = rc->next_frame_size_selector;
+  }
+
+  if (oxcf->pass == 0) {
+    if (cm->frame_type != KEY_FRAME)
+      compute_frame_low_motion(cpi);
+  }
+}
+
+void vp9_rc_postencode_update_drop_frame(VP9_COMP *cpi) {
+  // Update buffer level with zero size, update frame counters, and return.
+  update_buffer_level(cpi, 0);
+  cpi->rc.frames_since_key++;
+  cpi->rc.frames_to_key--;
+  cpi->rc.rc_2_frame = 0;
+  cpi->rc.rc_1_frame = 0;
+}
+
+// Use this macro to turn on/off use of alt-refs in one-pass mode.
+#define USE_ALTREF_FOR_ONE_PASS   1
+
+static int calc_pframe_target_size_one_pass_vbr(const VP9_COMP *const cpi) {
+  static const int af_ratio = 10;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  int target;
+#if USE_ALTREF_FOR_ONE_PASS
+  target = (!rc->is_src_frame_alt_ref &&
+            (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) ?
+      (rc->avg_frame_bandwidth * rc->baseline_gf_interval * af_ratio) /
+      (rc->baseline_gf_interval + af_ratio - 1) :
+      (rc->avg_frame_bandwidth * rc->baseline_gf_interval) /
+      (rc->baseline_gf_interval + af_ratio - 1);
+#else
+  target = rc->avg_frame_bandwidth;
+#endif
+  return vp9_rc_clamp_pframe_target_size(cpi, target);
+}
+
+static int calc_iframe_target_size_one_pass_vbr(const VP9_COMP *const cpi) {
+  static const int kf_ratio = 25;
+  const RATE_CONTROL *rc = &cpi->rc;
+  const int target = rc->avg_frame_bandwidth * kf_ratio;
+  return vp9_rc_clamp_iframe_target_size(cpi, target);
+}
+
+static void adjust_gf_key_frame(VP9_COMP *cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  rc->constrained_gf_group = 0;
+  // Reset gf interval to make more equal spacing for up-coming key frame.
+  if ((rc->frames_to_key <= 7 * rc->baseline_gf_interval >> 2) &&
+      (rc->frames_to_key > rc->baseline_gf_interval)) {
+    rc->baseline_gf_interval = rc->frames_to_key >> 1;
+    if (rc->baseline_gf_interval < 5)
+      rc->baseline_gf_interval = rc->frames_to_key;
+    rc->constrained_gf_group = 1;
+  } else {
+    // Reset since frames_till_gf_update_due must be <= frames_to_key.
+    if (rc->baseline_gf_interval > rc->frames_to_key) {
+      rc->baseline_gf_interval = rc->frames_to_key;
+      rc->constrained_gf_group = 1;
+    }
+  }
+}
+void vp9_rc_get_one_pass_vbr_params(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int target;
+  // TODO(yaowu): replace the "auto_key && 0" below with proper decision logic.
+  if (!cpi->refresh_alt_ref_frame &&
+      (cm->current_video_frame == 0 ||
+       (cpi->frame_flags & FRAMEFLAGS_KEY) ||
+       rc->frames_to_key == 0 ||
+       (cpi->oxcf.auto_key && 0))) {
+    cm->frame_type = KEY_FRAME;
+    rc->this_key_frame_forced = cm->current_video_frame != 0 &&
+                                rc->frames_to_key == 0;
+    rc->frames_to_key = cpi->oxcf.key_freq;
+    rc->kf_boost = DEFAULT_KF_BOOST;
+    rc->source_alt_ref_active = 0;
+  } else {
+    cm->frame_type = INTER_FRAME;
+  }
+  if (rc->frames_till_gf_update_due == 0) {
+    double rate_err = 1.0;
+    rc->gfu_boost = DEFAULT_GF_BOOST;
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cpi->oxcf.pass == 0) {
+      vp9_cyclic_refresh_set_golden_update(cpi);
+    } else {
+      rc->baseline_gf_interval =
+          (rc->min_gf_interval + rc->max_gf_interval) / 2;
+    }
+    if (rc->rolling_target_bits > 0)
+      rate_err =
+          (double)rc->rolling_actual_bits / (double)rc->rolling_target_bits;
+    // Increase gf interval at high Q and high overshoot.
+    if (cm->current_video_frame > 30 &&
+        rc->avg_frame_qindex[INTER_FRAME] > (7 * rc->worst_quality) >> 3 &&
+        rate_err > 3.5) {
+      rc->baseline_gf_interval =
+          VPXMIN(15, (3 * rc->baseline_gf_interval) >> 1);
+    } else if (cm->current_video_frame > 30 &&
+               rc->avg_frame_low_motion < 20) {
+      // Decrease boost and gf interval for high motion case.
+      rc->gfu_boost = DEFAULT_GF_BOOST >> 1;
+      rc->baseline_gf_interval = VPXMAX(5, rc->baseline_gf_interval >> 1);
+    }
+    adjust_gf_key_frame(cpi);
+    rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+    cpi->refresh_golden_frame = 1;
+    rc->source_alt_ref_pending = USE_ALTREF_FOR_ONE_PASS;
+  }
+  if (cm->frame_type == KEY_FRAME)
+    target = calc_iframe_target_size_one_pass_vbr(cpi);
+  else
+    target = calc_pframe_target_size_one_pass_vbr(cpi);
+  vp9_rc_set_frame_target(cpi, target);
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cpi->oxcf.pass == 0)
+    vp9_cyclic_refresh_update_parameters(cpi);
+}
+
+static int calc_pframe_target_size_one_pass_cbr(const VP9_COMP *cpi) {
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
+  const RATE_CONTROL *rc = &cpi->rc;
+  const SVC *const svc = &cpi->svc;
+  const int64_t diff = rc->optimal_buffer_level - rc->buffer_level;
+  const int64_t one_pct_bits = 1 + rc->optimal_buffer_level / 100;
+  int min_frame_target =
+      VPXMAX(rc->avg_frame_bandwidth >> 4, FRAME_OVERHEAD_BITS);
+  int target;
+
+  if (oxcf->gf_cbr_boost_pct) {
+    const int af_ratio_pct = oxcf->gf_cbr_boost_pct + 100;
+    target =  cpi->refresh_golden_frame ?
+      (rc->avg_frame_bandwidth * rc->baseline_gf_interval * af_ratio_pct) /
+      (rc->baseline_gf_interval * 100 + af_ratio_pct - 100) :
+      (rc->avg_frame_bandwidth * rc->baseline_gf_interval * 100) /
+      (rc->baseline_gf_interval * 100 + af_ratio_pct - 100);
+  } else {
+    target = rc->avg_frame_bandwidth;
+  }
+  if (is_one_pass_cbr_svc(cpi)) {
+    // Note that for layers, avg_frame_bandwidth is the cumulative
+    // per-frame-bandwidth. For the target size of this frame, use the
+    // layer average frame size (i.e., non-cumulative per-frame-bw).
+    int layer =
+        LAYER_IDS_TO_IDX(svc->spatial_layer_id,
+            svc->temporal_layer_id, svc->number_temporal_layers);
+    const LAYER_CONTEXT *lc = &svc->layer_context[layer];
+    target = lc->avg_frame_size;
+    min_frame_target = VPXMAX(lc->avg_frame_size >> 4, FRAME_OVERHEAD_BITS);
+  }
+  if (diff > 0) {
+    // Lower the target bandwidth for this frame.
+    const int pct_low = (int)VPXMIN(diff / one_pct_bits, oxcf->under_shoot_pct);
+    target -= (target * pct_low) / 200;
+  } else if (diff < 0) {
+    // Increase the target bandwidth for this frame.
+    const int pct_high =
+        (int)VPXMIN(-diff / one_pct_bits, oxcf->over_shoot_pct);
+    target += (target * pct_high) / 200;
+  }
+  if (oxcf->rc_max_inter_bitrate_pct) {
+    const int max_rate = rc->avg_frame_bandwidth *
+                         oxcf->rc_max_inter_bitrate_pct / 100;
+    target = VPXMIN(target, max_rate);
+  }
+  return VPXMAX(min_frame_target, target);
+}
+
+static int calc_iframe_target_size_one_pass_cbr(const VP9_COMP *cpi) {
+  const RATE_CONTROL *rc = &cpi->rc;
+  const VP9EncoderConfig *oxcf = &cpi->oxcf;
+  const SVC *const svc = &cpi->svc;
+  int target;
+  if (cpi->common.current_video_frame == 0) {
+    target = ((rc->starting_buffer_level / 2) > INT_MAX)
+      ? INT_MAX : (int)(rc->starting_buffer_level / 2);
+  } else {
+    int kf_boost = 32;
+    double framerate = cpi->framerate;
+    if (svc->number_temporal_layers > 1 &&
+        oxcf->rc_mode == VPX_CBR) {
+      // Use the layer framerate for temporal layers CBR mode.
+      const int layer = LAYER_IDS_TO_IDX(svc->spatial_layer_id,
+          svc->temporal_layer_id, svc->number_temporal_layers);
+      const LAYER_CONTEXT *lc = &svc->layer_context[layer];
+      framerate = lc->framerate;
+    }
+    kf_boost = VPXMAX(kf_boost, (int)(2 * framerate - 16));
+    if (rc->frames_since_key <  framerate / 2) {
+      kf_boost = (int)(kf_boost * rc->frames_since_key /
+                       (framerate / 2));
+    }
+    target = ((16 + kf_boost) * rc->avg_frame_bandwidth) >> 4;
+  }
+  return vp9_rc_clamp_iframe_target_size(cpi, target);
+}
+
+void vp9_rc_get_svc_params(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int target = rc->avg_frame_bandwidth;
+  int layer = LAYER_IDS_TO_IDX(cpi->svc.spatial_layer_id,
+      cpi->svc.temporal_layer_id, cpi->svc.number_temporal_layers);
+  // Periodic key frames is based on the super-frame counter
+  // (svc.current_superframe), also only base spatial layer is key frame.
+  if ((cm->current_video_frame == 0) ||
+      (cpi->frame_flags & FRAMEFLAGS_KEY) ||
+      (cpi->oxcf.auto_key &&
+       (cpi->svc.current_superframe % cpi->oxcf.key_freq == 0) &&
+       cpi->svc.spatial_layer_id == 0)) {
+    cm->frame_type = KEY_FRAME;
+    rc->source_alt_ref_active = 0;
+    if (is_two_pass_svc(cpi)) {
+      cpi->svc.layer_context[layer].is_key_frame = 1;
+      cpi->ref_frame_flags &=
+          (~VP9_LAST_FLAG & ~VP9_GOLD_FLAG & ~VP9_ALT_FLAG);
+    } else if (is_one_pass_cbr_svc(cpi)) {
+      if (cm->current_video_frame > 0)
+        vp9_svc_reset_key_frame(cpi);
+      layer = LAYER_IDS_TO_IDX(cpi->svc.spatial_layer_id,
+           cpi->svc.temporal_layer_id, cpi->svc.number_temporal_layers);
+      cpi->svc.layer_context[layer].is_key_frame = 1;
+      cpi->ref_frame_flags &=
+                (~VP9_LAST_FLAG & ~VP9_GOLD_FLAG & ~VP9_ALT_FLAG);
+      // Assumption here is that LAST_FRAME is being updated for a keyframe.
+      // Thus no change in update flags.
+      target = calc_iframe_target_size_one_pass_cbr(cpi);
+    }
+  } else {
+    cm->frame_type = INTER_FRAME;
+    if (is_two_pass_svc(cpi)) {
+      LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
+      if (cpi->svc.spatial_layer_id == 0) {
+        lc->is_key_frame = 0;
+      } else {
+        lc->is_key_frame =
+            cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame;
+        if (lc->is_key_frame)
+          cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
+      }
+      cpi->ref_frame_flags &= (~VP9_ALT_FLAG);
+    } else if (is_one_pass_cbr_svc(cpi)) {
+      LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
+      if (cpi->svc.spatial_layer_id == cpi->svc.first_spatial_layer_to_encode) {
+        lc->is_key_frame = 0;
+      } else {
+        lc->is_key_frame =
+            cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame;
+      }
+      target = calc_pframe_target_size_one_pass_cbr(cpi);
+    }
+  }
+
+  // Any update/change of global cyclic refresh parameters (amount/delta-qp)
+  // should be done here, before the frame qp is selected.
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
+    vp9_cyclic_refresh_update_parameters(cpi);
+
+  vp9_rc_set_frame_target(cpi, target);
+  rc->frames_till_gf_update_due = INT_MAX;
+  rc->baseline_gf_interval = INT_MAX;
+}
+
+void vp9_rc_get_one_pass_cbr_params(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int target;
+  // TODO(yaowu): replace the "auto_key && 0" below with proper decision logic.
+  if ((cm->current_video_frame == 0 ||
+      (cpi->frame_flags & FRAMEFLAGS_KEY) ||
+      rc->frames_to_key == 0 ||
+      (cpi->oxcf.auto_key && 0))) {
+    cm->frame_type = KEY_FRAME;
+    rc->this_key_frame_forced = cm->current_video_frame != 0 &&
+                                rc->frames_to_key == 0;
+    rc->frames_to_key = cpi->oxcf.key_freq;
+    rc->kf_boost = DEFAULT_KF_BOOST;
+    rc->source_alt_ref_active = 0;
+  } else {
+    cm->frame_type = INTER_FRAME;
+  }
+  if (rc->frames_till_gf_update_due == 0) {
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
+      vp9_cyclic_refresh_set_golden_update(cpi);
+    else
+      rc->baseline_gf_interval =
+          (rc->min_gf_interval + rc->max_gf_interval) / 2;
+    rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+    // NOTE: frames_till_gf_update_due must be <= frames_to_key.
+    if (rc->frames_till_gf_update_due > rc->frames_to_key)
+      rc->frames_till_gf_update_due = rc->frames_to_key;
+    cpi->refresh_golden_frame = 1;
+    rc->gfu_boost = DEFAULT_GF_BOOST;
+  }
+
+  // Any update/change of global cyclic refresh parameters (amount/delta-qp)
+  // should be done here, before the frame qp is selected.
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
+    vp9_cyclic_refresh_update_parameters(cpi);
+
+  if (cm->frame_type == KEY_FRAME)
+    target = calc_iframe_target_size_one_pass_cbr(cpi);
+  else
+    target = calc_pframe_target_size_one_pass_cbr(cpi);
+
+  vp9_rc_set_frame_target(cpi, target);
+  if (cpi->oxcf.resize_mode == RESIZE_DYNAMIC)
+    cpi->resize_pending = vp9_resize_one_pass_cbr(cpi);
+  else
+    cpi->resize_pending = 0;
+}
+
+int vp9_compute_qdelta(const RATE_CONTROL *rc, double qstart, double qtarget,
+                       vpx_bit_depth_t bit_depth) {
+  int start_index = rc->worst_quality;
+  int target_index = rc->worst_quality;
+  int i;
+
+  // Convert the average q value to an index.
+  for (i = rc->best_quality; i < rc->worst_quality; ++i) {
+    start_index = i;
+    if (vp9_convert_qindex_to_q(i, bit_depth) >= qstart)
+      break;
+  }
+
+  // Convert the q target to an index
+  for (i = rc->best_quality; i < rc->worst_quality; ++i) {
+    target_index = i;
+    if (vp9_convert_qindex_to_q(i, bit_depth) >= qtarget)
+      break;
+  }
+
+  return target_index - start_index;
+}
+
+int vp9_compute_qdelta_by_rate(const RATE_CONTROL *rc, FRAME_TYPE frame_type,
+                               int qindex, double rate_target_ratio,
+                               vpx_bit_depth_t bit_depth) {
+  int target_index = rc->worst_quality;
+  int i;
+
+  // Look up the current projected bits per block for the base index
+  const int base_bits_per_mb = vp9_rc_bits_per_mb(frame_type, qindex, 1.0,
+                                                  bit_depth);
+
+  // Find the target bits per mb based on the base value and given ratio.
+  const int target_bits_per_mb = (int)(rate_target_ratio * base_bits_per_mb);
+
+  // Convert the q target to an index
+  for (i = rc->best_quality; i < rc->worst_quality; ++i) {
+    if (vp9_rc_bits_per_mb(frame_type, i, 1.0, bit_depth) <=
+        target_bits_per_mb) {
+      target_index = i;
+      break;
+    }
+  }
+  return target_index - qindex;
+}
+
+void vp9_rc_set_gf_interval_range(const VP9_COMP *const cpi,
+                                  RATE_CONTROL *const rc) {
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+
+  // Special case code for 1 pass fixed Q mode tests
+  if ((oxcf->pass == 0) && (oxcf->rc_mode == VPX_Q)) {
+    rc->max_gf_interval = FIXED_GF_INTERVAL;
+    rc->min_gf_interval = FIXED_GF_INTERVAL;
+    rc->static_scene_max_gf_interval = FIXED_GF_INTERVAL;
+  } else {
+    // Set Maximum gf/arf interval
+    rc->max_gf_interval = oxcf->max_gf_interval;
+    rc->min_gf_interval = oxcf->min_gf_interval;
+    if (rc->min_gf_interval == 0)
+      rc->min_gf_interval = vp9_rc_get_default_min_gf_interval(
+          oxcf->width, oxcf->height, cpi->framerate);
+    if (rc->max_gf_interval == 0)
+      rc->max_gf_interval = vp9_rc_get_default_max_gf_interval(
+          cpi->framerate, rc->min_gf_interval);
+
+    // Extended interval for genuinely static scenes
+    rc->static_scene_max_gf_interval = MAX_LAG_BUFFERS * 2;
+
+    if (is_altref_enabled(cpi)) {
+      if (rc->static_scene_max_gf_interval > oxcf->lag_in_frames - 1)
+        rc->static_scene_max_gf_interval = oxcf->lag_in_frames - 1;
+    }
+
+    if (rc->max_gf_interval > rc->static_scene_max_gf_interval)
+      rc->max_gf_interval = rc->static_scene_max_gf_interval;
+
+    // Clamp min to max
+    rc->min_gf_interval = VPXMIN(rc->min_gf_interval, rc->max_gf_interval);
+  }
+}
+
+void vp9_rc_update_framerate(VP9_COMP *cpi) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int vbr_max_bits;
+
+  rc->avg_frame_bandwidth = (int)(oxcf->target_bandwidth / cpi->framerate);
+  rc->min_frame_bandwidth = (int)(rc->avg_frame_bandwidth *
+                                oxcf->two_pass_vbrmin_section / 100);
+
+  rc->min_frame_bandwidth =
+      VPXMAX(rc->min_frame_bandwidth, FRAME_OVERHEAD_BITS);
+
+  // A maximum bitrate for a frame is defined.
+  // The baseline for this aligns with HW implementations that
+  // can support decode of 1080P content up to a bitrate of MAX_MB_RATE bits
+  // per 16x16 MB (averaged over a frame). However this limit is extended if
+  // a very high rate is given on the command line or the the rate cannnot
+  // be acheived because of a user specificed max q (e.g. when the user
+  // specifies lossless encode.
+  vbr_max_bits = (int)(((int64_t)rc->avg_frame_bandwidth *
+                     oxcf->two_pass_vbrmax_section) / 100);
+  rc->max_frame_bandwidth =
+      VPXMAX(VPXMAX((cm->MBs * MAX_MB_RATE), MAXRATE_1080P), vbr_max_bits);
+
+  vp9_rc_set_gf_interval_range(cpi, rc);
+}
+
+#define VBR_PCT_ADJUSTMENT_LIMIT 50
+// For VBR...adjustment to the frame target based on error from previous frames
+static void vbr_rate_correction(VP9_COMP *cpi, int *this_frame_target) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  int64_t vbr_bits_off_target = rc->vbr_bits_off_target;
+  int max_delta;
+  int frame_window = VPXMIN(16,
+      ((int)cpi->twopass.total_stats.count - cpi->common.current_video_frame));
+
+  // Calcluate the adjustment to rate for this frame.
+  if (frame_window > 0) {
+    max_delta = (vbr_bits_off_target > 0)
+        ? (int)(vbr_bits_off_target / frame_window)
+        : (int)(-vbr_bits_off_target / frame_window);
+
+    max_delta = VPXMIN(max_delta,
+        ((*this_frame_target * VBR_PCT_ADJUSTMENT_LIMIT) / 100));
+
+    // vbr_bits_off_target > 0 means we have extra bits to spend
+    if (vbr_bits_off_target > 0) {
+      *this_frame_target +=
+        (vbr_bits_off_target > max_delta) ? max_delta
+                                          : (int)vbr_bits_off_target;
+    } else {
+      *this_frame_target -=
+        (vbr_bits_off_target < -max_delta) ? max_delta
+                                           : (int)-vbr_bits_off_target;
+    }
+  }
+
+  // Fast redistribution of bits arising from massive local undershoot.
+  // Dont do it for kf,arf,gf or overlay frames.
+  if (!frame_is_kf_gf_arf(cpi) && !rc->is_src_frame_alt_ref &&
+      rc->vbr_bits_off_target_fast) {
+    int one_frame_bits = VPXMAX(rc->avg_frame_bandwidth, *this_frame_target);
+    int fast_extra_bits;
+    fast_extra_bits = (int)VPXMIN(rc->vbr_bits_off_target_fast, one_frame_bits);
+    fast_extra_bits = (int)VPXMIN(
+        fast_extra_bits,
+        VPXMAX(one_frame_bits / 8, rc->vbr_bits_off_target_fast / 8));
+    *this_frame_target += (int)fast_extra_bits;
+    rc->vbr_bits_off_target_fast -= fast_extra_bits;
+  }
+}
+
+void vp9_set_target_rate(VP9_COMP *cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  int target_rate = rc->base_frame_target;
+
+  if (cpi->common.frame_type == KEY_FRAME)
+    target_rate = vp9_rc_clamp_iframe_target_size(cpi, target_rate);
+  else
+    target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
+
+  // Correction to rate target based on prior over or under shoot.
+  if (cpi->oxcf.rc_mode == VPX_VBR || cpi->oxcf.rc_mode == VPX_CQ)
+    vbr_rate_correction(cpi, &target_rate);
+  vp9_rc_set_frame_target(cpi, target_rate);
+}
+
+// Check if we should resize, based on average QP from past x frames.
+// Only allow for resize at most one scale down for now, scaling factor is 2.
+int vp9_resize_one_pass_cbr(VP9_COMP *cpi) {
+  const VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  RESIZE_ACTION resize_action = NO_RESIZE;
+  int avg_qp_thr1 = 70;
+  int avg_qp_thr2 = 50;
+  int min_width = 180;
+  int min_height = 180;
+  int down_size_on = 1;
+  cpi->resize_scale_num = 1;
+  cpi->resize_scale_den = 1;
+  // Don't resize on key frame; reset the counters on key frame.
+  if (cm->frame_type == KEY_FRAME) {
+    cpi->resize_avg_qp = 0;
+    cpi->resize_count = 0;
+    return 0;
+  }
+  // Check current frame reslution to avoid generating frames smaller than
+  // the minimum resolution.
+  if (ONEHALFONLY_RESIZE) {
+    if ((cm->width >> 1) < min_width || (cm->height >> 1) < min_height)
+      down_size_on = 0;
+  } else {
+    if (cpi->resize_state == ORIG &&
+        (cm->width * 3 / 4 < min_width ||
+         cm->height * 3 / 4 < min_height))
+      return 0;
+    else if (cpi->resize_state == THREE_QUARTER &&
+             ((cpi->oxcf.width >> 1) < min_width ||
+              (cpi->oxcf.height >> 1) < min_height))
+      down_size_on = 0;
+  }
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  // If denoiser is on, apply a smaller qp threshold.
+  if (cpi->oxcf.noise_sensitivity > 0) {
+    avg_qp_thr1 = 60;
+    avg_qp_thr2 = 40;
+  }
+#endif
+
+  // Resize based on average buffer underflow and QP over some window.
+  // Ignore samples close to key frame, since QP is usually high after key.
+  if (cpi->rc.frames_since_key > 2 * cpi->framerate) {
+    const int window = (int)(4 * cpi->framerate);
+    cpi->resize_avg_qp += cm->base_qindex;
+    if (cpi->rc.buffer_level < (int)(30 * rc->optimal_buffer_level / 100))
+      ++cpi->resize_buffer_underflow;
+    ++cpi->resize_count;
+    // Check for resize action every "window" frames.
+    if (cpi->resize_count >= window) {
+      int avg_qp = cpi->resize_avg_qp / cpi->resize_count;
+      // Resize down if buffer level has underflowed sufficient amount in past
+      // window, and we are at original or 3/4 of original resolution.
+      // Resize back up if average QP is low, and we are currently in a resized
+      // down state, i.e. 1/2 or 3/4 of original resolution.
+      // Currently, use a flag to turn 3/4 resizing feature on/off.
+      if (cpi->resize_buffer_underflow > (cpi->resize_count >> 2)) {
+        if (cpi->resize_state == THREE_QUARTER && down_size_on) {
+          resize_action = DOWN_ONEHALF;
+          cpi->resize_state = ONE_HALF;
+        } else if (cpi->resize_state == ORIG) {
+          resize_action = ONEHALFONLY_RESIZE ? DOWN_ONEHALF : DOWN_THREEFOUR;
+          cpi->resize_state = ONEHALFONLY_RESIZE ? ONE_HALF : THREE_QUARTER;
+        }
+      } else if (cpi->resize_state != ORIG &&
+                 avg_qp < avg_qp_thr1 * cpi->rc.worst_quality / 100) {
+        if (cpi->resize_state == THREE_QUARTER ||
+            avg_qp < avg_qp_thr2 * cpi->rc.worst_quality / 100 ||
+            ONEHALFONLY_RESIZE) {
+          resize_action = UP_ORIG;
+          cpi->resize_state = ORIG;
+        } else if (cpi->resize_state == ONE_HALF) {
+          resize_action = UP_THREEFOUR;
+          cpi->resize_state = THREE_QUARTER;
+        }
+      }
+      // Reset for next window measurement.
+      cpi->resize_avg_qp = 0;
+      cpi->resize_count = 0;
+      cpi->resize_buffer_underflow = 0;
+    }
+  }
+  // If decision is to resize, reset some quantities, and check is we should
+  // reduce rate correction factor,
+  if (resize_action != NO_RESIZE) {
+    int target_bits_per_frame;
+    int active_worst_quality;
+    int qindex;
+    int tot_scale_change;
+    if (resize_action == DOWN_THREEFOUR || resize_action == UP_THREEFOUR) {
+      cpi->resize_scale_num = 3;
+      cpi->resize_scale_den = 4;
+    } else if (resize_action == DOWN_ONEHALF) {
+      cpi->resize_scale_num = 1;
+      cpi->resize_scale_den = 2;
+    } else {  // UP_ORIG or anything else
+      cpi->resize_scale_num = 1;
+      cpi->resize_scale_den = 1;
+    }
+    tot_scale_change = (cpi->resize_scale_den * cpi->resize_scale_den) /
+        (cpi->resize_scale_num * cpi->resize_scale_num);
+    // Reset buffer level to optimal, update target size.
+    rc->buffer_level = rc->optimal_buffer_level;
+    rc->bits_off_target = rc->optimal_buffer_level;
+    rc->this_frame_target = calc_pframe_target_size_one_pass_cbr(cpi);
+    // Get the projected qindex, based on the scaled target frame size (scaled
+    // so target_bits_per_mb in vp9_rc_regulate_q will be correct target).
+    target_bits_per_frame = (resize_action >= 0) ?
+        rc->this_frame_target * tot_scale_change :
+        rc->this_frame_target / tot_scale_change;
+    active_worst_quality = calc_active_worst_quality_one_pass_cbr(cpi);
+    qindex = vp9_rc_regulate_q(cpi,
+                               target_bits_per_frame,
+                               rc->best_quality,
+                               active_worst_quality);
+    // If resize is down, check if projected q index is close to worst_quality,
+    // and if so, reduce the rate correction factor (since likely can afford
+    // lower q for resized frame).
+    if (resize_action > 0 &&
+        qindex > 90 * cpi->rc.worst_quality / 100) {
+      rc->rate_correction_factors[INTER_NORMAL] *= 0.85;
+    }
+    // If resize is back up, check if projected q index is too much above the
+    // current base_qindex, and if so, reduce the rate correction factor
+    // (since prefer to keep q for resized frame at least close to previous q).
+    if (resize_action < 0 &&
+       qindex > 130 * cm->base_qindex / 100) {
+      rc->rate_correction_factors[INTER_NORMAL] *= 0.9;
+    }
+  }
+  return resize_action;
+}
+
+// Compute average source sad (temporal sad: between current source and
+// previous source) over a subset of superblocks. Use this is detect big changes
+// in content and allow rate control to react.
+// TODO(marpan): Superblock sad is computed again in variance partition for
+// non-rd mode (but based on last reconstructed frame). Should try to reuse
+// these computations.
+void vp9_avg_source_sad(VP9_COMP *cpi) {
+  VP9_COMMON * const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  rc->high_source_sad = 0;
+  if (cpi->Last_Source != NULL &&
+      cpi->Last_Source->y_width == cpi->Source->y_width &&
+      cpi->Last_Source->y_height == cpi->Source->y_height) {
+    const uint8_t *src_y = cpi->Source->y_buffer;
+    const int src_ystride = cpi->Source->y_stride;
+    const uint8_t *last_src_y = cpi->Last_Source->y_buffer;
+    const int last_src_ystride = cpi->Last_Source->y_stride;
+    int sbi_row, sbi_col;
+    const BLOCK_SIZE bsize = BLOCK_64X64;
+    uint32_t min_thresh = 4000;
+    float thresh = 8.0f;
+    // Loop over sub-sample of frame, and compute average sad over 64x64 blocks.
+    uint64_t avg_sad = 0;
+    int num_samples = 0;
+    int sb_cols = (cm->mi_cols + MI_BLOCK_SIZE - 1) / MI_BLOCK_SIZE;
+    int sb_rows = (cm->mi_rows + MI_BLOCK_SIZE - 1) / MI_BLOCK_SIZE;
+    for (sbi_row = 0; sbi_row < sb_rows; sbi_row ++) {
+      for (sbi_col = 0; sbi_col < sb_cols; sbi_col ++) {
+        // Checker-board pattern, ignore boundary.
+        if ((sbi_row > 0 && sbi_col > 0) &&
+            (sbi_row < sb_rows - 1 && sbi_col < sb_cols - 1) &&
+            ((sbi_row % 2 == 0 && sbi_col % 2 == 0) ||
+            (sbi_row % 2 != 0 && sbi_col % 2 != 0))) {
+          num_samples++;
+          avg_sad += cpi->fn_ptr[bsize].sdf(src_y,
+                                            src_ystride,
+                                            last_src_y,
+                                            last_src_ystride);
+        }
+        src_y += 64;
+        last_src_y += 64;
+      }
+      src_y += (src_ystride << 6) - (sb_cols << 6);
+      last_src_y += (last_src_ystride << 6) - (sb_cols << 6);
+    }
+    if (num_samples > 0)
+      avg_sad = avg_sad / num_samples;
+    // Set high_source_sad flag if we detect very high increase in avg_sad
+    // between current and the previous frame value(s). Use a minimum threshold
+    // for cases where there is small change from content that is completely
+    // static.
+    if (cpi->oxcf.rc_mode == VPX_VBR) {
+      min_thresh = 60000;
+      thresh = 2.1f;
+    }
+    if (avg_sad >
+        VPXMAX(min_thresh, (unsigned int)(rc->avg_source_sad  * thresh)) &&
+        rc->frames_since_key > 1)
+      rc->high_source_sad = 1;
+    else
+      rc->high_source_sad = 0;
+    if (avg_sad > 0 || cpi->oxcf.rc_mode == VPX_CBR)
+      rc->avg_source_sad = (3 * rc->avg_source_sad + avg_sad) >> 2;
+    // For VBR, under scene change/high content change, force golden refresh.
+    if (cpi->oxcf.rc_mode == VPX_VBR &&
+        rc->high_source_sad &&
+        rc->frames_to_key > 3 &&
+        rc->count_last_scene_change > 4 &&
+        cpi->ext_refresh_frame_flags_pending == 0) {
+      int target;
+      cpi->refresh_golden_frame = 1;
+      rc->gfu_boost = DEFAULT_GF_BOOST >> 1;
+      rc->baseline_gf_interval = VPXMIN(20,
+          VPXMAX(10, rc->baseline_gf_interval));
+      adjust_gf_key_frame(cpi);
+      rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+      target = calc_pframe_target_size_one_pass_vbr(cpi);
+      vp9_rc_set_frame_target(cpi, target);
+      rc->count_last_scene_change = 0;
+    } else {
+      rc->count_last_scene_change++;
+    }
+  }
+}
+
+// Test if encoded frame will significantly overshoot the target bitrate, and
+// if so, set the QP, reset/adjust some rate control parameters, and return 1.
+int vp9_encodedframe_overshoot(VP9_COMP *cpi,
+                               int frame_size,
+                               int *q) {
+  VP9_COMMON * const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int thresh_qp = 3 * (rc->worst_quality >> 2);
+  int thresh_rate = rc->avg_frame_bandwidth * 10;
+  if (cm->base_qindex < thresh_qp &&
+      frame_size > thresh_rate) {
+    double rate_correction_factor =
+        cpi->rc.rate_correction_factors[INTER_NORMAL];
+    const int target_size = cpi->rc.avg_frame_bandwidth;
+    double new_correction_factor;
+    int target_bits_per_mb;
+    double q2;
+    int enumerator;
+    // Force a re-encode, and for now use max-QP.
+    *q = cpi->rc.worst_quality;
+    // Adjust avg_frame_qindex, buffer_level, and rate correction factors, as
+    // these parameters will affect QP selection for subsequent frames. If they
+    // have settled down to a very different (low QP) state, then not adjusting
+    // them may cause next frame to select low QP and overshoot again.
+    cpi->rc.avg_frame_qindex[INTER_FRAME] = *q;
+    rc->buffer_level = rc->optimal_buffer_level;
+    rc->bits_off_target = rc->optimal_buffer_level;
+    // Reset rate under/over-shoot flags.
+    cpi->rc.rc_1_frame = 0;
+    cpi->rc.rc_2_frame = 0;
+    // Adjust rate correction factor.
+    target_bits_per_mb = ((uint64_t)target_size << BPER_MB_NORMBITS) / cm->MBs;
+    // Rate correction factor based on target_bits_per_mb and qp (==max_QP).
+    // This comes from the inverse computation of vp9_rc_bits_per_mb().
+    q2 = vp9_convert_qindex_to_q(*q, cm->bit_depth);
+    enumerator = 1800000;  // Factor for inter frame.
+    enumerator += (int)(enumerator * q2) >> 12;
+    new_correction_factor = (double)target_bits_per_mb * q2 / enumerator;
+    if (new_correction_factor > rate_correction_factor) {
+      rate_correction_factor =
+          VPXMIN(2.0 * rate_correction_factor, new_correction_factor);
+      if (rate_correction_factor > MAX_BPB_FACTOR)
+        rate_correction_factor = MAX_BPB_FACTOR;
+      cpi->rc.rate_correction_factors[INTER_NORMAL] = rate_correction_factor;
+    }
+    // For temporal layers, reset the rate control parametes across all
+    // temporal layers.
+    if (cpi->use_svc) {
+      int i = 0;
+      SVC *svc = &cpi->svc;
+      for (i = 0; i < svc->number_temporal_layers; ++i) {
+        const int layer = LAYER_IDS_TO_IDX(svc->spatial_layer_id, i,
+                                           svc->number_temporal_layers);
+        LAYER_CONTEXT *lc = &svc->layer_context[layer];
+        RATE_CONTROL *lrc = &lc->rc;
+        lrc->avg_frame_qindex[INTER_FRAME] = *q;
+        lrc->buffer_level = rc->optimal_buffer_level;
+        lrc->bits_off_target = rc->optimal_buffer_level;
+        lrc->rc_1_frame = 0;
+        lrc->rc_2_frame = 0;
+        lrc->rate_correction_factors[INTER_NORMAL] =
+            rate_correction_factor;
+      }
+    }
+    return 1;
+  } else {
+    return 0;
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_ratectrl.h b/libvpx/libvpx/vp9/encoder/vp9_ratectrl.h
new file mode 100644
index 0000000..7024bcf
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_ratectrl.h
@@ -0,0 +1,288 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_RATECTRL_H_
+#define VP9_ENCODER_VP9_RATECTRL_H_
+
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Bits Per MB at different Q (Multiplied by 512)
+#define BPER_MB_NORMBITS    9
+
+#define MIN_GF_INTERVAL     4
+#define MAX_GF_INTERVAL     16
+#define FIXED_GF_INTERVAL   8    // Used in some testing modes only
+#define ONEHALFONLY_RESIZE  0
+
+typedef enum {
+  INTER_NORMAL = 0,
+  INTER_HIGH = 1,
+  GF_ARF_LOW = 2,
+  GF_ARF_STD = 3,
+  KF_STD = 4,
+  RATE_FACTOR_LEVELS = 5
+} RATE_FACTOR_LEVEL;
+
+// Internal frame scaling level.
+typedef enum {
+  UNSCALED = 0,     // Frame is unscaled.
+  SCALE_STEP1 = 1,  // First-level down-scaling.
+  FRAME_SCALE_STEPS
+} FRAME_SCALE_LEVEL;
+
+typedef enum {
+  NO_RESIZE = 0,
+  DOWN_THREEFOUR = 1,  // From orig to 3/4.
+  DOWN_ONEHALF = 2,    // From orig or 3/4 to 1/2.
+  UP_THREEFOUR = -1,   // From 1/2 to 3/4.
+  UP_ORIG = -2,        // From 1/2 or 3/4 to orig.
+} RESIZE_ACTION;
+
+typedef enum {
+  ORIG = 0,
+  THREE_QUARTER = 1,
+  ONE_HALF = 2
+} RESIZE_STATE;
+
+// Frame dimensions multiplier wrt the native frame size, in 1/16ths,
+// specified for the scale-up case.
+// e.g. 24 => 16/24 = 2/3 of native size. The restriction to 1/16th is
+// intended to match the capabilities of the normative scaling filters,
+// giving precedence to the up-scaling accuracy.
+static const int frame_scale_factor[FRAME_SCALE_STEPS] = {16, 24};
+
+// Multiplier of the target rate to be used as threshold for triggering scaling.
+static const double rate_thresh_mult[FRAME_SCALE_STEPS] = {1.0, 2.0};
+
+// Scale dependent Rate Correction Factor multipliers. Compensates for the
+// greater number of bits per pixel generated in down-scaled frames.
+static const double rcf_mult[FRAME_SCALE_STEPS] = {1.0, 2.0};
+
+typedef struct {
+  // Rate targetting variables
+  int base_frame_target;           // A baseline frame target before adjustment
+                                   // for previous under or over shoot.
+  int this_frame_target;           // Actual frame target after rc adjustment.
+  int projected_frame_size;
+  int sb64_target_rate;
+  int last_q[FRAME_TYPES];         // Separate values for Intra/Inter
+  int last_boosted_qindex;         // Last boosted GF/KF/ARF q
+  int last_kf_qindex;              // Q index of the last key frame coded.
+
+  int gfu_boost;
+  int last_boost;
+  int kf_boost;
+
+  double rate_correction_factors[RATE_FACTOR_LEVELS];
+
+  int frames_since_golden;
+  int frames_till_gf_update_due;
+  int min_gf_interval;
+  int max_gf_interval;
+  int static_scene_max_gf_interval;
+  int baseline_gf_interval;
+  int constrained_gf_group;
+  int frames_to_key;
+  int frames_since_key;
+  int this_key_frame_forced;
+  int next_key_frame_forced;
+  int source_alt_ref_pending;
+  int source_alt_ref_active;
+  int is_src_frame_alt_ref;
+
+  int avg_frame_bandwidth;  // Average frame size target for clip
+  int min_frame_bandwidth;  // Minimum allocation used for any frame
+  int max_frame_bandwidth;  // Maximum burst rate allowed for a frame.
+
+  int ni_av_qi;
+  int ni_tot_qi;
+  int ni_frames;
+  int avg_frame_qindex[FRAME_TYPES];
+  double tot_q;
+  double avg_q;
+
+  int64_t buffer_level;
+  int64_t bits_off_target;
+  int64_t vbr_bits_off_target;
+  int64_t vbr_bits_off_target_fast;
+
+  int decimation_factor;
+  int decimation_count;
+
+  int rolling_target_bits;
+  int rolling_actual_bits;
+
+  int long_rolling_target_bits;
+  int long_rolling_actual_bits;
+
+  int rate_error_estimate;
+
+  int64_t total_actual_bits;
+  int64_t total_target_bits;
+  int64_t total_target_vs_actual;
+
+  int worst_quality;
+  int best_quality;
+
+  int64_t starting_buffer_level;
+  int64_t optimal_buffer_level;
+  int64_t maximum_buffer_size;
+
+  // rate control history for last frame(1) and the frame before(2).
+  // -1: undershot
+  //  1: overshoot
+  //  0: not initialized.
+  int rc_1_frame;
+  int rc_2_frame;
+  int q_1_frame;
+  int q_2_frame;
+
+  // Auto frame-scaling variables.
+  FRAME_SCALE_LEVEL frame_size_selector;
+  FRAME_SCALE_LEVEL next_frame_size_selector;
+  int frame_width[FRAME_SCALE_STEPS];
+  int frame_height[FRAME_SCALE_STEPS];
+  int rf_level_maxq[RATE_FACTOR_LEVELS];
+
+  uint64_t avg_source_sad;
+  int high_source_sad;
+  int count_last_scene_change;
+  int avg_frame_low_motion;
+} RATE_CONTROL;
+
+struct VP9_COMP;
+struct VP9EncoderConfig;
+
+void vp9_rc_init(const struct VP9EncoderConfig *oxcf, int pass,
+                 RATE_CONTROL *rc);
+
+int vp9_estimate_bits_at_q(FRAME_TYPE frame_kind, int q, int mbs,
+                           double correction_factor,
+                           vpx_bit_depth_t bit_depth);
+
+double vp9_convert_qindex_to_q(int qindex, vpx_bit_depth_t bit_depth);
+
+void vp9_rc_init_minq_luts(void);
+
+int vp9_rc_get_default_min_gf_interval(int width, int height, double framerate);
+// Note vp9_rc_get_default_max_gf_interval() requires the min_gf_interval to
+// be passed in to ensure that the max_gf_interval returned is at least as bis
+// as that.
+int vp9_rc_get_default_max_gf_interval(double framerate, int min_frame_rate);
+
+// Generally at the high level, the following flow is expected
+// to be enforced for rate control:
+// First call per frame, one of:
+//   vp9_rc_get_one_pass_vbr_params()
+//   vp9_rc_get_one_pass_cbr_params()
+//   vp9_rc_get_svc_params()
+//   vp9_rc_get_first_pass_params()
+//   vp9_rc_get_second_pass_params()
+// depending on the usage to set the rate control encode parameters desired.
+//
+// Then, call encode_frame_to_data_rate() to perform the
+// actual encode. This function will in turn call encode_frame()
+// one or more times, followed by one of:
+//   vp9_rc_postencode_update()
+//   vp9_rc_postencode_update_drop_frame()
+//
+// The majority of rate control parameters are only expected
+// to be set in the vp9_rc_get_..._params() functions and
+// updated during the vp9_rc_postencode_update...() functions.
+// The only exceptions are vp9_rc_drop_frame() and
+// vp9_rc_update_rate_correction_factors() functions.
+
+// Functions to set parameters for encoding before the actual
+// encode_frame_to_data_rate() function.
+void vp9_rc_get_one_pass_vbr_params(struct VP9_COMP *cpi);
+void vp9_rc_get_one_pass_cbr_params(struct VP9_COMP *cpi);
+void vp9_rc_get_svc_params(struct VP9_COMP *cpi);
+
+// Post encode update of the rate control parameters based
+// on bytes used
+void vp9_rc_postencode_update(struct VP9_COMP *cpi, uint64_t bytes_used);
+// Post encode update of the rate control parameters for dropped frames
+void vp9_rc_postencode_update_drop_frame(struct VP9_COMP *cpi);
+
+// Updates rate correction factors
+// Changes only the rate correction factors in the rate control structure.
+void vp9_rc_update_rate_correction_factors(struct VP9_COMP *cpi);
+
+// Decide if we should drop this frame: For 1-pass CBR.
+// Changes only the decimation count in the rate control structure
+int vp9_rc_drop_frame(struct VP9_COMP *cpi);
+
+// Computes frame size bounds.
+void vp9_rc_compute_frame_size_bounds(const struct VP9_COMP *cpi,
+                                      int this_frame_target,
+                                      int *frame_under_shoot_limit,
+                                      int *frame_over_shoot_limit);
+
+// Picks q and q bounds given the target for bits
+int vp9_rc_pick_q_and_bounds(const struct VP9_COMP *cpi,
+                             int *bottom_index,
+                             int *top_index);
+
+// Estimates q to achieve a target bits per frame
+int vp9_rc_regulate_q(const struct VP9_COMP *cpi, int target_bits_per_frame,
+                      int active_best_quality, int active_worst_quality);
+
+// Estimates bits per mb for a given qindex and correction factor.
+int vp9_rc_bits_per_mb(FRAME_TYPE frame_type, int qindex,
+                       double correction_factor, vpx_bit_depth_t bit_depth);
+
+// Clamping utilities for bitrate targets for iframes and pframes.
+int vp9_rc_clamp_iframe_target_size(const struct VP9_COMP *const cpi,
+                                    int target);
+int vp9_rc_clamp_pframe_target_size(const struct VP9_COMP *const cpi,
+                                    int target);
+// Utility to set frame_target into the RATE_CONTROL structure
+// This function is called only from the vp9_rc_get_..._params() functions.
+void vp9_rc_set_frame_target(struct VP9_COMP *cpi, int target);
+
+// Computes a q delta (in "q index" terms) to get from a starting q value
+// to a target q value
+int vp9_compute_qdelta(const RATE_CONTROL *rc, double qstart, double qtarget,
+                       vpx_bit_depth_t bit_depth);
+
+// Computes a q delta (in "q index" terms) to get from a starting q value
+// to a value that should equate to the given rate ratio.
+int vp9_compute_qdelta_by_rate(const RATE_CONTROL *rc, FRAME_TYPE frame_type,
+                               int qindex, double rate_target_ratio,
+                               vpx_bit_depth_t bit_depth);
+
+int vp9_frame_type_qdelta(const struct VP9_COMP *cpi, int rf_level, int q);
+
+void vp9_rc_update_framerate(struct VP9_COMP *cpi);
+
+void vp9_rc_set_gf_interval_range(const struct VP9_COMP *const cpi,
+                                  RATE_CONTROL *const rc);
+
+void vp9_set_target_rate(struct VP9_COMP *cpi);
+
+int vp9_resize_one_pass_cbr(struct VP9_COMP *cpi);
+
+void vp9_avg_source_sad(struct VP9_COMP *cpi);
+
+int vp9_encodedframe_overshoot(struct VP9_COMP *cpi, int frame_size, int *q);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_RATECTRL_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_rd.c b/libvpx/libvpx/vp9/encoder/vp9_rd.c
new file mode 100644
index 0000000..91b2911
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_rd.c
@@ -0,0 +1,676 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./vp9_rtcd.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/bitops.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/system_state.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_encodemb.h"
+#include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_tokenize.h"
+
+#define RD_THRESH_POW      1.25
+
+// Factor to weigh the rate for switchable interp filters.
+#define SWITCHABLE_INTERP_RATE_FACTOR 1
+
+void vp9_rd_cost_reset(RD_COST *rd_cost) {
+  rd_cost->rate = INT_MAX;
+  rd_cost->dist = INT64_MAX;
+  rd_cost->rdcost = INT64_MAX;
+}
+
+void vp9_rd_cost_init(RD_COST *rd_cost) {
+  rd_cost->rate = 0;
+  rd_cost->dist = 0;
+  rd_cost->rdcost = 0;
+}
+
+// The baseline rd thresholds for breaking out of the rd loop for
+// certain modes are assumed to be based on 8x8 blocks.
+// This table is used to correct for block size.
+// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
+static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = {
+  2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32
+};
+
+static void fill_mode_costs(VP9_COMP *cpi) {
+  const FRAME_CONTEXT *const fc = cpi->common.fc;
+  int i, j;
+
+  for (i = 0; i < INTRA_MODES; ++i)
+    for (j = 0; j < INTRA_MODES; ++j)
+      vp9_cost_tokens(cpi->y_mode_costs[i][j], vp9_kf_y_mode_prob[i][j],
+                      vp9_intra_mode_tree);
+
+  vp9_cost_tokens(cpi->mbmode_cost, fc->y_mode_prob[1], vp9_intra_mode_tree);
+  for (i = 0; i < INTRA_MODES; ++i) {
+    vp9_cost_tokens(cpi->intra_uv_mode_cost[KEY_FRAME][i],
+                    vp9_kf_uv_mode_prob[i], vp9_intra_mode_tree);
+    vp9_cost_tokens(cpi->intra_uv_mode_cost[INTER_FRAME][i],
+                    fc->uv_mode_prob[i], vp9_intra_mode_tree);
+  }
+
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+    vp9_cost_tokens(cpi->switchable_interp_costs[i],
+                    fc->switchable_interp_prob[i], vp9_switchable_interp_tree);
+}
+
+static void fill_token_costs(vp9_coeff_cost *c,
+                             vp9_coeff_probs_model (*p)[PLANE_TYPES]) {
+  int i, j, k, l;
+  TX_SIZE t;
+  for (t = TX_4X4; t <= TX_32X32; ++t)
+    for (i = 0; i < PLANE_TYPES; ++i)
+      for (j = 0; j < REF_TYPES; ++j)
+        for (k = 0; k < COEF_BANDS; ++k)
+          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+            vpx_prob probs[ENTROPY_NODES];
+            vp9_model_to_full_probs(p[t][i][j][k][l], probs);
+            vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs,
+                            vp9_coef_tree);
+            vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
+                                 vp9_coef_tree);
+            assert(c[t][i][j][k][0][l][EOB_TOKEN] ==
+                   c[t][i][j][k][1][l][EOB_TOKEN]);
+          }
+}
+
+// Values are now correlated to quantizer.
+static int sad_per_bit16lut_8[QINDEX_RANGE];
+static int sad_per_bit4lut_8[QINDEX_RANGE];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static int sad_per_bit16lut_10[QINDEX_RANGE];
+static int sad_per_bit4lut_10[QINDEX_RANGE];
+static int sad_per_bit16lut_12[QINDEX_RANGE];
+static int sad_per_bit4lut_12[QINDEX_RANGE];
+#endif
+
+static void init_me_luts_bd(int *bit16lut, int *bit4lut, int range,
+                            vpx_bit_depth_t bit_depth) {
+  int i;
+  // Initialize the sad lut tables using a formulaic calculation for now.
+  // This is to make it easier to resolve the impact of experimental changes
+  // to the quantizer tables.
+  for (i = 0; i < range; i++) {
+    const double q = vp9_convert_qindex_to_q(i, bit_depth);
+    bit16lut[i] = (int)(0.0418 * q + 2.4107);
+    bit4lut[i] = (int)(0.063 * q + 2.742);
+  }
+}
+
+void vp9_init_me_luts(void) {
+  init_me_luts_bd(sad_per_bit16lut_8, sad_per_bit4lut_8, QINDEX_RANGE,
+                  VPX_BITS_8);
+#if CONFIG_VP9_HIGHBITDEPTH
+  init_me_luts_bd(sad_per_bit16lut_10, sad_per_bit4lut_10, QINDEX_RANGE,
+                  VPX_BITS_10);
+  init_me_luts_bd(sad_per_bit16lut_12, sad_per_bit4lut_12, QINDEX_RANGE,
+                  VPX_BITS_12);
+#endif
+}
+
+static const int rd_boost_factor[16] = {
+  64, 32, 32, 32, 24, 16, 12, 12,
+  8, 8, 4, 4, 2, 2, 1, 0
+};
+static const int rd_frame_type_factor[FRAME_UPDATE_TYPES] = {
+  128, 144, 128, 128, 144
+};
+
+int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex) {
+  const int64_t q = vp9_dc_quant(qindex, 0, cpi->common.bit_depth);
+#if CONFIG_VP9_HIGHBITDEPTH
+  int64_t rdmult = 0;
+  switch (cpi->common.bit_depth) {
+    case VPX_BITS_8:
+      rdmult = 88 * q * q / 24;
+      break;
+    case VPX_BITS_10:
+      rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 4);
+      break;
+    case VPX_BITS_12:
+      rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 8);
+      break;
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  int64_t rdmult = 88 * q * q / 24;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  if (cpi->oxcf.pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    const FRAME_UPDATE_TYPE frame_type = gf_group->update_type[gf_group->index];
+    const int boost_index = VPXMIN(15, (cpi->rc.gfu_boost / 100));
+
+    rdmult = (rdmult * rd_frame_type_factor[frame_type]) >> 7;
+    rdmult += ((rdmult * rd_boost_factor[boost_index]) >> 7);
+  }
+  if (rdmult < 1)
+    rdmult = 1;
+  return (int)rdmult;
+}
+
+static int compute_rd_thresh_factor(int qindex, vpx_bit_depth_t bit_depth) {
+  double q;
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0;
+      break;
+    case VPX_BITS_10:
+      q = vp9_dc_quant(qindex, 0, VPX_BITS_10) / 16.0;
+      break;
+    case VPX_BITS_12:
+      q = vp9_dc_quant(qindex, 0, VPX_BITS_12) / 64.0;
+      break;
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  (void) bit_depth;
+  q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  // TODO(debargha): Adjust the function below.
+  return VPXMAX((int)(pow(q, RD_THRESH_POW) * 5.12), 8);
+}
+
+void vp9_initialize_me_consts(VP9_COMP *cpi, MACROBLOCK *x, int qindex) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (cpi->common.bit_depth) {
+    case VPX_BITS_8:
+      x->sadperbit16 = sad_per_bit16lut_8[qindex];
+      x->sadperbit4 = sad_per_bit4lut_8[qindex];
+      break;
+    case VPX_BITS_10:
+      x->sadperbit16 = sad_per_bit16lut_10[qindex];
+      x->sadperbit4 = sad_per_bit4lut_10[qindex];
+      break;
+    case VPX_BITS_12:
+      x->sadperbit16 = sad_per_bit16lut_12[qindex];
+      x->sadperbit4 = sad_per_bit4lut_12[qindex];
+      break;
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+  }
+#else
+  (void)cpi;
+  x->sadperbit16 = sad_per_bit16lut_8[qindex];
+  x->sadperbit4 = sad_per_bit4lut_8[qindex];
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+static void set_block_thresholds(const VP9_COMMON *cm, RD_OPT *rd) {
+  int i, bsize, segment_id;
+
+  for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) {
+    const int qindex =
+        clamp(vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex) +
+              cm->y_dc_delta_q, 0, MAXQ);
+    const int q = compute_rd_thresh_factor(qindex, cm->bit_depth);
+
+    for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
+      // Threshold here seems unnecessarily harsh but fine given actual
+      // range of values used for cpi->sf.thresh_mult[].
+      const int t = q * rd_thresh_block_size_factor[bsize];
+      const int thresh_max = INT_MAX / t;
+
+      if (bsize >= BLOCK_8X8) {
+        for (i = 0; i < MAX_MODES; ++i)
+          rd->threshes[segment_id][bsize][i] =
+              rd->thresh_mult[i] < thresh_max
+                  ? rd->thresh_mult[i] * t / 4
+                  : INT_MAX;
+      } else {
+        for (i = 0; i < MAX_REFS; ++i)
+          rd->threshes[segment_id][bsize][i] =
+              rd->thresh_mult_sub8x8[i] < thresh_max
+                  ? rd->thresh_mult_sub8x8[i] * t / 4
+                  : INT_MAX;
+      }
+    }
+  }
+}
+
+void vp9_initialize_rd_consts(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+  RD_OPT *const rd = &cpi->rd;
+  int i;
+
+  vpx_clear_system_state();
+
+  rd->RDDIV = RDDIV_BITS;  // In bits (to multiply D by 128).
+  rd->RDMULT = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);
+
+  set_error_per_bit(x, rd->RDMULT);
+
+  x->select_tx_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
+                       cm->frame_type != KEY_FRAME) ? 0 : 1;
+
+  set_block_thresholds(cm, rd);
+  set_partition_probs(cm, xd);
+
+  if (cpi->oxcf.pass == 1) {
+    if (!frame_is_intra_only(cm))
+      vp9_build_nmv_cost_table(
+          x->nmvjointcost,
+          cm->allow_high_precision_mv ? x->nmvcost_hp : x->nmvcost,
+          &cm->fc->nmvc, cm->allow_high_precision_mv);
+  } else {
+    if (!cpi->sf.use_nonrd_pick_mode || cm->frame_type == KEY_FRAME)
+      fill_token_costs(x->token_costs, cm->fc->coef_probs);
+
+    if (cpi->sf.partition_search_type != VAR_BASED_PARTITION ||
+        cm->frame_type == KEY_FRAME) {
+      for (i = 0; i < PARTITION_CONTEXTS; ++i)
+        vp9_cost_tokens(cpi->partition_cost[i], get_partition_probs(xd, i),
+                        vp9_partition_tree);
+    }
+
+    if (!cpi->sf.use_nonrd_pick_mode || (cm->current_video_frame & 0x07) == 1 ||
+        cm->frame_type == KEY_FRAME) {
+      fill_mode_costs(cpi);
+
+      if (!frame_is_intra_only(cm)) {
+        vp9_build_nmv_cost_table(
+            x->nmvjointcost,
+            cm->allow_high_precision_mv ? x->nmvcost_hp : x->nmvcost,
+            &cm->fc->nmvc, cm->allow_high_precision_mv);
+
+        for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
+          vp9_cost_tokens((int *)cpi->inter_mode_cost[i],
+                          cm->fc->inter_mode_probs[i], vp9_inter_mode_tree);
+      }
+    }
+  }
+}
+
+static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {
+  // NOTE: The tables below must be of the same size.
+
+  // The functions described below are sampled at the four most significant
+  // bits of x^2 + 8 / 256.
+
+  // Normalized rate:
+  // This table models the rate for a Laplacian source with given variance
+  // when quantized with a uniform quantizer with given stepsize. The
+  // closed form expression is:
+  // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
+  // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
+  // and H(x) is the binary entropy function.
+  static const int rate_tab_q10[] = {
+    65536,  6086,  5574,  5275,  5063,  4899,  4764,  4651,
+     4553,  4389,  4255,  4142,  4044,  3958,  3881,  3811,
+     3748,  3635,  3538,  3453,  3376,  3307,  3244,  3186,
+     3133,  3037,  2952,  2877,  2809,  2747,  2690,  2638,
+     2589,  2501,  2423,  2353,  2290,  2232,  2179,  2130,
+     2084,  2001,  1928,  1862,  1802,  1748,  1698,  1651,
+     1608,  1530,  1460,  1398,  1342,  1290,  1243,  1199,
+     1159,  1086,  1021,   963,   911,   864,   821,   781,
+      745,   680,   623,   574,   530,   490,   455,   424,
+      395,   345,   304,   269,   239,   213,   190,   171,
+      154,   126,   104,    87,    73,    61,    52,    44,
+       38,    28,    21,    16,    12,    10,     8,     6,
+        5,     3,     2,     1,     1,     1,     0,     0,
+  };
+
+  // Normalized distortion:
+  // This table models the normalized distortion for a Laplacian source
+  // with given variance when quantized with a uniform quantizer
+  // with given stepsize. The closed form expression is:
+  // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
+  // where x = qpstep / sqrt(variance).
+  // Note the actual distortion is Dn * variance.
+  static const int dist_tab_q10[] = {
+       0,     0,     1,     1,     1,     2,     2,     2,
+       3,     3,     4,     5,     5,     6,     7,     7,
+       8,     9,    11,    12,    13,    15,    16,    17,
+      18,    21,    24,    26,    29,    31,    34,    36,
+      39,    44,    49,    54,    59,    64,    69,    73,
+      78,    88,    97,   106,   115,   124,   133,   142,
+     151,   167,   184,   200,   215,   231,   245,   260,
+     274,   301,   327,   351,   375,   397,   418,   439,
+     458,   495,   528,   559,   587,   613,   637,   659,
+     680,   717,   749,   777,   801,   823,   842,   859,
+     874,   899,   919,   936,   949,   960,   969,   977,
+     983,   994,  1001,  1006,  1010,  1013,  1015,  1017,
+    1018,  1020,  1022,  1022,  1023,  1023,  1023,  1024,
+  };
+  static const int xsq_iq_q10[] = {
+         0,      4,      8,     12,     16,     20,     24,     28,
+        32,     40,     48,     56,     64,     72,     80,     88,
+        96,    112,    128,    144,    160,    176,    192,    208,
+       224,    256,    288,    320,    352,    384,    416,    448,
+       480,    544,    608,    672,    736,    800,    864,    928,
+       992,   1120,   1248,   1376,   1504,   1632,   1760,   1888,
+      2016,   2272,   2528,   2784,   3040,   3296,   3552,   3808,
+      4064,   4576,   5088,   5600,   6112,   6624,   7136,   7648,
+      8160,   9184,  10208,  11232,  12256,  13280,  14304,  15328,
+     16352,  18400,  20448,  22496,  24544,  26592,  28640,  30688,
+     32736,  36832,  40928,  45024,  49120,  53216,  57312,  61408,
+     65504,  73696,  81888,  90080,  98272, 106464, 114656, 122848,
+    131040, 147424, 163808, 180192, 196576, 212960, 229344, 245728,
+  };
+  const int tmp = (xsq_q10 >> 2) + 8;
+  const int k = get_msb(tmp) - 3;
+  const int xq = (k << 3) + ((tmp >> k) & 0x7);
+  const int one_q10 = 1 << 10;
+  const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k);
+  const int b_q10 = one_q10 - a_q10;
+  *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10;
+  *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
+}
+
+void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n_log2,
+                                  unsigned int qstep, int *rate,
+                                  int64_t *dist) {
+  // This function models the rate and distortion for a Laplacian
+  // source with given variance when quantized with a uniform quantizer
+  // with given stepsize. The closed form expressions are in:
+  // Hang and Chen, "Source Model for transform video coder and its
+  // application - Part I: Fundamental Theory", IEEE Trans. Circ.
+  // Sys. for Video Tech., April 1997.
+  if (var == 0) {
+    *rate = 0;
+    *dist = 0;
+  } else {
+    int d_q10, r_q10;
+    static const uint32_t MAX_XSQ_Q10 = 245727;
+    const uint64_t xsq_q10_64 =
+        (((uint64_t)qstep * qstep << (n_log2 + 10)) + (var >> 1)) / var;
+    const int xsq_q10 = (int)VPXMIN(xsq_q10_64, MAX_XSQ_Q10);
+    model_rd_norm(xsq_q10, &r_q10, &d_q10);
+    *rate = ROUND_POWER_OF_TWO(r_q10 << n_log2, 10 - VP9_PROB_COST_SHIFT);
+    *dist = (var * (int64_t)d_q10 + 512) >> 10;
+  }
+}
+
+void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              const struct macroblockd_plane *pd,
+                              ENTROPY_CONTEXT t_above[16],
+                              ENTROPY_CONTEXT t_left[16]) {
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+  const ENTROPY_CONTEXT *const above = pd->above_context;
+  const ENTROPY_CONTEXT *const left = pd->left_context;
+
+  int i;
+  switch (tx_size) {
+    case TX_4X4:
+      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
+      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
+      break;
+    case TX_8X8:
+      for (i = 0; i < num_4x4_w; i += 2)
+        t_above[i] = !!*(const uint16_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 2)
+        t_left[i] = !!*(const uint16_t *)&left[i];
+      break;
+    case TX_16X16:
+      for (i = 0; i < num_4x4_w; i += 4)
+        t_above[i] = !!*(const uint32_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 4)
+        t_left[i] = !!*(const uint32_t *)&left[i];
+      break;
+    case TX_32X32:
+      for (i = 0; i < num_4x4_w; i += 8)
+        t_above[i] = !!*(const uint64_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 8)
+        t_left[i] = !!*(const uint64_t *)&left[i];
+      break;
+    default:
+      assert(0 && "Invalid transform size.");
+      break;
+  }
+}
+
+void vp9_mv_pred(VP9_COMP *cpi, MACROBLOCK *x,
+                 uint8_t *ref_y_buffer, int ref_y_stride,
+                 int ref_frame, BLOCK_SIZE block_size) {
+  int i;
+  int zero_seen = 0;
+  int best_index = 0;
+  int best_sad = INT_MAX;
+  int this_sad = INT_MAX;
+  int max_mv = 0;
+  int near_same_nearest;
+  uint8_t *src_y_ptr = x->plane[0].src.buf;
+  uint8_t *ref_y_ptr;
+  const int num_mv_refs = MAX_MV_REF_CANDIDATES +
+                    (cpi->sf.adaptive_motion_search &&
+                     block_size < x->max_partition_size);
+
+  MV pred_mv[3];
+  pred_mv[0] = x->mbmi_ext->ref_mvs[ref_frame][0].as_mv;
+  pred_mv[1] = x->mbmi_ext->ref_mvs[ref_frame][1].as_mv;
+  pred_mv[2] = x->pred_mv[ref_frame];
+  assert(num_mv_refs <= (int)(sizeof(pred_mv) / sizeof(pred_mv[0])));
+
+  near_same_nearest =
+      x->mbmi_ext->ref_mvs[ref_frame][0].as_int ==
+          x->mbmi_ext->ref_mvs[ref_frame][1].as_int;
+  // Get the sad for each candidate reference mv.
+  for (i = 0; i < num_mv_refs; ++i) {
+    const MV *this_mv = &pred_mv[i];
+    int fp_row, fp_col;
+
+    if (i == 1 && near_same_nearest)
+      continue;
+    fp_row = (this_mv->row + 3 + (this_mv->row >= 0)) >> 3;
+    fp_col = (this_mv->col + 3 + (this_mv->col >= 0)) >> 3;
+    max_mv = VPXMAX(max_mv, VPXMAX(abs(this_mv->row), abs(this_mv->col)) >> 3);
+
+    if (fp_row ==0 && fp_col == 0 && zero_seen)
+      continue;
+    zero_seen |= (fp_row ==0 && fp_col == 0);
+
+    ref_y_ptr =&ref_y_buffer[ref_y_stride * fp_row + fp_col];
+    // Find sad for current vector.
+    this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
+                                           ref_y_ptr, ref_y_stride);
+    // Note if it is the best so far.
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      best_index = i;
+    }
+  }
+
+  // Note the index of the mv that worked best in the reference list.
+  x->mv_best_ref_index[ref_frame] = best_index;
+  x->max_mv_context[ref_frame] = max_mv;
+  x->pred_mv_sad[ref_frame] = best_sad;
+}
+
+void vp9_setup_pred_block(const MACROBLOCKD *xd,
+                          struct buf_2d dst[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col,
+                          const struct scale_factors *scale,
+                          const struct scale_factors *scale_uv) {
+  int i;
+
+  dst[0].buf = src->y_buffer;
+  dst[0].stride = src->y_stride;
+  dst[1].buf = src->u_buffer;
+  dst[2].buf = src->v_buffer;
+  dst[1].stride = dst[2].stride = src->uv_stride;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    setup_pred_plane(dst + i, dst[i].buf, dst[i].stride, mi_row, mi_col,
+                     i ? scale_uv : scale,
+                     xd->plane[i].subsampling_x, xd->plane[i].subsampling_y);
+  }
+}
+
+int vp9_raster_block_offset(BLOCK_SIZE plane_bsize,
+                            int raster_block, int stride) {
+  const int bw = b_width_log2_lookup[plane_bsize];
+  const int y = 4 * (raster_block >> bw);
+  const int x = 4 * (raster_block & ((1 << bw) - 1));
+  return y * stride + x;
+}
+
+int16_t* vp9_raster_block_offset_int16(BLOCK_SIZE plane_bsize,
+                                       int raster_block, int16_t *base) {
+  const int stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
+  return base + vp9_raster_block_offset(plane_bsize, raster_block, stride);
+}
+
+YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi,
+                                             int ref_frame) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1];
+  const int ref_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+  return
+      (scaled_idx != ref_idx && scaled_idx != INVALID_IDX) ?
+          &cm->buffer_pool->frame_bufs[scaled_idx].buf : NULL;
+}
+
+int vp9_get_switchable_rate(const VP9_COMP *cpi, const MACROBLOCKD *const xd) {
+  const MODE_INFO *const mi = xd->mi[0];
+  const int ctx = vp9_get_pred_context_switchable_interp(xd);
+  return SWITCHABLE_INTERP_RATE_FACTOR *
+             cpi->switchable_interp_costs[ctx][mi->interp_filter];
+}
+
+void vp9_set_rd_speed_thresholds(VP9_COMP *cpi) {
+  int i;
+  RD_OPT *const rd = &cpi->rd;
+  SPEED_FEATURES *const sf = &cpi->sf;
+
+  // Set baseline threshold values.
+  for (i = 0; i < MAX_MODES; ++i)
+    rd->thresh_mult[i] = cpi->oxcf.mode == BEST ? -500 : 0;
+
+  if (sf->adaptive_rd_thresh) {
+    rd->thresh_mult[THR_NEARESTMV] = 300;
+    rd->thresh_mult[THR_NEARESTG] = 300;
+    rd->thresh_mult[THR_NEARESTA] = 300;
+  } else {
+    rd->thresh_mult[THR_NEARESTMV] = 0;
+    rd->thresh_mult[THR_NEARESTG] = 0;
+    rd->thresh_mult[THR_NEARESTA] = 0;
+  }
+
+  rd->thresh_mult[THR_DC] += 1000;
+
+  rd->thresh_mult[THR_NEWMV] += 1000;
+  rd->thresh_mult[THR_NEWA] += 1000;
+  rd->thresh_mult[THR_NEWG] += 1000;
+
+  rd->thresh_mult[THR_NEARMV] += 1000;
+  rd->thresh_mult[THR_NEARA] += 1000;
+  rd->thresh_mult[THR_COMP_NEARESTLA] += 1000;
+  rd->thresh_mult[THR_COMP_NEARESTGA] += 1000;
+
+  rd->thresh_mult[THR_TM] += 1000;
+
+  rd->thresh_mult[THR_COMP_NEARLA] += 1500;
+  rd->thresh_mult[THR_COMP_NEWLA] += 2000;
+  rd->thresh_mult[THR_NEARG] += 1000;
+  rd->thresh_mult[THR_COMP_NEARGA] += 1500;
+  rd->thresh_mult[THR_COMP_NEWGA] += 2000;
+
+  rd->thresh_mult[THR_ZEROMV] += 2000;
+  rd->thresh_mult[THR_ZEROG] += 2000;
+  rd->thresh_mult[THR_ZEROA] += 2000;
+  rd->thresh_mult[THR_COMP_ZEROLA] += 2500;
+  rd->thresh_mult[THR_COMP_ZEROGA] += 2500;
+
+  rd->thresh_mult[THR_H_PRED] += 2000;
+  rd->thresh_mult[THR_V_PRED] += 2000;
+  rd->thresh_mult[THR_D45_PRED ] += 2500;
+  rd->thresh_mult[THR_D135_PRED] += 2500;
+  rd->thresh_mult[THR_D117_PRED] += 2500;
+  rd->thresh_mult[THR_D153_PRED] += 2500;
+  rd->thresh_mult[THR_D207_PRED] += 2500;
+  rd->thresh_mult[THR_D63_PRED] += 2500;
+}
+
+void vp9_set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi) {
+  static const int thresh_mult[2][MAX_REFS] =
+      {{2500, 2500, 2500, 4500, 4500, 2500},
+       {2000, 2000, 2000, 4000, 4000, 2000}};
+  RD_OPT *const rd = &cpi->rd;
+  const int idx = cpi->oxcf.mode == BEST;
+  memcpy(rd->thresh_mult_sub8x8, thresh_mult[idx], sizeof(thresh_mult[idx]));
+}
+
+void vp9_update_rd_thresh_fact(int (*factor_buf)[MAX_MODES], int rd_thresh,
+                               int bsize, int best_mode_index) {
+  if (rd_thresh > 0) {
+    const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES;
+    int mode;
+    for (mode = 0; mode < top_mode; ++mode) {
+      const BLOCK_SIZE min_size = VPXMAX(bsize - 1, BLOCK_4X4);
+      const BLOCK_SIZE max_size = VPXMIN(bsize + 2, BLOCK_64X64);
+      BLOCK_SIZE bs;
+      for (bs = min_size; bs <= max_size; ++bs) {
+        int *const fact = &factor_buf[bs][mode];
+        if (mode == best_mode_index) {
+          *fact -= (*fact >> 4);
+        } else {
+          *fact = VPXMIN(*fact + RD_THRESH_INC, rd_thresh * RD_THRESH_MAX_FACT);
+        }
+      }
+    }
+  }
+}
+
+int vp9_get_intra_cost_penalty(int qindex, int qdelta,
+                               vpx_bit_depth_t bit_depth) {
+  const int q = vp9_dc_quant(qindex, qdelta, bit_depth);
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      return 20 * q;
+    case VPX_BITS_10:
+      return 5 * q;
+    case VPX_BITS_12:
+      return ROUND_POWER_OF_TWO(5 * q, 2);
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  return 20 * q;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
diff --git a/libvpx/libvpx/vp9/encoder/vp9_rd.h b/libvpx/libvpx/vp9/encoder/vp9_rd.h
new file mode 100644
index 0000000..9b8e273
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_rd.h
@@ -0,0 +1,204 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_RD_H_
+#define VP9_ENCODER_VP9_RD_H_
+
+#include <limits.h>
+
+#include "vp9/common/vp9_blockd.h"
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_cost.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define RDDIV_BITS          7
+#define RD_EPB_SHIFT        6
+
+#define RDCOST(RM, DM, R, D) \
+  (ROUND_POWER_OF_TWO(((int64_t)R) * (RM), VP9_PROB_COST_SHIFT) + (D << DM))
+#define QIDX_SKIP_THRESH     115
+
+#define MV_COST_WEIGHT      108
+#define MV_COST_WEIGHT_SUB  120
+
+#define INVALID_MV 0x80008000
+
+#define MAX_MODES 30
+#define MAX_REFS  6
+
+#define RD_THRESH_MAX_FACT 64
+#define RD_THRESH_INC      1
+
+// This enumerator type needs to be kept aligned with the mode order in
+// const MODE_DEFINITION vp9_mode_order[MAX_MODES] used in the rd code.
+typedef enum {
+  THR_NEARESTMV,
+  THR_NEARESTA,
+  THR_NEARESTG,
+
+  THR_DC,
+
+  THR_NEWMV,
+  THR_NEWA,
+  THR_NEWG,
+
+  THR_NEARMV,
+  THR_NEARA,
+  THR_NEARG,
+
+  THR_ZEROMV,
+  THR_ZEROG,
+  THR_ZEROA,
+
+  THR_COMP_NEARESTLA,
+  THR_COMP_NEARESTGA,
+
+  THR_TM,
+
+  THR_COMP_NEARLA,
+  THR_COMP_NEWLA,
+  THR_COMP_NEARGA,
+  THR_COMP_NEWGA,
+
+  THR_COMP_ZEROLA,
+  THR_COMP_ZEROGA,
+
+  THR_H_PRED,
+  THR_V_PRED,
+  THR_D135_PRED,
+  THR_D207_PRED,
+  THR_D153_PRED,
+  THR_D63_PRED,
+  THR_D117_PRED,
+  THR_D45_PRED,
+} THR_MODES;
+
+typedef enum {
+  THR_LAST,
+  THR_GOLD,
+  THR_ALTR,
+  THR_COMP_LA,
+  THR_COMP_GA,
+  THR_INTRA,
+} THR_MODES_SUB8X8;
+
+typedef struct RD_OPT {
+  // Thresh_mult is used to set a threshold for the rd score. A higher value
+  // means that we will accept the best mode so far more often. This number
+  // is used in combination with the current block size, and thresh_freq_fact
+  // to pick a threshold.
+  int thresh_mult[MAX_MODES];
+  int thresh_mult_sub8x8[MAX_REFS];
+
+  int threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
+
+  int64_t prediction_type_threshes[MAX_REF_FRAMES][REFERENCE_MODES];
+
+  int64_t filter_threshes[MAX_REF_FRAMES][SWITCHABLE_FILTER_CONTEXTS];
+
+  int RDMULT;
+  int RDDIV;
+} RD_OPT;
+
+typedef struct RD_COST {
+  int rate;
+  int64_t dist;
+  int64_t rdcost;
+} RD_COST;
+
+// Reset the rate distortion cost values to maximum (invalid) value.
+void vp9_rd_cost_reset(RD_COST *rd_cost);
+// Initialize the rate distortion cost values to zero.
+void vp9_rd_cost_init(RD_COST *rd_cost);
+
+struct TileInfo;
+struct TileDataEnc;
+struct VP9_COMP;
+struct macroblock;
+
+int vp9_compute_rd_mult(const struct VP9_COMP *cpi, int qindex);
+
+void vp9_initialize_rd_consts(struct VP9_COMP *cpi);
+
+void vp9_initialize_me_consts(struct VP9_COMP *cpi, MACROBLOCK *x, int qindex);
+
+void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n,
+                                  unsigned int qstep, int *rate,
+                                  int64_t *dist);
+
+int vp9_get_switchable_rate(const struct VP9_COMP *cpi,
+                            const MACROBLOCKD *const xd);
+
+int vp9_raster_block_offset(BLOCK_SIZE plane_bsize,
+                            int raster_block, int stride);
+
+int16_t* vp9_raster_block_offset_int16(BLOCK_SIZE plane_bsize,
+                                       int raster_block, int16_t *base);
+
+YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const struct VP9_COMP *cpi,
+                                             int ref_frame);
+
+void vp9_init_me_luts(void);
+
+void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              const struct macroblockd_plane *pd,
+                              ENTROPY_CONTEXT t_above[16],
+                              ENTROPY_CONTEXT t_left[16]);
+
+void vp9_set_rd_speed_thresholds(struct VP9_COMP *cpi);
+
+void vp9_set_rd_speed_thresholds_sub8x8(struct VP9_COMP *cpi);
+
+void vp9_update_rd_thresh_fact(int (*fact)[MAX_MODES], int rd_thresh,
+                               int bsize, int best_mode_index);
+
+static INLINE int rd_less_than_thresh(int64_t best_rd, int thresh,
+                                      int thresh_fact) {
+    return best_rd < ((int64_t)thresh * thresh_fact >> 5) || thresh == INT_MAX;
+}
+
+static INLINE void set_error_per_bit(MACROBLOCK *x, int rdmult) {
+  x->errorperbit = rdmult >> RD_EPB_SHIFT;
+  x->errorperbit += (x->errorperbit == 0);
+}
+
+void vp9_mv_pred(struct VP9_COMP *cpi, MACROBLOCK *x,
+                 uint8_t *ref_y_buffer, int ref_y_stride,
+                 int ref_frame, BLOCK_SIZE block_size);
+
+void vp9_setup_pred_block(const MACROBLOCKD *xd,
+                          struct buf_2d dst[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col,
+                          const struct scale_factors *scale,
+                          const struct scale_factors *scale_uv);
+
+int vp9_get_intra_cost_penalty(int qindex, int qdelta,
+                               vpx_bit_depth_t bit_depth);
+
+unsigned int vp9_get_sby_perpixel_variance(struct VP9_COMP *cpi,
+                                           const struct buf_2d *ref,
+                                           BLOCK_SIZE bs);
+#if CONFIG_VP9_HIGHBITDEPTH
+unsigned int vp9_high_get_sby_perpixel_variance(struct VP9_COMP *cpi,
+                                                const struct buf_2d *ref,
+                                                BLOCK_SIZE bs, int bd);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_RD_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_rdopt.c b/libvpx/libvpx/vp9/encoder/vp9_rdopt.c
new file mode 100644
index 0000000..e65e051
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_rdopt.c
@@ -0,0 +1,4398 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <math.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/system_state.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_idct.h"
+#include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_scan.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_encodemb.h"
+#include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_rdopt.h"
+#include "vp9/encoder/vp9_aq_variance.h"
+
+#define LAST_FRAME_MODE_MASK    ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | \
+                                 (1 << INTRA_FRAME))
+#define GOLDEN_FRAME_MODE_MASK  ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | \
+                                 (1 << INTRA_FRAME))
+#define ALT_REF_MODE_MASK       ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | \
+                                 (1 << INTRA_FRAME))
+
+#define SECOND_REF_FRAME_MASK   ((1 << ALTREF_FRAME) | 0x01)
+
+#define MIN_EARLY_TERM_INDEX    3
+#define NEW_MV_DISCOUNT_FACTOR  8
+
+typedef struct {
+  PREDICTION_MODE mode;
+  MV_REFERENCE_FRAME ref_frame[2];
+} MODE_DEFINITION;
+
+typedef struct {
+  MV_REFERENCE_FRAME ref_frame[2];
+} REF_DEFINITION;
+
+struct rdcost_block_args {
+  MACROBLOCK *x;
+  ENTROPY_CONTEXT t_above[16];
+  ENTROPY_CONTEXT t_left[16];
+  int this_rate;
+  int64_t this_dist;
+  int64_t this_sse;
+  int64_t this_rd;
+  int64_t best_rd;
+  int exit_early;
+  int use_fast_coef_costing;
+  const scan_order *so;
+  uint8_t skippable;
+};
+
+#define LAST_NEW_MV_INDEX 6
+static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
+  {NEARESTMV, {LAST_FRAME,   NONE}},
+  {NEARESTMV, {ALTREF_FRAME, NONE}},
+  {NEARESTMV, {GOLDEN_FRAME, NONE}},
+
+  {DC_PRED,   {INTRA_FRAME,  NONE}},
+
+  {NEWMV,     {LAST_FRAME,   NONE}},
+  {NEWMV,     {ALTREF_FRAME, NONE}},
+  {NEWMV,     {GOLDEN_FRAME, NONE}},
+
+  {NEARMV,    {LAST_FRAME,   NONE}},
+  {NEARMV,    {ALTREF_FRAME, NONE}},
+  {NEARMV,    {GOLDEN_FRAME, NONE}},
+
+  {ZEROMV,    {LAST_FRAME,   NONE}},
+  {ZEROMV,    {GOLDEN_FRAME, NONE}},
+  {ZEROMV,    {ALTREF_FRAME, NONE}},
+
+  {NEARESTMV, {LAST_FRAME,   ALTREF_FRAME}},
+  {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}},
+
+  {TM_PRED,   {INTRA_FRAME,  NONE}},
+
+  {NEARMV,    {LAST_FRAME,   ALTREF_FRAME}},
+  {NEWMV,     {LAST_FRAME,   ALTREF_FRAME}},
+  {NEARMV,    {GOLDEN_FRAME, ALTREF_FRAME}},
+  {NEWMV,     {GOLDEN_FRAME, ALTREF_FRAME}},
+
+  {ZEROMV,    {LAST_FRAME,   ALTREF_FRAME}},
+  {ZEROMV,    {GOLDEN_FRAME, ALTREF_FRAME}},
+
+  {H_PRED,    {INTRA_FRAME,  NONE}},
+  {V_PRED,    {INTRA_FRAME,  NONE}},
+  {D135_PRED, {INTRA_FRAME,  NONE}},
+  {D207_PRED, {INTRA_FRAME,  NONE}},
+  {D153_PRED, {INTRA_FRAME,  NONE}},
+  {D63_PRED,  {INTRA_FRAME,  NONE}},
+  {D117_PRED, {INTRA_FRAME,  NONE}},
+  {D45_PRED,  {INTRA_FRAME,  NONE}},
+};
+
+static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
+  {{LAST_FRAME,   NONE}},
+  {{GOLDEN_FRAME, NONE}},
+  {{ALTREF_FRAME, NONE}},
+  {{LAST_FRAME,   ALTREF_FRAME}},
+  {{GOLDEN_FRAME, ALTREF_FRAME}},
+  {{INTRA_FRAME,  NONE}},
+};
+
+static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
+                           int m, int n, int min_plane, int max_plane) {
+  int i;
+
+  for (i = min_plane; i < max_plane; ++i) {
+    struct macroblock_plane *const p = &x->plane[i];
+    struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
+
+    p->coeff    = ctx->coeff_pbuf[i][m];
+    p->qcoeff   = ctx->qcoeff_pbuf[i][m];
+    pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
+    p->eobs     = ctx->eobs_pbuf[i][m];
+
+    ctx->coeff_pbuf[i][m]   = ctx->coeff_pbuf[i][n];
+    ctx->qcoeff_pbuf[i][m]  = ctx->qcoeff_pbuf[i][n];
+    ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
+    ctx->eobs_pbuf[i][m]    = ctx->eobs_pbuf[i][n];
+
+    ctx->coeff_pbuf[i][n]   = p->coeff;
+    ctx->qcoeff_pbuf[i][n]  = p->qcoeff;
+    ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
+    ctx->eobs_pbuf[i][n]    = p->eobs;
+  }
+}
+
+static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
+                            MACROBLOCK *x, MACROBLOCKD *xd,
+                            int *out_rate_sum, int64_t *out_dist_sum,
+                            int *skip_txfm_sb, int64_t *skip_sse_sb) {
+  // Note our transform coeffs are 8 times an orthogonal transform.
+  // Hence quantizer step is also 8 times. To get effective quantizer
+  // we need to divide by 8 before sending to modeling function.
+  int i;
+  int64_t rate_sum = 0;
+  int64_t dist_sum = 0;
+  const int ref = xd->mi[0]->ref_frame[0];
+  unsigned int sse;
+  unsigned int var = 0;
+  unsigned int sum_sse = 0;
+  int64_t total_sse = 0;
+  int skip_flag = 1;
+  const int shift = 6;
+  int rate;
+  int64_t dist;
+  const int dequant_shift =
+#if CONFIG_VP9_HIGHBITDEPTH
+      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
+          xd->bd - 5 :
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+          3;
+
+  x->pred_sse[ref] = 0;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    struct macroblock_plane *const p = &x->plane[i];
+    struct macroblockd_plane *const pd = &xd->plane[i];
+    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
+    const TX_SIZE max_tx_size = max_txsize_lookup[bs];
+    const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
+    const int64_t dc_thr = p->quant_thred[0] >> shift;
+    const int64_t ac_thr = p->quant_thred[1] >> shift;
+    // The low thresholds are used to measure if the prediction errors are
+    // low enough so that we can skip the mode search.
+    const int64_t low_dc_thr = VPXMIN(50, dc_thr >> 2);
+    const int64_t low_ac_thr = VPXMIN(80, ac_thr >> 2);
+    int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
+    int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
+    int idx, idy;
+    int lw = b_width_log2_lookup[unit_size] + 2;
+    int lh = b_height_log2_lookup[unit_size] + 2;
+
+    sum_sse = 0;
+
+    for (idy = 0; idy < bh; ++idy) {
+      for (idx = 0; idx < bw; ++idx) {
+        uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
+        uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
+        int block_idx = (idy << 1) + idx;
+        int low_err_skip = 0;
+
+        var = cpi->fn_ptr[unit_size].vf(src, p->src.stride,
+                                        dst, pd->dst.stride, &sse);
+        x->bsse[(i << 2) + block_idx] = sse;
+        sum_sse += sse;
+
+        x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_NONE;
+        if (!x->select_tx_size) {
+          // Check if all ac coefficients can be quantized to zero.
+          if (var < ac_thr || var == 0) {
+            x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_ONLY;
+
+            // Check if dc coefficient can be quantized to zero.
+            if (sse - var < dc_thr || sse == var) {
+              x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_DC;
+
+              if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
+                low_err_skip = 1;
+            }
+          }
+        }
+
+        if (skip_flag && !low_err_skip)
+          skip_flag = 0;
+
+        if (i == 0)
+          x->pred_sse[ref] += sse;
+      }
+    }
+
+    total_sse += sum_sse;
+
+    // Fast approximate the modelling function.
+    if (cpi->sf.simple_model_rd_from_var) {
+      int64_t rate;
+      const int64_t square_error = sum_sse;
+      int quantizer = (pd->dequant[1] >> dequant_shift);
+
+      if (quantizer < 120)
+        rate = (square_error * (280 - quantizer)) >> (16 - VP9_PROB_COST_SHIFT);
+      else
+        rate = 0;
+      dist = (square_error * quantizer) >> 8;
+      rate_sum += rate;
+      dist_sum += dist;
+    } else {
+      vp9_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
+                                   pd->dequant[1] >> dequant_shift,
+                                   &rate, &dist);
+      rate_sum += rate;
+      dist_sum += dist;
+    }
+  }
+
+  *skip_txfm_sb = skip_flag;
+  *skip_sse_sb = total_sse << 4;
+  *out_rate_sum = (int)rate_sum;
+  *out_dist_sum = dist_sum << 4;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+int64_t vp9_highbd_block_error_c(const tran_low_t *coeff,
+                                 const tran_low_t *dqcoeff,
+                                 intptr_t block_size,
+                                 int64_t *ssz, int bd) {
+  int i;
+  int64_t error = 0, sqcoeff = 0;
+  int shift = 2 * (bd - 8);
+  int rounding = shift > 0 ? 1 << (shift - 1) : 0;
+
+  for (i = 0; i < block_size; i++) {
+    const int64_t diff = coeff[i] - dqcoeff[i];
+    error +=  diff * diff;
+    sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
+  }
+  assert(error >= 0 && sqcoeff >= 0);
+  error = (error + rounding) >> shift;
+  sqcoeff = (sqcoeff + rounding) >> shift;
+
+  *ssz = sqcoeff;
+  return error;
+}
+
+int64_t vp9_highbd_block_error_8bit_c(const tran_low_t *coeff,
+                                      const tran_low_t *dqcoeff,
+                                      intptr_t block_size,
+                                      int64_t *ssz) {
+  // Note that the C versions of these 2 functions (vp9_block_error and
+  // vp9_highbd_block_error_8bit are the same, but the optimized assembly
+  // routines are not compatible in the non high bitdepth configuration, so
+  // they still cannot share the same name.
+  return vp9_block_error_c(coeff, dqcoeff, block_size, ssz);
+}
+
+static int64_t vp9_highbd_block_error_dispatch(const tran_low_t *coeff,
+                                               const tran_low_t *dqcoeff,
+                                               intptr_t block_size,
+                                               int64_t *ssz, int bd) {
+  if (bd == 8) {
+    return vp9_highbd_block_error_8bit(coeff, dqcoeff, block_size, ssz);
+  } else {
+    return vp9_highbd_block_error(coeff, dqcoeff, block_size, ssz, bd);
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
+                          intptr_t block_size, int64_t *ssz) {
+  int i;
+  int64_t error = 0, sqcoeff = 0;
+
+  for (i = 0; i < block_size; i++) {
+    const int diff = coeff[i] - dqcoeff[i];
+    error +=  diff * diff;
+    sqcoeff += coeff[i] * coeff[i];
+  }
+
+  *ssz = sqcoeff;
+  return error;
+}
+
+int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff,
+                             int block_size) {
+  int i;
+  int64_t error = 0;
+
+  for (i = 0; i < block_size; i++) {
+    const int diff = coeff[i] - dqcoeff[i];
+    error +=  diff * diff;
+  }
+
+  return error;
+}
+
+/* The trailing '0' is a terminator which is used inside cost_coeffs() to
+ * decide whether to include cost of a trailing EOB node or not (i.e. we
+ * can skip this if the last coefficient in this transform block, e.g. the
+ * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
+ * were non-zero). */
+static const int16_t band_counts[TX_SIZES][8] = {
+  { 1, 2, 3, 4,  3,   16 - 13, 0 },
+  { 1, 2, 3, 4, 11,   64 - 21, 0 },
+  { 1, 2, 3, 4, 11,  256 - 21, 0 },
+  { 1, 2, 3, 4, 11, 1024 - 21, 0 },
+};
+static int cost_coeffs(MACROBLOCK *x,
+                       int plane, int block,
+                       ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
+                       TX_SIZE tx_size,
+                       const int16_t *scan, const int16_t *nb,
+                       int use_fast_coef_costing) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  const struct macroblock_plane *p = &x->plane[plane];
+  const PLANE_TYPE type = get_plane_type(plane);
+  const int16_t *band_count = &band_counts[tx_size][1];
+  const int eob = p->eobs[block];
+  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
+                   x->token_costs[tx_size][type][is_inter_block(mi)];
+  uint8_t token_cache[32 * 32];
+  int pt = combine_entropy_contexts(*A, *L);
+  int c, cost;
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
+#else
+  const int *cat6_high_cost = vp9_get_high_cost_table(8);
+#endif
+
+  // Check for consistency of tx_size with mode info
+  assert(type == PLANE_TYPE_Y ? mi->tx_size == tx_size :
+         get_uv_tx_size(mi, &xd->plane[plane]) == tx_size);
+
+  if (eob == 0) {
+    // single eob token
+    cost = token_costs[0][0][pt][EOB_TOKEN];
+    c = 0;
+  } else {
+    if (use_fast_coef_costing) {
+      int band_left = *band_count++;
+
+      // dc token
+      int v = qcoeff[0];
+      int16_t prev_t;
+      cost = vp9_get_token_cost(v, &prev_t, cat6_high_cost);
+      cost += (*token_costs)[0][pt][prev_t];
+
+      token_cache[0] = vp9_pt_energy_class[prev_t];
+      ++token_costs;
+
+      // ac tokens
+      for (c = 1; c < eob; c++) {
+        const int rc = scan[c];
+        int16_t t;
+
+        v = qcoeff[rc];
+        cost += vp9_get_token_cost(v, &t, cat6_high_cost);
+        cost += (*token_costs)[!prev_t][!prev_t][t];
+        prev_t = t;
+        if (!--band_left) {
+          band_left = *band_count++;
+          ++token_costs;
+        }
+      }
+
+      // eob token
+      if (band_left)
+        cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
+
+    } else {  // !use_fast_coef_costing
+      int band_left = *band_count++;
+
+      // dc token
+      int v = qcoeff[0];
+      int16_t tok;
+      unsigned int (*tok_cost_ptr)[COEFF_CONTEXTS][ENTROPY_TOKENS];
+      cost = vp9_get_token_cost(v, &tok, cat6_high_cost);
+      cost += (*token_costs)[0][pt][tok];
+
+      token_cache[0] = vp9_pt_energy_class[tok];
+      ++token_costs;
+
+      tok_cost_ptr = &((*token_costs)[!tok]);
+
+      // ac tokens
+      for (c = 1; c < eob; c++) {
+        const int rc = scan[c];
+
+        v = qcoeff[rc];
+        cost += vp9_get_token_cost(v, &tok, cat6_high_cost);
+        pt = get_coef_context(nb, token_cache, c);
+        cost += (*tok_cost_ptr)[pt][tok];
+        token_cache[rc] = vp9_pt_energy_class[tok];
+        if (!--band_left) {
+          band_left = *band_count++;
+          ++token_costs;
+        }
+        tok_cost_ptr = &((*token_costs)[!tok]);
+      }
+
+      // eob token
+      if (band_left) {
+        pt = get_coef_context(nb, token_cache, c);
+        cost += (*token_costs)[0][pt][EOB_TOKEN];
+      }
+    }
+  }
+
+  // is eob first coefficient;
+  *A = *L = (c > 0);
+
+  return cost;
+}
+
+static void dist_block(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
+                       int64_t *out_dist, int64_t *out_sse) {
+  const int ss_txfrm_size = tx_size << 1;
+  MACROBLOCKD* const xd = &x->e_mbd;
+  const struct macroblock_plane *const p = &x->plane[plane];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  int64_t this_sse;
+  int shift = tx_size == TX_32X32 ? 0 : 2;
+  tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
+  *out_dist = vp9_highbd_block_error_dispatch(coeff, dqcoeff,
+                                              16 << ss_txfrm_size,
+                                              &this_sse, bd) >> shift;
+#else
+  *out_dist = vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
+                              &this_sse) >> shift;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  *out_sse = this_sse >> shift;
+
+  if (x->skip_encode && !is_inter_block(xd->mi[0])) {
+    // TODO(jingning): tune the model to better capture the distortion.
+    int64_t p = (pd->dequant[1] * pd->dequant[1] *
+                    (1 << ss_txfrm_size)) >>
+#if CONFIG_VP9_HIGHBITDEPTH
+                        (shift + 2 + (bd - 8) * 2);
+#else
+                        (shift + 2);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    *out_dist += (p >> 4);
+    *out_sse  += p;
+  }
+}
+
+static int rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
+                      TX_SIZE tx_size, struct rdcost_block_args* args) {
+  int x_idx, y_idx;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx);
+
+  return cost_coeffs(args->x, plane, block, args->t_above + x_idx,
+                     args->t_left + y_idx, tx_size,
+                     args->so->scan, args->so->neighbors,
+                     args->use_fast_coef_costing);
+}
+
+static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
+                          TX_SIZE tx_size, void *arg) {
+  struct rdcost_block_args *args = arg;
+  MACROBLOCK *const x = args->x;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  int64_t rd1, rd2, rd;
+  int rate;
+  int64_t dist;
+  int64_t sse;
+
+  if (args->exit_early)
+    return;
+
+  if (!is_inter_block(mi)) {
+    struct encode_b_args arg = {x, NULL, &mi->skip};
+    vp9_encode_block_intra(plane, block, plane_bsize, tx_size, &arg);
+    dist_block(x, plane, block, tx_size, &dist, &sse);
+  } else if (max_txsize_lookup[plane_bsize] == tx_size) {
+    if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
+        SKIP_TXFM_NONE) {
+      // full forward transform and quantization
+      vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
+      dist_block(x, plane, block, tx_size, &dist, &sse);
+    } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
+               SKIP_TXFM_AC_ONLY) {
+      // compute DC coefficient
+      tran_low_t *const coeff   = BLOCK_OFFSET(x->plane[plane].coeff, block);
+      tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
+      vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
+      sse  = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
+      dist = sse;
+      if (x->plane[plane].eobs[block]) {
+        const int64_t orig_sse = (int64_t)coeff[0] * coeff[0];
+        const int64_t resd_sse = coeff[0] - dqcoeff[0];
+        int64_t dc_correct = orig_sse - resd_sse * resd_sse;
+#if CONFIG_VP9_HIGHBITDEPTH
+        dc_correct >>= ((xd->bd - 8) * 2);
+#endif
+        if (tx_size != TX_32X32)
+          dc_correct >>= 2;
+
+        dist = VPXMAX(0, sse - dc_correct);
+      }
+    } else {
+      // SKIP_TXFM_AC_DC
+      // skip forward transform
+      x->plane[plane].eobs[block] = 0;
+      sse  = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
+      dist = sse;
+    }
+  } else {
+    // full forward transform and quantization
+    vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
+    dist_block(x, plane, block, tx_size, &dist, &sse);
+  }
+
+  rd = RDCOST(x->rdmult, x->rddiv, 0, dist);
+  if (args->this_rd + rd > args->best_rd) {
+    args->exit_early = 1;
+    return;
+  }
+
+  rate = rate_block(plane, block, plane_bsize, tx_size, args);
+  rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
+  rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
+
+  // TODO(jingning): temporarily enabled only for luma component
+  rd = VPXMIN(rd1, rd2);
+  if (plane == 0)
+    x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
+                                    (rd1 > rd2 && !xd->lossless);
+
+  args->this_rate += rate;
+  args->this_dist += dist;
+  args->this_sse += sse;
+  args->this_rd += rd;
+
+  if (args->this_rd > args->best_rd) {
+    args->exit_early = 1;
+    return;
+  }
+
+  args->skippable &= !x->plane[plane].eobs[block];
+}
+
+static void txfm_rd_in_plane(MACROBLOCK *x,
+                             int *rate, int64_t *distortion,
+                             int *skippable, int64_t *sse,
+                             int64_t ref_best_rd, int plane,
+                             BLOCK_SIZE bsize, TX_SIZE tx_size,
+                             int use_fast_coef_casting) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  struct rdcost_block_args args;
+  vp9_zero(args);
+  args.x = x;
+  args.best_rd = ref_best_rd;
+  args.use_fast_coef_costing = use_fast_coef_casting;
+  args.skippable = 1;
+
+  if (plane == 0)
+    xd->mi[0]->tx_size = tx_size;
+
+  vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
+
+  args.so = get_scan(xd, tx_size, get_plane_type(plane), 0);
+
+  vp9_foreach_transformed_block_in_plane(xd, bsize, plane,
+                                         block_rd_txfm, &args);
+  if (args.exit_early) {
+    *rate       = INT_MAX;
+    *distortion = INT64_MAX;
+    *sse        = INT64_MAX;
+    *skippable  = 0;
+  } else {
+    *distortion = args.this_dist;
+    *rate       = args.this_rate;
+    *sse        = args.this_sse;
+    *skippable  = args.skippable;
+  }
+}
+
+static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x,
+                                   int *rate, int64_t *distortion,
+                                   int *skip, int64_t *sse,
+                                   int64_t ref_best_rd,
+                                   BLOCK_SIZE bs) {
+  const TX_SIZE max_tx_size = max_txsize_lookup[bs];
+  VP9_COMMON *const cm = &cpi->common;
+  const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+
+  mi->tx_size = VPXMIN(max_tx_size, largest_tx_size);
+
+  txfm_rd_in_plane(x, rate, distortion, skip,
+                   sse, ref_best_rd, 0, bs,
+                   mi->tx_size, cpi->sf.use_fast_coef_costing);
+}
+
+static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
+                                   int *rate,
+                                   int64_t *distortion,
+                                   int *skip,
+                                   int64_t *psse,
+                                   int64_t ref_best_rd,
+                                   BLOCK_SIZE bs) {
+  const TX_SIZE max_tx_size = max_txsize_lookup[bs];
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
+  int r[TX_SIZES][2], s[TX_SIZES];
+  int64_t d[TX_SIZES], sse[TX_SIZES];
+  int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
+                             {INT64_MAX, INT64_MAX},
+                             {INT64_MAX, INT64_MAX},
+                             {INT64_MAX, INT64_MAX}};
+  int n, m;
+  int s0, s1;
+  int64_t best_rd = INT64_MAX;
+  TX_SIZE best_tx = max_tx_size;
+  int start_tx, end_tx;
+
+  const vpx_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc->tx_probs);
+  assert(skip_prob > 0);
+  s0 = vp9_cost_bit(skip_prob, 0);
+  s1 = vp9_cost_bit(skip_prob, 1);
+
+  if (cm->tx_mode == TX_MODE_SELECT) {
+    start_tx = max_tx_size;
+    end_tx = 0;
+  } else {
+    TX_SIZE chosen_tx_size = VPXMIN(max_tx_size,
+                                    tx_mode_to_biggest_tx_size[cm->tx_mode]);
+    start_tx = chosen_tx_size;
+    end_tx = chosen_tx_size;
+  }
+
+  for (n = start_tx; n >= end_tx; n--) {
+    int r_tx_size = 0;
+    for (m = 0; m <= n - (n == (int) max_tx_size); m++) {
+      if (m == n)
+        r_tx_size += vp9_cost_zero(tx_probs[m]);
+      else
+        r_tx_size += vp9_cost_one(tx_probs[m]);
+    }
+    txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
+                     &sse[n], ref_best_rd, 0, bs, n,
+                     cpi->sf.use_fast_coef_costing);
+    r[n][1] = r[n][0];
+    if (r[n][0] < INT_MAX) {
+      r[n][1] += r_tx_size;
+    }
+    if (d[n] == INT64_MAX || r[n][0] == INT_MAX) {
+      rd[n][0] = rd[n][1] = INT64_MAX;
+    } else if (s[n]) {
+      if (is_inter_block(mi)) {
+        rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
+        r[n][1] -= r_tx_size;
+      } else {
+        rd[n][0] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
+        rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size, sse[n]);
+      }
+    } else {
+      rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
+      rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
+    }
+
+    if (is_inter_block(mi) && !xd->lossless && !s[n] && sse[n] != INT64_MAX) {
+      rd[n][0] = VPXMIN(rd[n][0], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
+      rd[n][1] = VPXMIN(rd[n][1], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
+    }
+
+    // Early termination in transform size search.
+    if (cpi->sf.tx_size_search_breakout &&
+        (rd[n][1] == INT64_MAX ||
+        (n < (int) max_tx_size && rd[n][1] > rd[n + 1][1]) ||
+        s[n] == 1))
+      break;
+
+    if (rd[n][1] < best_rd) {
+      best_tx = n;
+      best_rd = rd[n][1];
+    }
+  }
+  mi->tx_size = best_tx;
+
+  *distortion = d[mi->tx_size];
+  *rate       = r[mi->tx_size][cm->tx_mode == TX_MODE_SELECT];
+  *skip       = s[mi->tx_size];
+  *psse       = sse[mi->tx_size];
+}
+
+static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
+                            int64_t *distortion, int *skip,
+                            int64_t *psse, BLOCK_SIZE bs,
+                            int64_t ref_best_rd) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  int64_t sse;
+  int64_t *ret_sse = psse ? psse : &sse;
+
+  assert(bs == xd->mi[0]->sb_type);
+
+  if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
+    choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
+                           bs);
+  } else {
+    choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse,
+                           ref_best_rd, bs);
+  }
+}
+
+static int conditional_skipintra(PREDICTION_MODE mode,
+                                 PREDICTION_MODE best_intra_mode) {
+  if (mode == D117_PRED &&
+      best_intra_mode != V_PRED &&
+      best_intra_mode != D135_PRED)
+    return 1;
+  if (mode == D63_PRED &&
+      best_intra_mode != V_PRED &&
+      best_intra_mode != D45_PRED)
+    return 1;
+  if (mode == D207_PRED &&
+      best_intra_mode != H_PRED &&
+      best_intra_mode != D45_PRED)
+    return 1;
+  if (mode == D153_PRED &&
+      best_intra_mode != H_PRED &&
+      best_intra_mode != D135_PRED)
+    return 1;
+  return 0;
+}
+
+static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x,
+                                     int row, int col,
+                                     PREDICTION_MODE *best_mode,
+                                     const int *bmode_costs,
+                                     ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
+                                     int *bestrate, int *bestratey,
+                                     int64_t *bestdistortion,
+                                     BLOCK_SIZE bsize, int64_t rd_thresh) {
+  PREDICTION_MODE mode;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int64_t best_rd = rd_thresh;
+  struct macroblock_plane *p = &x->plane[0];
+  struct macroblockd_plane *pd = &xd->plane[0];
+  const int src_stride = p->src.stride;
+  const int dst_stride = pd->dst.stride;
+  const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
+  uint8_t *dst_init = &pd->dst.buf[row * 4 * src_stride + col * 4];
+  ENTROPY_CONTEXT ta[2], tempa[2];
+  ENTROPY_CONTEXT tl[2], templ[2];
+  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
+  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
+  int idx, idy;
+  uint8_t best_dst[8 * 8];
+#if CONFIG_VP9_HIGHBITDEPTH
+  uint16_t best_dst16[8 * 8];
+#endif
+  memcpy(ta, a, num_4x4_blocks_wide * sizeof(a[0]));
+  memcpy(tl, l, num_4x4_blocks_high * sizeof(l[0]));
+
+  xd->mi[0]->tx_size = TX_4X4;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
+      int64_t this_rd;
+      int ratey = 0;
+      int64_t distortion = 0;
+      int rate = bmode_costs[mode];
+
+      if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
+        continue;
+
+      // Only do the oblique modes if the best so far is
+      // one of the neighboring directional modes
+      if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
+        if (conditional_skipintra(mode, *best_mode))
+            continue;
+      }
+
+      memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
+      memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
+
+      for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
+        for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
+          const int block = (row + idy) * 2 + (col + idx);
+          const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
+          uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
+          int16_t *const src_diff = vp9_raster_block_offset_int16(BLOCK_8X8,
+                                                                  block,
+                                                                  p->src_diff);
+          tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
+          xd->mi[0]->bmi[block].as_mode = mode;
+          vp9_predict_intra_block(xd, 1, TX_4X4, mode,
+                                  x->skip_encode ? src : dst,
+                                  x->skip_encode ? src_stride : dst_stride,
+                                  dst, dst_stride,
+                                  col + idx, row + idy, 0);
+          vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride,
+                                    dst, dst_stride, xd->bd);
+          if (xd->lossless) {
+            const scan_order *so = &vp9_default_scan_orders[TX_4X4];
+            vp9_highbd_fwht4x4(src_diff, coeff, 8);
+            vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
+            ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
+                                 so->scan, so->neighbors,
+                                 cpi->sf.use_fast_coef_costing);
+            if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
+              goto next_highbd;
+            vp9_highbd_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block),
+                                   dst, dst_stride,
+                                   p->eobs[block], xd->bd);
+          } else {
+            int64_t unused;
+            const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
+            const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
+            if (tx_type == DCT_DCT)
+              vpx_highbd_fdct4x4(src_diff, coeff, 8);
+            else
+              vp9_highbd_fht4x4(src_diff, coeff, 8, tx_type);
+            vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
+            ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
+                                 so->scan, so->neighbors,
+                                 cpi->sf.use_fast_coef_costing);
+            distortion += vp9_highbd_block_error_dispatch(
+                coeff, BLOCK_OFFSET(pd->dqcoeff, block),
+                16, &unused, xd->bd) >> 2;
+            if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
+              goto next_highbd;
+            vp9_highbd_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
+                                  dst, dst_stride, p->eobs[block], xd->bd);
+          }
+        }
+      }
+
+      rate += ratey;
+      this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
+
+      if (this_rd < best_rd) {
+        *bestrate = rate;
+        *bestratey = ratey;
+        *bestdistortion = distortion;
+        best_rd = this_rd;
+        *best_mode = mode;
+        memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
+        memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
+        for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
+          memcpy(best_dst16 + idy * 8,
+                 CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
+                 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
+        }
+      }
+    next_highbd:
+      {}
+    }
+    if (best_rd >= rd_thresh || x->skip_encode)
+      return best_rd;
+
+    for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
+      memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
+             best_dst16 + idy * 8,
+             num_4x4_blocks_wide * 4 * sizeof(uint16_t));
+    }
+
+    return best_rd;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
+    int64_t this_rd;
+    int ratey = 0;
+    int64_t distortion = 0;
+    int rate = bmode_costs[mode];
+
+    if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
+      continue;
+
+    // Only do the oblique modes if the best so far is
+    // one of the neighboring directional modes
+    if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
+      if (conditional_skipintra(mode, *best_mode))
+          continue;
+    }
+
+    memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
+    memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
+
+    for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
+      for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
+        const int block = (row + idy) * 2 + (col + idx);
+        const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
+        uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
+        int16_t *const src_diff =
+            vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
+        tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
+        xd->mi[0]->bmi[block].as_mode = mode;
+        vp9_predict_intra_block(xd, 1, TX_4X4, mode,
+                                x->skip_encode ? src : dst,
+                                x->skip_encode ? src_stride : dst_stride,
+                                dst, dst_stride, col + idx, row + idy, 0);
+        vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
+
+        if (xd->lossless) {
+          const scan_order *so = &vp9_default_scan_orders[TX_4X4];
+          vp9_fwht4x4(src_diff, coeff, 8);
+          vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
+          ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
+                               so->scan, so->neighbors,
+                               cpi->sf.use_fast_coef_costing);
+          if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
+            goto next;
+          vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
+                          p->eobs[block]);
+        } else {
+          int64_t unused;
+          const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
+          const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
+          vp9_fht4x4(src_diff, coeff, 8, tx_type);
+          vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
+          ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
+                             so->scan, so->neighbors,
+                             cpi->sf.use_fast_coef_costing);
+#if CONFIG_VP9_HIGHBITDEPTH
+          distortion += vp9_highbd_block_error_8bit(
+              coeff, BLOCK_OFFSET(pd->dqcoeff, block), 16, &unused) >> 2;
+#else
+          distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
+                                        16, &unused) >> 2;
+#endif
+          if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
+            goto next;
+          vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
+                         dst, dst_stride, p->eobs[block]);
+        }
+      }
+    }
+
+    rate += ratey;
+    this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
+
+    if (this_rd < best_rd) {
+      *bestrate = rate;
+      *bestratey = ratey;
+      *bestdistortion = distortion;
+      best_rd = this_rd;
+      *best_mode = mode;
+      memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
+      memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
+      for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
+        memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
+               num_4x4_blocks_wide * 4);
+    }
+  next:
+    {}
+  }
+
+  if (best_rd >= rd_thresh || x->skip_encode)
+    return best_rd;
+
+  for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
+    memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
+           num_4x4_blocks_wide * 4);
+
+  return best_rd;
+}
+
+static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
+                                            int *rate, int *rate_y,
+                                            int64_t *distortion,
+                                            int64_t best_rd) {
+  int i, j;
+  const MACROBLOCKD *const xd = &mb->e_mbd;
+  MODE_INFO *const mic = xd->mi[0];
+  const MODE_INFO *above_mi = xd->above_mi;
+  const MODE_INFO *left_mi = xd->left_mi;
+  const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
+  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
+  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
+  int idx, idy;
+  int cost = 0;
+  int64_t total_distortion = 0;
+  int tot_rate_y = 0;
+  int64_t total_rd = 0;
+  const int *bmode_costs = cpi->mbmode_cost;
+
+  // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
+  for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
+    for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
+      PREDICTION_MODE best_mode = DC_PRED;
+      int r = INT_MAX, ry = INT_MAX;
+      int64_t d = INT64_MAX, this_rd = INT64_MAX;
+      i = idy * 2 + idx;
+      if (cpi->common.frame_type == KEY_FRAME) {
+        const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
+        const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
+
+        bmode_costs  = cpi->y_mode_costs[A][L];
+      }
+
+      this_rd = rd_pick_intra4x4block(cpi, mb, idy, idx, &best_mode,
+                                      bmode_costs,
+                                      xd->plane[0].above_context + idx,
+                                      xd->plane[0].left_context + idy,
+                                      &r, &ry, &d, bsize, best_rd - total_rd);
+
+      if (this_rd >= best_rd - total_rd)
+        return INT64_MAX;
+
+      total_rd += this_rd;
+      cost += r;
+      total_distortion += d;
+      tot_rate_y += ry;
+
+      mic->bmi[i].as_mode = best_mode;
+      for (j = 1; j < num_4x4_blocks_high; ++j)
+        mic->bmi[i + j * 2].as_mode = best_mode;
+      for (j = 1; j < num_4x4_blocks_wide; ++j)
+        mic->bmi[i + j].as_mode = best_mode;
+
+      if (total_rd >= best_rd)
+        return INT64_MAX;
+    }
+  }
+
+  *rate = cost;
+  *rate_y = tot_rate_y;
+  *distortion = total_distortion;
+  mic->mode = mic->bmi[3].as_mode;
+
+  return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
+}
+
+// This function is used only for intra_only frames
+static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
+                                      int *rate, int *rate_tokenonly,
+                                      int64_t *distortion, int *skippable,
+                                      BLOCK_SIZE bsize,
+                                      int64_t best_rd) {
+  PREDICTION_MODE mode;
+  PREDICTION_MODE mode_selected = DC_PRED;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mic = xd->mi[0];
+  int this_rate, this_rate_tokenonly, s;
+  int64_t this_distortion, this_rd;
+  TX_SIZE best_tx = TX_4X4;
+  int *bmode_costs;
+  const MODE_INFO *above_mi = xd->above_mi;
+  const MODE_INFO *left_mi = xd->left_mi;
+  const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
+  const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
+  bmode_costs = cpi->y_mode_costs[A][L];
+
+  memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
+  /* Y Search for intra prediction mode */
+  for (mode = DC_PRED; mode <= TM_PRED; mode++) {
+    if (cpi->sf.use_nonrd_pick_mode) {
+      // These speed features are turned on in hybrid non-RD and RD mode
+      // for key frame coding in the context of real-time setting.
+      if (conditional_skipintra(mode, mode_selected))
+          continue;
+      if (*skippable)
+        break;
+    }
+
+    mic->mode = mode;
+
+    super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion,
+        &s, NULL, bsize, best_rd);
+
+    if (this_rate_tokenonly == INT_MAX)
+      continue;
+
+    this_rate = this_rate_tokenonly + bmode_costs[mode];
+    this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
+
+    if (this_rd < best_rd) {
+      mode_selected   = mode;
+      best_rd         = this_rd;
+      best_tx         = mic->tx_size;
+      *rate           = this_rate;
+      *rate_tokenonly = this_rate_tokenonly;
+      *distortion     = this_distortion;
+      *skippable      = s;
+    }
+  }
+
+  mic->mode = mode_selected;
+  mic->tx_size = best_tx;
+
+  return best_rd;
+}
+
+// Return value 0: early termination triggered, no valid rd cost available;
+//              1: rd cost values are valid.
+static int super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x,
+                            int *rate, int64_t *distortion, int *skippable,
+                            int64_t *sse, BLOCK_SIZE bsize,
+                            int64_t ref_best_rd) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  const TX_SIZE uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
+  int plane;
+  int pnrate = 0, pnskip = 1;
+  int64_t pndist = 0, pnsse = 0;
+  int is_cost_valid = 1;
+
+  if (ref_best_rd < 0)
+    is_cost_valid = 0;
+
+  if (is_inter_block(mi) && is_cost_valid) {
+    int plane;
+    for (plane = 1; plane < MAX_MB_PLANE; ++plane)
+      vp9_subtract_plane(x, bsize, plane);
+  }
+
+  *rate = 0;
+  *distortion = 0;
+  *sse = 0;
+  *skippable = 1;
+
+  for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
+    txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse,
+                     ref_best_rd, plane, bsize, uv_tx_size,
+                     cpi->sf.use_fast_coef_costing);
+    if (pnrate == INT_MAX) {
+      is_cost_valid = 0;
+      break;
+    }
+    *rate += pnrate;
+    *distortion += pndist;
+    *sse += pnsse;
+    *skippable &= pnskip;
+  }
+
+  if (!is_cost_valid) {
+    // reset cost value
+    *rate = INT_MAX;
+    *distortion = INT64_MAX;
+    *sse = INT64_MAX;
+    *skippable = 0;
+  }
+
+  return is_cost_valid;
+}
+
+static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
+                                       PICK_MODE_CONTEXT *ctx,
+                                       int *rate, int *rate_tokenonly,
+                                       int64_t *distortion, int *skippable,
+                                       BLOCK_SIZE bsize, TX_SIZE max_tx_size) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  PREDICTION_MODE mode;
+  PREDICTION_MODE mode_selected = DC_PRED;
+  int64_t best_rd = INT64_MAX, this_rd;
+  int this_rate_tokenonly, this_rate, s;
+  int64_t this_distortion, this_sse;
+
+  memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
+  for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
+    if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode)))
+      continue;
+#if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
+    if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
+        (xd->above_mi == NULL || xd->left_mi == NULL) && need_top_left[mode])
+      continue;
+#endif  // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
+
+    xd->mi[0]->uv_mode = mode;
+
+    if (!super_block_uvrd(cpi, x, &this_rate_tokenonly,
+                          &this_distortion, &s, &this_sse, bsize, best_rd))
+      continue;
+    this_rate = this_rate_tokenonly +
+        cpi->intra_uv_mode_cost[cpi->common.frame_type]
+                                [xd->mi[0]->mode][mode];
+    this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
+
+    if (this_rd < best_rd) {
+      mode_selected   = mode;
+      best_rd         = this_rd;
+      *rate           = this_rate;
+      *rate_tokenonly = this_rate_tokenonly;
+      *distortion     = this_distortion;
+      *skippable      = s;
+      if (!x->select_tx_size)
+        swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
+    }
+  }
+
+  xd->mi[0]->uv_mode = mode_selected;
+  return best_rd;
+}
+
+static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x,
+                              int *rate, int *rate_tokenonly,
+                              int64_t *distortion, int *skippable,
+                              BLOCK_SIZE bsize) {
+  const VP9_COMMON *cm = &cpi->common;
+  int64_t unused;
+
+  x->e_mbd.mi[0]->uv_mode = DC_PRED;
+  memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
+  super_block_uvrd(cpi, x, rate_tokenonly, distortion,
+                   skippable, &unused, bsize, INT64_MAX);
+  *rate = *rate_tokenonly +
+      cpi->intra_uv_mode_cost[cm->frame_type]
+                              [x->e_mbd.mi[0]->mode][DC_PRED];
+  return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
+}
+
+static void choose_intra_uv_mode(VP9_COMP *cpi, MACROBLOCK *const x,
+                                 PICK_MODE_CONTEXT *ctx,
+                                 BLOCK_SIZE bsize, TX_SIZE max_tx_size,
+                                 int *rate_uv, int *rate_uv_tokenonly,
+                                 int64_t *dist_uv, int *skip_uv,
+                                 PREDICTION_MODE *mode_uv) {
+  // Use an estimated rd for uv_intra based on DC_PRED if the
+  // appropriate speed flag is set.
+  if (cpi->sf.use_uv_intra_rd_estimate) {
+    rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv,
+                   skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
+  // Else do a proper rd search for each possible transform size that may
+  // be considered in the main rd loop.
+  } else {
+    rd_pick_intra_sbuv_mode(cpi, x, ctx,
+                            rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
+                            bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, max_tx_size);
+  }
+  *mode_uv = x->e_mbd.mi[0]->uv_mode;
+}
+
+static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
+                       int mode_context) {
+  assert(is_inter_mode(mode));
+  return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
+}
+
+static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
+                                int i,
+                                PREDICTION_MODE mode, int_mv this_mv[2],
+                                int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
+                                int_mv seg_mvs[MAX_REF_FRAMES],
+                                int_mv *best_ref_mv[2], const int *mvjcost,
+                                int *mvcost[2]) {
+  MODE_INFO *const mi = xd->mi[0];
+  const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  int thismvcost = 0;
+  int idx, idy;
+  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mi->sb_type];
+  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mi->sb_type];
+  const int is_compound = has_second_ref(mi);
+
+  switch (mode) {
+    case NEWMV:
+      this_mv[0].as_int = seg_mvs[mi->ref_frame[0]].as_int;
+      thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
+                                    mvjcost, mvcost, MV_COST_WEIGHT_SUB);
+      if (is_compound) {
+        this_mv[1].as_int = seg_mvs[mi->ref_frame[1]].as_int;
+        thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
+                                      mvjcost, mvcost, MV_COST_WEIGHT_SUB);
+      }
+      break;
+    case NEARMV:
+    case NEARESTMV:
+      this_mv[0].as_int = frame_mv[mode][mi->ref_frame[0]].as_int;
+      if (is_compound)
+        this_mv[1].as_int = frame_mv[mode][mi->ref_frame[1]].as_int;
+      break;
+    case ZEROMV:
+      this_mv[0].as_int = 0;
+      if (is_compound)
+        this_mv[1].as_int = 0;
+      break;
+    default:
+      break;
+  }
+
+  mi->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
+  if (is_compound)
+    mi->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
+
+  mi->bmi[i].as_mode = mode;
+
+  for (idy = 0; idy < num_4x4_blocks_high; ++idy)
+    for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
+      memmove(&mi->bmi[i + idy * 2 + idx], &mi->bmi[i], sizeof(mi->bmi[i]));
+
+  return cost_mv_ref(cpi, mode, mbmi_ext->mode_context[mi->ref_frame[0]]) +
+            thismvcost;
+}
+
+static int64_t encode_inter_mb_segment(VP9_COMP *cpi,
+                                       MACROBLOCK *x,
+                                       int64_t best_yrd,
+                                       int i,
+                                       int *labelyrate,
+                                       int64_t *distortion, int64_t *sse,
+                                       ENTROPY_CONTEXT *ta,
+                                       ENTROPY_CONTEXT *tl,
+                                       int mi_row, int mi_col) {
+  int k;
+  MACROBLOCKD *xd = &x->e_mbd;
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  struct macroblock_plane *const p = &x->plane[0];
+  MODE_INFO *const mi = xd->mi[0];
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->sb_type, pd);
+  const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
+  const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
+  int idx, idy;
+
+  const uint8_t *const src =
+      &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
+  uint8_t *const dst = &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i,
+                                                            pd->dst.stride)];
+  int64_t thisdistortion = 0, thissse = 0;
+  int thisrate = 0, ref;
+  const scan_order *so = &vp9_default_scan_orders[TX_4X4];
+  const int is_compound = has_second_ref(mi);
+  const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
+
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    const int bw = b_width_log2_lookup[BLOCK_8X8];
+    const int h = 4 * (i >> bw);
+    const int w = 4 * (i & ((1 << bw) - 1));
+    const struct scale_factors *sf = &xd->block_refs[ref]->sf;
+    int y_stride = pd->pre[ref].stride;
+    uint8_t *pre = pd->pre[ref].buf + (h * pd->pre[ref].stride + w);
+
+    if (vp9_is_scaled(sf)) {
+      const int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
+      const int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
+
+      y_stride = xd->block_refs[ref]->buf->y_stride;
+      pre = xd->block_refs[ref]->buf->y_buffer;
+      pre += scaled_buffer_offset(x_start + w, y_start + h,
+                                  y_stride, sf);
+    }
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vp9_highbd_build_inter_predictor(pre, y_stride,
+                                     dst, pd->dst.stride,
+                                     &mi->bmi[i].as_mv[ref].as_mv,
+                                     &xd->block_refs[ref]->sf, width, height,
+                                     ref, kernel, MV_PRECISION_Q3,
+                                     mi_col * MI_SIZE + 4 * (i % 2),
+                                     mi_row * MI_SIZE + 4 * (i / 2), xd->bd);
+  } else {
+    vp9_build_inter_predictor(pre, y_stride,
+                              dst, pd->dst.stride,
+                              &mi->bmi[i].as_mv[ref].as_mv,
+                              &xd->block_refs[ref]->sf, width, height, ref,
+                              kernel, MV_PRECISION_Q3,
+                              mi_col * MI_SIZE + 4 * (i % 2),
+                              mi_row * MI_SIZE + 4 * (i / 2));
+  }
+#else
+    vp9_build_inter_predictor(pre, y_stride,
+                              dst, pd->dst.stride,
+                              &mi->bmi[i].as_mv[ref].as_mv,
+                              &xd->block_refs[ref]->sf, width, height, ref,
+                              kernel, MV_PRECISION_Q3,
+                              mi_col * MI_SIZE + 4 * (i % 2),
+                              mi_row * MI_SIZE + 4 * (i / 2));
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vpx_highbd_subtract_block(
+        height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
+        8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
+  } else {
+    vpx_subtract_block(
+        height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
+        8, src, p->src.stride, dst, pd->dst.stride);
+  }
+#else
+  vpx_subtract_block(height, width,
+                     vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
+                     8, src, p->src.stride, dst, pd->dst.stride);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  k = i;
+  for (idy = 0; idy < height / 4; ++idy) {
+    for (idx = 0; idx < width / 4; ++idx) {
+#if CONFIG_VP9_HIGHBITDEPTH
+      const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
+#endif
+      int64_t ssz, rd, rd1, rd2;
+      tran_low_t* coeff;
+
+      k += (idy * 2 + idx);
+      coeff = BLOCK_OFFSET(p->coeff, k);
+      x->fwd_txm4x4(vp9_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
+                    coeff, 8);
+      vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
+#if CONFIG_VP9_HIGHBITDEPTH
+      thisdistortion += vp9_highbd_block_error_dispatch(
+          coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz, bd);
+#else
+      thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
+                                        16, &ssz);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      thissse += ssz;
+      thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4,
+                              so->scan, so->neighbors,
+                              cpi->sf.use_fast_coef_costing);
+      rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
+      rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
+      rd = VPXMIN(rd1, rd2);
+      if (rd >= best_yrd)
+        return INT64_MAX;
+    }
+  }
+
+  *distortion = thisdistortion >> 2;
+  *labelyrate = thisrate;
+  *sse = thissse >> 2;
+
+  return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
+}
+
+typedef struct {
+  int eobs;
+  int brate;
+  int byrate;
+  int64_t bdist;
+  int64_t bsse;
+  int64_t brdcost;
+  int_mv mvs[2];
+  ENTROPY_CONTEXT ta[2];
+  ENTROPY_CONTEXT tl[2];
+} SEG_RDSTAT;
+
+typedef struct {
+  int_mv *ref_mv[2];
+  int_mv mvp;
+
+  int64_t segment_rd;
+  int r;
+  int64_t d;
+  int64_t sse;
+  int segment_yrate;
+  PREDICTION_MODE modes[4];
+  SEG_RDSTAT rdstat[4][INTER_MODES];
+  int mvthresh;
+} BEST_SEG_INFO;
+
+static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) {
+  return (mv->row >> 3) < x->mv_row_min ||
+         (mv->row >> 3) > x->mv_row_max ||
+         (mv->col >> 3) < x->mv_col_min ||
+         (mv->col >> 3) > x->mv_col_max;
+}
+
+static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
+  MODE_INFO *const mi = x->e_mbd.mi[0];
+  struct macroblock_plane *const p = &x->plane[0];
+  struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
+
+  p->src.buf = &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i,
+                                                   p->src.stride)];
+  assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
+  pd->pre[0].buf = &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8, i,
+                                                           pd->pre[0].stride)];
+  if (has_second_ref(mi))
+    pd->pre[1].buf = &pd->pre[1].buf[vp9_raster_block_offset(BLOCK_8X8, i,
+                                                           pd->pre[1].stride)];
+}
+
+static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
+                                  struct buf_2d orig_pre[2]) {
+  MODE_INFO *mi = x->e_mbd.mi[0];
+  x->plane[0].src = orig_src;
+  x->e_mbd.plane[0].pre[0] = orig_pre[0];
+  if (has_second_ref(mi))
+    x->e_mbd.plane[0].pre[1] = orig_pre[1];
+}
+
+static INLINE int mv_has_subpel(const MV *mv) {
+  return (mv->row & 0x0F) || (mv->col & 0x0F);
+}
+
+// Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
+// TODO(aconverse): Find out if this is still productive then clean up or remove
+static int check_best_zero_mv(
+    const VP9_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES],
+    int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int this_mode,
+    const MV_REFERENCE_FRAME ref_frames[2]) {
+  if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
+      frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
+      (ref_frames[1] == NONE ||
+       frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
+    int rfc = mode_context[ref_frames[0]];
+    int c1 = cost_mv_ref(cpi, NEARMV, rfc);
+    int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
+    int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
+
+    if (this_mode == NEARMV) {
+      if (c1 > c3) return 0;
+    } else if (this_mode == NEARESTMV) {
+      if (c2 > c3) return 0;
+    } else {
+      assert(this_mode == ZEROMV);
+      if (ref_frames[1] == NONE) {
+        if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
+            (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
+          return 0;
+      } else {
+        if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
+             frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
+            (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
+             frame_mv[NEARMV][ref_frames[1]].as_int == 0))
+          return 0;
+      }
+    }
+  }
+  return 1;
+}
+
+static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
+                                BLOCK_SIZE bsize,
+                                int_mv *frame_mv,
+                                int mi_row, int mi_col,
+                                int_mv single_newmv[MAX_REF_FRAMES],
+                                int *rate_mv) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
+  const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
+  MACROBLOCKD *xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  const int refs[2] = {mi->ref_frame[0],
+                       mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1]};
+  int_mv ref_mv[2];
+  int ite, ref;
+  const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
+  struct scale_factors sf;
+
+  // Do joint motion search in compound mode to get more accurate mv.
+  struct buf_2d backup_yv12[2][MAX_MB_PLANE];
+  uint32_t last_besterr[2] = {UINT32_MAX, UINT32_MAX};
+  const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
+    vp9_get_scaled_ref_frame(cpi, mi->ref_frame[0]),
+    vp9_get_scaled_ref_frame(cpi, mi->ref_frame[1])
+  };
+
+  // Prediction buffer from second frame.
+#if CONFIG_VP9_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[64 * 64]);
+  uint8_t *second_pred;
+#else
+  DECLARE_ALIGNED(16, uint8_t, second_pred[64 * 64]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  for (ref = 0; ref < 2; ++ref) {
+    ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
+
+    if (scaled_ref_frame[ref]) {
+      int i;
+      // Swap out the reference frame for a version that's been scaled to
+      // match the resolution of the current frame, allowing the existing
+      // motion search code to be used without additional modifications.
+      for (i = 0; i < MAX_MB_PLANE; i++)
+        backup_yv12[ref][i] = xd->plane[i].pre[ref];
+      vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
+                           NULL);
+    }
+
+    frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
+  }
+
+  // Since we have scaled the reference frames to match the size of the current
+  // frame we must use a unit scaling factor during mode selection.
+#if CONFIG_VP9_HIGHBITDEPTH
+  vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height,
+                                    cm->width, cm->height,
+                                    cm->use_highbitdepth);
+#else
+  vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height,
+                                    cm->width, cm->height);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  // Allow joint search multiple times iteratively for each reference frame
+  // and break out of the search loop if it couldn't find a better mv.
+  for (ite = 0; ite < 4; ite++) {
+    struct buf_2d ref_yv12[2];
+    uint32_t bestsme = UINT32_MAX;
+    int sadpb = x->sadperbit16;
+    MV tmp_mv;
+    int search_range = 3;
+
+    int tmp_col_min = x->mv_col_min;
+    int tmp_col_max = x->mv_col_max;
+    int tmp_row_min = x->mv_row_min;
+    int tmp_row_max = x->mv_row_max;
+    int id = ite % 2;  // Even iterations search in the first reference frame,
+                       // odd iterations search in the second. The predictor
+                       // found for the 'other' reference frame is factored in.
+
+    // Initialized here because of compiler problem in Visual Studio.
+    ref_yv12[0] = xd->plane[0].pre[0];
+    ref_yv12[1] = xd->plane[0].pre[1];
+
+    // Get the prediction block from the 'other' reference frame.
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
+      vp9_highbd_build_inter_predictor(ref_yv12[!id].buf,
+                                       ref_yv12[!id].stride,
+                                       second_pred, pw,
+                                       &frame_mv[refs[!id]].as_mv,
+                                       &sf, pw, ph, 0,
+                                       kernel, MV_PRECISION_Q3,
+                                       mi_col * MI_SIZE, mi_row * MI_SIZE,
+                                       xd->bd);
+    } else {
+      second_pred = (uint8_t *)second_pred_alloc_16;
+      vp9_build_inter_predictor(ref_yv12[!id].buf,
+                                ref_yv12[!id].stride,
+                                second_pred, pw,
+                                &frame_mv[refs[!id]].as_mv,
+                                &sf, pw, ph, 0,
+                                kernel, MV_PRECISION_Q3,
+                                mi_col * MI_SIZE, mi_row * MI_SIZE);
+    }
+#else
+    vp9_build_inter_predictor(ref_yv12[!id].buf,
+                              ref_yv12[!id].stride,
+                              second_pred, pw,
+                              &frame_mv[refs[!id]].as_mv,
+                              &sf, pw, ph, 0,
+                              kernel, MV_PRECISION_Q3,
+                              mi_col * MI_SIZE, mi_row * MI_SIZE);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    // Do compound motion search on the current reference frame.
+    if (id)
+      xd->plane[0].pre[0] = ref_yv12[id];
+    vp9_set_mv_search_range(x, &ref_mv[id].as_mv);
+
+    // Use the mv result from the single mode as mv predictor.
+    tmp_mv = frame_mv[refs[id]].as_mv;
+
+    tmp_mv.col >>= 3;
+    tmp_mv.row >>= 3;
+
+    // Small-range full-pixel motion search.
+    bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
+                                       search_range,
+                                       &cpi->fn_ptr[bsize],
+                                       &ref_mv[id].as_mv, second_pred);
+    if (bestsme < UINT32_MAX)
+      bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
+                                      second_pred, &cpi->fn_ptr[bsize], 1);
+
+    x->mv_col_min = tmp_col_min;
+    x->mv_col_max = tmp_col_max;
+    x->mv_row_min = tmp_row_min;
+    x->mv_row_max = tmp_row_max;
+
+    if (bestsme < UINT32_MAX) {
+      uint32_t dis; /* TODO: use dis in distortion calculation later. */
+      uint32_t sse;
+      bestsme = cpi->find_fractional_mv_step(
+          x, &tmp_mv,
+          &ref_mv[id].as_mv,
+          cpi->common.allow_high_precision_mv,
+          x->errorperbit,
+          &cpi->fn_ptr[bsize],
+          0, cpi->sf.mv.subpel_iters_per_step,
+          NULL,
+          x->nmvjointcost, x->mvcost,
+          &dis, &sse, second_pred,
+          pw, ph);
+    }
+
+    // Restore the pointer to the first (possibly scaled) prediction buffer.
+    if (id)
+      xd->plane[0].pre[0] = ref_yv12[0];
+
+    if (bestsme < last_besterr[id]) {
+      frame_mv[refs[id]].as_mv = tmp_mv;
+      last_besterr[id] = bestsme;
+    } else {
+      break;
+    }
+  }
+
+  *rate_mv = 0;
+
+  for (ref = 0; ref < 2; ++ref) {
+    if (scaled_ref_frame[ref]) {
+      // Restore the prediction frame pointers to their unscaled versions.
+      int i;
+      for (i = 0; i < MAX_MB_PLANE; i++)
+        xd->plane[i].pre[ref] = backup_yv12[ref][i];
+    }
+
+    *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
+                                &x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
+                                x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+  }
+}
+
+static int64_t rd_pick_best_sub8x8_mode(VP9_COMP *cpi, MACROBLOCK *x,
+                                        int_mv *best_ref_mv,
+                                        int_mv *second_best_ref_mv,
+                                        int64_t best_rd, int *returntotrate,
+                                        int *returnyrate,
+                                        int64_t *returndistortion,
+                                        int *skippable, int64_t *psse,
+                                        int mvthresh,
+                                        int_mv seg_mvs[4][MAX_REF_FRAMES],
+                                        BEST_SEG_INFO *bsi_buf, int filter_idx,
+                                        int mi_row, int mi_col) {
+  int i;
+  BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
+  MACROBLOCKD *xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  int mode_idx;
+  int k, br = 0, idx, idy;
+  int64_t bd = 0, block_sse = 0;
+  PREDICTION_MODE this_mode;
+  VP9_COMMON *cm = &cpi->common;
+  struct macroblock_plane *const p = &x->plane[0];
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  const int label_count = 4;
+  int64_t this_segment_rd = 0;
+  int label_mv_thresh;
+  int segmentyrate = 0;
+  const BLOCK_SIZE bsize = mi->sb_type;
+  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
+  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
+  ENTROPY_CONTEXT t_above[2], t_left[2];
+  int subpelmv = 1, have_ref = 0;
+  SPEED_FEATURES *const sf = &cpi->sf;
+  const int has_second_rf = has_second_ref(mi);
+  const int inter_mode_mask = sf->inter_mode_mask[bsize];
+  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+
+  vp9_zero(*bsi);
+
+  bsi->segment_rd = best_rd;
+  bsi->ref_mv[0] = best_ref_mv;
+  bsi->ref_mv[1] = second_best_ref_mv;
+  bsi->mvp.as_int = best_ref_mv->as_int;
+  bsi->mvthresh = mvthresh;
+
+  for (i = 0; i < 4; i++)
+    bsi->modes[i] = ZEROMV;
+
+  memcpy(t_above, pd->above_context, sizeof(t_above));
+  memcpy(t_left, pd->left_context, sizeof(t_left));
+
+  // 64 makes this threshold really big effectively
+  // making it so that we very rarely check mvs on
+  // segments.   setting this to 1 would make mv thresh
+  // roughly equal to what it is for macroblocks
+  label_mv_thresh = 1 * bsi->mvthresh / label_count;
+
+  // Segmentation method overheads
+  for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
+    for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
+      // TODO(jingning,rbultje): rewrite the rate-distortion optimization
+      // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
+      int_mv mode_mv[MB_MODE_COUNT][2];
+      int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
+      PREDICTION_MODE mode_selected = ZEROMV;
+      int64_t best_rd = INT64_MAX;
+      const int i = idy * 2 + idx;
+      int ref;
+
+      for (ref = 0; ref < 1 + has_second_rf; ++ref) {
+        const MV_REFERENCE_FRAME frame = mi->ref_frame[ref];
+        frame_mv[ZEROMV][frame].as_int = 0;
+        vp9_append_sub8x8_mvs_for_idx(cm, xd, i, ref, mi_row, mi_col,
+                                      &frame_mv[NEARESTMV][frame],
+                                      &frame_mv[NEARMV][frame],
+                                      mbmi_ext->mode_context);
+      }
+
+      // search for the best motion vector on this segment
+      for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
+        const struct buf_2d orig_src = x->plane[0].src;
+        struct buf_2d orig_pre[2];
+
+        mode_idx = INTER_OFFSET(this_mode);
+        bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
+        if (!(inter_mode_mask & (1 << this_mode)))
+          continue;
+
+        if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
+                                this_mode, mi->ref_frame))
+          continue;
+
+        memcpy(orig_pre, pd->pre, sizeof(orig_pre));
+        memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
+               sizeof(bsi->rdstat[i][mode_idx].ta));
+        memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
+               sizeof(bsi->rdstat[i][mode_idx].tl));
+
+        // motion search for newmv (single predictor case only)
+        if (!has_second_rf && this_mode == NEWMV &&
+            seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV) {
+          MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
+          int step_param = 0;
+          uint32_t bestsme = UINT32_MAX;
+          int sadpb = x->sadperbit4;
+          MV mvp_full;
+          int max_mv;
+          int cost_list[5];
+
+          /* Is the best so far sufficiently good that we cant justify doing
+           * and new motion search. */
+          if (best_rd < label_mv_thresh)
+            break;
+
+          if (cpi->oxcf.mode != BEST) {
+            // use previous block's result as next block's MV predictor.
+            if (i > 0) {
+              bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
+              if (i == 2)
+                bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
+            }
+          }
+          if (i == 0)
+            max_mv = x->max_mv_context[mi->ref_frame[0]];
+          else
+            max_mv =
+                VPXMAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
+
+          if (sf->mv.auto_mv_step_size && cm->show_frame) {
+            // Take wtd average of the step_params based on the last frame's
+            // max mv magnitude and the best ref mvs of the current block for
+            // the given reference.
+            step_param = (vp9_init_search_range(max_mv) +
+                              cpi->mv_step_param) / 2;
+          } else {
+            step_param = cpi->mv_step_param;
+          }
+
+          mvp_full.row = bsi->mvp.as_mv.row >> 3;
+          mvp_full.col = bsi->mvp.as_mv.col >> 3;
+
+          if (sf->adaptive_motion_search) {
+            mvp_full.row = x->pred_mv[mi->ref_frame[0]].row >> 3;
+            mvp_full.col = x->pred_mv[mi->ref_frame[0]].col >> 3;
+            step_param = VPXMAX(step_param, 8);
+          }
+
+          // adjust src pointer for this block
+          mi_buf_shift(x, i);
+
+          vp9_set_mv_search_range(x, &bsi->ref_mv[0]->as_mv);
+
+          bestsme = vp9_full_pixel_search(
+              cpi, x, bsize, &mvp_full, step_param, sadpb,
+              sf->mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
+              &bsi->ref_mv[0]->as_mv, new_mv,
+              INT_MAX, 1);
+
+          if (bestsme < UINT32_MAX) {
+            uint32_t distortion;
+            cpi->find_fractional_mv_step(
+                x,
+                new_mv,
+                &bsi->ref_mv[0]->as_mv,
+                cm->allow_high_precision_mv,
+                x->errorperbit, &cpi->fn_ptr[bsize],
+                sf->mv.subpel_force_stop,
+                sf->mv.subpel_iters_per_step,
+                cond_cost_list(cpi, cost_list),
+                x->nmvjointcost, x->mvcost,
+                &distortion,
+                &x->pred_sse[mi->ref_frame[0]],
+                NULL, 0, 0);
+
+            // save motion search result for use in compound prediction
+            seg_mvs[i][mi->ref_frame[0]].as_mv = *new_mv;
+          }
+
+          if (sf->adaptive_motion_search)
+            x->pred_mv[mi->ref_frame[0]] = *new_mv;
+
+          // restore src pointers
+          mi_buf_restore(x, orig_src, orig_pre);
+        }
+
+        if (has_second_rf) {
+          if (seg_mvs[i][mi->ref_frame[1]].as_int == INVALID_MV ||
+              seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV)
+            continue;
+        }
+
+        if (has_second_rf && this_mode == NEWMV &&
+            mi->interp_filter == EIGHTTAP) {
+          // adjust src pointers
+          mi_buf_shift(x, i);
+          if (sf->comp_inter_joint_search_thresh <= bsize) {
+            int rate_mv;
+            joint_motion_search(cpi, x, bsize, frame_mv[this_mode],
+                                mi_row, mi_col, seg_mvs[i],
+                                &rate_mv);
+            seg_mvs[i][mi->ref_frame[0]].as_int =
+                frame_mv[this_mode][mi->ref_frame[0]].as_int;
+            seg_mvs[i][mi->ref_frame[1]].as_int =
+                frame_mv[this_mode][mi->ref_frame[1]].as_int;
+          }
+          // restore src pointers
+          mi_buf_restore(x, orig_src, orig_pre);
+        }
+
+        bsi->rdstat[i][mode_idx].brate =
+            set_and_cost_bmi_mvs(cpi, x, xd, i, this_mode, mode_mv[this_mode],
+                                 frame_mv, seg_mvs[i], bsi->ref_mv,
+                                 x->nmvjointcost, x->mvcost);
+
+        for (ref = 0; ref < 1 + has_second_rf; ++ref) {
+          bsi->rdstat[i][mode_idx].mvs[ref].as_int =
+              mode_mv[this_mode][ref].as_int;
+          if (num_4x4_blocks_wide > 1)
+            bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
+                mode_mv[this_mode][ref].as_int;
+          if (num_4x4_blocks_high > 1)
+            bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
+                mode_mv[this_mode][ref].as_int;
+        }
+
+        // Trap vectors that reach beyond the UMV borders
+        if (mv_check_bounds(x, &mode_mv[this_mode][0].as_mv) ||
+            (has_second_rf &&
+             mv_check_bounds(x, &mode_mv[this_mode][1].as_mv)))
+          continue;
+
+        if (filter_idx > 0) {
+          BEST_SEG_INFO *ref_bsi = bsi_buf;
+          subpelmv = 0;
+          have_ref = 1;
+
+          for (ref = 0; ref < 1 + has_second_rf; ++ref) {
+            subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
+            have_ref &= mode_mv[this_mode][ref].as_int ==
+                ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
+          }
+
+          if (filter_idx > 1 && !subpelmv && !have_ref) {
+            ref_bsi = bsi_buf + 1;
+            have_ref = 1;
+            for (ref = 0; ref < 1 + has_second_rf; ++ref)
+              have_ref &= mode_mv[this_mode][ref].as_int ==
+                  ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
+          }
+
+          if (!subpelmv && have_ref &&
+              ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
+            memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
+                   sizeof(SEG_RDSTAT));
+            if (num_4x4_blocks_wide > 1)
+              bsi->rdstat[i + 1][mode_idx].eobs =
+                  ref_bsi->rdstat[i + 1][mode_idx].eobs;
+            if (num_4x4_blocks_high > 1)
+              bsi->rdstat[i + 2][mode_idx].eobs =
+                  ref_bsi->rdstat[i + 2][mode_idx].eobs;
+
+            if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
+              mode_selected = this_mode;
+              best_rd = bsi->rdstat[i][mode_idx].brdcost;
+            }
+            continue;
+          }
+        }
+
+        bsi->rdstat[i][mode_idx].brdcost =
+            encode_inter_mb_segment(cpi, x,
+                                    bsi->segment_rd - this_segment_rd, i,
+                                    &bsi->rdstat[i][mode_idx].byrate,
+                                    &bsi->rdstat[i][mode_idx].bdist,
+                                    &bsi->rdstat[i][mode_idx].bsse,
+                                    bsi->rdstat[i][mode_idx].ta,
+                                    bsi->rdstat[i][mode_idx].tl,
+                                    mi_row, mi_col);
+        if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
+          bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv,
+                                            bsi->rdstat[i][mode_idx].brate, 0);
+          bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
+          bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
+          if (num_4x4_blocks_wide > 1)
+            bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
+          if (num_4x4_blocks_high > 1)
+            bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
+        }
+
+        if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
+          mode_selected = this_mode;
+          best_rd = bsi->rdstat[i][mode_idx].brdcost;
+        }
+      } /*for each 4x4 mode*/
+
+      if (best_rd == INT64_MAX) {
+        int iy, midx;
+        for (iy = i + 1; iy < 4; ++iy)
+          for (midx = 0; midx < INTER_MODES; ++midx)
+            bsi->rdstat[iy][midx].brdcost = INT64_MAX;
+        bsi->segment_rd = INT64_MAX;
+        return INT64_MAX;
+      }
+
+      mode_idx = INTER_OFFSET(mode_selected);
+      memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
+      memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
+
+      set_and_cost_bmi_mvs(cpi, x, xd, i, mode_selected, mode_mv[mode_selected],
+                           frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
+                           x->mvcost);
+
+      br += bsi->rdstat[i][mode_idx].brate;
+      bd += bsi->rdstat[i][mode_idx].bdist;
+      block_sse += bsi->rdstat[i][mode_idx].bsse;
+      segmentyrate += bsi->rdstat[i][mode_idx].byrate;
+      this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
+
+      if (this_segment_rd > bsi->segment_rd) {
+        int iy, midx;
+        for (iy = i + 1; iy < 4; ++iy)
+          for (midx = 0; midx < INTER_MODES; ++midx)
+            bsi->rdstat[iy][midx].brdcost = INT64_MAX;
+        bsi->segment_rd = INT64_MAX;
+        return INT64_MAX;
+      }
+    }
+  } /* for each label */
+
+  bsi->r = br;
+  bsi->d = bd;
+  bsi->segment_yrate = segmentyrate;
+  bsi->segment_rd = this_segment_rd;
+  bsi->sse = block_sse;
+
+  // update the coding decisions
+  for (k = 0; k < 4; ++k)
+    bsi->modes[k] = mi->bmi[k].as_mode;
+
+  if (bsi->segment_rd > best_rd)
+    return INT64_MAX;
+  /* set it to the best */
+  for (i = 0; i < 4; i++) {
+    mode_idx = INTER_OFFSET(bsi->modes[i]);
+    mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
+    if (has_second_ref(mi))
+      mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
+    x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
+    mi->bmi[i].as_mode = bsi->modes[i];
+  }
+
+  /*
+   * used to set mbmi->mv.as_int
+   */
+  *returntotrate = bsi->r;
+  *returndistortion = bsi->d;
+  *returnyrate = bsi->segment_yrate;
+  *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
+  *psse = bsi->sse;
+  mi->mode = bsi->modes[3];
+
+  return bsi->segment_rd;
+}
+
+static void estimate_ref_frame_costs(const VP9_COMMON *cm,
+                                     const MACROBLOCKD *xd,
+                                     int segment_id,
+                                     unsigned int *ref_costs_single,
+                                     unsigned int *ref_costs_comp,
+                                     vpx_prob *comp_mode_p) {
+  int seg_ref_active = segfeature_active(&cm->seg, segment_id,
+                                         SEG_LVL_REF_FRAME);
+  if (seg_ref_active) {
+    memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
+    memset(ref_costs_comp,   0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
+    *comp_mode_p = 128;
+  } else {
+    vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
+    vpx_prob comp_inter_p = 128;
+
+    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+      comp_inter_p = vp9_get_reference_mode_prob(cm, xd);
+      *comp_mode_p = comp_inter_p;
+    } else {
+      *comp_mode_p = 128;
+    }
+
+    ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
+
+    if (cm->reference_mode != COMPOUND_REFERENCE) {
+      vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
+      vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
+      unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
+
+      if (cm->reference_mode == REFERENCE_MODE_SELECT)
+        base_cost += vp9_cost_bit(comp_inter_p, 0);
+
+      ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
+          ref_costs_single[ALTREF_FRAME] = base_cost;
+      ref_costs_single[LAST_FRAME]   += vp9_cost_bit(ref_single_p1, 0);
+      ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
+      ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
+      ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
+      ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
+    } else {
+      ref_costs_single[LAST_FRAME]   = 512;
+      ref_costs_single[GOLDEN_FRAME] = 512;
+      ref_costs_single[ALTREF_FRAME] = 512;
+    }
+    if (cm->reference_mode != SINGLE_REFERENCE) {
+      vpx_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
+      unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
+
+      if (cm->reference_mode == REFERENCE_MODE_SELECT)
+        base_cost += vp9_cost_bit(comp_inter_p, 1);
+
+      ref_costs_comp[LAST_FRAME]   = base_cost + vp9_cost_bit(ref_comp_p, 0);
+      ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
+    } else {
+      ref_costs_comp[LAST_FRAME]   = 512;
+      ref_costs_comp[GOLDEN_FRAME] = 512;
+    }
+  }
+}
+
+static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
+                         int mode_index,
+                         int64_t comp_pred_diff[REFERENCE_MODES],
+                         int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],
+                         int skippable) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+
+  // Take a snapshot of the coding context so it can be
+  // restored if we decide to encode this way
+  ctx->skip = x->skip;
+  ctx->skippable = skippable;
+  ctx->best_mode_index = mode_index;
+  ctx->mic = *xd->mi[0];
+  ctx->mbmi_ext = *x->mbmi_ext;
+  ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
+  ctx->comp_pred_diff   = (int)comp_pred_diff[COMPOUND_REFERENCE];
+  ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
+
+  memcpy(ctx->best_filter_diff, best_filter_diff,
+         sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
+}
+
+static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
+                               MV_REFERENCE_FRAME ref_frame,
+                               BLOCK_SIZE block_size,
+                               int mi_row, int mi_col,
+                               int_mv frame_nearest_mv[MAX_REF_FRAMES],
+                               int_mv frame_near_mv[MAX_REF_FRAMES],
+                               struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
+  const VP9_COMMON *cm = &cpi->common;
+  const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
+  const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
+  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+
+  assert(yv12 != NULL);
+
+  // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
+  // use the UV scaling factors.
+  vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
+
+  // Gets an initial list of candidate vectors from neighbours and orders them
+  vp9_find_mv_refs(cm, xd, mi, ref_frame, candidates, mi_row, mi_col,
+                   mbmi_ext->mode_context);
+
+  // Candidate refinement carried out at encoder and decoder
+  vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
+                        &frame_nearest_mv[ref_frame],
+                        &frame_near_mv[ref_frame]);
+
+  // Further refinement that is encode side only to test the top few candidates
+  // in full and choose the best as the centre point for subsequent searches.
+  // The current implementation doesn't support scaling.
+  if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
+    vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
+                ref_frame, block_size);
+}
+
+static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
+                                 BLOCK_SIZE bsize,
+                                 int mi_row, int mi_col,
+                                 int_mv *tmp_mv, int *rate_mv) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  const VP9_COMMON *cm = &cpi->common;
+  MODE_INFO *mi = xd->mi[0];
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
+  int bestsme = INT_MAX;
+  int step_param;
+  int sadpb = x->sadperbit16;
+  MV mvp_full;
+  int ref = mi->ref_frame[0];
+  MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
+
+  int tmp_col_min = x->mv_col_min;
+  int tmp_col_max = x->mv_col_max;
+  int tmp_row_min = x->mv_row_min;
+  int tmp_row_max = x->mv_row_max;
+  int cost_list[5];
+
+  const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
+                                                                        ref);
+
+  MV pred_mv[3];
+  pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
+  pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
+  pred_mv[2] = x->pred_mv[ref];
+
+  if (scaled_ref_frame) {
+    int i;
+    // Swap out the reference frame for a version that's been scaled to
+    // match the resolution of the current frame, allowing the existing
+    // motion search code to be used without additional modifications.
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      backup_yv12[i] = xd->plane[i].pre[0];
+
+    vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
+  }
+
+  vp9_set_mv_search_range(x, &ref_mv);
+
+  // Work out the size of the first step in the mv step search.
+  // 0 here is maximum length first step. 1 is VPXMAX >> 1 etc.
+  if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
+    // Take wtd average of the step_params based on the last frame's
+    // max mv magnitude and that based on the best ref mvs of the current
+    // block for the given reference.
+    step_param = (vp9_init_search_range(x->max_mv_context[ref]) +
+                    cpi->mv_step_param) / 2;
+  } else {
+    step_param = cpi->mv_step_param;
+  }
+
+  if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
+    int boffset =
+        2 * (b_width_log2_lookup[BLOCK_64X64] -
+             VPXMIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
+    step_param = VPXMAX(step_param, boffset);
+  }
+
+  if (cpi->sf.adaptive_motion_search) {
+    int bwl = b_width_log2_lookup[bsize];
+    int bhl = b_height_log2_lookup[bsize];
+    int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
+
+    if (tlevel < 5)
+      step_param += 2;
+
+    // prev_mv_sad is not setup for dynamically scaled frames.
+    if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
+      int i;
+      for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
+        if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
+          x->pred_mv[ref].row = 0;
+          x->pred_mv[ref].col = 0;
+          tmp_mv->as_int = INVALID_MV;
+
+          if (scaled_ref_frame) {
+            int i;
+            for (i = 0; i < MAX_MB_PLANE; ++i)
+              xd->plane[i].pre[0] = backup_yv12[i];
+          }
+          return;
+        }
+      }
+    }
+  }
+
+  mvp_full = pred_mv[x->mv_best_ref_index[ref]];
+
+  mvp_full.col >>= 3;
+  mvp_full.row >>= 3;
+
+  bestsme = vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
+                                  cond_cost_list(cpi, cost_list),
+                                  &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
+
+  x->mv_col_min = tmp_col_min;
+  x->mv_col_max = tmp_col_max;
+  x->mv_row_min = tmp_row_min;
+  x->mv_row_max = tmp_row_max;
+
+  if (bestsme < INT_MAX) {
+    uint32_t dis;  /* TODO: use dis in distortion calculation later. */
+    cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
+                                 cm->allow_high_precision_mv,
+                                 x->errorperbit,
+                                 &cpi->fn_ptr[bsize],
+                                 cpi->sf.mv.subpel_force_stop,
+                                 cpi->sf.mv.subpel_iters_per_step,
+                                 cond_cost_list(cpi, cost_list),
+                                 x->nmvjointcost, x->mvcost,
+                                 &dis, &x->pred_sse[ref], NULL, 0, 0);
+  }
+  *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
+                             x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+
+  if (cpi->sf.adaptive_motion_search)
+    x->pred_mv[ref] = tmp_mv->as_mv;
+
+  if (scaled_ref_frame) {
+    int i;
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      xd->plane[i].pre[0] = backup_yv12[i];
+  }
+}
+
+
+
+static INLINE void restore_dst_buf(MACROBLOCKD *xd,
+                                   uint8_t *orig_dst[MAX_MB_PLANE],
+                                   int orig_dst_stride[MAX_MB_PLANE]) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    xd->plane[i].dst.buf = orig_dst[i];
+    xd->plane[i].dst.stride = orig_dst_stride[i];
+  }
+}
+
+// In some situations we want to discount tha pparent cost of a new motion
+// vector. Where there is a subtle motion field and especially where there is
+// low spatial complexity then it can be hard to cover the cost of a new motion
+// vector in a single block, even if that motion vector reduces distortion.
+// However, once established that vector may be usable through the nearest and
+// near mv modes to reduce distortion in subsequent blocks and also improve
+// visual quality.
+static int discount_newmv_test(const VP9_COMP *cpi,
+                               int this_mode,
+                               int_mv this_mv,
+                               int_mv (*mode_mv)[MAX_REF_FRAMES],
+                               int ref_frame) {
+  return (!cpi->rc.is_src_frame_alt_ref &&
+          (this_mode == NEWMV) &&
+          (this_mv.as_int != 0) &&
+          ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
+           (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
+          ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
+           (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
+}
+
+static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
+                                 BLOCK_SIZE bsize,
+                                 int *rate2, int64_t *distortion,
+                                 int *skippable,
+                                 int *rate_y, int *rate_uv,
+                                 int *disable_skip,
+                                 int_mv (*mode_mv)[MAX_REF_FRAMES],
+                                 int mi_row, int mi_col,
+                                 int_mv single_newmv[MAX_REF_FRAMES],
+                                 INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
+                                 int (*single_skippable)[MAX_REF_FRAMES],
+                                 int64_t *psse,
+                                 const int64_t ref_best_rd,
+                                 int64_t *mask_filter,
+                                 int64_t filter_cache[]) {
+  VP9_COMMON *cm = &cpi->common;
+  MACROBLOCKD *xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  const int is_comp_pred = has_second_ref(mi);
+  const int this_mode = mi->mode;
+  int_mv *frame_mv = mode_mv[this_mode];
+  int i;
+  int refs[2] = { mi->ref_frame[0],
+    (mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1]) };
+  int_mv cur_mv[2];
+#if CONFIG_VP9_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, tmp_buf16[MAX_MB_PLANE * 64 * 64]);
+  uint8_t *tmp_buf;
+#else
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf[MAX_MB_PLANE * 64 * 64]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  int pred_exists = 0;
+  int intpel_mv;
+  int64_t rd, tmp_rd, best_rd = INT64_MAX;
+  int best_needs_copy = 0;
+  uint8_t *orig_dst[MAX_MB_PLANE];
+  int orig_dst_stride[MAX_MB_PLANE];
+  int rs = 0;
+  INTERP_FILTER best_filter = SWITCHABLE;
+  uint8_t skip_txfm[MAX_MB_PLANE << 2] = {0};
+  int64_t bsse[MAX_MB_PLANE << 2] = {0};
+
+  int bsl = mi_width_log2_lookup[bsize];
+  int pred_filter_search = cpi->sf.cb_pred_filter_search ?
+      (((mi_row + mi_col) >> bsl) +
+       get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
+
+  int skip_txfm_sb = 0;
+  int64_t skip_sse_sb = INT64_MAX;
+  int64_t distortion_y = 0, distortion_uv = 0;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf16);
+  } else {
+    tmp_buf = (uint8_t *)tmp_buf16;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  if (pred_filter_search) {
+    INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
+    if (xd->above_mi)
+      af = xd->above_mi->interp_filter;
+    if (xd->left_mi)
+      lf = xd->left_mi->interp_filter;
+
+    if ((this_mode != NEWMV) || (af == lf))
+      best_filter = af;
+  }
+
+  if (is_comp_pred) {
+    if (frame_mv[refs[0]].as_int == INVALID_MV ||
+        frame_mv[refs[1]].as_int == INVALID_MV)
+      return INT64_MAX;
+
+    if (cpi->sf.adaptive_mode_search) {
+      if (single_filter[this_mode][refs[0]] ==
+          single_filter[this_mode][refs[1]])
+        best_filter = single_filter[this_mode][refs[0]];
+    }
+  }
+
+  if (this_mode == NEWMV) {
+    int rate_mv;
+    if (is_comp_pred) {
+      // Initialize mv using single prediction mode result.
+      frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
+      frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
+
+      if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
+        joint_motion_search(cpi, x, bsize, frame_mv,
+                            mi_row, mi_col, single_newmv, &rate_mv);
+      } else {
+        rate_mv  = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv,
+                                   &x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
+                                   x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+        rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv,
+                                   &x->mbmi_ext->ref_mvs[refs[1]][0].as_mv,
+                                   x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+      }
+      *rate2 += rate_mv;
+    } else {
+      int_mv tmp_mv;
+      single_motion_search(cpi, x, bsize, mi_row, mi_col,
+                           &tmp_mv, &rate_mv);
+      if (tmp_mv.as_int == INVALID_MV)
+        return INT64_MAX;
+
+      frame_mv[refs[0]].as_int =
+          xd->mi[0]->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
+      single_newmv[refs[0]].as_int = tmp_mv.as_int;
+
+      // Estimate the rate implications of a new mv but discount this
+      // under certain circumstances where we want to help initiate a weak
+      // motion field, where the distortion gain for a single block may not
+      // be enough to overcome the cost of a new mv.
+      if (discount_newmv_test(cpi, this_mode, tmp_mv, mode_mv, refs[0])) {
+        *rate2 += VPXMAX((rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
+      } else {
+        *rate2 += rate_mv;
+      }
+    }
+  }
+
+  for (i = 0; i < is_comp_pred + 1; ++i) {
+    cur_mv[i] = frame_mv[refs[i]];
+    // Clip "next_nearest" so that it does not extend to far out of image
+    if (this_mode != NEWMV)
+      clamp_mv2(&cur_mv[i].as_mv, xd);
+
+    if (mv_check_bounds(x, &cur_mv[i].as_mv))
+      return INT64_MAX;
+    mi->mv[i].as_int = cur_mv[i].as_int;
+  }
+
+  // do first prediction into the destination buffer. Do the next
+  // prediction into a temporary buffer. Then keep track of which one
+  // of these currently holds the best predictor, and use the other
+  // one for future predictions. In the end, copy from tmp_buf to
+  // dst if necessary.
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    orig_dst[i] = xd->plane[i].dst.buf;
+    orig_dst_stride[i] = xd->plane[i].dst.stride;
+  }
+
+  // We don't include the cost of the second reference here, because there
+  // are only two options: Last/ARF or Golden/ARF; The second one is always
+  // known, which is ARF.
+  //
+  // Under some circumstances we discount the cost of new mv mode to encourage
+  // initiation of a motion field.
+  if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]],
+                          mode_mv, refs[0])) {
+    *rate2 += VPXMIN(cost_mv_ref(cpi, this_mode,
+                                 mbmi_ext->mode_context[refs[0]]),
+                     cost_mv_ref(cpi, NEARESTMV,
+                                 mbmi_ext->mode_context[refs[0]]));
+  } else {
+    *rate2 += cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]);
+  }
+
+  if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
+      mi->mode != NEARESTMV)
+    return INT64_MAX;
+
+  pred_exists = 0;
+  // Are all MVs integer pel for Y and UV
+  intpel_mv = !mv_has_subpel(&mi->mv[0].as_mv);
+  if (is_comp_pred)
+    intpel_mv &= !mv_has_subpel(&mi->mv[1].as_mv);
+
+  // Search for best switchable filter by checking the variance of
+  // pred error irrespective of whether the filter will be used
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+    filter_cache[i] = INT64_MAX;
+
+  if (cm->interp_filter != BILINEAR) {
+    if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
+      best_filter = EIGHTTAP;
+    } else if (best_filter == SWITCHABLE) {
+      int newbest;
+      int tmp_rate_sum = 0;
+      int64_t tmp_dist_sum = 0;
+
+      for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
+        int j;
+        int64_t rs_rd;
+        int tmp_skip_sb = 0;
+        int64_t tmp_skip_sse = INT64_MAX;
+
+        mi->interp_filter = i;
+        rs = vp9_get_switchable_rate(cpi, xd);
+        rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
+
+        if (i > 0 && intpel_mv) {
+          rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
+          filter_cache[i] = rd;
+          filter_cache[SWITCHABLE_FILTERS] =
+              VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
+          if (cm->interp_filter == SWITCHABLE)
+            rd += rs_rd;
+          *mask_filter = VPXMAX(*mask_filter, rd);
+        } else {
+          int rate_sum = 0;
+          int64_t dist_sum = 0;
+          if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
+              (cpi->sf.interp_filter_search_mask & (1 << i))) {
+            rate_sum = INT_MAX;
+            dist_sum = INT64_MAX;
+            continue;
+          }
+
+          if ((cm->interp_filter == SWITCHABLE &&
+               (!i || best_needs_copy)) ||
+              (cm->interp_filter != SWITCHABLE &&
+               (cm->interp_filter == mi->interp_filter ||
+                (i == 0 && intpel_mv)))) {
+            restore_dst_buf(xd, orig_dst, orig_dst_stride);
+          } else {
+            for (j = 0; j < MAX_MB_PLANE; j++) {
+              xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
+              xd->plane[j].dst.stride = 64;
+            }
+          }
+          vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
+          model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum,
+                          &tmp_skip_sb, &tmp_skip_sse);
+
+          rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
+          filter_cache[i] = rd;
+          filter_cache[SWITCHABLE_FILTERS] =
+              VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
+          if (cm->interp_filter == SWITCHABLE)
+            rd += rs_rd;
+          *mask_filter = VPXMAX(*mask_filter, rd);
+
+          if (i == 0 && intpel_mv) {
+            tmp_rate_sum = rate_sum;
+            tmp_dist_sum = dist_sum;
+          }
+        }
+
+        if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
+          if (rd / 2 > ref_best_rd) {
+            restore_dst_buf(xd, orig_dst, orig_dst_stride);
+            return INT64_MAX;
+          }
+        }
+        newbest = i == 0 || rd < best_rd;
+
+        if (newbest) {
+          best_rd = rd;
+          best_filter = mi->interp_filter;
+          if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
+            best_needs_copy = !best_needs_copy;
+        }
+
+        if ((cm->interp_filter == SWITCHABLE && newbest) ||
+            (cm->interp_filter != SWITCHABLE &&
+             cm->interp_filter == mi->interp_filter)) {
+          pred_exists = 1;
+          tmp_rd = best_rd;
+
+          skip_txfm_sb = tmp_skip_sb;
+          skip_sse_sb = tmp_skip_sse;
+          memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
+          memcpy(bsse, x->bsse, sizeof(bsse));
+        }
+      }
+      restore_dst_buf(xd, orig_dst, orig_dst_stride);
+    }
+  }
+  // Set the appropriate filter
+  mi->interp_filter = cm->interp_filter != SWITCHABLE ?
+      cm->interp_filter : best_filter;
+  rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi, xd) : 0;
+
+  if (pred_exists) {
+    if (best_needs_copy) {
+      // again temporarily set the buffers to local memory to prevent a memcpy
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
+        xd->plane[i].dst.stride = 64;
+      }
+    }
+    rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
+  } else {
+    int tmp_rate;
+    int64_t tmp_dist;
+    // Handles the special case when a filter that is not in the
+    // switchable list (ex. bilinear) is indicated at the frame level, or
+    // skip condition holds.
+    vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
+    model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist,
+                    &skip_txfm_sb, &skip_sse_sb);
+    rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
+    memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
+    memcpy(bsse, x->bsse, sizeof(bsse));
+  }
+
+  if (!is_comp_pred)
+    single_filter[this_mode][refs[0]] = mi->interp_filter;
+
+  if (cpi->sf.adaptive_mode_search)
+    if (is_comp_pred)
+      if (single_skippable[this_mode][refs[0]] &&
+          single_skippable[this_mode][refs[1]])
+        memset(skip_txfm, SKIP_TXFM_AC_DC, sizeof(skip_txfm));
+
+  if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
+    // if current pred_error modeled rd is substantially more than the best
+    // so far, do not bother doing full rd
+    if (rd / 2 > ref_best_rd) {
+      restore_dst_buf(xd, orig_dst, orig_dst_stride);
+      return INT64_MAX;
+    }
+  }
+
+  if (cm->interp_filter == SWITCHABLE)
+    *rate2 += rs;
+
+  memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
+  memcpy(x->bsse, bsse, sizeof(bsse));
+
+  if (!skip_txfm_sb) {
+    int skippable_y, skippable_uv;
+    int64_t sseuv = INT64_MAX;
+    int64_t rdcosty = INT64_MAX;
+
+    // Y cost and distortion
+    vp9_subtract_plane(x, bsize, 0);
+    super_block_yrd(cpi, x, rate_y, &distortion_y, &skippable_y, psse,
+                    bsize, ref_best_rd);
+
+    if (*rate_y == INT_MAX) {
+      *rate2 = INT_MAX;
+      *distortion = INT64_MAX;
+      restore_dst_buf(xd, orig_dst, orig_dst_stride);
+      return INT64_MAX;
+    }
+
+    *rate2 += *rate_y;
+    *distortion += distortion_y;
+
+    rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
+    rdcosty = VPXMIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
+
+    if (!super_block_uvrd(cpi, x, rate_uv, &distortion_uv, &skippable_uv,
+                          &sseuv, bsize, ref_best_rd - rdcosty)) {
+      *rate2 = INT_MAX;
+      *distortion = INT64_MAX;
+      restore_dst_buf(xd, orig_dst, orig_dst_stride);
+      return INT64_MAX;
+    }
+
+    *psse += sseuv;
+    *rate2 += *rate_uv;
+    *distortion += distortion_uv;
+    *skippable = skippable_y && skippable_uv;
+  } else {
+    x->skip = 1;
+    *disable_skip = 1;
+
+    // The cost of skip bit needs to be added.
+    *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+
+    *distortion = skip_sse_sb;
+  }
+
+  if (!is_comp_pred)
+    single_skippable[this_mode][refs[0]] = *skippable;
+
+  restore_dst_buf(xd, orig_dst, orig_dst_stride);
+  return 0;  // The rate-distortion cost will be re-calculated by caller.
+}
+
+void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
+                               RD_COST *rd_cost, BLOCK_SIZE bsize,
+                               PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblockd_plane *const pd = xd->plane;
+  int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
+  int y_skip = 0, uv_skip = 0;
+  int64_t dist_y = 0, dist_uv = 0;
+  TX_SIZE max_uv_tx_size;
+  x->skip_encode = 0;
+  ctx->skip = 0;
+  xd->mi[0]->ref_frame[0] = INTRA_FRAME;
+  xd->mi[0]->ref_frame[1] = NONE;
+
+  if (bsize >= BLOCK_8X8) {
+    if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
+                               &dist_y, &y_skip, bsize,
+                               best_rd) >= best_rd) {
+      rd_cost->rate = INT_MAX;
+      return;
+    }
+  } else {
+    y_skip = 0;
+    if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
+                                     &dist_y, best_rd) >= best_rd) {
+      rd_cost->rate = INT_MAX;
+      return;
+    }
+  }
+  max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->tx_size, bsize,
+                                       pd[1].subsampling_x,
+                                       pd[1].subsampling_y);
+  rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
+                          &dist_uv, &uv_skip, VPXMAX(BLOCK_8X8, bsize),
+                          max_uv_tx_size);
+
+  if (y_skip && uv_skip) {
+    rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
+                    vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+    rd_cost->dist = dist_y + dist_uv;
+  } else {
+    rd_cost->rate = rate_y + rate_uv +
+                      vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
+    rd_cost->dist = dist_y + dist_uv;
+  }
+
+  ctx->mic = *xd->mi[0];
+  ctx->mbmi_ext = *x->mbmi_ext;
+  rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
+}
+
+// This function is designed to apply a bias or adjustment to an rd value based
+// on the relative variance of the source and reconstruction.
+#define LOW_VAR_THRESH 16
+#define VLOW_ADJ_MAX 25
+#define VHIGH_ADJ_MAX 8
+static void rd_variance_adjustment(VP9_COMP *cpi,
+                                   MACROBLOCK *x,
+                                   BLOCK_SIZE bsize,
+                                   int64_t *this_rd,
+                                   MV_REFERENCE_FRAME ref_frame,
+                                   unsigned int source_variance) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  unsigned int recon_variance;
+  unsigned int absvar_diff = 0;
+  int64_t var_error = 0;
+  int64_t var_factor = 0;
+
+  if (*this_rd == INT64_MAX)
+    return;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    recon_variance =
+      vp9_high_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize, xd->bd);
+  } else {
+    recon_variance =
+      vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
+  }
+#else
+  recon_variance =
+    vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  if ((source_variance + recon_variance) > LOW_VAR_THRESH) {
+    absvar_diff = (source_variance > recon_variance)
+      ? (source_variance - recon_variance)
+      : (recon_variance - source_variance);
+
+    var_error = ((int64_t)200 * source_variance * recon_variance) /
+      (((int64_t)source_variance * source_variance) +
+       ((int64_t)recon_variance * recon_variance));
+    var_error = 100 - var_error;
+  }
+
+  // Source variance above a threshold and ref frame is intra.
+  // This case is targeted mainly at discouraging intra modes that give rise
+  // to a predictor with a low spatial complexity compared to the source.
+  if ((source_variance > LOW_VAR_THRESH) && (ref_frame == INTRA_FRAME) &&
+      (source_variance > recon_variance)) {
+    var_factor = VPXMIN(absvar_diff, VPXMIN(VLOW_ADJ_MAX, var_error));
+  // A second possible case of interest is where the source variance
+  // is very low and we wish to discourage false texture or motion trails.
+  } else if ((source_variance < (LOW_VAR_THRESH >> 1)) &&
+             (recon_variance > source_variance)) {
+    var_factor = VPXMIN(absvar_diff, VPXMIN(VHIGH_ADJ_MAX, var_error));
+  }
+  *this_rd += (*this_rd * var_factor) / 100;
+}
+
+
+// Do we have an internal image edge (e.g. formatting bars).
+int vp9_internal_image_edge(VP9_COMP *cpi) {
+  return (cpi->oxcf.pass == 2) &&
+    ((cpi->twopass.this_frame_stats.inactive_zone_rows > 0) ||
+    (cpi->twopass.this_frame_stats.inactive_zone_cols > 0));
+}
+
+// Checks to see if a super block is on a horizontal image edge.
+// In most cases this is the "real" edge unless there are formatting
+// bars embedded in the stream.
+int vp9_active_h_edge(VP9_COMP *cpi, int mi_row, int mi_step) {
+  int top_edge = 0;
+  int bottom_edge = cpi->common.mi_rows;
+  int is_active_h_edge = 0;
+
+  // For two pass account for any formatting bars detected.
+  if (cpi->oxcf.pass == 2) {
+    TWO_PASS *twopass = &cpi->twopass;
+
+    // The inactive region is specified in MBs not mi units.
+    // The image edge is in the following MB row.
+    top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
+
+    bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
+    bottom_edge = VPXMAX(top_edge, bottom_edge);
+  }
+
+  if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
+      ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
+    is_active_h_edge = 1;
+  }
+  return is_active_h_edge;
+}
+
+// Checks to see if a super block is on a vertical image edge.
+// In most cases this is the "real" edge unless there are formatting
+// bars embedded in the stream.
+int vp9_active_v_edge(VP9_COMP *cpi, int mi_col, int mi_step) {
+  int left_edge = 0;
+  int right_edge = cpi->common.mi_cols;
+  int is_active_v_edge = 0;
+
+  // For two pass account for any formatting bars detected.
+  if (cpi->oxcf.pass == 2) {
+    TWO_PASS *twopass = &cpi->twopass;
+
+    // The inactive region is specified in MBs not mi units.
+    // The image edge is in the following MB row.
+    left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
+
+    right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
+    right_edge = VPXMAX(left_edge, right_edge);
+  }
+
+  if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
+      ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
+    is_active_v_edge = 1;
+  }
+  return is_active_v_edge;
+}
+
+// Checks to see if a super block is at the edge of the active image.
+// In most cases this is the "real" edge unless there are formatting
+// bars embedded in the stream.
+int vp9_active_edge_sb(VP9_COMP *cpi,
+                       int mi_row, int mi_col) {
+  return vp9_active_h_edge(cpi, mi_row, MI_BLOCK_SIZE) ||
+         vp9_active_v_edge(cpi, mi_col, MI_BLOCK_SIZE);
+}
+
+void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi,
+                               TileDataEnc *tile_data,
+                               MACROBLOCK *x,
+                               int mi_row, int mi_col,
+                               RD_COST *rd_cost, BLOCK_SIZE bsize,
+                               PICK_MODE_CONTEXT *ctx,
+                               int64_t best_rd_so_far) {
+  VP9_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  RD_OPT *const rd_opt = &cpi->rd;
+  SPEED_FEATURES *const sf = &cpi->sf;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  const struct segmentation *const seg = &cm->seg;
+  PREDICTION_MODE this_mode;
+  MV_REFERENCE_FRAME ref_frame, second_ref_frame;
+  unsigned char segment_id = mi->segment_id;
+  int comp_pred, i, k;
+  int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
+  struct buf_2d yv12_mb[4][MAX_MB_PLANE];
+  int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
+  INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
+  int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
+  static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
+                                    VP9_ALT_FLAG };
+  int64_t best_rd = best_rd_so_far;
+  int64_t best_pred_diff[REFERENCE_MODES];
+  int64_t best_pred_rd[REFERENCE_MODES];
+  int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
+  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+  MODE_INFO best_mbmode;
+  int best_mode_skippable = 0;
+  int midx, best_mode_index = -1;
+  unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
+  vpx_prob comp_mode_p;
+  int64_t best_intra_rd = INT64_MAX;
+  unsigned int best_pred_sse = UINT_MAX;
+  PREDICTION_MODE best_intra_mode = DC_PRED;
+  int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
+  int64_t dist_uv[TX_SIZES];
+  int skip_uv[TX_SIZES];
+  PREDICTION_MODE mode_uv[TX_SIZES];
+  const int intra_cost_penalty = vp9_get_intra_cost_penalty(
+      cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
+  int best_skip2 = 0;
+  uint8_t ref_frame_skip_mask[2] = { 0 };
+  uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
+  int mode_skip_start = sf->mode_skip_start + 1;
+  const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
+  const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
+  int64_t mode_threshold[MAX_MODES];
+  int *mode_map = tile_data->mode_map[bsize];
+  const int mode_search_skip_flags = sf->mode_search_skip_flags;
+  int64_t mask_filter = 0;
+  int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
+
+  vp9_zero(best_mbmode);
+
+  x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
+
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+    filter_cache[i] = INT64_MAX;
+
+  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
+                           &comp_mode_p);
+
+  for (i = 0; i < REFERENCE_MODES; ++i)
+    best_pred_rd[i] = INT64_MAX;
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+    best_filter_rd[i] = INT64_MAX;
+  for (i = 0; i < TX_SIZES; i++)
+    rate_uv_intra[i] = INT_MAX;
+  for (i = 0; i < MAX_REF_FRAMES; ++i)
+    x->pred_sse[i] = INT_MAX;
+  for (i = 0; i < MB_MODE_COUNT; ++i) {
+    for (k = 0; k < MAX_REF_FRAMES; ++k) {
+      single_inter_filter[i][k] = SWITCHABLE;
+      single_skippable[i][k] = 0;
+    }
+  }
+
+  rd_cost->rate = INT_MAX;
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    x->pred_mv_sad[ref_frame] = INT_MAX;
+    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
+      assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
+      setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
+                         frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
+    }
+    frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
+    frame_mv[ZEROMV][ref_frame].as_int = 0;
+  }
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
+      // Skip checking missing references in both single and compound reference
+      // modes. Note that a mode will be skipped if both reference frames
+      // are masked out.
+      ref_frame_skip_mask[0] |= (1 << ref_frame);
+      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+    } else if (sf->reference_masking) {
+      for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
+        // Skip fixed mv modes for poor references
+        if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
+          mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
+          break;
+        }
+      }
+    }
+    // If the segment reference frame feature is enabled....
+    // then do nothing if the current ref frame is not allowed..
+    if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
+        get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
+      ref_frame_skip_mask[0] |= (1 << ref_frame);
+      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+    }
+  }
+
+  // Disable this drop out case if the ref frame
+  // segment level feature is enabled for this segment. This is to
+  // prevent the possibility that we end up unable to pick any mode.
+  if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
+    // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
+    // unless ARNR filtering is enabled in which case we want
+    // an unfiltered alternative. We allow near/nearest as well
+    // because they may result in zero-zero MVs but be cheaper.
+    if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
+      ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
+      ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
+      mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
+      if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
+        mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
+      if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
+        mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
+    }
+  }
+
+  if (cpi->rc.is_src_frame_alt_ref) {
+    if (sf->alt_ref_search_fp) {
+      mode_skip_mask[ALTREF_FRAME] = 0;
+      ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
+      ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
+    }
+  }
+
+  if (sf->alt_ref_search_fp)
+    if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
+      if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
+        mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
+
+  if (sf->adaptive_mode_search) {
+    if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
+        cpi->rc.frames_since_golden >= 3)
+      if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
+        mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
+  }
+
+  if (bsize > sf->max_intra_bsize) {
+    ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
+    ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
+  }
+
+  mode_skip_mask[INTRA_FRAME] |=
+      ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
+
+  for (i = 0; i <= LAST_NEW_MV_INDEX; ++i)
+    mode_threshold[i] = 0;
+  for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
+    mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
+
+  midx =  sf->schedule_mode_search ? mode_skip_start : 0;
+  while (midx > 4) {
+    uint8_t end_pos = 0;
+    for (i = 5; i < midx; ++i) {
+      if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
+        uint8_t tmp = mode_map[i];
+        mode_map[i] = mode_map[i - 1];
+        mode_map[i - 1] = tmp;
+        end_pos = i;
+      }
+    }
+    midx = end_pos;
+  }
+
+  for (midx = 0; midx < MAX_MODES; ++midx) {
+    int mode_index = mode_map[midx];
+    int mode_excluded = 0;
+    int64_t this_rd = INT64_MAX;
+    int disable_skip = 0;
+    int compmode_cost = 0;
+    int rate2 = 0, rate_y = 0, rate_uv = 0;
+    int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
+    int skippable = 0;
+    int this_skip2 = 0;
+    int64_t total_sse = INT64_MAX;
+    int early_term = 0;
+
+    this_mode = vp9_mode_order[mode_index].mode;
+    ref_frame = vp9_mode_order[mode_index].ref_frame[0];
+    second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
+
+    // Look at the reference frame of the best mode so far and set the
+    // skip mask to look at a subset of the remaining modes.
+    if (midx == mode_skip_start && best_mode_index >= 0) {
+      switch (best_mbmode.ref_frame[0]) {
+        case INTRA_FRAME:
+          break;
+        case LAST_FRAME:
+          ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
+          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+          break;
+        case GOLDEN_FRAME:
+          ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
+          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+          break;
+        case ALTREF_FRAME:
+          ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK;
+          break;
+        case NONE:
+        case MAX_REF_FRAMES:
+          assert(0 && "Invalid Reference frame");
+          break;
+      }
+    }
+
+    if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
+        (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
+      continue;
+
+    if (mode_skip_mask[ref_frame] & (1 << this_mode))
+      continue;
+
+    // Test best rd so far against threshold for trying this mode.
+    if (best_mode_skippable && sf->schedule_mode_search)
+      mode_threshold[mode_index] <<= 1;
+
+    if (best_rd < mode_threshold[mode_index])
+      continue;
+
+    if (sf->motion_field_mode_search) {
+      const int mi_width  = VPXMIN(num_8x8_blocks_wide_lookup[bsize],
+                                   tile_info->mi_col_end - mi_col);
+      const int mi_height = VPXMIN(num_8x8_blocks_high_lookup[bsize],
+                                   tile_info->mi_row_end - mi_row);
+      const int bsl = mi_width_log2_lookup[bsize];
+      int cb_partition_search_ctrl = (((mi_row + mi_col) >> bsl)
+          + get_chessboard_index(cm->current_video_frame)) & 0x1;
+      MODE_INFO *ref_mi;
+      int const_motion = 1;
+      int skip_ref_frame = !cb_partition_search_ctrl;
+      MV_REFERENCE_FRAME rf = NONE;
+      int_mv ref_mv;
+      ref_mv.as_int = INVALID_MV;
+
+      if ((mi_row - 1) >= tile_info->mi_row_start) {
+        ref_mv = xd->mi[-xd->mi_stride]->mv[0];
+        rf = xd->mi[-xd->mi_stride]->ref_frame[0];
+        for (i = 0; i < mi_width; ++i) {
+          ref_mi = xd->mi[-xd->mi_stride + i];
+          const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
+                          (ref_frame == ref_mi->ref_frame[0]);
+          skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
+        }
+      }
+
+      if ((mi_col - 1) >= tile_info->mi_col_start) {
+        if (ref_mv.as_int == INVALID_MV)
+          ref_mv = xd->mi[-1]->mv[0];
+        if (rf == NONE)
+          rf = xd->mi[-1]->ref_frame[0];
+        for (i = 0; i < mi_height; ++i) {
+          ref_mi = xd->mi[i * xd->mi_stride - 1];
+          const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
+                          (ref_frame == ref_mi->ref_frame[0]);
+          skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
+        }
+      }
+
+      if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
+        if (rf > INTRA_FRAME)
+          if (ref_frame != rf)
+            continue;
+
+      if (const_motion)
+        if (this_mode == NEARMV || this_mode == ZEROMV)
+          continue;
+    }
+
+    comp_pred = second_ref_frame > INTRA_FRAME;
+    if (comp_pred) {
+      if (!cpi->allow_comp_inter_inter)
+        continue;
+
+      // Skip compound inter modes if ARF is not available.
+      if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
+        continue;
+
+      // Do not allow compound prediction if the segment level reference frame
+      // feature is in use as in this case there can only be one reference.
+      if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
+        continue;
+
+      if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
+          best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
+        continue;
+
+      mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
+    } else {
+      if (ref_frame != INTRA_FRAME)
+        mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
+    }
+
+    if (ref_frame == INTRA_FRAME) {
+      if (sf->adaptive_mode_search)
+        if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
+          continue;
+
+      if (this_mode != DC_PRED) {
+        // Disable intra modes other than DC_PRED for blocks with low variance
+        // Threshold for intra skipping based on source variance
+        // TODO(debargha): Specialize the threshold for super block sizes
+        const unsigned int skip_intra_var_thresh = 64;
+        if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
+            x->source_variance < skip_intra_var_thresh)
+          continue;
+        // Only search the oblique modes if the best so far is
+        // one of the neighboring directional modes
+        if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
+            (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
+          if (best_mode_index >= 0 &&
+              best_mbmode.ref_frame[0] > INTRA_FRAME)
+            continue;
+        }
+        if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
+          if (conditional_skipintra(this_mode, best_intra_mode))
+              continue;
+        }
+      }
+    } else {
+      const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame};
+      if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
+                              this_mode, ref_frames))
+        continue;
+    }
+
+    mi->mode = this_mode;
+    mi->uv_mode = DC_PRED;
+    mi->ref_frame[0] = ref_frame;
+    mi->ref_frame[1] = second_ref_frame;
+    // Evaluate all sub-pel filters irrespective of whether we can use
+    // them for this frame.
+    mi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
+                                                          : cm->interp_filter;
+    mi->mv[0].as_int = mi->mv[1].as_int = 0;
+
+    x->skip = 0;
+    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
+
+    // Select prediction reference frames.
+    for (i = 0; i < MAX_MB_PLANE; i++) {
+      xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
+      if (comp_pred)
+        xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
+    }
+
+    if (ref_frame == INTRA_FRAME) {
+      TX_SIZE uv_tx;
+      struct macroblockd_plane *const pd = &xd->plane[1];
+      memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
+      super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable,
+                      NULL, bsize, best_rd);
+      if (rate_y == INT_MAX)
+        continue;
+
+      uv_tx = get_uv_tx_size_impl(mi->tx_size, bsize, pd->subsampling_x,
+                                  pd->subsampling_y);
+      if (rate_uv_intra[uv_tx] == INT_MAX) {
+        choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx,
+                             &rate_uv_intra[uv_tx], &rate_uv_tokenonly[uv_tx],
+                             &dist_uv[uv_tx], &skip_uv[uv_tx], &mode_uv[uv_tx]);
+      }
+
+      rate_uv = rate_uv_tokenonly[uv_tx];
+      distortion_uv = dist_uv[uv_tx];
+      skippable = skippable && skip_uv[uv_tx];
+      mi->uv_mode = mode_uv[uv_tx];
+
+      rate2 = rate_y + cpi->mbmode_cost[mi->mode] + rate_uv_intra[uv_tx];
+      if (this_mode != DC_PRED && this_mode != TM_PRED)
+        rate2 += intra_cost_penalty;
+      distortion2 = distortion_y + distortion_uv;
+    } else {
+      this_rd = handle_inter_mode(cpi, x, bsize,
+                                  &rate2, &distortion2, &skippable,
+                                  &rate_y, &rate_uv,
+                                  &disable_skip, frame_mv,
+                                  mi_row, mi_col,
+                                  single_newmv, single_inter_filter,
+                                  single_skippable, &total_sse, best_rd,
+                                  &mask_filter, filter_cache);
+      if (this_rd == INT64_MAX)
+        continue;
+
+      compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
+
+      if (cm->reference_mode == REFERENCE_MODE_SELECT)
+        rate2 += compmode_cost;
+    }
+
+    // Estimate the reference frame signaling cost and add it
+    // to the rolling cost variable.
+    if (comp_pred) {
+      rate2 += ref_costs_comp[ref_frame];
+    } else {
+      rate2 += ref_costs_single[ref_frame];
+    }
+
+    if (!disable_skip) {
+      const vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
+      const int skip_cost0 = vp9_cost_bit(skip_prob, 0);
+      const int skip_cost1 = vp9_cost_bit(skip_prob, 1);
+
+      if (skippable) {
+        // Back out the coefficient coding costs
+        rate2 -= (rate_y + rate_uv);
+
+        // Cost the skip mb case
+        rate2 += skip_cost1;
+      } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
+        if (RDCOST(x->rdmult, x->rddiv,
+                   rate_y + rate_uv + skip_cost0, distortion2) <
+            RDCOST(x->rdmult, x->rddiv, skip_cost1, total_sse)) {
+          // Add in the cost of the no skip flag.
+          rate2 += skip_cost0;
+        } else {
+          // FIXME(rbultje) make this work for splitmv also
+          assert(total_sse >= 0);
+
+          rate2 += skip_cost1;
+          distortion2 = total_sse;
+          rate2 -= (rate_y + rate_uv);
+          this_skip2 = 1;
+        }
+      } else {
+        // Add in the cost of the no skip flag.
+        rate2 += skip_cost0;
+      }
+
+      // Calculate the final RD estimate for this mode.
+      this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+    }
+
+    // Apply an adjustment to the rd value based on the similarity of the
+    // source variance and reconstructed variance.
+    rd_variance_adjustment(cpi, x, bsize, &this_rd,
+                           ref_frame, x->source_variance);
+
+    if (ref_frame == INTRA_FRAME) {
+    // Keep record of best intra rd
+      if (this_rd < best_intra_rd) {
+        best_intra_rd = this_rd;
+        best_intra_mode = mi->mode;
+      }
+    }
+
+    if (!disable_skip && ref_frame == INTRA_FRAME) {
+      for (i = 0; i < REFERENCE_MODES; ++i)
+        best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
+      for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+        best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
+    }
+
+    // Did this mode help.. i.e. is it the new best mode
+    if (this_rd < best_rd || x->skip) {
+      int max_plane = MAX_MB_PLANE;
+      if (!mode_excluded) {
+        // Note index of best mode so far
+        best_mode_index = mode_index;
+
+        if (ref_frame == INTRA_FRAME) {
+          /* required for left and above block mv */
+          mi->mv[0].as_int = 0;
+          max_plane = 1;
+        } else {
+          best_pred_sse = x->pred_sse[ref_frame];
+        }
+
+        rd_cost->rate = rate2;
+        rd_cost->dist = distortion2;
+        rd_cost->rdcost = this_rd;
+        best_rd = this_rd;
+        best_mbmode = *mi;
+        best_skip2 = this_skip2;
+        best_mode_skippable = skippable;
+
+        if (!x->select_tx_size)
+          swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
+        memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mi->tx_size],
+               sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
+
+        // TODO(debargha): enhance this test with a better distortion prediction
+        // based on qp, activity mask and history
+        if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
+            (mode_index > MIN_EARLY_TERM_INDEX)) {
+          int qstep = xd->plane[0].dequant[1];
+          // TODO(debargha): Enhance this by specializing for each mode_index
+          int scale = 4;
+#if CONFIG_VP9_HIGHBITDEPTH
+          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            qstep >>= (xd->bd - 8);
+          }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+          if (x->source_variance < UINT_MAX) {
+            const int var_adjust = (x->source_variance < 16);
+            scale -= var_adjust;
+          }
+          if (ref_frame > INTRA_FRAME &&
+              distortion2 * scale < qstep * qstep) {
+            early_term = 1;
+          }
+        }
+      }
+    }
+
+    /* keep record of best compound/single-only prediction */
+    if (!disable_skip && ref_frame != INTRA_FRAME) {
+      int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
+
+      if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+        single_rate = rate2 - compmode_cost;
+        hybrid_rate = rate2;
+      } else {
+        single_rate = rate2;
+        hybrid_rate = rate2 + compmode_cost;
+      }
+
+      single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
+      hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
+
+      if (!comp_pred) {
+        if (single_rd < best_pred_rd[SINGLE_REFERENCE])
+          best_pred_rd[SINGLE_REFERENCE] = single_rd;
+      } else {
+        if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
+          best_pred_rd[COMPOUND_REFERENCE] = single_rd;
+      }
+      if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
+        best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
+
+      /* keep record of best filter type */
+      if (!mode_excluded && cm->interp_filter != BILINEAR) {
+        int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
+                              SWITCHABLE_FILTERS : cm->interp_filter];
+
+        for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
+          int64_t adj_rd;
+          if (ref == INT64_MAX)
+            adj_rd = 0;
+          else if (filter_cache[i] == INT64_MAX)
+            // when early termination is triggered, the encoder does not have
+            // access to the rate-distortion cost. it only knows that the cost
+            // should be above the maximum valid value. hence it takes the known
+            // maximum plus an arbitrary constant as the rate-distortion cost.
+            adj_rd = mask_filter - ref + 10;
+          else
+            adj_rd = filter_cache[i] - ref;
+
+          adj_rd += this_rd;
+          best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
+        }
+      }
+    }
+
+    if (early_term)
+      break;
+
+    if (x->skip && !comp_pred)
+      break;
+  }
+
+  // The inter modes' rate costs are not calculated precisely in some cases.
+  // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
+  // ZEROMV. Here, checks are added for those cases, and the mode decisions
+  // are corrected.
+  if (best_mbmode.mode == NEWMV) {
+    const MV_REFERENCE_FRAME refs[2] = {best_mbmode.ref_frame[0],
+        best_mbmode.ref_frame[1]};
+    int comp_pred_mode = refs[1] > INTRA_FRAME;
+
+    if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
+        ((comp_pred_mode && frame_mv[NEARESTMV][refs[1]].as_int ==
+            best_mbmode.mv[1].as_int) || !comp_pred_mode))
+      best_mbmode.mode = NEARESTMV;
+    else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
+        ((comp_pred_mode && frame_mv[NEARMV][refs[1]].as_int ==
+            best_mbmode.mv[1].as_int) || !comp_pred_mode))
+      best_mbmode.mode = NEARMV;
+    else if (best_mbmode.mv[0].as_int == 0 &&
+        ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) || !comp_pred_mode))
+      best_mbmode.mode = ZEROMV;
+  }
+
+  if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
+    rd_cost->rate = INT_MAX;
+    rd_cost->rdcost = INT64_MAX;
+    return;
+  }
+
+  // If we used an estimate for the uv intra rd in the loop above...
+  if (sf->use_uv_intra_rd_estimate) {
+    // Do Intra UV best rd mode selection if best mode choice above was intra.
+    if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
+      TX_SIZE uv_tx_size;
+      *mi = best_mbmode;
+      uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
+      rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
+                              &rate_uv_tokenonly[uv_tx_size],
+                              &dist_uv[uv_tx_size],
+                              &skip_uv[uv_tx_size],
+                              bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
+                              uv_tx_size);
+    }
+  }
+
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == best_mbmode.interp_filter) ||
+         !is_inter_block(&best_mbmode));
+
+  if (!cpi->rc.is_src_frame_alt_ref)
+    vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
+                              sf->adaptive_rd_thresh, bsize, best_mode_index);
+
+  // macroblock modes
+  *mi = best_mbmode;
+  x->skip |= best_skip2;
+
+  for (i = 0; i < REFERENCE_MODES; ++i) {
+    if (best_pred_rd[i] == INT64_MAX)
+      best_pred_diff[i] = INT_MIN;
+    else
+      best_pred_diff[i] = best_rd - best_pred_rd[i];
+  }
+
+  if (!x->skip) {
+    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
+      if (best_filter_rd[i] == INT64_MAX)
+        best_filter_diff[i] = 0;
+      else
+        best_filter_diff[i] = best_rd - best_filter_rd[i];
+    }
+    if (cm->interp_filter == SWITCHABLE)
+      assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
+  } else {
+    vp9_zero(best_filter_diff);
+  }
+
+  // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
+  // updating code causes PSNR loss. Need to figure out the confliction.
+  x->skip |= best_mode_skippable;
+
+  if (!x->skip && !x->select_tx_size) {
+    int has_high_freq_coeff = 0;
+    int plane;
+    int max_plane = is_inter_block(xd->mi[0])
+                        ? MAX_MB_PLANE : 1;
+    for (plane = 0; plane < max_plane; ++plane) {
+      x->plane[plane].eobs = ctx->eobs_pbuf[plane][1];
+      has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
+    }
+
+    for (plane = max_plane; plane < MAX_MB_PLANE; ++plane) {
+      x->plane[plane].eobs = ctx->eobs_pbuf[plane][2];
+      has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
+    }
+
+    best_mode_skippable |= !has_high_freq_coeff;
+  }
+
+  assert(best_mode_index >= 0);
+
+  store_coding_context(x, ctx, best_mode_index, best_pred_diff,
+                       best_filter_diff, best_mode_skippable);
+}
+
+void vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi,
+                                        TileDataEnc *tile_data,
+                                        MACROBLOCK *x,
+                                        RD_COST *rd_cost,
+                                        BLOCK_SIZE bsize,
+                                        PICK_MODE_CONTEXT *ctx,
+                                        int64_t best_rd_so_far) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  unsigned char segment_id = mi->segment_id;
+  const int comp_pred = 0;
+  int i;
+  int64_t best_pred_diff[REFERENCE_MODES];
+  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+  unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
+  vpx_prob comp_mode_p;
+  INTERP_FILTER best_filter = SWITCHABLE;
+  int64_t this_rd = INT64_MAX;
+  int rate2 = 0;
+  const int64_t distortion2 = 0;
+
+  x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
+
+  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
+                           &comp_mode_p);
+
+  for (i = 0; i < MAX_REF_FRAMES; ++i)
+    x->pred_sse[i] = INT_MAX;
+  for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
+    x->pred_mv_sad[i] = INT_MAX;
+
+  rd_cost->rate = INT_MAX;
+
+  assert(segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
+
+  mi->mode = ZEROMV;
+  mi->uv_mode = DC_PRED;
+  mi->ref_frame[0] = LAST_FRAME;
+  mi->ref_frame[1] = NONE;
+  mi->mv[0].as_int = 0;
+  x->skip = 1;
+
+  if (cm->interp_filter != BILINEAR) {
+    best_filter = EIGHTTAP;
+    if (cm->interp_filter == SWITCHABLE &&
+        x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
+      int rs;
+      int best_rs = INT_MAX;
+      for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
+        mi->interp_filter = i;
+        rs = vp9_get_switchable_rate(cpi, xd);
+        if (rs < best_rs) {
+          best_rs = rs;
+          best_filter = mi->interp_filter;
+        }
+      }
+    }
+  }
+  // Set the appropriate filter
+  if (cm->interp_filter == SWITCHABLE) {
+    mi->interp_filter = best_filter;
+    rate2 += vp9_get_switchable_rate(cpi, xd);
+  } else {
+    mi->interp_filter = cm->interp_filter;
+  }
+
+  if (cm->reference_mode == REFERENCE_MODE_SELECT)
+    rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
+
+  // Estimate the reference frame signaling cost and add it
+  // to the rolling cost variable.
+  rate2 += ref_costs_single[LAST_FRAME];
+  this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+
+  rd_cost->rate = rate2;
+  rd_cost->dist = distortion2;
+  rd_cost->rdcost = this_rd;
+
+  if (this_rd >= best_rd_so_far) {
+    rd_cost->rate = INT_MAX;
+    rd_cost->rdcost = INT64_MAX;
+    return;
+  }
+
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == mi->interp_filter));
+
+  vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
+                            cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
+
+  vp9_zero(best_pred_diff);
+  vp9_zero(best_filter_diff);
+
+  if (!x->select_tx_size)
+    swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
+  store_coding_context(x, ctx, THR_ZEROMV,
+                       best_pred_diff, best_filter_diff, 0);
+}
+
+void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi,
+                                   TileDataEnc *tile_data,
+                                   MACROBLOCK *x,
+                                   int mi_row, int mi_col,
+                                   RD_COST *rd_cost,
+                                   BLOCK_SIZE bsize,
+                                   PICK_MODE_CONTEXT *ctx,
+                                   int64_t best_rd_so_far) {
+  VP9_COMMON *const cm = &cpi->common;
+  RD_OPT *const rd_opt = &cpi->rd;
+  SPEED_FEATURES *const sf = &cpi->sf;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  const struct segmentation *const seg = &cm->seg;
+  MV_REFERENCE_FRAME ref_frame, second_ref_frame;
+  unsigned char segment_id = mi->segment_id;
+  int comp_pred, i;
+  int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
+  struct buf_2d yv12_mb[4][MAX_MB_PLANE];
+  static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
+                                    VP9_ALT_FLAG };
+  int64_t best_rd = best_rd_so_far;
+  int64_t best_yrd = best_rd_so_far;  // FIXME(rbultje) more precise
+  int64_t best_pred_diff[REFERENCE_MODES];
+  int64_t best_pred_rd[REFERENCE_MODES];
+  int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
+  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+  MODE_INFO best_mbmode;
+  int ref_index, best_ref_index = 0;
+  unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
+  vpx_prob comp_mode_p;
+  INTERP_FILTER tmp_best_filter = SWITCHABLE;
+  int rate_uv_intra, rate_uv_tokenonly;
+  int64_t dist_uv;
+  int skip_uv;
+  PREDICTION_MODE mode_uv = DC_PRED;
+  const int intra_cost_penalty = vp9_get_intra_cost_penalty(
+    cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
+  int_mv seg_mvs[4][MAX_REF_FRAMES];
+  b_mode_info best_bmodes[4];
+  int best_skip2 = 0;
+  int ref_frame_skip_mask[2] = { 0 };
+  int64_t mask_filter = 0;
+  int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
+  int internal_active_edge =
+    vp9_active_edge_sb(cpi, mi_row, mi_col) && vp9_internal_image_edge(cpi);
+
+  x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
+  memset(x->zcoeff_blk[TX_4X4], 0, 4);
+  vp9_zero(best_mbmode);
+
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+    filter_cache[i] = INT64_MAX;
+
+  for (i = 0; i < 4; i++) {
+    int j;
+    for (j = 0; j < MAX_REF_FRAMES; j++)
+      seg_mvs[i][j].as_int = INVALID_MV;
+  }
+
+  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
+                           &comp_mode_p);
+
+  for (i = 0; i < REFERENCE_MODES; ++i)
+    best_pred_rd[i] = INT64_MAX;
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+    best_filter_rd[i] = INT64_MAX;
+  rate_uv_intra = INT_MAX;
+
+  rd_cost->rate = INT_MAX;
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
+    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
+      setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
+                         frame_mv[NEARESTMV], frame_mv[NEARMV],
+                         yv12_mb);
+    } else {
+      ref_frame_skip_mask[0] |= (1 << ref_frame);
+      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+    }
+    frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
+    frame_mv[ZEROMV][ref_frame].as_int = 0;
+  }
+
+  for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
+    int mode_excluded = 0;
+    int64_t this_rd = INT64_MAX;
+    int disable_skip = 0;
+    int compmode_cost = 0;
+    int rate2 = 0, rate_y = 0, rate_uv = 0;
+    int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
+    int skippable = 0;
+    int i;
+    int this_skip2 = 0;
+    int64_t total_sse = INT_MAX;
+    int early_term = 0;
+    struct buf_2d backup_yv12[2][MAX_MB_PLANE];
+
+    ref_frame = vp9_ref_order[ref_index].ref_frame[0];
+    second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
+
+#if CONFIG_BETTER_HW_COMPATIBILITY
+    // forbid 8X4 and 4X8 partitions if any reference frame is scaled.
+    if (bsize == BLOCK_8X4 || bsize == BLOCK_4X8) {
+      int ref_scaled = vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf);
+      if (second_ref_frame > INTRA_FRAME)
+        ref_scaled += vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf);
+      if (ref_scaled)
+        continue;
+    }
+#endif
+    // Look at the reference frame of the best mode so far and set the
+    // skip mask to look at a subset of the remaining modes.
+    if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
+      if (ref_index == 3) {
+        switch (best_mbmode.ref_frame[0]) {
+          case INTRA_FRAME:
+            break;
+          case LAST_FRAME:
+            ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
+            ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+            break;
+          case GOLDEN_FRAME:
+            ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
+            ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+            break;
+          case ALTREF_FRAME:
+            ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
+            break;
+          case NONE:
+          case MAX_REF_FRAMES:
+            assert(0 && "Invalid Reference frame");
+            break;
+        }
+      }
+    }
+
+    if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
+        (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
+      continue;
+
+    // Test best rd so far against threshold for trying this mode.
+    if (!internal_active_edge &&
+        rd_less_than_thresh(best_rd,
+                            rd_opt->threshes[segment_id][bsize][ref_index],
+                            tile_data->thresh_freq_fact[bsize][ref_index]))
+      continue;
+
+    comp_pred = second_ref_frame > INTRA_FRAME;
+    if (comp_pred) {
+      if (!cpi->allow_comp_inter_inter)
+        continue;
+      if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
+        continue;
+      // Do not allow compound prediction if the segment level reference frame
+      // feature is in use as in this case there can only be one reference.
+      if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
+        continue;
+
+      if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
+          best_mbmode.ref_frame[0] == INTRA_FRAME)
+        continue;
+    }
+
+    if (comp_pred)
+      mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
+    else if (ref_frame != INTRA_FRAME)
+      mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
+
+    // If the segment reference frame feature is enabled....
+    // then do nothing if the current ref frame is not allowed..
+    if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
+        get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
+      continue;
+    // Disable this drop out case if the ref frame
+    // segment level feature is enabled for this segment. This is to
+    // prevent the possibility that we end up unable to pick any mode.
+    } else if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
+      // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
+      // unless ARNR filtering is enabled in which case we want
+      // an unfiltered alternative. We allow near/nearest as well
+      // because they may result in zero-zero MVs but be cheaper.
+      if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
+        continue;
+    }
+
+    mi->tx_size = TX_4X4;
+    mi->uv_mode = DC_PRED;
+    mi->ref_frame[0] = ref_frame;
+    mi->ref_frame[1] = second_ref_frame;
+    // Evaluate all sub-pel filters irrespective of whether we can use
+    // them for this frame.
+    mi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
+                                                        : cm->interp_filter;
+    x->skip = 0;
+    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
+
+    // Select prediction reference frames.
+    for (i = 0; i < MAX_MB_PLANE; i++) {
+      xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
+      if (comp_pred)
+        xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
+    }
+
+    if (ref_frame == INTRA_FRAME) {
+      int rate;
+      if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
+                                       &distortion_y, best_rd) >= best_rd)
+        continue;
+      rate2 += rate;
+      rate2 += intra_cost_penalty;
+      distortion2 += distortion_y;
+
+      if (rate_uv_intra == INT_MAX) {
+        choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4,
+                             &rate_uv_intra,
+                             &rate_uv_tokenonly,
+                             &dist_uv, &skip_uv,
+                             &mode_uv);
+      }
+      rate2 += rate_uv_intra;
+      rate_uv = rate_uv_tokenonly;
+      distortion2 += dist_uv;
+      distortion_uv = dist_uv;
+      mi->uv_mode = mode_uv;
+    } else {
+      int rate;
+      int64_t distortion;
+      int64_t this_rd_thresh;
+      int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
+      int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
+      int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
+      int tmp_best_skippable = 0;
+      int switchable_filter_index;
+      int_mv *second_ref = comp_pred ?
+                             &x->mbmi_ext->ref_mvs[second_ref_frame][0] : NULL;
+      b_mode_info tmp_best_bmodes[16];
+      MODE_INFO tmp_best_mbmode;
+      BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
+      int pred_exists = 0;
+      int uv_skippable;
+
+      YV12_BUFFER_CONFIG *scaled_ref_frame[2] = {NULL, NULL};
+      int ref;
+
+      for (ref = 0; ref < 2; ++ref) {
+        scaled_ref_frame[ref] = mi->ref_frame[ref] > INTRA_FRAME ?
+            vp9_get_scaled_ref_frame(cpi, mi->ref_frame[ref]) : NULL;
+
+        if (scaled_ref_frame[ref]) {
+          int i;
+          // Swap out the reference frame for a version that's been scaled to
+          // match the resolution of the current frame, allowing the existing
+          // motion search code to be used without additional modifications.
+          for (i = 0; i < MAX_MB_PLANE; i++)
+            backup_yv12[ref][i] = xd->plane[i].pre[ref];
+          vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
+                               NULL);
+        }
+      }
+
+      this_rd_thresh = (ref_frame == LAST_FRAME) ?
+          rd_opt->threshes[segment_id][bsize][THR_LAST] :
+          rd_opt->threshes[segment_id][bsize][THR_ALTR];
+      this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
+      rd_opt->threshes[segment_id][bsize][THR_GOLD] : this_rd_thresh;
+      for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+        filter_cache[i] = INT64_MAX;
+
+      if (cm->interp_filter != BILINEAR) {
+        tmp_best_filter = EIGHTTAP;
+        if (x->source_variance < sf->disable_filter_search_var_thresh) {
+          tmp_best_filter = EIGHTTAP;
+        } else if (sf->adaptive_pred_interp_filter == 1 &&
+                   ctx->pred_interp_filter < SWITCHABLE) {
+          tmp_best_filter = ctx->pred_interp_filter;
+        } else if (sf->adaptive_pred_interp_filter == 2) {
+          tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE ?
+                              ctx->pred_interp_filter : 0;
+        } else {
+          for (switchable_filter_index = 0;
+               switchable_filter_index < SWITCHABLE_FILTERS;
+               ++switchable_filter_index) {
+            int newbest, rs;
+            int64_t rs_rd;
+            MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
+            mi->interp_filter = switchable_filter_index;
+            tmp_rd = rd_pick_best_sub8x8_mode(cpi, x,
+                                              &mbmi_ext->ref_mvs[ref_frame][0],
+                                              second_ref, best_yrd, &rate,
+                                              &rate_y, &distortion,
+                                              &skippable, &total_sse,
+                                              (int) this_rd_thresh, seg_mvs,
+                                              bsi, switchable_filter_index,
+                                              mi_row, mi_col);
+
+            if (tmp_rd == INT64_MAX)
+              continue;
+            rs = vp9_get_switchable_rate(cpi, xd);
+            rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
+            filter_cache[switchable_filter_index] = tmp_rd;
+            filter_cache[SWITCHABLE_FILTERS] =
+                VPXMIN(filter_cache[SWITCHABLE_FILTERS], tmp_rd + rs_rd);
+            if (cm->interp_filter == SWITCHABLE)
+              tmp_rd += rs_rd;
+
+            mask_filter = VPXMAX(mask_filter, tmp_rd);
+
+            newbest = (tmp_rd < tmp_best_rd);
+            if (newbest) {
+              tmp_best_filter = mi->interp_filter;
+              tmp_best_rd = tmp_rd;
+            }
+            if ((newbest && cm->interp_filter == SWITCHABLE) ||
+                (mi->interp_filter == cm->interp_filter &&
+                 cm->interp_filter != SWITCHABLE)) {
+              tmp_best_rdu = tmp_rd;
+              tmp_best_rate = rate;
+              tmp_best_ratey = rate_y;
+              tmp_best_distortion = distortion;
+              tmp_best_sse = total_sse;
+              tmp_best_skippable = skippable;
+              tmp_best_mbmode = *mi;
+              for (i = 0; i < 4; i++) {
+                tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
+                x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
+              }
+              pred_exists = 1;
+              if (switchable_filter_index == 0 &&
+                  sf->use_rd_breakout &&
+                  best_rd < INT64_MAX) {
+                if (tmp_best_rdu / 2 > best_rd) {
+                  // skip searching the other filters if the first is
+                  // already substantially larger than the best so far
+                  tmp_best_filter = mi->interp_filter;
+                  tmp_best_rdu = INT64_MAX;
+                  break;
+                }
+              }
+            }
+          }  // switchable_filter_index loop
+        }
+      }
+
+      if (tmp_best_rdu == INT64_MAX && pred_exists)
+        continue;
+
+      mi->interp_filter = (cm->interp_filter == SWITCHABLE ?
+                           tmp_best_filter : cm->interp_filter);
+      if (!pred_exists) {
+        // Handles the special case when a filter that is not in the
+        // switchable list (bilinear, 6-tap) is indicated at the frame level
+        tmp_rd = rd_pick_best_sub8x8_mode(cpi, x,
+                                          &x->mbmi_ext->ref_mvs[ref_frame][0],
+                                          second_ref, best_yrd, &rate, &rate_y,
+                                          &distortion, &skippable, &total_sse,
+                                          (int) this_rd_thresh, seg_mvs, bsi, 0,
+                                          mi_row, mi_col);
+        if (tmp_rd == INT64_MAX)
+          continue;
+      } else {
+        total_sse = tmp_best_sse;
+        rate = tmp_best_rate;
+        rate_y = tmp_best_ratey;
+        distortion = tmp_best_distortion;
+        skippable = tmp_best_skippable;
+        *mi = tmp_best_mbmode;
+        for (i = 0; i < 4; i++)
+          xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
+      }
+
+      rate2 += rate;
+      distortion2 += distortion;
+
+      if (cm->interp_filter == SWITCHABLE)
+        rate2 += vp9_get_switchable_rate(cpi, xd);
+
+      if (!mode_excluded)
+        mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
+                                  : cm->reference_mode == COMPOUND_REFERENCE;
+
+      compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
+
+      tmp_best_rdu =
+          best_rd - VPXMIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
+                           RDCOST(x->rdmult, x->rddiv, 0, total_sse));
+
+      if (tmp_best_rdu > 0) {
+        // If even the 'Y' rd value of split is higher than best so far
+        // then dont bother looking at UV
+        vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col,
+                                        BLOCK_8X8);
+        memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
+        if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
+                              &uv_sse, BLOCK_8X8, tmp_best_rdu)) {
+          for (ref = 0; ref < 2; ++ref) {
+            if (scaled_ref_frame[ref]) {
+              int i;
+              for (i = 0; i < MAX_MB_PLANE; ++i)
+                xd->plane[i].pre[ref] = backup_yv12[ref][i];
+            }
+          }
+          continue;
+        }
+
+        rate2 += rate_uv;
+        distortion2 += distortion_uv;
+        skippable = skippable && uv_skippable;
+        total_sse += uv_sse;
+      }
+
+      for (ref = 0; ref < 2; ++ref) {
+        if (scaled_ref_frame[ref]) {
+          // Restore the prediction frame pointers to their unscaled versions.
+          int i;
+          for (i = 0; i < MAX_MB_PLANE; ++i)
+            xd->plane[i].pre[ref] = backup_yv12[ref][i];
+        }
+      }
+    }
+
+    if (cm->reference_mode == REFERENCE_MODE_SELECT)
+      rate2 += compmode_cost;
+
+    // Estimate the reference frame signaling cost and add it
+    // to the rolling cost variable.
+    if (second_ref_frame > INTRA_FRAME) {
+      rate2 += ref_costs_comp[ref_frame];
+    } else {
+      rate2 += ref_costs_single[ref_frame];
+    }
+
+    if (!disable_skip) {
+      const vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
+      const int skip_cost0 = vp9_cost_bit(skip_prob, 0);
+      const int skip_cost1 = vp9_cost_bit(skip_prob, 1);
+
+      // Skip is never coded at the segment level for sub8x8 blocks and instead
+      // always coded in the bitstream at the mode info level.
+      if (ref_frame != INTRA_FRAME && !xd->lossless) {
+        if (RDCOST(x->rdmult, x->rddiv,
+                   rate_y + rate_uv + skip_cost0, distortion2) <
+            RDCOST(x->rdmult, x->rddiv, skip_cost1, total_sse)) {
+          // Add in the cost of the no skip flag.
+          rate2 += skip_cost0;
+        } else {
+          // FIXME(rbultje) make this work for splitmv also
+          rate2 += skip_cost1;
+          distortion2 = total_sse;
+          assert(total_sse >= 0);
+          rate2 -= (rate_y + rate_uv);
+          rate_y = 0;
+          rate_uv = 0;
+          this_skip2 = 1;
+        }
+      } else {
+        // Add in the cost of the no skip flag.
+        rate2 += skip_cost0;
+      }
+
+      // Calculate the final RD estimate for this mode.
+      this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+    }
+
+    if (!disable_skip && ref_frame == INTRA_FRAME) {
+      for (i = 0; i < REFERENCE_MODES; ++i)
+        best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
+      for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+        best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
+    }
+
+    // Did this mode help.. i.e. is it the new best mode
+    if (this_rd < best_rd || x->skip) {
+      if (!mode_excluded) {
+        int max_plane = MAX_MB_PLANE;
+        // Note index of best mode so far
+        best_ref_index = ref_index;
+
+        if (ref_frame == INTRA_FRAME) {
+          /* required for left and above block mv */
+          mi->mv[0].as_int = 0;
+          max_plane = 1;
+        }
+
+        rd_cost->rate = rate2;
+        rd_cost->dist = distortion2;
+        rd_cost->rdcost = this_rd;
+        best_rd = this_rd;
+        best_yrd = best_rd -
+                   RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
+        best_mbmode = *mi;
+        best_skip2 = this_skip2;
+        if (!x->select_tx_size)
+          swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
+        memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
+               sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
+
+        for (i = 0; i < 4; i++)
+          best_bmodes[i] = xd->mi[0]->bmi[i];
+
+        // TODO(debargha): enhance this test with a better distortion prediction
+        // based on qp, activity mask and history
+        if ((sf->mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
+            (ref_index > MIN_EARLY_TERM_INDEX)) {
+          int qstep = xd->plane[0].dequant[1];
+          // TODO(debargha): Enhance this by specializing for each mode_index
+          int scale = 4;
+#if CONFIG_VP9_HIGHBITDEPTH
+          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            qstep >>= (xd->bd - 8);
+          }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+          if (x->source_variance < UINT_MAX) {
+            const int var_adjust = (x->source_variance < 16);
+            scale -= var_adjust;
+          }
+          if (ref_frame > INTRA_FRAME &&
+              distortion2 * scale < qstep * qstep) {
+            early_term = 1;
+          }
+        }
+      }
+    }
+
+    /* keep record of best compound/single-only prediction */
+    if (!disable_skip && ref_frame != INTRA_FRAME) {
+      int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
+
+      if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+        single_rate = rate2 - compmode_cost;
+        hybrid_rate = rate2;
+      } else {
+        single_rate = rate2;
+        hybrid_rate = rate2 + compmode_cost;
+      }
+
+      single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
+      hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
+
+      if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
+        best_pred_rd[SINGLE_REFERENCE] = single_rd;
+      else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
+        best_pred_rd[COMPOUND_REFERENCE] = single_rd;
+
+      if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
+        best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
+    }
+
+    /* keep record of best filter type */
+    if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
+        cm->interp_filter != BILINEAR) {
+      int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
+                              SWITCHABLE_FILTERS : cm->interp_filter];
+      int64_t adj_rd;
+      for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
+        if (ref == INT64_MAX)
+          adj_rd = 0;
+        else if (filter_cache[i] == INT64_MAX)
+          // when early termination is triggered, the encoder does not have
+          // access to the rate-distortion cost. it only knows that the cost
+          // should be above the maximum valid value. hence it takes the known
+          // maximum plus an arbitrary constant as the rate-distortion cost.
+          adj_rd = mask_filter - ref + 10;
+        else
+          adj_rd = filter_cache[i] - ref;
+
+        adj_rd += this_rd;
+        best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
+      }
+    }
+
+    if (early_term)
+      break;
+
+    if (x->skip && !comp_pred)
+      break;
+  }
+
+  if (best_rd >= best_rd_so_far) {
+    rd_cost->rate = INT_MAX;
+    rd_cost->rdcost = INT64_MAX;
+    return;
+  }
+
+  // If we used an estimate for the uv intra rd in the loop above...
+  if (sf->use_uv_intra_rd_estimate) {
+    // Do Intra UV best rd mode selection if best mode choice above was intra.
+    if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
+      *mi = best_mbmode;
+      rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra,
+                              &rate_uv_tokenonly,
+                              &dist_uv,
+                              &skip_uv,
+                              BLOCK_8X8, TX_4X4);
+    }
+  }
+
+  if (best_rd == INT64_MAX) {
+    rd_cost->rate = INT_MAX;
+    rd_cost->dist = INT64_MAX;
+    rd_cost->rdcost = INT64_MAX;
+    return;
+  }
+
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == best_mbmode.interp_filter) ||
+         !is_inter_block(&best_mbmode));
+
+  vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
+                            sf->adaptive_rd_thresh, bsize, best_ref_index);
+
+  // macroblock modes
+  *mi = best_mbmode;
+  x->skip |= best_skip2;
+  if (!is_inter_block(&best_mbmode)) {
+    for (i = 0; i < 4; i++)
+      xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
+  } else {
+    for (i = 0; i < 4; ++i)
+      memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
+
+    mi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
+    mi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
+  }
+
+  for (i = 0; i < REFERENCE_MODES; ++i) {
+    if (best_pred_rd[i] == INT64_MAX)
+      best_pred_diff[i] = INT_MIN;
+    else
+      best_pred_diff[i] = best_rd - best_pred_rd[i];
+  }
+
+  if (!x->skip) {
+    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
+      if (best_filter_rd[i] == INT64_MAX)
+        best_filter_diff[i] = 0;
+      else
+        best_filter_diff[i] = best_rd - best_filter_rd[i];
+    }
+    if (cm->interp_filter == SWITCHABLE)
+      assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
+  } else {
+    vp9_zero(best_filter_diff);
+  }
+
+  store_coding_context(x, ctx, best_ref_index,
+                       best_pred_diff, best_filter_diff, 0);
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_rdopt.h b/libvpx/libvpx/vp9/encoder/vp9_rdopt.h
new file mode 100644
index 0000000..253e4a0
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_rdopt.h
@@ -0,0 +1,65 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_RDOPT_H_
+#define VP9_ENCODER_VP9_RDOPT_H_
+
+#include "vp9/common/vp9_blockd.h"
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_context_tree.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct TileInfo;
+struct VP9_COMP;
+struct macroblock;
+struct RD_COST;
+
+void vp9_rd_pick_intra_mode_sb(struct VP9_COMP *cpi, struct macroblock *x,
+                               struct RD_COST *rd_cost, BLOCK_SIZE bsize,
+                               PICK_MODE_CONTEXT *ctx, int64_t best_rd);
+
+void vp9_rd_pick_inter_mode_sb(struct VP9_COMP *cpi,
+                               struct TileDataEnc *tile_data,
+                               struct macroblock *x,
+                               int mi_row, int mi_col,
+                               struct RD_COST *rd_cost,
+                               BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                               int64_t best_rd_so_far);
+
+void vp9_rd_pick_inter_mode_sb_seg_skip(struct VP9_COMP *cpi,
+                                        struct TileDataEnc *tile_data,
+                                        struct macroblock *x,
+                                        struct RD_COST *rd_cost,
+                                        BLOCK_SIZE bsize,
+                                        PICK_MODE_CONTEXT *ctx,
+                                        int64_t best_rd_so_far);
+
+int vp9_internal_image_edge(struct VP9_COMP *cpi);
+int vp9_active_h_edge(struct VP9_COMP *cpi, int mi_row, int mi_step);
+int vp9_active_v_edge(struct VP9_COMP *cpi, int mi_col, int mi_step);
+int vp9_active_edge_sb(struct VP9_COMP *cpi, int mi_row, int mi_col);
+
+void vp9_rd_pick_inter_mode_sub8x8(struct VP9_COMP *cpi,
+                                   struct TileDataEnc *tile_data,
+                                   struct macroblock *x,
+                                   int mi_row, int mi_col,
+                                   struct RD_COST *rd_cost,
+                                   BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                                   int64_t best_rd_so_far);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_RDOPT_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_resize.c b/libvpx/libvpx/vp9/encoder/vp9_resize.c
new file mode 100644
index 0000000..307a112
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_resize.c
@@ -0,0 +1,929 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "./vpx_config.h"
+#if CONFIG_VP9_HIGHBITDEPTH
+#include "vpx_dsp/vpx_dsp_common.h"
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/encoder/vp9_resize.h"
+
+#define FILTER_BITS               7
+
+#define INTERP_TAPS               8
+#define SUBPEL_BITS               5
+#define SUBPEL_MASK               ((1 << SUBPEL_BITS) - 1)
+#define INTERP_PRECISION_BITS     32
+
+typedef int16_t interp_kernel[INTERP_TAPS];
+
+// Filters for interpolation (0.5-band) - note this also filters integer pels.
+static const interp_kernel filteredinterp_filters500[(1 << SUBPEL_BITS)] = {
+  {-3,  0, 35, 64, 35,  0, -3, 0},
+  {-3, -1, 34, 64, 36,  1, -3, 0},
+  {-3, -1, 32, 64, 38,  1, -3, 0},
+  {-2, -2, 31, 63, 39,  2, -3, 0},
+  {-2, -2, 29, 63, 41,  2, -3, 0},
+  {-2, -2, 28, 63, 42,  3, -4, 0},
+  {-2, -3, 27, 63, 43,  4, -4, 0},
+  {-2, -3, 25, 62, 45,  5, -4, 0},
+  {-2, -3, 24, 62, 46,  5, -4, 0},
+  {-2, -3, 23, 61, 47,  6, -4, 0},
+  {-2, -3, 21, 60, 49,  7, -4, 0},
+  {-1, -4, 20, 60, 50,  8, -4, -1},
+  {-1, -4, 19, 59, 51,  9, -4, -1},
+  {-1, -4, 17, 58, 52, 10, -4, 0},
+  {-1, -4, 16, 57, 53, 12, -4, -1},
+  {-1, -4, 15, 56, 54, 13, -4, -1},
+  {-1, -4, 14, 55, 55, 14, -4, -1},
+  {-1, -4, 13, 54, 56, 15, -4, -1},
+  {-1, -4, 12, 53, 57, 16, -4, -1},
+  {0, -4, 10, 52, 58, 17, -4, -1},
+  {-1, -4,  9, 51, 59, 19, -4, -1},
+  {-1, -4,  8, 50, 60, 20, -4, -1},
+  {0, -4,  7, 49, 60, 21, -3, -2},
+  {0, -4,  6, 47, 61, 23, -3, -2},
+  {0, -4,  5, 46, 62, 24, -3, -2},
+  {0, -4,  5, 45, 62, 25, -3, -2},
+  {0, -4,  4, 43, 63, 27, -3, -2},
+  {0, -4,  3, 42, 63, 28, -2, -2},
+  {0, -3,  2, 41, 63, 29, -2, -2},
+  {0, -3,  2, 39, 63, 31, -2, -2},
+  {0, -3,  1, 38, 64, 32, -1, -3},
+  {0, -3,  1, 36, 64, 34, -1, -3}
+};
+
+// Filters for interpolation (0.625-band) - note this also filters integer pels.
+static const interp_kernel filteredinterp_filters625[(1 << SUBPEL_BITS)] = {
+  {-1, -8, 33, 80, 33, -8, -1, 0},
+  {-1, -8, 30, 80, 35, -8, -1, 1},
+  {-1, -8, 28, 80, 37, -7, -2, 1},
+  {0, -8, 26, 79, 39, -7, -2, 1},
+  {0, -8, 24, 79, 41, -7, -2, 1},
+  {0, -8, 22, 78, 43, -6, -2, 1},
+  {0, -8, 20, 78, 45, -5, -3, 1},
+  {0, -8, 18, 77, 48, -5, -3, 1},
+  {0, -8, 16, 76, 50, -4, -3, 1},
+  {0, -8, 15, 75, 52, -3, -4, 1},
+  {0, -7, 13, 74, 54, -3, -4, 1},
+  {0, -7, 11, 73, 56, -2, -4, 1},
+  {0, -7, 10, 71, 58, -1, -4, 1},
+  {1, -7,  8, 70, 60,  0, -5, 1},
+  {1, -6,  6, 68, 62,  1, -5, 1},
+  {1, -6,  5, 67, 63,  2, -5, 1},
+  {1, -6,  4, 65, 65,  4, -6, 1},
+  {1, -5,  2, 63, 67,  5, -6, 1},
+  {1, -5,  1, 62, 68,  6, -6, 1},
+  {1, -5,  0, 60, 70,  8, -7, 1},
+  {1, -4, -1, 58, 71, 10, -7, 0},
+  {1, -4, -2, 56, 73, 11, -7, 0},
+  {1, -4, -3, 54, 74, 13, -7, 0},
+  {1, -4, -3, 52, 75, 15, -8, 0},
+  {1, -3, -4, 50, 76, 16, -8, 0},
+  {1, -3, -5, 48, 77, 18, -8, 0},
+  {1, -3, -5, 45, 78, 20, -8, 0},
+  {1, -2, -6, 43, 78, 22, -8, 0},
+  {1, -2, -7, 41, 79, 24, -8, 0},
+  {1, -2, -7, 39, 79, 26, -8, 0},
+  {1, -2, -7, 37, 80, 28, -8, -1},
+  {1, -1, -8, 35, 80, 30, -8, -1},
+};
+
+// Filters for interpolation (0.75-band) - note this also filters integer pels.
+static const interp_kernel filteredinterp_filters750[(1 << SUBPEL_BITS)] = {
+  {2, -11,  25,  96,  25, -11,   2, 0},
+  {2, -11,  22,  96,  28, -11,   2, 0},
+  {2, -10,  19,  95,  31, -11,   2, 0},
+  {2, -10,  17,  95,  34, -12,   2, 0},
+  {2,  -9,  14,  94,  37, -12,   2, 0},
+  {2,  -8,  12,  93,  40, -12,   1, 0},
+  {2,  -8,   9,  92,  43, -12,   1, 1},
+  {2,  -7,   7,  91,  46, -12,   1, 0},
+  {2,  -7,   5,  90,  49, -12,   1, 0},
+  {2,  -6,   3,  88,  52, -12,   0, 1},
+  {2,  -5,   1,  86,  55, -12,   0, 1},
+  {2,  -5,  -1,  84,  58, -11,   0, 1},
+  {2,  -4,  -2,  82,  61, -11,  -1, 1},
+  {2,  -4,  -4,  80,  64, -10,  -1, 1},
+  {1, -3, -5, 77, 67, -9, -1, 1},
+  {1, -3, -6, 75, 70, -8, -2, 1},
+  {1, -2, -7, 72, 72, -7, -2, 1},
+  {1, -2, -8, 70, 75, -6, -3, 1},
+  {1, -1, -9, 67, 77, -5, -3, 1},
+  {1,  -1, -10,  64,  80,  -4,  -4, 2},
+  {1,  -1, -11,  61,  82,  -2,  -4, 2},
+  {1,   0, -11,  58,  84,  -1,  -5, 2},
+  {1,   0, -12,  55,  86,   1,  -5, 2},
+  {1,   0, -12,  52,  88,   3,  -6, 2},
+  {0,   1, -12,  49,  90,   5,  -7, 2},
+  {0,   1, -12,  46,  91,   7,  -7, 2},
+  {1,   1, -12,  43,  92,   9,  -8, 2},
+  {0,   1, -12,  40,  93,  12,  -8, 2},
+  {0,   2, -12,  37,  94,  14,  -9, 2},
+  {0,   2, -12,  34,  95,  17, -10, 2},
+  {0,   2, -11,  31,  95,  19, -10, 2},
+  {0,   2, -11,  28,  96,  22, -11, 2}
+};
+
+// Filters for interpolation (0.875-band) - note this also filters integer pels.
+static const interp_kernel filteredinterp_filters875[(1 << SUBPEL_BITS)] = {
+  {3,  -8,  13, 112,  13,  -8,   3, 0},
+  {3,  -7,  10, 112,  17,  -9,   3, -1},
+  {2,  -6,   7, 111,  21,  -9,   3, -1},
+  {2,  -5,   4, 111,  24, -10,   3, -1},
+  {2,  -4,   1, 110,  28, -11,   3, -1},
+  {1,  -3,  -1, 108,  32, -12,   4, -1},
+  {1,  -2,  -3, 106,  36, -13,   4, -1},
+  {1,  -1,  -6, 105,  40, -14,   4, -1},
+  {1,  -1,  -7, 102,  44, -14,   4, -1},
+  {1,   0,  -9, 100,  48, -15,   4, -1},
+  {1,   1, -11,  97,  53, -16,   4, -1},
+  {0,   1, -12,  95,  57, -16,   4, -1},
+  {0,   2, -13,  91,  61, -16,   4, -1},
+  {0,   2, -14,  88,  65, -16,   4, -1},
+  {0,   3, -15,  84,  69, -17,   4, 0},
+  {0,   3, -16,  81,  73, -16,   3, 0},
+  {0,   3, -16,  77,  77, -16,   3, 0},
+  {0,   3, -16,  73,  81, -16,   3, 0},
+  {0,   4, -17,  69,  84, -15,   3, 0},
+  {-1,   4, -16,  65,  88, -14,   2, 0},
+  {-1,   4, -16,  61,  91, -13,   2, 0},
+  {-1,   4, -16,  57,  95, -12,   1, 0},
+  {-1,   4, -16,  53,  97, -11,   1, 1},
+  {-1,   4, -15,  48, 100,  -9,   0, 1},
+  {-1,   4, -14,  44, 102,  -7,  -1, 1},
+  {-1,   4, -14,  40, 105,  -6,  -1, 1},
+  {-1,   4, -13,  36, 106,  -3,  -2, 1},
+  {-1,   4, -12,  32, 108,  -1,  -3, 1},
+  {-1,   3, -11,  28, 110,   1,  -4, 2},
+  {-1,   3, -10,  24, 111,   4,  -5, 2},
+  {-1,   3,  -9,  21, 111,   7,  -6, 2},
+  {-1,   3,  -9,  17, 112,  10,  -7, 3}
+};
+
+// Filters for interpolation (full-band) - no filtering for integer pixels
+static const interp_kernel filteredinterp_filters1000[(1 << SUBPEL_BITS)] = {
+  {0,   0,   0, 128,   0,   0,   0, 0},
+  {0,   1,  -3, 128,   3,  -1,   0, 0},
+  {-1,   2,  -6, 127,   7,  -2,   1, 0},
+  {-1,   3,  -9, 126,  12,  -4,   1, 0},
+  {-1,   4, -12, 125,  16,  -5,   1, 0},
+  {-1,   4, -14, 123,  20,  -6,   2, 0},
+  {-1,   5, -15, 120,  25,  -8,   2, 0},
+  {-1,   5, -17, 118,  30,  -9,   3, -1},
+  {-1,   6, -18, 114,  35, -10,   3, -1},
+  {-1,   6, -19, 111,  41, -12,   3, -1},
+  {-1,   6, -20, 107,  46, -13,   4, -1},
+  {-1,   6, -21, 103,  52, -14,   4, -1},
+  {-1,   6, -21,  99,  57, -16,   5, -1},
+  {-1,   6, -21,  94,  63, -17,   5, -1},
+  {-1,   6, -20,  89,  68, -18,   5, -1},
+  {-1,   6, -20,  84,  73, -19,   6, -1},
+  {-1,   6, -20,  79,  79, -20,   6, -1},
+  {-1,   6, -19,  73,  84, -20,   6, -1},
+  {-1,   5, -18,  68,  89, -20,   6, -1},
+  {-1,   5, -17,  63,  94, -21,   6, -1},
+  {-1,   5, -16,  57,  99, -21,   6, -1},
+  {-1,   4, -14,  52, 103, -21,   6, -1},
+  {-1,   4, -13,  46, 107, -20,   6, -1},
+  {-1,   3, -12,  41, 111, -19,   6, -1},
+  {-1,   3, -10,  35, 114, -18,   6, -1},
+  {-1,   3,  -9,  30, 118, -17,   5, -1},
+  {0,   2,  -8,  25, 120, -15,   5, -1},
+  {0,   2,  -6,  20, 123, -14,   4, -1},
+  {0,   1,  -5,  16, 125, -12,   4, -1},
+  {0,   1,  -4,  12, 126,  -9,   3, -1},
+  {0,   1,  -2,   7, 127,  -6,   2, -1},
+  {0,   0,  -1,   3, 128,  -3,   1, 0}
+};
+
+// Filters for factor of 2 downsampling.
+static const int16_t vp9_down2_symeven_half_filter[] = {56, 12, -3, -1};
+static const int16_t vp9_down2_symodd_half_filter[] = {64, 35, 0, -3};
+
+static const interp_kernel *choose_interp_filter(int inlength, int outlength) {
+  int outlength16 = outlength * 16;
+  if (outlength16 >= inlength * 16)
+    return filteredinterp_filters1000;
+  else if (outlength16 >= inlength * 13)
+    return filteredinterp_filters875;
+  else if (outlength16 >= inlength * 11)
+    return filteredinterp_filters750;
+  else if (outlength16 >= inlength * 9)
+    return filteredinterp_filters625;
+  else
+    return filteredinterp_filters500;
+}
+
+static void interpolate(const uint8_t *const input, int inlength,
+                        uint8_t *output, int outlength) {
+  const int64_t delta = (((uint64_t)inlength << 32) + outlength / 2) /
+      outlength;
+  const int64_t offset = inlength > outlength ?
+      (((int64_t)(inlength - outlength) << 31) + outlength / 2) / outlength :
+      -(((int64_t)(outlength - inlength) << 31) + outlength / 2) / outlength;
+  uint8_t *optr = output;
+  int x, x1, x2, sum, k, int_pel, sub_pel;
+  int64_t y;
+
+  const interp_kernel *interp_filters =
+      choose_interp_filter(inlength, outlength);
+
+  x = 0;
+  y = offset;
+  while ((y >> INTERP_PRECISION_BITS) < (INTERP_TAPS / 2 - 1)) {
+    x++;
+    y += delta;
+  }
+  x1 = x;
+  x = outlength - 1;
+  y = delta * x + offset;
+  while ((y >> INTERP_PRECISION_BITS) +
+         (int64_t)(INTERP_TAPS / 2) >= inlength) {
+    x--;
+    y -= delta;
+  }
+  x2 = x;
+  if (x1 > x2) {
+    for (x = 0, y = offset; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k) {
+        const int pk = int_pel - INTERP_TAPS / 2 + 1 + k;
+        sum += filter[k] * input[(pk < 0 ? 0 :
+                                  (pk >= inlength ? inlength - 1 : pk))];
+      }
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+  } else {
+    // Initial part.
+    for (x = 0, y = offset; x < x1; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k < 0 ?
+                                  0 :
+                                  int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+    // Middle part.
+    for (; x <= x2; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[int_pel - INTERP_TAPS / 2 + 1 + k];
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+    // End part.
+    for (; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k >=
+                                  inlength ?  inlength - 1 :
+                                  int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+  }
+}
+
+static void down2_symeven(const uint8_t *const input, int length,
+                          uint8_t *output) {
+  // Actual filter len = 2 * filter_len_half.
+  const int16_t *filter = vp9_down2_symeven_half_filter;
+  const int filter_len_half = sizeof(vp9_down2_symeven_half_filter) / 2;
+  int i, j;
+  uint8_t *optr = output;
+  int l1 = filter_len_half;
+  int l2 = (length - filter_len_half);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  }
+}
+
+static void down2_symodd(const uint8_t *const input, int length,
+                         uint8_t *output) {
+  // Actual filter len = 2 * filter_len_half - 1.
+  const int16_t *filter = vp9_down2_symodd_half_filter;
+  const int filter_len_half = sizeof(vp9_down2_symodd_half_filter) / 2;
+  int i, j;
+  uint8_t *optr = output;
+  int l1 = filter_len_half - 1;
+  int l2 = (length - filter_len_half + 1);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + j >= length ? length - 1 : i + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[(i + j >= length ? length - 1 : i + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  }
+}
+
+static int get_down2_length(int length, int steps) {
+  int s;
+  for (s = 0; s < steps; ++s)
+    length = (length + 1) >> 1;
+  return length;
+}
+
+static int get_down2_steps(int in_length, int out_length) {
+  int steps = 0;
+  int proj_in_length;
+  while ((proj_in_length = get_down2_length(in_length, 1)) >= out_length) {
+    ++steps;
+    in_length = proj_in_length;
+  }
+  return steps;
+}
+
+static void resize_multistep(const uint8_t *const input,
+                             int length,
+                             uint8_t *output,
+                             int olength,
+                             uint8_t *otmp) {
+  int steps;
+  if (length == olength) {
+    memcpy(output, input, sizeof(output[0]) * length);
+    return;
+  }
+  steps = get_down2_steps(length, olength);
+
+  if (steps > 0) {
+    int s;
+    uint8_t *out = NULL;
+    uint8_t *otmp2;
+    int filteredlength = length;
+
+    assert(otmp != NULL);
+    otmp2 = otmp + get_down2_length(length, 1);
+    for (s = 0; s < steps; ++s) {
+      const int proj_filteredlength = get_down2_length(filteredlength, 1);
+      const uint8_t *const in = (s == 0 ? input : out);
+      if (s == steps - 1 && proj_filteredlength == olength)
+        out = output;
+      else
+        out = (s & 1 ? otmp2 : otmp);
+      if (filteredlength & 1)
+        down2_symodd(in, filteredlength, out);
+      else
+        down2_symeven(in, filteredlength, out);
+      filteredlength = proj_filteredlength;
+    }
+    if (filteredlength != olength) {
+      interpolate(out, filteredlength, output, olength);
+    }
+  } else {
+    interpolate(input, length, output, olength);
+  }
+}
+
+static void fill_col_to_arr(uint8_t *img, int stride, int len, uint8_t *arr) {
+  int i;
+  uint8_t *iptr = img;
+  uint8_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *aptr++ = *iptr;
+  }
+}
+
+static void fill_arr_to_col(uint8_t *img, int stride, int len, uint8_t *arr) {
+  int i;
+  uint8_t *iptr = img;
+  uint8_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *iptr = *aptr++;
+  }
+}
+
+void vp9_resize_plane(const uint8_t *const input,
+                      int height,
+                      int width,
+                      int in_stride,
+                      uint8_t *output,
+                      int height2,
+                      int width2,
+                      int out_stride) {
+  int i;
+  uint8_t *intbuf = (uint8_t *)malloc(sizeof(uint8_t) * width2 * height);
+  uint8_t *tmpbuf = (uint8_t *)malloc(sizeof(uint8_t) *
+                                      (width < height ? height : width));
+  uint8_t *arrbuf = (uint8_t *)malloc(sizeof(uint8_t) * height);
+  uint8_t *arrbuf2 = (uint8_t *)malloc(sizeof(uint8_t) * height2);
+  if (intbuf == NULL || tmpbuf == NULL ||
+      arrbuf == NULL || arrbuf2 == NULL)
+    goto Error;
+  assert(width > 0);
+  assert(height > 0);
+  assert(width2 > 0);
+  assert(height2 > 0);
+  for (i = 0; i < height; ++i)
+    resize_multistep(input + in_stride * i, width,
+                     intbuf + width2 * i, width2, tmpbuf);
+  for (i = 0; i < width2; ++i) {
+    fill_col_to_arr(intbuf + i, width2, height, arrbuf);
+    resize_multistep(arrbuf, height, arrbuf2, height2, tmpbuf);
+    fill_arr_to_col(output + i, out_stride, height2, arrbuf2);
+  }
+
+ Error:
+  free(intbuf);
+  free(tmpbuf);
+  free(arrbuf);
+  free(arrbuf2);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_interpolate(const uint16_t *const input, int inlength,
+                               uint16_t *output, int outlength, int bd) {
+  const int64_t delta =
+      (((uint64_t)inlength << 32) + outlength / 2) / outlength;
+  const int64_t offset = inlength > outlength ?
+      (((int64_t)(inlength - outlength) << 31) + outlength / 2) / outlength :
+      -(((int64_t)(outlength - inlength) << 31) + outlength / 2) / outlength;
+  uint16_t *optr = output;
+  int x, x1, x2, sum, k, int_pel, sub_pel;
+  int64_t y;
+
+  const interp_kernel *interp_filters =
+      choose_interp_filter(inlength, outlength);
+
+  x = 0;
+  y = offset;
+  while ((y >> INTERP_PRECISION_BITS) < (INTERP_TAPS / 2 - 1)) {
+    x++;
+    y += delta;
+  }
+  x1 = x;
+  x = outlength - 1;
+  y = delta * x + offset;
+  while ((y >> INTERP_PRECISION_BITS) +
+         (int64_t)(INTERP_TAPS / 2) >= inlength) {
+    x--;
+    y -= delta;
+  }
+  x2 = x;
+  if (x1 > x2) {
+    for (x = 0, y = offset; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k) {
+        const int pk = int_pel - INTERP_TAPS / 2 + 1 + k;
+        sum += filter[k] *
+            input[(pk < 0 ? 0 : (pk >= inlength ? inlength - 1 : pk))];
+      }
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+  } else {
+    // Initial part.
+    for (x = 0, y = offset; x < x1; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] *
+            input[(int_pel - INTERP_TAPS / 2 + 1 + k < 0 ?
+                   0 : int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+    // Middle part.
+    for (; x <= x2; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[int_pel - INTERP_TAPS / 2 + 1 + k];
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+    // End part.
+    for (; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k >=
+                                  inlength ?  inlength - 1 :
+                                  int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+  }
+}
+
+static void highbd_down2_symeven(const uint16_t *const input, int length,
+                                 uint16_t *output, int bd) {
+  // Actual filter len = 2 * filter_len_half.
+  static const int16_t *filter = vp9_down2_symeven_half_filter;
+  const int filter_len_half = sizeof(vp9_down2_symeven_half_filter) / 2;
+  int i, j;
+  uint16_t *optr = output;
+  int l1 = filter_len_half;
+  int l2 = (length - filter_len_half);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  }
+}
+
+static void highbd_down2_symodd(const uint16_t *const input, int length,
+                              uint16_t *output, int bd) {
+  // Actual filter len = 2 * filter_len_half - 1.
+  static const int16_t *filter = vp9_down2_symodd_half_filter;
+  const int filter_len_half = sizeof(vp9_down2_symodd_half_filter) / 2;
+  int i, j;
+  uint16_t *optr = output;
+  int l1 = filter_len_half - 1;
+  int l2 = (length - filter_len_half + 1);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + j >= length ? length - 1 : i + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[(i + j >= length ? length - 1 : i + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  }
+}
+
+static void highbd_resize_multistep(const uint16_t *const input,
+                                    int length,
+                                    uint16_t *output,
+                                    int olength,
+                                    uint16_t *otmp,
+                                    int bd) {
+  int steps;
+  if (length == olength) {
+    memcpy(output, input, sizeof(output[0]) * length);
+    return;
+  }
+  steps = get_down2_steps(length, olength);
+
+  if (steps > 0) {
+    int s;
+    uint16_t *out = NULL;
+    uint16_t *otmp2;
+    int filteredlength = length;
+
+    assert(otmp != NULL);
+    otmp2 = otmp + get_down2_length(length, 1);
+    for (s = 0; s < steps; ++s) {
+      const int proj_filteredlength = get_down2_length(filteredlength, 1);
+      const uint16_t *const in = (s == 0 ? input : out);
+      if (s == steps - 1 && proj_filteredlength == olength)
+        out = output;
+      else
+        out = (s & 1 ? otmp2 : otmp);
+      if (filteredlength & 1)
+        highbd_down2_symodd(in, filteredlength, out, bd);
+      else
+        highbd_down2_symeven(in, filteredlength, out, bd);
+      filteredlength = proj_filteredlength;
+    }
+    if (filteredlength != olength) {
+      highbd_interpolate(out, filteredlength, output, olength, bd);
+    }
+  } else {
+    highbd_interpolate(input, length, output, olength, bd);
+  }
+}
+
+static void highbd_fill_col_to_arr(uint16_t *img, int stride, int len,
+                                   uint16_t *arr) {
+  int i;
+  uint16_t *iptr = img;
+  uint16_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *aptr++ = *iptr;
+  }
+}
+
+static void highbd_fill_arr_to_col(uint16_t *img, int stride, int len,
+                                   uint16_t *arr) {
+  int i;
+  uint16_t *iptr = img;
+  uint16_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *iptr = *aptr++;
+  }
+}
+
+void vp9_highbd_resize_plane(const uint8_t *const input,
+                             int height,
+                             int width,
+                             int in_stride,
+                             uint8_t *output,
+                             int height2,
+                             int width2,
+                             int out_stride,
+                             int bd) {
+  int i;
+  uint16_t *intbuf = (uint16_t *)malloc(sizeof(uint16_t) * width2 * height);
+  uint16_t *tmpbuf = (uint16_t *)malloc(sizeof(uint16_t) *
+                                        (width < height ? height : width));
+  uint16_t *arrbuf = (uint16_t *)malloc(sizeof(uint16_t) * height);
+  uint16_t *arrbuf2 = (uint16_t *)malloc(sizeof(uint16_t) * height2);
+  if (intbuf == NULL || tmpbuf == NULL ||
+      arrbuf == NULL || arrbuf2 == NULL)
+    goto Error;
+  for (i = 0; i < height; ++i) {
+    highbd_resize_multistep(CONVERT_TO_SHORTPTR(input + in_stride * i), width,
+                            intbuf + width2 * i, width2, tmpbuf, bd);
+  }
+  for (i = 0; i < width2; ++i) {
+    highbd_fill_col_to_arr(intbuf + i, width2, height, arrbuf);
+    highbd_resize_multistep(arrbuf, height, arrbuf2, height2, tmpbuf,
+                            bd);
+    highbd_fill_arr_to_col(CONVERT_TO_SHORTPTR(output + i), out_stride, height2,
+                           arrbuf2);
+  }
+
+ Error:
+  free(intbuf);
+  free(tmpbuf);
+  free(arrbuf);
+  free(arrbuf2);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_resize_frame420(const uint8_t *const y,
+                         int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride,
+                         int height, int width,
+                         uint8_t *oy, int oy_stride,
+                         uint8_t *ou, uint8_t *ov, int ouv_stride,
+                         int oheight, int owidth) {
+  vp9_resize_plane(y, height, width, y_stride,
+                   oy, oheight, owidth, oy_stride);
+  vp9_resize_plane(u, height / 2, width / 2, uv_stride,
+                   ou, oheight / 2, owidth / 2, ouv_stride);
+  vp9_resize_plane(v, height / 2, width / 2, uv_stride,
+                   ov, oheight / 2, owidth / 2, ouv_stride);
+}
+
+void vp9_resize_frame422(const uint8_t *const y, int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride,
+                         int height, int width,
+                         uint8_t *oy, int oy_stride,
+                         uint8_t *ou, uint8_t *ov, int ouv_stride,
+                         int oheight, int owidth) {
+  vp9_resize_plane(y, height, width, y_stride,
+                   oy, oheight, owidth, oy_stride);
+  vp9_resize_plane(u, height, width / 2, uv_stride,
+                   ou, oheight, owidth / 2, ouv_stride);
+  vp9_resize_plane(v, height, width / 2, uv_stride,
+                   ov, oheight, owidth / 2, ouv_stride);
+}
+
+void vp9_resize_frame444(const uint8_t *const y, int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride,
+                         int height, int width,
+                         uint8_t *oy, int oy_stride,
+                         uint8_t *ou, uint8_t *ov, int ouv_stride,
+                         int oheight, int owidth) {
+  vp9_resize_plane(y, height, width, y_stride,
+                   oy, oheight, owidth, oy_stride);
+  vp9_resize_plane(u, height, width, uv_stride,
+                   ou, oheight, owidth, ouv_stride);
+  vp9_resize_plane(v, height, width, uv_stride,
+                   ov, oheight, owidth, ouv_stride);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_resize_frame420(const uint8_t *const y,
+                                int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride,
+                                int height, int width,
+                                uint8_t *oy, int oy_stride,
+                                uint8_t *ou, uint8_t *ov, int ouv_stride,
+                                int oheight, int owidth, int bd) {
+  vp9_highbd_resize_plane(y, height, width, y_stride,
+                          oy, oheight, owidth, oy_stride, bd);
+  vp9_highbd_resize_plane(u, height / 2, width / 2, uv_stride,
+                          ou, oheight / 2, owidth / 2, ouv_stride, bd);
+  vp9_highbd_resize_plane(v, height / 2, width / 2, uv_stride,
+                          ov, oheight / 2, owidth / 2, ouv_stride, bd);
+}
+
+void vp9_highbd_resize_frame422(const uint8_t *const y, int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride,
+                                int height, int width,
+                                uint8_t *oy, int oy_stride,
+                                uint8_t *ou, uint8_t *ov, int ouv_stride,
+                                int oheight, int owidth, int bd) {
+  vp9_highbd_resize_plane(y, height, width, y_stride,
+                          oy, oheight, owidth, oy_stride, bd);
+  vp9_highbd_resize_plane(u, height, width / 2, uv_stride,
+                          ou, oheight, owidth / 2, ouv_stride, bd);
+  vp9_highbd_resize_plane(v, height, width / 2, uv_stride,
+                          ov, oheight, owidth / 2, ouv_stride, bd);
+}
+
+void vp9_highbd_resize_frame444(const uint8_t *const y, int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride,
+                                int height, int width,
+                                uint8_t *oy, int oy_stride,
+                                uint8_t *ou, uint8_t *ov, int ouv_stride,
+                                int oheight, int owidth, int bd) {
+  vp9_highbd_resize_plane(y, height, width, y_stride,
+                          oy, oheight, owidth, oy_stride, bd);
+  vp9_highbd_resize_plane(u, height, width, uv_stride,
+                          ou, oheight, owidth, ouv_stride, bd);
+  vp9_highbd_resize_plane(v, height, width, uv_stride,
+                          ov, oheight, owidth, ouv_stride, bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vp9/encoder/vp9_resize.h b/libvpx/libvpx/vp9/encoder/vp9_resize.h
new file mode 100644
index 0000000..b5feb38
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_resize.h
@@ -0,0 +1,133 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_RESIZE_H_
+#define VP9_ENCODER_VP9_RESIZE_H_
+
+#include <stdio.h>
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_resize_plane(const uint8_t *const input,
+                      int height,
+                      int width,
+                      int in_stride,
+                      uint8_t *output,
+                      int height2,
+                      int width2,
+                      int out_stride);
+void vp9_resize_frame420(const uint8_t *const y,
+                         int y_stride,
+                         const uint8_t *const u,
+                         const uint8_t *const v,
+                         int uv_stride,
+                         int height,
+                         int width,
+                         uint8_t *oy,
+                         int oy_stride,
+                         uint8_t *ou,
+                         uint8_t *ov,
+                         int ouv_stride,
+                         int oheight,
+                         int owidth);
+void vp9_resize_frame422(const uint8_t *const y,
+                         int y_stride,
+                         const uint8_t *const u,
+                         const uint8_t *const v,
+                         int uv_stride,
+                         int height,
+                         int width,
+                         uint8_t *oy,
+                         int oy_stride,
+                         uint8_t *ou,
+                         uint8_t *ov,
+                         int ouv_stride,
+                         int oheight,
+                         int owidth);
+void vp9_resize_frame444(const uint8_t *const y,
+                         int y_stride,
+                         const uint8_t *const u,
+                         const uint8_t *const v,
+                         int uv_stride,
+                         int height,
+                         int width,
+                         uint8_t *oy,
+                         int oy_stride,
+                         uint8_t *ou,
+                         uint8_t *ov,
+                         int ouv_stride,
+                         int oheight,
+                         int owidth);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_resize_plane(const uint8_t *const input,
+                             int height,
+                             int width,
+                             int in_stride,
+                             uint8_t *output,
+                             int height2,
+                             int width2,
+                             int out_stride,
+                             int bd);
+void vp9_highbd_resize_frame420(const uint8_t *const y,
+                                int y_stride,
+                                const uint8_t *const u,
+                                const uint8_t *const v,
+                                int uv_stride,
+                                int height,
+                                int width,
+                                uint8_t *oy,
+                                int oy_stride,
+                                uint8_t *ou,
+                                uint8_t *ov,
+                                int ouv_stride,
+                                int oheight,
+                                int owidth,
+                                int bd);
+void vp9_highbd_resize_frame422(const uint8_t *const y,
+                                int y_stride,
+                                const uint8_t *const u,
+                                const uint8_t *const v,
+                                int uv_stride,
+                                int height,
+                                int width,
+                                uint8_t *oy,
+                                int oy_stride,
+                                uint8_t *ou,
+                                uint8_t *ov,
+                                int ouv_stride,
+                                int oheight,
+                                int owidth,
+                                int bd);
+void vp9_highbd_resize_frame444(const uint8_t *const y,
+                                int y_stride,
+                                const uint8_t *const u,
+                                const uint8_t *const v,
+                                int uv_stride,
+                                int height,
+                                int width,
+                                uint8_t *oy,
+                                int oy_stride,
+                                uint8_t *ou,
+                                uint8_t *ov,
+                                int ouv_stride,
+                                int oheight,
+                                int owidth,
+                                int bd);
+#endif    // CONFIG_VP9_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif    // VP9_ENCODER_VP9_RESIZE_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_segmentation.c b/libvpx/libvpx/vp9/encoder/vp9_segmentation.c
new file mode 100644
index 0000000..5a0a23d
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_segmentation.c
@@ -0,0 +1,281 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <limits.h>
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_tile_common.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_segmentation.h"
+
+void vp9_enable_segmentation(struct segmentation *seg) {
+  seg->enabled = 1;
+  seg->update_map = 1;
+  seg->update_data = 1;
+}
+
+void vp9_disable_segmentation(struct segmentation *seg) {
+  seg->enabled = 0;
+  seg->update_map = 0;
+  seg->update_data = 0;
+}
+
+void vp9_set_segment_data(struct segmentation *seg,
+                          signed char *feature_data,
+                          unsigned char abs_delta) {
+  seg->abs_delta = abs_delta;
+
+  memcpy(seg->feature_data, feature_data, sizeof(seg->feature_data));
+}
+void vp9_disable_segfeature(struct segmentation *seg, int segment_id,
+                            SEG_LVL_FEATURES feature_id) {
+  seg->feature_mask[segment_id] &= ~(1 << feature_id);
+}
+
+void vp9_clear_segdata(struct segmentation *seg, int segment_id,
+                       SEG_LVL_FEATURES feature_id) {
+  seg->feature_data[segment_id][feature_id] = 0;
+}
+
+// Based on set of segment counts calculate a probability tree
+static void calc_segtree_probs(int *segcounts, vpx_prob *segment_tree_probs) {
+  // Work out probabilities of each segment
+  const int c01 = segcounts[0] + segcounts[1];
+  const int c23 = segcounts[2] + segcounts[3];
+  const int c45 = segcounts[4] + segcounts[5];
+  const int c67 = segcounts[6] + segcounts[7];
+
+  segment_tree_probs[0] = get_binary_prob(c01 + c23, c45 + c67);
+  segment_tree_probs[1] = get_binary_prob(c01, c23);
+  segment_tree_probs[2] = get_binary_prob(c45, c67);
+  segment_tree_probs[3] = get_binary_prob(segcounts[0], segcounts[1]);
+  segment_tree_probs[4] = get_binary_prob(segcounts[2], segcounts[3]);
+  segment_tree_probs[5] = get_binary_prob(segcounts[4], segcounts[5]);
+  segment_tree_probs[6] = get_binary_prob(segcounts[6], segcounts[7]);
+}
+
+// Based on set of segment counts and probabilities calculate a cost estimate
+static int cost_segmap(int *segcounts, vpx_prob *probs) {
+  const int c01 = segcounts[0] + segcounts[1];
+  const int c23 = segcounts[2] + segcounts[3];
+  const int c45 = segcounts[4] + segcounts[5];
+  const int c67 = segcounts[6] + segcounts[7];
+  const int c0123 = c01 + c23;
+  const int c4567 = c45 + c67;
+
+  // Cost the top node of the tree
+  int cost = c0123 * vp9_cost_zero(probs[0]) +
+             c4567 * vp9_cost_one(probs[0]);
+
+  // Cost subsequent levels
+  if (c0123 > 0) {
+    cost += c01 * vp9_cost_zero(probs[1]) +
+            c23 * vp9_cost_one(probs[1]);
+
+    if (c01 > 0)
+      cost += segcounts[0] * vp9_cost_zero(probs[3]) +
+              segcounts[1] * vp9_cost_one(probs[3]);
+    if (c23 > 0)
+      cost += segcounts[2] * vp9_cost_zero(probs[4]) +
+              segcounts[3] * vp9_cost_one(probs[4]);
+  }
+
+  if (c4567 > 0) {
+    cost += c45 * vp9_cost_zero(probs[2]) +
+            c67 * vp9_cost_one(probs[2]);
+
+    if (c45 > 0)
+      cost += segcounts[4] * vp9_cost_zero(probs[5]) +
+              segcounts[5] * vp9_cost_one(probs[5]);
+    if (c67 > 0)
+      cost += segcounts[6] * vp9_cost_zero(probs[6]) +
+              segcounts[7] * vp9_cost_one(probs[6]);
+  }
+
+  return cost;
+}
+
+static void count_segs(const VP9_COMMON *cm, MACROBLOCKD *xd,
+                       const TileInfo *tile, MODE_INFO **mi,
+                       int *no_pred_segcounts,
+                       int (*temporal_predictor_count)[2],
+                       int *t_unpred_seg_counts,
+                       int bw, int bh, int mi_row, int mi_col) {
+  int segment_id;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  xd->mi = mi;
+  segment_id = xd->mi[0]->segment_id;
+
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
+
+  // Count the number of hits on each segment with no prediction
+  no_pred_segcounts[segment_id]++;
+
+  // Temporal prediction not allowed on key frames
+  if (cm->frame_type != KEY_FRAME) {
+    const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
+    // Test to see if the segment id matches the predicted value.
+    const int pred_segment_id = get_segment_id(cm, cm->last_frame_seg_map,
+                                               bsize, mi_row, mi_col);
+    const int pred_flag = pred_segment_id == segment_id;
+    const int pred_context = vp9_get_pred_context_seg_id(xd);
+
+    // Store the prediction status for this mb and update counts
+    // as appropriate
+    xd->mi[0]->seg_id_predicted = pred_flag;
+    temporal_predictor_count[pred_context][pred_flag]++;
+
+    // Update the "unpredicted" segment count
+    if (!pred_flag)
+      t_unpred_seg_counts[segment_id]++;
+  }
+}
+
+static void count_segs_sb(const VP9_COMMON *cm, MACROBLOCKD *xd,
+                          const TileInfo *tile, MODE_INFO **mi,
+                          int *no_pred_segcounts,
+                          int (*temporal_predictor_count)[2],
+                          int *t_unpred_seg_counts,
+                          int mi_row, int mi_col,
+                          BLOCK_SIZE bsize) {
+  const int mis = cm->mi_stride;
+  int bw, bh;
+  const int bs = num_8x8_blocks_wide_lookup[bsize], hbs = bs / 2;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  bw = num_8x8_blocks_wide_lookup[mi[0]->sb_type];
+  bh = num_8x8_blocks_high_lookup[mi[0]->sb_type];
+
+  if (bw == bs && bh == bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, bs, bs, mi_row, mi_col);
+  } else if (bw == bs && bh < bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
+    count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+               temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
+               mi_row + hbs, mi_col);
+  } else if (bw < bs && bh == bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
+    count_segs(cm, xd, tile, mi + hbs,
+               no_pred_segcounts, temporal_predictor_count, t_unpred_seg_counts,
+               hbs, bs, mi_row, mi_col + hbs);
+  } else {
+    const BLOCK_SIZE subsize = subsize_lookup[PARTITION_SPLIT][bsize];
+    int n;
+
+    assert(bw < bs && bh < bs);
+
+    for (n = 0; n < 4; n++) {
+      const int mi_dc = hbs * (n & 1);
+      const int mi_dr = hbs * (n >> 1);
+
+      count_segs_sb(cm, xd, tile, &mi[mi_dr * mis + mi_dc],
+                    no_pred_segcounts, temporal_predictor_count,
+                    t_unpred_seg_counts,
+                    mi_row + mi_dr, mi_col + mi_dc, subsize);
+    }
+  }
+}
+
+void vp9_choose_segmap_coding_method(VP9_COMMON *cm, MACROBLOCKD *xd) {
+  struct segmentation *seg = &cm->seg;
+
+  int no_pred_cost;
+  int t_pred_cost = INT_MAX;
+
+  int i, tile_col, mi_row, mi_col;
+
+  int temporal_predictor_count[PREDICTION_PROBS][2] = { { 0 } };
+  int no_pred_segcounts[MAX_SEGMENTS] = { 0 };
+  int t_unpred_seg_counts[MAX_SEGMENTS] = { 0 };
+
+  vpx_prob no_pred_tree[SEG_TREE_PROBS];
+  vpx_prob t_pred_tree[SEG_TREE_PROBS];
+  vpx_prob t_nopred_prob[PREDICTION_PROBS];
+
+  // Set default state for the segment tree probabilities and the
+  // temporal coding probabilities
+  memset(seg->tree_probs, 255, sizeof(seg->tree_probs));
+  memset(seg->pred_probs, 255, sizeof(seg->pred_probs));
+
+  // First of all generate stats regarding how well the last segment map
+  // predicts this one
+  for (tile_col = 0; tile_col < 1 << cm->log2_tile_cols; tile_col++) {
+    TileInfo tile;
+    MODE_INFO **mi_ptr;
+    vp9_tile_init(&tile, cm, 0, tile_col);
+
+    mi_ptr = cm->mi_grid_visible + tile.mi_col_start;
+    for (mi_row = 0; mi_row < cm->mi_rows;
+         mi_row += 8, mi_ptr += 8 * cm->mi_stride) {
+      MODE_INFO **mi = mi_ptr;
+      for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
+           mi_col += 8, mi += 8)
+        count_segs_sb(cm, xd, &tile, mi, no_pred_segcounts,
+                      temporal_predictor_count, t_unpred_seg_counts,
+                      mi_row, mi_col, BLOCK_64X64);
+    }
+  }
+
+  // Work out probability tree for coding segments without prediction
+  // and the cost.
+  calc_segtree_probs(no_pred_segcounts, no_pred_tree);
+  no_pred_cost = cost_segmap(no_pred_segcounts, no_pred_tree);
+
+  // Key frames cannot use temporal prediction
+  if (!frame_is_intra_only(cm)) {
+    // Work out probability tree for coding those segments not
+    // predicted using the temporal method and the cost.
+    calc_segtree_probs(t_unpred_seg_counts, t_pred_tree);
+    t_pred_cost = cost_segmap(t_unpred_seg_counts, t_pred_tree);
+
+    // Add in the cost of the signaling for each prediction context.
+    for (i = 0; i < PREDICTION_PROBS; i++) {
+      const int count0 = temporal_predictor_count[i][0];
+      const int count1 = temporal_predictor_count[i][1];
+
+      t_nopred_prob[i] = get_binary_prob(count0, count1);
+
+      // Add in the predictor signaling cost
+      t_pred_cost += count0 * vp9_cost_zero(t_nopred_prob[i]) +
+                     count1 * vp9_cost_one(t_nopred_prob[i]);
+    }
+  }
+
+  // Now choose which coding method to use.
+  if (t_pred_cost < no_pred_cost) {
+    seg->temporal_update = 1;
+    memcpy(seg->tree_probs, t_pred_tree, sizeof(t_pred_tree));
+    memcpy(seg->pred_probs, t_nopred_prob, sizeof(t_nopred_prob));
+  } else {
+    seg->temporal_update = 0;
+    memcpy(seg->tree_probs, no_pred_tree, sizeof(no_pred_tree));
+  }
+}
+
+void vp9_reset_segment_features(struct segmentation *seg) {
+  // Set up default state for MB feature flags
+  seg->enabled = 0;
+  seg->update_map = 0;
+  seg->update_data = 0;
+  memset(seg->tree_probs, 255, sizeof(seg->tree_probs));
+  vp9_clearall_segfeatures(seg);
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_segmentation.h b/libvpx/libvpx/vp9/encoder/vp9_segmentation.h
new file mode 100644
index 0000000..8c6944a
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_segmentation.h
@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_SEGMENTATION_H_
+#define VP9_ENCODER_VP9_SEGMENTATION_H_
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/encoder/vp9_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_enable_segmentation(struct segmentation *seg);
+void vp9_disable_segmentation(struct segmentation *seg);
+
+void vp9_disable_segfeature(struct segmentation *seg,
+                            int segment_id,
+                            SEG_LVL_FEATURES feature_id);
+void vp9_clear_segdata(struct segmentation *seg,
+                       int segment_id,
+                       SEG_LVL_FEATURES feature_id);
+
+// The values given for each segment can be either deltas (from the default
+// value chosen for the frame) or absolute values.
+//
+// Valid range for abs values is (0-127 for MB_LVL_ALT_Q), (0-63 for
+// SEGMENT_ALT_LF)
+// Valid range for delta values are (+/-127 for MB_LVL_ALT_Q), (+/-63 for
+// SEGMENT_ALT_LF)
+//
+// abs_delta = SEGMENT_DELTADATA (deltas) abs_delta = SEGMENT_ABSDATA (use
+// the absolute values given).
+void vp9_set_segment_data(struct segmentation *seg, signed char *feature_data,
+                          unsigned char abs_delta);
+
+void vp9_choose_segmap_coding_method(VP9_COMMON *cm, MACROBLOCKD *xd);
+
+void vp9_reset_segment_features(struct segmentation *seg);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SEGMENTATION_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_skin_detection.c b/libvpx/libvpx/vp9/encoder/vp9_skin_detection.c
new file mode 100644
index 0000000..23a5fc7
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_skin_detection.c
@@ -0,0 +1,208 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <math.h>
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_skin_detection.h"
+
+#define MODEL_MODE 1
+
+// Fixed-point skin color model parameters.
+static const int skin_mean[5][2] = {
+    {7463, 9614}, {6400, 10240}, {7040, 10240}, {8320, 9280}, {6800, 9614}};
+static const int skin_inv_cov[4] = {4107, 1663, 1663, 2157};  // q16
+static const int skin_threshold[6] = {1570636, 1400000, 800000, 800000, 800000,
+    800000};  // q18
+
+// Thresholds on luminance.
+static const int y_low = 40;
+static const int y_high = 220;
+
+// Evaluates the Mahalanobis distance measure for the input CbCr values.
+static int evaluate_skin_color_difference(int cb, int cr, int idx) {
+  const int cb_q6 = cb << 6;
+  const int cr_q6 = cr << 6;
+  const int cb_diff_q12 =
+      (cb_q6 - skin_mean[idx][0]) * (cb_q6 - skin_mean[idx][0]);
+  const int cbcr_diff_q12 =
+      (cb_q6 - skin_mean[idx][0]) * (cr_q6 - skin_mean[idx][1]);
+  const int cr_diff_q12 =
+      (cr_q6 - skin_mean[idx][1]) * (cr_q6 - skin_mean[idx][1]);
+  const int cb_diff_q2 = (cb_diff_q12 + (1 << 9)) >> 10;
+  const int cbcr_diff_q2 = (cbcr_diff_q12 + (1 << 9)) >> 10;
+  const int cr_diff_q2 = (cr_diff_q12 + (1 << 9)) >> 10;
+  const int skin_diff = skin_inv_cov[0] * cb_diff_q2 +
+      skin_inv_cov[1] * cbcr_diff_q2 +
+      skin_inv_cov[2] * cbcr_diff_q2 +
+      skin_inv_cov[3] * cr_diff_q2;
+  return skin_diff;
+}
+
+int vp9_skin_pixel(const uint8_t y, const uint8_t cb, const uint8_t cr,
+                   int motion) {
+  if (y < y_low || y > y_high) {
+    return 0;
+  } else {
+    if (MODEL_MODE == 0) {
+      return (evaluate_skin_color_difference(cb, cr, 0) < skin_threshold[0]);
+    } else {
+      int i = 0;
+      // Exit on grey.
+      if (cb == 128 && cr == 128)
+        return 0;
+      // Exit on very strong cb.
+      if (cb > 150 && cr < 110)
+        return 0;
+      for (; i < 5; i++) {
+        int skin_color_diff = evaluate_skin_color_difference(cb, cr, i);
+        if (skin_color_diff < skin_threshold[i + 1]) {
+           if (y < 60 && skin_color_diff > 3 * (skin_threshold[i + 1] >> 2))
+             return 0;
+           else if (motion == 0 &&
+                    skin_color_diff > (skin_threshold[i + 1] >> 1))
+             return 0;
+           else
+            return 1;
+        }
+        // Exit if difference is much large than the threshold.
+        if (skin_color_diff > (skin_threshold[i + 1] << 3)) {
+          return 0;
+        }
+      }
+      return 0;
+    }
+  }
+}
+
+int vp9_compute_skin_block(const uint8_t *y, const uint8_t *u, const uint8_t *v,
+                           int stride, int strideuv, int bsize,
+                           int consec_zeromv, int curr_motion_magn) {
+  // No skin if block has been zero/small motion for long consecutive time.
+  if (consec_zeromv > 60 && curr_motion_magn == 0) {
+    return 0;
+  } else {
+    int motion = 1;
+    // Take center pixel in block to determine is_skin.
+    const int y_width_shift = (4 << b_width_log2_lookup[bsize]) >> 1;
+    const int y_height_shift = (4 << b_height_log2_lookup[bsize]) >> 1;
+    const int uv_width_shift = y_width_shift >> 1;
+    const int uv_height_shift = y_height_shift >> 1;
+    const uint8_t ysource = y[y_height_shift * stride + y_width_shift];
+    const uint8_t usource = u[uv_height_shift * strideuv + uv_width_shift];
+    const uint8_t vsource = v[uv_height_shift * strideuv + uv_width_shift];
+    if (consec_zeromv > 25 && curr_motion_magn == 0)
+      motion = 0;
+    return vp9_skin_pixel(ysource, usource, vsource, motion);
+  }
+}
+
+
+#ifdef OUTPUT_YUV_SKINMAP
+// For viewing skin map on input source.
+void vp9_compute_skin_map(VP9_COMP *const cpi, FILE *yuv_skinmap_file) {
+  int i, j, mi_row, mi_col, num_bl;
+  VP9_COMMON *const cm = &cpi->common;
+  uint8_t *y;
+  const uint8_t *src_y = cpi->Source->y_buffer;
+  const uint8_t *src_u = cpi->Source->u_buffer;
+  const uint8_t *src_v = cpi->Source->v_buffer;
+  const int src_ystride = cpi->Source->y_stride;
+  const int src_uvstride = cpi->Source->uv_stride;
+  int y_bsize = 16;  // Use 8x8 or 16x16.
+  int uv_bsize = y_bsize >> 1;
+  int ypos = y_bsize >> 1;
+  int uvpos = uv_bsize >> 1;
+  int shy = (y_bsize == 8) ? 3 : 4;
+  int shuv = shy - 1;
+  int fac = y_bsize / 8;
+  // Use center pixel or average of center 2x2 pixels.
+  int mode_filter = 0;
+  YV12_BUFFER_CONFIG skinmap;
+  memset(&skinmap, 0, sizeof(YV12_BUFFER_CONFIG));
+  if (vpx_alloc_frame_buffer(&skinmap, cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment)) {
+      vpx_free_frame_buffer(&skinmap);
+      return;
+  }
+  memset(skinmap.buffer_alloc, 128, skinmap.frame_size);
+  y = skinmap.y_buffer;
+  // Loop through blocks and set skin map based on center pixel of block.
+  // Set y to white for skin block, otherwise set to source with gray scale.
+  // Ignore rightmost/bottom boundary blocks.
+  for (mi_row = 0; mi_row < cm->mi_rows - 1; mi_row += fac) {
+    num_bl = 0;
+    for (mi_col = 0; mi_col < cm->mi_cols - 1; mi_col += fac) {
+      int is_skin = 0;
+      if (mode_filter == 1) {
+        // Use 2x2 average at center.
+        uint8_t ysource = src_y[ypos * src_ystride + ypos];
+        uint8_t usource = src_u[uvpos * src_uvstride + uvpos];
+        uint8_t vsource = src_v[uvpos * src_uvstride + uvpos];
+        uint8_t ysource2 = src_y[(ypos + 1) * src_ystride + ypos];
+        uint8_t usource2 = src_u[(uvpos + 1) * src_uvstride + uvpos];
+        uint8_t vsource2 = src_v[(uvpos + 1) * src_uvstride + uvpos];
+        uint8_t ysource3 = src_y[ypos * src_ystride + (ypos + 1)];
+        uint8_t usource3 = src_u[uvpos * src_uvstride + (uvpos  + 1)];
+        uint8_t vsource3 = src_v[uvpos * src_uvstride + (uvpos +  1)];
+        uint8_t ysource4 = src_y[(ypos + 1) * src_ystride + (ypos + 1)];
+        uint8_t usource4 = src_u[(uvpos + 1) * src_uvstride + (uvpos  + 1)];
+        uint8_t vsource4 = src_v[(uvpos + 1) * src_uvstride + (uvpos +  1)];
+        ysource = (ysource + ysource2 + ysource3 + ysource4) >> 2;
+        usource = (usource + usource2 + usource3 + usource4) >> 2;
+        vsource = (vsource + vsource2 + vsource3 + vsource4) >> 2;
+        is_skin = vp9_skin_pixel(ysource, usource, vsource, 1);
+      } else {
+        int block_size = BLOCK_8X8;
+        int consec_zeromv = 0;
+        int bl_index = mi_row * cm->mi_cols + mi_col;
+        int bl_index1 = bl_index + 1;
+        int bl_index2 = bl_index + cm->mi_cols;
+        int bl_index3 = bl_index2 + 1;
+        if (y_bsize == 8)
+          consec_zeromv = cpi->consec_zero_mv[bl_index];
+        else
+          consec_zeromv = VPXMIN(cpi->consec_zero_mv[bl_index],
+                                 VPXMIN(cpi->consec_zero_mv[bl_index1],
+                                 VPXMIN(cpi->consec_zero_mv[bl_index2],
+                                 cpi->consec_zero_mv[bl_index3])));
+        if (y_bsize == 16)
+          block_size = BLOCK_16X16;
+        is_skin  = vp9_compute_skin_block(src_y, src_u, src_v, src_ystride,
+                                          src_uvstride, block_size,
+                                          consec_zeromv,
+                                          0);
+      }
+      for (i = 0; i < y_bsize; i++) {
+        for (j = 0; j < y_bsize; j++) {
+          if (is_skin)
+            y[i * src_ystride + j] = 255;
+          else
+            y[i * src_ystride + j] = src_y[i * src_ystride + j];
+        }
+      }
+      num_bl++;
+      y += y_bsize;
+      src_y += y_bsize;
+      src_u += uv_bsize;
+      src_v += uv_bsize;
+    }
+    y += (src_ystride << shy) - (num_bl << shy);
+    src_y += (src_ystride << shy) - (num_bl << shy);
+    src_u += (src_uvstride << shuv) - (num_bl << shuv);
+    src_v += (src_uvstride << shuv) - (num_bl << shuv);
+  }
+  vp9_write_yuv_frame_420(&skinmap, yuv_skinmap_file);
+  vpx_free_frame_buffer(&skinmap);
+}
+#endif
diff --git a/libvpx/libvpx/vp9/encoder/vp9_skin_detection.h b/libvpx/libvpx/vp9/encoder/vp9_skin_detection.h
new file mode 100644
index 0000000..c77382d
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_skin_detection.h
@@ -0,0 +1,41 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_SKIN_MAP_H_
+#define VP9_ENCODER_VP9_SKIN_MAP_H_
+
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9_COMP;
+
+// #define OUTPUT_YUV_SKINMAP
+
+int vp9_skin_pixel(const uint8_t y, const uint8_t cb, const uint8_t cr,
+                   int motion);
+
+int vp9_compute_skin_block(const uint8_t *y, const uint8_t *u, const uint8_t *v,
+                           int stride, int strideuv, int bsize,
+                           int consec_zeromv, int curr_motion_magn);
+
+#ifdef OUTPUT_YUV_SKINMAP
+// For viewing skin map on input source.
+void vp9_compute_skin_map(struct VP9_COMP *const cpi, FILE *yuv_skinmap_file);
+extern void vp9_write_yuv_frame_420(YV12_BUFFER_CONFIG *s, FILE *f);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SKIN_MAP_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_speed_features.c b/libvpx/libvpx/vp9/encoder/vp9_speed_features.c
new file mode 100644
index 0000000..e7f04a2
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_speed_features.c
@@ -0,0 +1,650 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_speed_features.h"
+#include "vp9/encoder/vp9_rdopt.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+// Mesh search patters for various speed settings
+static MESH_PATTERN best_quality_mesh_pattern[MAX_MESH_STEP] =
+    {{64, 4}, {28, 2}, {15, 1}, {7, 1}};
+
+#define MAX_MESH_SPEED 5  // Max speed setting for mesh motion method
+static MESH_PATTERN good_quality_mesh_patterns[MAX_MESH_SPEED + 1]
+                                              [MAX_MESH_STEP] =
+    {{{64, 8}, {28, 4}, {15, 1}, {7, 1}},
+     {{64, 8}, {28, 4}, {15, 1}, {7, 1}},
+     {{64, 8},  {14, 2}, {7, 1},  {7, 1}},
+     {{64, 16}, {24, 8}, {12, 4}, {7, 1}},
+     {{64, 16}, {24, 8}, {12, 4}, {7, 1}},
+     {{64, 16}, {24, 8}, {12, 4}, {7, 1}},
+    };
+static unsigned char good_quality_max_mesh_pct[MAX_MESH_SPEED + 1] =
+    {50, 25, 15, 5, 1, 1};
+
+// Intra only frames, golden frames (except alt ref overlays) and
+// alt ref frames tend to be coded at a higher than ambient quality
+static int frame_is_boosted(const VP9_COMP *cpi) {
+  return frame_is_kf_gf_arf(cpi) || vp9_is_upper_layer_key_frame(cpi);
+}
+
+// Sets a partition size down to which the auto partition code will always
+// search (can go lower), based on the image dimensions. The logic here
+// is that the extent to which ringing artefacts are offensive, depends
+// partly on the screen area that over which they propogate. Propogation is
+// limited by transform block size but the screen area take up by a given block
+// size will be larger for a small image format stretched to full screen.
+static BLOCK_SIZE set_partition_min_limit(VP9_COMMON *const cm) {
+  unsigned int screen_area = (cm->width * cm->height);
+
+  // Select block size based on image format size.
+  if (screen_area < 1280 * 720) {
+    // Formats smaller in area than 720P
+    return BLOCK_4X4;
+  } else if (screen_area < 1920 * 1080) {
+    // Format >= 720P and < 1080P
+    return BLOCK_8X8;
+  } else {
+    // Formats 1080P and up
+    return BLOCK_16X16;
+  }
+}
+
+static void set_good_speed_feature_framesize_dependent(VP9_COMP *cpi,
+                                                       SPEED_FEATURES *sf,
+                                                       int speed) {
+  VP9_COMMON *const cm = &cpi->common;
+
+  if (speed >= 1) {
+    if (VPXMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
+                                              : DISABLE_ALL_INTER_SPLIT;
+      sf->partition_search_breakout_dist_thr = (1 << 23);
+    } else {
+      sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
+      sf->partition_search_breakout_dist_thr = (1 << 21);
+    }
+  }
+
+  if (speed >= 2) {
+    if (VPXMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
+                                              : DISABLE_ALL_INTER_SPLIT;
+      sf->adaptive_pred_interp_filter = 0;
+      sf->partition_search_breakout_dist_thr = (1 << 24);
+      sf->partition_search_breakout_rate_thr = 120;
+    } else {
+      sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
+      sf->partition_search_breakout_dist_thr = (1 << 22);
+      sf->partition_search_breakout_rate_thr = 100;
+    }
+    sf->rd_auto_partition_min_limit = set_partition_min_limit(cm);
+  }
+
+  if (speed >= 3) {
+    if (VPXMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = DISABLE_ALL_SPLIT;
+      sf->schedule_mode_search = cm->base_qindex < 220 ? 1 : 0;
+      sf->partition_search_breakout_dist_thr = (1 << 25);
+      sf->partition_search_breakout_rate_thr = 200;
+    } else {
+      sf->max_intra_bsize = BLOCK_32X32;
+      sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT;
+      sf->schedule_mode_search = cm->base_qindex < 175 ? 1 : 0;
+      sf->partition_search_breakout_dist_thr = (1 << 23);
+      sf->partition_search_breakout_rate_thr = 120;
+    }
+  }
+
+  // If this is a two pass clip that fits the criteria for animated or
+  // graphics content then reset disable_split_mask for speeds 1-4.
+  // Also if the image edge is internal to the coded area.
+  if ((speed >= 1) && (cpi->oxcf.pass == 2) &&
+      ((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ||
+       (vp9_internal_image_edge(cpi)))) {
+    sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
+  }
+
+  if (speed >= 4) {
+    if (VPXMIN(cm->width, cm->height) >= 720) {
+      sf->partition_search_breakout_dist_thr = (1 << 26);
+    } else {
+      sf->partition_search_breakout_dist_thr = (1 << 24);
+    }
+    sf->disable_split_mask = DISABLE_ALL_SPLIT;
+  }
+}
+
+static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
+                                   SPEED_FEATURES *sf, int speed) {
+  const int boosted = frame_is_boosted(cpi);
+
+  sf->partition_search_breakout_dist_thr = (1 << 20);
+  sf->partition_search_breakout_rate_thr = 80;
+  sf->tx_size_search_breakout = 1;
+  sf->adaptive_rd_thresh = 1;
+  sf->allow_skip_recode = 1;
+  sf->less_rectangular_check = 1;
+  sf->use_square_partition_only = !frame_is_boosted(cpi);
+  sf->use_square_only_threshold = BLOCK_16X16;
+
+  if (speed >= 1) {
+    if ((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ||
+        vp9_internal_image_edge(cpi)) {
+      sf->use_square_partition_only = !frame_is_boosted(cpi);
+    } else {
+      sf->use_square_partition_only = !frame_is_intra_only(cm);
+    }
+    sf->use_square_only_threshold = BLOCK_4X4;
+
+    sf->less_rectangular_check  = 1;
+
+    sf->use_rd_breakout = 1;
+    sf->adaptive_motion_search = 1;
+    sf->mv.auto_mv_step_size = 1;
+    sf->adaptive_rd_thresh = 2;
+    sf->mv.subpel_iters_per_step = 1;
+    sf->mode_skip_start = 10;
+    sf->adaptive_pred_interp_filter = 1;
+
+    sf->recode_loop = ALLOW_RECODE_KFARFGF;
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
+  }
+
+  if (speed >= 2) {
+    sf->tx_size_search_method = frame_is_boosted(cpi) ? USE_FULL_RD
+                                                      : USE_LARGESTALL;
+
+    // Reference masking is not supported in dynamic scaling mode.
+    sf->reference_masking = cpi->oxcf.resize_mode != RESIZE_DYNAMIC ? 1 : 0;
+
+    sf->mode_search_skip_flags = (cm->frame_type == KEY_FRAME) ? 0 :
+                                 FLAG_SKIP_INTRA_DIRMISMATCH |
+                                 FLAG_SKIP_INTRA_BESTINTER |
+                                 FLAG_SKIP_COMP_BESTINTRA |
+                                 FLAG_SKIP_INTRA_LOWVAR;
+    sf->disable_filter_search_var_thresh = 100;
+    sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
+    sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
+    sf->allow_partition_search_skip = 1;
+  }
+
+  if (speed >= 3) {
+    sf->use_square_partition_only = !frame_is_intra_only(cm);
+    sf->tx_size_search_method = frame_is_intra_only(cm) ? USE_FULL_RD
+                                                        : USE_LARGESTALL;
+    sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED;
+    sf->adaptive_pred_interp_filter = 0;
+    sf->adaptive_mode_search = 1;
+    sf->cb_partition_search = !boosted;
+    sf->cb_pred_filter_search = 1;
+    sf->alt_ref_search_fp = 1;
+    sf->recode_loop = ALLOW_RECODE_KFMAXBW;
+    sf->adaptive_rd_thresh = 3;
+    sf->mode_skip_start = 6;
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
+    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC;
+    sf->adaptive_interp_filter_search = 1;
+  }
+
+  if (speed >= 4) {
+    sf->use_square_partition_only = 1;
+    sf->tx_size_search_method = USE_LARGESTALL;
+    sf->mv.search_method = BIGDIA;
+    sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
+    sf->adaptive_rd_thresh = 4;
+    if (cm->frame_type != KEY_FRAME)
+      sf->mode_search_skip_flags |= FLAG_EARLY_TERMINATE;
+    sf->disable_filter_search_var_thresh = 200;
+    sf->use_lp32x32fdct = 1;
+    sf->use_fast_coef_updates = ONE_LOOP_REDUCED;
+    sf->use_fast_coef_costing = 1;
+    sf->motion_field_mode_search = !boosted;
+    sf->partition_search_breakout_rate_thr = 300;
+  }
+
+  if (speed >= 5) {
+    int i;
+    sf->optimize_coefficients = 0;
+    sf->mv.search_method = HEX;
+    sf->disable_filter_search_var_thresh = 500;
+    for (i = 0; i < TX_SIZES; ++i) {
+      sf->intra_y_mode_mask[i] = INTRA_DC;
+      sf->intra_uv_mode_mask[i] = INTRA_DC;
+    }
+    sf->partition_search_breakout_rate_thr = 500;
+    sf->mv.reduce_first_step_size = 1;
+    sf->simple_model_rd_from_var = 1;
+  }
+}
+
+static void set_rt_speed_feature_framesize_dependent(VP9_COMP *cpi,
+    SPEED_FEATURES *sf, int speed) {
+  VP9_COMMON *const cm = &cpi->common;
+
+  if (speed >= 1) {
+    if (VPXMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
+                                              : DISABLE_ALL_INTER_SPLIT;
+    } else {
+      sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
+    }
+  }
+
+  if (speed >= 2) {
+    if (VPXMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
+                                              : DISABLE_ALL_INTER_SPLIT;
+    } else {
+      sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
+    }
+  }
+
+  if (speed >= 5) {
+    if (VPXMIN(cm->width, cm->height) >= 720) {
+      sf->partition_search_breakout_dist_thr = (1 << 25);
+    } else {
+      sf->partition_search_breakout_dist_thr = (1 << 23);
+    }
+  }
+
+  if (speed >= 7) {
+    sf->encode_breakout_thresh = (VPXMIN(cm->width, cm->height) >= 720) ?
+        800 : 300;
+  }
+}
+
+static void set_rt_speed_feature(VP9_COMP *cpi, SPEED_FEATURES *sf,
+                                 int speed, vp9e_tune_content content) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int is_keyframe = cm->frame_type == KEY_FRAME;
+  const int frames_since_key = is_keyframe ? 0 : cpi->rc.frames_since_key;
+  sf->static_segmentation = 0;
+  sf->adaptive_rd_thresh = 1;
+  sf->use_fast_coef_costing = 1;
+  sf->allow_exhaustive_searches = 0;
+  sf->exhaustive_searches_thresh = INT_MAX;
+
+  if (speed >= 1) {
+    sf->use_square_partition_only = !frame_is_intra_only(cm);
+    sf->less_rectangular_check = 1;
+    sf->tx_size_search_method = frame_is_intra_only(cm) ? USE_FULL_RD
+                                                        : USE_LARGESTALL;
+
+    sf->use_rd_breakout = 1;
+
+    sf->adaptive_motion_search = 1;
+    sf->adaptive_pred_interp_filter = 1;
+    sf->mv.auto_mv_step_size = 1;
+    sf->adaptive_rd_thresh = 2;
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
+  }
+
+  if (speed >= 2) {
+    sf->mode_search_skip_flags = (cm->frame_type == KEY_FRAME) ? 0 :
+                                 FLAG_SKIP_INTRA_DIRMISMATCH |
+                                 FLAG_SKIP_INTRA_BESTINTER |
+                                 FLAG_SKIP_COMP_BESTINTRA |
+                                 FLAG_SKIP_INTRA_LOWVAR;
+    sf->adaptive_pred_interp_filter = 2;
+
+    // Reference masking only enabled for 1 spatial layer, and if none of the
+    // references have been scaled. The latter condition needs to be checked
+    // for external or internal dynamic resize.
+    sf->reference_masking = (cpi->svc.number_spatial_layers == 1);
+    if (sf->reference_masking == 1 &&
+        (cpi->external_resize == 1 ||
+         cpi->oxcf.resize_mode == RESIZE_DYNAMIC)) {
+      MV_REFERENCE_FRAME ref_frame;
+      static const int flag_list[4] =
+          {0, VP9_LAST_FLAG, VP9_GOLD_FLAG, VP9_ALT_FLAG};
+      for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+        const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
+        if (yv12 != NULL && (cpi->ref_frame_flags & flag_list[ref_frame])) {
+          const struct scale_factors *const scale_fac =
+              &cm->frame_refs[ref_frame - 1].sf;
+          if (vp9_is_scaled(scale_fac))
+            sf->reference_masking = 0;
+        }
+      }
+    }
+
+    sf->disable_filter_search_var_thresh = 50;
+    sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
+    sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
+    sf->lf_motion_threshold = LOW_MOTION_THRESHOLD;
+    sf->adjust_partitioning_from_last_frame = 1;
+    sf->last_partitioning_redo_frequency = 3;
+    sf->use_lp32x32fdct = 1;
+    sf->mode_skip_start = 11;
+    sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
+  }
+
+  if (speed >= 3) {
+    sf->use_square_partition_only = 1;
+    sf->disable_filter_search_var_thresh = 100;
+    sf->use_uv_intra_rd_estimate = 1;
+    sf->skip_encode_sb = 1;
+    sf->mv.subpel_iters_per_step = 1;
+    sf->adaptive_rd_thresh = 4;
+    sf->mode_skip_start = 6;
+    sf->allow_skip_recode = 0;
+    sf->optimize_coefficients = 0;
+    sf->disable_split_mask = DISABLE_ALL_SPLIT;
+    sf->lpf_pick = LPF_PICK_FROM_Q;
+  }
+
+  if (speed >= 4) {
+    int i;
+    sf->last_partitioning_redo_frequency = 4;
+    sf->adaptive_rd_thresh = 5;
+    sf->use_fast_coef_costing = 0;
+    sf->auto_min_max_partition_size = STRICT_NEIGHBORING_MIN_MAX;
+    sf->adjust_partitioning_from_last_frame =
+        cm->last_frame_type != cm->frame_type || (0 ==
+        (frames_since_key + 1) % sf->last_partitioning_redo_frequency);
+    sf->mv.subpel_force_stop = 1;
+    for (i = 0; i < TX_SIZES; i++) {
+      sf->intra_y_mode_mask[i] = INTRA_DC_H_V;
+      sf->intra_uv_mode_mask[i] = INTRA_DC;
+    }
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
+    sf->frame_parameter_update = 0;
+    sf->mv.search_method = FAST_HEX;
+
+    sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEAR_NEW;
+    sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST;
+    sf->max_intra_bsize = BLOCK_32X32;
+    sf->allow_skip_recode = 1;
+  }
+
+  if (speed >= 5) {
+    sf->use_quant_fp = !is_keyframe;
+    sf->auto_min_max_partition_size = is_keyframe ? RELAXED_NEIGHBORING_MIN_MAX
+                                                  : STRICT_NEIGHBORING_MIN_MAX;
+    sf->default_max_partition_size = BLOCK_32X32;
+    sf->default_min_partition_size = BLOCK_8X8;
+    sf->force_frame_boost = is_keyframe ||
+        (frames_since_key % (sf->last_partitioning_redo_frequency << 1) == 1);
+    sf->max_delta_qindex = is_keyframe ? 20 : 15;
+    sf->partition_search_type = REFERENCE_PARTITION;
+    sf->use_nonrd_pick_mode = 1;
+    sf->allow_skip_recode = 0;
+    sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEW_ZERO;
+    sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST_NEW_ZERO;
+    sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST_NEW_ZERO;
+    sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST_NEW_ZERO;
+    sf->adaptive_rd_thresh = 2;
+    // This feature is only enabled when partition search is disabled.
+    sf->reuse_inter_pred_sby = 1;
+    sf->partition_search_breakout_rate_thr = 200;
+    sf->coeff_prob_appx_step = 4;
+    sf->use_fast_coef_updates = is_keyframe ? TWO_LOOP : ONE_LOOP_REDUCED;
+    sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH;
+    sf->tx_size_search_method = is_keyframe ? USE_LARGESTALL : USE_TX_8X8;
+    sf->simple_model_rd_from_var = 1;
+    if (cpi->oxcf.rc_mode == VPX_VBR)
+      sf->mv.search_method = NSTEP;
+
+    if (!is_keyframe) {
+      int i;
+      if (content == VP9E_CONTENT_SCREEN) {
+        for (i = 0; i < BLOCK_SIZES; ++i)
+          sf->intra_y_mode_bsize_mask[i] = INTRA_DC_TM_H_V;
+      } else {
+        for (i = 0; i < BLOCK_SIZES; ++i)
+          if (i > BLOCK_16X16)
+            sf->intra_y_mode_bsize_mask[i] = INTRA_DC;
+          else
+            // Use H and V intra mode for block sizes <= 16X16.
+            sf->intra_y_mode_bsize_mask[i] = INTRA_DC_H_V;
+      }
+    }
+    if (content == VP9E_CONTENT_SCREEN) {
+      sf->short_circuit_flat_blocks = 1;
+    }
+  }
+
+  if (speed >= 6) {
+    sf->partition_search_type = VAR_BASED_PARTITION;
+    // Turn on this to use non-RD key frame coding mode.
+    sf->use_nonrd_pick_mode = 1;
+    sf->mv.search_method = NSTEP;
+    sf->mv.reduce_first_step_size = 1;
+    sf->skip_encode_sb = 0;
+    if (!cpi->use_svc && cpi->oxcf.rc_mode == VPX_CBR &&
+        content != VP9E_CONTENT_SCREEN) {
+      // Enable short circuit for low temporal variance.
+      sf->short_circuit_low_temp_var = 1;
+    }
+  }
+
+  if (speed >= 7) {
+    sf->adaptive_rd_thresh = 3;
+    sf->mv.search_method = FAST_DIAMOND;
+    sf->mv.fullpel_search_step_param = 10;
+    if (cpi->svc.number_temporal_layers > 2 &&
+        cpi->svc.temporal_layer_id == 0) {
+      sf->mv.search_method = NSTEP;
+      sf->mv.fullpel_search_step_param = 6;
+    }
+  }
+  if (speed >= 8) {
+    sf->adaptive_rd_thresh = 4;
+    sf->mv.subpel_force_stop = (content == VP9E_CONTENT_SCREEN) ? 3 : 2;
+    sf->lpf_pick = LPF_PICK_MINIMAL_LPF;
+    // Only keep INTRA_DC mode for speed 8.
+    if (!is_keyframe) {
+      int i = 0;
+      for (i = 0; i < BLOCK_SIZES; ++i)
+        sf->intra_y_mode_bsize_mask[i] = INTRA_DC;
+    }
+    if (!cpi->use_svc && cpi->oxcf.rc_mode == VPX_CBR &&
+        content != VP9E_CONTENT_SCREEN) {
+      // More aggressive short circuit for speed 8.
+      sf->short_circuit_low_temp_var = 2;
+    }
+  }
+}
+
+void vp9_set_speed_features_framesize_dependent(VP9_COMP *cpi) {
+  SPEED_FEATURES *const sf = &cpi->sf;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  RD_OPT *const rd = &cpi->rd;
+  int i;
+
+  if (oxcf->mode == REALTIME) {
+    set_rt_speed_feature_framesize_dependent(cpi, sf, oxcf->speed);
+  } else if (oxcf->mode == GOOD) {
+    set_good_speed_feature_framesize_dependent(cpi, sf, oxcf->speed);
+  }
+
+  if (sf->disable_split_mask == DISABLE_ALL_SPLIT) {
+    sf->adaptive_pred_interp_filter = 0;
+  }
+
+  if (cpi->encode_breakout && oxcf->mode == REALTIME &&
+      sf->encode_breakout_thresh > cpi->encode_breakout) {
+    cpi->encode_breakout = sf->encode_breakout_thresh;
+  }
+
+  // Check for masked out split cases.
+  for (i = 0; i < MAX_REFS; ++i) {
+    if (sf->disable_split_mask & (1 << i)) {
+      rd->thresh_mult_sub8x8[i] = INT_MAX;
+    }
+  }
+}
+
+void vp9_set_speed_features_framesize_independent(VP9_COMP *cpi) {
+  SPEED_FEATURES *const sf = &cpi->sf;
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->td.mb;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  int i;
+
+  // best quality defaults
+  sf->frame_parameter_update = 1;
+  sf->mv.search_method = NSTEP;
+  sf->recode_loop = ALLOW_RECODE;
+  sf->mv.subpel_search_method = SUBPEL_TREE;
+  sf->mv.subpel_iters_per_step = 2;
+  sf->mv.subpel_force_stop = 0;
+  sf->optimize_coefficients = !is_lossless_requested(&cpi->oxcf);
+  sf->mv.reduce_first_step_size = 0;
+  sf->coeff_prob_appx_step = 1;
+  sf->mv.auto_mv_step_size = 0;
+  sf->mv.fullpel_search_step_param = 6;
+  sf->comp_inter_joint_search_thresh = BLOCK_4X4;
+  sf->tx_size_search_method = USE_FULL_RD;
+  sf->use_lp32x32fdct = 0;
+  sf->adaptive_motion_search = 0;
+  sf->adaptive_pred_interp_filter = 0;
+  sf->adaptive_mode_search = 0;
+  sf->cb_pred_filter_search = 0;
+  sf->cb_partition_search = 0;
+  sf->motion_field_mode_search = 0;
+  sf->alt_ref_search_fp = 0;
+  sf->use_quant_fp = 0;
+  sf->reference_masking = 0;
+  sf->partition_search_type = SEARCH_PARTITION;
+  sf->less_rectangular_check = 0;
+  sf->use_square_partition_only = 0;
+  sf->use_square_only_threshold = BLOCK_SIZES;
+  sf->auto_min_max_partition_size = NOT_IN_USE;
+  sf->rd_auto_partition_min_limit = BLOCK_4X4;
+  sf->default_max_partition_size = BLOCK_64X64;
+  sf->default_min_partition_size = BLOCK_4X4;
+  sf->adjust_partitioning_from_last_frame = 0;
+  sf->last_partitioning_redo_frequency = 4;
+  sf->disable_split_mask = 0;
+  sf->mode_search_skip_flags = 0;
+  sf->force_frame_boost = 0;
+  sf->max_delta_qindex = 0;
+  sf->disable_filter_search_var_thresh = 0;
+  sf->adaptive_interp_filter_search = 0;
+  sf->allow_partition_search_skip = 0;
+
+  for (i = 0; i < TX_SIZES; i++) {
+    sf->intra_y_mode_mask[i] = INTRA_ALL;
+    sf->intra_uv_mode_mask[i] = INTRA_ALL;
+  }
+  sf->use_rd_breakout = 0;
+  sf->skip_encode_sb = 0;
+  sf->use_uv_intra_rd_estimate = 0;
+  sf->allow_skip_recode = 0;
+  sf->lpf_pick = LPF_PICK_FROM_FULL_IMAGE;
+  sf->use_fast_coef_updates = TWO_LOOP;
+  sf->use_fast_coef_costing = 0;
+  sf->mode_skip_start = MAX_MODES;  // Mode index at which mode skip mask set
+  sf->schedule_mode_search = 0;
+  sf->use_nonrd_pick_mode = 0;
+  for (i = 0; i < BLOCK_SIZES; ++i)
+    sf->inter_mode_mask[i] = INTER_ALL;
+  sf->max_intra_bsize = BLOCK_64X64;
+  sf->reuse_inter_pred_sby = 0;
+  // This setting only takes effect when partition_search_type is set
+  // to FIXED_PARTITION.
+  sf->always_this_block_size = BLOCK_16X16;
+  sf->search_type_check_frequency = 50;
+  sf->encode_breakout_thresh = 0;
+  // Recode loop tolerance %.
+  sf->recode_tolerance = 25;
+  sf->default_interp_filter = SWITCHABLE;
+  sf->simple_model_rd_from_var = 0;
+  sf->short_circuit_flat_blocks = 0;
+  sf->short_circuit_low_temp_var = 0;
+
+  // Some speed-up features even for best quality as minimal impact on quality.
+  sf->adaptive_rd_thresh = 1;
+  sf->tx_size_search_breakout = 1;
+  sf->partition_search_breakout_dist_thr = (1 << 19);
+  sf->partition_search_breakout_rate_thr = 80;
+
+  if (oxcf->mode == REALTIME)
+    set_rt_speed_feature(cpi, sf, oxcf->speed, oxcf->content);
+  else if (oxcf->mode == GOOD)
+    set_good_speed_feature(cpi, cm, sf, oxcf->speed);
+
+  cpi->full_search_sad = vp9_full_search_sad;
+  cpi->diamond_search_sad = vp9_diamond_search_sad;
+
+  sf->allow_exhaustive_searches = 1;
+  if (oxcf->mode == BEST) {
+    if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION)
+      sf->exhaustive_searches_thresh = (1 << 20);
+    else
+      sf->exhaustive_searches_thresh = (1 << 21);
+    sf->max_exaustive_pct = 100;
+    for (i = 0; i < MAX_MESH_STEP; ++i) {
+      sf->mesh_patterns[i].range = best_quality_mesh_pattern[i].range;
+      sf->mesh_patterns[i].interval = best_quality_mesh_pattern[i].interval;
+    }
+  } else {
+    int speed = (oxcf->speed > MAX_MESH_SPEED) ? MAX_MESH_SPEED : oxcf->speed;
+    if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION)
+      sf->exhaustive_searches_thresh = (1 << 22);
+    else
+      sf->exhaustive_searches_thresh = (1 << 23);
+    sf->max_exaustive_pct = good_quality_max_mesh_pct[speed];
+    if (speed > 0)
+      sf->exhaustive_searches_thresh = sf->exhaustive_searches_thresh << 1;
+
+    for (i = 0; i < MAX_MESH_STEP; ++i) {
+      sf->mesh_patterns[i].range =
+          good_quality_mesh_patterns[speed][i].range;
+      sf->mesh_patterns[i].interval =
+          good_quality_mesh_patterns[speed][i].interval;
+    }
+  }
+
+  // Slow quant, dct and trellis not worthwhile for first pass
+  // so make sure they are always turned off.
+  if (oxcf->pass == 1)
+    sf->optimize_coefficients = 0;
+
+  // No recode for 1 pass.
+  if (oxcf->pass == 0) {
+    sf->recode_loop = DISALLOW_RECODE;
+    sf->optimize_coefficients = 0;
+  }
+
+  if (sf->mv.subpel_force_stop == 3) {
+    // Whole pel only
+    cpi->find_fractional_mv_step = vp9_skip_sub_pixel_tree;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE) {
+    cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED) {
+    cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_MORE) {
+    cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned_more;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_EVENMORE) {
+    cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned_evenmore;
+  }
+
+  x->optimize = sf->optimize_coefficients == 1 && oxcf->pass != 1;
+
+  x->min_partition_size = sf->default_min_partition_size;
+  x->max_partition_size = sf->default_max_partition_size;
+
+  if (!cpi->oxcf.frame_periodic_boost) {
+    sf->max_delta_qindex = 0;
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_speed_features.h b/libvpx/libvpx/vp9/encoder/vp9_speed_features.h
new file mode 100644
index 0000000..e88a7df
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_speed_features.h
@@ -0,0 +1,468 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_SPEED_FEATURES_H_
+#define VP9_ENCODER_VP9_SPEED_FEATURES_H_
+
+#include "vp9/common/vp9_enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum {
+  INTRA_ALL       = (1 << DC_PRED) |
+                    (1 << V_PRED) | (1 << H_PRED) |
+                    (1 << D45_PRED) | (1 << D135_PRED) |
+                    (1 << D117_PRED) | (1 << D153_PRED) |
+                    (1 << D207_PRED) | (1 << D63_PRED) |
+                    (1 << TM_PRED),
+  INTRA_DC        = (1 << DC_PRED),
+  INTRA_DC_TM     = (1 << DC_PRED) | (1 << TM_PRED),
+  INTRA_DC_H_V    = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
+  INTRA_DC_TM_H_V = (1 << DC_PRED) | (1 << TM_PRED) | (1 << V_PRED) |
+                    (1 << H_PRED)
+};
+
+enum {
+  INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) | (1 << NEWMV),
+  INTER_NEAREST = (1 << NEARESTMV),
+  INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV),
+  INTER_NEAREST_ZERO = (1 << NEARESTMV) | (1 << ZEROMV),
+  INTER_NEAREST_NEW_ZERO = (1 << NEARESTMV) | (1 << ZEROMV) | (1 << NEWMV),
+  INTER_NEAREST_NEAR_NEW = (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV),
+  INTER_NEAREST_NEAR_ZERO = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV),
+};
+
+enum {
+  DISABLE_ALL_INTER_SPLIT   = (1 << THR_COMP_GA) |
+                              (1 << THR_COMP_LA) |
+                              (1 << THR_ALTR) |
+                              (1 << THR_GOLD) |
+                              (1 << THR_LAST),
+
+  DISABLE_ALL_SPLIT         = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
+
+  DISABLE_COMPOUND_SPLIT    = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
+
+  LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) |
+                              (1 << THR_COMP_LA) |
+                              (1 << THR_ALTR) |
+                              (1 << THR_GOLD)
+};
+
+typedef enum {
+  DIAMOND = 0,
+  NSTEP = 1,
+  HEX = 2,
+  BIGDIA = 3,
+  SQUARE = 4,
+  FAST_HEX = 5,
+  FAST_DIAMOND = 6
+} SEARCH_METHODS;
+
+typedef enum {
+  // No recode.
+  DISALLOW_RECODE = 0,
+  // Allow recode for KF and exceeding maximum frame bandwidth.
+  ALLOW_RECODE_KFMAXBW = 1,
+  // Allow recode only for KF/ARF/GF frames.
+  ALLOW_RECODE_KFARFGF = 2,
+  // Allow recode for all frames based on bitrate constraints.
+  ALLOW_RECODE = 3,
+} RECODE_LOOP_TYPE;
+
+typedef enum {
+  SUBPEL_TREE = 0,
+  SUBPEL_TREE_PRUNED = 1,           // Prunes 1/2-pel searches
+  SUBPEL_TREE_PRUNED_MORE = 2,      // Prunes 1/2-pel searches more aggressively
+  SUBPEL_TREE_PRUNED_EVENMORE = 3,  // Prunes 1/2- and 1/4-pel searches
+  // Other methods to come
+} SUBPEL_SEARCH_METHODS;
+
+typedef enum {
+  NO_MOTION_THRESHOLD = 0,
+  LOW_MOTION_THRESHOLD = 7
+} MOTION_THRESHOLD;
+
+typedef enum {
+  USE_FULL_RD = 0,
+  USE_LARGESTALL,
+  USE_TX_8X8
+} TX_SIZE_SEARCH_METHOD;
+
+typedef enum {
+  NOT_IN_USE = 0,
+  RELAXED_NEIGHBORING_MIN_MAX = 1,
+  STRICT_NEIGHBORING_MIN_MAX = 2
+} AUTO_MIN_MAX_MODE;
+
+typedef enum {
+  // Try the full image with different values.
+  LPF_PICK_FROM_FULL_IMAGE,
+  // Try a small portion of the image with different values.
+  LPF_PICK_FROM_SUBIMAGE,
+  // Estimate the level based on quantizer and frame type
+  LPF_PICK_FROM_Q,
+  // Pick 0 to disable LPF if LPF was enabled last frame
+  LPF_PICK_MINIMAL_LPF
+} LPF_PICK_METHOD;
+
+typedef enum {
+  // Terminate search early based on distortion so far compared to
+  // qp step, distortion in the neighborhood of the frame, etc.
+  FLAG_EARLY_TERMINATE = 1 << 0,
+
+  // Skips comp inter modes if the best so far is an intra mode.
+  FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
+
+  // Skips oblique intra modes if the best so far is an inter mode.
+  FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
+
+  // Skips oblique intra modes  at angles 27, 63, 117, 153 if the best
+  // intra so far is not one of the neighboring directions.
+  FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
+
+  // Skips intra modes other than DC_PRED if the source variance is small
+  FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
+} MODE_SEARCH_SKIP_LOGIC;
+
+typedef enum {
+  FLAG_SKIP_EIGHTTAP = 1 << EIGHTTAP,
+  FLAG_SKIP_EIGHTTAP_SMOOTH = 1 << EIGHTTAP_SMOOTH,
+  FLAG_SKIP_EIGHTTAP_SHARP = 1 << EIGHTTAP_SHARP,
+} INTERP_FILTER_MASK;
+
+typedef enum {
+  // Search partitions using RD/NONRD criterion
+  SEARCH_PARTITION,
+
+  // Always use a fixed size partition
+  FIXED_PARTITION,
+
+  REFERENCE_PARTITION,
+
+  // Use an arbitrary partitioning scheme based on source variance within
+  // a 64X64 SB
+  VAR_BASED_PARTITION,
+
+  // Use non-fixed partitions based on source variance
+  SOURCE_VAR_BASED_PARTITION
+} PARTITION_SEARCH_TYPE;
+
+typedef enum {
+  // Does a dry run to see if any of the contexts need to be updated or not,
+  // before the final run.
+  TWO_LOOP = 0,
+
+  // No dry run, also only half the coef contexts and bands are updated.
+  // The rest are not updated at all.
+  ONE_LOOP_REDUCED = 1
+} FAST_COEFF_UPDATE;
+
+typedef struct MV_SPEED_FEATURES {
+  // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
+  SEARCH_METHODS search_method;
+
+  // This parameter controls which step in the n-step process we start at.
+  // It's changed adaptively based on circumstances.
+  int reduce_first_step_size;
+
+  // If this is set to 1, we limit the motion search range to 2 times the
+  // largest motion vector found in the last frame.
+  int auto_mv_step_size;
+
+  // Subpel_search_method can only be subpel_tree which does a subpixel
+  // logarithmic search that keeps stepping at 1/2 pixel units until
+  // you stop getting a gain, and then goes on to 1/4 and repeats
+  // the same process. Along the way it skips many diagonals.
+  SUBPEL_SEARCH_METHODS subpel_search_method;
+
+  // Maximum number of steps in logarithmic subpel search before giving up.
+  int subpel_iters_per_step;
+
+  // Control when to stop subpel search:
+  // 0: Full subpel search.
+  // 1: Stop at quarter pixel.
+  // 2: Stop at half pixel.
+  // 3: Stop at full pixel.
+  int subpel_force_stop;
+
+  // This variable sets the step_param used in full pel motion search.
+  int fullpel_search_step_param;
+} MV_SPEED_FEATURES;
+
+#define MAX_MESH_STEP 4
+
+typedef struct MESH_PATTERN {
+  int range;
+  int interval;
+} MESH_PATTERN;
+
+typedef struct SPEED_FEATURES {
+  MV_SPEED_FEATURES mv;
+
+  // Frame level coding parameter update
+  int frame_parameter_update;
+
+  RECODE_LOOP_TYPE recode_loop;
+
+  // Trellis (dynamic programming) optimization of quantized values (+1, 0).
+  int optimize_coefficients;
+
+  // Always set to 0. If on it enables 0 cost background transmission
+  // (except for the initial transmission of the segmentation). The feature is
+  // disabled because the addition of very large block sizes make the
+  // backgrounds very to cheap to encode, and the segmentation we have
+  // adds overhead.
+  int static_segmentation;
+
+  // If 1 we iterate finding a best reference for 2 ref frames together - via
+  // a log search that iterates 4 times (check around mv for last for best
+  // error of combined predictor then check around mv for alt). If 0 we
+  // we just use the best motion vector found for each frame by itself.
+  BLOCK_SIZE comp_inter_joint_search_thresh;
+
+  // This variable is used to cap the maximum number of times we skip testing a
+  // mode to be evaluated. A high value means we will be faster.
+  int adaptive_rd_thresh;
+
+  // Enables skipping the reconstruction step (idct, recon) in the
+  // intermediate steps assuming the last frame didn't have too many intra
+  // blocks and the q is less than a threshold.
+  int skip_encode_sb;
+  int skip_encode_frame;
+  // Speed feature to allow or disallow skipping of recode at block
+  // level within a frame.
+  int allow_skip_recode;
+
+  // Coefficient probability model approximation step size
+  int coeff_prob_appx_step;
+
+  // The threshold is to determine how slow the motino is, it is used when
+  // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
+  MOTION_THRESHOLD lf_motion_threshold;
+
+  // Determine which method we use to determine transform size. We can choose
+  // between options like full rd, largest for prediction size, largest
+  // for intra and model coefs for the rest.
+  TX_SIZE_SEARCH_METHOD tx_size_search_method;
+
+  // Low precision 32x32 fdct keeps everything in 16 bits and thus is less
+  // precise but significantly faster than the non lp version.
+  int use_lp32x32fdct;
+
+  // After looking at the first set of modes (set by index here), skip
+  // checking modes for reference frames that don't match the reference frame
+  // of the best so far.
+  int mode_skip_start;
+
+  // TODO(JBB): Remove this.
+  int reference_masking;
+
+  PARTITION_SEARCH_TYPE partition_search_type;
+
+  // Used if partition_search_type = FIXED_SIZE_PARTITION
+  BLOCK_SIZE always_this_block_size;
+
+  // Skip rectangular partition test when partition type none gives better
+  // rd than partition type split.
+  int less_rectangular_check;
+
+  // Disable testing non square partitions. (eg 16x32)
+  int use_square_partition_only;
+  BLOCK_SIZE use_square_only_threshold;
+
+  // Sets min and max partition sizes for this 64x64 region based on the
+  // same 64x64 in last encoded frame, and the left and above neighbor.
+  AUTO_MIN_MAX_MODE auto_min_max_partition_size;
+  // Ensures the rd based auto partition search will always
+  // go down at least to the specified level.
+  BLOCK_SIZE rd_auto_partition_min_limit;
+
+  // Min and max partition size we enable (block_size) as per auto
+  // min max, but also used by adjust partitioning, and pick_partitioning.
+  BLOCK_SIZE default_min_partition_size;
+  BLOCK_SIZE default_max_partition_size;
+
+  // Whether or not we allow partitions one smaller or one greater than the last
+  // frame's partitioning. Only used if use_lastframe_partitioning is set.
+  int adjust_partitioning_from_last_frame;
+
+  // How frequently we re do the partitioning from scratch. Only used if
+  // use_lastframe_partitioning is set.
+  int last_partitioning_redo_frequency;
+
+  // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable
+  // it always, to allow it for only Last frame and Intra, disable it for all
+  // inter modes or to enable it always.
+  int disable_split_mask;
+
+  // TODO(jingning): combine the related motion search speed features
+  // This allows us to use motion search at other sizes as a starting
+  // point for this motion search and limits the search range around it.
+  int adaptive_motion_search;
+
+  // Flag for allowing some use of exhaustive searches;
+  int allow_exhaustive_searches;
+
+  // Threshold for allowing exhaistive motion search.
+  int exhaustive_searches_thresh;
+
+  // Maximum number of exhaustive searches for a frame.
+  int max_exaustive_pct;
+
+  // Pattern to be used for any exhaustive mesh searches.
+  MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
+
+  int schedule_mode_search;
+
+  // Allows sub 8x8 modes to use the prediction filter that was determined
+  // best for 8x8 mode. If set to 0 we always re check all the filters for
+  // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter
+  // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
+  int adaptive_pred_interp_filter;
+
+  // Adaptive prediction mode search
+  int adaptive_mode_search;
+
+  // Chessboard pattern prediction filter type search
+  int cb_pred_filter_search;
+
+  int cb_partition_search;
+
+  int motion_field_mode_search;
+
+  int alt_ref_search_fp;
+
+  // Fast quantization process path
+  int use_quant_fp;
+
+  // Use finer quantizer in every other few frames that run variable block
+  // partition type search.
+  int force_frame_boost;
+
+  // Maximally allowed base quantization index fluctuation.
+  int max_delta_qindex;
+
+  // Implements various heuristics to skip searching modes
+  // The heuristics selected are based on  flags
+  // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
+  unsigned int mode_search_skip_flags;
+
+  // A source variance threshold below which filter search is disabled
+  // Choose a very large value (UINT_MAX) to use 8-tap always
+  unsigned int disable_filter_search_var_thresh;
+
+  // These bit masks allow you to enable or disable intra modes for each
+  // transform size separately.
+  int intra_y_mode_mask[TX_SIZES];
+  int intra_uv_mode_mask[TX_SIZES];
+
+  // These bit masks allow you to enable or disable intra modes for each
+  // prediction block size separately.
+  int intra_y_mode_bsize_mask[BLOCK_SIZES];
+
+  // This variable enables an early break out of mode testing if the model for
+  // rd built from the prediction signal indicates a value that's much
+  // higher than the best rd we've seen so far.
+  int use_rd_breakout;
+
+  // This enables us to use an estimate for intra rd based on dc mode rather
+  // than choosing an actual uv mode in the stage of encoding before the actual
+  // final encode.
+  int use_uv_intra_rd_estimate;
+
+  // This feature controls how the loop filter level is determined.
+  LPF_PICK_METHOD lpf_pick;
+
+  // This feature limits the number of coefficients updates we actually do
+  // by only looking at counts from 1/2 the bands.
+  FAST_COEFF_UPDATE use_fast_coef_updates;
+
+  // This flag controls the use of non-RD mode decision.
+  int use_nonrd_pick_mode;
+
+  // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
+  // modes are used in order from LSB to MSB for each BLOCK_SIZE.
+  int inter_mode_mask[BLOCK_SIZES];
+
+  // This feature controls whether we do the expensive context update and
+  // calculation in the rd coefficient costing loop.
+  int use_fast_coef_costing;
+
+  // This feature controls the tolerence vs target used in deciding whether to
+  // recode a frame. It has no meaning if recode is disabled.
+  int recode_tolerance;
+
+  // This variable controls the maximum block size where intra blocks can be
+  // used in inter frames.
+  // TODO(aconverse): Fold this into one of the other many mode skips
+  BLOCK_SIZE max_intra_bsize;
+
+  // The frequency that we check if SOURCE_VAR_BASED_PARTITION or
+  // FIXED_PARTITION search type should be used.
+  int search_type_check_frequency;
+
+  // When partition is pre-set, the inter prediction result from pick_inter_mode
+  // can be reused in final block encoding process. It is enabled only for real-
+  // time mode speed 6.
+  int reuse_inter_pred_sby;
+
+  // This variable sets the encode_breakout threshold. Currently, it is only
+  // enabled in real time mode.
+  int encode_breakout_thresh;
+
+  // default interp filter choice
+  INTERP_FILTER default_interp_filter;
+
+  // Early termination in transform size search, which only applies while
+  // tx_size_search_method is USE_FULL_RD.
+  int tx_size_search_breakout;
+
+  // adaptive interp_filter search to allow skip of certain filter types.
+  int adaptive_interp_filter_search;
+
+  // mask for skip evaluation of certain interp_filter type.
+  INTERP_FILTER_MASK interp_filter_search_mask;
+
+  // Partition search early breakout thresholds.
+  int64_t partition_search_breakout_dist_thr;
+  int partition_search_breakout_rate_thr;
+
+  // Allow skipping partition search for still image frame
+  int allow_partition_search_skip;
+
+  // Fast approximation of vp9_model_rd_from_var_lapndz
+  int simple_model_rd_from_var;
+
+  // Skip a number of expensive mode evaluations for blocks with zero source
+  // variance.
+  int short_circuit_flat_blocks;
+
+  // Skip a number of expensive mode evaluations for blocks with very low
+  // temporal variance.
+  // 1: Skip golden non-zeromv and ALL INTRA for bsize >= 32x32.
+  // 2: Skip golden non-zeromv and newmv-last for bsize >= 16x16, skip ALL
+  // INTRA for bsize >= 32x32 and vert/horz INTRA for bsize 16x16, 16x32 and
+  // 32x16.
+  int short_circuit_low_temp_var;
+} SPEED_FEATURES;
+
+struct VP9_COMP;
+
+void vp9_set_speed_features_framesize_independent(struct VP9_COMP *cpi);
+void vp9_set_speed_features_framesize_dependent(struct VP9_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SPEED_FEATURES_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_subexp.c b/libvpx/libvpx/vp9/encoder/vp9_subexp.c
new file mode 100644
index 0000000..29db015
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_subexp.c
@@ -0,0 +1,198 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "vpx_dsp/bitwriter.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_subexp.h"
+
+static const uint8_t update_bits[255] = {
+   5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
+   6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,
+   8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+   8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,  0,
+};
+#define MIN_DELP_BITS 5
+
+static int recenter_nonneg(int v, int m) {
+  if (v > (m << 1))
+    return v;
+  else if (v >= m)
+    return ((v - m) << 1);
+  else
+    return ((m - v) << 1) - 1;
+}
+
+static int remap_prob(int v, int m) {
+  int i;
+  static const uint8_t map_table[MAX_PROB - 1] = {
+    // generated by:
+    //   map_table[j] = split_index(j, MAX_PROB - 1, MODULUS_PARAM);
+     20,  21,  22,  23,  24,  25,   0,  26,  27,  28,  29,  30,  31,  32,  33,
+     34,  35,  36,  37,   1,  38,  39,  40,  41,  42,  43,  44,  45,  46,  47,
+     48,  49,   2,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,  60,  61,
+      3,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,   4,  74,
+     75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,   5,  86,  87,  88,
+     89,  90,  91,  92,  93,  94,  95,  96,  97,   6,  98,  99, 100, 101, 102,
+    103, 104, 105, 106, 107, 108, 109,   7, 110, 111, 112, 113, 114, 115, 116,
+    117, 118, 119, 120, 121,   8, 122, 123, 124, 125, 126, 127, 128, 129, 130,
+    131, 132, 133,   9, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144,
+    145,  10, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157,  11,
+    158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169,  12, 170, 171,
+    172, 173, 174, 175, 176, 177, 178, 179, 180, 181,  13, 182, 183, 184, 185,
+    186, 187, 188, 189, 190, 191, 192, 193,  14, 194, 195, 196, 197, 198, 199,
+    200, 201, 202, 203, 204, 205,  15, 206, 207, 208, 209, 210, 211, 212, 213,
+    214, 215, 216, 217,  16, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227,
+    228, 229,  17, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241,
+     18, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253,  19,
+  };
+  v--;
+  m--;
+  if ((m << 1) <= MAX_PROB)
+    i = recenter_nonneg(v, m) - 1;
+  else
+    i = recenter_nonneg(MAX_PROB - 1 - v, MAX_PROB - 1 - m) - 1;
+
+  i = map_table[i];
+  return i;
+}
+
+static int prob_diff_update_cost(vpx_prob newp, vpx_prob oldp) {
+  int delp = remap_prob(newp, oldp);
+  return update_bits[delp] << VP9_PROB_COST_SHIFT;
+}
+
+static void encode_uniform(vpx_writer *w, int v) {
+  const int l = 8;
+  const int m = (1 << l) - 191;
+  if (v < m) {
+    vpx_write_literal(w, v, l - 1);
+  } else {
+    vpx_write_literal(w, m + ((v - m) >> 1), l - 1);
+    vpx_write_literal(w, (v - m) & 1, 1);
+  }
+}
+
+static INLINE int write_bit_gte(vpx_writer *w, int word, int test) {
+  vpx_write_literal(w, word >= test, 1);
+  return word >= test;
+}
+
+static void encode_term_subexp(vpx_writer *w, int word) {
+  if (!write_bit_gte(w, word, 16)) {
+    vpx_write_literal(w, word, 4);
+  } else if (!write_bit_gte(w, word, 32)) {
+    vpx_write_literal(w, word - 16, 4);
+  } else if (!write_bit_gte(w, word, 64)) {
+    vpx_write_literal(w, word - 32, 5);
+  } else {
+    encode_uniform(w, word - 64);
+  }
+}
+
+void vp9_write_prob_diff_update(vpx_writer *w, vpx_prob newp, vpx_prob oldp) {
+  const int delp = remap_prob(newp, oldp);
+  encode_term_subexp(w, delp);
+}
+
+int vp9_prob_diff_update_savings_search(const unsigned int *ct,
+                                        vpx_prob oldp, vpx_prob *bestp,
+                                        vpx_prob upd) {
+  const int old_b = cost_branch256(ct, oldp);
+  int bestsavings = 0;
+  vpx_prob newp, bestnewp = oldp;
+  const int step = *bestp > oldp ? -1 : 1;
+  const int upd_cost = vp9_cost_one(upd) - vp9_cost_zero(upd);
+
+  if (old_b > upd_cost + (MIN_DELP_BITS << VP9_PROB_COST_SHIFT)) {
+    for (newp = *bestp; newp != oldp; newp += step) {
+      const int new_b = cost_branch256(ct, newp);
+      const int update_b = prob_diff_update_cost(newp, oldp) + upd_cost;
+      const int savings = old_b - new_b - update_b;
+      if (savings > bestsavings) {
+        bestsavings = savings;
+        bestnewp = newp;
+      }
+    }
+  }
+  *bestp = bestnewp;
+  return bestsavings;
+}
+
+int vp9_prob_diff_update_savings_search_model(const unsigned int *ct,
+                                              const vpx_prob oldp,
+                                              vpx_prob *bestp,
+                                              vpx_prob upd,
+                                              int stepsize) {
+  int i, old_b, new_b, update_b, savings, bestsavings;
+  int newp;
+  const int step_sign = *bestp > oldp ? -1 : 1;
+  const int step = stepsize * step_sign;
+  const int upd_cost = vp9_cost_one(upd) - vp9_cost_zero(upd);
+  const vpx_prob *newplist, *oldplist;
+  vpx_prob bestnewp;
+  oldplist = vp9_pareto8_full[oldp - 1];
+  old_b = cost_branch256(ct + 2 * PIVOT_NODE, oldp);
+  for (i = UNCONSTRAINED_NODES; i < ENTROPY_NODES; ++i)
+    old_b += cost_branch256(ct + 2 * i, oldplist[i - UNCONSTRAINED_NODES]);
+
+  bestsavings = 0;
+  bestnewp = oldp;
+
+  assert(stepsize > 0);
+
+  if (old_b > upd_cost + (MIN_DELP_BITS << VP9_PROB_COST_SHIFT)) {
+    for (newp = *bestp; (newp - oldp) * step_sign < 0; newp += step) {
+      if (newp < 1 || newp > 255) continue;
+      newplist = vp9_pareto8_full[newp - 1];
+      new_b = cost_branch256(ct + 2 * PIVOT_NODE, newp);
+      for (i = UNCONSTRAINED_NODES; i < ENTROPY_NODES; ++i)
+        new_b += cost_branch256(ct + 2 * i, newplist[i - UNCONSTRAINED_NODES]);
+      update_b = prob_diff_update_cost(newp, oldp) + upd_cost;
+      savings = old_b - new_b - update_b;
+      if (savings > bestsavings) {
+        bestsavings = savings;
+        bestnewp = newp;
+      }
+    }
+  }
+
+  *bestp = bestnewp;
+  return bestsavings;
+}
+
+void vp9_cond_prob_diff_update(vpx_writer *w, vpx_prob *oldp,
+                               const unsigned int ct[2]) {
+  const vpx_prob upd = DIFF_UPDATE_PROB;
+  vpx_prob newp = get_binary_prob(ct[0], ct[1]);
+  const int savings = vp9_prob_diff_update_savings_search(ct, *oldp, &newp,
+                                                          upd);
+  assert(newp >= 1);
+  if (savings > 0) {
+    vpx_write(w, 1, upd);
+    vp9_write_prob_diff_update(w, newp, *oldp);
+    *oldp = newp;
+  } else {
+    vpx_write(w, 0, upd);
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_subexp.h b/libvpx/libvpx/vp9/encoder/vp9_subexp.h
new file mode 100644
index 0000000..efe62c0
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_subexp.h
@@ -0,0 +1,44 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_SUBEXP_H_
+#define VP9_ENCODER_VP9_SUBEXP_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "vpx_dsp/prob.h"
+
+struct vpx_writer;
+
+void vp9_write_prob_diff_update(struct vpx_writer *w,
+                                vpx_prob newp, vpx_prob oldp);
+
+void vp9_cond_prob_diff_update(struct vpx_writer *w, vpx_prob *oldp,
+                               const unsigned int ct[2]);
+
+int vp9_prob_diff_update_savings_search(const unsigned int *ct,
+                                        vpx_prob oldp, vpx_prob *bestp,
+                                        vpx_prob upd);
+
+
+int vp9_prob_diff_update_savings_search_model(const unsigned int *ct,
+                                              const vpx_prob oldp,
+                                              vpx_prob *bestp,
+                                              vpx_prob upd,
+                                              int stepsize);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SUBEXP_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_svc_layercontext.c b/libvpx/libvpx/vp9/encoder/vp9_svc_layercontext.c
new file mode 100644
index 0000000..1814a32
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_svc_layercontext.c
@@ -0,0 +1,842 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_svc_layercontext.h"
+#include "vp9/encoder/vp9_extend.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+#define SMALL_FRAME_WIDTH  32
+#define SMALL_FRAME_HEIGHT 16
+
+void vp9_init_layer_context(VP9_COMP *const cpi) {
+  SVC *const svc = &cpi->svc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  int mi_rows = cpi->common.mi_rows;
+  int mi_cols = cpi->common.mi_cols;
+  int sl, tl, i;
+  int alt_ref_idx = svc->number_spatial_layers;
+
+  svc->spatial_layer_id = 0;
+  svc->temporal_layer_id = 0;
+  svc->first_spatial_layer_to_encode = 0;
+  svc->rc_drop_superframe = 0;
+  svc->force_zero_mode_spatial_ref = 0;
+  svc->use_base_mv = 0;
+  svc->current_superframe = 0;
+  for (i = 0; i < REF_FRAMES; ++i)
+    svc->ref_frame_index[i] = -1;
+  for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
+    cpi->svc.ext_frame_flags[sl] = 0;
+    cpi->svc.ext_lst_fb_idx[sl] = 0;
+    cpi->svc.ext_gld_fb_idx[sl] = 1;
+    cpi->svc.ext_alt_fb_idx[sl] = 2;
+  }
+
+  // For 1 pass cbr: allocate scaled_frame that may be used as an intermediate
+  // buffer for a 2 stage down-sampling: two stages of 1:2 down-sampling for a
+  // target of 1/4x1/4.
+  if (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR) {
+    if (vpx_realloc_frame_buffer(&cpi->svc.scaled_temp,
+                                 cpi->common.width >> 1,
+                                 cpi->common.height >> 1,
+                                 cpi->common.subsampling_x,
+                                 cpi->common.subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 cpi->common.use_highbitdepth,
+#endif
+                                 VP9_ENC_BORDER_IN_PIXELS,
+                                 cpi->common.byte_alignment,
+                                 NULL, NULL, NULL))
+      vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                         "Failed to allocate scaled_frame for svc ");
+  }
+
+
+  if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2) {
+    if (vpx_realloc_frame_buffer(&cpi->svc.empty_frame.img,
+                                 SMALL_FRAME_WIDTH, SMALL_FRAME_HEIGHT,
+                                 cpi->common.subsampling_x,
+                                 cpi->common.subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 cpi->common.use_highbitdepth,
+#endif
+                                 VP9_ENC_BORDER_IN_PIXELS,
+                                 cpi->common.byte_alignment,
+                                 NULL, NULL, NULL))
+      vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                         "Failed to allocate empty frame for multiple frame "
+                         "contexts");
+
+    memset(cpi->svc.empty_frame.img.buffer_alloc, 0x80,
+           cpi->svc.empty_frame.img.buffer_alloc_sz);
+  }
+
+  for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
+    for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
+      int layer = LAYER_IDS_TO_IDX(sl, tl, oxcf->ts_number_layers);
+      LAYER_CONTEXT *const lc = &svc->layer_context[layer];
+      RATE_CONTROL *const lrc = &lc->rc;
+      int i;
+      lc->current_video_frame_in_layer = 0;
+      lc->layer_size = 0;
+      lc->frames_from_key_frame = 0;
+      lc->last_frame_type = FRAME_TYPES;
+      lrc->ni_av_qi = oxcf->worst_allowed_q;
+      lrc->total_actual_bits = 0;
+      lrc->total_target_vs_actual = 0;
+      lrc->ni_tot_qi = 0;
+      lrc->tot_q = 0.0;
+      lrc->avg_q = 0.0;
+      lrc->ni_frames = 0;
+      lrc->decimation_count = 0;
+      lrc->decimation_factor = 0;
+
+      for (i = 0; i < RATE_FACTOR_LEVELS; ++i) {
+        lrc->rate_correction_factors[i] = 1.0;
+      }
+
+      if (cpi->oxcf.rc_mode == VPX_CBR) {
+        lc->target_bandwidth = oxcf->layer_target_bitrate[layer];
+        lrc->last_q[INTER_FRAME] = oxcf->worst_allowed_q;
+        lrc->avg_frame_qindex[INTER_FRAME] = oxcf->worst_allowed_q;
+        lrc->avg_frame_qindex[KEY_FRAME] = oxcf->worst_allowed_q;
+      } else {
+        lc->target_bandwidth = oxcf->layer_target_bitrate[layer];
+        lrc->last_q[KEY_FRAME] = oxcf->best_allowed_q;
+        lrc->last_q[INTER_FRAME] = oxcf->best_allowed_q;
+        lrc->avg_frame_qindex[KEY_FRAME] = (oxcf->worst_allowed_q +
+                                            oxcf->best_allowed_q) / 2;
+        lrc->avg_frame_qindex[INTER_FRAME] = (oxcf->worst_allowed_q +
+                                              oxcf->best_allowed_q) / 2;
+        if (oxcf->ss_enable_auto_arf[sl])
+          lc->alt_ref_idx = alt_ref_idx++;
+        else
+          lc->alt_ref_idx = INVALID_IDX;
+        lc->gold_ref_idx = INVALID_IDX;
+      }
+
+      lrc->buffer_level = oxcf->starting_buffer_level_ms *
+                              lc->target_bandwidth / 1000;
+      lrc->bits_off_target = lrc->buffer_level;
+
+      // Initialize the cyclic refresh parameters. If spatial layers are used
+      // (i.e., ss_number_layers > 1), these need to be updated per spatial
+      // layer.
+      // Cyclic refresh is only applied on base temporal layer.
+      if (oxcf->ss_number_layers > 1 &&
+          tl == 0) {
+        size_t last_coded_q_map_size;
+        size_t consec_zero_mv_size;
+        VP9_COMMON *const cm = &cpi->common;
+        lc->sb_index = 0;
+        CHECK_MEM_ERROR(cm, lc->map,
+                        vpx_malloc(mi_rows * mi_cols * sizeof(*lc->map)));
+        memset(lc->map, 0, mi_rows * mi_cols);
+        last_coded_q_map_size = mi_rows * mi_cols *
+                                sizeof(*lc->last_coded_q_map);
+        CHECK_MEM_ERROR(cm, lc->last_coded_q_map,
+                        vpx_malloc(last_coded_q_map_size));
+        assert(MAXQ <= 255);
+        memset(lc->last_coded_q_map, MAXQ, last_coded_q_map_size);
+        consec_zero_mv_size = mi_rows * mi_cols * sizeof(*lc->consec_zero_mv);
+        CHECK_MEM_ERROR(cm, lc->consec_zero_mv,
+                        vpx_malloc(consec_zero_mv_size));
+        memset(lc->consec_zero_mv, 0, consec_zero_mv_size);
+       }
+    }
+  }
+
+  // Still have extra buffer for base layer golden frame
+  if (!(svc->number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR)
+      && alt_ref_idx < REF_FRAMES)
+    svc->layer_context[0].gold_ref_idx = alt_ref_idx;
+}
+
+// Update the layer context from a change_config() call.
+void vp9_update_layer_context_change_config(VP9_COMP *const cpi,
+                                            const int target_bandwidth) {
+  SVC *const svc = &cpi->svc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  int sl, tl, layer = 0, spatial_layer_target;
+  float bitrate_alloc = 1.0;
+
+  if (svc->temporal_layering_mode != VP9E_TEMPORAL_LAYERING_MODE_NOLAYERING) {
+    for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
+      for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
+        layer = LAYER_IDS_TO_IDX(sl, tl, oxcf->ts_number_layers);
+        svc->layer_context[layer].target_bandwidth =
+            oxcf->layer_target_bitrate[layer];
+      }
+
+      layer = LAYER_IDS_TO_IDX(sl, ((oxcf->ts_number_layers - 1) < 0 ?
+          0 : (oxcf->ts_number_layers - 1)), oxcf->ts_number_layers);
+      spatial_layer_target =
+          svc->layer_context[layer].target_bandwidth =
+              oxcf->layer_target_bitrate[layer];
+
+      for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
+        LAYER_CONTEXT *const lc =
+            &svc->layer_context[sl * oxcf->ts_number_layers + tl];
+        RATE_CONTROL *const lrc = &lc->rc;
+
+        lc->spatial_layer_target_bandwidth = spatial_layer_target;
+        bitrate_alloc = (float)lc->target_bandwidth / spatial_layer_target;
+        lrc->starting_buffer_level =
+            (int64_t)(rc->starting_buffer_level * bitrate_alloc);
+        lrc->optimal_buffer_level =
+            (int64_t)(rc->optimal_buffer_level * bitrate_alloc);
+        lrc->maximum_buffer_size =
+            (int64_t)(rc->maximum_buffer_size * bitrate_alloc);
+        lrc->bits_off_target =
+            VPXMIN(lrc->bits_off_target, lrc->maximum_buffer_size);
+        lrc->buffer_level = VPXMIN(lrc->buffer_level, lrc->maximum_buffer_size);
+        lc->framerate = cpi->framerate / oxcf->ts_rate_decimator[tl];
+        lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+        lrc->max_frame_bandwidth = rc->max_frame_bandwidth;
+        lrc->worst_quality = rc->worst_quality;
+        lrc->best_quality = rc->best_quality;
+      }
+    }
+  } else {
+    int layer_end;
+
+    if (svc->number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) {
+      layer_end = svc->number_temporal_layers;
+    } else {
+      layer_end = svc->number_spatial_layers;
+    }
+
+    for (layer = 0; layer < layer_end; ++layer) {
+      LAYER_CONTEXT *const lc = &svc->layer_context[layer];
+      RATE_CONTROL *const lrc = &lc->rc;
+
+      lc->target_bandwidth = oxcf->layer_target_bitrate[layer];
+
+      bitrate_alloc = (float)lc->target_bandwidth / target_bandwidth;
+      // Update buffer-related quantities.
+      lrc->starting_buffer_level =
+          (int64_t)(rc->starting_buffer_level * bitrate_alloc);
+      lrc->optimal_buffer_level =
+          (int64_t)(rc->optimal_buffer_level * bitrate_alloc);
+      lrc->maximum_buffer_size =
+          (int64_t)(rc->maximum_buffer_size * bitrate_alloc);
+      lrc->bits_off_target = VPXMIN(lrc->bits_off_target,
+                                    lrc->maximum_buffer_size);
+      lrc->buffer_level = VPXMIN(lrc->buffer_level, lrc->maximum_buffer_size);
+      // Update framerate-related quantities.
+      if (svc->number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) {
+        lc->framerate = cpi->framerate / oxcf->ts_rate_decimator[layer];
+      } else {
+        lc->framerate = cpi->framerate;
+      }
+      lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+      lrc->max_frame_bandwidth = rc->max_frame_bandwidth;
+      // Update qp-related quantities.
+      lrc->worst_quality = rc->worst_quality;
+      lrc->best_quality = rc->best_quality;
+    }
+  }
+}
+
+static LAYER_CONTEXT *get_layer_context(VP9_COMP *const cpi) {
+  if (is_one_pass_cbr_svc(cpi))
+    return &cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+        cpi->svc.number_temporal_layers + cpi->svc.temporal_layer_id];
+  else
+    return (cpi->svc.number_temporal_layers > 1 &&
+            cpi->oxcf.rc_mode == VPX_CBR) ?
+             &cpi->svc.layer_context[cpi->svc.temporal_layer_id] :
+             &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
+}
+
+void vp9_update_temporal_layer_framerate(VP9_COMP *const cpi) {
+  SVC *const svc = &cpi->svc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  LAYER_CONTEXT *const lc = get_layer_context(cpi);
+  RATE_CONTROL *const lrc = &lc->rc;
+  // Index into spatial+temporal arrays.
+  const int st_idx = svc->spatial_layer_id * svc->number_temporal_layers +
+      svc->temporal_layer_id;
+  const int tl = svc->temporal_layer_id;
+
+  lc->framerate = cpi->framerate / oxcf->ts_rate_decimator[tl];
+  lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+  lrc->max_frame_bandwidth = cpi->rc.max_frame_bandwidth;
+  // Update the average layer frame size (non-cumulative per-frame-bw).
+  if (tl == 0) {
+    lc->avg_frame_size = lrc->avg_frame_bandwidth;
+  } else {
+    const double prev_layer_framerate =
+        cpi->framerate / oxcf->ts_rate_decimator[tl - 1];
+    const int prev_layer_target_bandwidth =
+        oxcf->layer_target_bitrate[st_idx - 1];
+    lc->avg_frame_size =
+        (int)((lc->target_bandwidth - prev_layer_target_bandwidth) /
+              (lc->framerate - prev_layer_framerate));
+  }
+}
+
+void vp9_update_spatial_layer_framerate(VP9_COMP *const cpi, double framerate) {
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  LAYER_CONTEXT *const lc = get_layer_context(cpi);
+  RATE_CONTROL *const lrc = &lc->rc;
+
+  lc->framerate = framerate;
+  lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+  lrc->min_frame_bandwidth = (int)(lrc->avg_frame_bandwidth *
+                                   oxcf->two_pass_vbrmin_section / 100);
+  lrc->max_frame_bandwidth = (int)(((int64_t)lrc->avg_frame_bandwidth *
+                                   oxcf->two_pass_vbrmax_section) / 100);
+  vp9_rc_set_gf_interval_range(cpi, lrc);
+}
+
+void vp9_restore_layer_context(VP9_COMP *const cpi) {
+  LAYER_CONTEXT *const lc = get_layer_context(cpi);
+  const int old_frame_since_key = cpi->rc.frames_since_key;
+  const int old_frame_to_key = cpi->rc.frames_to_key;
+
+  cpi->rc = lc->rc;
+  cpi->twopass = lc->twopass;
+  cpi->oxcf.target_bandwidth = lc->target_bandwidth;
+  cpi->alt_ref_source = lc->alt_ref_source;
+  // Reset the frames_since_key and frames_to_key counters to their values
+  // before the layer restore. Keep these defined for the stream (not layer).
+  if (cpi->svc.number_temporal_layers > 1 ||
+      (cpi->svc.number_spatial_layers > 1 && !is_two_pass_svc(cpi))) {
+    cpi->rc.frames_since_key = old_frame_since_key;
+    cpi->rc.frames_to_key = old_frame_to_key;
+  }
+
+  // For spatial-svc, allow cyclic-refresh to be applied on the spatial layers,
+  // for the base temporal layer.
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
+      cpi->svc.number_spatial_layers > 1 &&
+      cpi->svc.temporal_layer_id == 0) {
+    CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+    signed char *temp = cr->map;
+    uint8_t *temp2 = cr->last_coded_q_map;
+    uint8_t *temp3 = cpi->consec_zero_mv;
+    cr->map = lc->map;
+    lc->map = temp;
+    cr->last_coded_q_map = lc->last_coded_q_map;
+    lc->last_coded_q_map = temp2;
+    cpi->consec_zero_mv = lc->consec_zero_mv;
+    lc->consec_zero_mv = temp3;
+    cr->sb_index = lc->sb_index;
+  }
+}
+
+void vp9_save_layer_context(VP9_COMP *const cpi) {
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  LAYER_CONTEXT *const lc = get_layer_context(cpi);
+
+  lc->rc = cpi->rc;
+  lc->twopass = cpi->twopass;
+  lc->target_bandwidth = (int)oxcf->target_bandwidth;
+  lc->alt_ref_source = cpi->alt_ref_source;
+
+  // For spatial-svc, allow cyclic-refresh to be applied on the spatial layers,
+  // for the base temporal layer.
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
+      cpi->svc.number_spatial_layers > 1 &&
+      cpi->svc.temporal_layer_id == 0) {
+    CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+    signed char *temp = lc->map;
+    uint8_t *temp2 = lc->last_coded_q_map;
+    uint8_t *temp3 = lc->consec_zero_mv;
+    lc->map = cr->map;
+    cr->map = temp;
+    lc->last_coded_q_map = cr->last_coded_q_map;
+    cr->last_coded_q_map = temp2;
+    lc->consec_zero_mv = cpi->consec_zero_mv;
+    cpi->consec_zero_mv = temp3;
+    lc->sb_index = cr->sb_index;
+  }
+}
+
+void vp9_init_second_pass_spatial_svc(VP9_COMP *cpi) {
+  SVC *const svc = &cpi->svc;
+  int i;
+
+  for (i = 0; i < svc->number_spatial_layers; ++i) {
+    TWO_PASS *const twopass = &svc->layer_context[i].twopass;
+
+    svc->spatial_layer_id = i;
+    vp9_init_second_pass(cpi);
+
+    twopass->total_stats.spatial_layer_id = i;
+    twopass->total_left_stats.spatial_layer_id = i;
+  }
+  svc->spatial_layer_id = 0;
+}
+
+void vp9_inc_frame_in_layer(VP9_COMP *const cpi) {
+  LAYER_CONTEXT *const lc =
+      &cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+                              cpi->svc.number_temporal_layers];
+  ++lc->current_video_frame_in_layer;
+  ++lc->frames_from_key_frame;
+  if (cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)
+    ++cpi->svc.current_superframe;
+}
+
+int vp9_is_upper_layer_key_frame(const VP9_COMP *const cpi) {
+  return is_two_pass_svc(cpi) &&
+         cpi->svc.spatial_layer_id > 0 &&
+         cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+                                cpi->svc.number_temporal_layers +
+                                cpi->svc.temporal_layer_id].is_key_frame;
+}
+
+static void get_layer_resolution(const int width_org, const int height_org,
+                                 const int num, const int den,
+                                 int *width_out, int *height_out) {
+  int w, h;
+
+  if (width_out == NULL || height_out == NULL || den == 0)
+    return;
+
+  w = width_org * num / den;
+  h = height_org * num / den;
+
+  // make height and width even to make chrome player happy
+  w += w % 2;
+  h += h % 2;
+
+  *width_out = w;
+  *height_out = h;
+}
+
+// The function sets proper ref_frame_flags, buffer indices, and buffer update
+// variables for temporal layering mode 3 - that does 0-2-1-2 temporal layering
+// scheme.
+static void set_flags_and_fb_idx_for_temporal_mode3(VP9_COMP *const cpi) {
+  int frame_num_within_temporal_struct = 0;
+  int spatial_id, temporal_id;
+  spatial_id = cpi->svc.spatial_layer_id = cpi->svc.spatial_layer_to_encode;
+  frame_num_within_temporal_struct =
+      cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+      cpi->svc.number_temporal_layers].current_video_frame_in_layer % 4;
+  temporal_id = cpi->svc.temporal_layer_id =
+      (frame_num_within_temporal_struct & 1) ? 2 :
+      (frame_num_within_temporal_struct >> 1);
+  cpi->ext_refresh_last_frame = cpi->ext_refresh_golden_frame =
+      cpi->ext_refresh_alt_ref_frame = 0;
+  if (!temporal_id) {
+    cpi->ext_refresh_frame_flags_pending = 1;
+    cpi->ext_refresh_last_frame = 1;
+    if (!spatial_id) {
+      cpi->ref_frame_flags = VP9_LAST_FLAG;
+    } else if (cpi->svc.layer_context[temporal_id].is_key_frame) {
+      // base layer is a key frame.
+      cpi->ref_frame_flags = VP9_LAST_FLAG;
+      cpi->ext_refresh_last_frame = 0;
+      cpi->ext_refresh_golden_frame = 1;
+    } else {
+      cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+    }
+  } else if (temporal_id == 1) {
+    cpi->ext_refresh_frame_flags_pending = 1;
+    cpi->ext_refresh_alt_ref_frame = 1;
+    if (!spatial_id) {
+      cpi->ref_frame_flags = VP9_LAST_FLAG;
+    } else {
+      cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+    }
+  } else {
+    if (frame_num_within_temporal_struct == 1) {
+      // the first tl2 picture
+      if (spatial_id == cpi->svc.number_spatial_layers - 1) {  // top layer
+        cpi->ext_refresh_frame_flags_pending = 1;
+        if (!spatial_id)
+          cpi->ref_frame_flags = VP9_LAST_FLAG;
+        else
+          cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+      } else if (!spatial_id) {
+        cpi->ext_refresh_frame_flags_pending = 1;
+        cpi->ext_refresh_alt_ref_frame = 1;
+        cpi->ref_frame_flags = VP9_LAST_FLAG;
+      } else if (spatial_id < cpi->svc.number_spatial_layers - 1) {
+        cpi->ext_refresh_frame_flags_pending = 1;
+        cpi->ext_refresh_alt_ref_frame = 1;
+        cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+      }
+    } else {
+      //  The second tl2 picture
+      if (spatial_id == cpi->svc.number_spatial_layers - 1) {  // top layer
+        cpi->ext_refresh_frame_flags_pending = 1;
+        if (!spatial_id)
+        cpi->ref_frame_flags = VP9_LAST_FLAG;
+        else
+          cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+      } else if (!spatial_id) {
+        cpi->ext_refresh_frame_flags_pending = 1;
+        cpi->ref_frame_flags = VP9_LAST_FLAG;
+        cpi->ext_refresh_alt_ref_frame = 1;
+      } else {  // top layer
+        cpi->ext_refresh_frame_flags_pending = 1;
+        cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+        cpi->ext_refresh_alt_ref_frame = 1;
+      }
+    }
+  }
+  if (temporal_id == 0) {
+    cpi->lst_fb_idx = spatial_id;
+    if (spatial_id) {
+      if (cpi->svc.layer_context[temporal_id].is_key_frame) {
+        cpi->lst_fb_idx = spatial_id - 1;
+        cpi->gld_fb_idx = spatial_id;
+      } else {
+      cpi->gld_fb_idx = spatial_id - 1;
+      }
+    } else {
+      cpi->gld_fb_idx = 0;
+    }
+    cpi->alt_fb_idx = 0;
+  } else if (temporal_id == 1) {
+    cpi->lst_fb_idx = spatial_id;
+    cpi->gld_fb_idx = cpi->svc.number_spatial_layers + spatial_id - 1;
+    cpi->alt_fb_idx = cpi->svc.number_spatial_layers + spatial_id;
+  } else if (frame_num_within_temporal_struct == 1) {
+    cpi->lst_fb_idx = spatial_id;
+    cpi->gld_fb_idx = cpi->svc.number_spatial_layers + spatial_id - 1;
+    cpi->alt_fb_idx = cpi->svc.number_spatial_layers + spatial_id;
+  } else {
+    cpi->lst_fb_idx = cpi->svc.number_spatial_layers + spatial_id;
+    cpi->gld_fb_idx = cpi->svc.number_spatial_layers + spatial_id - 1;
+    cpi->alt_fb_idx = cpi->svc.number_spatial_layers + spatial_id;
+  }
+}
+
+// The function sets proper ref_frame_flags, buffer indices, and buffer update
+// variables for temporal layering mode 2 - that does 0-1-0-1 temporal layering
+// scheme.
+static void set_flags_and_fb_idx_for_temporal_mode2(VP9_COMP *const cpi) {
+  int spatial_id, temporal_id;
+  spatial_id = cpi->svc.spatial_layer_id = cpi->svc.spatial_layer_to_encode;
+  temporal_id = cpi->svc.temporal_layer_id =
+      cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+      cpi->svc.number_temporal_layers].current_video_frame_in_layer & 1;
+  cpi->ext_refresh_last_frame = cpi->ext_refresh_golden_frame =
+                                cpi->ext_refresh_alt_ref_frame = 0;
+  if (!temporal_id) {
+    cpi->ext_refresh_frame_flags_pending = 1;
+    cpi->ext_refresh_last_frame = 1;
+    if (!spatial_id) {
+      cpi->ref_frame_flags = VP9_LAST_FLAG;
+    } else if (cpi->svc.layer_context[temporal_id].is_key_frame) {
+      // base layer is a key frame.
+      cpi->ref_frame_flags = VP9_LAST_FLAG;
+      cpi->ext_refresh_last_frame = 0;
+      cpi->ext_refresh_golden_frame = 1;
+    } else {
+      cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+    }
+  } else if (temporal_id == 1) {
+    cpi->ext_refresh_frame_flags_pending = 1;
+    cpi->ext_refresh_alt_ref_frame = 1;
+    if (!spatial_id) {
+      cpi->ref_frame_flags = VP9_LAST_FLAG;
+    } else {
+      cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+    }
+  }
+
+  if (temporal_id == 0) {
+    cpi->lst_fb_idx = spatial_id;
+    if (spatial_id) {
+      if (cpi->svc.layer_context[temporal_id].is_key_frame) {
+        cpi->lst_fb_idx = spatial_id - 1;
+        cpi->gld_fb_idx = spatial_id;
+      } else {
+      cpi->gld_fb_idx = spatial_id - 1;
+      }
+    } else {
+      cpi->gld_fb_idx = 0;
+    }
+    cpi->alt_fb_idx = 0;
+  } else if (temporal_id == 1) {
+    cpi->lst_fb_idx = spatial_id;
+    cpi->gld_fb_idx = cpi->svc.number_spatial_layers + spatial_id - 1;
+    cpi->alt_fb_idx = cpi->svc.number_spatial_layers + spatial_id;
+  }
+}
+
+// The function sets proper ref_frame_flags, buffer indices, and buffer update
+// variables for temporal layering mode 0 - that has no temporal layering.
+static void set_flags_and_fb_idx_for_temporal_mode_noLayering(
+    VP9_COMP *const cpi) {
+  int spatial_id;
+  spatial_id = cpi->svc.spatial_layer_id = cpi->svc.spatial_layer_to_encode;
+  cpi->ext_refresh_last_frame =
+      cpi->ext_refresh_golden_frame = cpi->ext_refresh_alt_ref_frame = 0;
+  cpi->ext_refresh_frame_flags_pending = 1;
+  cpi->ext_refresh_last_frame = 1;
+  if (!spatial_id) {
+    cpi->ref_frame_flags = VP9_LAST_FLAG;
+  } else if (cpi->svc.layer_context[0].is_key_frame) {
+    cpi->ref_frame_flags = VP9_LAST_FLAG;
+    cpi->ext_refresh_last_frame = 0;
+    cpi->ext_refresh_golden_frame = 1;
+  } else {
+    cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+  }
+  cpi->lst_fb_idx = spatial_id;
+  if (spatial_id) {
+    if (cpi->svc.layer_context[0].is_key_frame) {
+      cpi->lst_fb_idx = spatial_id - 1;
+      cpi->gld_fb_idx = spatial_id;
+    } else {
+    cpi->gld_fb_idx = spatial_id - 1;
+    }
+  } else {
+    cpi->gld_fb_idx = 0;
+  }
+}
+
+int vp9_one_pass_cbr_svc_start_layer(VP9_COMP *const cpi) {
+  int width = 0, height = 0;
+  LAYER_CONTEXT *lc = NULL;
+  if (cpi->svc.number_spatial_layers > 1)
+    cpi->svc.use_base_mv = 1;
+  cpi->svc.force_zero_mode_spatial_ref = 1;
+
+  if (cpi->svc.temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_0212) {
+    set_flags_and_fb_idx_for_temporal_mode3(cpi);
+  } else if (cpi->svc.temporal_layering_mode ==
+           VP9E_TEMPORAL_LAYERING_MODE_NOLAYERING) {
+    set_flags_and_fb_idx_for_temporal_mode_noLayering(cpi);
+  } else if (cpi->svc.temporal_layering_mode ==
+           VP9E_TEMPORAL_LAYERING_MODE_0101) {
+    set_flags_and_fb_idx_for_temporal_mode2(cpi);
+  } else if (cpi->svc.temporal_layering_mode ==
+      VP9E_TEMPORAL_LAYERING_MODE_BYPASS) {
+    // In the BYPASS/flexible mode, the encoder is relying on the application
+    // to specify, for each spatial layer, the flags and buffer indices for the
+    // layering.
+    // Note that the check (cpi->ext_refresh_frame_flags_pending == 0) is
+    // needed to support the case where the frame flags may be passed in via
+    // vpx_codec_encode(), which can be used for the temporal-only svc case.
+    // TODO(marpan): Consider adding an enc_config parameter to better handle
+    // this case.
+    if (cpi->ext_refresh_frame_flags_pending == 0) {
+      int sl;
+      cpi->svc.spatial_layer_id = cpi->svc.spatial_layer_to_encode;
+      sl = cpi->svc.spatial_layer_id;
+      vp9_apply_encoding_flags(cpi, cpi->svc.ext_frame_flags[sl]);
+      cpi->lst_fb_idx = cpi->svc.ext_lst_fb_idx[sl];
+      cpi->gld_fb_idx = cpi->svc.ext_gld_fb_idx[sl];
+      cpi->alt_fb_idx = cpi->svc.ext_alt_fb_idx[sl];
+    }
+  }
+
+  if (cpi->svc.spatial_layer_id == cpi->svc.first_spatial_layer_to_encode)
+    cpi->svc.rc_drop_superframe = 0;
+
+  lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+                               cpi->svc.number_temporal_layers +
+                               cpi->svc.temporal_layer_id];
+
+  // Setting the worst/best_quality via the encoder control: SET_SVC_PARAMETERS,
+  // only for non-BYPASS mode for now.
+  if (cpi->svc.temporal_layering_mode != VP9E_TEMPORAL_LAYERING_MODE_BYPASS) {
+    RATE_CONTROL *const lrc = &lc->rc;
+    lrc->worst_quality = vp9_quantizer_to_qindex(lc->max_q);
+    lrc->best_quality =  vp9_quantizer_to_qindex(lc->min_q);
+  }
+
+  get_layer_resolution(cpi->oxcf.width, cpi->oxcf.height,
+                       lc->scaling_factor_num, lc->scaling_factor_den,
+                       &width, &height);
+
+  if (vp9_set_size_literal(cpi, width, height) != 0)
+    return VPX_CODEC_INVALID_PARAM;
+
+  return 0;
+}
+
+#if CONFIG_SPATIAL_SVC
+#define SMALL_FRAME_FB_IDX 7
+
+int vp9_svc_start_frame(VP9_COMP *const cpi) {
+  int width = 0, height = 0;
+  LAYER_CONTEXT *lc;
+  struct lookahead_entry *buf;
+  int count = 1 << (cpi->svc.number_temporal_layers - 1);
+
+  cpi->svc.spatial_layer_id = cpi->svc.spatial_layer_to_encode;
+  lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
+
+  cpi->svc.temporal_layer_id = 0;
+  while ((lc->current_video_frame_in_layer % count) != 0) {
+    ++cpi->svc.temporal_layer_id;
+    count >>= 1;
+  }
+
+  cpi->ref_frame_flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
+
+  cpi->lst_fb_idx = cpi->svc.spatial_layer_id;
+
+  if (cpi->svc.spatial_layer_id == 0)
+    cpi->gld_fb_idx = (lc->gold_ref_idx >= 0) ?
+                      lc->gold_ref_idx : cpi->lst_fb_idx;
+  else
+    cpi->gld_fb_idx = cpi->svc.spatial_layer_id - 1;
+
+  if (lc->current_video_frame_in_layer == 0) {
+    if (cpi->svc.spatial_layer_id >= 2) {
+      cpi->alt_fb_idx = cpi->svc.spatial_layer_id - 2;
+    } else {
+      cpi->alt_fb_idx = cpi->lst_fb_idx;
+      cpi->ref_frame_flags &= (~VP9_LAST_FLAG & ~VP9_ALT_FLAG);
+    }
+  } else {
+    if (cpi->oxcf.ss_enable_auto_arf[cpi->svc.spatial_layer_id]) {
+      cpi->alt_fb_idx = lc->alt_ref_idx;
+      if (!lc->has_alt_frame)
+        cpi->ref_frame_flags &= (~VP9_ALT_FLAG);
+    } else {
+      // Find a proper alt_fb_idx for layers that don't have alt ref frame
+      if (cpi->svc.spatial_layer_id == 0) {
+        cpi->alt_fb_idx = cpi->lst_fb_idx;
+      } else {
+        LAYER_CONTEXT *lc_lower =
+            &cpi->svc.layer_context[cpi->svc.spatial_layer_id - 1];
+
+        if (cpi->oxcf.ss_enable_auto_arf[cpi->svc.spatial_layer_id - 1] &&
+            lc_lower->alt_ref_source != NULL)
+          cpi->alt_fb_idx = lc_lower->alt_ref_idx;
+        else if (cpi->svc.spatial_layer_id >= 2)
+          cpi->alt_fb_idx = cpi->svc.spatial_layer_id - 2;
+        else
+          cpi->alt_fb_idx = cpi->lst_fb_idx;
+      }
+    }
+  }
+
+  get_layer_resolution(cpi->oxcf.width, cpi->oxcf.height,
+                       lc->scaling_factor_num, lc->scaling_factor_den,
+                       &width, &height);
+
+  // Workaround for multiple frame contexts. In some frames we can't use prev_mi
+  // since its previous frame could be changed during decoding time. The idea is
+  // we put a empty invisible frame in front of them, then we will not use
+  // prev_mi when encoding these frames.
+
+  buf = vp9_lookahead_peek(cpi->lookahead, 0);
+  if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2 &&
+      cpi->svc.encode_empty_frame_state == NEED_TO_ENCODE &&
+      lc->rc.frames_to_key != 0 &&
+      !(buf != NULL && (buf->flags & VPX_EFLAG_FORCE_KF))) {
+    if ((cpi->svc.number_temporal_layers > 1 &&
+         cpi->svc.temporal_layer_id < cpi->svc.number_temporal_layers - 1) ||
+        (cpi->svc.number_spatial_layers > 1 &&
+         cpi->svc.spatial_layer_id == 0)) {
+      struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead, 0);
+
+      if (buf != NULL) {
+        cpi->svc.empty_frame.ts_start = buf->ts_start;
+        cpi->svc.empty_frame.ts_end = buf->ts_end;
+        cpi->svc.encode_empty_frame_state = ENCODING;
+        cpi->common.show_frame = 0;
+        cpi->ref_frame_flags = 0;
+        cpi->common.frame_type = INTER_FRAME;
+        cpi->lst_fb_idx =
+            cpi->gld_fb_idx = cpi->alt_fb_idx = SMALL_FRAME_FB_IDX;
+
+        if (cpi->svc.encode_intra_empty_frame != 0)
+          cpi->common.intra_only = 1;
+
+        width = SMALL_FRAME_WIDTH;
+        height = SMALL_FRAME_HEIGHT;
+      }
+    }
+  }
+
+  cpi->oxcf.worst_allowed_q = vp9_quantizer_to_qindex(lc->max_q);
+  cpi->oxcf.best_allowed_q = vp9_quantizer_to_qindex(lc->min_q);
+
+  vp9_change_config(cpi, &cpi->oxcf);
+
+  if (vp9_set_size_literal(cpi, width, height) != 0)
+    return VPX_CODEC_INVALID_PARAM;
+
+  vp9_set_high_precision_mv(cpi, 1);
+
+  cpi->alt_ref_source = get_layer_context(cpi)->alt_ref_source;
+
+  return 0;
+}
+
+#undef SMALL_FRAME_FB_IDX
+#endif  // CONFIG_SPATIAL_SVC
+
+struct lookahead_entry *vp9_svc_lookahead_pop(VP9_COMP *const cpi,
+                                              struct lookahead_ctx *ctx,
+                                              int drain) {
+  struct lookahead_entry *buf = NULL;
+  if (ctx->sz && (drain || ctx->sz == ctx->max_sz - MAX_PRE_FRAMES)) {
+    buf = vp9_lookahead_peek(ctx, 0);
+    if (buf != NULL) {
+      // Only remove the buffer when pop the highest layer.
+      if (cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1) {
+        vp9_lookahead_pop(ctx, drain);
+      }
+    }
+  }
+  return buf;
+}
+
+void vp9_free_svc_cyclic_refresh(VP9_COMP *const cpi) {
+  int sl, tl;
+  SVC *const svc = &cpi->svc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
+    for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
+      int layer = LAYER_IDS_TO_IDX(sl, tl, oxcf->ts_number_layers);
+      LAYER_CONTEXT *const lc = &svc->layer_context[layer];
+        if (lc->map)
+          vpx_free(lc->map);
+        if (lc->last_coded_q_map)
+          vpx_free(lc->last_coded_q_map);
+        if (lc->consec_zero_mv)
+          vpx_free(lc->consec_zero_mv);
+    }
+  }
+}
+
+// Reset on key frame: reset counters, references and buffer updates.
+void vp9_svc_reset_key_frame(VP9_COMP *const cpi) {
+  int sl, tl;
+  SVC *const svc = &cpi->svc;
+  LAYER_CONTEXT *lc = NULL;
+  for (sl = 0; sl < svc->number_spatial_layers; ++sl) {
+    for (tl = 0; tl < svc->number_temporal_layers; ++tl) {
+      lc = &cpi->svc.layer_context[sl * svc->number_temporal_layers + tl];
+      lc->current_video_frame_in_layer = 0;
+      lc->frames_from_key_frame = 0;
+    }
+  }
+  if (svc->temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_0212) {
+    set_flags_and_fb_idx_for_temporal_mode3(cpi);
+  } else if (svc->temporal_layering_mode ==
+             VP9E_TEMPORAL_LAYERING_MODE_NOLAYERING) {
+     set_flags_and_fb_idx_for_temporal_mode_noLayering(cpi);
+  } else if (svc->temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_0101) {
+     set_flags_and_fb_idx_for_temporal_mode2(cpi);
+  }
+  vp9_update_temporal_layer_framerate(cpi);
+  vp9_restore_layer_context(cpi);
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_svc_layercontext.h b/libvpx/libvpx/vp9/encoder/vp9_svc_layercontext.h
new file mode 100644
index 0000000..9f386fb
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_svc_layercontext.h
@@ -0,0 +1,145 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_SVC_LAYERCONTEXT_H_
+#define VP9_ENCODER_VP9_SVC_LAYERCONTEXT_H_
+
+#include "vpx/vpx_encoder.h"
+
+#include "vp9/encoder/vp9_ratectrl.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  RATE_CONTROL rc;
+  int target_bandwidth;
+  int spatial_layer_target_bandwidth;  // Target for the spatial layer.
+  double framerate;
+  int avg_frame_size;
+  int max_q;
+  int min_q;
+  int scaling_factor_num;
+  int scaling_factor_den;
+  TWO_PASS twopass;
+  vpx_fixed_buf_t rc_twopass_stats_in;
+  unsigned int current_video_frame_in_layer;
+  int is_key_frame;
+  int frames_from_key_frame;
+  FRAME_TYPE last_frame_type;
+  struct lookahead_entry  *alt_ref_source;
+  int alt_ref_idx;
+  int gold_ref_idx;
+  int has_alt_frame;
+  size_t layer_size;
+  struct vpx_psnr_pkt psnr_pkt;
+  // Cyclic refresh parameters (aq-mode=3), that need to be updated per-frame.
+  int sb_index;
+  signed char *map;
+  uint8_t *last_coded_q_map;
+  uint8_t *consec_zero_mv;
+} LAYER_CONTEXT;
+
+typedef struct {
+  int spatial_layer_id;
+  int temporal_layer_id;
+  int number_spatial_layers;
+  int number_temporal_layers;
+
+  int spatial_layer_to_encode;
+  int first_spatial_layer_to_encode;
+  int rc_drop_superframe;
+
+  // Workaround for multiple frame contexts
+  enum {
+    ENCODED = 0,
+    ENCODING,
+    NEED_TO_ENCODE
+  }encode_empty_frame_state;
+  struct lookahead_entry empty_frame;
+  int encode_intra_empty_frame;
+
+  // Store scaled source frames to be used for temporal filter to generate
+  // a alt ref frame.
+  YV12_BUFFER_CONFIG scaled_frames[MAX_LAG_BUFFERS];
+  // Temp buffer used for 2-stage down-sampling, for real-time mode.
+  YV12_BUFFER_CONFIG scaled_temp;
+
+  // Layer context used for rate control in one pass temporal CBR mode or
+  // two pass spatial mode.
+  LAYER_CONTEXT layer_context[VPX_MAX_LAYERS];
+  // Indicates what sort of temporal layering is used.
+  // Currently, this only works for CBR mode.
+  VP9E_TEMPORAL_LAYERING_MODE temporal_layering_mode;
+  // Frame flags and buffer indexes for each spatial layer, set by the
+  // application (external settings).
+  int ext_frame_flags[VPX_MAX_LAYERS];
+  int ext_lst_fb_idx[VPX_MAX_LAYERS];
+  int ext_gld_fb_idx[VPX_MAX_LAYERS];
+  int ext_alt_fb_idx[VPX_MAX_LAYERS];
+  int ref_frame_index[REF_FRAMES];
+  int force_zero_mode_spatial_ref;
+  int current_superframe;
+  int use_base_mv;
+} SVC;
+
+struct VP9_COMP;
+
+// Initialize layer context data from init_config().
+void vp9_init_layer_context(struct VP9_COMP *const cpi);
+
+// Update the layer context from a change_config() call.
+void vp9_update_layer_context_change_config(struct VP9_COMP *const cpi,
+                                            const int target_bandwidth);
+
+// Prior to encoding the frame, update framerate-related quantities
+// for the current temporal layer.
+void vp9_update_temporal_layer_framerate(struct VP9_COMP *const cpi);
+
+// Update framerate-related quantities for the current spatial layer.
+void vp9_update_spatial_layer_framerate(struct VP9_COMP *const cpi,
+                                        double framerate);
+
+// Prior to encoding the frame, set the layer context, for the current layer
+// to be encoded, to the cpi struct.
+void vp9_restore_layer_context(struct VP9_COMP *const cpi);
+
+// Save the layer context after encoding the frame.
+void vp9_save_layer_context(struct VP9_COMP *const cpi);
+
+// Initialize second pass rc for spatial svc.
+void vp9_init_second_pass_spatial_svc(struct VP9_COMP *cpi);
+
+// Increment number of video frames in layer
+void vp9_inc_frame_in_layer(struct VP9_COMP *const cpi);
+
+// Check if current layer is key frame in spatial upper layer
+int vp9_is_upper_layer_key_frame(const struct VP9_COMP *const cpi);
+
+// Get the next source buffer to encode
+struct lookahead_entry *vp9_svc_lookahead_pop(struct VP9_COMP *const cpi,
+                                              struct lookahead_ctx *ctx,
+                                              int drain);
+
+// Start a frame and initialize svc parameters
+int vp9_svc_start_frame(struct VP9_COMP *const cpi);
+
+int vp9_one_pass_cbr_svc_start_layer(struct VP9_COMP *const cpi);
+
+void vp9_free_svc_cyclic_refresh(struct VP9_COMP *const cpi);
+
+void vp9_svc_reset_key_frame(struct VP9_COMP *const cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SVC_LAYERCONTEXT_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_temporal_filter.c b/libvpx/libvpx/vp9/encoder/vp9_temporal_filter.c
new file mode 100644
index 0000000..b6323e0
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_temporal_filter.c
@@ -0,0 +1,796 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <limits.h>
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_extend.h"
+#include "vp9/encoder/vp9_firstpass.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_segmentation.h"
+#include "vp9/encoder/vp9_temporal_filter.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/vpx_timer.h"
+#include "vpx_scale/vpx_scale.h"
+
+static int fixed_divide[512];
+
+static void temporal_filter_predictors_mb_c(MACROBLOCKD *xd,
+                                            uint8_t *y_mb_ptr,
+                                            uint8_t *u_mb_ptr,
+                                            uint8_t *v_mb_ptr,
+                                            int stride,
+                                            int uv_block_width,
+                                            int uv_block_height,
+                                            int mv_row,
+                                            int mv_col,
+                                            uint8_t *pred,
+                                            struct scale_factors *scale,
+                                            int x, int y) {
+  const int which_mv = 0;
+  const MV mv = { mv_row, mv_col };
+  const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP_SHARP];
+
+  enum mv_precision mv_precision_uv;
+  int uv_stride;
+  if (uv_block_width == 8) {
+    uv_stride = (stride + 1) >> 1;
+    mv_precision_uv = MV_PRECISION_Q4;
+  } else {
+    uv_stride = stride;
+    mv_precision_uv = MV_PRECISION_Q3;
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vp9_highbd_build_inter_predictor(y_mb_ptr, stride,
+                                     &pred[0], 16,
+                                     &mv,
+                                     scale,
+                                     16, 16,
+                                     which_mv,
+                                     kernel, MV_PRECISION_Q3, x, y, xd->bd);
+
+    vp9_highbd_build_inter_predictor(u_mb_ptr, uv_stride,
+                                     &pred[256], uv_block_width,
+                                     &mv,
+                                     scale,
+                                     uv_block_width, uv_block_height,
+                                     which_mv,
+                                     kernel, mv_precision_uv, x, y, xd->bd);
+
+    vp9_highbd_build_inter_predictor(v_mb_ptr, uv_stride,
+                                     &pred[512], uv_block_width,
+                                     &mv,
+                                     scale,
+                                     uv_block_width, uv_block_height,
+                                     which_mv,
+                                     kernel, mv_precision_uv, x, y, xd->bd);
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  (void)xd;
+  vp9_build_inter_predictor(y_mb_ptr, stride,
+                            &pred[0], 16,
+                            &mv,
+                            scale,
+                            16, 16,
+                            which_mv,
+                            kernel, MV_PRECISION_Q3, x, y);
+
+  vp9_build_inter_predictor(u_mb_ptr, uv_stride,
+                            &pred[256], uv_block_width,
+                            &mv,
+                            scale,
+                            uv_block_width, uv_block_height,
+                            which_mv,
+                            kernel, mv_precision_uv, x, y);
+
+  vp9_build_inter_predictor(v_mb_ptr, uv_stride,
+                            &pred[512], uv_block_width,
+                            &mv,
+                            scale,
+                            uv_block_width, uv_block_height,
+                            which_mv,
+                            kernel, mv_precision_uv, x, y);
+}
+
+void vp9_temporal_filter_init(void) {
+  int i;
+
+  fixed_divide[0] = 0;
+  for (i = 1; i < 512; ++i)
+    fixed_divide[i] = 0x80000 / i;
+}
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1,
+                                 unsigned int stride,
+                                 uint8_t *frame2,
+                                 unsigned int block_width,
+                                 unsigned int block_height,
+                                 int strength,
+                                 int filter_weight,
+                                 unsigned int *accumulator,
+                                 uint16_t *count) {
+  unsigned int i, j, k;
+  int modifier;
+  int byte = 0;
+  const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
+
+  for (i = 0, k = 0; i < block_height; i++) {
+    for (j = 0; j < block_width; j++, k++) {
+      int pixel_value = *frame2;
+
+      // non-local mean approach
+      int diff_sse[9] = { 0 };
+      int idx, idy, index = 0;
+
+      for (idy = -1; idy <= 1; ++idy) {
+        for (idx = -1; idx <= 1; ++idx) {
+          int row = (int)i + idy;
+          int col = (int)j + idx;
+
+          if (row >= 0 && row < (int)block_height &&
+              col >= 0 && col < (int)block_width) {
+            int diff = frame1[byte + idy * (int)stride + idx] -
+                frame2[idy * (int)block_width + idx];
+            diff_sse[index] = diff * diff;
+            ++index;
+          }
+        }
+      }
+
+      assert(index > 0);
+
+      modifier = 0;
+      for (idx = 0; idx < 9; ++idx)
+        modifier += diff_sse[idx];
+
+      modifier *= 3;
+      modifier /= index;
+
+      ++frame2;
+
+      modifier  += rounding;
+      modifier >>= strength;
+
+      if (modifier > 16)
+        modifier = 16;
+
+      modifier = 16 - modifier;
+      modifier *= filter_weight;
+
+      count[k] += modifier;
+      accumulator[k] += modifier * pixel_value;
+
+      byte++;
+    }
+
+    byte += stride - block_width;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_temporal_filter_apply_c(uint8_t *frame1_8,
+                                        unsigned int stride,
+                                        uint8_t *frame2_8,
+                                        unsigned int block_width,
+                                        unsigned int block_height,
+                                        int strength,
+                                        int filter_weight,
+                                        unsigned int *accumulator,
+                                        uint16_t *count) {
+  uint16_t *frame1 = CONVERT_TO_SHORTPTR(frame1_8);
+  uint16_t *frame2 = CONVERT_TO_SHORTPTR(frame2_8);
+  unsigned int i, j, k;
+  int modifier;
+  int byte = 0;
+  const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
+
+  for (i = 0, k = 0; i < block_height; i++) {
+    for (j = 0; j < block_width; j++, k++) {
+      int pixel_value = *frame2;
+      int diff_sse[9] = { 0 };
+      int idx, idy, index = 0;
+
+      for (idy = -1; idy <= 1; ++idy) {
+        for (idx = -1; idx <= 1; ++idx) {
+          int row = (int)i + idy;
+          int col = (int)j + idx;
+
+          if (row >= 0 && row < (int)block_height &&
+              col >= 0 && col < (int)block_width) {
+            int diff = frame1[byte + idy * (int)stride + idx] -
+                frame2[idy * (int)block_width + idx];
+            diff_sse[index] = diff * diff;
+            ++index;
+          }
+        }
+      }
+      assert(index > 0);
+
+      modifier = 0;
+      for (idx = 0; idx < 9; ++idx)
+        modifier += diff_sse[idx];
+
+      modifier *= 3;
+      modifier /= index;
+
+      ++frame2;
+      modifier += rounding;
+      modifier >>= strength;
+
+      if (modifier > 16)
+        modifier = 16;
+
+      modifier = 16 - modifier;
+      modifier *= filter_weight;
+
+      count[k] += modifier;
+      accumulator[k] += modifier * pixel_value;
+
+      byte++;
+    }
+
+    byte += stride - block_width;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static int temporal_filter_find_matching_mb_c(VP9_COMP *cpi,
+                                              uint8_t *arf_frame_buf,
+                                              uint8_t *frame_ptr_buf,
+                                              int stride) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
+  const SEARCH_METHODS old_search_method = mv_sf->search_method;
+  int step_param;
+  int sadpb = x->sadperbit16;
+  int bestsme = INT_MAX;
+  uint32_t distortion;
+  uint32_t sse;
+  int cost_list[5];
+
+  MV best_ref_mv1 = {0, 0};
+  MV best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */
+  MV *ref_mv = &x->e_mbd.mi[0]->bmi[0].as_mv[0].as_mv;
+
+  // Save input state
+  struct buf_2d src = x->plane[0].src;
+  struct buf_2d pre = xd->plane[0].pre[0];
+
+  best_ref_mv1_full.col = best_ref_mv1.col >> 3;
+  best_ref_mv1_full.row = best_ref_mv1.row >> 3;
+
+  // Setup frame pointers
+  x->plane[0].src.buf = arf_frame_buf;
+  x->plane[0].src.stride = stride;
+  xd->plane[0].pre[0].buf = frame_ptr_buf;
+  xd->plane[0].pre[0].stride = stride;
+
+  step_param = mv_sf->reduce_first_step_size;
+  step_param = VPXMIN(step_param, MAX_MVSEARCH_STEPS - 2);
+
+  mv_sf->search_method = HEX;
+  vp9_full_pixel_search(cpi, x, BLOCK_16X16, &best_ref_mv1_full, step_param,
+                        sadpb, cond_cost_list(cpi, cost_list), &best_ref_mv1,
+                        ref_mv, 0, 0);
+  mv_sf->search_method = old_search_method;
+
+  // Ignore mv costing by sending NULL pointer instead of cost array
+  bestsme = cpi->find_fractional_mv_step(x, ref_mv,
+                                         &best_ref_mv1,
+                                         cpi->common.allow_high_precision_mv,
+                                         x->errorperbit,
+                                         &cpi->fn_ptr[BLOCK_16X16],
+                                         0, mv_sf->subpel_iters_per_step,
+                                         cond_cost_list(cpi, cost_list),
+                                         NULL, NULL,
+                                         &distortion, &sse, NULL, 0, 0);
+
+  // Restore input state
+  x->plane[0].src = src;
+  xd->plane[0].pre[0] = pre;
+
+  return bestsme;
+}
+
+static void temporal_filter_iterate_c(VP9_COMP *cpi,
+                                      YV12_BUFFER_CONFIG **frames,
+                                      int frame_count,
+                                      int alt_ref_index,
+                                      int strength,
+                                      struct scale_factors *scale) {
+  int byte;
+  int frame;
+  int mb_col, mb_row;
+  unsigned int filter_weight;
+  int mb_cols = (frames[alt_ref_index]->y_crop_width + 15) >> 4;
+  int mb_rows = (frames[alt_ref_index]->y_crop_height + 15) >> 4;
+  int mb_y_offset = 0;
+  int mb_uv_offset = 0;
+  DECLARE_ALIGNED(16, unsigned int, accumulator[16 * 16 * 3]);
+  DECLARE_ALIGNED(16, uint16_t, count[16 * 16 * 3]);
+  MACROBLOCKD *mbd = &cpi->td.mb.e_mbd;
+  YV12_BUFFER_CONFIG *f = frames[alt_ref_index];
+  uint8_t *dst1, *dst2;
+#if CONFIG_VP9_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t,  predictor16[16 * 16 * 3]);
+  DECLARE_ALIGNED(16, uint8_t,  predictor8[16 * 16 * 3]);
+  uint8_t *predictor;
+#else
+  DECLARE_ALIGNED(16, uint8_t,  predictor[16 * 16 * 3]);
+#endif
+  const int mb_uv_height = 16 >> mbd->plane[1].subsampling_y;
+  const int mb_uv_width  = 16 >> mbd->plane[1].subsampling_x;
+
+  // Save input state
+  uint8_t* input_buffer[MAX_MB_PLANE];
+  int i;
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    predictor = CONVERT_TO_BYTEPTR(predictor16);
+  } else {
+    predictor = predictor8;
+  }
+#endif
+
+  for (i = 0; i < MAX_MB_PLANE; i++)
+    input_buffer[i] = mbd->plane[i].pre[0].buf;
+
+  for (mb_row = 0; mb_row < mb_rows; mb_row++) {
+    // Source frames are extended to 16 pixels. This is different than
+    //  L/A/G reference frames that have a border of 32 (VP9ENCBORDERINPIXELS)
+    // A 6/8 tap filter is used for motion search.  This requires 2 pixels
+    //  before and 3 pixels after.  So the largest Y mv on a border would
+    //  then be 16 - VP9_INTERP_EXTEND. The UV blocks are half the size of the
+    //  Y and therefore only extended by 8.  The largest mv that a UV block
+    //  can support is 8 - VP9_INTERP_EXTEND.  A UV mv is half of a Y mv.
+    //  (16 - VP9_INTERP_EXTEND) >> 1 which is greater than
+    //  8 - VP9_INTERP_EXTEND.
+    // To keep the mv in play for both Y and UV planes the max that it
+    //  can be on a border is therefore 16 - (2*VP9_INTERP_EXTEND+1).
+    cpi->td.mb.mv_row_min = -((mb_row * 16) + (17 - 2 * VP9_INTERP_EXTEND));
+    cpi->td.mb.mv_row_max = ((mb_rows - 1 - mb_row) * 16)
+                         + (17 - 2 * VP9_INTERP_EXTEND);
+
+    for (mb_col = 0; mb_col < mb_cols; mb_col++) {
+      int i, j, k;
+      int stride;
+
+      memset(accumulator, 0, 16 * 16 * 3 * sizeof(accumulator[0]));
+      memset(count, 0, 16 * 16 * 3 * sizeof(count[0]));
+
+      cpi->td.mb.mv_col_min = -((mb_col * 16) + (17 - 2 * VP9_INTERP_EXTEND));
+      cpi->td.mb.mv_col_max = ((mb_cols - 1 - mb_col) * 16)
+                           + (17 - 2 * VP9_INTERP_EXTEND);
+
+      for (frame = 0; frame < frame_count; frame++) {
+        const int thresh_low  = 10000;
+        const int thresh_high = 20000;
+
+        if (frames[frame] == NULL)
+          continue;
+
+        mbd->mi[0]->bmi[0].as_mv[0].as_mv.row = 0;
+        mbd->mi[0]->bmi[0].as_mv[0].as_mv.col = 0;
+
+        if (frame == alt_ref_index) {
+          filter_weight = 2;
+        } else {
+          // Find best match in this frame by MC
+          int err = temporal_filter_find_matching_mb_c(cpi,
+              frames[alt_ref_index]->y_buffer + mb_y_offset,
+              frames[frame]->y_buffer + mb_y_offset,
+              frames[frame]->y_stride);
+
+          // Assign higher weight to matching MB if it's error
+          // score is lower. If not applying MC default behavior
+          // is to weight all MBs equal.
+          filter_weight = err < thresh_low
+                          ? 2 : err < thresh_high ? 1 : 0;
+        }
+
+        if (filter_weight != 0) {
+          // Construct the predictors
+          temporal_filter_predictors_mb_c(mbd,
+              frames[frame]->y_buffer + mb_y_offset,
+              frames[frame]->u_buffer + mb_uv_offset,
+              frames[frame]->v_buffer + mb_uv_offset,
+              frames[frame]->y_stride,
+              mb_uv_width, mb_uv_height,
+              mbd->mi[0]->bmi[0].as_mv[0].as_mv.row,
+              mbd->mi[0]->bmi[0].as_mv[0].as_mv.col,
+              predictor, scale,
+              mb_col * 16, mb_row * 16);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+          if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            int adj_strength = strength + 2 * (mbd->bd - 8);
+            // Apply the filter (YUV)
+            vp9_highbd_temporal_filter_apply_c(f->y_buffer + mb_y_offset,
+                                               f->y_stride,
+                                               predictor, 16, 16, adj_strength,
+                                               filter_weight,
+                                               accumulator, count);
+            vp9_highbd_temporal_filter_apply_c(f->u_buffer + mb_uv_offset,
+                                               f->uv_stride, predictor + 256,
+                                               mb_uv_width, mb_uv_height,
+                                               adj_strength, filter_weight,
+                                               accumulator + 256, count + 256);
+            vp9_highbd_temporal_filter_apply_c(f->v_buffer + mb_uv_offset,
+                                               f->uv_stride, predictor + 512,
+                                               mb_uv_width, mb_uv_height,
+                                               adj_strength, filter_weight,
+                                               accumulator + 512, count + 512);
+          } else {
+            // Apply the filter (YUV)
+            vp9_temporal_filter_apply_c(f->y_buffer + mb_y_offset, f->y_stride,
+                                        predictor, 16, 16,
+                                        strength, filter_weight,
+                                        accumulator, count);
+            vp9_temporal_filter_apply_c(f->u_buffer + mb_uv_offset,
+                                        f->uv_stride,
+                                        predictor + 256,
+                                        mb_uv_width, mb_uv_height, strength,
+                                        filter_weight, accumulator + 256,
+                                        count + 256);
+            vp9_temporal_filter_apply_c(f->v_buffer + mb_uv_offset,
+                                        f->uv_stride,
+                                        predictor + 512,
+                                        mb_uv_width, mb_uv_height, strength,
+                                        filter_weight, accumulator + 512,
+                                        count + 512);
+          }
+#else
+          // Apply the filter (YUV)
+          // TODO(jingning): Need SIMD optimization for this.
+          vp9_temporal_filter_apply_c(f->y_buffer + mb_y_offset, f->y_stride,
+                                      predictor, 16, 16,
+                                      strength, filter_weight,
+                                      accumulator, count);
+          vp9_temporal_filter_apply_c(f->u_buffer + mb_uv_offset, f->uv_stride,
+                                      predictor + 256,
+                                      mb_uv_width, mb_uv_height, strength,
+                                      filter_weight, accumulator + 256,
+                                      count + 256);
+          vp9_temporal_filter_apply_c(f->v_buffer + mb_uv_offset, f->uv_stride,
+                                      predictor + 512,
+                                      mb_uv_width, mb_uv_height, strength,
+                                      filter_weight, accumulator + 512,
+                                      count + 512);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        }
+      }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        uint16_t *dst1_16;
+        uint16_t *dst2_16;
+        // Normalize filter output to produce AltRef frame
+        dst1 = cpi->alt_ref_buffer.y_buffer;
+        dst1_16 = CONVERT_TO_SHORTPTR(dst1);
+        stride = cpi->alt_ref_buffer.y_stride;
+        byte = mb_y_offset;
+        for (i = 0, k = 0; i < 16; i++) {
+          for (j = 0; j < 16; j++, k++) {
+            unsigned int pval = accumulator[k] + (count[k] >> 1);
+            pval *= fixed_divide[count[k]];
+            pval >>= 19;
+
+            dst1_16[byte] = (uint16_t)pval;
+
+            // move to next pixel
+            byte++;
+          }
+
+          byte += stride - 16;
+        }
+
+        dst1 = cpi->alt_ref_buffer.u_buffer;
+        dst2 = cpi->alt_ref_buffer.v_buffer;
+        dst1_16 = CONVERT_TO_SHORTPTR(dst1);
+        dst2_16 = CONVERT_TO_SHORTPTR(dst2);
+        stride = cpi->alt_ref_buffer.uv_stride;
+        byte = mb_uv_offset;
+        for (i = 0, k = 256; i < mb_uv_height; i++) {
+          for (j = 0; j < mb_uv_width; j++, k++) {
+            int m = k + 256;
+
+            // U
+            unsigned int pval = accumulator[k] + (count[k] >> 1);
+            pval *= fixed_divide[count[k]];
+            pval >>= 19;
+            dst1_16[byte] = (uint16_t)pval;
+
+            // V
+            pval = accumulator[m] + (count[m] >> 1);
+            pval *= fixed_divide[count[m]];
+            pval >>= 19;
+            dst2_16[byte] = (uint16_t)pval;
+
+            // move to next pixel
+            byte++;
+          }
+
+          byte += stride - mb_uv_width;
+        }
+      } else {
+        // Normalize filter output to produce AltRef frame
+        dst1 = cpi->alt_ref_buffer.y_buffer;
+        stride = cpi->alt_ref_buffer.y_stride;
+        byte = mb_y_offset;
+        for (i = 0, k = 0; i < 16; i++) {
+          for (j = 0; j < 16; j++, k++) {
+            unsigned int pval = accumulator[k] + (count[k] >> 1);
+            pval *= fixed_divide[count[k]];
+            pval >>= 19;
+
+            dst1[byte] = (uint8_t)pval;
+
+            // move to next pixel
+            byte++;
+          }
+          byte += stride - 16;
+        }
+
+        dst1 = cpi->alt_ref_buffer.u_buffer;
+        dst2 = cpi->alt_ref_buffer.v_buffer;
+        stride = cpi->alt_ref_buffer.uv_stride;
+        byte = mb_uv_offset;
+        for (i = 0, k = 256; i < mb_uv_height; i++) {
+          for (j = 0; j < mb_uv_width; j++, k++) {
+            int m = k + 256;
+
+            // U
+            unsigned int pval = accumulator[k] + (count[k] >> 1);
+            pval *= fixed_divide[count[k]];
+            pval >>= 19;
+            dst1[byte] = (uint8_t)pval;
+
+            // V
+            pval = accumulator[m] + (count[m] >> 1);
+            pval *= fixed_divide[count[m]];
+            pval >>= 19;
+            dst2[byte] = (uint8_t)pval;
+
+            // move to next pixel
+            byte++;
+          }
+          byte += stride - mb_uv_width;
+        }
+      }
+#else
+      // Normalize filter output to produce AltRef frame
+      dst1 = cpi->alt_ref_buffer.y_buffer;
+      stride = cpi->alt_ref_buffer.y_stride;
+      byte = mb_y_offset;
+      for (i = 0, k = 0; i < 16; i++) {
+        for (j = 0; j < 16; j++, k++) {
+          unsigned int pval = accumulator[k] + (count[k] >> 1);
+          pval *= fixed_divide[count[k]];
+          pval >>= 19;
+
+          dst1[byte] = (uint8_t)pval;
+
+          // move to next pixel
+          byte++;
+        }
+        byte += stride - 16;
+      }
+
+      dst1 = cpi->alt_ref_buffer.u_buffer;
+      dst2 = cpi->alt_ref_buffer.v_buffer;
+      stride = cpi->alt_ref_buffer.uv_stride;
+      byte = mb_uv_offset;
+      for (i = 0, k = 256; i < mb_uv_height; i++) {
+        for (j = 0; j < mb_uv_width; j++, k++) {
+          int m = k + 256;
+
+          // U
+          unsigned int pval = accumulator[k] + (count[k] >> 1);
+          pval *= fixed_divide[count[k]];
+          pval >>= 19;
+          dst1[byte] = (uint8_t)pval;
+
+          // V
+          pval = accumulator[m] + (count[m] >> 1);
+          pval *= fixed_divide[count[m]];
+          pval >>= 19;
+          dst2[byte] = (uint8_t)pval;
+
+          // move to next pixel
+          byte++;
+        }
+        byte += stride - mb_uv_width;
+      }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      mb_y_offset += 16;
+      mb_uv_offset += mb_uv_width;
+    }
+    mb_y_offset += 16 * (f->y_stride - mb_cols);
+    mb_uv_offset += mb_uv_height * f->uv_stride - mb_uv_width * mb_cols;
+  }
+
+  // Restore input state
+  for (i = 0; i < MAX_MB_PLANE; i++)
+    mbd->plane[i].pre[0].buf = input_buffer[i];
+}
+
+// Apply buffer limits and context specific adjustments to arnr filter.
+static void adjust_arnr_filter(VP9_COMP *cpi,
+                               int distance, int group_boost,
+                               int *arnr_frames, int *arnr_strength) {
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const int frames_after_arf =
+      vp9_lookahead_depth(cpi->lookahead) - distance - 1;
+  int frames_fwd = (cpi->oxcf.arnr_max_frames - 1) >> 1;
+  int frames_bwd;
+  int q, frames, strength;
+
+  // Define the forward and backwards filter limits for this arnr group.
+  if (frames_fwd > frames_after_arf)
+    frames_fwd = frames_after_arf;
+  if (frames_fwd > distance)
+    frames_fwd = distance;
+
+  frames_bwd = frames_fwd;
+
+  // For even length filter there is one more frame backward
+  // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff.
+  if (frames_bwd < distance)
+    frames_bwd += (oxcf->arnr_max_frames + 1) & 0x1;
+
+  // Set the baseline active filter size.
+  frames = frames_bwd + 1 + frames_fwd;
+
+  // Adjust the strength based on active max q.
+  if (cpi->common.current_video_frame > 1)
+    q = ((int)vp9_convert_qindex_to_q(
+        cpi->rc.avg_frame_qindex[INTER_FRAME], cpi->common.bit_depth));
+  else
+    q = ((int)vp9_convert_qindex_to_q(
+        cpi->rc.avg_frame_qindex[KEY_FRAME], cpi->common.bit_depth));
+  if (q > 16) {
+    strength = oxcf->arnr_strength;
+  } else {
+    strength = oxcf->arnr_strength - ((16 - q) / 2);
+    if (strength < 0)
+      strength = 0;
+  }
+
+  // Adjust number of frames in filter and strength based on gf boost level.
+  if (frames > group_boost / 150) {
+    frames = group_boost / 150;
+    frames += !(frames & 1);
+  }
+
+  if (strength > group_boost / 300) {
+    strength = group_boost / 300;
+  }
+
+  // Adjustments for second level arf in multi arf case.
+  if (cpi->oxcf.pass == 2 && cpi->multi_arf_allowed) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    if (gf_group->rf_level[gf_group->index] != GF_ARF_STD) {
+      strength >>= 1;
+    }
+  }
+
+  *arnr_frames = frames;
+  *arnr_strength = strength;
+}
+
+void vp9_temporal_filter(VP9_COMP *cpi, int distance) {
+  VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+  int frame;
+  int frames_to_blur;
+  int start_frame;
+  int strength;
+  int frames_to_blur_backward;
+  int frames_to_blur_forward;
+  struct scale_factors sf;
+  YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS] = {NULL};
+
+  // Apply context specific adjustments to the arnr filter parameters.
+  adjust_arnr_filter(cpi, distance, rc->gfu_boost, &frames_to_blur, &strength);
+  frames_to_blur_backward = (frames_to_blur / 2);
+  frames_to_blur_forward = ((frames_to_blur - 1) / 2);
+  start_frame = distance + frames_to_blur_forward;
+
+  // Setup frame pointers, NULL indicates frame not included in filter.
+  for (frame = 0; frame < frames_to_blur; ++frame) {
+    const int which_buffer = start_frame - frame;
+    struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead,
+                                                     which_buffer);
+    frames[frames_to_blur - 1 - frame] = &buf->img;
+  }
+
+  if (frames_to_blur > 0) {
+    // Setup scaling factors. Scaling on each of the arnr frames is not
+    // supported.
+    if (cpi->use_svc) {
+      // In spatial svc the scaling factors might be less then 1/2.
+      // So we will use non-normative scaling.
+      int frame_used = 0;
+#if CONFIG_VP9_HIGHBITDEPTH
+      vp9_setup_scale_factors_for_frame(
+          &sf,
+          get_frame_new_buffer(cm)->y_crop_width,
+          get_frame_new_buffer(cm)->y_crop_height,
+          get_frame_new_buffer(cm)->y_crop_width,
+          get_frame_new_buffer(cm)->y_crop_height,
+          cm->use_highbitdepth);
+#else
+      vp9_setup_scale_factors_for_frame(
+          &sf,
+          get_frame_new_buffer(cm)->y_crop_width,
+          get_frame_new_buffer(cm)->y_crop_height,
+          get_frame_new_buffer(cm)->y_crop_width,
+          get_frame_new_buffer(cm)->y_crop_height);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+      for (frame = 0; frame < frames_to_blur; ++frame) {
+        if (cm->mi_cols * MI_SIZE != frames[frame]->y_width ||
+            cm->mi_rows * MI_SIZE != frames[frame]->y_height) {
+          if (vpx_realloc_frame_buffer(&cpi->svc.scaled_frames[frame_used],
+                                       cm->width, cm->height,
+                                       cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                       cm->use_highbitdepth,
+#endif
+                                       VP9_ENC_BORDER_IN_PIXELS,
+                                       cm->byte_alignment,
+                                       NULL, NULL, NULL)) {
+            vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                               "Failed to reallocate alt_ref_buffer");
+          }
+          frames[frame] = vp9_scale_if_required(
+              cm, frames[frame], &cpi->svc.scaled_frames[frame_used], 0);
+          ++frame_used;
+        }
+      }
+      cm->mi = cm->mip + cm->mi_stride + 1;
+      xd->mi = cm->mi_grid_visible;
+      xd->mi[0] = cm->mi;
+    } else {
+      // ARF is produced at the native frame size and resized when coded.
+#if CONFIG_VP9_HIGHBITDEPTH
+      vp9_setup_scale_factors_for_frame(&sf,
+                                        frames[0]->y_crop_width,
+                                        frames[0]->y_crop_height,
+                                        frames[0]->y_crop_width,
+                                        frames[0]->y_crop_height,
+                                        cm->use_highbitdepth);
+#else
+      vp9_setup_scale_factors_for_frame(&sf,
+                                        frames[0]->y_crop_width,
+                                        frames[0]->y_crop_height,
+                                        frames[0]->y_crop_width,
+                                        frames[0]->y_crop_height);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    }
+  }
+
+  temporal_filter_iterate_c(cpi, frames, frames_to_blur,
+                            frames_to_blur_backward, strength, &sf);
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_temporal_filter.h b/libvpx/libvpx/vp9/encoder/vp9_temporal_filter.h
new file mode 100644
index 0000000..f537b88
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_temporal_filter.h
@@ -0,0 +1,25 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_TEMPORAL_FILTER_H_
+#define VP9_ENCODER_VP9_TEMPORAL_FILTER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_temporal_filter_init(void);
+void vp9_temporal_filter(VP9_COMP *cpi, int distance);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_TEMPORAL_FILTER_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_tokenize.c b/libvpx/libvpx/vp9/encoder/vp9_tokenize.c
new file mode 100644
index 0000000..edec755
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_tokenize.c
@@ -0,0 +1,500 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_scan.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_tokenize.h"
+
+static const TOKENVALUE dct_cat_lt_10_value_tokens[] = {
+  {9, 63}, {9, 61}, {9, 59}, {9, 57}, {9, 55}, {9, 53}, {9, 51}, {9, 49},
+  {9, 47}, {9, 45}, {9, 43}, {9, 41}, {9, 39}, {9, 37}, {9, 35}, {9, 33},
+  {9, 31}, {9, 29}, {9, 27}, {9, 25}, {9, 23}, {9, 21}, {9, 19}, {9, 17},
+  {9, 15}, {9, 13}, {9, 11}, {9, 9}, {9, 7}, {9, 5}, {9, 3}, {9, 1},
+  {8, 31}, {8, 29}, {8, 27}, {8, 25}, {8, 23}, {8, 21},
+  {8, 19}, {8, 17}, {8, 15}, {8, 13}, {8, 11}, {8, 9},
+  {8, 7}, {8, 5}, {8, 3}, {8, 1},
+  {7, 15}, {7, 13}, {7, 11}, {7, 9}, {7, 7}, {7, 5}, {7, 3}, {7, 1},
+  {6, 7}, {6, 5}, {6, 3}, {6, 1}, {5, 3}, {5, 1},
+  {4, 1}, {3, 1}, {2, 1}, {1, 1}, {0, 0},
+  {1, 0},  {2, 0}, {3, 0}, {4, 0},
+  {5, 0}, {5, 2}, {6, 0}, {6, 2}, {6, 4}, {6, 6},
+  {7, 0}, {7, 2}, {7, 4}, {7, 6}, {7, 8}, {7, 10}, {7, 12}, {7, 14},
+  {8, 0}, {8, 2}, {8, 4}, {8, 6}, {8, 8}, {8, 10}, {8, 12},
+  {8, 14}, {8, 16}, {8, 18}, {8, 20}, {8, 22}, {8, 24},
+  {8, 26}, {8, 28}, {8, 30}, {9, 0}, {9, 2},
+  {9, 4}, {9, 6}, {9, 8}, {9, 10}, {9, 12}, {9, 14}, {9, 16},
+  {9, 18}, {9, 20}, {9, 22}, {9, 24}, {9, 26}, {9, 28},
+  {9, 30}, {9, 32}, {9, 34}, {9, 36}, {9, 38}, {9, 40},
+  {9, 42}, {9, 44}, {9, 46}, {9, 48}, {9, 50}, {9, 52},
+  {9, 54}, {9, 56}, {9, 58}, {9, 60}, {9, 62}
+};
+const TOKENVALUE *vp9_dct_cat_lt_10_value_tokens = dct_cat_lt_10_value_tokens +
+    (sizeof(dct_cat_lt_10_value_tokens) / sizeof(*dct_cat_lt_10_value_tokens))
+    / 2;
+// The corresponding costs of the extrabits for the tokens in the above table
+// are stored in the table below. The values are obtained from looking up the
+// entry for the specified extrabits in the table corresponding to the token
+// (as defined in cost element vp9_extra_bits)
+// e.g. {9, 63} maps to cat5_cost[63 >> 1], {1, 1} maps to sign_cost[1 >> 1]
+static const int dct_cat_lt_10_value_cost[] = {
+  3773, 3750, 3704, 3681, 3623, 3600, 3554, 3531,
+  3432, 3409, 3363, 3340, 3282, 3259, 3213, 3190,
+  3136, 3113, 3067, 3044, 2986, 2963, 2917, 2894,
+  2795, 2772, 2726, 2703, 2645, 2622, 2576, 2553,
+  3197, 3116, 3058, 2977, 2881, 2800,
+  2742, 2661, 2615, 2534, 2476, 2395,
+  2299, 2218, 2160, 2079,
+  2566, 2427, 2334, 2195, 2023, 1884, 1791, 1652,
+  1893, 1696, 1453, 1256, 1229, 864,
+  512, 512, 512, 512, 0,
+  512, 512, 512, 512,
+  864, 1229, 1256, 1453, 1696, 1893,
+  1652, 1791, 1884, 2023, 2195, 2334, 2427, 2566,
+  2079, 2160, 2218, 2299, 2395, 2476, 2534, 2615,
+  2661, 2742, 2800, 2881, 2977, 3058, 3116, 3197,
+  2553, 2576, 2622, 2645, 2703, 2726, 2772, 2795,
+  2894, 2917, 2963, 2986, 3044, 3067, 3113, 3136,
+  3190, 3213, 3259, 3282, 3340, 3363, 3409, 3432,
+  3531, 3554, 3600, 3623, 3681, 3704, 3750, 3773,
+};
+const int *vp9_dct_cat_lt_10_value_cost = dct_cat_lt_10_value_cost +
+    (sizeof(dct_cat_lt_10_value_cost) / sizeof(*dct_cat_lt_10_value_cost))
+    / 2;
+
+// Array indices are identical to previously-existing CONTEXT_NODE indices
+const vpx_tree_index vp9_coef_tree[TREE_SIZE(ENTROPY_TOKENS)] = {
+  -EOB_TOKEN, 2,                       // 0  = EOB
+  -ZERO_TOKEN, 4,                      // 1  = ZERO
+  -ONE_TOKEN, 6,                       // 2  = ONE
+  8, 12,                               // 3  = LOW_VAL
+  -TWO_TOKEN, 10,                      // 4  = TWO
+  -THREE_TOKEN, -FOUR_TOKEN,           // 5  = THREE
+  14, 16,                              // 6  = HIGH_LOW
+  -CATEGORY1_TOKEN, -CATEGORY2_TOKEN,  // 7  = CAT_ONE
+  18, 20,                              // 8  = CAT_THREEFOUR
+  -CATEGORY3_TOKEN, -CATEGORY4_TOKEN,  // 9  = CAT_THREE
+  -CATEGORY5_TOKEN, -CATEGORY6_TOKEN   // 10 = CAT_FIVE
+};
+
+static const int16_t zero_cost[] = {0};
+static const int16_t sign_cost[1] = {512};
+static const int16_t cat1_cost[1 << 1] = {864, 1229};
+static const int16_t cat2_cost[1 << 2] = {1256, 1453, 1696, 1893};
+static const int16_t cat3_cost[1 << 3] = {1652, 1791, 1884, 2023,
+                                          2195, 2334, 2427, 2566};
+static const int16_t cat4_cost[1 << 4] = {2079, 2160, 2218, 2299, 2395, 2476,
+                                          2534, 2615, 2661, 2742, 2800, 2881,
+                                          2977, 3058, 3116, 3197};
+static const int16_t cat5_cost[1 << 5] = {
+    2553, 2576, 2622, 2645, 2703, 2726, 2772, 2795, 2894, 2917, 2963,
+    2986, 3044, 3067, 3113, 3136, 3190, 3213, 3259, 3282, 3340, 3363,
+    3409, 3432, 3531, 3554, 3600, 3623, 3681, 3704, 3750, 3773};
+const int16_t vp9_cat6_low_cost[256] = {
+    3378, 3390, 3401, 3413, 3435, 3447, 3458, 3470, 3517, 3529, 3540, 3552,
+    3574, 3586, 3597, 3609, 3671, 3683, 3694, 3706, 3728, 3740, 3751, 3763,
+    3810, 3822, 3833, 3845, 3867, 3879, 3890, 3902, 3973, 3985, 3996, 4008,
+    4030, 4042, 4053, 4065, 4112, 4124, 4135, 4147, 4169, 4181, 4192, 4204,
+    4266, 4278, 4289, 4301, 4323, 4335, 4346, 4358, 4405, 4417, 4428, 4440,
+    4462, 4474, 4485, 4497, 4253, 4265, 4276, 4288, 4310, 4322, 4333, 4345,
+    4392, 4404, 4415, 4427, 4449, 4461, 4472, 4484, 4546, 4558, 4569, 4581,
+    4603, 4615, 4626, 4638, 4685, 4697, 4708, 4720, 4742, 4754, 4765, 4777,
+    4848, 4860, 4871, 4883, 4905, 4917, 4928, 4940, 4987, 4999, 5010, 5022,
+    5044, 5056, 5067, 5079, 5141, 5153, 5164, 5176, 5198, 5210, 5221, 5233,
+    5280, 5292, 5303, 5315, 5337, 5349, 5360, 5372, 4988, 5000, 5011, 5023,
+    5045, 5057, 5068, 5080, 5127, 5139, 5150, 5162, 5184, 5196, 5207, 5219,
+    5281, 5293, 5304, 5316, 5338, 5350, 5361, 5373, 5420, 5432, 5443, 5455,
+    5477, 5489, 5500, 5512, 5583, 5595, 5606, 5618, 5640, 5652, 5663, 5675,
+    5722, 5734, 5745, 5757, 5779, 5791, 5802, 5814, 5876, 5888, 5899, 5911,
+    5933, 5945, 5956, 5968, 6015, 6027, 6038, 6050, 6072, 6084, 6095, 6107,
+    5863, 5875, 5886, 5898, 5920, 5932, 5943, 5955, 6002, 6014, 6025, 6037,
+    6059, 6071, 6082, 6094, 6156, 6168, 6179, 6191, 6213, 6225, 6236, 6248,
+    6295, 6307, 6318, 6330, 6352, 6364, 6375, 6387, 6458, 6470, 6481, 6493,
+    6515, 6527, 6538, 6550, 6597, 6609, 6620, 6632, 6654, 6666, 6677, 6689,
+    6751, 6763, 6774, 6786, 6808, 6820, 6831, 6843, 6890, 6902, 6913, 6925,
+    6947, 6959, 6970, 6982};
+const int vp9_cat6_high_cost[64] = {
+    88,    2251,  2727,  4890,  3148,  5311,  5787,  7950,  3666,  5829,  6305,
+    8468,  6726,  8889,  9365,  11528, 3666,  5829,  6305,  8468,  6726,  8889,
+    9365,  11528, 7244,  9407,  9883,  12046, 10304, 12467, 12943, 15106, 3666,
+    5829,  6305,  8468,  6726,  8889,  9365,  11528, 7244,  9407,  9883,  12046,
+    10304, 12467, 12943, 15106, 7244,  9407,  9883,  12046, 10304, 12467, 12943,
+    15106, 10822, 12985, 13461, 15624, 13882, 16045, 16521, 18684};
+
+#if CONFIG_VP9_HIGHBITDEPTH
+const int vp9_cat6_high10_high_cost[256] = {
+    94,    2257,  2733,  4896,  3154,  5317,  5793,  7956,  3672,  5835,  6311,
+    8474,  6732,  8895,  9371,  11534, 3672,  5835,  6311,  8474,  6732,  8895,
+    9371,  11534, 7250,  9413,  9889,  12052, 10310, 12473, 12949, 15112, 3672,
+    5835,  6311,  8474,  6732,  8895,  9371,  11534, 7250,  9413,  9889,  12052,
+    10310, 12473, 12949, 15112, 7250,  9413,  9889,  12052, 10310, 12473, 12949,
+    15112, 10828, 12991, 13467, 15630, 13888, 16051, 16527, 18690, 4187,  6350,
+    6826,  8989,  7247,  9410,  9886,  12049, 7765,  9928,  10404, 12567, 10825,
+    12988, 13464, 15627, 7765,  9928,  10404, 12567, 10825, 12988, 13464, 15627,
+    11343, 13506, 13982, 16145, 14403, 16566, 17042, 19205, 7765,  9928,  10404,
+    12567, 10825, 12988, 13464, 15627, 11343, 13506, 13982, 16145, 14403, 16566,
+    17042, 19205, 11343, 13506, 13982, 16145, 14403, 16566, 17042, 19205, 14921,
+    17084, 17560, 19723, 17981, 20144, 20620, 22783, 4187,  6350,  6826,  8989,
+    7247,  9410,  9886,  12049, 7765,  9928,  10404, 12567, 10825, 12988, 13464,
+    15627, 7765,  9928,  10404, 12567, 10825, 12988, 13464, 15627, 11343, 13506,
+    13982, 16145, 14403, 16566, 17042, 19205, 7765,  9928,  10404, 12567, 10825,
+    12988, 13464, 15627, 11343, 13506, 13982, 16145, 14403, 16566, 17042, 19205,
+    11343, 13506, 13982, 16145, 14403, 16566, 17042, 19205, 14921, 17084, 17560,
+    19723, 17981, 20144, 20620, 22783, 8280,  10443, 10919, 13082, 11340, 13503,
+    13979, 16142, 11858, 14021, 14497, 16660, 14918, 17081, 17557, 19720, 11858,
+    14021, 14497, 16660, 14918, 17081, 17557, 19720, 15436, 17599, 18075, 20238,
+    18496, 20659, 21135, 23298, 11858, 14021, 14497, 16660, 14918, 17081, 17557,
+    19720, 15436, 17599, 18075, 20238, 18496, 20659, 21135, 23298, 15436, 17599,
+    18075, 20238, 18496, 20659, 21135, 23298, 19014, 21177, 21653, 23816, 22074,
+    24237, 24713, 26876};
+const int vp9_cat6_high12_high_cost[1024] = {
+    100,   2263,  2739,  4902,  3160,  5323,  5799,  7962,  3678,  5841,  6317,
+    8480,  6738,  8901,  9377,  11540, 3678,  5841,  6317,  8480,  6738,  8901,
+    9377,  11540, 7256,  9419,  9895,  12058, 10316, 12479, 12955, 15118, 3678,
+    5841,  6317,  8480,  6738,  8901,  9377,  11540, 7256,  9419,  9895,  12058,
+    10316, 12479, 12955, 15118, 7256,  9419,  9895,  12058, 10316, 12479, 12955,
+    15118, 10834, 12997, 13473, 15636, 13894, 16057, 16533, 18696, 4193,  6356,
+    6832,  8995,  7253,  9416,  9892,  12055, 7771,  9934,  10410, 12573, 10831,
+    12994, 13470, 15633, 7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633,
+    11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 7771,  9934,  10410,
+    12573, 10831, 12994, 13470, 15633, 11349, 13512, 13988, 16151, 14409, 16572,
+    17048, 19211, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 14927,
+    17090, 17566, 19729, 17987, 20150, 20626, 22789, 4193,  6356,  6832,  8995,
+    7253,  9416,  9892,  12055, 7771,  9934,  10410, 12573, 10831, 12994, 13470,
+    15633, 7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633, 11349, 13512,
+    13988, 16151, 14409, 16572, 17048, 19211, 7771,  9934,  10410, 12573, 10831,
+    12994, 13470, 15633, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211,
+    11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 14927, 17090, 17566,
+    19729, 17987, 20150, 20626, 22789, 8286,  10449, 10925, 13088, 11346, 13509,
+    13985, 16148, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 11864,
+    14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244,
+    18502, 20665, 21141, 23304, 11864, 14027, 14503, 16666, 14924, 17087, 17563,
+    19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 15442, 17605,
+    18081, 20244, 18502, 20665, 21141, 23304, 19020, 21183, 21659, 23822, 22080,
+    24243, 24719, 26882, 4193,  6356,  6832,  8995,  7253,  9416,  9892,  12055,
+    7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633, 7771,  9934,  10410,
+    12573, 10831, 12994, 13470, 15633, 11349, 13512, 13988, 16151, 14409, 16572,
+    17048, 19211, 7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633, 11349,
+    13512, 13988, 16151, 14409, 16572, 17048, 19211, 11349, 13512, 13988, 16151,
+    14409, 16572, 17048, 19211, 14927, 17090, 17566, 19729, 17987, 20150, 20626,
+    22789, 8286,  10449, 10925, 13088, 11346, 13509, 13985, 16148, 11864, 14027,
+    14503, 16666, 14924, 17087, 17563, 19726, 11864, 14027, 14503, 16666, 14924,
+    17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304,
+    11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081,
+    20244, 18502, 20665, 21141, 23304, 15442, 17605, 18081, 20244, 18502, 20665,
+    21141, 23304, 19020, 21183, 21659, 23822, 22080, 24243, 24719, 26882, 8286,
+    10449, 10925, 13088, 11346, 13509, 13985, 16148, 11864, 14027, 14503, 16666,
+    14924, 17087, 17563, 19726, 11864, 14027, 14503, 16666, 14924, 17087, 17563,
+    19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 11864, 14027,
+    14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502,
+    20665, 21141, 23304, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304,
+    19020, 21183, 21659, 23822, 22080, 24243, 24719, 26882, 12379, 14542, 15018,
+    17181, 15439, 17602, 18078, 20241, 15957, 18120, 18596, 20759, 19017, 21180,
+    21656, 23819, 15957, 18120, 18596, 20759, 19017, 21180, 21656, 23819, 19535,
+    21698, 22174, 24337, 22595, 24758, 25234, 27397, 15957, 18120, 18596, 20759,
+    19017, 21180, 21656, 23819, 19535, 21698, 22174, 24337, 22595, 24758, 25234,
+    27397, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 23113, 25276,
+    25752, 27915, 26173, 28336, 28812, 30975, 4193,  6356,  6832,  8995,  7253,
+    9416,  9892,  12055, 7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633,
+    7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633, 11349, 13512, 13988,
+    16151, 14409, 16572, 17048, 19211, 7771,  9934,  10410, 12573, 10831, 12994,
+    13470, 15633, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 11349,
+    13512, 13988, 16151, 14409, 16572, 17048, 19211, 14927, 17090, 17566, 19729,
+    17987, 20150, 20626, 22789, 8286,  10449, 10925, 13088, 11346, 13509, 13985,
+    16148, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 11864, 14027,
+    14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502,
+    20665, 21141, 23304, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726,
+    15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 15442, 17605, 18081,
+    20244, 18502, 20665, 21141, 23304, 19020, 21183, 21659, 23822, 22080, 24243,
+    24719, 26882, 8286,  10449, 10925, 13088, 11346, 13509, 13985, 16148, 11864,
+    14027, 14503, 16666, 14924, 17087, 17563, 19726, 11864, 14027, 14503, 16666,
+    14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141,
+    23304, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605,
+    18081, 20244, 18502, 20665, 21141, 23304, 15442, 17605, 18081, 20244, 18502,
+    20665, 21141, 23304, 19020, 21183, 21659, 23822, 22080, 24243, 24719, 26882,
+    12379, 14542, 15018, 17181, 15439, 17602, 18078, 20241, 15957, 18120, 18596,
+    20759, 19017, 21180, 21656, 23819, 15957, 18120, 18596, 20759, 19017, 21180,
+    21656, 23819, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 15957,
+    18120, 18596, 20759, 19017, 21180, 21656, 23819, 19535, 21698, 22174, 24337,
+    22595, 24758, 25234, 27397, 19535, 21698, 22174, 24337, 22595, 24758, 25234,
+    27397, 23113, 25276, 25752, 27915, 26173, 28336, 28812, 30975, 8286,  10449,
+    10925, 13088, 11346, 13509, 13985, 16148, 11864, 14027, 14503, 16666, 14924,
+    17087, 17563, 19726, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726,
+    15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 11864, 14027, 14503,
+    16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665,
+    21141, 23304, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 19020,
+    21183, 21659, 23822, 22080, 24243, 24719, 26882, 12379, 14542, 15018, 17181,
+    15439, 17602, 18078, 20241, 15957, 18120, 18596, 20759, 19017, 21180, 21656,
+    23819, 15957, 18120, 18596, 20759, 19017, 21180, 21656, 23819, 19535, 21698,
+    22174, 24337, 22595, 24758, 25234, 27397, 15957, 18120, 18596, 20759, 19017,
+    21180, 21656, 23819, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397,
+    19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 23113, 25276, 25752,
+    27915, 26173, 28336, 28812, 30975, 12379, 14542, 15018, 17181, 15439, 17602,
+    18078, 20241, 15957, 18120, 18596, 20759, 19017, 21180, 21656, 23819, 15957,
+    18120, 18596, 20759, 19017, 21180, 21656, 23819, 19535, 21698, 22174, 24337,
+    22595, 24758, 25234, 27397, 15957, 18120, 18596, 20759, 19017, 21180, 21656,
+    23819, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 19535, 21698,
+    22174, 24337, 22595, 24758, 25234, 27397, 23113, 25276, 25752, 27915, 26173,
+    28336, 28812, 30975, 16472, 18635, 19111, 21274, 19532, 21695, 22171, 24334,
+    20050, 22213, 22689, 24852, 23110, 25273, 25749, 27912, 20050, 22213, 22689,
+    24852, 23110, 25273, 25749, 27912, 23628, 25791, 26267, 28430, 26688, 28851,
+    29327, 31490, 20050, 22213, 22689, 24852, 23110, 25273, 25749, 27912, 23628,
+    25791, 26267, 28430, 26688, 28851, 29327, 31490, 23628, 25791, 26267, 28430,
+    26688, 28851, 29327, 31490, 27206, 29369, 29845, 32008, 30266, 32429, 32905,
+    35068};
+#endif
+
+const vp9_extra_bit vp9_extra_bits[ENTROPY_TOKENS] = {
+  {0, 0, 0, zero_cost},                          // ZERO_TOKEN
+  {0, 0, 1, sign_cost},                          // ONE_TOKEN
+  {0, 0, 2, sign_cost},                          // TWO_TOKEN
+  {0, 0, 3, sign_cost},                          // THREE_TOKEN
+  {0, 0, 4, sign_cost},                          // FOUR_TOKEN
+  {vp9_cat1_prob, 1,  CAT1_MIN_VAL, cat1_cost},  // CATEGORY1_TOKEN
+  {vp9_cat2_prob, 2,  CAT2_MIN_VAL, cat2_cost},  // CATEGORY2_TOKEN
+  {vp9_cat3_prob, 3,  CAT3_MIN_VAL, cat3_cost},  // CATEGORY3_TOKEN
+  {vp9_cat4_prob, 4,  CAT4_MIN_VAL, cat4_cost},  // CATEGORY4_TOKEN
+  {vp9_cat5_prob, 5,  CAT5_MIN_VAL, cat5_cost},  // CATEGORY5_TOKEN
+  {vp9_cat6_prob, 14, CAT6_MIN_VAL, 0},          // CATEGORY6_TOKEN
+  {0, 0, 0, zero_cost}                           // EOB_TOKEN
+};
+
+#if CONFIG_VP9_HIGHBITDEPTH
+const vp9_extra_bit vp9_extra_bits_high10[ENTROPY_TOKENS] = {
+  {0, 0, 0, zero_cost},                             // ZERO
+  {0, 0, 1, sign_cost},                             // ONE
+  {0, 0, 2, sign_cost},                             // TWO
+  {0, 0, 3, sign_cost},                             // THREE
+  {0, 0, 4, sign_cost},                             // FOUR
+  {vp9_cat1_prob, 1,  CAT1_MIN_VAL, cat1_cost},     // CAT1
+  {vp9_cat2_prob, 2,  CAT2_MIN_VAL, cat2_cost},     // CAT2
+  {vp9_cat3_prob, 3,  CAT3_MIN_VAL, cat3_cost},     // CAT3
+  {vp9_cat4_prob, 4,  CAT4_MIN_VAL, cat4_cost},     // CAT4
+  {vp9_cat5_prob, 5,  CAT5_MIN_VAL, cat5_cost},     // CAT5
+  {vp9_cat6_prob_high12 + 2, 16, CAT6_MIN_VAL, 0},  // CAT6
+  {0, 0, 0, zero_cost}                              // EOB
+};
+const vp9_extra_bit vp9_extra_bits_high12[ENTROPY_TOKENS] = {
+  {0, 0, 0, zero_cost},                          // ZERO
+  {0, 0, 1, sign_cost},                          // ONE
+  {0, 0, 2, sign_cost},                          // TWO
+  {0, 0, 3, sign_cost},                          // THREE
+  {0, 0, 4, sign_cost},                          // FOUR
+  {vp9_cat1_prob, 1,  CAT1_MIN_VAL, cat1_cost},  // CAT1
+  {vp9_cat2_prob, 2,  CAT2_MIN_VAL, cat2_cost},  // CAT2
+  {vp9_cat3_prob, 3,  CAT3_MIN_VAL, cat3_cost},  // CAT3
+  {vp9_cat4_prob, 4,  CAT4_MIN_VAL, cat4_cost},  // CAT4
+  {vp9_cat5_prob, 5,  CAT5_MIN_VAL, cat5_cost},  // CAT5
+  {vp9_cat6_prob_high12, 18, CAT6_MIN_VAL, 0},   // CAT6
+  {0, 0, 0, zero_cost}                           // EOB
+};
+#endif
+
+const struct vp9_token vp9_coef_encodings[ENTROPY_TOKENS] = {
+  {2, 2}, {6, 3}, {28, 5}, {58, 6}, {59, 6}, {60, 6}, {61, 6}, {124, 7},
+  {125, 7}, {126, 7}, {127, 7}, {0, 1}
+};
+
+
+struct tokenize_b_args {
+  VP9_COMP *cpi;
+  ThreadData *td;
+  TOKENEXTRA **tp;
+};
+
+static void set_entropy_context_b(int plane, int block, BLOCK_SIZE plane_bsize,
+                                  TX_SIZE tx_size, void *arg) {
+  struct tokenize_b_args* const args = arg;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  int aoff, loff;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &aoff, &loff);
+  vp9_set_contexts(xd, pd, plane_bsize, tx_size, p->eobs[block] > 0,
+                   aoff, loff);
+}
+
+static INLINE void add_token(TOKENEXTRA **t, const vpx_prob *context_tree,
+                             int16_t token, EXTRABIT extra,
+                             unsigned int *counts) {
+  (*t)->context_tree = context_tree;
+  (*t)->token = token;
+  (*t)->extra = extra;
+  (*t)++;
+  ++counts[token];
+}
+
+static INLINE void add_token_no_extra(TOKENEXTRA **t,
+                                      const vpx_prob *context_tree,
+                                      int16_t token,
+                                      unsigned int *counts) {
+  (*t)->context_tree = context_tree;
+  (*t)->token = token;
+  (*t)++;
+  ++counts[token];
+}
+
+static void tokenize_b(int plane, int block, BLOCK_SIZE plane_bsize,
+                       TX_SIZE tx_size, void *arg) {
+  struct tokenize_b_args* const args = arg;
+  VP9_COMP *cpi = args->cpi;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  TOKENEXTRA **tp = args->tp;
+  uint8_t token_cache[32 * 32];
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  MODE_INFO *mi = xd->mi[0];
+  int pt; /* near block/prev token context index */
+  int c;
+  TOKENEXTRA *t = *tp;        /* store tokens starting here */
+  int eob = p->eobs[block];
+  const PLANE_TYPE type = get_plane_type(plane);
+  const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  const int16_t *scan, *nb;
+  const scan_order *so;
+  const int ref = is_inter_block(mi);
+  unsigned int (*const counts)[COEFF_CONTEXTS][ENTROPY_TOKENS] =
+      td->rd_counts.coef_counts[tx_size][type][ref];
+  vpx_prob (*const coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
+      cpi->common.fc->coef_probs[tx_size][type][ref];
+  unsigned int (*const eob_branch)[COEFF_CONTEXTS] =
+      td->counts->eob_branch[tx_size][type][ref];
+  const uint8_t *const band = get_band_translate(tx_size);
+  const int tx_eob = 16 << (tx_size << 1);
+  int16_t token;
+  EXTRABIT extra;
+  int aoff, loff;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &aoff, &loff);
+
+  pt = get_entropy_context(tx_size, pd->above_context + aoff,
+                           pd->left_context + loff);
+  so = get_scan(xd, tx_size, type, block);
+  scan = so->scan;
+  nb = so->neighbors;
+  c = 0;
+
+  while (c < eob) {
+    int v = 0;
+    v = qcoeff[scan[c]];
+    ++eob_branch[band[c]][pt];
+
+    while (!v) {
+      add_token_no_extra(&t, coef_probs[band[c]][pt], ZERO_TOKEN,
+                         counts[band[c]][pt]);
+
+      token_cache[scan[c]] = 0;
+      ++c;
+      pt = get_coef_context(nb, token_cache, c);
+      v = qcoeff[scan[c]];
+    }
+
+    vp9_get_token_extra(v, &token, &extra);
+
+    add_token(&t, coef_probs[band[c]][pt], token, extra,
+              counts[band[c]][pt]);
+
+    token_cache[scan[c]] = vp9_pt_energy_class[token];
+    ++c;
+    pt = get_coef_context(nb, token_cache, c);
+  }
+  if (c < tx_eob) {
+    ++eob_branch[band[c]][pt];
+    add_token_no_extra(&t, coef_probs[band[c]][pt], EOB_TOKEN,
+                       counts[band[c]][pt]);
+  }
+
+  *tp = t;
+
+  vp9_set_contexts(xd, pd, plane_bsize, tx_size, c > 0, aoff, loff);
+}
+
+struct is_skippable_args {
+  uint16_t *eobs;
+  int *skippable;
+};
+static void is_skippable(int plane, int block,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                         void *argv) {
+  struct is_skippable_args *args = argv;
+  (void)plane;
+  (void)plane_bsize;
+  (void)tx_size;
+  args->skippable[0] &= (!args->eobs[block]);
+}
+
+// TODO(yaowu): rewrite and optimize this function to remove the usage of
+//              vp9_foreach_transform_block() and simplify is_skippable().
+int vp9_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
+  int result = 1;
+  struct is_skippable_args args = {x->plane[plane].eobs, &result};
+  vp9_foreach_transformed_block_in_plane(&x->e_mbd, bsize, plane, is_skippable,
+                                         &args);
+  return result;
+}
+
+static void has_high_freq_coeff(int plane, int block,
+                                BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                                void *argv) {
+  struct is_skippable_args *args = argv;
+  int eobs = (tx_size == TX_4X4) ? 3 : 10;
+  (void) plane;
+  (void) plane_bsize;
+
+  *(args->skippable) |= (args->eobs[block] > eobs);
+}
+
+int vp9_has_high_freq_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
+  int result = 0;
+  struct is_skippable_args args = {x->plane[plane].eobs, &result};
+  vp9_foreach_transformed_block_in_plane(&x->e_mbd, bsize, plane,
+                                         has_high_freq_coeff, &args);
+  return result;
+}
+
+void vp9_tokenize_sb(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
+                     int dry_run, int seg_skip, BLOCK_SIZE bsize) {
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  const int ctx = vp9_get_skip_context(xd);
+  struct tokenize_b_args arg = {cpi, td, t};
+
+  if (seg_skip) {
+    assert(mi->skip);
+  }
+
+  if (mi->skip) {
+    if (!dry_run && !seg_skip)
+      ++td->counts->skip[ctx][1];
+    reset_skip_context(xd, bsize);
+    return;
+  }
+
+  if (!dry_run) {
+    ++td->counts->skip[ctx][0];
+    vp9_foreach_transformed_block(xd, bsize, tokenize_b, &arg);
+  } else {
+    vp9_foreach_transformed_block(xd, bsize, set_entropy_context_b, &arg);
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_tokenize.h b/libvpx/libvpx/vp9/encoder/vp9_tokenize.h
new file mode 100644
index 0000000..1caab2a
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_tokenize.h
@@ -0,0 +1,139 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_TOKENIZE_H_
+#define VP9_ENCODER_VP9_TOKENIZE_H_
+
+#include "vp9/common/vp9_entropy.h"
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_treewriter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define EOSB_TOKEN 127     // Not signalled, encoder only
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  typedef int32_t EXTRABIT;
+#else
+  typedef int16_t EXTRABIT;
+#endif
+
+
+typedef struct {
+  int16_t token;
+  EXTRABIT extra;
+} TOKENVALUE;
+
+typedef struct {
+  const vpx_prob *context_tree;
+  int16_t token;
+  EXTRABIT extra;
+} TOKENEXTRA;
+
+extern const vpx_tree_index vp9_coef_tree[];
+extern const vpx_tree_index vp9_coef_con_tree[];
+extern const struct vp9_token vp9_coef_encodings[];
+
+int vp9_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+int vp9_has_high_freq_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+
+struct VP9_COMP;
+struct ThreadData;
+
+void vp9_tokenize_sb(struct VP9_COMP *cpi, struct ThreadData *td,
+                     TOKENEXTRA **t, int dry_run, int seg_skip,
+                     BLOCK_SIZE bsize);
+
+typedef struct {
+  const vpx_prob *prob;
+  int len;
+  int base_val;
+  const int16_t *cost;
+} vp9_extra_bit;
+
+// indexed by token value
+extern const vp9_extra_bit vp9_extra_bits[ENTROPY_TOKENS];
+#if CONFIG_VP9_HIGHBITDEPTH
+extern const vp9_extra_bit vp9_extra_bits_high10[ENTROPY_TOKENS];
+extern const vp9_extra_bit vp9_extra_bits_high12[ENTROPY_TOKENS];
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+extern const int16_t *vp9_dct_value_cost_ptr;
+/* TODO: The Token field should be broken out into a separate char array to
+ *  improve cache locality, since it's needed for costing when the rest of the
+ *  fields are not.
+ */
+extern const TOKENVALUE *vp9_dct_value_tokens_ptr;
+extern const TOKENVALUE *vp9_dct_cat_lt_10_value_tokens;
+extern const int *vp9_dct_cat_lt_10_value_cost;
+extern const int16_t vp9_cat6_low_cost[256];
+extern const int vp9_cat6_high_cost[64];
+extern const int vp9_cat6_high10_high_cost[256];
+extern const int vp9_cat6_high12_high_cost[1024];
+static INLINE int vp9_get_cost(int16_t token, EXTRABIT extrabits,
+                               const int *cat6_high_table) {
+  if (token != CATEGORY6_TOKEN)
+    return vp9_extra_bits[token].cost[extrabits >> 1];
+  return vp9_cat6_low_cost[(extrabits >> 1) & 0xff]
+      + cat6_high_table[extrabits >> 9];
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE const int* vp9_get_high_cost_table(int bit_depth) {
+  return bit_depth == 8 ? vp9_cat6_high_cost
+      : (bit_depth == 10 ? vp9_cat6_high10_high_cost :
+         vp9_cat6_high12_high_cost);
+}
+#else
+static INLINE const int* vp9_get_high_cost_table(int bit_depth) {
+  (void) bit_depth;
+  return vp9_cat6_high_cost;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static INLINE void vp9_get_token_extra(int v, int16_t *token, EXTRABIT *extra) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL) {
+    *token = CATEGORY6_TOKEN;
+    if (v >= CAT6_MIN_VAL)
+      *extra = 2 * v - 2 * CAT6_MIN_VAL;
+    else
+      *extra = -2 * v - 2 * CAT6_MIN_VAL + 1;
+    return;
+  }
+  *token = vp9_dct_cat_lt_10_value_tokens[v].token;
+  *extra = vp9_dct_cat_lt_10_value_tokens[v].extra;
+}
+static INLINE int16_t vp9_get_token(int v) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL)
+    return 10;
+  return vp9_dct_cat_lt_10_value_tokens[v].token;
+}
+
+static INLINE int vp9_get_token_cost(int v, int16_t *token,
+                                     const int *cat6_high_table) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL) {
+    EXTRABIT extrabits;
+    *token = CATEGORY6_TOKEN;
+    extrabits = abs(v) - CAT6_MIN_VAL;
+    return vp9_cat6_low_cost[extrabits & 0xff] +
+           cat6_high_table[extrabits >> 8];
+  }
+  *token = vp9_dct_cat_lt_10_value_tokens[v].token;
+  return vp9_dct_cat_lt_10_value_cost[v];
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_TOKENIZE_H_
diff --git a/libvpx/libvpx/vp9/encoder/vp9_treewriter.c b/libvpx/libvpx/vp9/encoder/vp9_treewriter.c
new file mode 100644
index 0000000..0fc078e
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_treewriter.c
@@ -0,0 +1,58 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/encoder/vp9_treewriter.h"
+
+static void tree2tok(struct vp9_token *tokens, const vpx_tree_index *tree,
+                     int i, int v, int l) {
+  v += v;
+  ++l;
+
+  do {
+    const vpx_tree_index j = tree[i++];
+    if (j <= 0) {
+      tokens[-j].value = v;
+      tokens[-j].len = l;
+    } else {
+      tree2tok(tokens, tree, j, v, l);
+    }
+  } while (++v & 1);
+}
+
+void vp9_tokens_from_tree(struct vp9_token *tokens,
+                          const vpx_tree_index *tree) {
+  tree2tok(tokens, tree, 0, 0, 0);
+}
+
+static unsigned int convert_distribution(unsigned int i, vpx_tree tree,
+                                         unsigned int branch_ct[][2],
+                                         const unsigned int num_events[]) {
+  unsigned int left, right;
+
+  if (tree[i] <= 0)
+    left = num_events[-tree[i]];
+  else
+    left = convert_distribution(tree[i], tree, branch_ct, num_events);
+
+  if (tree[i + 1] <= 0)
+    right = num_events[-tree[i + 1]];
+  else
+    right = convert_distribution(tree[i + 1], tree, branch_ct, num_events);
+
+  branch_ct[i >> 1][0] = left;
+  branch_ct[i >> 1][1] = right;
+  return left + right;
+}
+
+void vp9_tree_probs_from_distribution(vpx_tree tree,
+                                      unsigned int branch_ct[/* n-1 */][2],
+                                      const unsigned int num_events[/* n */]) {
+  convert_distribution(0, tree, branch_ct, num_events);
+}
diff --git a/libvpx/libvpx/vp9/encoder/vp9_treewriter.h b/libvpx/libvpx/vp9/encoder/vp9_treewriter.h
new file mode 100644
index 0000000..0f89350
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/vp9_treewriter.h
@@ -0,0 +1,51 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_TREEWRITER_H_
+#define VP9_ENCODER_VP9_TREEWRITER_H_
+
+#include "vpx_dsp/bitwriter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_tree_probs_from_distribution(vpx_tree tree,
+                                      unsigned int branch_ct[ /* n - 1 */ ][2],
+                                      const unsigned int num_events[ /* n */ ]);
+
+struct vp9_token {
+  int value;
+  int len;
+};
+
+void vp9_tokens_from_tree(struct vp9_token*, const vpx_tree_index *);
+
+static INLINE void vp9_write_tree(vpx_writer *w, const vpx_tree_index *tree,
+                                  const vpx_prob *probs, int bits, int len,
+                                  vpx_tree_index i) {
+  do {
+    const int bit = (bits >> --len) & 1;
+    vpx_write(w, bit, probs[i >> 1]);
+    i = tree[i + bit];
+  } while (len);
+}
+
+static INLINE void vp9_write_token(vpx_writer *w, const vpx_tree_index *tree,
+                                   const vpx_prob *probs,
+                                   const struct vp9_token *token) {
+  vp9_write_tree(w, tree, probs, token->value, token->len, 0);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_TREEWRITER_H_
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_dct_intrin_sse2.c b/libvpx/libvpx/vp9/encoder/x86/vp9_dct_intrin_sse2.c
new file mode 100644
index 0000000..fa37b6f
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_dct_intrin_sse2.c
@@ -0,0 +1,2058 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>  // SSE2
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/txfm_common.h"
+#include "vpx_dsp/x86/fwd_txfm_sse2.h"
+#include "vpx_dsp/x86/txfm_common_sse2.h"
+#include "vpx_ports/mem.h"
+
+static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in,
+                                   int stride) {
+  const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
+  const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
+  __m128i mask;
+
+  in[0] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+  in[1] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+  in[2] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
+  in[3] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
+
+  in[0] = _mm_slli_epi16(in[0], 4);
+  in[1] = _mm_slli_epi16(in[1], 4);
+  in[2] = _mm_slli_epi16(in[2], 4);
+  in[3] = _mm_slli_epi16(in[3], 4);
+
+  mask = _mm_cmpeq_epi16(in[0], k__nonzero_bias_a);
+  in[0] = _mm_add_epi16(in[0], mask);
+  in[0] = _mm_add_epi16(in[0], k__nonzero_bias_b);
+}
+
+static INLINE void write_buffer_4x4(tran_low_t *output, __m128i *res) {
+  const __m128i kOne = _mm_set1_epi16(1);
+  __m128i in01 = _mm_unpacklo_epi64(res[0], res[1]);
+  __m128i in23 = _mm_unpacklo_epi64(res[2], res[3]);
+  __m128i out01 = _mm_add_epi16(in01, kOne);
+  __m128i out23 = _mm_add_epi16(in23, kOne);
+  out01 = _mm_srai_epi16(out01, 2);
+  out23 = _mm_srai_epi16(out23, 2);
+  store_output(&out01, (output + 0 * 8));
+  store_output(&out23, (output + 1 * 8));
+}
+
+static INLINE void transpose_4x4(__m128i *res) {
+  // Combine and transpose
+  // 00 01 02 03 20 21 22 23
+  // 10 11 12 13 30 31 32 33
+  const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
+  const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
+
+  // 00 10 01 11 02 12 03 13
+  // 20 30 21 31 22 32 23 33
+  res[0] = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  res[2] = _mm_unpackhi_epi32(tr0_0, tr0_1);
+
+  // 00 10 20 30 01 11 21 31
+  // 02 12 22 32 03 13 23 33
+  // only use the first 4 16-bit integers
+  res[1] = _mm_unpackhi_epi64(res[0], res[0]);
+  res[3] = _mm_unpackhi_epi64(res[2], res[2]);
+}
+
+static void fdct4_sse2(__m128i *in) {
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i u[4], v[4];
+  u[0]=_mm_unpacklo_epi16(in[0], in[1]);
+  u[1]=_mm_unpacklo_epi16(in[3], in[2]);
+
+  v[0] = _mm_add_epi16(u[0], u[1]);
+  v[1] = _mm_sub_epi16(u[0], u[1]);
+
+  u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);  // 0
+  u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16);  // 2
+  u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24);  // 1
+  u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08);  // 3
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[1]);
+  in[1] = _mm_packs_epi32(u[2], u[3]);
+  transpose_4x4(in);
+}
+
+static void fadst4_sse2(__m128i *in) {
+  const __m128i k__sinpi_p01_p02 = pair_set_epi16(sinpi_1_9, sinpi_2_9);
+  const __m128i k__sinpi_p04_m01 = pair_set_epi16(sinpi_4_9, -sinpi_1_9);
+  const __m128i k__sinpi_p03_p04 = pair_set_epi16(sinpi_3_9, sinpi_4_9);
+  const __m128i k__sinpi_m03_p02 = pair_set_epi16(-sinpi_3_9, sinpi_2_9);
+  const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8];
+  __m128i in7 = _mm_add_epi16(in[0], in[1]);
+
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpacklo_epi16(in[2], in[3]);
+  u[2] = _mm_unpacklo_epi16(in7, kZero);
+  u[3] = _mm_unpacklo_epi16(in[2], kZero);
+  u[4] = _mm_unpacklo_epi16(in[3], kZero);
+
+  v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02);  // s0 + s2
+  v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04);  // s4 + s5
+  v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x1
+  v[3] = _mm_madd_epi16(u[0], k__sinpi_p04_m01);  // s1 - s3
+  v[4] = _mm_madd_epi16(u[1], k__sinpi_m03_p02);  // -s4 + s6
+  v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s4
+  v[6] = _mm_madd_epi16(u[4], k__sinpi_p03_p03);
+
+  u[0] = _mm_add_epi32(v[0], v[1]);
+  u[1] = _mm_sub_epi32(v[2], v[6]);
+  u[2] = _mm_add_epi32(v[3], v[4]);
+  u[3] = _mm_sub_epi32(u[2], u[0]);
+  u[4] = _mm_slli_epi32(v[5], 2);
+  u[5] = _mm_sub_epi32(u[4], v[5]);
+  u[6] = _mm_add_epi32(u[3], u[5]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[2]);
+  in[1] = _mm_packs_epi32(u[1], u[3]);
+  transpose_4x4(in);
+}
+
+void vp9_fht4x4_sse2(const int16_t *input, tran_low_t *output,
+                     int stride, int tx_type) {
+  __m128i in[4];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      vpx_fdct4x4_sse2(input, output, stride);
+      break;
+    case ADST_DCT:
+      load_buffer_4x4(input, in, stride);
+      fadst4_sse2(in);
+      fdct4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case DCT_ADST:
+      load_buffer_4x4(input, in, stride);
+      fdct4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case ADST_ADST:
+      load_buffer_4x4(input, in, stride);
+      fadst4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+   default:
+     assert(0);
+     break;
+  }
+}
+
+void vp9_fdct8x8_quant_sse2(const int16_t *input, int stride,
+                            int16_t* coeff_ptr, intptr_t n_coeffs,
+                            int skip_block, const int16_t* zbin_ptr,
+                            const int16_t* round_ptr, const int16_t* quant_ptr,
+                            const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr,
+                            int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+                            uint16_t* eob_ptr,
+                            const int16_t* scan_ptr,
+                            const int16_t* iscan_ptr) {
+  __m128i zero;
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // Load input
+  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  __m128i *in[8];
+  int index = 0;
+
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)coeff_ptr;
+
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  in[0] = &in0;
+  in[1] = &in1;
+  in[2] = &in2;
+  in[3] = &in3;
+  in[4] = &in4;
+  in[5] = &in5;
+  in[6] = &in6;
+  in[7] = &in7;
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = _mm_add_epi16(in0, in7);
+    const __m128i q1 = _mm_add_epi16(in1, in6);
+    const __m128i q2 = _mm_add_epi16(in2, in5);
+    const __m128i q3 = _mm_add_epi16(in3, in4);
+    const __m128i q4 = _mm_sub_epi16(in3, in4);
+    const __m128i q5 = _mm_sub_epi16(in2, in5);
+    const __m128i q6 = _mm_sub_epi16(in1, in6);
+    const __m128i q7 = _mm_sub_epi16(in0, in7);
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = _mm_add_epi16(q0, q3);
+      const __m128i r1 = _mm_add_epi16(q1, q2);
+      const __m128i r2 = _mm_sub_epi16(q1, q2);
+      const __m128i r3 = _mm_sub_epi16(q0, q3);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+      const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+      const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+      const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+      // dct_const_round_shift
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+      res0 = _mm_packs_epi32(w0, w1);
+      res4 = _mm_packs_epi32(w2, w3);
+      res2 = _mm_packs_epi32(w4, w5);
+      res6 = _mm_packs_epi32(w6, w7);
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
+      const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
+      const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
+      const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
+      const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
+      const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
+      // dct_const_round_shift
+      const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
+      const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
+      const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
+      const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
+      const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
+      const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
+      const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
+      const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
+      // Combine
+      const __m128i r0 = _mm_packs_epi32(s0, s1);
+      const __m128i r1 = _mm_packs_epi32(s2, s3);
+      // Add/subtract
+      const __m128i x0 = _mm_add_epi16(q4, r0);
+      const __m128i x1 = _mm_sub_epi16(q4, r0);
+      const __m128i x2 = _mm_sub_epi16(q7, r1);
+      const __m128i x3 = _mm_add_epi16(q7, r1);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+      const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+      const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+      const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+      // dct_const_round_shift
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+      res1 = _mm_packs_epi32(w0, w1);
+      res7 = _mm_packs_epi32(w2, w3);
+      res5 = _mm_packs_epi32(w4, w5);
+      res3 = _mm_packs_epi32(w6, w7);
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+  }
+
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i*)round_ptr);
+        quant = _mm_load_si128((const __m128i*)quant_ptr);
+        dequant = _mm_load_si128((const __m128i*)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = *in[0];
+        coeff1 = *in[1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // AC only loop
+    index = 2;
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
+        coeff0 = *in[index];
+        coeff1 = *in[index + 1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+      index += 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
+
+// load 8x8 array
+static INLINE void load_buffer_8x8(const int16_t *input, __m128i *in,
+                                   int stride) {
+  in[0]  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  in[1]  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  in[2]  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  in[3]  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  in[4]  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  in[5]  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  in[6]  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  in[7]  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+
+  in[0] = _mm_slli_epi16(in[0], 2);
+  in[1] = _mm_slli_epi16(in[1], 2);
+  in[2] = _mm_slli_epi16(in[2], 2);
+  in[3] = _mm_slli_epi16(in[3], 2);
+  in[4] = _mm_slli_epi16(in[4], 2);
+  in[5] = _mm_slli_epi16(in[5], 2);
+  in[6] = _mm_slli_epi16(in[6], 2);
+  in[7] = _mm_slli_epi16(in[7], 2);
+}
+
+// right shift and rounding
+static INLINE void right_shift_8x8(__m128i *res, const int bit) {
+  __m128i sign0 = _mm_srai_epi16(res[0], 15);
+  __m128i sign1 = _mm_srai_epi16(res[1], 15);
+  __m128i sign2 = _mm_srai_epi16(res[2], 15);
+  __m128i sign3 = _mm_srai_epi16(res[3], 15);
+  __m128i sign4 = _mm_srai_epi16(res[4], 15);
+  __m128i sign5 = _mm_srai_epi16(res[5], 15);
+  __m128i sign6 = _mm_srai_epi16(res[6], 15);
+  __m128i sign7 = _mm_srai_epi16(res[7], 15);
+
+  if (bit == 2) {
+    const __m128i const_rounding = _mm_set1_epi16(1);
+    res[0] = _mm_add_epi16(res[0], const_rounding);
+    res[1] = _mm_add_epi16(res[1], const_rounding);
+    res[2] = _mm_add_epi16(res[2], const_rounding);
+    res[3] = _mm_add_epi16(res[3], const_rounding);
+    res[4] = _mm_add_epi16(res[4], const_rounding);
+    res[5] = _mm_add_epi16(res[5], const_rounding);
+    res[6] = _mm_add_epi16(res[6], const_rounding);
+    res[7] = _mm_add_epi16(res[7], const_rounding);
+  }
+
+  res[0] = _mm_sub_epi16(res[0], sign0);
+  res[1] = _mm_sub_epi16(res[1], sign1);
+  res[2] = _mm_sub_epi16(res[2], sign2);
+  res[3] = _mm_sub_epi16(res[3], sign3);
+  res[4] = _mm_sub_epi16(res[4], sign4);
+  res[5] = _mm_sub_epi16(res[5], sign5);
+  res[6] = _mm_sub_epi16(res[6], sign6);
+  res[7] = _mm_sub_epi16(res[7], sign7);
+
+  if (bit == 1) {
+    res[0] = _mm_srai_epi16(res[0], 1);
+    res[1] = _mm_srai_epi16(res[1], 1);
+    res[2] = _mm_srai_epi16(res[2], 1);
+    res[3] = _mm_srai_epi16(res[3], 1);
+    res[4] = _mm_srai_epi16(res[4], 1);
+    res[5] = _mm_srai_epi16(res[5], 1);
+    res[6] = _mm_srai_epi16(res[6], 1);
+    res[7] = _mm_srai_epi16(res[7], 1);
+  } else {
+    res[0] = _mm_srai_epi16(res[0], 2);
+    res[1] = _mm_srai_epi16(res[1], 2);
+    res[2] = _mm_srai_epi16(res[2], 2);
+    res[3] = _mm_srai_epi16(res[3], 2);
+    res[4] = _mm_srai_epi16(res[4], 2);
+    res[5] = _mm_srai_epi16(res[5], 2);
+    res[6] = _mm_srai_epi16(res[6], 2);
+    res[7] = _mm_srai_epi16(res[7], 2);
+  }
+}
+
+// write 8x8 array
+static INLINE void write_buffer_8x8(tran_low_t *output, __m128i *res,
+                                    int stride) {
+  store_output(&res[0], (output + 0 * stride));
+  store_output(&res[1], (output + 1 * stride));
+  store_output(&res[2], (output + 2 * stride));
+  store_output(&res[3], (output + 3 * stride));
+  store_output(&res[4], (output + 4 * stride));
+  store_output(&res[5], (output + 5 * stride));
+  store_output(&res[6], (output + 6 * stride));
+  store_output(&res[7], (output + 7 * stride));
+}
+
+// perform in-place transpose
+static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
+  // 00 10 01 11 02 12 03 13
+  // 20 30 21 31 22 32 23 33
+  // 04 14 05 15 06 16 07 17
+  // 24 34 25 35 26 36 27 37
+  // 40 50 41 51 42 52 43 53
+  // 60 70 61 71 62 72 63 73
+  // 44 54 45 55 46 56 47 57
+  // 64 74 65 75 66 76 67 77
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+  // 00 10 20 30 01 11 21 31
+  // 40 50 60 70 41 51 61 71
+  // 02 12 22 32 03 13 23 33
+  // 42 52 62 72 43 53 63 73
+  // 04 14 24 34 05 15 25 35
+  // 44 54 64 74 45 55 65 75
+  // 06 16 26 36 07 17 27 37
+  // 46 56 66 76 47 57 67 77
+  res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
+  res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
+  res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
+  res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
+  res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
+  res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
+  res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
+  res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
+  // 00 10 20 30 40 50 60 70
+  // 01 11 21 31 41 51 61 71
+  // 02 12 22 32 42 52 62 72
+  // 03 13 23 33 43 53 63 73
+  // 04 14 24 34 44 54 64 74
+  // 05 15 25 35 45 55 65 75
+  // 06 16 26 36 46 56 66 76
+  // 07 17 27 37 47 57 67 77
+}
+
+static void fdct8_sse2(__m128i *in) {
+  // constants
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+
+  // stage 1
+  s0 = _mm_add_epi16(in[0], in[7]);
+  s1 = _mm_add_epi16(in[1], in[6]);
+  s2 = _mm_add_epi16(in[2], in[5]);
+  s3 = _mm_add_epi16(in[3], in[4]);
+  s4 = _mm_sub_epi16(in[3], in[4]);
+  s5 = _mm_sub_epi16(in[2], in[5]);
+  s6 = _mm_sub_epi16(in[1], in[6]);
+  s7 = _mm_sub_epi16(in[0], in[7]);
+
+  u0 = _mm_add_epi16(s0, s3);
+  u1 = _mm_add_epi16(s1, s2);
+  u2 = _mm_sub_epi16(s1, s2);
+  u3 = _mm_sub_epi16(s0, s3);
+  // interleave and perform butterfly multiplication/addition
+  v0 = _mm_unpacklo_epi16(u0, u1);
+  v1 = _mm_unpackhi_epi16(u0, u1);
+  v2 = _mm_unpacklo_epi16(u2, u3);
+  v3 = _mm_unpackhi_epi16(u2, u3);
+
+  u0 = _mm_madd_epi16(v0, k__cospi_p16_p16);
+  u1 = _mm_madd_epi16(v1, k__cospi_p16_p16);
+  u2 = _mm_madd_epi16(v0, k__cospi_p16_m16);
+  u3 = _mm_madd_epi16(v1, k__cospi_p16_m16);
+  u4 = _mm_madd_epi16(v2, k__cospi_p24_p08);
+  u5 = _mm_madd_epi16(v3, k__cospi_p24_p08);
+  u6 = _mm_madd_epi16(v2, k__cospi_m08_p24);
+  u7 = _mm_madd_epi16(v3, k__cospi_m08_p24);
+
+  // shift and rounding
+  v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u0, u1);
+  in[2] = _mm_packs_epi32(u4, u5);
+  in[4] = _mm_packs_epi32(u2, u3);
+  in[6] = _mm_packs_epi32(u6, u7);
+
+  // stage 2
+  // interleave and perform butterfly multiplication/addition
+  u0 = _mm_unpacklo_epi16(s6, s5);
+  u1 = _mm_unpackhi_epi16(s6, s5);
+  v0 = _mm_madd_epi16(u0, k__cospi_p16_m16);
+  v1 = _mm_madd_epi16(u1, k__cospi_p16_m16);
+  v2 = _mm_madd_epi16(u0, k__cospi_p16_p16);
+  v3 = _mm_madd_epi16(u1, k__cospi_p16_p16);
+
+  // shift and rounding
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+
+  u0 = _mm_packs_epi32(v0, v1);
+  u1 = _mm_packs_epi32(v2, v3);
+
+  // stage 3
+  s0 = _mm_add_epi16(s4, u0);
+  s1 = _mm_sub_epi16(s4, u0);
+  s2 = _mm_sub_epi16(s7, u1);
+  s3 = _mm_add_epi16(s7, u1);
+
+  // stage 4
+  u0 = _mm_unpacklo_epi16(s0, s3);
+  u1 = _mm_unpackhi_epi16(s0, s3);
+  u2 = _mm_unpacklo_epi16(s1, s2);
+  u3 = _mm_unpackhi_epi16(s1, s2);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p28_p04);
+  v1 = _mm_madd_epi16(u1, k__cospi_p28_p04);
+  v2 = _mm_madd_epi16(u2, k__cospi_p12_p20);
+  v3 = _mm_madd_epi16(u3, k__cospi_p12_p20);
+  v4 = _mm_madd_epi16(u2, k__cospi_m20_p12);
+  v5 = _mm_madd_epi16(u3, k__cospi_m20_p12);
+  v6 = _mm_madd_epi16(u0, k__cospi_m04_p28);
+  v7 = _mm_madd_epi16(u1, k__cospi_m04_p28);
+
+  // shift and rounding
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  in[1] = _mm_packs_epi32(v0, v1);
+  in[3] = _mm_packs_epi32(v4, v5);
+  in[5] = _mm_packs_epi32(v2, v3);
+  in[7] = _mm_packs_epi32(v6, v7);
+
+  // transpose
+  array_transpose_8x8(in, in);
+}
+
+static void fadst8_sse2(__m128i *in) {
+  // Constants
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__const_0 = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+
+  // properly aligned for butterfly input
+  in0  = in[7];
+  in1  = in[0];
+  in2  = in[5];
+  in3  = in[2];
+  in4  = in[3];
+  in5  = in[4];
+  in6  = in[1];
+  in7  = in[6];
+
+  // column transformation
+  // stage 1
+  // interleave and multiply/add into 32-bit integer
+  s0 = _mm_unpacklo_epi16(in0, in1);
+  s1 = _mm_unpackhi_epi16(in0, in1);
+  s2 = _mm_unpacklo_epi16(in2, in3);
+  s3 = _mm_unpackhi_epi16(in2, in3);
+  s4 = _mm_unpacklo_epi16(in4, in5);
+  s5 = _mm_unpackhi_epi16(in4, in5);
+  s6 = _mm_unpacklo_epi16(in6, in7);
+  s7 = _mm_unpackhi_epi16(in6, in7);
+
+  u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
+  u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
+  u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
+  u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
+  u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
+  u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
+  u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
+  u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
+  u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
+  u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
+  u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
+  u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
+  u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
+  u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
+  u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
+  u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
+
+  // addition
+  w0 = _mm_add_epi32(u0, u8);
+  w1 = _mm_add_epi32(u1, u9);
+  w2 = _mm_add_epi32(u2, u10);
+  w3 = _mm_add_epi32(u3, u11);
+  w4 = _mm_add_epi32(u4, u12);
+  w5 = _mm_add_epi32(u5, u13);
+  w6 = _mm_add_epi32(u6, u14);
+  w7 = _mm_add_epi32(u7, u15);
+  w8 = _mm_sub_epi32(u0, u8);
+  w9 = _mm_sub_epi32(u1, u9);
+  w10 = _mm_sub_epi32(u2, u10);
+  w11 = _mm_sub_epi32(u3, u11);
+  w12 = _mm_sub_epi32(u4, u12);
+  w13 = _mm_sub_epi32(u5, u13);
+  w14 = _mm_sub_epi32(u6, u14);
+  w15 = _mm_sub_epi32(u7, u15);
+
+  // shift and rounding
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+  v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
+  v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
+  v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
+  v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
+  v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
+  v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
+  v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
+  v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+  u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
+  u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
+  u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
+  u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
+  u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
+  u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
+  u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
+  u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
+
+  // back to 16-bit and pack 8 integers into __m128i
+  in[0] = _mm_packs_epi32(u0, u1);
+  in[1] = _mm_packs_epi32(u2, u3);
+  in[2] = _mm_packs_epi32(u4, u5);
+  in[3] = _mm_packs_epi32(u6, u7);
+  in[4] = _mm_packs_epi32(u8, u9);
+  in[5] = _mm_packs_epi32(u10, u11);
+  in[6] = _mm_packs_epi32(u12, u13);
+  in[7] = _mm_packs_epi32(u14, u15);
+
+  // stage 2
+  s0 = _mm_add_epi16(in[0], in[2]);
+  s1 = _mm_add_epi16(in[1], in[3]);
+  s2 = _mm_sub_epi16(in[0], in[2]);
+  s3 = _mm_sub_epi16(in[1], in[3]);
+  u0 = _mm_unpacklo_epi16(in[4], in[5]);
+  u1 = _mm_unpackhi_epi16(in[4], in[5]);
+  u2 = _mm_unpacklo_epi16(in[6], in[7]);
+  u3 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
+  v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
+  v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
+  v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
+  v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
+  v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
+  v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
+  v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
+
+  w0 = _mm_add_epi32(v0, v4);
+  w1 = _mm_add_epi32(v1, v5);
+  w2 = _mm_add_epi32(v2, v6);
+  w3 = _mm_add_epi32(v3, v7);
+  w4 = _mm_sub_epi32(v0, v4);
+  w5 = _mm_sub_epi32(v1, v5);
+  w6 = _mm_sub_epi32(v2, v6);
+  w7 = _mm_sub_epi32(v3, v7);
+
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  // back to 16-bit intergers
+  s4 = _mm_packs_epi32(u0, u1);
+  s5 = _mm_packs_epi32(u2, u3);
+  s6 = _mm_packs_epi32(u4, u5);
+  s7 = _mm_packs_epi32(u6, u7);
+
+  // stage 3
+  u0 = _mm_unpacklo_epi16(s2, s3);
+  u1 = _mm_unpackhi_epi16(s2, s3);
+  u2 = _mm_unpacklo_epi16(s6, s7);
+  u3 = _mm_unpackhi_epi16(s6, s7);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
+  v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
+  v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
+  v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
+  v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
+  v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
+  v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
+  v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
+
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  s2 = _mm_packs_epi32(v0, v1);
+  s3 = _mm_packs_epi32(v2, v3);
+  s6 = _mm_packs_epi32(v4, v5);
+  s7 = _mm_packs_epi32(v6, v7);
+
+  // FIXME(jingning): do subtract using bit inversion?
+  in[0] = s0;
+  in[1] = _mm_sub_epi16(k__const_0, s4);
+  in[2] = s6;
+  in[3] = _mm_sub_epi16(k__const_0, s2);
+  in[4] = s3;
+  in[5] = _mm_sub_epi16(k__const_0, s7);
+  in[6] = s5;
+  in[7] = _mm_sub_epi16(k__const_0, s1);
+
+  // transpose
+  array_transpose_8x8(in, in);
+}
+
+void vp9_fht8x8_sse2(const int16_t *input, tran_low_t *output,
+                     int stride, int tx_type) {
+  __m128i in[8];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      vpx_fdct8x8_sse2(input, output, stride);
+      break;
+    case ADST_DCT:
+      load_buffer_8x8(input, in, stride);
+      fadst8_sse2(in);
+      fdct8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case DCT_ADST:
+      load_buffer_8x8(input, in, stride);
+      fdct8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case ADST_ADST:
+      load_buffer_8x8(input, in, stride);
+      fadst8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
+
+static INLINE void load_buffer_16x16(const int16_t* input, __m128i *in0,
+                                     __m128i *in1, int stride) {
+  // load first 8 columns
+  load_buffer_8x8(input, in0, stride);
+  load_buffer_8x8(input + 8 * stride, in0 + 8, stride);
+
+  input += 8;
+  // load second 8 columns
+  load_buffer_8x8(input, in1, stride);
+  load_buffer_8x8(input + 8 * stride, in1 + 8, stride);
+}
+
+static INLINE void write_buffer_16x16(tran_low_t *output, __m128i *in0,
+                                      __m128i *in1, int stride) {
+  // write first 8 columns
+  write_buffer_8x8(output, in0, stride);
+  write_buffer_8x8(output + 8 * stride, in0 + 8, stride);
+  // write second 8 columns
+  output += 8;
+  write_buffer_8x8(output, in1, stride);
+  write_buffer_8x8(output + 8 * stride, in1 + 8, stride);
+}
+
+static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
+  __m128i tbuf[8];
+  array_transpose_8x8(res0, res0);
+  array_transpose_8x8(res1, tbuf);
+  array_transpose_8x8(res0 + 8, res1);
+  array_transpose_8x8(res1 + 8, res1 + 8);
+
+  res0[8] = tbuf[0];
+  res0[9] = tbuf[1];
+  res0[10] = tbuf[2];
+  res0[11] = tbuf[3];
+  res0[12] = tbuf[4];
+  res0[13] = tbuf[5];
+  res0[14] = tbuf[6];
+  res0[15] = tbuf[7];
+}
+
+static INLINE void right_shift_16x16(__m128i *res0, __m128i *res1) {
+  // perform rounding operations
+  right_shift_8x8(res0, 2);
+  right_shift_8x8(res0 + 8, 2);
+  right_shift_8x8(res1, 2);
+  right_shift_8x8(res1 + 8, 2);
+}
+
+static void fdct16_8col(__m128i *in) {
+  // perform 16x16 1-D DCT for 8 columns
+  __m128i i[8], s[8], p[8], t[8], u[16], v[16];
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
+  const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
+  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
+  const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
+  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  // stage 1
+  i[0] = _mm_add_epi16(in[0], in[15]);
+  i[1] = _mm_add_epi16(in[1], in[14]);
+  i[2] = _mm_add_epi16(in[2], in[13]);
+  i[3] = _mm_add_epi16(in[3], in[12]);
+  i[4] = _mm_add_epi16(in[4], in[11]);
+  i[5] = _mm_add_epi16(in[5], in[10]);
+  i[6] = _mm_add_epi16(in[6], in[9]);
+  i[7] = _mm_add_epi16(in[7], in[8]);
+
+  s[0] = _mm_sub_epi16(in[7], in[8]);
+  s[1] = _mm_sub_epi16(in[6], in[9]);
+  s[2] = _mm_sub_epi16(in[5], in[10]);
+  s[3] = _mm_sub_epi16(in[4], in[11]);
+  s[4] = _mm_sub_epi16(in[3], in[12]);
+  s[5] = _mm_sub_epi16(in[2], in[13]);
+  s[6] = _mm_sub_epi16(in[1], in[14]);
+  s[7] = _mm_sub_epi16(in[0], in[15]);
+
+  p[0] = _mm_add_epi16(i[0], i[7]);
+  p[1] = _mm_add_epi16(i[1], i[6]);
+  p[2] = _mm_add_epi16(i[2], i[5]);
+  p[3] = _mm_add_epi16(i[3], i[4]);
+  p[4] = _mm_sub_epi16(i[3], i[4]);
+  p[5] = _mm_sub_epi16(i[2], i[5]);
+  p[6] = _mm_sub_epi16(i[1], i[6]);
+  p[7] = _mm_sub_epi16(i[0], i[7]);
+
+  u[0] = _mm_add_epi16(p[0], p[3]);
+  u[1] = _mm_add_epi16(p[1], p[2]);
+  u[2] = _mm_sub_epi16(p[1], p[2]);
+  u[3] = _mm_sub_epi16(p[0], p[3]);
+
+  v[0] = _mm_unpacklo_epi16(u[0], u[1]);
+  v[1] = _mm_unpackhi_epi16(u[0], u[1]);
+  v[2] = _mm_unpacklo_epi16(u[2], u[3]);
+  v[3] = _mm_unpackhi_epi16(u[2], u[3]);
+
+  u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);
+  u[1] = _mm_madd_epi16(v[1], k__cospi_p16_p16);
+  u[2] = _mm_madd_epi16(v[0], k__cospi_p16_m16);
+  u[3] = _mm_madd_epi16(v[1], k__cospi_p16_m16);
+  u[4] = _mm_madd_epi16(v[2], k__cospi_p24_p08);
+  u[5] = _mm_madd_epi16(v[3], k__cospi_p24_p08);
+  u[6] = _mm_madd_epi16(v[2], k__cospi_m08_p24);
+  u[7] = _mm_madd_epi16(v[3], k__cospi_m08_p24);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[1]);
+  in[4] = _mm_packs_epi32(u[4], u[5]);
+  in[8] = _mm_packs_epi32(u[2], u[3]);
+  in[12] = _mm_packs_epi32(u[6], u[7]);
+
+  u[0] = _mm_unpacklo_epi16(p[5], p[6]);
+  u[1] = _mm_unpackhi_epi16(p[5], p[6]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+
+  u[0] = _mm_packs_epi32(v[0], v[1]);
+  u[1] = _mm_packs_epi32(v[2], v[3]);
+
+  t[0] = _mm_add_epi16(p[4], u[0]);
+  t[1] = _mm_sub_epi16(p[4], u[0]);
+  t[2] = _mm_sub_epi16(p[7], u[1]);
+  t[3] = _mm_add_epi16(p[7], u[1]);
+
+  u[0] = _mm_unpacklo_epi16(t[0], t[3]);
+  u[1] = _mm_unpackhi_epi16(t[0], t[3]);
+  u[2] = _mm_unpacklo_epi16(t[1], t[2]);
+  u[3] = _mm_unpackhi_epi16(t[1], t[2]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p28_p04);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p28_p04);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_p12_p20);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_p12_p20);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m20_p12);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m20_p12);
+  v[6] = _mm_madd_epi16(u[0], k__cospi_m04_p28);
+  v[7] = _mm_madd_epi16(u[1], k__cospi_m04_p28);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  in[2] = _mm_packs_epi32(v[0], v[1]);
+  in[6] = _mm_packs_epi32(v[4], v[5]);
+  in[10] = _mm_packs_epi32(v[2], v[3]);
+  in[14] = _mm_packs_epi32(v[6], v[7]);
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[2], s[5]);
+  u[1] = _mm_unpackhi_epi16(s[2], s[5]);
+  u[2] = _mm_unpacklo_epi16(s[3], s[4]);
+  u[3] = _mm_unpackhi_epi16(s[3], s[4]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+  v[6] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[7] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[2] = _mm_packs_epi32(v[0], v[1]);
+  t[3] = _mm_packs_epi32(v[2], v[3]);
+  t[4] = _mm_packs_epi32(v[4], v[5]);
+  t[5] = _mm_packs_epi32(v[6], v[7]);
+
+  // stage 3
+  p[0] = _mm_add_epi16(s[0], t[3]);
+  p[1] = _mm_add_epi16(s[1], t[2]);
+  p[2] = _mm_sub_epi16(s[1], t[2]);
+  p[3] = _mm_sub_epi16(s[0], t[3]);
+  p[4] = _mm_sub_epi16(s[7], t[4]);
+  p[5] = _mm_sub_epi16(s[6], t[5]);
+  p[6] = _mm_add_epi16(s[6], t[5]);
+  p[7] = _mm_add_epi16(s[7], t[4]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(p[1], p[6]);
+  u[1] = _mm_unpackhi_epi16(p[1], p[6]);
+  u[2] = _mm_unpacklo_epi16(p[2], p[5]);
+  u[3] = _mm_unpackhi_epi16(p[2], p[5]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m08_p24);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m08_p24);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_p24_p08);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_p24_p08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p08_m24);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p08_m24);
+  v[6] = _mm_madd_epi16(u[0], k__cospi_p24_p08);
+  v[7] = _mm_madd_epi16(u[1], k__cospi_p24_p08);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[1] = _mm_packs_epi32(v[0], v[1]);
+  t[2] = _mm_packs_epi32(v[2], v[3]);
+  t[5] = _mm_packs_epi32(v[4], v[5]);
+  t[6] = _mm_packs_epi32(v[6], v[7]);
+
+  // stage 5
+  s[0] = _mm_add_epi16(p[0], t[1]);
+  s[1] = _mm_sub_epi16(p[0], t[1]);
+  s[2] = _mm_add_epi16(p[3], t[2]);
+  s[3] = _mm_sub_epi16(p[3], t[2]);
+  s[4] = _mm_sub_epi16(p[4], t[5]);
+  s[5] = _mm_add_epi16(p[4], t[5]);
+  s[6] = _mm_sub_epi16(p[7], t[6]);
+  s[7] = _mm_add_epi16(p[7], t[6]);
+
+  // stage 6
+  u[0] = _mm_unpacklo_epi16(s[0], s[7]);
+  u[1] = _mm_unpackhi_epi16(s[0], s[7]);
+  u[2] = _mm_unpacklo_epi16(s[1], s[6]);
+  u[3] = _mm_unpackhi_epi16(s[1], s[6]);
+  u[4] = _mm_unpacklo_epi16(s[2], s[5]);
+  u[5] = _mm_unpackhi_epi16(s[2], s[5]);
+  u[6] = _mm_unpacklo_epi16(s[3], s[4]);
+  u[7] = _mm_unpackhi_epi16(s[3], s[4]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p30_p02);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p30_p02);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_p14_p18);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_p14_p18);
+  v[4] = _mm_madd_epi16(u[4], k__cospi_p22_p10);
+  v[5] = _mm_madd_epi16(u[5], k__cospi_p22_p10);
+  v[6] = _mm_madd_epi16(u[6], k__cospi_p06_p26);
+  v[7] = _mm_madd_epi16(u[7], k__cospi_p06_p26);
+  v[8] = _mm_madd_epi16(u[6], k__cospi_m26_p06);
+  v[9] = _mm_madd_epi16(u[7], k__cospi_m26_p06);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_m10_p22);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_m10_p22);
+  v[12] = _mm_madd_epi16(u[2], k__cospi_m18_p14);
+  v[13] = _mm_madd_epi16(u[3], k__cospi_m18_p14);
+  v[14] = _mm_madd_epi16(u[0], k__cospi_m02_p30);
+  v[15] = _mm_madd_epi16(u[1], k__cospi_m02_p30);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  in[1]  = _mm_packs_epi32(v[0], v[1]);
+  in[9]  = _mm_packs_epi32(v[2], v[3]);
+  in[5]  = _mm_packs_epi32(v[4], v[5]);
+  in[13] = _mm_packs_epi32(v[6], v[7]);
+  in[3]  = _mm_packs_epi32(v[8], v[9]);
+  in[11] = _mm_packs_epi32(v[10], v[11]);
+  in[7]  = _mm_packs_epi32(v[12], v[13]);
+  in[15] = _mm_packs_epi32(v[14], v[15]);
+}
+
+static void fadst16_8col(__m128i *in) {
+  // perform 16x16 1-D ADST for 8 columns
+  __m128i s[16], x[16], u[32], v[32];
+  const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
+  const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kZero = _mm_set1_epi16(0);
+
+  u[0] = _mm_unpacklo_epi16(in[15], in[0]);
+  u[1] = _mm_unpackhi_epi16(in[15], in[0]);
+  u[2] = _mm_unpacklo_epi16(in[13], in[2]);
+  u[3] = _mm_unpackhi_epi16(in[13], in[2]);
+  u[4] = _mm_unpacklo_epi16(in[11], in[4]);
+  u[5] = _mm_unpackhi_epi16(in[11], in[4]);
+  u[6] = _mm_unpacklo_epi16(in[9], in[6]);
+  u[7] = _mm_unpackhi_epi16(in[9], in[6]);
+  u[8] = _mm_unpacklo_epi16(in[7], in[8]);
+  u[9] = _mm_unpackhi_epi16(in[7], in[8]);
+  u[10] = _mm_unpacklo_epi16(in[5], in[10]);
+  u[11] = _mm_unpackhi_epi16(in[5], in[10]);
+  u[12] = _mm_unpacklo_epi16(in[3], in[12]);
+  u[13] = _mm_unpackhi_epi16(in[3], in[12]);
+  u[14] = _mm_unpacklo_epi16(in[1], in[14]);
+  u[15] = _mm_unpackhi_epi16(in[1], in[14]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
+  v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
+  v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
+  v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
+  v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
+  v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
+  v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
+  v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
+  v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
+  v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
+  v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
+  v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
+  v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
+  v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
+  v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
+  v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
+  v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
+
+  u[0] = _mm_add_epi32(v[0], v[16]);
+  u[1] = _mm_add_epi32(v[1], v[17]);
+  u[2] = _mm_add_epi32(v[2], v[18]);
+  u[3] = _mm_add_epi32(v[3], v[19]);
+  u[4] = _mm_add_epi32(v[4], v[20]);
+  u[5] = _mm_add_epi32(v[5], v[21]);
+  u[6] = _mm_add_epi32(v[6], v[22]);
+  u[7] = _mm_add_epi32(v[7], v[23]);
+  u[8] = _mm_add_epi32(v[8], v[24]);
+  u[9] = _mm_add_epi32(v[9], v[25]);
+  u[10] = _mm_add_epi32(v[10], v[26]);
+  u[11] = _mm_add_epi32(v[11], v[27]);
+  u[12] = _mm_add_epi32(v[12], v[28]);
+  u[13] = _mm_add_epi32(v[13], v[29]);
+  u[14] = _mm_add_epi32(v[14], v[30]);
+  u[15] = _mm_add_epi32(v[15], v[31]);
+  u[16] = _mm_sub_epi32(v[0], v[16]);
+  u[17] = _mm_sub_epi32(v[1], v[17]);
+  u[18] = _mm_sub_epi32(v[2], v[18]);
+  u[19] = _mm_sub_epi32(v[3], v[19]);
+  u[20] = _mm_sub_epi32(v[4], v[20]);
+  u[21] = _mm_sub_epi32(v[5], v[21]);
+  u[22] = _mm_sub_epi32(v[6], v[22]);
+  u[23] = _mm_sub_epi32(v[7], v[23]);
+  u[24] = _mm_sub_epi32(v[8], v[24]);
+  u[25] = _mm_sub_epi32(v[9], v[25]);
+  u[26] = _mm_sub_epi32(v[10], v[26]);
+  u[27] = _mm_sub_epi32(v[11], v[27]);
+  u[28] = _mm_sub_epi32(v[12], v[28]);
+  u[29] = _mm_sub_epi32(v[13], v[29]);
+  u[30] = _mm_sub_epi32(v[14], v[30]);
+  u[31] = _mm_sub_epi32(v[15], v[31]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+  v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
+  v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
+  v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
+  v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
+  v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
+  v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
+  v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
+  v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
+  v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
+  v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
+  v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
+  v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
+  v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
+  v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
+  v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
+  v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+  u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
+  u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
+  u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
+  u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
+  u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
+  u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
+  u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
+  u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
+  u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
+  u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
+  u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
+  u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
+  u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
+  u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
+  u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
+  u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
+
+  s[0] = _mm_packs_epi32(u[0], u[1]);
+  s[1] = _mm_packs_epi32(u[2], u[3]);
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_packs_epi32(u[8], u[9]);
+  s[5] = _mm_packs_epi32(u[10], u[11]);
+  s[6] = _mm_packs_epi32(u[12], u[13]);
+  s[7] = _mm_packs_epi32(u[14], u[15]);
+  s[8] = _mm_packs_epi32(u[16], u[17]);
+  s[9] = _mm_packs_epi32(u[18], u[19]);
+  s[10] = _mm_packs_epi32(u[20], u[21]);
+  s[11] = _mm_packs_epi32(u[22], u[23]);
+  s[12] = _mm_packs_epi32(u[24], u[25]);
+  s[13] = _mm_packs_epi32(u[26], u[27]);
+  s[14] = _mm_packs_epi32(u[28], u[29]);
+  s[15] = _mm_packs_epi32(u[30], u[31]);
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[9]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[9]);
+  u[2] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[3] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[4] = _mm_unpacklo_epi16(s[12], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[12], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], v[8]);
+  u[1] = _mm_add_epi32(v[1], v[9]);
+  u[2] = _mm_add_epi32(v[2], v[10]);
+  u[3] = _mm_add_epi32(v[3], v[11]);
+  u[4] = _mm_add_epi32(v[4], v[12]);
+  u[5] = _mm_add_epi32(v[5], v[13]);
+  u[6] = _mm_add_epi32(v[6], v[14]);
+  u[7] = _mm_add_epi32(v[7], v[15]);
+  u[8] = _mm_sub_epi32(v[0], v[8]);
+  u[9] = _mm_sub_epi32(v[1], v[9]);
+  u[10] = _mm_sub_epi32(v[2], v[10]);
+  u[11] = _mm_sub_epi32(v[3], v[11]);
+  u[12] = _mm_sub_epi32(v[4], v[12]);
+  u[13] = _mm_sub_epi32(v[5], v[13]);
+  u[14] = _mm_sub_epi32(v[6], v[14]);
+  u[15] = _mm_sub_epi32(v[7], v[15]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+  x[0] = _mm_add_epi16(s[0], s[4]);
+  x[1] = _mm_add_epi16(s[1], s[5]);
+  x[2] = _mm_add_epi16(s[2], s[6]);
+  x[3] = _mm_add_epi16(s[3], s[7]);
+  x[4] = _mm_sub_epi16(s[0], s[4]);
+  x[5] = _mm_sub_epi16(s[1], s[5]);
+  x[6] = _mm_sub_epi16(s[2], s[6]);
+  x[7] = _mm_sub_epi16(s[3], s[7]);
+  x[8] = _mm_packs_epi32(u[0], u[1]);
+  x[9] = _mm_packs_epi32(u[2], u[3]);
+  x[10] = _mm_packs_epi32(u[4], u[5]);
+  x[11] = _mm_packs_epi32(u[6], u[7]);
+  x[12] = _mm_packs_epi32(u[8], u[9]);
+  x[13] = _mm_packs_epi32(u[10], u[11]);
+  x[14] = _mm_packs_epi32(u[12], u[13]);
+  x[15] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  u[0] = _mm_unpacklo_epi16(x[4], x[5]);
+  u[1] = _mm_unpackhi_epi16(x[4], x[5]);
+  u[2] = _mm_unpacklo_epi16(x[6], x[7]);
+  u[3] = _mm_unpackhi_epi16(x[6], x[7]);
+  u[4] = _mm_unpacklo_epi16(x[12], x[13]);
+  u[5] = _mm_unpackhi_epi16(x[12], x[13]);
+  u[6] = _mm_unpacklo_epi16(x[14], x[15]);
+  u[7] = _mm_unpackhi_epi16(x[14], x[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], v[4]);
+  u[1] = _mm_add_epi32(v[1], v[5]);
+  u[2] = _mm_add_epi32(v[2], v[6]);
+  u[3] = _mm_add_epi32(v[3], v[7]);
+  u[4] = _mm_sub_epi32(v[0], v[4]);
+  u[5] = _mm_sub_epi32(v[1], v[5]);
+  u[6] = _mm_sub_epi32(v[2], v[6]);
+  u[7] = _mm_sub_epi32(v[3], v[7]);
+  u[8] = _mm_add_epi32(v[8], v[12]);
+  u[9] = _mm_add_epi32(v[9], v[13]);
+  u[10] = _mm_add_epi32(v[10], v[14]);
+  u[11] = _mm_add_epi32(v[11], v[15]);
+  u[12] = _mm_sub_epi32(v[8], v[12]);
+  u[13] = _mm_sub_epi32(v[9], v[13]);
+  u[14] = _mm_sub_epi32(v[10], v[14]);
+  u[15] = _mm_sub_epi32(v[11], v[15]);
+
+  u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[0] = _mm_add_epi16(x[0], x[2]);
+  s[1] = _mm_add_epi16(x[1], x[3]);
+  s[2] = _mm_sub_epi16(x[0], x[2]);
+  s[3] = _mm_sub_epi16(x[1], x[3]);
+  s[4] = _mm_packs_epi32(v[0], v[1]);
+  s[5] = _mm_packs_epi32(v[2], v[3]);
+  s[6] = _mm_packs_epi32(v[4], v[5]);
+  s[7] = _mm_packs_epi32(v[6], v[7]);
+  s[8] = _mm_add_epi16(x[8], x[10]);
+  s[9] = _mm_add_epi16(x[9], x[11]);
+  s[10] = _mm_sub_epi16(x[8], x[10]);
+  s[11] = _mm_sub_epi16(x[9], x[11]);
+  s[12] = _mm_packs_epi32(v[8], v[9]);
+  s[13] = _mm_packs_epi32(v[10], v[11]);
+  s[14] = _mm_packs_epi32(v[12], v[13]);
+  s[15] = _mm_packs_epi32(v[14], v[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(s[2], s[3]);
+  u[1] = _mm_unpackhi_epi16(s[2], s[3]);
+  u[2] = _mm_unpacklo_epi16(s[6], s[7]);
+  u[3] = _mm_unpackhi_epi16(s[6], s[7]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  in[0] = s[0];
+  in[1] = _mm_sub_epi16(kZero, s[8]);
+  in[2] = s[12];
+  in[3] = _mm_sub_epi16(kZero, s[4]);
+  in[4] = _mm_packs_epi32(v[4], v[5]);
+  in[5] = _mm_packs_epi32(v[12], v[13]);
+  in[6] = _mm_packs_epi32(v[8], v[9]);
+  in[7] = _mm_packs_epi32(v[0], v[1]);
+  in[8] = _mm_packs_epi32(v[2], v[3]);
+  in[9] = _mm_packs_epi32(v[10], v[11]);
+  in[10] = _mm_packs_epi32(v[14], v[15]);
+  in[11] = _mm_packs_epi32(v[6], v[7]);
+  in[12] = s[5];
+  in[13] = _mm_sub_epi16(kZero, s[13]);
+  in[14] = s[9];
+  in[15] = _mm_sub_epi16(kZero, s[1]);
+}
+
+static void fdct16_sse2(__m128i *in0, __m128i *in1) {
+  fdct16_8col(in0);
+  fdct16_8col(in1);
+  array_transpose_16x16(in0, in1);
+}
+
+static void fadst16_sse2(__m128i *in0, __m128i *in1) {
+  fadst16_8col(in0);
+  fadst16_8col(in1);
+  array_transpose_16x16(in0, in1);
+}
+
+void vp9_fht16x16_sse2(const int16_t *input, tran_low_t *output,
+                       int stride, int tx_type) {
+  __m128i in0[16], in1[16];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      vpx_fdct16x16_sse2(input, output, stride);
+      break;
+    case ADST_DCT:
+      load_buffer_16x16(input, in0, in1, stride);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fdct16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case DCT_ADST:
+      load_buffer_16x16(input, in0, in1, stride);
+      fdct16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case ADST_ADST:
+      load_buffer_16x16(input, in0, in1, stride);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_dct_sse2.asm b/libvpx/libvpx/vp9/encoder/x86/vp9_dct_sse2.asm
new file mode 100644
index 0000000..ced37bd
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_dct_sse2.asm
@@ -0,0 +1,86 @@
+;
+;  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%define private_prefix vp9
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro TRANSFORM_COLS 0
+  paddw           m0,        m1
+  movq            m4,        m0
+  psubw           m3,        m2
+  psubw           m4,        m3
+  psraw           m4,        1
+  movq            m5,        m4
+  psubw           m5,        m1 ;b1
+  psubw           m4,        m2 ;c1
+  psubw           m0,        m4
+  paddw           m3,        m5
+                                ; m0 a0
+  SWAP            1,         4  ; m1 c1
+  SWAP            2,         3  ; m2 d1
+  SWAP            3,         5  ; m3 b1
+%endmacro
+
+%macro TRANSPOSE_4X4 0
+                                ; 00 01 02 03
+                                ; 10 11 12 13
+                                ; 20 21 22 23
+                                ; 30 31 32 33
+  punpcklwd       m0,        m1 ; 00 10 01 11  02 12 03 13
+  punpcklwd       m2,        m3 ; 20 30 21 31  22 32 23 33
+  mova            m1,        m0
+  punpckldq       m0,        m2 ; 00 10 20 30  01 11 21 31
+  punpckhdq       m1,        m2 ; 02 12 22 32  03 13 23 33
+%endmacro
+
+INIT_XMM sse2
+cglobal fwht4x4, 3, 4, 8, input, output, stride
+  lea             r3q,       [inputq + strideq*4]
+  movq            m0,        [inputq] ;a1
+  movq            m1,        [inputq + strideq*2] ;b1
+  movq            m2,        [r3q] ;c1
+  movq            m3,        [r3q + strideq*2] ;d1
+
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+  SWAP            1,         2
+  psrldq          m1,        m0, 8
+  psrldq          m3,        m2, 8
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+
+  psllw           m0,        2
+  psllw           m1,        2
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; sign extension
+  mova            m2,             m0
+  mova            m3,             m1
+  punpcklwd       m0,             m0
+  punpcklwd       m1,             m1
+  punpckhwd       m2,             m2
+  punpckhwd       m3,             m3
+  psrad           m0,             16
+  psrad           m1,             16
+  psrad           m2,             16
+  psrad           m3,             16
+  mova            [outputq],      m0
+  mova            [outputq + 16], m2
+  mova            [outputq + 32], m1
+  mova            [outputq + 48], m3
+%else
+  mova            [outputq],      m0
+  mova            [outputq + 16], m1
+%endif
+
+  RET
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_dct_ssse3.c b/libvpx/libvpx/vp9/encoder/x86/vp9_dct_ssse3.c
new file mode 100644
index 0000000..b09eac0
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_dct_ssse3.c
@@ -0,0 +1,472 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#if defined(_MSC_VER) && _MSC_VER <= 1500
+// Need to include math.h before calling tmmintrin.h/intrin.h
+// in certain versions of MSVS.
+#include <math.h>
+#endif
+#include <tmmintrin.h>  // SSSE3
+
+#include "./vp9_rtcd.h"
+#include "vpx_dsp/x86/inv_txfm_sse2.h"
+#include "vpx_dsp/x86/txfm_common_sse2.h"
+
+void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride,
+                             int16_t* coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t* zbin_ptr,
+                             const int16_t* round_ptr, const int16_t* quant_ptr,
+                             const int16_t* quant_shift_ptr,
+                             int16_t* qcoeff_ptr,
+                             int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+                             uint16_t* eob_ptr,
+                             const int16_t* scan_ptr,
+                             const int16_t* iscan_ptr) {
+  __m128i zero;
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__dual_p16_p16 = dual_set_epi16(23170, 23170);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // Load input
+  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  __m128i *in[8];
+  int index = 0;
+
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)coeff_ptr;
+
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  in[0] = &in0;
+  in[1] = &in1;
+  in[2] = &in2;
+  in[3] = &in3;
+  in[4] = &in4;
+  in[5] = &in5;
+  in[6] = &in6;
+  in[7] = &in7;
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = _mm_add_epi16(in0, in7);
+    const __m128i q1 = _mm_add_epi16(in1, in6);
+    const __m128i q2 = _mm_add_epi16(in2, in5);
+    const __m128i q3 = _mm_add_epi16(in3, in4);
+    const __m128i q4 = _mm_sub_epi16(in3, in4);
+    const __m128i q5 = _mm_sub_epi16(in2, in5);
+    const __m128i q6 = _mm_sub_epi16(in1, in6);
+    const __m128i q7 = _mm_sub_epi16(in0, in7);
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = _mm_add_epi16(q0, q3);
+      const __m128i r1 = _mm_add_epi16(q1, q2);
+      const __m128i r2 = _mm_sub_epi16(q1, q2);
+      const __m128i r3 = _mm_sub_epi16(q0, q3);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+      const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+      const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+      const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+      // dct_const_round_shift
+
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+
+      res0 = _mm_packs_epi32(w0, w1);
+      res4 = _mm_packs_epi32(w2, w3);
+      res2 = _mm_packs_epi32(w4, w5);
+      res6 = _mm_packs_epi32(w6, w7);
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_sub_epi16(q6, q5);
+      const __m128i d1 = _mm_add_epi16(q6, q5);
+      const __m128i r0 = _mm_mulhrs_epi16(d0, k__dual_p16_p16);
+      const __m128i r1 = _mm_mulhrs_epi16(d1, k__dual_p16_p16);
+
+      // Add/subtract
+      const __m128i x0 = _mm_add_epi16(q4, r0);
+      const __m128i x1 = _mm_sub_epi16(q4, r0);
+      const __m128i x2 = _mm_sub_epi16(q7, r1);
+      const __m128i x3 = _mm_add_epi16(q7, r1);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+      const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+      const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+      const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+      // dct_const_round_shift
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+      res1 = _mm_packs_epi32(w0, w1);
+      res7 = _mm_packs_epi32(w2, w3);
+      res5 = _mm_packs_epi32(w4, w5);
+      res3 = _mm_packs_epi32(w6, w7);
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+  }
+
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant, thr;
+    int16_t nzflag;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i*)round_ptr);
+        quant = _mm_load_si128((const __m128i*)quant_ptr);
+        dequant = _mm_load_si128((const __m128i*)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = *in[0];
+        coeff1 = *in[1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // AC only loop
+    index = 2;
+    thr = _mm_srai_epi16(dequant, 1);
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
+        coeff0 = *in[index];
+        coeff1 = *in[index + 1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+            _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
+
+        if (nzflag) {
+          qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+          qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+          qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+          qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+          // Reinsert signs
+          qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+          qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+          qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+          qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+          coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+          coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+        } else {
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+        }
+      }
+
+      if (nzflag) {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+      index += 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_denoiser_sse2.c b/libvpx/libvpx/vp9/encoder/x86/vp9_denoiser_sse2.c
new file mode 100644
index 0000000..883507a
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_denoiser_sse2.c
@@ -0,0 +1,341 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+
+#include "vpx_ports/emmintrin_compat.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_denoiser.h"
+#include "vpx_mem/vpx_mem.h"
+
+// Compute the sum of all pixel differences of this MB.
+static INLINE int sum_diff_16x1(__m128i acc_diff) {
+  const __m128i k_1 = _mm_set1_epi16(1);
+  const __m128i acc_diff_lo =
+      _mm_srai_epi16(_mm_unpacklo_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_hi =
+      _mm_srai_epi16(_mm_unpackhi_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
+  const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
+  const __m128i hgfe_dcba =
+      _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8));
+  const __m128i hgfedcba =
+      _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4));
+  return _mm_cvtsi128_si32(hgfedcba);
+}
+
+// Denoise a 16x1 vector.
+static INLINE __m128i vp9_denoiser_16x1_sse2(const uint8_t *sig,
+                                             const uint8_t *mc_running_avg_y,
+                                             uint8_t *running_avg_y,
+                                             const __m128i *k_0,
+                                             const __m128i *k_4,
+                                             const __m128i *k_8,
+                                             const __m128i *k_16,
+                                             const __m128i *l3,
+                                             const __m128i *l32,
+                                             const __m128i *l21,
+                                             __m128i acc_diff) {
+  // Calculate differences
+  const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0]));
+  const __m128i v_mc_running_avg_y =
+      _mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0]));
+  __m128i v_running_avg_y;
+  const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
+  const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
+  // Obtain the sign. FF if diff is negative.
+  const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, *k_0);
+  // Clamp absolute difference to 16 to be used to get mask. Doing this
+  // allows us to use _mm_cmpgt_epi8, which operates on signed byte.
+  const __m128i clamped_absdiff =
+      _mm_min_epu8(_mm_or_si128(pdiff, ndiff), *k_16);
+  // Get masks for l2 l1 and l0 adjustments.
+  const __m128i mask2 = _mm_cmpgt_epi8(*k_16, clamped_absdiff);
+  const __m128i mask1 = _mm_cmpgt_epi8(*k_8, clamped_absdiff);
+  const __m128i mask0 = _mm_cmpgt_epi8(*k_4, clamped_absdiff);
+  // Get adjustments for l2, l1, and l0.
+  __m128i adj2 = _mm_and_si128(mask2, *l32);
+  const __m128i adj1 = _mm_and_si128(mask1, *l21);
+  const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
+  __m128i adj,  padj, nadj;
+
+  // Combine the adjustments and get absolute adjustments.
+  adj2 = _mm_add_epi8(adj2, adj1);
+  adj = _mm_sub_epi8(*l3, adj2);
+  adj = _mm_andnot_si128(mask0, adj);
+  adj = _mm_or_si128(adj, adj0);
+
+  // Restore the sign and get positive and negative adjustments.
+  padj = _mm_andnot_si128(diff_sign, adj);
+  nadj = _mm_and_si128(diff_sign, adj);
+
+  // Calculate filtered value.
+  v_running_avg_y = _mm_adds_epu8(v_sig, padj);
+  v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
+  _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
+
+  // Adjustments <=7, and each element in acc_diff can fit in signed
+  // char.
+  acc_diff = _mm_adds_epi8(acc_diff, padj);
+  acc_diff = _mm_subs_epi8(acc_diff, nadj);
+  return acc_diff;
+}
+
+// Denoise a 16x1 vector with a weaker filter.
+static INLINE __m128i vp9_denoiser_adj_16x1_sse2(
+    const uint8_t *sig, const uint8_t *mc_running_avg_y,
+    uint8_t *running_avg_y, const __m128i k_0,
+    const __m128i k_delta, __m128i acc_diff) {
+  __m128i v_running_avg_y = _mm_loadu_si128((__m128i *)(&running_avg_y[0]));
+  // Calculate differences.
+  const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0]));
+  const __m128i v_mc_running_avg_y =
+      _mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0]));
+  const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
+  const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
+  // Obtain the sign. FF if diff is negative.
+  const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+  // Clamp absolute difference to delta to get the adjustment.
+  const __m128i adj =
+      _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
+  // Restore the sign and get positive and negative adjustments.
+  __m128i padj, nadj;
+  padj = _mm_andnot_si128(diff_sign, adj);
+  nadj = _mm_and_si128(diff_sign, adj);
+  // Calculate filtered value.
+  v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj);
+  v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj);
+  _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
+
+  // Accumulate the adjustments.
+  acc_diff = _mm_subs_epi8(acc_diff, padj);
+  acc_diff = _mm_adds_epi8(acc_diff, nadj);
+  return acc_diff;
+}
+
+// Denoise 8x8 and 8x16 blocks.
+static int vp9_denoiser_NxM_sse2_small(
+    const uint8_t *sig, int sig_stride, const uint8_t *mc_running_avg_y,
+    int mc_avg_y_stride, uint8_t *running_avg_y, int avg_y_stride,
+    int increase_denoising, BLOCK_SIZE bs, int motion_magnitude, int width) {
+  int sum_diff_thresh, r, sum_diff = 0;
+  const int shift_inc  = (increase_denoising &&
+                          motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ?
+                         1 : 0;
+  uint8_t sig_buffer[8][16], mc_running_buffer[8][16], running_buffer[8][16];
+  __m128i acc_diff = _mm_setzero_si128();
+  const __m128i k_0 = _mm_setzero_si128();
+  const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+  const __m128i k_8 = _mm_set1_epi8(8);
+  const __m128i k_16 = _mm_set1_epi8(16);
+  // Modify each level's adjustment according to motion_magnitude.
+  const __m128i l3 = _mm_set1_epi8(
+      (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
+  // Difference between level 3 and level 2 is 2.
+  const __m128i l32 = _mm_set1_epi8(2);
+  // Difference between level 2 and level 1 is 1.
+  const __m128i l21 = _mm_set1_epi8(1);
+  const int b_height = (4 << b_height_log2_lookup[bs]) >> 1;
+
+  for (r = 0; r < b_height; ++r) {
+    memcpy(sig_buffer[r], sig, width);
+    memcpy(sig_buffer[r] + width, sig + sig_stride, width);
+    memcpy(mc_running_buffer[r], mc_running_avg_y, width);
+    memcpy(mc_running_buffer[r] + width,
+           mc_running_avg_y + mc_avg_y_stride, width);
+    memcpy(running_buffer[r], running_avg_y, width);
+    memcpy(running_buffer[r] + width, running_avg_y + avg_y_stride, width);
+    acc_diff = vp9_denoiser_16x1_sse2(sig_buffer[r],
+                                      mc_running_buffer[r],
+                                      running_buffer[r],
+                                      &k_0, &k_4, &k_8, &k_16,
+                                      &l3, &l32, &l21, acc_diff);
+    memcpy(running_avg_y, running_buffer[r], width);
+    memcpy(running_avg_y + avg_y_stride, running_buffer[r] + width, width);
+    // Update pointers for next iteration.
+    sig += (sig_stride << 1);
+    mc_running_avg_y += (mc_avg_y_stride << 1);
+    running_avg_y += (avg_y_stride << 1);
+  }
+
+  {
+    sum_diff = sum_diff_16x1(acc_diff);
+    sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+    if (abs(sum_diff) > sum_diff_thresh) {
+      // Before returning to copy the block (i.e., apply no denoising),
+      // check if we can still apply some (weaker) temporal filtering to
+      // this block, that would otherwise not be denoised at all. Simplest
+      // is to apply an additional adjustment to running_avg_y to bring it
+      // closer to sig. The adjustment is capped by a maximum delta, and
+      // chosen such that in most cases the resulting sum_diff will be
+      // within the acceptable range given by sum_diff_thresh.
+
+      // The delta is set by the excess of absolute pixel diff over the
+      // threshold.
+      const int delta = ((abs(sum_diff) - sum_diff_thresh) >>
+                         num_pels_log2_lookup[bs]) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        const __m128i k_delta = _mm_set1_epi8(delta);
+        running_avg_y -= avg_y_stride * (b_height << 1);
+        for (r = 0; r < b_height; ++r) {
+          acc_diff = vp9_denoiser_adj_16x1_sse2(
+              sig_buffer[r], mc_running_buffer[r], running_buffer[r],
+              k_0, k_delta, acc_diff);
+          memcpy(running_avg_y, running_buffer[r], width);
+          memcpy(running_avg_y + avg_y_stride,
+                 running_buffer[r] + width, width);
+          // Update pointers for next iteration.
+          running_avg_y += (avg_y_stride << 1);
+        }
+        sum_diff = sum_diff_16x1(acc_diff);
+        if (abs(sum_diff) > sum_diff_thresh) {
+          return COPY_BLOCK;
+        }
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+  }
+  return FILTER_BLOCK;
+}
+
+// Denoise 16x16, 16x32, 32x16, 32x32, 32x64, 64x32 and 64x64 blocks.
+static int vp9_denoiser_NxM_sse2_big(const uint8_t *sig, int sig_stride,
+                                     const uint8_t *mc_running_avg_y,
+                                     int mc_avg_y_stride,
+                                     uint8_t *running_avg_y,
+                                     int avg_y_stride,
+                                     int increase_denoising, BLOCK_SIZE bs,
+                                     int motion_magnitude) {
+  int sum_diff_thresh, r, c, sum_diff = 0;
+  const int shift_inc  = (increase_denoising &&
+                          motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ?
+                         1 : 0;
+  __m128i acc_diff[4][4];
+  const __m128i k_0 = _mm_setzero_si128();
+  const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+  const __m128i k_8 = _mm_set1_epi8(8);
+  const __m128i k_16 = _mm_set1_epi8(16);
+  // Modify each level's adjustment according to motion_magnitude.
+  const __m128i l3 = _mm_set1_epi8(
+      (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
+  // Difference between level 3 and level 2 is 2.
+  const __m128i l32 = _mm_set1_epi8(2);
+  // Difference between level 2 and level 1 is 1.
+  const __m128i l21 = _mm_set1_epi8(1);
+  const int b_width = (4 << b_width_log2_lookup[bs]);
+  const int b_height = (4 << b_height_log2_lookup[bs]);
+  const int b_width_shift4 = b_width >> 4;
+
+  for (r = 0; r < 4; ++r) {
+    for (c = 0; c < b_width_shift4; ++c) {
+      acc_diff[c][r] = _mm_setzero_si128();
+    }
+  }
+
+  for (r = 0; r < b_height; ++r) {
+    for (c = 0; c < b_width_shift4; ++c) {
+      acc_diff[c][r>>4] = vp9_denoiser_16x1_sse2(
+          sig, mc_running_avg_y, running_avg_y, &k_0, &k_4,
+          &k_8, &k_16, &l3, &l32, &l21, acc_diff[c][r>>4]);
+      // Update pointers for next iteration.
+      sig += 16;
+      mc_running_avg_y += 16;
+      running_avg_y += 16;
+    }
+
+    if ((r & 0xf) == 0xf || (bs == BLOCK_16X8 && r == 7)) {
+      for (c = 0; c < b_width_shift4; ++c) {
+        sum_diff += sum_diff_16x1(acc_diff[c][r>>4]);
+      }
+    }
+
+    // Update pointers for next iteration.
+    sig = sig - b_width + sig_stride;
+    mc_running_avg_y = mc_running_avg_y - b_width + mc_avg_y_stride;
+    running_avg_y = running_avg_y - b_width + avg_y_stride;
+  }
+
+  {
+    sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+    if (abs(sum_diff) > sum_diff_thresh) {
+      const int delta = ((abs(sum_diff) - sum_diff_thresh) >>
+                         num_pels_log2_lookup[bs]) + 1;
+
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        const __m128i k_delta = _mm_set1_epi8(delta);
+        sig -= sig_stride * b_height;
+        mc_running_avg_y -= mc_avg_y_stride * b_height;
+        running_avg_y -= avg_y_stride * b_height;
+        sum_diff = 0;
+        for (r = 0; r < b_height; ++r) {
+          for (c = 0; c < b_width_shift4; ++c) {
+            acc_diff[c][r>>4] = vp9_denoiser_adj_16x1_sse2(
+                sig, mc_running_avg_y, running_avg_y, k_0,
+                k_delta, acc_diff[c][r>>4]);
+            // Update pointers for next iteration.
+            sig += 16;
+            mc_running_avg_y += 16;
+            running_avg_y += 16;
+          }
+
+          if ((r & 0xf) == 0xf || (bs == BLOCK_16X8 && r == 7)) {
+            for (c = 0; c < b_width_shift4; ++c) {
+              sum_diff += sum_diff_16x1(acc_diff[c][r>>4]);
+            }
+          }
+          sig = sig - b_width + sig_stride;
+          mc_running_avg_y = mc_running_avg_y - b_width + mc_avg_y_stride;
+          running_avg_y = running_avg_y - b_width + avg_y_stride;
+        }
+        if (abs(sum_diff) > sum_diff_thresh) {
+          return COPY_BLOCK;
+        }
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+  }
+  return FILTER_BLOCK;
+}
+
+int vp9_denoiser_filter_sse2(const uint8_t *sig, int sig_stride,
+                             const uint8_t *mc_avg,
+                             int mc_avg_stride,
+                             uint8_t *avg, int avg_stride,
+                             int increase_denoising,
+                             BLOCK_SIZE bs,
+                             int motion_magnitude) {
+  // Rank by frequency of the block type to have an early termination.
+  if (bs == BLOCK_16X16 || bs == BLOCK_32X32 || bs == BLOCK_64X64 ||
+      bs == BLOCK_16X32 || bs == BLOCK_16X8 || bs == BLOCK_32X16 ||
+      bs == BLOCK_32X64 || bs == BLOCK_64X32) {
+    return vp9_denoiser_NxM_sse2_big(sig, sig_stride,
+                                     mc_avg, mc_avg_stride,
+                                     avg, avg_stride,
+                                     increase_denoising,
+                                     bs, motion_magnitude);
+  } else if (bs == BLOCK_8X8 || bs == BLOCK_8X16) {
+    return vp9_denoiser_NxM_sse2_small(sig, sig_stride,
+                                       mc_avg, mc_avg_stride,
+                                       avg, avg_stride,
+                                       increase_denoising,
+                                       bs, motion_magnitude, 8);
+  } else {
+    return COPY_BLOCK;
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_diamond_search_sad_avx.c b/libvpx/libvpx/vp9/encoder/x86/vp9_diamond_search_sad_avx.c
new file mode 100644
index 0000000..cd3e87e
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_diamond_search_sad_avx.c
@@ -0,0 +1,314 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#if defined(_MSC_VER)
+# include <intrin.h>
+#endif
+#include <emmintrin.h>
+#include <smmintrin.h>
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vpx_ports/mem.h"
+
+#ifdef __GNUC__
+# define LIKELY(v)    __builtin_expect(v, 1)
+# define UNLIKELY(v)  __builtin_expect(v, 0)
+#else
+# define LIKELY(v)    (v)
+# define UNLIKELY(v)  (v)
+#endif
+
+static INLINE int_mv pack_int_mv(int16_t row, int16_t col) {
+  int_mv result;
+  result.as_mv.row = row;
+  result.as_mv.col = col;
+  return result;
+}
+
+static INLINE MV_JOINT_TYPE get_mv_joint(const int_mv mv) {
+  // This is simplified from the C implementation to utilise that
+  //  x->nmvjointsadcost[1] == x->nmvjointsadcost[2]  and
+  //  x->nmvjointsadcost[1] == x->nmvjointsadcost[3]
+  return mv.as_int == 0 ? 0 : 1;
+}
+
+static INLINE int mv_cost(const int_mv mv,
+                          const int *joint_cost, int *const comp_cost[2]) {
+  return joint_cost[get_mv_joint(mv)] +
+         comp_cost[0][mv.as_mv.row] + comp_cost[1][mv.as_mv.col];
+}
+
+static int mvsad_err_cost(const MACROBLOCK *x, const int_mv mv, const MV *ref,
+                          int sad_per_bit) {
+  const int_mv diff = pack_int_mv(mv.as_mv.row - ref->row,
+                                  mv.as_mv.col - ref->col);
+  return ROUND_POWER_OF_TWO((unsigned)mv_cost(diff, x->nmvjointsadcost,
+                                              x->nmvsadcost) *
+                                              sad_per_bit, VP9_PROB_COST_SHIFT);
+}
+
+/*****************************************************************************
+ * This function utilizes 3 properties of the cost function lookup tables,   *
+ * constructed in using 'cal_nmvjointsadcost' and 'cal_nmvsadcosts' in       *
+ * vp9_encoder.c.                                                            *
+ * For the joint cost:                                                       *
+ *   - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3]           *
+ * For the component costs:                                                  *
+ *   - For all i: mvsadcost[0][i] == mvsadcost[1][i]                         *
+ *         (Equal costs for both components)                                 *
+ *   - For all i: mvsadcost[0][i] == mvsadcost[0][-i]                        *
+ *         (Cost function is even)                                           *
+ * If these do not hold, then this function cannot be used without           *
+ * modification, in which case you can revert to using the C implementation, *
+ * which does not rely on these properties.                                  *
+ *****************************************************************************/
+int vp9_diamond_search_sad_avx(const MACROBLOCK *x,
+                               const search_site_config *cfg,
+                               MV *ref_mv, MV *best_mv, int search_param,
+                               int sad_per_bit, int *num00,
+                               const vp9_variance_fn_ptr_t *fn_ptr,
+                               const MV *center_mv) {
+  const int_mv maxmv = pack_int_mv(x->mv_row_max, x->mv_col_max);
+  const __m128i v_max_mv_w = _mm_set1_epi32(maxmv.as_int);
+  const int_mv minmv = pack_int_mv(x->mv_row_min, x->mv_col_min);
+  const __m128i v_min_mv_w = _mm_set1_epi32(minmv.as_int);
+
+  const __m128i v_spb_d = _mm_set1_epi32(sad_per_bit);
+
+  const __m128i v_joint_cost_0_d = _mm_set1_epi32(x->nmvjointsadcost[0]);
+  const __m128i v_joint_cost_1_d = _mm_set1_epi32(x->nmvjointsadcost[1]);
+
+  // search_param determines the length of the initial step and hence the number
+  // of iterations.
+  // 0 = initial step (MAX_FIRST_STEP) pel
+  // 1 = (MAX_FIRST_STEP/2) pel,
+  // 2 = (MAX_FIRST_STEP/4) pel...
+  const       MV *ss_mv = &cfg->ss_mv[cfg->searches_per_step * search_param];
+  const intptr_t *ss_os = &cfg->ss_os[cfg->searches_per_step * search_param];
+  const int tot_steps = cfg->total_steps - search_param;
+
+  const int_mv fcenter_mv = pack_int_mv(center_mv->row >> 3,
+                                        center_mv->col >> 3);
+  const __m128i vfcmv = _mm_set1_epi32(fcenter_mv.as_int);
+
+  const int ref_row = clamp(ref_mv->row, minmv.as_mv.row, maxmv.as_mv.row);
+  const int ref_col = clamp(ref_mv->col, minmv.as_mv.col, maxmv.as_mv.col);
+
+  int_mv bmv = pack_int_mv(ref_row, ref_col);
+  int_mv new_bmv = bmv;
+  __m128i v_bmv_w = _mm_set1_epi32(bmv.as_int);
+
+  const int what_stride = x->plane[0].src.stride;
+  const int in_what_stride = x->e_mbd.plane[0].pre[0].stride;
+  const uint8_t *const what = x->plane[0].src.buf;
+  const uint8_t *const in_what = x->e_mbd.plane[0].pre[0].buf +
+                                 ref_row * in_what_stride + ref_col;
+
+  // Work out the start point for the search
+  const uint8_t *best_address = in_what;
+  const uint8_t *new_best_address = best_address;
+#if ARCH_X86_64
+  __m128i v_ba_q = _mm_set1_epi64x((intptr_t)best_address);
+#else
+  __m128i v_ba_d = _mm_set1_epi32((intptr_t)best_address);
+#endif
+
+  unsigned int best_sad;
+  int i, j, step;
+
+  // Check the prerequisite cost function properties that are easy to check
+  // in an assert. See the function-level documentation for details on all
+  // prerequisites.
+  assert(x->nmvjointsadcost[1] == x->nmvjointsadcost[2]);
+  assert(x->nmvjointsadcost[1] == x->nmvjointsadcost[3]);
+
+  // Check the starting position
+  best_sad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride);
+  best_sad += mvsad_err_cost(x, bmv, &fcenter_mv.as_mv, sad_per_bit);
+
+  *num00 = 0;
+
+  for (i = 0, step = 0; step < tot_steps; step++) {
+    for (j = 0; j < cfg->searches_per_step; j += 4, i += 4) {
+      __m128i v_sad_d, v_cost_d, v_outside_d, v_inside_d, v_diff_mv_w;
+#if ARCH_X86_64
+      __m128i v_blocka[2];
+#else
+      __m128i v_blocka[1];
+#endif
+
+      // Compute the candidate motion vectors
+      const __m128i v_ss_mv_w = _mm_loadu_si128((const __m128i *)&ss_mv[i]);
+      const __m128i v_these_mv_w = _mm_add_epi16(v_bmv_w, v_ss_mv_w);
+      // Clamp them to the search bounds
+      __m128i v_these_mv_clamp_w = v_these_mv_w;
+      v_these_mv_clamp_w = _mm_min_epi16(v_these_mv_clamp_w, v_max_mv_w);
+      v_these_mv_clamp_w = _mm_max_epi16(v_these_mv_clamp_w, v_min_mv_w);
+      // The ones that did not change are inside the search area
+      v_inside_d = _mm_cmpeq_epi32(v_these_mv_clamp_w, v_these_mv_w);
+
+      // If none of them are inside, then move on
+      if (LIKELY(_mm_test_all_zeros(v_inside_d, v_inside_d))) {
+        continue;
+      }
+
+      // The inverse mask indicates which of the MVs are outside
+      v_outside_d = _mm_xor_si128(v_inside_d, _mm_set1_epi8(0xff));
+      // Shift right to keep the sign bit clear, we will use this later
+      // to set the cost to the maximum value.
+      v_outside_d = _mm_srli_epi32(v_outside_d, 1);
+
+      // Compute the difference MV
+      v_diff_mv_w = _mm_sub_epi16(v_these_mv_clamp_w, vfcmv);
+      // We utilise the fact that the cost function is even, and use the
+      // absolute difference. This allows us to use unsigned indexes later
+      // and reduces cache pressure somewhat as only a half of the table
+      // is ever referenced.
+      v_diff_mv_w = _mm_abs_epi16(v_diff_mv_w);
+
+      // Compute the SIMD pointer offsets.
+      {
+#if ARCH_X86_64  //  sizeof(intptr_t) == 8
+        // Load the offsets
+        __m128i v_bo10_q = _mm_loadu_si128((const __m128i *)&ss_os[i + 0]);
+        __m128i v_bo32_q = _mm_loadu_si128((const __m128i *)&ss_os[i + 2]);
+        // Set the ones falling outside to zero
+        v_bo10_q = _mm_and_si128(v_bo10_q,
+                                 _mm_cvtepi32_epi64(v_inside_d));
+        v_bo32_q = _mm_and_si128(v_bo32_q,
+                                 _mm_unpackhi_epi32(v_inside_d, v_inside_d));
+        // Compute the candidate addresses
+        v_blocka[0] = _mm_add_epi64(v_ba_q, v_bo10_q);
+        v_blocka[1] = _mm_add_epi64(v_ba_q, v_bo32_q);
+#else  // ARCH_X86 //  sizeof(intptr_t) == 4
+        __m128i v_bo_d = _mm_loadu_si128((const __m128i *)&ss_os[i]);
+        v_bo_d = _mm_and_si128(v_bo_d, v_inside_d);
+        v_blocka[0] = _mm_add_epi32(v_ba_d, v_bo_d);
+#endif
+      }
+
+      fn_ptr->sdx4df(what, what_stride,
+                     (const uint8_t **)&v_blocka[0], in_what_stride,
+                     (uint32_t*)&v_sad_d);
+
+      // Look up the component cost of the residual motion vector
+      {
+        const int32_t row0 = _mm_extract_epi16(v_diff_mv_w, 0);
+        const int32_t col0 = _mm_extract_epi16(v_diff_mv_w, 1);
+        const int32_t row1 = _mm_extract_epi16(v_diff_mv_w, 2);
+        const int32_t col1 = _mm_extract_epi16(v_diff_mv_w, 3);
+        const int32_t row2 = _mm_extract_epi16(v_diff_mv_w, 4);
+        const int32_t col2 = _mm_extract_epi16(v_diff_mv_w, 5);
+        const int32_t row3 = _mm_extract_epi16(v_diff_mv_w, 6);
+        const int32_t col3 = _mm_extract_epi16(v_diff_mv_w, 7);
+
+        // Note: This is a use case for vpgather in AVX2
+        const uint32_t cost0 = x->nmvsadcost[0][row0] + x->nmvsadcost[0][col0];
+        const uint32_t cost1 = x->nmvsadcost[0][row1] + x->nmvsadcost[0][col1];
+        const uint32_t cost2 = x->nmvsadcost[0][row2] + x->nmvsadcost[0][col2];
+        const uint32_t cost3 = x->nmvsadcost[0][row3] + x->nmvsadcost[0][col3];
+
+        __m128i v_cost_10_d, v_cost_32_d;
+        v_cost_10_d = _mm_cvtsi32_si128(cost0);
+        v_cost_10_d = _mm_insert_epi32(v_cost_10_d, cost1, 1);
+        v_cost_32_d = _mm_cvtsi32_si128(cost2);
+        v_cost_32_d = _mm_insert_epi32(v_cost_32_d, cost3, 1);
+        v_cost_d = _mm_unpacklo_epi64(v_cost_10_d, v_cost_32_d);
+      }
+
+      // Now add in the joint cost
+      {
+        const __m128i v_sel_d = _mm_cmpeq_epi32(v_diff_mv_w,
+                                                _mm_setzero_si128());
+        const __m128i v_joint_cost_d = _mm_blendv_epi8(v_joint_cost_1_d,
+                                                       v_joint_cost_0_d,
+                                                       v_sel_d);
+        v_cost_d = _mm_add_epi32(v_cost_d, v_joint_cost_d);
+      }
+
+      // Multiply by sad_per_bit
+      v_cost_d = _mm_mullo_epi32(v_cost_d, v_spb_d);
+      // ROUND_POWER_OF_TWO(v_cost_d, VP9_PROB_COST_SHIFT)
+      v_cost_d = _mm_add_epi32(v_cost_d,
+                               _mm_set1_epi32(1 << (VP9_PROB_COST_SHIFT - 1)));
+      v_cost_d = _mm_srai_epi32(v_cost_d, VP9_PROB_COST_SHIFT);
+      // Add the cost to the sad
+      v_sad_d = _mm_add_epi32(v_sad_d, v_cost_d);
+
+      // Make the motion vectors outside the search area have max cost
+      // by or'ing in the comparison mask, this way the minimum search won't
+      // pick them.
+      v_sad_d = _mm_or_si128(v_sad_d, v_outside_d);
+
+      // Find the minimum value and index horizontally in v_sad_d
+      {
+        // Try speculatively on 16 bits, so we can use the minpos intrinsic
+        const __m128i v_sad_w = _mm_packus_epi32(v_sad_d, v_sad_d);
+        const __m128i v_minp_w = _mm_minpos_epu16(v_sad_w);
+
+        uint32_t local_best_sad = _mm_extract_epi16(v_minp_w, 0);
+        uint32_t local_best_idx = _mm_extract_epi16(v_minp_w, 1);
+
+        // If the local best value is not saturated, just use it, otherwise
+        // find the horizontal minimum again the hard way on 32 bits.
+        // This is executed rarely.
+        if (UNLIKELY(local_best_sad == 0xffff)) {
+          __m128i v_loval_d, v_hival_d, v_loidx_d, v_hiidx_d, v_sel_d;
+
+          v_loval_d = v_sad_d;
+          v_loidx_d = _mm_set_epi32(3, 2, 1, 0);
+          v_hival_d = _mm_srli_si128(v_loval_d, 8);
+          v_hiidx_d = _mm_srli_si128(v_loidx_d, 8);
+
+          v_sel_d = _mm_cmplt_epi32(v_hival_d, v_loval_d);
+
+          v_loval_d = _mm_blendv_epi8(v_loval_d, v_hival_d, v_sel_d);
+          v_loidx_d = _mm_blendv_epi8(v_loidx_d, v_hiidx_d, v_sel_d);
+          v_hival_d = _mm_srli_si128(v_loval_d, 4);
+          v_hiidx_d = _mm_srli_si128(v_loidx_d, 4);
+
+          v_sel_d = _mm_cmplt_epi32(v_hival_d, v_loval_d);
+
+          v_loval_d = _mm_blendv_epi8(v_loval_d, v_hival_d, v_sel_d);
+          v_loidx_d = _mm_blendv_epi8(v_loidx_d, v_hiidx_d, v_sel_d);
+
+          local_best_sad = _mm_extract_epi32(v_loval_d, 0);
+          local_best_idx = _mm_extract_epi32(v_loidx_d, 0);
+        }
+
+        // Update the global minimum if the local minimum is smaller
+        if (LIKELY(local_best_sad < best_sad)) {
+          new_bmv = ((const int_mv *)&v_these_mv_w)[local_best_idx];
+          new_best_address = ((const uint8_t **)v_blocka)[local_best_idx];
+
+          best_sad = local_best_sad;
+        }
+      }
+    }
+
+    bmv = new_bmv;
+    best_address = new_best_address;
+
+    v_bmv_w = _mm_set1_epi32(bmv.as_int);
+#if ARCH_X86_64
+    v_ba_q = _mm_set1_epi64x((intptr_t)best_address);
+#else
+    v_ba_d = _mm_set1_epi32((intptr_t)best_address);
+#endif
+
+    if (UNLIKELY(best_address == in_what)) {
+      (*num00)++;
+    }
+  }
+
+  *best_mv = bmv.as_mv;
+  return best_sad;
+}
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_error_intrin_avx2.c b/libvpx/libvpx/vp9/encoder/x86/vp9_error_intrin_avx2.c
new file mode 100644
index 0000000..dfebaab
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_error_intrin_avx2.c
@@ -0,0 +1,73 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Usee of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "./vp9_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+int64_t vp9_block_error_avx2(const int16_t *coeff,
+                             const int16_t *dqcoeff,
+                             intptr_t block_size,
+                             int64_t *ssz) {
+  __m256i sse_reg, ssz_reg, coeff_reg, dqcoeff_reg;
+  __m256i exp_dqcoeff_lo, exp_dqcoeff_hi, exp_coeff_lo, exp_coeff_hi;
+  __m256i sse_reg_64hi, ssz_reg_64hi;
+  __m128i sse_reg128, ssz_reg128;
+  int64_t sse;
+  int i;
+  const __m256i zero_reg = _mm256_set1_epi16(0);
+
+  // init sse and ssz registerd to zero
+  sse_reg = _mm256_set1_epi16(0);
+  ssz_reg = _mm256_set1_epi16(0);
+
+  for (i = 0 ; i < block_size ; i+= 16) {
+    // load 32 bytes from coeff and dqcoeff
+    coeff_reg = _mm256_loadu_si256((const __m256i *)(coeff + i));
+    dqcoeff_reg = _mm256_loadu_si256((const __m256i *)(dqcoeff + i));
+    // dqcoeff - coeff
+    dqcoeff_reg = _mm256_sub_epi16(dqcoeff_reg, coeff_reg);
+    // madd (dqcoeff - coeff)
+    dqcoeff_reg = _mm256_madd_epi16(dqcoeff_reg, dqcoeff_reg);
+    // madd coeff
+    coeff_reg = _mm256_madd_epi16(coeff_reg, coeff_reg);
+    // expand each double word of madd (dqcoeff - coeff) to quad word
+    exp_dqcoeff_lo = _mm256_unpacklo_epi32(dqcoeff_reg, zero_reg);
+    exp_dqcoeff_hi = _mm256_unpackhi_epi32(dqcoeff_reg, zero_reg);
+    // expand each double word of madd (coeff) to quad word
+    exp_coeff_lo = _mm256_unpacklo_epi32(coeff_reg, zero_reg);
+    exp_coeff_hi = _mm256_unpackhi_epi32(coeff_reg, zero_reg);
+    // add each quad word of madd (dqcoeff - coeff) and madd (coeff)
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_lo);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_lo);
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_hi);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_hi);
+  }
+  // save the higher 64 bit of each 128 bit lane
+  sse_reg_64hi = _mm256_srli_si256(sse_reg, 8);
+  ssz_reg_64hi = _mm256_srli_si256(ssz_reg, 8);
+  // add the higher 64 bit to the low 64 bit
+  sse_reg = _mm256_add_epi64(sse_reg, sse_reg_64hi);
+  ssz_reg = _mm256_add_epi64(ssz_reg, ssz_reg_64hi);
+
+  // add each 64 bit from each of the 128 bit lane of the 256 bit
+  sse_reg128 = _mm_add_epi64(_mm256_castsi256_si128(sse_reg),
+                             _mm256_extractf128_si256(sse_reg, 1));
+
+  ssz_reg128 = _mm_add_epi64(_mm256_castsi256_si128(ssz_reg),
+                             _mm256_extractf128_si256(ssz_reg, 1));
+
+  // store the results
+  _mm_storel_epi64((__m128i*)(&sse), sse_reg128);
+
+  _mm_storel_epi64((__m128i*)(ssz), ssz_reg128);
+  return sse;
+}
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_error_sse2.asm b/libvpx/libvpx/vp9/encoder/x86/vp9_error_sse2.asm
new file mode 100644
index 0000000..5b02382
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_error_sse2.asm
@@ -0,0 +1,122 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%define private_prefix vp9
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; int64_t vp9_block_error(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size,
+;                         int64_t *ssz)
+
+INIT_XMM sse2
+cglobal block_error, 3, 3, 8, uqc, dqc, size, ssz
+  pxor      m4, m4                 ; sse accumulator
+  pxor      m6, m6                 ; ssz accumulator
+  pxor      m5, m5                 ; dedicated zero register
+  lea     uqcq, [uqcq+sizeq*2]
+  lea     dqcq, [dqcq+sizeq*2]
+  neg    sizeq
+.loop:
+  mova      m2, [uqcq+sizeq*2]
+  mova      m0, [dqcq+sizeq*2]
+  mova      m3, [uqcq+sizeq*2+mmsize]
+  mova      m1, [dqcq+sizeq*2+mmsize]
+  psubw     m0, m2
+  psubw     m1, m3
+  ; individual errors are max. 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit)
+  pmaddwd   m0, m0
+  pmaddwd   m1, m1
+  pmaddwd   m2, m2
+  pmaddwd   m3, m3
+  ; accumulate in 64bit
+  punpckldq m7, m0, m5
+  punpckhdq m0, m5
+  paddq     m4, m7
+  punpckldq m7, m1, m5
+  paddq     m4, m0
+  punpckhdq m1, m5
+  paddq     m4, m7
+  punpckldq m7, m2, m5
+  paddq     m4, m1
+  punpckhdq m2, m5
+  paddq     m6, m7
+  punpckldq m7, m3, m5
+  paddq     m6, m2
+  punpckhdq m3, m5
+  paddq     m6, m7
+  paddq     m6, m3
+  add    sizeq, mmsize
+  jl .loop
+
+  ; accumulate horizontally and store in return value
+  movhlps   m5, m4
+  movhlps   m7, m6
+  paddq     m4, m5
+  paddq     m6, m7
+%if ARCH_X86_64
+  movq    rax, m4
+  movq [sszq], m6
+%else
+  mov     eax, sszm
+  pshufd   m5, m4, 0x1
+  movq  [eax], m6
+  movd    eax, m4
+  movd    edx, m5
+%endif
+  RET
+
+; Compute the sum of squared difference between two int16_t vectors.
+; int64_t vp9_block_error_fp(int16_t *coeff, int16_t *dqcoeff,
+;                            intptr_t block_size)
+
+INIT_XMM sse2
+cglobal block_error_fp, 3, 3, 6, uqc, dqc, size
+  pxor      m4, m4                 ; sse accumulator
+  pxor      m5, m5                 ; dedicated zero register
+  lea     uqcq, [uqcq+sizeq*2]
+  lea     dqcq, [dqcq+sizeq*2]
+  neg    sizeq
+.loop:
+  mova      m2, [uqcq+sizeq*2]
+  mova      m0, [dqcq+sizeq*2]
+  mova      m3, [uqcq+sizeq*2+mmsize]
+  mova      m1, [dqcq+sizeq*2+mmsize]
+  psubw     m0, m2
+  psubw     m1, m3
+  ; individual errors are max. 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit)
+  pmaddwd   m0, m0
+  pmaddwd   m1, m1
+  ; accumulate in 64bit
+  punpckldq m3, m0, m5
+  punpckhdq m0, m5
+  paddq     m4, m3
+  punpckldq m3, m1, m5
+  paddq     m4, m0
+  punpckhdq m1, m5
+  paddq     m4, m3
+  paddq     m4, m1
+  add    sizeq, mmsize
+  jl .loop
+
+  ; accumulate horizontally and store in return value
+  movhlps   m5, m4
+  paddq     m4, m5
+%if ARCH_X86_64
+  movq    rax, m4
+%else
+  pshufd   m5, m4, 0x1
+  movd    eax, m4
+  movd    edx, m5
+%endif
+  RET
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_frame_scale_ssse3.c b/libvpx/libvpx/vp9/encoder/x86/vp9_frame_scale_ssse3.c
new file mode 100644
index 0000000..38af3b1
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_frame_scale_ssse3.c
@@ -0,0 +1,211 @@
+/*
+ *  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#if defined(_MSC_VER) && _MSC_VER <= 1500
+// Need to include math.h before calling tmmintrin.h/intrin.h
+// in certain versions of MSVS.
+#include <math.h>
+#endif
+#include <tmmintrin.h>  // SSSE3
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+#include "vpx_scale/yv12config.h"
+
+extern void vp9_scale_and_extend_frame_c(const YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst);
+
+static void downsample_2_to_1_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    int w, int h) {
+  const __m128i mask = _mm_set1_epi16(0x00FF);
+  const int max_width = w & ~15;
+  int y;
+  for (y = 0; y < h; ++y) {
+    int x;
+    for (x = 0; x < max_width; x += 16) {
+      const __m128i a = _mm_loadu_si128((const __m128i *)(src + x * 2 +  0));
+      const __m128i b = _mm_loadu_si128((const __m128i *)(src + x * 2 + 16));
+      const __m128i a_and = _mm_and_si128(a, mask);
+      const __m128i b_and = _mm_and_si128(b, mask);
+      const __m128i c = _mm_packus_epi16(a_and, b_and);
+      _mm_storeu_si128((__m128i *)(dst + x), c);
+    }
+    for (; x < w; ++x)
+      dst[x] = src[x * 2];
+    src += src_stride * 2;
+    dst += dst_stride;
+  }
+}
+
+static INLINE __m128i filter(const __m128i *const a, const __m128i *const b,
+                             const __m128i *const c, const __m128i *const d,
+                             const __m128i *const e, const __m128i *const f,
+                             const __m128i *const g, const __m128i *const h) {
+  const __m128i coeffs_ab =
+      _mm_set_epi8(6, -1, 6, -1, 6, -1, 6, -1, 6, -1, 6, -1, 6, -1, 6, -1);
+  const __m128i coeffs_cd =
+      _mm_set_epi8(78, -19, 78, -19, 78, -19, 78, -19, 78, -19, 78, -19,
+                   78, -19, 78, -19);
+  const __m128i const64_x16 = _mm_set1_epi16(64);
+  const __m128i ab = _mm_unpacklo_epi8(*a, *b);
+  const __m128i cd = _mm_unpacklo_epi8(*c, *d);
+  const __m128i fe = _mm_unpacklo_epi8(*f, *e);
+  const __m128i hg = _mm_unpacklo_epi8(*h, *g);
+  const __m128i ab_terms = _mm_maddubs_epi16(ab, coeffs_ab);
+  const __m128i cd_terms = _mm_maddubs_epi16(cd, coeffs_cd);
+  const __m128i fe_terms = _mm_maddubs_epi16(fe, coeffs_cd);
+  const __m128i hg_terms = _mm_maddubs_epi16(hg, coeffs_ab);
+  // can not overflow
+  const __m128i abcd_terms = _mm_add_epi16(ab_terms, cd_terms);
+  // can not overflow
+  const __m128i fehg_terms = _mm_add_epi16(fe_terms, hg_terms);
+  // can overflow, use saturating add
+  const __m128i terms = _mm_adds_epi16(abcd_terms, fehg_terms);
+  const __m128i round = _mm_adds_epi16(terms, const64_x16);
+  const __m128i shift = _mm_srai_epi16(round, 7);
+  return _mm_packus_epi16(shift, shift);
+}
+
+static void eight_tap_row_ssse3(const uint8_t *src, uint8_t *dst, int w) {
+  const int max_width = w & ~7;
+  int x = 0;
+  for (; x < max_width; x += 8) {
+    const __m128i a = _mm_loadl_epi64((const __m128i *)(src + x + 0));
+    const __m128i b = _mm_loadl_epi64((const __m128i *)(src + x + 1));
+    const __m128i c = _mm_loadl_epi64((const __m128i *)(src + x + 2));
+    const __m128i d = _mm_loadl_epi64((const __m128i *)(src + x + 3));
+    const __m128i e = _mm_loadl_epi64((const __m128i *)(src + x + 4));
+    const __m128i f = _mm_loadl_epi64((const __m128i *)(src + x + 5));
+    const __m128i g = _mm_loadl_epi64((const __m128i *)(src + x + 6));
+    const __m128i h = _mm_loadl_epi64((const __m128i *)(src + x + 7));
+    const __m128i pack = filter(&a, &b, &c, &d, &e, &f, &g, &h);
+    _mm_storel_epi64((__m128i *)(dst + x), pack);
+  }
+}
+
+static void upsample_1_to_2_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  int dst_w, int dst_h) {
+  dst_w /= 2;
+  dst_h /= 2;
+  {
+    DECLARE_ALIGNED(16, uint8_t, tmp[1920 * 8]);
+    uint8_t *tmp0 = tmp + dst_w * 0;
+    uint8_t *tmp1 = tmp + dst_w * 1;
+    uint8_t *tmp2 = tmp + dst_w * 2;
+    uint8_t *tmp3 = tmp + dst_w * 3;
+    uint8_t *tmp4 = tmp + dst_w * 4;
+    uint8_t *tmp5 = tmp + dst_w * 5;
+    uint8_t *tmp6 = tmp + dst_w * 6;
+    uint8_t *tmp7 = tmp + dst_w * 7;
+    uint8_t *tmp8 = NULL;
+    const int max_width = dst_w & ~7;
+    int y;
+    eight_tap_row_ssse3(src - src_stride * 3 - 3, tmp0, dst_w);
+    eight_tap_row_ssse3(src - src_stride * 2 - 3, tmp1, dst_w);
+    eight_tap_row_ssse3(src - src_stride * 1 - 3, tmp2, dst_w);
+    eight_tap_row_ssse3(src + src_stride * 0 - 3, tmp3, dst_w);
+    eight_tap_row_ssse3(src + src_stride * 1 - 3, tmp4, dst_w);
+    eight_tap_row_ssse3(src + src_stride * 2 - 3, tmp5, dst_w);
+    eight_tap_row_ssse3(src + src_stride * 3 - 3, tmp6, dst_w);
+    for (y = 0; y < dst_h; y++) {
+      int x;
+      eight_tap_row_ssse3(src + src_stride * 4 - 3, tmp7, dst_w);
+      for (x = 0; x < max_width; x += 8) {
+        const __m128i A = _mm_loadl_epi64((const __m128i *)(src  + x));
+        const __m128i B = _mm_loadl_epi64((const __m128i *)(tmp3 + x));
+        const __m128i AB = _mm_unpacklo_epi8(A, B);
+        __m128i C, D, CD;
+        _mm_storeu_si128((__m128i *)(dst + x * 2), AB);
+        {
+          const __m128i a =
+              _mm_loadl_epi64((const __m128i *)(src + x - src_stride * 3));
+          const __m128i b =
+              _mm_loadl_epi64((const __m128i *)(src + x - src_stride * 2));
+          const __m128i c =
+              _mm_loadl_epi64((const __m128i *)(src + x - src_stride * 1));
+          const __m128i d =
+              _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 0));
+          const __m128i e =
+              _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 1));
+          const __m128i f =
+              _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 2));
+          const __m128i g =
+              _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 3));
+          const __m128i h =
+              _mm_loadl_epi64((const __m128i *)(src + x + src_stride * 4));
+          C = filter(&a, &b, &c, &d, &e, &f, &g, &h);
+        }
+        {
+          const __m128i a = _mm_loadl_epi64((const __m128i *)(tmp0 + x));
+          const __m128i b = _mm_loadl_epi64((const __m128i *)(tmp1 + x));
+          const __m128i c = _mm_loadl_epi64((const __m128i *)(tmp2 + x));
+          const __m128i d = _mm_loadl_epi64((const __m128i *)(tmp3 + x));
+          const __m128i e = _mm_loadl_epi64((const __m128i *)(tmp4 + x));
+          const __m128i f = _mm_loadl_epi64((const __m128i *)(tmp5 + x));
+          const __m128i g = _mm_loadl_epi64((const __m128i *)(tmp6 + x));
+          const __m128i h = _mm_loadl_epi64((const __m128i *)(tmp7 + x));
+          D = filter(&a, &b, &c, &d, &e, &f, &g, &h);
+        }
+        CD = _mm_unpacklo_epi8(C, D);
+        _mm_storeu_si128((__m128i *)(dst + x * 2 + dst_stride), CD);
+      }
+      src += src_stride;
+      dst += dst_stride * 2;
+      tmp8 = tmp0;
+      tmp0 = tmp1;
+      tmp1 = tmp2;
+      tmp2 = tmp3;
+      tmp3 = tmp4;
+      tmp4 = tmp5;
+      tmp5 = tmp6;
+      tmp6 = tmp7;
+      tmp7 = tmp8;
+    }
+  }
+}
+
+void vp9_scale_and_extend_frame_ssse3(const YV12_BUFFER_CONFIG *src,
+                                      YV12_BUFFER_CONFIG *dst) {
+  const int src_w = src->y_crop_width;
+  const int src_h = src->y_crop_height;
+  const int dst_w = dst->y_crop_width;
+  const int dst_h = dst->y_crop_height;
+  const int dst_uv_w = dst_w / 2;
+  const int dst_uv_h = dst_h / 2;
+
+  if (dst_w * 2 == src_w && dst_h * 2 == src_h) {
+    downsample_2_to_1_ssse3(src->y_buffer, src->y_stride,
+                            dst->y_buffer, dst->y_stride, dst_w, dst_h);
+    downsample_2_to_1_ssse3(src->u_buffer, src->uv_stride,
+                            dst->u_buffer, dst->uv_stride, dst_uv_w, dst_uv_h);
+    downsample_2_to_1_ssse3(src->v_buffer, src->uv_stride,
+                            dst->v_buffer, dst->uv_stride, dst_uv_w, dst_uv_h);
+    vpx_extend_frame_borders(dst);
+  } else if (dst_w == src_w * 2 && dst_h == src_h * 2) {
+    // The upsample() supports widths up to 1920 * 2.  If greater, fall back
+    // to vp9_scale_and_extend_frame_c().
+    if (dst_w/2 <= 1920) {
+      upsample_1_to_2_ssse3(src->y_buffer, src->y_stride,
+                            dst->y_buffer, dst->y_stride, dst_w, dst_h);
+      upsample_1_to_2_ssse3(src->u_buffer, src->uv_stride,
+                            dst->u_buffer, dst->uv_stride, dst_uv_w, dst_uv_h);
+      upsample_1_to_2_ssse3(src->v_buffer, src->uv_stride,
+                            dst->v_buffer, dst->uv_stride, dst_uv_w, dst_uv_h);
+      vpx_extend_frame_borders(dst);
+    } else {
+      vp9_scale_and_extend_frame_c(src, dst);
+    }
+  } else {
+    vp9_scale_and_extend_frame_c(src, dst);
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_highbd_block_error_intrin_sse2.c b/libvpx/libvpx/vp9/encoder/x86/vp9_highbd_block_error_intrin_sse2.c
new file mode 100644
index 0000000..c245cca
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_highbd_block_error_intrin_sse2.c
@@ -0,0 +1,71 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+#include <stdio.h>
+
+#include "vp9/common/vp9_common.h"
+
+int64_t vp9_highbd_block_error_sse2(tran_low_t *coeff, tran_low_t *dqcoeff,
+                                    intptr_t block_size, int64_t *ssz,
+                                    int bps) {
+  int i, j, test;
+  uint32_t temp[4];
+  __m128i max, min, cmp0, cmp1, cmp2, cmp3;
+  int64_t error = 0, sqcoeff = 0;
+  const int shift = 2 * (bps - 8);
+  const int rounding = shift > 0 ? 1 << (shift - 1) : 0;
+
+  for (i = 0; i < block_size; i+=8) {
+    // Load the data into xmm registers
+    __m128i mm_coeff = _mm_load_si128((__m128i*) (coeff + i));
+    __m128i mm_coeff2 = _mm_load_si128((__m128i*) (coeff + i + 4));
+    __m128i mm_dqcoeff = _mm_load_si128((__m128i*) (dqcoeff + i));
+    __m128i mm_dqcoeff2 = _mm_load_si128((__m128i*) (dqcoeff + i + 4));
+    // Check if any values require more than 15 bit
+    max = _mm_set1_epi32(0x3fff);
+    min = _mm_set1_epi32(0xffffc000);
+    cmp0 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff, max),
+            _mm_cmplt_epi32(mm_coeff, min));
+    cmp1 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff2, max),
+            _mm_cmplt_epi32(mm_coeff2, min));
+    cmp2 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff, max),
+            _mm_cmplt_epi32(mm_dqcoeff, min));
+    cmp3 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff2, max),
+            _mm_cmplt_epi32(mm_dqcoeff2, min));
+    test = _mm_movemask_epi8(_mm_or_si128(_mm_or_si128(cmp0, cmp1),
+            _mm_or_si128(cmp2, cmp3)));
+
+    if (!test) {
+      __m128i mm_diff, error_sse2, sqcoeff_sse2;;
+      mm_coeff = _mm_packs_epi32(mm_coeff, mm_coeff2);
+      mm_dqcoeff = _mm_packs_epi32(mm_dqcoeff, mm_dqcoeff2);
+      mm_diff = _mm_sub_epi16(mm_coeff, mm_dqcoeff);
+      error_sse2 = _mm_madd_epi16(mm_diff, mm_diff);
+      sqcoeff_sse2 = _mm_madd_epi16(mm_coeff, mm_coeff);
+      _mm_storeu_si128((__m128i*)temp, error_sse2);
+      error = error + temp[0] + temp[1] + temp[2] + temp[3];
+      _mm_storeu_si128((__m128i*)temp, sqcoeff_sse2);
+      sqcoeff += temp[0] + temp[1] + temp[2] + temp[3];
+    } else {
+      for (j = 0; j < 8; j++) {
+        const int64_t diff = coeff[i + j] - dqcoeff[i + j];
+        error +=  diff * diff;
+        sqcoeff += (int64_t)coeff[i + j] * (int64_t)coeff[i + j];
+      }
+    }
+  }
+  assert(error >= 0 && sqcoeff >= 0);
+  error = (error + rounding) >> shift;
+  sqcoeff = (sqcoeff + rounding) >> shift;
+
+  *ssz = sqcoeff;
+  return error;
+}
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_highbd_error_avx.asm b/libvpx/libvpx/vp9/encoder/x86/vp9_highbd_error_avx.asm
new file mode 100644
index 0000000..e476323
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_highbd_error_avx.asm
@@ -0,0 +1,261 @@
+;
+;  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%define private_prefix vp9
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+ALIGN 16
+
+;
+; int64_t vp9_highbd_block_error_8bit(int32_t *coeff, int32_t *dqcoeff,
+;                                     intptr_t block_size, int64_t *ssz)
+;
+
+INIT_XMM avx
+cglobal highbd_block_error_8bit, 4, 5, 8, uqc, dqc, size, ssz
+  vzeroupper
+
+  ; If only one iteration is required, then handle this as a special case.
+  ; It is the most frequent case, so we can have a significant gain here
+  ; by not setting up a loop and accumulators.
+  cmp    sizeq, 16
+  jne   .generic
+
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+  ;; Common case of size == 16
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+  ; Load input vectors
+  mova      xm0, [dqcq]
+  packssdw  xm0, [dqcq+16]
+  mova      xm2, [uqcq]
+  packssdw  xm2, [uqcq+16]
+
+  mova      xm1, [dqcq+32]
+  packssdw  xm1, [dqcq+48]
+  mova      xm3, [uqcq+32]
+  packssdw  xm3, [uqcq+48]
+
+  ; Compute the errors.
+  psubw     xm0, xm2
+  psubw     xm1, xm3
+
+  ; Individual errors are max 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit).
+  pmaddwd   xm2, xm2
+  pmaddwd   xm3, xm3
+
+  pmaddwd   xm0, xm0
+  pmaddwd   xm1, xm1
+
+  ; Squares are always positive, so we can use unsigned arithmetic after
+  ; squaring. As mentioned earlier 2 sums fit in 31 bits, so 4 sums will
+  ; fit in 32bits
+  paddd     xm2, xm3
+  paddd     xm0, xm1
+
+  ; Accumulate horizontally in 64 bits, there is no chance of overflow here
+  pxor      xm5, xm5
+
+  pblendw   xm3, xm5, xm2, 0x33 ; Zero extended  low of a pair of 32 bits
+  psrlq     xm2, 32             ; Zero extended high of a pair of 32 bits
+
+  pblendw   xm1, xm5, xm0, 0x33 ; Zero extended  low of a pair of 32 bits
+  psrlq     xm0, 32             ; Zero extended high of a pair of 32 bits
+
+  paddq     xm2, xm3
+  paddq     xm0, xm1
+
+  psrldq    xm3, xm2, 8
+  psrldq    xm1, xm0, 8
+
+  paddq     xm2, xm3
+  paddq     xm0, xm1
+
+  ; Store the return value
+%if ARCH_X86_64
+  movq      rax, xm0
+  movq   [sszq], xm2
+%else
+  movd      eax, xm0
+  pextrd    edx, xm0, 1
+  movq   [sszd], xm2
+%endif
+  RET
+
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+  ;; Generic case of size != 16, speculative low precision
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+  ALIGN 16
+.generic:
+  pxor      xm4, xm4                ; sse accumulator
+  pxor      xm5, xm5                ; overflow detection register for xm4
+  pxor      xm6, xm6                ; ssz accumulator
+  pxor      xm7, xm7                ; overflow detection register for xm6
+  lea      uqcq, [uqcq+sizeq*4]
+  lea      dqcq, [dqcq+sizeq*4]
+  neg     sizeq
+
+  ; Push the negative size as the high precision code might need it
+  push    sizeq
+
+.loop:
+  ; Load input vectors
+  mova      xm0, [dqcq+sizeq*4]
+  packssdw  xm0, [dqcq+sizeq*4+16]
+  mova      xm2, [uqcq+sizeq*4]
+  packssdw  xm2, [uqcq+sizeq*4+16]
+
+  mova      xm1, [dqcq+sizeq*4+32]
+  packssdw  xm1, [dqcq+sizeq*4+48]
+  mova      xm3, [uqcq+sizeq*4+32]
+  packssdw  xm3, [uqcq+sizeq*4+48]
+
+  add     sizeq, 16
+
+  ; Compute the squared errors.
+  ; Individual errors are max 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit).
+  psubw     xm0, xm2
+  pmaddwd   xm2, xm2
+  pmaddwd   xm0, xm0
+
+  psubw     xm1, xm3
+  pmaddwd   xm3, xm3
+  pmaddwd   xm1, xm1
+
+  ; Squares are always positive, so we can use unsigned arithmetic after
+  ; squaring. As mentioned earlier 2 sums fit in 31 bits, so 4 sums will
+  ; fit in 32bits
+  paddd     xm2, xm3
+  paddd     xm0, xm1
+
+  ; We accumulate using 32 bit arithmetic, but detect potential overflow
+  ; by checking if the MSB of the accumulators have ever been a set bit.
+  ; If yes, we redo the whole compute at the end on higher precision, but
+  ; this happens extremely rarely, so we still achieve a net gain.
+  paddd     xm4, xm0
+  paddd     xm6, xm2
+  por       xm5, xm4  ; OR in the accumulator for overflow detection
+  por       xm7, xm6  ; OR in the accumulator for overflow detection
+
+  jnz .loop
+
+  ; Add pairs horizontally (still only on 32 bits)
+  phaddd    xm4, xm4
+  por       xm5, xm4  ; OR in the accumulator for overflow detection
+  phaddd    xm6, xm6
+  por       xm7, xm6  ; OR in the accumulator for overflow detection
+
+  ; Check for possibility of overflow by testing if bit 32 of each dword lane
+  ; have ever been set. If they were not, then there was no overflow and the
+  ; final sum will fit in 32 bits. If overflow happened, then
+  ; we redo the whole computation on higher precision.
+  por       xm7, xm5
+  pmovmskb   r4, xm7
+  test       r4, 0x8888
+  jnz .highprec
+
+  phaddd    xm4, xm4
+  phaddd    xm6, xm6
+  pmovzxdq  xm4, xm4
+  pmovzxdq  xm6, xm6
+
+  ; Restore stack
+  pop     sizeq
+
+  ; Store the return value
+%if ARCH_X86_64
+  movq      rax, xm4
+  movq   [sszq], xm6
+%else
+  movd      eax, xm4
+  pextrd    edx, xm4, 1
+  movq   [sszd], xm6
+%endif
+  RET
+
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+  ;; Generic case of size != 16, high precision case
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+.highprec:
+  pxor      xm4, xm4                 ; sse accumulator
+  pxor      xm5, xm5                 ; dedicated zero register
+  pxor      xm6, xm6                 ; ssz accumulator
+  pop     sizeq
+
+.loophp:
+  mova      xm0, [dqcq+sizeq*4]
+  packssdw  xm0, [dqcq+sizeq*4+16]
+  mova      xm2, [uqcq+sizeq*4]
+  packssdw  xm2, [uqcq+sizeq*4+16]
+
+  mova      xm1, [dqcq+sizeq*4+32]
+  packssdw  xm1, [dqcq+sizeq*4+48]
+  mova      xm3, [uqcq+sizeq*4+32]
+  packssdw  xm3, [uqcq+sizeq*4+48]
+
+  add     sizeq, 16
+
+  ; individual errors are max. 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit)
+
+  psubw     xm0, xm2
+  pmaddwd   xm2, xm2
+  pmaddwd   xm0, xm0
+
+  psubw     xm1, xm3
+  pmaddwd   xm3, xm3
+  pmaddwd   xm1, xm1
+
+  ; accumulate in 64bit
+  punpckldq xm7, xm0, xm5
+  punpckhdq xm0, xm5
+  paddq     xm4, xm7
+
+  punpckldq xm7, xm2, xm5
+  punpckhdq xm2, xm5
+  paddq     xm6, xm7
+
+  punpckldq xm7, xm1, xm5
+  punpckhdq xm1, xm5
+  paddq     xm4, xm7
+
+  punpckldq xm7, xm3, xm5
+  punpckhdq xm3, xm5
+  paddq     xm6, xm7
+
+  paddq     xm4, xm0
+  paddq     xm4, xm1
+  paddq     xm6, xm2
+  paddq     xm6, xm3
+
+  jnz .loophp
+
+  ; Accumulate horizontally
+  movhlps   xm5, xm4
+  movhlps   xm7, xm6
+  paddq     xm4, xm5
+  paddq     xm6, xm7
+
+  ; Store the return value
+%if ARCH_X86_64
+  movq      rax, xm4
+  movq   [sszq], xm6
+%else
+  movd      eax, xm4
+  pextrd    edx, xm4, 1
+  movq   [sszd], xm6
+%endif
+  RET
+
+END
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_highbd_error_sse2.asm b/libvpx/libvpx/vp9/encoder/x86/vp9_highbd_error_sse2.asm
new file mode 100644
index 0000000..f3b8f01
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_highbd_error_sse2.asm
@@ -0,0 +1,98 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%define private_prefix vp9
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+ALIGN 16
+
+;
+; int64_t vp9_highbd_block_error_8bit(int32_t *coeff, int32_t *dqcoeff,
+;                                     intptr_t block_size, int64_t *ssz)
+;
+
+INIT_XMM sse2
+cglobal highbd_block_error_8bit, 3, 3, 8, uqc, dqc, size, ssz
+  pxor      m4, m4                 ; sse accumulator
+  pxor      m6, m6                 ; ssz accumulator
+  pxor      m5, m5                 ; dedicated zero register
+  lea     uqcq, [uqcq+sizeq*4]
+  lea     dqcq, [dqcq+sizeq*4]
+  neg    sizeq
+
+  ALIGN 16
+
+.loop:
+  mova      m0, [dqcq+sizeq*4]
+  packssdw  m0, [dqcq+sizeq*4+mmsize]
+  mova      m2, [uqcq+sizeq*4]
+  packssdw  m2, [uqcq+sizeq*4+mmsize]
+
+  mova      m1, [dqcq+sizeq*4+mmsize*2]
+  packssdw  m1, [dqcq+sizeq*4+mmsize*3]
+  mova      m3, [uqcq+sizeq*4+mmsize*2]
+  packssdw  m3, [uqcq+sizeq*4+mmsize*3]
+
+  add    sizeq, mmsize
+
+  ; individual errors are max. 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit)
+
+  psubw     m0, m2
+  pmaddwd   m2, m2
+  pmaddwd   m0, m0
+
+  psubw     m1, m3
+  pmaddwd   m3, m3
+  pmaddwd   m1, m1
+
+  ; accumulate in 64bit
+  punpckldq m7, m0, m5
+  punpckhdq m0, m5
+  paddq     m4, m7
+
+  punpckldq m7, m2, m5
+  punpckhdq m2, m5
+  paddq     m6, m7
+
+  punpckldq m7, m1, m5
+  punpckhdq m1, m5
+  paddq     m4, m7
+
+  punpckldq m7, m3, m5
+  punpckhdq m3, m5
+  paddq     m6, m7
+
+  paddq     m4, m0
+  paddq     m4, m1
+  paddq     m6, m2
+  paddq     m6, m3
+
+  jnz .loop
+
+  ; accumulate horizontally and store in return value
+  movhlps   m5, m4
+  movhlps   m7, m6
+  paddq     m4, m5
+  paddq     m6, m7
+
+%if ARCH_X86_64
+  movq    rax, m4
+  movq [sszq], m6
+%else
+  mov     eax, sszm
+  pshufd   m5, m4, 0x1
+  movq  [eax], m6
+  movd    eax, m4
+  movd    edx, m5
+%endif
+  RET
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_quantize_sse2.c b/libvpx/libvpx/vp9/encoder/x86/vp9_quantize_sse2.c
new file mode 100644
index 0000000..2071dfe
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_quantize_sse2.c
@@ -0,0 +1,211 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+#include <xmmintrin.h>
+
+#include "./vp9_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+void vp9_quantize_fp_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
+                          int skip_block, const int16_t* zbin_ptr,
+                          const int16_t* round_ptr, const int16_t* quant_ptr,
+                          const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr,
+                          int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+                          uint16_t* eob_ptr,
+                          const int16_t* scan_ptr,
+                          const int16_t* iscan_ptr) {
+  __m128i zero;
+  __m128i thr;
+  int16_t nzflag;
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+
+  coeff_ptr += n_coeffs;
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i*)round_ptr);
+        quant = _mm_load_si128((const __m128i*)quant_ptr);
+        dequant = _mm_load_si128((const __m128i*)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    thr = _mm_srai_epi16(dequant, 1);
+
+    // AC only loop
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+            _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
+
+        if (nzflag) {
+          qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+          qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+          qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+          qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+          // Reinsert signs
+          qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+          qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+          qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+          qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+          coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+          coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+        } else {
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+        }
+      }
+
+      if (nzflag) {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm b/libvpx/libvpx/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
new file mode 100644
index 0000000..ec61c0c
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
@@ -0,0 +1,201 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%define private_prefix vp9
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_1: times 8 dw 1
+
+SECTION .text
+
+%macro QUANTIZE_FP 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, \
+                                eob, scan, iscan
+  cmp                    dword skipm, 0
+  jne .blank
+
+  ; actual quantize loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, dequantmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+%ifidn %1, fp_32x32
+  pcmpeqw                         m5, m5
+  psrlw                           m5, 15
+  paddw                           m1, m5
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  mova                            m3, [r2q]                ; m3 = dequant
+  mov                             r3, qcoeffmp
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+%ifidn %1, fp_32x32
+  psllw                           m2, 1
+%endif
+  pxor                            m5, m5                   ; m5 = dedicated zero
+
+  lea                         coeffq, [  coeffq+ncoeffq*2]
+  lea                            r5q, [  r5q+ncoeffq*2]
+  lea                            r3q, [ r3q+ncoeffq*2]
+  lea                            r4q, [r4q+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpeqw                         m7, m7
+
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  mova            [r3q+ncoeffq*2+ 0], m8
+  mova            [r3q+ncoeffq*2+16], m13
+%ifidn %1, fp_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; r4[i] = r3[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; r4[i] = r3[i] * q
+%ifidn %1, fp_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+  psrlw                           m0, m3, 2
+%else
+  psrlw                           m0, m3, 1
+%endif
+  mova            [r4q+ncoeffq*2+ 0], m8
+  mova            [r4q+ncoeffq*2+16], m13
+  pcmpeqw                         m8, m5                   ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  r5q+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  r5q+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m7                   ; m11 = scan[i] + 1
+  pandn                           m8, m6                   ; m8 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jz .accumulate_eob
+
+.ac_only_loop:
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+
+  pcmpgtw                         m7, m6,  m0
+  pcmpgtw                        m12, m11, m0
+  pmovmskb                       r6d, m7
+  pmovmskb                       r2d, m12
+
+  or                              r6, r2
+  jz .skip_iter
+
+  pcmpeqw                         m7, m7
+
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  mova            [r3q+ncoeffq*2+ 0], m14
+  mova            [r3q+ncoeffq*2+16], m13
+%ifidn %1, fp_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; r4[i] = r3[i] * q
+  pmullw                         m13, m3                   ; r4[i] = r3[i] * q
+%ifidn %1, fp_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+  mova            [r4q+ncoeffq*2+ 0], m14
+  mova            [r4q+ncoeffq*2+16], m13
+  pcmpeqw                        m14, m5                   ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  r5q+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  r5q+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m7                   ; m11 = scan[i] + 1
+  pandn                          m14, m6                   ; m14 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+  jmp .accumulate_eob
+.skip_iter:
+  mova            [r3q+ncoeffq*2+ 0], m5
+  mova            [r3q+ncoeffq*2+16], m5
+  mova            [r4q+ncoeffq*2+ 0], m5
+  mova            [r4q+ncoeffq*2+16], m5
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+.accumulate_eob:
+  ; horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  pextrw                          r6, m8, 0
+  mov                           [r2], r6
+  RET
+
+  ; skip-block, i.e. just write all zeroes
+.blank:
+  mov                             r0, dqcoeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, qcoeffmp
+  mov                             r3, eobmp
+
+  lea                            r0q, [r0q+ncoeffq*2]
+  lea                            r2q, [r2q+ncoeffq*2]
+  neg                        ncoeffq
+  pxor                            m7, m7
+.blank_loop:
+  mova            [r0q+ncoeffq*2+ 0], m7
+  mova            [r0q+ncoeffq*2+16], m7
+  mova            [r2q+ncoeffq*2+ 0], m7
+  mova            [r2q+ncoeffq*2+16], m7
+  add                        ncoeffq, mmsize
+  jl .blank_loop
+  mov                     word [r3q], 0
+  RET
+%endmacro
+
+INIT_XMM ssse3
+QUANTIZE_FP fp, 7
+QUANTIZE_FP fp_32x32, 7
diff --git a/libvpx/libvpx/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm b/libvpx/libvpx/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
new file mode 100644
index 0000000..21aaa93
--- /dev/null
+++ b/libvpx/libvpx/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
@@ -0,0 +1,212 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+; void vp9_temporal_filter_apply_sse2 | arg
+;  (unsigned char  *frame1,           |  0
+;   unsigned int    stride,           |  1
+;   unsigned char  *frame2,           |  2
+;   unsigned int    block_width,      |  3
+;   unsigned int    block_height,     |  4
+;   int             strength,         |  5
+;   int             filter_weight,    |  6
+;   unsigned int   *accumulator,      |  7
+;   unsigned short *count)            |  8
+global sym(vp9_temporal_filter_apply_sse2) PRIVATE
+sym(vp9_temporal_filter_apply_sse2):
+
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ALIGN_STACK 16, rax
+    %define block_width    0
+    %define block_height  16
+    %define strength      32
+    %define filter_weight 48
+    %define rounding_bit  64
+    %define rbp_backup    80
+    %define stack_size    96
+    sub         rsp,           stack_size
+    mov         [rsp + rbp_backup], rbp
+    ; end prolog
+
+        mov         edx,            arg(3)
+        mov         [rsp + block_width], rdx
+        mov         edx,            arg(4)
+        mov         [rsp + block_height], rdx
+        movd        xmm6,           arg(5)
+        movdqa      [rsp + strength], xmm6 ; where strength is used, all 16 bytes are read
+
+        ; calculate the rounding bit outside the loop
+        ; 0x8000 >> (16 - strength)
+        mov         rdx,            16
+        sub         rdx,            arg(5) ; 16 - strength
+        movq        xmm4,           rdx    ; can't use rdx w/ shift
+        movdqa      xmm5,           [GLOBAL(_const_top_bit)]
+        psrlw       xmm5,           xmm4
+        movdqa      [rsp + rounding_bit], xmm5
+
+        mov         rsi,            arg(0) ; src/frame1
+        mov         rdx,            arg(2) ; predictor frame
+        mov         rdi,            arg(7) ; accumulator
+        mov         rax,            arg(8) ; count
+
+        ; dup the filter weight and store for later
+        movd        xmm0,           arg(6) ; filter_weight
+        pshuflw     xmm0,           xmm0, 0
+        punpcklwd   xmm0,           xmm0
+        movdqa      [rsp + filter_weight], xmm0
+
+        mov         rbp,            arg(1) ; stride
+        pxor        xmm7,           xmm7   ; zero for extraction
+
+        mov         rcx,            [rsp + block_width]
+        imul        rcx,            [rsp + block_height]
+        add         rcx,            rdx
+        cmp         dword ptr [rsp + block_width], 8
+        jne         .temporal_filter_apply_load_16
+
+.temporal_filter_apply_load_8:
+        movq        xmm0,           [rsi]  ; first row
+        lea         rsi,            [rsi + rbp] ; += stride
+        punpcklbw   xmm0,           xmm7   ; src[ 0- 7]
+        movq        xmm1,           [rsi]  ; second row
+        lea         rsi,            [rsi + rbp] ; += stride
+        punpcklbw   xmm1,           xmm7   ; src[ 8-15]
+        jmp         .temporal_filter_apply_load_finished
+
+.temporal_filter_apply_load_16:
+        movdqa      xmm0,           [rsi]  ; src (frame1)
+        lea         rsi,            [rsi + rbp] ; += stride
+        movdqa      xmm1,           xmm0
+        punpcklbw   xmm0,           xmm7   ; src[ 0- 7]
+        punpckhbw   xmm1,           xmm7   ; src[ 8-15]
+
+.temporal_filter_apply_load_finished:
+        movdqa      xmm2,           [rdx]  ; predictor (frame2)
+        movdqa      xmm3,           xmm2
+        punpcklbw   xmm2,           xmm7   ; pred[ 0- 7]
+        punpckhbw   xmm3,           xmm7   ; pred[ 8-15]
+
+        ; modifier = src_byte - pixel_value
+        psubw       xmm0,           xmm2   ; src - pred[ 0- 7]
+        psubw       xmm1,           xmm3   ; src - pred[ 8-15]
+
+        ; modifier *= modifier
+        pmullw      xmm0,           xmm0   ; modifer[ 0- 7]^2
+        pmullw      xmm1,           xmm1   ; modifer[ 8-15]^2
+
+        ; modifier *= 3
+        pmullw      xmm0,           [GLOBAL(_const_3w)]
+        pmullw      xmm1,           [GLOBAL(_const_3w)]
+
+        ; modifer += 0x8000 >> (16 - strength)
+        paddw       xmm0,           [rsp + rounding_bit]
+        paddw       xmm1,           [rsp + rounding_bit]
+
+        ; modifier >>= strength
+        psrlw       xmm0,           [rsp + strength]
+        psrlw       xmm1,           [rsp + strength]
+
+        ; modifier = 16 - modifier
+        ; saturation takes care of modifier > 16
+        movdqa      xmm3,           [GLOBAL(_const_16w)]
+        movdqa      xmm2,           [GLOBAL(_const_16w)]
+        psubusw     xmm3,           xmm1
+        psubusw     xmm2,           xmm0
+
+        ; modifier *= filter_weight
+        pmullw      xmm2,           [rsp + filter_weight]
+        pmullw      xmm3,           [rsp + filter_weight]
+
+        ; count
+        movdqa      xmm4,           [rax]
+        movdqa      xmm5,           [rax+16]
+        ; += modifier
+        paddw       xmm4,           xmm2
+        paddw       xmm5,           xmm3
+        ; write back
+        movdqa      [rax],          xmm4
+        movdqa      [rax+16],       xmm5
+        lea         rax,            [rax + 16*2] ; count += 16*(sizeof(short))
+
+        ; load and extract the predictor up to shorts
+        pxor        xmm7,           xmm7
+        movdqa      xmm0,           [rdx]
+        lea         rdx,            [rdx + 16*1] ; pred += 16*(sizeof(char))
+        movdqa      xmm1,           xmm0
+        punpcklbw   xmm0,           xmm7   ; pred[ 0- 7]
+        punpckhbw   xmm1,           xmm7   ; pred[ 8-15]
+
+        ; modifier *= pixel_value
+        pmullw      xmm0,           xmm2
+        pmullw      xmm1,           xmm3
+
+        ; expand to double words
+        movdqa      xmm2,           xmm0
+        punpcklwd   xmm0,           xmm7   ; [ 0- 3]
+        punpckhwd   xmm2,           xmm7   ; [ 4- 7]
+        movdqa      xmm3,           xmm1
+        punpcklwd   xmm1,           xmm7   ; [ 8-11]
+        punpckhwd   xmm3,           xmm7   ; [12-15]
+
+        ; accumulator
+        movdqa      xmm4,           [rdi]
+        movdqa      xmm5,           [rdi+16]
+        movdqa      xmm6,           [rdi+32]
+        movdqa      xmm7,           [rdi+48]
+        ; += modifier
+        paddd       xmm4,           xmm0
+        paddd       xmm5,           xmm2
+        paddd       xmm6,           xmm1
+        paddd       xmm7,           xmm3
+        ; write back
+        movdqa      [rdi],          xmm4
+        movdqa      [rdi+16],       xmm5
+        movdqa      [rdi+32],       xmm6
+        movdqa      [rdi+48],       xmm7
+        lea         rdi,            [rdi + 16*4] ; accumulator += 16*(sizeof(int))
+
+        cmp         rdx,            rcx
+        je          .temporal_filter_apply_epilog
+        pxor        xmm7,           xmm7   ; zero for extraction
+        cmp         dword ptr [rsp + block_width], 16
+        je          .temporal_filter_apply_load_16
+        jmp         .temporal_filter_apply_load_8
+
+.temporal_filter_apply_epilog:
+    ; begin epilog
+    mov         rbp,            [rsp + rbp_backup]
+    add         rsp,            stack_size
+    pop         rsp
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+_const_3w:
+    times 8 dw 3
+align 16
+_const_top_bit:
+    times 8 dw 1<<15
+align 16
+_const_16w
+    times 8 dw 16
diff --git a/libvpx/libvpx/vp9/exports_dec b/libvpx/libvpx/vp9/exports_dec
new file mode 100644
index 0000000..0a61fde
--- /dev/null
+++ b/libvpx/libvpx/vp9/exports_dec
@@ -0,0 +1,2 @@
+data vpx_codec_vp9_dx_algo
+text vpx_codec_vp9_dx
diff --git a/libvpx/libvpx/vp9/exports_enc b/libvpx/libvpx/vp9/exports_enc
new file mode 100644
index 0000000..2a0fef3
--- /dev/null
+++ b/libvpx/libvpx/vp9/exports_enc
@@ -0,0 +1,2 @@
+data vpx_codec_vp9_cx_algo
+text vpx_codec_vp9_cx
diff --git a/libvpx/libvpx/vp9/vp9_common.mk b/libvpx/libvpx/vp9/vp9_common.mk
new file mode 100644
index 0000000..d0135c6
--- /dev/null
+++ b/libvpx/libvpx/vp9/vp9_common.mk
@@ -0,0 +1,95 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+VP9_COMMON_SRCS-yes += vp9_common.mk
+VP9_COMMON_SRCS-yes += vp9_iface_common.h
+VP9_COMMON_SRCS-yes += common/vp9_ppflags.h
+VP9_COMMON_SRCS-yes += common/vp9_alloccommon.c
+VP9_COMMON_SRCS-yes += common/vp9_blockd.c
+VP9_COMMON_SRCS-yes += common/vp9_debugmodes.c
+VP9_COMMON_SRCS-yes += common/vp9_entropy.c
+VP9_COMMON_SRCS-yes += common/vp9_entropymode.c
+VP9_COMMON_SRCS-yes += common/vp9_entropymv.c
+VP9_COMMON_SRCS-yes += common/vp9_frame_buffers.c
+VP9_COMMON_SRCS-yes += common/vp9_frame_buffers.h
+VP9_COMMON_SRCS-yes += common/vp9_idct.c
+VP9_COMMON_SRCS-yes += common/vp9_alloccommon.h
+VP9_COMMON_SRCS-yes += common/vp9_blockd.h
+VP9_COMMON_SRCS-yes += common/vp9_common.h
+VP9_COMMON_SRCS-yes += common/vp9_entropy.h
+VP9_COMMON_SRCS-yes += common/vp9_entropymode.h
+VP9_COMMON_SRCS-yes += common/vp9_entropymv.h
+VP9_COMMON_SRCS-yes += common/vp9_enums.h
+VP9_COMMON_SRCS-yes += common/vp9_filter.h
+VP9_COMMON_SRCS-yes += common/vp9_filter.c
+VP9_COMMON_SRCS-yes += common/vp9_idct.h
+VP9_COMMON_SRCS-yes += common/vp9_loopfilter.h
+VP9_COMMON_SRCS-yes += common/vp9_thread_common.h
+VP9_COMMON_SRCS-yes += common/vp9_mv.h
+VP9_COMMON_SRCS-yes += common/vp9_onyxc_int.h
+VP9_COMMON_SRCS-yes += common/vp9_pred_common.h
+VP9_COMMON_SRCS-yes += common/vp9_pred_common.c
+VP9_COMMON_SRCS-yes += common/vp9_quant_common.h
+VP9_COMMON_SRCS-yes += common/vp9_reconinter.h
+VP9_COMMON_SRCS-yes += common/vp9_reconintra.h
+VP9_COMMON_SRCS-yes += common/vp9_rtcd.c
+VP9_COMMON_SRCS-yes += common/vp9_rtcd_defs.pl
+VP9_COMMON_SRCS-yes += common/vp9_scale.h
+VP9_COMMON_SRCS-yes += common/vp9_scale.c
+VP9_COMMON_SRCS-yes += common/vp9_seg_common.h
+VP9_COMMON_SRCS-yes += common/vp9_seg_common.c
+VP9_COMMON_SRCS-yes += common/vp9_textblit.h
+VP9_COMMON_SRCS-yes += common/vp9_tile_common.h
+VP9_COMMON_SRCS-yes += common/vp9_tile_common.c
+VP9_COMMON_SRCS-yes += common/vp9_loopfilter.c
+VP9_COMMON_SRCS-yes += common/vp9_thread_common.c
+VP9_COMMON_SRCS-yes += common/vp9_mvref_common.c
+VP9_COMMON_SRCS-yes += common/vp9_mvref_common.h
+VP9_COMMON_SRCS-yes += common/vp9_quant_common.c
+VP9_COMMON_SRCS-yes += common/vp9_reconinter.c
+VP9_COMMON_SRCS-yes += common/vp9_reconintra.c
+VP9_COMMON_SRCS-$(CONFIG_POSTPROC_VISUALIZER) += common/vp9_textblit.c
+VP9_COMMON_SRCS-yes += common/vp9_common_data.c
+VP9_COMMON_SRCS-yes += common/vp9_common_data.h
+VP9_COMMON_SRCS-yes += common/vp9_scan.c
+VP9_COMMON_SRCS-yes += common/vp9_scan.h
+
+VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_postproc.h
+VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_postproc.c
+VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_mfqe.h
+VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_mfqe.c
+ifeq ($(CONFIG_VP9_POSTPROC),yes)
+VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_mfqe_sse2.asm
+VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_postproc_sse2.asm
+endif
+
+ifneq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+VP9_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/vp9_itrans4_dspr2.c
+VP9_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/vp9_itrans8_dspr2.c
+VP9_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/vp9_itrans16_dspr2.c
+endif
+
+# common (msa)
+VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_idct4x4_msa.c
+VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_idct8x8_msa.c
+VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_idct16x16_msa.c
+
+ifeq ($(CONFIG_VP9_POSTPROC),yes)
+VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_mfqe_msa.c
+endif
+
+VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_idct_intrin_sse2.c
+
+ifneq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_iht4x4_add_neon.c
+VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_iht8x8_add_neon.c
+endif
+
+$(eval $(call rtcd_h_template,vp9_rtcd,vp9/common/vp9_rtcd_defs.pl))
diff --git a/libvpx/libvpx/vp9/vp9_cx_iface.c b/libvpx/libvpx/vp9/vp9_cx_iface.c
new file mode 100644
index 0000000..10d6893
--- /dev/null
+++ b/libvpx/libvpx/vp9/vp9_cx_iface.c
@@ -0,0 +1,1659 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "./vpx_config.h"
+#include "vpx/vpx_encoder.h"
+#include "vpx_ports/vpx_once.h"
+#include "vpx_ports/system_state.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "./vpx_version.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vpx/vp8cx.h"
+#include "vp9/encoder/vp9_firstpass.h"
+#include "vp9/vp9_iface_common.h"
+
+struct vp9_extracfg {
+  int                         cpu_used;  // available cpu percentage in 1/16
+  unsigned int                enable_auto_alt_ref;
+  unsigned int                noise_sensitivity;
+  unsigned int                sharpness;
+  unsigned int                static_thresh;
+  unsigned int                tile_columns;
+  unsigned int                tile_rows;
+  unsigned int                arnr_max_frames;
+  unsigned int                arnr_strength;
+  unsigned int                min_gf_interval;
+  unsigned int                max_gf_interval;
+  vp8e_tuning                 tuning;
+  unsigned int                cq_level;  // constrained quality level
+  unsigned int                rc_max_intra_bitrate_pct;
+  unsigned int                rc_max_inter_bitrate_pct;
+  unsigned int                gf_cbr_boost_pct;
+  unsigned int                lossless;
+  unsigned int                target_level;
+  unsigned int                frame_parallel_decoding_mode;
+  AQ_MODE                     aq_mode;
+  unsigned int                frame_periodic_boost;
+  vpx_bit_depth_t             bit_depth;
+  vp9e_tune_content           content;
+  vpx_color_space_t           color_space;
+  vpx_color_range_t           color_range;
+  int                         render_width;
+  int                         render_height;
+};
+
+static struct vp9_extracfg default_extra_cfg = {
+  0,                          // cpu_used
+  1,                          // enable_auto_alt_ref
+  0,                          // noise_sensitivity
+  0,                          // sharpness
+  0,                          // static_thresh
+  6,                          // tile_columns
+  0,                          // tile_rows
+  7,                          // arnr_max_frames
+  5,                          // arnr_strength
+  0,                          // min_gf_interval; 0 -> default decision
+  0,                          // max_gf_interval; 0 -> default decision
+  VP8_TUNE_PSNR,              // tuning
+  10,                         // cq_level
+  0,                          // rc_max_intra_bitrate_pct
+  0,                          // rc_max_inter_bitrate_pct
+  0,                          // gf_cbr_boost_pct
+  0,                          // lossless
+  255,                        // target_level
+  1,                          // frame_parallel_decoding_mode
+  NO_AQ,                      // aq_mode
+  0,                          // frame_periodic_delta_q
+  VPX_BITS_8,                 // Bit depth
+  VP9E_CONTENT_DEFAULT,       // content
+  VPX_CS_UNKNOWN,             // color space
+  0,                          // color range
+  0,                          // render width
+  0,                          // render height
+};
+
+struct vpx_codec_alg_priv {
+  vpx_codec_priv_t        base;
+  vpx_codec_enc_cfg_t     cfg;
+  struct vp9_extracfg     extra_cfg;
+  VP9EncoderConfig        oxcf;
+  VP9_COMP               *cpi;
+  unsigned char          *cx_data;
+  size_t                  cx_data_sz;
+  unsigned char          *pending_cx_data;
+  size_t                  pending_cx_data_sz;
+  int                     pending_frame_count;
+  size_t                  pending_frame_sizes[8];
+  size_t                  pending_frame_magnitude;
+  vpx_image_t             preview_img;
+  vpx_enc_frame_flags_t   next_frame_flags;
+  vp8_postproc_cfg_t      preview_ppcfg;
+  vpx_codec_pkt_list_decl(256) pkt_list;
+  unsigned int                 fixed_kf_cntr;
+  vpx_codec_priv_output_cx_pkt_cb_pair_t output_cx_pkt_cb;
+  // BufferPool that holds all reference frames.
+  BufferPool              *buffer_pool;
+};
+
+static VP9_REFFRAME ref_frame_to_vp9_reframe(vpx_ref_frame_type_t frame) {
+  switch (frame) {
+    case VP8_LAST_FRAME:
+      return VP9_LAST_FLAG;
+    case VP8_GOLD_FRAME:
+      return VP9_GOLD_FLAG;
+    case VP8_ALTR_FRAME:
+      return VP9_ALT_FLAG;
+  }
+  assert(0 && "Invalid Reference Frame");
+  return VP9_LAST_FLAG;
+}
+
+static vpx_codec_err_t update_error_state(vpx_codec_alg_priv_t *ctx,
+    const struct vpx_internal_error_info *error) {
+  const vpx_codec_err_t res = error->error_code;
+
+  if (res != VPX_CODEC_OK)
+    ctx->base.err_detail = error->has_detail ? error->detail : NULL;
+
+  return res;
+}
+
+
+#undef ERROR
+#define ERROR(str) do {\
+    ctx->base.err_detail = str;\
+    return VPX_CODEC_INVALID_PARAM;\
+  } while (0)
+
+#define RANGE_CHECK(p, memb, lo, hi) do {\
+    if (!(((p)->memb == lo || (p)->memb > (lo)) && (p)->memb <= hi)) \
+      ERROR(#memb " out of range ["#lo".."#hi"]");\
+  } while (0)
+
+#define RANGE_CHECK_HI(p, memb, hi) do {\
+    if (!((p)->memb <= (hi))) \
+      ERROR(#memb " out of range [.."#hi"]");\
+  } while (0)
+
+#define RANGE_CHECK_LO(p, memb, lo) do {\
+    if (!((p)->memb >= (lo))) \
+      ERROR(#memb " out of range ["#lo"..]");\
+  } while (0)
+
+#define RANGE_CHECK_BOOL(p, memb) do {\
+    if (!!((p)->memb) != (p)->memb) ERROR(#memb " expected boolean");\
+  } while (0)
+
+static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
+                                       const vpx_codec_enc_cfg_t *cfg,
+                                       const struct vp9_extracfg *extra_cfg) {
+  RANGE_CHECK(cfg, g_w,                   1, 65535);  // 16 bits available
+  RANGE_CHECK(cfg, g_h,                   1, 65535);  // 16 bits available
+  RANGE_CHECK(cfg, g_timebase.den,        1, 1000000000);
+  RANGE_CHECK(cfg, g_timebase.num,        1, 1000000000);
+  RANGE_CHECK_HI(cfg, g_profile,          3);
+
+  RANGE_CHECK_HI(cfg, rc_max_quantizer,   63);
+  RANGE_CHECK_HI(cfg, rc_min_quantizer,   cfg->rc_max_quantizer);
+  RANGE_CHECK_BOOL(extra_cfg, lossless);
+  RANGE_CHECK(extra_cfg, aq_mode,           0, AQ_MODE_COUNT - 1);
+  RANGE_CHECK(extra_cfg, frame_periodic_boost, 0, 1);
+  RANGE_CHECK_HI(cfg, g_threads,          64);
+  RANGE_CHECK_HI(cfg, g_lag_in_frames,    MAX_LAG_BUFFERS);
+  RANGE_CHECK(cfg, rc_end_usage,          VPX_VBR, VPX_Q);
+  RANGE_CHECK_HI(cfg, rc_undershoot_pct,  100);
+  RANGE_CHECK_HI(cfg, rc_overshoot_pct,   100);
+  RANGE_CHECK_HI(cfg, rc_2pass_vbr_bias_pct, 100);
+  RANGE_CHECK(cfg, kf_mode,               VPX_KF_DISABLED, VPX_KF_AUTO);
+  RANGE_CHECK_BOOL(cfg,                   rc_resize_allowed);
+  RANGE_CHECK_HI(cfg, rc_dropframe_thresh,   100);
+  RANGE_CHECK_HI(cfg, rc_resize_up_thresh,   100);
+  RANGE_CHECK_HI(cfg, rc_resize_down_thresh, 100);
+  RANGE_CHECK(cfg,        g_pass,         VPX_RC_ONE_PASS, VPX_RC_LAST_PASS);
+  RANGE_CHECK(extra_cfg, min_gf_interval, 0, (MAX_LAG_BUFFERS - 1));
+  RANGE_CHECK(extra_cfg, max_gf_interval, 0, (MAX_LAG_BUFFERS - 1));
+  if (extra_cfg->max_gf_interval > 0) {
+    RANGE_CHECK(extra_cfg, max_gf_interval, 2, (MAX_LAG_BUFFERS - 1));
+  }
+  if (extra_cfg->min_gf_interval > 0 && extra_cfg->max_gf_interval > 0) {
+    RANGE_CHECK(extra_cfg, max_gf_interval, extra_cfg->min_gf_interval,
+      (MAX_LAG_BUFFERS - 1));
+  }
+
+  if (cfg->rc_resize_allowed == 1) {
+    RANGE_CHECK(cfg, rc_scaled_width, 0, cfg->g_w);
+    RANGE_CHECK(cfg, rc_scaled_height, 0, cfg->g_h);
+  }
+
+  RANGE_CHECK(cfg, ss_number_layers, 1, VPX_SS_MAX_LAYERS);
+  RANGE_CHECK(cfg, ts_number_layers, 1, VPX_TS_MAX_LAYERS);
+
+  {
+    unsigned int level = extra_cfg->target_level;
+    if (level != LEVEL_1 && level != LEVEL_1_1 && level != LEVEL_2 &&
+        level != LEVEL_2_1 && level != LEVEL_3 && level != LEVEL_3_1 &&
+        level != LEVEL_4 && level != LEVEL_4_1 && level != LEVEL_5 &&
+        level != LEVEL_5_1 && level != LEVEL_5_2 && level != LEVEL_6 &&
+        level != LEVEL_6_1 && level != LEVEL_6_2 &&
+        level != LEVEL_UNKNOWN && level != LEVEL_MAX)
+    ERROR("target_level is invalid");
+  }
+
+  if (cfg->ss_number_layers * cfg->ts_number_layers > VPX_MAX_LAYERS)
+    ERROR("ss_number_layers * ts_number_layers is out of range");
+  if (cfg->ts_number_layers > 1) {
+    unsigned int sl, tl;
+    for (sl = 1; sl < cfg->ss_number_layers; ++sl) {
+      for (tl = 1; tl < cfg->ts_number_layers; ++tl) {
+        const int layer =
+            LAYER_IDS_TO_IDX(sl, tl, cfg->ts_number_layers);
+        if (cfg->layer_target_bitrate[layer] <
+            cfg->layer_target_bitrate[layer - 1])
+        ERROR("ts_target_bitrate entries are not increasing");
+      }
+    }
+
+    RANGE_CHECK(cfg, ts_rate_decimator[cfg->ts_number_layers - 1], 1, 1);
+    for (tl = cfg->ts_number_layers - 2; tl > 0; --tl)
+      if (cfg->ts_rate_decimator[tl - 1] != 2 * cfg->ts_rate_decimator[tl])
+        ERROR("ts_rate_decimator factors are not powers of 2");
+  }
+
+#if CONFIG_SPATIAL_SVC
+
+  if ((cfg->ss_number_layers > 1 || cfg->ts_number_layers > 1) &&
+      cfg->g_pass == VPX_RC_LAST_PASS) {
+    unsigned int i, alt_ref_sum = 0;
+    for (i = 0; i < cfg->ss_number_layers; ++i) {
+      if (cfg->ss_enable_auto_alt_ref[i])
+        ++alt_ref_sum;
+    }
+    if (alt_ref_sum > REF_FRAMES - cfg->ss_number_layers)
+      ERROR("Not enough ref buffers for svc alt ref frames");
+    if (cfg->ss_number_layers * cfg->ts_number_layers > 3 &&
+        cfg->g_error_resilient == 0)
+    ERROR("Multiple frame context are not supported for more than 3 layers");
+  }
+#endif
+
+  // VP9 does not support a lower bound on the keyframe interval in
+  // automatic keyframe placement mode.
+  if (cfg->kf_mode != VPX_KF_DISABLED &&
+      cfg->kf_min_dist != cfg->kf_max_dist &&
+      cfg->kf_min_dist > 0)
+    ERROR("kf_min_dist not supported in auto mode, use 0 "
+          "or kf_max_dist instead.");
+
+  RANGE_CHECK(extra_cfg, enable_auto_alt_ref, 0, 2);
+  RANGE_CHECK(extra_cfg, cpu_used, -8, 8);
+  RANGE_CHECK_HI(extra_cfg, noise_sensitivity, 6);
+  RANGE_CHECK(extra_cfg, tile_columns, 0, 6);
+  RANGE_CHECK(extra_cfg, tile_rows, 0, 2);
+  RANGE_CHECK_HI(extra_cfg, sharpness, 7);
+  RANGE_CHECK(extra_cfg, arnr_max_frames, 0, 15);
+  RANGE_CHECK_HI(extra_cfg, arnr_strength, 6);
+  RANGE_CHECK(extra_cfg, cq_level, 0, 63);
+  RANGE_CHECK(cfg, g_bit_depth, VPX_BITS_8, VPX_BITS_12);
+  RANGE_CHECK(cfg, g_input_bit_depth, 8, 12);
+  RANGE_CHECK(extra_cfg, content,
+              VP9E_CONTENT_DEFAULT, VP9E_CONTENT_INVALID - 1);
+
+  // TODO(yaowu): remove this when ssim tuning is implemented for vp9
+  if (extra_cfg->tuning == VP8_TUNE_SSIM)
+      ERROR("Option --tune=ssim is not currently supported in VP9.");
+
+  if (cfg->g_pass == VPX_RC_LAST_PASS) {
+    const size_t packet_sz = sizeof(FIRSTPASS_STATS);
+    const int n_packets = (int)(cfg->rc_twopass_stats_in.sz / packet_sz);
+    const FIRSTPASS_STATS *stats;
+
+    if (cfg->rc_twopass_stats_in.buf == NULL)
+      ERROR("rc_twopass_stats_in.buf not set.");
+
+    if (cfg->rc_twopass_stats_in.sz % packet_sz)
+      ERROR("rc_twopass_stats_in.sz indicates truncated packet.");
+
+    if (cfg->ss_number_layers > 1 || cfg->ts_number_layers > 1) {
+      int i;
+      unsigned int n_packets_per_layer[VPX_SS_MAX_LAYERS] = {0};
+
+      stats = cfg->rc_twopass_stats_in.buf;
+      for (i = 0; i < n_packets; ++i) {
+        const int layer_id = (int)stats[i].spatial_layer_id;
+        if (layer_id >= 0 && layer_id < (int)cfg->ss_number_layers) {
+          ++n_packets_per_layer[layer_id];
+        }
+      }
+
+      for (i = 0; i < (int)cfg->ss_number_layers; ++i) {
+        unsigned int layer_id;
+        if (n_packets_per_layer[i] < 2) {
+          ERROR("rc_twopass_stats_in requires at least two packets for each "
+                "layer.");
+        }
+
+        stats = (const FIRSTPASS_STATS *)cfg->rc_twopass_stats_in.buf +
+                n_packets - cfg->ss_number_layers + i;
+        layer_id = (int)stats->spatial_layer_id;
+
+        if (layer_id >= cfg->ss_number_layers
+            ||(unsigned int)(stats->count + 0.5) !=
+               n_packets_per_layer[layer_id] - 1)
+          ERROR("rc_twopass_stats_in missing EOS stats packet");
+      }
+    } else {
+      if (cfg->rc_twopass_stats_in.sz < 2 * packet_sz)
+        ERROR("rc_twopass_stats_in requires at least two packets.");
+
+      stats =
+          (const FIRSTPASS_STATS *)cfg->rc_twopass_stats_in.buf + n_packets - 1;
+
+      if ((int)(stats->count + 0.5) != n_packets - 1)
+        ERROR("rc_twopass_stats_in missing EOS stats packet");
+    }
+  }
+
+#if !CONFIG_VP9_HIGHBITDEPTH
+  if (cfg->g_profile > (unsigned int)PROFILE_1) {
+    ERROR("Profile > 1 not supported in this build configuration");
+  }
+#endif
+  if (cfg->g_profile <= (unsigned int)PROFILE_1 &&
+      cfg->g_bit_depth > VPX_BITS_8) {
+    ERROR("Codec high bit-depth not supported in profile < 2");
+  }
+  if (cfg->g_profile <= (unsigned int)PROFILE_1 &&
+      cfg->g_input_bit_depth > 8) {
+    ERROR("Source high bit-depth not supported in profile < 2");
+  }
+  if (cfg->g_profile > (unsigned int)PROFILE_1 &&
+      cfg->g_bit_depth == VPX_BITS_8) {
+    ERROR("Codec bit-depth 8 not supported in profile > 1");
+  }
+  RANGE_CHECK(extra_cfg, color_space, VPX_CS_UNKNOWN, VPX_CS_SRGB);
+  RANGE_CHECK(extra_cfg, color_range,
+              VPX_CR_STUDIO_RANGE, VPX_CR_FULL_RANGE);
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t validate_img(vpx_codec_alg_priv_t *ctx,
+                                    const vpx_image_t *img) {
+  switch (img->fmt) {
+    case VPX_IMG_FMT_YV12:
+    case VPX_IMG_FMT_I420:
+    case VPX_IMG_FMT_I42016:
+      break;
+    case VPX_IMG_FMT_I422:
+    case VPX_IMG_FMT_I444:
+    case VPX_IMG_FMT_I440:
+      if (ctx->cfg.g_profile != (unsigned int)PROFILE_1) {
+        ERROR("Invalid image format. I422, I444, I440 images are "
+              "not supported in profile.");
+      }
+      break;
+    case VPX_IMG_FMT_I42216:
+    case VPX_IMG_FMT_I44416:
+    case VPX_IMG_FMT_I44016:
+      if (ctx->cfg.g_profile != (unsigned int)PROFILE_1 &&
+          ctx->cfg.g_profile != (unsigned int)PROFILE_3) {
+        ERROR("Invalid image format. 16-bit I422, I444, I440 images are "
+              "not supported in profile.");
+      }
+      break;
+    default:
+      ERROR("Invalid image format. Only YV12, I420, I422, I444 images are "
+            "supported.");
+      break;
+  }
+
+  if (img->d_w != ctx->cfg.g_w || img->d_h != ctx->cfg.g_h)
+    ERROR("Image size must match encoder init configuration size");
+
+  return VPX_CODEC_OK;
+}
+
+static int get_image_bps(const vpx_image_t *img) {
+  switch (img->fmt) {
+    case VPX_IMG_FMT_YV12:
+    case VPX_IMG_FMT_I420: return 12;
+    case VPX_IMG_FMT_I422: return 16;
+    case VPX_IMG_FMT_I444: return 24;
+    case VPX_IMG_FMT_I440: return 16;
+    case VPX_IMG_FMT_I42016: return 24;
+    case VPX_IMG_FMT_I42216: return 32;
+    case VPX_IMG_FMT_I44416: return 48;
+    case VPX_IMG_FMT_I44016: return 32;
+    default: assert(0 && "Invalid image format"); break;
+  }
+  return 0;
+}
+
+static vpx_codec_err_t set_encoder_config(
+  VP9EncoderConfig *oxcf,
+  const vpx_codec_enc_cfg_t *cfg,
+  const struct vp9_extracfg *extra_cfg) {
+  const int is_vbr = cfg->rc_end_usage == VPX_VBR;
+  int sl, tl;
+  oxcf->profile = cfg->g_profile;
+  oxcf->max_threads = (int)cfg->g_threads;
+  oxcf->width   = cfg->g_w;
+  oxcf->height  = cfg->g_h;
+  oxcf->bit_depth = cfg->g_bit_depth;
+  oxcf->input_bit_depth = cfg->g_input_bit_depth;
+  // guess a frame rate if out of whack, use 30
+  oxcf->init_framerate = (double)cfg->g_timebase.den / cfg->g_timebase.num;
+  if (oxcf->init_framerate > 180)
+    oxcf->init_framerate = 30;
+
+  oxcf->mode = GOOD;
+
+  switch (cfg->g_pass) {
+    case VPX_RC_ONE_PASS:
+      oxcf->pass = 0;
+      break;
+    case VPX_RC_FIRST_PASS:
+      oxcf->pass = 1;
+      break;
+    case VPX_RC_LAST_PASS:
+      oxcf->pass = 2;
+      break;
+  }
+
+  oxcf->lag_in_frames = cfg->g_pass == VPX_RC_FIRST_PASS ? 0
+                                                         : cfg->g_lag_in_frames;
+  oxcf->rc_mode = cfg->rc_end_usage;
+
+  // Convert target bandwidth from Kbit/s to Bit/s
+  oxcf->target_bandwidth = 1000 * cfg->rc_target_bitrate;
+  oxcf->rc_max_intra_bitrate_pct = extra_cfg->rc_max_intra_bitrate_pct;
+  oxcf->rc_max_inter_bitrate_pct = extra_cfg->rc_max_inter_bitrate_pct;
+  oxcf->gf_cbr_boost_pct = extra_cfg->gf_cbr_boost_pct;
+
+  oxcf->best_allowed_q =
+      extra_cfg->lossless ? 0 : vp9_quantizer_to_qindex(cfg->rc_min_quantizer);
+  oxcf->worst_allowed_q =
+      extra_cfg->lossless ? 0 : vp9_quantizer_to_qindex(cfg->rc_max_quantizer);
+  oxcf->cq_level        = vp9_quantizer_to_qindex(extra_cfg->cq_level);
+  oxcf->fixed_q = -1;
+
+  oxcf->under_shoot_pct         = cfg->rc_undershoot_pct;
+  oxcf->over_shoot_pct          = cfg->rc_overshoot_pct;
+
+  oxcf->scaled_frame_width  = cfg->rc_scaled_width;
+  oxcf->scaled_frame_height = cfg->rc_scaled_height;
+  if (cfg->rc_resize_allowed == 1) {
+    oxcf->resize_mode =
+        (oxcf->scaled_frame_width == 0 || oxcf->scaled_frame_height == 0) ?
+            RESIZE_DYNAMIC : RESIZE_FIXED;
+  } else {
+    oxcf->resize_mode = RESIZE_NONE;
+  }
+
+  oxcf->maximum_buffer_size_ms   = is_vbr ? 240000 : cfg->rc_buf_sz;
+  oxcf->starting_buffer_level_ms = is_vbr ? 60000 : cfg->rc_buf_initial_sz;
+  oxcf->optimal_buffer_level_ms  = is_vbr ? 60000 : cfg->rc_buf_optimal_sz;
+
+  oxcf->drop_frames_water_mark   = cfg->rc_dropframe_thresh;
+
+  oxcf->two_pass_vbrbias         = cfg->rc_2pass_vbr_bias_pct;
+  oxcf->two_pass_vbrmin_section  = cfg->rc_2pass_vbr_minsection_pct;
+  oxcf->two_pass_vbrmax_section  = cfg->rc_2pass_vbr_maxsection_pct;
+
+  oxcf->auto_key               = cfg->kf_mode == VPX_KF_AUTO &&
+                                 cfg->kf_min_dist != cfg->kf_max_dist;
+
+  oxcf->key_freq               = cfg->kf_max_dist;
+
+  oxcf->speed                  =  abs(extra_cfg->cpu_used);
+  oxcf->encode_breakout        =  extra_cfg->static_thresh;
+  oxcf->enable_auto_arf        =  extra_cfg->enable_auto_alt_ref;
+  oxcf->noise_sensitivity      =  extra_cfg->noise_sensitivity;
+  oxcf->sharpness              =  extra_cfg->sharpness;
+
+  oxcf->two_pass_stats_in      =  cfg->rc_twopass_stats_in;
+
+#if CONFIG_FP_MB_STATS
+  oxcf->firstpass_mb_stats_in  = cfg->rc_firstpass_mb_stats_in;
+#endif
+
+  oxcf->color_space = extra_cfg->color_space;
+  oxcf->color_range = extra_cfg->color_range;
+  oxcf->render_width  = extra_cfg->render_width;
+  oxcf->render_height = extra_cfg->render_height;
+  oxcf->arnr_max_frames = extra_cfg->arnr_max_frames;
+  oxcf->arnr_strength   = extra_cfg->arnr_strength;
+  oxcf->min_gf_interval = extra_cfg->min_gf_interval;
+  oxcf->max_gf_interval = extra_cfg->max_gf_interval;
+
+  oxcf->tuning = extra_cfg->tuning;
+  oxcf->content = extra_cfg->content;
+
+  oxcf->tile_columns = extra_cfg->tile_columns;
+
+  // TODO(yunqing): The dependencies between row tiles cause error in multi-
+  // threaded encoding. For now, tile_rows is forced to be 0 in this case.
+  // The further fix can be done by adding synchronizations after a tile row
+  // is encoded. But this will hurt multi-threaded encoder performance. So,
+  // it is recommended to use tile-rows=0 while encoding with threads > 1.
+  if (oxcf->max_threads > 1 && oxcf->tile_columns > 0)
+    oxcf->tile_rows  = 0;
+  else
+    oxcf->tile_rows  = extra_cfg->tile_rows;
+
+  oxcf->error_resilient_mode         = cfg->g_error_resilient;
+  oxcf->frame_parallel_decoding_mode = extra_cfg->frame_parallel_decoding_mode;
+
+  oxcf->aq_mode = extra_cfg->aq_mode;
+
+  oxcf->frame_periodic_boost =  extra_cfg->frame_periodic_boost;
+
+  oxcf->ss_number_layers = cfg->ss_number_layers;
+  oxcf->ts_number_layers = cfg->ts_number_layers;
+  oxcf->temporal_layering_mode = (enum vp9e_temporal_layering_mode)
+      cfg->temporal_layering_mode;
+
+  oxcf->target_level = extra_cfg->target_level;
+
+  for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
+#if CONFIG_SPATIAL_SVC
+    oxcf->ss_enable_auto_arf[sl] = cfg->ss_enable_auto_alt_ref[sl];
+#endif
+    for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
+      oxcf->layer_target_bitrate[sl * oxcf->ts_number_layers + tl] =
+          1000 * cfg->layer_target_bitrate[sl * oxcf->ts_number_layers + tl];
+    }
+  }
+  if (oxcf->ss_number_layers == 1 && oxcf->pass != 0) {
+    oxcf->ss_target_bitrate[0] = (int)oxcf->target_bandwidth;
+#if CONFIG_SPATIAL_SVC
+    oxcf->ss_enable_auto_arf[0] = extra_cfg->enable_auto_alt_ref;
+#endif
+  }
+  if (oxcf->ts_number_layers > 1) {
+    for (tl = 0; tl < VPX_TS_MAX_LAYERS; ++tl) {
+      oxcf->ts_rate_decimator[tl] = cfg->ts_rate_decimator[tl] ?
+          cfg->ts_rate_decimator[tl] : 1;
+    }
+  } else if (oxcf->ts_number_layers == 1) {
+    oxcf->ts_rate_decimator[0] = 1;
+  }
+  /*
+  printf("Current VP9 Settings: \n");
+  printf("target_bandwidth: %d\n", oxcf->target_bandwidth);
+  printf("target_level: %d\n", oxcf->target_level);
+  printf("noise_sensitivity: %d\n", oxcf->noise_sensitivity);
+  printf("sharpness: %d\n",    oxcf->sharpness);
+  printf("cpu_used: %d\n",  oxcf->cpu_used);
+  printf("Mode: %d\n",     oxcf->mode);
+  printf("auto_key: %d\n",  oxcf->auto_key);
+  printf("key_freq: %d\n", oxcf->key_freq);
+  printf("end_usage: %d\n", oxcf->end_usage);
+  printf("under_shoot_pct: %d\n", oxcf->under_shoot_pct);
+  printf("over_shoot_pct: %d\n", oxcf->over_shoot_pct);
+  printf("starting_buffer_level: %d\n", oxcf->starting_buffer_level);
+  printf("optimal_buffer_level: %d\n",  oxcf->optimal_buffer_level);
+  printf("maximum_buffer_size: %d\n", oxcf->maximum_buffer_size);
+  printf("fixed_q: %d\n",  oxcf->fixed_q);
+  printf("worst_allowed_q: %d\n", oxcf->worst_allowed_q);
+  printf("best_allowed_q: %d\n", oxcf->best_allowed_q);
+  printf("allow_spatial_resampling: %d\n", oxcf->allow_spatial_resampling);
+  printf("scaled_frame_width: %d\n", oxcf->scaled_frame_width);
+  printf("scaled_frame_height: %d\n", oxcf->scaled_frame_height);
+  printf("two_pass_vbrbias: %d\n",  oxcf->two_pass_vbrbias);
+  printf("two_pass_vbrmin_section: %d\n", oxcf->two_pass_vbrmin_section);
+  printf("two_pass_vbrmax_section: %d\n", oxcf->two_pass_vbrmax_section);
+  printf("lag_in_frames: %d\n", oxcf->lag_in_frames);
+  printf("enable_auto_arf: %d\n", oxcf->enable_auto_arf);
+  printf("Version: %d\n", oxcf->Version);
+  printf("encode_breakout: %d\n", oxcf->encode_breakout);
+  printf("error resilient: %d\n", oxcf->error_resilient_mode);
+  printf("frame parallel detokenization: %d\n",
+         oxcf->frame_parallel_decoding_mode);
+  */
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t encoder_set_config(vpx_codec_alg_priv_t *ctx,
+                                          const vpx_codec_enc_cfg_t  *cfg) {
+  vpx_codec_err_t res;
+  int force_key = 0;
+
+  if (cfg->g_w != ctx->cfg.g_w || cfg->g_h != ctx->cfg.g_h) {
+    if (cfg->g_lag_in_frames > 1 || cfg->g_pass != VPX_RC_ONE_PASS)
+      ERROR("Cannot change width or height after initialization");
+    if (!valid_ref_frame_size(ctx->cfg.g_w, ctx->cfg.g_h, cfg->g_w, cfg->g_h) ||
+        (ctx->cpi->initial_width && (int)cfg->g_w > ctx->cpi->initial_width) ||
+        (ctx->cpi->initial_height && (int)cfg->g_h > ctx->cpi->initial_height))
+      force_key = 1;
+  }
+
+  // Prevent increasing lag_in_frames. This check is stricter than it needs
+  // to be -- the limit is not increasing past the first lag_in_frames
+  // value, but we don't track the initial config, only the last successful
+  // config.
+  if (cfg->g_lag_in_frames > ctx->cfg.g_lag_in_frames)
+    ERROR("Cannot increase lag_in_frames");
+
+  res = validate_config(ctx, cfg, &ctx->extra_cfg);
+
+  if (res == VPX_CODEC_OK) {
+    ctx->cfg = *cfg;
+    set_encoder_config(&ctx->oxcf, &ctx->cfg, &ctx->extra_cfg);
+    // On profile change, request a key frame
+    force_key |= ctx->cpi->common.profile != ctx->oxcf.profile;
+    vp9_change_config(ctx->cpi, &ctx->oxcf);
+  }
+
+  if (force_key)
+    ctx->next_frame_flags |= VPX_EFLAG_FORCE_KF;
+
+  return res;
+}
+
+static vpx_codec_err_t ctrl_get_quantizer(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL)
+    return VPX_CODEC_INVALID_PARAM;
+  *arg = vp9_get_quantizer(ctx->cpi);
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_get_quantizer64(vpx_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL)
+    return VPX_CODEC_INVALID_PARAM;
+  *arg = vp9_qindex_to_quantizer(vp9_get_quantizer(ctx->cpi));
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t update_extra_cfg(vpx_codec_alg_priv_t *ctx,
+                                        const struct vp9_extracfg *extra_cfg) {
+  const vpx_codec_err_t res = validate_config(ctx, &ctx->cfg, extra_cfg);
+  if (res == VPX_CODEC_OK) {
+    ctx->extra_cfg = *extra_cfg;
+    set_encoder_config(&ctx->oxcf, &ctx->cfg, &ctx->extra_cfg);
+    vp9_change_config(ctx->cpi, &ctx->oxcf);
+  }
+  return res;
+}
+
+static vpx_codec_err_t ctrl_set_cpuused(vpx_codec_alg_priv_t *ctx,
+                                        va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.cpu_used = CAST(VP8E_SET_CPUUSED, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_enable_auto_alt_ref(vpx_codec_alg_priv_t *ctx,
+                                                    va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.enable_auto_alt_ref = CAST(VP8E_SET_ENABLEAUTOALTREF, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_noise_sensitivity(vpx_codec_alg_priv_t *ctx,
+                                                  va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.noise_sensitivity = CAST(VP9E_SET_NOISE_SENSITIVITY, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_sharpness(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.sharpness = CAST(VP8E_SET_SHARPNESS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_static_thresh(vpx_codec_alg_priv_t *ctx,
+                                              va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.static_thresh = CAST(VP8E_SET_STATIC_THRESHOLD, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_tile_columns(vpx_codec_alg_priv_t *ctx,
+                                             va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.tile_columns = CAST(VP9E_SET_TILE_COLUMNS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_tile_rows(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.tile_rows = CAST(VP9E_SET_TILE_ROWS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_arnr_max_frames(vpx_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.arnr_max_frames = CAST(VP8E_SET_ARNR_MAXFRAMES, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_arnr_strength(vpx_codec_alg_priv_t *ctx,
+                                              va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.arnr_strength = CAST(VP8E_SET_ARNR_STRENGTH, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_arnr_type(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_set_tuning(vpx_codec_alg_priv_t *ctx,
+                                       va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.tuning = CAST(VP8E_SET_TUNING, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_cq_level(vpx_codec_alg_priv_t *ctx,
+                                         va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.cq_level = CAST(VP8E_SET_CQ_LEVEL, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_rc_max_intra_bitrate_pct(
+    vpx_codec_alg_priv_t *ctx, va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.rc_max_intra_bitrate_pct =
+      CAST(VP8E_SET_MAX_INTRA_BITRATE_PCT, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_rc_max_inter_bitrate_pct(
+    vpx_codec_alg_priv_t *ctx, va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.rc_max_inter_bitrate_pct =
+      CAST(VP8E_SET_MAX_INTER_BITRATE_PCT, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_rc_gf_cbr_boost_pct(
+    vpx_codec_alg_priv_t *ctx, va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.gf_cbr_boost_pct =
+      CAST(VP9E_SET_GF_CBR_BOOST_PCT, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_lossless(vpx_codec_alg_priv_t *ctx,
+                                         va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.lossless = CAST(VP9E_SET_LOSSLESS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_frame_parallel_decoding_mode(
+    vpx_codec_alg_priv_t *ctx, va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.frame_parallel_decoding_mode =
+      CAST(VP9E_SET_FRAME_PARALLEL_DECODING, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_aq_mode(vpx_codec_alg_priv_t *ctx,
+                                        va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.aq_mode = CAST(VP9E_SET_AQ_MODE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_min_gf_interval(vpx_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.min_gf_interval = CAST(VP9E_SET_MIN_GF_INTERVAL, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_max_gf_interval(vpx_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.max_gf_interval = CAST(VP9E_SET_MAX_GF_INTERVAL, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_frame_periodic_boost(vpx_codec_alg_priv_t *ctx,
+                                                     va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.frame_periodic_boost = CAST(VP9E_SET_FRAME_PERIODIC_BOOST, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_target_level(vpx_codec_alg_priv_t *ctx,
+                                             va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.target_level = CAST(VP9E_SET_TARGET_LEVEL, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_get_level(vpx_codec_alg_priv_t *ctx, va_list args) {
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL) return VPX_CODEC_INVALID_PARAM;
+  *arg = (int)vp9_get_level(&ctx->cpi->level_info.level_spec);
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t encoder_init(vpx_codec_ctx_t *ctx,
+                                    vpx_codec_priv_enc_mr_cfg_t *data) {
+  vpx_codec_err_t res = VPX_CODEC_OK;
+  (void)data;
+
+  if (ctx->priv == NULL) {
+    vpx_codec_alg_priv_t *const priv = vpx_calloc(1, sizeof(*priv));
+    if (priv == NULL)
+      return VPX_CODEC_MEM_ERROR;
+
+    ctx->priv = (vpx_codec_priv_t *)priv;
+    ctx->priv->init_flags = ctx->init_flags;
+    ctx->priv->enc.total_encoders = 1;
+    priv->buffer_pool =
+        (BufferPool *)vpx_calloc(1, sizeof(BufferPool));
+    if (priv->buffer_pool == NULL)
+      return VPX_CODEC_MEM_ERROR;
+
+#if CONFIG_MULTITHREAD
+    if (pthread_mutex_init(&priv->buffer_pool->pool_mutex, NULL)) {
+      return VPX_CODEC_MEM_ERROR;
+    }
+#endif
+
+    if (ctx->config.enc) {
+      // Update the reference to the config structure to an internal copy.
+      priv->cfg = *ctx->config.enc;
+      ctx->config.enc = &priv->cfg;
+    }
+
+    priv->extra_cfg = default_extra_cfg;
+    once(vp9_initialize_enc);
+
+    res = validate_config(priv, &priv->cfg, &priv->extra_cfg);
+
+    if (res == VPX_CODEC_OK) {
+      set_encoder_config(&priv->oxcf, &priv->cfg, &priv->extra_cfg);
+#if CONFIG_VP9_HIGHBITDEPTH
+      priv->oxcf.use_highbitdepth =
+          (ctx->init_flags & VPX_CODEC_USE_HIGHBITDEPTH) ? 1 : 0;
+#endif
+      priv->cpi = vp9_create_compressor(&priv->oxcf, priv->buffer_pool);
+      if (priv->cpi == NULL)
+        res = VPX_CODEC_MEM_ERROR;
+      else
+        priv->cpi->output_pkt_list = &priv->pkt_list.head;
+    }
+  }
+
+  return res;
+}
+
+static vpx_codec_err_t encoder_destroy(vpx_codec_alg_priv_t *ctx) {
+  free(ctx->cx_data);
+  vp9_remove_compressor(ctx->cpi);
+#if CONFIG_MULTITHREAD
+  pthread_mutex_destroy(&ctx->buffer_pool->pool_mutex);
+#endif
+  vpx_free(ctx->buffer_pool);
+  vpx_free(ctx);
+  return VPX_CODEC_OK;
+}
+
+static void pick_quickcompress_mode(vpx_codec_alg_priv_t *ctx,
+                                    unsigned long duration,
+                                    unsigned long deadline) {
+  MODE new_mode = BEST;
+
+  switch (ctx->cfg.g_pass) {
+    case VPX_RC_ONE_PASS:
+      if (deadline > 0) {
+        const vpx_codec_enc_cfg_t *const cfg = &ctx->cfg;
+
+        // Convert duration parameter from stream timebase to microseconds.
+        const uint64_t duration_us = (uint64_t)duration * 1000000 *
+           (uint64_t)cfg->g_timebase.num /(uint64_t)cfg->g_timebase.den;
+
+        // If the deadline is more that the duration this frame is to be shown,
+        // use good quality mode. Otherwise use realtime mode.
+        new_mode = (deadline > duration_us) ? GOOD : REALTIME;
+      } else {
+        new_mode = BEST;
+      }
+      break;
+    case VPX_RC_FIRST_PASS:
+      break;
+    case VPX_RC_LAST_PASS:
+      new_mode = deadline > 0 ? GOOD : BEST;
+      break;
+  }
+
+  if (deadline == VPX_DL_REALTIME) {
+    ctx->oxcf.pass = 0;
+    new_mode = REALTIME;
+  }
+
+  if (ctx->oxcf.mode != new_mode) {
+    ctx->oxcf.mode = new_mode;
+    vp9_change_config(ctx->cpi, &ctx->oxcf);
+  }
+}
+
+// Turn on to test if supplemental superframe data breaks decoding
+// #define TEST_SUPPLEMENTAL_SUPERFRAME_DATA
+static int write_superframe_index(vpx_codec_alg_priv_t *ctx) {
+  uint8_t marker = 0xc0;
+  unsigned int mask;
+  int mag, index_sz;
+
+  assert(ctx->pending_frame_count);
+  assert(ctx->pending_frame_count <= 8);
+
+  // Add the number of frames to the marker byte
+  marker |= ctx->pending_frame_count - 1;
+
+  // Choose the magnitude
+  for (mag = 0, mask = 0xff; mag < 4; mag++) {
+    if (ctx->pending_frame_magnitude < mask)
+      break;
+    mask <<= 8;
+    mask |= 0xff;
+  }
+  marker |= mag << 3;
+
+  // Write the index
+  index_sz = 2 + (mag + 1) * ctx->pending_frame_count;
+  if (ctx->pending_cx_data_sz + index_sz < ctx->cx_data_sz) {
+    uint8_t *x = ctx->pending_cx_data + ctx->pending_cx_data_sz;
+    int i, j;
+#ifdef TEST_SUPPLEMENTAL_SUPERFRAME_DATA
+    uint8_t marker_test = 0xc0;
+    int mag_test = 2;     // 1 - 4
+    int frames_test = 4;  // 1 - 8
+    int index_sz_test = 2 + mag_test * frames_test;
+    marker_test |= frames_test - 1;
+    marker_test |= (mag_test - 1) << 3;
+    *x++ = marker_test;
+    for (i = 0; i < mag_test * frames_test; ++i)
+      *x++ = 0;  // fill up with arbitrary data
+    *x++ = marker_test;
+    ctx->pending_cx_data_sz += index_sz_test;
+    printf("Added supplemental superframe data\n");
+#endif
+
+    *x++ = marker;
+    for (i = 0; i < ctx->pending_frame_count; i++) {
+      unsigned int this_sz = (unsigned int)ctx->pending_frame_sizes[i];
+
+      for (j = 0; j <= mag; j++) {
+        *x++ = this_sz & 0xff;
+        this_sz >>= 8;
+      }
+    }
+    *x++ = marker;
+    ctx->pending_cx_data_sz += index_sz;
+#ifdef TEST_SUPPLEMENTAL_SUPERFRAME_DATA
+    index_sz += index_sz_test;
+#endif
+  }
+  return index_sz;
+}
+
+static int64_t timebase_units_to_ticks(const vpx_rational_t *timebase,
+                                       int64_t n) {
+  return n * TICKS_PER_SEC * timebase->num / timebase->den;
+}
+
+static int64_t ticks_to_timebase_units(const vpx_rational_t *timebase,
+                                       int64_t n) {
+  const int64_t round = (int64_t)TICKS_PER_SEC * timebase->num / 2 - 1;
+  return (n * timebase->den + round) / timebase->num / TICKS_PER_SEC;
+}
+
+static vpx_codec_frame_flags_t get_frame_pkt_flags(const VP9_COMP *cpi,
+                                                   unsigned int lib_flags) {
+  vpx_codec_frame_flags_t flags = lib_flags << 16;
+
+  if (lib_flags & FRAMEFLAGS_KEY ||
+      (cpi->use_svc &&
+          cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+              cpi->svc.number_temporal_layers +
+              cpi->svc.temporal_layer_id].is_key_frame)
+     )
+    flags |= VPX_FRAME_IS_KEY;
+
+  if (cpi->droppable)
+    flags |= VPX_FRAME_IS_DROPPABLE;
+
+  return flags;
+}
+
+const size_t kMinCompressedSize = 8192;
+static vpx_codec_err_t encoder_encode(vpx_codec_alg_priv_t  *ctx,
+                                      const vpx_image_t *img,
+                                      vpx_codec_pts_t pts,
+                                      unsigned long duration,
+                                      vpx_enc_frame_flags_t enc_flags,
+                                      unsigned long deadline) {
+  volatile vpx_codec_err_t res = VPX_CODEC_OK;
+  volatile vpx_enc_frame_flags_t flags = enc_flags;
+  VP9_COMP *const cpi = ctx->cpi;
+  const vpx_rational_t *const timebase = &ctx->cfg.g_timebase;
+  size_t data_sz;
+
+  if (cpi == NULL) return VPX_CODEC_INVALID_PARAM;
+
+  if (img != NULL) {
+    res = validate_img(ctx, img);
+    if (res == VPX_CODEC_OK) {
+      // There's no codec control for multiple alt-refs so check the encoder
+      // instance for its status to determine the compressed data size.
+      data_sz = ctx->cfg.g_w * ctx->cfg.g_h * get_image_bps(img) / 8 *
+                (cpi->multi_arf_allowed ? 8 : 2);
+      if (data_sz < kMinCompressedSize)
+        data_sz = kMinCompressedSize;
+      if (ctx->cx_data == NULL || ctx->cx_data_sz < data_sz) {
+        ctx->cx_data_sz = data_sz;
+        free(ctx->cx_data);
+        ctx->cx_data = (unsigned char*)malloc(ctx->cx_data_sz);
+        if (ctx->cx_data == NULL) {
+          return VPX_CODEC_MEM_ERROR;
+        }
+      }
+    }
+  }
+
+  pick_quickcompress_mode(ctx, duration, deadline);
+  vpx_codec_pkt_list_init(&ctx->pkt_list);
+
+  // Handle Flags
+  if (((flags & VP8_EFLAG_NO_UPD_GF) && (flags & VP8_EFLAG_FORCE_GF)) ||
+       ((flags & VP8_EFLAG_NO_UPD_ARF) && (flags & VP8_EFLAG_FORCE_ARF))) {
+    ctx->base.err_detail = "Conflicting flags.";
+    return VPX_CODEC_INVALID_PARAM;
+  }
+
+  if (setjmp(cpi->common.error.jmp)) {
+    cpi->common.error.setjmp = 0;
+    res = update_error_state(ctx, &cpi->common.error);
+    vpx_clear_system_state();
+    return res;
+  }
+  cpi->common.error.setjmp = 1;
+
+  vp9_apply_encoding_flags(cpi, flags);
+
+  // Handle fixed keyframe intervals
+  if (ctx->cfg.kf_mode == VPX_KF_AUTO &&
+      ctx->cfg.kf_min_dist == ctx->cfg.kf_max_dist) {
+    if (++ctx->fixed_kf_cntr > ctx->cfg.kf_min_dist) {
+      flags |= VPX_EFLAG_FORCE_KF;
+      ctx->fixed_kf_cntr = 1;
+    }
+  }
+
+  if (res == VPX_CODEC_OK) {
+    unsigned int lib_flags = 0;
+    YV12_BUFFER_CONFIG sd;
+    int64_t dst_time_stamp = timebase_units_to_ticks(timebase, pts);
+    int64_t dst_end_time_stamp =
+        timebase_units_to_ticks(timebase, pts + duration);
+    size_t size, cx_data_sz;
+    unsigned char *cx_data;
+
+    // Set up internal flags
+    if (ctx->base.init_flags & VPX_CODEC_USE_PSNR)
+      cpi->b_calculate_psnr = 1;
+
+    if (img != NULL) {
+      res = image2yuvconfig(img, &sd);
+
+      // Store the original flags in to the frame buffer. Will extract the
+      // key frame flag when we actually encode this frame.
+      if (vp9_receive_raw_frame(cpi, flags | ctx->next_frame_flags,
+                                &sd, dst_time_stamp, dst_end_time_stamp)) {
+        res = update_error_state(ctx, &cpi->common.error);
+      }
+      ctx->next_frame_flags = 0;
+    }
+
+    cx_data = ctx->cx_data;
+    cx_data_sz = ctx->cx_data_sz;
+
+    /* Any pending invisible frames? */
+    if (ctx->pending_cx_data) {
+      memmove(cx_data, ctx->pending_cx_data, ctx->pending_cx_data_sz);
+      ctx->pending_cx_data = cx_data;
+      cx_data += ctx->pending_cx_data_sz;
+      cx_data_sz -= ctx->pending_cx_data_sz;
+
+      /* TODO: this is a minimal check, the underlying codec doesn't respect
+       * the buffer size anyway.
+       */
+      if (cx_data_sz < ctx->cx_data_sz / 2) {
+        vpx_internal_error(&cpi->common.error, VPX_CODEC_ERROR,
+                           "Compressed data buffer too small");
+        return VPX_CODEC_ERROR;
+      }
+    }
+
+    while (cx_data_sz >= ctx->cx_data_sz / 2 &&
+           -1 != vp9_get_compressed_data(cpi, &lib_flags, &size,
+                                         cx_data, &dst_time_stamp,
+                                         &dst_end_time_stamp, !img)) {
+      if (size) {
+        vpx_codec_cx_pkt_t pkt;
+
+#if CONFIG_SPATIAL_SVC
+        if (cpi->use_svc)
+          cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+              cpi->svc.number_temporal_layers].layer_size += size;
+#endif
+
+        // Pack invisible frames with the next visible frame
+        if (!cpi->common.show_frame ||
+            (cpi->use_svc &&
+             cpi->svc.spatial_layer_id < cpi->svc.number_spatial_layers - 1)
+            ) {
+          if (ctx->pending_cx_data == 0)
+            ctx->pending_cx_data = cx_data;
+          ctx->pending_cx_data_sz += size;
+          ctx->pending_frame_sizes[ctx->pending_frame_count++] = size;
+          ctx->pending_frame_magnitude |= size;
+          cx_data += size;
+          cx_data_sz -= size;
+
+          if (ctx->output_cx_pkt_cb.output_cx_pkt) {
+            pkt.kind = VPX_CODEC_CX_FRAME_PKT;
+            pkt.data.frame.pts = ticks_to_timebase_units(timebase,
+                                                         dst_time_stamp);
+            pkt.data.frame.duration =
+               (unsigned long)ticks_to_timebase_units(timebase,
+                   dst_end_time_stamp - dst_time_stamp);
+            pkt.data.frame.flags = get_frame_pkt_flags(cpi, lib_flags);
+            pkt.data.frame.buf = ctx->pending_cx_data;
+            pkt.data.frame.sz  = size;
+            ctx->pending_cx_data = NULL;
+            ctx->pending_cx_data_sz = 0;
+            ctx->pending_frame_count = 0;
+            ctx->pending_frame_magnitude = 0;
+            ctx->output_cx_pkt_cb.output_cx_pkt(
+                &pkt, ctx->output_cx_pkt_cb.user_priv);
+          }
+          continue;
+        }
+
+        // Add the frame packet to the list of returned packets.
+        pkt.kind = VPX_CODEC_CX_FRAME_PKT;
+        pkt.data.frame.pts = ticks_to_timebase_units(timebase, dst_time_stamp);
+        pkt.data.frame.duration =
+           (unsigned long)ticks_to_timebase_units(timebase,
+               dst_end_time_stamp - dst_time_stamp);
+        pkt.data.frame.flags = get_frame_pkt_flags(cpi, lib_flags);
+
+        if (ctx->pending_cx_data) {
+          ctx->pending_frame_sizes[ctx->pending_frame_count++] = size;
+          ctx->pending_frame_magnitude |= size;
+          ctx->pending_cx_data_sz += size;
+          // write the superframe only for the case when
+          if (!ctx->output_cx_pkt_cb.output_cx_pkt)
+            size += write_superframe_index(ctx);
+          pkt.data.frame.buf = ctx->pending_cx_data;
+          pkt.data.frame.sz  = ctx->pending_cx_data_sz;
+          ctx->pending_cx_data = NULL;
+          ctx->pending_cx_data_sz = 0;
+          ctx->pending_frame_count = 0;
+          ctx->pending_frame_magnitude = 0;
+        } else {
+          pkt.data.frame.buf = cx_data;
+          pkt.data.frame.sz  = size;
+        }
+        pkt.data.frame.partition_id = -1;
+
+        if(ctx->output_cx_pkt_cb.output_cx_pkt)
+          ctx->output_cx_pkt_cb.output_cx_pkt(&pkt,
+                                              ctx->output_cx_pkt_cb.user_priv);
+        else
+          vpx_codec_pkt_list_add(&ctx->pkt_list.head, &pkt);
+
+        cx_data += size;
+        cx_data_sz -= size;
+#if VPX_ENCODER_ABI_VERSION > (5 + VPX_CODEC_ABI_VERSION)
+#if CONFIG_SPATIAL_SVC
+        if (cpi->use_svc && !ctx->output_cx_pkt_cb.output_cx_pkt) {
+          vpx_codec_cx_pkt_t pkt_sizes, pkt_psnr;
+          int sl;
+          vp9_zero(pkt_sizes);
+          vp9_zero(pkt_psnr);
+          pkt_sizes.kind = VPX_CODEC_SPATIAL_SVC_LAYER_SIZES;
+          pkt_psnr.kind = VPX_CODEC_SPATIAL_SVC_LAYER_PSNR;
+          for (sl = 0; sl < cpi->svc.number_spatial_layers; ++sl) {
+            LAYER_CONTEXT *lc =
+                &cpi->svc.layer_context[sl * cpi->svc.number_temporal_layers];
+            pkt_sizes.data.layer_sizes[sl] = lc->layer_size;
+            pkt_psnr.data.layer_psnr[sl] = lc->psnr_pkt;
+            lc->layer_size = 0;
+          }
+
+          vpx_codec_pkt_list_add(&ctx->pkt_list.head, &pkt_sizes);
+
+          vpx_codec_pkt_list_add(&ctx->pkt_list.head, &pkt_psnr);
+        }
+#endif
+#endif
+        if (is_one_pass_cbr_svc(cpi) &&
+            (cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)) {
+          // Encoded all spatial layers; exit loop.
+          break;
+        }
+      }
+    }
+  }
+
+  cpi->common.error.setjmp = 0;
+  return res;
+}
+
+static const vpx_codec_cx_pkt_t *encoder_get_cxdata(vpx_codec_alg_priv_t *ctx,
+                                                    vpx_codec_iter_t *iter) {
+  return vpx_codec_pkt_list_get(&ctx->pkt_list.head, iter);
+}
+
+static vpx_codec_err_t ctrl_set_reference(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  vpx_ref_frame_t *const frame = va_arg(args, vpx_ref_frame_t *);
+
+  if (frame != NULL) {
+    YV12_BUFFER_CONFIG sd;
+
+    image2yuvconfig(&frame->img, &sd);
+    vp9_set_reference_enc(ctx->cpi, ref_frame_to_vp9_reframe(frame->frame_type),
+                          &sd);
+    return VPX_CODEC_OK;
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+}
+
+static vpx_codec_err_t ctrl_copy_reference(vpx_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  vpx_ref_frame_t *const frame = va_arg(args, vpx_ref_frame_t *);
+
+  if (frame != NULL) {
+    YV12_BUFFER_CONFIG sd;
+
+    image2yuvconfig(&frame->img, &sd);
+    vp9_copy_reference_enc(ctx->cpi,
+                           ref_frame_to_vp9_reframe(frame->frame_type), &sd);
+    return VPX_CODEC_OK;
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+}
+
+static vpx_codec_err_t ctrl_get_reference(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  vp9_ref_frame_t *const frame = va_arg(args, vp9_ref_frame_t *);
+
+  if (frame != NULL) {
+    YV12_BUFFER_CONFIG *fb = get_ref_frame(&ctx->cpi->common, frame->idx);
+    if (fb == NULL) return VPX_CODEC_ERROR;
+
+    yuvconfig2image(&frame->img, fb, NULL);
+    return VPX_CODEC_OK;
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+}
+
+static vpx_codec_err_t ctrl_set_previewpp(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+#if CONFIG_VP9_POSTPROC
+  vp8_postproc_cfg_t *config = va_arg(args, vp8_postproc_cfg_t *);
+  if (config != NULL) {
+    ctx->preview_ppcfg = *config;
+    return VPX_CODEC_OK;
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+#else
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
+#endif
+}
+
+
+static vpx_image_t *encoder_get_preview(vpx_codec_alg_priv_t *ctx) {
+  YV12_BUFFER_CONFIG sd;
+  vp9_ppflags_t flags;
+  vp9_zero(flags);
+
+  if (ctx->preview_ppcfg.post_proc_flag) {
+    flags.post_proc_flag   = ctx->preview_ppcfg.post_proc_flag;
+    flags.deblocking_level = ctx->preview_ppcfg.deblocking_level;
+    flags.noise_level      = ctx->preview_ppcfg.noise_level;
+  }
+
+  if (vp9_get_preview_raw_frame(ctx->cpi, &sd, &flags) == 0) {
+    yuvconfig2image(&ctx->preview_img, &sd, NULL);
+    return &ctx->preview_img;
+  } else {
+    return NULL;
+  }
+}
+
+static vpx_codec_err_t ctrl_set_roi_map(vpx_codec_alg_priv_t *ctx,
+                                        va_list args) {
+  (void)ctx;
+  (void)args;
+
+  // TODO(yaowu): Need to re-implement and test for VP9.
+  return VPX_CODEC_INVALID_PARAM;
+}
+
+
+static vpx_codec_err_t ctrl_set_active_map(vpx_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  vpx_active_map_t *const map = va_arg(args, vpx_active_map_t *);
+
+  if (map) {
+    if (!vp9_set_active_map(ctx->cpi, map->active_map,
+                            (int)map->rows, (int)map->cols))
+      return VPX_CODEC_OK;
+    else
+      return VPX_CODEC_INVALID_PARAM;
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+}
+
+static vpx_codec_err_t ctrl_get_active_map(vpx_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  vpx_active_map_t *const map = va_arg(args, vpx_active_map_t *);
+
+  if (map) {
+    if (!vp9_get_active_map(ctx->cpi, map->active_map,
+                            (int)map->rows, (int)map->cols))
+      return VPX_CODEC_OK;
+    else
+      return VPX_CODEC_INVALID_PARAM;
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+}
+
+static vpx_codec_err_t ctrl_set_scale_mode(vpx_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  vpx_scaling_mode_t *const mode = va_arg(args, vpx_scaling_mode_t *);
+
+  if (mode) {
+    const int res = vp9_set_internal_size(ctx->cpi,
+                                          (VPX_SCALING)mode->h_scaling_mode,
+                                          (VPX_SCALING)mode->v_scaling_mode);
+    return (res == 0) ? VPX_CODEC_OK : VPX_CODEC_INVALID_PARAM;
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+}
+
+static vpx_codec_err_t ctrl_set_svc(vpx_codec_alg_priv_t *ctx, va_list args) {
+  int data = va_arg(args, int);
+  const vpx_codec_enc_cfg_t *cfg = &ctx->cfg;
+  // Both one-pass and two-pass RC are supported now.
+  // User setting this has to make sure of the following.
+  // In two-pass setting: either (but not both)
+  //      cfg->ss_number_layers > 1, or cfg->ts_number_layers > 1
+  // In one-pass setting:
+  //      either or both cfg->ss_number_layers > 1, or cfg->ts_number_layers > 1
+
+  vp9_set_svc(ctx->cpi, data);
+
+  if (data == 1 &&
+      (cfg->g_pass == VPX_RC_FIRST_PASS ||
+       cfg->g_pass == VPX_RC_LAST_PASS) &&
+       cfg->ss_number_layers > 1 &&
+       cfg->ts_number_layers > 1) {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_set_svc_layer_id(vpx_codec_alg_priv_t *ctx,
+                                             va_list args) {
+  vpx_svc_layer_id_t *const data = va_arg(args, vpx_svc_layer_id_t *);
+  VP9_COMP *const cpi = (VP9_COMP *)ctx->cpi;
+  SVC *const svc = &cpi->svc;
+
+  svc->first_spatial_layer_to_encode = data->spatial_layer_id;
+  svc->spatial_layer_to_encode = data->spatial_layer_id;
+  svc->temporal_layer_id = data->temporal_layer_id;
+  // Checks on valid layer_id input.
+  if (svc->temporal_layer_id < 0 ||
+      svc->temporal_layer_id >= (int)ctx->cfg.ts_number_layers) {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+  if (svc->first_spatial_layer_to_encode < 0 ||
+      svc->first_spatial_layer_to_encode >= (int)ctx->cfg.ss_number_layers) {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+  // First spatial layer to encode not implemented for two-pass.
+  if (is_two_pass_svc(cpi) && svc->first_spatial_layer_to_encode > 0)
+    return VPX_CODEC_INVALID_PARAM;
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_get_svc_layer_id(vpx_codec_alg_priv_t *ctx,
+                                             va_list args) {
+  vpx_svc_layer_id_t *data = va_arg(args, vpx_svc_layer_id_t *);
+  VP9_COMP *const cpi = (VP9_COMP *)ctx->cpi;
+  SVC *const svc = &cpi->svc;
+
+  data->spatial_layer_id = svc->spatial_layer_id;
+  data->temporal_layer_id = svc->temporal_layer_id;
+
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_set_svc_parameters(vpx_codec_alg_priv_t *ctx,
+                                               va_list args) {
+  VP9_COMP *const cpi = ctx->cpi;
+  vpx_svc_extra_cfg_t *const params = va_arg(args, vpx_svc_extra_cfg_t *);
+  int sl, tl;
+
+  // Number of temporal layers and number of spatial layers have to be set
+  // properly before calling this control function.
+  for (sl = 0; sl < cpi->svc.number_spatial_layers; ++sl) {
+    for (tl = 0; tl < cpi->svc.number_temporal_layers; ++tl) {
+      const int layer =
+          LAYER_IDS_TO_IDX(sl, tl, cpi->svc.number_temporal_layers);
+      LAYER_CONTEXT *lc =
+          &cpi->svc.layer_context[layer];
+      lc->max_q = params->max_quantizers[layer];
+      lc->min_q = params->min_quantizers[layer];
+      lc->scaling_factor_num = params->scaling_factor_num[sl];
+      lc->scaling_factor_den = params->scaling_factor_den[sl];
+    }
+  }
+
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_set_svc_ref_frame_config(vpx_codec_alg_priv_t *ctx,
+                                                     va_list args) {
+  VP9_COMP *const cpi = ctx->cpi;
+  vpx_svc_ref_frame_config_t *data = va_arg(args, vpx_svc_ref_frame_config_t *);
+  int sl;
+  for (sl = 0; sl < cpi->svc.number_spatial_layers; ++sl) {
+    cpi->svc.ext_frame_flags[sl] = data->frame_flags[sl];
+    cpi->svc.ext_lst_fb_idx[sl] = data->lst_fb_idx[sl];
+    cpi->svc.ext_gld_fb_idx[sl] = data->gld_fb_idx[sl];
+    cpi->svc.ext_alt_fb_idx[sl] = data->alt_fb_idx[sl];
+  }
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_register_cx_callback(vpx_codec_alg_priv_t *ctx,
+                                                 va_list args) {
+  vpx_codec_priv_output_cx_pkt_cb_pair_t *cbp =
+      (vpx_codec_priv_output_cx_pkt_cb_pair_t *)va_arg(args, void *);
+  ctx->output_cx_pkt_cb.output_cx_pkt = cbp->output_cx_pkt;
+  ctx->output_cx_pkt_cb.user_priv = cbp->user_priv;
+
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_set_tune_content(vpx_codec_alg_priv_t *ctx,
+                                             va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.content = CAST(VP9E_SET_TUNE_CONTENT, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_color_space(vpx_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.color_space = CAST(VP9E_SET_COLOR_SPACE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_color_range(vpx_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.color_range = CAST(VP9E_SET_COLOR_RANGE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_render_size(vpx_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  int *const render_size = va_arg(args, int *);
+  extra_cfg.render_width  = render_size[0];
+  extra_cfg.render_height = render_size[1];
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_ctrl_fn_map_t encoder_ctrl_maps[] = {
+  {VP8_COPY_REFERENCE,                ctrl_copy_reference},
+
+  // Setters
+  {VP8_SET_REFERENCE,                 ctrl_set_reference},
+  {VP8_SET_POSTPROC,                  ctrl_set_previewpp},
+  {VP8E_SET_ROI_MAP,                  ctrl_set_roi_map},
+  {VP8E_SET_ACTIVEMAP,                ctrl_set_active_map},
+  {VP8E_SET_SCALEMODE,                ctrl_set_scale_mode},
+  {VP8E_SET_CPUUSED,                  ctrl_set_cpuused},
+  {VP8E_SET_ENABLEAUTOALTREF,         ctrl_set_enable_auto_alt_ref},
+  {VP8E_SET_SHARPNESS,                ctrl_set_sharpness},
+  {VP8E_SET_STATIC_THRESHOLD,         ctrl_set_static_thresh},
+  {VP9E_SET_TILE_COLUMNS,             ctrl_set_tile_columns},
+  {VP9E_SET_TILE_ROWS,                ctrl_set_tile_rows},
+  {VP8E_SET_ARNR_MAXFRAMES,           ctrl_set_arnr_max_frames},
+  {VP8E_SET_ARNR_STRENGTH,            ctrl_set_arnr_strength},
+  {VP8E_SET_ARNR_TYPE,                ctrl_set_arnr_type},
+  {VP8E_SET_TUNING,                   ctrl_set_tuning},
+  {VP8E_SET_CQ_LEVEL,                 ctrl_set_cq_level},
+  {VP8E_SET_MAX_INTRA_BITRATE_PCT,    ctrl_set_rc_max_intra_bitrate_pct},
+  {VP9E_SET_MAX_INTER_BITRATE_PCT,    ctrl_set_rc_max_inter_bitrate_pct},
+  {VP9E_SET_GF_CBR_BOOST_PCT,         ctrl_set_rc_gf_cbr_boost_pct},
+  {VP9E_SET_LOSSLESS,                 ctrl_set_lossless},
+  {VP9E_SET_FRAME_PARALLEL_DECODING,  ctrl_set_frame_parallel_decoding_mode},
+  {VP9E_SET_AQ_MODE,                  ctrl_set_aq_mode},
+  {VP9E_SET_FRAME_PERIODIC_BOOST,     ctrl_set_frame_periodic_boost},
+  {VP9E_SET_SVC,                      ctrl_set_svc},
+  {VP9E_SET_SVC_PARAMETERS,           ctrl_set_svc_parameters},
+  {VP9E_REGISTER_CX_CALLBACK,         ctrl_register_cx_callback},
+  {VP9E_SET_SVC_LAYER_ID,             ctrl_set_svc_layer_id},
+  {VP9E_SET_TUNE_CONTENT,             ctrl_set_tune_content},
+  {VP9E_SET_COLOR_SPACE,              ctrl_set_color_space},
+  {VP9E_SET_COLOR_RANGE,              ctrl_set_color_range},
+  {VP9E_SET_NOISE_SENSITIVITY,        ctrl_set_noise_sensitivity},
+  {VP9E_SET_MIN_GF_INTERVAL,          ctrl_set_min_gf_interval},
+  {VP9E_SET_MAX_GF_INTERVAL,          ctrl_set_max_gf_interval},
+  {VP9E_SET_SVC_REF_FRAME_CONFIG,     ctrl_set_svc_ref_frame_config},
+  {VP9E_SET_RENDER_SIZE,              ctrl_set_render_size},
+  {VP9E_SET_TARGET_LEVEL,             ctrl_set_target_level},
+
+  // Getters
+  {VP8E_GET_LAST_QUANTIZER,           ctrl_get_quantizer},
+  {VP8E_GET_LAST_QUANTIZER_64,        ctrl_get_quantizer64},
+  {VP9_GET_REFERENCE,                 ctrl_get_reference},
+  {VP9E_GET_SVC_LAYER_ID,             ctrl_get_svc_layer_id},
+  {VP9E_GET_ACTIVEMAP,                ctrl_get_active_map},
+  {VP9E_GET_LEVEL,                    ctrl_get_level},
+
+  { -1, NULL},
+};
+
+static vpx_codec_enc_cfg_map_t encoder_usage_cfg_map[] = {
+  {
+    0,
+    {  // NOLINT
+      0,                  // g_usage
+      8,                  // g_threads
+      0,                  // g_profile
+
+      320,                // g_width
+      240,                // g_height
+      VPX_BITS_8,         // g_bit_depth
+      8,                  // g_input_bit_depth
+
+      {1, 30},            // g_timebase
+
+      0,                  // g_error_resilient
+
+      VPX_RC_ONE_PASS,    // g_pass
+
+      25,                 // g_lag_in_frames
+
+      0,                  // rc_dropframe_thresh
+      0,                  // rc_resize_allowed
+      0,                  // rc_scaled_width
+      0,                  // rc_scaled_height
+      60,                 // rc_resize_down_thresold
+      30,                 // rc_resize_up_thresold
+
+      VPX_VBR,            // rc_end_usage
+      {NULL, 0},          // rc_twopass_stats_in
+      {NULL, 0},          // rc_firstpass_mb_stats_in
+      256,                // rc_target_bandwidth
+      0,                  // rc_min_quantizer
+      63,                 // rc_max_quantizer
+      25,                 // rc_undershoot_pct
+      25,                 // rc_overshoot_pct
+
+      6000,               // rc_max_buffer_size
+      4000,               // rc_buffer_initial_size
+      5000,               // rc_buffer_optimal_size
+
+      50,                 // rc_two_pass_vbrbias
+      0,                  // rc_two_pass_vbrmin_section
+      2000,               // rc_two_pass_vbrmax_section
+
+      // keyframing settings (kf)
+      VPX_KF_AUTO,        // g_kfmode
+      0,                  // kf_min_dist
+      128,                // kf_max_dist
+
+      VPX_SS_DEFAULT_LAYERS,  // ss_number_layers
+      {0},
+      {0},                    // ss_target_bitrate
+      1,                      // ts_number_layers
+      {0},                    // ts_target_bitrate
+      {0},                    // ts_rate_decimator
+      0,                      // ts_periodicity
+      {0},                    // ts_layer_id
+      {0},                  // layer_taget_bitrate
+      0                     // temporal_layering_mode
+    }
+  },
+};
+
+#ifndef VERSION_STRING
+#define VERSION_STRING
+#endif
+CODEC_INTERFACE(vpx_codec_vp9_cx) = {
+  "WebM Project VP9 Encoder" VERSION_STRING,
+  VPX_CODEC_INTERNAL_ABI_VERSION,
+#if CONFIG_VP9_HIGHBITDEPTH
+  VPX_CODEC_CAP_HIGHBITDEPTH |
+#endif
+  VPX_CODEC_CAP_ENCODER | VPX_CODEC_CAP_PSNR,  // vpx_codec_caps_t
+  encoder_init,       // vpx_codec_init_fn_t
+  encoder_destroy,    // vpx_codec_destroy_fn_t
+  encoder_ctrl_maps,  // vpx_codec_ctrl_fn_map_t
+  {  // NOLINT
+    NULL,  // vpx_codec_peek_si_fn_t
+    NULL,  // vpx_codec_get_si_fn_t
+    NULL,  // vpx_codec_decode_fn_t
+    NULL,  // vpx_codec_frame_get_fn_t
+    NULL   // vpx_codec_set_fb_fn_t
+  },
+  {  // NOLINT
+    1,                      // 1 cfg map
+    encoder_usage_cfg_map,  // vpx_codec_enc_cfg_map_t
+    encoder_encode,         // vpx_codec_encode_fn_t
+    encoder_get_cxdata,     // vpx_codec_get_cx_data_fn_t
+    encoder_set_config,     // vpx_codec_enc_config_set_fn_t
+    NULL,        // vpx_codec_get_global_headers_fn_t
+    encoder_get_preview,    // vpx_codec_get_preview_frame_fn_t
+    NULL         // vpx_codec_enc_mr_get_mem_loc_fn_t
+  }
+};
diff --git a/libvpx/libvpx/vp9/vp9_dx_iface.c b/libvpx/libvpx/vp9/vp9_dx_iface.c
new file mode 100644
index 0000000..6531e2c
--- /dev/null
+++ b/libvpx/libvpx/vp9/vp9_dx_iface.c
@@ -0,0 +1,1093 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "./vpx_config.h"
+#include "./vpx_version.h"
+
+#include "vpx/internal/vpx_codec_internal.h"
+#include "vpx/vp8dx.h"
+#include "vpx/vpx_decoder.h"
+#include "vpx_dsp/bitreader_buffer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_util/vpx_thread.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_frame_buffers.h"
+
+#include "vp9/decoder/vp9_decodeframe.h"
+
+#include "vp9/vp9_dx_iface.h"
+#include "vp9/vp9_iface_common.h"
+
+#define VP9_CAP_POSTPROC (CONFIG_VP9_POSTPROC ? VPX_CODEC_CAP_POSTPROC : 0)
+
+static vpx_codec_err_t decoder_init(vpx_codec_ctx_t *ctx,
+                                    vpx_codec_priv_enc_mr_cfg_t *data) {
+  // This function only allocates space for the vpx_codec_alg_priv_t
+  // structure. More memory may be required at the time the stream
+  // information becomes known.
+  (void)data;
+
+  if (!ctx->priv) {
+    vpx_codec_alg_priv_t *const priv =
+        (vpx_codec_alg_priv_t *)vpx_calloc(1, sizeof(*priv));
+    if (priv == NULL)
+      return VPX_CODEC_MEM_ERROR;
+
+    ctx->priv = (vpx_codec_priv_t *)priv;
+    ctx->priv->init_flags = ctx->init_flags;
+    priv->si.sz = sizeof(priv->si);
+    priv->flushed = 0;
+    // Only do frame parallel decode when threads > 1.
+    priv->frame_parallel_decode =
+        (ctx->config.dec && (ctx->config.dec->threads > 1) &&
+         (ctx->init_flags & VPX_CODEC_USE_FRAME_THREADING)) ? 1 : 0;
+    if (ctx->config.dec) {
+      priv->cfg = *ctx->config.dec;
+      ctx->config.dec = &priv->cfg;
+    }
+  }
+
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t decoder_destroy(vpx_codec_alg_priv_t *ctx) {
+  if (ctx->frame_workers != NULL) {
+    int i;
+    for (i = 0; i < ctx->num_frame_workers; ++i) {
+      VPxWorker *const worker = &ctx->frame_workers[i];
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      vpx_get_worker_interface()->end(worker);
+      vp9_remove_common(&frame_worker_data->pbi->common);
+#if CONFIG_VP9_POSTPROC
+      vp9_free_postproc_buffers(&frame_worker_data->pbi->common);
+#endif
+      vp9_decoder_remove(frame_worker_data->pbi);
+      vpx_free(frame_worker_data->scratch_buffer);
+#if CONFIG_MULTITHREAD
+      pthread_mutex_destroy(&frame_worker_data->stats_mutex);
+      pthread_cond_destroy(&frame_worker_data->stats_cond);
+#endif
+      vpx_free(frame_worker_data);
+    }
+#if CONFIG_MULTITHREAD
+    pthread_mutex_destroy(&ctx->buffer_pool->pool_mutex);
+#endif
+  }
+
+  if (ctx->buffer_pool) {
+    vp9_free_ref_frame_buffers(ctx->buffer_pool);
+    vp9_free_internal_frame_buffers(&ctx->buffer_pool->int_frame_buffers);
+  }
+
+  vpx_free(ctx->frame_workers);
+  vpx_free(ctx->buffer_pool);
+  vpx_free(ctx);
+  return VPX_CODEC_OK;
+}
+
+static int parse_bitdepth_colorspace_sampling(
+    BITSTREAM_PROFILE profile, struct vpx_read_bit_buffer *rb) {
+  vpx_color_space_t color_space;
+  if (profile >= PROFILE_2)
+    rb->bit_offset += 1;  // Bit-depth 10 or 12.
+  color_space = (vpx_color_space_t)vpx_rb_read_literal(rb, 3);
+  if (color_space != VPX_CS_SRGB) {
+    rb->bit_offset += 1;  // [16,235] (including xvycc) vs [0,255] range.
+    if (profile == PROFILE_1 || profile == PROFILE_3) {
+      rb->bit_offset += 2;  // subsampling x/y.
+      rb->bit_offset += 1;  // unused.
+    }
+  } else {
+    if (profile == PROFILE_1 || profile == PROFILE_3) {
+      rb->bit_offset += 1;  // unused
+    } else {
+      // RGB is only available in version 1.
+      return 0;
+    }
+  }
+  return 1;
+}
+
+static vpx_codec_err_t decoder_peek_si_internal(const uint8_t *data,
+                                                unsigned int data_sz,
+                                                vpx_codec_stream_info_t *si,
+                                                int *is_intra_only,
+                                                vpx_decrypt_cb decrypt_cb,
+                                                void *decrypt_state) {
+  int intra_only_flag = 0;
+  uint8_t clear_buffer[10];
+
+  if (data + data_sz <= data)
+    return VPX_CODEC_INVALID_PARAM;
+
+  si->is_kf = 0;
+  si->w = si->h = 0;
+
+  if (decrypt_cb) {
+    data_sz = VPXMIN(sizeof(clear_buffer), data_sz);
+    decrypt_cb(decrypt_state, data, clear_buffer, data_sz);
+    data = clear_buffer;
+  }
+
+  // A maximum of 6 bits are needed to read the frame marker, profile and
+  // show_existing_frame.
+  if (data_sz < 1)
+    return VPX_CODEC_UNSUP_BITSTREAM;
+
+  {
+    int show_frame;
+    int error_resilient;
+    struct vpx_read_bit_buffer rb = { data, data + data_sz, 0, NULL, NULL };
+    const int frame_marker = vpx_rb_read_literal(&rb, 2);
+    const BITSTREAM_PROFILE profile = vp9_read_profile(&rb);
+
+    if (frame_marker != VP9_FRAME_MARKER)
+      return VPX_CODEC_UNSUP_BITSTREAM;
+
+    if (profile >= MAX_PROFILES)
+      return VPX_CODEC_UNSUP_BITSTREAM;
+
+    if (vpx_rb_read_bit(&rb)) {  // show an existing frame
+      // If profile is > 2 and show_existing_frame is true, then at least 1 more
+      // byte (6+3=9 bits) is needed.
+      if (profile > 2 && data_sz < 2)
+        return VPX_CODEC_UNSUP_BITSTREAM;
+      vpx_rb_read_literal(&rb, 3);  // Frame buffer to show.
+      return VPX_CODEC_OK;
+    }
+
+    // For the rest of the function, a maximum of 9 more bytes are needed
+    // (computed by taking the maximum possible bits needed in each case). Note
+    // that this has to be updated if we read any more bits in this function.
+    if (data_sz < 10)
+      return VPX_CODEC_UNSUP_BITSTREAM;
+
+    si->is_kf = !vpx_rb_read_bit(&rb);
+    show_frame = vpx_rb_read_bit(&rb);
+    error_resilient = vpx_rb_read_bit(&rb);
+
+    if (si->is_kf) {
+      if (!vp9_read_sync_code(&rb))
+        return VPX_CODEC_UNSUP_BITSTREAM;
+
+      if (!parse_bitdepth_colorspace_sampling(profile, &rb))
+        return VPX_CODEC_UNSUP_BITSTREAM;
+      vp9_read_frame_size(&rb, (int *)&si->w, (int *)&si->h);
+    } else {
+      intra_only_flag = show_frame ? 0 : vpx_rb_read_bit(&rb);
+
+      rb.bit_offset += error_resilient ? 0 : 2;  // reset_frame_context
+
+      if (intra_only_flag) {
+        if (!vp9_read_sync_code(&rb))
+          return VPX_CODEC_UNSUP_BITSTREAM;
+        if (profile > PROFILE_0) {
+          if (!parse_bitdepth_colorspace_sampling(profile, &rb))
+            return VPX_CODEC_UNSUP_BITSTREAM;
+        }
+        rb.bit_offset += REF_FRAMES;  // refresh_frame_flags
+        vp9_read_frame_size(&rb, (int *)&si->w, (int *)&si->h);
+      }
+    }
+  }
+  if (is_intra_only != NULL)
+    *is_intra_only = intra_only_flag;
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t decoder_peek_si(const uint8_t *data,
+                                       unsigned int data_sz,
+                                       vpx_codec_stream_info_t *si) {
+  return decoder_peek_si_internal(data, data_sz, si, NULL, NULL, NULL);
+}
+
+static vpx_codec_err_t decoder_get_si(vpx_codec_alg_priv_t *ctx,
+                                      vpx_codec_stream_info_t *si) {
+  const size_t sz = (si->sz >= sizeof(vp9_stream_info_t))
+                       ? sizeof(vp9_stream_info_t)
+                       : sizeof(vpx_codec_stream_info_t);
+  memcpy(si, &ctx->si, sz);
+  si->sz = (unsigned int)sz;
+
+  return VPX_CODEC_OK;
+}
+
+static void set_error_detail(vpx_codec_alg_priv_t *ctx,
+                             const char *const error) {
+  ctx->base.err_detail = error;
+}
+
+static vpx_codec_err_t update_error_state(vpx_codec_alg_priv_t *ctx,
+                           const struct vpx_internal_error_info *error) {
+  if (error->error_code)
+    set_error_detail(ctx, error->has_detail ? error->detail : NULL);
+
+  return error->error_code;
+}
+
+static void init_buffer_callbacks(vpx_codec_alg_priv_t *ctx) {
+  int i;
+
+  for (i = 0; i < ctx->num_frame_workers; ++i) {
+    VPxWorker *const worker = &ctx->frame_workers[i];
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    VP9_COMMON *const cm = &frame_worker_data->pbi->common;
+    BufferPool *const pool = cm->buffer_pool;
+
+    cm->new_fb_idx = INVALID_IDX;
+    cm->byte_alignment = ctx->byte_alignment;
+    cm->skip_loop_filter = ctx->skip_loop_filter;
+
+    if (ctx->get_ext_fb_cb != NULL && ctx->release_ext_fb_cb != NULL) {
+      pool->get_fb_cb = ctx->get_ext_fb_cb;
+      pool->release_fb_cb = ctx->release_ext_fb_cb;
+      pool->cb_priv = ctx->ext_priv;
+    } else {
+      pool->get_fb_cb = vp9_get_frame_buffer;
+      pool->release_fb_cb = vp9_release_frame_buffer;
+
+      if (vp9_alloc_internal_frame_buffers(&pool->int_frame_buffers))
+        vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                           "Failed to initialize internal frame buffers");
+
+      pool->cb_priv = &pool->int_frame_buffers;
+    }
+  }
+}
+
+static void set_default_ppflags(vp8_postproc_cfg_t *cfg) {
+  cfg->post_proc_flag = VP8_DEBLOCK | VP8_DEMACROBLOCK;
+  cfg->deblocking_level = 4;
+  cfg->noise_level = 0;
+}
+
+static void set_ppflags(const vpx_codec_alg_priv_t *ctx,
+                        vp9_ppflags_t *flags) {
+  flags->post_proc_flag =
+      ctx->postproc_cfg.post_proc_flag;
+
+  flags->deblocking_level = ctx->postproc_cfg.deblocking_level;
+  flags->noise_level = ctx->postproc_cfg.noise_level;
+}
+
+static int frame_worker_hook(void *arg1, void *arg2) {
+  FrameWorkerData *const frame_worker_data = (FrameWorkerData *)arg1;
+  const uint8_t *data = frame_worker_data->data;
+  (void)arg2;
+
+  frame_worker_data->result =
+      vp9_receive_compressed_data(frame_worker_data->pbi,
+                                  frame_worker_data->data_size,
+                                  &data);
+  frame_worker_data->data_end = data;
+
+  if (frame_worker_data->pbi->frame_parallel_decode) {
+    // In frame parallel decoding, a worker thread must successfully decode all
+    // the compressed data.
+    if (frame_worker_data->result != 0 ||
+        frame_worker_data->data + frame_worker_data->data_size - 1 > data) {
+      VPxWorker *const worker = frame_worker_data->pbi->frame_worker_owner;
+      BufferPool *const pool = frame_worker_data->pbi->common.buffer_pool;
+      // Signal all the other threads that are waiting for this frame.
+      vp9_frameworker_lock_stats(worker);
+      frame_worker_data->frame_context_ready = 1;
+      lock_buffer_pool(pool);
+      frame_worker_data->pbi->cur_buf->buf.corrupted = 1;
+      unlock_buffer_pool(pool);
+      frame_worker_data->pbi->need_resync = 1;
+      vp9_frameworker_signal_stats(worker);
+      vp9_frameworker_unlock_stats(worker);
+      return 0;
+    }
+  } else if (frame_worker_data->result != 0) {
+    // Check decode result in serial decode.
+    frame_worker_data->pbi->cur_buf->buf.corrupted = 1;
+    frame_worker_data->pbi->need_resync = 1;
+  }
+  return !frame_worker_data->result;
+}
+
+static vpx_codec_err_t init_decoder(vpx_codec_alg_priv_t *ctx) {
+  int i;
+  const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
+
+  ctx->last_show_frame = -1;
+  ctx->next_submit_worker_id = 0;
+  ctx->last_submit_worker_id = 0;
+  ctx->next_output_worker_id = 0;
+  ctx->frame_cache_read = 0;
+  ctx->frame_cache_write = 0;
+  ctx->num_cache_frames = 0;
+  ctx->need_resync = 1;
+  ctx->num_frame_workers =
+      (ctx->frame_parallel_decode == 1) ? ctx->cfg.threads: 1;
+  if (ctx->num_frame_workers > MAX_DECODE_THREADS)
+    ctx->num_frame_workers = MAX_DECODE_THREADS;
+  ctx->available_threads = ctx->num_frame_workers;
+  ctx->flushed = 0;
+
+  ctx->buffer_pool = (BufferPool *)vpx_calloc(1, sizeof(BufferPool));
+  if (ctx->buffer_pool == NULL)
+    return VPX_CODEC_MEM_ERROR;
+
+#if CONFIG_MULTITHREAD
+    if (pthread_mutex_init(&ctx->buffer_pool->pool_mutex, NULL)) {
+      set_error_detail(ctx, "Failed to allocate buffer pool mutex");
+      return VPX_CODEC_MEM_ERROR;
+    }
+#endif
+
+  ctx->frame_workers = (VPxWorker *)
+      vpx_malloc(ctx->num_frame_workers * sizeof(*ctx->frame_workers));
+  if (ctx->frame_workers == NULL) {
+    set_error_detail(ctx, "Failed to allocate frame_workers");
+    return VPX_CODEC_MEM_ERROR;
+  }
+
+  for (i = 0; i < ctx->num_frame_workers; ++i) {
+    VPxWorker *const worker = &ctx->frame_workers[i];
+    FrameWorkerData *frame_worker_data = NULL;
+    winterface->init(worker);
+    worker->data1 = vpx_memalign(32, sizeof(FrameWorkerData));
+    if (worker->data1 == NULL) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data");
+      return VPX_CODEC_MEM_ERROR;
+    }
+    frame_worker_data = (FrameWorkerData *)worker->data1;
+    frame_worker_data->pbi = vp9_decoder_create(ctx->buffer_pool);
+    if (frame_worker_data->pbi == NULL) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data");
+      return VPX_CODEC_MEM_ERROR;
+    }
+    frame_worker_data->pbi->frame_worker_owner = worker;
+    frame_worker_data->worker_id = i;
+    frame_worker_data->scratch_buffer = NULL;
+    frame_worker_data->scratch_buffer_size = 0;
+    frame_worker_data->frame_context_ready = 0;
+    frame_worker_data->received_frame = 0;
+#if CONFIG_MULTITHREAD
+    if (pthread_mutex_init(&frame_worker_data->stats_mutex, NULL)) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data mutex");
+      return VPX_CODEC_MEM_ERROR;
+    }
+
+    if (pthread_cond_init(&frame_worker_data->stats_cond, NULL)) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data cond");
+      return VPX_CODEC_MEM_ERROR;
+    }
+#endif
+    // If decoding in serial mode, FrameWorker thread could create tile worker
+    // thread or loopfilter thread.
+    frame_worker_data->pbi->max_threads =
+        (ctx->frame_parallel_decode == 0) ? ctx->cfg.threads : 0;
+
+    frame_worker_data->pbi->inv_tile_order = ctx->invert_tile_order;
+    frame_worker_data->pbi->frame_parallel_decode = ctx->frame_parallel_decode;
+    frame_worker_data->pbi->common.frame_parallel_decode =
+        ctx->frame_parallel_decode;
+    worker->hook = (VPxWorkerHook)frame_worker_hook;
+    if (!winterface->reset(worker)) {
+      set_error_detail(ctx, "Frame Worker thread creation failed");
+      return VPX_CODEC_MEM_ERROR;
+    }
+  }
+
+  // If postprocessing was enabled by the application and a
+  // configuration has not been provided, default it.
+  if (!ctx->postproc_cfg_set &&
+      (ctx->base.init_flags & VPX_CODEC_USE_POSTPROC))
+    set_default_ppflags(&ctx->postproc_cfg);
+
+  init_buffer_callbacks(ctx);
+
+  return VPX_CODEC_OK;
+}
+
+static INLINE void check_resync(vpx_codec_alg_priv_t *const ctx,
+                                const VP9Decoder *const pbi) {
+  // Clear resync flag if worker got a key frame or intra only frame.
+  if (ctx->need_resync == 1 && pbi->need_resync == 0 &&
+      (pbi->common.intra_only || pbi->common.frame_type == KEY_FRAME))
+    ctx->need_resync = 0;
+}
+
+static vpx_codec_err_t decode_one(vpx_codec_alg_priv_t *ctx,
+                                  const uint8_t **data, unsigned int data_sz,
+                                  void *user_priv, int64_t deadline) {
+  const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
+  (void)deadline;
+
+  // Determine the stream parameters. Note that we rely on peek_si to
+  // validate that we have a buffer that does not wrap around the top
+  // of the heap.
+  if (!ctx->si.h) {
+    int is_intra_only = 0;
+    const vpx_codec_err_t res =
+        decoder_peek_si_internal(*data, data_sz, &ctx->si, &is_intra_only,
+                                 ctx->decrypt_cb, ctx->decrypt_state);
+    if (res != VPX_CODEC_OK)
+      return res;
+
+    if (!ctx->si.is_kf && !is_intra_only)
+      return VPX_CODEC_ERROR;
+  }
+
+  if (!ctx->frame_parallel_decode) {
+    VPxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    frame_worker_data->data = *data;
+    frame_worker_data->data_size = data_sz;
+    frame_worker_data->user_priv = user_priv;
+    frame_worker_data->received_frame = 1;
+
+    // Set these even if already initialized.  The caller may have changed the
+    // decrypt config between frames.
+    frame_worker_data->pbi->decrypt_cb = ctx->decrypt_cb;
+    frame_worker_data->pbi->decrypt_state = ctx->decrypt_state;
+
+    worker->had_error = 0;
+    winterface->execute(worker);
+
+    // Update data pointer after decode.
+    *data = frame_worker_data->data_end;
+
+    if (worker->had_error)
+      return update_error_state(ctx, &frame_worker_data->pbi->common.error);
+
+    check_resync(ctx, frame_worker_data->pbi);
+  } else {
+    VPxWorker *const worker = &ctx->frame_workers[ctx->next_submit_worker_id];
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    // Copy context from last worker thread to next worker thread.
+    if (ctx->next_submit_worker_id != ctx->last_submit_worker_id)
+      vp9_frameworker_copy_context(
+          &ctx->frame_workers[ctx->next_submit_worker_id],
+          &ctx->frame_workers[ctx->last_submit_worker_id]);
+
+    frame_worker_data->pbi->ready_for_new_data = 0;
+    // Copy the compressed data into worker's internal buffer.
+    // TODO(hkuang): Will all the workers allocate the same size
+    // as the size of the first intra frame be better? This will
+    // avoid too many deallocate and allocate.
+    if (frame_worker_data->scratch_buffer_size < data_sz) {
+      frame_worker_data->scratch_buffer =
+          (uint8_t *)vpx_realloc(frame_worker_data->scratch_buffer, data_sz);
+      if (frame_worker_data->scratch_buffer == NULL) {
+        set_error_detail(ctx, "Failed to reallocate scratch buffer");
+        return VPX_CODEC_MEM_ERROR;
+      }
+      frame_worker_data->scratch_buffer_size = data_sz;
+    }
+    frame_worker_data->data_size = data_sz;
+    memcpy(frame_worker_data->scratch_buffer, *data, data_sz);
+
+    frame_worker_data->frame_decoded = 0;
+    frame_worker_data->frame_context_ready = 0;
+    frame_worker_data->received_frame = 1;
+    frame_worker_data->data = frame_worker_data->scratch_buffer;
+    frame_worker_data->user_priv = user_priv;
+
+    if (ctx->next_submit_worker_id != ctx->last_submit_worker_id)
+      ctx->last_submit_worker_id =
+          (ctx->last_submit_worker_id + 1) % ctx->num_frame_workers;
+
+    ctx->next_submit_worker_id =
+        (ctx->next_submit_worker_id + 1) % ctx->num_frame_workers;
+    --ctx->available_threads;
+    worker->had_error = 0;
+    winterface->launch(worker);
+  }
+
+  return VPX_CODEC_OK;
+}
+
+static void wait_worker_and_cache_frame(vpx_codec_alg_priv_t *ctx) {
+  YV12_BUFFER_CONFIG sd;
+  vp9_ppflags_t flags = {0, 0, 0};
+  const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
+  VPxWorker *const worker = &ctx->frame_workers[ctx->next_output_worker_id];
+  FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+  ctx->next_output_worker_id =
+      (ctx->next_output_worker_id + 1) % ctx->num_frame_workers;
+  // TODO(hkuang): Add worker error handling here.
+  winterface->sync(worker);
+  frame_worker_data->received_frame = 0;
+  ++ctx->available_threads;
+
+  check_resync(ctx, frame_worker_data->pbi);
+
+  if (vp9_get_raw_frame(frame_worker_data->pbi, &sd, &flags) == 0) {
+    VP9_COMMON *const cm = &frame_worker_data->pbi->common;
+    RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+    ctx->frame_cache[ctx->frame_cache_write].fb_idx = cm->new_fb_idx;
+    yuvconfig2image(&ctx->frame_cache[ctx->frame_cache_write].img, &sd,
+                    frame_worker_data->user_priv);
+    ctx->frame_cache[ctx->frame_cache_write].img.fb_priv =
+        frame_bufs[cm->new_fb_idx].raw_frame_buffer.priv;
+    ctx->frame_cache_write =
+        (ctx->frame_cache_write + 1) % FRAME_CACHE_SIZE;
+    ++ctx->num_cache_frames;
+  }
+}
+
+static vpx_codec_err_t decoder_decode(vpx_codec_alg_priv_t *ctx,
+                                      const uint8_t *data, unsigned int data_sz,
+                                      void *user_priv, long deadline) {
+  const uint8_t *data_start = data;
+  const uint8_t * const data_end = data + data_sz;
+  vpx_codec_err_t res;
+  uint32_t frame_sizes[8];
+  int frame_count;
+
+  if (data == NULL && data_sz == 0) {
+    ctx->flushed = 1;
+    return VPX_CODEC_OK;
+  }
+
+  // Reset flushed when receiving a valid frame.
+  ctx->flushed = 0;
+
+  // Initialize the decoder workers on the first frame.
+  if (ctx->frame_workers == NULL) {
+    const vpx_codec_err_t res = init_decoder(ctx);
+    if (res != VPX_CODEC_OK)
+      return res;
+  }
+
+  res = vp9_parse_superframe_index(data, data_sz, frame_sizes, &frame_count,
+                                   ctx->decrypt_cb, ctx->decrypt_state);
+  if (res != VPX_CODEC_OK)
+    return res;
+
+  if (ctx->frame_parallel_decode) {
+    // Decode in frame parallel mode. When decoding in this mode, the frame
+    // passed to the decoder must be either a normal frame or a superframe with
+    // superframe index so the decoder could get each frame's start position
+    // in the superframe.
+    if (frame_count > 0) {
+      int i;
+
+      for (i = 0; i < frame_count; ++i) {
+        const uint8_t *data_start_copy = data_start;
+        const uint32_t frame_size = frame_sizes[i];
+        if (data_start < data
+            || frame_size > (uint32_t) (data_end - data_start)) {
+          set_error_detail(ctx, "Invalid frame size in index");
+          return VPX_CODEC_CORRUPT_FRAME;
+        }
+
+        if (ctx->available_threads == 0) {
+          // No more threads for decoding. Wait until the next output worker
+          // finishes decoding. Then copy the decoded frame into cache.
+          if (ctx->num_cache_frames < FRAME_CACHE_SIZE) {
+            wait_worker_and_cache_frame(ctx);
+          } else {
+            // TODO(hkuang): Add unit test to test this path.
+            set_error_detail(ctx, "Frame output cache is full.");
+            return VPX_CODEC_ERROR;
+          }
+        }
+
+        res = decode_one(ctx, &data_start_copy, frame_size, user_priv,
+                         deadline);
+        if (res != VPX_CODEC_OK)
+          return res;
+        data_start += frame_size;
+      }
+    } else {
+      if (ctx->available_threads == 0) {
+        // No more threads for decoding. Wait until the next output worker
+        // finishes decoding. Then copy the decoded frame into cache.
+        if (ctx->num_cache_frames < FRAME_CACHE_SIZE) {
+          wait_worker_and_cache_frame(ctx);
+        } else {
+          // TODO(hkuang): Add unit test to test this path.
+          set_error_detail(ctx, "Frame output cache is full.");
+          return VPX_CODEC_ERROR;
+        }
+      }
+
+      res = decode_one(ctx, &data, data_sz, user_priv, deadline);
+      if (res != VPX_CODEC_OK)
+        return res;
+    }
+  } else {
+    // Decode in serial mode.
+    if (frame_count > 0) {
+      int i;
+
+      for (i = 0; i < frame_count; ++i) {
+        const uint8_t *data_start_copy = data_start;
+        const uint32_t frame_size = frame_sizes[i];
+        vpx_codec_err_t res;
+        if (data_start < data
+            || frame_size > (uint32_t) (data_end - data_start)) {
+          set_error_detail(ctx, "Invalid frame size in index");
+          return VPX_CODEC_CORRUPT_FRAME;
+        }
+
+        res = decode_one(ctx, &data_start_copy, frame_size, user_priv,
+                         deadline);
+        if (res != VPX_CODEC_OK)
+          return res;
+
+        data_start += frame_size;
+      }
+    } else {
+      while (data_start < data_end) {
+        const uint32_t frame_size = (uint32_t) (data_end - data_start);
+        const vpx_codec_err_t res = decode_one(ctx, &data_start, frame_size,
+                                               user_priv, deadline);
+        if (res != VPX_CODEC_OK)
+          return res;
+
+        // Account for suboptimal termination by the encoder.
+        while (data_start < data_end) {
+          const uint8_t marker = read_marker(ctx->decrypt_cb,
+                                             ctx->decrypt_state, data_start);
+          if (marker)
+            break;
+          ++data_start;
+        }
+      }
+    }
+  }
+
+  return res;
+}
+
+static void release_last_output_frame(vpx_codec_alg_priv_t *ctx) {
+  RefCntBuffer *const frame_bufs = ctx->buffer_pool->frame_bufs;
+  // Decrease reference count of last output frame in frame parallel mode.
+  if (ctx->frame_parallel_decode && ctx->last_show_frame >= 0) {
+    BufferPool *const pool = ctx->buffer_pool;
+    lock_buffer_pool(pool);
+    decrease_ref_count(ctx->last_show_frame, frame_bufs, pool);
+    unlock_buffer_pool(pool);
+  }
+}
+
+static vpx_image_t *decoder_get_frame(vpx_codec_alg_priv_t *ctx,
+                                      vpx_codec_iter_t *iter) {
+  vpx_image_t *img = NULL;
+
+  // Only return frame when all the cpu are busy or
+  // application fluhsed the decoder in frame parallel decode.
+  if (ctx->frame_parallel_decode && ctx->available_threads > 0 &&
+      !ctx->flushed) {
+    return NULL;
+  }
+
+  // Output the frames in the cache first.
+  if (ctx->num_cache_frames > 0) {
+    release_last_output_frame(ctx);
+    ctx->last_show_frame  = ctx->frame_cache[ctx->frame_cache_read].fb_idx;
+    if (ctx->need_resync)
+      return NULL;
+    img = &ctx->frame_cache[ctx->frame_cache_read].img;
+    ctx->frame_cache_read = (ctx->frame_cache_read + 1) % FRAME_CACHE_SIZE;
+    --ctx->num_cache_frames;
+    return img;
+  }
+
+  // iter acts as a flip flop, so an image is only returned on the first
+  // call to get_frame.
+  if (*iter == NULL && ctx->frame_workers != NULL) {
+    do {
+      YV12_BUFFER_CONFIG sd;
+      vp9_ppflags_t flags = {0, 0, 0};
+      const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
+      VPxWorker *const worker =
+          &ctx->frame_workers[ctx->next_output_worker_id];
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      ctx->next_output_worker_id =
+          (ctx->next_output_worker_id + 1) % ctx->num_frame_workers;
+      if (ctx->base.init_flags & VPX_CODEC_USE_POSTPROC)
+        set_ppflags(ctx, &flags);
+      // Wait for the frame from worker thread.
+      if (winterface->sync(worker)) {
+        // Check if worker has received any frames.
+        if (frame_worker_data->received_frame == 1) {
+          ++ctx->available_threads;
+          frame_worker_data->received_frame = 0;
+          check_resync(ctx, frame_worker_data->pbi);
+        }
+        if (vp9_get_raw_frame(frame_worker_data->pbi, &sd, &flags) == 0) {
+          VP9_COMMON *const cm = &frame_worker_data->pbi->common;
+          RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+          release_last_output_frame(ctx);
+          ctx->last_show_frame = frame_worker_data->pbi->common.new_fb_idx;
+          if (ctx->need_resync)
+            return NULL;
+          yuvconfig2image(&ctx->img, &sd, frame_worker_data->user_priv);
+          ctx->img.fb_priv = frame_bufs[cm->new_fb_idx].raw_frame_buffer.priv;
+          img = &ctx->img;
+          return img;
+        }
+      } else {
+        // Decoding failed. Release the worker thread.
+        frame_worker_data->received_frame = 0;
+        ++ctx->available_threads;
+        ctx->need_resync = 1;
+        if (ctx->flushed != 1)
+          return NULL;
+      }
+    } while (ctx->next_output_worker_id != ctx->next_submit_worker_id);
+  }
+  return NULL;
+}
+
+static vpx_codec_err_t decoder_set_fb_fn(
+    vpx_codec_alg_priv_t *ctx,
+    vpx_get_frame_buffer_cb_fn_t cb_get,
+    vpx_release_frame_buffer_cb_fn_t cb_release, void *cb_priv) {
+  if (cb_get == NULL || cb_release == NULL) {
+    return VPX_CODEC_INVALID_PARAM;
+  } else if (ctx->frame_workers == NULL) {
+    // If the decoder has already been initialized, do not accept changes to
+    // the frame buffer functions.
+    ctx->get_ext_fb_cb = cb_get;
+    ctx->release_ext_fb_cb = cb_release;
+    ctx->ext_priv = cb_priv;
+    return VPX_CODEC_OK;
+  }
+
+  return VPX_CODEC_ERROR;
+}
+
+static vpx_codec_err_t ctrl_set_reference(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  vpx_ref_frame_t *const data = va_arg(args, vpx_ref_frame_t *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
+  if (data) {
+    vpx_ref_frame_t *const frame = (vpx_ref_frame_t *)data;
+    YV12_BUFFER_CONFIG sd;
+    VPxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    image2yuvconfig(&frame->img, &sd);
+    return vp9_set_reference_dec(&frame_worker_data->pbi->common,
+                                 (VP9_REFFRAME)frame->frame_type, &sd);
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+}
+
+static vpx_codec_err_t ctrl_copy_reference(vpx_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
+  if (data) {
+    vpx_ref_frame_t *frame = (vpx_ref_frame_t *) data;
+    YV12_BUFFER_CONFIG sd;
+    VPxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    image2yuvconfig(&frame->img, &sd);
+    return vp9_copy_reference_dec(frame_worker_data->pbi,
+                                  (VP9_REFFRAME)frame->frame_type, &sd);
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+}
+
+static vpx_codec_err_t ctrl_get_reference(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  vp9_ref_frame_t *data = va_arg(args, vp9_ref_frame_t *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
+  if (data) {
+    YV12_BUFFER_CONFIG* fb;
+    VPxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    fb = get_ref_frame(&frame_worker_data->pbi->common, data->idx);
+    if (fb == NULL) return VPX_CODEC_ERROR;
+    yuvconfig2image(&data->img, fb, NULL);
+    return VPX_CODEC_OK;
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+}
+
+static vpx_codec_err_t ctrl_set_postproc(vpx_codec_alg_priv_t *ctx,
+                                         va_list args) {
+#if CONFIG_VP9_POSTPROC
+  vp8_postproc_cfg_t *data = va_arg(args, vp8_postproc_cfg_t *);
+
+  if (data) {
+    ctx->postproc_cfg_set = 1;
+    ctx->postproc_cfg = *((vp8_postproc_cfg_t *)data);
+    return VPX_CODEC_OK;
+  } else {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+#else
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
+#endif
+}
+
+static vpx_codec_err_t ctrl_set_dbg_options(vpx_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  (void)ctx;
+  (void)args;
+  return VPX_CODEC_INCAPABLE;
+}
+
+static vpx_codec_err_t ctrl_get_last_ref_updates(vpx_codec_alg_priv_t *ctx,
+                                                 va_list args) {
+  int *const update_info = va_arg(args, int *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
+  if (update_info) {
+    if (ctx->frame_workers) {
+      VPxWorker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      *update_info = frame_worker_data->pbi->refresh_frame_flags;
+      return VPX_CODEC_OK;
+    } else {
+      return VPX_CODEC_ERROR;
+    }
+  }
+
+  return VPX_CODEC_INVALID_PARAM;
+}
+
+static vpx_codec_err_t ctrl_get_frame_corrupted(vpx_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  int *corrupted = va_arg(args, int *);
+
+  if (corrupted) {
+    if (ctx->frame_workers) {
+      VPxWorker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      RefCntBuffer *const frame_bufs =
+          frame_worker_data->pbi->common.buffer_pool->frame_bufs;
+      if (frame_worker_data->pbi->common.frame_to_show == NULL)
+        return VPX_CODEC_ERROR;
+      if (ctx->last_show_frame >= 0)
+        *corrupted = frame_bufs[ctx->last_show_frame].buf.corrupted;
+      return VPX_CODEC_OK;
+    } else {
+      return VPX_CODEC_ERROR;
+    }
+  }
+
+  return VPX_CODEC_INVALID_PARAM;
+}
+
+static vpx_codec_err_t ctrl_get_frame_size(vpx_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  int *const frame_size = va_arg(args, int *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
+  if (frame_size) {
+    if (ctx->frame_workers) {
+      VPxWorker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      const VP9_COMMON *const cm = &frame_worker_data->pbi->common;
+      frame_size[0] = cm->width;
+      frame_size[1] = cm->height;
+      return VPX_CODEC_OK;
+    } else {
+      return VPX_CODEC_ERROR;
+    }
+  }
+
+  return VPX_CODEC_INVALID_PARAM;
+}
+
+static vpx_codec_err_t ctrl_get_render_size(vpx_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  int *const render_size = va_arg(args, int *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
+  if (render_size) {
+    if (ctx->frame_workers) {
+      VPxWorker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      const VP9_COMMON *const cm = &frame_worker_data->pbi->common;
+      render_size[0] = cm->render_width;
+      render_size[1] = cm->render_height;
+      return VPX_CODEC_OK;
+    } else {
+      return VPX_CODEC_ERROR;
+    }
+  }
+
+  return VPX_CODEC_INVALID_PARAM;
+}
+
+static vpx_codec_err_t ctrl_get_bit_depth(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  unsigned int *const bit_depth = va_arg(args, unsigned int *);
+  VPxWorker *const worker = &ctx->frame_workers[ctx->next_output_worker_id];
+
+  if (bit_depth) {
+    if (worker) {
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      const VP9_COMMON *const cm = &frame_worker_data->pbi->common;
+      *bit_depth = cm->bit_depth;
+      return VPX_CODEC_OK;
+    } else {
+      return VPX_CODEC_ERROR;
+    }
+  }
+
+  return VPX_CODEC_INVALID_PARAM;
+}
+
+static vpx_codec_err_t ctrl_set_invert_tile_order(vpx_codec_alg_priv_t *ctx,
+                                                  va_list args) {
+  ctx->invert_tile_order = va_arg(args, int);
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_set_decryptor(vpx_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  vpx_decrypt_init *init = va_arg(args, vpx_decrypt_init *);
+  ctx->decrypt_cb = init ? init->decrypt_cb : NULL;
+  ctx->decrypt_state = init ? init->decrypt_state : NULL;
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_set_byte_alignment(vpx_codec_alg_priv_t *ctx,
+                                               va_list args) {
+  const int legacy_byte_alignment = 0;
+  const int min_byte_alignment = 32;
+  const int max_byte_alignment = 1024;
+  const int byte_alignment = va_arg(args, int);
+
+  if (byte_alignment != legacy_byte_alignment &&
+      (byte_alignment < min_byte_alignment ||
+       byte_alignment > max_byte_alignment ||
+       (byte_alignment & (byte_alignment - 1)) != 0))
+    return VPX_CODEC_INVALID_PARAM;
+
+  ctx->byte_alignment = byte_alignment;
+  if (ctx->frame_workers) {
+    VPxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data =
+        (FrameWorkerData *)worker->data1;
+    frame_worker_data->pbi->common.byte_alignment = byte_alignment;
+  }
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t ctrl_set_skip_loop_filter(vpx_codec_alg_priv_t *ctx,
+                                                 va_list args) {
+  ctx->skip_loop_filter = va_arg(args, int);
+
+  if (ctx->frame_workers) {
+    VPxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    frame_worker_data->pbi->common.skip_loop_filter = ctx->skip_loop_filter;
+  }
+
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_ctrl_fn_map_t decoder_ctrl_maps[] = {
+  {VP8_COPY_REFERENCE,            ctrl_copy_reference},
+
+  // Setters
+  {VP8_SET_REFERENCE,             ctrl_set_reference},
+  {VP8_SET_POSTPROC,              ctrl_set_postproc},
+  {VP8_SET_DBG_COLOR_REF_FRAME,   ctrl_set_dbg_options},
+  {VP8_SET_DBG_COLOR_MB_MODES,    ctrl_set_dbg_options},
+  {VP8_SET_DBG_COLOR_B_MODES,     ctrl_set_dbg_options},
+  {VP8_SET_DBG_DISPLAY_MV,        ctrl_set_dbg_options},
+  {VP9_INVERT_TILE_DECODE_ORDER,  ctrl_set_invert_tile_order},
+  {VPXD_SET_DECRYPTOR,            ctrl_set_decryptor},
+  {VP9_SET_BYTE_ALIGNMENT,        ctrl_set_byte_alignment},
+  {VP9_SET_SKIP_LOOP_FILTER,      ctrl_set_skip_loop_filter},
+
+  // Getters
+  {VP8D_GET_LAST_REF_UPDATES,     ctrl_get_last_ref_updates},
+  {VP8D_GET_FRAME_CORRUPTED,      ctrl_get_frame_corrupted},
+  {VP9_GET_REFERENCE,             ctrl_get_reference},
+  {VP9D_GET_DISPLAY_SIZE,         ctrl_get_render_size},
+  {VP9D_GET_BIT_DEPTH,            ctrl_get_bit_depth},
+  {VP9D_GET_FRAME_SIZE,           ctrl_get_frame_size},
+
+  { -1, NULL},
+};
+
+#ifndef VERSION_STRING
+#define VERSION_STRING
+#endif
+CODEC_INTERFACE(vpx_codec_vp9_dx) = {
+  "WebM Project VP9 Decoder" VERSION_STRING,
+  VPX_CODEC_INTERNAL_ABI_VERSION,
+  VPX_CODEC_CAP_DECODER | VP9_CAP_POSTPROC |
+      VPX_CODEC_CAP_EXTERNAL_FRAME_BUFFER,  // vpx_codec_caps_t
+  decoder_init,       // vpx_codec_init_fn_t
+  decoder_destroy,    // vpx_codec_destroy_fn_t
+  decoder_ctrl_maps,  // vpx_codec_ctrl_fn_map_t
+  { // NOLINT
+    decoder_peek_si,    // vpx_codec_peek_si_fn_t
+    decoder_get_si,     // vpx_codec_get_si_fn_t
+    decoder_decode,     // vpx_codec_decode_fn_t
+    decoder_get_frame,  // vpx_codec_frame_get_fn_t
+    decoder_set_fb_fn,  // vpx_codec_set_fb_fn_t
+  },
+  { // NOLINT
+    0,
+    NULL,  // vpx_codec_enc_cfg_map_t
+    NULL,  // vpx_codec_encode_fn_t
+    NULL,  // vpx_codec_get_cx_data_fn_t
+    NULL,  // vpx_codec_enc_config_set_fn_t
+    NULL,  // vpx_codec_get_global_headers_fn_t
+    NULL,  // vpx_codec_get_preview_frame_fn_t
+    NULL   // vpx_codec_enc_mr_get_mem_loc_fn_t
+  }
+};
diff --git a/libvpx/libvpx/vp9/vp9_dx_iface.h b/libvpx/libvpx/vp9/vp9_dx_iface.h
new file mode 100644
index 0000000..e0e948e
--- /dev/null
+++ b/libvpx/libvpx/vp9/vp9_dx_iface.h
@@ -0,0 +1,65 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_VP9_DX_IFACE_H_
+#define VP9_VP9_DX_IFACE_H_
+
+#include "vp9/decoder/vp9_decoder.h"
+
+typedef vpx_codec_stream_info_t vp9_stream_info_t;
+
+// This limit is due to framebuffer numbers.
+// TODO(hkuang): Remove this limit after implementing ondemand framebuffers.
+#define FRAME_CACHE_SIZE 6   // Cache maximum 6 decoded frames.
+
+typedef struct cache_frame {
+  int fb_idx;
+  vpx_image_t img;
+} cache_frame;
+
+struct vpx_codec_alg_priv {
+  vpx_codec_priv_t        base;
+  vpx_codec_dec_cfg_t     cfg;
+  vp9_stream_info_t       si;
+  int                     postproc_cfg_set;
+  vp8_postproc_cfg_t      postproc_cfg;
+  vpx_decrypt_cb          decrypt_cb;
+  void                    *decrypt_state;
+  vpx_image_t             img;
+  int                     img_avail;
+  int                     flushed;
+  int                     invert_tile_order;
+  int                     last_show_frame;  // Index of last output frame.
+  int                     byte_alignment;
+  int                     skip_loop_filter;
+
+  // Frame parallel related.
+  int                     frame_parallel_decode;  // frame-based threading.
+  VPxWorker               *frame_workers;
+  int                     num_frame_workers;
+  int                     next_submit_worker_id;
+  int                     last_submit_worker_id;
+  int                     next_output_worker_id;
+  int                     available_threads;
+  cache_frame             frame_cache[FRAME_CACHE_SIZE];
+  int                     frame_cache_write;
+  int                     frame_cache_read;
+  int                     num_cache_frames;
+  int                     need_resync;      // wait for key/intra-only frame
+  // BufferPool that holds all reference frames. Shared by all the FrameWorkers.
+  BufferPool              *buffer_pool;
+
+  // External frame buffer info to save for VP9 common.
+  void *ext_priv;  // Private data associated with the external frame buffers.
+  vpx_get_frame_buffer_cb_fn_t get_ext_fb_cb;
+  vpx_release_frame_buffer_cb_fn_t release_ext_fb_cb;
+};
+
+#endif  // VP9_VP9_DX_IFACE_H_
diff --git a/libvpx/libvpx/vp9/vp9_iface_common.h b/libvpx/libvpx/vp9/vp9_iface_common.h
new file mode 100644
index 0000000..938d422
--- /dev/null
+++ b/libvpx/libvpx/vp9/vp9_iface_common.h
@@ -0,0 +1,136 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VP9_VP9_IFACE_COMMON_H_
+#define VP9_VP9_IFACE_COMMON_H_
+
+#include "vpx_ports/mem.h"
+
+static void yuvconfig2image(vpx_image_t *img, const YV12_BUFFER_CONFIG  *yv12,
+                            void *user_priv) {
+  /** vpx_img_wrap() doesn't allow specifying independent strides for
+    * the Y, U, and V planes, nor other alignment adjustments that
+    * might be representable by a YV12_BUFFER_CONFIG, so we just
+    * initialize all the fields.*/
+  int bps;
+  if (!yv12->subsampling_y) {
+    if (!yv12->subsampling_x) {
+      img->fmt = VPX_IMG_FMT_I444;
+      bps = 24;
+    } else {
+      img->fmt = VPX_IMG_FMT_I422;
+      bps = 16;
+    }
+  } else {
+    if (!yv12->subsampling_x) {
+      img->fmt = VPX_IMG_FMT_I440;
+      bps = 16;
+    } else {
+      img->fmt = VPX_IMG_FMT_I420;
+      bps = 12;
+    }
+  }
+  img->cs = yv12->color_space;
+  img->range = yv12->color_range;
+  img->bit_depth = 8;
+  img->w = yv12->y_stride;
+  img->h = ALIGN_POWER_OF_TWO(yv12->y_height + 2 * VP9_ENC_BORDER_IN_PIXELS, 3);
+  img->d_w = yv12->y_crop_width;
+  img->d_h = yv12->y_crop_height;
+  img->r_w = yv12->render_width;
+  img->r_h = yv12->render_height;
+  img->x_chroma_shift = yv12->subsampling_x;
+  img->y_chroma_shift = yv12->subsampling_y;
+  img->planes[VPX_PLANE_Y] = yv12->y_buffer;
+  img->planes[VPX_PLANE_U] = yv12->u_buffer;
+  img->planes[VPX_PLANE_V] = yv12->v_buffer;
+  img->planes[VPX_PLANE_ALPHA] = NULL;
+  img->stride[VPX_PLANE_Y] = yv12->y_stride;
+  img->stride[VPX_PLANE_U] = yv12->uv_stride;
+  img->stride[VPX_PLANE_V] = yv12->uv_stride;
+  img->stride[VPX_PLANE_ALPHA] = yv12->y_stride;
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (yv12->flags & YV12_FLAG_HIGHBITDEPTH) {
+    // vpx_image_t uses byte strides and a pointer to the first byte
+    // of the image.
+    img->fmt = (vpx_img_fmt_t)(img->fmt | VPX_IMG_FMT_HIGHBITDEPTH);
+    img->bit_depth = yv12->bit_depth;
+    img->planes[VPX_PLANE_Y] = (uint8_t*)CONVERT_TO_SHORTPTR(yv12->y_buffer);
+    img->planes[VPX_PLANE_U] = (uint8_t*)CONVERT_TO_SHORTPTR(yv12->u_buffer);
+    img->planes[VPX_PLANE_V] = (uint8_t*)CONVERT_TO_SHORTPTR(yv12->v_buffer);
+    img->planes[VPX_PLANE_ALPHA] = NULL;
+    img->stride[VPX_PLANE_Y] = 2 * yv12->y_stride;
+    img->stride[VPX_PLANE_U] = 2 * yv12->uv_stride;
+    img->stride[VPX_PLANE_V] = 2 * yv12->uv_stride;
+    img->stride[VPX_PLANE_ALPHA] = 2 * yv12->y_stride;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  img->bps = bps;
+  img->user_priv = user_priv;
+  img->img_data = yv12->buffer_alloc;
+  img->img_data_owner = 0;
+  img->self_allocd = 0;
+}
+
+static vpx_codec_err_t image2yuvconfig(const vpx_image_t *img,
+                                       YV12_BUFFER_CONFIG *yv12) {
+  yv12->y_buffer = img->planes[VPX_PLANE_Y];
+  yv12->u_buffer = img->planes[VPX_PLANE_U];
+  yv12->v_buffer = img->planes[VPX_PLANE_V];
+
+  yv12->y_crop_width  = img->d_w;
+  yv12->y_crop_height = img->d_h;
+  yv12->render_width  = img->r_w;
+  yv12->render_height = img->r_h;
+  yv12->y_width  = img->d_w;
+  yv12->y_height = img->d_h;
+
+  yv12->uv_width = img->x_chroma_shift == 1 ? (1 + yv12->y_width) / 2
+                                            : yv12->y_width;
+  yv12->uv_height = img->y_chroma_shift == 1 ? (1 + yv12->y_height) / 2
+                                             : yv12->y_height;
+  yv12->uv_crop_width = yv12->uv_width;
+  yv12->uv_crop_height = yv12->uv_height;
+
+  yv12->y_stride = img->stride[VPX_PLANE_Y];
+  yv12->uv_stride = img->stride[VPX_PLANE_U];
+  yv12->color_space = img->cs;
+  yv12->color_range = img->range;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (img->fmt & VPX_IMG_FMT_HIGHBITDEPTH) {
+    // In vpx_image_t
+    //     planes point to uint8 address of start of data
+    //     stride counts uint8s to reach next row
+    // In YV12_BUFFER_CONFIG
+    //     y_buffer, u_buffer, v_buffer point to uint16 address of data
+    //     stride and border counts in uint16s
+    // This means that all the address calculations in the main body of code
+    // should work correctly.
+    // However, before we do any pixel operations we need to cast the address
+    // to a uint16 ponter and double its value.
+    yv12->y_buffer = CONVERT_TO_BYTEPTR(yv12->y_buffer);
+    yv12->u_buffer = CONVERT_TO_BYTEPTR(yv12->u_buffer);
+    yv12->v_buffer = CONVERT_TO_BYTEPTR(yv12->v_buffer);
+    yv12->y_stride >>= 1;
+    yv12->uv_stride >>= 1;
+    yv12->flags = YV12_FLAG_HIGHBITDEPTH;
+  } else {
+    yv12->flags = 0;
+  }
+  yv12->border  = (yv12->y_stride - img->w) / 2;
+#else
+  yv12->border  = (img->stride[VPX_PLANE_Y] - img->w) / 2;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  yv12->subsampling_x = img->x_chroma_shift;
+  yv12->subsampling_y = img->y_chroma_shift;
+  return VPX_CODEC_OK;
+}
+
+#endif  // VP9_VP9_IFACE_COMMON_H_
diff --git a/libvpx/libvpx/vp9/vp9cx.mk b/libvpx/libvpx/vp9/vp9cx.mk
new file mode 100644
index 0000000..5f3de8f
--- /dev/null
+++ b/libvpx/libvpx/vp9/vp9cx.mk
@@ -0,0 +1,145 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+VP9_CX_EXPORTS += exports_enc
+
+VP9_CX_SRCS-yes += $(VP9_COMMON_SRCS-yes)
+VP9_CX_SRCS-no  += $(VP9_COMMON_SRCS-no)
+VP9_CX_SRCS_REMOVE-yes += $(VP9_COMMON_SRCS_REMOVE-yes)
+VP9_CX_SRCS_REMOVE-no  += $(VP9_COMMON_SRCS_REMOVE-no)
+
+VP9_CX_SRCS-yes += vp9_cx_iface.c
+
+VP9_CX_SRCS-yes += encoder/vp9_bitstream.c
+VP9_CX_SRCS-yes += encoder/vp9_context_tree.c
+VP9_CX_SRCS-yes += encoder/vp9_context_tree.h
+VP9_CX_SRCS-yes += encoder/vp9_cost.h
+VP9_CX_SRCS-yes += encoder/vp9_cost.c
+VP9_CX_SRCS-yes += encoder/vp9_dct.c
+VP9_CX_SRCS-$(CONFIG_VP9_TEMPORAL_DENOISING) += encoder/vp9_denoiser.c
+VP9_CX_SRCS-$(CONFIG_VP9_TEMPORAL_DENOISING) += encoder/vp9_denoiser.h
+VP9_CX_SRCS-yes += encoder/vp9_encodeframe.c
+VP9_CX_SRCS-yes += encoder/vp9_encodeframe.h
+VP9_CX_SRCS-yes += encoder/vp9_encodemb.c
+VP9_CX_SRCS-yes += encoder/vp9_encodemv.c
+VP9_CX_SRCS-yes += encoder/vp9_ethread.h
+VP9_CX_SRCS-yes += encoder/vp9_ethread.c
+VP9_CX_SRCS-yes += encoder/vp9_extend.c
+VP9_CX_SRCS-yes += encoder/vp9_firstpass.c
+VP9_CX_SRCS-yes += encoder/vp9_block.h
+VP9_CX_SRCS-yes += encoder/vp9_bitstream.h
+VP9_CX_SRCS-yes += encoder/vp9_encodemb.h
+VP9_CX_SRCS-yes += encoder/vp9_encodemv.h
+VP9_CX_SRCS-yes += encoder/vp9_extend.h
+VP9_CX_SRCS-yes += encoder/vp9_firstpass.h
+VP9_CX_SRCS-yes += encoder/vp9_lookahead.c
+VP9_CX_SRCS-yes += encoder/vp9_lookahead.h
+VP9_CX_SRCS-yes += encoder/vp9_mcomp.h
+VP9_CX_SRCS-yes += encoder/vp9_encoder.h
+VP9_CX_SRCS-yes += encoder/vp9_quantize.h
+VP9_CX_SRCS-yes += encoder/vp9_ratectrl.h
+VP9_CX_SRCS-yes += encoder/vp9_rd.h
+VP9_CX_SRCS-yes += encoder/vp9_rdopt.h
+VP9_CX_SRCS-yes += encoder/vp9_pickmode.h
+VP9_CX_SRCS-yes += encoder/vp9_svc_layercontext.h
+VP9_CX_SRCS-yes += encoder/vp9_tokenize.h
+VP9_CX_SRCS-yes += encoder/vp9_treewriter.h
+VP9_CX_SRCS-yes += encoder/vp9_mcomp.c
+VP9_CX_SRCS-yes += encoder/vp9_encoder.c
+VP9_CX_SRCS-yes += encoder/vp9_picklpf.c
+VP9_CX_SRCS-yes += encoder/vp9_picklpf.h
+VP9_CX_SRCS-yes += encoder/vp9_quantize.c
+VP9_CX_SRCS-yes += encoder/vp9_ratectrl.c
+VP9_CX_SRCS-yes += encoder/vp9_rd.c
+VP9_CX_SRCS-yes += encoder/vp9_rdopt.c
+VP9_CX_SRCS-yes += encoder/vp9_pickmode.c
+VP9_CX_SRCS-yes += encoder/vp9_segmentation.c
+VP9_CX_SRCS-yes += encoder/vp9_segmentation.h
+VP9_CX_SRCS-yes += encoder/vp9_speed_features.c
+VP9_CX_SRCS-yes += encoder/vp9_speed_features.h
+VP9_CX_SRCS-yes += encoder/vp9_subexp.c
+VP9_CX_SRCS-yes += encoder/vp9_subexp.h
+VP9_CX_SRCS-yes += encoder/vp9_svc_layercontext.c
+VP9_CX_SRCS-yes += encoder/vp9_resize.c
+VP9_CX_SRCS-yes += encoder/vp9_resize.h
+VP9_CX_SRCS-$(CONFIG_INTERNAL_STATS) += encoder/vp9_blockiness.c
+
+VP9_CX_SRCS-yes += encoder/vp9_tokenize.c
+VP9_CX_SRCS-yes += encoder/vp9_treewriter.c
+VP9_CX_SRCS-yes += encoder/vp9_aq_variance.c
+VP9_CX_SRCS-yes += encoder/vp9_aq_variance.h
+VP9_CX_SRCS-yes += encoder/vp9_aq_360.c
+VP9_CX_SRCS-yes += encoder/vp9_aq_360.h
+VP9_CX_SRCS-yes += encoder/vp9_aq_cyclicrefresh.c
+VP9_CX_SRCS-yes += encoder/vp9_aq_cyclicrefresh.h
+VP9_CX_SRCS-yes += encoder/vp9_aq_complexity.c
+VP9_CX_SRCS-yes += encoder/vp9_aq_complexity.h
+VP9_CX_SRCS-yes += encoder/vp9_skin_detection.c
+VP9_CX_SRCS-yes += encoder/vp9_skin_detection.h
+VP9_CX_SRCS-yes += encoder/vp9_noise_estimate.c
+VP9_CX_SRCS-yes += encoder/vp9_noise_estimate.h
+ifeq ($(CONFIG_VP9_POSTPROC),yes)
+VP9_CX_SRCS-$(CONFIG_INTERNAL_STATS) += common/vp9_postproc.h
+VP9_CX_SRCS-$(CONFIG_INTERNAL_STATS) += common/vp9_postproc.c
+endif
+VP9_CX_SRCS-yes += encoder/vp9_temporal_filter.c
+VP9_CX_SRCS-yes += encoder/vp9_temporal_filter.h
+VP9_CX_SRCS-yes += encoder/vp9_mbgraph.c
+VP9_CX_SRCS-yes += encoder/vp9_mbgraph.h
+
+VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_temporal_filter_apply_sse2.asm
+VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_quantize_sse2.c
+VP9_CX_SRCS-$(HAVE_AVX) += encoder/x86/vp9_diamond_search_sad_avx.c
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_highbd_block_error_intrin_sse2.c
+endif
+
+ifeq ($(CONFIG_USE_X86INC),yes)
+VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_dct_sse2.asm
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_highbd_error_sse2.asm
+VP9_CX_SRCS-$(HAVE_AVX) += encoder/x86/vp9_highbd_error_avx.asm
+else
+VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_error_sse2.asm
+endif
+endif
+
+ifeq ($(ARCH_X86_64),yes)
+ifeq ($(CONFIG_USE_X86INC),yes)
+VP9_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/vp9_quantize_ssse3_x86_64.asm
+endif
+endif
+
+VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_dct_intrin_sse2.c
+VP9_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/vp9_dct_ssse3.c
+ifneq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+VP9_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/vp9_frame_scale_ssse3.c
+endif
+
+ifeq ($(CONFIG_VP9_TEMPORAL_DENOISING),yes)
+VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_denoiser_sse2.c
+endif
+
+VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_error_intrin_avx2.c
+
+ifneq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_dct_neon.c
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_error_neon.c
+endif
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_quantize_neon.c
+
+VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_error_msa.c
+VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct4x4_msa.c
+VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct8x8_msa.c
+VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct16x16_msa.c
+VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct_msa.h
+VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_temporal_filter_msa.c
+
+VP9_CX_SRCS-yes := $(filter-out $(VP9_CX_SRCS_REMOVE-yes),$(VP9_CX_SRCS-yes))
diff --git a/libvpx/libvpx/vp9/vp9dx.mk b/libvpx/libvpx/vp9/vp9dx.mk
new file mode 100644
index 0000000..4c6fd00
--- /dev/null
+++ b/libvpx/libvpx/vp9/vp9dx.mk
@@ -0,0 +1,34 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+VP9_DX_EXPORTS += exports_dec
+
+VP9_DX_SRCS-yes += $(VP9_COMMON_SRCS-yes)
+VP9_DX_SRCS-no  += $(VP9_COMMON_SRCS-no)
+VP9_DX_SRCS_REMOVE-yes += $(VP9_COMMON_SRCS_REMOVE-yes)
+VP9_DX_SRCS_REMOVE-no  += $(VP9_COMMON_SRCS_REMOVE-no)
+
+VP9_DX_SRCS-yes += vp9_dx_iface.c
+VP9_DX_SRCS-yes += vp9_dx_iface.h
+
+VP9_DX_SRCS-yes += decoder/vp9_decodemv.c
+VP9_DX_SRCS-yes += decoder/vp9_decodeframe.c
+VP9_DX_SRCS-yes += decoder/vp9_decodeframe.h
+VP9_DX_SRCS-yes += decoder/vp9_detokenize.c
+VP9_DX_SRCS-yes += decoder/vp9_decodemv.h
+VP9_DX_SRCS-yes += decoder/vp9_detokenize.h
+VP9_DX_SRCS-yes += decoder/vp9_dthread.c
+VP9_DX_SRCS-yes += decoder/vp9_dthread.h
+VP9_DX_SRCS-yes += decoder/vp9_decoder.c
+VP9_DX_SRCS-yes += decoder/vp9_decoder.h
+VP9_DX_SRCS-yes += decoder/vp9_dsubexp.c
+VP9_DX_SRCS-yes += decoder/vp9_dsubexp.h
+
+VP9_DX_SRCS-yes := $(filter-out $(VP9_DX_SRCS_REMOVE-yes),$(VP9_DX_SRCS-yes))
diff --git a/libvpx/libvpx/vpx/exports_com b/libvpx/libvpx/vpx/exports_com
new file mode 100644
index 0000000..2ab0509
--- /dev/null
+++ b/libvpx/libvpx/vpx/exports_com
@@ -0,0 +1,16 @@
+text vpx_codec_build_config
+text vpx_codec_control_
+text vpx_codec_destroy
+text vpx_codec_err_to_string
+text vpx_codec_error
+text vpx_codec_error_detail
+text vpx_codec_get_caps
+text vpx_codec_iface_name
+text vpx_codec_version
+text vpx_codec_version_extra_str
+text vpx_codec_version_str
+text vpx_img_alloc
+text vpx_img_flip
+text vpx_img_free
+text vpx_img_set_rect
+text vpx_img_wrap
diff --git a/libvpx/libvpx/vpx/exports_dec b/libvpx/libvpx/vpx/exports_dec
new file mode 100644
index 0000000..c694eba
--- /dev/null
+++ b/libvpx/libvpx/vpx/exports_dec
@@ -0,0 +1,8 @@
+text vpx_codec_dec_init_ver
+text vpx_codec_decode
+text vpx_codec_get_frame
+text vpx_codec_get_stream_info
+text vpx_codec_peek_stream_info
+text vpx_codec_register_put_frame_cb
+text vpx_codec_register_put_slice_cb
+text vpx_codec_set_frame_buffer_functions
diff --git a/libvpx/libvpx/vpx/exports_enc b/libvpx/libvpx/vpx/exports_enc
new file mode 100644
index 0000000..914e36c
--- /dev/null
+++ b/libvpx/libvpx/vpx/exports_enc
@@ -0,0 +1,9 @@
+text vpx_codec_enc_config_default
+text vpx_codec_enc_config_set
+text vpx_codec_enc_init_multi_ver
+text vpx_codec_enc_init_ver
+text vpx_codec_encode
+text vpx_codec_get_cx_data
+text vpx_codec_get_global_headers
+text vpx_codec_get_preview_frame
+text vpx_codec_set_cx_data_buf
diff --git a/libvpx/libvpx/vpx/exports_spatial_svc b/libvpx/libvpx/vpx/exports_spatial_svc
new file mode 100644
index 0000000..d258a1d
--- /dev/null
+++ b/libvpx/libvpx/vpx/exports_spatial_svc
@@ -0,0 +1,6 @@
+text vpx_svc_dump_statistics
+text vpx_svc_encode
+text vpx_svc_get_message
+text vpx_svc_init
+text vpx_svc_release
+text vpx_svc_set_options
diff --git a/libvpx/libvpx/vpx/internal/vpx_codec_internal.h b/libvpx/libvpx/vpx/internal/vpx_codec_internal.h
new file mode 100644
index 0000000..7380fcc
--- /dev/null
+++ b/libvpx/libvpx/vpx/internal/vpx_codec_internal.h
@@ -0,0 +1,445 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+/*!\file
+ * \brief Describes the decoder algorithm interface for algorithm
+ *        implementations.
+ *
+ * This file defines the private structures and data types that are only
+ * relevant to implementing an algorithm, as opposed to using it.
+ *
+ * To create a decoder algorithm class, an interface structure is put
+ * into the global namespace:
+ *     <pre>
+ *     my_codec.c:
+ *       vpx_codec_iface_t my_codec = {
+ *           "My Codec v1.0",
+ *           VPX_CODEC_ALG_ABI_VERSION,
+ *           ...
+ *       };
+ *     </pre>
+ *
+ * An application instantiates a specific decoder instance by using
+ * vpx_codec_init() and a pointer to the algorithm's interface structure:
+ *     <pre>
+ *     my_app.c:
+ *       extern vpx_codec_iface_t my_codec;
+ *       {
+ *           vpx_codec_ctx_t algo;
+ *           res = vpx_codec_init(&algo, &my_codec);
+ *       }
+ *     </pre>
+ *
+ * Once initialized, the instance is manged using other functions from
+ * the vpx_codec_* family.
+ */
+#ifndef VPX_INTERNAL_VPX_CODEC_INTERNAL_H_
+#define VPX_INTERNAL_VPX_CODEC_INTERNAL_H_
+#include "../vpx_decoder.h"
+#include "../vpx_encoder.h"
+#include <stdarg.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*!\brief Current ABI version number
+ *
+ * \internal
+ * If this file is altered in any way that changes the ABI, this value
+ * must be bumped.  Examples include, but are not limited to, changing
+ * types, removing or reassigning enums, adding/removing/rearranging
+ * fields to structures
+ */
+#define VPX_CODEC_INTERNAL_ABI_VERSION (5) /**<\hideinitializer*/
+
+typedef struct vpx_codec_alg_priv  vpx_codec_alg_priv_t;
+typedef struct vpx_codec_priv_enc_mr_cfg vpx_codec_priv_enc_mr_cfg_t;
+
+/*!\brief init function pointer prototype
+ *
+ * Performs algorithm-specific initialization of the decoder context. This
+ * function is called by the generic vpx_codec_init() wrapper function, so
+ * plugins implementing this interface may trust the input parameters to be
+ * properly initialized.
+ *
+ * \param[in] ctx   Pointer to this instance's context
+ * \retval #VPX_CODEC_OK
+ *     The input stream was recognized and decoder initialized.
+ * \retval #VPX_CODEC_MEM_ERROR
+ *     Memory operation failed.
+ */
+typedef vpx_codec_err_t (*vpx_codec_init_fn_t)(vpx_codec_ctx_t *ctx,
+                                               vpx_codec_priv_enc_mr_cfg_t *data);
+
+/*!\brief destroy function pointer prototype
+ *
+ * Performs algorithm-specific destruction of the decoder context. This
+ * function is called by the generic vpx_codec_destroy() wrapper function,
+ * so plugins implementing this interface may trust the input parameters
+ * to be properly initialized.
+ *
+ * \param[in] ctx   Pointer to this instance's context
+ * \retval #VPX_CODEC_OK
+ *     The input stream was recognized and decoder initialized.
+ * \retval #VPX_CODEC_MEM_ERROR
+ *     Memory operation failed.
+ */
+typedef vpx_codec_err_t (*vpx_codec_destroy_fn_t)(vpx_codec_alg_priv_t *ctx);
+
+/*!\brief parse stream info function pointer prototype
+ *
+ * Performs high level parsing of the bitstream. This function is called by the
+ * generic vpx_codec_peek_stream_info() wrapper function, so plugins
+ * implementing this interface may trust the input parameters to be properly
+ * initialized.
+ *
+ * \param[in]      data    Pointer to a block of data to parse
+ * \param[in]      data_sz Size of the data buffer
+ * \param[in,out]  si      Pointer to stream info to update. The size member
+ *                         \ref MUST be properly initialized, but \ref MAY be
+ *                         clobbered by the algorithm. This parameter \ref MAY
+ *                         be NULL.
+ *
+ * \retval #VPX_CODEC_OK
+ *     Bitstream is parsable and stream information updated
+ */
+typedef vpx_codec_err_t (*vpx_codec_peek_si_fn_t)(const uint8_t         *data,
+                                                  unsigned int           data_sz,
+                                                  vpx_codec_stream_info_t *si);
+
+/*!\brief Return information about the current stream.
+ *
+ * Returns information about the stream that has been parsed during decoding.
+ *
+ * \param[in]      ctx     Pointer to this instance's context
+ * \param[in,out]  si      Pointer to stream info to update. The size member
+ *                         \ref MUST be properly initialized, but \ref MAY be
+ *                         clobbered by the algorithm. This parameter \ref MAY
+ *                         be NULL.
+ *
+ * \retval #VPX_CODEC_OK
+ *     Bitstream is parsable and stream information updated
+ */
+typedef vpx_codec_err_t (*vpx_codec_get_si_fn_t)(vpx_codec_alg_priv_t    *ctx,
+                                                 vpx_codec_stream_info_t *si);
+
+/*!\brief control function pointer prototype
+ *
+ * This function is used to exchange algorithm specific data with the decoder
+ * instance. This can be used to implement features specific to a particular
+ * algorithm.
+ *
+ * This function is called by the generic vpx_codec_control() wrapper
+ * function, so plugins implementing this interface may trust the input
+ * parameters to be properly initialized. However,  this interface does not
+ * provide type safety for the exchanged data or assign meanings to the
+ * control codes. Those details should be specified in the algorithm's
+ * header file. In particular, the ctrl_id parameter is guaranteed to exist
+ * in the algorithm's control mapping table, and the data parameter may be NULL.
+ *
+ *
+ * \param[in]     ctx              Pointer to this instance's context
+ * \param[in]     ctrl_id          Algorithm specific control identifier
+ * \param[in,out] data             Data to exchange with algorithm instance.
+ *
+ * \retval #VPX_CODEC_OK
+ *     The internal state data was deserialized.
+ */
+typedef vpx_codec_err_t (*vpx_codec_control_fn_t)(vpx_codec_alg_priv_t *ctx,
+                                                  va_list ap);
+
+/*!\brief control function pointer mapping
+ *
+ * This structure stores the mapping between control identifiers and
+ * implementing functions. Each algorithm provides a list of these
+ * mappings. This list is searched by the vpx_codec_control() wrapper
+ * function to determine which function to invoke. The special
+ * value {0, NULL} is used to indicate end-of-list, and must be
+ * present. The special value {0, <non-null>} can be used as a catch-all
+ * mapping. This implies that ctrl_id values chosen by the algorithm
+ * \ref MUST be non-zero.
+ */
+typedef const struct vpx_codec_ctrl_fn_map {
+  int ctrl_id;
+  vpx_codec_control_fn_t fn;
+} vpx_codec_ctrl_fn_map_t;
+
+/*!\brief decode data function pointer prototype
+ *
+ * Processes a buffer of coded data. If the processing results in a new
+ * decoded frame becoming available, #VPX_CODEC_CB_PUT_SLICE and
+ * #VPX_CODEC_CB_PUT_FRAME events are generated as appropriate. This
+ * function is called by the generic vpx_codec_decode() wrapper function,
+ * so plugins implementing this interface may trust the input parameters
+ * to be properly initialized.
+ *
+ * \param[in] ctx          Pointer to this instance's context
+ * \param[in] data         Pointer to this block of new coded data. If
+ *                         NULL, a #VPX_CODEC_CB_PUT_FRAME event is posted
+ *                         for the previously decoded frame.
+ * \param[in] data_sz      Size of the coded data, in bytes.
+ *
+ * \return Returns #VPX_CODEC_OK if the coded data was processed completely
+ *         and future pictures can be decoded without error. Otherwise,
+ *         see the descriptions of the other error codes in ::vpx_codec_err_t
+ *         for recoverability capabilities.
+ */
+typedef vpx_codec_err_t (*vpx_codec_decode_fn_t)(vpx_codec_alg_priv_t  *ctx,
+                                                 const uint8_t         *data,
+                                                 unsigned int     data_sz,
+                                                 void        *user_priv,
+                                                 long         deadline);
+
+/*!\brief Decoded frames iterator
+ *
+ * Iterates over a list of the frames available for display. The iterator
+ * storage should be initialized to NULL to start the iteration. Iteration is
+ * complete when this function returns NULL.
+ *
+ * The list of available frames becomes valid upon completion of the
+ * vpx_codec_decode call, and remains valid until the next call to vpx_codec_decode.
+ *
+ * \param[in]     ctx      Pointer to this instance's context
+ * \param[in out] iter     Iterator storage, initialized to NULL
+ *
+ * \return Returns a pointer to an image, if one is ready for display. Frames
+ *         produced will always be in PTS (presentation time stamp) order.
+ */
+typedef vpx_image_t *(*vpx_codec_get_frame_fn_t)(vpx_codec_alg_priv_t *ctx,
+                                                 vpx_codec_iter_t     *iter);
+
+/*!\brief Pass in external frame buffers for the decoder to use.
+ *
+ * Registers functions to be called when libvpx needs a frame buffer
+ * to decode the current frame and a function to be called when libvpx does
+ * not internally reference the frame buffer. This set function must
+ * be called before the first call to decode or libvpx will assume the
+ * default behavior of allocating frame buffers internally.
+ *
+ * \param[in] ctx          Pointer to this instance's context
+ * \param[in] cb_get       Pointer to the get callback function
+ * \param[in] cb_release   Pointer to the release callback function
+ * \param[in] cb_priv      Callback's private data
+ *
+ * \retval #VPX_CODEC_OK
+ *     External frame buffers will be used by libvpx.
+ * \retval #VPX_CODEC_INVALID_PARAM
+ *     One or more of the callbacks were NULL.
+ * \retval #VPX_CODEC_ERROR
+ *     Decoder context not initialized, or algorithm not capable of
+ *     using external frame buffers.
+ *
+ * \note
+ * When decoding VP9, the application may be required to pass in at least
+ * #VP9_MAXIMUM_REF_BUFFERS + #VPX_MAXIMUM_WORK_BUFFERS external frame
+ * buffers.
+ */
+typedef vpx_codec_err_t (*vpx_codec_set_fb_fn_t)(
+    vpx_codec_alg_priv_t *ctx,
+    vpx_get_frame_buffer_cb_fn_t cb_get,
+    vpx_release_frame_buffer_cb_fn_t cb_release, void *cb_priv);
+
+
+typedef vpx_codec_err_t (*vpx_codec_encode_fn_t)(vpx_codec_alg_priv_t  *ctx,
+                                                 const vpx_image_t     *img,
+                                                 vpx_codec_pts_t        pts,
+                                                 unsigned long          duration,
+                                                 vpx_enc_frame_flags_t  flags,
+                                                 unsigned long          deadline);
+typedef const vpx_codec_cx_pkt_t *(*vpx_codec_get_cx_data_fn_t)(vpx_codec_alg_priv_t *ctx,
+                                                                vpx_codec_iter_t     *iter);
+
+typedef vpx_codec_err_t
+(*vpx_codec_enc_config_set_fn_t)(vpx_codec_alg_priv_t       *ctx,
+                                 const vpx_codec_enc_cfg_t  *cfg);
+typedef vpx_fixed_buf_t *
+(*vpx_codec_get_global_headers_fn_t)(vpx_codec_alg_priv_t   *ctx);
+
+typedef vpx_image_t *
+(*vpx_codec_get_preview_frame_fn_t)(vpx_codec_alg_priv_t   *ctx);
+
+typedef vpx_codec_err_t
+(*vpx_codec_enc_mr_get_mem_loc_fn_t)(const vpx_codec_enc_cfg_t     *cfg,
+                                     void **mem_loc);
+
+/*!\brief usage configuration mapping
+ *
+ * This structure stores the mapping between usage identifiers and
+ * configuration structures. Each algorithm provides a list of these
+ * mappings. This list is searched by the vpx_codec_enc_config_default()
+ * wrapper function to determine which config to return. The special value
+ * {-1, {0}} is used to indicate end-of-list, and must be present. At least
+ * one mapping must be present, in addition to the end-of-list.
+ *
+ */
+typedef const struct vpx_codec_enc_cfg_map {
+  int                 usage;
+  vpx_codec_enc_cfg_t cfg;
+} vpx_codec_enc_cfg_map_t;
+
+/*!\brief Decoder algorithm interface interface
+ *
+ * All decoders \ref MUST expose a variable of this type.
+ */
+struct vpx_codec_iface {
+  const char               *name;        /**< Identification String  */
+  int                       abi_version; /**< Implemented ABI version */
+  vpx_codec_caps_t          caps;    /**< Decoder capabilities */
+  vpx_codec_init_fn_t       init;    /**< \copydoc ::vpx_codec_init_fn_t */
+  vpx_codec_destroy_fn_t    destroy;     /**< \copydoc ::vpx_codec_destroy_fn_t */
+  vpx_codec_ctrl_fn_map_t  *ctrl_maps;   /**< \copydoc ::vpx_codec_ctrl_fn_map_t */
+  struct vpx_codec_dec_iface {
+    vpx_codec_peek_si_fn_t    peek_si;     /**< \copydoc ::vpx_codec_peek_si_fn_t */
+    vpx_codec_get_si_fn_t     get_si;      /**< \copydoc ::vpx_codec_get_si_fn_t */
+    vpx_codec_decode_fn_t     decode;      /**< \copydoc ::vpx_codec_decode_fn_t */
+    vpx_codec_get_frame_fn_t  get_frame;   /**< \copydoc ::vpx_codec_get_frame_fn_t */
+    vpx_codec_set_fb_fn_t     set_fb_fn;   /**< \copydoc ::vpx_codec_set_fb_fn_t */
+  } dec;
+  struct vpx_codec_enc_iface {
+    int                                cfg_map_count;
+    vpx_codec_enc_cfg_map_t           *cfg_maps;      /**< \copydoc ::vpx_codec_enc_cfg_map_t */
+    vpx_codec_encode_fn_t              encode;        /**< \copydoc ::vpx_codec_encode_fn_t */
+    vpx_codec_get_cx_data_fn_t         get_cx_data;   /**< \copydoc ::vpx_codec_get_cx_data_fn_t */
+    vpx_codec_enc_config_set_fn_t      cfg_set;       /**< \copydoc ::vpx_codec_enc_config_set_fn_t */
+    vpx_codec_get_global_headers_fn_t  get_glob_hdrs; /**< \copydoc ::vpx_codec_get_global_headers_fn_t */
+    vpx_codec_get_preview_frame_fn_t   get_preview;   /**< \copydoc ::vpx_codec_get_preview_frame_fn_t */
+    vpx_codec_enc_mr_get_mem_loc_fn_t  mr_get_mem_loc;   /**< \copydoc ::vpx_codec_enc_mr_get_mem_loc_fn_t */
+  } enc;
+};
+
+/*!\brief Callback function pointer / user data pair storage */
+typedef struct vpx_codec_priv_cb_pair {
+  union {
+    vpx_codec_put_frame_cb_fn_t    put_frame;
+    vpx_codec_put_slice_cb_fn_t    put_slice;
+  } u;
+  void                            *user_priv;
+} vpx_codec_priv_cb_pair_t;
+
+
+/*!\brief Instance private storage
+ *
+ * This structure is allocated by the algorithm's init function. It can be
+ * extended in one of two ways. First, a second, algorithm specific structure
+ * can be allocated and the priv member pointed to it. Alternatively, this
+ * structure can be made the first member of the algorithm specific structure,
+ * and the pointer cast to the proper type.
+ */
+struct vpx_codec_priv {
+  const char                     *err_detail;
+  vpx_codec_flags_t               init_flags;
+  struct {
+    vpx_codec_priv_cb_pair_t    put_frame_cb;
+    vpx_codec_priv_cb_pair_t    put_slice_cb;
+  } dec;
+  struct {
+    vpx_fixed_buf_t             cx_data_dst_buf;
+    unsigned int                cx_data_pad_before;
+    unsigned int                cx_data_pad_after;
+    vpx_codec_cx_pkt_t          cx_data_pkt;
+    unsigned int                total_encoders;
+  } enc;
+};
+
+/*
+ * Multi-resolution encoding internal configuration
+ */
+struct vpx_codec_priv_enc_mr_cfg
+{
+    unsigned int           mr_total_resolutions;
+    unsigned int           mr_encoder_id;
+    struct vpx_rational    mr_down_sampling_factor;
+    void*                  mr_low_res_mode_info;
+};
+
+#undef VPX_CTRL_USE_TYPE
+#define VPX_CTRL_USE_TYPE(id, typ) \
+  static VPX_INLINE typ id##__value(va_list args) {return va_arg(args, typ);}
+
+#undef VPX_CTRL_USE_TYPE_DEPRECATED
+#define VPX_CTRL_USE_TYPE_DEPRECATED(id, typ) \
+  static VPX_INLINE typ id##__value(va_list args) {return va_arg(args, typ);}
+
+#define CAST(id, arg) id##__value(arg)
+
+/* CODEC_INTERFACE convenience macro
+ *
+ * By convention, each codec interface is a struct with extern linkage, where
+ * the symbol is suffixed with _algo. A getter function is also defined to
+ * return a pointer to the struct, since in some cases it's easier to work
+ * with text symbols than data symbols (see issue #169). This function has
+ * the same name as the struct, less the _algo suffix. The CODEC_INTERFACE
+ * macro is provided to define this getter function automatically.
+ */
+#define CODEC_INTERFACE(id)\
+  vpx_codec_iface_t* id(void) { return &id##_algo; }\
+  vpx_codec_iface_t  id##_algo
+
+
+/* Internal Utility Functions
+ *
+ * The following functions are intended to be used inside algorithms as
+ * utilities for manipulating vpx_codec_* data structures.
+ */
+struct vpx_codec_pkt_list {
+  unsigned int            cnt;
+  unsigned int            max;
+  struct vpx_codec_cx_pkt pkts[1];
+};
+
+#define vpx_codec_pkt_list_decl(n)\
+  union {struct vpx_codec_pkt_list head;\
+    struct {struct vpx_codec_pkt_list head;\
+      struct vpx_codec_cx_pkt    pkts[n];} alloc;}
+
+#define vpx_codec_pkt_list_init(m)\
+  (m)->alloc.head.cnt = 0,\
+                        (m)->alloc.head.max = sizeof((m)->alloc.pkts) / sizeof((m)->alloc.pkts[0])
+
+int
+vpx_codec_pkt_list_add(struct vpx_codec_pkt_list *,
+                       const struct vpx_codec_cx_pkt *);
+
+const vpx_codec_cx_pkt_t *
+vpx_codec_pkt_list_get(struct vpx_codec_pkt_list *list,
+                       vpx_codec_iter_t           *iter);
+
+
+#include <stdio.h>
+#include <setjmp.h>
+
+struct vpx_internal_error_info {
+  vpx_codec_err_t  error_code;
+  int              has_detail;
+  char             detail[80];
+  int              setjmp;
+  jmp_buf          jmp;
+};
+
+#define CLANG_ANALYZER_NORETURN
+#if defined(__has_feature)
+#if __has_feature(attribute_analyzer_noreturn)
+#undef CLANG_ANALYZER_NORETURN
+#define CLANG_ANALYZER_NORETURN __attribute__((analyzer_noreturn))
+#endif
+#endif
+
+void vpx_internal_error(struct vpx_internal_error_info *info,
+                        vpx_codec_err_t                 error,
+                        const char                     *fmt,
+                        ...) CLANG_ANALYZER_NORETURN;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_INTERNAL_VPX_CODEC_INTERNAL_H_
diff --git a/libvpx/libvpx/vpx/internal/vpx_psnr.h b/libvpx/libvpx/vpx/internal/vpx_psnr.h
new file mode 100644
index 0000000..07d81bb
--- /dev/null
+++ b/libvpx/libvpx/vpx/internal/vpx_psnr.h
@@ -0,0 +1,34 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_INTERNAL_VPX_PSNR_H_
+#define VPX_INTERNAL_VPX_PSNR_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// TODO(dkovalev) change vpx_sse_to_psnr signature: double -> int64_t
+
+/*!\brief Converts SSE to PSNR
+ *
+ * Converts sum of squared errros (SSE) to peak signal-to-noise ratio (PNSR).
+ *
+ * \param[in]    samples       Number of samples
+ * \param[in]    peak          Max sample value
+ * \param[in]    sse           Sum of squared errors
+ */
+double vpx_sse_to_psnr(double samples, double peak, double sse);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_INTERNAL_VPX_PSNR_H_
diff --git a/libvpx/libvpx/vpx/src/svc_encodeframe.c b/libvpx/libvpx/vpx/src/svc_encodeframe.c
new file mode 100644
index 0000000..ef9b352
--- /dev/null
+++ b/libvpx/libvpx/vpx/src/svc_encodeframe.c
@@ -0,0 +1,686 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+/**
+ * @file
+ * VP9 SVC encoding support via libvpx
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#define VPX_DISABLE_CTRL_TYPECHECKS 1
+#include "./vpx_config.h"
+#include "vpx/svc_context.h"
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp9/common/vp9_onyxc_int.h"
+
+#ifdef __MINGW32__
+#define strtok_r strtok_s
+#ifndef MINGW_HAS_SECURE_API
+// proto from /usr/x86_64-w64-mingw32/include/sec_api/string_s.h
+_CRTIMP char *__cdecl strtok_s(char *str, const char *delim, char **context);
+#endif  /* MINGW_HAS_SECURE_API */
+#endif  /* __MINGW32__ */
+
+#ifdef _MSC_VER
+#define strdup _strdup
+#define strtok_r strtok_s
+#endif
+
+#define SVC_REFERENCE_FRAMES 8
+#define SUPERFRAME_SLOTS (8)
+#define SUPERFRAME_BUFFER_SIZE (SUPERFRAME_SLOTS * sizeof(uint32_t) + 2)
+
+#define MAX_QUANTIZER 63
+
+static const int DEFAULT_SCALE_FACTORS_NUM[VPX_SS_MAX_LAYERS] = {
+  4, 5, 7, 11, 16
+};
+
+static const int DEFAULT_SCALE_FACTORS_DEN[VPX_SS_MAX_LAYERS] = {
+  16, 16, 16, 16, 16
+};
+
+typedef enum {
+  QUANTIZER = 0,
+  BITRATE,
+  SCALE_FACTOR,
+  AUTO_ALT_REF,
+  ALL_OPTION_TYPES
+} LAYER_OPTION_TYPE;
+
+static const int option_max_values[ALL_OPTION_TYPES] = {
+  63, INT_MAX, INT_MAX, 1
+};
+
+static const int option_min_values[ALL_OPTION_TYPES] = {
+  0, 0, 1, 0
+};
+
+// One encoded frame
+typedef struct FrameData {
+  void                     *buf;    // compressed data buffer
+  size_t                    size;  // length of compressed data
+  vpx_codec_frame_flags_t   flags;    /**< flags for this frame */
+  struct FrameData         *next;
+} FrameData;
+
+static SvcInternal_t *get_svc_internal(SvcContext *svc_ctx) {
+  if (svc_ctx == NULL) return NULL;
+  if (svc_ctx->internal == NULL) {
+    SvcInternal_t *const si = (SvcInternal_t *)malloc(sizeof(*si));
+    if (si != NULL) {
+      memset(si, 0, sizeof(*si));
+    }
+    svc_ctx->internal = si;
+  }
+  return (SvcInternal_t *)svc_ctx->internal;
+}
+
+static const SvcInternal_t *get_const_svc_internal(
+    const SvcContext *svc_ctx) {
+  if (svc_ctx == NULL) return NULL;
+  return (const SvcInternal_t *)svc_ctx->internal;
+}
+
+static void svc_log_reset(SvcContext *svc_ctx) {
+  SvcInternal_t *const si = (SvcInternal_t *)svc_ctx->internal;
+  si->message_buffer[0] = '\0';
+}
+
+static int svc_log(SvcContext *svc_ctx, SVC_LOG_LEVEL level,
+                   const char *fmt, ...) {
+  char buf[512];
+  int retval = 0;
+  va_list ap;
+  SvcInternal_t *const si = get_svc_internal(svc_ctx);
+
+  if (level > svc_ctx->log_level) {
+    return retval;
+  }
+
+  va_start(ap, fmt);
+  retval = vsnprintf(buf, sizeof(buf), fmt, ap);
+  va_end(ap);
+
+  if (svc_ctx->log_print) {
+    printf("%s", buf);
+  } else {
+    strncat(si->message_buffer, buf,
+            sizeof(si->message_buffer) - strlen(si->message_buffer) - 1);
+  }
+
+  if (level == SVC_LOG_ERROR) {
+    si->codec_ctx->err_detail = si->message_buffer;
+  }
+  return retval;
+}
+
+static vpx_codec_err_t extract_option(LAYER_OPTION_TYPE type,
+                                      char *input,
+                                      int *value0,
+                                      int *value1) {
+  if (type == SCALE_FACTOR) {
+    *value0 = strtol(input, &input, 10);
+    if (*input++ != '/')
+      return VPX_CODEC_INVALID_PARAM;
+    *value1 = strtol(input, &input, 10);
+
+    if (*value0 < option_min_values[SCALE_FACTOR] ||
+        *value1 < option_min_values[SCALE_FACTOR] ||
+        *value0 > option_max_values[SCALE_FACTOR] ||
+        *value1 > option_max_values[SCALE_FACTOR] ||
+        *value0 > *value1)  // num shouldn't be greater than den
+      return VPX_CODEC_INVALID_PARAM;
+  } else {
+    *value0 = atoi(input);
+    if (*value0 < option_min_values[type] ||
+        *value0 > option_max_values[type])
+      return VPX_CODEC_INVALID_PARAM;
+  }
+  return VPX_CODEC_OK;
+}
+
+static vpx_codec_err_t parse_layer_options_from_string(SvcContext *svc_ctx,
+                                                       LAYER_OPTION_TYPE type,
+                                                       const char *input,
+                                                       int *option0,
+                                                       int *option1) {
+  int i;
+  vpx_codec_err_t res = VPX_CODEC_OK;
+  char *input_string;
+  char *token;
+  const char *delim = ",";
+  char *save_ptr;
+
+  if (input == NULL || option0 == NULL ||
+      (option1 == NULL && type == SCALE_FACTOR))
+    return VPX_CODEC_INVALID_PARAM;
+
+  input_string = strdup(input);
+  token = strtok_r(input_string, delim, &save_ptr);
+  for (i = 0; i < svc_ctx->spatial_layers; ++i) {
+    if (token != NULL) {
+      res = extract_option(type, token, option0 + i, option1 + i);
+      if (res != VPX_CODEC_OK)
+        break;
+      token = strtok_r(NULL, delim, &save_ptr);
+    } else {
+      break;
+    }
+  }
+  if (res == VPX_CODEC_OK && i != svc_ctx->spatial_layers) {
+    svc_log(svc_ctx, SVC_LOG_ERROR,
+            "svc: layer params type: %d    %d values required, "
+            "but only %d specified\n", type, svc_ctx->spatial_layers, i);
+    res = VPX_CODEC_INVALID_PARAM;
+  }
+  free(input_string);
+  return res;
+}
+
+/**
+ * Parse SVC encoding options
+ * Format: encoding-mode=<svc_mode>,layers=<layer_count>
+ *         scale-factors=<n1>/<d1>,<n2>/<d2>,...
+ *         quantizers=<q1>,<q2>,...
+ * svc_mode = [i|ip|alt_ip|gf]
+ */
+static vpx_codec_err_t parse_options(SvcContext *svc_ctx, const char *options) {
+  char *input_string;
+  char *option_name;
+  char *option_value;
+  char *input_ptr;
+  SvcInternal_t *const si = get_svc_internal(svc_ctx);
+  vpx_codec_err_t res = VPX_CODEC_OK;
+  int i, alt_ref_enabled = 0;
+
+  if (options == NULL) return VPX_CODEC_OK;
+  input_string = strdup(options);
+
+  // parse option name
+  option_name = strtok_r(input_string, "=", &input_ptr);
+  while (option_name != NULL) {
+    // parse option value
+    option_value = strtok_r(NULL, " ", &input_ptr);
+    if (option_value == NULL) {
+      svc_log(svc_ctx, SVC_LOG_ERROR, "option missing value: %s\n",
+              option_name);
+      res = VPX_CODEC_INVALID_PARAM;
+      break;
+    }
+    if (strcmp("spatial-layers", option_name) == 0) {
+      svc_ctx->spatial_layers = atoi(option_value);
+    } else if (strcmp("temporal-layers", option_name) == 0) {
+      svc_ctx->temporal_layers = atoi(option_value);
+    } else if (strcmp("scale-factors", option_name) == 0) {
+      res = parse_layer_options_from_string(svc_ctx, SCALE_FACTOR, option_value,
+                                            si->svc_params.scaling_factor_num,
+                                            si->svc_params.scaling_factor_den);
+      if (res != VPX_CODEC_OK) break;
+    } else if (strcmp("max-quantizers", option_name) == 0) {
+      res = parse_layer_options_from_string(svc_ctx, QUANTIZER, option_value,
+                                            si->svc_params.max_quantizers,
+                                            NULL);
+      if (res != VPX_CODEC_OK) break;
+    } else if (strcmp("min-quantizers", option_name) == 0) {
+      res = parse_layer_options_from_string(svc_ctx, QUANTIZER, option_value,
+                                            si->svc_params.min_quantizers,
+                                            NULL);
+      if (res != VPX_CODEC_OK) break;
+    } else if (strcmp("auto-alt-refs", option_name) == 0) {
+      res = parse_layer_options_from_string(svc_ctx, AUTO_ALT_REF, option_value,
+                                            si->enable_auto_alt_ref, NULL);
+      if (res != VPX_CODEC_OK) break;
+    } else if (strcmp("bitrates", option_name) == 0) {
+      res = parse_layer_options_from_string(svc_ctx, BITRATE, option_value,
+                                            si->bitrates, NULL);
+      if (res != VPX_CODEC_OK) break;
+    } else if (strcmp("multi-frame-contexts", option_name) == 0) {
+      si->use_multiple_frame_contexts = atoi(option_value);
+    } else {
+      svc_log(svc_ctx, SVC_LOG_ERROR, "invalid option: %s\n", option_name);
+      res = VPX_CODEC_INVALID_PARAM;
+      break;
+    }
+    option_name = strtok_r(NULL, "=", &input_ptr);
+  }
+  free(input_string);
+
+  for (i = 0; i < svc_ctx->spatial_layers; ++i) {
+    if (si->svc_params.max_quantizers[i] > MAX_QUANTIZER ||
+        si->svc_params.max_quantizers[i] < 0 ||
+        si->svc_params.min_quantizers[i] > si->svc_params.max_quantizers[i] ||
+        si->svc_params.min_quantizers[i] < 0)
+      res = VPX_CODEC_INVALID_PARAM;
+  }
+
+  if (si->use_multiple_frame_contexts &&
+      (svc_ctx->spatial_layers > 3 ||
+       svc_ctx->spatial_layers * svc_ctx->temporal_layers > 4))
+    res = VPX_CODEC_INVALID_PARAM;
+
+  for (i = 0; i < svc_ctx->spatial_layers; ++i)
+    alt_ref_enabled += si->enable_auto_alt_ref[i];
+  if (alt_ref_enabled > REF_FRAMES - svc_ctx->spatial_layers) {
+    svc_log(svc_ctx, SVC_LOG_ERROR,
+            "svc: auto alt ref: Maxinum %d(REF_FRAMES - layers) layers could"
+            "enabled auto alt reference frame, but % layers are enabled\n",
+            REF_FRAMES - svc_ctx->spatial_layers, alt_ref_enabled);
+    res = VPX_CODEC_INVALID_PARAM;
+  }
+
+  return res;
+}
+
+vpx_codec_err_t vpx_svc_set_options(SvcContext *svc_ctx,
+                                    const char *options) {
+  SvcInternal_t *const si = get_svc_internal(svc_ctx);
+  if (svc_ctx == NULL || options == NULL || si == NULL) {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+  strncpy(si->options, options, sizeof(si->options));
+  si->options[sizeof(si->options) - 1] = '\0';
+  return VPX_CODEC_OK;
+}
+
+void assign_layer_bitrates(const SvcContext *svc_ctx,
+                           vpx_codec_enc_cfg_t *const enc_cfg) {
+  int i;
+  const SvcInternal_t *const si = get_const_svc_internal(svc_ctx);
+  int sl, tl, spatial_layer_target;
+
+  if (svc_ctx->temporal_layering_mode != 0) {
+    if (si->bitrates[0] != 0) {
+      enc_cfg->rc_target_bitrate = 0;
+      for (sl = 0; sl < svc_ctx->spatial_layers; ++sl) {
+        enc_cfg->ss_target_bitrate[sl*svc_ctx->temporal_layers] = 0;
+        for (tl = 0; tl < svc_ctx->temporal_layers; ++tl) {
+          enc_cfg->ss_target_bitrate[sl*svc_ctx->temporal_layers]
+              += (unsigned int)si->bitrates[sl * svc_ctx->temporal_layers + tl];
+          enc_cfg->layer_target_bitrate[sl*svc_ctx->temporal_layers + tl]
+              = si->bitrates[sl * svc_ctx->temporal_layers + tl];
+        }
+      }
+    } else {
+      float total = 0;
+      float alloc_ratio[VPX_MAX_LAYERS] = {0};
+
+      for (sl = 0; sl < svc_ctx->spatial_layers; ++sl) {
+        if (si->svc_params.scaling_factor_den[sl] > 0) {
+          alloc_ratio[sl] = (float)( pow(2, sl) );
+          total += alloc_ratio[sl];
+        }
+      }
+
+      for (sl = 0; sl < svc_ctx->spatial_layers; ++sl) {
+        enc_cfg->ss_target_bitrate[sl] = spatial_layer_target =
+            (unsigned int)(enc_cfg->rc_target_bitrate *
+                alloc_ratio[sl] / total);
+        if (svc_ctx->temporal_layering_mode == 3) {
+          enc_cfg->layer_target_bitrate[sl * svc_ctx->temporal_layers] =
+              (spatial_layer_target*6)/10;  // 60%
+          enc_cfg->layer_target_bitrate[sl * svc_ctx->temporal_layers + 1] =
+              (spatial_layer_target*8)/10;  // 80%
+          enc_cfg->layer_target_bitrate[sl * svc_ctx->temporal_layers + 2] =
+              spatial_layer_target;
+        } else if (svc_ctx->temporal_layering_mode == 2 ||
+                   svc_ctx->temporal_layering_mode == 1) {
+          enc_cfg->layer_target_bitrate[sl * svc_ctx->temporal_layers] =
+              spatial_layer_target * 2 / 3;
+          enc_cfg->layer_target_bitrate[sl * svc_ctx->temporal_layers + 1] =
+              spatial_layer_target;
+        } else {
+          // User should explicitly assign bitrates in this case.
+          assert(0);
+        }
+      }
+    }
+  } else {
+    if (si->bitrates[0] != 0) {
+      enc_cfg->rc_target_bitrate = 0;
+      for (i = 0; i < svc_ctx->spatial_layers; ++i) {
+        enc_cfg->ss_target_bitrate[i] = (unsigned int)si->bitrates[i];
+        enc_cfg->rc_target_bitrate += si->bitrates[i];
+      }
+    } else {
+      float total = 0;
+      float alloc_ratio[VPX_MAX_LAYERS] = {0};
+
+      for (i = 0; i < svc_ctx->spatial_layers; ++i) {
+        if (si->svc_params.scaling_factor_den[i] > 0) {
+          alloc_ratio[i] = (float)(si->svc_params.scaling_factor_num[i] * 1.0 /
+                                   si->svc_params.scaling_factor_den[i]);
+
+          alloc_ratio[i] *= alloc_ratio[i];
+          total += alloc_ratio[i];
+        }
+      }
+      for (i = 0; i < VPX_SS_MAX_LAYERS; ++i) {
+        if (total > 0) {
+          enc_cfg->layer_target_bitrate[i] = (unsigned int)
+              (enc_cfg->rc_target_bitrate * alloc_ratio[i] / total);
+        }
+      }
+    }
+  }
+}
+
+vpx_codec_err_t vpx_svc_init(SvcContext *svc_ctx, vpx_codec_ctx_t *codec_ctx,
+                             vpx_codec_iface_t *iface,
+                             vpx_codec_enc_cfg_t *enc_cfg) {
+  vpx_codec_err_t res;
+  int i, sl , tl;
+  SvcInternal_t *const si = get_svc_internal(svc_ctx);
+  if (svc_ctx == NULL || codec_ctx == NULL || iface == NULL ||
+      enc_cfg == NULL) {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+  if (si == NULL) return VPX_CODEC_MEM_ERROR;
+
+  si->codec_ctx = codec_ctx;
+
+  si->width = enc_cfg->g_w;
+  si->height = enc_cfg->g_h;
+
+  si->kf_dist = enc_cfg->kf_max_dist;
+
+  if (svc_ctx->spatial_layers == 0)
+    svc_ctx->spatial_layers = VPX_SS_DEFAULT_LAYERS;
+  if (svc_ctx->spatial_layers < 1 ||
+      svc_ctx->spatial_layers > VPX_SS_MAX_LAYERS) {
+    svc_log(svc_ctx, SVC_LOG_ERROR, "spatial layers: invalid value: %d\n",
+            svc_ctx->spatial_layers);
+    return VPX_CODEC_INVALID_PARAM;
+  }
+
+  // Note: temporal_layering_mode only applies to one-pass CBR
+  // si->svc_params.temporal_layering_mode = svc_ctx->temporal_layering_mode;
+  if (svc_ctx->temporal_layering_mode == 3) {
+    svc_ctx->temporal_layers = 3;
+  } else if (svc_ctx->temporal_layering_mode == 2 ||
+             svc_ctx->temporal_layering_mode == 1) {
+    svc_ctx->temporal_layers = 2;
+  }
+
+  for (sl = 0; sl < VPX_SS_MAX_LAYERS; ++sl) {
+    si->svc_params.scaling_factor_num[sl] = DEFAULT_SCALE_FACTORS_NUM[sl];
+    si->svc_params.scaling_factor_den[sl] = DEFAULT_SCALE_FACTORS_DEN[sl];
+  }
+  for (tl = 0; tl < svc_ctx->temporal_layers; ++tl) {
+    for (sl = 0; sl < svc_ctx->spatial_layers; ++sl) {
+      i = sl * svc_ctx->temporal_layers + tl;
+      si->svc_params.max_quantizers[i] = MAX_QUANTIZER;
+      si->svc_params.min_quantizers[i] = 0;
+    }
+  }
+
+  // Parse aggregate command line options. Options must start with
+  // "layers=xx" then followed by other options
+  res = parse_options(svc_ctx, si->options);
+  if (res != VPX_CODEC_OK) return res;
+
+  if (svc_ctx->spatial_layers < 1)
+    svc_ctx->spatial_layers = 1;
+  if (svc_ctx->spatial_layers > VPX_SS_MAX_LAYERS)
+    svc_ctx->spatial_layers = VPX_SS_MAX_LAYERS;
+
+  if (svc_ctx->temporal_layers < 1)
+    svc_ctx->temporal_layers = 1;
+  if (svc_ctx->temporal_layers > VPX_TS_MAX_LAYERS)
+    svc_ctx->temporal_layers = VPX_TS_MAX_LAYERS;
+
+  if (svc_ctx->temporal_layers * svc_ctx->spatial_layers > VPX_MAX_LAYERS) {
+      svc_log(svc_ctx, SVC_LOG_ERROR,
+          "spatial layers * temporal layers exceeds the maximum number of "
+          "allowed layers of %d\n",
+          svc_ctx->spatial_layers * svc_ctx->temporal_layers,
+          (int) VPX_MAX_LAYERS);
+      return VPX_CODEC_INVALID_PARAM;
+  }
+  assign_layer_bitrates(svc_ctx, enc_cfg);
+
+#if CONFIG_SPATIAL_SVC
+  for (i = 0; i < svc_ctx->spatial_layers; ++i)
+    enc_cfg->ss_enable_auto_alt_ref[i] = si->enable_auto_alt_ref[i];
+#endif
+
+  if (svc_ctx->temporal_layers > 1) {
+    int i;
+    for (i = 0; i < svc_ctx->temporal_layers; ++i) {
+      enc_cfg->ts_target_bitrate[i] = enc_cfg->rc_target_bitrate /
+                                      svc_ctx->temporal_layers;
+      enc_cfg->ts_rate_decimator[i] = 1 << (svc_ctx->temporal_layers - 1 - i);
+    }
+  }
+
+  if (svc_ctx->threads)
+    enc_cfg->g_threads = svc_ctx->threads;
+
+  // Modify encoder configuration
+  enc_cfg->ss_number_layers = svc_ctx->spatial_layers;
+  enc_cfg->ts_number_layers = svc_ctx->temporal_layers;
+
+  if (enc_cfg->rc_end_usage == VPX_CBR) {
+    enc_cfg->rc_resize_allowed = 0;
+    enc_cfg->rc_min_quantizer = 2;
+    enc_cfg->rc_max_quantizer = 56;
+    enc_cfg->rc_undershoot_pct = 50;
+    enc_cfg->rc_overshoot_pct = 50;
+    enc_cfg->rc_buf_initial_sz = 500;
+    enc_cfg->rc_buf_optimal_sz = 600;
+    enc_cfg->rc_buf_sz = 1000;
+    enc_cfg->rc_dropframe_thresh = 0;
+  }
+
+  if (enc_cfg->g_error_resilient == 0 && si->use_multiple_frame_contexts == 0)
+    enc_cfg->g_error_resilient = 1;
+
+  // Initialize codec
+  res = vpx_codec_enc_init(codec_ctx, iface, enc_cfg, VPX_CODEC_USE_PSNR);
+  if (res != VPX_CODEC_OK) {
+    svc_log(svc_ctx, SVC_LOG_ERROR, "svc_enc_init error\n");
+    return res;
+  }
+  if (svc_ctx->spatial_layers > 1 || svc_ctx->temporal_layers > 1) {
+    vpx_codec_control(codec_ctx, VP9E_SET_SVC, 1);
+    vpx_codec_control(codec_ctx, VP9E_SET_SVC_PARAMETERS, &si->svc_params);
+  }
+  return VPX_CODEC_OK;
+}
+
+/**
+ * Encode a frame into multiple layers
+ * Create a superframe containing the individual layers
+ */
+vpx_codec_err_t vpx_svc_encode(SvcContext *svc_ctx,
+                               vpx_codec_ctx_t *codec_ctx,
+                               struct vpx_image *rawimg,
+                               vpx_codec_pts_t pts,
+                               int64_t duration, int deadline) {
+  vpx_codec_err_t res;
+  vpx_codec_iter_t iter;
+  const vpx_codec_cx_pkt_t *cx_pkt;
+  SvcInternal_t *const si = get_svc_internal(svc_ctx);
+  if (svc_ctx == NULL || codec_ctx == NULL || si == NULL) {
+    return VPX_CODEC_INVALID_PARAM;
+  }
+
+  svc_log_reset(svc_ctx);
+
+  res = vpx_codec_encode(codec_ctx, rawimg, pts, (uint32_t)duration, 0,
+                         deadline);
+  if (res != VPX_CODEC_OK) {
+    return res;
+  }
+  // save compressed data
+  iter = NULL;
+  while ((cx_pkt = vpx_codec_get_cx_data(codec_ctx, &iter))) {
+    switch (cx_pkt->kind) {
+#if VPX_ENCODER_ABI_VERSION > (5 + VPX_CODEC_ABI_VERSION)
+#if CONFIG_SPATIAL_SVC
+      case VPX_CODEC_SPATIAL_SVC_LAYER_PSNR: {
+        int i;
+        for (i = 0; i < svc_ctx->spatial_layers; ++i) {
+          int j;
+          svc_log(svc_ctx, SVC_LOG_DEBUG,
+                  "SVC frame: %d, layer: %d, PSNR(Total/Y/U/V): "
+                  "%2.3f  %2.3f  %2.3f  %2.3f \n",
+                  si->psnr_pkt_received, i,
+                  cx_pkt->data.layer_psnr[i].psnr[0],
+                  cx_pkt->data.layer_psnr[i].psnr[1],
+                  cx_pkt->data.layer_psnr[i].psnr[2],
+                  cx_pkt->data.layer_psnr[i].psnr[3]);
+          svc_log(svc_ctx, SVC_LOG_DEBUG,
+                  "SVC frame: %d, layer: %d, SSE(Total/Y/U/V): "
+                  "%2.3f  %2.3f  %2.3f  %2.3f \n",
+                  si->psnr_pkt_received, i,
+                  cx_pkt->data.layer_psnr[i].sse[0],
+                  cx_pkt->data.layer_psnr[i].sse[1],
+                  cx_pkt->data.layer_psnr[i].sse[2],
+                  cx_pkt->data.layer_psnr[i].sse[3]);
+
+          for (j = 0; j < COMPONENTS; ++j) {
+            si->psnr_sum[i][j] +=
+                cx_pkt->data.layer_psnr[i].psnr[j];
+            si->sse_sum[i][j] += cx_pkt->data.layer_psnr[i].sse[j];
+          }
+        }
+        ++si->psnr_pkt_received;
+        break;
+      }
+      case VPX_CODEC_SPATIAL_SVC_LAYER_SIZES: {
+        int i;
+        for (i = 0; i < svc_ctx->spatial_layers; ++i)
+          si->bytes_sum[i] += cx_pkt->data.layer_sizes[i];
+        break;
+      }
+#endif
+#endif
+      case VPX_CODEC_PSNR_PKT:
+      {
+#if VPX_ENCODER_ABI_VERSION > (5 + VPX_CODEC_ABI_VERSION)
+        int j;
+        svc_log(svc_ctx, SVC_LOG_DEBUG,
+                "frame: %d, layer: %d, PSNR(Total/Y/U/V): "
+                "%2.3f  %2.3f  %2.3f  %2.3f \n",
+                si->psnr_pkt_received, 0,
+                cx_pkt->data.layer_psnr[0].psnr[0],
+                cx_pkt->data.layer_psnr[0].psnr[1],
+                cx_pkt->data.layer_psnr[0].psnr[2],
+                cx_pkt->data.layer_psnr[0].psnr[3]);
+        for (j = 0; j < COMPONENTS; ++j) {
+          si->psnr_sum[0][j] +=
+              cx_pkt->data.layer_psnr[0].psnr[j];
+          si->sse_sum[0][j] += cx_pkt->data.layer_psnr[0].sse[j];
+        }
+#endif
+      }
+      ++si->psnr_pkt_received;
+      break;
+      default: {
+        break;
+      }
+    }
+  }
+
+  return VPX_CODEC_OK;
+}
+
+const char *vpx_svc_get_message(const SvcContext *svc_ctx) {
+  const SvcInternal_t *const si = get_const_svc_internal(svc_ctx);
+  if (svc_ctx == NULL || si == NULL) return NULL;
+  return si->message_buffer;
+}
+
+static double calc_psnr(double d) {
+  if (d == 0) return 100;
+  return -10.0 * log(d) / log(10.0);
+}
+
+// dump accumulated statistics and reset accumulated values
+const char *vpx_svc_dump_statistics(SvcContext *svc_ctx) {
+  int number_of_frames;
+  int i, j;
+  uint32_t bytes_total = 0;
+  double scale[COMPONENTS];
+  double psnr[COMPONENTS];
+  double mse[COMPONENTS];
+  double y_scale;
+
+  SvcInternal_t *const si = get_svc_internal(svc_ctx);
+  if (svc_ctx == NULL || si == NULL) return NULL;
+
+  svc_log_reset(svc_ctx);
+
+  number_of_frames = si->psnr_pkt_received;
+  if (number_of_frames <= 0) return vpx_svc_get_message(svc_ctx);
+
+  svc_log(svc_ctx, SVC_LOG_INFO, "\n");
+  for (i = 0; i < svc_ctx->spatial_layers; ++i) {
+
+    svc_log(svc_ctx, SVC_LOG_INFO,
+            "Layer %d Average PSNR=[%2.3f, %2.3f, %2.3f, %2.3f], Bytes=[%u]\n",
+            i, (double)si->psnr_sum[i][0] / number_of_frames,
+            (double)si->psnr_sum[i][1] / number_of_frames,
+            (double)si->psnr_sum[i][2] / number_of_frames,
+            (double)si->psnr_sum[i][3] / number_of_frames, si->bytes_sum[i]);
+    // the following psnr calculation is deduced from ffmpeg.c#print_report
+    y_scale = si->width * si->height * 255.0 * 255.0 * number_of_frames;
+    scale[1] = y_scale;
+    scale[2] = scale[3] = y_scale / 4;  // U or V
+    scale[0] = y_scale * 1.5;           // total
+
+    for (j = 0; j < COMPONENTS; j++) {
+      psnr[j] = calc_psnr(si->sse_sum[i][j] / scale[j]);
+      mse[j] = si->sse_sum[i][j] * 255.0 * 255.0 / scale[j];
+    }
+    svc_log(svc_ctx, SVC_LOG_INFO,
+            "Layer %d Overall PSNR=[%2.3f, %2.3f, %2.3f, %2.3f]\n", i, psnr[0],
+            psnr[1], psnr[2], psnr[3]);
+    svc_log(svc_ctx, SVC_LOG_INFO,
+            "Layer %d Overall MSE=[%2.3f, %2.3f, %2.3f, %2.3f]\n", i, mse[0],
+            mse[1], mse[2], mse[3]);
+
+    bytes_total += si->bytes_sum[i];
+    // Clear sums for next time.
+    si->bytes_sum[i] = 0;
+    for (j = 0; j < COMPONENTS; ++j) {
+      si->psnr_sum[i][j] = 0;
+      si->sse_sum[i][j] = 0;
+    }
+  }
+
+  // only display statistics once
+  si->psnr_pkt_received = 0;
+
+  svc_log(svc_ctx, SVC_LOG_INFO, "Total Bytes=[%u]\n", bytes_total);
+  return vpx_svc_get_message(svc_ctx);
+}
+
+void vpx_svc_release(SvcContext *svc_ctx) {
+  SvcInternal_t *si;
+  if (svc_ctx == NULL) return;
+  // do not use get_svc_internal as it will unnecessarily allocate an
+  // SvcInternal_t if it was not already allocated
+  si = (SvcInternal_t *)svc_ctx->internal;
+  if (si != NULL) {
+    free(si);
+    svc_ctx->internal = NULL;
+  }
+}
+
diff --git a/libvpx/libvpx/vpx/src/vpx_codec.c b/libvpx/libvpx/vpx/src/vpx_codec.c
new file mode 100644
index 0000000..5a495ce
--- /dev/null
+++ b/libvpx/libvpx/vpx/src/vpx_codec.c
@@ -0,0 +1,158 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+/*!\file
+ * \brief Provides the high level interface to wrap decoder algorithms.
+ *
+ */
+#include <stdarg.h>
+#include <stdlib.h>
+#include "vpx/vpx_integer.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "vpx_version.h"
+
+#define SAVE_STATUS(ctx,var) (ctx?(ctx->err = var):var)
+
+int vpx_codec_version(void) {
+  return VERSION_PACKED;
+}
+
+
+const char *vpx_codec_version_str(void) {
+  return VERSION_STRING_NOSP;
+}
+
+
+const char *vpx_codec_version_extra_str(void) {
+  return VERSION_EXTRA;
+}
+
+
+const char *vpx_codec_iface_name(vpx_codec_iface_t *iface) {
+  return iface ? iface->name : "<invalid interface>";
+}
+
+const char *vpx_codec_err_to_string(vpx_codec_err_t  err) {
+  switch (err) {
+    case VPX_CODEC_OK:
+      return "Success";
+    case VPX_CODEC_ERROR:
+      return "Unspecified internal error";
+    case VPX_CODEC_MEM_ERROR:
+      return "Memory allocation error";
+    case VPX_CODEC_ABI_MISMATCH:
+      return "ABI version mismatch";
+    case VPX_CODEC_INCAPABLE:
+      return "Codec does not implement requested capability";
+    case VPX_CODEC_UNSUP_BITSTREAM:
+      return "Bitstream not supported by this decoder";
+    case VPX_CODEC_UNSUP_FEATURE:
+      return "Bitstream required feature not supported by this decoder";
+    case VPX_CODEC_CORRUPT_FRAME:
+      return "Corrupt frame detected";
+    case  VPX_CODEC_INVALID_PARAM:
+      return "Invalid parameter";
+    case VPX_CODEC_LIST_END:
+      return "End of iterated list";
+  }
+
+  return "Unrecognized error code";
+}
+
+const char *vpx_codec_error(vpx_codec_ctx_t  *ctx) {
+  return (ctx) ? vpx_codec_err_to_string(ctx->err)
+         : vpx_codec_err_to_string(VPX_CODEC_INVALID_PARAM);
+}
+
+const char *vpx_codec_error_detail(vpx_codec_ctx_t  *ctx) {
+  if (ctx && ctx->err)
+    return ctx->priv ? ctx->priv->err_detail : ctx->err_detail;
+
+  return NULL;
+}
+
+
+vpx_codec_err_t vpx_codec_destroy(vpx_codec_ctx_t *ctx) {
+  vpx_codec_err_t res;
+
+  if (!ctx)
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv)
+    res = VPX_CODEC_ERROR;
+  else {
+    ctx->iface->destroy((vpx_codec_alg_priv_t *)ctx->priv);
+
+    ctx->iface = NULL;
+    ctx->name = NULL;
+    ctx->priv = NULL;
+    res = VPX_CODEC_OK;
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+
+vpx_codec_caps_t vpx_codec_get_caps(vpx_codec_iface_t *iface) {
+  return (iface) ? iface->caps : 0;
+}
+
+
+vpx_codec_err_t vpx_codec_control_(vpx_codec_ctx_t  *ctx,
+                                   int               ctrl_id,
+                                   ...) {
+  vpx_codec_err_t res;
+
+  if (!ctx || !ctrl_id)
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv || !ctx->iface->ctrl_maps)
+    res = VPX_CODEC_ERROR;
+  else {
+    vpx_codec_ctrl_fn_map_t *entry;
+
+    res = VPX_CODEC_ERROR;
+
+    for (entry = ctx->iface->ctrl_maps; entry && entry->fn; entry++) {
+      if (!entry->ctrl_id || entry->ctrl_id == ctrl_id) {
+        va_list  ap;
+
+        va_start(ap, ctrl_id);
+        res = entry->fn((vpx_codec_alg_priv_t *)ctx->priv, ap);
+        va_end(ap);
+        break;
+      }
+    }
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+void vpx_internal_error(struct vpx_internal_error_info *info,
+                        vpx_codec_err_t                 error,
+                        const char                     *fmt,
+                        ...) {
+  va_list ap;
+
+  info->error_code = error;
+  info->has_detail = 0;
+
+  if (fmt) {
+    size_t  sz = sizeof(info->detail);
+
+    info->has_detail = 1;
+    va_start(ap, fmt);
+    vsnprintf(info->detail, sz - 1, fmt, ap);
+    va_end(ap);
+    info->detail[sz - 1] = '\0';
+  }
+
+  if (info->setjmp)
+    longjmp(info->jmp, info->error_code);
+}
diff --git a/libvpx/libvpx/vpx/src/vpx_decoder.c b/libvpx/libvpx/vpx/src/vpx_decoder.c
new file mode 100644
index 0000000..802d8ed
--- /dev/null
+++ b/libvpx/libvpx/vpx/src/vpx_decoder.c
@@ -0,0 +1,197 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+/*!\file
+ * \brief Provides the high level interface to wrap decoder algorithms.
+ *
+ */
+#include <string.h>
+#include "vpx/internal/vpx_codec_internal.h"
+
+#define SAVE_STATUS(ctx,var) (ctx?(ctx->err = var):var)
+
+static vpx_codec_alg_priv_t *get_alg_priv(vpx_codec_ctx_t *ctx) {
+  return (vpx_codec_alg_priv_t *)ctx->priv;
+}
+
+vpx_codec_err_t vpx_codec_dec_init_ver(vpx_codec_ctx_t      *ctx,
+                                       vpx_codec_iface_t    *iface,
+                                       const vpx_codec_dec_cfg_t *cfg,
+                                       vpx_codec_flags_t     flags,
+                                       int                   ver) {
+  vpx_codec_err_t res;
+
+  if (ver != VPX_DECODER_ABI_VERSION)
+    res = VPX_CODEC_ABI_MISMATCH;
+  else if (!ctx || !iface)
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (iface->abi_version != VPX_CODEC_INTERNAL_ABI_VERSION)
+    res = VPX_CODEC_ABI_MISMATCH;
+  else if ((flags & VPX_CODEC_USE_POSTPROC) && !(iface->caps & VPX_CODEC_CAP_POSTPROC))
+    res = VPX_CODEC_INCAPABLE;
+  else if ((flags & VPX_CODEC_USE_ERROR_CONCEALMENT) &&
+           !(iface->caps & VPX_CODEC_CAP_ERROR_CONCEALMENT))
+    res = VPX_CODEC_INCAPABLE;
+  else if ((flags & VPX_CODEC_USE_INPUT_FRAGMENTS) &&
+           !(iface->caps & VPX_CODEC_CAP_INPUT_FRAGMENTS))
+    res = VPX_CODEC_INCAPABLE;
+  else if (!(iface->caps & VPX_CODEC_CAP_DECODER))
+    res = VPX_CODEC_INCAPABLE;
+  else {
+    memset(ctx, 0, sizeof(*ctx));
+    ctx->iface = iface;
+    ctx->name = iface->name;
+    ctx->priv = NULL;
+    ctx->init_flags = flags;
+    ctx->config.dec = cfg;
+
+    res = ctx->iface->init(ctx, NULL);
+    if (res) {
+      ctx->err_detail = ctx->priv ? ctx->priv->err_detail : NULL;
+      vpx_codec_destroy(ctx);
+    }
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+
+vpx_codec_err_t vpx_codec_peek_stream_info(vpx_codec_iface_t       *iface,
+                                           const uint8_t         *data,
+                                           unsigned int           data_sz,
+                                           vpx_codec_stream_info_t *si) {
+  vpx_codec_err_t res;
+
+  if (!iface || !data || !data_sz || !si
+      || si->sz < sizeof(vpx_codec_stream_info_t))
+    res = VPX_CODEC_INVALID_PARAM;
+  else {
+    /* Set default/unknown values */
+    si->w = 0;
+    si->h = 0;
+
+    res = iface->dec.peek_si(data, data_sz, si);
+  }
+
+  return res;
+}
+
+
+vpx_codec_err_t vpx_codec_get_stream_info(vpx_codec_ctx_t         *ctx,
+                                          vpx_codec_stream_info_t *si) {
+  vpx_codec_err_t res;
+
+  if (!ctx || !si || si->sz < sizeof(vpx_codec_stream_info_t))
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv)
+    res = VPX_CODEC_ERROR;
+  else {
+    /* Set default/unknown values */
+    si->w = 0;
+    si->h = 0;
+
+    res = ctx->iface->dec.get_si(get_alg_priv(ctx), si);
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+
+vpx_codec_err_t vpx_codec_decode(vpx_codec_ctx_t    *ctx,
+                                 const uint8_t        *data,
+                                 unsigned int    data_sz,
+                                 void       *user_priv,
+                                 long        deadline) {
+  vpx_codec_err_t res;
+
+  /* Sanity checks */
+  /* NULL data ptr allowed if data_sz is 0 too */
+  if (!ctx || (!data && data_sz) || (data && !data_sz))
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv)
+    res = VPX_CODEC_ERROR;
+  else {
+    res = ctx->iface->dec.decode(get_alg_priv(ctx), data, data_sz, user_priv,
+                                 deadline);
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+vpx_image_t *vpx_codec_get_frame(vpx_codec_ctx_t  *ctx,
+                                 vpx_codec_iter_t *iter) {
+  vpx_image_t *img;
+
+  if (!ctx || !iter || !ctx->iface || !ctx->priv)
+    img = NULL;
+  else
+    img = ctx->iface->dec.get_frame(get_alg_priv(ctx), iter);
+
+  return img;
+}
+
+
+vpx_codec_err_t vpx_codec_register_put_frame_cb(vpx_codec_ctx_t             *ctx,
+                                                vpx_codec_put_frame_cb_fn_t  cb,
+                                                void                      *user_priv) {
+  vpx_codec_err_t res;
+
+  if (!ctx || !cb)
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv
+           || !(ctx->iface->caps & VPX_CODEC_CAP_PUT_FRAME))
+    res = VPX_CODEC_ERROR;
+  else {
+    ctx->priv->dec.put_frame_cb.u.put_frame = cb;
+    ctx->priv->dec.put_frame_cb.user_priv = user_priv;
+    res = VPX_CODEC_OK;
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+
+vpx_codec_err_t vpx_codec_register_put_slice_cb(vpx_codec_ctx_t             *ctx,
+                                                vpx_codec_put_slice_cb_fn_t  cb,
+                                                void                      *user_priv) {
+  vpx_codec_err_t res;
+
+  if (!ctx || !cb)
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv
+           || !(ctx->iface->caps & VPX_CODEC_CAP_PUT_SLICE))
+    res = VPX_CODEC_ERROR;
+  else {
+    ctx->priv->dec.put_slice_cb.u.put_slice = cb;
+    ctx->priv->dec.put_slice_cb.user_priv = user_priv;
+    res = VPX_CODEC_OK;
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+vpx_codec_err_t vpx_codec_set_frame_buffer_functions(
+    vpx_codec_ctx_t *ctx, vpx_get_frame_buffer_cb_fn_t cb_get,
+    vpx_release_frame_buffer_cb_fn_t cb_release, void *cb_priv) {
+  vpx_codec_err_t res;
+
+  if (!ctx || !cb_get || !cb_release) {
+    res = VPX_CODEC_INVALID_PARAM;
+  } else if (!ctx->iface || !ctx->priv ||
+             !(ctx->iface->caps & VPX_CODEC_CAP_EXTERNAL_FRAME_BUFFER)) {
+    res = VPX_CODEC_ERROR;
+  } else {
+    res = ctx->iface->dec.set_fb_fn(get_alg_priv(ctx), cb_get, cb_release,
+                                    cb_priv);
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
diff --git a/libvpx/libvpx/vpx/src/vpx_encoder.c b/libvpx/libvpx/vpx/src/vpx_encoder.c
new file mode 100644
index 0000000..cd10c41
--- /dev/null
+++ b/libvpx/libvpx/vpx/src/vpx_encoder.c
@@ -0,0 +1,401 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+/*!\file
+ * \brief Provides the high level interface to wrap encoder algorithms.
+ *
+ */
+#include <limits.h>
+#include <string.h>
+#include "vpx_config.h"
+#include "vpx/internal/vpx_codec_internal.h"
+
+#define SAVE_STATUS(ctx,var) (ctx?(ctx->err = var):var)
+
+static vpx_codec_alg_priv_t *get_alg_priv(vpx_codec_ctx_t *ctx) {
+  return (vpx_codec_alg_priv_t *)ctx->priv;
+}
+
+vpx_codec_err_t vpx_codec_enc_init_ver(vpx_codec_ctx_t      *ctx,
+                                       vpx_codec_iface_t    *iface,
+                                       const vpx_codec_enc_cfg_t *cfg,
+                                       vpx_codec_flags_t     flags,
+                                       int                   ver) {
+  vpx_codec_err_t res;
+
+  if (ver != VPX_ENCODER_ABI_VERSION)
+    res = VPX_CODEC_ABI_MISMATCH;
+  else if (!ctx || !iface || !cfg)
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (iface->abi_version != VPX_CODEC_INTERNAL_ABI_VERSION)
+    res = VPX_CODEC_ABI_MISMATCH;
+  else if (!(iface->caps & VPX_CODEC_CAP_ENCODER))
+    res = VPX_CODEC_INCAPABLE;
+  else if ((flags & VPX_CODEC_USE_PSNR)
+           && !(iface->caps & VPX_CODEC_CAP_PSNR))
+    res = VPX_CODEC_INCAPABLE;
+  else if ((flags & VPX_CODEC_USE_OUTPUT_PARTITION)
+           && !(iface->caps & VPX_CODEC_CAP_OUTPUT_PARTITION))
+    res = VPX_CODEC_INCAPABLE;
+  else {
+    ctx->iface = iface;
+    ctx->name = iface->name;
+    ctx->priv = NULL;
+    ctx->init_flags = flags;
+    ctx->config.enc = cfg;
+    res = ctx->iface->init(ctx, NULL);
+
+    if (res) {
+      ctx->err_detail = ctx->priv ? ctx->priv->err_detail : NULL;
+      vpx_codec_destroy(ctx);
+    }
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+vpx_codec_err_t vpx_codec_enc_init_multi_ver(vpx_codec_ctx_t      *ctx,
+                                             vpx_codec_iface_t    *iface,
+                                             vpx_codec_enc_cfg_t  *cfg,
+                                             int                   num_enc,
+                                             vpx_codec_flags_t     flags,
+                                             vpx_rational_t       *dsf,
+                                             int                   ver) {
+  vpx_codec_err_t res = VPX_CODEC_OK;
+
+  if (ver != VPX_ENCODER_ABI_VERSION)
+    res = VPX_CODEC_ABI_MISMATCH;
+  else if (!ctx || !iface || !cfg || (num_enc > 16 || num_enc < 1))
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (iface->abi_version != VPX_CODEC_INTERNAL_ABI_VERSION)
+    res = VPX_CODEC_ABI_MISMATCH;
+  else if (!(iface->caps & VPX_CODEC_CAP_ENCODER))
+    res = VPX_CODEC_INCAPABLE;
+  else if ((flags & VPX_CODEC_USE_PSNR)
+           && !(iface->caps & VPX_CODEC_CAP_PSNR))
+    res = VPX_CODEC_INCAPABLE;
+  else if ((flags & VPX_CODEC_USE_OUTPUT_PARTITION)
+           && !(iface->caps & VPX_CODEC_CAP_OUTPUT_PARTITION))
+    res = VPX_CODEC_INCAPABLE;
+  else {
+    int i;
+    void *mem_loc = NULL;
+
+    if (!(res = iface->enc.mr_get_mem_loc(cfg, &mem_loc))) {
+      for (i = 0; i < num_enc; i++) {
+        vpx_codec_priv_enc_mr_cfg_t mr_cfg;
+
+        /* Validate down-sampling factor. */
+        if (dsf->num < 1 || dsf->num > 4096 || dsf->den < 1 ||
+            dsf->den > dsf->num) {
+          res = VPX_CODEC_INVALID_PARAM;
+          break;
+        }
+
+        mr_cfg.mr_low_res_mode_info = mem_loc;
+        mr_cfg.mr_total_resolutions = num_enc;
+        mr_cfg.mr_encoder_id = num_enc - 1 - i;
+        mr_cfg.mr_down_sampling_factor.num = dsf->num;
+        mr_cfg.mr_down_sampling_factor.den = dsf->den;
+
+        /* Force Key-frame synchronization. Namely, encoder at higher
+         * resolution always use the same frame_type chosen by the
+         * lowest-resolution encoder.
+         */
+        if (mr_cfg.mr_encoder_id)
+          cfg->kf_mode = VPX_KF_DISABLED;
+
+        ctx->iface = iface;
+        ctx->name = iface->name;
+        ctx->priv = NULL;
+        ctx->init_flags = flags;
+        ctx->config.enc = cfg;
+        res = ctx->iface->init(ctx, &mr_cfg);
+
+        if (res) {
+          const char *error_detail =
+            ctx->priv ? ctx->priv->err_detail : NULL;
+          /* Destroy current ctx */
+          ctx->err_detail = error_detail;
+          vpx_codec_destroy(ctx);
+
+          /* Destroy already allocated high-level ctx */
+          while (i) {
+            ctx--;
+            ctx->err_detail = error_detail;
+            vpx_codec_destroy(ctx);
+            i--;
+          }
+        }
+
+        if (res)
+          break;
+
+        ctx++;
+        cfg++;
+        dsf++;
+      }
+      ctx--;
+    }
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+
+vpx_codec_err_t  vpx_codec_enc_config_default(vpx_codec_iface_t    *iface,
+                                              vpx_codec_enc_cfg_t  *cfg,
+                                              unsigned int          usage) {
+  vpx_codec_err_t res;
+  vpx_codec_enc_cfg_map_t *map;
+  int i;
+
+  if (!iface || !cfg || usage > INT_MAX)
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (!(iface->caps & VPX_CODEC_CAP_ENCODER))
+    res = VPX_CODEC_INCAPABLE;
+  else {
+    res = VPX_CODEC_INVALID_PARAM;
+
+    for (i = 0; i < iface->enc.cfg_map_count; ++i) {
+      map = iface->enc.cfg_maps + i;
+      if (map->usage == (int)usage) {
+        *cfg = map->cfg;
+        cfg->g_usage = usage;
+        res = VPX_CODEC_OK;
+        break;
+      }
+    }
+  }
+
+  return res;
+}
+
+
+#if ARCH_X86 || ARCH_X86_64
+/* On X86, disable the x87 unit's internal 80 bit precision for better
+ * consistency with the SSE unit's 64 bit precision.
+ */
+#include "vpx_ports/x86.h"
+#define FLOATING_POINT_INIT() do {\
+    unsigned short x87_orig_mode = x87_set_double_precision();
+#define FLOATING_POINT_RESTORE() \
+  x87_set_control_word(x87_orig_mode); }while(0)
+
+
+#else
+static void FLOATING_POINT_INIT() {}
+static void FLOATING_POINT_RESTORE() {}
+#endif
+
+
+vpx_codec_err_t  vpx_codec_encode(vpx_codec_ctx_t            *ctx,
+                                  const vpx_image_t          *img,
+                                  vpx_codec_pts_t             pts,
+                                  unsigned long               duration,
+                                  vpx_enc_frame_flags_t       flags,
+                                  unsigned long               deadline) {
+  vpx_codec_err_t res = VPX_CODEC_OK;
+
+  if (!ctx || (img && !duration))
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv)
+    res = VPX_CODEC_ERROR;
+  else if (!(ctx->iface->caps & VPX_CODEC_CAP_ENCODER))
+    res = VPX_CODEC_INCAPABLE;
+  else {
+    unsigned int num_enc = ctx->priv->enc.total_encoders;
+
+    /* Execute in a normalized floating point environment, if the platform
+     * requires it.
+     */
+    FLOATING_POINT_INIT();
+
+    if (num_enc == 1)
+      res = ctx->iface->enc.encode(get_alg_priv(ctx), img, pts,
+                                   duration, flags, deadline);
+    else {
+      /* Multi-resolution encoding:
+       * Encode multi-levels in reverse order. For example,
+       * if mr_total_resolutions = 3, first encode level 2,
+       * then encode level 1, and finally encode level 0.
+       */
+      int i;
+
+      ctx += num_enc - 1;
+      if (img) img += num_enc - 1;
+
+      for (i = num_enc - 1; i >= 0; i--) {
+        if ((res = ctx->iface->enc.encode(get_alg_priv(ctx), img, pts,
+                                          duration, flags, deadline)))
+          break;
+
+        ctx--;
+        if (img) img--;
+      }
+      ctx++;
+    }
+
+    FLOATING_POINT_RESTORE();
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+
+const vpx_codec_cx_pkt_t *vpx_codec_get_cx_data(vpx_codec_ctx_t *ctx,
+                                                vpx_codec_iter_t *iter) {
+  const vpx_codec_cx_pkt_t *pkt = NULL;
+
+  if (ctx) {
+    if (!iter)
+      ctx->err = VPX_CODEC_INVALID_PARAM;
+    else if (!ctx->iface || !ctx->priv)
+      ctx->err = VPX_CODEC_ERROR;
+    else if (!(ctx->iface->caps & VPX_CODEC_CAP_ENCODER))
+      ctx->err = VPX_CODEC_INCAPABLE;
+    else
+      pkt = ctx->iface->enc.get_cx_data(get_alg_priv(ctx), iter);
+  }
+
+  if (pkt && pkt->kind == VPX_CODEC_CX_FRAME_PKT) {
+    // If the application has specified a destination area for the
+    // compressed data, and the codec has not placed the data there,
+    // and it fits, copy it.
+    vpx_codec_priv_t *const priv = ctx->priv;
+    char *const dst_buf = (char *)priv->enc.cx_data_dst_buf.buf;
+
+    if (dst_buf &&
+        pkt->data.raw.buf != dst_buf &&
+        pkt->data.raw.sz + priv->enc.cx_data_pad_before +
+            priv->enc.cx_data_pad_after <= priv->enc.cx_data_dst_buf.sz) {
+      vpx_codec_cx_pkt_t *modified_pkt = &priv->enc.cx_data_pkt;
+
+      memcpy(dst_buf + priv->enc.cx_data_pad_before, pkt->data.raw.buf,
+             pkt->data.raw.sz);
+      *modified_pkt = *pkt;
+      modified_pkt->data.raw.buf = dst_buf;
+      modified_pkt->data.raw.sz += priv->enc.cx_data_pad_before +
+                                       priv->enc.cx_data_pad_after;
+      pkt = modified_pkt;
+    }
+
+    if (dst_buf == pkt->data.raw.buf) {
+      priv->enc.cx_data_dst_buf.buf = dst_buf + pkt->data.raw.sz;
+      priv->enc.cx_data_dst_buf.sz -= pkt->data.raw.sz;
+    }
+  }
+
+  return pkt;
+}
+
+
+vpx_codec_err_t vpx_codec_set_cx_data_buf(vpx_codec_ctx_t       *ctx,
+                                          const vpx_fixed_buf_t *buf,
+                                          unsigned int           pad_before,
+                                          unsigned int           pad_after) {
+  if (!ctx || !ctx->priv)
+    return VPX_CODEC_INVALID_PARAM;
+
+  if (buf) {
+    ctx->priv->enc.cx_data_dst_buf = *buf;
+    ctx->priv->enc.cx_data_pad_before = pad_before;
+    ctx->priv->enc.cx_data_pad_after = pad_after;
+  } else {
+    ctx->priv->enc.cx_data_dst_buf.buf = NULL;
+    ctx->priv->enc.cx_data_dst_buf.sz = 0;
+    ctx->priv->enc.cx_data_pad_before = 0;
+    ctx->priv->enc.cx_data_pad_after = 0;
+  }
+
+  return VPX_CODEC_OK;
+}
+
+
+const vpx_image_t *vpx_codec_get_preview_frame(vpx_codec_ctx_t   *ctx) {
+  vpx_image_t *img = NULL;
+
+  if (ctx) {
+    if (!ctx->iface || !ctx->priv)
+      ctx->err = VPX_CODEC_ERROR;
+    else if (!(ctx->iface->caps & VPX_CODEC_CAP_ENCODER))
+      ctx->err = VPX_CODEC_INCAPABLE;
+    else if (!ctx->iface->enc.get_preview)
+      ctx->err = VPX_CODEC_INCAPABLE;
+    else
+      img = ctx->iface->enc.get_preview(get_alg_priv(ctx));
+  }
+
+  return img;
+}
+
+
+vpx_fixed_buf_t *vpx_codec_get_global_headers(vpx_codec_ctx_t   *ctx) {
+  vpx_fixed_buf_t *buf = NULL;
+
+  if (ctx) {
+    if (!ctx->iface || !ctx->priv)
+      ctx->err = VPX_CODEC_ERROR;
+    else if (!(ctx->iface->caps & VPX_CODEC_CAP_ENCODER))
+      ctx->err = VPX_CODEC_INCAPABLE;
+    else if (!ctx->iface->enc.get_glob_hdrs)
+      ctx->err = VPX_CODEC_INCAPABLE;
+    else
+      buf = ctx->iface->enc.get_glob_hdrs(get_alg_priv(ctx));
+  }
+
+  return buf;
+}
+
+
+vpx_codec_err_t  vpx_codec_enc_config_set(vpx_codec_ctx_t            *ctx,
+                                          const vpx_codec_enc_cfg_t  *cfg) {
+  vpx_codec_err_t res;
+
+  if (!ctx || !ctx->iface || !ctx->priv || !cfg)
+    res = VPX_CODEC_INVALID_PARAM;
+  else if (!(ctx->iface->caps & VPX_CODEC_CAP_ENCODER))
+    res = VPX_CODEC_INCAPABLE;
+  else
+    res = ctx->iface->enc.cfg_set(get_alg_priv(ctx), cfg);
+
+  return SAVE_STATUS(ctx, res);
+}
+
+
+int vpx_codec_pkt_list_add(struct vpx_codec_pkt_list *list,
+                           const struct vpx_codec_cx_pkt *pkt) {
+  if (list->cnt < list->max) {
+    list->pkts[list->cnt++] = *pkt;
+    return 0;
+  }
+
+  return 1;
+}
+
+
+const vpx_codec_cx_pkt_t *vpx_codec_pkt_list_get(struct vpx_codec_pkt_list *list,
+                                                 vpx_codec_iter_t           *iter) {
+  const vpx_codec_cx_pkt_t *pkt;
+
+  if (!(*iter)) {
+    *iter = list->pkts;
+  }
+
+  pkt = (const vpx_codec_cx_pkt_t *)*iter;
+
+  if ((size_t)(pkt - list->pkts) < list->cnt)
+    *iter = pkt + 1;
+  else
+    pkt = NULL;
+
+  return pkt;
+}
diff --git a/libvpx/libvpx/vpx/src/vpx_image.c b/libvpx/libvpx/vpx/src/vpx_image.c
new file mode 100644
index 0000000..9aae12c
--- /dev/null
+++ b/libvpx/libvpx/vpx/src/vpx_image.c
@@ -0,0 +1,285 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vpx_image.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+
+static vpx_image_t *img_alloc_helper(vpx_image_t *img,
+                                     vpx_img_fmt_t fmt,
+                                     unsigned int d_w,
+                                     unsigned int d_h,
+                                     unsigned int buf_align,
+                                     unsigned int stride_align,
+                                     unsigned char *img_data) {
+  unsigned int h, w, s, xcs, ycs, bps;
+  unsigned int stride_in_bytes;
+  int align;
+
+  /* Treat align==0 like align==1 */
+  if (!buf_align)
+    buf_align = 1;
+
+  /* Validate alignment (must be power of 2) */
+  if (buf_align & (buf_align - 1))
+    goto fail;
+
+  /* Treat align==0 like align==1 */
+  if (!stride_align)
+    stride_align = 1;
+
+  /* Validate alignment (must be power of 2) */
+  if (stride_align & (stride_align - 1))
+    goto fail;
+
+  /* Get sample size for this format */
+  switch (fmt) {
+    case VPX_IMG_FMT_RGB32:
+    case VPX_IMG_FMT_RGB32_LE:
+    case VPX_IMG_FMT_ARGB:
+    case VPX_IMG_FMT_ARGB_LE:
+      bps = 32;
+      break;
+    case VPX_IMG_FMT_RGB24:
+    case VPX_IMG_FMT_BGR24:
+      bps = 24;
+      break;
+    case VPX_IMG_FMT_RGB565:
+    case VPX_IMG_FMT_RGB565_LE:
+    case VPX_IMG_FMT_RGB555:
+    case VPX_IMG_FMT_RGB555_LE:
+    case VPX_IMG_FMT_UYVY:
+    case VPX_IMG_FMT_YUY2:
+    case VPX_IMG_FMT_YVYU:
+      bps = 16;
+      break;
+    case VPX_IMG_FMT_I420:
+    case VPX_IMG_FMT_YV12:
+    case VPX_IMG_FMT_VPXI420:
+    case VPX_IMG_FMT_VPXYV12:
+      bps = 12;
+      break;
+    case VPX_IMG_FMT_I422:
+    case VPX_IMG_FMT_I440:
+      bps = 16;
+      break;
+    case VPX_IMG_FMT_I444:
+      bps = 24;
+      break;
+    case VPX_IMG_FMT_I42016:
+      bps = 24;
+      break;
+    case VPX_IMG_FMT_I42216:
+    case VPX_IMG_FMT_I44016:
+      bps = 32;
+      break;
+    case VPX_IMG_FMT_I44416:
+      bps = 48;
+      break;
+    default:
+      bps = 16;
+      break;
+  }
+
+  /* Get chroma shift values for this format */
+  switch (fmt) {
+    case VPX_IMG_FMT_I420:
+    case VPX_IMG_FMT_YV12:
+    case VPX_IMG_FMT_VPXI420:
+    case VPX_IMG_FMT_VPXYV12:
+    case VPX_IMG_FMT_I422:
+    case VPX_IMG_FMT_I42016:
+    case VPX_IMG_FMT_I42216:
+      xcs = 1;
+      break;
+    default:
+      xcs = 0;
+      break;
+  }
+
+  switch (fmt) {
+    case VPX_IMG_FMT_I420:
+    case VPX_IMG_FMT_I440:
+    case VPX_IMG_FMT_YV12:
+    case VPX_IMG_FMT_VPXI420:
+    case VPX_IMG_FMT_VPXYV12:
+    case VPX_IMG_FMT_I42016:
+    case VPX_IMG_FMT_I44016:
+      ycs = 1;
+      break;
+    default:
+      ycs = 0;
+      break;
+  }
+
+  /* Calculate storage sizes given the chroma subsampling */
+  align = (1 << xcs) - 1;
+  w = (d_w + align) & ~align;
+  align = (1 << ycs) - 1;
+  h = (d_h + align) & ~align;
+  s = (fmt & VPX_IMG_FMT_PLANAR) ? w : bps * w / 8;
+  s = (s + stride_align - 1) & ~(stride_align - 1);
+  stride_in_bytes = (fmt & VPX_IMG_FMT_HIGHBITDEPTH) ? s * 2 : s;
+
+  /* Allocate the new image */
+  if (!img) {
+    img = (vpx_image_t *)calloc(1, sizeof(vpx_image_t));
+
+    if (!img)
+      goto fail;
+
+    img->self_allocd = 1;
+  } else {
+    memset(img, 0, sizeof(vpx_image_t));
+  }
+
+  img->img_data = img_data;
+
+  if (!img_data) {
+    const uint64_t alloc_size = (fmt & VPX_IMG_FMT_PLANAR) ?
+                                (uint64_t)h * s * bps / 8 : (uint64_t)h * s;
+
+    if (alloc_size != (size_t)alloc_size)
+      goto fail;
+
+    img->img_data = (uint8_t *)vpx_memalign(buf_align, (size_t)alloc_size);
+    img->img_data_owner = 1;
+  }
+
+  if (!img->img_data)
+    goto fail;
+
+  img->fmt = fmt;
+  img->bit_depth = (fmt & VPX_IMG_FMT_HIGHBITDEPTH) ? 16 : 8;
+  img->w = w;
+  img->h = h;
+  img->x_chroma_shift = xcs;
+  img->y_chroma_shift = ycs;
+  img->bps = bps;
+
+  /* Calculate strides */
+  img->stride[VPX_PLANE_Y] = img->stride[VPX_PLANE_ALPHA] = stride_in_bytes;
+  img->stride[VPX_PLANE_U] = img->stride[VPX_PLANE_V] = stride_in_bytes >> xcs;
+
+  /* Default viewport to entire image */
+  if (!vpx_img_set_rect(img, 0, 0, d_w, d_h))
+    return img;
+
+fail:
+  vpx_img_free(img);
+  return NULL;
+}
+
+vpx_image_t *vpx_img_alloc(vpx_image_t  *img,
+                           vpx_img_fmt_t fmt,
+                           unsigned int  d_w,
+                           unsigned int  d_h,
+                           unsigned int  align) {
+  return img_alloc_helper(img, fmt, d_w, d_h, align, align, NULL);
+}
+
+vpx_image_t *vpx_img_wrap(vpx_image_t  *img,
+                          vpx_img_fmt_t fmt,
+                          unsigned int  d_w,
+                          unsigned int  d_h,
+                          unsigned int  stride_align,
+                          unsigned char       *img_data) {
+  /* By setting buf_align = 1, we don't change buffer alignment in this
+   * function. */
+  return img_alloc_helper(img, fmt, d_w, d_h, 1, stride_align, img_data);
+}
+
+int vpx_img_set_rect(vpx_image_t  *img,
+                     unsigned int  x,
+                     unsigned int  y,
+                     unsigned int  w,
+                     unsigned int  h) {
+  unsigned char      *data;
+
+  if (x + w <= img->w && y + h <= img->h) {
+    img->d_w = w;
+    img->d_h = h;
+
+    /* Calculate plane pointers */
+    if (!(img->fmt & VPX_IMG_FMT_PLANAR)) {
+      img->planes[VPX_PLANE_PACKED] =
+        img->img_data + x * img->bps / 8 + y * img->stride[VPX_PLANE_PACKED];
+    } else {
+      const int bytes_per_sample =
+          (img->fmt & VPX_IMG_FMT_HIGHBITDEPTH) ? 2 : 1;
+      data = img->img_data;
+
+      if (img->fmt & VPX_IMG_FMT_HAS_ALPHA) {
+        img->planes[VPX_PLANE_ALPHA] =
+            data + x * bytes_per_sample + y * img->stride[VPX_PLANE_ALPHA];
+        data += img->h * img->stride[VPX_PLANE_ALPHA];
+      }
+
+      img->planes[VPX_PLANE_Y] = data + x * bytes_per_sample +
+          y * img->stride[VPX_PLANE_Y];
+      data += img->h * img->stride[VPX_PLANE_Y];
+
+      if (!(img->fmt & VPX_IMG_FMT_UV_FLIP)) {
+        img->planes[VPX_PLANE_U] =
+            data + (x >> img->x_chroma_shift) * bytes_per_sample +
+            (y >> img->y_chroma_shift) * img->stride[VPX_PLANE_U];
+        data += (img->h >> img->y_chroma_shift) * img->stride[VPX_PLANE_U];
+        img->planes[VPX_PLANE_V] =
+            data + (x >> img->x_chroma_shift) * bytes_per_sample +
+            (y >> img->y_chroma_shift) * img->stride[VPX_PLANE_V];
+      } else {
+        img->planes[VPX_PLANE_V] =
+            data + (x >> img->x_chroma_shift) * bytes_per_sample +
+            (y >> img->y_chroma_shift) * img->stride[VPX_PLANE_V];
+        data += (img->h >> img->y_chroma_shift) * img->stride[VPX_PLANE_V];
+        img->planes[VPX_PLANE_U] =
+            data + (x >> img->x_chroma_shift) * bytes_per_sample +
+            (y >> img->y_chroma_shift) * img->stride[VPX_PLANE_U];
+      }
+    }
+    return 0;
+  }
+  return -1;
+}
+
+void vpx_img_flip(vpx_image_t *img) {
+  /* Note: In the calculation pointer adjustment calculation, we want the
+   * rhs to be promoted to a signed type. Section 6.3.1.8 of the ISO C99
+   * standard indicates that if the adjustment parameter is unsigned, the
+   * stride parameter will be promoted to unsigned, causing errors when
+   * the lhs is a larger type than the rhs.
+   */
+  img->planes[VPX_PLANE_Y] += (signed)(img->d_h - 1) * img->stride[VPX_PLANE_Y];
+  img->stride[VPX_PLANE_Y] = -img->stride[VPX_PLANE_Y];
+
+  img->planes[VPX_PLANE_U] += (signed)((img->d_h >> img->y_chroma_shift) - 1)
+                              * img->stride[VPX_PLANE_U];
+  img->stride[VPX_PLANE_U] = -img->stride[VPX_PLANE_U];
+
+  img->planes[VPX_PLANE_V] += (signed)((img->d_h >> img->y_chroma_shift) - 1)
+                              * img->stride[VPX_PLANE_V];
+  img->stride[VPX_PLANE_V] = -img->stride[VPX_PLANE_V];
+
+  img->planes[VPX_PLANE_ALPHA] += (signed)(img->d_h - 1) * img->stride[VPX_PLANE_ALPHA];
+  img->stride[VPX_PLANE_ALPHA] = -img->stride[VPX_PLANE_ALPHA];
+}
+
+void vpx_img_free(vpx_image_t *img) {
+  if (img) {
+    if (img->img_data && img->img_data_owner)
+      vpx_free(img->img_data);
+
+    if (img->self_allocd)
+      free(img);
+  }
+}
diff --git a/libvpx/libvpx/vpx/src/vpx_psnr.c b/libvpx/libvpx/vpx/src/vpx_psnr.c
new file mode 100644
index 0000000..05843ac
--- /dev/null
+++ b/libvpx/libvpx/vpx/src/vpx_psnr.c
@@ -0,0 +1,24 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "vpx/internal/vpx_psnr.h"
+
+#define MAX_PSNR 100.0
+
+double vpx_sse_to_psnr(double samples, double peak, double sse) {
+  if (sse > 0.0) {
+    const double psnr = 10.0 * log10(samples * peak * peak / sse);
+    return psnr > MAX_PSNR ? MAX_PSNR : psnr;
+  } else {
+    return MAX_PSNR;
+  }
+}
diff --git a/libvpx/libvpx/vpx/svc_context.h b/libvpx/libvpx/vpx/svc_context.h
new file mode 100644
index 0000000..5bc2518
--- /dev/null
+++ b/libvpx/libvpx/vpx/svc_context.h
@@ -0,0 +1,123 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+/**
+ * SvcContext - input parameters and state to encode a multi-layered
+ * spatial SVC frame
+ */
+
+#ifndef VPX_SVC_CONTEXT_H_
+#define VPX_SVC_CONTEXT_H_
+
+#include "./vp8cx.h"
+#include "./vpx_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef enum SVC_LOG_LEVEL {
+  SVC_LOG_ERROR,
+  SVC_LOG_INFO,
+  SVC_LOG_DEBUG
+} SVC_LOG_LEVEL;
+
+typedef struct {
+  // public interface to svc_command options
+  int spatial_layers;               // number of spatial layers
+  int temporal_layers;               // number of temporal layers
+  int temporal_layering_mode;
+  SVC_LOG_LEVEL log_level;  // amount of information to display
+  int log_print;  // when set, printf log messages instead of returning the
+                  // message with svc_get_message
+  int output_rc_stat;  // for outputting rc stats
+  int speed;  // speed setting for codec
+  int threads;
+  int aqmode;  // turns on aq-mode=3 (cyclic_refresh): 0=off, 1=on.
+  // private storage for vpx_svc_encode
+  void *internal;
+} SvcContext;
+
+#define OPTION_BUFFER_SIZE 1024
+#define COMPONENTS 4  // psnr & sse statistics maintained for total, y, u, v
+
+typedef struct SvcInternal {
+  char options[OPTION_BUFFER_SIZE];        // set by vpx_svc_set_options
+
+  // values extracted from option, quantizers
+  vpx_svc_extra_cfg_t svc_params;
+  int enable_auto_alt_ref[VPX_SS_MAX_LAYERS];
+  int bitrates[VPX_SS_MAX_LAYERS];
+
+  // accumulated statistics
+  double psnr_sum[VPX_SS_MAX_LAYERS][COMPONENTS];   // total/Y/U/V
+  uint64_t sse_sum[VPX_SS_MAX_LAYERS][COMPONENTS];
+  uint32_t bytes_sum[VPX_SS_MAX_LAYERS];
+
+  // codec encoding values
+  int width;    // width of highest layer
+  int height;   // height of highest layer
+  int kf_dist;  // distance between keyframes
+
+  // state variables
+  int psnr_pkt_received;
+  int layer;
+  int use_multiple_frame_contexts;
+
+  char message_buffer[2048];
+  vpx_codec_ctx_t *codec_ctx;
+} SvcInternal_t;
+
+/**
+ * Set SVC options
+ * options are supplied as a single string separated by spaces
+ * Format: encoding-mode=<i|ip|alt-ip|gf>
+ *         layers=<layer_count>
+ *         scaling-factors=<n1>/<d1>,<n2>/<d2>,...
+ *         quantizers=<q1>,<q2>,...
+ */
+vpx_codec_err_t vpx_svc_set_options(SvcContext *svc_ctx, const char *options);
+
+/**
+ * initialize SVC encoding
+ */
+vpx_codec_err_t vpx_svc_init(SvcContext *svc_ctx,
+                             vpx_codec_ctx_t *codec_ctx,
+                             vpx_codec_iface_t *iface,
+                             vpx_codec_enc_cfg_t *cfg);
+/**
+ * encode a frame of video with multiple layers
+ */
+vpx_codec_err_t vpx_svc_encode(SvcContext *svc_ctx,
+                               vpx_codec_ctx_t *codec_ctx,
+                               struct vpx_image *rawimg,
+                               vpx_codec_pts_t pts,
+                               int64_t duration, int deadline);
+
+/**
+ * finished with svc encoding, release allocated resources
+ */
+void vpx_svc_release(SvcContext *svc_ctx);
+
+/**
+ * dump accumulated statistics and reset accumulated values
+ */
+const char *vpx_svc_dump_statistics(SvcContext *svc_ctx);
+
+/**
+ *  get status message from previous encode
+ */
+const char *vpx_svc_get_message(const SvcContext *svc_ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_SVC_CONTEXT_H_
diff --git a/libvpx/libvpx/vpx/vp8.h b/libvpx/libvpx/vpx/vp8.h
new file mode 100644
index 0000000..8a035f9
--- /dev/null
+++ b/libvpx/libvpx/vpx/vp8.h
@@ -0,0 +1,148 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+/*!\defgroup vp8 VP8
+ * \ingroup codecs
+ * VP8 is vpx's newest video compression algorithm that uses motion
+ * compensated prediction, Discrete Cosine Transform (DCT) coding of the
+ * prediction error signal and context dependent entropy coding techniques
+ * based on arithmetic principles. It features:
+ *  - YUV 4:2:0 image format
+ *  - Macro-block based coding (16x16 luma plus two 8x8 chroma)
+ *  - 1/4 (1/8) pixel accuracy motion compensated prediction
+ *  - 4x4 DCT transform
+ *  - 128 level linear quantizer
+ *  - In loop deblocking filter
+ *  - Context-based entropy coding
+ *
+ * @{
+ */
+/*!\file
+ * \brief Provides controls common to both the VP8 encoder and decoder.
+ */
+#ifndef VPX_VP8_H_
+#define VPX_VP8_H_
+
+#include "./vpx_codec.h"
+#include "./vpx_image.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*!\brief Control functions
+ *
+ * The set of macros define the control functions of VP8 interface
+ */
+enum vp8_com_control_id {
+  VP8_SET_REFERENCE           = 1,    /**< pass in an external frame into decoder to be used as reference frame */
+  VP8_COPY_REFERENCE          = 2,    /**< get a copy of reference frame from the decoder */
+  VP8_SET_POSTPROC            = 3,    /**< set the decoder's post processing settings  */
+  VP8_SET_DBG_COLOR_REF_FRAME = 4,    /**< set the reference frames to color for each macroblock */
+  VP8_SET_DBG_COLOR_MB_MODES  = 5,    /**< set which macro block modes to color */
+  VP8_SET_DBG_COLOR_B_MODES   = 6,    /**< set which blocks modes to color */
+  VP8_SET_DBG_DISPLAY_MV      = 7,    /**< set which motion vector modes to draw */
+
+  /* TODO(jkoleszar): The encoder incorrectly reuses some of these values (5+)
+   * for its control ids. These should be migrated to something like the
+   * VP8_DECODER_CTRL_ID_START range next time we're ready to break the ABI.
+   */
+  VP9_GET_REFERENCE           = 128,  /**< get a pointer to a reference frame */
+  VP8_COMMON_CTRL_ID_MAX,
+  VP8_DECODER_CTRL_ID_START   = 256
+};
+
+/*!\brief post process flags
+ *
+ * The set of macros define VP8 decoder post processing flags
+ */
+enum vp8_postproc_level {
+  VP8_NOFILTERING             = 0,
+  VP8_DEBLOCK                 = 1 << 0,
+  VP8_DEMACROBLOCK            = 1 << 1,
+  VP8_ADDNOISE                = 1 << 2,
+  VP8_DEBUG_TXT_FRAME_INFO    = 1 << 3, /**< print frame information */
+  VP8_DEBUG_TXT_MBLK_MODES    = 1 << 4, /**< print macro block modes over each macro block */
+  VP8_DEBUG_TXT_DC_DIFF       = 1 << 5, /**< print dc diff for each macro block */
+  VP8_DEBUG_TXT_RATE_INFO     = 1 << 6, /**< print video rate info (encoder only) */
+  VP8_MFQE                    = 1 << 10
+};
+
+/*!\brief post process flags
+ *
+ * This define a structure that describe the post processing settings. For
+ * the best objective measure (using the PSNR metric) set post_proc_flag
+ * to VP8_DEBLOCK and deblocking_level to 1.
+ */
+
+typedef struct vp8_postproc_cfg {
+  int post_proc_flag;         /**< the types of post processing to be done, should be combination of "vp8_postproc_level" */
+  int deblocking_level;       /**< the strength of deblocking, valid range [0, 16] */
+  int noise_level;            /**< the strength of additive noise, valid range [0, 16] */
+} vp8_postproc_cfg_t;
+
+/*!\brief reference frame type
+ *
+ * The set of macros define the type of VP8 reference frames
+ */
+typedef enum vpx_ref_frame_type {
+  VP8_LAST_FRAME = 1,
+  VP8_GOLD_FRAME = 2,
+  VP8_ALTR_FRAME = 4
+} vpx_ref_frame_type_t;
+
+/*!\brief reference frame data struct
+ *
+ * Define the data struct to access vp8 reference frames.
+ */
+typedef struct vpx_ref_frame {
+  vpx_ref_frame_type_t  frame_type;   /**< which reference frame */
+  vpx_image_t           img;          /**< reference frame data in image format */
+} vpx_ref_frame_t;
+
+/*!\brief VP9 specific reference frame data struct
+ *
+ * Define the data struct to access vp9 reference frames.
+ */
+typedef struct vp9_ref_frame {
+  int idx; /**< frame index to get (input) */
+  vpx_image_t  img; /**< img structure to populate (output) */
+} vp9_ref_frame_t;
+
+/*!\cond */
+/*!\brief vp8 decoder control function parameter type
+ *
+ * defines the data type for each of VP8 decoder control function requires
+ */
+VPX_CTRL_USE_TYPE(VP8_SET_REFERENCE,           vpx_ref_frame_t *)
+#define VPX_CTRL_VP8_SET_REFERENCE
+VPX_CTRL_USE_TYPE(VP8_COPY_REFERENCE,          vpx_ref_frame_t *)
+#define VPX_CTRL_VP8_COPY_REFERENCE
+VPX_CTRL_USE_TYPE(VP8_SET_POSTPROC,            vp8_postproc_cfg_t *)
+#define VPX_CTRL_VP8_SET_POSTPROC
+VPX_CTRL_USE_TYPE(VP8_SET_DBG_COLOR_REF_FRAME, int)
+#define VPX_CTRL_VP8_SET_DBG_COLOR_REF_FRAME
+VPX_CTRL_USE_TYPE(VP8_SET_DBG_COLOR_MB_MODES,  int)
+#define VPX_CTRL_VP8_SET_DBG_COLOR_MB_MODES
+VPX_CTRL_USE_TYPE(VP8_SET_DBG_COLOR_B_MODES,   int)
+#define VPX_CTRL_VP8_SET_DBG_COLOR_B_MODES
+VPX_CTRL_USE_TYPE(VP8_SET_DBG_DISPLAY_MV,      int)
+#define VPX_CTRL_VP8_SET_DBG_DISPLAY_MV
+VPX_CTRL_USE_TYPE(VP9_GET_REFERENCE,           vp9_ref_frame_t *)
+#define VPX_CTRL_VP9_GET_REFERENCE
+
+/*!\endcond */
+/*! @} - end defgroup vp8 */
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_VP8_H_
diff --git a/libvpx/libvpx/vpx/vp8cx.h b/libvpx/libvpx/vpx/vp8cx.h
new file mode 100644
index 0000000..61882e6
--- /dev/null
+++ b/libvpx/libvpx/vpx/vp8cx.h
@@ -0,0 +1,830 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VPX_VP8CX_H_
+#define VPX_VP8CX_H_
+
+/*!\defgroup vp8_encoder WebM VP8/VP9 Encoder
+ * \ingroup vp8
+ *
+ * @{
+ */
+#include "./vp8.h"
+#include "./vpx_encoder.h"
+
+/*!\file
+ * \brief Provides definitions for using VP8 or VP9 encoder algorithm within the
+ *        vpx Codec Interface.
+ */
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*!\name Algorithm interface for VP8
+ *
+ * This interface provides the capability to encode raw VP8 streams.
+ * @{
+ */
+extern vpx_codec_iface_t  vpx_codec_vp8_cx_algo;
+extern vpx_codec_iface_t *vpx_codec_vp8_cx(void);
+/*!@} - end algorithm interface member group*/
+
+/*!\name Algorithm interface for VP9
+ *
+ * This interface provides the capability to encode raw VP9 streams.
+ * @{
+ */
+extern vpx_codec_iface_t  vpx_codec_vp9_cx_algo;
+extern vpx_codec_iface_t *vpx_codec_vp9_cx(void);
+/*!@} - end algorithm interface member group*/
+
+/*
+ * Algorithm Flags
+ */
+
+/*!\brief Don't reference the last frame
+ *
+ * When this flag is set, the encoder will not use the last frame as a
+ * predictor. When not set, the encoder will choose whether to use the
+ * last frame or not automatically.
+ */
+#define VP8_EFLAG_NO_REF_LAST      (1<<16)
+
+
+/*!\brief Don't reference the golden frame
+ *
+ * When this flag is set, the encoder will not use the golden frame as a
+ * predictor. When not set, the encoder will choose whether to use the
+ * golden frame or not automatically.
+ */
+#define VP8_EFLAG_NO_REF_GF        (1<<17)
+
+
+/*!\brief Don't reference the alternate reference frame
+ *
+ * When this flag is set, the encoder will not use the alt ref frame as a
+ * predictor. When not set, the encoder will choose whether to use the
+ * alt ref frame or not automatically.
+ */
+#define VP8_EFLAG_NO_REF_ARF       (1<<21)
+
+
+/*!\brief Don't update the last frame
+ *
+ * When this flag is set, the encoder will not update the last frame with
+ * the contents of the current frame.
+ */
+#define VP8_EFLAG_NO_UPD_LAST      (1<<18)
+
+
+/*!\brief Don't update the golden frame
+ *
+ * When this flag is set, the encoder will not update the golden frame with
+ * the contents of the current frame.
+ */
+#define VP8_EFLAG_NO_UPD_GF        (1<<22)
+
+
+/*!\brief Don't update the alternate reference frame
+ *
+ * When this flag is set, the encoder will not update the alt ref frame with
+ * the contents of the current frame.
+ */
+#define VP8_EFLAG_NO_UPD_ARF       (1<<23)
+
+
+/*!\brief Force golden frame update
+ *
+ * When this flag is set, the encoder copy the contents of the current frame
+ * to the golden frame buffer.
+ */
+#define VP8_EFLAG_FORCE_GF         (1<<19)
+
+
+/*!\brief Force alternate reference frame update
+ *
+ * When this flag is set, the encoder copy the contents of the current frame
+ * to the alternate reference frame buffer.
+ */
+#define VP8_EFLAG_FORCE_ARF        (1<<24)
+
+
+/*!\brief Disable entropy update
+ *
+ * When this flag is set, the encoder will not update its internal entropy
+ * model based on the entropy of this frame.
+ */
+#define VP8_EFLAG_NO_UPD_ENTROPY   (1<<20)
+
+
+/*!\brief VPx encoder control functions
+ *
+ * This set of macros define the control functions available for VPx
+ * encoder interface.
+ *
+ * \sa #vpx_codec_control
+ */
+enum vp8e_enc_control_id {
+  /*!\brief Codec control function to pass an ROI map to encoder.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_ROI_MAP           = 8,
+
+  /*!\brief Codec control function to pass an Active map to encoder.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_ACTIVEMAP,
+
+  /*!\brief Codec control function to set encoder scaling mode.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_SCALEMODE         = 11,
+
+  /*!\brief Codec control function to set encoder internal speed settings.
+   *
+   * Changes in this value influences, among others, the encoder's selection
+   * of motion estimation methods. Values greater than 0 will increase encoder
+   * speed at the expense of quality.
+   *
+   * \note Valid range for VP8: -16..16
+   * \note Valid range for VP9: -8..8
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_CPUUSED           = 13,
+
+  /*!\brief Codec control function to enable automatic set and use alf frames.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_ENABLEAUTOALTREF,
+
+  /*!\brief control function to set noise sensitivity
+   *
+   * 0: off, 1: OnYOnly, 2: OnYUV,
+   * 3: OnYUVAggressive, 4: Adaptive
+   *
+   * Supported in codecs: VP8
+   */
+  VP8E_SET_NOISE_SENSITIVITY,
+
+  /*!\brief Codec control function to set sharpness.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_SHARPNESS,
+
+  /*!\brief Codec control function to set the threshold for MBs treated static.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_STATIC_THRESHOLD,
+
+  /*!\brief Codec control function to set the number of token partitions.
+   *
+   * Supported in codecs: VP8
+   */
+  VP8E_SET_TOKEN_PARTITIONS,
+
+  /*!\brief Codec control function to get last quantizer chosen by the encoder.
+   *
+   * Return value uses internal quantizer scale defined by the codec.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_GET_LAST_QUANTIZER,
+
+  /*!\brief Codec control function to get last quantizer chosen by the encoder.
+   *
+   * Return value uses the 0..63 scale as used by the rc_*_quantizer config
+   * parameters.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_GET_LAST_QUANTIZER_64,
+
+  /*!\brief Codec control function to set the max no of frames to create arf.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_ARNR_MAXFRAMES,
+
+  /*!\brief Codec control function to set the filter strength for the arf.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_ARNR_STRENGTH,
+
+  /*!\deprecated control function to set the filter type to use for the arf. */
+  VP8E_SET_ARNR_TYPE,
+
+  /*!\brief Codec control function to set visual tuning.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_TUNING,
+
+  /*!\brief Codec control function to set constrained quality level.
+   *
+   * \attention For this value to be used vpx_codec_enc_cfg_t::g_usage must be
+   *            set to #VPX_CQ.
+   * \note Valid range: 0..63
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_CQ_LEVEL,
+
+  /*!\brief Codec control function to set Max data rate for Intra frames.
+   *
+   * This value controls additional clamping on the maximum size of a
+   * keyframe. It is expressed as a percentage of the average
+   * per-frame bitrate, with the special (and default) value 0 meaning
+   * unlimited, or no additional clamping beyond the codec's built-in
+   * algorithm.
+   *
+   * For example, to allocate no more than 4.5 frames worth of bitrate
+   * to a keyframe, set this to 450.
+   *
+   * Supported in codecs: VP8, VP9
+   */
+  VP8E_SET_MAX_INTRA_BITRATE_PCT,
+
+  /*!\brief Codec control function to set reference and update frame flags.
+   *
+   *  Supported in codecs: VP8
+   */
+  VP8E_SET_FRAME_FLAGS,
+
+  /*!\brief Codec control function to set max data rate for Inter frames.
+   *
+   * This value controls additional clamping on the maximum size of an
+   * inter frame. It is expressed as a percentage of the average
+   * per-frame bitrate, with the special (and default) value 0 meaning
+   * unlimited, or no additional clamping beyond the codec's built-in
+   * algorithm.
+   *
+   * For example, to allow no more than 4.5 frames worth of bitrate
+   * to an inter frame, set this to 450.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_MAX_INTER_BITRATE_PCT,
+
+  /*!\brief Boost percentage for Golden Frame in CBR mode.
+   *
+   * This value controls the amount of boost given to Golden Frame in
+   * CBR mode. It is expressed as a percentage of the average
+   * per-frame bitrate, with the special (and default) value 0 meaning
+   * the feature is off, i.e., no golden frame boost in CBR mode and
+   * average bitrate target is used.
+   *
+   * For example, to allow 100% more bits, i.e, 2X, in a golden frame
+   * than average frame, set this to 100.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_GF_CBR_BOOST_PCT,
+
+  /*!\brief Codec control function to set the temporal layer id.
+   *
+   * For temporal scalability: this control allows the application to set the
+   * layer id for each frame to be encoded. Note that this control must be set
+   * for every frame prior to encoding. The usage of this control function
+   * supersedes the internal temporal pattern counter, which is now deprecated.
+   *
+   * Supported in codecs: VP8
+   */
+  VP8E_SET_TEMPORAL_LAYER_ID,
+
+  /*!\brief Codec control function to set encoder screen content mode.
+   *
+   * 0: off, 1: On, 2: On with more aggressive rate control.
+   *
+   * Supported in codecs: VP8
+   */
+  VP8E_SET_SCREEN_CONTENT_MODE,
+
+  /*!\brief Codec control function to set lossless encoding mode.
+   *
+   * VP9 can operate in lossless encoding mode, in which the bitstream
+   * produced will be able to decode and reconstruct a perfect copy of
+   * input source. This control function provides a mean to switch encoder
+   * into lossless coding mode(1) or normal coding mode(0) that may be lossy.
+   *                          0 = lossy coding mode
+   *                          1 = lossless coding mode
+   *
+   *  By default, encoder operates in normal coding mode (maybe lossy).
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_LOSSLESS,
+
+  /*!\brief Codec control function to set number of tile columns.
+   *
+   * In encoding and decoding, VP9 allows an input image frame be partitioned
+   * into separated vertical tile columns, which can be encoded or decoded
+   * independently. This enables easy implementation of parallel encoding and
+   * decoding. This control requests the encoder to use column tiles in
+   * encoding an input frame, with number of tile columns (in Log2 unit) as
+   * the parameter:
+   *             0 = 1 tile column
+   *             1 = 2 tile columns
+   *             2 = 4 tile columns
+   *             .....
+   *             n = 2**n tile columns
+   * The requested tile columns will be capped by encoder based on image size
+   * limitation (The minimum width of a tile column is 256 pixel, the maximum
+   * is 4096).
+   *
+   * By default, the value is 0, i.e. one single column tile for entire image.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_TILE_COLUMNS,
+
+  /*!\brief Codec control function to set number of tile rows.
+   *
+   * In encoding and decoding, VP9 allows an input image frame be partitioned
+   * into separated horizontal tile rows. Tile rows are encoded or decoded
+   * sequentially. Even though encoding/decoding of later tile rows depends on
+   * earlier ones, this allows the encoder to output data packets for tile rows
+   * prior to completely processing all tile rows in a frame, thereby reducing
+   * the latency in processing between input and output. The parameter
+   * for this control describes the number of tile rows, which has a valid
+   * range [0, 2]:
+   *            0 = 1 tile row
+   *            1 = 2 tile rows
+   *            2 = 4 tile rows
+   *
+   * By default, the value is 0, i.e. one single row tile for entire image.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_TILE_ROWS,
+
+  /*!\brief Codec control function to enable frame parallel decoding feature.
+   *
+   * VP9 has a bitstream feature to reduce decoding dependency between frames
+   * by turning off backward update of probability context used in encoding
+   * and decoding. This allows staged parallel processing of more than one
+   * video frames in the decoder. This control function provides a mean to
+   * turn this feature on or off for bitstreams produced by encoder.
+   *
+   * By default, this feature is off.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_FRAME_PARALLEL_DECODING,
+
+  /*!\brief Codec control function to set adaptive quantization mode.
+   *
+   * VP9 has a segment based feature that allows encoder to adaptively change
+   * quantization parameter for each segment within a frame to improve the
+   * subjective quality. This control makes encoder operate in one of the
+   * several AQ_modes supported.
+   *
+   * By default, encoder operates with AQ_Mode 0(adaptive quantization off).
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_AQ_MODE,
+
+  /*!\brief Codec control function to enable/disable periodic Q boost.
+   *
+   * One VP9 encoder speed feature is to enable quality boost by lowering
+   * frame level Q periodically. This control function provides a mean to
+   * turn on/off this feature.
+   *               0 = off
+   *               1 = on
+   *
+   * By default, the encoder is allowed to use this feature for appropriate
+   * encoding modes.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_FRAME_PERIODIC_BOOST,
+
+  /*!\brief Codec control function to set noise sensitivity.
+   *
+   *  0: off, 1: On(YOnly)
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_NOISE_SENSITIVITY,
+
+  /*!\brief Codec control function to turn on/off SVC in encoder.
+   * \note Return value is VPX_CODEC_INVALID_PARAM if the encoder does not
+   *       support SVC in its current encoding mode
+   *  0: off, 1: on
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_SVC,
+
+  /*!\brief Codec control function to set parameters for SVC.
+   * \note Parameters contain min_q, max_q, scaling factor for each of the
+   *       SVC layers.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_SVC_PARAMETERS,
+
+  /*!\brief Codec control function to set svc layer for spatial and temporal.
+   * \note Valid ranges: 0..#vpx_codec_enc_cfg::ss_number_layers for spatial
+   *                     layer and 0..#vpx_codec_enc_cfg::ts_number_layers for
+   *                     temporal layer.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_SVC_LAYER_ID,
+
+  /*!\brief Codec control function to set content type.
+   * \note Valid parameter range:
+   *              VP9E_CONTENT_DEFAULT = Regular video content (Default)
+   *              VP9E_CONTENT_SCREEN  = Screen capture content
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_TUNE_CONTENT,
+
+  /*!\brief Codec control function to get svc layer ID.
+   * \note The layer ID returned is for the data packet from the registered
+   *       callback function.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_GET_SVC_LAYER_ID,
+
+  /*!\brief Codec control function to register callback to get per layer packet.
+   * \note Parameter for this control function is a structure with a callback
+   *       function and a pointer to private data used by the callback.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_REGISTER_CX_CALLBACK,
+
+  /*!\brief Codec control function to set color space info.
+   * \note Valid ranges: 0..7, default is "UNKNOWN".
+   *                     0 = UNKNOWN,
+   *                     1 = BT_601
+   *                     2 = BT_709
+   *                     3 = SMPTE_170
+   *                     4 = SMPTE_240
+   *                     5 = BT_2020
+   *                     6 = RESERVED
+   *                     7 = SRGB
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_COLOR_SPACE,
+
+  /*!\brief Codec control function to set temporal layering mode.
+   * \note Valid ranges: 0..3, default is "0" (VP9E_TEMPORAL_LAYERING_MODE_NOLAYERING).
+   *                     0 = VP9E_TEMPORAL_LAYERING_MODE_NOLAYERING
+   *                     1 = VP9E_TEMPORAL_LAYERING_MODE_BYPASS
+   *                     2 = VP9E_TEMPORAL_LAYERING_MODE_0101
+   *                     3 = VP9E_TEMPORAL_LAYERING_MODE_0212
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_TEMPORAL_LAYERING_MODE,
+
+  /*!\brief Codec control function to set minimum interval between GF/ARF frames
+   *
+   * By default the value is set as 4.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_MIN_GF_INTERVAL,
+
+  /*!\brief Codec control function to set minimum interval between GF/ARF frames
+   *
+   * By default the value is set as 16.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_MAX_GF_INTERVAL,
+
+  /*!\brief Codec control function to get an Active map back from the encoder.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_GET_ACTIVEMAP,
+
+  /*!\brief Codec control function to set color range bit.
+   * \note Valid ranges: 0..1, default is 0
+   *                     0 = Limited range (16..235 or HBD equivalent)
+   *                     1 = Full range (0..255 or HBD equivalent)
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_COLOR_RANGE,
+
+  /*!\brief Codec control function to set the frame flags and buffer indices
+   * for spatial layers. The frame flags and buffer indices are set using the
+   * struct #vpx_svc_ref_frame_config defined below.
+   *
+   * Supported in codecs: VP9
+  */
+  VP9E_SET_SVC_REF_FRAME_CONFIG,
+
+  /*!\brief Codec control function to set intended rendering image size.
+   *
+   * By default, this is identical to the image size in pixels.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_RENDER_SIZE,
+
+  /*!\brief Codec control function to set target level.
+   *
+   * 255: off (default); 0: only keep level stats; 10: target for level 1.0;
+   * 11: target for level 1.1; ... 62: target for level 6.2
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_TARGET_LEVEL,
+
+  /*!\brief Codec control function to get bitstream level.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_GET_LEVEL
+};
+
+/*!\brief vpx 1-D scaling mode
+ *
+ * This set of constants define 1-D vpx scaling modes
+ */
+typedef enum vpx_scaling_mode_1d {
+  VP8E_NORMAL      = 0,
+  VP8E_FOURFIVE    = 1,
+  VP8E_THREEFIVE   = 2,
+  VP8E_ONETWO      = 3
+} VPX_SCALING_MODE;
+
+/*!\brief Temporal layering mode enum for VP9 SVC.
+ *
+ * This set of macros define the different temporal layering modes.
+ * Supported codecs: VP9 (in SVC mode)
+ *
+ */
+typedef enum vp9e_temporal_layering_mode {
+  /*!\brief No temporal layering.
+   * Used when only spatial layering is used.
+   */
+  VP9E_TEMPORAL_LAYERING_MODE_NOLAYERING   = 0,
+
+  /*!\brief Bypass mode.
+   * Used when application needs to control temporal layering.
+   * This will only work when the number of spatial layers equals 1.
+   */
+  VP9E_TEMPORAL_LAYERING_MODE_BYPASS       = 1,
+
+  /*!\brief 0-1-0-1... temporal layering scheme with two temporal layers.
+   */
+  VP9E_TEMPORAL_LAYERING_MODE_0101         = 2,
+
+  /*!\brief 0-2-1-2... temporal layering scheme with three temporal layers.
+   */
+  VP9E_TEMPORAL_LAYERING_MODE_0212         = 3
+} VP9E_TEMPORAL_LAYERING_MODE;
+
+/*!\brief  vpx region of interest map
+ *
+ * These defines the data structures for the region of interest map
+ *
+ */
+
+typedef struct vpx_roi_map {
+  /*! An id between 0 and 3 for each 16x16 region within a frame. */
+  unsigned char *roi_map;
+  unsigned int rows;       /**< Number of rows. */
+  unsigned int cols;       /**< Number of columns. */
+  // TODO(paulwilkins): broken for VP9 which has 8 segments
+  // q and loop filter deltas for each segment
+  // (see MAX_MB_SEGMENTS)
+  int delta_q[4];          /**< Quantizer deltas. */
+  int delta_lf[4];         /**< Loop filter deltas. */
+  /*! Static breakout threshold for each segment. */
+  unsigned int static_threshold[4];
+} vpx_roi_map_t;
+
+/*!\brief  vpx active region map
+ *
+ * These defines the data structures for active region map
+ *
+ */
+
+
+typedef struct vpx_active_map {
+  unsigned char  *active_map; /**< specify an on (1) or off (0) each 16x16 region within a frame */
+  unsigned int    rows;       /**< number of rows */
+  unsigned int    cols;       /**< number of cols */
+} vpx_active_map_t;
+
+/*!\brief  vpx image scaling mode
+ *
+ * This defines the data structure for image scaling mode
+ *
+ */
+typedef struct vpx_scaling_mode {
+  VPX_SCALING_MODE    h_scaling_mode;  /**< horizontal scaling mode */
+  VPX_SCALING_MODE    v_scaling_mode;  /**< vertical scaling mode   */
+} vpx_scaling_mode_t;
+
+/*!\brief VP8 token partition mode
+ *
+ * This defines VP8 partitioning mode for compressed data, i.e., the number of
+ * sub-streams in the bitstream. Used for parallelized decoding.
+ *
+ */
+
+typedef enum {
+  VP8_ONE_TOKENPARTITION   = 0,
+  VP8_TWO_TOKENPARTITION   = 1,
+  VP8_FOUR_TOKENPARTITION  = 2,
+  VP8_EIGHT_TOKENPARTITION = 3
+} vp8e_token_partitions;
+
+/*!brief VP9 encoder content type */
+typedef enum {
+  VP9E_CONTENT_DEFAULT,
+  VP9E_CONTENT_SCREEN,
+  VP9E_CONTENT_INVALID
+} vp9e_tune_content;
+
+/*!\brief VP8 model tuning parameters
+ *
+ * Changes the encoder to tune for certain types of input material.
+ *
+ */
+typedef enum {
+  VP8_TUNE_PSNR,
+  VP8_TUNE_SSIM
+} vp8e_tuning;
+
+/*!\brief  vp9 svc layer parameters
+ *
+ * This defines the spatial and temporal layer id numbers for svc encoding.
+ * This is used with the #VP9E_SET_SVC_LAYER_ID control to set the spatial and
+ * temporal layer id for the current frame.
+ *
+ */
+typedef struct vpx_svc_layer_id {
+  int spatial_layer_id;       /**< Spatial layer id number. */
+  int temporal_layer_id;      /**< Temporal layer id number. */
+} vpx_svc_layer_id_t;
+
+/*!\brief  vp9 svc frame flag parameters.
+ *
+ * This defines the frame flags and buffer indices for each spatial layer for
+ * svc encoding.
+ * This is used with the #VP9E_SET_SVC_REF_FRAME_CONFIG control to set frame
+ * flags and buffer indices for each spatial layer for the current (super)frame.
+ *
+ */
+typedef struct vpx_svc_ref_frame_config {
+  int frame_flags[VPX_TS_MAX_LAYERS];  /**< Frame flags. */
+  int lst_fb_idx[VPX_TS_MAX_LAYERS];  /**< Last buffer index. */
+  int gld_fb_idx[VPX_TS_MAX_LAYERS];  /**< Golden buffer index. */
+  int alt_fb_idx[VPX_TS_MAX_LAYERS];  /**< Altref buffer index. */
+} vpx_svc_ref_frame_config_t;
+
+/*!\cond */
+/*!\brief VP8 encoder control function parameter type
+ *
+ * Defines the data types that VP8E control functions take. Note that
+ * additional common controls are defined in vp8.h
+ *
+ */
+
+VPX_CTRL_USE_TYPE(VP8E_SET_FRAME_FLAGS,        int)
+#define VPX_CTRL_VP8E_SET_FRAME_FLAGS
+VPX_CTRL_USE_TYPE(VP8E_SET_TEMPORAL_LAYER_ID,  int)
+#define VPX_CTRL_VP8E_SET_TEMPORAL_LAYER_ID
+VPX_CTRL_USE_TYPE(VP8E_SET_ROI_MAP,            vpx_roi_map_t *)
+#define VPX_CTRL_VP8E_SET_ROI_MAP
+VPX_CTRL_USE_TYPE(VP8E_SET_ACTIVEMAP,          vpx_active_map_t *)
+#define VPX_CTRL_VP8E_SET_ACTIVEMAP
+VPX_CTRL_USE_TYPE(VP8E_SET_SCALEMODE,          vpx_scaling_mode_t *)
+#define VPX_CTRL_VP8E_SET_SCALEMODE
+
+VPX_CTRL_USE_TYPE(VP9E_SET_SVC,                int)
+#define VPX_CTRL_VP9E_SET_SVC
+VPX_CTRL_USE_TYPE(VP9E_SET_SVC_PARAMETERS,     void *)
+#define VPX_CTRL_VP9E_SET_SVC_PARAMETERS
+VPX_CTRL_USE_TYPE(VP9E_REGISTER_CX_CALLBACK,   void *)
+#define VPX_CTRL_VP9E_REGISTER_CX_CALLBACK
+VPX_CTRL_USE_TYPE(VP9E_SET_SVC_LAYER_ID,       vpx_svc_layer_id_t *)
+#define VPX_CTRL_VP9E_SET_SVC_LAYER_ID
+
+VPX_CTRL_USE_TYPE(VP8E_SET_CPUUSED,            int)
+#define VPX_CTRL_VP8E_SET_CPUUSED
+VPX_CTRL_USE_TYPE(VP8E_SET_ENABLEAUTOALTREF,   unsigned int)
+#define VPX_CTRL_VP8E_SET_ENABLEAUTOALTREF
+VPX_CTRL_USE_TYPE(VP8E_SET_NOISE_SENSITIVITY,  unsigned int)
+#define VPX_CTRL_VP8E_SET_NOISE_SENSITIVITY
+VPX_CTRL_USE_TYPE(VP8E_SET_SHARPNESS,          unsigned int)
+#define VPX_CTRL_VP8E_SET_SHARPNESS
+VPX_CTRL_USE_TYPE(VP8E_SET_STATIC_THRESHOLD,   unsigned int)
+#define VPX_CTRL_VP8E_SET_STATIC_THRESHOLD
+VPX_CTRL_USE_TYPE(VP8E_SET_TOKEN_PARTITIONS,   int) /* vp8e_token_partitions */
+#define VPX_CTRL_VP8E_SET_TOKEN_PARTITIONS
+
+VPX_CTRL_USE_TYPE(VP8E_SET_ARNR_MAXFRAMES,     unsigned int)
+#define VPX_CTRL_VP8E_SET_ARNR_MAXFRAMES
+VPX_CTRL_USE_TYPE(VP8E_SET_ARNR_STRENGTH,     unsigned int)
+#define VPX_CTRL_VP8E_SET_ARNR_STRENGTH
+VPX_CTRL_USE_TYPE_DEPRECATED(VP8E_SET_ARNR_TYPE,     unsigned int)
+#define VPX_CTRL_VP8E_SET_ARNR_TYPE
+VPX_CTRL_USE_TYPE(VP8E_SET_TUNING,             int) /* vp8e_tuning */
+#define VPX_CTRL_VP8E_SET_TUNING
+VPX_CTRL_USE_TYPE(VP8E_SET_CQ_LEVEL,      unsigned int)
+#define VPX_CTRL_VP8E_SET_CQ_LEVEL
+
+VPX_CTRL_USE_TYPE(VP9E_SET_TILE_COLUMNS,  int)
+#define VPX_CTRL_VP9E_SET_TILE_COLUMNS
+VPX_CTRL_USE_TYPE(VP9E_SET_TILE_ROWS,  int)
+#define VPX_CTRL_VP9E_SET_TILE_ROWS
+
+VPX_CTRL_USE_TYPE(VP8E_GET_LAST_QUANTIZER,     int *)
+#define VPX_CTRL_VP8E_GET_LAST_QUANTIZER
+VPX_CTRL_USE_TYPE(VP8E_GET_LAST_QUANTIZER_64,  int *)
+#define VPX_CTRL_VP8E_GET_LAST_QUANTIZER_64
+VPX_CTRL_USE_TYPE(VP9E_GET_SVC_LAYER_ID,  vpx_svc_layer_id_t *)
+#define VPX_CTRL_VP9E_GET_SVC_LAYER_ID
+
+VPX_CTRL_USE_TYPE(VP8E_SET_MAX_INTRA_BITRATE_PCT, unsigned int)
+#define VPX_CTRL_VP8E_SET_MAX_INTRA_BITRATE_PCT
+VPX_CTRL_USE_TYPE(VP8E_SET_MAX_INTER_BITRATE_PCT, unsigned int)
+#define VPX_CTRL_VP8E_SET_MAX_INTER_BITRATE_PCT
+
+VPX_CTRL_USE_TYPE(VP8E_SET_SCREEN_CONTENT_MODE, unsigned int)
+#define VPX_CTRL_VP8E_SET_SCREEN_CONTENT_MODE
+
+VPX_CTRL_USE_TYPE(VP9E_SET_GF_CBR_BOOST_PCT, unsigned int)
+#define VPX_CTRL_VP9E_SET_GF_CBR_BOOST_PCT
+
+VPX_CTRL_USE_TYPE(VP9E_SET_LOSSLESS, unsigned int)
+#define VPX_CTRL_VP9E_SET_LOSSLESS
+
+VPX_CTRL_USE_TYPE(VP9E_SET_FRAME_PARALLEL_DECODING, unsigned int)
+#define VPX_CTRL_VP9E_SET_FRAME_PARALLEL_DECODING
+
+VPX_CTRL_USE_TYPE(VP9E_SET_AQ_MODE, unsigned int)
+#define VPX_CTRL_VP9E_SET_AQ_MODE
+
+VPX_CTRL_USE_TYPE(VP9E_SET_FRAME_PERIODIC_BOOST, unsigned int)
+#define VPX_CTRL_VP9E_SET_FRAME_PERIODIC_BOOST
+
+VPX_CTRL_USE_TYPE(VP9E_SET_NOISE_SENSITIVITY,  unsigned int)
+#define VPX_CTRL_VP9E_SET_NOISE_SENSITIVITY
+
+VPX_CTRL_USE_TYPE(VP9E_SET_TUNE_CONTENT, int) /* vp9e_tune_content */
+#define VPX_CTRL_VP9E_SET_TUNE_CONTENT
+
+VPX_CTRL_USE_TYPE(VP9E_SET_COLOR_SPACE, int)
+#define VPX_CTRL_VP9E_SET_COLOR_SPACE
+
+VPX_CTRL_USE_TYPE(VP9E_SET_MIN_GF_INTERVAL,  unsigned int)
+#define VPX_CTRL_VP9E_SET_MIN_GF_INTERVAL
+
+VPX_CTRL_USE_TYPE(VP9E_SET_MAX_GF_INTERVAL,  unsigned int)
+#define VPX_CTRL_VP9E_SET_MAX_GF_INTERVAL
+
+VPX_CTRL_USE_TYPE(VP9E_GET_ACTIVEMAP, vpx_active_map_t *)
+#define VPX_CTRL_VP9E_GET_ACTIVEMAP
+
+VPX_CTRL_USE_TYPE(VP9E_SET_COLOR_RANGE, int)
+#define VPX_CTRL_VP9E_SET_COLOR_RANGE
+
+VPX_CTRL_USE_TYPE(VP9E_SET_SVC_REF_FRAME_CONFIG, vpx_svc_ref_frame_config_t *)
+#define VPX_CTRL_VP9E_SET_SVC_REF_FRAME_CONFIG
+
+VPX_CTRL_USE_TYPE(VP9E_SET_RENDER_SIZE, int *)
+#define VPX_CTRL_VP9E_SET_RENDER_SIZE
+
+VPX_CTRL_USE_TYPE(VP9E_SET_TARGET_LEVEL,  unsigned int)
+#define VPX_CTRL_VP9E_SET_TARGET_LEVEL
+
+VPX_CTRL_USE_TYPE(VP9E_GET_LEVEL, int *)
+#define VPX_CTRL_VP9E_GET_LEVEL
+
+/*!\endcond */
+/*! @} - end defgroup vp8_encoder */
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_VP8CX_H_
diff --git a/libvpx/libvpx/vpx/vp8dx.h b/libvpx/libvpx/vpx/vp8dx.h
new file mode 100644
index 0000000..67c97bb
--- /dev/null
+++ b/libvpx/libvpx/vpx/vp8dx.h
@@ -0,0 +1,176 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+/*!\defgroup vp8_decoder WebM VP8/VP9 Decoder
+ * \ingroup vp8
+ *
+ * @{
+ */
+/*!\file
+ * \brief Provides definitions for using VP8 or VP9 within the vpx Decoder
+ *        interface.
+ */
+#ifndef VPX_VP8DX_H_
+#define VPX_VP8DX_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Include controls common to both the encoder and decoder */
+#include "./vp8.h"
+
+/*!\name Algorithm interface for VP8
+ *
+ * This interface provides the capability to decode VP8 streams.
+ * @{
+ */
+extern vpx_codec_iface_t  vpx_codec_vp8_dx_algo;
+extern vpx_codec_iface_t *vpx_codec_vp8_dx(void);
+/*!@} - end algorithm interface member group*/
+
+/*!\name Algorithm interface for VP9
+ *
+ * This interface provides the capability to decode VP9 streams.
+ * @{
+ */
+extern vpx_codec_iface_t  vpx_codec_vp9_dx_algo;
+extern vpx_codec_iface_t *vpx_codec_vp9_dx(void);
+/*!@} - end algorithm interface member group*/
+
+/*!\enum vp8_dec_control_id
+ * \brief VP8 decoder control functions
+ *
+ * This set of macros define the control functions available for the VP8
+ * decoder interface.
+ *
+ * \sa #vpx_codec_control
+ */
+enum vp8_dec_control_id {
+  /** control function to get info on which reference frames were updated
+   *  by the last decode
+   */
+  VP8D_GET_LAST_REF_UPDATES = VP8_DECODER_CTRL_ID_START,
+
+  /** check if the indicated frame is corrupted */
+  VP8D_GET_FRAME_CORRUPTED,
+
+  /** control function to get info on which reference frames were used
+   *  by the last decode
+   */
+  VP8D_GET_LAST_REF_USED,
+
+  /** decryption function to decrypt encoded buffer data immediately
+   * before decoding. Takes a vpx_decrypt_init, which contains
+   * a callback function and opaque context pointer.
+   */
+  VPXD_SET_DECRYPTOR,
+  VP8D_SET_DECRYPTOR = VPXD_SET_DECRYPTOR,
+
+  /** control function to get the dimensions that the current frame is decoded
+   * at. This may be different to the intended display size for the frame as
+   * specified in the wrapper or frame header (see VP9D_GET_DISPLAY_SIZE). */
+  VP9D_GET_FRAME_SIZE,
+
+  /** control function to get the current frame's intended display dimensions
+   * (as specified in the wrapper or frame header). This may be different to
+   * the decoded dimensions of this frame (see VP9D_GET_FRAME_SIZE). */
+  VP9D_GET_DISPLAY_SIZE,
+
+  /** control function to get the bit depth of the stream. */
+  VP9D_GET_BIT_DEPTH,
+
+  /** control function to set the byte alignment of the planes in the reference
+   * buffers. Valid values are power of 2, from 32 to 1024. A value of 0 sets
+   * legacy alignment. I.e. Y plane is aligned to 32 bytes, U plane directly
+   * follows Y plane, and V plane directly follows U plane. Default value is 0.
+   */
+  VP9_SET_BYTE_ALIGNMENT,
+
+  /** control function to invert the decoding order to from right to left. The
+   * function is used in a test to confirm the decoding independence of tile
+   * columns. The function may be used in application where this order
+   * of decoding is desired.
+   *
+   * TODO(yaowu): Rework the unit test that uses this control, and in a future
+   *              release, this test-only control shall be removed.
+   */
+  VP9_INVERT_TILE_DECODE_ORDER,
+
+  /** control function to set the skip loop filter flag. Valid values are
+   * integers. The decoder will skip the loop filter when its value is set to
+   * nonzero. If the loop filter is skipped the decoder may accumulate decode
+   * artifacts. The default value is 0.
+   */
+  VP9_SET_SKIP_LOOP_FILTER,
+
+  VP8_DECODER_CTRL_ID_MAX
+};
+
+/** Decrypt n bytes of data from input -> output, using the decrypt_state
+ *  passed in VPXD_SET_DECRYPTOR.
+ */
+typedef void (*vpx_decrypt_cb)(void *decrypt_state, const unsigned char *input,
+                               unsigned char *output, int count);
+
+/*!\brief Structure to hold decryption state
+ *
+ * Defines a structure to hold the decryption state and access function.
+ */
+typedef struct vpx_decrypt_init {
+    /*! Decrypt callback. */
+    vpx_decrypt_cb decrypt_cb;
+
+    /*! Decryption state. */
+    void *decrypt_state;
+} vpx_decrypt_init;
+
+/*!\brief A deprecated alias for vpx_decrypt_init.
+ */
+typedef vpx_decrypt_init vp8_decrypt_init;
+
+
+/*!\cond */
+/*!\brief VP8 decoder control function parameter type
+ *
+ * Defines the data types that VP8D control functions take. Note that
+ * additional common controls are defined in vp8.h
+ *
+ */
+
+
+VPX_CTRL_USE_TYPE(VP8D_GET_LAST_REF_UPDATES,    int *)
+#define VPX_CTRL_VP8D_GET_LAST_REF_UPDATES
+VPX_CTRL_USE_TYPE(VP8D_GET_FRAME_CORRUPTED,     int *)
+#define VPX_CTRL_VP8D_GET_FRAME_CORRUPTED
+VPX_CTRL_USE_TYPE(VP8D_GET_LAST_REF_USED,       int *)
+#define VPX_CTRL_VP8D_GET_LAST_REF_USED
+VPX_CTRL_USE_TYPE(VPXD_SET_DECRYPTOR,           vpx_decrypt_init *)
+#define VPX_CTRL_VPXD_SET_DECRYPTOR
+VPX_CTRL_USE_TYPE(VP8D_SET_DECRYPTOR,           vpx_decrypt_init *)
+#define VPX_CTRL_VP8D_SET_DECRYPTOR
+VPX_CTRL_USE_TYPE(VP9D_GET_DISPLAY_SIZE,        int *)
+#define VPX_CTRL_VP9D_GET_DISPLAY_SIZE
+VPX_CTRL_USE_TYPE(VP9D_GET_BIT_DEPTH,           unsigned int *)
+#define VPX_CTRL_VP9D_GET_BIT_DEPTH
+VPX_CTRL_USE_TYPE(VP9D_GET_FRAME_SIZE,          int *)
+#define VPX_CTRL_VP9D_GET_FRAME_SIZE
+VPX_CTRL_USE_TYPE(VP9_INVERT_TILE_DECODE_ORDER, int)
+#define VPX_CTRL_VP9_INVERT_TILE_DECODE_ORDER
+
+/*!\endcond */
+/*! @} - end defgroup vp8_decoder */
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_VP8DX_H_
diff --git a/libvpx/libvpx/vpx/vpx_codec.h b/libvpx/libvpx/vpx/vpx_codec.h
new file mode 100644
index 0000000..b6037bb
--- /dev/null
+++ b/libvpx/libvpx/vpx/vpx_codec.h
@@ -0,0 +1,479 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+/*!\defgroup codec Common Algorithm Interface
+ * This abstraction allows applications to easily support multiple video
+ * formats with minimal code duplication. This section describes the interface
+ * common to all codecs (both encoders and decoders).
+ * @{
+ */
+
+/*!\file
+ * \brief Describes the codec algorithm interface to applications.
+ *
+ * This file describes the interface between an application and a
+ * video codec algorithm.
+ *
+ * An application instantiates a specific codec instance by using
+ * vpx_codec_init() and a pointer to the algorithm's interface structure:
+ *     <pre>
+ *     my_app.c:
+ *       extern vpx_codec_iface_t my_codec;
+ *       {
+ *           vpx_codec_ctx_t algo;
+ *           res = vpx_codec_init(&algo, &my_codec);
+ *       }
+ *     </pre>
+ *
+ * Once initialized, the instance is manged using other functions from
+ * the vpx_codec_* family.
+ */
+#ifndef VPX_VPX_CODEC_H_
+#define VPX_VPX_CODEC_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./vpx_integer.h"
+#include "./vpx_image.h"
+
+  /*!\brief Decorator indicating a function is deprecated */
+#ifndef DEPRECATED
+#if defined(__GNUC__) && __GNUC__
+#define DEPRECATED          __attribute__ ((deprecated))
+#elif defined(_MSC_VER)
+#define DEPRECATED
+#else
+#define DEPRECATED
+#endif
+#endif  /* DEPRECATED */
+
+#ifndef DECLSPEC_DEPRECATED
+#if defined(__GNUC__) && __GNUC__
+#define DECLSPEC_DEPRECATED /**< \copydoc #DEPRECATED */
+#elif defined(_MSC_VER)
+#define DECLSPEC_DEPRECATED __declspec(deprecated) /**< \copydoc #DEPRECATED */
+#else
+#define DECLSPEC_DEPRECATED /**< \copydoc #DEPRECATED */
+#endif
+#endif  /* DECLSPEC_DEPRECATED */
+
+  /*!\brief Decorator indicating a function is potentially unused */
+#ifdef UNUSED
+#elif defined(__GNUC__) || defined(__clang__)
+#define UNUSED __attribute__ ((unused))
+#else
+#define UNUSED
+#endif
+
+  /*!\brief Current ABI version number
+   *
+   * \internal
+   * If this file is altered in any way that changes the ABI, this value
+   * must be bumped.  Examples include, but are not limited to, changing
+   * types, removing or reassigning enums, adding/removing/rearranging
+   * fields to structures
+   */
+#define VPX_CODEC_ABI_VERSION (3 + VPX_IMAGE_ABI_VERSION) /**<\hideinitializer*/
+
+  /*!\brief Algorithm return codes */
+  typedef enum {
+    /*!\brief Operation completed without error */
+    VPX_CODEC_OK,
+
+    /*!\brief Unspecified error */
+    VPX_CODEC_ERROR,
+
+    /*!\brief Memory operation failed */
+    VPX_CODEC_MEM_ERROR,
+
+    /*!\brief ABI version mismatch */
+    VPX_CODEC_ABI_MISMATCH,
+
+    /*!\brief Algorithm does not have required capability */
+    VPX_CODEC_INCAPABLE,
+
+    /*!\brief The given bitstream is not supported.
+     *
+     * The bitstream was unable to be parsed at the highest level. The decoder
+     * is unable to proceed. This error \ref SHOULD be treated as fatal to the
+     * stream. */
+    VPX_CODEC_UNSUP_BITSTREAM,
+
+    /*!\brief Encoded bitstream uses an unsupported feature
+     *
+     * The decoder does not implement a feature required by the encoder. This
+     * return code should only be used for features that prevent future
+     * pictures from being properly decoded. This error \ref MAY be treated as
+     * fatal to the stream or \ref MAY be treated as fatal to the current GOP.
+     */
+    VPX_CODEC_UNSUP_FEATURE,
+
+    /*!\brief The coded data for this stream is corrupt or incomplete
+     *
+     * There was a problem decoding the current frame.  This return code
+     * should only be used for failures that prevent future pictures from
+     * being properly decoded. This error \ref MAY be treated as fatal to the
+     * stream or \ref MAY be treated as fatal to the current GOP. If decoding
+     * is continued for the current GOP, artifacts may be present.
+     */
+    VPX_CODEC_CORRUPT_FRAME,
+
+    /*!\brief An application-supplied parameter is not valid.
+     *
+     */
+    VPX_CODEC_INVALID_PARAM,
+
+    /*!\brief An iterator reached the end of list.
+     *
+     */
+    VPX_CODEC_LIST_END
+
+  }
+  vpx_codec_err_t;
+
+
+  /*! \brief Codec capabilities bitfield
+   *
+   *  Each codec advertises the capabilities it supports as part of its
+   *  ::vpx_codec_iface_t interface structure. Capabilities are extra interfaces
+   *  or functionality, and are not required to be supported.
+   *
+   *  The available flags are specified by VPX_CODEC_CAP_* defines.
+   */
+  typedef long vpx_codec_caps_t;
+#define VPX_CODEC_CAP_DECODER 0x1 /**< Is a decoder */
+#define VPX_CODEC_CAP_ENCODER 0x2 /**< Is an encoder */
+
+
+  /*! \brief Initialization-time Feature Enabling
+   *
+   *  Certain codec features must be known at initialization time, to allow for
+   *  proper memory allocation.
+   *
+   *  The available flags are specified by VPX_CODEC_USE_* defines.
+   */
+  typedef long vpx_codec_flags_t;
+
+
+  /*!\brief Codec interface structure.
+   *
+   * Contains function pointers and other data private to the codec
+   * implementation. This structure is opaque to the application.
+   */
+  typedef const struct vpx_codec_iface vpx_codec_iface_t;
+
+
+  /*!\brief Codec private data structure.
+   *
+   * Contains data private to the codec implementation. This structure is opaque
+   * to the application.
+   */
+  typedef       struct vpx_codec_priv  vpx_codec_priv_t;
+
+
+  /*!\brief Iterator
+   *
+   * Opaque storage used for iterating over lists.
+   */
+  typedef const void *vpx_codec_iter_t;
+
+
+  /*!\brief Codec context structure
+   *
+   * All codecs \ref MUST support this context structure fully. In general,
+   * this data should be considered private to the codec algorithm, and
+   * not be manipulated or examined by the calling application. Applications
+   * may reference the 'name' member to get a printable description of the
+   * algorithm.
+   */
+  typedef struct vpx_codec_ctx {
+    const char              *name;        /**< Printable interface name */
+    vpx_codec_iface_t       *iface;       /**< Interface pointers */
+    vpx_codec_err_t          err;         /**< Last returned error */
+    const char              *err_detail;  /**< Detailed info, if available */
+    vpx_codec_flags_t        init_flags;  /**< Flags passed at init time */
+    union {
+      /**< Decoder Configuration Pointer */
+      const struct vpx_codec_dec_cfg *dec;
+      /**< Encoder Configuration Pointer */
+      const struct vpx_codec_enc_cfg *enc;
+      const void                     *raw;
+    }                        config;      /**< Configuration pointer aliasing union */
+    vpx_codec_priv_t        *priv;        /**< Algorithm private storage */
+  } vpx_codec_ctx_t;
+
+  /*!\brief Bit depth for codec
+   * *
+   * This enumeration determines the bit depth of the codec.
+   */
+  typedef enum vpx_bit_depth {
+    VPX_BITS_8  =  8,  /**<  8 bits */
+    VPX_BITS_10 = 10,  /**< 10 bits */
+    VPX_BITS_12 = 12,  /**< 12 bits */
+  } vpx_bit_depth_t;
+
+  /*
+   * Library Version Number Interface
+   *
+   * For example, see the following sample return values:
+   *     vpx_codec_version()           (1<<16 | 2<<8 | 3)
+   *     vpx_codec_version_str()       "v1.2.3-rc1-16-gec6a1ba"
+   *     vpx_codec_version_extra_str() "rc1-16-gec6a1ba"
+   */
+
+  /*!\brief Return the version information (as an integer)
+   *
+   * Returns a packed encoding of the library version number. This will only include
+   * the major.minor.patch component of the version number. Note that this encoded
+   * value should be accessed through the macros provided, as the encoding may change
+   * in the future.
+   *
+   */
+  int vpx_codec_version(void);
+#define VPX_VERSION_MAJOR(v) ((v>>16)&0xff) /**< extract major from packed version */
+#define VPX_VERSION_MINOR(v) ((v>>8)&0xff)  /**< extract minor from packed version */
+#define VPX_VERSION_PATCH(v) ((v>>0)&0xff)  /**< extract patch from packed version */
+
+  /*!\brief Return the version major number */
+#define vpx_codec_version_major() ((vpx_codec_version()>>16)&0xff)
+
+  /*!\brief Return the version minor number */
+#define vpx_codec_version_minor() ((vpx_codec_version()>>8)&0xff)
+
+  /*!\brief Return the version patch number */
+#define vpx_codec_version_patch() ((vpx_codec_version()>>0)&0xff)
+
+
+  /*!\brief Return the version information (as a string)
+   *
+   * Returns a printable string containing the full library version number. This may
+   * contain additional text following the three digit version number, as to indicate
+   * release candidates, prerelease versions, etc.
+   *
+   */
+  const char *vpx_codec_version_str(void);
+
+
+  /*!\brief Return the version information (as a string)
+   *
+   * Returns a printable "extra string". This is the component of the string returned
+   * by vpx_codec_version_str() following the three digit version number.
+   *
+   */
+  const char *vpx_codec_version_extra_str(void);
+
+
+  /*!\brief Return the build configuration
+   *
+   * Returns a printable string containing an encoded version of the build
+   * configuration. This may be useful to vpx support.
+   *
+   */
+  const char *vpx_codec_build_config(void);
+
+
+  /*!\brief Return the name for a given interface
+   *
+   * Returns a human readable string for name of the given codec interface.
+   *
+   * \param[in]    iface     Interface pointer
+   *
+   */
+  const char *vpx_codec_iface_name(vpx_codec_iface_t *iface);
+
+
+  /*!\brief Convert error number to printable string
+   *
+   * Returns a human readable string for the last error returned by the
+   * algorithm. The returned error will be one line and will not contain
+   * any newline characters.
+   *
+   *
+   * \param[in]    err     Error number.
+   *
+   */
+  const char *vpx_codec_err_to_string(vpx_codec_err_t  err);
+
+
+  /*!\brief Retrieve error synopsis for codec context
+   *
+   * Returns a human readable string for the last error returned by the
+   * algorithm. The returned error will be one line and will not contain
+   * any newline characters.
+   *
+   *
+   * \param[in]    ctx     Pointer to this instance's context.
+   *
+   */
+  const char *vpx_codec_error(vpx_codec_ctx_t  *ctx);
+
+
+  /*!\brief Retrieve detailed error information for codec context
+   *
+   * Returns a human readable string providing detailed information about
+   * the last error.
+   *
+   * \param[in]    ctx     Pointer to this instance's context.
+   *
+   * \retval NULL
+   *     No detailed information is available.
+   */
+  const char *vpx_codec_error_detail(vpx_codec_ctx_t  *ctx);
+
+
+  /* REQUIRED FUNCTIONS
+   *
+   * The following functions are required to be implemented for all codecs.
+   * They represent the base case functionality expected of all codecs.
+   */
+
+  /*!\brief Destroy a codec instance
+   *
+   * Destroys a codec context, freeing any associated memory buffers.
+   *
+   * \param[in] ctx   Pointer to this instance's context
+   *
+   * \retval #VPX_CODEC_OK
+   *     The codec algorithm initialized.
+   * \retval #VPX_CODEC_MEM_ERROR
+   *     Memory allocation failed.
+   */
+  vpx_codec_err_t vpx_codec_destroy(vpx_codec_ctx_t *ctx);
+
+
+  /*!\brief Get the capabilities of an algorithm.
+   *
+   * Retrieves the capabilities bitfield from the algorithm's interface.
+   *
+   * \param[in] iface   Pointer to the algorithm interface
+   *
+   */
+  vpx_codec_caps_t vpx_codec_get_caps(vpx_codec_iface_t *iface);
+
+
+  /*!\brief Control algorithm
+   *
+   * This function is used to exchange algorithm specific data with the codec
+   * instance. This can be used to implement features specific to a particular
+   * algorithm.
+   *
+   * This wrapper function dispatches the request to the helper function
+   * associated with the given ctrl_id. It tries to call this function
+   * transparently, but will return #VPX_CODEC_ERROR if the request could not
+   * be dispatched.
+   *
+   * Note that this function should not be used directly. Call the
+   * #vpx_codec_control wrapper macro instead.
+   *
+   * \param[in]     ctx              Pointer to this instance's context
+   * \param[in]     ctrl_id          Algorithm specific control identifier
+   *
+   * \retval #VPX_CODEC_OK
+   *     The control request was processed.
+   * \retval #VPX_CODEC_ERROR
+   *     The control request was not processed.
+   * \retval #VPX_CODEC_INVALID_PARAM
+   *     The data was not valid.
+   */
+  vpx_codec_err_t vpx_codec_control_(vpx_codec_ctx_t  *ctx,
+                                     int               ctrl_id,
+                                     ...);
+#if defined(VPX_DISABLE_CTRL_TYPECHECKS) && VPX_DISABLE_CTRL_TYPECHECKS
+#    define vpx_codec_control(ctx,id,data) vpx_codec_control_(ctx,id,data)
+#    define VPX_CTRL_USE_TYPE(id, typ)
+#    define VPX_CTRL_USE_TYPE_DEPRECATED(id, typ)
+#    define VPX_CTRL_VOID(id, typ)
+
+#else
+  /*!\brief vpx_codec_control wrapper macro
+   *
+   * This macro allows for type safe conversions across the variadic parameter
+   * to vpx_codec_control_().
+   *
+   * \internal
+   * It works by dispatching the call to the control function through a wrapper
+   * function named with the id parameter.
+   */
+#    define vpx_codec_control(ctx,id,data) vpx_codec_control_##id(ctx,id,data)\
+  /**<\hideinitializer*/
+
+
+  /*!\brief vpx_codec_control type definition macro
+   *
+   * This macro allows for type safe conversions across the variadic parameter
+   * to vpx_codec_control_(). It defines the type of the argument for a given
+   * control identifier.
+   *
+   * \internal
+   * It defines a static function with
+   * the correctly typed arguments as a wrapper to the type-unsafe internal
+   * function.
+   */
+#    define VPX_CTRL_USE_TYPE(id, typ) \
+  static vpx_codec_err_t \
+  vpx_codec_control_##id(vpx_codec_ctx_t*, int, typ) UNUSED;\
+  \
+  static vpx_codec_err_t \
+  vpx_codec_control_##id(vpx_codec_ctx_t  *ctx, int ctrl_id, typ data) {\
+    return vpx_codec_control_(ctx, ctrl_id, data);\
+  } /**<\hideinitializer*/
+
+
+  /*!\brief vpx_codec_control deprecated type definition macro
+   *
+   * Like #VPX_CTRL_USE_TYPE, but indicates that the specified control is
+   * deprecated and should not be used. Consult the documentation for your
+   * codec for more information.
+   *
+   * \internal
+   * It defines a static function with the correctly typed arguments as a
+   * wrapper to the type-unsafe internal function.
+   */
+#    define VPX_CTRL_USE_TYPE_DEPRECATED(id, typ) \
+  DECLSPEC_DEPRECATED static vpx_codec_err_t \
+  vpx_codec_control_##id(vpx_codec_ctx_t*, int, typ) DEPRECATED UNUSED;\
+  \
+  DECLSPEC_DEPRECATED static vpx_codec_err_t \
+  vpx_codec_control_##id(vpx_codec_ctx_t  *ctx, int ctrl_id, typ data) {\
+    return vpx_codec_control_(ctx, ctrl_id, data);\
+  } /**<\hideinitializer*/
+
+
+  /*!\brief vpx_codec_control void type definition macro
+   *
+   * This macro allows for type safe conversions across the variadic parameter
+   * to vpx_codec_control_(). It indicates that a given control identifier takes
+   * no argument.
+   *
+   * \internal
+   * It defines a static function without a data argument as a wrapper to the
+   * type-unsafe internal function.
+   */
+#    define VPX_CTRL_VOID(id) \
+  static vpx_codec_err_t \
+  vpx_codec_control_##id(vpx_codec_ctx_t*, int) UNUSED;\
+  \
+  static vpx_codec_err_t \
+  vpx_codec_control_##id(vpx_codec_ctx_t  *ctx, int ctrl_id) {\
+    return vpx_codec_control_(ctx, ctrl_id);\
+  } /**<\hideinitializer*/
+
+
+#endif
+
+  /*!@} - end defgroup codec*/
+#ifdef __cplusplus
+}
+#endif
+#endif  // VPX_VPX_CODEC_H_
+
diff --git a/libvpx/libvpx/vpx/vpx_codec.mk b/libvpx/libvpx/vpx/vpx_codec.mk
new file mode 100644
index 0000000..ccdef04
--- /dev/null
+++ b/libvpx/libvpx/vpx/vpx_codec.mk
@@ -0,0 +1,47 @@
+##
+##  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+API_EXPORTS += exports
+
+API_SRCS-$(CONFIG_VP8_ENCODER) += vp8.h
+API_SRCS-$(CONFIG_VP8_ENCODER) += vp8cx.h
+API_DOC_SRCS-$(CONFIG_VP8_ENCODER) += vp8.h
+API_DOC_SRCS-$(CONFIG_VP8_ENCODER) += vp8cx.h
+ifeq ($(CONFIG_VP9_ENCODER),yes)
+  API_SRCS-$(CONFIG_SPATIAL_SVC) += src/svc_encodeframe.c
+  API_SRCS-$(CONFIG_SPATIAL_SVC) += svc_context.h
+endif
+
+API_SRCS-$(CONFIG_VP8_DECODER) += vp8.h
+API_SRCS-$(CONFIG_VP8_DECODER) += vp8dx.h
+API_DOC_SRCS-$(CONFIG_VP8_DECODER) += vp8.h
+API_DOC_SRCS-$(CONFIG_VP8_DECODER) += vp8dx.h
+
+API_DOC_SRCS-yes += vpx_codec.h
+API_DOC_SRCS-yes += vpx_decoder.h
+API_DOC_SRCS-yes += vpx_encoder.h
+API_DOC_SRCS-yes += vpx_frame_buffer.h
+API_DOC_SRCS-yes += vpx_image.h
+
+API_SRCS-yes += src/vpx_decoder.c
+API_SRCS-yes += vpx_decoder.h
+API_SRCS-yes += src/vpx_encoder.c
+API_SRCS-yes += vpx_encoder.h
+API_SRCS-yes += internal/vpx_codec_internal.h
+API_SRCS-yes += internal/vpx_psnr.h
+API_SRCS-yes += src/vpx_codec.c
+API_SRCS-yes += src/vpx_image.c
+API_SRCS-yes += src/vpx_psnr.c
+API_SRCS-yes += vpx_codec.h
+API_SRCS-yes += vpx_codec.mk
+API_SRCS-yes += vpx_frame_buffer.h
+API_SRCS-yes += vpx_image.h
+API_SRCS-yes += vpx_integer.h
diff --git a/libvpx/libvpx/vpx/vpx_decoder.h b/libvpx/libvpx/vpx/vpx_decoder.h
new file mode 100644
index 0000000..62fd919
--- /dev/null
+++ b/libvpx/libvpx/vpx/vpx_decoder.h
@@ -0,0 +1,378 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VPX_VPX_DECODER_H_
+#define VPX_VPX_DECODER_H_
+
+/*!\defgroup decoder Decoder Algorithm Interface
+ * \ingroup codec
+ * This abstraction allows applications using this decoder to easily support
+ * multiple video formats with minimal code duplication. This section describes
+ * the interface common to all decoders.
+ * @{
+ */
+
+/*!\file
+ * \brief Describes the decoder algorithm interface to applications.
+ *
+ * This file describes the interface between an application and a
+ * video decoder algorithm.
+ *
+ */
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./vpx_codec.h"
+#include "./vpx_frame_buffer.h"
+
+  /*!\brief Current ABI version number
+   *
+   * \internal
+   * If this file is altered in any way that changes the ABI, this value
+   * must be bumped.  Examples include, but are not limited to, changing
+   * types, removing or reassigning enums, adding/removing/rearranging
+   * fields to structures
+   */
+#define VPX_DECODER_ABI_VERSION (3 + VPX_CODEC_ABI_VERSION) /**<\hideinitializer*/
+
+  /*! \brief Decoder capabilities bitfield
+   *
+   *  Each decoder advertises the capabilities it supports as part of its
+   *  ::vpx_codec_iface_t interface structure. Capabilities are extra interfaces
+   *  or functionality, and are not required to be supported by a decoder.
+   *
+   *  The available flags are specified by VPX_CODEC_CAP_* defines.
+   */
+#define VPX_CODEC_CAP_PUT_SLICE  0x10000 /**< Will issue put_slice callbacks */
+#define VPX_CODEC_CAP_PUT_FRAME  0x20000 /**< Will issue put_frame callbacks */
+#define VPX_CODEC_CAP_POSTPROC   0x40000 /**< Can postprocess decoded frame */
+#define VPX_CODEC_CAP_ERROR_CONCEALMENT   0x80000 /**< Can conceal errors due to
+  packet loss */
+#define VPX_CODEC_CAP_INPUT_FRAGMENTS   0x100000 /**< Can receive encoded frames
+  one fragment at a time */
+
+  /*! \brief Initialization-time Feature Enabling
+   *
+   *  Certain codec features must be known at initialization time, to allow for
+   *  proper memory allocation.
+   *
+   *  The available flags are specified by VPX_CODEC_USE_* defines.
+   */
+#define VPX_CODEC_CAP_FRAME_THREADING   0x200000 /**< Can support frame-based
+                                                      multi-threading */
+#define VPX_CODEC_CAP_EXTERNAL_FRAME_BUFFER 0x400000 /**< Can support external
+                                                          frame buffers */
+
+#define VPX_CODEC_USE_POSTPROC   0x10000 /**< Postprocess decoded frame */
+#define VPX_CODEC_USE_ERROR_CONCEALMENT 0x20000 /**< Conceal errors in decoded
+  frames */
+#define VPX_CODEC_USE_INPUT_FRAGMENTS   0x40000 /**< The input frame should be
+  passed to the decoder one
+  fragment at a time */
+#define VPX_CODEC_USE_FRAME_THREADING   0x80000 /**< Enable frame-based
+                                                     multi-threading */
+
+  /*!\brief Stream properties
+   *
+   * This structure is used to query or set properties of the decoded
+   * stream. Algorithms may extend this structure with data specific
+   * to their bitstream by setting the sz member appropriately.
+   */
+  typedef struct vpx_codec_stream_info {
+    unsigned int sz;     /**< Size of this structure */
+    unsigned int w;      /**< Width (or 0 for unknown/default) */
+    unsigned int h;      /**< Height (or 0 for unknown/default) */
+    unsigned int is_kf;  /**< Current frame is a keyframe */
+  } vpx_codec_stream_info_t;
+
+  /* REQUIRED FUNCTIONS
+   *
+   * The following functions are required to be implemented for all decoders.
+   * They represent the base case functionality expected of all decoders.
+   */
+
+
+  /*!\brief Initialization Configurations
+   *
+   * This structure is used to pass init time configuration options to the
+   * decoder.
+   */
+  typedef struct vpx_codec_dec_cfg {
+    unsigned int threads; /**< Maximum number of threads to use, default 1 */
+    unsigned int w;      /**< Width */
+    unsigned int h;      /**< Height */
+  } vpx_codec_dec_cfg_t; /**< alias for struct vpx_codec_dec_cfg */
+
+
+  /*!\brief Initialize a decoder instance
+   *
+   * Initializes a decoder context using the given interface. Applications
+   * should call the vpx_codec_dec_init convenience macro instead of this
+   * function directly, to ensure that the ABI version number parameter
+   * is properly initialized.
+   *
+   * If the library was configured with --disable-multithread, this call
+   * is not thread safe and should be guarded with a lock if being used
+   * in a multithreaded context.
+   *
+   * \param[in]    ctx     Pointer to this instance's context.
+   * \param[in]    iface   Pointer to the algorithm interface to use.
+   * \param[in]    cfg     Configuration to use, if known. May be NULL.
+   * \param[in]    flags   Bitfield of VPX_CODEC_USE_* flags
+   * \param[in]    ver     ABI version number. Must be set to
+   *                       VPX_DECODER_ABI_VERSION
+   * \retval #VPX_CODEC_OK
+   *     The decoder algorithm initialized.
+   * \retval #VPX_CODEC_MEM_ERROR
+   *     Memory allocation failed.
+   */
+  vpx_codec_err_t vpx_codec_dec_init_ver(vpx_codec_ctx_t      *ctx,
+                                         vpx_codec_iface_t    *iface,
+                                         const vpx_codec_dec_cfg_t *cfg,
+                                         vpx_codec_flags_t     flags,
+                                         int                   ver);
+
+  /*!\brief Convenience macro for vpx_codec_dec_init_ver()
+   *
+   * Ensures the ABI version parameter is properly set.
+   */
+#define vpx_codec_dec_init(ctx, iface, cfg, flags) \
+  vpx_codec_dec_init_ver(ctx, iface, cfg, flags, VPX_DECODER_ABI_VERSION)
+
+
+  /*!\brief Parse stream info from a buffer
+   *
+   * Performs high level parsing of the bitstream. Construction of a decoder
+   * context is not necessary. Can be used to determine if the bitstream is
+   * of the proper format, and to extract information from the stream.
+   *
+   * \param[in]      iface   Pointer to the algorithm interface
+   * \param[in]      data    Pointer to a block of data to parse
+   * \param[in]      data_sz Size of the data buffer
+   * \param[in,out]  si      Pointer to stream info to update. The size member
+   *                         \ref MUST be properly initialized, but \ref MAY be
+   *                         clobbered by the algorithm. This parameter \ref MAY
+   *                         be NULL.
+   *
+   * \retval #VPX_CODEC_OK
+   *     Bitstream is parsable and stream information updated
+   */
+  vpx_codec_err_t vpx_codec_peek_stream_info(vpx_codec_iface_t       *iface,
+                                             const uint8_t           *data,
+                                             unsigned int             data_sz,
+                                             vpx_codec_stream_info_t *si);
+
+
+  /*!\brief Return information about the current stream.
+   *
+   * Returns information about the stream that has been parsed during decoding.
+   *
+   * \param[in]      ctx     Pointer to this instance's context
+   * \param[in,out]  si      Pointer to stream info to update. The size member
+   *                         \ref MUST be properly initialized, but \ref MAY be
+   *                         clobbered by the algorithm. This parameter \ref MAY
+   *                         be NULL.
+   *
+   * \retval #VPX_CODEC_OK
+   *     Bitstream is parsable and stream information updated
+   */
+  vpx_codec_err_t vpx_codec_get_stream_info(vpx_codec_ctx_t         *ctx,
+                                            vpx_codec_stream_info_t *si);
+
+
+  /*!\brief Decode data
+   *
+   * Processes a buffer of coded data. If the processing results in a new
+   * decoded frame becoming available, PUT_SLICE and PUT_FRAME events may be
+   * generated, as appropriate. Encoded data \ref MUST be passed in DTS (decode
+   * time stamp) order. Frames produced will always be in PTS (presentation
+   * time stamp) order.
+   * If the decoder is configured with VPX_CODEC_USE_INPUT_FRAGMENTS enabled,
+   * data and data_sz can contain a fragment of the encoded frame. Fragment
+   * \#n must contain at least partition \#n, but can also contain subsequent
+   * partitions (\#n+1 - \#n+i), and if so, fragments \#n+1, .., \#n+i must
+   * be empty. When no more data is available, this function should be called
+   * with NULL as data and 0 as data_sz. The memory passed to this function
+   * must be available until the frame has been decoded.
+   *
+   * \param[in] ctx          Pointer to this instance's context
+   * \param[in] data         Pointer to this block of new coded data. If
+   *                         NULL, a VPX_CODEC_CB_PUT_FRAME event is posted
+   *                         for the previously decoded frame.
+   * \param[in] data_sz      Size of the coded data, in bytes.
+   * \param[in] user_priv    Application specific data to associate with
+   *                         this frame.
+   * \param[in] deadline     Soft deadline the decoder should attempt to meet,
+   *                         in us. Set to zero for unlimited.
+   *
+   * \return Returns #VPX_CODEC_OK if the coded data was processed completely
+   *         and future pictures can be decoded without error. Otherwise,
+   *         see the descriptions of the other error codes in ::vpx_codec_err_t
+   *         for recoverability capabilities.
+   */
+  vpx_codec_err_t vpx_codec_decode(vpx_codec_ctx_t    *ctx,
+                                   const uint8_t        *data,
+                                   unsigned int            data_sz,
+                                   void               *user_priv,
+                                   long                deadline);
+
+
+  /*!\brief Decoded frames iterator
+   *
+   * Iterates over a list of the frames available for display. The iterator
+   * storage should be initialized to NULL to start the iteration. Iteration is
+   * complete when this function returns NULL.
+   *
+   * The list of available frames becomes valid upon completion of the
+   * vpx_codec_decode call, and remains valid until the next call to vpx_codec_decode.
+   *
+   * \param[in]     ctx      Pointer to this instance's context
+   * \param[in,out] iter     Iterator storage, initialized to NULL
+   *
+   * \return Returns a pointer to an image, if one is ready for display. Frames
+   *         produced will always be in PTS (presentation time stamp) order.
+   */
+  vpx_image_t *vpx_codec_get_frame(vpx_codec_ctx_t  *ctx,
+                                   vpx_codec_iter_t *iter);
+
+
+  /*!\defgroup cap_put_frame Frame-Based Decoding Functions
+   *
+   * The following functions are required to be implemented for all decoders
+   * that advertise the VPX_CODEC_CAP_PUT_FRAME capability. Calling these functions
+   * for codecs that don't advertise this capability will result in an error
+   * code being returned, usually VPX_CODEC_ERROR
+   * @{
+   */
+
+  /*!\brief put frame callback prototype
+   *
+   * This callback is invoked by the decoder to notify the application of
+   * the availability of decoded image data.
+   */
+  typedef void (*vpx_codec_put_frame_cb_fn_t)(void        *user_priv,
+                                              const vpx_image_t *img);
+
+
+  /*!\brief Register for notification of frame completion.
+   *
+   * Registers a given function to be called when a decoded frame is
+   * available.
+   *
+   * \param[in] ctx          Pointer to this instance's context
+   * \param[in] cb           Pointer to the callback function
+   * \param[in] user_priv    User's private data
+   *
+   * \retval #VPX_CODEC_OK
+   *     Callback successfully registered.
+   * \retval #VPX_CODEC_ERROR
+   *     Decoder context not initialized, or algorithm not capable of
+   *     posting slice completion.
+   */
+  vpx_codec_err_t vpx_codec_register_put_frame_cb(vpx_codec_ctx_t             *ctx,
+                                                  vpx_codec_put_frame_cb_fn_t  cb,
+                                                  void                        *user_priv);
+
+
+  /*!@} - end defgroup cap_put_frame */
+
+  /*!\defgroup cap_put_slice Slice-Based Decoding Functions
+   *
+   * The following functions are required to be implemented for all decoders
+   * that advertise the VPX_CODEC_CAP_PUT_SLICE capability. Calling these functions
+   * for codecs that don't advertise this capability will result in an error
+   * code being returned, usually VPX_CODEC_ERROR
+   * @{
+   */
+
+  /*!\brief put slice callback prototype
+   *
+   * This callback is invoked by the decoder to notify the application of
+   * the availability of partially decoded image data. The
+   */
+  typedef void (*vpx_codec_put_slice_cb_fn_t)(void         *user_priv,
+                                              const vpx_image_t      *img,
+                                              const vpx_image_rect_t *valid,
+                                              const vpx_image_rect_t *update);
+
+
+  /*!\brief Register for notification of slice completion.
+   *
+   * Registers a given function to be called when a decoded slice is
+   * available.
+   *
+   * \param[in] ctx          Pointer to this instance's context
+   * \param[in] cb           Pointer to the callback function
+   * \param[in] user_priv    User's private data
+   *
+   * \retval #VPX_CODEC_OK
+   *     Callback successfully registered.
+   * \retval #VPX_CODEC_ERROR
+   *     Decoder context not initialized, or algorithm not capable of
+   *     posting slice completion.
+   */
+  vpx_codec_err_t vpx_codec_register_put_slice_cb(vpx_codec_ctx_t             *ctx,
+                                                  vpx_codec_put_slice_cb_fn_t  cb,
+                                                  void                        *user_priv);
+
+
+  /*!@} - end defgroup cap_put_slice*/
+
+  /*!\defgroup cap_external_frame_buffer External Frame Buffer Functions
+   *
+   * The following section is required to be implemented for all decoders
+   * that advertise the VPX_CODEC_CAP_EXTERNAL_FRAME_BUFFER capability.
+   * Calling this function for codecs that don't advertise this capability
+   * will result in an error code being returned, usually VPX_CODEC_ERROR.
+   *
+   * \note
+   * Currently this only works with VP9.
+   * @{
+   */
+
+  /*!\brief Pass in external frame buffers for the decoder to use.
+   *
+   * Registers functions to be called when libvpx needs a frame buffer
+   * to decode the current frame and a function to be called when libvpx does
+   * not internally reference the frame buffer. This set function must
+   * be called before the first call to decode or libvpx will assume the
+   * default behavior of allocating frame buffers internally.
+   *
+   * \param[in] ctx          Pointer to this instance's context
+   * \param[in] cb_get       Pointer to the get callback function
+   * \param[in] cb_release   Pointer to the release callback function
+   * \param[in] cb_priv      Callback's private data
+   *
+   * \retval #VPX_CODEC_OK
+   *     External frame buffers will be used by libvpx.
+   * \retval #VPX_CODEC_INVALID_PARAM
+   *     One or more of the callbacks were NULL.
+   * \retval #VPX_CODEC_ERROR
+   *     Decoder context not initialized, or algorithm not capable of
+   *     using external frame buffers.
+   *
+   * \note
+   * When decoding VP9, the application may be required to pass in at least
+   * #VP9_MAXIMUM_REF_BUFFERS + #VPX_MAXIMUM_WORK_BUFFERS external frame
+   * buffers.
+   */
+  vpx_codec_err_t vpx_codec_set_frame_buffer_functions(
+      vpx_codec_ctx_t *ctx,
+      vpx_get_frame_buffer_cb_fn_t cb_get,
+      vpx_release_frame_buffer_cb_fn_t cb_release, void *cb_priv);
+
+  /*!@} - end defgroup cap_external_frame_buffer */
+
+  /*!@} - end defgroup decoder*/
+#ifdef __cplusplus
+}
+#endif
+#endif  // VPX_VPX_DECODER_H_
+
diff --git a/libvpx/libvpx/vpx/vpx_encoder.h b/libvpx/libvpx/vpx/vpx_encoder.h
new file mode 100644
index 0000000..955e873
--- /dev/null
+++ b/libvpx/libvpx/vpx/vpx_encoder.h
@@ -0,0 +1,1043 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VPX_VPX_ENCODER_H_
+#define VPX_VPX_ENCODER_H_
+
+/*!\defgroup encoder Encoder Algorithm Interface
+ * \ingroup codec
+ * This abstraction allows applications using this encoder to easily support
+ * multiple video formats with minimal code duplication. This section describes
+ * the interface common to all encoders.
+ * @{
+ */
+
+/*!\file
+ * \brief Describes the encoder algorithm interface to applications.
+ *
+ * This file describes the interface between an application and a
+ * video encoder algorithm.
+ *
+ */
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./vpx_codec.h"
+
+  /*! Temporal Scalability: Maximum length of the sequence defining frame
+   * layer membership
+   */
+#define VPX_TS_MAX_PERIODICITY 16
+
+  /*! Temporal Scalability: Maximum number of coding layers */
+#define VPX_TS_MAX_LAYERS       5
+
+  /*!\deprecated Use #VPX_TS_MAX_PERIODICITY instead. */
+#define MAX_PERIODICITY VPX_TS_MAX_PERIODICITY
+
+/*! Temporal+Spatial Scalability: Maximum number of coding layers */
+#define VPX_MAX_LAYERS  12  // 3 temporal + 4 spatial layers are allowed.
+
+/*!\deprecated Use #VPX_MAX_LAYERS instead. */
+#define MAX_LAYERS    VPX_MAX_LAYERS  // 3 temporal + 4 spatial layers allowed.
+
+/*! Spatial Scalability: Maximum number of coding layers */
+#define VPX_SS_MAX_LAYERS       5
+
+/*! Spatial Scalability: Default number of coding layers */
+#define VPX_SS_DEFAULT_LAYERS       1
+
+  /*!\brief Current ABI version number
+   *
+   * \internal
+   * If this file is altered in any way that changes the ABI, this value
+   * must be bumped.  Examples include, but are not limited to, changing
+   * types, removing or reassigning enums, adding/removing/rearranging
+   * fields to structures
+   */
+#define VPX_ENCODER_ABI_VERSION (5 + VPX_CODEC_ABI_VERSION) /**<\hideinitializer*/
+
+
+  /*! \brief Encoder capabilities bitfield
+   *
+   *  Each encoder advertises the capabilities it supports as part of its
+   *  ::vpx_codec_iface_t interface structure. Capabilities are extra
+   *  interfaces or functionality, and are not required to be supported
+   *  by an encoder.
+   *
+   *  The available flags are specified by VPX_CODEC_CAP_* defines.
+   */
+#define VPX_CODEC_CAP_PSNR  0x10000 /**< Can issue PSNR packets */
+
+  /*! Can output one partition at a time. Each partition is returned in its
+   *  own VPX_CODEC_CX_FRAME_PKT, with the FRAME_IS_FRAGMENT flag set for
+   *  every partition but the last. In this mode all frames are always
+   *  returned partition by partition.
+   */
+#define VPX_CODEC_CAP_OUTPUT_PARTITION  0x20000
+
+/*! Can support input images at greater than 8 bitdepth.
+ */
+#define VPX_CODEC_CAP_HIGHBITDEPTH  0x40000
+
+  /*! \brief Initialization-time Feature Enabling
+   *
+   *  Certain codec features must be known at initialization time, to allow
+   *  for proper memory allocation.
+   *
+   *  The available flags are specified by VPX_CODEC_USE_* defines.
+   */
+#define VPX_CODEC_USE_PSNR  0x10000 /**< Calculate PSNR on each frame */
+#define VPX_CODEC_USE_OUTPUT_PARTITION  0x20000 /**< Make the encoder output one
+  partition at a time. */
+#define VPX_CODEC_USE_HIGHBITDEPTH 0x40000 /**< Use high bitdepth */
+
+
+  /*!\brief Generic fixed size buffer structure
+   *
+   * This structure is able to hold a reference to any fixed size buffer.
+   */
+  typedef struct vpx_fixed_buf {
+    void          *buf; /**< Pointer to the data */
+    size_t         sz;  /**< Length of the buffer, in chars */
+  } vpx_fixed_buf_t; /**< alias for struct vpx_fixed_buf */
+
+
+  /*!\brief Time Stamp Type
+   *
+   * An integer, which when multiplied by the stream's time base, provides
+   * the absolute time of a sample.
+   */
+  typedef int64_t vpx_codec_pts_t;
+
+
+  /*!\brief Compressed Frame Flags
+   *
+   * This type represents a bitfield containing information about a compressed
+   * frame that may be useful to an application. The most significant 16 bits
+   * can be used by an algorithm to provide additional detail, for example to
+   * support frame types that are codec specific (MPEG-1 D-frames for example)
+   */
+  typedef uint32_t vpx_codec_frame_flags_t;
+#define VPX_FRAME_IS_KEY       0x1 /**< frame is the start of a GOP */
+#define VPX_FRAME_IS_DROPPABLE 0x2 /**< frame can be dropped without affecting
+  the stream (no future frame depends on
+              this one) */
+#define VPX_FRAME_IS_INVISIBLE 0x4 /**< frame should be decoded but will not
+  be shown */
+#define VPX_FRAME_IS_FRAGMENT  0x8 /**< this is a fragment of the encoded
+  frame */
+
+  /*!\brief Error Resilient flags
+   *
+   * These flags define which error resilient features to enable in the
+   * encoder. The flags are specified through the
+   * vpx_codec_enc_cfg::g_error_resilient variable.
+   */
+  typedef uint32_t vpx_codec_er_flags_t;
+#define VPX_ERROR_RESILIENT_DEFAULT     0x1 /**< Improve resiliency against
+  losses of whole frames */
+#define VPX_ERROR_RESILIENT_PARTITIONS  0x2 /**< The frame partitions are
+  independently decodable by the
+  bool decoder, meaning that
+  partitions can be decoded even
+  though earlier partitions have
+  been lost. Note that intra
+  prediction is still done over
+  the partition boundary. */
+
+  /*!\brief Encoder output packet variants
+   *
+   * This enumeration lists the different kinds of data packets that can be
+   * returned by calls to vpx_codec_get_cx_data(). Algorithms \ref MAY
+   * extend this list to provide additional functionality.
+   */
+  enum vpx_codec_cx_pkt_kind {
+    VPX_CODEC_CX_FRAME_PKT,    /**< Compressed video frame */
+    VPX_CODEC_STATS_PKT,       /**< Two-pass statistics for this frame */
+    VPX_CODEC_FPMB_STATS_PKT,  /**< first pass mb statistics for this frame */
+    VPX_CODEC_PSNR_PKT,        /**< PSNR statistics for this frame */
+    // Spatial SVC is still experimental and may be removed before the next ABI
+    // bump.
+#if VPX_ENCODER_ABI_VERSION > (5 + VPX_CODEC_ABI_VERSION)
+    VPX_CODEC_SPATIAL_SVC_LAYER_SIZES, /**< Sizes for each layer in this frame*/
+    VPX_CODEC_SPATIAL_SVC_LAYER_PSNR, /**< PSNR for each layer in this frame*/
+#endif
+    VPX_CODEC_CUSTOM_PKT = 256 /**< Algorithm extensions  */
+  };
+
+
+  /*!\brief Encoder output packet
+   *
+   * This structure contains the different kinds of output data the encoder
+   * may produce while compressing a frame.
+   */
+  typedef struct vpx_codec_cx_pkt {
+    enum vpx_codec_cx_pkt_kind  kind; /**< packet variant */
+    union {
+      struct {
+        void                    *buf;      /**< compressed data buffer */
+        size_t                   sz;       /**< length of compressed data */
+        vpx_codec_pts_t          pts;      /**< time stamp to show frame
+                                                    (in timebase units) */
+        unsigned long            duration; /**< duration to show frame
+                                                    (in timebase units) */
+        vpx_codec_frame_flags_t  flags;    /**< flags for this frame */
+        int                      partition_id; /**< the partition id
+                                              defines the decoding order
+                                              of the partitions. Only
+                                              applicable when "output partition"
+                                              mode is enabled. First partition
+                                              has id 0.*/
+
+      } frame;  /**< data for compressed frame packet */
+      vpx_fixed_buf_t twopass_stats;  /**< data for two-pass packet */
+      vpx_fixed_buf_t firstpass_mb_stats; /**< first pass mb packet */
+      struct vpx_psnr_pkt {
+        unsigned int samples[4];  /**< Number of samples, total/y/u/v */
+        uint64_t     sse[4];      /**< sum squared error, total/y/u/v */
+        double       psnr[4];     /**< PSNR, total/y/u/v */
+      } psnr;                       /**< data for PSNR packet */
+      vpx_fixed_buf_t raw;     /**< data for arbitrary packets */
+      // Spatial SVC is still experimental and may be removed before the next
+      // ABI bump.
+#if VPX_ENCODER_ABI_VERSION > (5 + VPX_CODEC_ABI_VERSION)
+      size_t layer_sizes[VPX_SS_MAX_LAYERS];
+      struct vpx_psnr_pkt layer_psnr[VPX_SS_MAX_LAYERS];
+#endif
+
+      /* This packet size is fixed to allow codecs to extend this
+       * interface without having to manage storage for raw packets,
+       * i.e., if it's smaller than 128 bytes, you can store in the
+       * packet list directly.
+       */
+      char pad[128 - sizeof(enum vpx_codec_cx_pkt_kind)]; /**< fixed sz */
+    } data; /**< packet data */
+  } vpx_codec_cx_pkt_t; /**< alias for struct vpx_codec_cx_pkt */
+
+
+  /*!\brief Encoder return output buffer callback
+   *
+   * This callback function, when registered, returns with packets when each
+   * spatial layer is encoded.
+   */
+  // putting the definitions here for now. (agrange: find if there
+  // is a better place for this)
+  typedef void (* vpx_codec_enc_output_cx_pkt_cb_fn_t)(vpx_codec_cx_pkt_t *pkt,
+                                                       void *user_data);
+
+  /*!\brief Callback function pointer / user data pair storage */
+  typedef struct vpx_codec_enc_output_cx_cb_pair {
+    vpx_codec_enc_output_cx_pkt_cb_fn_t output_cx_pkt; /**< Callback function */
+    void                            *user_priv; /**< Pointer to private data */
+  } vpx_codec_priv_output_cx_pkt_cb_pair_t;
+
+  /*!\brief Rational Number
+   *
+   * This structure holds a fractional value.
+   */
+  typedef struct vpx_rational {
+    int num; /**< fraction numerator */
+    int den; /**< fraction denominator */
+  } vpx_rational_t; /**< alias for struct vpx_rational */
+
+
+  /*!\brief Multi-pass Encoding Pass */
+  enum vpx_enc_pass {
+    VPX_RC_ONE_PASS,   /**< Single pass mode */
+    VPX_RC_FIRST_PASS, /**< First pass of multi-pass mode */
+    VPX_RC_LAST_PASS   /**< Final pass of multi-pass mode */
+  };
+
+
+  /*!\brief Rate control mode */
+  enum vpx_rc_mode {
+    VPX_VBR,  /**< Variable Bit Rate (VBR) mode */
+    VPX_CBR,  /**< Constant Bit Rate (CBR) mode */
+    VPX_CQ,   /**< Constrained Quality (CQ)  mode */
+    VPX_Q,    /**< Constant Quality (Q) mode */
+  };
+
+
+  /*!\brief Keyframe placement mode.
+   *
+   * This enumeration determines whether keyframes are placed automatically by
+   * the encoder or whether this behavior is disabled. Older releases of this
+   * SDK were implemented such that VPX_KF_FIXED meant keyframes were disabled.
+   * This name is confusing for this behavior, so the new symbols to be used
+   * are VPX_KF_AUTO and VPX_KF_DISABLED.
+   */
+  enum vpx_kf_mode {
+    VPX_KF_FIXED, /**< deprecated, implies VPX_KF_DISABLED */
+    VPX_KF_AUTO,  /**< Encoder determines optimal placement automatically */
+    VPX_KF_DISABLED = 0 /**< Encoder does not place keyframes. */
+  };
+
+
+  /*!\brief Encoded Frame Flags
+   *
+   * This type indicates a bitfield to be passed to vpx_codec_encode(), defining
+   * per-frame boolean values. By convention, bits common to all codecs will be
+   * named VPX_EFLAG_*, and bits specific to an algorithm will be named
+   * /algo/_eflag_*. The lower order 16 bits are reserved for common use.
+   */
+  typedef long vpx_enc_frame_flags_t;
+#define VPX_EFLAG_FORCE_KF (1<<0)  /**< Force this frame to be a keyframe */
+
+
+  /*!\brief Encoder configuration structure
+   *
+   * This structure contains the encoder settings that have common representations
+   * across all codecs. This doesn't imply that all codecs support all features,
+   * however.
+   */
+  typedef struct vpx_codec_enc_cfg {
+    /*
+     * generic settings (g)
+     */
+
+    /*!\brief Algorithm specific "usage" value
+     *
+     * Algorithms may define multiple values for usage, which may convey the
+     * intent of how the application intends to use the stream. If this value
+     * is non-zero, consult the documentation for the codec to determine its
+     * meaning.
+     */
+    unsigned int           g_usage;
+
+
+    /*!\brief Maximum number of threads to use
+     *
+     * For multi-threaded implementations, use no more than this number of
+     * threads. The codec may use fewer threads than allowed. The value
+     * 0 is equivalent to the value 1.
+     */
+    unsigned int           g_threads;
+
+
+    /*!\brief Bitstream profile to use
+     *
+     * Some codecs support a notion of multiple bitstream profiles. Typically
+     * this maps to a set of features that are turned on or off. Often the
+     * profile to use is determined by the features of the intended decoder.
+     * Consult the documentation for the codec to determine the valid values
+     * for this parameter, or set to zero for a sane default.
+     */
+    unsigned int           g_profile;  /**< profile of bitstream to use */
+
+
+
+    /*!\brief Width of the frame
+     *
+     * This value identifies the presentation resolution of the frame,
+     * in pixels. Note that the frames passed as input to the encoder must
+     * have this resolution. Frames will be presented by the decoder in this
+     * resolution, independent of any spatial resampling the encoder may do.
+     */
+    unsigned int           g_w;
+
+
+    /*!\brief Height of the frame
+     *
+     * This value identifies the presentation resolution of the frame,
+     * in pixels. Note that the frames passed as input to the encoder must
+     * have this resolution. Frames will be presented by the decoder in this
+     * resolution, independent of any spatial resampling the encoder may do.
+     */
+    unsigned int           g_h;
+
+    /*!\brief Bit-depth of the codec
+     *
+     * This value identifies the bit_depth of the codec,
+     * Only certain bit-depths are supported as identified in the
+     * vpx_bit_depth_t enum.
+     */
+    vpx_bit_depth_t        g_bit_depth;
+
+    /*!\brief Bit-depth of the input frames
+     *
+     * This value identifies the bit_depth of the input frames in bits.
+     * Note that the frames passed as input to the encoder must have
+     * this bit-depth.
+     */
+    unsigned int           g_input_bit_depth;
+
+    /*!\brief Stream timebase units
+     *
+     * Indicates the smallest interval of time, in seconds, used by the stream.
+     * For fixed frame rate material, or variable frame rate material where
+     * frames are timed at a multiple of a given clock (ex: video capture),
+     * the \ref RECOMMENDED method is to set the timebase to the reciprocal
+     * of the frame rate (ex: 1001/30000 for 29.970 Hz NTSC). This allows the
+     * pts to correspond to the frame number, which can be handy. For
+     * re-encoding video from containers with absolute time timestamps, the
+     * \ref RECOMMENDED method is to set the timebase to that of the parent
+     * container or multimedia framework (ex: 1/1000 for ms, as in FLV).
+     */
+    struct vpx_rational    g_timebase;
+
+
+    /*!\brief Enable error resilient modes.
+     *
+     * The error resilient bitfield indicates to the encoder which features
+     * it should enable to take measures for streaming over lossy or noisy
+     * links.
+     */
+    vpx_codec_er_flags_t   g_error_resilient;
+
+
+    /*!\brief Multi-pass Encoding Mode
+     *
+     * This value should be set to the current phase for multi-pass encoding.
+     * For single pass, set to #VPX_RC_ONE_PASS.
+     */
+    enum vpx_enc_pass      g_pass;
+
+
+    /*!\brief Allow lagged encoding
+     *
+     * If set, this value allows the encoder to consume a number of input
+     * frames before producing output frames. This allows the encoder to
+     * base decisions for the current frame on future frames. This does
+     * increase the latency of the encoding pipeline, so it is not appropriate
+     * in all situations (ex: realtime encoding).
+     *
+     * Note that this is a maximum value -- the encoder may produce frames
+     * sooner than the given limit. Set this value to 0 to disable this
+     * feature.
+     */
+    unsigned int           g_lag_in_frames;
+
+
+    /*
+     * rate control settings (rc)
+     */
+
+    /*!\brief Temporal resampling configuration, if supported by the codec.
+     *
+     * Temporal resampling allows the codec to "drop" frames as a strategy to
+     * meet its target data rate. This can cause temporal discontinuities in
+     * the encoded video, which may appear as stuttering during playback. This
+     * trade-off is often acceptable, but for many applications is not. It can
+     * be disabled in these cases.
+     *
+     * Note that not all codecs support this feature. All vpx VPx codecs do.
+     * For other codecs, consult the documentation for that algorithm.
+     *
+     * This threshold is described as a percentage of the target data buffer.
+     * When the data buffer falls below this percentage of fullness, a
+     * dropped frame is indicated. Set the threshold to zero (0) to disable
+     * this feature.
+     */
+    unsigned int           rc_dropframe_thresh;
+
+
+    /*!\brief Enable/disable spatial resampling, if supported by the codec.
+     *
+     * Spatial resampling allows the codec to compress a lower resolution
+     * version of the frame, which is then upscaled by the encoder to the
+     * correct presentation resolution. This increases visual quality at
+     * low data rates, at the expense of CPU time on the encoder/decoder.
+     */
+    unsigned int           rc_resize_allowed;
+
+    /*!\brief Internal coded frame width.
+     *
+     * If spatial resampling is enabled this specifies the width of the
+     * encoded frame.
+     */
+    unsigned int           rc_scaled_width;
+
+    /*!\brief Internal coded frame height.
+     *
+     * If spatial resampling is enabled this specifies the height of the
+     * encoded frame.
+     */
+    unsigned int           rc_scaled_height;
+
+    /*!\brief Spatial resampling up watermark.
+     *
+     * This threshold is described as a percentage of the target data buffer.
+     * When the data buffer rises above this percentage of fullness, the
+     * encoder will step up to a higher resolution version of the frame.
+     */
+    unsigned int           rc_resize_up_thresh;
+
+
+    /*!\brief Spatial resampling down watermark.
+     *
+     * This threshold is described as a percentage of the target data buffer.
+     * When the data buffer falls below this percentage of fullness, the
+     * encoder will step down to a lower resolution version of the frame.
+     */
+    unsigned int           rc_resize_down_thresh;
+
+
+    /*!\brief Rate control algorithm to use.
+     *
+     * Indicates whether the end usage of this stream is to be streamed over
+     * a bandwidth constrained link, indicating that Constant Bit Rate (CBR)
+     * mode should be used, or whether it will be played back on a high
+     * bandwidth link, as from a local disk, where higher variations in
+     * bitrate are acceptable.
+     */
+    enum vpx_rc_mode       rc_end_usage;
+
+
+    /*!\brief Two-pass stats buffer.
+     *
+     * A buffer containing all of the stats packets produced in the first
+     * pass, concatenated.
+     */
+    vpx_fixed_buf_t   rc_twopass_stats_in;
+
+    /*!\brief first pass mb stats buffer.
+     *
+     * A buffer containing all of the first pass mb stats packets produced
+     * in the first pass, concatenated.
+     */
+    vpx_fixed_buf_t   rc_firstpass_mb_stats_in;
+
+    /*!\brief Target data rate
+     *
+     * Target bandwidth to use for this stream, in kilobits per second.
+     */
+    unsigned int           rc_target_bitrate;
+
+
+    /*
+     * quantizer settings
+     */
+
+
+    /*!\brief Minimum (Best Quality) Quantizer
+     *
+     * The quantizer is the most direct control over the quality of the
+     * encoded image. The range of valid values for the quantizer is codec
+     * specific. Consult the documentation for the codec to determine the
+     * values to use. To determine the range programmatically, call
+     * vpx_codec_enc_config_default() with a usage value of 0.
+     */
+    unsigned int           rc_min_quantizer;
+
+
+    /*!\brief Maximum (Worst Quality) Quantizer
+     *
+     * The quantizer is the most direct control over the quality of the
+     * encoded image. The range of valid values for the quantizer is codec
+     * specific. Consult the documentation for the codec to determine the
+     * values to use. To determine the range programmatically, call
+     * vpx_codec_enc_config_default() with a usage value of 0.
+     */
+    unsigned int           rc_max_quantizer;
+
+
+    /*
+     * bitrate tolerance
+     */
+
+
+    /*!\brief Rate control adaptation undershoot control
+     *
+     * This value, expressed as a percentage of the target bitrate,
+     * controls the maximum allowed adaptation speed of the codec.
+     * This factor controls the maximum amount of bits that can
+     * be subtracted from the target bitrate in order to compensate
+     * for prior overshoot.
+     *
+     * Valid values in the range 0-1000.
+     */
+    unsigned int           rc_undershoot_pct;
+
+
+    /*!\brief Rate control adaptation overshoot control
+     *
+     * This value, expressed as a percentage of the target bitrate,
+     * controls the maximum allowed adaptation speed of the codec.
+     * This factor controls the maximum amount of bits that can
+     * be added to the target bitrate in order to compensate for
+     * prior undershoot.
+     *
+     * Valid values in the range 0-1000.
+     */
+    unsigned int           rc_overshoot_pct;
+
+
+    /*
+     * decoder buffer model parameters
+     */
+
+
+    /*!\brief Decoder Buffer Size
+     *
+     * This value indicates the amount of data that may be buffered by the
+     * decoding application. Note that this value is expressed in units of
+     * time (milliseconds). For example, a value of 5000 indicates that the
+     * client will buffer (at least) 5000ms worth of encoded data. Use the
+     * target bitrate (#rc_target_bitrate) to convert to bits/bytes, if
+     * necessary.
+     */
+    unsigned int           rc_buf_sz;
+
+
+    /*!\brief Decoder Buffer Initial Size
+     *
+     * This value indicates the amount of data that will be buffered by the
+     * decoding application prior to beginning playback. This value is
+     * expressed in units of time (milliseconds). Use the target bitrate
+     * (#rc_target_bitrate) to convert to bits/bytes, if necessary.
+     */
+    unsigned int           rc_buf_initial_sz;
+
+
+    /*!\brief Decoder Buffer Optimal Size
+     *
+     * This value indicates the amount of data that the encoder should try
+     * to maintain in the decoder's buffer. This value is expressed in units
+     * of time (milliseconds). Use the target bitrate (#rc_target_bitrate)
+     * to convert to bits/bytes, if necessary.
+     */
+    unsigned int           rc_buf_optimal_sz;
+
+
+    /*
+     * 2 pass rate control parameters
+     */
+
+
+    /*!\brief Two-pass mode CBR/VBR bias
+     *
+     * Bias, expressed on a scale of 0 to 100, for determining target size
+     * for the current frame. The value 0 indicates the optimal CBR mode
+     * value should be used. The value 100 indicates the optimal VBR mode
+     * value should be used. Values in between indicate which way the
+     * encoder should "lean."
+     */
+    unsigned int           rc_2pass_vbr_bias_pct;       /**< RC mode bias between CBR and VBR(0-100: 0->CBR, 100->VBR)   */
+
+
+    /*!\brief Two-pass mode per-GOP minimum bitrate
+     *
+     * This value, expressed as a percentage of the target bitrate, indicates
+     * the minimum bitrate to be used for a single GOP (aka "section")
+     */
+    unsigned int           rc_2pass_vbr_minsection_pct;
+
+
+    /*!\brief Two-pass mode per-GOP maximum bitrate
+     *
+     * This value, expressed as a percentage of the target bitrate, indicates
+     * the maximum bitrate to be used for a single GOP (aka "section")
+     */
+    unsigned int           rc_2pass_vbr_maxsection_pct;
+
+
+    /*
+     * keyframing settings (kf)
+     */
+
+    /*!\brief Keyframe placement mode
+     *
+     * This value indicates whether the encoder should place keyframes at a
+     * fixed interval, or determine the optimal placement automatically
+     * (as governed by the #kf_min_dist and #kf_max_dist parameters)
+     */
+    enum vpx_kf_mode       kf_mode;
+
+
+    /*!\brief Keyframe minimum interval
+     *
+     * This value, expressed as a number of frames, prevents the encoder from
+     * placing a keyframe nearer than kf_min_dist to the previous keyframe. At
+     * least kf_min_dist frames non-keyframes will be coded before the next
+     * keyframe. Set kf_min_dist equal to kf_max_dist for a fixed interval.
+     */
+    unsigned int           kf_min_dist;
+
+
+    /*!\brief Keyframe maximum interval
+     *
+     * This value, expressed as a number of frames, forces the encoder to code
+     * a keyframe if one has not been coded in the last kf_max_dist frames.
+     * A value of 0 implies all frames will be keyframes. Set kf_min_dist
+     * equal to kf_max_dist for a fixed interval.
+     */
+    unsigned int           kf_max_dist;
+
+    /*
+     * Spatial scalability settings (ss)
+     */
+
+    /*!\brief Number of spatial coding layers.
+     *
+     * This value specifies the number of spatial coding layers to be used.
+     */
+    unsigned int           ss_number_layers;
+
+    /*!\brief Enable auto alt reference flags for each spatial layer.
+     *
+     * These values specify if auto alt reference frame is enabled for each
+     * spatial layer.
+     */
+    int                    ss_enable_auto_alt_ref[VPX_SS_MAX_LAYERS];
+
+    /*!\brief Target bitrate for each spatial layer.
+     *
+     * These values specify the target coding bitrate to be used for each
+     * spatial layer.
+     */
+    unsigned int           ss_target_bitrate[VPX_SS_MAX_LAYERS];
+
+    /*!\brief Number of temporal coding layers.
+     *
+     * This value specifies the number of temporal layers to be used.
+     */
+    unsigned int           ts_number_layers;
+
+    /*!\brief Target bitrate for each temporal layer.
+     *
+     * These values specify the target coding bitrate to be used for each
+     * temporal layer.
+     */
+    unsigned int           ts_target_bitrate[VPX_TS_MAX_LAYERS];
+
+    /*!\brief Frame rate decimation factor for each temporal layer.
+     *
+     * These values specify the frame rate decimation factors to apply
+     * to each temporal layer.
+     */
+    unsigned int           ts_rate_decimator[VPX_TS_MAX_LAYERS];
+
+    /*!\brief Length of the sequence defining frame temporal layer membership.
+     *
+     * This value specifies the length of the sequence that defines the
+     * membership of frames to temporal layers. For example, if the
+     * ts_periodicity = 8, then the frames are assigned to coding layers with a
+     * repeated sequence of length 8.
+    */
+    unsigned int           ts_periodicity;
+
+    /*!\brief Template defining the membership of frames to temporal layers.
+     *
+     * This array defines the membership of frames to temporal coding layers.
+     * For a 2-layer encoding that assigns even numbered frames to one temporal
+     * layer (0) and odd numbered frames to a second temporal layer (1) with
+     * ts_periodicity=8, then ts_layer_id = (0,1,0,1,0,1,0,1).
+    */
+    unsigned int           ts_layer_id[VPX_TS_MAX_PERIODICITY];
+
+    /*!\brief Target bitrate for each spatial/temporal layer.
+     *
+     * These values specify the target coding bitrate to be used for each
+     * spatial/temporal layer.
+     *
+     */
+    unsigned int           layer_target_bitrate[VPX_MAX_LAYERS];
+
+    /*!\brief Temporal layering mode indicating which temporal layering scheme to use.
+     *
+     * The value (refer to VP9E_TEMPORAL_LAYERING_MODE) specifies the
+     * temporal layering mode to use.
+     *
+     */
+    int                    temporal_layering_mode;
+  } vpx_codec_enc_cfg_t; /**< alias for struct vpx_codec_enc_cfg */
+
+  /*!\brief  vp9 svc extra configure parameters
+   *
+   * This defines max/min quantizers and scale factors for each layer
+   *
+   */
+  typedef struct vpx_svc_parameters {
+    int max_quantizers[VPX_MAX_LAYERS]; /**< Max Q for each layer */
+    int min_quantizers[VPX_MAX_LAYERS]; /**< Min Q for each layer */
+    int scaling_factor_num[VPX_MAX_LAYERS]; /**< Scaling factor-numerator */
+    int scaling_factor_den[VPX_MAX_LAYERS]; /**< Scaling factor-denominator */
+    int temporal_layering_mode; /**< Temporal layering mode */
+  } vpx_svc_extra_cfg_t;
+
+
+  /*!\brief Initialize an encoder instance
+   *
+   * Initializes a encoder context using the given interface. Applications
+   * should call the vpx_codec_enc_init convenience macro instead of this
+   * function directly, to ensure that the ABI version number parameter
+   * is properly initialized.
+   *
+   * If the library was configured with --disable-multithread, this call
+   * is not thread safe and should be guarded with a lock if being used
+   * in a multithreaded context.
+   *
+   * \param[in]    ctx     Pointer to this instance's context.
+   * \param[in]    iface   Pointer to the algorithm interface to use.
+   * \param[in]    cfg     Configuration to use, if known. May be NULL.
+   * \param[in]    flags   Bitfield of VPX_CODEC_USE_* flags
+   * \param[in]    ver     ABI version number. Must be set to
+   *                       VPX_ENCODER_ABI_VERSION
+   * \retval #VPX_CODEC_OK
+   *     The decoder algorithm initialized.
+   * \retval #VPX_CODEC_MEM_ERROR
+   *     Memory allocation failed.
+   */
+  vpx_codec_err_t vpx_codec_enc_init_ver(vpx_codec_ctx_t      *ctx,
+                                         vpx_codec_iface_t    *iface,
+                                         const vpx_codec_enc_cfg_t *cfg,
+                                         vpx_codec_flags_t     flags,
+                                         int                   ver);
+
+
+  /*!\brief Convenience macro for vpx_codec_enc_init_ver()
+   *
+   * Ensures the ABI version parameter is properly set.
+   */
+#define vpx_codec_enc_init(ctx, iface, cfg, flags) \
+  vpx_codec_enc_init_ver(ctx, iface, cfg, flags, VPX_ENCODER_ABI_VERSION)
+
+
+  /*!\brief Initialize multi-encoder instance
+   *
+   * Initializes multi-encoder context using the given interface.
+   * Applications should call the vpx_codec_enc_init_multi convenience macro
+   * instead of this function directly, to ensure that the ABI version number
+   * parameter is properly initialized.
+   *
+   * \param[in]    ctx     Pointer to this instance's context.
+   * \param[in]    iface   Pointer to the algorithm interface to use.
+   * \param[in]    cfg     Configuration to use, if known. May be NULL.
+   * \param[in]    num_enc Total number of encoders.
+   * \param[in]    flags   Bitfield of VPX_CODEC_USE_* flags
+   * \param[in]    dsf     Pointer to down-sampling factors.
+   * \param[in]    ver     ABI version number. Must be set to
+   *                       VPX_ENCODER_ABI_VERSION
+   * \retval #VPX_CODEC_OK
+   *     The decoder algorithm initialized.
+   * \retval #VPX_CODEC_MEM_ERROR
+   *     Memory allocation failed.
+   */
+  vpx_codec_err_t vpx_codec_enc_init_multi_ver(vpx_codec_ctx_t      *ctx,
+                                               vpx_codec_iface_t    *iface,
+                                               vpx_codec_enc_cfg_t  *cfg,
+                                               int                   num_enc,
+                                               vpx_codec_flags_t     flags,
+                                               vpx_rational_t       *dsf,
+                                               int                   ver);
+
+
+  /*!\brief Convenience macro for vpx_codec_enc_init_multi_ver()
+   *
+   * Ensures the ABI version parameter is properly set.
+   */
+#define vpx_codec_enc_init_multi(ctx, iface, cfg, num_enc, flags, dsf) \
+  vpx_codec_enc_init_multi_ver(ctx, iface, cfg, num_enc, flags, dsf, \
+                               VPX_ENCODER_ABI_VERSION)
+
+
+  /*!\brief Get a default configuration
+   *
+   * Initializes a encoder configuration structure with default values. Supports
+   * the notion of "usages" so that an algorithm may offer different default
+   * settings depending on the user's intended goal. This function \ref SHOULD
+   * be called by all applications to initialize the configuration structure
+   * before specializing the configuration with application specific values.
+   *
+   * \param[in]    iface     Pointer to the algorithm interface to use.
+   * \param[out]   cfg       Configuration buffer to populate.
+   * \param[in]    reserved  Must set to 0 for VP8 and VP9.
+   *
+   * \retval #VPX_CODEC_OK
+   *     The configuration was populated.
+   * \retval #VPX_CODEC_INCAPABLE
+   *     Interface is not an encoder interface.
+   * \retval #VPX_CODEC_INVALID_PARAM
+   *     A parameter was NULL, or the usage value was not recognized.
+   */
+  vpx_codec_err_t  vpx_codec_enc_config_default(vpx_codec_iface_t    *iface,
+                                                vpx_codec_enc_cfg_t  *cfg,
+                                                unsigned int          reserved);
+
+
+  /*!\brief Set or change configuration
+   *
+   * Reconfigures an encoder instance according to the given configuration.
+   *
+   * \param[in]    ctx     Pointer to this instance's context
+   * \param[in]    cfg     Configuration buffer to use
+   *
+   * \retval #VPX_CODEC_OK
+   *     The configuration was populated.
+   * \retval #VPX_CODEC_INCAPABLE
+   *     Interface is not an encoder interface.
+   * \retval #VPX_CODEC_INVALID_PARAM
+   *     A parameter was NULL, or the usage value was not recognized.
+   */
+  vpx_codec_err_t  vpx_codec_enc_config_set(vpx_codec_ctx_t            *ctx,
+                                            const vpx_codec_enc_cfg_t  *cfg);
+
+
+  /*!\brief Get global stream headers
+   *
+   * Retrieves a stream level global header packet, if supported by the codec.
+   *
+   * \param[in]    ctx     Pointer to this instance's context
+   *
+   * \retval NULL
+   *     Encoder does not support global header
+   * \retval Non-NULL
+   *     Pointer to buffer containing global header packet
+   */
+  vpx_fixed_buf_t *vpx_codec_get_global_headers(vpx_codec_ctx_t   *ctx);
+
+
+#define VPX_DL_REALTIME     (1)        /**< deadline parameter analogous to
+  *   VPx REALTIME mode. */
+#define VPX_DL_GOOD_QUALITY (1000000)  /**< deadline parameter analogous to
+  *   VPx GOOD QUALITY mode. */
+#define VPX_DL_BEST_QUALITY (0)        /**< deadline parameter analogous to
+  *   VPx BEST QUALITY mode. */
+  /*!\brief Encode a frame
+   *
+   * Encodes a video frame at the given "presentation time." The presentation
+   * time stamp (PTS) \ref MUST be strictly increasing.
+   *
+   * The encoder supports the notion of a soft real-time deadline. Given a
+   * non-zero value to the deadline parameter, the encoder will make a "best
+   * effort" guarantee to  return before the given time slice expires. It is
+   * implicit that limiting the available time to encode will degrade the
+   * output quality. The encoder can be given an unlimited time to produce the
+   * best possible frame by specifying a deadline of '0'. This deadline
+   * supercedes the VPx notion of "best quality, good quality, realtime".
+   * Applications that wish to map these former settings to the new deadline
+   * based system can use the symbols #VPX_DL_REALTIME, #VPX_DL_GOOD_QUALITY,
+   * and #VPX_DL_BEST_QUALITY.
+   *
+   * When the last frame has been passed to the encoder, this function should
+   * continue to be called, with the img parameter set to NULL. This will
+   * signal the end-of-stream condition to the encoder and allow it to encode
+   * any held buffers. Encoding is complete when vpx_codec_encode() is called
+   * and vpx_codec_get_cx_data() returns no data.
+   *
+   * \param[in]    ctx       Pointer to this instance's context
+   * \param[in]    img       Image data to encode, NULL to flush.
+   * \param[in]    pts       Presentation time stamp, in timebase units.
+   * \param[in]    duration  Duration to show frame, in timebase units.
+   * \param[in]    flags     Flags to use for encoding this frame.
+   * \param[in]    deadline  Time to spend encoding, in microseconds. (0=infinite)
+   *
+   * \retval #VPX_CODEC_OK
+   *     The configuration was populated.
+   * \retval #VPX_CODEC_INCAPABLE
+   *     Interface is not an encoder interface.
+   * \retval #VPX_CODEC_INVALID_PARAM
+   *     A parameter was NULL, the image format is unsupported, etc.
+   */
+  vpx_codec_err_t  vpx_codec_encode(vpx_codec_ctx_t            *ctx,
+                                    const vpx_image_t          *img,
+                                    vpx_codec_pts_t             pts,
+                                    unsigned long               duration,
+                                    vpx_enc_frame_flags_t       flags,
+                                    unsigned long               deadline);
+
+  /*!\brief Set compressed data output buffer
+   *
+   * Sets the buffer that the codec should output the compressed data
+   * into. This call effectively sets the buffer pointer returned in the
+   * next VPX_CODEC_CX_FRAME_PKT packet. Subsequent packets will be
+   * appended into this buffer. The buffer is preserved across frames,
+   * so applications must periodically call this function after flushing
+   * the accumulated compressed data to disk or to the network to reset
+   * the pointer to the buffer's head.
+   *
+   * `pad_before` bytes will be skipped before writing the compressed
+   * data, and `pad_after` bytes will be appended to the packet. The size
+   * of the packet will be the sum of the size of the actual compressed
+   * data, pad_before, and pad_after. The padding bytes will be preserved
+   * (not overwritten).
+   *
+   * Note that calling this function does not guarantee that the returned
+   * compressed data will be placed into the specified buffer. In the
+   * event that the encoded data will not fit into the buffer provided,
+   * the returned packet \ref MAY point to an internal buffer, as it would
+   * if this call were never used. In this event, the output packet will
+   * NOT have any padding, and the application must free space and copy it
+   * to the proper place. This is of particular note in configurations
+   * that may output multiple packets for a single encoded frame (e.g., lagged
+   * encoding) or if the application does not reset the buffer periodically.
+   *
+   * Applications may restore the default behavior of the codec providing
+   * the compressed data buffer by calling this function with a NULL
+   * buffer.
+   *
+   * Applications \ref MUSTNOT call this function during iteration of
+   * vpx_codec_get_cx_data().
+   *
+   * \param[in]    ctx         Pointer to this instance's context
+   * \param[in]    buf         Buffer to store compressed data into
+   * \param[in]    pad_before  Bytes to skip before writing compressed data
+   * \param[in]    pad_after   Bytes to skip after writing compressed data
+   *
+   * \retval #VPX_CODEC_OK
+   *     The buffer was set successfully.
+   * \retval #VPX_CODEC_INVALID_PARAM
+   *     A parameter was NULL, the image format is unsupported, etc.
+   */
+  vpx_codec_err_t vpx_codec_set_cx_data_buf(vpx_codec_ctx_t       *ctx,
+                                            const vpx_fixed_buf_t *buf,
+                                            unsigned int           pad_before,
+                                            unsigned int           pad_after);
+
+
+  /*!\brief Encoded data iterator
+   *
+   * Iterates over a list of data packets to be passed from the encoder to the
+   * application. The different kinds of packets available are enumerated in
+   * #vpx_codec_cx_pkt_kind.
+   *
+   * #VPX_CODEC_CX_FRAME_PKT packets should be passed to the application's
+   * muxer. Multiple compressed frames may be in the list.
+   * #VPX_CODEC_STATS_PKT packets should be appended to a global buffer.
+   *
+   * The application \ref MUST silently ignore any packet kinds that it does
+   * not recognize or support.
+   *
+   * The data buffers returned from this function are only guaranteed to be
+   * valid until the application makes another call to any vpx_codec_* function.
+   *
+   * \param[in]     ctx      Pointer to this instance's context
+   * \param[in,out] iter     Iterator storage, initialized to NULL
+   *
+   * \return Returns a pointer to an output data packet (compressed frame data,
+   *         two-pass statistics, etc.) or NULL to signal end-of-list.
+   *
+   */
+  const vpx_codec_cx_pkt_t *vpx_codec_get_cx_data(vpx_codec_ctx_t   *ctx,
+                                                  vpx_codec_iter_t  *iter);
+
+
+  /*!\brief Get Preview Frame
+   *
+   * Returns an image that can be used as a preview. Shows the image as it would
+   * exist at the decompressor. The application \ref MUST NOT write into this
+   * image buffer.
+   *
+   * \param[in]     ctx      Pointer to this instance's context
+   *
+   * \return Returns a pointer to a preview image, or NULL if no image is
+   *         available.
+   *
+   */
+  const vpx_image_t *vpx_codec_get_preview_frame(vpx_codec_ctx_t   *ctx);
+
+
+  /*!@} - end defgroup encoder*/
+#ifdef __cplusplus
+}
+#endif
+#endif  // VPX_VPX_ENCODER_H_
+
diff --git a/libvpx/libvpx/vpx/vpx_frame_buffer.h b/libvpx/libvpx/vpx/vpx_frame_buffer.h
new file mode 100644
index 0000000..9036459
--- /dev/null
+++ b/libvpx/libvpx/vpx/vpx_frame_buffer.h
@@ -0,0 +1,83 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_VPX_FRAME_BUFFER_H_
+#define VPX_VPX_FRAME_BUFFER_H_
+
+/*!\file
+ * \brief Describes the decoder external frame buffer interface.
+ */
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./vpx_integer.h"
+
+/*!\brief The maximum number of work buffers used by libvpx.
+ *  Support maximum 4 threads to decode video in parallel.
+ *  Each thread will use one work buffer.
+ * TODO(hkuang): Add support to set number of worker threads dynamically.
+ */
+#define VPX_MAXIMUM_WORK_BUFFERS 8
+
+/*!\brief The maximum number of reference buffers that a VP9 encoder may use.
+ */
+#define VP9_MAXIMUM_REF_BUFFERS 8
+
+/*!\brief External frame buffer
+ *
+ * This structure holds allocated frame buffers used by the decoder.
+ */
+typedef struct vpx_codec_frame_buffer {
+  uint8_t *data;  /**< Pointer to the data buffer */
+  size_t size;  /**< Size of data in bytes */
+  void *priv;  /**< Frame's private data */
+} vpx_codec_frame_buffer_t;
+
+/*!\brief get frame buffer callback prototype
+ *
+ * This callback is invoked by the decoder to retrieve data for the frame
+ * buffer in order for the decode call to complete. The callback must
+ * allocate at least min_size in bytes and assign it to fb->data. The callback
+ * must zero out all the data allocated. Then the callback must set fb->size
+ * to the allocated size. The application does not need to align the allocated
+ * data. The callback is triggered when the decoder needs a frame buffer to
+ * decode a compressed image into. This function may be called more than once
+ * for every call to vpx_codec_decode. The application may set fb->priv to
+ * some data which will be passed back in the ximage and the release function
+ * call. |fb| is guaranteed to not be NULL. On success the callback must
+ * return 0. Any failure the callback must return a value less than 0.
+ *
+ * \param[in] priv         Callback's private data
+ * \param[in] new_size     Size in bytes needed by the buffer
+ * \param[in,out] fb       Pointer to vpx_codec_frame_buffer_t
+ */
+typedef int (*vpx_get_frame_buffer_cb_fn_t)(
+    void *priv, size_t min_size, vpx_codec_frame_buffer_t *fb);
+
+/*!\brief release frame buffer callback prototype
+ *
+ * This callback is invoked by the decoder when the frame buffer is not
+ * referenced by any other buffers. |fb| is guaranteed to not be NULL. On
+ * success the callback must return 0. Any failure the callback must return
+ * a value less than 0.
+ *
+ * \param[in] priv         Callback's private data
+ * \param[in] fb           Pointer to vpx_codec_frame_buffer_t
+ */
+typedef int (*vpx_release_frame_buffer_cb_fn_t)(
+    void *priv, vpx_codec_frame_buffer_t *fb);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_VPX_FRAME_BUFFER_H_
diff --git a/libvpx/libvpx/vpx/vpx_image.h b/libvpx/libvpx/vpx/vpx_image.h
new file mode 100644
index 0000000..7958c69
--- /dev/null
+++ b/libvpx/libvpx/vpx/vpx_image.h
@@ -0,0 +1,235 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+/*!\file
+ * \brief Describes the vpx image descriptor and associated operations
+ *
+ */
+#ifndef VPX_VPX_IMAGE_H_
+#define VPX_VPX_IMAGE_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+  /*!\brief Current ABI version number
+   *
+   * \internal
+   * If this file is altered in any way that changes the ABI, this value
+   * must be bumped.  Examples include, but are not limited to, changing
+   * types, removing or reassigning enums, adding/removing/rearranging
+   * fields to structures
+   */
+#define VPX_IMAGE_ABI_VERSION (4) /**<\hideinitializer*/
+
+
+#define VPX_IMG_FMT_PLANAR     0x100  /**< Image is a planar format. */
+#define VPX_IMG_FMT_UV_FLIP    0x200  /**< V plane precedes U in memory. */
+#define VPX_IMG_FMT_HAS_ALPHA  0x400  /**< Image has an alpha channel. */
+#define VPX_IMG_FMT_HIGHBITDEPTH 0x800  /**< Image uses 16bit framebuffer. */
+
+  /*!\brief List of supported image formats */
+  typedef enum vpx_img_fmt {
+    VPX_IMG_FMT_NONE,
+    VPX_IMG_FMT_RGB24,   /**< 24 bit per pixel packed RGB */
+    VPX_IMG_FMT_RGB32,   /**< 32 bit per pixel packed 0RGB */
+    VPX_IMG_FMT_RGB565,  /**< 16 bit per pixel, 565 */
+    VPX_IMG_FMT_RGB555,  /**< 16 bit per pixel, 555 */
+    VPX_IMG_FMT_UYVY,    /**< UYVY packed YUV */
+    VPX_IMG_FMT_YUY2,    /**< YUYV packed YUV */
+    VPX_IMG_FMT_YVYU,    /**< YVYU packed YUV */
+    VPX_IMG_FMT_BGR24,   /**< 24 bit per pixel packed BGR */
+    VPX_IMG_FMT_RGB32_LE, /**< 32 bit packed BGR0 */
+    VPX_IMG_FMT_ARGB,     /**< 32 bit packed ARGB, alpha=255 */
+    VPX_IMG_FMT_ARGB_LE,  /**< 32 bit packed BGRA, alpha=255 */
+    VPX_IMG_FMT_RGB565_LE,  /**< 16 bit per pixel, gggbbbbb rrrrrggg */
+    VPX_IMG_FMT_RGB555_LE,  /**< 16 bit per pixel, gggbbbbb 0rrrrrgg */
+    VPX_IMG_FMT_YV12    = VPX_IMG_FMT_PLANAR | VPX_IMG_FMT_UV_FLIP | 1, /**< planar YVU */
+    VPX_IMG_FMT_I420    = VPX_IMG_FMT_PLANAR | 2,
+    VPX_IMG_FMT_VPXYV12 = VPX_IMG_FMT_PLANAR | VPX_IMG_FMT_UV_FLIP | 3, /** < planar 4:2:0 format with vpx color space */
+    VPX_IMG_FMT_VPXI420 = VPX_IMG_FMT_PLANAR | 4,
+    VPX_IMG_FMT_I422    = VPX_IMG_FMT_PLANAR | 5,
+    VPX_IMG_FMT_I444    = VPX_IMG_FMT_PLANAR | 6,
+    VPX_IMG_FMT_I440    = VPX_IMG_FMT_PLANAR | 7,
+    VPX_IMG_FMT_444A    = VPX_IMG_FMT_PLANAR | VPX_IMG_FMT_HAS_ALPHA | 6,
+    VPX_IMG_FMT_I42016    = VPX_IMG_FMT_I420 | VPX_IMG_FMT_HIGHBITDEPTH,
+    VPX_IMG_FMT_I42216    = VPX_IMG_FMT_I422 | VPX_IMG_FMT_HIGHBITDEPTH,
+    VPX_IMG_FMT_I44416    = VPX_IMG_FMT_I444 | VPX_IMG_FMT_HIGHBITDEPTH,
+    VPX_IMG_FMT_I44016    = VPX_IMG_FMT_I440 | VPX_IMG_FMT_HIGHBITDEPTH
+  } vpx_img_fmt_t; /**< alias for enum vpx_img_fmt */
+
+  /*!\brief List of supported color spaces */
+  typedef enum vpx_color_space {
+    VPX_CS_UNKNOWN    = 0,  /**< Unknown */
+    VPX_CS_BT_601     = 1,  /**< BT.601 */
+    VPX_CS_BT_709     = 2,  /**< BT.709 */
+    VPX_CS_SMPTE_170  = 3,  /**< SMPTE.170 */
+    VPX_CS_SMPTE_240  = 4,  /**< SMPTE.240 */
+    VPX_CS_BT_2020    = 5,  /**< BT.2020 */
+    VPX_CS_RESERVED   = 6,  /**< Reserved */
+    VPX_CS_SRGB       = 7   /**< sRGB */
+  } vpx_color_space_t; /**< alias for enum vpx_color_space */
+
+  /*!\brief List of supported color range */
+  typedef enum vpx_color_range {
+    VPX_CR_STUDIO_RANGE = 0,    /**< Y [16..235], UV [16..240] */
+    VPX_CR_FULL_RANGE   = 1     /**< YUV/RGB [0..255] */
+  } vpx_color_range_t; /**< alias for enum vpx_color_range */
+
+  /**\brief Image Descriptor */
+  typedef struct vpx_image {
+    vpx_img_fmt_t fmt; /**< Image Format */
+    vpx_color_space_t cs; /**< Color Space */
+    vpx_color_range_t range; /**< Color Range */
+
+    /* Image storage dimensions */
+    unsigned int  w;           /**< Stored image width */
+    unsigned int  h;           /**< Stored image height */
+    unsigned int  bit_depth;   /**< Stored image bit-depth */
+
+    /* Image display dimensions */
+    unsigned int  d_w;   /**< Displayed image width */
+    unsigned int  d_h;   /**< Displayed image height */
+
+    /* Image intended rendering dimensions */
+    unsigned int  r_w;   /**< Intended rendering image width */
+    unsigned int  r_h;   /**< Intended rendering image height */
+
+    /* Chroma subsampling info */
+    unsigned int  x_chroma_shift;   /**< subsampling order, X */
+    unsigned int  y_chroma_shift;   /**< subsampling order, Y */
+
+    /* Image data pointers. */
+#define VPX_PLANE_PACKED 0   /**< To be used for all packed formats */
+#define VPX_PLANE_Y      0   /**< Y (Luminance) plane */
+#define VPX_PLANE_U      1   /**< U (Chroma) plane */
+#define VPX_PLANE_V      2   /**< V (Chroma) plane */
+#define VPX_PLANE_ALPHA  3   /**< A (Transparency) plane */
+    unsigned char *planes[4];  /**< pointer to the top left pixel for each plane */
+    int      stride[4];  /**< stride between rows for each plane */
+
+    int     bps; /**< bits per sample (for packed formats) */
+
+    /* The following member may be set by the application to associate data
+     * with this image.
+     */
+    void    *user_priv; /**< may be set by the application to associate data
+                         *   with this image. */
+
+    /* The following members should be treated as private. */
+    unsigned char *img_data;       /**< private */
+    int      img_data_owner; /**< private */
+    int      self_allocd;    /**< private */
+
+    void    *fb_priv; /**< Frame buffer data associated with the image. */
+  } vpx_image_t; /**< alias for struct vpx_image */
+
+  /**\brief Representation of a rectangle on a surface */
+  typedef struct vpx_image_rect {
+    unsigned int x; /**< leftmost column */
+    unsigned int y; /**< topmost row */
+    unsigned int w; /**< width */
+    unsigned int h; /**< height */
+  } vpx_image_rect_t; /**< alias for struct vpx_image_rect */
+
+  /*!\brief Open a descriptor, allocating storage for the underlying image
+   *
+   * Returns a descriptor for storing an image of the given format. The
+   * storage for the descriptor is allocated on the heap.
+   *
+   * \param[in]    img       Pointer to storage for descriptor. If this parameter
+   *                         is NULL, the storage for the descriptor will be
+   *                         allocated on the heap.
+   * \param[in]    fmt       Format for the image
+   * \param[in]    d_w       Width of the image
+   * \param[in]    d_h       Height of the image
+   * \param[in]    align     Alignment, in bytes, of the image buffer and
+   *                         each row in the image(stride).
+   *
+   * \return Returns a pointer to the initialized image descriptor. If the img
+   *         parameter is non-null, the value of the img parameter will be
+   *         returned.
+   */
+  vpx_image_t *vpx_img_alloc(vpx_image_t  *img,
+                             vpx_img_fmt_t fmt,
+                             unsigned int d_w,
+                             unsigned int d_h,
+                             unsigned int align);
+
+  /*!\brief Open a descriptor, using existing storage for the underlying image
+   *
+   * Returns a descriptor for storing an image of the given format. The
+   * storage for descriptor has been allocated elsewhere, and a descriptor is
+   * desired to "wrap" that storage.
+   *
+   * \param[in]    img       Pointer to storage for descriptor. If this parameter
+   *                         is NULL, the storage for the descriptor will be
+   *                         allocated on the heap.
+   * \param[in]    fmt       Format for the image
+   * \param[in]    d_w       Width of the image
+   * \param[in]    d_h       Height of the image
+   * \param[in]    align     Alignment, in bytes, of each row in the image.
+   * \param[in]    img_data  Storage to use for the image
+   *
+   * \return Returns a pointer to the initialized image descriptor. If the img
+   *         parameter is non-null, the value of the img parameter will be
+   *         returned.
+   */
+  vpx_image_t *vpx_img_wrap(vpx_image_t  *img,
+                            vpx_img_fmt_t fmt,
+                            unsigned int d_w,
+                            unsigned int d_h,
+                            unsigned int align,
+                            unsigned char      *img_data);
+
+
+  /*!\brief Set the rectangle identifying the displayed portion of the image
+   *
+   * Updates the displayed rectangle (aka viewport) on the image surface to
+   * match the specified coordinates and size.
+   *
+   * \param[in]    img       Image descriptor
+   * \param[in]    x         leftmost column
+   * \param[in]    y         topmost row
+   * \param[in]    w         width
+   * \param[in]    h         height
+   *
+   * \return 0 if the requested rectangle is valid, nonzero otherwise.
+   */
+  int vpx_img_set_rect(vpx_image_t  *img,
+                       unsigned int  x,
+                       unsigned int  y,
+                       unsigned int  w,
+                       unsigned int  h);
+
+
+  /*!\brief Flip the image vertically (top for bottom)
+   *
+   * Adjusts the image descriptor's pointers and strides to make the image
+   * be referenced upside-down.
+   *
+   * \param[in]    img       Image descriptor
+   */
+  void vpx_img_flip(vpx_image_t *img);
+
+  /*!\brief Close an image descriptor
+   *
+   * Frees all allocated storage associated with an image descriptor.
+   *
+   * \param[in]    img       Image descriptor
+   */
+  void vpx_img_free(vpx_image_t *img);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_VPX_IMAGE_H_
diff --git a/libvpx/libvpx/vpx/vpx_integer.h b/libvpx/libvpx/vpx/vpx_integer.h
new file mode 100644
index 0000000..829c9d1
--- /dev/null
+++ b/libvpx/libvpx/vpx/vpx_integer.h
@@ -0,0 +1,74 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VPX_VPX_INTEGER_H_
+#define VPX_VPX_INTEGER_H_
+
+/* get ptrdiff_t, size_t, wchar_t, NULL */
+#include <stddef.h>
+
+#if defined(_MSC_VER)
+#define VPX_FORCE_INLINE __forceinline
+#define VPX_INLINE __inline
+#else
+#define VPX_FORCE_INLINE __inline__ __attribute__(always_inline)
+// TODO(jbb): Allow a way to force inline off for older compilers.
+#define VPX_INLINE inline
+#endif
+
+#if (defined(_MSC_VER) && (_MSC_VER < 1600)) || defined(VPX_EMULATE_INTTYPES)
+typedef signed char  int8_t;
+typedef signed short int16_t;
+typedef signed int   int32_t;
+
+typedef unsigned char  uint8_t;
+typedef unsigned short uint16_t;
+typedef unsigned int   uint32_t;
+
+#if (defined(_MSC_VER) && (_MSC_VER < 1600))
+typedef signed __int64   int64_t;
+typedef unsigned __int64 uint64_t;
+#define INT64_MAX _I64_MAX
+#define INT32_MAX _I32_MAX
+#define INT32_MIN _I32_MIN
+#define INT16_MAX _I16_MAX
+#define INT16_MIN _I16_MIN
+#endif
+
+#ifndef _UINTPTR_T_DEFINED
+typedef size_t uintptr_t;
+#endif
+
+#else
+
+/* Most platforms have the C99 standard integer types. */
+
+#if defined(__cplusplus)
+# if !defined(__STDC_FORMAT_MACROS)
+#  define __STDC_FORMAT_MACROS
+# endif
+# if !defined(__STDC_LIMIT_MACROS)
+#  define __STDC_LIMIT_MACROS
+# endif
+#endif  // __cplusplus
+
+#include <stdint.h>
+
+#endif
+
+/* VS2010 defines stdint.h, but not inttypes.h */
+#if defined(_MSC_VER) && _MSC_VER < 1800
+#define PRId64 "I64d"
+#else
+#include <inttypes.h>
+#endif
+
+#endif  // VPX_VPX_INTEGER_H_
diff --git a/libvpx/libvpx/vpx_dsp/add_noise.c b/libvpx/libvpx/vpx_dsp/add_noise.c
new file mode 100644
index 0000000..682b444
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/add_noise.c
@@ -0,0 +1,40 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+void vpx_plane_add_noise_c(uint8_t *start, char *noise,
+                           char blackclamp[16],
+                           char whiteclamp[16],
+                           char bothclamp[16],
+                           unsigned int width, unsigned int height, int pitch) {
+  unsigned int i, j;
+
+  for (i = 0; i < height; i++) {
+    uint8_t *pos = start + i * pitch;
+    char  *ref = (char *)(noise + (rand() & 0xff));  // NOLINT
+
+    for (j = 0; j < width; j++) {
+      int v = pos[j];
+
+      v = clamp(v - blackclamp[0], 0, 255);
+      v = clamp(v + bothclamp[0], 0, 255);
+      v = clamp(v - whiteclamp[0], 0, 255);
+
+      pos[j] = v + ref[j];
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/avg_neon.c b/libvpx/libvpx/vpx_dsp/arm/avg_neon.c
new file mode 100644
index 0000000..e52958c
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/avg_neon.c
@@ -0,0 +1,256 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+static INLINE unsigned int horizontal_add_u16x8(const uint16x8_t v_16x8) {
+  const uint32x4_t a = vpaddlq_u16(v_16x8);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+
+unsigned int vpx_avg_4x4_neon(const uint8_t *s, int p) {
+  uint16x8_t v_sum;
+  uint32x2_t v_s0 = vdup_n_u32(0);
+  uint32x2_t v_s1 = vdup_n_u32(0);
+  v_s0 = vld1_lane_u32((const uint32_t *)s, v_s0, 0);
+  v_s0 = vld1_lane_u32((const uint32_t *)(s + p), v_s0, 1);
+  v_s1 = vld1_lane_u32((const uint32_t *)(s + 2 * p), v_s1, 0);
+  v_s1 = vld1_lane_u32((const uint32_t *)(s + 3 * p), v_s1, 1);
+  v_sum = vaddl_u8(vreinterpret_u8_u32(v_s0), vreinterpret_u8_u32(v_s1));
+  return (horizontal_add_u16x8(v_sum) + 8) >> 4;
+}
+
+unsigned int vpx_avg_8x8_neon(const uint8_t *s, int p) {
+  uint8x8_t v_s0 = vld1_u8(s);
+  const uint8x8_t v_s1 = vld1_u8(s + p);
+  uint16x8_t v_sum = vaddl_u8(v_s0, v_s1);
+
+  v_s0 = vld1_u8(s + 2 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 3 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 4 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 5 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 6 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 7 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  return (horizontal_add_u16x8(v_sum) + 32) >> 6;
+}
+
+// coeff: 16 bits, dynamic range [-32640, 32640].
+// length: value range {16, 64, 256, 1024}.
+int vpx_satd_neon(const int16_t *coeff, int length) {
+  const int16x4_t zero = vdup_n_s16(0);
+  int32x4_t accum = vdupq_n_s32(0);
+
+  do {
+    const int16x8_t src0 = vld1q_s16(coeff);
+    const int16x8_t src8 = vld1q_s16(coeff + 8);
+    accum = vabal_s16(accum, vget_low_s16(src0), zero);
+    accum = vabal_s16(accum, vget_high_s16(src0), zero);
+    accum = vabal_s16(accum, vget_low_s16(src8), zero);
+    accum = vabal_s16(accum, vget_high_s16(src8), zero);
+    length -= 16;
+    coeff += 16;
+  } while (length != 0);
+
+  {
+    // satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
+    const int64x2_t s0 = vpaddlq_s32(accum);  // cascading summation of 'accum'.
+    const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
+                                  vreinterpret_s32_s64(vget_high_s64(s0)));
+    const int satd = vget_lane_s32(s1, 0);
+    return satd;
+  }
+}
+
+void vpx_int_pro_row_neon(int16_t hbuf[16], uint8_t const *ref,
+                          const int ref_stride, const int height) {
+  int i;
+  uint16x8_t vec_sum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_hi = vdupq_n_u16(0);
+  const int shift_factor = ((height >> 5) + 3) * -1;
+  const int16x8_t vec_shift = vdupq_n_s16(shift_factor);
+
+  for (i = 0; i < height; i += 8) {
+    const uint8x16_t vec_row1 = vld1q_u8(ref);
+    const uint8x16_t vec_row2 = vld1q_u8(ref + ref_stride);
+    const uint8x16_t vec_row3 = vld1q_u8(ref + ref_stride * 2);
+    const uint8x16_t vec_row4 = vld1q_u8(ref + ref_stride * 3);
+    const uint8x16_t vec_row5 = vld1q_u8(ref + ref_stride * 4);
+    const uint8x16_t vec_row6 = vld1q_u8(ref + ref_stride * 5);
+    const uint8x16_t vec_row7 = vld1q_u8(ref + ref_stride * 6);
+    const uint8x16_t vec_row8 = vld1q_u8(ref + ref_stride * 7);
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row1));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row1));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row2));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row2));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row3));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row3));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row4));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row4));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row5));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row5));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row6));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row6));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row7));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row7));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row8));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row8));
+
+    ref += ref_stride * 8;
+  }
+
+  vec_sum_lo = vshlq_u16(vec_sum_lo, vec_shift);
+  vec_sum_hi = vshlq_u16(vec_sum_hi, vec_shift);
+
+  vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_lo));
+  hbuf += 8;
+  vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_hi));
+}
+
+int16_t vpx_int_pro_col_neon(uint8_t const *ref, const int width) {
+  int i;
+  uint16x8_t vec_sum = vdupq_n_u16(0);
+
+  for (i = 0; i < width; i += 16) {
+    const uint8x16_t vec_row = vld1q_u8(ref);
+    vec_sum = vaddw_u8(vec_sum, vget_low_u8(vec_row));
+    vec_sum = vaddw_u8(vec_sum, vget_high_u8(vec_row));
+    ref += 16;
+  }
+
+  return horizontal_add_u16x8(vec_sum);
+}
+
+// ref, src = [0, 510] - max diff = 16-bits
+// bwl = {2, 3, 4}, width = {16, 32, 64}
+int vpx_vector_var_neon(int16_t const *ref, int16_t const *src, const int bwl) {
+  int width = 4 << bwl;
+  int32x4_t sse = vdupq_n_s32(0);
+  int16x8_t total = vdupq_n_s16(0);
+
+  assert(width >= 8);
+  assert((width % 8) == 0);
+
+  do {
+    const int16x8_t r = vld1q_s16(ref);
+    const int16x8_t s = vld1q_s16(src);
+    const int16x8_t diff = vsubq_s16(r, s);  // [-510, 510], 10 bits.
+    const int16x4_t diff_lo = vget_low_s16(diff);
+    const int16x4_t diff_hi = vget_high_s16(diff);
+    sse = vmlal_s16(sse, diff_lo, diff_lo);  // dynamic range 26 bits.
+    sse = vmlal_s16(sse, diff_hi, diff_hi);
+    total = vaddq_s16(total, diff);  // dynamic range 16 bits.
+
+    ref += 8;
+    src += 8;
+    width -= 8;
+  } while (width != 0);
+
+  {
+    // Note: 'total''s pairwise addition could be implemented similarly to
+    // horizontal_add_u16x8(), but one less vpaddl with 'total' when paired
+    // with the summation of 'sse' performed better on a Cortex-A15.
+    const int32x4_t t0 = vpaddlq_s16(total);  // cascading summation of 'total'
+    const int32x2_t t1 = vadd_s32(vget_low_s32(t0), vget_high_s32(t0));
+    const int32x2_t t2 = vpadd_s32(t1, t1);
+    const int t = vget_lane_s32(t2, 0);
+    const int64x2_t s0 = vpaddlq_s32(sse);  // cascading summation of 'sse'.
+    const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
+                                  vreinterpret_s32_s64(vget_high_s64(s0)));
+    const int s = vget_lane_s32(s1, 0);
+    const int shift_factor = bwl + 2;
+    return s - ((t * t) >> shift_factor);
+  }
+}
+
+void vpx_minmax_8x8_neon(const uint8_t *a, int a_stride,
+                         const uint8_t *b, int b_stride,
+                         int *min, int *max) {
+  // Load and concatenate.
+  const uint8x16_t a01 = vcombine_u8(vld1_u8(a),
+                                     vld1_u8(a + a_stride));
+  const uint8x16_t a23 = vcombine_u8(vld1_u8(a + 2 * a_stride),
+                                     vld1_u8(a + 3 * a_stride));
+  const uint8x16_t a45 = vcombine_u8(vld1_u8(a + 4 * a_stride),
+                                     vld1_u8(a + 5 * a_stride));
+  const uint8x16_t a67 = vcombine_u8(vld1_u8(a + 6 * a_stride),
+                                     vld1_u8(a + 7 * a_stride));
+
+  const uint8x16_t b01 = vcombine_u8(vld1_u8(b),
+                                     vld1_u8(b + b_stride));
+  const uint8x16_t b23 = vcombine_u8(vld1_u8(b + 2 * b_stride),
+                                     vld1_u8(b + 3 * b_stride));
+  const uint8x16_t b45 = vcombine_u8(vld1_u8(b + 4 * b_stride),
+                                     vld1_u8(b + 5 * b_stride));
+  const uint8x16_t b67 = vcombine_u8(vld1_u8(b + 6 * b_stride),
+                                     vld1_u8(b + 7 * b_stride));
+
+  // Absolute difference.
+  const uint8x16_t ab01_diff = vabdq_u8(a01, b01);
+  const uint8x16_t ab23_diff = vabdq_u8(a23, b23);
+  const uint8x16_t ab45_diff = vabdq_u8(a45, b45);
+  const uint8x16_t ab67_diff = vabdq_u8(a67, b67);
+
+  // Max values between the Q vectors.
+  const uint8x16_t ab0123_max = vmaxq_u8(ab01_diff, ab23_diff);
+  const uint8x16_t ab4567_max = vmaxq_u8(ab45_diff, ab67_diff);
+  const uint8x16_t ab0123_min = vminq_u8(ab01_diff, ab23_diff);
+  const uint8x16_t ab4567_min = vminq_u8(ab45_diff, ab67_diff);
+
+  const uint8x16_t ab07_max = vmaxq_u8(ab0123_max, ab4567_max);
+  const uint8x16_t ab07_min = vminq_u8(ab0123_min, ab4567_min);
+
+  // Split to D and start doing pairwise.
+  uint8x8_t ab_max = vmax_u8(vget_high_u8(ab07_max), vget_low_u8(ab07_max));
+  uint8x8_t ab_min = vmin_u8(vget_high_u8(ab07_min), vget_low_u8(ab07_min));
+
+  // Enough runs of vpmax/min propogate the max/min values to every position.
+  ab_max = vpmax_u8(ab_max, ab_max);
+  ab_min = vpmin_u8(ab_min, ab_min);
+
+  ab_max = vpmax_u8(ab_max, ab_max);
+  ab_min = vpmin_u8(ab_min, ab_min);
+
+  ab_max = vpmax_u8(ab_max, ab_max);
+  ab_min = vpmin_u8(ab_min, ab_min);
+
+  *min = *max = 0;  // Clear high bits
+  // Store directly to avoid costly neon->gpr transfer.
+  vst1_lane_u8((uint8_t *)max, ab_max, 0);
+  vst1_lane_u8((uint8_t *)min, ab_min, 0);
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/bilinear_filter_media.asm b/libvpx/libvpx/vpx_dsp/arm/bilinear_filter_media.asm
new file mode 100644
index 0000000..f3f9754
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/bilinear_filter_media.asm
@@ -0,0 +1,237 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vpx_filter_block2d_bil_first_pass_media|
+    EXPORT  |vpx_filter_block2d_bil_second_pass_media|
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+
+;-------------------------------------
+; r0    unsigned char  *src_ptr,
+; r1    unsigned short *dst_ptr,
+; r2    unsigned int    src_pitch,
+; r3    unsigned int    height,
+; stack unsigned int    width,
+; stack const short    *vpx_filter
+;-------------------------------------
+; The output is transposed stroed in output array to make it easy for second pass filtering.
+|vpx_filter_block2d_bil_first_pass_media| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; vpx_filter address
+    ldr     r4, [sp, #36]                   ; width
+
+    mov     r12, r3                         ; outer-loop counter
+
+    add     r7, r2, r4                      ; preload next row
+    pld     [r0, r7]
+
+    sub     r2, r2, r4                      ; src increment for height loop
+
+    ldr     r5, [r11]                       ; load up filter coefficients
+
+    mov     r3, r3, lsl #1                  ; height*2
+    add     r3, r3, #2                      ; plus 2 to make output buffer 4-bit aligned since height is actually (height+1)
+
+    mov     r11, r1                         ; save dst_ptr for each row
+
+    cmp     r5, #128                        ; if filter coef = 128, then skip the filter
+    beq     bil_null_1st_filter
+
+|bil_height_loop_1st_v6|
+    ldrb    r6, [r0]                        ; load source data
+    ldrb    r7, [r0, #1]
+    ldrb    r8, [r0, #2]
+    mov     lr, r4, lsr #2                  ; 4-in-parellel loop counter
+
+|bil_width_loop_1st_v6|
+    ldrb    r9, [r0, #3]
+    ldrb    r10, [r0, #4]
+
+    pkhbt   r6, r6, r7, lsl #16             ; src[1] | src[0]
+    pkhbt   r7, r7, r8, lsl #16             ; src[2] | src[1]
+
+    smuad   r6, r6, r5                      ; apply the filter
+    pkhbt   r8, r8, r9, lsl #16             ; src[3] | src[2]
+    smuad   r7, r7, r5
+    pkhbt   r9, r9, r10, lsl #16            ; src[4] | src[3]
+
+    smuad   r8, r8, r5
+    smuad   r9, r9, r5
+
+    add     r0, r0, #4
+    subs    lr, lr, #1
+
+    add     r6, r6, #0x40                   ; round_shift_and_clamp
+    add     r7, r7, #0x40
+    usat    r6, #16, r6, asr #7
+    usat    r7, #16, r7, asr #7
+
+    strh    r6, [r1], r3                    ; result is transposed and stored
+
+    add     r8, r8, #0x40                   ; round_shift_and_clamp
+    strh    r7, [r1], r3
+    add     r9, r9, #0x40
+    usat    r8, #16, r8, asr #7
+    usat    r9, #16, r9, asr #7
+
+    strh    r8, [r1], r3                    ; result is transposed and stored
+
+    ldrneb  r6, [r0]                        ; load source data
+    strh    r9, [r1], r3
+
+    ldrneb  r7, [r0, #1]
+    ldrneb  r8, [r0, #2]
+
+    bne     bil_width_loop_1st_v6
+
+    add     r0, r0, r2                      ; move to next input row
+    subs    r12, r12, #1
+
+    add     r9, r2, r4, lsl #1              ; adding back block width
+    pld     [r0, r9]                        ; preload next row
+
+    add     r11, r11, #2                    ; move over to next column
+    mov     r1, r11
+
+    bne     bil_height_loop_1st_v6
+
+    ldmia   sp!, {r4 - r11, pc}
+
+|bil_null_1st_filter|
+|bil_height_loop_null_1st|
+    mov     lr, r4, lsr #2                  ; loop counter
+
+|bil_width_loop_null_1st|
+    ldrb    r6, [r0]                        ; load data
+    ldrb    r7, [r0, #1]
+    ldrb    r8, [r0, #2]
+    ldrb    r9, [r0, #3]
+
+    strh    r6, [r1], r3                    ; store it to immediate buffer
+    add     r0, r0, #4
+    strh    r7, [r1], r3
+    subs    lr, lr, #1
+    strh    r8, [r1], r3
+    strh    r9, [r1], r3
+
+    bne     bil_width_loop_null_1st
+
+    subs    r12, r12, #1
+    add     r0, r0, r2                      ; move to next input line
+    add     r11, r11, #2                    ; move over to next column
+    mov     r1, r11
+
+    bne     bil_height_loop_null_1st
+
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP  ; |vpx_filter_block2d_bil_first_pass_media|
+
+
+;---------------------------------
+; r0    unsigned short *src_ptr,
+; r1    unsigned char  *dst_ptr,
+; r2    int             dst_pitch,
+; r3    unsigned int    height,
+; stack unsigned int    width,
+; stack const short    *vpx_filter
+;---------------------------------
+|vpx_filter_block2d_bil_second_pass_media| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; vpx_filter address
+    ldr     r4, [sp, #36]                   ; width
+
+    ldr     r5, [r11]                       ; load up filter coefficients
+    mov     r12, r4                         ; outer-loop counter = width, since we work on transposed data matrix
+    mov     r11, r1
+
+    cmp     r5, #128                        ; if filter coef = 128, then skip the filter
+    beq     bil_null_2nd_filter
+
+|bil_height_loop_2nd|
+    ldr     r6, [r0]                        ; load the data
+    ldr     r8, [r0, #4]
+    ldrh    r10, [r0, #8]
+    mov     lr, r3, lsr #2                  ; loop counter
+
+|bil_width_loop_2nd|
+    pkhtb   r7, r6, r8                      ; src[1] | src[2]
+    pkhtb   r9, r8, r10                     ; src[3] | src[4]
+
+    smuad   r6, r6, r5                      ; apply filter
+    smuad   r8, r8, r5                      ; apply filter
+
+    subs    lr, lr, #1
+
+    smuadx  r7, r7, r5                      ; apply filter
+    smuadx  r9, r9, r5                      ; apply filter
+
+    add     r0, r0, #8
+
+    add     r6, r6, #0x40                   ; round_shift_and_clamp
+    add     r7, r7, #0x40
+    usat    r6, #8, r6, asr #7
+    usat    r7, #8, r7, asr #7
+    strb    r6, [r1], r2                    ; the result is transposed back and stored
+
+    add     r8, r8, #0x40                   ; round_shift_and_clamp
+    strb    r7, [r1], r2
+    add     r9, r9, #0x40
+    usat    r8, #8, r8, asr #7
+    usat    r9, #8, r9, asr #7
+    strb    r8, [r1], r2                    ; the result is transposed back and stored
+
+    ldrne   r6, [r0]                        ; load data
+    strb    r9, [r1], r2
+    ldrne   r8, [r0, #4]
+    ldrneh  r10, [r0, #8]
+
+    bne     bil_width_loop_2nd
+
+    subs    r12, r12, #1
+    add     r0, r0, #4                      ; update src for next row
+    add     r11, r11, #1
+    mov     r1, r11
+
+    bne     bil_height_loop_2nd
+    ldmia   sp!, {r4 - r11, pc}
+
+|bil_null_2nd_filter|
+|bil_height_loop_null_2nd|
+    mov     lr, r3, lsr #2
+
+|bil_width_loop_null_2nd|
+    ldr     r6, [r0], #4                    ; load data
+    subs    lr, lr, #1
+    ldr     r8, [r0], #4
+
+    strb    r6, [r1], r2                    ; store data
+    mov     r7, r6, lsr #16
+    strb    r7, [r1], r2
+    mov     r9, r8, lsr #16
+    strb    r8, [r1], r2
+    strb    r9, [r1], r2
+
+    bne     bil_width_loop_null_2nd
+
+    subs    r12, r12, #1
+    add     r0, r0, #4
+    add     r11, r11, #1
+    mov     r1, r11
+
+    bne     bil_height_loop_null_2nd
+
+    ldmia   sp!, {r4 - r11, pc}
+    ENDP  ; |vpx_filter_block2d_second_pass_media|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/fwd_txfm_neon.c b/libvpx/libvpx/vpx_dsp/arm/fwd_txfm_neon.c
new file mode 100644
index 0000000..9f9de98
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/fwd_txfm_neon.c
@@ -0,0 +1,220 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+#include "vpx_dsp/txfm_common.h"
+
+void vpx_fdct8x8_neon(const int16_t *input, int16_t *final_output, int stride) {
+  int i;
+  // stage 1
+  int16x8_t input_0 = vshlq_n_s16(vld1q_s16(&input[0 * stride]), 2);
+  int16x8_t input_1 = vshlq_n_s16(vld1q_s16(&input[1 * stride]), 2);
+  int16x8_t input_2 = vshlq_n_s16(vld1q_s16(&input[2 * stride]), 2);
+  int16x8_t input_3 = vshlq_n_s16(vld1q_s16(&input[3 * stride]), 2);
+  int16x8_t input_4 = vshlq_n_s16(vld1q_s16(&input[4 * stride]), 2);
+  int16x8_t input_5 = vshlq_n_s16(vld1q_s16(&input[5 * stride]), 2);
+  int16x8_t input_6 = vshlq_n_s16(vld1q_s16(&input[6 * stride]), 2);
+  int16x8_t input_7 = vshlq_n_s16(vld1q_s16(&input[7 * stride]), 2);
+  for (i = 0; i < 2; ++i) {
+    int16x8_t out_0, out_1, out_2, out_3, out_4, out_5, out_6, out_7;
+    const int16x8_t v_s0 = vaddq_s16(input_0, input_7);
+    const int16x8_t v_s1 = vaddq_s16(input_1, input_6);
+    const int16x8_t v_s2 = vaddq_s16(input_2, input_5);
+    const int16x8_t v_s3 = vaddq_s16(input_3, input_4);
+    const int16x8_t v_s4 = vsubq_s16(input_3, input_4);
+    const int16x8_t v_s5 = vsubq_s16(input_2, input_5);
+    const int16x8_t v_s6 = vsubq_s16(input_1, input_6);
+    const int16x8_t v_s7 = vsubq_s16(input_0, input_7);
+    // fdct4(step, step);
+    int16x8_t v_x0 = vaddq_s16(v_s0, v_s3);
+    int16x8_t v_x1 = vaddq_s16(v_s1, v_s2);
+    int16x8_t v_x2 = vsubq_s16(v_s1, v_s2);
+    int16x8_t v_x3 = vsubq_s16(v_s0, v_s3);
+    // fdct4(step, step);
+    int32x4_t v_t0_lo = vaddl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
+    int32x4_t v_t0_hi = vaddl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
+    int32x4_t v_t1_lo = vsubl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
+    int32x4_t v_t1_hi = vsubl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
+    int32x4_t v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_24_64);
+    int32x4_t v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_24_64);
+    int32x4_t v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_24_64);
+    int32x4_t v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_24_64);
+    v_t2_lo = vmlal_n_s16(v_t2_lo, vget_low_s16(v_x3), (int16_t)cospi_8_64);
+    v_t2_hi = vmlal_n_s16(v_t2_hi, vget_high_s16(v_x3), (int16_t)cospi_8_64);
+    v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x2), (int16_t)cospi_8_64);
+    v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x2), (int16_t)cospi_8_64);
+    v_t0_lo = vmulq_n_s32(v_t0_lo, cospi_16_64);
+    v_t0_hi = vmulq_n_s32(v_t0_hi, cospi_16_64);
+    v_t1_lo = vmulq_n_s32(v_t1_lo, cospi_16_64);
+    v_t1_hi = vmulq_n_s32(v_t1_hi, cospi_16_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
+      const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
+      const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
+      const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
+      out_0 = vcombine_s16(a, c);  // 00 01 02 03 40 41 42 43
+      out_2 = vcombine_s16(e, g);  // 20 21 22 23 60 61 62 63
+      out_4 = vcombine_s16(b, d);  // 04 05 06 07 44 45 46 47
+      out_6 = vcombine_s16(f, h);  // 24 25 26 27 64 65 66 67
+    }
+    // Stage 2
+    v_x0 = vsubq_s16(v_s6, v_s5);
+    v_x1 = vaddq_s16(v_s6, v_s5);
+    v_t0_lo = vmull_n_s16(vget_low_s16(v_x0), (int16_t)cospi_16_64);
+    v_t0_hi = vmull_n_s16(vget_high_s16(v_x0), (int16_t)cospi_16_64);
+    v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_16_64);
+    v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_16_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x8_t ab = vcombine_s16(a, b);
+      const int16x8_t cd = vcombine_s16(c, d);
+      // Stage 3
+      v_x0 = vaddq_s16(v_s4, ab);
+      v_x1 = vsubq_s16(v_s4, ab);
+      v_x2 = vsubq_s16(v_s7, cd);
+      v_x3 = vaddq_s16(v_s7, cd);
+    }
+    // Stage 4
+    v_t0_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_4_64);
+    v_t0_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_4_64);
+    v_t0_lo = vmlal_n_s16(v_t0_lo, vget_low_s16(v_x0), (int16_t)cospi_28_64);
+    v_t0_hi = vmlal_n_s16(v_t0_hi, vget_high_s16(v_x0), (int16_t)cospi_28_64);
+    v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_12_64);
+    v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_12_64);
+    v_t1_lo = vmlal_n_s16(v_t1_lo, vget_low_s16(v_x2), (int16_t)cospi_20_64);
+    v_t1_hi = vmlal_n_s16(v_t1_hi, vget_high_s16(v_x2), (int16_t)cospi_20_64);
+    v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_12_64);
+    v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_12_64);
+    v_t2_lo = vmlsl_n_s16(v_t2_lo, vget_low_s16(v_x1), (int16_t)cospi_20_64);
+    v_t2_hi = vmlsl_n_s16(v_t2_hi, vget_high_s16(v_x1), (int16_t)cospi_20_64);
+    v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_28_64);
+    v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_28_64);
+    v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x0), (int16_t)cospi_4_64);
+    v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x0), (int16_t)cospi_4_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
+      const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
+      const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
+      const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
+      out_1 = vcombine_s16(a, c);  // 10 11 12 13 50 51 52 53
+      out_3 = vcombine_s16(e, g);  // 30 31 32 33 70 71 72 73
+      out_5 = vcombine_s16(b, d);  // 14 15 16 17 54 55 56 57
+      out_7 = vcombine_s16(f, h);  // 34 35 36 37 74 75 76 77
+    }
+    // transpose 8x8
+    {
+      // 00 01 02 03 40 41 42 43
+      // 10 11 12 13 50 51 52 53
+      // 20 21 22 23 60 61 62 63
+      // 30 31 32 33 70 71 72 73
+      // 04 05 06 07 44 45 46 47
+      // 14 15 16 17 54 55 56 57
+      // 24 25 26 27 64 65 66 67
+      // 34 35 36 37 74 75 76 77
+      const int32x4x2_t r02_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_0),
+                                            vreinterpretq_s32_s16(out_2));
+      const int32x4x2_t r13_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_1),
+                                            vreinterpretq_s32_s16(out_3));
+      const int32x4x2_t r46_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_4),
+                                            vreinterpretq_s32_s16(out_6));
+      const int32x4x2_t r57_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_5),
+                                            vreinterpretq_s32_s16(out_7));
+      const int16x8x2_t r01_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[0]),
+                    vreinterpretq_s16_s32(r13_s32.val[0]));
+      const int16x8x2_t r23_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[1]),
+                    vreinterpretq_s16_s32(r13_s32.val[1]));
+      const int16x8x2_t r45_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[0]),
+                    vreinterpretq_s16_s32(r57_s32.val[0]));
+      const int16x8x2_t r67_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[1]),
+                    vreinterpretq_s16_s32(r57_s32.val[1]));
+      input_0 = r01_s16.val[0];
+      input_1 = r01_s16.val[1];
+      input_2 = r23_s16.val[0];
+      input_3 = r23_s16.val[1];
+      input_4 = r45_s16.val[0];
+      input_5 = r45_s16.val[1];
+      input_6 = r67_s16.val[0];
+      input_7 = r67_s16.val[1];
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }  // for
+  {
+    // from vpx_dct_sse2.c
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const int16x8_t sign_in0 = vshrq_n_s16(input_0, 15);
+    const int16x8_t sign_in1 = vshrq_n_s16(input_1, 15);
+    const int16x8_t sign_in2 = vshrq_n_s16(input_2, 15);
+    const int16x8_t sign_in3 = vshrq_n_s16(input_3, 15);
+    const int16x8_t sign_in4 = vshrq_n_s16(input_4, 15);
+    const int16x8_t sign_in5 = vshrq_n_s16(input_5, 15);
+    const int16x8_t sign_in6 = vshrq_n_s16(input_6, 15);
+    const int16x8_t sign_in7 = vshrq_n_s16(input_7, 15);
+    input_0 = vhsubq_s16(input_0, sign_in0);
+    input_1 = vhsubq_s16(input_1, sign_in1);
+    input_2 = vhsubq_s16(input_2, sign_in2);
+    input_3 = vhsubq_s16(input_3, sign_in3);
+    input_4 = vhsubq_s16(input_4, sign_in4);
+    input_5 = vhsubq_s16(input_5, sign_in5);
+    input_6 = vhsubq_s16(input_6, sign_in6);
+    input_7 = vhsubq_s16(input_7, sign_in7);
+    // store results
+    vst1q_s16(&final_output[0 * 8], input_0);
+    vst1q_s16(&final_output[1 * 8], input_1);
+    vst1q_s16(&final_output[2 * 8], input_2);
+    vst1q_s16(&final_output[3 * 8], input_3);
+    vst1q_s16(&final_output[4 * 8], input_4);
+    vst1q_s16(&final_output[5 * 8], input_5);
+    vst1q_s16(&final_output[6 * 8], input_6);
+    vst1q_s16(&final_output[7 * 8], input_7);
+  }
+}
+
+void vpx_fdct8x8_1_neon(const int16_t *input, int16_t *output, int stride) {
+  int r;
+  int16x8_t sum = vld1q_s16(&input[0]);
+  for (r = 1; r < 8; ++r) {
+    const int16x8_t input_00 = vld1q_s16(&input[r * stride]);
+    sum = vaddq_s16(sum, input_00);
+  }
+  {
+    const int32x4_t a = vpaddlq_s16(sum);
+    const int64x2_t b = vpaddlq_s32(a);
+    const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                                 vreinterpret_s32_s64(vget_high_s64(b)));
+    output[0] = vget_lane_s16(vreinterpret_s16_s32(c), 0);
+    output[1] = 0;
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/hadamard_neon.c b/libvpx/libvpx/vpx_dsp/arm/hadamard_neon.c
new file mode 100644
index 0000000..21e3e3d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/hadamard_neon.c
@@ -0,0 +1,201 @@
+/*
+ *  Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+
+static void hadamard8x8_one_pass(int16x8_t *a0, int16x8_t *a1,
+                                 int16x8_t *a2, int16x8_t *a3,
+                                 int16x8_t *a4, int16x8_t *a5,
+                                 int16x8_t *a6, int16x8_t *a7) {
+  const int16x8_t b0 = vaddq_s16(*a0, *a1);
+  const int16x8_t b1 = vsubq_s16(*a0, *a1);
+  const int16x8_t b2 = vaddq_s16(*a2, *a3);
+  const int16x8_t b3 = vsubq_s16(*a2, *a3);
+  const int16x8_t b4 = vaddq_s16(*a4, *a5);
+  const int16x8_t b5 = vsubq_s16(*a4, *a5);
+  const int16x8_t b6 = vaddq_s16(*a6, *a7);
+  const int16x8_t b7 = vsubq_s16(*a6, *a7);
+
+  const int16x8_t c0 = vaddq_s16(b0, b2);
+  const int16x8_t c1 = vaddq_s16(b1, b3);
+  const int16x8_t c2 = vsubq_s16(b0, b2);
+  const int16x8_t c3 = vsubq_s16(b1, b3);
+  const int16x8_t c4 = vaddq_s16(b4, b6);
+  const int16x8_t c5 = vaddq_s16(b5, b7);
+  const int16x8_t c6 = vsubq_s16(b4, b6);
+  const int16x8_t c7 = vsubq_s16(b5, b7);
+
+  *a0 = vaddq_s16(c0, c4);
+  *a1 = vsubq_s16(c2, c6);
+  *a2 = vsubq_s16(c0, c4);
+  *a3 = vaddq_s16(c2, c6);
+  *a4 = vaddq_s16(c3, c7);
+  *a5 = vsubq_s16(c3, c7);
+  *a6 = vsubq_s16(c1, c5);
+  *a7 = vaddq_s16(c1, c5);
+}
+
+// TODO(johannkoenig): Make a transpose library and dedup with idct. Consider
+// reversing transpose order which may make it easier for the compiler to
+// reconcile the vtrn.64 moves.
+static void transpose8x8(int16x8_t *a0, int16x8_t *a1,
+                         int16x8_t *a2, int16x8_t *a3,
+                         int16x8_t *a4, int16x8_t *a5,
+                         int16x8_t *a6, int16x8_t *a7) {
+  // Swap 64 bit elements. Goes from:
+  // a0: 00 01 02 03 04 05 06 07
+  // a1: 08 09 10 11 12 13 14 15
+  // a2: 16 17 18 19 20 21 22 23
+  // a3: 24 25 26 27 28 29 30 31
+  // a4: 32 33 34 35 36 37 38 39
+  // a5: 40 41 42 43 44 45 46 47
+  // a6: 48 49 50 51 52 53 54 55
+  // a7: 56 57 58 59 60 61 62 63
+  // to:
+  // a04_lo: 00 01 02 03 32 33 34 35
+  // a15_lo: 08 09 10 11 40 41 42 43
+  // a26_lo: 16 17 18 19 48 49 50 51
+  // a37_lo: 24 25 26 27 56 57 58 59
+  // a04_hi: 04 05 06 07 36 37 38 39
+  // a15_hi: 12 13 14 15 44 45 46 47
+  // a26_hi: 20 21 22 23 52 53 54 55
+  // a37_hi: 28 29 30 31 60 61 62 63
+  const int16x8_t a04_lo = vcombine_s16(vget_low_s16(*a0), vget_low_s16(*a4));
+  const int16x8_t a15_lo = vcombine_s16(vget_low_s16(*a1), vget_low_s16(*a5));
+  const int16x8_t a26_lo = vcombine_s16(vget_low_s16(*a2), vget_low_s16(*a6));
+  const int16x8_t a37_lo = vcombine_s16(vget_low_s16(*a3), vget_low_s16(*a7));
+  const int16x8_t a04_hi = vcombine_s16(vget_high_s16(*a0), vget_high_s16(*a4));
+  const int16x8_t a15_hi = vcombine_s16(vget_high_s16(*a1), vget_high_s16(*a5));
+  const int16x8_t a26_hi = vcombine_s16(vget_high_s16(*a2), vget_high_s16(*a6));
+  const int16x8_t a37_hi = vcombine_s16(vget_high_s16(*a3), vget_high_s16(*a7));
+
+  // Swap 32 bit elements resulting in:
+  // a0246_lo:
+  // 00 01 16 17 32 33 48 49
+  // 02 03 18 19 34 35 50 51
+  // a1357_lo:
+  // 08 09 24 25 40 41 56 57
+  // 10 11 26 27 42 43 58 59
+  // a0246_hi:
+  // 04 05 20 21 36 37 52 53
+  // 06 07 22 23 38 39 54 55
+  // a1657_hi:
+  // 12 13 28 29 44 45 60 61
+  // 14 15 30 31 46 47 62 63
+  const int32x4x2_t a0246_lo = vtrnq_s32(vreinterpretq_s32_s16(a04_lo),
+                                         vreinterpretq_s32_s16(a26_lo));
+  const int32x4x2_t a1357_lo = vtrnq_s32(vreinterpretq_s32_s16(a15_lo),
+                                         vreinterpretq_s32_s16(a37_lo));
+  const int32x4x2_t a0246_hi = vtrnq_s32(vreinterpretq_s32_s16(a04_hi),
+                                         vreinterpretq_s32_s16(a26_hi));
+  const int32x4x2_t a1357_hi = vtrnq_s32(vreinterpretq_s32_s16(a15_hi),
+                                         vreinterpretq_s32_s16(a37_hi));
+
+  // Swap 16 bit elements resulting in:
+  // b0:
+  // 00 08 16 24 32 40 48 56
+  // 01 09 17 25 33 41 49 57
+  // b1:
+  // 02 10 18 26 34 42 50 58
+  // 03 11 19 27 35 43 51 59
+  // b2:
+  // 04 12 20 28 36 44 52 60
+  // 05 13 21 29 37 45 53 61
+  // b3:
+  // 06 14 22 30 38 46 54 62
+  // 07 15 23 31 39 47 55 63
+  const int16x8x2_t b0 = vtrnq_s16(vreinterpretq_s16_s32(a0246_lo.val[0]),
+                                   vreinterpretq_s16_s32(a1357_lo.val[0]));
+  const int16x8x2_t b1 = vtrnq_s16(vreinterpretq_s16_s32(a0246_lo.val[1]),
+                                   vreinterpretq_s16_s32(a1357_lo.val[1]));
+  const int16x8x2_t b2 = vtrnq_s16(vreinterpretq_s16_s32(a0246_hi.val[0]),
+                                   vreinterpretq_s16_s32(a1357_hi.val[0]));
+  const int16x8x2_t b3 = vtrnq_s16(vreinterpretq_s16_s32(a0246_hi.val[1]),
+                                   vreinterpretq_s16_s32(a1357_hi.val[1]));
+
+  *a0 = b0.val[0];
+  *a1 = b0.val[1];
+  *a2 = b1.val[0];
+  *a3 = b1.val[1];
+  *a4 = b2.val[0];
+  *a5 = b2.val[1];
+  *a6 = b3.val[0];
+  *a7 = b3.val[1];
+}
+
+void vpx_hadamard_8x8_neon(const int16_t *src_diff, int src_stride,
+                           int16_t *coeff) {
+  int16x8_t a0 = vld1q_s16(src_diff);
+  int16x8_t a1 = vld1q_s16(src_diff + src_stride);
+  int16x8_t a2 = vld1q_s16(src_diff + 2 * src_stride);
+  int16x8_t a3 = vld1q_s16(src_diff + 3 * src_stride);
+  int16x8_t a4 = vld1q_s16(src_diff + 4 * src_stride);
+  int16x8_t a5 = vld1q_s16(src_diff + 5 * src_stride);
+  int16x8_t a6 = vld1q_s16(src_diff + 6 * src_stride);
+  int16x8_t a7 = vld1q_s16(src_diff + 7 * src_stride);
+
+  hadamard8x8_one_pass(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+  transpose8x8(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+  hadamard8x8_one_pass(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+  // Skip the second transpose because it is not required.
+
+  vst1q_s16(coeff + 0, a0);
+  vst1q_s16(coeff + 8, a1);
+  vst1q_s16(coeff + 16, a2);
+  vst1q_s16(coeff + 24, a3);
+  vst1q_s16(coeff + 32, a4);
+  vst1q_s16(coeff + 40, a5);
+  vst1q_s16(coeff + 48, a6);
+  vst1q_s16(coeff + 56, a7);
+}
+
+void vpx_hadamard_16x16_neon(const int16_t *src_diff, int src_stride,
+                             int16_t *coeff) {
+  int i;
+
+  /* Rearrange 16x16 to 8x32 and remove stride.
+   * Top left first. */
+  vpx_hadamard_8x8_neon(src_diff + 0 + 0 * src_stride, src_stride, coeff + 0);
+  /* Top right. */
+  vpx_hadamard_8x8_neon(src_diff + 8 + 0 * src_stride, src_stride, coeff + 64);
+  /* Bottom left. */
+  vpx_hadamard_8x8_neon(src_diff + 0 + 8 * src_stride, src_stride, coeff + 128);
+  /* Bottom right. */
+  vpx_hadamard_8x8_neon(src_diff + 8 + 8 * src_stride, src_stride, coeff + 192);
+
+  for (i = 0; i < 64; i += 8) {
+    const int16x8_t a0 = vld1q_s16(coeff + 0);
+    const int16x8_t a1 = vld1q_s16(coeff + 64);
+    const int16x8_t a2 = vld1q_s16(coeff + 128);
+    const int16x8_t a3 = vld1q_s16(coeff + 192);
+
+    const int16x8_t b0 = vhaddq_s16(a0, a1);
+    const int16x8_t b1 = vhsubq_s16(a0, a1);
+    const int16x8_t b2 = vhaddq_s16(a2, a3);
+    const int16x8_t b3 = vhsubq_s16(a2, a3);
+
+    const int16x8_t c0 = vaddq_s16(b0, b2);
+    const int16x8_t c1 = vaddq_s16(b1, b3);
+    const int16x8_t c2 = vsubq_s16(b0, b2);
+    const int16x8_t c3 = vsubq_s16(b1, b3);
+
+    vst1q_s16(coeff + 0, c0);
+    vst1q_s16(coeff + 64, c1);
+    vst1q_s16(coeff + 128, c2);
+    vst1q_s16(coeff + 192, c3);
+
+    coeff += 8;
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct16x16_1_add_neon.asm b/libvpx/libvpx/vpx_dsp/arm/idct16x16_1_add_neon.asm
new file mode 100644
index 0000000..dc459e2
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct16x16_1_add_neon.asm
@@ -0,0 +1,198 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+
+    EXPORT  |vpx_idct16x16_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vpx_idct16x16_1_add_neon(int16_t *input, uint8_t *dest,
+;                                    int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vpx_idct16x16_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 6)
+    add              r0, r0, #32               ; + (1 <<((6) - 1))
+    asr              r0, r0, #6                ; >> 6
+
+    vdup.s16         q0, r0                    ; duplicate a1
+    mov              r0, #8
+    sub              r2, #8
+
+    ; load destination data row0 - row3
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row4 - row7
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row8 - row11
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row12 - row15
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    bx               lr
+    ENDP             ; |vpx_idct16x16_1_add_neon|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct16x16_1_add_neon.c b/libvpx/libvpx/vpx_dsp/arm/idct16x16_1_add_neon.c
new file mode 100644
index 0000000..f734e48
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct16x16_1_add_neon.c
@@ -0,0 +1,61 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "vpx_dsp/inv_txfm.h"
+#include "vpx_ports/mem.h"
+
+void vpx_idct16x16_1_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8x8_t d2u8, d3u8, d30u8, d31u8;
+    uint64x1_t d2u64, d3u64, d4u64, d5u64;
+    uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16;
+    int16x8_t q0s16;
+    uint8_t *d1, *d2;
+    int16_t i, j, a1, cospi_16_64 = 11585;
+    int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+    out = dct_const_round_shift(out * cospi_16_64);
+    a1 = ROUND_POWER_OF_TWO(out, 6);
+
+    q0s16 = vdupq_n_s16(a1);
+    q0u16 = vreinterpretq_u16_s16(q0s16);
+
+    for (d1 = d2 = dest, i = 0; i < 4; i++) {
+        for (j = 0; j < 2; j++) {
+            d2u64 = vld1_u64((const uint64_t *)d1);
+            d3u64 = vld1_u64((const uint64_t *)(d1 + 8));
+            d1 += dest_stride;
+            d4u64 = vld1_u64((const uint64_t *)d1);
+            d5u64 = vld1_u64((const uint64_t *)(d1 + 8));
+            d1 += dest_stride;
+
+            q9u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64));
+            q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64));
+            q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64));
+            q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64));
+
+            d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+            d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+            d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+            d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+
+            vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+            vst1_u64((uint64_t *)(d2 + 8), vreinterpret_u64_u8(d3u8));
+            d2 += dest_stride;
+            vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8));
+            vst1_u64((uint64_t *)(d2 + 8), vreinterpret_u64_u8(d31u8));
+            d2 += dest_stride;
+        }
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct16x16_add_neon.asm b/libvpx/libvpx/vpx_dsp/arm/idct16x16_add_neon.asm
new file mode 100644
index 0000000..22a0c95
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct16x16_add_neon.asm
@@ -0,0 +1,1179 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vpx_idct16x16_256_add_neon_pass1|
+    EXPORT  |vpx_idct16x16_256_add_neon_pass2|
+    EXPORT  |vpx_idct16x16_10_add_neon_pass1|
+    EXPORT  |vpx_idct16x16_10_add_neon_pass2|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Transpose a 8x8 16bit data matrix. Datas are loaded in q8-q15.
+    MACRO
+    TRANSPOSE8X8
+    vswp            d17, d24
+    vswp            d23, d30
+    vswp            d21, d28
+    vswp            d19, d26
+    vtrn.32         q8, q10
+    vtrn.32         q9, q11
+    vtrn.32         q12, q14
+    vtrn.32         q13, q15
+    vtrn.16         q8, q9
+    vtrn.16         q10, q11
+    vtrn.16         q12, q13
+    vtrn.16         q14, q15
+    MEND
+
+    AREA    Block, CODE, READONLY ; name this block of code
+;void |vpx_idct16x16_256_add_neon_pass1|(int16_t *input,
+;                                          int16_t *output, int output_stride)
+;
+; r0  int16_t input
+; r1  int16_t *output
+; r2  int  output_stride)
+
+; idct16 stage1 - stage6 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|vpx_idct16x16_256_add_neon_pass1| PROC
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 0, 2, 4, 6, 8, 10, 12, 14 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q1,q2}, [r0]!
+    vmov.s16        q15, q1
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0xc00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r12, #0x3e00
+    add             r12, #0xc5
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         d0, r3                    ; duplicate cospi_28_64
+    vdup.16         d1, r12                   ; duplicate cospi_4_64
+
+    ; preloading to avoid stall
+    ; generate cospi_12_64 = 13623
+    mov             r3, #0x3500
+    add             r3, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r12, #0x2300
+    add             r12, #0x8e
+
+    ; step2[4] * cospi_28_64
+    vmull.s16       q2, d18, d0
+    vmull.s16       q3, d19, d0
+
+    ; step2[4] * cospi_4_64
+    vmull.s16       q5, d18, d1
+    vmull.s16       q6, d19, d1
+
+    ; temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64
+    vmlsl.s16       q2, d30, d1
+    vmlsl.s16       q3, d31, d1
+
+    ; temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64
+    vmlal.s16       q5, d30, d0
+    vmlal.s16       q6, d31, d0
+
+    vdup.16         d2, r3                    ; duplicate cospi_12_64
+    vdup.16         d3, r12                   ; duplicate cospi_20_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d8, q2, #14               ; >> 14
+    vqrshrn.s32     d9, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d14, q5, #14              ; >> 14
+    vqrshrn.s32     d15, q6, #14              ; >> 14
+
+    ; preloading to avoid stall
+    ; generate cospi_16_64 = 11585
+    mov             r3, #0x2d00
+    add             r3, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r12, #0x1800
+    add             r12, #0x7e
+
+    ; step2[5] * cospi_12_64
+    vmull.s16       q2, d26, d2
+    vmull.s16       q3, d27, d2
+
+    ; step2[5] * cospi_20_64
+    vmull.s16       q9, d26, d3
+    vmull.s16       q15, d27, d3
+
+    ; temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64
+    vmlsl.s16       q2, d22, d3
+    vmlsl.s16       q3, d23, d3
+
+    ; temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64
+    vmlal.s16       q9, d22, d2
+    vmlal.s16       q15, d23, d2
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d10, q2, #14              ; >> 14
+    vqrshrn.s32     d11, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q15, #14             ; >> 14
+
+    ; stage 4
+    vdup.16         d30, r3                   ; cospi_16_64
+
+    ; step1[0] * cospi_16_64
+    vmull.s16       q2, d16, d30
+    vmull.s16       q11, d17, d30
+
+    ; step1[1] * cospi_16_64
+    vmull.s16       q0, d24, d30
+    vmull.s16       q1, d25, d30
+
+    ; generate cospi_8_64 = 15137
+    mov             r3, #0x3b00
+    add             r3, #0x21
+
+    vdup.16         d30, r12                  ; duplicate cospi_24_64
+    vdup.16         d31, r3                   ; duplicate cospi_8_64
+
+    ; temp1 = (step1[0] + step1[1]) * cospi_16_64
+    vadd.s32        q3, q2, q0
+    vadd.s32        q12, q11, q1
+
+    ; temp2 = (step1[0] - step1[1]) * cospi_16_64
+    vsub.s32        q13, q2, q0
+    vsub.s32        q1, q11, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d16, q3, #14              ; >> 14
+    vqrshrn.s32     d17, q12, #14             ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d18, q13, #14             ; >> 14
+    vqrshrn.s32     d19, q1, #14              ; >> 14
+
+    ; step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+    ; step1[2] * cospi_8_64
+    vmull.s16       q0, d20, d31
+    vmull.s16       q1, d21, d31
+
+    ; step1[2] * cospi_24_64
+    vmull.s16       q12, d20, d30
+    vmull.s16       q13, d21, d30
+
+    ; temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64
+    vmlal.s16       q0, d28, d30
+    vmlal.s16       q1, d29, d30
+
+    ; temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64
+    vmlsl.s16       q12, d28, d31
+    vmlsl.s16       q13, d29, d31
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d22, q0, #14              ; >> 14
+    vqrshrn.s32     d23, q1, #14              ; >> 14
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d20, q12, #14             ; >> 14
+    vqrshrn.s32     d21, q13, #14             ; >> 14
+
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5];
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5];
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7];
+    vadd.s16        q15, q6, q7               ; step2[7] = step1[6] + step1[7];
+
+    ; generate cospi_16_64 = 11585
+    mov             r3, #0x2d00
+    add             r3, #0x41
+
+    ; stage 5
+    vadd.s16        q0, q8, q11               ; step1[0] = step2[0] + step2[3];
+    vadd.s16        q1, q9, q10               ; step1[1] = step2[1] + step2[2];
+    vsub.s16        q2, q9, q10               ; step1[2] = step2[1] - step2[2];
+    vsub.s16        q3, q8, q11               ; step1[3] = step2[0] - step2[3];
+
+    vdup.16         d16, r3;                  ; duplicate cospi_16_64
+
+    ; step2[5] * cospi_16_64
+    vmull.s16       q11, d26, d16
+    vmull.s16       q12, d27, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; temp1 = (step2[6] - step2[5]) * cospi_16_64
+    vsub.s32        q6, q9, q11
+    vsub.s32        q13, q10, q12
+
+    ; temp2 = (step2[5] + step2[6]) * cospi_16_64
+    vadd.s32        q9, q9, q11
+    vadd.s32        q10, q10, q12
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d10, q6, #14              ; >> 14
+    vqrshrn.s32     d11, q13, #14             ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q10, #14             ; >> 14
+
+    ; stage 6
+    vadd.s16        q8, q0, q15                ; step2[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; step2[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; step2[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; step2[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; step2[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; step2[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; step2[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q15              ; step2[7] = step1[0] - step1[7];
+
+    ; store the data
+    vst1.64         {d16}, [r1], r2
+    vst1.64         {d17}, [r1], r2
+    vst1.64         {d18}, [r1], r2
+    vst1.64         {d19}, [r1], r2
+    vst1.64         {d20}, [r1], r2
+    vst1.64         {d21}, [r1], r2
+    vst1.64         {d22}, [r1], r2
+    vst1.64         {d23}, [r1], r2
+    vst1.64         {d24}, [r1], r2
+    vst1.64         {d25}, [r1], r2
+    vst1.64         {d26}, [r1], r2
+    vst1.64         {d27}, [r1], r2
+    vst1.64         {d28}, [r1], r2
+    vst1.64         {d29}, [r1], r2
+    vst1.64         {d30}, [r1], r2
+    vst1.64         {d31}, [r1], r2
+
+    bx              lr
+    ENDP  ; |vpx_idct16x16_256_add_neon_pass1|
+
+;void vpx_idct16x16_256_add_neon_pass2(int16_t *src,
+;                                        int16_t *output,
+;                                        int16_t *pass1Output,
+;                                        int16_t skip_adding,
+;                                        uint8_t *dest,
+;                                        int dest_stride)
+;
+; r0  int16_t *src
+; r1  int16_t *output,
+; r2  int16_t *pass1Output,
+; r3  int16_t skip_adding,
+; r4  uint8_t *dest,
+; r5  int dest_stride)
+
+; idct16 stage1 - stage7 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|vpx_idct16x16_256_add_neon_pass2| PROC
+    push            {r3-r9}
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 1, 3, 5, 7, 9, 11, 13, 15 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q0,q1}, [r0]!
+    vmov.s16        q15, q0;
+
+    ; generate  cospi_30_64 = 1606
+    mov             r3, #0x0600
+    add             r3, #0x46
+
+    ; generate cospi_2_64  = 16305
+    mov             r12, #0x3f00
+    add             r12, #0xb1
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         d12, r3                   ; duplicate cospi_30_64
+    vdup.16         d13, r12                  ; duplicate cospi_2_64
+
+    ; preloading to avoid stall
+    ; generate cospi_14_64 = 12665
+    mov             r3, #0x3100
+    add             r3, #0x79
+
+    ; generate cospi_18_64 = 10394
+    mov             r12, #0x2800
+    add             r12, #0x9a
+
+    ; step1[8] * cospi_30_64
+    vmull.s16       q2, d16, d12
+    vmull.s16       q3, d17, d12
+
+    ; step1[8] * cospi_2_64
+    vmull.s16       q1, d16, d13
+    vmull.s16       q4, d17, d13
+
+    ; temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64
+    vmlsl.s16       q2, d30, d13
+    vmlsl.s16       q3, d31, d13
+
+    ; temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64
+    vmlal.s16       q1, d30, d12
+    vmlal.s16       q4, d31, d12
+
+    vdup.16         d30, r3                   ; duplicate cospi_14_64
+    vdup.16         d31, r12                  ; duplicate cospi_18_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d0, q2, #14               ; >> 14
+    vqrshrn.s32     d1, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d14, q1, #14              ; >> 14
+    vqrshrn.s32     d15, q4, #14              ; >> 14
+
+    ; preloading to avoid stall
+    ; generate cospi_22_64 = 7723
+    mov             r3, #0x1e00
+    add             r3, #0x2b
+
+    ; generate cospi_10_64 = 14449
+    mov             r12, #0x3800
+    add             r12, #0x71
+
+    ; step1[9] * cospi_14_64
+    vmull.s16       q2, d24, d30
+    vmull.s16       q3, d25, d30
+
+    ; step1[9] * cospi_18_64
+    vmull.s16       q4, d24, d31
+    vmull.s16       q5, d25, d31
+
+    ; temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64
+    vmlsl.s16       q2, d22, d31
+    vmlsl.s16       q3, d23, d31
+
+    ; temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64
+    vmlal.s16       q4, d22, d30
+    vmlal.s16       q5, d23, d30
+
+    vdup.16         d30, r3                   ; duplicate cospi_22_64
+    vdup.16         d31, r12                  ; duplicate cospi_10_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d2, q2, #14               ; >> 14
+    vqrshrn.s32     d3, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q4, #14              ; >> 14
+    vqrshrn.s32     d13, q5, #14              ; >> 14
+
+    ; step1[10] * cospi_22_64
+    vmull.s16       q11, d20, d30
+    vmull.s16       q12, d21, d30
+
+    ; step1[10] * cospi_10_64
+    vmull.s16       q4, d20, d31
+    vmull.s16       q5, d21, d31
+
+    ; temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64
+    vmlsl.s16       q11, d26, d31
+    vmlsl.s16       q12, d27, d31
+
+    ; temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64
+    vmlal.s16       q4, d26, d30
+    vmlal.s16       q5, d27, d30
+
+    ; preloading to avoid stall
+    ; generate cospi_6_64 = 15679
+    mov             r3, #0x3d00
+    add             r3, #0x3f
+
+    ; generate cospi_26_64 = 4756
+    mov             r12, #0x1200
+    add             r12, #0x94
+
+    vdup.16         d30, r3                   ; duplicate cospi_6_64
+    vdup.16         d31, r12                  ; duplicate cospi_26_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q11, #14              ; >> 14
+    vqrshrn.s32     d5, q12, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d11, q5, #14              ; >> 14
+    vqrshrn.s32     d10, q4, #14              ; >> 14
+
+    ; step1[11] * cospi_6_64
+    vmull.s16       q10, d28, d30
+    vmull.s16       q11, d29, d30
+
+    ; step1[11] * cospi_26_64
+    vmull.s16       q12, d28, d31
+    vmull.s16       q13, d29, d31
+
+    ; temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64
+    vmlsl.s16       q10, d18, d31
+    vmlsl.s16       q11, d19, d31
+
+    ; temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64
+    vmlal.s16       q12, d18, d30
+    vmlal.s16       q13, d19, d30
+
+    vsub.s16        q9, q0, q1                ; step1[9]=step2[8]-step2[9]
+    vadd.s16        q0, q0, q1                ; step1[8]=step2[8]+step2[9]
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d6, q10, #14              ; >> 14
+    vqrshrn.s32     d7, q11, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d8, q12, #14              ; >> 14
+    vqrshrn.s32     d9, q13, #14              ; >> 14
+
+    ; stage 3
+    vsub.s16        q10, q3, q2               ; step1[10]=-step2[10]+step2[11]
+    vadd.s16        q11, q2, q3               ; step1[11]=step2[10]+step2[11]
+    vadd.s16        q12, q4, q5               ; step1[12]=step2[12]+step2[13]
+    vsub.s16        q13, q4, q5               ; step1[13]=step2[12]-step2[13]
+    vsub.s16        q14, q7, q6               ; step1[14]=-step2[14]+tep2[15]
+    vadd.s16        q7, q6, q7                ; step1[15]=step2[14]+step2[15]
+
+    ; stage 4
+    ; generate cospi_24_64 = 6270
+    mov             r3, #0x1800
+    add             r3, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r12, #0x3b00
+    add             r12, #0x21
+
+    ; -step1[9] * cospi_8_64 + step1[14] * cospi_24_64
+    vdup.16         d30, r12                  ; duplicate cospi_8_64
+    vdup.16         d31, r3                   ; duplicate cospi_24_64
+
+    ; step1[9] * cospi_24_64
+    vmull.s16       q2, d18, d31
+    vmull.s16       q3, d19, d31
+
+    ; step1[14] * cospi_24_64
+    vmull.s16       q4, d28, d31
+    vmull.s16       q5, d29, d31
+
+    ; temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64
+    vmlal.s16       q2, d28, d30
+    vmlal.s16       q3, d29, d30
+
+    ; temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64
+    vmlsl.s16       q4, d18, d30
+    vmlsl.s16       q5, d19, d30
+
+    rsb             r12, #0
+    vdup.16         d30, r12                  ; duplicate -cospi_8_64
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q2, #14              ; >> 14
+    vqrshrn.s32     d13, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d2, q4, #14               ; >> 14
+    vqrshrn.s32     d3, q5, #14               ; >> 14
+
+    vmov.s16        q3, q11
+    vmov.s16        q4, q12
+
+    ; - step1[13] * cospi_8_64
+    vmull.s16       q11, d26, d30
+    vmull.s16       q12, d27, d30
+
+    ; -step1[10] * cospi_8_64
+    vmull.s16       q8, d20, d30
+    vmull.s16       q9, d21, d30
+
+    ; temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64
+    vmlsl.s16       q11, d20, d31
+    vmlsl.s16       q12, d21, d31
+
+    ; temp1 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64
+    vmlal.s16       q8, d26, d31
+    vmlal.s16       q9, d27, d31
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d4, q11, #14              ; >> 14
+    vqrshrn.s32     d5, q12, #14              ; >> 14
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d10, q8, #14              ; >> 14
+    vqrshrn.s32     d11, q9, #14              ; >> 14
+
+    ; stage 5
+    vadd.s16        q8, q0, q3                ; step1[8] = step2[8]+step2[11];
+    vadd.s16        q9, q1, q2                ; step1[9] = step2[9]+step2[10];
+    vsub.s16        q10, q1, q2               ; step1[10] = step2[9]-step2[10];
+    vsub.s16        q11, q0, q3               ; step1[11] = step2[8]-step2[11];
+    vsub.s16        q12, q7, q4               ; step1[12] =-step2[12]+step2[15];
+    vsub.s16        q13, q6, q5               ; step1[13] =-step2[13]+step2[14];
+    vadd.s16        q14, q6, q5               ; step1[14] =step2[13]+step2[14];
+    vadd.s16        q15, q7, q4               ; step1[15] =step2[12]+step2[15];
+
+    ; stage 6.
+    ; generate cospi_16_64 = 11585
+    mov             r12, #0x2d00
+    add             r12, #0x41
+
+    vdup.16         d14, r12                  ; duplicate cospi_16_64
+
+    ; step1[13] * cospi_16_64
+    vmull.s16       q3, d26, d14
+    vmull.s16       q4, d27, d14
+
+    ; step1[10] * cospi_16_64
+    vmull.s16       q0, d20, d14
+    vmull.s16       q1, d21, d14
+
+    ; temp1 = (-step1[10] + step1[13]) * cospi_16_64
+    vsub.s32        q5, q3, q0
+    vsub.s32        q6, q4, q1
+
+    ; temp2 = (step1[10] + step1[13]) * cospi_16_64
+    vadd.s32        q10, q3, q0
+    vadd.s32        q4, q4, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q5, #14               ; >> 14
+    vqrshrn.s32     d5, q6, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d10, q10, #14             ; >> 14
+    vqrshrn.s32     d11, q4, #14              ; >> 14
+
+    ; step1[11] * cospi_16_64
+    vmull.s16       q0, d22, d14
+    vmull.s16       q1, d23, d14
+
+    ; step1[12] * cospi_16_64
+    vmull.s16       q13, d24, d14
+    vmull.s16       q6, d25, d14
+
+    ; temp1 = (-step1[11] + step1[12]) * cospi_16_64
+    vsub.s32        q10, q13, q0
+    vsub.s32        q4, q6, q1
+
+    ; temp2 = (step1[11] + step1[12]) * cospi_16_64
+    vadd.s32        q13, q13, q0
+    vadd.s32        q6, q6, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d6, q10, #14              ; >> 14
+    vqrshrn.s32     d7, q4, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d8, q13, #14              ; >> 14
+    vqrshrn.s32     d9, q6, #14               ; >> 14
+
+    mov              r4, #16                  ; pass1Output stride
+    ldr              r3, [sp]                 ; load skip_adding
+    cmp              r3, #0                   ; check if need adding dest data
+    beq              skip_adding_dest
+
+    ldr              r7, [sp, #28]            ; dest used to save element 0-7
+    mov              r9, r7                   ; save dest pointer for later use
+    ldr              r8, [sp, #32]            ; load dest_stride
+
+    ; stage 7
+    ; load the data in pass1
+    vld1.s16        {q0}, [r2], r4            ; load data step2[0]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[1]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[2]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[3]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q0, q15              ; step2[0] + step2[15]
+    vadd.s16        q13, q1, q14              ; step2[1] + step2[14]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vsub.s16        q14, q1, q14              ; step2[1] - step2[14]
+    vsub.s16        q15, q0, q15              ; step2[0] - step2[15]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q10, q5              ; step2[2] + step2[13]
+    vadd.s16        q13, q11, q4              ; step2[3] + step2[12]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vsub.s16        q4, q11, q4               ; step2[3] - step2[12]
+    vsub.s16        q5, q10, q5               ; step2[2] - step2[13]
+    vld1.s16        {q0}, [r2], r4            ; load data step2[4]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[5]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[6]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[7]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q0, q3               ; step2[4] + step2[11]
+    vadd.s16        q13, q1, q2               ; step2[5] + step2[10]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vsub.s16        q2, q1, q2                ; step2[5] - step2[10]
+    vsub.s16        q3, q0, q3                ; step2[4] - step2[11]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q10, q9              ; step2[6] + step2[9]
+    vadd.s16        q13, q11, q8              ; step2[7] + step2[8]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vsub.s16        q8, q11, q8               ; step2[7] - step2[8]
+    vsub.s16        q9, q10, q9               ; step2[6] - step2[9]
+
+    ; store the data  output 8,9,10,11,12,13,14,15
+    vrshr.s16       q8, q8, #6                ; ROUND_POWER_OF_TWO
+    vaddw.u8        q8, q8, d12               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q8                   ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q9, q9, #6
+    vaddw.u8        q9, q9, d13               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q9                   ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q2, q2, #6
+    vaddw.u8        q2, q2, d12               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q2                   ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q3, q3, #6
+    vaddw.u8        q3, q3, d13               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q3                   ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q4, q4, #6
+    vaddw.u8        q4, q4, d12               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q4                   ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q5, q5, #6
+    vaddw.u8        q5, q5, d13               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q5                   ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q14, q14, #6
+    vaddw.u8        q14, q14, d12             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q14                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q15, q15, #6
+    vaddw.u8        q15, q15, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q15                  ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    b               end_idct16x16_pass2
+
+skip_adding_dest
+    ; stage 7
+    ; load the data in pass1
+    mov              r5, #24
+    mov              r3, #8
+
+    vld1.s16        {q0}, [r2], r4            ; load data step2[0]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[1]
+    vadd.s16        q12, q0, q15              ; step2[0] + step2[15]
+    vadd.s16        q13, q1, q14              ; step2[1] + step2[14]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[2]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[3]
+    vst1.64         {d24}, [r1], r3           ; store output[0]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[1]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q5              ; step2[2] + step2[13]
+    vadd.s16        q13, q11, q4              ; step2[3] + step2[12]
+    vsub.s16        q14, q1, q14              ; step2[1] - step2[14]
+    vsub.s16        q15, q0, q15              ; step2[0] - step2[15]
+    vst1.64         {d24}, [r1], r3           ; store output[2]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[3]
+    vst1.64         {d27}, [r1], r5
+    vsub.s16        q4, q11, q4               ; step2[3] - step2[12]
+    vsub.s16        q5, q10, q5               ; step2[2] - step2[13]
+    vld1.s16        {q0}, [r2], r4            ; load data step2[4]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[5]
+    vadd.s16        q12, q0, q3               ; step2[4] + step2[11]
+    vadd.s16        q13, q1, q2               ; step2[5] + step2[10]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[6]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[7]
+    vst1.64         {d24}, [r1], r3           ; store output[4]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[5]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q9              ; step2[6] + step2[9]
+    vadd.s16        q13, q11, q8              ; step2[7] + step2[8]
+    vsub.s16        q2, q1, q2                ; step2[5] - step2[10]
+    vsub.s16        q3, q0, q3                ; step2[4] - step2[11]
+    vsub.s16        q8, q11, q8               ; step2[7] - step2[8]
+    vsub.s16        q9, q10, q9               ; step2[6] - step2[9]
+    vst1.64         {d24}, [r1], r3           ; store output[6]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[7]
+    vst1.64         {d27}, [r1], r5
+
+    ; store the data  output 8,9,10,11,12,13,14,15
+    vst1.64         {d16}, [r1], r3
+    vst1.64         {d17}, [r1], r5
+    vst1.64         {d18}, [r1], r3
+    vst1.64         {d19}, [r1], r5
+    vst1.64         {d4}, [r1], r3
+    vst1.64         {d5}, [r1], r5
+    vst1.64         {d6}, [r1], r3
+    vst1.64         {d7}, [r1], r5
+    vst1.64         {d8}, [r1], r3
+    vst1.64         {d9}, [r1], r5
+    vst1.64         {d10}, [r1], r3
+    vst1.64         {d11}, [r1], r5
+    vst1.64         {d28}, [r1], r3
+    vst1.64         {d29}, [r1], r5
+    vst1.64         {d30}, [r1], r3
+    vst1.64         {d31}, [r1], r5
+end_idct16x16_pass2
+    pop             {r3-r9}
+    bx              lr
+    ENDP  ; |vpx_idct16x16_256_add_neon_pass2|
+
+;void |vpx_idct16x16_10_add_neon_pass1|(int16_t *input,
+;                                             int16_t *output, int output_stride)
+;
+; r0  int16_t input
+; r1  int16_t *output
+; r2  int  output_stride)
+
+; idct16 stage1 - stage6 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|vpx_idct16x16_10_add_neon_pass1| PROC
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 0, 2, 4, 6, 8, 10, 12, 14 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q1,q2}, [r0]!
+    vmov.s16        q15, q1
+
+    ; generate  cospi_28_64*2 = 6392
+    mov             r3, #0x1800
+    add             r3, #0xf8
+
+    ; generate cospi_4_64*2  = 32138
+    mov             r12, #0x7d00
+    add             r12, #0x8a
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         q0, r3                    ; duplicate cospi_28_64*2
+    vdup.16         q1, r12                   ; duplicate cospi_4_64*2
+
+    ; The following instructions use vqrdmulh to do the
+    ; dct_const_round_shift(step2[4] * cospi_28_64). vvqrdmulh will multiply,
+    ; double, and return the high 16 bits, effectively giving >> 15. Doubling
+    ; the constant will change this to >> 14.
+    ; dct_const_round_shift(step2[4] * cospi_28_64);
+    vqrdmulh.s16    q4, q9, q0
+
+    ; preloading to avoid stall
+    ; generate cospi_16_64*2 = 23170
+    mov             r3, #0x5a00
+    add             r3, #0x82
+
+    ; dct_const_round_shift(step2[4] * cospi_4_64);
+    vqrdmulh.s16    q7, q9, q1
+
+    ; stage 4
+    vdup.16         q1, r3                    ; cospi_16_64*2
+
+    ; generate cospi_16_64 = 11585
+    mov             r3, #0x2d00
+    add             r3, #0x41
+
+    vdup.16         d4, r3;                   ; duplicate cospi_16_64
+
+    ; dct_const_round_shift(step1[0] * cospi_16_64)
+    vqrdmulh.s16    q8, q8, q1
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d14, d4
+    vmull.s16       q10, d15, d4
+
+    ; step2[5] * cospi_16_64
+    vmull.s16       q12, d9, d4
+    vmull.s16       q11, d8, d4
+
+    ; temp1 = (step2[6] - step2[5]) * cospi_16_64
+    vsub.s32        q15, q10, q12
+    vsub.s32        q6, q9, q11
+
+    ; temp2 = (step2[5] + step2[6]) * cospi_16_64
+    vadd.s32        q9, q9, q11
+    vadd.s32        q10, q10, q12
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d11, q15, #14             ; >> 14
+    vqrshrn.s32     d10, q6, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q10, #14             ; >> 14
+
+    ; stage 6
+    vadd.s16        q2, q8, q7                ; step2[0] = step1[0] + step1[7];
+    vadd.s16        q10, q8, q5               ; step2[2] = step1[2] + step1[5];
+    vadd.s16        q11, q8, q4               ; step2[3] = step1[3] + step1[4];
+    vadd.s16        q9, q8, q6                ; step2[1] = step1[1] + step1[6];
+    vsub.s16        q12, q8, q4               ; step2[4] = step1[3] - step1[4];
+    vsub.s16        q13, q8, q5               ; step2[5] = step1[2] - step1[5];
+    vsub.s16        q14, q8, q6               ; step2[6] = step1[1] - step1[6];
+    vsub.s16        q15, q8, q7               ; step2[7] = step1[0] - step1[7];
+
+    ; store the data
+    vst1.64         {d4}, [r1], r2
+    vst1.64         {d5}, [r1], r2
+    vst1.64         {d18}, [r1], r2
+    vst1.64         {d19}, [r1], r2
+    vst1.64         {d20}, [r1], r2
+    vst1.64         {d21}, [r1], r2
+    vst1.64         {d22}, [r1], r2
+    vst1.64         {d23}, [r1], r2
+    vst1.64         {d24}, [r1], r2
+    vst1.64         {d25}, [r1], r2
+    vst1.64         {d26}, [r1], r2
+    vst1.64         {d27}, [r1], r2
+    vst1.64         {d28}, [r1], r2
+    vst1.64         {d29}, [r1], r2
+    vst1.64         {d30}, [r1], r2
+    vst1.64         {d31}, [r1], r2
+
+    bx              lr
+    ENDP  ; |vpx_idct16x16_10_add_neon_pass1|
+
+;void vpx_idct16x16_10_add_neon_pass2(int16_t *src,
+;                                           int16_t *output,
+;                                           int16_t *pass1Output,
+;                                           int16_t skip_adding,
+;                                           uint8_t *dest,
+;                                           int dest_stride)
+;
+; r0  int16_t *src
+; r1  int16_t *output,
+; r2  int16_t *pass1Output,
+; r3  int16_t skip_adding,
+; r4  uint8_t *dest,
+; r5  int dest_stride)
+
+; idct16 stage1 - stage7 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|vpx_idct16x16_10_add_neon_pass2| PROC
+    push            {r3-r9}
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 1, 3, 5, 7, 9, 11, 13, 15 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q0,q1}, [r0]!
+    vmov.s16        q15, q0;
+
+    ; generate 2*cospi_30_64 = 3212
+    mov             r3, #0xc00
+    add             r3, #0x8c
+
+    ; generate 2*cospi_2_64  = 32610
+    mov             r12, #0x7f00
+    add             r12, #0x62
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         q6, r3                    ; duplicate 2*cospi_30_64
+
+    ; dct_const_round_shift(step1[8] * cospi_30_64)
+    vqrdmulh.s16    q0, q8, q6
+
+    vdup.16         q6, r12                   ; duplicate 2*cospi_2_64
+
+    ; dct_const_round_shift(step1[8] * cospi_2_64)
+    vqrdmulh.s16    q7, q8, q6
+
+    ; preloading to avoid stall
+    ; generate 2*cospi_26_64 = 9512
+    mov             r12, #0x2500
+    add             r12, #0x28
+    rsb             r12, #0
+    vdup.16         q15, r12                  ; duplicate -2*cospi_26_64
+
+    ; generate 2*cospi_6_64 = 31358
+    mov             r3, #0x7a00
+    add             r3, #0x7e
+    vdup.16         q14, r3                   ; duplicate 2*cospi_6_64
+
+    ; dct_const_round_shift(- step1[12] * cospi_26_64)
+    vqrdmulh.s16    q3, q9, q15
+
+    ; dct_const_round_shift(step1[12] * cospi_6_64)
+    vqrdmulh.s16    q4, q9, q14
+
+    ; stage 4
+    ; generate cospi_24_64 = 6270
+    mov             r3, #0x1800
+    add             r3, #0x7e
+    vdup.16         d31, r3                   ; duplicate cospi_24_64
+
+    ; generate cospi_8_64 = 15137
+    mov             r12, #0x3b00
+    add             r12, #0x21
+    vdup.16         d30, r12                  ; duplicate cospi_8_64
+
+    ; step1[14] * cospi_24_64
+    vmull.s16       q12, d14, d31
+    vmull.s16       q5, d15, d31
+
+    ; step1[9] * cospi_24_64
+    vmull.s16       q2, d0, d31
+    vmull.s16       q11, d1, d31
+
+    ; temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64
+    vmlsl.s16       q12, d0, d30
+    vmlsl.s16       q5, d1, d30
+
+    ; temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64
+    vmlal.s16       q2, d14, d30
+    vmlal.s16       q11, d15, d30
+
+    rsb              r12, #0
+    vdup.16          d30, r12                 ; duplicate -cospi_8_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d2, q12, #14              ; >> 14
+    vqrshrn.s32     d3, q5, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q2, #14              ; >> 14
+    vqrshrn.s32     d13, q11, #14             ; >> 14
+
+    ; - step1[13] * cospi_8_64
+    vmull.s16       q10, d8, d30
+    vmull.s16       q13, d9, d30
+
+    ; -step1[10] * cospi_8_64
+    vmull.s16       q8, d6, d30
+    vmull.s16       q9, d7, d30
+
+    ; temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64
+    vmlsl.s16       q10, d6, d31
+    vmlsl.s16       q13, d7, d31
+
+    ; temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64
+    vmlal.s16       q8, d8, d31
+    vmlal.s16       q9, d9, d31
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q10, #14              ; >> 14
+    vqrshrn.s32     d5, q13, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d10, q8, #14              ; >> 14
+    vqrshrn.s32     d11, q9, #14              ; >> 14
+
+    ; stage 5
+    vadd.s16        q8, q0, q3                ; step1[8] = step2[8]+step2[11];
+    vadd.s16        q9, q1, q2                ; step1[9] = step2[9]+step2[10];
+    vsub.s16        q10, q1, q2               ; step1[10] = step2[9]-step2[10];
+    vsub.s16        q11, q0, q3               ; step1[11] = step2[8]-step2[11];
+    vsub.s16        q12, q7, q4               ; step1[12] =-step2[12]+step2[15];
+    vsub.s16        q13, q6, q5               ; step1[13] =-step2[13]+step2[14];
+    vadd.s16        q14, q6, q5               ; step1[14] =step2[13]+step2[14];
+    vadd.s16        q15, q7, q4               ; step1[15] =step2[12]+step2[15];
+
+    ; stage 6.
+    ; generate cospi_16_64 = 11585
+    mov             r12, #0x2d00
+    add             r12, #0x41
+
+    vdup.16         d14, r12                  ; duplicate cospi_16_64
+
+    ; step1[13] * cospi_16_64
+    vmull.s16       q3, d26, d14
+    vmull.s16       q4, d27, d14
+
+    ; step1[10] * cospi_16_64
+    vmull.s16       q0, d20, d14
+    vmull.s16       q1, d21, d14
+
+    ; temp1 = (-step1[10] + step1[13]) * cospi_16_64
+    vsub.s32        q5, q3, q0
+    vsub.s32        q6, q4, q1
+
+    ; temp2 = (step1[10] + step1[13]) * cospi_16_64
+    vadd.s32        q0, q3, q0
+    vadd.s32        q1, q4, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q5, #14               ; >> 14
+    vqrshrn.s32     d5, q6, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d10, q0, #14              ; >> 14
+    vqrshrn.s32     d11, q1, #14              ; >> 14
+
+    ; step1[11] * cospi_16_64
+    vmull.s16       q0, d22, d14
+    vmull.s16       q1, d23, d14
+
+    ; step1[12] * cospi_16_64
+    vmull.s16       q13, d24, d14
+    vmull.s16       q6, d25, d14
+
+    ; temp1 = (-step1[11] + step1[12]) * cospi_16_64
+    vsub.s32        q10, q13, q0
+    vsub.s32        q4, q6, q1
+
+    ; temp2 = (step1[11] + step1[12]) * cospi_16_64
+    vadd.s32        q13, q13, q0
+    vadd.s32        q6, q6, q1
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d6, q10, #14              ; >> 14
+    vqrshrn.s32     d7, q4, #14               ; >> 14
+
+    ; dct_const_round_shift((step1[11] + step1[12]) * cospi_16_64);
+    vqrshrn.s32     d8, q13, #14              ; >> 14
+    vqrshrn.s32     d9, q6, #14               ; >> 14
+
+    mov              r4, #16                  ; pass1Output stride
+    ldr              r3, [sp]                 ; load skip_adding
+
+    ; stage 7
+    ; load the data in pass1
+    mov              r5, #24
+    mov              r3, #8
+
+    vld1.s16        {q0}, [r2], r4            ; load data step2[0]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[1]
+    vadd.s16        q12, q0, q15              ; step2[0] + step2[15]
+    vadd.s16        q13, q1, q14              ; step2[1] + step2[14]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[2]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[3]
+    vst1.64         {d24}, [r1], r3           ; store output[0]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[1]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q5              ; step2[2] + step2[13]
+    vadd.s16        q13, q11, q4              ; step2[3] + step2[12]
+    vsub.s16        q14, q1, q14              ; step2[1] - step2[14]
+    vsub.s16        q15, q0, q15              ; step2[0] - step2[15]
+    vst1.64         {d24}, [r1], r3           ; store output[2]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[3]
+    vst1.64         {d27}, [r1], r5
+    vsub.s16        q4, q11, q4               ; step2[3] - step2[12]
+    vsub.s16        q5, q10, q5               ; step2[2] - step2[13]
+    vld1.s16        {q0}, [r2], r4            ; load data step2[4]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[5]
+    vadd.s16        q12, q0, q3               ; step2[4] + step2[11]
+    vadd.s16        q13, q1, q2               ; step2[5] + step2[10]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[6]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[7]
+    vst1.64         {d24}, [r1], r3           ; store output[4]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[5]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q9              ; step2[6] + step2[9]
+    vadd.s16        q13, q11, q8              ; step2[7] + step2[8]
+    vsub.s16        q2, q1, q2                ; step2[5] - step2[10]
+    vsub.s16        q3, q0, q3                ; step2[4] - step2[11]
+    vsub.s16        q8, q11, q8               ; step2[7] - step2[8]
+    vsub.s16        q9, q10, q9               ; step2[6] - step2[9]
+    vst1.64         {d24}, [r1], r3           ; store output[6]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[7]
+    vst1.64         {d27}, [r1], r5
+
+    ; store the data  output 8,9,10,11,12,13,14,15
+    vst1.64         {d16}, [r1], r3
+    vst1.64         {d17}, [r1], r5
+    vst1.64         {d18}, [r1], r3
+    vst1.64         {d19}, [r1], r5
+    vst1.64         {d4}, [r1], r3
+    vst1.64         {d5}, [r1], r5
+    vst1.64         {d6}, [r1], r3
+    vst1.64         {d7}, [r1], r5
+    vst1.64         {d8}, [r1], r3
+    vst1.64         {d9}, [r1], r5
+    vst1.64         {d10}, [r1], r3
+    vst1.64         {d11}, [r1], r5
+    vst1.64         {d28}, [r1], r3
+    vst1.64         {d29}, [r1], r5
+    vst1.64         {d30}, [r1], r3
+    vst1.64         {d31}, [r1], r5
+end_idct10_16x16_pass2
+    pop             {r3-r9}
+    bx              lr
+    ENDP  ; |vpx_idct16x16_10_add_neon_pass2|
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct16x16_add_neon.c b/libvpx/libvpx/vpx_dsp/arm/idct16x16_add_neon.c
new file mode 100644
index 0000000..651ebb2
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct16x16_add_neon.c
@@ -0,0 +1,1317 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+#include "vpx_dsp/txfm_common.h"
+
+static INLINE void TRANSPOSE8X8(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+    int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    *q8s16  = vcombine_s16(d16s16, d24s16);  // vswp d17, d24
+    *q9s16  = vcombine_s16(d18s16, d26s16);  // vswp d19, d26
+    *q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+    *q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+    *q12s16 = vcombine_s16(d17s16, d25s16);
+    *q13s16 = vcombine_s16(d19s16, d27s16);
+    *q14s16 = vcombine_s16(d21s16, d29s16);
+    *q15s16 = vcombine_s16(d23s16, d31s16);
+
+    q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q8s16),
+                        vreinterpretq_s32_s16(*q10s16));
+    q1x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q9s16),
+                        vreinterpretq_s32_s16(*q11s16));
+    q2x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q12s16),
+                        vreinterpretq_s32_s16(*q14s16));
+    q3x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q13s16),
+                        vreinterpretq_s32_s16(*q15s16));
+
+    q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                        vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+    q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                        vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+    q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                        vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+    q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                        vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+    *q8s16  = q0x2s16.val[0];
+    *q9s16  = q0x2s16.val[1];
+    *q10s16 = q1x2s16.val[0];
+    *q11s16 = q1x2s16.val[1];
+    *q12s16 = q2x2s16.val[0];
+    *q13s16 = q2x2s16.val[1];
+    *q14s16 = q3x2s16.val[0];
+    *q15s16 = q3x2s16.val[1];
+    return;
+}
+
+void vpx_idct16x16_256_add_neon_pass1(
+        int16_t *in,
+        int16_t *out,
+        int output_stride) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    uint64x1_t d16u64, d17u64, d18u64, d19u64, d20u64, d21u64, d22u64, d23u64;
+    uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int32x4_t q0s32, q1s32, q2s32, q3s32, q5s32, q6s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
+    int16x8x2_t q0x2s16;
+
+    q0x2s16 = vld2q_s16(in);
+    q8s16  = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q9s16  = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q10s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q11s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q12s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q13s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q14s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q15s16 = q0x2s16.val[0];
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    d16s16 = vget_low_s16(q8s16);
+    d17s16 = vget_high_s16(q8s16);
+    d18s16 = vget_low_s16(q9s16);
+    d19s16 = vget_high_s16(q9s16);
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d22s16 = vget_low_s16(q11s16);
+    d23s16 = vget_high_s16(q11s16);
+    d24s16 = vget_low_s16(q12s16);
+    d25s16 = vget_high_s16(q12s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+    d30s16 = vget_low_s16(q15s16);
+    d31s16 = vget_high_s16(q15s16);
+
+    // stage 3
+    d0s16 = vdup_n_s16(cospi_28_64);
+    d1s16 = vdup_n_s16(cospi_4_64);
+
+    q2s32 = vmull_s16(d18s16, d0s16);
+    q3s32 = vmull_s16(d19s16, d0s16);
+    q5s32 = vmull_s16(d18s16, d1s16);
+    q6s32 = vmull_s16(d19s16, d1s16);
+
+    q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
+    q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
+    q5s32 = vmlal_s16(q5s32, d30s16, d0s16);
+    q6s32 = vmlal_s16(q6s32, d31s16, d0s16);
+
+    d2s16 = vdup_n_s16(cospi_12_64);
+    d3s16 = vdup_n_s16(cospi_20_64);
+
+    d8s16 = vqrshrn_n_s32(q2s32, 14);
+    d9s16 = vqrshrn_n_s32(q3s32, 14);
+    d14s16 = vqrshrn_n_s32(q5s32, 14);
+    d15s16 = vqrshrn_n_s32(q6s32, 14);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+    q7s16 = vcombine_s16(d14s16, d15s16);
+
+    q2s32 = vmull_s16(d26s16, d2s16);
+    q3s32 = vmull_s16(d27s16, d2s16);
+    q9s32 = vmull_s16(d26s16, d3s16);
+    q15s32 = vmull_s16(d27s16, d3s16);
+
+    q2s32 = vmlsl_s16(q2s32, d22s16, d3s16);
+    q3s32 = vmlsl_s16(q3s32, d23s16, d3s16);
+    q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
+    q15s32 = vmlal_s16(q15s32, d23s16, d2s16);
+
+    d10s16 = vqrshrn_n_s32(q2s32, 14);
+    d11s16 = vqrshrn_n_s32(q3s32, 14);
+    d12s16 = vqrshrn_n_s32(q9s32, 14);
+    d13s16 = vqrshrn_n_s32(q15s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    // stage 4
+    d30s16 = vdup_n_s16(cospi_16_64);
+
+    q2s32 = vmull_s16(d16s16, d30s16);
+    q11s32 = vmull_s16(d17s16, d30s16);
+    q0s32 = vmull_s16(d24s16, d30s16);
+    q1s32 = vmull_s16(d25s16, d30s16);
+
+    d30s16 = vdup_n_s16(cospi_24_64);
+    d31s16 = vdup_n_s16(cospi_8_64);
+
+    q3s32 = vaddq_s32(q2s32, q0s32);
+    q12s32 = vaddq_s32(q11s32, q1s32);
+    q13s32 = vsubq_s32(q2s32, q0s32);
+    q1s32 = vsubq_s32(q11s32, q1s32);
+
+    d16s16 = vqrshrn_n_s32(q3s32, 14);
+    d17s16 = vqrshrn_n_s32(q12s32, 14);
+    d18s16 = vqrshrn_n_s32(q13s32, 14);
+    d19s16 = vqrshrn_n_s32(q1s32, 14);
+    q8s16 = vcombine_s16(d16s16, d17s16);
+    q9s16 = vcombine_s16(d18s16, d19s16);
+
+    q0s32 = vmull_s16(d20s16, d31s16);
+    q1s32 = vmull_s16(d21s16, d31s16);
+    q12s32 = vmull_s16(d20s16, d30s16);
+    q13s32 = vmull_s16(d21s16, d30s16);
+
+    q0s32 = vmlal_s16(q0s32, d28s16, d30s16);
+    q1s32 = vmlal_s16(q1s32, d29s16, d30s16);
+    q12s32 = vmlsl_s16(q12s32, d28s16, d31s16);
+    q13s32 = vmlsl_s16(q13s32, d29s16, d31s16);
+
+    d22s16 = vqrshrn_n_s32(q0s32, 14);
+    d23s16 = vqrshrn_n_s32(q1s32, 14);
+    d20s16 = vqrshrn_n_s32(q12s32, 14);
+    d21s16 = vqrshrn_n_s32(q13s32, 14);
+    q10s16 = vcombine_s16(d20s16, d21s16);
+    q11s16 = vcombine_s16(d22s16, d23s16);
+
+    q13s16 = vsubq_s16(q4s16, q5s16);
+    q4s16 = vaddq_s16(q4s16, q5s16);
+    q14s16 = vsubq_s16(q7s16, q6s16);
+    q15s16 = vaddq_s16(q6s16, q7s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+
+    // stage 5
+    q0s16 = vaddq_s16(q8s16, q11s16);
+    q1s16 = vaddq_s16(q9s16, q10s16);
+    q2s16 = vsubq_s16(q9s16, q10s16);
+    q3s16 = vsubq_s16(q8s16, q11s16);
+
+    d16s16 = vdup_n_s16(cospi_16_64);
+
+    q11s32 = vmull_s16(d26s16, d16s16);
+    q12s32 = vmull_s16(d27s16, d16s16);
+    q9s32 = vmull_s16(d28s16, d16s16);
+    q10s32 = vmull_s16(d29s16, d16s16);
+
+    q6s32 = vsubq_s32(q9s32, q11s32);
+    q13s32 = vsubq_s32(q10s32, q12s32);
+    q9s32 = vaddq_s32(q9s32, q11s32);
+    q10s32 = vaddq_s32(q10s32, q12s32);
+
+    d10s16 = vqrshrn_n_s32(q6s32, 14);
+    d11s16 = vqrshrn_n_s32(q13s32, 14);
+    d12s16 = vqrshrn_n_s32(q9s32, 14);
+    d13s16 = vqrshrn_n_s32(q10s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    // stage 6
+    q8s16 = vaddq_s16(q0s16, q15s16);
+    q9s16 = vaddq_s16(q1s16, q6s16);
+    q10s16 = vaddq_s16(q2s16, q5s16);
+    q11s16 = vaddq_s16(q3s16, q4s16);
+    q12s16 = vsubq_s16(q3s16, q4s16);
+    q13s16 = vsubq_s16(q2s16, q5s16);
+    q14s16 = vsubq_s16(q1s16, q6s16);
+    q15s16 = vsubq_s16(q0s16, q15s16);
+
+    d16u64 = vreinterpret_u64_s16(vget_low_s16(q8s16));
+    d17u64 = vreinterpret_u64_s16(vget_high_s16(q8s16));
+    d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16));
+    d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16));
+    d20u64 = vreinterpret_u64_s16(vget_low_s16(q10s16));
+    d21u64 = vreinterpret_u64_s16(vget_high_s16(q10s16));
+    d22u64 = vreinterpret_u64_s16(vget_low_s16(q11s16));
+    d23u64 = vreinterpret_u64_s16(vget_high_s16(q11s16));
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16));
+    d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16));
+    d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16));
+    d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16));
+
+    // store the data
+    output_stride >>= 1;  // output_stride / 2, out is int16_t
+    vst1_u64((uint64_t *)out, d16u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d17u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d18u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d19u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d20u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d21u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d22u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d23u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d24u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d28u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d29u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d30u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d31u64);
+    return;
+}
+
+void vpx_idct16x16_256_add_neon_pass2(
+        int16_t *src,
+        int16_t *out,
+        int16_t *pass1Output,
+        int16_t skip_adding,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8_t *d;
+    uint8x8_t d12u8, d13u8;
+    int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    uint64x1_t d24u64, d25u64, d26u64, d27u64;
+    int64x1_t d12s64, d13s64;
+    uint16x8_t q2u16, q3u16, q4u16, q5u16, q8u16;
+    uint16x8_t q9u16, q12u16, q13u16, q14u16, q15u16;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q8s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q13s32;
+    int16x8x2_t q0x2s16;
+
+    q0x2s16 = vld2q_s16(src);
+    q8s16  = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q9s16  = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q10s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q11s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q12s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q13s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q14s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q15s16 = q0x2s16.val[0];
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    d16s16 = vget_low_s16(q8s16);
+    d17s16 = vget_high_s16(q8s16);
+    d18s16 = vget_low_s16(q9s16);
+    d19s16 = vget_high_s16(q9s16);
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d22s16 = vget_low_s16(q11s16);
+    d23s16 = vget_high_s16(q11s16);
+    d24s16 = vget_low_s16(q12s16);
+    d25s16 = vget_high_s16(q12s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+    d30s16 = vget_low_s16(q15s16);
+    d31s16 = vget_high_s16(q15s16);
+
+    // stage 3
+    d12s16 = vdup_n_s16(cospi_30_64);
+    d13s16 = vdup_n_s16(cospi_2_64);
+
+    q2s32 = vmull_s16(d16s16, d12s16);
+    q3s32 = vmull_s16(d17s16, d12s16);
+    q1s32 = vmull_s16(d16s16, d13s16);
+    q4s32 = vmull_s16(d17s16, d13s16);
+
+    q2s32 = vmlsl_s16(q2s32, d30s16, d13s16);
+    q3s32 = vmlsl_s16(q3s32, d31s16, d13s16);
+    q1s32 = vmlal_s16(q1s32, d30s16, d12s16);
+    q4s32 = vmlal_s16(q4s32, d31s16, d12s16);
+
+    d0s16 = vqrshrn_n_s32(q2s32, 14);
+    d1s16 = vqrshrn_n_s32(q3s32, 14);
+    d14s16 = vqrshrn_n_s32(q1s32, 14);
+    d15s16 = vqrshrn_n_s32(q4s32, 14);
+    q0s16 = vcombine_s16(d0s16, d1s16);
+    q7s16 = vcombine_s16(d14s16, d15s16);
+
+    d30s16 = vdup_n_s16(cospi_14_64);
+    d31s16 = vdup_n_s16(cospi_18_64);
+
+    q2s32 = vmull_s16(d24s16, d30s16);
+    q3s32 = vmull_s16(d25s16, d30s16);
+    q4s32 = vmull_s16(d24s16, d31s16);
+    q5s32 = vmull_s16(d25s16, d31s16);
+
+    q2s32 = vmlsl_s16(q2s32, d22s16, d31s16);
+    q3s32 = vmlsl_s16(q3s32, d23s16, d31s16);
+    q4s32 = vmlal_s16(q4s32, d22s16, d30s16);
+    q5s32 = vmlal_s16(q5s32, d23s16, d30s16);
+
+    d2s16 = vqrshrn_n_s32(q2s32, 14);
+    d3s16 = vqrshrn_n_s32(q3s32, 14);
+    d12s16 = vqrshrn_n_s32(q4s32, 14);
+    d13s16 = vqrshrn_n_s32(q5s32, 14);
+    q1s16 = vcombine_s16(d2s16, d3s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    d30s16 = vdup_n_s16(cospi_22_64);
+    d31s16 = vdup_n_s16(cospi_10_64);
+
+    q11s32 = vmull_s16(d20s16, d30s16);
+    q12s32 = vmull_s16(d21s16, d30s16);
+    q4s32 = vmull_s16(d20s16, d31s16);
+    q5s32 = vmull_s16(d21s16, d31s16);
+
+    q11s32 = vmlsl_s16(q11s32, d26s16, d31s16);
+    q12s32 = vmlsl_s16(q12s32, d27s16, d31s16);
+    q4s32 = vmlal_s16(q4s32, d26s16, d30s16);
+    q5s32 = vmlal_s16(q5s32, d27s16, d30s16);
+
+    d4s16 = vqrshrn_n_s32(q11s32, 14);
+    d5s16 = vqrshrn_n_s32(q12s32, 14);
+    d11s16 = vqrshrn_n_s32(q5s32, 14);
+    d10s16 = vqrshrn_n_s32(q4s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    d30s16 = vdup_n_s16(cospi_6_64);
+    d31s16 = vdup_n_s16(cospi_26_64);
+
+    q10s32 = vmull_s16(d28s16, d30s16);
+    q11s32 = vmull_s16(d29s16, d30s16);
+    q12s32 = vmull_s16(d28s16, d31s16);
+    q13s32 = vmull_s16(d29s16, d31s16);
+
+    q10s32 = vmlsl_s16(q10s32, d18s16, d31s16);
+    q11s32 = vmlsl_s16(q11s32, d19s16, d31s16);
+    q12s32 = vmlal_s16(q12s32, d18s16, d30s16);
+    q13s32 = vmlal_s16(q13s32, d19s16, d30s16);
+
+    d6s16 = vqrshrn_n_s32(q10s32, 14);
+    d7s16 = vqrshrn_n_s32(q11s32, 14);
+    d8s16 = vqrshrn_n_s32(q12s32, 14);
+    d9s16 = vqrshrn_n_s32(q13s32, 14);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+
+    // stage 3
+    q9s16  = vsubq_s16(q0s16, q1s16);
+    q0s16  = vaddq_s16(q0s16, q1s16);
+    q10s16 = vsubq_s16(q3s16, q2s16);
+    q11s16 = vaddq_s16(q2s16, q3s16);
+    q12s16 = vaddq_s16(q4s16, q5s16);
+    q13s16 = vsubq_s16(q4s16, q5s16);
+    q14s16 = vsubq_s16(q7s16, q6s16);
+    q7s16  = vaddq_s16(q6s16, q7s16);
+
+    // stage 4
+    d18s16 = vget_low_s16(q9s16);
+    d19s16 = vget_high_s16(q9s16);
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+
+    d30s16 = vdup_n_s16(cospi_8_64);
+    d31s16 = vdup_n_s16(cospi_24_64);
+
+    q2s32 = vmull_s16(d18s16, d31s16);
+    q3s32 = vmull_s16(d19s16, d31s16);
+    q4s32 = vmull_s16(d28s16, d31s16);
+    q5s32 = vmull_s16(d29s16, d31s16);
+
+    q2s32 = vmlal_s16(q2s32, d28s16, d30s16);
+    q3s32 = vmlal_s16(q3s32, d29s16, d30s16);
+    q4s32 = vmlsl_s16(q4s32, d18s16, d30s16);
+    q5s32 = vmlsl_s16(q5s32, d19s16, d30s16);
+
+    d12s16 = vqrshrn_n_s32(q2s32, 14);
+    d13s16 = vqrshrn_n_s32(q3s32, 14);
+    d2s16 = vqrshrn_n_s32(q4s32, 14);
+    d3s16 = vqrshrn_n_s32(q5s32, 14);
+    q1s16 = vcombine_s16(d2s16, d3s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    q3s16 = q11s16;
+    q4s16 = q12s16;
+
+    d30s16 = vdup_n_s16(-cospi_8_64);
+    q11s32 = vmull_s16(d26s16, d30s16);
+    q12s32 = vmull_s16(d27s16, d30s16);
+    q8s32 = vmull_s16(d20s16, d30s16);
+    q9s32 = vmull_s16(d21s16, d30s16);
+
+    q11s32 = vmlsl_s16(q11s32, d20s16, d31s16);
+    q12s32 = vmlsl_s16(q12s32, d21s16, d31s16);
+    q8s32 = vmlal_s16(q8s32, d26s16, d31s16);
+    q9s32 = vmlal_s16(q9s32, d27s16, d31s16);
+
+    d4s16 = vqrshrn_n_s32(q11s32, 14);
+    d5s16 = vqrshrn_n_s32(q12s32, 14);
+    d10s16 = vqrshrn_n_s32(q8s32, 14);
+    d11s16 = vqrshrn_n_s32(q9s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    // stage 5
+    q8s16  = vaddq_s16(q0s16, q3s16);
+    q9s16  = vaddq_s16(q1s16, q2s16);
+    q10s16 = vsubq_s16(q1s16, q2s16);
+    q11s16 = vsubq_s16(q0s16, q3s16);
+    q12s16 = vsubq_s16(q7s16, q4s16);
+    q13s16 = vsubq_s16(q6s16, q5s16);
+    q14s16 = vaddq_s16(q6s16, q5s16);
+    q15s16 = vaddq_s16(q7s16, q4s16);
+
+    // stage 6
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d22s16 = vget_low_s16(q11s16);
+    d23s16 = vget_high_s16(q11s16);
+    d24s16 = vget_low_s16(q12s16);
+    d25s16 = vget_high_s16(q12s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+
+    d14s16 = vdup_n_s16(cospi_16_64);
+
+    q3s32 = vmull_s16(d26s16, d14s16);
+    q4s32 = vmull_s16(d27s16, d14s16);
+    q0s32 = vmull_s16(d20s16, d14s16);
+    q1s32 = vmull_s16(d21s16, d14s16);
+
+    q5s32 = vsubq_s32(q3s32, q0s32);
+    q6s32 = vsubq_s32(q4s32, q1s32);
+    q10s32 = vaddq_s32(q3s32, q0s32);
+    q4s32 = vaddq_s32(q4s32, q1s32);
+
+    d4s16 = vqrshrn_n_s32(q5s32, 14);
+    d5s16 = vqrshrn_n_s32(q6s32, 14);
+    d10s16 = vqrshrn_n_s32(q10s32, 14);
+    d11s16 = vqrshrn_n_s32(q4s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    q0s32 = vmull_s16(d22s16, d14s16);
+    q1s32 = vmull_s16(d23s16, d14s16);
+    q13s32 = vmull_s16(d24s16, d14s16);
+    q6s32 = vmull_s16(d25s16, d14s16);
+
+    q10s32 = vsubq_s32(q13s32, q0s32);
+    q4s32 = vsubq_s32(q6s32, q1s32);
+    q13s32 = vaddq_s32(q13s32, q0s32);
+    q6s32 = vaddq_s32(q6s32, q1s32);
+
+    d6s16 = vqrshrn_n_s32(q10s32, 14);
+    d7s16 = vqrshrn_n_s32(q4s32, 14);
+    d8s16 = vqrshrn_n_s32(q13s32, 14);
+    d9s16 = vqrshrn_n_s32(q6s32, 14);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+
+    // stage 7
+    if (skip_adding != 0) {
+        d = dest;
+        // load the data in pass1
+        q0s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q1s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        d13s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+
+        q12s16 = vaddq_s16(q0s16, q15s16);
+        q13s16 = vaddq_s16(q1s16, q14s16);
+        q12s16 = vrshrq_n_s16(q12s16, 6);
+        q13s16 = vrshrq_n_s16(q13s16, 6);
+        q12u16 = vaddw_u8(vreinterpretq_u16_s16(q12s16),
+                          vreinterpret_u8_s64(d12s64));
+        q13u16 = vaddw_u8(vreinterpretq_u16_s16(q13s16),
+                          vreinterpret_u8_s64(d13s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+        d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+        d += dest_stride;
+        q14s16 = vsubq_s16(q1s16, q14s16);
+        q15s16 = vsubq_s16(q0s16, q15s16);
+
+        q10s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q11s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        d13s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q12s16 = vaddq_s16(q10s16, q5s16);
+        q13s16 = vaddq_s16(q11s16, q4s16);
+        q12s16 = vrshrq_n_s16(q12s16, 6);
+        q13s16 = vrshrq_n_s16(q13s16, 6);
+        q12u16 = vaddw_u8(vreinterpretq_u16_s16(q12s16),
+                          vreinterpret_u8_s64(d12s64));
+        q13u16 = vaddw_u8(vreinterpretq_u16_s16(q13s16),
+                          vreinterpret_u8_s64(d13s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+        d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+        d += dest_stride;
+        q4s16 = vsubq_s16(q11s16, q4s16);
+        q5s16 = vsubq_s16(q10s16, q5s16);
+
+        q0s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q1s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        d13s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q12s16 = vaddq_s16(q0s16, q3s16);
+        q13s16 = vaddq_s16(q1s16, q2s16);
+        q12s16 = vrshrq_n_s16(q12s16, 6);
+        q13s16 = vrshrq_n_s16(q13s16, 6);
+        q12u16 = vaddw_u8(vreinterpretq_u16_s16(q12s16),
+                          vreinterpret_u8_s64(d12s64));
+        q13u16 = vaddw_u8(vreinterpretq_u16_s16(q13s16),
+                          vreinterpret_u8_s64(d13s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+        d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+        d += dest_stride;
+        q2s16 = vsubq_s16(q1s16, q2s16);
+        q3s16 = vsubq_s16(q0s16, q3s16);
+
+        q10s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q11s16 = vld1q_s16(pass1Output);
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        d13s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q12s16 = vaddq_s16(q10s16, q9s16);
+        q13s16 = vaddq_s16(q11s16, q8s16);
+        q12s16 = vrshrq_n_s16(q12s16, 6);
+        q13s16 = vrshrq_n_s16(q13s16, 6);
+        q12u16 = vaddw_u8(vreinterpretq_u16_s16(q12s16),
+                          vreinterpret_u8_s64(d12s64));
+        q13u16 = vaddw_u8(vreinterpretq_u16_s16(q13s16),
+                          vreinterpret_u8_s64(d13s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+        d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+        d += dest_stride;
+        q8s16 = vsubq_s16(q11s16, q8s16);
+        q9s16 = vsubq_s16(q10s16, q9s16);
+
+        // store the data  out 8,9,10,11,12,13,14,15
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q8s16 = vrshrq_n_s16(q8s16, 6);
+        q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                         vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q9s16 = vrshrq_n_s16(q9s16, 6);
+        q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                          vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q2s16 = vrshrq_n_s16(q2s16, 6);
+        q2u16 = vaddw_u8(vreinterpretq_u16_s16(q2s16),
+                          vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q2u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q3s16 = vrshrq_n_s16(q3s16, 6);
+        q3u16 = vaddw_u8(vreinterpretq_u16_s16(q3s16),
+                         vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q3u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q4s16 = vrshrq_n_s16(q4s16, 6);
+        q4u16 = vaddw_u8(vreinterpretq_u16_s16(q4s16),
+                         vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q4u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q5s16 = vrshrq_n_s16(q5s16, 6);
+        q5u16 = vaddw_u8(vreinterpretq_u16_s16(q5s16),
+                         vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q5u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q14s16 = vrshrq_n_s16(q14s16, 6);
+        q14u16 = vaddw_u8(vreinterpretq_u16_s16(q14s16),
+                          vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q14u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        q15s16 = vrshrq_n_s16(q15s16, 6);
+        q15u16 = vaddw_u8(vreinterpretq_u16_s16(q15s16),
+                          vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q15u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    } else {  // skip_adding_dest
+        q0s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q1s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q12s16 = vaddq_s16(q0s16, q15s16);
+        q13s16 = vaddq_s16(q1s16, q14s16);
+        d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+        d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+        d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+        d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+        vst1_u64((uint64_t *)out, d24u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d25u64);
+        out += 12;
+        vst1_u64((uint64_t *)out, d26u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d27u64);
+        out += 12;
+        q14s16 = vsubq_s16(q1s16, q14s16);
+        q15s16 = vsubq_s16(q0s16, q15s16);
+
+        q10s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q11s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q12s16 = vaddq_s16(q10s16, q5s16);
+        q13s16 = vaddq_s16(q11s16, q4s16);
+        d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+        d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+        d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+        d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+        vst1_u64((uint64_t *)out, d24u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d25u64);
+        out += 12;
+        vst1_u64((uint64_t *)out, d26u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d27u64);
+        out += 12;
+        q4s16 = vsubq_s16(q11s16, q4s16);
+        q5s16 = vsubq_s16(q10s16, q5s16);
+
+        q0s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q1s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q12s16 = vaddq_s16(q0s16, q3s16);
+        q13s16 = vaddq_s16(q1s16, q2s16);
+        d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+        d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+        d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+        d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+        vst1_u64((uint64_t *)out, d24u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d25u64);
+        out += 12;
+        vst1_u64((uint64_t *)out, d26u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d27u64);
+        out += 12;
+        q2s16 = vsubq_s16(q1s16, q2s16);
+        q3s16 = vsubq_s16(q0s16, q3s16);
+
+        q10s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q11s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q12s16 = vaddq_s16(q10s16, q9s16);
+        q13s16 = vaddq_s16(q11s16, q8s16);
+        d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+        d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+        d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+        d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+        vst1_u64((uint64_t *)out, d24u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d25u64);
+        out += 12;
+        vst1_u64((uint64_t *)out, d26u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d27u64);
+        out += 12;
+        q8s16 = vsubq_s16(q11s16, q8s16);
+        q9s16 = vsubq_s16(q10s16, q9s16);
+
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q8s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q8s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q9s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q9s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q2s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q2s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q3s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q3s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q4s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q4s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q5s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q5s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q14s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q14s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q15s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q15s16)));
+    }
+    return;
+}
+
+void vpx_idct16x16_10_add_neon_pass1(
+        int16_t *in,
+        int16_t *out,
+        int output_stride) {
+    int16x4_t d4s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    uint64x1_t d4u64, d5u64, d18u64, d19u64, d20u64, d21u64, d22u64, d23u64;
+    uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64;
+    int16x8_t q0s16, q1s16, q2s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int32x4_t q6s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q15s32;
+    int16x8x2_t q0x2s16;
+
+    q0x2s16 = vld2q_s16(in);
+    q8s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q9s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q10s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q11s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q12s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q13s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q14s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q15s16 = q0x2s16.val[0];
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    // stage 3
+    q0s16 = vdupq_n_s16(cospi_28_64 * 2);
+    q1s16 = vdupq_n_s16(cospi_4_64 * 2);
+
+    q4s16 = vqrdmulhq_s16(q9s16, q0s16);
+    q7s16 = vqrdmulhq_s16(q9s16, q1s16);
+
+    // stage 4
+    q1s16 = vdupq_n_s16(cospi_16_64 * 2);
+    d4s16 = vdup_n_s16(cospi_16_64);
+
+    q8s16 = vqrdmulhq_s16(q8s16, q1s16);
+
+    d8s16 = vget_low_s16(q4s16);
+    d9s16 = vget_high_s16(q4s16);
+    d14s16 = vget_low_s16(q7s16);
+    d15s16 = vget_high_s16(q7s16);
+    q9s32  = vmull_s16(d14s16, d4s16);
+    q10s32 = vmull_s16(d15s16, d4s16);
+    q12s32 = vmull_s16(d9s16, d4s16);
+    q11s32 = vmull_s16(d8s16, d4s16);
+
+    q15s32 = vsubq_s32(q10s32, q12s32);
+    q6s32 = vsubq_s32(q9s32, q11s32);
+    q9s32 = vaddq_s32(q9s32, q11s32);
+    q10s32 = vaddq_s32(q10s32, q12s32);
+
+    d11s16 = vqrshrn_n_s32(q15s32, 14);
+    d10s16 = vqrshrn_n_s32(q6s32, 14);
+    d12s16 = vqrshrn_n_s32(q9s32, 14);
+    d13s16 = vqrshrn_n_s32(q10s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    // stage 6
+    q2s16 = vaddq_s16(q8s16, q7s16);
+    q9s16 = vaddq_s16(q8s16, q6s16);
+    q10s16 = vaddq_s16(q8s16, q5s16);
+    q11s16 = vaddq_s16(q8s16, q4s16);
+    q12s16 = vsubq_s16(q8s16, q4s16);
+    q13s16 = vsubq_s16(q8s16, q5s16);
+    q14s16 = vsubq_s16(q8s16, q6s16);
+    q15s16 = vsubq_s16(q8s16, q7s16);
+
+    d4u64 = vreinterpret_u64_s16(vget_low_s16(q2s16));
+    d5u64 = vreinterpret_u64_s16(vget_high_s16(q2s16));
+    d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16));
+    d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16));
+    d20u64 = vreinterpret_u64_s16(vget_low_s16(q10s16));
+    d21u64 = vreinterpret_u64_s16(vget_high_s16(q10s16));
+    d22u64 = vreinterpret_u64_s16(vget_low_s16(q11s16));
+    d23u64 = vreinterpret_u64_s16(vget_high_s16(q11s16));
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16));
+    d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16));
+    d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16));
+    d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16));
+
+    // store the data
+    output_stride >>= 1;  // output_stride / 2, out is int16_t
+    vst1_u64((uint64_t *)out, d4u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d5u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d18u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d19u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d20u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d21u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d22u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d23u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d24u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d28u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d29u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d30u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d31u64);
+    return;
+}
+
+void vpx_idct16x16_10_add_neon_pass2(
+        int16_t *src,
+        int16_t *out,
+        int16_t *pass1Output,
+        int16_t skip_adding,
+        uint8_t *dest,
+        int dest_stride) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d30s16, d31s16;
+    uint64x1_t d4u64, d5u64, d6u64, d7u64, d8u64, d9u64, d10u64, d11u64;
+    uint64x1_t d16u64, d17u64, d18u64, d19u64;
+    uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q8s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q13s32;
+    int16x8x2_t q0x2s16;
+    (void)skip_adding;
+    (void)dest;
+    (void)dest_stride;
+
+    q0x2s16 = vld2q_s16(src);
+    q8s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q9s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q10s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q11s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q12s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q13s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q14s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q15s16 = q0x2s16.val[0];
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    // stage 3
+    q6s16 = vdupq_n_s16(cospi_30_64 * 2);
+    q0s16 = vqrdmulhq_s16(q8s16, q6s16);
+    q6s16 = vdupq_n_s16(cospi_2_64 * 2);
+    q7s16 = vqrdmulhq_s16(q8s16, q6s16);
+
+    q15s16 = vdupq_n_s16(-cospi_26_64 * 2);
+    q14s16 = vdupq_n_s16(cospi_6_64 * 2);
+    q3s16 = vqrdmulhq_s16(q9s16, q15s16);
+    q4s16 = vqrdmulhq_s16(q9s16, q14s16);
+
+    // stage 4
+    d0s16 = vget_low_s16(q0s16);
+    d1s16 = vget_high_s16(q0s16);
+    d6s16 = vget_low_s16(q3s16);
+    d7s16 = vget_high_s16(q3s16);
+    d8s16 = vget_low_s16(q4s16);
+    d9s16 = vget_high_s16(q4s16);
+    d14s16 = vget_low_s16(q7s16);
+    d15s16 = vget_high_s16(q7s16);
+
+    d30s16 = vdup_n_s16(cospi_8_64);
+    d31s16 = vdup_n_s16(cospi_24_64);
+
+    q12s32 = vmull_s16(d14s16, d31s16);
+    q5s32 = vmull_s16(d15s16, d31s16);
+    q2s32 = vmull_s16(d0s16, d31s16);
+    q11s32 = vmull_s16(d1s16, d31s16);
+
+    q12s32 = vmlsl_s16(q12s32, d0s16, d30s16);
+    q5s32 = vmlsl_s16(q5s32, d1s16, d30s16);
+    q2s32 = vmlal_s16(q2s32, d14s16, d30s16);
+    q11s32 = vmlal_s16(q11s32, d15s16, d30s16);
+
+    d2s16 = vqrshrn_n_s32(q12s32, 14);
+    d3s16 = vqrshrn_n_s32(q5s32, 14);
+    d12s16 = vqrshrn_n_s32(q2s32, 14);
+    d13s16 = vqrshrn_n_s32(q11s32, 14);
+    q1s16 = vcombine_s16(d2s16, d3s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    d30s16 = vdup_n_s16(-cospi_8_64);
+    q10s32 = vmull_s16(d8s16, d30s16);
+    q13s32 = vmull_s16(d9s16, d30s16);
+    q8s32 = vmull_s16(d6s16, d30s16);
+    q9s32 = vmull_s16(d7s16, d30s16);
+
+    q10s32 = vmlsl_s16(q10s32, d6s16, d31s16);
+    q13s32 = vmlsl_s16(q13s32, d7s16, d31s16);
+    q8s32 = vmlal_s16(q8s32, d8s16, d31s16);
+    q9s32 = vmlal_s16(q9s32, d9s16, d31s16);
+
+    d4s16 = vqrshrn_n_s32(q10s32, 14);
+    d5s16 = vqrshrn_n_s32(q13s32, 14);
+    d10s16 = vqrshrn_n_s32(q8s32, 14);
+    d11s16 = vqrshrn_n_s32(q9s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    // stage 5
+    q8s16  = vaddq_s16(q0s16, q3s16);
+    q9s16  = vaddq_s16(q1s16, q2s16);
+    q10s16 = vsubq_s16(q1s16, q2s16);
+    q11s16 = vsubq_s16(q0s16, q3s16);
+    q12s16 = vsubq_s16(q7s16, q4s16);
+    q13s16 = vsubq_s16(q6s16, q5s16);
+    q14s16 = vaddq_s16(q6s16, q5s16);
+    q15s16 = vaddq_s16(q7s16, q4s16);
+
+    // stage 6
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d22s16 = vget_low_s16(q11s16);
+    d23s16 = vget_high_s16(q11s16);
+    d24s16 = vget_low_s16(q12s16);
+    d25s16 = vget_high_s16(q12s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+
+    d14s16 = vdup_n_s16(cospi_16_64);
+    q3s32 = vmull_s16(d26s16, d14s16);
+    q4s32 = vmull_s16(d27s16, d14s16);
+    q0s32 = vmull_s16(d20s16, d14s16);
+    q1s32 = vmull_s16(d21s16, d14s16);
+
+    q5s32 = vsubq_s32(q3s32, q0s32);
+    q6s32 = vsubq_s32(q4s32, q1s32);
+    q0s32 = vaddq_s32(q3s32, q0s32);
+    q4s32 = vaddq_s32(q4s32, q1s32);
+
+    d4s16 = vqrshrn_n_s32(q5s32, 14);
+    d5s16 = vqrshrn_n_s32(q6s32, 14);
+    d10s16 = vqrshrn_n_s32(q0s32, 14);
+    d11s16 = vqrshrn_n_s32(q4s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    q0s32 = vmull_s16(d22s16, d14s16);
+    q1s32 = vmull_s16(d23s16, d14s16);
+    q13s32 = vmull_s16(d24s16, d14s16);
+    q6s32 = vmull_s16(d25s16, d14s16);
+
+    q10s32 = vsubq_s32(q13s32, q0s32);
+    q4s32 = vsubq_s32(q6s32, q1s32);
+    q13s32 = vaddq_s32(q13s32, q0s32);
+    q6s32 = vaddq_s32(q6s32, q1s32);
+
+    d6s16 = vqrshrn_n_s32(q10s32, 14);
+    d7s16 = vqrshrn_n_s32(q4s32, 14);
+    d8s16 = vqrshrn_n_s32(q13s32, 14);
+    d9s16 = vqrshrn_n_s32(q6s32, 14);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+
+    // stage 7
+    q0s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q1s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q12s16 = vaddq_s16(q0s16, q15s16);
+    q13s16 = vaddq_s16(q1s16, q14s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q14s16 = vsubq_s16(q1s16, q14s16);
+    q15s16 = vsubq_s16(q0s16, q15s16);
+
+    q10s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q11s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q12s16 = vaddq_s16(q10s16, q5s16);
+    q13s16 = vaddq_s16(q11s16, q4s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q4s16 = vsubq_s16(q11s16, q4s16);
+    q5s16 = vsubq_s16(q10s16, q5s16);
+
+    q0s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q1s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q12s16 = vaddq_s16(q0s16, q3s16);
+    q13s16 = vaddq_s16(q1s16, q2s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q2s16 = vsubq_s16(q1s16, q2s16);
+    q3s16 = vsubq_s16(q0s16, q3s16);
+
+    q10s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q11s16 = vld1q_s16(pass1Output);
+    q12s16 = vaddq_s16(q10s16, q9s16);
+    q13s16 = vaddq_s16(q11s16, q8s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q8s16 = vsubq_s16(q11s16, q8s16);
+    q9s16 = vsubq_s16(q10s16, q9s16);
+
+    d4u64  = vreinterpret_u64_s16(vget_low_s16(q2s16));
+    d5u64  = vreinterpret_u64_s16(vget_high_s16(q2s16));
+    d6u64  = vreinterpret_u64_s16(vget_low_s16(q3s16));
+    d7u64  = vreinterpret_u64_s16(vget_high_s16(q3s16));
+    d8u64  = vreinterpret_u64_s16(vget_low_s16(q4s16));
+    d9u64  = vreinterpret_u64_s16(vget_high_s16(q4s16));
+    d10u64 = vreinterpret_u64_s16(vget_low_s16(q5s16));
+    d11u64 = vreinterpret_u64_s16(vget_high_s16(q5s16));
+    d16u64 = vreinterpret_u64_s16(vget_low_s16(q8s16));
+    d17u64 = vreinterpret_u64_s16(vget_high_s16(q8s16));
+    d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16));
+    d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16));
+    d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16));
+    d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16));
+    d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16));
+    d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16));
+
+    vst1_u64((uint64_t *)out, d16u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d17u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d18u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d19u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d4u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d5u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d6u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d7u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d8u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d9u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d10u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d11u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d28u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d29u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d30u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d31u64);
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct16x16_neon.c b/libvpx/libvpx/vpx_dsp/arm/idct16x16_neon.c
new file mode 100644
index 0000000..352979a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct16x16_neon.c
@@ -0,0 +1,185 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/vpx_dsp_common.h"
+
+void vpx_idct16x16_256_add_neon_pass1(const int16_t *input,
+                                      int16_t *output,
+                                      int output_stride);
+void vpx_idct16x16_256_add_neon_pass2(const int16_t *src,
+                                      int16_t *output,
+                                      int16_t *pass1Output,
+                                      int16_t skip_adding,
+                                      uint8_t *dest,
+                                      int dest_stride);
+void vpx_idct16x16_10_add_neon_pass1(const int16_t *input,
+                                     int16_t *output,
+                                     int output_stride);
+void vpx_idct16x16_10_add_neon_pass2(const int16_t *src,
+                                     int16_t *output,
+                                     int16_t *pass1Output,
+                                     int16_t skip_adding,
+                                     uint8_t *dest,
+                                     int dest_stride);
+
+#if HAVE_NEON_ASM
+/* For ARM NEON, d8-d15 are callee-saved registers, and need to be saved. */
+extern void vpx_push_neon(int64_t *store);
+extern void vpx_pop_neon(int64_t *store);
+#endif  // HAVE_NEON_ASM
+
+void vpx_idct16x16_256_add_neon(const int16_t *input,
+                                uint8_t *dest, int dest_stride) {
+#if HAVE_NEON_ASM
+  int64_t store_reg[8];
+#endif
+  int16_t pass1_output[16*16] = {0};
+  int16_t row_idct_output[16*16] = {0};
+
+#if HAVE_NEON_ASM
+  // save d8-d15 register values.
+  vpx_push_neon(store_reg);
+#endif
+
+  /* Parallel idct on the upper 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vpx_idct16x16_256_add_neon_pass1(input, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  vpx_idct16x16_256_add_neon_pass2(input+1,
+                                     row_idct_output,
+                                     pass1_output,
+                                     0,
+                                     dest,
+                                     dest_stride);
+
+  /* Parallel idct on the lower 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vpx_idct16x16_256_add_neon_pass1(input+8*16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  vpx_idct16x16_256_add_neon_pass2(input+8*16+1,
+                                     row_idct_output+8,
+                                     pass1_output,
+                                     0,
+                                     dest,
+                                     dest_stride);
+
+  /* Parallel idct on the left 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vpx_idct16x16_256_add_neon_pass1(row_idct_output, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  vpx_idct16x16_256_add_neon_pass2(row_idct_output+1,
+                                     row_idct_output,
+                                     pass1_output,
+                                     1,
+                                     dest,
+                                     dest_stride);
+
+  /* Parallel idct on the right 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vpx_idct16x16_256_add_neon_pass1(row_idct_output+8*16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  vpx_idct16x16_256_add_neon_pass2(row_idct_output+8*16+1,
+                                     row_idct_output+8,
+                                     pass1_output,
+                                     1,
+                                     dest+8,
+                                     dest_stride);
+
+#if HAVE_NEON_ASM
+  // restore d8-d15 register values.
+  vpx_pop_neon(store_reg);
+#endif
+
+  return;
+}
+
+void vpx_idct16x16_10_add_neon(const int16_t *input,
+                               uint8_t *dest, int dest_stride) {
+#if HAVE_NEON_ASM
+  int64_t store_reg[8];
+#endif
+  int16_t pass1_output[16*16] = {0};
+  int16_t row_idct_output[16*16] = {0};
+
+#if HAVE_NEON_ASM
+  // save d8-d15 register values.
+  vpx_push_neon(store_reg);
+#endif
+
+  /* Parallel idct on the upper 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vpx_idct16x16_10_add_neon_pass1(input, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  vpx_idct16x16_10_add_neon_pass2(input+1,
+                                        row_idct_output,
+                                        pass1_output,
+                                        0,
+                                        dest,
+                                        dest_stride);
+
+  /* Skip Parallel idct on the lower 8 rows as they are all 0s */
+
+  /* Parallel idct on the left 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vpx_idct16x16_256_add_neon_pass1(row_idct_output, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  vpx_idct16x16_256_add_neon_pass2(row_idct_output+1,
+                                     row_idct_output,
+                                     pass1_output,
+                                     1,
+                                     dest,
+                                     dest_stride);
+
+  /* Parallel idct on the right 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vpx_idct16x16_256_add_neon_pass1(row_idct_output+8*16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  vpx_idct16x16_256_add_neon_pass2(row_idct_output+8*16+1,
+                                     row_idct_output+8,
+                                     pass1_output,
+                                     1,
+                                     dest+8,
+                                     dest_stride);
+
+#if HAVE_NEON_ASM
+  // restore d8-d15 register values.
+  vpx_pop_neon(store_reg);
+#endif
+
+  return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct32x32_1_add_neon.asm b/libvpx/libvpx/vpx_dsp/arm/idct32x32_1_add_neon.asm
new file mode 100644
index 0000000..96d276b
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct32x32_1_add_neon.asm
@@ -0,0 +1,144 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+    EXPORT  |vpx_idct32x32_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ;TODO(hkuang): put the following macros in a seperate
+    ;file so other idct function could also use them.
+    MACRO
+    LD_16x8          $src, $stride
+    vld1.8           {q8}, [$src], $stride
+    vld1.8           {q9}, [$src], $stride
+    vld1.8           {q10}, [$src], $stride
+    vld1.8           {q11}, [$src], $stride
+    vld1.8           {q12}, [$src], $stride
+    vld1.8           {q13}, [$src], $stride
+    vld1.8           {q14}, [$src], $stride
+    vld1.8           {q15}, [$src], $stride
+    MEND
+
+    MACRO
+    ADD_DIFF_16x8    $diff
+    vqadd.u8         q8, q8, $diff
+    vqadd.u8         q9, q9, $diff
+    vqadd.u8         q10, q10, $diff
+    vqadd.u8         q11, q11, $diff
+    vqadd.u8         q12, q12, $diff
+    vqadd.u8         q13, q13, $diff
+    vqadd.u8         q14, q14, $diff
+    vqadd.u8         q15, q15, $diff
+    MEND
+
+    MACRO
+    SUB_DIFF_16x8    $diff
+    vqsub.u8         q8, q8, $diff
+    vqsub.u8         q9, q9, $diff
+    vqsub.u8         q10, q10, $diff
+    vqsub.u8         q11, q11, $diff
+    vqsub.u8         q12, q12, $diff
+    vqsub.u8         q13, q13, $diff
+    vqsub.u8         q14, q14, $diff
+    vqsub.u8         q15, q15, $diff
+    MEND
+
+    MACRO
+    ST_16x8          $dst, $stride
+    vst1.8           {q8}, [$dst], $stride
+    vst1.8           {q9}, [$dst], $stride
+    vst1.8           {q10},[$dst], $stride
+    vst1.8           {q11},[$dst], $stride
+    vst1.8           {q12},[$dst], $stride
+    vst1.8           {q13},[$dst], $stride
+    vst1.8           {q14},[$dst], $stride
+    vst1.8           {q15},[$dst], $stride
+    MEND
+
+;void vpx_idct32x32_1_add_neon(int16_t *input, uint8_t *dest,
+;                              int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride
+
+|vpx_idct32x32_1_add_neon| PROC
+    push             {lr}
+    pld              [r1]
+    add              r3, r1, #16               ; r3 dest + 16 for second loop
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 6)
+    add              r0, r0, #32               ; + (1 <<((6) - 1))
+    asrs             r0, r0, #6                ; >> 6
+    bge              diff_positive_32_32
+
+diff_negative_32_32
+    neg              r0, r0
+    usat             r0, #8, r0
+    vdup.u8          q0, r0
+    mov              r0, #4
+
+diff_negative_32_32_loop
+    sub              r0, #1
+    LD_16x8          r1, r2
+    SUB_DIFF_16x8    q0
+    ST_16x8          r12, r2
+
+    LD_16x8          r1, r2
+    SUB_DIFF_16x8    q0
+    ST_16x8          r12, r2
+    cmp              r0, #2
+    moveq            r1, r3
+    moveq            r12, r3
+    cmp              r0, #0
+    bne              diff_negative_32_32_loop
+    pop              {pc}
+
+diff_positive_32_32
+    usat             r0, #8, r0
+    vdup.u8          q0, r0
+    mov              r0, #4
+
+diff_positive_32_32_loop
+    sub              r0, #1
+    LD_16x8          r1, r2
+    ADD_DIFF_16x8    q0
+    ST_16x8          r12, r2
+
+    LD_16x8          r1, r2
+    ADD_DIFF_16x8    q0
+    ST_16x8          r12, r2
+    cmp              r0, #2
+    moveq            r1, r3
+    moveq            r12, r3
+    cmp              r0, #0
+    bne              diff_positive_32_32_loop
+    pop              {pc}
+
+    ENDP             ; |vpx_idct32x32_1_add_neon|
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct32x32_1_add_neon.c b/libvpx/libvpx/vpx_dsp/arm/idct32x32_1_add_neon.c
new file mode 100644
index 0000000..c25c0c4
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct32x32_1_add_neon.c
@@ -0,0 +1,165 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+
+#include "vpx_dsp/inv_txfm.h"
+#include "vpx_ports/mem.h"
+
+static INLINE void LD_16x8(
+        uint8_t *d,
+        int d_stride,
+        uint8x16_t *q8u8,
+        uint8x16_t *q9u8,
+        uint8x16_t *q10u8,
+        uint8x16_t *q11u8,
+        uint8x16_t *q12u8,
+        uint8x16_t *q13u8,
+        uint8x16_t *q14u8,
+        uint8x16_t *q15u8) {
+    *q8u8 = vld1q_u8(d);
+    d += d_stride;
+    *q9u8 = vld1q_u8(d);
+    d += d_stride;
+    *q10u8 = vld1q_u8(d);
+    d += d_stride;
+    *q11u8 = vld1q_u8(d);
+    d += d_stride;
+    *q12u8 = vld1q_u8(d);
+    d += d_stride;
+    *q13u8 = vld1q_u8(d);
+    d += d_stride;
+    *q14u8 = vld1q_u8(d);
+    d += d_stride;
+    *q15u8 = vld1q_u8(d);
+    return;
+}
+
+static INLINE void ADD_DIFF_16x8(
+        uint8x16_t qdiffu8,
+        uint8x16_t *q8u8,
+        uint8x16_t *q9u8,
+        uint8x16_t *q10u8,
+        uint8x16_t *q11u8,
+        uint8x16_t *q12u8,
+        uint8x16_t *q13u8,
+        uint8x16_t *q14u8,
+        uint8x16_t *q15u8) {
+    *q8u8 = vqaddq_u8(*q8u8, qdiffu8);
+    *q9u8 = vqaddq_u8(*q9u8, qdiffu8);
+    *q10u8 = vqaddq_u8(*q10u8, qdiffu8);
+    *q11u8 = vqaddq_u8(*q11u8, qdiffu8);
+    *q12u8 = vqaddq_u8(*q12u8, qdiffu8);
+    *q13u8 = vqaddq_u8(*q13u8, qdiffu8);
+    *q14u8 = vqaddq_u8(*q14u8, qdiffu8);
+    *q15u8 = vqaddq_u8(*q15u8, qdiffu8);
+    return;
+}
+
+static INLINE void SUB_DIFF_16x8(
+        uint8x16_t qdiffu8,
+        uint8x16_t *q8u8,
+        uint8x16_t *q9u8,
+        uint8x16_t *q10u8,
+        uint8x16_t *q11u8,
+        uint8x16_t *q12u8,
+        uint8x16_t *q13u8,
+        uint8x16_t *q14u8,
+        uint8x16_t *q15u8) {
+    *q8u8 = vqsubq_u8(*q8u8, qdiffu8);
+    *q9u8 = vqsubq_u8(*q9u8, qdiffu8);
+    *q10u8 = vqsubq_u8(*q10u8, qdiffu8);
+    *q11u8 = vqsubq_u8(*q11u8, qdiffu8);
+    *q12u8 = vqsubq_u8(*q12u8, qdiffu8);
+    *q13u8 = vqsubq_u8(*q13u8, qdiffu8);
+    *q14u8 = vqsubq_u8(*q14u8, qdiffu8);
+    *q15u8 = vqsubq_u8(*q15u8, qdiffu8);
+    return;
+}
+
+static INLINE void ST_16x8(
+        uint8_t *d,
+        int d_stride,
+        uint8x16_t *q8u8,
+        uint8x16_t *q9u8,
+        uint8x16_t *q10u8,
+        uint8x16_t *q11u8,
+        uint8x16_t *q12u8,
+        uint8x16_t *q13u8,
+        uint8x16_t *q14u8,
+        uint8x16_t *q15u8) {
+    vst1q_u8(d, *q8u8);
+    d += d_stride;
+    vst1q_u8(d, *q9u8);
+    d += d_stride;
+    vst1q_u8(d, *q10u8);
+    d += d_stride;
+    vst1q_u8(d, *q11u8);
+    d += d_stride;
+    vst1q_u8(d, *q12u8);
+    d += d_stride;
+    vst1q_u8(d, *q13u8);
+    d += d_stride;
+    vst1q_u8(d, *q14u8);
+    d += d_stride;
+    vst1q_u8(d, *q15u8);
+    return;
+}
+
+void vpx_idct32x32_1_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8x16_t q0u8, q8u8, q9u8, q10u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+    int i, j, dest_stride8;
+    uint8_t *d;
+    int16_t a1, cospi_16_64 = 11585;
+    int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+
+    out = dct_const_round_shift(out * cospi_16_64);
+    a1 = ROUND_POWER_OF_TWO(out, 6);
+
+    dest_stride8 = dest_stride * 8;
+    if (a1 >= 0) {  // diff_positive_32_32
+        a1 = a1 < 0 ? 0 : a1 > 255 ? 255 : a1;
+        q0u8 = vdupq_n_u8(a1);
+        for (i = 0; i < 2; i++, dest += 16) {  // diff_positive_32_32_loop
+            d = dest;
+            for (j = 0; j < 4; j++) {
+                LD_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8,
+                                        &q12u8, &q13u8, &q14u8, &q15u8);
+                ADD_DIFF_16x8(q0u8, &q8u8, &q9u8, &q10u8, &q11u8,
+                                    &q12u8, &q13u8, &q14u8, &q15u8);
+                ST_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8,
+                                        &q12u8, &q13u8, &q14u8, &q15u8);
+                d += dest_stride8;
+            }
+        }
+    } else {  // diff_negative_32_32
+        a1 = -a1;
+        a1 = a1 < 0 ? 0 : a1 > 255 ? 255 : a1;
+        q0u8 = vdupq_n_u8(a1);
+        for (i = 0; i < 2; i++, dest += 16) {  // diff_negative_32_32_loop
+            d = dest;
+            for (j = 0; j < 4; j++) {
+                LD_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8,
+                                        &q12u8, &q13u8, &q14u8, &q15u8);
+                SUB_DIFF_16x8(q0u8, &q8u8, &q9u8, &q10u8, &q11u8,
+                                    &q12u8, &q13u8, &q14u8, &q15u8);
+                ST_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8,
+                                        &q12u8, &q13u8, &q14u8, &q15u8);
+                d += dest_stride8;
+            }
+        }
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct32x32_add_neon.asm b/libvpx/libvpx/vpx_dsp/arm/idct32x32_add_neon.asm
new file mode 100644
index 0000000..7483ee7
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct32x32_add_neon.asm
@@ -0,0 +1,1299 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+;TODO(cd): adjust these constant to be able to use vqdmulh for faster
+;          dct_const_round_shift(a * b) within butterfly calculations.
+cospi_1_64  EQU 16364
+cospi_2_64  EQU 16305
+cospi_3_64  EQU 16207
+cospi_4_64  EQU 16069
+cospi_5_64  EQU 15893
+cospi_6_64  EQU 15679
+cospi_7_64  EQU 15426
+cospi_8_64  EQU 15137
+cospi_9_64  EQU 14811
+cospi_10_64 EQU 14449
+cospi_11_64 EQU 14053
+cospi_12_64 EQU 13623
+cospi_13_64 EQU 13160
+cospi_14_64 EQU 12665
+cospi_15_64 EQU 12140
+cospi_16_64 EQU 11585
+cospi_17_64 EQU 11003
+cospi_18_64 EQU 10394
+cospi_19_64 EQU  9760
+cospi_20_64 EQU  9102
+cospi_21_64 EQU  8423
+cospi_22_64 EQU  7723
+cospi_23_64 EQU  7005
+cospi_24_64 EQU  6270
+cospi_25_64 EQU  5520
+cospi_26_64 EQU  4756
+cospi_27_64 EQU  3981
+cospi_28_64 EQU  3196
+cospi_29_64 EQU  2404
+cospi_30_64 EQU  1606
+cospi_31_64 EQU   804
+
+
+    EXPORT  |vpx_idct32x32_1024_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    AREA     Block, CODE, READONLY
+
+    ; --------------------------------------------------------------------------
+    ; Load from transposed_buffer
+    ;   q13 = transposed_buffer[first_offset]
+    ;   q14 = transposed_buffer[second_offset]
+    ;   for proper address calculation, the last offset used when manipulating
+    ;   transposed_buffer must be passed in. use 0 for first use.
+    MACRO
+    LOAD_FROM_TRANSPOSED $prev_offset, $first_offset, $second_offset
+    ; address calculation with proper stride and loading
+    add r0, #($first_offset  - $prev_offset )*8*2
+    vld1.s16        {q14}, [r0]
+    add r0, #($second_offset - $first_offset)*8*2
+    vld1.s16        {q13}, [r0]
+    ; (used) two registers (q14, q13)
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Load from output (used as temporary storage)
+    ;   reg1 = output[first_offset]
+    ;   reg2 = output[second_offset]
+    ;   for proper address calculation, the last offset used when manipulating
+    ;   output, whether reading or storing) must be passed in. use 0 for first
+    ;   use.
+    MACRO
+    LOAD_FROM_OUTPUT $prev_offset, $first_offset, $second_offset, $reg1, $reg2
+    ; address calculation with proper stride and loading
+    add r1, #($first_offset  - $prev_offset )*32*2
+    vld1.s16        {$reg1}, [r1]
+    add r1, #($second_offset - $first_offset)*32*2
+    vld1.s16        {$reg2}, [r1]
+    ; (used) two registers ($reg1, $reg2)
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Store into output (sometimes as as temporary storage)
+    ;   output[first_offset] = reg1
+    ;   output[second_offset] = reg2
+    ;   for proper address calculation, the last offset used when manipulating
+    ;   output, whether reading or storing) must be passed in. use 0 for first
+    ;   use.
+    MACRO
+    STORE_IN_OUTPUT $prev_offset, $first_offset, $second_offset, $reg1, $reg2
+    ; address calculation with proper stride and storing
+    add r1, #($first_offset  - $prev_offset )*32*2
+    vst1.16 {$reg1}, [r1]
+    add r1, #($second_offset - $first_offset)*32*2
+    vst1.16 {$reg2}, [r1]
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q6-q9 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_CENTER_RESULTS
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d8}, [r10], r2
+    vld1.s16        {d11}, [r9], r11
+    vld1.s16        {d9}, [r10]
+    vld1.s16        {d10}, [r9]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q8, q8, #6
+    vrshr.s16       q9, q9, #6
+    vrshr.s16       q6, q6, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q7, q7, d9
+    vaddw.u8        q8, q8, d10
+    vaddw.u8        q9, q9, d11
+    vaddw.u8        q6, q6, d8
+    ; clip pixel
+    vqmovun.s16     d9,  q7
+    vqmovun.s16     d10, q8
+    vqmovun.s16     d11, q9
+    vqmovun.s16     d8,  q6
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d9}, [r10], r11
+    vst1.16         {d10}, [r9], r2
+    vst1.16         {d8}, [r10]
+    vst1.16         {d11}, [r9]
+    ; update pointers (by dest_stride * 2)
+    sub r9,  r9,  r2, lsl #1
+    add r10, r10, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q6-q9 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_CENTER_RESULTS_LAST
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d8}, [r10], r2
+    vld1.s16        {d11}, [r9], r11
+    vld1.s16        {d9}, [r10]
+    vld1.s16        {d10}, [r9]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q8, q8, #6
+    vrshr.s16       q9, q9, #6
+    vrshr.s16       q6, q6, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q7, q7, d9
+    vaddw.u8        q8, q8, d10
+    vaddw.u8        q9, q9, d11
+    vaddw.u8        q6, q6, d8
+    ; clip pixel
+    vqmovun.s16     d9,  q7
+    vqmovun.s16     d10, q8
+    vqmovun.s16     d11, q9
+    vqmovun.s16     d8,  q6
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d9}, [r10], r11
+    vst1.16         {d10}, [r9], r2
+    vst1.16         {d8}, [r10]!
+    vst1.16         {d11}, [r9]!
+    ; update pointers (by dest_stride * 2)
+    sub r9,  r9,  r2, lsl #1
+    add r10, r10, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q4-q7 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_EXTREME_RESULTS
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d4}, [r7], r2
+    vld1.s16        {d7}, [r6], r11
+    vld1.s16        {d5}, [r7]
+    vld1.s16        {d6}, [r6]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q5, q5, #6
+    vrshr.s16       q6, q6, #6
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q4, q4, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q5, q5, d5
+    vaddw.u8        q6, q6, d6
+    vaddw.u8        q7, q7, d7
+    vaddw.u8        q4, q4, d4
+    ; clip pixel
+    vqmovun.s16     d5, q5
+    vqmovun.s16     d6, q6
+    vqmovun.s16     d7, q7
+    vqmovun.s16     d4, q4
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d5}, [r7], r11
+    vst1.16         {d6}, [r6], r2
+    vst1.16         {d7}, [r6]
+    vst1.16         {d4}, [r7]
+    ; update pointers (by dest_stride * 2)
+    sub r6, r6, r2, lsl #1
+    add r7, r7, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q4-q7 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_EXTREME_RESULTS_LAST
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d4}, [r7], r2
+    vld1.s16        {d7}, [r6], r11
+    vld1.s16        {d5}, [r7]
+    vld1.s16        {d6}, [r6]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q5, q5, #6
+    vrshr.s16       q6, q6, #6
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q4, q4, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q5, q5, d5
+    vaddw.u8        q6, q6, d6
+    vaddw.u8        q7, q7, d7
+    vaddw.u8        q4, q4, d4
+    ; clip pixel
+    vqmovun.s16     d5, q5
+    vqmovun.s16     d6, q6
+    vqmovun.s16     d7, q7
+    vqmovun.s16     d4, q4
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d5}, [r7], r11
+    vst1.16         {d6}, [r6], r2
+    vst1.16         {d7}, [r6]!
+    vst1.16         {d4}, [r7]!
+    ; update pointers (by dest_stride * 2)
+    sub r6, r6, r2, lsl #1
+    add r7, r7, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Touches q8-q12, q15 (q13-q14 are preserved)
+    ; valid output registers are anything but q8-q11
+    MACRO
+    DO_BUTTERFLY $regC, $regD, $regA, $regB, $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4
+    ; TODO(cd): have special case to re-use constants when they are similar for
+    ;           consecutive butterflies
+    ; TODO(cd): have special case when both constants are the same, do the
+    ;           additions/subtractions before the multiplies.
+    ; generate the constants
+    ;   generate scalar constants
+    mov             r8,  #$first_constant  & 0xFF00
+    mov             r12, #$second_constant & 0xFF00
+    add             r8,  #$first_constant  & 0x00FF
+    add             r12, #$second_constant & 0x00FF
+    ;   generate vector constants
+    vdup.16         d30, r8
+    vdup.16         d31, r12
+    ; (used) two for inputs (regA-regD), one for constants (q15)
+    ; do some multiplications (ordered for maximum latency hiding)
+    vmull.s16 q8,  $regC, d30
+    vmull.s16 q10, $regA, d31
+    vmull.s16 q9,  $regD, d30
+    vmull.s16 q11, $regB, d31
+    vmull.s16 q12, $regC, d31
+    ; (used) five for intermediate (q8-q12), one for constants (q15)
+    ; do some addition/subtractions (to get back two register)
+    vsub.s32  q8, q8, q10
+    vsub.s32  q9, q9, q11
+    ; do more multiplications (ordered for maximum latency hiding)
+    vmull.s16 q10, $regD, d31
+    vmull.s16 q11, $regA, d30
+    vmull.s16 q15, $regB, d30
+    ; (used) six for intermediate (q8-q12, q15)
+    ; do more addition/subtractions
+    vadd.s32  q11, q12, q11
+    vadd.s32  q10, q10, q15
+    ; (used) four for intermediate (q8-q11)
+    ; dct_const_round_shift
+    vqrshrn.s32 $reg1, q8,  #14
+    vqrshrn.s32 $reg2, q9,  #14
+    vqrshrn.s32 $reg3, q11, #14
+    vqrshrn.s32 $reg4, q10, #14
+    ; (used) two for results, well four d registers
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Touches q8-q12, q15 (q13-q14 are preserved)
+    ; valid output registers are anything but q8-q11
+    MACRO
+    DO_BUTTERFLY_STD $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4
+    DO_BUTTERFLY d28, d29, d26, d27, $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4
+    MEND
+    ; --------------------------------------------------------------------------
+
+;void vpx_idct32x32_1024_add_neon(int16_t *input, uint8_t *dest, int dest_stride);
+;
+;   r0  int16_t *input,
+;   r1  uint8_t *dest,
+;   r2  int dest_stride)
+; loop counters
+;   r4  bands loop counter
+;   r5  pass loop counter
+;   r8  transpose loop counter
+; combine-add pointers
+;   r6  dest + 31 * dest_stride, descending (30, 29, 28, ...)
+;   r7  dest +  0 * dest_stride, ascending  (1, 2, 3, ...)
+;   r9  dest + 15 * dest_stride, descending (14, 13, 12, ...)
+;   r10 dest + 16 * dest_stride, ascending  (17, 18, 19, ...)
+
+|vpx_idct32x32_1024_add_neon| PROC
+    ; This function does one pass of idct32x32 transform.
+    ;
+    ; This is done by transposing the input and then doing a 1d transform on
+    ; columns. In the first pass, the transposed columns are the original
+    ; rows. In the second pass, after the transposition, the colums are the
+    ; original columns.
+    ; The 1d transform is done by looping over bands of eight columns (the
+    ; idct32_bands loop). For each band, the transform input transposition
+    ; is done on demand, one band of four 8x8 matrices at a time. The four
+    ; matrices are transposed by pairs (the idct32_transpose_pair loop).
+    push  {r4-r11}
+    vpush {d8-d15}
+    ; stack operation
+    ; internal buffer used to transpose 8 lines into before transforming them
+    ;   int16_t transpose_buffer[32 * 8];
+    ;   at sp + [4096, 4607]
+    ; results of the first pass (transpose and transform rows)
+    ;   int16_t pass1[32 * 32];
+    ;   at sp + [0, 2047]
+    ; results of the second pass (transpose and transform columns)
+    ;   int16_t pass2[32 * 32];
+    ;   at sp + [2048, 4095]
+    sub sp, sp, #512+2048+2048
+
+    ; r6  = dest + 31 * dest_stride
+    ; r7  = dest +  0 * dest_stride
+    ; r9  = dest + 15 * dest_stride
+    ; r10 = dest + 16 * dest_stride
+    rsb r6,  r2, r2, lsl #5
+    rsb r9,  r2, r2, lsl #4
+    add r10, r1, r2, lsl #4
+    mov r7, r1
+    add r6, r6, r1
+    add r9, r9, r1
+    ; r11 = -dest_stride
+    neg r11, r2
+    ; r3 = input
+    mov r3, r0
+    ; parameters for first pass
+      ; r0 = transpose_buffer[32 * 8]
+    add r0, sp, #4096
+      ; r1 = pass1[32 * 32]
+    mov r1, sp
+
+    mov r5, #0          ; initialize pass loop counter
+idct32_pass_loop
+    mov r4, #4          ; initialize bands loop counter
+idct32_bands_loop
+    mov r8, #2          ; initialize transpose loop counter
+idct32_transpose_pair_loop
+    ; Load two horizontally consecutive 8x8 16bit data matrices. The first one
+    ; into q0-q7 and the second one into q8-q15. There is a stride of 64,
+    ; adjusted to 32 because of the two post-increments.
+    vld1.s16        {q8},  [r3]!
+    vld1.s16        {q0},  [r3]!
+    add r3, #32
+    vld1.s16        {q9},  [r3]!
+    vld1.s16        {q1},  [r3]!
+    add r3, #32
+    vld1.s16        {q10}, [r3]!
+    vld1.s16        {q2},  [r3]!
+    add r3, #32
+    vld1.s16        {q11}, [r3]!
+    vld1.s16        {q3},  [r3]!
+    add r3, #32
+    vld1.s16        {q12}, [r3]!
+    vld1.s16        {q4},  [r3]!
+    add r3, #32
+    vld1.s16        {q13}, [r3]!
+    vld1.s16        {q5},  [r3]!
+    add r3, #32
+    vld1.s16        {q14}, [r3]!
+    vld1.s16        {q6},  [r3]!
+    add r3, #32
+    vld1.s16        {q15}, [r3]!
+    vld1.s16        {q7},  [r3]!
+
+    ; Transpose the two 8x8 16bit data matrices.
+    vswp            d17, d24
+    vswp            d23, d30
+    vswp            d21, d28
+    vswp            d19, d26
+    vswp            d1,  d8
+    vswp            d7,  d14
+    vswp            d5,  d12
+    vswp            d3,  d10
+    vtrn.32         q8,  q10
+    vtrn.32         q9,  q11
+    vtrn.32         q12, q14
+    vtrn.32         q13, q15
+    vtrn.32         q0,  q2
+    vtrn.32         q1,  q3
+    vtrn.32         q4,  q6
+    vtrn.32         q5,  q7
+    vtrn.16         q8,  q9
+    vtrn.16         q10, q11
+    vtrn.16         q12, q13
+    vtrn.16         q14, q15
+    vtrn.16         q0,  q1
+    vtrn.16         q2,  q3
+    vtrn.16         q4,  q5
+    vtrn.16         q6,  q7
+
+    ; Store both matrices after each other. There is a stride of 32, which
+    ; adjusts to nothing because of the post-increments.
+    vst1.16        {q8},  [r0]!
+    vst1.16        {q9},  [r0]!
+    vst1.16        {q10}, [r0]!
+    vst1.16        {q11}, [r0]!
+    vst1.16        {q12}, [r0]!
+    vst1.16        {q13}, [r0]!
+    vst1.16        {q14}, [r0]!
+    vst1.16        {q15}, [r0]!
+    vst1.16        {q0},  [r0]!
+    vst1.16        {q1},  [r0]!
+    vst1.16        {q2},  [r0]!
+    vst1.16        {q3},  [r0]!
+    vst1.16        {q4},  [r0]!
+    vst1.16        {q5},  [r0]!
+    vst1.16        {q6},  [r0]!
+    vst1.16        {q7},  [r0]!
+
+    ; increment pointers by adjusted stride (not necessary for r0/out)
+    ;   go back by 7*32 for the seven lines moved fully by read and add
+    ;   go back by 32 for the eigth line only read
+    ;   advance by 16*2 to go the next pair
+    sub r3,  r3,  #7*32*2 + 32 - 16*2
+    ; transpose pair loop processing
+    subs r8, r8, #1
+    bne idct32_transpose_pair_loop
+
+    ; restore r0/input to its original value
+    sub r0, r0, #32*8*2
+
+    ; Instead of doing the transforms stage by stage, it is done by loading
+    ; some input values and doing as many stages as possible to minimize the
+    ; storing/loading of intermediate results. To fit within registers, the
+    ; final coefficients are cut into four blocks:
+    ; BLOCK A: 16-19,28-31
+    ; BLOCK B: 20-23,24-27
+    ; BLOCK C: 8-10,11-15
+    ; BLOCK D: 0-3,4-7
+    ; Blocks A and C are straight calculation through the various stages. In
+    ; block B, further calculations are performed using the results from
+    ; block A. In block D, further calculations are performed using the results
+    ; from block C and then the final calculations are done using results from
+    ; block A and B which have been combined at the end of block B.
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK A: 16-19,28-31
+    ; --------------------------------------------------------------------------
+    ; generate 16,17,30,31
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[1 * 32] * cospi_31_64 - input[31 * 32] *  cospi_1_64;
+    ;temp2 = input[1 * 32] *  cospi_1_64 + input[31 * 32] * cospi_31_64;
+    ;step1b[16][i] = dct_const_round_shift(temp1);
+    ;step1b[31][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 0, 1, 31
+    DO_BUTTERFLY_STD cospi_31_64, cospi_1_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[17 * 32] * cospi_15_64 - input[15 * 32] * cospi_17_64;
+    ;temp2 = input[17 * 32] * cospi_17_64 + input[15 * 32] * cospi_15_64;
+    ;step1b[17][i] = dct_const_round_shift(temp1);
+    ;step1b[30][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 31, 17, 15
+    DO_BUTTERFLY_STD cospi_15_64, cospi_17_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[16] =  step1b[16][i] + step1b[17][i];
+    ;step2[17] =  step1b[16][i] - step1b[17][i];
+    ;step2[30] = -step1b[30][i] + step1b[31][i];
+    ;step2[31] =  step1b[30][i] + step1b[31][i];
+    vadd.s16  q4, q0, q1
+    vsub.s16  q13, q0, q1
+    vadd.s16  q6, q2, q3
+    vsub.s16  q14, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[30][i] * cospi_28_64 - step1b[17][i] * cospi_4_64;
+    ;temp2 = step1b[30][i] * cospi_4_64  - step1b[17][i] * cospi_28_64;
+    ;step3[17] = dct_const_round_shift(temp1);
+    ;step3[30] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_28_64, cospi_4_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; generate 18,19,28,29
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[9 * 32] * cospi_23_64 - input[23 * 32] * cospi_9_64;
+    ;temp2 = input[9 * 32] *  cospi_9_64 + input[23 * 32] * cospi_23_64;
+    ;step1b[18][i] = dct_const_round_shift(temp1);
+    ;step1b[29][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 15, 9, 23
+    DO_BUTTERFLY_STD cospi_23_64, cospi_9_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[25 * 32] *  cospi_7_64 - input[7 * 32] * cospi_25_64;
+    ;temp2 = input[25 * 32] * cospi_25_64 + input[7 * 32] * cospi_7_64;
+    ;step1b[19][i] = dct_const_round_shift(temp1);
+    ;step1b[28][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 23, 25, 7
+    DO_BUTTERFLY_STD cospi_7_64, cospi_25_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[18] = -step1b[18][i] + step1b[19][i];
+    ;step2[19] =  step1b[18][i] + step1b[19][i];
+    ;step2[28] =  step1b[28][i] + step1b[29][i];
+    ;step2[29] =  step1b[28][i] - step1b[29][i];
+    vsub.s16  q13, q3, q2
+    vadd.s16  q3,  q3, q2
+    vsub.s16  q14, q1, q0
+    vadd.s16  q2,  q1, q0
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[18][i] * (-cospi_4_64)  - step1b[29][i] * (-cospi_28_64);
+    ;temp2 = step1b[18][i] * (-cospi_28_64) + step1b[29][i] * (-cospi_4_64);
+    ;step3[29] = dct_const_round_shift(temp1);
+    ;step3[18] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_4_64), (-cospi_28_64), d2, d3, d0, d1
+    ; --------------------------------------------------------------------------
+    ; combine 16-19,28-31
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[16] = step1b[16][i] + step1b[19][i];
+    ;step1[17] = step1b[17][i] + step1b[18][i];
+    ;step1[18] = step1b[17][i] - step1b[18][i];
+    ;step1[29] = step1b[30][i] - step1b[29][i];
+    ;step1[30] = step1b[30][i] + step1b[29][i];
+    ;step1[31] = step1b[31][i] + step1b[28][i];
+    vadd.s16  q8,  q4, q2
+    vadd.s16  q9,  q5, q0
+    vadd.s16  q10, q7, q1
+    vadd.s16  q15, q6, q3
+    vsub.s16  q13, q5, q0
+    vsub.s16  q14, q7, q1
+    STORE_IN_OUTPUT 0,  16, 31, q8,  q15
+    STORE_IN_OUTPUT 31, 17, 30, q9,  q10
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[29][i] * cospi_24_64 - step1b[18][i] * cospi_8_64;
+    ;temp2 = step1b[29][i] * cospi_8_64  + step1b[18][i] * cospi_24_64;
+    ;step2[18] = dct_const_round_shift(temp1);
+    ;step2[29] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d0, d1, d2, d3
+    STORE_IN_OUTPUT 30, 29, 18, q1, q0
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[19] = step1b[16][i] - step1b[19][i];
+    ;step1[28] = step1b[31][i] - step1b[28][i];
+    vsub.s16  q13, q4, q2
+    vsub.s16  q14, q6, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[28][i] * cospi_24_64 - step1b[19][i] * cospi_8_64;
+    ;temp2 = step1b[28][i] * cospi_8_64  + step1b[19][i] * cospi_24_64;
+    ;step2[19] = dct_const_round_shift(temp1);
+    ;step2[28] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d8, d9, d12, d13
+    STORE_IN_OUTPUT 18, 19, 28, q4, q6
+    ; --------------------------------------------------------------------------
+
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK B: 20-23,24-27
+    ; --------------------------------------------------------------------------
+    ; generate 20,21,26,27
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[5 * 32] * cospi_27_64 - input[27 * 32] * cospi_5_64;
+    ;temp2 = input[5 * 32] *  cospi_5_64 + input[27 * 32] * cospi_27_64;
+    ;step1b[20][i] = dct_const_round_shift(temp1);
+    ;step1b[27][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 7, 5, 27
+    DO_BUTTERFLY_STD cospi_27_64, cospi_5_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[21 * 32] * cospi_11_64 - input[11 * 32] * cospi_21_64;
+    ;temp2 = input[21 * 32] * cospi_21_64 + input[11 * 32] * cospi_11_64;
+    ;step1b[21][i] = dct_const_round_shift(temp1);
+    ;step1b[26][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 27, 21, 11
+    DO_BUTTERFLY_STD cospi_11_64, cospi_21_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[20] =  step1b[20][i] + step1b[21][i];
+    ;step2[21] =  step1b[20][i] - step1b[21][i];
+    ;step2[26] = -step1b[26][i] + step1b[27][i];
+    ;step2[27] =  step1b[26][i] + step1b[27][i];
+    vsub.s16  q13, q0, q1
+    vadd.s16  q0, q0, q1
+    vsub.s16  q14, q2, q3
+    vadd.s16  q2, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[26][i] * cospi_12_64 - step1b[21][i] * cospi_20_64;
+    ;temp2 = step1b[26][i] * cospi_20_64 + step1b[21][i] * cospi_12_64;
+    ;step3[21] = dct_const_round_shift(temp1);
+    ;step3[26] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_12_64, cospi_20_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; generate 22,23,24,25
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[13 * 32] * cospi_19_64 - input[19 * 32] * cospi_13_64;
+    ;temp2 = input[13 * 32] * cospi_13_64 + input[19 * 32] * cospi_19_64;
+    ;step1b[22][i] = dct_const_round_shift(temp1);
+    ;step1b[25][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 11, 13, 19
+    DO_BUTTERFLY_STD cospi_19_64, cospi_13_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[29 * 32] *  cospi_3_64 - input[3 * 32] * cospi_29_64;
+    ;temp2 = input[29 * 32] * cospi_29_64 + input[3 * 32] * cospi_3_64;
+    ;step1b[23][i] = dct_const_round_shift(temp1);
+    ;step1b[24][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 19, 29, 3
+    DO_BUTTERFLY_STD cospi_3_64, cospi_29_64, d8, d9, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[22] = -step1b[22][i] + step1b[23][i];
+    ;step2[23] =  step1b[22][i] + step1b[23][i];
+    ;step2[24] =  step1b[24][i] + step1b[25][i];
+    ;step2[25] =  step1b[24][i] - step1b[25][i];
+    vsub.s16  q14, q4, q5
+    vadd.s16  q5, q4, q5
+    vsub.s16  q13, q6, q7
+    vadd.s16  q6, q6, q7
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[22][i] * (-cospi_20_64) - step1b[25][i] * (-cospi_12_64);
+    ;temp2 = step1b[22][i] * (-cospi_12_64) + step1b[25][i] * (-cospi_20_64);
+    ;step3[25] = dct_const_round_shift(temp1);
+    ;step3[22] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_20_64), (-cospi_12_64), d8, d9, d14, d15
+    ; --------------------------------------------------------------------------
+    ; combine 20-23,24-27
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[22] = step1b[22][i] + step1b[21][i];
+    ;step1[23] = step1b[23][i] + step1b[20][i];
+    vadd.s16  q10, q7, q1
+    vadd.s16  q11, q5, q0
+    ;step1[24] = step1b[24][i] + step1b[27][i];
+    ;step1[25] = step1b[25][i] + step1b[26][i];
+    vadd.s16  q12, q6, q2
+    vadd.s16  q15, q4, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[16] = step1b[16][i] + step1b[23][i];
+    ;step3[17] = step1b[17][i] + step1b[22][i];
+    ;step3[22] = step1b[17][i] - step1b[22][i];
+    ;step3[23] = step1b[16][i] - step1b[23][i];
+    LOAD_FROM_OUTPUT 28, 16, 17, q14, q13
+    vadd.s16  q8,  q14, q11
+    vadd.s16  q9,  q13, q10
+    vsub.s16  q13, q13, q10
+    vsub.s16  q11, q14, q11
+    STORE_IN_OUTPUT 17, 17, 16, q9, q8
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[24] = step1b[31][i] - step1b[24][i];
+    ;step3[25] = step1b[30][i] - step1b[25][i];
+    ;step3[30] = step1b[30][i] + step1b[25][i];
+    ;step3[31] = step1b[31][i] + step1b[24][i];
+    LOAD_FROM_OUTPUT 16, 30, 31, q14, q9
+    vsub.s16  q8,  q9,  q12
+    vadd.s16  q10, q14, q15
+    vsub.s16  q14, q14, q15
+    vadd.s16  q12, q9,  q12
+    STORE_IN_OUTPUT 31, 30, 31, q10, q12
+    ; --------------------------------------------------------------------------
+    ; TODO(cd) do some register allocation change to remove these push/pop
+    vpush {q8}  ; [24]
+    vpush {q11} ; [23]
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[25][i] - step1b[22][i]) * cospi_16_64;
+    ;temp2 = (step1b[25][i] + step1b[22][i]) * cospi_16_64;
+    ;step1[22] = dct_const_round_shift(temp1);
+    ;step1[25] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29
+    STORE_IN_OUTPUT 31, 25, 22, q14, q13
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[24][i] - step1b[23][i]) * cospi_16_64;
+    ;temp2 = (step1b[24][i] + step1b[23][i]) * cospi_16_64;
+    ;step1[23] = dct_const_round_shift(temp1);
+    ;step1[24] = dct_const_round_shift(temp2);
+    ; TODO(cd) do some register allocation change to remove these push/pop
+    vpop  {q13} ; [23]
+    vpop  {q14} ; [24]
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29
+    STORE_IN_OUTPUT 22, 24, 23, q14, q13
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[20] = step1b[23][i] - step1b[20][i];
+    ;step1[27] = step1b[24][i] - step1b[27][i];
+    vsub.s16  q14, q5, q0
+    vsub.s16  q13, q6, q2
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[20][i] * (-cospi_8_64)  - step1b[27][i] * (-cospi_24_64);
+    ;temp2 = step1b[20][i] * (-cospi_24_64) + step1b[27][i] * (-cospi_8_64);
+    ;step2[27] = dct_const_round_shift(temp1);
+    ;step2[20] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d10, d11, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[21] = step1b[22][i] - step1b[21][i];
+    ;step1[26] = step1b[25][i] - step1b[26][i];
+    vsub.s16  q14,  q7, q1
+    vsub.s16  q13,  q4, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[21][i] * (-cospi_8_64)  - step1b[26][i] * (-cospi_24_64);
+    ;temp2 = step1b[21][i] * (-cospi_24_64) + step1b[26][i] * (-cospi_8_64);
+    ;step2[26] = dct_const_round_shift(temp1);
+    ;step2[21] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d0, d1, d2, d3
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[18] = step1b[18][i] + step1b[21][i];
+    ;step3[19] = step1b[19][i] + step1b[20][i];
+    ;step3[20] = step1b[19][i] - step1b[20][i];
+    ;step3[21] = step1b[18][i] - step1b[21][i];
+    LOAD_FROM_OUTPUT 23, 18, 19, q14, q13
+    vadd.s16  q8,  q14, q1
+    vadd.s16  q9,  q13, q6
+    vsub.s16  q13, q13, q6
+    vsub.s16  q1,  q14, q1
+    STORE_IN_OUTPUT 19, 18, 19, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[27] = step1b[28][i] - step1b[27][i];
+    ;step3[28] = step1b[28][i] + step1b[27][i];
+    ;step3[29] = step1b[29][i] + step1b[26][i];
+    ;step3[26] = step1b[29][i] - step1b[26][i];
+    LOAD_FROM_OUTPUT 19, 28, 29, q8, q9
+    vsub.s16  q14, q8, q5
+    vadd.s16  q10, q8, q5
+    vadd.s16  q11, q9, q0
+    vsub.s16  q0, q9, q0
+    STORE_IN_OUTPUT 29, 28, 29, q10, q11
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[27][i] - step1b[20][i]) * cospi_16_64;
+    ;temp2 = (step1b[27][i] + step1b[20][i]) * cospi_16_64;
+    ;step1[20] = dct_const_round_shift(temp1);
+    ;step1[27] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29
+    STORE_IN_OUTPUT 29, 20, 27, q13, q14
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[26][i] - step1b[21][i]) * cospi_16_64;
+    ;temp2 = (step1b[26][i] + step1b[21][i]) * cospi_16_64;
+    ;step1[21] = dct_const_round_shift(temp1);
+    ;step1[26] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY d0, d1, d2, d3, cospi_16_64, cospi_16_64, d2, d3, d0, d1
+    STORE_IN_OUTPUT 27, 21, 26, q1, q0
+    ; --------------------------------------------------------------------------
+
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK C: 8-10,11-15
+    ; --------------------------------------------------------------------------
+    ; generate 8,9,14,15
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[2 * 32] * cospi_30_64 - input[30 * 32] * cospi_2_64;
+    ;temp2 = input[2 * 32] * cospi_2_64 + input[30 * 32] * cospi_30_64;
+    ;step2[8] = dct_const_round_shift(temp1);
+    ;step2[15] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 3, 2, 30
+    DO_BUTTERFLY_STD cospi_30_64, cospi_2_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[18 * 32] * cospi_14_64 - input[14 * 32] * cospi_18_64;
+    ;temp2 = input[18 * 32] * cospi_18_64 + input[14 * 32] * cospi_14_64;
+    ;step2[9] = dct_const_round_shift(temp1);
+    ;step2[14] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 30, 18, 14
+    DO_BUTTERFLY_STD cospi_14_64, cospi_18_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;step3[8] = step1b[8][i] + step1b[9][i];
+    ;step3[9] = step1b[8][i] - step1b[9][i];
+    ;step3[14] = step1b[15][i] - step1b[14][i];
+    ;step3[15] = step1b[15][i] + step1b[14][i];
+    vsub.s16  q13, q0, q1
+    vadd.s16  q0, q0, q1
+    vsub.s16  q14, q2, q3
+    vadd.s16  q2, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = step1b[14][i] * cospi_24_64 - step1b[9][i] * cospi_8_64;
+    ;temp2 = step1b[14][i] * cospi_8_64  + step1b[9][i] * cospi_24_64;
+    ;step1[9]  = dct_const_round_shift(temp1);
+    ;step1[14] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; generate 10,11,12,13
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[10 * 32] * cospi_22_64 - input[22 * 32] * cospi_10_64;
+    ;temp2 = input[10 * 32] * cospi_10_64 + input[22 * 32] * cospi_22_64;
+    ;step2[10] = dct_const_round_shift(temp1);
+    ;step2[13] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 14, 10, 22
+    DO_BUTTERFLY_STD cospi_22_64, cospi_10_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[26 * 32] * cospi_6_64 - input[6 * 32] * cospi_26_64;
+    ;temp2 = input[26 * 32] * cospi_26_64 + input[6 * 32] * cospi_6_64;
+    ;step2[11] = dct_const_round_shift(temp1);
+    ;step2[12] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 22, 26, 6
+    DO_BUTTERFLY_STD cospi_6_64, cospi_26_64, d8, d9, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;step3[10] = step1b[11][i] - step1b[10][i];
+    ;step3[11] = step1b[11][i] + step1b[10][i];
+    ;step3[12] = step1b[12][i] + step1b[13][i];
+    ;step3[13] = step1b[12][i] - step1b[13][i];
+    vsub.s16  q14, q4, q5
+    vadd.s16  q5, q4, q5
+    vsub.s16  q13, q6, q7
+    vadd.s16  q6, q6, q7
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = step1b[10][i] * (-cospi_8_64)  - step1b[13][i] * (-cospi_24_64);
+    ;temp2 = step1b[10][i] * (-cospi_24_64) + step1b[13][i] * (-cospi_8_64);
+    ;step1[13] = dct_const_round_shift(temp1);
+    ;step1[10] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d8, d9, d14, d15
+    ; --------------------------------------------------------------------------
+    ; combine 8-10,11-15
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;step2[8]  = step1b[8][i] + step1b[11][i];
+    ;step2[9]  = step1b[9][i] + step1b[10][i];
+    ;step2[10] = step1b[9][i] - step1b[10][i];
+    vadd.s16  q8,  q0, q5
+    vadd.s16  q9,  q1, q7
+    vsub.s16  q13, q1, q7
+    ;step2[13] = step1b[14][i] - step1b[13][i];
+    ;step2[14] = step1b[14][i] + step1b[13][i];
+    ;step2[15] = step1b[15][i] + step1b[12][i];
+    vsub.s16  q14, q3, q4
+    vadd.s16  q10, q3, q4
+    vadd.s16  q15, q2, q6
+    STORE_IN_OUTPUT 26, 8, 15, q8, q15
+    STORE_IN_OUTPUT 15, 9, 14, q9, q10
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;temp1 = (step1b[13][i] - step1b[10][i]) * cospi_16_64;
+    ;temp2 = (step1b[13][i] + step1b[10][i]) * cospi_16_64;
+    ;step3[10] = dct_const_round_shift(temp1);
+    ;step3[13] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7
+    STORE_IN_OUTPUT 14, 13, 10, q3, q1
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;step2[11] = step1b[8][i] - step1b[11][i];
+    ;step2[12] = step1b[15][i] - step1b[12][i];
+    vsub.s16  q13, q0, q5
+    vsub.s16  q14,  q2, q6
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;temp1 = (step1b[12][i] - step1b[11][i]) * cospi_16_64;
+    ;temp2 = (step1b[12][i] + step1b[11][i]) * cospi_16_64;
+    ;step3[11] = dct_const_round_shift(temp1);
+    ;step3[12] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7
+    STORE_IN_OUTPUT 10, 11, 12, q1, q3
+    ; --------------------------------------------------------------------------
+
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK D: 0-3,4-7
+    ; --------------------------------------------------------------------------
+    ; generate 4,5,6,7
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = input[4 * 32] * cospi_28_64 - input[28 * 32] * cospi_4_64;
+    ;temp2 = input[4 * 32] * cospi_4_64 + input[28 * 32] * cospi_28_64;
+    ;step3[4] = dct_const_round_shift(temp1);
+    ;step3[7] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 6, 4, 28
+    DO_BUTTERFLY_STD cospi_28_64, cospi_4_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = input[20 * 32] * cospi_12_64 - input[12 * 32] * cospi_20_64;
+    ;temp2 = input[20 * 32] * cospi_20_64 + input[12 * 32] * cospi_12_64;
+    ;step3[5] = dct_const_round_shift(temp1);
+    ;step3[6] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 28, 20, 12
+    DO_BUTTERFLY_STD cospi_12_64, cospi_20_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[4] = step1b[4][i] + step1b[5][i];
+    ;step1[5] = step1b[4][i] - step1b[5][i];
+    ;step1[6] = step1b[7][i] - step1b[6][i];
+    ;step1[7] = step1b[7][i] + step1b[6][i];
+    vsub.s16  q13, q0, q1
+    vadd.s16  q0, q0, q1
+    vsub.s16  q14, q2, q3
+    vadd.s16  q2, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = (step1b[6][i] - step1b[5][i]) * cospi_16_64;
+    ;temp2 = (step1b[5][i] + step1b[6][i]) * cospi_16_64;
+    ;step2[5] = dct_const_round_shift(temp1);
+    ;step2[6] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; generate 0,1,2,3
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = (input[0 * 32] - input[16 * 32]) * cospi_16_64;
+    ;temp2 = (input[0 * 32] + input[16 * 32]) * cospi_16_64;
+    ;step1[1] = dct_const_round_shift(temp1);
+    ;step1[0] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 12, 0, 16
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = input[8 * 32] * cospi_24_64 - input[24 * 32] * cospi_8_64;
+    ;temp2 = input[8 * 32] * cospi_8_64 + input[24 * 32] * cospi_24_64;
+    ;step1[2] = dct_const_round_shift(temp1);
+    ;step1[3] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 16, 8, 24
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d28, d29, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;step2[0] = step1b[0][i] + step1b[3][i];
+    ;step2[1] = step1b[1][i] + step1b[2][i];
+    ;step2[2] = step1b[1][i] - step1b[2][i];
+    ;step2[3] = step1b[0][i] - step1b[3][i];
+    vadd.s16  q4, q7, q6
+    vsub.s16  q7, q7, q6
+    vsub.s16  q6, q5, q14
+    vadd.s16  q5, q5, q14
+    ; --------------------------------------------------------------------------
+    ; combine 0-3,4-7
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[0] = step1b[0][i] + step1b[7][i];
+    ;step3[1] = step1b[1][i] + step1b[6][i];
+    ;step3[2] = step1b[2][i] + step1b[5][i];
+    ;step3[3] = step1b[3][i] + step1b[4][i];
+    vadd.s16  q8,  q4, q2
+    vadd.s16  q9,  q5, q3
+    vadd.s16  q10, q6, q1
+    vadd.s16  q11, q7, q0
+    ;step3[4] = step1b[3][i] - step1b[4][i];
+    ;step3[5] = step1b[2][i] - step1b[5][i];
+    ;step3[6] = step1b[1][i] - step1b[6][i];
+    ;step3[7] = step1b[0][i] - step1b[7][i];
+    vsub.s16  q12, q7, q0
+    vsub.s16  q13, q6, q1
+    vsub.s16  q14, q5, q3
+    vsub.s16  q15, q4, q2
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[0] = step1b[0][i] + step1b[15][i];
+    ;step1[1] = step1b[1][i] + step1b[14][i];
+    ;step1[14] = step1b[1][i] - step1b[14][i];
+    ;step1[15] = step1b[0][i] - step1b[15][i];
+    LOAD_FROM_OUTPUT 12, 14, 15, q0, q1
+    vadd.s16  q2, q8, q1
+    vadd.s16  q3, q9, q0
+    vsub.s16  q4, q9, q0
+    vsub.s16  q5, q8, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[14 * 32] = step1b[14][i] + step1b[17][i];
+    ;output[15 * 32] = step1b[15][i] + step1b[16][i];
+    ;output[16 * 32] = step1b[15][i] - step1b[16][i];
+    ;output[17 * 32] = step1b[14][i] - step1b[17][i];
+    LOAD_FROM_OUTPUT 15, 16, 17, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+
+    cmp r5, #0
+    bgt idct32_bands_end_2nd_pass
+
+idct32_bands_end_1st_pass
+    STORE_IN_OUTPUT 17, 16, 17, q6, q7
+    STORE_IN_OUTPUT 17, 14, 15, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 0 * 32] = step1b[0][i] + step1b[31][i];
+    ;output[ 1 * 32] = step1b[1][i] + step1b[30][i];
+    ;output[30 * 32] = step1b[1][i] - step1b[30][i];
+    ;output[31 * 32] = step1b[0][i] - step1b[31][i];
+    LOAD_FROM_OUTPUT 15, 30, 31, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 31, 30, 31, q6, q7
+    STORE_IN_OUTPUT 31,  0,  1, q4, q5
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[2] = step1b[2][i] + step1b[13][i];
+    ;step1[3] = step1b[3][i] + step1b[12][i];
+    ;step1[12] = step1b[3][i] - step1b[12][i];
+    ;step1[13] = step1b[2][i] - step1b[13][i];
+    LOAD_FROM_OUTPUT 1, 12, 13, q0, q1
+    vadd.s16  q2, q10, q1
+    vadd.s16  q3, q11, q0
+    vsub.s16  q4, q11, q0
+    vsub.s16  q5, q10, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[12 * 32] = step1b[12][i] + step1b[19][i];
+    ;output[13 * 32] = step1b[13][i] + step1b[18][i];
+    ;output[18 * 32] = step1b[13][i] - step1b[18][i];
+    ;output[19 * 32] = step1b[12][i] - step1b[19][i];
+    LOAD_FROM_OUTPUT 13, 18, 19, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_IN_OUTPUT 19, 18, 19, q6, q7
+    STORE_IN_OUTPUT 19, 12, 13, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 2 * 32] = step1b[2][i] + step1b[29][i];
+    ;output[ 3 * 32] = step1b[3][i] + step1b[28][i];
+    ;output[28 * 32] = step1b[3][i] - step1b[28][i];
+    ;output[29 * 32] = step1b[2][i] - step1b[29][i];
+    LOAD_FROM_OUTPUT 13, 28, 29, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 29, 28, 29, q6, q7
+    STORE_IN_OUTPUT 29,  2,  3, q4, q5
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[4] = step1b[4][i] + step1b[11][i];
+    ;step1[5] = step1b[5][i] + step1b[10][i];
+    ;step1[10] = step1b[5][i] - step1b[10][i];
+    ;step1[11] = step1b[4][i] - step1b[11][i];
+    LOAD_FROM_OUTPUT 3, 10, 11, q0, q1
+    vadd.s16  q2, q12, q1
+    vadd.s16  q3, q13, q0
+    vsub.s16  q4, q13, q0
+    vsub.s16  q5, q12, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[10 * 32] = step1b[10][i] + step1b[21][i];
+    ;output[11 * 32] = step1b[11][i] + step1b[20][i];
+    ;output[20 * 32] = step1b[11][i] - step1b[20][i];
+    ;output[21 * 32] = step1b[10][i] - step1b[21][i];
+    LOAD_FROM_OUTPUT 11, 20, 21, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_IN_OUTPUT 21, 20, 21, q6, q7
+    STORE_IN_OUTPUT 21, 10, 11, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 4 * 32] = step1b[4][i] + step1b[27][i];
+    ;output[ 5 * 32] = step1b[5][i] + step1b[26][i];
+    ;output[26 * 32] = step1b[5][i] - step1b[26][i];
+    ;output[27 * 32] = step1b[4][i] - step1b[27][i];
+    LOAD_FROM_OUTPUT 11, 26, 27, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 27, 26, 27, q6, q7
+    STORE_IN_OUTPUT 27,  4,  5, q4, q5
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[6] = step1b[6][i] + step1b[9][i];
+    ;step1[7] = step1b[7][i] + step1b[8][i];
+    ;step1[8] = step1b[7][i] - step1b[8][i];
+    ;step1[9] = step1b[6][i] - step1b[9][i];
+    LOAD_FROM_OUTPUT 5, 8, 9, q0, q1
+    vadd.s16  q2, q14, q1
+    vadd.s16  q3, q15, q0
+    vsub.s16  q4, q15, q0
+    vsub.s16  q5, q14, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 8 * 32] = step1b[8][i] + step1b[23][i];
+    ;output[ 9 * 32] = step1b[9][i] + step1b[22][i];
+    ;output[22 * 32] = step1b[9][i] - step1b[22][i];
+    ;output[23 * 32] = step1b[8][i] - step1b[23][i];
+    LOAD_FROM_OUTPUT 9, 22, 23, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_IN_OUTPUT 23, 22, 23, q6, q7
+    STORE_IN_OUTPUT 23, 8, 9, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 6 * 32] = step1b[6][i] + step1b[25][i];
+    ;output[ 7 * 32] = step1b[7][i] + step1b[24][i];
+    ;output[24 * 32] = step1b[7][i] - step1b[24][i];
+    ;output[25 * 32] = step1b[6][i] - step1b[25][i];
+    LOAD_FROM_OUTPUT 9, 24, 25, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 25, 24, 25, q6, q7
+    STORE_IN_OUTPUT 25,  6,  7, q4, q5
+
+    ; restore r0 by removing the last offset from the last
+    ;     operation (LOAD_FROM_TRANSPOSED 16, 8, 24) => 24*8*2
+    sub r0, r0, #24*8*2
+    ; restore r1 by removing the last offset from the last
+    ;     operation (STORE_IN_OUTPUT 24,  6,  7) => 7*32*2
+    ; advance by 8 columns => 8*2
+    sub r1, r1, #7*32*2 - 8*2
+    ;   advance by 8 lines (8*32*2)
+    ;   go back by the two pairs from the loop (32*2)
+    add r3, r3, #8*32*2 - 32*2
+
+    ; bands loop processing
+    subs r4, r4, #1
+    bne idct32_bands_loop
+
+    ; parameters for second pass
+    ; the input of pass2 is the result of pass1. we have to remove the offset
+    ;   of 32 columns induced by the above idct32_bands_loop
+    sub r3, r1, #32*2
+      ; r1 = pass2[32 * 32]
+    add r1, sp, #2048
+
+    ; pass loop processing
+    add r5, r5, #1
+    b idct32_pass_loop
+
+idct32_bands_end_2nd_pass
+    STORE_COMBINE_CENTER_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 0 * 32] = step1b[0][i] + step1b[31][i];
+    ;output[ 1 * 32] = step1b[1][i] + step1b[30][i];
+    ;output[30 * 32] = step1b[1][i] - step1b[30][i];
+    ;output[31 * 32] = step1b[0][i] - step1b[31][i];
+    LOAD_FROM_OUTPUT 17, 30, 31, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[2] = step1b[2][i] + step1b[13][i];
+    ;step1[3] = step1b[3][i] + step1b[12][i];
+    ;step1[12] = step1b[3][i] - step1b[12][i];
+    ;step1[13] = step1b[2][i] - step1b[13][i];
+    LOAD_FROM_OUTPUT 31, 12, 13, q0, q1
+    vadd.s16  q2, q10, q1
+    vadd.s16  q3, q11, q0
+    vsub.s16  q4, q11, q0
+    vsub.s16  q5, q10, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[12 * 32] = step1b[12][i] + step1b[19][i];
+    ;output[13 * 32] = step1b[13][i] + step1b[18][i];
+    ;output[18 * 32] = step1b[13][i] - step1b[18][i];
+    ;output[19 * 32] = step1b[12][i] - step1b[19][i];
+    LOAD_FROM_OUTPUT 13, 18, 19, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_COMBINE_CENTER_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 2 * 32] = step1b[2][i] + step1b[29][i];
+    ;output[ 3 * 32] = step1b[3][i] + step1b[28][i];
+    ;output[28 * 32] = step1b[3][i] - step1b[28][i];
+    ;output[29 * 32] = step1b[2][i] - step1b[29][i];
+    LOAD_FROM_OUTPUT 19, 28, 29, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[4] = step1b[4][i] + step1b[11][i];
+    ;step1[5] = step1b[5][i] + step1b[10][i];
+    ;step1[10] = step1b[5][i] - step1b[10][i];
+    ;step1[11] = step1b[4][i] - step1b[11][i];
+    LOAD_FROM_OUTPUT 29, 10, 11, q0, q1
+    vadd.s16  q2, q12, q1
+    vadd.s16  q3, q13, q0
+    vsub.s16  q4, q13, q0
+    vsub.s16  q5, q12, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[10 * 32] = step1b[10][i] + step1b[21][i];
+    ;output[11 * 32] = step1b[11][i] + step1b[20][i];
+    ;output[20 * 32] = step1b[11][i] - step1b[20][i];
+    ;output[21 * 32] = step1b[10][i] - step1b[21][i];
+    LOAD_FROM_OUTPUT 11, 20, 21, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_COMBINE_CENTER_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 4 * 32] = step1b[4][i] + step1b[27][i];
+    ;output[ 5 * 32] = step1b[5][i] + step1b[26][i];
+    ;output[26 * 32] = step1b[5][i] - step1b[26][i];
+    ;output[27 * 32] = step1b[4][i] - step1b[27][i];
+    LOAD_FROM_OUTPUT 21, 26, 27, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[6] = step1b[6][i] + step1b[9][i];
+    ;step1[7] = step1b[7][i] + step1b[8][i];
+    ;step1[8] = step1b[7][i] - step1b[8][i];
+    ;step1[9] = step1b[6][i] - step1b[9][i];
+    LOAD_FROM_OUTPUT 27, 8, 9, q0, q1
+    vadd.s16  q2, q14, q1
+    vadd.s16  q3, q15, q0
+    vsub.s16  q4, q15, q0
+    vsub.s16  q5, q14, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 8 * 32] = step1b[8][i] + step1b[23][i];
+    ;output[ 9 * 32] = step1b[9][i] + step1b[22][i];
+    ;output[22 * 32] = step1b[9][i] - step1b[22][i];
+    ;output[23 * 32] = step1b[8][i] - step1b[23][i];
+    LOAD_FROM_OUTPUT 9, 22, 23, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_COMBINE_CENTER_RESULTS_LAST
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 6 * 32] = step1b[6][i] + step1b[25][i];
+    ;output[ 7 * 32] = step1b[7][i] + step1b[24][i];
+    ;output[24 * 32] = step1b[7][i] - step1b[24][i];
+    ;output[25 * 32] = step1b[6][i] - step1b[25][i];
+    LOAD_FROM_OUTPUT 23, 24, 25, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS_LAST
+    ; --------------------------------------------------------------------------
+    ; restore pointers to their initial indices for next band pass by
+    ;     removing/adding dest_stride * 8. The actual increment by eight
+    ;     is taken care of within the _LAST macros.
+    add r6,  r6,  r2, lsl #3
+    add r9,  r9,  r2, lsl #3
+    sub r7,  r7,  r2, lsl #3
+    sub r10, r10, r2, lsl #3
+
+    ; restore r0 by removing the last offset from the last
+    ;     operation (LOAD_FROM_TRANSPOSED 16, 8, 24) => 24*8*2
+    sub r0, r0, #24*8*2
+    ; restore r1 by removing the last offset from the last
+    ;     operation (LOAD_FROM_OUTPUT 23, 24, 25) => 25*32*2
+    ; advance by 8 columns => 8*2
+    sub r1, r1, #25*32*2 - 8*2
+    ;   advance by 8 lines (8*32*2)
+    ;   go back by the two pairs from the loop (32*2)
+    add r3, r3, #8*32*2 - 32*2
+
+    ; bands loop processing
+    subs r4, r4, #1
+    bne idct32_bands_loop
+
+    ; stack operation
+    add sp, sp, #512+2048+2048
+    vpop {d8-d15}
+    pop  {r4-r11}
+    bx              lr
+    ENDP  ; |vpx_idct32x32_1024_add_neon|
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct32x32_add_neon.c b/libvpx/libvpx/vpx_dsp/arm/idct32x32_add_neon.c
new file mode 100644
index 0000000..025437e
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct32x32_add_neon.c
@@ -0,0 +1,719 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+#include "vpx_dsp/txfm_common.h"
+
+#define LOAD_FROM_TRANSPOSED(prev, first, second) \
+    q14s16 = vld1q_s16(trans_buf + first * 8); \
+    q13s16 = vld1q_s16(trans_buf + second * 8);
+
+#define LOAD_FROM_OUTPUT(prev, first, second, qA, qB) \
+    qA = vld1q_s16(out + first * 32); \
+    qB = vld1q_s16(out + second * 32);
+
+#define STORE_IN_OUTPUT(prev, first, second, qA, qB) \
+    vst1q_s16(out + first * 32, qA); \
+    vst1q_s16(out + second * 32, qB);
+
+#define  STORE_COMBINE_CENTER_RESULTS(r10, r9) \
+       __STORE_COMBINE_CENTER_RESULTS(r10, r9, stride, \
+                                      q6s16, q7s16, q8s16, q9s16);
+static INLINE void __STORE_COMBINE_CENTER_RESULTS(
+        uint8_t *p1,
+        uint8_t *p2,
+        int stride,
+        int16x8_t q6s16,
+        int16x8_t q7s16,
+        int16x8_t q8s16,
+        int16x8_t q9s16) {
+    int16x4_t d8s16, d9s16, d10s16, d11s16;
+
+    d8s16 = vld1_s16((int16_t *)p1);
+    p1 += stride;
+    d11s16 = vld1_s16((int16_t *)p2);
+    p2 -= stride;
+    d9s16 = vld1_s16((int16_t *)p1);
+    d10s16 = vld1_s16((int16_t *)p2);
+
+    q7s16 = vrshrq_n_s16(q7s16, 6);
+    q8s16 = vrshrq_n_s16(q8s16, 6);
+    q9s16 = vrshrq_n_s16(q9s16, 6);
+    q6s16 = vrshrq_n_s16(q6s16, 6);
+
+    q7s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q7s16),
+                                           vreinterpret_u8_s16(d9s16)));
+    q8s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                                           vreinterpret_u8_s16(d10s16)));
+    q9s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                                           vreinterpret_u8_s16(d11s16)));
+    q6s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q6s16),
+                                           vreinterpret_u8_s16(d8s16)));
+
+    d9s16  = vreinterpret_s16_u8(vqmovun_s16(q7s16));
+    d10s16 = vreinterpret_s16_u8(vqmovun_s16(q8s16));
+    d11s16 = vreinterpret_s16_u8(vqmovun_s16(q9s16));
+    d8s16  = vreinterpret_s16_u8(vqmovun_s16(q6s16));
+
+    vst1_s16((int16_t *)p1, d9s16);
+    p1 -= stride;
+    vst1_s16((int16_t *)p2, d10s16);
+    p2 += stride;
+    vst1_s16((int16_t *)p1, d8s16);
+    vst1_s16((int16_t *)p2, d11s16);
+    return;
+}
+
+#define  STORE_COMBINE_EXTREME_RESULTS(r7, r6); \
+       __STORE_COMBINE_EXTREME_RESULTS(r7, r6, stride, \
+                                      q4s16, q5s16, q6s16, q7s16);
+static INLINE void __STORE_COMBINE_EXTREME_RESULTS(
+        uint8_t *p1,
+        uint8_t *p2,
+        int stride,
+        int16x8_t q4s16,
+        int16x8_t q5s16,
+        int16x8_t q6s16,
+        int16x8_t q7s16) {
+    int16x4_t d4s16, d5s16, d6s16, d7s16;
+
+    d4s16 = vld1_s16((int16_t *)p1);
+    p1 += stride;
+    d7s16 = vld1_s16((int16_t *)p2);
+    p2 -= stride;
+    d5s16 = vld1_s16((int16_t *)p1);
+    d6s16 = vld1_s16((int16_t *)p2);
+
+    q5s16 = vrshrq_n_s16(q5s16, 6);
+    q6s16 = vrshrq_n_s16(q6s16, 6);
+    q7s16 = vrshrq_n_s16(q7s16, 6);
+    q4s16 = vrshrq_n_s16(q4s16, 6);
+
+    q5s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q5s16),
+                                           vreinterpret_u8_s16(d5s16)));
+    q6s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q6s16),
+                                           vreinterpret_u8_s16(d6s16)));
+    q7s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q7s16),
+                                           vreinterpret_u8_s16(d7s16)));
+    q4s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q4s16),
+                                           vreinterpret_u8_s16(d4s16)));
+
+    d5s16 = vreinterpret_s16_u8(vqmovun_s16(q5s16));
+    d6s16 = vreinterpret_s16_u8(vqmovun_s16(q6s16));
+    d7s16 = vreinterpret_s16_u8(vqmovun_s16(q7s16));
+    d4s16 = vreinterpret_s16_u8(vqmovun_s16(q4s16));
+
+    vst1_s16((int16_t *)p1, d5s16);
+    p1 -= stride;
+    vst1_s16((int16_t *)p2, d6s16);
+    p2 += stride;
+    vst1_s16((int16_t *)p2, d7s16);
+    vst1_s16((int16_t *)p1, d4s16);
+    return;
+}
+
+#define DO_BUTTERFLY_STD(const_1, const_2, qA, qB) \
+        DO_BUTTERFLY(q14s16, q13s16, const_1, const_2, qA, qB);
+static INLINE void DO_BUTTERFLY(
+        int16x8_t q14s16,
+        int16x8_t q13s16,
+        int16_t first_const,
+        int16_t second_const,
+        int16x8_t *qAs16,
+        int16x8_t *qBs16) {
+    int16x4_t d30s16, d31s16;
+    int32x4_t q8s32, q9s32, q10s32, q11s32, q12s32, q15s32;
+    int16x4_t dCs16, dDs16, dAs16, dBs16;
+
+    dCs16 = vget_low_s16(q14s16);
+    dDs16 = vget_high_s16(q14s16);
+    dAs16 = vget_low_s16(q13s16);
+    dBs16 = vget_high_s16(q13s16);
+
+    d30s16 = vdup_n_s16(first_const);
+    d31s16 = vdup_n_s16(second_const);
+
+    q8s32 = vmull_s16(dCs16, d30s16);
+    q10s32 = vmull_s16(dAs16, d31s16);
+    q9s32 = vmull_s16(dDs16, d30s16);
+    q11s32 = vmull_s16(dBs16, d31s16);
+    q12s32 = vmull_s16(dCs16, d31s16);
+
+    q8s32 = vsubq_s32(q8s32, q10s32);
+    q9s32 = vsubq_s32(q9s32, q11s32);
+
+    q10s32 = vmull_s16(dDs16, d31s16);
+    q11s32 = vmull_s16(dAs16, d30s16);
+    q15s32 = vmull_s16(dBs16, d30s16);
+
+    q11s32 = vaddq_s32(q12s32, q11s32);
+    q10s32 = vaddq_s32(q10s32, q15s32);
+
+    *qAs16 = vcombine_s16(vqrshrn_n_s32(q8s32, 14),
+                          vqrshrn_n_s32(q9s32, 14));
+    *qBs16 = vcombine_s16(vqrshrn_n_s32(q11s32, 14),
+                          vqrshrn_n_s32(q10s32, 14));
+    return;
+}
+
+static INLINE void idct32_transpose_pair(
+        int16_t *input,
+        int16_t *t_buf) {
+    int16_t *in;
+    int i;
+    const int stride = 32;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+    int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+    for (i = 0; i < 4; i++, input += 8) {
+        in = input;
+        q8s16 = vld1q_s16(in);
+        in += stride;
+        q9s16 = vld1q_s16(in);
+        in += stride;
+        q10s16 = vld1q_s16(in);
+        in += stride;
+        q11s16 = vld1q_s16(in);
+        in += stride;
+        q12s16 = vld1q_s16(in);
+        in += stride;
+        q13s16 = vld1q_s16(in);
+        in += stride;
+        q14s16 = vld1q_s16(in);
+        in += stride;
+        q15s16 = vld1q_s16(in);
+
+        d16s16 = vget_low_s16(q8s16);
+        d17s16 = vget_high_s16(q8s16);
+        d18s16 = vget_low_s16(q9s16);
+        d19s16 = vget_high_s16(q9s16);
+        d20s16 = vget_low_s16(q10s16);
+        d21s16 = vget_high_s16(q10s16);
+        d22s16 = vget_low_s16(q11s16);
+        d23s16 = vget_high_s16(q11s16);
+        d24s16 = vget_low_s16(q12s16);
+        d25s16 = vget_high_s16(q12s16);
+        d26s16 = vget_low_s16(q13s16);
+        d27s16 = vget_high_s16(q13s16);
+        d28s16 = vget_low_s16(q14s16);
+        d29s16 = vget_high_s16(q14s16);
+        d30s16 = vget_low_s16(q15s16);
+        d31s16 = vget_high_s16(q15s16);
+
+        q8s16  = vcombine_s16(d16s16, d24s16);  // vswp d17, d24
+        q9s16  = vcombine_s16(d18s16, d26s16);  // vswp d19, d26
+        q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+        q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+        q12s16 = vcombine_s16(d17s16, d25s16);
+        q13s16 = vcombine_s16(d19s16, d27s16);
+        q14s16 = vcombine_s16(d21s16, d29s16);
+        q15s16 = vcombine_s16(d23s16, d31s16);
+
+        q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q8s16),
+                            vreinterpretq_s32_s16(q10s16));
+        q1x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q9s16),
+                            vreinterpretq_s32_s16(q11s16));
+        q2x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q12s16),
+                            vreinterpretq_s32_s16(q14s16));
+        q3x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q13s16),
+                            vreinterpretq_s32_s16(q15s16));
+
+        q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                            vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+        q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                            vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+        q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                            vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+        q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                            vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+        vst1q_s16(t_buf, q0x2s16.val[0]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q0x2s16.val[1]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q1x2s16.val[0]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q1x2s16.val[1]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q2x2s16.val[0]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q2x2s16.val[1]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q3x2s16.val[0]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q3x2s16.val[1]);
+        t_buf += 8;
+    }
+    return;
+}
+
+static INLINE void idct32_bands_end_1st_pass(
+        int16_t *out,
+        int16x8_t q2s16,
+        int16x8_t q3s16,
+        int16x8_t q6s16,
+        int16x8_t q7s16,
+        int16x8_t q8s16,
+        int16x8_t q9s16,
+        int16x8_t q10s16,
+        int16x8_t q11s16,
+        int16x8_t q12s16,
+        int16x8_t q13s16,
+        int16x8_t q14s16,
+        int16x8_t q15s16) {
+    int16x8_t q0s16, q1s16, q4s16, q5s16;
+
+    STORE_IN_OUTPUT(17, 16, 17, q6s16, q7s16);
+    STORE_IN_OUTPUT(17, 14, 15, q8s16, q9s16);
+
+    LOAD_FROM_OUTPUT(15, 30, 31, q0s16, q1s16);
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_IN_OUTPUT(31, 30, 31, q6s16, q7s16);
+    STORE_IN_OUTPUT(31, 0, 1, q4s16, q5s16);
+
+    LOAD_FROM_OUTPUT(1, 12, 13, q0s16, q1s16);
+    q2s16 = vaddq_s16(q10s16, q1s16);
+    q3s16 = vaddq_s16(q11s16, q0s16);
+    q4s16 = vsubq_s16(q11s16, q0s16);
+    q5s16 = vsubq_s16(q10s16, q1s16);
+
+    LOAD_FROM_OUTPUT(13, 18, 19, q0s16, q1s16);
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_IN_OUTPUT(19, 18, 19, q6s16, q7s16);
+    STORE_IN_OUTPUT(19, 12, 13, q8s16, q9s16);
+
+    LOAD_FROM_OUTPUT(13, 28, 29, q0s16, q1s16);
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_IN_OUTPUT(29, 28, 29, q6s16, q7s16);
+    STORE_IN_OUTPUT(29, 2, 3, q4s16, q5s16);
+
+    LOAD_FROM_OUTPUT(3, 10, 11, q0s16, q1s16);
+    q2s16 = vaddq_s16(q12s16, q1s16);
+    q3s16 = vaddq_s16(q13s16, q0s16);
+    q4s16 = vsubq_s16(q13s16, q0s16);
+    q5s16 = vsubq_s16(q12s16, q1s16);
+
+    LOAD_FROM_OUTPUT(11, 20, 21, q0s16, q1s16);
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_IN_OUTPUT(21, 20, 21, q6s16, q7s16);
+    STORE_IN_OUTPUT(21, 10, 11, q8s16, q9s16);
+
+    LOAD_FROM_OUTPUT(11, 26, 27, q0s16, q1s16);
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_IN_OUTPUT(27, 26, 27, q6s16, q7s16);
+    STORE_IN_OUTPUT(27, 4, 5, q4s16, q5s16);
+
+    LOAD_FROM_OUTPUT(5, 8, 9, q0s16, q1s16);
+    q2s16 = vaddq_s16(q14s16, q1s16);
+    q3s16 = vaddq_s16(q15s16, q0s16);
+    q4s16 = vsubq_s16(q15s16, q0s16);
+    q5s16 = vsubq_s16(q14s16, q1s16);
+
+    LOAD_FROM_OUTPUT(9, 22, 23, q0s16, q1s16);
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_IN_OUTPUT(23, 22, 23, q6s16, q7s16);
+    STORE_IN_OUTPUT(23, 8, 9, q8s16, q9s16);
+
+    LOAD_FROM_OUTPUT(9, 24, 25, q0s16, q1s16);
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_IN_OUTPUT(25, 24, 25, q6s16, q7s16);
+    STORE_IN_OUTPUT(25, 6, 7, q4s16, q5s16);
+    return;
+}
+
+static INLINE void idct32_bands_end_2nd_pass(
+        int16_t *out,
+        uint8_t *dest,
+        int stride,
+        int16x8_t q2s16,
+        int16x8_t q3s16,
+        int16x8_t q6s16,
+        int16x8_t q7s16,
+        int16x8_t q8s16,
+        int16x8_t q9s16,
+        int16x8_t q10s16,
+        int16x8_t q11s16,
+        int16x8_t q12s16,
+        int16x8_t q13s16,
+        int16x8_t q14s16,
+        int16x8_t q15s16) {
+    uint8_t *r6  = dest + 31 * stride;
+    uint8_t *r7  = dest/* +  0 * stride*/;
+    uint8_t *r9  = dest + 15 * stride;
+    uint8_t *r10 = dest + 16 * stride;
+    int str2 = stride << 1;
+    int16x8_t q0s16, q1s16, q4s16, q5s16;
+
+    STORE_COMBINE_CENTER_RESULTS(r10, r9);
+    r10 += str2; r9 -= str2;
+
+    LOAD_FROM_OUTPUT(17, 30, 31, q0s16, q1s16)
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+    r7 += str2; r6 -= str2;
+
+    LOAD_FROM_OUTPUT(31, 12, 13, q0s16, q1s16)
+    q2s16 = vaddq_s16(q10s16, q1s16);
+    q3s16 = vaddq_s16(q11s16, q0s16);
+    q4s16 = vsubq_s16(q11s16, q0s16);
+    q5s16 = vsubq_s16(q10s16, q1s16);
+
+    LOAD_FROM_OUTPUT(13, 18, 19, q0s16, q1s16)
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_COMBINE_CENTER_RESULTS(r10, r9);
+    r10 += str2; r9 -= str2;
+
+    LOAD_FROM_OUTPUT(19, 28, 29, q0s16, q1s16)
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+    r7 += str2; r6 -= str2;
+
+    LOAD_FROM_OUTPUT(29, 10, 11, q0s16, q1s16)
+    q2s16 = vaddq_s16(q12s16, q1s16);
+    q3s16 = vaddq_s16(q13s16, q0s16);
+    q4s16 = vsubq_s16(q13s16, q0s16);
+    q5s16 = vsubq_s16(q12s16, q1s16);
+
+    LOAD_FROM_OUTPUT(11, 20, 21, q0s16, q1s16)
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_COMBINE_CENTER_RESULTS(r10, r9);
+    r10 += str2; r9 -= str2;
+
+    LOAD_FROM_OUTPUT(21, 26, 27, q0s16, q1s16)
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+    r7 += str2; r6 -= str2;
+
+    LOAD_FROM_OUTPUT(27, 8, 9, q0s16, q1s16)
+    q2s16 = vaddq_s16(q14s16, q1s16);
+    q3s16 = vaddq_s16(q15s16, q0s16);
+    q4s16 = vsubq_s16(q15s16, q0s16);
+    q5s16 = vsubq_s16(q14s16, q1s16);
+
+    LOAD_FROM_OUTPUT(9, 22, 23, q0s16, q1s16)
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_COMBINE_CENTER_RESULTS(r10, r9);
+
+    LOAD_FROM_OUTPUT(23, 24, 25, q0s16, q1s16)
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+    return;
+}
+
+void vpx_idct32x32_1024_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int stride) {
+    int i, idct32_pass_loop;
+    int16_t trans_buf[32 * 8];
+    int16_t pass1[32 * 32];
+    int16_t pass2[32 * 32];
+    int16_t *out;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+
+    for (idct32_pass_loop = 0, out = pass1;
+         idct32_pass_loop < 2;
+         idct32_pass_loop++,
+         input = pass1,  // the input of pass2 is the result of pass1
+         out = pass2) {
+        for (i = 0;
+             i < 4; i++,
+             input += 32 * 8, out += 8) {  // idct32_bands_loop
+            idct32_transpose_pair(input, trans_buf);
+
+            // -----------------------------------------
+            // BLOCK A: 16-19,28-31
+            // -----------------------------------------
+            // generate 16,17,30,31
+            // part of stage 1
+            LOAD_FROM_TRANSPOSED(0, 1, 31)
+            DO_BUTTERFLY_STD(cospi_31_64, cospi_1_64, &q0s16, &q2s16)
+            LOAD_FROM_TRANSPOSED(31, 17, 15)
+            DO_BUTTERFLY_STD(cospi_15_64, cospi_17_64, &q1s16, &q3s16)
+            // part of stage 2
+            q4s16 = vaddq_s16(q0s16, q1s16);
+            q13s16 = vsubq_s16(q0s16, q1s16);
+            q6s16 = vaddq_s16(q2s16, q3s16);
+            q14s16 = vsubq_s16(q2s16, q3s16);
+            // part of stage 3
+            DO_BUTTERFLY_STD(cospi_28_64, cospi_4_64, &q5s16, &q7s16)
+
+            // generate 18,19,28,29
+            // part of stage 1
+            LOAD_FROM_TRANSPOSED(15, 9, 23)
+            DO_BUTTERFLY_STD(cospi_23_64, cospi_9_64, &q0s16, &q2s16)
+            LOAD_FROM_TRANSPOSED(23, 25, 7)
+            DO_BUTTERFLY_STD(cospi_7_64, cospi_25_64, &q1s16, &q3s16)
+            // part of stage 2
+            q13s16 = vsubq_s16(q3s16, q2s16);
+            q3s16 = vaddq_s16(q3s16, q2s16);
+            q14s16 = vsubq_s16(q1s16, q0s16);
+            q2s16 = vaddq_s16(q1s16, q0s16);
+            // part of stage 3
+            DO_BUTTERFLY_STD(-cospi_4_64, -cospi_28_64, &q1s16, &q0s16)
+            // part of stage 4
+            q8s16 = vaddq_s16(q4s16, q2s16);
+            q9s16 = vaddq_s16(q5s16, q0s16);
+            q10s16 = vaddq_s16(q7s16, q1s16);
+            q15s16 = vaddq_s16(q6s16, q3s16);
+            q13s16 = vsubq_s16(q5s16, q0s16);
+            q14s16 = vsubq_s16(q7s16, q1s16);
+            STORE_IN_OUTPUT(0, 16, 31, q8s16, q15s16)
+            STORE_IN_OUTPUT(31, 17, 30, q9s16, q10s16)
+            // part of stage 5
+            DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q0s16, &q1s16)
+            STORE_IN_OUTPUT(30, 29, 18, q1s16, q0s16)
+            // part of stage 4
+            q13s16 = vsubq_s16(q4s16, q2s16);
+            q14s16 = vsubq_s16(q6s16, q3s16);
+            // part of stage 5
+            DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q4s16, &q6s16)
+            STORE_IN_OUTPUT(18, 19, 28, q4s16, q6s16)
+
+            // -----------------------------------------
+            // BLOCK B: 20-23,24-27
+            // -----------------------------------------
+            // generate 20,21,26,27
+            // part of stage 1
+            LOAD_FROM_TRANSPOSED(7, 5, 27)
+            DO_BUTTERFLY_STD(cospi_27_64, cospi_5_64, &q0s16, &q2s16)
+            LOAD_FROM_TRANSPOSED(27, 21, 11)
+            DO_BUTTERFLY_STD(cospi_11_64, cospi_21_64, &q1s16, &q3s16)
+            // part of stage 2
+            q13s16 = vsubq_s16(q0s16, q1s16);
+            q0s16 = vaddq_s16(q0s16, q1s16);
+            q14s16 = vsubq_s16(q2s16, q3s16);
+            q2s16 = vaddq_s16(q2s16, q3s16);
+            // part of stage 3
+            DO_BUTTERFLY_STD(cospi_12_64, cospi_20_64, &q1s16, &q3s16)
+
+            // generate 22,23,24,25
+            // part of stage 1
+            LOAD_FROM_TRANSPOSED(11, 13, 19)
+            DO_BUTTERFLY_STD(cospi_19_64, cospi_13_64, &q5s16, &q7s16)
+            LOAD_FROM_TRANSPOSED(19, 29, 3)
+            DO_BUTTERFLY_STD(cospi_3_64, cospi_29_64, &q4s16, &q6s16)
+            // part of stage 2
+            q14s16 = vsubq_s16(q4s16, q5s16);
+            q5s16  = vaddq_s16(q4s16, q5s16);
+            q13s16 = vsubq_s16(q6s16, q7s16);
+            q6s16  = vaddq_s16(q6s16, q7s16);
+            // part of stage 3
+            DO_BUTTERFLY_STD(-cospi_20_64, -cospi_12_64, &q4s16, &q7s16)
+            // part of stage 4
+            q10s16 = vaddq_s16(q7s16, q1s16);
+            q11s16 = vaddq_s16(q5s16, q0s16);
+            q12s16 = vaddq_s16(q6s16, q2s16);
+            q15s16 = vaddq_s16(q4s16, q3s16);
+            // part of stage 6
+            LOAD_FROM_OUTPUT(28, 16, 17, q14s16, q13s16)
+            q8s16 = vaddq_s16(q14s16, q11s16);
+            q9s16 = vaddq_s16(q13s16, q10s16);
+            q13s16 = vsubq_s16(q13s16, q10s16);
+            q11s16 = vsubq_s16(q14s16, q11s16);
+            STORE_IN_OUTPUT(17, 17, 16, q9s16, q8s16)
+            LOAD_FROM_OUTPUT(16, 30, 31, q14s16, q9s16)
+            q8s16  = vsubq_s16(q9s16, q12s16);
+            q10s16 = vaddq_s16(q14s16, q15s16);
+            q14s16 = vsubq_s16(q14s16, q15s16);
+            q12s16 = vaddq_s16(q9s16, q12s16);
+            STORE_IN_OUTPUT(31, 30, 31, q10s16, q12s16)
+            // part of stage 7
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+            STORE_IN_OUTPUT(31, 25, 22, q14s16, q13s16)
+            q13s16 = q11s16;
+            q14s16 = q8s16;
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+            STORE_IN_OUTPUT(22, 24, 23, q14s16, q13s16)
+            // part of stage 4
+            q14s16 = vsubq_s16(q5s16, q0s16);
+            q13s16 = vsubq_s16(q6s16, q2s16);
+            DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q5s16, &q6s16);
+            q14s16 = vsubq_s16(q7s16, q1s16);
+            q13s16 = vsubq_s16(q4s16, q3s16);
+            DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q0s16, &q1s16);
+            // part of stage 6
+            LOAD_FROM_OUTPUT(23, 18, 19, q14s16, q13s16)
+            q8s16 = vaddq_s16(q14s16, q1s16);
+            q9s16 = vaddq_s16(q13s16, q6s16);
+            q13s16 = vsubq_s16(q13s16, q6s16);
+            q1s16 = vsubq_s16(q14s16, q1s16);
+            STORE_IN_OUTPUT(19, 18, 19, q8s16, q9s16)
+            LOAD_FROM_OUTPUT(19, 28, 29, q8s16, q9s16)
+            q14s16 = vsubq_s16(q8s16, q5s16);
+            q10s16 = vaddq_s16(q8s16, q5s16);
+            q11s16 = vaddq_s16(q9s16, q0s16);
+            q0s16 = vsubq_s16(q9s16, q0s16);
+            STORE_IN_OUTPUT(29, 28, 29, q10s16, q11s16)
+            // part of stage 7
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+            STORE_IN_OUTPUT(29, 20, 27, q13s16, q14s16)
+            DO_BUTTERFLY(q0s16, q1s16, cospi_16_64, cospi_16_64,
+                                                         &q1s16, &q0s16);
+            STORE_IN_OUTPUT(27, 21, 26, q1s16, q0s16)
+
+            // -----------------------------------------
+            // BLOCK C: 8-10,11-15
+            // -----------------------------------------
+            // generate 8,9,14,15
+            // part of stage 2
+            LOAD_FROM_TRANSPOSED(3, 2, 30)
+            DO_BUTTERFLY_STD(cospi_30_64, cospi_2_64, &q0s16, &q2s16)
+            LOAD_FROM_TRANSPOSED(30, 18, 14)
+            DO_BUTTERFLY_STD(cospi_14_64, cospi_18_64, &q1s16, &q3s16)
+            // part of stage 3
+            q13s16 = vsubq_s16(q0s16, q1s16);
+            q0s16 = vaddq_s16(q0s16, q1s16);
+            q14s16 = vsubq_s16(q2s16, q3s16);
+            q2s16 = vaddq_s16(q2s16, q3s16);
+            // part of stage 4
+            DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q1s16, &q3s16)
+
+            // generate 10,11,12,13
+            // part of stage 2
+            LOAD_FROM_TRANSPOSED(14, 10, 22)
+            DO_BUTTERFLY_STD(cospi_22_64, cospi_10_64, &q5s16, &q7s16)
+            LOAD_FROM_TRANSPOSED(22, 26, 6)
+            DO_BUTTERFLY_STD(cospi_6_64, cospi_26_64, &q4s16, &q6s16)
+            // part of stage 3
+            q14s16 = vsubq_s16(q4s16, q5s16);
+            q5s16 = vaddq_s16(q4s16, q5s16);
+            q13s16 = vsubq_s16(q6s16, q7s16);
+            q6s16 = vaddq_s16(q6s16, q7s16);
+            // part of stage 4
+            DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q4s16, &q7s16)
+            // part of stage 5
+            q8s16 = vaddq_s16(q0s16, q5s16);
+            q9s16 = vaddq_s16(q1s16, q7s16);
+            q13s16 = vsubq_s16(q1s16, q7s16);
+            q14s16 = vsubq_s16(q3s16, q4s16);
+            q10s16 = vaddq_s16(q3s16, q4s16);
+            q15s16 = vaddq_s16(q2s16, q6s16);
+            STORE_IN_OUTPUT(26, 8, 15, q8s16, q15s16)
+            STORE_IN_OUTPUT(15, 9, 14, q9s16, q10s16)
+            // part of stage 6
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+            STORE_IN_OUTPUT(14, 13, 10, q3s16, q1s16)
+            q13s16 = vsubq_s16(q0s16, q5s16);
+            q14s16 = vsubq_s16(q2s16, q6s16);
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+            STORE_IN_OUTPUT(10, 11, 12, q1s16, q3s16)
+
+            // -----------------------------------------
+            // BLOCK D: 0-3,4-7
+            // -----------------------------------------
+            // generate 4,5,6,7
+            // part of stage 3
+            LOAD_FROM_TRANSPOSED(6, 4, 28)
+            DO_BUTTERFLY_STD(cospi_28_64, cospi_4_64, &q0s16, &q2s16)
+            LOAD_FROM_TRANSPOSED(28, 20, 12)
+            DO_BUTTERFLY_STD(cospi_12_64, cospi_20_64, &q1s16, &q3s16)
+            // part of stage 4
+            q13s16 = vsubq_s16(q0s16, q1s16);
+            q0s16 = vaddq_s16(q0s16, q1s16);
+            q14s16 = vsubq_s16(q2s16, q3s16);
+            q2s16 = vaddq_s16(q2s16, q3s16);
+            // part of stage 5
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+
+            // generate 0,1,2,3
+            // part of stage 4
+            LOAD_FROM_TRANSPOSED(12, 0, 16)
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q5s16, &q7s16)
+            LOAD_FROM_TRANSPOSED(16, 8, 24)
+            DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q14s16, &q6s16)
+            // part of stage 5
+            q4s16 = vaddq_s16(q7s16, q6s16);
+            q7s16 = vsubq_s16(q7s16, q6s16);
+            q6s16 = vsubq_s16(q5s16, q14s16);
+            q5s16 = vaddq_s16(q5s16, q14s16);
+            // part of stage 6
+            q8s16 = vaddq_s16(q4s16, q2s16);
+            q9s16 = vaddq_s16(q5s16, q3s16);
+            q10s16 = vaddq_s16(q6s16, q1s16);
+            q11s16 = vaddq_s16(q7s16, q0s16);
+            q12s16 = vsubq_s16(q7s16, q0s16);
+            q13s16 = vsubq_s16(q6s16, q1s16);
+            q14s16 = vsubq_s16(q5s16, q3s16);
+            q15s16 = vsubq_s16(q4s16, q2s16);
+            // part of stage 7
+            LOAD_FROM_OUTPUT(12, 14, 15, q0s16, q1s16)
+            q2s16 = vaddq_s16(q8s16, q1s16);
+            q3s16 = vaddq_s16(q9s16, q0s16);
+            q4s16 = vsubq_s16(q9s16, q0s16);
+            q5s16 = vsubq_s16(q8s16, q1s16);
+            LOAD_FROM_OUTPUT(15, 16, 17, q0s16, q1s16)
+            q8s16 = vaddq_s16(q4s16, q1s16);
+            q9s16 = vaddq_s16(q5s16, q0s16);
+            q6s16 = vsubq_s16(q5s16, q0s16);
+            q7s16 = vsubq_s16(q4s16, q1s16);
+
+            if (idct32_pass_loop == 0) {
+                idct32_bands_end_1st_pass(out,
+                         q2s16, q3s16, q6s16, q7s16, q8s16, q9s16,
+                         q10s16, q11s16, q12s16, q13s16, q14s16, q15s16);
+            } else {
+                idct32_bands_end_2nd_pass(out, dest, stride,
+                         q2s16, q3s16, q6s16, q7s16, q8s16, q9s16,
+                         q10s16, q11s16, q12s16, q13s16, q14s16, q15s16);
+                dest += 8;
+            }
+        }
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct4x4_1_add_neon.asm b/libvpx/libvpx/vpx_dsp/arm/idct4x4_1_add_neon.asm
new file mode 100644
index 0000000..adab715
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct4x4_1_add_neon.asm
@@ -0,0 +1,68 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+
+    EXPORT  |vpx_idct4x4_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vpx_idct4x4_1_add_neon(int16_t *input, uint8_t *dest,
+;                                  int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vpx_idct4x4_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 4)
+    add              r0, r0, #8                ; + (1 <<((4) - 1))
+    asr              r0, r0, #4                ; >> 4
+
+    vdup.s16         q0, r0                    ; duplicate a1
+
+    vld1.32          {d2[0]}, [r1], r2
+    vld1.32          {d2[1]}, [r1], r2
+    vld1.32          {d4[0]}, [r1], r2
+    vld1.32          {d4[1]}, [r1]
+
+    vaddw.u8         q8, q0, d2                ; dest[x] + a1
+    vaddw.u8         q9, q0, d4
+
+    vqmovun.s16      d6, q8                    ; clip_pixel
+    vqmovun.s16      d7, q9
+
+    vst1.32          {d6[0]}, [r12], r2
+    vst1.32          {d6[1]}, [r12], r2
+    vst1.32          {d7[0]}, [r12], r2
+    vst1.32          {d7[1]}, [r12]
+
+    bx               lr
+    ENDP             ; |vpx_idct4x4_1_add_neon|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct4x4_1_add_neon.c b/libvpx/libvpx/vpx_dsp/arm/idct4x4_1_add_neon.c
new file mode 100644
index 0000000..ea61870
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct4x4_1_add_neon.c
@@ -0,0 +1,50 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "vpx_dsp/inv_txfm.h"
+#include "vpx_ports/mem.h"
+
+void vpx_idct4x4_1_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8x8_t d6u8;
+    uint32x2_t d2u32 = vdup_n_u32(0);
+    uint16x8_t q8u16;
+    int16x8_t q0s16;
+    uint8_t *d1, *d2;
+    int16_t i, a1, cospi_16_64 = 11585;
+    int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+    out = dct_const_round_shift(out * cospi_16_64);
+    a1 = ROUND_POWER_OF_TWO(out, 4);
+
+    q0s16 = vdupq_n_s16(a1);
+
+    // dc_only_idct_add
+    d1 = d2 = dest;
+    for (i = 0; i < 2; i++) {
+        d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 0);
+        d1 += dest_stride;
+        d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 1);
+        d1 += dest_stride;
+
+        q8u16 = vaddw_u8(vreinterpretq_u16_s16(q0s16),
+                         vreinterpret_u8_u32(d2u32));
+        d6u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+
+        vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 0);
+        d2 += dest_stride;
+        vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 1);
+        d2 += dest_stride;
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct4x4_add_neon.asm b/libvpx/libvpx/vpx_dsp/arm/idct4x4_add_neon.asm
new file mode 100644
index 0000000..877fbd6
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct4x4_add_neon.asm
@@ -0,0 +1,190 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vpx_idct4x4_16_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    AREA     Block, CODE, READONLY ; name this block of code
+;void vpx_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vpx_idct4x4_16_add_neon| PROC
+
+    ; The 2D transform is done with two passes which are actually pretty
+    ; similar. We first transform the rows. This is done by transposing
+    ; the inputs, doing an SIMD column transform (the columns are the
+    ; transposed rows) and then transpose the results (so that it goes back
+    ; in normal/row positions). Then, we transform the columns by doing
+    ; another SIMD column transform.
+    ; So, two passes of a transpose followed by a column transform.
+
+    ; load the inputs into q8-q9, d16-d19
+    vld1.s16        {q8,q9}, [r0]!
+
+    ; generate scalar constants
+    ; cospi_8_64 = 15137 = 0x3b21
+    mov             r0, #0x3b00
+    add             r0, #0x21
+    ; cospi_16_64 = 11585 = 0x2d41
+    mov             r3, #0x2d00
+    add             r3, #0x41
+    ; cospi_24_64 = 6270 = 0x 187e
+    mov             r12, #0x1800
+    add             r12, #0x7e
+
+    ; transpose the input data
+    ; 00 01 02 03   d16
+    ; 10 11 12 13   d17
+    ; 20 21 22 23   d18
+    ; 30 31 32 33   d19
+    vtrn.16         d16, d17
+    vtrn.16         d18, d19
+
+    ; generate constant vectors
+    vdup.16         d20, r0         ; replicate cospi_8_64
+    vdup.16         d21, r3         ; replicate cospi_16_64
+
+    ; 00 10 02 12   d16
+    ; 01 11 03 13   d17
+    ; 20 30 22 32   d18
+    ; 21 31 23 33   d19
+    vtrn.32         q8, q9
+    ; 00 10 20 30   d16
+    ; 01 11 21 31   d17
+    ; 02 12 22 32   d18
+    ; 03 13 23 33   d19
+
+    vdup.16         d22, r12        ; replicate cospi_24_64
+
+    ; do the transform on transposed rows
+
+    ; stage 1
+    vadd.s16  d23, d16, d18         ; (input[0] + input[2])
+    vsub.s16  d24, d16, d18         ; (input[0] - input[2])
+
+    vmull.s16 q15, d17, d22         ; input[1] * cospi_24_64
+    vmull.s16 q1,  d17, d20         ; input[1] * cospi_8_64
+
+    ; (input[0] + input[2]) * cospi_16_64;
+    ; (input[0] - input[2]) * cospi_16_64;
+    vmull.s16 q13, d23, d21
+    vmull.s16 q14, d24, d21
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64;
+    ; input[1] * cospi_8_64  + input[3] * cospi_24_64;
+    vmlsl.s16 q15, d19, d20
+    vmlal.s16 q1,  d19, d22
+
+    ; dct_const_round_shift
+    vqrshrn.s32 d26, q13, #14
+    vqrshrn.s32 d27, q14, #14
+    vqrshrn.s32 d29, q15, #14
+    vqrshrn.s32 d28, q1,  #14
+
+    ; stage 2
+    ; output[0] = step[0] + step[3];
+    ; output[1] = step[1] + step[2];
+    ; output[3] = step[0] - step[3];
+    ; output[2] = step[1] - step[2];
+    vadd.s16 q8,  q13, q14
+    vsub.s16 q9,  q13, q14
+    vswp     d18, d19
+
+    ; transpose the results
+    ; 00 01 02 03   d16
+    ; 10 11 12 13   d17
+    ; 20 21 22 23   d18
+    ; 30 31 32 33   d19
+    vtrn.16         d16, d17
+    vtrn.16         d18, d19
+    ; 00 10 02 12   d16
+    ; 01 11 03 13   d17
+    ; 20 30 22 32   d18
+    ; 21 31 23 33   d19
+    vtrn.32         q8, q9
+    ; 00 10 20 30   d16
+    ; 01 11 21 31   d17
+    ; 02 12 22 32   d18
+    ; 03 13 23 33   d19
+
+    ; do the transform on columns
+
+    ; stage 1
+    vadd.s16  d23, d16, d18         ; (input[0] + input[2])
+    vsub.s16  d24, d16, d18         ; (input[0] - input[2])
+
+    vmull.s16 q15, d17, d22         ; input[1] * cospi_24_64
+    vmull.s16 q1,  d17, d20         ; input[1] * cospi_8_64
+
+    ; (input[0] + input[2]) * cospi_16_64;
+    ; (input[0] - input[2]) * cospi_16_64;
+    vmull.s16 q13, d23, d21
+    vmull.s16 q14, d24, d21
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64;
+    ; input[1] * cospi_8_64  + input[3] * cospi_24_64;
+    vmlsl.s16 q15, d19, d20
+    vmlal.s16 q1,  d19, d22
+
+    ; dct_const_round_shift
+    vqrshrn.s32 d26, q13, #14
+    vqrshrn.s32 d27, q14, #14
+    vqrshrn.s32 d29, q15, #14
+    vqrshrn.s32 d28, q1,  #14
+
+    ; stage 2
+    ; output[0] = step[0] + step[3];
+    ; output[1] = step[1] + step[2];
+    ; output[3] = step[0] - step[3];
+    ; output[2] = step[1] - step[2];
+    vadd.s16 q8,  q13, q14
+    vsub.s16 q9,  q13, q14
+
+    ; The results are in two registers, one of them being swapped. This will
+    ; be taken care of by loading the 'dest' value in a swapped fashion and
+    ; also storing them in the same swapped fashion.
+    ; temp_out[0, 1] = d16, d17 = q8
+    ; temp_out[2, 3] = d19, d18 = q9 swapped
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 4)
+    vrshr.s16 q8, q8, #4
+    vrshr.s16 q9, q9, #4
+
+    vld1.32 {d26[0]}, [r1], r2
+    vld1.32 {d26[1]}, [r1], r2
+    vld1.32 {d27[1]}, [r1], r2
+    vld1.32 {d27[0]}, [r1]  ; no post-increment
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 4) + dest[j * dest_stride + i]
+    vaddw.u8 q8, q8, d26
+    vaddw.u8 q9, q9, d27
+
+    ; clip_pixel
+    vqmovun.s16 d26, q8
+    vqmovun.s16 d27, q9
+
+    ; do the stores in reverse order with negative post-increment, by changing
+    ; the sign of the stride
+    rsb r2, r2, #0
+    vst1.32 {d27[0]}, [r1], r2
+    vst1.32 {d27[1]}, [r1], r2
+    vst1.32 {d26[1]}, [r1], r2
+    vst1.32 {d26[0]}, [r1]  ; no post-increment
+    bx              lr
+    ENDP  ; |vpx_idct4x4_16_add_neon|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct4x4_add_neon.c b/libvpx/libvpx/vpx_dsp/arm/idct4x4_add_neon.c
new file mode 100644
index 0000000..3c975c9
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct4x4_add_neon.c
@@ -0,0 +1,151 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+void vpx_idct4x4_16_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8x8_t d26u8, d27u8;
+    uint32x2_t d26u32, d27u32;
+    uint16x8_t q8u16, q9u16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16;
+    int16x4_t d22s16, d23s16, d24s16, d26s16, d27s16, d28s16, d29s16;
+    int16x8_t q8s16, q9s16, q13s16, q14s16;
+    int32x4_t q1s32, q13s32, q14s32, q15s32;
+    int16x4x2_t d0x2s16, d1x2s16;
+    int32x4x2_t q0x2s32;
+    uint8_t *d;
+    int16_t cospi_8_64 = 15137;
+    int16_t cospi_16_64 = 11585;
+    int16_t cospi_24_64 = 6270;
+
+    d26u32 = d27u32 = vdup_n_u32(0);
+
+    q8s16 = vld1q_s16(input);
+    q9s16 = vld1q_s16(input + 8);
+
+    d16s16 = vget_low_s16(q8s16);
+    d17s16 = vget_high_s16(q8s16);
+    d18s16 = vget_low_s16(q9s16);
+    d19s16 = vget_high_s16(q9s16);
+
+    d0x2s16 = vtrn_s16(d16s16, d17s16);
+    d1x2s16 = vtrn_s16(d18s16, d19s16);
+    q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
+    q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
+
+    d20s16 = vdup_n_s16(cospi_8_64);
+    d21s16 = vdup_n_s16(cospi_16_64);
+
+    q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q8s16),
+                        vreinterpretq_s32_s16(q9s16));
+    d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+    d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+    d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+    d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+
+    d22s16 = vdup_n_s16(cospi_24_64);
+
+    // stage 1
+    d23s16 = vadd_s16(d16s16, d18s16);
+    d24s16 = vsub_s16(d16s16, d18s16);
+
+    q15s32 = vmull_s16(d17s16, d22s16);
+    q1s32  = vmull_s16(d17s16, d20s16);
+    q13s32 = vmull_s16(d23s16, d21s16);
+    q14s32 = vmull_s16(d24s16, d21s16);
+
+    q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
+    q1s32  = vmlal_s16(q1s32,  d19s16, d22s16);
+
+    d26s16 = vqrshrn_n_s32(q13s32, 14);
+    d27s16 = vqrshrn_n_s32(q14s32, 14);
+    d29s16 = vqrshrn_n_s32(q15s32, 14);
+    d28s16 = vqrshrn_n_s32(q1s32,  14);
+    q13s16 = vcombine_s16(d26s16, d27s16);
+    q14s16 = vcombine_s16(d28s16, d29s16);
+
+    // stage 2
+    q8s16 = vaddq_s16(q13s16, q14s16);
+    q9s16 = vsubq_s16(q13s16, q14s16);
+
+    d16s16 = vget_low_s16(q8s16);
+    d17s16 = vget_high_s16(q8s16);
+    d18s16 = vget_high_s16(q9s16);  // vswp d18 d19
+    d19s16 = vget_low_s16(q9s16);
+
+    d0x2s16 = vtrn_s16(d16s16, d17s16);
+    d1x2s16 = vtrn_s16(d18s16, d19s16);
+    q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
+    q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
+
+    q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q8s16),
+                        vreinterpretq_s32_s16(q9s16));
+    d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+    d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+    d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+    d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+
+    // do the transform on columns
+    // stage 1
+    d23s16 = vadd_s16(d16s16, d18s16);
+    d24s16 = vsub_s16(d16s16, d18s16);
+
+    q15s32 = vmull_s16(d17s16, d22s16);
+    q1s32  = vmull_s16(d17s16, d20s16);
+    q13s32 = vmull_s16(d23s16, d21s16);
+    q14s32 = vmull_s16(d24s16, d21s16);
+
+    q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
+    q1s32  = vmlal_s16(q1s32,  d19s16, d22s16);
+
+    d26s16 = vqrshrn_n_s32(q13s32, 14);
+    d27s16 = vqrshrn_n_s32(q14s32, 14);
+    d29s16 = vqrshrn_n_s32(q15s32, 14);
+    d28s16 = vqrshrn_n_s32(q1s32,  14);
+    q13s16 = vcombine_s16(d26s16, d27s16);
+    q14s16 = vcombine_s16(d28s16, d29s16);
+
+    // stage 2
+    q8s16 = vaddq_s16(q13s16, q14s16);
+    q9s16 = vsubq_s16(q13s16, q14s16);
+
+    q8s16 = vrshrq_n_s16(q8s16, 4);
+    q9s16 = vrshrq_n_s16(q9s16, 4);
+
+    d = dest;
+    d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 0);
+    d += dest_stride;
+    d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 1);
+    d += dest_stride;
+    d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 1);
+    d += dest_stride;
+    d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 0);
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                     vreinterpret_u8_u32(d26u32));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                     vreinterpret_u8_u32(d27u32));
+
+    d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+
+    d = dest;
+    vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 0);
+    d += dest_stride;
+    vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 1);
+    d += dest_stride;
+    vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 1);
+    d += dest_stride;
+    vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 0);
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct8x8_1_add_neon.asm b/libvpx/libvpx/vpx_dsp/arm/idct8x8_1_add_neon.asm
new file mode 100644
index 0000000..dbbff36
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct8x8_1_add_neon.asm
@@ -0,0 +1,88 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+
+    EXPORT  |vpx_idct8x8_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vpx_idct8x8_1_add_neon(int16_t *input, uint8_t *dest,
+;                                  int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vpx_idct8x8_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 5)
+    add              r0, r0, #16               ; + (1 <<((5) - 1))
+    asr              r0, r0, #5                ; >> 5
+
+    vdup.s16         q0, r0                    ; duplicate a1
+
+    ; load destination data
+    vld1.64          {d2}, [r1], r2
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r2
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r2
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r2
+    vld1.64          {d17}, [r1]
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r2
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r2
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r2
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r2
+    vst1.64          {d31}, [r12], r2
+
+    bx               lr
+    ENDP             ; |vpx_idct8x8_1_add_neon|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct8x8_1_add_neon.c b/libvpx/libvpx/vpx_dsp/arm/idct8x8_1_add_neon.c
new file mode 100644
index 0000000..c1b801f
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct8x8_1_add_neon.c
@@ -0,0 +1,64 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "vpx_dsp/inv_txfm.h"
+#include "vpx_ports/mem.h"
+
+void vpx_idct8x8_1_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8x8_t d2u8, d3u8, d30u8, d31u8;
+    uint64x1_t d2u64, d3u64, d4u64, d5u64;
+    uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16;
+    int16x8_t q0s16;
+    uint8_t *d1, *d2;
+    int16_t i, a1, cospi_16_64 = 11585;
+    int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+    out = dct_const_round_shift(out * cospi_16_64);
+    a1 = ROUND_POWER_OF_TWO(out, 5);
+
+    q0s16 = vdupq_n_s16(a1);
+    q0u16 = vreinterpretq_u16_s16(q0s16);
+
+    d1 = d2 = dest;
+    for (i = 0; i < 2; i++) {
+        d2u64 = vld1_u64((const uint64_t *)d1);
+        d1 += dest_stride;
+        d3u64 = vld1_u64((const uint64_t *)d1);
+        d1 += dest_stride;
+        d4u64 = vld1_u64((const uint64_t *)d1);
+        d1 += dest_stride;
+        d5u64 = vld1_u64((const uint64_t *)d1);
+        d1 += dest_stride;
+
+        q9u16  = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64));
+        q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64));
+        q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64));
+        q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64));
+
+        d2u8  = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+        d3u8  = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+        d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+        d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d31u8));
+        d2 += dest_stride;
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct8x8_add_neon.asm b/libvpx/libvpx/vpx_dsp/arm/idct8x8_add_neon.asm
new file mode 100644
index 0000000..6ab59b4
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct8x8_add_neon.asm
@@ -0,0 +1,519 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vpx_idct8x8_64_add_neon|
+    EXPORT  |vpx_idct8x8_12_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Parallel 1D IDCT on all the columns of a 8x8 16bit data matrix which are
+    ; loaded in q8-q15. The output will be stored back into q8-q15 registers.
+    ; This macro will touch q0-q7 registers and use them as buffer during
+    ; calculation.
+    MACRO
+    IDCT8x8_1D
+    ; stage 1
+    vdup.16         d0, r3                    ; duplicate cospi_28_64
+    vdup.16         d1, r4                    ; duplicate cospi_4_64
+    vdup.16         d2, r5                    ; duplicate cospi_12_64
+    vdup.16         d3, r6                    ; duplicate cospi_20_64
+
+    ; input[1] * cospi_28_64
+    vmull.s16       q2, d18, d0
+    vmull.s16       q3, d19, d0
+
+    ; input[5] * cospi_12_64
+    vmull.s16       q5, d26, d2
+    vmull.s16       q6, d27, d2
+
+    ; input[1]*cospi_28_64-input[7]*cospi_4_64
+    vmlsl.s16       q2, d30, d1
+    vmlsl.s16       q3, d31, d1
+
+    ; input[5] * cospi_12_64 - input[3] * cospi_20_64
+    vmlsl.s16       q5, d22, d3
+    vmlsl.s16       q6, d23, d3
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d8, q2, #14               ; >> 14
+    vqrshrn.s32     d9, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q5, #14              ; >> 14
+    vqrshrn.s32     d11, q6, #14              ; >> 14
+
+    ; input[1] * cospi_4_64
+    vmull.s16       q2, d18, d1
+    vmull.s16       q3, d19, d1
+
+    ; input[5] * cospi_20_64
+    vmull.s16       q9, d26, d3
+    vmull.s16       q13, d27, d3
+
+    ; input[1]*cospi_4_64+input[7]*cospi_28_64
+    vmlal.s16       q2, d30, d0
+    vmlal.s16       q3, d31, d0
+
+    ; input[5] * cospi_20_64 + input[3] * cospi_12_64
+    vmlal.s16       q9, d22, d2
+    vmlal.s16       q13, d23, d2
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d14, q2, #14              ; >> 14
+    vqrshrn.s32     d15, q3, #14              ; >> 14
+
+    ; stage 2 & stage 3 - even half
+    vdup.16         d0, r7                    ; duplicate cospi_16_64
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q13, #14              ; >> 14
+
+    ; input[0] * cospi_16_64
+    vmull.s16       q2, d16, d0
+    vmull.s16       q3, d17, d0
+
+    ; input[0] * cospi_16_64
+    vmull.s16       q13, d16, d0
+    vmull.s16       q15, d17, d0
+
+    ; (input[0] + input[2]) * cospi_16_64
+    vmlal.s16       q2,  d24, d0
+    vmlal.s16       q3, d25, d0
+
+    ; (input[0] - input[2]) * cospi_16_64
+    vmlsl.s16       q13, d24, d0
+    vmlsl.s16       q15, d25, d0
+
+    vdup.16         d0, r8                    ; duplicate cospi_24_64
+    vdup.16         d1, r9                    ; duplicate cospi_8_64
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d18, q2, #14              ; >> 14
+    vqrshrn.s32     d19, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d22, q13, #14              ; >> 14
+    vqrshrn.s32     d23, q15, #14              ; >> 14
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64
+    ; input[1] * cospi_24_64
+    vmull.s16       q2, d20, d0
+    vmull.s16       q3, d21, d0
+
+    ; input[1] * cospi_8_64
+    vmull.s16       q8, d20, d1
+    vmull.s16       q12, d21, d1
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64
+    vmlsl.s16       q2, d28, d1
+    vmlsl.s16       q3, d29, d1
+
+    ; input[1] * cospi_8_64 + input[3] * cospi_24_64
+    vmlal.s16       q8, d28, d0
+    vmlal.s16       q12, d29, d0
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d26, q2, #14              ; >> 14
+    vqrshrn.s32     d27, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d30, q8, #14              ; >> 14
+    vqrshrn.s32     d31, q12, #14              ; >> 14
+
+    vadd.s16        q0, q9, q15               ; output[0] = step[0] + step[3]
+    vadd.s16        q1, q11, q13              ; output[1] = step[1] + step[2]
+    vsub.s16        q2, q11, q13              ; output[2] = step[1] - step[2]
+    vsub.s16        q3, q9, q15               ; output[3] = step[0] - step[3]
+
+    ; stage 3 -odd half
+    vdup.16         d16, r7                   ; duplicate cospi_16_64
+
+    ; stage 2 - odd half
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5]
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5]
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7]
+    vadd.s16        q7, q7, q6                ; step2[7] = step1[6] + step1[7]
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q11, d28, d16
+    vmull.s16       q12, d29, d16
+
+    ; (step2[6] - step2[5]) * cospi_16_64
+    vmlsl.s16       q9, d26, d16
+    vmlsl.s16       q10, d27, d16
+
+    ; (step2[5] + step2[6]) * cospi_16_64
+    vmlal.s16       q11, d26, d16
+    vmlal.s16       q12, d27, d16
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q9, #14              ; >> 14
+    vqrshrn.s32     d11, q10, #14             ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q11, #14              ; >> 14
+    vqrshrn.s32     d13, q12, #14             ; >> 14
+
+    ; stage 4
+    vadd.s16        q8, q0, q7                ; output[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; output[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; output[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; output[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; output[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; output[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; output[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q7               ; output[7] = step1[0] - step1[7];
+    MEND
+
+    ; Transpose a 8x8 16bit data matrix. Datas are loaded in q8-q15.
+    MACRO
+    TRANSPOSE8X8
+    vswp            d17, d24
+    vswp            d23, d30
+    vswp            d21, d28
+    vswp            d19, d26
+    vtrn.32         q8, q10
+    vtrn.32         q9, q11
+    vtrn.32         q12, q14
+    vtrn.32         q13, q15
+    vtrn.16         q8, q9
+    vtrn.16         q10, q11
+    vtrn.16         q12, q13
+    vtrn.16         q14, q15
+    MEND
+
+    AREA    Block, CODE, READONLY ; name this block of code
+;void vpx_idct8x8_64_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vpx_idct8x8_64_add_neon| PROC
+    push            {r4-r9}
+    vpush           {d8-d15}
+    vld1.s16        {q8,q9}, [r0]!
+    vld1.s16        {q10,q11}, [r0]!
+    vld1.s16        {q12,q13}, [r0]!
+    vld1.s16        {q14,q15}, [r0]!
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0x0c00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r4, #0x3e00
+    add             r4, #0xc5
+
+    ; generate cospi_12_64 = 13623
+    mov             r5, #0x3500
+    add             r5, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r6, #0x2300
+    add             r6, #0x8e
+
+    ; generate cospi_16_64 = 11585
+    mov             r7, #0x2d00
+    add             r7, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r8, #0x1800
+    add             r8, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r9, #0x3b00
+    add             r9, #0x21
+
+    ; First transform rows
+    IDCT8x8_1D
+
+    ; Transpose the matrix
+    TRANSPOSE8X8
+
+    ; Then transform columns
+    IDCT8x8_1D
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5)
+    vrshr.s16       q8, q8, #5
+    vrshr.s16       q9, q9, #5
+    vrshr.s16       q10, q10, #5
+    vrshr.s16       q11, q11, #5
+    vrshr.s16       q12, q12, #5
+    vrshr.s16       q13, q13, #5
+    vrshr.s16       q14, q14, #5
+    vrshr.s16       q15, q15, #5
+
+    ; save dest pointer
+    mov             r0, r1
+
+    ; load destination data
+    vld1.64         {d0}, [r1], r2
+    vld1.64         {d1}, [r1], r2
+    vld1.64         {d2}, [r1], r2
+    vld1.64         {d3}, [r1], r2
+    vld1.64         {d4}, [r1], r2
+    vld1.64         {d5}, [r1], r2
+    vld1.64         {d6}, [r1], r2
+    vld1.64         {d7}, [r1]
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i]
+    vaddw.u8        q8, q8, d0
+    vaddw.u8        q9, q9, d1
+    vaddw.u8        q10, q10, d2
+    vaddw.u8        q11, q11, d3
+    vaddw.u8        q12, q12, d4
+    vaddw.u8        q13, q13, d5
+    vaddw.u8        q14, q14, d6
+    vaddw.u8        q15, q15, d7
+
+    ; clip_pixel
+    vqmovun.s16     d0, q8
+    vqmovun.s16     d1, q9
+    vqmovun.s16     d2, q10
+    vqmovun.s16     d3, q11
+    vqmovun.s16     d4, q12
+    vqmovun.s16     d5, q13
+    vqmovun.s16     d6, q14
+    vqmovun.s16     d7, q15
+
+    ; store the data
+    vst1.64         {d0}, [r0], r2
+    vst1.64         {d1}, [r0], r2
+    vst1.64         {d2}, [r0], r2
+    vst1.64         {d3}, [r0], r2
+    vst1.64         {d4}, [r0], r2
+    vst1.64         {d5}, [r0], r2
+    vst1.64         {d6}, [r0], r2
+    vst1.64         {d7}, [r0], r2
+
+    vpop            {d8-d15}
+    pop             {r4-r9}
+    bx              lr
+    ENDP  ; |vpx_idct8x8_64_add_neon|
+
+;void vpx_idct8x8_12_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vpx_idct8x8_12_add_neon| PROC
+    push            {r4-r9}
+    vpush           {d8-d15}
+    vld1.s16        {q8,q9}, [r0]!
+    vld1.s16        {q10,q11}, [r0]!
+    vld1.s16        {q12,q13}, [r0]!
+    vld1.s16        {q14,q15}, [r0]!
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0x0c00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r4, #0x3e00
+    add             r4, #0xc5
+
+    ; generate cospi_12_64 = 13623
+    mov             r5, #0x3500
+    add             r5, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r6, #0x2300
+    add             r6, #0x8e
+
+    ; generate cospi_16_64 = 11585
+    mov             r7, #0x2d00
+    add             r7, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r8, #0x1800
+    add             r8, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r9, #0x3b00
+    add             r9, #0x21
+
+    ; First transform rows
+    ; stage 1
+    ; The following instructions use vqrdmulh to do the
+    ; dct_const_round_shift(input[1] * cospi_28_64). vqrdmulh will do doubling
+    ; multiply and shift the result by 16 bits instead of 14 bits. So we need
+    ; to double the constants before multiplying to compensate this.
+    mov             r12, r3, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_28_64*2
+    mov             r12, r4, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_4_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_28_64)
+    vqrdmulh.s16    q4, q9, q0
+
+    mov             r12, r6, lsl #1
+    rsb             r12, #0
+    vdup.16         q0, r12                   ; duplicate -cospi_20_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_4_64)
+    vqrdmulh.s16    q7, q9, q1
+
+    mov             r12, r5, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_12_64*2
+
+    ; dct_const_round_shift(- input[3] * cospi_20_64)
+    vqrdmulh.s16    q5, q11, q0
+
+    mov             r12, r7, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_16_64*2
+
+    ; dct_const_round_shift(input[3] * cospi_12_64)
+    vqrdmulh.s16    q6, q11, q1
+
+    ; stage 2 & stage 3 - even half
+    mov             r12, r8, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_24_64*2
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrdmulh.s16    q9, q8, q0
+
+    mov             r12, r9, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_8_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_24_64)
+    vqrdmulh.s16    q13, q10, q1
+
+    ; dct_const_round_shift(input[1] * cospi_8_64)
+    vqrdmulh.s16    q15, q10, q0
+
+    ; stage 3 -odd half
+    vdup.16         d16, r7                   ; duplicate cospi_16_64
+
+    vadd.s16        q0, q9, q15               ; output[0] = step[0] + step[3]
+    vadd.s16        q1, q9, q13               ; output[1] = step[1] + step[2]
+    vsub.s16        q2, q9, q13               ; output[2] = step[1] - step[2]
+    vsub.s16        q3, q9, q15               ; output[3] = step[0] - step[3]
+
+    ; stage 2 - odd half
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5]
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5]
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7]
+    vadd.s16        q7, q7, q6                ; step2[7] = step1[6] + step1[7]
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q11, d28, d16
+    vmull.s16       q12, d29, d16
+
+    ; (step2[6] - step2[5]) * cospi_16_64
+    vmlsl.s16       q9, d26, d16
+    vmlsl.s16       q10, d27, d16
+
+    ; (step2[5] + step2[6]) * cospi_16_64
+    vmlal.s16       q11, d26, d16
+    vmlal.s16       q12, d27, d16
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q9, #14              ; >> 14
+    vqrshrn.s32     d11, q10, #14             ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q11, #14              ; >> 14
+    vqrshrn.s32     d13, q12, #14             ; >> 14
+
+    ; stage 4
+    vadd.s16        q8, q0, q7                ; output[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; output[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; output[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; output[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; output[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; output[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; output[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q7               ; output[7] = step1[0] - step1[7];
+
+    ; Transpose the matrix
+    TRANSPOSE8X8
+
+    ; Then transform columns
+    IDCT8x8_1D
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5)
+    vrshr.s16       q8, q8, #5
+    vrshr.s16       q9, q9, #5
+    vrshr.s16       q10, q10, #5
+    vrshr.s16       q11, q11, #5
+    vrshr.s16       q12, q12, #5
+    vrshr.s16       q13, q13, #5
+    vrshr.s16       q14, q14, #5
+    vrshr.s16       q15, q15, #5
+
+    ; save dest pointer
+    mov             r0, r1
+
+    ; load destination data
+    vld1.64         {d0}, [r1], r2
+    vld1.64         {d1}, [r1], r2
+    vld1.64         {d2}, [r1], r2
+    vld1.64         {d3}, [r1], r2
+    vld1.64         {d4}, [r1], r2
+    vld1.64         {d5}, [r1], r2
+    vld1.64         {d6}, [r1], r2
+    vld1.64         {d7}, [r1]
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i]
+    vaddw.u8        q8, q8, d0
+    vaddw.u8        q9, q9, d1
+    vaddw.u8        q10, q10, d2
+    vaddw.u8        q11, q11, d3
+    vaddw.u8        q12, q12, d4
+    vaddw.u8        q13, q13, d5
+    vaddw.u8        q14, q14, d6
+    vaddw.u8        q15, q15, d7
+
+    ; clip_pixel
+    vqmovun.s16     d0, q8
+    vqmovun.s16     d1, q9
+    vqmovun.s16     d2, q10
+    vqmovun.s16     d3, q11
+    vqmovun.s16     d4, q12
+    vqmovun.s16     d5, q13
+    vqmovun.s16     d6, q14
+    vqmovun.s16     d7, q15
+
+    ; store the data
+    vst1.64         {d0}, [r0], r2
+    vst1.64         {d1}, [r0], r2
+    vst1.64         {d2}, [r0], r2
+    vst1.64         {d3}, [r0], r2
+    vst1.64         {d4}, [r0], r2
+    vst1.64         {d5}, [r0], r2
+    vst1.64         {d6}, [r0], r2
+    vst1.64         {d7}, [r0], r2
+
+    vpop            {d8-d15}
+    pop             {r4-r9}
+    bx              lr
+    ENDP  ; |vpx_idct8x8_12_add_neon|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/idct8x8_add_neon.c b/libvpx/libvpx/vpx_dsp/arm/idct8x8_add_neon.c
new file mode 100644
index 0000000..4b2c2a6
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/idct8x8_add_neon.c
@@ -0,0 +1,540 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+#include "vpx_dsp/txfm_common.h"
+
+static INLINE void TRANSPOSE8X8(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+    int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    *q8s16  = vcombine_s16(d16s16, d24s16);  // vswp d17, d24
+    *q9s16  = vcombine_s16(d18s16, d26s16);  // vswp d19, d26
+    *q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+    *q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+    *q12s16 = vcombine_s16(d17s16, d25s16);
+    *q13s16 = vcombine_s16(d19s16, d27s16);
+    *q14s16 = vcombine_s16(d21s16, d29s16);
+    *q15s16 = vcombine_s16(d23s16, d31s16);
+
+    q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q8s16),
+                        vreinterpretq_s32_s16(*q10s16));
+    q1x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q9s16),
+                        vreinterpretq_s32_s16(*q11s16));
+    q2x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q12s16),
+                        vreinterpretq_s32_s16(*q14s16));
+    q3x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q13s16),
+                        vreinterpretq_s32_s16(*q15s16));
+
+    q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                        vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+    q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                        vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+    q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                        vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+    q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                        vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+    *q8s16  = q0x2s16.val[0];
+    *q9s16  = q0x2s16.val[1];
+    *q10s16 = q1x2s16.val[0];
+    *q11s16 = q1x2s16.val[1];
+    *q12s16 = q2x2s16.val[0];
+    *q13s16 = q2x2s16.val[1];
+    *q14s16 = q3x2s16.val[0];
+    *q15s16 = q3x2s16.val[1];
+    return;
+}
+
+static INLINE void IDCT8x8_1D(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
+
+    d0s16 = vdup_n_s16(cospi_28_64);
+    d1s16 = vdup_n_s16(cospi_4_64);
+    d2s16 = vdup_n_s16(cospi_12_64);
+    d3s16 = vdup_n_s16(cospi_20_64);
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    q2s32 = vmull_s16(d18s16, d0s16);
+    q3s32 = vmull_s16(d19s16, d0s16);
+    q5s32 = vmull_s16(d26s16, d2s16);
+    q6s32 = vmull_s16(d27s16, d2s16);
+
+    q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
+    q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
+    q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
+    q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
+
+    d8s16 = vqrshrn_n_s32(q2s32, 14);
+    d9s16 = vqrshrn_n_s32(q3s32, 14);
+    d10s16 = vqrshrn_n_s32(q5s32, 14);
+    d11s16 = vqrshrn_n_s32(q6s32, 14);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    q2s32 = vmull_s16(d18s16, d1s16);
+    q3s32 = vmull_s16(d19s16, d1s16);
+    q9s32 = vmull_s16(d26s16, d3s16);
+    q13s32 = vmull_s16(d27s16, d3s16);
+
+    q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
+    q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
+    q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
+    q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
+
+    d14s16 = vqrshrn_n_s32(q2s32, 14);
+    d15s16 = vqrshrn_n_s32(q3s32, 14);
+    d12s16 = vqrshrn_n_s32(q9s32, 14);
+    d13s16 = vqrshrn_n_s32(q13s32, 14);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+    q7s16 = vcombine_s16(d14s16, d15s16);
+
+    d0s16 = vdup_n_s16(cospi_16_64);
+
+    q2s32 = vmull_s16(d16s16, d0s16);
+    q3s32 = vmull_s16(d17s16, d0s16);
+    q13s32 = vmull_s16(d16s16, d0s16);
+    q15s32 = vmull_s16(d17s16, d0s16);
+
+    q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
+    q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
+    q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
+    q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
+
+    d0s16 = vdup_n_s16(cospi_24_64);
+    d1s16 = vdup_n_s16(cospi_8_64);
+
+    d18s16 = vqrshrn_n_s32(q2s32, 14);
+    d19s16 = vqrshrn_n_s32(q3s32, 14);
+    d22s16 = vqrshrn_n_s32(q13s32, 14);
+    d23s16 = vqrshrn_n_s32(q15s32, 14);
+    *q9s16 = vcombine_s16(d18s16, d19s16);
+    *q11s16 = vcombine_s16(d22s16, d23s16);
+
+    q2s32 = vmull_s16(d20s16, d0s16);
+    q3s32 = vmull_s16(d21s16, d0s16);
+    q8s32 = vmull_s16(d20s16, d1s16);
+    q12s32 = vmull_s16(d21s16, d1s16);
+
+    q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
+    q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
+    q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
+    q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
+
+    d26s16 = vqrshrn_n_s32(q2s32, 14);
+    d27s16 = vqrshrn_n_s32(q3s32, 14);
+    d30s16 = vqrshrn_n_s32(q8s32, 14);
+    d31s16 = vqrshrn_n_s32(q12s32, 14);
+    *q13s16 = vcombine_s16(d26s16, d27s16);
+    *q15s16 = vcombine_s16(d30s16, d31s16);
+
+    q0s16 = vaddq_s16(*q9s16, *q15s16);
+    q1s16 = vaddq_s16(*q11s16, *q13s16);
+    q2s16 = vsubq_s16(*q11s16, *q13s16);
+    q3s16 = vsubq_s16(*q9s16, *q15s16);
+
+    *q13s16 = vsubq_s16(q4s16, q5s16);
+    q4s16 = vaddq_s16(q4s16, q5s16);
+    *q14s16 = vsubq_s16(q7s16, q6s16);
+    q7s16 = vaddq_s16(q7s16, q6s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+
+    d16s16 = vdup_n_s16(cospi_16_64);
+
+    q9s32 = vmull_s16(d28s16, d16s16);
+    q10s32 = vmull_s16(d29s16, d16s16);
+    q11s32 = vmull_s16(d28s16, d16s16);
+    q12s32 = vmull_s16(d29s16, d16s16);
+
+    q9s32 = vmlsl_s16(q9s32,  d26s16, d16s16);
+    q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+    q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+    q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+    d10s16 = vqrshrn_n_s32(q9s32, 14);
+    d11s16 = vqrshrn_n_s32(q10s32, 14);
+    d12s16 = vqrshrn_n_s32(q11s32, 14);
+    d13s16 = vqrshrn_n_s32(q12s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    *q8s16 = vaddq_s16(q0s16, q7s16);
+    *q9s16 = vaddq_s16(q1s16, q6s16);
+    *q10s16 = vaddq_s16(q2s16, q5s16);
+    *q11s16 = vaddq_s16(q3s16, q4s16);
+    *q12s16 = vsubq_s16(q3s16, q4s16);
+    *q13s16 = vsubq_s16(q2s16, q5s16);
+    *q14s16 = vsubq_s16(q1s16, q6s16);
+    *q15s16 = vsubq_s16(q0s16, q7s16);
+    return;
+}
+
+void vpx_idct8x8_64_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8_t *d1, *d2;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8;
+    uint64x1_t d0u64, d1u64, d2u64, d3u64;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16;
+
+    q8s16 = vld1q_s16(input);
+    q9s16 = vld1q_s16(input + 8);
+    q10s16 = vld1q_s16(input + 16);
+    q11s16 = vld1q_s16(input + 24);
+    q12s16 = vld1q_s16(input + 32);
+    q13s16 = vld1q_s16(input + 40);
+    q14s16 = vld1q_s16(input + 48);
+    q15s16 = vld1q_s16(input + 56);
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+               &q12s16, &q13s16, &q14s16, &q15s16);
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+               &q12s16, &q13s16, &q14s16, &q15s16);
+
+    q8s16 = vrshrq_n_s16(q8s16, 5);
+    q9s16 = vrshrq_n_s16(q9s16, 5);
+    q10s16 = vrshrq_n_s16(q10s16, 5);
+    q11s16 = vrshrq_n_s16(q11s16, 5);
+    q12s16 = vrshrq_n_s16(q12s16, 5);
+    q13s16 = vrshrq_n_s16(q13s16, 5);
+    q14s16 = vrshrq_n_s16(q14s16, 5);
+    q15s16 = vrshrq_n_s16(q15s16, 5);
+
+    d1 = d2 = dest;
+
+    d0u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d1u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d2u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                     vreinterpret_u8_u64(d0u64));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                     vreinterpret_u8_u64(d1u64));
+    q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
+                      vreinterpret_u8_u64(d2u64));
+    q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
+                      vreinterpret_u8_u64(d3u64));
+
+    d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+
+    q8s16 = q12s16;
+    q9s16 = q13s16;
+    q10s16 = q14s16;
+    q11s16 = q15s16;
+
+    d0u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d1u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d2u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                     vreinterpret_u8_u64(d0u64));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                     vreinterpret_u8_u64(d1u64));
+    q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
+                      vreinterpret_u8_u64(d2u64));
+    q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
+                      vreinterpret_u8_u64(d3u64));
+
+    d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+    return;
+}
+
+void vpx_idct8x8_12_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8_t *d1, *d2;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8;
+    int16x4_t d10s16, d11s16, d12s16, d13s16, d16s16;
+    int16x4_t d26s16, d27s16, d28s16, d29s16;
+    uint64x1_t d0u64, d1u64, d2u64, d3u64;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16;
+    int32x4_t q9s32, q10s32, q11s32, q12s32;
+
+    q8s16 = vld1q_s16(input);
+    q9s16 = vld1q_s16(input + 8);
+    q10s16 = vld1q_s16(input + 16);
+    q11s16 = vld1q_s16(input + 24);
+    q12s16 = vld1q_s16(input + 32);
+    q13s16 = vld1q_s16(input + 40);
+    q14s16 = vld1q_s16(input + 48);
+    q15s16 = vld1q_s16(input + 56);
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    // First transform rows
+    // stage 1
+    q0s16 = vdupq_n_s16(cospi_28_64 * 2);
+    q1s16 = vdupq_n_s16(cospi_4_64 * 2);
+
+    q4s16 = vqrdmulhq_s16(q9s16, q0s16);
+
+    q0s16 = vdupq_n_s16(-cospi_20_64 * 2);
+
+    q7s16 = vqrdmulhq_s16(q9s16, q1s16);
+
+    q1s16 = vdupq_n_s16(cospi_12_64 * 2);
+
+    q5s16 = vqrdmulhq_s16(q11s16, q0s16);
+
+    q0s16 = vdupq_n_s16(cospi_16_64 * 2);
+
+    q6s16 = vqrdmulhq_s16(q11s16, q1s16);
+
+    // stage 2 & stage 3 - even half
+    q1s16 = vdupq_n_s16(cospi_24_64 * 2);
+
+    q9s16 = vqrdmulhq_s16(q8s16, q0s16);
+
+    q0s16 = vdupq_n_s16(cospi_8_64 * 2);
+
+    q13s16 = vqrdmulhq_s16(q10s16, q1s16);
+
+    q15s16 = vqrdmulhq_s16(q10s16, q0s16);
+
+    // stage 3 -odd half
+    q0s16 = vaddq_s16(q9s16, q15s16);
+    q1s16 = vaddq_s16(q9s16, q13s16);
+    q2s16 = vsubq_s16(q9s16, q13s16);
+    q3s16 = vsubq_s16(q9s16, q15s16);
+
+    // stage 2 - odd half
+    q13s16 = vsubq_s16(q4s16, q5s16);
+    q4s16 = vaddq_s16(q4s16, q5s16);
+    q14s16 = vsubq_s16(q7s16, q6s16);
+    q7s16 = vaddq_s16(q7s16, q6s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+
+    d16s16 = vdup_n_s16(cospi_16_64);
+    q9s32 = vmull_s16(d28s16, d16s16);
+    q10s32 = vmull_s16(d29s16, d16s16);
+    q11s32 = vmull_s16(d28s16, d16s16);
+    q12s32 = vmull_s16(d29s16, d16s16);
+
+    q9s32 = vmlsl_s16(q9s32,  d26s16, d16s16);
+    q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+    q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+    q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+    d10s16 = vqrshrn_n_s32(q9s32, 14);
+    d11s16 = vqrshrn_n_s32(q10s32, 14);
+    d12s16 = vqrshrn_n_s32(q11s32, 14);
+    d13s16 = vqrshrn_n_s32(q12s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    // stage 4
+    q8s16 = vaddq_s16(q0s16, q7s16);
+    q9s16 = vaddq_s16(q1s16, q6s16);
+    q10s16 = vaddq_s16(q2s16, q5s16);
+    q11s16 = vaddq_s16(q3s16, q4s16);
+    q12s16 = vsubq_s16(q3s16, q4s16);
+    q13s16 = vsubq_s16(q2s16, q5s16);
+    q14s16 = vsubq_s16(q1s16, q6s16);
+    q15s16 = vsubq_s16(q0s16, q7s16);
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+               &q12s16, &q13s16, &q14s16, &q15s16);
+
+    q8s16 = vrshrq_n_s16(q8s16, 5);
+    q9s16 = vrshrq_n_s16(q9s16, 5);
+    q10s16 = vrshrq_n_s16(q10s16, 5);
+    q11s16 = vrshrq_n_s16(q11s16, 5);
+    q12s16 = vrshrq_n_s16(q12s16, 5);
+    q13s16 = vrshrq_n_s16(q13s16, 5);
+    q14s16 = vrshrq_n_s16(q14s16, 5);
+    q15s16 = vrshrq_n_s16(q15s16, 5);
+
+    d1 = d2 = dest;
+
+    d0u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d1u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d2u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                     vreinterpret_u8_u64(d0u64));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                     vreinterpret_u8_u64(d1u64));
+    q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
+                      vreinterpret_u8_u64(d2u64));
+    q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
+                      vreinterpret_u8_u64(d3u64));
+
+    d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+
+    q8s16 = q12s16;
+    q9s16 = q13s16;
+    q10s16 = q14s16;
+    q11s16 = q15s16;
+
+    d0u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d1u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d2u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                     vreinterpret_u8_u64(d0u64));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                     vreinterpret_u8_u64(d1u64));
+    q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
+                      vreinterpret_u8_u64(d2u64));
+    q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
+                      vreinterpret_u8_u64(d3u64));
+
+    d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/intrapred_neon.c b/libvpx/libvpx/vpx_dsp/arm/intrapred_neon.c
new file mode 100644
index 0000000..0a37610
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/intrapred_neon.c
@@ -0,0 +1,822 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+//------------------------------------------------------------------------------
+// DC 4x4
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_4x4(uint8_t *dst, ptrdiff_t stride,
+                          const uint8_t *above, const uint8_t *left,
+                          int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x8_t A = vld1_u8(above);  // top row
+    const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    sum_top = vcombine_u16(p1, p1);
+  }
+
+  if (do_left) {
+    const uint8x8_t L = vld1_u8(left);  // left border
+    const uint16x4_t p0 = vpaddl_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    sum_left = vcombine_u16(p1, p1);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 3);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 2);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 2);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x8_t dc = vdup_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 4; ++i) {
+      vst1_lane_u32((uint32_t*)(dst + i * stride), vreinterpret_u32_u8(dc), 0);
+    }
+  }
+}
+
+void vpx_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  dc_4x4(dst, stride, above, left, 1, 1);
+}
+
+void vpx_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  dc_4x4(dst, stride, NULL, left, 0, 1);
+}
+
+void vpx_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  dc_4x4(dst, stride, above, NULL, 1, 0);
+}
+
+void vpx_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_4x4(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+// DC 8x8
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_8x8(uint8_t *dst, ptrdiff_t stride,
+                          const uint8_t *above, const uint8_t *left,
+                          int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x8_t A = vld1_u8(above);  // top row
+    const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    sum_top = vcombine_u16(p2, p2);
+  }
+
+  if (do_left) {
+    const uint8x8_t L = vld1_u8(left);  // left border
+    const uint16x4_t p0 = vpaddl_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    sum_left = vcombine_u16(p2, p2);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 4);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 3);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 3);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x8_t dc = vdup_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 8; ++i) {
+      vst1_u32((uint32_t*)(dst + i * stride), vreinterpret_u32_u8(dc));
+    }
+  }
+}
+
+void vpx_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  dc_8x8(dst, stride, above, left, 1, 1);
+}
+
+void vpx_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  dc_8x8(dst, stride, NULL, left, 0, 1);
+}
+
+void vpx_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  dc_8x8(dst, stride, above, NULL, 1, 0);
+}
+
+void vpx_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_8x8(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+// DC 16x16
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_16x16(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left,
+                            int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x16_t A = vld1q_u8(above);  // top row
+    const uint16x8_t p0 = vpaddlq_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    const uint16x4_t p3 = vpadd_u16(p2, p2);
+    sum_top = vcombine_u16(p3, p3);
+  }
+
+  if (do_left) {
+    const uint8x16_t L = vld1q_u8(left);  // left row
+    const uint16x8_t p0 = vpaddlq_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    const uint16x4_t p3 = vpadd_u16(p2, p2);
+    sum_left = vcombine_u16(p3, p3);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 5);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 4);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 4);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 16; ++i) {
+      vst1q_u8(dst + i * stride, dc);
+    }
+  }
+}
+
+void vpx_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  dc_16x16(dst, stride, above, left, 1, 1);
+}
+
+void vpx_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  dc_16x16(dst, stride, NULL, left, 0, 1);
+}
+
+void vpx_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  dc_16x16(dst, stride, above, NULL, 1, 0);
+}
+
+void vpx_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_16x16(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+// DC 32x32
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_32x32(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left,
+                            int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x16_t A0 = vld1q_u8(above);  // top row
+    const uint8x16_t A1 = vld1q_u8(above + 16);
+    const uint16x8_t p0 = vpaddlq_u8(A0);  // cascading summation of the top
+    const uint16x8_t p1 = vpaddlq_u8(A1);
+    const uint16x8_t p2 = vaddq_u16(p0, p1);
+    const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
+    const uint16x4_t p4 = vpadd_u16(p3, p3);
+    const uint16x4_t p5 = vpadd_u16(p4, p4);
+    sum_top = vcombine_u16(p5, p5);
+  }
+
+  if (do_left) {
+    const uint8x16_t L0 = vld1q_u8(left);  // left row
+    const uint8x16_t L1 = vld1q_u8(left + 16);
+    const uint16x8_t p0 = vpaddlq_u8(L0);  // cascading summation of the left
+    const uint16x8_t p1 = vpaddlq_u8(L1);
+    const uint16x8_t p2 = vaddq_u16(p0, p1);
+    const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
+    const uint16x4_t p4 = vpadd_u16(p3, p3);
+    const uint16x4_t p5 = vpadd_u16(p4, p4);
+    sum_left = vcombine_u16(p5, p5);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 6);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 5);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 5);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 32; ++i) {
+      vst1q_u8(dst + i * stride, dc);
+      vst1q_u8(dst + i * stride + 16, dc);
+    }
+  }
+}
+
+void vpx_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  dc_32x32(dst, stride, above, left, 1, 1);
+}
+
+void vpx_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  dc_32x32(dst, stride, NULL, left, 0, 1);
+}
+
+void vpx_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  dc_32x32(dst, stride, above, NULL, 1, 0);
+}
+
+void vpx_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_32x32(dst, stride, NULL, NULL, 0, 0);
+}
+
+// -----------------------------------------------------------------------------
+
+void vpx_d45_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const uint64x1_t A0 = vreinterpret_u64_u8(vld1_u8(above));  // top row
+  const uint64x1_t A1 = vshr_n_u64(A0, 8);
+  const uint64x1_t A2 = vshr_n_u64(A0, 16);
+  const uint8x8_t ABCDEFGH = vreinterpret_u8_u64(A0);
+  const uint8x8_t BCDEFGH0 = vreinterpret_u8_u64(A1);
+  const uint8x8_t CDEFGH00 = vreinterpret_u8_u64(A2);
+  const uint8x8_t avg1 = vhadd_u8(ABCDEFGH, CDEFGH00);
+  const uint8x8_t avg2 = vrhadd_u8(avg1, BCDEFGH0);
+  const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
+  const uint32x2_t r0 = vreinterpret_u32_u8(avg2);
+  const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
+  const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
+  const uint32x2_t r3 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
+  (void)left;
+  vst1_lane_u32((uint32_t *)(dst + 0 * stride), r0, 0);
+  vst1_lane_u32((uint32_t *)(dst + 1 * stride), r1, 0);
+  vst1_lane_u32((uint32_t *)(dst + 2 * stride), r2, 0);
+  vst1_lane_u32((uint32_t *)(dst + 3 * stride), r3, 0);
+  dst[3 * stride + 3] = above[7];
+}
+
+void vpx_d45_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  static const uint8_t shuffle1[8] = { 1, 2, 3, 4, 5, 6, 7, 7 };
+  static const uint8_t shuffle2[8] = { 2, 3, 4, 5, 6, 7, 7, 7 };
+  const uint8x8_t sh_12345677 = vld1_u8(shuffle1);
+  const uint8x8_t sh_23456777 = vld1_u8(shuffle2);
+  const uint8x8_t A0 = vld1_u8(above);  // top row
+  const uint8x8_t A1 = vtbl1_u8(A0, sh_12345677);
+  const uint8x8_t A2 = vtbl1_u8(A0, sh_23456777);
+  const uint8x8_t avg1 = vhadd_u8(A0, A2);
+  uint8x8_t row = vrhadd_u8(avg1, A1);
+  int i;
+  (void)left;
+  for (i = 0; i < 7; ++i) {
+    vst1_u8(dst + i * stride, row);
+    row = vtbl1_u8(row, sh_12345677);
+  }
+  vst1_u8(dst + i * stride, row);
+}
+
+void vpx_d45_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  const uint8x16_t A0 = vld1q_u8(above);  // top row
+  const uint8x16_t above_right = vld1q_dup_u8(above + 15);
+  const uint8x16_t A1 = vextq_u8(A0, above_right, 1);
+  const uint8x16_t A2 = vextq_u8(A0, above_right, 2);
+  const uint8x16_t avg1 = vhaddq_u8(A0, A2);
+  uint8x16_t row = vrhaddq_u8(avg1, A1);
+  int i;
+  (void)left;
+  for (i = 0; i < 15; ++i) {
+    vst1q_u8(dst + i * stride, row);
+    row = vextq_u8(row, above_right, 1);
+  }
+  vst1q_u8(dst + i * stride, row);
+}
+
+// -----------------------------------------------------------------------------
+
+void vpx_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const uint8x8_t XABCD_u8 = vld1_u8(above - 1);
+  const uint64x1_t XABCD = vreinterpret_u64_u8(XABCD_u8);
+  const uint64x1_t ____XABC = vshl_n_u64(XABCD, 32);
+  const uint32x2_t zero = vdup_n_u32(0);
+  const uint32x2_t IJKL = vld1_lane_u32((const uint32_t *)left, zero, 0);
+  const uint8x8_t IJKL_u8 = vreinterpret_u8_u32(IJKL);
+  const uint64x1_t LKJI____ = vreinterpret_u64_u8(vrev32_u8(IJKL_u8));
+  const uint64x1_t LKJIXABC = vorr_u64(LKJI____, ____XABC);
+  const uint8x8_t KJIXABC_ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 8));
+  const uint8x8_t JIXABC__ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 16));
+  const uint8_t D = vget_lane_u8(XABCD_u8, 4);
+  const uint8x8_t JIXABCD_ = vset_lane_u8(D, JIXABC__, 6);
+  const uint8x8_t LKJIXABC_u8 = vreinterpret_u8_u64(LKJIXABC);
+  const uint8x8_t avg1 = vhadd_u8(JIXABCD_, LKJIXABC_u8);
+  const uint8x8_t avg2 = vrhadd_u8(avg1, KJIXABC_);
+  const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
+  const uint32x2_t r3 = vreinterpret_u32_u8(avg2);
+  const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
+  const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
+  const uint32x2_t r0 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
+  vst1_lane_u32((uint32_t *)(dst + 0 * stride), r0, 0);
+  vst1_lane_u32((uint32_t *)(dst + 1 * stride), r1, 0);
+  vst1_lane_u32((uint32_t *)(dst + 2 * stride), r2, 0);
+  vst1_lane_u32((uint32_t *)(dst + 3 * stride), r3, 0);
+}
+
+#if !HAVE_NEON_ASM
+
+void vpx_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint32x2_t d0u32 = vdup_n_u32(0);
+  (void)left;
+
+  d0u32 = vld1_lane_u32((const uint32_t *)above, d0u32, 0);
+  for (i = 0; i < 4; i++, dst += stride)
+    vst1_lane_u32((uint32_t *)dst, d0u32, 0);
+}
+
+void vpx_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  (void)left;
+
+  d0u8 = vld1_u8(above);
+  for (i = 0; i < 8; i++, dst += stride)
+    vst1_u8(dst, d0u8);
+}
+
+void vpx_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  (void)left;
+
+  q0u8 = vld1q_u8(above);
+  for (i = 0; i < 16; i++, dst += stride)
+    vst1q_u8(dst, q0u8);
+}
+
+void vpx_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)left;
+
+  q0u8 = vld1q_u8(above);
+  q1u8 = vld1q_u8(above + 16);
+  for (i = 0; i < 32; i++, dst += stride) {
+    vst1q_u8(dst, q0u8);
+    vst1q_u8(dst + 16, q1u8);
+  }
+}
+
+void vpx_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint32x2_t d1u32 = vdup_n_u32(0);
+  (void)above;
+
+  d1u32 = vld1_lane_u32((const uint32_t *)left, d1u32, 0);
+
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 0);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 1);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 2);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 3);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+}
+
+void vpx_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint64x1_t d1u64 = vdup_n_u64(0);
+  (void)above;
+
+  d1u64 = vld1_u64((const uint64_t *)left);
+
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 0);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 1);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 2);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 3);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 4);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 5);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 6);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 7);
+  vst1_u8(dst, d0u8);
+}
+
+void vpx_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int j;
+  uint8x8_t d2u8 = vdup_n_u8(0);
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)above;
+
+  q1u8 = vld1q_u8(left);
+  d2u8 = vget_low_u8(q1u8);
+  for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) {
+    q0u8 = vdupq_lane_u8(d2u8, 0);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 1);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 2);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 3);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 4);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 5);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 6);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 7);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+  }
+}
+
+void vpx_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint8x8_t d2u8 = vdup_n_u8(0);
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)above;
+
+  for (k = 0; k < 2; k++, left += 16) {
+    q1u8 = vld1q_u8(left);
+    d2u8 = vget_low_u8(q1u8);
+    for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) {
+      q0u8 = vdupq_lane_u8(d2u8, 0);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 1);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 2);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 3);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 4);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 5);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 6);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 7);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+    }
+  }
+}
+
+void vpx_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint16x8_t q1u16, q3u16;
+  int16x8_t q1s16;
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint32x2_t d2u32 = vdup_n_u32(0);
+
+  d0u8 = vld1_dup_u8(above - 1);
+  d2u32 = vld1_lane_u32((const uint32_t *)above, d2u32, 0);
+  q3u16 = vsubl_u8(vreinterpret_u8_u32(d2u32), d0u8);
+  for (i = 0; i < 4; i++, dst += stride) {
+    q1u16 = vdupq_n_u16((uint16_t)left[i]);
+    q1s16 = vaddq_s16(vreinterpretq_s16_u16(q1u16),
+                      vreinterpretq_s16_u16(q3u16));
+    d0u8 = vqmovun_s16(q1s16);
+    vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  }
+}
+
+void vpx_tm_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int j;
+  uint16x8_t q0u16, q3u16, q10u16;
+  int16x8_t q0s16;
+  uint16x4_t d20u16;
+  uint8x8_t d0u8, d2u8, d30u8;
+
+  d0u8 = vld1_dup_u8(above - 1);
+  d30u8 = vld1_u8(left);
+  d2u8 = vld1_u8(above);
+  q10u16 = vmovl_u8(d30u8);
+  q3u16 = vsubl_u8(d2u8, d0u8);
+  d20u16 = vget_low_u16(q10u16);
+  for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
+    q0u16 = vdupq_lane_u16(d20u16, 0);
+    q0s16 = vaddq_s16(vreinterpretq_s16_u16(q3u16),
+                      vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 1);
+    q0s16 = vaddq_s16(vreinterpretq_s16_u16(q3u16),
+                      vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 2);
+    q0s16 = vaddq_s16(vreinterpretq_s16_u16(q3u16),
+                      vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 3);
+    q0s16 = vaddq_s16(vreinterpretq_s16_u16(q3u16),
+                      vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+  }
+}
+
+void vpx_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint16x8_t q0u16, q2u16, q3u16, q8u16, q10u16;
+  uint8x16_t q0u8, q1u8;
+  int16x8_t q0s16, q1s16, q8s16, q11s16;
+  uint16x4_t d20u16;
+  uint8x8_t d2u8, d3u8, d18u8, d22u8, d23u8;
+
+  q0u8 = vld1q_dup_u8(above - 1);
+  q1u8 = vld1q_u8(above);
+  q2u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
+  q3u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
+  for (k = 0; k < 2; k++, left += 8) {
+    d18u8 = vld1_u8(left);
+    q10u16 = vmovl_u8(d18u8);
+    d20u16 = vget_low_u16(q10u16);
+    for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
+      q0u16 = vdupq_lane_u16(d20u16, 0);
+      q8u16 = vdupq_lane_u16(d20u16, 1);
+      q1s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                        vreinterpretq_s16_u16(q2u16));
+      q0s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                        vreinterpretq_s16_u16(q3u16));
+      q11s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
+                         vreinterpretq_s16_u16(q2u16));
+      q8s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
+                        vreinterpretq_s16_u16(q3u16));
+      d2u8 = vqmovun_s16(q1s16);
+      d3u8 = vqmovun_s16(q0s16);
+      d22u8 = vqmovun_s16(q11s16);
+      d23u8 = vqmovun_s16(q8s16);
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
+      dst += stride;
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d20u16, 2);
+      q8u16 = vdupq_lane_u16(d20u16, 3);
+      q1s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                        vreinterpretq_s16_u16(q2u16));
+      q0s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                        vreinterpretq_s16_u16(q3u16));
+      q11s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
+                         vreinterpretq_s16_u16(q2u16));
+      q8s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
+                        vreinterpretq_s16_u16(q3u16));
+      d2u8 = vqmovun_s16(q1s16);
+      d3u8 = vqmovun_s16(q0s16);
+      d22u8 = vqmovun_s16(q11s16);
+      d23u8 = vqmovun_s16(q8s16);
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
+      dst += stride;
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
+      dst += stride;
+    }
+  }
+}
+
+void vpx_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint16x8_t q0u16, q3u16, q8u16, q9u16, q10u16, q11u16;
+  uint8x16_t q0u8, q1u8, q2u8;
+  int16x8_t q12s16, q13s16, q14s16, q15s16;
+  uint16x4_t d6u16;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8, d26u8;
+
+  q0u8 = vld1q_dup_u8(above - 1);
+  q1u8 = vld1q_u8(above);
+  q2u8 = vld1q_u8(above + 16);
+  q8u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
+  q9u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
+  q10u16 = vsubl_u8(vget_low_u8(q2u8), vget_low_u8(q0u8));
+  q11u16 = vsubl_u8(vget_high_u8(q2u8), vget_high_u8(q0u8));
+  for (k = 0; k < 4; k++, left += 8) {
+    d26u8 = vld1_u8(left);
+    q3u16 = vmovl_u8(d26u8);
+    d6u16 = vget_low_u16(q3u16);
+    for (j = 0; j < 2; j++, d6u16 = vget_high_u16(q3u16)) {
+      q0u16 = vdupq_lane_u16(d6u16, 0);
+      q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q8u16));
+      q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 1);
+      q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q8u16));
+      q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 2);
+      q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q8u16));
+      q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 3);
+      q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q8u16));
+      q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+    }
+  }
+}
+#endif  // !HAVE_NEON_ASM
diff --git a/libvpx/libvpx/vpx_dsp/arm/intrapred_neon_asm.asm b/libvpx/libvpx/vpx_dsp/arm/intrapred_neon_asm.asm
new file mode 100644
index 0000000..115790d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/intrapred_neon_asm.asm
@@ -0,0 +1,630 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vpx_v_predictor_4x4_neon|
+    EXPORT  |vpx_v_predictor_8x8_neon|
+    EXPORT  |vpx_v_predictor_16x16_neon|
+    EXPORT  |vpx_v_predictor_32x32_neon|
+    EXPORT  |vpx_h_predictor_4x4_neon|
+    EXPORT  |vpx_h_predictor_8x8_neon|
+    EXPORT  |vpx_h_predictor_16x16_neon|
+    EXPORT  |vpx_h_predictor_32x32_neon|
+    EXPORT  |vpx_tm_predictor_4x4_neon|
+    EXPORT  |vpx_tm_predictor_8x8_neon|
+    EXPORT  |vpx_tm_predictor_16x16_neon|
+    EXPORT  |vpx_tm_predictor_32x32_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vpx_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_v_predictor_4x4_neon| PROC
+    vld1.32             {d0[0]}, [r2]
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |vpx_v_predictor_4x4_neon|
+
+;void vpx_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_v_predictor_8x8_neon| PROC
+    vld1.8              {d0}, [r2]
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    bx                  lr
+    ENDP                ; |vpx_v_predictor_8x8_neon|
+
+;void vpx_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_v_predictor_16x16_neon| PROC
+    vld1.8              {q0}, [r2]
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    bx                  lr
+    ENDP                ; |vpx_v_predictor_16x16_neon|
+
+;void vpx_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_v_predictor_32x32_neon| PROC
+    vld1.8              {q0, q1}, [r2]
+    mov                 r2, #2
+loop_v
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    subs                r2, r2, #1
+    bgt                 loop_v
+    bx                  lr
+    ENDP                ; |vpx_v_predictor_32x32_neon|
+
+;void vpx_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_h_predictor_4x4_neon| PROC
+    vld1.32             {d1[0]}, [r3]
+    vdup.8              d0, d1[0]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[1]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[2]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[3]
+    vst1.32             {d0[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |vpx_h_predictor_4x4_neon|
+
+;void vpx_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_h_predictor_8x8_neon| PROC
+    vld1.64             {d1}, [r3]
+    vdup.8              d0, d1[0]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[1]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[2]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[3]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[4]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[5]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[6]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[7]
+    vst1.64             {d0}, [r0], r1
+    bx                  lr
+    ENDP                ; |vpx_h_predictor_8x8_neon|
+
+;void vpx_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_h_predictor_16x16_neon| PROC
+    vld1.8              {q1}, [r3]
+    vdup.8              q0, d2[0]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[1]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[2]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[3]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[4]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[5]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[6]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[7]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[0]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[1]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[2]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[3]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[4]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[5]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[6]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[7]
+    vst1.8              {q0}, [r0], r1
+    bx                  lr
+    ENDP                ; |vpx_h_predictor_16x16_neon|
+
+;void vpx_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_h_predictor_32x32_neon| PROC
+    sub                 r1, r1, #16
+    mov                 r2, #2
+loop_h
+    vld1.8              {q1}, [r3]!
+    vdup.8              q0, d2[0]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[1]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[2]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[3]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[4]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[5]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[6]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[7]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[0]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[1]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[2]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[3]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[4]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[5]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[6]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[7]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    subs                r2, r2, #1
+    bgt                 loop_h
+    bx                  lr
+    ENDP                ; |vpx_h_predictor_32x32_neon|
+
+;void vpx_tm_predictor_4x4_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_tm_predictor_4x4_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.u8             {d0[]}, [r12]
+
+    ; Load above 4 pixels
+    vld1.32             {d2[0]}, [r2]
+
+    ; Compute above - ytop_left
+    vsubl.u8            q3, d2, d0
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; 1st row and 2nd row
+    vld1.u8             {d2[]}, [r3]!
+    vld1.u8             {d4[]}, [r3]!
+    vmovl.u8            q1, d2
+    vmovl.u8            q2, d4
+    vadd.s16            q1, q1, q3
+    vadd.s16            q2, q2, q3
+    vqmovun.s16         d0, q1
+    vqmovun.s16         d1, q2
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d1[0]}, [r0], r1
+
+    ; 3rd row and 4th row
+    vld1.u8             {d2[]}, [r3]!
+    vld1.u8             {d4[]}, [r3]
+    vmovl.u8            q1, d2
+    vmovl.u8            q2, d4
+    vadd.s16            q1, q1, q3
+    vadd.s16            q2, q2, q3
+    vqmovun.s16         d0, q1
+    vqmovun.s16         d1, q2
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d1[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |vpx_tm_predictor_4x4_neon|
+
+;void vpx_tm_predictor_8x8_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_tm_predictor_8x8_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; preload 8 left
+    vld1.8              {d30}, [r3]
+
+    ; Load above 8 pixels
+    vld1.64             {d2}, [r2]
+
+    vmovl.u8            q10, d30
+
+    ; Compute above - ytop_left
+    vsubl.u8            q3, d2, d0
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; 1st row and 2nd row
+    vdup.16             q0, d20[0]
+    vdup.16             q1, d20[1]
+    vadd.s16            q0, q3, q0
+    vadd.s16            q1, q3, q1
+
+    ; 3rd row and 4th row
+    vdup.16             q8, d20[2]
+    vdup.16             q9, d20[3]
+    vadd.s16            q8, q3, q8
+    vadd.s16            q9, q3, q9
+
+    vqmovun.s16         d0, q0
+    vqmovun.s16         d1, q1
+    vqmovun.s16         d2, q8
+    vqmovun.s16         d3, q9
+
+    vst1.64             {d0}, [r0], r1
+    vst1.64             {d1}, [r0], r1
+    vst1.64             {d2}, [r0], r1
+    vst1.64             {d3}, [r0], r1
+
+    ; 5th row and 6th row
+    vdup.16             q0, d21[0]
+    vdup.16             q1, d21[1]
+    vadd.s16            q0, q3, q0
+    vadd.s16            q1, q3, q1
+
+    ; 7th row and 8th row
+    vdup.16             q8, d21[2]
+    vdup.16             q9, d21[3]
+    vadd.s16            q8, q3, q8
+    vadd.s16            q9, q3, q9
+
+    vqmovun.s16         d0, q0
+    vqmovun.s16         d1, q1
+    vqmovun.s16         d2, q8
+    vqmovun.s16         d3, q9
+
+    vst1.64             {d0}, [r0], r1
+    vst1.64             {d1}, [r0], r1
+    vst1.64             {d2}, [r0], r1
+    vst1.64             {d3}, [r0], r1
+
+    bx                  lr
+    ENDP                ; |vpx_tm_predictor_8x8_neon|
+
+;void vpx_tm_predictor_16x16_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_tm_predictor_16x16_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; Load above 8 pixels
+    vld1.8              {q1}, [r2]
+
+    ; preload 8 left into r12
+    vld1.8              {d18}, [r3]!
+
+    ; Compute above - ytop_left
+    vsubl.u8            q2, d2, d0
+    vsubl.u8            q3, d3, d0
+
+    vmovl.u8            q10, d18
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; Process 8 rows in each single loop and loop 2 times to process 16 rows.
+    mov                 r2, #2
+
+loop_16x16_neon
+    ; Process two rows.
+    vdup.16             q0, d20[0]
+    vdup.16             q8, d20[1]
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d20[2]                  ; proload next 2 rows data
+    vdup.16             q8, d20[3]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d21[0]                  ; proload next 2 rows data
+    vdup.16             q8, d21[1]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d21[2]                  ; proload next 2 rows data
+    vdup.16             q8, d21[3]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vld1.8              {d18}, [r3]!                  ; preload 8 left into r12
+    vmovl.u8            q10, d18
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    subs                r2, r2, #1
+    bgt                 loop_16x16_neon
+
+    bx                  lr
+    ENDP                ; |vpx_tm_predictor_16x16_neon|
+
+;void vpx_tm_predictor_32x32_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                  const uint8_t *above,
+;                                  const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vpx_tm_predictor_32x32_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; Load above 32 pixels
+    vld1.8              {q1}, [r2]!
+    vld1.8              {q2}, [r2]
+
+    ; preload 8 left pixels
+    vld1.8              {d26}, [r3]!
+
+    ; Compute above - ytop_left
+    vsubl.u8            q8, d2, d0
+    vsubl.u8            q9, d3, d0
+    vsubl.u8            q10, d4, d0
+    vsubl.u8            q11, d5, d0
+
+    vmovl.u8            q3, d26
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; Process 8 rows in each single loop and loop 4 times to process 32 rows.
+    mov                 r2, #4
+
+loop_32x32_neon
+    ; Process two rows.
+    vdup.16             q0, d6[0]
+    vdup.16             q2, d6[1]
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q1, d6[2]
+    vdup.16             q2, d6[3]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q1, q8
+    vadd.s16            q13, q1, q9
+    vadd.s16            q14, q1, q10
+    vadd.s16            q15, q1, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q0, d7[0]
+    vdup.16             q2, d7[1]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q0, d7[2]
+    vdup.16             q2, d7[3]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vld1.8              {d0}, [r3]!                   ; preload 8 left pixels
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vmovl.u8            q3, d0
+    vst1.64             {d24-d27}, [r0], r1
+
+    subs                r2, r2, #1
+    bgt                 loop_32x32_neon
+
+    bx                  lr
+    ENDP                ; |vpx_tm_predictor_32x32_neon|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/loopfilter_16_neon.asm b/libvpx/libvpx/vpx_dsp/arm/loopfilter_16_neon.asm
new file mode 100644
index 0000000..5a8fdd6
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/loopfilter_16_neon.asm
@@ -0,0 +1,199 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vpx_lpf_horizontal_4_dual_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vpx_lpf_horizontal_4_dual_neon(uint8_t *s, int p,
+;                                    const uint8_t *blimit0,
+;                                    const uint8_t *limit0,
+;                                    const uint8_t *thresh0,
+;                                    const uint8_t *blimit1,
+;                                    const uint8_t *limit1,
+;                                    const uint8_t *thresh1)
+; r0    uint8_t *s,
+; r1    int p,
+; r2    const uint8_t *blimit0,
+; r3    const uint8_t *limit0,
+; sp    const uint8_t *thresh0,
+; sp+4  const uint8_t *blimit1,
+; sp+8  const uint8_t *limit1,
+; sp+12 const uint8_t *thresh1,
+
+|vpx_lpf_horizontal_4_dual_neon| PROC
+    push        {lr}
+
+    ldr         r12, [sp, #4]              ; load thresh0
+    vld1.8      {d0}, [r2]                 ; load blimit0 to first half q
+    vld1.8      {d2}, [r3]                 ; load limit0 to first half q
+
+    add         r1, r1, r1                 ; double pitch
+    ldr         r2, [sp, #8]               ; load blimit1
+
+    vld1.8      {d4}, [r12]                ; load thresh0 to first half q
+
+    ldr         r3, [sp, #12]              ; load limit1
+    ldr         r12, [sp, #16]             ; load thresh1
+    vld1.8      {d1}, [r2]                 ; load blimit1 to 2nd half q
+
+    sub         r2, r0, r1, lsl #1         ; s[-4 * p]
+
+    vld1.8      {d3}, [r3]                 ; load limit1 to 2nd half q
+    vld1.8      {d5}, [r12]                ; load thresh1 to 2nd half q
+
+    vpush       {d8-d15}                   ; save neon registers
+
+    add         r3, r2, r1, lsr #1         ; s[-3 * p]
+
+    vld1.u8     {q3}, [r2@64], r1          ; p3
+    vld1.u8     {q4}, [r3@64], r1          ; p2
+    vld1.u8     {q5}, [r2@64], r1          ; p1
+    vld1.u8     {q6}, [r3@64], r1          ; p0
+    vld1.u8     {q7}, [r2@64], r1          ; q0
+    vld1.u8     {q8}, [r3@64], r1          ; q1
+    vld1.u8     {q9}, [r2@64]              ; q2
+    vld1.u8     {q10}, [r3@64]             ; q3
+
+    sub         r2, r2, r1, lsl #1
+    sub         r3, r3, r1, lsl #1
+
+    bl          vpx_loop_filter_neon_16
+
+    vst1.u8     {q5}, [r2@64], r1          ; store op1
+    vst1.u8     {q6}, [r3@64], r1          ; store op0
+    vst1.u8     {q7}, [r2@64], r1          ; store oq0
+    vst1.u8     {q8}, [r3@64], r1          ; store oq1
+
+    vpop        {d8-d15}                   ; restore neon registers
+
+    pop         {pc}
+    ENDP        ; |vpx_lpf_horizontal_4_dual_neon|
+
+; void vpx_loop_filter_neon_16();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. This function uses
+; registers d8-d15, so the calling function must save those registers.
+;
+; r0-r3, r12 PRESERVE
+; q0    blimit
+; q1    limit
+; q2    thresh
+; q3    p3
+; q4    p2
+; q5    p1
+; q6    p0
+; q7    q0
+; q8    q1
+; q9    q2
+; q10   q3
+;
+; Outputs:
+; q5    op1
+; q6    op0
+; q7    oq0
+; q8    oq1
+|vpx_loop_filter_neon_16| PROC
+
+    ; filter_mask
+    vabd.u8     q11, q3, q4                 ; m1 = abs(p3 - p2)
+    vabd.u8     q12, q4, q5                 ; m2 = abs(p2 - p1)
+    vabd.u8     q13, q5, q6                 ; m3 = abs(p1 - p0)
+    vabd.u8     q14, q8, q7                 ; m4 = abs(q1 - q0)
+    vabd.u8     q3, q9, q8                  ; m5 = abs(q2 - q1)
+    vabd.u8     q4, q10, q9                 ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     q11, q11, q12               ; m7 = max(m1, m2)
+    vmax.u8     q12, q13, q14               ; m8 = max(m3, m4)
+
+    vabd.u8     q9, q6, q7                  ; abs(p0 - q0)
+
+    vmax.u8     q3, q3, q4                  ; m9 = max(m5, m6)
+
+    vmov.u8     q10, #0x80
+
+    vmax.u8     q15, q11, q12               ; m10 = max(m7, m8)
+
+    vcgt.u8     q13, q13, q2                ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     q14, q14, q2                ; (abs(q1 - q0) > thresh)*-1
+    vmax.u8     q15, q15, q3                ; m11 = max(m10, m9)
+
+    vabd.u8     q2, q5, q8                  ; a = abs(p1 - q1)
+    vqadd.u8    q9, q9, q9                  ; b = abs(p0 - q0) * 2
+
+    veor        q7, q7, q10                 ; qs0
+
+    vcge.u8     q15, q1, q15                ; abs(m11) > limit
+
+    vshr.u8     q2, q2, #1                  ; a = a / 2
+    veor        q6, q6, q10                 ; ps0
+
+    veor        q5, q5, q10                 ; ps1
+    vqadd.u8    q9, q9, q2                  ; a = b + a
+
+    veor        q8, q8, q10                 ; qs1
+
+    vmov.u16    q4, #3
+
+    vsubl.s8    q2, d14, d12                ; ( qs0 - ps0)
+    vsubl.s8    q11, d15, d13
+
+    vcge.u8     q9, q0, q9                  ; a > blimit
+
+    vqsub.s8    q1, q5, q8                  ; filter = clamp(ps1-qs1)
+    vorr        q14, q13, q14               ; hev
+
+    vmul.i16    q2, q2, q4                  ; 3 * ( qs0 - ps0)
+    vmul.i16    q11, q11, q4
+
+    vand        q1, q1, q14                 ; filter &= hev
+    vand        q15, q15, q9                ; mask
+
+    vmov.u8     q4, #3
+
+    vaddw.s8    q2, q2, d2                  ; filter + 3 * (qs0 - ps0)
+    vaddw.s8    q11, q11, d3
+
+    vmov.u8     q9, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d2, q2
+    vqmovn.s16  d3, q11
+    vand        q1, q1, q15                 ; filter &= mask
+
+    vqadd.s8    q2, q1, q4                  ; filter2 = clamp(filter+3)
+    vqadd.s8    q1, q1, q9                  ; filter1 = clamp(filter+4)
+    vshr.s8     q2, q2, #3                  ; filter2 >>= 3
+    vshr.s8     q1, q1, #3                  ; filter1 >>= 3
+
+
+    vqadd.s8    q11, q6, q2                 ; u = clamp(ps0 + filter2)
+    vqsub.s8    q0, q7, q1                  ; u = clamp(qs0 - filter1)
+
+    ; outer tap adjustments
+    vrshr.s8    q1, q1, #1                  ; filter = ++filter1 >> 1
+
+    veor        q7, q0,  q10                ; *oq0 = u^0x80
+
+    vbic        q1, q1, q14                 ; filter &= ~hev
+
+    vqadd.s8    q13, q5, q1                 ; u = clamp(ps1 + filter)
+    vqsub.s8    q12, q8, q1                 ; u = clamp(qs1 - filter)
+
+    veor        q6, q11, q10                ; *op0 = u^0x80
+    veor        q5, q13, q10                ; *op1 = u^0x80
+    veor        q8, q12, q10                ; *oq1 = u^0x80
+
+    bx          lr
+    ENDP        ; |vpx_loop_filter_neon_16|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/loopfilter_16_neon.c b/libvpx/libvpx/vpx_dsp/arm/loopfilter_16_neon.c
new file mode 100644
index 0000000..d24e6ad
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/loopfilter_16_neon.c
@@ -0,0 +1,179 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+static INLINE void loop_filter_neon_16(
+        uint8x16_t qblimit,  // blimit
+        uint8x16_t qlimit,   // limit
+        uint8x16_t qthresh,  // thresh
+        uint8x16_t q3,       // p3
+        uint8x16_t q4,       // p2
+        uint8x16_t q5,       // p1
+        uint8x16_t q6,       // p0
+        uint8x16_t q7,       // q0
+        uint8x16_t q8,       // q1
+        uint8x16_t q9,       // q2
+        uint8x16_t q10,      // q3
+        uint8x16_t *q5r,     // p1
+        uint8x16_t *q6r,     // p0
+        uint8x16_t *q7r,     // q0
+        uint8x16_t *q8r) {   // q1
+    uint8x16_t q1u8, q2u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+    int16x8_t q2s16, q11s16;
+    uint16x8_t q4u16;
+    int8x16_t q0s8, q1s8, q2s8, q11s8, q12s8, q13s8;
+    int8x8_t d2s8, d3s8;
+
+    q11u8 = vabdq_u8(q3, q4);
+    q12u8 = vabdq_u8(q4, q5);
+    q13u8 = vabdq_u8(q5, q6);
+    q14u8 = vabdq_u8(q8, q7);
+    q3 = vabdq_u8(q9, q8);
+    q4 = vabdq_u8(q10, q9);
+
+    q11u8 = vmaxq_u8(q11u8, q12u8);
+    q12u8 = vmaxq_u8(q13u8, q14u8);
+    q3 = vmaxq_u8(q3, q4);
+    q15u8 = vmaxq_u8(q11u8, q12u8);
+
+    q9 = vabdq_u8(q6, q7);
+
+    // vp8_hevmask
+    q13u8 = vcgtq_u8(q13u8, qthresh);
+    q14u8 = vcgtq_u8(q14u8, qthresh);
+    q15u8 = vmaxq_u8(q15u8, q3);
+
+    q2u8 = vabdq_u8(q5, q8);
+    q9 = vqaddq_u8(q9, q9);
+
+    q15u8 = vcgeq_u8(qlimit, q15u8);
+
+    // vp8_filter() function
+    // convert to signed
+    q10 = vdupq_n_u8(0x80);
+    q8 = veorq_u8(q8, q10);
+    q7 = veorq_u8(q7, q10);
+    q6 = veorq_u8(q6, q10);
+    q5 = veorq_u8(q5, q10);
+
+    q2u8 = vshrq_n_u8(q2u8, 1);
+    q9 = vqaddq_u8(q9, q2u8);
+
+    q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
+                     vget_low_s8(vreinterpretq_s8_u8(q6)));
+    q11s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
+                      vget_high_s8(vreinterpretq_s8_u8(q6)));
+
+    q9 = vcgeq_u8(qblimit, q9);
+
+    q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5),
+                    vreinterpretq_s8_u8(q8));
+
+    q14u8 = vorrq_u8(q13u8, q14u8);
+
+    q4u16 = vdupq_n_u16(3);
+    q2s16 = vmulq_s16(q2s16, vreinterpretq_s16_u16(q4u16));
+    q11s16 = vmulq_s16(q11s16, vreinterpretq_s16_u16(q4u16));
+
+    q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q14u8);
+    q15u8 = vandq_u8(q15u8, q9);
+
+    q1s8 = vreinterpretq_s8_u8(q1u8);
+    q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
+    q11s16 = vaddw_s8(q11s16, vget_high_s8(q1s8));
+
+    q4 = vdupq_n_u8(3);
+    q9 = vdupq_n_u8(4);
+    // vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
+    d2s8 = vqmovn_s16(q2s16);
+    d3s8 = vqmovn_s16(q11s16);
+    q1s8 = vcombine_s8(d2s8, d3s8);
+    q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q15u8);
+    q1s8 = vreinterpretq_s8_u8(q1u8);
+
+    q2s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q4));
+    q1s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q9));
+    q2s8 = vshrq_n_s8(q2s8, 3);
+    q1s8 = vshrq_n_s8(q1s8, 3);
+
+    q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8);
+    q0s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q1s8);
+
+    q1s8 = vrshrq_n_s8(q1s8, 1);
+    q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+    q13s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q1s8);
+    q12s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q1s8);
+
+    *q8r = veorq_u8(vreinterpretq_u8_s8(q12s8), q10);
+    *q7r = veorq_u8(vreinterpretq_u8_s8(q0s8),  q10);
+    *q6r = veorq_u8(vreinterpretq_u8_s8(q11s8), q10);
+    *q5r = veorq_u8(vreinterpretq_u8_s8(q13s8), q10);
+    return;
+}
+
+void vpx_lpf_horizontal_4_dual_neon(uint8_t *s, int p /* pitch */,
+                                    const uint8_t *blimit0,
+                                    const uint8_t *limit0,
+                                    const uint8_t *thresh0,
+                                    const uint8_t *blimit1,
+                                    const uint8_t *limit1,
+                                    const uint8_t *thresh1) {
+    uint8x8_t dblimit0, dlimit0, dthresh0, dblimit1, dlimit1, dthresh1;
+    uint8x16_t qblimit, qlimit, qthresh;
+    uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8;
+
+    dblimit0 = vld1_u8(blimit0);
+    dlimit0 = vld1_u8(limit0);
+    dthresh0 = vld1_u8(thresh0);
+    dblimit1 = vld1_u8(blimit1);
+    dlimit1 = vld1_u8(limit1);
+    dthresh1 = vld1_u8(thresh1);
+    qblimit = vcombine_u8(dblimit0, dblimit1);
+    qlimit = vcombine_u8(dlimit0, dlimit1);
+    qthresh = vcombine_u8(dthresh0, dthresh1);
+
+    s -= (p << 2);
+
+    q3u8 = vld1q_u8(s);
+    s += p;
+    q4u8 = vld1q_u8(s);
+    s += p;
+    q5u8 = vld1q_u8(s);
+    s += p;
+    q6u8 = vld1q_u8(s);
+    s += p;
+    q7u8 = vld1q_u8(s);
+    s += p;
+    q8u8 = vld1q_u8(s);
+    s += p;
+    q9u8 = vld1q_u8(s);
+    s += p;
+    q10u8 = vld1q_u8(s);
+
+    loop_filter_neon_16(qblimit, qlimit, qthresh,
+                        q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8,
+                        &q5u8, &q6u8, &q7u8, &q8u8);
+
+    s -= (p * 5);
+    vst1q_u8(s, q5u8);
+    s += p;
+    vst1q_u8(s, q6u8);
+    s += p;
+    vst1q_u8(s, q7u8);
+    s += p;
+    vst1q_u8(s, q8u8);
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/loopfilter_4_neon.asm b/libvpx/libvpx/vpx_dsp/arm/loopfilter_4_neon.asm
new file mode 100644
index 0000000..9371158
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/loopfilter_4_neon.asm
@@ -0,0 +1,249 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vpx_lpf_horizontal_4_neon|
+    EXPORT  |vpx_lpf_vertical_4_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; Currently vpx only works on iterations 8 at a time. The vp8 loop filter
+; works on 16 iterations at a time.
+;
+; void vpx_lpf_horizontal_4_neon(uint8_t *s,
+;                                int p /* pitch */,
+;                                const uint8_t *blimit,
+;                                const uint8_t *limit,
+;                                const uint8_t *thresh)
+;
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|vpx_lpf_horizontal_4_neon| PROC
+    push        {lr}
+
+    vld1.8      {d0[]}, [r2]               ; duplicate *blimit
+    ldr         r2, [sp, #4]               ; load thresh
+    add         r1, r1, r1                 ; double pitch
+
+    vld1.8      {d1[]}, [r3]               ; duplicate *limit
+    vld1.8      {d2[]}, [r2]               ; duplicate *thresh
+
+    sub         r2, r0, r1, lsl #1         ; move src pointer down by 4 lines
+    add         r3, r2, r1, lsr #1         ; set to 3 lines down
+
+    vld1.u8     {d3}, [r2@64], r1          ; p3
+    vld1.u8     {d4}, [r3@64], r1          ; p2
+    vld1.u8     {d5}, [r2@64], r1          ; p1
+    vld1.u8     {d6}, [r3@64], r1          ; p0
+    vld1.u8     {d7}, [r2@64], r1          ; q0
+    vld1.u8     {d16}, [r3@64], r1         ; q1
+    vld1.u8     {d17}, [r2@64]             ; q2
+    vld1.u8     {d18}, [r3@64]             ; q3
+
+    sub         r2, r2, r1, lsl #1
+    sub         r3, r3, r1, lsl #1
+
+    bl          vpx_loop_filter_neon
+
+    vst1.u8     {d4}, [r2@64], r1          ; store op1
+    vst1.u8     {d5}, [r3@64], r1          ; store op0
+    vst1.u8     {d6}, [r2@64], r1          ; store oq0
+    vst1.u8     {d7}, [r3@64], r1          ; store oq1
+
+    pop         {pc}
+    ENDP        ; |vpx_lpf_horizontal_4_neon|
+
+; Currently vpx only works on iterations 8 at a time. The vp8 loop filter
+; works on 16 iterations at a time.
+;
+; void vpx_lpf_vertical_4_neon(uint8_t *s,
+;                              int p /* pitch */,
+;                              const uint8_t *blimit,
+;                              const uint8_t *limit,
+;                              const uint8_t *thresh)
+;
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|vpx_lpf_vertical_4_neon| PROC
+    push        {lr}
+
+    vld1.8      {d0[]}, [r2]              ; duplicate *blimit
+    vld1.8      {d1[]}, [r3]              ; duplicate *limit
+
+    ldr         r3, [sp, #4]              ; load thresh
+    sub         r2, r0, #4                ; move s pointer down by 4 columns
+
+    vld1.8      {d2[]}, [r3]              ; duplicate *thresh
+
+    vld1.u8     {d3}, [r2], r1             ; load s data
+    vld1.u8     {d4}, [r2], r1
+    vld1.u8     {d5}, [r2], r1
+    vld1.u8     {d6}, [r2], r1
+    vld1.u8     {d7}, [r2], r1
+    vld1.u8     {d16}, [r2], r1
+    vld1.u8     {d17}, [r2], r1
+    vld1.u8     {d18}, [r2]
+
+    ;transpose to 8x16 matrix
+    vtrn.32     d3, d7
+    vtrn.32     d4, d16
+    vtrn.32     d5, d17
+    vtrn.32     d6, d18
+
+    vtrn.16     d3, d5
+    vtrn.16     d4, d6
+    vtrn.16     d7, d17
+    vtrn.16     d16, d18
+
+    vtrn.8      d3, d4
+    vtrn.8      d5, d6
+    vtrn.8      d7, d16
+    vtrn.8      d17, d18
+
+    bl          vpx_loop_filter_neon
+
+    sub         r0, r0, #2
+
+    ;store op1, op0, oq0, oq1
+    vst4.8      {d4[0], d5[0], d6[0], d7[0]}, [r0], r1
+    vst4.8      {d4[1], d5[1], d6[1], d7[1]}, [r0], r1
+    vst4.8      {d4[2], d5[2], d6[2], d7[2]}, [r0], r1
+    vst4.8      {d4[3], d5[3], d6[3], d7[3]}, [r0], r1
+    vst4.8      {d4[4], d5[4], d6[4], d7[4]}, [r0], r1
+    vst4.8      {d4[5], d5[5], d6[5], d7[5]}, [r0], r1
+    vst4.8      {d4[6], d5[6], d6[6], d7[6]}, [r0], r1
+    vst4.8      {d4[7], d5[7], d6[7], d7[7]}, [r0]
+
+    pop         {pc}
+    ENDP        ; |vpx_lpf_vertical_4_neon|
+
+; void vpx_loop_filter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. The function does not use
+; registers d8-d15.
+;
+; Inputs:
+; r0-r3, r12 PRESERVE
+; d0    blimit
+; d1    limit
+; d2    thresh
+; d3    p3
+; d4    p2
+; d5    p1
+; d6    p0
+; d7    q0
+; d16   q1
+; d17   q2
+; d18   q3
+;
+; Outputs:
+; d4    op1
+; d5    op0
+; d6    oq0
+; d7    oq1
+|vpx_loop_filter_neon| PROC
+    ; filter_mask
+    vabd.u8     d19, d3, d4                 ; m1 = abs(p3 - p2)
+    vabd.u8     d20, d4, d5                 ; m2 = abs(p2 - p1)
+    vabd.u8     d21, d5, d6                 ; m3 = abs(p1 - p0)
+    vabd.u8     d22, d16, d7                ; m4 = abs(q1 - q0)
+    vabd.u8     d3, d17, d16                ; m5 = abs(q2 - q1)
+    vabd.u8     d4, d18, d17                ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20               ; m1 = max(m1, m2)
+    vmax.u8     d20, d21, d22               ; m2 = max(m3, m4)
+
+    vabd.u8     d17, d6, d7                 ; abs(p0 - q0)
+
+    vmax.u8     d3, d3, d4                  ; m3 = max(m5, m6)
+
+    vmov.u8     d18, #0x80
+
+    vmax.u8     d23, d19, d20               ; m1 = max(m1, m2)
+
+    ; hevmask
+    vcgt.u8     d21, d21, d2                ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     d22, d22, d2                ; (abs(q1 - q0) > thresh)*-1
+    vmax.u8     d23, d23, d3                ; m1 = max(m1, m3)
+
+    vabd.u8     d28, d5, d16                ; a = abs(p1 - q1)
+    vqadd.u8    d17, d17, d17               ; b = abs(p0 - q0) * 2
+
+    veor        d7, d7, d18                 ; qs0
+
+    vcge.u8     d23, d1, d23                ; abs(m1) > limit
+
+    ; filter() function
+    ; convert to signed
+
+    vshr.u8     d28, d28, #1                ; a = a / 2
+    veor        d6, d6, d18                 ; ps0
+
+    veor        d5, d5, d18                 ; ps1
+    vqadd.u8    d17, d17, d28               ; a = b + a
+
+    veor        d16, d16, d18               ; qs1
+
+    vmov.u8     d19, #3
+
+    vsub.s8     d28, d7, d6                 ; ( qs0 - ps0)
+
+    vcge.u8     d17, d0, d17                ; a > blimit
+
+    vqsub.s8    d27, d5, d16                ; filter = clamp(ps1-qs1)
+    vorr        d22, d21, d22               ; hevmask
+
+    vmull.s8    q12, d28, d19               ; 3 * ( qs0 - ps0)
+
+    vand        d27, d27, d22               ; filter &= hev
+    vand        d23, d23, d17               ; filter_mask
+
+    vaddw.s8    q12, q12, d27               ; filter + 3 * (qs0 - ps0)
+
+    vmov.u8     d17, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d27, q12
+
+    vand        d27, d27, d23               ; filter &= mask
+
+    vqadd.s8    d28, d27, d19               ; filter2 = clamp(filter+3)
+    vqadd.s8    d27, d27, d17               ; filter1 = clamp(filter+4)
+    vshr.s8     d28, d28, #3                ; filter2 >>= 3
+    vshr.s8     d27, d27, #3                ; filter1 >>= 3
+
+    vqadd.s8    d19, d6, d28                ; u = clamp(ps0 + filter2)
+    vqsub.s8    d26, d7, d27                ; u = clamp(qs0 - filter1)
+
+    ; outer tap adjustments
+    vrshr.s8    d27, d27, #1                ; filter = ++filter1 >> 1
+
+    veor        d6, d26, d18                ; *oq0 = u^0x80
+
+    vbic        d27, d27, d22               ; filter &= ~hev
+
+    vqadd.s8    d21, d5, d27                ; u = clamp(ps1 + filter)
+    vqsub.s8    d20, d16, d27               ; u = clamp(qs1 - filter)
+
+    veor        d5, d19, d18                ; *op0 = u^0x80
+    veor        d4, d21, d18                ; *op1 = u^0x80
+    veor        d7, d20, d18                ; *oq1 = u^0x80
+
+    bx          lr
+    ENDP        ; |vpx_loop_filter_neon|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/loopfilter_4_neon.c b/libvpx/libvpx/vpx_dsp/arm/loopfilter_4_neon.c
new file mode 100644
index 0000000..7f3ee70
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/loopfilter_4_neon.c
@@ -0,0 +1,266 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+
+static INLINE void loop_filter_neon(
+        uint8x8_t dblimit,    // flimit
+        uint8x8_t dlimit,     // limit
+        uint8x8_t dthresh,    // thresh
+        uint8x8_t d3u8,       // p3
+        uint8x8_t d4u8,       // p2
+        uint8x8_t d5u8,       // p1
+        uint8x8_t d6u8,       // p0
+        uint8x8_t d7u8,       // q0
+        uint8x8_t d16u8,      // q1
+        uint8x8_t d17u8,      // q2
+        uint8x8_t d18u8,      // q3
+        uint8x8_t *d4ru8,     // p1
+        uint8x8_t *d5ru8,     // p0
+        uint8x8_t *d6ru8,     // q0
+        uint8x8_t *d7ru8) {   // q1
+    uint8x8_t d19u8, d20u8, d21u8, d22u8, d23u8, d27u8, d28u8;
+    int16x8_t q12s16;
+    int8x8_t d19s8, d20s8, d21s8, d26s8, d27s8, d28s8;
+
+    d19u8 = vabd_u8(d3u8, d4u8);
+    d20u8 = vabd_u8(d4u8, d5u8);
+    d21u8 = vabd_u8(d5u8, d6u8);
+    d22u8 = vabd_u8(d16u8, d7u8);
+    d3u8  = vabd_u8(d17u8, d16u8);
+    d4u8  = vabd_u8(d18u8, d17u8);
+
+    d19u8 = vmax_u8(d19u8, d20u8);
+    d20u8 = vmax_u8(d21u8, d22u8);
+    d3u8  = vmax_u8(d3u8,  d4u8);
+    d23u8 = vmax_u8(d19u8, d20u8);
+
+    d17u8 = vabd_u8(d6u8, d7u8);
+
+    d21u8 = vcgt_u8(d21u8, dthresh);
+    d22u8 = vcgt_u8(d22u8, dthresh);
+    d23u8 = vmax_u8(d23u8, d3u8);
+
+    d28u8 = vabd_u8(d5u8, d16u8);
+    d17u8 = vqadd_u8(d17u8, d17u8);
+
+    d23u8 = vcge_u8(dlimit, d23u8);
+
+    d18u8 = vdup_n_u8(0x80);
+    d5u8  = veor_u8(d5u8,  d18u8);
+    d6u8  = veor_u8(d6u8,  d18u8);
+    d7u8  = veor_u8(d7u8,  d18u8);
+    d16u8 = veor_u8(d16u8, d18u8);
+
+    d28u8 = vshr_n_u8(d28u8, 1);
+    d17u8 = vqadd_u8(d17u8, d28u8);
+
+    d19u8 = vdup_n_u8(3);
+
+    d28s8 = vsub_s8(vreinterpret_s8_u8(d7u8),
+                    vreinterpret_s8_u8(d6u8));
+
+    d17u8 = vcge_u8(dblimit, d17u8);
+
+    d27s8 = vqsub_s8(vreinterpret_s8_u8(d5u8),
+                     vreinterpret_s8_u8(d16u8));
+
+    d22u8 = vorr_u8(d21u8, d22u8);
+
+    q12s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d19u8));
+
+    d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d22u8);
+    d23u8 = vand_u8(d23u8, d17u8);
+
+    q12s16 = vaddw_s8(q12s16, vreinterpret_s8_u8(d27u8));
+
+    d17u8 = vdup_n_u8(4);
+
+    d27s8 = vqmovn_s16(q12s16);
+    d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d23u8);
+    d27s8 = vreinterpret_s8_u8(d27u8);
+
+    d28s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d19u8));
+    d27s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d17u8));
+    d28s8 = vshr_n_s8(d28s8, 3);
+    d27s8 = vshr_n_s8(d27s8, 3);
+
+    d19s8 = vqadd_s8(vreinterpret_s8_u8(d6u8), d28s8);
+    d26s8 = vqsub_s8(vreinterpret_s8_u8(d7u8), d27s8);
+
+    d27s8 = vrshr_n_s8(d27s8, 1);
+    d27s8 = vbic_s8(d27s8, vreinterpret_s8_u8(d22u8));
+
+    d21s8 = vqadd_s8(vreinterpret_s8_u8(d5u8), d27s8);
+    d20s8 = vqsub_s8(vreinterpret_s8_u8(d16u8), d27s8);
+
+    *d4ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d18u8);
+    *d5ru8 = veor_u8(vreinterpret_u8_s8(d19s8), d18u8);
+    *d6ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d18u8);
+    *d7ru8 = veor_u8(vreinterpret_u8_s8(d20s8), d18u8);
+    return;
+}
+
+void vpx_lpf_horizontal_4_neon(
+        uint8_t *src,
+        int pitch,
+        const uint8_t *blimit,
+        const uint8_t *limit,
+        const uint8_t *thresh) {
+    int i;
+    uint8_t *s, *psrc;
+    uint8x8_t dblimit, dlimit, dthresh;
+    uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
+
+    dblimit = vld1_u8(blimit);
+    dlimit = vld1_u8(limit);
+    dthresh = vld1_u8(thresh);
+
+    psrc = src - (pitch << 2);
+    for (i = 0; i < 1; i++) {
+        s = psrc + i * 8;
+
+        d3u8 = vld1_u8(s);
+        s += pitch;
+        d4u8 = vld1_u8(s);
+        s += pitch;
+        d5u8 = vld1_u8(s);
+        s += pitch;
+        d6u8 = vld1_u8(s);
+        s += pitch;
+        d7u8 = vld1_u8(s);
+        s += pitch;
+        d16u8 = vld1_u8(s);
+        s += pitch;
+        d17u8 = vld1_u8(s);
+        s += pitch;
+        d18u8 = vld1_u8(s);
+
+        loop_filter_neon(dblimit, dlimit, dthresh,
+                         d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+                         &d4u8, &d5u8, &d6u8, &d7u8);
+
+        s -= (pitch * 5);
+        vst1_u8(s, d4u8);
+        s += pitch;
+        vst1_u8(s, d5u8);
+        s += pitch;
+        vst1_u8(s, d6u8);
+        s += pitch;
+        vst1_u8(s, d7u8);
+    }
+    return;
+}
+
+void vpx_lpf_vertical_4_neon(
+        uint8_t *src,
+        int pitch,
+        const uint8_t *blimit,
+        const uint8_t *limit,
+        const uint8_t *thresh) {
+    int i, pitch8;
+    uint8_t *s;
+    uint8x8_t dblimit, dlimit, dthresh;
+    uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
+    uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
+    uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
+    uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
+    uint8x8x4_t d4Result;
+
+    dblimit = vld1_u8(blimit);
+    dlimit = vld1_u8(limit);
+    dthresh = vld1_u8(thresh);
+
+    pitch8 = pitch * 8;
+    for (i = 0; i < 1; i++, src += pitch8) {
+        s = src - (i + 1) * 4;
+
+        d3u8 = vld1_u8(s);
+        s += pitch;
+        d4u8 = vld1_u8(s);
+        s += pitch;
+        d5u8 = vld1_u8(s);
+        s += pitch;
+        d6u8 = vld1_u8(s);
+        s += pitch;
+        d7u8 = vld1_u8(s);
+        s += pitch;
+        d16u8 = vld1_u8(s);
+        s += pitch;
+        d17u8 = vld1_u8(s);
+        s += pitch;
+        d18u8 = vld1_u8(s);
+
+        d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8),
+                      vreinterpret_u32_u8(d7u8));
+        d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8),
+                      vreinterpret_u32_u8(d16u8));
+        d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8),
+                      vreinterpret_u32_u8(d17u8));
+        d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8),
+                      vreinterpret_u32_u8(d18u8));
+
+        d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
+                          vreinterpret_u16_u32(d2tmp2.val[0]));
+        d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
+                          vreinterpret_u16_u32(d2tmp3.val[0]));
+        d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
+                          vreinterpret_u16_u32(d2tmp2.val[1]));
+        d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
+                          vreinterpret_u16_u32(d2tmp3.val[1]));
+
+        d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
+                         vreinterpret_u8_u16(d2tmp5.val[0]));
+        d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
+                         vreinterpret_u8_u16(d2tmp5.val[1]));
+        d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
+                          vreinterpret_u8_u16(d2tmp7.val[0]));
+        d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
+                          vreinterpret_u8_u16(d2tmp7.val[1]));
+
+        d3u8 = d2tmp8.val[0];
+        d4u8 = d2tmp8.val[1];
+        d5u8 = d2tmp9.val[0];
+        d6u8 = d2tmp9.val[1];
+        d7u8 = d2tmp10.val[0];
+        d16u8 = d2tmp10.val[1];
+        d17u8 = d2tmp11.val[0];
+        d18u8 = d2tmp11.val[1];
+
+        loop_filter_neon(dblimit, dlimit, dthresh,
+                         d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+                         &d4u8, &d5u8, &d6u8, &d7u8);
+
+        d4Result.val[0] = d4u8;
+        d4Result.val[1] = d5u8;
+        d4Result.val[2] = d6u8;
+        d4Result.val[3] = d7u8;
+
+        src -= 2;
+        vst4_lane_u8(src, d4Result, 0);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 1);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 2);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 3);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 4);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 5);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 6);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 7);
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/loopfilter_8_neon.asm b/libvpx/libvpx/vpx_dsp/arm/loopfilter_8_neon.asm
new file mode 100644
index 0000000..a2f20e1
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/loopfilter_8_neon.asm
@@ -0,0 +1,425 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vpx_lpf_horizontal_8_neon|
+    EXPORT  |vpx_lpf_vertical_8_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; Currently vpx only works on iterations 8 at a time. The vp8 loop filter
+; works on 16 iterations at a time.
+;
+; void vpx_lpf_horizontal_8_neon(uint8_t *s, int p,
+;                                const uint8_t *blimit,
+;                                const uint8_t *limit,
+;                                const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|vpx_lpf_horizontal_8_neon| PROC
+    push        {r4-r5, lr}
+
+    vld1.8      {d0[]}, [r2]               ; duplicate *blimit
+    ldr         r2, [sp, #12]              ; load thresh
+    add         r1, r1, r1                 ; double pitch
+
+    vld1.8      {d1[]}, [r3]               ; duplicate *limit
+    vld1.8      {d2[]}, [r2]               ; duplicate *thresh
+
+    sub         r3, r0, r1, lsl #1         ; move src pointer down by 4 lines
+    add         r2, r3, r1, lsr #1         ; set to 3 lines down
+
+    vld1.u8     {d3}, [r3@64], r1          ; p3
+    vld1.u8     {d4}, [r2@64], r1          ; p2
+    vld1.u8     {d5}, [r3@64], r1          ; p1
+    vld1.u8     {d6}, [r2@64], r1          ; p0
+    vld1.u8     {d7}, [r3@64], r1          ; q0
+    vld1.u8     {d16}, [r2@64], r1         ; q1
+    vld1.u8     {d17}, [r3@64]             ; q2
+    vld1.u8     {d18}, [r2@64], r1         ; q3
+
+    sub         r3, r3, r1, lsl #1
+    sub         r2, r2, r1, lsl #2
+
+    bl          vpx_mbloop_filter_neon
+
+    vst1.u8     {d0}, [r2@64], r1          ; store op2
+    vst1.u8     {d1}, [r3@64], r1          ; store op1
+    vst1.u8     {d2}, [r2@64], r1          ; store op0
+    vst1.u8     {d3}, [r3@64], r1          ; store oq0
+    vst1.u8     {d4}, [r2@64], r1          ; store oq1
+    vst1.u8     {d5}, [r3@64], r1          ; store oq2
+
+    pop         {r4-r5, pc}
+
+    ENDP        ; |vpx_lpf_horizontal_8_neon|
+
+; void vpx_lpf_vertical_8_neon(uint8_t *s,
+;                              int pitch,
+;                              const uint8_t *blimit,
+;                              const uint8_t *limit,
+;                              const uint8_t *thresh)
+;
+; r0    uint8_t *s,
+; r1    int pitch,
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|vpx_lpf_vertical_8_neon| PROC
+    push        {r4-r5, lr}
+
+    vld1.8      {d0[]}, [r2]              ; duplicate *blimit
+    vld1.8      {d1[]}, [r3]              ; duplicate *limit
+
+    ldr         r3, [sp, #12]             ; load thresh
+    sub         r2, r0, #4                ; move s pointer down by 4 columns
+
+    vld1.8      {d2[]}, [r3]              ; duplicate *thresh
+
+    vld1.u8     {d3}, [r2], r1             ; load s data
+    vld1.u8     {d4}, [r2], r1
+    vld1.u8     {d5}, [r2], r1
+    vld1.u8     {d6}, [r2], r1
+    vld1.u8     {d7}, [r2], r1
+    vld1.u8     {d16}, [r2], r1
+    vld1.u8     {d17}, [r2], r1
+    vld1.u8     {d18}, [r2]
+
+    ;transpose to 8x16 matrix
+    vtrn.32     d3, d7
+    vtrn.32     d4, d16
+    vtrn.32     d5, d17
+    vtrn.32     d6, d18
+
+    vtrn.16     d3, d5
+    vtrn.16     d4, d6
+    vtrn.16     d7, d17
+    vtrn.16     d16, d18
+
+    vtrn.8      d3, d4
+    vtrn.8      d5, d6
+    vtrn.8      d7, d16
+    vtrn.8      d17, d18
+
+    sub         r2, r0, #3
+    add         r3, r0, #1
+
+    bl          vpx_mbloop_filter_neon
+
+    ;store op2, op1, op0, oq0
+    vst4.8      {d0[0], d1[0], d2[0], d3[0]}, [r2], r1
+    vst4.8      {d0[1], d1[1], d2[1], d3[1]}, [r2], r1
+    vst4.8      {d0[2], d1[2], d2[2], d3[2]}, [r2], r1
+    vst4.8      {d0[3], d1[3], d2[3], d3[3]}, [r2], r1
+    vst4.8      {d0[4], d1[4], d2[4], d3[4]}, [r2], r1
+    vst4.8      {d0[5], d1[5], d2[5], d3[5]}, [r2], r1
+    vst4.8      {d0[6], d1[6], d2[6], d3[6]}, [r2], r1
+    vst4.8      {d0[7], d1[7], d2[7], d3[7]}, [r2]
+
+    ;store oq1, oq2
+    vst2.8      {d4[0], d5[0]}, [r3], r1
+    vst2.8      {d4[1], d5[1]}, [r3], r1
+    vst2.8      {d4[2], d5[2]}, [r3], r1
+    vst2.8      {d4[3], d5[3]}, [r3], r1
+    vst2.8      {d4[4], d5[4]}, [r3], r1
+    vst2.8      {d4[5], d5[5]}, [r3], r1
+    vst2.8      {d4[6], d5[6]}, [r3], r1
+    vst2.8      {d4[7], d5[7]}, [r3]
+
+    pop         {r4-r5, pc}
+    ENDP        ; |vpx_lpf_vertical_8_neon|
+
+; void vpx_mbloop_filter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. The function does not use
+; registers d8-d15.
+;
+; Inputs:
+; r0-r3, r12 PRESERVE
+; d0    blimit
+; d1    limit
+; d2    thresh
+; d3    p3
+; d4    p2
+; d5    p1
+; d6    p0
+; d7    q0
+; d16   q1
+; d17   q2
+; d18   q3
+;
+; Outputs:
+; d0    op2
+; d1    op1
+; d2    op0
+; d3    oq0
+; d4    oq1
+; d5    oq2
+|vpx_mbloop_filter_neon| PROC
+    ; filter_mask
+    vabd.u8     d19, d3, d4                ; m1 = abs(p3 - p2)
+    vabd.u8     d20, d4, d5                ; m2 = abs(p2 - p1)
+    vabd.u8     d21, d5, d6                ; m3 = abs(p1 - p0)
+    vabd.u8     d22, d16, d7               ; m4 = abs(q1 - q0)
+    vabd.u8     d23, d17, d16              ; m5 = abs(q2 - q1)
+    vabd.u8     d24, d18, d17              ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20              ; m1 = max(m1, m2)
+    vmax.u8     d20, d21, d22              ; m2 = max(m3, m4)
+
+    vabd.u8     d25, d6, d4                ; m7 = abs(p0 - p2)
+
+    vmax.u8     d23, d23, d24              ; m3 = max(m5, m6)
+
+    vabd.u8     d26, d7, d17               ; m8 = abs(q0 - q2)
+
+    vmax.u8     d19, d19, d20
+
+    vabd.u8     d24, d6, d7                ; m9 = abs(p0 - q0)
+    vabd.u8     d27, d3, d6                ; m10 = abs(p3 - p0)
+    vabd.u8     d28, d18, d7               ; m11 = abs(q3 - q0)
+
+    vmax.u8     d19, d19, d23
+
+    vabd.u8     d23, d5, d16               ; a = abs(p1 - q1)
+    vqadd.u8    d24, d24, d24              ; b = abs(p0 - q0) * 2
+
+    ; abs () > limit
+    vcge.u8     d19, d1, d19
+
+    ; only compare the largest value to thresh
+    vmax.u8     d25, d25, d26              ; m4 = max(m7, m8)
+    vmax.u8     d26, d27, d28              ; m5 = max(m10, m11)
+
+    vshr.u8     d23, d23, #1               ; a = a / 2
+
+    vmax.u8     d25, d25, d26              ; m4 = max(m4, m5)
+
+    vqadd.u8    d24, d24, d23              ; a = b + a
+
+    vmax.u8     d20, d20, d25              ; m2 = max(m2, m4)
+
+    vmov.u8     d23, #1
+    vcge.u8     d24, d0, d24               ; a > blimit
+
+    vcgt.u8     d21, d21, d2               ; (abs(p1 - p0) > thresh)*-1
+
+    vcge.u8     d20, d23, d20              ; flat
+
+    vand        d19, d19, d24              ; mask
+
+    vcgt.u8     d23, d22, d2               ; (abs(q1 - q0) > thresh)*-1
+
+    vand        d20, d20, d19              ; flat & mask
+
+    vmov.u8     d22, #0x80
+
+    vorr        d23, d21, d23              ; hev
+
+    ; This instruction will truncate the "flat & mask" masks down to 4 bits
+    ; each to fit into one 32 bit arm register. The values are stored in
+    ; q10.64[0].
+    vshrn.u16   d30, q10, #4
+    vmov.u32    r4, d30[0]                 ; flat & mask 4bits
+
+    adds        r5, r4, #1                 ; Check for all 1's
+
+    ; If mask and flat are 1's for all vectors, then we only need to execute
+    ; the power branch for all vectors.
+    beq         power_branch_only
+
+    cmp         r4, #0                     ; Check for 0, set flag for later
+
+    ; mbfilter() function
+    ; filter() function
+    ; convert to signed
+    veor        d21, d7, d22               ; qs0
+    veor        d24, d6, d22               ; ps0
+    veor        d25, d5, d22               ; ps1
+    veor        d26, d16, d22              ; qs1
+
+    vmov.u8     d27, #3
+
+    vsub.s8     d28, d21, d24              ; ( qs0 - ps0)
+
+    vqsub.s8    d29, d25, d26              ; filter = clamp(ps1-qs1)
+
+    vmull.s8    q15, d28, d27              ; 3 * ( qs0 - ps0)
+
+    vand        d29, d29, d23              ; filter &= hev
+
+    vaddw.s8    q15, q15, d29              ; filter + 3 * (qs0 - ps0)
+
+    vmov.u8     d29, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d28, q15
+
+    vand        d28, d28, d19              ; filter &= mask
+
+    vqadd.s8    d30, d28, d27              ; filter2 = clamp(filter+3)
+    vqadd.s8    d29, d28, d29              ; filter1 = clamp(filter+4)
+    vshr.s8     d30, d30, #3               ; filter2 >>= 3
+    vshr.s8     d29, d29, #3               ; filter1 >>= 3
+
+    vqadd.s8    d24, d24, d30              ; op0 = clamp(ps0 + filter2)
+    vqsub.s8    d21, d21, d29              ; oq0 = clamp(qs0 - filter1)
+
+    ; outer tap adjustments: ++filter1 >> 1
+    vrshr.s8    d29, d29, #1
+    vbic        d29, d29, d23              ; filter &= ~hev
+
+    vqadd.s8    d25, d25, d29              ; op1 = clamp(ps1 + filter)
+    vqsub.s8    d26, d26, d29              ; oq1 = clamp(qs1 - filter)
+
+    ; If mask and flat are 0's for all vectors, then we only need to execute
+    ; the filter branch for all vectors.
+    beq         filter_branch_only
+
+    ; If mask and flat are mixed then we must perform both branches and
+    ; combine the data.
+    veor        d24, d24, d22              ; *f_op0 = u^0x80
+    veor        d21, d21, d22              ; *f_oq0 = u^0x80
+    veor        d25, d25, d22              ; *f_op1 = u^0x80
+    veor        d26, d26, d22              ; *f_oq1 = u^0x80
+
+    ; At this point we have already executed the filter branch. The filter
+    ; branch does not set op2 or oq2, so use p2 and q2. Execute the power
+    ; branch and combine the data.
+    vmov.u8     d23, #2
+    vaddl.u8    q14, d6, d7                ; r_op2 = p0 + q0
+    vmlal.u8    q14, d3, d27               ; r_op2 += p3 * 3
+    vmlal.u8    q14, d4, d23               ; r_op2 += p2 * 2
+
+    vbif        d0, d4, d20                ; op2 |= p2 & ~(flat & mask)
+
+    vaddw.u8    q14, d5                    ; r_op2 += p1
+
+    vbif        d1, d25, d20               ; op1 |= f_op1 & ~(flat & mask)
+
+    vqrshrn.u16 d30, q14, #3               ; r_op2
+
+    vsubw.u8    q14, d3                    ; r_op1 = r_op2 - p3
+    vsubw.u8    q14, d4                    ; r_op1 -= p2
+    vaddw.u8    q14, d5                    ; r_op1 += p1
+    vaddw.u8    q14, d16                   ; r_op1 += q1
+
+    vbif        d2, d24, d20               ; op0 |= f_op0 & ~(flat & mask)
+
+    vqrshrn.u16 d31, q14, #3               ; r_op1
+
+    vsubw.u8    q14, d3                    ; r_op0 = r_op1 - p3
+    vsubw.u8    q14, d5                    ; r_op0 -= p1
+    vaddw.u8    q14, d6                    ; r_op0 += p0
+    vaddw.u8    q14, d17                   ; r_op0 += q2
+
+    vbit        d0, d30, d20               ; op2 |= r_op2 & (flat & mask)
+
+    vqrshrn.u16 d23, q14, #3               ; r_op0
+
+    vsubw.u8    q14, d3                    ; r_oq0 = r_op0 - p3
+    vsubw.u8    q14, d6                    ; r_oq0 -= p0
+    vaddw.u8    q14, d7                    ; r_oq0 += q0
+
+    vbit        d1, d31, d20               ; op1 |= r_op1 & (flat & mask)
+
+    vaddw.u8    q14, d18                   ; oq0 += q3
+
+    vbit        d2, d23, d20               ; op0 |= r_op0 & (flat & mask)
+
+    vqrshrn.u16 d22, q14, #3               ; r_oq0
+
+    vsubw.u8    q14, d4                    ; r_oq1 = r_oq0 - p2
+    vsubw.u8    q14, d7                    ; r_oq1 -= q0
+    vaddw.u8    q14, d16                   ; r_oq1 += q1
+
+    vbif        d3, d21, d20               ; oq0 |= f_oq0 & ~(flat & mask)
+
+    vaddw.u8    q14, d18                   ; r_oq1 += q3
+
+    vbif        d4, d26, d20               ; oq1 |= f_oq1 & ~(flat & mask)
+
+    vqrshrn.u16 d6, q14, #3                ; r_oq1
+
+    vsubw.u8    q14, d5                    ; r_oq2 = r_oq1 - p1
+    vsubw.u8    q14, d16                   ; r_oq2 -= q1
+    vaddw.u8    q14, d17                   ; r_oq2 += q2
+    vaddw.u8    q14, d18                   ; r_oq2 += q3
+
+    vbif        d5, d17, d20               ; oq2 |= q2 & ~(flat & mask)
+
+    vqrshrn.u16 d7, q14, #3                ; r_oq2
+
+    vbit        d3, d22, d20               ; oq0 |= r_oq0 & (flat & mask)
+    vbit        d4, d6, d20                ; oq1 |= r_oq1 & (flat & mask)
+    vbit        d5, d7, d20                ; oq2 |= r_oq2 & (flat & mask)
+
+    bx          lr
+
+power_branch_only
+    vmov.u8     d27, #3
+    vmov.u8     d21, #2
+    vaddl.u8    q14, d6, d7                ; op2 = p0 + q0
+    vmlal.u8    q14, d3, d27               ; op2 += p3 * 3
+    vmlal.u8    q14, d4, d21               ; op2 += p2 * 2
+    vaddw.u8    q14, d5                    ; op2 += p1
+    vqrshrn.u16 d0, q14, #3                ; op2
+
+    vsubw.u8    q14, d3                    ; op1 = op2 - p3
+    vsubw.u8    q14, d4                    ; op1 -= p2
+    vaddw.u8    q14, d5                    ; op1 += p1
+    vaddw.u8    q14, d16                   ; op1 += q1
+    vqrshrn.u16 d1, q14, #3                ; op1
+
+    vsubw.u8    q14, d3                    ; op0 = op1 - p3
+    vsubw.u8    q14, d5                    ; op0 -= p1
+    vaddw.u8    q14, d6                    ; op0 += p0
+    vaddw.u8    q14, d17                   ; op0 += q2
+    vqrshrn.u16 d2, q14, #3                ; op0
+
+    vsubw.u8    q14, d3                    ; oq0 = op0 - p3
+    vsubw.u8    q14, d6                    ; oq0 -= p0
+    vaddw.u8    q14, d7                    ; oq0 += q0
+    vaddw.u8    q14, d18                   ; oq0 += q3
+    vqrshrn.u16 d3, q14, #3                ; oq0
+
+    vsubw.u8    q14, d4                    ; oq1 = oq0 - p2
+    vsubw.u8    q14, d7                    ; oq1 -= q0
+    vaddw.u8    q14, d16                   ; oq1 += q1
+    vaddw.u8    q14, d18                   ; oq1 += q3
+    vqrshrn.u16 d4, q14, #3                ; oq1
+
+    vsubw.u8    q14, d5                    ; oq2 = oq1 - p1
+    vsubw.u8    q14, d16                   ; oq2 -= q1
+    vaddw.u8    q14, d17                   ; oq2 += q2
+    vaddw.u8    q14, d18                   ; oq2 += q3
+    vqrshrn.u16 d5, q14, #3                ; oq2
+
+    bx          lr
+
+filter_branch_only
+    ; TODO(fgalligan): See if we can rearange registers so we do not need to
+    ; do the 2 vswp.
+    vswp        d0, d4                      ; op2
+    vswp        d5, d17                     ; oq2
+    veor        d2, d24, d22                ; *op0 = u^0x80
+    veor        d3, d21, d22                ; *oq0 = u^0x80
+    veor        d1, d25, d22                ; *op1 = u^0x80
+    veor        d4, d26, d22                ; *oq1 = u^0x80
+
+    bx          lr
+
+    ENDP        ; |vpx_mbloop_filter_neon|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/loopfilter_8_neon.c b/libvpx/libvpx/vpx_dsp/arm/loopfilter_8_neon.c
new file mode 100644
index 0000000..ec37573
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/loopfilter_8_neon.c
@@ -0,0 +1,445 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+
+static INLINE void mbloop_filter_neon(
+        uint8x8_t dblimit,   // mblimit
+        uint8x8_t dlimit,    // limit
+        uint8x8_t dthresh,   // thresh
+        uint8x8_t d3u8,      // p2
+        uint8x8_t d4u8,      // p2
+        uint8x8_t d5u8,      // p1
+        uint8x8_t d6u8,      // p0
+        uint8x8_t d7u8,      // q0
+        uint8x8_t d16u8,     // q1
+        uint8x8_t d17u8,     // q2
+        uint8x8_t d18u8,     // q3
+        uint8x8_t *d0ru8,    // p1
+        uint8x8_t *d1ru8,    // p1
+        uint8x8_t *d2ru8,    // p0
+        uint8x8_t *d3ru8,    // q0
+        uint8x8_t *d4ru8,    // q1
+        uint8x8_t *d5ru8) {  // q1
+    uint32_t flat;
+    uint8x8_t d0u8, d1u8, d2u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8;
+    uint8x8_t d25u8, d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
+    int16x8_t q15s16;
+    uint16x8_t q10u16, q14u16;
+    int8x8_t d21s8, d24s8, d25s8, d26s8, d28s8, d29s8, d30s8;
+
+    d19u8 = vabd_u8(d3u8, d4u8);
+    d20u8 = vabd_u8(d4u8, d5u8);
+    d21u8 = vabd_u8(d5u8, d6u8);
+    d22u8 = vabd_u8(d16u8, d7u8);
+    d23u8 = vabd_u8(d17u8, d16u8);
+    d24u8 = vabd_u8(d18u8, d17u8);
+
+    d19u8 = vmax_u8(d19u8, d20u8);
+    d20u8 = vmax_u8(d21u8, d22u8);
+
+    d25u8 = vabd_u8(d6u8, d4u8);
+
+    d23u8 = vmax_u8(d23u8, d24u8);
+
+    d26u8 = vabd_u8(d7u8, d17u8);
+
+    d19u8 = vmax_u8(d19u8, d20u8);
+
+    d24u8 = vabd_u8(d6u8, d7u8);
+    d27u8 = vabd_u8(d3u8, d6u8);
+    d28u8 = vabd_u8(d18u8, d7u8);
+
+    d19u8 = vmax_u8(d19u8, d23u8);
+
+    d23u8 = vabd_u8(d5u8, d16u8);
+    d24u8 = vqadd_u8(d24u8, d24u8);
+
+
+    d19u8 = vcge_u8(dlimit, d19u8);
+
+
+    d25u8 = vmax_u8(d25u8, d26u8);
+    d26u8 = vmax_u8(d27u8, d28u8);
+
+    d23u8 = vshr_n_u8(d23u8, 1);
+
+    d25u8 = vmax_u8(d25u8, d26u8);
+
+    d24u8 = vqadd_u8(d24u8, d23u8);
+
+    d20u8 = vmax_u8(d20u8, d25u8);
+
+    d23u8 = vdup_n_u8(1);
+    d24u8 = vcge_u8(dblimit, d24u8);
+
+    d21u8 = vcgt_u8(d21u8, dthresh);
+
+    d20u8 = vcge_u8(d23u8, d20u8);
+
+    d19u8 = vand_u8(d19u8, d24u8);
+
+    d23u8 = vcgt_u8(d22u8, dthresh);
+
+    d20u8 = vand_u8(d20u8, d19u8);
+
+    d22u8 = vdup_n_u8(0x80);
+
+    d23u8 = vorr_u8(d21u8, d23u8);
+
+    q10u16 = vcombine_u16(vreinterpret_u16_u8(d20u8),
+                          vreinterpret_u16_u8(d21u8));
+
+    d30u8 = vshrn_n_u16(q10u16, 4);
+    flat = vget_lane_u32(vreinterpret_u32_u8(d30u8), 0);
+
+    if (flat == 0xffffffff) {  // Check for all 1's, power_branch_only
+        d27u8 = vdup_n_u8(3);
+        d21u8 = vdup_n_u8(2);
+        q14u16 = vaddl_u8(d6u8, d7u8);
+        q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
+        q14u16 = vmlal_u8(q14u16, d4u8, d21u8);
+        q14u16 = vaddw_u8(q14u16, d5u8);
+        *d0ru8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d4u8);
+        q14u16 = vaddw_u8(q14u16, d5u8);
+        q14u16 = vaddw_u8(q14u16, d16u8);
+        *d1ru8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d5u8);
+        q14u16 = vaddw_u8(q14u16, d6u8);
+        q14u16 = vaddw_u8(q14u16, d17u8);
+        *d2ru8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d6u8);
+        q14u16 = vaddw_u8(q14u16, d7u8);
+        q14u16 = vaddw_u8(q14u16, d18u8);
+        *d3ru8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d4u8);
+        q14u16 = vsubw_u8(q14u16, d7u8);
+        q14u16 = vaddw_u8(q14u16, d16u8);
+        q14u16 = vaddw_u8(q14u16, d18u8);
+        *d4ru8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d5u8);
+        q14u16 = vsubw_u8(q14u16, d16u8);
+        q14u16 = vaddw_u8(q14u16, d17u8);
+        q14u16 = vaddw_u8(q14u16, d18u8);
+        *d5ru8 = vqrshrn_n_u16(q14u16, 3);
+    } else {
+        d21u8 = veor_u8(d7u8,  d22u8);
+        d24u8 = veor_u8(d6u8,  d22u8);
+        d25u8 = veor_u8(d5u8,  d22u8);
+        d26u8 = veor_u8(d16u8, d22u8);
+
+        d27u8 = vdup_n_u8(3);
+
+        d28s8 = vsub_s8(vreinterpret_s8_u8(d21u8), vreinterpret_s8_u8(d24u8));
+        d29s8 = vqsub_s8(vreinterpret_s8_u8(d25u8), vreinterpret_s8_u8(d26u8));
+
+        q15s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d27u8));
+
+        d29s8 = vand_s8(d29s8, vreinterpret_s8_u8(d23u8));
+
+        q15s16 = vaddw_s8(q15s16, d29s8);
+
+        d29u8 = vdup_n_u8(4);
+
+        d28s8 = vqmovn_s16(q15s16);
+
+        d28s8 = vand_s8(d28s8, vreinterpret_s8_u8(d19u8));
+
+        d30s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d27u8));
+        d29s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d29u8));
+        d30s8 = vshr_n_s8(d30s8, 3);
+        d29s8 = vshr_n_s8(d29s8, 3);
+
+        d24s8 = vqadd_s8(vreinterpret_s8_u8(d24u8), d30s8);
+        d21s8 = vqsub_s8(vreinterpret_s8_u8(d21u8), d29s8);
+
+        d29s8 = vrshr_n_s8(d29s8, 1);
+        d29s8 = vbic_s8(d29s8, vreinterpret_s8_u8(d23u8));
+
+        d25s8 = vqadd_s8(vreinterpret_s8_u8(d25u8), d29s8);
+        d26s8 = vqsub_s8(vreinterpret_s8_u8(d26u8), d29s8);
+
+        if (flat == 0) {  // filter_branch_only
+            *d0ru8 = d4u8;
+            *d1ru8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
+            *d2ru8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
+            *d3ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
+            *d4ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
+            *d5ru8 = d17u8;
+            return;
+        }
+
+        d21u8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
+        d24u8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
+        d25u8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
+        d26u8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
+
+        d23u8 = vdup_n_u8(2);
+        q14u16 = vaddl_u8(d6u8, d7u8);
+        q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
+        q14u16 = vmlal_u8(q14u16, d4u8, d23u8);
+
+        d0u8 = vbsl_u8(d20u8, dblimit, d4u8);
+
+        q14u16 = vaddw_u8(q14u16, d5u8);
+
+        d1u8 = vbsl_u8(d20u8, dlimit, d25u8);
+
+        d30u8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d4u8);
+        q14u16 = vaddw_u8(q14u16, d5u8);
+        q14u16 = vaddw_u8(q14u16, d16u8);
+
+        d2u8 = vbsl_u8(d20u8, dthresh, d24u8);
+
+        d31u8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d5u8);
+        q14u16 = vaddw_u8(q14u16, d6u8);
+        q14u16 = vaddw_u8(q14u16, d17u8);
+
+        *d0ru8 = vbsl_u8(d20u8, d30u8, d0u8);
+
+        d23u8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d6u8);
+        q14u16 = vaddw_u8(q14u16, d7u8);
+
+        *d1ru8 = vbsl_u8(d20u8, d31u8, d1u8);
+
+        q14u16 = vaddw_u8(q14u16, d18u8);
+
+        *d2ru8 = vbsl_u8(d20u8, d23u8, d2u8);
+
+        d22u8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d4u8);
+        q14u16 = vsubw_u8(q14u16, d7u8);
+        q14u16 = vaddw_u8(q14u16, d16u8);
+
+        d3u8 = vbsl_u8(d20u8, d3u8, d21u8);
+
+        q14u16 = vaddw_u8(q14u16, d18u8);
+
+        d4u8 = vbsl_u8(d20u8, d4u8, d26u8);
+
+        d6u8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d5u8);
+        q14u16 = vsubw_u8(q14u16, d16u8);
+        q14u16 = vaddw_u8(q14u16, d17u8);
+        q14u16 = vaddw_u8(q14u16, d18u8);
+
+        d5u8 = vbsl_u8(d20u8, d5u8, d17u8);
+
+        d7u8 = vqrshrn_n_u16(q14u16, 3);
+
+        *d3ru8 = vbsl_u8(d20u8, d22u8, d3u8);
+        *d4ru8 = vbsl_u8(d20u8, d6u8, d4u8);
+        *d5ru8 = vbsl_u8(d20u8, d7u8, d5u8);
+    }
+    return;
+}
+
+void vpx_lpf_horizontal_8_neon(
+        uint8_t *src,
+        int pitch,
+        const uint8_t *blimit,
+        const uint8_t *limit,
+        const uint8_t *thresh) {
+    int i;
+    uint8_t *s, *psrc;
+    uint8x8_t dblimit, dlimit, dthresh;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+    uint8x8_t d16u8, d17u8, d18u8;
+
+    dblimit = vld1_u8(blimit);
+    dlimit = vld1_u8(limit);
+    dthresh = vld1_u8(thresh);
+
+    psrc = src - (pitch << 2);
+    for (i = 0; i < 1; i++) {
+        s = psrc + i * 8;
+
+        d3u8  = vld1_u8(s);
+        s += pitch;
+        d4u8  = vld1_u8(s);
+        s += pitch;
+        d5u8  = vld1_u8(s);
+        s += pitch;
+        d6u8  = vld1_u8(s);
+        s += pitch;
+        d7u8  = vld1_u8(s);
+        s += pitch;
+        d16u8 = vld1_u8(s);
+        s += pitch;
+        d17u8 = vld1_u8(s);
+        s += pitch;
+        d18u8 = vld1_u8(s);
+
+        mbloop_filter_neon(dblimit, dlimit, dthresh,
+                           d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+                           &d0u8, &d1u8, &d2u8, &d3u8, &d4u8, &d5u8);
+
+        s -= (pitch * 6);
+        vst1_u8(s, d0u8);
+        s += pitch;
+        vst1_u8(s, d1u8);
+        s += pitch;
+        vst1_u8(s, d2u8);
+        s += pitch;
+        vst1_u8(s, d3u8);
+        s += pitch;
+        vst1_u8(s, d4u8);
+        s += pitch;
+        vst1_u8(s, d5u8);
+    }
+    return;
+}
+
+void vpx_lpf_vertical_8_neon(
+        uint8_t *src,
+        int pitch,
+        const uint8_t *blimit,
+        const uint8_t *limit,
+        const uint8_t *thresh) {
+    int i;
+    uint8_t *s;
+    uint8x8_t dblimit, dlimit, dthresh;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+    uint8x8_t d16u8, d17u8, d18u8;
+    uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
+    uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
+    uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
+    uint8x8x4_t d4Result;
+    uint8x8x2_t d2Result;
+
+    dblimit = vld1_u8(blimit);
+    dlimit = vld1_u8(limit);
+    dthresh = vld1_u8(thresh);
+
+    for (i = 0; i < 1; i++) {
+        s = src + (i * (pitch << 3)) - 4;
+
+        d3u8 = vld1_u8(s);
+        s += pitch;
+        d4u8 = vld1_u8(s);
+        s += pitch;
+        d5u8 = vld1_u8(s);
+        s += pitch;
+        d6u8 = vld1_u8(s);
+        s += pitch;
+        d7u8 = vld1_u8(s);
+        s += pitch;
+        d16u8 = vld1_u8(s);
+        s += pitch;
+        d17u8 = vld1_u8(s);
+        s += pitch;
+        d18u8 = vld1_u8(s);
+
+        d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8),
+                          vreinterpret_u32_u8(d7u8));
+        d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8),
+                          vreinterpret_u32_u8(d16u8));
+        d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8),
+                          vreinterpret_u32_u8(d17u8));
+        d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8),
+                          vreinterpret_u32_u8(d18u8));
+
+        d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
+                          vreinterpret_u16_u32(d2tmp2.val[0]));
+        d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
+                          vreinterpret_u16_u32(d2tmp3.val[0]));
+        d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
+                          vreinterpret_u16_u32(d2tmp2.val[1]));
+        d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
+                          vreinterpret_u16_u32(d2tmp3.val[1]));
+
+        d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
+                         vreinterpret_u8_u16(d2tmp5.val[0]));
+        d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
+                         vreinterpret_u8_u16(d2tmp5.val[1]));
+        d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
+                          vreinterpret_u8_u16(d2tmp7.val[0]));
+        d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
+                          vreinterpret_u8_u16(d2tmp7.val[1]));
+
+        d3u8 = d2tmp8.val[0];
+        d4u8 = d2tmp8.val[1];
+        d5u8 = d2tmp9.val[0];
+        d6u8 = d2tmp9.val[1];
+        d7u8 = d2tmp10.val[0];
+        d16u8 = d2tmp10.val[1];
+        d17u8 = d2tmp11.val[0];
+        d18u8 = d2tmp11.val[1];
+
+        mbloop_filter_neon(dblimit, dlimit, dthresh,
+                           d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+                           &d0u8, &d1u8, &d2u8, &d3u8, &d4u8, &d5u8);
+
+        d4Result.val[0] = d0u8;
+        d4Result.val[1] = d1u8;
+        d4Result.val[2] = d2u8;
+        d4Result.val[3] = d3u8;
+
+        d2Result.val[0] = d4u8;
+        d2Result.val[1] = d5u8;
+
+        s = src - 3;
+        vst4_lane_u8(s, d4Result, 0);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 1);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 2);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 3);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 4);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 5);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 6);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 7);
+
+        s = src + 1;
+        vst2_lane_u8(s, d2Result, 0);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 1);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 2);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 3);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 4);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 5);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 6);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 7);
+    }
+    return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/loopfilter_mb_neon.asm b/libvpx/libvpx/vpx_dsp/arm/loopfilter_mb_neon.asm
new file mode 100644
index 0000000..d5da7a8
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/loopfilter_mb_neon.asm
@@ -0,0 +1,635 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vpx_lpf_horizontal_edge_8_neon|
+    EXPORT  |vpx_lpf_horizontal_edge_16_neon|
+    EXPORT  |vpx_lpf_vertical_16_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; void mb_lpf_horizontal_edge(uint8_t *s, int p,
+;                             const uint8_t *blimit,
+;                             const uint8_t *limit,
+;                             const uint8_t *thresh,
+;                             int count)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+; r12   int count
+|mb_lpf_horizontal_edge| PROC
+    push        {r4-r8, lr}
+    vpush       {d8-d15}
+    ldr         r4, [sp, #88]              ; load thresh
+
+h_count
+    vld1.8      {d16[]}, [r2]              ; load *blimit
+    vld1.8      {d17[]}, [r3]              ; load *limit
+    vld1.8      {d18[]}, [r4]              ; load *thresh
+
+    sub         r8, r0, r1, lsl #3         ; move src pointer down by 8 lines
+
+    vld1.u8     {d0}, [r8@64], r1          ; p7
+    vld1.u8     {d1}, [r8@64], r1          ; p6
+    vld1.u8     {d2}, [r8@64], r1          ; p5
+    vld1.u8     {d3}, [r8@64], r1          ; p4
+    vld1.u8     {d4}, [r8@64], r1          ; p3
+    vld1.u8     {d5}, [r8@64], r1          ; p2
+    vld1.u8     {d6}, [r8@64], r1          ; p1
+    vld1.u8     {d7}, [r8@64], r1          ; p0
+    vld1.u8     {d8}, [r8@64], r1          ; q0
+    vld1.u8     {d9}, [r8@64], r1          ; q1
+    vld1.u8     {d10}, [r8@64], r1         ; q2
+    vld1.u8     {d11}, [r8@64], r1         ; q3
+    vld1.u8     {d12}, [r8@64], r1         ; q4
+    vld1.u8     {d13}, [r8@64], r1         ; q5
+    vld1.u8     {d14}, [r8@64], r1         ; q6
+    vld1.u8     {d15}, [r8@64], r1         ; q7
+
+    bl          vpx_wide_mbfilter_neon
+
+    tst         r7, #1
+    beq         h_mbfilter
+
+    ; flat && mask were not set for any of the channels. Just store the values
+    ; from filter.
+    sub         r8, r0, r1, lsl #1
+
+    vst1.u8     {d25}, [r8@64], r1         ; store op1
+    vst1.u8     {d24}, [r8@64], r1         ; store op0
+    vst1.u8     {d23}, [r8@64], r1         ; store oq0
+    vst1.u8     {d26}, [r8@64], r1         ; store oq1
+
+    b           h_next
+
+h_mbfilter
+    tst         r7, #2
+    beq         h_wide_mbfilter
+
+    ; flat2 was not set for any of the channels. Just store the values from
+    ; mbfilter.
+    sub         r8, r0, r1, lsl #1
+    sub         r8, r8, r1
+
+    vst1.u8     {d18}, [r8@64], r1         ; store op2
+    vst1.u8     {d19}, [r8@64], r1         ; store op1
+    vst1.u8     {d20}, [r8@64], r1         ; store op0
+    vst1.u8     {d21}, [r8@64], r1         ; store oq0
+    vst1.u8     {d22}, [r8@64], r1         ; store oq1
+    vst1.u8     {d23}, [r8@64], r1         ; store oq2
+
+    b           h_next
+
+h_wide_mbfilter
+    sub         r8, r0, r1, lsl #3
+    add         r8, r8, r1
+
+    vst1.u8     {d16}, [r8@64], r1         ; store op6
+    vst1.u8     {d24}, [r8@64], r1         ; store op5
+    vst1.u8     {d25}, [r8@64], r1         ; store op4
+    vst1.u8     {d26}, [r8@64], r1         ; store op3
+    vst1.u8     {d27}, [r8@64], r1         ; store op2
+    vst1.u8     {d18}, [r8@64], r1         ; store op1
+    vst1.u8     {d19}, [r8@64], r1         ; store op0
+    vst1.u8     {d20}, [r8@64], r1         ; store oq0
+    vst1.u8     {d21}, [r8@64], r1         ; store oq1
+    vst1.u8     {d22}, [r8@64], r1         ; store oq2
+    vst1.u8     {d23}, [r8@64], r1         ; store oq3
+    vst1.u8     {d1}, [r8@64], r1          ; store oq4
+    vst1.u8     {d2}, [r8@64], r1          ; store oq5
+    vst1.u8     {d3}, [r8@64], r1          ; store oq6
+
+h_next
+    add         r0, r0, #8
+    subs        r12, r12, #1
+    bne         h_count
+
+    vpop        {d8-d15}
+    pop         {r4-r8, pc}
+
+    ENDP        ; |mb_lpf_horizontal_edge|
+
+; void vpx_lpf_horizontal_edge_8_neon(uint8_t *s, int pitch,
+;                                     const uint8_t *blimit,
+;                                     const uint8_t *limit,
+;                                     const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int pitch,
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh
+|vpx_lpf_horizontal_edge_8_neon| PROC
+    mov r12, #1
+    b mb_lpf_horizontal_edge
+    ENDP        ; |vpx_lpf_horizontal_edge_8_neon|
+
+; void vpx_lpf_horizontal_edge_16_neon(uint8_t *s, int pitch,
+;                                      const uint8_t *blimit,
+;                                      const uint8_t *limit,
+;                                      const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int pitch,
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh
+|vpx_lpf_horizontal_edge_16_neon| PROC
+    mov r12, #2
+    b mb_lpf_horizontal_edge
+    ENDP        ; |vpx_lpf_horizontal_edge_16_neon|
+
+; void vpx_lpf_vertical_16_neon(uint8_t *s, int p,
+;                               const uint8_t *blimit,
+;                               const uint8_t *limit,
+;                               const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|vpx_lpf_vertical_16_neon| PROC
+    push        {r4-r8, lr}
+    vpush       {d8-d15}
+    ldr         r4, [sp, #88]              ; load thresh
+
+    vld1.8      {d16[]}, [r2]              ; load *blimit
+    vld1.8      {d17[]}, [r3]              ; load *limit
+    vld1.8      {d18[]}, [r4]              ; load *thresh
+
+    sub         r8, r0, #8
+
+    vld1.8      {d0}, [r8@64], r1
+    vld1.8      {d8}, [r0@64], r1
+    vld1.8      {d1}, [r8@64], r1
+    vld1.8      {d9}, [r0@64], r1
+    vld1.8      {d2}, [r8@64], r1
+    vld1.8      {d10}, [r0@64], r1
+    vld1.8      {d3}, [r8@64], r1
+    vld1.8      {d11}, [r0@64], r1
+    vld1.8      {d4}, [r8@64], r1
+    vld1.8      {d12}, [r0@64], r1
+    vld1.8      {d5}, [r8@64], r1
+    vld1.8      {d13}, [r0@64], r1
+    vld1.8      {d6}, [r8@64], r1
+    vld1.8      {d14}, [r0@64], r1
+    vld1.8      {d7}, [r8@64], r1
+    vld1.8      {d15}, [r0@64], r1
+
+    sub         r0, r0, r1, lsl #3
+
+    vtrn.32     q0, q2
+    vtrn.32     q1, q3
+    vtrn.32     q4, q6
+    vtrn.32     q5, q7
+
+    vtrn.16     q0, q1
+    vtrn.16     q2, q3
+    vtrn.16     q4, q5
+    vtrn.16     q6, q7
+
+    vtrn.8      d0, d1
+    vtrn.8      d2, d3
+    vtrn.8      d4, d5
+    vtrn.8      d6, d7
+
+    vtrn.8      d8, d9
+    vtrn.8      d10, d11
+    vtrn.8      d12, d13
+    vtrn.8      d14, d15
+
+    bl          vpx_wide_mbfilter_neon
+
+    tst         r7, #1
+    beq         v_mbfilter
+
+    ; flat && mask were not set for any of the channels. Just store the values
+    ; from filter.
+    sub         r8, r0, #2
+
+    vswp        d23, d25
+
+    vst4.8      {d23[0], d24[0], d25[0], d26[0]}, [r8], r1
+    vst4.8      {d23[1], d24[1], d25[1], d26[1]}, [r8], r1
+    vst4.8      {d23[2], d24[2], d25[2], d26[2]}, [r8], r1
+    vst4.8      {d23[3], d24[3], d25[3], d26[3]}, [r8], r1
+    vst4.8      {d23[4], d24[4], d25[4], d26[4]}, [r8], r1
+    vst4.8      {d23[5], d24[5], d25[5], d26[5]}, [r8], r1
+    vst4.8      {d23[6], d24[6], d25[6], d26[6]}, [r8], r1
+    vst4.8      {d23[7], d24[7], d25[7], d26[7]}, [r8], r1
+
+    b           v_end
+
+v_mbfilter
+    tst         r7, #2
+    beq         v_wide_mbfilter
+
+    ; flat2 was not set for any of the channels. Just store the values from
+    ; mbfilter.
+    sub         r8, r0, #3
+
+    vst3.8      {d18[0], d19[0], d20[0]}, [r8], r1
+    vst3.8      {d21[0], d22[0], d23[0]}, [r0], r1
+    vst3.8      {d18[1], d19[1], d20[1]}, [r8], r1
+    vst3.8      {d21[1], d22[1], d23[1]}, [r0], r1
+    vst3.8      {d18[2], d19[2], d20[2]}, [r8], r1
+    vst3.8      {d21[2], d22[2], d23[2]}, [r0], r1
+    vst3.8      {d18[3], d19[3], d20[3]}, [r8], r1
+    vst3.8      {d21[3], d22[3], d23[3]}, [r0], r1
+    vst3.8      {d18[4], d19[4], d20[4]}, [r8], r1
+    vst3.8      {d21[4], d22[4], d23[4]}, [r0], r1
+    vst3.8      {d18[5], d19[5], d20[5]}, [r8], r1
+    vst3.8      {d21[5], d22[5], d23[5]}, [r0], r1
+    vst3.8      {d18[6], d19[6], d20[6]}, [r8], r1
+    vst3.8      {d21[6], d22[6], d23[6]}, [r0], r1
+    vst3.8      {d18[7], d19[7], d20[7]}, [r8], r1
+    vst3.8      {d21[7], d22[7], d23[7]}, [r0], r1
+
+    b           v_end
+
+v_wide_mbfilter
+    sub         r8, r0, #8
+
+    vtrn.32     d0,  d26
+    vtrn.32     d16, d27
+    vtrn.32     d24, d18
+    vtrn.32     d25, d19
+
+    vtrn.16     d0,  d24
+    vtrn.16     d16, d25
+    vtrn.16     d26, d18
+    vtrn.16     d27, d19
+
+    vtrn.8      d0,  d16
+    vtrn.8      d24, d25
+    vtrn.8      d26, d27
+    vtrn.8      d18, d19
+
+    vtrn.32     d20, d1
+    vtrn.32     d21, d2
+    vtrn.32     d22, d3
+    vtrn.32     d23, d15
+
+    vtrn.16     d20, d22
+    vtrn.16     d21, d23
+    vtrn.16     d1,  d3
+    vtrn.16     d2,  d15
+
+    vtrn.8      d20, d21
+    vtrn.8      d22, d23
+    vtrn.8      d1,  d2
+    vtrn.8      d3,  d15
+
+    vst1.8      {d0}, [r8@64], r1
+    vst1.8      {d20}, [r0@64], r1
+    vst1.8      {d16}, [r8@64], r1
+    vst1.8      {d21}, [r0@64], r1
+    vst1.8      {d24}, [r8@64], r1
+    vst1.8      {d22}, [r0@64], r1
+    vst1.8      {d25}, [r8@64], r1
+    vst1.8      {d23}, [r0@64], r1
+    vst1.8      {d26}, [r8@64], r1
+    vst1.8      {d1}, [r0@64], r1
+    vst1.8      {d27}, [r8@64], r1
+    vst1.8      {d2}, [r0@64], r1
+    vst1.8      {d18}, [r8@64], r1
+    vst1.8      {d3}, [r0@64], r1
+    vst1.8      {d19}, [r8@64], r1
+    vst1.8      {d15}, [r0@64], r1
+
+v_end
+    vpop        {d8-d15}
+    pop         {r4-r8, pc}
+
+    ENDP        ; |vpx_lpf_vertical_16_neon|
+
+; void vpx_wide_mbfilter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store.
+;
+; r0-r3 PRESERVE
+; d16    blimit
+; d17    limit
+; d18    thresh
+; d0    p7
+; d1    p6
+; d2    p5
+; d3    p4
+; d4    p3
+; d5    p2
+; d6    p1
+; d7    p0
+; d8    q0
+; d9    q1
+; d10   q2
+; d11   q3
+; d12   q4
+; d13   q5
+; d14   q6
+; d15   q7
+|vpx_wide_mbfilter_neon| PROC
+    mov         r7, #0
+
+    ; filter_mask
+    vabd.u8     d19, d4, d5                ; abs(p3 - p2)
+    vabd.u8     d20, d5, d6                ; abs(p2 - p1)
+    vabd.u8     d21, d6, d7                ; abs(p1 - p0)
+    vabd.u8     d22, d9, d8                ; abs(q1 - q0)
+    vabd.u8     d23, d10, d9               ; abs(q2 - q1)
+    vabd.u8     d24, d11, d10              ; abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20              ; max(abs(p3 - p2), abs(p2 - p1))
+    vmax.u8     d20, d21, d22              ; max(abs(p1 - p0), abs(q1 - q0))
+    vmax.u8     d23, d23, d24              ; max(abs(q2 - q1), abs(q3 - q2))
+    vmax.u8     d19, d19, d20
+
+    vabd.u8     d24, d7, d8                ; abs(p0 - q0)
+
+    vmax.u8     d19, d19, d23
+
+    vabd.u8     d23, d6, d9                ; a = abs(p1 - q1)
+    vqadd.u8    d24, d24, d24              ; b = abs(p0 - q0) * 2
+
+    ; abs () > limit
+    vcge.u8     d19, d17, d19
+
+    ; flatmask4
+    vabd.u8     d25, d7, d5                ; abs(p0 - p2)
+    vabd.u8     d26, d8, d10               ; abs(q0 - q2)
+    vabd.u8     d27, d4, d7                ; abs(p3 - p0)
+    vabd.u8     d28, d11, d8               ; abs(q3 - q0)
+
+    ; only compare the largest value to thresh
+    vmax.u8     d25, d25, d26              ; max(abs(p0 - p2), abs(q0 - q2))
+    vmax.u8     d26, d27, d28              ; max(abs(p3 - p0), abs(q3 - q0))
+    vmax.u8     d25, d25, d26
+    vmax.u8     d20, d20, d25
+
+    vshr.u8     d23, d23, #1               ; a = a / 2
+    vqadd.u8    d24, d24, d23              ; a = b + a
+
+    vmov.u8     d30, #1
+    vcge.u8     d24, d16, d24              ; (a > blimit * 2 + limit) * -1
+
+    vcge.u8     d20, d30, d20              ; flat
+
+    vand        d19, d19, d24              ; mask
+
+    ; hevmask
+    vcgt.u8     d21, d21, d18              ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     d22, d22, d18              ; (abs(q1 - q0) > thresh)*-1
+    vorr        d21, d21, d22              ; hev
+
+    vand        d16, d20, d19              ; flat && mask
+    vmov        r5, r6, d16
+
+    ; flatmask5(1, p7, p6, p5, p4, p0, q0, q4, q5, q6, q7)
+    vabd.u8     d22, d3, d7                ; abs(p4 - p0)
+    vabd.u8     d23, d12, d8               ; abs(q4 - q0)
+    vabd.u8     d24, d7, d2                ; abs(p0 - p5)
+    vabd.u8     d25, d8, d13               ; abs(q0 - q5)
+    vabd.u8     d26, d1, d7                ; abs(p6 - p0)
+    vabd.u8     d27, d14, d8               ; abs(q6 - q0)
+    vabd.u8     d28, d0, d7                ; abs(p7 - p0)
+    vabd.u8     d29, d15, d8               ; abs(q7 - q0)
+
+    ; only compare the largest value to thresh
+    vmax.u8     d22, d22, d23              ; max(abs(p4 - p0), abs(q4 - q0))
+    vmax.u8     d23, d24, d25              ; max(abs(p0 - p5), abs(q0 - q5))
+    vmax.u8     d24, d26, d27              ; max(abs(p6 - p0), abs(q6 - q0))
+    vmax.u8     d25, d28, d29              ; max(abs(p7 - p0), abs(q7 - q0))
+
+    vmax.u8     d26, d22, d23
+    vmax.u8     d27, d24, d25
+    vmax.u8     d23, d26, d27
+
+    vcge.u8     d18, d30, d23              ; flat2
+
+    vmov.u8     d22, #0x80
+
+    orrs        r5, r5, r6                 ; Check for 0
+    orreq       r7, r7, #1                 ; Only do filter branch
+
+    vand        d17, d18, d16              ; flat2 && flat && mask
+    vmov        r5, r6, d17
+
+    ; mbfilter() function
+
+    ; filter() function
+    ; convert to signed
+    veor        d23, d8, d22               ; qs0
+    veor        d24, d7, d22               ; ps0
+    veor        d25, d6, d22               ; ps1
+    veor        d26, d9, d22               ; qs1
+
+    vmov.u8     d27, #3
+
+    vsub.s8     d28, d23, d24              ; ( qs0 - ps0)
+    vqsub.s8    d29, d25, d26              ; filter = clamp(ps1-qs1)
+    vmull.s8    q15, d28, d27              ; 3 * ( qs0 - ps0)
+    vand        d29, d29, d21              ; filter &= hev
+    vaddw.s8    q15, q15, d29              ; filter + 3 * (qs0 - ps0)
+    vmov.u8     d29, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d28, q15
+
+    vand        d28, d28, d19              ; filter &= mask
+
+    vqadd.s8    d30, d28, d27              ; filter2 = clamp(filter+3)
+    vqadd.s8    d29, d28, d29              ; filter1 = clamp(filter+4)
+    vshr.s8     d30, d30, #3               ; filter2 >>= 3
+    vshr.s8     d29, d29, #3               ; filter1 >>= 3
+
+
+    vqadd.s8    d24, d24, d30              ; op0 = clamp(ps0 + filter2)
+    vqsub.s8    d23, d23, d29              ; oq0 = clamp(qs0 - filter1)
+
+    ; outer tap adjustments: ++filter1 >> 1
+    vrshr.s8    d29, d29, #1
+    vbic        d29, d29, d21              ; filter &= ~hev
+
+    vqadd.s8    d25, d25, d29              ; op1 = clamp(ps1 + filter)
+    vqsub.s8    d26, d26, d29              ; oq1 = clamp(qs1 - filter)
+
+    veor        d24, d24, d22              ; *f_op0 = u^0x80
+    veor        d23, d23, d22              ; *f_oq0 = u^0x80
+    veor        d25, d25, d22              ; *f_op1 = u^0x80
+    veor        d26, d26, d22              ; *f_oq1 = u^0x80
+
+    tst         r7, #1
+    bxne        lr
+
+    orrs        r5, r5, r6                 ; Check for 0
+    orreq       r7, r7, #2                 ; Only do mbfilter branch
+
+    ; mbfilter flat && mask branch
+    ; TODO(fgalligan): Can I decrease the cycles shifting to consective d's
+    ; and using vibt on the q's?
+    vmov.u8     d29, #2
+    vaddl.u8    q15, d7, d8                ; op2 = p0 + q0
+    vmlal.u8    q15, d4, d27               ; op2 = p0 + q0 + p3 * 3
+    vmlal.u8    q15, d5, d29               ; op2 = p0 + q0 + p3 * 3 + p2 * 2
+    vaddl.u8    q10, d4, d5
+    vaddw.u8    q15, d6                    ; op2=p1 + p0 + q0 + p3 * 3 + p2 *2
+    vaddl.u8    q14, d6, d9
+    vqrshrn.u16 d18, q15, #3               ; r_op2
+
+    vsub.i16    q15, q10
+    vaddl.u8    q10, d4, d6
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d7, d10
+    vqrshrn.u16 d19, q15, #3               ; r_op1
+
+    vsub.i16    q15, q10
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d8, d11
+    vqrshrn.u16 d20, q15, #3               ; r_op0
+
+    vsubw.u8    q15, d4                    ; oq0 = op0 - p3
+    vsubw.u8    q15, d7                    ; oq0 -= p0
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d9, d11
+    vqrshrn.u16 d21, q15, #3               ; r_oq0
+
+    vsubw.u8    q15, d5                    ; oq1 = oq0 - p2
+    vsubw.u8    q15, d8                    ; oq1 -= q0
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d10, d11
+    vqrshrn.u16 d22, q15, #3               ; r_oq1
+
+    vsubw.u8    q15, d6                    ; oq2 = oq0 - p1
+    vsubw.u8    q15, d9                    ; oq2 -= q1
+    vadd.i16    q15, q14
+    vqrshrn.u16 d27, q15, #3               ; r_oq2
+
+    ; Filter does not set op2 or oq2, so use p2 and q2.
+    vbif        d18, d5, d16               ; t_op2 |= p2 & ~(flat & mask)
+    vbif        d19, d25, d16              ; t_op1 |= f_op1 & ~(flat & mask)
+    vbif        d20, d24, d16              ; t_op0 |= f_op0 & ~(flat & mask)
+    vbif        d21, d23, d16              ; t_oq0 |= f_oq0 & ~(flat & mask)
+    vbif        d22, d26, d16              ; t_oq1 |= f_oq1 & ~(flat & mask)
+
+    vbit        d23, d27, d16              ; t_oq2 |= r_oq2 & (flat & mask)
+    vbif        d23, d10, d16              ; t_oq2 |= q2 & ~(flat & mask)
+
+    tst         r7, #2
+    bxne        lr
+
+    ; wide_mbfilter flat2 && flat && mask branch
+    vmov.u8     d16, #7
+    vaddl.u8    q15, d7, d8                ; op6 = p0 + q0
+    vaddl.u8    q12, d2, d3
+    vaddl.u8    q13, d4, d5
+    vaddl.u8    q14, d1, d6
+    vmlal.u8    q15, d0, d16               ; op6 += p7 * 3
+    vadd.i16    q12, q13
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d2, d9
+    vadd.i16    q15, q12
+    vaddl.u8    q12, d0, d1
+    vaddw.u8    q15, d1
+    vaddl.u8    q13, d0, d2
+    vadd.i16    q14, q15, q14
+    vqrshrn.u16 d16, q15, #4               ; w_op6
+
+    vsub.i16    q15, q14, q12
+    vaddl.u8    q14, d3, d10
+    vqrshrn.u16 d24, q15, #4               ; w_op5
+
+    vsub.i16    q15, q13
+    vaddl.u8    q13, d0, d3
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d4, d11
+    vqrshrn.u16 d25, q15, #4               ; w_op4
+
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d0, d4
+    vsub.i16    q15, q13
+    vsub.i16    q14, q15, q14
+    vqrshrn.u16 d26, q15, #4               ; w_op3
+
+    vaddw.u8    q15, q14, d5               ; op2 += p2
+    vaddl.u8    q14, d0, d5
+    vaddw.u8    q15, d12                   ; op2 += q4
+    vbif        d26, d4, d17               ; op3 |= p3 & ~(f2 & f & m)
+    vqrshrn.u16 d27, q15, #4               ; w_op2
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d0, d6
+    vaddw.u8    q15, d6                    ; op1 += p1
+    vaddw.u8    q15, d13                   ; op1 += q5
+    vbif        d27, d18, d17              ; op2 |= t_op2 & ~(f2 & f & m)
+    vqrshrn.u16 d18, q15, #4               ; w_op1
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d0, d7
+    vaddw.u8    q15, d7                    ; op0 += p0
+    vaddw.u8    q15, d14                   ; op0 += q6
+    vbif        d18, d19, d17              ; op1 |= t_op1 & ~(f2 & f & m)
+    vqrshrn.u16 d19, q15, #4               ; w_op0
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d1, d8
+    vaddw.u8    q15, d8                    ; oq0 += q0
+    vaddw.u8    q15, d15                   ; oq0 += q7
+    vbif        d19, d20, d17              ; op0 |= t_op0 & ~(f2 & f & m)
+    vqrshrn.u16 d20, q15, #4               ; w_oq0
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d2, d9
+    vaddw.u8    q15, d9                    ; oq1 += q1
+    vaddl.u8    q4, d10, d15
+    vaddw.u8    q15, d15                   ; oq1 += q7
+    vbif        d20, d21, d17              ; oq0 |= t_oq0 & ~(f2 & f & m)
+    vqrshrn.u16 d21, q15, #4               ; w_oq1
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d3, d10
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d11, d15
+    vbif        d21, d22, d17              ; oq1 |= t_oq1 & ~(f2 & f & m)
+    vqrshrn.u16 d22, q15, #4               ; w_oq2
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d4, d11
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d12, d15
+    vbif        d22, d23, d17              ; oq2 |= t_oq2 & ~(f2 & f & m)
+    vqrshrn.u16 d23, q15, #4               ; w_oq3
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d5, d12
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d13, d15
+    vbif        d16, d1, d17               ; op6 |= p6 & ~(f2 & f & m)
+    vqrshrn.u16 d1, q15, #4                ; w_oq4
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d6, d13
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d14, d15
+    vbif        d24, d2, d17               ; op5 |= p5 & ~(f2 & f & m)
+    vqrshrn.u16 d2, q15, #4                ; w_oq5
+
+    vsub.i16    q15, q14
+    vbif        d25, d3, d17               ; op4 |= p4 & ~(f2 & f & m)
+    vadd.i16    q15, q4
+    vbif        d23, d11, d17              ; oq3 |= q3 & ~(f2 & f & m)
+    vqrshrn.u16 d3, q15, #4                ; w_oq6
+    vbif        d1, d12, d17               ; oq4 |= q4 & ~(f2 & f & m)
+    vbif        d2, d13, d17               ; oq5 |= q5 & ~(f2 & f & m)
+    vbif        d3, d14, d17               ; oq6 |= q6 & ~(f2 & f & m)
+
+    bx          lr
+    ENDP        ; |vpx_wide_mbfilter_neon|
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/loopfilter_neon.c b/libvpx/libvpx/vpx_dsp/arm/loopfilter_neon.c
new file mode 100644
index 0000000..aa31f29
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/loopfilter_neon.c
@@ -0,0 +1,58 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+void vpx_lpf_vertical_4_dual_neon(uint8_t *s, int p,
+                                  const uint8_t *blimit0,
+                                  const uint8_t *limit0,
+                                  const uint8_t *thresh0,
+                                  const uint8_t *blimit1,
+                                  const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  vpx_lpf_vertical_4_neon(s, p, blimit0, limit0, thresh0);
+  vpx_lpf_vertical_4_neon(s + 8 * p, p, blimit1, limit1, thresh1);
+}
+
+#if HAVE_NEON_ASM
+void vpx_lpf_horizontal_8_dual_neon(uint8_t *s, int p /* pitch */,
+                                    const uint8_t *blimit0,
+                                    const uint8_t *limit0,
+                                    const uint8_t *thresh0,
+                                    const uint8_t *blimit1,
+                                    const uint8_t *limit1,
+                                    const uint8_t *thresh1) {
+  vpx_lpf_horizontal_8_neon(s, p, blimit0, limit0, thresh0);
+  vpx_lpf_horizontal_8_neon(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void vpx_lpf_vertical_8_dual_neon(uint8_t *s, int p,
+                                  const uint8_t *blimit0,
+                                  const uint8_t *limit0,
+                                  const uint8_t *thresh0,
+                                  const uint8_t *blimit1,
+                                  const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  vpx_lpf_vertical_8_neon(s, p, blimit0, limit0, thresh0);
+  vpx_lpf_vertical_8_neon(s + 8 * p, p, blimit1, limit1, thresh1);
+}
+
+void vpx_lpf_vertical_16_dual_neon(uint8_t *s, int p,
+                                   const uint8_t *blimit,
+                                   const uint8_t *limit,
+                                   const uint8_t *thresh) {
+  vpx_lpf_vertical_16_neon(s, p, blimit, limit, thresh);
+  vpx_lpf_vertical_16_neon(s + 8 * p, p, blimit, limit, thresh);
+}
+#endif  // HAVE_NEON_ASM
diff --git a/libvpx/libvpx/vpx_dsp/arm/sad4d_neon.c b/libvpx/libvpx/vpx_dsp/arm/sad4d_neon.c
new file mode 100644
index 0000000..c7704dc
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/sad4d_neon.c
@@ -0,0 +1,226 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
+                                                    const uint16x8_t vec_hi) {
+  const uint32x4_t vec_l_lo = vaddl_u16(vget_low_u16(vec_lo),
+                                        vget_high_u16(vec_lo));
+  const uint32x4_t vec_l_hi = vaddl_u16(vget_low_u16(vec_hi),
+                                        vget_high_u16(vec_hi));
+  const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+
+// Calculate the absolute difference of 64 bytes from vec_src_00, vec_src_16,
+// vec_src_32, vec_src_48 and ref. Accumulate partial sums in vec_sum_ref_lo
+// and vec_sum_ref_hi.
+static void sad_neon_64(const uint8x16_t vec_src_00,
+                        const uint8x16_t vec_src_16,
+                        const uint8x16_t vec_src_32,
+                        const uint8x16_t vec_src_48,
+                        const uint8_t *ref,
+                        uint16x8_t *vec_sum_ref_lo,
+                        uint16x8_t *vec_sum_ref_hi) {
+  const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+  const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+  const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32);
+  const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48);
+
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_00),
+                             vget_low_u8(vec_ref_00));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_00),
+                             vget_high_u8(vec_ref_00));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_16),
+                             vget_low_u8(vec_ref_16));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_16),
+                             vget_high_u8(vec_ref_16));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_32),
+                             vget_low_u8(vec_ref_32));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_32),
+                             vget_high_u8(vec_ref_32));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_48),
+                             vget_low_u8(vec_ref_48));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_48),
+                             vget_high_u8(vec_ref_48));
+}
+
+// Calculate the absolute difference of 32 bytes from vec_src_00, vec_src_16,
+// and ref. Accumulate partial sums in vec_sum_ref_lo and vec_sum_ref_hi.
+static void sad_neon_32(const uint8x16_t vec_src_00,
+                        const uint8x16_t vec_src_16,
+                        const uint8_t *ref,
+                        uint16x8_t *vec_sum_ref_lo,
+                        uint16x8_t *vec_sum_ref_hi) {
+  const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+  const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_00),
+                             vget_low_u8(vec_ref_00));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_00),
+                             vget_high_u8(vec_ref_00));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_16),
+                             vget_low_u8(vec_ref_16));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_16),
+                             vget_high_u8(vec_ref_16));
+}
+
+void vpx_sad64x64x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t* const ref[4], int ref_stride,
+                          uint32_t *res) {
+  int i;
+  uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+
+  for (i = 0; i < 64; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_src_32 = vld1q_u8(src + 32);
+    const uint8x16_t vec_src_48 = vld1q_u8(src + 48);
+
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref0,
+                &vec_sum_ref0_lo, &vec_sum_ref0_hi);
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref1,
+                &vec_sum_ref1_lo, &vec_sum_ref1_hi);
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref2,
+                &vec_sum_ref2_lo, &vec_sum_ref2_hi);
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref3,
+                &vec_sum_ref3_lo, &vec_sum_ref3_hi);
+
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+
+  res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
+  res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
+  res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
+  res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
+}
+
+void vpx_sad32x32x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t* const ref[4], int ref_stride,
+                          uint32_t *res) {
+  int i;
+  uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+
+  for (i = 0; i < 32; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+
+    sad_neon_32(vec_src_00, vec_src_16, ref0,
+                &vec_sum_ref0_lo, &vec_sum_ref0_hi);
+    sad_neon_32(vec_src_00, vec_src_16, ref1,
+                &vec_sum_ref1_lo, &vec_sum_ref1_hi);
+    sad_neon_32(vec_src_00, vec_src_16, ref2,
+                &vec_sum_ref2_lo, &vec_sum_ref2_hi);
+    sad_neon_32(vec_src_00, vec_src_16, ref3,
+                &vec_sum_ref3_lo, &vec_sum_ref3_hi);
+
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+
+  res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
+  res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
+  res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
+  res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
+}
+
+void vpx_sad16x16x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t* const ref[4], int ref_stride,
+                          uint32_t *res) {
+  int i;
+  uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+
+  for (i = 0; i < 16; ++i) {
+    const uint8x16_t vec_src = vld1q_u8(src);
+    const uint8x16_t vec_ref0 = vld1q_u8(ref0);
+    const uint8x16_t vec_ref1 = vld1q_u8(ref1);
+    const uint8x16_t vec_ref2 = vld1q_u8(ref2);
+    const uint8x16_t vec_ref3 = vld1q_u8(ref3);
+
+    vec_sum_ref0_lo = vabal_u8(vec_sum_ref0_lo, vget_low_u8(vec_src),
+                               vget_low_u8(vec_ref0));
+    vec_sum_ref0_hi = vabal_u8(vec_sum_ref0_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref0));
+    vec_sum_ref1_lo = vabal_u8(vec_sum_ref1_lo, vget_low_u8(vec_src),
+                               vget_low_u8(vec_ref1));
+    vec_sum_ref1_hi = vabal_u8(vec_sum_ref1_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref1));
+    vec_sum_ref2_lo = vabal_u8(vec_sum_ref2_lo, vget_low_u8(vec_src),
+                               vget_low_u8(vec_ref2));
+    vec_sum_ref2_hi = vabal_u8(vec_sum_ref2_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref2));
+    vec_sum_ref3_lo = vabal_u8(vec_sum_ref3_lo, vget_low_u8(vec_src),
+                               vget_low_u8(vec_ref3));
+    vec_sum_ref3_hi = vabal_u8(vec_sum_ref3_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref3));
+
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+
+  res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
+  res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
+  res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
+  res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/sad_media.asm b/libvpx/libvpx/vpx_dsp/arm/sad_media.asm
new file mode 100644
index 0000000..aed1d3a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/sad_media.asm
@@ -0,0 +1,95 @@
+;
+;  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vpx_sad16x16_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    const unsigned char *src_ptr
+; r1    int  src_stride
+; r2    const unsigned char *ref_ptr
+; r3    int  ref_stride
+|vpx_sad16x16_media| PROC
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+    pld     [r0, r1, lsl #1]
+    pld     [r2, r3, lsl #1]
+
+    mov     r4, #0              ; sad = 0;
+    mov     r5, #8              ; loop count
+
+loop
+    ; 1st row
+    ldr     r6, [r0, #0x0]      ; load 4 src pixels (1A)
+    ldr     r8, [r2, #0x0]      ; load 4 ref pixels (1A)
+    ldr     r7, [r0, #0x4]      ; load 4 src pixels (1A)
+    ldr     r9, [r2, #0x4]      ; load 4 ref pixels (1A)
+    ldr     r10, [r0, #0x8]     ; load 4 src pixels (1B)
+    ldr     r11, [r0, #0xC]     ; load 4 src pixels (1B)
+
+    usada8  r4, r8, r6, r4      ; calculate sad for 4 pixels
+    usad8   r8, r7, r9          ; calculate sad for 4 pixels
+
+    ldr     r12, [r2, #0x8]     ; load 4 ref pixels (1B)
+    ldr     lr, [r2, #0xC]      ; load 4 ref pixels (1B)
+
+    add     r0, r0, r1          ; set src pointer to next row
+    add     r2, r2, r3          ; set dst pointer to next row
+
+    pld     [r0, r1, lsl #1]
+    pld     [r2, r3, lsl #1]
+
+    usada8  r4, r10, r12, r4    ; calculate sad for 4 pixels
+    usada8  r8, r11, lr, r8     ; calculate sad for 4 pixels
+
+    ldr     r6, [r0, #0x0]      ; load 4 src pixels (2A)
+    ldr     r7, [r0, #0x4]      ; load 4 src pixels (2A)
+    add     r4, r4, r8          ; add partial sad values
+
+    ; 2nd row
+    ldr     r8, [r2, #0x0]      ; load 4 ref pixels (2A)
+    ldr     r9, [r2, #0x4]      ; load 4 ref pixels (2A)
+    ldr     r10, [r0, #0x8]     ; load 4 src pixels (2B)
+    ldr     r11, [r0, #0xC]     ; load 4 src pixels (2B)
+
+    usada8  r4, r6, r8, r4      ; calculate sad for 4 pixels
+    usad8   r8, r7, r9          ; calculate sad for 4 pixels
+
+    ldr     r12, [r2, #0x8]     ; load 4 ref pixels (2B)
+    ldr     lr, [r2, #0xC]      ; load 4 ref pixels (2B)
+
+    add     r0, r0, r1          ; set src pointer to next row
+    add     r2, r2, r3          ; set dst pointer to next row
+
+    usada8  r4, r10, r12, r4    ; calculate sad for 4 pixels
+    usada8  r8, r11, lr, r8     ; calculate sad for 4 pixels
+
+    pld     [r0, r1, lsl #1]
+    pld     [r2, r3, lsl #1]
+
+    subs    r5, r5, #1          ; decrement loop counter
+    add     r4, r4, r8          ; add partial sad values
+
+    bne     loop
+
+    mov     r0, r4              ; return sad
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+    END
+
diff --git a/libvpx/libvpx/vpx_dsp/arm/sad_neon.c b/libvpx/libvpx/vpx_dsp/arm/sad_neon.c
new file mode 100644
index 0000000..173f08a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/sad_neon.c
@@ -0,0 +1,232 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+unsigned int vpx_sad8x16_neon(
+        unsigned char *src_ptr,
+        int src_stride,
+        unsigned char *ref_ptr,
+        int ref_stride) {
+    uint8x8_t d0, d8;
+    uint16x8_t q12;
+    uint32x4_t q1;
+    uint64x2_t q3;
+    uint32x2_t d5;
+    int i;
+
+    d0 = vld1_u8(src_ptr);
+    src_ptr += src_stride;
+    d8 = vld1_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabdl_u8(d0, d8);
+
+    for (i = 0; i < 15; i++) {
+        d0 = vld1_u8(src_ptr);
+        src_ptr += src_stride;
+        d8 = vld1_u8(ref_ptr);
+        ref_ptr += ref_stride;
+        q12 = vabal_u8(q12, d0, d8);
+    }
+
+    q1 = vpaddlq_u16(q12);
+    q3 = vpaddlq_u32(q1);
+    d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
+                  vreinterpret_u32_u64(vget_high_u64(q3)));
+
+    return vget_lane_u32(d5, 0);
+}
+
+unsigned int vpx_sad4x4_neon(
+        unsigned char *src_ptr,
+        int src_stride,
+        unsigned char *ref_ptr,
+        int ref_stride) {
+    uint8x8_t d0, d8;
+    uint16x8_t q12;
+    uint32x2_t d1;
+    uint64x1_t d3;
+    int i;
+
+    d0 = vld1_u8(src_ptr);
+    src_ptr += src_stride;
+    d8 = vld1_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabdl_u8(d0, d8);
+
+    for (i = 0; i < 3; i++) {
+        d0 = vld1_u8(src_ptr);
+        src_ptr += src_stride;
+        d8 = vld1_u8(ref_ptr);
+        ref_ptr += ref_stride;
+        q12 = vabal_u8(q12, d0, d8);
+    }
+
+    d1 = vpaddl_u16(vget_low_u16(q12));
+    d3 = vpaddl_u32(d1);
+
+    return vget_lane_u32(vreinterpret_u32_u64(d3), 0);
+}
+
+unsigned int vpx_sad16x8_neon(
+        unsigned char *src_ptr,
+        int src_stride,
+        unsigned char *ref_ptr,
+        int ref_stride) {
+    uint8x16_t q0, q4;
+    uint16x8_t q12, q13;
+    uint32x4_t q1;
+    uint64x2_t q3;
+    uint32x2_t d5;
+    int i;
+
+    q0 = vld1q_u8(src_ptr);
+    src_ptr += src_stride;
+    q4 = vld1q_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabdl_u8(vget_low_u8(q0), vget_low_u8(q4));
+    q13 = vabdl_u8(vget_high_u8(q0), vget_high_u8(q4));
+
+    for (i = 0; i < 7; i++) {
+        q0 = vld1q_u8(src_ptr);
+        src_ptr += src_stride;
+        q4 = vld1q_u8(ref_ptr);
+        ref_ptr += ref_stride;
+        q12 = vabal_u8(q12, vget_low_u8(q0), vget_low_u8(q4));
+        q13 = vabal_u8(q13, vget_high_u8(q0), vget_high_u8(q4));
+    }
+
+    q12 = vaddq_u16(q12, q13);
+    q1 = vpaddlq_u16(q12);
+    q3 = vpaddlq_u32(q1);
+    d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
+                  vreinterpret_u32_u64(vget_high_u64(q3)));
+
+    return vget_lane_u32(d5, 0);
+}
+
+static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
+                                                    const uint16x8_t vec_hi) {
+  const uint32x4_t vec_l_lo = vaddl_u16(vget_low_u16(vec_lo),
+                                        vget_high_u16(vec_lo));
+  const uint32x4_t vec_l_hi = vaddl_u16(vget_low_u16(vec_hi),
+                                        vget_high_u16(vec_hi));
+  const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+static INLINE unsigned int horizontal_add_16x8(const uint16x8_t vec_16x8) {
+  const uint32x4_t a = vpaddlq_u16(vec_16x8);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+
+unsigned int vpx_sad64x64_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+  for (i = 0; i < 64; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_src_32 = vld1q_u8(src + 32);
+    const uint8x16_t vec_src_48 = vld1q_u8(src + 48);
+    const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+    const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+    const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32);
+    const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
+                            vget_low_u8(vec_ref_00));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
+                            vget_high_u8(vec_ref_00));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
+                            vget_low_u8(vec_ref_16));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
+                            vget_high_u8(vec_ref_16));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_32),
+                            vget_low_u8(vec_ref_32));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_32),
+                            vget_high_u8(vec_ref_32));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_48),
+                            vget_low_u8(vec_ref_48));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_48),
+                            vget_high_u8(vec_ref_48));
+  }
+  return horizontal_long_add_16x8(vec_accum_lo, vec_accum_hi);
+}
+
+unsigned int vpx_sad32x32_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+
+  for (i = 0; i < 32; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+    const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
+                            vget_low_u8(vec_ref_00));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
+                            vget_high_u8(vec_ref_00));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
+                            vget_low_u8(vec_ref_16));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
+                            vget_high_u8(vec_ref_16));
+  }
+  return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
+}
+
+unsigned int vpx_sad16x16_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+
+  for (i = 0; i < 16; ++i) {
+    const uint8x16_t vec_src = vld1q_u8(src);
+    const uint8x16_t vec_ref = vld1q_u8(ref);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src),
+                            vget_low_u8(vec_ref));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src),
+                            vget_high_u8(vec_ref));
+  }
+  return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
+}
+
+unsigned int vpx_sad8x8_neon(const uint8_t *src, int src_stride,
+                             const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum = vdupq_n_u16(0);
+
+  for (i = 0; i < 8; ++i) {
+    const uint8x8_t vec_src = vld1_u8(src);
+    const uint8x8_t vec_ref = vld1_u8(ref);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum = vabal_u8(vec_accum, vec_src, vec_ref);
+  }
+  return horizontal_add_16x8(vec_accum);
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/save_reg_neon.asm b/libvpx/libvpx/vpx_dsp/arm/save_reg_neon.asm
new file mode 100644
index 0000000..c9ca108
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/save_reg_neon.asm
@@ -0,0 +1,36 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vpx_push_neon|
+    EXPORT  |vpx_pop_neon|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|vpx_push_neon| PROC
+    vst1.i64            {d8, d9, d10, d11}, [r0]!
+    vst1.i64            {d12, d13, d14, d15}, [r0]!
+    bx              lr
+
+    ENDP
+
+|vpx_pop_neon| PROC
+    vld1.i64            {d8, d9, d10, d11}, [r0]!
+    vld1.i64            {d12, d13, d14, d15}, [r0]!
+    bx              lr
+
+    ENDP
+
+    END
+
diff --git a/libvpx/libvpx/vpx_dsp/arm/subpel_variance_media.c b/libvpx/libvpx/vpx_dsp/arm/subpel_variance_media.c
new file mode 100644
index 0000000..e7d8c85
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/subpel_variance_media.c
@@ -0,0 +1,105 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_MEDIA
+static const int16_t bilinear_filters_media[8][2] = {
+  { 128,   0 },
+  { 112,  16 },
+  {  96,  32 },
+  {  80,  48 },
+  {  64,  64 },
+  {  48,  80 },
+  {  32,  96 },
+  {  16, 112 }
+};
+
+extern void vpx_filter_block2d_bil_first_pass_media(const uint8_t *src_ptr,
+                                                    uint16_t *dst_ptr,
+                                                    uint32_t src_pitch,
+                                                    uint32_t height,
+                                                    uint32_t width,
+                                                    const int16_t *filter);
+
+extern void vpx_filter_block2d_bil_second_pass_media(const uint16_t *src_ptr,
+                                                     uint8_t *dst_ptr,
+                                                     int32_t src_pitch,
+                                                     uint32_t height,
+                                                     uint32_t width,
+                                                     const int16_t *filter);
+
+
+unsigned int vpx_sub_pixel_variance8x8_media(const uint8_t *src_ptr,
+                                             int src_pixels_per_line,
+                                             int xoffset, int yoffset,
+                                             const uint8_t *dst_ptr,
+                                             int dst_pixels_per_line,
+                                             unsigned int *sse) {
+  uint16_t first_pass[10*8];
+  uint8_t  second_pass[8*8];
+  const int16_t *HFilter, *VFilter;
+
+  HFilter = bilinear_filters_media[xoffset];
+  VFilter = bilinear_filters_media[yoffset];
+
+  vpx_filter_block2d_bil_first_pass_media(src_ptr, first_pass,
+                                          src_pixels_per_line,
+                                          9, 8, HFilter);
+  vpx_filter_block2d_bil_second_pass_media(first_pass, second_pass,
+                                           8, 8, 8, VFilter);
+
+  return vpx_variance8x8_media(second_pass, 8, dst_ptr,
+                               dst_pixels_per_line, sse);
+}
+
+unsigned int vpx_sub_pixel_variance16x16_media(const uint8_t *src_ptr,
+                                               int src_pixels_per_line,
+                                               int xoffset,
+                                               int yoffset,
+                                               const uint8_t *dst_ptr,
+                                               int dst_pixels_per_line,
+                                               unsigned int *sse) {
+  uint16_t first_pass[36*16];
+  uint8_t  second_pass[20*16];
+  const int16_t *HFilter, *VFilter;
+  unsigned int var;
+
+  if (xoffset == 4 && yoffset == 0) {
+    var = vpx_variance_halfpixvar16x16_h_media(src_ptr, src_pixels_per_line,
+                                               dst_ptr, dst_pixels_per_line,
+                                               sse);
+  } else if (xoffset == 0 && yoffset == 4) {
+    var = vpx_variance_halfpixvar16x16_v_media(src_ptr, src_pixels_per_line,
+                                               dst_ptr, dst_pixels_per_line,
+                                               sse);
+  } else if (xoffset == 4 && yoffset == 4) {
+    var = vpx_variance_halfpixvar16x16_hv_media(src_ptr, src_pixels_per_line,
+                                                dst_ptr, dst_pixels_per_line,
+                                                sse);
+  } else {
+    HFilter = bilinear_filters_media[xoffset];
+    VFilter = bilinear_filters_media[yoffset];
+
+    vpx_filter_block2d_bil_first_pass_media(src_ptr, first_pass,
+                                            src_pixels_per_line,
+                                            17, 16, HFilter);
+    vpx_filter_block2d_bil_second_pass_media(first_pass, second_pass,
+                                             16, 16, 16, VFilter);
+
+    var = vpx_variance16x16_media(second_pass, 16, dst_ptr,
+                                  dst_pixels_per_line, sse);
+  }
+  return var;
+}
+#endif  // HAVE_MEDIA
diff --git a/libvpx/libvpx/vpx_dsp/arm/subpel_variance_neon.c b/libvpx/libvpx/vpx_dsp/arm/subpel_variance_neon.c
new file mode 100644
index 0000000..40e2cc8
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/subpel_variance_neon.c
@@ -0,0 +1,152 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx_ports/mem.h"
+#include "vpx/vpx_integer.h"
+
+#include "vpx_dsp/variance.h"
+
+static const uint8_t bilinear_filters[8][2] = {
+  { 128,   0, },
+  { 112,  16, },
+  {  96,  32, },
+  {  80,  48, },
+  {  64,  64, },
+  {  48,  80, },
+  {  32,  96, },
+  {  16, 112, },
+};
+
+static void var_filter_block2d_bil_w8(const uint8_t *src_ptr,
+                                      uint8_t *output_ptr,
+                                      unsigned int src_pixels_per_line,
+                                      int pixel_step,
+                                      unsigned int output_height,
+                                      unsigned int output_width,
+                                      const uint8_t *filter) {
+  const uint8x8_t f0 = vmov_n_u8(filter[0]);
+  const uint8x8_t f1 = vmov_n_u8(filter[1]);
+  unsigned int i;
+  for (i = 0; i < output_height; ++i) {
+    const uint8x8_t src_0 = vld1_u8(&src_ptr[0]);
+    const uint8x8_t src_1 = vld1_u8(&src_ptr[pixel_step]);
+    const uint16x8_t a = vmull_u8(src_0, f0);
+    const uint16x8_t b = vmlal_u8(a, src_1, f1);
+    const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS);
+    vst1_u8(&output_ptr[0], out);
+    // Next row...
+    src_ptr += src_pixels_per_line;
+    output_ptr += output_width;
+  }
+}
+
+static void var_filter_block2d_bil_w16(const uint8_t *src_ptr,
+                                       uint8_t *output_ptr,
+                                       unsigned int src_pixels_per_line,
+                                       int pixel_step,
+                                       unsigned int output_height,
+                                       unsigned int output_width,
+                                       const uint8_t *filter) {
+  const uint8x8_t f0 = vmov_n_u8(filter[0]);
+  const uint8x8_t f1 = vmov_n_u8(filter[1]);
+  unsigned int i, j;
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; j += 16) {
+      const uint8x16_t src_0 = vld1q_u8(&src_ptr[j]);
+      const uint8x16_t src_1 = vld1q_u8(&src_ptr[j + pixel_step]);
+      const uint16x8_t a = vmull_u8(vget_low_u8(src_0), f0);
+      const uint16x8_t b = vmlal_u8(a, vget_low_u8(src_1), f1);
+      const uint8x8_t out_lo = vrshrn_n_u16(b, FILTER_BITS);
+      const uint16x8_t c = vmull_u8(vget_high_u8(src_0), f0);
+      const uint16x8_t d = vmlal_u8(c, vget_high_u8(src_1), f1);
+      const uint8x8_t out_hi = vrshrn_n_u16(d, FILTER_BITS);
+      vst1q_u8(&output_ptr[j], vcombine_u8(out_lo, out_hi));
+    }
+    // Next row...
+    src_ptr += src_pixels_per_line;
+    output_ptr += output_width;
+  }
+}
+
+unsigned int vpx_sub_pixel_variance8x8_neon(const uint8_t *src,
+                                            int src_stride,
+                                            int xoffset,
+                                            int yoffset,
+                                            const uint8_t *dst,
+                                            int dst_stride,
+                                            unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[8 * 8]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[9 * 8]);
+
+  var_filter_block2d_bil_w8(src, fdata3, src_stride, 1,
+                            9, 8,
+                            bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w8(fdata3, temp2, 8, 8, 8,
+                            8, bilinear_filters[yoffset]);
+  return vpx_variance8x8_neon(temp2, 8, dst, dst_stride, sse);
+}
+
+unsigned int vpx_sub_pixel_variance16x16_neon(const uint8_t *src,
+                                              int src_stride,
+                                              int xoffset,
+                                              int yoffset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[16 * 16]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[17 * 16]);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, 1,
+                             17, 16,
+                             bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w16(fdata3, temp2, 16, 16, 16,
+                             16, bilinear_filters[yoffset]);
+  return vpx_variance16x16_neon(temp2, 16, dst, dst_stride, sse);
+}
+
+unsigned int vpx_sub_pixel_variance32x32_neon(const uint8_t *src,
+                                              int src_stride,
+                                              int xoffset,
+                                              int yoffset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[32 * 32]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[33 * 32]);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, 1,
+                             33, 32,
+                             bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w16(fdata3, temp2, 32, 32, 32,
+                             32, bilinear_filters[yoffset]);
+  return vpx_variance32x32_neon(temp2, 32, dst, dst_stride, sse);
+}
+
+unsigned int vpx_sub_pixel_variance64x64_neon(const uint8_t *src,
+                                              int src_stride,
+                                              int xoffset,
+                                              int yoffset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[64 * 64]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[65 * 64]);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, 1,
+                             65, 64,
+                             bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w16(fdata3, temp2, 64, 64, 64,
+                             64, bilinear_filters[yoffset]);
+  return vpx_variance64x64_neon(temp2, 64, dst, dst_stride, sse);
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/subtract_neon.c b/libvpx/libvpx/vpx_dsp/arm/subtract_neon.c
new file mode 100644
index 0000000..7b14609
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/subtract_neon.c
@@ -0,0 +1,80 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+void vpx_subtract_block_neon(int rows, int cols,
+                             int16_t *diff, ptrdiff_t diff_stride,
+                             const uint8_t *src, ptrdiff_t src_stride,
+                             const uint8_t *pred, ptrdiff_t pred_stride) {
+  int r, c;
+
+  if (cols > 16) {
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; c += 32) {
+        const uint8x16_t v_src_00 = vld1q_u8(&src[c + 0]);
+        const uint8x16_t v_src_16 = vld1q_u8(&src[c + 16]);
+        const uint8x16_t v_pred_00 = vld1q_u8(&pred[c +  0]);
+        const uint8x16_t v_pred_16 = vld1q_u8(&pred[c + 16]);
+        const uint16x8_t v_diff_lo_00 = vsubl_u8(vget_low_u8(v_src_00),
+                                                 vget_low_u8(v_pred_00));
+        const uint16x8_t v_diff_hi_00 = vsubl_u8(vget_high_u8(v_src_00),
+                                                 vget_high_u8(v_pred_00));
+        const uint16x8_t v_diff_lo_16 = vsubl_u8(vget_low_u8(v_src_16),
+                                                 vget_low_u8(v_pred_16));
+        const uint16x8_t v_diff_hi_16 = vsubl_u8(vget_high_u8(v_src_16),
+                                                 vget_high_u8(v_pred_16));
+        vst1q_s16(&diff[c +  0], vreinterpretq_s16_u16(v_diff_lo_00));
+        vst1q_s16(&diff[c +  8], vreinterpretq_s16_u16(v_diff_hi_00));
+        vst1q_s16(&diff[c + 16], vreinterpretq_s16_u16(v_diff_lo_16));
+        vst1q_s16(&diff[c + 24], vreinterpretq_s16_u16(v_diff_hi_16));
+      }
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  } else if (cols > 8) {
+    for (r = 0; r < rows; ++r) {
+      const uint8x16_t v_src = vld1q_u8(&src[0]);
+      const uint8x16_t v_pred = vld1q_u8(&pred[0]);
+      const uint16x8_t v_diff_lo = vsubl_u8(vget_low_u8(v_src),
+                                            vget_low_u8(v_pred));
+      const uint16x8_t v_diff_hi = vsubl_u8(vget_high_u8(v_src),
+                                            vget_high_u8(v_pred));
+      vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff_lo));
+      vst1q_s16(&diff[8], vreinterpretq_s16_u16(v_diff_hi));
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  } else if (cols > 4) {
+    for (r = 0; r < rows; ++r) {
+      const uint8x8_t v_src = vld1_u8(&src[0]);
+      const uint8x8_t v_pred = vld1_u8(&pred[0]);
+      const uint16x8_t v_diff = vsubl_u8(v_src, v_pred);
+      vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff));
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  } else {
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; ++c)
+        diff[c] = src[c] - pred[c];
+
+      diff += diff_stride;
+      pred += pred_stride;
+      src  += src_stride;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/variance_halfpixvar16x16_h_media.asm b/libvpx/libvpx/vpx_dsp/arm/variance_halfpixvar16x16_h_media.asm
new file mode 100644
index 0000000..dab845a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/variance_halfpixvar16x16_h_media.asm
@@ -0,0 +1,182 @@
+;
+;  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vpx_variance_halfpixvar16x16_h_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|vpx_variance_halfpixvar16x16_h_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    ldr     r10, c80808080
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     lr, #0              ; constant zero
+loop
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load 4 src pixels
+    ldr     r6, [r0, #1]        ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load 4 src pixels
+    ldr     r6, [r0, #5]        ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load 4 src pixels
+    ldr     r6, [r0, #9]        ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load 4 src pixels
+    ldr     r6, [r0, #13]       ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    subs    r12, r12, #1
+
+    bne     loop
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+c80808080
+    DCD     0x80808080
+
+    END
+
diff --git a/libvpx/libvpx/vpx_dsp/arm/variance_halfpixvar16x16_hv_media.asm b/libvpx/libvpx/vpx_dsp/arm/variance_halfpixvar16x16_hv_media.asm
new file mode 100644
index 0000000..01953b7
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/variance_halfpixvar16x16_hv_media.asm
@@ -0,0 +1,222 @@
+;
+;  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vpx_variance_halfpixvar16x16_hv_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|vpx_variance_halfpixvar16x16_hv_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    ldr     r10, c80808080
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     lr, #0              ; constant zero
+loop
+    add     r9, r0, r1          ; pointer to pixels on the next row
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load source pixels a, row N
+    ldr     r6, [r0, #1]        ; load source pixels b, row N
+    ldr     r5, [r9, #0]        ; load source pixels c, row N+1
+    ldr     r7, [r9, #1]        ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load source pixels a, row N
+    ldr     r6, [r0, #5]        ; load source pixels b, row N
+    ldr     r5, [r9, #4]        ; load source pixels c, row N+1
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    ldr     r7, [r9, #5]        ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load source pixels a, row N
+    ldr     r6, [r0, #9]        ; load source pixels b, row N
+    ldr     r5, [r9, #8]        ; load source pixels c, row N+1
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    ldr     r7, [r9, #9]        ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load source pixels a, row N
+    ldr     r6, [r0, #13]       ; load source pixels b, row N
+    ldr     r5, [r9, #12]       ; load source pixels c, row N+1
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+    ldr     r7, [r9, #13]       ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    subs    r12, r12, #1
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    bne     loop
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+c80808080
+    DCD     0x80808080
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/variance_halfpixvar16x16_v_media.asm b/libvpx/libvpx/vpx_dsp/arm/variance_halfpixvar16x16_v_media.asm
new file mode 100644
index 0000000..0d17acb
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/variance_halfpixvar16x16_v_media.asm
@@ -0,0 +1,184 @@
+;
+;  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vpx_variance_halfpixvar16x16_v_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|vpx_variance_halfpixvar16x16_v_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    ldr     r10, c80808080
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     lr, #0              ; constant zero
+loop
+    add     r9, r0, r1          ; set src pointer to next row
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load 4 src pixels
+    ldr     r6, [r9, #0]        ; load 4 src pixels from next row
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load 4 src pixels
+    ldr     r6, [r9, #4]        ; load 4 src pixels from next row
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load 4 src pixels
+    ldr     r6, [r9, #8]        ; load 4 src pixels from next row
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load 4 src pixels
+    ldr     r6, [r9, #12]       ; load 4 src pixels from next row
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+
+    subs    r12, r12, #1
+
+    bne     loop
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+c80808080
+    DCD     0x80808080
+
+    END
+
diff --git a/libvpx/libvpx/vpx_dsp/arm/variance_media.asm b/libvpx/libvpx/vpx_dsp/arm/variance_media.asm
new file mode 100644
index 0000000..f7f9e14
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/variance_media.asm
@@ -0,0 +1,358 @@
+;
+;  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vpx_variance16x16_media|
+    EXPORT  |vpx_variance8x8_media|
+    EXPORT  |vpx_mse16x16_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|vpx_variance16x16_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+
+loop16x16
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load 4 src pixels
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+
+    mov     lr, #0              ; constant zero
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load 4 src pixels
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load 4 src pixels
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load 4 src pixels
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+
+    subs    r12, r12, #1
+
+    bne     loop16x16
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|vpx_variance8x8_media| PROC
+
+    push    {r4-r10, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r12, #8             ; set loop counter to 8 (=block height)
+    mov     r4, #0              ; initialize sum = 0
+    mov     r5, #0              ; initialize sse = 0
+
+loop8x8
+    ; 1st 4 pixels
+    ldr     r6, [r0, #0x0]      ; load 4 src pixels
+    ldr     r7, [r2, #0x0]      ; load 4 ref pixels
+
+    mov     lr, #0              ; constant zero
+
+    usub8   r8, r6, r7          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r10, r8, lr         ; select bytes with positive difference
+    usub8   r9, r7, r6          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r6, r10, lr         ; calculate sum of positive differences
+    usad8   r7, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r10         ; differences of all 4 pixels
+    ; calculate total sum
+    add    r4, r4, r6           ; add positive differences to sum
+    sub    r4, r4, r7           ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r7, r8              ; byte (two pixels) to halfwords
+    uxtb16  r10, r8, ror #8     ; another two pixels to halfwords
+    smlad   r5, r7, r7, r5      ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r6, [r0, #0x4]      ; load 4 src pixels
+    ldr     r7, [r2, #0x4]      ; load 4 ref pixels
+    smlad   r5, r10, r10, r5    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r6, r7          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r10, r8, lr         ; select bytes with positive difference
+    usub8   r9, r7, r6          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r6, r10, lr         ; calculate sum of positive differences
+    usad8   r7, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r10         ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r4, r4, r6          ; add positive differences to sum
+    sub     r4, r4, r7          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r7, r8              ; byte (two pixels) to halfwords
+    uxtb16  r10, r8, ror #8     ; another two pixels to halfwords
+    smlad   r5, r7, r7, r5      ; dual signed multiply, add and accumulate (1)
+    subs    r12, r12, #1        ; next row
+    smlad   r5, r10, r10, r5    ; dual signed multiply, add and accumulate (2)
+
+    bne     loop8x8
+
+    ; return stuff
+    ldr     r8, [sp, #32]       ; get address of sse
+    mul     r1, r4, r4          ; sum * sum
+    str     r5, [r8]            ; store sse
+    sub     r0, r5, r1, ASR #6  ; return (sse - ((sum * sum) >> 6))
+
+    pop     {r4-r10, pc}
+
+    ENDP
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+;
+;note: Based on vpx_variance16x16_media. In this function, sum is never used.
+;      So, we can remove this part of calculation.
+
+|vpx_mse16x16_media| PROC
+
+    push    {r4-r9, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     r4, #0              ; initialize sse = 0
+
+loopmse
+    ; 1st 4 pixels
+    ldr     r5, [r0, #0x0]      ; load 4 src pixels
+    ldr     r6, [r2, #0x0]      ; load 4 ref pixels
+
+    mov     lr, #0              ; constant zero
+
+    usub8   r8, r5, r6          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+
+    ldr     r5, [r0, #0x4]      ; load 4 src pixels
+
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r6, [r2, #0x4]      ; load 4 ref pixels
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r5, r6          ; calculate difference
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+    ldr     r5, [r0, #0x8]      ; load 4 src pixels
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r6, [r2, #0x8]      ; load 4 ref pixels
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r5, r6          ; calculate difference
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+
+    ldr     r5, [r0, #0xc]      ; load 4 src pixels
+
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r6, [r2, #0xc]      ; load 4 ref pixels
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r5, r6          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+
+    subs    r12, r12, #1        ; next row
+
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    bne     loopmse
+
+    ; return stuff
+    ldr     r1, [sp, #28]       ; get address of sse
+    mov     r0, r4              ; return sse
+    str     r4, [r1]            ; store sse
+
+    pop     {r4-r9, pc}
+
+    ENDP
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/variance_neon.c b/libvpx/libvpx/vpx_dsp/arm/variance_neon.c
new file mode 100644
index 0000000..ede6e7b
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/variance_neon.c
@@ -0,0 +1,418 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+static INLINE int horizontal_add_s16x8(const int16x8_t v_16x8) {
+  const int32x4_t a = vpaddlq_s16(v_16x8);
+  const int64x2_t b = vpaddlq_s32(a);
+  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                               vreinterpret_s32_s64(vget_high_s64(b)));
+  return vget_lane_s32(c, 0);
+}
+
+static INLINE int horizontal_add_s32x4(const int32x4_t v_32x4) {
+  const int64x2_t b = vpaddlq_s32(v_32x4);
+  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                               vreinterpret_s32_s64(vget_high_s64(b)));
+  return vget_lane_s32(c, 0);
+}
+
+// w * h must be less than 2048 or local variable v_sum may overflow.
+static void variance_neon_w8(const uint8_t *a, int a_stride,
+                             const uint8_t *b, int b_stride,
+                             int w, int h, uint32_t *sse, int *sum) {
+  int i, j;
+  int16x8_t v_sum = vdupq_n_s16(0);
+  int32x4_t v_sse_lo = vdupq_n_s32(0);
+  int32x4_t v_sse_hi = vdupq_n_s32(0);
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; j += 8) {
+      const uint8x8_t v_a = vld1_u8(&a[j]);
+      const uint8x8_t v_b = vld1_u8(&b[j]);
+      const uint16x8_t v_diff = vsubl_u8(v_a, v_b);
+      const int16x8_t sv_diff = vreinterpretq_s16_u16(v_diff);
+      v_sum = vaddq_s16(v_sum, sv_diff);
+      v_sse_lo = vmlal_s16(v_sse_lo,
+                           vget_low_s16(sv_diff),
+                           vget_low_s16(sv_diff));
+      v_sse_hi = vmlal_s16(v_sse_hi,
+                           vget_high_s16(sv_diff),
+                           vget_high_s16(sv_diff));
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+
+  *sum = horizontal_add_s16x8(v_sum);
+  *sse = (unsigned int)horizontal_add_s32x4(vaddq_s32(v_sse_lo, v_sse_hi));
+}
+
+void vpx_get8x8var_neon(const uint8_t *a, int a_stride,
+                        const uint8_t *b, int b_stride,
+                        unsigned int *sse, int *sum) {
+  variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, sum);
+}
+
+void vpx_get16x16var_neon(const uint8_t *a, int a_stride,
+                          const uint8_t *b, int b_stride,
+                          unsigned int *sse, int *sum) {
+  variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, sum);
+}
+
+unsigned int vpx_variance8x8_neon(const uint8_t *a, int a_stride,
+                                  const uint8_t *b, int b_stride,
+                                  unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, &sum);
+  return *sse - (((int64_t)sum * sum) >> 6);  //  >> 6 = / 8 * 8
+}
+
+unsigned int vpx_variance16x16_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, &sum);
+  return *sse - (((int64_t)sum * sum) >> 8);  //  >> 8 = / 16 * 16
+}
+
+unsigned int vpx_variance32x32_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, 32, 32, sse, &sum);
+  return *sse - (((int64_t)sum * sum) >> 10);  // >> 10 = / 32 * 32
+}
+
+unsigned int vpx_variance32x64_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum1, sum2;
+  uint32_t sse1, sse2;
+  variance_neon_w8(a, a_stride, b, b_stride, 32, 32, &sse1, &sum1);
+  variance_neon_w8(a + (32 * a_stride), a_stride,
+                   b + (32 * b_stride), b_stride, 32, 32,
+                   &sse2, &sum2);
+  *sse = sse1 + sse2;
+  sum1 += sum2;
+  return *sse - (((int64_t)sum1 * sum1) >> 11);  // >> 11 = / 32 * 64
+}
+
+unsigned int vpx_variance64x32_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum1, sum2;
+  uint32_t sse1, sse2;
+  variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1);
+  variance_neon_w8(a + (16 * a_stride), a_stride,
+                   b + (16 * b_stride), b_stride, 64, 16,
+                   &sse2, &sum2);
+  *sse = sse1 + sse2;
+  sum1 += sum2;
+  return *sse - (((int64_t)sum1 * sum1) >> 11);  // >> 11 = / 32 * 64
+}
+
+unsigned int vpx_variance64x64_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum1, sum2;
+  uint32_t sse1, sse2;
+
+  variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1);
+  variance_neon_w8(a + (16 * a_stride), a_stride,
+                   b + (16 * b_stride), b_stride, 64, 16,
+                   &sse2, &sum2);
+  sse1 += sse2;
+  sum1 += sum2;
+
+  variance_neon_w8(a + (16 * 2 * a_stride), a_stride,
+                   b + (16 * 2 * b_stride), b_stride,
+                   64, 16, &sse2, &sum2);
+  sse1 += sse2;
+  sum1 += sum2;
+
+  variance_neon_w8(a + (16 * 3 * a_stride), a_stride,
+                   b + (16 * 3 * b_stride), b_stride,
+                   64, 16, &sse2, &sum2);
+  *sse = sse1 + sse2;
+  sum1 += sum2;
+  return *sse - (((int64_t)sum1 * sum1) >> 12);  // >> 12 = / 64 * 64
+}
+
+unsigned int vpx_variance16x8_neon(
+        const unsigned char *src_ptr,
+        int source_stride,
+        const unsigned char *ref_ptr,
+        int recon_stride,
+        unsigned int *sse) {
+    int i;
+    int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
+    uint32x2_t d0u32, d10u32;
+    int64x1_t d0s64, d1s64;
+    uint8x16_t q0u8, q1u8, q2u8, q3u8;
+    uint16x8_t q11u16, q12u16, q13u16, q14u16;
+    int32x4_t q8s32, q9s32, q10s32;
+    int64x2_t q0s64, q1s64, q5s64;
+
+    q8s32 = vdupq_n_s32(0);
+    q9s32 = vdupq_n_s32(0);
+    q10s32 = vdupq_n_s32(0);
+
+    for (i = 0; i < 4; i++) {
+        q0u8 = vld1q_u8(src_ptr);
+        src_ptr += source_stride;
+        q1u8 = vld1q_u8(src_ptr);
+        src_ptr += source_stride;
+        __builtin_prefetch(src_ptr);
+
+        q2u8 = vld1q_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        q3u8 = vld1q_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        __builtin_prefetch(ref_ptr);
+
+        q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
+        q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
+        q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
+        q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
+
+        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
+        q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
+        q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
+
+        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
+        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+
+        d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+        d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
+        q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
+        q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
+
+        d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
+        d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
+        q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
+        q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
+    }
+
+    q10s32 = vaddq_s32(q10s32, q9s32);
+    q0s64 = vpaddlq_s32(q8s32);
+    q1s64 = vpaddlq_s32(q10s32);
+
+    d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
+    d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+    q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
+                      vreinterpret_s32_s64(d0s64));
+    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
+
+    d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
+    d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
+
+    return vget_lane_u32(d0u32, 0);
+}
+
+unsigned int vpx_variance8x16_neon(
+        const unsigned char *src_ptr,
+        int source_stride,
+        const unsigned char *ref_ptr,
+        int recon_stride,
+        unsigned int *sse) {
+    int i;
+    uint8x8_t d0u8, d2u8, d4u8, d6u8;
+    int16x4_t d22s16, d23s16, d24s16, d25s16;
+    uint32x2_t d0u32, d10u32;
+    int64x1_t d0s64, d1s64;
+    uint16x8_t q11u16, q12u16;
+    int32x4_t q8s32, q9s32, q10s32;
+    int64x2_t q0s64, q1s64, q5s64;
+
+    q8s32 = vdupq_n_s32(0);
+    q9s32 = vdupq_n_s32(0);
+    q10s32 = vdupq_n_s32(0);
+
+    for (i = 0; i < 8; i++) {
+        d0u8 = vld1_u8(src_ptr);
+        src_ptr += source_stride;
+        d2u8 = vld1_u8(src_ptr);
+        src_ptr += source_stride;
+        __builtin_prefetch(src_ptr);
+
+        d4u8 = vld1_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        d6u8 = vld1_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        __builtin_prefetch(ref_ptr);
+
+        q11u16 = vsubl_u8(d0u8, d4u8);
+        q12u16 = vsubl_u8(d2u8, d6u8);
+
+        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
+        q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
+        q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
+
+        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
+        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+    }
+
+    q10s32 = vaddq_s32(q10s32, q9s32);
+    q0s64 = vpaddlq_s32(q8s32);
+    q1s64 = vpaddlq_s32(q10s32);
+
+    d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
+    d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+    q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
+                      vreinterpret_s32_s64(d0s64));
+    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
+
+    d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
+    d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
+
+    return vget_lane_u32(d0u32, 0);
+}
+
+unsigned int vpx_mse16x16_neon(
+        const unsigned char *src_ptr,
+        int source_stride,
+        const unsigned char *ref_ptr,
+        int recon_stride,
+        unsigned int *sse) {
+    int i;
+    int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
+    int64x1_t d0s64;
+    uint8x16_t q0u8, q1u8, q2u8, q3u8;
+    int32x4_t q7s32, q8s32, q9s32, q10s32;
+    uint16x8_t q11u16, q12u16, q13u16, q14u16;
+    int64x2_t q1s64;
+
+    q7s32 = vdupq_n_s32(0);
+    q8s32 = vdupq_n_s32(0);
+    q9s32 = vdupq_n_s32(0);
+    q10s32 = vdupq_n_s32(0);
+
+    for (i = 0; i < 8; i++) {  // mse16x16_neon_loop
+        q0u8 = vld1q_u8(src_ptr);
+        src_ptr += source_stride;
+        q1u8 = vld1q_u8(src_ptr);
+        src_ptr += source_stride;
+        q2u8 = vld1q_u8(ref_ptr);
+        ref_ptr += recon_stride;
+        q3u8 = vld1q_u8(ref_ptr);
+        ref_ptr += recon_stride;
+
+        q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
+        q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
+        q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
+        q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
+
+        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+        q7s32 = vmlal_s16(q7s32, d22s16, d22s16);
+        q8s32 = vmlal_s16(q8s32, d23s16, d23s16);
+
+        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+
+        d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+        d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+        q7s32 = vmlal_s16(q7s32, d26s16, d26s16);
+        q8s32 = vmlal_s16(q8s32, d27s16, d27s16);
+
+        d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
+        d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
+        q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
+        q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
+    }
+
+    q7s32 = vaddq_s32(q7s32, q8s32);
+    q9s32 = vaddq_s32(q9s32, q10s32);
+    q10s32 = vaddq_s32(q7s32, q9s32);
+
+    q1s64 = vpaddlq_s32(q10s32);
+    d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d0s64), 0);
+    return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
+}
+
+unsigned int vpx_get4x4sse_cs_neon(
+        const unsigned char *src_ptr,
+        int source_stride,
+        const unsigned char *ref_ptr,
+        int recon_stride) {
+    int16x4_t d22s16, d24s16, d26s16, d28s16;
+    int64x1_t d0s64;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+    int32x4_t q7s32, q8s32, q9s32, q10s32;
+    uint16x8_t q11u16, q12u16, q13u16, q14u16;
+    int64x2_t q1s64;
+
+    d0u8 = vld1_u8(src_ptr);
+    src_ptr += source_stride;
+    d4u8 = vld1_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    d1u8 = vld1_u8(src_ptr);
+    src_ptr += source_stride;
+    d5u8 = vld1_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    d2u8 = vld1_u8(src_ptr);
+    src_ptr += source_stride;
+    d6u8 = vld1_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    d3u8 = vld1_u8(src_ptr);
+    src_ptr += source_stride;
+    d7u8 = vld1_u8(ref_ptr);
+    ref_ptr += recon_stride;
+
+    q11u16 = vsubl_u8(d0u8, d4u8);
+    q12u16 = vsubl_u8(d1u8, d5u8);
+    q13u16 = vsubl_u8(d2u8, d6u8);
+    q14u16 = vsubl_u8(d3u8, d7u8);
+
+    d22s16 = vget_low_s16(vreinterpretq_s16_u16(q11u16));
+    d24s16 = vget_low_s16(vreinterpretq_s16_u16(q12u16));
+    d26s16 = vget_low_s16(vreinterpretq_s16_u16(q13u16));
+    d28s16 = vget_low_s16(vreinterpretq_s16_u16(q14u16));
+
+    q7s32 = vmull_s16(d22s16, d22s16);
+    q8s32 = vmull_s16(d24s16, d24s16);
+    q9s32 = vmull_s16(d26s16, d26s16);
+    q10s32 = vmull_s16(d28s16, d28s16);
+
+    q7s32 = vaddq_s32(q7s32, q8s32);
+    q9s32 = vaddq_s32(q9s32, q10s32);
+    q9s32 = vaddq_s32(q7s32, q9s32);
+
+    q1s64 = vpaddlq_s32(q9s32);
+    d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+    return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_avg_neon.c b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_avg_neon.c
new file mode 100644
index 0000000..8632250
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_avg_neon.c
@@ -0,0 +1,373 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+static INLINE int32x4_t MULTIPLY_BY_Q0(
+    int16x4_t dsrc0,
+    int16x4_t dsrc1,
+    int16x4_t dsrc2,
+    int16x4_t dsrc3,
+    int16x4_t dsrc4,
+    int16x4_t dsrc5,
+    int16x4_t dsrc6,
+    int16x4_t dsrc7,
+    int16x8_t q0s16) {
+  int32x4_t qdst;
+  int16x4_t d0s16, d1s16;
+
+  d0s16 = vget_low_s16(q0s16);
+  d1s16 = vget_high_s16(q0s16);
+
+  qdst = vmull_lane_s16(dsrc0, d0s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc1, d0s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc2, d0s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc3, d0s16, 3);
+  qdst = vmlal_lane_s16(qdst, dsrc4, d1s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc5, d1s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc6, d1s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc7, d1s16, 3);
+  return qdst;
+}
+
+void vpx_convolve8_avg_horiz_neon(
+    const uint8_t *src,
+    ptrdiff_t src_stride,
+    uint8_t *dst,
+    ptrdiff_t dst_stride,
+    const int16_t *filter_x,
+    int x_step_q4,
+    const int16_t *filter_y,  // unused
+    int y_step_q4,            // unused
+    int w,
+    int h) {
+  int width;
+  const uint8_t *s;
+  uint8_t *d;
+  uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
+  uint32x2_t d2u32, d3u32, d6u32, d7u32, d28u32, d29u32, d30u32, d31u32;
+  uint8x16_t q1u8, q3u8, q12u8, q13u8, q14u8, q15u8;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+  uint16x8x2_t q0x2u16;
+  uint8x8x2_t d0x2u8, d1x2u8;
+  uint32x2x2_t d0x2u32;
+  uint16x4x2_t d0x2u16, d1x2u16;
+  uint32x4x2_t q0x2u32;
+
+  assert(x_step_q4 == 16);
+
+  q0s16 = vld1q_s16(filter_x);
+
+  src -= 3;  // adjust for taps
+  for (; h > 0; h -= 4) {  // loop_horiz_v
+    s = src;
+    d24u8 = vld1_u8(s);
+    s += src_stride;
+    d25u8 = vld1_u8(s);
+    s += src_stride;
+    d26u8 = vld1_u8(s);
+    s += src_stride;
+    d27u8 = vld1_u8(s);
+
+    q12u8 = vcombine_u8(d24u8, d25u8);
+    q13u8 = vcombine_u8(d26u8, d27u8);
+
+    q0x2u16 = vtrnq_u16(vreinterpretq_u16_u8(q12u8),
+                        vreinterpretq_u16_u8(q13u8));
+    d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
+    d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
+    d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
+    d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
+    d0x2u8 = vtrn_u8(d24u8, d25u8);
+    d1x2u8 = vtrn_u8(d26u8, d27u8);
+
+    __builtin_prefetch(src + src_stride * 4);
+    __builtin_prefetch(src + src_stride * 5);
+
+    q8u16 = vmovl_u8(d0x2u8.val[0]);
+    q9u16 = vmovl_u8(d0x2u8.val[1]);
+    q10u16 = vmovl_u8(d1x2u8.val[0]);
+    q11u16 = vmovl_u8(d1x2u8.val[1]);
+
+    src += 7;
+    d16u16 = vget_low_u16(q8u16);
+    d17u16 = vget_high_u16(q8u16);
+    d18u16 = vget_low_u16(q9u16);
+    d19u16 = vget_high_u16(q9u16);
+    q8u16 = vcombine_u16(d16u16, d18u16);  // vswp 17 18
+    q9u16 = vcombine_u16(d17u16, d19u16);
+
+    d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));  // vmov 23 21
+    for (width = w;
+         width > 0;
+         width -= 4, src += 4, dst += 4) {  // loop_horiz
+      s = src;
+      d28u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d29u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d31u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d30u32 = vld1_dup_u32((const uint32_t *)s);
+
+      __builtin_prefetch(src + 64);
+
+      d0x2u16 = vtrn_u16(vreinterpret_u16_u32(d28u32),
+                         vreinterpret_u16_u32(d31u32));
+      d1x2u16 = vtrn_u16(vreinterpret_u16_u32(d29u32),
+                         vreinterpret_u16_u32(d30u32));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]),   // d28
+                       vreinterpret_u8_u16(d1x2u16.val[0]));  // d29
+      d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]),   // d31
+                       vreinterpret_u8_u16(d1x2u16.val[1]));  // d30
+
+      __builtin_prefetch(src + 64 + src_stride);
+
+      q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+      q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
+      q0x2u32 = vtrnq_u32(vreinterpretq_u32_u8(q14u8),
+                          vreinterpretq_u32_u8(q15u8));
+
+      d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
+      d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
+      q12u16 = vmovl_u8(d28u8);
+      q13u16 = vmovl_u8(d29u8);
+
+      __builtin_prefetch(src + 64 + src_stride * 2);
+
+      d = dst;
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 0);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 0);
+      d += dst_stride;
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 1);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 1);
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+      d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+      d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      q1s32  = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16,
+                              d18s16, d19s16, d23s16, d24s16, q0s16);
+      q2s32  = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16,
+                              d19s16, d23s16, d24s16, d26s16, q0s16);
+      q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16,
+                              d23s16, d24s16, d26s16, d27s16, q0s16);
+      q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16,
+                              d24s16, d26s16, d27s16, d25s16, q0s16);
+
+      __builtin_prefetch(src + 64 + src_stride * 3);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u8 = vqmovn_u16(q1u16);
+      d3u8 = vqmovn_u16(q2u16);
+
+      d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8),
+                         vreinterpret_u16_u8(d3u8));
+      d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
+                         vreinterpret_u32_u16(d0x2u16.val[1]));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
+                       vreinterpret_u8_u32(d0x2u32.val[1]));
+
+      q1u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+      q3u8 = vreinterpretq_u8_u32(vcombine_u32(d6u32, d7u32));
+
+      q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+      d2u32 = vreinterpret_u32_u8(vget_low_u8(q1u8));
+      d3u32 = vreinterpret_u32_u8(vget_high_u8(q1u8));
+
+      d = dst;
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+
+      q8u16 = q9u16;
+      d20s16 = d23s16;
+      q11u16 = q12u16;
+      q9u16 = q13u16;
+      d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    }
+    src += src_stride * 4 - w - 7;
+    dst += dst_stride * 4 - w;
+  }
+  return;
+}
+
+void vpx_convolve8_avg_vert_neon(
+    const uint8_t *src,
+    ptrdiff_t src_stride,
+    uint8_t *dst,
+    ptrdiff_t dst_stride,
+    const int16_t *filter_x,  // unused
+    int x_step_q4,            // unused
+    const int16_t *filter_y,
+    int y_step_q4,
+    int w,
+    int h) {
+  int height;
+  const uint8_t *s;
+  uint8_t *d;
+  uint8x8_t d2u8, d3u8;
+  uint32x2_t d2u32, d3u32, d6u32, d7u32;
+  uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
+  uint8x16_t q1u8, q3u8;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+
+  assert(y_step_q4 == 16);
+
+  src -= src_stride * 3;
+  q0s16 = vld1q_s16(filter_y);
+  for (; w > 0; w -= 4, src += 4, dst += 4) {  // loop_vert_h
+    s = src;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
+    s += src_stride;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
+    s += src_stride;
+    d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
+    s += src_stride;
+
+    q8u16  = vmovl_u8(vreinterpret_u8_u32(d16u32));
+    q9u16  = vmovl_u8(vreinterpret_u8_u32(d18u32));
+    q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
+    q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
+
+    d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+    d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d = dst;
+    for (height = h; height > 0; height -= 4) {  // loop_vert
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
+      s += src_stride;
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
+      s += src_stride;
+
+      q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
+      q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
+
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 0);
+      d += dst_stride;
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 1);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 0);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 1);
+      d -= dst_stride * 3;
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+      d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      __builtin_prefetch(s);
+      __builtin_prefetch(s + src_stride);
+      q1s32  = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16,
+                              d20s16, d21s16, d22s16, d24s16, q0s16);
+      __builtin_prefetch(s + src_stride * 2);
+      __builtin_prefetch(s + src_stride * 3);
+      q2s32  = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16,
+                              d21s16, d22s16, d24s16, d26s16, q0s16);
+      __builtin_prefetch(d);
+      __builtin_prefetch(d + dst_stride);
+      q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16,
+                              d22s16, d24s16, d26s16, d27s16, q0s16);
+      __builtin_prefetch(d + dst_stride * 2);
+      __builtin_prefetch(d + dst_stride * 3);
+      q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16,
+                              d24s16, d26s16, d27s16, d25s16, q0s16);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u8 = vqmovn_u16(q1u16);
+      d3u8 = vqmovn_u16(q2u16);
+
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      q3u8 = vreinterpretq_u8_u32(vcombine_u32(d6u32, d7u32));
+
+      q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+      d2u32 = vreinterpret_u32_u8(vget_low_u8(q1u8));
+      d3u32 = vreinterpret_u32_u8(vget_high_u8(q1u8));
+
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+      d += dst_stride;
+
+      q8u16 = q10u16;
+      d18s16 = d22s16;
+      d19s16 = d24s16;
+      q10u16 = q13u16;
+      d22s16 = d25s16;
+    }
+  }
+  return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_avg_neon_asm.asm b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_avg_neon_asm.asm
new file mode 100644
index 0000000..e279d57
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_avg_neon_asm.asm
@@ -0,0 +1,292 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    ; These functions are only valid when:
+    ; x_step_q4 == 16
+    ; w%4 == 0
+    ; h%4 == 0
+    ; taps == 8
+    ; VP9_FILTER_WEIGHT == 128
+    ; VP9_FILTER_SHIFT == 7
+
+    EXPORT  |vpx_convolve8_avg_horiz_neon|
+    EXPORT  |vpx_convolve8_avg_vert_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Multiply and accumulate by q0
+    MACRO
+    MULTIPLY_BY_Q0 $dst, $src0, $src1, $src2, $src3, $src4, $src5, $src6, $src7
+    vmull.s16 $dst, $src0, d0[0]
+    vmlal.s16 $dst, $src1, d0[1]
+    vmlal.s16 $dst, $src2, d0[2]
+    vmlal.s16 $dst, $src3, d0[3]
+    vmlal.s16 $dst, $src4, d1[0]
+    vmlal.s16 $dst, $src5, d1[1]
+    vmlal.s16 $dst, $src6, d1[2]
+    vmlal.s16 $dst, $src7, d1[3]
+    MEND
+
+; r0    const uint8_t *src
+; r1    int src_stride
+; r2    uint8_t *dst
+; r3    int dst_stride
+; sp[]const int16_t *filter_x
+; sp[]int x_step_q4
+; sp[]const int16_t *filter_y ; unused
+; sp[]int y_step_q4           ; unused
+; sp[]int w
+; sp[]int h
+
+|vpx_convolve8_avg_horiz_neon| PROC
+    push            {r4-r10, lr}
+
+    sub             r0, r0, #3              ; adjust for taps
+
+    ldr             r5, [sp, #32]           ; filter_x
+    ldr             r6, [sp, #48]           ; w
+    ldr             r7, [sp, #52]           ; h
+
+    vld1.s16        {q0}, [r5]              ; filter_x
+
+    sub             r8, r1, r1, lsl #2      ; -src_stride * 3
+    add             r8, r8, #4              ; -src_stride * 3 + 4
+
+    sub             r4, r3, r3, lsl #2      ; -dst_stride * 3
+    add             r4, r4, #4              ; -dst_stride * 3 + 4
+
+    rsb             r9, r6, r1, lsl #2      ; reset src for outer loop
+    sub             r9, r9, #7
+    rsb             r12, r6, r3, lsl #2     ; reset dst for outer loop
+
+    mov             r10, r6                 ; w loop counter
+
+vpx_convolve8_avg_loop_horiz_v
+    vld1.8          {d24}, [r0], r1
+    vld1.8          {d25}, [r0], r1
+    vld1.8          {d26}, [r0], r1
+    vld1.8          {d27}, [r0], r8
+
+    vtrn.16         q12, q13
+    vtrn.8          d24, d25
+    vtrn.8          d26, d27
+
+    pld             [r0, r1, lsl #2]
+
+    vmovl.u8        q8, d24
+    vmovl.u8        q9, d25
+    vmovl.u8        q10, d26
+    vmovl.u8        q11, d27
+
+    ; save a few instructions in the inner loop
+    vswp            d17, d18
+    vmov            d23, d21
+
+    add             r0, r0, #3
+
+vpx_convolve8_avg_loop_horiz
+    add             r5, r0, #64
+
+    vld1.32         {d28[]}, [r0], r1
+    vld1.32         {d29[]}, [r0], r1
+    vld1.32         {d31[]}, [r0], r1
+    vld1.32         {d30[]}, [r0], r8
+
+    pld             [r5]
+
+    vtrn.16         d28, d31
+    vtrn.16         d29, d30
+    vtrn.8          d28, d29
+    vtrn.8          d31, d30
+
+    pld             [r5, r1]
+
+    ; extract to s16
+    vtrn.32         q14, q15
+    vmovl.u8        q12, d28
+    vmovl.u8        q13, d29
+
+    pld             [r5, r1, lsl #1]
+
+    ; slightly out of order load to match the existing data
+    vld1.u32        {d6[0]}, [r2], r3
+    vld1.u32        {d7[0]}, [r2], r3
+    vld1.u32        {d6[1]}, [r2], r3
+    vld1.u32        {d7[1]}, [r2], r3
+
+    sub             r2, r2, r3, lsl #2      ; reset for store
+
+    ; src[] * filter_x
+    MULTIPLY_BY_Q0  q1,  d16, d17, d20, d22, d18, d19, d23, d24
+    MULTIPLY_BY_Q0  q2,  d17, d20, d22, d18, d19, d23, d24, d26
+    MULTIPLY_BY_Q0  q14, d20, d22, d18, d19, d23, d24, d26, d27
+    MULTIPLY_BY_Q0  q15, d22, d18, d19, d23, d24, d26, d27, d25
+
+    pld             [r5, -r8]
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; transpose
+    vtrn.16         d2, d3
+    vtrn.32         d2, d3
+    vtrn.8          d2, d3
+
+    ; average the new value and the dst value
+    vrhadd.u8       q1, q1, q3
+
+    vst1.u32        {d2[0]}, [r2@32], r3
+    vst1.u32        {d3[0]}, [r2@32], r3
+    vst1.u32        {d2[1]}, [r2@32], r3
+    vst1.u32        {d3[1]}, [r2@32], r4
+
+    vmov            q8,  q9
+    vmov            d20, d23
+    vmov            q11, q12
+    vmov            q9,  q13
+
+    subs            r6, r6, #4              ; w -= 4
+    bgt             vpx_convolve8_avg_loop_horiz
+
+    ; outer loop
+    mov             r6, r10                 ; restore w counter
+    add             r0, r0, r9              ; src += src_stride * 4 - w
+    add             r2, r2, r12             ; dst += dst_stride * 4 - w
+    subs            r7, r7, #4              ; h -= 4
+    bgt vpx_convolve8_avg_loop_horiz_v
+
+    pop             {r4-r10, pc}
+
+    ENDP
+
+|vpx_convolve8_avg_vert_neon| PROC
+    push            {r4-r8, lr}
+
+    ; adjust for taps
+    sub             r0, r0, r1
+    sub             r0, r0, r1, lsl #1
+
+    ldr             r4, [sp, #32]           ; filter_y
+    ldr             r6, [sp, #40]           ; w
+    ldr             lr, [sp, #44]           ; h
+
+    vld1.s16        {q0}, [r4]              ; filter_y
+
+    lsl             r1, r1, #1
+    lsl             r3, r3, #1
+
+vpx_convolve8_avg_loop_vert_h
+    mov             r4, r0
+    add             r7, r0, r1, asr #1
+    mov             r5, r2
+    add             r8, r2, r3, asr #1
+    mov             r12, lr                 ; h loop counter
+
+    vld1.u32        {d16[0]}, [r4], r1
+    vld1.u32        {d16[1]}, [r7], r1
+    vld1.u32        {d18[0]}, [r4], r1
+    vld1.u32        {d18[1]}, [r7], r1
+    vld1.u32        {d20[0]}, [r4], r1
+    vld1.u32        {d20[1]}, [r7], r1
+    vld1.u32        {d22[0]}, [r4], r1
+
+    vmovl.u8        q8, d16
+    vmovl.u8        q9, d18
+    vmovl.u8        q10, d20
+    vmovl.u8        q11, d22
+
+vpx_convolve8_avg_loop_vert
+    ; always process a 4x4 block at a time
+    vld1.u32        {d24[0]}, [r7], r1
+    vld1.u32        {d26[0]}, [r4], r1
+    vld1.u32        {d26[1]}, [r7], r1
+    vld1.u32        {d24[1]}, [r4], r1
+
+    ; extract to s16
+    vmovl.u8        q12, d24
+    vmovl.u8        q13, d26
+
+    vld1.u32        {d6[0]}, [r5@32], r3
+    vld1.u32        {d6[1]}, [r8@32], r3
+    vld1.u32        {d7[0]}, [r5@32], r3
+    vld1.u32        {d7[1]}, [r8@32], r3
+
+    pld             [r7]
+    pld             [r4]
+
+    ; src[] * filter_y
+    MULTIPLY_BY_Q0  q1,  d16, d17, d18, d19, d20, d21, d22, d24
+
+    pld             [r7, r1]
+    pld             [r4, r1]
+
+    MULTIPLY_BY_Q0  q2,  d17, d18, d19, d20, d21, d22, d24, d26
+
+    pld             [r5]
+    pld             [r8]
+
+    MULTIPLY_BY_Q0  q14, d18, d19, d20, d21, d22, d24, d26, d27
+
+    pld             [r5, r3]
+    pld             [r8, r3]
+
+    MULTIPLY_BY_Q0  q15, d19, d20, d21, d22, d24, d26, d27, d25
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; average the new value and the dst value
+    vrhadd.u8       q1, q1, q3
+
+    sub             r5, r5, r3, lsl #1      ; reset for store
+    sub             r8, r8, r3, lsl #1
+
+    vst1.u32        {d2[0]}, [r5@32], r3
+    vst1.u32        {d2[1]}, [r8@32], r3
+    vst1.u32        {d3[0]}, [r5@32], r3
+    vst1.u32        {d3[1]}, [r8@32], r3
+
+    vmov            q8, q10
+    vmov            d18, d22
+    vmov            d19, d24
+    vmov            q10, q13
+    vmov            d22, d25
+
+    subs            r12, r12, #4            ; h -= 4
+    bgt             vpx_convolve8_avg_loop_vert
+
+    ; outer loop
+    add             r0, r0, #4
+    add             r2, r2, #4
+    subs            r6, r6, #4              ; w -= 4
+    bgt             vpx_convolve8_avg_loop_vert_h
+
+    pop             {r4-r8, pc}
+
+    ENDP
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon.c b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon.c
new file mode 100644
index 0000000..9bd715e
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon.c
@@ -0,0 +1,340 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+static INLINE int32x4_t MULTIPLY_BY_Q0(
+    int16x4_t dsrc0,
+    int16x4_t dsrc1,
+    int16x4_t dsrc2,
+    int16x4_t dsrc3,
+    int16x4_t dsrc4,
+    int16x4_t dsrc5,
+    int16x4_t dsrc6,
+    int16x4_t dsrc7,
+    int16x8_t q0s16) {
+  int32x4_t qdst;
+  int16x4_t d0s16, d1s16;
+
+  d0s16 = vget_low_s16(q0s16);
+  d1s16 = vget_high_s16(q0s16);
+
+  qdst = vmull_lane_s16(dsrc0, d0s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc1, d0s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc2, d0s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc3, d0s16, 3);
+  qdst = vmlal_lane_s16(qdst, dsrc4, d1s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc5, d1s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc6, d1s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc7, d1s16, 3);
+  return qdst;
+}
+
+void vpx_convolve8_horiz_neon(
+    const uint8_t *src,
+    ptrdiff_t src_stride,
+    uint8_t *dst,
+    ptrdiff_t dst_stride,
+    const int16_t *filter_x,
+    int x_step_q4,
+    const int16_t *filter_y,  // unused
+    int y_step_q4,            // unused
+    int w,
+    int h) {
+  int width;
+  const uint8_t *s, *psrc;
+  uint8_t *d, *pdst;
+  uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
+  uint32x2_t d2u32, d3u32, d28u32, d29u32, d30u32, d31u32;
+  uint8x16_t q12u8, q13u8, q14u8, q15u8;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+  uint16x8x2_t q0x2u16;
+  uint8x8x2_t d0x2u8, d1x2u8;
+  uint32x2x2_t d0x2u32;
+  uint16x4x2_t d0x2u16, d1x2u16;
+  uint32x4x2_t q0x2u32;
+
+  assert(x_step_q4 == 16);
+
+  q0s16 = vld1q_s16(filter_x);
+
+  src -= 3;  // adjust for taps
+  for (; h > 0; h -= 4,
+    src += src_stride * 4,
+    dst += dst_stride * 4) {  // loop_horiz_v
+    s = src;
+    d24u8 = vld1_u8(s);
+    s += src_stride;
+    d25u8 = vld1_u8(s);
+    s += src_stride;
+    d26u8 = vld1_u8(s);
+    s += src_stride;
+    d27u8 = vld1_u8(s);
+
+    q12u8 = vcombine_u8(d24u8, d25u8);
+    q13u8 = vcombine_u8(d26u8, d27u8);
+
+    q0x2u16 = vtrnq_u16(vreinterpretq_u16_u8(q12u8),
+                        vreinterpretq_u16_u8(q13u8));
+    d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
+    d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
+    d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
+    d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
+    d0x2u8 = vtrn_u8(d24u8, d25u8);
+    d1x2u8 = vtrn_u8(d26u8, d27u8);
+
+    __builtin_prefetch(src + src_stride * 4);
+    __builtin_prefetch(src + src_stride * 5);
+    __builtin_prefetch(src + src_stride * 6);
+
+    q8u16  = vmovl_u8(d0x2u8.val[0]);
+    q9u16  = vmovl_u8(d0x2u8.val[1]);
+    q10u16 = vmovl_u8(d1x2u8.val[0]);
+    q11u16 = vmovl_u8(d1x2u8.val[1]);
+
+    d16u16 = vget_low_u16(q8u16);
+    d17u16 = vget_high_u16(q8u16);
+    d18u16 = vget_low_u16(q9u16);
+    d19u16 = vget_high_u16(q9u16);
+    q8u16 = vcombine_u16(d16u16, d18u16);  // vswp 17 18
+    q9u16 = vcombine_u16(d17u16, d19u16);
+
+    d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));  // vmov 23 21
+    for (width = w, psrc = src + 7, pdst = dst;
+         width > 0;
+         width -= 4, psrc += 4, pdst += 4) {  // loop_horiz
+      s = psrc;
+      d28u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d29u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d31u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d30u32 = vld1_dup_u32((const uint32_t *)s);
+
+      __builtin_prefetch(psrc + 64);
+
+      d0x2u16 = vtrn_u16(vreinterpret_u16_u32(d28u32),
+                         vreinterpret_u16_u32(d31u32));
+      d1x2u16 = vtrn_u16(vreinterpret_u16_u32(d29u32),
+                         vreinterpret_u16_u32(d30u32));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]),   // d28
+                       vreinterpret_u8_u16(d1x2u16.val[0]));  // d29
+      d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]),   // d31
+                       vreinterpret_u8_u16(d1x2u16.val[1]));  // d30
+
+      __builtin_prefetch(psrc + 64 + src_stride);
+
+      q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+      q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
+      q0x2u32 = vtrnq_u32(vreinterpretq_u32_u8(q14u8),
+                          vreinterpretq_u32_u8(q15u8));
+
+      d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
+      d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
+      q12u16 = vmovl_u8(d28u8);
+      q13u16 = vmovl_u8(d29u8);
+
+      __builtin_prefetch(psrc + 64 + src_stride * 2);
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+      d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+      d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      q1s32  = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16,
+                              d18s16, d19s16, d23s16, d24s16, q0s16);
+      q2s32  = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16,
+                              d19s16, d23s16, d24s16, d26s16, q0s16);
+      q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16,
+                              d23s16, d24s16, d26s16, d27s16, q0s16);
+      q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16,
+                              d24s16, d26s16, d27s16, d25s16, q0s16);
+
+      __builtin_prefetch(psrc + 60 + src_stride * 3);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u8 = vqmovn_u16(q1u16);
+      d3u8 = vqmovn_u16(q2u16);
+
+      d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8),
+                         vreinterpret_u16_u8(d3u8));
+      d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
+                         vreinterpret_u32_u16(d0x2u16.val[1]));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
+                       vreinterpret_u8_u32(d0x2u32.val[1]));
+
+      d2u32 = vreinterpret_u32_u8(d0x2u8.val[0]);
+      d3u32 = vreinterpret_u32_u8(d0x2u8.val[1]);
+
+      d = pdst;
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+
+      q8u16 = q9u16;
+      d20s16 = d23s16;
+      q11u16 = q12u16;
+      q9u16 = q13u16;
+      d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    }
+  }
+  return;
+}
+
+void vpx_convolve8_vert_neon(
+    const uint8_t *src,
+    ptrdiff_t src_stride,
+    uint8_t *dst,
+    ptrdiff_t dst_stride,
+    const int16_t *filter_x,  // unused
+    int x_step_q4,            // unused
+    const int16_t *filter_y,
+    int y_step_q4,
+    int w,
+    int h) {
+  int height;
+  const uint8_t *s;
+  uint8_t *d;
+  uint32x2_t d2u32, d3u32;
+  uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+
+  assert(y_step_q4 == 16);
+
+  src -= src_stride * 3;
+  q0s16 = vld1q_s16(filter_y);
+  for (; w > 0; w -= 4, src += 4, dst += 4) {  // loop_vert_h
+    s = src;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
+    s += src_stride;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
+    s += src_stride;
+    d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
+    s += src_stride;
+
+    q8u16  = vmovl_u8(vreinterpret_u8_u32(d16u32));
+    q9u16  = vmovl_u8(vreinterpret_u8_u32(d18u32));
+    q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
+    q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
+
+    d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+    d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d = dst;
+    for (height = h; height > 0; height -= 4) {  // loop_vert
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
+      s += src_stride;
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
+      s += src_stride;
+
+      q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
+      q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+      d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      __builtin_prefetch(d);
+      __builtin_prefetch(d + dst_stride);
+      q1s32  = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16,
+                              d20s16, d21s16, d22s16, d24s16, q0s16);
+      __builtin_prefetch(d + dst_stride * 2);
+      __builtin_prefetch(d + dst_stride * 3);
+      q2s32  = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16,
+                              d21s16, d22s16, d24s16, d26s16, q0s16);
+      __builtin_prefetch(s);
+      __builtin_prefetch(s + src_stride);
+      q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16,
+                              d22s16, d24s16, d26s16, d27s16, q0s16);
+      __builtin_prefetch(s + src_stride * 2);
+      __builtin_prefetch(s + src_stride * 3);
+      q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16,
+                              d24s16, d26s16, d27s16, d25s16, q0s16);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u32 = vreinterpret_u32_u8(vqmovn_u16(q1u16));
+      d3u32 = vreinterpret_u32_u8(vqmovn_u16(q2u16));
+
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+      d += dst_stride;
+
+      q8u16 = q10u16;
+      d18s16 = d22s16;
+      d19s16 = d24s16;
+      q10u16 = q13u16;
+      d22s16 = d25s16;
+    }
+  }
+  return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon_asm.asm b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon_asm.asm
new file mode 100644
index 0000000..2d0f2ae
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon_asm.asm
@@ -0,0 +1,270 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    ; These functions are only valid when:
+    ; x_step_q4 == 16
+    ; w%4 == 0
+    ; h%4 == 0
+    ; taps == 8
+    ; VP9_FILTER_WEIGHT == 128
+    ; VP9_FILTER_SHIFT == 7
+
+    EXPORT  |vpx_convolve8_horiz_neon|
+    EXPORT  |vpx_convolve8_vert_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Multiply and accumulate by q0
+    MACRO
+    MULTIPLY_BY_Q0 $dst, $src0, $src1, $src2, $src3, $src4, $src5, $src6, $src7
+    vmull.s16 $dst, $src0, d0[0]
+    vmlal.s16 $dst, $src1, d0[1]
+    vmlal.s16 $dst, $src2, d0[2]
+    vmlal.s16 $dst, $src3, d0[3]
+    vmlal.s16 $dst, $src4, d1[0]
+    vmlal.s16 $dst, $src5, d1[1]
+    vmlal.s16 $dst, $src6, d1[2]
+    vmlal.s16 $dst, $src7, d1[3]
+    MEND
+
+; r0    const uint8_t *src
+; r1    int src_stride
+; r2    uint8_t *dst
+; r3    int dst_stride
+; sp[]const int16_t *filter_x
+; sp[]int x_step_q4
+; sp[]const int16_t *filter_y ; unused
+; sp[]int y_step_q4           ; unused
+; sp[]int w
+; sp[]int h
+
+|vpx_convolve8_horiz_neon| PROC
+    push            {r4-r10, lr}
+
+    sub             r0, r0, #3              ; adjust for taps
+
+    ldr             r5, [sp, #32]           ; filter_x
+    ldr             r6, [sp, #48]           ; w
+    ldr             r7, [sp, #52]           ; h
+
+    vld1.s16        {q0}, [r5]              ; filter_x
+
+    sub             r8, r1, r1, lsl #2      ; -src_stride * 3
+    add             r8, r8, #4              ; -src_stride * 3 + 4
+
+    sub             r4, r3, r3, lsl #2      ; -dst_stride * 3
+    add             r4, r4, #4              ; -dst_stride * 3 + 4
+
+    rsb             r9, r6, r1, lsl #2      ; reset src for outer loop
+    sub             r9, r9, #7
+    rsb             r12, r6, r3, lsl #2     ; reset dst for outer loop
+
+    mov             r10, r6                 ; w loop counter
+
+vpx_convolve8_loop_horiz_v
+    vld1.8          {d24}, [r0], r1
+    vld1.8          {d25}, [r0], r1
+    vld1.8          {d26}, [r0], r1
+    vld1.8          {d27}, [r0], r8
+
+    vtrn.16         q12, q13
+    vtrn.8          d24, d25
+    vtrn.8          d26, d27
+
+    pld             [r0, r1, lsl #2]
+
+    vmovl.u8        q8, d24
+    vmovl.u8        q9, d25
+    vmovl.u8        q10, d26
+    vmovl.u8        q11, d27
+
+    ; save a few instructions in the inner loop
+    vswp            d17, d18
+    vmov            d23, d21
+
+    add             r0, r0, #3
+
+vpx_convolve8_loop_horiz
+    add             r5, r0, #64
+
+    vld1.32         {d28[]}, [r0], r1
+    vld1.32         {d29[]}, [r0], r1
+    vld1.32         {d31[]}, [r0], r1
+    vld1.32         {d30[]}, [r0], r8
+
+    pld             [r5]
+
+    vtrn.16         d28, d31
+    vtrn.16         d29, d30
+    vtrn.8          d28, d29
+    vtrn.8          d31, d30
+
+    pld             [r5, r1]
+
+    ; extract to s16
+    vtrn.32         q14, q15
+    vmovl.u8        q12, d28
+    vmovl.u8        q13, d29
+
+    pld             [r5, r1, lsl #1]
+
+    ; src[] * filter_x
+    MULTIPLY_BY_Q0  q1,  d16, d17, d20, d22, d18, d19, d23, d24
+    MULTIPLY_BY_Q0  q2,  d17, d20, d22, d18, d19, d23, d24, d26
+    MULTIPLY_BY_Q0  q14, d20, d22, d18, d19, d23, d24, d26, d27
+    MULTIPLY_BY_Q0  q15, d22, d18, d19, d23, d24, d26, d27, d25
+
+    pld             [r5, -r8]
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; transpose
+    vtrn.16         d2, d3
+    vtrn.32         d2, d3
+    vtrn.8          d2, d3
+
+    vst1.u32        {d2[0]}, [r2@32], r3
+    vst1.u32        {d3[0]}, [r2@32], r3
+    vst1.u32        {d2[1]}, [r2@32], r3
+    vst1.u32        {d3[1]}, [r2@32], r4
+
+    vmov            q8,  q9
+    vmov            d20, d23
+    vmov            q11, q12
+    vmov            q9,  q13
+
+    subs            r6, r6, #4              ; w -= 4
+    bgt             vpx_convolve8_loop_horiz
+
+    ; outer loop
+    mov             r6, r10                 ; restore w counter
+    add             r0, r0, r9              ; src += src_stride * 4 - w
+    add             r2, r2, r12             ; dst += dst_stride * 4 - w
+    subs            r7, r7, #4              ; h -= 4
+    bgt vpx_convolve8_loop_horiz_v
+
+    pop             {r4-r10, pc}
+
+    ENDP
+
+|vpx_convolve8_vert_neon| PROC
+    push            {r4-r8, lr}
+
+    ; adjust for taps
+    sub             r0, r0, r1
+    sub             r0, r0, r1, lsl #1
+
+    ldr             r4, [sp, #32]           ; filter_y
+    ldr             r6, [sp, #40]           ; w
+    ldr             lr, [sp, #44]           ; h
+
+    vld1.s16        {q0}, [r4]              ; filter_y
+
+    lsl             r1, r1, #1
+    lsl             r3, r3, #1
+
+vpx_convolve8_loop_vert_h
+    mov             r4, r0
+    add             r7, r0, r1, asr #1
+    mov             r5, r2
+    add             r8, r2, r3, asr #1
+    mov             r12, lr                 ; h loop counter
+
+    vld1.u32        {d16[0]}, [r4], r1
+    vld1.u32        {d16[1]}, [r7], r1
+    vld1.u32        {d18[0]}, [r4], r1
+    vld1.u32        {d18[1]}, [r7], r1
+    vld1.u32        {d20[0]}, [r4], r1
+    vld1.u32        {d20[1]}, [r7], r1
+    vld1.u32        {d22[0]}, [r4], r1
+
+    vmovl.u8        q8, d16
+    vmovl.u8        q9, d18
+    vmovl.u8        q10, d20
+    vmovl.u8        q11, d22
+
+vpx_convolve8_loop_vert
+    ; always process a 4x4 block at a time
+    vld1.u32        {d24[0]}, [r7], r1
+    vld1.u32        {d26[0]}, [r4], r1
+    vld1.u32        {d26[1]}, [r7], r1
+    vld1.u32        {d24[1]}, [r4], r1
+
+    ; extract to s16
+    vmovl.u8        q12, d24
+    vmovl.u8        q13, d26
+
+    pld             [r5]
+    pld             [r8]
+
+    ; src[] * filter_y
+    MULTIPLY_BY_Q0  q1,  d16, d17, d18, d19, d20, d21, d22, d24
+
+    pld             [r5, r3]
+    pld             [r8, r3]
+
+    MULTIPLY_BY_Q0  q2,  d17, d18, d19, d20, d21, d22, d24, d26
+
+    pld             [r7]
+    pld             [r4]
+
+    MULTIPLY_BY_Q0  q14, d18, d19, d20, d21, d22, d24, d26, d27
+
+    pld             [r7, r1]
+    pld             [r4, r1]
+
+    MULTIPLY_BY_Q0  q15, d19, d20, d21, d22, d24, d26, d27, d25
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    vst1.u32        {d2[0]}, [r5@32], r3
+    vst1.u32        {d2[1]}, [r8@32], r3
+    vst1.u32        {d3[0]}, [r5@32], r3
+    vst1.u32        {d3[1]}, [r8@32], r3
+
+    vmov            q8, q10
+    vmov            d18, d22
+    vmov            d19, d24
+    vmov            q10, q13
+    vmov            d22, d25
+
+    subs            r12, r12, #4            ; h -= 4
+    bgt             vpx_convolve8_loop_vert
+
+    ; outer loop
+    add             r0, r0, #4
+    add             r2, r2, #4
+    subs            r6, r6, #4              ; w -= 4
+    bgt             vpx_convolve8_loop_vert_h
+
+    pop             {r4-r8, pc}
+
+    ENDP
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_avg_neon.c b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_avg_neon.c
new file mode 100644
index 0000000..dc58a33
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_avg_neon.c
@@ -0,0 +1,147 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+void vpx_convolve_avg_neon(
+    const uint8_t *src,    // r0
+    ptrdiff_t src_stride,  // r1
+    uint8_t *dst,          // r2
+    ptrdiff_t dst_stride,  // r3
+    const int16_t *filter_x,
+    int filter_x_stride,
+    const int16_t *filter_y,
+    int filter_y_stride,
+    int w,
+    int h) {
+  uint8_t *d;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8;
+  uint32x2_t d0u32, d2u32;
+  uint8x16_t q0u8, q1u8, q2u8, q3u8, q8u8, q9u8, q10u8, q11u8;
+  (void)filter_x;  (void)filter_x_stride;
+  (void)filter_y;  (void)filter_y_stride;
+
+  d = dst;
+  if (w > 32) {  // avg64
+    for (; h > 0; h -= 1) {
+      q0u8  = vld1q_u8(src);
+      q1u8  = vld1q_u8(src + 16);
+      q2u8  = vld1q_u8(src + 32);
+      q3u8  = vld1q_u8(src + 48);
+      src += src_stride;
+      q8u8  = vld1q_u8(d);
+      q9u8  = vld1q_u8(d + 16);
+      q10u8 = vld1q_u8(d + 32);
+      q11u8 = vld1q_u8(d + 48);
+      d += dst_stride;
+
+      q0u8 = vrhaddq_u8(q0u8, q8u8);
+      q1u8 = vrhaddq_u8(q1u8, q9u8);
+      q2u8 = vrhaddq_u8(q2u8, q10u8);
+      q3u8 = vrhaddq_u8(q3u8, q11u8);
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      vst1q_u8(dst + 32, q2u8);
+      vst1q_u8(dst + 48, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w == 32) {  // avg32
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      src += src_stride;
+      q2u8 = vld1q_u8(src);
+      q3u8 = vld1q_u8(src + 16);
+      src += src_stride;
+      q8u8 = vld1q_u8(d);
+      q9u8 = vld1q_u8(d + 16);
+      d += dst_stride;
+      q10u8 = vld1q_u8(d);
+      q11u8 = vld1q_u8(d + 16);
+      d += dst_stride;
+
+      q0u8 = vrhaddq_u8(q0u8, q8u8);
+      q1u8 = vrhaddq_u8(q1u8, q9u8);
+      q2u8 = vrhaddq_u8(q2u8, q10u8);
+      q3u8 = vrhaddq_u8(q3u8, q11u8);
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q2u8);
+      vst1q_u8(dst + 16, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w > 8) {  // avg16
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      src += src_stride;
+      q1u8 = vld1q_u8(src);
+      src += src_stride;
+      q2u8 = vld1q_u8(d);
+      d += dst_stride;
+      q3u8 = vld1q_u8(d);
+      d += dst_stride;
+
+      q0u8 = vrhaddq_u8(q0u8, q2u8);
+      q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+      vst1q_u8(dst, q0u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q1u8);
+      dst += dst_stride;
+    }
+  } else if (w == 8) {  // avg8
+    for (; h > 0; h -= 2) {
+      d0u8 = vld1_u8(src);
+      src += src_stride;
+      d1u8 = vld1_u8(src);
+      src += src_stride;
+      d2u8 = vld1_u8(d);
+      d += dst_stride;
+      d3u8 = vld1_u8(d);
+      d += dst_stride;
+
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      q0u8 = vrhaddq_u8(q0u8, q1u8);
+
+      vst1_u8(dst, vget_low_u8(q0u8));
+      dst += dst_stride;
+      vst1_u8(dst, vget_high_u8(q0u8));
+      dst += dst_stride;
+    }
+  } else {  // avg4
+    for (; h > 0; h -= 2) {
+      d0u32 = vld1_lane_u32((const uint32_t *)src, d0u32, 0);
+      src += src_stride;
+      d0u32 = vld1_lane_u32((const uint32_t *)src, d0u32, 1);
+      src += src_stride;
+      d2u32 = vld1_lane_u32((const uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      d2u32 = vld1_lane_u32((const uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+
+      d0u8 = vrhadd_u8(vreinterpret_u8_u32(d0u32),
+                       vreinterpret_u8_u32(d2u32));
+
+      d0u32 = vreinterpret_u32_u8(d0u8);
+      vst1_lane_u32((uint32_t *)dst, d0u32, 0);
+      dst += dst_stride;
+      vst1_lane_u32((uint32_t *)dst, d0u32, 1);
+      dst += dst_stride;
+    }
+  }
+  return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_avg_neon_asm.asm b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_avg_neon_asm.asm
new file mode 100644
index 0000000..97e6189
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_avg_neon_asm.asm
@@ -0,0 +1,116 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vpx_convolve_avg_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|vpx_convolve_avg_neon| PROC
+    push                {r4-r6, lr}
+    ldrd                r4, r5, [sp, #32]
+    mov                 r6, r2
+
+    cmp                 r4, #32
+    bgt                 avg64
+    beq                 avg32
+    cmp                 r4, #8
+    bgt                 avg16
+    beq                 avg8
+    b                   avg4
+
+avg64
+    sub                 lr, r1, #32
+    sub                 r4, r3, #32
+avg64_h
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0]!
+    vld1.8              {q2-q3}, [r0], lr
+    pld                 [r2, r3]
+    vld1.8              {q8-q9},   [r6@128]!
+    vld1.8              {q10-q11}, [r6@128], r4
+    vrhadd.u8           q0, q0, q8
+    vrhadd.u8           q1, q1, q9
+    vrhadd.u8           q2, q2, q10
+    vrhadd.u8           q3, q3, q11
+    vst1.8              {q0-q1}, [r2@128]!
+    vst1.8              {q2-q3}, [r2@128], r4
+    subs                r5, r5, #1
+    bgt                 avg64_h
+    pop                 {r4-r6, pc}
+
+avg32
+    vld1.8              {q0-q1}, [r0], r1
+    vld1.8              {q2-q3}, [r0], r1
+    vld1.8              {q8-q9},   [r6@128], r3
+    vld1.8              {q10-q11}, [r6@128], r3
+    pld                 [r0]
+    vrhadd.u8           q0, q0, q8
+    pld                 [r0, r1]
+    vrhadd.u8           q1, q1, q9
+    pld                 [r6]
+    vrhadd.u8           q2, q2, q10
+    pld                 [r6, r3]
+    vrhadd.u8           q3, q3, q11
+    vst1.8              {q0-q1}, [r2@128], r3
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 avg32
+    pop                 {r4-r6, pc}
+
+avg16
+    vld1.8              {q0}, [r0], r1
+    vld1.8              {q1}, [r0], r1
+    vld1.8              {q2}, [r6@128], r3
+    vld1.8              {q3}, [r6@128], r3
+    pld                 [r0]
+    pld                 [r0, r1]
+    vrhadd.u8           q0, q0, q2
+    pld                 [r6]
+    pld                 [r6, r3]
+    vrhadd.u8           q1, q1, q3
+    vst1.8              {q0}, [r2@128], r3
+    vst1.8              {q1}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 avg16
+    pop                 {r4-r6, pc}
+
+avg8
+    vld1.8              {d0}, [r0], r1
+    vld1.8              {d1}, [r0], r1
+    vld1.8              {d2}, [r6@64], r3
+    vld1.8              {d3}, [r6@64], r3
+    pld                 [r0]
+    pld                 [r0, r1]
+    vrhadd.u8           q0, q0, q1
+    pld                 [r6]
+    pld                 [r6, r3]
+    vst1.8              {d0}, [r2@64], r3
+    vst1.8              {d1}, [r2@64], r3
+    subs                r5, r5, #2
+    bgt                 avg8
+    pop                 {r4-r6, pc}
+
+avg4
+    vld1.32             {d0[0]}, [r0], r1
+    vld1.32             {d0[1]}, [r0], r1
+    vld1.32             {d2[0]}, [r6@32], r3
+    vld1.32             {d2[1]}, [r6@32], r3
+    vrhadd.u8           d0, d0, d2
+    vst1.32             {d0[0]}, [r2@32], r3
+    vst1.32             {d0[1]}, [r2@32], r3
+    subs                r5, r5, #2
+    bgt                 avg4
+    pop                 {r4-r6, pc}
+    ENDP
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_copy_neon.c b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_copy_neon.c
new file mode 100644
index 0000000..d8fb97a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_copy_neon.c
@@ -0,0 +1,94 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+void vpx_convolve_copy_neon(
+    const uint8_t *src,    // r0
+    ptrdiff_t src_stride,  // r1
+    uint8_t *dst,          // r2
+    ptrdiff_t dst_stride,  // r3
+    const int16_t *filter_x,
+    int filter_x_stride,
+    const int16_t *filter_y,
+    int filter_y_stride,
+    int w,
+    int h) {
+  uint8x8_t d0u8, d2u8;
+  uint8x16_t q0u8, q1u8, q2u8, q3u8;
+  (void)filter_x;  (void)filter_x_stride;
+  (void)filter_y;  (void)filter_y_stride;
+
+  if (w > 32) {  // copy64
+    for (; h > 0; h--) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      q2u8 = vld1q_u8(src + 32);
+      q3u8 = vld1q_u8(src + 48);
+      src += src_stride;
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      vst1q_u8(dst + 32, q2u8);
+      vst1q_u8(dst + 48, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w == 32) {  // copy32
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      src += src_stride;
+      q2u8 = vld1q_u8(src);
+      q3u8 = vld1q_u8(src + 16);
+      src += src_stride;
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q2u8);
+      vst1q_u8(dst + 16, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w > 8) {  // copy16
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      src += src_stride;
+      q1u8 = vld1q_u8(src);
+      src += src_stride;
+
+      vst1q_u8(dst, q0u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q1u8);
+      dst += dst_stride;
+    }
+  } else if (w == 8) {  // copy8
+    for (; h > 0; h -= 2) {
+      d0u8 = vld1_u8(src);
+      src += src_stride;
+      d2u8 = vld1_u8(src);
+      src += src_stride;
+
+      vst1_u8(dst, d0u8);
+      dst += dst_stride;
+      vst1_u8(dst, d2u8);
+      dst += dst_stride;
+    }
+  } else {  // copy4
+    for (; h > 0; h--) {
+      *(uint32_t *)dst = *(const uint32_t *)src;
+      src += src_stride;
+      dst += dst_stride;
+    }
+  }
+  return;
+}
diff --git a/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_copy_neon_asm.asm b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_copy_neon_asm.asm
new file mode 100644
index 0000000..89164ad
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_copy_neon_asm.asm
@@ -0,0 +1,84 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vpx_convolve_copy_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|vpx_convolve_copy_neon| PROC
+    push                {r4-r5, lr}
+    ldrd                r4, r5, [sp, #28]
+
+    cmp                 r4, #32
+    bgt                 copy64
+    beq                 copy32
+    cmp                 r4, #8
+    bgt                 copy16
+    beq                 copy8
+    b                   copy4
+
+copy64
+    sub                 lr, r1, #32
+    sub                 r3, r3, #32
+copy64_h
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0]!
+    vld1.8              {q2-q3}, [r0], lr
+    vst1.8              {q0-q1}, [r2@128]!
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #1
+    bgt                 copy64_h
+    pop                 {r4-r5, pc}
+
+copy32
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q2-q3}, [r0], r1
+    vst1.8              {q0-q1}, [r2@128], r3
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 copy32
+    pop                 {r4-r5, pc}
+
+copy16
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q1}, [r0], r1
+    vst1.8              {q0}, [r2@128], r3
+    vst1.8              {q1}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 copy16
+    pop                 {r4-r5, pc}
+
+copy8
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {d0}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {d2}, [r0], r1
+    vst1.8              {d0}, [r2@64], r3
+    vst1.8              {d2}, [r2@64], r3
+    subs                r5, r5, #2
+    bgt                 copy8
+    pop                 {r4-r5, pc}
+
+copy4
+    ldr                 r12, [r0], r1
+    str                 r12, [r2], r3
+    subs                r5, r5, #1
+    bgt                 copy4
+    pop                 {r4-r5, pc}
+    ENDP
+
+    END
diff --git a/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_neon.c b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_neon.c
new file mode 100644
index 0000000..1506ce6
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/arm/vpx_convolve_neon.c
@@ -0,0 +1,72 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+
+void vpx_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride,
+                        uint8_t *dst, ptrdiff_t dst_stride,
+                        const int16_t *filter_x, int x_step_q4,
+                        const int16_t *filter_y, int y_step_q4,
+                        int w, int h) {
+  /* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the
+   * maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4).
+   */
+  DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
+
+  // Account for the vertical phase needing 3 lines prior and 4 lines post
+  int intermediate_height = h + 7;
+
+  assert(y_step_q4 == 16);
+  assert(x_step_q4 == 16);
+
+  /* Filter starting 3 lines back. The neon implementation will ignore the
+   * given height and filter a multiple of 4 lines. Since this goes in to
+   * the temp buffer which has lots of extra room and is subsequently discarded
+   * this is safe if somewhat less than ideal.
+   */
+  vpx_convolve8_horiz_neon(src - src_stride * 3, src_stride,
+                           temp, 64,
+                           filter_x, x_step_q4, filter_y, y_step_q4,
+                           w, intermediate_height);
+
+  /* Step into the temp buffer 3 lines to get the actual frame data */
+  vpx_convolve8_vert_neon(temp + 64 * 3, 64,
+                          dst, dst_stride,
+                          filter_x, x_step_q4, filter_y, y_step_q4,
+                          w, h);
+}
+
+void vpx_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4,
+                            int w, int h) {
+  DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
+  int intermediate_height = h + 7;
+
+  assert(y_step_q4 == 16);
+  assert(x_step_q4 == 16);
+
+  /* This implementation has the same issues as above. In addition, we only want
+   * to average the values after both passes.
+   */
+  vpx_convolve8_horiz_neon(src - src_stride * 3, src_stride,
+                           temp, 64,
+                           filter_x, x_step_q4, filter_y, y_step_q4,
+                           w, intermediate_height);
+  vpx_convolve8_avg_vert_neon(temp + 64 * 3,
+                              64, dst, dst_stride,
+                              filter_x, x_step_q4, filter_y, y_step_q4,
+                              w, h);
+}
diff --git a/libvpx/libvpx/vpx_dsp/avg.c b/libvpx/libvpx/vpx_dsp/avg.c
new file mode 100644
index 0000000..a8c9966
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/avg.c
@@ -0,0 +1,233 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <stdlib.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+
+unsigned int vpx_avg_8x8_c(const uint8_t *s, int p) {
+  int i, j;
+  int sum = 0;
+  for (i = 0; i < 8; ++i, s+=p)
+    for (j = 0; j < 8; sum += s[j], ++j) {}
+
+  return (sum + 32) >> 6;
+}
+
+unsigned int vpx_avg_4x4_c(const uint8_t *s, int p) {
+  int i, j;
+  int sum = 0;
+  for (i = 0; i < 4; ++i, s+=p)
+    for (j = 0; j < 4; sum += s[j], ++j) {}
+
+  return (sum + 8) >> 4;
+}
+
+// src_diff: first pass, 9 bit, dynamic range [-255, 255]
+//           second pass, 12 bit, dynamic range [-2040, 2040]
+static void hadamard_col8(const int16_t *src_diff, int src_stride,
+                          int16_t *coeff) {
+  int16_t b0 = src_diff[0 * src_stride] + src_diff[1 * src_stride];
+  int16_t b1 = src_diff[0 * src_stride] - src_diff[1 * src_stride];
+  int16_t b2 = src_diff[2 * src_stride] + src_diff[3 * src_stride];
+  int16_t b3 = src_diff[2 * src_stride] - src_diff[3 * src_stride];
+  int16_t b4 = src_diff[4 * src_stride] + src_diff[5 * src_stride];
+  int16_t b5 = src_diff[4 * src_stride] - src_diff[5 * src_stride];
+  int16_t b6 = src_diff[6 * src_stride] + src_diff[7 * src_stride];
+  int16_t b7 = src_diff[6 * src_stride] - src_diff[7 * src_stride];
+
+  int16_t c0 = b0 + b2;
+  int16_t c1 = b1 + b3;
+  int16_t c2 = b0 - b2;
+  int16_t c3 = b1 - b3;
+  int16_t c4 = b4 + b6;
+  int16_t c5 = b5 + b7;
+  int16_t c6 = b4 - b6;
+  int16_t c7 = b5 - b7;
+
+  coeff[0] = c0 + c4;
+  coeff[7] = c1 + c5;
+  coeff[3] = c2 + c6;
+  coeff[4] = c3 + c7;
+  coeff[2] = c0 - c4;
+  coeff[6] = c1 - c5;
+  coeff[1] = c2 - c6;
+  coeff[5] = c3 - c7;
+}
+
+// The order of the output coeff of the hadamard is not important. For
+// optimization purposes the final transpose may be skipped.
+void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride,
+                        int16_t *coeff) {
+  int idx;
+  int16_t buffer[64];
+  int16_t *tmp_buf = &buffer[0];
+  for (idx = 0; idx < 8; ++idx) {
+    hadamard_col8(src_diff, src_stride, tmp_buf);  // src_diff: 9 bit
+                                                   // dynamic range [-255, 255]
+    tmp_buf += 8;
+    ++src_diff;
+  }
+
+  tmp_buf = &buffer[0];
+  for (idx = 0; idx < 8; ++idx) {
+    hadamard_col8(tmp_buf, 8, coeff);  // tmp_buf: 12 bit
+                                       // dynamic range [-2040, 2040]
+    coeff += 8;  // coeff: 15 bit
+                 // dynamic range [-16320, 16320]
+    ++tmp_buf;
+  }
+}
+
+// In place 16x16 2D Hadamard transform
+void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride,
+                          int16_t *coeff) {
+  int idx;
+  for (idx = 0; idx < 4; ++idx) {
+    // src_diff: 9 bit, dynamic range [-255, 255]
+    const int16_t *src_ptr = src_diff + (idx >> 1) * 8 * src_stride
+                                + (idx & 0x01) * 8;
+    vpx_hadamard_8x8_c(src_ptr, src_stride, coeff + idx * 64);
+  }
+
+  // coeff: 15 bit, dynamic range [-16320, 16320]
+  for (idx = 0; idx < 64; ++idx) {
+    int16_t a0 = coeff[0];
+    int16_t a1 = coeff[64];
+    int16_t a2 = coeff[128];
+    int16_t a3 = coeff[192];
+
+    int16_t b0 = (a0 + a1) >> 1;  // (a0 + a1): 16 bit, [-32640, 32640]
+    int16_t b1 = (a0 - a1) >> 1;  // b0-b3: 15 bit, dynamic range
+    int16_t b2 = (a2 + a3) >> 1;  // [-16320, 16320]
+    int16_t b3 = (a2 - a3) >> 1;
+
+    coeff[0]   = b0 + b2;  // 16 bit, [-32640, 32640]
+    coeff[64]  = b1 + b3;
+    coeff[128] = b0 - b2;
+    coeff[192] = b1 - b3;
+
+    ++coeff;
+  }
+}
+
+// coeff: 16 bits, dynamic range [-32640, 32640].
+// length: value range {16, 64, 256, 1024}.
+int vpx_satd_c(const int16_t *coeff, int length) {
+  int i;
+  int satd = 0;
+  for (i = 0; i < length; ++i)
+    satd += abs(coeff[i]);
+
+  // satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
+  return satd;
+}
+
+// Integer projection onto row vectors.
+// height: value range {16, 32, 64}.
+void vpx_int_pro_row_c(int16_t hbuf[16], const uint8_t *ref,
+                       const int ref_stride, const int height) {
+  int idx;
+  const int norm_factor = height >> 1;
+  for (idx = 0; idx < 16; ++idx) {
+    int i;
+    hbuf[idx] = 0;
+    // hbuf[idx]: 14 bit, dynamic range [0, 16320].
+    for (i = 0; i < height; ++i)
+      hbuf[idx] += ref[i * ref_stride];
+    // hbuf[idx]: 9 bit, dynamic range [0, 510].
+    hbuf[idx] /= norm_factor;
+    ++ref;
+  }
+}
+
+// width: value range {16, 32, 64}.
+int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width) {
+  int idx;
+  int16_t sum = 0;
+  // sum: 14 bit, dynamic range [0, 16320]
+  for (idx = 0; idx < width; ++idx)
+    sum += ref[idx];
+  return sum;
+}
+
+// ref: [0 - 510]
+// src: [0 - 510]
+// bwl: {2, 3, 4}
+int vpx_vector_var_c(const int16_t *ref, const int16_t *src,
+                     const int bwl) {
+  int i;
+  int width = 4 << bwl;
+  int sse = 0, mean = 0, var;
+
+  for (i = 0; i < width; ++i) {
+    int diff = ref[i] - src[i];  // diff: dynamic range [-510, 510], 10 bits.
+    mean += diff;                // mean: dynamic range 16 bits.
+    sse += diff * diff;          // sse:  dynamic range 26 bits.
+  }
+
+  // (mean * mean): dynamic range 31 bits.
+  var = sse - ((mean * mean) >> (bwl + 2));
+  return var;
+}
+
+void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp,
+                      int *min, int *max) {
+  int i, j;
+  *min = 255;
+  *max = 0;
+  for (i = 0; i < 8; ++i, s += p, d += dp) {
+    for (j = 0; j < 8; ++j) {
+      int diff = abs(s[j]-d[j]);
+      *min = diff < *min ? diff : *min;
+      *max = diff > *max ? diff : *max;
+    }
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+unsigned int vpx_highbd_avg_8x8_c(const uint8_t *s8, int p) {
+  int i, j;
+  int sum = 0;
+  const uint16_t* s = CONVERT_TO_SHORTPTR(s8);
+  for (i = 0; i < 8; ++i, s+=p)
+    for (j = 0; j < 8; sum += s[j], ++j) {}
+
+  return (sum + 32) >> 6;
+}
+
+unsigned int vpx_highbd_avg_4x4_c(const uint8_t *s8, int p) {
+  int i, j;
+  int sum = 0;
+  const uint16_t* s = CONVERT_TO_SHORTPTR(s8);
+  for (i = 0; i < 4; ++i, s+=p)
+    for (j = 0; j < 4; sum += s[j], ++j) {}
+
+  return (sum + 8) >> 4;
+}
+
+void vpx_highbd_minmax_8x8_c(const uint8_t *s8, int p, const uint8_t *d8,
+                             int dp, int *min, int *max) {
+  int i, j;
+  const uint16_t* s = CONVERT_TO_SHORTPTR(s8);
+  const uint16_t* d = CONVERT_TO_SHORTPTR(d8);
+  *min = 255;
+  *max = 0;
+  for (i = 0; i < 8; ++i, s += p, d += dp) {
+    for (j = 0; j < 8; ++j) {
+      int diff = abs(s[j]-d[j]);
+      *min = diff < *min ? diff : *min;
+      *max = diff > *max ? diff : *max;
+    }
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+
diff --git a/libvpx/libvpx/vpx_dsp/bitreader.c b/libvpx/libvpx/vpx_dsp/bitreader.c
new file mode 100644
index 0000000..8140e78
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/bitreader.c
@@ -0,0 +1,103 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <stdlib.h>
+
+#include "./vpx_config.h"
+
+#include "vpx_dsp/bitreader.h"
+#include "vpx_dsp/prob.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_util/endian_inl.h"
+
+int vpx_reader_init(vpx_reader *r,
+                    const uint8_t *buffer,
+                    size_t size,
+                    vpx_decrypt_cb decrypt_cb,
+                    void *decrypt_state) {
+  if (size && !buffer) {
+    return 1;
+  } else {
+    r->buffer_end = buffer + size;
+    r->buffer = buffer;
+    r->value = 0;
+    r->count = -8;
+    r->range = 255;
+    r->decrypt_cb = decrypt_cb;
+    r->decrypt_state = decrypt_state;
+    vpx_reader_fill(r);
+    return vpx_read_bit(r) != 0;  // marker bit
+  }
+}
+
+void vpx_reader_fill(vpx_reader *r) {
+  const uint8_t *const buffer_end = r->buffer_end;
+  const uint8_t *buffer = r->buffer;
+  const uint8_t *buffer_start = buffer;
+  BD_VALUE value = r->value;
+  int count = r->count;
+  const size_t bytes_left = buffer_end - buffer;
+  const size_t bits_left = bytes_left * CHAR_BIT;
+  int shift = BD_VALUE_SIZE - CHAR_BIT - (count + CHAR_BIT);
+
+  if (r->decrypt_cb) {
+    size_t n = VPXMIN(sizeof(r->clear_buffer), bytes_left);
+    r->decrypt_cb(r->decrypt_state, buffer, r->clear_buffer, (int)n);
+    buffer = r->clear_buffer;
+    buffer_start = r->clear_buffer;
+  }
+  if (bits_left > BD_VALUE_SIZE) {
+      const int bits = (shift & 0xfffffff8) + CHAR_BIT;
+      BD_VALUE nv;
+      BD_VALUE big_endian_values;
+      memcpy(&big_endian_values, buffer, sizeof(BD_VALUE));
+#if SIZE_MAX == 0xffffffffffffffffULL
+        big_endian_values = HToBE64(big_endian_values);
+#else
+        big_endian_values = HToBE32(big_endian_values);
+#endif
+      nv = big_endian_values >> (BD_VALUE_SIZE - bits);
+      count += bits;
+      buffer += (bits >> 3);
+      value = r->value | (nv << (shift & 0x7));
+  } else {
+    const int bits_over = (int)(shift + CHAR_BIT - (int)bits_left);
+    int loop_end = 0;
+    if (bits_over >= 0) {
+      count += LOTS_OF_BITS;
+      loop_end = bits_over;
+    }
+
+    if (bits_over < 0 || bits_left) {
+      while (shift >= loop_end) {
+        count += CHAR_BIT;
+        value |= (BD_VALUE)*buffer++ << shift;
+        shift -= CHAR_BIT;
+      }
+    }
+  }
+
+  // NOTE: Variable 'buffer' may not relate to 'r->buffer' after decryption,
+  // so we increase 'r->buffer' by the amount that 'buffer' moved, rather than
+  // assign 'buffer' to 'r->buffer'.
+  r->buffer += buffer - buffer_start;
+  r->value = value;
+  r->count = count;
+}
+
+const uint8_t *vpx_reader_find_end(vpx_reader *r) {
+  // Find the end of the coded buffer
+  while (r->count > CHAR_BIT && r->count < BD_VALUE_SIZE) {
+    r->count -= CHAR_BIT;
+    r->buffer--;
+  }
+  return r->buffer;
+}
diff --git a/libvpx/libvpx/vpx_dsp/bitreader.h b/libvpx/libvpx/vpx_dsp/bitreader.h
new file mode 100644
index 0000000..9a441b4
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/bitreader.h
@@ -0,0 +1,140 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_BITREADER_H_
+#define VPX_DSP_BITREADER_H_
+
+#include <stddef.h>
+#include <limits.h>
+
+#include "./vpx_config.h"
+#include "vpx_ports/mem.h"
+#include "vpx/vp8dx.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef size_t BD_VALUE;
+
+#define BD_VALUE_SIZE ((int)sizeof(BD_VALUE) * CHAR_BIT)
+
+// This is meant to be a large, positive constant that can still be efficiently
+// loaded as an immediate (on platforms like ARM, for example).
+// Even relatively modest values like 100 would work fine.
+#define LOTS_OF_BITS 0x40000000
+
+typedef struct {
+  // Be careful when reordering this struct, it may impact the cache negatively.
+  BD_VALUE value;
+  unsigned int range;
+  int count;
+  const uint8_t *buffer_end;
+  const uint8_t *buffer;
+  vpx_decrypt_cb decrypt_cb;
+  void *decrypt_state;
+  uint8_t clear_buffer[sizeof(BD_VALUE) + 1];
+} vpx_reader;
+
+int vpx_reader_init(vpx_reader *r,
+                    const uint8_t *buffer,
+                    size_t size,
+                    vpx_decrypt_cb decrypt_cb,
+                    void *decrypt_state);
+
+void vpx_reader_fill(vpx_reader *r);
+
+const uint8_t *vpx_reader_find_end(vpx_reader *r);
+
+static INLINE int vpx_reader_has_error(vpx_reader *r) {
+  // Check if we have reached the end of the buffer.
+  //
+  // Variable 'count' stores the number of bits in the 'value' buffer, minus
+  // 8. The top byte is part of the algorithm, and the remainder is buffered
+  // to be shifted into it. So if count == 8, the top 16 bits of 'value' are
+  // occupied, 8 for the algorithm and 8 in the buffer.
+  //
+  // When reading a byte from the user's buffer, count is filled with 8 and
+  // one byte is filled into the value buffer. When we reach the end of the
+  // data, count is additionally filled with LOTS_OF_BITS. So when
+  // count == LOTS_OF_BITS - 1, the user's data has been exhausted.
+  //
+  // 1 if we have tried to decode bits after the end of stream was encountered.
+  // 0 No error.
+  return r->count > BD_VALUE_SIZE && r->count < LOTS_OF_BITS;
+}
+
+static INLINE int vpx_read(vpx_reader *r, int prob) {
+  unsigned int bit = 0;
+  BD_VALUE value;
+  BD_VALUE bigsplit;
+  int count;
+  unsigned int range;
+  unsigned int split = (r->range * prob + (256 - prob)) >> CHAR_BIT;
+
+  if (r->count < 0)
+    vpx_reader_fill(r);
+
+  value = r->value;
+  count = r->count;
+
+  bigsplit = (BD_VALUE)split << (BD_VALUE_SIZE - CHAR_BIT);
+
+  range = split;
+
+  if (value >= bigsplit) {
+    range = r->range - split;
+    value = value - bigsplit;
+    bit = 1;
+  }
+
+  {
+    register int shift = vpx_norm[range];
+    range <<= shift;
+    value <<= shift;
+    count -= shift;
+  }
+  r->value = value;
+  r->count = count;
+  r->range = range;
+
+  return bit;
+}
+
+static INLINE int vpx_read_bit(vpx_reader *r) {
+  return vpx_read(r, 128);  // vpx_prob_half
+}
+
+static INLINE int vpx_read_literal(vpx_reader *r, int bits) {
+  int literal = 0, bit;
+
+  for (bit = bits - 1; bit >= 0; bit--)
+    literal |= vpx_read_bit(r) << bit;
+
+  return literal;
+}
+
+static INLINE int vpx_read_tree(vpx_reader *r, const vpx_tree_index *tree,
+                                const vpx_prob *probs) {
+  vpx_tree_index i = 0;
+
+  while ((i = tree[i + vpx_read(r, probs[i >> 1])]) > 0)
+    continue;
+
+  return -i;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_BITREADER_H_
diff --git a/libvpx/libvpx/vpx_dsp/bitreader_buffer.c b/libvpx/libvpx/vpx_dsp/bitreader_buffer.c
new file mode 100644
index 0000000..d7b55cf
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/bitreader_buffer.c
@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_config.h"
+#include "./bitreader_buffer.h"
+
+size_t vpx_rb_bytes_read(struct vpx_read_bit_buffer *rb) {
+  return (rb->bit_offset + 7) >> 3;
+}
+
+int vpx_rb_read_bit(struct vpx_read_bit_buffer *rb) {
+  const size_t off = rb->bit_offset;
+  const size_t p = off >> 3;
+  const int q = 7 - (int)(off & 0x7);
+  if (rb->bit_buffer + p < rb->bit_buffer_end) {
+    const int bit = (rb->bit_buffer[p] >> q) & 1;
+    rb->bit_offset = off + 1;
+    return bit;
+  } else {
+    rb->error_handler(rb->error_handler_data);
+    return 0;
+  }
+}
+
+int vpx_rb_read_literal(struct vpx_read_bit_buffer *rb, int bits) {
+  int value = 0, bit;
+  for (bit = bits - 1; bit >= 0; bit--)
+    value |= vpx_rb_read_bit(rb) << bit;
+  return value;
+}
+
+int vpx_rb_read_signed_literal(struct vpx_read_bit_buffer *rb,
+                               int bits) {
+  const int value = vpx_rb_read_literal(rb, bits);
+  return vpx_rb_read_bit(rb) ? -value : value;
+}
+
+int vpx_rb_read_inv_signed_literal(struct vpx_read_bit_buffer *rb,
+                                   int bits) {
+#if CONFIG_MISC_FIXES
+  const int nbits = sizeof(unsigned) * 8 - bits - 1;
+  const unsigned value = (unsigned)vpx_rb_read_literal(rb, bits + 1) << nbits;
+  return ((int) value) >> nbits;
+#else
+  return vpx_rb_read_signed_literal(rb, bits);
+#endif
+}
diff --git a/libvpx/libvpx/vpx_dsp/bitreader_buffer.h b/libvpx/libvpx/vpx_dsp/bitreader_buffer.h
new file mode 100644
index 0000000..8a48a95
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/bitreader_buffer.h
@@ -0,0 +1,47 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_BITREADER_BUFFER_H_
+#define VPX_DSP_BITREADER_BUFFER_H_
+
+#include <limits.h>
+
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef void (*vpx_rb_error_handler)(void *data);
+
+struct vpx_read_bit_buffer {
+  const uint8_t *bit_buffer;
+  const uint8_t *bit_buffer_end;
+  size_t bit_offset;
+
+  void *error_handler_data;
+  vpx_rb_error_handler error_handler;
+};
+
+size_t vpx_rb_bytes_read(struct vpx_read_bit_buffer *rb);
+
+int vpx_rb_read_bit(struct vpx_read_bit_buffer *rb);
+
+int vpx_rb_read_literal(struct vpx_read_bit_buffer *rb, int bits);
+
+int vpx_rb_read_signed_literal(struct vpx_read_bit_buffer *rb, int bits);
+
+int vpx_rb_read_inv_signed_literal(struct vpx_read_bit_buffer *rb, int bits);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_BITREADER_BUFFER_H_
diff --git a/libvpx/libvpx/vpx_dsp/bitwriter.c b/libvpx/libvpx/vpx_dsp/bitwriter.c
new file mode 100644
index 0000000..5b232e3
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/bitwriter.c
@@ -0,0 +1,34 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "./bitwriter.h"
+
+void vpx_start_encode(vpx_writer *br, uint8_t *source) {
+  br->lowvalue = 0;
+  br->range    = 255;
+  br->count    = -24;
+  br->buffer   = source;
+  br->pos      = 0;
+  vpx_write_bit(br, 0);
+}
+
+void vpx_stop_encode(vpx_writer *br) {
+  int i;
+
+  for (i = 0; i < 32; i++)
+    vpx_write_bit(br, 0);
+
+  // Ensure there's no ambigous collision with any index marker bytes
+  if ((br->buffer[br->pos - 1] & 0xe0) == 0xc0)
+    br->buffer[br->pos++] = 0;
+}
+
diff --git a/libvpx/libvpx/vpx_dsp/bitwriter.h b/libvpx/libvpx/vpx_dsp/bitwriter.h
new file mode 100644
index 0000000..d904997
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/bitwriter.h
@@ -0,0 +1,98 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_BITWRITER_H_
+#define VPX_DSP_BITWRITER_H_
+
+#include "vpx_ports/mem.h"
+
+#include "vpx_dsp/prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct vpx_writer {
+  unsigned int lowvalue;
+  unsigned int range;
+  int count;
+  unsigned int pos;
+  uint8_t *buffer;
+} vpx_writer;
+
+void vpx_start_encode(vpx_writer *bc, uint8_t *buffer);
+void vpx_stop_encode(vpx_writer *bc);
+
+static INLINE void vpx_write(vpx_writer *br, int bit, int probability) {
+  unsigned int split;
+  int count = br->count;
+  unsigned int range = br->range;
+  unsigned int lowvalue = br->lowvalue;
+  register int shift;
+
+  split = 1 + (((range - 1) * probability) >> 8);
+
+  range = split;
+
+  if (bit) {
+    lowvalue += split;
+    range = br->range - split;
+  }
+
+  shift = vpx_norm[range];
+
+  range <<= shift;
+  count += shift;
+
+  if (count >= 0) {
+    int offset = shift - count;
+
+    if ((lowvalue << (offset - 1)) & 0x80000000) {
+      int x = br->pos - 1;
+
+      while (x >= 0 && br->buffer[x] == 0xff) {
+        br->buffer[x] = 0;
+        x--;
+      }
+
+      br->buffer[x] += 1;
+    }
+
+    br->buffer[br->pos++] = (lowvalue >> (24 - offset));
+    lowvalue <<= offset;
+    shift = count;
+    lowvalue &= 0xffffff;
+    count -= 8;
+  }
+
+  lowvalue <<= shift;
+  br->count = count;
+  br->lowvalue = lowvalue;
+  br->range = range;
+}
+
+static INLINE void vpx_write_bit(vpx_writer *w, int bit) {
+  vpx_write(w, bit, 128);  // vpx_prob_half
+}
+
+static INLINE void vpx_write_literal(vpx_writer *w, int data, int bits) {
+  int bit;
+
+  for (bit = bits - 1; bit >= 0; bit--)
+    vpx_write_bit(w, 1 & (data >> bit));
+}
+
+#define vpx_write_prob(w, v) vpx_write_literal((w), (v), 8)
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_BITWRITER_H_
diff --git a/libvpx/libvpx/vpx_dsp/bitwriter_buffer.c b/libvpx/libvpx/vpx_dsp/bitwriter_buffer.c
new file mode 100644
index 0000000..6182a72
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/bitwriter_buffer.c
@@ -0,0 +1,48 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <stdlib.h>
+
+#include "./vpx_config.h"
+#include "./bitwriter_buffer.h"
+
+size_t vpx_wb_bytes_written(const struct vpx_write_bit_buffer *wb) {
+  return wb->bit_offset / CHAR_BIT + (wb->bit_offset % CHAR_BIT > 0);
+}
+
+void vpx_wb_write_bit(struct vpx_write_bit_buffer *wb, int bit) {
+  const int off = (int)wb->bit_offset;
+  const int p = off / CHAR_BIT;
+  const int q = CHAR_BIT - 1 - off % CHAR_BIT;
+  if (q == CHAR_BIT -1) {
+    wb->bit_buffer[p] = bit << q;
+  } else {
+    wb->bit_buffer[p] &= ~(1 << q);
+    wb->bit_buffer[p] |= bit << q;
+  }
+  wb->bit_offset = off + 1;
+}
+
+void vpx_wb_write_literal(struct vpx_write_bit_buffer *wb, int data, int bits) {
+  int bit;
+  for (bit = bits - 1; bit >= 0; bit--)
+    vpx_wb_write_bit(wb, (data >> bit) & 1);
+}
+
+void vpx_wb_write_inv_signed_literal(struct vpx_write_bit_buffer *wb,
+                                     int data, int bits) {
+#if CONFIG_MISC_FIXES
+  vpx_wb_write_literal(wb, data, bits + 1);
+#else
+  vpx_wb_write_literal(wb, abs(data), bits);
+  vpx_wb_write_bit(wb, data < 0);
+#endif
+}
diff --git a/libvpx/libvpx/vpx_dsp/bitwriter_buffer.h b/libvpx/libvpx/vpx_dsp/bitwriter_buffer.h
new file mode 100644
index 0000000..a123a2f
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/bitwriter_buffer.h
@@ -0,0 +1,38 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_BITWRITER_BUFFER_H_
+#define VPX_DSP_BITWRITER_BUFFER_H_
+
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct vpx_write_bit_buffer {
+  uint8_t *bit_buffer;
+  size_t bit_offset;
+};
+
+size_t vpx_wb_bytes_written(const struct vpx_write_bit_buffer *wb);
+
+void vpx_wb_write_bit(struct vpx_write_bit_buffer *wb, int bit);
+
+void vpx_wb_write_literal(struct vpx_write_bit_buffer *wb, int data, int bits);
+
+void vpx_wb_write_inv_signed_literal(struct vpx_write_bit_buffer *wb, int data,
+                                     int bits);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_BITWRITER_BUFFER_H_
diff --git a/libvpx/libvpx/vpx_dsp/fastssim.c b/libvpx/libvpx/vpx_dsp/fastssim.c
new file mode 100644
index 0000000..1405a30
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/fastssim.c
@@ -0,0 +1,468 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ *
+ *  This code was originally written by: Nathan E. Egge, at the Daala
+ *  project.
+ */
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/ssim.h"
+#include "vpx_ports/system_state.h"
+/* TODO(jbb): High bit depth version of this code needed */
+typedef struct fs_level fs_level;
+typedef struct fs_ctx fs_ctx;
+
+#define SSIM_C1 (255 * 255 * 0.01 * 0.01)
+#define SSIM_C2 (255 * 255 * 0.03 * 0.03)
+
+#define FS_MINI(_a, _b) ((_a) < (_b) ? (_a) : (_b))
+#define FS_MAXI(_a, _b) ((_a) > (_b) ? (_a) : (_b))
+
+struct fs_level {
+  uint16_t *im1;
+  uint16_t *im2;
+  double *ssim;
+  int w;
+  int h;
+};
+
+struct fs_ctx {
+  fs_level *level;
+  int nlevels;
+  unsigned *col_buf;
+};
+
+static void fs_ctx_init(fs_ctx *_ctx, int _w, int _h, int _nlevels) {
+  unsigned char *data;
+  size_t data_size;
+  int lw;
+  int lh;
+  int l;
+  lw = (_w + 1) >> 1;
+  lh = (_h + 1) >> 1;
+  data_size = _nlevels * sizeof(fs_level)
+      + 2 * (lw + 8) * 8 * sizeof(*_ctx->col_buf);
+  for (l = 0; l < _nlevels; l++) {
+    size_t im_size;
+    size_t level_size;
+    im_size = lw * (size_t) lh;
+    level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
+    level_size += sizeof(*_ctx->level[l].ssim) - 1;
+    level_size /= sizeof(*_ctx->level[l].ssim);
+    level_size += im_size;
+    level_size *= sizeof(*_ctx->level[l].ssim);
+    data_size += level_size;
+    lw = (lw + 1) >> 1;
+    lh = (lh + 1) >> 1;
+  }
+  data = (unsigned char *) malloc(data_size);
+  _ctx->level = (fs_level *) data;
+  _ctx->nlevels = _nlevels;
+  data += _nlevels * sizeof(*_ctx->level);
+  lw = (_w + 1) >> 1;
+  lh = (_h + 1) >> 1;
+  for (l = 0; l < _nlevels; l++) {
+    size_t im_size;
+    size_t level_size;
+    _ctx->level[l].w = lw;
+    _ctx->level[l].h = lh;
+    im_size = lw * (size_t) lh;
+    level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
+    level_size += sizeof(*_ctx->level[l].ssim) - 1;
+    level_size /= sizeof(*_ctx->level[l].ssim);
+    level_size *= sizeof(*_ctx->level[l].ssim);
+    _ctx->level[l].im1 = (uint16_t *) data;
+    _ctx->level[l].im2 = _ctx->level[l].im1 + im_size;
+    data += level_size;
+    _ctx->level[l].ssim = (double *) data;
+    data += im_size * sizeof(*_ctx->level[l].ssim);
+    lw = (lw + 1) >> 1;
+    lh = (lh + 1) >> 1;
+  }
+  _ctx->col_buf = (unsigned *) data;
+}
+
+static void fs_ctx_clear(fs_ctx *_ctx) {
+  free(_ctx->level);
+}
+
+static void fs_downsample_level(fs_ctx *_ctx, int _l) {
+  const uint16_t *src1;
+  const uint16_t *src2;
+  uint16_t *dst1;
+  uint16_t *dst2;
+  int w2;
+  int h2;
+  int w;
+  int h;
+  int i;
+  int j;
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  dst1 = _ctx->level[_l].im1;
+  dst2 = _ctx->level[_l].im2;
+  w2 = _ctx->level[_l - 1].w;
+  h2 = _ctx->level[_l - 1].h;
+  src1 = _ctx->level[_l - 1].im1;
+  src2 = _ctx->level[_l - 1].im2;
+  for (j = 0; j < h; j++) {
+    int j0offs;
+    int j1offs;
+    j0offs = 2 * j * w2;
+    j1offs = FS_MINI(2 * j + 1, h2) * w2;
+    for (i = 0; i < w; i++) {
+      int i0;
+      int i1;
+      i0 = 2 * i;
+      i1 = FS_MINI(i0 + 1, w2);
+      dst1[j * w + i] = src1[j0offs + i0] + src1[j0offs + i1]
+          + src1[j1offs + i0] + src1[j1offs + i1];
+      dst2[j * w + i] = src2[j0offs + i0] + src2[j0offs + i1]
+          + src2[j1offs + i0] + src2[j1offs + i1];
+    }
+  }
+}
+
+static void fs_downsample_level0(fs_ctx *_ctx, const unsigned char *_src1,
+                                 int _s1ystride, const unsigned char *_src2,
+                                 int _s2ystride, int _w, int _h) {
+  uint16_t *dst1;
+  uint16_t *dst2;
+  int w;
+  int h;
+  int i;
+  int j;
+  w = _ctx->level[0].w;
+  h = _ctx->level[0].h;
+  dst1 = _ctx->level[0].im1;
+  dst2 = _ctx->level[0].im2;
+  for (j = 0; j < h; j++) {
+    int j0;
+    int j1;
+    j0 = 2 * j;
+    j1 = FS_MINI(j0 + 1, _h);
+    for (i = 0; i < w; i++) {
+      int i0;
+      int i1;
+      i0 = 2 * i;
+      i1 = FS_MINI(i0 + 1, _w);
+      dst1[j * w + i] = _src1[j0 * _s1ystride + i0]
+          + _src1[j0 * _s1ystride + i1] + _src1[j1 * _s1ystride + i0]
+          + _src1[j1 * _s1ystride + i1];
+      dst2[j * w + i] = _src2[j0 * _s2ystride + i0]
+          + _src2[j0 * _s2ystride + i1] + _src2[j1 * _s2ystride + i0]
+          + _src2[j1 * _s2ystride + i1];
+    }
+  }
+}
+
+static void fs_apply_luminance(fs_ctx *_ctx, int _l) {
+  unsigned *col_sums_x;
+  unsigned *col_sums_y;
+  uint16_t *im1;
+  uint16_t *im2;
+  double *ssim;
+  double c1;
+  int w;
+  int h;
+  int j0offs;
+  int j1offs;
+  int i;
+  int j;
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  col_sums_x = _ctx->col_buf;
+  col_sums_y = col_sums_x + w;
+  im1 = _ctx->level[_l].im1;
+  im2 = _ctx->level[_l].im2;
+  for (i = 0; i < w; i++)
+    col_sums_x[i] = 5 * im1[i];
+  for (i = 0; i < w; i++)
+    col_sums_y[i] = 5 * im2[i];
+  for (j = 1; j < 4; j++) {
+    j1offs = FS_MINI(j, h - 1) * w;
+    for (i = 0; i < w; i++)
+      col_sums_x[i] += im1[j1offs + i];
+    for (i = 0; i < w; i++)
+      col_sums_y[i] += im2[j1offs + i];
+  }
+  ssim = _ctx->level[_l].ssim;
+  c1 = (double) (SSIM_C1 * 4096 * (1 << 4 * _l));
+  for (j = 0; j < h; j++) {
+    unsigned mux;
+    unsigned muy;
+    int i0;
+    int i1;
+    mux = 5 * col_sums_x[0];
+    muy = 5 * col_sums_y[0];
+    for (i = 1; i < 4; i++) {
+      i1 = FS_MINI(i, w - 1);
+      mux += col_sums_x[i1];
+      muy += col_sums_y[i1];
+    }
+    for (i = 0; i < w; i++) {
+      ssim[j * w + i] *= (2 * mux * (double) muy + c1)
+          / (mux * (double) mux + muy * (double) muy + c1);
+      if (i + 1 < w) {
+        i0 = FS_MAXI(0, i - 4);
+        i1 = FS_MINI(i + 4, w - 1);
+        mux += col_sums_x[i1] - col_sums_x[i0];
+        muy += col_sums_x[i1] - col_sums_x[i0];
+      }
+    }
+    if (j + 1 < h) {
+      j0offs = FS_MAXI(0, j - 4) * w;
+      for (i = 0; i < w; i++)
+        col_sums_x[i] -= im1[j0offs + i];
+      for (i = 0; i < w; i++)
+        col_sums_y[i] -= im2[j0offs + i];
+      j1offs = FS_MINI(j + 4, h - 1) * w;
+      for (i = 0; i < w; i++)
+        col_sums_x[i] += im1[j1offs + i];
+      for (i = 0; i < w; i++)
+        col_sums_y[i] += im2[j1offs + i];
+    }
+  }
+}
+
+#define FS_COL_SET(_col, _joffs, _ioffs) \
+  do { \
+    unsigned gx; \
+    unsigned gy; \
+    gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    col_sums_gx2[(_col)] = gx * (double)gx; \
+    col_sums_gy2[(_col)] = gy * (double)gy; \
+    col_sums_gxgy[(_col)] = gx * (double)gy; \
+  } \
+  while (0)
+
+#define FS_COL_ADD(_col, _joffs, _ioffs) \
+  do { \
+    unsigned gx; \
+    unsigned gy; \
+    gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    col_sums_gx2[(_col)] += gx * (double)gx; \
+    col_sums_gy2[(_col)] += gy * (double)gy; \
+    col_sums_gxgy[(_col)] += gx * (double)gy; \
+  } \
+  while (0)
+
+#define FS_COL_SUB(_col, _joffs, _ioffs) \
+  do { \
+    unsigned gx; \
+    unsigned gy; \
+    gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    col_sums_gx2[(_col)] -= gx * (double)gx; \
+    col_sums_gy2[(_col)] -= gy * (double)gy; \
+    col_sums_gxgy[(_col)] -= gx * (double)gy; \
+  } \
+  while (0)
+
+#define FS_COL_COPY(_col1, _col2) \
+  do { \
+    col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)]; \
+    col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)]; \
+    col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)]; \
+  } \
+  while (0)
+
+#define FS_COL_HALVE(_col1, _col2) \
+  do { \
+    col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 0.5; \
+    col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 0.5; \
+    col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 0.5; \
+  } \
+  while (0)
+
+#define FS_COL_DOUBLE(_col1, _col2) \
+  do { \
+    col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 2; \
+    col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 2; \
+    col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 2; \
+  } \
+  while (0)
+
+static void fs_calc_structure(fs_ctx *_ctx, int _l) {
+  uint16_t *im1;
+  uint16_t *im2;
+  unsigned *gx_buf;
+  unsigned *gy_buf;
+  double *ssim;
+  double col_sums_gx2[8];
+  double col_sums_gy2[8];
+  double col_sums_gxgy[8];
+  double c2;
+  int stride;
+  int w;
+  int h;
+  int i;
+  int j;
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  im1 = _ctx->level[_l].im1;
+  im2 = _ctx->level[_l].im2;
+  ssim = _ctx->level[_l].ssim;
+  gx_buf = _ctx->col_buf;
+  stride = w + 8;
+  gy_buf = gx_buf + 8 * stride;
+  memset(gx_buf, 0, 2 * 8 * stride * sizeof(*gx_buf));
+  c2 = SSIM_C2 * (1 << 4 * _l) * 16 * 104;
+  for (j = 0; j < h + 4; j++) {
+    if (j < h - 1) {
+      for (i = 0; i < w - 1; i++) {
+        unsigned g1;
+        unsigned g2;
+        unsigned gx;
+        unsigned gy;
+        g1 = abs(im1[(j + 1) * w + i + 1] - im1[j * w + i]);
+        g2 = abs(im1[(j + 1) * w + i] - im1[j * w + i + 1]);
+        gx = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
+        g1 = abs(im2[(j + 1) * w + i + 1] - im2[j * w + i]);
+        g2 = abs(im2[(j + 1) * w + i] - im2[j * w + i + 1]);
+        gy = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
+        gx_buf[(j & 7) * stride + i + 4] = gx;
+        gy_buf[(j & 7) * stride + i + 4] = gy;
+      }
+    } else {
+      memset(gx_buf + (j & 7) * stride, 0, stride * sizeof(*gx_buf));
+      memset(gy_buf + (j & 7) * stride, 0, stride * sizeof(*gy_buf));
+    }
+    if (j >= 4) {
+      int k;
+      col_sums_gx2[3] = col_sums_gx2[2] = col_sums_gx2[1] = col_sums_gx2[0] = 0;
+      col_sums_gy2[3] = col_sums_gy2[2] = col_sums_gy2[1] = col_sums_gy2[0] = 0;
+      col_sums_gxgy[3] = col_sums_gxgy[2] = col_sums_gxgy[1] =
+          col_sums_gxgy[0] = 0;
+      for (i = 4; i < 8; i++) {
+        FS_COL_SET(i, -1, 0);
+        FS_COL_ADD(i, 0, 0);
+        for (k = 1; k < 8 - i; k++) {
+          FS_COL_DOUBLE(i, i);
+          FS_COL_ADD(i, -k - 1, 0);
+          FS_COL_ADD(i, k, 0);
+        }
+      }
+      for (i = 0; i < w; i++) {
+        double mugx2;
+        double mugy2;
+        double mugxgy;
+        mugx2 = col_sums_gx2[0];
+        for (k = 1; k < 8; k++)
+          mugx2 += col_sums_gx2[k];
+        mugy2 = col_sums_gy2[0];
+        for (k = 1; k < 8; k++)
+          mugy2 += col_sums_gy2[k];
+        mugxgy = col_sums_gxgy[0];
+        for (k = 1; k < 8; k++)
+          mugxgy += col_sums_gxgy[k];
+        ssim[(j - 4) * w + i] = (2 * mugxgy + c2) / (mugx2 + mugy2 + c2);
+        if (i + 1 < w) {
+          FS_COL_SET(0, -1, 1);
+          FS_COL_ADD(0, 0, 1);
+          FS_COL_SUB(2, -3, 2);
+          FS_COL_SUB(2, 2, 2);
+          FS_COL_HALVE(1, 2);
+          FS_COL_SUB(3, -4, 3);
+          FS_COL_SUB(3, 3, 3);
+          FS_COL_HALVE(2, 3);
+          FS_COL_COPY(3, 4);
+          FS_COL_DOUBLE(4, 5);
+          FS_COL_ADD(4, -4, 5);
+          FS_COL_ADD(4, 3, 5);
+          FS_COL_DOUBLE(5, 6);
+          FS_COL_ADD(5, -3, 6);
+          FS_COL_ADD(5, 2, 6);
+          FS_COL_DOUBLE(6, 7);
+          FS_COL_ADD(6, -2, 7);
+          FS_COL_ADD(6, 1, 7);
+          FS_COL_SET(7, -1, 8);
+          FS_COL_ADD(7, 0, 8);
+        }
+      }
+    }
+  }
+}
+
+#define FS_NLEVELS (4)
+
+/*These weights were derived from the default weights found in Wang's original
+ Matlab implementation: {0.0448, 0.2856, 0.2363, 0.1333}.
+ We drop the finest scale and renormalize the rest to sum to 1.*/
+
+static const double FS_WEIGHTS[FS_NLEVELS] = {0.2989654541015625,
+    0.3141326904296875, 0.2473602294921875, 0.1395416259765625};
+
+static double fs_average(fs_ctx *_ctx, int _l) {
+  double *ssim;
+  double ret;
+  int w;
+  int h;
+  int i;
+  int j;
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  ssim = _ctx->level[_l].ssim;
+  ret = 0;
+  for (j = 0; j < h; j++)
+    for (i = 0; i < w; i++)
+      ret += ssim[j * w + i];
+  return pow(ret / (w * h), FS_WEIGHTS[_l]);
+}
+
+static double calc_ssim(const unsigned char *_src, int _systride,
+                 const unsigned char *_dst, int _dystride, int _w, int _h) {
+  fs_ctx ctx;
+  double ret;
+  int l;
+  ret = 1;
+  fs_ctx_init(&ctx, _w, _h, FS_NLEVELS);
+  fs_downsample_level0(&ctx, _src, _systride, _dst, _dystride, _w, _h);
+  for (l = 0; l < FS_NLEVELS - 1; l++) {
+    fs_calc_structure(&ctx, l);
+    ret *= fs_average(&ctx, l);
+    fs_downsample_level(&ctx, l + 1);
+  }
+  fs_calc_structure(&ctx, l);
+  fs_apply_luminance(&ctx, l);
+  ret *= fs_average(&ctx, l);
+  fs_ctx_clear(&ctx);
+  return ret;
+}
+
+static double convert_ssim_db(double _ssim, double _weight) {
+  return 10 * (log10(_weight) - log10(_weight - _ssim));
+}
+
+double vpx_calc_fastssim(const YV12_BUFFER_CONFIG *source,
+                         const YV12_BUFFER_CONFIG *dest,
+                         double *ssim_y, double *ssim_u, double *ssim_v) {
+  double ssimv;
+  vpx_clear_system_state();
+
+  *ssim_y = calc_ssim(source->y_buffer, source->y_stride, dest->y_buffer,
+                      dest->y_stride, source->y_crop_width,
+                      source->y_crop_height);
+
+  *ssim_u = calc_ssim(source->u_buffer, source->uv_stride, dest->u_buffer,
+                      dest->uv_stride, source->uv_crop_width,
+                      source->uv_crop_height);
+
+  *ssim_v = calc_ssim(source->v_buffer, source->uv_stride, dest->v_buffer,
+                      dest->uv_stride, source->uv_crop_width,
+                      source->uv_crop_height);
+  ssimv = (*ssim_y) * .8 + .1 * ((*ssim_u) + (*ssim_v));
+
+  return convert_ssim_db(ssimv, 1.0);
+}
diff --git a/libvpx/libvpx/vpx_dsp/fwd_txfm.c b/libvpx/libvpx/vpx_dsp/fwd_txfm.c
new file mode 100644
index 0000000..4c0d5db
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/fwd_txfm.c
@@ -0,0 +1,819 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/fwd_txfm.h"
+
+void vpx_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride) {
+  // The 2D transform is done with two passes which are actually pretty
+  // similar. In the first one, we transform the columns and transpose
+  // the results. In the second one, we transform the rows. To achieve that,
+  // as the first pass results are transposed, we transpose the columns (that
+  // is the transposed rows) and transpose the results (so that it goes back
+  // in normal/row positions).
+  int pass;
+  // We need an intermediate buffer between passes.
+  tran_low_t intermediate[4 * 4];
+  const int16_t *in_pass0 = input;
+  const tran_low_t *in = NULL;
+  tran_low_t *out = intermediate;
+  // Do the two transform/transpose passes
+  for (pass = 0; pass < 2; ++pass) {
+    tran_high_t input[4];      // canbe16
+    tran_high_t step[4];       // canbe16
+    tran_high_t temp1, temp2;  // needs32
+    int i;
+    for (i = 0; i < 4; ++i) {
+      // Load inputs.
+      if (0 == pass) {
+        input[0] = in_pass0[0 * stride] * 16;
+        input[1] = in_pass0[1 * stride] * 16;
+        input[2] = in_pass0[2 * stride] * 16;
+        input[3] = in_pass0[3 * stride] * 16;
+        if (i == 0 && input[0]) {
+          input[0] += 1;
+        }
+      } else {
+        input[0] = in[0 * 4];
+        input[1] = in[1 * 4];
+        input[2] = in[2 * 4];
+        input[3] = in[3 * 4];
+      }
+      // Transform.
+      step[0] = input[0] + input[3];
+      step[1] = input[1] + input[2];
+      step[2] = input[1] - input[2];
+      step[3] = input[0] - input[3];
+      temp1 = (step[0] + step[1]) * cospi_16_64;
+      temp2 = (step[0] - step[1]) * cospi_16_64;
+      out[0] = (tran_low_t)fdct_round_shift(temp1);
+      out[2] = (tran_low_t)fdct_round_shift(temp2);
+      temp1 = step[2] * cospi_24_64 + step[3] * cospi_8_64;
+      temp2 = -step[2] * cospi_8_64 + step[3] * cospi_24_64;
+      out[1] = (tran_low_t)fdct_round_shift(temp1);
+      out[3] = (tran_low_t)fdct_round_shift(temp2);
+      // Do next column (which is a transposed row in second/horizontal pass)
+      in_pass0++;
+      in++;
+      out += 4;
+    }
+    // Setup in/out for next pass.
+    in = intermediate;
+    out = output;
+  }
+
+  {
+    int i, j;
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j)
+        output[j + i * 4] = (output[j + i * 4] + 1) >> 2;
+    }
+  }
+}
+
+void vpx_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  tran_low_t sum = 0;
+  for (r = 0; r < 4; ++r)
+    for (c = 0; c < 4; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = sum << 1;
+}
+
+void vpx_fdct8x8_c(const int16_t *input, tran_low_t *final_output, int stride) {
+  int i, j;
+  tran_low_t intermediate[64];
+  int pass;
+  tran_low_t *output = intermediate;
+  const tran_low_t *in = NULL;
+
+  // Transform columns
+  for (pass = 0; pass < 2; ++pass) {
+    tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
+    tran_high_t t0, t1, t2, t3;                  // needs32
+    tran_high_t x0, x1, x2, x3;                  // canbe16
+
+    int i;
+    for (i = 0; i < 8; i++) {
+      // stage 1
+      if (pass == 0) {
+        s0 = (input[0 * stride] + input[7 * stride]) * 4;
+        s1 = (input[1 * stride] + input[6 * stride]) * 4;
+        s2 = (input[2 * stride] + input[5 * stride]) * 4;
+        s3 = (input[3 * stride] + input[4 * stride]) * 4;
+        s4 = (input[3 * stride] - input[4 * stride]) * 4;
+        s5 = (input[2 * stride] - input[5 * stride]) * 4;
+        s6 = (input[1 * stride] - input[6 * stride]) * 4;
+        s7 = (input[0 * stride] - input[7 * stride]) * 4;
+        ++input;
+      } else {
+        s0 = in[0 * 8] + in[7 * 8];
+        s1 = in[1 * 8] + in[6 * 8];
+        s2 = in[2 * 8] + in[5 * 8];
+        s3 = in[3 * 8] + in[4 * 8];
+        s4 = in[3 * 8] - in[4 * 8];
+        s5 = in[2 * 8] - in[5 * 8];
+        s6 = in[1 * 8] - in[6 * 8];
+        s7 = in[0 * 8] - in[7 * 8];
+        ++in;
+      }
+
+      // fdct4(step, step);
+      x0 = s0 + s3;
+      x1 = s1 + s2;
+      x2 = s1 - s2;
+      x3 = s0 - s3;
+      t0 = (x0 + x1) * cospi_16_64;
+      t1 = (x0 - x1) * cospi_16_64;
+      t2 =  x2 * cospi_24_64 + x3 *  cospi_8_64;
+      t3 = -x2 * cospi_8_64  + x3 * cospi_24_64;
+      output[0] = (tran_low_t)fdct_round_shift(t0);
+      output[2] = (tran_low_t)fdct_round_shift(t2);
+      output[4] = (tran_low_t)fdct_round_shift(t1);
+      output[6] = (tran_low_t)fdct_round_shift(t3);
+
+      // Stage 2
+      t0 = (s6 - s5) * cospi_16_64;
+      t1 = (s6 + s5) * cospi_16_64;
+      t2 = fdct_round_shift(t0);
+      t3 = fdct_round_shift(t1);
+
+      // Stage 3
+      x0 = s4 + t2;
+      x1 = s4 - t2;
+      x2 = s7 - t3;
+      x3 = s7 + t3;
+
+      // Stage 4
+      t0 = x0 * cospi_28_64 + x3 *   cospi_4_64;
+      t1 = x1 * cospi_12_64 + x2 *  cospi_20_64;
+      t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
+      t3 = x3 * cospi_28_64 + x0 *  -cospi_4_64;
+      output[1] = (tran_low_t)fdct_round_shift(t0);
+      output[3] = (tran_low_t)fdct_round_shift(t2);
+      output[5] = (tran_low_t)fdct_round_shift(t1);
+      output[7] = (tran_low_t)fdct_round_shift(t3);
+      output += 8;
+    }
+    in  = intermediate;
+    output = final_output;
+  }
+
+  // Rows
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      final_output[j + i * 8] /= 2;
+  }
+}
+
+void vpx_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  tran_low_t sum = 0;
+  for (r = 0; r < 8; ++r)
+    for (c = 0; c < 8; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = sum;
+}
+
+void vpx_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride) {
+  // The 2D transform is done with two passes which are actually pretty
+  // similar. In the first one, we transform the columns and transpose
+  // the results. In the second one, we transform the rows. To achieve that,
+  // as the first pass results are transposed, we transpose the columns (that
+  // is the transposed rows) and transpose the results (so that it goes back
+  // in normal/row positions).
+  int pass;
+  // We need an intermediate buffer between passes.
+  tran_low_t intermediate[256];
+  const int16_t *in_pass0 = input;
+  const tran_low_t *in = NULL;
+  tran_low_t *out = intermediate;
+  // Do the two transform/transpose passes
+  for (pass = 0; pass < 2; ++pass) {
+    tran_high_t step1[8];      // canbe16
+    tran_high_t step2[8];      // canbe16
+    tran_high_t step3[8];      // canbe16
+    tran_high_t input[8];      // canbe16
+    tran_high_t temp1, temp2;  // needs32
+    int i;
+    for (i = 0; i < 16; i++) {
+      if (0 == pass) {
+        // Calculate input for the first 8 results.
+        input[0] = (in_pass0[0 * stride] + in_pass0[15 * stride]) * 4;
+        input[1] = (in_pass0[1 * stride] + in_pass0[14 * stride]) * 4;
+        input[2] = (in_pass0[2 * stride] + in_pass0[13 * stride]) * 4;
+        input[3] = (in_pass0[3 * stride] + in_pass0[12 * stride]) * 4;
+        input[4] = (in_pass0[4 * stride] + in_pass0[11 * stride]) * 4;
+        input[5] = (in_pass0[5 * stride] + in_pass0[10 * stride]) * 4;
+        input[6] = (in_pass0[6 * stride] + in_pass0[ 9 * stride]) * 4;
+        input[7] = (in_pass0[7 * stride] + in_pass0[ 8 * stride]) * 4;
+        // Calculate input for the next 8 results.
+        step1[0] = (in_pass0[7 * stride] - in_pass0[ 8 * stride]) * 4;
+        step1[1] = (in_pass0[6 * stride] - in_pass0[ 9 * stride]) * 4;
+        step1[2] = (in_pass0[5 * stride] - in_pass0[10 * stride]) * 4;
+        step1[3] = (in_pass0[4 * stride] - in_pass0[11 * stride]) * 4;
+        step1[4] = (in_pass0[3 * stride] - in_pass0[12 * stride]) * 4;
+        step1[5] = (in_pass0[2 * stride] - in_pass0[13 * stride]) * 4;
+        step1[6] = (in_pass0[1 * stride] - in_pass0[14 * stride]) * 4;
+        step1[7] = (in_pass0[0 * stride] - in_pass0[15 * stride]) * 4;
+      } else {
+        // Calculate input for the first 8 results.
+        input[0] = ((in[0 * 16] + 1) >> 2) + ((in[15 * 16] + 1) >> 2);
+        input[1] = ((in[1 * 16] + 1) >> 2) + ((in[14 * 16] + 1) >> 2);
+        input[2] = ((in[2 * 16] + 1) >> 2) + ((in[13 * 16] + 1) >> 2);
+        input[3] = ((in[3 * 16] + 1) >> 2) + ((in[12 * 16] + 1) >> 2);
+        input[4] = ((in[4 * 16] + 1) >> 2) + ((in[11 * 16] + 1) >> 2);
+        input[5] = ((in[5 * 16] + 1) >> 2) + ((in[10 * 16] + 1) >> 2);
+        input[6] = ((in[6 * 16] + 1) >> 2) + ((in[ 9 * 16] + 1) >> 2);
+        input[7] = ((in[7 * 16] + 1) >> 2) + ((in[ 8 * 16] + 1) >> 2);
+        // Calculate input for the next 8 results.
+        step1[0] = ((in[7 * 16] + 1) >> 2) - ((in[ 8 * 16] + 1) >> 2);
+        step1[1] = ((in[6 * 16] + 1) >> 2) - ((in[ 9 * 16] + 1) >> 2);
+        step1[2] = ((in[5 * 16] + 1) >> 2) - ((in[10 * 16] + 1) >> 2);
+        step1[3] = ((in[4 * 16] + 1) >> 2) - ((in[11 * 16] + 1) >> 2);
+        step1[4] = ((in[3 * 16] + 1) >> 2) - ((in[12 * 16] + 1) >> 2);
+        step1[5] = ((in[2 * 16] + 1) >> 2) - ((in[13 * 16] + 1) >> 2);
+        step1[6] = ((in[1 * 16] + 1) >> 2) - ((in[14 * 16] + 1) >> 2);
+        step1[7] = ((in[0 * 16] + 1) >> 2) - ((in[15 * 16] + 1) >> 2);
+      }
+      // Work on the first eight values; fdct8(input, even_results);
+      {
+        tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
+        tran_high_t t0, t1, t2, t3;                  // needs32
+        tran_high_t x0, x1, x2, x3;                  // canbe16
+
+        // stage 1
+        s0 = input[0] + input[7];
+        s1 = input[1] + input[6];
+        s2 = input[2] + input[5];
+        s3 = input[3] + input[4];
+        s4 = input[3] - input[4];
+        s5 = input[2] - input[5];
+        s6 = input[1] - input[6];
+        s7 = input[0] - input[7];
+
+        // fdct4(step, step);
+        x0 = s0 + s3;
+        x1 = s1 + s2;
+        x2 = s1 - s2;
+        x3 = s0 - s3;
+        t0 = (x0 + x1) * cospi_16_64;
+        t1 = (x0 - x1) * cospi_16_64;
+        t2 = x3 * cospi_8_64  + x2 * cospi_24_64;
+        t3 = x3 * cospi_24_64 - x2 * cospi_8_64;
+        out[0] = (tran_low_t)fdct_round_shift(t0);
+        out[4] = (tran_low_t)fdct_round_shift(t2);
+        out[8] = (tran_low_t)fdct_round_shift(t1);
+        out[12] = (tran_low_t)fdct_round_shift(t3);
+
+        // Stage 2
+        t0 = (s6 - s5) * cospi_16_64;
+        t1 = (s6 + s5) * cospi_16_64;
+        t2 = fdct_round_shift(t0);
+        t3 = fdct_round_shift(t1);
+
+        // Stage 3
+        x0 = s4 + t2;
+        x1 = s4 - t2;
+        x2 = s7 - t3;
+        x3 = s7 + t3;
+
+        // Stage 4
+        t0 = x0 * cospi_28_64 + x3 *   cospi_4_64;
+        t1 = x1 * cospi_12_64 + x2 *  cospi_20_64;
+        t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
+        t3 = x3 * cospi_28_64 + x0 *  -cospi_4_64;
+        out[2] = (tran_low_t)fdct_round_shift(t0);
+        out[6] = (tran_low_t)fdct_round_shift(t2);
+        out[10] = (tran_low_t)fdct_round_shift(t1);
+        out[14] = (tran_low_t)fdct_round_shift(t3);
+      }
+      // Work on the next eight values; step1 -> odd_results
+      {
+        // step 2
+        temp1 = (step1[5] - step1[2]) * cospi_16_64;
+        temp2 = (step1[4] - step1[3]) * cospi_16_64;
+        step2[2] = fdct_round_shift(temp1);
+        step2[3] = fdct_round_shift(temp2);
+        temp1 = (step1[4] + step1[3]) * cospi_16_64;
+        temp2 = (step1[5] + step1[2]) * cospi_16_64;
+        step2[4] = fdct_round_shift(temp1);
+        step2[5] = fdct_round_shift(temp2);
+        // step 3
+        step3[0] = step1[0] + step2[3];
+        step3[1] = step1[1] + step2[2];
+        step3[2] = step1[1] - step2[2];
+        step3[3] = step1[0] - step2[3];
+        step3[4] = step1[7] - step2[4];
+        step3[5] = step1[6] - step2[5];
+        step3[6] = step1[6] + step2[5];
+        step3[7] = step1[7] + step2[4];
+        // step 4
+        temp1 = step3[1] *  -cospi_8_64 + step3[6] * cospi_24_64;
+        temp2 = step3[2] * cospi_24_64 + step3[5] *  cospi_8_64;
+        step2[1] = fdct_round_shift(temp1);
+        step2[2] = fdct_round_shift(temp2);
+        temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
+        temp2 = step3[1] * cospi_24_64 + step3[6] *  cospi_8_64;
+        step2[5] = fdct_round_shift(temp1);
+        step2[6] = fdct_round_shift(temp2);
+        // step 5
+        step1[0] = step3[0] + step2[1];
+        step1[1] = step3[0] - step2[1];
+        step1[2] = step3[3] + step2[2];
+        step1[3] = step3[3] - step2[2];
+        step1[4] = step3[4] - step2[5];
+        step1[5] = step3[4] + step2[5];
+        step1[6] = step3[7] - step2[6];
+        step1[7] = step3[7] + step2[6];
+        // step 6
+        temp1 = step1[0] * cospi_30_64 + step1[7] *  cospi_2_64;
+        temp2 = step1[1] * cospi_14_64 + step1[6] * cospi_18_64;
+        out[1] = (tran_low_t)fdct_round_shift(temp1);
+        out[9] = (tran_low_t)fdct_round_shift(temp2);
+        temp1 = step1[2] * cospi_22_64 + step1[5] * cospi_10_64;
+        temp2 = step1[3] *  cospi_6_64 + step1[4] * cospi_26_64;
+        out[5] = (tran_low_t)fdct_round_shift(temp1);
+        out[13] = (tran_low_t)fdct_round_shift(temp2);
+        temp1 = step1[3] * -cospi_26_64 + step1[4] *  cospi_6_64;
+        temp2 = step1[2] * -cospi_10_64 + step1[5] * cospi_22_64;
+        out[3] = (tran_low_t)fdct_round_shift(temp1);
+        out[11] = (tran_low_t)fdct_round_shift(temp2);
+        temp1 = step1[1] * -cospi_18_64 + step1[6] * cospi_14_64;
+        temp2 = step1[0] *  -cospi_2_64 + step1[7] * cospi_30_64;
+        out[7] = (tran_low_t)fdct_round_shift(temp1);
+        out[15] = (tran_low_t)fdct_round_shift(temp2);
+      }
+      // Do next column (which is a transposed row in second/horizontal pass)
+      in++;
+      in_pass0++;
+      out += 16;
+    }
+    // Setup in/out for next pass.
+    in = intermediate;
+    out = output;
+  }
+}
+
+void vpx_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  int sum = 0;
+  for (r = 0; r < 16; ++r)
+    for (c = 0; c < 16; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = (tran_low_t)(sum >> 1);
+}
+
+static INLINE tran_high_t dct_32_round(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  // TODO(debargha, peter.derivaz): Find new bounds for this assert,
+  // and make the bounds consts.
+  // assert(-131072 <= rv && rv <= 131071);
+  return rv;
+}
+
+static INLINE tran_high_t half_round_shift(tran_high_t input) {
+  tran_high_t rv = (input + 1 + (input < 0)) >> 2;
+  return rv;
+}
+
+void vpx_fdct32(const tran_high_t *input, tran_high_t *output, int round) {
+  tran_high_t step[32];
+  // Stage 1
+  step[0] = input[0] + input[(32 - 1)];
+  step[1] = input[1] + input[(32 - 2)];
+  step[2] = input[2] + input[(32 - 3)];
+  step[3] = input[3] + input[(32 - 4)];
+  step[4] = input[4] + input[(32 - 5)];
+  step[5] = input[5] + input[(32 - 6)];
+  step[6] = input[6] + input[(32 - 7)];
+  step[7] = input[7] + input[(32 - 8)];
+  step[8] = input[8] + input[(32 - 9)];
+  step[9] = input[9] + input[(32 - 10)];
+  step[10] = input[10] + input[(32 - 11)];
+  step[11] = input[11] + input[(32 - 12)];
+  step[12] = input[12] + input[(32 - 13)];
+  step[13] = input[13] + input[(32 - 14)];
+  step[14] = input[14] + input[(32 - 15)];
+  step[15] = input[15] + input[(32 - 16)];
+  step[16] = -input[16] + input[(32 - 17)];
+  step[17] = -input[17] + input[(32 - 18)];
+  step[18] = -input[18] + input[(32 - 19)];
+  step[19] = -input[19] + input[(32 - 20)];
+  step[20] = -input[20] + input[(32 - 21)];
+  step[21] = -input[21] + input[(32 - 22)];
+  step[22] = -input[22] + input[(32 - 23)];
+  step[23] = -input[23] + input[(32 - 24)];
+  step[24] = -input[24] + input[(32 - 25)];
+  step[25] = -input[25] + input[(32 - 26)];
+  step[26] = -input[26] + input[(32 - 27)];
+  step[27] = -input[27] + input[(32 - 28)];
+  step[28] = -input[28] + input[(32 - 29)];
+  step[29] = -input[29] + input[(32 - 30)];
+  step[30] = -input[30] + input[(32 - 31)];
+  step[31] = -input[31] + input[(32 - 32)];
+
+  // Stage 2
+  output[0] = step[0] + step[16 - 1];
+  output[1] = step[1] + step[16 - 2];
+  output[2] = step[2] + step[16 - 3];
+  output[3] = step[3] + step[16 - 4];
+  output[4] = step[4] + step[16 - 5];
+  output[5] = step[5] + step[16 - 6];
+  output[6] = step[6] + step[16 - 7];
+  output[7] = step[7] + step[16 - 8];
+  output[8] = -step[8] + step[16 - 9];
+  output[9] = -step[9] + step[16 - 10];
+  output[10] = -step[10] + step[16 - 11];
+  output[11] = -step[11] + step[16 - 12];
+  output[12] = -step[12] + step[16 - 13];
+  output[13] = -step[13] + step[16 - 14];
+  output[14] = -step[14] + step[16 - 15];
+  output[15] = -step[15] + step[16 - 16];
+
+  output[16] = step[16];
+  output[17] = step[17];
+  output[18] = step[18];
+  output[19] = step[19];
+
+  output[20] = dct_32_round((-step[20] + step[27]) * cospi_16_64);
+  output[21] = dct_32_round((-step[21] + step[26]) * cospi_16_64);
+  output[22] = dct_32_round((-step[22] + step[25]) * cospi_16_64);
+  output[23] = dct_32_round((-step[23] + step[24]) * cospi_16_64);
+
+  output[24] = dct_32_round((step[24] + step[23]) * cospi_16_64);
+  output[25] = dct_32_round((step[25] + step[22]) * cospi_16_64);
+  output[26] = dct_32_round((step[26] + step[21]) * cospi_16_64);
+  output[27] = dct_32_round((step[27] + step[20]) * cospi_16_64);
+
+  output[28] = step[28];
+  output[29] = step[29];
+  output[30] = step[30];
+  output[31] = step[31];
+
+  // dump the magnitude by 4, hence the intermediate values are within
+  // the range of 16 bits.
+  if (round) {
+    output[0] = half_round_shift(output[0]);
+    output[1] = half_round_shift(output[1]);
+    output[2] = half_round_shift(output[2]);
+    output[3] = half_round_shift(output[3]);
+    output[4] = half_round_shift(output[4]);
+    output[5] = half_round_shift(output[5]);
+    output[6] = half_round_shift(output[6]);
+    output[7] = half_round_shift(output[7]);
+    output[8] = half_round_shift(output[8]);
+    output[9] = half_round_shift(output[9]);
+    output[10] = half_round_shift(output[10]);
+    output[11] = half_round_shift(output[11]);
+    output[12] = half_round_shift(output[12]);
+    output[13] = half_round_shift(output[13]);
+    output[14] = half_round_shift(output[14]);
+    output[15] = half_round_shift(output[15]);
+
+    output[16] = half_round_shift(output[16]);
+    output[17] = half_round_shift(output[17]);
+    output[18] = half_round_shift(output[18]);
+    output[19] = half_round_shift(output[19]);
+    output[20] = half_round_shift(output[20]);
+    output[21] = half_round_shift(output[21]);
+    output[22] = half_round_shift(output[22]);
+    output[23] = half_round_shift(output[23]);
+    output[24] = half_round_shift(output[24]);
+    output[25] = half_round_shift(output[25]);
+    output[26] = half_round_shift(output[26]);
+    output[27] = half_round_shift(output[27]);
+    output[28] = half_round_shift(output[28]);
+    output[29] = half_round_shift(output[29]);
+    output[30] = half_round_shift(output[30]);
+    output[31] = half_round_shift(output[31]);
+  }
+
+  // Stage 3
+  step[0] = output[0] + output[(8 - 1)];
+  step[1] = output[1] + output[(8 - 2)];
+  step[2] = output[2] + output[(8 - 3)];
+  step[3] = output[3] + output[(8 - 4)];
+  step[4] = -output[4] + output[(8 - 5)];
+  step[5] = -output[5] + output[(8 - 6)];
+  step[6] = -output[6] + output[(8 - 7)];
+  step[7] = -output[7] + output[(8 - 8)];
+  step[8] = output[8];
+  step[9] = output[9];
+  step[10] = dct_32_round((-output[10] + output[13]) * cospi_16_64);
+  step[11] = dct_32_round((-output[11] + output[12]) * cospi_16_64);
+  step[12] = dct_32_round((output[12] + output[11]) * cospi_16_64);
+  step[13] = dct_32_round((output[13] + output[10]) * cospi_16_64);
+  step[14] = output[14];
+  step[15] = output[15];
+
+  step[16] = output[16] + output[23];
+  step[17] = output[17] + output[22];
+  step[18] = output[18] + output[21];
+  step[19] = output[19] + output[20];
+  step[20] = -output[20] + output[19];
+  step[21] = -output[21] + output[18];
+  step[22] = -output[22] + output[17];
+  step[23] = -output[23] + output[16];
+  step[24] = -output[24] + output[31];
+  step[25] = -output[25] + output[30];
+  step[26] = -output[26] + output[29];
+  step[27] = -output[27] + output[28];
+  step[28] = output[28] + output[27];
+  step[29] = output[29] + output[26];
+  step[30] = output[30] + output[25];
+  step[31] = output[31] + output[24];
+
+  // Stage 4
+  output[0] = step[0] + step[3];
+  output[1] = step[1] + step[2];
+  output[2] = -step[2] + step[1];
+  output[3] = -step[3] + step[0];
+  output[4] = step[4];
+  output[5] = dct_32_round((-step[5] + step[6]) * cospi_16_64);
+  output[6] = dct_32_round((step[6] + step[5]) * cospi_16_64);
+  output[7] = step[7];
+  output[8] = step[8] + step[11];
+  output[9] = step[9] + step[10];
+  output[10] = -step[10] + step[9];
+  output[11] = -step[11] + step[8];
+  output[12] = -step[12] + step[15];
+  output[13] = -step[13] + step[14];
+  output[14] = step[14] + step[13];
+  output[15] = step[15] + step[12];
+
+  output[16] = step[16];
+  output[17] = step[17];
+  output[18] = dct_32_round(step[18] * -cospi_8_64 + step[29] * cospi_24_64);
+  output[19] = dct_32_round(step[19] * -cospi_8_64 + step[28] * cospi_24_64);
+  output[20] = dct_32_round(step[20] * -cospi_24_64 + step[27] * -cospi_8_64);
+  output[21] = dct_32_round(step[21] * -cospi_24_64 + step[26] * -cospi_8_64);
+  output[22] = step[22];
+  output[23] = step[23];
+  output[24] = step[24];
+  output[25] = step[25];
+  output[26] = dct_32_round(step[26] * cospi_24_64 + step[21] * -cospi_8_64);
+  output[27] = dct_32_round(step[27] * cospi_24_64 + step[20] * -cospi_8_64);
+  output[28] = dct_32_round(step[28] * cospi_8_64 + step[19] * cospi_24_64);
+  output[29] = dct_32_round(step[29] * cospi_8_64 + step[18] * cospi_24_64);
+  output[30] = step[30];
+  output[31] = step[31];
+
+  // Stage 5
+  step[0] = dct_32_round((output[0] + output[1]) * cospi_16_64);
+  step[1] = dct_32_round((-output[1] + output[0]) * cospi_16_64);
+  step[2] = dct_32_round(output[2] * cospi_24_64 + output[3] * cospi_8_64);
+  step[3] = dct_32_round(output[3] * cospi_24_64 - output[2] * cospi_8_64);
+  step[4] = output[4] + output[5];
+  step[5] = -output[5] + output[4];
+  step[6] = -output[6] + output[7];
+  step[7] = output[7] + output[6];
+  step[8] = output[8];
+  step[9] = dct_32_round(output[9] * -cospi_8_64 + output[14] * cospi_24_64);
+  step[10] = dct_32_round(output[10] * -cospi_24_64 + output[13] * -cospi_8_64);
+  step[11] = output[11];
+  step[12] = output[12];
+  step[13] = dct_32_round(output[13] * cospi_24_64 + output[10] * -cospi_8_64);
+  step[14] = dct_32_round(output[14] * cospi_8_64 + output[9] * cospi_24_64);
+  step[15] = output[15];
+
+  step[16] = output[16] + output[19];
+  step[17] = output[17] + output[18];
+  step[18] = -output[18] + output[17];
+  step[19] = -output[19] + output[16];
+  step[20] = -output[20] + output[23];
+  step[21] = -output[21] + output[22];
+  step[22] = output[22] + output[21];
+  step[23] = output[23] + output[20];
+  step[24] = output[24] + output[27];
+  step[25] = output[25] + output[26];
+  step[26] = -output[26] + output[25];
+  step[27] = -output[27] + output[24];
+  step[28] = -output[28] + output[31];
+  step[29] = -output[29] + output[30];
+  step[30] = output[30] + output[29];
+  step[31] = output[31] + output[28];
+
+  // Stage 6
+  output[0] = step[0];
+  output[1] = step[1];
+  output[2] = step[2];
+  output[3] = step[3];
+  output[4] = dct_32_round(step[4] * cospi_28_64 + step[7] * cospi_4_64);
+  output[5] = dct_32_round(step[5] * cospi_12_64 + step[6] * cospi_20_64);
+  output[6] = dct_32_round(step[6] * cospi_12_64 + step[5] * -cospi_20_64);
+  output[7] = dct_32_round(step[7] * cospi_28_64 + step[4] * -cospi_4_64);
+  output[8] = step[8] + step[9];
+  output[9] = -step[9] + step[8];
+  output[10] = -step[10] + step[11];
+  output[11] = step[11] + step[10];
+  output[12] = step[12] + step[13];
+  output[13] = -step[13] + step[12];
+  output[14] = -step[14] + step[15];
+  output[15] = step[15] + step[14];
+
+  output[16] = step[16];
+  output[17] = dct_32_round(step[17] * -cospi_4_64 + step[30] * cospi_28_64);
+  output[18] = dct_32_round(step[18] * -cospi_28_64 + step[29] * -cospi_4_64);
+  output[19] = step[19];
+  output[20] = step[20];
+  output[21] = dct_32_round(step[21] * -cospi_20_64 + step[26] * cospi_12_64);
+  output[22] = dct_32_round(step[22] * -cospi_12_64 + step[25] * -cospi_20_64);
+  output[23] = step[23];
+  output[24] = step[24];
+  output[25] = dct_32_round(step[25] * cospi_12_64 + step[22] * -cospi_20_64);
+  output[26] = dct_32_round(step[26] * cospi_20_64 + step[21] * cospi_12_64);
+  output[27] = step[27];
+  output[28] = step[28];
+  output[29] = dct_32_round(step[29] * cospi_28_64 + step[18] * -cospi_4_64);
+  output[30] = dct_32_round(step[30] * cospi_4_64 + step[17] * cospi_28_64);
+  output[31] = step[31];
+
+  // Stage 7
+  step[0] = output[0];
+  step[1] = output[1];
+  step[2] = output[2];
+  step[3] = output[3];
+  step[4] = output[4];
+  step[5] = output[5];
+  step[6] = output[6];
+  step[7] = output[7];
+  step[8] = dct_32_round(output[8] * cospi_30_64 + output[15] * cospi_2_64);
+  step[9] = dct_32_round(output[9] * cospi_14_64 + output[14] * cospi_18_64);
+  step[10] = dct_32_round(output[10] * cospi_22_64 + output[13] * cospi_10_64);
+  step[11] = dct_32_round(output[11] * cospi_6_64 + output[12] * cospi_26_64);
+  step[12] = dct_32_round(output[12] * cospi_6_64 + output[11] * -cospi_26_64);
+  step[13] = dct_32_round(output[13] * cospi_22_64 + output[10] * -cospi_10_64);
+  step[14] = dct_32_round(output[14] * cospi_14_64 + output[9] * -cospi_18_64);
+  step[15] = dct_32_round(output[15] * cospi_30_64 + output[8] * -cospi_2_64);
+
+  step[16] = output[16] + output[17];
+  step[17] = -output[17] + output[16];
+  step[18] = -output[18] + output[19];
+  step[19] = output[19] + output[18];
+  step[20] = output[20] + output[21];
+  step[21] = -output[21] + output[20];
+  step[22] = -output[22] + output[23];
+  step[23] = output[23] + output[22];
+  step[24] = output[24] + output[25];
+  step[25] = -output[25] + output[24];
+  step[26] = -output[26] + output[27];
+  step[27] = output[27] + output[26];
+  step[28] = output[28] + output[29];
+  step[29] = -output[29] + output[28];
+  step[30] = -output[30] + output[31];
+  step[31] = output[31] + output[30];
+
+  // Final stage --- outputs indices are bit-reversed.
+  output[0]  = step[0];
+  output[16] = step[1];
+  output[8]  = step[2];
+  output[24] = step[3];
+  output[4]  = step[4];
+  output[20] = step[5];
+  output[12] = step[6];
+  output[28] = step[7];
+  output[2]  = step[8];
+  output[18] = step[9];
+  output[10] = step[10];
+  output[26] = step[11];
+  output[6]  = step[12];
+  output[22] = step[13];
+  output[14] = step[14];
+  output[30] = step[15];
+
+  output[1]  = dct_32_round(step[16] * cospi_31_64 + step[31] * cospi_1_64);
+  output[17] = dct_32_round(step[17] * cospi_15_64 + step[30] * cospi_17_64);
+  output[9]  = dct_32_round(step[18] * cospi_23_64 + step[29] * cospi_9_64);
+  output[25] = dct_32_round(step[19] * cospi_7_64 + step[28] * cospi_25_64);
+  output[5]  = dct_32_round(step[20] * cospi_27_64 + step[27] * cospi_5_64);
+  output[21] = dct_32_round(step[21] * cospi_11_64 + step[26] * cospi_21_64);
+  output[13] = dct_32_round(step[22] * cospi_19_64 + step[25] * cospi_13_64);
+  output[29] = dct_32_round(step[23] * cospi_3_64 + step[24] * cospi_29_64);
+  output[3]  = dct_32_round(step[24] * cospi_3_64 + step[23] * -cospi_29_64);
+  output[19] = dct_32_round(step[25] * cospi_19_64 + step[22] * -cospi_13_64);
+  output[11] = dct_32_round(step[26] * cospi_11_64 + step[21] * -cospi_21_64);
+  output[27] = dct_32_round(step[27] * cospi_27_64 + step[20] * -cospi_5_64);
+  output[7]  = dct_32_round(step[28] * cospi_7_64 + step[19] * -cospi_25_64);
+  output[23] = dct_32_round(step[29] * cospi_23_64 + step[18] * -cospi_9_64);
+  output[15] = dct_32_round(step[30] * cospi_15_64 + step[17] * -cospi_17_64);
+  output[31] = dct_32_round(step[31] * cospi_31_64 + step[16] * -cospi_1_64);
+}
+
+void vpx_fdct32x32_c(const int16_t *input, tran_low_t *out, int stride) {
+  int i, j;
+  tran_high_t output[32 * 32];
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = input[j * stride + i] * 4;
+    vpx_fdct32(temp_in, temp_out, 0);
+    for (j = 0; j < 32; ++j)
+      output[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
+  }
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = output[j + i * 32];
+    vpx_fdct32(temp_in, temp_out, 0);
+    for (j = 0; j < 32; ++j)
+      out[j + i * 32] =
+          (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2);
+  }
+}
+
+// Note that although we use dct_32_round in dct32 computation flow,
+// this 2d fdct32x32 for rate-distortion optimization loop is operating
+// within 16 bits precision.
+void vpx_fdct32x32_rd_c(const int16_t *input, tran_low_t *out, int stride) {
+  int i, j;
+  tran_high_t output[32 * 32];
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = input[j * stride + i] * 4;
+    vpx_fdct32(temp_in, temp_out, 0);
+    for (j = 0; j < 32; ++j)
+      // TODO(cd): see quality impact of only doing
+      //           output[j * 32 + i] = (temp_out[j] + 1) >> 2;
+      //           PS: also change code in vpx_dsp/x86/vpx_dct_sse2.c
+      output[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
+  }
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = output[j + i * 32];
+    vpx_fdct32(temp_in, temp_out, 1);
+    for (j = 0; j < 32; ++j)
+      out[j + i * 32] = (tran_low_t)temp_out[j];
+  }
+}
+
+void vpx_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  int sum = 0;
+  for (r = 0; r < 32; ++r)
+    for (c = 0; c < 32; ++c)
+      sum += input[r * stride + c];
+
+  output[0] = (tran_low_t)(sum >> 3);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_fdct4x4_c(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  vpx_fdct4x4_c(input, output, stride);
+}
+
+void vpx_highbd_fdct8x8_c(const int16_t *input, tran_low_t *final_output,
+                          int stride) {
+  vpx_fdct8x8_c(input, final_output, stride);
+}
+
+void vpx_highbd_fdct8x8_1_c(const int16_t *input, tran_low_t *final_output,
+                            int stride) {
+  vpx_fdct8x8_1_c(input, final_output, stride);
+}
+
+void vpx_highbd_fdct16x16_c(const int16_t *input, tran_low_t *output,
+                            int stride) {
+  vpx_fdct16x16_c(input, output, stride);
+}
+
+void vpx_highbd_fdct16x16_1_c(const int16_t *input, tran_low_t *output,
+                              int stride) {
+  vpx_fdct16x16_1_c(input, output, stride);
+}
+
+void vpx_highbd_fdct32x32_c(const int16_t *input, tran_low_t *out, int stride) {
+  vpx_fdct32x32_c(input, out, stride);
+}
+
+void vpx_highbd_fdct32x32_rd_c(const int16_t *input, tran_low_t *out,
+                               int stride) {
+  vpx_fdct32x32_rd_c(input, out, stride);
+}
+
+void vpx_highbd_fdct32x32_1_c(const int16_t *input, tran_low_t *out,
+                              int stride) {
+  vpx_fdct32x32_1_c(input, out, stride);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vpx_dsp/fwd_txfm.h b/libvpx/libvpx/vpx_dsp/fwd_txfm.h
new file mode 100644
index 0000000..29e139c
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/fwd_txfm.h
@@ -0,0 +1,25 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_FWD_TXFM_H_
+#define VPX_DSP_FWD_TXFM_H_
+
+#include "vpx_dsp/txfm_common.h"
+
+static INLINE tran_high_t fdct_round_shift(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  // TODO(debargha, peter.derivaz): Find new bounds for this assert
+  // and make the bounds consts.
+  // assert(INT16_MIN <= rv && rv <= INT16_MAX);
+  return rv;
+}
+
+void vpx_fdct32(const tran_high_t *input, tran_high_t *output, int round);
+#endif  // VPX_DSP_FWD_TXFM_H_
diff --git a/libvpx/libvpx/vpx_dsp/intrapred.c b/libvpx/libvpx/vpx_dsp/intrapred.c
new file mode 100644
index 0000000..cc4a74b
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/intrapred.c
@@ -0,0 +1,870 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+
+#define DST(x, y) dst[(x) + (y) * stride]
+#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
+#define AVG2(a, b) (((a) + (b) + 1) >> 1)
+
+static INLINE void d207_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  (void) above;
+  // first column
+  for (r = 0; r < bs - 1; ++r)
+    dst[r * stride] = AVG2(left[r], left[r + 1]);
+  dst[(bs - 1) * stride] = left[bs - 1];
+  dst++;
+
+  // second column
+  for (r = 0; r < bs - 2; ++r)
+    dst[r * stride] = AVG3(left[r], left[r + 1], left[r + 2]);
+  dst[(bs - 2) * stride] = AVG3(left[bs - 2], left[bs - 1], left[bs - 1]);
+  dst[(bs - 1) * stride] = left[bs - 1];
+  dst++;
+
+  // rest of last row
+  for (c = 0; c < bs - 2; ++c)
+    dst[(bs - 1) * stride + c] = left[bs - 1];
+
+  for (r = bs - 2; r >= 0; --r)
+    for (c = 0; c < bs - 2; ++c)
+      dst[r * stride + c] = dst[(r + 1) * stride + c - 2];
+}
+
+#if CONFIG_MISC_FIXES
+static INLINE void d207e_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                   const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  (void) above;
+
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = c & 1 ? AVG3(left[(c >> 1) + r], left[(c >> 1) + r + 1],
+                            left[(c >> 1) + r + 2])
+          : AVG2(left[(c >> 1) + r], left[(c >> 1) + r + 1]);
+    }
+    dst += stride;
+  }
+}
+#endif  // CONFIG_MISC_FIXES
+
+static INLINE void d63_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                 const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  int size;
+  (void)left;
+  for (c = 0; c < bs; ++c) {
+    dst[c] = AVG2(above[c], above[c + 1]);
+    dst[stride + c] = AVG3(above[c], above[c + 1], above[c + 2]);
+  }
+  for (r = 2, size = bs - 2; r < bs; r += 2, --size) {
+    memcpy(dst + (r + 0) * stride, dst + (r >> 1), size);
+    memset(dst + (r + 0) * stride + size, above[bs - 1], bs - size);
+    memcpy(dst + (r + 1) * stride, dst + stride + (r >> 1), size);
+    memset(dst + (r + 1) * stride + size, above[bs - 1], bs - size);
+  }
+}
+
+#if CONFIG_MISC_FIXES
+static INLINE void d63e_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  (void) left;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = r & 1 ? AVG3(above[(r >> 1) + c], above[(r >> 1) + c + 1],
+                            above[(r >> 1) + c + 2])
+          : AVG2(above[(r >> 1) + c], above[(r >> 1) + c + 1]);
+    }
+    dst += stride;
+  }
+}
+#endif  // CONFIG_MISC_FIXES
+
+static INLINE void d45_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                 const uint8_t *above, const uint8_t *left) {
+  const uint8_t above_right = above[bs - 1];
+  const uint8_t *const dst_row0 = dst;
+  int x, size;
+  (void)left;
+
+  for (x = 0; x < bs - 1; ++x) {
+    dst[x] = AVG3(above[x], above[x + 1], above[x + 2]);
+  }
+  dst[bs - 1] = above_right;
+  dst += stride;
+  for (x = 1, size = bs - 2; x < bs; ++x, --size) {
+    memcpy(dst, dst_row0 + x, size);
+    memset(dst + size, above_right, x + 1);
+    dst += stride;
+  }
+}
+
+#if CONFIG_MISC_FIXES
+static INLINE void d45e_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  (void) left;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = AVG3(above[r + c], above[r + c + 1],
+                    above[r + c + 1 + (r + c + 2 < bs * 2)]);
+    }
+    dst += stride;
+  }
+}
+#endif  // CONFIG_MISC_FIXES
+
+static INLINE void d117_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+
+  // first row
+  for (c = 0; c < bs; c++)
+    dst[c] = AVG2(above[c - 1], above[c]);
+  dst += stride;
+
+  // second row
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  for (c = 1; c < bs; c++)
+    dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);
+  dst += stride;
+
+  // the rest of first col
+  dst[0] = AVG3(above[-1], left[0], left[1]);
+  for (r = 3; r < bs; ++r)
+    dst[(r - 2) * stride] = AVG3(left[r - 3], left[r - 2], left[r - 1]);
+
+  // the rest of the block
+  for (r = 2; r < bs; ++r) {
+    for (c = 1; c < bs; c++)
+      dst[c] = dst[-2 * stride + c - 1];
+    dst += stride;
+  }
+}
+
+static INLINE void d135_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int i;
+#if defined(__GNUC__) && __GNUC__ == 4 && __GNUC_MINOR__ > 7
+  // silence a spurious -Warray-bounds warning, possibly related to:
+  // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56273
+  uint8_t border[69];
+#else
+  uint8_t border[32 + 32 - 1];  // outer border from bottom-left to top-right
+#endif
+
+  // dst(bs, bs - 2)[0], i.e., border starting at bottom-left
+  for (i = 0; i < bs - 2; ++i) {
+    border[i] = AVG3(left[bs - 3 - i], left[bs - 2 - i], left[bs - 1 - i]);
+  }
+  border[bs - 2] = AVG3(above[-1], left[0], left[1]);
+  border[bs - 1] = AVG3(left[0], above[-1], above[0]);
+  border[bs - 0] = AVG3(above[-1], above[0], above[1]);
+  // dst[0][2, size), i.e., remaining top border ascending
+  for (i = 0; i < bs - 2; ++i) {
+    border[bs + 1 + i] = AVG3(above[i], above[i + 1], above[i + 2]);
+  }
+
+  for (i = 0; i < bs; ++i) {
+    memcpy(dst + i * stride, border + bs - 1 - i, bs);
+  }
+}
+
+static INLINE void d153_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  dst[0] = AVG2(above[-1], left[0]);
+  for (r = 1; r < bs; r++)
+    dst[r * stride] = AVG2(left[r - 1], left[r]);
+  dst++;
+
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  dst[stride] = AVG3(above[-1], left[0], left[1]);
+  for (r = 2; r < bs; r++)
+    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);
+  dst++;
+
+  for (c = 0; c < bs - 2; c++)
+    dst[c] = AVG3(above[c - 1], above[c], above[c + 1]);
+  dst += stride;
+
+  for (r = 1; r < bs; ++r) {
+    for (c = 0; c < bs - 2; c++)
+      dst[c] = dst[-stride + c - 2];
+    dst += stride;
+  }
+}
+
+static INLINE void v_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                               const uint8_t *above, const uint8_t *left) {
+  int r;
+  (void) left;
+
+  for (r = 0; r < bs; r++) {
+    memcpy(dst, above, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void h_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                               const uint8_t *above, const uint8_t *left) {
+  int r;
+  (void) above;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, left[r], bs);
+    dst += stride;
+  }
+}
+
+static INLINE void tm_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  int ytop_left = above[-1];
+
+  for (r = 0; r < bs; r++) {
+    for (c = 0; c < bs; c++)
+      dst[c] = clip_pixel(left[r] + above[c] - ytop_left);
+    dst += stride;
+  }
+}
+
+static INLINE void dc_128_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                    const uint8_t *above, const uint8_t *left) {
+  int r;
+  (void) above;
+  (void) left;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, 128, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void dc_left_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  int i, r, expected_dc, sum = 0;
+  (void) above;
+
+  for (i = 0; i < bs; i++)
+    sum += left[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void dc_top_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                    const uint8_t *above, const uint8_t *left) {
+  int i, r, expected_dc, sum = 0;
+  (void) left;
+
+  for (i = 0; i < bs; i++)
+    sum += above[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void dc_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                const uint8_t *above, const uint8_t *left) {
+  int i, r, expected_dc, sum = 0;
+  const int count = 2 * bs;
+
+  for (i = 0; i < bs; i++) {
+    sum += above[i];
+    sum += left[i];
+  }
+
+  expected_dc = (sum + (count >> 1)) / count;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+void vpx_he_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left) {
+  const int H = above[-1];
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int L = left[3];
+
+  memset(dst + stride * 0, AVG3(H, I, J), 4);
+  memset(dst + stride * 1, AVG3(I, J, K), 4);
+  memset(dst + stride * 2, AVG3(J, K, L), 4);
+  memset(dst + stride * 3, AVG3(K, L, L), 4);
+}
+
+void vpx_ve_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left) {
+  const int H = above[-1];
+  const int I = above[0];
+  const int J = above[1];
+  const int K = above[2];
+  const int L = above[3];
+  const int M = above[4];
+  (void)left;
+
+  dst[0] = AVG3(H, I, J);
+  dst[1] = AVG3(I, J, K);
+  dst[2] = AVG3(J, K, L);
+  dst[3] = AVG3(K, L, M);
+  memcpy(dst + stride * 1, dst, 4);
+  memcpy(dst + stride * 2, dst, 4);
+  memcpy(dst + stride * 3, dst, 4);
+}
+
+void vpx_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int L = left[3];
+  (void)above;
+  DST(0, 0) =             AVG2(I, J);
+  DST(2, 0) = DST(0, 1) = AVG2(J, K);
+  DST(2, 1) = DST(0, 2) = AVG2(K, L);
+  DST(1, 0) =             AVG3(I, J, K);
+  DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
+  DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
+  DST(3, 2) = DST(2, 2) =
+      DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
+}
+
+void vpx_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                             const uint8_t *above, const uint8_t *left) {
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  const int E = above[4];
+  const int F = above[5];
+  const int G = above[6];
+  (void)left;
+  DST(0, 0) =             AVG2(A, B);
+  DST(1, 0) = DST(0, 2) = AVG2(B, C);
+  DST(2, 0) = DST(1, 2) = AVG2(C, D);
+  DST(3, 0) = DST(2, 2) = AVG2(D, E);
+              DST(3, 2) = AVG2(E, F);  // differs from vp8
+
+  DST(0, 1) =             AVG3(A, B, C);
+  DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
+  DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
+  DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
+              DST(3, 3) = AVG3(E, F, G);  // differs from vp8
+}
+
+void vpx_d63f_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  const int E = above[4];
+  const int F = above[5];
+  const int G = above[6];
+  const int H = above[7];
+  (void)left;
+  DST(0, 0) =             AVG2(A, B);
+  DST(1, 0) = DST(0, 2) = AVG2(B, C);
+  DST(2, 0) = DST(1, 2) = AVG2(C, D);
+  DST(3, 0) = DST(2, 2) = AVG2(D, E);
+              DST(3, 2) = AVG3(E, F, G);
+
+  DST(0, 1) =             AVG3(A, B, C);
+  DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
+  DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
+  DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
+              DST(3, 3) = AVG3(F, G, H);
+}
+
+void vpx_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                             const uint8_t *above, const uint8_t *left) {
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  const int E = above[4];
+  const int F = above[5];
+  const int G = above[6];
+  const int H = above[7];
+  (void)stride;
+  (void)left;
+  DST(0, 0)                                     = AVG3(A, B, C);
+  DST(1, 0) = DST(0, 1)                         = AVG3(B, C, D);
+  DST(2, 0) = DST(1, 1) = DST(0, 2)             = AVG3(C, D, E);
+  DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
+              DST(3, 1) = DST(2, 2) = DST(1, 3) = AVG3(E, F, G);
+                          DST(3, 2) = DST(2, 3) = AVG3(F, G, H);
+                                      DST(3, 3) = H;  // differs from vp8
+}
+
+void vpx_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  const int E = above[4];
+  const int F = above[5];
+  const int G = above[6];
+  const int H = above[7];
+  (void)stride;
+  (void)left;
+  DST(0, 0)                                     = AVG3(A, B, C);
+  DST(1, 0) = DST(0, 1)                         = AVG3(B, C, D);
+  DST(2, 0) = DST(1, 1) = DST(0, 2)             = AVG3(C, D, E);
+  DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
+              DST(3, 1) = DST(2, 2) = DST(1, 3) = AVG3(E, F, G);
+                          DST(3, 2) = DST(2, 3) = AVG3(F, G, H);
+                                      DST(3, 3) = AVG3(G, H, H);
+}
+
+void vpx_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  DST(0, 0) = DST(1, 2) = AVG2(X, A);
+  DST(1, 0) = DST(2, 2) = AVG2(A, B);
+  DST(2, 0) = DST(3, 2) = AVG2(B, C);
+  DST(3, 0)             = AVG2(C, D);
+
+  DST(0, 3) =             AVG3(K, J, I);
+  DST(0, 2) =             AVG3(J, I, X);
+  DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
+  DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
+  DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
+  DST(3, 1) =             AVG3(B, C, D);
+}
+
+void vpx_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int L = left[3];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  (void)stride;
+  DST(0, 3)                                     = AVG3(J, K, L);
+  DST(1, 3) = DST(0, 2)                         = AVG3(I, J, K);
+  DST(2, 3) = DST(1, 2) = DST(0, 1)             = AVG3(X, I, J);
+  DST(3, 3) = DST(2, 2) = DST(1, 1) = DST(0, 0) = AVG3(A, X, I);
+              DST(3, 2) = DST(2, 1) = DST(1, 0) = AVG3(B, A, X);
+                          DST(3, 1) = DST(2, 0) = AVG3(C, B, A);
+                                      DST(3, 0) = AVG3(D, C, B);
+}
+
+void vpx_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int L = left[3];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+
+  DST(0, 0) = DST(2, 1) = AVG2(I, X);
+  DST(0, 1) = DST(2, 2) = AVG2(J, I);
+  DST(0, 2) = DST(2, 3) = AVG2(K, J);
+  DST(0, 3)             = AVG2(L, K);
+
+  DST(3, 0)             = AVG3(A, B, C);
+  DST(2, 0)             = AVG3(X, A, B);
+  DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
+  DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
+  DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
+  DST(1, 3)             = AVG3(L, K, J);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE void highbd_d207_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void) above;
+  (void) bd;
+
+  // First column.
+  for (r = 0; r < bs - 1; ++r) {
+    dst[r * stride] = AVG2(left[r], left[r + 1]);
+  }
+  dst[(bs - 1) * stride] = left[bs - 1];
+  dst++;
+
+  // Second column.
+  for (r = 0; r < bs - 2; ++r) {
+    dst[r * stride] = AVG3(left[r], left[r + 1], left[r + 2]);
+  }
+  dst[(bs - 2) * stride] = AVG3(left[bs - 2], left[bs - 1], left[bs - 1]);
+  dst[(bs - 1) * stride] = left[bs - 1];
+  dst++;
+
+  // Rest of last row.
+  for (c = 0; c < bs - 2; ++c)
+    dst[(bs - 1) * stride + c] = left[bs - 1];
+
+  for (r = bs - 2; r >= 0; --r) {
+    for (c = 0; c < bs - 2; ++c)
+      dst[r * stride + c] = dst[(r + 1) * stride + c - 2];
+  }
+}
+
+#if CONFIG_MISC_FIXES
+static INLINE void highbd_d207e_predictor(uint16_t *dst, ptrdiff_t stride,
+                                          int bs, const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  int r, c;
+  (void) above;
+  (void) bd;
+
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = c & 1 ? AVG3(left[(c >> 1) + r], left[(c >> 1) + r + 1],
+                            left[(c >> 1) + r + 2])
+          : AVG2(left[(c >> 1) + r], left[(c >> 1) + r + 1]);
+    }
+    dst += stride;
+  }
+}
+#endif  // CONFIG_MISC_FIXES
+
+static INLINE void highbd_d63_predictor(uint16_t *dst, ptrdiff_t stride,
+                                        int bs, const uint16_t *above,
+                                        const uint16_t *left, int bd) {
+  int r, c;
+  (void) left;
+  (void) bd;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = r & 1 ? AVG3(above[(r >> 1) + c], above[(r >> 1) + c + 1],
+                            above[(r >> 1) + c + 2])
+          : AVG2(above[(r >> 1) + c], above[(r >> 1) + c + 1]);
+    }
+    dst += stride;
+  }
+}
+
+#define highbd_d63e_predictor highbd_d63_predictor
+
+static INLINE void highbd_d45_predictor(uint16_t *dst, ptrdiff_t stride, int bs,
+                                        const uint16_t *above,
+                                        const uint16_t *left, int bd) {
+  int r, c;
+  (void) left;
+  (void) bd;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = r + c + 2 < bs * 2 ? AVG3(above[r + c], above[r + c + 1],
+                                         above[r + c + 2])
+          : above[bs * 2 - 1];
+    }
+    dst += stride;
+  }
+}
+
+#if CONFIG_MISC_FIXES
+static INLINE void highbd_d45e_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void) left;
+  (void) bd;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = AVG3(above[r + c], above[r + c + 1],
+                    above[r + c + 1 + (r + c + 2 < bs * 2)]);
+    }
+    dst += stride;
+  }
+}
+#endif  // CONFIG_MISC_FIXES
+
+static INLINE void highbd_d117_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void) bd;
+
+  // first row
+  for (c = 0; c < bs; c++)
+    dst[c] = AVG2(above[c - 1], above[c]);
+  dst += stride;
+
+  // second row
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  for (c = 1; c < bs; c++)
+    dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);
+  dst += stride;
+
+  // the rest of first col
+  dst[0] = AVG3(above[-1], left[0], left[1]);
+  for (r = 3; r < bs; ++r)
+    dst[(r - 2) * stride] = AVG3(left[r - 3], left[r - 2], left[r - 1]);
+
+  // the rest of the block
+  for (r = 2; r < bs; ++r) {
+    for (c = 1; c < bs; c++)
+      dst[c] = dst[-2 * stride + c - 1];
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d135_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void) bd;
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  for (c = 1; c < bs; c++)
+    dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);
+
+  dst[stride] = AVG3(above[-1], left[0], left[1]);
+  for (r = 2; r < bs; ++r)
+    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);
+
+  dst += stride;
+  for (r = 1; r < bs; ++r) {
+    for (c = 1; c < bs; c++)
+      dst[c] = dst[-stride + c - 1];
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d153_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void) bd;
+  dst[0] = AVG2(above[-1], left[0]);
+  for (r = 1; r < bs; r++)
+    dst[r * stride] = AVG2(left[r - 1], left[r]);
+  dst++;
+
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  dst[stride] = AVG3(above[-1], left[0], left[1]);
+  for (r = 2; r < bs; r++)
+    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);
+  dst++;
+
+  for (c = 0; c < bs - 2; c++)
+    dst[c] = AVG3(above[c - 1], above[c], above[c + 1]);
+  dst += stride;
+
+  for (r = 1; r < bs; ++r) {
+    for (c = 0; c < bs - 2; c++)
+      dst[c] = dst[-stride + c - 2];
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_v_predictor(uint16_t *dst, ptrdiff_t stride,
+                                      int bs, const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  int r;
+  (void) left;
+  (void) bd;
+  for (r = 0; r < bs; r++) {
+    memcpy(dst, above, bs * sizeof(uint16_t));
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_h_predictor(uint16_t *dst, ptrdiff_t stride,
+                                      int bs, const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  int r;
+  (void) above;
+  (void) bd;
+  for (r = 0; r < bs; r++) {
+    vpx_memset16(dst, left[r], bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_tm_predictor(uint16_t *dst, ptrdiff_t stride,
+                                       int bs, const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int r, c;
+  int ytop_left = above[-1];
+  (void) bd;
+
+  for (r = 0; r < bs; r++) {
+    for (c = 0; c < bs; c++)
+      dst[c] = clip_pixel_highbd(left[r] + above[c] - ytop_left, bd);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_128_predictor(uint16_t *dst, ptrdiff_t stride,
+                                           int bs, const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  int r;
+  (void) above;
+  (void) left;
+
+  for (r = 0; r < bs; r++) {
+    vpx_memset16(dst, 128 << (bd - 8), bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_left_predictor(uint16_t *dst, ptrdiff_t stride,
+                                            int bs, const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  int i, r, expected_dc, sum = 0;
+  (void) above;
+  (void) bd;
+
+  for (i = 0; i < bs; i++)
+    sum += left[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    vpx_memset16(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_top_predictor(uint16_t *dst, ptrdiff_t stride,
+                                           int bs, const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  int i, r, expected_dc, sum = 0;
+  (void) left;
+  (void) bd;
+
+  for (i = 0; i < bs; i++)
+    sum += above[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    vpx_memset16(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_predictor(uint16_t *dst, ptrdiff_t stride,
+                                       int bs, const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int i, r, expected_dc, sum = 0;
+  const int count = 2 * bs;
+  (void) bd;
+
+  for (i = 0; i < bs; i++) {
+    sum += above[i];
+    sum += left[i];
+  }
+
+  expected_dc = (sum + (count >> 1)) / count;
+
+  for (r = 0; r < bs; r++) {
+    vpx_memset16(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+// This serves as a wrapper function, so that all the prediction functions
+// can be unified and accessed as a pointer array. Note that the boundary
+// above and left are not necessarily used all the time.
+#define intra_pred_sized(type, size) \
+  void vpx_##type##_predictor_##size##x##size##_c(uint8_t *dst, \
+                                                  ptrdiff_t stride, \
+                                                  const uint8_t *above, \
+                                                  const uint8_t *left) { \
+    type##_predictor(dst, stride, size, above, left); \
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+#define intra_pred_highbd_sized(type, size) \
+  void vpx_highbd_##type##_predictor_##size##x##size##_c( \
+      uint16_t *dst, ptrdiff_t stride, const uint16_t *above, \
+      const uint16_t *left, int bd) { \
+    highbd_##type##_predictor(dst, stride, size, above, left, bd); \
+  }
+
+#define intra_pred_allsizes(type) \
+  intra_pred_sized(type, 4) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32) \
+  intra_pred_highbd_sized(type, 4) \
+  intra_pred_highbd_sized(type, 8) \
+  intra_pred_highbd_sized(type, 16) \
+  intra_pred_highbd_sized(type, 32)
+
+#define intra_pred_no_4x4(type) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32) \
+  intra_pred_highbd_sized(type, 4) \
+  intra_pred_highbd_sized(type, 8) \
+  intra_pred_highbd_sized(type, 16) \
+  intra_pred_highbd_sized(type, 32)
+
+#else
+#define intra_pred_allsizes(type) \
+  intra_pred_sized(type, 4) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32)
+
+#define intra_pred_no_4x4(type) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32)
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+intra_pred_no_4x4(d207)
+intra_pred_no_4x4(d63)
+intra_pred_no_4x4(d45)
+#if CONFIG_MISC_FIXES
+intra_pred_allsizes(d207e)
+intra_pred_allsizes(d63e)
+intra_pred_no_4x4(d45e)
+#endif
+intra_pred_no_4x4(d117)
+intra_pred_no_4x4(d135)
+intra_pred_no_4x4(d153)
+intra_pred_allsizes(v)
+intra_pred_allsizes(h)
+intra_pred_allsizes(tm)
+intra_pred_allsizes(dc_128)
+intra_pred_allsizes(dc_left)
+intra_pred_allsizes(dc_top)
+intra_pred_allsizes(dc)
+#undef intra_pred_allsizes
diff --git a/libvpx/libvpx/vpx_dsp/inv_txfm.c b/libvpx/libvpx/vpx_dsp/inv_txfm.c
new file mode 100644
index 0000000..e18d31d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/inv_txfm.c
@@ -0,0 +1,2518 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <string.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/inv_txfm.h"
+
+void vpx_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+/* 4-point reversible, orthonormal inverse Walsh-Hadamard in 3.5 adds,
+   0.5 shifts per pixel. */
+  int i;
+  tran_low_t output[16];
+  tran_high_t a1, b1, c1, d1, e1;
+  const tran_low_t *ip = input;
+  tran_low_t *op = output;
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip[0] >> UNIT_QUANT_SHIFT;
+    c1 = ip[1] >> UNIT_QUANT_SHIFT;
+    d1 = ip[2] >> UNIT_QUANT_SHIFT;
+    b1 = ip[3] >> UNIT_QUANT_SHIFT;
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    op[0] = WRAPLOW(a1);
+    op[1] = WRAPLOW(b1);
+    op[2] = WRAPLOW(c1);
+    op[3] = WRAPLOW(d1);
+    ip += 4;
+    op += 4;
+  }
+
+  ip = output;
+  for (i = 0; i < 4; i++) {
+    a1 = ip[4 * 0];
+    c1 = ip[4 * 1];
+    d1 = ip[4 * 2];
+    b1 = ip[4 * 3];
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    dest[stride * 0] = clip_pixel_add(dest[stride * 0], WRAPLOW(a1));
+    dest[stride * 1] = clip_pixel_add(dest[stride * 1], WRAPLOW(b1));
+    dest[stride * 2] = clip_pixel_add(dest[stride * 2], WRAPLOW(c1));
+    dest[stride * 3] = clip_pixel_add(dest[stride * 3], WRAPLOW(d1));
+
+    ip++;
+    dest++;
+  }
+}
+
+void vpx_iwht4x4_1_add_c(const tran_low_t *in, uint8_t *dest, int dest_stride) {
+  int i;
+  tran_high_t a1, e1;
+  tran_low_t tmp[4];
+  const tran_low_t *ip = in;
+  tran_low_t *op = tmp;
+
+  a1 = ip[0] >> UNIT_QUANT_SHIFT;
+  e1 = a1 >> 1;
+  a1 -= e1;
+  op[0] = WRAPLOW(a1);
+  op[1] = op[2] = op[3] = WRAPLOW(e1);
+
+  ip = tmp;
+  for (i = 0; i < 4; i++) {
+    e1 = ip[0] >> 1;
+    a1 = ip[0] - e1;
+    dest[dest_stride * 0] = clip_pixel_add(dest[dest_stride * 0], a1);
+    dest[dest_stride * 1] = clip_pixel_add(dest[dest_stride * 1], e1);
+    dest[dest_stride * 2] = clip_pixel_add(dest[dest_stride * 2], e1);
+    dest[dest_stride * 3] = clip_pixel_add(dest[dest_stride * 3], e1);
+    ip++;
+    dest++;
+  }
+}
+
+void idct4_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step[4];
+  tran_high_t temp1, temp2;
+  // stage 1
+  temp1 = (input[0] + input[2]) * cospi_16_64;
+  temp2 = (input[0] - input[2]) * cospi_16_64;
+  step[0] = WRAPLOW(dct_const_round_shift(temp1));
+  step[1] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+  temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+  step[2] = WRAPLOW(dct_const_round_shift(temp1));
+  step[3] = WRAPLOW(dct_const_round_shift(temp2));
+
+  // stage 2
+  output[0] = WRAPLOW(step[0] + step[3]);
+  output[1] = WRAPLOW(step[1] + step[2]);
+  output[2] = WRAPLOW(step[1] - step[2]);
+  output[3] = WRAPLOW(step[0] - step[3]);
+}
+
+void vpx_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[4], temp_out[4];
+
+  // Rows
+  for (i = 0; i < 4; ++i) {
+    idct4_c(input, outptr);
+    input += 4;
+    outptr += 4;
+  }
+
+  // Columns
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j)
+      temp_in[j] = out[j * 4 + i];
+    idct4_c(temp_in, temp_out);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 4));
+    }
+  }
+}
+
+void vpx_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest,
+                         int dest_stride) {
+  int i;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64));
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
+  a1 = ROUND_POWER_OF_TWO(out, 4);
+
+  for (i = 0; i < 4; i++) {
+    dest[0] = clip_pixel_add(dest[0], a1);
+    dest[1] = clip_pixel_add(dest[1], a1);
+    dest[2] = clip_pixel_add(dest[2], a1);
+    dest[3] = clip_pixel_add(dest[3], a1);
+    dest += dest_stride;
+  }
+}
+
+void idct8_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[8], step2[8];
+  tran_high_t temp1, temp2;
+  // stage 1
+  step1[0] = input[0];
+  step1[2] = input[4];
+  step1[1] = input[2];
+  step1[3] = input[6];
+  temp1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+  temp2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+  temp2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+
+  // stage 2
+  temp1 = (step1[0] + step1[2]) * cospi_16_64;
+  temp2 = (step1[0] - step1[2]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = step1[1] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[1] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[4] = WRAPLOW(step1[4] + step1[5]);
+  step2[5] = WRAPLOW(step1[4] - step1[5]);
+  step2[6] = WRAPLOW(-step1[6] + step1[7]);
+  step2[7] = WRAPLOW(step1[6] + step1[7]);
+
+  // stage 3
+  step1[0] = WRAPLOW(step2[0] + step2[3]);
+  step1[1] = WRAPLOW(step2[1] + step2[2]);
+  step1[2] = WRAPLOW(step2[1] - step2[2]);
+  step1[3] = WRAPLOW(step2[0] - step2[3]);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[7] = step2[7];
+
+  // stage 4
+  output[0] = WRAPLOW(step1[0] + step1[7]);
+  output[1] = WRAPLOW(step1[1] + step1[6]);
+  output[2] = WRAPLOW(step1[2] + step1[5]);
+  output[3] = WRAPLOW(step1[3] + step1[4]);
+  output[4] = WRAPLOW(step1[3] - step1[4]);
+  output[5] = WRAPLOW(step1[2] - step1[5]);
+  output[6] = WRAPLOW(step1[1] - step1[6]);
+  output[7] = WRAPLOW(step1[0] - step1[7]);
+}
+
+void vpx_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+
+  // First transform rows
+  for (i = 0; i < 8; ++i) {
+    idct8_c(input, outptr);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    idct8_c(temp_in, temp_out);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 5));
+    }
+  }
+}
+
+void vpx_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64));
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
+  a1 = ROUND_POWER_OF_TWO(out, 5);
+  for (j = 0; j < 8; ++j) {
+    for (i = 0; i < 8; ++i)
+      dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+void iadst4_c(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_low_t x0 = input[0];
+  tran_low_t x1 = input[1];
+  tran_low_t x2 = input[2];
+  tran_low_t x3 = input[3];
+
+  if (!(x0 | x1 | x2 | x3)) {
+    output[0] = output[1] = output[2] = output[3] = 0;
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_2_9 * x0;
+  s2 = sinpi_3_9 * x1;
+  s3 = sinpi_4_9 * x2;
+  s4 = sinpi_1_9 * x2;
+  s5 = sinpi_2_9 * x3;
+  s6 = sinpi_4_9 * x3;
+  s7 = WRAPLOW(x0 - x2 + x3);
+
+  s0 = s0 + s3 + s5;
+  s1 = s1 - s4 - s6;
+  s3 = s2;
+  s2 = sinpi_3_9 * s7;
+
+  // 1-D transform scaling factor is sqrt(2).
+  // The overall dynamic range is 14b (input) + 14b (multiplication scaling)
+  // + 1b (addition) = 29b.
+  // Hence the output bit depth is 15b.
+  output[0] = WRAPLOW(dct_const_round_shift(s0 + s3));
+  output[1] = WRAPLOW(dct_const_round_shift(s1 + s3));
+  output[2] = WRAPLOW(dct_const_round_shift(s2));
+  output[3] = WRAPLOW(dct_const_round_shift(s0 + s1 - s3));
+}
+
+void iadst8_c(const tran_low_t *input, tran_low_t *output) {
+  int s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_high_t x0 = input[7];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[5];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[3];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[1];
+  tran_high_t x7 = input[6];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7)) {
+    output[0] = output[1] = output[2] = output[3] = output[4]
+              = output[5] = output[6] = output[7] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = (int)(cospi_2_64  * x0 + cospi_30_64 * x1);
+  s1 = (int)(cospi_30_64 * x0 - cospi_2_64  * x1);
+  s2 = (int)(cospi_10_64 * x2 + cospi_22_64 * x3);
+  s3 = (int)(cospi_22_64 * x2 - cospi_10_64 * x3);
+  s4 = (int)(cospi_18_64 * x4 + cospi_14_64 * x5);
+  s5 = (int)(cospi_14_64 * x4 - cospi_18_64 * x5);
+  s6 = (int)(cospi_26_64 * x6 + cospi_6_64  * x7);
+  s7 = (int)(cospi_6_64  * x6 - cospi_26_64 * x7);
+
+  x0 = WRAPLOW(dct_const_round_shift(s0 + s4));
+  x1 = WRAPLOW(dct_const_round_shift(s1 + s5));
+  x2 = WRAPLOW(dct_const_round_shift(s2 + s6));
+  x3 = WRAPLOW(dct_const_round_shift(s3 + s7));
+  x4 = WRAPLOW(dct_const_round_shift(s0 - s4));
+  x5 = WRAPLOW(dct_const_round_shift(s1 - s5));
+  x6 = WRAPLOW(dct_const_round_shift(s2 - s6));
+  x7 = WRAPLOW(dct_const_round_shift(s3 - s7));
+
+  // stage 2
+  s0 = (int)x0;
+  s1 = (int)x1;
+  s2 = (int)x2;
+  s3 = (int)x3;
+  s4 = (int)(cospi_8_64 * x4 + cospi_24_64 * x5);
+  s5 = (int)(cospi_24_64 * x4 - cospi_8_64 * x5);
+  s6 = (int)(-cospi_24_64 * x6 + cospi_8_64 * x7);
+  s7 = (int)(cospi_8_64 * x6 + cospi_24_64 * x7);
+
+  x0 = WRAPLOW(s0 + s2);
+  x1 = WRAPLOW(s1 + s3);
+  x2 = WRAPLOW(s0 - s2);
+  x3 = WRAPLOW(s1 - s3);
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s6));
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s7));
+  x6 = WRAPLOW(dct_const_round_shift(s4 - s6));
+  x7 = WRAPLOW(dct_const_round_shift(s5 - s7));
+
+  // stage 3
+  s2 = (int)(cospi_16_64 * (x2 + x3));
+  s3 = (int)(cospi_16_64 * (x2 - x3));
+  s6 = (int)(cospi_16_64 * (x6 + x7));
+  s7 = (int)(cospi_16_64 * (x6 - x7));
+
+  x2 = WRAPLOW(dct_const_round_shift(s2));
+  x3 = WRAPLOW(dct_const_round_shift(s3));
+  x6 = WRAPLOW(dct_const_round_shift(s6));
+  x7 = WRAPLOW(dct_const_round_shift(s7));
+
+  output[0] = WRAPLOW(x0);
+  output[1] = WRAPLOW(-x4);
+  output[2] = WRAPLOW(x6);
+  output[3] = WRAPLOW(-x2);
+  output[4] = WRAPLOW(x3);
+  output[5] = WRAPLOW(-x7);
+  output[6] = WRAPLOW(x5);
+  output[7] = WRAPLOW(-x1);
+}
+
+void vpx_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[8 * 8] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+
+  // First transform rows
+  // only first 4 row has non-zero coefs
+  for (i = 0; i < 4; ++i) {
+    idct8_c(input, outptr);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    idct8_c(temp_in, temp_out);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 5));
+    }
+  }
+}
+
+void idct16_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[16], step2[16];
+  tran_high_t temp1, temp2;
+
+  // stage 1
+  step1[0] = input[0/2];
+  step1[1] = input[16/2];
+  step1[2] = input[8/2];
+  step1[3] = input[24/2];
+  step1[4] = input[4/2];
+  step1[5] = input[20/2];
+  step1[6] = input[12/2];
+  step1[7] = input[28/2];
+  step1[8] = input[2/2];
+  step1[9] = input[18/2];
+  step1[10] = input[10/2];
+  step1[11] = input[26/2];
+  step1[12] = input[6/2];
+  step1[13] = input[22/2];
+  step1[14] = input[14/2];
+  step1[15] = input[30/2];
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[15] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2));
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+
+  step1[8] = WRAPLOW(step2[8] + step2[9]);
+  step1[9] = WRAPLOW(step2[8] - step2[9]);
+  step1[10] = WRAPLOW(-step2[10] + step2[11]);
+  step1[11] = WRAPLOW(step2[10] + step2[11]);
+  step1[12] = WRAPLOW(step2[12] + step2[13]);
+  step1[13] = WRAPLOW(step2[12] - step2[13]);
+  step1[14] = WRAPLOW(-step2[14] + step2[15]);
+  step1[15] = WRAPLOW(step2[14] + step2[15]);
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[4] = WRAPLOW(step1[4] + step1[5]);
+  step2[5] = WRAPLOW(step1[4] - step1[5]);
+  step2[6] = WRAPLOW(-step1[6] + step1[7]);
+  step2[7] = WRAPLOW(step1[6] + step1[7]);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3]);
+  step1[1] = WRAPLOW(step2[1] + step2[2]);
+  step1[2] = WRAPLOW(step2[1] - step2[2]);
+  step1[3] = WRAPLOW(step2[0] - step2[3]);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11]);
+  step1[9] = WRAPLOW(step2[9] + step2[10]);
+  step1[10] = WRAPLOW(step2[9] - step2[10]);
+  step1[11] = WRAPLOW(step2[8] - step2[11]);
+  step1[12] = WRAPLOW(-step2[12] + step2[15]);
+  step1[13] = WRAPLOW(-step2[13] + step2[14]);
+  step1[14] = WRAPLOW(step2[13] + step2[14]);
+  step1[15] = WRAPLOW(step2[12] + step2[15]);
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7]);
+  step2[1] = WRAPLOW(step1[1] + step1[6]);
+  step2[2] = WRAPLOW(step1[2] + step1[5]);
+  step2[3] = WRAPLOW(step1[3] + step1[4]);
+  step2[4] = WRAPLOW(step1[3] - step1[4]);
+  step2[5] = WRAPLOW(step1[2] - step1[5]);
+  step2[6] = WRAPLOW(step1[1] - step1[6]);
+  step2[7] = WRAPLOW(step1[0] - step1[7]);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  // stage 7
+  output[0] = WRAPLOW(step2[0] + step2[15]);
+  output[1] = WRAPLOW(step2[1] + step2[14]);
+  output[2] = WRAPLOW(step2[2] + step2[13]);
+  output[3] = WRAPLOW(step2[3] + step2[12]);
+  output[4] = WRAPLOW(step2[4] + step2[11]);
+  output[5] = WRAPLOW(step2[5] + step2[10]);
+  output[6] = WRAPLOW(step2[6] + step2[9]);
+  output[7] = WRAPLOW(step2[7] + step2[8]);
+  output[8] = WRAPLOW(step2[7] - step2[8]);
+  output[9] = WRAPLOW(step2[6] - step2[9]);
+  output[10] = WRAPLOW(step2[5] - step2[10]);
+  output[11] = WRAPLOW(step2[4] - step2[11]);
+  output[12] = WRAPLOW(step2[3] - step2[12]);
+  output[13] = WRAPLOW(step2[2] - step2[13]);
+  output[14] = WRAPLOW(step2[1] - step2[14]);
+  output[15] = WRAPLOW(step2[0] - step2[15]);
+}
+
+void vpx_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+
+  // First transform rows
+  for (i = 0; i < 16; ++i) {
+    idct16_c(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j * 16 + i];
+    idct16_c(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void iadst16_c(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7, s8;
+  tran_high_t s9, s10, s11, s12, s13, s14, s15;
+
+  tran_high_t x0 = input[15];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[13];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[11];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[9];
+  tran_high_t x7 = input[6];
+  tran_high_t x8 = input[7];
+  tran_high_t x9 = input[8];
+  tran_high_t x10 = input[5];
+  tran_high_t x11 = input[10];
+  tran_high_t x12 = input[3];
+  tran_high_t x13 = input[12];
+  tran_high_t x14 = input[1];
+  tran_high_t x15 = input[14];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7 | x8
+           | x9 | x10 | x11 | x12 | x13 | x14 | x15)) {
+    output[0] = output[1] = output[2] = output[3] = output[4]
+              = output[5] = output[6] = output[7] = output[8]
+              = output[9] = output[10] = output[11] = output[12]
+              = output[13] = output[14] = output[15] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = x0 * cospi_1_64  + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64  + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64  + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64  - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64  - x15 * cospi_29_64;
+
+  x0 = WRAPLOW(dct_const_round_shift(s0 + s8));
+  x1 = WRAPLOW(dct_const_round_shift(s1 + s9));
+  x2 = WRAPLOW(dct_const_round_shift(s2 + s10));
+  x3 = WRAPLOW(dct_const_round_shift(s3 + s11));
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s12));
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s13));
+  x6 = WRAPLOW(dct_const_round_shift(s6 + s14));
+  x7 = WRAPLOW(dct_const_round_shift(s7 + s15));
+  x8 = WRAPLOW(dct_const_round_shift(s0 - s8));
+  x9 = WRAPLOW(dct_const_round_shift(s1 - s9));
+  x10 = WRAPLOW(dct_const_round_shift(s2 - s10));
+  x11 = WRAPLOW(dct_const_round_shift(s3 - s11));
+  x12 = WRAPLOW(dct_const_round_shift(s4 - s12));
+  x13 = WRAPLOW(dct_const_round_shift(s5 - s13));
+  x14 = WRAPLOW(dct_const_round_shift(s6 - s14));
+  x15 = WRAPLOW(dct_const_round_shift(s7 - s15));
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 =    x8 * cospi_4_64   + x9 * cospi_28_64;
+  s9 =    x8 * cospi_28_64  - x9 * cospi_4_64;
+  s10 =   x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 =   x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = - x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 =   x12 * cospi_4_64  + x13 * cospi_28_64;
+  s14 = - x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 =   x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = WRAPLOW(s0 + s4);
+  x1 = WRAPLOW(s1 + s5);
+  x2 = WRAPLOW(s2 + s6);
+  x3 = WRAPLOW(s3 + s7);
+  x4 = WRAPLOW(s0 - s4);
+  x5 = WRAPLOW(s1 - s5);
+  x6 = WRAPLOW(s2 - s6);
+  x7 = WRAPLOW(s3 - s7);
+  x8 = WRAPLOW(dct_const_round_shift(s8 + s12));
+  x9 = WRAPLOW(dct_const_round_shift(s9 + s13));
+  x10 = WRAPLOW(dct_const_round_shift(s10 + s14));
+  x11 = WRAPLOW(dct_const_round_shift(s11 + s15));
+  x12 = WRAPLOW(dct_const_round_shift(s8 - s12));
+  x13 = WRAPLOW(dct_const_round_shift(s9 - s13));
+  x14 = WRAPLOW(dct_const_round_shift(s10 - s14));
+  x15 = WRAPLOW(dct_const_round_shift(s11 - s15));
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64  + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = - x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 =   x6 * cospi_8_64  + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64  + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = - x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 =   x14 * cospi_8_64  + x15 * cospi_24_64;
+
+  x0 = WRAPLOW(s0 + s2);
+  x1 = WRAPLOW(s1 + s3);
+  x2 = WRAPLOW(s0 - s2);
+  x3 = WRAPLOW(s1 - s3);
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s6));
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s7));
+  x6 = WRAPLOW(dct_const_round_shift(s4 - s6));
+  x7 = WRAPLOW(dct_const_round_shift(s5 - s7));
+  x8 = WRAPLOW(s8 + s10);
+  x9 = WRAPLOW(s9 + s11);
+  x10 = WRAPLOW(s8 - s10);
+  x11 = WRAPLOW(s9 - s11);
+  x12 = WRAPLOW(dct_const_round_shift(s12 + s14));
+  x13 = WRAPLOW(dct_const_round_shift(s13 + s15));
+  x14 = WRAPLOW(dct_const_round_shift(s12 - s14));
+  x15 = WRAPLOW(dct_const_round_shift(s13 - s15));
+
+  // stage 4
+  s2 = (- cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (- x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (- x10 + x11);
+  s14 = (- cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = WRAPLOW(dct_const_round_shift(s2));
+  x3 = WRAPLOW(dct_const_round_shift(s3));
+  x6 = WRAPLOW(dct_const_round_shift(s6));
+  x7 = WRAPLOW(dct_const_round_shift(s7));
+  x10 = WRAPLOW(dct_const_round_shift(s10));
+  x11 = WRAPLOW(dct_const_round_shift(s11));
+  x14 = WRAPLOW(dct_const_round_shift(s14));
+  x15 = WRAPLOW(dct_const_round_shift(s15));
+
+  output[0] = WRAPLOW(x0);
+  output[1] = WRAPLOW(-x8);
+  output[2] = WRAPLOW(x12);
+  output[3] = WRAPLOW(-x4);
+  output[4] = WRAPLOW(x6);
+  output[5] = WRAPLOW(x14);
+  output[6] = WRAPLOW(x10);
+  output[7] = WRAPLOW(x2);
+  output[8] = WRAPLOW(x3);
+  output[9] = WRAPLOW(x11);
+  output[10] = WRAPLOW(x15);
+  output[11] = WRAPLOW(x7);
+  output[12] = WRAPLOW(x5);
+  output[13] = WRAPLOW(-x13);
+  output[14] = WRAPLOW(x9);
+  output[15] = WRAPLOW(-x1);
+}
+
+void vpx_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+
+  // First transform rows. Since all non-zero dct coefficients are in
+  // upper-left 4x4 area, we only need to calculate first 4 rows here.
+  for (i = 0; i < 4; ++i) {
+    idct16_c(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j*16 + i];
+    idct16_c(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vpx_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64));
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+  for (j = 0; j < 16; ++j) {
+    for (i = 0; i < 16; ++i)
+      dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+void idct32_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[32], step2[32];
+  tran_high_t temp1, temp2;
+
+  // stage 1
+  step1[0] = input[0];
+  step1[1] = input[16];
+  step1[2] = input[8];
+  step1[3] = input[24];
+  step1[4] = input[4];
+  step1[5] = input[20];
+  step1[6] = input[12];
+  step1[7] = input[28];
+  step1[8] = input[2];
+  step1[9] = input[18];
+  step1[10] = input[10];
+  step1[11] = input[26];
+  step1[12] = input[6];
+  step1[13] = input[22];
+  step1[14] = input[14];
+  step1[15] = input[30];
+
+  temp1 = input[1] * cospi_31_64 - input[31] * cospi_1_64;
+  temp2 = input[1] * cospi_1_64 + input[31] * cospi_31_64;
+  step1[16] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[31] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[17] * cospi_15_64 - input[15] * cospi_17_64;
+  temp2 = input[17] * cospi_17_64 + input[15] * cospi_15_64;
+  step1[17] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[30] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[9] * cospi_23_64 - input[23] * cospi_9_64;
+  temp2 = input[9] * cospi_9_64 + input[23] * cospi_23_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[25] * cospi_7_64 - input[7] * cospi_25_64;
+  temp2 = input[25] * cospi_25_64 + input[7] * cospi_7_64;
+  step1[19] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[28] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[5] * cospi_27_64 - input[27] * cospi_5_64;
+  temp2 = input[5] * cospi_5_64 + input[27] * cospi_27_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[21] * cospi_11_64 - input[11] * cospi_21_64;
+  temp2 = input[21] * cospi_21_64 + input[11] * cospi_11_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[13] * cospi_19_64 - input[19] * cospi_13_64;
+  temp2 = input[13] * cospi_13_64 + input[19] * cospi_19_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[29] * cospi_3_64 - input[3] * cospi_29_64;
+  temp2 = input[29] * cospi_29_64 + input[3] * cospi_3_64;
+  step1[23] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[24] = WRAPLOW(dct_const_round_shift(temp2));
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[15] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2));
+
+  step2[16] = WRAPLOW(step1[16] + step1[17]);
+  step2[17] = WRAPLOW(step1[16] - step1[17]);
+  step2[18] = WRAPLOW(-step1[18] + step1[19]);
+  step2[19] = WRAPLOW(step1[18] + step1[19]);
+  step2[20] = WRAPLOW(step1[20] + step1[21]);
+  step2[21] = WRAPLOW(step1[20] - step1[21]);
+  step2[22] = WRAPLOW(-step1[22] + step1[23]);
+  step2[23] = WRAPLOW(step1[22] + step1[23]);
+  step2[24] = WRAPLOW(step1[24] + step1[25]);
+  step2[25] = WRAPLOW(step1[24] - step1[25]);
+  step2[26] = WRAPLOW(-step1[26] + step1[27]);
+  step2[27] = WRAPLOW(step1[26] + step1[27]);
+  step2[28] = WRAPLOW(step1[28] + step1[29]);
+  step2[29] = WRAPLOW(step1[28] - step1[29]);
+  step2[30] = WRAPLOW(-step1[30] + step1[31]);
+  step2[31] = WRAPLOW(step1[30] + step1[31]);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+
+  step1[8] = WRAPLOW(step2[8] + step2[9]);
+  step1[9] = WRAPLOW(step2[8] - step2[9]);
+  step1[10] = WRAPLOW(-step2[10] + step2[11]);
+  step1[11] = WRAPLOW(step2[10] + step2[11]);
+  step1[12] = WRAPLOW(step2[12] + step2[13]);
+  step1[13] = WRAPLOW(step2[12] - step2[13]);
+  step1[14] = WRAPLOW(-step2[14] + step2[15]);
+  step1[15] = WRAPLOW(step2[14] + step2[15]);
+
+  step1[16] = step2[16];
+  step1[31] = step2[31];
+  temp1 = -step2[17] * cospi_4_64 + step2[30] * cospi_28_64;
+  temp2 = step2[17] * cospi_28_64 + step2[30] * cospi_4_64;
+  step1[17] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[30] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step2[18] * cospi_28_64 - step2[29] * cospi_4_64;
+  temp2 = -step2[18] * cospi_4_64 + step2[29] * cospi_28_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[19] = step2[19];
+  step1[20] = step2[20];
+  temp1 = -step2[21] * cospi_20_64 + step2[26] * cospi_12_64;
+  temp2 = step2[21] * cospi_12_64 + step2[26] * cospi_20_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step2[22] * cospi_12_64 - step2[25] * cospi_20_64;
+  temp2 = -step2[22] * cospi_20_64 + step2[25] * cospi_12_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[27] = step2[27];
+  step1[28] = step2[28];
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[4] = WRAPLOW(step1[4] + step1[5]);
+  step2[5] = WRAPLOW(step1[4] - step1[5]);
+  step2[6] = WRAPLOW(-step1[6] + step1[7]);
+  step2[7] = WRAPLOW(step1[6] + step1[7]);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  step2[16] = WRAPLOW(step1[16] + step1[19]);
+  step2[17] = WRAPLOW(step1[17] + step1[18]);
+  step2[18] = WRAPLOW(step1[17] - step1[18]);
+  step2[19] = WRAPLOW(step1[16] - step1[19]);
+  step2[20] = WRAPLOW(-step1[20] + step1[23]);
+  step2[21] = WRAPLOW(-step1[21] + step1[22]);
+  step2[22] = WRAPLOW(step1[21] + step1[22]);
+  step2[23] = WRAPLOW(step1[20] + step1[23]);
+
+  step2[24] = WRAPLOW(step1[24] + step1[27]);
+  step2[25] = WRAPLOW(step1[25] + step1[26]);
+  step2[26] = WRAPLOW(step1[25] - step1[26]);
+  step2[27] = WRAPLOW(step1[24] - step1[27]);
+  step2[28] = WRAPLOW(-step1[28] + step1[31]);
+  step2[29] = WRAPLOW(-step1[29] + step1[30]);
+  step2[30] = WRAPLOW(step1[29] + step1[30]);
+  step2[31] = WRAPLOW(step1[28] + step1[31]);
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3]);
+  step1[1] = WRAPLOW(step2[1] + step2[2]);
+  step1[2] = WRAPLOW(step2[1] - step2[2]);
+  step1[3] = WRAPLOW(step2[0] - step2[3]);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11]);
+  step1[9] = WRAPLOW(step2[9] + step2[10]);
+  step1[10] = WRAPLOW(step2[9] - step2[10]);
+  step1[11] = WRAPLOW(step2[8] - step2[11]);
+  step1[12] = WRAPLOW(-step2[12] + step2[15]);
+  step1[13] = WRAPLOW(-step2[13] + step2[14]);
+  step1[14] = WRAPLOW(step2[13] + step2[14]);
+  step1[15] = WRAPLOW(step2[12] + step2[15]);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  temp1 = -step2[18] * cospi_8_64 + step2[29] * cospi_24_64;
+  temp2 = step2[18] * cospi_24_64 + step2[29] * cospi_8_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step2[19] * cospi_8_64 + step2[28] * cospi_24_64;
+  temp2 = step2[19] * cospi_24_64 + step2[28] * cospi_8_64;
+  step1[19] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[28] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step2[20] * cospi_24_64 - step2[27] * cospi_8_64;
+  temp2 = -step2[20] * cospi_8_64 + step2[27] * cospi_24_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step2[21] * cospi_24_64 - step2[26] * cospi_8_64;
+  temp2 = -step2[21] * cospi_8_64 + step2[26] * cospi_24_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[22] = step2[22];
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[25] = step2[25];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7]);
+  step2[1] = WRAPLOW(step1[1] + step1[6]);
+  step2[2] = WRAPLOW(step1[2] + step1[5]);
+  step2[3] = WRAPLOW(step1[3] + step1[4]);
+  step2[4] = WRAPLOW(step1[3] - step1[4]);
+  step2[5] = WRAPLOW(step1[2] - step1[5]);
+  step2[6] = WRAPLOW(step1[1] - step1[6]);
+  step2[7] = WRAPLOW(step1[0] - step1[7]);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  step2[16] = WRAPLOW(step1[16] + step1[23]);
+  step2[17] = WRAPLOW(step1[17] + step1[22]);
+  step2[18] = WRAPLOW(step1[18] + step1[21]);
+  step2[19] = WRAPLOW(step1[19] + step1[20]);
+  step2[20] = WRAPLOW(step1[19] - step1[20]);
+  step2[21] = WRAPLOW(step1[18] - step1[21]);
+  step2[22] = WRAPLOW(step1[17] - step1[22]);
+  step2[23] = WRAPLOW(step1[16] - step1[23]);
+
+  step2[24] = WRAPLOW(-step1[24] + step1[31]);
+  step2[25] = WRAPLOW(-step1[25] + step1[30]);
+  step2[26] = WRAPLOW(-step1[26] + step1[29]);
+  step2[27] = WRAPLOW(-step1[27] + step1[28]);
+  step2[28] = WRAPLOW(step1[27] + step1[28]);
+  step2[29] = WRAPLOW(step1[26] + step1[29]);
+  step2[30] = WRAPLOW(step1[25] + step1[30]);
+  step2[31] = WRAPLOW(step1[24] + step1[31]);
+
+  // stage 7
+  step1[0] = WRAPLOW(step2[0] + step2[15]);
+  step1[1] = WRAPLOW(step2[1] + step2[14]);
+  step1[2] = WRAPLOW(step2[2] + step2[13]);
+  step1[3] = WRAPLOW(step2[3] + step2[12]);
+  step1[4] = WRAPLOW(step2[4] + step2[11]);
+  step1[5] = WRAPLOW(step2[5] + step2[10]);
+  step1[6] = WRAPLOW(step2[6] + step2[9]);
+  step1[7] = WRAPLOW(step2[7] + step2[8]);
+  step1[8] = WRAPLOW(step2[7] - step2[8]);
+  step1[9] = WRAPLOW(step2[6] - step2[9]);
+  step1[10] = WRAPLOW(step2[5] - step2[10]);
+  step1[11] = WRAPLOW(step2[4] - step2[11]);
+  step1[12] = WRAPLOW(step2[3] - step2[12]);
+  step1[13] = WRAPLOW(step2[2] - step2[13]);
+  step1[14] = WRAPLOW(step2[1] - step2[14]);
+  step1[15] = WRAPLOW(step2[0] - step2[15]);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  step1[18] = step2[18];
+  step1[19] = step2[19];
+  temp1 = (-step2[20] + step2[27]) * cospi_16_64;
+  temp2 = (step2[20] + step2[27]) * cospi_16_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = (-step2[21] + step2[26]) * cospi_16_64;
+  temp2 = (step2[21] + step2[26]) * cospi_16_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = (-step2[22] + step2[25]) * cospi_16_64;
+  temp2 = (step2[22] + step2[25]) * cospi_16_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = (-step2[23] + step2[24]) * cospi_16_64;
+  temp2 = (step2[23] + step2[24]) * cospi_16_64;
+  step1[23] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[24] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[28] = step2[28];
+  step1[29] = step2[29];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // final stage
+  output[0] = WRAPLOW(step1[0] + step1[31]);
+  output[1] = WRAPLOW(step1[1] + step1[30]);
+  output[2] = WRAPLOW(step1[2] + step1[29]);
+  output[3] = WRAPLOW(step1[3] + step1[28]);
+  output[4] = WRAPLOW(step1[4] + step1[27]);
+  output[5] = WRAPLOW(step1[5] + step1[26]);
+  output[6] = WRAPLOW(step1[6] + step1[25]);
+  output[7] = WRAPLOW(step1[7] + step1[24]);
+  output[8] = WRAPLOW(step1[8] + step1[23]);
+  output[9] = WRAPLOW(step1[9] + step1[22]);
+  output[10] = WRAPLOW(step1[10] + step1[21]);
+  output[11] = WRAPLOW(step1[11] + step1[20]);
+  output[12] = WRAPLOW(step1[12] + step1[19]);
+  output[13] = WRAPLOW(step1[13] + step1[18]);
+  output[14] = WRAPLOW(step1[14] + step1[17]);
+  output[15] = WRAPLOW(step1[15] + step1[16]);
+  output[16] = WRAPLOW(step1[15] - step1[16]);
+  output[17] = WRAPLOW(step1[14] - step1[17]);
+  output[18] = WRAPLOW(step1[13] - step1[18]);
+  output[19] = WRAPLOW(step1[12] - step1[19]);
+  output[20] = WRAPLOW(step1[11] - step1[20]);
+  output[21] = WRAPLOW(step1[10] - step1[21]);
+  output[22] = WRAPLOW(step1[9] - step1[22]);
+  output[23] = WRAPLOW(step1[8] - step1[23]);
+  output[24] = WRAPLOW(step1[7] - step1[24]);
+  output[25] = WRAPLOW(step1[6] - step1[25]);
+  output[26] = WRAPLOW(step1[5] - step1[26]);
+  output[27] = WRAPLOW(step1[4] - step1[27]);
+  output[28] = WRAPLOW(step1[3] - step1[28]);
+  output[29] = WRAPLOW(step1[2] - step1[29]);
+  output[30] = WRAPLOW(step1[1] - step1[30]);
+  output[31] = WRAPLOW(step1[0] - step1[31]);
+}
+
+void vpx_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest,
+                              int stride) {
+  tran_low_t out[32 * 32];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    int16_t zero_coeff[16];
+    for (j = 0; j < 16; ++j)
+      zero_coeff[j] = input[2 * j] | input[2 * j + 1];
+    for (j = 0; j < 8; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 4; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 2; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+
+    if (zero_coeff[0] | zero_coeff[1])
+      idct32_c(input, outptr);
+    else
+      memset(outptr, 0, sizeof(tran_low_t) * 32);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    idct32_c(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vpx_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  tran_low_t out[32 * 32] = {0};
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+  // Rows
+  // only upper-left 16x16 has non-zero coeff
+  for (i = 0; i < 16; ++i) {
+    idct32_c(input, outptr);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    idct32_c(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vpx_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  tran_low_t out[32 * 32] = {0};
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+  // Rows
+  // only upper-left 8x8 has non-zero coeff
+  for (i = 0; i < 8; ++i) {
+    idct32_c(input, outptr);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    idct32_c(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vpx_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  int i, j;
+  tran_high_t a1;
+
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64));
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+
+  for (j = 0; j < 32; ++j) {
+    for (i = 0; i < 32; ++i)
+      dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  /* 4-point reversible, orthonormal inverse Walsh-Hadamard in 3.5 adds,
+     0.5 shifts per pixel. */
+  int i;
+  tran_low_t output[16];
+  tran_high_t a1, b1, c1, d1, e1;
+  const tran_low_t *ip = input;
+  tran_low_t *op = output;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip[0] >> UNIT_QUANT_SHIFT;
+    c1 = ip[1] >> UNIT_QUANT_SHIFT;
+    d1 = ip[2] >> UNIT_QUANT_SHIFT;
+    b1 = ip[3] >> UNIT_QUANT_SHIFT;
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    op[0] = HIGHBD_WRAPLOW(a1, bd);
+    op[1] = HIGHBD_WRAPLOW(b1, bd);
+    op[2] = HIGHBD_WRAPLOW(c1, bd);
+    op[3] = HIGHBD_WRAPLOW(d1, bd);
+    ip += 4;
+    op += 4;
+  }
+
+  ip = output;
+  for (i = 0; i < 4; i++) {
+    a1 = ip[4 * 0];
+    c1 = ip[4 * 1];
+    d1 = ip[4 * 2];
+    b1 = ip[4 * 3];
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    dest[stride * 0] = highbd_clip_pixel_add(dest[stride * 0],
+                                             HIGHBD_WRAPLOW(a1, bd), bd);
+    dest[stride * 1] = highbd_clip_pixel_add(dest[stride * 1],
+                                             HIGHBD_WRAPLOW(b1, bd), bd);
+    dest[stride * 2] = highbd_clip_pixel_add(dest[stride * 2],
+                                             HIGHBD_WRAPLOW(c1, bd), bd);
+    dest[stride * 3] = highbd_clip_pixel_add(dest[stride * 3],
+                                             HIGHBD_WRAPLOW(d1, bd), bd);
+
+    ip++;
+    dest++;
+  }
+}
+
+void vpx_highbd_iwht4x4_1_add_c(const tran_low_t *in, uint8_t *dest8,
+                                int dest_stride, int bd) {
+  int i;
+  tran_high_t a1, e1;
+  tran_low_t tmp[4];
+  const tran_low_t *ip = in;
+  tran_low_t *op = tmp;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  (void) bd;
+
+  a1 = ip[0] >> UNIT_QUANT_SHIFT;
+  e1 = a1 >> 1;
+  a1 -= e1;
+  op[0] = HIGHBD_WRAPLOW(a1, bd);
+  op[1] = op[2] = op[3] = HIGHBD_WRAPLOW(e1, bd);
+
+  ip = tmp;
+  for (i = 0; i < 4; i++) {
+    e1 = ip[0] >> 1;
+    a1 = ip[0] - e1;
+    dest[dest_stride * 0] = highbd_clip_pixel_add(
+        dest[dest_stride * 0], a1, bd);
+    dest[dest_stride * 1] = highbd_clip_pixel_add(
+        dest[dest_stride * 1], e1, bd);
+    dest[dest_stride * 2] = highbd_clip_pixel_add(
+        dest[dest_stride * 2], e1, bd);
+    dest[dest_stride * 3] = highbd_clip_pixel_add(
+        dest[dest_stride * 3], e1, bd);
+    ip++;
+    dest++;
+  }
+}
+
+void vpx_highbd_idct4_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step[4];
+  tran_high_t temp1, temp2;
+  (void) bd;
+  // stage 1
+  temp1 = (input[0] + input[2]) * cospi_16_64;
+  temp2 = (input[0] - input[2]) * cospi_16_64;
+  step[0] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step[1] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+  temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+  step[2] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step[3] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  // stage 2
+  output[0] = HIGHBD_WRAPLOW(step[0] + step[3], bd);
+  output[1] = HIGHBD_WRAPLOW(step[1] + step[2], bd);
+  output[2] = HIGHBD_WRAPLOW(step[1] - step[2], bd);
+  output[3] = HIGHBD_WRAPLOW(step[0] - step[3], bd);
+}
+
+void vpx_highbd_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[4], temp_out[4];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  for (i = 0; i < 4; ++i) {
+    vpx_highbd_idct4_c(input, outptr, bd);
+    input += 4;
+    outptr += 4;
+  }
+
+  // Columns
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j)
+      temp_in[j] = out[j * 4 + i];
+    vpx_highbd_idct4_c(temp_in, temp_out, bd);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
+    }
+  }
+}
+
+void vpx_highbd_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int dest_stride, int bd) {
+  int i;
+  tran_high_t a1;
+  tran_low_t out = HIGHBD_WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64), bd);
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  out = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 4);
+
+  for (i = 0; i < 4; i++) {
+    dest[0] = highbd_clip_pixel_add(dest[0], a1, bd);
+    dest[1] = highbd_clip_pixel_add(dest[1], a1, bd);
+    dest[2] = highbd_clip_pixel_add(dest[2], a1, bd);
+    dest[3] = highbd_clip_pixel_add(dest[3], a1, bd);
+    dest += dest_stride;
+  }
+}
+
+void vpx_highbd_idct8_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step1[8], step2[8];
+  tran_high_t temp1, temp2;
+  // stage 1
+  step1[0] = input[0];
+  step1[2] = input[4];
+  step1[1] = input[2];
+  step1[3] = input[6];
+  temp1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+  temp2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+  step1[4] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[7] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+  temp2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+  step1[5] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[6] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  // stage 2 & stage 3 - even half
+  vpx_highbd_idct4_c(step1, step1, bd);
+
+  // stage 2 - odd half
+  step2[4] = HIGHBD_WRAPLOW(step1[4] + step1[5], bd);
+  step2[5] = HIGHBD_WRAPLOW(step1[4] - step1[5], bd);
+  step2[6] = HIGHBD_WRAPLOW(-step1[6] + step1[7], bd);
+  step2[7] = HIGHBD_WRAPLOW(step1[6] + step1[7], bd);
+
+  // stage 3 - odd half
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[6] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step1[7] = step2[7];
+
+  // stage 4
+  output[0] = HIGHBD_WRAPLOW(step1[0] + step1[7], bd);
+  output[1] = HIGHBD_WRAPLOW(step1[1] + step1[6], bd);
+  output[2] = HIGHBD_WRAPLOW(step1[2] + step1[5], bd);
+  output[3] = HIGHBD_WRAPLOW(step1[3] + step1[4], bd);
+  output[4] = HIGHBD_WRAPLOW(step1[3] - step1[4], bd);
+  output[5] = HIGHBD_WRAPLOW(step1[2] - step1[5], bd);
+  output[6] = HIGHBD_WRAPLOW(step1[1] - step1[6], bd);
+  output[7] = HIGHBD_WRAPLOW(step1[0] - step1[7], bd);
+}
+
+void vpx_highbd_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows.
+  for (i = 0; i < 8; ++i) {
+    vpx_highbd_idct8_c(input, outptr, bd);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns.
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    vpx_highbd_idct8_c(temp_in, temp_out, bd);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+    }
+  }
+}
+
+void vpx_highbd_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int bd) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = HIGHBD_WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64), bd);
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  out = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 5);
+  for (j = 0; j < 8; ++j) {
+    for (i = 0; i < 8; ++i)
+      dest[i] = highbd_clip_pixel_add(dest[i], a1, bd);
+    dest += stride;
+  }
+}
+
+void vpx_highbd_iadst4_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_low_t x0 = input[0];
+  tran_low_t x1 = input[1];
+  tran_low_t x2 = input[2];
+  tran_low_t x3 = input[3];
+  (void) bd;
+
+  if (!(x0 | x1 | x2 | x3)) {
+    memset(output, 0, 4 * sizeof(*output));
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_2_9 * x0;
+  s2 = sinpi_3_9 * x1;
+  s3 = sinpi_4_9 * x2;
+  s4 = sinpi_1_9 * x2;
+  s5 = sinpi_2_9 * x3;
+  s6 = sinpi_4_9 * x3;
+  s7 = (tran_high_t)HIGHBD_WRAPLOW(x0 - x2 + x3, bd);
+
+  s0 = s0 + s3 + s5;
+  s1 = s1 - s4 - s6;
+  s3 = s2;
+  s2 = sinpi_3_9 * s7;
+
+  // 1-D transform scaling factor is sqrt(2).
+  // The overall dynamic range is 14b (input) + 14b (multiplication scaling)
+  // + 1b (addition) = 29b.
+  // Hence the output bit depth is 15b.
+  output[0] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s0 + s3), bd);
+  output[1] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s1 + s3), bd);
+  output[2] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s2), bd);
+  output[3] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s0 + s1 - s3), bd);
+}
+
+void vpx_highbd_iadst8_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_low_t x0 = input[7];
+  tran_low_t x1 = input[0];
+  tran_low_t x2 = input[5];
+  tran_low_t x3 = input[2];
+  tran_low_t x4 = input[3];
+  tran_low_t x5 = input[4];
+  tran_low_t x6 = input[1];
+  tran_low_t x7 = input[6];
+  (void) bd;
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7)) {
+    memset(output, 0, 8 * sizeof(*output));
+    return;
+  }
+
+  // stage 1
+  s0 = cospi_2_64  * x0 + cospi_30_64 * x1;
+  s1 = cospi_30_64 * x0 - cospi_2_64  * x1;
+  s2 = cospi_10_64 * x2 + cospi_22_64 * x3;
+  s3 = cospi_22_64 * x2 - cospi_10_64 * x3;
+  s4 = cospi_18_64 * x4 + cospi_14_64 * x5;
+  s5 = cospi_14_64 * x4 - cospi_18_64 * x5;
+  s6 = cospi_26_64 * x6 + cospi_6_64  * x7;
+  s7 = cospi_6_64  * x6 - cospi_26_64 * x7;
+
+  x0 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s0 + s4), bd);
+  x1 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s1 + s5), bd);
+  x2 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s2 + s6), bd);
+  x3 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s3 + s7), bd);
+  x4 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s0 - s4), bd);
+  x5 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s1 - s5), bd);
+  x6 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s2 - s6), bd);
+  x7 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s3 - s7), bd);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 =  cospi_8_64  * x4 + cospi_24_64 * x5;
+  s5 =  cospi_24_64 * x4 - cospi_8_64  * x5;
+  s6 = -cospi_24_64 * x6 + cospi_8_64  * x7;
+  s7 =  cospi_8_64  * x6 + cospi_24_64 * x7;
+
+  x0 = HIGHBD_WRAPLOW(s0 + s2, bd);
+  x1 = HIGHBD_WRAPLOW(s1 + s3, bd);
+  x2 = HIGHBD_WRAPLOW(s0 - s2, bd);
+  x3 = HIGHBD_WRAPLOW(s1 - s3, bd);
+  x4 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s4 + s6), bd);
+  x5 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s5 + s7), bd);
+  x6 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s4 - s6), bd);
+  x7 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s5 - s7), bd);
+
+  // stage 3
+  s2 = cospi_16_64 * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (x6 - x7);
+
+  x2 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s2), bd);
+  x3 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s3), bd);
+  x6 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s6), bd);
+  x7 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s7), bd);
+
+  output[0] = HIGHBD_WRAPLOW(x0, bd);
+  output[1] = HIGHBD_WRAPLOW(-x4, bd);
+  output[2] = HIGHBD_WRAPLOW(x6, bd);
+  output[3] = HIGHBD_WRAPLOW(-x2, bd);
+  output[4] = HIGHBD_WRAPLOW(x3, bd);
+  output[5] = HIGHBD_WRAPLOW(-x7, bd);
+  output[6] = HIGHBD_WRAPLOW(x5, bd);
+  output[7] = HIGHBD_WRAPLOW(-x1, bd);
+}
+
+void vpx_highbd_idct8x8_10_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  tran_low_t out[8 * 8] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows.
+  // Only first 4 row has non-zero coefs.
+  for (i = 0; i < 4; ++i) {
+    vpx_highbd_idct8_c(input, outptr, bd);
+    input += 8;
+    outptr += 8;
+  }
+  // Then transform columns.
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    vpx_highbd_idct8_c(temp_in, temp_out, bd);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+    }
+  }
+}
+
+void vpx_highbd_idct16_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step1[16], step2[16];
+  tran_high_t temp1, temp2;
+  (void) bd;
+
+  // stage 1
+  step1[0] = input[0/2];
+  step1[1] = input[16/2];
+  step1[2] = input[8/2];
+  step1[3] = input[24/2];
+  step1[4] = input[4/2];
+  step1[5] = input[20/2];
+  step1[6] = input[12/2];
+  step1[7] = input[28/2];
+  step1[8] = input[2/2];
+  step1[9] = input[18/2];
+  step1[10] = input[10/2];
+  step1[11] = input[26/2];
+  step1[12] = input[6/2];
+  step1[13] = input[22/2];
+  step1[14] = input[14/2];
+  step1[15] = input[30/2];
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[15] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[14] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[13] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[12] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[7] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[6] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  step1[8] = HIGHBD_WRAPLOW(step2[8] + step2[9], bd);
+  step1[9] = HIGHBD_WRAPLOW(step2[8] - step2[9], bd);
+  step1[10] = HIGHBD_WRAPLOW(-step2[10] + step2[11], bd);
+  step1[11] = HIGHBD_WRAPLOW(step2[10] + step2[11], bd);
+  step1[12] = HIGHBD_WRAPLOW(step2[12] + step2[13], bd);
+  step1[13] = HIGHBD_WRAPLOW(step2[12] - step2[13], bd);
+  step1[14] = HIGHBD_WRAPLOW(-step2[14] + step2[15], bd);
+  step1[15] = HIGHBD_WRAPLOW(step2[14] + step2[15], bd);
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[1] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[3] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step2[4] = HIGHBD_WRAPLOW(step1[4] + step1[5], bd);
+  step2[5] = HIGHBD_WRAPLOW(step1[4] - step1[5], bd);
+  step2[6] = HIGHBD_WRAPLOW(-step1[6] + step1[7], bd);
+  step2[7] = HIGHBD_WRAPLOW(step1[6] + step1[7], bd);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[14] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[13] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  // stage 5
+  step1[0] = HIGHBD_WRAPLOW(step2[0] + step2[3], bd);
+  step1[1] = HIGHBD_WRAPLOW(step2[1] + step2[2], bd);
+  step1[2] = HIGHBD_WRAPLOW(step2[1] - step2[2], bd);
+  step1[3] = HIGHBD_WRAPLOW(step2[0] - step2[3], bd);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[6] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step1[7] = step2[7];
+
+  step1[8] = HIGHBD_WRAPLOW(step2[8] + step2[11], bd);
+  step1[9] = HIGHBD_WRAPLOW(step2[9] + step2[10], bd);
+  step1[10] = HIGHBD_WRAPLOW(step2[9] - step2[10], bd);
+  step1[11] = HIGHBD_WRAPLOW(step2[8] - step2[11], bd);
+  step1[12] = HIGHBD_WRAPLOW(-step2[12] + step2[15], bd);
+  step1[13] = HIGHBD_WRAPLOW(-step2[13] + step2[14], bd);
+  step1[14] = HIGHBD_WRAPLOW(step2[13] + step2[14], bd);
+  step1[15] = HIGHBD_WRAPLOW(step2[12] + step2[15], bd);
+
+  // stage 6
+  step2[0] = HIGHBD_WRAPLOW(step1[0] + step1[7], bd);
+  step2[1] = HIGHBD_WRAPLOW(step1[1] + step1[6], bd);
+  step2[2] = HIGHBD_WRAPLOW(step1[2] + step1[5], bd);
+  step2[3] = HIGHBD_WRAPLOW(step1[3] + step1[4], bd);
+  step2[4] = HIGHBD_WRAPLOW(step1[3] - step1[4], bd);
+  step2[5] = HIGHBD_WRAPLOW(step1[2] - step1[5], bd);
+  step2[6] = HIGHBD_WRAPLOW(step1[1] - step1[6], bd);
+  step2[7] = HIGHBD_WRAPLOW(step1[0] - step1[7], bd);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[13] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[12] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  // stage 7
+  output[0] = HIGHBD_WRAPLOW(step2[0] + step2[15], bd);
+  output[1] = HIGHBD_WRAPLOW(step2[1] + step2[14], bd);
+  output[2] = HIGHBD_WRAPLOW(step2[2] + step2[13], bd);
+  output[3] = HIGHBD_WRAPLOW(step2[3] + step2[12], bd);
+  output[4] = HIGHBD_WRAPLOW(step2[4] + step2[11], bd);
+  output[5] = HIGHBD_WRAPLOW(step2[5] + step2[10], bd);
+  output[6] = HIGHBD_WRAPLOW(step2[6] + step2[9], bd);
+  output[7] = HIGHBD_WRAPLOW(step2[7] + step2[8], bd);
+  output[8] = HIGHBD_WRAPLOW(step2[7] - step2[8], bd);
+  output[9] = HIGHBD_WRAPLOW(step2[6] - step2[9], bd);
+  output[10] = HIGHBD_WRAPLOW(step2[5] - step2[10], bd);
+  output[11] = HIGHBD_WRAPLOW(step2[4] - step2[11], bd);
+  output[12] = HIGHBD_WRAPLOW(step2[3] - step2[12], bd);
+  output[13] = HIGHBD_WRAPLOW(step2[2] - step2[13], bd);
+  output[14] = HIGHBD_WRAPLOW(step2[1] - step2[14], bd);
+  output[15] = HIGHBD_WRAPLOW(step2[0] - step2[15], bd);
+}
+
+void vpx_highbd_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows.
+  for (i = 0; i < 16; ++i) {
+    vpx_highbd_idct16_c(input, outptr, bd);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns.
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j * 16 + i];
+    vpx_highbd_idct16_c(temp_in, temp_out, bd);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vpx_highbd_iadst16_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7, s8;
+  tran_high_t s9, s10, s11, s12, s13, s14, s15;
+
+  tran_low_t x0 = input[15];
+  tran_low_t x1 = input[0];
+  tran_low_t x2 = input[13];
+  tran_low_t x3 = input[2];
+  tran_low_t x4 = input[11];
+  tran_low_t x5 = input[4];
+  tran_low_t x6 = input[9];
+  tran_low_t x7 = input[6];
+  tran_low_t x8 = input[7];
+  tran_low_t x9 = input[8];
+  tran_low_t x10 = input[5];
+  tran_low_t x11 = input[10];
+  tran_low_t x12 = input[3];
+  tran_low_t x13 = input[12];
+  tran_low_t x14 = input[1];
+  tran_low_t x15 = input[14];
+  (void) bd;
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7 | x8
+           | x9 | x10 | x11 | x12 | x13 | x14 | x15)) {
+    memset(output, 0, 16 * sizeof(*output));
+    return;
+  }
+
+  // stage 1
+  s0 = x0 * cospi_1_64  + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64  + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64  + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64  - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64  - x15 * cospi_29_64;
+
+  x0 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s0 + s8), bd);
+  x1 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s1 + s9), bd);
+  x2 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s2 + s10), bd);
+  x3 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s3 + s11), bd);
+  x4 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s4 + s12), bd);
+  x5 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s5 + s13), bd);
+  x6 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s6 + s14), bd);
+  x7 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s7 + s15), bd);
+  x8  = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s0 - s8), bd);
+  x9  = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s1 - s9), bd);
+  x10 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s2 - s10), bd);
+  x11 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s3 - s11), bd);
+  x12 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s4 - s12), bd);
+  x13 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s5 - s13), bd);
+  x14 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s6 - s14), bd);
+  x15 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s7 - s15), bd);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 = x8 * cospi_4_64 + x9 * cospi_28_64;
+  s9 = x8 * cospi_28_64 - x9 * cospi_4_64;
+  s10 = x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 = x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = -x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 = x12 * cospi_4_64 + x13 * cospi_28_64;
+  s14 = -x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 = x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = HIGHBD_WRAPLOW(s0 + s4, bd);
+  x1 = HIGHBD_WRAPLOW(s1 + s5, bd);
+  x2 = HIGHBD_WRAPLOW(s2 + s6, bd);
+  x3 = HIGHBD_WRAPLOW(s3 + s7, bd);
+  x4 = HIGHBD_WRAPLOW(s0 - s4, bd);
+  x5 = HIGHBD_WRAPLOW(s1 - s5, bd);
+  x6 = HIGHBD_WRAPLOW(s2 - s6, bd);
+  x7 = HIGHBD_WRAPLOW(s3 - s7, bd);
+  x8 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s8 + s12), bd);
+  x9 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s9 + s13), bd);
+  x10 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s10 + s14), bd);
+  x11 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s11 + s15), bd);
+  x12 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s8 - s12), bd);
+  x13 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s9 - s13), bd);
+  x14 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s10 - s14), bd);
+  x15 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s11 - s15), bd);
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64 + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = -x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 = x6 * cospi_8_64 + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64 + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = -x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 = x14 * cospi_8_64 + x15 * cospi_24_64;
+
+  x0 = HIGHBD_WRAPLOW(s0 + s2, bd);
+  x1 = HIGHBD_WRAPLOW(s1 + s3, bd);
+  x2 = HIGHBD_WRAPLOW(s0 - s2, bd);
+  x3 = HIGHBD_WRAPLOW(s1 - s3, bd);
+  x4 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s4 + s6), bd);
+  x5 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s5 + s7), bd);
+  x6 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s4 - s6), bd);
+  x7 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s5 - s7), bd);
+  x8 = HIGHBD_WRAPLOW(s8 + s10, bd);
+  x9 = HIGHBD_WRAPLOW(s9 + s11, bd);
+  x10 = HIGHBD_WRAPLOW(s8 - s10, bd);
+  x11 = HIGHBD_WRAPLOW(s9 - s11, bd);
+  x12 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s12 + s14), bd);
+  x13 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s13 + s15), bd);
+  x14 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s12 - s14), bd);
+  x15 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s13 - s15), bd);
+
+  // stage 4
+  s2 = (- cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (-x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (-x10 + x11);
+  s14 = (- cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s2), bd);
+  x3 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s3), bd);
+  x6 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s6), bd);
+  x7 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s7), bd);
+  x10 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s10), bd);
+  x11 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s11), bd);
+  x14 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s14), bd);
+  x15 = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(s15), bd);
+
+  output[0] = HIGHBD_WRAPLOW(x0, bd);
+  output[1] = HIGHBD_WRAPLOW(-x8, bd);
+  output[2] = HIGHBD_WRAPLOW(x12, bd);
+  output[3] = HIGHBD_WRAPLOW(-x4, bd);
+  output[4] = HIGHBD_WRAPLOW(x6, bd);
+  output[5] = HIGHBD_WRAPLOW(x14, bd);
+  output[6] = HIGHBD_WRAPLOW(x10, bd);
+  output[7] = HIGHBD_WRAPLOW(x2, bd);
+  output[8] = HIGHBD_WRAPLOW(x3, bd);
+  output[9] = HIGHBD_WRAPLOW(x11, bd);
+  output[10] = HIGHBD_WRAPLOW(x15, bd);
+  output[11] = HIGHBD_WRAPLOW(x7, bd);
+  output[12] = HIGHBD_WRAPLOW(x5, bd);
+  output[13] = HIGHBD_WRAPLOW(-x13, bd);
+  output[14] = HIGHBD_WRAPLOW(x9, bd);
+  output[15] = HIGHBD_WRAPLOW(-x1, bd);
+}
+
+void vpx_highbd_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int bd) {
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows. Since all non-zero dct coefficients are in
+  // upper-left 4x4 area, we only need to calculate first 4 rows here.
+  for (i = 0; i < 4; ++i) {
+    vpx_highbd_idct16_c(input, outptr, bd);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns.
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j*16 + i];
+    vpx_highbd_idct16_c(temp_in, temp_out, bd);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vpx_highbd_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int bd) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = HIGHBD_WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64), bd);
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  out = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+  for (j = 0; j < 16; ++j) {
+    for (i = 0; i < 16; ++i)
+      dest[i] = highbd_clip_pixel_add(dest[i], a1, bd);
+    dest += stride;
+  }
+}
+
+static void highbd_idct32_c(const tran_low_t *input,
+                            tran_low_t *output, int bd) {
+  tran_low_t step1[32], step2[32];
+  tran_high_t temp1, temp2;
+  (void) bd;
+
+  // stage 1
+  step1[0] = input[0];
+  step1[1] = input[16];
+  step1[2] = input[8];
+  step1[3] = input[24];
+  step1[4] = input[4];
+  step1[5] = input[20];
+  step1[6] = input[12];
+  step1[7] = input[28];
+  step1[8] = input[2];
+  step1[9] = input[18];
+  step1[10] = input[10];
+  step1[11] = input[26];
+  step1[12] = input[6];
+  step1[13] = input[22];
+  step1[14] = input[14];
+  step1[15] = input[30];
+
+  temp1 = input[1] * cospi_31_64 - input[31] * cospi_1_64;
+  temp2 = input[1] * cospi_1_64 + input[31] * cospi_31_64;
+  step1[16] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[31] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = input[17] * cospi_15_64 - input[15] * cospi_17_64;
+  temp2 = input[17] * cospi_17_64 + input[15] * cospi_15_64;
+  step1[17] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[30] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = input[9] * cospi_23_64 - input[23] * cospi_9_64;
+  temp2 = input[9] * cospi_9_64 + input[23] * cospi_23_64;
+  step1[18] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[29] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = input[25] * cospi_7_64 - input[7] * cospi_25_64;
+  temp2 = input[25] * cospi_25_64 + input[7] * cospi_7_64;
+  step1[19] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[28] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = input[5] * cospi_27_64 - input[27] * cospi_5_64;
+  temp2 = input[5] * cospi_5_64 + input[27] * cospi_27_64;
+  step1[20] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[27] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = input[21] * cospi_11_64 - input[11] * cospi_21_64;
+  temp2 = input[21] * cospi_21_64 + input[11] * cospi_11_64;
+  step1[21] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[26] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = input[13] * cospi_19_64 - input[19] * cospi_13_64;
+  temp2 = input[13] * cospi_13_64 + input[19] * cospi_19_64;
+  step1[22] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[25] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = input[29] * cospi_3_64 - input[3] * cospi_29_64;
+  temp2 = input[29] * cospi_29_64 + input[3] * cospi_3_64;
+  step1[23] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[24] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[15] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[14] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[13] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[12] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  step2[16] = HIGHBD_WRAPLOW(step1[16] + step1[17], bd);
+  step2[17] = HIGHBD_WRAPLOW(step1[16] - step1[17], bd);
+  step2[18] = HIGHBD_WRAPLOW(-step1[18] + step1[19], bd);
+  step2[19] = HIGHBD_WRAPLOW(step1[18] + step1[19], bd);
+  step2[20] = HIGHBD_WRAPLOW(step1[20] + step1[21], bd);
+  step2[21] = HIGHBD_WRAPLOW(step1[20] - step1[21], bd);
+  step2[22] = HIGHBD_WRAPLOW(-step1[22] + step1[23], bd);
+  step2[23] = HIGHBD_WRAPLOW(step1[22] + step1[23], bd);
+  step2[24] = HIGHBD_WRAPLOW(step1[24] + step1[25], bd);
+  step2[25] = HIGHBD_WRAPLOW(step1[24] - step1[25], bd);
+  step2[26] = HIGHBD_WRAPLOW(-step1[26] + step1[27], bd);
+  step2[27] = HIGHBD_WRAPLOW(step1[26] + step1[27], bd);
+  step2[28] = HIGHBD_WRAPLOW(step1[28] + step1[29], bd);
+  step2[29] = HIGHBD_WRAPLOW(step1[28] - step1[29], bd);
+  step2[30] = HIGHBD_WRAPLOW(-step1[30] + step1[31], bd);
+  step2[31] = HIGHBD_WRAPLOW(step1[30] + step1[31], bd);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[7] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[6] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+
+  step1[8] = HIGHBD_WRAPLOW(step2[8] + step2[9], bd);
+  step1[9] = HIGHBD_WRAPLOW(step2[8] - step2[9], bd);
+  step1[10] = HIGHBD_WRAPLOW(-step2[10] + step2[11], bd);
+  step1[11] = HIGHBD_WRAPLOW(step2[10] + step2[11], bd);
+  step1[12] = HIGHBD_WRAPLOW(step2[12] + step2[13], bd);
+  step1[13] = HIGHBD_WRAPLOW(step2[12] - step2[13], bd);
+  step1[14] = HIGHBD_WRAPLOW(-step2[14] + step2[15], bd);
+  step1[15] = HIGHBD_WRAPLOW(step2[14] + step2[15], bd);
+
+  step1[16] = step2[16];
+  step1[31] = step2[31];
+  temp1 = -step2[17] * cospi_4_64 + step2[30] * cospi_28_64;
+  temp2 = step2[17] * cospi_28_64 + step2[30] * cospi_4_64;
+  step1[17] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[30] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = -step2[18] * cospi_28_64 - step2[29] * cospi_4_64;
+  temp2 = -step2[18] * cospi_4_64 + step2[29] * cospi_28_64;
+  step1[18] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[29] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step1[19] = step2[19];
+  step1[20] = step2[20];
+  temp1 = -step2[21] * cospi_20_64 + step2[26] * cospi_12_64;
+  temp2 = step2[21] * cospi_12_64 + step2[26] * cospi_20_64;
+  step1[21] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[26] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = -step2[22] * cospi_12_64 - step2[25] * cospi_20_64;
+  temp2 = -step2[22] * cospi_20_64 + step2[25] * cospi_12_64;
+  step1[22] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[25] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[27] = step2[27];
+  step1[28] = step2[28];
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[1] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[3] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step2[4] = HIGHBD_WRAPLOW(step1[4] + step1[5], bd);
+  step2[5] = HIGHBD_WRAPLOW(step1[4] - step1[5], bd);
+  step2[6] = HIGHBD_WRAPLOW(-step1[6] + step1[7], bd);
+  step2[7] = HIGHBD_WRAPLOW(step1[6] + step1[7], bd);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[14] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[13] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  step2[16] = HIGHBD_WRAPLOW(step1[16] + step1[19], bd);
+  step2[17] = HIGHBD_WRAPLOW(step1[17] + step1[18], bd);
+  step2[18] = HIGHBD_WRAPLOW(step1[17] - step1[18], bd);
+  step2[19] = HIGHBD_WRAPLOW(step1[16] - step1[19], bd);
+  step2[20] = HIGHBD_WRAPLOW(-step1[20] + step1[23], bd);
+  step2[21] = HIGHBD_WRAPLOW(-step1[21] + step1[22], bd);
+  step2[22] = HIGHBD_WRAPLOW(step1[21] + step1[22], bd);
+  step2[23] = HIGHBD_WRAPLOW(step1[20] + step1[23], bd);
+
+  step2[24] = HIGHBD_WRAPLOW(step1[24] + step1[27], bd);
+  step2[25] = HIGHBD_WRAPLOW(step1[25] + step1[26], bd);
+  step2[26] = HIGHBD_WRAPLOW(step1[25] - step1[26], bd);
+  step2[27] = HIGHBD_WRAPLOW(step1[24] - step1[27], bd);
+  step2[28] = HIGHBD_WRAPLOW(-step1[28] + step1[31], bd);
+  step2[29] = HIGHBD_WRAPLOW(-step1[29] + step1[30], bd);
+  step2[30] = HIGHBD_WRAPLOW(step1[29] + step1[30], bd);
+  step2[31] = HIGHBD_WRAPLOW(step1[28] + step1[31], bd);
+
+  // stage 5
+  step1[0] = HIGHBD_WRAPLOW(step2[0] + step2[3], bd);
+  step1[1] = HIGHBD_WRAPLOW(step2[1] + step2[2], bd);
+  step1[2] = HIGHBD_WRAPLOW(step2[1] - step2[2], bd);
+  step1[3] = HIGHBD_WRAPLOW(step2[0] - step2[3], bd);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[6] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step1[7] = step2[7];
+
+  step1[8] = HIGHBD_WRAPLOW(step2[8] + step2[11], bd);
+  step1[9] = HIGHBD_WRAPLOW(step2[9] + step2[10], bd);
+  step1[10] = HIGHBD_WRAPLOW(step2[9] - step2[10], bd);
+  step1[11] = HIGHBD_WRAPLOW(step2[8] - step2[11], bd);
+  step1[12] = HIGHBD_WRAPLOW(-step2[12] + step2[15], bd);
+  step1[13] = HIGHBD_WRAPLOW(-step2[13] + step2[14], bd);
+  step1[14] = HIGHBD_WRAPLOW(step2[13] + step2[14], bd);
+  step1[15] = HIGHBD_WRAPLOW(step2[12] + step2[15], bd);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  temp1 = -step2[18] * cospi_8_64 + step2[29] * cospi_24_64;
+  temp2 = step2[18] * cospi_24_64 + step2[29] * cospi_8_64;
+  step1[18] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[29] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = -step2[19] * cospi_8_64 + step2[28] * cospi_24_64;
+  temp2 = step2[19] * cospi_24_64 + step2[28] * cospi_8_64;
+  step1[19] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[28] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = -step2[20] * cospi_24_64 - step2[27] * cospi_8_64;
+  temp2 = -step2[20] * cospi_8_64 + step2[27] * cospi_24_64;
+  step1[20] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[27] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = -step2[21] * cospi_24_64 - step2[26] * cospi_8_64;
+  temp2 = -step2[21] * cospi_8_64 + step2[26] * cospi_24_64;
+  step1[21] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[26] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step1[22] = step2[22];
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[25] = step2[25];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // stage 6
+  step2[0] = HIGHBD_WRAPLOW(step1[0] + step1[7], bd);
+  step2[1] = HIGHBD_WRAPLOW(step1[1] + step1[6], bd);
+  step2[2] = HIGHBD_WRAPLOW(step1[2] + step1[5], bd);
+  step2[3] = HIGHBD_WRAPLOW(step1[3] + step1[4], bd);
+  step2[4] = HIGHBD_WRAPLOW(step1[3] - step1[4], bd);
+  step2[5] = HIGHBD_WRAPLOW(step1[2] - step1[5], bd);
+  step2[6] = HIGHBD_WRAPLOW(step1[1] - step1[6], bd);
+  step2[7] = HIGHBD_WRAPLOW(step1[0] - step1[7], bd);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[13] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step2[12] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  step2[16] = HIGHBD_WRAPLOW(step1[16] + step1[23], bd);
+  step2[17] = HIGHBD_WRAPLOW(step1[17] + step1[22], bd);
+  step2[18] = HIGHBD_WRAPLOW(step1[18] + step1[21], bd);
+  step2[19] = HIGHBD_WRAPLOW(step1[19] + step1[20], bd);
+  step2[20] = HIGHBD_WRAPLOW(step1[19] - step1[20], bd);
+  step2[21] = HIGHBD_WRAPLOW(step1[18] - step1[21], bd);
+  step2[22] = HIGHBD_WRAPLOW(step1[17] - step1[22], bd);
+  step2[23] = HIGHBD_WRAPLOW(step1[16] - step1[23], bd);
+
+  step2[24] = HIGHBD_WRAPLOW(-step1[24] + step1[31], bd);
+  step2[25] = HIGHBD_WRAPLOW(-step1[25] + step1[30], bd);
+  step2[26] = HIGHBD_WRAPLOW(-step1[26] + step1[29], bd);
+  step2[27] = HIGHBD_WRAPLOW(-step1[27] + step1[28], bd);
+  step2[28] = HIGHBD_WRAPLOW(step1[27] + step1[28], bd);
+  step2[29] = HIGHBD_WRAPLOW(step1[26] + step1[29], bd);
+  step2[30] = HIGHBD_WRAPLOW(step1[25] + step1[30], bd);
+  step2[31] = HIGHBD_WRAPLOW(step1[24] + step1[31], bd);
+
+  // stage 7
+  step1[0] = HIGHBD_WRAPLOW(step2[0] + step2[15], bd);
+  step1[1] = HIGHBD_WRAPLOW(step2[1] + step2[14], bd);
+  step1[2] = HIGHBD_WRAPLOW(step2[2] + step2[13], bd);
+  step1[3] = HIGHBD_WRAPLOW(step2[3] + step2[12], bd);
+  step1[4] = HIGHBD_WRAPLOW(step2[4] + step2[11], bd);
+  step1[5] = HIGHBD_WRAPLOW(step2[5] + step2[10], bd);
+  step1[6] = HIGHBD_WRAPLOW(step2[6] + step2[9], bd);
+  step1[7] = HIGHBD_WRAPLOW(step2[7] + step2[8], bd);
+  step1[8] = HIGHBD_WRAPLOW(step2[7] - step2[8], bd);
+  step1[9] = HIGHBD_WRAPLOW(step2[6] - step2[9], bd);
+  step1[10] = HIGHBD_WRAPLOW(step2[5] - step2[10], bd);
+  step1[11] = HIGHBD_WRAPLOW(step2[4] - step2[11], bd);
+  step1[12] = HIGHBD_WRAPLOW(step2[3] - step2[12], bd);
+  step1[13] = HIGHBD_WRAPLOW(step2[2] - step2[13], bd);
+  step1[14] = HIGHBD_WRAPLOW(step2[1] - step2[14], bd);
+  step1[15] = HIGHBD_WRAPLOW(step2[0] - step2[15], bd);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  step1[18] = step2[18];
+  step1[19] = step2[19];
+  temp1 = (-step2[20] + step2[27]) * cospi_16_64;
+  temp2 = (step2[20] + step2[27]) * cospi_16_64;
+  step1[20] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[27] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = (-step2[21] + step2[26]) * cospi_16_64;
+  temp2 = (step2[21] + step2[26]) * cospi_16_64;
+  step1[21] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[26] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = (-step2[22] + step2[25]) * cospi_16_64;
+  temp2 = (step2[22] + step2[25]) * cospi_16_64;
+  step1[22] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[25] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  temp1 = (-step2[23] + step2[24]) * cospi_16_64;
+  temp2 = (step2[23] + step2[24]) * cospi_16_64;
+  step1[23] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp1), bd);
+  step1[24] = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(temp2), bd);
+  step1[28] = step2[28];
+  step1[29] = step2[29];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // final stage
+  output[0] = HIGHBD_WRAPLOW(step1[0] + step1[31], bd);
+  output[1] = HIGHBD_WRAPLOW(step1[1] + step1[30], bd);
+  output[2] = HIGHBD_WRAPLOW(step1[2] + step1[29], bd);
+  output[3] = HIGHBD_WRAPLOW(step1[3] + step1[28], bd);
+  output[4] = HIGHBD_WRAPLOW(step1[4] + step1[27], bd);
+  output[5] = HIGHBD_WRAPLOW(step1[5] + step1[26], bd);
+  output[6] = HIGHBD_WRAPLOW(step1[6] + step1[25], bd);
+  output[7] = HIGHBD_WRAPLOW(step1[7] + step1[24], bd);
+  output[8] = HIGHBD_WRAPLOW(step1[8] + step1[23], bd);
+  output[9] = HIGHBD_WRAPLOW(step1[9] + step1[22], bd);
+  output[10] = HIGHBD_WRAPLOW(step1[10] + step1[21], bd);
+  output[11] = HIGHBD_WRAPLOW(step1[11] + step1[20], bd);
+  output[12] = HIGHBD_WRAPLOW(step1[12] + step1[19], bd);
+  output[13] = HIGHBD_WRAPLOW(step1[13] + step1[18], bd);
+  output[14] = HIGHBD_WRAPLOW(step1[14] + step1[17], bd);
+  output[15] = HIGHBD_WRAPLOW(step1[15] + step1[16], bd);
+  output[16] = HIGHBD_WRAPLOW(step1[15] - step1[16], bd);
+  output[17] = HIGHBD_WRAPLOW(step1[14] - step1[17], bd);
+  output[18] = HIGHBD_WRAPLOW(step1[13] - step1[18], bd);
+  output[19] = HIGHBD_WRAPLOW(step1[12] - step1[19], bd);
+  output[20] = HIGHBD_WRAPLOW(step1[11] - step1[20], bd);
+  output[21] = HIGHBD_WRAPLOW(step1[10] - step1[21], bd);
+  output[22] = HIGHBD_WRAPLOW(step1[9] - step1[22], bd);
+  output[23] = HIGHBD_WRAPLOW(step1[8] - step1[23], bd);
+  output[24] = HIGHBD_WRAPLOW(step1[7] - step1[24], bd);
+  output[25] = HIGHBD_WRAPLOW(step1[6] - step1[25], bd);
+  output[26] = HIGHBD_WRAPLOW(step1[5] - step1[26], bd);
+  output[27] = HIGHBD_WRAPLOW(step1[4] - step1[27], bd);
+  output[28] = HIGHBD_WRAPLOW(step1[3] - step1[28], bd);
+  output[29] = HIGHBD_WRAPLOW(step1[2] - step1[29], bd);
+  output[30] = HIGHBD_WRAPLOW(step1[1] - step1[30], bd);
+  output[31] = HIGHBD_WRAPLOW(step1[0] - step1[31], bd);
+}
+
+void vpx_highbd_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest8,
+                                     int stride, int bd) {
+  tran_low_t out[32 * 32];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    tran_low_t zero_coeff[16];
+    for (j = 0; j < 16; ++j)
+      zero_coeff[j] = input[2 * j] | input[2 * j + 1];
+    for (j = 0; j < 8; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 4; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 2; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+
+    if (zero_coeff[0] | zero_coeff[1])
+      highbd_idct32_c(input, outptr, bd);
+    else
+      memset(outptr, 0, sizeof(tran_low_t) * 32);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    highbd_idct32_c(temp_in, temp_out, bd);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vpx_highbd_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int bd) {
+  tran_low_t out[32 * 32] = {0};
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  // Only upper-left 8x8 has non-zero coeff.
+  for (i = 0; i < 8; ++i) {
+    highbd_idct32_c(input, outptr, bd);
+    input += 32;
+    outptr += 32;
+  }
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    highbd_idct32_c(temp_in, temp_out, bd);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vpx_highbd_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int bd) {
+  int i, j;
+  int a1;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  tran_low_t out = HIGHBD_WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64), bd);
+  out = HIGHBD_WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+
+  for (j = 0; j < 32; ++j) {
+    for (i = 0; i < 32; ++i)
+      dest[i] = highbd_clip_pixel_add(dest[i], a1, bd);
+    dest += stride;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vpx_dsp/inv_txfm.h b/libvpx/libvpx/vpx_dsp/inv_txfm.h
new file mode 100644
index 0000000..9cfe1be
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/inv_txfm.h
@@ -0,0 +1,133 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_INV_TXFM_H_
+#define VPX_DSP_INV_TXFM_H_
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx_dsp/txfm_common.h"
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE tran_high_t check_range(tran_high_t input) {
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+  // For valid VP9 input streams, intermediate stage coefficients should always
+  // stay within the range of a signed 16 bit integer. Coefficients can go out
+  // of this range for invalid/corrupt VP9 streams. However, strictly checking
+  // this range for every intermediate coefficient can burdensome for a decoder,
+  // therefore the following assertion is only enabled when configured with
+  // --enable-coefficient-range-checking.
+  assert(INT16_MIN <= input);
+  assert(input <= INT16_MAX);
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+  return input;
+}
+
+static INLINE tran_high_t dct_const_round_shift(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  return (tran_high_t)rv;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE tran_high_t highbd_check_range(tran_high_t input,
+                                             int bd) {
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+  // For valid highbitdepth VP9 streams, intermediate stage coefficients will
+  // stay within the ranges:
+  // - 8 bit: signed 16 bit integer
+  // - 10 bit: signed 18 bit integer
+  // - 12 bit: signed 20 bit integer
+  const int32_t int_max = (1 << (7 + bd)) - 1;
+  const int32_t int_min = -int_max - 1;
+  assert(int_min <= input);
+  assert(input <= int_max);
+  (void) int_min;
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+  (void) bd;
+  return input;
+}
+
+static INLINE tran_high_t highbd_dct_const_round_shift(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  return (tran_high_t)rv;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_EMULATE_HARDWARE
+// When CONFIG_EMULATE_HARDWARE is 1 the transform performs a
+// non-normative method to handle overflows. A stream that causes
+// overflows  in the inverse transform is considered invalid in VP9,
+// and a hardware implementer is free to choose any reasonable
+// method to handle overflows. However to aid in hardware
+// verification they can use a specific implementation of the
+// WRAPLOW() macro below that is identical to their intended
+// hardware implementation (and also use configure options to trigger
+// the C-implementation of the transform).
+//
+// The particular WRAPLOW implementation below performs strict
+// overflow wrapping to match common hardware implementations.
+// bd of 8 uses trans_low with 16bits, need to remove 16bits
+// bd of 10 uses trans_low with 18bits, need to remove 14bits
+// bd of 12 uses trans_low with 20bits, need to remove 12bits
+// bd of x uses trans_low with 8+x bits, need to remove 24-x bits
+
+#define WRAPLOW(x) ((((int32_t)check_range(x)) << 16) >> 16)
+#if CONFIG_VP9_HIGHBITDEPTH
+#define HIGHBD_WRAPLOW(x, bd) \
+    ((((int32_t)highbd_check_range((x), bd)) << (24 - bd)) >> (24 - bd))
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#else   // CONFIG_EMULATE_HARDWARE
+
+#define WRAPLOW(x) ((int32_t)check_range(x))
+#if CONFIG_VP9_HIGHBITDEPTH
+#define HIGHBD_WRAPLOW(x, bd) \
+    ((int32_t)highbd_check_range((x), bd))
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#endif  // CONFIG_EMULATE_HARDWARE
+
+void idct4_c(const tran_low_t *input, tran_low_t *output);
+void idct8_c(const tran_low_t *input, tran_low_t *output);
+void idct16_c(const tran_low_t *input, tran_low_t *output);
+void idct32_c(const tran_low_t *input, tran_low_t *output);
+void iadst4_c(const tran_low_t *input, tran_low_t *output);
+void iadst8_c(const tran_low_t *input, tran_low_t *output);
+void iadst16_c(const tran_low_t *input, tran_low_t *output);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_idct4_c(const tran_low_t *input, tran_low_t *output, int bd);
+void vpx_highbd_idct8_c(const tran_low_t *input, tran_low_t *output, int bd);
+void vpx_highbd_idct16_c(const tran_low_t *input, tran_low_t *output, int bd);
+
+void vpx_highbd_iadst4_c(const tran_low_t *input, tran_low_t *output, int bd);
+void vpx_highbd_iadst8_c(const tran_low_t *input, tran_low_t *output, int bd);
+void vpx_highbd_iadst16_c(const tran_low_t *input, tran_low_t *output, int bd);
+
+static INLINE uint16_t highbd_clip_pixel_add(uint16_t dest, tran_high_t trans,
+                                             int bd) {
+  trans = HIGHBD_WRAPLOW(trans, bd);
+  return clip_pixel_highbd(dest + (int)trans, bd);
+}
+#endif
+
+static INLINE uint8_t clip_pixel_add(uint8_t dest, tran_high_t trans) {
+  trans = WRAPLOW(trans);
+  return clip_pixel(dest + (int)trans);
+}
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_INV_TXFM_H_
diff --git a/libvpx/libvpx/vpx_dsp/loopfilter.c b/libvpx/libvpx/vpx_dsp/loopfilter.c
new file mode 100644
index 0000000..645a1ab
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/loopfilter.c
@@ -0,0 +1,767 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+
+static INLINE int8_t signed_char_clamp(int t) {
+  return (int8_t)clamp(t, -128, 127);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE int16_t signed_char_clamp_high(int t, int bd) {
+  switch (bd) {
+    case 10:
+      return (int16_t)clamp(t, -128*4, 128*4-1);
+    case 12:
+      return (int16_t)clamp(t, -128*16, 128*16-1);
+    case 8:
+    default:
+      return (int16_t)clamp(t, -128, 128-1);
+  }
+}
+#endif
+
+// should we apply any filter at all: 11111111 yes, 00000000 no
+static INLINE int8_t filter_mask(uint8_t limit, uint8_t blimit,
+                                 uint8_t p3, uint8_t p2,
+                                 uint8_t p1, uint8_t p0,
+                                 uint8_t q0, uint8_t q1,
+                                 uint8_t q2, uint8_t q3) {
+  int8_t mask = 0;
+  mask |= (abs(p3 - p2) > limit) * -1;
+  mask |= (abs(p2 - p1) > limit) * -1;
+  mask |= (abs(p1 - p0) > limit) * -1;
+  mask |= (abs(q1 - q0) > limit) * -1;
+  mask |= (abs(q2 - q1) > limit) * -1;
+  mask |= (abs(q3 - q2) > limit) * -1;
+  mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t flat_mask4(uint8_t thresh,
+                                uint8_t p3, uint8_t p2,
+                                uint8_t p1, uint8_t p0,
+                                uint8_t q0, uint8_t q1,
+                                uint8_t q2, uint8_t q3) {
+  int8_t mask = 0;
+  mask |= (abs(p1 - p0) > thresh) * -1;
+  mask |= (abs(q1 - q0) > thresh) * -1;
+  mask |= (abs(p2 - p0) > thresh) * -1;
+  mask |= (abs(q2 - q0) > thresh) * -1;
+  mask |= (abs(p3 - p0) > thresh) * -1;
+  mask |= (abs(q3 - q0) > thresh) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t flat_mask5(uint8_t thresh,
+                                uint8_t p4, uint8_t p3,
+                                uint8_t p2, uint8_t p1,
+                                uint8_t p0, uint8_t q0,
+                                uint8_t q1, uint8_t q2,
+                                uint8_t q3, uint8_t q4) {
+  int8_t mask = ~flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3);
+  mask |= (abs(p4 - p0) > thresh) * -1;
+  mask |= (abs(q4 - q0) > thresh) * -1;
+  return ~mask;
+}
+
+// is there high edge variance internal edge: 11111111 yes, 00000000 no
+static INLINE int8_t hev_mask(uint8_t thresh, uint8_t p1, uint8_t p0,
+                              uint8_t q0, uint8_t q1) {
+  int8_t hev = 0;
+  hev  |= (abs(p1 - p0) > thresh) * -1;
+  hev  |= (abs(q1 - q0) > thresh) * -1;
+  return hev;
+}
+
+static INLINE void filter4(int8_t mask, uint8_t thresh, uint8_t *op1,
+                           uint8_t *op0, uint8_t *oq0, uint8_t *oq1) {
+  int8_t filter1, filter2;
+
+  const int8_t ps1 = (int8_t) *op1 ^ 0x80;
+  const int8_t ps0 = (int8_t) *op0 ^ 0x80;
+  const int8_t qs0 = (int8_t) *oq0 ^ 0x80;
+  const int8_t qs1 = (int8_t) *oq1 ^ 0x80;
+  const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1);
+
+  // add outer taps if we have high edge variance
+  int8_t filter = signed_char_clamp(ps1 - qs1) & hev;
+
+  // inner taps
+  filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask;
+
+  // save bottom 3 bits so that we round one side +4 and the other +3
+  // if it equals 4 we'll set to adjust by -1 to account for the fact
+  // we'd round 3 the other way
+  filter1 = signed_char_clamp(filter + 4) >> 3;
+  filter2 = signed_char_clamp(filter + 3) >> 3;
+
+  *oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80;
+  *op0 = signed_char_clamp(ps0 + filter2) ^ 0x80;
+
+  // outer tap adjustments
+  filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+
+  *oq1 = signed_char_clamp(qs1 - filter) ^ 0x80;
+  *op1 = signed_char_clamp(ps1 + filter) ^ 0x80;
+}
+
+void vpx_lpf_horizontal_4_c(uint8_t *s, int p /* pitch */,
+                            const uint8_t *blimit, const uint8_t *limit,
+                            const uint8_t *thresh) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+    const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint8_t q0 = s[0 * p],  q1 = s[1 * p],  q2 = s[2 * p],  q3 = s[3 * p];
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p);
+    ++s;
+  }
+}
+
+void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
+                                 const uint8_t *limit0, const uint8_t *thresh0,
+                                 const uint8_t *blimit1, const uint8_t *limit1,
+                                 const uint8_t *thresh1) {
+  vpx_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0);
+  vpx_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit,
+                          const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+    const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint8_t q0 = s[0],  q1 = s[1],  q2 = s[2],  q3 = s[3];
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    filter4(mask, *thresh, s - 2, s - 1, s, s + 1);
+    s += pitch;
+  }
+}
+
+void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
+                               const uint8_t *limit0, const uint8_t *thresh0,
+                               const uint8_t *blimit1, const uint8_t *limit1,
+                               const uint8_t *thresh1) {
+  vpx_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0);
+  vpx_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1, thresh1);
+}
+
+static INLINE void filter8(int8_t mask, uint8_t thresh, uint8_t flat,
+                           uint8_t *op3, uint8_t *op2,
+                           uint8_t *op1, uint8_t *op0,
+                           uint8_t *oq0, uint8_t *oq1,
+                           uint8_t *oq2, uint8_t *oq3) {
+  if (flat && mask) {
+    const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+    const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+
+    // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
+    *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
+    *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
+    *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
+    *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
+    *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
+    *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
+  } else {
+    filter4(mask, thresh, op1,  op0, oq0, oq1);
+  }
+}
+
+void vpx_lpf_horizontal_8_c(uint8_t *s, int p, const uint8_t *blimit,
+                            const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+    const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+    filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+                                 s,         s + 1 * p, s + 2 * p, s + 3 * p);
+    ++s;
+  }
+}
+
+void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
+                                 const uint8_t *limit0, const uint8_t *thresh0,
+                                 const uint8_t *blimit1, const uint8_t *limit1,
+                                 const uint8_t *thresh1) {
+  vpx_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0);
+  vpx_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit,
+                          const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+
+  for (i = 0; i < 8; ++i) {
+    const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+    filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1,
+                                 s,     s + 1, s + 2, s + 3);
+    s += pitch;
+  }
+}
+
+void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
+                               const uint8_t *limit0, const uint8_t *thresh0,
+                               const uint8_t *blimit1, const uint8_t *limit1,
+                               const uint8_t *thresh1) {
+  vpx_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0);
+  vpx_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1, thresh1);
+}
+
+static INLINE void filter16(int8_t mask, uint8_t thresh,
+                            uint8_t flat, uint8_t flat2,
+                            uint8_t *op7, uint8_t *op6,
+                            uint8_t *op5, uint8_t *op4,
+                            uint8_t *op3, uint8_t *op2,
+                            uint8_t *op1, uint8_t *op0,
+                            uint8_t *oq0, uint8_t *oq1,
+                            uint8_t *oq2, uint8_t *oq3,
+                            uint8_t *oq4, uint8_t *oq5,
+                            uint8_t *oq6, uint8_t *oq7) {
+  if (flat2 && flat && mask) {
+    const uint8_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4,
+                  p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+
+    const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3,
+                  q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7;
+
+    // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
+    *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0, 4);
+    *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1, 4);
+    *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2, 4);
+    *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3, 4);
+    *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4, 4);
+    *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5, 4);
+    *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
+                              q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
+    *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
+    *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
+    *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
+    *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
+    *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
+    *oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
+    *oq6 = ROUND_POWER_OF_TWO(p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
+  } else {
+    filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
+  }
+}
+
+static void mb_lpf_horizontal_edge_w(uint8_t *s, int p, const uint8_t *blimit,
+                                     const uint8_t *limit,
+                                     const uint8_t *thresh, int count) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat2 = flat_mask5(1,
+                             s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
+                             q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p]);
+
+    filter16(mask, *thresh, flat, flat2,
+             s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
+             s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+             s,         s + 1 * p, s + 2 * p, s + 3 * p,
+             s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p);
+    ++s;
+  }
+}
+
+void vpx_lpf_horizontal_edge_8_c(uint8_t *s, int p, const uint8_t *blimit,
+                                 const uint8_t *limit, const uint8_t *thresh) {
+  mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 1);
+}
+
+void vpx_lpf_horizontal_edge_16_c(uint8_t *s, int p, const uint8_t *blimit,
+                                  const uint8_t *limit, const uint8_t *thresh) {
+  mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 2);
+}
+
+static void mb_lpf_vertical_edge_w(uint8_t *s, int p,
+                                   const uint8_t *blimit,
+                                   const uint8_t *limit,
+                                   const uint8_t *thresh,
+                                   int count) {
+  int i;
+
+  for (i = 0; i < count; ++i) {
+    const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint8_t q0 = s[0], q1 = s[1],  q2 = s[2], q3 = s[3];
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat2 = flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
+                                    q0, s[4], s[5], s[6], s[7]);
+
+    filter16(mask, *thresh, flat, flat2,
+             s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
+             s,     s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7);
+    s += p;
+  }
+}
+
+void vpx_lpf_vertical_16_c(uint8_t *s, int p, const uint8_t *blimit,
+                           const uint8_t *limit, const uint8_t *thresh) {
+  mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8);
+}
+
+void vpx_lpf_vertical_16_dual_c(uint8_t *s, int p, const uint8_t *blimit,
+                                const uint8_t *limit, const uint8_t *thresh) {
+  mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+// Should we apply any filter at all: 11111111 yes, 00000000 no ?
+static INLINE int8_t highbd_filter_mask(uint8_t limit, uint8_t blimit,
+                                        uint16_t p3, uint16_t p2,
+                                        uint16_t p1, uint16_t p0,
+                                        uint16_t q0, uint16_t q1,
+                                        uint16_t q2, uint16_t q3, int bd) {
+  int8_t mask = 0;
+  int16_t limit16 = (uint16_t)limit << (bd - 8);
+  int16_t blimit16 = (uint16_t)blimit << (bd - 8);
+  mask |= (abs(p3 - p2) > limit16) * -1;
+  mask |= (abs(p2 - p1) > limit16) * -1;
+  mask |= (abs(p1 - p0) > limit16) * -1;
+  mask |= (abs(q1 - q0) > limit16) * -1;
+  mask |= (abs(q2 - q1) > limit16) * -1;
+  mask |= (abs(q3 - q2) > limit16) * -1;
+  mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit16) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t highbd_flat_mask4(uint8_t thresh,
+                                       uint16_t p3, uint16_t p2,
+                                       uint16_t p1, uint16_t p0,
+                                       uint16_t q0, uint16_t q1,
+                                       uint16_t q2, uint16_t q3, int bd) {
+  int8_t mask = 0;
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  mask |= (abs(p1 - p0) > thresh16) * -1;
+  mask |= (abs(q1 - q0) > thresh16) * -1;
+  mask |= (abs(p2 - p0) > thresh16) * -1;
+  mask |= (abs(q2 - q0) > thresh16) * -1;
+  mask |= (abs(p3 - p0) > thresh16) * -1;
+  mask |= (abs(q3 - q0) > thresh16) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t highbd_flat_mask5(uint8_t thresh,
+                                       uint16_t p4, uint16_t p3,
+                                       uint16_t p2, uint16_t p1,
+                                       uint16_t p0, uint16_t q0,
+                                       uint16_t q1, uint16_t q2,
+                                       uint16_t q3, uint16_t q4, int bd) {
+  int8_t mask = ~highbd_flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  mask |= (abs(p4 - p0) > thresh16) * -1;
+  mask |= (abs(q4 - q0) > thresh16) * -1;
+  return ~mask;
+}
+
+// Is there high edge variance internal edge:
+// 11111111_11111111 yes, 00000000_00000000 no ?
+static INLINE int16_t highbd_hev_mask(uint8_t thresh, uint16_t p1, uint16_t p0,
+                                      uint16_t q0, uint16_t q1, int bd) {
+  int16_t hev = 0;
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  hev |= (abs(p1 - p0) > thresh16) * -1;
+  hev |= (abs(q1 - q0) > thresh16) * -1;
+  return hev;
+}
+
+static INLINE void highbd_filter4(int8_t mask, uint8_t thresh, uint16_t *op1,
+                                  uint16_t *op0, uint16_t *oq0, uint16_t *oq1,
+                                  int bd) {
+  int16_t filter1, filter2;
+  // ^0x80 equivalent to subtracting 0x80 from the values to turn them
+  // into -128 to +127 instead of 0 to 255.
+  int shift = bd - 8;
+  const int16_t ps1 = (int16_t)*op1 - (0x80 << shift);
+  const int16_t ps0 = (int16_t)*op0 - (0x80 << shift);
+  const int16_t qs0 = (int16_t)*oq0 - (0x80 << shift);
+  const int16_t qs1 = (int16_t)*oq1 - (0x80 << shift);
+  const uint16_t hev = highbd_hev_mask(thresh, *op1, *op0, *oq0, *oq1, bd);
+
+  // Add outer taps if we have high edge variance.
+  int16_t filter = signed_char_clamp_high(ps1 - qs1, bd) & hev;
+
+  // Inner taps.
+  filter = signed_char_clamp_high(filter + 3 * (qs0 - ps0), bd) & mask;
+
+  // Save bottom 3 bits so that we round one side +4 and the other +3
+  // if it equals 4 we'll set to adjust by -1 to account for the fact
+  // we'd round 3 the other way.
+  filter1 = signed_char_clamp_high(filter + 4, bd) >> 3;
+  filter2 = signed_char_clamp_high(filter + 3, bd) >> 3;
+
+  *oq0 = signed_char_clamp_high(qs0 - filter1, bd) + (0x80 << shift);
+  *op0 = signed_char_clamp_high(ps0 + filter2, bd) + (0x80 << shift);
+
+  // Outer tap adjustments.
+  filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+
+  *oq1 = signed_char_clamp_high(qs1 - filter, bd) + (0x80 << shift);
+  *op1 = signed_char_clamp_high(ps1 + filter, bd) + (0x80 << shift);
+}
+
+void vpx_highbd_lpf_horizontal_4_c(uint16_t *s, int p /* pitch */,
+                                   const uint8_t *blimit, const uint8_t *limit,
+                                   const uint8_t *thresh, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+    const uint16_t p3 = s[-4 * p];
+    const uint16_t p2 = s[-3 * p];
+    const uint16_t p1 = s[-2 * p];
+    const uint16_t p0 = s[-p];
+    const uint16_t q0 = s[0 * p];
+    const uint16_t q1 = s[1 * p];
+    const uint16_t q2 = s[2 * p];
+    const uint16_t q3 = s[3 * p];
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                           p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    highbd_filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p, bd);
+    ++s;
+  }
+}
+
+void vpx_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int p,
+                                        const uint8_t *blimit0,
+                                        const uint8_t *limit0,
+                                        const uint8_t *thresh0,
+                                        const uint8_t *blimit1,
+                                        const uint8_t *limit1,
+                                        const uint8_t *thresh1,
+                                        int bd) {
+  vpx_highbd_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, bd);
+  vpx_highbd_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, bd);
+}
+
+void vpx_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit,
+                                 const uint8_t *limit, const uint8_t *thresh,
+                                 int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+    const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint16_t q0 = s[0],  q1 = s[1],  q2 = s[2],  q3 = s[3];
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                           p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    highbd_filter4(mask, *thresh, s - 2, s - 1, s, s + 1, bd);
+    s += pitch;
+  }
+}
+
+void vpx_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch,
+                                      const uint8_t *blimit0,
+                                      const uint8_t *limit0,
+                                      const uint8_t *thresh0,
+                                      const uint8_t *blimit1,
+                                      const uint8_t *limit1,
+                                      const uint8_t *thresh1,
+                                      int bd) {
+  vpx_highbd_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, bd);
+  vpx_highbd_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1,
+                              thresh1, bd);
+}
+
+static INLINE void highbd_filter8(int8_t mask, uint8_t thresh, uint8_t flat,
+                                  uint16_t *op3, uint16_t *op2,
+                                  uint16_t *op1, uint16_t *op0,
+                                  uint16_t *oq0, uint16_t *oq1,
+                                  uint16_t *oq2, uint16_t *oq3, int bd) {
+  if (flat && mask) {
+    const uint16_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+    const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+
+    // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
+    *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
+    *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
+    *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
+    *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
+    *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
+    *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
+  } else {
+    highbd_filter4(mask, thresh, op1,  op0, oq0, oq1, bd);
+  }
+}
+
+void vpx_highbd_lpf_horizontal_8_c(uint16_t *s, int p, const uint8_t *blimit,
+                                   const uint8_t *limit, const uint8_t *thresh,
+                                   int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+    const uint16_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                         p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+                                          bd);
+    highbd_filter8(mask, *thresh, flat,
+                 s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+                 s, s + 1 * p, s + 2 * p, s + 3 * p, bd);
+    ++s;
+  }
+}
+
+void vpx_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int p,
+                                        const uint8_t *blimit0,
+                                        const uint8_t *limit0,
+                                        const uint8_t *thresh0,
+                                        const uint8_t *blimit1,
+                                        const uint8_t *limit1,
+                                        const uint8_t *thresh1,
+                                        int bd) {
+  vpx_highbd_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, bd);
+  vpx_highbd_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, bd);
+}
+
+void vpx_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit,
+                                 const uint8_t *limit, const uint8_t *thresh,
+                                 int bd) {
+  int i;
+
+  for (i = 0; i < 8; ++i) {
+    const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                           p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+                                          bd);
+    highbd_filter8(mask, *thresh, flat,
+                 s - 4, s - 3, s - 2, s - 1,
+                 s, s + 1, s + 2, s + 3,
+                 bd);
+    s += pitch;
+  }
+}
+
+void vpx_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch,
+                                      const uint8_t *blimit0,
+                                      const uint8_t *limit0,
+                                      const uint8_t *thresh0,
+                                      const uint8_t *blimit1,
+                                      const uint8_t *limit1,
+                                      const uint8_t *thresh1,
+                                      int bd) {
+  vpx_highbd_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, bd);
+  vpx_highbd_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1,
+                              thresh1, bd);
+}
+
+static INLINE void highbd_filter16(int8_t mask, uint8_t thresh,
+                                   uint8_t flat, uint8_t flat2,
+                                   uint16_t *op7, uint16_t *op6,
+                                   uint16_t *op5, uint16_t *op4,
+                                   uint16_t *op3, uint16_t *op2,
+                                   uint16_t *op1, uint16_t *op0,
+                                   uint16_t *oq0, uint16_t *oq1,
+                                   uint16_t *oq2, uint16_t *oq3,
+                                   uint16_t *oq4, uint16_t *oq5,
+                                   uint16_t *oq6, uint16_t *oq7, int bd) {
+  if (flat2 && flat && mask) {
+    const uint16_t p7 = *op7;
+    const uint16_t p6 = *op6;
+    const uint16_t p5 = *op5;
+    const uint16_t p4 = *op4;
+    const uint16_t p3 = *op3;
+    const uint16_t p2 = *op2;
+    const uint16_t p1 = *op1;
+    const uint16_t p0 = *op0;
+    const uint16_t q0 = *oq0;
+    const uint16_t q1 = *oq1;
+    const uint16_t q2 = *oq2;
+    const uint16_t q3 = *oq3;
+    const uint16_t q4 = *oq4;
+    const uint16_t q5 = *oq5;
+    const uint16_t q6 = *oq6;
+    const uint16_t q7 = *oq7;
+
+    // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
+    *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0, 4);
+    *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1, 4);
+    *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2, 4);
+    *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3, 4);
+    *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4, 4);
+    *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5, 4);
+    *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
+                              q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
+    *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
+    *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
+    *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
+    *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
+    *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
+    *oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
+    *oq6 = ROUND_POWER_OF_TWO(p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
+  } else {
+    highbd_filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3,
+                   bd);
+  }
+}
+
+static void highbd_mb_lpf_horizontal_edge_w(uint16_t *s, int p,
+                                            const uint8_t *blimit,
+                                            const uint8_t *limit,
+                                            const uint8_t *thresh,
+                                            int count, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4 * p];
+    const uint16_t p2 = s[-3 * p];
+    const uint16_t p1 = s[-2 * p];
+    const uint16_t p0 = s[-p];
+    const uint16_t q0 = s[0 * p];
+    const uint16_t q1 = s[1 * p];
+    const uint16_t q2 = s[2 * p];
+    const uint16_t q3 = s[3 * p];
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                           p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+                                          bd);
+    const int8_t flat2 = highbd_flat_mask5(
+        1, s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
+        q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p], bd);
+
+    highbd_filter16(mask, *thresh, flat, flat2,
+                    s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
+                    s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+                    s, s + 1 * p, s + 2 * p, s + 3 * p,
+                    s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p,
+                    bd);
+    ++s;
+  }
+}
+
+void vpx_highbd_lpf_horizontal_edge_8_c(uint16_t *s, int p,
+                                        const uint8_t *blimit,
+                                        const uint8_t *limit,
+                                        const uint8_t *thresh, int bd) {
+  highbd_mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 1, bd);
+}
+
+void vpx_highbd_lpf_horizontal_edge_16_c(uint16_t *s, int p,
+                                         const uint8_t *blimit,
+                                         const uint8_t *limit,
+                                         const uint8_t *thresh, int bd) {
+  highbd_mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 2, bd);
+}
+
+static void highbd_mb_lpf_vertical_edge_w(uint16_t *s, int p,
+                                          const uint8_t *blimit,
+                                          const uint8_t *limit,
+                                          const uint8_t *thresh,
+                                          int count, int bd) {
+  int i;
+
+  for (i = 0; i < count; ++i) {
+    const uint16_t p3 = s[-4];
+    const uint16_t p2 = s[-3];
+    const uint16_t p1 = s[-2];
+    const uint16_t p0 = s[-1];
+    const uint16_t q0 = s[0];
+    const uint16_t q1 = s[1];
+    const uint16_t q2 = s[2];
+    const uint16_t q3 = s[3];
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                           p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+                                          bd);
+    const int8_t flat2 = highbd_flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
+                                           q0, s[4], s[5], s[6], s[7], bd);
+
+    highbd_filter16(mask, *thresh, flat, flat2,
+                    s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
+                    s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7,
+                    bd);
+    s += p;
+  }
+}
+
+void vpx_highbd_lpf_vertical_16_c(uint16_t *s, int p, const uint8_t *blimit,
+                                  const uint8_t *limit, const uint8_t *thresh,
+                                  int bd) {
+  highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8, bd);
+}
+
+void vpx_highbd_lpf_vertical_16_dual_c(uint16_t *s, int p,
+                                       const uint8_t *blimit,
+                                       const uint8_t *limit,
+                                       const uint8_t *thresh,
+                                       int bd) {
+  highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16, bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vpx_dsp/mips/add_noise_msa.c b/libvpx/libvpx/vpx_dsp/mips/add_noise_msa.c
new file mode 100644
index 0000000..366770c
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/add_noise_msa.c
@@ -0,0 +1,59 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include "./macros_msa.h"
+
+void vpx_plane_add_noise_msa(uint8_t *start_ptr, char *noise,
+                             char blackclamp[16], char whiteclamp[16],
+                             char bothclamp[16], uint32_t width,
+                             uint32_t height, int32_t pitch) {
+  uint32_t i, j;
+
+  for (i = 0; i < height / 2; ++i) {
+    uint8_t *pos0_ptr = start_ptr + (2 * i) * pitch;
+    int8_t *ref0_ptr = (int8_t *)(noise + (rand() & 0xff));
+    uint8_t *pos1_ptr = start_ptr + (2 * i + 1) * pitch;
+    int8_t *ref1_ptr = (int8_t *)(noise + (rand() & 0xff));
+    for (j = width / 16; j--;) {
+      v16i8 temp00_s, temp01_s;
+      v16u8 temp00, temp01, black_clamp, white_clamp;
+      v16u8 pos0, ref0, pos1, ref1;
+      v16i8 const127 = __msa_ldi_b(127);
+
+      pos0 = LD_UB(pos0_ptr);
+      ref0 = LD_UB(ref0_ptr);
+      pos1 = LD_UB(pos1_ptr);
+      ref1 = LD_UB(ref1_ptr);
+      black_clamp = (v16u8)__msa_fill_b(blackclamp[0]);
+      white_clamp = (v16u8)__msa_fill_b(whiteclamp[0]);
+      temp00 = (pos0 < black_clamp);
+      pos0 = __msa_bmnz_v(pos0, black_clamp, temp00);
+      temp01 = (pos1 < black_clamp);
+      pos1 = __msa_bmnz_v(pos1, black_clamp, temp01);
+      XORI_B2_128_UB(pos0, pos1);
+      temp00_s = __msa_adds_s_b((v16i8)white_clamp, const127);
+      temp00 = (v16u8)(temp00_s < pos0);
+      pos0 = (v16u8)__msa_bmnz_v((v16u8)pos0, (v16u8)temp00_s, temp00);
+      temp01_s = __msa_adds_s_b((v16i8)white_clamp, const127);
+      temp01 = (temp01_s < pos1);
+      pos1 = (v16u8)__msa_bmnz_v((v16u8)pos1, (v16u8)temp01_s, temp01);
+      XORI_B2_128_UB(pos0, pos1);
+      pos0 += ref0;
+      ST_UB(pos0, pos0_ptr);
+      pos1 += ref1;
+      ST_UB(pos1, pos1_ptr);
+      pos0_ptr += 16;
+      pos1_ptr += 16;
+      ref0_ptr += 16;
+      ref1_ptr += 16;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/avg_msa.c b/libvpx/libvpx/vpx_dsp/mips/avg_msa.c
new file mode 100644
index 0000000..52a24ed
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/avg_msa.c
@@ -0,0 +1,56 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/macros_msa.h"
+
+uint32_t vpx_avg_8x8_msa(const uint8_t *src, int32_t src_stride) {
+  uint32_t sum_out;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v8u16 sum0, sum1, sum2, sum3, sum4, sum5, sum6, sum7;
+  v4u32 sum = { 0 };
+
+  LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  HADD_UB4_UH(src0, src1, src2, src3, sum0, sum1, sum2, sum3);
+  HADD_UB4_UH(src4, src5, src6, src7, sum4, sum5, sum6, sum7);
+  ADD4(sum0, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum0, sum2, sum4, sum6);
+  ADD2(sum0, sum2, sum4, sum6, sum0, sum4);
+  sum0 += sum4;
+
+  sum = __msa_hadd_u_w(sum0, sum0);
+  sum0 = (v8u16)__msa_pckev_h((v8i16)sum, (v8i16)sum);
+  sum = __msa_hadd_u_w(sum0, sum0);
+  sum = (v4u32)__msa_srari_w((v4i32)sum, 6);
+  sum_out = __msa_copy_u_w((v4i32)sum, 0);
+
+  return sum_out;
+}
+
+uint32_t vpx_avg_4x4_msa(const uint8_t *src, int32_t src_stride) {
+  uint32_t sum_out;
+  uint32_t src0, src1, src2, src3;
+  v16u8 vec = { 0 };
+  v8u16 sum0;
+  v4u32 sum1;
+  v2u64 sum2;
+
+  LW4(src, src_stride, src0, src1, src2, src3);
+  INSERT_W4_UB(src0, src1, src2, src3, vec);
+
+  sum0 = __msa_hadd_u_h(vec, vec);
+  sum1 = __msa_hadd_u_w(sum0, sum0);
+  sum0 = (v8u16)__msa_pckev_h((v8i16)sum1, (v8i16)sum1);
+  sum1 = __msa_hadd_u_w(sum0, sum0);
+  sum2 = __msa_hadd_u_d(sum1, sum1);
+  sum1 = (v4u32)__msa_srari_w((v4i32)sum2, 4);
+  sum_out = __msa_copy_u_w((v4i32)sum1, 0);
+
+  return sum_out;
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/common_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/common_dspr2.c
new file mode 100644
index 0000000..b22f084
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/common_dspr2.c
@@ -0,0 +1,30 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/common_dspr2.h"
+
+#if HAVE_DSPR2
+uint8_t vpx_ff_cropTbl_a[256 + 2 * CROP_WIDTH];
+uint8_t *vpx_ff_cropTbl;
+
+void vpx_dsputil_static_init(void) {
+  int i;
+
+  for (i = 0; i < 256; i++) vpx_ff_cropTbl_a[i + CROP_WIDTH] = i;
+
+  for (i = 0; i < CROP_WIDTH; i++) {
+    vpx_ff_cropTbl_a[i] = 0;
+    vpx_ff_cropTbl_a[i + CROP_WIDTH + 256] = 255;
+  }
+
+  vpx_ff_cropTbl = &vpx_ff_cropTbl_a[CROP_WIDTH];
+}
+
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/common_dspr2.h b/libvpx/libvpx/vpx_dsp/mips/common_dspr2.h
new file mode 100644
index 0000000..7a10bf1
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/common_dspr2.h
@@ -0,0 +1,64 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_COMMON_MIPS_DSPR2_H_
+#define VPX_COMMON_MIPS_DSPR2_H_
+
+#include <assert.h>
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+#if HAVE_DSPR2
+#define CROP_WIDTH 512
+
+extern uint8_t *vpx_ff_cropTbl;  // From "vpx_dsp/mips/intrapred4_dspr2.c"
+
+static INLINE void prefetch_load(const unsigned char *src) {
+  __asm__ __volatile__ (
+      "pref   0,  0(%[src])   \n\t"
+      :
+      : [src] "r" (src)
+  );
+}
+
+/* prefetch data for store */
+static INLINE void prefetch_store(unsigned char *dst) {
+  __asm__ __volatile__ (
+      "pref   1,  0(%[dst])   \n\t"
+      :
+      : [dst] "r" (dst)
+  );
+}
+
+static INLINE void prefetch_load_streamed(const unsigned char *src) {
+  __asm__ __volatile__ (
+      "pref   4,  0(%[src])   \n\t"
+      :
+      : [src] "r" (src)
+  );
+}
+
+/* prefetch data for store */
+static INLINE void prefetch_store_streamed(unsigned char *dst) {
+  __asm__ __volatile__ (
+      "pref   5,  0(%[dst])   \n\t"
+      :
+      : [dst] "r" (dst)
+  );
+}
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_COMMON_MIPS_DSPR2_H_
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve2_avg_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/convolve2_avg_dspr2.c
new file mode 100644
index 0000000..3c76767
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve2_avg_dspr2.c
@@ -0,0 +1,273 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/convolve_common_dspr2.h"
+#include "vpx_dsp/vpx_convolve.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_bi_avg_vert_4_dspr2(const uint8_t *src,
+                                         int32_t src_stride,
+                                         uint8_t *dst,
+                                         int32_t dst_stride,
+                                         const int16_t *filter_y,
+                                         int32_t w,
+                                         int32_t h) {
+  int32_t       x, y;
+  const uint8_t *src_ptr;
+  uint8_t       *dst_ptr;
+  uint8_t       *cm = vpx_ff_cropTbl;
+  uint32_t      vector4a = 64;
+  uint32_t      load1, load2;
+  uint32_t      p1, p2;
+  uint32_t      scratch1, scratch2;
+  uint32_t      store1, store2;
+  int32_t       Temp1, Temp2;
+  const int16_t *filter = &filter_y[3];
+  uint32_t      filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+
+    for (x = 0; x < w; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__ (
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]     \n\t"
+
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "lbu              %[scratch1],  0(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 1 */
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 2 */
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+          "lbu              %[scratch1],  2(%[dst_ptr])                   \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  3(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 3 */
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 4 */
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r" (load1), [load2] "=&r" (load2),
+            [p1] "=&r" (p1), [p2] "=&r" (p2),
+            [scratch1] "=&r" (scratch1), [scratch2] "=&r" (scratch2),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+            [store1] "=&r" (store1), [store2] "=&r" (store2),
+            [src_ptr] "+r" (src_ptr)
+          : [filter45] "r" (filter45), [vector4a] "r" (vector4a),
+            [src_stride] "r" (src_stride), [cm] "r" (cm),
+            [dst_ptr] "r" (dst_ptr)
+      );
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_avg_vert_64_dspr2(const uint8_t *src,
+                                          int32_t src_stride,
+                                          uint8_t *dst,
+                                          int32_t dst_stride,
+                                          const int16_t *filter_y,
+                                          int32_t h) {
+  int32_t       x, y;
+  const uint8_t *src_ptr;
+  uint8_t       *dst_ptr;
+  uint8_t       *cm = vpx_ff_cropTbl;
+  uint32_t      vector4a = 64;
+  uint32_t      load1, load2;
+  uint32_t      p1, p2;
+  uint32_t      scratch1, scratch2;
+  uint32_t      store1, store2;
+  int32_t       Temp1, Temp2;
+  const int16_t *filter = &filter_y[3];
+  uint32_t filter45;;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+    prefetch_store(dst + dst_stride + 32);
+
+    for (x = 0; x < 64; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__ (
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]     \n\t"
+
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "lbu              %[scratch1],  0(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 1 */
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 2 */
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+          "lbu              %[scratch1],  2(%[dst_ptr])                   \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  3(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 3 */
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 4 */
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r" (load1), [load2] "=&r" (load2),
+            [p1] "=&r" (p1), [p2] "=&r" (p2),
+            [scratch1] "=&r" (scratch1), [scratch2] "=&r" (scratch2),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+            [store1] "=&r" (store1), [store2] "=&r" (store2),
+            [src_ptr] "+r" (src_ptr)
+          : [filter45] "r" (filter45), [vector4a] "r" (vector4a),
+            [src_stride] "r" (src_stride), [cm] "r" (cm),
+            [dst_ptr] "r" (dst_ptr)
+      );
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vpx_convolve2_avg_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4,
+                                  int w, int h) {
+  uint32_t pos = 38;
+
+  assert(y_step_q4 == 16);
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4:
+    case 8:
+    case 16:
+    case 32:
+      convolve_bi_avg_vert_4_dspr2(src, src_stride,
+                                   dst, dst_stride,
+                                   filter_y, w, h);
+      break;
+    case 64:
+      prefetch_store(dst + 32);
+      convolve_bi_avg_vert_64_dspr2(src, src_stride,
+                                    dst, dst_stride,
+                                    filter_y, h);
+      break;
+    default:
+      vpx_convolve8_avg_vert_c(src, src_stride,
+                               dst, dst_stride,
+                               filter_x, x_step_q4,
+                               filter_y, y_step_q4,
+                               w, h);
+      break;
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve2_avg_horiz_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/convolve2_avg_horiz_dspr2.c
new file mode 100644
index 0000000..932a73d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve2_avg_horiz_dspr2.c
@@ -0,0 +1,825 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/convolve_common_dspr2.h"
+#include "vpx_dsp/vpx_convolve.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_bi_avg_horiz_4_dspr2(const uint8_t *src,
+                                          int32_t src_stride,
+                                          uint8_t *dst,
+                                          int32_t dst_stride,
+                                          const int16_t *filter_x0,
+                                          int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  int32_t  Temp1, Temp2, Temp3, Temp4;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3;
+  uint32_t tn1, tn2;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t      filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        "lbu              %[p2],          3(%[dst])                      \n\t"  /* load odd 2 */
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],         %[Temp1](%[cm])                \n\t"  /* even 1 */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "lbu              %[Temp1],       1(%[dst])                      \n\t"  /* load odd 1 */
+        "preceu.ph.qbr    %[p1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p3],          %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tn2],         0(%[dst])                      \n\t"  /* load even 1 */
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],         %[Temp3](%[cm])                \n\t"  /* even 2 */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "lbux             %[tn1],         %[Temp2](%[cm])                \n\t"  /* odd 1 */
+        "addqh_r.w        %[tn2],         %[tn2],         %[tp1]         \n\t"  /* average even 1 */
+        "dpa.w.ph         $ac2,           %[p3],          %[filter45]    \n\t"
+        "extp             %[Temp4],       $ac2,           31             \n\t"
+
+        "lbu              %[tp1],         2(%[dst])                      \n\t"  /* load even 2 */
+        "sb               %[tn2],         0(%[dst])                      \n\t"  /* store even 1 */
+
+        /* clamp */
+        "addqh_r.w        %[Temp1],       %[Temp1],       %[tn1]         \n\t"  /* average odd 1 */
+        "lbux             %[p3],          %[Temp4](%[cm])                \n\t"  /* odd 2 */
+        "sb               %[Temp1],       1(%[dst])                      \n\t"  /* store odd 1 */
+
+        "addqh_r.w        %[tp1],         %[tp1],         %[tp2]         \n\t"  /* average even 2 */
+        "sb               %[tp1],         2(%[dst])                      \n\t"  /* store even 2 */
+
+        "addqh_r.w        %[p2],          %[p2],          %[p3]          \n\t"  /* average odd 2 */
+        "sb               %[p2],          3(%[dst])                      \n\t"  /* store odd 2 */
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+          [tn1] "=&r" (tn1), [tn2] "=&r" (tn2),
+          [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+          [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4)
+        : [filter45] "r" (filter45), [vector4a] "r" (vector4a),
+          [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_avg_horiz_8_dspr2(const uint8_t *src,
+                                         int32_t src_stride,
+                                         uint8_t *dst,
+                                         int32_t dst_stride,
+                                         const int16_t *filter_x0,
+                                         int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector4a = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2, tp3, tp4;
+  uint32_t p1, p2, p3, p4, n1;
+  uint32_t st0, st1;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp2]                         \n\t"
+        "ulw              %[tp3],         8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+        "lbu              %[Temp2],       0(%[dst])                      \n\t"
+        "lbu              %[tp4],         2(%[dst])                      \n\t"
+
+        /* even 2. pixel */
+        "dpa.w.ph         $ac2,           %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[st0],         %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac1                           \n\t"
+        "mthi             $zero,          $ac1                           \n\t"
+        "lbux             %[st1],         %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac1,           31             \n\t"
+
+        "addqh_r.w        %[Temp2],       %[Temp2],       %[st0]         \n\t"
+        "addqh_r.w        %[tp4],         %[tp4],         %[st1]         \n\t"
+        "sb               %[Temp2],       0(%[dst])                      \n\t"
+        "sb               %[tp4],         2(%[dst])                      \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+
+        "balign           %[tp3],         %[tp2],         3              \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+
+        "lbux             %[st0],         %[Temp1](%[cm])                \n\t"
+        "lbu              %[Temp2],       4(%[dst])                      \n\t"
+        "addqh_r.w        %[Temp2],       %[Temp2],       %[st0]         \n\t"
+
+        "dpa.w.ph         $ac2,           %[p4],          %[filter45]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a],    $ac1                           \n\t"
+        "mthi             $zero,          $ac1                           \n\t"
+        "sb               %[Temp2],       4(%[dst])                      \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp3]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp3]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tp1],         6(%[dst])                      \n\t"
+
+        /* odd 2. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "lbux             %[st0],         %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,           %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],       $ac1,           31             \n\t"
+
+        "lbu              %[tp2],         1(%[dst])                      \n\t"
+        "lbu              %[tp3],         3(%[dst])                      \n\t"
+        "addqh_r.w        %[tp1],         %[tp1],         %[st0]         \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[st1],         %[Temp2](%[cm])                \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[filter45]    \n\t"
+        "addqh_r.w        %[tp2],         %[tp2],         %[st1]         \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tp4],         5(%[dst])                      \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[tp2],         1(%[dst])                      \n\t"
+        "sb               %[tp1],         6(%[dst])                      \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac2,           31             \n\t"
+
+        "lbu              %[tp1],         7(%[dst])                      \n\t"
+
+        /* clamp */
+        "lbux             %[p4],          %[Temp3](%[cm])                \n\t"
+        "addqh_r.w        %[tp3],         %[tp3],         %[p4]          \n\t"
+
+        "lbux             %[p2],          %[Temp2](%[cm])                \n\t"
+        "addqh_r.w        %[tp4],         %[tp4],         %[p2]          \n\t"
+
+        "lbux             %[p1],          %[Temp1](%[cm])                \n\t"
+        "addqh_r.w        %[tp1],         %[tp1],         %[p1]          \n\t"
+
+        /* store bytes */
+        "sb               %[tp3],         3(%[dst])                      \n\t"
+        "sb               %[tp4],         5(%[dst])                      \n\t"
+        "sb               %[tp1],         7(%[dst])                      \n\t"
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+          [tp3] "=&r" (tp3), [tp4] "=&r" (tp4),
+          [st0] "=&r" (st0), [st1] "=&r" (st1),
+          [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+          [n1] "=&r" (n1),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+        : [filter45] "r" (filter45), [vector4a] "r" (vector4a),
+          [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_avg_horiz_16_dspr2(const uint8_t *src_ptr,
+                                          int32_t src_stride,
+                                          uint8_t *dst_ptr,
+                                          int32_t dst_stride,
+                                          const int16_t *filter_x0,
+                                          int32_t h,
+                                          int32_t count) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_store(dst_ptr + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+          "lbu              %[st2],       0(%[dst])                    \n\t" /* load even 1 from dst */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          "lbu              %[qload3],    2(%[dst])                    \n\t" /* load even 2 from dst */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st2],       0(%[dst])                    \n\t" /* store even 1 to dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter45]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st2]       \n\t" /* average even 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    2(%[dst])                    \n\t" /* store even 2 to dst */
+          "lbu              %[qload3],    4(%[dst])                    \n\t" /* load even 3 from dst */
+          "lbu              %[qload1],    6(%[dst])                    \n\t" /* load even 4 from dst */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter45]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 3 */
+          "sb               %[qload3],    4(%[dst])                    \n\t" /* store even 3 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average even 4 */
+          "sb               %[qload1],    6(%[dst])                    \n\t" /* store even 4 to dst */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter45]  \n\t" /* even 6 */
+          "lbu              %[qload2],    8(%[dst])                    \n\t" /* load even 5 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 5 */
+          "sb               %[qload2],    8(%[dst])                    \n\t" /* store even 5 to dst */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* even 7 */
+          "lbu              %[qload3],    10(%[dst])                   \n\t" /* load even 6 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          "lbu              %[st2],       12(%[dst])                   \n\t" /* load even 7 from dst */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* even 8 */
+          "sb               %[qload3],    10(%[dst])                   \n\t" /* store even 6 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                   \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 7 */
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       12(%[dst])                   \n\t" /* store even 7 to dst */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter45]  \n\t" /* odd 1 */
+          "lbu              %[qload2],    14(%[dst])                   \n\t" /* load even 8 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          "lbu              %[st1],       1(%[dst])                    \n\t" /* load odd 1 from dst */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 8 */
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    14(%[dst])                   \n\t" /* store even 8 to dst */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* odd 2 */
+          "lbu              %[qload3],    3(%[dst])                    \n\t" /* load odd 2 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st3],       %[st3],         %[st1]       \n\t" /* average odd 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* odd 3 */
+          "sb               %[st3],       1(%[dst])                    \n\t" /* store odd 1 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st1]       \n\t" /* average odd 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    3(%[dst])                    \n\t" /* store odd 2 to dst */
+          "lbu              %[qload1],    5(%[dst])                    \n\t" /* load odd 3 from dst */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter45]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          "lbu              %[st1],       7(%[dst])                    \n\t" /* load odd 4 from dst */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st2]       \n\t" /* average odd 3 */
+          "sb               %[qload1],    5(%[dst])                    \n\t" /* store odd 3 to dst */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          "lbu              %[qload1],    9(%[dst])                    \n\t" /* load odd 5 from dst */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st1],       %[st1],         %[st3]       \n\t" /* average odd 4 */
+          "sb               %[st1],       7(%[dst])                    \n\t" /* store odd 4 to dst */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter45]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 5 */
+          "sb               %[qload1],    9(%[dst])                    \n\t" /* store odd 5 to dst */
+          "lbu              %[qload2],    11(%[dst])                   \n\t" /* load odd 6 from dst */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          "lbu              %[qload3],    13(%[dst])                   \n\t" /* load odd 7 from dst */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter45]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbu              %[qload1],    15(%[dst])                   \n\t" /* load odd 8 from dst */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average odd 6 */
+
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average odd 7 */
+
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 8 */
+
+          "sb               %[qload2],    11(%[dst])                   \n\t" /* store odd 6 to dst */
+          "sb               %[qload3],    13(%[dst])                   \n\t" /* store odd 7 to dst */
+          "sb               %[qload1],    15(%[dst])                   \n\t" /* store odd 8 to dst */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [qload3] "=&r" (qload3), [p5] "=&r" (p5),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+          : [filter45] "r" (filter45), [vector_64] "r" (vector_64),
+            [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+      );
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void convolve_bi_avg_horiz_64_dspr2(const uint8_t *src_ptr,
+                                          int32_t src_stride,
+                                          uint8_t *dst_ptr,
+                                          int32_t dst_stride,
+                                          const int16_t *filter_x0,
+                                          int32_t h) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+    prefetch_store(dst_ptr + dst_stride);
+    prefetch_store(dst_ptr + dst_stride + 32);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+          "lbu              %[st2],       0(%[dst])                    \n\t" /* load even 1 from dst */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          "lbu              %[qload3],    2(%[dst])                    \n\t" /* load even 2 from dst */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st2],       0(%[dst])                    \n\t" /* store even 1 to dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter45]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st2]       \n\t" /* average even 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    2(%[dst])                    \n\t" /* store even 2 to dst */
+          "lbu              %[qload3],    4(%[dst])                    \n\t" /* load even 3 from dst */
+          "lbu              %[qload1],    6(%[dst])                    \n\t" /* load even 4 from dst */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter45]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 3 */
+          "sb               %[qload3],    4(%[dst])                    \n\t" /* store even 3 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average even 4 */
+          "sb               %[qload1],    6(%[dst])                    \n\t" /* store even 4 to dst */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter45]  \n\t" /* even 6 */
+          "lbu              %[qload2],    8(%[dst])                    \n\t" /* load even 5 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 5 */
+          "sb               %[qload2],    8(%[dst])                    \n\t" /* store even 5 to dst */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* even 7 */
+          "lbu              %[qload3],    10(%[dst])                   \n\t" /* load even 6 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          "lbu              %[st2],       12(%[dst])                   \n\t" /* load even 7 from dst */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* even 8 */
+          "sb               %[qload3],    10(%[dst])                   \n\t" /* store even 6 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                   \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 7 */
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       12(%[dst])                   \n\t" /* store even 7 to dst */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter45]  \n\t" /* odd 1 */
+          "lbu              %[qload2],    14(%[dst])                   \n\t" /* load even 8 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          "lbu              %[st1],       1(%[dst])                    \n\t" /* load odd 1 from dst */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 8 */
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    14(%[dst])                   \n\t" /* store even 8 to dst */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* odd 2 */
+          "lbu              %[qload3],    3(%[dst])                    \n\t" /* load odd 2 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st3],       %[st3],         %[st1]       \n\t" /* average odd 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* odd 3 */
+          "sb               %[st3],       1(%[dst])                    \n\t" /* store odd 1 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st1]       \n\t" /* average odd 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    3(%[dst])                    \n\t" /* store odd 2 to dst */
+          "lbu              %[qload1],    5(%[dst])                    \n\t" /* load odd 3 from dst */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter45]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          "lbu              %[st1],       7(%[dst])                    \n\t" /* load odd 4 from dst */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st2]       \n\t" /* average odd 3 */
+          "sb               %[qload1],    5(%[dst])                    \n\t" /* store odd 3 to dst */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          "lbu              %[qload1],    9(%[dst])                    \n\t" /* load odd 5 from dst */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st1],       %[st1],         %[st3]       \n\t" /* average odd 4 */
+          "sb               %[st1],       7(%[dst])                    \n\t" /* store odd 4 to dst */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter45]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 5 */
+          "sb               %[qload1],    9(%[dst])                    \n\t" /* store odd 5 to dst */
+          "lbu              %[qload2],    11(%[dst])                   \n\t" /* load odd 6 from dst */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          "lbu              %[qload3],    13(%[dst])                   \n\t" /* load odd 7 from dst */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter45]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbu              %[qload1],    15(%[dst])                   \n\t" /* load odd 8 from dst */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average odd 6 */
+
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average odd 7 */
+
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 8 */
+
+          "sb               %[qload2],    11(%[dst])                   \n\t" /* store odd 6 to dst */
+          "sb               %[qload3],    13(%[dst])                   \n\t" /* store odd 7 to dst */
+          "sb               %[qload1],    15(%[dst])                   \n\t" /* store odd 8 to dst */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [qload3] "=&r" (qload3), [p5] "=&r" (p5),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+          : [filter45] "r" (filter45), [vector_64] "r" (vector_64),
+            [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+      );
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+void vpx_convolve2_avg_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int x_step_q4,
+                                   const int16_t *filter_y, int y_step_q4,
+                                   int w, int h) {
+  uint32_t pos = 38;
+
+  assert(x_step_q4 == 16);
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  /* prefetch data to cache memory */
+  prefetch_load(src);
+  prefetch_load(src + 32);
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4:
+      convolve_bi_avg_horiz_4_dspr2(src, src_stride,
+                                   dst, dst_stride,
+                                   filter_x, h);
+      break;
+    case 8:
+      convolve_bi_avg_horiz_8_dspr2(src, src_stride,
+                                   dst, dst_stride,
+                                   filter_x, h);
+      break;
+    case 16:
+      convolve_bi_avg_horiz_16_dspr2(src, src_stride,
+                                    dst, dst_stride,
+                                    filter_x, h, 1);
+      break;
+    case 32:
+      convolve_bi_avg_horiz_16_dspr2(src, src_stride,
+                                    dst, dst_stride,
+                                    filter_x, h, 2);
+      break;
+    case 64:
+      prefetch_load(src + 64);
+      prefetch_store(dst + 32);
+
+      convolve_bi_avg_horiz_64_dspr2(src, src_stride,
+                                    dst, dst_stride,
+                                    filter_x, h);
+      break;
+    default:
+      vpx_convolve8_avg_horiz_c(src, src_stride,
+                                dst, dst_stride,
+                                filter_x, x_step_q4,
+                                filter_y, y_step_q4,
+                                w, h);
+      break;
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve2_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/convolve2_dspr2.c
new file mode 100644
index 0000000..d111029
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve2_dspr2.c
@@ -0,0 +1,782 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/convolve_common_dspr2.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/vpx_filter.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_bi_horiz_4_transposed_dspr2(const uint8_t *src,
+                                                 int32_t src_stride,
+                                                 uint8_t *dst,
+                                                 int32_t dst_stride,
+                                                 const int16_t *filter_x0,
+                                                 int32_t h) {
+  int32_t       y;
+  uint8_t       *cm = vpx_ff_cropTbl;
+  uint8_t       *dst_ptr;
+  int32_t       Temp1, Temp2;
+  uint32_t      vector4a = 64;
+  uint32_t      tp1, tp2;
+  uint32_t      p1, p2;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t      filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    dst_ptr = dst;
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp2],       $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],         %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],         %[Temp2](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp2],       $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[p1],          %[Temp1](%[cm])                \n\t"
+        "lbux             %[p2],          %[Temp2](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[tp1],         0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[p1],          0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[tp2],         0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[p2],          0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+          [p1] "=&r" (p1), [p2] "=&r" (p2),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+          [dst_ptr] "+r" (dst_ptr)
+        : [filter45] "r" (filter45),[vector4a] "r" (vector4a),
+          [cm] "r" (cm), [src] "r" (src), [dst_stride] "r" (dst_stride)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+static void convolve_bi_horiz_8_transposed_dspr2(const uint8_t *src,
+                                                 int32_t src_stride,
+                                                 uint8_t *dst,
+                                                 int32_t dst_stride,
+                                                 const int16_t *filter_x0,
+                                                 int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint8_t *dst_ptr;
+  uint32_t vector4a = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2, tp3;
+  uint32_t p1, p2, p3, p4;
+  uint8_t *odd_dst;
+  uint32_t dst_pitch_2 = (dst_stride << 1);
+  const int16_t *filter = &filter_x0[3];
+  uint32_t      filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+
+    dst_ptr = dst;
+    odd_dst = (dst_ptr + dst_stride);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],         0(%[src])                       \n\t"
+        "ulw              %[tp2],         4(%[src])                       \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                          \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                          \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp2]                          \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp2]                          \n\t"
+        "ulw              %[tp3],         8(%[src])                       \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]     \n\t"
+        "extp             %[Temp1],       $ac3,           31              \n\t"
+
+        /* even 2. pixel */
+        "dpa.w.ph         $ac2,           %[p2],          %[filter45]     \n\t"
+        "extp             %[Temp3],       $ac2,           31              \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[Temp2],       %[Temp1](%[cm])                 \n\t"
+        "mtlo             %[vector4a],    $ac1                            \n\t"
+        "mthi             $zero,          $ac1                            \n\t"
+        "balign           %[tp3],         %[tp2],         3              \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[filter45]     \n\t"
+        "lbux             %[tp1],         %[Temp3](%[cm])                 \n\t"
+        "extp             %[p3],          $ac1,           31              \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "sb               %[Temp2],       0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+        "sb               %[tp1],         0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+
+        "dpa.w.ph         $ac2,           %[p4],          %[filter45]     \n\t"
+        "extp             %[Temp3],       $ac2,           31              \n\t"
+
+        "lbux             %[Temp1],         %[p3](%[cm])                    \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a],    $ac1                            \n\t"
+        "mthi             $zero,          $ac1                            \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                          \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                          \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp3]                          \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp3]                          \n\t"
+        "sb               %[Temp1],       0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]     \n\t"
+        "extp             %[Temp2],       $ac3,           31              \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp1],         %[Temp3](%[cm])                 \n\t"
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "dpa.w.ph         $ac1,           %[p2],          %[filter45]     \n\t"
+        "sb               %[tp1],         0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+        "extp             %[Temp3],       $ac1,           31              \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[tp3],         %[Temp2](%[cm])                 \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[filter45]     \n\t"
+        "extp             %[Temp2],       $ac3,           31              \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[tp3],         0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[filter45]     \n\t"
+        "extp             %[Temp1],       $ac2,           31              \n\t"
+
+        /* clamp */
+        "lbux             %[p4],          %[Temp3](%[cm])                 \n\t"
+        "lbux             %[p2],          %[Temp2](%[cm])                 \n\t"
+        "lbux             %[p1],          %[Temp1](%[cm])                 \n\t"
+
+        /* store bytes */
+        "sb               %[p4],          0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+        "sb               %[p2],          0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+        "sb               %[p1],          0(%[odd_dst])                   \n\t"
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2), [tp3] "=&r" (tp3),
+          [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+          [dst_ptr] "+r" (dst_ptr), [odd_dst] "+r" (odd_dst)
+        : [filter45] "r" (filter45),[vector4a] "r" (vector4a), [cm] "r" (cm),
+          [src] "r" (src), [dst_pitch_2] "r" (dst_pitch_2)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+static void convolve_bi_horiz_16_transposed_dspr2(const uint8_t *src_ptr,
+                                                  int32_t src_stride,
+                                                  uint8_t *dst_ptr,
+                                                  int32_t dst_stride,
+                                                  const int16_t *filter_x0,
+                                                  int32_t h,
+                                                  int32_t count) {
+  int32_t       c, y;
+  const uint8_t *src;
+  uint8_t       *dst;
+  uint8_t       *cm = vpx_ff_cropTbl;
+  uint32_t      vector_64 = 64;
+  int32_t       Temp1, Temp2, Temp3;
+  uint32_t      qload1, qload2;
+  uint32_t      p1, p2, p3, p4, p5;
+  uint32_t      st1, st2, st3;
+  uint32_t      dst_pitch_2 = (dst_stride << 1);
+  uint8_t       *odd_dst;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t      filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+
+    src = src_ptr;
+    dst = dst_ptr;
+
+    odd_dst = (dst + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],        0(%[src])                       \n\t"
+          "ulw              %[qload2],        4(%[src])                       \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 1 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 2 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       \n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       \n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       \n\t"
+          "ulw              %[qload1],        8(%[src])                       \n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter45]     \n\t" /* even 1 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* even 3 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       \n\t"
+          "preceu.ph.qbl    %[p5],            %[qload1]                       \n\t"
+          "ulw              %[qload2],        12(%[src])                      \n\t"
+          "dpa.w.ph         $ac2,             %[p2],          %[filter45]     \n\t" /* even 1 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 1 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 4 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p2],            %[qload2]                       \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 1 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]             \n\t"
+          "dpa.w.ph         $ac3,             %[p3],          %[filter45]     \n\t" /* even 3 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* even 3 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 5 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbl    %[p3],            %[qload2]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 2 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac1,             %[p4],          %[filter45]     \n\t" /* even 4 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 4 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* even 6 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "sb               %[st3],           0(%[dst])                       \n\t" /* even 3 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter45]     \n\t" /* even 5 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 5 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 7 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 4 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "ulw              %[qload1],        20(%[src])                      \n\t"
+          "dpa.w.ph         $ac3,             %[p5],          %[filter45]     \n\t" /* even 6 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* even 6 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 8 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 5 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter45]     \n\t" /* even 7 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 7 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 1 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter45]     \n\t" /* even 8 */
+          "sb               %[st3],           0(%[dst])                       \n\t" /* even 6 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 8 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],        1(%[src])                       \n\t"
+          "ulw              %[qload2],        5(%[src])                       \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 2 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       \n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       \n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 7 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "ulw              %[qload2],        9(%[src])                       \n\t"
+          "dpa.w.ph         $ac3,             %[p1],          %[filter45]     \n\t" /* odd 1 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 1 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* odd 3 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 8 */
+          "ulw              %[qload1],        13(%[src])                      \n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter45]     \n\t" /* odd 2 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 2 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 4 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       \n\t"
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 1 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter45]     \n\t" /* odd 3 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* odd 3 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 5 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       \n\t"
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 2 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac3,             %[p4],          %[filter45]     \n\t" /* odd 4 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 4 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* odd 6 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "sb               %[st2],           0(%[odd_dst])                   \n\t" /* odd 3 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter45]     \n\t" /* odd 5 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 5 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 7 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 4 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "ulw              %[qload1],        21(%[src])                      \n\t"
+          "dpa.w.ph         $ac2,             %[p5],          %[filter45]     \n\t" /* odd 6 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* odd 6 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 8 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       \n\t"
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 5 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac3,             %[p2],          %[filter45]     \n\t" /* odd 7 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter45]     \n\t" /* odd 8 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 8 */
+
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* odd 6 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 7 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 8 */
+
+          "sb               %[st2],           0(%[odd_dst])                   \n\t" /* odd 6 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 7 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2), [p5] "=&r" (p5),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+            [dst] "+r" (dst), [odd_dst] "+r" (odd_dst)
+          : [filter45] "r" (filter45), [vector_64] "r" (vector_64),
+            [cm] "r" (cm),
+            [src] "r" (src), [dst_pitch_2] "r" (dst_pitch_2)
+      );
+
+      src += 16;
+      dst = (dst_ptr + ((c + 1) * 16 * dst_stride));
+      odd_dst = (dst + dst_stride);
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += 1;
+  }
+}
+
+static void convolve_bi_horiz_64_transposed_dspr2(const uint8_t *src_ptr,
+                                                  int32_t src_stride,
+                                                  uint8_t *dst_ptr,
+                                                  int32_t dst_stride,
+                                                  const int16_t *filter_x0,
+                                                  int32_t h) {
+  int32_t       c, y;
+  const uint8_t *src;
+  uint8_t       *dst;
+  uint8_t       *cm = vpx_ff_cropTbl;
+  uint32_t      vector_64 = 64;
+  int32_t       Temp1, Temp2, Temp3;
+  uint32_t      qload1, qload2;
+  uint32_t      p1, p2, p3, p4, p5;
+  uint32_t      st1, st2, st3;
+  uint32_t      dst_pitch_2 = (dst_stride << 1);
+  uint8_t       *odd_dst;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t      filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+
+    src = src_ptr;
+    dst = dst_ptr;
+
+    odd_dst = (dst + dst_stride);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],        0(%[src])                       \n\t"
+          "ulw              %[qload2],        4(%[src])                       \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 1 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 2 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       \n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       \n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       \n\t"
+          "ulw              %[qload1],        8(%[src])                       \n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter45]     \n\t" /* even 1 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* even 3 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       \n\t"
+          "preceu.ph.qbl    %[p5],            %[qload1]                       \n\t"
+          "ulw              %[qload2],        12(%[src])                      \n\t"
+          "dpa.w.ph         $ac2,             %[p2],          %[filter45]     \n\t" /* even 1 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 1 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 4 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p2],            %[qload2]                       \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 1 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]             \n\t"
+          "dpa.w.ph         $ac3,             %[p3],          %[filter45]     \n\t" /* even 3 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* even 3 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 5 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbl    %[p3],            %[qload2]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 2 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac1,             %[p4],          %[filter45]     \n\t" /* even 4 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 4 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* even 6 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "sb               %[st3],           0(%[dst])                       \n\t" /* even 3 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter45]     \n\t" /* even 5 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 5 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 7 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 4 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "ulw              %[qload1],        20(%[src])                      \n\t"
+          "dpa.w.ph         $ac3,             %[p5],          %[filter45]     \n\t" /* even 6 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* even 6 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 8 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 5 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter45]     \n\t" /* even 7 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 7 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 1 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter45]     \n\t" /* even 8 */
+          "sb               %[st3],           0(%[dst])                       \n\t" /* even 6 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 8 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],        1(%[src])                       \n\t"
+          "ulw              %[qload2],        5(%[src])                       \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 2 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       \n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       \n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 7 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "ulw              %[qload2],        9(%[src])                       \n\t"
+          "dpa.w.ph         $ac3,             %[p1],          %[filter45]     \n\t" /* odd 1 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 1 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* odd 3 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 8 */
+          "ulw              %[qload1],        13(%[src])                      \n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter45]     \n\t" /* odd 2 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 2 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 4 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       \n\t"
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 1 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter45]     \n\t" /* odd 3 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* odd 3 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 5 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       \n\t"
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 2 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac3,             %[p4],          %[filter45]     \n\t" /* odd 4 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 4 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* odd 6 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "sb               %[st2],           0(%[odd_dst])                   \n\t" /* odd 3 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter45]     \n\t" /* odd 5 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 5 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 7 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 4 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "ulw              %[qload1],        21(%[src])                      \n\t"
+          "dpa.w.ph         $ac2,             %[p5],          %[filter45]     \n\t" /* odd 6 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* odd 6 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 8 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       \n\t"
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 5 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac3,             %[p2],          %[filter45]     \n\t" /* odd 7 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter45]     \n\t" /* odd 8 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 8 */
+
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* odd 6 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 7 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 8 */
+
+          "sb               %[st2],           0(%[odd_dst])                   \n\t" /* odd 6 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 7 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2), [p5] "=&r" (p5),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+            [dst] "+r" (dst), [odd_dst] "+r" (odd_dst)
+          : [filter45] "r" (filter45), [vector_64] "r" (vector_64),
+            [cm] "r" (cm),
+            [src] "r" (src), [dst_pitch_2] "r" (dst_pitch_2)
+      );
+
+      src += 16;
+      dst = (dst_ptr + ((c + 1) * 16 * dst_stride));
+      odd_dst = (dst + dst_stride);
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += 1;
+  }
+}
+
+void convolve_bi_horiz_transposed(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter, int w, int h) {
+  int x, y;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      int sum = 0;
+
+      sum += src[x] * filter[3];
+      sum += src[x + 1] * filter[4];
+
+      dst[x * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+void vpx_convolve2_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                         uint8_t *dst, ptrdiff_t dst_stride,
+                         const int16_t *filter,
+                         int w, int h) {
+  uint32_t pos = 38;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  /* prefetch data to cache memory */
+  prefetch_load(src);
+  prefetch_load(src + 32);
+
+  switch (w) {
+    case 4:
+      convolve_bi_horiz_4_transposed_dspr2(src, src_stride,
+                                           dst, dst_stride,
+                                           filter, h);
+      break;
+    case 8:
+      convolve_bi_horiz_8_transposed_dspr2(src, src_stride,
+                                           dst, dst_stride,
+                                           filter, h);
+      break;
+    case 16:
+    case 32:
+      convolve_bi_horiz_16_transposed_dspr2(src, src_stride,
+                                            dst, dst_stride,
+                                            filter, h,
+                                            (w/16));
+      break;
+    case 64:
+      prefetch_load(src + 32);
+      convolve_bi_horiz_64_transposed_dspr2(src, src_stride,
+                                            dst, dst_stride,
+                                            filter, h);
+      break;
+    default:
+      convolve_bi_horiz_transposed(src, src_stride,
+                                   dst, dst_stride,
+                                   filter, w, h);
+      break;
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve2_horiz_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/convolve2_horiz_dspr2.c
new file mode 100644
index 0000000..9fe1a34
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve2_horiz_dspr2.c
@@ -0,0 +1,705 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/convolve_common_dspr2.h"
+#include "vpx_dsp/vpx_convolve.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_bi_horiz_4_dspr2(const uint8_t *src,
+                                      int32_t src_stride,
+                                      uint8_t *dst,
+                                      int32_t dst_stride,
+                                      const int16_t *filter_x0,
+                                      int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  int32_t Temp1, Temp2, Temp3, Temp4;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],      0(%[src])                      \n\t"
+        "ulw              %[tp2],      4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],    $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "balign           %[tp2],      %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],      %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],      %[Temp3](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp4],    $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[p1],       %[Temp2](%[cm])                \n\t"
+        "lbux             %[p2],       %[Temp4](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[tp1],      0(%[dst])                      \n\t"
+        "sb               %[p1],       1(%[dst])                      \n\t"
+        "sb               %[tp2],      2(%[dst])                      \n\t"
+        "sb               %[p2],       3(%[dst])                      \n\t"
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+          [p1] "=&r" (p1), [p2] "=&r" (p2),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+          [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4)
+        : [filter45] "r" (filter45), [vector4a] "r" (vector4a),
+          [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_horiz_8_dspr2(const uint8_t *src,
+                                      int32_t src_stride,
+                                      uint8_t *dst,
+                                      int32_t dst_stride,
+                                      const int16_t *filter_x0,
+                                      int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector4a = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2, tp3;
+  uint32_t p1, p2, p3, p4;
+  uint32_t st0, st1;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],      0(%[src])                      \n\t"
+        "ulw              %[tp2],      4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],       %[tp2]                         \n\t"
+        "ulw              %[tp3],      8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],    $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "dpa.w.ph         $ac2,        %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[st0],      %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac1                           \n\t"
+        "mthi             $zero,       $ac1                           \n\t"
+        "dpa.w.ph         $ac1,        %[p3],          %[filter45]    \n\t"
+        "extp             %[Temp1],    $ac1,           31             \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "sb               %[st0],      0(%[dst])                      \n\t"
+        "lbux             %[st1],      %[Temp3](%[cm])                \n\t"
+
+        "balign           %[tp3],      %[tp2],         3              \n\t"
+        "balign           %[tp2],      %[tp1],         3              \n\t"
+
+        "dpa.w.ph         $ac2,        %[p4],          %[filter45]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        "lbux             %[st0],      %[Temp1](%[cm])                \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a], $ac1                           \n\t"
+        "mthi             $zero,       $ac1                           \n\t"
+        "sb               %[st1],      2(%[dst])                      \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[p3],       %[tp3]                         \n\t"
+        "preceu.ph.qbl    %[p4],       %[tp3]                         \n\t"
+        "sb               %[st0],      4(%[dst])                      \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "lbux             %[st0],      %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,        %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],    $ac1,           31             \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[st1],      %[Temp2](%[cm])                \n\t"
+        "dpa.w.ph         $ac3,        %[p3],          %[filter45]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[st1],      1(%[dst])                      \n\t"
+        "sb               %[st0],      6(%[dst])                      \n\t"
+        "dpa.w.ph         $ac2,        %[p4],          %[filter45]    \n\t"
+        "extp             %[Temp1],    $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[p4],       %[Temp3](%[cm])                \n\t"
+        "lbux             %[p2],       %[Temp2](%[cm])                \n\t"
+        "lbux             %[p1],       %[Temp1](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[p4],       3(%[dst])                      \n\t"
+        "sb               %[p2],       5(%[dst])                      \n\t"
+        "sb               %[p1],       7(%[dst])                      \n\t"
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2), [tp3] "=&r" (tp3),
+          [st0] "=&r" (st0), [st1] "=&r" (st1),
+          [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+        : [filter45] "r" (filter45), [vector4a] "r" (vector4a),
+          [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_horiz_16_dspr2(const uint8_t *src_ptr,
+                                       int32_t src_stride,
+                                       uint8_t *dst_ptr,
+                                       int32_t dst_stride,
+                                       const int16_t *filter_x0,
+                                       int32_t h,
+                                       int32_t count) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_store(dst_ptr + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st1],       0(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter45]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st2],       2(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter45]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "sb               %[st3],       4(%[dst])                    \n\t" /* even 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "sb               %[st1],       6(%[dst])                    \n\t" /* even 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter45]  \n\t" /* even 6 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "sb               %[st2],       8(%[dst])                    \n\t" /* even 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* even 7 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* even 8 */
+          "sb               %[st3],       10(%[dst])                   \n\t" /* even 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                    \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st1],       12(%[dst])                   \n\t" /* even 7 */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter45]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       14(%[dst])                   \n\t" /* even 8 */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* odd 2 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st3],       1(%[dst])                    \n\t" /* odd 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* odd 3 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st1],       3(%[dst])                    \n\t" /* odd 2 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter45]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "sb               %[st2],       5(%[dst])                    \n\t" /* odd 3 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "sb               %[st3],       7(%[dst])                    \n\t" /* odd 4 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter45]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "sb               %[st1],       9(%[dst])                    \n\t" /* odd 5 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter45]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+
+          "sb               %[st2],       11(%[dst])                   \n\t" /* odd 6 */
+          "sb               %[st3],       13(%[dst])                   \n\t" /* odd 7 */
+          "sb               %[st1],       15(%[dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2), [qload3] "=&r" (qload3),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [p5] "=&r" (p5),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+          : [filter45] "r" (filter45), [vector_64] "r" (vector_64),
+            [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+      );
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void convolve_bi_horiz_64_dspr2(const uint8_t *src_ptr,
+                                       int32_t src_stride,
+                                       uint8_t *dst_ptr,
+                                       int32_t dst_stride,
+                                       const int16_t *filter_x0,
+                                       int32_t h) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+    prefetch_store(dst_ptr + dst_stride);
+    prefetch_store(dst_ptr + dst_stride + 32);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st1],       0(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter45]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st2],       2(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter45]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "sb               %[st3],       4(%[dst])                    \n\t" /* even 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "sb               %[st1],       6(%[dst])                    \n\t" /* even 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter45]  \n\t" /* even 6 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "sb               %[st2],       8(%[dst])                    \n\t" /* even 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* even 7 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* even 8 */
+          "sb               %[st3],       10(%[dst])                   \n\t" /* even 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                    \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st1],       12(%[dst])                   \n\t" /* even 7 */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter45]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       14(%[dst])                   \n\t" /* even 8 */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* odd 2 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st3],       1(%[dst])                    \n\t" /* odd 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* odd 3 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st1],       3(%[dst])                    \n\t" /* odd 2 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter45]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "sb               %[st2],       5(%[dst])                    \n\t" /* odd 3 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "sb               %[st3],       7(%[dst])                    \n\t" /* odd 4 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter45]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "sb               %[st1],       9(%[dst])                    \n\t" /* odd 5 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter45]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+
+          "sb               %[st2],       11(%[dst])                   \n\t" /* odd 6 */
+          "sb               %[st3],       13(%[dst])                   \n\t" /* odd 7 */
+          "sb               %[st1],       15(%[dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2), [qload3] "=&r" (qload3),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [p5] "=&r" (p5),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+          : [filter45] "r" (filter45), [vector_64] "r" (vector_64),
+            [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+      );
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+void vpx_convolve2_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4,
+                               int w, int h) {
+  uint32_t pos = 38;
+
+  assert(x_step_q4 == 16);
+
+  prefetch_load((const uint8_t *)filter_x);
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  /* prefetch data to cache memory */
+  prefetch_load(src);
+  prefetch_load(src + 32);
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4:
+      convolve_bi_horiz_4_dspr2(src, (int32_t)src_stride,
+                                dst, (int32_t)dst_stride,
+                                filter_x, (int32_t)h);
+      break;
+    case 8:
+      convolve_bi_horiz_8_dspr2(src, (int32_t)src_stride,
+                                dst, (int32_t)dst_stride,
+                                filter_x, (int32_t)h);
+      break;
+    case 16:
+      convolve_bi_horiz_16_dspr2(src, (int32_t)src_stride,
+                                 dst, (int32_t)dst_stride,
+                                 filter_x, (int32_t)h, 1);
+      break;
+    case 32:
+      convolve_bi_horiz_16_dspr2(src, (int32_t)src_stride,
+                                 dst, (int32_t)dst_stride,
+                                 filter_x, (int32_t)h, 2);
+      break;
+    case 64:
+      prefetch_load(src + 64);
+      prefetch_store(dst + 32);
+
+      convolve_bi_horiz_64_dspr2(src, (int32_t)src_stride,
+                                 dst, (int32_t)dst_stride,
+                                 filter_x, (int32_t)h);
+      break;
+    default:
+      vpx_convolve8_horiz_c(src, src_stride,
+                            dst, dst_stride,
+                            filter_x, x_step_q4,
+                            filter_y, y_step_q4,
+                            w, h);
+      break;
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve2_vert_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/convolve2_vert_dspr2.c
new file mode 100644
index 0000000..dde6ffd
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve2_vert_dspr2.c
@@ -0,0 +1,258 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/convolve_common_dspr2.h"
+#include "vpx_dsp/vpx_convolve.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_bi_vert_4_dspr2(const uint8_t *src,
+                                     int32_t src_stride,
+                                     uint8_t *dst,
+                                     int32_t dst_stride,
+                                     const int16_t *filter_y,
+                                     int32_t w,
+                                     int32_t h) {
+  int32_t       x, y;
+  const uint8_t *src_ptr;
+  uint8_t       *dst_ptr;
+  uint8_t       *cm = vpx_ff_cropTbl;
+  uint32_t      vector4a = 64;
+  uint32_t      load1, load2;
+  uint32_t      p1, p2;
+  uint32_t      scratch1;
+  uint32_t      store1, store2;
+  int32_t       Temp1, Temp2;
+  const int16_t *filter = &filter_y[3];
+  uint32_t      filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+
+    for (x = 0; x < w; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__ (
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]     \n\t"
+
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r" (load1), [load2] "=&r" (load2),
+            [p1] "=&r" (p1), [p2] "=&r" (p2),
+            [scratch1] "=&r" (scratch1),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+            [store1] "=&r" (store1), [store2] "=&r" (store2),
+            [src_ptr] "+r" (src_ptr)
+          : [filter45] "r" (filter45),[vector4a] "r" (vector4a),
+            [src_stride] "r" (src_stride),
+            [cm] "r" (cm), [dst_ptr] "r" (dst_ptr)
+      );
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_vert_64_dspr2(const uint8_t *src,
+                                      int32_t src_stride,
+                                      uint8_t *dst,
+                                      int32_t dst_stride,
+                                      const int16_t *filter_y,
+                                      int32_t h) {
+  int32_t       x, y;
+  const uint8_t *src_ptr;
+  uint8_t       *dst_ptr;
+  uint8_t       *cm = vpx_ff_cropTbl;
+  uint32_t      vector4a = 64;
+  uint32_t      load1, load2;
+  uint32_t      p1, p2;
+  uint32_t      scratch1;
+  uint32_t      store1, store2;
+  int32_t       Temp1, Temp2;
+  const int16_t *filter = &filter_y[3];
+  uint32_t      filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+
+    for (x = 0; x < 64; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__ (
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]     \n\t"
+
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r" (load1), [load2] "=&r" (load2),
+            [p1] "=&r" (p1), [p2] "=&r" (p2),
+            [scratch1] "=&r" (scratch1),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+            [store1] "=&r" (store1), [store2] "=&r" (store2),
+            [src_ptr] "+r" (src_ptr)
+          : [filter45] "r" (filter45),[vector4a] "r" (vector4a),
+            [src_stride] "r" (src_stride),
+            [cm] "r" (cm), [dst_ptr] "r" (dst_ptr)
+      );
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vpx_convolve2_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4,
+                              int w, int h) {
+  uint32_t pos = 38;
+
+  assert(y_step_q4 == 16);
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4 :
+    case 8 :
+    case 16 :
+    case 32 :
+      convolve_bi_vert_4_dspr2(src, src_stride,
+                               dst, dst_stride,
+                               filter_y, w, h);
+      break;
+    case 64 :
+      prefetch_store(dst + 32);
+      convolve_bi_vert_64_dspr2(src, src_stride,
+                                dst, dst_stride,
+                                filter_y, h);
+      break;
+    default:
+      vpx_convolve8_vert_c(src, src_stride,
+                           dst, dst_stride,
+                           filter_x, x_step_q4,
+                           filter_y, y_step_q4,
+                           w, h);
+      break;
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve8_avg_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/convolve8_avg_dspr2.c
new file mode 100644
index 0000000..43da9e5
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve8_avg_dspr2.c
@@ -0,0 +1,677 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/convolve_common_dspr2.h"
+#include "vpx_dsp/vpx_convolve.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_avg_vert_4_dspr2(const uint8_t *src,
+                                      int32_t src_stride,
+                                      uint8_t *dst,
+                                      int32_t dst_stride,
+                                      const int16_t *filter_y,
+                                      int32_t w,
+                                      int32_t h) {
+  int32_t       x, y;
+  const uint8_t *src_ptr;
+  uint8_t       *dst_ptr;
+  uint8_t       *cm = vpx_ff_cropTbl;
+  uint32_t      vector4a = 64;
+  uint32_t      load1, load2, load3, load4;
+  uint32_t      p1, p2;
+  uint32_t      n1, n2;
+  uint32_t      scratch1, scratch2;
+  uint32_t      store1, store2;
+  int32_t       vector1b, vector2b, vector3b, vector4b;
+  int32_t       Temp1, Temp2;
+
+  vector1b = ((const int32_t *)filter_y)[0];
+  vector2b = ((const int32_t *)filter_y)[1];
+  vector3b = ((const int32_t *)filter_y)[2];
+  vector4b = ((const int32_t *)filter_y)[3];
+
+  src -= 3 * src_stride;
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+
+    for (x = 0; x < w; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__ (
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector2b]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector2b]     \n\t"
+
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "lbu              %[scratch1],  0(%[dst_ptr])                   \n\t"
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+          "lbu              %[scratch2],  1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "dpa.w.ph         $ac2,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector4b]     \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 1 */
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector4b]     \n\t"
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 2 */
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+          "lbu              %[scratch1],  2(%[dst_ptr])                   \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  3(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 3 */
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 4 */
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r" (load1), [load2] "=&r" (load2),
+            [load3] "=&r" (load3), [load4] "=&r" (load4),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [n1] "=&r" (n1), [n2] "=&r" (n2),
+            [scratch1] "=&r" (scratch1), [scratch2] "=&r" (scratch2),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+            [store1] "=&r" (store1), [store2] "=&r" (store2),
+            [src_ptr] "+r" (src_ptr)
+          : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+            [vector3b] "r" (vector3b), [vector4b] "r" (vector4b),
+            [vector4a] "r" (vector4a),
+            [src_stride] "r" (src_stride), [cm] "r" (cm), [dst_ptr] "r" (dst_ptr)
+      );
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_avg_vert_64_dspr2(const uint8_t *src,
+                                       int32_t src_stride,
+                                       uint8_t *dst,
+                                       int32_t dst_stride,
+                                       const int16_t *filter_y,
+                                       int32_t h) {
+  int32_t       x, y;
+  const uint8_t *src_ptr;
+  uint8_t       *dst_ptr;
+  uint8_t       *cm = vpx_ff_cropTbl;
+  uint32_t      vector4a = 64;
+  uint32_t      load1, load2, load3, load4;
+  uint32_t      p1, p2;
+  uint32_t      n1, n2;
+  uint32_t      scratch1, scratch2;
+  uint32_t      store1, store2;
+  int32_t       vector1b, vector2b, vector3b, vector4b;
+  int32_t       Temp1, Temp2;
+
+  vector1b = ((const int32_t *)filter_y)[0];
+  vector2b = ((const int32_t *)filter_y)[1];
+  vector3b = ((const int32_t *)filter_y)[2];
+  vector4b = ((const int32_t *)filter_y)[3];
+
+  src -= 3 * src_stride;
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+    prefetch_store(dst + dst_stride + 32);
+
+    for (x = 0; x < 64; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__ (
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector2b]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector2b]     \n\t"
+
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "lbu              %[scratch1],  0(%[dst_ptr])                   \n\t"
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+          "lbu              %[scratch2],  1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "dpa.w.ph         $ac2,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector4b]     \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 1 */
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector4b]     \n\t"
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 2 */
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+          "lbu              %[scratch1],  2(%[dst_ptr])                   \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  3(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 3 */
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 4 */
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r" (load1), [load2] "=&r" (load2),
+            [load3] "=&r" (load3), [load4] "=&r" (load4),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [n1] "=&r" (n1), [n2] "=&r" (n2),
+            [scratch1] "=&r" (scratch1), [scratch2] "=&r" (scratch2),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+            [store1] "=&r" (store1), [store2] "=&r" (store2),
+            [src_ptr] "+r" (src_ptr)
+          : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+            [vector3b] "r" (vector3b), [vector4b] "r" (vector4b),
+            [vector4a] "r" (vector4a),
+            [src_stride] "r" (src_stride), [cm] "r" (cm), [dst_ptr] "r" (dst_ptr)
+      );
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vpx_convolve8_avg_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4,
+                                  int w, int h) {
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  if (((const int32_t *)filter_y)[0] == 0) {
+    vpx_convolve2_avg_vert_dspr2(src, src_stride,
+                                 dst, dst_stride,
+                                 filter_x, x_step_q4,
+                                 filter_y, y_step_q4,
+                                 w, h);
+  } else {
+    uint32_t pos = 38;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__ (
+      "wrdsp      %[pos],     1           \n\t"
+      :
+      : [pos] "r" (pos)
+    );
+
+    prefetch_store(dst);
+
+    switch (w) {
+      case 4:
+      case 8:
+      case 16:
+      case 32:
+        convolve_avg_vert_4_dspr2(src, src_stride,
+                                  dst, dst_stride,
+                                  filter_y, w, h);
+        break;
+      case 64:
+        prefetch_store(dst + 32);
+        convolve_avg_vert_64_dspr2(src, src_stride,
+                                   dst, dst_stride,
+                                   filter_y, h);
+        break;
+      default:
+        vpx_convolve8_avg_vert_c(src, src_stride,
+                                 dst, dst_stride,
+                                 filter_x, x_step_q4,
+                                 filter_y, y_step_q4,
+                                 w, h);
+        break;
+    }
+  }
+}
+
+void vpx_convolve8_avg_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int x_step_q4,
+                             const int16_t *filter_y, int y_step_q4,
+                             int w, int h) {
+  /* Fixed size intermediate buffer places limits on parameters. */
+  DECLARE_ALIGNED(32, uint8_t, temp[64 * 135]);
+  int32_t intermediate_height = ((h * y_step_q4) >> 4) + 7;
+
+  assert(w <= 64);
+  assert(h <= 64);
+  assert(x_step_q4 == 16);
+  assert(y_step_q4 == 16);
+
+  if (intermediate_height < h)
+    intermediate_height = h;
+
+  vpx_convolve8_horiz(src - (src_stride * 3), src_stride,
+                      temp, 64,
+                      filter_x, x_step_q4,
+                      filter_y, y_step_q4,
+                      w, intermediate_height);
+
+  vpx_convolve8_avg_vert(temp + 64 * 3, 64,
+                         dst, dst_stride,
+                         filter_x, x_step_q4,
+                         filter_y, y_step_q4,
+                         w, h);
+}
+
+void vpx_convolve_avg_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int filter_x_stride,
+                            const int16_t *filter_y, int filter_y_stride,
+                            int w, int h) {
+  int x, y;
+  uint32_t tp1, tp2, tn1;
+  uint32_t tp3, tp4, tn2;
+
+  /* prefetch data to cache memory */
+  prefetch_load(src);
+  prefetch_load(src + 32);
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4:
+      /* 1 word storage */
+      for (y = h; y--; ) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__ (
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         0(%[dst])      \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "sw               %[tn1],         0(%[dst])      \n\t"  /* store */
+
+            : [tn1] "=&r" (tn1), [tp1] "=&r" (tp1),
+              [tp2] "=&r" (tp2)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    case 8:
+      /* 2 word storage */
+      for (y = h; y--; ) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__ (
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         0(%[dst])      \n\t"
+            "ulw              %[tp3],         4(%[src])      \n\t"
+            "ulw              %[tp4],         4(%[dst])      \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "sw               %[tn1],         0(%[dst])      \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         4(%[dst])      \n\t"  /* store */
+
+            : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+              [tp3] "=&r" (tp3), [tp4] "=&r" (tp4),
+              [tn1] "=&r" (tn1), [tn2] "=&r" (tn2)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    case 16:
+      /* 4 word storage */
+      for (y = h; y--; ) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__ (
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         0(%[dst])      \n\t"
+            "ulw              %[tp3],         4(%[src])      \n\t"
+            "ulw              %[tp4],         4(%[dst])      \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         8(%[src])      \n\t"
+            "ulw              %[tp2],         8(%[dst])      \n\t"
+            "sw               %[tn1],         0(%[dst])      \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         4(%[dst])      \n\t"  /* store */
+            "ulw              %[tp3],         12(%[src])     \n\t"
+            "ulw              %[tp4],         12(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "sw               %[tn1],         8(%[dst])      \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         12(%[dst])     \n\t"  /* store */
+
+            : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+              [tp3] "=&r" (tp3), [tp4] "=&r" (tp4),
+              [tn1] "=&r" (tn1), [tn2] "=&r" (tn2)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    case 32:
+      /* 8 word storage */
+      for (y = h; y--; ) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__ (
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         0(%[dst])      \n\t"
+            "ulw              %[tp3],         4(%[src])      \n\t"
+            "ulw              %[tp4],         4(%[dst])      \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         8(%[src])      \n\t"
+            "ulw              %[tp2],         8(%[dst])      \n\t"
+            "sw               %[tn1],         0(%[dst])      \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         4(%[dst])      \n\t"  /* store */
+            "ulw              %[tp3],         12(%[src])     \n\t"
+            "ulw              %[tp4],         12(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         16(%[src])     \n\t"
+            "ulw              %[tp2],         16(%[dst])     \n\t"
+            "sw               %[tn1],         8(%[dst])      \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         12(%[dst])     \n\t"  /* store */
+            "ulw              %[tp3],         20(%[src])     \n\t"
+            "ulw              %[tp4],         20(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         24(%[src])     \n\t"
+            "ulw              %[tp2],         24(%[dst])     \n\t"
+            "sw               %[tn1],         16(%[dst])     \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         20(%[dst])     \n\t"  /* store */
+            "ulw              %[tp3],         28(%[src])     \n\t"
+            "ulw              %[tp4],         28(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "sw               %[tn1],         24(%[dst])     \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         28(%[dst])     \n\t"  /* store */
+
+            : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+              [tp3] "=&r" (tp3), [tp4] "=&r" (tp4),
+              [tn1] "=&r" (tn1), [tn2] "=&r" (tn2)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    case 64:
+      prefetch_load(src + 64);
+      prefetch_store(dst + 32);
+
+      /* 16 word storage */
+      for (y = h; y--; ) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_load(src + src_stride + 64);
+        prefetch_store(dst + dst_stride);
+        prefetch_store(dst + dst_stride + 32);
+
+        __asm__ __volatile__ (
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         0(%[dst])      \n\t"
+            "ulw              %[tp3],         4(%[src])      \n\t"
+            "ulw              %[tp4],         4(%[dst])      \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         8(%[src])      \n\t"
+            "ulw              %[tp2],         8(%[dst])      \n\t"
+            "sw               %[tn1],         0(%[dst])      \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         4(%[dst])      \n\t"  /* store */
+            "ulw              %[tp3],         12(%[src])     \n\t"
+            "ulw              %[tp4],         12(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         16(%[src])     \n\t"
+            "ulw              %[tp2],         16(%[dst])     \n\t"
+            "sw               %[tn1],         8(%[dst])      \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         12(%[dst])     \n\t"  /* store */
+            "ulw              %[tp3],         20(%[src])     \n\t"
+            "ulw              %[tp4],         20(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         24(%[src])     \n\t"
+            "ulw              %[tp2],         24(%[dst])     \n\t"
+            "sw               %[tn1],         16(%[dst])     \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         20(%[dst])     \n\t"  /* store */
+            "ulw              %[tp3],         28(%[src])     \n\t"
+            "ulw              %[tp4],         28(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         32(%[src])     \n\t"
+            "ulw              %[tp2],         32(%[dst])     \n\t"
+            "sw               %[tn1],         24(%[dst])     \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         28(%[dst])     \n\t"  /* store */
+            "ulw              %[tp3],         36(%[src])     \n\t"
+            "ulw              %[tp4],         36(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         40(%[src])     \n\t"
+            "ulw              %[tp2],         40(%[dst])     \n\t"
+            "sw               %[tn1],         32(%[dst])     \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         36(%[dst])     \n\t"  /* store */
+            "ulw              %[tp3],         44(%[src])     \n\t"
+            "ulw              %[tp4],         44(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         48(%[src])     \n\t"
+            "ulw              %[tp2],         48(%[dst])     \n\t"
+            "sw               %[tn1],         40(%[dst])     \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         44(%[dst])     \n\t"  /* store */
+            "ulw              %[tp3],         52(%[src])     \n\t"
+            "ulw              %[tp4],         52(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "ulw              %[tp1],         56(%[src])     \n\t"
+            "ulw              %[tp2],         56(%[dst])     \n\t"
+            "sw               %[tn1],         48(%[dst])     \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         52(%[dst])     \n\t"  /* store */
+            "ulw              %[tp3],         60(%[src])     \n\t"
+            "ulw              %[tp4],         60(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t"  /* average */
+            "sw               %[tn1],         56(%[dst])     \n\t"  /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t"  /* average */
+            "sw               %[tn2],         60(%[dst])     \n\t"  /* store */
+
+            : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+              [tp3] "=&r" (tp3), [tp4] "=&r" (tp4),
+              [tn1] "=&r" (tn1), [tn2] "=&r" (tn2)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    default:
+      for (y = h; y > 0; --y) {
+        for (x = 0; x < w; ++x) {
+          dst[x] = (dst[x] + src[x] + 1) >> 1;
+        }
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve8_avg_horiz_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/convolve8_avg_horiz_dspr2.c
new file mode 100644
index 0000000..db0c2a4
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve8_avg_horiz_dspr2.c
@@ -0,0 +1,1025 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/convolve_common_dspr2.h"
+#include "vpx_dsp/vpx_convolve.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_avg_horiz_4_dspr2(const uint8_t *src,
+                                       int32_t src_stride,
+                                       uint8_t *dst,
+                                       int32_t dst_stride,
+                                       const int16_t *filter_x0,
+                                       int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  int32_t  vector1b, vector2b, vector3b, vector4b;
+  int32_t  Temp1, Temp2, Temp3, Temp4;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3, p4;
+  uint32_t n1, n2, n3, n4;
+  uint32_t tn1, tn2;
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]    \n\t"
+        "ulw              %[tn2],         8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tn2]                         \n\t"
+        "balign           %[tn1],         %[tn2],         3              \n\t"
+        "balign           %[tn2],         %[tp2],         3              \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        "lbu              %[p2],          3(%[dst])                      \n\t"  /* load odd 2 */
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],         %[Temp1](%[cm])                \n\t"  /* even 1 */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "lbu              %[Temp1],       1(%[dst])                      \n\t"  /* load odd 1 */
+        "preceu.ph.qbr    %[n1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[n2],          %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[n3],          %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[n4],          %[tn2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[n1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[n2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[n3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,           %[n4],          %[vector4b]    \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tn2],         0(%[dst])                      \n\t"  /* load even 1 */
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],         %[Temp3](%[cm])                \n\t"  /* even 2 */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[n1],          %[tn1]                         \n\t"
+        "lbux             %[tn1],         %[Temp2](%[cm])                \n\t"  /* odd 1 */
+        "addqh_r.w        %[tn2],         %[tn2],         %[tp1]         \n\t"  /* average even 1 */
+        "dpa.w.ph         $ac2,           %[n2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[n3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[n4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[n1],          %[vector4b]    \n\t"
+        "extp             %[Temp4],       $ac2,           31             \n\t"
+
+        "lbu              %[tp1],         2(%[dst])                      \n\t"  /* load even 2 */
+        "sb               %[tn2],         0(%[dst])                      \n\t"  /* store even 1 */
+
+        /* clamp */
+        "addqh_r.w        %[Temp1],       %[Temp1],       %[tn1]         \n\t"  /* average odd 1 */
+        "lbux             %[n2],          %[Temp4](%[cm])                \n\t"  /* odd 2 */
+        "sb               %[Temp1],       1(%[dst])                      \n\t"  /* store odd 1 */
+
+        "addqh_r.w        %[tp1],         %[tp1],         %[tp2]         \n\t"  /* average even 2 */
+        "sb               %[tp1],         2(%[dst])                      \n\t"  /* store even 2 */
+
+        "addqh_r.w        %[p2],          %[p2],          %[n2]          \n\t"  /* average odd 2 */
+        "sb               %[p2],          3(%[dst])                      \n\t"  /* store odd 2 */
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+          [tn1] "=&r" (tn1), [tn2] "=&r" (tn2),
+          [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+          [n1] "=&r" (n1), [n2] "=&r" (n2), [n3] "=&r" (n3), [n4] "=&r" (n4),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+          [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4)
+        : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+          [vector3b] "r" (vector3b), [vector4b] "r" (vector4b),
+          [vector4a] "r" (vector4a),
+          [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_avg_horiz_8_dspr2(const uint8_t *src,
+                                       int32_t src_stride,
+                                       uint8_t *dst,
+                                       int32_t dst_stride,
+                                       const int16_t *filter_x0,
+                                       int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector4a = 64;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3, p4, n1;
+  uint32_t tn1, tn2, tn3;
+  uint32_t st0, st1;
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp2]                         \n\t"
+        "ulw              %[tn2],         8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+        "lbu              %[Temp2],       0(%[dst])                      \n\t"
+        "lbu              %[tn3],         2(%[dst])                      \n\t"
+
+        /* even 2. pixel */
+        "preceu.ph.qbr    %[p1],          %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[n1],          %[tn2]                         \n\t"
+        "ulw              %[tn1],         12(%[src])                     \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[st0],         %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac1                           \n\t"
+        "mthi             $zero,          $ac1                           \n\t"
+        "preceu.ph.qbr    %[p2],          %[tn1]                         \n\t"
+        "lbux             %[st1],         %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac1,           %[p4],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac1,           %[p1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac1,           %[n1],          %[vector4b]    \n\t"
+        "extp             %[Temp1],       $ac1,           31             \n\t"
+
+        "addqh_r.w        %[Temp2],       %[Temp2],       %[st0]         \n\t"
+        "addqh_r.w        %[tn3],         %[tn3],         %[st1]         \n\t"
+        "sb               %[Temp2],       0(%[dst])                      \n\t"
+        "sb               %[tn3],         2(%[dst])                      \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+
+        "balign           %[tn3],         %[tn1],         3              \n\t"
+        "balign           %[tn1],         %[tn2],         3              \n\t"
+        "balign           %[tn2],         %[tp2],         3              \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+
+        "lbux             %[st0],         %[Temp1](%[cm])                \n\t"
+        "lbu              %[Temp2],       4(%[dst])                      \n\t"
+        "addqh_r.w        %[Temp2],       %[Temp2],       %[st0]         \n\t"
+
+        "dpa.w.ph         $ac2,           %[p4],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[n1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector4b]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a],    $ac1                           \n\t"
+        "mthi             $zero,          $ac1                           \n\t"
+        "sb               %[Temp2],       4(%[dst])                      \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tn2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tp1],         6(%[dst])                      \n\t"
+
+        /* odd 2. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tn1]                         \n\t"
+        "preceu.ph.qbl    %[n1],          %[tn1]                         \n\t"
+        "lbux             %[st0],         %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,           %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac1,           %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac1,           %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],       $ac1,           31             \n\t"
+
+        "lbu              %[tp2],         1(%[dst])                      \n\t"
+        "lbu              %[tn2],         3(%[dst])                      \n\t"
+        "addqh_r.w        %[tp1],         %[tp1],         %[st0]         \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[st1],         %[Temp2](%[cm])                \n\t"
+        "preceu.ph.qbr    %[p2],          %[tn3]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,           %[n1],          %[vector4b]    \n\t"
+        "addqh_r.w        %[tp2],         %[tp2],         %[st1]         \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tn3],         5(%[dst])                      \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[tp2],         1(%[dst])                      \n\t"
+        "sb               %[tp1],         6(%[dst])                      \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[n1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector4b]    \n\t"
+        "extp             %[Temp1],       $ac2,           31             \n\t"
+
+        "lbu              %[tn1],         7(%[dst])                      \n\t"
+
+        /* clamp */
+        "lbux             %[p4],          %[Temp3](%[cm])                \n\t"
+        "addqh_r.w        %[tn2],         %[tn2],         %[p4]          \n\t"
+
+        "lbux             %[p2],          %[Temp2](%[cm])                \n\t"
+        "addqh_r.w        %[tn3],         %[tn3],         %[p2]          \n\t"
+
+        "lbux             %[n1],          %[Temp1](%[cm])                \n\t"
+        "addqh_r.w        %[tn1],         %[tn1],         %[n1]          \n\t"
+
+        /* store bytes */
+        "sb               %[tn2],         3(%[dst])                      \n\t"
+        "sb               %[tn3],         5(%[dst])                      \n\t"
+        "sb               %[tn1],         7(%[dst])                      \n\t"
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+          [tn1] "=&r" (tn1), [tn2] "=&r" (tn2), [tn3] "=&r" (tn3),
+          [st0] "=&r" (st0), [st1] "=&r" (st1),
+          [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+          [n1] "=&r" (n1),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+        : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+          [vector3b] "r" (vector3b), [vector4b] "r" (vector4b),
+          [vector4a] "r" (vector4a),
+          [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_avg_horiz_16_dspr2(const uint8_t *src_ptr,
+                                        int32_t src_stride,
+                                        uint8_t *dst_ptr,
+                                        int32_t dst_stride,
+                                        const int16_t *filter_x0,
+                                        int32_t h,
+                                        int32_t count) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t filter12, filter34, filter56, filter78;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_store(dst_ptr + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+          "lbu              %[st2],       0(%[dst])                    \n\t" /* load even 1 from dst */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          "lbu              %[qload3],    2(%[dst])                    \n\t" /* load even 2 from dst */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st2],       0(%[dst])                    \n\t" /* store even 1 to dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter12]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter34]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter56]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter78]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st2]       \n\t" /* average even 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    2(%[dst])                    \n\t" /* store even 2 to dst */
+          "ulw              %[qload2],    16(%[src])                   \n\t"
+          "lbu              %[qload3],    4(%[dst])                    \n\t" /* load even 3 from dst */
+          "lbu              %[qload1],    6(%[dst])                    \n\t" /* load even 4 from dst */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter12]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter34]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter56]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter78]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 3 */
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[qload3],    4(%[dst])                    \n\t" /* store even 3 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter12]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter34]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter56]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter78]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average even 4 */
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[qload1],    6(%[dst])                    \n\t" /* store even 4 to dst */
+          "ulw              %[qload3],    20(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p5],          %[filter12]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* even 6 */
+          "lbu              %[qload2],    8(%[dst])                    \n\t" /* load even 5 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 5 */
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    8(%[dst])                    \n\t" /* store even 5 to dst */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* even 7 */
+          "lbu              %[qload3],    10(%[dst])                   \n\t" /* load even 6 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          "lbu              %[st2],       12(%[dst])                   \n\t" /* load even 7 from dst */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* even 8 */
+          "sb               %[qload3],    10(%[dst])                   \n\t" /* store even 6 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* even 8 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                   \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 7 */
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       12(%[dst])                   \n\t" /* store even 7 to dst */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter12]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* odd 1 */
+          "lbu              %[qload2],    14(%[dst])                   \n\t" /* load even 8 from dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          "lbu              %[st1],       1(%[dst])                    \n\t" /* load odd 1 from dst */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 8 */
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    14(%[dst])                   \n\t" /* store even 8 to dst */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* odd 2 */
+          "lbu              %[qload3],    3(%[dst])                    \n\t" /* load odd 2 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st3],       %[st3],         %[st1]       \n\t" /* average odd 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* odd 3 */
+          "sb               %[st3],       1(%[dst])                    \n\t" /* store odd 1 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st1]       \n\t" /* average odd 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    3(%[dst])                    \n\t" /* store odd 2 to dst */
+          "lbu              %[qload1],    5(%[dst])                    \n\t" /* load odd 3 from dst */
+          "ulw              %[qload2],    17(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p4],          %[filter12]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter34]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter56]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter78]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          "lbu              %[st1],       7(%[dst])                    \n\t" /* load odd 4 from dst */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st2]       \n\t" /* average odd 3 */
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[qload1],    5(%[dst])                    \n\t" /* store odd 3 to dst */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter34]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter56]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter78]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          "lbu              %[qload1],    9(%[dst])                    \n\t" /* load odd 5 from dst */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st1],       %[st1],         %[st3]       \n\t" /* average odd 4 */
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st1],       7(%[dst])                    \n\t" /* store odd 4 to dst */
+          "ulw              %[qload3],    21(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p5],          %[filter12]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter34]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter56]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter78]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 5 */
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload1],    9(%[dst])                    \n\t" /* store odd 5 to dst */
+          "lbu              %[qload2],    11(%[dst])                   \n\t" /* load odd 6 from dst */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter12]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter34]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter56]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter78]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          "lbu              %[qload3],    13(%[dst])                   \n\t" /* load odd 7 from dst */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter12]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter34]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter56]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter78]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbu              %[qload1],    15(%[dst])                   \n\t" /* load odd 8 from dst */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average odd 6 */
+
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average odd 7 */
+
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 8 */
+
+          "sb               %[qload2],    11(%[dst])                   \n\t" /* store odd 6 to dst */
+          "sb               %[qload3],    13(%[dst])                   \n\t" /* store odd 7 to dst */
+          "sb               %[qload1],    15(%[dst])                   \n\t" /* store odd 8 to dst */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [qload3] "=&r" (qload3), [p5] "=&r" (p5),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+          : [filter12] "r" (filter12), [filter34] "r" (filter34),
+            [filter56] "r" (filter56), [filter78] "r" (filter78),
+            [vector_64] "r" (vector_64),
+            [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+      );
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void convolve_avg_horiz_64_dspr2(const uint8_t *src_ptr,
+                                        int32_t src_stride,
+                                        uint8_t *dst_ptr,
+                                        int32_t dst_stride,
+                                        const int16_t *filter_x0,
+                                        int32_t h) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t filter12, filter34, filter56, filter78;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+    prefetch_store(dst_ptr + dst_stride);
+    prefetch_store(dst_ptr + dst_stride + 32);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+          "lbu              %[st2],       0(%[dst])                    \n\t" /* load even 1 from dst */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          "lbu              %[qload3],    2(%[dst])                    \n\t" /* load even 2 from dst */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st2],       0(%[dst])                    \n\t" /* store even 1 to dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter12]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter34]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter56]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter78]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st2]       \n\t" /* average even 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    2(%[dst])                    \n\t" /* store even 2 to dst */
+          "ulw              %[qload2],    16(%[src])                   \n\t"
+          "lbu              %[qload3],    4(%[dst])                    \n\t" /* load even 3 from dst */
+          "lbu              %[qload1],    6(%[dst])                    \n\t" /* load even 4 from dst */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter12]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter34]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter56]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter78]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 3 */
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[qload3],    4(%[dst])                    \n\t" /* store even 3 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter12]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter34]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter56]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter78]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average even 4 */
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[qload1],    6(%[dst])                    \n\t" /* store even 4 to dst */
+          "ulw              %[qload3],    20(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p5],          %[filter12]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* even 6 */
+          "lbu              %[qload2],    8(%[dst])                    \n\t" /* load even 5 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 5 */
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    8(%[dst])                    \n\t" /* store even 5 to dst */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* even 7 */
+          "lbu              %[qload3],    10(%[dst])                   \n\t" /* load even 6 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          "lbu              %[st2],       12(%[dst])                   \n\t" /* load even 7 from dst */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* even 8 */
+          "sb               %[qload3],    10(%[dst])                   \n\t" /* store even 6 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* even 8 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                   \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 7 */
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       12(%[dst])                   \n\t" /* store even 7 to dst */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter12]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* odd 1 */
+          "lbu              %[qload2],    14(%[dst])                   \n\t" /* load even 8 from dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          "lbu              %[st1],       1(%[dst])                    \n\t" /* load odd 1 from dst */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 8 */
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    14(%[dst])                   \n\t" /* store even 8 to dst */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* odd 2 */
+          "lbu              %[qload3],    3(%[dst])                    \n\t" /* load odd 2 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st3],       %[st3],         %[st1]       \n\t" /* average odd 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* odd 3 */
+          "sb               %[st3],       1(%[dst])                    \n\t" /* store odd 1 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st1]       \n\t" /* average odd 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    3(%[dst])                    \n\t" /* store odd 2 to dst */
+          "lbu              %[qload1],    5(%[dst])                    \n\t" /* load odd 3 from dst */
+          "ulw              %[qload2],    17(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p4],          %[filter12]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter34]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter56]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter78]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          "lbu              %[st1],       7(%[dst])                    \n\t" /* load odd 4 from dst */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st2]       \n\t" /* average odd 3 */
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[qload1],    5(%[dst])                    \n\t" /* store odd 3 to dst */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter34]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter56]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter78]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          "lbu              %[qload1],    9(%[dst])                    \n\t" /* load odd 5 from dst */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st1],       %[st1],         %[st3]       \n\t" /* average odd 4 */
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st1],       7(%[dst])                    \n\t" /* store odd 4 to dst */
+          "ulw              %[qload3],    21(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p5],          %[filter12]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter34]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter56]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter78]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 5 */
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload1],    9(%[dst])                    \n\t" /* store odd 5 to dst */
+          "lbu              %[qload2],    11(%[dst])                   \n\t" /* load odd 6 from dst */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter12]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter34]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter56]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter78]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          "lbu              %[qload3],    13(%[dst])                   \n\t" /* load odd 7 from dst */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter12]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter34]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter56]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter78]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbu              %[qload1],    15(%[dst])                   \n\t" /* load odd 8 from dst */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average odd 6 */
+
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average odd 7 */
+
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 8 */
+
+          "sb               %[qload2],    11(%[dst])                   \n\t" /* store odd 6 to dst */
+          "sb               %[qload3],    13(%[dst])                   \n\t" /* store odd 7 to dst */
+          "sb               %[qload1],    15(%[dst])                   \n\t" /* store odd 8 to dst */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [qload3] "=&r" (qload3), [p5] "=&r" (p5),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+          : [filter12] "r" (filter12), [filter34] "r" (filter34),
+            [filter56] "r" (filter56), [filter78] "r" (filter78),
+            [vector_64] "r" (vector_64),
+            [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+      );
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+void vpx_convolve8_avg_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int x_step_q4,
+                                   const int16_t *filter_y, int y_step_q4,
+                                   int w, int h) {
+  assert(x_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+
+  if (((const int32_t *)filter_x)[0] == 0) {
+    vpx_convolve2_avg_horiz_dspr2(src, src_stride,
+                                  dst, dst_stride,
+                                  filter_x, x_step_q4,
+                                  filter_y, y_step_q4,
+                                  w, h);
+  } else {
+    uint32_t pos = 38;
+
+    src -= 3;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__ (
+      "wrdsp      %[pos],     1           \n\t"
+      :
+      : [pos] "r" (pos)
+    );
+
+    /* prefetch data to cache memory */
+    prefetch_load(src);
+    prefetch_load(src + 32);
+    prefetch_store(dst);
+
+    switch (w) {
+      case 4:
+        convolve_avg_horiz_4_dspr2(src, src_stride,
+                                   dst, dst_stride,
+                                   filter_x, h);
+        break;
+      case 8:
+        convolve_avg_horiz_8_dspr2(src, src_stride,
+                                   dst, dst_stride,
+                                   filter_x, h);
+        break;
+      case 16:
+        convolve_avg_horiz_16_dspr2(src, src_stride,
+                                    dst, dst_stride,
+                                    filter_x, h, 1);
+        break;
+      case 32:
+        convolve_avg_horiz_16_dspr2(src, src_stride,
+                                    dst, dst_stride,
+                                    filter_x, h, 2);
+        break;
+      case 64:
+        prefetch_load(src + 64);
+        prefetch_store(dst + 32);
+
+        convolve_avg_horiz_64_dspr2(src, src_stride,
+                                    dst, dst_stride,
+                                    filter_x, h);
+        break;
+      default:
+        vpx_convolve8_avg_horiz_c(src + 3, src_stride,
+                                  dst, dst_stride,
+                                  filter_x, x_step_q4,
+                                  filter_y, y_step_q4,
+                                  w, h);
+        break;
+    }
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve8_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/convolve8_dspr2.c
new file mode 100644
index 0000000..ddad186
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve8_dspr2.c
@@ -0,0 +1,1257 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/convolve_common_dspr2.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/vpx_filter.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_horiz_4_transposed_dspr2(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              const int16_t *filter_x0,
+                                              int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint8_t *dst_ptr;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3, Temp4;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3, p4;
+  uint32_t tn1, tn2;
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    dst_ptr = dst;
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]    \n\t"
+        "ulw              %[tn2],         8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tn2]                         \n\t"
+        "balign           %[tn1],         %[tn2],         3              \n\t"
+        "balign           %[tn2],         %[tp2],         3              \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],         %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tn2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],         %[Temp3](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tn1]                         \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp4],       $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[tn1],         %[Temp2](%[cm])                \n\t"
+        "lbux             %[p2],          %[Temp4](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[tp1],         0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[tn1],         0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[tp2],         0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[p2],          0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2), [tn1] "=&r" (tn1), [tn2] "=&r" (tn2),
+          [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4),
+          [dst_ptr] "+r" (dst_ptr)
+        : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+          [vector3b] "r" (vector3b), [vector4b] "r" (vector4b),
+          [vector4a] "r" (vector4a),
+          [cm] "r" (cm), [src] "r" (src), [dst_stride] "r" (dst_stride)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+static void convolve_horiz_8_transposed_dspr2(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              const int16_t *filter_x0,
+                                              int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint8_t *dst_ptr;
+  uint32_t vector4a = 64;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2, tp3;
+  uint32_t p1, p2, p3, p4, n1;
+  uint8_t *odd_dst;
+  uint32_t dst_pitch_2 = (dst_stride << 1);
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+
+    dst_ptr = dst;
+    odd_dst = (dst_ptr + dst_stride);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp2],         0(%[src])                       \n\t"
+        "ulw              %[tp1],         4(%[src])                       \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                          \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                          \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp1]                          \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp1]                          \n\t"
+        "ulw              %[tp3],         8(%[src])                       \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]     \n\t"
+        "extp             %[Temp1],       $ac3,           31              \n\t"
+
+        /* even 2. pixel */
+        "preceu.ph.qbr    %[p1],          %[tp3]                          \n\t"
+        "preceu.ph.qbl    %[n1],          %[tp3]                          \n\t"
+        "ulw              %[tp2],         12(%[src])                      \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p3],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector4b]     \n\t"
+        "extp             %[Temp3],       $ac2,           31              \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[Temp2],       %[Temp1](%[cm])                 \n\t"
+        "mtlo             %[vector4a],    $ac1                            \n\t"
+        "mthi             $zero,          $ac1                            \n\t"
+        "preceu.ph.qbr    %[p2],          %[tp2]                          \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac1,           %[p4],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac1,           %[p1],          %[vector3b]     \n\t"
+        "lbux             %[tp3],         %[Temp3](%[cm])                 \n\t"
+        "dpa.w.ph         $ac1,           %[n1],          %[vector4b]     \n\t"
+        "extp             %[p3],          $ac1,           31              \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "sb               %[Temp2],       0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+        "sb               %[tp3],         0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+
+        "ulw              %[tp1],         1(%[src])                       \n\t"
+        "ulw              %[tp3],         5(%[src])                       \n\t"
+
+        "dpa.w.ph         $ac2,           %[p4],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac2,           %[n1],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector4b]     \n\t"
+        "extp             %[Temp3],       $ac2,           31              \n\t"
+
+        "lbux             %[tp2],         %[p3](%[cm])                    \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a],    $ac1                            \n\t"
+        "mthi             $zero,          $ac1                            \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                          \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                          \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp3]                          \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp3]                          \n\t"
+        "sb               %[tp2],         0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+        "ulw              %[tp2],         9(%[src])                       \n\t"
+
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]     \n\t"
+        "extp             %[Temp2],       $ac3,           31              \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp1],         %[Temp3](%[cm])                 \n\t"
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                          \n\t"
+        "preceu.ph.qbl    %[n1],          %[tp2]                          \n\t"
+        "ulw              %[Temp1],       13(%[src])                      \n\t"
+        "dpa.w.ph         $ac1,           %[p2],          %[vector1b]     \n\t"
+        "sb               %[tp1],         0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac1,           %[p4],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac1,           %[p1],          %[vector4b]     \n\t"
+        "extp             %[Temp3],       $ac1,           31              \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[tp3],         %[Temp2](%[cm])                 \n\t"
+        "preceu.ph.qbr    %[p2],          %[Temp1]                        \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac3,           %[n1],          %[vector4b]     \n\t"
+        "extp             %[Temp2],       $ac3,           31              \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[tp3],         0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac2,           %[n1],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector4b]     \n\t"
+        "extp             %[Temp1],       $ac2,           31              \n\t"
+
+        /* clamp */
+        "lbux             %[p4],          %[Temp3](%[cm])                 \n\t"
+        "lbux             %[p2],          %[Temp2](%[cm])                 \n\t"
+        "lbux             %[n1],          %[Temp1](%[cm])                 \n\t"
+
+        /* store bytes */
+        "sb               %[p4],          0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+        "sb               %[p2],          0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+        "sb               %[n1],          0(%[odd_dst])                   \n\t"
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2), [tp3] "=&r" (tp3),
+          [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+          [n1] "=&r" (n1),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+          [dst_ptr] "+r" (dst_ptr), [odd_dst] "+r" (odd_dst)
+        : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+          [vector3b] "r" (vector3b), [vector4b] "r" (vector4b),
+          [vector4a] "r" (vector4a), [cm] "r" (cm),
+          [src] "r" (src), [dst_pitch_2] "r" (dst_pitch_2)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+static void convolve_horiz_16_transposed_dspr2(const uint8_t *src_ptr,
+                                               int32_t src_stride,
+                                               uint8_t *dst_ptr,
+                                               int32_t dst_stride,
+                                               const int16_t *filter_x0,
+                                               int32_t h,
+                                               int32_t count) {
+  int32_t c, y;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t  filter12, filter34, filter56, filter78;
+  int32_t  Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  uint32_t dst_pitch_2 = (dst_stride << 1);
+  uint8_t  *odd_dst;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+
+    src = src_ptr;
+    dst = dst_ptr;
+
+    odd_dst = (dst + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],        0(%[src])                       \n\t"
+          "ulw              %[qload2],        4(%[src])                       \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 1 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 2 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       \n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       \n\t"
+          "ulw              %[qload2],        8(%[src])                       \n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter12]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter34]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter56]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter78]     \n\t" /* even 1 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* even 3 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       \n\t"
+          "ulw              %[qload1],        12(%[src])                      \n\t"
+          "dpa.w.ph         $ac2,             %[p2],          %[filter12]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter34]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter56]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p1],          %[filter78]     \n\t" /* even 1 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 1 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 4 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 1 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]             \n\t"
+          "dpa.w.ph         $ac3,             %[p3],          %[filter12]     \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter34]     \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter56]     \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p5],          %[filter78]     \n\t" /* even 3 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* even 3 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 5 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 2 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "ulw              %[qload2],        16(%[src])                      \n\t"
+          "dpa.w.ph         $ac1,             %[p4],          %[filter12]     \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter34]     \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter56]     \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter78]     \n\t" /* even 4 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 4 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* even 6 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbr    %[p4],            %[qload2]                       \n\t"
+          "sb               %[st3],           0(%[dst])                       \n\t" /* even 3 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter12]     \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter34]     \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p2],          %[filter56]     \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter78]     \n\t" /* even 5 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 5 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 7 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbl    %[p1],            %[qload2]                       \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 4 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "ulw              %[qload1],        20(%[src])                      \n\t"
+          "dpa.w.ph         $ac3,             %[p5],          %[filter12]     \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter34]     \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter56]     \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter78]     \n\t" /* even 6 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* even 6 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 8 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 5 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter12]     \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter34]     \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter56]     \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter78]     \n\t" /* even 7 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 7 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 1 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter12]     \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter34]     \n\t" /* even 8 */
+          "sb               %[st3],           0(%[dst])                       \n\t" /* even 6 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter56]     \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter78]     \n\t" /* even 8 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 8 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],        1(%[src])                       \n\t"
+          "ulw              %[qload2],        5(%[src])                       \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 2 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       \n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       \n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 7 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "ulw              %[qload2],        9(%[src])                       \n\t"
+          "dpa.w.ph         $ac3,             %[p1],          %[filter12]     \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter34]     \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter56]     \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter78]     \n\t" /* odd 1 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 1 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* odd 3 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 8 */
+          "ulw              %[qload1],        13(%[src])                      \n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter12]     \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter34]     \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter56]     \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter78]     \n\t" /* odd 2 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 2 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 4 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       \n\t"
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 1 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter12]     \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter34]     \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p1],          %[filter56]     \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter78]     \n\t" /* odd 3 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* odd 3 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 5 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       \n\t"
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 2 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "ulw              %[qload2],        17(%[src])                      \n\t"
+          "dpa.w.ph         $ac3,             %[p4],          %[filter12]     \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter34]     \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p5],          %[filter56]     \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter78]     \n\t" /* odd 4 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 4 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* odd 6 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p4],            %[qload2]                       \n\t"
+          "sb               %[st2],           0(%[odd_dst])                   \n\t" /* odd 3 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter12]     \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter34]     \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter56]     \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter78]     \n\t" /* odd 5 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 5 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 7 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbl    %[p1],            %[qload2]                       \n\t"
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 4 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "ulw              %[qload1],        21(%[src])                      \n\t"
+          "dpa.w.ph         $ac2,             %[p5],          %[filter12]     \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p2],          %[filter34]     \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter56]     \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter78]     \n\t" /* odd 6 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* odd 6 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 8 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       \n\t"
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 5 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac3,             %[p2],          %[filter12]     \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter34]     \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter56]     \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter78]     \n\t" /* odd 7 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter12]     \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter34]     \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter56]     \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter78]     \n\t" /* odd 8 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 8 */
+
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* odd 6 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 7 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 8 */
+
+          "sb               %[st2],           0(%[odd_dst])                   \n\t" /* odd 6 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 7 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2), [p5] "=&r" (p5),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+            [dst] "+r" (dst), [odd_dst] "+r" (odd_dst)
+          : [filter12] "r" (filter12), [filter34] "r" (filter34),
+            [filter56] "r" (filter56), [filter78] "r" (filter78),
+            [vector_64] "r" (vector_64), [cm] "r" (cm),
+            [src] "r" (src), [dst_pitch_2] "r" (dst_pitch_2)
+      );
+
+      src += 16;
+      dst = (dst_ptr + ((c + 1) * 16 * dst_stride));
+      odd_dst = (dst + dst_stride);
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+
+    dst_ptr += 1;
+  }
+}
+
+static void convolve_horiz_64_transposed_dspr2(const uint8_t *src_ptr,
+                                               int32_t src_stride,
+                                               uint8_t *dst_ptr,
+                                               int32_t dst_stride,
+                                               const int16_t *filter_x0,
+                                               int32_t h) {
+  int32_t c, y;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t  filter12, filter34, filter56, filter78;
+  int32_t  Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  uint32_t dst_pitch_2 = (dst_stride << 1);
+  uint8_t  *odd_dst;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+
+    src = src_ptr;
+    dst = dst_ptr;
+
+    odd_dst = (dst + dst_stride);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],        0(%[src])                       \n\t"
+          "ulw              %[qload2],        4(%[src])                       \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 1 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 2 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       \n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       \n\t"
+          "ulw              %[qload2],        8(%[src])                       \n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter12]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter34]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter56]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter78]     \n\t" /* even 1 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* even 3 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       \n\t"
+          "ulw              %[qload1],        12(%[src])                      \n\t"
+          "dpa.w.ph         $ac2,             %[p2],          %[filter12]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter34]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter56]     \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p1],          %[filter78]     \n\t" /* even 1 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 1 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 4 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 1 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]             \n\t"
+          "dpa.w.ph         $ac3,             %[p3],          %[filter12]     \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter34]     \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter56]     \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p5],          %[filter78]     \n\t" /* even 3 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* even 3 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 5 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 2 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "ulw              %[qload2],        16(%[src])                      \n\t"
+          "dpa.w.ph         $ac1,             %[p4],          %[filter12]     \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter34]     \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter56]     \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter78]     \n\t" /* even 4 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 4 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* even 6 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbr    %[p4],            %[qload2]                       \n\t"
+          "sb               %[st3],           0(%[dst])                       \n\t" /* even 3 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter12]     \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter34]     \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p2],          %[filter56]     \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter78]     \n\t" /* even 5 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 5 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* even 7 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbl    %[p1],            %[qload2]                       \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 4 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "ulw              %[qload1],        20(%[src])                      \n\t"
+          "dpa.w.ph         $ac3,             %[p5],          %[filter12]     \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter34]     \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter56]     \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter78]     \n\t" /* even 6 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* even 6 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* even 8 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 5 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter12]     \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter34]     \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter56]     \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter78]     \n\t" /* even 7 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* even 7 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 1 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter12]     \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter34]     \n\t" /* even 8 */
+          "sb               %[st3],           0(%[dst])                       \n\t" /* even 6 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter56]     \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter78]     \n\t" /* even 8 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* even 8 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],        1(%[src])                       \n\t"
+          "ulw              %[qload2],        5(%[src])                       \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 2 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       \n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       \n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       \n\t"
+          "sb               %[st1],           0(%[dst])                       \n\t" /* even 7 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  \n\t"
+          "ulw              %[qload2],        9(%[src])                       \n\t"
+          "dpa.w.ph         $ac3,             %[p1],          %[filter12]     \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter34]     \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter56]     \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter78]     \n\t" /* odd 1 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 1 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* odd 3 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       \n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       \n\t"
+          "sb               %[st2],           0(%[dst])                       \n\t" /* even 8 */
+          "ulw              %[qload1],        13(%[src])                      \n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter12]     \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter34]     \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter56]     \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter78]     \n\t" /* odd 2 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 2 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 4 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       \n\t"
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 1 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter12]     \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter34]     \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p1],          %[filter56]     \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter78]     \n\t" /* odd 3 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* odd 3 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 5 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       \n\t"
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 2 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "ulw              %[qload2],        17(%[src])                      \n\t"
+          "dpa.w.ph         $ac3,             %[p4],          %[filter12]     \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter34]     \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p5],          %[filter56]     \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter78]     \n\t" /* odd 4 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 4 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64],     $ac2                            \n\t" /* odd 6 */
+          "mthi             $zero,            $ac2                            \n\t"
+          "preceu.ph.qbr    %[p4],            %[qload2]                       \n\t"
+          "sb               %[st2],           0(%[odd_dst])                   \n\t" /* odd 3 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter12]     \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter34]     \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter56]     \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter78]     \n\t" /* odd 5 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 5 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64],     $ac3                            \n\t" /* odd 7 */
+          "mthi             $zero,            $ac3                            \n\t"
+          "preceu.ph.qbl    %[p1],            %[qload2]                       \n\t"
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 4 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "ulw              %[qload1],        21(%[src])                      \n\t"
+          "dpa.w.ph         $ac2,             %[p5],          %[filter12]     \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p2],          %[filter34]     \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter56]     \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter78]     \n\t" /* odd 6 */
+          "extp             %[Temp2],         $ac2,           31              \n\t" /* odd 6 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64],     $ac1                            \n\t" /* odd 8 */
+          "mthi             $zero,            $ac1                            \n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       \n\t"
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 5 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+          "dpa.w.ph         $ac3,             %[p2],          %[filter12]     \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter34]     \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter56]     \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter78]     \n\t" /* odd 7 */
+          "extp             %[Temp3],         $ac3,           31              \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter12]     \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter34]     \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter56]     \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter78]     \n\t" /* odd 8 */
+          "extp             %[Temp1],         $ac1,           31              \n\t" /* odd 8 */
+
+          "lbux             %[st2],           %[Temp2](%[cm])                 \n\t" /* odd 6 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 \n\t" /* odd 7 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 \n\t" /* odd 8 */
+
+          "sb               %[st2],           0(%[odd_dst])                   \n\t" /* odd 6 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+          "sb               %[st3],           0(%[odd_dst])                   \n\t" /* odd 7 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+          "sb               %[st1],           0(%[odd_dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2), [p5] "=&r" (p5),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+            [dst] "+r" (dst), [odd_dst] "+r" (odd_dst)
+          : [filter12] "r" (filter12), [filter34] "r" (filter34),
+            [filter56] "r" (filter56), [filter78] "r" (filter78),
+            [vector_64] "r" (vector_64), [cm] "r" (cm),
+            [src] "r" (src), [dst_pitch_2] "r" (dst_pitch_2)
+      );
+
+      src += 16;
+      dst = (dst_ptr + ((c + 1) * 16 * dst_stride));
+      odd_dst = (dst + dst_stride);
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+
+    dst_ptr += 1;
+  }
+}
+
+void convolve_horiz_transposed(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter, int w, int h) {
+  int x, y, k;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      int sum = 0;
+
+      for (k = 0; k < 8; ++k)
+        sum += src[x + k] * filter[k];
+
+      dst[x * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+void copy_horiz_transposed(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           int w, int h) {
+  int x, y;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      dst[x * dst_stride] = src[x];
+    }
+
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+void vpx_convolve8_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                         uint8_t *dst, ptrdiff_t dst_stride,
+                         const int16_t *filter_x, int x_step_q4,
+                         const int16_t *filter_y, int y_step_q4,
+                         int w, int h) {
+  DECLARE_ALIGNED(32, uint8_t, temp[64 * 135]);
+  int32_t intermediate_height = ((h * y_step_q4) >> 4) + 7;
+  uint32_t pos = 38;
+
+  assert(x_step_q4 == 16);
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  if (intermediate_height < h)
+    intermediate_height = h;
+
+  /* copy the src to dst */
+  if (filter_x[3] == 0x80) {
+    copy_horiz_transposed(src - src_stride * 3, src_stride,
+                          temp, intermediate_height,
+                          w, intermediate_height);
+  } else if (((const int32_t *)filter_x)[0] == 0) {
+    vpx_convolve2_dspr2(src - src_stride * 3, src_stride,
+                        temp, intermediate_height,
+                        filter_x,
+                        w, intermediate_height);
+  } else {
+    src -= (src_stride * 3 + 3);
+
+    /* prefetch data to cache memory */
+    prefetch_load(src);
+    prefetch_load(src + 32);
+
+    switch (w) {
+      case 4:
+        convolve_horiz_4_transposed_dspr2(src, src_stride,
+                                          temp, intermediate_height,
+                                          filter_x, intermediate_height);
+        break;
+      case 8:
+        convolve_horiz_8_transposed_dspr2(src, src_stride,
+                                          temp, intermediate_height,
+                                          filter_x, intermediate_height);
+        break;
+      case 16:
+      case 32:
+        convolve_horiz_16_transposed_dspr2(src, src_stride,
+                                           temp, intermediate_height,
+                                           filter_x, intermediate_height,
+                                           (w/16));
+        break;
+      case 64:
+        prefetch_load(src + 32);
+        convolve_horiz_64_transposed_dspr2(src, src_stride,
+                                           temp, intermediate_height,
+                                           filter_x, intermediate_height);
+        break;
+      default:
+        convolve_horiz_transposed(src, src_stride,
+                                  temp, intermediate_height,
+                                  filter_x, w, intermediate_height);
+        break;
+    }
+  }
+
+  /* copy the src to dst */
+  if (filter_y[3] == 0x80) {
+    copy_horiz_transposed(temp + 3, intermediate_height,
+                          dst, dst_stride,
+                          h, w);
+  } else if (((const int32_t *)filter_y)[0] == 0) {
+    vpx_convolve2_dspr2(temp + 3, intermediate_height,
+                        dst, dst_stride,
+                        filter_y,
+                        h, w);
+  } else {
+    switch (h) {
+      case 4:
+        convolve_horiz_4_transposed_dspr2(temp, intermediate_height,
+                                          dst, dst_stride,
+                                          filter_y, w);
+        break;
+      case 8:
+        convolve_horiz_8_transposed_dspr2(temp, intermediate_height,
+                                          dst, dst_stride,
+                                          filter_y, w);
+        break;
+      case 16:
+      case 32:
+        convolve_horiz_16_transposed_dspr2(temp, intermediate_height,
+                                           dst, dst_stride,
+                                           filter_y, w, (h/16));
+        break;
+      case 64:
+        convolve_horiz_64_transposed_dspr2(temp, intermediate_height,
+                                           dst, dst_stride,
+                                           filter_y, w);
+        break;
+      default:
+        convolve_horiz_transposed(temp, intermediate_height,
+                                  dst, dst_stride,
+                                  filter_y, h, w);
+        break;
+    }
+  }
+}
+
+void vpx_convolve_copy_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int filter_x_stride,
+                             const int16_t *filter_y, int filter_y_stride,
+                             int w, int h) {
+  int x, y;
+
+  /* prefetch data to cache memory */
+  prefetch_load(src);
+  prefetch_load(src + 32);
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4:
+      {
+      uint32_t tp1;
+
+      /* 1 word storage */
+      for (y = h; y--; ) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__ (
+            "ulw              %[tp1],         (%[src])      \n\t"
+            "sw               %[tp1],         (%[dst])      \n\t"  /* store */
+
+            : [tp1] "=&r" (tp1)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      }
+      break;
+    case 8:
+      {
+      uint32_t tp1, tp2;
+
+      /* 2 word storage */
+      for (y = h; y--; ) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__ (
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         4(%[src])      \n\t"
+            "sw               %[tp1],         0(%[dst])      \n\t"  /* store */
+            "sw               %[tp2],         4(%[dst])      \n\t"  /* store */
+
+            : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      }
+      break;
+    case 16:
+      {
+      uint32_t tp1, tp2, tp3, tp4;
+
+      /* 4 word storage */
+      for (y = h; y--; ) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__ (
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         4(%[src])      \n\t"
+            "ulw              %[tp3],         8(%[src])      \n\t"
+            "ulw              %[tp4],         12(%[src])     \n\t"
+
+            "sw               %[tp1],         0(%[dst])      \n\t"  /* store */
+            "sw               %[tp2],         4(%[dst])      \n\t"  /* store */
+            "sw               %[tp3],         8(%[dst])      \n\t"  /* store */
+            "sw               %[tp4],         12(%[dst])     \n\t"  /* store */
+
+            : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+              [tp3] "=&r" (tp3), [tp4] "=&r" (tp4)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      }
+      break;
+    case 32:
+      {
+      uint32_t tp1, tp2, tp3, tp4;
+      uint32_t tp5, tp6, tp7, tp8;
+
+      /* 8 word storage */
+      for (y = h; y--; ) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__ (
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         4(%[src])      \n\t"
+            "ulw              %[tp3],         8(%[src])      \n\t"
+            "ulw              %[tp4],         12(%[src])     \n\t"
+            "ulw              %[tp5],         16(%[src])     \n\t"
+            "ulw              %[tp6],         20(%[src])     \n\t"
+            "ulw              %[tp7],         24(%[src])     \n\t"
+            "ulw              %[tp8],         28(%[src])     \n\t"
+
+            "sw               %[tp1],         0(%[dst])      \n\t"  /* store */
+            "sw               %[tp2],         4(%[dst])      \n\t"  /* store */
+            "sw               %[tp3],         8(%[dst])      \n\t"  /* store */
+            "sw               %[tp4],         12(%[dst])     \n\t"  /* store */
+            "sw               %[tp5],         16(%[dst])     \n\t"  /* store */
+            "sw               %[tp6],         20(%[dst])     \n\t"  /* store */
+            "sw               %[tp7],         24(%[dst])     \n\t"  /* store */
+            "sw               %[tp8],         28(%[dst])     \n\t"  /* store */
+
+            : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+              [tp3] "=&r" (tp3), [tp4] "=&r" (tp4),
+              [tp5] "=&r" (tp5), [tp6] "=&r" (tp6),
+              [tp7] "=&r" (tp7), [tp8] "=&r" (tp8)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      }
+      break;
+    case 64:
+      {
+      uint32_t tp1, tp2, tp3, tp4;
+      uint32_t tp5, tp6, tp7, tp8;
+
+      prefetch_load(src + 64);
+      prefetch_store(dst + 32);
+
+      /* 16 word storage */
+      for (y = h; y--; ) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_load(src + src_stride + 64);
+        prefetch_store(dst + dst_stride);
+        prefetch_store(dst + dst_stride + 32);
+
+        __asm__ __volatile__ (
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         4(%[src])      \n\t"
+            "ulw              %[tp3],         8(%[src])      \n\t"
+            "ulw              %[tp4],         12(%[src])     \n\t"
+            "ulw              %[tp5],         16(%[src])     \n\t"
+            "ulw              %[tp6],         20(%[src])     \n\t"
+            "ulw              %[tp7],         24(%[src])     \n\t"
+            "ulw              %[tp8],         28(%[src])     \n\t"
+
+            "sw               %[tp1],         0(%[dst])      \n\t"  /* store */
+            "sw               %[tp2],         4(%[dst])      \n\t"  /* store */
+            "sw               %[tp3],         8(%[dst])      \n\t"  /* store */
+            "sw               %[tp4],         12(%[dst])     \n\t"  /* store */
+            "sw               %[tp5],         16(%[dst])     \n\t"  /* store */
+            "sw               %[tp6],         20(%[dst])     \n\t"  /* store */
+            "sw               %[tp7],         24(%[dst])     \n\t"  /* store */
+            "sw               %[tp8],         28(%[dst])     \n\t"  /* store */
+
+            "ulw              %[tp1],         32(%[src])     \n\t"
+            "ulw              %[tp2],         36(%[src])     \n\t"
+            "ulw              %[tp3],         40(%[src])     \n\t"
+            "ulw              %[tp4],         44(%[src])     \n\t"
+            "ulw              %[tp5],         48(%[src])     \n\t"
+            "ulw              %[tp6],         52(%[src])     \n\t"
+            "ulw              %[tp7],         56(%[src])     \n\t"
+            "ulw              %[tp8],         60(%[src])     \n\t"
+
+            "sw               %[tp1],         32(%[dst])     \n\t"  /* store */
+            "sw               %[tp2],         36(%[dst])     \n\t"  /* store */
+            "sw               %[tp3],         40(%[dst])     \n\t"  /* store */
+            "sw               %[tp4],         44(%[dst])     \n\t"  /* store */
+            "sw               %[tp5],         48(%[dst])     \n\t"  /* store */
+            "sw               %[tp6],         52(%[dst])     \n\t"  /* store */
+            "sw               %[tp7],         56(%[dst])     \n\t"  /* store */
+            "sw               %[tp8],         60(%[dst])     \n\t"  /* store */
+
+            : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+              [tp3] "=&r" (tp3), [tp4] "=&r" (tp4),
+              [tp5] "=&r" (tp5), [tp6] "=&r" (tp6),
+              [tp7] "=&r" (tp7), [tp8] "=&r" (tp8)
+            : [src] "r" (src), [dst] "r" (dst)
+        );
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      }
+      break;
+    default:
+      for (y = h; y--; ) {
+        for (x = 0; x < w; ++x) {
+          dst[x] = src[x];
+        }
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve8_horiz_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/convolve8_horiz_dspr2.c
new file mode 100644
index 0000000..ae78bab
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve8_horiz_dspr2.c
@@ -0,0 +1,910 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/convolve_common_dspr2.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/vpx_filter.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_horiz_4_dspr2(const uint8_t *src,
+                                   int32_t src_stride,
+                                   uint8_t *dst,
+                                   int32_t dst_stride,
+                                   const int16_t *filter_x0,
+                                   int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3, Temp4;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3, p4;
+  uint32_t n1, n2, n3, n4;
+  uint32_t tn1, tn2;
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],      0(%[src])                      \n\t"
+        "ulw              %[tp2],      4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],       %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p3],          %[vector3b]    \n\t"
+        "ulw              %[tn2],      8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,        %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp1],    $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tn2]                         \n\t"
+        "balign           %[tn1],      %[tn2],         3              \n\t"
+        "balign           %[tn2],      %[tp2],         3              \n\t"
+        "balign           %[tp2],      %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],      %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "preceu.ph.qbr    %[n1],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[n2],       %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[n3],       %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[n4],       %[tn2]                         \n\t"
+        "dpa.w.ph         $ac3,        %[n1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,        %[n2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,        %[n3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,        %[n4],          %[vector4b]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],      %[Temp3](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "preceu.ph.qbr    %[n1],       %[tn1]                         \n\t"
+        "dpa.w.ph         $ac2,        %[n2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,        %[n3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,        %[n4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,        %[n1],          %[vector4b]    \n\t"
+        "extp             %[Temp4],    $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[tn1],      %[Temp2](%[cm])                \n\t"
+        "lbux             %[n2],       %[Temp4](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[tp1],      0(%[dst])                      \n\t"
+        "sb               %[tn1],      1(%[dst])                      \n\t"
+        "sb               %[tp2],      2(%[dst])                      \n\t"
+        "sb               %[n2],       3(%[dst])                      \n\t"
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+          [tn1] "=&r" (tn1), [tn2] "=&r" (tn2),
+          [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+          [n1] "=&r" (n1), [n2] "=&r" (n2), [n3] "=&r" (n3), [n4] "=&r" (n4),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+          [Temp3] "=&r" (Temp3), [Temp4] "=&r" (Temp4)
+        : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+          [vector3b] "r" (vector3b), [vector4b] "r" (vector4b),
+          [vector4a] "r" (vector4a),
+          [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_horiz_8_dspr2(const uint8_t *src,
+                                   int32_t src_stride,
+                                   uint8_t *dst,
+                                   int32_t dst_stride,
+                                   const int16_t *filter_x0,
+                                   int32_t h) {
+  int32_t y;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector4a = 64;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3, p4, n1;
+  uint32_t tn1, tn2, tn3;
+  uint32_t st0, st1;
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__ (
+        "ulw              %[tp1],      0(%[src])                      \n\t"
+        "ulw              %[tp2],      4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],       %[tp2]                         \n\t"
+        "ulw              %[tn2],      8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp1],    $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "preceu.ph.qbr    %[p1],       %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[n1],       %[tn2]                         \n\t"
+        "ulw              %[tn1],      12(%[src])                     \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[st0],      %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac1                           \n\t"
+        "mthi             $zero,       $ac1                           \n\t"
+        "preceu.ph.qbr    %[p2],       %[tn1]                         \n\t"
+        "dpa.w.ph         $ac1,        %[p3],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac1,        %[p4],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac1,        %[p1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac1,        %[n1],          %[vector4b]    \n\t"
+        "extp             %[Temp1],    $ac1,           31             \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "sb               %[st0],      0(%[dst])                      \n\t"
+        "lbux             %[st1],      %[Temp3](%[cm])                \n\t"
+
+        "balign           %[tn3],      %[tn1],         3              \n\t"
+        "balign           %[tn1],      %[tn2],         3              \n\t"
+        "balign           %[tn2],      %[tp2],         3              \n\t"
+        "balign           %[tp2],      %[tp1],         3              \n\t"
+
+        "dpa.w.ph         $ac2,        %[p4],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p1],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,        %[n1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[vector4b]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        "lbux             %[st0],      %[Temp1](%[cm])                \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a], $ac1                           \n\t"
+        "mthi             $zero,       $ac1                           \n\t"
+        "sb               %[st1],      2(%[dst])                      \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[p3],       %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[p4],       %[tn2]                         \n\t"
+        "sb               %[st0],      4(%[dst])                      \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tn1]                         \n\t"
+        "preceu.ph.qbl    %[n1],       %[tn1]                         \n\t"
+        "lbux             %[st0],      %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,        %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac1,        %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac1,        %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac1,        %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],    $ac1,           31             \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[st1],      %[Temp2](%[cm])                \n\t"
+        "preceu.ph.qbr    %[p2],       %[tn3]                         \n\t"
+        "dpa.w.ph         $ac3,        %[p3],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p4],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,        %[n1],          %[vector4b]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[st1],      1(%[dst])                      \n\t"
+        "sb               %[st0],      6(%[dst])                      \n\t"
+        "dpa.w.ph         $ac2,        %[p4],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p1],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,        %[n1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[vector4b]    \n\t"
+        "extp             %[Temp1],    $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[p4],       %[Temp3](%[cm])                \n\t"
+        "lbux             %[p2],       %[Temp2](%[cm])                \n\t"
+        "lbux             %[n1],       %[Temp1](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[p4],       3(%[dst])                      \n\t"
+        "sb               %[p2],       5(%[dst])                      \n\t"
+        "sb               %[n1],       7(%[dst])                      \n\t"
+
+        : [tp1] "=&r" (tp1), [tp2] "=&r" (tp2),
+          [tn1] "=&r" (tn1), [tn2] "=&r" (tn2), [tn3] "=&r" (tn3),
+          [st0] "=&r" (st0), [st1] "=&r" (st1),
+          [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+          [n1] "=&r" (n1),
+          [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+        : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+          [vector3b] "r" (vector3b), [vector4b] "r" (vector4b),
+          [vector4a] "r" (vector4a),
+          [cm] "r" (cm), [dst] "r" (dst), [src] "r" (src)
+    );
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_horiz_16_dspr2(const uint8_t *src_ptr,
+                                    int32_t src_stride,
+                                    uint8_t *dst_ptr,
+                                    int32_t dst_stride,
+                                    const int16_t *filter_x0,
+                                    int32_t h,
+                                    int32_t count) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t filter12, filter34, filter56, filter78;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_store(dst_ptr + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st1],       0(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter12]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter34]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter56]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter78]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st2],       2(%[dst])                    \n\t" /* even 1 */
+          "ulw              %[qload2],    16(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p4],          %[filter12]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter34]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter56]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter78]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st3],       4(%[dst])                    \n\t" /* even 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter12]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter34]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter56]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter78]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st1],       6(%[dst])                    \n\t" /* even 4 */
+          "ulw              %[qload3],    20(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p5],          %[filter12]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* even 6 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       8(%[dst])                    \n\t" /* even 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* even 7 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* even 8 */
+          "sb               %[st3],       10(%[dst])                   \n\t" /* even 6 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* even 8 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                    \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st1],       12(%[dst])                   \n\t" /* even 7 */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter12]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       14(%[dst])                   \n\t" /* even 8 */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* odd 2 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st3],       1(%[dst])                    \n\t" /* odd 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* odd 3 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st1],       3(%[dst])                    \n\t" /* odd 2 */
+          "ulw              %[qload2],    17(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p4],          %[filter12]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter34]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter56]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter78]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       5(%[dst])                    \n\t" /* odd 3 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter34]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter56]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter78]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st3],       7(%[dst])                    \n\t" /* odd 4 */
+          "ulw              %[qload3],    21(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p5],          %[filter12]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter34]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter56]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter78]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st1],       9(%[dst])                    \n\t" /* odd 5 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter12]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter34]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter56]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter78]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter12]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter34]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter56]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter78]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+
+          "sb               %[st2],       11(%[dst])                   \n\t" /* odd 6 */
+          "sb               %[st3],       13(%[dst])                   \n\t" /* odd 7 */
+          "sb               %[st1],       15(%[dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2), [qload3] "=&r" (qload3),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [p5] "=&r" (p5),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+          : [filter12] "r" (filter12), [filter34] "r" (filter34),
+            [filter56] "r" (filter56), [filter78] "r" (filter78),
+            [vector_64] "r" (vector_64),
+            [cm] "r" (cm), [dst] "r" (dst),
+            [src] "r" (src)
+      );
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void convolve_horiz_64_dspr2(const uint8_t *src_ptr,
+                                    int32_t src_stride,
+                                    uint8_t *dst_ptr,
+                                    int32_t dst_stride,
+                                    const int16_t *filter_x0,
+                                    int32_t h) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t filter12, filter34, filter56, filter78;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+    prefetch_store(dst_ptr + dst_stride);
+    prefetch_store(dst_ptr + dst_stride + 32);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__ (
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st1],       0(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter12]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter34]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter56]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter78]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st2],       2(%[dst])                    \n\t" /* even 1 */
+          "ulw              %[qload2],    16(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p4],          %[filter12]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter34]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter56]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter78]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st3],       4(%[dst])                    \n\t" /* even 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter12]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter34]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter56]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter78]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st1],       6(%[dst])                    \n\t" /* even 4 */
+          "ulw              %[qload3],    20(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p5],          %[filter12]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* even 6 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       8(%[dst])                    \n\t" /* even 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* even 7 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* even 8 */
+          "sb               %[st3],       10(%[dst])                   \n\t" /* even 6 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* even 8 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                    \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st1],       12(%[dst])                   \n\t" /* even 7 */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter12]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       14(%[dst])                   \n\t" /* even 8 */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* odd 2 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st3],       1(%[dst])                    \n\t" /* odd 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* odd 3 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st1],       3(%[dst])                    \n\t" /* odd 2 */
+          "ulw              %[qload2],    17(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p4],          %[filter12]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter34]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter56]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter78]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       5(%[dst])                    \n\t" /* odd 3 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter34]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter56]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter78]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st3],       7(%[dst])                    \n\t" /* odd 4 */
+          "ulw              %[qload3],    21(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p5],          %[filter12]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter34]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter56]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter78]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st1],       9(%[dst])                    \n\t" /* odd 5 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter12]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter34]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter56]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter78]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter12]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter34]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter56]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter78]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+
+          "sb               %[st2],       11(%[dst])                   \n\t" /* odd 6 */
+          "sb               %[st3],       13(%[dst])                   \n\t" /* odd 7 */
+          "sb               %[st1],       15(%[dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r" (qload1), [qload2] "=&r" (qload2), [qload3] "=&r" (qload3),
+            [st1] "=&r" (st1), [st2] "=&r" (st2), [st3] "=&r" (st3),
+            [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4),
+            [p5] "=&r" (p5),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3)
+          : [filter12] "r" (filter12), [filter34] "r" (filter34),
+            [filter56] "r" (filter56), [filter78] "r" (filter78),
+            [vector_64] "r" (vector_64),
+            [cm] "r" (cm), [dst] "r" (dst),
+            [src] "r" (src)
+      );
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+void vpx_convolve8_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4,
+                               int w, int h) {
+  assert(x_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+
+  if (((const int32_t *)filter_x)[0] == 0) {
+    vpx_convolve2_horiz_dspr2(src, src_stride,
+                              dst, dst_stride,
+                              filter_x, x_step_q4,
+                              filter_y, y_step_q4,
+                              w, h);
+  } else {
+    uint32_t pos = 38;
+
+    prefetch_load((const uint8_t *)filter_x);
+    src -= 3;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__ (
+      "wrdsp      %[pos],     1           \n\t"
+      :
+      : [pos] "r" (pos)
+    );
+
+    /* prefetch data to cache memory */
+    prefetch_load(src);
+    prefetch_load(src + 32);
+    prefetch_store(dst);
+
+    switch (w) {
+      case 4:
+        convolve_horiz_4_dspr2(src, (int32_t)src_stride,
+                               dst, (int32_t)dst_stride,
+                               filter_x, (int32_t)h);
+        break;
+      case 8:
+        convolve_horiz_8_dspr2(src, (int32_t)src_stride,
+                               dst, (int32_t)dst_stride,
+                               filter_x, (int32_t)h);
+        break;
+      case 16:
+        convolve_horiz_16_dspr2(src, (int32_t)src_stride,
+                                dst, (int32_t)dst_stride,
+                                filter_x, (int32_t)h, 1);
+        break;
+      case 32:
+        convolve_horiz_16_dspr2(src, (int32_t)src_stride,
+                                dst, (int32_t)dst_stride,
+                                filter_x, (int32_t)h, 2);
+        break;
+      case 64:
+        prefetch_load(src + 64);
+        prefetch_store(dst + 32);
+
+        convolve_horiz_64_dspr2(src, (int32_t)src_stride,
+                                dst, (int32_t)dst_stride,
+                                filter_x, (int32_t)h);
+        break;
+      default:
+        vpx_convolve8_horiz_c(src + 3, src_stride,
+                              dst, dst_stride,
+                              filter_x, x_step_q4,
+                              filter_y, y_step_q4,
+                              w, h);
+        break;
+    }
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve8_vert_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/convolve8_vert_dspr2.c
new file mode 100644
index 0000000..d553828
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve8_vert_dspr2.c
@@ -0,0 +1,383 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/convolve_common_dspr2.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/vpx_filter.h"
+#include "vpx_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_vert_4_dspr2(const uint8_t *src,
+                                  int32_t src_stride,
+                                  uint8_t *dst,
+                                  int32_t dst_stride,
+                                  const int16_t *filter_y,
+                                  int32_t w,
+                                  int32_t h) {
+  int32_t x, y;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector4a = 64;
+  uint32_t load1, load2, load3, load4;
+  uint32_t p1, p2;
+  uint32_t n1, n2;
+  uint32_t scratch1, scratch2;
+  uint32_t store1, store2;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2;
+
+  vector1b = ((const int32_t *)filter_y)[0];
+  vector2b = ((const int32_t *)filter_y)[1];
+  vector3b = ((const int32_t *)filter_y)[2];
+  vector4b = ((const int32_t *)filter_y)[3];
+
+  src -= 3 * src_stride;
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+
+    for (x = 0; x < w; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__ (
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector2b]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector2b]     \n\t"
+
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "dpa.w.ph         $ac2,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r" (load1), [load2] "=&r" (load2),
+            [load3] "=&r" (load3), [load4] "=&r" (load4),
+            [p1] "=&r" (p1), [p2] "=&r" (p2),
+            [n1] "=&r" (n1), [n2] "=&r" (n2),
+            [scratch1] "=&r" (scratch1), [scratch2] "=&r" (scratch2),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+            [store1] "=&r" (store1), [store2] "=&r" (store2),
+            [src_ptr] "+r" (src_ptr)
+          : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+            [vector3b] "r" (vector3b), [vector4b] "r" (vector4b),
+            [vector4a] "r" (vector4a), [src_stride] "r" (src_stride),
+            [cm] "r" (cm), [dst_ptr] "r" (dst_ptr)
+      );
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_vert_64_dspr2(const uint8_t *src,
+                                   int32_t src_stride,
+                                   uint8_t *dst,
+                                   int32_t dst_stride,
+                                   const int16_t *filter_y,
+                                   int32_t h) {
+  int32_t x, y;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr;
+  uint8_t *cm = vpx_ff_cropTbl;
+  uint32_t vector4a = 64;
+  uint32_t load1, load2, load3, load4;
+  uint32_t p1, p2;
+  uint32_t n1, n2;
+  uint32_t scratch1, scratch2;
+  uint32_t store1, store2;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2;
+
+  vector1b = ((const int32_t *)filter_y)[0];
+  vector2b = ((const int32_t *)filter_y)[1];
+  vector3b = ((const int32_t *)filter_y)[2];
+  vector4b = ((const int32_t *)filter_y)[3];
+
+  src -= 3 * src_stride;
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+    prefetch_store(dst + dst_stride + 32);
+
+    for (x = 0; x < 64; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__ (
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector2b]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector2b]     \n\t"
+
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "dpa.w.ph         $ac2,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r" (load1), [load2] "=&r" (load2),
+            [load3] "=&r" (load3), [load4] "=&r" (load4),
+            [p1] "=&r" (p1), [p2] "=&r" (p2),
+            [n1] "=&r" (n1), [n2] "=&r" (n2),
+            [scratch1] "=&r" (scratch1), [scratch2] "=&r" (scratch2),
+            [Temp1] "=&r" (Temp1), [Temp2] "=&r" (Temp2),
+            [store1] "=&r" (store1), [store2] "=&r" (store2),
+            [src_ptr] "+r" (src_ptr)
+          : [vector1b] "r" (vector1b), [vector2b] "r" (vector2b),
+            [vector3b] "r" (vector3b), [vector4b] "r" (vector4b),
+            [vector4a] "r" (vector4a), [src_stride] "r" (src_stride),
+            [cm] "r" (cm), [dst_ptr] "r" (dst_ptr)
+      );
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vpx_convolve8_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4,
+                              int w, int h) {
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  if (((const int32_t *)filter_y)[0] == 0) {
+    vpx_convolve2_vert_dspr2(src, src_stride,
+                             dst, dst_stride,
+                             filter_x, x_step_q4,
+                             filter_y, y_step_q4,
+                             w, h);
+  } else {
+    uint32_t pos = 38;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__ (
+      "wrdsp      %[pos],     1           \n\t"
+      :
+      : [pos] "r" (pos)
+    );
+
+    prefetch_store(dst);
+
+    switch (w) {
+      case 4 :
+      case 8 :
+      case 16 :
+      case 32 :
+        convolve_vert_4_dspr2(src, src_stride,
+                              dst, dst_stride,
+                              filter_y, w, h);
+        break;
+      case 64 :
+        prefetch_store(dst + 32);
+        convolve_vert_64_dspr2(src, src_stride,
+                               dst, dst_stride,
+                               filter_y, h);
+        break;
+      default:
+        vpx_convolve8_vert_c(src, src_stride,
+                             dst, dst_stride,
+                             filter_x, x_step_q4,
+                             filter_y, y_step_q4,
+                             w, h);
+        break;
+    }
+  }
+}
+
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/mips/convolve_common_dspr2.h b/libvpx/libvpx/vpx_dsp/mips/convolve_common_dspr2.h
new file mode 100644
index 0000000..66d77a2
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/convolve_common_dspr2.h
@@ -0,0 +1,59 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_MIPS_VPX_COMMON_DSPR2_H_
+#define VPX_DSP_MIPS_VPX_COMMON_DSPR2_H_
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/mips/common_dspr2.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if HAVE_DSPR2
+void vpx_convolve2_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4,
+                               int w, int h);
+
+void vpx_convolve2_avg_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int x_step_q4,
+                                   const int16_t *filter_y, int y_step_q4,
+                                   int w, int h);
+
+void vpx_convolve2_avg_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4,
+                                  int w, int h);
+
+void vpx_convolve2_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                         uint8_t *dst, ptrdiff_t dst_stride,
+                         const int16_t *filter,
+                         int w, int h);
+
+void vpx_convolve2_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4,
+                              int w, int h);
+
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_MIPS_VPX_COMMON_DSPR2_H_
diff --git a/libvpx/libvpx/vpx_dsp/mips/fwd_dct32x32_msa.c b/libvpx/libvpx/vpx_dsp/mips/fwd_dct32x32_msa.c
new file mode 100644
index 0000000..f29c14b
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/fwd_dct32x32_msa.c
@@ -0,0 +1,953 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/fwd_txfm_msa.h"
+
+static void fdct8x32_1d_column_load_butterfly(const int16_t *input,
+                                              int32_t src_stride,
+                                              int16_t *temp_buff) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 step0, step1, step2, step3;
+  v8i16 in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1;
+  v8i16 step0_1, step1_1, step2_1, step3_1;
+
+  /* 1st and 2nd set */
+  LD_SH4(input, src_stride, in0, in1, in2, in3);
+  LD_SH4(input + (28 * src_stride), src_stride, in4, in5, in6, in7);
+  LD_SH4(input + (4 * src_stride), src_stride, in0_1, in1_1, in2_1, in3_1);
+  LD_SH4(input + (24 * src_stride), src_stride, in4_1, in5_1, in6_1, in7_1);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+  SLLI_4V(in0_1, in1_1, in2_1, in3_1, 2);
+  SLLI_4V(in4_1, in5_1, in6_1, in7_1, 2);
+  BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7,
+              step0, step1, step2, step3, in4, in5, in6, in7);
+  BUTTERFLY_8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+              step0_1, step1_1, step2_1, step3_1, in4_1, in5_1, in6_1, in7_1);
+  ST_SH4(step0, step1, step2, step3, temp_buff, 8);
+  ST_SH4(in4, in5, in6, in7, temp_buff + (28 * 8), 8);
+  ST_SH4(step0_1, step1_1, step2_1, step3_1, temp_buff + (4 * 8), 8);
+  ST_SH4(in4_1, in5_1, in6_1, in7_1, temp_buff + (24 * 8), 8);
+
+  /* 3rd and 4th set */
+  LD_SH4(input + (8 * src_stride), src_stride, in0, in1, in2, in3);
+  LD_SH4(input + (20 * src_stride), src_stride, in4, in5, in6, in7);
+  LD_SH4(input + (12 * src_stride), src_stride, in0_1, in1_1, in2_1, in3_1);
+  LD_SH4(input + (16 * src_stride), src_stride, in4_1, in5_1, in6_1, in7_1);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+  SLLI_4V(in0_1, in1_1, in2_1, in3_1, 2);
+  SLLI_4V(in4_1, in5_1, in6_1, in7_1, 2);
+  BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7,
+              step0, step1, step2, step3, in4, in5, in6, in7);
+  BUTTERFLY_8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+              step0_1, step1_1, step2_1, step3_1, in4_1, in5_1, in6_1, in7_1);
+  ST_SH4(step0, step1, step2, step3, temp_buff + (8 * 8), 8);
+  ST_SH4(in4, in5, in6, in7, temp_buff + (20 * 8), 8);
+  ST_SH4(step0_1, step1_1, step2_1, step3_1, temp_buff + (12 * 8), 8);
+  ST_SH4(in4_1, in5_1, in6_1, in7_1, temp_buff + (15 * 8) + 8, 8);
+}
+
+static void fdct8x32_1d_column_even_store(int16_t *input, int16_t *temp) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8i16 temp0, temp1;
+
+  /* fdct even */
+  LD_SH4(input, 8, in0, in1, in2, in3);
+  LD_SH4(input + 96, 8, in12, in13, in14, in15);
+  BUTTERFLY_8(in0, in1, in2, in3, in12, in13, in14, in15,
+              vec0, vec1, vec2, vec3, in12, in13, in14, in15);
+  LD_SH4(input + 32, 8, in4, in5, in6, in7);
+  LD_SH4(input + 64, 8, in8, in9, in10, in11);
+  BUTTERFLY_8(in4, in5, in6, in7, in8, in9, in10, in11,
+              vec4, vec5, vec6, vec7, in8, in9, in10, in11);
+
+  /* Stage 3 */
+  ADD4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, in0, in1, in2, in3);
+  BUTTERFLY_4(in0, in1, in2, in3, temp0, in4, in1, in0);
+  DOTP_CONST_PAIR(temp0, in4, cospi_16_64, cospi_16_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp);
+  ST_SH(temp1, temp + 512);
+
+  DOTP_CONST_PAIR(in0, in1, cospi_24_64, cospi_8_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 256);
+  ST_SH(temp1, temp + 768);
+
+  SUB4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, vec7, vec6, vec5, vec4);
+  DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+  ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 128);
+  ST_SH(temp1, temp + 896);
+
+  SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+  DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 640);
+  ST_SH(temp1, temp + 384);
+
+  DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+  DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+  ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+  DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+  ADD2(in0, in1, in2, in3, vec0, vec7);
+  DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 64);
+  ST_SH(temp1, temp + 960);
+
+  SUB2(in0, in1, in2, in3, in0, in2);
+  DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 576);
+  ST_SH(temp1, temp + 448);
+
+  SUB2(in9, vec2, in14, vec5, vec2, vec5);
+  DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+  SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5);
+  DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 320);
+  ST_SH(temp1, temp + 704);
+
+  ADD2(in3, in2, in0, in1, vec3, vec4);
+  DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, temp0, temp1);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 192);
+  ST_SH(temp1, temp + 832);
+}
+
+static void fdct8x32_1d_column_odd_store(int16_t *input, int16_t *temp_ptr) {
+  v8i16 in16, in17, in18, in19, in20, in21, in22, in23;
+  v8i16 in24, in25, in26, in27, in28, in29, in30, in31, vec4, vec5;
+
+  in20 = LD_SH(input + 32);
+  in21 = LD_SH(input + 40);
+  in26 = LD_SH(input + 80);
+  in27 = LD_SH(input + 88);
+
+  DOTP_CONST_PAIR(in27, in20, cospi_16_64, cospi_16_64, in20, in27);
+  DOTP_CONST_PAIR(in26, in21, cospi_16_64, cospi_16_64, in21, in26);
+
+  in18 = LD_SH(input + 16);
+  in19 = LD_SH(input + 24);
+  in28 = LD_SH(input + 96);
+  in29 = LD_SH(input + 104);
+
+  vec4 = in19 - in20;
+  ST_SH(vec4, input + 32);
+  vec4 = in18 - in21;
+  ST_SH(vec4, input + 40);
+  vec4 = in29 - in26;
+  ST_SH(vec4, input + 80);
+  vec4 = in28 - in27;
+  ST_SH(vec4, input + 88);
+
+  in21 = in18 + in21;
+  in20 = in19 + in20;
+  in27 = in28 + in27;
+  in26 = in29 + in26;
+
+  LD_SH4(input + 48, 8, in22, in23, in24, in25);
+  DOTP_CONST_PAIR(in25, in22, cospi_16_64, cospi_16_64, in22, in25);
+  DOTP_CONST_PAIR(in24, in23, cospi_16_64, cospi_16_64, in23, in24);
+
+  in16 = LD_SH(input);
+  in17 = LD_SH(input + 8);
+  in30 = LD_SH(input + 112);
+  in31 = LD_SH(input + 120);
+
+  vec4 = in17 - in22;
+  ST_SH(vec4, input + 16);
+  vec4 = in16 - in23;
+  ST_SH(vec4, input + 24);
+  vec4 = in31 - in24;
+  ST_SH(vec4, input + 96);
+  vec4 = in30 - in25;
+  ST_SH(vec4, input + 104);
+
+  ADD4(in16, in23, in17, in22, in30, in25, in31, in24, in16, in17, in30, in31);
+  DOTP_CONST_PAIR(in26, in21, cospi_24_64, cospi_8_64, in18, in29);
+  DOTP_CONST_PAIR(in27, in20, cospi_24_64, cospi_8_64, in19, in28);
+  ADD4(in16, in19, in17, in18, in30, in29, in31, in28, in27, in22, in21, in25);
+  DOTP_CONST_PAIR(in21, in22, cospi_28_64, cospi_4_64, in26, in24);
+  ADD2(in27, in26, in25, in24, in23, in20);
+  DOTP_CONST_PAIR(in20, in23, cospi_31_64, cospi_1_64, vec4, vec5);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr);
+  ST_SH(vec4, temp_ptr + 960);
+
+  SUB2(in27, in26, in25, in24, in22, in21);
+  DOTP_CONST_PAIR(in21, in22, cospi_15_64, cospi_17_64, vec5, vec4);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr + 448);
+  ST_SH(vec4, temp_ptr + 512);
+
+  SUB4(in17, in18, in16, in19, in31, in28, in30, in29, in23, in26, in24, in20);
+  DOTP_CONST_PAIR((-in23), in20, cospi_28_64, cospi_4_64, in27, in25);
+  SUB2(in26, in27, in24, in25, in23, in20);
+  DOTP_CONST_PAIR(in20, in23, cospi_23_64, cospi_9_64, vec4, vec5);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec4, temp_ptr + 704);
+  ST_SH(vec5, temp_ptr + 256);
+
+  ADD2(in26, in27, in24, in25, in22, in21);
+  DOTP_CONST_PAIR(in21, in22, cospi_7_64, cospi_25_64, vec4, vec5);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec4, temp_ptr + 192);
+  ST_SH(vec5, temp_ptr + 768);
+
+  LD_SH4(input + 16, 8, in22, in23, in20, in21);
+  LD_SH4(input + 80, 8, in26, in27, in24, in25);
+  in16 = in20;
+  in17 = in21;
+  DOTP_CONST_PAIR(-in16, in27, cospi_24_64, cospi_8_64, in20, in27);
+  DOTP_CONST_PAIR(-in17, in26, cospi_24_64, cospi_8_64, in21, in26);
+  SUB4(in23, in20, in22, in21, in25, in26, in24, in27, in28, in17, in18, in31);
+  DOTP_CONST_PAIR(in18, in17, cospi_12_64, cospi_20_64, in29, in30);
+  ADD2(in28, in29, in31, in30, in16, in19);
+  DOTP_CONST_PAIR(in19, in16, cospi_27_64, cospi_5_64, vec5, vec4);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr + 832);
+  ST_SH(vec4, temp_ptr + 128);
+
+  SUB2(in28, in29, in31, in30, in17, in18);
+  DOTP_CONST_PAIR(in18, in17, cospi_11_64, cospi_21_64, vec5, vec4);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr + 320);
+  ST_SH(vec4, temp_ptr + 640);
+  ADD4(in22, in21, in23, in20, in24, in27, in25, in26, in16, in29, in30, in19);
+  DOTP_CONST_PAIR(-in16, in19, cospi_12_64, cospi_20_64, in28, in31);
+  SUB2(in29, in28, in30, in31, in16, in19);
+  DOTP_CONST_PAIR(in19, in16, cospi_19_64, cospi_13_64, vec5, vec4);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr + 576);
+  ST_SH(vec4, temp_ptr + 384);
+
+  ADD2(in29, in28, in30, in31, in17, in18);
+  DOTP_CONST_PAIR(in18, in17, cospi_3_64, cospi_29_64, vec5, vec4);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr + 64);
+  ST_SH(vec4, temp_ptr + 896);
+}
+
+static void fdct8x32_1d_column(const int16_t *input, int32_t src_stride,
+                               int16_t *tmp_buf, int16_t *tmp_buf_big) {
+  fdct8x32_1d_column_load_butterfly(input, src_stride, tmp_buf);
+  fdct8x32_1d_column_even_store(tmp_buf, tmp_buf_big);
+  fdct8x32_1d_column_odd_store(tmp_buf + 128, (tmp_buf_big + 32));
+}
+
+static void fdct8x32_1d_row_load_butterfly(int16_t *temp_buff,
+                                           int16_t *output) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 step0, step1, step2, step3, step4, step5, step6, step7;
+
+  LD_SH8(temp_buff, 32, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8(temp_buff + 24, 32, in8, in9, in10, in11, in12, in13, in14, in15);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15,
+                     in8, in9, in10, in11, in12, in13, in14, in15);
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+               in8, in9, in10, in11, in12, in13, in14, in15,
+               step0, step1, step2, step3, step4, step5, step6, step7,
+               in8, in9, in10, in11, in12, in13, in14, in15);
+  ST_SH8(step0, step1, step2, step3, step4, step5, step6, step7, output, 8);
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, (output + 24 * 8), 8);
+
+  /* 2nd set */
+  LD_SH8(temp_buff + 8, 32, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8(temp_buff + 16, 32, in8, in9, in10, in11, in12, in13, in14, in15);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15,
+                     in8, in9, in10, in11, in12, in13, in14, in15);
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+               in8, in9, in10, in11, in12, in13, in14, in15,
+               step0, step1, step2, step3, step4, step5, step6, step7,
+               in8, in9, in10, in11, in12, in13, in14, in15);
+  ST_SH8(step0, step1, step2, step3, step4, step5, step6, step7,
+         (output + 8 * 8), 8);
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, (output + 16 * 8), 8);
+}
+
+static void fdct8x32_1d_row_even_4x(int16_t *input, int16_t *interm_ptr,
+                                    int16_t *out) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v4i32 vec0_l, vec1_l, vec2_l, vec3_l, vec4_l, vec5_l, vec6_l, vec7_l;
+  v4i32 vec0_r, vec1_r, vec2_r, vec3_r, vec4_r, vec5_r, vec6_r, vec7_r;
+  v4i32 tmp0_w, tmp1_w, tmp2_w, tmp3_w;
+
+  /* fdct32 even */
+  /* stage 2 */
+  LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8(input + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+               in8, in9, in10, in11, in12, in13, in14, in15,
+               vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7,
+               in8, in9, in10, in11, in12, in13, in14, in15);
+  ST_SH8(vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, interm_ptr, 8);
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, interm_ptr + 64, 8);
+
+  /* Stage 3 */
+  UNPCK_SH_SW(vec0, vec0_l, vec0_r);
+  UNPCK_SH_SW(vec1, vec1_l, vec1_r);
+  UNPCK_SH_SW(vec2, vec2_l, vec2_r);
+  UNPCK_SH_SW(vec3, vec3_l, vec3_r);
+  UNPCK_SH_SW(vec4, vec4_l, vec4_r);
+  UNPCK_SH_SW(vec5, vec5_l, vec5_r);
+  UNPCK_SH_SW(vec6, vec6_l, vec6_r);
+  UNPCK_SH_SW(vec7, vec7_l, vec7_r);
+  ADD4(vec0_r, vec7_r, vec1_r, vec6_r, vec2_r, vec5_r, vec3_r, vec4_r,
+       tmp0_w, tmp1_w, tmp2_w, tmp3_w);
+  BUTTERFLY_4(tmp0_w, tmp1_w, tmp2_w, tmp3_w, vec4_r, vec6_r, vec7_r, vec5_r);
+  ADD4(vec0_l, vec7_l, vec1_l, vec6_l, vec2_l, vec5_l, vec3_l, vec4_l,
+       vec0_r, vec1_r, vec2_r, vec3_r);
+
+  tmp3_w = vec0_r + vec3_r;
+  vec0_r = vec0_r - vec3_r;
+  vec3_r = vec1_r + vec2_r;
+  vec1_r = vec1_r - vec2_r;
+
+  DOTP_CONST_PAIR_W(vec4_r, vec6_r, tmp3_w, vec3_r, cospi_16_64,
+                    cospi_16_64, vec4_r, tmp3_w, vec6_r, vec3_r);
+  FDCT32_POSTPROC_NEG_W(vec4_r);
+  FDCT32_POSTPROC_NEG_W(tmp3_w);
+  FDCT32_POSTPROC_NEG_W(vec6_r);
+  FDCT32_POSTPROC_NEG_W(vec3_r);
+  PCKEV_H2_SH(vec4_r, tmp3_w, vec6_r, vec3_r, vec4, vec5);
+  ST_SH2(vec5, vec4, out, 8);
+
+  DOTP_CONST_PAIR_W(vec5_r, vec7_r, vec0_r, vec1_r, cospi_24_64,
+                    cospi_8_64, vec4_r, tmp3_w, vec6_r, vec3_r);
+  FDCT32_POSTPROC_NEG_W(vec4_r);
+  FDCT32_POSTPROC_NEG_W(tmp3_w);
+  FDCT32_POSTPROC_NEG_W(vec6_r);
+  FDCT32_POSTPROC_NEG_W(vec3_r);
+  PCKEV_H2_SH(vec4_r, tmp3_w, vec6_r, vec3_r, vec4, vec5);
+  ST_SH2(vec5, vec4, out + 16, 8);
+
+  LD_SH8(interm_ptr, 8, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7);
+  SUB4(vec3, vec4, vec2, vec5, vec1, vec6, vec0, vec7, vec4, vec5, vec6, vec7);
+  DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+  ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, in5, in4);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 32);
+  ST_SH(in5, out + 56);
+
+  SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+  DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, in5, in4);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 40);
+  ST_SH(in5, out + 48);
+
+  LD_SH8(interm_ptr + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+  DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+  DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+  ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+  DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+  ADD2(in0, in1, in2, in3, vec0, vec7);
+  DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, in5, in4);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 64);
+  ST_SH(in5, out + 120);
+
+  SUB2(in0, in1, in2, in3, in0, in2);
+  DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, in5, in4);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 72);
+  ST_SH(in5, out + 112);
+
+  SUB2(in9, vec2, in14, vec5, vec2, vec5);
+  DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+  SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5);
+  DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, in5, in4);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 80);
+  ST_SH(in5, out + 104);
+
+  ADD2(in3, in2, in0, in1, vec3, vec4);
+  DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, in4, in5);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 96);
+  ST_SH(in5, out + 88);
+}
+
+static void fdct8x32_1d_row_even(int16_t *temp, int16_t *out) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, temp0, temp1;
+
+  /* fdct32 even */
+  /* stage 2 */
+  LD_SH8(temp, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8(temp + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+               in8, in9, in10, in11, in12, in13, in14, in15,
+               vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7,
+               in8, in9, in10, in11, in12, in13, in14, in15);
+
+  /* Stage 3 */
+  ADD4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, in0, in1, in2, in3);
+  BUTTERFLY_4(in0, in1, in2, in3, temp0, in4, in1, in0);
+  DOTP_CONST_PAIR(temp0, in4, cospi_16_64, cospi_16_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out);
+  ST_SH(temp1, out + 8);
+
+  DOTP_CONST_PAIR(in0, in1, cospi_24_64, cospi_8_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 16);
+  ST_SH(temp1, out + 24);
+
+  SUB4(vec3, vec4, vec2, vec5, vec1, vec6, vec0, vec7, vec4, vec5, vec6, vec7);
+  DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+  ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 32);
+  ST_SH(temp1, out + 56);
+
+  SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+  DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 40);
+  ST_SH(temp1, out + 48);
+
+  DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+  DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+  ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+  DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+  ADD2(in0, in1, in2, in3, vec0, vec7);
+  DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 64);
+  ST_SH(temp1, out + 120);
+
+  SUB2(in0, in1, in2, in3, in0, in2);
+  DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 72);
+  ST_SH(temp1, out + 112);
+
+  SUB2(in9, vec2, in14, vec5, vec2, vec5);
+  DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+  SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5)
+  DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 80);
+  ST_SH(temp1, out + 104);
+
+  ADD2(in3, in2, in0, in1, vec3, vec4);
+  DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, temp0, temp1);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 96);
+  ST_SH(temp1, out + 88);
+}
+
+static void fdct8x32_1d_row_odd(int16_t *temp, int16_t *interm_ptr,
+                                int16_t *out) {
+  v8i16 in16, in17, in18, in19, in20, in21, in22, in23;
+  v8i16 in24, in25, in26, in27, in28, in29, in30, in31, vec4, vec5;
+
+  in20 = LD_SH(temp + 32);
+  in21 = LD_SH(temp + 40);
+  in26 = LD_SH(temp + 80);
+  in27 = LD_SH(temp + 88);
+
+  DOTP_CONST_PAIR(in27, in20, cospi_16_64, cospi_16_64, in20, in27);
+  DOTP_CONST_PAIR(in26, in21, cospi_16_64, cospi_16_64, in21, in26);
+
+  in18 = LD_SH(temp + 16);
+  in19 = LD_SH(temp + 24);
+  in28 = LD_SH(temp + 96);
+  in29 = LD_SH(temp + 104);
+
+  vec4 = in19 - in20;
+  ST_SH(vec4, interm_ptr + 32);
+  vec4 = in18 - in21;
+  ST_SH(vec4, interm_ptr + 88);
+  vec4 = in28 - in27;
+  ST_SH(vec4, interm_ptr + 56);
+  vec4 = in29 - in26;
+  ST_SH(vec4, interm_ptr + 64);
+
+  ADD4(in18, in21, in19, in20, in28, in27, in29, in26, in21, in20, in27, in26);
+
+  in22 = LD_SH(temp + 48);
+  in23 = LD_SH(temp + 56);
+  in24 = LD_SH(temp + 64);
+  in25 = LD_SH(temp + 72);
+
+  DOTP_CONST_PAIR(in25, in22, cospi_16_64, cospi_16_64, in22, in25);
+  DOTP_CONST_PAIR(in24, in23, cospi_16_64, cospi_16_64, in23, in24);
+
+  in16 = LD_SH(temp);
+  in17 = LD_SH(temp + 8);
+  in30 = LD_SH(temp + 112);
+  in31 = LD_SH(temp + 120);
+
+  vec4 = in17 - in22;
+  ST_SH(vec4, interm_ptr + 40);
+  vec4 = in30 - in25;
+  ST_SH(vec4, interm_ptr + 48);
+  vec4 = in31 - in24;
+  ST_SH(vec4, interm_ptr + 72);
+  vec4 = in16 - in23;
+  ST_SH(vec4, interm_ptr + 80);
+
+  ADD4(in16, in23, in17, in22, in30, in25, in31, in24, in16, in17, in30, in31);
+  DOTP_CONST_PAIR(in26, in21, cospi_24_64, cospi_8_64, in18, in29);
+  DOTP_CONST_PAIR(in27, in20, cospi_24_64, cospi_8_64, in19, in28);
+
+  ADD4(in16, in19, in17, in18, in30, in29, in31, in28, in27, in22, in21, in25);
+  DOTP_CONST_PAIR(in21, in22, cospi_28_64, cospi_4_64, in26, in24);
+  ADD2(in27, in26, in25, in24, in23, in20);
+
+  DOTP_CONST_PAIR(in20, in23, cospi_31_64, cospi_1_64, vec4, vec5);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec5, out);
+  ST_SH(vec4, out + 120);
+
+  SUB2(in27, in26, in25, in24, in22, in21);
+
+  DOTP_CONST_PAIR(in21, in22, cospi_15_64, cospi_17_64, vec5, vec4);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec5, out + 112);
+  ST_SH(vec4, out + 8);
+
+  SUB4(in17, in18, in16, in19, in31, in28, in30, in29, in23, in26, in24, in20);
+  DOTP_CONST_PAIR((-in23), in20, cospi_28_64, cospi_4_64, in27, in25);
+  SUB2(in26, in27, in24, in25, in23, in20);
+
+  DOTP_CONST_PAIR(in20, in23, cospi_23_64, cospi_9_64, vec4, vec5);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec4, out + 16);
+  ST_SH(vec5, out + 104);
+
+  ADD2(in26, in27, in24, in25, in22, in21);
+  DOTP_CONST_PAIR(in21, in22, cospi_7_64, cospi_25_64, vec4, vec5);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec4, out + 24);
+  ST_SH(vec5, out + 96);
+
+  in20 = LD_SH(interm_ptr + 32);
+  in21 = LD_SH(interm_ptr + 88);
+  in27 = LD_SH(interm_ptr + 56);
+  in26 = LD_SH(interm_ptr + 64);
+
+  in16 = in20;
+  in17 = in21;
+  DOTP_CONST_PAIR(-in16, in27, cospi_24_64, cospi_8_64, in20, in27);
+  DOTP_CONST_PAIR(-in17, in26, cospi_24_64, cospi_8_64, in21, in26);
+
+  in22 = LD_SH(interm_ptr + 40);
+  in25 = LD_SH(interm_ptr + 48);
+  in24 = LD_SH(interm_ptr + 72);
+  in23 = LD_SH(interm_ptr + 80);
+
+  SUB4(in23, in20, in22, in21, in25, in26, in24, in27, in28, in17, in18, in31);
+  DOTP_CONST_PAIR(in18, in17, cospi_12_64, cospi_20_64, in29, in30);
+  ADD2(in28, in29, in31, in30, in16, in19);
+  DOTP_CONST_PAIR(in19, in16, cospi_27_64, cospi_5_64, vec5, vec4);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec5, out + 32);
+  ST_SH(vec4, out + 88);
+
+  SUB2(in28, in29, in31, in30, in17, in18);
+  DOTP_CONST_PAIR(in18, in17, cospi_11_64, cospi_21_64, vec5, vec4);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec5, out + 40);
+  ST_SH(vec4, out + 80);
+
+  ADD4(in22, in21, in23, in20, in24, in27, in25, in26, in16, in29, in30, in19);
+  DOTP_CONST_PAIR(-in16, in19, cospi_12_64, cospi_20_64, in28, in31);
+  SUB2(in29, in28, in30, in31, in16, in19);
+
+  DOTP_CONST_PAIR(in19, in16, cospi_19_64, cospi_13_64, vec5, vec4);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec5, out + 72);
+  ST_SH(vec4, out + 48);
+
+  ADD2(in29, in28, in30, in31, in17, in18);
+
+  DOTP_CONST_PAIR(in18, in17, cospi_3_64, cospi_29_64, vec5, vec4);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec4, out + 56);
+  ST_SH(vec5, out + 64);
+}
+
+static void fdct8x32_1d_row_transpose_store(int16_t *temp, int16_t *output) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1;
+
+  /* 1st set */
+  in0 = LD_SH(temp);
+  in4 = LD_SH(temp + 32);
+  in2 = LD_SH(temp + 64);
+  in6 = LD_SH(temp + 96);
+  in1 = LD_SH(temp + 128);
+  in7 = LD_SH(temp + 152);
+  in3 = LD_SH(temp + 192);
+  in5 = LD_SH(temp + 216);
+
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+
+  /* 2nd set */
+  in0_1 = LD_SH(temp + 16);
+  in1_1 = LD_SH(temp + 232);
+  in2_1 = LD_SH(temp + 80);
+  in3_1 = LD_SH(temp + 168);
+  in4_1 = LD_SH(temp + 48);
+  in5_1 = LD_SH(temp + 176);
+  in6_1 = LD_SH(temp + 112);
+  in7_1 = LD_SH(temp + 240);
+
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output, 32);
+  TRANSPOSE8x8_SH_SH(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+                     in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1);
+
+  /* 3rd set */
+  in0 = LD_SH(temp + 8);
+  in1 = LD_SH(temp + 136);
+  in2 = LD_SH(temp + 72);
+  in3 = LD_SH(temp + 200);
+  in4 = LD_SH(temp + 40);
+  in5 = LD_SH(temp + 208);
+  in6 = LD_SH(temp + 104);
+  in7 = LD_SH(temp + 144);
+
+  ST_SH8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+         output + 8, 32);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output + 16, 32);
+
+  /* 4th set */
+  in0_1 = LD_SH(temp + 24);
+  in1_1 = LD_SH(temp + 224);
+  in2_1 = LD_SH(temp + 88);
+  in3_1 = LD_SH(temp + 160);
+  in4_1 = LD_SH(temp + 56);
+  in5_1 = LD_SH(temp + 184);
+  in6_1 = LD_SH(temp + 120);
+  in7_1 = LD_SH(temp + 248);
+
+  TRANSPOSE8x8_SH_SH(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+                     in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1);
+  ST_SH8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+         output + 24, 32);
+}
+
+static void fdct32x8_1d_row(int16_t *temp, int16_t *temp_buf,
+                            int16_t *output) {
+  fdct8x32_1d_row_load_butterfly(temp, temp_buf);
+  fdct8x32_1d_row_even(temp_buf, temp_buf);
+  fdct8x32_1d_row_odd(temp_buf + 128, temp, temp_buf + 128);
+  fdct8x32_1d_row_transpose_store(temp_buf, output);
+}
+
+static void fdct32x8_1d_row_4x(int16_t *tmp_buf_big, int16_t *tmp_buf,
+                               int16_t *output) {
+  fdct8x32_1d_row_load_butterfly(tmp_buf_big, tmp_buf);
+  fdct8x32_1d_row_even_4x(tmp_buf, tmp_buf_big, tmp_buf);
+  fdct8x32_1d_row_odd(tmp_buf + 128, tmp_buf_big, tmp_buf + 128);
+  fdct8x32_1d_row_transpose_store(tmp_buf, output);
+}
+
+void vpx_fdct32x32_msa(const int16_t *input, int16_t *output,
+                       int32_t src_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, tmp_buf_big[1024]);
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[256]);
+
+  /* column transform */
+  for (i = 0; i < 4; ++i) {
+    fdct8x32_1d_column(input + (8 * i), src_stride, tmp_buf,
+                       tmp_buf_big + (8 * i));
+  }
+
+  /* row transform */
+  fdct32x8_1d_row_4x(tmp_buf_big, tmp_buf, output);
+
+  /* row transform */
+  for (i = 1; i < 4; ++i) {
+    fdct32x8_1d_row(tmp_buf_big + (i * 256), tmp_buf, output + (i * 256));
+  }
+}
+
+static void fdct8x32_1d_row_even_rd(int16_t *temp, int16_t *out) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, temp0, temp1;
+
+  /* fdct32 even */
+  /* stage 2 */
+  LD_SH8(temp, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8(temp + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,
+               in8, in9, in10, in11, in12, in13, in14, in15,
+               vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7,
+               in8, in9, in10, in11, in12, in13, in14, in15);
+  FDCT_POSTPROC_2V_NEG_H(vec0, vec1);
+  FDCT_POSTPROC_2V_NEG_H(vec2, vec3);
+  FDCT_POSTPROC_2V_NEG_H(vec4, vec5);
+  FDCT_POSTPROC_2V_NEG_H(vec6, vec7);
+  FDCT_POSTPROC_2V_NEG_H(in8, in9);
+  FDCT_POSTPROC_2V_NEG_H(in10, in11);
+  FDCT_POSTPROC_2V_NEG_H(in12, in13);
+  FDCT_POSTPROC_2V_NEG_H(in14, in15);
+
+  /* Stage 3 */
+  ADD4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, in0, in1, in2, in3);
+
+  temp0 = in0 + in3;
+  in0 = in0 - in3;
+  in3 = in1 + in2;
+  in1 = in1 - in2;
+
+  DOTP_CONST_PAIR(temp0, in3, cospi_16_64, cospi_16_64, temp1, temp0);
+  ST_SH(temp0, out);
+  ST_SH(temp1, out + 8);
+
+  DOTP_CONST_PAIR(in0, in1, cospi_24_64, cospi_8_64, temp1, temp0);
+  ST_SH(temp0, out + 16);
+  ST_SH(temp1, out + 24);
+
+  SUB4(vec3, vec4, vec2, vec5, vec1, vec6, vec0, vec7, vec4, vec5, vec6, vec7);
+  DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+  ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, temp1, temp0);
+  ST_SH(temp0, out + 32);
+  ST_SH(temp1, out + 56);
+
+  SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+  DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, temp1, temp0);
+  ST_SH(temp0, out + 40);
+  ST_SH(temp1, out + 48);
+
+  DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+  DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+  ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+  DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+  ADD2(in0, in1, in2, in3, vec0, vec7);
+  DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, temp1, temp0);
+  ST_SH(temp0, out + 64);
+  ST_SH(temp1, out + 120);
+
+  SUB2(in0, in1, in2, in3, in0, in2);
+  DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, temp1, temp0);
+  ST_SH(temp0, out + 72);
+  ST_SH(temp1, out + 112);
+
+  SUB2(in9, vec2, in14, vec5, vec2, vec5);
+  DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+  SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5);
+  DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, temp1, temp0);
+  ST_SH(temp0, out + 80);
+  ST_SH(temp1, out + 104);
+
+  ADD2(in3, in2, in0, in1, vec3, vec4);
+  DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, temp0, temp1);
+  ST_SH(temp0, out + 96);
+  ST_SH(temp1, out + 88);
+}
+
+static void fdct8x32_1d_row_odd_rd(int16_t *temp, int16_t *interm_ptr,
+                                   int16_t *out) {
+  v8i16 in16, in17, in18, in19, in20, in21, in22, in23;
+  v8i16 in24, in25, in26, in27, in28, in29, in30, in31;
+  v8i16 vec4, vec5;
+
+  in20 = LD_SH(temp + 32);
+  in21 = LD_SH(temp + 40);
+  in26 = LD_SH(temp + 80);
+  in27 = LD_SH(temp + 88);
+
+  DOTP_CONST_PAIR(in27, in20, cospi_16_64, cospi_16_64, in20, in27);
+  DOTP_CONST_PAIR(in26, in21, cospi_16_64, cospi_16_64, in21, in26);
+
+  FDCT_POSTPROC_2V_NEG_H(in20, in21);
+  FDCT_POSTPROC_2V_NEG_H(in26, in27);
+
+  in18 = LD_SH(temp + 16);
+  in19 = LD_SH(temp + 24);
+  in28 = LD_SH(temp + 96);
+  in29 = LD_SH(temp + 104);
+
+  FDCT_POSTPROC_2V_NEG_H(in18, in19);
+  FDCT_POSTPROC_2V_NEG_H(in28, in29);
+
+  vec4 = in19 - in20;
+  ST_SH(vec4, interm_ptr + 32);
+  vec4 = in18 - in21;
+  ST_SH(vec4, interm_ptr + 88);
+  vec4 = in29 - in26;
+  ST_SH(vec4, interm_ptr + 64);
+  vec4 = in28 - in27;
+  ST_SH(vec4, interm_ptr + 56);
+
+  ADD4(in18, in21, in19, in20, in28, in27, in29, in26, in21, in20, in27, in26);
+
+  in22 = LD_SH(temp + 48);
+  in23 = LD_SH(temp + 56);
+  in24 = LD_SH(temp + 64);
+  in25 = LD_SH(temp + 72);
+
+  DOTP_CONST_PAIR(in25, in22, cospi_16_64, cospi_16_64, in22, in25);
+  DOTP_CONST_PAIR(in24, in23, cospi_16_64, cospi_16_64, in23, in24);
+  FDCT_POSTPROC_2V_NEG_H(in22, in23);
+  FDCT_POSTPROC_2V_NEG_H(in24, in25);
+
+  in16 = LD_SH(temp);
+  in17 = LD_SH(temp + 8);
+  in30 = LD_SH(temp + 112);
+  in31 = LD_SH(temp + 120);
+
+  FDCT_POSTPROC_2V_NEG_H(in16, in17);
+  FDCT_POSTPROC_2V_NEG_H(in30, in31);
+
+  vec4 = in17 - in22;
+  ST_SH(vec4, interm_ptr + 40);
+  vec4 = in30 - in25;
+  ST_SH(vec4, interm_ptr + 48);
+  vec4 = in31 - in24;
+  ST_SH(vec4, interm_ptr + 72);
+  vec4 = in16 - in23;
+  ST_SH(vec4, interm_ptr + 80);
+
+  ADD4(in16, in23, in17, in22, in30, in25, in31, in24, in16, in17, in30, in31);
+  DOTP_CONST_PAIR(in26, in21, cospi_24_64, cospi_8_64, in18, in29);
+  DOTP_CONST_PAIR(in27, in20, cospi_24_64, cospi_8_64, in19, in28);
+  ADD4(in16, in19, in17, in18, in30, in29, in31, in28, in27, in22, in21, in25);
+  DOTP_CONST_PAIR(in21, in22, cospi_28_64, cospi_4_64, in26, in24);
+  ADD2(in27, in26, in25, in24, in23, in20);
+  DOTP_CONST_PAIR(in20, in23, cospi_31_64, cospi_1_64, vec4, vec5);
+  ST_SH(vec5, out);
+  ST_SH(vec4, out + 120);
+
+  SUB2(in27, in26, in25, in24, in22, in21);
+  DOTP_CONST_PAIR(in21, in22, cospi_15_64, cospi_17_64, vec5, vec4);
+  ST_SH(vec5, out + 112);
+  ST_SH(vec4, out + 8);
+
+  SUB4(in17, in18, in16, in19, in31, in28, in30, in29, in23, in26, in24, in20);
+  DOTP_CONST_PAIR((-in23), in20, cospi_28_64, cospi_4_64, in27, in25);
+  SUB2(in26, in27, in24, in25, in23, in20);
+  DOTP_CONST_PAIR(in20, in23, cospi_23_64, cospi_9_64, vec4, vec5);
+  ST_SH(vec4, out + 16);
+  ST_SH(vec5, out + 104);
+
+  ADD2(in26, in27, in24, in25, in22, in21);
+  DOTP_CONST_PAIR(in21, in22, cospi_7_64, cospi_25_64, vec4, vec5);
+  ST_SH(vec4, out + 24);
+  ST_SH(vec5, out + 96);
+
+  in20 = LD_SH(interm_ptr + 32);
+  in21 = LD_SH(interm_ptr + 88);
+  in27 = LD_SH(interm_ptr + 56);
+  in26 = LD_SH(interm_ptr + 64);
+
+  in16 = in20;
+  in17 = in21;
+  DOTP_CONST_PAIR(-in16, in27, cospi_24_64, cospi_8_64, in20, in27);
+  DOTP_CONST_PAIR(-in17, in26, cospi_24_64, cospi_8_64, in21, in26);
+
+  in22 = LD_SH(interm_ptr + 40);
+  in25 = LD_SH(interm_ptr + 48);
+  in24 = LD_SH(interm_ptr + 72);
+  in23 = LD_SH(interm_ptr + 80);
+
+  SUB4(in23, in20, in22, in21, in25, in26, in24, in27, in28, in17, in18, in31);
+  DOTP_CONST_PAIR(in18, in17, cospi_12_64, cospi_20_64, in29, in30);
+  in16 = in28 + in29;
+  in19 = in31 + in30;
+  DOTP_CONST_PAIR(in19, in16, cospi_27_64, cospi_5_64, vec5, vec4);
+  ST_SH(vec5, out + 32);
+  ST_SH(vec4, out + 88);
+
+  SUB2(in28, in29, in31, in30, in17, in18);
+  DOTP_CONST_PAIR(in18, in17, cospi_11_64, cospi_21_64, vec5, vec4);
+  ST_SH(vec5, out + 40);
+  ST_SH(vec4, out + 80);
+
+  ADD4(in22, in21, in23, in20, in24, in27, in25, in26, in16, in29, in30, in19);
+  DOTP_CONST_PAIR(-in16, in19, cospi_12_64, cospi_20_64, in28, in31);
+  SUB2(in29, in28, in30, in31, in16, in19);
+  DOTP_CONST_PAIR(in19, in16, cospi_19_64, cospi_13_64, vec5, vec4);
+  ST_SH(vec5, out + 72);
+  ST_SH(vec4, out + 48);
+
+  ADD2(in29, in28, in30, in31, in17, in18);
+  DOTP_CONST_PAIR(in18, in17, cospi_3_64, cospi_29_64, vec5, vec4);
+  ST_SH(vec4, out + 56);
+  ST_SH(vec5, out + 64);
+}
+
+static void fdct32x8_1d_row_rd(int16_t *tmp_buf_big, int16_t *tmp_buf,
+                               int16_t *output) {
+  fdct8x32_1d_row_load_butterfly(tmp_buf_big, tmp_buf);
+  fdct8x32_1d_row_even_rd(tmp_buf, tmp_buf);
+  fdct8x32_1d_row_odd_rd((tmp_buf + 128), tmp_buf_big, (tmp_buf + 128));
+  fdct8x32_1d_row_transpose_store(tmp_buf, output);
+}
+
+void vpx_fdct32x32_rd_msa(const int16_t *input, int16_t *out,
+                          int32_t src_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, tmp_buf_big[1024]);
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[256]);
+
+  /* column transform */
+  for (i = 0; i < 4; ++i) {
+    fdct8x32_1d_column(input + (8 * i), src_stride, &tmp_buf[0],
+                       &tmp_buf_big[0] + (8 * i));
+  }
+
+  /* row transform */
+  for (i = 0; i < 4; ++i) {
+    fdct32x8_1d_row_rd(&tmp_buf_big[0] + (8 * i * 32), &tmp_buf[0],
+                       out + (8 * i * 32));
+  }
+}
+
+void vpx_fdct32x32_1_msa(const int16_t *input, int16_t *out, int32_t stride) {
+  int sum = LD_HADD(input, stride);
+  sum += LD_HADD(input + 8, stride);
+  sum += LD_HADD(input + 16, stride);
+  sum += LD_HADD(input + 24, stride);
+  sum += LD_HADD(input + 32 * 8, stride);
+  sum += LD_HADD(input + 32 * 8 + 8, stride);
+  sum += LD_HADD(input + 32 * 8 + 16, stride);
+  sum += LD_HADD(input + 32 * 8 + 24, stride);
+  sum += LD_HADD(input + 32 * 16, stride);
+  sum += LD_HADD(input + 32 * 16 + 8, stride);
+  sum += LD_HADD(input + 32 * 16 + 16, stride);
+  sum += LD_HADD(input + 32 * 16 + 24, stride);
+  sum += LD_HADD(input + 32 * 24, stride);
+  sum += LD_HADD(input + 32 * 24 + 8, stride);
+  sum += LD_HADD(input + 32 * 24 + 16, stride);
+  sum += LD_HADD(input + 32 * 24 + 24, stride);
+  out[0] = (int16_t)(sum >> 3);
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/fwd_txfm_msa.c b/libvpx/libvpx/vpx_dsp/mips/fwd_txfm_msa.c
new file mode 100644
index 0000000..0dd141f
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/fwd_txfm_msa.c
@@ -0,0 +1,245 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/fwd_txfm_msa.h"
+
+void fdct8x16_1d_column(const int16_t *input, int16_t *tmp_ptr,
+                        int32_t src_stride) {
+  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 stp21, stp22, stp23, stp24, stp25, stp26, stp30;
+  v8i16 stp31, stp32, stp33, stp34, stp35, stp36, stp37;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, cnst0, cnst1, cnst4, cnst5;
+  v8i16 coeff = { cospi_16_64, -cospi_16_64, cospi_8_64, cospi_24_64,
+                 -cospi_8_64, -cospi_24_64, cospi_12_64, cospi_20_64 };
+  v8i16 coeff1 = { cospi_2_64, cospi_30_64, cospi_14_64, cospi_18_64,
+                   cospi_10_64, cospi_22_64, cospi_6_64, cospi_26_64 };
+  v8i16 coeff2 = { -cospi_2_64, -cospi_10_64, -cospi_18_64, -cospi_26_64,
+                   0, 0, 0, 0 };
+
+  LD_SH16(input, src_stride,
+          in0, in1, in2, in3, in4, in5, in6, in7,
+          in8, in9, in10, in11, in12, in13, in14, in15);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+  SLLI_4V(in8, in9, in10, in11, 2);
+  SLLI_4V(in12, in13, in14, in15, 2);
+  ADD4(in0, in15, in1, in14, in2, in13, in3, in12, tmp0, tmp1, tmp2, tmp3);
+  ADD4(in4, in11, in5, in10, in6, in9, in7, in8, tmp4, tmp5, tmp6, tmp7);
+  FDCT8x16_EVEN(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7,
+                tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7);
+  ST_SH8(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp_ptr, 32);
+  SUB4(in0, in15, in1, in14, in2, in13, in3, in12, in15, in14, in13, in12);
+  SUB4(in4, in11, in5, in10, in6, in9, in7, in8, in11, in10, in9, in8);
+
+  tmp_ptr += 16;
+
+  /* stp 1 */
+  ILVL_H2_SH(in10, in13, in11, in12, vec2, vec4);
+  ILVR_H2_SH(in10, in13, in11, in12, vec3, vec5);
+
+  cnst4 = __msa_splati_h(coeff, 0);
+  stp25 = DOT_SHIFT_RIGHT_PCK_H(vec2, vec3, cnst4);
+
+  cnst5 = __msa_splati_h(coeff, 1);
+  cnst5 = __msa_ilvev_h(cnst5, cnst4);
+  stp22 = DOT_SHIFT_RIGHT_PCK_H(vec2, vec3, cnst5);
+  stp24 = DOT_SHIFT_RIGHT_PCK_H(vec4, vec5, cnst4);
+  stp23 = DOT_SHIFT_RIGHT_PCK_H(vec4, vec5, cnst5);
+
+  /* stp2 */
+  BUTTERFLY_4(in8, in9, stp22, stp23, stp30, stp31, stp32, stp33);
+  BUTTERFLY_4(in15, in14, stp25, stp24, stp37, stp36, stp35, stp34);
+  ILVL_H2_SH(stp36, stp31, stp35, stp32, vec2, vec4);
+  ILVR_H2_SH(stp36, stp31, stp35, stp32, vec3, vec5);
+  SPLATI_H2_SH(coeff, 2, 3, cnst0, cnst1);
+  cnst0 = __msa_ilvev_h(cnst0, cnst1);
+  stp26 = DOT_SHIFT_RIGHT_PCK_H(vec2, vec3, cnst0);
+
+  cnst0 = __msa_splati_h(coeff, 4);
+  cnst1 = __msa_ilvev_h(cnst1, cnst0);
+  stp21 = DOT_SHIFT_RIGHT_PCK_H(vec2, vec3, cnst1);
+
+  BUTTERFLY_4(stp30, stp37, stp26, stp21, in8, in15, in14, in9);
+  ILVRL_H2_SH(in15, in8, vec1, vec0);
+  SPLATI_H2_SH(coeff1, 0, 1, cnst0, cnst1);
+  cnst0 = __msa_ilvev_h(cnst0, cnst1);
+
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr);
+
+  cnst0 = __msa_splati_h(coeff2, 0);
+  cnst0 = __msa_ilvev_h(cnst1, cnst0);
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr + 224);
+
+  ILVRL_H2_SH(in14, in9, vec1, vec0);
+  SPLATI_H2_SH(coeff1, 2, 3, cnst0, cnst1);
+  cnst1 = __msa_ilvev_h(cnst1, cnst0);
+
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst1);
+  ST_SH(in8, tmp_ptr + 128);
+
+  cnst1 = __msa_splati_h(coeff2, 2);
+  cnst0 = __msa_ilvev_h(cnst0, cnst1);
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr + 96);
+
+  SPLATI_H2_SH(coeff, 2, 5, cnst0, cnst1);
+  cnst1 = __msa_ilvev_h(cnst1, cnst0);
+
+  stp25 = DOT_SHIFT_RIGHT_PCK_H(vec4, vec5, cnst1);
+
+  cnst1 = __msa_splati_h(coeff, 3);
+  cnst1 = __msa_ilvev_h(cnst0, cnst1);
+  stp22 = DOT_SHIFT_RIGHT_PCK_H(vec4, vec5, cnst1);
+
+  /* stp4 */
+  ADD2(stp34, stp25, stp33, stp22, in13, in10);
+
+  ILVRL_H2_SH(in13, in10, vec1, vec0);
+  SPLATI_H2_SH(coeff1, 4, 5, cnst0, cnst1);
+  cnst0 = __msa_ilvev_h(cnst0, cnst1);
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr + 64);
+
+  cnst0 = __msa_splati_h(coeff2, 1);
+  cnst0 = __msa_ilvev_h(cnst1, cnst0);
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr + 160);
+
+  SUB2(stp34, stp25, stp33, stp22, in12, in11);
+  ILVRL_H2_SH(in12, in11, vec1, vec0);
+  SPLATI_H2_SH(coeff1, 6, 7, cnst0, cnst1);
+  cnst1 = __msa_ilvev_h(cnst1, cnst0);
+
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst1);
+  ST_SH(in8, tmp_ptr + 192);
+
+  cnst1 = __msa_splati_h(coeff2, 3);
+  cnst0 = __msa_ilvev_h(cnst0, cnst1);
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr + 32);
+}
+
+void fdct16x8_1d_row(int16_t *input, int16_t *output) {
+  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+
+  LD_SH8(input, 16, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8((input + 8), 16, in8, in9, in10, in11, in12, in13, in14, in15);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15,
+                     in8, in9, in10, in11, in12, in13, in14, in15);
+  ADD4(in0, 1, in1, 1, in2, 1, in3, 1, in0, in1, in2, in3);
+  ADD4(in4, 1, in5, 1, in6, 1, in7, 1, in4, in5, in6, in7);
+  ADD4(in8, 1, in9, 1, in10, 1, in11, 1, in8, in9, in10, in11);
+  ADD4(in12, 1, in13, 1, in14, 1, in15, 1, in12, in13, in14, in15);
+  SRA_4V(in0, in1, in2, in3, 2);
+  SRA_4V(in4, in5, in6, in7, 2);
+  SRA_4V(in8, in9, in10, in11, 2);
+  SRA_4V(in12, in13, in14, in15, 2);
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,
+               in12, in13, in14, in15, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5,
+               tmp6, tmp7, in8, in9, in10, in11, in12, in13, in14, in15);
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, input, 16);
+  FDCT8x16_EVEN(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7,
+                tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7);
+  LD_SH8(input, 16, in8, in9, in10, in11, in12, in13, in14, in15);
+  FDCT8x16_ODD(in8, in9, in10, in11, in12, in13, in14, in15,
+                   in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3,
+                     tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3);
+  ST_SH8(tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3, output, 16);
+  TRANSPOSE8x8_SH_SH(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7,
+                     tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7);
+  ST_SH8(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7, output + 8, 16);
+}
+
+void vpx_fdct4x4_msa(const int16_t *input, int16_t *output,
+                     int32_t src_stride) {
+  v8i16 in0, in1, in2, in3;
+
+  LD_SH4(input, src_stride, in0, in1, in2, in3);
+
+  /* fdct4 pre-process */
+  {
+    v8i16 vec, mask;
+    v16i8 zero = { 0 };
+    v16i8 one = __msa_ldi_b(1);
+
+    mask = (v8i16)__msa_sldi_b(zero, one, 15);
+    SLLI_4V(in0, in1, in2, in3, 4);
+    vec = __msa_ceqi_h(in0, 0);
+    vec = vec ^ 255;
+    vec = mask & vec;
+    in0 += vec;
+  }
+
+  VP9_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  VP9_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  ADD4(in0, 1, in1, 1, in2, 1, in3, 1, in0, in1, in2, in3);
+  SRA_4V(in0, in1, in2, in3, 2);
+  PCKEV_D2_SH(in1, in0, in3, in2, in0, in2);
+  ST_SH2(in0, in2, output, 8);
+}
+
+void vpx_fdct8x8_msa(const int16_t *input, int16_t *output,
+                     int32_t src_stride) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  LD_SH8(input, src_stride, in0, in1, in2, in3, in4, in5, in6, in7);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+  VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+            in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+            in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7);
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output, 8);
+}
+
+void vpx_fdct8x8_1_msa(const int16_t *input, int16_t *out, int32_t stride) {
+  out[0] = LD_HADD(input, stride);
+  out[1] = 0;
+}
+
+void vpx_fdct16x16_msa(const int16_t *input, int16_t *output,
+                       int32_t src_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[16 * 16]);
+
+  /* column transform */
+  for (i = 0; i < 2; ++i) {
+    fdct8x16_1d_column((input + 8 * i), (&tmp_buf[0] + 8 * i), src_stride);
+  }
+
+  /* row transform */
+  for (i = 0; i < 2; ++i) {
+    fdct16x8_1d_row((&tmp_buf[0] + (128 * i)), (output + (128 * i)));
+  }
+}
+
+void vpx_fdct16x16_1_msa(const int16_t *input, int16_t *out, int32_t stride) {
+  int sum = LD_HADD(input, stride);
+  sum += LD_HADD(input + 8, stride);
+  sum += LD_HADD(input + 16 * 8, stride);
+  sum += LD_HADD(input + 16 * 8 + 8, stride);
+  out[0] = (int16_t)(sum >> 1);
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/fwd_txfm_msa.h b/libvpx/libvpx/vpx_dsp/mips/fwd_txfm_msa.h
new file mode 100644
index 0000000..d1e160e
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/fwd_txfm_msa.h
@@ -0,0 +1,373 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_MIPS_FWD_TXFM_MSA_H_
+#define VPX_DSP_MIPS_FWD_TXFM_MSA_H_
+
+#include "vpx_dsp/mips/txfm_macros_msa.h"
+#include "vpx_dsp/txfm_common.h"
+
+#define LD_HADD(psrc, stride) ({                                      \
+  v8i16 in0_m, in1_m, in2_m, in3_m, in4_m, in5_m, in6_m, in7_m;       \
+  v4i32 vec_w_m;                                                      \
+                                                                      \
+  LD_SH4((psrc), stride, in0_m, in1_m, in2_m, in3_m);                 \
+  ADD2(in0_m, in1_m, in2_m, in3_m, in0_m, in2_m);                     \
+  LD_SH4(((psrc) + 4 * stride), stride, in4_m, in5_m, in6_m, in7_m);  \
+  ADD4(in4_m, in5_m, in6_m, in7_m, in0_m, in2_m, in4_m, in6_m,        \
+       in4_m, in6_m, in0_m, in4_m);                                   \
+  in0_m += in4_m;                                                     \
+                                                                      \
+  vec_w_m = __msa_hadd_s_w(in0_m, in0_m);                             \
+  HADD_SW_S32(vec_w_m);                                               \
+})
+
+#define VP9_FDCT4(in0, in1, in2, in3, out0, out1, out2, out3) {     \
+  v8i16 cnst0_m, cnst1_m, cnst2_m, cnst3_m;                         \
+  v8i16 vec0_m, vec1_m, vec2_m, vec3_m;                             \
+  v4i32 vec4_m, vec5_m, vec6_m, vec7_m;                             \
+  v8i16 coeff_m = { cospi_16_64, -cospi_16_64, cospi_8_64,          \
+                    cospi_24_64, -cospi_8_64, 0, 0, 0 };            \
+                                                                    \
+  BUTTERFLY_4(in0, in1, in2, in3, vec0_m, vec1_m, vec2_m, vec3_m);  \
+  ILVR_H2_SH(vec1_m, vec0_m, vec3_m, vec2_m, vec0_m, vec2_m);       \
+  SPLATI_H2_SH(coeff_m, 0, 1, cnst0_m, cnst1_m);                    \
+  cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                        \
+  vec5_m = __msa_dotp_s_w(vec0_m, cnst1_m);                         \
+                                                                    \
+  SPLATI_H2_SH(coeff_m, 4, 3, cnst2_m, cnst3_m);                    \
+  cnst2_m = __msa_ilvev_h(cnst3_m, cnst2_m);                        \
+  vec7_m = __msa_dotp_s_w(vec2_m, cnst2_m);                         \
+                                                                    \
+  vec4_m = __msa_dotp_s_w(vec0_m, cnst0_m);                         \
+  cnst2_m = __msa_splati_h(coeff_m, 2);                             \
+  cnst2_m = __msa_ilvev_h(cnst2_m, cnst3_m);                        \
+  vec6_m = __msa_dotp_s_w(vec2_m, cnst2_m);                         \
+                                                                    \
+  SRARI_W4_SW(vec4_m, vec5_m, vec6_m, vec7_m, DCT_CONST_BITS);      \
+  PCKEV_H4_SH(vec4_m, vec4_m, vec5_m, vec5_m, vec6_m, vec6_m,       \
+              vec7_m, vec7_m, out0, out2, out1, out3);              \
+}
+
+#define SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7) {        \
+  v8i16 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;  \
+                                                                         \
+  SRLI_H4_SH(in0, in1, in2, in3, vec0_m, vec1_m, vec2_m, vec3_m, 15);    \
+  SRLI_H4_SH(in4, in5, in6, in7, vec4_m, vec5_m, vec6_m, vec7_m, 15);    \
+  AVE_SH4_SH(vec0_m, in0, vec1_m, in1, vec2_m, in2, vec3_m, in3,         \
+             in0, in1, in2, in3);                                        \
+  AVE_SH4_SH(vec4_m, in4, vec5_m, in5, vec6_m, in6, vec7_m, in7,         \
+             in4, in5, in6, in7);                                        \
+}
+
+#define VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,            \
+                  out0, out1, out2, out3, out4, out5, out6, out7) {  \
+  v8i16 s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m;                    \
+  v8i16 s7_m, x0_m, x1_m, x2_m, x3_m;                                \
+  v8i16 coeff_m = { cospi_16_64, -cospi_16_64, cospi_8_64,           \
+                    cospi_24_64, cospi_4_64, cospi_28_64,            \
+                    cospi_12_64, cospi_20_64 };                      \
+                                                                     \
+  /* FDCT stage1 */                                                  \
+  BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7,                \
+              s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m, s7_m);       \
+  BUTTERFLY_4(s0_m, s1_m, s2_m, s3_m, x0_m, x1_m, x2_m, x3_m);       \
+  ILVL_H2_SH(x1_m, x0_m, x3_m, x2_m, s0_m, s2_m);                    \
+  ILVR_H2_SH(x1_m, x0_m, x3_m, x2_m, s1_m, s3_m);                    \
+  SPLATI_H2_SH(coeff_m, 0, 1, x0_m, x1_m);                           \
+  x1_m = __msa_ilvev_h(x1_m, x0_m);                                  \
+  out4 = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x1_m);                    \
+                                                                     \
+  SPLATI_H2_SH(coeff_m, 2, 3, x2_m, x3_m);                           \
+  x2_m = -x2_m;                                                      \
+  x2_m = __msa_ilvev_h(x3_m, x2_m);                                  \
+  out6 = DOT_SHIFT_RIGHT_PCK_H(s2_m, s3_m, x2_m);                    \
+                                                                     \
+  out0 = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x0_m);                    \
+  x2_m = __msa_splati_h(coeff_m, 2);                                 \
+  x2_m = __msa_ilvev_h(x2_m, x3_m);                                  \
+  out2 = DOT_SHIFT_RIGHT_PCK_H(s2_m, s3_m, x2_m);                    \
+                                                                     \
+  /* stage2 */                                                       \
+  ILVRL_H2_SH(s5_m, s6_m, s1_m, s0_m);                               \
+                                                                     \
+  s6_m = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x0_m);                    \
+  s5_m = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x1_m);                    \
+                                                                     \
+  /* stage3 */                                                       \
+  BUTTERFLY_4(s4_m, s7_m, s6_m, s5_m, x0_m, x3_m, x2_m, x1_m);       \
+                                                                     \
+  /* stage4 */                                                       \
+  ILVL_H2_SH(x3_m, x0_m, x2_m, x1_m, s4_m, s6_m);                    \
+  ILVR_H2_SH(x3_m, x0_m, x2_m, x1_m, s5_m, s7_m);                    \
+                                                                     \
+  SPLATI_H2_SH(coeff_m, 4, 5, x0_m, x1_m);                           \
+  x1_m = __msa_ilvev_h(x0_m, x1_m);                                  \
+  out1 = DOT_SHIFT_RIGHT_PCK_H(s4_m, s5_m, x1_m);                    \
+                                                                     \
+  SPLATI_H2_SH(coeff_m, 6, 7, x2_m, x3_m);                           \
+  x2_m = __msa_ilvev_h(x3_m, x2_m);                                  \
+  out5 = DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m);                    \
+                                                                     \
+  x1_m = __msa_splati_h(coeff_m, 5);                                 \
+  x0_m = -x0_m;                                                      \
+  x0_m = __msa_ilvev_h(x1_m, x0_m);                                  \
+  out7 = DOT_SHIFT_RIGHT_PCK_H(s4_m, s5_m, x0_m);                    \
+                                                                     \
+  x2_m = __msa_splati_h(coeff_m, 6);                                 \
+  x3_m = -x3_m;                                                      \
+  x2_m = __msa_ilvev_h(x2_m, x3_m);                                  \
+  out3 = DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m);                    \
+}
+
+#define FDCT8x16_EVEN(in0, in1, in2, in3, in4, in5, in6, in7,                \
+                      out0, out1, out2, out3, out4, out5, out6, out7) {      \
+  v8i16 s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m, s7_m;                      \
+  v8i16 x0_m, x1_m, x2_m, x3_m;                                              \
+  v8i16 coeff_m = { cospi_16_64, -cospi_16_64, cospi_8_64, cospi_24_64,      \
+                    cospi_4_64, cospi_28_64, cospi_12_64, cospi_20_64 };     \
+                                                                             \
+  /* FDCT stage1 */                                                          \
+  BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7,                        \
+              s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m, s7_m);               \
+  BUTTERFLY_4(s0_m, s1_m, s2_m, s3_m, x0_m, x1_m, x2_m, x3_m);               \
+  ILVL_H2_SH(x1_m, x0_m, x3_m, x2_m, s0_m, s2_m);                            \
+  ILVR_H2_SH(x1_m, x0_m, x3_m, x2_m, s1_m, s3_m);                            \
+  SPLATI_H2_SH(coeff_m, 0, 1, x0_m, x1_m);                                   \
+  x1_m = __msa_ilvev_h(x1_m, x0_m);                                          \
+  out4 = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x1_m);                            \
+                                                                             \
+  SPLATI_H2_SH(coeff_m, 2, 3, x2_m, x3_m);                                   \
+  x2_m = -x2_m;                                                              \
+  x2_m = __msa_ilvev_h(x3_m, x2_m);                                          \
+  out6 = DOT_SHIFT_RIGHT_PCK_H(s2_m, s3_m, x2_m);                            \
+                                                                             \
+  out0 = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x0_m);                            \
+  x2_m = __msa_splati_h(coeff_m, 2);                                         \
+  x2_m = __msa_ilvev_h(x2_m, x3_m);                                          \
+  out2 = DOT_SHIFT_RIGHT_PCK_H(s2_m, s3_m, x2_m);                            \
+                                                                             \
+  /* stage2 */                                                               \
+  ILVRL_H2_SH(s5_m, s6_m, s1_m, s0_m);                                       \
+                                                                             \
+  s6_m = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x0_m);                            \
+  s5_m = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x1_m);                            \
+                                                                             \
+  /* stage3 */                                                               \
+  BUTTERFLY_4(s4_m, s7_m, s6_m, s5_m, x0_m, x3_m, x2_m, x1_m);               \
+                                                                             \
+  /* stage4 */                                                               \
+  ILVL_H2_SH(x3_m, x0_m, x2_m, x1_m, s4_m, s6_m);                            \
+  ILVR_H2_SH(x3_m, x0_m, x2_m, x1_m, s5_m, s7_m);                            \
+                                                                             \
+  SPLATI_H2_SH(coeff_m, 4, 5, x0_m, x1_m);                                   \
+  x1_m = __msa_ilvev_h(x0_m, x1_m);                                          \
+  out1 = DOT_SHIFT_RIGHT_PCK_H(s4_m, s5_m, x1_m);                            \
+                                                                             \
+  SPLATI_H2_SH(coeff_m, 6, 7, x2_m, x3_m);                                   \
+  x2_m = __msa_ilvev_h(x3_m, x2_m);                                          \
+  out5 = DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m);                            \
+                                                                             \
+  x1_m = __msa_splati_h(coeff_m, 5);                                         \
+  x0_m = -x0_m;                                                              \
+  x0_m = __msa_ilvev_h(x1_m, x0_m);                                          \
+  out7 = DOT_SHIFT_RIGHT_PCK_H(s4_m, s5_m, x0_m);                            \
+                                                                             \
+  x2_m = __msa_splati_h(coeff_m, 6);                                         \
+  x3_m = -x3_m;                                                              \
+  x2_m = __msa_ilvev_h(x2_m, x3_m);                                          \
+  out3 = DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m);                            \
+}
+
+#define FDCT8x16_ODD(input0, input1, input2, input3,               \
+                     input4, input5, input6, input7,               \
+                     out1, out3, out5, out7,                       \
+                     out9, out11, out13, out15) {                  \
+  v8i16 stp21_m, stp22_m, stp23_m, stp24_m, stp25_m, stp26_m;      \
+  v8i16 stp30_m, stp31_m, stp32_m, stp33_m, stp34_m, stp35_m;      \
+  v8i16 stp36_m, stp37_m, vec0_m, vec1_m;                          \
+  v8i16 vec2_m, vec3_m, vec4_m, vec5_m, vec6_m;                    \
+  v8i16 cnst0_m, cnst1_m, cnst4_m, cnst5_m;                        \
+  v8i16 coeff_m = { cospi_16_64, -cospi_16_64, cospi_8_64,         \
+                    cospi_24_64, -cospi_8_64, -cospi_24_64,        \
+                    cospi_12_64, cospi_20_64 };                    \
+  v8i16 coeff1_m = { cospi_2_64, cospi_30_64, cospi_14_64,         \
+                     cospi_18_64, cospi_10_64, cospi_22_64,        \
+                     cospi_6_64, cospi_26_64 };                    \
+  v8i16 coeff2_m = { -cospi_2_64, -cospi_10_64, -cospi_18_64,      \
+                     -cospi_26_64, 0, 0, 0, 0 };                   \
+                                                                   \
+  /* stp 1 */                                                      \
+  ILVL_H2_SH(input2, input5, input3, input4, vec2_m, vec4_m);      \
+  ILVR_H2_SH(input2, input5, input3, input4, vec3_m, vec5_m);      \
+                                                                   \
+  cnst4_m = __msa_splati_h(coeff_m, 0);                            \
+  stp25_m = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst4_m);        \
+                                                                   \
+  cnst5_m = __msa_splati_h(coeff_m, 1);                            \
+  cnst5_m = __msa_ilvev_h(cnst5_m, cnst4_m);                       \
+  stp22_m = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst5_m);        \
+  stp24_m = DOT_SHIFT_RIGHT_PCK_H(vec4_m, vec5_m, cnst4_m);        \
+  stp23_m = DOT_SHIFT_RIGHT_PCK_H(vec4_m, vec5_m, cnst5_m);        \
+                                                                   \
+  /* stp2 */                                                       \
+  BUTTERFLY_4(input0, input1, stp22_m, stp23_m,                    \
+              stp30_m, stp31_m, stp32_m, stp33_m);                 \
+  BUTTERFLY_4(input7, input6, stp25_m, stp24_m,                    \
+              stp37_m, stp36_m, stp35_m, stp34_m);                 \
+                                                                   \
+  ILVL_H2_SH(stp36_m, stp31_m, stp35_m, stp32_m, vec2_m, vec4_m);  \
+  ILVR_H2_SH(stp36_m, stp31_m, stp35_m, stp32_m, vec3_m, vec5_m);  \
+                                                                   \
+  SPLATI_H2_SH(coeff_m, 2, 3, cnst0_m, cnst1_m);                   \
+  cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m);                       \
+  stp26_m = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst0_m);        \
+                                                                   \
+  cnst0_m = __msa_splati_h(coeff_m, 4);                            \
+  cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                       \
+  stp21_m = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst1_m);        \
+                                                                   \
+  SPLATI_H2_SH(coeff_m, 5, 2, cnst0_m, cnst1_m);                   \
+  cnst1_m = __msa_ilvev_h(cnst0_m, cnst1_m);                       \
+  stp25_m = DOT_SHIFT_RIGHT_PCK_H(vec4_m, vec5_m, cnst1_m);        \
+                                                                   \
+  cnst0_m = __msa_splati_h(coeff_m, 3);                            \
+  cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                       \
+  stp22_m = DOT_SHIFT_RIGHT_PCK_H(vec4_m, vec5_m, cnst1_m);        \
+                                                                   \
+  /* stp4 */                                                       \
+  BUTTERFLY_4(stp30_m, stp37_m, stp26_m, stp21_m,                  \
+              vec6_m, vec2_m, vec4_m, vec5_m);                     \
+  BUTTERFLY_4(stp33_m, stp34_m, stp25_m, stp22_m,                  \
+              stp21_m, stp23_m, stp24_m, stp31_m);                 \
+                                                                   \
+  ILVRL_H2_SH(vec2_m, vec6_m, vec1_m, vec0_m);                     \
+  SPLATI_H2_SH(coeff1_m, 0, 1, cnst0_m, cnst1_m);                  \
+  cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m);                       \
+                                                                   \
+  out1 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);           \
+                                                                   \
+  cnst0_m = __msa_splati_h(coeff2_m, 0);                           \
+  cnst0_m = __msa_ilvev_h(cnst1_m, cnst0_m);                       \
+  out15 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);          \
+                                                                   \
+  ILVRL_H2_SH(vec4_m, vec5_m, vec1_m, vec0_m);                     \
+  SPLATI_H2_SH(coeff1_m, 2, 3, cnst0_m, cnst1_m);                  \
+  cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                       \
+                                                                   \
+  out9 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst1_m);           \
+                                                                   \
+  cnst1_m = __msa_splati_h(coeff2_m, 2);                           \
+  cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m);                       \
+  out7 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);           \
+                                                                   \
+  ILVRL_H2_SH(stp23_m, stp21_m, vec1_m, vec0_m);                   \
+  SPLATI_H2_SH(coeff1_m, 4, 5, cnst0_m, cnst1_m);                  \
+  cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m);                       \
+  out5 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);           \
+                                                                   \
+  cnst0_m = __msa_splati_h(coeff2_m, 1);                           \
+  cnst0_m = __msa_ilvev_h(cnst1_m, cnst0_m);                       \
+  out11 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);          \
+                                                                   \
+  ILVRL_H2_SH(stp24_m, stp31_m, vec1_m, vec0_m);                   \
+  SPLATI_H2_SH(coeff1_m, 6, 7, cnst0_m, cnst1_m);                  \
+  cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                       \
+                                                                   \
+  out13 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst1_m);          \
+                                                                   \
+  cnst1_m = __msa_splati_h(coeff2_m, 3);                           \
+  cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m);                       \
+  out3 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);           \
+}
+
+#define FDCT_POSTPROC_2V_NEG_H(vec0, vec1) {      \
+  v8i16 tp0_m, tp1_m;                             \
+  v8i16 one_m = __msa_ldi_h(1);                   \
+                                                  \
+  tp0_m = __msa_clti_s_h(vec0, 0);                \
+  tp1_m = __msa_clti_s_h(vec1, 0);                \
+  vec0 += 1;                                      \
+  vec1 += 1;                                      \
+  tp0_m = one_m & tp0_m;                          \
+  tp1_m = one_m & tp1_m;                          \
+  vec0 += tp0_m;                                  \
+  vec1 += tp1_m;                                  \
+  vec0 >>= 2;                                     \
+  vec1 >>= 2;                                     \
+}
+
+#define FDCT32_POSTPROC_NEG_W(vec) {      \
+  v4i32 temp_m;                           \
+  v4i32 one_m = __msa_ldi_w(1);           \
+                                          \
+  temp_m = __msa_clti_s_w(vec, 0);        \
+  vec += 1;                               \
+  temp_m = one_m & temp_m;                \
+  vec += temp_m;                          \
+  vec >>= 2;                              \
+}
+
+#define FDCT32_POSTPROC_2V_POS_H(vec0, vec1) {      \
+  v8i16 tp0_m, tp1_m;                               \
+  v8i16 one = __msa_ldi_h(1);                       \
+                                                    \
+  tp0_m = __msa_clei_s_h(vec0, 0);                  \
+  tp1_m = __msa_clei_s_h(vec1, 0);                  \
+  tp0_m = (v8i16)__msa_xori_b((v16u8)tp0_m, 255);   \
+  tp1_m = (v8i16)__msa_xori_b((v16u8)tp1_m, 255);   \
+  vec0 += 1;                                        \
+  vec1 += 1;                                        \
+  tp0_m = one & tp0_m;                              \
+  tp1_m = one & tp1_m;                              \
+  vec0 += tp0_m;                                    \
+  vec1 += tp1_m;                                    \
+  vec0 >>= 2;                                       \
+  vec1 >>= 2;                                       \
+}
+
+#define DOTP_CONST_PAIR_W(reg0_left, reg1_left, reg0_right,      \
+                          reg1_right, const0, const1,            \
+                          out0, out1, out2, out3) {              \
+  v4i32 s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m, s7_m;          \
+  v2i64 tp0_m, tp1_m, tp2_m, tp3_m;                              \
+  v4i32 k0_m = __msa_fill_w((int32_t) const0);                   \
+                                                                 \
+  s0_m = __msa_fill_w((int32_t) const1);                         \
+  k0_m = __msa_ilvev_w(s0_m, k0_m);                              \
+                                                                 \
+  ILVRL_W2_SW(-reg1_left, reg0_left, s1_m, s0_m);                \
+  ILVRL_W2_SW(reg0_left, reg1_left, s3_m, s2_m);                 \
+  ILVRL_W2_SW(-reg1_right, reg0_right, s5_m, s4_m);              \
+  ILVRL_W2_SW(reg0_right, reg1_right, s7_m, s6_m);               \
+                                                                 \
+  DOTP_SW2_SD(s0_m, s1_m, k0_m, k0_m, tp0_m, tp1_m);             \
+  DOTP_SW2_SD(s4_m, s5_m, k0_m, k0_m, tp2_m, tp3_m);             \
+  tp0_m = __msa_srari_d(tp0_m, DCT_CONST_BITS);                  \
+  tp1_m = __msa_srari_d(tp1_m, DCT_CONST_BITS);                  \
+  tp2_m = __msa_srari_d(tp2_m, DCT_CONST_BITS);                  \
+  tp3_m = __msa_srari_d(tp3_m, DCT_CONST_BITS);                  \
+  out0 = __msa_pckev_w((v4i32)tp0_m, (v4i32)tp1_m);              \
+  out1 = __msa_pckev_w((v4i32)tp2_m, (v4i32)tp3_m);              \
+                                                                 \
+  DOTP_SW2_SD(s2_m, s3_m, k0_m, k0_m, tp0_m, tp1_m);             \
+  DOTP_SW2_SD(s6_m, s7_m, k0_m, k0_m, tp2_m, tp3_m);             \
+  tp0_m = __msa_srari_d(tp0_m, DCT_CONST_BITS);                  \
+  tp1_m = __msa_srari_d(tp1_m, DCT_CONST_BITS);                  \
+  tp2_m = __msa_srari_d(tp2_m, DCT_CONST_BITS);                  \
+  tp3_m = __msa_srari_d(tp3_m, DCT_CONST_BITS);                  \
+  out2 = __msa_pckev_w((v4i32)tp0_m, (v4i32)tp1_m);              \
+  out3 = __msa_pckev_w((v4i32)tp2_m, (v4i32)tp3_m);              \
+}
+
+void fdct8x16_1d_column(const int16_t *input, int16_t *tmp_ptr,
+                        int32_t src_stride);
+void fdct16x8_1d_row(int16_t *input, int16_t *output);
+#endif  // VPX_DSP_MIPS_FWD_TXFM_MSA_H_
diff --git a/libvpx/libvpx/vpx_dsp/mips/idct16x16_msa.c b/libvpx/libvpx/vpx_dsp/mips/idct16x16_msa.c
new file mode 100644
index 0000000..5faac71
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/idct16x16_msa.c
@@ -0,0 +1,487 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/inv_txfm_msa.h"
+
+void vpx_idct16_1d_rows_msa(const int16_t *input, int16_t *output) {
+  v8i16 loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg2, reg4, reg6, reg8, reg10, reg12, reg14;
+  v8i16 reg3, reg13, reg11, reg5, reg7, reg9, reg1, reg15;
+  v8i16 tmp5, tmp6, tmp7;
+
+  LD_SH8(input, 16, reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+  input += 8;
+  LD_SH8(input, 16, reg8, reg9, reg10, reg11, reg12, reg13, reg14, reg15);
+
+  TRANSPOSE8x8_SH_SH(reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7,
+                     reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+  TRANSPOSE8x8_SH_SH(reg8, reg9, reg10, reg11, reg12, reg13, reg14, reg15,
+                     reg8, reg9, reg10, reg11, reg12, reg13, reg14, reg15);
+  DOTP_CONST_PAIR(reg2, reg14, cospi_28_64, cospi_4_64, reg2, reg14);
+  DOTP_CONST_PAIR(reg10, reg6, cospi_12_64, cospi_20_64, reg10, reg6);
+  BUTTERFLY_4(reg2, reg14, reg6, reg10, loc0, loc1, reg14, reg2);
+  DOTP_CONST_PAIR(reg14, reg2, cospi_16_64, cospi_16_64, loc2, loc3);
+  DOTP_CONST_PAIR(reg0, reg8, cospi_16_64, cospi_16_64, reg0, reg8);
+  DOTP_CONST_PAIR(reg4, reg12, cospi_24_64, cospi_8_64, reg4, reg12);
+  BUTTERFLY_4(reg8, reg0, reg4, reg12, reg2, reg6, reg10, reg14);
+  SUB4(reg2, loc1, reg14, loc0, reg6, loc3, reg10, loc2, reg0, reg12, reg4,
+       reg8);
+  ADD4(reg2, loc1, reg14, loc0, reg6, loc3, reg10, loc2, reg2, reg14, reg6,
+       reg10);
+
+  /* stage 2 */
+  DOTP_CONST_PAIR(reg1, reg15, cospi_30_64, cospi_2_64, reg1, reg15);
+  DOTP_CONST_PAIR(reg9, reg7, cospi_14_64, cospi_18_64, loc2, loc3);
+
+  reg9 = reg1 - loc2;
+  reg1 = reg1 + loc2;
+  reg7 = reg15 - loc3;
+  reg15 = reg15 + loc3;
+
+  DOTP_CONST_PAIR(reg5, reg11, cospi_22_64, cospi_10_64, reg5, reg11);
+  DOTP_CONST_PAIR(reg13, reg3, cospi_6_64, cospi_26_64, loc0, loc1);
+  BUTTERFLY_4(loc0, loc1, reg11, reg5, reg13, reg3, reg11, reg5);
+
+  loc1 = reg15 + reg3;
+  reg3 = reg15 - reg3;
+  loc2 = reg2 + loc1;
+  reg15 = reg2 - loc1;
+
+  loc1 = reg1 + reg13;
+  reg13 = reg1 - reg13;
+  loc0 = reg0 + loc1;
+  loc1 = reg0 - loc1;
+  tmp6 = loc0;
+  tmp7 = loc1;
+  reg0 = loc2;
+
+  DOTP_CONST_PAIR(reg7, reg9, cospi_24_64, cospi_8_64, reg7, reg9);
+  DOTP_CONST_PAIR((-reg5), (-reg11), cospi_8_64, cospi_24_64, reg5, reg11);
+
+  loc0 = reg9 + reg5;
+  reg5 = reg9 - reg5;
+  reg2 = reg6 + loc0;
+  reg1 = reg6 - loc0;
+
+  loc0 = reg7 + reg11;
+  reg11 = reg7 - reg11;
+  loc1 = reg4 + loc0;
+  loc2 = reg4 - loc0;
+  tmp5 = loc1;
+
+  DOTP_CONST_PAIR(reg5, reg11, cospi_16_64, cospi_16_64, reg5, reg11);
+  BUTTERFLY_4(reg8, reg10, reg11, reg5, loc0, reg4, reg9, loc1);
+
+  reg10 = loc0;
+  reg11 = loc1;
+
+  DOTP_CONST_PAIR(reg3, reg13, cospi_16_64, cospi_16_64, reg3, reg13);
+  BUTTERFLY_4(reg12, reg14, reg13, reg3, reg8, reg6, reg7, reg5);
+
+  reg13 = loc2;
+
+  /* Transpose and store the output */
+  reg12 = tmp5;
+  reg14 = tmp6;
+  reg3 = tmp7;
+
+  /* transpose block */
+  TRANSPOSE8x8_SH_SH(reg0, reg2, reg4, reg6, reg8, reg10, reg12, reg14,
+                     reg0, reg2, reg4, reg6, reg8, reg10, reg12, reg14);
+  ST_SH8(reg0, reg2, reg4, reg6, reg8, reg10, reg12, reg14, output, 16);
+
+  /* transpose block */
+  TRANSPOSE8x8_SH_SH(reg3, reg13, reg11, reg5, reg7, reg9, reg1, reg15,
+                     reg3, reg13, reg11, reg5, reg7, reg9, reg1, reg15);
+  ST_SH8(reg3, reg13, reg11, reg5, reg7, reg9, reg1, reg15, (output + 8), 16);
+}
+
+void vpx_idct16_1d_columns_addblk_msa(int16_t *input, uint8_t *dst,
+                                      int32_t dst_stride) {
+  v8i16 loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg2, reg4, reg6, reg8, reg10, reg12, reg14;
+  v8i16 reg3, reg13, reg11, reg5, reg7, reg9, reg1, reg15;
+  v8i16 tmp5, tmp6, tmp7;
+
+  /* load up 8x8 */
+  LD_SH8(input, 16, reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+  input += 8 * 16;
+  /* load bottom 8x8 */
+  LD_SH8(input, 16, reg8, reg9, reg10, reg11, reg12, reg13, reg14, reg15);
+
+  DOTP_CONST_PAIR(reg2, reg14, cospi_28_64, cospi_4_64, reg2, reg14);
+  DOTP_CONST_PAIR(reg10, reg6, cospi_12_64, cospi_20_64, reg10, reg6);
+  BUTTERFLY_4(reg2, reg14, reg6, reg10, loc0, loc1, reg14, reg2);
+  DOTP_CONST_PAIR(reg14, reg2, cospi_16_64, cospi_16_64, loc2, loc3);
+  DOTP_CONST_PAIR(reg0, reg8, cospi_16_64, cospi_16_64, reg0, reg8);
+  DOTP_CONST_PAIR(reg4, reg12, cospi_24_64, cospi_8_64, reg4, reg12);
+  BUTTERFLY_4(reg8, reg0, reg4, reg12, reg2, reg6, reg10, reg14);
+
+  reg0 = reg2 - loc1;
+  reg2 = reg2 + loc1;
+  reg12 = reg14 - loc0;
+  reg14 = reg14 + loc0;
+  reg4 = reg6 - loc3;
+  reg6 = reg6 + loc3;
+  reg8 = reg10 - loc2;
+  reg10 = reg10 + loc2;
+
+  /* stage 2 */
+  DOTP_CONST_PAIR(reg1, reg15, cospi_30_64, cospi_2_64, reg1, reg15);
+  DOTP_CONST_PAIR(reg9, reg7, cospi_14_64, cospi_18_64, loc2, loc3);
+
+  reg9 = reg1 - loc2;
+  reg1 = reg1 + loc2;
+  reg7 = reg15 - loc3;
+  reg15 = reg15 + loc3;
+
+  DOTP_CONST_PAIR(reg5, reg11, cospi_22_64, cospi_10_64, reg5, reg11);
+  DOTP_CONST_PAIR(reg13, reg3, cospi_6_64, cospi_26_64, loc0, loc1);
+  BUTTERFLY_4(loc0, loc1, reg11, reg5, reg13, reg3, reg11, reg5);
+
+  loc1 = reg15 + reg3;
+  reg3 = reg15 - reg3;
+  loc2 = reg2 + loc1;
+  reg15 = reg2 - loc1;
+
+  loc1 = reg1 + reg13;
+  reg13 = reg1 - reg13;
+  loc0 = reg0 + loc1;
+  loc1 = reg0 - loc1;
+  tmp6 = loc0;
+  tmp7 = loc1;
+  reg0 = loc2;
+
+  DOTP_CONST_PAIR(reg7, reg9, cospi_24_64, cospi_8_64, reg7, reg9);
+  DOTP_CONST_PAIR((-reg5), (-reg11), cospi_8_64, cospi_24_64, reg5, reg11);
+
+  loc0 = reg9 + reg5;
+  reg5 = reg9 - reg5;
+  reg2 = reg6 + loc0;
+  reg1 = reg6 - loc0;
+
+  loc0 = reg7 + reg11;
+  reg11 = reg7 - reg11;
+  loc1 = reg4 + loc0;
+  loc2 = reg4 - loc0;
+  tmp5 = loc1;
+
+  DOTP_CONST_PAIR(reg5, reg11, cospi_16_64, cospi_16_64, reg5, reg11);
+  BUTTERFLY_4(reg8, reg10, reg11, reg5, loc0, reg4, reg9, loc1);
+
+  reg10 = loc0;
+  reg11 = loc1;
+
+  DOTP_CONST_PAIR(reg3, reg13, cospi_16_64, cospi_16_64, reg3, reg13);
+  BUTTERFLY_4(reg12, reg14, reg13, reg3, reg8, reg6, reg7, reg5);
+  reg13 = loc2;
+
+  /* Transpose and store the output */
+  reg12 = tmp5;
+  reg14 = tmp6;
+  reg3 = tmp7;
+
+  SRARI_H4_SH(reg0, reg2, reg4, reg6, 6);
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, reg0, reg2, reg4, reg6);
+  dst += (4 * dst_stride);
+  SRARI_H4_SH(reg8, reg10, reg12, reg14, 6);
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, reg8, reg10, reg12, reg14);
+  dst += (4 * dst_stride);
+  SRARI_H4_SH(reg3, reg13, reg11, reg5, 6);
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, reg3, reg13, reg11, reg5);
+  dst += (4 * dst_stride);
+  SRARI_H4_SH(reg7, reg9, reg1, reg15, 6);
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, reg7, reg9, reg1, reg15);
+}
+
+void vpx_idct16x16_256_add_msa(const int16_t *input, uint8_t *dst,
+                               int32_t dst_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, out_arr[16 * 16]);
+  int16_t *out = out_arr;
+
+  /* transform rows */
+  for (i = 0; i < 2; ++i) {
+    /* process 16 * 8 block */
+    vpx_idct16_1d_rows_msa((input + (i << 7)), (out + (i << 7)));
+  }
+
+  /* transform columns */
+  for (i = 0; i < 2; ++i) {
+    /* process 8 * 16 block */
+    vpx_idct16_1d_columns_addblk_msa((out + (i << 3)), (dst + (i << 3)),
+                                     dst_stride);
+  }
+}
+
+void vpx_idct16x16_10_add_msa(const int16_t *input, uint8_t *dst,
+                              int32_t dst_stride) {
+  uint8_t i;
+  DECLARE_ALIGNED(32, int16_t, out_arr[16 * 16]);
+  int16_t *out = out_arr;
+
+  /* process 16 * 8 block */
+  vpx_idct16_1d_rows_msa(input, out);
+
+  /* short case just considers top 4 rows as valid output */
+  out += 4 * 16;
+  for (i = 12; i--;) {
+    __asm__ __volatile__ (
+        "sw     $zero,   0(%[out])     \n\t"
+        "sw     $zero,   4(%[out])     \n\t"
+        "sw     $zero,   8(%[out])     \n\t"
+        "sw     $zero,  12(%[out])     \n\t"
+        "sw     $zero,  16(%[out])     \n\t"
+        "sw     $zero,  20(%[out])     \n\t"
+        "sw     $zero,  24(%[out])     \n\t"
+        "sw     $zero,  28(%[out])     \n\t"
+
+        :
+        : [out] "r" (out)
+    );
+
+    out += 16;
+  }
+
+  out = out_arr;
+
+  /* transform columns */
+  for (i = 0; i < 2; ++i) {
+    /* process 8 * 16 block */
+    vpx_idct16_1d_columns_addblk_msa((out + (i << 3)), (dst + (i << 3)),
+                                     dst_stride);
+  }
+}
+
+void vpx_idct16x16_1_add_msa(const int16_t *input, uint8_t *dst,
+                             int32_t dst_stride) {
+  uint8_t i;
+  int16_t out;
+  v8i16 vec, res0, res1, res2, res3, res4, res5, res6, res7;
+  v16u8 dst0, dst1, dst2, dst3, tmp0, tmp1, tmp2, tmp3;
+
+  out = ROUND_POWER_OF_TWO((input[0] * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO((out * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO(out, 6);
+
+  vec = __msa_fill_h(out);
+
+  for (i = 4; i--;) {
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    UNPCK_UB_SH(dst0, res0, res4);
+    UNPCK_UB_SH(dst1, res1, res5);
+    UNPCK_UB_SH(dst2, res2, res6);
+    UNPCK_UB_SH(dst3, res3, res7);
+    ADD4(res0, vec, res1, vec, res2, vec, res3, vec, res0, res1, res2, res3);
+    ADD4(res4, vec, res5, vec, res6, vec, res7, vec, res4, res5, res6, res7);
+    CLIP_SH4_0_255(res0, res1, res2, res3);
+    CLIP_SH4_0_255(res4, res5, res6, res7);
+    PCKEV_B4_UB(res4, res0, res5, res1, res6, res2, res7, res3,
+                tmp0, tmp1, tmp2, tmp3);
+    ST_UB4(tmp0, tmp1, tmp2, tmp3, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+void vpx_iadst16_1d_rows_msa(const int16_t *input, int16_t *output) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 l0, l1, l2, l3, l4, l5, l6, l7, l8, l9, l10, l11, l12, l13, l14, l15;
+
+  /* load input data */
+  LD_SH16(input, 8,
+          l0, l8, l1, l9, l2, l10, l3, l11, l4, l12, l5, l13, l6, l14, l7, l15);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7,
+                     l0, l1, l2, l3, l4, l5, l6, l7);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15,
+                     l8, l9, l10, l11, l12, l13, l14, l15);
+
+  /* ADST in horizontal */
+  VP9_IADST8x16_1D(l0, l1, l2, l3, l4, l5, l6, l7,
+                   l8, l9, l10, l11, l12, l13, l14, l15,
+                   r0, r1, r2, r3, r4, r5, r6, r7,
+                   r8, r9, r10, r11, r12, r13, r14, r15);
+
+  l1 = -r8;
+  l3 = -r4;
+  l13 = -r13;
+  l15 = -r1;
+
+  TRANSPOSE8x8_SH_SH(r0, l1, r12, l3, r6, r14, r10, r2,
+                     l0, l1, l2, l3, l4, l5, l6, l7);
+  ST_SH8(l0, l1, l2, l3, l4, l5, l6, l7, output, 16);
+  TRANSPOSE8x8_SH_SH(r3, r11, r15, r7, r5, l13, r9, l15,
+                     l8, l9, l10, l11, l12, l13, l14, l15);
+  ST_SH8(l8, l9, l10, l11, l12, l13, l14, l15, (output + 8), 16);
+}
+
+void vpx_iadst16_1d_columns_addblk_msa(int16_t *input, uint8_t *dst,
+                                       int32_t dst_stride) {
+  v8i16 v0, v2, v4, v6, k0, k1, k2, k3;
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 out0, out1, out2, out3, out4, out5, out6, out7;
+  v8i16 out8, out9, out10, out11, out12, out13, out14, out15;
+  v8i16 g0, g1, g2, g3, g4, g5, g6, g7, g8, g9, g10, g11, g12, g13, g14, g15;
+  v8i16 h0, h1, h2, h3, h4, h5, h6, h7, h8, h9, h10, h11;
+  v8i16 res0, res1, res2, res3, res4, res5, res6, res7;
+  v8i16 res8, res9, res10, res11, res12, res13, res14, res15;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16u8 dst8, dst9, dst10, dst11, dst12, dst13, dst14, dst15;
+  v16i8 zero = { 0 };
+
+  r0 = LD_SH(input + 0 * 16);
+  r3 = LD_SH(input + 3 * 16);
+  r4 = LD_SH(input + 4 * 16);
+  r7 = LD_SH(input + 7 * 16);
+  r8 = LD_SH(input + 8 * 16);
+  r11 = LD_SH(input + 11 * 16);
+  r12 = LD_SH(input + 12 * 16);
+  r15 = LD_SH(input + 15 * 16);
+
+  /* stage 1 */
+  k0 = VP9_SET_COSPI_PAIR(cospi_1_64, cospi_31_64);
+  k1 = VP9_SET_COSPI_PAIR(cospi_31_64, -cospi_1_64);
+  k2 = VP9_SET_COSPI_PAIR(cospi_17_64, cospi_15_64);
+  k3 = VP9_SET_COSPI_PAIR(cospi_15_64, -cospi_17_64);
+  MADD_BF(r15, r0, r7, r8, k0, k1, k2, k3, g0, g1, g2, g3);
+  k0 = VP9_SET_COSPI_PAIR(cospi_9_64, cospi_23_64);
+  k1 = VP9_SET_COSPI_PAIR(cospi_23_64, -cospi_9_64);
+  k2 = VP9_SET_COSPI_PAIR(cospi_25_64, cospi_7_64);
+  k3 = VP9_SET_COSPI_PAIR(cospi_7_64, -cospi_25_64);
+  MADD_BF(r11, r4, r3, r12, k0, k1, k2, k3, g8, g9, g10, g11);
+  BUTTERFLY_4(g0, g2, g10, g8, h8, h9, v2, v0);
+  k0 = VP9_SET_COSPI_PAIR(cospi_4_64, cospi_28_64);
+  k1 = VP9_SET_COSPI_PAIR(cospi_28_64, -cospi_4_64);
+  k2 = VP9_SET_COSPI_PAIR(-cospi_28_64, cospi_4_64);
+  MADD_BF(g1, g3, g9, g11, k0, k1, k2, k0, h0, h1, h2, h3);
+
+  r1 = LD_SH(input + 1 * 16);
+  r2 = LD_SH(input + 2 * 16);
+  r5 = LD_SH(input + 5 * 16);
+  r6 = LD_SH(input + 6 * 16);
+  r9 = LD_SH(input + 9 * 16);
+  r10 = LD_SH(input + 10 * 16);
+  r13 = LD_SH(input + 13 * 16);
+  r14 = LD_SH(input + 14 * 16);
+
+  k0 = VP9_SET_COSPI_PAIR(cospi_5_64, cospi_27_64);
+  k1 = VP9_SET_COSPI_PAIR(cospi_27_64, -cospi_5_64);
+  k2 = VP9_SET_COSPI_PAIR(cospi_21_64, cospi_11_64);
+  k3 = VP9_SET_COSPI_PAIR(cospi_11_64, -cospi_21_64);
+  MADD_BF(r13, r2, r5, r10, k0, k1, k2, k3, g4, g5, g6, g7);
+  k0 = VP9_SET_COSPI_PAIR(cospi_13_64, cospi_19_64);
+  k1 = VP9_SET_COSPI_PAIR(cospi_19_64, -cospi_13_64);
+  k2 = VP9_SET_COSPI_PAIR(cospi_29_64, cospi_3_64);
+  k3 = VP9_SET_COSPI_PAIR(cospi_3_64, -cospi_29_64);
+  MADD_BF(r9, r6, r1, r14, k0, k1, k2, k3, g12, g13, g14, g15);
+  BUTTERFLY_4(g4, g6, g14, g12, h10, h11, v6, v4);
+  BUTTERFLY_4(h8, h9, h11, h10, out0, out1, h11, h10);
+  out1 = -out1;
+  SRARI_H2_SH(out0, out1, 6);
+  dst0 = LD_UB(dst + 0 * dst_stride);
+  dst1 = LD_UB(dst + 15 * dst_stride);
+  ILVR_B2_SH(zero, dst0, zero, dst1, res0, res1);
+  ADD2(res0, out0, res1, out1, res0, res1);
+  CLIP_SH2_0_255(res0, res1);
+  PCKEV_B2_SH(res0, res0, res1, res1, res0, res1);
+  ST8x1_UB(res0, dst);
+  ST8x1_UB(res1, dst + 15 * dst_stride);
+
+  k0 = VP9_SET_COSPI_PAIR(cospi_12_64, cospi_20_64);
+  k1 = VP9_SET_COSPI_PAIR(-cospi_20_64, cospi_12_64);
+  k2 = VP9_SET_COSPI_PAIR(cospi_20_64, -cospi_12_64);
+  MADD_BF(g7, g5, g15, g13, k0, k1, k2, k0, h4, h5, h6, h7);
+  BUTTERFLY_4(h0, h2, h6, h4, out8, out9, out11, out10);
+  out8 = -out8;
+
+  SRARI_H2_SH(out8, out9, 6);
+  dst8 = LD_UB(dst + 1 * dst_stride);
+  dst9 = LD_UB(dst + 14 * dst_stride);
+  ILVR_B2_SH(zero, dst8, zero, dst9, res8, res9);
+  ADD2(res8, out8, res9, out9, res8, res9);
+  CLIP_SH2_0_255(res8, res9);
+  PCKEV_B2_SH(res8, res8, res9, res9, res8, res9);
+  ST8x1_UB(res8, dst + dst_stride);
+  ST8x1_UB(res9, dst + 14 * dst_stride);
+
+  k0 = VP9_SET_COSPI_PAIR(cospi_8_64, cospi_24_64);
+  k1 = VP9_SET_COSPI_PAIR(cospi_24_64, -cospi_8_64);
+  k2 = VP9_SET_COSPI_PAIR(-cospi_24_64, cospi_8_64);
+  MADD_BF(v0, v2, v4, v6, k0, k1, k2, k0, out4, out6, out5, out7);
+  out4 = -out4;
+  SRARI_H2_SH(out4, out5, 6);
+  dst4 = LD_UB(dst + 3 * dst_stride);
+  dst5 = LD_UB(dst + 12 * dst_stride);
+  ILVR_B2_SH(zero, dst4, zero, dst5, res4, res5);
+  ADD2(res4, out4, res5, out5, res4, res5);
+  CLIP_SH2_0_255(res4, res5);
+  PCKEV_B2_SH(res4, res4, res5, res5, res4, res5);
+  ST8x1_UB(res4, dst + 3 * dst_stride);
+  ST8x1_UB(res5, dst + 12 * dst_stride);
+
+  MADD_BF(h1, h3, h5, h7, k0, k1, k2, k0, out12, out14, out13, out15);
+  out13 = -out13;
+  SRARI_H2_SH(out12, out13, 6);
+  dst12 = LD_UB(dst + 2 * dst_stride);
+  dst13 = LD_UB(dst + 13 * dst_stride);
+  ILVR_B2_SH(zero, dst12, zero, dst13, res12, res13);
+  ADD2(res12, out12, res13, out13, res12, res13);
+  CLIP_SH2_0_255(res12, res13);
+  PCKEV_B2_SH(res12, res12, res13, res13, res12, res13);
+  ST8x1_UB(res12, dst + 2 * dst_stride);
+  ST8x1_UB(res13, dst + 13 * dst_stride);
+
+  k0 = VP9_SET_COSPI_PAIR(cospi_16_64, cospi_16_64);
+  k3 = VP9_SET_COSPI_PAIR(-cospi_16_64, cospi_16_64);
+  MADD_SHORT(out6, out7, k0, k3, out6, out7);
+  SRARI_H2_SH(out6, out7, 6);
+  dst6 = LD_UB(dst + 4 * dst_stride);
+  dst7 = LD_UB(dst + 11 * dst_stride);
+  ILVR_B2_SH(zero, dst6, zero, dst7, res6, res7);
+  ADD2(res6, out6, res7, out7, res6, res7);
+  CLIP_SH2_0_255(res6, res7);
+  PCKEV_B2_SH(res6, res6, res7, res7, res6, res7);
+  ST8x1_UB(res6, dst + 4 * dst_stride);
+  ST8x1_UB(res7, dst + 11 * dst_stride);
+
+  MADD_SHORT(out10, out11, k0, k3, out10, out11);
+  SRARI_H2_SH(out10, out11, 6);
+  dst10 = LD_UB(dst + 6 * dst_stride);
+  dst11 = LD_UB(dst + 9 * dst_stride);
+  ILVR_B2_SH(zero, dst10, zero, dst11, res10, res11);
+  ADD2(res10, out10, res11, out11, res10, res11);
+  CLIP_SH2_0_255(res10, res11);
+  PCKEV_B2_SH(res10, res10, res11, res11, res10, res11);
+  ST8x1_UB(res10, dst + 6 * dst_stride);
+  ST8x1_UB(res11, dst + 9 * dst_stride);
+
+  k1 = VP9_SET_COSPI_PAIR(-cospi_16_64, -cospi_16_64);
+  k2 = VP9_SET_COSPI_PAIR(cospi_16_64, -cospi_16_64);
+  MADD_SHORT(h10, h11, k1, k2, out2, out3);
+  SRARI_H2_SH(out2, out3, 6);
+  dst2 = LD_UB(dst + 7 * dst_stride);
+  dst3 = LD_UB(dst + 8 * dst_stride);
+  ILVR_B2_SH(zero, dst2, zero, dst3, res2, res3);
+  ADD2(res2, out2, res3, out3, res2, res3);
+  CLIP_SH2_0_255(res2, res3);
+  PCKEV_B2_SH(res2, res2, res3, res3, res2, res3);
+  ST8x1_UB(res2, dst + 7 * dst_stride);
+  ST8x1_UB(res3, dst + 8 * dst_stride);
+
+  MADD_SHORT(out14, out15, k1, k2, out14, out15);
+  SRARI_H2_SH(out14, out15, 6);
+  dst14 = LD_UB(dst + 5 * dst_stride);
+  dst15 = LD_UB(dst + 10 * dst_stride);
+  ILVR_B2_SH(zero, dst14, zero, dst15, res14, res15);
+  ADD2(res14, out14, res15, out15, res14, res15);
+  CLIP_SH2_0_255(res14, res15);
+  PCKEV_B2_SH(res14, res14, res15, res15, res14, res15);
+  ST8x1_UB(res14, dst + 5 * dst_stride);
+  ST8x1_UB(res15, dst + 10 * dst_stride);
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/idct32x32_msa.c b/libvpx/libvpx/vpx_dsp/mips/idct32x32_msa.c
new file mode 100644
index 0000000..d5b3966
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/idct32x32_msa.c
@@ -0,0 +1,739 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/inv_txfm_msa.h"
+
+static void idct32x8_row_transpose_store(const int16_t *input,
+                                         int16_t *tmp_buf) {
+  v8i16 m0, m1, m2, m3, m4, m5, m6, m7, n0, n1, n2, n3, n4, n5, n6, n7;
+
+  /* 1st & 2nd 8x8 */
+  LD_SH8(input, 32, m0, n0, m1, n1, m2, n2, m3, n3);
+  LD_SH8((input + 8), 32, m4, n4, m5, n5, m6, n6, m7, n7);
+  TRANSPOSE8x8_SH_SH(m0, n0, m1, n1, m2, n2, m3, n3,
+                     m0, n0, m1, n1, m2, n2, m3, n3);
+  TRANSPOSE8x8_SH_SH(m4, n4, m5, n5, m6, n6, m7, n7,
+                     m4, n4, m5, n5, m6, n6, m7, n7);
+  ST_SH8(m0, n0, m1, n1, m2, n2, m3, n3, (tmp_buf), 8);
+  ST_SH4(m4, n4, m5, n5, (tmp_buf + 8 * 8), 8);
+  ST_SH4(m6, n6, m7, n7, (tmp_buf + 12 * 8), 8);
+
+  /* 3rd & 4th 8x8 */
+  LD_SH8((input + 16), 32, m0, n0, m1, n1, m2, n2, m3, n3);
+  LD_SH8((input + 24), 32, m4, n4, m5, n5, m6, n6, m7, n7);
+  TRANSPOSE8x8_SH_SH(m0, n0, m1, n1, m2, n2, m3, n3,
+                     m0, n0, m1, n1, m2, n2, m3, n3);
+  TRANSPOSE8x8_SH_SH(m4, n4, m5, n5, m6, n6, m7, n7,
+                     m4, n4, m5, n5, m6, n6, m7, n7);
+  ST_SH4(m0, n0, m1, n1, (tmp_buf + 16 * 8), 8);
+  ST_SH4(m2, n2, m3, n3, (tmp_buf + 20 * 8), 8);
+  ST_SH4(m4, n4, m5, n5, (tmp_buf + 24 * 8), 8);
+  ST_SH4(m6, n6, m7, n7, (tmp_buf + 28 * 8), 8);
+}
+
+static void idct32x8_row_even_process_store(int16_t *tmp_buf,
+                                            int16_t *tmp_eve_buf) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v8i16 stp0, stp1, stp2, stp3, stp4, stp5, stp6, stp7;
+
+  /* Even stage 1 */
+  LD_SH8(tmp_buf, 32, reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+
+  DOTP_CONST_PAIR(reg1, reg7, cospi_28_64, cospi_4_64, reg1, reg7);
+  DOTP_CONST_PAIR(reg5, reg3, cospi_12_64, cospi_20_64, reg5, reg3);
+  BUTTERFLY_4(reg1, reg7, reg3, reg5, vec1, vec3, vec2, vec0);
+  DOTP_CONST_PAIR(vec2, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+
+  loc1 = vec3;
+  loc0 = vec1;
+
+  DOTP_CONST_PAIR(reg0, reg4, cospi_16_64, cospi_16_64, reg0, reg4);
+  DOTP_CONST_PAIR(reg2, reg6, cospi_24_64, cospi_8_64, reg2, reg6);
+  BUTTERFLY_4(reg4, reg0, reg2, reg6, vec1, vec3, vec2, vec0);
+  BUTTERFLY_4(vec0, vec1, loc1, loc0, stp3, stp0, stp7, stp4);
+  BUTTERFLY_4(vec2, vec3, loc3, loc2, stp2, stp1, stp6, stp5);
+
+  /* Even stage 2 */
+  LD_SH8((tmp_buf + 16), 32, reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+  DOTP_CONST_PAIR(reg0, reg7, cospi_30_64, cospi_2_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_14_64, cospi_18_64, reg4, reg3);
+  DOTP_CONST_PAIR(reg2, reg5, cospi_22_64, cospi_10_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg6, reg1, cospi_6_64, cospi_26_64, reg6, reg1);
+
+  vec0 = reg0 + reg4;
+  reg0 = reg0 - reg4;
+  reg4 = reg6 + reg2;
+  reg6 = reg6 - reg2;
+  reg2 = reg1 + reg5;
+  reg1 = reg1 - reg5;
+  reg5 = reg7 + reg3;
+  reg7 = reg7 - reg3;
+  reg3 = vec0;
+
+  vec1 = reg2;
+  reg2 = reg3 + reg4;
+  reg3 = reg3 - reg4;
+  reg4 = reg5 - vec1;
+  reg5 = reg5 + vec1;
+
+  DOTP_CONST_PAIR(reg7, reg0, cospi_24_64, cospi_8_64, reg0, reg7);
+  DOTP_CONST_PAIR((-reg6), reg1, cospi_24_64, cospi_8_64, reg6, reg1);
+
+  vec0 = reg0 - reg6;
+  reg0 = reg0 + reg6;
+  vec1 = reg7 - reg1;
+  reg7 = reg7 + reg1;
+
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, reg6, reg1);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_16_64, cospi_16_64, reg3, reg4);
+
+  /* Even stage 3 : Dependency on Even stage 1 & Even stage 2 */
+  BUTTERFLY_4(stp0, stp1, reg7, reg5, loc1, loc3, loc2, loc0);
+  ST_SH(loc0, (tmp_eve_buf + 15 * 8));
+  ST_SH(loc1, (tmp_eve_buf));
+  ST_SH(loc2, (tmp_eve_buf + 14 * 8));
+  ST_SH(loc3, (tmp_eve_buf + 8));
+
+  BUTTERFLY_4(stp2, stp3, reg4, reg1, loc1, loc3, loc2, loc0);
+  ST_SH(loc0, (tmp_eve_buf + 13 * 8));
+  ST_SH(loc1, (tmp_eve_buf + 2 * 8));
+  ST_SH(loc2, (tmp_eve_buf + 12 * 8));
+  ST_SH(loc3, (tmp_eve_buf + 3 * 8));
+
+  /* Store 8 */
+  BUTTERFLY_4(stp4, stp5, reg6, reg3, loc1, loc3, loc2, loc0);
+  ST_SH(loc0, (tmp_eve_buf + 11 * 8));
+  ST_SH(loc1, (tmp_eve_buf + 4 * 8));
+  ST_SH(loc2, (tmp_eve_buf + 10 * 8));
+  ST_SH(loc3, (tmp_eve_buf + 5 * 8));
+
+  BUTTERFLY_4(stp6, stp7, reg2, reg0, loc1, loc3, loc2, loc0);
+  ST_SH(loc0, (tmp_eve_buf + 9 * 8));
+  ST_SH(loc1, (tmp_eve_buf + 6 * 8));
+  ST_SH(loc2, (tmp_eve_buf + 8 * 8));
+  ST_SH(loc3, (tmp_eve_buf + 7 * 8));
+}
+
+static void idct32x8_row_odd_process_store(int16_t *tmp_buf,
+                                           int16_t *tmp_odd_buf) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+
+  /* Odd stage 1 */
+  reg0 = LD_SH(tmp_buf + 8);
+  reg1 = LD_SH(tmp_buf + 7 * 8);
+  reg2 = LD_SH(tmp_buf + 9 * 8);
+  reg3 = LD_SH(tmp_buf + 15 * 8);
+  reg4 = LD_SH(tmp_buf + 17 * 8);
+  reg5 = LD_SH(tmp_buf + 23 * 8);
+  reg6 = LD_SH(tmp_buf + 25 * 8);
+  reg7 = LD_SH(tmp_buf + 31 * 8);
+
+  DOTP_CONST_PAIR(reg0, reg7, cospi_31_64, cospi_1_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_15_64, cospi_17_64, reg3, reg4);
+  DOTP_CONST_PAIR(reg2, reg5, cospi_23_64, cospi_9_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg6, reg1, cospi_7_64, cospi_25_64, reg1, reg6);
+
+  vec0 = reg0 + reg3;
+  reg0 = reg0 - reg3;
+  reg3 = reg7 + reg4;
+  reg7 = reg7 - reg4;
+  reg4 = reg1 + reg2;
+  reg1 = reg1 - reg2;
+  reg2 = reg6 + reg5;
+  reg6 = reg6 - reg5;
+  reg5 = vec0;
+
+  /* 4 Stores */
+  ADD2(reg5, reg4, reg3, reg2, vec0, vec1);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 4 * 8), 8);
+
+  SUB2(reg5, reg4, reg3, reg2, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_24_64, cospi_8_64, vec0, vec1);
+  ST_SH2(vec0, vec1, (tmp_odd_buf), 8);
+
+  /* 4 Stores */
+  DOTP_CONST_PAIR(reg7, reg0, cospi_28_64, cospi_4_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg6, reg1, -cospi_4_64, cospi_28_64, reg1, reg6);
+  BUTTERFLY_4(reg0, reg7, reg6, reg1, vec0, vec1, vec2, vec3);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 6 * 8), 8);
+
+  DOTP_CONST_PAIR(vec2, vec3, cospi_24_64, cospi_8_64, vec2, vec3);
+  ST_SH2(vec2, vec3, (tmp_odd_buf + 2 * 8), 8);
+
+  /* Odd stage 2 */
+  /* 8 loads */
+  reg0 = LD_SH(tmp_buf + 3 * 8);
+  reg1 = LD_SH(tmp_buf + 5 * 8);
+  reg2 = LD_SH(tmp_buf + 11 * 8);
+  reg3 = LD_SH(tmp_buf + 13 * 8);
+  reg4 = LD_SH(tmp_buf + 19 * 8);
+  reg5 = LD_SH(tmp_buf + 21 * 8);
+  reg6 = LD_SH(tmp_buf + 27 * 8);
+  reg7 = LD_SH(tmp_buf + 29 * 8);
+
+  DOTP_CONST_PAIR(reg1, reg6, cospi_27_64, cospi_5_64, reg1, reg6);
+  DOTP_CONST_PAIR(reg5, reg2, cospi_11_64, cospi_21_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg3, reg4, cospi_19_64, cospi_13_64, reg3, reg4);
+  DOTP_CONST_PAIR(reg7, reg0, cospi_3_64, cospi_29_64, reg0, reg7);
+
+  /* 4 Stores */
+  SUB4(reg1, reg2, reg6, reg5, reg0, reg3, reg7, reg4,
+       vec0, vec1, vec2, vec3);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_12_64, cospi_20_64, loc0, loc1);
+  DOTP_CONST_PAIR(vec3, vec2, -cospi_20_64, cospi_12_64, loc2, loc3);
+
+  BUTTERFLY_4(loc3, loc2, loc0, loc1, vec1, vec0, vec2, vec3);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 12 * 8), 3 * 8);
+
+  DOTP_CONST_PAIR(vec3, vec2, -cospi_8_64, cospi_24_64, vec0, vec1);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 10 * 8), 8);
+
+  /* 4 Stores */
+  ADD4(reg1, reg2, reg6, reg5, reg0, reg3, reg7, reg4,
+       vec1, vec2, vec0, vec3);
+  BUTTERFLY_4(vec0, vec3, vec2, vec1, reg0, reg1, reg3, reg2);
+  ST_SH(reg0, (tmp_odd_buf + 13 * 8));
+  ST_SH(reg1, (tmp_odd_buf + 14 * 8));
+
+  DOTP_CONST_PAIR(reg3, reg2, -cospi_8_64, cospi_24_64, reg0, reg1);
+  ST_SH2(reg0, reg1, (tmp_odd_buf + 8 * 8), 8);
+
+  /* Odd stage 3 : Dependency on Odd stage 1 & Odd stage 2 */
+
+  /* Load 8 & Store 8 */
+  LD_SH4(tmp_odd_buf, 8, reg0, reg1, reg2, reg3);
+  LD_SH4((tmp_odd_buf + 8 * 8), 8, reg4, reg5, reg6, reg7);
+
+  ADD4(reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7,
+       loc0, loc1, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, tmp_odd_buf, 8);
+
+  SUB2(reg0, reg4, reg1, reg5, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc0, loc1);
+
+  SUB2(reg2, reg6, reg3, reg7, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 8 * 8), 8);
+
+  /* Load 8 & Store 8 */
+  LD_SH4((tmp_odd_buf + 4 * 8), 8, reg1, reg2, reg0, reg3);
+  LD_SH4((tmp_odd_buf + 12 * 8), 8, reg4, reg5, reg6, reg7);
+
+  ADD4(reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7,
+       loc0, loc1, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 4 * 8), 8);
+
+  SUB2(reg0, reg4, reg3, reg7, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc0, loc1);
+
+  SUB2(reg1, reg5, reg2, reg6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 12 * 8), 8);
+}
+
+static void idct_butterfly_transpose_store(int16_t *tmp_buf,
+                                           int16_t *tmp_eve_buf,
+                                           int16_t *tmp_odd_buf,
+                                           int16_t *dst) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 m0, m1, m2, m3, m4, m5, m6, m7, n0, n1, n2, n3, n4, n5, n6, n7;
+
+  /* FINAL BUTTERFLY : Dependency on Even & Odd */
+  vec0 = LD_SH(tmp_odd_buf);
+  vec1 = LD_SH(tmp_odd_buf + 9 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 14 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 6 * 8);
+  loc0 = LD_SH(tmp_eve_buf);
+  loc1 = LD_SH(tmp_eve_buf + 8 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 4 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 12 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m0, m4, m2, m6);
+
+  ST_SH((loc0 - vec3), (tmp_buf + 31 * 8));
+  ST_SH((loc1 - vec2), (tmp_buf + 23 * 8));
+  ST_SH((loc2 - vec1), (tmp_buf + 27 * 8));
+  ST_SH((loc3 - vec0), (tmp_buf + 19 * 8));
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 4 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 13 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 10 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 3 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 2 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 10 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 6 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 14 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m1, m5, m3, m7);
+
+  ST_SH((loc0 - vec3), (tmp_buf + 29 * 8));
+  ST_SH((loc1 - vec2), (tmp_buf + 21 * 8));
+  ST_SH((loc2 - vec1), (tmp_buf + 25 * 8));
+  ST_SH((loc3 - vec0), (tmp_buf + 17 * 8));
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 2 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 11 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 12 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 7 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 1 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 9 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 5 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 13 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n0, n4, n2, n6);
+
+  ST_SH((loc0 - vec3), (tmp_buf + 30 * 8));
+  ST_SH((loc1 - vec2), (tmp_buf + 22 * 8));
+  ST_SH((loc2 - vec1), (tmp_buf + 26 * 8));
+  ST_SH((loc3 - vec0), (tmp_buf + 18 * 8));
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 5 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 15 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 8 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 1 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 3 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 11 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 7 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 15 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n1, n5, n3, n7);
+
+  ST_SH((loc0 - vec3), (tmp_buf + 28 * 8));
+  ST_SH((loc1 - vec2), (tmp_buf + 20 * 8));
+  ST_SH((loc2 - vec1), (tmp_buf + 24 * 8));
+  ST_SH((loc3 - vec0), (tmp_buf + 16 * 8));
+
+  /* Transpose : 16 vectors */
+  /* 1st & 2nd 8x8 */
+  TRANSPOSE8x8_SH_SH(m0, n0, m1, n1, m2, n2, m3, n3,
+                     m0, n0, m1, n1, m2, n2, m3, n3);
+  ST_SH4(m0, n0, m1, n1, (dst + 0), 32);
+  ST_SH4(m2, n2, m3, n3, (dst + 4 * 32), 32);
+
+  TRANSPOSE8x8_SH_SH(m4, n4, m5, n5, m6, n6, m7, n7,
+                     m4, n4, m5, n5, m6, n6, m7, n7);
+  ST_SH4(m4, n4, m5, n5, (dst + 8), 32);
+  ST_SH4(m6, n6, m7, n7, (dst + 8 + 4 * 32), 32);
+
+  /* 3rd & 4th 8x8 */
+  LD_SH8((tmp_buf + 8 * 16), 8, m0, n0, m1, n1, m2, n2, m3, n3);
+  LD_SH8((tmp_buf + 12 * 16), 8, m4, n4, m5, n5, m6, n6, m7, n7);
+  TRANSPOSE8x8_SH_SH(m0, n0, m1, n1, m2, n2, m3, n3,
+                     m0, n0, m1, n1, m2, n2, m3, n3);
+  ST_SH4(m0, n0, m1, n1, (dst + 16), 32);
+  ST_SH4(m2, n2, m3, n3, (dst + 16 + 4 * 32), 32);
+
+  TRANSPOSE8x8_SH_SH(m4, n4, m5, n5, m6, n6, m7, n7,
+                     m4, n4, m5, n5, m6, n6, m7, n7);
+  ST_SH4(m4, n4, m5, n5, (dst + 24), 32);
+  ST_SH4(m6, n6, m7, n7, (dst + 24 + 4 * 32), 32);
+}
+
+static void idct32x8_1d_rows_msa(const int16_t *input, int16_t *output) {
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[8 * 32]);
+  DECLARE_ALIGNED(32, int16_t, tmp_odd_buf[16 * 8]);
+  DECLARE_ALIGNED(32, int16_t, tmp_eve_buf[16 * 8]);
+
+  idct32x8_row_transpose_store(input, &tmp_buf[0]);
+  idct32x8_row_even_process_store(&tmp_buf[0], &tmp_eve_buf[0]);
+  idct32x8_row_odd_process_store(&tmp_buf[0], &tmp_odd_buf[0]);
+  idct_butterfly_transpose_store(&tmp_buf[0], &tmp_eve_buf[0],
+                                 &tmp_odd_buf[0], output);
+}
+
+static void idct8x32_column_even_process_store(int16_t *tmp_buf,
+                                               int16_t *tmp_eve_buf) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v8i16 stp0, stp1, stp2, stp3, stp4, stp5, stp6, stp7;
+
+  /* Even stage 1 */
+  LD_SH8(tmp_buf, (4 * 32), reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+  tmp_buf += (2 * 32);
+
+  DOTP_CONST_PAIR(reg1, reg7, cospi_28_64, cospi_4_64, reg1, reg7);
+  DOTP_CONST_PAIR(reg5, reg3, cospi_12_64, cospi_20_64, reg5, reg3);
+  BUTTERFLY_4(reg1, reg7, reg3, reg5, vec1, vec3, vec2, vec0);
+  DOTP_CONST_PAIR(vec2, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+
+  loc1 = vec3;
+  loc0 = vec1;
+
+  DOTP_CONST_PAIR(reg0, reg4, cospi_16_64, cospi_16_64, reg0, reg4);
+  DOTP_CONST_PAIR(reg2, reg6, cospi_24_64, cospi_8_64, reg2, reg6);
+  BUTTERFLY_4(reg4, reg0, reg2, reg6, vec1, vec3, vec2, vec0);
+  BUTTERFLY_4(vec0, vec1, loc1, loc0, stp3, stp0, stp7, stp4);
+  BUTTERFLY_4(vec2, vec3, loc3, loc2, stp2, stp1, stp6, stp5);
+
+  /* Even stage 2 */
+  /* Load 8 */
+  LD_SH8(tmp_buf, (4 * 32), reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+
+  DOTP_CONST_PAIR(reg0, reg7, cospi_30_64, cospi_2_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_14_64, cospi_18_64, reg4, reg3);
+  DOTP_CONST_PAIR(reg2, reg5, cospi_22_64, cospi_10_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg6, reg1, cospi_6_64, cospi_26_64, reg6, reg1);
+
+  vec0 = reg0 + reg4;
+  reg0 = reg0 - reg4;
+  reg4 = reg6 + reg2;
+  reg6 = reg6 - reg2;
+  reg2 = reg1 + reg5;
+  reg1 = reg1 - reg5;
+  reg5 = reg7 + reg3;
+  reg7 = reg7 - reg3;
+  reg3 = vec0;
+
+  vec1 = reg2;
+  reg2 = reg3 + reg4;
+  reg3 = reg3 - reg4;
+  reg4 = reg5 - vec1;
+  reg5 = reg5 + vec1;
+
+  DOTP_CONST_PAIR(reg7, reg0, cospi_24_64, cospi_8_64, reg0, reg7);
+  DOTP_CONST_PAIR((-reg6), reg1, cospi_24_64, cospi_8_64, reg6, reg1);
+
+  vec0 = reg0 - reg6;
+  reg0 = reg0 + reg6;
+  vec1 = reg7 - reg1;
+  reg7 = reg7 + reg1;
+
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, reg6, reg1);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_16_64, cospi_16_64, reg3, reg4);
+
+  /* Even stage 3 : Dependency on Even stage 1 & Even stage 2 */
+  /* Store 8 */
+  BUTTERFLY_4(stp0, stp1, reg7, reg5, loc1, loc3, loc2, loc0);
+  ST_SH2(loc1, loc3, tmp_eve_buf, 8);
+  ST_SH2(loc2, loc0, (tmp_eve_buf + 14 * 8), 8);
+
+  BUTTERFLY_4(stp2, stp3, reg4, reg1, loc1, loc3, loc2, loc0);
+  ST_SH2(loc1, loc3, (tmp_eve_buf + 2 * 8), 8);
+  ST_SH2(loc2, loc0, (tmp_eve_buf + 12 * 8), 8);
+
+  /* Store 8 */
+  BUTTERFLY_4(stp4, stp5, reg6, reg3, loc1, loc3, loc2, loc0);
+  ST_SH2(loc1, loc3, (tmp_eve_buf + 4 * 8), 8);
+  ST_SH2(loc2, loc0, (tmp_eve_buf + 10 * 8), 8);
+
+  BUTTERFLY_4(stp6, stp7, reg2, reg0, loc1, loc3, loc2, loc0);
+  ST_SH2(loc1, loc3, (tmp_eve_buf + 6 * 8), 8);
+  ST_SH2(loc2, loc0, (tmp_eve_buf + 8 * 8), 8);
+}
+
+static void idct8x32_column_odd_process_store(int16_t *tmp_buf,
+                                              int16_t *tmp_odd_buf) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+
+  /* Odd stage 1 */
+  reg0 = LD_SH(tmp_buf + 32);
+  reg1 = LD_SH(tmp_buf + 7 * 32);
+  reg2 = LD_SH(tmp_buf + 9 * 32);
+  reg3 = LD_SH(tmp_buf + 15 * 32);
+  reg4 = LD_SH(tmp_buf + 17 * 32);
+  reg5 = LD_SH(tmp_buf + 23 * 32);
+  reg6 = LD_SH(tmp_buf + 25 * 32);
+  reg7 = LD_SH(tmp_buf + 31 * 32);
+
+  DOTP_CONST_PAIR(reg0, reg7, cospi_31_64, cospi_1_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_15_64, cospi_17_64, reg3, reg4);
+  DOTP_CONST_PAIR(reg2, reg5, cospi_23_64, cospi_9_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg6, reg1, cospi_7_64, cospi_25_64, reg1, reg6);
+
+  vec0 = reg0 + reg3;
+  reg0 = reg0 - reg3;
+  reg3 = reg7 + reg4;
+  reg7 = reg7 - reg4;
+  reg4 = reg1 + reg2;
+  reg1 = reg1 - reg2;
+  reg2 = reg6 + reg5;
+  reg6 = reg6 - reg5;
+  reg5 = vec0;
+
+  /* 4 Stores */
+  ADD2(reg5, reg4, reg3, reg2, vec0, vec1);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 4 * 8), 8);
+  SUB2(reg5, reg4, reg3, reg2, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_24_64, cospi_8_64, vec0, vec1);
+  ST_SH2(vec0, vec1, tmp_odd_buf, 8);
+
+  /* 4 Stores */
+  DOTP_CONST_PAIR(reg7, reg0, cospi_28_64, cospi_4_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg6, reg1, -cospi_4_64, cospi_28_64, reg1, reg6);
+  BUTTERFLY_4(reg0, reg7, reg6, reg1, vec0, vec1, vec2, vec3);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 6 * 8), 8);
+  DOTP_CONST_PAIR(vec2, vec3, cospi_24_64, cospi_8_64, vec2, vec3);
+  ST_SH2(vec2, vec3, (tmp_odd_buf + 2 * 8), 8);
+
+  /* Odd stage 2 */
+  /* 8 loads */
+  reg0 = LD_SH(tmp_buf + 3 * 32);
+  reg1 = LD_SH(tmp_buf + 5 * 32);
+  reg2 = LD_SH(tmp_buf + 11 * 32);
+  reg3 = LD_SH(tmp_buf + 13 * 32);
+  reg4 = LD_SH(tmp_buf + 19 * 32);
+  reg5 = LD_SH(tmp_buf + 21 * 32);
+  reg6 = LD_SH(tmp_buf + 27 * 32);
+  reg7 = LD_SH(tmp_buf + 29 * 32);
+
+  DOTP_CONST_PAIR(reg1, reg6, cospi_27_64, cospi_5_64, reg1, reg6);
+  DOTP_CONST_PAIR(reg5, reg2, cospi_11_64, cospi_21_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg3, reg4, cospi_19_64, cospi_13_64, reg3, reg4);
+  DOTP_CONST_PAIR(reg7, reg0, cospi_3_64, cospi_29_64, reg0, reg7);
+
+  /* 4 Stores */
+  SUB4(reg1, reg2, reg6, reg5, reg0, reg3, reg7, reg4, vec0, vec1, vec2, vec3);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_12_64, cospi_20_64, loc0, loc1);
+  DOTP_CONST_PAIR(vec3, vec2, -cospi_20_64, cospi_12_64, loc2, loc3);
+  BUTTERFLY_4(loc2, loc3, loc1, loc0, vec0, vec1, vec3, vec2);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 12 * 8), 3 * 8);
+  DOTP_CONST_PAIR(vec3, vec2, -cospi_8_64, cospi_24_64, vec0, vec1);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 10 * 8), 8);
+
+  /* 4 Stores */
+  ADD4(reg0, reg3, reg1, reg2, reg5, reg6, reg4, reg7, vec0, vec1, vec2, vec3);
+  BUTTERFLY_4(vec0, vec3, vec2, vec1, reg0, reg1, reg3, reg2);
+  ST_SH2(reg0, reg1, (tmp_odd_buf + 13 * 8), 8);
+  DOTP_CONST_PAIR(reg3, reg2, -cospi_8_64, cospi_24_64, reg0, reg1);
+  ST_SH2(reg0, reg1, (tmp_odd_buf + 8 * 8), 8);
+
+  /* Odd stage 3 : Dependency on Odd stage 1 & Odd stage 2 */
+  /* Load 8 & Store 8 */
+  LD_SH4(tmp_odd_buf, 8, reg0, reg1, reg2, reg3);
+  LD_SH4((tmp_odd_buf + 8 * 8), 8, reg4, reg5, reg6, reg7);
+
+  ADD4(reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7, loc0, loc1, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, tmp_odd_buf, 8);
+
+  SUB2(reg0, reg4, reg1, reg5, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc0, loc1);
+
+  SUB2(reg2, reg6, reg3, reg7, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 8 * 8), 8);
+
+  /* Load 8 & Store 8 */
+  LD_SH4((tmp_odd_buf + 4 * 8), 8, reg1, reg2, reg0, reg3);
+  LD_SH4((tmp_odd_buf + 12 * 8), 8, reg4, reg5, reg6, reg7);
+
+  ADD4(reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7, loc0, loc1, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 4 * 8), 8);
+
+  SUB2(reg0, reg4, reg3, reg7, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc0, loc1);
+
+  SUB2(reg1, reg5, reg2, reg6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 12 * 8), 8);
+}
+
+static void idct8x32_column_butterfly_addblk(int16_t *tmp_eve_buf,
+                                             int16_t *tmp_odd_buf,
+                                             uint8_t *dst,
+                                             int32_t dst_stride) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 m0, m1, m2, m3, m4, m5, m6, m7, n0, n1, n2, n3, n4, n5, n6, n7;
+
+  /* FINAL BUTTERFLY : Dependency on Even & Odd */
+  vec0 = LD_SH(tmp_odd_buf);
+  vec1 = LD_SH(tmp_odd_buf + 9 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 14 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 6 * 8);
+  loc0 = LD_SH(tmp_eve_buf);
+  loc1 = LD_SH(tmp_eve_buf + 8 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 4 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 12 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m0, m4, m2, m6);
+  SRARI_H4_SH(m0, m2, m4, m6, 6);
+  VP9_ADDBLK_ST8x4_UB(dst, (4 * dst_stride), m0, m2, m4, m6);
+
+  SUB4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m6, m2, m4, m0);
+  SRARI_H4_SH(m0, m2, m4, m6, 6);
+  VP9_ADDBLK_ST8x4_UB((dst + 19 * dst_stride), (4 * dst_stride),
+                      m0, m2, m4, m6);
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 4 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 13 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 10 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 3 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 2 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 10 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 6 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 14 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m1, m5, m3, m7);
+  SRARI_H4_SH(m1, m3, m5, m7, 6);
+  VP9_ADDBLK_ST8x4_UB((dst + 2 * dst_stride), (4 * dst_stride),
+                      m1, m3, m5, m7);
+
+  SUB4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m7, m3, m5, m1);
+  SRARI_H4_SH(m1, m3, m5, m7, 6);
+  VP9_ADDBLK_ST8x4_UB((dst + 17 * dst_stride), (4 * dst_stride),
+                      m1, m3, m5, m7);
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 2 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 11 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 12 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 7 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 1 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 9 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 5 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 13 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n0, n4, n2, n6);
+  SRARI_H4_SH(n0, n2, n4, n6, 6);
+  VP9_ADDBLK_ST8x4_UB((dst + 1 * dst_stride), (4 * dst_stride),
+                      n0, n2, n4, n6);
+
+  SUB4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n6, n2, n4, n0);
+  SRARI_H4_SH(n0, n2, n4, n6, 6);
+  VP9_ADDBLK_ST8x4_UB((dst + 18 * dst_stride), (4 * dst_stride),
+                      n0, n2, n4, n6);
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 5 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 15 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 8 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 1 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 3 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 11 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 7 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 15 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n1, n5, n3, n7);
+  SRARI_H4_SH(n1, n3, n5, n7, 6);
+  VP9_ADDBLK_ST8x4_UB((dst + 3 * dst_stride), (4 * dst_stride),
+                      n1, n3, n5, n7);
+
+  SUB4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n7, n3, n5, n1);
+  SRARI_H4_SH(n1, n3, n5, n7, 6);
+  VP9_ADDBLK_ST8x4_UB((dst + 16 * dst_stride), (4 * dst_stride),
+                      n1, n3, n5, n7);
+}
+
+static void idct8x32_1d_columns_addblk_msa(int16_t *input, uint8_t *dst,
+                                           int32_t dst_stride) {
+  DECLARE_ALIGNED(32, int16_t, tmp_odd_buf[16 * 8]);
+  DECLARE_ALIGNED(32, int16_t, tmp_eve_buf[16 * 8]);
+
+  idct8x32_column_even_process_store(input, &tmp_eve_buf[0]);
+  idct8x32_column_odd_process_store(input, &tmp_odd_buf[0]);
+  idct8x32_column_butterfly_addblk(&tmp_eve_buf[0], &tmp_odd_buf[0],
+                                   dst, dst_stride);
+}
+
+void vpx_idct32x32_1024_add_msa(const int16_t *input, uint8_t *dst,
+                                int32_t dst_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, out_arr[32 * 32]);
+  int16_t *out_ptr = out_arr;
+
+  /* transform rows */
+  for (i = 0; i < 4; ++i) {
+    /* process 32 * 8 block */
+    idct32x8_1d_rows_msa((input + (i << 8)), (out_ptr + (i << 8)));
+  }
+
+  /* transform columns */
+  for (i = 0; i < 4; ++i) {
+    /* process 8 * 32 block */
+    idct8x32_1d_columns_addblk_msa((out_ptr + (i << 3)), (dst + (i << 3)),
+                                   dst_stride);
+  }
+}
+
+void vpx_idct32x32_34_add_msa(const int16_t *input, uint8_t *dst,
+                              int32_t dst_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, out_arr[32 * 32]);
+  int16_t *out_ptr = out_arr;
+
+  for (i = 32; i--;) {
+    __asm__ __volatile__ (
+        "sw     $zero,      0(%[out_ptr])     \n\t"
+        "sw     $zero,      4(%[out_ptr])     \n\t"
+        "sw     $zero,      8(%[out_ptr])     \n\t"
+        "sw     $zero,     12(%[out_ptr])     \n\t"
+        "sw     $zero,     16(%[out_ptr])     \n\t"
+        "sw     $zero,     20(%[out_ptr])     \n\t"
+        "sw     $zero,     24(%[out_ptr])     \n\t"
+        "sw     $zero,     28(%[out_ptr])     \n\t"
+        "sw     $zero,     32(%[out_ptr])     \n\t"
+        "sw     $zero,     36(%[out_ptr])     \n\t"
+        "sw     $zero,     40(%[out_ptr])     \n\t"
+        "sw     $zero,     44(%[out_ptr])     \n\t"
+        "sw     $zero,     48(%[out_ptr])     \n\t"
+        "sw     $zero,     52(%[out_ptr])     \n\t"
+        "sw     $zero,     56(%[out_ptr])     \n\t"
+        "sw     $zero,     60(%[out_ptr])     \n\t"
+
+        :
+        : [out_ptr] "r" (out_ptr)
+    );
+
+    out_ptr += 32;
+  }
+
+  out_ptr = out_arr;
+
+  /* rows: only upper-left 8x8 has non-zero coeff */
+  idct32x8_1d_rows_msa(input, out_ptr);
+
+  /* transform columns */
+  for (i = 0; i < 4; ++i) {
+    /* process 8 * 32 block */
+    idct8x32_1d_columns_addblk_msa((out_ptr + (i << 3)), (dst + (i << 3)),
+                                   dst_stride);
+  }
+}
+
+void vpx_idct32x32_1_add_msa(const int16_t *input, uint8_t *dst,
+                             int32_t dst_stride) {
+  int32_t i;
+  int16_t out;
+  v16u8 dst0, dst1, dst2, dst3, tmp0, tmp1, tmp2, tmp3;
+  v8i16 res0, res1, res2, res3, res4, res5, res6, res7, vec;
+
+  out = ROUND_POWER_OF_TWO((input[0] * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO((out * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO(out, 6);
+
+  vec = __msa_fill_h(out);
+
+  for (i = 16; i--;) {
+    LD_UB2(dst, 16, dst0, dst1);
+    LD_UB2(dst + dst_stride, 16, dst2, dst3);
+
+    UNPCK_UB_SH(dst0, res0, res4);
+    UNPCK_UB_SH(dst1, res1, res5);
+    UNPCK_UB_SH(dst2, res2, res6);
+    UNPCK_UB_SH(dst3, res3, res7);
+    ADD4(res0, vec, res1, vec, res2, vec, res3, vec, res0, res1, res2, res3);
+    ADD4(res4, vec, res5, vec, res6, vec, res7, vec, res4, res5, res6, res7);
+    CLIP_SH4_0_255(res0, res1, res2, res3);
+    CLIP_SH4_0_255(res4, res5, res6, res7);
+    PCKEV_B4_UB(res4, res0, res5, res1, res6, res2, res7, res3,
+                tmp0, tmp1, tmp2, tmp3);
+
+    ST_UB2(tmp0, tmp1, dst, 16);
+    dst += dst_stride;
+    ST_UB2(tmp2, tmp3, dst, 16);
+    dst += dst_stride;
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/idct4x4_msa.c b/libvpx/libvpx/vpx_dsp/mips/idct4x4_msa.c
new file mode 100644
index 0000000..f289d8e
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/idct4x4_msa.c
@@ -0,0 +1,98 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/inv_txfm_msa.h"
+
+void vpx_iwht4x4_16_add_msa(const int16_t *input, uint8_t *dst,
+                            int32_t dst_stride) {
+  v8i16 in0, in1, in2, in3;
+  v4i32 in0_r, in1_r, in2_r, in3_r, in4_r;
+
+  /* load vector elements of 4x4 block */
+  LD4x4_SH(input, in0, in2, in3, in1);
+  TRANSPOSE4x4_SH_SH(in0, in2, in3, in1, in0, in2, in3, in1);
+  UNPCK_R_SH_SW(in0, in0_r);
+  UNPCK_R_SH_SW(in2, in2_r);
+  UNPCK_R_SH_SW(in3, in3_r);
+  UNPCK_R_SH_SW(in1, in1_r);
+  SRA_4V(in0_r, in1_r, in2_r, in3_r, UNIT_QUANT_SHIFT);
+
+  in0_r += in2_r;
+  in3_r -= in1_r;
+  in4_r = (in0_r - in3_r) >> 1;
+  in1_r = in4_r - in1_r;
+  in2_r = in4_r - in2_r;
+  in0_r -= in1_r;
+  in3_r += in2_r;
+
+  TRANSPOSE4x4_SW_SW(in0_r, in1_r, in2_r, in3_r, in0_r, in1_r, in2_r, in3_r);
+
+  in0_r += in1_r;
+  in2_r -= in3_r;
+  in4_r = (in0_r - in2_r) >> 1;
+  in3_r = in4_r - in3_r;
+  in1_r = in4_r - in1_r;
+  in0_r -= in3_r;
+  in2_r += in1_r;
+
+  PCKEV_H4_SH(in0_r, in0_r, in1_r, in1_r, in2_r, in2_r, in3_r, in3_r,
+              in0, in1, in2, in3);
+  ADDBLK_ST4x4_UB(in0, in3, in1, in2, dst, dst_stride);
+}
+
+void vpx_iwht4x4_1_add_msa(const int16_t *input, uint8_t *dst,
+                           int32_t dst_stride) {
+  int16_t a1, e1;
+  v8i16 in1, in0 = { 0 };
+
+  a1 = input[0] >> UNIT_QUANT_SHIFT;
+  e1 = a1 >> 1;
+  a1 -= e1;
+
+  in0 = __msa_insert_h(in0, 0, a1);
+  in0 = __msa_insert_h(in0, 1, e1);
+  in0 = __msa_insert_h(in0, 2, e1);
+  in0 = __msa_insert_h(in0, 3, e1);
+
+  in1 = in0 >> 1;
+  in0 -= in1;
+
+  ADDBLK_ST4x4_UB(in0, in1, in1, in1, dst, dst_stride);
+}
+
+void vpx_idct4x4_16_add_msa(const int16_t *input, uint8_t *dst,
+                            int32_t dst_stride) {
+  v8i16 in0, in1, in2, in3;
+
+  /* load vector elements of 4x4 block */
+  LD4x4_SH(input, in0, in1, in2, in3);
+  /* rows */
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  VP9_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+  /* columns */
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  VP9_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+  /* rounding (add 2^3, divide by 2^4) */
+  SRARI_H4_SH(in0, in1, in2, in3, 4);
+  ADDBLK_ST4x4_UB(in0, in1, in2, in3, dst, dst_stride);
+}
+
+void vpx_idct4x4_1_add_msa(const int16_t *input, uint8_t *dst,
+                           int32_t dst_stride) {
+  int16_t out;
+  v8i16 vec;
+
+  out = ROUND_POWER_OF_TWO((input[0] * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO((out * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO(out, 4);
+  vec = __msa_fill_h(out);
+
+  ADDBLK_ST4x4_UB(vec, vec, vec, vec, dst, dst_stride);
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/idct8x8_msa.c b/libvpx/libvpx/vpx_dsp/mips/idct8x8_msa.c
new file mode 100644
index 0000000..fd667e4
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/idct8x8_msa.c
@@ -0,0 +1,116 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/inv_txfm_msa.h"
+
+void vpx_idct8x8_64_add_msa(const int16_t *input, uint8_t *dst,
+                            int32_t dst_stride) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  /* load vector elements of 8x8 block */
+  LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+
+  /* rows transform */
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  /* 1D idct8x8 */
+  VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7,
+                 in0, in1, in2, in3, in4, in5, in6, in7);
+  /* columns transform */
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  /* 1D idct8x8 */
+  VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7,
+                 in0, in1, in2, in3, in4, in5, in6, in7);
+  /* final rounding (add 2^4, divide by 2^5) and shift */
+  SRARI_H4_SH(in0, in1, in2, in3, 5);
+  SRARI_H4_SH(in4, in5, in6, in7, 5);
+  /* add block and store 8x8 */
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, in0, in1, in2, in3);
+  dst += (4 * dst_stride);
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, in4, in5, in6, in7);
+}
+
+void vpx_idct8x8_12_add_msa(const int16_t *input, uint8_t *dst,
+                            int32_t dst_stride) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 s0, s1, s2, s3, s4, s5, s6, s7, k0, k1, k2, k3, m0, m1, m2, m3;
+  v4i32 tmp0, tmp1, tmp2, tmp3;
+  v8i16 zero = { 0 };
+
+  /* load vector elements of 8x8 block */
+  LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8X4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+
+  /* stage1 */
+  ILVL_H2_SH(in3, in0, in2, in1, s0, s1);
+  k0 = VP9_SET_COSPI_PAIR(cospi_28_64, -cospi_4_64);
+  k1 = VP9_SET_COSPI_PAIR(cospi_4_64, cospi_28_64);
+  k2 = VP9_SET_COSPI_PAIR(-cospi_20_64, cospi_12_64);
+  k3 = VP9_SET_COSPI_PAIR(cospi_12_64, cospi_20_64);
+  DOTP_SH4_SW(s0, s0, s1, s1, k0, k1, k2, k3, tmp0, tmp1, tmp2, tmp3);
+  SRARI_W4_SW(tmp0, tmp1, tmp2, tmp3, DCT_CONST_BITS);
+  PCKEV_H2_SH(zero, tmp0, zero, tmp1, s0, s1);
+  PCKEV_H2_SH(zero, tmp2, zero, tmp3, s2, s3);
+  BUTTERFLY_4(s0, s1, s3, s2, s4, s7, s6, s5);
+
+  /* stage2 */
+  ILVR_H2_SH(in3, in1, in2, in0, s1, s0);
+  k0 = VP9_SET_COSPI_PAIR(cospi_16_64, cospi_16_64);
+  k1 = VP9_SET_COSPI_PAIR(cospi_16_64, -cospi_16_64);
+  k2 = VP9_SET_COSPI_PAIR(cospi_24_64, -cospi_8_64);
+  k3 = VP9_SET_COSPI_PAIR(cospi_8_64, cospi_24_64);
+  DOTP_SH4_SW(s0, s0, s1, s1, k0, k1, k2, k3, tmp0, tmp1, tmp2, tmp3);
+  SRARI_W4_SW(tmp0, tmp1, tmp2, tmp3, DCT_CONST_BITS);
+  PCKEV_H2_SH(zero, tmp0, zero, tmp1, s0, s1);
+  PCKEV_H2_SH(zero, tmp2, zero, tmp3, s2, s3);
+  BUTTERFLY_4(s0, s1, s2, s3, m0, m1, m2, m3);
+
+  /* stage3 */
+  s0 = __msa_ilvr_h(s6, s5);
+
+  k1 = VP9_SET_COSPI_PAIR(-cospi_16_64, cospi_16_64);
+  DOTP_SH2_SW(s0, s0, k1, k0, tmp0, tmp1);
+  SRARI_W2_SW(tmp0, tmp1, DCT_CONST_BITS);
+  PCKEV_H2_SH(zero, tmp0, zero, tmp1, s2, s3);
+
+  /* stage4 */
+  BUTTERFLY_8(m0, m1, m2, m3, s4, s2, s3, s7,
+              in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE4X8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7,
+                 in0, in1, in2, in3, in4, in5, in6, in7);
+
+  /* final rounding (add 2^4, divide by 2^5) and shift */
+  SRARI_H4_SH(in0, in1, in2, in3, 5);
+  SRARI_H4_SH(in4, in5, in6, in7, 5);
+
+  /* add block and store 8x8 */
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, in0, in1, in2, in3);
+  dst += (4 * dst_stride);
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, in4, in5, in6, in7);
+}
+
+void vpx_idct8x8_1_add_msa(const int16_t *input, uint8_t *dst,
+                           int32_t dst_stride) {
+  int16_t out;
+  int32_t val;
+  v8i16 vec;
+
+  out = ROUND_POWER_OF_TWO((input[0] * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO((out * cospi_16_64), DCT_CONST_BITS);
+  val = ROUND_POWER_OF_TWO(out, 5);
+  vec = __msa_fill_h(val);
+
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, vec, vec, vec, vec);
+  dst += (4 * dst_stride);
+  VP9_ADDBLK_ST8x4_UB(dst, dst_stride, vec, vec, vec, vec);
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/intrapred16_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/intrapred16_dspr2.c
new file mode 100644
index 0000000..11444c7
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/intrapred16_dspr2.c
@@ -0,0 +1,329 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/common_dspr2.h"
+
+#if HAVE_DSPR2
+void vpx_h_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  int32_t  tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
+  int32_t  tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+
+  __asm__ __volatile__ (
+      "lb         %[tmp1],      (%[left])                    \n\t"
+      "lb         %[tmp2],      1(%[left])                   \n\t"
+      "lb         %[tmp3],      2(%[left])                   \n\t"
+      "lb         %[tmp4],      3(%[left])                   \n\t"
+      "lb         %[tmp5],      4(%[left])                   \n\t"
+      "lb         %[tmp6],      5(%[left])                   \n\t"
+      "lb         %[tmp7],      6(%[left])                   \n\t"
+      "lb         %[tmp8],      7(%[left])                   \n\t"
+      "lb         %[tmp9],      8(%[left])                   \n\t"
+      "lb         %[tmp10],     9(%[left])                   \n\t"
+      "lb         %[tmp11],     10(%[left])                  \n\t"
+      "lb         %[tmp12],     11(%[left])                  \n\t"
+      "lb         %[tmp13],     12(%[left])                  \n\t"
+      "lb         %[tmp14],     13(%[left])                  \n\t"
+      "lb         %[tmp15],     14(%[left])                  \n\t"
+      "lb         %[tmp16],     15(%[left])                  \n\t"
+
+      "replv.qb   %[tmp1],      %[tmp1]                      \n\t"
+      "replv.qb   %[tmp2],      %[tmp2]                      \n\t"
+      "replv.qb   %[tmp3],      %[tmp3]                      \n\t"
+      "replv.qb   %[tmp4],      %[tmp4]                      \n\t"
+      "replv.qb   %[tmp5],      %[tmp5]                      \n\t"
+      "replv.qb   %[tmp6],      %[tmp6]                      \n\t"
+      "replv.qb   %[tmp7],      %[tmp7]                      \n\t"
+      "replv.qb   %[tmp8],      %[tmp8]                      \n\t"
+      "replv.qb   %[tmp9],      %[tmp9]                      \n\t"
+      "replv.qb   %[tmp10],     %[tmp10]                     \n\t"
+      "replv.qb   %[tmp11],     %[tmp11]                     \n\t"
+      "replv.qb   %[tmp12],     %[tmp12]                     \n\t"
+      "replv.qb   %[tmp13],     %[tmp13]                     \n\t"
+      "replv.qb   %[tmp14],     %[tmp14]                     \n\t"
+      "replv.qb   %[tmp15],     %[tmp15]                     \n\t"
+      "replv.qb   %[tmp16],     %[tmp16]                     \n\t"
+
+      "sw         %[tmp1],      (%[dst])                     \n\t"
+      "sw         %[tmp1],      4(%[dst])                    \n\t"
+      "sw         %[tmp1],      8(%[dst])                    \n\t"
+      "sw         %[tmp1],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp2],      (%[dst])                     \n\t"
+      "sw         %[tmp2],      4(%[dst])                    \n\t"
+      "sw         %[tmp2],      8(%[dst])                    \n\t"
+      "sw         %[tmp2],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp3],      (%[dst])                     \n\t"
+      "sw         %[tmp3],      4(%[dst])                    \n\t"
+      "sw         %[tmp3],      8(%[dst])                    \n\t"
+      "sw         %[tmp3],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp4],      (%[dst])                     \n\t"
+      "sw         %[tmp4],      4(%[dst])                    \n\t"
+      "sw         %[tmp4],      8(%[dst])                    \n\t"
+      "sw         %[tmp4],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp5],      (%[dst])                     \n\t"
+      "sw         %[tmp5],      4(%[dst])                    \n\t"
+      "sw         %[tmp5],      8(%[dst])                    \n\t"
+      "sw         %[tmp5],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp6],      (%[dst])                     \n\t"
+      "sw         %[tmp6],      4(%[dst])                    \n\t"
+      "sw         %[tmp6],      8(%[dst])                    \n\t"
+      "sw         %[tmp6],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp7],      (%[dst])                     \n\t"
+      "sw         %[tmp7],      4(%[dst])                    \n\t"
+      "sw         %[tmp7],      8(%[dst])                    \n\t"
+      "sw         %[tmp7],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp8],      (%[dst])                     \n\t"
+      "sw         %[tmp8],      4(%[dst])                    \n\t"
+      "sw         %[tmp8],      8(%[dst])                    \n\t"
+      "sw         %[tmp8],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp9],      (%[dst])                     \n\t"
+      "sw         %[tmp9],      4(%[dst])                    \n\t"
+      "sw         %[tmp9],      8(%[dst])                    \n\t"
+      "sw         %[tmp9],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp10],     (%[dst])                     \n\t"
+      "sw         %[tmp10],     4(%[dst])                    \n\t"
+      "sw         %[tmp10],     8(%[dst])                    \n\t"
+      "sw         %[tmp10],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp11],     (%[dst])                     \n\t"
+      "sw         %[tmp11],     4(%[dst])                    \n\t"
+      "sw         %[tmp11],     8(%[dst])                    \n\t"
+      "sw         %[tmp11],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp12],     (%[dst])                     \n\t"
+      "sw         %[tmp12],     4(%[dst])                    \n\t"
+      "sw         %[tmp12],     8(%[dst])                    \n\t"
+      "sw         %[tmp12],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp13],     (%[dst])                     \n\t"
+      "sw         %[tmp13],     4(%[dst])                    \n\t"
+      "sw         %[tmp13],     8(%[dst])                    \n\t"
+      "sw         %[tmp13],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp14],     (%[dst])                     \n\t"
+      "sw         %[tmp14],     4(%[dst])                    \n\t"
+      "sw         %[tmp14],     8(%[dst])                    \n\t"
+      "sw         %[tmp14],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp15],     (%[dst])                     \n\t"
+      "sw         %[tmp15],     4(%[dst])                    \n\t"
+      "sw         %[tmp15],     8(%[dst])                    \n\t"
+      "sw         %[tmp15],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp16],     (%[dst])                     \n\t"
+      "sw         %[tmp16],     4(%[dst])                    \n\t"
+      "sw         %[tmp16],     8(%[dst])                    \n\t"
+      "sw         %[tmp16],     12(%[dst])                   \n\t"
+
+      : [tmp1] "=&r" (tmp1),   [tmp2] "=&r" (tmp2),
+        [tmp3] "=&r" (tmp3),   [tmp4] "=&r" (tmp4),
+        [tmp5] "=&r" (tmp5),   [tmp7] "=&r" (tmp7),
+        [tmp6] "=&r" (tmp6),   [tmp8] "=&r" (tmp8),
+        [tmp9] "=&r" (tmp9),   [tmp10] "=&r" (tmp10),
+        [tmp11] "=&r" (tmp11), [tmp12] "=&r" (tmp12),
+        [tmp13] "=&r" (tmp13), [tmp14] "=&r" (tmp14),
+        [tmp15] "=&r" (tmp15), [tmp16] "=&r" (tmp16)
+      : [left] "r" (left), [dst] "r" (dst), [stride] "r" (stride)
+  );
+}
+
+void vpx_dc_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  int32_t  expected_dc;
+  int32_t  average;
+  int32_t  tmp, above1, above_l1, above_r1, left1, left_r1, left_l1;
+  int32_t  above2, left2;
+
+  __asm__ __volatile__ (
+      "lw              %[above1],           (%[above])                    \n\t"
+      "lw              %[above2],           4(%[above])                   \n\t"
+      "lw              %[left1],            (%[left])                     \n\t"
+      "lw              %[left2],            4(%[left])                    \n\t"
+
+      "preceu.ph.qbl   %[above_l1],         %[above1]                     \n\t"
+      "preceu.ph.qbr   %[above_r1],         %[above1]                     \n\t"
+      "preceu.ph.qbl   %[left_l1],          %[left1]                      \n\t"
+      "preceu.ph.qbr   %[left_r1],          %[left1]                      \n\t"
+
+      "addu.ph         %[average],          %[above_r1],     %[above_l1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[left_l1]   \n\t"
+      "addu.ph         %[average],          %[average],      %[left_r1]   \n\t"
+
+      "preceu.ph.qbl   %[above_l1],         %[above2]                     \n\t"
+      "preceu.ph.qbr   %[above_r1],         %[above2]                     \n\t"
+      "preceu.ph.qbl   %[left_l1],          %[left2]                      \n\t"
+      "preceu.ph.qbr   %[left_r1],          %[left2]                      \n\t"
+
+      "addu.ph         %[average],          %[average],      %[above_l1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[above_r1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[left_l1]   \n\t"
+      "addu.ph         %[average],          %[average],      %[left_r1]   \n\t"
+
+      "lw              %[above1],           8(%[above])                   \n\t"
+      "lw              %[above2],           12(%[above])                  \n\t"
+      "lw              %[left1],            8(%[left])                    \n\t"
+      "lw              %[left2],            12(%[left])                   \n\t"
+
+      "preceu.ph.qbl   %[above_l1],         %[above1]                     \n\t"
+      "preceu.ph.qbr   %[above_r1],         %[above1]                     \n\t"
+      "preceu.ph.qbl   %[left_l1],          %[left1]                      \n\t"
+      "preceu.ph.qbr   %[left_r1],          %[left1]                      \n\t"
+
+      "addu.ph         %[average],          %[average],      %[above_l1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[above_r1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[left_l1]   \n\t"
+      "addu.ph         %[average],          %[average],      %[left_r1]   \n\t"
+
+      "preceu.ph.qbl   %[above_l1],         %[above2]                     \n\t"
+      "preceu.ph.qbr   %[above_r1],         %[above2]                     \n\t"
+      "preceu.ph.qbl   %[left_l1],          %[left2]                      \n\t"
+      "preceu.ph.qbr   %[left_r1],          %[left2]                      \n\t"
+
+      "addu.ph         %[average],          %[average],      %[above_l1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[above_r1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[left_l1]   \n\t"
+      "addu.ph         %[average],          %[average],      %[left_r1]   \n\t"
+
+      "addiu           %[average],          %[average],      16           \n\t"
+      "srl             %[tmp],              %[average],      16           \n\t"
+      "addu.ph         %[average],          %[tmp],          %[average]   \n\t"
+      "srl             %[expected_dc],      %[average],      5            \n\t"
+      "replv.qb        %[expected_dc],      %[expected_dc]                \n\t"
+
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      : [left1] "=&r" (left1), [above1] "=&r" (above1),
+        [left_l1] "=&r" (left_l1), [above_l1] "=&r" (above_l1),
+        [left_r1] "=&r" (left_r1), [above_r1] "=&r" (above_r1),
+        [above2] "=&r" (above2), [left2] "=&r" (left2),
+        [average] "=&r" (average), [tmp] "=&r" (tmp),
+        [expected_dc] "=&r" (expected_dc)
+      : [above] "r" (above), [left] "r" (left),
+        [dst] "r" (dst), [stride] "r" (stride)
+  );
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/intrapred4_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/intrapred4_dspr2.c
new file mode 100644
index 0000000..03baf4c
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/intrapred4_dspr2.c
@@ -0,0 +1,229 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/common_dspr2.h"
+
+#if HAVE_DSPR2
+void vpx_h_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int32_t  tmp1, tmp2, tmp3, tmp4;
+
+  __asm__ __volatile__ (
+      "lb         %[tmp1],      (%[left])                    \n\t"
+      "lb         %[tmp2],      1(%[left])                   \n\t"
+      "lb         %[tmp3],      2(%[left])                   \n\t"
+      "lb         %[tmp4],      3(%[left])                   \n\t"
+      "replv.qb   %[tmp1],      %[tmp1]                      \n\t"
+      "replv.qb   %[tmp2],      %[tmp2]                      \n\t"
+      "replv.qb   %[tmp3],      %[tmp3]                      \n\t"
+      "replv.qb   %[tmp4],      %[tmp4]                      \n\t"
+      "sw         %[tmp1],      (%[dst])                     \n\t"
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp2],      (%[dst])                     \n\t"
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp3],      (%[dst])                     \n\t"
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp4],      (%[dst])                     \n\t"
+
+      : [tmp1] "=&r" (tmp1), [tmp2] "=&r" (tmp2),
+        [tmp3] "=&r" (tmp3), [tmp4] "=&r" (tmp4)
+      : [left] "r" (left), [dst] "r" (dst), [stride] "r" (stride)
+  );
+}
+
+void vpx_dc_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int32_t  expected_dc;
+  int32_t  average;
+  int32_t  tmp, above_c, above_l, above_r, left_c, left_r, left_l;
+
+  __asm__ __volatile__ (
+      "lw              %[above_c],         (%[above])                    \n\t"
+      "lw              %[left_c],          (%[left])                     \n\t"
+
+      "preceu.ph.qbl   %[above_l],         %[above_c]                    \n\t"
+      "preceu.ph.qbr   %[above_r],         %[above_c]                    \n\t"
+      "preceu.ph.qbl   %[left_l],          %[left_c]                     \n\t"
+      "preceu.ph.qbr   %[left_r],          %[left_c]                     \n\t"
+
+      "addu.ph         %[average],         %[above_r],       %[above_l]  \n\t"
+      "addu.ph         %[average],         %[average],       %[left_l]   \n\t"
+      "addu.ph         %[average],         %[average],       %[left_r]   \n\t"
+      "addiu           %[average],         %[average],       4           \n\t"
+      "srl             %[tmp],             %[average],       16          \n\t"
+      "addu.ph         %[average],         %[tmp],           %[average]  \n\t"
+      "srl             %[expected_dc],     %[average],       3           \n\t"
+      "replv.qb        %[expected_dc],     %[expected_dc]                \n\t"
+
+      "sw              %[expected_dc],     (%[dst])                      \n\t"
+      "add             %[dst],              %[dst],          %[stride]   \n\t"
+      "sw              %[expected_dc],     (%[dst])                      \n\t"
+      "add             %[dst],              %[dst],          %[stride]   \n\t"
+      "sw              %[expected_dc],     (%[dst])                      \n\t"
+      "add             %[dst],              %[dst],          %[stride]   \n\t"
+      "sw              %[expected_dc],     (%[dst])                      \n\t"
+
+      : [above_c] "=&r" (above_c), [above_l] "=&r" (above_l),
+        [above_r] "=&r" (above_r), [left_c] "=&r" (left_c),
+        [left_l] "=&r" (left_l), [left_r] "=&r" (left_r),
+        [average] "=&r" (average), [tmp] "=&r" (tmp),
+        [expected_dc] "=&r" (expected_dc)
+      : [above] "r" (above), [left] "r" (left),
+        [dst] "r" (dst), [stride] "r" (stride)
+  );
+}
+
+void vpx_tm_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int32_t  abovel, abover;
+  int32_t  left0, left1, left2, left3;
+  int32_t  res0, res1;
+  int32_t  resl;
+  int32_t  resr;
+  int32_t  top_left;
+  uint8_t  *cm = vpx_ff_cropTbl;
+
+  __asm__ __volatile__ (
+      "ulw             %[resl],       (%[above])                         \n\t"
+
+      "lbu             %[left0],       (%[left])                         \n\t"
+      "lbu             %[left1],       1(%[left])                        \n\t"
+      "lbu             %[left2],       2(%[left])                        \n\t"
+      "lbu             %[left3],       3(%[left])                        \n\t"
+
+      "lbu             %[top_left],    -1(%[above])                      \n\t"
+
+      "preceu.ph.qbl   %[abovel],      %[resl]                           \n\t"
+      "preceu.ph.qbr   %[abover],      %[resl]                           \n\t"
+
+      "replv.ph        %[left0],       %[left0]                          \n\t"
+      "replv.ph        %[left1],       %[left1]                          \n\t"
+      "replv.ph        %[left2],       %[left2]                          \n\t"
+      "replv.ph        %[left3],       %[left3]                          \n\t"
+
+      "replv.ph        %[top_left],    %[top_left]                       \n\t"
+
+      "addu.ph         %[resl],        %[abovel],         %[left0]       \n\t"
+      "subu.ph         %[resl],        %[resl],           %[top_left]    \n\t"
+
+      "addu.ph         %[resr],        %[abover],         %[left0]       \n\t"
+      "subu.ph         %[resr],        %[resr],           %[top_left]    \n\t"
+
+      "sll             %[res0],        %[resr],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+      "sra             %[res1],        %[resr],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+      "sb              %[res0],        (%[dst])                          \n\t"
+
+      "sll             %[res0],        %[resl],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+      "sb              %[res1],        1(%[dst])                         \n\t"
+
+      "sra             %[res1],        %[resl],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+
+      "addu.ph         %[resl],        %[abovel],         %[left1]       \n\t"
+      "subu.ph         %[resl],        %[resl],           %[top_left]    \n\t"
+
+      "addu.ph         %[resr],        %[abover],         %[left1]       \n\t"
+      "subu.ph         %[resr],        %[resr],           %[top_left]    \n\t"
+
+      "sb              %[res0],        2(%[dst])                         \n\t"
+      "sb              %[res1],        3(%[dst])                         \n\t"
+
+      "add             %[dst],          %[dst],           %[stride]      \n\t"
+
+      "sll             %[res0],        %[resr],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+      "sra             %[res1],        %[resr],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+      "sb              %[res0],        (%[dst])                          \n\t"
+
+      "sll             %[res0],        %[resl],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+      "sb              %[res1],        1(%[dst])                         \n\t"
+      "sra             %[res1],        %[resl],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+
+      "addu.ph         %[resl],        %[abovel],         %[left2]       \n\t"
+      "subu.ph         %[resl],        %[resl],           %[top_left]    \n\t"
+
+      "addu.ph         %[resr],        %[abover],         %[left2]       \n\t"
+      "subu.ph         %[resr],        %[resr],           %[top_left]    \n\t"
+
+      "sb              %[res0],        2(%[dst])                         \n\t"
+      "sb              %[res1],        3(%[dst])                         \n\t"
+
+      "add             %[dst],          %[dst],           %[stride]      \n\t"
+
+      "sll             %[res0],        %[resr],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+
+      "sra             %[res1],        %[resr],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+      "sb              %[res0],        (%[dst])                          \n\t"
+
+      "sll             %[res0],        %[resl],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+
+      "sb              %[res1],        1(%[dst])                         \n\t"
+      "sra             %[res1],        %[resl],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+
+      "addu.ph         %[resl],        %[abovel],        %[left3]        \n\t"
+      "subu.ph         %[resl],        %[resl],          %[top_left]     \n\t"
+
+      "addu.ph         %[resr],        %[abover],        %[left3]        \n\t"
+      "subu.ph         %[resr],        %[resr],          %[top_left]     \n\t"
+
+      "sb              %[res0],        2(%[dst])                         \n\t"
+      "sb              %[res1],        3(%[dst])                         \n\t"
+
+      "add             %[dst],          %[dst],          %[stride]       \n\t"
+
+      "sll             %[res0],        %[resr],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+      "sra             %[res1],        %[resr],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+      "sb              %[res0],        (%[dst])                          \n\t"
+
+      "sll             %[res0],        %[resl],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+      "sb              %[res1],        1(%[dst])                         \n\t"
+
+      "sra             %[res1],        %[resl],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+
+      "sb              %[res0],        2(%[dst])                         \n\t"
+      "sb              %[res1],        3(%[dst])                         \n\t"
+
+      : [abovel] "=&r" (abovel), [abover] "=&r" (abover),
+        [left0] "=&r" (left0), [left1] "=&r" (left1), [left2] "=&r" (left2),
+        [res0] "=&r" (res0), [res1] "=&r" (res1), [left3] "=&r" (left3),
+        [resl] "=&r" (resl), [resr] "=&r" (resr), [top_left] "=&r" (top_left)
+      : [above] "r" (above), [left] "r" (left),
+        [dst] "r" (dst), [stride] "r" (stride), [cm] "r" (cm)
+  );
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/intrapred8_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/intrapred8_dspr2.c
new file mode 100644
index 0000000..196ff5a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/intrapred8_dspr2.c
@@ -0,0 +1,607 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/common_dspr2.h"
+
+#if HAVE_DSPR2
+void vpx_h_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int32_t  tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
+
+  __asm__ __volatile__ (
+      "lb         %[tmp1],      (%[left])                   \n\t"
+      "lb         %[tmp2],      1(%[left])                  \n\t"
+      "lb         %[tmp3],      2(%[left])                  \n\t"
+      "lb         %[tmp4],      3(%[left])                  \n\t"
+      "lb         %[tmp5],      4(%[left])                  \n\t"
+      "lb         %[tmp6],      5(%[left])                  \n\t"
+      "lb         %[tmp7],      6(%[left])                  \n\t"
+      "lb         %[tmp8],      7(%[left])                  \n\t"
+
+      "replv.qb   %[tmp1],      %[tmp1]                     \n\t"
+      "replv.qb   %[tmp2],      %[tmp2]                     \n\t"
+      "replv.qb   %[tmp3],      %[tmp3]                     \n\t"
+      "replv.qb   %[tmp4],      %[tmp4]                     \n\t"
+      "replv.qb   %[tmp5],      %[tmp5]                     \n\t"
+      "replv.qb   %[tmp6],      %[tmp6]                     \n\t"
+      "replv.qb   %[tmp7],      %[tmp7]                     \n\t"
+      "replv.qb   %[tmp8],      %[tmp8]                     \n\t"
+
+      "sw         %[tmp1],      (%[dst])                    \n\t"
+      "sw         %[tmp1],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp2],      (%[dst])                    \n\t"
+      "sw         %[tmp2],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp3],      (%[dst])                    \n\t"
+      "sw         %[tmp3],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp4],      (%[dst])                    \n\t"
+      "sw         %[tmp4],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp5],      (%[dst])                    \n\t"
+      "sw         %[tmp5],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp6],      (%[dst])                    \n\t"
+      "sw         %[tmp6],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp7],      (%[dst])                    \n\t"
+      "sw         %[tmp7],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp8],      (%[dst])                    \n\t"
+      "sw         %[tmp8],      4(%[dst])                   \n\t"
+
+      : [tmp1] "=&r" (tmp1), [tmp2] "=&r" (tmp2),
+        [tmp3] "=&r" (tmp3), [tmp4] "=&r" (tmp4),
+        [tmp5] "=&r" (tmp5), [tmp7] "=&r" (tmp7),
+        [tmp6] "=&r" (tmp6), [tmp8] "=&r" (tmp8)
+      : [left] "r" (left), [dst] "r" (dst),
+        [stride] "r" (stride)
+  );
+}
+
+void vpx_dc_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int32_t  expected_dc;
+  int32_t  average;
+  int32_t  tmp, above1, above_l1, above_r1, left1, left_r1, left_l1;
+  int32_t  above2, above_l2, above_r2, left2, left_r2, left_l2;
+
+  __asm__ __volatile__ (
+      "lw              %[above1],         (%[above])                      \n\t"
+      "lw              %[above2],         4(%[above])                     \n\t"
+      "lw              %[left1],          (%[left])                       \n\t"
+      "lw              %[left2],          4(%[left])                      \n\t"
+
+      "preceu.ph.qbl   %[above_l1],       %[above1]                       \n\t"
+      "preceu.ph.qbr   %[above_r1],       %[above1]                       \n\t"
+      "preceu.ph.qbl   %[left_l1],        %[left1]                        \n\t"
+      "preceu.ph.qbr   %[left_r1],        %[left1]                        \n\t"
+
+      "preceu.ph.qbl   %[above_l2],       %[above2]                       \n\t"
+      "preceu.ph.qbr   %[above_r2],       %[above2]                       \n\t"
+      "preceu.ph.qbl   %[left_l2],        %[left2]                        \n\t"
+      "preceu.ph.qbr   %[left_r2],        %[left2]                        \n\t"
+
+      "addu.ph         %[average],        %[above_r1],      %[above_l1]   \n\t"
+      "addu.ph         %[average],        %[average],       %[left_l1]    \n\t"
+      "addu.ph         %[average],        %[average],       %[left_r1]    \n\t"
+
+      "addu.ph         %[average],        %[average],       %[above_l2]   \n\t"
+      "addu.ph         %[average],        %[average],       %[above_r2]   \n\t"
+      "addu.ph         %[average],        %[average],       %[left_l2]    \n\t"
+      "addu.ph         %[average],        %[average],       %[left_r2]    \n\t"
+
+      "addiu           %[average],        %[average],       8             \n\t"
+
+      "srl             %[tmp],            %[average],       16            \n\t"
+      "addu.ph         %[average],        %[tmp],           %[average]    \n\t"
+      "srl             %[expected_dc],    %[average],       4             \n\t"
+      "replv.qb        %[expected_dc],    %[expected_dc]                  \n\t"
+
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      : [above1] "=&r" (above1), [above_l1] "=&r" (above_l1),
+        [above_r1] "=&r" (above_r1), [left1] "=&r" (left1),
+        [left_l1] "=&r" (left_l1), [left_r1] "=&r" (left_r1),
+        [above2] "=&r" (above2), [above_l2] "=&r" (above_l2),
+        [above_r2] "=&r" (above_r2), [left2] "=&r" (left2),
+        [left_l2] "=&r" (left_l2), [left_r2] "=&r" (left_r2),
+        [average] "=&r" (average), [tmp] "=&r" (tmp),
+        [expected_dc] "=&r" (expected_dc)
+      : [above] "r" (above), [left] "r" (left), [dst] "r" (dst),
+        [stride] "r" (stride)
+  );
+}
+
+void vpx_tm_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int32_t   abovel, abover;
+  int32_t   abovel_1, abover_1;
+  int32_t   left0;
+  int32_t   res0, res1, res2, res3;
+  int32_t   reshw;
+  int32_t   top_left;
+  uint8_t   *cm = vpx_ff_cropTbl;
+
+  __asm__ __volatile__ (
+      "ulw             %[reshw],       (%[above])                         \n\t"
+      "ulw             %[top_left],    4(%[above])                        \n\t"
+
+      "lbu             %[left0],       (%[left])                          \n\t"
+
+      "preceu.ph.qbl   %[abovel],      %[reshw]                           \n\t"
+      "preceu.ph.qbr   %[abover],      %[reshw]                           \n\t"
+      "preceu.ph.qbl   %[abovel_1],    %[top_left]                        \n\t"
+      "preceu.ph.qbr   %[abover_1],    %[top_left]                        \n\t"
+
+      "lbu             %[top_left],    -1(%[above])                       \n\t"
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+
+      "replv.ph        %[top_left],    %[top_left]                        \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       1(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       2(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       3(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       4(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       5(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       6(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       7(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      : [abovel] "=&r" (abovel), [abover] "=&r" (abover),
+        [abovel_1] "=&r" (abovel_1), [abover_1] "=&r" (abover_1),
+        [left0] "=&r" (left0), [res2] "=&r" (res2), [res3] "=&r" (res3),
+        [res0] "=&r" (res0), [res1] "=&r" (res1),
+        [reshw] "=&r" (reshw), [top_left] "=&r" (top_left)
+      : [above] "r" (above), [left] "r" (left),
+        [dst] "r" (dst), [stride] "r" (stride), [cm] "r" (cm)
+  );
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/intrapred_msa.c b/libvpx/libvpx/vpx_dsp/mips/intrapred_msa.c
new file mode 100644
index 0000000..f6fbe40
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/intrapred_msa.c
@@ -0,0 +1,737 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/macros_msa.h"
+
+#define IPRED_SUBS_UH2_UH(in0, in1, out0, out1) {  \
+  out0 = __msa_subs_u_h(out0, in0);                \
+  out1 = __msa_subs_u_h(out1, in1);                \
+}
+
+static void intra_predict_vert_4x4_msa(const uint8_t *src, uint8_t *dst,
+                                       int32_t dst_stride) {
+  uint32_t src_data;
+
+  src_data = LW(src);
+
+  SW4(src_data, src_data, src_data, src_data, dst, dst_stride);
+}
+
+static void intra_predict_vert_8x8_msa(const uint8_t *src, uint8_t *dst,
+                                       int32_t dst_stride) {
+  uint32_t row;
+  uint32_t src_data1, src_data2;
+
+  src_data1 = LW(src);
+  src_data2 = LW(src + 4);
+
+  for (row = 8; row--;) {
+    SW(src_data1, dst);
+    SW(src_data2, (dst + 4));
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_vert_16x16_msa(const uint8_t *src, uint8_t *dst,
+                                         int32_t dst_stride) {
+  uint32_t row;
+  v16u8 src0;
+
+  src0 = LD_UB(src);
+
+  for (row = 16; row--;) {
+    ST_UB(src0, dst);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_vert_32x32_msa(const uint8_t *src, uint8_t *dst,
+                                         int32_t dst_stride) {
+  uint32_t row;
+  v16u8 src1, src2;
+
+  src1 = LD_UB(src);
+  src2 = LD_UB(src + 16);
+
+  for (row = 32; row--;) {
+    ST_UB2(src1, src2, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_horiz_4x4_msa(const uint8_t *src, uint8_t *dst,
+                                        int32_t dst_stride) {
+  uint32_t out0, out1, out2, out3;
+
+  out0 = src[0] * 0x01010101;
+  out1 = src[1] * 0x01010101;
+  out2 = src[2] * 0x01010101;
+  out3 = src[3] * 0x01010101;
+
+  SW4(out0, out1, out2, out3, dst, dst_stride);
+}
+
+static void intra_predict_horiz_8x8_msa(const uint8_t *src, uint8_t *dst,
+                                        int32_t dst_stride) {
+  uint64_t out0, out1, out2, out3, out4, out5, out6, out7;
+
+  out0 = src[0] * 0x0101010101010101ull;
+  out1 = src[1] * 0x0101010101010101ull;
+  out2 = src[2] * 0x0101010101010101ull;
+  out3 = src[3] * 0x0101010101010101ull;
+  out4 = src[4] * 0x0101010101010101ull;
+  out5 = src[5] * 0x0101010101010101ull;
+  out6 = src[6] * 0x0101010101010101ull;
+  out7 = src[7] * 0x0101010101010101ull;
+
+  SD4(out0, out1, out2, out3, dst, dst_stride);
+  dst += (4 * dst_stride);
+  SD4(out4, out5, out6, out7, dst, dst_stride);
+}
+
+static void intra_predict_horiz_16x16_msa(const uint8_t *src, uint8_t *dst,
+                                          int32_t dst_stride) {
+  uint32_t row;
+  uint8_t inp0, inp1, inp2, inp3;
+  v16u8 src0, src1, src2, src3;
+
+  for (row = 4; row--;) {
+    inp0 = src[0];
+    inp1 = src[1];
+    inp2 = src[2];
+    inp3 = src[3];
+    src += 4;
+
+    src0 = (v16u8)__msa_fill_b(inp0);
+    src1 = (v16u8)__msa_fill_b(inp1);
+    src2 = (v16u8)__msa_fill_b(inp2);
+    src3 = (v16u8)__msa_fill_b(inp3);
+
+    ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void intra_predict_horiz_32x32_msa(const uint8_t *src, uint8_t *dst,
+                                          int32_t dst_stride) {
+  uint32_t row;
+  uint8_t inp0, inp1, inp2, inp3;
+  v16u8 src0, src1, src2, src3;
+
+  for (row = 8; row--;) {
+    inp0 = src[0];
+    inp1 = src[1];
+    inp2 = src[2];
+    inp3 = src[3];
+    src += 4;
+
+    src0 = (v16u8)__msa_fill_b(inp0);
+    src1 = (v16u8)__msa_fill_b(inp1);
+    src2 = (v16u8)__msa_fill_b(inp2);
+    src3 = (v16u8)__msa_fill_b(inp3);
+
+    ST_UB2(src0, src0, dst, 16);
+    dst += dst_stride;
+    ST_UB2(src1, src1, dst, 16);
+    dst += dst_stride;
+    ST_UB2(src2, src2, dst, 16);
+    dst += dst_stride;
+    ST_UB2(src3, src3, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_dc_4x4_msa(const uint8_t *src_top,
+                                     const uint8_t *src_left,
+                                     uint8_t *dst, int32_t dst_stride) {
+  uint32_t val0, val1;
+  v16i8 store, src = { 0 };
+  v8u16 sum_h;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  val0 = LW(src_top);
+  val1 = LW(src_left);
+  INSERT_W2_SB(val0, val1, src);
+  sum_h = __msa_hadd_u_h((v16u8)src, (v16u8)src);
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 3);
+  store = __msa_splati_b((v16i8)sum_w, 0);
+  val0 = __msa_copy_u_w((v4i32)store, 0);
+
+  SW4(val0, val0, val0, val0, dst, dst_stride);
+}
+
+static void intra_predict_dc_tl_4x4_msa(const uint8_t *src, uint8_t *dst,
+                                        int32_t dst_stride) {
+  uint32_t val0;
+  v16i8 store, data = { 0 };
+  v8u16 sum_h;
+  v4u32 sum_w;
+
+  val0 = LW(src);
+  data = (v16i8)__msa_insert_w((v4i32)data, 0, val0);
+  sum_h = __msa_hadd_u_h((v16u8)data, (v16u8)data);
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_w, 2);
+  store = __msa_splati_b((v16i8)sum_w, 0);
+  val0 = __msa_copy_u_w((v4i32)store, 0);
+
+  SW4(val0, val0, val0, val0, dst, dst_stride);
+}
+
+static void intra_predict_128dc_4x4_msa(uint8_t *dst, int32_t dst_stride) {
+  uint32_t out;
+  const v16i8 store = __msa_ldi_b(128);
+
+  out = __msa_copy_u_w((v4i32)store, 0);
+
+  SW4(out, out, out, out, dst, dst_stride);
+}
+
+static void intra_predict_dc_8x8_msa(const uint8_t *src_top,
+                                     const uint8_t *src_left,
+                                     uint8_t *dst, int32_t dst_stride) {
+  uint64_t val0, val1;
+  v16i8 store;
+  v16u8 src = { 0 };
+  v8u16 sum_h;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  val0 = LD(src_top);
+  val1 = LD(src_left);
+  INSERT_D2_UB(val0, val1, src);
+  sum_h = __msa_hadd_u_h(src, src);
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 4);
+  store = __msa_splati_b((v16i8)sum_w, 0);
+  val0 = __msa_copy_u_d((v2i64)store, 0);
+
+  SD4(val0, val0, val0, val0, dst, dst_stride);
+  dst += (4 * dst_stride);
+  SD4(val0, val0, val0, val0, dst, dst_stride);
+}
+
+static void intra_predict_dc_tl_8x8_msa(const uint8_t *src, uint8_t *dst,
+                                        int32_t dst_stride) {
+  uint64_t val0;
+  v16i8 store;
+  v16u8 data = { 0 };
+  v8u16 sum_h;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  val0 = LD(src);
+  data = (v16u8)__msa_insert_d((v2i64)data, 0, val0);
+  sum_h = __msa_hadd_u_h(data, data);
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 3);
+  store = __msa_splati_b((v16i8)sum_w, 0);
+  val0 = __msa_copy_u_d((v2i64)store, 0);
+
+  SD4(val0, val0, val0, val0, dst, dst_stride);
+  dst += (4 * dst_stride);
+  SD4(val0, val0, val0, val0, dst, dst_stride);
+}
+
+static void intra_predict_128dc_8x8_msa(uint8_t *dst, int32_t dst_stride) {
+  uint64_t out;
+  const v16i8 store = __msa_ldi_b(128);
+
+  out = __msa_copy_u_d((v2i64)store, 0);
+
+  SD4(out, out, out, out, dst, dst_stride);
+  dst += (4 * dst_stride);
+  SD4(out, out, out, out, dst, dst_stride);
+}
+
+static void intra_predict_dc_16x16_msa(const uint8_t *src_top,
+                                       const uint8_t *src_left,
+                                       uint8_t *dst, int32_t dst_stride) {
+  v16u8 top, left, out;
+  v8u16 sum_h, sum_top, sum_left;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  top = LD_UB(src_top);
+  left = LD_UB(src_left);
+  HADD_UB2_UH(top, left, sum_top, sum_left);
+  sum_h = sum_top + sum_left;
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 5);
+  out = (v16u8)__msa_splati_b((v16i8)sum_w, 0);
+
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+  dst += (8 * dst_stride);
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+}
+
+static void intra_predict_dc_tl_16x16_msa(const uint8_t *src, uint8_t *dst,
+                                          int32_t dst_stride) {
+  v16u8 data, out;
+  v8u16 sum_h;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  data = LD_UB(src);
+  sum_h = __msa_hadd_u_h(data, data);
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 4);
+  out = (v16u8)__msa_splati_b((v16i8)sum_w, 0);
+
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+  dst += (8 * dst_stride);
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+}
+
+static void intra_predict_128dc_16x16_msa(uint8_t *dst, int32_t dst_stride) {
+  const v16u8 out = (v16u8)__msa_ldi_b(128);
+
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+  dst += (8 * dst_stride);
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+}
+
+static void intra_predict_dc_32x32_msa(const uint8_t *src_top,
+                                       const uint8_t *src_left,
+                                       uint8_t *dst, int32_t dst_stride) {
+  uint32_t row;
+  v16u8 top0, top1, left0, left1, out;
+  v8u16 sum_h, sum_top0, sum_top1, sum_left0, sum_left1;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  LD_UB2(src_top, 16, top0, top1);
+  LD_UB2(src_left, 16, left0, left1);
+  HADD_UB2_UH(top0, top1, sum_top0, sum_top1);
+  HADD_UB2_UH(left0, left1, sum_left0, sum_left1);
+  sum_h = sum_top0 + sum_top1;
+  sum_h += sum_left0 + sum_left1;
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 6);
+  out = (v16u8)__msa_splati_b((v16i8)sum_w, 0);
+
+  for (row = 16; row--;) {
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_dc_tl_32x32_msa(const uint8_t *src, uint8_t *dst,
+                                          int32_t dst_stride) {
+  uint32_t row;
+  v16u8 data0, data1, out;
+  v8u16 sum_h, sum_data0, sum_data1;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  LD_UB2(src, 16, data0, data1);
+  HADD_UB2_UH(data0, data1, sum_data0, sum_data1);
+  sum_h = sum_data0 + sum_data1;
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 5);
+  out = (v16u8)__msa_splati_b((v16i8)sum_w, 0);
+
+  for (row = 16; row--;) {
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_128dc_32x32_msa(uint8_t *dst, int32_t dst_stride) {
+  uint32_t row;
+  const v16u8 out = (v16u8)__msa_ldi_b(128);
+
+  for (row = 16; row--;) {
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_tm_4x4_msa(const uint8_t *src_top_ptr,
+                                     const uint8_t *src_left,
+                                     uint8_t *dst, int32_t dst_stride) {
+  uint32_t val;
+  uint8_t top_left = src_top_ptr[-1];
+  v16i8 src_left0, src_left1, src_left2, src_left3, tmp0, tmp1, src_top = { 0 };
+  v16u8 src0, src1, src2, src3;
+  v8u16 src_top_left, vec0, vec1, vec2, vec3;
+
+  src_top_left = (v8u16)__msa_fill_h(top_left);
+  val = LW(src_top_ptr);
+  src_top = (v16i8)__msa_insert_w((v4i32)src_top, 0, val);
+
+  src_left0 = __msa_fill_b(src_left[0]);
+  src_left1 = __msa_fill_b(src_left[1]);
+  src_left2 = __msa_fill_b(src_left[2]);
+  src_left3 = __msa_fill_b(src_left[3]);
+
+  ILVR_B4_UB(src_left0, src_top, src_left1, src_top, src_left2, src_top,
+             src_left3, src_top, src0, src1, src2, src3);
+  HADD_UB4_UH(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+  IPRED_SUBS_UH2_UH(src_top_left, src_top_left, vec0, vec1);
+  IPRED_SUBS_UH2_UH(src_top_left, src_top_left, vec2, vec3);
+  SAT_UH4_UH(vec0, vec1, vec2, vec3, 7);
+  PCKEV_B2_SB(vec1, vec0, vec3, vec2, tmp0, tmp1);
+  ST4x4_UB(tmp0, tmp1, 0, 2, 0, 2, dst, dst_stride);
+}
+
+static void intra_predict_tm_8x8_msa(const uint8_t *src_top_ptr,
+                                     const uint8_t *src_left,
+                                     uint8_t *dst, int32_t dst_stride) {
+  uint64_t val;
+  uint8_t top_left = src_top_ptr[-1];
+  uint32_t loop_cnt;
+  v16i8 src_left0, src_left1, src_left2, src_left3, tmp0, tmp1, src_top = { 0 };
+  v8u16 src_top_left, vec0, vec1, vec2, vec3;
+  v16u8 src0, src1, src2, src3;
+
+  val = LD(src_top_ptr);
+  src_top = (v16i8)__msa_insert_d((v2i64)src_top, 0, val);
+  src_top_left = (v8u16)__msa_fill_h(top_left);
+
+  for (loop_cnt = 2; loop_cnt--;) {
+    src_left0 = __msa_fill_b(src_left[0]);
+    src_left1 = __msa_fill_b(src_left[1]);
+    src_left2 = __msa_fill_b(src_left[2]);
+    src_left3 = __msa_fill_b(src_left[3]);
+    src_left += 4;
+
+    ILVR_B4_UB(src_left0, src_top, src_left1, src_top, src_left2, src_top,
+               src_left3, src_top, src0, src1, src2, src3);
+    HADD_UB4_UH(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, vec0, vec1);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, vec2, vec3);
+    SAT_UH4_UH(vec0, vec1, vec2, vec3, 7);
+    PCKEV_B2_SB(vec1, vec0, vec3, vec2, tmp0, tmp1);
+    ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void intra_predict_tm_16x16_msa(const uint8_t *src_top_ptr,
+                                       const uint8_t *src_left,
+                                       uint8_t *dst, int32_t dst_stride) {
+  uint8_t top_left = src_top_ptr[-1];
+  uint32_t loop_cnt;
+  v16i8 src_top, src_left0, src_left1, src_left2, src_left3;
+  v8u16 src_top_left, res_r, res_l;
+
+  src_top = LD_SB(src_top_ptr);
+  src_top_left = (v8u16)__msa_fill_h(top_left);
+
+  for (loop_cnt = 4; loop_cnt--;) {
+    src_left0 = __msa_fill_b(src_left[0]);
+    src_left1 = __msa_fill_b(src_left[1]);
+    src_left2 = __msa_fill_b(src_left[2]);
+    src_left3 = __msa_fill_b(src_left[3]);
+    src_left += 4;
+
+    ILVRL_B2_UH(src_left0, src_top, res_r, res_l);
+    HADD_UB2_UH(res_r, res_l, res_r, res_l);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r, res_l);
+
+    SAT_UH2_UH(res_r, res_l, 7);
+    PCKEV_ST_SB(res_r, res_l, dst);
+    dst += dst_stride;
+
+    ILVRL_B2_UH(src_left1, src_top, res_r, res_l);
+    HADD_UB2_UH(res_r, res_l, res_r, res_l);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r, res_l);
+    SAT_UH2_UH(res_r, res_l, 7);
+    PCKEV_ST_SB(res_r, res_l, dst);
+    dst += dst_stride;
+
+    ILVRL_B2_UH(src_left2, src_top, res_r, res_l);
+    HADD_UB2_UH(res_r, res_l, res_r, res_l);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r, res_l);
+    SAT_UH2_UH(res_r, res_l, 7);
+    PCKEV_ST_SB(res_r, res_l, dst);
+    dst += dst_stride;
+
+    ILVRL_B2_UH(src_left3, src_top, res_r, res_l);
+    HADD_UB2_UH(res_r, res_l, res_r, res_l);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r, res_l);
+    SAT_UH2_UH(res_r, res_l, 7);
+    PCKEV_ST_SB(res_r, res_l, dst);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_tm_32x32_msa(const uint8_t *src_top,
+                                       const uint8_t *src_left,
+                                       uint8_t *dst, int32_t dst_stride) {
+  uint8_t top_left = src_top[-1];
+  uint32_t loop_cnt;
+  v16i8 src_top0, src_top1, src_left0, src_left1, src_left2, src_left3;
+  v8u16 src_top_left, res_r0, res_r1, res_l0, res_l1;
+
+  LD_SB2(src_top, 16, src_top0, src_top1);
+  src_top_left = (v8u16)__msa_fill_h(top_left);
+
+  for (loop_cnt = 8; loop_cnt--;) {
+    src_left0 = __msa_fill_b(src_left[0]);
+    src_left1 = __msa_fill_b(src_left[1]);
+    src_left2 = __msa_fill_b(src_left[2]);
+    src_left3 = __msa_fill_b(src_left[3]);
+    src_left += 4;
+
+    ILVR_B2_UH(src_left0, src_top0, src_left0, src_top1, res_r0, res_r1);
+    ILVL_B2_UH(src_left0, src_top0, src_left0, src_top1, res_l0, res_l1);
+    HADD_UB4_UH(res_r0, res_l0, res_r1, res_l1, res_r0, res_l0, res_r1, res_l1);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r0, res_l0);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r1, res_l1);
+    SAT_UH4_UH(res_r0, res_l0, res_r1, res_l1, 7);
+    PCKEV_ST_SB(res_r0, res_l0, dst);
+    PCKEV_ST_SB(res_r1, res_l1, dst + 16);
+    dst += dst_stride;
+
+    ILVR_B2_UH(src_left1, src_top0, src_left1, src_top1, res_r0, res_r1);
+    ILVL_B2_UH(src_left1, src_top0, src_left1, src_top1, res_l0, res_l1);
+    HADD_UB4_UH(res_r0, res_l0, res_r1, res_l1, res_r0, res_l0, res_r1, res_l1);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r0, res_l0);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r1, res_l1);
+    SAT_UH4_UH(res_r0, res_l0, res_r1, res_l1, 7);
+    PCKEV_ST_SB(res_r0, res_l0, dst);
+    PCKEV_ST_SB(res_r1, res_l1, dst + 16);
+    dst += dst_stride;
+
+    ILVR_B2_UH(src_left2, src_top0, src_left2, src_top1, res_r0, res_r1);
+    ILVL_B2_UH(src_left2, src_top0, src_left2, src_top1, res_l0, res_l1);
+    HADD_UB4_UH(res_r0, res_l0, res_r1, res_l1, res_r0, res_l0, res_r1, res_l1);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r0, res_l0);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r1, res_l1);
+    SAT_UH4_UH(res_r0, res_l0, res_r1, res_l1, 7);
+    PCKEV_ST_SB(res_r0, res_l0, dst);
+    PCKEV_ST_SB(res_r1, res_l1, dst + 16);
+    dst += dst_stride;
+
+    ILVR_B2_UH(src_left3, src_top0, src_left3, src_top1, res_r0, res_r1);
+    ILVL_B2_UH(src_left3, src_top0, src_left3, src_top1, res_l0, res_l1);
+    HADD_UB4_UH(res_r0, res_l0, res_r1, res_l1, res_r0, res_l0, res_r1, res_l1);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r0, res_l0);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r1, res_l1);
+    SAT_UH4_UH(res_r0, res_l0, res_r1, res_l1, 7);
+    PCKEV_ST_SB(res_r0, res_l0, dst);
+    PCKEV_ST_SB(res_r1, res_l1, dst + 16);
+    dst += dst_stride;
+  }
+}
+
+void vpx_v_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                             const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_vert_4x4_msa(above, dst, y_stride);
+}
+
+void vpx_v_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                             const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_vert_8x8_msa(above, dst, y_stride);
+}
+
+void vpx_v_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_vert_16x16_msa(above, dst, y_stride);
+}
+
+void vpx_v_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_vert_32x32_msa(above, dst, y_stride);
+}
+
+void vpx_h_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                             const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_horiz_4x4_msa(left, dst, y_stride);
+}
+
+void vpx_h_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                             const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_horiz_8x8_msa(left, dst, y_stride);
+}
+
+void vpx_h_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_horiz_16x16_msa(left, dst, y_stride);
+}
+
+void vpx_h_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_horiz_32x32_msa(left, dst, y_stride);
+}
+
+void vpx_dc_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                              const uint8_t *above, const uint8_t *left) {
+  intra_predict_dc_4x4_msa(above, left, dst, y_stride);
+}
+
+void vpx_dc_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                              const uint8_t *above, const uint8_t *left) {
+  intra_predict_dc_8x8_msa(above, left, dst, y_stride);
+}
+
+void vpx_dc_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                const uint8_t *above, const uint8_t *left) {
+  intra_predict_dc_16x16_msa(above, left, dst, y_stride);
+}
+
+void vpx_dc_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                const uint8_t *above, const uint8_t *left) {
+  intra_predict_dc_32x32_msa(above, left, dst, y_stride);
+}
+
+void vpx_dc_top_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_dc_tl_4x4_msa(above, dst, y_stride);
+}
+
+void vpx_dc_top_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_dc_tl_8x8_msa(above, dst, y_stride);
+}
+
+void vpx_dc_top_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_dc_tl_16x16_msa(above, dst, y_stride);
+}
+
+void vpx_dc_top_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_dc_tl_32x32_msa(above, dst, y_stride);
+}
+
+void vpx_dc_left_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_dc_tl_4x4_msa(left, dst, y_stride);
+}
+
+void vpx_dc_left_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_dc_tl_8x8_msa(left, dst, y_stride);
+}
+
+void vpx_dc_left_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+
+  intra_predict_dc_tl_16x16_msa(left, dst, y_stride);
+}
+
+void vpx_dc_left_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+
+  intra_predict_dc_tl_32x32_msa(left, dst, y_stride);
+}
+
+void vpx_dc_128_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+
+  intra_predict_128dc_4x4_msa(dst, y_stride);
+}
+
+void vpx_dc_128_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+
+  intra_predict_128dc_8x8_msa(dst, y_stride);
+}
+
+void vpx_dc_128_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+
+  intra_predict_128dc_16x16_msa(dst, y_stride);
+}
+
+void vpx_dc_128_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+
+  intra_predict_128dc_32x32_msa(dst, y_stride);
+}
+
+void vpx_tm_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                              const uint8_t *above, const uint8_t *left) {
+  intra_predict_tm_4x4_msa(above, left, dst, y_stride);
+}
+
+void vpx_tm_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                              const uint8_t *above, const uint8_t *left) {
+  intra_predict_tm_8x8_msa(above, left, dst, y_stride);
+}
+
+void vpx_tm_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                const uint8_t *above, const uint8_t *left) {
+  intra_predict_tm_16x16_msa(above, left, dst, y_stride);
+}
+
+void vpx_tm_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                const uint8_t *above, const uint8_t *left) {
+  intra_predict_tm_32x32_msa(above, left, dst, y_stride);
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/inv_txfm_dspr2.h b/libvpx/libvpx/vpx_dsp/mips/inv_txfm_dspr2.h
new file mode 100644
index 0000000..abd8509
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/inv_txfm_dspr2.h
@@ -0,0 +1,73 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_MIPS_INV_TXFM_DSPR2_H_
+#define VPX_DSP_MIPS_INV_TXFM_DSPR2_H_
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/inv_txfm.h"
+#include "vpx_dsp/mips/common_dspr2.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if HAVE_DSPR2
+#define DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input)                    ({   \
+                                                                               \
+  int32_t tmp, out;                                                            \
+  int     dct_cost_rounding = DCT_CONST_ROUNDING;                              \
+  int     in = input;                                                          \
+                                                                               \
+  __asm__ __volatile__ (                                                       \
+      /* out = dct_const_round_shift(input_dc * cospi_16_64); */               \
+      "mtlo     %[dct_cost_rounding],   $ac1                              \n\t"\
+      "mthi     $zero,                  $ac1                              \n\t"\
+      "madd     $ac1,                   %[in],            %[cospi_16_64]  \n\t"\
+      "extp     %[tmp],                 $ac1,             31              \n\t"\
+                                                                               \
+      /* out = dct_const_round_shift(out * cospi_16_64); */                    \
+      "mtlo     %[dct_cost_rounding],   $ac2                              \n\t"\
+      "mthi     $zero,                  $ac2                              \n\t"\
+      "madd     $ac2,                   %[tmp],           %[cospi_16_64]  \n\t"\
+      "extp     %[out],                 $ac2,             31              \n\t"\
+                                                                               \
+      : [tmp] "=&r" (tmp), [out] "=r" (out)                                    \
+      : [in] "r" (in),                                                         \
+        [dct_cost_rounding] "r" (dct_cost_rounding),                           \
+        [cospi_16_64] "r" (cospi_16_64)                                        \
+   );                                                                          \
+  out;                                                                    })
+
+void vpx_idct32_cols_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                   int dest_stride);
+void vpx_idct4_rows_dspr2(const int16_t *input, int16_t *output);
+void vpx_idct4_columns_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                     int dest_stride);
+void iadst4_dspr2(const int16_t *input, int16_t *output);
+void idct8_rows_dspr2(const int16_t *input, int16_t *output, uint32_t no_rows);
+void idct8_columns_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                 int dest_stride);
+void iadst8_dspr2(const int16_t *input, int16_t *output);
+void idct16_rows_dspr2(const int16_t *input, int16_t *output,
+                       uint32_t no_rows);
+void idct16_cols_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                               int dest_stride);
+void iadst16_dspr2(const int16_t *input, int16_t *output);
+
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_MIPS_INV_TXFM_DSPR2_H_
diff --git a/libvpx/libvpx/vpx_dsp/mips/inv_txfm_msa.h b/libvpx/libvpx/vpx_dsp/mips/inv_txfm_msa.h
new file mode 100644
index 0000000..1458561
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/inv_txfm_msa.h
@@ -0,0 +1,410 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_MIPS_INV_TXFM_MSA_H_
+#define VPX_DSP_MIPS_INV_TXFM_MSA_H_
+
+#include "vpx_dsp/mips/macros_msa.h"
+#include "vpx_dsp/mips/txfm_macros_msa.h"
+#include "vpx_dsp/txfm_common.h"
+
+#define VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,               \
+                  out0, out1, out2, out3, out4, out5, out6, out7) {     \
+  v8i16 cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst4_m;                    \
+  v8i16 vec0_m, vec1_m, vec2_m, vec3_m, s0_m, s1_m;                     \
+  v8i16 coeff0_m = { cospi_2_64, cospi_6_64, cospi_10_64, cospi_14_64,  \
+    cospi_18_64, cospi_22_64, cospi_26_64, cospi_30_64 };               \
+  v8i16 coeff1_m = { cospi_8_64, -cospi_8_64, cospi_16_64,              \
+    -cospi_16_64, cospi_24_64, -cospi_24_64, 0, 0 };                    \
+                                                                        \
+  SPLATI_H2_SH(coeff0_m, 0, 7, cnst0_m, cnst1_m);                       \
+  cnst2_m = -cnst0_m;                                                   \
+  ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);    \
+  SPLATI_H2_SH(coeff0_m, 4, 3, cnst2_m, cnst3_m);                       \
+  cnst4_m = -cnst2_m;                                                   \
+  ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);    \
+                                                                        \
+  ILVRL_H2_SH(in0, in7, vec1_m, vec0_m);                                \
+  ILVRL_H2_SH(in4, in3, vec3_m, vec2_m);                                \
+  DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m,        \
+                        cnst1_m, cnst2_m, cnst3_m, in7, in0,            \
+                        in4, in3);                                      \
+                                                                        \
+  SPLATI_H2_SH(coeff0_m, 2, 5, cnst0_m, cnst1_m);                       \
+  cnst2_m = -cnst0_m;                                                   \
+  ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);    \
+  SPLATI_H2_SH(coeff0_m, 6, 1, cnst2_m, cnst3_m);                       \
+  cnst4_m = -cnst2_m;                                                   \
+  ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);    \
+                                                                        \
+  ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                \
+  ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                \
+                                                                        \
+  DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m,        \
+                        cnst1_m, cnst2_m, cnst3_m, in5, in2,            \
+                        in6, in1);                                      \
+  BUTTERFLY_4(in7, in0, in2, in5, s1_m, s0_m, in2, in5);                \
+  out7 = -s0_m;                                                         \
+  out0 = s1_m;                                                          \
+                                                                        \
+  SPLATI_H4_SH(coeff1_m, 0, 4, 1, 5,                                    \
+               cnst0_m, cnst1_m, cnst2_m, cnst3_m);                     \
+                                                                        \
+  ILVEV_H2_SH(cnst3_m, cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst2_m);    \
+  cnst0_m = __msa_ilvev_h(cnst1_m, cnst0_m);                            \
+  cnst1_m = cnst0_m;                                                    \
+                                                                        \
+  ILVRL_H2_SH(in4, in3, vec1_m, vec0_m);                                \
+  ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                \
+  DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m,        \
+                        cnst2_m, cnst3_m, cnst1_m, out1, out6,          \
+                        s0_m, s1_m);                                    \
+                                                                        \
+  SPLATI_H2_SH(coeff1_m, 2, 3, cnst0_m, cnst1_m);                       \
+  cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                            \
+                                                                        \
+  ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                \
+  ILVRL_H2_SH(s0_m, s1_m, vec3_m, vec2_m);                              \
+  out3 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                \
+  out4 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst1_m);                \
+  out2 = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst0_m);                \
+  out5 = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst1_m);                \
+                                                                        \
+  out1 = -out1;                                                         \
+  out3 = -out3;                                                         \
+  out5 = -out5;                                                         \
+}
+
+#define VP9_SET_COSPI_PAIR(c0_h, c1_h) ({  \
+  v8i16 out0_m, r0_m, r1_m;                \
+                                           \
+  r0_m = __msa_fill_h(c0_h);               \
+  r1_m = __msa_fill_h(c1_h);               \
+  out0_m = __msa_ilvev_h(r1_m, r0_m);      \
+                                           \
+  out0_m;                                  \
+})
+
+#define VP9_ADDBLK_ST8x4_UB(dst, dst_stride, in0, in1, in2, in3) {  \
+  uint8_t *dst_m = (uint8_t *) (dst);                               \
+  v16u8 dst0_m, dst1_m, dst2_m, dst3_m;                             \
+  v16i8 tmp0_m, tmp1_m;                                             \
+  v16i8 zero_m = { 0 };                                             \
+  v8i16 res0_m, res1_m, res2_m, res3_m;                             \
+                                                                    \
+  LD_UB4(dst_m, dst_stride, dst0_m, dst1_m, dst2_m, dst3_m);        \
+  ILVR_B4_SH(zero_m, dst0_m, zero_m, dst1_m, zero_m, dst2_m,        \
+             zero_m, dst3_m, res0_m, res1_m, res2_m, res3_m);       \
+  ADD4(res0_m, in0, res1_m, in1, res2_m, in2, res3_m, in3,          \
+       res0_m, res1_m, res2_m, res3_m);                             \
+  CLIP_SH4_0_255(res0_m, res1_m, res2_m, res3_m);                   \
+  PCKEV_B2_SB(res1_m, res0_m, res3_m, res2_m, tmp0_m, tmp1_m);      \
+  ST8x4_UB(tmp0_m, tmp1_m, dst_m, dst_stride);                      \
+}
+
+#define VP9_IDCT4x4(in0, in1, in2, in3, out0, out1, out2, out3) {   \
+  v8i16 c0_m, c1_m, c2_m, c3_m;                                     \
+  v8i16 step0_m, step1_m;                                           \
+  v4i32 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                             \
+                                                                    \
+  c0_m = VP9_SET_COSPI_PAIR(cospi_16_64, cospi_16_64);              \
+  c1_m = VP9_SET_COSPI_PAIR(cospi_16_64, -cospi_16_64);             \
+  step0_m = __msa_ilvr_h(in2, in0);                                 \
+  DOTP_SH2_SW(step0_m, step0_m, c0_m, c1_m, tmp0_m, tmp1_m);        \
+                                                                    \
+  c2_m = VP9_SET_COSPI_PAIR(cospi_24_64, -cospi_8_64);              \
+  c3_m = VP9_SET_COSPI_PAIR(cospi_8_64, cospi_24_64);               \
+  step1_m = __msa_ilvr_h(in3, in1);                                 \
+  DOTP_SH2_SW(step1_m, step1_m, c2_m, c3_m, tmp2_m, tmp3_m);        \
+  SRARI_W4_SW(tmp0_m, tmp1_m, tmp2_m, tmp3_m, DCT_CONST_BITS);      \
+                                                                    \
+  PCKEV_H2_SW(tmp1_m, tmp0_m, tmp3_m, tmp2_m, tmp0_m, tmp2_m);      \
+  SLDI_B2_0_SW(tmp0_m, tmp2_m, tmp1_m, tmp3_m, 8);                  \
+  BUTTERFLY_4((v8i16)tmp0_m, (v8i16)tmp1_m,                         \
+              (v8i16)tmp2_m, (v8i16)tmp3_m,                         \
+              out0, out1, out2, out3);                              \
+}
+
+#define VP9_IADST4x4(in0, in1, in2, in3, out0, out1, out2, out3) {  \
+  v8i16 res0_m, res1_m, c0_m, c1_m;                                 \
+  v8i16 k1_m, k2_m, k3_m, k4_m;                                     \
+  v8i16 zero_m = { 0 };                                             \
+  v4i32 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                             \
+  v4i32 int0_m, int1_m, int2_m, int3_m;                             \
+  v8i16 mask_m = { sinpi_1_9, sinpi_2_9, sinpi_3_9,                 \
+    sinpi_4_9, -sinpi_1_9, -sinpi_2_9, -sinpi_3_9,                  \
+    -sinpi_4_9 };                                                   \
+                                                                    \
+  SPLATI_H4_SH(mask_m, 3, 0, 1, 2, c0_m, c1_m, k1_m, k2_m);         \
+  ILVEV_H2_SH(c0_m, c1_m, k1_m, k2_m, c0_m, c1_m);                  \
+  ILVR_H2_SH(in0, in2, in1, in3, res0_m, res1_m);                   \
+  DOTP_SH2_SW(res0_m, res1_m, c0_m, c1_m, tmp2_m, tmp1_m);          \
+  int0_m = tmp2_m + tmp1_m;                                         \
+                                                                    \
+  SPLATI_H2_SH(mask_m, 4, 7, k4_m, k3_m);                           \
+  ILVEV_H2_SH(k4_m, k1_m, k3_m, k2_m, c0_m, c1_m);                  \
+  DOTP_SH2_SW(res0_m, res1_m, c0_m, c1_m, tmp0_m, tmp1_m);          \
+  int1_m = tmp0_m + tmp1_m;                                         \
+                                                                    \
+  c0_m = __msa_splati_h(mask_m, 6);                                 \
+  ILVL_H2_SH(k2_m, c0_m, zero_m, k2_m, c0_m, c1_m);                 \
+  ILVR_H2_SH(in0, in2, in1, in3, res0_m, res1_m);                   \
+  DOTP_SH2_SW(res0_m, res1_m, c0_m, c1_m, tmp0_m, tmp1_m);          \
+  int2_m = tmp0_m + tmp1_m;                                         \
+                                                                    \
+  c0_m = __msa_splati_h(mask_m, 6);                                 \
+  c0_m = __msa_ilvev_h(c0_m, k1_m);                                 \
+                                                                    \
+  res0_m = __msa_ilvr_h((in1), (in3));                              \
+  tmp0_m = __msa_dotp_s_w(res0_m, c0_m);                            \
+  int3_m = tmp2_m + tmp0_m;                                         \
+                                                                    \
+  res0_m = __msa_ilvr_h((in2), (in3));                              \
+  c1_m = __msa_ilvev_h(k4_m, k3_m);                                 \
+                                                                    \
+  tmp2_m = __msa_dotp_s_w(res0_m, c1_m);                            \
+  res1_m = __msa_ilvr_h((in0), (in2));                              \
+  c1_m = __msa_ilvev_h(k1_m, zero_m);                               \
+                                                                    \
+  tmp3_m = __msa_dotp_s_w(res1_m, c1_m);                            \
+  int3_m += tmp2_m;                                                 \
+  int3_m += tmp3_m;                                                 \
+                                                                    \
+  SRARI_W4_SW(int0_m, int1_m, int2_m, int3_m, DCT_CONST_BITS);      \
+  PCKEV_H2_SH(int0_m, int0_m, int1_m, int1_m, out0, out1);          \
+  PCKEV_H2_SH(int2_m, int2_m, int3_m, int3_m, out2, out3);          \
+}
+
+#define VP9_SET_CONST_PAIR(mask_h, idx1_h, idx2_h) ({  \
+  v8i16 c0_m, c1_m;                                    \
+                                                       \
+  SPLATI_H2_SH(mask_h, idx1_h, idx2_h, c0_m, c1_m);    \
+  c0_m = __msa_ilvev_h(c1_m, c0_m);                    \
+                                                       \
+  c0_m;                                                \
+})
+
+/* multiply and add macro */
+#define VP9_MADD(inp0, inp1, inp2, inp3, cst0, cst1, cst2, cst3,        \
+                 out0, out1, out2, out3) {                              \
+  v8i16 madd_s0_m, madd_s1_m, madd_s2_m, madd_s3_m;                     \
+  v4i32 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                 \
+                                                                        \
+  ILVRL_H2_SH(inp1, inp0, madd_s1_m, madd_s0_m);                        \
+  ILVRL_H2_SH(inp3, inp2, madd_s3_m, madd_s2_m);                        \
+  DOTP_SH4_SW(madd_s1_m, madd_s0_m, madd_s1_m, madd_s0_m,               \
+              cst0, cst0, cst1, cst1, tmp0_m, tmp1_m, tmp2_m, tmp3_m);  \
+  SRARI_W4_SW(tmp0_m, tmp1_m, tmp2_m, tmp3_m, DCT_CONST_BITS);          \
+  PCKEV_H2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out0, out1);              \
+  DOTP_SH4_SW(madd_s3_m, madd_s2_m, madd_s3_m, madd_s2_m,               \
+              cst2, cst2, cst3, cst3, tmp0_m, tmp1_m, tmp2_m, tmp3_m);  \
+  SRARI_W4_SW(tmp0_m, tmp1_m, tmp2_m, tmp3_m, DCT_CONST_BITS);          \
+  PCKEV_H2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out2, out3);              \
+}
+
+/* idct 8x8 macro */
+#define VP9_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7,               \
+                       out0, out1, out2, out3, out4, out5, out6, out7) {     \
+  v8i16 tp0_m, tp1_m, tp2_m, tp3_m, tp4_m, tp5_m, tp6_m, tp7_m;              \
+  v8i16 k0_m, k1_m, k2_m, k3_m, res0_m, res1_m, res2_m, res3_m;              \
+  v4i32 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                      \
+  v8i16 mask_m = { cospi_28_64, cospi_4_64, cospi_20_64, cospi_12_64,        \
+    cospi_16_64, -cospi_4_64, -cospi_20_64, -cospi_16_64 };                  \
+                                                                             \
+  k0_m = VP9_SET_CONST_PAIR(mask_m, 0, 5);                                   \
+  k1_m = VP9_SET_CONST_PAIR(mask_m, 1, 0);                                   \
+  k2_m = VP9_SET_CONST_PAIR(mask_m, 6, 3);                                   \
+  k3_m = VP9_SET_CONST_PAIR(mask_m, 3, 2);                                   \
+  VP9_MADD(in1, in7, in3, in5, k0_m, k1_m, k2_m, k3_m, in1, in7, in3, in5);  \
+  SUB2(in1, in3, in7, in5, res0_m, res1_m);                                  \
+  k0_m = VP9_SET_CONST_PAIR(mask_m, 4, 7);                                   \
+  k1_m = __msa_splati_h(mask_m, 4);                                          \
+                                                                             \
+  ILVRL_H2_SH(res0_m, res1_m, res2_m, res3_m);                               \
+  DOTP_SH4_SW(res2_m, res3_m, res2_m, res3_m, k0_m, k0_m, k1_m, k1_m,        \
+              tmp0_m, tmp1_m, tmp2_m, tmp3_m);                               \
+  SRARI_W4_SW(tmp0_m, tmp1_m, tmp2_m, tmp3_m, DCT_CONST_BITS);               \
+  tp4_m = in1 + in3;                                                         \
+  PCKEV_H2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, tp5_m, tp6_m);                 \
+  tp7_m = in7 + in5;                                                         \
+  k2_m = VP9_SET_COSPI_PAIR(cospi_24_64, -cospi_8_64);                       \
+  k3_m = VP9_SET_COSPI_PAIR(cospi_8_64, cospi_24_64);                        \
+  VP9_MADD(in0, in4, in2, in6, k1_m, k0_m, k2_m, k3_m,                       \
+           in0, in4, in2, in6);                                              \
+  BUTTERFLY_4(in0, in4, in2, in6, tp0_m, tp1_m, tp2_m, tp3_m);               \
+  BUTTERFLY_8(tp0_m, tp1_m, tp2_m, tp3_m, tp4_m, tp5_m, tp6_m, tp7_m,        \
+              out0, out1, out2, out3, out4, out5, out6, out7);               \
+}
+
+#define VP9_IADST8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7,            \
+                        out0, out1, out2, out3, out4, out5, out6, out7) {  \
+  v4i32 r0_m, r1_m, r2_m, r3_m, r4_m, r5_m, r6_m, r7_m;                    \
+  v4i32 m0_m, m1_m, m2_m, m3_m, t0_m, t1_m;                                \
+  v8i16 res0_m, res1_m, res2_m, res3_m, k0_m, k1_m, in_s0, in_s1;          \
+  v8i16 mask1_m = { cospi_2_64, cospi_30_64, -cospi_2_64,                  \
+    cospi_10_64, cospi_22_64, -cospi_10_64, cospi_18_64, cospi_14_64 };    \
+  v8i16 mask2_m = { cospi_14_64, -cospi_18_64, cospi_26_64,                \
+    cospi_6_64, -cospi_26_64, cospi_8_64, cospi_24_64, -cospi_8_64 };      \
+  v8i16 mask3_m = { -cospi_24_64, cospi_8_64, cospi_16_64,                 \
+    -cospi_16_64, 0, 0, 0, 0 };                                            \
+                                                                           \
+  k0_m = VP9_SET_CONST_PAIR(mask1_m, 0, 1);                                \
+  k1_m = VP9_SET_CONST_PAIR(mask1_m, 1, 2);                                \
+  ILVRL_H2_SH(in1, in0, in_s1, in_s0);                                     \
+  DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m,          \
+              r0_m, r1_m, r2_m, r3_m);                                     \
+  k0_m = VP9_SET_CONST_PAIR(mask1_m, 6, 7);                                \
+  k1_m = VP9_SET_CONST_PAIR(mask2_m, 0, 1);                                \
+  ILVRL_H2_SH(in5, in4, in_s1, in_s0);                                     \
+  DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m,          \
+              r4_m, r5_m, r6_m, r7_m);                                     \
+  ADD4(r0_m, r4_m, r1_m, r5_m, r2_m, r6_m, r3_m, r7_m,                     \
+       m0_m, m1_m, m2_m, m3_m);                                            \
+  SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                     \
+  PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, res0_m, res1_m);                     \
+  SUB4(r0_m, r4_m, r1_m, r5_m, r2_m, r6_m, r3_m, r7_m,                     \
+       m0_m, m1_m, m2_m, m3_m);                                            \
+  SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                     \
+  PCKEV_H2_SW(m1_m, m0_m, m3_m, m2_m, t0_m, t1_m);                         \
+  k0_m = VP9_SET_CONST_PAIR(mask1_m, 3, 4);                                \
+  k1_m = VP9_SET_CONST_PAIR(mask1_m, 4, 5);                                \
+  ILVRL_H2_SH(in3, in2, in_s1, in_s0);                                     \
+  DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m,          \
+              r0_m, r1_m, r2_m, r3_m);                                     \
+  k0_m = VP9_SET_CONST_PAIR(mask2_m, 2, 3);                                \
+  k1_m = VP9_SET_CONST_PAIR(mask2_m, 3, 4);                                \
+  ILVRL_H2_SH(in7, in6, in_s1, in_s0);                                     \
+  DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m,          \
+              r4_m, r5_m, r6_m, r7_m);                                     \
+  ADD4(r0_m, r4_m, r1_m, r5_m, r2_m, r6_m, r3_m, r7_m,                     \
+       m0_m, m1_m, m2_m, m3_m);                                            \
+  SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                     \
+  PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, res2_m, res3_m);                     \
+  SUB4(r0_m, r4_m, r1_m, r5_m, r2_m, r6_m, r3_m, r7_m,                     \
+       m0_m, m1_m, m2_m, m3_m);                                            \
+  SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                     \
+  PCKEV_H2_SW(m1_m, m0_m, m3_m, m2_m, r2_m, r3_m);                         \
+  ILVRL_H2_SW(r3_m, r2_m, m2_m, m3_m);                                     \
+  BUTTERFLY_4(res0_m, res1_m, res3_m, res2_m, out0, in7, in4, in3);        \
+  k0_m = VP9_SET_CONST_PAIR(mask2_m, 5, 6);                                \
+  k1_m = VP9_SET_CONST_PAIR(mask2_m, 6, 7);                                \
+  ILVRL_H2_SH(t1_m, t0_m, in_s1, in_s0);                                   \
+  DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m,          \
+              r0_m, r1_m, r2_m, r3_m);                                     \
+  k1_m = VP9_SET_CONST_PAIR(mask3_m, 0, 1);                                \
+  DOTP_SH4_SW(m2_m, m3_m, m2_m, m3_m, k0_m, k0_m, k1_m, k1_m,              \
+              r4_m, r5_m, r6_m, r7_m);                                     \
+  ADD4(r0_m, r6_m, r1_m, r7_m, r2_m, r4_m, r3_m, r5_m,                     \
+       m0_m, m1_m, m2_m, m3_m);                                            \
+  SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                     \
+  PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, in1, out6);                          \
+  SUB4(r0_m, r6_m, r1_m, r7_m, r2_m, r4_m, r3_m, r5_m,                     \
+       m0_m, m1_m, m2_m, m3_m);                                            \
+  SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                     \
+  PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, in2, in5);                           \
+  k0_m = VP9_SET_CONST_PAIR(mask3_m, 2, 2);                                \
+  k1_m = VP9_SET_CONST_PAIR(mask3_m, 2, 3);                                \
+  ILVRL_H2_SH(in4, in3, in_s1, in_s0);                                     \
+  DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m,          \
+              m0_m, m1_m, m2_m, m3_m);                                     \
+  SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                     \
+  PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, in3, out4);                          \
+  ILVRL_H2_SW(in5, in2, m2_m, m3_m);                                       \
+  DOTP_SH4_SW(m2_m, m3_m, m2_m, m3_m, k0_m, k0_m, k1_m, k1_m,              \
+              m0_m, m1_m, m2_m, m3_m);                                     \
+  SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                     \
+  PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, out2, in5);                          \
+                                                                           \
+  out1 = -in1;                                                             \
+  out3 = -in3;                                                             \
+  out5 = -in5;                                                             \
+  out7 = -in7;                                                             \
+}
+
+#define VP9_IADST8x16_1D(r0, r1, r2, r3, r4, r5, r6, r7, r8,        \
+                         r9, r10, r11, r12, r13, r14, r15,          \
+                         out0, out1, out2, out3, out4, out5,        \
+                         out6, out7, out8, out9, out10, out11,      \
+                         out12, out13, out14, out15) {              \
+  v8i16 g0_m, g1_m, g2_m, g3_m, g4_m, g5_m, g6_m, g7_m;             \
+  v8i16 g8_m, g9_m, g10_m, g11_m, g12_m, g13_m, g14_m, g15_m;       \
+  v8i16 h0_m, h1_m, h2_m, h3_m, h4_m, h5_m, h6_m, h7_m;             \
+  v8i16 h8_m, h9_m, h10_m, h11_m;                                   \
+  v8i16 k0_m, k1_m, k2_m, k3_m;                                     \
+                                                                    \
+  /* stage 1 */                                                     \
+  k0_m = VP9_SET_COSPI_PAIR(cospi_1_64, cospi_31_64);               \
+  k1_m = VP9_SET_COSPI_PAIR(cospi_31_64, -cospi_1_64);              \
+  k2_m = VP9_SET_COSPI_PAIR(cospi_17_64, cospi_15_64);              \
+  k3_m = VP9_SET_COSPI_PAIR(cospi_15_64, -cospi_17_64);             \
+  MADD_BF(r15, r0, r7, r8, k0_m, k1_m, k2_m, k3_m,                  \
+          g0_m, g1_m, g2_m, g3_m);                                  \
+  k0_m = VP9_SET_COSPI_PAIR(cospi_5_64, cospi_27_64);               \
+  k1_m = VP9_SET_COSPI_PAIR(cospi_27_64, -cospi_5_64);              \
+  k2_m = VP9_SET_COSPI_PAIR(cospi_21_64, cospi_11_64);              \
+  k3_m = VP9_SET_COSPI_PAIR(cospi_11_64, -cospi_21_64);             \
+  MADD_BF(r13, r2, r5, r10, k0_m, k1_m, k2_m, k3_m,                 \
+          g4_m, g5_m, g6_m, g7_m);                                  \
+  k0_m = VP9_SET_COSPI_PAIR(cospi_9_64, cospi_23_64);               \
+  k1_m = VP9_SET_COSPI_PAIR(cospi_23_64, -cospi_9_64);              \
+  k2_m = VP9_SET_COSPI_PAIR(cospi_25_64, cospi_7_64);               \
+  k3_m = VP9_SET_COSPI_PAIR(cospi_7_64, -cospi_25_64);              \
+  MADD_BF(r11, r4, r3, r12, k0_m, k1_m, k2_m, k3_m,                 \
+          g8_m, g9_m, g10_m, g11_m);                                \
+  k0_m = VP9_SET_COSPI_PAIR(cospi_13_64, cospi_19_64);              \
+  k1_m = VP9_SET_COSPI_PAIR(cospi_19_64, -cospi_13_64);             \
+  k2_m = VP9_SET_COSPI_PAIR(cospi_29_64, cospi_3_64);               \
+  k3_m = VP9_SET_COSPI_PAIR(cospi_3_64, -cospi_29_64);              \
+  MADD_BF(r9, r6, r1, r14, k0_m, k1_m, k2_m, k3_m,                  \
+          g12_m, g13_m, g14_m, g15_m);                              \
+                                                                    \
+  /* stage 2 */                                                     \
+  k0_m = VP9_SET_COSPI_PAIR(cospi_4_64, cospi_28_64);               \
+  k1_m = VP9_SET_COSPI_PAIR(cospi_28_64, -cospi_4_64);              \
+  k2_m = VP9_SET_COSPI_PAIR(-cospi_28_64, cospi_4_64);              \
+  MADD_BF(g1_m, g3_m, g9_m, g11_m, k0_m, k1_m, k2_m, k0_m,          \
+          h0_m, h1_m, h2_m, h3_m);                                  \
+  k0_m = VP9_SET_COSPI_PAIR(cospi_12_64, cospi_20_64);              \
+  k1_m = VP9_SET_COSPI_PAIR(-cospi_20_64, cospi_12_64);             \
+  k2_m = VP9_SET_COSPI_PAIR(cospi_20_64, -cospi_12_64);             \
+  MADD_BF(g7_m, g5_m, g15_m, g13_m, k0_m, k1_m, k2_m, k0_m,         \
+          h4_m, h5_m, h6_m, h7_m);                                  \
+  BUTTERFLY_4(h0_m, h2_m, h6_m, h4_m, out8, out9, out11, out10);    \
+  BUTTERFLY_8(g0_m, g2_m, g4_m, g6_m, g14_m, g12_m, g10_m, g8_m,    \
+              h8_m, h9_m, h10_m, h11_m, h6_m, h4_m, h2_m, h0_m);    \
+                                                                    \
+  /* stage 3 */                                                     \
+  BUTTERFLY_4(h8_m, h9_m, h11_m, h10_m, out0, out1, h11_m, h10_m);  \
+  k0_m = VP9_SET_COSPI_PAIR(cospi_8_64, cospi_24_64);               \
+  k1_m = VP9_SET_COSPI_PAIR(cospi_24_64, -cospi_8_64);              \
+  k2_m = VP9_SET_COSPI_PAIR(-cospi_24_64, cospi_8_64);              \
+  MADD_BF(h0_m, h2_m, h4_m, h6_m, k0_m, k1_m, k2_m, k0_m,           \
+          out4, out6, out5, out7);                                  \
+  MADD_BF(h1_m, h3_m, h5_m, h7_m, k0_m, k1_m, k2_m, k0_m,           \
+          out12, out14, out13, out15);                              \
+                                                                    \
+  /* stage 4 */                                                     \
+  k0_m = VP9_SET_COSPI_PAIR(cospi_16_64, cospi_16_64);              \
+  k1_m = VP9_SET_COSPI_PAIR(-cospi_16_64, -cospi_16_64);            \
+  k2_m = VP9_SET_COSPI_PAIR(cospi_16_64, -cospi_16_64);             \
+  k3_m = VP9_SET_COSPI_PAIR(-cospi_16_64, cospi_16_64);             \
+  MADD_SHORT(h10_m, h11_m, k1_m, k2_m, out2, out3);                 \
+  MADD_SHORT(out6, out7, k0_m, k3_m, out6, out7);                   \
+  MADD_SHORT(out10, out11, k0_m, k3_m, out10, out11);               \
+  MADD_SHORT(out14, out15, k1_m, k2_m, out14, out15);               \
+}
+
+void vpx_idct16_1d_columns_addblk_msa(int16_t *input, uint8_t *dst,
+                                      int32_t dst_stride);
+void vpx_idct16_1d_rows_msa(const int16_t *input, int16_t *output);
+void vpx_iadst16_1d_columns_addblk_msa(int16_t *input, uint8_t *dst,
+                                       int32_t dst_stride);
+void vpx_iadst16_1d_rows_msa(const int16_t *input, int16_t *output);
+#endif  // VPX_DSP_MIPS_INV_TXFM_MSA_H_
diff --git a/libvpx/libvpx/vpx_dsp/mips/itrans16_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/itrans16_dspr2.c
new file mode 100644
index 0000000..6d41e61
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/itrans16_dspr2.c
@@ -0,0 +1,1227 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/inv_txfm_dspr2.h"
+#include "vpx_dsp/txfm_common.h"
+
+#if HAVE_DSPR2
+void idct16_rows_dspr2(const int16_t *input, int16_t *output,
+                       uint32_t no_rows) {
+  int i;
+  int step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6, step1_7;
+  int step1_10, step1_11, step1_12, step1_13;
+  int step2_0, step2_1, step2_2, step2_3;
+  int step2_8, step2_9, step2_10, step2_11;
+  int step2_12, step2_13, step2_14, step2_15;
+  int load1, load2, load3, load4, load5, load6, load7, load8;
+  int result1, result2, result3, result4;
+  const int const_2_power_13 = 8192;
+
+  for (i = no_rows; i--; ) {
+    /* prefetch row */
+    prefetch_load((const uint8_t *)(input + 16));
+
+    __asm__ __volatile__ (
+        "lh       %[load1],              0(%[input])                    \n\t"
+        "lh       %[load2],             16(%[input])                    \n\t"
+        "lh       %[load3],              8(%[input])                    \n\t"
+        "lh       %[load4],             24(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "add      %[result1],           %[load1],       %[load2]        \n\t"
+        "sub      %[result2],           %[load1],       %[load2]        \n\t"
+        "madd     $ac1,                 %[result1],     %[cospi_16_64]  \n\t"
+        "madd     $ac2,                 %[result2],     %[cospi_16_64]  \n\t"
+        "extp     %[step2_0],           $ac1,           31              \n\t"
+        "extp     %[step2_1],           $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "madd     $ac3,                 %[load3],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load4],       %[cospi_8_64]   \n\t"
+        "extp     %[step2_2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_24_64]  \n\t"
+        "extp     %[step2_3],           $ac1,           31              \n\t"
+
+        "add      %[step1_0],           %[step2_0],     %[step2_3]      \n\t"
+        "add      %[step1_1],           %[step2_1],     %[step2_2]      \n\t"
+        "sub      %[step1_2],           %[step2_1],     %[step2_2]      \n\t"
+        "sub      %[step1_3],           %[step2_0],     %[step2_3]      \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [result1] "=&r" (result1), [result2] "=&r" (result2),
+          [step2_0] "=&r" (step2_0), [step2_1] "=&r" (step2_1),
+          [step2_2] "=&r" (step2_2), [step2_3] "=&r" (step2_3),
+          [step1_0] "=r" (step1_0), [step1_1] "=r" (step1_1),
+          [step1_2] "=r" (step1_2), [step1_3] "=r" (step1_3)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load5],             2(%[input])                     \n\t"
+        "lh       %[load6],             30(%[input])                    \n\t"
+        "lh       %[load7],             18(%[input])                    \n\t"
+        "lh       %[load8],             14(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load5],       %[cospi_30_64]  \n\t"
+        "msub     $ac1,                 %[load6],       %[cospi_2_64]   \n\t"
+        "extp     %[result1],           $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load7],       %[cospi_14_64]  \n\t"
+        "msub     $ac3,                 %[load8],       %[cospi_18_64]  \n\t"
+        "extp     %[result2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac1,                 %[load7],       %[cospi_18_64]  \n\t"
+        "madd     $ac1,                 %[load8],       %[cospi_14_64]  \n\t"
+        "extp     %[result3],           $ac1,           31              \n\t"
+
+        "madd     $ac2,                 %[load5],       %[cospi_2_64]   \n\t"
+        "madd     $ac2,                 %[load6],       %[cospi_30_64]  \n\t"
+        "extp     %[result4],           $ac2,           31              \n\t"
+
+        "sub      %[load5],             %[result1],     %[result2]      \n\t"
+        "sub      %[load6],             %[result4],     %[result3]      \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load6],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load5],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load5],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load6],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_9],           $ac1,           31              \n\t"
+        "extp     %[step2_14],          $ac3,           31              \n\t"
+        "add      %[step2_8],           %[result1],     %[result2]      \n\t"
+        "add      %[step2_15],          %[result4],     %[result3]      \n\t"
+
+        : [load5] "=&r" (load5), [load6] "=&r" (load6),
+          [load7] "=&r" (load7), [load8] "=&r" (load8),
+          [result1] "=&r" (result1), [result2] "=&r" (result2),
+          [result3] "=&r" (result3), [result4] "=&r" (result4),
+          [step2_8] "=r" (step2_8), [step2_15] "=r" (step2_15),
+          [step2_9] "=r" (step2_9), [step2_14] "=r" (step2_14)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_30_64] "r" (cospi_30_64), [cospi_2_64] "r" (cospi_2_64),
+          [cospi_14_64] "r" (cospi_14_64), [cospi_18_64] "r" (cospi_18_64),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             10(%[input])                    \n\t"
+        "lh       %[load2],             22(%[input])                    \n\t"
+        "lh       %[load3],             26(%[input])                    \n\t"
+        "lh       %[load4],             6(%[input])                     \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_22_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_10_64]  \n\t"
+        "extp     %[result1],           $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load3],       %[cospi_6_64]   \n\t"
+        "msub     $ac3,                 %[load4],       %[cospi_26_64]  \n\t"
+        "extp     %[result2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_10_64]  \n\t"
+        "madd     $ac1,                 %[load2],       %[cospi_22_64]  \n\t"
+        "extp     %[result3],           $ac1,           31              \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_26_64]  \n\t"
+        "madd     $ac2,                 %[load4],       %[cospi_6_64]   \n\t"
+        "extp     %[result4],           $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[result2],     %[result1]      \n\t"
+        "sub      %[load2],             %[result4],     %[result3]      \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_10],          $ac1,           31              \n\t"
+        "extp     %[step2_13],          $ac3,           31              \n\t"
+        "add      %[step2_11],          %[result1],     %[result2]      \n\t"
+        "add      %[step2_12],          %[result4],     %[result3]      \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [result1] "=&r" (result1), [result2] "=&r" (result2),
+          [result3] "=&r" (result3), [result4] "=&r" (result4),
+          [step2_10] "=r" (step2_10), [step2_11] "=r" (step2_11),
+          [step2_12] "=r" (step2_12), [step2_13] "=r" (step2_13)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_22_64] "r" (cospi_22_64), [cospi_10_64] "r" (cospi_10_64),
+          [cospi_6_64] "r" (cospi_6_64), [cospi_26_64] "r" (cospi_26_64),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load5],             4(%[input])                     \n\t"
+        "lh       %[load6],             28(%[input])                    \n\t"
+        "lh       %[load7],             20(%[input])                    \n\t"
+        "lh       %[load8],             12(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load5],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load6],       %[cospi_4_64]   \n\t"
+        "extp     %[result1],           $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load7],       %[cospi_12_64]  \n\t"
+        "msub     $ac3,                 %[load8],       %[cospi_20_64]  \n\t"
+        "extp     %[result2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac1,                 %[load7],       %[cospi_20_64]  \n\t"
+        "madd     $ac1,                 %[load8],       %[cospi_12_64]  \n\t"
+        "extp     %[result3],           $ac1,           31              \n\t"
+
+        "madd     $ac2,                 %[load5],       %[cospi_4_64]   \n\t"
+        "madd     $ac2,                 %[load6],       %[cospi_28_64]  \n\t"
+        "extp     %[result4],           $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load5],             %[result4],     %[result3]      \n\t"
+        "sub      %[load5],             %[load5],       %[result1]      \n\t"
+        "add      %[load5],             %[load5],       %[result2]      \n\t"
+
+        "sub      %[load6],             %[result1],     %[result2]      \n\t"
+        "sub      %[load6],             %[load6],       %[result3]      \n\t"
+        "add      %[load6],             %[load6],       %[result4]      \n\t"
+
+        "madd     $ac1,                 %[load5],       %[cospi_16_64]  \n\t"
+        "madd     $ac3,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_5],           $ac1,           31              \n\t"
+        "extp     %[step1_6],           $ac3,           31              \n\t"
+        "add      %[step1_4],           %[result1],     %[result2]      \n\t"
+        "add      %[step1_7],           %[result4],     %[result3]      \n\t"
+
+        : [load5] "=&r" (load5), [load6] "=&r" (load6),
+          [load7] "=&r" (load7), [load8] "=&r" (load8),
+          [result1] "=&r" (result1), [result2] "=&r" (result2),
+          [result3] "=&r" (result3), [result4] "=&r" (result4),
+          [step1_4] "=r" (step1_4), [step1_5] "=r" (step1_5),
+          [step1_6] "=r" (step1_6), [step1_7] "=r" (step1_7)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_20_64] "r" (cospi_20_64), [cospi_12_64] "r" (cospi_12_64),
+          [cospi_4_64] "r" (cospi_4_64), [cospi_28_64] "r" (cospi_28_64),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        "sub      %[load5],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[load5],             %[load5],       %[step2_9]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_10]     \n\t"
+
+        "madd     $ac0,                 %[load5],       %[cospi_16_64]  \n\t"
+
+        "sub      %[load6],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[load6],             %[load6],       %[step2_10]     \n\t"
+        "add      %[load6],             %[load6],       %[step2_9]      \n\t"
+
+        "madd     $ac1,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load5],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[load5],             %[load5],       %[step2_8]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_11]     \n\t"
+
+        "madd     $ac2,                 %[load5],       %[cospi_16_64]  \n\t"
+
+        "sub      %[load6],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[load6],             %[load6],       %[step2_11]     \n\t"
+        "add      %[load6],             %[load6],       %[step2_8]      \n\t"
+
+        "madd     $ac3,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_10],          $ac0,           31              \n\t"
+        "extp     %[step1_13],          $ac1,           31              \n\t"
+        "extp     %[step1_11],          $ac2,           31              \n\t"
+        "extp     %[step1_12],          $ac3,           31              \n\t"
+
+        : [load5] "=&r" (load5), [load6] "=&r" (load6),
+          [step1_10] "=r" (step1_10), [step1_11] "=r" (step1_11),
+          [step1_12] "=r" (step1_12), [step1_13] "=r" (step1_13)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_14] "r" (step2_14), [step2_13] "r" (step2_13),
+          [step2_9] "r" (step2_9), [step2_10] "r" (step2_10),
+          [step2_15] "r" (step2_15), [step2_12] "r" (step2_12),
+          [step2_8] "r" (step2_8), [step2_11] "r" (step2_11),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    __asm__ __volatile__ (
+        "add      %[load5],             %[step1_0],     %[step1_7]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_12]     \n\t"
+        "add      %[load5],             %[load5],       %[step2_15]     \n\t"
+        "add      %[load6],             %[step1_1],     %[step1_6]      \n\t"
+        "add      %[load6],             %[load6],       %[step2_13]     \n\t"
+        "add      %[load6],             %[load6],       %[step2_14]     \n\t"
+        "sh       %[load5],             0(%[output])                    \n\t"
+        "sh       %[load6],             32(%[output])                   \n\t"
+        "sub      %[load5],             %[step1_1],     %[step1_6]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_9]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_10]     \n\t"
+        "sub      %[load6],             %[step1_0],     %[step1_7]      \n\t"
+        "add      %[load6],             %[load6],       %[step2_8]      \n\t"
+        "add      %[load6],             %[load6],       %[step2_11]     \n\t"
+        "sh       %[load5],             192(%[output])                  \n\t"
+        "sh       %[load6],             224(%[output])                  \n\t"
+        "sub      %[load5],             %[step1_0],     %[step1_7]      \n\t"
+        "sub      %[load5],             %[load5],       %[step2_8]      \n\t"
+        "sub      %[load5],             %[load5],       %[step2_11]     \n\t"
+        "sub      %[load6],             %[step1_1],     %[step1_6]      \n\t"
+        "sub      %[load6],             %[load6],       %[step2_9]      \n\t"
+        "sub      %[load6],             %[load6],       %[step2_10]     \n\t"
+        "sh       %[load5],             256(%[output])                  \n\t"
+        "sh       %[load6],             288(%[output])                  \n\t"
+        "add      %[load5],             %[step1_1],     %[step1_6]      \n\t"
+        "sub      %[load5],             %[load5],       %[step2_13]     \n\t"
+        "sub      %[load5],             %[load5],       %[step2_14]     \n\t"
+        "add      %[load6],             %[step1_0],     %[step1_7]      \n\t"
+        "sub      %[load6],             %[load6],       %[step2_12]     \n\t"
+        "sub      %[load6],             %[load6],       %[step2_15]     \n\t"
+        "sh       %[load5],             448(%[output])                  \n\t"
+        "sh       %[load6],             480(%[output])                  \n\t"
+
+        : [load5] "=&r" (load5), [load6] "=&r" (load6)
+        : [output] "r" (output),
+          [step1_0] "r" (step1_0), [step1_1] "r" (step1_1),
+          [step1_6] "r" (step1_6), [step1_7] "r" (step1_7),
+          [step2_8] "r" (step2_8), [step2_9] "r" (step2_9),
+          [step2_10] "r" (step2_10), [step2_11] "r" (step2_11),
+          [step2_12] "r" (step2_12), [step2_13] "r" (step2_13),
+          [step2_14] "r" (step2_14), [step2_15] "r" (step2_15)
+    );
+
+    __asm__ __volatile__ (
+        "add      %[load5],             %[step1_2],     %[step1_5]      \n\t"
+        "add      %[load5],             %[load5],       %[step1_13]     \n\t"
+        "add      %[load6],             %[step1_3],     %[step1_4]      \n\t"
+        "add      %[load6],             %[load6],       %[step1_12]     \n\t"
+        "sh       %[load5],             64(%[output])                   \n\t"
+        "sh       %[load6],             96(%[output])                   \n\t"
+        "sub      %[load5],             %[step1_3],     %[step1_4]      \n\t"
+        "add      %[load5],             %[load5],       %[step1_11]     \n\t"
+        "sub      %[load6],             %[step1_2],     %[step1_5]      \n\t"
+        "add      %[load6],             %[load6],       %[step1_10]     \n\t"
+        "sh       %[load5],             128(%[output])                  \n\t"
+        "sh       %[load6],             160(%[output])                  \n\t"
+        "sub      %[load5],             %[step1_2],     %[step1_5]      \n\t"
+        "sub      %[load5],             %[load5],       %[step1_10]     \n\t"
+        "sub      %[load6],             %[step1_3],     %[step1_4]      \n\t"
+        "sub      %[load6],             %[load6],       %[step1_11]     \n\t"
+        "sh       %[load5],             320(%[output])                  \n\t"
+        "sh       %[load6],             352(%[output])                  \n\t"
+        "add      %[load5],             %[step1_3],     %[step1_4]      \n\t"
+        "sub      %[load5],             %[load5],       %[step1_12]     \n\t"
+        "add      %[load6],             %[step1_2],     %[step1_5]      \n\t"
+        "sub      %[load6],             %[load6],       %[step1_13]     \n\t"
+        "sh       %[load5],             384(%[output])                  \n\t"
+        "sh       %[load6],             416(%[output])                  \n\t"
+
+        : [load5] "=&r" (load5), [load6] "=&r" (load6)
+        : [output] "r" (output),
+          [step1_2] "r" (step1_2), [step1_3] "r" (step1_3),
+          [step1_4] "r" (step1_4), [step1_5] "r" (step1_5),
+          [step1_10] "r" (step1_10), [step1_11] "r" (step1_11),
+          [step1_12] "r" (step1_12), [step1_13] "r" (step1_13)
+    );
+
+    input += 16;
+    output += 1;
+  }
+}
+
+void idct16_cols_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                               int dest_stride) {
+  int i;
+  int step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6, step1_7;
+  int step1_8, step1_9, step1_10, step1_11;
+  int step1_12, step1_13, step1_14, step1_15;
+  int step2_0, step2_1, step2_2, step2_3;
+  int step2_8, step2_9, step2_10, step2_11;
+  int step2_12, step2_13, step2_14, step2_15;
+  int load1, load2, load3, load4, load5, load6, load7, load8;
+  int result1, result2, result3, result4;
+  const int const_2_power_13 = 8192;
+  uint8_t *dest_pix;
+  uint8_t *cm = vpx_ff_cropTbl;
+
+  /* prefetch vpx_ff_cropTbl */
+  prefetch_load(vpx_ff_cropTbl);
+  prefetch_load(vpx_ff_cropTbl +  32);
+  prefetch_load(vpx_ff_cropTbl +  64);
+  prefetch_load(vpx_ff_cropTbl +  96);
+  prefetch_load(vpx_ff_cropTbl + 128);
+  prefetch_load(vpx_ff_cropTbl + 160);
+  prefetch_load(vpx_ff_cropTbl + 192);
+  prefetch_load(vpx_ff_cropTbl + 224);
+
+  for (i = 0; i < 16; ++i) {
+    dest_pix = (dest + i);
+    __asm__ __volatile__ (
+        "lh       %[load1],              0(%[input])                    \n\t"
+        "lh       %[load2],             16(%[input])                    \n\t"
+        "lh       %[load3],              8(%[input])                    \n\t"
+        "lh       %[load4],             24(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "add      %[result1],           %[load1],       %[load2]        \n\t"
+        "sub      %[result2],           %[load1],       %[load2]        \n\t"
+        "madd     $ac1,                 %[result1],     %[cospi_16_64]  \n\t"
+        "madd     $ac2,                 %[result2],     %[cospi_16_64]  \n\t"
+        "extp     %[step2_0],           $ac1,           31              \n\t"
+        "extp     %[step2_1],           $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "madd     $ac3,                 %[load3],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load4],       %[cospi_8_64]   \n\t"
+        "extp     %[step2_2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_24_64]  \n\t"
+        "extp     %[step2_3],           $ac1,           31              \n\t"
+
+        "add      %[step1_0],           %[step2_0],     %[step2_3]      \n\t"
+        "add      %[step1_1],           %[step2_1],     %[step2_2]      \n\t"
+        "sub      %[step1_2],           %[step2_1],     %[step2_2]      \n\t"
+        "sub      %[step1_3],           %[step2_0],     %[step2_3]      \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [result1] "=&r" (result1), [result2] "=&r" (result2),
+          [step2_0] "=&r" (step2_0), [step2_1] "=&r" (step2_1),
+          [step2_2] "=&r" (step2_2), [step2_3] "=&r" (step2_3),
+          [step1_0] "=r" (step1_0), [step1_1] "=r" (step1_1),
+          [step1_2] "=r" (step1_2), [step1_3] "=r" (step1_3)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load5],             2(%[input])                     \n\t"
+        "lh       %[load6],             30(%[input])                    \n\t"
+        "lh       %[load7],             18(%[input])                    \n\t"
+        "lh       %[load8],             14(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load5],       %[cospi_30_64]  \n\t"
+        "msub     $ac1,                 %[load6],       %[cospi_2_64]   \n\t"
+        "extp     %[result1],           $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load7],       %[cospi_14_64]  \n\t"
+        "msub     $ac3,                 %[load8],       %[cospi_18_64]  \n\t"
+        "extp     %[result2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac1,                 %[load7],       %[cospi_18_64]  \n\t"
+        "madd     $ac1,                 %[load8],       %[cospi_14_64]  \n\t"
+        "extp     %[result3],           $ac1,           31              \n\t"
+
+        "madd     $ac2,                 %[load5],        %[cospi_2_64]  \n\t"
+        "madd     $ac2,                 %[load6],        %[cospi_30_64] \n\t"
+        "extp     %[result4],           $ac2,            31             \n\t"
+
+        "sub      %[load5],             %[result1],     %[result2]      \n\t"
+        "sub      %[load6],             %[result4],     %[result3]      \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load6],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load5],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load5],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load6],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_9],           $ac1,           31              \n\t"
+        "extp     %[step2_14],          $ac3,           31              \n\t"
+        "add      %[step2_8],           %[result1],     %[result2]      \n\t"
+        "add      %[step2_15],          %[result4],     %[result3]      \n\t"
+
+        : [load5] "=&r" (load5), [load6] "=&r" (load6),
+          [load7] "=&r" (load7), [load8] "=&r" (load8),
+          [result1] "=&r" (result1), [result2] "=&r" (result2),
+          [result3] "=&r" (result3), [result4] "=&r" (result4),
+          [step2_8] "=r" (step2_8), [step2_15] "=r" (step2_15),
+          [step2_9] "=r" (step2_9), [step2_14] "=r" (step2_14)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_30_64] "r" (cospi_30_64), [cospi_2_64] "r" (cospi_2_64),
+          [cospi_14_64] "r" (cospi_14_64), [cospi_18_64] "r" (cospi_18_64),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             10(%[input])                    \n\t"
+        "lh       %[load2],             22(%[input])                    \n\t"
+        "lh       %[load3],             26(%[input])                    \n\t"
+        "lh       %[load4],             6(%[input])                     \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],    %[cospi_22_64]     \n\t"
+        "msub     $ac1,                 %[load2],    %[cospi_10_64]     \n\t"
+        "extp     %[result1],           $ac1,        31                 \n\t"
+
+        "madd     $ac3,                 %[load3],    %[cospi_6_64]      \n\t"
+        "msub     $ac3,                 %[load4],    %[cospi_26_64]     \n\t"
+        "extp     %[result2],           $ac3,        31                 \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac1,                 %[load1],    %[cospi_10_64]     \n\t"
+        "madd     $ac1,                 %[load2],    %[cospi_22_64]     \n\t"
+        "extp     %[result3],           $ac1,        31                 \n\t"
+
+        "madd     $ac2,                 %[load3],    %[cospi_26_64]     \n\t"
+        "madd     $ac2,                 %[load4],    %[cospi_6_64]      \n\t"
+        "extp     %[result4],           $ac2,        31                 \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[result2],     %[result1]      \n\t"
+        "sub      %[load2],             %[result4],     %[result3]      \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_10],          $ac1,           31              \n\t"
+        "extp     %[step2_13],          $ac3,           31              \n\t"
+        "add      %[step2_11],          %[result1],     %[result2]      \n\t"
+        "add      %[step2_12],          %[result4],     %[result3]      \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [result1] "=&r" (result1), [result2] "=&r" (result2),
+          [result3] "=&r" (result3), [result4] "=&r" (result4),
+          [step2_10] "=r" (step2_10), [step2_11] "=r" (step2_11),
+          [step2_12] "=r" (step2_12), [step2_13] "=r" (step2_13)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_22_64] "r" (cospi_22_64), [cospi_10_64] "r" (cospi_10_64),
+          [cospi_6_64] "r" (cospi_6_64), [cospi_26_64] "r" (cospi_26_64),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load5],             4(%[input])                   \n\t"
+        "lh       %[load6],             28(%[input])                  \n\t"
+        "lh       %[load7],             20(%[input])                  \n\t"
+        "lh       %[load8],             12(%[input])                  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                          \n\t"
+        "mthi     $zero,                $ac1                          \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                          \n\t"
+        "mthi     $zero,                $ac3                          \n\t"
+
+        "madd     $ac1,                 %[load5],    %[cospi_28_64]   \n\t"
+        "msub     $ac1,                 %[load6],    %[cospi_4_64]    \n\t"
+        "extp     %[result1],           $ac1,        31               \n\t"
+
+        "madd     $ac3,                 %[load7],    %[cospi_12_64]   \n\t"
+        "msub     $ac3,                 %[load8],    %[cospi_20_64]   \n\t"
+        "extp     %[result2],           $ac3,        31               \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                          \n\t"
+        "mthi     $zero,                $ac1                          \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                          \n\t"
+        "mthi     $zero,                $ac2                          \n\t"
+
+        "madd     $ac1,                 %[load7],    %[cospi_20_64]   \n\t"
+        "madd     $ac1,                 %[load8],    %[cospi_12_64]   \n\t"
+        "extp     %[result3],           $ac1,        31               \n\t"
+
+        "madd     $ac2,                 %[load5],    %[cospi_4_64]    \n\t"
+        "madd     $ac2,                 %[load6],    %[cospi_28_64]   \n\t"
+        "extp     %[result4],           $ac2,        31               \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load5],             %[result4],     %[result3]      \n\t"
+        "sub      %[load5],             %[load5],       %[result1]      \n\t"
+        "add      %[load5],             %[load5],       %[result2]      \n\t"
+
+        "sub      %[load6],             %[result1],     %[result2]      \n\t"
+        "sub      %[load6],             %[load6],       %[result3]      \n\t"
+        "add      %[load6],             %[load6],       %[result4]      \n\t"
+
+        "madd     $ac1,                 %[load5],       %[cospi_16_64]  \n\t"
+        "madd     $ac3,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_5],           $ac1,           31              \n\t"
+        "extp     %[step1_6],           $ac3,           31              \n\t"
+
+        "add      %[step1_4],           %[result1],     %[result2]      \n\t"
+        "add      %[step1_7],           %[result4],     %[result3]      \n\t"
+
+        : [load5] "=&r" (load5), [load6] "=&r" (load6),
+          [load7] "=&r" (load7), [load8] "=&r" (load8),
+          [result1] "=&r" (result1), [result2] "=&r" (result2),
+          [result3] "=&r" (result3), [result4] "=&r" (result4),
+          [step1_4] "=r" (step1_4), [step1_5] "=r" (step1_5),
+          [step1_6] "=r" (step1_6), [step1_7] "=r" (step1_7)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_20_64] "r" (cospi_20_64), [cospi_12_64] "r" (cospi_12_64),
+          [cospi_4_64] "r" (cospi_4_64), [cospi_28_64] "r" (cospi_28_64),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        "sub      %[load5],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[load5],             %[load5],       %[step2_9]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_10]     \n\t"
+
+        "madd     $ac0,                 %[load5],       %[cospi_16_64]  \n\t"
+
+        "sub      %[load6],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[load6],             %[load6],       %[step2_10]     \n\t"
+        "add      %[load6],             %[load6],       %[step2_9]      \n\t"
+
+        "madd     $ac1,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load5],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[load5],             %[load5],       %[step2_8]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_11]     \n\t"
+
+        "madd     $ac2,                 %[load5],       %[cospi_16_64]  \n\t"
+
+        "sub      %[load6],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[load6],             %[load6],       %[step2_11]     \n\t"
+        "add      %[load6],             %[load6],       %[step2_8]      \n\t"
+
+        "madd     $ac3,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_10],          $ac0,           31              \n\t"
+        "extp     %[step1_13],          $ac1,           31              \n\t"
+        "extp     %[step1_11],          $ac2,           31              \n\t"
+        "extp     %[step1_12],          $ac3,           31              \n\t"
+
+        : [load5] "=&r" (load5), [load6] "=&r" (load6),
+          [step1_10] "=r" (step1_10), [step1_11] "=r" (step1_11),
+          [step1_12] "=r" (step1_12), [step1_13] "=r" (step1_13)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_14] "r" (step2_14), [step2_13] "r" (step2_13),
+          [step2_9] "r" (step2_9), [step2_10] "r" (step2_10),
+          [step2_15] "r" (step2_15), [step2_12] "r" (step2_12),
+          [step2_8] "r" (step2_8), [step2_11] "r" (step2_11),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    step1_8 = step2_8 + step2_11;
+    step1_9 = step2_9 + step2_10;
+    step1_14 = step2_13 + step2_14;
+    step1_15 = step2_12 + step2_15;
+
+    __asm__ __volatile__ (
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "add      %[load5],         %[step1_0],         %[step1_7]      \n\t"
+        "add      %[load5],         %[load5],           %[step1_15]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "add      %[load6],         %[step1_1],         %[step1_6]      \n\t"
+        "add      %[load6],         %[load6],           %[step1_14]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "add      %[load5],         %[step1_2],         %[step1_5]      \n\t"
+        "add      %[load5],         %[load5],           %[step1_13]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "add      %[load6],         %[step1_3],         %[step1_4]      \n\t"
+        "add      %[load6],         %[load6],           %[step1_12]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "sub      %[load5],         %[step1_3],         %[step1_4]      \n\t"
+        "add      %[load5],         %[load5],           %[step1_11]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "sub      %[load6],         %[step1_2],         %[step1_5]      \n\t"
+        "add      %[load6],         %[load6],           %[step1_10]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "sub      %[load5],         %[step1_1],         %[step1_6]      \n\t"
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "add      %[load5],         %[load5],           %[step1_9]      \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "sub      %[load6],         %[step1_0],         %[step1_7]      \n\t"
+        "add      %[load6],         %[load6],           %[step1_8]      \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "sub      %[load5],         %[step1_0],         %[step1_7]      \n\t"
+        "sub      %[load5],         %[load5],           %[step1_8]      \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "sub      %[load6],         %[step1_1],         %[step1_6]      \n\t"
+        "sub      %[load6],         %[load6],           %[step1_9]      \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "sub      %[load5],         %[step1_2],         %[step1_5]      \n\t"
+        "sub      %[load5],         %[load5],           %[step1_10]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "sub      %[load6],         %[step1_3],         %[step1_4]      \n\t"
+        "sub      %[load6],         %[load6],           %[step1_11]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "add      %[load5],         %[step1_3],         %[step1_4]      \n\t"
+        "sub      %[load5],         %[load5],           %[step1_12]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "add      %[load6],         %[step1_2],         %[step1_5]      \n\t"
+        "sub      %[load6],         %[load6],           %[step1_13]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "add      %[load5],         %[step1_1],         %[step1_6]      \n\t"
+        "sub      %[load5],         %[load5],           %[step1_14]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "add      %[load6],         %[step1_0],         %[step1_7]      \n\t"
+        "sub      %[load6],         %[load6],           %[step1_15]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+
+        : [load5] "=&r" (load5), [load6] "=&r" (load6), [load7] "=&r" (load7),
+          [load8] "=&r" (load8), [dest_pix] "+r" (dest_pix)
+        : [cm] "r" (cm), [dest_stride] "r" (dest_stride),
+          [step1_0] "r" (step1_0), [step1_1] "r" (step1_1),
+          [step1_2] "r" (step1_2), [step1_3] "r" (step1_3),
+          [step1_4] "r" (step1_4), [step1_5] "r" (step1_5),
+          [step1_6] "r" (step1_6), [step1_7] "r" (step1_7),
+          [step1_8] "r" (step1_8), [step1_9] "r" (step1_9),
+          [step1_10] "r" (step1_10), [step1_11] "r" (step1_11),
+          [step1_12] "r" (step1_12), [step1_13] "r" (step1_13),
+          [step1_14] "r" (step1_14), [step1_15] "r" (step1_15)
+    );
+
+    input += 16;
+  }
+}
+
+void vpx_idct16x16_256_add_dspr2(const int16_t *input, uint8_t *dest,
+                                 int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t,  out[16 * 16]);
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp    %[pos],    1    \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  // First transform rows
+  idct16_rows_dspr2(input, out, 16);
+
+  // Then transform columns and add to dest
+  idct16_cols_add_blk_dspr2(out, dest, dest_stride);
+}
+
+void vpx_idct16x16_10_add_dspr2(const int16_t *input, uint8_t *dest,
+                                int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t,  out[16 * 16]);
+  int16_t *outptr = out;
+  uint32_t i;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp    %[pos],    1    \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  // First transform rows. Since all non-zero dct coefficients are in
+  // upper-left 4x4 area, we only need to calculate first 4 rows here.
+  idct16_rows_dspr2(input, outptr, 4);
+
+  outptr += 4;
+  for (i = 0; i < 6; ++i) {
+    __asm__ __volatile__ (
+        "sw     $zero,    0(%[outptr])     \n\t"
+        "sw     $zero,   32(%[outptr])     \n\t"
+        "sw     $zero,   64(%[outptr])     \n\t"
+        "sw     $zero,   96(%[outptr])     \n\t"
+        "sw     $zero,  128(%[outptr])     \n\t"
+        "sw     $zero,  160(%[outptr])     \n\t"
+        "sw     $zero,  192(%[outptr])     \n\t"
+        "sw     $zero,  224(%[outptr])     \n\t"
+        "sw     $zero,  256(%[outptr])     \n\t"
+        "sw     $zero,  288(%[outptr])     \n\t"
+        "sw     $zero,  320(%[outptr])     \n\t"
+        "sw     $zero,  352(%[outptr])     \n\t"
+        "sw     $zero,  384(%[outptr])     \n\t"
+        "sw     $zero,  416(%[outptr])     \n\t"
+        "sw     $zero,  448(%[outptr])     \n\t"
+        "sw     $zero,  480(%[outptr])     \n\t"
+
+        :
+        : [outptr] "r" (outptr)
+    );
+
+    outptr += 2;
+  }
+
+  // Then transform columns
+  idct16_cols_add_blk_dspr2(out, dest, dest_stride);
+}
+
+void vpx_idct16x16_1_add_dspr2(const int16_t *input, uint8_t *dest,
+                               int dest_stride) {
+  uint32_t pos = 45;
+  int32_t out;
+  int32_t r;
+  int32_t a1, absa1;
+  int32_t vector_a1;
+  int32_t t1, t2, t3, t4;
+  int32_t vector_1, vector_2, vector_3, vector_4;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+
+    :
+    : [pos] "r" (pos)
+  );
+
+  out = DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input[0]);
+  __asm__ __volatile__ (
+      "addi     %[out],     %[out],     32      \n\t"
+      "sra      %[a1],      %[out],     6       \n\t"
+
+      : [out] "+r" (out), [a1] "=r" (a1)
+      :
+  );
+
+  if (a1 < 0) {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__ (
+        "abs        %[absa1],       %[a1]       \n\t"
+        "replv.qb   %[vector_a1],   %[absa1]    \n\t"
+
+        : [absa1] "=r" (absa1), [vector_a1] "=r" (vector_a1)
+        : [a1] "r" (a1)
+    );
+
+    for (r = 16; r--;) {
+      __asm__ __volatile__ (
+          "lw             %[t1],          0(%[dest])                      \n\t"
+          "lw             %[t2],          4(%[dest])                      \n\t"
+          "lw             %[t3],          8(%[dest])                      \n\t"
+          "lw             %[t4],          12(%[dest])                     \n\t"
+          "subu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    0(%[dest])                      \n\t"
+          "sw             %[vector_2],    4(%[dest])                      \n\t"
+          "sw             %[vector_3],    8(%[dest])                      \n\t"
+          "sw             %[vector_4],    12(%[dest])                     \n\t"
+          "add            %[dest],        %[dest],        %[dest_stride]  \n\t"
+
+          : [t1] "=&r" (t1), [t2] "=&r" (t2), [t3] "=&r" (t3), [t4] "=&r" (t4),
+            [vector_1] "=&r" (vector_1), [vector_2] "=&r" (vector_2),
+            [vector_3] "=&r" (vector_3), [vector_4] "=&r" (vector_4),
+            [dest] "+&r" (dest)
+          : [dest_stride] "r" (dest_stride), [vector_a1] "r" (vector_a1)
+      );
+    }
+  } else {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__ (
+        "replv.qb   %[vector_a1],   %[a1]   \n\t"
+
+        : [vector_a1] "=r" (vector_a1)
+        : [a1] "r" (a1)
+    );
+
+    for (r = 16; r--;) {
+      __asm__ __volatile__ (
+          "lw             %[t1],          0(%[dest])                      \n\t"
+          "lw             %[t2],          4(%[dest])                      \n\t"
+          "lw             %[t3],          8(%[dest])                      \n\t"
+          "lw             %[t4],          12(%[dest])                     \n\t"
+          "addu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    0(%[dest])                      \n\t"
+          "sw             %[vector_2],    4(%[dest])                      \n\t"
+          "sw             %[vector_3],    8(%[dest])                      \n\t"
+          "sw             %[vector_4],    12(%[dest])                     \n\t"
+          "add            %[dest],        %[dest],        %[dest_stride]  \n\t"
+
+          : [t1] "=&r" (t1), [t2] "=&r" (t2), [t3] "=&r" (t3), [t4] "=&r" (t4),
+            [vector_1] "=&r" (vector_1), [vector_2] "=&r" (vector_2),
+            [vector_3] "=&r" (vector_3), [vector_4] "=&r" (vector_4),
+            [dest] "+&r" (dest)
+          : [dest_stride] "r" (dest_stride), [vector_a1] "r" (vector_a1)
+      );
+    }
+  }
+}
+
+void iadst16_dspr2(const int16_t *input, int16_t *output) {
+  int s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14, s15;
+
+  int x0 = input[15];
+  int x1 = input[0];
+  int x2 = input[13];
+  int x3 = input[2];
+  int x4 = input[11];
+  int x5 = input[4];
+  int x6 = input[9];
+  int x7 = input[6];
+  int x8 = input[7];
+  int x9 = input[8];
+  int x10 = input[5];
+  int x11 = input[10];
+  int x12 = input[3];
+  int x13 = input[12];
+  int x14 = input[1];
+  int x15 = input[14];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7 | x8
+           | x9 | x10 | x11 | x12 | x13 | x14 | x15)) {
+    output[0] = output[1] = output[2] = output[3] = output[4]
+              = output[5] = output[6] = output[7] = output[8]
+              = output[9] = output[10] = output[11] = output[12]
+              = output[13] = output[14] = output[15] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = x0 * cospi_1_64  + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64  + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64  + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64  - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64  - x15 * cospi_29_64;
+
+  x0 = dct_const_round_shift(s0 + s8);
+  x1 = dct_const_round_shift(s1 + s9);
+  x2 = dct_const_round_shift(s2 + s10);
+  x3 = dct_const_round_shift(s3 + s11);
+  x4 = dct_const_round_shift(s4 + s12);
+  x5 = dct_const_round_shift(s5 + s13);
+  x6 = dct_const_round_shift(s6 + s14);
+  x7 = dct_const_round_shift(s7 + s15);
+  x8  = dct_const_round_shift(s0 - s8);
+  x9  = dct_const_round_shift(s1 - s9);
+  x10 = dct_const_round_shift(s2 - s10);
+  x11 = dct_const_round_shift(s3 - s11);
+  x12 = dct_const_round_shift(s4 - s12);
+  x13 = dct_const_round_shift(s5 - s13);
+  x14 = dct_const_round_shift(s6 - s14);
+  x15 = dct_const_round_shift(s7 - s15);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 =    x8 * cospi_4_64   + x9 * cospi_28_64;
+  s9 =    x8 * cospi_28_64  - x9 * cospi_4_64;
+  s10 =   x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 =   x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = - x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 =   x12 * cospi_4_64  + x13 * cospi_28_64;
+  s14 = - x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 =   x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = s0 + s4;
+  x1 = s1 + s5;
+  x2 = s2 + s6;
+  x3 = s3 + s7;
+  x4 = s0 - s4;
+  x5 = s1 - s5;
+  x6 = s2 - s6;
+  x7 = s3 - s7;
+  x8 = dct_const_round_shift(s8 + s12);
+  x9 = dct_const_round_shift(s9 + s13);
+  x10 = dct_const_round_shift(s10 + s14);
+  x11 = dct_const_round_shift(s11 + s15);
+  x12 = dct_const_round_shift(s8 - s12);
+  x13 = dct_const_round_shift(s9 - s13);
+  x14 = dct_const_round_shift(s10 - s14);
+  x15 = dct_const_round_shift(s11 - s15);
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64  + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = - x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 =   x6 * cospi_8_64  + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64  + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = - x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 =   x14 * cospi_8_64  + x15 * cospi_24_64;
+
+  x0 = s0 + s2;
+  x1 = s1 + s3;
+  x2 = s0 - s2;
+  x3 = s1 - s3;
+  x4 = dct_const_round_shift(s4 + s6);
+  x5 = dct_const_round_shift(s5 + s7);
+  x6 = dct_const_round_shift(s4 - s6);
+  x7 = dct_const_round_shift(s5 - s7);
+  x8 = s8 + s10;
+  x9 = s9 + s11;
+  x10 = s8 - s10;
+  x11 = s9 - s11;
+  x12 = dct_const_round_shift(s12 + s14);
+  x13 = dct_const_round_shift(s13 + s15);
+  x14 = dct_const_round_shift(s12 - s14);
+  x15 = dct_const_round_shift(s13 - s15);
+
+  // stage 4
+  s2 = (- cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (- x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (- x10 + x11);
+  s14 = (- cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = dct_const_round_shift(s2);
+  x3 = dct_const_round_shift(s3);
+  x6 = dct_const_round_shift(s6);
+  x7 = dct_const_round_shift(s7);
+  x10 = dct_const_round_shift(s10);
+  x11 = dct_const_round_shift(s11);
+  x14 = dct_const_round_shift(s14);
+  x15 = dct_const_round_shift(s15);
+
+  output[0] =  x0;
+  output[1] = -x8;
+  output[2] =  x12;
+  output[3] = -x4;
+  output[4] =  x6;
+  output[5] =  x14;
+  output[6] =  x10;
+  output[7] =  x2;
+  output[8] =  x3;
+  output[9] =  x11;
+  output[10] =  x15;
+  output[11] =  x7;
+  output[12] =  x5;
+  output[13] = -x13;
+  output[14] =  x9;
+  output[15] = -x1;
+}
+
+
+#endif  // HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/itrans32_cols_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/itrans32_cols_dspr2.c
new file mode 100644
index 0000000..553acb0
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/itrans32_cols_dspr2.c
@@ -0,0 +1,1068 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "vpx_dsp/mips/inv_txfm_dspr2.h"
+#include "vpx_dsp/txfm_common.h"
+
+#if HAVE_DSPR2
+void vpx_idct32_cols_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                   int dest_stride) {
+  int16_t step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6;
+  int16_t step1_7, step1_8, step1_9, step1_10, step1_11, step1_12, step1_13;
+  int16_t step1_14, step1_15, step1_16, step1_17, step1_18, step1_19;
+  int16_t step1_20, step1_21, step1_22, step1_23, step1_24, step1_25, step1_26;
+  int16_t step1_27, step1_28, step1_29, step1_30, step1_31;
+  int16_t step2_0, step2_1, step2_2, step2_3, step2_4, step2_5, step2_6;
+  int16_t step2_7, step2_8, step2_9, step2_10, step2_11, step2_12, step2_13;
+  int16_t step2_14, step2_15, step2_16, step2_17, step2_18, step2_19, step2_20;
+  int16_t step2_21, step2_22, step2_23, step2_24, step2_25, step2_26, step2_27;
+  int16_t step2_28, step2_29, step2_30, step2_31;
+  int16_t step3_8, step3_9, step3_10, step3_11, step3_12, step3_13, step3_14;
+  int16_t step3_15, step3_16, step3_17, step3_18, step3_19, step3_20, step3_21;
+  int16_t step3_22, step3_23, step3_24, step3_25, step3_26, step3_27;
+  int16_t step3_28, step3_29, step3_30, step3_31;
+  int temp0, temp1, temp2, temp3;
+  int load1, load2, load3, load4;
+  int result1, result2;
+  int i, temp21;
+  uint8_t *dest_pix, *dest_pix1;
+  const int const_2_power_13 = 8192;
+  uint8_t *cm = vpx_ff_cropTbl;
+
+  /* prefetch vpx_ff_cropTbl */
+  prefetch_load(vpx_ff_cropTbl);
+  prefetch_load(vpx_ff_cropTbl +  32);
+  prefetch_load(vpx_ff_cropTbl +  64);
+  prefetch_load(vpx_ff_cropTbl +  96);
+  prefetch_load(vpx_ff_cropTbl + 128);
+  prefetch_load(vpx_ff_cropTbl + 160);
+  prefetch_load(vpx_ff_cropTbl + 192);
+  prefetch_load(vpx_ff_cropTbl + 224);
+
+  for (i = 0; i < 32; ++i) {
+    dest_pix = dest + i;
+    dest_pix1 = dest + i + 31 * dest_stride;
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             2(%[input])                     \n\t"
+        "lh       %[load2],             62(%[input])                    \n\t"
+        "lh       %[load3],             34(%[input])                    \n\t"
+        "lh       %[load4],             30(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_31_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_1_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_1_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_31_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_15_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_17_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_17_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_15_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp3],       %[temp2]        \n\t"
+        "sub      %[load2],             %[temp0],       %[temp1]        \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+
+        "extp     %[step1_17],          $ac1,           31              \n\t"
+        "extp     %[step1_30],          $ac3,           31              \n\t"
+        "add      %[step1_16],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_31],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2), [load3] "=&r" (load3),
+          [load4] "=&r" (load4), [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_16] "=r" (step1_16), [step1_17] "=r" (step1_17),
+          [step1_30] "=r" (step1_30), [step1_31] "=r" (step1_31)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_31_64] "r" (cospi_31_64), [cospi_1_64] "r" (cospi_1_64),
+          [cospi_4_64] "r" (cospi_4_64), [cospi_17_64] "r" (cospi_17_64),
+          [cospi_15_64] "r" (cospi_15_64), [cospi_28_64] "r" (cospi_28_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             18(%[input])                    \n\t"
+        "lh       %[load2],             46(%[input])                    \n\t"
+        "lh       %[load3],             50(%[input])                    \n\t"
+        "lh       %[load4],             14(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_23_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_9_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_9_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_23_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_7_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_25_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_25_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_7_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+
+        "extp     %[step1_18],          $ac1,           31              \n\t"
+        "extp     %[step1_29],          $ac3,           31              \n\t"
+        "add      %[step1_19],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_28],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2), [load3] "=&r" (load3),
+          [load4] "=&r" (load4), [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_18] "=r" (step1_18), [step1_19] "=r" (step1_19),
+          [step1_28] "=r" (step1_28), [step1_29] "=r" (step1_29)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_23_64] "r" (cospi_23_64), [cospi_9_64] "r" (cospi_9_64),
+          [cospi_4_64] "r" (cospi_4_64), [cospi_7_64] "r" (cospi_7_64),
+          [cospi_25_64] "r" (cospi_25_64), [cospi_28_64] "r" (cospi_28_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             10(%[input])                    \n\t"
+        "lh       %[load2],             54(%[input])                    \n\t"
+        "lh       %[load3],             42(%[input])                    \n\t"
+        "lh       %[load4],             22(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_27_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_5_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_5_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_27_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_11_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_21_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_21_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_11_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[temp3],       %[temp2]        \n\t"
+
+        "madd     $ac1,                 %[load2],       %[cospi_12_64]  \n\t"
+        "msub     $ac1,                 %[load1],       %[cospi_20_64]  \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_12_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_20_64]  \n\t"
+
+        "extp     %[step1_21],          $ac1,           31              \n\t"
+        "extp     %[step1_26],          $ac3,           31              \n\t"
+        "add      %[step1_20],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_27],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2), [load3] "=&r" (load3),
+          [load4] "=&r" (load4), [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_20] "=r" (step1_20), [step1_21] "=r" (step1_21),
+          [step1_26] "=r" (step1_26), [step1_27] "=r" (step1_27)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_27_64] "r" (cospi_27_64), [cospi_5_64] "r" (cospi_5_64),
+          [cospi_11_64] "r" (cospi_11_64), [cospi_21_64] "r" (cospi_21_64),
+          [cospi_12_64] "r" (cospi_12_64), [cospi_20_64] "r" (cospi_20_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             26(%[input])                    \n\t"
+        "lh       %[load2],             38(%[input])                    \n\t"
+        "lh       %[load3],             58(%[input])                    \n\t"
+        "lh       %[load4],              6(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_19_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_13_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_13_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_19_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_3_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_29_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_29_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_3_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+        "msub     $ac1,                 %[load1],       %[cospi_12_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_20_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_20_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_12_64]  \n\t"
+        "extp     %[step1_22],          $ac1,           31              \n\t"
+        "extp     %[step1_25],          $ac3,           31              \n\t"
+        "add      %[step1_23],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_24],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2), [load3] "=&r" (load3),
+          [load4] "=&r" (load4), [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_22] "=r" (step1_22), [step1_23] "=r" (step1_23),
+          [step1_24] "=r" (step1_24), [step1_25] "=r" (step1_25)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_19_64] "r" (cospi_19_64), [cospi_13_64] "r" (cospi_13_64),
+          [cospi_3_64] "r" (cospi_3_64), [cospi_29_64] "r" (cospi_29_64),
+          [cospi_12_64] "r" (cospi_12_64), [cospi_20_64] "r" (cospi_20_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],              4(%[input])                    \n\t"
+        "lh       %[load2],             60(%[input])                    \n\t"
+        "lh       %[load3],             36(%[input])                    \n\t"
+        "lh       %[load4],             28(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_30_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_2_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_2_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_30_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_14_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_18_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_18_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_14_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[temp3],       %[temp2]        \n\t"
+        "msub     $ac1,                 %[load1],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load2],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_8_64]   \n\t"
+        "extp     %[step2_9],           $ac1,           31              \n\t"
+        "extp     %[step2_14],          $ac3,           31              \n\t"
+        "add      %[step2_8],           %[temp0],       %[temp1]        \n\t"
+        "add      %[step2_15],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2), [load3] "=&r" (load3),
+          [load4] "=&r" (load4), [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step2_8] "=r" (step2_8), [step2_9] "=r" (step2_9),
+          [step2_14] "=r" (step2_14), [step2_15] "=r" (step2_15)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_30_64] "r" (cospi_30_64), [cospi_2_64] "r" (cospi_2_64),
+          [cospi_14_64] "r" (cospi_14_64), [cospi_18_64] "r" (cospi_18_64),
+          [cospi_8_64] "r" (cospi_8_64), [cospi_24_64] "r" (cospi_24_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             20(%[input])                    \n\t"
+        "lh       %[load2],             44(%[input])                    \n\t"
+        "lh       %[load3],             52(%[input])                    \n\t"
+        "lh       %[load4],             12(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_22_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_10_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_10_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_22_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_6_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_26_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_26_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_6_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+        "msub     $ac1,                 %[load1],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_8_64]   \n\t"
+        "extp     %[step2_10],          $ac1,           31              \n\t"
+        "extp     %[step2_13],          $ac3,           31              \n\t"
+        "add      %[step2_11],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step2_12],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2), [load3] "=&r" (load3),
+          [load4] "=&r" (load4), [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step2_10] "=r" (step2_10), [step2_11] "=r" (step2_11),
+          [step2_12] "=r" (step2_12), [step2_13] "=r" (step2_13)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_22_64] "r" (cospi_22_64), [cospi_10_64] "r" (cospi_10_64),
+          [cospi_6_64] "r" (cospi_6_64), [cospi_26_64] "r" (cospi_26_64),
+          [cospi_8_64] "r" (cospi_8_64), [cospi_24_64] "r" (cospi_24_64)
+    );
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "sub      %[temp0],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[temp0],             %[temp0],       %[step2_9]      \n\t"
+        "add      %[temp0],             %[temp0],       %[step2_10]     \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "sub      %[temp1],             %[step2_14],    %[step2_13]     \n\t"
+        "add      %[temp1],             %[temp1],       %[step2_9]      \n\t"
+        "sub      %[temp1],             %[temp1],       %[step2_10]     \n\t"
+        "madd     $ac1,                 %[temp1],       %[cospi_16_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "sub      %[temp0],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[temp0],             %[temp0],       %[step2_8]      \n\t"
+        "add      %[temp0],             %[temp0],       %[step2_11]     \n\t"
+        "madd     $ac2,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "sub      %[temp1],             %[step2_15],    %[step2_12]     \n\t"
+        "add      %[temp1],             %[temp1],       %[step2_8]      \n\t"
+        "sub      %[temp1],             %[temp1],       %[step2_11]     \n\t"
+        "madd     $ac3,                 %[temp1],       %[cospi_16_64]  \n\t"
+
+        "add      %[step3_8],           %[step2_8],     %[step2_11]     \n\t"
+        "add      %[step3_9],           %[step2_9],     %[step2_10]     \n\t"
+        "add      %[step3_14],          %[step2_13],    %[step2_14]     \n\t"
+        "add      %[step3_15],          %[step2_12],    %[step2_15]     \n\t"
+        "extp     %[step3_10],          $ac0,           31              \n\t"
+        "extp     %[step3_13],          $ac1,           31              \n\t"
+        "extp     %[step3_11],          $ac2,           31              \n\t"
+        "extp     %[step3_12],          $ac3,           31              \n\t"
+
+        : [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [step3_8] "=r" (step3_8), [step3_9] "=r" (step3_9),
+          [step3_10] "=r" (step3_10), [step3_11] "=r" (step3_11),
+          [step3_12] "=r" (step3_12), [step3_13] "=r" (step3_13),
+          [step3_14] "=r" (step3_14), [step3_15] "=r" (step3_15)
+        : [const_2_power_13] "r" (const_2_power_13), [step2_8] "r" (step2_8),
+          [step2_9] "r" (step2_9), [step2_10] "r" (step2_10),
+          [step2_11] "r" (step2_11), [step2_12] "r" (step2_12),
+          [step2_13] "r" (step2_13), [step2_14] "r" (step2_14),
+          [step2_15] "r" (step2_15), [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    step2_18 = step1_17 - step1_18;
+    step2_29 = step1_30 - step1_29;
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_18],    %[cospi_8_64]   \n\t"
+        "madd     $ac0,                 %[step2_29],    %[cospi_24_64]  \n\t"
+        "extp     %[step3_18],          $ac0,           31              \n\t"
+
+        : [step3_18] "=r" (step3_18)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_18] "r" (step2_18), [step2_29] "r" (step2_29),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    temp21 = step2_18 * cospi_24_64 + step2_29 * cospi_8_64;
+    step3_29 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step2_19 = step1_16 - step1_19;
+    step2_28 = step1_31 - step1_28;
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_19],    %[cospi_8_64]   \n\t"
+        "madd     $ac0,                 %[step2_28],    %[cospi_24_64]  \n\t"
+        "extp     %[step3_19],          $ac0,           31              \n\t"
+
+        : [step3_19] "=r" (step3_19)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_19] "r" (step2_19), [step2_28] "r" (step2_28),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    temp21 = step2_19 * cospi_24_64 + step2_28 * cospi_8_64;
+    step3_28 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step3_16 = step1_16 + step1_19;
+    step3_17 = step1_17 + step1_18;
+    step3_30 = step1_29 + step1_30;
+    step3_31 = step1_28 + step1_31;
+
+    step2_20 = step1_23 - step1_20;
+    step2_27 = step1_24 - step1_27;
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_20],    %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[step2_27],    %[cospi_8_64]   \n\t"
+        "extp     %[step3_20],          $ac0,           31              \n\t"
+
+        : [step3_20] "=r" (step3_20)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_20] "r" (step2_20), [step2_27] "r" (step2_27),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    temp21 = -step2_20 * cospi_8_64 + step2_27 * cospi_24_64;
+    step3_27 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step2_21 = step1_22 - step1_21;
+    step2_26 = step1_25 - step1_26;
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "msub     $ac1,                 %[step2_21],    %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[step2_26],    %[cospi_8_64]   \n\t"
+        "extp     %[step3_21],          $ac1,           31              \n\t"
+
+        : [step3_21] "=r" (step3_21)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_21] "r" (step2_21), [step2_26] "r" (step2_26),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    temp21 = -step2_21 * cospi_8_64 + step2_26 * cospi_24_64;
+    step3_26 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step3_22 = step1_21 + step1_22;
+    step3_23 = step1_20 + step1_23;
+    step3_24 = step1_24 + step1_27;
+    step3_25 = step1_25 + step1_26;
+
+    step2_16 = step3_16 + step3_23;
+    step2_17 = step3_17 + step3_22;
+    step2_18 = step3_18 + step3_21;
+    step2_19 = step3_19 + step3_20;
+    step2_20 = step3_19 - step3_20;
+    step2_21 = step3_18 - step3_21;
+    step2_22 = step3_17 - step3_22;
+    step2_23 = step3_16 - step3_23;
+
+    step2_24 = step3_31 - step3_24;
+    step2_25 = step3_30 - step3_25;
+    step2_26 = step3_29 - step3_26;
+    step2_27 = step3_28 - step3_27;
+    step2_28 = step3_28 + step3_27;
+    step2_29 = step3_29 + step3_26;
+    step2_30 = step3_30 + step3_25;
+    step2_31 = step3_31 + step3_24;
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             0(%[input])                     \n\t"
+        "lh       %[load2],             32(%[input])                    \n\t"
+        "lh       %[load3],             16(%[input])                    \n\t"
+        "lh       %[load4],             48(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "add      %[result1],           %[load1],       %[load2]        \n\t"
+        "sub      %[result2],           %[load1],       %[load2]        \n\t"
+        "madd     $ac1,                 %[result1],     %[cospi_16_64]  \n\t"
+        "madd     $ac2,                 %[result2],     %[cospi_16_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "madd     $ac3,                 %[load3],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load4],       %[cospi_8_64]   \n\t"
+        "extp     %[temp2],             $ac3,           31              \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_24_64]  \n\t"
+        "extp     %[temp3],             $ac1,           31              \n\t"
+        "add      %[step1_0],           %[temp0],       %[temp3]        \n\t"
+        "add      %[step1_1],           %[temp1],       %[temp2]        \n\t"
+        "sub      %[step1_2],           %[temp1],       %[temp2]        \n\t"
+        "sub      %[step1_3],           %[temp0],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [result1] "=&r" (result1), [result2] "=&r" (result2),
+          [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_0] "=r" (step1_0), [step1_1] "=r" (step1_1),
+          [step1_2] "=r" (step1_2), [step1_3] "=r" (step1_3)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             8(%[input])                     \n\t"
+        "lh       %[load2],             56(%[input])                    \n\t"
+        "lh       %[load3],             40(%[input])                    \n\t"
+        "lh       %[load4],             24(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_12_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_20_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_20_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_12_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp3],       %[temp2]        \n\t"
+        "sub      %[load1],             %[load1],       %[temp0]        \n\t"
+        "add      %[load1],             %[load1],       %[temp1]        \n\t"
+        "sub      %[load2],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[load2],       %[temp2]        \n\t"
+        "add      %[load2],             %[load2],       %[temp3]        \n\t"
+        "madd     $ac1,                 %[load1],       %[cospi_16_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_5],           $ac1,           31              \n\t"
+        "extp     %[step1_6],           $ac3,           31              \n\t"
+        "add      %[step1_4],           %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_7],           %[temp3],       %[temp2]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_4] "=r" (step1_4), [step1_5] "=r" (step1_5),
+          [step1_6] "=r" (step1_6), [step1_7] "=r" (step1_7)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_20_64] "r" (cospi_20_64), [cospi_12_64] "r" (cospi_12_64),
+          [cospi_4_64] "r" (cospi_4_64), [cospi_28_64] "r" (cospi_28_64),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    step2_0 = step1_0 + step1_7;
+    step2_1 = step1_1 + step1_6;
+    step2_2 = step1_2 + step1_5;
+    step2_3 = step1_3 + step1_4;
+    step2_4 = step1_3 - step1_4;
+    step2_5 = step1_2 - step1_5;
+    step2_6 = step1_1 - step1_6;
+    step2_7 = step1_0 - step1_7;
+
+    // stage 7
+    step1_0 = step2_0 + step3_15;
+    step1_1 = step2_1 + step3_14;
+    step1_2 = step2_2 + step3_13;
+    step1_3 = step2_3 + step3_12;
+    step1_4 = step2_4 + step3_11;
+    step1_5 = step2_5 + step3_10;
+    step1_6 = step2_6 + step3_9;
+    step1_7 = step2_7 + step3_8;
+    step1_8 = step2_7 - step3_8;
+    step1_9 = step2_6 - step3_9;
+    step1_10 = step2_5 - step3_10;
+    step1_11 = step2_4 - step3_11;
+    step1_12 = step2_3 - step3_12;
+    step1_13 = step2_2 - step3_13;
+    step1_14 = step2_1 - step3_14;
+    step1_15 = step2_0 - step3_15;
+
+    __asm__ __volatile__ (
+        "sub      %[temp0],             %[step2_27],    %[step2_20]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_20],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r" (temp0), [step1_20] "=r" (step1_20)
+        : [const_2_power_13] "r" (const_2_power_13), [step2_20] "r" (step2_20),
+          [step2_27] "r" (step2_27), [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    temp21 = (step2_20 + step2_27) * cospi_16_64;
+    step1_27 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__ (
+        "sub      %[temp0],             %[step2_26],    %[step2_21]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_21],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r" (temp0), [step1_21] "=r" (step1_21)
+        : [const_2_power_13] "r" (const_2_power_13), [step2_26] "r" (step2_26),
+          [step2_21] "r" (step2_21), [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    temp21 = (step2_21 + step2_26) * cospi_16_64;
+    step1_26 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__ (
+        "sub      %[temp0],             %[step2_25],    %[step2_22]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_22],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r" (temp0), [step1_22] "=r" (step1_22)
+        : [const_2_power_13] "r" (const_2_power_13), [step2_25] "r" (step2_25),
+          [step2_22] "r" (step2_22), [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    temp21 = (step2_22 + step2_25) * cospi_16_64;
+    step1_25 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__ (
+        "sub      %[temp0],             %[step2_24],    %[step2_23]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_23],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r" (temp0), [step1_23] "=r" (step1_23)
+        : [const_2_power_13] "r" (const_2_power_13), [step2_24] "r" (step2_24),
+          [step2_23] "r" (step2_23), [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    temp21 = (step2_23 + step2_24) * cospi_16_64;
+    step1_24 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__ (
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_0],         %[step2_31]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_1],         %[step2_30]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_2],         %[step2_29]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_3],         %[step2_28]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        : [temp0] "=&r" (temp0), [temp1] "=&r" (temp1), [temp2] "=&r" (temp2),
+          [temp3] "=&r" (temp3), [dest_pix] "+r" (dest_pix)
+        : [cm] "r" (cm), [dest_stride] "r" (dest_stride),
+          [step1_0] "r" (step1_0), [step1_1] "r" (step1_1),
+          [step1_2] "r" (step1_2), [step1_3] "r" (step1_3),
+          [step2_28] "r" (step2_28), [step2_29] "r" (step2_29),
+          [step2_30] "r" (step2_30), [step2_31] "r" (step2_31)
+    );
+
+    step3_12 = ROUND_POWER_OF_TWO((step1_3 - step2_28), 6);
+    step3_13 = ROUND_POWER_OF_TWO((step1_2 - step2_29), 6);
+    step3_14 = ROUND_POWER_OF_TWO((step1_1 - step2_30), 6);
+    step3_15 = ROUND_POWER_OF_TWO((step1_0 - step2_31), 6);
+
+    __asm__ __volatile__ (
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_15]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_14]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_13]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_12]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        : [temp0] "=&r" (temp0), [temp1] "=&r" (temp1), [temp2] "=&r" (temp2),
+          [temp3] "=&r" (temp3), [dest_pix1] "+r" (dest_pix1)
+        : [cm] "r" (cm), [dest_stride] "r" (dest_stride),
+          [step3_12] "r" (step3_12), [step3_13] "r" (step3_13),
+          [step3_14] "r" (step3_14), [step3_15] "r" (step3_15)
+    );
+
+    __asm__ __volatile__ (
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_4],         %[step1_27]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_5],         %[step1_26]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_6],         %[step1_25]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_7],         %[step1_24]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        : [temp0] "=&r" (temp0), [temp1] "=&r" (temp1), [temp2] "=&r" (temp2),
+          [temp3] "=&r" (temp3), [dest_pix] "+r" (dest_pix)
+        : [cm] "r" (cm), [dest_stride] "r" (dest_stride),
+          [step1_4] "r" (step1_4), [step1_5] "r" (step1_5),
+          [step1_6] "r" (step1_6), [step1_7] "r" (step1_7),
+          [step1_24] "r" (step1_24), [step1_25] "r" (step1_25),
+          [step1_26] "r" (step1_26), [step1_27] "r" (step1_27)
+    );
+
+    step3_12 = ROUND_POWER_OF_TWO((step1_7 - step1_24), 6);
+    step3_13 = ROUND_POWER_OF_TWO((step1_6 - step1_25), 6);
+    step3_14 = ROUND_POWER_OF_TWO((step1_5 - step1_26), 6);
+    step3_15 = ROUND_POWER_OF_TWO((step1_4 - step1_27), 6);
+
+    __asm__ __volatile__ (
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_15]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_14]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_13]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_12]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        : [temp0] "=&r" (temp0), [temp1] "=&r" (temp1), [temp2] "=&r" (temp2),
+          [temp3] "=&r" (temp3), [dest_pix1] "+r" (dest_pix1)
+        : [cm] "r" (cm), [dest_stride] "r" (dest_stride),
+          [step3_12] "r" (step3_12), [step3_13] "r" (step3_13),
+          [step3_14] "r" (step3_14), [step3_15] "r" (step3_15)
+    );
+
+    __asm__ __volatile__ (
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_8],         %[step1_23]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_9],         %[step1_22]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_10],        %[step1_21]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_11],        %[step1_20]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        : [temp0] "=&r" (temp0), [temp1] "=&r" (temp1), [temp2] "=&r" (temp2),
+          [temp3] "=&r" (temp3), [dest_pix] "+r" (dest_pix)
+        : [cm] "r" (cm), [dest_stride] "r" (dest_stride),
+          [step1_8] "r" (step1_8), [step1_9] "r" (step1_9),
+          [step1_10] "r" (step1_10), [step1_11] "r" (step1_11),
+          [step1_20] "r" (step1_20), [step1_21] "r" (step1_21),
+          [step1_22] "r" (step1_22), [step1_23] "r" (step1_23)
+    );
+
+    step3_12 = ROUND_POWER_OF_TWO((step1_11 - step1_20), 6);
+    step3_13 = ROUND_POWER_OF_TWO((step1_10 - step1_21), 6);
+    step3_14 = ROUND_POWER_OF_TWO((step1_9 - step1_22), 6);
+    step3_15 = ROUND_POWER_OF_TWO((step1_8 - step1_23), 6);
+
+    __asm__ __volatile__ (
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_15]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_14]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_13]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_12]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        : [temp0] "=&r" (temp0), [temp1] "=&r" (temp1), [temp2] "=&r" (temp2),
+          [temp3] "=&r" (temp3), [dest_pix1] "+r" (dest_pix1)
+        : [cm] "r" (cm), [dest_stride] "r" (dest_stride),
+          [step3_12] "r" (step3_12), [step3_13] "r" (step3_13),
+          [step3_14] "r" (step3_14), [step3_15] "r" (step3_15)
+    );
+
+    __asm__ __volatile__ (
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_12],        %[step2_19]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_13],        %[step2_18]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_14],        %[step2_17]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_15],        %[step2_16]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+
+        : [temp0] "=&r" (temp0), [temp1] "=&r" (temp1), [temp2] "=&r" (temp2),
+          [temp3] "=&r" (temp3), [dest_pix] "+r" (dest_pix)
+        : [cm] "r" (cm), [dest_stride] "r" (dest_stride),
+          [step1_12] "r" (step1_12), [step1_13] "r" (step1_13),
+          [step1_14] "r" (step1_14), [step1_15] "r" (step1_15),
+          [step2_16] "r" (step2_16), [step2_17] "r" (step2_17),
+          [step2_18] "r" (step2_18), [step2_19] "r" (step2_19)
+    );
+
+    step3_12 = ROUND_POWER_OF_TWO((step1_15 - step2_16), 6);
+    step3_13 = ROUND_POWER_OF_TWO((step1_14 - step2_17), 6);
+    step3_14 = ROUND_POWER_OF_TWO((step1_13 - step2_18), 6);
+    step3_15 = ROUND_POWER_OF_TWO((step1_12 - step2_19), 6);
+
+    __asm__ __volatile__ (
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_15]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_14]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_13]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_12]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+
+        : [temp0] "=&r" (temp0), [temp1] "=&r" (temp1), [temp2] "=&r" (temp2),
+          [temp3] "=&r" (temp3), [dest_pix1] "+r" (dest_pix1)
+        : [cm] "r" (cm), [dest_stride] "r" (dest_stride),
+          [step3_12] "r" (step3_12), [step3_13] "r" (step3_13),
+          [step3_14] "r" (step3_14), [step3_15] "r" (step3_15)
+    );
+
+    input += 32;
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/itrans32_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/itrans32_dspr2.c
new file mode 100644
index 0000000..523da1d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/itrans32_dspr2.c
@@ -0,0 +1,1073 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./vpx_config.h"
+#include "vpx_dsp/mips/inv_txfm_dspr2.h"
+#include "vpx_dsp/txfm_common.h"
+
+#if HAVE_DSPR2
+static void idct32_rows_dspr2(const int16_t *input, int16_t *output,
+                              uint32_t no_rows) {
+  int16_t step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6;
+  int16_t step1_7, step1_8, step1_9, step1_10, step1_11, step1_12, step1_13;
+  int16_t step1_14, step1_15, step1_16, step1_17, step1_18, step1_19, step1_20;
+  int16_t step1_21, step1_22, step1_23, step1_24, step1_25, step1_26, step1_27;
+  int16_t step1_28, step1_29, step1_30, step1_31;
+  int16_t step2_0, step2_1, step2_2, step2_3, step2_4, step2_5, step2_6;
+  int16_t step2_7, step2_8, step2_9, step2_10, step2_11, step2_12, step2_13;
+  int16_t step2_14, step2_15, step2_16, step2_17, step2_18, step2_19, step2_20;
+  int16_t step2_21, step2_22, step2_23, step2_24, step2_25, step2_26, step2_27;
+  int16_t step2_28, step2_29, step2_30, step2_31;
+  int16_t step3_8, step3_9, step3_10, step3_11, step3_12, step3_13, step3_14;
+  int16_t step3_15, step3_16, step3_17, step3_18, step3_19, step3_20, step3_21;
+  int16_t step3_22, step3_23, step3_24, step3_25, step3_26, step3_27, step3_28;
+  int16_t step3_29, step3_30, step3_31;
+  int temp0, temp1, temp2, temp3;
+  int load1, load2, load3, load4;
+  int result1, result2;
+  int temp21;
+  int i;
+  const int const_2_power_13 = 8192;
+  const int32_t *input_int;
+
+  for (i = no_rows; i--; ) {
+    input_int = (const int32_t *)input;
+
+    if (!(input_int[0]  | input_int[1]  | input_int[2]  | input_int[3]  |
+          input_int[4]  | input_int[5]  | input_int[6]  | input_int[7]  |
+          input_int[8]  | input_int[9]  | input_int[10] | input_int[11] |
+          input_int[12] | input_int[13] | input_int[14] | input_int[15])) {
+      input += 32;
+
+      __asm__ __volatile__ (
+          "sh     $zero,     0(%[output])     \n\t"
+          "sh     $zero,    64(%[output])     \n\t"
+          "sh     $zero,   128(%[output])     \n\t"
+          "sh     $zero,   192(%[output])     \n\t"
+          "sh     $zero,   256(%[output])     \n\t"
+          "sh     $zero,   320(%[output])     \n\t"
+          "sh     $zero,   384(%[output])     \n\t"
+          "sh     $zero,   448(%[output])     \n\t"
+          "sh     $zero,   512(%[output])     \n\t"
+          "sh     $zero,   576(%[output])     \n\t"
+          "sh     $zero,   640(%[output])     \n\t"
+          "sh     $zero,   704(%[output])     \n\t"
+          "sh     $zero,   768(%[output])     \n\t"
+          "sh     $zero,   832(%[output])     \n\t"
+          "sh     $zero,   896(%[output])     \n\t"
+          "sh     $zero,   960(%[output])     \n\t"
+          "sh     $zero,  1024(%[output])     \n\t"
+          "sh     $zero,  1088(%[output])     \n\t"
+          "sh     $zero,  1152(%[output])     \n\t"
+          "sh     $zero,  1216(%[output])     \n\t"
+          "sh     $zero,  1280(%[output])     \n\t"
+          "sh     $zero,  1344(%[output])     \n\t"
+          "sh     $zero,  1408(%[output])     \n\t"
+          "sh     $zero,  1472(%[output])     \n\t"
+          "sh     $zero,  1536(%[output])     \n\t"
+          "sh     $zero,  1600(%[output])     \n\t"
+          "sh     $zero,  1664(%[output])     \n\t"
+          "sh     $zero,  1728(%[output])     \n\t"
+          "sh     $zero,  1792(%[output])     \n\t"
+          "sh     $zero,  1856(%[output])     \n\t"
+          "sh     $zero,  1920(%[output])     \n\t"
+          "sh     $zero,  1984(%[output])     \n\t"
+
+          :
+          : [output] "r" (output)
+      );
+
+      output += 1;
+
+      continue;
+    }
+
+    /* prefetch row */
+    prefetch_load((const uint8_t *)(input + 32));
+    prefetch_load((const uint8_t *)(input + 48));
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             2(%[input])                     \n\t"
+        "lh       %[load2],             62(%[input])                    \n\t"
+        "lh       %[load3],             34(%[input])                    \n\t"
+        "lh       %[load4],             30(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_31_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_1_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_1_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_31_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_15_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_17_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_17_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_15_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp3],       %[temp2]        \n\t"
+        "sub      %[load2],             %[temp0],       %[temp1]        \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+
+        "extp     %[step1_17],          $ac1,           31              \n\t"
+        "extp     %[step1_30],          $ac3,           31              \n\t"
+        "add      %[step1_16],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_31],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_16] "=r" (step1_16), [step1_17] "=r" (step1_17),
+          [step1_30] "=r" (step1_30), [step1_31] "=r" (step1_31)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_31_64] "r" (cospi_31_64), [cospi_1_64] "r" (cospi_1_64),
+          [cospi_4_64] "r" (cospi_4_64), [cospi_17_64] "r" (cospi_17_64),
+          [cospi_15_64] "r" (cospi_15_64), [cospi_28_64] "r" (cospi_28_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             18(%[input])                    \n\t"
+        "lh       %[load2],             46(%[input])                    \n\t"
+        "lh       %[load3],             50(%[input])                    \n\t"
+        "lh       %[load4],             14(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_23_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_9_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_9_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_23_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_7_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_25_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_25_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_7_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+
+        "extp     %[step1_18],          $ac1,           31              \n\t"
+        "extp     %[step1_29],          $ac3,           31              \n\t"
+        "add      %[step1_19],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_28],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_18] "=r" (step1_18), [step1_19] "=r" (step1_19),
+          [step1_28] "=r" (step1_28), [step1_29] "=r" (step1_29)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_23_64] "r" (cospi_23_64), [cospi_9_64] "r" (cospi_9_64),
+          [cospi_4_64] "r" (cospi_4_64), [cospi_7_64] "r" (cospi_7_64),
+          [cospi_25_64] "r" (cospi_25_64), [cospi_28_64] "r" (cospi_28_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             10(%[input])                    \n\t"
+        "lh       %[load2],             54(%[input])                    \n\t"
+        "lh       %[load3],             42(%[input])                    \n\t"
+        "lh       %[load4],             22(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_27_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_5_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_5_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_27_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_11_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_21_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_21_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_11_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[temp3],       %[temp2]        \n\t"
+
+        "madd     $ac1,                 %[load2],       %[cospi_12_64]  \n\t"
+        "msub     $ac1,                 %[load1],       %[cospi_20_64]  \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_12_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_20_64]  \n\t"
+
+        "extp     %[step1_21],          $ac1,           31              \n\t"
+        "extp     %[step1_26],          $ac3,           31              \n\t"
+        "add      %[step1_20],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_27],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_20] "=r" (step1_20), [step1_21] "=r" (step1_21),
+          [step1_26] "=r" (step1_26), [step1_27] "=r" (step1_27)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_27_64] "r" (cospi_27_64), [cospi_5_64] "r" (cospi_5_64),
+          [cospi_11_64] "r" (cospi_11_64), [cospi_21_64] "r" (cospi_21_64),
+          [cospi_12_64] "r" (cospi_12_64), [cospi_20_64] "r" (cospi_20_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             26(%[input])                    \n\t"
+        "lh       %[load2],             38(%[input])                    \n\t"
+        "lh       %[load3],             58(%[input])                    \n\t"
+        "lh       %[load4],              6(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_19_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_13_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_13_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_19_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_3_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_29_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_29_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_3_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_12_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_20_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_20_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_12_64]  \n\t"
+
+        "extp     %[step1_22],          $ac1,           31              \n\t"
+        "extp     %[step1_25],          $ac3,           31              \n\t"
+        "add      %[step1_23],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_24],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_22] "=r" (step1_22), [step1_23] "=r" (step1_23),
+          [step1_24] "=r" (step1_24), [step1_25] "=r" (step1_25)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_19_64] "r" (cospi_19_64), [cospi_13_64] "r" (cospi_13_64),
+          [cospi_3_64] "r" (cospi_3_64), [cospi_29_64] "r" (cospi_29_64),
+          [cospi_12_64] "r" (cospi_12_64), [cospi_20_64] "r" (cospi_20_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],              4(%[input])                    \n\t"
+        "lh       %[load2],             60(%[input])                    \n\t"
+        "lh       %[load3],             36(%[input])                    \n\t"
+        "lh       %[load4],             28(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_30_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_2_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_2_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_30_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_14_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_18_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_18_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_14_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[temp3],       %[temp2]        \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load2],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_9],           $ac1,           31              \n\t"
+        "extp     %[step2_14],          $ac3,           31              \n\t"
+        "add      %[step2_8],           %[temp0],       %[temp1]        \n\t"
+        "add      %[step2_15],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step2_8] "=r" (step2_8), [step2_9] "=r" (step2_9),
+          [step2_14] "=r" (step2_14), [step2_15] "=r" (step2_15)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_30_64] "r" (cospi_30_64), [cospi_2_64] "r" (cospi_2_64),
+          [cospi_14_64] "r" (cospi_14_64), [cospi_18_64] "r" (cospi_18_64),
+          [cospi_8_64] "r" (cospi_8_64), [cospi_24_64] "r" (cospi_24_64)
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             20(%[input])                    \n\t"
+        "lh       %[load2],             44(%[input])                    \n\t"
+        "lh       %[load3],             52(%[input])                    \n\t"
+        "lh       %[load4],             12(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_22_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_10_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_10_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_22_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_6_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_26_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_26_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_6_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_10],          $ac1,           31              \n\t"
+        "extp     %[step2_13],          $ac3,           31              \n\t"
+        "add      %[step2_11],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step2_12],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step2_10] "=r" (step2_10), [step2_11] "=r" (step2_11),
+          [step2_12] "=r" (step2_12), [step2_13] "=r" (step2_13)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_22_64] "r" (cospi_22_64), [cospi_10_64] "r" (cospi_10_64),
+          [cospi_6_64] "r" (cospi_6_64), [cospi_26_64] "r" (cospi_26_64),
+          [cospi_8_64] "r" (cospi_8_64), [cospi_24_64] "r" (cospi_24_64)
+    );
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "sub      %[temp0],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[temp0],             %[temp0],       %[step2_9]      \n\t"
+        "add      %[temp0],             %[temp0],       %[step2_10]     \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "sub      %[temp1],             %[step2_14],    %[step2_13]     \n\t"
+        "add      %[temp1],             %[temp1],       %[step2_9]      \n\t"
+        "sub      %[temp1],             %[temp1],       %[step2_10]     \n\t"
+        "madd     $ac1,                 %[temp1],       %[cospi_16_64]  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "sub      %[temp0],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[temp0],             %[temp0],       %[step2_8]      \n\t"
+        "add      %[temp0],             %[temp0],       %[step2_11]     \n\t"
+        "madd     $ac2,                 %[temp0],       %[cospi_16_64]  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "sub      %[temp1],             %[step2_15],    %[step2_12]     \n\t"
+        "add      %[temp1],             %[temp1],       %[step2_8]      \n\t"
+        "sub      %[temp1],             %[temp1],       %[step2_11]     \n\t"
+        "madd     $ac3,                 %[temp1],       %[cospi_16_64]  \n\t"
+
+        "add      %[step3_8],           %[step2_8],     %[step2_11]     \n\t"
+        "add      %[step3_9],           %[step2_9],     %[step2_10]     \n\t"
+        "add      %[step3_14],          %[step2_13],    %[step2_14]     \n\t"
+        "add      %[step3_15],          %[step2_12],    %[step2_15]     \n\t"
+
+        "extp     %[step3_10],          $ac0,           31              \n\t"
+        "extp     %[step3_13],          $ac1,           31              \n\t"
+        "extp     %[step3_11],          $ac2,           31              \n\t"
+        "extp     %[step3_12],          $ac3,           31              \n\t"
+
+        : [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [step3_8] "=r" (step3_8), [step3_9] "=r" (step3_9),
+          [step3_10] "=r" (step3_10), [step3_11] "=r" (step3_11),
+          [step3_12] "=r" (step3_12), [step3_13] "=r" (step3_13),
+          [step3_14] "=r" (step3_14), [step3_15] "=r" (step3_15)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_8] "r" (step2_8), [step2_9] "r" (step2_9),
+          [step2_10] "r" (step2_10), [step2_11] "r" (step2_11),
+          [step2_12] "r" (step2_12), [step2_13] "r" (step2_13),
+          [step2_14] "r" (step2_14), [step2_15] "r" (step2_15),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    step2_18 = step1_17 - step1_18;
+    step2_29 = step1_30 - step1_29;
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_18],    %[cospi_8_64]   \n\t"
+        "madd     $ac0,                 %[step2_29],    %[cospi_24_64]  \n\t"
+        "extp     %[step3_18],          $ac0,           31              \n\t"
+
+        : [step3_18] "=r" (step3_18)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_18] "r" (step2_18), [step2_29] "r" (step2_29),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    temp21 = step2_18 * cospi_24_64 + step2_29 * cospi_8_64;
+    step3_29 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step2_19 = step1_16 - step1_19;
+    step2_28 = step1_31 - step1_28;
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_19],    %[cospi_8_64]   \n\t"
+        "madd     $ac0,                 %[step2_28],    %[cospi_24_64]  \n\t"
+        "extp     %[step3_19],          $ac0,           31              \n\t"
+
+        : [step3_19] "=r" (step3_19)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_19] "r" (step2_19), [step2_28] "r" (step2_28),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    temp21 = step2_19 * cospi_24_64 + step2_28 * cospi_8_64;
+    step3_28 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step3_16 = step1_16 + step1_19;
+    step3_17 = step1_17 + step1_18;
+    step3_30 = step1_29 + step1_30;
+    step3_31 = step1_28 + step1_31;
+
+    step2_20 = step1_23 - step1_20;
+    step2_27 = step1_24 - step1_27;
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_20],    %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[step2_27],    %[cospi_8_64]   \n\t"
+        "extp     %[step3_20],          $ac0,           31              \n\t"
+
+        : [step3_20] "=r" (step3_20)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_20] "r" (step2_20), [step2_27] "r" (step2_27),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    temp21 = -step2_20 * cospi_8_64 + step2_27 * cospi_24_64;
+    step3_27 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step2_21 = step1_22 - step1_21;
+    step2_26 = step1_25 - step1_26;
+
+    __asm__ __volatile__ (
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "msub     $ac1,                 %[step2_21],    %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[step2_26],    %[cospi_8_64]   \n\t"
+        "extp     %[step3_21],          $ac1,           31              \n\t"
+
+        : [step3_21] "=r" (step3_21)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_21] "r" (step2_21), [step2_26] "r" (step2_26),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+    );
+
+    temp21 = -step2_21 * cospi_8_64 + step2_26 * cospi_24_64;
+    step3_26 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step3_22 = step1_21 + step1_22;
+    step3_23 = step1_20 + step1_23;
+    step3_24 = step1_24 + step1_27;
+    step3_25 = step1_25 + step1_26;
+
+    step2_16 = step3_16 + step3_23;
+    step2_17 = step3_17 + step3_22;
+    step2_18 = step3_18 + step3_21;
+    step2_19 = step3_19 + step3_20;
+    step2_20 = step3_19 - step3_20;
+    step2_21 = step3_18 - step3_21;
+    step2_22 = step3_17 - step3_22;
+    step2_23 = step3_16 - step3_23;
+
+    step2_24 = step3_31 - step3_24;
+    step2_25 = step3_30 - step3_25;
+    step2_26 = step3_29 - step3_26;
+    step2_27 = step3_28 - step3_27;
+    step2_28 = step3_28 + step3_27;
+    step2_29 = step3_29 + step3_26;
+    step2_30 = step3_30 + step3_25;
+    step2_31 = step3_31 + step3_24;
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             0(%[input])                     \n\t"
+        "lh       %[load2],             32(%[input])                    \n\t"
+        "lh       %[load3],             16(%[input])                    \n\t"
+        "lh       %[load4],             48(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "add      %[result1],           %[load1],       %[load2]        \n\t"
+        "sub      %[result2],           %[load1],       %[load2]        \n\t"
+        "madd     $ac1,                 %[result1],     %[cospi_16_64]  \n\t"
+        "madd     $ac2,                 %[result2],     %[cospi_16_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "madd     $ac3,                 %[load3],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load4],       %[cospi_8_64]   \n\t"
+        "extp     %[temp2],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_24_64]  \n\t"
+        "extp     %[temp3],             $ac1,           31              \n\t"
+
+        "add      %[step1_0],          %[temp0],        %[temp3]        \n\t"
+        "add      %[step1_1],          %[temp1],        %[temp2]        \n\t"
+        "sub      %[step1_2],          %[temp1],        %[temp2]        \n\t"
+        "sub      %[step1_3],          %[temp0],        %[temp3]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [result1] "=&r" (result1), [result2] "=&r" (result2),
+          [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_0] "=r" (step1_0), [step1_1] "=r" (step1_1),
+          [step1_2] "=r" (step1_2), [step1_3] "=r" (step1_3)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_16_64] "r" (cospi_16_64),
+          [cospi_24_64] "r" (cospi_24_64), [cospi_8_64] "r" (cospi_8_64)
+
+    );
+
+    __asm__ __volatile__ (
+        "lh       %[load1],             8(%[input])                     \n\t"
+        "lh       %[load2],             56(%[input])                    \n\t"
+        "lh       %[load3],             40(%[input])                    \n\t"
+        "lh       %[load4],             24(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_12_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_20_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_20_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_12_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp3],       %[temp2]        \n\t"
+        "sub      %[load1],             %[load1],       %[temp0]        \n\t"
+        "add      %[load1],             %[load1],       %[temp1]        \n\t"
+
+        "sub      %[load2],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[load2],       %[temp2]        \n\t"
+        "add      %[load2],             %[load2],       %[temp3]        \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_16_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_5],           $ac1,           31              \n\t"
+        "extp     %[step1_6],           $ac3,           31              \n\t"
+        "add      %[step1_4],           %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_7],           %[temp3],       %[temp2]        \n\t"
+
+        : [load1] "=&r" (load1), [load2] "=&r" (load2),
+          [load3] "=&r" (load3), [load4] "=&r" (load4),
+          [temp0] "=&r" (temp0), [temp1] "=&r" (temp1),
+          [temp2] "=&r" (temp2), [temp3] "=&r" (temp3),
+          [step1_4] "=r" (step1_4), [step1_5] "=r" (step1_5),
+          [step1_6] "=r" (step1_6), [step1_7] "=r" (step1_7)
+        : [const_2_power_13] "r" (const_2_power_13), [input] "r" (input),
+          [cospi_20_64] "r" (cospi_20_64), [cospi_12_64] "r" (cospi_12_64),
+          [cospi_4_64] "r" (cospi_4_64), [cospi_28_64] "r" (cospi_28_64),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    step2_0 = step1_0 + step1_7;
+    step2_1 = step1_1 + step1_6;
+    step2_2 = step1_2 + step1_5;
+    step2_3 = step1_3 + step1_4;
+    step2_4 = step1_3 - step1_4;
+    step2_5 = step1_2 - step1_5;
+    step2_6 = step1_1 - step1_6;
+    step2_7 = step1_0 - step1_7;
+
+    step1_0 = step2_0 + step3_15;
+    step1_1 = step2_1 + step3_14;
+    step1_2 = step2_2 + step3_13;
+    step1_3 = step2_3 + step3_12;
+    step1_4 = step2_4 + step3_11;
+    step1_5 = step2_5 + step3_10;
+    step1_6 = step2_6 + step3_9;
+    step1_7 = step2_7 + step3_8;
+    step1_8 = step2_7 - step3_8;
+    step1_9 = step2_6 - step3_9;
+    step1_10 = step2_5 - step3_10;
+    step1_11 = step2_4 - step3_11;
+    step1_12 = step2_3 - step3_12;
+    step1_13 = step2_2 - step3_13;
+    step1_14 = step2_1 - step3_14;
+    step1_15 = step2_0 - step3_15;
+
+    __asm__ __volatile__ (
+        "sub      %[temp0],             %[step2_27],    %[step2_20]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_20],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r" (temp0), [step1_20] "=r" (step1_20)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_20] "r" (step2_20), [step2_27] "r" (step2_27),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    temp21 = (step2_20 + step2_27) * cospi_16_64;
+    step1_27 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__ (
+        "sub      %[temp0],             %[step2_26],    %[step2_21]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_21],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r" (temp0), [step1_21] "=r" (step1_21)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_26] "r" (step2_26), [step2_21] "r" (step2_21),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    temp21 = (step2_21 + step2_26) * cospi_16_64;
+    step1_26 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__ (
+        "sub      %[temp0],             %[step2_25],    %[step2_22]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_22],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r" (temp0), [step1_22] "=r" (step1_22)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_25] "r" (step2_25), [step2_22] "r" (step2_22),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    temp21 = (step2_22 + step2_25) * cospi_16_64;
+    step1_25 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__ (
+        "sub      %[temp0],             %[step2_24],    %[step2_23]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_23],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r" (temp0), [step1_23] "=r" (step1_23)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [step2_24] "r" (step2_24), [step2_23] "r" (step2_23),
+          [cospi_16_64] "r" (cospi_16_64)
+    );
+
+    temp21 = (step2_23 + step2_24) * cospi_16_64;
+    step1_24 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    // final stage
+    output[0 * 32] = step1_0 + step2_31;
+    output[1 * 32] = step1_1 + step2_30;
+    output[2 * 32] = step1_2 + step2_29;
+    output[3 * 32] = step1_3 + step2_28;
+    output[4 * 32] = step1_4 + step1_27;
+    output[5 * 32] = step1_5 + step1_26;
+    output[6 * 32] = step1_6 + step1_25;
+    output[7 * 32] = step1_7 + step1_24;
+    output[8 * 32] = step1_8 + step1_23;
+    output[9 * 32] = step1_9 + step1_22;
+    output[10 * 32] = step1_10 + step1_21;
+    output[11 * 32] = step1_11 + step1_20;
+    output[12 * 32] = step1_12 + step2_19;
+    output[13 * 32] = step1_13 + step2_18;
+    output[14 * 32] = step1_14 + step2_17;
+    output[15 * 32] = step1_15 + step2_16;
+    output[16 * 32] = step1_15 - step2_16;
+    output[17 * 32] = step1_14 - step2_17;
+    output[18 * 32] = step1_13 - step2_18;
+    output[19 * 32] = step1_12 - step2_19;
+    output[20 * 32] = step1_11 - step1_20;
+    output[21 * 32] = step1_10 - step1_21;
+    output[22 * 32] = step1_9 - step1_22;
+    output[23 * 32] = step1_8 - step1_23;
+    output[24 * 32] = step1_7 - step1_24;
+    output[25 * 32] = step1_6 - step1_25;
+    output[26 * 32] = step1_5 - step1_26;
+    output[27 * 32] = step1_4 - step1_27;
+    output[28 * 32] = step1_3 - step2_28;
+    output[29 * 32] = step1_2 - step2_29;
+    output[30 * 32] = step1_1 - step2_30;
+    output[31 * 32] = step1_0 - step2_31;
+
+    input += 32;
+    output += 1;
+  }
+}
+
+void vpx_idct32x32_1024_add_dspr2(const int16_t *input, uint8_t *dest,
+                                  int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t,  out[32 * 32]);
+  int16_t *outptr = out;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  // Rows
+  idct32_rows_dspr2(input, outptr, 32);
+
+  // Columns
+  vpx_idct32_cols_add_blk_dspr2(out, dest, dest_stride);
+}
+
+void vpx_idct32x32_34_add_dspr2(const int16_t *input, uint8_t *dest,
+                                int stride) {
+  DECLARE_ALIGNED(32, int16_t,  out[32 * 32]);
+  int16_t *outptr = out;
+  uint32_t i;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  // Rows
+  idct32_rows_dspr2(input, outptr, 8);
+
+  outptr += 8;
+  __asm__ __volatile__ (
+      "sw     $zero,      0(%[outptr])     \n\t"
+      "sw     $zero,      4(%[outptr])     \n\t"
+      "sw     $zero,      8(%[outptr])     \n\t"
+      "sw     $zero,     12(%[outptr])     \n\t"
+      "sw     $zero,     16(%[outptr])     \n\t"
+      "sw     $zero,     20(%[outptr])     \n\t"
+      "sw     $zero,     24(%[outptr])     \n\t"
+      "sw     $zero,     28(%[outptr])     \n\t"
+      "sw     $zero,     32(%[outptr])     \n\t"
+      "sw     $zero,     36(%[outptr])     \n\t"
+      "sw     $zero,     40(%[outptr])     \n\t"
+      "sw     $zero,     44(%[outptr])     \n\t"
+
+      :
+      : [outptr] "r" (outptr)
+  );
+
+  for (i = 0; i < 31; ++i) {
+    outptr += 32;
+
+    __asm__ __volatile__ (
+        "sw     $zero,      0(%[outptr])     \n\t"
+        "sw     $zero,      4(%[outptr])     \n\t"
+        "sw     $zero,      8(%[outptr])     \n\t"
+        "sw     $zero,     12(%[outptr])     \n\t"
+        "sw     $zero,     16(%[outptr])     \n\t"
+        "sw     $zero,     20(%[outptr])     \n\t"
+        "sw     $zero,     24(%[outptr])     \n\t"
+        "sw     $zero,     28(%[outptr])     \n\t"
+        "sw     $zero,     32(%[outptr])     \n\t"
+        "sw     $zero,     36(%[outptr])     \n\t"
+        "sw     $zero,     40(%[outptr])     \n\t"
+        "sw     $zero,     44(%[outptr])     \n\t"
+
+        :
+        : [outptr] "r" (outptr)
+    );
+  }
+
+  // Columns
+  vpx_idct32_cols_add_blk_dspr2(out, dest, stride);
+}
+
+void vpx_idct32x32_1_add_dspr2(const int16_t *input, uint8_t *dest,
+                               int stride) {
+  int       r, out;
+  int32_t   a1, absa1;
+  int32_t   vector_a1;
+  int32_t   t1, t2, t3, t4;
+  int32_t   vector_1, vector_2, vector_3, vector_4;
+  uint32_t  pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+
+    :
+    : [pos] "r" (pos)
+  );
+
+  out = DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input[0]);
+  __asm__ __volatile__ (
+      "addi     %[out],    %[out],    32      \n\t"
+      "sra      %[a1],     %[out],    6       \n\t"
+
+      : [out] "+r" (out), [a1] "=r" (a1)
+      :
+  );
+
+  if (a1 < 0) {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__ (
+        "abs        %[absa1],     %[a1]         \n\t"
+        "replv.qb   %[vector_a1], %[absa1]      \n\t"
+
+        : [absa1] "=r" (absa1), [vector_a1] "=r" (vector_a1)
+        : [a1] "r" (a1)
+    );
+
+    for (r = 32; r--;) {
+      __asm__ __volatile__ (
+          "lw             %[t1],          0(%[dest])                      \n\t"
+          "lw             %[t2],          4(%[dest])                      \n\t"
+          "lw             %[t3],          8(%[dest])                      \n\t"
+          "lw             %[t4],          12(%[dest])                     \n\t"
+          "subu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    0(%[dest])                      \n\t"
+          "sw             %[vector_2],    4(%[dest])                      \n\t"
+          "sw             %[vector_3],    8(%[dest])                      \n\t"
+          "sw             %[vector_4],    12(%[dest])                     \n\t"
+
+          "lw             %[t1],          16(%[dest])                     \n\t"
+          "lw             %[t2],          20(%[dest])                     \n\t"
+          "lw             %[t3],          24(%[dest])                     \n\t"
+          "lw             %[t4],          28(%[dest])                     \n\t"
+          "subu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    16(%[dest])                     \n\t"
+          "sw             %[vector_2],    20(%[dest])                     \n\t"
+          "sw             %[vector_3],    24(%[dest])                     \n\t"
+          "sw             %[vector_4],    28(%[dest])                     \n\t"
+
+          "add            %[dest],        %[dest],        %[stride]       \n\t"
+
+          : [t1] "=&r" (t1), [t2] "=&r" (t2), [t3] "=&r" (t3), [t4] "=&r" (t4),
+            [vector_1] "=&r" (vector_1), [vector_2] "=&r" (vector_2),
+            [vector_3] "=&r" (vector_3), [vector_4] "=&r" (vector_4),
+            [dest] "+&r" (dest)
+          : [stride] "r" (stride), [vector_a1] "r" (vector_a1)
+      );
+    }
+  } else {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__ (
+        "replv.qb       %[vector_a1],   %[a1]     \n\t"
+
+        : [vector_a1] "=r" (vector_a1)
+        : [a1] "r" (a1)
+    );
+
+    for (r = 32; r--;) {
+      __asm__ __volatile__ (
+          "lw             %[t1],          0(%[dest])                      \n\t"
+          "lw             %[t2],          4(%[dest])                      \n\t"
+          "lw             %[t3],          8(%[dest])                      \n\t"
+          "lw             %[t4],          12(%[dest])                     \n\t"
+          "addu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    0(%[dest])                      \n\t"
+          "sw             %[vector_2],    4(%[dest])                      \n\t"
+          "sw             %[vector_3],    8(%[dest])                      \n\t"
+          "sw             %[vector_4],    12(%[dest])                     \n\t"
+
+          "lw             %[t1],          16(%[dest])                     \n\t"
+          "lw             %[t2],          20(%[dest])                     \n\t"
+          "lw             %[t3],          24(%[dest])                     \n\t"
+          "lw             %[t4],          28(%[dest])                     \n\t"
+          "addu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    16(%[dest])                     \n\t"
+          "sw             %[vector_2],    20(%[dest])                     \n\t"
+          "sw             %[vector_3],    24(%[dest])                     \n\t"
+          "sw             %[vector_4],    28(%[dest])                     \n\t"
+
+          "add            %[dest],        %[dest],        %[stride]       \n\t"
+
+          : [t1] "=&r" (t1), [t2] "=&r" (t2), [t3] "=&r" (t3), [t4] "=&r" (t4),
+            [vector_1] "=&r" (vector_1), [vector_2] "=&r" (vector_2),
+            [vector_3] "=&r" (vector_3), [vector_4] "=&r" (vector_4),
+            [dest] "+&r" (dest)
+          : [stride] "r" (stride), [vector_a1] "r" (vector_a1)
+      );
+    }
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/itrans4_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/itrans4_dspr2.c
new file mode 100644
index 0000000..ecb8bd3
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/itrans4_dspr2.c
@@ -0,0 +1,359 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/inv_txfm_dspr2.h"
+#include "vpx_dsp/txfm_common.h"
+
+#if HAVE_DSPR2
+void vpx_idct4_rows_dspr2(const int16_t *input, int16_t *output) {
+  int16_t   step_0, step_1, step_2, step_3;
+  int       Temp0, Temp1, Temp2, Temp3;
+  const int const_2_power_13 = 8192;
+  int       i;
+
+  for (i = 4; i--; ) {
+    __asm__ __volatile__ (
+        /*
+          temp_1 = (input[0] + input[2]) * cospi_16_64;
+          step_0 = dct_const_round_shift(temp_1);
+
+          temp_2 = (input[0] - input[2]) * cospi_16_64;
+          step_1 = dct_const_round_shift(temp_2);
+        */
+        "lh       %[Temp0],             0(%[input])                     \n\t"
+        "lh       %[Temp1],             4(%[input])                     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "add      %[Temp2],             %[Temp0],       %[Temp1]        \n\t"
+        "sub      %[Temp3],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac0,                 %[Temp2],       %[cospi_16_64]  \n\t"
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "extp     %[step_0],            $ac0,           31              \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+
+        "madd     $ac1,                 %[Temp3],       %[cospi_16_64]  \n\t"
+        "extp     %[step_1],            $ac1,           31              \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        /*
+          temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+          step_2 = dct_const_round_shift(temp1);
+        */
+        "madd     $ac0,                 %[Temp0],       %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_8_64]   \n\t"
+        "extp     %[step_2],            $ac0,           31              \n\t"
+
+        /*
+          temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+          step_3 = dct_const_round_shift(temp2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_24_64]  \n\t"
+        "extp     %[step_3],            $ac1,           31              \n\t"
+
+        /*
+          output[0]  = step_0 + step_3;
+          output[4]  = step_1 + step_2;
+          output[8]  = step_1 - step_2;
+          output[12] = step_0 - step_3;
+        */
+        "add      %[Temp0],             %[step_0],      %[step_3]       \n\t"
+        "sh       %[Temp0],             0(%[output])                    \n\t"
+
+        "add      %[Temp1],             %[step_1],      %[step_2]       \n\t"
+        "sh       %[Temp1],             8(%[output])                    \n\t"
+
+        "sub      %[Temp2],             %[step_1],      %[step_2]       \n\t"
+        "sh       %[Temp2],             16(%[output])                   \n\t"
+
+        "sub      %[Temp3],             %[step_0],      %[step_3]       \n\t"
+        "sh       %[Temp3],             24(%[output])                   \n\t"
+
+      : [Temp0] "=&r" (Temp0), [Temp1] "=&r" (Temp1),
+        [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+        [step_0] "=&r" (step_0), [step_1] "=&r" (step_1),
+        [step_2] "=&r" (step_2), [step_3] "=&r" (step_3),
+        [output] "+r" (output)
+      : [const_2_power_13] "r" (const_2_power_13),
+        [cospi_8_64] "r" (cospi_8_64), [cospi_16_64] "r" (cospi_16_64),
+        [cospi_24_64] "r" (cospi_24_64),
+        [input] "r" (input)
+    );
+
+    input += 4;
+    output += 1;
+  }
+}
+
+void vpx_idct4_columns_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                     int dest_stride) {
+  int16_t   step_0, step_1, step_2, step_3;
+  int       Temp0, Temp1, Temp2, Temp3;
+  const int const_2_power_13 = 8192;
+  int       i;
+  uint8_t   *dest_pix;
+  uint8_t   *cm = vpx_ff_cropTbl;
+
+  /* prefetch vpx_ff_cropTbl */
+  prefetch_load(vpx_ff_cropTbl);
+  prefetch_load(vpx_ff_cropTbl +  32);
+  prefetch_load(vpx_ff_cropTbl +  64);
+  prefetch_load(vpx_ff_cropTbl +  96);
+  prefetch_load(vpx_ff_cropTbl + 128);
+  prefetch_load(vpx_ff_cropTbl + 160);
+  prefetch_load(vpx_ff_cropTbl + 192);
+  prefetch_load(vpx_ff_cropTbl + 224);
+
+  for (i = 0; i < 4; ++i) {
+      dest_pix = (dest + i);
+
+    __asm__ __volatile__ (
+        /*
+          temp_1 = (input[0] + input[2]) * cospi_16_64;
+          step_0 = dct_const_round_shift(temp_1);
+
+          temp_2 = (input[0] - input[2]) * cospi_16_64;
+          step_1 = dct_const_round_shift(temp_2);
+        */
+        "lh       %[Temp0],             0(%[input])                     \n\t"
+        "lh       %[Temp1],             4(%[input])                     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "add      %[Temp2],             %[Temp0],       %[Temp1]        \n\t"
+        "sub      %[Temp3],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac0,                 %[Temp2],       %[cospi_16_64]  \n\t"
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "extp     %[step_0],            $ac0,           31              \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+
+        "madd     $ac1,                 %[Temp3],       %[cospi_16_64]  \n\t"
+        "extp     %[step_1],            $ac1,           31              \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        /*
+          temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+          step_2 = dct_const_round_shift(temp1);
+        */
+        "madd     $ac0,                 %[Temp0],       %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_8_64]   \n\t"
+        "extp     %[step_2],            $ac0,           31              \n\t"
+
+        /*
+          temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+          step_3 = dct_const_round_shift(temp2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_24_64]  \n\t"
+        "extp     %[step_3],            $ac1,           31              \n\t"
+
+        /*
+          output[0]  = step_0 + step_3;
+          output[4]  = step_1 + step_2;
+          output[8]  = step_1 - step_2;
+          output[12] = step_0 - step_3;
+        */
+        "add      %[Temp0],             %[step_0],      %[step_3]       \n\t"
+        "addi     %[Temp0],             %[Temp0],       8               \n\t"
+        "sra      %[Temp0],             %[Temp0],       4               \n\t"
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "add      %[Temp0],             %[step_1],      %[step_2]       \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "addi     %[Temp0],             %[Temp0],       8               \n\t"
+        "sra      %[Temp0],             %[Temp0],       4               \n\t"
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step_1],      %[step_2]       \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "addi     %[Temp0],             %[Temp0],       8               \n\t"
+        "sra      %[Temp0],             %[Temp0],       4               \n\t"
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step_0],      %[step_3]       \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "addi     %[Temp0],             %[Temp0],       8               \n\t"
+        "sra      %[Temp0],             %[Temp0],       4               \n\t"
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+
+      : [Temp0] "=&r" (Temp0), [Temp1] "=&r" (Temp1),
+        [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+        [step_0] "=&r" (step_0), [step_1] "=&r" (step_1),
+        [step_2] "=&r" (step_2), [step_3] "=&r" (step_3),
+        [dest_pix] "+r" (dest_pix)
+      : [const_2_power_13] "r" (const_2_power_13),
+        [cospi_8_64] "r" (cospi_8_64), [cospi_16_64] "r" (cospi_16_64),
+        [cospi_24_64] "r" (cospi_24_64),
+        [input] "r" (input), [cm] "r" (cm), [dest_stride] "r" (dest_stride)
+    );
+
+    input += 4;
+  }
+}
+
+void vpx_idct4x4_16_add_dspr2(const int16_t *input, uint8_t *dest,
+                              int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t, out[4 * 4]);
+  int16_t *outptr = out;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  // Rows
+  vpx_idct4_rows_dspr2(input, outptr);
+
+  // Columns
+  vpx_idct4_columns_add_blk_dspr2(&out[0], dest, dest_stride);
+}
+
+void vpx_idct4x4_1_add_dspr2(const int16_t *input, uint8_t *dest,
+                             int dest_stride) {
+  int       a1, absa1;
+  int       r;
+  int32_t   out;
+  int       t2, vector_a1, vector_a;
+  uint32_t  pos = 45;
+  int16_t   input_dc = input[0];
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+
+    :
+    : [pos] "r" (pos)
+  );
+
+  out = DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input_dc);
+  __asm__ __volatile__ (
+      "addi     %[out],     %[out],    8       \n\t"
+      "sra      %[a1],      %[out],    4       \n\t"
+
+      : [out] "+r" (out), [a1] "=r" (a1)
+      :
+  );
+
+  if (a1 < 0) {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__ (
+        "abs        %[absa1],     %[a1]         \n\t"
+        "replv.qb   %[vector_a1], %[absa1]      \n\t"
+
+        : [absa1] "=r" (absa1), [vector_a1] "=r" (vector_a1)
+        : [a1] "r" (a1)
+    );
+
+    for (r = 4; r--;) {
+      __asm__ __volatile__ (
+          "lw             %[t2],          0(%[dest])                      \n\t"
+          "subu_s.qb      %[vector_a],    %[t2],          %[vector_a1]    \n\t"
+          "sw             %[vector_a],    0(%[dest])                      \n\t"
+          "add            %[dest],        %[dest],        %[dest_stride]  \n\t"
+
+          : [t2] "=&r" (t2), [vector_a] "=&r" (vector_a),
+            [dest] "+&r" (dest)
+          : [dest_stride] "r" (dest_stride), [vector_a1] "r" (vector_a1)
+      );
+    }
+  } else {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__ (
+        "replv.qb       %[vector_a1],   %[a1]     \n\t"
+        : [vector_a1] "=r" (vector_a1)
+        : [a1] "r" (a1)
+    );
+
+    for (r = 4; r--;) {
+      __asm__ __volatile__ (
+          "lw           %[t2],          0(%[dest])                        \n\t"
+          "addu_s.qb    %[vector_a],    %[t2],            %[vector_a1]    \n\t"
+          "sw           %[vector_a],    0(%[dest])                        \n\t"
+          "add          %[dest],        %[dest],          %[dest_stride]  \n\t"
+
+          : [t2] "=&r" (t2), [vector_a] "=&r" (vector_a),
+            [dest] "+&r" (dest)
+          : [dest_stride] "r" (dest_stride), [vector_a1] "r" (vector_a1)
+      );
+    }
+  }
+}
+
+void iadst4_dspr2(const int16_t *input, int16_t *output) {
+  int s0, s1, s2, s3, s4, s5, s6, s7;
+  int x0, x1, x2, x3;
+
+  x0 = input[0];
+  x1 = input[1];
+  x2 = input[2];
+  x3 = input[3];
+
+  if (!(x0 | x1 | x2 | x3)) {
+    output[0] = output[1] = output[2] = output[3] = 0;
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_2_9 * x0;
+  s2 = sinpi_3_9 * x1;
+  s3 = sinpi_4_9 * x2;
+  s4 = sinpi_1_9 * x2;
+  s5 = sinpi_2_9 * x3;
+  s6 = sinpi_4_9 * x3;
+  s7 = x0 - x2 + x3;
+
+  x0 = s0 + s3 + s5;
+  x1 = s1 - s4 - s6;
+  x2 = sinpi_3_9 * s7;
+  x3 = s2;
+
+  s0 = x0 + x3;
+  s1 = x1 + x3;
+  s2 = x2;
+  s3 = x0 + x1 - x3;
+
+  // 1-D transform scaling factor is sqrt(2).
+  // The overall dynamic range is 14b (input) + 14b (multiplication scaling)
+  // + 1b (addition) = 29b.
+  // Hence the output bit depth is 15b.
+  output[0] = dct_const_round_shift(s0);
+  output[1] = dct_const_round_shift(s1);
+  output[2] = dct_const_round_shift(s2);
+  output[3] = dct_const_round_shift(s3);
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/itrans8_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/itrans8_dspr2.c
new file mode 100644
index 0000000..823e845
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/itrans8_dspr2.c
@@ -0,0 +1,668 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/inv_txfm_dspr2.h"
+#include "vpx_dsp/txfm_common.h"
+
+#if HAVE_DSPR2
+void idct8_rows_dspr2(const int16_t *input, int16_t *output, uint32_t no_rows) {
+  int step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6, step1_7;
+  const int const_2_power_13 = 8192;
+  int Temp0, Temp1, Temp2, Temp3, Temp4;
+  int i;
+
+  for (i = no_rows; i--; ) {
+    __asm__ __volatile__ (
+        /*
+          temp_1 = (input[0] + input[4]) * cospi_16_64;
+          step2_0 = dct_const_round_shift(temp_1);
+
+          temp_2 = (input[0] - input[4]) * cospi_16_64;
+          step2_1 = dct_const_round_shift(temp_2);
+        */
+        "lh       %[Temp0],             0(%[input])                     \n\t"
+        "lh       %[Temp1],             8(%[input])                     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "add      %[Temp2],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac0,                 %[Temp2],       %[cospi_16_64]  \n\t"
+        "extp     %[Temp4],             $ac0,           31              \n\t"
+
+        "sub      %[Temp3],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac1,                 %[Temp3],       %[cospi_16_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "extp     %[Temp2],             $ac1,           31              \n\t"
+
+        /*
+          temp_1 = input[2] * cospi_24_64 - input[6] * cospi_8_64;
+          step2_2 = dct_const_round_shift(temp_1);
+        */
+        "lh       %[Temp0],             4(%[input])                     \n\t"
+        "lh       %[Temp1],             12(%[input])                    \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_8_64]   \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "extp     %[Temp3],             $ac0,           31              \n\t"
+
+        /*
+          step1_1 = step2_1 + step2_2;
+          step1_2 = step2_1 - step2_2;
+        */
+        "add      %[step1_1],           %[Temp2],       %[Temp3]        \n\t"
+        "sub      %[step1_2],           %[Temp2],       %[Temp3]        \n\t"
+
+        /*
+          temp_2 = input[2] * cospi_8_64 + input[6] * cospi_24_64;
+          step2_3 = dct_const_round_shift(temp_2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_24_64]  \n\t"
+        "extp     %[Temp1],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+
+        /*
+          step1_0 = step2_0 + step2_3;
+          step1_3 = step2_0 - step2_3;
+        */
+        "add      %[step1_0],           %[Temp4],       %[Temp1]        \n\t"
+        "sub      %[step1_3],           %[Temp4],       %[Temp1]        \n\t"
+
+        /*
+          temp_1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+          step1_4 = dct_const_round_shift(temp_1);
+        */
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_28_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "lh       %[Temp1],             14(%[input])                    \n\t"
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_4_64]   \n\t"
+        "extp     %[step1_4],           $ac0,           31              \n\t"
+
+        /*
+          temp_2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+          step1_7 = dct_const_round_shift(temp_2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_4_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_28_64]  \n\t"
+        "extp     %[step1_7],           $ac1,           31              \n\t"
+
+        /*
+          temp_1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+          step1_5 = dct_const_round_shift(temp_1);
+        */
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "lh       %[Temp0],             10(%[input])                    \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_12_64]  \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_20_64]  \n\t"
+        "extp     %[step1_5],           $ac0,           31              \n\t"
+
+        /*
+          temp_2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+          step1_6 = dct_const_round_shift(temp_2);
+        */
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "lh       %[Temp0],             10(%[input])                    \n\t"
+        "madd     $ac1,                 %[Temp0],       %[cospi_20_64]  \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_12_64]  \n\t"
+        "extp     %[step1_6],           $ac1,           31              \n\t"
+
+        /*
+          temp_1 = (step1_7 - step1_6 - step1_4 + step1_5) * cospi_16_64;
+          temp_2 = (step1_4 - step1_5 - step1_6 + step1_7) * cospi_16_64;
+        */
+        "sub      %[Temp0],             %[step1_7],     %[step1_6]      \n\t"
+        "sub      %[Temp0],             %[Temp0],       %[step1_4]      \n\t"
+        "add      %[Temp0],             %[Temp0],       %[step1_5]      \n\t"
+        "sub      %[Temp1],             %[step1_4],     %[step1_5]      \n\t"
+        "sub      %[Temp1],             %[Temp1],       %[step1_6]      \n\t"
+        "add      %[Temp1],             %[Temp1],       %[step1_7]      \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        "madd     $ac0,                 %[Temp0],       %[cospi_16_64]  \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_16_64]  \n\t"
+
+        /*
+          step1_4 = step1_4 + step1_5;
+          step1_7 = step1_6 + step1_7;
+        */
+        "add      %[step1_4],           %[step1_4],     %[step1_5]      \n\t"
+        "add      %[step1_7],           %[step1_7],     %[step1_6]      \n\t"
+
+        "extp     %[step1_5],           $ac0,           31              \n\t"
+        "extp     %[step1_6],           $ac1,           31              \n\t"
+
+        "add      %[Temp0],             %[step1_0],     %[step1_7]      \n\t"
+        "sh       %[Temp0],             0(%[output])                    \n\t"
+        "add      %[Temp1],             %[step1_1],     %[step1_6]      \n\t"
+        "sh       %[Temp1],             16(%[output])                   \n\t"
+        "add      %[Temp0],             %[step1_2],     %[step1_5]      \n\t"
+        "sh       %[Temp0],             32(%[output])                   \n\t"
+        "add      %[Temp1],             %[step1_3],     %[step1_4]      \n\t"
+        "sh       %[Temp1],             48(%[output])                   \n\t"
+
+        "sub      %[Temp0],             %[step1_3],     %[step1_4]      \n\t"
+        "sh       %[Temp0],             64(%[output])                   \n\t"
+        "sub      %[Temp1],             %[step1_2],     %[step1_5]      \n\t"
+        "sh       %[Temp1],             80(%[output])                   \n\t"
+        "sub      %[Temp0],             %[step1_1],     %[step1_6]      \n\t"
+        "sh       %[Temp0],             96(%[output])                   \n\t"
+        "sub      %[Temp1],             %[step1_0],     %[step1_7]      \n\t"
+        "sh       %[Temp1],             112(%[output])                  \n\t"
+
+        : [step1_0] "=&r" (step1_0), [step1_1] "=&r" (step1_1),
+          [step1_2] "=&r" (step1_2), [step1_3] "=&r" (step1_3),
+          [step1_4] "=&r" (step1_4), [step1_5] "=&r" (step1_5),
+          [step1_6] "=&r" (step1_6), [step1_7] "=&r" (step1_7),
+          [Temp0] "=&r" (Temp0), [Temp1] "=&r" (Temp1),
+          [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+          [Temp4] "=&r" (Temp4)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [cospi_16_64] "r" (cospi_16_64), [cospi_28_64] "r" (cospi_28_64),
+          [cospi_4_64] "r" (cospi_4_64), [cospi_12_64] "r" (cospi_12_64),
+          [cospi_20_64] "r" (cospi_20_64), [cospi_8_64] "r" (cospi_8_64),
+          [cospi_24_64] "r" (cospi_24_64),
+          [output] "r" (output), [input] "r" (input)
+    );
+
+    input += 8;
+    output += 1;
+  }
+}
+
+void idct8_columns_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                 int dest_stride) {
+  int step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6, step1_7;
+  int Temp0, Temp1, Temp2, Temp3;
+  int i;
+  const int const_2_power_13 = 8192;
+  uint8_t *dest_pix;
+  uint8_t *cm = vpx_ff_cropTbl;
+
+  /* prefetch vpx_ff_cropTbl */
+  prefetch_load(vpx_ff_cropTbl);
+  prefetch_load(vpx_ff_cropTbl +  32);
+  prefetch_load(vpx_ff_cropTbl +  64);
+  prefetch_load(vpx_ff_cropTbl +  96);
+  prefetch_load(vpx_ff_cropTbl + 128);
+  prefetch_load(vpx_ff_cropTbl + 160);
+  prefetch_load(vpx_ff_cropTbl + 192);
+  prefetch_load(vpx_ff_cropTbl + 224);
+
+  for (i = 0; i < 8; ++i) {
+      dest_pix = (dest + i);
+
+    __asm__ __volatile__ (
+        /*
+          temp_1 = (input[0] + input[4]) * cospi_16_64;
+          step2_0 = dct_const_round_shift(temp_1);
+
+          temp_2 = (input[0] - input[4]) * cospi_16_64;
+          step2_1 = dct_const_round_shift(temp_2);
+        */
+        "lh       %[Temp0],             0(%[input])                     \n\t"
+        "lh       %[Temp1],             8(%[input])                     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "add      %[Temp2],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac0,                 %[Temp2],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_6],           $ac0,           31              \n\t"
+
+        "sub      %[Temp3],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac1,                 %[Temp3],       %[cospi_16_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "extp     %[Temp2],             $ac1,           31              \n\t"
+
+        /*
+          temp_1 = input[2] * cospi_24_64 - input[6] * cospi_8_64;
+          step2_2 = dct_const_round_shift(temp_1);
+        */
+        "lh       %[Temp0],             4(%[input])                     \n\t"
+        "lh       %[Temp1],             12(%[input])                    \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_8_64]   \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "extp     %[Temp3],             $ac0,           31              \n\t"
+
+        /*
+          step1_1 = step2_1 + step2_2;
+          step1_2 = step2_1 - step2_2;
+        */
+        "add      %[step1_1],           %[Temp2],       %[Temp3]        \n\t"
+        "sub      %[step1_2],           %[Temp2],       %[Temp3]        \n\t"
+
+        /*
+          temp_2 = input[2] * cospi_8_64 + input[6] * cospi_24_64;
+          step2_3 = dct_const_round_shift(temp_2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_24_64]  \n\t"
+        "extp     %[Temp1],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+
+        /*
+          step1_0 = step2_0 + step2_3;
+          step1_3 = step2_0 - step2_3;
+        */
+        "add      %[step1_0],           %[step1_6],     %[Temp1]        \n\t"
+        "sub      %[step1_3],           %[step1_6],     %[Temp1]        \n\t"
+
+        /*
+          temp_1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+          step1_4 = dct_const_round_shift(temp_1);
+        */
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_28_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "lh       %[Temp1],             14(%[input])                    \n\t"
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_4_64]   \n\t"
+        "extp     %[step1_4],           $ac0,           31              \n\t"
+
+        /*
+          temp_2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+          step1_7 = dct_const_round_shift(temp_2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_4_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_28_64]  \n\t"
+        "extp     %[step1_7],           $ac1,           31              \n\t"
+
+        /*
+          temp_1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+          step1_5 = dct_const_round_shift(temp_1);
+        */
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "lh       %[Temp0],             10(%[input])                    \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_12_64]  \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_20_64]  \n\t"
+        "extp     %[step1_5],           $ac0,           31              \n\t"
+
+        /*
+          temp_2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+          step1_6 = dct_const_round_shift(temp_2);
+        */
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "lh       %[Temp0],             10(%[input])                    \n\t"
+        "madd     $ac1,                 %[Temp0],       %[cospi_20_64]  \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_12_64]  \n\t"
+        "extp     %[step1_6],           $ac1,           31              \n\t"
+
+        /*
+          temp_1 = (step1_7 - step1_6 - step1_4 + step1_5) * cospi_16_64;
+          temp_2 = (step1_4 - step1_5 - step1_6 + step1_7) * cospi_16_64;
+        */
+        "sub      %[Temp0],             %[step1_7],     %[step1_6]      \n\t"
+        "sub      %[Temp0],             %[Temp0],       %[step1_4]      \n\t"
+        "add      %[Temp0],             %[Temp0],       %[step1_5]      \n\t"
+        "sub      %[Temp1],             %[step1_4],     %[step1_5]      \n\t"
+        "sub      %[Temp1],             %[Temp1],       %[step1_6]      \n\t"
+        "add      %[Temp1],             %[Temp1],       %[step1_7]      \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        "madd     $ac0,                 %[Temp0],       %[cospi_16_64]  \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_16_64]  \n\t"
+
+        /*
+          step1_4 = step1_4 + step1_5;
+          step1_7 = step1_6 + step1_7;
+        */
+        "add      %[step1_4],           %[step1_4],     %[step1_5]      \n\t"
+        "add      %[step1_7],           %[step1_7],     %[step1_6]      \n\t"
+
+        "extp     %[step1_5],           $ac0,           31              \n\t"
+        "extp     %[step1_6],           $ac1,           31              \n\t"
+
+        /* add block */
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "add      %[Temp0],             %[step1_0],     %[step1_7]      \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "add      %[Temp0],             %[step1_1],     %[step1_6]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "add      %[Temp0],             %[step1_2],     %[step1_5]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "add      %[Temp0],             %[step1_3],     %[step1_4]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step1_3],     %[step1_4]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step1_2],     %[step1_5]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step1_1],     %[step1_6]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step1_0],     %[step1_7]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+
+        : [step1_0] "=&r" (step1_0), [step1_1] "=&r" (step1_1),
+          [step1_2] "=&r" (step1_2), [step1_3] "=&r" (step1_3),
+          [step1_4] "=&r" (step1_4), [step1_5] "=&r" (step1_5),
+          [step1_6] "=&r" (step1_6), [step1_7] "=&r" (step1_7),
+          [Temp0] "=&r" (Temp0), [Temp1] "=&r" (Temp1),
+          [Temp2] "=&r" (Temp2), [Temp3] "=&r" (Temp3),
+          [dest_pix] "+r" (dest_pix)
+        : [const_2_power_13] "r" (const_2_power_13),
+          [cospi_16_64] "r" (cospi_16_64), [cospi_28_64] "r" (cospi_28_64),
+          [cospi_4_64] "r" (cospi_4_64), [cospi_12_64] "r" (cospi_12_64),
+          [cospi_20_64] "r" (cospi_20_64), [cospi_8_64] "r" (cospi_8_64),
+          [cospi_24_64] "r" (cospi_24_64),
+          [input] "r" (input), [cm] "r" (cm), [dest_stride] "r" (dest_stride)
+    );
+
+    input += 8;
+  }
+}
+
+void vpx_idct8x8_64_add_dspr2(const int16_t *input, uint8_t *dest,
+                              int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t, out[8 * 8]);
+  int16_t *outptr = out;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp    %[pos],    1    \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  // First transform rows
+  idct8_rows_dspr2(input, outptr, 8);
+
+  // Then transform columns and add to dest
+  idct8_columns_add_blk_dspr2(&out[0], dest, dest_stride);
+}
+
+void vpx_idct8x8_12_add_dspr2(const int16_t *input, uint8_t *dest,
+                              int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t, out[8 * 8]);
+  int16_t *outptr = out;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp    %[pos],    1    \n\t"
+    :
+    : [pos] "r" (pos)
+  );
+
+  // First transform rows
+  idct8_rows_dspr2(input, outptr, 4);
+
+  outptr += 4;
+
+  __asm__ __volatile__ (
+      "sw  $zero,   0(%[outptr])  \n\t"
+      "sw  $zero,   4(%[outptr])  \n\t"
+      "sw  $zero,  16(%[outptr])  \n\t"
+      "sw  $zero,  20(%[outptr])  \n\t"
+      "sw  $zero,  32(%[outptr])  \n\t"
+      "sw  $zero,  36(%[outptr])  \n\t"
+      "sw  $zero,  48(%[outptr])  \n\t"
+      "sw  $zero,  52(%[outptr])  \n\t"
+      "sw  $zero,  64(%[outptr])  \n\t"
+      "sw  $zero,  68(%[outptr])  \n\t"
+      "sw  $zero,  80(%[outptr])  \n\t"
+      "sw  $zero,  84(%[outptr])  \n\t"
+      "sw  $zero,  96(%[outptr])  \n\t"
+      "sw  $zero, 100(%[outptr])  \n\t"
+      "sw  $zero, 112(%[outptr])  \n\t"
+      "sw  $zero, 116(%[outptr])  \n\t"
+
+      :
+      : [outptr] "r" (outptr)
+  );
+
+
+  // Then transform columns and add to dest
+  idct8_columns_add_blk_dspr2(&out[0], dest, dest_stride);
+}
+
+void vpx_idct8x8_1_add_dspr2(const int16_t *input, uint8_t *dest,
+                             int dest_stride) {
+  uint32_t pos = 45;
+  int32_t out;
+  int32_t r;
+  int32_t a1, absa1;
+  int32_t t1, t2, vector_a1, vector_1, vector_2;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__ (
+    "wrdsp      %[pos],     1           \n\t"
+
+    :
+    : [pos] "r" (pos)
+  );
+
+  out = DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input[0]);
+  __asm__ __volatile__ (
+      "addi     %[out],     %[out],     16      \n\t"
+      "sra      %[a1],      %[out],     5       \n\t"
+
+      : [out] "+r" (out), [a1] "=r" (a1)
+      :
+  );
+
+  if (a1 < 0) {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__ (
+        "abs        %[absa1],       %[a1]       \n\t"
+        "replv.qb   %[vector_a1],   %[absa1]    \n\t"
+
+        : [absa1] "=r" (absa1), [vector_a1] "=r" (vector_a1)
+        : [a1] "r" (a1)
+    );
+
+    for (r = 8; r--;) {
+      __asm__ __volatile__ (
+          "lw           %[t1],          0(%[dest])                      \n\t"
+          "lw           %[t2],          4(%[dest])                      \n\t"
+          "subu_s.qb    %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "subu_s.qb    %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "sw           %[vector_1],    0(%[dest])                      \n\t"
+          "sw           %[vector_2],    4(%[dest])                      \n\t"
+          "add          %[dest],        %[dest],        %[dest_stride]  \n\t"
+
+          : [t1] "=&r" (t1), [t2] "=&r" (t2),
+            [vector_1] "=&r" (vector_1), [vector_2] "=&r" (vector_2),
+            [dest] "+&r" (dest)
+          : [dest_stride] "r" (dest_stride), [vector_a1] "r" (vector_a1)
+      );
+    }
+  } else {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__ (
+        "replv.qb   %[vector_a1],   %[a1]   \n\t"
+
+        : [vector_a1] "=r" (vector_a1)
+        : [a1] "r" (a1)
+    );
+
+    for (r = 8; r--;) {
+      __asm__ __volatile__ (
+          "lw           %[t1],          0(%[dest])                      \n\t"
+          "lw           %[t2],          4(%[dest])                      \n\t"
+          "addu_s.qb    %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "addu_s.qb    %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "sw           %[vector_1],    0(%[dest])                      \n\t"
+          "sw           %[vector_2],    4(%[dest])                      \n\t"
+          "add          %[dest],        %[dest],        %[dest_stride]  \n\t"
+
+          : [t1] "=&r" (t1), [t2] "=&r" (t2),
+            [vector_1] "=&r" (vector_1), [vector_2] "=&r" (vector_2),
+            [dest] "+r" (dest)
+          : [dest_stride] "r" (dest_stride), [vector_a1] "r" (vector_a1)
+      );
+    }
+  }
+}
+
+void iadst8_dspr2(const int16_t *input, int16_t *output) {
+  int s0, s1, s2, s3, s4, s5, s6, s7;
+  int x0, x1, x2, x3, x4, x5, x6, x7;
+
+  x0 = input[7];
+  x1 = input[0];
+  x2 = input[5];
+  x3 = input[2];
+  x4 = input[3];
+  x5 = input[4];
+  x6 = input[1];
+  x7 = input[6];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7)) {
+    output[0] = output[1] = output[2] = output[3] = output[4]
+              = output[5] = output[6] = output[7] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = cospi_2_64  * x0 + cospi_30_64 * x1;
+  s1 = cospi_30_64 * x0 - cospi_2_64  * x1;
+  s2 = cospi_10_64 * x2 + cospi_22_64 * x3;
+  s3 = cospi_22_64 * x2 - cospi_10_64 * x3;
+  s4 = cospi_18_64 * x4 + cospi_14_64 * x5;
+  s5 = cospi_14_64 * x4 - cospi_18_64 * x5;
+  s6 = cospi_26_64 * x6 + cospi_6_64  * x7;
+  s7 = cospi_6_64  * x6 - cospi_26_64 * x7;
+
+  x0 = ROUND_POWER_OF_TWO((s0 + s4), DCT_CONST_BITS);
+  x1 = ROUND_POWER_OF_TWO((s1 + s5), DCT_CONST_BITS);
+  x2 = ROUND_POWER_OF_TWO((s2 + s6), DCT_CONST_BITS);
+  x3 = ROUND_POWER_OF_TWO((s3 + s7), DCT_CONST_BITS);
+  x4 = ROUND_POWER_OF_TWO((s0 - s4), DCT_CONST_BITS);
+  x5 = ROUND_POWER_OF_TWO((s1 - s5), DCT_CONST_BITS);
+  x6 = ROUND_POWER_OF_TWO((s2 - s6), DCT_CONST_BITS);
+  x7 = ROUND_POWER_OF_TWO((s3 - s7), DCT_CONST_BITS);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 =  cospi_8_64  * x4 + cospi_24_64 * x5;
+  s5 =  cospi_24_64 * x4 - cospi_8_64  * x5;
+  s6 = -cospi_24_64 * x6 + cospi_8_64  * x7;
+  s7 =  cospi_8_64  * x6 + cospi_24_64 * x7;
+
+  x0 = s0 + s2;
+  x1 = s1 + s3;
+  x2 = s0 - s2;
+  x3 = s1 - s3;
+  x4 = ROUND_POWER_OF_TWO((s4 + s6), DCT_CONST_BITS);
+  x5 = ROUND_POWER_OF_TWO((s5 + s7), DCT_CONST_BITS);
+  x6 = ROUND_POWER_OF_TWO((s4 - s6), DCT_CONST_BITS);
+  x7 = ROUND_POWER_OF_TWO((s5 - s7), DCT_CONST_BITS);
+
+  // stage 3
+  s2 = cospi_16_64 * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (x6 - x7);
+
+  x2 = ROUND_POWER_OF_TWO((s2), DCT_CONST_BITS);
+  x3 = ROUND_POWER_OF_TWO((s3), DCT_CONST_BITS);
+  x6 = ROUND_POWER_OF_TWO((s6), DCT_CONST_BITS);
+  x7 = ROUND_POWER_OF_TWO((s7), DCT_CONST_BITS);
+
+  output[0] =  x0;
+  output[1] = -x4;
+  output[2] =  x6;
+  output[3] = -x2;
+  output[4] =  x3;
+  output[5] = -x7;
+  output[6] =  x5;
+  output[7] = -x1;
+}
+#endif  // HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_16_msa.c b/libvpx/libvpx/vpx_dsp/mips/loopfilter_16_msa.c
new file mode 100644
index 0000000..a6c581d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_16_msa.c
@@ -0,0 +1,1494 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_ports/mem.h"
+#include "vpx_dsp/mips/loopfilter_msa.h"
+
+int32_t vpx_hz_lpf_t4_and_t8_16w(uint8_t *src, int32_t pitch,
+                                 uint8_t *filter48,
+                                 const uint8_t *b_limit_ptr,
+                                 const uint8_t *limit_ptr,
+                                 const uint8_t *thresh_ptr) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+  v16u8 zero = { 0 };
+
+  /* load vector elements */
+  LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  if (__msa_test_bz_v(flat)) {
+    ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch);
+
+    return 1;
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
+               q2_r, q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+    ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+    VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+                p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+                p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
+    filter48 += (4 * 16);
+    ST_UB2(q1_out, q2_out, filter48, 16);
+    filter48 += (2 * 16);
+    ST_UB(flat, filter48);
+
+    return 0;
+  }
+}
+
+void vpx_hz_lpf_t16_16w(uint8_t *src, int32_t pitch, uint8_t *filter48) {
+  v16u8 flat, flat2, filter8;
+  v16i8 zero = { 0 };
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
+  v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
+  v8u16 p7_l_in, p6_l_in, p5_l_in, p4_l_in, p3_l_in, p2_l_in, p1_l_in, p0_l_in;
+  v8u16 q7_l_in, q6_l_in, q5_l_in, q4_l_in, q3_l_in, q2_l_in, q1_l_in, q0_l_in;
+  v8u16 tmp0_r, tmp1_r, tmp0_l, tmp1_l;
+  v8i16 l_out, r_out;
+
+  flat = LD_UB(filter48 + 96);
+
+  LD_UB8((src - 8 * pitch), pitch, p7, p6, p5, p4, p3, p2, p1, p0);
+  LD_UB8(src, pitch, q0, q1, q2, q3, q4, q5, q6, q7);
+  VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+  if (__msa_test_bz_v(flat2)) {
+    LD_UB4(filter48, 16, p2, p1, p0, q0);
+    LD_UB2(filter48 + 4 * 16, 16, q1, q2);
+
+    src -= 3 * pitch;
+    ST_UB4(p2, p1, p0, q0, src, pitch);
+    src += (4 * pitch);
+    ST_UB2(q1, q2, src, pitch);
+  } else {
+    src -= 7 * pitch;
+
+    ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2,
+               zero, p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in,
+               p2_r_in, p1_r_in, p0_r_in);
+
+    q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
+
+    tmp0_r = p7_r_in << 3;
+    tmp0_r -= p7_r_in;
+    tmp0_r += p6_r_in;
+    tmp0_r += q0_r_in;
+    tmp1_r = p6_r_in + p5_r_in;
+    tmp1_r += p4_r_in;
+    tmp1_r += p3_r_in;
+    tmp1_r += p2_r_in;
+    tmp1_r += p1_r_in;
+    tmp1_r += p0_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    ILVL_B4_UH(zero, p7, zero, p6, zero, p5, zero, p4, p7_l_in, p6_l_in,
+               p5_l_in, p4_l_in);
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l_in, p2_l_in,
+               p1_l_in, p0_l_in);
+    q0_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q0);
+
+    tmp0_l = p7_l_in << 3;
+    tmp0_l -= p7_l_in;
+    tmp0_l += p6_l_in;
+    tmp0_l += q0_l_in;
+    tmp1_l = p6_l_in + p5_l_in;
+    tmp1_l += p4_l_in;
+    tmp1_l += p3_l_in;
+    tmp1_l += p2_l_in;
+    tmp1_l += p1_l_in;
+    tmp1_l += p0_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
+    ST_UB(p6, src);
+    src += pitch;
+
+    /* p5 */
+    q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
+    tmp0_r = p5_r_in - p6_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q1_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q1);
+    tmp0_l = p5_l_in - p6_l_in;
+    tmp0_l += q1_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
+    ST_UB(p5, src);
+    src += pitch;
+
+    /* p4 */
+    q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
+    tmp0_r = p4_r_in - p5_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = (v8i16)__msa_srari_h((v8i16)tmp1_r, 4);
+
+    q2_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q2);
+    tmp0_l = p4_l_in - p5_l_in;
+    tmp0_l += q2_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
+    ST_UB(p4, src);
+    src += pitch;
+
+    /* p3 */
+    q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
+    tmp0_r = p3_r_in - p4_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q3_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q3);
+    tmp0_l = p3_l_in - p4_l_in;
+    tmp0_l += q3_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
+    ST_UB(p3, src);
+    src += pitch;
+
+    /* p2 */
+    q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
+    filter8 = LD_UB(filter48);
+    tmp0_r = p2_r_in - p3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q4_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q4);
+    tmp0_l = p2_l_in - p3_l_in;
+    tmp0_l += q4_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* p1 */
+    q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
+    filter8 = LD_UB(filter48 + 16);
+    tmp0_r = p1_r_in - p2_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q5_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q5);
+    tmp0_l = p1_l_in - p2_l_in;
+    tmp0_l += q5_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* p0 */
+    q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
+    filter8 = LD_UB(filter48 + 32);
+    tmp0_r = p0_r_in - p1_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q6_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q6);
+    tmp0_l = p0_l_in - p1_l_in;
+    tmp0_l += q6_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* q0 */
+    q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
+    filter8 = LD_UB(filter48 + 48);
+    tmp0_r = q7_r_in - p0_r_in;
+    tmp0_r += q0_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q7_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q7);
+    tmp0_l = q7_l_in - p0_l_in;
+    tmp0_l += q0_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* q1 */
+    filter8 = LD_UB(filter48 + 64);
+    tmp0_r = q7_r_in - q0_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p6_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q0_l_in;
+    tmp0_l += q1_l_in;
+    tmp0_l -= p6_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* q2 */
+    filter8 = LD_UB(filter48 + 80);
+    tmp0_r = q7_r_in - q1_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p5_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q1_l_in;
+    tmp0_l += q2_l_in;
+    tmp0_l -= p5_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* q3 */
+    tmp0_r = q7_r_in - q2_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p4_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q2_l_in;
+    tmp0_l += q3_l_in;
+    tmp0_l -= p4_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
+    ST_UB(q3, src);
+    src += pitch;
+
+    /* q4 */
+    tmp0_r = q7_r_in - q3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p3_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q3_l_in;
+    tmp0_l += q4_l_in;
+    tmp0_l -= p3_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
+    ST_UB(q4, src);
+    src += pitch;
+
+    /* q5 */
+    tmp0_r = q7_r_in - q4_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p2_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q4_l_in;
+    tmp0_l += q5_l_in;
+    tmp0_l -= p2_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
+    ST_UB(q5, src);
+    src += pitch;
+
+    /* q6 */
+    tmp0_r = q7_r_in - q5_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p1_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q5_l_in;
+    tmp0_l += q6_l_in;
+    tmp0_l -= p1_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
+    ST_UB(q6, src);
+  }
+}
+
+void vpx_lpf_horizontal_16_dual_msa(uint8_t *src, int32_t pitch,
+                                    const uint8_t *b_limit_ptr,
+                                    const uint8_t *limit_ptr,
+                                    const uint8_t *thresh_ptr,
+                                    int32_t count) {
+  DECLARE_ALIGNED(32, uint8_t, filter48[16 * 8]);
+  uint8_t early_exit = 0;
+
+  (void)count;
+
+  early_exit = vpx_hz_lpf_t4_and_t8_16w(src, pitch, &filter48[0], b_limit_ptr,
+                                        limit_ptr, thresh_ptr);
+
+  if (0 == early_exit) {
+    vpx_hz_lpf_t16_16w(src, pitch, filter48);
+  }
+}
+
+static void mb_lpf_horizontal_edge(uint8_t *src, int32_t pitch,
+                                   const uint8_t *b_limit_ptr,
+                                   const uint8_t *limit_ptr,
+                                   const uint8_t *thresh_ptr,
+                                   int32_t count) {
+  if (1 == count) {
+    uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
+    uint64_t dword0, dword1;
+    v16u8 flat2, mask, hev, flat, thresh, b_limit, limit;
+    v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p7, p6, p5, p4, q4, q5, q6, q7;
+    v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+    v16u8 p0_filter16, p1_filter16;
+    v8i16 p2_filter8, p1_filter8, p0_filter8;
+    v8i16 q0_filter8, q1_filter8, q2_filter8;
+    v8u16 p7_r, p6_r, p5_r, p4_r, q7_r, q6_r, q5_r, q4_r;
+    v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r;
+    v16i8 zero = { 0 };
+    v8u16 tmp0, tmp1, tmp2;
+
+    /* load vector elements */
+    LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+    thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+    b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+    limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+    LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+                 hev, mask, flat);
+    VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+    VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out,
+                       q1_out);
+
+    flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+    if (__msa_test_bz_v(flat)) {
+      p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+      p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+      q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+      q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+      SD4(p1_d, p0_d, q0_d, q1_d, src - 2 * pitch, pitch);
+    } else {
+      /* convert 8 bit input data into 16 bit */
+      ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+                 zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+                 q3_r);
+      VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8,
+                  p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8);
+
+      /* convert 16 bit output data into 8 bit */
+      PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8,
+                  zero, q0_filter8, p2_filter8, p1_filter8, p0_filter8,
+                  q0_filter8);
+      PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8);
+
+      /* store pixel values */
+      p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat);
+      p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat);
+      p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat);
+      q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat);
+      q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat);
+      q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat);
+
+      /* load 16 vector elements */
+      LD_UB4((src - 8 * pitch), pitch, p7, p6, p5, p4);
+      LD_UB4(src + (4 * pitch), pitch, q4, q5, q6, q7);
+
+      VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+      if (__msa_test_bz_v(flat2)) {
+        p2_d = __msa_copy_u_d((v2i64)p2_out, 0);
+        p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+        p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+        q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+        q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+        q2_d = __msa_copy_u_d((v2i64)q2_out, 0);
+
+        SD4(p2_d, p1_d, p0_d, q0_d, src - 3 * pitch, pitch);
+        SD(q1_d, src + pitch);
+        SD(q2_d, src + 2 * pitch);
+      } else {
+        /* LSB(right) 8 pixel operation */
+        ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, q4, zero, q5,
+                   zero, q6, zero, q7, p7_r, p6_r, p5_r, p4_r, q4_r, q5_r, q6_r,
+                   q7_r);
+
+        tmp0 = p7_r << 3;
+        tmp0 -= p7_r;
+        tmp0 += p6_r;
+        tmp0 += q0_r;
+
+        src -= 7 * pitch;
+
+        /* calculation of p6 and p5 */
+        tmp1 = p6_r + p5_r + p4_r + p3_r;
+        tmp1 += (p2_r + p1_r + p0_r);
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp0 = p5_r - p6_r + q1_r - p7_r;
+        tmp1 += tmp0;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(p6, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(p5, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of p4 and p3 */
+        tmp0 = p4_r - p5_r + q2_r - p7_r;
+        tmp2 = p3_r - p4_r + q3_r - p7_r;
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(p4, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(p3, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of p2 and p1 */
+        tmp0 = p2_r - p3_r + q4_r - p7_r;
+        tmp2 = p1_r - p2_r + q5_r - p7_r;
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(p2_out, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(p1_out, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of p0 and q0 */
+        tmp0 = (p0_r - p1_r) + (q6_r - p7_r);
+        tmp2 = (q7_r - p0_r) + (q0_r - p7_r);
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(p0_out, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(q0_out, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of q1 and q2 */
+        tmp0 = q7_r - q0_r + q1_r - p6_r;
+        tmp2 = q7_r - q1_r + q2_r - p5_r;
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(q1_out, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(q2_out, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of q3 and q4 */
+        tmp0 = (q7_r - q2_r) + (q3_r - p4_r);
+        tmp2 = (q7_r - q3_r) + (q4_r - p3_r);
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(q3, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(q4, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of q5 and q6 */
+        tmp0 = (q7_r - q4_r) + (q5_r - p2_r);
+        tmp2 = (q7_r - q5_r) + (q6_r - p1_r);
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(q5, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(q6, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+      }
+    }
+  } else {
+    vpx_lpf_horizontal_16_dual_msa(src, pitch, b_limit_ptr, limit_ptr,
+                                   thresh_ptr, count);
+  }
+}
+
+void vpx_lpf_horizontal_edge_8_msa(uint8_t *src, int32_t pitch,
+                                   const uint8_t *b_limit_ptr,
+                                   const uint8_t *limit_ptr,
+                                   const uint8_t *thresh_ptr) {
+  mb_lpf_horizontal_edge(src, pitch, b_limit_ptr, limit_ptr, thresh_ptr, 1);
+}
+
+void vpx_lpf_horizontal_edge_16_msa(uint8_t *src, int32_t pitch,
+                                    const uint8_t *b_limit_ptr,
+                                    const uint8_t *limit_ptr,
+                                    const uint8_t *thresh_ptr) {
+  mb_lpf_horizontal_edge(src, pitch, b_limit_ptr, limit_ptr, thresh_ptr, 2);
+}
+
+static void transpose_16x8_to_8x16(uint8_t *input, int32_t in_pitch,
+                                   uint8_t *output, int32_t out_pitch) {
+  v16u8 p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org, p0_org;
+  v16i8 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+
+  LD_UB8(input, in_pitch,
+         p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org, p0_org);
+  /* 8x8 transpose */
+  TRANSPOSE8x8_UB_UB(p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org,
+                     p0_org, p7, p6, p5, p4, p3, p2, p1, p0);
+  /* 8x8 transpose */
+  ILVL_B4_SB(p5_org, p7_org, p4_org, p6_org, p1_org, p3_org, p0_org, p2_org,
+             tmp0, tmp1, tmp2, tmp3);
+  ILVR_B2_SB(tmp1, tmp0, tmp3, tmp2, tmp4, tmp6);
+  ILVL_B2_SB(tmp1, tmp0, tmp3, tmp2, tmp5, tmp7);
+  ILVR_W2_UB(tmp6, tmp4, tmp7, tmp5, q0, q4);
+  ILVL_W2_UB(tmp6, tmp4, tmp7, tmp5, q2, q6);
+  SLDI_B4_0_UB(q0, q2, q4, q6, q1, q3, q5, q7, 8);
+
+  ST_UB8(p7, p6, p5, p4, p3, p2, p1, p0, output, out_pitch);
+  output += (8 * out_pitch);
+  ST_UB8(q0, q1, q2, q3, q4, q5, q6, q7, output, out_pitch);
+}
+
+static void transpose_8x16_to_16x8(uint8_t *input, int32_t in_pitch,
+                                   uint8_t *output, int32_t out_pitch) {
+  v16u8 p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o;
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+
+  LD_UB8(input, in_pitch, p7, p6, p5, p4, p3, p2, p1, p0);
+  LD_UB8(input + (8 * in_pitch), in_pitch, q0, q1, q2, q3, q4, q5, q6, q7);
+  TRANSPOSE16x8_UB_UB(p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5,
+                      q6, q7, p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o);
+  ST_UB8(p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o, output, out_pitch);
+}
+
+static void transpose_16x16(uint8_t *input, int32_t in_pitch,
+                            uint8_t *output, int32_t out_pitch) {
+  v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
+  v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  v8i16 tmp0, tmp1, tmp4, tmp5, tmp6, tmp7;
+  v4i32 tmp2, tmp3;
+
+  LD_UB8(input, in_pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+  input += (8 * in_pitch);
+  LD_UB8(input, in_pitch,
+         row8, row9, row10, row11, row12, row13, row14, row15);
+
+  TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
+                      row8, row9, row10, row11, row12, row13, row14, row15,
+                      p7, p6, p5, p4, p3, p2, p1, p0);
+
+  /* transpose 16x8 matrix into 8x16 */
+  /* total 8 intermediate register and 32 instructions */
+  q7 = (v16u8)__msa_ilvod_d((v2i64)row8, (v2i64)row0);
+  q6 = (v16u8)__msa_ilvod_d((v2i64)row9, (v2i64)row1);
+  q5 = (v16u8)__msa_ilvod_d((v2i64)row10, (v2i64)row2);
+  q4 = (v16u8)__msa_ilvod_d((v2i64)row11, (v2i64)row3);
+  q3 = (v16u8)__msa_ilvod_d((v2i64)row12, (v2i64)row4);
+  q2 = (v16u8)__msa_ilvod_d((v2i64)row13, (v2i64)row5);
+  q1 = (v16u8)__msa_ilvod_d((v2i64)row14, (v2i64)row6);
+  q0 = (v16u8)__msa_ilvod_d((v2i64)row15, (v2i64)row7);
+
+  ILVEV_B2_SH(q7, q6, q5, q4, tmp0, tmp1);
+  tmp4 = (v8i16)__msa_ilvod_b((v16i8)q6, (v16i8)q7);
+  tmp5 = (v8i16)__msa_ilvod_b((v16i8)q4, (v16i8)q5);
+
+  ILVEV_B2_UB(q3, q2, q1, q0, q5, q7);
+  tmp6 = (v8i16)__msa_ilvod_b((v16i8)q2, (v16i8)q3);
+  tmp7 = (v8i16)__msa_ilvod_b((v16i8)q0, (v16i8)q1);
+
+  ILVEV_H2_SW(tmp0, tmp1, q5, q7, tmp2, tmp3);
+  q0 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+  q4 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+  tmp2 = (v4i32)__msa_ilvod_h(tmp1, tmp0);
+  tmp3 = (v4i32)__msa_ilvod_h((v8i16)q7, (v8i16)q5);
+  q2 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+  q6 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+  ILVEV_H2_SW(tmp4, tmp5, tmp6, tmp7, tmp2, tmp3);
+  q1 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+  q5 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+  tmp2 = (v4i32)__msa_ilvod_h(tmp5, tmp4);
+  tmp3 = (v4i32)__msa_ilvod_h(tmp7, tmp6);
+  q3 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+  q7 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+  ST_UB8(p7, p6, p5, p4, p3, p2, p1, p0, output, out_pitch);
+  output += (8 * out_pitch);
+  ST_UB8(q0, q1, q2, q3, q4, q5, q6, q7, output, out_pitch);
+}
+
+int32_t vpx_vt_lpf_t4_and_t8_8w(uint8_t *src, uint8_t *filter48,
+                                uint8_t *src_org, int32_t pitch_org,
+                                const uint8_t *b_limit_ptr,
+                                const uint8_t *limit_ptr,
+                                const uint8_t *thresh_ptr) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v16i8 zero = { 0 };
+  v8i16 vec0, vec1, vec2, vec3;
+
+  /* load vector elements */
+  LD_UB8(src - (4 * 16), 16, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  /* flat4 */
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  /* filter4 */
+  VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+  if (__msa_test_bz_v(flat)) {
+    ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+    ST4x8_UB(vec2, vec3, (src_org - 2), pitch_org);
+    return 1;
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+               q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+    /* convert 16 bit output data into 8 bit */
+    p2_r = (v8u16)__msa_pckev_b((v16i8)p2_filt8_r, (v16i8)p2_filt8_r);
+    p1_r = (v8u16)__msa_pckev_b((v16i8)p1_filt8_r, (v16i8)p1_filt8_r);
+    p0_r = (v8u16)__msa_pckev_b((v16i8)p0_filt8_r, (v16i8)p0_filt8_r);
+    q0_r = (v8u16)__msa_pckev_b((v16i8)q0_filt8_r, (v16i8)q0_filt8_r);
+    q1_r = (v8u16)__msa_pckev_b((v16i8)q1_filt8_r, (v16i8)q1_filt8_r);
+    q2_r = (v8u16)__msa_pckev_b((v16i8)q2_filt8_r, (v16i8)q2_filt8_r);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_r, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_r, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_r, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_r, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_r, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_r, flat);
+
+    ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
+    filter48 += (4 * 16);
+    ST_UB2(q1_out, q2_out, filter48, 16);
+    filter48 += (2 * 16);
+    ST_UB(flat, filter48);
+
+    return 0;
+  }
+}
+
+int32_t vpx_vt_lpf_t16_8w(uint8_t *src, uint8_t *src_org, int32_t pitch,
+                          uint8_t *filter48) {
+  v16i8 zero = { 0 };
+  v16u8 filter8, flat, flat2;
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
+  v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
+  v8u16 tmp0_r, tmp1_r;
+  v8i16 r_out;
+
+  flat = LD_UB(filter48 + 6 * 16);
+
+  LD_UB8((src - 8 * 16), 16, p7, p6, p5, p4, p3, p2, p1, p0);
+  LD_UB8(src, 16, q0, q1, q2, q3, q4, q5, q6, q7);
+
+  VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+  if (__msa_test_bz_v(flat2)) {
+    v8i16 vec0, vec1, vec2, vec3, vec4;
+
+    LD_UB4(filter48, 16, p2, p1, p0, q0);
+    LD_UB2(filter48 + 4 * 16, 16, q1, q2);
+
+    ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec3, vec4);
+    vec2 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1);
+
+    src_org -= 3;
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec2, 0, (src_org + 4), pitch);
+    src_org += (4 * pitch);
+    ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec2, 4, (src_org + 4), pitch);
+
+    return 1;
+  } else {
+    src -= 7 * 16;
+
+    ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2,
+               zero, p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in,
+               p3_r_in, p2_r_in, p1_r_in, p0_r_in);
+    q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
+
+    tmp0_r = p7_r_in << 3;
+    tmp0_r -= p7_r_in;
+    tmp0_r += p6_r_in;
+    tmp0_r += q0_r_in;
+    tmp1_r = p6_r_in + p5_r_in;
+    tmp1_r += p4_r_in;
+    tmp1_r += p3_r_in;
+    tmp1_r += p2_r_in;
+    tmp1_r += p1_r_in;
+    tmp1_r += p0_r_in;
+    tmp1_r += tmp0_r;
+
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
+    ST8x1_UB(p6, src);
+    src += 16;
+
+    /* p5 */
+    q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
+    tmp0_r = p5_r_in - p6_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
+    ST8x1_UB(p5, src);
+    src += 16;
+
+    /* p4 */
+    q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
+    tmp0_r = p4_r_in - p5_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
+    ST8x1_UB(p4, src);
+    src += 16;
+
+    /* p3 */
+    q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
+    tmp0_r = p3_r_in - p4_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
+    ST8x1_UB(p3, src);
+    src += 16;
+
+    /* p2 */
+    q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
+    filter8 = LD_UB(filter48);
+    tmp0_r = p2_r_in - p3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* p1 */
+    q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
+    filter8 = LD_UB(filter48 + 16);
+    tmp0_r = p1_r_in - p2_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* p0 */
+    q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
+    filter8 = LD_UB(filter48 + 32);
+    tmp0_r = p0_r_in - p1_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* q0 */
+    q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
+    filter8 = LD_UB(filter48 + 48);
+    tmp0_r = q7_r_in - p0_r_in;
+    tmp0_r += q0_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* q1 */
+    filter8 = LD_UB(filter48 + 64);
+    tmp0_r = q7_r_in - q0_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p6_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* q2 */
+    filter8 = LD_UB(filter48 + 80);
+    tmp0_r = q7_r_in - q1_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p5_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* q3 */
+    tmp0_r = q7_r_in - q2_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p4_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
+    ST8x1_UB(q3, src);
+    src += 16;
+
+    /* q4 */
+    tmp0_r = q7_r_in - q3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p3_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
+    ST8x1_UB(q4, src);
+    src += 16;
+
+    /* q5 */
+    tmp0_r = q7_r_in - q4_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p2_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
+    ST8x1_UB(q5, src);
+    src += 16;
+
+    /* q6 */
+    tmp0_r = q7_r_in - q5_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p1_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
+    ST8x1_UB(q6, src);
+
+    return 0;
+  }
+}
+
+void vpx_lpf_vertical_16_msa(uint8_t *src, int32_t pitch,
+                             const uint8_t *b_limit_ptr,
+                             const uint8_t *limit_ptr,
+                             const uint8_t *thresh_ptr) {
+  uint8_t early_exit = 0;
+  DECLARE_ALIGNED(32, uint8_t, transposed_input[16 * 24]);
+  uint8_t *filter48 = &transposed_input[16 * 16];
+
+  transpose_16x8_to_8x16(src - 8, pitch, transposed_input, 16);
+
+  early_exit = vpx_vt_lpf_t4_and_t8_8w((transposed_input + 16 * 8),
+                                       &filter48[0], src, pitch, b_limit_ptr,
+                                       limit_ptr, thresh_ptr);
+
+  if (0 == early_exit) {
+    early_exit = vpx_vt_lpf_t16_8w((transposed_input + 16 * 8), src, pitch,
+                                   &filter48[0]);
+
+    if (0 == early_exit) {
+      transpose_8x16_to_16x8(transposed_input, 16, src - 8, pitch);
+    }
+  }
+}
+
+int32_t vpx_vt_lpf_t4_and_t8_16w(uint8_t *src, uint8_t *filter48,
+                                 uint8_t *src_org, int32_t pitch,
+                                 const uint8_t *b_limit_ptr,
+                                 const uint8_t *limit_ptr,
+                                 const uint8_t *thresh_ptr) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+  v16i8 zero = { 0 };
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5;
+
+  /* load vector elements */
+  LD_UB8(src - (4 * 16), 16, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  /* flat4 */
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  /* filter4 */
+  VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  if (__msa_test_bz_v(flat)) {
+    ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+    ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec4, vec5);
+
+    src_org -= 2;
+    ST4x8_UB(vec2, vec3, src_org, pitch);
+    src_org += 8 * pitch;
+    ST4x8_UB(vec4, vec5, src_org, pitch);
+
+    return 1;
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+               q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+    ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+    VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+                p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+                p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
+    filter48 += (4 * 16);
+    ST_UB2(q1_out, q2_out, filter48, 16);
+    filter48 += (2 * 16);
+    ST_UB(flat, filter48);
+
+    return 0;
+  }
+}
+
+int32_t vpx_vt_lpf_t16_16w(uint8_t *src, uint8_t *src_org, int32_t pitch,
+                           uint8_t *filter48) {
+  v16u8 flat, flat2, filter8;
+  v16i8 zero = { 0 };
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
+  v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
+  v8u16 p7_l_in, p6_l_in, p5_l_in, p4_l_in, p3_l_in, p2_l_in, p1_l_in, p0_l_in;
+  v8u16 q7_l_in, q6_l_in, q5_l_in, q4_l_in, q3_l_in, q2_l_in, q1_l_in, q0_l_in;
+  v8u16 tmp0_r, tmp1_r, tmp0_l, tmp1_l;
+  v8i16 l_out, r_out;
+
+  flat = LD_UB(filter48 + 6 * 16);
+
+  LD_UB8((src - 8 * 16), 16, p7, p6, p5, p4, p3, p2, p1, p0);
+  LD_UB8(src, 16, q0, q1, q2, q3, q4, q5, q6, q7);
+
+  VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+  if (__msa_test_bz_v(flat2)) {
+    v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+
+    LD_UB4(filter48, 16, p2, p1, p0, q0);
+    LD_UB2(filter48 + 4 * 16, 16, q1, q2);
+
+    ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec3, vec4);
+    ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec6, vec7);
+    ILVRL_B2_SH(q2, q1, vec2, vec5);
+
+    src_org -= 3;
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec2, 0, (src_org + 4), pitch);
+    src_org += (4 * pitch);
+    ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec2, 4, (src_org + 4), pitch);
+    src_org += (4 * pitch);
+    ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec5, 0, (src_org + 4), pitch);
+    src_org += (4 * pitch);
+    ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec5, 4, (src_org + 4), pitch);
+
+    return 1;
+  } else {
+    src -= 7 * 16;
+
+    ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2,
+               zero, p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in,
+               p3_r_in, p2_r_in, p1_r_in, p0_r_in);
+    q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
+
+    tmp0_r = p7_r_in << 3;
+    tmp0_r -= p7_r_in;
+    tmp0_r += p6_r_in;
+    tmp0_r += q0_r_in;
+    tmp1_r = p6_r_in + p5_r_in;
+    tmp1_r += p4_r_in;
+    tmp1_r += p3_r_in;
+    tmp1_r += p2_r_in;
+    tmp1_r += p1_r_in;
+    tmp1_r += p0_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    ILVL_B4_UH(zero, p7, zero, p6, zero, p5, zero, p4, p7_l_in, p6_l_in,
+               p5_l_in, p4_l_in);
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l_in, p2_l_in,
+               p1_l_in, p0_l_in);
+    q0_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q0);
+
+    tmp0_l = p7_l_in << 3;
+    tmp0_l -= p7_l_in;
+    tmp0_l += p6_l_in;
+    tmp0_l += q0_l_in;
+    tmp1_l = p6_l_in + p5_l_in;
+    tmp1_l += p4_l_in;
+    tmp1_l += p3_l_in;
+    tmp1_l += p2_l_in;
+    tmp1_l += p1_l_in;
+    tmp1_l += p0_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
+    ST_UB(p6, src);
+    src += 16;
+
+    /* p5 */
+    q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
+    tmp0_r = p5_r_in - p6_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q1_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q1);
+    tmp0_l = p5_l_in - p6_l_in;
+    tmp0_l += q1_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
+    ST_UB(p5, src);
+    src += 16;
+
+    /* p4 */
+    q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
+    tmp0_r = p4_r_in - p5_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q2_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q2);
+    tmp0_l = p4_l_in - p5_l_in;
+    tmp0_l += q2_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
+    ST_UB(p4, src);
+    src += 16;
+
+    /* p3 */
+    q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
+    tmp0_r = p3_r_in - p4_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q3_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q3);
+    tmp0_l = p3_l_in - p4_l_in;
+    tmp0_l += q3_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
+    ST_UB(p3, src);
+    src += 16;
+
+    /* p2 */
+    q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
+    filter8 = LD_UB(filter48);
+    tmp0_r = p2_r_in - p3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q4_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q4);
+    tmp0_l = p2_l_in - p3_l_in;
+    tmp0_l += q4_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* p1 */
+    q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
+    filter8 = LD_UB(filter48 + 16);
+    tmp0_r = p1_r_in - p2_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q5_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q5);
+    tmp0_l = p1_l_in - p2_l_in;
+    tmp0_l += q5_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)(tmp1_l), 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* p0 */
+    q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
+    filter8 = LD_UB(filter48 + 32);
+    tmp0_r = p0_r_in - p1_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q6_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q6);
+    tmp0_l = p0_l_in - p1_l_in;
+    tmp0_l += q6_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* q0 */
+    q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
+    filter8 = LD_UB(filter48 + 48);
+    tmp0_r = q7_r_in - p0_r_in;
+    tmp0_r += q0_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q7_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q7);
+    tmp0_l = q7_l_in - p0_l_in;
+    tmp0_l += q0_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* q1 */
+    filter8 = LD_UB(filter48 + 64);
+    tmp0_r = q7_r_in - q0_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p6_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q0_l_in;
+    tmp0_l += q1_l_in;
+    tmp0_l -= p6_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* q2 */
+    filter8 = LD_UB(filter48 + 80);
+    tmp0_r = q7_r_in - q1_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p5_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q1_l_in;
+    tmp0_l += q2_l_in;
+    tmp0_l -= p5_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* q3 */
+    tmp0_r = q7_r_in - q2_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p4_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q2_l_in;
+    tmp0_l += q3_l_in;
+    tmp0_l -= p4_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
+    ST_UB(q3, src);
+    src += 16;
+
+    /* q4 */
+    tmp0_r = q7_r_in - q3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p3_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q3_l_in;
+    tmp0_l += q4_l_in;
+    tmp0_l -= p3_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
+    ST_UB(q4, src);
+    src += 16;
+
+    /* q5 */
+    tmp0_r = q7_r_in - q4_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p2_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q4_l_in;
+    tmp0_l += q5_l_in;
+    tmp0_l -= p2_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
+    ST_UB(q5, src);
+    src += 16;
+
+    /* q6 */
+    tmp0_r = q7_r_in - q5_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p1_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q5_l_in;
+    tmp0_l += q6_l_in;
+    tmp0_l -= p1_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
+    ST_UB(q6, src);
+
+    return 0;
+  }
+}
+
+void vpx_lpf_vertical_16_dual_msa(uint8_t *src, int32_t pitch,
+                                  const uint8_t *b_limit_ptr,
+                                  const uint8_t *limit_ptr,
+                                  const uint8_t *thresh_ptr) {
+  uint8_t early_exit = 0;
+  DECLARE_ALIGNED(32, uint8_t, transposed_input[16 * 24]);
+  uint8_t *filter48 = &transposed_input[16 * 16];
+
+  transpose_16x16((src - 8), pitch, &transposed_input[0], 16);
+
+  early_exit = vpx_vt_lpf_t4_and_t8_16w((transposed_input + 16 * 8),
+                                        &filter48[0], src, pitch, b_limit_ptr,
+                                        limit_ptr, thresh_ptr);
+
+  if (0 == early_exit) {
+    early_exit = vpx_vt_lpf_t16_16w((transposed_input + 16 * 8), src, pitch,
+                                    &filter48[0]);
+
+    if (0 == early_exit) {
+      transpose_16x16(transposed_input, 16, (src - 8), pitch);
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_4_msa.c b/libvpx/libvpx/vpx_dsp/mips/loopfilter_4_msa.c
new file mode 100644
index 0000000..9363470
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_4_msa.c
@@ -0,0 +1,146 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/loopfilter_msa.h"
+
+void vpx_lpf_horizontal_4_msa(uint8_t *src, int32_t pitch,
+                              const uint8_t *b_limit_ptr,
+                              const uint8_t *limit_ptr,
+                              const uint8_t *thresh_ptr) {
+  uint64_t p1_d, p0_d, q0_d, q1_d;
+  v16u8 mask, hev, flat, thresh, b_limit, limit;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p1_out, p0_out, q0_out, q1_out;
+
+  /* load vector elements */
+  LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+  p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+  q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+  q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+  SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch);
+}
+
+void vpx_lpf_horizontal_4_dual_msa(uint8_t *src, int32_t pitch,
+                                   const uint8_t *b_limit0_ptr,
+                                   const uint8_t *limit0_ptr,
+                                   const uint8_t *thresh0_ptr,
+                                   const uint8_t *b_limit1_ptr,
+                                   const uint8_t *limit1_ptr,
+                                   const uint8_t *thresh1_ptr) {
+  v16u8 mask, hev, flat, thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+
+  /* load vector elements */
+  LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
+  thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
+  thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
+
+  b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
+  b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
+  b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
+
+  limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
+  limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
+  limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
+               hev, mask, flat);
+  VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+
+  ST_UB4(p1, p0, q0, q1, (src - 2 * pitch), pitch);
+}
+
+void vpx_lpf_vertical_4_msa(uint8_t *src, int32_t pitch,
+                            const uint8_t *b_limit_ptr,
+                            const uint8_t *limit_ptr,
+                            const uint8_t *thresh_ptr) {
+  v16u8 mask, hev, flat, limit, thresh, b_limit;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v8i16 vec0, vec1, vec2, vec3;
+
+  LD_UB8((src - 4), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3,
+                     p3, p2, p1, p0, q0, q1, q2, q3);
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+  ILVR_B2_SH(p0, p1, q1, q0, vec0, vec1);
+  ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+
+  src -= 2;
+  ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+  src += 4 * pitch;
+  ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+}
+
+void vpx_lpf_vertical_4_dual_msa(uint8_t *src, int32_t pitch,
+                                 const uint8_t *b_limit0_ptr,
+                                 const uint8_t *limit0_ptr,
+                                 const uint8_t *thresh0_ptr,
+                                 const uint8_t *b_limit1_ptr,
+                                 const uint8_t *limit1_ptr,
+                                 const uint8_t *thresh1_ptr) {
+  v16u8 mask, hev, flat;
+  v16u8 thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
+  v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
+  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+
+  LD_UB8(src - 4, pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+  LD_UB8(src - 4 + (8 * pitch), pitch,
+         row8, row9, row10, row11, row12, row13, row14, row15);
+
+  TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
+                      row8, row9, row10, row11, row12, row13, row14, row15,
+                      p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
+  thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
+  thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
+
+  b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
+  b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
+  b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
+
+  limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
+  limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
+  limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
+               hev, mask, flat);
+  VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+  ILVR_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
+  ILVRL_H2_SH(tmp1, tmp0, tmp2, tmp3);
+  ILVL_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
+  ILVRL_H2_SH(tmp1, tmp0, tmp4, tmp5);
+
+  src -= 2;
+
+  ST4x8_UB(tmp2, tmp3, src, pitch);
+  src += (8 * pitch);
+  ST4x8_UB(tmp4, tmp5, src, pitch);
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_8_msa.c b/libvpx/libvpx/vpx_dsp/mips/loopfilter_8_msa.c
new file mode 100644
index 0000000..5b22bd0
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_8_msa.c
@@ -0,0 +1,342 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/loopfilter_msa.h"
+
+void vpx_lpf_horizontal_8_msa(uint8_t *src, int32_t pitch,
+                              const uint8_t *b_limit_ptr,
+                              const uint8_t *limit_ptr,
+                              const uint8_t *thresh_ptr) {
+  uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
+  v16u8 mask, hev, flat, thresh, b_limit, limit;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v8i16 p2_filter8, p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r;
+  v16i8 zero = { 0 };
+
+  /* load vector elements */
+  LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+  if (__msa_test_bz_v(flat)) {
+    p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+    p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+    q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+    q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+    SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
+               q2_r, q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8,
+                p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8,
+                zero, q0_filter8, p2_filter8, p1_filter8, p0_filter8,
+                q0_filter8);
+    PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat);
+
+    p2_d = __msa_copy_u_d((v2i64)p2_out, 0);
+    p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+    p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+    q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+    q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+    q2_d = __msa_copy_u_d((v2i64)q2_out, 0);
+
+    src -= 3 * pitch;
+
+    SD4(p2_d, p1_d, p0_d, q0_d, src, pitch);
+    src += (4 * pitch);
+    SD(q1_d, src);
+    src += pitch;
+    SD(q2_d, src);
+  }
+}
+
+void vpx_lpf_horizontal_8_dual_msa(uint8_t *src, int32_t pitch,
+                                   const uint8_t *b_limit0,
+                                   const uint8_t *limit0,
+                                   const uint8_t *thresh0,
+                                   const uint8_t *b_limit1,
+                                   const uint8_t *limit1,
+                                   const uint8_t *thresh1) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v16u8 flat, mask, hev, tmp, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+  v16u8 zero = { 0 };
+
+  /* load vector elements */
+  LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh0);
+  tmp = (v16u8)__msa_fill_b(*thresh1);
+  thresh = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)thresh);
+
+  b_limit = (v16u8)__msa_fill_b(*b_limit0);
+  tmp = (v16u8)__msa_fill_b(*b_limit1);
+  b_limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)b_limit);
+
+  limit = (v16u8)__msa_fill_b(*limit0);
+  tmp = (v16u8)__msa_fill_b(*limit1);
+  limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)limit);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  if (__msa_test_bz_v(flat)) {
+    ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
+               q2_r, q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+    ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+    VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+                p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+                p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    src -= 3 * pitch;
+
+    ST_UB4(p2_out, p1_out, p0_out, q0_out, src, pitch);
+    src += (4 * pitch);
+    ST_UB2(q1_out, q2_out, src, pitch);
+    src += (2 * pitch);
+  }
+}
+
+void vpx_lpf_vertical_8_msa(uint8_t *src, int32_t pitch,
+                            const uint8_t *b_limit_ptr,
+                            const uint8_t *limit_ptr,
+                            const uint8_t *thresh_ptr) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p1_out, p0_out, q0_out, q1_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v16u8 zero = { 0 };
+  v8i16 vec0, vec1, vec2, vec3, vec4;
+
+  /* load vector elements */
+  LD_UB8(src - 4, pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3,
+                     p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  /* flat4 */
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  /* filter4 */
+  VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+  if (__msa_test_bz_v(flat)) {
+    /* Store 4 pixels p1-_q1 */
+    ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+
+    src -= 2;
+    ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+    src += 4 * pitch;
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+               q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_r, p2_filt8_r, p1_filt8_r, p1_filt8_r, p0_filt8_r,
+                p0_filt8_r, q0_filt8_r, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_r, q1_filt8_r, q2_filt8_r, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    /* Store 6 pixels p2-_q2 */
+    ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+    vec4 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1);
+
+    src -= 3;
+    ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec4, 0, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec4, 4, src + 4, pitch);
+  }
+}
+
+void vpx_lpf_vertical_8_dual_msa(uint8_t *src, int32_t pitch,
+                                 const uint8_t *b_limit0,
+                                 const uint8_t *limit0,
+                                 const uint8_t *thresh0,
+                                 const uint8_t *b_limit1,
+                                 const uint8_t *limit1,
+                                 const uint8_t *thresh1) {
+  uint8_t *temp_src;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p1_out, p0_out, q0_out, q1_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v16u8 row4, row5, row6, row7, row12, row13, row14, row15;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+  v16u8 zero = { 0 };
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+
+  temp_src = src - 4;
+
+  LD_UB8(temp_src, pitch, p0, p1, p2, p3, row4, row5, row6, row7);
+  temp_src += (8 * pitch);
+  LD_UB8(temp_src, pitch, q3, q2, q1, q0, row12, row13, row14, row15);
+
+  /* transpose 16x8 matrix into 8x16 */
+  TRANSPOSE16x8_UB_UB(p0, p1, p2, p3, row4, row5, row6, row7,
+                      q3, q2, q1, q0, row12, row13, row14, row15,
+                      p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh0);
+  vec0 = (v8i16)__msa_fill_b(*thresh1);
+  thresh = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)thresh);
+
+  b_limit = (v16u8)__msa_fill_b(*b_limit0);
+  vec0 = (v8i16)__msa_fill_b(*b_limit1);
+  b_limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)b_limit);
+
+  limit = (v16u8)__msa_fill_b(*limit0);
+  vec0 = (v8i16)__msa_fill_b(*limit1);
+  limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)limit);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+               hev, mask, flat);
+  /* flat4 */
+  VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  /* filter4 */
+  VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  if (__msa_test_bz_v(flat)) {
+    ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+    ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec4, vec5);
+
+    src -= 2;
+    ST4x8_UB(vec2, vec3, src, pitch);
+    src += 8 * pitch;
+    ST4x8_UB(vec4, vec5, src, pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+               zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+               q3_r);
+    VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+    ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+
+    /* filter8 */
+    VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+                p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+                p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec3, vec4);
+    ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec6, vec7);
+    ILVRL_B2_SH(q2, q1, vec2, vec5);
+
+    src -= 3;
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec2, 0, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec2, 4, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec5, 0, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec5, 4, src + 4, pitch);
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_filters_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/loopfilter_filters_dspr2.c
new file mode 100644
index 0000000..8414b9e
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_filters_dspr2.c
@@ -0,0 +1,358 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/mips/common_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_filters_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_macros_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_masks_dspr2.h"
+#include "vpx_mem/vpx_mem.h"
+
+#if HAVE_DSPR2
+void vpx_lpf_horizontal_4_dspr2(unsigned char *s,
+                                int pitch,
+                                const uint8_t *blimit,
+                                const uint8_t *limit,
+                                const uint8_t *thresh) {
+  uint8_t   i;
+  uint32_t  mask;
+  uint32_t  hev;
+  uint32_t  pm1, p0, p1, p2, p3, p4, p5, p6;
+  uint8_t   *sm1, *s0, *s1, *s2, *s3, *s4, *s5, *s6;
+  uint32_t  thresh_vec, flimit_vec, limit_vec;
+  uint32_t  uflimit, ulimit, uthresh;
+
+  uflimit = *blimit;
+  ulimit = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__ (
+      "replv.qb       %[thresh_vec],    %[uthresh]    \n\t"
+      "replv.qb       %[flimit_vec],    %[uflimit]    \n\t"
+      "replv.qb       %[limit_vec],     %[ulimit]     \n\t"
+
+      : [thresh_vec] "=&r" (thresh_vec), [flimit_vec] "=&r" (flimit_vec),
+        [limit_vec] "=r" (limit_vec)
+      : [uthresh] "r" (uthresh), [uflimit] "r" (uflimit), [ulimit] "r" (ulimit)
+  );
+
+  /* prefetch data for store */
+  prefetch_store(s);
+
+  /* loop filter designed to work using chars so that we can make maximum use
+     of 8 bit simd instructions. */
+  for (i = 0; i < 2; i++) {
+    sm1 = s - (pitch << 2);
+    s0 = sm1 + pitch;
+    s1 = s0 + pitch;
+    s2 = s - pitch;
+    s3 = s;
+    s4 = s + pitch;
+    s5 = s4 + pitch;
+    s6 = s5 + pitch;
+
+    __asm__ __volatile__ (
+        "lw     %[p1],  (%[s1])    \n\t"
+        "lw     %[p2],  (%[s2])    \n\t"
+        "lw     %[p3],  (%[s3])    \n\t"
+        "lw     %[p4],  (%[s4])    \n\t"
+
+        : [p1] "=&r" (p1), [p2] "=&r" (p2), [p3] "=&r" (p3), [p4] "=&r" (p4)
+        : [s1] "r" (s1), [s2] "r" (s2), [s3] "r" (s3), [s4] "r" (s4)
+    );
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+       mask will be zero and filtering is not needed */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0))) {
+      __asm__ __volatile__ (
+          "lw       %[pm1], (%[sm1])   \n\t"
+          "lw       %[p0],  (%[s0])    \n\t"
+          "lw       %[p5],  (%[s5])    \n\t"
+          "lw       %[p6],  (%[s6])    \n\t"
+
+          : [pm1] "=&r" (pm1), [p0] "=&r" (p0), [p5] "=&r" (p5),
+            [p6] "=&r" (p6)
+          : [sm1] "r" (sm1), [s0] "r" (s0), [s5] "r" (s5), [s6] "r" (s6)
+      );
+
+      filter_hev_mask_dspr2(limit_vec, flimit_vec, p1, p2,
+                            pm1, p0, p3, p4, p5, p6,
+                            thresh_vec, &hev, &mask);
+
+      /* if mask == 0 do filtering is not needed */
+      if (mask) {
+        /* filtering */
+        filter_dspr2(mask, hev, &p1, &p2, &p3, &p4);
+
+        __asm__ __volatile__ (
+            "sw     %[p1],  (%[s1])    \n\t"
+            "sw     %[p2],  (%[s2])    \n\t"
+            "sw     %[p3],  (%[s3])    \n\t"
+            "sw     %[p4],  (%[s4])    \n\t"
+
+            :
+            : [p1] "r" (p1), [p2] "r" (p2), [p3] "r" (p3), [p4] "r" (p4),
+              [s1] "r" (s1), [s2] "r" (s2), [s3] "r" (s3), [s4] "r" (s4)
+        );
+      }
+    }
+
+    s = s + 4;
+  }
+}
+
+void vpx_lpf_vertical_4_dspr2(unsigned char *s,
+                              int pitch,
+                              const uint8_t *blimit,
+                              const uint8_t *limit,
+                              const uint8_t *thresh) {
+  uint8_t   i;
+  uint32_t  mask, hev;
+  uint32_t  pm1, p0, p1, p2, p3, p4, p5, p6;
+  uint8_t   *s1, *s2, *s3, *s4;
+  uint32_t  prim1, prim2, sec3, sec4, prim3, prim4;
+  uint32_t  thresh_vec, flimit_vec, limit_vec;
+  uint32_t  uflimit, ulimit, uthresh;
+
+  uflimit = *blimit;
+  ulimit = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__ (
+      "replv.qb       %[thresh_vec],    %[uthresh]    \n\t"
+      "replv.qb       %[flimit_vec],    %[uflimit]    \n\t"
+      "replv.qb       %[limit_vec],     %[ulimit]     \n\t"
+
+      : [thresh_vec] "=&r" (thresh_vec), [flimit_vec] "=&r" (flimit_vec),
+        [limit_vec] "=r" (limit_vec)
+      : [uthresh] "r" (uthresh), [uflimit] "r" (uflimit), [ulimit] "r" (ulimit)
+  );
+
+  /* prefetch data for store */
+  prefetch_store(s + pitch);
+
+  for (i = 0; i < 2; i++) {
+    s1 = s;
+    s2 = s + pitch;
+    s3 = s2 + pitch;
+    s4 = s3 + pitch;
+    s  = s4 + pitch;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p2  = *((uint32_t *)(s1 - 4));
+    p6  = *((uint32_t *)(s1));
+    p1  = *((uint32_t *)(s2 - 4));
+    p5  = *((uint32_t *)(s2));
+    p0  = *((uint32_t *)(s3 - 4));
+    p4  = *((uint32_t *)(s3));
+    pm1 = *((uint32_t *)(s4 - 4));
+    p3  = *((uint32_t *)(s4));
+
+    /* transpose pm1, p0, p1, p2 */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p2],      %[p1]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p2],      %[p1]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p0],      %[pm1]      \n\t"
+        "precr.qb.ph    %[prim4],   %[p0],      %[pm1]      \n\t"
+
+        "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[pm1],     %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p2],      %[p1],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p0],      %[pm1],     %[sec4]     \n\t"
+        "append         %[p1],      %[sec3],    16          \n\t"
+        "append         %[pm1],     %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0), [pm1] "+r" (pm1),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* transpose p3, p4, p5, p6 */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p6],      %[p5]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p6],      %[p5]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p4],      %[p3]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p4],      %[p3]       \n\t"
+
+        "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p6],      %[p5],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p4],      %[p3],      %[sec4]     \n\t"
+        "append         %[p5],      %[sec3],    16          \n\t"
+        "append         %[p3],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p6] "+r" (p6), [p5] "+r" (p5), [p4] "+r" (p4), [p3] "+r" (p3),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0))) {
+      filter_hev_mask_dspr2(limit_vec, flimit_vec, p1, p2, pm1,
+                            p0, p3, p4, p5, p6, thresh_vec,
+                            &hev, &mask);
+
+      /* if mask == 0 do filtering is not needed */
+      if (mask) {
+        /* filtering */
+        filter_dspr2(mask, hev, &p1, &p2, &p3, &p4);
+
+        /* unpack processed 4x4 neighborhood
+         * don't use transpose on output data
+         * because memory isn't aligned
+         */
+        __asm__ __volatile__ (
+            "sb     %[p4],   1(%[s4])    \n\t"
+            "sb     %[p3],   0(%[s4])    \n\t"
+            "sb     %[p2],  -1(%[s4])    \n\t"
+            "sb     %[p1],  -2(%[s4])    \n\t"
+
+            :
+            : [p4] "r" (p4), [p3] "r" (p3), [p2] "r" (p2), [p1] "r" (p1),
+              [s4] "r" (s4)
+        );
+
+        __asm__ __volatile__ (
+            "srl    %[p4],  %[p4],  8     \n\t"
+            "srl    %[p3],  %[p3],  8     \n\t"
+            "srl    %[p2],  %[p2],  8     \n\t"
+            "srl    %[p1],  %[p1],  8     \n\t"
+
+            : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+            :
+        );
+
+        __asm__ __volatile__ (
+            "sb     %[p4],   1(%[s3])    \n\t"
+            "sb     %[p3],   0(%[s3])    \n\t"
+            "sb     %[p2],  -1(%[s3])    \n\t"
+            "sb     %[p1],  -2(%[s3])    \n\t"
+
+            : [p1] "+r" (p1)
+            : [p4] "r" (p4), [p3] "r" (p3), [p2] "r" (p2), [s3] "r" (s3)
+        );
+
+        __asm__ __volatile__ (
+            "srl    %[p4],  %[p4],  8     \n\t"
+            "srl    %[p3],  %[p3],  8     \n\t"
+            "srl    %[p2],  %[p2],  8     \n\t"
+            "srl    %[p1],  %[p1],  8     \n\t"
+
+            : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+            :
+        );
+
+        __asm__ __volatile__ (
+            "sb     %[p4],   1(%[s2])    \n\t"
+            "sb     %[p3],   0(%[s2])    \n\t"
+            "sb     %[p2],  -1(%[s2])    \n\t"
+            "sb     %[p1],  -2(%[s2])    \n\t"
+
+            :
+            : [p4] "r" (p4), [p3] "r" (p3), [p2] "r" (p2), [p1] "r" (p1),
+              [s2] "r" (s2)
+        );
+
+        __asm__ __volatile__ (
+            "srl    %[p4],  %[p4],  8     \n\t"
+            "srl    %[p3],  %[p3],  8     \n\t"
+            "srl    %[p2],  %[p2],  8     \n\t"
+            "srl    %[p1],  %[p1],  8     \n\t"
+
+            : [p4] "+r" (p4), [p3] "+r" (p3), [p2] "+r" (p2), [p1] "+r" (p1)
+            :
+        );
+
+        __asm__ __volatile__ (
+            "sb     %[p4],   1(%[s1])    \n\t"
+            "sb     %[p3],   0(%[s1])    \n\t"
+            "sb     %[p2],  -1(%[s1])    \n\t"
+            "sb     %[p1],  -2(%[s1])    \n\t"
+
+            :
+            : [p4] "r" (p4), [p3] "r" (p3), [p2] "r" (p2), [p1] "r" (p1),
+              [s1] "r" (s1)
+        );
+      }
+    }
+  }
+}
+
+void vpx_lpf_horizontal_4_dual_dspr2(uint8_t *s, int p /* pitch */,
+                                     const uint8_t *blimit0,
+                                     const uint8_t *limit0,
+                                     const uint8_t *thresh0,
+                                     const uint8_t *blimit1,
+                                     const uint8_t *limit1,
+                                     const uint8_t *thresh1) {
+  vpx_lpf_horizontal_4_dspr2(s, p, blimit0, limit0, thresh0);
+  vpx_lpf_horizontal_4_dspr2(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void vpx_lpf_horizontal_8_dual_dspr2(uint8_t *s, int p /* pitch */,
+                                     const uint8_t *blimit0,
+                                     const uint8_t *limit0,
+                                     const uint8_t *thresh0,
+                                     const uint8_t *blimit1,
+                                     const uint8_t *limit1,
+                                     const uint8_t *thresh1) {
+  vpx_lpf_horizontal_8_dspr2(s, p, blimit0, limit0, thresh0);
+  vpx_lpf_horizontal_8_dspr2(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void vpx_lpf_vertical_4_dual_dspr2(uint8_t *s, int p,
+                                   const uint8_t *blimit0,
+                                   const uint8_t *limit0,
+                                   const uint8_t *thresh0,
+                                   const uint8_t *blimit1,
+                                   const uint8_t *limit1,
+                                   const uint8_t *thresh1) {
+  vpx_lpf_vertical_4_dspr2(s, p, blimit0, limit0, thresh0);
+  vpx_lpf_vertical_4_dspr2(s + 8 * p, p, blimit1, limit1, thresh1);
+}
+
+void vpx_lpf_vertical_8_dual_dspr2(uint8_t *s, int p,
+                                   const uint8_t *blimit0,
+                                   const uint8_t *limit0,
+                                   const uint8_t *thresh0,
+                                   const uint8_t *blimit1,
+                                   const uint8_t *limit1,
+                                   const uint8_t *thresh1) {
+  vpx_lpf_vertical_8_dspr2(s, p, blimit0, limit0, thresh0);
+  vpx_lpf_vertical_8_dspr2(s + 8 * p, p, blimit1, limit1, thresh1);
+}
+
+void vpx_lpf_vertical_16_dual_dspr2(uint8_t *s, int p,
+                                    const uint8_t *blimit,
+                                    const uint8_t *limit,
+                                    const uint8_t *thresh) {
+  vpx_lpf_vertical_16_dspr2(s, p, blimit, limit, thresh);
+  vpx_lpf_vertical_16_dspr2(s + 8 * p, p, blimit, limit, thresh);
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_filters_dspr2.h b/libvpx/libvpx/vpx_dsp/mips/loopfilter_filters_dspr2.h
new file mode 100644
index 0000000..4a1506b
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_filters_dspr2.h
@@ -0,0 +1,764 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_MIPS_DSPR2_VP9_LOOPFILTER_FILTERS_DSPR2_H_
+#define VP9_COMMON_MIPS_DSPR2_VP9_LOOPFILTER_FILTERS_DSPR2_H_
+
+#include <stdlib.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if HAVE_DSPR2
+/* inputs & outputs are quad-byte vectors */
+static INLINE void filter_dspr2(uint32_t mask, uint32_t hev,
+                                uint32_t *ps1, uint32_t *ps0,
+                                uint32_t *qs0, uint32_t *qs1) {
+  int32_t   vpx_filter_l, vpx_filter_r;
+  int32_t   Filter1_l, Filter1_r, Filter2_l, Filter2_r;
+  int32_t   subr_r, subr_l;
+  uint32_t  t1, t2, HWM, t3;
+  uint32_t  hev_l, hev_r, mask_l, mask_r, invhev_l, invhev_r;
+  int32_t   vps1, vps0, vqs0, vqs1;
+  int32_t   vps1_l, vps1_r, vps0_l, vps0_r, vqs0_l, vqs0_r, vqs1_l, vqs1_r;
+  uint32_t  N128;
+
+  N128 = 0x80808080;
+  t1  = 0x03000300;
+  t2  = 0x04000400;
+  t3  = 0x01000100;
+  HWM = 0xFF00FF00;
+
+  vps0 = (*ps0) ^ N128;
+  vps1 = (*ps1) ^ N128;
+  vqs0 = (*qs0) ^ N128;
+  vqs1 = (*qs1) ^ N128;
+
+  /* use halfword pairs instead quad-bytes because of accuracy */
+  vps0_l = vps0 & HWM;
+  vps0_r = vps0 << 8;
+  vps0_r = vps0_r & HWM;
+
+  vps1_l = vps1 & HWM;
+  vps1_r = vps1 << 8;
+  vps1_r = vps1_r & HWM;
+
+  vqs0_l = vqs0 & HWM;
+  vqs0_r = vqs0 << 8;
+  vqs0_r = vqs0_r & HWM;
+
+  vqs1_l = vqs1 & HWM;
+  vqs1_r = vqs1 << 8;
+  vqs1_r = vqs1_r & HWM;
+
+  mask_l = mask & HWM;
+  mask_r = mask << 8;
+  mask_r = mask_r & HWM;
+
+  hev_l = hev & HWM;
+  hev_r = hev << 8;
+  hev_r = hev_r & HWM;
+
+  __asm__ __volatile__ (
+      /* vpx_filter = vp8_signed_char_clamp(ps1 - qs1); */
+      "subq_s.ph    %[vpx_filter_l], %[vps1_l],       %[vqs1_l]       \n\t"
+      "subq_s.ph    %[vpx_filter_r], %[vps1_r],       %[vqs1_r]       \n\t"
+
+      /* qs0 - ps0 */
+      "subq_s.ph    %[subr_l],       %[vqs0_l],       %[vps0_l]       \n\t"
+      "subq_s.ph    %[subr_r],       %[vqs0_r],       %[vps0_r]       \n\t"
+
+      /* vpx_filter &= hev; */
+      "and          %[vpx_filter_l], %[vpx_filter_l], %[hev_l]        \n\t"
+      "and          %[vpx_filter_r], %[vpx_filter_r], %[hev_r]        \n\t"
+
+      /* vpx_filter = vp8_signed_char_clamp(vpx_filter + 3 * (qs0 - ps0)); */
+      "addq_s.ph    %[vpx_filter_l], %[vpx_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[vpx_filter_r], %[vpx_filter_r], %[subr_r]       \n\t"
+      "xor          %[invhev_l],     %[hev_l],        %[HWM]          \n\t"
+      "addq_s.ph    %[vpx_filter_l], %[vpx_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[vpx_filter_r], %[vpx_filter_r], %[subr_r]       \n\t"
+      "xor          %[invhev_r],     %[hev_r],        %[HWM]          \n\t"
+      "addq_s.ph    %[vpx_filter_l], %[vpx_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[vpx_filter_r], %[vpx_filter_r], %[subr_r]       \n\t"
+
+      /* vpx_filter &= mask; */
+      "and          %[vpx_filter_l], %[vpx_filter_l], %[mask_l]       \n\t"
+      "and          %[vpx_filter_r], %[vpx_filter_r], %[mask_r]       \n\t"
+
+      : [vpx_filter_l] "=&r" (vpx_filter_l),
+        [vpx_filter_r] "=&r" (vpx_filter_r),
+        [subr_l] "=&r" (subr_l), [subr_r] "=&r" (subr_r),
+        [invhev_l] "=&r" (invhev_l), [invhev_r] "=&r" (invhev_r)
+      : [vps0_l] "r" (vps0_l), [vps0_r] "r" (vps0_r), [vps1_l] "r" (vps1_l),
+        [vps1_r] "r" (vps1_r), [vqs0_l] "r" (vqs0_l), [vqs0_r] "r" (vqs0_r),
+        [vqs1_l] "r" (vqs1_l), [vqs1_r] "r" (vqs1_r),
+        [mask_l] "r" (mask_l), [mask_r] "r" (mask_r),
+        [hev_l] "r" (hev_l), [hev_r] "r" (hev_r),
+        [HWM] "r" (HWM)
+  );
+
+  /* save bottom 3 bits so that we round one side +4 and the other +3 */
+  __asm__ __volatile__ (
+      /* Filter2 = vp8_signed_char_clamp(vpx_filter + 3) >>= 3; */
+      "addq_s.ph    %[Filter1_l],    %[vpx_filter_l], %[t2]           \n\t"
+      "addq_s.ph    %[Filter1_r],    %[vpx_filter_r], %[t2]           \n\t"
+
+      /* Filter1 = vp8_signed_char_clamp(vpx_filter + 4) >>= 3; */
+      "addq_s.ph    %[Filter2_l],    %[vpx_filter_l], %[t1]           \n\t"
+      "addq_s.ph    %[Filter2_r],    %[vpx_filter_r], %[t1]           \n\t"
+      "shra.ph      %[Filter1_r],    %[Filter1_r],    3               \n\t"
+      "shra.ph      %[Filter1_l],    %[Filter1_l],    3               \n\t"
+
+      "shra.ph      %[Filter2_l],    %[Filter2_l],    3               \n\t"
+      "shra.ph      %[Filter2_r],    %[Filter2_r],    3               \n\t"
+
+      "and          %[Filter1_l],    %[Filter1_l],    %[HWM]          \n\t"
+      "and          %[Filter1_r],    %[Filter1_r],    %[HWM]          \n\t"
+
+      /* vps0 = vp8_signed_char_clamp(ps0 + Filter2); */
+      "addq_s.ph    %[vps0_l],       %[vps0_l],       %[Filter2_l]    \n\t"
+      "addq_s.ph    %[vps0_r],       %[vps0_r],       %[Filter2_r]    \n\t"
+
+      /* vqs0 = vp8_signed_char_clamp(qs0 - Filter1); */
+      "subq_s.ph    %[vqs0_l],       %[vqs0_l],       %[Filter1_l]    \n\t"
+      "subq_s.ph    %[vqs0_r],       %[vqs0_r],       %[Filter1_r]    \n\t"
+
+      : [Filter1_l] "=&r" (Filter1_l), [Filter1_r] "=&r" (Filter1_r),
+        [Filter2_l] "=&r" (Filter2_l), [Filter2_r] "=&r" (Filter2_r),
+        [vps0_l] "+r" (vps0_l), [vps0_r] "+r" (vps0_r),
+        [vqs0_l] "+r" (vqs0_l), [vqs0_r] "+r" (vqs0_r)
+      : [t1] "r" (t1), [t2] "r" (t2), [HWM] "r" (HWM),
+        [vpx_filter_l] "r" (vpx_filter_l), [vpx_filter_r] "r" (vpx_filter_r)
+  );
+
+  __asm__ __volatile__ (
+      /* (vpx_filter += 1) >>= 1 */
+      "addqh.ph    %[Filter1_l],    %[Filter1_l],     %[t3]           \n\t"
+      "addqh.ph    %[Filter1_r],    %[Filter1_r],     %[t3]           \n\t"
+
+      /* vpx_filter &= ~hev; */
+      "and          %[Filter1_l],    %[Filter1_l],    %[invhev_l]     \n\t"
+      "and          %[Filter1_r],    %[Filter1_r],    %[invhev_r]     \n\t"
+
+      /* vps1 = vp8_signed_char_clamp(ps1 + vpx_filter); */
+      "addq_s.ph    %[vps1_l],       %[vps1_l],       %[Filter1_l]    \n\t"
+      "addq_s.ph    %[vps1_r],       %[vps1_r],       %[Filter1_r]    \n\t"
+
+      /* vqs1 = vp8_signed_char_clamp(qs1 - vpx_filter); */
+      "subq_s.ph    %[vqs1_l],       %[vqs1_l],       %[Filter1_l]    \n\t"
+      "subq_s.ph    %[vqs1_r],       %[vqs1_r],       %[Filter1_r]    \n\t"
+
+      : [Filter1_l] "+r" (Filter1_l), [Filter1_r] "+r" (Filter1_r),
+        [vps1_l] "+r" (vps1_l), [vps1_r] "+r" (vps1_r),
+        [vqs1_l] "+r" (vqs1_l), [vqs1_r] "+r" (vqs1_r)
+      : [t3] "r" (t3), [invhev_l] "r" (invhev_l), [invhev_r] "r" (invhev_r)
+  );
+
+  /* Create quad-bytes from halfword pairs */
+  vqs0_l = vqs0_l & HWM;
+  vqs1_l = vqs1_l & HWM;
+  vps0_l = vps0_l & HWM;
+  vps1_l = vps1_l & HWM;
+
+  __asm__ __volatile__ (
+      "shrl.ph      %[vqs0_r],       %[vqs0_r],       8   \n\t"
+      "shrl.ph      %[vps0_r],       %[vps0_r],       8   \n\t"
+      "shrl.ph      %[vqs1_r],       %[vqs1_r],       8   \n\t"
+      "shrl.ph      %[vps1_r],       %[vps1_r],       8   \n\t"
+
+      : [vps1_r] "+r" (vps1_r), [vqs1_r] "+r" (vqs1_r),
+        [vps0_r] "+r" (vps0_r), [vqs0_r] "+r" (vqs0_r)
+      :
+  );
+
+  vqs0 = vqs0_l | vqs0_r;
+  vqs1 = vqs1_l | vqs1_r;
+  vps0 = vps0_l | vps0_r;
+  vps1 = vps1_l | vps1_r;
+
+  *ps0 = vps0 ^ N128;
+  *ps1 = vps1 ^ N128;
+  *qs0 = vqs0 ^ N128;
+  *qs1 = vqs1 ^ N128;
+}
+
+static INLINE void filter1_dspr2(uint32_t mask, uint32_t hev,
+                                 uint32_t ps1, uint32_t ps0,
+                                 uint32_t qs0, uint32_t qs1,
+                                 uint32_t *p1_f0, uint32_t *p0_f0,
+                                 uint32_t *q0_f0, uint32_t *q1_f0) {
+  int32_t   vpx_filter_l, vpx_filter_r;
+  int32_t   Filter1_l, Filter1_r, Filter2_l, Filter2_r;
+  int32_t   subr_r, subr_l;
+  uint32_t  t1, t2, HWM, t3;
+  uint32_t  hev_l, hev_r, mask_l, mask_r, invhev_l, invhev_r;
+  int32_t   vps1, vps0, vqs0, vqs1;
+  int32_t   vps1_l, vps1_r, vps0_l, vps0_r, vqs0_l, vqs0_r, vqs1_l, vqs1_r;
+  uint32_t  N128;
+
+  N128 = 0x80808080;
+  t1  = 0x03000300;
+  t2  = 0x04000400;
+  t3  = 0x01000100;
+  HWM = 0xFF00FF00;
+
+  vps0 = (ps0) ^ N128;
+  vps1 = (ps1) ^ N128;
+  vqs0 = (qs0) ^ N128;
+  vqs1 = (qs1) ^ N128;
+
+  /* use halfword pairs instead quad-bytes because of accuracy */
+  vps0_l = vps0 & HWM;
+  vps0_r = vps0 << 8;
+  vps0_r = vps0_r & HWM;
+
+  vps1_l = vps1 & HWM;
+  vps1_r = vps1 << 8;
+  vps1_r = vps1_r & HWM;
+
+  vqs0_l = vqs0 & HWM;
+  vqs0_r = vqs0 << 8;
+  vqs0_r = vqs0_r & HWM;
+
+  vqs1_l = vqs1 & HWM;
+  vqs1_r = vqs1 << 8;
+  vqs1_r = vqs1_r & HWM;
+
+  mask_l = mask & HWM;
+  mask_r = mask << 8;
+  mask_r = mask_r & HWM;
+
+  hev_l = hev & HWM;
+  hev_r = hev << 8;
+  hev_r = hev_r & HWM;
+
+  __asm__ __volatile__ (
+      /* vpx_filter = vp8_signed_char_clamp(ps1 - qs1); */
+      "subq_s.ph    %[vpx_filter_l], %[vps1_l],       %[vqs1_l]       \n\t"
+      "subq_s.ph    %[vpx_filter_r], %[vps1_r],       %[vqs1_r]       \n\t"
+
+      /* qs0 - ps0 */
+      "subq_s.ph    %[subr_l],       %[vqs0_l],       %[vps0_l]       \n\t"
+      "subq_s.ph    %[subr_r],       %[vqs0_r],       %[vps0_r]       \n\t"
+
+      /* vpx_filter &= hev; */
+      "and          %[vpx_filter_l], %[vpx_filter_l], %[hev_l]        \n\t"
+      "and          %[vpx_filter_r], %[vpx_filter_r], %[hev_r]        \n\t"
+
+      /* vpx_filter = vp8_signed_char_clamp(vpx_filter + 3 * (qs0 - ps0)); */
+      "addq_s.ph    %[vpx_filter_l], %[vpx_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[vpx_filter_r], %[vpx_filter_r], %[subr_r]       \n\t"
+      "xor          %[invhev_l],     %[hev_l],        %[HWM]          \n\t"
+      "addq_s.ph    %[vpx_filter_l], %[vpx_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[vpx_filter_r], %[vpx_filter_r], %[subr_r]       \n\t"
+      "xor          %[invhev_r],     %[hev_r],        %[HWM]          \n\t"
+      "addq_s.ph    %[vpx_filter_l], %[vpx_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[vpx_filter_r], %[vpx_filter_r], %[subr_r]       \n\t"
+
+      /* vpx_filter &= mask; */
+      "and          %[vpx_filter_l], %[vpx_filter_l], %[mask_l]       \n\t"
+      "and          %[vpx_filter_r], %[vpx_filter_r], %[mask_r]       \n\t"
+
+      : [vpx_filter_l] "=&r" (vpx_filter_l),
+        [vpx_filter_r] "=&r" (vpx_filter_r),
+        [subr_l] "=&r" (subr_l), [subr_r] "=&r" (subr_r),
+        [invhev_l] "=&r" (invhev_l), [invhev_r] "=&r" (invhev_r)
+      : [vps0_l] "r" (vps0_l), [vps0_r] "r" (vps0_r), [vps1_l] "r" (vps1_l),
+        [vps1_r] "r" (vps1_r), [vqs0_l] "r" (vqs0_l), [vqs0_r] "r" (vqs0_r),
+        [vqs1_l] "r" (vqs1_l), [vqs1_r] "r" (vqs1_r),
+        [mask_l] "r" (mask_l), [mask_r] "r" (mask_r),
+        [hev_l] "r" (hev_l), [hev_r] "r" (hev_r), [HWM] "r" (HWM)
+  );
+
+  /* save bottom 3 bits so that we round one side +4 and the other +3 */
+  __asm__ __volatile__ (
+      /* Filter2 = vp8_signed_char_clamp(vpx_filter + 3) >>= 3; */
+      "addq_s.ph    %[Filter1_l],    %[vpx_filter_l], %[t2]           \n\t"
+      "addq_s.ph    %[Filter1_r],    %[vpx_filter_r], %[t2]           \n\t"
+
+      /* Filter1 = vp8_signed_char_clamp(vpx_filter + 4) >>= 3; */
+      "addq_s.ph    %[Filter2_l],    %[vpx_filter_l], %[t1]           \n\t"
+      "addq_s.ph    %[Filter2_r],    %[vpx_filter_r], %[t1]           \n\t"
+      "shra.ph      %[Filter1_r],    %[Filter1_r],    3               \n\t"
+      "shra.ph      %[Filter1_l],    %[Filter1_l],    3               \n\t"
+
+      "shra.ph      %[Filter2_l],    %[Filter2_l],    3               \n\t"
+      "shra.ph      %[Filter2_r],    %[Filter2_r],    3               \n\t"
+
+      "and          %[Filter1_l],    %[Filter1_l],    %[HWM]          \n\t"
+      "and          %[Filter1_r],    %[Filter1_r],    %[HWM]          \n\t"
+
+      /* vps0 = vp8_signed_char_clamp(ps0 + Filter2); */
+      "addq_s.ph    %[vps0_l],       %[vps0_l],       %[Filter2_l]    \n\t"
+      "addq_s.ph    %[vps0_r],       %[vps0_r],       %[Filter2_r]    \n\t"
+
+      /* vqs0 = vp8_signed_char_clamp(qs0 - Filter1); */
+      "subq_s.ph    %[vqs0_l],       %[vqs0_l],       %[Filter1_l]    \n\t"
+      "subq_s.ph    %[vqs0_r],       %[vqs0_r],       %[Filter1_r]    \n\t"
+
+      : [Filter1_l] "=&r" (Filter1_l), [Filter1_r] "=&r" (Filter1_r),
+        [Filter2_l] "=&r" (Filter2_l), [Filter2_r] "=&r" (Filter2_r),
+        [vps0_l] "+r" (vps0_l), [vps0_r] "+r" (vps0_r),
+        [vqs0_l] "+r" (vqs0_l), [vqs0_r] "+r" (vqs0_r)
+      : [t1] "r" (t1), [t2] "r" (t2), [HWM] "r" (HWM),
+        [vpx_filter_l] "r" (vpx_filter_l), [vpx_filter_r] "r" (vpx_filter_r)
+  );
+
+  __asm__ __volatile__ (
+      /* (vpx_filter += 1) >>= 1 */
+      "addqh.ph    %[Filter1_l],    %[Filter1_l],     %[t3]           \n\t"
+      "addqh.ph    %[Filter1_r],    %[Filter1_r],     %[t3]           \n\t"
+
+      /* vpx_filter &= ~hev; */
+      "and          %[Filter1_l],    %[Filter1_l],    %[invhev_l]     \n\t"
+      "and          %[Filter1_r],    %[Filter1_r],    %[invhev_r]     \n\t"
+
+      /* vps1 = vp8_signed_char_clamp(ps1 + vpx_filter); */
+      "addq_s.ph    %[vps1_l],       %[vps1_l],       %[Filter1_l]    \n\t"
+      "addq_s.ph    %[vps1_r],       %[vps1_r],       %[Filter1_r]    \n\t"
+
+      /* vqs1 = vp8_signed_char_clamp(qs1 - vpx_filter); */
+      "subq_s.ph    %[vqs1_l],       %[vqs1_l],       %[Filter1_l]    \n\t"
+      "subq_s.ph    %[vqs1_r],       %[vqs1_r],       %[Filter1_r]    \n\t"
+
+      : [Filter1_l] "+r" (Filter1_l), [Filter1_r] "+r" (Filter1_r),
+        [vps1_l] "+r" (vps1_l), [vps1_r] "+r" (vps1_r),
+        [vqs1_l] "+r" (vqs1_l), [vqs1_r] "+r" (vqs1_r)
+      : [t3] "r" (t3), [invhev_l] "r" (invhev_l), [invhev_r] "r" (invhev_r)
+  );
+
+  /* Create quad-bytes from halfword pairs */
+  vqs0_l = vqs0_l & HWM;
+  vqs1_l = vqs1_l & HWM;
+  vps0_l = vps0_l & HWM;
+  vps1_l = vps1_l & HWM;
+
+  __asm__ __volatile__ (
+      "shrl.ph      %[vqs0_r],       %[vqs0_r],       8   \n\t"
+      "shrl.ph      %[vps0_r],       %[vps0_r],       8   \n\t"
+      "shrl.ph      %[vqs1_r],       %[vqs1_r],       8   \n\t"
+      "shrl.ph      %[vps1_r],       %[vps1_r],       8   \n\t"
+
+      : [vps1_r] "+r" (vps1_r), [vqs1_r] "+r" (vqs1_r),
+        [vps0_r] "+r" (vps0_r), [vqs0_r] "+r" (vqs0_r)
+      :
+  );
+
+  vqs0 = vqs0_l | vqs0_r;
+  vqs1 = vqs1_l | vqs1_r;
+  vps0 = vps0_l | vps0_r;
+  vps1 = vps1_l | vps1_r;
+
+  *p0_f0 = vps0 ^ N128;
+  *p1_f0 = vps1 ^ N128;
+  *q0_f0 = vqs0 ^ N128;
+  *q1_f0 = vqs1 ^ N128;
+}
+
+static INLINE void mbfilter_dspr2(uint32_t *op3, uint32_t *op2,
+                                  uint32_t *op1, uint32_t *op0,
+                                  uint32_t *oq0, uint32_t *oq1,
+                                  uint32_t *oq2, uint32_t *oq3) {
+  /* use a 7 tap filter [1, 1, 1, 2, 1, 1, 1] for flat line */
+  const uint32_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+  const uint32_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+  uint32_t       res_op2, res_op1, res_op0;
+  uint32_t       res_oq0, res_oq1, res_oq2;
+  uint32_t       tmp;
+  uint32_t       add_p210_q012;
+  uint32_t       u32Four = 0x00040004;
+
+  /* *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + p2 + p2 + p1 + p0 + q0, 3)  1 */
+  /* *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + p1 + p1 + p0 + q0 + q1, 3)  2 */
+  /* *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + p0 + q0 + q1 + q2, 3)  3 */
+  /* *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + q0 + q0 + q1 + q2 + q3, 3)  4 */
+  /* *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q1 + q2 + q3 + q3, 3)  5 */
+  /* *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q2 + q3 + q3 + q3, 3)  6 */
+
+  __asm__ __volatile__ (
+      "addu.ph    %[add_p210_q012],  %[p2],             %[p1]            \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[p0]            \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q0]            \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q1]            \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q2]            \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[u32Four]       \n\t"
+
+      "shll.ph    %[tmp],            %[p3],             1                \n\t"
+      "addu.ph    %[res_op2],        %[tmp],            %[p3]            \n\t"
+      "addu.ph    %[res_op1],        %[p3],             %[p3]            \n\t"
+      "addu.ph    %[res_op2],        %[res_op2],        %[p2]            \n\t"
+      "addu.ph    %[res_op1],        %[res_op1],        %[p1]            \n\t"
+      "addu.ph    %[res_op2],        %[res_op2],        %[add_p210_q012] \n\t"
+      "addu.ph    %[res_op1],        %[res_op1],        %[add_p210_q012] \n\t"
+      "subu.ph    %[res_op2],        %[res_op2],        %[q1]            \n\t"
+      "subu.ph    %[res_op1],        %[res_op1],        %[q2]            \n\t"
+      "subu.ph    %[res_op2],        %[res_op2],        %[q2]            \n\t"
+      "shrl.ph    %[res_op1],        %[res_op1],        3                \n\t"
+      "shrl.ph    %[res_op2],        %[res_op2],        3                \n\t"
+      "addu.ph    %[res_op0],        %[p3],             %[p0]            \n\t"
+      "addu.ph    %[res_oq0],        %[q0],             %[q3]            \n\t"
+      "addu.ph    %[res_op0],        %[res_op0],        %[add_p210_q012] \n\t"
+      "addu.ph    %[res_oq0],        %[res_oq0],        %[add_p210_q012] \n\t"
+      "addu.ph    %[res_oq1],        %[q3],             %[q3]            \n\t"
+      "shll.ph    %[tmp],            %[q3],             1                \n\t"
+      "addu.ph    %[res_oq1],        %[res_oq1],        %[q1]            \n\t"
+      "addu.ph    %[res_oq2],        %[tmp],            %[q3]            \n\t"
+      "addu.ph    %[res_oq1],        %[res_oq1],        %[add_p210_q012] \n\t"
+      "addu.ph    %[res_oq2],        %[res_oq2],        %[add_p210_q012] \n\t"
+      "subu.ph    %[res_oq1],        %[res_oq1],        %[p2]            \n\t"
+      "addu.ph    %[res_oq2],        %[res_oq2],        %[q2]            \n\t"
+      "shrl.ph    %[res_oq1],        %[res_oq1],        3                \n\t"
+      "subu.ph    %[res_oq2],        %[res_oq2],        %[p2]            \n\t"
+      "shrl.ph    %[res_oq0],        %[res_oq0],        3                \n\t"
+      "subu.ph    %[res_oq2],        %[res_oq2],        %[p1]            \n\t"
+      "shrl.ph    %[res_op0],        %[res_op0],        3                \n\t"
+      "shrl.ph    %[res_oq2],        %[res_oq2],        3                \n\t"
+
+      : [add_p210_q012] "=&r" (add_p210_q012),
+        [tmp] "=&r" (tmp), [res_op2] "=&r" (res_op2),
+        [res_op1] "=&r" (res_op1), [res_op0] "=&r" (res_op0),
+        [res_oq0] "=&r" (res_oq0), [res_oq1] "=&r" (res_oq1),
+        [res_oq2] "=&r" (res_oq2)
+      : [p0] "r" (p0), [q0] "r" (q0), [p1] "r" (p1), [q1] "r" (q1),
+        [p2] "r" (p2), [q2] "r" (q2), [p3] "r" (p3), [q3] "r" (q3),
+        [u32Four] "r" (u32Four)
+  );
+
+  *op2 = res_op2;
+  *op1 = res_op1;
+  *op0 = res_op0;
+  *oq0 = res_oq0;
+  *oq1 = res_oq1;
+  *oq2 = res_oq2;
+}
+
+static INLINE void mbfilter1_dspr2(uint32_t p3, uint32_t p2,
+                                   uint32_t p1, uint32_t p0,
+                                   uint32_t q0, uint32_t q1,
+                                   uint32_t q2, uint32_t q3,
+                                   uint32_t *op2_f1,
+                                   uint32_t *op1_f1, uint32_t *op0_f1,
+                                   uint32_t *oq0_f1, uint32_t *oq1_f1,
+                                   uint32_t *oq2_f1) {
+  /* use a 7 tap filter [1, 1, 1, 2, 1, 1, 1] for flat line */
+  uint32_t  res_op2, res_op1, res_op0;
+  uint32_t  res_oq0, res_oq1, res_oq2;
+  uint32_t  tmp;
+  uint32_t  add_p210_q012;
+  uint32_t  u32Four = 0x00040004;
+
+  /* *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + p2 + p2 + p1 + p0 + q0, 3)   1 */
+  /* *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + p1 + p1 + p0 + q0 + q1, 3)   2 */
+  /* *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + p0 + q0 + q1 + q2, 3)   3 */
+  /* *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + q0 + q0 + q1 + q2 + q3, 3)   4 */
+  /* *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q1 + q2 + q3 + q3, 3)   5 */
+  /* *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q2 + q3 + q3 + q3, 3)   6 */
+
+  __asm__ __volatile__ (
+      "addu.ph    %[add_p210_q012],  %[p2],             %[p1]             \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[p0]             \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q0]             \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q1]             \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q2]             \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[u32Four]        \n\t"
+
+      "shll.ph    %[tmp],            %[p3],             1                 \n\t"
+      "addu.ph    %[res_op2],        %[tmp],            %[p3]             \n\t"
+      "addu.ph    %[res_op1],        %[p3],             %[p3]             \n\t"
+      "addu.ph    %[res_op2],        %[res_op2],        %[p2]             \n\t"
+      "addu.ph    %[res_op1],        %[res_op1],        %[p1]             \n\t"
+      "addu.ph    %[res_op2],        %[res_op2],        %[add_p210_q012]  \n\t"
+      "addu.ph    %[res_op1],        %[res_op1],        %[add_p210_q012]  \n\t"
+      "subu.ph    %[res_op2],        %[res_op2],        %[q1]             \n\t"
+      "subu.ph    %[res_op1],        %[res_op1],        %[q2]             \n\t"
+      "subu.ph    %[res_op2],        %[res_op2],        %[q2]             \n\t"
+      "shrl.ph    %[res_op1],        %[res_op1],        3                 \n\t"
+      "shrl.ph    %[res_op2],        %[res_op2],        3                 \n\t"
+      "addu.ph    %[res_op0],        %[p3],             %[p0]             \n\t"
+      "addu.ph    %[res_oq0],        %[q0],             %[q3]             \n\t"
+      "addu.ph    %[res_op0],        %[res_op0],        %[add_p210_q012]  \n\t"
+      "addu.ph    %[res_oq0],        %[res_oq0],        %[add_p210_q012]  \n\t"
+      "addu.ph    %[res_oq1],        %[q3],             %[q3]             \n\t"
+      "shll.ph    %[tmp],            %[q3],             1                 \n\t"
+      "addu.ph    %[res_oq1],        %[res_oq1],        %[q1]             \n\t"
+      "addu.ph    %[res_oq2],        %[tmp],            %[q3]             \n\t"
+      "addu.ph    %[res_oq1],        %[res_oq1],        %[add_p210_q012]  \n\t"
+      "addu.ph    %[res_oq2],        %[res_oq2],        %[add_p210_q012]  \n\t"
+      "subu.ph    %[res_oq1],        %[res_oq1],        %[p2]             \n\t"
+      "addu.ph    %[res_oq2],        %[res_oq2],        %[q2]             \n\t"
+      "shrl.ph    %[res_oq1],        %[res_oq1],        3                 \n\t"
+      "subu.ph    %[res_oq2],        %[res_oq2],        %[p2]             \n\t"
+      "shrl.ph    %[res_oq0],        %[res_oq0],        3                 \n\t"
+      "subu.ph    %[res_oq2],        %[res_oq2],        %[p1]             \n\t"
+      "shrl.ph    %[res_op0],        %[res_op0],        3                 \n\t"
+      "shrl.ph    %[res_oq2],        %[res_oq2],        3                 \n\t"
+
+      : [add_p210_q012] "=&r" (add_p210_q012), [tmp] "=&r" (tmp),
+        [res_op2] "=&r" (res_op2), [res_op1] "=&r" (res_op1),
+        [res_op0] "=&r" (res_op0), [res_oq0] "=&r" (res_oq0),
+        [res_oq1] "=&r" (res_oq1), [res_oq2] "=&r" (res_oq2)
+      : [p0] "r" (p0), [q0] "r" (q0), [p1] "r" (p1), [q1] "r" (q1),
+        [p2] "r" (p2), [q2] "r" (q2), [p3] "r" (p3), [q3] "r" (q3),
+        [u32Four] "r" (u32Four)
+  );
+
+  *op2_f1 = res_op2;
+  *op1_f1 = res_op1;
+  *op0_f1 = res_op0;
+  *oq0_f1 = res_oq0;
+  *oq1_f1 = res_oq1;
+  *oq2_f1 = res_oq2;
+}
+
+static INLINE void wide_mbfilter_dspr2(uint32_t *op7, uint32_t *op6,
+                                       uint32_t *op5, uint32_t *op4,
+                                       uint32_t *op3, uint32_t *op2,
+                                       uint32_t *op1, uint32_t *op0,
+                                       uint32_t *oq0, uint32_t *oq1,
+                                       uint32_t *oq2, uint32_t *oq3,
+                                       uint32_t *oq4, uint32_t *oq5,
+                                       uint32_t *oq6, uint32_t *oq7) {
+  const uint32_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4;
+  const uint32_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+  const uint32_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+  const uint32_t q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7;
+  uint32_t       res_op6, res_op5, res_op4, res_op3, res_op2, res_op1, res_op0;
+  uint32_t       res_oq0, res_oq1, res_oq2, res_oq3, res_oq4, res_oq5, res_oq6;
+  uint32_t       tmp;
+  uint32_t       add_p6toq6;
+  uint32_t       u32Eight = 0x00080008;
+
+  __asm__ __volatile__ (
+      /* addition of p6,p5,p4,p3,p2,p1,p0,q0,q1,q2,q3,q4,q5,q6
+         which is used most of the time */
+      "addu.ph      %[add_p6toq6],     %[p6],              %[p5]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[p4]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[p3]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[p2]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[p1]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[p0]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q0]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q1]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q2]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q3]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q4]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q5]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q6]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[u32Eight]   \n\t"
+
+      : [add_p6toq6] "=&r" (add_p6toq6)
+      : [p6] "r" (p6), [p5] "r" (p5), [p4] "r" (p4),
+        [p3] "r" (p3), [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0),
+        [q0] "r" (q0), [q1] "r" (q1), [q2] "r" (q2), [q3] "r" (q3),
+        [q4] "r" (q4), [q5] "r" (q5), [q6] "r" (q6),
+        [u32Eight] "r" (u32Eight)
+  );
+
+  __asm__ __volatile__ (
+      /* *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 +
+                                   p3 + p2 + p1 + p0 + q0, 4) */
+      "shll.ph       %[tmp],            %[p7],            3               \n\t"
+      "subu.ph       %[res_op6],        %[tmp],           %[p7]           \n\t"
+      "addu.ph       %[res_op6],        %[res_op6],       %[p6]           \n\t"
+      "addu.ph       %[res_op6],        %[res_op6],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q1]           \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q2]           \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q3]           \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q4]           \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q5]           \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q6]           \n\t"
+      "shrl.ph       %[res_op6],        %[res_op6],       4               \n\t"
+
+      /* *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 +
+                                   p2 + p1 + p0 + q0 + q1, 4) */
+      "shll.ph       %[tmp],            %[p7],            2               \n\t"
+      "addu.ph       %[res_op5],        %[tmp],           %[p7]           \n\t"
+      "addu.ph       %[res_op5],        %[res_op5],       %[p7]           \n\t"
+      "addu.ph       %[res_op5],        %[res_op5],       %[p5]           \n\t"
+      "addu.ph       %[res_op5],        %[res_op5],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op5],        %[res_op5],       %[q2]           \n\t"
+      "subu.ph       %[res_op5],        %[res_op5],       %[q3]           \n\t"
+      "subu.ph       %[res_op5],        %[res_op5],       %[q4]           \n\t"
+      "subu.ph       %[res_op5],        %[res_op5],       %[q5]           \n\t"
+      "subu.ph       %[res_op5],        %[res_op5],       %[q6]           \n\t"
+      "shrl.ph       %[res_op5],        %[res_op5],       4               \n\t"
+
+      /* *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 +
+                                   p1 + p0 + q0 + q1 + q2, 4) */
+      "shll.ph       %[tmp],            %[p7],            2               \n\t"
+      "addu.ph       %[res_op4],        %[tmp],           %[p7]           \n\t"
+      "addu.ph       %[res_op4],        %[res_op4],       %[p4]           \n\t"
+      "addu.ph       %[res_op4],        %[res_op4],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op4],        %[res_op4],       %[q3]           \n\t"
+      "subu.ph       %[res_op4],        %[res_op4],       %[q4]           \n\t"
+      "subu.ph       %[res_op4],        %[res_op4],       %[q5]           \n\t"
+      "subu.ph       %[res_op4],        %[res_op4],       %[q6]           \n\t"
+      "shrl.ph       %[res_op4],        %[res_op4],       4               \n\t"
+
+      /* *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 +
+                                   p1 + p0 + q0 + q1 + q2 + q3, 4) */
+      "shll.ph       %[tmp],            %[p7],            2               \n\t"
+      "addu.ph       %[res_op3],        %[tmp],           %[p3]           \n\t"
+      "addu.ph       %[res_op3],        %[res_op3],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op3],        %[res_op3],       %[q4]           \n\t"
+      "subu.ph       %[res_op3],        %[res_op3],       %[q5]           \n\t"
+      "subu.ph       %[res_op3],        %[res_op3],       %[q6]           \n\t"
+      "shrl.ph       %[res_op3],        %[res_op3],       4               \n\t"
+
+      /* *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 +
+                                   p0 + q0 + q1 + q2 + q3 + q4, 4) */
+      "shll.ph       %[tmp],            %[p7],            1               \n\t"
+      "addu.ph       %[res_op2],        %[tmp],           %[p7]           \n\t"
+      "addu.ph       %[res_op2],        %[res_op2],       %[p2]           \n\t"
+      "addu.ph       %[res_op2],        %[res_op2],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op2],        %[res_op2],       %[q5]           \n\t"
+      "subu.ph       %[res_op2],        %[res_op2],       %[q6]           \n\t"
+      "shrl.ph       %[res_op2],        %[res_op2],       4               \n\t"
+
+      /* *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 +
+                                   p0 + q0 + q1 + q2 + q3 + q4 + q5, 4); */
+      "shll.ph       %[tmp],            %[p7],            1               \n\t"
+      "addu.ph       %[res_op1],        %[tmp],           %[p1]           \n\t"
+      "addu.ph       %[res_op1],        %[res_op1],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op1],        %[res_op1],       %[q6]           \n\t"
+      "shrl.ph       %[res_op1],        %[res_op1],       4               \n\t"
+
+      /* *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
+                                  q0 + q1 + q2 + q3 + q4 + q5 + q6, 4) */
+      "addu.ph       %[res_op0],        %[p7],            %[p0]           \n\t"
+      "addu.ph       %[res_op0],        %[res_op0],       %[add_p6toq6]   \n\t"
+      "shrl.ph       %[res_op0],        %[res_op0],       4               \n\t"
+
+      : [res_op6] "=&r" (res_op6), [res_op5] "=&r" (res_op5),
+        [res_op4] "=&r" (res_op4), [res_op3] "=&r" (res_op3),
+        [res_op2] "=&r" (res_op2), [res_op1] "=&r" (res_op1),
+        [res_op0] "=&r" (res_op0), [tmp] "=&r" (tmp)
+      : [p7] "r" (p7), [p6] "r" (p6), [p5] "r" (p5), [p4] "r" (p4),
+        [p3] "r" (p3), [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0),
+        [q2] "r" (q2), [q1] "r" (q1),
+        [q3] "r" (q3), [q4] "r" (q4), [q5] "r" (q5), [q6] "r" (q6),
+        [add_p6toq6] "r" (add_p6toq6)
+  );
+
+  *op6 = res_op6;
+  *op5 = res_op5;
+  *op4 = res_op4;
+  *op3 = res_op3;
+  *op2 = res_op2;
+  *op1 = res_op1;
+  *op0 = res_op0;
+
+  __asm__ __volatile__ (
+      /* *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 + q0 * 2 +
+                                   q1 + q2 + q3 + q4 + q5 + q6 + q7, 4); */
+      "addu.ph       %[res_oq0],        %[q7],            %[q0]           \n\t"
+      "addu.ph       %[res_oq0],        %[res_oq0],       %[add_p6toq6]   \n\t"
+      "shrl.ph       %[res_oq0],        %[res_oq0],       4               \n\t"
+
+      /* *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 + q0 + q1 * 2 +
+                                   q2 + q3 + q4 + q5 + q6 + q7 * 2, 4) */
+      "shll.ph       %[tmp],            %[q7],            1               \n\t"
+      "addu.ph       %[res_oq1],        %[tmp],           %[q1]           \n\t"
+      "addu.ph       %[res_oq1],        %[res_oq1],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq1],        %[res_oq1],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq1],        %[res_oq1],       4               \n\t"
+
+      /* *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 + q0 + q1 + q2 * 2 +
+                                   q3 + q4 + q5 + q6 + q7 * 3, 4) */
+      "shll.ph       %[tmp],            %[q7],            1               \n\t"
+      "addu.ph       %[res_oq2],        %[tmp],           %[q7]           \n\t"
+      "addu.ph       %[res_oq2],        %[res_oq2],       %[q2]           \n\t"
+      "addu.ph       %[res_oq2],        %[res_oq2],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq2],        %[res_oq2],       %[p5]           \n\t"
+      "subu.ph       %[res_oq2],        %[res_oq2],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq2],        %[res_oq2],       4               \n\t"
+
+      /* *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + q0 + q1 + q2 +
+                                   q3 * 2 + q4 + q5 + q6 + q7 * 4, 4) */
+      "shll.ph       %[tmp],            %[q7],            2               \n\t"
+      "addu.ph       %[res_oq3],        %[tmp],           %[q3]           \n\t"
+      "addu.ph       %[res_oq3],        %[res_oq3],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq3],        %[res_oq3],       %[p4]           \n\t"
+      "subu.ph       %[res_oq3],        %[res_oq3],       %[p5]           \n\t"
+      "subu.ph       %[res_oq3],        %[res_oq3],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq3],        %[res_oq3],       4               \n\t"
+
+      /* *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + q0 + q1 + q2 + q3 +
+                                   q4 * 2 + q5 + q6 + q7 * 5, 4) */
+      "shll.ph       %[tmp],            %[q7],            2               \n\t"
+      "addu.ph       %[res_oq4],        %[tmp],           %[q7]           \n\t"
+      "addu.ph       %[res_oq4],        %[res_oq4],       %[q4]           \n\t"
+      "addu.ph       %[res_oq4],        %[res_oq4],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq4],        %[res_oq4],       %[p3]           \n\t"
+      "subu.ph       %[res_oq4],        %[res_oq4],       %[p4]           \n\t"
+      "subu.ph       %[res_oq4],        %[res_oq4],       %[p5]           \n\t"
+      "subu.ph       %[res_oq4],        %[res_oq4],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq4],        %[res_oq4],       4               \n\t"
+
+      /* *oq5 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q2 + q3 + q4 +
+                                   q5 * 2 + q6 + q7 * 6, 4) */
+      "shll.ph       %[tmp],            %[q7],            2               \n\t"
+      "addu.ph       %[res_oq5],        %[tmp],           %[q7]           \n\t"
+      "addu.ph       %[res_oq5],        %[res_oq5],       %[q7]           \n\t"
+      "addu.ph       %[res_oq5],        %[res_oq5],       %[q5]           \n\t"
+      "addu.ph       %[res_oq5],        %[res_oq5],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq5],        %[res_oq5],       %[p2]           \n\t"
+      "subu.ph       %[res_oq5],        %[res_oq5],       %[p3]           \n\t"
+      "subu.ph       %[res_oq5],        %[res_oq5],       %[p4]           \n\t"
+      "subu.ph       %[res_oq5],        %[res_oq5],       %[p5]           \n\t"
+      "subu.ph       %[res_oq5],        %[res_oq5],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq5],        %[res_oq5],       4               \n\t"
+
+      /* *oq6 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q3 +
+                                   q4 + q5 + q6 * 2 + q7 * 7, 4) */
+      "shll.ph       %[tmp],            %[q7],            3               \n\t"
+      "subu.ph       %[res_oq6],        %[tmp],           %[q7]           \n\t"
+      "addu.ph       %[res_oq6],        %[res_oq6],       %[q6]           \n\t"
+      "addu.ph       %[res_oq6],        %[res_oq6],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p1]           \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p2]           \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p3]           \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p4]           \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p5]           \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq6],        %[res_oq6],       4               \n\t"
+
+      : [res_oq6] "=&r" (res_oq6), [res_oq5] "=&r" (res_oq5),
+        [res_oq4] "=&r" (res_oq4), [res_oq3] "=&r" (res_oq3),
+        [res_oq2] "=&r" (res_oq2), [res_oq1] "=&r" (res_oq1),
+        [res_oq0] "=&r" (res_oq0), [tmp] "=&r" (tmp)
+      : [q7] "r" (q7), [q6] "r" (q6), [q5] "r" (q5), [q4] "r" (q4),
+        [q3] "r" (q3), [q2] "r" (q2), [q1] "r" (q1), [q0] "r" (q0),
+        [p1] "r" (p1), [p2] "r" (p2),
+        [p3] "r" (p3), [p4] "r" (p4), [p5] "r" (p5), [p6] "r" (p6),
+        [add_p6toq6] "r" (add_p6toq6)
+  );
+
+  *oq0 = res_oq0;
+  *oq1 = res_oq1;
+  *oq2 = res_oq2;
+  *oq3 = res_oq3;
+  *oq4 = res_oq4;
+  *oq5 = res_oq5;
+  *oq6 = res_oq6;
+}
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_MIPS_DSPR2_VP9_LOOPFILTER_FILTERS_DSPR2_H_
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_macros_dspr2.h b/libvpx/libvpx/vpx_dsp/mips/loopfilter_macros_dspr2.h
new file mode 100644
index 0000000..994ff18
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_macros_dspr2.h
@@ -0,0 +1,478 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_MIPS_DSPR2_VP9_LOOPFILTER_MACROS_DSPR2_H_
+#define VP9_COMMON_MIPS_DSPR2_VP9_LOOPFILTER_MACROS_DSPR2_H_
+
+#include <stdlib.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if HAVE_DSPR2
+#define STORE_F0() {                                                    \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q1_f0],    1(%[s4])           \n\t"                   \
+        "sb     %[q0_f0],    0(%[s4])           \n\t"                   \
+        "sb     %[p0_f0],   -1(%[s4])           \n\t"                   \
+        "sb     %[p1_f0],   -2(%[s4])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q1_f0] "r" (q1_f0), [q0_f0] "r" (q0_f0),                     \
+          [p0_f0] "r" (p0_f0), [p1_f0] "r" (p1_f0),                     \
+          [s4] "r" (s4)                                                 \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "srl    %[q1_f0],   %[q1_f0],   8       \n\t"                   \
+        "srl    %[q0_f0],   %[q0_f0],   8       \n\t"                   \
+        "srl    %[p0_f0],   %[p0_f0],   8       \n\t"                   \
+        "srl    %[p1_f0],   %[p1_f0],   8       \n\t"                   \
+                                                                        \
+        : [q1_f0] "+r" (q1_f0), [q0_f0] "+r" (q0_f0),                   \
+          [p0_f0] "+r" (p0_f0), [p1_f0] "+r" (p1_f0)                    \
+        :                                                               \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q1_f0],    1(%[s3])           \n\t"                   \
+        "sb     %[q0_f0],    0(%[s3])           \n\t"                   \
+        "sb     %[p0_f0],   -1(%[s3])           \n\t"                   \
+        "sb     %[p1_f0],   -2(%[s3])           \n\t"                   \
+                                                                        \
+        : [p1_f0] "+r" (p1_f0)                                          \
+        : [q1_f0] "r" (q1_f0), [q0_f0] "r" (q0_f0),                     \
+          [s3] "r" (s3), [p0_f0] "r" (p0_f0)                            \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "srl    %[q1_f0],   %[q1_f0],   8       \n\t"                   \
+        "srl    %[q0_f0],   %[q0_f0],   8       \n\t"                   \
+        "srl    %[p0_f0],   %[p0_f0],   8       \n\t"                   \
+        "srl    %[p1_f0],   %[p1_f0],   8       \n\t"                   \
+                                                                        \
+        : [q1_f0] "+r" (q1_f0), [q0_f0] "+r" (q0_f0),                   \
+          [p0_f0] "+r" (p0_f0), [p1_f0] "+r" (p1_f0)                    \
+        :                                                               \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q1_f0],    1(%[s2])           \n\t"                   \
+        "sb     %[q0_f0],    0(%[s2])           \n\t"                   \
+        "sb     %[p0_f0],   -1(%[s2])           \n\t"                   \
+        "sb     %[p1_f0],   -2(%[s2])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q1_f0] "r" (q1_f0), [q0_f0] "r" (q0_f0),                     \
+          [p0_f0] "r" (p0_f0), [p1_f0] "r" (p1_f0),                     \
+          [s2] "r" (s2)                                                 \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "srl    %[q1_f0],   %[q1_f0],   8       \n\t"                   \
+        "srl    %[q0_f0],   %[q0_f0],   8       \n\t"                   \
+        "srl    %[p0_f0],   %[p0_f0],   8       \n\t"                   \
+        "srl    %[p1_f0],   %[p1_f0],   8       \n\t"                   \
+                                                                        \
+        : [q1_f0] "+r" (q1_f0), [q0_f0] "+r" (q0_f0),                   \
+          [p0_f0] "+r" (p0_f0), [p1_f0] "+r" (p1_f0)                    \
+        :                                                               \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q1_f0],    1(%[s1])           \n\t"                   \
+        "sb     %[q0_f0],    0(%[s1])           \n\t"                   \
+        "sb     %[p0_f0],   -1(%[s1])           \n\t"                   \
+        "sb     %[p1_f0],   -2(%[s1])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q1_f0] "r" (q1_f0), [q0_f0] "r" (q0_f0),                     \
+          [p0_f0] "r" (p0_f0), [p1_f0] "r" (p1_f0),                     \
+          [s1] "r" (s1)                                                 \
+    );                                                                  \
+}
+
+#define STORE_F1() {                                                    \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q2_r],     2(%[s4])           \n\t"                   \
+        "sb     %[q1_r],     1(%[s4])           \n\t"                   \
+        "sb     %[q0_r],     0(%[s4])           \n\t"                   \
+        "sb     %[p0_r],    -1(%[s4])           \n\t"                   \
+        "sb     %[p1_r],    -2(%[s4])           \n\t"                   \
+        "sb     %[p2_r],    -3(%[s4])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q2_r] "r" (q2_r), [q1_r] "r" (q1_r), [q0_r] "r" (q0_r),      \
+          [p0_r] "r" (p0_r), [p1_r] "r" (p1_r), [p2_r] "r" (p2_r),      \
+          [s4] "r" (s4)                                                 \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "srl    %[q2_r],    %[q2_r],    16      \n\t"                   \
+        "srl    %[q1_r],    %[q1_r],    16      \n\t"                   \
+        "srl    %[q0_r],    %[q0_r],    16      \n\t"                   \
+        "srl    %[p0_r],    %[p0_r],    16      \n\t"                   \
+        "srl    %[p1_r],    %[p1_r],    16      \n\t"                   \
+        "srl    %[p2_r],    %[p2_r],    16      \n\t"                   \
+                                                                        \
+        : [q2_r] "+r" (q2_r), [q1_r] "+r" (q1_r), [q0_r] "+r" (q0_r),   \
+          [p0_r] "+r" (p0_r), [p1_r] "+r" (p1_r), [p2_r] "+r" (p2_r)    \
+        :                                                               \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q2_r],     2(%[s3])           \n\t"                   \
+        "sb     %[q1_r],     1(%[s3])           \n\t"                   \
+        "sb     %[q0_r],     0(%[s3])           \n\t"                   \
+        "sb     %[p0_r],    -1(%[s3])           \n\t"                   \
+        "sb     %[p1_r],    -2(%[s3])           \n\t"                   \
+        "sb     %[p2_r],    -3(%[s3])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q2_r] "r" (q2_r), [q1_r] "r" (q1_r), [q0_r] "r" (q0_r),      \
+          [p0_r] "r" (p0_r), [p1_r] "r" (p1_r), [p2_r] "r" (p2_r),      \
+          [s3] "r" (s3)                                                 \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q2_l],     2(%[s2])           \n\t"                   \
+        "sb     %[q1_l],     1(%[s2])           \n\t"                   \
+        "sb     %[q0_l],     0(%[s2])           \n\t"                   \
+        "sb     %[p0_l],    -1(%[s2])           \n\t"                   \
+        "sb     %[p1_l],    -2(%[s2])           \n\t"                   \
+        "sb     %[p2_l],    -3(%[s2])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q2_l] "r" (q2_l), [q1_l] "r" (q1_l), [q0_l] "r" (q0_l),      \
+          [p0_l] "r" (p0_l), [p1_l] "r" (p1_l), [p2_l] "r" (p2_l),      \
+          [s2] "r" (s2)                                                 \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "srl    %[q2_l],    %[q2_l],    16      \n\t"                   \
+        "srl    %[q1_l],    %[q1_l],    16      \n\t"                   \
+        "srl    %[q0_l],    %[q0_l],    16      \n\t"                   \
+        "srl    %[p0_l],    %[p0_l],    16      \n\t"                   \
+        "srl    %[p1_l],    %[p1_l],    16      \n\t"                   \
+        "srl    %[p2_l],    %[p2_l],    16      \n\t"                   \
+                                                                        \
+        : [q2_l] "+r" (q2_l), [q1_l] "+r" (q1_l), [q0_l] "+r" (q0_l),   \
+          [p0_l] "+r" (p0_l), [p1_l] "+r" (p1_l), [p2_l] "+r" (p2_l)    \
+        :                                                               \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q2_l],     2(%[s1])           \n\t"                   \
+        "sb     %[q1_l],     1(%[s1])           \n\t"                   \
+        "sb     %[q0_l],     0(%[s1])           \n\t"                   \
+        "sb     %[p0_l],    -1(%[s1])           \n\t"                   \
+        "sb     %[p1_l],    -2(%[s1])           \n\t"                   \
+        "sb     %[p2_l],    -3(%[s1])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q2_l] "r" (q2_l), [q1_l] "r" (q1_l), [q0_l] "r" (q0_l),      \
+          [p0_l] "r" (p0_l), [p1_l] "r" (p1_l), [p2_l] "r" (p2_l),      \
+          [s1] "r" (s1)                                                 \
+    );                                                                  \
+}
+
+#define STORE_F2() {                                                    \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q6_r],     6(%[s4])           \n\t"                   \
+        "sb     %[q5_r],     5(%[s4])           \n\t"                   \
+        "sb     %[q4_r],     4(%[s4])           \n\t"                   \
+        "sb     %[q3_r],     3(%[s4])           \n\t"                   \
+        "sb     %[q2_r],     2(%[s4])           \n\t"                   \
+        "sb     %[q1_r],     1(%[s4])           \n\t"                   \
+        "sb     %[q0_r],     0(%[s4])           \n\t"                   \
+        "sb     %[p0_r],    -1(%[s4])           \n\t"                   \
+        "sb     %[p1_r],    -2(%[s4])           \n\t"                   \
+        "sb     %[p2_r],    -3(%[s4])           \n\t"                   \
+        "sb     %[p3_r],    -4(%[s4])           \n\t"                   \
+        "sb     %[p4_r],    -5(%[s4])           \n\t"                   \
+        "sb     %[p5_r],    -6(%[s4])           \n\t"                   \
+        "sb     %[p6_r],    -7(%[s4])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q6_r] "r" (q6_r), [q5_r] "r" (q5_r), [q4_r] "r" (q4_r),      \
+          [q3_r] "r" (q3_r), [q2_r] "r" (q2_r), [q1_r] "r" (q1_r),      \
+          [q0_r] "r" (q0_r),                                            \
+          [p0_r] "r" (p0_r), [p1_r] "r" (p1_r), [p2_r] "r" (p2_r),      \
+          [p3_r] "r" (p3_r), [p4_r] "r" (p4_r), [p5_r] "r" (p5_r),      \
+          [p6_r] "r" (p6_r),                                            \
+          [s4] "r" (s4)                                                 \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "srl    %[q6_r],    %[q6_r],    16      \n\t"                   \
+        "srl    %[q5_r],    %[q5_r],    16      \n\t"                   \
+        "srl    %[q4_r],    %[q4_r],    16      \n\t"                   \
+        "srl    %[q3_r],    %[q3_r],    16      \n\t"                   \
+        "srl    %[q2_r],    %[q2_r],    16      \n\t"                   \
+        "srl    %[q1_r],    %[q1_r],    16      \n\t"                   \
+        "srl    %[q0_r],    %[q0_r],    16      \n\t"                   \
+        "srl    %[p0_r],    %[p0_r],    16      \n\t"                   \
+        "srl    %[p1_r],    %[p1_r],    16      \n\t"                   \
+        "srl    %[p2_r],    %[p2_r],    16      \n\t"                   \
+        "srl    %[p3_r],    %[p3_r],    16      \n\t"                   \
+        "srl    %[p4_r],    %[p4_r],    16      \n\t"                   \
+        "srl    %[p5_r],    %[p5_r],    16      \n\t"                   \
+        "srl    %[p6_r],    %[p6_r],    16      \n\t"                   \
+                                                                        \
+        : [q6_r] "+r" (q6_r), [q5_r] "+r" (q5_r), [q4_r] "+r" (q4_r),   \
+          [q3_r] "+r" (q3_r), [q2_r] "+r" (q2_r), [q1_r] "+r" (q1_r),   \
+          [q0_r] "+r" (q0_r),                                           \
+          [p0_r] "+r" (p0_r), [p1_r] "+r" (p1_r), [p2_r] "+r" (p2_r),   \
+          [p3_r] "+r" (p3_r), [p4_r] "+r" (p4_r), [p5_r] "+r" (p5_r),   \
+          [p6_r] "+r" (p6_r)                                            \
+        :                                                               \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q6_r],     6(%[s3])           \n\t"                   \
+        "sb     %[q5_r],     5(%[s3])           \n\t"                   \
+        "sb     %[q4_r],     4(%[s3])           \n\t"                   \
+        "sb     %[q3_r],     3(%[s3])           \n\t"                   \
+        "sb     %[q2_r],     2(%[s3])           \n\t"                   \
+        "sb     %[q1_r],     1(%[s3])           \n\t"                   \
+        "sb     %[q0_r],     0(%[s3])           \n\t"                   \
+        "sb     %[p0_r],    -1(%[s3])           \n\t"                   \
+        "sb     %[p1_r],    -2(%[s3])           \n\t"                   \
+        "sb     %[p2_r],    -3(%[s3])           \n\t"                   \
+        "sb     %[p3_r],    -4(%[s3])           \n\t"                   \
+        "sb     %[p4_r],    -5(%[s3])           \n\t"                   \
+        "sb     %[p5_r],    -6(%[s3])           \n\t"                   \
+        "sb     %[p6_r],    -7(%[s3])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q6_r] "r" (q6_r), [q5_r] "r" (q5_r), [q4_r] "r" (q4_r),      \
+          [q3_r] "r" (q3_r), [q2_r] "r" (q2_r), [q1_r] "r" (q1_r),      \
+          [q0_r] "r" (q0_r),                                            \
+          [p0_r] "r" (p0_r), [p1_r] "r" (p1_r), [p2_r] "r" (p2_r),      \
+          [p3_r] "r" (p3_r), [p4_r] "r" (p4_r), [p5_r] "r" (p5_r),      \
+          [p6_r] "r" (p6_r),                                            \
+          [s3] "r" (s3)                                                 \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q6_l],     6(%[s2])           \n\t"                   \
+        "sb     %[q5_l],     5(%[s2])           \n\t"                   \
+        "sb     %[q4_l],     4(%[s2])           \n\t"                   \
+        "sb     %[q3_l],     3(%[s2])           \n\t"                   \
+        "sb     %[q2_l],     2(%[s2])           \n\t"                   \
+        "sb     %[q1_l],     1(%[s2])           \n\t"                   \
+        "sb     %[q0_l],     0(%[s2])           \n\t"                   \
+        "sb     %[p0_l],    -1(%[s2])           \n\t"                   \
+        "sb     %[p1_l],    -2(%[s2])           \n\t"                   \
+        "sb     %[p2_l],    -3(%[s2])           \n\t"                   \
+        "sb     %[p3_l],    -4(%[s2])           \n\t"                   \
+        "sb     %[p4_l],    -5(%[s2])           \n\t"                   \
+        "sb     %[p5_l],    -6(%[s2])           \n\t"                   \
+        "sb     %[p6_l],    -7(%[s2])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q6_l] "r" (q6_l), [q5_l] "r" (q5_l), [q4_l] "r" (q4_l),      \
+          [q3_l] "r" (q3_l), [q2_l] "r" (q2_l), [q1_l] "r" (q1_l),      \
+          [q0_l] "r" (q0_l),                                            \
+          [p0_l] "r" (p0_l), [p1_l] "r" (p1_l), [p2_l] "r" (p2_l),      \
+          [p3_l] "r" (p3_l), [p4_l] "r" (p4_l), [p5_l] "r" (p5_l),      \
+          [p6_l] "r" (p6_l),                                            \
+          [s2] "r" (s2)                                                 \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "srl    %[q6_l],    %[q6_l],    16     \n\t"                    \
+        "srl    %[q5_l],    %[q5_l],    16     \n\t"                    \
+        "srl    %[q4_l],    %[q4_l],    16     \n\t"                    \
+        "srl    %[q3_l],    %[q3_l],    16     \n\t"                    \
+        "srl    %[q2_l],    %[q2_l],    16     \n\t"                    \
+        "srl    %[q1_l],    %[q1_l],    16     \n\t"                    \
+        "srl    %[q0_l],    %[q0_l],    16     \n\t"                    \
+        "srl    %[p0_l],    %[p0_l],    16     \n\t"                    \
+        "srl    %[p1_l],    %[p1_l],    16     \n\t"                    \
+        "srl    %[p2_l],    %[p2_l],    16     \n\t"                    \
+        "srl    %[p3_l],    %[p3_l],    16     \n\t"                    \
+        "srl    %[p4_l],    %[p4_l],    16     \n\t"                    \
+        "srl    %[p5_l],    %[p5_l],    16     \n\t"                    \
+        "srl    %[p6_l],    %[p6_l],    16     \n\t"                    \
+                                                                        \
+        : [q6_l] "+r" (q6_l), [q5_l] "+r" (q5_l), [q4_l] "+r" (q4_l),   \
+          [q3_l] "+r" (q3_l), [q2_l] "+r" (q2_l), [q1_l] "+r" (q1_l),   \
+          [q0_l] "+r" (q0_l),                                           \
+          [p0_l] "+r" (p0_l), [p1_l] "+r" (p1_l), [p2_l] "+r" (p2_l),   \
+          [p3_l] "+r" (p3_l), [p4_l] "+r" (p4_l), [p5_l] "+r" (p5_l),   \
+          [p6_l] "+r" (p6_l)                                            \
+        :                                                               \
+    );                                                                  \
+                                                                        \
+    __asm__ __volatile__ (                                              \
+        "sb     %[q6_l],     6(%[s1])           \n\t"                   \
+        "sb     %[q5_l],     5(%[s1])           \n\t"                   \
+        "sb     %[q4_l],     4(%[s1])           \n\t"                   \
+        "sb     %[q3_l],     3(%[s1])           \n\t"                   \
+        "sb     %[q2_l],     2(%[s1])           \n\t"                   \
+        "sb     %[q1_l],     1(%[s1])           \n\t"                   \
+        "sb     %[q0_l],     0(%[s1])           \n\t"                   \
+        "sb     %[p0_l],    -1(%[s1])           \n\t"                   \
+        "sb     %[p1_l],    -2(%[s1])           \n\t"                   \
+        "sb     %[p2_l],    -3(%[s1])           \n\t"                   \
+        "sb     %[p3_l],    -4(%[s1])           \n\t"                   \
+        "sb     %[p4_l],    -5(%[s1])           \n\t"                   \
+        "sb     %[p5_l],    -6(%[s1])           \n\t"                   \
+        "sb     %[p6_l],    -7(%[s1])           \n\t"                   \
+                                                                        \
+        :                                                               \
+        : [q6_l] "r" (q6_l), [q5_l] "r" (q5_l), [q4_l] "r" (q4_l),      \
+          [q3_l] "r" (q3_l), [q2_l] "r" (q2_l), [q1_l] "r" (q1_l),      \
+          [q0_l] "r" (q0_l),                                            \
+          [p0_l] "r" (p0_l), [p1_l] "r" (p1_l), [p2_l] "r" (p2_l),      \
+          [p3_l] "r" (p3_l), [p4_l] "r" (p4_l), [p5_l] "r" (p5_l),      \
+          [p6_l] "r" (p6_l),                                            \
+          [s1] "r" (s1)                                                 \
+    );                                                                  \
+}
+
+#define PACK_LEFT_0TO3() {                                              \
+    __asm__ __volatile__ (                                              \
+        "preceu.ph.qbl   %[p3_l],   %[p3]   \n\t"                       \
+        "preceu.ph.qbl   %[p2_l],   %[p2]   \n\t"                       \
+        "preceu.ph.qbl   %[p1_l],   %[p1]   \n\t"                       \
+        "preceu.ph.qbl   %[p0_l],   %[p0]   \n\t"                       \
+        "preceu.ph.qbl   %[q0_l],   %[q0]   \n\t"                       \
+        "preceu.ph.qbl   %[q1_l],   %[q1]   \n\t"                       \
+        "preceu.ph.qbl   %[q2_l],   %[q2]   \n\t"                       \
+        "preceu.ph.qbl   %[q3_l],   %[q3]   \n\t"                       \
+                                                                        \
+        : [p3_l] "=&r" (p3_l), [p2_l] "=&r" (p2_l),                     \
+          [p1_l] "=&r" (p1_l), [p0_l] "=&r" (p0_l),                     \
+          [q0_l] "=&r" (q0_l), [q1_l] "=&r" (q1_l),                     \
+          [q2_l] "=&r" (q2_l), [q3_l] "=&r" (q3_l)                      \
+        : [p3] "r" (p3), [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0),   \
+          [q0] "r" (q0), [q1] "r" (q1), [q2] "r" (q2), [q3] "r" (q3)    \
+    );                                                                  \
+}
+
+#define PACK_LEFT_4TO7() {                                              \
+    __asm__ __volatile__ (                                              \
+        "preceu.ph.qbl   %[p7_l],   %[p7]   \n\t"                       \
+        "preceu.ph.qbl   %[p6_l],   %[p6]   \n\t"                       \
+        "preceu.ph.qbl   %[p5_l],   %[p5]   \n\t"                       \
+        "preceu.ph.qbl   %[p4_l],   %[p4]   \n\t"                       \
+        "preceu.ph.qbl   %[q4_l],   %[q4]   \n\t"                       \
+        "preceu.ph.qbl   %[q5_l],   %[q5]   \n\t"                       \
+        "preceu.ph.qbl   %[q6_l],   %[q6]   \n\t"                       \
+        "preceu.ph.qbl   %[q7_l],   %[q7]   \n\t"                       \
+                                                                        \
+        : [p7_l] "=&r" (p7_l), [p6_l] "=&r" (p6_l),                     \
+          [p5_l] "=&r" (p5_l), [p4_l] "=&r" (p4_l),                     \
+          [q4_l] "=&r" (q4_l), [q5_l] "=&r" (q5_l),                     \
+          [q6_l] "=&r" (q6_l), [q7_l] "=&r" (q7_l)                      \
+        : [p7] "r" (p7), [p6] "r" (p6), [p5] "r" (p5), [p4] "r" (p4),   \
+          [q4] "r" (q4), [q5] "r" (q5), [q6] "r" (q6), [q7] "r" (q7)    \
+    );                                                                  \
+}
+
+#define PACK_RIGHT_0TO3() {                                             \
+    __asm__ __volatile__ (                                              \
+        "preceu.ph.qbr   %[p3_r],   %[p3]  \n\t"                        \
+        "preceu.ph.qbr   %[p2_r],   %[p2]   \n\t"                       \
+        "preceu.ph.qbr   %[p1_r],   %[p1]   \n\t"                       \
+        "preceu.ph.qbr   %[p0_r],   %[p0]   \n\t"                       \
+        "preceu.ph.qbr   %[q0_r],   %[q0]   \n\t"                       \
+        "preceu.ph.qbr   %[q1_r],   %[q1]   \n\t"                       \
+        "preceu.ph.qbr   %[q2_r],   %[q2]   \n\t"                       \
+        "preceu.ph.qbr   %[q3_r],   %[q3]   \n\t"                       \
+                                                                        \
+        : [p3_r] "=&r" (p3_r), [p2_r] "=&r" (p2_r),                     \
+          [p1_r] "=&r" (p1_r), [p0_r] "=&r" (p0_r),                     \
+          [q0_r] "=&r" (q0_r), [q1_r] "=&r" (q1_r),                     \
+          [q2_r] "=&r" (q2_r), [q3_r] "=&r" (q3_r)                      \
+        : [p3] "r" (p3), [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0),   \
+          [q0] "r" (q0), [q1] "r" (q1), [q2] "r" (q2), [q3] "r" (q3)    \
+    );                                                                  \
+}
+
+#define PACK_RIGHT_4TO7() {                                             \
+    __asm__ __volatile__ (                                              \
+        "preceu.ph.qbr   %[p7_r],   %[p7]   \n\t"                       \
+        "preceu.ph.qbr   %[p6_r],   %[p6]   \n\t"                       \
+        "preceu.ph.qbr   %[p5_r],   %[p5]   \n\t"                       \
+        "preceu.ph.qbr   %[p4_r],   %[p4]   \n\t"                       \
+        "preceu.ph.qbr   %[q4_r],   %[q4]   \n\t"                       \
+        "preceu.ph.qbr   %[q5_r],   %[q5]   \n\t"                       \
+        "preceu.ph.qbr   %[q6_r],   %[q6]   \n\t"                       \
+        "preceu.ph.qbr   %[q7_r],   %[q7]   \n\t"                       \
+                                                                        \
+        : [p7_r] "=&r" (p7_r), [p6_r] "=&r" (p6_r),                     \
+          [p5_r] "=&r" (p5_r), [p4_r] "=&r" (p4_r),                     \
+          [q4_r] "=&r" (q4_r), [q5_r] "=&r" (q5_r),                     \
+          [q6_r] "=&r" (q6_r), [q7_r] "=&r" (q7_r)                      \
+        : [p7] "r" (p7), [p6] "r" (p6), [p5] "r" (p5), [p4] "r" (p4),   \
+          [q4] "r" (q4), [q5] "r" (q5), [q6] "r" (q6), [q7] "r" (q7)    \
+    );                                                                  \
+}
+
+#define COMBINE_LEFT_RIGHT_0TO2() {                                     \
+    __asm__ __volatile__ (                                              \
+        "precr.qb.ph    %[p2],  %[p2_l],    %[p2_r]    \n\t"            \
+        "precr.qb.ph    %[p1],  %[p1_l],    %[p1_r]    \n\t"            \
+        "precr.qb.ph    %[p0],  %[p0_l],    %[p0_r]    \n\t"            \
+        "precr.qb.ph    %[q0],  %[q0_l],    %[q0_r]    \n\t"            \
+        "precr.qb.ph    %[q1],  %[q1_l],    %[q1_r]    \n\t"            \
+        "precr.qb.ph    %[q2],  %[q2_l],    %[q2_r]    \n\t"            \
+                                                                        \
+        : [p2] "=&r" (p2), [p1] "=&r" (p1), [p0] "=&r" (p0),            \
+          [q0] "=&r" (q0), [q1] "=&r" (q1), [q2] "=&r" (q2)             \
+        : [p2_l] "r" (p2_l), [p2_r] "r" (p2_r),                         \
+          [p1_l] "r" (p1_l), [p1_r] "r" (p1_r),                         \
+          [p0_l] "r" (p0_l), [p0_r] "r" (p0_r),                         \
+          [q0_l] "r" (q0_l), [q0_r] "r" (q0_r),                         \
+          [q1_l] "r" (q1_l), [q1_r] "r" (q1_r),                         \
+          [q2_l] "r" (q2_l), [q2_r] "r" (q2_r)                          \
+    );                                                                  \
+}
+
+#define COMBINE_LEFT_RIGHT_3TO6() {                                     \
+    __asm__ __volatile__ (                                              \
+        "precr.qb.ph    %[p6],  %[p6_l],    %[p6_r]    \n\t"            \
+        "precr.qb.ph    %[p5],  %[p5_l],    %[p5_r]    \n\t"            \
+        "precr.qb.ph    %[p4],  %[p4_l],    %[p4_r]    \n\t"            \
+        "precr.qb.ph    %[p3],  %[p3_l],    %[p3_r]    \n\t"            \
+        "precr.qb.ph    %[q3],  %[q3_l],    %[q3_r]    \n\t"            \
+        "precr.qb.ph    %[q4],  %[q4_l],    %[q4_r]    \n\t"            \
+        "precr.qb.ph    %[q5],  %[q5_l],    %[q5_r]    \n\t"            \
+        "precr.qb.ph    %[q6],  %[q6_l],    %[q6_r]    \n\t"            \
+                                                                        \
+        : [p6] "=&r" (p6),[p5] "=&r" (p5),                              \
+          [p4] "=&r" (p4),[p3] "=&r" (p3),                              \
+          [q3] "=&r" (q3),[q4] "=&r" (q4),                              \
+          [q5] "=&r" (q5),[q6] "=&r" (q6)                               \
+        : [p6_l] "r" (p6_l), [p5_l] "r" (p5_l),                         \
+          [p4_l] "r" (p4_l), [p3_l] "r" (p3_l),                         \
+          [p6_r] "r" (p6_r), [p5_r] "r" (p5_r),                         \
+          [p4_r] "r" (p4_r), [p3_r] "r" (p3_r),                         \
+          [q3_l] "r" (q3_l), [q4_l] "r" (q4_l),                         \
+          [q5_l] "r" (q5_l), [q6_l] "r" (q6_l),                         \
+          [q3_r] "r" (q3_r), [q4_r] "r" (q4_r),                         \
+          [q5_r] "r" (q5_r), [q6_r] "r" (q6_r)                          \
+    );                                                                  \
+}
+
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_MIPS_DSPR2_VP9_LOOPFILTER_MACROS_DSPR2_H_
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_masks_dspr2.h b/libvpx/libvpx/vpx_dsp/mips/loopfilter_masks_dspr2.h
new file mode 100644
index 0000000..2c964af
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_masks_dspr2.h
@@ -0,0 +1,373 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_MIPS_DSPR2_VP9_LOOPFILTER_MASKS_DSPR2_H_
+#define VP9_COMMON_MIPS_DSPR2_VP9_LOOPFILTER_MASKS_DSPR2_H_
+
+#include <stdlib.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if HAVE_DSPR2
+/* processing 4 pixels at the same time
+ * compute hev and mask in the same function */
+static INLINE void filter_hev_mask_dspr2(uint32_t limit, uint32_t flimit,
+                                         uint32_t p1, uint32_t p0,
+                                         uint32_t p3, uint32_t p2,
+                                         uint32_t q0, uint32_t q1,
+                                         uint32_t q2, uint32_t q3,
+                                         uint32_t thresh, uint32_t *hev,
+                                         uint32_t *mask) {
+  uint32_t  c, r, r3, r_k;
+  uint32_t  s1, s2, s3;
+  uint32_t  ones = 0xFFFFFFFF;
+  uint32_t  hev1;
+
+  __asm__ __volatile__ (
+      /* mask |= (abs(p3 - p2) > limit) */
+      "subu_s.qb      %[c],   %[p3],     %[p2]        \n\t"
+      "subu_s.qb      %[r_k], %[p2],     %[p3]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   $0,        %[c]         \n\t"
+
+      /* mask |= (abs(p2 - p1) > limit) */
+      "subu_s.qb      %[c],   %[p2],     %[p1]        \n\t"
+      "subu_s.qb      %[r_k], %[p1],     %[p2]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   %[r],      %[c]         \n\t"
+
+      /* mask |= (abs(p1 - p0) > limit)
+       * hev  |= (abs(p1 - p0) > thresh)
+       */
+      "subu_s.qb      %[c],   %[p1],     %[p0]        \n\t"
+      "subu_s.qb      %[r_k], %[p0],     %[p1]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[thresh], %[r_k]       \n\t"
+      "or             %[r3],  $0,        %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   %[r],      %[c]         \n\t"
+
+      /* mask |= (abs(q1 - q0) > limit)
+       * hev  |= (abs(q1 - q0) > thresh)
+       */
+      "subu_s.qb      %[c],   %[q1],     %[q0]        \n\t"
+      "subu_s.qb      %[r_k], %[q0],     %[q1]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[thresh], %[r_k]       \n\t"
+      "or             %[r3],  %[r3],     %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   %[r],      %[c]         \n\t"
+
+      /* mask |= (abs(q2 - q1) > limit) */
+      "subu_s.qb      %[c],   %[q2],     %[q1]        \n\t"
+      "subu_s.qb      %[r_k], %[q1],     %[q2]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   %[r],      %[c]         \n\t"
+      "sll            %[r3],    %[r3],    24          \n\t"
+
+      /* mask |= (abs(q3 - q2) > limit) */
+      "subu_s.qb      %[c],   %[q3],     %[q2]        \n\t"
+      "subu_s.qb      %[r_k], %[q2],     %[q3]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   %[r],      %[c]         \n\t"
+
+      : [c] "=&r" (c), [r_k] "=&r" (r_k),
+        [r] "=&r" (r), [r3] "=&r" (r3)
+      : [limit] "r" (limit), [p3] "r" (p3), [p2] "r" (p2),
+        [p1] "r" (p1), [p0] "r" (p0), [q1] "r" (q1), [q0] "r" (q0),
+        [q2] "r" (q2), [q3] "r" (q3), [thresh] "r" (thresh)
+  );
+
+  __asm__ __volatile__ (
+      /* abs(p0 - q0) */
+      "subu_s.qb      %[c],   %[p0],     %[q0]        \n\t"
+      "subu_s.qb      %[r_k], %[q0],     %[p0]        \n\t"
+      "wrdsp          %[r3]                           \n\t"
+      "or             %[s1],  %[r_k],    %[c]         \n\t"
+
+      /* abs(p1 - q1) */
+      "subu_s.qb      %[c],    %[p1],    %[q1]        \n\t"
+      "addu_s.qb      %[s3],   %[s1],    %[s1]        \n\t"
+      "pick.qb        %[hev1], %[ones],  $0           \n\t"
+      "subu_s.qb      %[r_k],  %[q1],    %[p1]        \n\t"
+      "or             %[s2],   %[r_k],   %[c]         \n\t"
+
+      /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > flimit * 2 + limit */
+      "shrl.qb        %[s2],   %[s2],     1           \n\t"
+      "addu_s.qb      %[s1],   %[s2],     %[s3]       \n\t"
+      "cmpgu.lt.qb    %[c],    %[flimit], %[s1]       \n\t"
+      "or             %[r],    %[r],      %[c]        \n\t"
+      "sll            %[r],    %[r],      24          \n\t"
+
+      "wrdsp          %[r]                            \n\t"
+      "pick.qb        %[s2],  $0,         %[ones]     \n\t"
+
+      : [c] "=&r" (c), [r_k] "=&r" (r_k), [s1] "=&r" (s1), [hev1] "=&r" (hev1),
+        [s2] "=&r" (s2), [r] "+r" (r), [s3] "=&r" (s3)
+      : [p0] "r" (p0), [q0] "r" (q0), [p1] "r" (p1), [r3] "r" (r3),
+        [q1] "r" (q1), [ones] "r" (ones), [flimit] "r" (flimit)
+  );
+
+  *hev = hev1;
+  *mask = s2;
+}
+
+static INLINE void filter_hev_mask_flatmask4_dspr2(uint32_t limit,
+                                                   uint32_t flimit,
+                                                   uint32_t thresh,
+                                                   uint32_t p1, uint32_t p0,
+                                                   uint32_t p3, uint32_t p2,
+                                                   uint32_t q0, uint32_t q1,
+                                                   uint32_t q2, uint32_t q3,
+                                                   uint32_t *hev,
+                                                   uint32_t *mask,
+                                                   uint32_t *flat) {
+  uint32_t  c, r, r3, r_k, r_flat;
+  uint32_t  s1, s2, s3;
+  uint32_t  ones = 0xFFFFFFFF;
+  uint32_t  flat_thresh = 0x01010101;
+  uint32_t  hev1;
+  uint32_t  flat1;
+
+  __asm__ __volatile__ (
+      /* mask |= (abs(p3 - p2) > limit) */
+      "subu_s.qb      %[c],       %[p3],          %[p2]        \n\t"
+      "subu_s.qb      %[r_k],     %[p2],          %[p3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       $0,             %[c]         \n\t"
+
+      /* mask |= (abs(p2 - p1) > limit) */
+      "subu_s.qb      %[c],       %[p2],          %[p1]        \n\t"
+      "subu_s.qb      %[r_k],     %[p1],          %[p2]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       %[r],           %[c]         \n\t"
+
+      /* mask |= (abs(p1 - p0) > limit)
+       * hev  |= (abs(p1 - p0) > thresh)
+       * flat |= (abs(p1 - p0) > thresh)
+       */
+      "subu_s.qb      %[c],       %[p1],          %[p0]        \n\t"
+      "subu_s.qb      %[r_k],     %[p0],          %[p1]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[thresh],      %[r_k]       \n\t"
+      "or             %[r3],      $0,             %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       %[r],           %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  $0,             %[c]         \n\t"
+
+      /* mask |= (abs(q1 - q0) > limit)
+       * hev  |= (abs(q1 - q0) > thresh)
+       * flat |= (abs(q1 - q0) > thresh)
+       */
+      "subu_s.qb      %[c],       %[q1],          %[q0]        \n\t"
+      "subu_s.qb      %[r_k],     %[q0],          %[q1]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[thresh],      %[r_k]       \n\t"
+      "or             %[r3],      %[r3],          %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       %[r],           %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(p0 - p2) > thresh) */
+      "subu_s.qb      %[c],       %[p0],          %[p2]        \n\t"
+      "subu_s.qb      %[r_k],     %[p2],          %[p0]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(q0 - q2) > thresh) */
+      "subu_s.qb      %[c],       %[q0],          %[q2]        \n\t"
+      "subu_s.qb      %[r_k],     %[q2],          %[q0]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(p3 - p0) > thresh) */
+      "subu_s.qb      %[c],       %[p3],          %[p0]        \n\t"
+      "subu_s.qb      %[r_k],     %[p0],          %[p3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(q3 - q0) > thresh) */
+      "subu_s.qb      %[c],       %[q3],          %[q0]        \n\t"
+      "subu_s.qb      %[r_k],     %[q0],          %[q3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+      "sll            %[r_flat],  %[r_flat],      24           \n\t"
+      /* look at stall here */
+      "wrdsp          %[r_flat]                                \n\t"
+      "pick.qb        %[flat1],   $0,             %[ones]      \n\t"
+
+      /* mask |= (abs(q2 - q1) > limit) */
+      "subu_s.qb      %[c],       %[q2],          %[q1]        \n\t"
+      "subu_s.qb      %[r_k],     %[q1],          %[q2]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       %[r],           %[c]         \n\t"
+      "sll            %[r3],      %[r3],          24           \n\t"
+
+      /* mask |= (abs(q3 - q2) > limit) */
+      "subu_s.qb      %[c],       %[q3],          %[q2]        \n\t"
+      "subu_s.qb      %[r_k],     %[q2],          %[q3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       %[r],           %[c]         \n\t"
+
+      : [c] "=&r" (c), [r_k] "=&r" (r_k), [r] "=&r" (r), [r3] "=&r" (r3),
+        [r_flat] "=&r" (r_flat), [flat1] "=&r" (flat1)
+      : [limit] "r" (limit), [p3] "r" (p3), [p2] "r" (p2),
+        [p1] "r" (p1), [p0] "r" (p0), [q1] "r" (q1), [q0] "r" (q0),
+        [q2] "r" (q2), [q3] "r" (q3), [thresh] "r" (thresh),
+        [flat_thresh] "r" (flat_thresh), [ones] "r" (ones)
+  );
+
+  __asm__ __volatile__ (
+      /* abs(p0 - q0) */
+      "subu_s.qb      %[c],   %[p0],     %[q0]        \n\t"
+      "subu_s.qb      %[r_k], %[q0],     %[p0]        \n\t"
+      "wrdsp          %[r3]                           \n\t"
+      "or             %[s1],  %[r_k],    %[c]         \n\t"
+
+      /* abs(p1 - q1) */
+      "subu_s.qb      %[c],    %[p1],    %[q1]        \n\t"
+      "addu_s.qb      %[s3],   %[s1],    %[s1]        \n\t"
+      "pick.qb        %[hev1], %[ones],  $0           \n\t"
+      "subu_s.qb      %[r_k],  %[q1],    %[p1]        \n\t"
+      "or             %[s2],   %[r_k],   %[c]         \n\t"
+
+      /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > flimit * 2 + limit */
+      "shrl.qb        %[s2],   %[s2],     1           \n\t"
+      "addu_s.qb      %[s1],   %[s2],     %[s3]       \n\t"
+      "cmpgu.lt.qb    %[c],    %[flimit], %[s1]       \n\t"
+      "or             %[r],    %[r],      %[c]        \n\t"
+      "sll            %[r],    %[r],      24          \n\t"
+
+      "wrdsp          %[r]                            \n\t"
+      "pick.qb        %[s2],   $0,        %[ones]     \n\t"
+
+      : [c] "=&r" (c), [r_k] "=&r" (r_k), [s1] "=&r" (s1), [hev1] "=&r" (hev1),
+        [s2] "=&r" (s2), [r] "+r" (r), [s3] "=&r" (s3)
+      : [p0] "r" (p0), [q0] "r" (q0), [p1] "r" (p1), [r3] "r" (r3),
+        [q1] "r" (q1), [ones] "r" (ones), [flimit] "r" (flimit)
+  );
+
+  *hev = hev1;
+  *mask = s2;
+  *flat = flat1;
+}
+
+static INLINE void flatmask5(uint32_t p4, uint32_t p3,
+                             uint32_t p2, uint32_t p1,
+                             uint32_t p0, uint32_t q0,
+                             uint32_t q1, uint32_t q2,
+                             uint32_t q3, uint32_t q4,
+                             uint32_t *flat2) {
+  uint32_t  c, r, r_k, r_flat;
+  uint32_t  ones = 0xFFFFFFFF;
+  uint32_t  flat_thresh = 0x01010101;
+  uint32_t  flat1, flat3;
+
+  __asm__ __volatile__ (
+      /* flat |= (abs(p4 - p0) > thresh) */
+      "subu_s.qb      %[c],   %[p4],           %[p0]        \n\t"
+      "subu_s.qb      %[r_k], %[p0],           %[p4]        \n\t"
+      "or             %[r_k], %[r_k],          %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[flat_thresh],  %[r_k]       \n\t"
+      "or             %[r],   $0,              %[c]         \n\t"
+
+      /* flat |= (abs(q4 - q0) > thresh) */
+      "subu_s.qb      %[c],     %[q4],           %[q0]     \n\t"
+      "subu_s.qb      %[r_k],   %[q0],           %[q4]     \n\t"
+      "or             %[r_k],   %[r_k],          %[c]      \n\t"
+      "cmpgu.lt.qb    %[c],     %[flat_thresh],  %[r_k]    \n\t"
+      "or             %[r],     %[r],            %[c]      \n\t"
+      "sll            %[r],     %[r],            24        \n\t"
+      "wrdsp          %[r]                                 \n\t"
+      "pick.qb        %[flat3], $0,           %[ones]      \n\t"
+
+      /* flat |= (abs(p1 - p0) > thresh) */
+      "subu_s.qb      %[c],       %[p1],          %[p0]        \n\t"
+      "subu_s.qb      %[r_k],     %[p0],          %[p1]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  $0,             %[c]         \n\t"
+
+      /* flat |= (abs(q1 - q0) > thresh) */
+      "subu_s.qb      %[c],      %[q1],           %[q0]        \n\t"
+      "subu_s.qb      %[r_k],    %[q0],           %[q1]        \n\t"
+      "or             %[r_k],    %[r_k],          %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],      %[flat_thresh],  %[r_k]       \n\t"
+      "or             %[r_flat], %[r_flat],       %[c]         \n\t"
+
+      /* flat |= (abs(p0 - p2) > thresh) */
+      "subu_s.qb      %[c],       %[p0],          %[p2]        \n\t"
+      "subu_s.qb      %[r_k],     %[p2],          %[p0]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(q0 - q2) > thresh) */
+      "subu_s.qb      %[c],       %[q0],          %[q2]        \n\t"
+      "subu_s.qb      %[r_k],     %[q2],          %[q0]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(p3 - p0) > thresh) */
+      "subu_s.qb      %[c],       %[p3],          %[p0]        \n\t"
+      "subu_s.qb      %[r_k],     %[p0],          %[p3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(q3 - q0) > thresh) */
+      "subu_s.qb      %[c],       %[q3],          %[q0]        \n\t"
+      "subu_s.qb      %[r_k],     %[q0],          %[q3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+      "sll            %[r_flat],  %[r_flat],      24           \n\t"
+      "wrdsp          %[r_flat]                                \n\t"
+      "pick.qb        %[flat1],   $0,             %[ones]      \n\t"
+      /* flat & flatmask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3) */
+      "and            %[flat1],  %[flat3],        %[flat1]     \n\t"
+
+      : [c] "=&r" (c), [r_k] "=&r" (r_k), [r] "=&r" (r),
+        [r_flat] "=&r" (r_flat), [flat1] "=&r" (flat1), [flat3] "=&r" (flat3)
+      : [p4] "r" (p4), [p3] "r" (p3), [p2] "r" (p2),
+        [p1] "r" (p1), [p0] "r" (p0), [q0] "r" (q0), [q1] "r" (q1),
+        [q2] "r" (q2), [q3] "r" (q3), [q4] "r" (q4),
+        [flat_thresh] "r" (flat_thresh), [ones] "r" (ones)
+  );
+
+  *flat2 = flat1;
+}
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_MIPS_DSPR2_VP9_LOOPFILTER_MASKS_DSPR2_H_
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_mb_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/loopfilter_mb_dspr2.c
new file mode 100644
index 0000000..dd0545e
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_mb_dspr2.c
@@ -0,0 +1,649 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/mips/common_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_filters_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_macros_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_masks_dspr2.h"
+#include "vpx_mem/vpx_mem.h"
+
+#if HAVE_DSPR2
+void vpx_lpf_horizontal_8_dspr2(unsigned char *s,
+                                int pitch,
+                                const uint8_t *blimit,
+                                const uint8_t *limit,
+                                const uint8_t *thresh) {
+  uint32_t  mask;
+  uint32_t  hev, flat;
+  uint8_t   i;
+  uint8_t   *sp3, *sp2, *sp1, *sp0, *sq0, *sq1, *sq2, *sq3;
+  uint32_t  thresh_vec, flimit_vec, limit_vec;
+  uint32_t  uflimit, ulimit, uthresh;
+  uint32_t  p1_f0, p0_f0, q0_f0, q1_f0;
+  uint32_t  p3, p2, p1, p0, q0, q1, q2, q3;
+  uint32_t  p0_l, p1_l, p2_l, p3_l, q0_l, q1_l, q2_l, q3_l;
+  uint32_t  p0_r, p1_r, p2_r, p3_r, q0_r, q1_r, q2_r, q3_r;
+
+  uflimit = *blimit;
+  ulimit  = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__ (
+      "replv.qb       %[thresh_vec],    %[uthresh]    \n\t"
+      "replv.qb       %[flimit_vec],    %[uflimit]    \n\t"
+      "replv.qb       %[limit_vec],     %[ulimit]     \n\t"
+
+      : [thresh_vec] "=&r" (thresh_vec), [flimit_vec] "=&r" (flimit_vec),
+        [limit_vec] "=r" (limit_vec)
+      : [uthresh] "r" (uthresh), [uflimit] "r" (uflimit), [ulimit] "r" (ulimit)
+  );
+
+  /* prefetch data for store */
+  prefetch_store(s);
+
+  for (i = 0; i < 2; i++) {
+    sp3 = s - (pitch << 2);
+    sp2 = sp3 + pitch;
+    sp1 = sp2 + pitch;
+    sp0 = sp1 + pitch;
+    sq0 = s;
+    sq1 = s + pitch;
+    sq2 = sq1 + pitch;
+    sq3 = sq2 + pitch;
+
+    __asm__ __volatile__ (
+        "lw     %[p3],      (%[sp3])    \n\t"
+        "lw     %[p2],      (%[sp2])    \n\t"
+        "lw     %[p1],      (%[sp1])    \n\t"
+        "lw     %[p0],      (%[sp0])    \n\t"
+        "lw     %[q0],      (%[sq0])    \n\t"
+        "lw     %[q1],      (%[sq1])    \n\t"
+        "lw     %[q2],      (%[sq2])    \n\t"
+        "lw     %[q3],      (%[sq3])    \n\t"
+
+        : [p3] "=&r" (p3), [p2] "=&r" (p2), [p1] "=&r" (p1), [p0] "=&r" (p0),
+          [q3] "=&r" (q3), [q2] "=&r" (q2), [q1] "=&r" (q1), [q0] "=&r" (q0)
+        : [sp3] "r" (sp3), [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+          [sq3] "r" (sq3), [sq2] "r" (sq2), [sq1] "r" (sq1), [sq0] "r" (sq0)
+    );
+
+    filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec,
+                                    p1, p0, p3, p2, q0, q1, q2, q3,
+                                    &hev, &mask, &flat);
+
+    if ((flat == 0) && (mask != 0)) {
+      filter1_dspr2(mask, hev, p1, p0, q0, q1,
+                    &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      __asm__ __volatile__ (
+          "sw       %[p1_f0],   (%[sp1])    \n\t"
+          "sw       %[p0_f0],   (%[sp0])    \n\t"
+          "sw       %[q0_f0],   (%[sq0])    \n\t"
+          "sw       %[q1_f0],   (%[sq1])    \n\t"
+
+          :
+          : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+            [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+            [sp1] "r" (sp1), [sp0] "r" (sp0),
+            [sq0] "r" (sq0), [sq1] "r" (sq1)
+      );
+    } else if ((mask & flat) == 0xFFFFFFFF) {
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l,
+                     &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r,
+                     &q0_r, &q1_r, &q2_r, &q3_r);
+
+      COMBINE_LEFT_RIGHT_0TO2()
+
+      __asm__ __volatile__ (
+          "sw       %[p2],      (%[sp2])    \n\t"
+          "sw       %[p1],      (%[sp1])    \n\t"
+          "sw       %[p0],      (%[sp0])    \n\t"
+          "sw       %[q0],      (%[sq0])    \n\t"
+          "sw       %[q1],      (%[sq1])    \n\t"
+          "sw       %[q2],      (%[sq2])    \n\t"
+
+          :
+          : [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0),
+            [q0] "r" (q0), [q1] "r" (q1), [q2] "r" (q2),
+            [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+            [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+      );
+    } else if ((flat != 0) && (mask != 0)) {
+      /* filtering */
+      filter1_dspr2(mask, hev, p1, p0, q0, q1,
+                    &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l,
+                     &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r,
+                     &q0_r, &q1_r, &q2_r, &q3_r);
+
+      if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb     %[p2_r],    (%[sp2])    \n\t"
+            "sb     %[p1_r],    (%[sp1])    \n\t"
+            "sb     %[p0_r],    (%[sp0])    \n\t"
+            "sb     %[q0_r],    (%[sq0])    \n\t"
+            "sb     %[q1_r],    (%[sq1])    \n\t"
+            "sb     %[q2_r],    (%[sq2])    \n\t"
+
+            :
+            : [p2_r] "r" (p2_r), [p1_r] "r" (p1_r), [p0_r] "r" (p0_r),
+              [q0_r] "r" (q0_r), [q1_r] "r" (q1_r), [q2_r] "r" (q2_r),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  (%[sp1])    \n\t"
+            "sb         %[p0_f0],  (%[sp0])    \n\t"
+            "sb         %[q0_f0],  (%[sq0])    \n\t"
+            "sb         %[q1_f0],  (%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p2_r],    %[p2_r],    16      \n\t"
+          "srl      %[p1_r],    %[p1_r],    16      \n\t"
+          "srl      %[p0_r],    %[p0_r],    16      \n\t"
+          "srl      %[q0_r],    %[q0_r],    16      \n\t"
+          "srl      %[q1_r],    %[q1_r],    16      \n\t"
+          "srl      %[q2_r],    %[q2_r],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_r] "+r" (p2_r), [p1_r] "+r" (p1_r), [p0_r] "+r" (p0_r),
+            [q0_r] "+r" (q0_r), [q1_r] "+r" (q1_r), [q2_r] "+r" (q2_r),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb     %[p2_r],    +1(%[sp2])    \n\t"
+            "sb     %[p1_r],    +1(%[sp1])    \n\t"
+            "sb     %[p0_r],    +1(%[sp0])    \n\t"
+            "sb     %[q0_r],    +1(%[sq0])    \n\t"
+            "sb     %[q1_r],    +1(%[sq1])    \n\t"
+            "sb     %[q2_r],    +1(%[sq2])    \n\t"
+
+            :
+            : [p2_r] "r" (p2_r), [p1_r] "r" (p1_r), [p0_r] "r" (p0_r),
+              [q0_r] "r" (q0_r), [q1_r] "r" (q1_r), [q2_r] "r" (q2_r),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   +1(%[sp1])    \n\t"
+            "sb     %[p0_f0],   +1(%[sp0])    \n\t"
+            "sb     %[q0_f0],   +1(%[sq0])    \n\t"
+            "sb     %[q1_f0],   +1(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p1_f0],   %[p1_f0],   8     \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8     \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8     \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8     \n\t"
+
+          : [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0),
+            [q0] "+r" (q0), [q1] "+r" (q1), [q2] "+r" (q2),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb     %[p2_l],    +2(%[sp2])    \n\t"
+            "sb     %[p1_l],    +2(%[sp1])    \n\t"
+            "sb     %[p0_l],    +2(%[sp0])    \n\t"
+            "sb     %[q0_l],    +2(%[sq0])    \n\t"
+            "sb     %[q1_l],    +2(%[sq1])    \n\t"
+            "sb     %[q2_l],    +2(%[sq2])    \n\t"
+
+            :
+            : [p2_l] "r" (p2_l), [p1_l] "r" (p1_l), [p0_l] "r" (p0_l),
+              [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   +2(%[sp1])    \n\t"
+            "sb     %[p0_f0],   +2(%[sp0])    \n\t"
+            "sb     %[q0_f0],   +2(%[sq0])    \n\t"
+            "sb     %[q1_f0],   +2(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p2_l],    %[p2_l],    16      \n\t"
+          "srl      %[p1_l],    %[p1_l],    16      \n\t"
+          "srl      %[p0_l],    %[p0_l],    16      \n\t"
+          "srl      %[q0_l],    %[q0_l],    16      \n\t"
+          "srl      %[q1_l],    %[q1_l],    16      \n\t"
+          "srl      %[q2_l],    %[q2_l],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_l] "+r" (p2_l), [p1_l] "+r" (p1_l), [p0_l] "+r" (p0_l),
+            [q0_l] "+r" (q0_l), [q1_l] "+r" (q1_l), [q2_l] "+r" (q2_l),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb     %[p2_l],    +3(%[sp2])    \n\t"
+            "sb     %[p1_l],    +3(%[sp1])    \n\t"
+            "sb     %[p0_l],    +3(%[sp0])    \n\t"
+            "sb     %[q0_l],    +3(%[sq0])    \n\t"
+            "sb     %[q1_l],    +3(%[sq1])    \n\t"
+            "sb     %[q2_l],    +3(%[sq2])    \n\t"
+
+            :
+            : [p2_l] "r" (p2_l), [p1_l] "r" (p1_l), [p0_l] "r" (p0_l),
+              [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   +3(%[sp1])    \n\t"
+            "sb     %[p0_f0],   +3(%[sp0])    \n\t"
+            "sb     %[q0_f0],   +3(%[sq0])    \n\t"
+            "sb     %[q1_f0],   +3(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+    }
+
+    s = s + 4;
+  }
+}
+
+void vpx_lpf_vertical_8_dspr2(unsigned char *s,
+                              int pitch,
+                              const uint8_t *blimit,
+                              const uint8_t *limit,
+                              const uint8_t *thresh) {
+  uint8_t   i;
+  uint32_t  mask, hev, flat;
+  uint8_t   *s1, *s2, *s3, *s4;
+  uint32_t  prim1, prim2, sec3, sec4, prim3, prim4;
+  uint32_t  thresh_vec, flimit_vec, limit_vec;
+  uint32_t  uflimit, ulimit, uthresh;
+  uint32_t  p3, p2, p1, p0, q3, q2, q1, q0;
+  uint32_t  p1_f0, p0_f0, q0_f0, q1_f0;
+  uint32_t  p0_l, p1_l, p2_l, p3_l, q0_l, q1_l, q2_l, q3_l;
+  uint32_t  p0_r, p1_r, p2_r, p3_r, q0_r, q1_r, q2_r, q3_r;
+
+  uflimit = *blimit;
+  ulimit  = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__ (
+      "replv.qb     %[thresh_vec],  %[uthresh]    \n\t"
+      "replv.qb     %[flimit_vec],  %[uflimit]    \n\t"
+      "replv.qb     %[limit_vec],   %[ulimit]     \n\t"
+
+      : [thresh_vec] "=&r" (thresh_vec), [flimit_vec] "=&r" (flimit_vec),
+        [limit_vec] "=r" (limit_vec)
+      : [uthresh] "r" (uthresh), [uflimit] "r" (uflimit), [ulimit] "r" (ulimit)
+  );
+
+  prefetch_store(s + pitch);
+
+  for (i = 0; i < 2; i++) {
+    s1 = s;
+    s2 = s + pitch;
+    s3 = s2 + pitch;
+    s4 = s3 + pitch;
+    s  = s4 + pitch;
+
+    __asm__ __volatile__ (
+        "lw     %[p0],  -4(%[s1])    \n\t"
+        "lw     %[p1],  -4(%[s2])    \n\t"
+        "lw     %[p2],  -4(%[s3])    \n\t"
+        "lw     %[p3],  -4(%[s4])    \n\t"
+        "lw     %[q3],    (%[s1])    \n\t"
+        "lw     %[q2],    (%[s2])    \n\t"
+        "lw     %[q1],    (%[s3])    \n\t"
+        "lw     %[q0],    (%[s4])    \n\t"
+
+        : [p3] "=&r" (p3), [p2] "=&r" (p2), [p1] "=&r" (p1), [p0] "=&r" (p0),
+          [q0] "=&r" (q0), [q1] "=&r" (q1), [q2] "=&r" (q2), [q3] "=&r" (q3)
+        : [s1] "r" (s1), [s2] "r" (s2), [s3] "r" (s3), [s4] "r" (s4)
+    );
+
+    /* transpose p3, p2, p1, p0
+       original (when loaded from memory)
+       register       -4    -3   -2     -1
+         p0         p0_0  p0_1  p0_2  p0_3
+         p1         p1_0  p1_1  p1_2  p1_3
+         p2         p2_0  p2_1  p2_2  p2_3
+         p3         p3_0  p3_1  p3_2  p3_3
+
+       after transpose
+       register
+         p0         p3_3  p2_3  p1_3  p0_3
+         p1         p3_2  p2_2  p1_2  p0_2
+         p2         p3_1  p2_1  p1_1  p0_1
+         p3         p3_0  p2_0  p1_0  p0_0
+    */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p0],      %[p1]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p0],      %[p1]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p2],      %[p3]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p2],      %[p3]       \n\t"
+
+        "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p0],      %[p1],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p2],      %[p3],      %[sec4]     \n\t"
+        "append         %[p1],      %[sec3],    16          \n\t"
+        "append         %[p3],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p0] "+r" (p0), [p1] "+r" (p1), [p2] "+r" (p2), [p3] "+r" (p3),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* transpose q0, q1, q2, q3
+       original (when loaded from memory)
+       register       +1    +2    +3    +4
+         q3         q3_0  q3_1  q3_2  q3_3
+         q2         q2_0  q2_1  q2_2  q2_3
+         q1         q1_0  q1_1  q1_2  q1_3
+         q0         q0_0  q0_1  q0_2  q0_3
+
+       after transpose
+       register
+         q3         q0_3  q1_3  q2_3  q3_3
+         q2         q0_2  q1_2  q2_2  q3_2
+         q1         q0_1  q1_1  q2_1  q3_1
+         q0         q0_0  q1_0  q2_0  q3_0
+    */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[q3],      %[q2]       \n\t"
+        "precr.qb.ph    %[prim2],   %[q3],      %[q2]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[q1],      %[q0]       \n\t"
+        "precr.qb.ph    %[prim4],   %[q1],      %[q0]       \n\t"
+
+        "precrq.qb.ph   %[q2],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[q0],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[q3],      %[q2],      %[sec3]     \n\t"
+        "precrq.ph.w    %[q1],      %[q0],      %[sec4]     \n\t"
+        "append         %[q2],      %[sec3],    16          \n\t"
+        "append         %[q0],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [q3] "+r" (q3), [q2] "+r" (q2), [q1] "+r" (q1), [q0] "+r" (q0),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec,
+                                    p1, p0, p3, p2, q0, q1, q2, q3,
+                                    &hev, &mask, &flat);
+
+    if ((flat == 0) && (mask != 0)) {
+      filter1_dspr2(mask, hev, p1, p0, q0, q1,
+                    &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+      STORE_F0()
+    } else if ((mask & flat) == 0xFFFFFFFF) {
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l,
+                     &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r,
+                     &q0_r, &q1_r, &q2_r, &q3_r);
+
+      STORE_F1()
+    } else if ((flat != 0) && (mask != 0)) {
+      filter1_dspr2(mask, hev, p1, p0, q0, q1,
+                    &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l,
+                     &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r,
+                     &q0_r, &q1_r, &q2_r, &q3_r);
+
+      if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb         %[p2_r],  -3(%[s4])    \n\t"
+            "sb         %[p1_r],  -2(%[s4])    \n\t"
+            "sb         %[p0_r],  -1(%[s4])    \n\t"
+            "sb         %[q0_r],    (%[s4])    \n\t"
+            "sb         %[q1_r],  +1(%[s4])    \n\t"
+            "sb         %[q2_r],  +2(%[s4])    \n\t"
+
+            :
+            : [p2_r] "r" (p2_r), [p1_r] "r" (p1_r), [p0_r] "r" (p0_r),
+              [q0_r] "r" (q0_r), [q1_r] "r" (q1_r), [q2_r] "r" (q2_r),
+              [s4] "r" (s4)
+        );
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  -2(%[s4])    \n\t"
+            "sb         %[p0_f0],  -1(%[s4])    \n\t"
+            "sb         %[q0_f0],    (%[s4])    \n\t"
+            "sb         %[q1_f0],  +1(%[s4])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s4] "r" (s4)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p2_r],    %[p2_r],    16      \n\t"
+          "srl      %[p1_r],    %[p1_r],    16      \n\t"
+          "srl      %[p0_r],    %[p0_r],    16      \n\t"
+          "srl      %[q0_r],    %[q0_r],    16      \n\t"
+          "srl      %[q1_r],    %[q1_r],    16      \n\t"
+          "srl      %[q2_r],    %[q2_r],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_r] "+r" (p2_r), [p1_r] "+r" (p1_r), [p0_r] "+r" (p0_r),
+            [q0_r] "+r" (q0_r), [q1_r] "+r" (q1_r), [q2_r] "+r" (q2_r),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb         %[p2_r],  -3(%[s3])    \n\t"
+            "sb         %[p1_r],  -2(%[s3])    \n\t"
+            "sb         %[p0_r],  -1(%[s3])    \n\t"
+            "sb         %[q0_r],    (%[s3])    \n\t"
+            "sb         %[q1_r],  +1(%[s3])    \n\t"
+            "sb         %[q2_r],  +2(%[s3])    \n\t"
+
+            :
+            : [p2_r] "r" (p2_r), [p1_r] "r" (p1_r), [p0_r] "r" (p0_r),
+              [q0_r] "r" (q0_r), [q1_r] "r" (q1_r), [q2_r] "r" (q2_r),
+              [s3] "r" (s3)
+        );
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  -2(%[s3])    \n\t"
+            "sb         %[p0_f0],  -1(%[s3])    \n\t"
+            "sb         %[q0_f0],    (%[s3])    \n\t"
+            "sb         %[q1_f0],  +1(%[s3])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s3] "r" (s3)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p1_f0],   %[p1_f0],   8     \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8     \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8     \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8     \n\t"
+
+          : [p2] "+r" (p2), [p1] "+r" (p1), [p0] "+r" (p0),
+            [q0] "+r" (q0), [q1] "+r" (q1), [q2] "+r" (q2),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__ (
+          "sb         %[p2_l],  -3(%[s2])    \n\t"
+          "sb         %[p1_l],  -2(%[s2])    \n\t"
+          "sb         %[p0_l],  -1(%[s2])    \n\t"
+          "sb         %[q0_l],    (%[s2])    \n\t"
+          "sb         %[q1_l],  +1(%[s2])    \n\t"
+          "sb         %[q2_l],  +2(%[s2])    \n\t"
+
+          :
+          : [p2_l] "r" (p2_l), [p1_l] "r" (p1_l), [p0_l] "r" (p0_l),
+            [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+            [s2] "r" (s2)
+        );
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  -2(%[s2])    \n\t"
+            "sb         %[p0_f0],  -1(%[s2])    \n\t"
+            "sb         %[q0_f0],    (%[s2])    \n\t"
+            "sb         %[q1_f0],  +1(%[s2])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s2] "r" (s2)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p2_l],    %[p2_l],    16      \n\t"
+          "srl      %[p1_l],    %[p1_l],    16      \n\t"
+          "srl      %[p0_l],    %[p0_l],    16      \n\t"
+          "srl      %[q0_l],    %[q0_l],    16      \n\t"
+          "srl      %[q1_l],    %[q1_l],    16      \n\t"
+          "srl      %[q2_l],    %[q2_l],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_l] "+r" (p2_l), [p1_l] "+r" (p1_l), [p0_l] "+r" (p0_l),
+            [q0_l] "+r" (q0_l), [q1_l] "+r" (q1_l), [q2_l] "+r" (q2_l),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb         %[p2_l],  -3(%[s1])    \n\t"
+            "sb         %[p1_l],  -2(%[s1])    \n\t"
+            "sb         %[p0_l],  -1(%[s1])    \n\t"
+            "sb         %[q0_l],    (%[s1])    \n\t"
+            "sb         %[q1_l],  +1(%[s1])    \n\t"
+            "sb         %[q2_l],  +2(%[s1])    \n\t"
+
+            :
+            : [p2_l] "r" (p2_l), [p1_l] "r" (p1_l), [p0_l] "r" (p0_l),
+              [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+              [s1] "r" (s1)
+        );
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  -2(%[s1])    \n\t"
+            "sb         %[p0_f0],  -1(%[s1])    \n\t"
+            "sb         %[q0_f0],    (%[s1])    \n\t"
+            "sb         %[q1_f0],  +1(%[s1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0), [q0_f0] "r" (q0_f0),
+              [q1_f0] "r" (q1_f0), [s1] "r" (s1)
+        );
+      }
+    }
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_mb_horiz_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/loopfilter_mb_horiz_dspr2.c
new file mode 100644
index 0000000..85e167c
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_mb_horiz_dspr2.c
@@ -0,0 +1,808 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/mips/common_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_filters_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_macros_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_masks_dspr2.h"
+#include "vpx_mem/vpx_mem.h"
+
+#if HAVE_DSPR2
+static void mb_lpf_horizontal_edge(unsigned char *s,
+                                   int pitch,
+                                   const uint8_t *blimit,
+                                   const uint8_t *limit,
+                                   const uint8_t *thresh,
+                                   int count) {
+  uint32_t  mask;
+  uint32_t  hev, flat, flat2;
+  uint8_t   i;
+  uint8_t   *sp7, *sp6, *sp5, *sp4, *sp3, *sp2, *sp1, *sp0;
+  uint8_t   *sq0, *sq1, *sq2, *sq3, *sq4, *sq5, *sq6, *sq7;
+  uint32_t  thresh_vec, flimit_vec, limit_vec;
+  uint32_t  uflimit, ulimit, uthresh;
+  uint32_t  p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  uint32_t  p1_f0, p0_f0, q0_f0, q1_f0;
+  uint32_t  p7_l, p6_l, p5_l, p4_l, p3_l, p2_l, p1_l, p0_l;
+  uint32_t  q0_l, q1_l, q2_l, q3_l, q4_l, q5_l, q6_l, q7_l;
+  uint32_t  p7_r, p6_r, p5_r, p4_r, p3_r, p2_r, p1_r, p0_r;
+  uint32_t  q0_r, q1_r, q2_r, q3_r, q4_r, q5_r, q6_r, q7_r;
+  uint32_t  p2_l_f1, p1_l_f1, p0_l_f1, p2_r_f1, p1_r_f1, p0_r_f1;
+  uint32_t  q0_l_f1, q1_l_f1, q2_l_f1, q0_r_f1, q1_r_f1, q2_r_f1;
+
+  uflimit = *blimit;
+  ulimit  = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__ (
+      "replv.qb       %[thresh_vec],    %[uthresh]      \n\t"
+      "replv.qb       %[flimit_vec],    %[uflimit]      \n\t"
+      "replv.qb       %[limit_vec],     %[ulimit]       \n\t"
+
+      : [thresh_vec] "=&r" (thresh_vec), [flimit_vec] "=&r" (flimit_vec),
+        [limit_vec] "=r" (limit_vec)
+      : [uthresh] "r" (uthresh), [uflimit] "r" (uflimit), [ulimit] "r" (ulimit)
+  );
+
+  /* prefetch data for store */
+  prefetch_store(s);
+
+  for (i = 0; i < (2 * count); i++) {
+    sp7 = s - (pitch << 3);
+    sp6 = sp7 + pitch;
+    sp5 = sp6 + pitch;
+    sp4 = sp5 + pitch;
+    sp3 = sp4 + pitch;
+    sp2 = sp3 + pitch;
+    sp1 = sp2 + pitch;
+    sp0 = sp1 + pitch;
+    sq0 = s;
+    sq1 = s + pitch;
+    sq2 = sq1 + pitch;
+    sq3 = sq2 + pitch;
+    sq4 = sq3 + pitch;
+    sq5 = sq4 + pitch;
+    sq6 = sq5 + pitch;
+    sq7 = sq6 + pitch;
+
+    __asm__ __volatile__ (
+        "lw     %[p7],      (%[sp7])            \n\t"
+        "lw     %[p6],      (%[sp6])            \n\t"
+        "lw     %[p5],      (%[sp5])            \n\t"
+        "lw     %[p4],      (%[sp4])            \n\t"
+        "lw     %[p3],      (%[sp3])            \n\t"
+        "lw     %[p2],      (%[sp2])            \n\t"
+        "lw     %[p1],      (%[sp1])            \n\t"
+        "lw     %[p0],      (%[sp0])            \n\t"
+
+        : [p3] "=&r" (p3), [p2] "=&r" (p2), [p1] "=&r" (p1), [p0] "=&r" (p0),
+          [p7] "=&r" (p7), [p6] "=&r" (p6), [p5] "=&r" (p5), [p4] "=&r" (p4)
+        : [sp3] "r" (sp3), [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+          [sp4] "r" (sp4), [sp5] "r" (sp5), [sp6] "r" (sp6), [sp7] "r" (sp7)
+    );
+
+    __asm__ __volatile__ (
+        "lw     %[q0],      (%[sq0])            \n\t"
+        "lw     %[q1],      (%[sq1])            \n\t"
+        "lw     %[q2],      (%[sq2])            \n\t"
+        "lw     %[q3],      (%[sq3])            \n\t"
+        "lw     %[q4],      (%[sq4])            \n\t"
+        "lw     %[q5],      (%[sq5])            \n\t"
+        "lw     %[q6],      (%[sq6])            \n\t"
+        "lw     %[q7],      (%[sq7])            \n\t"
+
+        : [q3] "=&r" (q3), [q2] "=&r" (q2), [q1] "=&r" (q1), [q0] "=&r" (q0),
+          [q7] "=&r" (q7), [q6] "=&r" (q6), [q5] "=&r" (q5), [q4] "=&r" (q4)
+        : [sq3] "r" (sq3), [sq2] "r" (sq2), [sq1] "r" (sq1), [sq0] "r" (sq0),
+          [sq4] "r" (sq4), [sq5] "r" (sq5), [sq6] "r" (sq6), [sq7] "r" (sq7)
+    );
+
+    filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec,
+                                    p1, p0, p3, p2, q0, q1, q2, q3,
+                                    &hev, &mask, &flat);
+
+    flatmask5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, &flat2);
+
+    /* f0 */
+    if (((flat2 == 0) && (flat == 0) && (mask != 0)) ||
+        ((flat2 != 0) && (flat == 0) && (mask != 0))) {
+      filter1_dspr2(mask, hev, p1, p0, q0, q1,
+                    &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      __asm__ __volatile__ (
+          "sw       %[p1_f0],   (%[sp1])            \n\t"
+          "sw       %[p0_f0],   (%[sp0])            \n\t"
+          "sw       %[q0_f0],   (%[sq0])            \n\t"
+          "sw       %[q1_f0],   (%[sq1])            \n\t"
+
+          :
+          : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+            [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+            [sp1] "r" (sp1), [sp0] "r" (sp0),
+            [sq0] "r" (sq0), [sq1] "r" (sq1)
+      );
+    } else if ((flat2 == 0XFFFFFFFF) && (flat == 0xFFFFFFFF) &&
+               (mask == 0xFFFFFFFF)) {
+      /* f2 */
+      PACK_LEFT_0TO3()
+      PACK_LEFT_4TO7()
+      wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l,
+                          &p3_l, &p2_l, &p1_l, &p0_l,
+                          &q0_l, &q1_l, &q2_l, &q3_l,
+                          &q4_l, &q5_l, &q6_l, &q7_l);
+
+      PACK_RIGHT_0TO3()
+      PACK_RIGHT_4TO7()
+      wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r,
+                          &p3_r, &p2_r, &p1_r, &p0_r,
+                          &q0_r, &q1_r, &q2_r, &q3_r,
+                          &q4_r, &q5_r, &q6_r, &q7_r);
+
+      COMBINE_LEFT_RIGHT_0TO2()
+      COMBINE_LEFT_RIGHT_3TO6()
+
+      __asm__ __volatile__ (
+          "sw         %[p6], (%[sp6])    \n\t"
+          "sw         %[p5], (%[sp5])    \n\t"
+          "sw         %[p4], (%[sp4])    \n\t"
+          "sw         %[p3], (%[sp3])    \n\t"
+          "sw         %[p2], (%[sp2])    \n\t"
+          "sw         %[p1], (%[sp1])    \n\t"
+          "sw         %[p0], (%[sp0])    \n\t"
+
+          :
+          : [p6] "r" (p6), [p5] "r" (p5), [p4] "r" (p4), [p3] "r" (p3),
+            [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0),
+            [sp6] "r" (sp6), [sp5] "r" (sp5), [sp4] "r" (sp4), [sp3] "r" (sp3),
+            [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0)
+      );
+
+      __asm__ __volatile__ (
+          "sw         %[q6], (%[sq6])    \n\t"
+          "sw         %[q5], (%[sq5])    \n\t"
+          "sw         %[q4], (%[sq4])    \n\t"
+          "sw         %[q3], (%[sq3])    \n\t"
+          "sw         %[q2], (%[sq2])    \n\t"
+          "sw         %[q1], (%[sq1])    \n\t"
+          "sw         %[q0], (%[sq0])    \n\t"
+
+          :
+          : [q6] "r" (q6), [q5] "r" (q5), [q4] "r" (q4), [q3] "r" (q3),
+            [q2] "r" (q2), [q1] "r" (q1), [q0] "r" (q0),
+            [sq6] "r" (sq6), [sq5] "r" (sq5), [sq4] "r" (sq4), [sq3] "r" (sq3),
+            [sq2] "r" (sq2), [sq1] "r" (sq1), [sq0] "r" (sq0)
+      );
+    } else if ((flat2 == 0) && (flat == 0xFFFFFFFF) && (mask == 0xFFFFFFFF)) {
+      /* f1 */
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l,
+                     &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r,
+                     &q0_r, &q1_r, &q2_r, &q3_r);
+
+      COMBINE_LEFT_RIGHT_0TO2()
+
+      __asm__ __volatile__ (
+          "sw         %[p2], (%[sp2])    \n\t"
+          "sw         %[p1], (%[sp1])    \n\t"
+          "sw         %[p0], (%[sp0])    \n\t"
+          "sw         %[q0], (%[sq0])    \n\t"
+          "sw         %[q1], (%[sq1])    \n\t"
+          "sw         %[q2], (%[sq2])    \n\t"
+
+          :
+          : [p2] "r" (p2), [p1] "r" (p1), [p0] "r" (p0),
+            [q0] "r" (q0), [q1] "r" (q1), [q2] "r" (q2),
+            [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+            [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+      );
+    } else if ((flat2 == 0) && (flat != 0) && (mask != 0)) {
+      /* f0+f1 */
+      filter1_dspr2(mask, hev, p1, p0, q0, q1,
+                    &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l,
+                     &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r,
+                     &q0_r, &q1_r, &q2_r, &q3_r);
+
+      if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb         %[p2_r],  (%[sp2])    \n\t"
+            "sb         %[p1_r],  (%[sp1])    \n\t"
+            "sb         %[p0_r],  (%[sp0])    \n\t"
+            "sb         %[q0_r],  (%[sq0])    \n\t"
+            "sb         %[q1_r],  (%[sq1])    \n\t"
+            "sb         %[q2_r],  (%[sq2])    \n\t"
+
+            :
+            : [p2_r] "r" (p2_r), [p1_r] "r" (p1_r), [p0_r] "r" (p0_r),
+              [q0_r] "r" (q0_r), [q1_r] "r" (q1_r), [q2_r] "r" (q2_r),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  (%[sp1])    \n\t"
+            "sb         %[p0_f0],  (%[sp0])    \n\t"
+            "sb         %[q0_f0],  (%[sq0])    \n\t"
+            "sb         %[q1_f0],  (%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p2_r],    %[p2_r],    16      \n\t"
+          "srl      %[p1_r],    %[p1_r],    16      \n\t"
+          "srl      %[p0_r],    %[p0_r],    16      \n\t"
+          "srl      %[q0_r],    %[q0_r],    16      \n\t"
+          "srl      %[q1_r],    %[q1_r],    16      \n\t"
+          "srl      %[q2_r],    %[q2_r],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_r] "+r" (p2_r), [p1_r] "+r" (p1_r), [p0_r] "+r" (p0_r),
+            [q0_r] "+r" (q0_r), [q1_r] "+r" (q1_r), [q2_r] "+r" (q2_r),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb         %[p2_r],  +1(%[sp2])    \n\t"
+            "sb         %[p1_r],  +1(%[sp1])    \n\t"
+            "sb         %[p0_r],  +1(%[sp0])    \n\t"
+            "sb         %[q0_r],  +1(%[sq0])    \n\t"
+            "sb         %[q1_r],  +1(%[sq1])    \n\t"
+            "sb         %[q2_r],  +1(%[sq2])    \n\t"
+
+            :
+            : [p2_r] "r" (p2_r), [p1_r] "r" (p1_r), [p0_r] "r" (p0_r),
+              [q0_r] "r" (q0_r), [q1_r] "r" (q1_r), [q2_r] "r" (q2_r),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  +1(%[sp1])    \n\t"
+            "sb         %[p0_f0],  +1(%[sp0])    \n\t"
+            "sb         %[q0_f0],  +1(%[sq0])    \n\t"
+            "sb         %[q1_f0],  +1(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p1_f0],   %[p1_f0],   8     \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8     \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8     \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8     \n\t"
+
+          : [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb         %[p2_l],  +2(%[sp2])    \n\t"
+            "sb         %[p1_l],  +2(%[sp1])    \n\t"
+            "sb         %[p0_l],  +2(%[sp0])    \n\t"
+            "sb         %[q0_l],  +2(%[sq0])    \n\t"
+            "sb         %[q1_l],  +2(%[sq1])    \n\t"
+            "sb         %[q2_l],  +2(%[sq2])    \n\t"
+
+            :
+            : [p2_l] "r" (p2_l), [p1_l] "r" (p1_l), [p0_l] "r" (p0_l),
+              [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  +2(%[sp1])    \n\t"
+            "sb         %[p0_f0],  +2(%[sp0])    \n\t"
+            "sb         %[q0_f0],  +2(%[sq0])    \n\t"
+            "sb         %[q1_f0],  +2(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p2_l],    %[p2_l],    16      \n\t"
+          "srl      %[p1_l],    %[p1_l],    16      \n\t"
+          "srl      %[p0_l],    %[p0_l],    16      \n\t"
+          "srl      %[q0_l],    %[q0_l],    16      \n\t"
+          "srl      %[q1_l],    %[q1_l],    16      \n\t"
+          "srl      %[q2_l],    %[q2_l],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_l] "+r" (p2_l), [p1_l] "+r" (p1_l), [p0_l] "+r" (p0_l),
+            [q0_l] "+r" (q0_l), [q1_l] "+r" (q1_l), [q2_l] "+r" (q2_l),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb         %[p2_l],  +3(%[sp2])    \n\t"
+            "sb         %[p1_l],  +3(%[sp1])    \n\t"
+            "sb         %[p0_l],  +3(%[sp0])    \n\t"
+            "sb         %[q0_l],  +3(%[sq0])    \n\t"
+            "sb         %[q1_l],  +3(%[sq1])    \n\t"
+            "sb         %[q2_l],  +3(%[sq2])    \n\t"
+
+            :
+            : [p2_l] "r" (p2_l), [p1_l] "r" (p1_l), [p0_l] "r" (p0_l),
+              [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  +3(%[sp1])    \n\t"
+            "sb         %[p0_f0],  +3(%[sp0])    \n\t"
+            "sb         %[q0_f0],  +3(%[sq0])    \n\t"
+            "sb         %[q1_f0],  +3(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+    } else if ((flat2 != 0) && (flat != 0) && (mask != 0)) {
+      /* f0 + f1 + f2 */
+      /* f0  function */
+      filter1_dspr2(mask, hev, p1, p0, q0, q1,
+                    &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      /* f1  function */
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter1_dspr2(p3_l, p2_l, p1_l, p0_l,
+                      q0_l, q1_l, q2_l, q3_l,
+                      &p2_l_f1, &p1_l_f1, &p0_l_f1,
+                      &q0_l_f1, &q1_l_f1, &q2_l_f1);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter1_dspr2(p3_r, p2_r, p1_r, p0_r,
+                      q0_r, q1_r, q2_r, q3_r,
+                      &p2_r_f1, &p1_r_f1, &p0_r_f1,
+                      &q0_r_f1, &q1_r_f1, &q2_r_f1);
+
+      /* f2  function */
+      PACK_LEFT_4TO7()
+      wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l,
+                          &p3_l, &p2_l, &p1_l, &p0_l,
+                          &q0_l, &q1_l, &q2_l, &q3_l,
+                          &q4_l, &q5_l, &q6_l, &q7_l);
+
+      PACK_RIGHT_4TO7()
+      wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r,
+                          &p3_r, &p2_r, &p1_r, &p0_r,
+                          &q0_r, &q1_r, &q2_r, &q3_r,
+                          &q4_r, &q5_r, &q6_r, &q7_r);
+
+      if (mask & flat & flat2 & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb         %[p6_r],  (%[sp6])    \n\t"
+            "sb         %[p5_r],  (%[sp5])    \n\t"
+            "sb         %[p4_r],  (%[sp4])    \n\t"
+            "sb         %[p3_r],  (%[sp3])    \n\t"
+            "sb         %[p2_r],  (%[sp2])    \n\t"
+            "sb         %[p1_r],  (%[sp1])    \n\t"
+            "sb         %[p0_r],  (%[sp0])    \n\t"
+
+            :
+            : [p6_r] "r" (p6_r), [p5_r] "r" (p5_r), [p4_r] "r" (p4_r),
+              [p3_r] "r" (p3_r), [p2_r] "r" (p2_r), [p1_r] "r" (p1_r),
+              [sp6] "r" (sp6), [sp5] "r" (sp5), [sp4] "r" (sp4),
+              [sp3] "r" (sp3), [sp2] "r" (sp2), [sp1] "r" (sp1),
+              [p0_r] "r" (p0_r), [sp0] "r" (sp0)
+        );
+
+        __asm__ __volatile__ (
+            "sb         %[q0_r],  (%[sq0])    \n\t"
+            "sb         %[q1_r],  (%[sq1])    \n\t"
+            "sb         %[q2_r],  (%[sq2])    \n\t"
+            "sb         %[q3_r],  (%[sq3])    \n\t"
+            "sb         %[q4_r],  (%[sq4])    \n\t"
+            "sb         %[q5_r],  (%[sq5])    \n\t"
+            "sb         %[q6_r],  (%[sq6])    \n\t"
+
+            :
+            : [q0_r] "r" (q0_r), [q1_r] "r" (q1_r), [q2_r] "r" (q2_r),
+              [q3_r] "r" (q3_r), [q4_r] "r" (q4_r), [q5_r] "r" (q5_r),
+              [q6_r] "r" (q6_r),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2),
+              [sq3] "r" (sq3), [sq4] "r" (sq4), [sq5] "r" (sq5),
+              [sq6] "r" (sq6)
+        );
+      } else if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb         %[p2_r_f1],  (%[sp2])    \n\t"
+            "sb         %[p1_r_f1],  (%[sp1])    \n\t"
+            "sb         %[p0_r_f1],  (%[sp0])    \n\t"
+            "sb         %[q0_r_f1],  (%[sq0])    \n\t"
+            "sb         %[q1_r_f1],  (%[sq1])    \n\t"
+            "sb         %[q2_r_f1],  (%[sq2])    \n\t"
+
+            :
+            : [p2_r_f1] "r" (p2_r_f1), [p1_r_f1] "r" (p1_r_f1),
+              [p0_r_f1] "r" (p0_r_f1), [q0_r_f1] "r" (q0_r_f1),
+              [q1_r_f1] "r" (q1_r_f1), [q2_r_f1] "r" (q2_r_f1),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  (%[sp1])    \n\t"
+            "sb         %[p0_f0],  (%[sp0])    \n\t"
+            "sb         %[q0_f0],  (%[sq0])    \n\t"
+            "sb         %[q1_f0],  (%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0), [q0_f0] "r" (q0_f0),
+              [q1_f0] "r" (q1_f0), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl        %[p6_r], %[p6_r], 16     \n\t"
+          "srl        %[p5_r], %[p5_r], 16     \n\t"
+          "srl        %[p4_r], %[p4_r], 16     \n\t"
+          "srl        %[p3_r], %[p3_r], 16     \n\t"
+          "srl        %[p2_r], %[p2_r], 16     \n\t"
+          "srl        %[p1_r], %[p1_r], 16     \n\t"
+          "srl        %[p0_r], %[p0_r], 16     \n\t"
+          "srl        %[q0_r], %[q0_r], 16     \n\t"
+          "srl        %[q1_r], %[q1_r], 16     \n\t"
+          "srl        %[q2_r], %[q2_r], 16     \n\t"
+          "srl        %[q3_r], %[q3_r], 16     \n\t"
+          "srl        %[q4_r], %[q4_r], 16     \n\t"
+          "srl        %[q5_r], %[q5_r], 16     \n\t"
+          "srl        %[q6_r], %[q6_r], 16     \n\t"
+
+          : [q0_r] "+r" (q0_r), [q1_r] "+r" (q1_r), [q2_r] "+r" (q2_r),
+            [q3_r] "+r" (q3_r), [q4_r] "+r" (q4_r), [q5_r] "+r" (q5_r),
+            [p6_r] "+r" (p6_r), [p5_r] "+r" (p5_r), [p4_r] "+r" (p4_r),
+            [p3_r] "+r" (p3_r), [p2_r] "+r" (p2_r), [p1_r] "+r" (p1_r),
+            [q6_r] "+r" (q6_r), [p0_r] "+r" (p0_r)
+          :
+      );
+
+      __asm__ __volatile__ (
+          "srl        %[p2_r_f1], %[p2_r_f1], 16     \n\t"
+          "srl        %[p1_r_f1], %[p1_r_f1], 16     \n\t"
+          "srl        %[p0_r_f1], %[p0_r_f1], 16     \n\t"
+          "srl        %[q0_r_f1], %[q0_r_f1], 16     \n\t"
+          "srl        %[q1_r_f1], %[q1_r_f1], 16     \n\t"
+          "srl        %[q2_r_f1], %[q2_r_f1], 16     \n\t"
+          "srl        %[p1_f0],   %[p1_f0],   8      \n\t"
+          "srl        %[p0_f0],   %[p0_f0],   8      \n\t"
+          "srl        %[q0_f0],   %[q0_f0],   8      \n\t"
+          "srl        %[q1_f0],   %[q1_f0],   8      \n\t"
+
+          : [p2_r_f1] "+r" (p2_r_f1), [p1_r_f1] "+r" (p1_r_f1),
+            [p0_r_f1] "+r" (p0_r_f1), [q0_r_f1] "+r" (q0_r_f1),
+            [q1_r_f1] "+r" (q1_r_f1), [q2_r_f1] "+r" (q2_r_f1),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & flat2 & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb         %[p6_r],  +1(%[sp6])    \n\t"
+            "sb         %[p5_r],  +1(%[sp5])    \n\t"
+            "sb         %[p4_r],  +1(%[sp4])    \n\t"
+            "sb         %[p3_r],  +1(%[sp3])    \n\t"
+            "sb         %[p2_r],  +1(%[sp2])    \n\t"
+            "sb         %[p1_r],  +1(%[sp1])    \n\t"
+            "sb         %[p0_r],  +1(%[sp0])    \n\t"
+
+            :
+            : [p6_r] "r" (p6_r), [p5_r] "r" (p5_r), [p4_r] "r" (p4_r),
+              [p3_r] "r" (p3_r), [p2_r] "r" (p2_r), [p1_r] "r" (p1_r),
+              [p0_r] "r" (p0_r), [sp6] "r" (sp6), [sp5] "r" (sp5),
+              [sp4] "r" (sp4), [sp3] "r" (sp3),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0)
+        );
+
+        __asm__ __volatile__ (
+            "sb         %[q0_r],  +1(%[sq0])    \n\t"
+            "sb         %[q1_r],  +1(%[sq1])    \n\t"
+            "sb         %[q2_r],  +1(%[sq2])    \n\t"
+            "sb         %[q3_r],  +1(%[sq3])    \n\t"
+            "sb         %[q4_r],  +1(%[sq4])    \n\t"
+            "sb         %[q5_r],  +1(%[sq5])    \n\t"
+            "sb         %[q6_r],  +1(%[sq6])    \n\t"
+
+            :
+            : [q0_r] "r" (q0_r), [q1_r] "r" (q1_r), [q2_r] "r" (q2_r),
+              [q3_r] "r" (q3_r), [q4_r] "r" (q4_r), [q5_r] "r" (q5_r),
+              [q6_r] "r" (q6_r), [sq0] "r" (sq0), [sq1] "r" (sq1),
+              [sq2] "r" (sq2), [sq3] "r" (sq3),
+              [sq4] "r" (sq4), [sq5] "r" (sq5), [sq6] "r" (sq6)
+        );
+      } else if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb         %[p2_r_f1],  +1(%[sp2])    \n\t"
+            "sb         %[p1_r_f1],  +1(%[sp1])    \n\t"
+            "sb         %[p0_r_f1],  +1(%[sp0])    \n\t"
+            "sb         %[q0_r_f1],  +1(%[sq0])    \n\t"
+            "sb         %[q1_r_f1],  +1(%[sq1])    \n\t"
+            "sb         %[q2_r_f1],  +1(%[sq2])    \n\t"
+
+            :
+            : [p2_r_f1] "r" (p2_r_f1), [p1_r_f1] "r" (p1_r_f1),
+              [p0_r_f1] "r" (p0_r_f1), [q0_r_f1] "r" (q0_r_f1),
+              [q1_r_f1] "r" (q1_r_f1), [q2_r_f1] "r" (q2_r_f1),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  +1(%[sp1])    \n\t"
+            "sb         %[p0_f0],  +1(%[sp0])    \n\t"
+            "sb         %[q0_f0],  +1(%[sq0])    \n\t"
+            "sb         %[q1_f0],  +1(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0), [q0_f0] "r" (q0_f0),
+              [q1_f0] "r" (q1_f0), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl        %[p1_f0], %[p1_f0], 8     \n\t"
+          "srl        %[p0_f0], %[p0_f0], 8     \n\t"
+          "srl        %[q0_f0], %[q0_f0], 8     \n\t"
+          "srl        %[q1_f0], %[q1_f0], 8     \n\t"
+
+          : [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & flat2 & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb         %[p6_l],  +2(%[sp6])    \n\t"
+            "sb         %[p5_l],  +2(%[sp5])    \n\t"
+            "sb         %[p4_l],  +2(%[sp4])    \n\t"
+            "sb         %[p3_l],  +2(%[sp3])    \n\t"
+            "sb         %[p2_l],  +2(%[sp2])    \n\t"
+            "sb         %[p1_l],  +2(%[sp1])    \n\t"
+            "sb         %[p0_l],  +2(%[sp0])    \n\t"
+
+            :
+            : [p6_l] "r" (p6_l), [p5_l] "r" (p5_l), [p4_l] "r" (p4_l),
+              [p3_l] "r" (p3_l), [p2_l] "r" (p2_l), [p1_l] "r" (p1_l),
+              [p0_l] "r" (p0_l), [sp6] "r" (sp6), [sp5] "r" (sp5),
+              [sp4] "r" (sp4), [sp3] "r" (sp3),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0)
+        );
+
+        __asm__ __volatile__ (
+            "sb         %[q0_l],  +2(%[sq0])    \n\t"
+            "sb         %[q1_l],  +2(%[sq1])    \n\t"
+            "sb         %[q2_l],  +2(%[sq2])    \n\t"
+            "sb         %[q3_l],  +2(%[sq3])    \n\t"
+            "sb         %[q4_l],  +2(%[sq4])    \n\t"
+            "sb         %[q5_l],  +2(%[sq5])    \n\t"
+            "sb         %[q6_l],  +2(%[sq6])    \n\t"
+
+            :
+            : [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+              [q3_l] "r" (q3_l), [q4_l] "r" (q4_l), [q5_l] "r" (q5_l),
+              [q6_l] "r" (q6_l), [sq0] "r" (sq0), [sq1] "r" (sq1),
+              [sq2] "r" (sq2), [sq3] "r" (sq3),
+              [sq4] "r" (sq4), [sq5] "r" (sq5), [sq6] "r" (sq6)
+        );
+      } else if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb         %[p2_l_f1],  +2(%[sp2])    \n\t"
+            "sb         %[p1_l_f1],  +2(%[sp1])    \n\t"
+            "sb         %[p0_l_f1],  +2(%[sp0])    \n\t"
+            "sb         %[q0_l_f1],  +2(%[sq0])    \n\t"
+            "sb         %[q1_l_f1],  +2(%[sq1])    \n\t"
+            "sb         %[q2_l_f1],  +2(%[sq2])    \n\t"
+
+            :
+            : [p2_l_f1] "r" (p2_l_f1), [p1_l_f1] "r" (p1_l_f1),
+              [p0_l_f1] "r" (p0_l_f1), [q0_l_f1] "r" (q0_l_f1),
+              [q1_l_f1] "r" (q1_l_f1), [q2_l_f1] "r" (q2_l_f1),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  +2(%[sp1])    \n\t"
+            "sb         %[p0_f0],  +2(%[sp0])    \n\t"
+            "sb         %[q0_f0],  +2(%[sq0])    \n\t"
+            "sb         %[q1_f0],  +2(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0), [q0_f0] "r" (q0_f0),
+              [q1_f0] "r" (q1_f0), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p6_l],    %[p6_l],    16   \n\t"
+          "srl      %[p5_l],    %[p5_l],    16   \n\t"
+          "srl      %[p4_l],    %[p4_l],    16   \n\t"
+          "srl      %[p3_l],    %[p3_l],    16   \n\t"
+          "srl      %[p2_l],    %[p2_l],    16   \n\t"
+          "srl      %[p1_l],    %[p1_l],    16   \n\t"
+          "srl      %[p0_l],    %[p0_l],    16   \n\t"
+          "srl      %[q0_l],    %[q0_l],    16   \n\t"
+          "srl      %[q1_l],    %[q1_l],    16   \n\t"
+          "srl      %[q2_l],    %[q2_l],    16   \n\t"
+          "srl      %[q3_l],    %[q3_l],    16   \n\t"
+          "srl      %[q4_l],    %[q4_l],    16   \n\t"
+          "srl      %[q5_l],    %[q5_l],    16   \n\t"
+          "srl      %[q6_l],    %[q6_l],    16   \n\t"
+
+          : [q0_l] "+r" (q0_l), [q1_l] "+r" (q1_l), [q2_l] "+r" (q2_l),
+            [q3_l] "+r" (q3_l), [q4_l] "+r" (q4_l), [q5_l] "+r" (q5_l),
+            [q6_l] "+r" (q6_l), [p6_l] "+r" (p6_l), [p5_l] "+r" (p5_l),
+            [p4_l] "+r" (p4_l), [p3_l] "+r" (p3_l), [p2_l] "+r" (p2_l),
+            [p1_l] "+r" (p1_l), [p0_l] "+r" (p0_l)
+          :
+      );
+
+      __asm__ __volatile__ (
+          "srl      %[p2_l_f1],   %[p2_l_f1],   16   \n\t"
+          "srl      %[p1_l_f1],   %[p1_l_f1],   16   \n\t"
+          "srl      %[p0_l_f1],   %[p0_l_f1],   16   \n\t"
+          "srl      %[q0_l_f1],   %[q0_l_f1],   16   \n\t"
+          "srl      %[q1_l_f1],   %[q1_l_f1],   16   \n\t"
+          "srl      %[q2_l_f1],   %[q2_l_f1],   16   \n\t"
+          "srl      %[p1_f0],     %[p1_f0],     8    \n\t"
+          "srl      %[p0_f0],     %[p0_f0],     8    \n\t"
+          "srl      %[q0_f0],     %[q0_f0],     8    \n\t"
+          "srl      %[q1_f0],     %[q1_f0],     8    \n\t"
+
+          : [p2_l_f1] "+r" (p2_l_f1), [p1_l_f1] "+r" (p1_l_f1),
+            [p0_l_f1] "+r" (p0_l_f1), [q0_l_f1] "+r" (q0_l_f1),
+            [q1_l_f1] "+r" (q1_l_f1), [q2_l_f1] "+r" (q2_l_f1),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & flat2 & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb     %[p6_l],    +3(%[sp6])    \n\t"
+            "sb     %[p5_l],    +3(%[sp5])    \n\t"
+            "sb     %[p4_l],    +3(%[sp4])    \n\t"
+            "sb     %[p3_l],    +3(%[sp3])    \n\t"
+            "sb     %[p2_l],    +3(%[sp2])    \n\t"
+            "sb     %[p1_l],    +3(%[sp1])    \n\t"
+            "sb     %[p0_l],    +3(%[sp0])    \n\t"
+
+            :
+            : [p6_l] "r" (p6_l), [p5_l] "r" (p5_l), [p4_l] "r" (p4_l),
+              [p3_l] "r" (p3_l), [p2_l] "r" (p2_l), [p1_l] "r" (p1_l),
+              [p0_l] "r" (p0_l), [sp6] "r" (sp6), [sp5] "r" (sp5),
+              [sp4] "r" (sp4), [sp3] "r" (sp3), [sp2] "r" (sp2),
+              [sp1] "r" (sp1), [sp0] "r" (sp0)
+        );
+
+        __asm__ __volatile__ (
+            "sb     %[q0_l],    +3(%[sq0])    \n\t"
+            "sb     %[q1_l],    +3(%[sq1])    \n\t"
+            "sb     %[q2_l],    +3(%[sq2])    \n\t"
+            "sb     %[q3_l],    +3(%[sq3])    \n\t"
+            "sb     %[q4_l],    +3(%[sq4])    \n\t"
+            "sb     %[q5_l],    +3(%[sq5])    \n\t"
+            "sb     %[q6_l],    +3(%[sq6])    \n\t"
+
+            :
+            : [q0_l] "r" (q0_l), [q1_l] "r" (q1_l),
+              [q2_l] "r" (q2_l), [q3_l] "r" (q3_l),
+              [q4_l] "r" (q4_l), [q5_l] "r" (q5_l),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2),
+              [sq3] "r" (sq3), [sq4] "r" (sq4), [sq5] "r" (sq5),
+              [q6_l] "r" (q6_l), [sq6] "r" (sq6)
+        );
+      } else if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb     %[p2_l_f1],     +3(%[sp2])    \n\t"
+            "sb     %[p1_l_f1],     +3(%[sp1])    \n\t"
+            "sb     %[p0_l_f1],     +3(%[sp0])    \n\t"
+            "sb     %[q0_l_f1],     +3(%[sq0])    \n\t"
+            "sb     %[q1_l_f1],     +3(%[sq1])    \n\t"
+            "sb     %[q2_l_f1],     +3(%[sq2])    \n\t"
+
+            :
+            : [p2_l_f1] "r" (p2_l_f1), [p1_l_f1] "r" (p1_l_f1),
+              [p0_l_f1] "r" (p0_l_f1), [q0_l_f1] "r" (q0_l_f1),
+              [q1_l_f1] "r" (q1_l_f1), [q2_l_f1] "r" (q2_l_f1),
+              [sp2] "r" (sp2), [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1), [sq2] "r" (sq2)
+        );
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   +3(%[sp1])    \n\t"
+            "sb     %[p0_f0],   +3(%[sp0])    \n\t"
+            "sb     %[q0_f0],   +3(%[sq0])    \n\t"
+            "sb     %[q1_f0],   +3(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [sp1] "r" (sp1), [sp0] "r" (sp0),
+              [sq0] "r" (sq0), [sq1] "r" (sq1)
+        );
+      }
+    }
+
+    s = s + 4;
+  }
+}
+
+void vpx_lpf_horizontal_edge_8_dspr2(unsigned char *s, int pitch,
+                                     const uint8_t *blimit,
+                                     const uint8_t *limit,
+                                     const uint8_t *thresh) {
+  mb_lpf_horizontal_edge(s, pitch, blimit, limit, thresh, 1);
+}
+
+void vpx_lpf_horizontal_edge_16_dspr2(unsigned char *s, int pitch,
+                                      const uint8_t *blimit,
+                                      const uint8_t *limit,
+                                      const uint8_t *thresh) {
+  mb_lpf_horizontal_edge(s, pitch, blimit, limit, thresh, 2);
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_mb_vert_dspr2.c b/libvpx/libvpx/vpx_dsp/mips/loopfilter_mb_vert_dspr2.c
new file mode 100644
index 0000000..e580f01
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_mb_vert_dspr2.c
@@ -0,0 +1,839 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/mips/common_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_filters_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_macros_dspr2.h"
+#include "vpx_dsp/mips/loopfilter_masks_dspr2.h"
+#include "vpx_mem/vpx_mem.h"
+
+#if HAVE_DSPR2
+void vpx_lpf_vertical_16_dspr2(uint8_t *s,
+                               int pitch,
+                               const uint8_t *blimit,
+                               const uint8_t *limit,
+                               const uint8_t *thresh) {
+  uint8_t   i;
+  uint32_t  mask, hev, flat, flat2;
+  uint8_t   *s1, *s2, *s3, *s4;
+  uint32_t  prim1, prim2, sec3, sec4, prim3, prim4;
+  uint32_t  thresh_vec, flimit_vec, limit_vec;
+  uint32_t  uflimit, ulimit, uthresh;
+  uint32_t  p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  uint32_t  p1_f0, p0_f0, q0_f0, q1_f0;
+  uint32_t  p7_l, p6_l, p5_l, p4_l, p3_l, p2_l, p1_l, p0_l;
+  uint32_t  q0_l, q1_l, q2_l, q3_l, q4_l, q5_l, q6_l, q7_l;
+  uint32_t  p7_r, p6_r, p5_r, p4_r, p3_r, p2_r, p1_r, p0_r;
+  uint32_t  q0_r, q1_r, q2_r, q3_r, q4_r, q5_r, q6_r, q7_r;
+  uint32_t  p2_l_f1, p1_l_f1, p0_l_f1, p2_r_f1, p1_r_f1, p0_r_f1;
+  uint32_t  q0_l_f1, q1_l_f1, q2_l_f1, q0_r_f1, q1_r_f1, q2_r_f1;
+
+  uflimit = *blimit;
+  ulimit = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__ (
+      "replv.qb     %[thresh_vec],     %[uthresh]    \n\t"
+      "replv.qb     %[flimit_vec],     %[uflimit]    \n\t"
+      "replv.qb     %[limit_vec],      %[ulimit]     \n\t"
+
+      : [thresh_vec] "=&r" (thresh_vec), [flimit_vec] "=&r" (flimit_vec),
+        [limit_vec] "=r" (limit_vec)
+      : [uthresh] "r" (uthresh), [uflimit] "r" (uflimit), [ulimit] "r" (ulimit)
+  );
+
+  prefetch_store(s + pitch);
+
+  for (i = 0; i < 2; i++) {
+    s1 = s;
+    s2 = s + pitch;
+    s3 = s2 + pitch;
+    s4 = s3 + pitch;
+    s  = s4 + pitch;
+
+    __asm__ __volatile__ (
+        "lw     %[p0],  -4(%[s1])    \n\t"
+        "lw     %[p1],  -4(%[s2])    \n\t"
+        "lw     %[p2],  -4(%[s3])    \n\t"
+        "lw     %[p3],  -4(%[s4])    \n\t"
+        "lw     %[p4],  -8(%[s1])    \n\t"
+        "lw     %[p5],  -8(%[s2])    \n\t"
+        "lw     %[p6],  -8(%[s3])    \n\t"
+        "lw     %[p7],  -8(%[s4])    \n\t"
+
+        : [p3] "=&r" (p3), [p2] "=&r" (p2), [p1] "=&r" (p1),
+          [p0] "=&r" (p0), [p7] "=&r" (p7), [p6] "=&r" (p6),
+          [p5] "=&r" (p5), [p4] "=&r" (p4)
+        : [s1] "r" (s1), [s2] "r" (s2), [s3] "r" (s3), [s4] "r" (s4)
+    );
+
+    __asm__ __volatile__ (
+        "lw     %[q3],  (%[s1])     \n\t"
+        "lw     %[q2],  (%[s2])     \n\t"
+        "lw     %[q1],  (%[s3])     \n\t"
+        "lw     %[q0],  (%[s4])     \n\t"
+        "lw     %[q7],  +4(%[s1])   \n\t"
+        "lw     %[q6],  +4(%[s2])   \n\t"
+        "lw     %[q5],  +4(%[s3])   \n\t"
+        "lw     %[q4],  +4(%[s4])   \n\t"
+
+        : [q3] "=&r" (q3), [q2] "=&r" (q2), [q1] "=&r" (q1),
+          [q0] "=&r" (q0), [q7] "=&r" (q7), [q6] "=&r" (q6),
+          [q5] "=&r" (q5), [q4] "=&r" (q4)
+        : [s1] "r" (s1), [s2] "r" (s2), [s3] "r" (s3), [s4] "r" (s4)
+    );
+
+    /* transpose p3, p2, p1, p0
+       original (when loaded from memory)
+       register       -4    -3   -2     -1
+         p0         p0_0  p0_1  p0_2  p0_3
+         p1         p1_0  p1_1  p1_2  p1_3
+         p2         p2_0  p2_1  p2_2  p2_3
+         p3         p3_0  p3_1  p3_2  p3_3
+
+       after transpose
+       register
+         p0         p3_3  p2_3  p1_3  p0_3
+         p1         p3_2  p2_2  p1_2  p0_2
+         p2         p3_1  p2_1  p1_1  p0_1
+         p3         p3_0  p2_0  p1_0  p0_0
+    */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p0],      %[p1]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p0],      %[p1]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p2],      %[p3]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p2],      %[p3]       \n\t"
+
+        "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p0],      %[p1],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p2],      %[p3],      %[sec4]     \n\t"
+        "append         %[p1],      %[sec3],    16          \n\t"
+        "append         %[p3],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p0] "+r" (p0), [p1] "+r" (p1), [p2] "+r" (p2), [p3] "+r" (p3),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* transpose q0, q1, q2, q3
+       original (when loaded from memory)
+       register       +1    +2    +3    +4
+         q3         q3_0  q3_1  q3_2  q3_3
+         q2         q2_0  q2_1  q2_2  q2_3
+         q1         q1_0  q1_1  q1_2  q1_3
+         q0         q0_0  q0_1  q0_2  q0_3
+
+       after transpose
+       register
+         q3         q0_3  q1_3  q2_3  q3_3
+         q2         q0_2  q1_2  q2_2  q3_2
+         q1         q0_1  q1_1  q2_1  q3_1
+         q0         q0_0  q1_0  q2_0  q3_0
+    */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[q3],      %[q2]       \n\t"
+        "precr.qb.ph    %[prim2],   %[q3],      %[q2]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[q1],      %[q0]       \n\t"
+        "precr.qb.ph    %[prim4],   %[q1],      %[q0]       \n\t"
+
+        "precrq.qb.ph   %[q2],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[q0],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[q3],      %[q2],      %[sec3]     \n\t"
+        "precrq.ph.w    %[q1],      %[q0],      %[sec4]     \n\t"
+        "append         %[q2],      %[sec3],    16          \n\t"
+        "append         %[q0],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [q3] "+r" (q3), [q2] "+r" (q2), [q1] "+r" (q1), [q0] "+r" (q0),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* transpose p7, p6, p5, p4
+       original (when loaded from memory)
+       register      -8    -7   -6     -5
+         p4         p4_0  p4_1  p4_2  p4_3
+         p5         p5_0  p5_1  p5_2  p5_3
+         p6         p6_0  p6_1  p6_2  p6_3
+         p7         p7_0  p7_1  p7_2  p7_3
+
+       after transpose
+       register
+         p4         p7_3  p6_3  p5_3  p4_3
+         p5         p7_2  p6_2  p5_2  p4_2
+         p6         p7_1  p6_1  p5_1  p4_1
+         p7         p7_0  p6_0  p5_0  p4_0
+    */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[p4],      %[p5]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p4],      %[p5]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p6],      %[p7]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p6],      %[p7]       \n\t"
+
+        "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p7],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p4],      %[p5],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p6],      %[p7],      %[sec4]     \n\t"
+        "append         %[p5],      %[sec3],    16          \n\t"
+        "append         %[p7],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [p4] "+r" (p4), [p5] "+r" (p5), [p6] "+r" (p6), [p7] "+r" (p7),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    /* transpose q4, q5, q6, q7
+       original (when loaded from memory)
+       register      +5    +6    +7    +8
+         q7         q7_0  q7_1  q7_2  q7_3
+         q6         q6_0  q6_1  q6_2  q6_3
+         q5         q5_0  q5_1  q5_2  q5_3
+         q4         q4_0  q4_1  q4_2  q4_3
+
+       after transpose
+       register
+         q7         q4_3  q5_3  q26_3  q7_3
+         q6         q4_2  q5_2  q26_2  q7_2
+         q5         q4_1  q5_1  q26_1  q7_1
+         q4         q4_0  q5_0  q26_0  q7_0
+    */
+    __asm__ __volatile__ (
+        "precrq.qb.ph   %[prim1],   %[q7],      %[q6]       \n\t"
+        "precr.qb.ph    %[prim2],   %[q7],      %[q6]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[q5],      %[q4]       \n\t"
+        "precr.qb.ph    %[prim4],   %[q5],      %[q4]       \n\t"
+
+        "precrq.qb.ph   %[q6],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[q4],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[q7],      %[q6],      %[sec3]     \n\t"
+        "precrq.ph.w    %[q5],      %[q4],      %[sec4]     \n\t"
+        "append         %[q6],      %[sec3],    16          \n\t"
+        "append         %[q4],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r" (prim1), [prim2] "=&r" (prim2),
+          [prim3] "=&r" (prim3), [prim4] "=&r" (prim4),
+          [q7] "+r" (q7), [q6] "+r" (q6), [q5] "+r" (q5), [q4] "+r" (q4),
+          [sec3] "=&r" (sec3), [sec4] "=&r" (sec4)
+        :
+    );
+
+    filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec,
+                                    p1, p0, p3, p2, q0, q1, q2, q3,
+                                    &hev, &mask, &flat);
+
+    flatmask5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, &flat2);
+
+    /* f0 */
+    if (((flat2 == 0) && (flat == 0) && (mask != 0)) ||
+        ((flat2 != 0) && (flat == 0) && (mask != 0))) {
+      filter1_dspr2(mask, hev, p1, p0, q0, q1,
+                    &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+      STORE_F0()
+    } else if ((flat2 == 0XFFFFFFFF) && (flat == 0xFFFFFFFF) &&
+               (mask == 0xFFFFFFFF)) {
+      /* f2 */
+      PACK_LEFT_0TO3()
+      PACK_LEFT_4TO7()
+      wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l,
+                          &p3_l, &p2_l, &p1_l, &p0_l,
+                          &q0_l, &q1_l, &q2_l, &q3_l,
+                          &q4_l, &q5_l, &q6_l, &q7_l);
+
+      PACK_RIGHT_0TO3()
+      PACK_RIGHT_4TO7()
+      wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r,
+                          &p3_r, &p2_r, &p1_r, &p0_r,
+                          &q0_r, &q1_r, &q2_r, &q3_r,
+                          &q4_r, &q5_r, &q6_r, &q7_r);
+
+      STORE_F2()
+    } else if ((flat2 == 0) && (flat == 0xFFFFFFFF) && (mask == 0xFFFFFFFF)) {
+      /* f1 */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l,
+                     &q0_l, &q1_l, &q2_l, &q3_l);
+
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r,
+                     &q0_r, &q1_r, &q2_r, &q3_r);
+
+      STORE_F1()
+    } else if ((flat2 == 0) && (flat != 0) && (mask != 0)) {
+      /* f0 + f1 */
+      filter1_dspr2(mask, hev, p1, p0, q0, q1,
+                    &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l,
+                     &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r,
+                     &q0_r, &q1_r, &q2_r, &q3_r);
+
+      if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb     %[p2_r],    -3(%[s4])    \n\t"
+            "sb     %[p1_r],    -2(%[s4])    \n\t"
+            "sb     %[p0_r],    -1(%[s4])    \n\t"
+            "sb     %[q0_r],      (%[s4])    \n\t"
+            "sb     %[q1_r],    +1(%[s4])    \n\t"
+            "sb     %[q2_r],    +2(%[s4])    \n\t"
+
+            :
+            : [p2_r] "r" (p2_r), [p1_r] "r" (p1_r), [p0_r] "r" (p0_r),
+              [q0_r] "r" (q0_r), [q1_r] "r" (q1_r), [q2_r] "r" (q2_r),
+              [s4] "r" (s4)
+        );
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb         %[p1_f0],  -2(%[s4])    \n\t"
+            "sb         %[p0_f0],  -1(%[s4])    \n\t"
+            "sb         %[q0_f0],    (%[s4])    \n\t"
+            "sb         %[q1_f0],  +1(%[s4])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s4] "r" (s4)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p2_r],    %[p2_r],    16      \n\t"
+          "srl      %[p1_r],    %[p1_r],    16      \n\t"
+          "srl      %[p0_r],    %[p0_r],    16      \n\t"
+          "srl      %[q0_r],    %[q0_r],    16      \n\t"
+          "srl      %[q1_r],    %[q1_r],    16      \n\t"
+          "srl      %[q2_r],    %[q2_r],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_r] "+r" (p2_r), [p1_r] "+r" (p1_r), [p0_r] "+r" (p0_r),
+            [q0_r] "+r" (q0_r), [q1_r] "+r" (q1_r), [q2_r] "+r" (q2_r),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb     %[p2_r],    -3(%[s3])    \n\t"
+            "sb     %[p1_r],    -2(%[s3])    \n\t"
+            "sb     %[p0_r],    -1(%[s3])    \n\t"
+            "sb     %[q0_r],      (%[s3])    \n\t"
+            "sb     %[q1_r],    +1(%[s3])    \n\t"
+            "sb     %[q2_r],    +2(%[s3])    \n\t"
+
+            :
+            : [p2_r] "r" (p2_r), [p1_r] "r" (p1_r), [p0_r] "r" (p0_r),
+              [q0_r] "r" (q0_r), [q1_r] "r" (q1_r), [q2_r] "r" (q2_r),
+              [s3] "r" (s3)
+        );
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   -2(%[s3])    \n\t"
+            "sb     %[p0_f0],   -1(%[s3])    \n\t"
+            "sb     %[q0_f0],     (%[s3])    \n\t"
+            "sb     %[q1_f0],   +1(%[s3])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s3] "r" (s3)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p1_f0],   %[p1_f0],   8     \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8     \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8     \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8     \n\t"
+
+          : [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__ (
+          "sb       %[p2_l],    -3(%[s2])    \n\t"
+          "sb       %[p1_l],    -2(%[s2])    \n\t"
+          "sb       %[p0_l],    -1(%[s2])    \n\t"
+          "sb       %[q0_l],      (%[s2])    \n\t"
+          "sb       %[q1_l],    +1(%[s2])    \n\t"
+          "sb       %[q2_l],    +2(%[s2])    \n\t"
+
+          :
+          : [p2_l] "r" (p2_l), [p1_l] "r" (p1_l), [p0_l] "r" (p0_l),
+            [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+            [s2] "r" (s2)
+        );
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   -2(%[s2])    \n\t"
+            "sb     %[p0_f0],   -1(%[s2])    \n\t"
+            "sb     %[q0_f0],     (%[s2])    \n\t"
+            "sb     %[q1_f0],   +1(%[s2])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s2] "r" (s2)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p2_l],    %[p2_l],    16      \n\t"
+          "srl      %[p1_l],    %[p1_l],    16      \n\t"
+          "srl      %[p0_l],    %[p0_l],    16      \n\t"
+          "srl      %[q0_l],    %[q0_l],    16      \n\t"
+          "srl      %[q1_l],    %[q1_l],    16      \n\t"
+          "srl      %[q2_l],    %[q2_l],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_l] "+r" (p2_l), [p1_l] "+r" (p1_l), [p0_l] "+r" (p0_l),
+            [q0_l] "+r" (q0_l), [q1_l] "+r" (q1_l), [q2_l] "+r" (q2_l),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb     %[p2_l],    -3(%[s1])    \n\t"
+            "sb     %[p1_l],    -2(%[s1])    \n\t"
+            "sb     %[p0_l],    -1(%[s1])    \n\t"
+            "sb     %[q0_l],      (%[s1])    \n\t"
+            "sb     %[q1_l],    +1(%[s1])    \n\t"
+            "sb     %[q2_l],    +2(%[s1])    \n\t"
+
+            :
+            : [p2_l] "r" (p2_l), [p1_l] "r" (p1_l), [p0_l] "r" (p0_l),
+              [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+              [s1] "r" (s1)
+        );
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   -2(%[s1])    \n\t"
+            "sb     %[p0_f0],   -1(%[s1])    \n\t"
+            "sb     %[q0_f0],     (%[s1])    \n\t"
+            "sb     %[q1_f0],   +1(%[s1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s1] "r" (s1)
+        );
+      }
+    } else if ((flat2 != 0) && (flat != 0) && (mask != 0)) {
+      /* f0+f1+f2 */
+      filter1_dspr2(mask, hev, p1, p0, q0, q1,
+                    &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      PACK_LEFT_0TO3()
+      mbfilter1_dspr2(p3_l, p2_l, p1_l, p0_l,
+                      q0_l, q1_l, q2_l, q3_l,
+                      &p2_l_f1, &p1_l_f1, &p0_l_f1,
+                      &q0_l_f1, &q1_l_f1, &q2_l_f1);
+
+      PACK_RIGHT_0TO3()
+      mbfilter1_dspr2(p3_r, p2_r, p1_r, p0_r,
+                      q0_r, q1_r, q2_r, q3_r,
+                      &p2_r_f1, &p1_r_f1, &p0_r_f1,
+                      &q0_r_f1, &q1_r_f1, &q2_r_f1);
+
+      PACK_LEFT_4TO7()
+      wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l,
+                          &p3_l, &p2_l, &p1_l, &p0_l,
+                          &q0_l, &q1_l, &q2_l, &q3_l,
+                          &q4_l, &q5_l, &q6_l, &q7_l);
+
+      PACK_RIGHT_4TO7()
+      wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r,
+                          &p3_r, &p2_r, &p1_r, &p0_r,
+                          &q0_r, &q1_r, &q2_r, &q3_r,
+                          &q4_r, &q5_r, &q6_r, &q7_r);
+
+      if (mask & flat & flat2 & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb     %[p6_r],    -7(%[s4])    \n\t"
+            "sb     %[p5_r],    -6(%[s4])    \n\t"
+            "sb     %[p4_r],    -5(%[s4])    \n\t"
+            "sb     %[p3_r],    -4(%[s4])    \n\t"
+            "sb     %[p2_r],    -3(%[s4])    \n\t"
+            "sb     %[p1_r],    -2(%[s4])    \n\t"
+            "sb     %[p0_r],    -1(%[s4])    \n\t"
+
+            :
+            : [p6_r] "r" (p6_r), [p5_r] "r" (p5_r),
+              [p4_r] "r" (p4_r), [p3_r] "r" (p3_r),
+              [p2_r] "r" (p2_r), [p1_r] "r" (p1_r),
+              [p0_r] "r" (p0_r), [s4] "r" (s4)
+        );
+
+        __asm__ __volatile__ (
+            "sb     %[q0_r],      (%[s4])    \n\t"
+            "sb     %[q1_r],    +1(%[s4])    \n\t"
+            "sb     %[q2_r],    +2(%[s4])    \n\t"
+            "sb     %[q3_r],    +3(%[s4])    \n\t"
+            "sb     %[q4_r],    +4(%[s4])    \n\t"
+            "sb     %[q5_r],    +5(%[s4])    \n\t"
+            "sb     %[q6_r],    +6(%[s4])    \n\t"
+
+            :
+            : [q0_r] "r" (q0_r), [q1_r] "r" (q1_r),
+              [q2_r] "r" (q2_r), [q3_r] "r" (q3_r),
+              [q4_r] "r" (q4_r), [q5_r] "r" (q5_r),
+              [q6_r] "r" (q6_r), [s4] "r" (s4)
+        );
+      } else if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb     %[p2_r_f1],     -3(%[s4])    \n\t"
+            "sb     %[p1_r_f1],     -2(%[s4])    \n\t"
+            "sb     %[p0_r_f1],     -1(%[s4])    \n\t"
+            "sb     %[q0_r_f1],       (%[s4])    \n\t"
+            "sb     %[q1_r_f1],     +1(%[s4])    \n\t"
+            "sb     %[q2_r_f1],     +2(%[s4])    \n\t"
+
+            :
+            : [p2_r_f1] "r" (p2_r_f1), [p1_r_f1] "r" (p1_r_f1),
+              [p0_r_f1] "r" (p0_r_f1), [q0_r_f1] "r" (q0_r_f1),
+              [q1_r_f1] "r" (q1_r_f1), [q2_r_f1] "r" (q2_r_f1),
+              [s4] "r" (s4)
+        );
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   -2(%[s4])    \n\t"
+            "sb     %[p0_f0],   -1(%[s4])    \n\t"
+            "sb     %[q0_f0],     (%[s4])    \n\t"
+            "sb     %[q1_f0],   +1(%[s4])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s4] "r" (s4)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p6_r],        %[p6_r],        16     \n\t"
+          "srl      %[p5_r],        %[p5_r],        16     \n\t"
+          "srl      %[p4_r],        %[p4_r],        16     \n\t"
+          "srl      %[p3_r],        %[p3_r],        16     \n\t"
+          "srl      %[p2_r],        %[p2_r],        16     \n\t"
+          "srl      %[p1_r],        %[p1_r],        16     \n\t"
+          "srl      %[p0_r],        %[p0_r],        16     \n\t"
+          "srl      %[q0_r],        %[q0_r],        16     \n\t"
+          "srl      %[q1_r],        %[q1_r],        16     \n\t"
+          "srl      %[q2_r],        %[q2_r],        16     \n\t"
+          "srl      %[q3_r],        %[q3_r],        16     \n\t"
+          "srl      %[q4_r],        %[q4_r],        16     \n\t"
+          "srl      %[q5_r],        %[q5_r],        16     \n\t"
+          "srl      %[q6_r],        %[q6_r],        16     \n\t"
+
+          : [q0_r] "+r" (q0_r), [q1_r] "+r" (q1_r),
+            [q2_r] "+r" (q2_r), [q3_r] "+r" (q3_r),
+            [q4_r] "+r" (q4_r), [q5_r] "+r" (q5_r),
+            [q6_r] "+r" (q6_r), [p6_r] "+r" (p6_r),
+            [p5_r] "+r" (p5_r), [p4_r] "+r" (p4_r),
+            [p3_r] "+r" (p3_r), [p2_r] "+r" (p2_r),
+            [p1_r] "+r" (p1_r), [p0_r] "+r" (p0_r)
+          :
+      );
+
+      __asm__ __volatile__ (
+          "srl      %[p2_r_f1],     %[p2_r_f1],     16      \n\t"
+          "srl      %[p1_r_f1],     %[p1_r_f1],     16      \n\t"
+          "srl      %[p0_r_f1],     %[p0_r_f1],     16      \n\t"
+          "srl      %[q0_r_f1],     %[q0_r_f1],     16      \n\t"
+          "srl      %[q1_r_f1],     %[q1_r_f1],     16      \n\t"
+          "srl      %[q2_r_f1],     %[q2_r_f1],     16      \n\t"
+          "srl      %[p1_f0],       %[p1_f0],       8       \n\t"
+          "srl      %[p0_f0],       %[p0_f0],       8       \n\t"
+          "srl      %[q0_f0],       %[q0_f0],       8       \n\t"
+          "srl      %[q1_f0],       %[q1_f0],       8       \n\t"
+
+          : [p2_r_f1] "+r" (p2_r_f1), [p1_r_f1] "+r" (p1_r_f1),
+            [p0_r_f1] "+r" (p0_r_f1), [q0_r_f1] "+r" (q0_r_f1),
+            [q1_r_f1] "+r" (q1_r_f1), [q2_r_f1] "+r" (q2_r_f1),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & flat2 & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb     %[p6_r],    -7(%[s3])    \n\t"
+            "sb     %[p5_r],    -6(%[s3])    \n\t"
+            "sb     %[p4_r],    -5(%[s3])    \n\t"
+            "sb     %[p3_r],    -4(%[s3])    \n\t"
+            "sb     %[p2_r],    -3(%[s3])    \n\t"
+            "sb     %[p1_r],    -2(%[s3])    \n\t"
+            "sb     %[p0_r],    -1(%[s3])    \n\t"
+
+            :
+            : [p6_r] "r" (p6_r), [p5_r] "r" (p5_r), [p4_r] "r" (p4_r),
+              [p3_r] "r" (p3_r), [p2_r] "r" (p2_r), [p1_r] "r" (p1_r),
+              [p0_r] "r" (p0_r), [s3] "r" (s3)
+        );
+
+        __asm__ __volatile__ (
+            "sb     %[q0_r],      (%[s3])    \n\t"
+            "sb     %[q1_r],    +1(%[s3])    \n\t"
+            "sb     %[q2_r],    +2(%[s3])    \n\t"
+            "sb     %[q3_r],    +3(%[s3])    \n\t"
+            "sb     %[q4_r],    +4(%[s3])    \n\t"
+            "sb     %[q5_r],    +5(%[s3])    \n\t"
+            "sb     %[q6_r],    +6(%[s3])    \n\t"
+
+            :
+            : [q0_r] "r" (q0_r), [q1_r] "r" (q1_r),
+              [q2_r] "r" (q2_r), [q3_r] "r" (q3_r),
+              [q4_r] "r" (q4_r), [q5_r] "r" (q5_r),
+              [q6_r] "r" (q6_r), [s3] "r" (s3)
+        );
+      } else if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb     %[p2_r_f1],     -3(%[s3])    \n\t"
+            "sb     %[p1_r_f1],     -2(%[s3])    \n\t"
+            "sb     %[p0_r_f1],     -1(%[s3])    \n\t"
+            "sb     %[q0_r_f1],       (%[s3])    \n\t"
+            "sb     %[q1_r_f1],     +1(%[s3])    \n\t"
+            "sb     %[q2_r_f1],     +2(%[s3])    \n\t"
+
+            :
+            : [p2_r_f1] "r" (p2_r_f1), [p1_r_f1] "r" (p1_r_f1),
+              [p0_r_f1] "r" (p0_r_f1), [q0_r_f1] "r" (q0_r_f1),
+              [q1_r_f1] "r" (q1_r_f1), [q2_r_f1] "r" (q2_r_f1),
+              [s3] "r" (s3)
+        );
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   -2(%[s3])    \n\t"
+            "sb     %[p0_f0],   -1(%[s3])    \n\t"
+            "sb     %[q0_f0],     (%[s3])    \n\t"
+            "sb     %[q1_f0],   +1(%[s3])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s3] "r" (s3)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p1_f0],   %[p1_f0],   8     \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8     \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8     \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8     \n\t"
+
+          : [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & flat2 & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb     %[p6_l],    -7(%[s2])    \n\t"
+            "sb     %[p5_l],    -6(%[s2])    \n\t"
+            "sb     %[p4_l],    -5(%[s2])    \n\t"
+            "sb     %[p3_l],    -4(%[s2])    \n\t"
+            "sb     %[p2_l],    -3(%[s2])    \n\t"
+            "sb     %[p1_l],    -2(%[s2])    \n\t"
+            "sb     %[p0_l],    -1(%[s2])    \n\t"
+
+            :
+            : [p6_l] "r" (p6_l), [p5_l] "r" (p5_l), [p4_l] "r" (p4_l),
+              [p3_l] "r" (p3_l), [p2_l] "r" (p2_l), [p1_l] "r" (p1_l),
+              [p0_l] "r" (p0_l), [s2] "r" (s2)
+        );
+
+        __asm__ __volatile__ (
+            "sb     %[q0_l],      (%[s2])    \n\t"
+            "sb     %[q1_l],    +1(%[s2])    \n\t"
+            "sb     %[q2_l],    +2(%[s2])    \n\t"
+            "sb     %[q3_l],    +3(%[s2])    \n\t"
+            "sb     %[q4_l],    +4(%[s2])    \n\t"
+            "sb     %[q5_l],    +5(%[s2])    \n\t"
+            "sb     %[q6_l],    +6(%[s2])    \n\t"
+
+            :
+            : [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+              [q3_l] "r" (q3_l), [q4_l] "r" (q4_l), [q5_l] "r" (q5_l),
+              [q6_l] "r" (q6_l), [s2] "r" (s2)
+        );
+      } else if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb     %[p2_l_f1],     -3(%[s2])    \n\t"
+            "sb     %[p1_l_f1],     -2(%[s2])    \n\t"
+            "sb     %[p0_l_f1],     -1(%[s2])    \n\t"
+            "sb     %[q0_l_f1],       (%[s2])    \n\t"
+            "sb     %[q1_l_f1],     +1(%[s2])    \n\t"
+            "sb     %[q2_l_f1],     +2(%[s2])    \n\t"
+
+            :
+            : [p2_l_f1] "r" (p2_l_f1), [p1_l_f1] "r" (p1_l_f1),
+              [p0_l_f1] "r" (p0_l_f1), [q0_l_f1] "r" (q0_l_f1),
+              [q1_l_f1] "r" (q1_l_f1), [q2_l_f1] "r" (q2_l_f1),
+              [s2] "r" (s2)
+        );
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   -2(%[s2])    \n\t"
+            "sb     %[p0_f0],   -1(%[s2])    \n\t"
+            "sb     %[q0_f0],     (%[s2])    \n\t"
+            "sb     %[q1_f0],   +1(%[s2])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s2] "r" (s2)
+        );
+      }
+
+      __asm__ __volatile__ (
+          "srl      %[p6_l],        %[p6_l],        16     \n\t"
+          "srl      %[p5_l],        %[p5_l],        16     \n\t"
+          "srl      %[p4_l],        %[p4_l],        16     \n\t"
+          "srl      %[p3_l],        %[p3_l],        16     \n\t"
+          "srl      %[p2_l],        %[p2_l],        16     \n\t"
+          "srl      %[p1_l],        %[p1_l],        16     \n\t"
+          "srl      %[p0_l],        %[p0_l],        16     \n\t"
+          "srl      %[q0_l],        %[q0_l],        16     \n\t"
+          "srl      %[q1_l],        %[q1_l],        16     \n\t"
+          "srl      %[q2_l],        %[q2_l],        16     \n\t"
+          "srl      %[q3_l],        %[q3_l],        16     \n\t"
+          "srl      %[q4_l],        %[q4_l],        16     \n\t"
+          "srl      %[q5_l],        %[q5_l],        16     \n\t"
+          "srl      %[q6_l],        %[q6_l],        16     \n\t"
+
+          : [q0_l] "+r" (q0_l), [q1_l] "+r" (q1_l), [q2_l] "+r" (q2_l),
+            [q3_l] "+r" (q3_l), [q4_l] "+r" (q4_l), [q5_l] "+r" (q5_l),
+            [q6_l] "+r" (q6_l), [p6_l] "+r" (p6_l), [p5_l] "+r" (p5_l),
+            [p4_l] "+r" (p4_l), [p3_l] "+r" (p3_l), [p2_l] "+r" (p2_l),
+            [p1_l] "+r" (p1_l), [p0_l] "+r" (p0_l)
+          :
+      );
+
+      __asm__ __volatile__ (
+          "srl      %[p2_l_f1],     %[p2_l_f1],     16      \n\t"
+          "srl      %[p1_l_f1],     %[p1_l_f1],     16      \n\t"
+          "srl      %[p0_l_f1],     %[p0_l_f1],     16      \n\t"
+          "srl      %[q0_l_f1],     %[q0_l_f1],     16      \n\t"
+          "srl      %[q1_l_f1],     %[q1_l_f1],     16      \n\t"
+          "srl      %[q2_l_f1],     %[q2_l_f1],     16      \n\t"
+          "srl      %[p1_f0],       %[p1_f0],       8       \n\t"
+          "srl      %[p0_f0],       %[p0_f0],       8       \n\t"
+          "srl      %[q0_f0],       %[q0_f0],       8       \n\t"
+          "srl      %[q1_f0],       %[q1_f0],       8       \n\t"
+
+          : [p2_l_f1] "+r" (p2_l_f1), [p1_l_f1] "+r" (p1_l_f1),
+            [p0_l_f1] "+r" (p0_l_f1), [q0_l_f1] "+r" (q0_l_f1),
+            [q1_l_f1] "+r" (q1_l_f1), [q2_l_f1] "+r" (q2_l_f1),
+            [p1_f0] "+r" (p1_f0), [p0_f0] "+r" (p0_f0),
+            [q0_f0] "+r" (q0_f0), [q1_f0] "+r" (q1_f0)
+          :
+      );
+
+      if (mask & flat & flat2 & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb     %[p6_l],    -7(%[s1])    \n\t"
+            "sb     %[p5_l],    -6(%[s1])    \n\t"
+            "sb     %[p4_l],    -5(%[s1])    \n\t"
+            "sb     %[p3_l],    -4(%[s1])    \n\t"
+            "sb     %[p2_l],    -3(%[s1])    \n\t"
+            "sb     %[p1_l],    -2(%[s1])    \n\t"
+            "sb     %[p0_l],    -1(%[s1])    \n\t"
+
+            :
+            : [p6_l] "r" (p6_l), [p5_l] "r" (p5_l), [p4_l] "r" (p4_l),
+              [p3_l] "r" (p3_l), [p2_l] "r" (p2_l), [p1_l] "r" (p1_l),
+              [p0_l] "r" (p0_l),
+              [s1] "r" (s1)
+        );
+
+        __asm__ __volatile__ (
+            "sb     %[q0_l],     (%[s1])    \n\t"
+            "sb     %[q1_l],    1(%[s1])    \n\t"
+            "sb     %[q2_l],    2(%[s1])    \n\t"
+            "sb     %[q3_l],    3(%[s1])    \n\t"
+            "sb     %[q4_l],    4(%[s1])    \n\t"
+            "sb     %[q5_l],    5(%[s1])    \n\t"
+            "sb     %[q6_l],    6(%[s1])    \n\t"
+
+            :
+            : [q0_l] "r" (q0_l), [q1_l] "r" (q1_l), [q2_l] "r" (q2_l),
+              [q3_l] "r" (q3_l), [q4_l] "r" (q4_l), [q5_l] "r" (q5_l),
+              [q6_l] "r" (q6_l),
+              [s1] "r" (s1)
+        );
+      } else if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb     %[p2_l_f1],     -3(%[s1])    \n\t"
+            "sb     %[p1_l_f1],     -2(%[s1])    \n\t"
+            "sb     %[p0_l_f1],     -1(%[s1])    \n\t"
+            "sb     %[q0_l_f1],       (%[s1])    \n\t"
+            "sb     %[q1_l_f1],     +1(%[s1])    \n\t"
+            "sb     %[q2_l_f1],     +2(%[s1])    \n\t"
+
+            :
+            : [p2_l_f1] "r" (p2_l_f1), [p1_l_f1] "r" (p1_l_f1),
+              [p0_l_f1] "r" (p0_l_f1), [q0_l_f1] "r" (q0_l_f1),
+              [q1_l_f1] "r" (q1_l_f1), [q2_l_f1] "r" (q2_l_f1),
+              [s1] "r" (s1)
+        );
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__ (
+            "sb     %[p1_f0],   -2(%[s1])    \n\t"
+            "sb     %[p0_f0],   -1(%[s1])    \n\t"
+            "sb     %[q0_f0],     (%[s1])    \n\t"
+            "sb     %[q1_f0],   +1(%[s1])    \n\t"
+
+            :
+            : [p1_f0] "r" (p1_f0), [p0_f0] "r" (p0_f0),
+              [q0_f0] "r" (q0_f0), [q1_f0] "r" (q1_f0),
+              [s1] "r" (s1)
+        );
+      }
+    }
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/libvpx/libvpx/vpx_dsp/mips/loopfilter_msa.h b/libvpx/libvpx/vpx_dsp/mips/loopfilter_msa.h
new file mode 100644
index 0000000..62b1706
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/loopfilter_msa.h
@@ -0,0 +1,246 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_LOOPFILTER_MSA_H_
+#define VPX_DSP_LOOPFILTER_MSA_H_
+
+#include "vpx_dsp/mips/macros_msa.h"
+
+#define VP9_LPF_FILTER4_8W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in,  \
+                           p1_out, p0_out, q0_out, q1_out) {             \
+  v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign;                    \
+  v16i8 filt, filt1, filt2, cnst4b, cnst3b;                              \
+  v8i16 q0_sub_p0_r, filt_r, cnst3h;                                     \
+                                                                         \
+  p1_m = (v16i8)__msa_xori_b(p1_in, 0x80);                               \
+  p0_m = (v16i8)__msa_xori_b(p0_in, 0x80);                               \
+  q0_m = (v16i8)__msa_xori_b(q0_in, 0x80);                               \
+  q1_m = (v16i8)__msa_xori_b(q1_in, 0x80);                               \
+                                                                         \
+  filt = __msa_subs_s_b(p1_m, q1_m);                                     \
+  filt = filt & (v16i8)hev_in;                                           \
+  q0_sub_p0 = q0_m - p0_m;                                               \
+  filt_sign = __msa_clti_s_b(filt, 0);                                   \
+                                                                         \
+  cnst3h = __msa_ldi_h(3);                                               \
+  q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0);               \
+  q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h);       \
+  filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt);                         \
+  filt_r += q0_sub_p0_r;                                                 \
+  filt_r = __msa_sat_s_h(filt_r, 7);                                     \
+                                                                         \
+  /* combine left and right part */                                      \
+  filt = __msa_pckev_b((v16i8)filt_r, (v16i8)filt_r);                    \
+                                                                         \
+  filt = filt & (v16i8)mask_in;                                          \
+  cnst4b = __msa_ldi_b(4);                                               \
+  filt1 = __msa_adds_s_b(filt, cnst4b);                                  \
+  filt1 >>= 3;                                                           \
+                                                                         \
+  cnst3b = __msa_ldi_b(3);                                               \
+  filt2 = __msa_adds_s_b(filt, cnst3b);                                  \
+  filt2 >>= 3;                                                           \
+                                                                         \
+  q0_m = __msa_subs_s_b(q0_m, filt1);                                    \
+  q0_out = __msa_xori_b((v16u8)q0_m, 0x80);                              \
+  p0_m = __msa_adds_s_b(p0_m, filt2);                                    \
+  p0_out = __msa_xori_b((v16u8)p0_m, 0x80);                              \
+                                                                         \
+  filt = __msa_srari_b(filt1, 1);                                        \
+  hev_in = __msa_xori_b((v16u8)hev_in, 0xff);                            \
+  filt = filt & (v16i8)hev_in;                                           \
+                                                                         \
+  q1_m = __msa_subs_s_b(q1_m, filt);                                     \
+  q1_out = __msa_xori_b((v16u8)q1_m, 0x80);                              \
+  p1_m = __msa_adds_s_b(p1_m, filt);                                     \
+  p1_out = __msa_xori_b((v16u8)p1_m, 0x80);                              \
+}
+
+#define VP9_LPF_FILTER4_4W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in,  \
+                           p1_out, p0_out, q0_out, q1_out) {             \
+  v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign;                    \
+  v16i8 filt, filt1, filt2, cnst4b, cnst3b;                              \
+  v8i16 q0_sub_p0_r, q0_sub_p0_l, filt_l, filt_r, cnst3h;                \
+                                                                         \
+  p1_m = (v16i8)__msa_xori_b(p1_in, 0x80);                               \
+  p0_m = (v16i8)__msa_xori_b(p0_in, 0x80);                               \
+  q0_m = (v16i8)__msa_xori_b(q0_in, 0x80);                               \
+  q1_m = (v16i8)__msa_xori_b(q1_in, 0x80);                               \
+                                                                         \
+  filt = __msa_subs_s_b(p1_m, q1_m);                                     \
+                                                                         \
+  filt = filt & (v16i8)hev_in;                                           \
+                                                                         \
+  q0_sub_p0 = q0_m - p0_m;                                               \
+  filt_sign = __msa_clti_s_b(filt, 0);                                   \
+                                                                         \
+  cnst3h = __msa_ldi_h(3);                                               \
+  q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0);               \
+  q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h);       \
+  filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt);                         \
+  filt_r += q0_sub_p0_r;                                                 \
+  filt_r = __msa_sat_s_h(filt_r, 7);                                     \
+                                                                         \
+  q0_sub_p0_l = (v8i16)__msa_ilvl_b(q0_sub_p0, q0_sub_p0);               \
+  q0_sub_p0_l = __msa_dotp_s_h((v16i8)q0_sub_p0_l, (v16i8)cnst3h);       \
+  filt_l = (v8i16)__msa_ilvl_b(filt_sign, filt);                         \
+  filt_l += q0_sub_p0_l;                                                 \
+  filt_l = __msa_sat_s_h(filt_l, 7);                                     \
+                                                                         \
+  filt = __msa_pckev_b((v16i8)filt_l, (v16i8)filt_r);                    \
+  filt = filt & (v16i8)mask_in;                                          \
+                                                                         \
+  cnst4b = __msa_ldi_b(4);                                               \
+  filt1 = __msa_adds_s_b(filt, cnst4b);                                  \
+  filt1 >>= 3;                                                           \
+                                                                         \
+  cnst3b = __msa_ldi_b(3);                                               \
+  filt2 = __msa_adds_s_b(filt, cnst3b);                                  \
+  filt2 >>= 3;                                                           \
+                                                                         \
+  q0_m = __msa_subs_s_b(q0_m, filt1);                                    \
+  q0_out = __msa_xori_b((v16u8)q0_m, 0x80);                              \
+  p0_m = __msa_adds_s_b(p0_m, filt2);                                    \
+  p0_out = __msa_xori_b((v16u8)p0_m, 0x80);                              \
+                                                                         \
+  filt = __msa_srari_b(filt1, 1);                                        \
+  hev_in = __msa_xori_b((v16u8)hev_in, 0xff);                            \
+  filt = filt & (v16i8)hev_in;                                           \
+                                                                         \
+  q1_m = __msa_subs_s_b(q1_m, filt);                                     \
+  q1_out = __msa_xori_b((v16u8)q1_m, 0x80);                              \
+  p1_m = __msa_adds_s_b(p1_m, filt);                                     \
+  p1_out = __msa_xori_b((v16u8)p1_m, 0x80);                              \
+}
+
+#define VP9_FLAT4(p3_in, p2_in, p0_in, q0_in, q2_in, q3_in, flat_out) {  \
+  v16u8 tmp, p2_a_sub_p0, q2_a_sub_q0, p3_a_sub_p0, q3_a_sub_q0;         \
+  v16u8 zero_in = { 0 };                                                 \
+                                                                         \
+  tmp = __msa_ori_b(zero_in, 1);                                         \
+  p2_a_sub_p0 = __msa_asub_u_b(p2_in, p0_in);                            \
+  q2_a_sub_q0 = __msa_asub_u_b(q2_in, q0_in);                            \
+  p3_a_sub_p0 = __msa_asub_u_b(p3_in, p0_in);                            \
+  q3_a_sub_q0 = __msa_asub_u_b(q3_in, q0_in);                            \
+                                                                         \
+  p2_a_sub_p0 = __msa_max_u_b(p2_a_sub_p0, q2_a_sub_q0);                 \
+  flat_out = __msa_max_u_b(p2_a_sub_p0, flat_out);                       \
+  p3_a_sub_p0 = __msa_max_u_b(p3_a_sub_p0, q3_a_sub_q0);                 \
+  flat_out = __msa_max_u_b(p3_a_sub_p0, flat_out);                       \
+                                                                         \
+  flat_out = (tmp < (v16u8)flat_out);                                    \
+  flat_out = __msa_xori_b(flat_out, 0xff);                               \
+  flat_out = flat_out & (mask);                                          \
+}
+
+#define VP9_FLAT5(p7_in, p6_in, p5_in, p4_in, p0_in, q0_in, q4_in,  \
+                  q5_in, q6_in, q7_in, flat_in, flat2_out) {        \
+  v16u8 tmp, zero_in = { 0 };                                       \
+  v16u8 p4_a_sub_p0, q4_a_sub_q0, p5_a_sub_p0, q5_a_sub_q0;         \
+  v16u8 p6_a_sub_p0, q6_a_sub_q0, p7_a_sub_p0, q7_a_sub_q0;         \
+                                                                    \
+  tmp = __msa_ori_b(zero_in, 1);                                    \
+  p4_a_sub_p0 = __msa_asub_u_b(p4_in, p0_in);                       \
+  q4_a_sub_q0 = __msa_asub_u_b(q4_in, q0_in);                       \
+  p5_a_sub_p0 = __msa_asub_u_b(p5_in, p0_in);                       \
+  q5_a_sub_q0 = __msa_asub_u_b(q5_in, q0_in);                       \
+  p6_a_sub_p0 = __msa_asub_u_b(p6_in, p0_in);                       \
+  q6_a_sub_q0 = __msa_asub_u_b(q6_in, q0_in);                       \
+  p7_a_sub_p0 = __msa_asub_u_b(p7_in, p0_in);                       \
+  q7_a_sub_q0 = __msa_asub_u_b(q7_in, q0_in);                       \
+                                                                    \
+  p4_a_sub_p0 = __msa_max_u_b(p4_a_sub_p0, q4_a_sub_q0);            \
+  flat2_out = __msa_max_u_b(p5_a_sub_p0, q5_a_sub_q0);              \
+  flat2_out = __msa_max_u_b(p4_a_sub_p0, flat2_out);                \
+  p6_a_sub_p0 = __msa_max_u_b(p6_a_sub_p0, q6_a_sub_q0);            \
+  flat2_out = __msa_max_u_b(p6_a_sub_p0, flat2_out);                \
+  p7_a_sub_p0 = __msa_max_u_b(p7_a_sub_p0, q7_a_sub_q0);            \
+  flat2_out = __msa_max_u_b(p7_a_sub_p0, flat2_out);                \
+                                                                    \
+  flat2_out = (tmp < (v16u8)flat2_out);                             \
+  flat2_out = __msa_xori_b(flat2_out, 0xff);                        \
+  flat2_out = flat2_out & flat_in;                                  \
+}
+
+#define VP9_FILTER8(p3_in, p2_in, p1_in, p0_in,                  \
+                    q0_in, q1_in, q2_in, q3_in,                  \
+                    p2_filt8_out, p1_filt8_out, p0_filt8_out,    \
+                    q0_filt8_out, q1_filt8_out, q2_filt8_out) {  \
+  v8u16 tmp0, tmp1, tmp2;                                        \
+                                                                 \
+  tmp2 = p2_in + p1_in + p0_in;                                  \
+  tmp0 = p3_in << 1;                                             \
+                                                                 \
+  tmp0 = tmp0 + tmp2 + q0_in;                                    \
+  tmp1 = tmp0 + p3_in + p2_in;                                   \
+  p2_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3);           \
+                                                                 \
+  tmp1 = tmp0 + p1_in + q1_in;                                   \
+  p1_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3);           \
+                                                                 \
+  tmp1 = q2_in + q1_in + q0_in;                                  \
+  tmp2 = tmp2 + tmp1;                                            \
+  tmp0 = tmp2 + (p0_in);                                         \
+  tmp0 = tmp0 + (p3_in);                                         \
+  p0_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp0, 3);           \
+                                                                 \
+  tmp0 = q2_in + q3_in;                                          \
+  tmp0 = p0_in + tmp1 + tmp0;                                    \
+  tmp1 = q3_in + q3_in;                                          \
+  tmp1 = tmp1 + tmp0;                                            \
+  q2_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3);           \
+                                                                 \
+  tmp0 = tmp2 + q3_in;                                           \
+  tmp1 = tmp0 + q0_in;                                           \
+  q0_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3);           \
+                                                                 \
+  tmp1 = tmp0 - p2_in;                                           \
+  tmp0 = q1_in + q3_in;                                          \
+  tmp1 = tmp0 + tmp1;                                            \
+  q1_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3);           \
+}
+
+#define LPF_MASK_HEV(p3_in, p2_in, p1_in, p0_in,                 \
+                     q0_in, q1_in, q2_in, q3_in,                 \
+                     limit_in, b_limit_in, thresh_in,            \
+                     hev_out, mask_out, flat_out) {              \
+  v16u8 p3_asub_p2_m, p2_asub_p1_m, p1_asub_p0_m, q1_asub_q0_m;  \
+  v16u8 p1_asub_q1_m, p0_asub_q0_m, q3_asub_q2_m, q2_asub_q1_m;  \
+                                                                 \
+  /* absolute subtraction of pixel values */                     \
+  p3_asub_p2_m = __msa_asub_u_b(p3_in, p2_in);                   \
+  p2_asub_p1_m = __msa_asub_u_b(p2_in, p1_in);                   \
+  p1_asub_p0_m = __msa_asub_u_b(p1_in, p0_in);                   \
+  q1_asub_q0_m = __msa_asub_u_b(q1_in, q0_in);                   \
+  q2_asub_q1_m = __msa_asub_u_b(q2_in, q1_in);                   \
+  q3_asub_q2_m = __msa_asub_u_b(q3_in, q2_in);                   \
+  p0_asub_q0_m = __msa_asub_u_b(p0_in, q0_in);                   \
+  p1_asub_q1_m = __msa_asub_u_b(p1_in, q1_in);                   \
+                                                                 \
+  /* calculation of hev */                                       \
+  flat_out = __msa_max_u_b(p1_asub_p0_m, q1_asub_q0_m);          \
+  hev_out = thresh_in < (v16u8)flat_out;                         \
+                                                                 \
+  /* calculation of mask */                                      \
+  p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p0_asub_q0_m);     \
+  p1_asub_q1_m >>= 1;                                            \
+  p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p1_asub_q1_m);     \
+                                                                 \
+  mask_out = b_limit_in < p0_asub_q0_m;                          \
+  mask_out = __msa_max_u_b(flat_out, mask_out);                  \
+  p3_asub_p2_m = __msa_max_u_b(p3_asub_p2_m, p2_asub_p1_m);      \
+  mask_out = __msa_max_u_b(p3_asub_p2_m, mask_out);              \
+  q2_asub_q1_m = __msa_max_u_b(q2_asub_q1_m, q3_asub_q2_m);      \
+  mask_out = __msa_max_u_b(q2_asub_q1_m, mask_out);              \
+                                                                 \
+  mask_out = limit_in < (v16u8)mask_out;                         \
+  mask_out = __msa_xori_b(mask_out, 0xff);                       \
+}
+#endif  /* VPX_DSP_LOOPFILTER_MSA_H_ */
diff --git a/libvpx/libvpx/vpx_dsp/mips/macros_msa.h b/libvpx/libvpx/vpx_dsp/mips/macros_msa.h
new file mode 100644
index 0000000..91e3615
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/macros_msa.h
@@ -0,0 +1,1932 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_MIPS_MACROS_MSA_H_
+#define VPX_DSP_MIPS_MACROS_MSA_H_
+
+#include <msa.h>
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+#define LD_B(RTYPE, psrc) *((const RTYPE *)(psrc))
+#define LD_UB(...) LD_B(v16u8, __VA_ARGS__)
+#define LD_SB(...) LD_B(v16i8, __VA_ARGS__)
+
+#define LD_H(RTYPE, psrc) *((const RTYPE *)(psrc))
+#define LD_UH(...) LD_H(v8u16, __VA_ARGS__)
+#define LD_SH(...) LD_H(v8i16, __VA_ARGS__)
+
+#define LD_W(RTYPE, psrc) *((const RTYPE *)(psrc))
+#define LD_SW(...) LD_W(v4i32, __VA_ARGS__)
+
+#define ST_B(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_UB(...) ST_B(v16u8, __VA_ARGS__)
+#define ST_SB(...) ST_B(v16i8, __VA_ARGS__)
+
+#define ST_H(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_SH(...) ST_H(v8i16, __VA_ARGS__)
+
+#define ST_W(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_SW(...) ST_W(v4i32, __VA_ARGS__)
+
+#if (__mips_isa_rev >= 6)
+#define LH(psrc) ({                                 \
+  const uint8_t *psrc_m = (const uint8_t *)(psrc);  \
+  uint16_t val_m;                                   \
+                                                    \
+  __asm__ __volatile__ (                            \
+      "lh  %[val_m],  %[psrc_m]  \n\t"              \
+                                                    \
+      : [val_m] "=r" (val_m)                        \
+      : [psrc_m] "m" (*psrc_m)                      \
+  );                                                \
+                                                    \
+  val_m;                                            \
+})
+
+#define LW(psrc) ({                                 \
+  const uint8_t *psrc_m = (const uint8_t *)(psrc);  \
+  uint32_t val_m;                                   \
+                                                    \
+  __asm__ __volatile__ (                            \
+      "lw  %[val_m],  %[psrc_m]  \n\t"              \
+                                                    \
+      : [val_m] "=r" (val_m)                        \
+      : [psrc_m] "m" (*psrc_m)                      \
+  );                                                \
+                                                    \
+  val_m;                                            \
+})
+
+#if (__mips == 64)
+#define LD(psrc) ({                                 \
+  const uint8_t *psrc_m = (const uint8_t *)(psrc);  \
+  uint64_t val_m = 0;                               \
+                                                    \
+  __asm__ __volatile__ (                            \
+      "ld  %[val_m],  %[psrc_m]  \n\t"              \
+                                                    \
+      : [val_m] "=r" (val_m)                        \
+      : [psrc_m] "m" (*psrc_m)                      \
+  );                                                \
+                                                    \
+  val_m;                                            \
+})
+#else  // !(__mips == 64)
+#define LD(psrc) ({                                        \
+  const uint8_t *psrc_m = (const uint8_t *)(psrc);         \
+  uint32_t val0_m, val1_m;                                 \
+  uint64_t val_m = 0;                                      \
+                                                           \
+  val0_m = LW(psrc_m);                                     \
+  val1_m = LW(psrc_m + 4);                                 \
+                                                           \
+  val_m = (uint64_t)(val1_m);                              \
+  val_m = (uint64_t)((val_m << 32) & 0xFFFFFFFF00000000);  \
+  val_m = (uint64_t)(val_m | (uint64_t)val0_m);            \
+                                                           \
+  val_m;                                                   \
+})
+#endif  // (__mips == 64)
+
+#define SH(val, pdst) {                 \
+  uint8_t *pdst_m = (uint8_t *)(pdst);  \
+  const uint16_t val_m = (val);         \
+                                        \
+  __asm__ __volatile__ (                \
+      "sh  %[val_m],  %[pdst_m]  \n\t"  \
+                                        \
+      : [pdst_m] "=m" (*pdst_m)         \
+      : [val_m] "r" (val_m)             \
+  );                                    \
+}
+
+#define SW(val, pdst) {                 \
+  uint8_t *pdst_m = (uint8_t *)(pdst);  \
+  const uint32_t val_m = (val);         \
+                                        \
+  __asm__ __volatile__ (                \
+      "sw  %[val_m],  %[pdst_m]  \n\t"  \
+                                        \
+      : [pdst_m] "=m" (*pdst_m)         \
+      : [val_m] "r" (val_m)             \
+  );                                    \
+}
+
+#define SD(val, pdst) {                 \
+  uint8_t *pdst_m = (uint8_t *)(pdst);  \
+  const uint64_t val_m = (val);         \
+                                        \
+  __asm__ __volatile__ (                \
+      "sd  %[val_m],  %[pdst_m]  \n\t"  \
+                                        \
+      : [pdst_m] "=m" (*pdst_m)         \
+      : [val_m] "r" (val_m)             \
+  );                                    \
+}
+#else  // !(__mips_isa_rev >= 6)
+#define LH(psrc) ({                                 \
+  const uint8_t *psrc_m = (const uint8_t *)(psrc);  \
+  uint16_t val_m;                                   \
+                                                    \
+  __asm__ __volatile__ (                            \
+      "ulh  %[val_m],  %[psrc_m]  \n\t"             \
+                                                    \
+      : [val_m] "=r" (val_m)                        \
+      : [psrc_m] "m" (*psrc_m)                      \
+  );                                                \
+                                                    \
+  val_m;                                            \
+})
+
+#define LW(psrc) ({                                 \
+  const uint8_t *psrc_m = (const uint8_t *)(psrc);  \
+  uint32_t val_m;                                   \
+                                                    \
+  __asm__ __volatile__ (                            \
+      "ulw  %[val_m],  %[psrc_m]  \n\t"             \
+                                                    \
+      : [val_m] "=r" (val_m)                        \
+      : [psrc_m] "m" (*psrc_m)                      \
+  );                                                \
+                                                    \
+  val_m;                                            \
+})
+
+#if (__mips == 64)
+#define LD(psrc) ({                                 \
+  const uint8_t *psrc_m = (const uint8_t *)(psrc);  \
+  uint64_t val_m = 0;                               \
+                                                    \
+  __asm__ __volatile__ (                            \
+      "uld  %[val_m],  %[psrc_m]  \n\t"             \
+                                                    \
+      : [val_m] "=r" (val_m)                        \
+      : [psrc_m] "m" (*psrc_m)                      \
+  );                                                \
+                                                    \
+  val_m;                                            \
+})
+#else  // !(__mips == 64)
+#define LD(psrc) ({                                        \
+  const uint8_t *psrc_m1 = (const uint8_t *)(psrc);        \
+  uint32_t val0_m, val1_m;                                 \
+  uint64_t val_m = 0;                                      \
+                                                           \
+  val0_m = LW(psrc_m1);                                    \
+  val1_m = LW(psrc_m1 + 4);                                \
+                                                           \
+  val_m = (uint64_t)(val1_m);                              \
+  val_m = (uint64_t)((val_m << 32) & 0xFFFFFFFF00000000);  \
+  val_m = (uint64_t)(val_m | (uint64_t)val0_m);            \
+                                                           \
+  val_m;                                                   \
+})
+#endif  // (__mips == 64)
+
+#define SH(val, pdst) {                  \
+  uint8_t *pdst_m = (uint8_t *)(pdst);   \
+  const uint16_t val_m = (val);          \
+                                         \
+  __asm__ __volatile__ (                 \
+      "ush  %[val_m],  %[pdst_m]  \n\t"  \
+                                         \
+      : [pdst_m] "=m" (*pdst_m)          \
+      : [val_m] "r" (val_m)              \
+  );                                     \
+}
+
+#define SW(val, pdst) {                  \
+  uint8_t *pdst_m = (uint8_t *)(pdst);   \
+  const uint32_t val_m = (val);          \
+                                         \
+  __asm__ __volatile__ (                 \
+      "usw  %[val_m],  %[pdst_m]  \n\t"  \
+                                         \
+      : [pdst_m] "=m" (*pdst_m)          \
+      : [val_m] "r" (val_m)              \
+  );                                     \
+}
+
+#define SD(val, pdst) {                                     \
+  uint8_t *pdst_m1 = (uint8_t *)(pdst);                     \
+  uint32_t val0_m, val1_m;                                  \
+                                                            \
+  val0_m = (uint32_t)((val) & 0x00000000FFFFFFFF);          \
+  val1_m = (uint32_t)(((val) >> 32) & 0x00000000FFFFFFFF);  \
+                                                            \
+  SW(val0_m, pdst_m1);                                      \
+  SW(val1_m, pdst_m1 + 4);                                  \
+}
+#endif  // (__mips_isa_rev >= 6)
+
+/* Description : Load 4 words with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1, out2, out3
+   Details     : Load word in 'out0' from (psrc)
+                 Load word in 'out1' from (psrc + stride)
+                 Load word in 'out2' from (psrc + 2 * stride)
+                 Load word in 'out3' from (psrc + 3 * stride)
+*/
+#define LW4(psrc, stride, out0, out1, out2, out3) {  \
+  out0 = LW((psrc));                                 \
+  out1 = LW((psrc) + stride);                        \
+  out2 = LW((psrc) + 2 * stride);                    \
+  out3 = LW((psrc) + 3 * stride);                    \
+}
+
+/* Description : Load double words with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+   Details     : Load double word in 'out0' from (psrc)
+                 Load double word in 'out1' from (psrc + stride)
+*/
+#define LD2(psrc, stride, out0, out1) {  \
+  out0 = LD((psrc));                     \
+  out1 = LD((psrc) + stride);            \
+}
+#define LD4(psrc, stride, out0, out1, out2, out3) {  \
+  LD2((psrc), stride, out0, out1);                   \
+  LD2((psrc) + 2 * stride, stride, out2, out3);      \
+}
+
+/* Description : Store 4 words with stride
+   Arguments   : Inputs - in0, in1, in2, in3, pdst, stride
+   Details     : Store word from 'in0' to (pdst)
+                 Store word from 'in1' to (pdst + stride)
+                 Store word from 'in2' to (pdst + 2 * stride)
+                 Store word from 'in3' to (pdst + 3 * stride)
+*/
+#define SW4(in0, in1, in2, in3, pdst, stride) {  \
+  SW(in0, (pdst))                                \
+  SW(in1, (pdst) + stride);                      \
+  SW(in2, (pdst) + 2 * stride);                  \
+  SW(in3, (pdst) + 3 * stride);                  \
+}
+
+/* Description : Store 4 double words with stride
+   Arguments   : Inputs - in0, in1, in2, in3, pdst, stride
+   Details     : Store double word from 'in0' to (pdst)
+                 Store double word from 'in1' to (pdst + stride)
+                 Store double word from 'in2' to (pdst + 2 * stride)
+                 Store double word from 'in3' to (pdst + 3 * stride)
+*/
+#define SD4(in0, in1, in2, in3, pdst, stride) {  \
+  SD(in0, (pdst))                                \
+  SD(in1, (pdst) + stride);                      \
+  SD(in2, (pdst) + 2 * stride);                  \
+  SD(in3, (pdst) + 3 * stride);                  \
+}
+
+/* Description : Load vectors with 16 byte elements with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Load 16 byte elements in 'out0' from (psrc)
+                 Load 16 byte elements in 'out1' from (psrc + stride)
+*/
+#define LD_B2(RTYPE, psrc, stride, out0, out1) {  \
+  out0 = LD_B(RTYPE, (psrc));                     \
+  out1 = LD_B(RTYPE, (psrc) + stride);            \
+}
+#define LD_UB2(...) LD_B2(v16u8, __VA_ARGS__)
+#define LD_SB2(...) LD_B2(v16i8, __VA_ARGS__)
+
+#define LD_B3(RTYPE, psrc, stride, out0, out1, out2) {  \
+  LD_B2(RTYPE, (psrc), stride, out0, out1);             \
+  out2 = LD_B(RTYPE, (psrc) + 2 * stride);              \
+}
+#define LD_UB3(...) LD_B3(v16u8, __VA_ARGS__)
+
+#define LD_B4(RTYPE, psrc, stride, out0, out1, out2, out3) {  \
+  LD_B2(RTYPE, (psrc), stride, out0, out1);                   \
+  LD_B2(RTYPE, (psrc) + 2 * stride , stride, out2, out3);     \
+}
+#define LD_UB4(...) LD_B4(v16u8, __VA_ARGS__)
+#define LD_SB4(...) LD_B4(v16i8, __VA_ARGS__)
+
+#define LD_B5(RTYPE, psrc, stride, out0, out1, out2, out3, out4) {  \
+  LD_B4(RTYPE, (psrc), stride, out0, out1, out2, out3);             \
+  out4 = LD_B(RTYPE, (psrc) + 4 * stride);                          \
+}
+#define LD_UB5(...) LD_B5(v16u8, __VA_ARGS__)
+#define LD_SB5(...) LD_B5(v16i8, __VA_ARGS__)
+
+#define LD_B7(RTYPE, psrc, stride,                             \
+              out0, out1, out2, out3, out4, out5, out6) {      \
+  LD_B5(RTYPE, (psrc), stride, out0, out1, out2, out3, out4);  \
+  LD_B2(RTYPE, (psrc) + 5 * stride, stride, out5, out6);       \
+}
+#define LD_SB7(...) LD_B7(v16i8, __VA_ARGS__)
+
+#define LD_B8(RTYPE, psrc, stride,                                    \
+              out0, out1, out2, out3, out4, out5, out6, out7) {       \
+  LD_B4(RTYPE, (psrc), stride, out0, out1, out2, out3);               \
+  LD_B4(RTYPE, (psrc) + 4 * stride, stride, out4, out5, out6, out7);  \
+}
+#define LD_UB8(...) LD_B8(v16u8, __VA_ARGS__)
+#define LD_SB8(...) LD_B8(v16i8, __VA_ARGS__)
+
+/* Description : Load vectors with 8 halfword elements with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+   Details     : Load 8 halfword elements in 'out0' from (psrc)
+                 Load 8 halfword elements in 'out1' from (psrc + stride)
+*/
+#define LD_H2(RTYPE, psrc, stride, out0, out1) {  \
+  out0 = LD_H(RTYPE, (psrc));                     \
+  out1 = LD_H(RTYPE, (psrc) + (stride));          \
+}
+#define LD_SH2(...) LD_H2(v8i16, __VA_ARGS__)
+
+#define LD_H4(RTYPE, psrc, stride, out0, out1, out2, out3) {  \
+  LD_H2(RTYPE, (psrc), stride, out0, out1);                   \
+  LD_H2(RTYPE, (psrc) + 2 * stride, stride, out2, out3);      \
+}
+#define LD_SH4(...) LD_H4(v8i16, __VA_ARGS__)
+
+#define LD_H8(RTYPE, psrc, stride,                                    \
+              out0, out1, out2, out3, out4, out5, out6, out7) {       \
+  LD_H4(RTYPE, (psrc), stride, out0, out1, out2, out3);               \
+  LD_H4(RTYPE, (psrc) + 4 * stride, stride, out4, out5, out6, out7);  \
+}
+#define LD_SH8(...) LD_H8(v8i16, __VA_ARGS__)
+
+#define LD_H16(RTYPE, psrc, stride,                                     \
+               out0, out1, out2, out3, out4, out5, out6, out7,          \
+               out8, out9, out10, out11, out12, out13, out14, out15) {  \
+  LD_H8(RTYPE, (psrc), stride,                                          \
+        out0, out1, out2, out3, out4, out5, out6, out7);                \
+  LD_H8(RTYPE, (psrc) + 8 * stride, stride,                             \
+        out8, out9, out10, out11, out12, out13, out14, out15);          \
+}
+#define LD_SH16(...) LD_H16(v8i16, __VA_ARGS__)
+
+/* Description : Load 4x4 block of signed halfword elements from 1D source
+                 data into 4 vectors (Each vector with 4 signed halfwords)
+   Arguments   : Input   - psrc
+                 Outputs - out0, out1, out2, out3
+*/
+#define LD4x4_SH(psrc, out0, out1, out2, out3) {         \
+  out0 = LD_SH(psrc);                                    \
+  out2 = LD_SH(psrc + 8);                                \
+  out1 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out0);  \
+  out3 = (v8i16)__msa_ilvl_d((v2i64)out2, (v2i64)out2);  \
+}
+
+/* Description : Load 2 vectors of signed word elements with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+                 Return Type - signed word
+*/
+#define LD_SW2(psrc, stride, out0, out1) {  \
+  out0 = LD_SW((psrc));                     \
+  out1 = LD_SW((psrc) + stride);            \
+}
+
+/* Description : Store vectors of 16 byte elements with stride
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 16 byte elements from 'in0' to (pdst)
+                 Store 16 byte elements from 'in1' to (pdst + stride)
+*/
+#define ST_B2(RTYPE, in0, in1, pdst, stride) {  \
+  ST_B(RTYPE, in0, (pdst));                     \
+  ST_B(RTYPE, in1, (pdst) + stride);            \
+}
+#define ST_UB2(...) ST_B2(v16u8, __VA_ARGS__)
+
+#define ST_B4(RTYPE, in0, in1, in2, in3, pdst, stride) {  \
+  ST_B2(RTYPE, in0, in1, (pdst), stride);                 \
+  ST_B2(RTYPE, in2, in3, (pdst) + 2 * stride, stride);    \
+}
+#define ST_UB4(...) ST_B4(v16u8, __VA_ARGS__)
+
+#define ST_B8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,      \
+              pdst, stride) {                                     \
+  ST_B4(RTYPE, in0, in1, in2, in3, pdst, stride);                 \
+  ST_B4(RTYPE, in4, in5, in6, in7, (pdst) + 4 * stride, stride);  \
+}
+#define ST_UB8(...) ST_B8(v16u8, __VA_ARGS__)
+
+/* Description : Store vectors of 8 halfword elements with stride
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 8 halfword elements from 'in0' to (pdst)
+                 Store 8 halfword elements from 'in1' to (pdst + stride)
+*/
+#define ST_H2(RTYPE, in0, in1, pdst, stride) {  \
+  ST_H(RTYPE, in0, (pdst));                     \
+  ST_H(RTYPE, in1, (pdst) + stride);            \
+}
+#define ST_SH2(...) ST_H2(v8i16, __VA_ARGS__)
+
+#define ST_H4(RTYPE, in0, in1, in2, in3, pdst, stride) {  \
+  ST_H2(RTYPE, in0, in1, (pdst), stride);                 \
+  ST_H2(RTYPE, in2, in3, (pdst) + 2 * stride, stride);    \
+}
+#define ST_SH4(...) ST_H4(v8i16, __VA_ARGS__)
+
+#define ST_H8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, pdst, stride) {  \
+  ST_H4(RTYPE, in0, in1, in2, in3, (pdst), stride);                           \
+  ST_H4(RTYPE, in4, in5, in6, in7, (pdst) + 4 * stride, stride);              \
+}
+#define ST_SH8(...) ST_H8(v8i16, __VA_ARGS__)
+
+/* Description : Store vectors of word elements with stride
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 4 word elements from 'in0' to (pdst)
+                 Store 4 word elements from 'in1' to (pdst + stride)
+*/
+#define ST_SW2(in0, in1, pdst, stride) {  \
+  ST_SW(in0, (pdst));                     \
+  ST_SW(in1, (pdst) + stride);            \
+}
+
+/* Description : Store 2x4 byte block to destination memory from input vector
+   Arguments   : Inputs - in, stidx, pdst, stride
+   Details     : Index 'stidx' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst)
+                 Index 'stidx+1' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst + stride)
+                 Index 'stidx+2' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst + 2 * stride)
+                 Index 'stidx+3' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst + 3 * stride)
+*/
+#define ST2x4_UB(in, stidx, pdst, stride) {         \
+  uint16_t out0_m, out1_m, out2_m, out3_m;          \
+  uint8_t *pblk_2x4_m = (uint8_t *)(pdst);          \
+                                                    \
+  out0_m = __msa_copy_u_h((v8i16)in, (stidx));      \
+  out1_m = __msa_copy_u_h((v8i16)in, (stidx + 1));  \
+  out2_m = __msa_copy_u_h((v8i16)in, (stidx + 2));  \
+  out3_m = __msa_copy_u_h((v8i16)in, (stidx + 3));  \
+                                                    \
+  SH(out0_m, pblk_2x4_m);                           \
+  SH(out1_m, pblk_2x4_m + stride);                  \
+  SH(out2_m, pblk_2x4_m + 2 * stride);              \
+  SH(out3_m, pblk_2x4_m + 3 * stride);              \
+}
+
+/* Description : Store 4x2 byte block to destination memory from input vector
+   Arguments   : Inputs - in, pdst, stride
+   Details     : Index 0 word element from 'in' vector is copied to the GP
+                 register and stored to (pdst)
+                 Index 1 word element from 'in' vector is copied to the GP
+                 register and stored to (pdst + stride)
+*/
+#define ST4x2_UB(in, pdst, stride) {        \
+  uint32_t out0_m, out1_m;                  \
+  uint8_t *pblk_4x2_m = (uint8_t *)(pdst);  \
+                                            \
+  out0_m = __msa_copy_u_w((v4i32)in, 0);    \
+  out1_m = __msa_copy_u_w((v4i32)in, 1);    \
+                                            \
+  SW(out0_m, pblk_4x2_m);                   \
+  SW(out1_m, pblk_4x2_m + stride);          \
+}
+
+/* Description : Store 4x4 byte block to destination memory from input vector
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : 'Idx0' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst)
+                 'Idx1' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst + stride)
+                 'Idx2' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst + 2 * stride)
+                 'Idx3' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst + 3 * stride)
+*/
+#define ST4x4_UB(in0, in1, idx0, idx1, idx2, idx3, pdst, stride) {  \
+  uint32_t out0_m, out1_m, out2_m, out3_m;                          \
+  uint8_t *pblk_4x4_m = (uint8_t *)(pdst);                          \
+                                                                    \
+  out0_m = __msa_copy_u_w((v4i32)in0, idx0);                        \
+  out1_m = __msa_copy_u_w((v4i32)in0, idx1);                        \
+  out2_m = __msa_copy_u_w((v4i32)in1, idx2);                        \
+  out3_m = __msa_copy_u_w((v4i32)in1, idx3);                        \
+                                                                    \
+  SW4(out0_m, out1_m, out2_m, out3_m, pblk_4x4_m, stride);          \
+}
+#define ST4x8_UB(in0, in1, pdst, stride) {                        \
+  uint8_t *pblk_4x8 = (uint8_t *)(pdst);                          \
+                                                                  \
+  ST4x4_UB(in0, in0, 0, 1, 2, 3, pblk_4x8, stride);               \
+  ST4x4_UB(in1, in1, 0, 1, 2, 3, pblk_4x8 + 4 * stride, stride);  \
+}
+
+/* Description : Store 8x1 byte block to destination memory from input vector
+   Arguments   : Inputs - in, pdst
+   Details     : Index 0 double word element from 'in' vector is copied to the
+                 GP register and stored to (pdst)
+*/
+#define ST8x1_UB(in, pdst) {              \
+  uint64_t out0_m;                        \
+                                          \
+  out0_m = __msa_copy_u_d((v2i64)in, 0);  \
+  SD(out0_m, pdst);                       \
+}
+
+/* Description : Store 8x2 byte block to destination memory from input vector
+   Arguments   : Inputs - in, pdst, stride
+   Details     : Index 0 double word element from 'in' vector is copied to the
+                 GP register and stored to (pdst)
+                 Index 1 double word element from 'in' vector is copied to the
+                 GP register and stored to (pdst + stride)
+*/
+#define ST8x2_UB(in, pdst, stride) {        \
+  uint64_t out0_m, out1_m;                  \
+  uint8_t *pblk_8x2_m = (uint8_t *)(pdst);  \
+                                            \
+  out0_m = __msa_copy_u_d((v2i64)in, 0);    \
+  out1_m = __msa_copy_u_d((v2i64)in, 1);    \
+                                            \
+  SD(out0_m, pblk_8x2_m);                   \
+  SD(out1_m, pblk_8x2_m + stride);          \
+}
+
+/* Description : Store 8x4 byte block to destination memory from input
+                 vectors
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Index 0 double word element from 'in0' vector is copied to the
+                 GP register and stored to (pdst)
+                 Index 1 double word element from 'in0' vector is copied to the
+                 GP register and stored to (pdst + stride)
+                 Index 0 double word element from 'in1' vector is copied to the
+                 GP register and stored to (pdst + 2 * stride)
+                 Index 1 double word element from 'in1' vector is copied to the
+                 GP register and stored to (pdst + 3 * stride)
+*/
+#define ST8x4_UB(in0, in1, pdst, stride) {                  \
+  uint64_t out0_m, out1_m, out2_m, out3_m;                  \
+  uint8_t *pblk_8x4_m = (uint8_t *)(pdst);                  \
+                                                            \
+  out0_m = __msa_copy_u_d((v2i64)in0, 0);                   \
+  out1_m = __msa_copy_u_d((v2i64)in0, 1);                   \
+  out2_m = __msa_copy_u_d((v2i64)in1, 0);                   \
+  out3_m = __msa_copy_u_d((v2i64)in1, 1);                   \
+                                                            \
+  SD4(out0_m, out1_m, out2_m, out3_m, pblk_8x4_m, stride);  \
+}
+
+/* Description : average with rounding (in0 + in1 + 1) / 2.
+   Arguments   : Inputs  - in0, in1, in2, in3,
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each unsigned byte element from 'in0' vector is added with
+                 each unsigned byte element from 'in1' vector. Then the average
+                 with rounding is calculated and written to 'out0'
+*/
+#define AVER_UB2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_aver_u_b((v16u8)in0, (v16u8)in1);    \
+  out1 = (RTYPE)__msa_aver_u_b((v16u8)in2, (v16u8)in3);    \
+}
+#define AVER_UB2_UB(...) AVER_UB2(v16u8, __VA_ARGS__)
+
+#define AVER_UB4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                 out0, out1, out2, out3) {                       \
+  AVER_UB2(RTYPE, in0, in1, in2, in3, out0, out1)                \
+  AVER_UB2(RTYPE, in4, in5, in6, in7, out2, out3)                \
+}
+#define AVER_UB4_UB(...) AVER_UB4(v16u8, __VA_ARGS__)
+
+/* Description : Immediate number of elements to slide with zero
+   Arguments   : Inputs  - in0, in1, slide_val
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'zero_m' vector are slid into 'in0' by
+                 value specified in the 'slide_val'
+*/
+#define SLDI_B2_0(RTYPE, in0, in1, out0, out1, slide_val) {          \
+  v16i8 zero_m = { 0 };                                              \
+  out0 = (RTYPE)__msa_sldi_b((v16i8)zero_m, (v16i8)in0, slide_val);  \
+  out1 = (RTYPE)__msa_sldi_b((v16i8)zero_m, (v16i8)in1, slide_val);  \
+}
+#define SLDI_B2_0_SW(...) SLDI_B2_0(v4i32, __VA_ARGS__)
+
+#define SLDI_B4_0(RTYPE, in0, in1, in2, in3,            \
+                  out0, out1, out2, out3, slide_val) {  \
+  SLDI_B2_0(RTYPE, in0, in1, out0, out1, slide_val);    \
+  SLDI_B2_0(RTYPE, in2, in3, out2, out3, slide_val);    \
+}
+#define SLDI_B4_0_UB(...) SLDI_B4_0(v16u8, __VA_ARGS__)
+
+/* Description : Immediate number of elements to slide
+   Arguments   : Inputs  - in0_0, in0_1, in1_0, in1_1, slide_val
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'in0_0' vector are slid into 'in1_0' by
+                 value specified in the 'slide_val'
+*/
+#define SLDI_B2(RTYPE, in0_0, in0_1, in1_0, in1_1, out0, out1, slide_val) {  \
+  out0 = (RTYPE)__msa_sldi_b((v16i8)in0_0, (v16i8)in1_0, slide_val);         \
+  out1 = (RTYPE)__msa_sldi_b((v16i8)in0_1, (v16i8)in1_1, slide_val);         \
+}
+#define SLDI_B2_UB(...) SLDI_B2(v16u8, __VA_ARGS__)
+#define SLDI_B2_SH(...) SLDI_B2(v8i16, __VA_ARGS__)
+
+#define SLDI_B3(RTYPE, in0_0, in0_1, in0_2, in1_0, in1_1, in1_2,      \
+                out0, out1, out2, slide_val) {                        \
+  SLDI_B2(RTYPE, in0_0, in0_1, in1_0, in1_1, out0, out1, slide_val)   \
+  out2 = (RTYPE)__msa_sldi_b((v16i8)in0_2, (v16i8)in1_2, slide_val);  \
+}
+#define SLDI_B3_SB(...) SLDI_B3(v16i8, __VA_ARGS__)
+#define SLDI_B3_UH(...) SLDI_B3(v8u16, __VA_ARGS__)
+
+/* Description : Shuffle byte vector elements as per mask vector
+   Arguments   : Inputs  - in0, in1, in2, in3, mask0, mask1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'in0' & 'in1' are copied selectively to
+                 'out0' as per control vector 'mask0'
+*/
+#define VSHF_B2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1) {  \
+  out0 = (RTYPE)__msa_vshf_b((v16i8)mask0, (v16i8)in1, (v16i8)in0);     \
+  out1 = (RTYPE)__msa_vshf_b((v16i8)mask1, (v16i8)in3, (v16i8)in2);     \
+}
+#define VSHF_B2_UB(...) VSHF_B2(v16u8, __VA_ARGS__)
+#define VSHF_B2_SB(...) VSHF_B2(v16i8, __VA_ARGS__)
+#define VSHF_B2_UH(...) VSHF_B2(v8u16, __VA_ARGS__)
+
+#define VSHF_B4(RTYPE, in0, in1, mask0, mask1, mask2, mask3,     \
+                out0, out1, out2, out3) {                        \
+  VSHF_B2(RTYPE, in0, in1, in0, in1, mask0, mask1, out0, out1);  \
+  VSHF_B2(RTYPE, in0, in1, in0, in1, mask2, mask3, out2, out3);  \
+}
+#define VSHF_B4_SB(...) VSHF_B4(v16i8, __VA_ARGS__)
+#define VSHF_B4_SH(...) VSHF_B4(v8i16, __VA_ARGS__)
+
+/* Description : Dot product of byte vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Unsigned byte elements from 'mult0' are multiplied with
+                 unsigned byte elements from 'cnst0' producing a result
+                 twice the size of input i.e. unsigned halfword.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_UB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) {  \
+  out0 = (RTYPE)__msa_dotp_u_h((v16u8)mult0, (v16u8)cnst0);        \
+  out1 = (RTYPE)__msa_dotp_u_h((v16u8)mult1, (v16u8)cnst1);        \
+}
+#define DOTP_UB2_UH(...) DOTP_UB2(v8u16, __VA_ARGS__)
+
+#define DOTP_UB4(RTYPE, mult0, mult1, mult2, mult3,         \
+                 cnst0, cnst1, cnst2, cnst3,                \
+                 out0, out1, out2, out3) {                  \
+  DOTP_UB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);  \
+  DOTP_UB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);  \
+}
+#define DOTP_UB4_UH(...) DOTP_UB4(v8u16, __VA_ARGS__)
+
+/* Description : Dot product of byte vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed byte elements from 'mult0' are multiplied with
+                 signed byte elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed halfword.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) {  \
+  out0 = (RTYPE)__msa_dotp_s_h((v16i8)mult0, (v16i8)cnst0);        \
+  out1 = (RTYPE)__msa_dotp_s_h((v16i8)mult1, (v16i8)cnst1);        \
+}
+#define DOTP_SB2_SH(...) DOTP_SB2(v8i16, __VA_ARGS__)
+
+#define DOTP_SB4(RTYPE, mult0, mult1, mult2, mult3,                     \
+                 cnst0, cnst1, cnst2, cnst3, out0, out1, out2, out3) {  \
+  DOTP_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);              \
+  DOTP_SB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);              \
+}
+#define DOTP_SB4_SH(...) DOTP_SB4(v8i16, __VA_ARGS__)
+
+/* Description : Dot product of halfword vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed halfword elements from 'mult0' are multiplied with
+                 signed halfword elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed word.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) {  \
+  out0 = (RTYPE)__msa_dotp_s_w((v8i16)mult0, (v8i16)cnst0);        \
+  out1 = (RTYPE)__msa_dotp_s_w((v8i16)mult1, (v8i16)cnst1);        \
+}
+#define DOTP_SH2_SW(...) DOTP_SH2(v4i32, __VA_ARGS__)
+
+#define DOTP_SH4(RTYPE, mult0, mult1, mult2, mult3,         \
+                 cnst0, cnst1, cnst2, cnst3,                \
+                 out0, out1, out2, out3) {                  \
+  DOTP_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);  \
+  DOTP_SH2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);  \
+}
+#define DOTP_SH4_SW(...) DOTP_SH4(v4i32, __VA_ARGS__)
+
+/* Description : Dot product of word vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed word elements from 'mult0' are multiplied with
+                 signed word elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed double word.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_SW2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) {  \
+  out0 = (RTYPE)__msa_dotp_s_d((v4i32)mult0, (v4i32)cnst0);        \
+  out1 = (RTYPE)__msa_dotp_s_d((v4i32)mult1, (v4i32)cnst1);        \
+}
+#define DOTP_SW2_SD(...) DOTP_SW2(v2i64, __VA_ARGS__)
+
+/* Description : Dot product & addition of byte vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed byte elements from 'mult0' are multiplied with
+                 signed byte elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed halfword.
+                 The multiplication result of adjacent odd-even elements
+                 are added to the 'out0' vector
+*/
+#define DPADD_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) {         \
+  out0 = (RTYPE)__msa_dpadd_s_h((v8i16)out0, (v16i8)mult0, (v16i8)cnst0);  \
+  out1 = (RTYPE)__msa_dpadd_s_h((v8i16)out1, (v16i8)mult1, (v16i8)cnst1);  \
+}
+#define DPADD_SB2_SH(...) DPADD_SB2(v8i16, __VA_ARGS__)
+
+#define DPADD_SB4(RTYPE, mult0, mult1, mult2, mult3,                     \
+                  cnst0, cnst1, cnst2, cnst3, out0, out1, out2, out3) {  \
+  DPADD_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);              \
+  DPADD_SB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);              \
+}
+#define DPADD_SB4_SH(...) DPADD_SB4(v8i16, __VA_ARGS__)
+
+/* Description : Dot product & addition of halfword vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed halfword elements from 'mult0' are multiplied with
+                 signed halfword elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed word.
+                 The multiplication result of adjacent odd-even elements
+                 are added to the 'out0' vector
+*/
+#define DPADD_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) {         \
+  out0 = (RTYPE)__msa_dpadd_s_w((v4i32)out0, (v8i16)mult0, (v8i16)cnst0);  \
+  out1 = (RTYPE)__msa_dpadd_s_w((v4i32)out1, (v8i16)mult1, (v8i16)cnst1);  \
+}
+#define DPADD_SH2_SW(...) DPADD_SH2(v4i32, __VA_ARGS__)
+
+/* Description : Dot product & addition of double word vector elements
+   Arguments   : Inputs  - mult0, mult1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each signed word element from 'mult0' is multiplied with itself
+                 producing an intermediate result twice the size of input
+                 i.e. signed double word
+                 The multiplication result of adjacent odd-even elements
+                 are added to the 'out0' vector
+*/
+#define DPADD_SD2(RTYPE, mult0, mult1, out0, out1) {                       \
+  out0 = (RTYPE)__msa_dpadd_s_d((v2i64)out0, (v4i32)mult0, (v4i32)mult0);  \
+  out1 = (RTYPE)__msa_dpadd_s_d((v2i64)out1, (v4i32)mult1, (v4i32)mult1);  \
+}
+#define DPADD_SD2_SD(...) DPADD_SD2(v2i64, __VA_ARGS__)
+
+/* Description : Minimum values between unsigned elements of
+                 either vector are copied to the output vector
+   Arguments   : Inputs  - in0, in1, min_vec
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Minimum of unsigned halfword element values from 'in0' and
+                 'min_vec' are written to output vector 'in0'
+*/
+#define MIN_UH2(RTYPE, in0, in1, min_vec) {         \
+  in0 = (RTYPE)__msa_min_u_h((v8u16)in0, min_vec);  \
+  in1 = (RTYPE)__msa_min_u_h((v8u16)in1, min_vec);  \
+}
+#define MIN_UH2_UH(...) MIN_UH2(v8u16, __VA_ARGS__)
+
+#define MIN_UH4(RTYPE, in0, in1, in2, in3, min_vec) {  \
+  MIN_UH2(RTYPE, in0, in1, min_vec);                   \
+  MIN_UH2(RTYPE, in2, in3, min_vec);                   \
+}
+#define MIN_UH4_UH(...) MIN_UH4(v8u16, __VA_ARGS__)
+
+/* Description : Clips all signed halfword elements of input vector
+                 between 0 & 255
+   Arguments   : Input  - in
+                 Output - out_m
+                 Return Type - signed halfword
+*/
+#define CLIP_SH_0_255(in) ({                          \
+  v8i16 max_m = __msa_ldi_h(255);                     \
+  v8i16 out_m;                                        \
+                                                      \
+  out_m = __msa_maxi_s_h((v8i16)in, 0);               \
+  out_m = __msa_min_s_h((v8i16)max_m, (v8i16)out_m);  \
+  out_m;                                              \
+})
+#define CLIP_SH2_0_255(in0, in1) {  \
+  in0 = CLIP_SH_0_255(in0);         \
+  in1 = CLIP_SH_0_255(in1);         \
+}
+#define CLIP_SH4_0_255(in0, in1, in2, in3) {  \
+  CLIP_SH2_0_255(in0, in1);                   \
+  CLIP_SH2_0_255(in2, in3);                   \
+}
+
+/* Description : Horizontal addition of 4 signed word elements of input vector
+   Arguments   : Input  - in       (signed word vector)
+                 Output - sum_m    (i32 sum)
+                 Return Type - signed word (GP)
+   Details     : 4 signed word elements of 'in' vector are added together and
+                 the resulting integer sum is returned
+*/
+#define HADD_SW_S32(in) ({                        \
+  v2i64 res0_m, res1_m;                           \
+  int32_t sum_m;                                  \
+                                                  \
+  res0_m = __msa_hadd_s_d((v4i32)in, (v4i32)in);  \
+  res1_m = __msa_splati_d(res0_m, 1);             \
+  res0_m = res0_m + res1_m;                       \
+  sum_m = __msa_copy_s_w((v4i32)res0_m, 0);       \
+  sum_m;                                          \
+})
+
+/* Description : Horizontal addition of 8 unsigned halfword elements
+   Arguments   : Inputs  - in       (unsigned halfword vector)
+                 Outputs - sum_m    (u32 sum)
+                 Return Type - unsigned word
+   Details     : 8 unsigned halfword elements of input vector are added
+                 together and the resulting integer sum is returned
+*/
+#define HADD_UH_U32(in) ({                           \
+  v4u32 res_m;                                       \
+  v2u64 res0_m, res1_m;                              \
+  uint32_t sum_m;                                    \
+                                                     \
+  res_m = __msa_hadd_u_w((v8u16)in, (v8u16)in);      \
+  res0_m = __msa_hadd_u_d(res_m, res_m);             \
+  res1_m = (v2u64)__msa_splati_d((v2i64)res0_m, 1);  \
+  res0_m = res0_m + res1_m;                          \
+  sum_m = __msa_copy_u_w((v4i32)res0_m, 0);          \
+  sum_m;                                             \
+})
+
+/* Description : Horizontal addition of unsigned byte vector elements
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each unsigned odd byte element from 'in0' is added to
+                 even unsigned byte element from 'in0' (pairwise) and the
+                 halfword result is written to 'out0'
+*/
+#define HADD_UB2(RTYPE, in0, in1, out0, out1) {          \
+  out0 = (RTYPE)__msa_hadd_u_h((v16u8)in0, (v16u8)in0);  \
+  out1 = (RTYPE)__msa_hadd_u_h((v16u8)in1, (v16u8)in1);  \
+}
+#define HADD_UB2_UH(...) HADD_UB2(v8u16, __VA_ARGS__)
+
+#define HADD_UB4(RTYPE, in0, in1, in2, in3, out0, out1, out2, out3) {  \
+  HADD_UB2(RTYPE, in0, in1, out0, out1);                               \
+  HADD_UB2(RTYPE, in2, in3, out2, out3);                               \
+}
+#define HADD_UB4_UH(...) HADD_UB4(v8u16, __VA_ARGS__)
+
+/* Description : Horizontal subtraction of unsigned byte vector elements
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each unsigned odd byte element from 'in0' is subtracted from
+                 even unsigned byte element from 'in0' (pairwise) and the
+                 halfword result is written to 'out0'
+*/
+#define HSUB_UB2(RTYPE, in0, in1, out0, out1) {          \
+  out0 = (RTYPE)__msa_hsub_u_h((v16u8)in0, (v16u8)in0);  \
+  out1 = (RTYPE)__msa_hsub_u_h((v16u8)in1, (v16u8)in1);  \
+}
+#define HSUB_UB2_SH(...) HSUB_UB2(v8i16, __VA_ARGS__)
+
+/* Description : SAD (Sum of Absolute Difference)
+   Arguments   : Inputs  - in0, in1, ref0, ref1
+                 Outputs - sad_m                 (halfword vector)
+                 Return Type - unsigned halfword
+   Details     : Absolute difference of all the byte elements from 'in0' with
+                 'ref0' is calculated and preserved in 'diff0'. Then even-odd
+                 pairs are added together to generate 8 halfword results.
+*/
+#define SAD_UB2_UH(in0, in1, ref0, ref1) ({                 \
+  v16u8 diff0_m, diff1_m;                                   \
+  v8u16 sad_m = { 0 };                                      \
+                                                            \
+  diff0_m = __msa_asub_u_b((v16u8)in0, (v16u8)ref0);        \
+  diff1_m = __msa_asub_u_b((v16u8)in1, (v16u8)ref1);        \
+                                                            \
+  sad_m += __msa_hadd_u_h((v16u8)diff0_m, (v16u8)diff0_m);  \
+  sad_m += __msa_hadd_u_h((v16u8)diff1_m, (v16u8)diff1_m);  \
+                                                            \
+  sad_m;                                                    \
+})
+
+/* Description : Horizontal subtraction of signed halfword vector elements
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each signed odd halfword element from 'in0' is subtracted from
+                 even signed halfword element from 'in0' (pairwise) and the
+                 word result is written to 'out0'
+*/
+#define HSUB_UH2(RTYPE, in0, in1, out0, out1) {          \
+  out0 = (RTYPE)__msa_hsub_s_w((v8i16)in0, (v8i16)in0);  \
+  out1 = (RTYPE)__msa_hsub_s_w((v8i16)in1, (v8i16)in1);  \
+}
+#define HSUB_UH2_SW(...) HSUB_UH2(v4i32, __VA_ARGS__)
+
+/* Description : Set element n input vector to GPR value
+   Arguments   : Inputs - in0, in1, in2, in3
+                 Output - out
+                 Return Type - as per RTYPE
+   Details     : Set element 0 in vector 'out' to value specified in 'in0'
+*/
+#define INSERT_W2(RTYPE, in0, in1, out) {           \
+  out = (RTYPE)__msa_insert_w((v4i32)out, 0, in0);  \
+  out = (RTYPE)__msa_insert_w((v4i32)out, 1, in1);  \
+}
+#define INSERT_W2_SB(...) INSERT_W2(v16i8, __VA_ARGS__)
+
+#define INSERT_W4(RTYPE, in0, in1, in2, in3, out) {  \
+  out = (RTYPE)__msa_insert_w((v4i32)out, 0, in0);   \
+  out = (RTYPE)__msa_insert_w((v4i32)out, 1, in1);   \
+  out = (RTYPE)__msa_insert_w((v4i32)out, 2, in2);   \
+  out = (RTYPE)__msa_insert_w((v4i32)out, 3, in3);   \
+}
+#define INSERT_W4_UB(...) INSERT_W4(v16u8, __VA_ARGS__)
+#define INSERT_W4_SB(...) INSERT_W4(v16i8, __VA_ARGS__)
+
+#define INSERT_D2(RTYPE, in0, in1, out) {           \
+  out = (RTYPE)__msa_insert_d((v2i64)out, 0, in0);  \
+  out = (RTYPE)__msa_insert_d((v2i64)out, 1, in1);  \
+}
+#define INSERT_D2_UB(...) INSERT_D2(v16u8, __VA_ARGS__)
+#define INSERT_D2_SB(...) INSERT_D2(v16i8, __VA_ARGS__)
+
+/* Description : Interleave even byte elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even byte elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_B2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvev_b((v16i8)in1, (v16i8)in0);     \
+  out1 = (RTYPE)__msa_ilvev_b((v16i8)in3, (v16i8)in2);     \
+}
+#define ILVEV_B2_UB(...) ILVEV_B2(v16u8, __VA_ARGS__)
+#define ILVEV_B2_SH(...) ILVEV_B2(v8i16, __VA_ARGS__)
+
+/* Description : Interleave even halfword elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even halfword elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_H2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvev_h((v8i16)in1, (v8i16)in0);     \
+  out1 = (RTYPE)__msa_ilvev_h((v8i16)in3, (v8i16)in2);     \
+}
+#define ILVEV_H2_UB(...) ILVEV_H2(v16u8, __VA_ARGS__)
+#define ILVEV_H2_SH(...) ILVEV_H2(v8i16, __VA_ARGS__)
+#define ILVEV_H2_SW(...) ILVEV_H2(v4i32, __VA_ARGS__)
+
+/* Description : Interleave even word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even word elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_W2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvev_w((v4i32)in1, (v4i32)in0);     \
+  out1 = (RTYPE)__msa_ilvev_w((v4i32)in3, (v4i32)in2);     \
+}
+#define ILVEV_W2_SB(...) ILVEV_W2(v16i8, __VA_ARGS__)
+
+/* Description : Interleave even double word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even double word elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_D2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvev_d((v2i64)in1, (v2i64)in0);     \
+  out1 = (RTYPE)__msa_ilvev_d((v2i64)in3, (v2i64)in2);     \
+}
+#define ILVEV_D2_UB(...) ILVEV_D2(v16u8, __VA_ARGS__)
+
+/* Description : Interleave left half of byte elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Left half of byte elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'.
+*/
+#define ILVL_B2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvl_b((v16i8)in0, (v16i8)in1);     \
+  out1 = (RTYPE)__msa_ilvl_b((v16i8)in2, (v16i8)in3);     \
+}
+#define ILVL_B2_UB(...) ILVL_B2(v16u8, __VA_ARGS__)
+#define ILVL_B2_SB(...) ILVL_B2(v16i8, __VA_ARGS__)
+#define ILVL_B2_UH(...) ILVL_B2(v8u16, __VA_ARGS__)
+#define ILVL_B2_SH(...) ILVL_B2(v8i16, __VA_ARGS__)
+
+#define ILVL_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3) {                       \
+  ILVL_B2(RTYPE, in0, in1, in2, in3, out0, out1);               \
+  ILVL_B2(RTYPE, in4, in5, in6, in7, out2, out3);               \
+}
+#define ILVL_B4_SB(...) ILVL_B4(v16i8, __VA_ARGS__)
+#define ILVL_B4_UH(...) ILVL_B4(v8u16, __VA_ARGS__)
+
+/* Description : Interleave left half of halfword elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Left half of halfword elements of 'in0' and 'in1' are
+                 interleaved and written to 'out0'.
+*/
+#define ILVL_H2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvl_h((v8i16)in0, (v8i16)in1);     \
+  out1 = (RTYPE)__msa_ilvl_h((v8i16)in2, (v8i16)in3);     \
+}
+#define ILVL_H2_SH(...) ILVL_H2(v8i16, __VA_ARGS__)
+
+/* Description : Interleave left half of word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Left half of word elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'.
+*/
+#define ILVL_W2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvl_w((v4i32)in0, (v4i32)in1);     \
+  out1 = (RTYPE)__msa_ilvl_w((v4i32)in2, (v4i32)in3);     \
+}
+#define ILVL_W2_UB(...) ILVL_W2(v16u8, __VA_ARGS__)
+#define ILVL_W2_SH(...) ILVL_W2(v8i16, __VA_ARGS__)
+
+/* Description : Interleave right half of byte elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of byte elements of 'in0' and 'in1' are interleaved
+                 and written to out0.
+*/
+#define ILVR_B2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvr_b((v16i8)in0, (v16i8)in1);     \
+  out1 = (RTYPE)__msa_ilvr_b((v16i8)in2, (v16i8)in3);     \
+}
+#define ILVR_B2_UB(...) ILVR_B2(v16u8, __VA_ARGS__)
+#define ILVR_B2_SB(...) ILVR_B2(v16i8, __VA_ARGS__)
+#define ILVR_B2_UH(...) ILVR_B2(v8u16, __VA_ARGS__)
+#define ILVR_B2_SH(...) ILVR_B2(v8i16, __VA_ARGS__)
+
+#define ILVR_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3) {                       \
+  ILVR_B2(RTYPE, in0, in1, in2, in3, out0, out1);               \
+  ILVR_B2(RTYPE, in4, in5, in6, in7, out2, out3);               \
+}
+#define ILVR_B4_UB(...) ILVR_B4(v16u8, __VA_ARGS__)
+#define ILVR_B4_SB(...) ILVR_B4(v16i8, __VA_ARGS__)
+#define ILVR_B4_UH(...) ILVR_B4(v8u16, __VA_ARGS__)
+#define ILVR_B4_SH(...) ILVR_B4(v8i16, __VA_ARGS__)
+
+#define ILVR_B8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,     \
+                in8, in9, in10, in11, in12, in13, in14, in15,      \
+                out0, out1, out2, out3, out4, out5, out6, out7) {  \
+  ILVR_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,           \
+          out0, out1, out2, out3);                                 \
+  ILVR_B4(RTYPE, in8, in9, in10, in11, in12, in13, in14, in15,     \
+          out4, out5, out6, out7);                                 \
+}
+#define ILVR_B8_UH(...) ILVR_B8(v8u16, __VA_ARGS__)
+
+/* Description : Interleave right half of halfword elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of halfword elements of 'in0' and 'in1' are
+                 interleaved and written to 'out0'.
+*/
+#define ILVR_H2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvr_h((v8i16)in0, (v8i16)in1);     \
+  out1 = (RTYPE)__msa_ilvr_h((v8i16)in2, (v8i16)in3);     \
+}
+#define ILVR_H2_SH(...) ILVR_H2(v8i16, __VA_ARGS__)
+
+#define ILVR_H4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3) {                       \
+  ILVR_H2(RTYPE, in0, in1, in2, in3, out0, out1);               \
+  ILVR_H2(RTYPE, in4, in5, in6, in7, out2, out3);               \
+}
+#define ILVR_H4_SH(...) ILVR_H4(v8i16, __VA_ARGS__)
+
+#define ILVR_W2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvr_w((v4i32)in0, (v4i32)in1);     \
+  out1 = (RTYPE)__msa_ilvr_w((v4i32)in2, (v4i32)in3);     \
+}
+#define ILVR_W2_UB(...) ILVR_W2(v16u8, __VA_ARGS__)
+#define ILVR_W2_SH(...) ILVR_W2(v8i16, __VA_ARGS__)
+
+#define ILVR_W4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3) {                       \
+  ILVR_W2(RTYPE, in0, in1, in2, in3, out0, out1);               \
+  ILVR_W2(RTYPE, in4, in5, in6, in7, out2, out3);               \
+}
+#define ILVR_W4_UB(...) ILVR_W4(v16u8, __VA_ARGS__)
+
+/* Description : Interleave right half of double word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of double word elements of 'in0' and 'in1' are
+                 interleaved and written to 'out0'.
+*/
+#define ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1) {   \
+  out0 = (RTYPE)__msa_ilvr_d((v2i64)(in0), (v2i64)(in1));  \
+  out1 = (RTYPE)__msa_ilvr_d((v2i64)(in2), (v2i64)(in3));  \
+}
+#define ILVR_D2_UB(...) ILVR_D2(v16u8, __VA_ARGS__)
+#define ILVR_D2_SB(...) ILVR_D2(v16i8, __VA_ARGS__)
+#define ILVR_D2_SH(...) ILVR_D2(v8i16, __VA_ARGS__)
+
+#define ILVR_D3(RTYPE, in0, in1, in2, in3, in4, in5, out0, out1, out2) {  \
+  ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1);                         \
+  out2 = (RTYPE)__msa_ilvr_d((v2i64)(in4), (v2i64)(in5));                 \
+}
+#define ILVR_D3_SB(...) ILVR_D3(v16i8, __VA_ARGS__)
+
+#define ILVR_D4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3) {                       \
+  ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1);               \
+  ILVR_D2(RTYPE, in4, in5, in6, in7, out2, out3);               \
+}
+#define ILVR_D4_SB(...) ILVR_D4(v16i8, __VA_ARGS__)
+#define ILVR_D4_UB(...) ILVR_D4(v16u8, __VA_ARGS__)
+
+/* Description : Interleave both left and right half of input vectors
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of byte elements from 'in0' and 'in1' are
+                 interleaved and written to 'out0'
+*/
+#define ILVRL_B2(RTYPE, in0, in1, out0, out1) {        \
+  out0 = (RTYPE)__msa_ilvr_b((v16i8)in0, (v16i8)in1);  \
+  out1 = (RTYPE)__msa_ilvl_b((v16i8)in0, (v16i8)in1);  \
+}
+#define ILVRL_B2_UB(...) ILVRL_B2(v16u8, __VA_ARGS__)
+#define ILVRL_B2_SB(...) ILVRL_B2(v16i8, __VA_ARGS__)
+#define ILVRL_B2_UH(...) ILVRL_B2(v8u16, __VA_ARGS__)
+#define ILVRL_B2_SH(...) ILVRL_B2(v8i16, __VA_ARGS__)
+
+#define ILVRL_H2(RTYPE, in0, in1, out0, out1) {        \
+  out0 = (RTYPE)__msa_ilvr_h((v8i16)in0, (v8i16)in1);  \
+  out1 = (RTYPE)__msa_ilvl_h((v8i16)in0, (v8i16)in1);  \
+}
+#define ILVRL_H2_SH(...) ILVRL_H2(v8i16, __VA_ARGS__)
+#define ILVRL_H2_SW(...) ILVRL_H2(v4i32, __VA_ARGS__)
+
+#define ILVRL_W2(RTYPE, in0, in1, out0, out1) {        \
+  out0 = (RTYPE)__msa_ilvr_w((v4i32)in0, (v4i32)in1);  \
+  out1 = (RTYPE)__msa_ilvl_w((v4i32)in0, (v4i32)in1);  \
+}
+#define ILVRL_W2_SH(...) ILVRL_W2(v8i16, __VA_ARGS__)
+#define ILVRL_W2_SW(...) ILVRL_W2(v4i32, __VA_ARGS__)
+
+/* Description : Saturate the halfword element values to the max
+                 unsigned value of (sat_val + 1) bits
+                 The element data width remains unchanged
+   Arguments   : Inputs  - in0, in1, sat_val
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each unsigned halfword element from 'in0' is saturated to the
+                 value generated with (sat_val + 1) bit range.
+                 The results are written in place
+*/
+#define SAT_UH2(RTYPE, in0, in1, sat_val) {         \
+  in0 = (RTYPE)__msa_sat_u_h((v8u16)in0, sat_val);  \
+  in1 = (RTYPE)__msa_sat_u_h((v8u16)in1, sat_val);  \
+}
+#define SAT_UH2_UH(...) SAT_UH2(v8u16, __VA_ARGS__)
+
+#define SAT_UH4(RTYPE, in0, in1, in2, in3, sat_val) {  \
+  SAT_UH2(RTYPE, in0, in1, sat_val);                   \
+  SAT_UH2(RTYPE, in2, in3, sat_val)                    \
+}
+#define SAT_UH4_UH(...) SAT_UH4(v8u16, __VA_ARGS__)
+
+/* Description : Saturate the halfword element values to the max
+                 unsigned value of (sat_val + 1) bits
+                 The element data width remains unchanged
+   Arguments   : Inputs  - in0, in1, sat_val
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each unsigned halfword element from 'in0' is saturated to the
+                 value generated with (sat_val + 1) bit range
+                 The results are written in place
+*/
+#define SAT_SH2(RTYPE, in0, in1, sat_val) {         \
+  in0 = (RTYPE)__msa_sat_s_h((v8i16)in0, sat_val);  \
+  in1 = (RTYPE)__msa_sat_s_h((v8i16)in1, sat_val);  \
+}
+#define SAT_SH2_SH(...) SAT_SH2(v8i16, __VA_ARGS__)
+
+#define SAT_SH4(RTYPE, in0, in1, in2, in3, sat_val) {  \
+  SAT_SH2(RTYPE, in0, in1, sat_val);                   \
+  SAT_SH2(RTYPE, in2, in3, sat_val);                   \
+}
+#define SAT_SH4_SH(...) SAT_SH4(v8i16, __VA_ARGS__)
+
+/* Description : Indexed halfword element values are replicated to all
+                 elements in output vector
+   Arguments   : Inputs  - in, idx0, idx1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : 'idx0' element value from 'in' vector is replicated to all
+                  elements in 'out0' vector
+                  Valid index range for halfword operation is 0-7
+*/
+#define SPLATI_H2(RTYPE, in, idx0, idx1, out0, out1) {  \
+  out0 = (RTYPE)__msa_splati_h((v8i16)in, idx0);        \
+  out1 = (RTYPE)__msa_splati_h((v8i16)in, idx1);        \
+}
+#define SPLATI_H2_SH(...) SPLATI_H2(v8i16, __VA_ARGS__)
+
+#define SPLATI_H4(RTYPE, in, idx0, idx1, idx2, idx3,  \
+                  out0, out1, out2, out3) {           \
+  SPLATI_H2(RTYPE, in, idx0, idx1, out0, out1);       \
+  SPLATI_H2(RTYPE, in, idx2, idx3, out2, out3);       \
+}
+#define SPLATI_H4_SB(...) SPLATI_H4(v16i8, __VA_ARGS__)
+#define SPLATI_H4_SH(...) SPLATI_H4(v8i16, __VA_ARGS__)
+
+/* Description : Pack even byte elements of vector pairs
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even byte elements of 'in0' are copied to the left half of
+                 'out0' & even byte elements of 'in1' are copied to the right
+                 half of 'out0'.
+*/
+#define PCKEV_B2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_pckev_b((v16i8)in0, (v16i8)in1);     \
+  out1 = (RTYPE)__msa_pckev_b((v16i8)in2, (v16i8)in3);     \
+}
+#define PCKEV_B2_SB(...) PCKEV_B2(v16i8, __VA_ARGS__)
+#define PCKEV_B2_UB(...) PCKEV_B2(v16u8, __VA_ARGS__)
+#define PCKEV_B2_SH(...) PCKEV_B2(v8i16, __VA_ARGS__)
+
+#define PCKEV_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                 out0, out1, out2, out3) {                       \
+  PCKEV_B2(RTYPE, in0, in1, in2, in3, out0, out1);               \
+  PCKEV_B2(RTYPE, in4, in5, in6, in7, out2, out3);               \
+}
+#define PCKEV_B4_SB(...) PCKEV_B4(v16i8, __VA_ARGS__)
+#define PCKEV_B4_UB(...) PCKEV_B4(v16u8, __VA_ARGS__)
+#define PCKEV_B4_SH(...) PCKEV_B4(v8i16, __VA_ARGS__)
+
+/* Description : Pack even halfword elements of vector pairs
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even halfword elements of 'in0' are copied to the left half of
+                 'out0' & even halfword elements of 'in1' are copied to the
+                 right half of 'out0'.
+*/
+#define PCKEV_H2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_pckev_h((v8i16)in0, (v8i16)in1);     \
+  out1 = (RTYPE)__msa_pckev_h((v8i16)in2, (v8i16)in3);     \
+}
+#define PCKEV_H2_SH(...) PCKEV_H2(v8i16, __VA_ARGS__)
+#define PCKEV_H2_SW(...) PCKEV_H2(v4i32, __VA_ARGS__)
+
+#define PCKEV_H4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                 out0, out1, out2, out3) {                       \
+  PCKEV_H2(RTYPE, in0, in1, in2, in3, out0, out1);               \
+  PCKEV_H2(RTYPE, in4, in5, in6, in7, out2, out3);               \
+}
+#define PCKEV_H4_SH(...) PCKEV_H4(v8i16, __VA_ARGS__)
+
+/* Description : Pack even double word elements of vector pairs
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even double elements of 'in0' are copied to the left half of
+                 'out0' & even double elements of 'in1' are copied to the right
+                 half of 'out0'.
+*/
+#define PCKEV_D2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_pckev_d((v2i64)in0, (v2i64)in1);     \
+  out1 = (RTYPE)__msa_pckev_d((v2i64)in2, (v2i64)in3);     \
+}
+#define PCKEV_D2_UB(...) PCKEV_D2(v16u8, __VA_ARGS__)
+#define PCKEV_D2_SH(...) PCKEV_D2(v8i16, __VA_ARGS__)
+
+#define PCKEV_D4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                 out0, out1, out2, out3) {                       \
+  PCKEV_D2(RTYPE, in0, in1, in2, in3, out0, out1);               \
+  PCKEV_D2(RTYPE, in4, in5, in6, in7, out2, out3);               \
+}
+#define PCKEV_D4_UB(...) PCKEV_D4(v16u8, __VA_ARGS__)
+
+/* Description : Each byte element is logically xor'ed with immediate 128
+   Arguments   : Inputs  - in0, in1
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each unsigned byte element from input vector 'in0' is
+                 logically xor'ed with 128 and the result is stored in-place.
+*/
+#define XORI_B2_128(RTYPE, in0, in1) {         \
+  in0 = (RTYPE)__msa_xori_b((v16u8)in0, 128);  \
+  in1 = (RTYPE)__msa_xori_b((v16u8)in1, 128);  \
+}
+#define XORI_B2_128_UB(...) XORI_B2_128(v16u8, __VA_ARGS__)
+#define XORI_B2_128_SB(...) XORI_B2_128(v16i8, __VA_ARGS__)
+
+#define XORI_B3_128(RTYPE, in0, in1, in2) {    \
+  XORI_B2_128(RTYPE, in0, in1);                \
+  in2 = (RTYPE)__msa_xori_b((v16u8)in2, 128);  \
+}
+#define XORI_B3_128_SB(...) XORI_B3_128(v16i8, __VA_ARGS__)
+
+#define XORI_B4_128(RTYPE, in0, in1, in2, in3) {  \
+  XORI_B2_128(RTYPE, in0, in1);                   \
+  XORI_B2_128(RTYPE, in2, in3);                   \
+}
+#define XORI_B4_128_UB(...) XORI_B4_128(v16u8, __VA_ARGS__)
+#define XORI_B4_128_SB(...) XORI_B4_128(v16i8, __VA_ARGS__)
+
+#define XORI_B7_128(RTYPE, in0, in1, in2, in3, in4, in5, in6) {  \
+  XORI_B4_128(RTYPE, in0, in1, in2, in3);                        \
+  XORI_B3_128(RTYPE, in4, in5, in6);                             \
+}
+#define XORI_B7_128_SB(...) XORI_B7_128(v16i8, __VA_ARGS__)
+
+/* Description : Average of signed halfword elements -> (a + b) / 2
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3
+                 Return Type - as per RTYPE
+   Details     : Each signed halfword element from 'in0' is added to each
+                 signed halfword element of 'in1' with full precision resulting
+                 in one extra bit in the result. The result is then divided by
+                 2 and written to 'out0'
+*/
+#define AVE_SH4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3) {                       \
+  out0 = (RTYPE)__msa_ave_s_h((v8i16)in0, (v8i16)in1);          \
+  out1 = (RTYPE)__msa_ave_s_h((v8i16)in2, (v8i16)in3);          \
+  out2 = (RTYPE)__msa_ave_s_h((v8i16)in4, (v8i16)in5);          \
+  out3 = (RTYPE)__msa_ave_s_h((v8i16)in6, (v8i16)in7);          \
+}
+#define AVE_SH4_SH(...) AVE_SH4(v8i16, __VA_ARGS__)
+
+/* Description : Addition of signed halfword elements and signed saturation
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed halfword elements from 'in0' are added to signed
+                 halfword elements of 'in1'. The result is then signed saturated
+                 between halfword data type range
+*/
+#define ADDS_SH2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_adds_s_h((v8i16)in0, (v8i16)in1);    \
+  out1 = (RTYPE)__msa_adds_s_h((v8i16)in2, (v8i16)in3);    \
+}
+#define ADDS_SH2_SH(...) ADDS_SH2(v8i16, __VA_ARGS__)
+
+#define ADDS_SH4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                 out0, out1, out2, out3) {                       \
+  ADDS_SH2(RTYPE, in0, in1, in2, in3, out0, out1);               \
+  ADDS_SH2(RTYPE, in4, in5, in6, in7, out2, out3);               \
+}
+#define ADDS_SH4_SH(...) ADDS_SH4(v8i16, __VA_ARGS__)
+
+/* Description : Shift left all elements of vector (generic for all data types)
+   Arguments   : Inputs  - in0, in1, in2, in3, shift
+                 Outputs - in place operation
+                 Return Type - as per input vector RTYPE
+   Details     : Each element of vector 'in0' is left shifted by 'shift' and
+                 the result is written in-place.
+*/
+#define SLLI_4V(in0, in1, in2, in3, shift) {  \
+  in0 = in0 << shift;                         \
+  in1 = in1 << shift;                         \
+  in2 = in2 << shift;                         \
+  in3 = in3 << shift;                         \
+}
+
+/* Description : Arithmetic shift right all elements of vector
+                 (generic for all data types)
+   Arguments   : Inputs  - in0, in1, in2, in3, shift
+                 Outputs - in place operation
+                 Return Type - as per input vector RTYPE
+   Details     : Each element of vector 'in0' is right shifted by 'shift' and
+                 the result is written in-place. 'shift' is a GP variable.
+*/
+#define SRA_4V(in0, in1, in2, in3, shift) {  \
+  in0 = in0 >> shift;                        \
+  in1 = in1 >> shift;                        \
+  in2 = in2 >> shift;                        \
+  in3 = in3 >> shift;                        \
+}
+
+/* Description : Shift right arithmetic rounded words
+   Arguments   : Inputs  - in0, in1, shift
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is shifted right arithmetically by
+                 the number of bits in the corresponding element in the vector
+                 'shift'. The last discarded bit is added to shifted value for
+                 rounding and the result is written in-place.
+                 'shift' is a vector.
+*/
+#define SRAR_W2(RTYPE, in0, in1, shift) {               \
+  in0 = (RTYPE)__msa_srar_w((v4i32)in0, (v4i32)shift);  \
+  in1 = (RTYPE)__msa_srar_w((v4i32)in1, (v4i32)shift);  \
+}
+
+#define SRAR_W4(RTYPE, in0, in1, in2, in3, shift) {  \
+  SRAR_W2(RTYPE, in0, in1, shift)                    \
+  SRAR_W2(RTYPE, in2, in3, shift)                    \
+}
+#define SRAR_W4_SW(...) SRAR_W4(v4i32, __VA_ARGS__)
+
+/* Description : Shift right arithmetic rounded (immediate)
+   Arguments   : Inputs  - in0, in1, shift
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is shifted right arithmetically by
+                 the value in 'shift'. The last discarded bit is added to the
+                 shifted value for rounding and the result is written in-place.
+                 'shift' is an immediate value.
+*/
+#define SRARI_H2(RTYPE, in0, in1, shift) {        \
+  in0 = (RTYPE)__msa_srari_h((v8i16)in0, shift);  \
+  in1 = (RTYPE)__msa_srari_h((v8i16)in1, shift);  \
+}
+#define SRARI_H2_UH(...) SRARI_H2(v8u16, __VA_ARGS__)
+#define SRARI_H2_SH(...) SRARI_H2(v8i16, __VA_ARGS__)
+
+#define SRARI_H4(RTYPE, in0, in1, in2, in3, shift) {  \
+  SRARI_H2(RTYPE, in0, in1, shift);                   \
+  SRARI_H2(RTYPE, in2, in3, shift);                   \
+}
+#define SRARI_H4_UH(...) SRARI_H4(v8u16, __VA_ARGS__)
+#define SRARI_H4_SH(...) SRARI_H4(v8i16, __VA_ARGS__)
+
+#define SRARI_W2(RTYPE, in0, in1, shift) {        \
+  in0 = (RTYPE)__msa_srari_w((v4i32)in0, shift);  \
+  in1 = (RTYPE)__msa_srari_w((v4i32)in1, shift);  \
+}
+#define SRARI_W2_SW(...) SRARI_W2(v4i32, __VA_ARGS__)
+
+#define SRARI_W4(RTYPE, in0, in1, in2, in3, shift) {  \
+  SRARI_W2(RTYPE, in0, in1, shift);                   \
+  SRARI_W2(RTYPE, in2, in3, shift);                   \
+}
+#define SRARI_W4_SW(...) SRARI_W4(v4i32, __VA_ARGS__)
+
+/* Description : Logical shift right all elements of vector (immediate)
+   Arguments   : Inputs  - in0, in1, in2, in3, shift
+                 Outputs - out0, out1, out2, out3
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is right shifted by 'shift' and
+                 the result is written in-place. 'shift' is an immediate value.
+*/
+#define SRLI_H4(RTYPE, in0, in1, in2, in3, out0, out1, out2, out3, shift) {  \
+  out0 = (RTYPE)__msa_srli_h((v8i16)in0, shift);                             \
+  out1 = (RTYPE)__msa_srli_h((v8i16)in1, shift);                             \
+  out2 = (RTYPE)__msa_srli_h((v8i16)in2, shift);                             \
+  out3 = (RTYPE)__msa_srli_h((v8i16)in3, shift);                             \
+}
+#define SRLI_H4_SH(...) SRLI_H4(v8i16, __VA_ARGS__)
+
+/* Description : Multiplication of pairs of vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+   Details     : Each element from 'in0' is multiplied with elements from 'in1'
+                 and the result is written to 'out0'
+*/
+#define MUL2(in0, in1, in2, in3, out0, out1) {  \
+  out0 = in0 * in1;                             \
+  out1 = in2 * in3;                             \
+}
+#define MUL4(in0, in1, in2, in3, in4, in5, in6, in7,  \
+             out0, out1, out2, out3) {                \
+  MUL2(in0, in1, in2, in3, out0, out1);               \
+  MUL2(in4, in5, in6, in7, out2, out3);               \
+}
+
+/* Description : Addition of 2 pairs of vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+   Details     : Each element in 'in0' is added to 'in1' and result is written
+                 to 'out0'.
+*/
+#define ADD2(in0, in1, in2, in3, out0, out1) {  \
+  out0 = in0 + in1;                             \
+  out1 = in2 + in3;                             \
+}
+#define ADD4(in0, in1, in2, in3, in4, in5, in6, in7,  \
+             out0, out1, out2, out3) {                \
+  ADD2(in0, in1, in2, in3, out0, out1);               \
+  ADD2(in4, in5, in6, in7, out2, out3);               \
+}
+
+/* Description : Subtraction of 2 pairs of vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+   Details     : Each element in 'in1' is subtracted from 'in0' and result is
+                 written to 'out0'.
+*/
+#define SUB2(in0, in1, in2, in3, out0, out1) {  \
+  out0 = in0 - in1;                             \
+  out1 = in2 - in3;                             \
+}
+#define SUB4(in0, in1, in2, in3, in4, in5, in6, in7,  \
+             out0, out1, out2, out3) {                \
+  out0 = in0 - in1;                                   \
+  out1 = in2 - in3;                                   \
+  out2 = in4 - in5;                                   \
+  out3 = in6 - in7;                                   \
+}
+
+/* Description : Sign extend halfword elements from right half of the vector
+   Arguments   : Input  - in    (halfword vector)
+                 Output - out   (sign extended word vector)
+                 Return Type - signed word
+   Details     : Sign bit of halfword elements from input vector 'in' is
+                 extracted and interleaved with same vector 'in0' to generate
+                 4 word elements keeping sign intact
+*/
+#define UNPCK_R_SH_SW(in, out) {                 \
+  v8i16 sign_m;                                  \
+                                                 \
+  sign_m = __msa_clti_s_h((v8i16)in, 0);         \
+  out = (v4i32)__msa_ilvr_h(sign_m, (v8i16)in);  \
+}
+
+/* Description : Zero extend unsigned byte elements to halfword elements
+   Arguments   : Input   - in          (unsigned byte vector)
+                 Outputs - out0, out1  (unsigned  halfword vectors)
+                 Return Type - signed halfword
+   Details     : Zero extended right half of vector is returned in 'out0'
+                 Zero extended left half of vector is returned in 'out1'
+*/
+#define UNPCK_UB_SH(in, out0, out1) {   \
+  v16i8 zero_m = { 0 };                 \
+                                        \
+  ILVRL_B2_SH(zero_m, in, out0, out1);  \
+}
+
+/* Description : Sign extend halfword elements from input vector and return
+                 the result in pair of vectors
+   Arguments   : Input   - in            (halfword vector)
+                 Outputs - out0, out1   (sign extended word vectors)
+                 Return Type - signed word
+   Details     : Sign bit of halfword elements from input vector 'in' is
+                 extracted and interleaved right with same vector 'in0' to
+                 generate 4 signed word elements in 'out0'
+                 Then interleaved left with same vector 'in0' to
+                 generate 4 signed word elements in 'out1'
+*/
+#define UNPCK_SH_SW(in, out0, out1) {    \
+  v8i16 tmp_m;                           \
+                                         \
+  tmp_m = __msa_clti_s_h((v8i16)in, 0);  \
+  ILVRL_H2_SW(tmp_m, in, out0, out1);    \
+}
+
+/* Description : Butterfly of 4 input vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1, out2, out3
+   Details     : Butterfly operation
+*/
+#define BUTTERFLY_4(in0, in1, in2, in3, out0, out1, out2, out3) {  \
+  out0 = in0 + in3;                                                \
+  out1 = in1 + in2;                                                \
+                                                                   \
+  out2 = in1 - in2;                                                \
+  out3 = in0 - in3;                                                \
+}
+
+/* Description : Butterfly of 8 input vectors
+   Arguments   : Inputs  - in0 ...  in7
+                 Outputs - out0 .. out7
+   Details     : Butterfly operation
+*/
+#define BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7,            \
+                    out0, out1, out2, out3, out4, out5, out6, out7) {  \
+  out0 = in0 + in7;                                                    \
+  out1 = in1 + in6;                                                    \
+  out2 = in2 + in5;                                                    \
+  out3 = in3 + in4;                                                    \
+                                                                       \
+  out4 = in3 - in4;                                                    \
+  out5 = in2 - in5;                                                    \
+  out6 = in1 - in6;                                                    \
+  out7 = in0 - in7;                                                    \
+}
+
+/* Description : Butterfly of 16 input vectors
+   Arguments   : Inputs  - in0 ...  in15
+                 Outputs - out0 .. out15
+   Details     : Butterfly operation
+*/
+#define BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7,                  \
+                     in8, in9,  in10, in11, in12, in13, in14, in15,           \
+                     out0, out1, out2, out3, out4, out5, out6, out7,          \
+                     out8, out9, out10, out11, out12, out13, out14, out15) {  \
+  out0 = in0 + in15;                                                          \
+  out1 = in1 + in14;                                                          \
+  out2 = in2 + in13;                                                          \
+  out3 = in3 + in12;                                                          \
+  out4 = in4 + in11;                                                          \
+  out5 = in5 + in10;                                                          \
+  out6 = in6 + in9;                                                           \
+  out7 = in7 + in8;                                                           \
+                                                                              \
+  out8 = in7 - in8;                                                           \
+  out9 = in6 - in9;                                                           \
+  out10 = in5 - in10;                                                         \
+  out11 = in4 - in11;                                                         \
+  out12 = in3 - in12;                                                         \
+  out13 = in2 - in13;                                                         \
+  out14 = in1 - in14;                                                         \
+  out15 = in0 - in15;                                                         \
+}
+
+/* Description : Transpose input 8x8 byte block
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - as per RTYPE
+*/
+#define TRANSPOSE8x8_UB(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,     \
+                        out0, out1, out2, out3, out4, out5, out6, out7) {  \
+  v16i8 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                    \
+  v16i8 tmp4_m, tmp5_m, tmp6_m, tmp7_m;                                    \
+                                                                           \
+  ILVR_B4_SB(in2, in0, in3, in1, in6, in4, in7, in5,                       \
+             tmp0_m, tmp1_m, tmp2_m, tmp3_m);                              \
+  ILVRL_B2_SB(tmp1_m, tmp0_m, tmp4_m, tmp5_m);                             \
+  ILVRL_B2_SB(tmp3_m, tmp2_m, tmp6_m, tmp7_m);                             \
+  ILVRL_W2(RTYPE, tmp6_m, tmp4_m, out0, out2);                             \
+  ILVRL_W2(RTYPE, tmp7_m, tmp5_m, out4, out6);                             \
+  SLDI_B2_0(RTYPE, out0, out2, out1, out3, 8);                             \
+  SLDI_B2_0(RTYPE, out4, out6, out5, out7, 8);                             \
+}
+#define TRANSPOSE8x8_UB_UB(...) TRANSPOSE8x8_UB(v16u8, __VA_ARGS__)
+
+/* Description : Transpose 16x8 block into 8x16 with byte elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7,
+                           in8, in9, in10, in11, in12, in13, in14, in15
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - unsigned byte
+*/
+#define TRANSPOSE16x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7,            \
+                            in8, in9, in10, in11, in12, in13, in14, in15,      \
+                            out0, out1, out2, out3, out4, out5, out6, out7) {  \
+  v16u8 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                        \
+  v16u8 tmp4_m, tmp5_m, tmp6_m, tmp7_m;                                        \
+                                                                               \
+  ILVEV_D2_UB(in0, in8, in1, in9, out7, out6);                                 \
+  ILVEV_D2_UB(in2, in10, in3, in11, out5, out4);                               \
+  ILVEV_D2_UB(in4, in12, in5, in13, out3, out2);                               \
+  ILVEV_D2_UB(in6, in14, in7, in15, out1, out0);                               \
+                                                                               \
+  tmp0_m = (v16u8)__msa_ilvev_b((v16i8)out6, (v16i8)out7);                     \
+  tmp4_m = (v16u8)__msa_ilvod_b((v16i8)out6, (v16i8)out7);                     \
+  tmp1_m = (v16u8)__msa_ilvev_b((v16i8)out4, (v16i8)out5);                     \
+  tmp5_m = (v16u8)__msa_ilvod_b((v16i8)out4, (v16i8)out5);                     \
+  out5 = (v16u8)__msa_ilvev_b((v16i8)out2, (v16i8)out3);                       \
+  tmp6_m = (v16u8)__msa_ilvod_b((v16i8)out2, (v16i8)out3);                     \
+  out7 = (v16u8)__msa_ilvev_b((v16i8)out0, (v16i8)out1);                       \
+  tmp7_m = (v16u8)__msa_ilvod_b((v16i8)out0, (v16i8)out1);                     \
+                                                                               \
+  ILVEV_H2_UB(tmp0_m, tmp1_m, out5, out7, tmp2_m, tmp3_m);                     \
+  out0 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);                   \
+  out4 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);                   \
+                                                                               \
+  tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp1_m, (v8i16)tmp0_m);                 \
+  tmp3_m = (v16u8)__msa_ilvod_h((v8i16)out7, (v8i16)out5);                     \
+  out2 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);                   \
+  out6 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);                   \
+                                                                               \
+  ILVEV_H2_UB(tmp4_m, tmp5_m, tmp6_m, tmp7_m, tmp2_m, tmp3_m);                 \
+  out1 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);                   \
+  out5 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);                   \
+                                                                               \
+  tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp5_m, (v8i16)tmp4_m);                 \
+  tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp5_m, (v8i16)tmp4_m);                 \
+  tmp3_m = (v16u8)__msa_ilvod_h((v8i16)tmp7_m, (v8i16)tmp6_m);                 \
+  tmp3_m = (v16u8)__msa_ilvod_h((v8i16)tmp7_m, (v8i16)tmp6_m);                 \
+  out3 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);                   \
+  out7 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);                   \
+}
+
+/* Description : Transpose 4x4 block with half word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1, out2, out3
+                 Return Type - signed halfword
+*/
+#define TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, out0, out1, out2, out3) {  \
+  v8i16 s0_m, s1_m;                                                       \
+                                                                          \
+  ILVR_H2_SH(in1, in0, in3, in2, s0_m, s1_m);                             \
+  ILVRL_W2_SH(s1_m, s0_m, out0, out2);                                    \
+  out1 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out0);                   \
+  out3 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out2);                   \
+}
+
+/* Description : Transpose 4x8 block with half word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - signed halfword
+*/
+#define TRANSPOSE4X8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,            \
+                           out0, out1, out2, out3, out4, out5, out6, out7) {  \
+  v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                       \
+  v8i16 tmp0_n, tmp1_n, tmp2_n, tmp3_n;                                       \
+  v8i16 zero_m = { 0 };                                                       \
+                                                                              \
+  ILVR_H4_SH(in1, in0, in3, in2, in5, in4, in7, in6,                          \
+             tmp0_n, tmp1_n, tmp2_n, tmp3_n);                                 \
+  ILVRL_W2_SH(tmp1_n, tmp0_n, tmp0_m, tmp2_m);                                \
+  ILVRL_W2_SH(tmp3_n, tmp2_n, tmp1_m, tmp3_m);                                \
+                                                                              \
+  out0 = (v8i16)__msa_ilvr_d((v2i64)tmp1_m, (v2i64)tmp0_m);                   \
+  out1 = (v8i16)__msa_ilvl_d((v2i64)tmp1_m, (v2i64)tmp0_m);                   \
+  out2 = (v8i16)__msa_ilvr_d((v2i64)tmp3_m, (v2i64)tmp2_m);                   \
+  out3 = (v8i16)__msa_ilvl_d((v2i64)tmp3_m, (v2i64)tmp2_m);                   \
+                                                                              \
+  out4 = zero_m;                                                              \
+  out5 = zero_m;                                                              \
+  out6 = zero_m;                                                              \
+  out7 = zero_m;                                                              \
+}
+
+/* Description : Transpose 8x4 block with half word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - signed halfword
+*/
+#define TRANSPOSE8X4_SH_SH(in0, in1, in2, in3, out0, out1, out2, out3) {  \
+  v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                   \
+                                                                          \
+  ILVR_H2_SH(in1, in0, in3, in2, tmp0_m, tmp1_m);                         \
+  ILVL_H2_SH(in1, in0, in3, in2, tmp2_m, tmp3_m);                         \
+  ILVR_W2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out0, out2);                 \
+  ILVL_W2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out1, out3);                 \
+}
+
+/* Description : Transpose 8x8 block with half word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - as per RTYPE
+*/
+#define TRANSPOSE8x8_H(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,     \
+                       out0, out1, out2, out3, out4, out5, out6, out7) {  \
+  v8i16 s0_m, s1_m;                                                       \
+  v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                   \
+  v8i16 tmp4_m, tmp5_m, tmp6_m, tmp7_m;                                   \
+                                                                          \
+  ILVR_H2_SH(in6, in4, in7, in5, s0_m, s1_m);                             \
+  ILVRL_H2_SH(s1_m, s0_m, tmp0_m, tmp1_m);                                \
+  ILVL_H2_SH(in6, in4, in7, in5, s0_m, s1_m);                             \
+  ILVRL_H2_SH(s1_m, s0_m, tmp2_m, tmp3_m);                                \
+  ILVR_H2_SH(in2, in0, in3, in1, s0_m, s1_m);                             \
+  ILVRL_H2_SH(s1_m, s0_m, tmp4_m, tmp5_m);                                \
+  ILVL_H2_SH(in2, in0, in3, in1, s0_m, s1_m);                             \
+  ILVRL_H2_SH(s1_m, s0_m, tmp6_m, tmp7_m);                                \
+  PCKEV_D4(RTYPE, tmp0_m, tmp4_m, tmp1_m, tmp5_m, tmp2_m, tmp6_m,         \
+           tmp3_m, tmp7_m, out0, out2, out4, out6);                       \
+  out1 = (RTYPE)__msa_pckod_d((v2i64)tmp0_m, (v2i64)tmp4_m);              \
+  out3 = (RTYPE)__msa_pckod_d((v2i64)tmp1_m, (v2i64)tmp5_m);              \
+  out5 = (RTYPE)__msa_pckod_d((v2i64)tmp2_m, (v2i64)tmp6_m);              \
+  out7 = (RTYPE)__msa_pckod_d((v2i64)tmp3_m, (v2i64)tmp7_m);              \
+}
+#define TRANSPOSE8x8_SH_SH(...) TRANSPOSE8x8_H(v8i16, __VA_ARGS__)
+
+/* Description : Transpose 4x4 block with word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1, out2, out3
+                 Return Type - signed word
+*/
+#define TRANSPOSE4x4_SW_SW(in0, in1, in2, in3, out0, out1, out2, out3) {  \
+  v4i32 s0_m, s1_m, s2_m, s3_m;                                           \
+                                                                          \
+  ILVRL_W2_SW(in1, in0, s0_m, s1_m);                                      \
+  ILVRL_W2_SW(in3, in2, s2_m, s3_m);                                      \
+                                                                          \
+  out0 = (v4i32)__msa_ilvr_d((v2i64)s2_m, (v2i64)s0_m);                   \
+  out1 = (v4i32)__msa_ilvl_d((v2i64)s2_m, (v2i64)s0_m);                   \
+  out2 = (v4i32)__msa_ilvr_d((v2i64)s3_m, (v2i64)s1_m);                   \
+  out3 = (v4i32)__msa_ilvl_d((v2i64)s3_m, (v2i64)s1_m);                   \
+}
+
+/* Description : Add block 4x4
+   Arguments   : Inputs - in0, in1, in2, in3, pdst, stride
+   Details     : Least significant 4 bytes from each input vector are added to
+                 the destination bytes, clipped between 0-255 and stored.
+*/
+#define ADDBLK_ST4x4_UB(in0, in1, in2, in3, pdst, stride) {     \
+  uint32_t src0_m, src1_m, src2_m, src3_m;                      \
+  v8i16 inp0_m, inp1_m, res0_m, res1_m;                         \
+  v16i8 dst0_m = { 0 };                                         \
+  v16i8 dst1_m = { 0 };                                         \
+  v16i8 zero_m = { 0 };                                         \
+                                                                \
+  ILVR_D2_SH(in1, in0, in3, in2, inp0_m, inp1_m)                \
+  LW4(pdst, stride,  src0_m, src1_m, src2_m, src3_m);           \
+  INSERT_W2_SB(src0_m, src1_m, dst0_m);                         \
+  INSERT_W2_SB(src2_m, src3_m, dst1_m);                         \
+  ILVR_B2_SH(zero_m, dst0_m, zero_m, dst1_m, res0_m, res1_m);   \
+  ADD2(res0_m, inp0_m, res1_m, inp1_m, res0_m, res1_m);         \
+  CLIP_SH2_0_255(res0_m, res1_m);                               \
+  PCKEV_B2_SB(res0_m, res0_m, res1_m, res1_m, dst0_m, dst1_m);  \
+  ST4x4_UB(dst0_m, dst1_m, 0, 1, 0, 1, pdst, stride);           \
+}
+
+/* Description : Pack even elements of input vectors & xor with 128
+   Arguments   : Inputs - in0, in1
+                 Output - out_m
+                 Return Type - unsigned byte
+   Details     : Signed byte even elements from 'in0' and 'in1' are packed
+                 together in one vector and the resulting vector is xor'ed with
+                 128 to shift the range from signed to unsigned byte
+*/
+#define PCKEV_XORI128_UB(in0, in1) ({                    \
+  v16u8 out_m;                                           \
+                                                         \
+  out_m = (v16u8)__msa_pckev_b((v16i8)in1, (v16i8)in0);  \
+  out_m = (v16u8)__msa_xori_b((v16u8)out_m, 128);        \
+  out_m;                                                 \
+})
+
+/* Description : Converts inputs to unsigned bytes, interleave, average & store
+                 as 8x4 unsigned byte block
+   Arguments   : Inputs - in0, in1, in2, in3, dst0, dst1, dst2, dst3,
+                          pdst, stride
+*/
+#define CONVERT_UB_AVG_ST8x4_UB(in0, in1, in2, in3,                      \
+                                dst0, dst1, dst2, dst3, pdst, stride) {  \
+  v16u8 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                  \
+  uint8_t *pdst_m = (uint8_t *)(pdst);                                   \
+                                                                         \
+  tmp0_m = PCKEV_XORI128_UB(in0, in1);                                   \
+  tmp1_m = PCKEV_XORI128_UB(in2, in3);                                   \
+  ILVR_D2_UB(dst1, dst0, dst3, dst2, tmp2_m, tmp3_m);                    \
+  AVER_UB2_UB(tmp0_m, tmp2_m, tmp1_m, tmp3_m, tmp0_m, tmp1_m);           \
+  ST8x4_UB(tmp0_m, tmp1_m, pdst_m, stride);                              \
+}
+
+/* Description : Pack even byte elements and store byte vector in destination
+                 memory
+   Arguments   : Inputs - in0, in1, pdst
+*/
+#define PCKEV_ST_SB(in0, in1, pdst) {             \
+  v16i8 tmp_m;                                    \
+                                                  \
+  tmp_m = __msa_pckev_b((v16i8)in1, (v16i8)in0);  \
+  ST_SB(tmp_m, (pdst));                           \
+}
+
+/* Description : Horizontal 2 tap filter kernel code
+   Arguments   : Inputs - in0, in1, mask, coeff, shift
+*/
+#define HORIZ_2TAP_FILT_UH(in0, in1, mask, coeff, shift) ({    \
+  v16i8 tmp0_m;                                                \
+  v8u16 tmp1_m;                                                \
+                                                               \
+  tmp0_m = __msa_vshf_b((v16i8)mask, (v16i8)in1, (v16i8)in0);  \
+  tmp1_m = __msa_dotp_u_h((v16u8)tmp0_m, (v16u8)coeff);        \
+  tmp1_m = (v8u16)__msa_srari_h((v8i16)tmp1_m, shift);         \
+                                                               \
+  tmp1_m;                                                      \
+})
+#endif  /* VPX_DSP_MIPS_MACROS_MSA_H_ */
diff --git a/libvpx/libvpx/vpx_dsp/mips/sad_msa.c b/libvpx/libvpx/vpx_dsp/mips/sad_msa.c
new file mode 100644
index 0000000..3bdec28
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/sad_msa.c
@@ -0,0 +1,1525 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/macros_msa.h"
+
+#define SAD_INSVE_W4(RTYPE, in0, in1, in2, in3, out) {    \
+  out = (RTYPE)__msa_insve_w((v4i32)out, 0, (v4i32)in0);  \
+  out = (RTYPE)__msa_insve_w((v4i32)out, 1, (v4i32)in1);  \
+  out = (RTYPE)__msa_insve_w((v4i32)out, 2, (v4i32)in2);  \
+  out = (RTYPE)__msa_insve_w((v4i32)out, 3, (v4i32)in3);  \
+}
+#define SAD_INSVE_W4_UB(...) SAD_INSVE_W4(v16u8, __VA_ARGS__)
+
+static uint32_t sad_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 diff;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad += __msa_hadd_u_h(diff, diff);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_8width_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2,
+                src0, src1, ref0, ref1);
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_16width_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *ref, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(src, src_stride, src0, src1);
+    src += (2 * src_stride);
+    LD_UB2(ref, ref_stride, ref0, ref1);
+    ref += (2 * ref_stride);
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, src_stride, src0, src1);
+    src += (2 * src_stride);
+    LD_UB2(ref, ref_stride, ref0, ref1);
+    ref += (2 * ref_stride);
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_32width_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *ref, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_64width_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *ref, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  uint32_t sad = 0;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+
+  for (ht_cnt = (height >> 1); ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad = HADD_UH_U32(sad0);
+  sad += HADD_UH_U32(sad1);
+
+  return sad;
+}
+
+static void sad_4width_x3_msa(const uint8_t *src_ptr, int32_t src_stride,
+                              const uint8_t *ref_ptr, int32_t ref_stride,
+                              int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 ref0, ref1, ref2, ref3, diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+
+    LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+}
+
+static void sad_8width_x3_msa(const uint8_t *src, int32_t src_stride,
+                              const uint8_t *ref, int32_t ref_stride,
+                              int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref00, ref11, ref22, ref33;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref00, ref11, ref22, ref33);
+    ref += (4 * ref_stride);
+    PCKEV_D4_UB(src1, src0, src3, src2, ref11, ref00, ref33, ref22,
+                src0, src1, ref0, ref1);
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+}
+
+static void sad_16width_x3_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src, ref, ref0, ref1, diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+
+  for (ht_cnt = (height >> 1); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+}
+
+static void sad_32width_x3_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0_0, ref0_1, ref0_2, ref0, ref1;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+
+  for (ht_cnt = height >> 1; ht_cnt--;) {
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB3(ref, 16, ref0_0, ref0_1, ref0_2);
+    ref += ref_stride;
+
+    sad0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB3(ref, 16, ref0_0, ref0_1, ref0_2);
+    ref += ref_stride;
+
+    sad0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+}
+
+static void sad_64width_x3_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0_0, ref0_1, ref0_2, ref0_3, ref0_4, ref0, ref1, ref2, ref3;
+  v8u16 sad0_0 = { 0 };
+  v8u16 sad0_1 = { 0 };
+  v8u16 sad1_0 = { 0 };
+  v8u16 sad1_1 = { 0 };
+  v8u16 sad2_0 = { 0 };
+  v8u16 sad2_1 = { 0 };
+  v4u32 sad;
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0_0, ref0_1, ref0_2, ref0_3);
+    ref0_4 = LD_UB(ref + 64);
+    ref += ref_stride;
+
+    sad0_0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
+    sad0_1 += SAD_UB2_UH(src2, src3, ref0_2, ref0_3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 1);
+    sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 2);
+    sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad = __msa_hadd_u_w(sad0_0, sad0_0);
+  sad += __msa_hadd_u_w(sad0_1, sad0_1);
+  sad_array[0] = HADD_SW_S32((v4i32)sad);
+
+  sad = __msa_hadd_u_w(sad1_0, sad1_0);
+  sad += __msa_hadd_u_w(sad1_1, sad1_1);
+  sad_array[1] = HADD_SW_S32((v4i32)sad);
+
+  sad = __msa_hadd_u_w(sad2_0, sad2_0);
+  sad += __msa_hadd_u_w(sad2_1, sad2_1);
+  sad_array[2] = HADD_SW_S32((v4i32)sad);
+}
+
+static void sad_4width_x8_msa(const uint8_t *src_ptr, int32_t src_stride,
+                              const uint8_t *ref_ptr, int32_t ref_stride,
+                              int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3, diff;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+  v8u16 sad4 = { 0 };
+  v8u16 sad5 = { 0 };
+  v8u16 sad6 = { 0 };
+  v8u16 sad7 = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad3 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad4 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad5 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad6 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad7 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+  sad_array[4] = HADD_UH_U32(sad4);
+  sad_array[5] = HADD_UH_U32(sad5);
+  sad_array[6] = HADD_UH_U32(sad6);
+  sad_array[7] = HADD_UH_U32(sad7);
+}
+
+static void sad_8width_x8_msa(const uint8_t *src, int32_t src_stride,
+                              const uint8_t *ref, int32_t ref_stride,
+                              int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref00, ref11, ref22, ref33;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+  v8u16 sad4 = { 0 };
+  v8u16 sad5 = { 0 };
+  v8u16 sad6 = { 0 };
+  v8u16 sad7 = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref00, ref11, ref22, ref33);
+    ref += (4 * ref_stride);
+    PCKEV_D4_UB(src1, src0, src3, src2, ref11, ref00, ref33, ref22,
+                src0, src1, ref0, ref1);
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad4 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad5 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad6 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad7 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+  sad_array[4] = HADD_UH_U32(sad4);
+  sad_array[5] = HADD_UH_U32(sad5);
+  sad_array[6] = HADD_UH_U32(sad6);
+  sad_array[7] = HADD_UH_U32(sad7);
+}
+
+static void sad_16width_x8_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src, ref0, ref1, ref;
+  v16u8 diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+  v8u16 sad4 = { 0 };
+  v8u16 sad5 = { 0 };
+  v8u16 sad6 = { 0 };
+  v8u16 sad7 = { 0 };
+
+  for (ht_cnt = (height >> 1); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 3);
+    diff = __msa_asub_u_b(src, ref);
+    sad3 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 4);
+    diff = __msa_asub_u_b(src, ref);
+    sad4 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 5);
+    diff = __msa_asub_u_b(src, ref);
+    sad5 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 6);
+    diff = __msa_asub_u_b(src, ref);
+    sad6 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 7);
+    diff = __msa_asub_u_b(src, ref);
+    sad7 += __msa_hadd_u_h(diff, diff);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 3);
+    diff = __msa_asub_u_b(src, ref);
+    sad3 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 4);
+    diff = __msa_asub_u_b(src, ref);
+    sad4 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 5);
+    diff = __msa_asub_u_b(src, ref);
+    sad5 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 6);
+    diff = __msa_asub_u_b(src, ref);
+    sad6 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 7);
+    diff = __msa_asub_u_b(src, ref);
+    sad7 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+  sad_array[4] = HADD_UH_U32(sad4);
+  sad_array[5] = HADD_UH_U32(sad5);
+  sad_array[6] = HADD_UH_U32(sad6);
+  sad_array[7] = HADD_UH_U32(sad7);
+}
+
+static void sad_32width_x8_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src0, src1;
+  v16u8 ref0, ref1, ref0_0, ref0_1, ref0_2;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+  v8u16 sad4 = { 0 };
+  v8u16 sad5 = { 0 };
+  v8u16 sad6 = { 0 };
+  v8u16 sad7 = { 0 };
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB3(ref, 16, ref0_0, ref0_1, ref0_2);
+    ref += ref_stride;
+
+    sad0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 3);
+    sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 4);
+    sad4 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 5);
+    sad5 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 6);
+    sad6 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 7);
+    sad7 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+  sad_array[4] = HADD_UH_U32(sad4);
+  sad_array[5] = HADD_UH_U32(sad5);
+  sad_array[6] = HADD_UH_U32(sad6);
+  sad_array[7] = HADD_UH_U32(sad7);
+}
+
+static void sad_64width_x8_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  const uint8_t *src_dup, *ref_dup;
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0_0, ref0_1, ref0_2, ref0_3, ref0_4;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 sad0_0 = { 0 };
+  v8u16 sad0_1 = { 0 };
+  v8u16 sad1_0 = { 0 };
+  v8u16 sad1_1 = { 0 };
+  v8u16 sad2_0 = { 0 };
+  v8u16 sad2_1 = { 0 };
+  v8u16 sad3_0 = { 0 };
+  v8u16 sad3_1 = { 0 };
+  v4u32 sad;
+
+  src_dup = src;
+  ref_dup = ref;
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB5(ref, 16, ref0_0, ref0_1, ref0_2, ref0_3, ref0_4);
+    ref += ref_stride;
+
+    sad0_0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
+    sad0_1 += SAD_UB2_UH(src2, src3, ref0_2, ref0_3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 1);
+    sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 2);
+    sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 3);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 3);
+    sad3_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad3_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad = __msa_hadd_u_w(sad0_0, sad0_0);
+  sad += __msa_hadd_u_w(sad0_1, sad0_1);
+  sad_array[0] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad1_0, sad1_0);
+  sad += __msa_hadd_u_w(sad1_1, sad1_1);
+  sad_array[1] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad2_0, sad2_0);
+  sad += __msa_hadd_u_w(sad2_1, sad2_1);
+  sad_array[2] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad3_0, sad3_0);
+  sad += __msa_hadd_u_w(sad3_1, sad3_1);
+  sad_array[3] = HADD_SW_S32(sad);
+
+  sad0_0 = (v8u16)__msa_ldi_h(0);
+  sad0_1 = (v8u16)__msa_ldi_h(0);
+  sad1_0 = (v8u16)__msa_ldi_h(0);
+  sad1_1 = (v8u16)__msa_ldi_h(0);
+  sad2_0 = (v8u16)__msa_ldi_h(0);
+  sad2_1 = (v8u16)__msa_ldi_h(0);
+  sad3_0 = (v8u16)__msa_ldi_h(0);
+  sad3_1 = (v8u16)__msa_ldi_h(0);
+
+  for (ht_cnt = 64; ht_cnt--;) {
+    LD_UB4(src_dup, 16, src0, src1, src2, src3);
+    src_dup += src_stride;
+    LD_UB5(ref_dup, 16, ref0_0, ref0_1, ref0_2, ref0_3, ref0_4);
+    ref_dup += ref_stride;
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 4);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 4);
+    sad0_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad0_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 5);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 5);
+    sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 6);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 6);
+    sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 7);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 7);
+    sad3_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad3_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad = __msa_hadd_u_w(sad0_0, sad0_0);
+  sad += __msa_hadd_u_w(sad0_1, sad0_1);
+  sad_array[4] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad1_0, sad1_0);
+  sad += __msa_hadd_u_w(sad1_1, sad1_1);
+  sad_array[5] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad2_0, sad2_0);
+  sad += __msa_hadd_u_w(sad2_1, sad2_1);
+  sad_array[6] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad3_0, sad3_0);
+  sad += __msa_hadd_u_w(sad3_1, sad3_1);
+  sad_array[7] = HADD_SW_S32(sad);
+}
+
+static void sad_4width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t * const aref_ptr[],
+                               int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    src_ptr += (4 * src_stride);
+
+    LW4(ref0_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref0_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    LW4(ref1_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref1_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    LW4(ref2_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref2_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    LW4(ref3_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref3_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad3 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_8width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t * const aref_ptr[],
+                               int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
+  v16u8 ref8, ref9, ref10, ref11, ref12, ref13, ref14, ref15;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref0_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref0_ptr += (4 * ref_stride);
+    LD_UB4(ref1_ptr, ref_stride, ref4, ref5, ref6, ref7);
+    ref1_ptr += (4 * ref_stride);
+    LD_UB4(ref2_ptr, ref_stride, ref8, ref9, ref10, ref11);
+    ref2_ptr += (4 * ref_stride);
+    LD_UB4(ref3_ptr, ref_stride, ref12, ref13, ref14, ref15);
+    ref3_ptr += (4 * ref_stride);
+
+    PCKEV_D2_UB(src1, src0, src3, src2, src0, src1);
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    PCKEV_D2_UB(ref5, ref4, ref7, ref6, ref0, ref1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    PCKEV_D2_UB(ref9, ref8, ref11, ref10, ref0, ref1);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    PCKEV_D2_UB(ref13, ref12, ref15, ref14, ref0, ref1);
+    sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_16width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                const uint8_t * const aref_ptr[],
+                                int32_t ref_stride,
+                                int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  v16u8 src, ref0, ref1, ref2, ref3, diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = (height >> 1); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref0 = LD_UB(ref0_ptr);
+    ref0_ptr += ref_stride;
+    ref1 = LD_UB(ref1_ptr);
+    ref1_ptr += ref_stride;
+    ref2 = LD_UB(ref2_ptr);
+    ref2_ptr += ref_stride;
+    ref3 = LD_UB(ref3_ptr);
+    ref3_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref1);
+    sad1 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref2);
+    sad2 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref3);
+    sad3 += __msa_hadd_u_h(diff, diff);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref0 = LD_UB(ref0_ptr);
+    ref0_ptr += ref_stride;
+    ref1 = LD_UB(ref1_ptr);
+    ref1_ptr += ref_stride;
+    ref2 = LD_UB(ref2_ptr);
+    ref2_ptr += ref_stride;
+    ref3 = LD_UB(ref3_ptr);
+    ref3_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref1);
+    sad1 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref2);
+    sad2 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref3);
+    sad3 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_32width_x4d_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t * const aref_ptr[],
+                                int32_t ref_stride,
+                                int32_t height, uint32_t *sad_array) {
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+
+    LD_UB2(ref0_ptr, 16, ref0, ref1);
+    ref0_ptr += ref_stride;
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(ref1_ptr, 16, ref0, ref1);
+    ref1_ptr += ref_stride;
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(ref2_ptr, 16, ref0, ref1);
+    ref2_ptr += ref_stride;
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(ref3_ptr, 16, ref0, ref1);
+    ref3_ptr += ref_stride;
+    sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_64width_x4d_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t * const aref_ptr[],
+                                int32_t ref_stride,
+                                int32_t height, uint32_t *sad_array) {
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 sad0_0 = { 0 };
+  v8u16 sad0_1 = { 0 };
+  v8u16 sad1_0 = { 0 };
+  v8u16 sad1_1 = { 0 };
+  v8u16 sad2_0 = { 0 };
+  v8u16 sad2_1 = { 0 };
+  v8u16 sad3_0 = { 0 };
+  v8u16 sad3_1 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+
+    LD_UB4(ref0_ptr, 16, ref0, ref1, ref2, ref3);
+    ref0_ptr += ref_stride;
+    sad0_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad0_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(ref1_ptr, 16, ref0, ref1, ref2, ref3);
+    ref1_ptr += ref_stride;
+    sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(ref2_ptr, 16, ref0, ref1, ref2, ref3);
+    ref2_ptr += ref_stride;
+    sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(ref3_ptr, 16, ref0, ref1, ref2, ref3);
+    ref3_ptr += ref_stride;
+    sad3_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad3_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0_0);
+  sad_array[0] += HADD_UH_U32(sad0_1);
+  sad_array[1] = HADD_UH_U32(sad1_0);
+  sad_array[1] += HADD_UH_U32(sad1_1);
+  sad_array[2] = HADD_UH_U32(sad2_0);
+  sad_array[2] += HADD_UH_U32(sad2_1);
+  sad_array[3] = HADD_UH_U32(sad3_0);
+  sad_array[3] += HADD_UH_U32(sad3_1);
+}
+
+static uint32_t avgsad_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                  const uint8_t *ref_ptr, int32_t ref_stride,
+                                  int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 diff, pred, comp;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+
+    comp = __msa_aver_u_b(pred, ref);
+    diff = __msa_asub_u_b(src, comp);
+    sad += __msa_hadd_u_h(diff, diff);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_8width_msa(const uint8_t *src, int32_t src_stride,
+                                  const uint8_t *ref, int32_t ref_stride,
+                                  int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 diff0, diff1, pred0, pred1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2,
+                src0, src1, ref0, ref1);
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, diff0, diff1);
+    sad += SAD_UB2_UH(src0, src1, diff0, diff1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_16width_msa(const uint8_t *src, int32_t src_stride,
+                                   const uint8_t *ref, int32_t ref_stride,
+                                   int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3, comp0, comp1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 3); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * 16);
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1);
+    sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1);
+    sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * 16);
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1);
+    sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1);
+    sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_32width_msa(const uint8_t *src, int32_t src_stride,
+                                   const uint8_t *ref, int32_t ref_stride,
+                                   int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
+  v16u8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7;
+  v16u8 comp0, comp1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src2, src4, src6);
+    LD_UB4(src + 16, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+
+    LD_UB4(ref, ref_stride, ref0, ref2, ref4, ref6);
+    LD_UB4(ref + 16, ref_stride, ref1, ref3, ref5, ref7);
+    ref += (4 * ref_stride);
+
+    LD_UB4(sec_pred, 32, pred0, pred2, pred4, pred6);
+    LD_UB4(sec_pred + 16, 32, pred1, pred3, pred5, pred7);
+    sec_pred += (4 * 32);
+
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1);
+    sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1);
+    sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+    AVER_UB2_UB(pred4, ref4, pred5, ref5, comp0, comp1);
+    sad += SAD_UB2_UH(src4, src5, comp0, comp1);
+    AVER_UB2_UB(pred6, ref6, pred7, ref7, comp0, comp1);
+    sad += SAD_UB2_UH(src6, src7, comp0, comp1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_64width_msa(const uint8_t *src, int32_t src_stride,
+                                   const uint8_t *ref, int32_t ref_stride,
+                                   int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 comp0, comp1, comp2, comp3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v4u32 sad;
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3,
+                comp0, comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3,
+                comp0, comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3,
+                comp0, comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3,
+                comp0, comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+  }
+
+  sad = __msa_hadd_u_w(sad0, sad0);
+  sad += __msa_hadd_u_w(sad1, sad1);
+
+  return HADD_SW_S32(sad);
+}
+
+#define VPX_SAD_4xHEIGHT_MSA(height)                                        \
+uint32_t vpx_sad4x##height##_msa(const uint8_t *src, int32_t src_stride,    \
+                                 const uint8_t *ref, int32_t ref_stride) {  \
+  return sad_4width_msa(src, src_stride,  ref, ref_stride, height);         \
+}
+
+#define VPX_SAD_8xHEIGHT_MSA(height)                                        \
+uint32_t vpx_sad8x##height##_msa(const uint8_t *src, int32_t src_stride,    \
+                                 const uint8_t *ref, int32_t ref_stride) {  \
+  return sad_8width_msa(src, src_stride, ref, ref_stride, height);          \
+}
+
+#define VPX_SAD_16xHEIGHT_MSA(height)                                        \
+uint32_t vpx_sad16x##height##_msa(const uint8_t *src, int32_t src_stride,    \
+                                  const uint8_t *ref, int32_t ref_stride) {  \
+  return sad_16width_msa(src, src_stride, ref, ref_stride, height);          \
+}
+
+#define VPX_SAD_32xHEIGHT_MSA(height)                                        \
+uint32_t vpx_sad32x##height##_msa(const uint8_t *src, int32_t src_stride,    \
+                                  const uint8_t *ref, int32_t ref_stride) {  \
+  return sad_32width_msa(src, src_stride, ref, ref_stride, height);          \
+}
+
+#define VPX_SAD_64xHEIGHT_MSA(height)                                        \
+uint32_t vpx_sad64x##height##_msa(const uint8_t *src, int32_t src_stride,    \
+                                  const uint8_t *ref, int32_t ref_stride) {  \
+  return sad_64width_msa(src, src_stride, ref, ref_stride, height);          \
+}
+
+#define VPX_SAD_4xHEIGHTx3_MSA(height)                                  \
+void vpx_sad4x##height##x3_msa(const uint8_t *src, int32_t src_stride,  \
+                               const uint8_t *ref, int32_t ref_stride,  \
+                               uint32_t *sads) {                        \
+  sad_4width_x3_msa(src, src_stride, ref, ref_stride, height, sads);    \
+}
+
+#define VPX_SAD_8xHEIGHTx3_MSA(height)                                  \
+void vpx_sad8x##height##x3_msa(const uint8_t *src, int32_t src_stride,  \
+                               const uint8_t *ref, int32_t ref_stride,  \
+                               uint32_t *sads) {                        \
+  sad_8width_x3_msa(src, src_stride, ref, ref_stride, height, sads);    \
+}
+
+#define VPX_SAD_16xHEIGHTx3_MSA(height)                                  \
+void vpx_sad16x##height##x3_msa(const uint8_t *src, int32_t src_stride,  \
+                                const uint8_t *ref, int32_t ref_stride,  \
+                                uint32_t *sads) {                        \
+  sad_16width_x3_msa(src, src_stride, ref, ref_stride, height, sads);    \
+}
+
+#define VPX_SAD_32xHEIGHTx3_MSA(height)                                  \
+void vpx_sad32x##height##x3_msa(const uint8_t *src, int32_t src_stride,  \
+                                const uint8_t *ref, int32_t ref_stride,  \
+                                uint32_t *sads) {                        \
+  sad_32width_x3_msa(src, src_stride, ref, ref_stride, height, sads);    \
+}
+
+#define VPX_SAD_64xHEIGHTx3_MSA(height)                                  \
+void vpx_sad64x##height##x3_msa(const uint8_t *src, int32_t src_stride,  \
+                                const uint8_t *ref, int32_t ref_stride,  \
+                                uint32_t *sads) {                        \
+  sad_64width_x3_msa(src, src_stride, ref, ref_stride, height, sads);    \
+}
+
+#define VPX_SAD_4xHEIGHTx8_MSA(height)                                  \
+void vpx_sad4x##height##x8_msa(const uint8_t *src, int32_t src_stride,  \
+                               const uint8_t *ref, int32_t ref_stride,  \
+                               uint32_t *sads) {                        \
+  sad_4width_x8_msa(src, src_stride, ref, ref_stride, height, sads);    \
+}
+
+#define VPX_SAD_8xHEIGHTx8_MSA(height)                                  \
+void vpx_sad8x##height##x8_msa(const uint8_t *src, int32_t src_stride,  \
+                               const uint8_t *ref, int32_t ref_stride,  \
+                               uint32_t *sads) {                        \
+  sad_8width_x8_msa(src, src_stride, ref, ref_stride, height, sads);    \
+}
+
+#define VPX_SAD_16xHEIGHTx8_MSA(height)                                  \
+void vpx_sad16x##height##x8_msa(const uint8_t *src, int32_t src_stride,  \
+                                const uint8_t *ref, int32_t ref_stride,  \
+                                uint32_t *sads) {                        \
+  sad_16width_x8_msa(src, src_stride, ref, ref_stride, height, sads);    \
+}
+
+#define VPX_SAD_32xHEIGHTx8_MSA(height)                                  \
+void vpx_sad32x##height##x8_msa(const uint8_t *src, int32_t src_stride,  \
+                                const uint8_t *ref, int32_t ref_stride,  \
+                                uint32_t *sads) {                        \
+  sad_32width_x8_msa(src, src_stride, ref, ref_stride, height, sads);    \
+}
+
+#define VPX_SAD_64xHEIGHTx8_MSA(height)                                  \
+void vpx_sad64x##height##x8_msa(const uint8_t *src, int32_t src_stride,  \
+                                const uint8_t *ref, int32_t ref_stride,  \
+                                uint32_t *sads) {                        \
+  sad_64width_x8_msa(src, src_stride, ref, ref_stride, height, sads);    \
+}
+
+#define VPX_SAD_4xHEIGHTx4D_MSA(height)                                  \
+void vpx_sad4x##height##x4d_msa(const uint8_t *src, int32_t src_stride,  \
+                                const uint8_t *const refs[],             \
+                                int32_t ref_stride, uint32_t *sads) {    \
+  sad_4width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);   \
+}
+
+#define VPX_SAD_8xHEIGHTx4D_MSA(height)                                  \
+void vpx_sad8x##height##x4d_msa(const uint8_t *src, int32_t src_stride,  \
+                                const uint8_t *const refs[],             \
+                                int32_t ref_stride, uint32_t *sads) {    \
+  sad_8width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);   \
+}
+
+#define VPX_SAD_16xHEIGHTx4D_MSA(height)                                  \
+void vpx_sad16x##height##x4d_msa(const uint8_t *src, int32_t src_stride,  \
+                                 const uint8_t *const refs[],             \
+                                 int32_t ref_stride, uint32_t *sads) {    \
+  sad_16width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);   \
+}
+
+#define VPX_SAD_32xHEIGHTx4D_MSA(height)                                  \
+void vpx_sad32x##height##x4d_msa(const uint8_t *src, int32_t src_stride,  \
+                                 const uint8_t *const refs[],             \
+                                 int32_t ref_stride, uint32_t *sads) {    \
+  sad_32width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);   \
+}
+
+#define VPX_SAD_64xHEIGHTx4D_MSA(height)                                  \
+void vpx_sad64x##height##x4d_msa(const uint8_t *src, int32_t src_stride,  \
+                                 const uint8_t *const refs[],             \
+                                 int32_t ref_stride, uint32_t *sads) {    \
+  sad_64width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);   \
+}
+
+#define VPX_AVGSAD_4xHEIGHT_MSA(height)                                       \
+uint32_t vpx_sad4x##height##_avg_msa(const uint8_t *src, int32_t src_stride,  \
+                                     const uint8_t *ref, int32_t ref_stride,  \
+                                     const uint8_t *second_pred) {            \
+  return avgsad_4width_msa(src, src_stride, ref, ref_stride,                  \
+                           height, second_pred);                              \
+}
+
+#define VPX_AVGSAD_8xHEIGHT_MSA(height)                                       \
+uint32_t vpx_sad8x##height##_avg_msa(const uint8_t *src, int32_t src_stride,  \
+                                     const uint8_t *ref, int32_t ref_stride,  \
+                                     const uint8_t *second_pred) {            \
+  return avgsad_8width_msa(src, src_stride, ref, ref_stride,                  \
+                           height, second_pred);                              \
+}
+
+#define VPX_AVGSAD_16xHEIGHT_MSA(height)                                       \
+uint32_t vpx_sad16x##height##_avg_msa(const uint8_t *src, int32_t src_stride,  \
+                                      const uint8_t *ref, int32_t ref_stride,  \
+                                      const uint8_t *second_pred) {            \
+  return avgsad_16width_msa(src, src_stride, ref, ref_stride,                  \
+                            height, second_pred);                              \
+}
+
+#define VPX_AVGSAD_32xHEIGHT_MSA(height)                                       \
+uint32_t vpx_sad32x##height##_avg_msa(const uint8_t *src, int32_t src_stride,  \
+                                      const uint8_t *ref, int32_t ref_stride,  \
+                                      const uint8_t *second_pred) {            \
+  return avgsad_32width_msa(src, src_stride, ref, ref_stride,                  \
+                            height, second_pred);                              \
+}
+
+#define VPX_AVGSAD_64xHEIGHT_MSA(height)                                       \
+uint32_t vpx_sad64x##height##_avg_msa(const uint8_t *src, int32_t src_stride,  \
+                                      const uint8_t *ref, int32_t ref_stride,  \
+                                      const uint8_t *second_pred) {            \
+  return avgsad_64width_msa(src, src_stride, ref, ref_stride,                  \
+                            height, second_pred);                              \
+}
+
+// 64x64
+VPX_SAD_64xHEIGHT_MSA(64);
+VPX_SAD_64xHEIGHTx3_MSA(64);
+VPX_SAD_64xHEIGHTx8_MSA(64);
+VPX_SAD_64xHEIGHTx4D_MSA(64);
+VPX_AVGSAD_64xHEIGHT_MSA(64);
+
+// 64x32
+VPX_SAD_64xHEIGHT_MSA(32);
+VPX_SAD_64xHEIGHTx3_MSA(32);
+VPX_SAD_64xHEIGHTx8_MSA(32);
+VPX_SAD_64xHEIGHTx4D_MSA(32);
+VPX_AVGSAD_64xHEIGHT_MSA(32);
+
+// 32x64
+VPX_SAD_32xHEIGHT_MSA(64);
+VPX_SAD_32xHEIGHTx3_MSA(64);
+VPX_SAD_32xHEIGHTx8_MSA(64);
+VPX_SAD_32xHEIGHTx4D_MSA(64);
+VPX_AVGSAD_32xHEIGHT_MSA(64);
+
+// 32x32
+VPX_SAD_32xHEIGHT_MSA(32);
+VPX_SAD_32xHEIGHTx3_MSA(32);
+VPX_SAD_32xHEIGHTx8_MSA(32);
+VPX_SAD_32xHEIGHTx4D_MSA(32);
+VPX_AVGSAD_32xHEIGHT_MSA(32);
+
+// 32x16
+VPX_SAD_32xHEIGHT_MSA(16);
+VPX_SAD_32xHEIGHTx3_MSA(16);
+VPX_SAD_32xHEIGHTx8_MSA(16);
+VPX_SAD_32xHEIGHTx4D_MSA(16);
+VPX_AVGSAD_32xHEIGHT_MSA(16);
+
+// 16x32
+VPX_SAD_16xHEIGHT_MSA(32);
+VPX_SAD_16xHEIGHTx3_MSA(32);
+VPX_SAD_16xHEIGHTx8_MSA(32);
+VPX_SAD_16xHEIGHTx4D_MSA(32);
+VPX_AVGSAD_16xHEIGHT_MSA(32);
+
+// 16x16
+VPX_SAD_16xHEIGHT_MSA(16);
+VPX_SAD_16xHEIGHTx3_MSA(16);
+VPX_SAD_16xHEIGHTx8_MSA(16);
+VPX_SAD_16xHEIGHTx4D_MSA(16);
+VPX_AVGSAD_16xHEIGHT_MSA(16);
+
+// 16x8
+VPX_SAD_16xHEIGHT_MSA(8);
+VPX_SAD_16xHEIGHTx3_MSA(8);
+VPX_SAD_16xHEIGHTx8_MSA(8);
+VPX_SAD_16xHEIGHTx4D_MSA(8);
+VPX_AVGSAD_16xHEIGHT_MSA(8);
+
+// 8x16
+VPX_SAD_8xHEIGHT_MSA(16);
+VPX_SAD_8xHEIGHTx3_MSA(16);
+VPX_SAD_8xHEIGHTx8_MSA(16);
+VPX_SAD_8xHEIGHTx4D_MSA(16);
+VPX_AVGSAD_8xHEIGHT_MSA(16);
+
+// 8x8
+VPX_SAD_8xHEIGHT_MSA(8);
+VPX_SAD_8xHEIGHTx3_MSA(8);
+VPX_SAD_8xHEIGHTx8_MSA(8);
+VPX_SAD_8xHEIGHTx4D_MSA(8);
+VPX_AVGSAD_8xHEIGHT_MSA(8);
+
+// 8x4
+VPX_SAD_8xHEIGHT_MSA(4);
+VPX_SAD_8xHEIGHTx3_MSA(4);
+VPX_SAD_8xHEIGHTx8_MSA(4);
+VPX_SAD_8xHEIGHTx4D_MSA(4);
+VPX_AVGSAD_8xHEIGHT_MSA(4);
+
+// 4x8
+VPX_SAD_4xHEIGHT_MSA(8);
+VPX_SAD_4xHEIGHTx3_MSA(8);
+VPX_SAD_4xHEIGHTx8_MSA(8);
+VPX_SAD_4xHEIGHTx4D_MSA(8);
+VPX_AVGSAD_4xHEIGHT_MSA(8);
+
+// 4x4
+VPX_SAD_4xHEIGHT_MSA(4);
+VPX_SAD_4xHEIGHTx3_MSA(4);
+VPX_SAD_4xHEIGHTx8_MSA(4);
+VPX_SAD_4xHEIGHTx4D_MSA(4);
+VPX_AVGSAD_4xHEIGHT_MSA(4);
diff --git a/libvpx/libvpx/vpx_dsp/mips/sub_pixel_variance_msa.c b/libvpx/libvpx/vpx_dsp/mips/sub_pixel_variance_msa.c
new file mode 100644
index 0000000..a592a2d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/sub_pixel_variance_msa.c
@@ -0,0 +1,1952 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vpx_dsp/mips/macros_msa.h"
+#include "vpx_dsp/variance.h"
+
+static const uint8_t bilinear_filters_msa[8][2] = {
+  { 128,   0, },
+  { 112,  16, },
+  {  96,  32, },
+  {  80,  48, },
+  {  64,  64, },
+  {  48,  80, },
+  {  32,  96, },
+  {  16, 112, },
+};
+
+#define CALC_MSE_AVG_B(src, ref, var, sub) {                       \
+  v16u8 src_l0_m, src_l1_m;                                        \
+  v8i16 res_l0_m, res_l1_m;                                        \
+                                                                   \
+  ILVRL_B2_UB(src, ref, src_l0_m, src_l1_m);                       \
+  HSUB_UB2_SH(src_l0_m, src_l1_m, res_l0_m, res_l1_m);             \
+  DPADD_SH2_SW(res_l0_m, res_l1_m, res_l0_m, res_l1_m, var, var);  \
+                                                                   \
+  sub += res_l0_m + res_l1_m;                                      \
+}
+
+#define VARIANCE_WxH(sse, diff, shift) \
+  sse - (((uint32_t)diff * diff) >> shift)
+
+#define VARIANCE_LARGE_WxH(sse, diff, shift) \
+  sse - (((int64_t)diff * diff) >> shift)
+
+static uint32_t avg_sse_diff_4width_msa(const uint8_t *src_ptr,
+                                        int32_t src_stride,
+                                        const uint8_t *ref_ptr,
+                                        int32_t ref_stride,
+                                        const uint8_t *sec_pred,
+                                        int32_t height,
+                                        int32_t *diff) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 pred, src = { 0 };
+  v16u8 ref = { 0 };
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+
+    src = __msa_aver_u_b(src, pred);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_8width_msa(const uint8_t *src_ptr,
+                                        int32_t src_stride,
+                                        const uint8_t *ref_ptr,
+                                        int32_t ref_stride,
+                                        const uint8_t *sec_pred,
+                                        int32_t height,
+                                        int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2,
+                src0, src1, ref0, ref1);
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_16width_msa(const uint8_t *src_ptr,
+                                         int32_t src_stride,
+                                         const uint8_t *ref_ptr,
+                                         int32_t ref_stride,
+                                         const uint8_t *sec_pred,
+                                         int32_t height,
+                                         int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src, ref, pred;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    src = __msa_aver_u_b(src, pred);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    src = __msa_aver_u_b(src, pred);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    src = __msa_aver_u_b(src, pred);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    src = __msa_aver_u_b(src, pred);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_32width_msa(const uint8_t *src_ptr,
+                                         int32_t src_stride,
+                                         const uint8_t *ref_ptr,
+                                         int32_t ref_stride,
+                                         const uint8_t *sec_pred,
+                                         int32_t height,
+                                         int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1, pred0, pred1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_32x64_msa(const uint8_t *src_ptr,
+                                       int32_t src_stride,
+                                       const uint8_t *ref_ptr,
+                                       int32_t ref_stride,
+                                       const uint8_t *sec_pred,
+                                       int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1, pred0, pred1;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 16; ht_cnt--;) {
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_64x32_msa(const uint8_t *src_ptr,
+                                       int32_t src_stride,
+                                       const uint8_t *ref_ptr,
+                                       int32_t ref_stride,
+                                       const uint8_t *sec_pred,
+                                       int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 16; ht_cnt--;) {
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3,
+                src0, src1, src2, src3);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src2, ref2, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src3, ref3, var, avg1);
+
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3,
+                src0, src1, src2, src3);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src2, ref2, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src3, ref3, var, avg1);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_64x64_msa(const uint8_t *src_ptr,
+                                       int32_t src_stride,
+                                       const uint8_t *ref_ptr,
+                                       int32_t ref_stride,
+                                       const uint8_t *sec_pred,
+                                       int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v8i16 avg2 = { 0 };
+  v8i16 avg3 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 32; ht_cnt--;) {
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3,
+                src0, src1, src2, src3);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src2, ref2, var, avg2);
+    CALC_MSE_AVG_B(src3, ref3, var, avg3);
+
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3,
+                src0, src1, src2, src3);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src2, ref2, var, avg2);
+    CALC_MSE_AVG_B(src3, ref3, var, avg3);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+  vec += __msa_hadd_s_w(avg2, avg2);
+  vec += __msa_hadd_s_w(avg3, avg3);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_4width_h_msa(const uint8_t *src,
+                                                int32_t src_stride,
+                                                const uint8_t *dst,
+                                                int32_t dst_stride,
+                                                const uint8_t *filter,
+                                                int32_t height,
+                                                int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 filt0, ref = { 0 };
+  v16i8 src0, src1, src2, src3;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 vec0, vec1, vec2, vec3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                vec0, vec1, vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3,
+                src0, src1, src2, src3);
+    ILVEV_W2_SB(src0, src1, src2, src3, src0, src2);
+    src0 = (v16i8)__msa_ilvev_d((v2i64)src2, (v2i64)src0);
+    CALC_MSE_AVG_B(src0, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_8width_h_msa(const uint8_t *src,
+                                                int32_t src_stride,
+                                                const uint8_t *dst,
+                                                int32_t dst_stride,
+                                                const uint8_t *filter,
+                                                int32_t height,
+                                                int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 filt0, out, ref0, ref1, ref2, ref3;
+  v16i8 src0, src1, src2, src3;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 vec0, vec1, vec2, vec3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                vec0, vec1, vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3,
+                src0, src1, src2, src3);
+    out = (v16u8)__msa_ilvev_d((v2i64)src1, (v2i64)src0);
+    CALC_MSE_AVG_B(out, ref0, var, avg);
+    out = (v16u8)__msa_ilvev_d((v2i64)src3, (v2i64)src2);
+    CALC_MSE_AVG_B(out, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_16width_h_msa(const uint8_t *src,
+                                                 int32_t src_stride,
+                                                 const uint8_t *dst,
+                                                 int32_t dst_stride,
+                                                 const uint8_t *filter,
+                                                 int32_t height,
+                                                 int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v16u8 dst0, dst1, dst2, dst3, filt0;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    dst += (4 * dst_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UH(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UH(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                out0, out1, out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0,
+                out4, out5, out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    PCKEV_B4_SB(out1, out0, out3, out2, out5, out4, out7, out6,
+                src0, src1, src2, src3);
+    CALC_MSE_AVG_B(src0, dst0, var, avg);
+    CALC_MSE_AVG_B(src1, dst1, var, avg);
+    CALC_MSE_AVG_B(src2, dst2, var, avg);
+    CALC_MSE_AVG_B(src3, dst3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_32width_h_msa(const uint8_t *src,
+                                                 int32_t src_stride,
+                                                 const uint8_t *dst,
+                                                 int32_t dst_stride,
+                                                 const uint8_t *filter,
+                                                 int32_t height,
+                                                 int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_h_msa(src, src_stride, dst, dst_stride,
+                                            filter, height, &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_sse_diff_64width_h_msa(const uint8_t *src,
+                                                 int32_t src_stride,
+                                                 const uint8_t *dst,
+                                                 int32_t dst_stride,
+                                                 const uint8_t *filter,
+                                                 int32_t height,
+                                                 int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_h_msa(src, src_stride, dst, dst_stride,
+                                            filter, height, &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_sse_diff_4width_v_msa(const uint8_t *src,
+                                                int32_t src_stride,
+                                                const uint8_t *dst,
+                                                int32_t dst_stride,
+                                                const uint8_t *filter,
+                                                int32_t height,
+                                                int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4, out;
+  v16u8 src10_r, src32_r, src21_r, src43_r;
+  v16u8 ref = { 0 };
+  v16u8 src2110, src4332;
+  v16u8 filt0;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+  v8u16 tmp0, tmp1;
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3,
+               src10_r, src21_r, src32_r, src43_r);
+    ILVR_D2_UB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+    DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    CALC_MSE_AVG_B(out, ref, var, avg);
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_8width_v_msa(const uint8_t *src,
+                                                int32_t src_stride,
+                                                const uint8_t *dst,
+                                                int32_t dst_stride,
+                                                const uint8_t *filter,
+                                                int32_t height,
+                                                int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 vec0, vec1, vec2, vec3;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 filt0;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    ILVR_B4_UH(src1, src0, src2, src1, src3, src2, src4, src3,
+               vec0, vec1, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                tmp0, tmp1, tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_16width_v_msa(const uint8_t *src,
+                                                 int32_t src_stride,
+                                                 const uint8_t *dst,
+                                                 int32_t dst_stride,
+                                                 const uint8_t *filter,
+                                                 int32_t height,
+                                                 int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 out0, out1, out2, out3;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 filt0;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    out1 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    out3 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+
+    src0 = src4;
+
+    CALC_MSE_AVG_B(out0, ref0, var, avg);
+    CALC_MSE_AVG_B(out1, ref1, var, avg);
+    CALC_MSE_AVG_B(out2, ref2, var, avg);
+    CALC_MSE_AVG_B(out3, ref3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_32width_v_msa(const uint8_t *src,
+                                                 int32_t src_stride,
+                                                 const uint8_t *dst,
+                                                 int32_t dst_stride,
+                                                 const uint8_t *filter,
+                                                 int32_t height,
+                                                 int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_v_msa(src, src_stride, dst, dst_stride,
+                                            filter, height, &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_sse_diff_64width_v_msa(const uint8_t *src,
+                                                 int32_t src_stride,
+                                                 const uint8_t *dst,
+                                                 int32_t dst_stride,
+                                                 const uint8_t *filter,
+                                                 int32_t height,
+                                                 int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_v_msa(src, src_stride, dst, dst_stride,
+                                            filter, height, &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_sse_diff_4width_hv_msa(const uint8_t *src,
+                                                 int32_t src_stride,
+                                                 const uint8_t *dst,
+                                                 int32_t dst_stride,
+                                                 const uint8_t *filter_horiz,
+                                                 const uint8_t *filter_vert,
+                                                 int32_t height,
+                                                 int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 out, ref = { 0 };
+  v16u8 filt_vt, filt_hz, vec0, vec1;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20 };
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4;
+  v8u16 tmp0, tmp1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+    hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+    hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    CALC_MSE_AVG_B(out, ref, var, avg);
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_8width_hv_msa(const uint8_t *src,
+                                                 int32_t src_stride,
+                                                 const uint8_t *dst,
+                                                 int32_t dst_stride,
+                                                 const uint8_t *filter_horiz,
+                                                 const uint8_t *filter_vert,
+                                                 int32_t height,
+                                                 int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 out0, out1;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 hz_out0, hz_out1;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 filt_vt, filt_hz, vec0;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp1 = __msa_dotp_u_h(vec0, filt_vt);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp2 = __msa_dotp_u_h(vec0, filt_vt);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp3 = __msa_dotp_u_h(vec0, filt_vt);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+    CALC_MSE_AVG_B(out0, ref0, var, avg);
+    CALC_MSE_AVG_B(out1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_16width_hv_msa(const uint8_t *src,
+                                                  int32_t src_stride,
+                                                  const uint8_t *dst,
+                                                  int32_t dst_stride,
+                                                  const uint8_t *filter_horiz,
+                                                  const uint8_t *filter_vert,
+                                                  int32_t height,
+                                                  int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 filt_hz, filt_vt, vec0, vec1;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3;
+  v8u16 tmp0, tmp1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  LD_UB2(src, 8, src0, src1);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src0, src2, src4, src6);
+    LD_UB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    src0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    src1 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    src2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    src3 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+    CALC_MSE_AVG_B(src2, ref2, var, avg);
+    CALC_MSE_AVG_B(src3, ref3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_32width_hv_msa(const uint8_t *src,
+                                                  int32_t src_stride,
+                                                  const uint8_t *dst,
+                                                  int32_t dst_stride,
+                                                  const uint8_t *filter_horiz,
+                                                  const uint8_t *filter_vert,
+                                                  int32_t height,
+                                                  int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_hv_msa(src, src_stride, dst, dst_stride,
+                                             filter_horiz, filter_vert, height,
+                                             &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_sse_diff_64width_hv_msa(const uint8_t *src,
+                                                  int32_t src_stride,
+                                                  const uint8_t *dst,
+                                                  int32_t dst_stride,
+                                                  const uint8_t *filter_horiz,
+                                                  const uint8_t *filter_vert,
+                                                  int32_t height,
+                                                  int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_hv_msa(src, src_stride, dst, dst_stride,
+                                             filter_horiz, filter_vert, height,
+                                             &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_4width_h_msa(const uint8_t *src,
+                                                    int32_t src_stride,
+                                                    const uint8_t *dst,
+                                                    int32_t dst_stride,
+                                                    const uint8_t *sec_pred,
+                                                    const uint8_t *filter,
+                                                    int32_t height,
+                                                    int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 out, pred, filt0, ref = { 0 };
+  v16i8 src0, src1, src2, src3;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 vec0, vec1, vec2, vec3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                vec0, vec1, vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3,
+                src0, src1, src2, src3);
+    ILVEV_W2_SB(src0, src1, src2, src3, src0, src2);
+    out = (v16u8)__msa_ilvev_d((v2i64)src2, (v2i64)src0);
+    out = __msa_aver_u_b(out, pred);
+    CALC_MSE_AVG_B(out, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_8width_h_msa(const uint8_t *src,
+                                                    int32_t src_stride,
+                                                    const uint8_t *dst,
+                                                    int32_t dst_stride,
+                                                    const uint8_t *sec_pred,
+                                                    const uint8_t *filter,
+                                                    int32_t height,
+                                                    int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 out, pred, filt0;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16i8 src0, src1, src2, src3;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 vec0, vec1, vec2, vec3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                vec0, vec1, vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3,
+                src0, src1, src2, src3);
+    out = (v16u8)__msa_ilvev_d((v2i64)src1, (v2i64)src0);
+
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    out = __msa_aver_u_b(out, pred);
+    CALC_MSE_AVG_B(out, ref0, var, avg);
+    out = (v16u8)__msa_ilvev_d((v2i64)src3, (v2i64)src2);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    out = __msa_aver_u_b(out, pred);
+    CALC_MSE_AVG_B(out, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t subpel_avg_ssediff_16w_h_msa(const uint8_t *src,
+                                             int32_t src_stride,
+                                             const uint8_t *dst,
+                                             int32_t dst_stride,
+                                             const uint8_t *sec_pred,
+                                             const uint8_t *filter,
+                                             int32_t height,
+                                             int32_t *diff,
+                                             int32_t width) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v16u8 dst0, dst1, dst2, dst3;
+  v16u8 tmp0, tmp1, tmp2, tmp3;
+  v16u8 pred0, pred1, pred2, pred3, filt0;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    dst += (4 * dst_stride);
+    LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * width);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UH(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UH(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                out0, out1, out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0,
+                out4, out5, out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    PCKEV_B4_UB(out1, out0, out3, out2, out5, out4, out7, out6,
+                tmp0, tmp1, tmp2, tmp3);
+    AVER_UB4_UB(tmp0, pred0, tmp1, pred1, tmp2, pred2, tmp3, pred3,
+                tmp0, tmp1, tmp2, tmp3);
+
+    CALC_MSE_AVG_B(tmp0, dst0, var, avg);
+    CALC_MSE_AVG_B(tmp1, dst1, var, avg);
+    CALC_MSE_AVG_B(tmp2, dst2, var, avg);
+    CALC_MSE_AVG_B(tmp3, dst3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_16width_h_msa(const uint8_t *src,
+                                                     int32_t src_stride,
+                                                     const uint8_t *dst,
+                                                     int32_t dst_stride,
+                                                     const uint8_t *sec_pred,
+                                                     const uint8_t *filter,
+                                                     int32_t height,
+                                                     int32_t *diff) {
+  return subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride,
+                                      sec_pred, filter, height, diff, 16);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_32width_h_msa(const uint8_t *src,
+                                                     int32_t src_stride,
+                                                     const uint8_t *dst,
+                                                     int32_t dst_stride,
+                                                     const uint8_t *sec_pred,
+                                                     const uint8_t *filter,
+                                                     int32_t height,
+                                                     int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse += subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride,
+                                        sec_pred, filter, height,
+                                        &diff0[loop_cnt], 32);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_64width_h_msa(const uint8_t *src,
+                                                     int32_t src_stride,
+                                                     const uint8_t *dst,
+                                                     int32_t dst_stride,
+                                                     const uint8_t *sec_pred,
+                                                     const uint8_t *filter,
+                                                     int32_t height,
+                                                     int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse += subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride,
+                                        sec_pred, filter, height,
+                                        &diff0[loop_cnt], 64);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_4width_v_msa(const uint8_t *src,
+                                                    int32_t src_stride,
+                                                    const uint8_t *dst,
+                                                    int32_t dst_stride,
+                                                    const uint8_t *sec_pred,
+                                                    const uint8_t *filter,
+                                                    int32_t height,
+                                                    int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 src10_r, src32_r, src21_r, src43_r;
+  v16u8 out, pred, ref = { 0 };
+  v16u8 src2110, src4332, filt0;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+  v8u16 tmp0, tmp1;
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3,
+               src10_r, src21_r, src32_r, src43_r);
+    ILVR_D2_UB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+    DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+
+    out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    out = __msa_aver_u_b(out, pred);
+    CALC_MSE_AVG_B(out, ref, var, avg);
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_8width_v_msa(const uint8_t *src,
+                                                    int32_t src_stride,
+                                                    const uint8_t *dst,
+                                                    int32_t dst_stride,
+                                                    const uint8_t *sec_pred,
+                                                    const uint8_t *filter,
+                                                    int32_t height,
+                                                    int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, filt0;
+  v8u16 vec0, vec1, vec2, vec3;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    ILVR_B4_UH(src1, src0, src2, src1, src3, src2, src4, src3,
+               vec0, vec1, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0,
+                tmp0, tmp1, tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, src0, src1);
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t subpel_avg_ssediff_16w_v_msa(const uint8_t *src,
+                                             int32_t src_stride,
+                                             const uint8_t *dst,
+                                             int32_t dst_stride,
+                                             const uint8_t *sec_pred,
+                                             const uint8_t *filter,
+                                             int32_t height,
+                                             int32_t *diff,
+                                             int32_t width) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 out0, out1, out2, out3, filt0;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * width);
+
+    ILVR_B2_UH(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UH(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    ILVR_B2_UH(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UH(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    out1 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    out3 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+
+    src0 = src4;
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    AVER_UB4_UB(out0, pred0, out1, pred1, out2, pred2, out3, pred3,
+                out0, out1, out2, out3);
+
+    CALC_MSE_AVG_B(out0, ref0, var, avg);
+    CALC_MSE_AVG_B(out1, ref1, var, avg);
+    CALC_MSE_AVG_B(out2, ref2, var, avg);
+    CALC_MSE_AVG_B(out3, ref3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_16width_v_msa(const uint8_t *src,
+                                                     int32_t src_stride,
+                                                     const uint8_t *dst,
+                                                     int32_t dst_stride,
+                                                     const uint8_t *sec_pred,
+                                                     const uint8_t *filter,
+                                                     int32_t height,
+                                                     int32_t *diff) {
+  return subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride,
+                                      sec_pred, filter, height, diff, 16);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_32width_v_msa(const uint8_t *src,
+                                                     int32_t src_stride,
+                                                     const uint8_t *dst,
+                                                     int32_t dst_stride,
+                                                     const uint8_t *sec_pred,
+                                                     const uint8_t *filter,
+                                                     int32_t height,
+                                                     int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse += subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride,
+                                        sec_pred, filter, height,
+                                        &diff0[loop_cnt], 32);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_64width_v_msa(const uint8_t *src,
+                                                     int32_t src_stride,
+                                                     const uint8_t *dst,
+                                                     int32_t dst_stride,
+                                                     const uint8_t *sec_pred,
+                                                     const uint8_t *filter,
+                                                     int32_t height,
+                                                     int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse += subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride,
+                                        sec_pred, filter, height,
+                                        &diff0[loop_cnt], 64);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_4width_hv_msa(
+  const uint8_t *src, int32_t src_stride,
+  const uint8_t *dst, int32_t dst_stride,
+  const uint8_t *sec_pred,
+  const uint8_t *filter_horiz, const uint8_t *filter_vert,
+  int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20 };
+  v16u8 filt_hz, filt_vt, vec0, vec1;
+  v16u8 out, pred, ref = { 0 };
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, tmp0, tmp1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+    hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+    hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    out = __msa_aver_u_b(out, pred);
+    CALC_MSE_AVG_B(out, ref, var, avg);
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_8width_hv_msa(
+  const uint8_t *src, int32_t src_stride,
+  const uint8_t *dst, int32_t dst_stride,
+  const uint8_t *sec_pred,
+  const uint8_t *filter_horiz, const uint8_t *filter_vert,
+  int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 pred0, pred1, out0, out1;
+  v16u8 filt_hz, filt_vt, vec0;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp1 = __msa_dotp_u_h(vec0, filt_vt);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp2 = __msa_dotp_u_h(vec0, filt_vt);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp3 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+    AVER_UB2_UB(out0, pred0, out1, pred1, out0, out1);
+
+    CALC_MSE_AVG_B(out0, ref0, var, avg);
+    CALC_MSE_AVG_B(out1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t subpel_avg_ssediff_16w_hv_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              const uint8_t *dst,
+                                              int32_t dst_stride,
+                                              const uint8_t *sec_pred,
+                                              const uint8_t *filter_horiz,
+                                              const uint8_t *filter_vert,
+                                              int32_t height,
+                                              int32_t *diff,
+                                              int32_t width) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v16u8 out0, out1, out2, out3;
+  v16u8 filt_hz, filt_vt, vec0, vec1;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, tmp0, tmp1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  LD_UB2(src, 8, src0, src1);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src0, src2, src4, src6);
+    LD_UB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * width);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out1 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out3 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    AVER_UB4_UB(out0, pred0, out1, pred1, out2, pred2, out3, pred3,
+                out0, out1, out2, out3);
+
+    CALC_MSE_AVG_B(out0, ref0, var, avg);
+    CALC_MSE_AVG_B(out1, ref1, var, avg);
+    CALC_MSE_AVG_B(out2, ref2, var, avg);
+    CALC_MSE_AVG_B(out3, ref3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_16width_hv_msa(
+  const uint8_t *src, int32_t src_stride,
+  const uint8_t *dst, int32_t dst_stride,
+  const uint8_t *sec_pred,
+  const uint8_t *filter_horiz, const uint8_t *filter_vert,
+  int32_t height, int32_t *diff) {
+  return subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride,
+                                       sec_pred, filter_horiz, filter_vert,
+                                       height, diff, 16);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_32width_hv_msa(
+  const uint8_t *src, int32_t src_stride,
+  const uint8_t *dst, int32_t dst_stride,
+  const uint8_t *sec_pred,
+  const uint8_t *filter_horiz, const uint8_t *filter_vert,
+  int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse += subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride,
+                                         sec_pred, filter_horiz, filter_vert,
+                                         height, &diff0[loop_cnt], 32);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_64width_hv_msa(
+  const uint8_t *src, int32_t src_stride,
+  const uint8_t *dst, int32_t dst_stride,
+  const uint8_t *sec_pred,
+  const uint8_t *filter_horiz, const uint8_t *filter_vert,
+  int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse += subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride,
+                                         sec_pred, filter_horiz, filter_vert,
+                                         height, &diff0[loop_cnt], 64);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+#define VARIANCE_4Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 4);
+#define VARIANCE_4Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 5);
+#define VARIANCE_8Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 5);
+#define VARIANCE_8Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 6);
+#define VARIANCE_8Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 7);
+#define VARIANCE_16Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 7);
+#define VARIANCE_16Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 8);
+
+#define VARIANCE_16Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
+#define VARIANCE_32Wx16H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
+#define VARIANCE_32Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 10);
+#define VARIANCE_32Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
+#define VARIANCE_64Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
+#define VARIANCE_64Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 12);
+
+#define VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(wd, ht)                         \
+uint32_t vpx_sub_pixel_variance##wd##x##ht##_msa(const uint8_t *src,     \
+                                                 int32_t src_stride,     \
+                                                 int32_t xoffset,        \
+                                                 int32_t yoffset,        \
+                                                 const uint8_t *ref,     \
+                                                 int32_t ref_stride,     \
+                                                 uint32_t *sse) {        \
+  int32_t diff;                                                          \
+  uint32_t var;                                                          \
+  const uint8_t *h_filter = bilinear_filters_msa[xoffset];               \
+  const uint8_t *v_filter = bilinear_filters_msa[yoffset];               \
+                                                                         \
+  if (yoffset) {                                                         \
+    if (xoffset) {                                                       \
+      *sse = sub_pixel_sse_diff_##wd##width_hv_msa(src, src_stride,      \
+                                                   ref, ref_stride,      \
+                                                   h_filter, v_filter,   \
+                                                   ht, &diff);           \
+    } else {                                                             \
+      *sse = sub_pixel_sse_diff_##wd##width_v_msa(src, src_stride,       \
+                                                  ref, ref_stride,       \
+                                                  v_filter, ht, &diff);  \
+    }                                                                    \
+                                                                         \
+    var = VARIANCE_##wd##Wx##ht##H(*sse, diff);                          \
+  } else {                                                               \
+    if (xoffset) {                                                       \
+      *sse = sub_pixel_sse_diff_##wd##width_h_msa(src, src_stride,       \
+                                                  ref, ref_stride,       \
+                                                  h_filter, ht, &diff);  \
+                                                                         \
+      var = VARIANCE_##wd##Wx##ht##H(*sse, diff);                        \
+    } else {                                                             \
+      var = vpx_variance##wd##x##ht##_msa(src, src_stride,               \
+                                          ref, ref_stride, sse);         \
+    }                                                                    \
+  }                                                                      \
+                                                                         \
+  return var;                                                            \
+}
+
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(4, 4);
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(4, 8);
+
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 4);
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 8);
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 16);
+
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 8);
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 16);
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 32);
+
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 16);
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 32);
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 64);
+
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(64, 32);
+VPX_SUB_PIXEL_VARIANCE_WDXHT_MSA(64, 64);
+
+#define VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(wd, ht)                          \
+uint32_t vpx_sub_pixel_avg_variance##wd##x##ht##_msa(                         \
+  const uint8_t *src_ptr, int32_t src_stride,                                 \
+  int32_t xoffset, int32_t yoffset,                                           \
+  const uint8_t *ref_ptr, int32_t ref_stride,                                 \
+  uint32_t *sse, const uint8_t *sec_pred) {                                   \
+  int32_t diff;                                                               \
+  const uint8_t *h_filter = bilinear_filters_msa[xoffset];                    \
+  const uint8_t *v_filter = bilinear_filters_msa[yoffset];                    \
+                                                                              \
+  if (yoffset) {                                                              \
+    if (xoffset) {                                                            \
+      *sse = sub_pixel_avg_sse_diff_##wd##width_hv_msa(src_ptr, src_stride,   \
+                                                       ref_ptr, ref_stride,   \
+                                                       sec_pred, h_filter,    \
+                                                       v_filter, ht, &diff);  \
+    } else {                                                                  \
+      *sse = sub_pixel_avg_sse_diff_##wd##width_v_msa(src_ptr, src_stride,    \
+                                                      ref_ptr, ref_stride,    \
+                                                      sec_pred, v_filter,     \
+                                                      ht, &diff);             \
+    }                                                                         \
+  } else {                                                                    \
+    if (xoffset) {                                                            \
+      *sse = sub_pixel_avg_sse_diff_##wd##width_h_msa(src_ptr, src_stride,    \
+                                                      ref_ptr, ref_stride,    \
+                                                      sec_pred, h_filter,     \
+                                                      ht, &diff);             \
+    } else {                                                                  \
+      *sse = avg_sse_diff_##wd##width_msa(src_ptr, src_stride,                \
+                                          ref_ptr, ref_stride,                \
+                                          sec_pred, ht, &diff);               \
+    }                                                                         \
+  }                                                                           \
+                                                                              \
+  return VARIANCE_##wd##Wx##ht##H(*sse, diff);                                \
+}
+
+VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(4, 4);
+VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(4, 8);
+
+VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 4);
+VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 8);
+VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 16);
+
+VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 8);
+VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 16);
+VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 32);
+
+VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(32, 16);
+VPX_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(32, 32);
+
+uint32_t vpx_sub_pixel_avg_variance32x64_msa(const uint8_t *src_ptr,
+                                             int32_t src_stride,
+                                             int32_t xoffset,
+                                             int32_t yoffset,
+                                             const uint8_t *ref_ptr,
+                                             int32_t ref_stride,
+                                             uint32_t *sse,
+                                             const uint8_t *sec_pred) {
+  int32_t diff;
+  const uint8_t *h_filter = bilinear_filters_msa[xoffset];
+  const uint8_t *v_filter = bilinear_filters_msa[yoffset];
+
+  if (yoffset) {
+    if (xoffset) {
+      *sse = sub_pixel_avg_sse_diff_32width_hv_msa(src_ptr, src_stride,
+                                                   ref_ptr, ref_stride,
+                                                   sec_pred, h_filter,
+                                                   v_filter, 64, &diff);
+    } else {
+      *sse = sub_pixel_avg_sse_diff_32width_v_msa(src_ptr, src_stride,
+                                                  ref_ptr, ref_stride,
+                                                  sec_pred, v_filter,
+                                                  64, &diff);
+    }
+  } else {
+    if (xoffset) {
+      *sse = sub_pixel_avg_sse_diff_32width_h_msa(src_ptr, src_stride,
+                                                  ref_ptr, ref_stride,
+                                                  sec_pred, h_filter,
+                                                  64, &diff);
+    } else {
+      *sse = avg_sse_diff_32x64_msa(src_ptr, src_stride, ref_ptr, ref_stride,
+                                    sec_pred, &diff);
+    }
+  }
+
+  return VARIANCE_32Wx64H(*sse, diff);
+}
+
+#define VPX_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(ht)                          \
+uint32_t vpx_sub_pixel_avg_variance64x##ht##_msa(const uint8_t *src_ptr,     \
+                                                 int32_t src_stride,         \
+                                                 int32_t xoffset,            \
+                                                 int32_t yoffset,            \
+                                                 const uint8_t *ref_ptr,     \
+                                                 int32_t ref_stride,         \
+                                                 uint32_t *sse,              \
+                                                 const uint8_t *sec_pred) {  \
+  int32_t diff;                                                              \
+  const uint8_t *h_filter = bilinear_filters_msa[xoffset];                   \
+  const uint8_t *v_filter = bilinear_filters_msa[yoffset];                   \
+                                                                             \
+  if (yoffset) {                                                             \
+    if (xoffset) {                                                           \
+      *sse = sub_pixel_avg_sse_diff_64width_hv_msa(src_ptr, src_stride,      \
+                                                   ref_ptr, ref_stride,      \
+                                                   sec_pred, h_filter,       \
+                                                   v_filter, ht, &diff);     \
+    } else {                                                                 \
+      *sse = sub_pixel_avg_sse_diff_64width_v_msa(src_ptr, src_stride,       \
+                                                  ref_ptr, ref_stride,       \
+                                                  sec_pred, v_filter,        \
+                                                  ht, &diff);                \
+    }                                                                        \
+  } else {                                                                   \
+    if (xoffset) {                                                           \
+      *sse = sub_pixel_avg_sse_diff_64width_h_msa(src_ptr, src_stride,       \
+                                                  ref_ptr, ref_stride,       \
+                                                  sec_pred, h_filter,        \
+                                                  ht, &diff);                \
+    } else {                                                                 \
+      *sse = avg_sse_diff_64x##ht##_msa(src_ptr, src_stride,                 \
+                                        ref_ptr, ref_stride,                 \
+                                        sec_pred, &diff);                    \
+    }                                                                        \
+  }                                                                          \
+                                                                             \
+  return VARIANCE_64Wx##ht##H(*sse, diff);                                   \
+}
+
+VPX_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(32);
+VPX_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(64);
diff --git a/libvpx/libvpx/vpx_dsp/mips/subtract_msa.c b/libvpx/libvpx/vpx_dsp/mips/subtract_msa.c
new file mode 100644
index 0000000..9ac43c5
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/subtract_msa.c
@@ -0,0 +1,264 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/macros_msa.h"
+
+static void sub_blk_4x4_msa(const uint8_t *src_ptr, int32_t src_stride,
+                            const uint8_t *pred_ptr, int32_t pred_stride,
+                            int16_t *diff_ptr, int32_t diff_stride) {
+  uint32_t src0, src1, src2, src3;
+  uint32_t pred0, pred1, pred2, pred3;
+  v16i8 src = { 0 };
+  v16i8 pred = { 0 };
+  v16u8 src_l0, src_l1;
+  v8i16 diff0, diff1;
+
+  LW4(src_ptr, src_stride, src0, src1, src2, src3);
+  LW4(pred_ptr, pred_stride, pred0, pred1, pred2, pred3);
+  INSERT_W4_SB(src0, src1, src2, src3, src);
+  INSERT_W4_SB(pred0, pred1, pred2, pred3, pred);
+  ILVRL_B2_UB(src, pred, src_l0, src_l1);
+  HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+  ST8x4_UB(diff0, diff1, diff_ptr, (2 * diff_stride));
+}
+
+static void sub_blk_8x8_msa(const uint8_t *src_ptr, int32_t src_stride,
+                            const uint8_t *pred_ptr, int32_t pred_stride,
+                            int16_t *diff_ptr, int32_t diff_stride) {
+  uint32_t loop_cnt;
+  uint64_t src0, src1, pred0, pred1;
+  v16i8 src = { 0 };
+  v16i8 pred = { 0 };
+  v16u8 src_l0, src_l1;
+  v8i16 diff0, diff1;
+
+  for (loop_cnt = 4; loop_cnt--;) {
+    LD2(src_ptr, src_stride, src0, src1);
+    src_ptr += (2 * src_stride);
+    LD2(pred_ptr, pred_stride, pred0, pred1);
+    pred_ptr += (2 * pred_stride);
+
+    INSERT_D2_SB(src0, src1, src);
+    INSERT_D2_SB(pred0, pred1, pred);
+    ILVRL_B2_UB(src, pred, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff_ptr, diff_stride);
+    diff_ptr += (2 * diff_stride);
+  }
+}
+
+static void sub_blk_16x16_msa(const uint8_t *src, int32_t src_stride,
+                              const uint8_t *pred, int32_t pred_stride,
+                              int16_t *diff, int32_t diff_stride) {
+  int8_t count;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16i8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7;
+  v16u8 src_l0, src_l1;
+  v8i16 diff0, diff1;
+
+  for (count = 2; count--;) {
+    LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    src += (8 * src_stride);
+
+    LD_SB8(pred, pred_stride,
+           pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7);
+    pred += (8 * pred_stride);
+
+    ILVRL_B2_UB(src0, pred0, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src1, pred1, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src2, pred2, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src3, pred3, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src4, pred4, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src5, pred5, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src6, pred6, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src7, pred7, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+  }
+}
+
+static void sub_blk_32x32_msa(const uint8_t *src, int32_t src_stride,
+                              const uint8_t *pred, int32_t pred_stride,
+                              int16_t *diff, int32_t diff_stride) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16i8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7;
+  v16u8 src_l0, src_l1;
+  v8i16 diff0, diff1;
+
+  for (loop_cnt = 8; loop_cnt--;) {
+    LD_SB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_SB2(src, 16, src2, src3);
+    src += src_stride;
+    LD_SB2(src, 16, src4, src5);
+    src += src_stride;
+    LD_SB2(src, 16, src6, src7);
+    src += src_stride;
+
+    LD_SB2(pred, 16, pred0, pred1);
+    pred += pred_stride;
+    LD_SB2(pred, 16, pred2, pred3);
+    pred += pred_stride;
+    LD_SB2(pred, 16, pred4, pred5);
+    pred += pred_stride;
+    LD_SB2(pred, 16, pred6, pred7);
+    pred += pred_stride;
+
+    ILVRL_B2_UB(src0, pred0, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src1, pred1, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src2, pred2, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src3, pred3, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src4, pred4, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src5, pred5, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src6, pred6, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src7, pred7, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    diff += diff_stride;
+  }
+}
+
+static void sub_blk_64x64_msa(const uint8_t *src, int32_t src_stride,
+                              const uint8_t *pred, int32_t pred_stride,
+                              int16_t *diff, int32_t diff_stride) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16i8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7;
+  v16u8 src_l0, src_l1;
+  v8i16 diff0, diff1;
+
+  for (loop_cnt = 32; loop_cnt--;) {
+    LD_SB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_SB4(src, 16, src4, src5, src6, src7);
+    src += src_stride;
+
+    LD_SB4(pred, 16, pred0, pred1, pred2, pred3);
+    pred += pred_stride;
+    LD_SB4(pred, 16, pred4, pred5, pred6, pred7);
+    pred += pred_stride;
+
+    ILVRL_B2_UB(src0, pred0, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src1, pred1, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    ILVRL_B2_UB(src2, pred2, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 32, 8);
+    ILVRL_B2_UB(src3, pred3, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 48, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src4, pred4, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src5, pred5, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    ILVRL_B2_UB(src6, pred6, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 32, 8);
+    ILVRL_B2_UB(src7, pred7, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 48, 8);
+    diff += diff_stride;
+  }
+}
+
+void vpx_subtract_block_msa(int32_t rows, int32_t cols,
+                            int16_t *diff_ptr, ptrdiff_t diff_stride,
+                            const uint8_t *src_ptr, ptrdiff_t src_stride,
+                            const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
+  if (rows == cols) {
+    switch (rows) {
+      case 4:
+        sub_blk_4x4_msa(src_ptr, src_stride, pred_ptr, pred_stride,
+                        diff_ptr, diff_stride);
+        break;
+      case 8:
+        sub_blk_8x8_msa(src_ptr, src_stride, pred_ptr, pred_stride,
+                        diff_ptr, diff_stride);
+        break;
+      case 16:
+        sub_blk_16x16_msa(src_ptr, src_stride, pred_ptr, pred_stride,
+                          diff_ptr, diff_stride);
+        break;
+      case 32:
+        sub_blk_32x32_msa(src_ptr, src_stride, pred_ptr, pred_stride,
+                          diff_ptr, diff_stride);
+        break;
+      case 64:
+        sub_blk_64x64_msa(src_ptr, src_stride, pred_ptr, pred_stride,
+                          diff_ptr, diff_stride);
+        break;
+      default:
+        vpx_subtract_block_c(rows, cols, diff_ptr, diff_stride, src_ptr,
+                             src_stride, pred_ptr, pred_stride);
+        break;
+    }
+  } else {
+    vpx_subtract_block_c(rows, cols, diff_ptr, diff_stride, src_ptr, src_stride,
+                         pred_ptr, pred_stride);
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/txfm_macros_msa.h b/libvpx/libvpx/vpx_dsp/mips/txfm_macros_msa.h
new file mode 100644
index 0000000..68c63d5
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/txfm_macros_msa.h
@@ -0,0 +1,93 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_MIPS_TXFM_MACROS_MIPS_MSA_H_
+#define VPX_DSP_MIPS_TXFM_MACROS_MIPS_MSA_H_
+
+#include "vpx_dsp/mips/macros_msa.h"
+
+#define DOTP_CONST_PAIR(reg0, reg1, cnst0, cnst1, out0, out1) {      \
+  v8i16 k0_m = __msa_fill_h(cnst0);                                  \
+  v4i32 s0_m, s1_m, s2_m, s3_m;                                      \
+                                                                     \
+  s0_m = (v4i32)__msa_fill_h(cnst1);                                 \
+  k0_m = __msa_ilvev_h((v8i16)s0_m, k0_m);                           \
+                                                                     \
+  ILVRL_H2_SW((-reg1), reg0, s1_m, s0_m);                            \
+  ILVRL_H2_SW(reg0, reg1, s3_m, s2_m);                               \
+  DOTP_SH2_SW(s1_m, s0_m, k0_m, k0_m, s1_m, s0_m);                   \
+  SRARI_W2_SW(s1_m, s0_m, DCT_CONST_BITS);                           \
+  out0 = __msa_pckev_h((v8i16)s0_m, (v8i16)s1_m);                    \
+                                                                     \
+  DOTP_SH2_SW(s3_m, s2_m, k0_m, k0_m, s1_m, s0_m);                   \
+  SRARI_W2_SW(s1_m, s0_m, DCT_CONST_BITS);                           \
+  out1 = __msa_pckev_h((v8i16)s0_m, (v8i16)s1_m);                    \
+}
+
+#define DOT_ADD_SUB_SRARI_PCK(in0, in1, in2, in3, in4, in5, in6, in7,      \
+                              dst0, dst1, dst2, dst3) {                    \
+  v4i32 tp0_m, tp1_m, tp2_m, tp3_m, tp4_m;                                 \
+  v4i32 tp5_m, tp6_m, tp7_m, tp8_m, tp9_m;                                 \
+                                                                           \
+  DOTP_SH4_SW(in0, in1, in0, in1, in4, in4, in5, in5,                      \
+              tp0_m, tp2_m, tp3_m, tp4_m);                                 \
+  DOTP_SH4_SW(in2, in3, in2, in3, in6, in6, in7, in7,                      \
+              tp5_m, tp6_m, tp7_m, tp8_m);                                 \
+  BUTTERFLY_4(tp0_m, tp3_m, tp7_m, tp5_m, tp1_m, tp9_m, tp7_m, tp5_m);     \
+  BUTTERFLY_4(tp2_m, tp4_m, tp8_m, tp6_m, tp3_m, tp0_m, tp4_m, tp2_m);     \
+  SRARI_W4_SW(tp1_m, tp9_m, tp7_m, tp5_m, DCT_CONST_BITS);                 \
+  SRARI_W4_SW(tp3_m, tp0_m, tp4_m, tp2_m, DCT_CONST_BITS);                 \
+  PCKEV_H4_SH(tp1_m, tp3_m, tp9_m, tp0_m, tp7_m, tp4_m, tp5_m, tp2_m,      \
+              dst0, dst1, dst2, dst3);                                     \
+}
+
+#define DOT_SHIFT_RIGHT_PCK_H(in0, in1, in2) ({       \
+  v8i16 dst_m;                                        \
+  v4i32 tp0_m, tp1_m;                                 \
+                                                      \
+  DOTP_SH2_SW(in0, in1, in2, in2, tp1_m, tp0_m);      \
+  SRARI_W2_SW(tp1_m, tp0_m, DCT_CONST_BITS);          \
+  dst_m = __msa_pckev_h((v8i16)tp1_m, (v8i16)tp0_m);  \
+                                                      \
+  dst_m;                                              \
+})
+
+#define MADD_SHORT(m0, m1, c0, c1, res0, res1) {                    \
+  v4i32 madd0_m, madd1_m, madd2_m, madd3_m;                         \
+  v8i16 madd_s0_m, madd_s1_m;                                       \
+                                                                    \
+  ILVRL_H2_SH(m1, m0, madd_s0_m, madd_s1_m);                        \
+  DOTP_SH4_SW(madd_s0_m, madd_s1_m, madd_s0_m, madd_s1_m,           \
+              c0, c0, c1, c1, madd0_m, madd1_m, madd2_m, madd3_m);  \
+  SRARI_W4_SW(madd0_m, madd1_m, madd2_m, madd3_m, DCT_CONST_BITS);  \
+  PCKEV_H2_SH(madd1_m, madd0_m, madd3_m, madd2_m, res0, res1);      \
+}
+
+#define MADD_BF(inp0, inp1, inp2, inp3, cst0, cst1, cst2, cst3,         \
+                out0, out1, out2, out3) {                               \
+  v8i16 madd_s0_m, madd_s1_m, madd_s2_m, madd_s3_m;                     \
+  v4i32 tmp0_m, tmp1_m, tmp2_m, tmp3_m, m4_m, m5_m;                     \
+                                                                        \
+  ILVRL_H2_SH(inp1, inp0, madd_s0_m, madd_s1_m);                        \
+  ILVRL_H2_SH(inp3, inp2, madd_s2_m, madd_s3_m);                        \
+  DOTP_SH4_SW(madd_s0_m, madd_s1_m, madd_s2_m, madd_s3_m,               \
+              cst0, cst0, cst2, cst2, tmp0_m, tmp1_m, tmp2_m, tmp3_m);  \
+  BUTTERFLY_4(tmp0_m, tmp1_m, tmp3_m, tmp2_m,                           \
+              m4_m, m5_m, tmp3_m, tmp2_m);                              \
+  SRARI_W4_SW(m4_m, m5_m, tmp2_m, tmp3_m, DCT_CONST_BITS);              \
+  PCKEV_H2_SH(m5_m, m4_m, tmp3_m, tmp2_m, out0, out1);                  \
+  DOTP_SH4_SW(madd_s0_m, madd_s1_m, madd_s2_m, madd_s3_m,               \
+              cst1, cst1, cst3, cst3, tmp0_m, tmp1_m, tmp2_m, tmp3_m);  \
+  BUTTERFLY_4(tmp0_m, tmp1_m, tmp3_m, tmp2_m,                           \
+              m4_m, m5_m, tmp3_m, tmp2_m);                              \
+  SRARI_W4_SW(m4_m, m5_m, tmp2_m, tmp3_m, DCT_CONST_BITS);              \
+  PCKEV_H2_SH(m5_m, m4_m, tmp3_m, tmp2_m, out2, out3);                  \
+}
+#endif  // VPX_DSP_MIPS_TXFM_MACROS_MIPS_MSA_H_
diff --git a/libvpx/libvpx/vpx_dsp/mips/variance_msa.c b/libvpx/libvpx/vpx_dsp/mips/variance_msa.c
new file mode 100644
index 0000000..33e1755
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/variance_msa.c
@@ -0,0 +1,633 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/macros_msa.h"
+
+#define CALC_MSE_B(src, ref, var) {                                \
+  v16u8 src_l0_m, src_l1_m;                                        \
+  v8i16 res_l0_m, res_l1_m;                                        \
+                                                                   \
+  ILVRL_B2_UB(src, ref, src_l0_m, src_l1_m);                       \
+  HSUB_UB2_SH(src_l0_m, src_l1_m, res_l0_m, res_l1_m);             \
+  DPADD_SH2_SW(res_l0_m, res_l1_m, res_l0_m, res_l1_m, var, var);  \
+}
+
+#define CALC_MSE_AVG_B(src, ref, var, sub) {                       \
+  v16u8 src_l0_m, src_l1_m;                                        \
+  v8i16 res_l0_m, res_l1_m;                                        \
+                                                                   \
+  ILVRL_B2_UB(src, ref, src_l0_m, src_l1_m);                       \
+  HSUB_UB2_SH(src_l0_m, src_l1_m, res_l0_m, res_l1_m);             \
+  DPADD_SH2_SW(res_l0_m, res_l1_m, res_l0_m, res_l1_m, var, var);  \
+                                                                   \
+  sub += res_l0_m + res_l1_m;                                      \
+}
+
+#define VARIANCE_WxH(sse, diff, shift) \
+  sse - (((uint32_t)diff * diff) >> shift)
+
+#define VARIANCE_LARGE_WxH(sse, diff, shift) \
+  sse - (((int64_t)diff * diff) >> shift)
+
+static uint32_t sse_diff_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                    const uint8_t *ref_ptr, int32_t ref_stride,
+                                    int32_t height, int32_t *diff) {
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  int32_t ht_cnt;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_8width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                    const uint8_t *ref_ptr, int32_t ref_stride,
+                                    int32_t height, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2,
+                src0, src1, ref0, ref1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_16width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                     const uint8_t *ref_ptr, int32_t ref_stride,
+                                     int32_t height, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src, ref;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_32width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                     const uint8_t *ref_ptr, int32_t ref_stride,
+                                     int32_t height, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_32x64_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                   const uint8_t *ref_ptr, int32_t ref_stride,
+                                   int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 16; ht_cnt--;) {
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_64x32_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                   const uint8_t *ref_ptr, int32_t ref_stride,
+                                   int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 16; ht_cnt--;) {
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src2, ref2, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src3, ref3, var, avg1);
+
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src2, ref2, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src3, ref3, var, avg1);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_64x64_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                   const uint8_t *ref_ptr, int32_t ref_stride,
+                                   int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v8i16 avg2 = { 0 };
+  v8i16 avg3 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 32; ht_cnt--;) {
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src2, ref2, var, avg2);
+    CALC_MSE_AVG_B(src3, ref3, var, avg3);
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src2, ref2, var, avg2);
+    CALC_MSE_AVG_B(src3, ref3, var, avg3);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+  vec += __msa_hadd_s_w(avg2, avg2);
+  vec += __msa_hadd_s_w(avg3, avg3);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t get_mb_ss_msa(const int16_t *src) {
+  uint32_t sum, cnt;
+  v8i16 src0, src1, src2, src3;
+  v4i32 src0_l, src1_l, src2_l, src3_l;
+  v4i32 src0_r, src1_r, src2_r, src3_r;
+  v2i64 sq_src_l = { 0 };
+  v2i64 sq_src_r = { 0 };
+
+  for (cnt = 8; cnt--;) {
+    LD_SH4(src, 8, src0, src1, src2, src3);
+    src += 4 * 8;
+
+    UNPCK_SH_SW(src0, src0_l, src0_r);
+    UNPCK_SH_SW(src1, src1_l, src1_r);
+    UNPCK_SH_SW(src2, src2_l, src2_r);
+    UNPCK_SH_SW(src3, src3_l, src3_r);
+
+    DPADD_SD2_SD(src0_l, src0_r, sq_src_l, sq_src_r);
+    DPADD_SD2_SD(src1_l, src1_r, sq_src_l, sq_src_r);
+    DPADD_SD2_SD(src2_l, src2_r, sq_src_l, sq_src_r);
+    DPADD_SD2_SD(src3_l, src3_r, sq_src_l, sq_src_r);
+  }
+
+  sq_src_l += __msa_splati_d(sq_src_l, 1);
+  sq_src_r += __msa_splati_d(sq_src_r, 1);
+
+  sum = __msa_copy_s_d(sq_src_l, 0);
+  sum += __msa_copy_s_d(sq_src_r, 0);
+
+  return sum;
+}
+
+static uint32_t sse_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v4i32 var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    CALC_MSE_B(src, ref, var);
+  }
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_8width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v4i32 var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2,
+                src0, src1, ref0, ref1);
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src1, ref1, var);
+  }
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_16width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                const uint8_t *ref_ptr, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src, ref;
+  v4i32 var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src, ref, var);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src, ref, var);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src, ref, var);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src, ref, var);
+  }
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_32width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                const uint8_t *ref_ptr, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v4i32 var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src1, ref1, var);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src1, ref1, var);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src1, ref1, var);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src1, ref1, var);
+  }
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_64width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                const uint8_t *ref_ptr, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v4i32 var = { 0 };
+
+  for (ht_cnt = height >> 1; ht_cnt--;) {
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src2, ref2, var);
+    CALC_MSE_B(src1, ref1, var);
+    CALC_MSE_B(src3, ref3, var);
+
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src2, ref2, var);
+    CALC_MSE_B(src1, ref1, var);
+    CALC_MSE_B(src3, ref3, var);
+  }
+
+  return HADD_SW_S32(var);
+}
+
+uint32_t vpx_get4x4sse_cs_msa(const uint8_t *src_ptr, int32_t src_stride,
+                              const uint8_t *ref_ptr, int32_t ref_stride) {
+  uint32_t err = 0;
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16i8 src = { 0 };
+  v16i8 ref = { 0 };
+  v16u8 src_vec0, src_vec1;
+  v8i16 diff0, diff1;
+  v4i32 err0 = { 0 };
+  v4i32 err1 = { 0 };
+
+  LW4(src_ptr, src_stride, src0, src1, src2, src3);
+  LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+  INSERT_W4_SB(src0, src1, src2, src3, src);
+  INSERT_W4_SB(ref0, ref1, ref2, ref3, ref);
+  ILVRL_B2_UB(src, ref, src_vec0, src_vec1);
+  HSUB_UB2_SH(src_vec0, src_vec1, diff0, diff1);
+  DPADD_SH2_SW(diff0, diff1, diff0, diff1, err0, err1);
+  err = HADD_SW_S32(err0);
+  err += HADD_SW_S32(err1);
+
+  return err;
+}
+
+#define VARIANCE_4Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 4);
+#define VARIANCE_4Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 5);
+#define VARIANCE_8Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 5);
+#define VARIANCE_8Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 6);
+#define VARIANCE_8Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 7);
+#define VARIANCE_16Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 7);
+#define VARIANCE_16Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 8);
+
+#define VARIANCE_16Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
+#define VARIANCE_32Wx16H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
+#define VARIANCE_32Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 10);
+#define VARIANCE_32Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
+#define VARIANCE_64Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
+#define VARIANCE_64Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 12);
+
+#define VPX_VARIANCE_WDXHT_MSA(wd, ht)                               \
+uint32_t vpx_variance##wd##x##ht##_msa(const uint8_t *src,           \
+                                       int32_t src_stride,           \
+                                       const uint8_t *ref,           \
+                                       int32_t ref_stride,           \
+                                       uint32_t *sse) {              \
+  int32_t diff;                                                      \
+                                                                     \
+  *sse = sse_diff_##wd##width_msa(src, src_stride, ref, ref_stride,  \
+                                  ht, &diff);                        \
+                                                                     \
+  return VARIANCE_##wd##Wx##ht##H(*sse, diff);                       \
+}
+
+VPX_VARIANCE_WDXHT_MSA(4, 4);
+VPX_VARIANCE_WDXHT_MSA(4, 8);
+
+VPX_VARIANCE_WDXHT_MSA(8, 4)
+VPX_VARIANCE_WDXHT_MSA(8, 8)
+VPX_VARIANCE_WDXHT_MSA(8, 16)
+
+VPX_VARIANCE_WDXHT_MSA(16, 8)
+VPX_VARIANCE_WDXHT_MSA(16, 16)
+VPX_VARIANCE_WDXHT_MSA(16, 32)
+
+VPX_VARIANCE_WDXHT_MSA(32, 16)
+VPX_VARIANCE_WDXHT_MSA(32, 32)
+
+uint32_t vpx_variance32x64_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               uint32_t *sse) {
+  int32_t diff;
+
+  *sse = sse_diff_32x64_msa(src, src_stride, ref, ref_stride, &diff);
+
+  return VARIANCE_32Wx64H(*sse, diff);
+}
+
+uint32_t vpx_variance64x32_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               uint32_t *sse) {
+  int32_t diff;
+
+  *sse = sse_diff_64x32_msa(src, src_stride, ref, ref_stride, &diff);
+
+  return VARIANCE_64Wx32H(*sse, diff);
+}
+
+uint32_t vpx_variance64x64_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               uint32_t *sse) {
+  int32_t diff;
+
+  *sse = sse_diff_64x64_msa(src, src_stride, ref, ref_stride, &diff);
+
+  return VARIANCE_64Wx64H(*sse, diff);
+}
+
+uint32_t vpx_mse8x8_msa(const uint8_t *src, int32_t src_stride,
+                        const uint8_t *ref, int32_t ref_stride,
+                        uint32_t *sse) {
+  *sse = sse_8width_msa(src, src_stride, ref, ref_stride, 8);
+
+  return *sse;
+}
+
+uint32_t vpx_mse8x16_msa(const uint8_t *src, int32_t src_stride,
+                         const uint8_t *ref, int32_t ref_stride,
+                         uint32_t *sse) {
+  *sse = sse_8width_msa(src, src_stride, ref, ref_stride, 16);
+
+  return *sse;
+}
+
+uint32_t vpx_mse16x8_msa(const uint8_t *src, int32_t src_stride,
+                         const uint8_t *ref, int32_t ref_stride,
+                         uint32_t *sse) {
+  *sse = sse_16width_msa(src, src_stride, ref, ref_stride, 8);
+
+  return *sse;
+}
+
+uint32_t vpx_mse16x16_msa(const uint8_t *src, int32_t src_stride,
+                          const uint8_t *ref, int32_t ref_stride,
+                          uint32_t *sse) {
+  *sse = sse_16width_msa(src, src_stride, ref, ref_stride, 16);
+
+  return *sse;
+}
+
+void vpx_get8x8var_msa(const uint8_t *src, int32_t src_stride,
+                       const uint8_t *ref, int32_t ref_stride,
+                       uint32_t *sse, int32_t *sum) {
+  *sse = sse_diff_8width_msa(src, src_stride, ref, ref_stride, 8, sum);
+}
+
+void vpx_get16x16var_msa(const uint8_t *src, int32_t src_stride,
+                         const uint8_t *ref, int32_t ref_stride,
+                         uint32_t *sse, int32_t *sum) {
+  *sse = sse_diff_16width_msa(src, src_stride, ref, ref_stride, 16, sum);
+}
+
+uint32_t vpx_get_mb_ss_msa(const int16_t *src) {
+  return get_mb_ss_msa(src);
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_avg_horiz_msa.c b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_avg_horiz_msa.c
new file mode 100644
index 0000000..f6244d8
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_avg_horiz_msa.c
@@ -0,0 +1,743 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/vpx_convolve_msa.h"
+
+static void common_hz_8t_and_aver_dst_4x4_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 dst0, dst1, dst2, dst3, res2, res3;
+  v16u8 mask0, mask1, mask2, mask3;
+  v8i16 filt, res0, res1;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, res0, res1);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  SRARI_H2_SH(res0, res1, FILTER_BITS);
+  SAT_SH2_SH(res0, res1, 7);
+  PCKEV_B2_UB(res0, res0, res1, res1, res2, res3);
+  ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
+  XORI_B2_128_UB(res2, res3);
+  AVER_UB2_UB(res2, dst0, res3, dst2, res2, res3);
+  ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_8t_and_aver_dst_4x8_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, res0, res1, res2, res3;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v8i16 filt, vec0, vec1, vec2, vec3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  src += (4 * src_stride);
+  LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, vec0, vec1);
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, vec2, vec3);
+  SRARI_H4_SH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  SAT_SH4_SH(vec0, vec1, vec2, vec3, 7);
+  PCKEV_B4_UB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3, res0, res1, res2,
+              res3);
+  ILVR_D2_UB(res1, res0, res3, res2, res0, res2);
+  XORI_B2_128_UB(res0, res2);
+  ILVR_W4_UB(dst1, dst0, dst3, dst2, dst5, dst4, dst7, dst6, dst0, dst2, dst4,
+             dst6);
+  ILVR_D2_UB(dst2, dst0, dst6, dst4, dst0, dst4);
+  AVER_UB2_UB(res0, dst0, res2, dst4, res0, res2);
+  ST4x8_UB(res0, res2, dst, dst_stride);
+}
+
+static void common_hz_8t_and_aver_dst_4w_msa(const uint8_t *src,
+                                             int32_t src_stride,
+                                             uint8_t *dst,
+                                             int32_t dst_stride,
+                                             int8_t *filter,
+                                             int32_t height) {
+  if (4 == height) {
+    common_hz_8t_and_aver_dst_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_hz_8t_and_aver_dst_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_hz_8t_and_aver_dst_8w_msa(const uint8_t *src,
+                                             int32_t src_stride,
+                                             uint8_t *dst,
+                                             int32_t dst_stride,
+                                             int8_t *filter,
+                                             int32_t height) {
+  int32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, dst0, dst1, dst2, dst3;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    src += (4 * src_stride);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    CONVERT_UB_AVG_ST8x4_UB(out0, out1, out2, out3, dst0, dst1, dst2, dst3,
+                            dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void common_hz_8t_and_aver_dst_16w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  int32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, dst0, dst1;
+  v8i16 filt, out0, out1, out2, out3;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8i16 vec8, vec9, vec10, vec11, vec12, vec13, vec14, vec15;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = height >> 1; loop_cnt--;) {
+    LD_SB2(src, src_stride, src0, src2);
+    LD_SB2(src + 8, src_stride, src1, src3);
+    src += (2 * src_stride);
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    VSHF_B4_SH(src0, src0, mask0, mask1, mask2, mask3, vec0, vec4, vec8, vec12);
+    VSHF_B4_SH(src1, src1, mask0, mask1, mask2, mask3, vec1, vec5, vec9, vec13);
+    VSHF_B4_SH(src2, src2, mask0, mask1, mask2, mask3, vec2, vec6, vec10,
+               vec14);
+    VSHF_B4_SH(src3, src3, mask0, mask1, mask2, mask3, vec3, vec7, vec11,
+               vec15);
+    DOTP_SB4_SH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    DOTP_SB4_SH(vec8, vec9, vec10, vec11, filt2, filt2, filt2, filt2, vec8,
+                vec9, vec10, vec11);
+    DPADD_SB4_SH(vec4, vec5, vec6, vec7, filt1, filt1, filt1, filt1, vec0, vec1,
+                 vec2, vec3);
+    DPADD_SB4_SH(vec12, vec13, vec14, vec15, filt3, filt3, filt3, filt3, vec8,
+                 vec9, vec10, vec11);
+    ADDS_SH4_SH(vec0, vec8, vec1, vec9, vec2, vec10, vec3, vec11, out0, out1,
+                out2, out3);
+    LD_UB2(dst, dst_stride, dst0, dst1);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    PCKEV_XORI128_AVG_ST_UB(out1, out0, dst0, dst);
+    dst += dst_stride;
+    PCKEV_XORI128_AVG_ST_UB(out3, out2, dst1, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_8t_and_aver_dst_32w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 dst1, dst2, mask0, mask1, mask2, mask3;
+  v8i16 filt, out0, out1, out2, out3;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8i16 vec8, vec9, vec10, vec11, vec12, vec13, vec14, vec15;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = height; loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    VSHF_B4_SH(src0, src0, mask0, mask1, mask2, mask3, vec0, vec4, vec8, vec12);
+    VSHF_B4_SH(src1, src1, mask0, mask1, mask2, mask3, vec1, vec5, vec9, vec13);
+    VSHF_B4_SH(src2, src2, mask0, mask1, mask2, mask3, vec2, vec6, vec10,
+               vec14);
+    VSHF_B4_SH(src3, src3, mask0, mask1, mask2, mask3, vec3, vec7, vec11,
+               vec15);
+    DOTP_SB4_SH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    DOTP_SB4_SH(vec8, vec9, vec10, vec11, filt2, filt2, filt2, filt2, vec8,
+                vec9, vec10, vec11);
+    DPADD_SB4_SH(vec4, vec5, vec6, vec7, filt1, filt1, filt1, filt1, vec0, vec1,
+                 vec2, vec3);
+    DPADD_SB4_SH(vec12, vec13, vec14, vec15, filt3, filt3, filt3, filt3, vec8,
+                 vec9, vec10, vec11);
+    ADDS_SH4_SH(vec0, vec8, vec1, vec9, vec2, vec10, vec3, vec11, out0, out1,
+                out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    LD_UB2(dst, 16, dst1, dst2);
+    PCKEV_XORI128_AVG_ST_UB(out1, out0, dst1, dst);
+    PCKEV_XORI128_AVG_ST_UB(out3, out2, dst2, dst + 16);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_8t_and_aver_dst_64w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  uint32_t loop_cnt, cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 dst1, dst2, mask0, mask1, mask2, mask3;
+  v8i16 filt, out0, out1, out2, out3;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8i16 vec8, vec9, vec10, vec11, vec12, vec13, vec14, vec15;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = height; loop_cnt--;) {
+    for (cnt = 0; cnt < 2; ++cnt) {
+      src0 = LD_SB(&src[cnt << 5]);
+      src2 = LD_SB(&src[16 + (cnt << 5)]);
+      src3 = LD_SB(&src[24 + (cnt << 5)]);
+      src1 = __msa_sldi_b(src2, src0, 8);
+
+      XORI_B4_128_SB(src0, src1, src2, src3);
+      VSHF_B4_SH(src0, src0, mask0, mask1, mask2, mask3, vec0, vec4, vec8,
+                 vec12);
+      VSHF_B4_SH(src1, src1, mask0, mask1, mask2, mask3, vec1, vec5, vec9,
+                 vec13);
+      VSHF_B4_SH(src2, src2, mask0, mask1, mask2, mask3, vec2, vec6, vec10,
+                 vec14);
+      VSHF_B4_SH(src3, src3, mask0, mask1, mask2, mask3, vec3, vec7, vec11,
+                 vec15);
+      DOTP_SB4_SH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0,
+                  vec1, vec2, vec3);
+      DOTP_SB4_SH(vec8, vec9, vec10, vec11, filt2, filt2, filt2, filt2, vec8,
+                  vec9, vec10, vec11);
+      DPADD_SB4_SH(vec4, vec5, vec6, vec7, filt1, filt1, filt1, filt1, vec0,
+                   vec1, vec2, vec3);
+      DPADD_SB4_SH(vec12, vec13, vec14, vec15, filt3, filt3, filt3, filt3, vec8,
+                   vec9, vec10, vec11);
+      ADDS_SH4_SH(vec0, vec8, vec1, vec9, vec2, vec10, vec3, vec11, out0, out1,
+                  out2, out3);
+      SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+      SAT_SH4_SH(out0, out1, out2, out3, 7);
+      LD_UB2(&dst[cnt << 5], 16, dst1, dst2);
+      PCKEV_XORI128_AVG_ST_UB(out1, out0, dst1, &dst[cnt << 5]);
+      PCKEV_XORI128_AVG_ST_UB(out3, out2, dst2, &dst[16 + (cnt << 5)]);
+    }
+
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_and_aver_dst_4x4_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3, vec0, vec1, res0, res1;
+  v8u16 vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
+  DOTP_UB2_UH(vec0, vec1, filt0, filt0, vec2, vec3);
+  SRARI_H2_UH(vec2, vec3, FILTER_BITS);
+  PCKEV_B2_UB(vec2, vec2, vec3, vec3, res0, res1);
+  ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
+  AVER_UB2_UB(res0, dst0, res1, dst2, res0, res1);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_2t_and_aver_dst_4x8_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, vec0, vec1, vec2, vec3, res0, res1, res2, res3;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v8u16 vec4, vec5, vec6, vec7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+  VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
+  VSHF_B2_UB(src4, src5, src6, src7, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec4, vec5,
+              vec6, vec7);
+  SRARI_H4_UH(vec4, vec5, vec6, vec7, FILTER_BITS);
+  PCKEV_B4_UB(vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7, res0, res1, res2,
+              res3);
+  ILVR_W4_UB(dst1, dst0, dst3, dst2, dst5, dst4, dst7, dst6, dst0, dst2, dst4,
+             dst6);
+  AVER_UB4_UB(res0, dst0, res1, dst2, res2, dst4, res3, dst6, res0, res1, res2,
+              res3);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+  dst += (4 * dst_stride);
+  ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_2t_and_aver_dst_4w_msa(const uint8_t *src,
+                                             int32_t src_stride,
+                                             uint8_t *dst,
+                                             int32_t dst_stride,
+                                             int8_t *filter,
+                                             int32_t height) {
+  if (4 == height) {
+    common_hz_2t_and_aver_dst_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_hz_2t_and_aver_dst_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_hz_2t_and_aver_dst_8x4_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3;
+  v8u16 vec0, vec1, vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  PCKEV_AVG_ST8x4_UB(vec0, dst0, vec1, dst1, vec2, dst2, vec3, dst3,
+                     dst, dst_stride);
+}
+
+static void common_hz_2t_and_aver_dst_8x8mult_msa(const uint8_t *src,
+                                                  int32_t src_stride,
+                                                  uint8_t *dst,
+                                                  int32_t dst_stride,
+                                                  int8_t *filter,
+                                                  int32_t height) {
+  v16i8 src0, src1, src2, src3, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3;
+  v8u16 vec0, vec1, vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  src += (4 * src_stride);
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  src += (4 * src_stride);
+  PCKEV_AVG_ST8x4_UB(vec0, dst0, vec1, dst1, vec2, dst2, vec3, dst3,
+                     dst, dst_stride);
+  dst += (4 * dst_stride);
+
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  PCKEV_AVG_ST8x4_UB(vec0, dst0, vec1, dst1, vec2, dst2, vec3, dst3,
+                     dst, dst_stride);
+  dst += (4 * dst_stride);
+
+  if (16 == height) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    PCKEV_AVG_ST8x4_UB(vec0, dst0, vec1, dst1, vec2, dst2, vec3, dst3,
+                       dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    PCKEV_AVG_ST8x4_UB(vec0, dst0, vec1, dst1, vec2, dst2, vec3, dst3,
+                       dst, dst_stride);
+  }
+}
+
+static void common_hz_2t_and_aver_dst_8w_msa(const uint8_t *src,
+                                             int32_t src_stride,
+                                             uint8_t *dst,
+                                             int32_t dst_stride,
+                                             int8_t *filter,
+                                             int32_t height) {
+  if (4 == height) {
+    common_hz_2t_and_aver_dst_8x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else {
+    common_hz_2t_and_aver_dst_8x8mult_msa(src, src_stride, dst, dst_stride,
+                                          filter, height);
+  }
+}
+
+static void common_hz_2t_and_aver_dst_16w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 res0, res1, res2, res3, res4, res5, res6, res7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src2, src4, src6);
+  LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+  src += (4 * src_stride);
+
+  VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+  VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, res0, res1,
+              res2, res3);
+  DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, res4, res5,
+              res6, res7);
+  SRARI_H4_UH(res0, res1, res2, res3, FILTER_BITS);
+  SRARI_H4_UH(res4, res5, res6, res7, FILTER_BITS);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  PCKEV_AVG_ST_UB(res1, res0, dst0, dst);
+  dst += dst_stride;
+  PCKEV_AVG_ST_UB(res3, res2, dst1, dst);
+  dst += dst_stride;
+  PCKEV_AVG_ST_UB(res5, res4, dst2, dst);
+  dst += dst_stride;
+  PCKEV_AVG_ST_UB(res7, res6, dst3, dst);
+  dst += dst_stride;
+
+  for (loop_cnt = (height >> 2) - 1; loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, res0, res1,
+                res2, res3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, res4, res5,
+                res6, res7);
+    SRARI_H4_UH(res0, res1, res2, res3, FILTER_BITS);
+    SRARI_H4_UH(res4, res5, res6, res7, FILTER_BITS);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    PCKEV_AVG_ST_UB(res1, res0, dst0, dst);
+    dst += dst_stride;
+    PCKEV_AVG_ST_UB(res3, res2, dst1, dst);
+    dst += dst_stride;
+    PCKEV_AVG_ST_UB(res5, res4, dst2, dst);
+    dst += dst_stride;
+    PCKEV_AVG_ST_UB(res7, res6, dst3, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_and_aver_dst_32w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 res0, res1, res2, res3, res4, res5, res6, res7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  for (loop_cnt = (height >> 1); loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+    src4 = LD_SB(src);
+    src6 = LD_SB(src + 16);
+    src7 = LD_SB(src + 24);
+    src5 = __msa_sldi_b(src6, src4, 8);
+    src += src_stride;
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, res0, res1,
+                res2, res3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, res4, res5,
+                res6, res7);
+    SRARI_H4_UH(res0, res1, res2, res3, FILTER_BITS);
+    SRARI_H4_UH(res4, res5, res6, res7, FILTER_BITS);
+    LD_UB2(dst, 16, dst0, dst1);
+    PCKEV_AVG_ST_UB(res1, res0, dst0, dst);
+    PCKEV_AVG_ST_UB(res3, res2, dst1, (dst + 16));
+    dst += dst_stride;
+    LD_UB2(dst, 16, dst2, dst3);
+    PCKEV_AVG_ST_UB(res5, res4, dst2, dst);
+    PCKEV_AVG_ST_UB(res7, res6, dst3, (dst + 16));
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_and_aver_dst_64w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  for (loop_cnt = height; loop_cnt--;) {
+    LD_SB4(src, 16, src0, src2, src4, src6);
+    src7 = LD_SB(src + 56);
+    SLDI_B3_SB(src2, src4, src6, src0, src2, src4, src1, src3, src5, 8);
+    src += src_stride;
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+                out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+                out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    LD_UB4(dst, 16, dst0, dst1, dst2, dst3);
+    PCKEV_AVG_ST_UB(out1, out0, dst0, dst);
+    PCKEV_AVG_ST_UB(out3, out2, dst1, dst + 16);
+    PCKEV_AVG_ST_UB(out5, out4, dst2, dst + 32);
+    PCKEV_AVG_ST_UB(out7, out6, dst3, dst + 48);
+    dst += dst_stride;
+  }
+}
+
+void vpx_convolve8_avg_horiz_msa(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const int16_t *filter_x, int x_step_q4,
+                                 const int16_t *filter_y, int y_step_q4,
+                                 int w, int h) {
+  int8_t cnt, filt_hor[8];
+
+  assert(x_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+
+  for (cnt = 0; cnt < 8; ++cnt) {
+    filt_hor[cnt] = filter_x[cnt];
+  }
+
+  if (((const int32_t *)filter_x)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_hz_2t_and_aver_dst_4w_msa(src, (int32_t)src_stride,
+                                         dst, (int32_t)dst_stride,
+                                         &filt_hor[3], h);
+        break;
+      case 8:
+        common_hz_2t_and_aver_dst_8w_msa(src, (int32_t)src_stride,
+                                         dst, (int32_t)dst_stride,
+                                         &filt_hor[3], h);
+        break;
+      case 16:
+        common_hz_2t_and_aver_dst_16w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          &filt_hor[3], h);
+        break;
+      case 32:
+        common_hz_2t_and_aver_dst_32w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          &filt_hor[3], h);
+        break;
+      case 64:
+        common_hz_2t_and_aver_dst_64w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          &filt_hor[3], h);
+        break;
+      default:
+        vpx_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride,
+                                  filter_x, x_step_q4, filter_y, y_step_q4,
+                                  w, h);
+        break;
+    }
+  } else {
+    switch (w) {
+      case 4:
+        common_hz_8t_and_aver_dst_4w_msa(src, (int32_t)src_stride,
+                                         dst, (int32_t)dst_stride,
+                                         filt_hor, h);
+        break;
+      case 8:
+        common_hz_8t_and_aver_dst_8w_msa(src, (int32_t)src_stride,
+                                         dst, (int32_t)dst_stride,
+                                         filt_hor, h);
+        break;
+      case 16:
+        common_hz_8t_and_aver_dst_16w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          filt_hor, h);
+        break;
+      case 32:
+        common_hz_8t_and_aver_dst_32w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          filt_hor, h);
+        break;
+      case 64:
+        common_hz_8t_and_aver_dst_64w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          filt_hor, h);
+        break;
+      default:
+        vpx_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride,
+                                  filter_x, x_step_q4, filter_y, y_step_q4,
+                                  w, h);
+        break;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_avg_msa.c b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_avg_msa.c
new file mode 100644
index 0000000..2abde6d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_avg_msa.c
@@ -0,0 +1,661 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/vpx_convolve_msa.h"
+
+static void common_hv_8ht_8vt_and_aver_dst_4w_msa(const uint8_t *src,
+                                                  int32_t src_stride,
+                                                  uint8_t *dst,
+                                                  int32_t dst_stride,
+                                                  int8_t *filter_horiz,
+                                                  int8_t *filter_vert,
+                                                  int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16u8 dst0, dst1, dst2, dst3, mask0, mask1, mask2, mask3, tmp0, tmp1;
+  v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3;
+  v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8i16 hz_out7, hz_out8, hz_out9, res0, res1, vec0, vec1, vec2, vec3, vec4;
+  v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= (3 + 3 * src_stride);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  hz_out0 = HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out2 = HORIZ_8TAP_FILT(src2, src3, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out4 = HORIZ_8TAP_FILT(src4, src5, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out5 = HORIZ_8TAP_FILT(src5, src6, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  SLDI_B2_SH(hz_out2, hz_out4, hz_out0, hz_out2, hz_out1, hz_out3, 8);
+
+  filt = LD_SH(filter_vert);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
+
+  ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+  vec2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    hz_out7 = HORIZ_8TAP_FILT(src7, src8, mask0, mask1, mask2, mask3,
+                              filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    hz_out6 = (v8i16)__msa_sldi_b((v16i8)hz_out7, (v16i8)hz_out5, 8);
+    vec3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+    res0 = FILT_8TAP_DPADD_S_H(vec0, vec1, vec2, vec3, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out9 = HORIZ_8TAP_FILT(src9, src10, mask0, mask1, mask2, mask3,
+                              filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    hz_out8 = (v8i16)__msa_sldi_b((v16i8)hz_out9, (v16i8)hz_out7, 8);
+    vec4 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8);
+    res1 = FILT_8TAP_DPADD_S_H(vec1, vec2, vec3, vec4, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+    ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
+
+    SRARI_H2_SH(res0, res1, FILTER_BITS);
+    SAT_SH2_SH(res0, res1, 7);
+    PCKEV_B2_UB(res0, res0, res1, res1, tmp0, tmp1);
+    XORI_B2_128_UB(tmp0, tmp1);
+    AVER_UB2_UB(tmp0, dst0, tmp1, dst2, tmp0, tmp1);
+    ST4x4_UB(tmp0, tmp1, 0, 1, 0, 1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    hz_out5 = hz_out9;
+    vec0 = vec2;
+    vec1 = vec3;
+    vec2 = vec4;
+  }
+}
+
+static void common_hv_8ht_8vt_and_aver_dst_8w_msa(const uint8_t *src,
+                                                  int32_t src_stride,
+                                                  uint8_t *dst,
+                                                  int32_t dst_stride,
+                                                  int8_t *filter_horiz,
+                                                  int8_t *filter_vert,
+                                                  int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3;
+  v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3;
+  v16u8 dst0, dst1, dst2, dst3, mask0, mask1, mask2, mask3;
+  v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8i16 hz_out7, hz_out8, hz_out9, hz_out10, tmp0, tmp1, tmp2, tmp3;
+  v8i16 out0, out1, out2, out3, out4, out5, out6, out7, out8, out9;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= (3 + 3 * src_stride);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  hz_out0 = HORIZ_8TAP_FILT(src0, src0, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out1 = HORIZ_8TAP_FILT(src1, src1, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out2 = HORIZ_8TAP_FILT(src2, src2, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out3 = HORIZ_8TAP_FILT(src3, src3, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out4 = HORIZ_8TAP_FILT(src4, src4, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out5 = HORIZ_8TAP_FILT(src5, src5, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out6 = HORIZ_8TAP_FILT(src6, src6, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+
+  filt = LD_SH(filter_vert);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
+
+  ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
+  ILVEV_B2_SH(hz_out4, hz_out5, hz_out1, hz_out2, out2, out4);
+  ILVEV_B2_SH(hz_out3, hz_out4, hz_out5, hz_out6, out5, out6);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+
+    hz_out7 = HORIZ_8TAP_FILT(src7, src7, mask0, mask1, mask2, mask3,
+                              filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+    tmp0 = FILT_8TAP_DPADD_S_H(out0, out1, out2, out3, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out8 = HORIZ_8TAP_FILT(src8, src8, mask0, mask1, mask2, mask3,
+                              filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    out7 = (v8i16)__msa_ilvev_b((v16i8)hz_out8, (v16i8)hz_out7);
+    tmp1 = FILT_8TAP_DPADD_S_H(out4, out5, out6, out7, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out9 = HORIZ_8TAP_FILT(src9, src9, mask0, mask1, mask2, mask3,
+                              filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    out8 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8);
+    tmp2 = FILT_8TAP_DPADD_S_H(out1, out2, out3, out8, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out10 = HORIZ_8TAP_FILT(src10, src10, mask0, mask1, mask2, mask3,
+                               filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    out9 = (v8i16)__msa_ilvev_b((v16i8)hz_out10, (v16i8)hz_out9);
+    tmp3 = FILT_8TAP_DPADD_S_H(out5, out6, out7, out9, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    SRARI_H4_SH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    SAT_SH4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+    CONVERT_UB_AVG_ST8x4_UB(tmp0, tmp1, tmp2, tmp3, dst0, dst1, dst2, dst3,
+                            dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    hz_out6 = hz_out10;
+    out0 = out2;
+    out1 = out3;
+    out2 = out8;
+    out4 = out6;
+    out5 = out7;
+    out6 = out9;
+  }
+}
+
+static void common_hv_8ht_8vt_and_aver_dst_16w_msa(const uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *dst,
+                                                   int32_t dst_stride,
+                                                   int8_t *filter_horiz,
+                                                   int8_t *filter_vert,
+                                                   int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 2; multiple8_cnt--;) {
+    common_hv_8ht_8vt_and_aver_dst_8w_msa(src, src_stride, dst, dst_stride,
+                                          filter_horiz, filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_8ht_8vt_and_aver_dst_32w_msa(const uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *dst,
+                                                   int32_t dst_stride,
+                                                   int8_t *filter_horiz,
+                                                   int8_t *filter_vert,
+                                                   int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 4; multiple8_cnt--;) {
+    common_hv_8ht_8vt_and_aver_dst_8w_msa(src, src_stride, dst, dst_stride,
+                                          filter_horiz, filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_8ht_8vt_and_aver_dst_64w_msa(const uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *dst,
+                                                   int32_t dst_stride,
+                                                   int8_t *filter_horiz,
+                                                   int8_t *filter_vert,
+                                                   int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 8; multiple8_cnt--;) {
+    common_hv_8ht_8vt_and_aver_dst_8w_msa(src, src_stride, dst, dst_stride,
+                                          filter_horiz, filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_4x4_msa(const uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *dst,
+                                                   int32_t dst_stride,
+                                                   int8_t *filter_horiz,
+                                                   int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, mask;
+  v16u8 filt_hz, filt_vt, vec0, vec1;
+  v16u8 dst0, dst1, dst2, dst3, res0, res1;
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, tmp0, tmp1, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  filt = LD_UH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+  hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+  hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+  hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2);
+  ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
+  DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+  SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+  PCKEV_B2_UB(tmp0, tmp0, tmp1, tmp1, res0, res1);
+  AVER_UB2_UB(res0, dst0, res1, dst2, res0, res1);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_4x8_msa(const uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *dst,
+                                                   int32_t dst_stride,
+                                                   int8_t *filter_horiz,
+                                                   int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, mask;
+  v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3, res0, res1, res2, res3;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8u16 hz_out7, hz_out8, tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  src += (8 * src_stride);
+  src8 = LD_SB(src);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+  hz_out4 = HORIZ_2TAP_FILT_UH(src4, src5, mask, filt_hz, FILTER_BITS);
+  hz_out6 = HORIZ_2TAP_FILT_UH(src6, src7, mask, filt_hz, FILTER_BITS);
+  hz_out8 = HORIZ_2TAP_FILT_UH(src8, src8, mask, filt_hz, FILTER_BITS);
+  SLDI_B3_UH(hz_out2, hz_out4, hz_out6, hz_out0, hz_out2, hz_out4, hz_out1,
+             hz_out3, hz_out5, 8);
+  hz_out7 = (v8u16)__msa_pckod_d((v2i64)hz_out8, (v2i64)hz_out6);
+
+  LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+  ILVR_W4_UB(dst1, dst0, dst3, dst2, dst5, dst4, dst7, dst6, dst0, dst2,
+             dst4, dst6);
+  ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+  ILVEV_B2_UB(hz_out4, hz_out5, hz_out6, hz_out7, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt_vt, filt_vt, filt_vt, filt_vt,
+              tmp0, tmp1, tmp2, tmp3);
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_B4_UB(tmp0, tmp0, tmp1, tmp1, tmp2, tmp2, tmp3, tmp3, res0, res1,
+              res2, res3);
+  AVER_UB4_UB(res0, dst0, res1, dst2, res2, dst4, res3, dst6, res0, res1,
+              res2, res3);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+  dst += (4 * dst_stride);
+  ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_4w_msa(const uint8_t *src,
+                                                  int32_t src_stride,
+                                                  uint8_t *dst,
+                                                  int32_t dst_stride,
+                                                  int8_t *filter_horiz,
+                                                  int8_t *filter_vert,
+                                                  int32_t height) {
+  if (4 == height) {
+    common_hv_2ht_2vt_and_aver_dst_4x4_msa(src, src_stride, dst, dst_stride,
+                                           filter_horiz, filter_vert);
+  } else if (8 == height) {
+    common_hv_2ht_2vt_and_aver_dst_4x8_msa(src, src_stride, dst, dst_stride,
+                                           filter_horiz, filter_vert);
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_8x4_msa(const uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *dst,
+                                                   int32_t dst_stride,
+                                                   int8_t *filter_horiz,
+                                                   int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, mask;
+  v16u8 filt_hz, filt_vt, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3;
+  v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+  src += (5 * src_stride);
+
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+  vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+  tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+  vec1 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+  tmp1 = __msa_dotp_u_h(vec1, filt_vt);
+
+  hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+  vec2 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+  tmp2 = __msa_dotp_u_h(vec2, filt_vt);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+  vec3 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+  tmp3 = __msa_dotp_u_h(vec3, filt_vt);
+
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_AVG_ST8x4_UB(tmp0, dst0, tmp1, dst1, tmp2, dst2, tmp3, dst3,
+                     dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_8x8mult_msa(const uint8_t *src,
+                                                       int32_t src_stride,
+                                                       uint8_t *dst,
+                                                       int32_t dst_stride,
+                                                       int8_t *filter_horiz,
+                                                       int8_t *filter_vert,
+                                                       int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, mask;
+  v16u8 filt_hz, filt_vt, vec0, dst0, dst1, dst2, dst3;
+  v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_SB(src);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp1 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp2 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp3 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    PCKEV_AVG_ST8x4_UB(tmp0, dst0, tmp1, dst1, tmp2, dst2, tmp3, dst3,
+                       dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_8w_msa(const uint8_t *src,
+                                                  int32_t src_stride,
+                                                  uint8_t *dst,
+                                                  int32_t dst_stride,
+                                                  int8_t *filter_horiz,
+                                                  int8_t *filter_vert,
+                                                  int32_t height) {
+  if (4 == height) {
+    common_hv_2ht_2vt_and_aver_dst_8x4_msa(src, src_stride, dst, dst_stride,
+                                           filter_horiz, filter_vert);
+  } else {
+    common_hv_2ht_2vt_and_aver_dst_8x8mult_msa(src, src_stride, dst, dst_stride,
+                                               filter_horiz, filter_vert,
+                                               height);
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_16w_msa(const uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *dst,
+                                                   int32_t dst_stride,
+                                                   int8_t *filter_horiz,
+                                                   int8_t *filter_vert,
+                                                   int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt_hz, filt_vt, vec0, vec1, dst0, dst1, dst2, dst3;
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, tmp0, tmp1;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB2(src, 8, src0, src1);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst0, dst);
+    dst += dst_stride;
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst1, dst);
+    dst += dst_stride;
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst2, dst);
+    dst += dst_stride;
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst3, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_32w_msa(const uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *dst,
+                                                   int32_t dst_stride,
+                                                   int8_t *filter_horiz,
+                                                   int8_t *filter_vert,
+                                                   int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 2; multiple8_cnt--;) {
+    common_hv_2ht_2vt_and_aver_dst_16w_msa(src, src_stride, dst, dst_stride,
+                                           filter_horiz, filter_vert, height);
+    src += 16;
+    dst += 16;
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_64w_msa(const uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *dst,
+                                                   int32_t dst_stride,
+                                                   int8_t *filter_horiz,
+                                                   int8_t *filter_vert,
+                                                   int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 4; multiple8_cnt--;) {
+    common_hv_2ht_2vt_and_aver_dst_16w_msa(src, src_stride, dst, dst_stride,
+                                           filter_horiz, filter_vert, height);
+    src += 16;
+    dst += 16;
+  }
+}
+
+void vpx_convolve8_avg_msa(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4,
+                           int w, int h) {
+  int8_t cnt, filt_hor[8], filt_ver[8];
+
+  assert(x_step_q4 == 16);
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  for (cnt = 0; cnt < 8; ++cnt) {
+    filt_hor[cnt] = filter_x[cnt];
+    filt_ver[cnt] = filter_y[cnt];
+  }
+
+  if (((const int32_t *)filter_x)[0] == 0 &&
+      ((const int32_t *)filter_y)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_hv_2ht_2vt_and_aver_dst_4w_msa(src, (int32_t)src_stride,
+                                              dst, (int32_t)dst_stride,
+                                              &filt_hor[3], &filt_ver[3], h);
+        break;
+      case 8:
+        common_hv_2ht_2vt_and_aver_dst_8w_msa(src, (int32_t)src_stride,
+                                              dst, (int32_t)dst_stride,
+                                              &filt_hor[3], &filt_ver[3], h);
+        break;
+      case 16:
+        common_hv_2ht_2vt_and_aver_dst_16w_msa(src, (int32_t)src_stride,
+                                               dst, (int32_t)dst_stride,
+                                               &filt_hor[3], &filt_ver[3], h);
+        break;
+      case 32:
+        common_hv_2ht_2vt_and_aver_dst_32w_msa(src, (int32_t)src_stride,
+                                               dst, (int32_t)dst_stride,
+                                               &filt_hor[3], &filt_ver[3], h);
+        break;
+      case 64:
+        common_hv_2ht_2vt_and_aver_dst_64w_msa(src, (int32_t)src_stride,
+                                               dst, (int32_t)dst_stride,
+                                               &filt_hor[3], &filt_ver[3], h);
+        break;
+      default:
+        vpx_convolve8_avg_c(src, src_stride, dst, dst_stride,
+                            filter_x, x_step_q4, filter_y, y_step_q4,
+                            w, h);
+        break;
+    }
+  } else if (((const int32_t *)filter_x)[0] == 0 ||
+             ((const int32_t *)filter_y)[0] == 0) {
+    vpx_convolve8_avg_c(src, src_stride, dst, dst_stride,
+                        filter_x, x_step_q4, filter_y, y_step_q4,
+                        w, h);
+  } else {
+    switch (w) {
+      case 4:
+        common_hv_8ht_8vt_and_aver_dst_4w_msa(src, (int32_t)src_stride,
+                                              dst, (int32_t)dst_stride,
+                                              filt_hor, filt_ver, h);
+        break;
+      case 8:
+        common_hv_8ht_8vt_and_aver_dst_8w_msa(src, (int32_t)src_stride,
+                                              dst, (int32_t)dst_stride,
+                                              filt_hor, filt_ver, h);
+        break;
+      case 16:
+        common_hv_8ht_8vt_and_aver_dst_16w_msa(src, (int32_t)src_stride,
+                                               dst, (int32_t)dst_stride,
+                                               filt_hor, filt_ver, h);
+        break;
+      case 32:
+        common_hv_8ht_8vt_and_aver_dst_32w_msa(src, (int32_t)src_stride,
+                                               dst, (int32_t)dst_stride,
+                                               filt_hor, filt_ver, h);
+        break;
+      case 64:
+        common_hv_8ht_8vt_and_aver_dst_64w_msa(src, (int32_t)src_stride,
+                                               dst, (int32_t)dst_stride,
+                                               filt_hor, filt_ver, h);
+        break;
+      default:
+        vpx_convolve8_avg_c(src, src_stride, dst, dst_stride,
+                            filter_x, x_step_q4, filter_y, y_step_q4,
+                            w, h);
+        break;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_avg_vert_msa.c b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_avg_vert_msa.c
new file mode 100644
index 0000000..0164e41
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_avg_vert_msa.c
@@ -0,0 +1,718 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/vpx_convolve_msa.h"
+
+static void common_vt_8t_and_aver_dst_4w_msa(const uint8_t *src,
+                                             int32_t src_stride,
+                                             uint8_t *dst,
+                                             int32_t dst_stride,
+                                             int8_t *filter,
+                                             int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16u8 dst0, dst1, dst2, dst3, out;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, src2110, src4332, src6554, src8776;
+  v16i8 src10998, filt0, filt1, filt2, filt3;
+  v8i16 filt, out10, out32;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+             src54_r, src21_r);
+  ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+  ILVR_D3_SB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r, src2110,
+             src4332, src6554);
+  XORI_B3_128_SB(src2110, src4332, src6554);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+               src87_r, src98_r, src109_r);
+    ILVR_D2_SB(src87_r, src76_r, src109_r, src98_r, src8776, src10998);
+    XORI_B2_128_SB(src8776, src10998);
+    out10 = FILT_8TAP_DPADD_S_H(src2110, src4332, src6554, src8776, filt0,
+                                filt1, filt2, filt3);
+    out32 = FILT_8TAP_DPADD_S_H(src4332, src6554, src8776, src10998, filt0,
+                                filt1, filt2, filt3);
+    SRARI_H2_SH(out10, out32, FILTER_BITS);
+    SAT_SH2_SH(out10, out32, 7);
+    out = PCKEV_XORI128_UB(out10, out32);
+    ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
+
+    dst0 = (v16u8)__msa_ilvr_d((v2i64)dst2, (v2i64)dst0);
+    out = __msa_aver_u_b(out, dst0);
+
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src2110 = src6554;
+    src4332 = src8776;
+    src6554 = src10998;
+    src6 = src10;
+  }
+}
+
+static void common_vt_8t_and_aver_dst_8w_msa(const uint8_t *src,
+                                             int32_t src_stride,
+                                             uint8_t *dst,
+                                             int32_t dst_stride,
+                                             int8_t *filter,
+                                             int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16u8 dst0, dst1, dst2, dst3;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, filt0, filt1, filt2, filt3;
+  v8i16 filt, out0, out1, out2, out3;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+             src54_r, src21_r);
+  ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+               src87_r, src98_r, src109_r);
+    out0 = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0,
+                               filt1, filt2, filt3);
+    out1 = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0,
+                               filt1, filt2, filt3);
+    out2 = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0,
+                               filt1, filt2, filt3);
+    out3 = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0,
+                               filt1, filt2, filt3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    CONVERT_UB_AVG_ST8x4_UB(out0, out1, out2, out3, dst0, dst1, dst2, dst3,
+                            dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src10_r = src54_r;
+    src32_r = src76_r;
+    src54_r = src98_r;
+    src21_r = src65_r;
+    src43_r = src87_r;
+    src65_r = src109_r;
+    src6 = src10;
+  }
+}
+
+static void common_vt_8t_and_aver_dst_16w_mult_msa(const uint8_t *src,
+                                                   int32_t src_stride,
+                                                   uint8_t *dst,
+                                                   int32_t dst_stride,
+                                                   int8_t *filter,
+                                                   int32_t height,
+                                                   int32_t width) {
+  const uint8_t *src_tmp;
+  uint8_t *dst_tmp;
+  uint32_t loop_cnt, cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, src10_l, src32_l, src54_l, src76_l;
+  v16i8 src98_l, src21_l, src43_l, src65_l, src87_l, src109_l;
+  v16i8 filt0, filt1, filt2, filt3;
+  v16u8 dst0, dst1, dst2, dst3, tmp0, tmp1, tmp2, tmp3;
+  v8i16 out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l, filt;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  for (cnt = (width >> 4); cnt--;) {
+    src_tmp = src;
+    dst_tmp = dst;
+
+    LD_SB7(src_tmp, src_stride, src0, src1, src2, src3, src4, src5, src6);
+    XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+    src_tmp += (7 * src_stride);
+
+    ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+               src54_r, src21_r);
+    ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+    ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l, src32_l,
+               src54_l, src21_l);
+    ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;) {
+      LD_SB4(src_tmp, src_stride, src7, src8, src9, src10);
+      src_tmp += (4 * src_stride);
+
+      LD_UB4(dst_tmp, dst_stride, dst0, dst1, dst2, dst3);
+      XORI_B4_128_SB(src7, src8, src9, src10);
+      ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+                 src87_r, src98_r, src109_r);
+      ILVL_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_l,
+                 src87_l, src98_l, src109_l);
+      out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0,
+                                   filt1, filt2, filt3);
+      out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0,
+                                   filt1, filt2, filt3);
+      out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0,
+                                   filt1, filt2, filt3);
+      out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0,
+                                   filt1, filt2, filt3);
+      out0_l = FILT_8TAP_DPADD_S_H(src10_l, src32_l, src54_l, src76_l, filt0,
+                                   filt1, filt2, filt3);
+      out1_l = FILT_8TAP_DPADD_S_H(src21_l, src43_l, src65_l, src87_l, filt0,
+                                   filt1, filt2, filt3);
+      out2_l = FILT_8TAP_DPADD_S_H(src32_l, src54_l, src76_l, src98_l, filt0,
+                                   filt1, filt2, filt3);
+      out3_l = FILT_8TAP_DPADD_S_H(src43_l, src65_l, src87_l, src109_l, filt0,
+                                   filt1, filt2, filt3);
+      SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS);
+      SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, FILTER_BITS);
+      SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+      SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7);
+      PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l,
+                  out3_r, tmp0, tmp1, tmp2, tmp3);
+      XORI_B4_128_UB(tmp0, tmp1, tmp2, tmp3);
+      AVER_UB4_UB(tmp0, dst0, tmp1, dst1, tmp2, dst2, tmp3, dst3, dst0, dst1,
+                  dst2, dst3);
+      ST_UB4(dst0, dst1, dst2, dst3, dst_tmp, dst_stride);
+      dst_tmp += (4 * dst_stride);
+
+      src10_r = src54_r;
+      src32_r = src76_r;
+      src54_r = src98_r;
+      src21_r = src65_r;
+      src43_r = src87_r;
+      src65_r = src109_r;
+      src10_l = src54_l;
+      src32_l = src76_l;
+      src54_l = src98_l;
+      src21_l = src65_l;
+      src43_l = src87_l;
+      src65_l = src109_l;
+      src6 = src10;
+    }
+
+    src += 16;
+    dst += 16;
+  }
+}
+
+static void common_vt_8t_and_aver_dst_16w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  common_vt_8t_and_aver_dst_16w_mult_msa(src, src_stride, dst, dst_stride,
+                                         filter, height, 16);
+}
+
+static void common_vt_8t_and_aver_dst_32w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  common_vt_8t_and_aver_dst_16w_mult_msa(src, src_stride, dst, dst_stride,
+                                         filter, height, 32);
+}
+
+static void common_vt_8t_and_aver_dst_64w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  common_vt_8t_and_aver_dst_16w_mult_msa(src, src_stride, dst, dst_stride,
+                                         filter, height, 64);
+}
+
+static void common_vt_2t_and_aver_dst_4x4_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, src4;
+  v16u8 dst0, dst1, dst2, dst3, out, filt0, src2110, src4332;
+  v16i8 src10_r, src32_r, src21_r, src43_r;
+  v8i16 filt;
+  v8u16 tmp0, tmp1;
+
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  src += (4 * src_stride);
+
+  src4 = LD_SB(src);
+  src += src_stride;
+
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst1);
+  dst0 = (v16u8)__msa_ilvr_d((v2i64)dst1, (v2i64)dst0);
+  ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+             src32_r, src43_r);
+  ILVR_D2_UB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+  DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1);
+  SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+
+  out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+  out = __msa_aver_u_b(out, dst0);
+
+  ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_vt_2t_and_aver_dst_4x8_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src87_r;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r;
+  v16u8 src2110, src4332, src6554, src8776, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  src += (8 * src_stride);
+  src8 = LD_SB(src);
+
+  LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+  ILVR_W4_UB(dst1, dst0, dst3, dst2, dst5, dst4, dst7, dst6, dst0, dst1,
+             dst2, dst3);
+  ILVR_D2_UB(dst1, dst0, dst3, dst2, dst0, dst1);
+  ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+             src32_r, src43_r);
+  ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r, src65_r,
+             src76_r, src87_r);
+  ILVR_D4_UB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r,
+             src87_r, src76_r, src2110, src4332, src6554, src8776);
+  DOTP_UB4_UH(src2110, src4332, src6554, src8776, filt0, filt0, filt0, filt0,
+              tmp0, tmp1, tmp2, tmp3);
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, src2110, src4332);
+  AVER_UB2_UB(src2110, dst0, src4332, dst1, src2110, src4332);
+  ST4x4_UB(src2110, src2110, 0, 1, 2, 3, dst, dst_stride);
+  dst += (4 * dst_stride);
+  ST4x4_UB(src4332, src4332, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_vt_2t_and_aver_dst_4w_msa(const uint8_t *src,
+                                             int32_t src_stride,
+                                             uint8_t *dst,
+                                             int32_t dst_stride,
+                                             int8_t *filter,
+                                             int32_t height) {
+  if (4 == height) {
+    common_vt_2t_and_aver_dst_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_vt_2t_and_aver_dst_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_vt_2t_and_aver_dst_8x4_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_UB5(src, src_stride, src0, src1, src2, src3, src4);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  ILVR_B2_UB(src1, src0, src2, src1, vec0, vec1);
+  ILVR_B2_UB(src3, src2, src4, src3, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1,
+              tmp2, tmp3);
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_AVG_ST8x4_UB(tmp0, dst0, tmp1, dst1, tmp2, dst2, tmp3, dst3,
+                     dst, dst_stride);
+}
+
+static void common_vt_2t_and_aver_dst_8x8mult_msa(const uint8_t *src,
+                                                  int32_t src_stride,
+                                                  uint8_t *dst,
+                                                  int32_t dst_stride,
+                                                  int8_t *filter,
+                                                  int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+  v16u8 dst1, dst2, dst3, dst4, dst5, dst6, dst7, dst8;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 3); loop_cnt--;) {
+    LD_UB8(src, src_stride, src1, src2, src3, src4, src5, src6, src7, src8);
+    src += (8 * src_stride);
+    LD_UB8(dst, dst_stride, dst1, dst2, dst3, dst4, dst5, dst6, dst7, dst8);
+
+    ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, vec0, vec1,
+               vec2, vec3);
+    ILVR_B4_UB(src5, src4, src6, src5, src7, src6, src8, src7, vec4, vec5,
+               vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1,
+                tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST8x4_UB(tmp0, dst1, tmp1, dst2, tmp2, dst3, tmp3, dst4,
+                       dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, tmp0, tmp1,
+                tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST8x4_UB(tmp0, dst5, tmp1, dst6, tmp2, dst7, tmp3, dst8,
+                       dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src0 = src8;
+  }
+}
+
+static void common_vt_2t_and_aver_dst_8w_msa(const uint8_t *src,
+                                             int32_t src_stride,
+                                             uint8_t *dst,
+                                             int32_t dst_stride,
+                                             int8_t *filter,
+                                             int32_t height) {
+  if (4 == height) {
+    common_vt_2t_and_aver_dst_8x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else {
+    common_vt_2t_and_aver_dst_8x8mult_msa(src, src_stride, dst, dst_stride,
+                                          filter, height);
+  }
+}
+
+static void common_vt_2t_and_aver_dst_16w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, dst0, dst1, dst2, dst3, filt0;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 tmp0, tmp1, tmp2, tmp3, filt;
+
+  /* rearranging filter_y */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst0, dst);
+    dst += dst_stride;
+
+    ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst1, dst);
+    dst += dst_stride;
+
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst2, dst);
+    dst += dst_stride;
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst3, dst);
+    dst += dst_stride;
+
+    src0 = src4;
+  }
+}
+
+static void common_vt_2t_and_aver_dst_32w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3, filt;
+
+  /* rearranging filter_y */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_UB2(src, 16, src0, src5);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+
+    LD_UB4(src + 16, src_stride, src6, src7, src8, src9);
+    LD_UB4(dst + 16, dst_stride, dst4, dst5, dst6, dst7);
+    src += (4 * src_stride);
+
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst0, dst);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst1, dst + dst_stride);
+
+    ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst2, dst + 2 * dst_stride);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst3, dst + 3 * dst_stride);
+
+    ILVR_B2_UB(src6, src5, src7, src6, vec0, vec2);
+    ILVL_B2_UB(src6, src5, src7, src6, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst4, dst + 16);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst5, dst + 16 + dst_stride);
+
+    ILVR_B2_UB(src8, src7, src9, src8, vec4, vec6);
+    ILVL_B2_UB(src8, src7, src9, src8, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst6, dst + 16 + 2 * dst_stride);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst7, dst + 16 + 3 * dst_stride);
+    dst += (4 * dst_stride);
+
+    src0 = src4;
+    src5 = src9;
+  }
+}
+
+static void common_vt_2t_and_aver_dst_64w_msa(const uint8_t *src,
+                                              int32_t src_stride,
+                                              uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter,
+                                              int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5;
+  v16u8 src6, src7, src8, src9, src10, src11, filt0;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8u16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_UB4(src, 16, src0, src3, src6, src9);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 1); loop_cnt--;) {
+    LD_UB2(src, src_stride, src1, src2);
+    LD_UB2(dst, dst_stride, dst0, dst1);
+    LD_UB2(src + 16, src_stride, src4, src5);
+    LD_UB2(dst + 16, dst_stride, dst2, dst3);
+    LD_UB2(src + 32, src_stride, src7, src8);
+    LD_UB2(dst + 32, dst_stride, dst4, dst5);
+    LD_UB2(src + 48, src_stride, src10, src11);
+    LD_UB2(dst + 48, dst_stride, dst6, dst7);
+    src += (2 * src_stride);
+
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst0, dst);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst1, dst + dst_stride);
+
+    ILVR_B2_UB(src4, src3, src5, src4, vec4, vec6);
+    ILVL_B2_UB(src4, src3, src5, src4, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp4, tmp5);
+    SRARI_H2_UH(tmp4, tmp5, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp5, tmp4, dst2, dst + 16);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp6, tmp7);
+    SRARI_H2_UH(tmp6, tmp7, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp7, tmp6, dst3, dst + 16 + dst_stride);
+
+    ILVR_B2_UB(src7, src6, src8, src7, vec0, vec2);
+    ILVL_B2_UB(src7, src6, src8, src7, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst4, dst + 32);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst5, dst + 32 + dst_stride);
+
+    ILVR_B2_UB(src10, src9, src11, src10, vec4, vec6);
+    ILVL_B2_UB(src10, src9, src11, src10, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp4, tmp5);
+    SRARI_H2_UH(tmp4, tmp5, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp5, tmp4, dst6, (dst + 48));
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp6, tmp7);
+    SRARI_H2_UH(tmp6, tmp7, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp7, tmp6, dst7, dst + 48 + dst_stride);
+    dst += (2 * dst_stride);
+
+    src0 = src2;
+    src3 = src5;
+    src6 = src8;
+    src9 = src11;
+  }
+}
+
+void vpx_convolve8_avg_vert_msa(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int x_step_q4,
+                                const int16_t *filter_y, int y_step_q4,
+                                int w, int h) {
+  int8_t cnt, filt_ver[8];
+
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  for (cnt = 0; cnt < 8; ++cnt) {
+    filt_ver[cnt] = filter_y[cnt];
+  }
+
+  if (((const int32_t *)filter_y)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_vt_2t_and_aver_dst_4w_msa(src, (int32_t)src_stride,
+                                         dst, (int32_t)dst_stride,
+                                         &filt_ver[3], h);
+        break;
+      case 8:
+        common_vt_2t_and_aver_dst_8w_msa(src, (int32_t)src_stride,
+                                         dst, (int32_t)dst_stride,
+                                         &filt_ver[3], h);
+        break;
+      case 16:
+        common_vt_2t_and_aver_dst_16w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          &filt_ver[3], h);
+        break;
+      case 32:
+        common_vt_2t_and_aver_dst_32w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          &filt_ver[3], h);
+        break;
+      case 64:
+        common_vt_2t_and_aver_dst_64w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          &filt_ver[3], h);
+        break;
+      default:
+        vpx_convolve8_avg_vert_c(src, src_stride, dst, dst_stride,
+                                 filter_x, x_step_q4, filter_y, y_step_q4,
+                                 w, h);
+        break;
+    }
+  } else {
+    switch (w) {
+      case 4:
+        common_vt_8t_and_aver_dst_4w_msa(src, (int32_t)src_stride,
+                                         dst, (int32_t)dst_stride,
+                                         filt_ver, h);
+        break;
+      case 8:
+        common_vt_8t_and_aver_dst_8w_msa(src, (int32_t)src_stride,
+                                         dst, (int32_t)dst_stride,
+                                         filt_ver, h);
+        break;
+      case 16:
+        common_vt_8t_and_aver_dst_16w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          filt_ver, h);
+
+        break;
+      case 32:
+        common_vt_8t_and_aver_dst_32w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          filt_ver, h);
+        break;
+      case 64:
+        common_vt_8t_and_aver_dst_64w_msa(src, (int32_t)src_stride,
+                                          dst, (int32_t)dst_stride,
+                                          filt_ver, h);
+        break;
+      default:
+        vpx_convolve8_avg_vert_c(src, src_stride, dst, dst_stride,
+                                 filter_x, x_step_q4, filter_y, y_step_q4,
+                                 w, h);
+        break;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_horiz_msa.c b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_horiz_msa.c
new file mode 100644
index 0000000..dbd120b
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_horiz_msa.c
@@ -0,0 +1,703 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/vpx_convolve_msa.h"
+
+static void common_hz_8t_4x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v8i16 filt, out0, out1;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, out0, out1);
+  SRARI_H2_SH(out0, out1, FILTER_BITS);
+  SAT_SH2_SH(out0, out1, 7);
+  out = PCKEV_XORI128_UB(out0, out1);
+  ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_hz_8t_4x8_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16i8 filt0, filt1, filt2, filt3;
+  v16i8 src0, src1, src2, src3;
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  src += (4 * src_stride);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, out0, out1);
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, out2, out3);
+  SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+  SAT_SH4_SH(out0, out1, out2, out3, 7);
+  out = PCKEV_XORI128_UB(out0, out1);
+  ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+  dst += (4 * dst_stride);
+  out = PCKEV_XORI128_UB(out2, out3);
+  ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_hz_8t_4w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_hz_8t_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_hz_8t_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_hz_8t_8x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, tmp0, tmp1;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, out0, out1, out2,
+                             out3);
+  SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+  SAT_SH4_SH(out0, out1, out2, out3, 7);
+  tmp0 = PCKEV_XORI128_UB(out0, out1);
+  tmp1 = PCKEV_XORI128_UB(out2, out3);
+  ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+}
+
+static void common_hz_8t_8x8mult_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, tmp0, tmp1;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    src += (4 * src_stride);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    tmp0 = PCKEV_XORI128_UB(out0, out1);
+    tmp1 = PCKEV_XORI128_UB(out2, out3);
+    ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void common_hz_8t_8w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_hz_8t_8x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else {
+    common_hz_8t_8x8mult_msa(src, src_stride, dst, dst_stride, filter, height);
+  }
+}
+
+static void common_hz_8t_16w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = (height >> 1); loop_cnt--;) {
+    LD_SB2(src, src_stride, src0, src2);
+    LD_SB2(src + 8, src_stride, src1, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    src += (2 * src_stride);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST_UB(out, dst);
+    dst += dst_stride;
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST_UB(out, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_8t_32w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = (height >> 1); loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST_UB(out, dst);
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST_UB(out, dst + 16);
+    dst += dst_stride;
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST_UB(out, dst);
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST_UB(out, dst + 16);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_8t_64w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  int32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = height; loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST_UB(out, dst);
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST_UB(out, dst + 16);
+
+    src0 = LD_SB(src + 32);
+    src2 = LD_SB(src + 48);
+    src3 = LD_SB(src + 56);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST_UB(out, dst + 32);
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST_UB(out, dst + 48);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_4x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16i8 src0, src1, src2, src3, mask;
+  v16u8 filt0, vec0, vec1, res0, res1;
+  v8u16 vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8) __msa_splati_h((v8i16) filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
+  DOTP_UB2_UH(vec0, vec1, filt0, filt0, vec2, vec3);
+  SRARI_H2_UH(vec2, vec3, FILTER_BITS);
+  PCKEV_B2_UB(vec2, vec2, vec3, vec3, res0, res1);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_2t_4x8_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16u8 vec0, vec1, vec2, vec3, filt0;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16i8 res0, res1, res2, res3;
+  v8u16 vec4, vec5, vec6, vec7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8) __msa_splati_h((v8i16) filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
+  VSHF_B2_UB(src4, src5, src6, src7, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec4, vec5,
+              vec6, vec7);
+  SRARI_H4_UH(vec4, vec5, vec6, vec7, FILTER_BITS);
+  PCKEV_B4_SB(vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7, res0, res1,
+              res2, res3);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+  dst += (4 * dst_stride);
+  ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_2t_4w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_hz_2t_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_hz_2t_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_hz_2t_8x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16u8 filt0;
+  v16i8 src0, src1, src2, src3, mask;
+  v8u16 vec0, vec1, vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8) __msa_splati_h((v8i16) filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  PCKEV_B2_SB(vec1, vec0, vec3, vec2, src0, src1);
+  ST8x4_UB(src0, src1, dst, dst_stride);
+}
+
+static void common_hz_2t_8x8mult_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter, int32_t height) {
+  v16u8 filt0;
+  v16i8 src0, src1, src2, src3, mask, out0, out1;
+  v8u16 vec0, vec1, vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8) __msa_splati_h((v8i16) filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  src += (4 * src_stride);
+
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  src += (4 * src_stride);
+
+  PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+  ST8x4_UB(out0, out1, dst, dst_stride);
+  dst += (4 * dst_stride);
+
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+  ST8x4_UB(out0, out1, dst, dst_stride);
+  dst += (4 * dst_stride);
+
+  if (16 == height) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+
+    PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+    ST8x4_UB(out0, out1, dst + 4 * dst_stride, dst_stride);
+  }
+}
+
+static void common_hz_2t_8w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_hz_2t_8x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else {
+    common_hz_2t_8x8mult_msa(src, src_stride, dst, dst_stride, filter, height);
+  }
+}
+
+static void common_hz_2t_16w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  loop_cnt = (height >> 2) - 1;
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8) __msa_splati_h((v8i16) filt, 0);
+
+  LD_SB4(src, src_stride, src0, src2, src4, src6);
+  LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+  src += (4 * src_stride);
+
+  VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+  VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+              out2, out3);
+  DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+              out6, out7);
+  SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+  SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+  PCKEV_ST_SB(out0, out1, dst);
+  dst += dst_stride;
+  PCKEV_ST_SB(out2, out3, dst);
+  dst += dst_stride;
+  PCKEV_ST_SB(out4, out5, dst);
+  dst += dst_stride;
+  PCKEV_ST_SB(out6, out7, dst);
+  dst += dst_stride;
+
+  for (; loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+                out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+                out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    PCKEV_ST_SB(out0, out1, dst);
+    dst += dst_stride;
+    PCKEV_ST_SB(out2, out3, dst);
+    dst += dst_stride;
+    PCKEV_ST_SB(out4, out5, dst);
+    dst += dst_stride;
+    PCKEV_ST_SB(out6, out7, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_32w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8) __msa_splati_h((v8i16) filt, 0);
+
+  for (loop_cnt = height >> 1; loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+    src4 = LD_SB(src);
+    src6 = LD_SB(src + 16);
+    src7 = LD_SB(src + 24);
+    src5 = __msa_sldi_b(src6, src4, 8);
+    src += src_stride;
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+                out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+                out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    PCKEV_ST_SB(out0, out1, dst);
+    PCKEV_ST_SB(out2, out3, dst + 16);
+    dst += dst_stride;
+    PCKEV_ST_SB(out4, out5, dst);
+    PCKEV_ST_SB(out6, out7, dst + 16);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_64w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8) __msa_splati_h((v8i16) filt, 0);
+
+  for (loop_cnt = height; loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src4 = LD_SB(src + 32);
+    src6 = LD_SB(src + 48);
+    src7 = LD_SB(src + 56);
+    SLDI_B3_SB(src2, src4, src6, src0, src2, src4, src1, src3, src5, 8);
+    src += src_stride;
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+                out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+                out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    PCKEV_ST_SB(out0, out1, dst);
+    PCKEV_ST_SB(out2, out3, dst + 16);
+    PCKEV_ST_SB(out4, out5, dst + 32);
+    PCKEV_ST_SB(out6, out7, dst + 48);
+    dst += dst_stride;
+  }
+}
+
+void vpx_convolve8_horiz_msa(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int x_step_q4,
+                             const int16_t *filter_y, int y_step_q4,
+                             int w, int h) {
+  int8_t cnt, filt_hor[8];
+
+  assert(x_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+
+  for (cnt = 0; cnt < 8; ++cnt) {
+    filt_hor[cnt] = filter_x[cnt];
+  }
+
+  if (((const int32_t *)filter_x)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_hz_2t_4w_msa(src, (int32_t)src_stride,
+                            dst, (int32_t)dst_stride,
+                            &filt_hor[3], h);
+        break;
+      case 8:
+        common_hz_2t_8w_msa(src, (int32_t)src_stride,
+                            dst, (int32_t)dst_stride,
+                            &filt_hor[3], h);
+        break;
+      case 16:
+        common_hz_2t_16w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             &filt_hor[3], h);
+        break;
+      case 32:
+        common_hz_2t_32w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             &filt_hor[3], h);
+        break;
+      case 64:
+        common_hz_2t_64w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             &filt_hor[3], h);
+        break;
+      default:
+        vpx_convolve8_horiz_c(src, src_stride, dst, dst_stride,
+                              filter_x, x_step_q4, filter_y, y_step_q4,
+                              w, h);
+        break;
+    }
+  } else {
+    switch (w) {
+      case 4:
+        common_hz_8t_4w_msa(src, (int32_t)src_stride,
+                            dst, (int32_t)dst_stride,
+                            filt_hor, h);
+        break;
+      case 8:
+        common_hz_8t_8w_msa(src, (int32_t)src_stride,
+                            dst, (int32_t)dst_stride,
+                            filt_hor, h);
+        break;
+      case 16:
+        common_hz_8t_16w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             filt_hor, h);
+        break;
+      case 32:
+        common_hz_8t_32w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             filt_hor, h);
+        break;
+      case 64:
+        common_hz_8t_64w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             filt_hor, h);
+        break;
+      default:
+        vpx_convolve8_horiz_c(src, src_stride, dst, dst_stride,
+                              filter_x, x_step_q4, filter_y, y_step_q4,
+                              w, h);
+        break;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_msa.c b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_msa.c
new file mode 100644
index 0000000..7546f13
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_msa.c
@@ -0,0 +1,635 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/vpx_convolve_msa.h"
+
+const uint8_t mc_filt_mask_arr[16 * 3] = {
+  /* 8 width cases */
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+  /* 4 width cases */
+  0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20,
+  /* 4 width cases */
+  8, 9, 9, 10, 10, 11, 11, 12, 24, 25, 25, 26, 26, 27, 27, 28
+};
+
+static void common_hv_8ht_8vt_4w_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter_horiz, int8_t *filter_vert,
+                                     int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3;
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8i16 hz_out7, hz_out8, hz_out9, tmp0, tmp1, out0, out1, out2, out3, out4;
+  v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= (3 + 3 * src_stride);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  hz_out0 = HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out2 = HORIZ_8TAP_FILT(src2, src3, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out4 = HORIZ_8TAP_FILT(src4, src5, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out5 = HORIZ_8TAP_FILT(src5, src6, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  SLDI_B2_SH(hz_out2, hz_out4, hz_out0, hz_out2, hz_out1, hz_out3, 8);
+
+  filt = LD_SH(filter_vert);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
+
+  ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
+  out2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    hz_out7 = HORIZ_8TAP_FILT(src7, src8, mask0, mask1, mask2, mask3,
+                              filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    hz_out6 = (v8i16)__msa_sldi_b((v16i8)hz_out7, (v16i8)hz_out5, 8);
+    out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+    tmp0 = FILT_8TAP_DPADD_S_H(out0, out1, out2, out3, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out9 = HORIZ_8TAP_FILT(src9, src10, mask0, mask1, mask2, mask3,
+                              filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    hz_out8 = (v8i16)__msa_sldi_b((v16i8)hz_out9, (v16i8)hz_out7, 8);
+    out4 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8);
+    tmp1 = FILT_8TAP_DPADD_S_H(out1, out2, out3, out4, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+    SRARI_H2_SH(tmp0, tmp1, FILTER_BITS);
+    SAT_SH2_SH(tmp0, tmp1, 7);
+    out = PCKEV_XORI128_UB(tmp0, tmp1);
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    hz_out5 = hz_out9;
+    out0 = out2;
+    out1 = out3;
+    out2 = out4;
+  }
+}
+
+static void common_hv_8ht_8vt_8w_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter_horiz, int8_t *filter_vert,
+                                     int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3;
+  v16u8 mask0, mask1, mask2, mask3, vec0, vec1;
+  v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3;
+  v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8i16 hz_out7, hz_out8, hz_out9, hz_out10, tmp0, tmp1, tmp2, tmp3;
+  v8i16 out0, out1, out2, out3, out4, out5, out6, out7, out8, out9;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= (3 + 3 * src_stride);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  hz_out0 = HORIZ_8TAP_FILT(src0, src0, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out1 = HORIZ_8TAP_FILT(src1, src1, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out2 = HORIZ_8TAP_FILT(src2, src2, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out3 = HORIZ_8TAP_FILT(src3, src3, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out4 = HORIZ_8TAP_FILT(src4, src4, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out5 = HORIZ_8TAP_FILT(src5, src5, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out6 = HORIZ_8TAP_FILT(src6, src6, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+
+  filt = LD_SH(filter_vert);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
+
+  ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
+  ILVEV_B2_SH(hz_out4, hz_out5, hz_out1, hz_out2, out2, out4);
+  ILVEV_B2_SH(hz_out3, hz_out4, hz_out5, hz_out6, out5, out6);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    XORI_B4_128_SB(src7, src8, src9, src10);
+
+    hz_out7 = HORIZ_8TAP_FILT(src7, src7, mask0, mask1, mask2, mask3,
+                              filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+    tmp0 = FILT_8TAP_DPADD_S_H(out0, out1, out2, out3, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out8 = HORIZ_8TAP_FILT(src8, src8, mask0, mask1, mask2, mask3,
+                              filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    out7 = (v8i16)__msa_ilvev_b((v16i8)hz_out8, (v16i8)hz_out7);
+    tmp1 = FILT_8TAP_DPADD_S_H(out4, out5, out6, out7, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out9 = HORIZ_8TAP_FILT(src9, src9, mask0, mask1, mask2, mask3,
+                              filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    out8 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8);
+    tmp2 = FILT_8TAP_DPADD_S_H(out1, out2, out3, out8, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out10 = HORIZ_8TAP_FILT(src10, src10, mask0, mask1, mask2, mask3,
+                               filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    out9 = (v8i16)__msa_ilvev_b((v16i8)hz_out10, (v16i8)hz_out9);
+    tmp3 = FILT_8TAP_DPADD_S_H(out5, out6, out7, out9, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+    SRARI_H4_SH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    SAT_SH4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+    vec0 = PCKEV_XORI128_UB(tmp0, tmp1);
+    vec1 = PCKEV_XORI128_UB(tmp2, tmp3);
+    ST8x4_UB(vec0, vec1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    hz_out6 = hz_out10;
+    out0 = out2;
+    out1 = out3;
+    out2 = out8;
+    out4 = out6;
+    out5 = out7;
+    out6 = out9;
+  }
+}
+
+static void common_hv_8ht_8vt_16w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 2; multiple8_cnt--;) {
+    common_hv_8ht_8vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                             filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_8ht_8vt_32w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 4; multiple8_cnt--;) {
+    common_hv_8ht_8vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                             filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_8ht_8vt_64w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 8; multiple8_cnt--;) {
+    common_hv_8ht_8vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                             filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_2ht_2vt_4x4_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz,
+                                      int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, mask;
+  v16u8 filt_vt, filt_hz, vec0, vec1, res0, res1;
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, filt, tmp0, tmp1;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  filt = LD_UH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+  hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+  hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+  hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2);
+
+  ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+  DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+  SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+  PCKEV_B2_UB(tmp0, tmp0, tmp1, tmp1, res0, res1);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_4x8_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz,
+                                      int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, mask;
+  v16i8 res0, res1, res2, res3;
+  v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3;
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8u16 hz_out7, hz_out8, vec4, vec5, vec6, vec7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  filt = LD_UH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  src += (8 * src_stride);
+  src8 = LD_SB(src);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+  hz_out4 = HORIZ_2TAP_FILT_UH(src4, src5, mask, filt_hz, FILTER_BITS);
+  hz_out6 = HORIZ_2TAP_FILT_UH(src6, src7, mask, filt_hz, FILTER_BITS);
+  hz_out8 = HORIZ_2TAP_FILT_UH(src8, src8, mask, filt_hz, FILTER_BITS);
+  SLDI_B3_UH(hz_out2, hz_out4, hz_out6, hz_out0, hz_out2, hz_out4, hz_out1,
+             hz_out3, hz_out5, 8);
+  hz_out7 = (v8u16)__msa_pckod_d((v2i64)hz_out8, (v2i64)hz_out6);
+
+  ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+  ILVEV_B2_UB(hz_out4, hz_out5, hz_out6, hz_out7, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt_vt, filt_vt, filt_vt, filt_vt,
+              vec4, vec5, vec6, vec7);
+  SRARI_H4_UH(vec4, vec5, vec6, vec7, FILTER_BITS);
+  PCKEV_B4_SB(vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7, res0, res1,
+              res2, res3);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+  dst += (4 * dst_stride);
+  ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_4w_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter_horiz, int8_t *filter_vert,
+                                     int32_t height) {
+  if (4 == height) {
+    common_hv_2ht_2vt_4x4_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                              filter_vert);
+  } else if (8 == height) {
+    common_hv_2ht_2vt_4x8_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                              filter_vert);
+  }
+}
+
+static void common_hv_2ht_2vt_8x4_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz,
+                                      int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, mask, out0, out1;
+  v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3;
+  v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+  vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+  tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+  vec1 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+  tmp1 = __msa_dotp_u_h(vec1, filt_vt);
+
+  hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+  vec2 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+  tmp2 = __msa_dotp_u_h(vec2, filt_vt);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+  vec3 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+  tmp3 = __msa_dotp_u_h(vec3, filt_vt);
+
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+  ST8x4_UB(out0, out1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_8x8mult_msa(const uint8_t *src,
+                                          int32_t src_stride,
+                                          uint8_t *dst,
+                                          int32_t dst_stride,
+                                          int8_t *filter_horiz,
+                                          int8_t *filter_vert,
+                                          int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, mask, out0, out1;
+  v16u8 filt_hz, filt_vt, vec0;
+  v8u16 hz_out0, hz_out1, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_SB(src);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 3); loop_cnt--;) {
+    LD_SB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp1 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp2 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H2_UH(tmp1, tmp2, FILTER_BITS);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp3 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    LD_SB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp4 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H2_UH(tmp3, tmp4, FILTER_BITS);
+    PCKEV_B2_SB(tmp2, tmp1, tmp4, tmp3, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp5 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp6 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp7 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp8 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H4_UH(tmp5, tmp6, tmp7, tmp8, FILTER_BITS);
+    PCKEV_B2_SB(tmp6, tmp5, tmp8, tmp7, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void common_hv_2ht_2vt_8w_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter_horiz, int8_t *filter_vert,
+                                     int32_t height) {
+  if (4 == height) {
+    common_hv_2ht_2vt_8x4_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                              filter_vert);
+  } else {
+    common_hv_2ht_2vt_8x8mult_msa(src, src_stride, dst, dst_stride,
+                                  filter_horiz, filter_vert, height);
+  }
+}
+
+static void common_hv_2ht_2vt_16w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt_hz, filt_vt, vec0, vec1;
+  v8u16 tmp1, tmp2, hz_out0, hz_out1, hz_out2, hz_out3;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB2(src, 8, src0, src1);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+    SRARI_H2_UH(tmp1, tmp2, FILTER_BITS);
+    PCKEV_ST_SB(tmp1, tmp2, dst);
+    dst += dst_stride;
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+    SRARI_H2_UH(tmp1, tmp2, FILTER_BITS);
+    PCKEV_ST_SB(tmp1, tmp2, dst);
+    dst += dst_stride;
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+    SRARI_H2_UH(tmp1, tmp2, FILTER_BITS);
+    PCKEV_ST_SB(tmp1, tmp2, dst);
+    dst += dst_stride;
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+    SRARI_H2_UH(tmp1, tmp2, FILTER_BITS);
+    PCKEV_ST_SB(tmp1, tmp2, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hv_2ht_2vt_32w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 2; multiple8_cnt--;) {
+    common_hv_2ht_2vt_16w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                              filter_vert, height);
+    src += 16;
+    dst += 16;
+  }
+}
+
+static void common_hv_2ht_2vt_64w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 4; multiple8_cnt--;) {
+    common_hv_2ht_2vt_16w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                              filter_vert, height);
+    src += 16;
+    dst += 16;
+  }
+}
+
+void vpx_convolve8_msa(const uint8_t *src, ptrdiff_t src_stride,
+                       uint8_t *dst, ptrdiff_t dst_stride,
+                       const int16_t *filter_x, int32_t x_step_q4,
+                       const int16_t *filter_y, int32_t y_step_q4,
+                       int32_t w, int32_t h) {
+  int8_t cnt, filt_hor[8], filt_ver[8];
+
+  assert(x_step_q4 == 16);
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  for (cnt = 0; cnt < 8; ++cnt) {
+    filt_hor[cnt] = filter_x[cnt];
+    filt_ver[cnt] = filter_y[cnt];
+  }
+
+  if (((const int32_t *)filter_x)[0] == 0 &&
+      ((const int32_t *)filter_y)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_hv_2ht_2vt_4w_msa(src, (int32_t)src_stride,
+                                 dst, (int32_t)dst_stride,
+                                 &filt_hor[3], &filt_ver[3], (int32_t)h);
+        break;
+      case 8:
+        common_hv_2ht_2vt_8w_msa(src, (int32_t)src_stride,
+                                 dst, (int32_t)dst_stride,
+                                 &filt_hor[3], &filt_ver[3], (int32_t)h);
+        break;
+      case 16:
+        common_hv_2ht_2vt_16w_msa(src, (int32_t)src_stride,
+                                  dst, (int32_t)dst_stride,
+                                  &filt_hor[3], &filt_ver[3], (int32_t)h);
+        break;
+      case 32:
+        common_hv_2ht_2vt_32w_msa(src, (int32_t)src_stride,
+                                  dst, (int32_t)dst_stride,
+                                  &filt_hor[3], &filt_ver[3], (int32_t)h);
+        break;
+      case 64:
+        common_hv_2ht_2vt_64w_msa(src, (int32_t)src_stride,
+                                  dst, (int32_t)dst_stride,
+                                  &filt_hor[3], &filt_ver[3], (int32_t)h);
+        break;
+      default:
+        vpx_convolve8_c(src, src_stride, dst, dst_stride,
+                        filter_x, x_step_q4, filter_y, y_step_q4,
+                        w, h);
+        break;
+    }
+  } else if (((const int32_t *)filter_x)[0] == 0 ||
+             ((const int32_t *)filter_y)[0] == 0) {
+    vpx_convolve8_c(src, src_stride, dst, dst_stride,
+                    filter_x, x_step_q4, filter_y, y_step_q4,
+                    w, h);
+  } else {
+    switch (w) {
+      case 4:
+        common_hv_8ht_8vt_4w_msa(src, (int32_t)src_stride,
+                                 dst, (int32_t)dst_stride,
+                                 filt_hor, filt_ver, (int32_t)h);
+        break;
+      case 8:
+        common_hv_8ht_8vt_8w_msa(src, (int32_t)src_stride,
+                                 dst, (int32_t)dst_stride,
+                                 filt_hor, filt_ver, (int32_t)h);
+        break;
+      case 16:
+        common_hv_8ht_8vt_16w_msa(src, (int32_t)src_stride,
+                                  dst, (int32_t)dst_stride,
+                                  filt_hor, filt_ver, (int32_t)h);
+        break;
+      case 32:
+        common_hv_8ht_8vt_32w_msa(src, (int32_t)src_stride,
+                                  dst, (int32_t)dst_stride,
+                                  filt_hor, filt_ver, (int32_t)h);
+        break;
+      case 64:
+        common_hv_8ht_8vt_64w_msa(src, (int32_t)src_stride,
+                                  dst, (int32_t)dst_stride,
+                                  filt_hor, filt_ver, (int32_t)h);
+        break;
+      default:
+        vpx_convolve8_c(src, src_stride, dst, dst_stride,
+                        filter_x, x_step_q4, filter_y, y_step_q4,
+                        w, h);
+        break;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_vert_msa.c b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_vert_msa.c
new file mode 100644
index 0000000..527d457
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve8_vert_msa.c
@@ -0,0 +1,710 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/vpx_convolve_msa.h"
+
+static void common_vt_8t_4w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, src2110, src4332, src6554, src8776;
+  v16i8 src10998, filt0, filt1, filt2, filt3;
+  v16u8 out;
+  v8i16 filt, out10, out32;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+             src54_r, src21_r);
+  ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+  ILVR_D3_SB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r, src2110,
+             src4332, src6554);
+  XORI_B3_128_SB(src2110, src4332, src6554);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+               src87_r, src98_r, src109_r);
+    ILVR_D2_SB(src87_r, src76_r, src109_r, src98_r, src8776, src10998);
+    XORI_B2_128_SB(src8776, src10998);
+    out10 = FILT_8TAP_DPADD_S_H(src2110, src4332, src6554, src8776, filt0,
+                                filt1, filt2, filt3);
+    out32 = FILT_8TAP_DPADD_S_H(src4332, src6554, src8776, src10998, filt0,
+                                filt1, filt2, filt3);
+    SRARI_H2_SH(out10, out32, FILTER_BITS);
+    SAT_SH2_SH(out10, out32, 7);
+    out = PCKEV_XORI128_UB(out10, out32);
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src2110 = src6554;
+    src4332 = src8776;
+    src6554 = src10998;
+    src6 = src10;
+  }
+}
+
+static void common_vt_8t_8w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, filt0, filt1, filt2, filt3;
+  v16u8 tmp0, tmp1;
+  v8i16 filt, out0_r, out1_r, out2_r, out3_r;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+  ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+             src54_r, src21_r);
+  ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+               src87_r, src98_r, src109_r);
+    out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0,
+                                 filt1, filt2, filt3);
+    out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0,
+                                 filt1, filt2, filt3);
+    out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0,
+                                 filt1, filt2, filt3);
+    out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0,
+                                 filt1, filt2, filt3);
+    SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS);
+    SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+    tmp0 = PCKEV_XORI128_UB(out0_r, out1_r);
+    tmp1 = PCKEV_XORI128_UB(out2_r, out3_r);
+    ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src10_r = src54_r;
+    src32_r = src76_r;
+    src54_r = src98_r;
+    src21_r = src65_r;
+    src43_r = src87_r;
+    src65_r = src109_r;
+    src6 = src10;
+  }
+}
+
+static void common_vt_8t_16w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 filt0, filt1, filt2, filt3;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, src10_l, src32_l, src54_l, src76_l;
+  v16i8 src98_l, src21_l, src43_l, src65_l, src87_l, src109_l;
+  v16u8 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt, out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+  ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+             src54_r, src21_r);
+  ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+  ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l, src32_l,
+             src54_l, src21_l);
+  ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+               src87_r, src98_r, src109_r);
+    ILVL_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_l,
+               src87_l, src98_l, src109_l);
+    out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0,
+                                 filt1, filt2, filt3);
+    out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0,
+                                 filt1, filt2, filt3);
+    out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0,
+                                 filt1, filt2, filt3);
+    out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0,
+                                 filt1, filt2, filt3);
+    out0_l = FILT_8TAP_DPADD_S_H(src10_l, src32_l, src54_l, src76_l, filt0,
+                                 filt1, filt2, filt3);
+    out1_l = FILT_8TAP_DPADD_S_H(src21_l, src43_l, src65_l, src87_l, filt0,
+                                 filt1, filt2, filt3);
+    out2_l = FILT_8TAP_DPADD_S_H(src32_l, src54_l, src76_l, src98_l, filt0,
+                                 filt1, filt2, filt3);
+    out3_l = FILT_8TAP_DPADD_S_H(src43_l, src65_l, src87_l, src109_l, filt0,
+                                 filt1, filt2, filt3);
+    SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS);
+    SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, FILTER_BITS);
+    SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+    SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7);
+    PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l, out3_r,
+                tmp0, tmp1, tmp2, tmp3);
+    XORI_B4_128_UB(tmp0, tmp1, tmp2, tmp3);
+    ST_UB4(tmp0, tmp1, tmp2, tmp3, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src10_r = src54_r;
+    src32_r = src76_r;
+    src54_r = src98_r;
+    src21_r = src65_r;
+    src43_r = src87_r;
+    src65_r = src109_r;
+    src10_l = src54_l;
+    src32_l = src76_l;
+    src54_l = src98_l;
+    src21_l = src65_l;
+    src43_l = src87_l;
+    src65_l = src109_l;
+    src6 = src10;
+  }
+}
+
+static void common_vt_8t_16w_mult_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter, int32_t height,
+                                      int32_t width) {
+  const uint8_t *src_tmp;
+  uint8_t *dst_tmp;
+  uint32_t loop_cnt, cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 filt0, filt1, filt2, filt3;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, src10_l, src32_l, src54_l, src76_l;
+  v16i8 src98_l, src21_l, src43_l, src65_l, src87_l, src109_l;
+  v16u8 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt, out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  for (cnt = (width >> 4); cnt--;) {
+    src_tmp = src;
+    dst_tmp = dst;
+
+    LD_SB7(src_tmp, src_stride, src0, src1, src2, src3, src4, src5, src6);
+    XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+    src_tmp += (7 * src_stride);
+    ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r,
+               src32_r, src54_r, src21_r);
+    ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+    ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l,
+               src32_l, src54_l, src21_l);
+    ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;) {
+      LD_SB4(src_tmp, src_stride, src7, src8, src9, src10);
+      XORI_B4_128_SB(src7, src8, src9, src10);
+      src_tmp += (4 * src_stride);
+      ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+                 src87_r, src98_r, src109_r);
+      ILVL_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_l,
+                 src87_l, src98_l, src109_l);
+      out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0,
+                                   filt1, filt2, filt3);
+      out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0,
+                                   filt1, filt2, filt3);
+      out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0,
+                                   filt1, filt2, filt3);
+      out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0,
+                                   filt1, filt2, filt3);
+      out0_l = FILT_8TAP_DPADD_S_H(src10_l, src32_l, src54_l, src76_l, filt0,
+                                   filt1, filt2, filt3);
+      out1_l = FILT_8TAP_DPADD_S_H(src21_l, src43_l, src65_l, src87_l, filt0,
+                                   filt1, filt2, filt3);
+      out2_l = FILT_8TAP_DPADD_S_H(src32_l, src54_l, src76_l, src98_l, filt0,
+                                   filt1, filt2, filt3);
+      out3_l = FILT_8TAP_DPADD_S_H(src43_l, src65_l, src87_l, src109_l, filt0,
+                                   filt1, filt2, filt3);
+      SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS);
+      SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, FILTER_BITS);
+      SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+      SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7);
+      PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l,
+                  out3_r, tmp0, tmp1, tmp2, tmp3);
+      XORI_B4_128_UB(tmp0, tmp1, tmp2, tmp3);
+      ST_UB4(tmp0, tmp1, tmp2, tmp3, dst_tmp, dst_stride);
+      dst_tmp += (4 * dst_stride);
+
+      src10_r = src54_r;
+      src32_r = src76_r;
+      src54_r = src98_r;
+      src21_r = src65_r;
+      src43_r = src87_r;
+      src65_r = src109_r;
+      src10_l = src54_l;
+      src32_l = src76_l;
+      src54_l = src98_l;
+      src21_l = src65_l;
+      src43_l = src87_l;
+      src65_l = src109_l;
+      src6 = src10;
+    }
+
+    src += 16;
+    dst += 16;
+  }
+}
+
+static void common_vt_8t_32w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  common_vt_8t_16w_mult_msa(src, src_stride, dst, dst_stride, filter, height,
+                            32);
+}
+
+static void common_vt_8t_64w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  common_vt_8t_16w_mult_msa(src, src_stride, dst, dst_stride, filter, height,
+                            64);
+}
+
+static void common_vt_2t_4x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16i8 src0, src1, src2, src3, src4;
+  v16i8 src10_r, src32_r, src21_r, src43_r, src2110, src4332;
+  v16u8 filt0;
+  v8i16 filt;
+  v8u16 tmp0, tmp1;
+
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+  src += (5 * src_stride);
+
+  ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+             src32_r, src43_r);
+  ILVR_D2_SB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+  DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1);
+  SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+  src2110 = __msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+  ST4x4_UB(src2110, src2110, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_vt_2t_4x8_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src2110, src4332, src6554, src8776;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 filt0;
+  v8i16 filt;
+
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  src += (8 * src_stride);
+
+  src8 = LD_SB(src);
+  src += src_stride;
+
+  ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+             src32_r, src43_r);
+  ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r, src65_r,
+             src76_r, src87_r);
+  ILVR_D4_SB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r,
+             src87_r, src76_r, src2110, src4332, src6554, src8776);
+  DOTP_UB4_UH(src2110, src4332, src6554, src8776, filt0, filt0, filt0, filt0,
+              tmp0, tmp1, tmp2, tmp3);
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, src2110, src4332);
+  ST4x4_UB(src2110, src2110, 0, 1, 2, 3, dst, dst_stride);
+  ST4x4_UB(src4332, src4332, 0, 1, 2, 3, dst + 4 * dst_stride, dst_stride);
+}
+
+static void common_vt_2t_4w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_vt_2t_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_vt_2t_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_vt_2t_8x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16u8 src0, src1, src2, src3, src4, vec0, vec1, vec2, vec3, filt0;
+  v16i8 out0, out1;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_UB5(src, src_stride, src0, src1, src2, src3, src4);
+  ILVR_B2_UB(src1, src0, src2, src1, vec0, vec1);
+  ILVR_B2_UB(src3, src2, src4, src3, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1,
+              tmp2, tmp3);
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+  ST8x4_UB(out0, out1, dst, dst_stride);
+}
+
+static void common_vt_2t_8x8mult_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v16i8 out0, out1;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 3); loop_cnt--;) {
+    LD_UB8(src, src_stride, src1, src2, src3, src4, src5, src6, src7, src8);
+    src += (8 * src_stride);
+
+    ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, vec0, vec1,
+               vec2, vec3);
+    ILVR_B4_UB(src5, src4, src6, src5, src7, src6, src8, src7, vec4, vec5,
+               vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1,
+                tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, tmp0, tmp1,
+                tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src0 = src8;
+  }
+}
+
+static void common_vt_2t_8w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_vt_2t_8x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else {
+    common_vt_2t_8x8mult_msa(src, src_stride, dst, dst_stride, filter, height);
+  }
+}
+
+static void common_vt_2t_16w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst);
+    dst += dst_stride;
+
+    ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst);
+    dst += dst_stride;
+
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst);
+    dst += dst_stride;
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst);
+    dst += dst_stride;
+
+    src0 = src4;
+  }
+}
+
+static void common_vt_2t_32w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_UB(src);
+  src5 = LD_UB(src + 16);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+
+    LD_UB4(src + 16, src_stride, src6, src7, src8, src9);
+    src += (4 * src_stride);
+
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst);
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + dst_stride);
+
+    ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst + 2 * dst_stride);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + 3 * dst_stride);
+
+    ILVR_B2_UB(src6, src5, src7, src6, vec0, vec2);
+    ILVL_B2_UB(src6, src5, src7, src6, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst + 16);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + 16 + dst_stride);
+
+    ILVR_B2_UB(src8, src7, src9, src8, vec4, vec6);
+    ILVL_B2_UB(src8, src7, src9, src8, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst + 16 + 2 * dst_stride);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + 16 + 3 * dst_stride);
+    dst += (4 * dst_stride);
+
+    src0 = src4;
+    src5 = src9;
+  }
+}
+
+static void common_vt_2t_64w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16u8 src11, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_UB4(src, 16, src0, src3, src6, src9);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 1); loop_cnt--;) {
+    LD_UB2(src, src_stride, src1, src2);
+    LD_UB2(src + 16, src_stride, src4, src5);
+    LD_UB2(src + 32, src_stride, src7, src8);
+    LD_UB2(src + 48, src_stride, src10, src11);
+    src += (2 * src_stride);
+
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + dst_stride);
+
+    ILVR_B2_UB(src4, src3, src5, src4, vec4, vec6);
+    ILVL_B2_UB(src4, src3, src5, src4, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp4, tmp5);
+    SRARI_H2_UH(tmp4, tmp5, FILTER_BITS);
+    PCKEV_ST_SB(tmp4, tmp5, dst + 16);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp6, tmp7);
+    SRARI_H2_UH(tmp6, tmp7, FILTER_BITS);
+    PCKEV_ST_SB(tmp6, tmp7, dst + 16 + dst_stride);
+
+    ILVR_B2_UB(src7, src6, src8, src7, vec0, vec2);
+    ILVL_B2_UB(src7, src6, src8, src7, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst + 32);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + 32 + dst_stride);
+
+    ILVR_B2_UB(src10, src9, src11, src10, vec4, vec6);
+    ILVL_B2_UB(src10, src9, src11, src10, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp4, tmp5);
+    SRARI_H2_UH(tmp4, tmp5, FILTER_BITS);
+    PCKEV_ST_SB(tmp4, tmp5, dst + 48);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp6, tmp7);
+    SRARI_H2_UH(tmp6, tmp7, FILTER_BITS);
+    PCKEV_ST_SB(tmp6, tmp7, dst + 48 + dst_stride);
+    dst += (2 * dst_stride);
+
+    src0 = src2;
+    src3 = src5;
+    src6 = src8;
+    src9 = src11;
+  }
+}
+
+void vpx_convolve8_vert_msa(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4,
+                            int w, int h) {
+  int8_t cnt, filt_ver[8];
+
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  for (cnt = 8; cnt--;) {
+    filt_ver[cnt] = filter_y[cnt];
+  }
+
+  if (((const int32_t *)filter_y)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_vt_2t_4w_msa(src, (int32_t)src_stride,
+                            dst, (int32_t)dst_stride,
+                            &filt_ver[3], h);
+        break;
+      case 8:
+        common_vt_2t_8w_msa(src, (int32_t)src_stride,
+                            dst, (int32_t)dst_stride,
+                            &filt_ver[3], h);
+        break;
+      case 16:
+        common_vt_2t_16w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             &filt_ver[3], h);
+        break;
+      case 32:
+        common_vt_2t_32w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             &filt_ver[3], h);
+        break;
+      case 64:
+        common_vt_2t_64w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             &filt_ver[3], h);
+        break;
+      default:
+        vpx_convolve8_vert_c(src, src_stride, dst, dst_stride,
+                             filter_x, x_step_q4, filter_y, y_step_q4,
+                             w, h);
+        break;
+    }
+  } else {
+    switch (w) {
+      case 4:
+        common_vt_8t_4w_msa(src, (int32_t)src_stride,
+                            dst, (int32_t)dst_stride,
+                            filt_ver, h);
+        break;
+      case 8:
+        common_vt_8t_8w_msa(src, (int32_t)src_stride,
+                            dst, (int32_t)dst_stride,
+                            filt_ver, h);
+        break;
+      case 16:
+        common_vt_8t_16w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             filt_ver, h);
+        break;
+      case 32:
+        common_vt_8t_32w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             filt_ver, h);
+        break;
+      case 64:
+        common_vt_8t_64w_msa(src, (int32_t)src_stride,
+                             dst, (int32_t)dst_stride,
+                             filt_ver, h);
+        break;
+      default:
+        vpx_convolve8_vert_c(src, src_stride, dst, dst_stride,
+                             filter_x, x_step_q4, filter_y, y_step_q4,
+                             w, h);
+        break;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/vpx_convolve_avg_msa.c b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve_avg_msa.c
new file mode 100644
index 0000000..4c3d978
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve_avg_msa.c
@@ -0,0 +1,232 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/macros_msa.h"
+
+static void avg_width4_msa(const uint8_t *src, int32_t src_stride,
+                           uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  uint32_t out0, out1, out2, out3;
+  v16u8 src0, src1, src2, src3;
+  v16u8 dst0, dst1, dst2, dst3;
+
+  if (0 == (height % 4)) {
+    for (cnt = (height / 4); cnt--;) {
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      src += (4 * src_stride);
+
+      LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+
+      AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3,
+                  dst0, dst1, dst2, dst3);
+
+      out0 = __msa_copy_u_w((v4i32)dst0, 0);
+      out1 = __msa_copy_u_w((v4i32)dst1, 0);
+      out2 = __msa_copy_u_w((v4i32)dst2, 0);
+      out3 = __msa_copy_u_w((v4i32)dst3, 0);
+      SW4(out0, out1, out2, out3, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == (height % 2)) {
+    for (cnt = (height / 2); cnt--;) {
+      LD_UB2(src, src_stride, src0, src1);
+      src += (2 * src_stride);
+
+      LD_UB2(dst, dst_stride, dst0, dst1);
+
+      AVER_UB2_UB(src0, dst0, src1, dst1, dst0, dst1);
+
+      out0 = __msa_copy_u_w((v4i32)dst0, 0);
+      out1 = __msa_copy_u_w((v4i32)dst1, 0);
+      SW(out0, dst);
+      dst += dst_stride;
+      SW(out1, dst);
+      dst += dst_stride;
+    }
+  }
+}
+
+static void avg_width8_msa(const uint8_t *src, int32_t src_stride,
+                           uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  uint64_t out0, out1, out2, out3;
+  v16u8 src0, src1, src2, src3;
+  v16u8 dst0, dst1, dst2, dst3;
+
+  for (cnt = (height / 4); cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+
+    AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3,
+                dst0, dst1, dst2, dst3);
+
+    out0 = __msa_copy_u_d((v2i64)dst0, 0);
+    out1 = __msa_copy_u_d((v2i64)dst1, 0);
+    out2 = __msa_copy_u_d((v2i64)dst2, 0);
+    out3 = __msa_copy_u_d((v2i64)dst3, 0);
+    SD4(out0, out1, out2, out3, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void avg_width16_msa(const uint8_t *src, int32_t src_stride,
+                            uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+
+  for (cnt = (height / 8); cnt--;) {
+    LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    src += (8 * src_stride);
+    LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+
+    AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3,
+                dst0, dst1, dst2, dst3);
+    AVER_UB4_UB(src4, dst4, src5, dst5, src6, dst6, src7, dst7,
+                dst4, dst5, dst6, dst7);
+    ST_UB8(dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7, dst, dst_stride);
+    dst += (8 * dst_stride);
+  }
+}
+
+static void avg_width32_msa(const uint8_t *src, int32_t src_stride,
+                            uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  uint8_t *dst_dup = dst;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 src8, src9, src10, src11, src12, src13, src14, src15;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16u8 dst8, dst9, dst10, dst11, dst12, dst13, dst14, dst15;
+
+  for (cnt = (height / 8); cnt--;) {
+    LD_UB4(src, src_stride, src0, src2, src4, src6);
+    LD_UB4(src + 16, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(dst_dup, dst_stride, dst0, dst2, dst4, dst6);
+    LD_UB4(dst_dup + 16, dst_stride, dst1, dst3, dst5, dst7);
+    dst_dup += (4 * dst_stride);
+    LD_UB4(src, src_stride, src8, src10, src12, src14);
+    LD_UB4(src + 16, src_stride, src9, src11, src13, src15);
+    src += (4 * src_stride);
+    LD_UB4(dst_dup, dst_stride, dst8, dst10, dst12, dst14);
+    LD_UB4(dst_dup + 16, dst_stride, dst9, dst11, dst13, dst15);
+    dst_dup += (4 * dst_stride);
+
+    AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3,
+                dst0, dst1, dst2, dst3);
+    AVER_UB4_UB(src4, dst4, src5, dst5, src6, dst6, src7, dst7,
+                dst4, dst5, dst6, dst7);
+    AVER_UB4_UB(src8, dst8, src9, dst9, src10, dst10, src11, dst11,
+                dst8, dst9, dst10, dst11);
+    AVER_UB4_UB(src12, dst12, src13, dst13, src14, dst14, src15, dst15,
+                dst12, dst13, dst14, dst15);
+
+    ST_UB4(dst0, dst2, dst4, dst6, dst, dst_stride);
+    ST_UB4(dst1, dst3, dst5, dst7, dst + 16, dst_stride);
+    dst += (4 * dst_stride);
+    ST_UB4(dst8, dst10, dst12, dst14, dst, dst_stride);
+    ST_UB4(dst9, dst11, dst13, dst15, dst + 16, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void avg_width64_msa(const uint8_t *src, int32_t src_stride,
+                            uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  uint8_t *dst_dup = dst;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 src8, src9, src10, src11, src12, src13, src14, src15;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16u8 dst8, dst9, dst10, dst11, dst12, dst13, dst14, dst15;
+
+  for (cnt = (height / 4); cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(src, 16, src4, src5, src6, src7);
+    src += src_stride;
+    LD_UB4(src, 16, src8, src9, src10, src11);
+    src += src_stride;
+    LD_UB4(src, 16, src12, src13, src14, src15);
+    src += src_stride;
+
+    LD_UB4(dst_dup, 16, dst0, dst1, dst2, dst3);
+    dst_dup += dst_stride;
+    LD_UB4(dst_dup, 16, dst4, dst5, dst6, dst7);
+    dst_dup += dst_stride;
+    LD_UB4(dst_dup, 16, dst8, dst9, dst10, dst11);
+    dst_dup += dst_stride;
+    LD_UB4(dst_dup, 16, dst12, dst13, dst14, dst15);
+    dst_dup += dst_stride;
+
+    AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3,
+                dst0, dst1, dst2, dst3);
+    AVER_UB4_UB(src4, dst4, src5, dst5, src6, dst6, src7, dst7,
+                dst4, dst5, dst6, dst7);
+    AVER_UB4_UB(src8, dst8, src9, dst9, src10, dst10, src11, dst11,
+                dst8, dst9, dst10, dst11);
+    AVER_UB4_UB(src12, dst12, src13, dst13, src14, dst14, src15, dst15,
+                dst12, dst13, dst14, dst15);
+
+    ST_UB4(dst0, dst1, dst2, dst3, dst, 16);
+    dst += dst_stride;
+    ST_UB4(dst4, dst5, dst6, dst7, dst, 16);
+    dst += dst_stride;
+    ST_UB4(dst8, dst9, dst10, dst11, dst, 16);
+    dst += dst_stride;
+    ST_UB4(dst12, dst13, dst14, dst15, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+void vpx_convolve_avg_msa(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const int16_t *filter_x, int32_t filter_x_stride,
+                          const int16_t *filter_y, int32_t filter_y_stride,
+                          int32_t w, int32_t h) {
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+
+  switch (w) {
+    case 4: {
+      avg_width4_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 8: {
+      avg_width8_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 16: {
+      avg_width16_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 32: {
+      avg_width32_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 64: {
+      avg_width64_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    default: {
+      int32_t lp, cnt;
+      for (cnt = h; cnt--;) {
+        for (lp = 0; lp < w; ++lp) {
+          dst[lp] = (((dst[lp] + src[lp]) + 1) >> 1);
+        }
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/vpx_convolve_copy_msa.c b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve_copy_msa.c
new file mode 100644
index 0000000..ba40122
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve_copy_msa.c
@@ -0,0 +1,247 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <string.h>
+#include "vpx_dsp/mips/macros_msa.h"
+
+static void copy_width8_msa(const uint8_t *src, int32_t src_stride,
+                            uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  uint64_t out0, out1, out2, out3, out4, out5, out6, out7;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+
+  if (0 == height % 12) {
+    for (cnt = (height / 12); cnt--;) {
+      LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+      src += (8 * src_stride);
+
+      out0 = __msa_copy_u_d((v2i64)src0, 0);
+      out1 = __msa_copy_u_d((v2i64)src1, 0);
+      out2 = __msa_copy_u_d((v2i64)src2, 0);
+      out3 = __msa_copy_u_d((v2i64)src3, 0);
+      out4 = __msa_copy_u_d((v2i64)src4, 0);
+      out5 = __msa_copy_u_d((v2i64)src5, 0);
+      out6 = __msa_copy_u_d((v2i64)src6, 0);
+      out7 = __msa_copy_u_d((v2i64)src7, 0);
+
+      SD4(out0, out1, out2, out3, dst, dst_stride);
+      dst += (4 * dst_stride);
+      SD4(out4, out5, out6, out7, dst, dst_stride);
+      dst += (4 * dst_stride);
+
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      src += (4 * src_stride);
+
+      out0 = __msa_copy_u_d((v2i64)src0, 0);
+      out1 = __msa_copy_u_d((v2i64)src1, 0);
+      out2 = __msa_copy_u_d((v2i64)src2, 0);
+      out3 = __msa_copy_u_d((v2i64)src3, 0);
+      SD4(out0, out1, out2, out3, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == height % 8) {
+    for (cnt = height >> 3; cnt--;) {
+      LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+      src += (8 * src_stride);
+
+      out0 = __msa_copy_u_d((v2i64)src0, 0);
+      out1 = __msa_copy_u_d((v2i64)src1, 0);
+      out2 = __msa_copy_u_d((v2i64)src2, 0);
+      out3 = __msa_copy_u_d((v2i64)src3, 0);
+      out4 = __msa_copy_u_d((v2i64)src4, 0);
+      out5 = __msa_copy_u_d((v2i64)src5, 0);
+      out6 = __msa_copy_u_d((v2i64)src6, 0);
+      out7 = __msa_copy_u_d((v2i64)src7, 0);
+
+      SD4(out0, out1, out2, out3, dst, dst_stride);
+      dst += (4 * dst_stride);
+      SD4(out4, out5, out6, out7, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == height % 4) {
+    for (cnt = (height / 4); cnt--;) {
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      src += (4 * src_stride);
+      out0 = __msa_copy_u_d((v2i64)src0, 0);
+      out1 = __msa_copy_u_d((v2i64)src1, 0);
+      out2 = __msa_copy_u_d((v2i64)src2, 0);
+      out3 = __msa_copy_u_d((v2i64)src3, 0);
+
+      SD4(out0, out1, out2, out3, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == height % 2) {
+    for (cnt = (height / 2); cnt--;) {
+      LD_UB2(src, src_stride, src0, src1);
+      src += (2 * src_stride);
+      out0 = __msa_copy_u_d((v2i64)src0, 0);
+      out1 = __msa_copy_u_d((v2i64)src1, 0);
+
+      SD(out0, dst);
+      dst += dst_stride;
+      SD(out1, dst);
+      dst += dst_stride;
+    }
+  }
+}
+
+static void copy_16multx8mult_msa(const uint8_t *src, int32_t src_stride,
+                                  uint8_t *dst, int32_t dst_stride,
+                                  int32_t height, int32_t width) {
+  int32_t cnt, loop_cnt;
+  const uint8_t *src_tmp;
+  uint8_t *dst_tmp;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+
+  for (cnt = (width >> 4); cnt--;) {
+    src_tmp = src;
+    dst_tmp = dst;
+
+    for (loop_cnt = (height >> 3); loop_cnt--;) {
+      LD_UB8(src_tmp, src_stride,
+             src0, src1, src2, src3, src4, src5, src6, src7);
+      src_tmp += (8 * src_stride);
+
+      ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7,
+             dst_tmp, dst_stride);
+      dst_tmp += (8 * dst_stride);
+    }
+
+    src += 16;
+    dst += 16;
+  }
+}
+
+static void copy_width16_msa(const uint8_t *src, int32_t src_stride,
+                             uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+
+  if (0 == height % 12) {
+    for (cnt = (height / 12); cnt--;) {
+      LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+      src += (8 * src_stride);
+      ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
+      dst += (8 * dst_stride);
+
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      src += (4 * src_stride);
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == height % 8) {
+    copy_16multx8mult_msa(src, src_stride, dst, dst_stride, height, 16);
+  } else if (0 == height % 4) {
+    for (cnt = (height >> 2); cnt--;) {
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      src += (4 * src_stride);
+
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  }
+}
+
+static void copy_width32_msa(const uint8_t *src, int32_t src_stride,
+                             uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+
+  if (0 == height % 12) {
+    for (cnt = (height / 12); cnt--;) {
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      LD_UB4(src + 16, src_stride, src4, src5, src6, src7);
+      src += (4 * src_stride);
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride);
+      dst += (4 * dst_stride);
+
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      LD_UB4(src + 16, src_stride, src4, src5, src6, src7);
+      src += (4 * src_stride);
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride);
+      dst += (4 * dst_stride);
+
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      LD_UB4(src + 16, src_stride, src4, src5, src6, src7);
+      src += (4 * src_stride);
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == height % 8) {
+    copy_16multx8mult_msa(src, src_stride, dst, dst_stride, height, 32);
+  } else if (0 == height % 4) {
+    for (cnt = (height >> 2); cnt--;) {
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      LD_UB4(src + 16, src_stride, src4, src5, src6, src7);
+      src += (4 * src_stride);
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  }
+}
+
+static void copy_width64_msa(const uint8_t *src, int32_t src_stride,
+                             uint8_t *dst, int32_t dst_stride, int32_t height) {
+  copy_16multx8mult_msa(src, src_stride, dst, dst_stride, height, 64);
+}
+
+void vpx_convolve_copy_msa(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int32_t filter_x_stride,
+                           const int16_t *filter_y, int32_t filter_y_stride,
+                           int32_t w, int32_t h) {
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+
+  switch (w) {
+    case 4: {
+      uint32_t cnt, tmp;
+      /* 1 word storage */
+      for (cnt = h; cnt--;) {
+        tmp = LW(src);
+        SW(tmp, dst);
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    }
+    case 8: {
+      copy_width8_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 16: {
+      copy_width16_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 32: {
+      copy_width32_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 64: {
+      copy_width64_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    default: {
+      uint32_t cnt;
+      for (cnt = h; cnt--;) {
+        memcpy(dst, src, w);
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    }
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/mips/vpx_convolve_msa.h b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve_msa.h
new file mode 100644
index 0000000..e001398
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/mips/vpx_convolve_msa.h
@@ -0,0 +1,119 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_MIPS_VPX_CONVOLVE_MSA_H_
+#define VPX_DSP_MIPS_VPX_CONVOLVE_MSA_H_
+
+#include "vpx_dsp/mips/macros_msa.h"
+#include "vpx_dsp/vpx_filter.h"
+
+extern const uint8_t mc_filt_mask_arr[16 * 3];
+
+#define FILT_8TAP_DPADD_S_H(vec0, vec1, vec2, vec3,         \
+                            filt0, filt1, filt2, filt3) ({  \
+  v8i16 tmp0, tmp1;                                         \
+                                                            \
+  tmp0 = __msa_dotp_s_h((v16i8)vec0, (v16i8)filt0);         \
+  tmp0 = __msa_dpadd_s_h(tmp0, (v16i8)vec1, (v16i8)filt1);  \
+  tmp1 = __msa_dotp_s_h((v16i8)vec2, (v16i8)filt2);         \
+  tmp1 = __msa_dpadd_s_h(tmp1, (v16i8)vec3, (v16i8)filt3);  \
+  tmp0 = __msa_adds_s_h(tmp0, tmp1);                        \
+                                                            \
+  tmp0;                                                     \
+})
+
+#define HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3,        \
+                        filt_h0, filt_h1, filt_h2, filt_h3) ({         \
+  v16i8 vec0_m, vec1_m, vec2_m, vec3_m;                                \
+  v8i16 hz_out_m;                                                      \
+                                                                       \
+  VSHF_B4_SB(src0, src1, mask0, mask1, mask2, mask3,                   \
+             vec0_m, vec1_m, vec2_m, vec3_m);                          \
+  hz_out_m = FILT_8TAP_DPADD_S_H(vec0_m, vec1_m, vec2_m, vec3_m,       \
+                                 filt_h0, filt_h1, filt_h2, filt_h3);  \
+                                                                       \
+  hz_out_m = __msa_srari_h(hz_out_m, FILTER_BITS);                     \
+  hz_out_m = __msa_sat_s_h(hz_out_m, 7);                               \
+                                                                       \
+  hz_out_m;                                                            \
+})
+
+#define HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3,               \
+                                   mask0, mask1, mask2, mask3,           \
+                                   filt0, filt1, filt2, filt3,           \
+                                   out0, out1) {                         \
+  v16i8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;  \
+  v8i16 res0_m, res1_m, res2_m, res3_m;                                  \
+                                                                         \
+  VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0_m, vec1_m);      \
+  DOTP_SB2_SH(vec0_m, vec1_m, filt0, filt0, res0_m, res1_m);             \
+  VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2_m, vec3_m);      \
+  DPADD_SB2_SH(vec2_m, vec3_m, filt1, filt1, res0_m, res1_m);            \
+  VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4_m, vec5_m);      \
+  DOTP_SB2_SH(vec4_m, vec5_m, filt2, filt2, res2_m, res3_m);             \
+  VSHF_B2_SB(src0, src1, src2, src3, mask3, mask3, vec6_m, vec7_m);      \
+  DPADD_SB2_SH(vec6_m, vec7_m, filt3, filt3, res2_m, res3_m);            \
+  ADDS_SH2_SH(res0_m, res2_m, res1_m, res3_m, out0, out1);               \
+}
+
+#define HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3,                  \
+                                   mask0, mask1, mask2, mask3,              \
+                                   filt0, filt1, filt2, filt3,              \
+                                   out0, out1, out2, out3) {                \
+  v16i8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;     \
+  v8i16 res0_m, res1_m, res2_m, res3_m, res4_m, res5_m, res6_m, res7_m;     \
+                                                                            \
+  VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0_m, vec1_m);         \
+  VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2_m, vec3_m);         \
+  DOTP_SB4_SH(vec0_m, vec1_m, vec2_m, vec3_m, filt0, filt0, filt0, filt0,   \
+              res0_m, res1_m, res2_m, res3_m);                              \
+  VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec0_m, vec1_m);         \
+  VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec2_m, vec3_m);         \
+  DOTP_SB4_SH(vec0_m, vec1_m, vec2_m, vec3_m, filt2, filt2, filt2, filt2,   \
+              res4_m, res5_m, res6_m, res7_m);                              \
+  VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec4_m, vec5_m);         \
+  VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec6_m, vec7_m);         \
+  DPADD_SB4_SH(vec4_m, vec5_m, vec6_m, vec7_m, filt1, filt1, filt1, filt1,  \
+               res0_m, res1_m, res2_m, res3_m);                             \
+  VSHF_B2_SB(src0, src0, src1, src1, mask3, mask3, vec4_m, vec5_m);         \
+  VSHF_B2_SB(src2, src2, src3, src3, mask3, mask3, vec6_m, vec7_m);         \
+  DPADD_SB4_SH(vec4_m, vec5_m, vec6_m, vec7_m, filt3, filt3, filt3, filt3,  \
+               res4_m, res5_m, res6_m, res7_m);                             \
+  ADDS_SH4_SH(res0_m, res4_m, res1_m, res5_m, res2_m, res6_m, res3_m,       \
+              res7_m, out0, out1, out2, out3);                              \
+}
+
+#define PCKEV_XORI128_AVG_ST_UB(in0, in1, dst, pdst) {  \
+  v16u8 tmp_m;                                          \
+                                                        \
+  tmp_m = PCKEV_XORI128_UB(in1, in0);                   \
+  tmp_m = __msa_aver_u_b(tmp_m, (v16u8)dst);            \
+  ST_UB(tmp_m, (pdst));                                 \
+}
+
+#define PCKEV_AVG_ST_UB(in0, in1, dst, pdst) {           \
+  v16u8 tmp_m;                                           \
+                                                         \
+  tmp_m = (v16u8)__msa_pckev_b((v16i8)in0, (v16i8)in1);  \
+  tmp_m = __msa_aver_u_b(tmp_m, (v16u8)dst);             \
+  ST_UB(tmp_m, (pdst));                                  \
+}
+
+#define PCKEV_AVG_ST8x4_UB(in1, dst0, in2, dst1, in3, dst2, in4, dst3,  \
+                           pdst, stride) {                              \
+  v16u8 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                 \
+  uint8_t *pdst_m = (uint8_t *)(pdst);                                  \
+                                                                        \
+  PCKEV_B2_UB(in2, in1, in4, in3, tmp0_m, tmp1_m);                      \
+  PCKEV_D2_UB(dst1, dst0, dst3, dst2, tmp2_m, tmp3_m);                  \
+  AVER_UB2_UB(tmp0_m, tmp2_m, tmp1_m, tmp3_m, tmp0_m, tmp1_m);          \
+  ST8x4_UB(tmp0_m, tmp1_m, pdst_m, stride);                             \
+}
+#endif  /* VPX_DSP_MIPS_VPX_CONVOLVE_MSA_H_ */
diff --git a/libvpx/libvpx/vpx_dsp/prob.c b/libvpx/libvpx/vpx_dsp/prob.c
new file mode 100644
index 0000000..639d24d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/prob.c
@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./prob.h"
+
+const uint8_t vpx_norm[256] = {
+  0, 7, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4,
+  3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
+  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
+};
+
+static unsigned int tree_merge_probs_impl(unsigned int i,
+                                          const vpx_tree_index *tree,
+                                          const vpx_prob *pre_probs,
+                                          const unsigned int *counts,
+                                          vpx_prob *probs) {
+  const int l = tree[i];
+  const unsigned int left_count = (l <= 0)
+                 ? counts[-l]
+                 : tree_merge_probs_impl(l, tree, pre_probs, counts, probs);
+  const int r = tree[i + 1];
+  const unsigned int right_count = (r <= 0)
+                 ? counts[-r]
+                 : tree_merge_probs_impl(r, tree, pre_probs, counts, probs);
+  const unsigned int ct[2] = { left_count, right_count };
+  probs[i >> 1] = mode_mv_merge_probs(pre_probs[i >> 1], ct);
+  return left_count + right_count;
+}
+
+void vpx_tree_merge_probs(const vpx_tree_index *tree, const vpx_prob *pre_probs,
+                          const unsigned int *counts, vpx_prob *probs) {
+  tree_merge_probs_impl(0, tree, pre_probs, counts, probs);
+}
diff --git a/libvpx/libvpx/vpx_dsp/prob.h b/libvpx/libvpx/vpx_dsp/prob.h
new file mode 100644
index 0000000..c3cb103
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/prob.h
@@ -0,0 +1,103 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_PROB_H_
+#define VPX_DSP_PROB_H_
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_common.h"
+
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef uint8_t vpx_prob;
+
+#define MAX_PROB 255
+
+#define vpx_prob_half ((vpx_prob) 128)
+
+typedef int8_t vpx_tree_index;
+
+#define TREE_SIZE(leaf_count) (2 * (leaf_count) - 2)
+
+#define vpx_complement(x) (255 - x)
+
+#define MODE_MV_COUNT_SAT 20
+
+/* We build coding trees compactly in arrays.
+   Each node of the tree is a pair of vpx_tree_indices.
+   Array index often references a corresponding probability table.
+   Index <= 0 means done encoding/decoding and value = -Index,
+   Index > 0 means need another bit, specification at index.
+   Nonnegative indices are always even;  processing begins at node 0. */
+
+typedef const vpx_tree_index vpx_tree[];
+
+static INLINE vpx_prob clip_prob(int p) {
+  return (p > 255) ? 255 : (p < 1) ? 1 : p;
+}
+
+static INLINE vpx_prob get_prob(int num, int den) {
+  return (den == 0) ? 128u : clip_prob(((int64_t)num * 256 + (den >> 1)) / den);
+}
+
+static INLINE vpx_prob get_binary_prob(int n0, int n1) {
+  return get_prob(n0, n0 + n1);
+}
+
+/* This function assumes prob1 and prob2 are already within [1,255] range. */
+static INLINE vpx_prob weighted_prob(int prob1, int prob2, int factor) {
+  return ROUND_POWER_OF_TWO(prob1 * (256 - factor) + prob2 * factor, 8);
+}
+
+static INLINE vpx_prob merge_probs(vpx_prob pre_prob,
+                                   const unsigned int ct[2],
+                                   unsigned int count_sat,
+                                   unsigned int max_update_factor) {
+  const vpx_prob prob = get_binary_prob(ct[0], ct[1]);
+  const unsigned int count = VPXMIN(ct[0] + ct[1], count_sat);
+  const unsigned int factor = max_update_factor * count / count_sat;
+  return weighted_prob(pre_prob, prob, factor);
+}
+
+// MODE_MV_MAX_UPDATE_FACTOR (128) * count / MODE_MV_COUNT_SAT;
+static const int count_to_update_factor[MODE_MV_COUNT_SAT + 1] = {
+  0, 6, 12, 19, 25, 32, 38, 44, 51, 57, 64,
+  70, 76, 83, 89, 96, 102, 108, 115, 121, 128
+};
+
+static INLINE vpx_prob mode_mv_merge_probs(vpx_prob pre_prob,
+                                           const unsigned int ct[2]) {
+  const unsigned int den = ct[0] + ct[1];
+  if (den == 0) {
+    return pre_prob;
+  } else {
+    const unsigned int count = VPXMIN(den, MODE_MV_COUNT_SAT);
+    const unsigned int factor = count_to_update_factor[count];
+    const vpx_prob prob =
+        clip_prob(((int64_t)(ct[0]) * 256 + (den >> 1)) / den);
+    return weighted_prob(pre_prob, prob, factor);
+  }
+}
+
+void vpx_tree_merge_probs(const vpx_tree_index *tree, const vpx_prob *pre_probs,
+                          const unsigned int *counts, vpx_prob *probs);
+
+
+DECLARE_ALIGNED(16, extern const uint8_t, vpx_norm[256]);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_PROB_H_
diff --git a/libvpx/libvpx/vpx_dsp/psnrhvs.c b/libvpx/libvpx/vpx_dsp/psnrhvs.c
new file mode 100644
index 0000000..0ffa1b2
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/psnrhvs.c
@@ -0,0 +1,229 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ *
+ *  This code was originally written by: Gregory Maxwell, at the Daala
+ *  project.
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <math.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/ssim.h"
+#include "vpx_ports/system_state.h"
+
+#if !defined(M_PI)
+# define M_PI (3.141592653589793238462643)
+#endif
+#include <string.h>
+
+static void od_bin_fdct8x8(tran_low_t *y, int ystride, const int16_t *x,
+                           int xstride) {
+  (void) xstride;
+  vpx_fdct8x8(x, y, ystride);
+}
+
+/* Normalized inverse quantization matrix for 8x8 DCT at the point of
+ * transparency. This is not the JPEG based matrix from the paper,
+ this one gives a slightly higher MOS agreement.*/
+static const float csf_y[8][8] = {
+    {1.6193873005, 2.2901594831, 2.08509755623, 1.48366094411, 1.00227514334,
+     0.678296995242, 0.466224900598, 0.3265091542},
+    {2.2901594831, 1.94321815382, 2.04793073064, 1.68731108984, 1.2305666963,
+     0.868920337363, 0.61280991668, 0.436405793551},
+    {2.08509755623, 2.04793073064, 1.34329019223, 1.09205635862, 0.875748795257,
+     0.670882927016, 0.501731932449, 0.372504254596},
+    {1.48366094411, 1.68731108984, 1.09205635862, 0.772819797575,
+     0.605636379554, 0.48309405692, 0.380429446972, 0.295774038565},
+    {1.00227514334, 1.2305666963, 0.875748795257, 0.605636379554,
+     0.448996256676, 0.352889268808, 0.283006984131, 0.226951348204},
+    {0.678296995242, 0.868920337363, 0.670882927016, 0.48309405692,
+     0.352889268808, 0.27032073436, 0.215017739696, 0.17408067321},
+    {0.466224900598, 0.61280991668, 0.501731932449, 0.380429446972,
+     0.283006984131, 0.215017739696, 0.168869545842, 0.136153931001},
+    {0.3265091542, 0.436405793551, 0.372504254596, 0.295774038565,
+     0.226951348204, 0.17408067321, 0.136153931001, 0.109083846276}};
+static const float csf_cb420[8][8] = {
+    {1.91113096927, 2.46074210438, 1.18284184739, 1.14982565193, 1.05017074788,
+     0.898018824055, 0.74725392039, 0.615105596242},
+    {2.46074210438, 1.58529308355, 1.21363250036, 1.38190029285, 1.33100189972,
+     1.17428548929, 0.996404342439, 0.830890433625},
+    {1.18284184739, 1.21363250036, 0.978712413627, 1.02624506078, 1.03145147362,
+     0.960060382087, 0.849823426169, 0.731221236837},
+    {1.14982565193, 1.38190029285, 1.02624506078, 0.861317501629,
+     0.801821139099, 0.751437590932, 0.685398513368, 0.608694761374},
+    {1.05017074788, 1.33100189972, 1.03145147362, 0.801821139099,
+     0.676555426187, 0.605503172737, 0.55002013668, 0.495804539034},
+    {0.898018824055, 1.17428548929, 0.960060382087, 0.751437590932,
+     0.605503172737, 0.514674450957, 0.454353482512, 0.407050308965},
+    {0.74725392039, 0.996404342439, 0.849823426169, 0.685398513368,
+     0.55002013668, 0.454353482512, 0.389234902883, 0.342353999733},
+    {0.615105596242, 0.830890433625, 0.731221236837, 0.608694761374,
+     0.495804539034, 0.407050308965, 0.342353999733, 0.295530605237}};
+static const float csf_cr420[8][8] = {
+    {2.03871978502, 2.62502345193, 1.26180942886, 1.11019789803, 1.01397751469,
+     0.867069376285, 0.721500455585, 0.593906509971},
+    {2.62502345193, 1.69112867013, 1.17180569821, 1.3342742857, 1.28513006198,
+     1.13381474809, 0.962064122248, 0.802254508198},
+    {1.26180942886, 1.17180569821, 0.944981930573, 0.990876405848,
+     0.995903384143, 0.926972725286, 0.820534991409, 0.706020324706},
+    {1.11019789803, 1.3342742857, 0.990876405848, 0.831632933426, 0.77418706195,
+     0.725539939514, 0.661776842059, 0.587716619023},
+    {1.01397751469, 1.28513006198, 0.995903384143, 0.77418706195,
+     0.653238524286, 0.584635025748, 0.531064164893, 0.478717061273},
+    {0.867069376285, 1.13381474809, 0.926972725286, 0.725539939514,
+     0.584635025748, 0.496936637883, 0.438694579826, 0.393021669543},
+    {0.721500455585, 0.962064122248, 0.820534991409, 0.661776842059,
+     0.531064164893, 0.438694579826, 0.375820256136, 0.330555063063},
+    {0.593906509971, 0.802254508198, 0.706020324706, 0.587716619023,
+     0.478717061273, 0.393021669543, 0.330555063063, 0.285345396658}};
+
+static double convert_score_db(double _score, double _weight) {
+  return 10 * (log10(255 * 255) - log10(_weight * _score));
+}
+
+static double calc_psnrhvs(const unsigned char *_src, int _systride,
+                           const unsigned char *_dst, int _dystride,
+                           double _par, int _w, int _h, int _step,
+                           const float _csf[8][8]) {
+  float ret;
+  int16_t dct_s[8 * 8], dct_d[8 * 8];
+  tran_low_t dct_s_coef[8 * 8], dct_d_coef[8 * 8];
+  float mask[8][8];
+  int pixels;
+  int x;
+  int y;
+  (void) _par;
+  ret = pixels = 0;
+  /*In the PSNR-HVS-M paper[1] the authors describe the construction of
+   their masking table as "we have used the quantization table for the
+   color component Y of JPEG [6] that has been also obtained on the
+   basis of CSF. Note that the values in quantization table JPEG have
+   been normalized and then squared." Their CSF matrix (from PSNR-HVS)
+   was also constructed from the JPEG matrices. I can not find any obvious
+   scheme of normalizing to produce their table, but if I multiply their
+   CSF by 0.38857 and square the result I get their masking table.
+   I have no idea where this constant comes from, but deviating from it
+   too greatly hurts MOS agreement.
+
+   [1] Nikolay Ponomarenko, Flavia Silvestri, Karen Egiazarian, Marco Carli,
+   Jaakko Astola, Vladimir Lukin, "On between-coefficient contrast masking
+   of DCT basis functions", CD-ROM Proceedings of the Third
+   International Workshop on Video Processing and Quality Metrics for Consumer
+   Electronics VPQM-07, Scottsdale, Arizona, USA, 25-26 January, 2007, 4 p.*/
+  for (x = 0; x < 8; x++)
+    for (y = 0; y < 8; y++)
+      mask[x][y] = (_csf[x][y] * 0.3885746225901003)
+          * (_csf[x][y] * 0.3885746225901003);
+  for (y = 0; y < _h - 7; y += _step) {
+    for (x = 0; x < _w - 7; x += _step) {
+      int i;
+      int j;
+      float s_means[4];
+      float d_means[4];
+      float s_vars[4];
+      float d_vars[4];
+      float s_gmean = 0;
+      float d_gmean = 0;
+      float s_gvar = 0;
+      float d_gvar = 0;
+      float s_mask = 0;
+      float d_mask = 0;
+      for (i = 0; i < 4; i++)
+        s_means[i] = d_means[i] = s_vars[i] = d_vars[i] = 0;
+      for (i = 0; i < 8; i++) {
+        for (j = 0; j < 8; j++) {
+          int sub = ((i & 12) >> 2) + ((j & 12) >> 1);
+          dct_s[i * 8 + j] = _src[(y + i) * _systride + (j + x)];
+          dct_d[i * 8 + j] = _dst[(y + i) * _dystride + (j + x)];
+          s_gmean += dct_s[i * 8 + j];
+          d_gmean += dct_d[i * 8 + j];
+          s_means[sub] += dct_s[i * 8 + j];
+          d_means[sub] += dct_d[i * 8 + j];
+        }
+      }
+      s_gmean /= 64.f;
+      d_gmean /= 64.f;
+      for (i = 0; i < 4; i++)
+        s_means[i] /= 16.f;
+      for (i = 0; i < 4; i++)
+        d_means[i] /= 16.f;
+      for (i = 0; i < 8; i++) {
+        for (j = 0; j < 8; j++) {
+          int sub = ((i & 12) >> 2) + ((j & 12) >> 1);
+          s_gvar += (dct_s[i * 8 + j] - s_gmean) * (dct_s[i * 8 + j] - s_gmean);
+          d_gvar += (dct_d[i * 8 + j] - d_gmean) * (dct_d[i * 8 + j] - d_gmean);
+          s_vars[sub] += (dct_s[i * 8 + j] - s_means[sub])
+              * (dct_s[i * 8 + j] - s_means[sub]);
+          d_vars[sub] += (dct_d[i * 8 + j] - d_means[sub])
+              * (dct_d[i * 8 + j] - d_means[sub]);
+        }
+      }
+      s_gvar *= 1 / 63.f * 64;
+      d_gvar *= 1 / 63.f * 64;
+      for (i = 0; i < 4; i++)
+        s_vars[i] *= 1 / 15.f * 16;
+      for (i = 0; i < 4; i++)
+        d_vars[i] *= 1 / 15.f * 16;
+      if (s_gvar > 0)
+        s_gvar = (s_vars[0] + s_vars[1] + s_vars[2] + s_vars[3]) / s_gvar;
+      if (d_gvar > 0)
+        d_gvar = (d_vars[0] + d_vars[1] + d_vars[2] + d_vars[3]) / d_gvar;
+      od_bin_fdct8x8(dct_s_coef, 8, dct_s, 8);
+      od_bin_fdct8x8(dct_d_coef, 8, dct_d, 8);
+      for (i = 0; i < 8; i++)
+        for (j = (i == 0); j < 8; j++)
+          s_mask += dct_s_coef[i * 8 + j] * dct_s_coef[i * 8 + j] * mask[i][j];
+      for (i = 0; i < 8; i++)
+        for (j = (i == 0); j < 8; j++)
+          d_mask += dct_d_coef[i * 8 + j] * dct_d_coef[i * 8 + j] * mask[i][j];
+      s_mask = sqrt(s_mask * s_gvar) / 32.f;
+      d_mask = sqrt(d_mask * d_gvar) / 32.f;
+      if (d_mask > s_mask)
+        s_mask = d_mask;
+      for (i = 0; i < 8; i++) {
+        for (j = 0; j < 8; j++) {
+          float err;
+          err = fabs((float)(dct_s_coef[i * 8 + j] - dct_d_coef[i * 8 + j]));
+          if (i != 0 || j != 0)
+            err = err < s_mask / mask[i][j] ? 0 : err - s_mask / mask[i][j];
+          ret += (err * _csf[i][j]) * (err * _csf[i][j]);
+          pixels++;
+        }
+      }
+    }
+  }
+  if (pixels <=0)
+      return 0;
+  ret /= pixels;
+  return ret;
+}
+double vpx_psnrhvs(const YV12_BUFFER_CONFIG *source,
+                   const YV12_BUFFER_CONFIG *dest, double *y_psnrhvs,
+                   double *u_psnrhvs, double *v_psnrhvs) {
+  double psnrhvs;
+  const double par = 1.0;
+  const int step = 7;
+  vpx_clear_system_state();
+  *y_psnrhvs = calc_psnrhvs(source->y_buffer, source->y_stride, dest->y_buffer,
+                            dest->y_stride, par, source->y_crop_width,
+                            source->y_crop_height, step, csf_y);
+
+  *u_psnrhvs = calc_psnrhvs(source->u_buffer, source->uv_stride, dest->u_buffer,
+                            dest->uv_stride, par, source->uv_crop_width,
+                            source->uv_crop_height, step, csf_cb420);
+
+  *v_psnrhvs = calc_psnrhvs(source->v_buffer, source->uv_stride, dest->v_buffer,
+                            dest->uv_stride, par, source->uv_crop_width,
+                            source->uv_crop_height, step, csf_cr420);
+  psnrhvs = (*y_psnrhvs) * .8 + .1 * ((*u_psnrhvs) + (*v_psnrhvs));
+
+  return convert_score_db(psnrhvs, 1.0);
+}
diff --git a/libvpx/libvpx/vpx_dsp/quantize.c b/libvpx/libvpx/vpx_dsp/quantize.c
new file mode 100644
index 0000000..80fcd66
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/quantize.c
@@ -0,0 +1,338 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/quantize.h"
+#include "vpx_mem/vpx_mem.h"
+
+void vpx_quantize_dc(const tran_low_t *coeff_ptr,
+                     int n_coeffs, int skip_block,
+                     const int16_t *round_ptr, const int16_t quant,
+                     tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                     const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int tmp, eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+    tmp = (tmp * quant) >> 16;
+    qcoeff_ptr[rc]  = (tmp ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr;
+    if (tmp)
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_quantize_dc(const tran_low_t *coeff_ptr,
+                            int n_coeffs, int skip_block,
+                            const int16_t *round_ptr, const int16_t quant,
+                            tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                            const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  int eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    const int coeff = coeff_ptr[0];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp = abs_coeff + round_ptr[0];
+    const int abs_qcoeff = (int)((tmp * quant) >> 16);
+    qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+    dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr;
+    if (abs_qcoeff)
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+void vpx_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  const int n_coeffs = 1024;
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int tmp, eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
+                INT16_MIN, INT16_MAX);
+    tmp = (tmp * quant) >> 15;
+    qcoeff_ptr[rc]  = (tmp ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr / 2;
+    if (tmp)
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr,
+                                  int skip_block,
+                                  const int16_t *round_ptr,
+                                  const int16_t quant,
+                                  tran_low_t *qcoeff_ptr,
+                                  tran_low_t *dqcoeff_ptr,
+                                  const int16_t dequant_ptr,
+                                  uint16_t *eob_ptr) {
+  const int n_coeffs = 1024;
+  int eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    const int coeff = coeff_ptr[0];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[0], 1);
+    const int abs_qcoeff = (int)((tmp * quant) >> 15);
+    qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+    dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr / 2;
+    if (abs_qcoeff)
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                      int skip_block,
+                      const int16_t *zbin_ptr, const int16_t *round_ptr,
+                      const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
+                      tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                      const int16_t *dequant_ptr,
+                      uint16_t *eob_ptr,
+                      const int16_t *scan, const int16_t *iscan) {
+  int i, non_zero_count = (int)n_coeffs, eob = -1;
+  const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};
+  const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = (int)n_coeffs - 1; i >= 0; i--) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
+        non_zero_count--;
+      else
+        break;
+    }
+
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < non_zero_count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff >= zbins[rc != 0]) {
+        int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+        tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+                  quant_shift_ptr[rc != 0]) >> 16;  // quantization
+        qcoeff_ptr[rc]  = (tmp ^ coeff_sign) - coeff_sign;
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+
+        if (tmp)
+          eob = i;
+      }
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t *zbin_ptr,
+                             const int16_t *round_ptr, const int16_t *quant_ptr,
+                             const int16_t *quant_shift_ptr,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             const int16_t *dequant_ptr,
+                             uint16_t *eob_ptr, const int16_t *scan,
+                             const int16_t *iscan) {
+  int i, non_zero_count = (int)n_coeffs, eob = -1;
+  const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};
+  const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = (int)n_coeffs - 1; i >= 0; i--) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
+        non_zero_count--;
+      else
+        break;
+    }
+
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < non_zero_count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff >= zbins[rc != 0]) {
+        const int64_t tmp1 = abs_coeff + round_ptr[rc != 0];
+        const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+        const uint32_t abs_qcoeff =
+            (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 16);
+        qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+        if (abs_qcoeff)
+          eob = i;
+      }
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            int skip_block,
+                            const int16_t *zbin_ptr, const int16_t *round_ptr,
+                            const int16_t *quant_ptr,
+                            const int16_t *quant_shift_ptr,
+                            tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                            const int16_t *dequant_ptr,
+                            uint16_t *eob_ptr,
+                            const int16_t *scan, const int16_t *iscan) {
+  const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
+                        ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
+  const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
+
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      // If the coefficient is out of the base ZBIN range, keep it for
+      // quantization.
+      if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
+        idx_arr[idx++] = i;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = scan[idx_arr[i]];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      int tmp;
+      int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+      abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+      tmp = ((((abs_coeff * quant_ptr[rc != 0]) >> 16) + abs_coeff) *
+               quant_shift_ptr[rc != 0]) >> 15;
+
+      qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+
+      if (tmp)
+        eob = idx_arr[i];
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr,
+                                   intptr_t n_coeffs, int skip_block,
+                                   const int16_t *zbin_ptr,
+                                   const int16_t *round_ptr,
+                                   const int16_t *quant_ptr,
+                                   const int16_t *quant_shift_ptr,
+                                   tran_low_t *qcoeff_ptr,
+                                   tran_low_t *dqcoeff_ptr,
+                                   const int16_t *dequant_ptr,
+                                   uint16_t *eob_ptr,
+                                   const int16_t *scan, const int16_t *iscan) {
+  const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
+                        ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
+  const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
+
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      // If the coefficient is out of the base ZBIN range, keep it for
+      // quantization.
+      if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
+        idx_arr[idx++] = i;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = scan[idx_arr[i]];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int64_t tmp1 = abs_coeff
+                         + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+      const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+      const uint32_t abs_qcoeff =
+          (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
+      qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+      if (abs_qcoeff)
+        eob = idx_arr[i];
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/quantize.h b/libvpx/libvpx/vpx_dsp/quantize.h
new file mode 100644
index 0000000..89ec597
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/quantize.h
@@ -0,0 +1,51 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_QUANTIZE_H_
+#define VPX_DSP_QUANTIZE_H_
+
+#include "./vpx_config.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vpx_quantize_dc(const tran_low_t *coeff_ptr,
+                     int n_coeffs, int skip_block,
+                     const int16_t *round_ptr, const int16_t quant_ptr,
+                     tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                     const int16_t dequant_ptr, uint16_t *eob_ptr);
+void vpx_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant_ptr,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_quantize_dc(const tran_low_t *coeff_ptr,
+                            int n_coeffs, int skip_block,
+                            const int16_t *round_ptr, const int16_t quant_ptr,
+                            tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                            const int16_t dequant_ptr, uint16_t *eob_ptr);
+void vpx_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr,
+                                  int skip_block,
+                                  const int16_t *round_ptr,
+                                  const int16_t quant_ptr,
+                                  tran_low_t *qcoeff_ptr,
+                                  tran_low_t *dqcoeff_ptr,
+                                  const int16_t dequant_ptr,
+                                  uint16_t *eob_ptr);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_QUANTIZE_H_
diff --git a/libvpx/libvpx/vpx_dsp/sad.c b/libvpx/libvpx/vpx_dsp/sad.c
new file mode 100644
index 0000000..f1f951f
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/sad.c
@@ -0,0 +1,277 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+/* Sum the difference between every corresponding element of the buffers. */
+static INLINE unsigned int sad(const uint8_t *a, int a_stride,
+                               const uint8_t *b, int b_stride,
+                               int width, int height) {
+  int y, x;
+  unsigned int sad = 0;
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x++)
+      sad += abs(a[x] - b[x]);
+
+    a += a_stride;
+    b += b_stride;
+  }
+  return sad;
+}
+
+#define sadMxN(m, n) \
+unsigned int vpx_sad##m##x##n##_c(const uint8_t *src, int src_stride, \
+                                  const uint8_t *ref, int ref_stride) { \
+  return sad(src, src_stride, ref, ref_stride, m, n); \
+} \
+unsigned int vpx_sad##m##x##n##_avg_c(const uint8_t *src, int src_stride, \
+                                      const uint8_t *ref, int ref_stride, \
+                                      const uint8_t *second_pred) { \
+  uint8_t comp_pred[m * n]; \
+  vpx_comp_avg_pred_c(comp_pred, second_pred, m, n, ref, ref_stride); \
+  return sad(src, src_stride, comp_pred, m, m, n); \
+}
+
+// depending on call sites, pass **ref_array to avoid & in subsequent call and
+// de-dup with 4D below.
+#define sadMxNxK(m, n, k) \
+void vpx_sad##m##x##n##x##k##_c(const uint8_t *src, int src_stride, \
+                                const uint8_t *ref_array, int ref_stride, \
+                                uint32_t *sad_array) { \
+  int i; \
+  for (i = 0; i < k; ++i) \
+    sad_array[i] = vpx_sad##m##x##n##_c(src, src_stride, &ref_array[i], ref_stride); \
+}
+
+// This appears to be equivalent to the above when k == 4 and refs is const
+#define sadMxNx4D(m, n) \
+void vpx_sad##m##x##n##x4d_c(const uint8_t *src, int src_stride, \
+                             const uint8_t *const ref_array[], int ref_stride, \
+                             uint32_t *sad_array) { \
+  int i; \
+  for (i = 0; i < 4; ++i) \
+    sad_array[i] = vpx_sad##m##x##n##_c(src, src_stride, ref_array[i], ref_stride); \
+}
+
+// 64x64
+sadMxN(64, 64)
+sadMxNxK(64, 64, 3)
+sadMxNxK(64, 64, 8)
+sadMxNx4D(64, 64)
+
+// 64x32
+sadMxN(64, 32)
+sadMxNx4D(64, 32)
+
+// 32x64
+sadMxN(32, 64)
+sadMxNx4D(32, 64)
+
+// 32x32
+sadMxN(32, 32)
+sadMxNxK(32, 32, 3)
+sadMxNxK(32, 32, 8)
+sadMxNx4D(32, 32)
+
+// 32x16
+sadMxN(32, 16)
+sadMxNx4D(32, 16)
+
+// 16x32
+sadMxN(16, 32)
+sadMxNx4D(16, 32)
+
+// 16x16
+sadMxN(16, 16)
+sadMxNxK(16, 16, 3)
+sadMxNxK(16, 16, 8)
+sadMxNx4D(16, 16)
+
+// 16x8
+sadMxN(16, 8)
+sadMxNxK(16, 8, 3)
+sadMxNxK(16, 8, 8)
+sadMxNx4D(16, 8)
+
+// 8x16
+sadMxN(8, 16)
+sadMxNxK(8, 16, 3)
+sadMxNxK(8, 16, 8)
+sadMxNx4D(8, 16)
+
+// 8x8
+sadMxN(8, 8)
+sadMxNxK(8, 8, 3)
+sadMxNxK(8, 8, 8)
+sadMxNx4D(8, 8)
+
+// 8x4
+sadMxN(8, 4)
+sadMxNxK(8, 4, 8)
+sadMxNx4D(8, 4)
+
+// 4x8
+sadMxN(4, 8)
+sadMxNxK(4, 8, 8)
+sadMxNx4D(4, 8)
+
+// 4x4
+sadMxN(4, 4)
+sadMxNxK(4, 4, 3)
+sadMxNxK(4, 4, 8)
+sadMxNx4D(4, 4)
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE unsigned int highbd_sad(const uint8_t *a8, int a_stride,
+                                      const uint8_t *b8, int b_stride,
+                                      int width, int height) {
+  int y, x;
+  unsigned int sad = 0;
+  const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  const uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x++)
+      sad += abs(a[x] - b[x]);
+
+    a += a_stride;
+    b += b_stride;
+  }
+  return sad;
+}
+
+static INLINE unsigned int highbd_sadb(const uint8_t *a8, int a_stride,
+                                       const uint16_t *b, int b_stride,
+                                       int width, int height) {
+  int y, x;
+  unsigned int sad = 0;
+  const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x++)
+      sad += abs(a[x] - b[x]);
+
+    a += a_stride;
+    b += b_stride;
+  }
+  return sad;
+}
+
+#define highbd_sadMxN(m, n) \
+unsigned int vpx_highbd_sad##m##x##n##_c(const uint8_t *src, int src_stride, \
+                                         const uint8_t *ref, int ref_stride) { \
+  return highbd_sad(src, src_stride, ref, ref_stride, m, n); \
+} \
+unsigned int vpx_highbd_sad##m##x##n##_avg_c(const uint8_t *src, \
+                                             int src_stride, \
+                                             const uint8_t *ref, \
+                                             int ref_stride, \
+                                             const uint8_t *second_pred) { \
+  uint16_t comp_pred[m * n]; \
+  vpx_highbd_comp_avg_pred_c(comp_pred, second_pred, m, n, ref, ref_stride); \
+  return highbd_sadb(src, src_stride, comp_pred, m, m, n); \
+}
+
+#define highbd_sadMxNxK(m, n, k) \
+void vpx_highbd_sad##m##x##n##x##k##_c(const uint8_t *src, int src_stride, \
+                                       const uint8_t *ref_array, int ref_stride, \
+                                       uint32_t *sad_array) { \
+  int i; \
+  for (i = 0; i < k; ++i) { \
+    sad_array[i] = vpx_highbd_sad##m##x##n##_c(src, src_stride, &ref_array[i], \
+                                               ref_stride); \
+  } \
+}
+
+#define highbd_sadMxNx4D(m, n) \
+void vpx_highbd_sad##m##x##n##x4d_c(const uint8_t *src, int src_stride, \
+                                    const uint8_t *const ref_array[], \
+                                    int ref_stride, uint32_t *sad_array) { \
+  int i; \
+  for (i = 0; i < 4; ++i) { \
+    sad_array[i] = vpx_highbd_sad##m##x##n##_c(src, src_stride, ref_array[i], \
+                                               ref_stride); \
+  } \
+}
+
+// 64x64
+highbd_sadMxN(64, 64)
+highbd_sadMxNxK(64, 64, 3)
+highbd_sadMxNxK(64, 64, 8)
+highbd_sadMxNx4D(64, 64)
+
+// 64x32
+highbd_sadMxN(64, 32)
+highbd_sadMxNx4D(64, 32)
+
+// 32x64
+highbd_sadMxN(32, 64)
+highbd_sadMxNx4D(32, 64)
+
+// 32x32
+highbd_sadMxN(32, 32)
+highbd_sadMxNxK(32, 32, 3)
+highbd_sadMxNxK(32, 32, 8)
+highbd_sadMxNx4D(32, 32)
+
+// 32x16
+highbd_sadMxN(32, 16)
+highbd_sadMxNx4D(32, 16)
+
+// 16x32
+highbd_sadMxN(16, 32)
+highbd_sadMxNx4D(16, 32)
+
+// 16x16
+highbd_sadMxN(16, 16)
+highbd_sadMxNxK(16, 16, 3)
+highbd_sadMxNxK(16, 16, 8)
+highbd_sadMxNx4D(16, 16)
+
+// 16x8
+highbd_sadMxN(16, 8)
+highbd_sadMxNxK(16, 8, 3)
+highbd_sadMxNxK(16, 8, 8)
+highbd_sadMxNx4D(16, 8)
+
+// 8x16
+highbd_sadMxN(8, 16)
+highbd_sadMxNxK(8, 16, 3)
+highbd_sadMxNxK(8, 16, 8)
+highbd_sadMxNx4D(8, 16)
+
+// 8x8
+highbd_sadMxN(8, 8)
+highbd_sadMxNxK(8, 8, 3)
+highbd_sadMxNxK(8, 8, 8)
+highbd_sadMxNx4D(8, 8)
+
+// 8x4
+highbd_sadMxN(8, 4)
+highbd_sadMxNxK(8, 4, 8)
+highbd_sadMxNx4D(8, 4)
+
+// 4x8
+highbd_sadMxN(4, 8)
+highbd_sadMxNxK(4, 8, 8)
+highbd_sadMxNx4D(4, 8)
+
+// 4x4
+highbd_sadMxN(4, 4)
+highbd_sadMxNxK(4, 4, 3)
+highbd_sadMxNxK(4, 4, 8)
+highbd_sadMxNx4D(4, 4)
+
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vpx_dsp/ssim.c b/libvpx/libvpx/vpx_dsp/ssim.c
new file mode 100644
index 0000000..cfe5bb3
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/ssim.c
@@ -0,0 +1,505 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/ssim.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/system_state.h"
+
+void vpx_ssim_parms_16x16_c(const uint8_t *s, int sp, const uint8_t *r,
+                            int rp, uint32_t *sum_s, uint32_t *sum_r,
+                            uint32_t *sum_sq_s, uint32_t *sum_sq_r,
+                            uint32_t *sum_sxr) {
+  int i, j;
+  for (i = 0; i < 16; i++, s += sp, r += rp) {
+    for (j = 0; j < 16; j++) {
+      *sum_s += s[j];
+      *sum_r += r[j];
+      *sum_sq_s += s[j] * s[j];
+      *sum_sq_r += r[j] * r[j];
+      *sum_sxr += s[j] * r[j];
+    }
+  }
+}
+void vpx_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp,
+                          uint32_t *sum_s, uint32_t *sum_r,
+                          uint32_t *sum_sq_s, uint32_t *sum_sq_r,
+                          uint32_t *sum_sxr) {
+  int i, j;
+  for (i = 0; i < 8; i++, s += sp, r += rp) {
+    for (j = 0; j < 8; j++) {
+      *sum_s += s[j];
+      *sum_r += r[j];
+      *sum_sq_s += s[j] * s[j];
+      *sum_sq_r += r[j] * r[j];
+      *sum_sxr += s[j] * r[j];
+    }
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp,
+                                 const uint16_t *r, int rp,
+                                 uint32_t *sum_s, uint32_t *sum_r,
+                                 uint32_t *sum_sq_s, uint32_t *sum_sq_r,
+                                 uint32_t *sum_sxr) {
+  int i, j;
+  for (i = 0; i < 8; i++, s += sp, r += rp) {
+    for (j = 0; j < 8; j++) {
+      *sum_s += s[j];
+      *sum_r += r[j];
+      *sum_sq_s += s[j] * s[j];
+      *sum_sq_r += r[j] * r[j];
+      *sum_sxr += s[j] * r[j];
+    }
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static const int64_t cc1 =  26634;  // (64^2*(.01*255)^2
+static const int64_t cc2 = 239708;  // (64^2*(.03*255)^2
+
+static double similarity(uint32_t sum_s, uint32_t sum_r,
+                         uint32_t sum_sq_s, uint32_t sum_sq_r,
+                         uint32_t sum_sxr, int count) {
+  int64_t ssim_n, ssim_d;
+  int64_t c1, c2;
+
+  // scale the constants by number of pixels
+  c1 = (cc1 * count * count) >> 12;
+  c2 = (cc2 * count * count) >> 12;
+
+  ssim_n = (2 * sum_s * sum_r + c1) * ((int64_t) 2 * count * sum_sxr -
+                                       (int64_t) 2 * sum_s * sum_r + c2);
+
+  ssim_d = (sum_s * sum_s + sum_r * sum_r + c1) *
+           ((int64_t)count * sum_sq_s - (int64_t)sum_s * sum_s +
+            (int64_t)count * sum_sq_r - (int64_t) sum_r * sum_r + c2);
+
+  return ssim_n * 1.0 / ssim_d;
+}
+
+static double ssim_8x8(const uint8_t *s, int sp, const uint8_t *r, int rp) {
+  uint32_t sum_s = 0, sum_r = 0, sum_sq_s = 0, sum_sq_r = 0, sum_sxr = 0;
+  vpx_ssim_parms_8x8(s, sp, r, rp, &sum_s, &sum_r, &sum_sq_s, &sum_sq_r,
+                     &sum_sxr);
+  return similarity(sum_s, sum_r, sum_sq_s, sum_sq_r, sum_sxr, 64);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static double highbd_ssim_8x8(const uint16_t *s, int sp, const uint16_t *r,
+                              int rp, unsigned int bd) {
+  uint32_t sum_s = 0, sum_r = 0, sum_sq_s = 0, sum_sq_r = 0, sum_sxr = 0;
+  const int oshift = bd - 8;
+  vpx_highbd_ssim_parms_8x8(s, sp, r, rp, &sum_s, &sum_r, &sum_sq_s, &sum_sq_r,
+                            &sum_sxr);
+  return similarity(sum_s >> oshift,
+                    sum_r >> oshift,
+                    sum_sq_s >> (2 * oshift),
+                    sum_sq_r >> (2 * oshift),
+                    sum_sxr >> (2 * oshift),
+                    64);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+// We are using a 8x8 moving window with starting location of each 8x8 window
+// on the 4x4 pixel grid. Such arrangement allows the windows to overlap
+// block boundaries to penalize blocking artifacts.
+static double vpx_ssim2(const uint8_t *img1, const uint8_t *img2,
+                        int stride_img1, int stride_img2, int width,
+                        int height) {
+  int i, j;
+  int samples = 0;
+  double ssim_total = 0;
+
+  // sample point start with each 4x4 location
+  for (i = 0; i <= height - 8;
+       i += 4, img1 += stride_img1 * 4, img2 += stride_img2 * 4) {
+    for (j = 0; j <= width - 8; j += 4) {
+      double v = ssim_8x8(img1 + j, stride_img1, img2 + j, stride_img2);
+      ssim_total += v;
+      samples++;
+    }
+  }
+  ssim_total /= samples;
+  return ssim_total;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static double vpx_highbd_ssim2(const uint8_t *img1, const uint8_t *img2,
+                               int stride_img1, int stride_img2, int width,
+                               int height, unsigned int bd) {
+  int i, j;
+  int samples = 0;
+  double ssim_total = 0;
+
+  // sample point start with each 4x4 location
+  for (i = 0; i <= height - 8;
+       i += 4, img1 += stride_img1 * 4, img2 += stride_img2 * 4) {
+    for (j = 0; j <= width - 8; j += 4) {
+      double v = highbd_ssim_8x8(CONVERT_TO_SHORTPTR(img1 + j), stride_img1,
+                                 CONVERT_TO_SHORTPTR(img2 + j), stride_img2,
+                                 bd);
+      ssim_total += v;
+      samples++;
+    }
+  }
+  ssim_total /= samples;
+  return ssim_total;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+double vpx_calc_ssim(const YV12_BUFFER_CONFIG *source,
+                     const YV12_BUFFER_CONFIG *dest,
+                     double *weight) {
+  double a, b, c;
+  double ssimv;
+
+  a = vpx_ssim2(source->y_buffer, dest->y_buffer,
+                source->y_stride, dest->y_stride,
+                source->y_crop_width, source->y_crop_height);
+
+  b = vpx_ssim2(source->u_buffer, dest->u_buffer,
+                source->uv_stride, dest->uv_stride,
+                source->uv_crop_width, source->uv_crop_height);
+
+  c = vpx_ssim2(source->v_buffer, dest->v_buffer,
+                source->uv_stride, dest->uv_stride,
+                source->uv_crop_width, source->uv_crop_height);
+
+  ssimv = a * .8 + .1 * (b + c);
+
+  *weight = 1;
+
+  return ssimv;
+}
+
+double vpx_calc_ssimg(const YV12_BUFFER_CONFIG *source,
+                      const YV12_BUFFER_CONFIG *dest,
+                      double *ssim_y, double *ssim_u, double *ssim_v) {
+  double ssim_all = 0;
+  double a, b, c;
+
+  a = vpx_ssim2(source->y_buffer, dest->y_buffer,
+                source->y_stride, dest->y_stride,
+                source->y_crop_width, source->y_crop_height);
+
+  b = vpx_ssim2(source->u_buffer, dest->u_buffer,
+                source->uv_stride, dest->uv_stride,
+                source->uv_crop_width, source->uv_crop_height);
+
+  c = vpx_ssim2(source->v_buffer, dest->v_buffer,
+                source->uv_stride, dest->uv_stride,
+                source->uv_crop_width, source->uv_crop_height);
+  *ssim_y = a;
+  *ssim_u = b;
+  *ssim_v = c;
+  ssim_all = (a * 4 + b + c) / 6;
+
+  return ssim_all;
+}
+
+// traditional ssim as per: http://en.wikipedia.org/wiki/Structural_similarity
+//
+// Re working out the math ->
+//
+// ssim(x,y) =  (2*mean(x)*mean(y) + c1)*(2*cov(x,y)+c2) /
+//   ((mean(x)^2+mean(y)^2+c1)*(var(x)+var(y)+c2))
+//
+// mean(x) = sum(x) / n
+//
+// cov(x,y) = (n*sum(xi*yi)-sum(x)*sum(y))/(n*n)
+//
+// var(x) = (n*sum(xi*xi)-sum(xi)*sum(xi))/(n*n)
+//
+// ssim(x,y) =
+//   (2*sum(x)*sum(y)/(n*n) + c1)*(2*(n*sum(xi*yi)-sum(x)*sum(y))/(n*n)+c2) /
+//   (((sum(x)*sum(x)+sum(y)*sum(y))/(n*n) +c1) *
+//    ((n*sum(xi*xi) - sum(xi)*sum(xi))/(n*n)+
+//     (n*sum(yi*yi) - sum(yi)*sum(yi))/(n*n)+c2)))
+//
+// factoring out n*n
+//
+// ssim(x,y) =
+//   (2*sum(x)*sum(y) + n*n*c1)*(2*(n*sum(xi*yi)-sum(x)*sum(y))+n*n*c2) /
+//   (((sum(x)*sum(x)+sum(y)*sum(y)) + n*n*c1) *
+//    (n*sum(xi*xi)-sum(xi)*sum(xi)+n*sum(yi*yi)-sum(yi)*sum(yi)+n*n*c2))
+//
+// Replace c1 with n*n * c1 for the final step that leads to this code:
+// The final step scales by 12 bits so we don't lose precision in the constants.
+
+static double ssimv_similarity(const Ssimv *sv, int64_t n) {
+  // Scale the constants by number of pixels.
+  const int64_t c1 = (cc1 * n * n) >> 12;
+  const int64_t c2 = (cc2 * n * n) >> 12;
+
+  const double l = 1.0 * (2 * sv->sum_s * sv->sum_r + c1) /
+      (sv->sum_s * sv->sum_s + sv->sum_r * sv->sum_r + c1);
+
+  // Since these variables are unsigned sums, convert to double so
+  // math is done in double arithmetic.
+  const double v = (2.0 * n * sv->sum_sxr - 2 * sv->sum_s * sv->sum_r + c2)
+      / (n * sv->sum_sq_s - sv->sum_s * sv->sum_s + n * sv->sum_sq_r
+         - sv->sum_r * sv->sum_r + c2);
+
+  return l * v;
+}
+
+// The first term of the ssim metric is a luminance factor.
+//
+// (2*mean(x)*mean(y) + c1)/ (mean(x)^2+mean(y)^2+c1)
+//
+// This luminance factor is super sensitive to the dark side of luminance
+// values and completely insensitive on the white side.  check out 2 sets
+// (1,3) and (250,252) the term gives ( 2*1*3/(1+9) = .60
+// 2*250*252/ (250^2+252^2) => .99999997
+//
+// As a result in this tweaked version of the calculation in which the
+// luminance is taken as percentage off from peak possible.
+//
+// 255 * 255 - (sum_s - sum_r) / count * (sum_s - sum_r) / count
+//
+static double ssimv_similarity2(const Ssimv *sv, int64_t n) {
+  // Scale the constants by number of pixels.
+  const int64_t c1 = (cc1 * n * n) >> 12;
+  const int64_t c2 = (cc2 * n * n) >> 12;
+
+  const double mean_diff = (1.0 * sv->sum_s - sv->sum_r) / n;
+  const double l = (255 * 255 - mean_diff * mean_diff + c1) / (255 * 255 + c1);
+
+  // Since these variables are unsigned, sums convert to double so
+  // math is done in double arithmetic.
+  const double v = (2.0 * n * sv->sum_sxr - 2 * sv->sum_s * sv->sum_r + c2)
+      / (n * sv->sum_sq_s - sv->sum_s * sv->sum_s +
+         n * sv->sum_sq_r - sv->sum_r * sv->sum_r + c2);
+
+  return l * v;
+}
+static void ssimv_parms(uint8_t *img1, int img1_pitch, uint8_t *img2,
+                        int img2_pitch, Ssimv *sv) {
+  vpx_ssim_parms_8x8(img1, img1_pitch, img2, img2_pitch,
+                     &sv->sum_s, &sv->sum_r, &sv->sum_sq_s, &sv->sum_sq_r,
+                     &sv->sum_sxr);
+}
+
+double vpx_get_ssim_metrics(uint8_t *img1, int img1_pitch,
+                            uint8_t *img2, int img2_pitch,
+                            int width, int height,
+                            Ssimv *sv2, Metrics *m,
+                            int do_inconsistency) {
+  double dssim_total = 0;
+  double ssim_total = 0;
+  double ssim2_total = 0;
+  double inconsistency_total = 0;
+  int i, j;
+  int c = 0;
+  double norm;
+  double old_ssim_total = 0;
+  vpx_clear_system_state();
+  // We can sample points as frequently as we like start with 1 per 4x4.
+  for (i = 0; i < height; i += 4,
+       img1 += img1_pitch * 4, img2 += img2_pitch * 4) {
+    for (j = 0; j < width; j += 4, ++c) {
+      Ssimv sv = {0};
+      double ssim;
+      double ssim2;
+      double dssim;
+      uint32_t var_new;
+      uint32_t var_old;
+      uint32_t mean_new;
+      uint32_t mean_old;
+      double ssim_new;
+      double ssim_old;
+
+      // Not sure there's a great way to handle the edge pixels
+      // in ssim when using a window. Seems biased against edge pixels
+      // however you handle this. This uses only samples that are
+      // fully in the frame.
+      if (j + 8 <= width && i + 8 <= height) {
+        ssimv_parms(img1 + j, img1_pitch, img2 + j, img2_pitch, &sv);
+      }
+
+      ssim = ssimv_similarity(&sv, 64);
+      ssim2 = ssimv_similarity2(&sv, 64);
+
+      sv.ssim = ssim2;
+
+      // dssim is calculated to use as an actual error metric and
+      // is scaled up to the same range as sum square error.
+      // Since we are subsampling every 16th point maybe this should be
+      // *16 ?
+      dssim = 255 * 255 * (1 - ssim2) / 2;
+
+      // Here I introduce a new error metric: consistency-weighted
+      // SSIM-inconsistency.  This metric isolates frames where the
+      // SSIM 'suddenly' changes, e.g. if one frame in every 8 is much
+      // sharper or blurrier than the others. Higher values indicate a
+      // temporally inconsistent SSIM. There are two ideas at work:
+      //
+      // 1) 'SSIM-inconsistency': the total inconsistency value
+      // reflects how much SSIM values are changing between this
+      // source / reference frame pair and the previous pair.
+      //
+      // 2) 'consistency-weighted': weights de-emphasize areas in the
+      // frame where the scene content has changed. Changes in scene
+      // content are detected via changes in local variance and local
+      // mean.
+      //
+      // Thus the overall measure reflects how inconsistent the SSIM
+      // values are, over consistent regions of the frame.
+      //
+      // The metric has three terms:
+      //
+      // term 1 -> uses change in scene Variance to weight error score
+      //  2 * var(Fi)*var(Fi-1) / (var(Fi)^2+var(Fi-1)^2)
+      //  larger changes from one frame to the next mean we care
+      //  less about consistency.
+      //
+      // term 2 -> uses change in local scene luminance to weight error
+      //  2 * avg(Fi)*avg(Fi-1) / (avg(Fi)^2+avg(Fi-1)^2)
+      //  larger changes from one frame to the next mean we care
+      //  less about consistency.
+      //
+      // term3 -> measures inconsistency in ssim scores between frames
+      //   1 - ( 2 * ssim(Fi)*ssim(Fi-1)/(ssim(Fi)^2+sssim(Fi-1)^2).
+      //
+      // This term compares the ssim score for the same location in 2
+      // subsequent frames.
+      var_new = sv.sum_sq_s - sv.sum_s * sv.sum_s / 64;
+      var_old = sv2[c].sum_sq_s - sv2[c].sum_s * sv2[c].sum_s / 64;
+      mean_new = sv.sum_s;
+      mean_old = sv2[c].sum_s;
+      ssim_new = sv.ssim;
+      ssim_old = sv2[c].ssim;
+
+      if (do_inconsistency) {
+        // We do the metric once for every 4x4 block in the image. Since
+        // we are scaling the error to SSE for use in a psnr calculation
+        // 1.0 = 4x4x255x255 the worst error we can possibly have.
+        static const double kScaling = 4. * 4 * 255 * 255;
+
+        // The constants have to be non 0 to avoid potential divide by 0
+        // issues other than that they affect kind of a weighting between
+        // the terms.  No testing of what the right terms should be has been
+        // done.
+        static const double c1 = 1, c2 = 1, c3 = 1;
+
+        // This measures how much consistent variance is in two consecutive
+        // source frames. 1.0 means they have exactly the same variance.
+        const double variance_term = (2.0 * var_old * var_new + c1) /
+            (1.0 * var_old * var_old + 1.0 * var_new * var_new + c1);
+
+        // This measures how consistent the local mean are between two
+        // consecutive frames. 1.0 means they have exactly the same mean.
+        const double mean_term = (2.0 * mean_old * mean_new + c2) /
+            (1.0 * mean_old * mean_old + 1.0 * mean_new * mean_new + c2);
+
+        // This measures how consistent the ssims of two
+        // consecutive frames is. 1.0 means they are exactly the same.
+        double ssim_term = pow((2.0 * ssim_old * ssim_new + c3) /
+                               (ssim_old * ssim_old + ssim_new * ssim_new + c3),
+                               5);
+
+        double this_inconsistency;
+
+        // Floating point math sometimes makes this > 1 by a tiny bit.
+        // We want the metric to scale between 0 and 1.0 so we can convert
+        // it to an snr scaled value.
+        if (ssim_term > 1)
+          ssim_term = 1;
+
+        // This converts the consistency metric to an inconsistency metric
+        // ( so we can scale it like psnr to something like sum square error.
+        // The reason for the variance and mean terms is the assumption that
+        // if there are big changes in the source we shouldn't penalize
+        // inconsistency in ssim scores a bit less as it will be less visible
+        // to the user.
+        this_inconsistency = (1 - ssim_term) * variance_term * mean_term;
+
+        this_inconsistency *= kScaling;
+        inconsistency_total += this_inconsistency;
+      }
+      sv2[c] = sv;
+      ssim_total += ssim;
+      ssim2_total += ssim2;
+      dssim_total += dssim;
+
+      old_ssim_total += ssim_old;
+    }
+    old_ssim_total += 0;
+  }
+
+  norm = 1. / (width / 4) / (height / 4);
+  ssim_total *= norm;
+  ssim2_total *= norm;
+  m->ssim2 = ssim2_total;
+  m->ssim = ssim_total;
+  if (old_ssim_total == 0)
+    inconsistency_total = 0;
+
+  m->ssimc = inconsistency_total;
+
+  m->dssim = dssim_total;
+  return inconsistency_total;
+}
+
+
+#if CONFIG_VP9_HIGHBITDEPTH
+double vpx_highbd_calc_ssim(const YV12_BUFFER_CONFIG *source,
+                            const YV12_BUFFER_CONFIG *dest,
+                            double *weight, unsigned int bd) {
+  double a, b, c;
+  double ssimv;
+
+  a = vpx_highbd_ssim2(source->y_buffer, dest->y_buffer,
+                       source->y_stride, dest->y_stride,
+                       source->y_crop_width, source->y_crop_height, bd);
+
+  b = vpx_highbd_ssim2(source->u_buffer, dest->u_buffer,
+                       source->uv_stride, dest->uv_stride,
+                       source->uv_crop_width, source->uv_crop_height, bd);
+
+  c = vpx_highbd_ssim2(source->v_buffer, dest->v_buffer,
+                       source->uv_stride, dest->uv_stride,
+                       source->uv_crop_width, source->uv_crop_height, bd);
+
+  ssimv = a * .8 + .1 * (b + c);
+
+  *weight = 1;
+
+  return ssimv;
+}
+
+double vpx_highbd_calc_ssimg(const YV12_BUFFER_CONFIG *source,
+                             const YV12_BUFFER_CONFIG *dest, double *ssim_y,
+                             double *ssim_u, double *ssim_v, unsigned int bd) {
+  double ssim_all = 0;
+  double a, b, c;
+
+  a = vpx_highbd_ssim2(source->y_buffer, dest->y_buffer,
+                       source->y_stride, dest->y_stride,
+                       source->y_crop_width, source->y_crop_height, bd);
+
+  b = vpx_highbd_ssim2(source->u_buffer, dest->u_buffer,
+                       source->uv_stride, dest->uv_stride,
+                       source->uv_crop_width, source->uv_crop_height, bd);
+
+  c = vpx_highbd_ssim2(source->v_buffer, dest->v_buffer,
+                       source->uv_stride, dest->uv_stride,
+                       source->uv_crop_width, source->uv_crop_height, bd);
+  *ssim_y = a;
+  *ssim_u = b;
+  *ssim_v = c;
+  ssim_all = (a * 4 + b + c) / 6;
+
+  return ssim_all;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vpx_dsp/ssim.h b/libvpx/libvpx/vpx_dsp/ssim.h
new file mode 100644
index 0000000..132f7f9
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/ssim.h
@@ -0,0 +1,101 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_SSIM_H_
+#define VPX_DSP_SSIM_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./vpx_config.h"
+#include "vpx_scale/yv12config.h"
+
+// metrics used for calculating ssim, ssim2, dssim, and ssimc
+typedef struct {
+  // source sum ( over 8x8 region )
+  uint32_t sum_s;
+
+  // reference sum (over 8x8 region )
+  uint32_t sum_r;
+
+  // source sum squared ( over 8x8 region )
+  uint32_t sum_sq_s;
+
+  // reference sum squared (over 8x8 region )
+  uint32_t sum_sq_r;
+
+  // sum of source times reference (over 8x8 region)
+  uint32_t sum_sxr;
+
+  // calculated ssim score between source and reference
+  double ssim;
+} Ssimv;
+
+// metrics collected on a frame basis
+typedef struct {
+  // ssim consistency error metric ( see code for explanation )
+  double ssimc;
+
+  // standard ssim
+  double ssim;
+
+  // revised ssim ( see code for explanation)
+  double ssim2;
+
+  // ssim restated as an error metric like sse
+  double dssim;
+
+  // dssim converted to decibels
+  double dssimd;
+
+  // ssimc converted to decibels
+  double ssimcd;
+} Metrics;
+
+double vpx_get_ssim_metrics(uint8_t *img1, int img1_pitch, uint8_t *img2,
+                      int img2_pitch, int width, int height, Ssimv *sv2,
+                      Metrics *m, int do_inconsistency);
+
+double vpx_calc_ssim(const YV12_BUFFER_CONFIG *source,
+                     const YV12_BUFFER_CONFIG *dest,
+                     double *weight);
+
+double vpx_calc_ssimg(const YV12_BUFFER_CONFIG *source,
+                      const YV12_BUFFER_CONFIG *dest,
+                      double *ssim_y, double *ssim_u, double *ssim_v);
+
+double vpx_calc_fastssim(const YV12_BUFFER_CONFIG *source,
+                         const YV12_BUFFER_CONFIG *dest,
+                         double *ssim_y, double *ssim_u, double *ssim_v);
+
+double vpx_psnrhvs(const YV12_BUFFER_CONFIG *source,
+                   const YV12_BUFFER_CONFIG *dest,
+                   double *ssim_y, double *ssim_u, double *ssim_v);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+double vpx_highbd_calc_ssim(const YV12_BUFFER_CONFIG *source,
+                            const YV12_BUFFER_CONFIG *dest,
+                            double *weight,
+                            unsigned int bd);
+
+double vpx_highbd_calc_ssimg(const YV12_BUFFER_CONFIG *source,
+                             const YV12_BUFFER_CONFIG *dest,
+                             double *ssim_y,
+                             double *ssim_u,
+                             double *ssim_v,
+                             unsigned int bd);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_SSIM_H_
diff --git a/libvpx/libvpx/vpx_dsp/subtract.c b/libvpx/libvpx/vpx_dsp/subtract.c
new file mode 100644
index 0000000..556e013
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/subtract.c
@@ -0,0 +1,56 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+void vpx_subtract_block_c(int rows, int cols,
+                          int16_t *diff, ptrdiff_t diff_stride,
+                          const uint8_t *src, ptrdiff_t src_stride,
+                          const uint8_t *pred, ptrdiff_t pred_stride) {
+  int r, c;
+
+  for (r = 0; r < rows; r++) {
+    for (c = 0; c < cols; c++)
+      diff[c] = src[c] - pred[c];
+
+    diff += diff_stride;
+    pred += pred_stride;
+    src  += src_stride;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_subtract_block_c(int rows, int cols,
+                                 int16_t *diff, ptrdiff_t diff_stride,
+                                 const uint8_t *src8, ptrdiff_t src_stride,
+                                 const uint8_t *pred8, ptrdiff_t pred_stride,
+                                 int bd) {
+  int r, c;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  (void) bd;
+
+  for (r = 0; r < rows; r++) {
+    for (c = 0; c < cols; c++) {
+      diff[c] = src[c] - pred[c];
+    }
+
+    diff += diff_stride;
+    pred += pred_stride;
+    src  += src_stride;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vpx_dsp/txfm_common.h b/libvpx/libvpx/vpx_dsp/txfm_common.h
new file mode 100644
index 0000000..442e6a5
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/txfm_common.h
@@ -0,0 +1,66 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_TXFM_COMMON_H_
+#define VPX_DSP_TXFM_COMMON_H_
+
+#include "vpx_dsp/vpx_dsp_common.h"
+
+// Constants and Macros used by all idct/dct functions
+#define DCT_CONST_BITS 14
+#define DCT_CONST_ROUNDING  (1 << (DCT_CONST_BITS - 1))
+
+#define UNIT_QUANT_SHIFT 2
+#define UNIT_QUANT_FACTOR (1 << UNIT_QUANT_SHIFT)
+
+// Constants:
+//  for (int i = 1; i< 32; ++i)
+//    printf("static const int cospi_%d_64 = %.0f;\n", i,
+//           round(16384 * cos(i*M_PI/64)));
+// Note: sin(k*Pi/64) = cos((32-k)*Pi/64)
+static const tran_high_t cospi_1_64  = 16364;
+static const tran_high_t cospi_2_64  = 16305;
+static const tran_high_t cospi_3_64  = 16207;
+static const tran_high_t cospi_4_64  = 16069;
+static const tran_high_t cospi_5_64  = 15893;
+static const tran_high_t cospi_6_64  = 15679;
+static const tran_high_t cospi_7_64  = 15426;
+static const tran_high_t cospi_8_64  = 15137;
+static const tran_high_t cospi_9_64  = 14811;
+static const tran_high_t cospi_10_64 = 14449;
+static const tran_high_t cospi_11_64 = 14053;
+static const tran_high_t cospi_12_64 = 13623;
+static const tran_high_t cospi_13_64 = 13160;
+static const tran_high_t cospi_14_64 = 12665;
+static const tran_high_t cospi_15_64 = 12140;
+static const tran_high_t cospi_16_64 = 11585;
+static const tran_high_t cospi_17_64 = 11003;
+static const tran_high_t cospi_18_64 = 10394;
+static const tran_high_t cospi_19_64 = 9760;
+static const tran_high_t cospi_20_64 = 9102;
+static const tran_high_t cospi_21_64 = 8423;
+static const tran_high_t cospi_22_64 = 7723;
+static const tran_high_t cospi_23_64 = 7005;
+static const tran_high_t cospi_24_64 = 6270;
+static const tran_high_t cospi_25_64 = 5520;
+static const tran_high_t cospi_26_64 = 4756;
+static const tran_high_t cospi_27_64 = 3981;
+static const tran_high_t cospi_28_64 = 3196;
+static const tran_high_t cospi_29_64 = 2404;
+static const tran_high_t cospi_30_64 = 1606;
+static const tran_high_t cospi_31_64 = 804;
+
+//  16384 * sqrt(2) * sin(kPi/9) * 2 / 3
+static const tran_high_t sinpi_1_9 = 5283;
+static const tran_high_t sinpi_2_9 = 9929;
+static const tran_high_t sinpi_3_9 = 13377;
+static const tran_high_t sinpi_4_9 = 15212;
+
+#endif  // VPX_DSP_TXFM_COMMON_H_
diff --git a/libvpx/libvpx/vpx_dsp/variance.c b/libvpx/libvpx/vpx_dsp/variance.c
new file mode 100644
index 0000000..d960c54
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/variance.c
@@ -0,0 +1,625 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_ports/mem.h"
+#include "vpx/vpx_integer.h"
+
+#include "vpx_dsp/variance.h"
+
+static const uint8_t bilinear_filters[8][2] = {
+  { 128,   0  },
+  { 112,  16  },
+  {  96,  32  },
+  {  80,  48  },
+  {  64,  64  },
+  {  48,  80  },
+  {  32,  96  },
+  {  16, 112  },
+};
+
+uint32_t vpx_get4x4sse_cs_c(const uint8_t *a, int  a_stride,
+                            const uint8_t *b, int  b_stride) {
+  int distortion = 0;
+  int r, c;
+
+  for (r = 0; r < 4; ++r) {
+    for (c = 0; c < 4; ++c) {
+      int diff = a[c] - b[c];
+      distortion += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+
+  return distortion;
+}
+
+uint32_t vpx_get_mb_ss_c(const int16_t *a) {
+  unsigned int i, sum = 0;
+
+  for (i = 0; i < 256; ++i) {
+    sum += a[i] * a[i];
+  }
+
+  return sum;
+}
+
+uint32_t vpx_variance_halfpixvar16x16_h_c(const uint8_t *a, int a_stride,
+                                          const uint8_t *b, int b_stride,
+                                          uint32_t *sse) {
+  return vpx_sub_pixel_variance16x16_c(a, a_stride, 4, 0,
+                                       b, b_stride, sse);
+}
+
+
+uint32_t vpx_variance_halfpixvar16x16_v_c(const uint8_t *a, int a_stride,
+                                          const uint8_t *b, int b_stride,
+                                          uint32_t *sse) {
+  return vpx_sub_pixel_variance16x16_c(a, a_stride, 0, 4,
+                                       b, b_stride, sse);
+}
+
+uint32_t vpx_variance_halfpixvar16x16_hv_c(const uint8_t *a, int a_stride,
+                                           const uint8_t *b, int b_stride,
+                                           uint32_t *sse) {
+  return vpx_sub_pixel_variance16x16_c(a, a_stride, 4, 4,
+                                       b, b_stride, sse);
+}
+
+static void variance(const uint8_t *a, int  a_stride,
+                     const uint8_t *b, int  b_stride,
+                     int  w, int  h, uint32_t *sse, int *sum) {
+  int i, j;
+
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+// Applies a 1-D 2-tap bilinear filter to the source block in either horizontal
+// or vertical direction to produce the filtered output block. Used to implement
+// the first-pass of 2-D separable filter.
+//
+// Produces int16_t output to retain precision for the next pass. Two filter
+// taps should sum to FILTER_WEIGHT. pixel_step defines whether the filter is
+// applied horizontally (pixel_step = 1) or vertically (pixel_step = stride).
+// It defines the offset required to move from one input to the next.
+static void var_filter_block2d_bil_first_pass(const uint8_t *a, uint16_t *b,
+                                              unsigned int src_pixels_per_line,
+                                              int pixel_step,
+                                              unsigned int output_height,
+                                              unsigned int output_width,
+                                              const uint8_t *filter) {
+  unsigned int i, j;
+
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      b[j] = ROUND_POWER_OF_TWO((int)a[0] * filter[0] +
+                          (int)a[pixel_step] * filter[1],
+                          FILTER_BITS);
+
+      ++a;
+    }
+
+    a += src_pixels_per_line - output_width;
+    b += output_width;
+  }
+}
+
+// Applies a 1-D 2-tap bilinear filter to the source block in either horizontal
+// or vertical direction to produce the filtered output block. Used to implement
+// the second-pass of 2-D separable filter.
+//
+// Requires 16-bit input as produced by filter_block2d_bil_first_pass. Two
+// filter taps should sum to FILTER_WEIGHT. pixel_step defines whether the
+// filter is applied horizontally (pixel_step = 1) or vertically
+// (pixel_step = stride). It defines the offset required to move from one input
+// to the next. Output is 8-bit.
+static void var_filter_block2d_bil_second_pass(const uint16_t *a, uint8_t *b,
+                                               unsigned int src_pixels_per_line,
+                                               unsigned int pixel_step,
+                                               unsigned int output_height,
+                                               unsigned int output_width,
+                                               const uint8_t *filter) {
+  unsigned int  i, j;
+
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      b[j] = ROUND_POWER_OF_TWO((int)a[0] * filter[0] +
+                          (int)a[pixel_step] * filter[1],
+                          FILTER_BITS);
+      ++a;
+    }
+
+    a += src_pixels_per_line - output_width;
+    b += output_width;
+  }
+}
+
+#define VAR(W, H) \
+uint32_t vpx_variance##W##x##H##_c(const uint8_t *a, int a_stride, \
+                                   const uint8_t *b, int b_stride, \
+                                   uint32_t *sse) { \
+  int sum; \
+  variance(a, a_stride, b, b_stride, W, H, sse, &sum); \
+  return *sse - (((int64_t)sum * sum) / (W * H)); \
+}
+
+#define SUBPIX_VAR(W, H) \
+uint32_t vpx_sub_pixel_variance##W##x##H##_c(const uint8_t *a, int a_stride, \
+                                             int xoffset, int  yoffset, \
+                                             const uint8_t *b, int b_stride, \
+                                             uint32_t *sse) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint8_t temp2[H * W]; \
+\
+  var_filter_block2d_bil_first_pass(a, fdata3, a_stride, 1, H + 1, W, \
+                                    bilinear_filters[xoffset]); \
+  var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                     bilinear_filters[yoffset]); \
+\
+  return vpx_variance##W##x##H##_c(temp2, W, b, b_stride, sse); \
+}
+
+#define SUBPIX_AVG_VAR(W, H) \
+uint32_t vpx_sub_pixel_avg_variance##W##x##H##_c(const uint8_t *a, \
+                                                 int  a_stride, \
+                                                 int xoffset, int  yoffset, \
+                                                 const uint8_t *b, \
+                                                 int b_stride, \
+                                                 uint32_t *sse, \
+                                                 const uint8_t *second_pred) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint8_t temp2[H * W]; \
+  DECLARE_ALIGNED(16, uint8_t, temp3[H * W]); \
+\
+  var_filter_block2d_bil_first_pass(a, fdata3, a_stride, 1, H + 1, W, \
+                                    bilinear_filters[xoffset]); \
+  var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                     bilinear_filters[yoffset]); \
+\
+  vpx_comp_avg_pred(temp3, second_pred, W, H, temp2, W); \
+\
+  return vpx_variance##W##x##H##_c(temp3, W, b, b_stride, sse); \
+}
+
+/* Identical to the variance call except it takes an additional parameter, sum,
+ * and returns that value using pass-by-reference instead of returning
+ * sse - sum^2 / w*h
+ */
+#define GET_VAR(W, H) \
+void vpx_get##W##x##H##var_c(const uint8_t *a, int a_stride, \
+                             const uint8_t *b, int b_stride, \
+                             uint32_t *sse, int *sum) { \
+  variance(a, a_stride, b, b_stride, W, H, sse, sum); \
+}
+
+/* Identical to the variance call except it does not calculate the
+ * sse - sum^2 / w*h and returns sse in addtion to modifying the passed in
+ * variable.
+ */
+#define MSE(W, H) \
+uint32_t vpx_mse##W##x##H##_c(const uint8_t *a, int a_stride, \
+                              const uint8_t *b, int b_stride, \
+                              uint32_t *sse) { \
+  int sum; \
+  variance(a, a_stride, b, b_stride, W, H, sse, &sum); \
+  return *sse; \
+}
+
+/* All three forms of the variance are available in the same sizes. */
+#define VARIANCES(W, H) \
+    VAR(W, H) \
+    SUBPIX_VAR(W, H) \
+    SUBPIX_AVG_VAR(W, H)
+
+VARIANCES(64, 64)
+VARIANCES(64, 32)
+VARIANCES(32, 64)
+VARIANCES(32, 32)
+VARIANCES(32, 16)
+VARIANCES(16, 32)
+VARIANCES(16, 16)
+VARIANCES(16, 8)
+VARIANCES(8, 16)
+VARIANCES(8, 8)
+VARIANCES(8, 4)
+VARIANCES(4, 8)
+VARIANCES(4, 4)
+
+GET_VAR(16, 16)
+GET_VAR(8, 8)
+
+MSE(16, 16)
+MSE(16, 8)
+MSE(8, 16)
+MSE(8, 8)
+
+void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred,
+                         int width, int height,
+                         const uint8_t *ref, int ref_stride) {
+  int i, j;
+
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      const int tmp = pred[j] + ref[j];
+      comp_pred[j] = ROUND_POWER_OF_TWO(tmp, 1);
+    }
+    comp_pred += width;
+    pred += width;
+    ref += ref_stride;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_variance64(const uint8_t *a8, int  a_stride,
+                              const uint8_t *b8, int  b_stride,
+                              int w, int h, uint64_t *sse, int64_t *sum) {
+  int i, j;
+
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+static void highbd_8_variance(const uint8_t *a8, int  a_stride,
+                              const uint8_t *b8, int  b_stride,
+                              int w, int h, uint32_t *sse, int *sum) {
+  uint64_t sse_long = 0;
+  int64_t sum_long = 0;
+  highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
+  *sse = (uint32_t)sse_long;
+  *sum = (int)sum_long;
+}
+
+static void highbd_10_variance(const uint8_t *a8, int  a_stride,
+                               const uint8_t *b8, int  b_stride,
+                               int w, int h, uint32_t *sse, int *sum) {
+  uint64_t sse_long = 0;
+  int64_t sum_long = 0;
+  highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4);
+  *sum = (int)ROUND_POWER_OF_TWO(sum_long, 2);
+}
+
+static void highbd_12_variance(const uint8_t *a8, int  a_stride,
+                               const uint8_t *b8, int  b_stride,
+                               int w, int h, uint32_t *sse, int *sum) {
+  uint64_t sse_long = 0;
+  int64_t sum_long = 0;
+  highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8);
+  *sum = (int)ROUND_POWER_OF_TWO(sum_long, 4);
+}
+
+#define HIGHBD_VAR(W, H) \
+uint32_t vpx_highbd_8_variance##W##x##H##_c(const uint8_t *a, \
+                                            int a_stride, \
+                                            const uint8_t *b, \
+                                            int b_stride, \
+                                            uint32_t *sse) { \
+  int sum; \
+  highbd_8_variance(a, a_stride, b, b_stride, W, H, sse, &sum); \
+  return *sse - (((int64_t)sum * sum) / (W * H)); \
+} \
+\
+uint32_t vpx_highbd_10_variance##W##x##H##_c(const uint8_t *a, \
+                                             int a_stride, \
+                                             const uint8_t *b, \
+                                             int b_stride, \
+                                             uint32_t *sse) { \
+  int sum; \
+  int64_t var; \
+  highbd_10_variance(a, a_stride, b, b_stride, W, H, sse, &sum); \
+  var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \
+  return (var >= 0) ? (uint32_t)var : 0; \
+} \
+\
+uint32_t vpx_highbd_12_variance##W##x##H##_c(const uint8_t *a, \
+                                             int a_stride, \
+                                             const uint8_t *b, \
+                                             int b_stride, \
+                                             uint32_t *sse) { \
+  int sum; \
+  int64_t var; \
+  highbd_12_variance(a, a_stride, b, b_stride, W, H, sse, &sum); \
+  var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \
+  return (var >= 0) ? (uint32_t)var : 0; \
+}
+
+#define HIGHBD_GET_VAR(S) \
+void vpx_highbd_8_get##S##x##S##var_c(const uint8_t *src, int src_stride, \
+                                      const uint8_t *ref, int ref_stride, \
+                                      uint32_t *sse, int *sum) { \
+  highbd_8_variance(src, src_stride, ref, ref_stride, S, S, sse, sum); \
+} \
+\
+void vpx_highbd_10_get##S##x##S##var_c(const uint8_t *src, int src_stride, \
+                                       const uint8_t *ref, int ref_stride, \
+                                       uint32_t *sse, int *sum) { \
+  highbd_10_variance(src, src_stride, ref, ref_stride, S, S, sse, sum); \
+} \
+\
+void vpx_highbd_12_get##S##x##S##var_c(const uint8_t *src, int src_stride, \
+                                       const uint8_t *ref, int ref_stride, \
+                                       uint32_t *sse, int *sum) { \
+  highbd_12_variance(src, src_stride, ref, ref_stride, S, S, sse, sum); \
+}
+
+#define HIGHBD_MSE(W, H) \
+uint32_t vpx_highbd_8_mse##W##x##H##_c(const uint8_t *src, \
+                                       int src_stride, \
+                                       const uint8_t *ref, \
+                                       int ref_stride, \
+                                       uint32_t *sse) { \
+  int sum; \
+  highbd_8_variance(src, src_stride, ref, ref_stride, W, H, sse, &sum); \
+  return *sse; \
+} \
+\
+uint32_t vpx_highbd_10_mse##W##x##H##_c(const uint8_t *src, \
+                                        int src_stride, \
+                                        const uint8_t *ref, \
+                                        int ref_stride, \
+                                        uint32_t *sse) { \
+  int sum; \
+  highbd_10_variance(src, src_stride, ref, ref_stride, W, H, sse, &sum); \
+  return *sse; \
+} \
+\
+uint32_t vpx_highbd_12_mse##W##x##H##_c(const uint8_t *src, \
+                                        int src_stride, \
+                                        const uint8_t *ref, \
+                                        int ref_stride, \
+                                        uint32_t *sse) { \
+  int sum; \
+  highbd_12_variance(src, src_stride, ref, ref_stride, W, H, sse, &sum); \
+  return *sse; \
+}
+
+static void highbd_var_filter_block2d_bil_first_pass(
+    const uint8_t *src_ptr8,
+    uint16_t *output_ptr,
+    unsigned int src_pixels_per_line,
+    int pixel_step,
+    unsigned int output_height,
+    unsigned int output_width,
+    const uint8_t *filter) {
+  unsigned int i, j;
+  uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src_ptr8);
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      output_ptr[j] =
+          ROUND_POWER_OF_TWO((int)src_ptr[0] * filter[0] +
+                             (int)src_ptr[pixel_step] * filter[1],
+                             FILTER_BITS);
+
+      ++src_ptr;
+    }
+
+    // Next row...
+    src_ptr += src_pixels_per_line - output_width;
+    output_ptr += output_width;
+  }
+}
+
+static void highbd_var_filter_block2d_bil_second_pass(
+    const uint16_t *src_ptr,
+    uint16_t *output_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int pixel_step,
+    unsigned int output_height,
+    unsigned int output_width,
+    const uint8_t *filter) {
+  unsigned int  i, j;
+
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      output_ptr[j] =
+          ROUND_POWER_OF_TWO((int)src_ptr[0] * filter[0] +
+                             (int)src_ptr[pixel_step] * filter[1],
+                             FILTER_BITS);
+      ++src_ptr;
+    }
+
+    src_ptr += src_pixels_per_line - output_width;
+    output_ptr += output_width;
+  }
+}
+
+#define HIGHBD_SUBPIX_VAR(W, H) \
+uint32_t vpx_highbd_8_sub_pixel_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  uint32_t *sse) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  return vpx_highbd_8_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), W, dst, \
+                                          dst_stride, sse); \
+} \
+\
+uint32_t vpx_highbd_10_sub_pixel_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  uint32_t *sse) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  return vpx_highbd_10_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), \
+                                             W, dst, dst_stride, sse); \
+} \
+\
+uint32_t vpx_highbd_12_sub_pixel_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  uint32_t *sse) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  return vpx_highbd_12_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), \
+                                             W, dst, dst_stride, sse); \
+}
+
+#define HIGHBD_SUBPIX_AVG_VAR(W, H) \
+uint32_t vpx_highbd_8_sub_pixel_avg_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  uint32_t *sse, \
+  const uint8_t *second_pred) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+  DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  vpx_highbd_comp_avg_pred(temp3, second_pred, W, H, \
+                           CONVERT_TO_BYTEPTR(temp2), W); \
+\
+  return vpx_highbd_8_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, dst, \
+                                          dst_stride, sse); \
+} \
+\
+uint32_t vpx_highbd_10_sub_pixel_avg_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  uint32_t *sse, \
+  const uint8_t *second_pred) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+  DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  vpx_highbd_comp_avg_pred(temp3, second_pred, W, H, \
+                           CONVERT_TO_BYTEPTR(temp2), W); \
+\
+  return vpx_highbd_10_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), \
+                                             W, dst, dst_stride, sse); \
+} \
+\
+uint32_t vpx_highbd_12_sub_pixel_avg_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  uint32_t *sse, \
+  const uint8_t *second_pred) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+  DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  vpx_highbd_comp_avg_pred(temp3, second_pred, W, H, \
+                           CONVERT_TO_BYTEPTR(temp2), W); \
+\
+  return vpx_highbd_12_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), \
+                                             W, dst, dst_stride, sse); \
+}
+
+/* All three forms of the variance are available in the same sizes. */
+#define HIGHBD_VARIANCES(W, H) \
+    HIGHBD_VAR(W, H) \
+    HIGHBD_SUBPIX_VAR(W, H) \
+    HIGHBD_SUBPIX_AVG_VAR(W, H)
+
+HIGHBD_VARIANCES(64, 64)
+HIGHBD_VARIANCES(64, 32)
+HIGHBD_VARIANCES(32, 64)
+HIGHBD_VARIANCES(32, 32)
+HIGHBD_VARIANCES(32, 16)
+HIGHBD_VARIANCES(16, 32)
+HIGHBD_VARIANCES(16, 16)
+HIGHBD_VARIANCES(16, 8)
+HIGHBD_VARIANCES(8, 16)
+HIGHBD_VARIANCES(8, 8)
+HIGHBD_VARIANCES(8, 4)
+HIGHBD_VARIANCES(4, 8)
+HIGHBD_VARIANCES(4, 4)
+
+HIGHBD_GET_VAR(8)
+HIGHBD_GET_VAR(16)
+
+HIGHBD_MSE(16, 16)
+HIGHBD_MSE(16, 8)
+HIGHBD_MSE(8, 16)
+HIGHBD_MSE(8, 8)
+
+void vpx_highbd_comp_avg_pred(uint16_t *comp_pred, const uint8_t *pred8,
+                              int width, int height, const uint8_t *ref8,
+                              int ref_stride) {
+  int i, j;
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      const int tmp = pred[j] + ref[j];
+      comp_pred[j] = ROUND_POWER_OF_TWO(tmp, 1);
+    }
+    comp_pred += width;
+    pred += width;
+    ref += ref_stride;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vpx_dsp/variance.h b/libvpx/libvpx/vpx_dsp/variance.h
new file mode 100644
index 0000000..c18d9b4
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/variance.h
@@ -0,0 +1,94 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_VARIANCE_H_
+#define VPX_DSP_VARIANCE_H_
+
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define FILTER_BITS 7
+#define FILTER_WEIGHT 128
+
+typedef unsigned int(*vpx_sad_fn_t)(const uint8_t *a, int a_stride,
+                                    const uint8_t *b_ptr, int b_stride);
+
+typedef unsigned int(*vpx_sad_avg_fn_t)(const uint8_t *a_ptr, int a_stride,
+                                        const uint8_t *b_ptr, int b_stride,
+                                        const uint8_t *second_pred);
+
+typedef void (*vp8_copy32xn_fn_t)(const uint8_t *a, int a_stride,
+                                  uint8_t *b, int b_stride, int n);
+
+typedef void (*vpx_sad_multi_fn_t)(const uint8_t *a, int a_stride,
+                                   const uint8_t *b, int b_stride,
+                                   unsigned int *sad_array);
+
+typedef void (*vpx_sad_multi_d_fn_t)(const uint8_t *a, int a_stride,
+                                     const uint8_t *const b_array[],
+                                     int b_stride,
+                                     unsigned int *sad_array);
+
+typedef unsigned int (*vpx_variance_fn_t)(const uint8_t *a, int a_stride,
+                                          const uint8_t *b, int b_stride,
+                                          unsigned int *sse);
+
+typedef unsigned int (*vpx_subpixvariance_fn_t)(const uint8_t *a, int a_stride,
+                                                int xoffset, int yoffset,
+                                                const uint8_t *b, int b_stride,
+                                                unsigned int *sse);
+
+typedef unsigned int (*vpx_subp_avg_variance_fn_t)(const uint8_t *a_ptr,
+                                                   int a_stride,
+                                                   int xoffset, int yoffset,
+                                                   const uint8_t *b_ptr,
+                                                   int b_stride,
+                                                   unsigned int *sse,
+                                                   const uint8_t *second_pred);
+#if CONFIG_VP8
+typedef struct variance_vtable {
+  vpx_sad_fn_t            sdf;
+  vpx_variance_fn_t       vf;
+  vpx_subpixvariance_fn_t svf;
+  vpx_variance_fn_t       svf_halfpix_h;
+  vpx_variance_fn_t       svf_halfpix_v;
+  vpx_variance_fn_t       svf_halfpix_hv;
+  vpx_sad_multi_fn_t      sdx3f;
+  vpx_sad_multi_fn_t      sdx8f;
+  vpx_sad_multi_d_fn_t    sdx4df;
+#if ARCH_X86 || ARCH_X86_64
+  vp8_copy32xn_fn_t       copymem;
+#endif
+} vp8_variance_fn_ptr_t;
+#endif  // CONFIG_VP8
+
+#if CONFIG_VP9
+typedef struct vp9_variance_vtable {
+  vpx_sad_fn_t               sdf;
+  vpx_sad_avg_fn_t           sdaf;
+  vpx_variance_fn_t          vf;
+  vpx_subpixvariance_fn_t    svf;
+  vpx_subp_avg_variance_fn_t svaf;
+  vpx_sad_multi_fn_t         sdx3f;
+  vpx_sad_multi_fn_t         sdx8f;
+  vpx_sad_multi_d_fn_t       sdx4df;
+} vp9_variance_fn_ptr_t;
+#endif  // CONFIG_VP9
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_VARIANCE_H_
diff --git a/libvpx/libvpx/vpx_dsp/vpx_convolve.c b/libvpx/libvpx/vpx_dsp/vpx_convolve.c
new file mode 100644
index 0000000..2d1c927
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/vpx_convolve.c
@@ -0,0 +1,612 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <string.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_convolve.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/vpx_filter.h"
+#include "vpx_ports/mem.h"
+
+static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const InterpKernel *x_filters,
+                           int x0_q4, int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const InterpKernel *x_filters,
+                               int x0_q4, int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] +
+          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const InterpKernel *y_filters,
+                          int y0_q4, int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve_avg_vert(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const InterpKernel *y_filters,
+                              int y0_q4, int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
+          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve(const uint8_t *src, ptrdiff_t src_stride,
+                     uint8_t *dst, ptrdiff_t dst_stride,
+                     const InterpKernel *const x_filters,
+                     int x0_q4, int x_step_q4,
+                     const InterpKernel *const y_filters,
+                     int y0_q4, int y_step_q4,
+                     int w, int h) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint8_t temp[135 * 64];
+  int intermediate_height =
+          (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= 64);
+  assert(h <= 64);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, 64,
+                 x_filters, x0_q4, x_step_q4, w, intermediate_height);
+  convolve_vert(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, dst_stride,
+                y_filters, y0_q4, y_step_q4, w, h);
+}
+
+static const InterpKernel *get_filter_base(const int16_t *filter) {
+  // NOTE: This assumes that the filter table is 256-byte aligned.
+  // TODO(agrange) Modify to make independent of table alignment.
+  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
+}
+
+static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
+  return (int)((const InterpKernel *)(intptr_t)f - base);
+}
+
+void vpx_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4,
+                           int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_horiz(src, src_stride, dst, dst_stride, filters_x,
+                 x0_q4, x_step_q4, w, h);
+}
+
+void vpx_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4,
+                               int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x,
+                     x0_q4, x_step_q4, w, h);
+}
+
+void vpx_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const int16_t *filter_x, int x_step_q4,
+                          const int16_t *filter_y, int y_step_q4,
+                          int w, int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_vert(src, src_stride, dst, dst_stride, filters_y,
+                y0_q4, y_step_q4, w, h);
+}
+
+void vpx_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4,
+                              int w, int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y,
+                    y0_q4, y_step_q4, w, h);
+}
+
+void vpx_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
+                     uint8_t *dst, ptrdiff_t dst_stride,
+                     const int16_t *filter_x, int x_step_q4,
+                     const int16_t *filter_y, int y_step_q4,
+                     int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  convolve(src, src_stride, dst, dst_stride,
+           filters_x, x0_q4, x_step_q4,
+           filters_y, y0_q4, y_step_q4, w, h);
+}
+
+void vpx_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
+                         uint8_t *dst, ptrdiff_t dst_stride,
+                         const int16_t *filter_x, int x_step_q4,
+                         const int16_t *filter_y, int y_step_q4,
+                         int w, int h) {
+  /* Fixed size intermediate buffer places limits on parameters. */
+  DECLARE_ALIGNED(16, uint8_t, temp[64 * 64]);
+  assert(w <= 64);
+  assert(h <= 64);
+
+  vpx_convolve8_c(src, src_stride, temp, 64,
+                  filter_x, x_step_q4, filter_y, y_step_q4, w, h);
+  vpx_convolve_avg_c(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
+}
+
+void vpx_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride,
+                         uint8_t *dst, ptrdiff_t dst_stride,
+                         const int16_t *filter_x, int filter_x_stride,
+                         const int16_t *filter_y, int filter_y_stride,
+                         int w, int h) {
+  int r;
+
+  (void)filter_x;  (void)filter_x_stride;
+  (void)filter_y;  (void)filter_y_stride;
+
+  for (r = h; r > 0; --r) {
+    memcpy(dst, src, w);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vpx_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride,
+                        uint8_t *dst, ptrdiff_t dst_stride,
+                        const int16_t *filter_x, int filter_x_stride,
+                        const int16_t *filter_y, int filter_y_stride,
+                        int w, int h) {
+  int x, y;
+
+  (void)filter_x;  (void)filter_x_stride;
+  (void)filter_y;  (void)filter_y_stride;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x)
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
+
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vpx_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                        uint8_t *dst, ptrdiff_t dst_stride,
+                        const int16_t *filter_x, int x_step_q4,
+                        const int16_t *filter_y, int y_step_q4,
+                        int w, int h) {
+  vpx_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                        filter_y, y_step_q4, w, h);
+}
+
+void vpx_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                       uint8_t *dst, ptrdiff_t dst_stride,
+                       const int16_t *filter_x, int x_step_q4,
+                       const int16_t *filter_y, int y_step_q4,
+                       int w, int h) {
+  vpx_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                       filter_y, y_step_q4, w, h);
+}
+
+void vpx_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride,
+                     uint8_t *dst, ptrdiff_t dst_stride,
+                     const int16_t *filter_x, int x_step_q4,
+                     const int16_t *filter_y, int y_step_q4,
+                     int w, int h) {
+  vpx_convolve8_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                  filter_y, y_step_q4, w, h);
+}
+
+void vpx_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4,
+                            int w, int h) {
+  vpx_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                            x_step_q4, filter_y, y_step_q4, w, h);
+}
+
+void vpx_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4,
+                           int w, int h) {
+  vpx_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                           x_step_q4, filter_y, y_step_q4, w, h);
+}
+
+void vpx_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride,
+                     uint8_t *dst, ptrdiff_t dst_stride,
+                     const int16_t *filter_x, int x_step_q4,
+                     const int16_t *filter_y, int y_step_q4,
+                     int w, int h) {
+  vpx_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                      filter_y, y_step_q4, w, h);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+                                  uint8_t *dst8, ptrdiff_t dst_stride,
+                                  const InterpKernel *x_filters,
+                                  int x0_q4, int x_step_q4,
+                                  int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_avg_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+                                      uint8_t *dst8, ptrdiff_t dst_stride,
+                                      const InterpKernel *x_filters,
+                                      int x0_q4, int x_step_q4,
+                                      int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] +
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd), 1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride,
+                                 uint8_t *dst8, ptrdiff_t dst_stride,
+                                 const InterpKernel *y_filters,
+                                 int y0_q4, int y_step_q4, int w, int h,
+                                 int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = clip_pixel_highbd(
+          ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve_avg_vert(const uint8_t *src8, ptrdiff_t src_stride,
+                                     uint8_t *dst8, ptrdiff_t dst_stride,
+                                     const InterpKernel *y_filters,
+                                     int y0_q4, int y_step_q4, int w, int h,
+                                     int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd), 1);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const InterpKernel *const x_filters,
+                            int x0_q4, int x_step_q4,
+                            const InterpKernel *const y_filters,
+                            int y0_q4, int y_step_q4,
+                            int w, int h, int bd) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint16_t temp[64 * 135];
+  int intermediate_height =
+          (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= 64);
+  assert(h <= 64);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  highbd_convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                        src_stride, CONVERT_TO_BYTEPTR(temp), 64,
+                        x_filters, x0_q4, x_step_q4, w,
+                        intermediate_height, bd);
+  highbd_convolve_vert(CONVERT_TO_BYTEPTR(temp) + 64 * (SUBPEL_TAPS / 2 - 1),
+                       64, dst, dst_stride, y_filters, y0_q4, y_step_q4,
+                       w, h, bd);
+}
+
+
+void vpx_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4,
+                                  int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_horiz(src, src_stride, dst, dst_stride, filters_x,
+                        x0_q4, x_step_q4, w, h, bd);
+}
+
+void vpx_highbd_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                      uint8_t *dst, ptrdiff_t dst_stride,
+                                      const int16_t *filter_x, int x_step_q4,
+                                      const int16_t *filter_y, int y_step_q4,
+                                      int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x,
+                            x0_q4, x_step_q4, w, h, bd);
+}
+
+void vpx_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const int16_t *filter_x, int x_step_q4,
+                                 const int16_t *filter_y, int y_step_q4,
+                                 int w, int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_vert(src, src_stride, dst, dst_stride, filters_y,
+                       y0_q4, y_step_q4, w, h, bd);
+}
+
+void vpx_highbd_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                     uint8_t *dst, ptrdiff_t dst_stride,
+                                     const int16_t *filter_x, int x_step_q4,
+                                     const int16_t *filter_y, int y_step_q4,
+                                     int w, int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y,
+                           y0_q4, y_step_q4, w, h, bd);
+}
+
+void vpx_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4,
+                            int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  highbd_convolve(src, src_stride, dst, dst_stride,
+                  filters_x, x0_q4, x_step_q4,
+                  filters_y, y0_q4, y_step_q4, w, h, bd);
+}
+
+void vpx_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int x_step_q4,
+                                const int16_t *filter_y, int y_step_q4,
+                                int w, int h, int bd) {
+  // Fixed size intermediate buffer places limits on parameters.
+  DECLARE_ALIGNED(16, uint16_t, temp[64 * 64]);
+  assert(w <= 64);
+  assert(h <= 64);
+
+  vpx_highbd_convolve8_c(src, src_stride, CONVERT_TO_BYTEPTR(temp), 64,
+                         filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);
+  vpx_highbd_convolve_avg_c(CONVERT_TO_BYTEPTR(temp), 64, dst, dst_stride,
+                            NULL, 0, NULL, 0, w, h, bd);
+}
+
+void vpx_highbd_convolve_copy_c(const uint8_t *src8, ptrdiff_t src_stride,
+                                uint8_t *dst8, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int filter_x_stride,
+                                const int16_t *filter_y, int filter_y_stride,
+                                int w, int h, int bd) {
+  int r;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  for (r = h; r > 0; --r) {
+    memcpy(dst, src, w * sizeof(uint16_t));
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vpx_highbd_convolve_avg_c(const uint8_t *src8, ptrdiff_t src_stride,
+                               uint8_t *dst8, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int filter_x_stride,
+                               const int16_t *filter_y, int filter_y_stride,
+                               int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/vpx_convolve.h b/libvpx/libvpx/vpx_dsp/vpx_convolve.h
new file mode 100644
index 0000000..9ed3f17
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/vpx_convolve.h
@@ -0,0 +1,38 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VPX_DSP_VPX_CONVOLVE_H_
+#define VPX_DSP_VPX_CONVOLVE_H_
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef void (*convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4,
+                              int w, int h);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*highbd_convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
+                                     uint8_t *dst, ptrdiff_t dst_stride,
+                                     const int16_t *filter_x, int x_step_q4,
+                                     const int16_t *filter_y, int y_step_q4,
+                                     int w, int h, int bd);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_VPX_CONVOLVE_H_
diff --git a/libvpx/libvpx/vpx_dsp/vpx_dsp.mk b/libvpx/libvpx/vpx_dsp/vpx_dsp.mk
new file mode 100644
index 0000000..84b5291
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/vpx_dsp.mk
@@ -0,0 +1,354 @@
+##
+## Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+DSP_SRCS-yes += vpx_dsp.mk
+DSP_SRCS-yes += vpx_dsp_common.h
+
+DSP_SRCS-$(HAVE_MSA)    += mips/macros_msa.h
+
+# bit reader
+DSP_SRCS-yes += prob.h
+DSP_SRCS-yes += prob.c
+
+ifeq ($(CONFIG_ENCODERS),yes)
+DSP_SRCS-yes += bitwriter.h
+DSP_SRCS-yes += bitwriter.c
+DSP_SRCS-yes += bitwriter_buffer.c
+DSP_SRCS-yes += bitwriter_buffer.h
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += ssim.c
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += ssim.h
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += psnrhvs.c
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += fastssim.c
+endif
+
+ifeq ($(CONFIG_DECODERS),yes)
+DSP_SRCS-yes += bitreader.h
+DSP_SRCS-yes += bitreader.c
+DSP_SRCS-yes += bitreader_buffer.c
+DSP_SRCS-yes += bitreader_buffer.h
+endif
+
+# intra predictions
+DSP_SRCS-yes += intrapred.c
+
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSE) += x86/intrapred_sse2.asm
+DSP_SRCS-$(HAVE_SSE2) += x86/intrapred_sse2.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/intrapred_ssse3.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/vpx_subpixel_8t_ssse3.asm
+endif  # CONFIG_USE_X86INC
+
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSE)  += x86/highbd_intrapred_sse2.asm
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_intrapred_sse2.asm
+endif  # CONFIG_USE_X86INC
+endif  # CONFIG_VP9_HIGHBITDEPTH
+
+ifneq ($(filter yes,$(CONFIG_POSTPROC) $(CONFIG_VP9_POSTPROC)),)
+DSP_SRCS-yes += add_noise.c
+DSP_SRCS-$(HAVE_MSA) += mips/add_noise_msa.c
+DSP_SRCS-$(HAVE_SSE2) += x86/add_noise_sse2.asm
+endif # CONFIG_POSTPROC
+
+DSP_SRCS-$(HAVE_NEON_ASM) += arm/intrapred_neon_asm$(ASM)
+DSP_SRCS-$(HAVE_NEON) += arm/intrapred_neon.c
+DSP_SRCS-$(HAVE_MSA) += mips/intrapred_msa.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/intrapred4_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/intrapred8_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/intrapred16_dspr2.c
+
+DSP_SRCS-$(HAVE_DSPR2)  += mips/common_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/common_dspr2.c
+
+# interpolation filters
+DSP_SRCS-yes += vpx_convolve.c
+DSP_SRCS-yes += vpx_convolve.h
+DSP_SRCS-yes += vpx_filter.h
+
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/convolve.h
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/vpx_asm_stubs.c
+DSP_SRCS-$(HAVE_SSE2)  += x86/vpx_subpixel_8t_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)  += x86/vpx_subpixel_bilinear_sse2.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/vpx_subpixel_8t_ssse3.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/vpx_subpixel_bilinear_ssse3.asm
+DSP_SRCS-$(HAVE_AVX2)  += x86/vpx_subpixel_8t_intrin_avx2.c
+DSP_SRCS-$(HAVE_SSSE3) += x86/vpx_subpixel_8t_intrin_ssse3.c
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)  += x86/vpx_high_subpixel_8t_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)  += x86/vpx_high_subpixel_bilinear_sse2.asm
+endif
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSE2)  += x86/vpx_convolve_copy_sse2.asm
+endif
+
+ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes += arm/vpx_convolve_copy_neon_asm$(ASM)
+DSP_SRCS-yes += arm/vpx_convolve8_avg_neon_asm$(ASM)
+DSP_SRCS-yes += arm/vpx_convolve8_neon_asm$(ASM)
+DSP_SRCS-yes += arm/vpx_convolve_avg_neon_asm$(ASM)
+DSP_SRCS-yes += arm/vpx_convolve_neon.c
+else
+ifeq ($(HAVE_NEON),yes)
+DSP_SRCS-yes += arm/vpx_convolve_copy_neon.c
+DSP_SRCS-yes += arm/vpx_convolve8_avg_neon.c
+DSP_SRCS-yes += arm/vpx_convolve8_neon.c
+DSP_SRCS-yes += arm/vpx_convolve_avg_neon.c
+DSP_SRCS-yes += arm/vpx_convolve_neon.c
+endif  # HAVE_NEON
+endif  # HAVE_NEON_ASM
+
+# common (msa)
+DSP_SRCS-$(HAVE_MSA) += mips/vpx_convolve8_avg_horiz_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/vpx_convolve8_avg_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/vpx_convolve8_avg_vert_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/vpx_convolve8_horiz_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/vpx_convolve8_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/vpx_convolve8_vert_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/vpx_convolve_avg_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/vpx_convolve_copy_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/vpx_convolve_msa.h
+
+# common (dspr2)
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve_common_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_avg_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_avg_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_vert_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_avg_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_avg_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_vert_dspr2.c
+
+# loop filters
+DSP_SRCS-yes += loopfilter.c
+
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64)   += x86/loopfilter_sse2.c
+DSP_SRCS-$(HAVE_AVX2)                += x86/loopfilter_avx2.c
+
+DSP_SRCS-$(HAVE_NEON)   += arm/loopfilter_neon.c
+ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes  += arm/loopfilter_mb_neon$(ASM)
+DSP_SRCS-yes  += arm/loopfilter_16_neon$(ASM)
+DSP_SRCS-yes  += arm/loopfilter_8_neon$(ASM)
+DSP_SRCS-yes  += arm/loopfilter_4_neon$(ASM)
+else
+ifeq ($(HAVE_NEON),yes)
+DSP_SRCS-yes   += arm/loopfilter_16_neon.c
+DSP_SRCS-yes   += arm/loopfilter_8_neon.c
+DSP_SRCS-yes   += arm/loopfilter_4_neon.c
+endif  # HAVE_NEON
+endif  # HAVE_NEON_ASM
+
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_msa.h
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_16_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_8_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_4_msa.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_filters_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_filters_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_macros_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_masks_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_mb_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_mb_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_mb_vert_dspr2.c
+
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_loopfilter_sse2.c
+endif  # CONFIG_VP9_HIGHBITDEPTH
+
+DSP_SRCS-yes            += txfm_common.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/txfm_common_sse2.h
+DSP_SRCS-$(HAVE_MSA)    += mips/txfm_macros_msa.h
+# forward transform
+ifeq ($(CONFIG_VP9_ENCODER),yes)
+DSP_SRCS-yes            += fwd_txfm.c
+DSP_SRCS-yes            += fwd_txfm.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_txfm_sse2.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_txfm_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_txfm_impl_sse2.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_dct32x32_impl_sse2.h
+ifeq ($(ARCH_X86_64),yes)
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSSE3)  += x86/fwd_txfm_ssse3_x86_64.asm
+endif
+endif
+DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_txfm_avx2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_dct32x32_impl_avx2.h
+DSP_SRCS-$(HAVE_NEON)   += arm/fwd_txfm_neon.c
+DSP_SRCS-$(HAVE_MSA)    += mips/fwd_txfm_msa.h
+DSP_SRCS-$(HAVE_MSA)    += mips/fwd_txfm_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/fwd_dct32x32_msa.c
+endif  # CONFIG_VP9_ENCODER
+
+# inverse transform
+ifeq ($(CONFIG_VP9),yes)
+DSP_SRCS-yes            += inv_txfm.h
+DSP_SRCS-yes            += inv_txfm.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/inv_txfm_sse2.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/inv_txfm_sse2.c
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/inv_wht_sse2.asm
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_SSSE3)  += x86/inv_txfm_ssse3_x86_64.asm
+endif  # ARCH_X86_64
+endif  # CONFIG_USE_X86INC
+
+ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes  += arm/save_reg_neon$(ASM)
+DSP_SRCS-yes  += arm/idct4x4_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct4x4_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct8x8_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct8x8_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct16x16_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct16x16_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct32x32_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct32x32_add_neon$(ASM)
+else
+ifeq ($(HAVE_NEON),yes)
+DSP_SRCS-yes  += arm/idct4x4_1_add_neon.c
+DSP_SRCS-yes  += arm/idct4x4_add_neon.c
+DSP_SRCS-yes  += arm/idct8x8_1_add_neon.c
+DSP_SRCS-yes  += arm/idct8x8_add_neon.c
+DSP_SRCS-yes  += arm/idct16x16_1_add_neon.c
+DSP_SRCS-yes  += arm/idct16x16_add_neon.c
+DSP_SRCS-yes  += arm/idct32x32_1_add_neon.c
+DSP_SRCS-yes  += arm/idct32x32_add_neon.c
+endif  # HAVE_NEON
+endif  # HAVE_NEON_ASM
+DSP_SRCS-$(HAVE_NEON)  += arm/idct16x16_neon.c
+
+DSP_SRCS-$(HAVE_MSA)   += mips/inv_txfm_msa.h
+DSP_SRCS-$(HAVE_MSA)   += mips/idct4x4_msa.c
+DSP_SRCS-$(HAVE_MSA)   += mips/idct8x8_msa.c
+DSP_SRCS-$(HAVE_MSA)   += mips/idct16x16_msa.c
+DSP_SRCS-$(HAVE_MSA)   += mips/idct32x32_msa.c
+
+ifneq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_DSPR2) += mips/inv_txfm_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans4_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans8_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans16_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans32_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans32_cols_dspr2.c
+endif  # CONFIG_VP9_HIGHBITDEPTH
+endif  # CONFIG_VP9
+
+# quantization
+ifeq ($(CONFIG_VP9_ENCODER),yes)
+DSP_SRCS-yes            += quantize.c
+DSP_SRCS-yes            += quantize.h
+
+DSP_SRCS-$(HAVE_SSE2)   += x86/quantize_sse2.c
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_quantize_intrin_sse2.c
+endif
+ifeq ($(ARCH_X86_64),yes)
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSSE3)  += x86/quantize_ssse3_x86_64.asm
+DSP_SRCS-$(HAVE_AVX)    += x86/quantize_avx_x86_64.asm
+endif
+endif
+
+# avg
+DSP_SRCS-yes           += avg.c
+DSP_SRCS-$(HAVE_SSE2)  += x86/avg_intrin_sse2.c
+DSP_SRCS-$(HAVE_NEON)  += arm/avg_neon.c
+DSP_SRCS-$(HAVE_MSA)   += mips/avg_msa.c
+DSP_SRCS-$(HAVE_NEON)  += arm/hadamard_neon.c
+ifeq ($(ARCH_X86_64),yes)
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSSE3) += x86/avg_ssse3_x86_64.asm
+endif
+endif
+
+endif  # CONFIG_VP9_ENCODER
+
+ifeq ($(CONFIG_ENCODERS),yes)
+DSP_SRCS-yes            += sad.c
+DSP_SRCS-yes            += subtract.c
+
+DSP_SRCS-$(HAVE_MEDIA)  += arm/sad_media$(ASM)
+DSP_SRCS-$(HAVE_NEON)   += arm/sad4d_neon.c
+DSP_SRCS-$(HAVE_NEON)   += arm/sad_neon.c
+DSP_SRCS-$(HAVE_NEON)   += arm/subtract_neon.c
+
+DSP_SRCS-$(HAVE_MSA)    += mips/sad_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/subtract_msa.c
+
+DSP_SRCS-$(HAVE_SSE3)   += x86/sad_sse3.asm
+DSP_SRCS-$(HAVE_SSSE3)  += x86/sad_ssse3.asm
+DSP_SRCS-$(HAVE_SSE4_1) += x86/sad_sse4.asm
+DSP_SRCS-$(HAVE_AVX2)   += x86/sad4d_avx2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/sad_avx2.c
+
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSE)    += x86/sad4d_sse2.asm
+DSP_SRCS-$(HAVE_SSE)    += x86/sad_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/sad4d_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/sad_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/subtract_sse2.asm
+
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad4d_sse2.asm
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad_sse2.asm
+endif  # CONFIG_VP9_HIGHBITDEPTH
+endif  # CONFIG_USE_X86INC
+
+endif  # CONFIG_ENCODERS
+
+ifneq ($(filter yes,$(CONFIG_ENCODERS) $(CONFIG_POSTPROC) $(CONFIG_VP9_POSTPROC)),)
+DSP_SRCS-yes            += variance.c
+DSP_SRCS-yes            += variance.h
+
+DSP_SRCS-$(HAVE_MEDIA)  += arm/bilinear_filter_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/subpel_variance_media.c
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_halfpixvar16x16_h_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_halfpixvar16x16_hv_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_halfpixvar16x16_v_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_media$(ASM)
+DSP_SRCS-$(HAVE_NEON)   += arm/subpel_variance_neon.c
+DSP_SRCS-$(HAVE_NEON)   += arm/variance_neon.c
+
+DSP_SRCS-$(HAVE_MSA)    += mips/variance_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/sub_pixel_variance_msa.c
+
+DSP_SRCS-$(HAVE_SSE)    += x86/variance_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/variance_sse2.c  # Contains SSE2 and SSSE3
+DSP_SRCS-$(HAVE_SSE2)   += x86/halfpix_variance_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/halfpix_variance_impl_sse2.asm
+DSP_SRCS-$(HAVE_AVX2)   += x86/variance_avx2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/variance_impl_avx2.c
+
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/ssim_opt_x86_64.asm
+endif  # ARCH_X86_64
+
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSE)    += x86/subpel_variance_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/subpel_variance_sse2.asm  # Contains SSE2 and SSSE3
+endif  # CONFIG_USE_X86INC
+
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_variance_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_variance_impl_sse2.asm
+ifeq ($(CONFIG_USE_X86INC),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_subpel_variance_impl_sse2.asm
+endif  # CONFIG_USE_X86INC
+endif  # CONFIG_VP9_HIGHBITDEPTH
+endif  # CONFIG_ENCODERS || CONFIG_POSTPROC || CONFIG_VP9_POSTPROC
+
+DSP_SRCS-no += $(DSP_SRCS_REMOVE-yes)
+
+DSP_SRCS-yes += vpx_dsp_rtcd.c
+DSP_SRCS-yes += vpx_dsp_rtcd_defs.pl
+
+$(eval $(call rtcd_h_template,vpx_dsp_rtcd,vpx_dsp/vpx_dsp_rtcd_defs.pl))
diff --git a/libvpx/libvpx/vpx_dsp/vpx_dsp_common.h b/libvpx/libvpx/vpx_dsp/vpx_dsp_common.h
new file mode 100644
index 0000000..a1d0a51
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/vpx_dsp_common.h
@@ -0,0 +1,69 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_VPX_DSP_COMMON_H_
+#define VPX_DSP_VPX_DSP_COMMON_H_
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define VPXMIN(x, y) (((x) < (y)) ? (x) : (y))
+#define VPXMAX(x, y) (((x) > (y)) ? (x) : (y))
+
+#if CONFIG_VP9_HIGHBITDEPTH
+// Note:
+// tran_low_t  is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+typedef int64_t tran_high_t;
+typedef int32_t tran_low_t;
+#else
+// Note:
+// tran_low_t  is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+typedef int32_t tran_high_t;
+typedef int16_t tran_low_t;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static INLINE uint8_t clip_pixel(int val) {
+  return (val > 255) ? 255 : (val < 0) ? 0 : val;
+}
+
+static INLINE int clamp(int value, int low, int high) {
+  return value < low ? low : (value > high ? high : value);
+}
+
+static INLINE double fclamp(double value, double low, double high) {
+  return value < low ? low : (value > high ? high : value);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE uint16_t clip_pixel_highbd(int val, int bd) {
+  switch (bd) {
+    case 8:
+    default:
+      return (uint16_t)clamp(val, 0, 255);
+    case 10:
+      return (uint16_t)clamp(val, 0, 1023);
+    case 12:
+      return (uint16_t)clamp(val, 0, 4095);
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_VPX_DSP_COMMON_H_
diff --git a/libvpx/libvpx/vpx_dsp/vpx_dsp_rtcd.c b/libvpx/libvpx/vpx_dsp/vpx_dsp_rtcd.c
new file mode 100644
index 0000000..5fe27b6
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/vpx_dsp_rtcd.c
@@ -0,0 +1,17 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_config.h"
+#define RTCD_C
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/vpx_once.h"
+
+void vpx_dsp_rtcd() {
+  once(setup_rtcd_internal);
+}
diff --git a/libvpx/libvpx/vpx_dsp/vpx_dsp_rtcd_defs.pl b/libvpx/libvpx/vpx_dsp/vpx_dsp_rtcd_defs.pl
new file mode 100644
index 0000000..37239a1
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -0,0 +1,1921 @@
+sub vpx_dsp_forward_decls() {
+print <<EOF
+/*
+ * DSP
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+
+EOF
+}
+forward_decls qw/vpx_dsp_forward_decls/;
+
+# x86inc.asm had specific constraints. break it out so it's easy to disable.
+# zero all the variables to avoid tricky else conditions.
+$mmx_x86inc = $sse_x86inc = $sse2_x86inc = $ssse3_x86inc = $avx_x86inc =
+  $avx2_x86inc = '';
+$mmx_x86_64_x86inc = $sse_x86_64_x86inc = $sse2_x86_64_x86inc =
+  $ssse3_x86_64_x86inc = $avx_x86_64_x86inc = $avx2_x86_64_x86inc = '';
+if (vpx_config("CONFIG_USE_X86INC") eq "yes") {
+  $mmx_x86inc = 'mmx';
+  $sse_x86inc = 'sse';
+  $sse2_x86inc = 'sse2';
+  $ssse3_x86inc = 'ssse3';
+  $avx_x86inc = 'avx';
+  $avx2_x86inc = 'avx2';
+  if ($opts{arch} eq "x86_64") {
+    $mmx_x86_64_x86inc = 'mmx';
+    $sse_x86_64_x86inc = 'sse';
+    $sse2_x86_64_x86inc = 'sse2';
+    $ssse3_x86_64_x86inc = 'ssse3';
+    $avx_x86_64_x86inc = 'avx';
+    $avx2_x86_64_x86inc = 'avx2';
+  }
+}
+
+# optimizations which depend on multiple features
+$avx2_ssse3 = '';
+if ((vpx_config("HAVE_AVX2") eq "yes") && (vpx_config("HAVE_SSSE3") eq "yes")) {
+  $avx2_ssse3 = 'avx2';
+}
+
+# functions that are 64 bit only.
+$mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = '';
+if ($opts{arch} eq "x86_64") {
+  $mmx_x86_64 = 'mmx';
+  $sse2_x86_64 = 'sse2';
+  $ssse3_x86_64 = 'ssse3';
+  $avx_x86_64 = 'avx';
+  $avx2_x86_64 = 'avx2';
+}
+
+#
+# Intra prediction
+#
+
+add_proto qw/void vpx_d207_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d207_predictor_4x4/, "$sse2_x86inc";
+
+add_proto qw/void vpx_d207e_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d207e_predictor_4x4/;
+
+add_proto qw/void vpx_d45_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d45_predictor_4x4 neon/, "$sse2_x86inc";
+
+add_proto qw/void vpx_d45e_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d45e_predictor_4x4/;
+
+add_proto qw/void vpx_d63_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d63_predictor_4x4/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_d63e_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d63e_predictor_4x4/;
+
+add_proto qw/void vpx_d63f_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d63f_predictor_4x4/;
+
+add_proto qw/void vpx_h_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_h_predictor_4x4 neon dspr2 msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_he_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_he_predictor_4x4/;
+
+add_proto qw/void vpx_d117_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d117_predictor_4x4/;
+
+add_proto qw/void vpx_d135_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d135_predictor_4x4 neon/;
+
+add_proto qw/void vpx_d153_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d153_predictor_4x4/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_v_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_v_predictor_4x4 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_ve_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_ve_predictor_4x4/;
+
+add_proto qw/void vpx_tm_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_predictor_4x4 dspr2 msa neon/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_top_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_top_predictor_4x4 msa neon/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_left_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_left_predictor_4x4 msa neon/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_128_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_128_predictor_4x4 msa neon/, "$sse2_x86inc";
+
+add_proto qw/void vpx_d207_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d207_predictor_8x8/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_d207e_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d207e_predictor_8x8/;
+
+add_proto qw/void vpx_d45_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d45_predictor_8x8 neon/, "$sse2_x86inc";
+
+add_proto qw/void vpx_d45e_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d45e_predictor_8x8/;
+
+add_proto qw/void vpx_d63_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d63_predictor_8x8/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_d63e_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d63e_predictor_8x8/;
+
+add_proto qw/void vpx_h_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_h_predictor_8x8 neon dspr2 msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_d117_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d117_predictor_8x8/;
+
+add_proto qw/void vpx_d135_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d135_predictor_8x8/;
+
+add_proto qw/void vpx_d153_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d153_predictor_8x8/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_v_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_v_predictor_8x8 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_tm_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_tm_predictor_8x8 neon dspr2 msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_predictor_8x8 dspr2 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_top_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_top_predictor_8x8 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_left_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_left_predictor_8x8 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_128_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_128_predictor_8x8 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_d207_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d207_predictor_16x16/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_d207e_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d207e_predictor_16x16/;
+
+add_proto qw/void vpx_d45_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d45_predictor_16x16 neon/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_d45e_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d45e_predictor_16x16/;
+
+add_proto qw/void vpx_d63_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d63_predictor_16x16/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_d63e_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d63e_predictor_16x16/;
+
+add_proto qw/void vpx_h_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_h_predictor_16x16 neon dspr2 msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_d117_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d117_predictor_16x16/;
+
+add_proto qw/void vpx_d135_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d135_predictor_16x16/;
+
+add_proto qw/void vpx_d153_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d153_predictor_16x16/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_v_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_v_predictor_16x16 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_tm_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_tm_predictor_16x16 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_predictor_16x16 dspr2 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_top_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_top_predictor_16x16 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_left_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_left_predictor_16x16 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_128_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_128_predictor_16x16 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_d207_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d207_predictor_32x32/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_d207e_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d207e_predictor_32x32/;
+
+add_proto qw/void vpx_d45_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d45_predictor_32x32/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_d45e_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d45e_predictor_32x32/;
+
+add_proto qw/void vpx_d63_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d63_predictor_32x32/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_d63e_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d63e_predictor_32x32/;
+
+add_proto qw/void vpx_h_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_h_predictor_32x32 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_d117_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d117_predictor_32x32/;
+
+add_proto qw/void vpx_d135_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d135_predictor_32x32/;
+
+add_proto qw/void vpx_d153_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_d153_predictor_32x32/, "$ssse3_x86inc";
+
+add_proto qw/void vpx_v_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_v_predictor_32x32 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_tm_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_tm_predictor_32x32 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_predictor_32x32 msa neon/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_top_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_top_predictor_32x32 msa neon/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_left_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_left_predictor_32x32 msa neon/, "$sse2_x86inc";
+
+add_proto qw/void vpx_dc_128_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+specialize qw/vpx_dc_128_predictor_32x32 msa neon/, "$sse2_x86inc";
+
+# High bitdepth functions
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void vpx_highbd_d207_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d207_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_d207e_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d207e_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_d45_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d45_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_d45e_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d45e_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_d63_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d63_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_d63e_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d63e_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_h_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_h_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_d117_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d117_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_d135_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d135_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_d153_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d153_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_v_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_v_predictor_4x4/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_tm_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_tm_predictor_4x4/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_dc_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_predictor_4x4/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_dc_top_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_top_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_dc_left_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_left_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_dc_128_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_128_predictor_4x4/;
+
+  add_proto qw/void vpx_highbd_d207_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d207_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_d207e_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d207e_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_d45_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d45_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_d45e_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d45e_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_d63_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d63_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_d63e_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d63e_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_h_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_h_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_d117_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d117_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_d135_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d135_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_d153_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d153_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_v_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_v_predictor_8x8/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_tm_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_tm_predictor_8x8/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_dc_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_predictor_8x8/, "$sse2_x86inc";;
+
+  add_proto qw/void vpx_highbd_dc_top_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_top_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_dc_left_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_left_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_dc_128_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_128_predictor_8x8/;
+
+  add_proto qw/void vpx_highbd_d207_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d207_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_d207e_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d207e_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_d45_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d45_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_d45e_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d45e_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_d63_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d63_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_d63e_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d63e_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_h_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_h_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_d117_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d117_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_d135_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d135_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_d153_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d153_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_v_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_v_predictor_16x16/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_tm_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_tm_predictor_16x16/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_dc_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_predictor_16x16/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_dc_top_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_top_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_dc_left_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_left_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_dc_128_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_128_predictor_16x16/;
+
+  add_proto qw/void vpx_highbd_d207_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d207_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_d207e_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d207e_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_d45_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d45_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_d45e_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d45e_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_d63_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d63_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_d63e_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d63e_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_h_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_h_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_d117_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d117_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_d135_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d135_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_d153_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_d153_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_v_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_v_predictor_32x32/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_tm_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_tm_predictor_32x32/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_dc_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_predictor_32x32/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_dc_top_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_top_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_dc_left_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_left_predictor_32x32/;
+
+  add_proto qw/void vpx_highbd_dc_128_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+  specialize qw/vpx_highbd_dc_128_predictor_32x32/;
+}  # CONFIG_VP9_HIGHBITDEPTH
+
+#
+# Sub Pixel Filters
+#
+add_proto qw/void vpx_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_convolve_copy neon dspr2 msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_convolve_avg neon dspr2 msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_convolve8 sse2 ssse3 neon dspr2 msa/, "$avx2_ssse3";
+
+add_proto qw/void vpx_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_convolve8_horiz sse2 ssse3 neon dspr2 msa/, "$avx2_ssse3";
+
+add_proto qw/void vpx_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_convolve8_vert sse2 ssse3 neon dspr2 msa/, "$avx2_ssse3";
+
+add_proto qw/void vpx_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_convolve8_avg sse2 ssse3 neon dspr2 msa/;
+
+add_proto qw/void vpx_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_convolve8_avg_horiz sse2 ssse3 neon dspr2 msa/;
+
+add_proto qw/void vpx_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_convolve8_avg_vert sse2 ssse3 neon dspr2 msa/;
+
+add_proto qw/void vpx_scaled_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_scaled_2d ssse3/;
+
+add_proto qw/void vpx_scaled_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_scaled_horiz/;
+
+add_proto qw/void vpx_scaled_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_scaled_vert/;
+
+add_proto qw/void vpx_scaled_avg_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_scaled_avg_2d/;
+
+add_proto qw/void vpx_scaled_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_scaled_avg_horiz/;
+
+add_proto qw/void vpx_scaled_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+specialize qw/vpx_scaled_avg_vert/;
+
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  #
+  # Sub Pixel Filters
+  #
+  add_proto qw/void vpx_highbd_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vpx_highbd_convolve_copy/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vpx_highbd_convolve_avg/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vpx_highbd_convolve8/, "$sse2_x86_64";
+
+  add_proto qw/void vpx_highbd_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vpx_highbd_convolve8_horiz/, "$sse2_x86_64";
+
+  add_proto qw/void vpx_highbd_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vpx_highbd_convolve8_vert/, "$sse2_x86_64";
+
+  add_proto qw/void vpx_highbd_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vpx_highbd_convolve8_avg/, "$sse2_x86_64";
+
+  add_proto qw/void vpx_highbd_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vpx_highbd_convolve8_avg_horiz/, "$sse2_x86_64";
+
+  add_proto qw/void vpx_highbd_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vpx_highbd_convolve8_avg_vert/, "$sse2_x86_64";
+}  # CONFIG_VP9_HIGHBITDEPTH
+
+#
+# Loopfilter
+#
+add_proto qw/void vpx_lpf_vertical_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/vpx_lpf_vertical_16 sse2 neon_asm dspr2 msa/;
+$vpx_lpf_vertical_16_neon_asm=vpx_lpf_vertical_16_neon;
+
+add_proto qw/void vpx_lpf_vertical_16_dual/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/vpx_lpf_vertical_16_dual sse2 neon_asm dspr2 msa/;
+$vpx_lpf_vertical_16_dual_neon_asm=vpx_lpf_vertical_16_dual_neon;
+
+add_proto qw/void vpx_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/vpx_lpf_vertical_8 sse2 neon dspr2 msa/;
+
+add_proto qw/void vpx_lpf_vertical_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/vpx_lpf_vertical_8_dual sse2 neon_asm dspr2 msa/;
+$vpx_lpf_vertical_8_dual_neon_asm=vpx_lpf_vertical_8_dual_neon;
+
+add_proto qw/void vpx_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/vpx_lpf_vertical_4 sse2 neon dspr2 msa/;
+
+add_proto qw/void vpx_lpf_vertical_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/vpx_lpf_vertical_4_dual sse2 neon dspr2 msa/;
+
+add_proto qw/void vpx_lpf_horizontal_edge_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/vpx_lpf_horizontal_edge_8 sse2 avx2 neon_asm dspr2 msa/;
+$vpx_lpf_horizontal_edge_8_neon_asm=vpx_lpf_horizontal_edge_8_neon;
+
+add_proto qw/void vpx_lpf_horizontal_edge_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/vpx_lpf_horizontal_edge_16 sse2 avx2 neon_asm dspr2 msa/;
+$vpx_lpf_horizontal_edge_16_neon_asm=vpx_lpf_horizontal_edge_16_neon;
+
+add_proto qw/void vpx_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/vpx_lpf_horizontal_8 sse2 neon dspr2 msa/;
+
+add_proto qw/void vpx_lpf_horizontal_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/vpx_lpf_horizontal_8_dual sse2 neon_asm dspr2 msa/;
+$vpx_lpf_horizontal_8_dual_neon_asm=vpx_lpf_horizontal_8_dual_neon;
+
+add_proto qw/void vpx_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/vpx_lpf_horizontal_4 sse2 neon dspr2 msa/;
+
+add_proto qw/void vpx_lpf_horizontal_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/vpx_lpf_horizontal_4_dual sse2 neon dspr2 msa/;
+
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void vpx_highbd_lpf_vertical_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/vpx_highbd_lpf_vertical_16 sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_vertical_16_dual/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/vpx_highbd_lpf_vertical_16_dual sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_vertical_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/vpx_highbd_lpf_vertical_8 sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_vertical_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+  specialize qw/vpx_highbd_lpf_vertical_8_dual sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_vertical_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/vpx_highbd_lpf_vertical_4 sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_vertical_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+  specialize qw/vpx_highbd_lpf_vertical_4_dual sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_horizontal_edge_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/vpx_highbd_lpf_horizontal_edge_8 sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_horizontal_edge_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/vpx_highbd_lpf_horizontal_edge_16 sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_horizontal_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/vpx_highbd_lpf_horizontal_8 sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_horizontal_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+  specialize qw/vpx_highbd_lpf_horizontal_8_dual sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_horizontal_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/vpx_highbd_lpf_horizontal_4 sse2/;
+
+  add_proto qw/void vpx_highbd_lpf_horizontal_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+  specialize qw/vpx_highbd_lpf_horizontal_4_dual sse2/;
+}  # CONFIG_VP9_HIGHBITDEPTH
+
+#
+# Encoder functions.
+#
+
+#
+# Forward transform
+#
+if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct4x4 sse2/;
+
+  add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct4x4_1 sse2/;
+
+  add_proto qw/void vpx_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct8x8 sse2/;
+
+  add_proto qw/void vpx_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct8x8_1 sse2/;
+
+  add_proto qw/void vpx_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct16x16 sse2/;
+
+  add_proto qw/void vpx_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct16x16_1 sse2/;
+
+  add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct32x32 sse2/;
+
+  add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct32x32_rd sse2/;
+
+  add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct32x32_1 sse2/;
+
+  add_proto qw/void vpx_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_highbd_fdct4x4 sse2/;
+
+  add_proto qw/void vpx_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_highbd_fdct8x8 sse2/;
+
+  add_proto qw/void vpx_highbd_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_highbd_fdct8x8_1/;
+
+  add_proto qw/void vpx_highbd_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_highbd_fdct16x16 sse2/;
+
+  add_proto qw/void vpx_highbd_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_highbd_fdct16x16_1/;
+
+  add_proto qw/void vpx_highbd_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_highbd_fdct32x32 sse2/;
+
+  add_proto qw/void vpx_highbd_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_highbd_fdct32x32_rd sse2/;
+
+  add_proto qw/void vpx_highbd_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_highbd_fdct32x32_1/;
+} else {
+  add_proto qw/void vpx_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct4x4 sse2 msa/;
+
+  add_proto qw/void vpx_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct4x4_1 sse2/;
+
+  add_proto qw/void vpx_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct8x8 sse2 neon msa/, "$ssse3_x86_64_x86inc";
+
+  add_proto qw/void vpx_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct8x8_1 sse2 neon msa/;
+
+  add_proto qw/void vpx_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct16x16 sse2 msa/;
+
+  add_proto qw/void vpx_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct16x16_1 sse2 msa/;
+
+  add_proto qw/void vpx_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct32x32 sse2 avx2 msa/;
+
+  add_proto qw/void vpx_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct32x32_rd sse2 avx2 msa/;
+
+  add_proto qw/void vpx_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/vpx_fdct32x32_1 sse2 msa/;
+}  # CONFIG_VP9_HIGHBITDEPTH
+}  # CONFIG_VP9_ENCODER
+
+#
+# Inverse transform
+if (vpx_config("CONFIG_VP9") eq "yes") {
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  # Note as optimized versions of these functions are added we need to add a check to ensure
+  # that when CONFIG_EMULATE_HARDWARE is on, it defaults to the C versions only.
+  add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vpx_iwht4x4_1_add/;
+
+  add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vpx_iwht4x4_16_add/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vpx_highbd_idct4x4_1_add/;
+
+  add_proto qw/void vpx_highbd_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vpx_highbd_idct8x8_1_add/;
+
+  add_proto qw/void vpx_highbd_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vpx_highbd_idct16x16_1_add/;
+
+  add_proto qw/void vpx_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vpx_highbd_idct32x32_1024_add/;
+
+  add_proto qw/void vpx_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vpx_highbd_idct32x32_34_add/;
+
+  add_proto qw/void vpx_highbd_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vpx_highbd_idct32x32_1_add/;
+
+  add_proto qw/void vpx_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vpx_highbd_iwht4x4_1_add/;
+
+  add_proto qw/void vpx_highbd_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vpx_highbd_iwht4x4_16_add/;
+
+  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
+  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
+    add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct4x4_16_add/;
+
+    add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct4x4_1_add/;
+
+    add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_64_add/;
+
+    add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_12_add/;
+
+    add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_1_add/;
+
+    add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_256_add/;
+
+    add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_10_add/;
+
+    add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_1_add/;
+
+    add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_1024_add/;
+
+    add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_135_add/;
+
+    add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_34_add/;
+
+    add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_1_add/;
+
+    add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vpx_highbd_idct4x4_16_add/;
+
+    add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vpx_highbd_idct8x8_64_add/;
+
+    add_proto qw/void vpx_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vpx_highbd_idct8x8_10_add/;
+
+    add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vpx_highbd_idct16x16_256_add/;
+
+    add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vpx_highbd_idct16x16_10_add/;
+  } else {
+    add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct4x4_16_add sse2/;
+
+    add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct4x4_1_add sse2/;
+
+    add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_64_add sse2/, "$ssse3_x86_64_x86inc";
+
+    add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_12_add sse2/, "$ssse3_x86_64_x86inc";
+
+    add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_1_add sse2/;
+
+    add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_256_add sse2/;
+
+    add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_10_add sse2/;
+
+    add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_1_add sse2/;
+
+    add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_1024_add sse2/, "$ssse3_x86_64_x86inc";
+
+    add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_135_add sse2/, "$ssse3_x86_64_x86inc";
+    # Need to add 135 eob idct32x32 implementations.
+    $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;
+
+    add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_34_add sse2/, "$ssse3_x86_64_x86inc";
+
+    add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_1_add sse2/;
+
+    add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vpx_highbd_idct4x4_16_add sse2/;
+
+    add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vpx_highbd_idct8x8_64_add sse2/;
+
+    add_proto qw/void vpx_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vpx_highbd_idct8x8_10_add sse2/;
+
+    add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vpx_highbd_idct16x16_256_add sse2/;
+
+    add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vpx_highbd_idct16x16_10_add sse2/;
+  }  # CONFIG_EMULATE_HARDWARE
+} else {
+  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
+  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
+    add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct4x4_1_add/;
+
+    add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct4x4_16_add/;
+
+    add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_1_add/;
+
+    add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_64_add/;
+
+    add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_12_add/;
+
+    add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_1_add/;
+
+    add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_256_add/;
+
+    add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_10_add/;
+
+    add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_1024_add/;
+
+    add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_135_add/;
+
+    add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_34_add/;
+
+    add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_1_add/;
+
+    add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_iwht4x4_1_add/;
+
+    add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_iwht4x4_16_add/;
+  } else {
+    add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct4x4_1_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct4x4_16_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_1_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_64_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
+
+    add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct8x8_12_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
+
+    add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_1_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_256_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct16x16_10_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_1024_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
+
+    add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_135_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
+    # Need to add 135 eob idct32x32 implementations.
+    $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;
+    $vpx_idct32x32_135_add_neon=vpx_idct32x32_1024_add_neon;
+    $vpx_idct32x32_135_add_dspr2=vpx_idct32x32_1024_add_dspr2;
+    $vpx_idct32x32_135_add_msa=vpx_idct32x32_1024_add_msa;
+
+    add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_34_add sse2 neon dspr2 msa/, "$ssse3_x86_64_x86inc";
+    # Need to add 34 eob idct32x32 neon implementation.
+    $vpx_idct32x32_34_add_neon=vpx_idct32x32_1024_add_neon;
+
+    add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_idct32x32_1_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_iwht4x4_1_add msa/;
+
+    add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vpx_iwht4x4_16_add msa/, "$sse2_x86inc";
+  }  # CONFIG_EMULATE_HARDWARE
+}  # CONFIG_VP9_HIGHBITDEPTH
+}  # CONFIG_VP9
+
+#
+# Quantization
+#
+if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
+  add_proto qw/void vpx_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  specialize qw/vpx_quantize_b sse2/, "$ssse3_x86_64_x86inc", "$avx_x86_64_x86inc";
+
+  add_proto qw/void vpx_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  specialize qw/vpx_quantize_b_32x32/, "$ssse3_x86_64_x86inc", "$avx_x86_64_x86inc";
+
+  if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+    add_proto qw/void vpx_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    specialize qw/vpx_highbd_quantize_b sse2/;
+
+    add_proto qw/void vpx_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    specialize qw/vpx_highbd_quantize_b_32x32 sse2/;
+  }  # CONFIG_VP9_HIGHBITDEPTH
+}  # CONFIG_VP9_ENCODER
+
+if (vpx_config("CONFIG_ENCODERS") eq "yes") {
+#
+# Block subtraction
+#
+add_proto qw/void vpx_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
+specialize qw/vpx_subtract_block neon msa/, "$sse2_x86inc";
+
+#
+# Single block SAD
+#
+add_proto qw/unsigned int vpx_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad64x64 avx2 neon msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad64x32 avx2 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad32x64 avx2 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad32x32 avx2 neon msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad32x16 avx2 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad16x32 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad16x16 media neon msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad16x8 neon msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad8x16 neon msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad8x8 neon msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad8x4 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad4x8 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad4x4 neon msa/, "$sse2_x86inc";
+
+#
+# Avg
+#
+if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
+  add_proto qw/unsigned int vpx_avg_8x8/, "const uint8_t *, int p";
+  specialize qw/vpx_avg_8x8 sse2 neon msa/;
+
+  add_proto qw/unsigned int vpx_avg_4x4/, "const uint8_t *, int p";
+  specialize qw/vpx_avg_4x4 sse2 neon msa/;
+
+  add_proto qw/void vpx_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
+  specialize qw/vpx_minmax_8x8 sse2 neon/;
+
+  add_proto qw/void vpx_hadamard_8x8/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
+  specialize qw/vpx_hadamard_8x8 sse2 neon/, "$ssse3_x86_64_x86inc";
+
+  add_proto qw/void vpx_hadamard_16x16/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
+  specialize qw/vpx_hadamard_16x16 sse2 neon/;
+
+  add_proto qw/int vpx_satd/, "const int16_t *coeff, int length";
+  specialize qw/vpx_satd sse2 neon/;
+
+  add_proto qw/void vpx_int_pro_row/, "int16_t *hbuf, const uint8_t *ref, const int ref_stride, const int height";
+  specialize qw/vpx_int_pro_row sse2 neon/;
+
+  add_proto qw/int16_t vpx_int_pro_col/, "const uint8_t *ref, const int width";
+  specialize qw/vpx_int_pro_col sse2 neon/;
+
+  add_proto qw/int vpx_vector_var/, "const int16_t *ref, const int16_t *src, const int bwl";
+  specialize qw/vpx_vector_var neon sse2/;
+}  # CONFIG_VP9_ENCODER
+
+add_proto qw/unsigned int vpx_sad64x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad64x64_avg avx2 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad64x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad64x32_avg avx2 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad32x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad32x64_avg avx2 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad32x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad32x32_avg avx2 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad32x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad32x16_avg avx2 msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad16x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad16x32_avg msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad16x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad16x16_avg msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad16x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad16x8_avg msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad8x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad8x16_avg msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad8x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad8x8_avg msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad8x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad8x4_avg msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad4x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad4x8_avg msa/, "$sse2_x86inc";
+
+add_proto qw/unsigned int vpx_sad4x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vpx_sad4x4_avg msa/, "$sse2_x86inc";
+
+#
+# Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
+#
+# Blocks of 3
+add_proto qw/void vpx_sad64x64x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad64x64x3 msa/;
+
+add_proto qw/void vpx_sad32x32x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad32x32x3 msa/;
+
+add_proto qw/void vpx_sad16x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad16x16x3 sse3 ssse3 msa/;
+
+add_proto qw/void vpx_sad16x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad16x8x3 sse3 ssse3 msa/;
+
+add_proto qw/void vpx_sad8x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad8x16x3 sse3 msa/;
+
+add_proto qw/void vpx_sad8x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad8x8x3 sse3 msa/;
+
+add_proto qw/void vpx_sad4x4x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad4x4x3 sse3 msa/;
+
+# Blocks of 8
+add_proto qw/void vpx_sad64x64x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad64x64x8 msa/;
+
+add_proto qw/void vpx_sad32x32x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad32x32x8 msa/;
+
+add_proto qw/void vpx_sad16x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad16x16x8 sse4_1 msa/;
+
+add_proto qw/void vpx_sad16x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad16x8x8 sse4_1 msa/;
+
+add_proto qw/void vpx_sad8x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad8x16x8 sse4_1 msa/;
+
+add_proto qw/void vpx_sad8x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad8x8x8 sse4_1 msa/;
+
+add_proto qw/void vpx_sad8x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad8x4x8 msa/;
+
+add_proto qw/void vpx_sad4x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad4x8x8 msa/;
+
+add_proto qw/void vpx_sad4x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad4x4x8 sse4_1 msa/;
+
+#
+# Multi-block SAD, comparing a reference to N independent blocks
+#
+add_proto qw/void vpx_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad64x64x4d avx2 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad64x32x4d msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad32x64x4d msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad32x32x4d avx2 neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad32x16x4d msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad16x32x4d msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad16x16x4d neon msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad16x8x4d msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad8x16x4d msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad8x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad8x8x4d msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad8x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad8x4x4d msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad4x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad4x8x4d msa/, "$sse2_x86inc";
+
+add_proto qw/void vpx_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+specialize qw/vpx_sad4x4x4d msa/, "$sse2_x86inc";
+
+#
+# Structured Similarity (SSIM)
+#
+if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
+    add_proto qw/void vpx_ssim_parms_8x8/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
+    specialize qw/vpx_ssim_parms_8x8/, "$sse2_x86_64";
+
+    add_proto qw/void vpx_ssim_parms_16x16/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
+    specialize qw/vpx_ssim_parms_16x16/, "$sse2_x86_64";
+}
+
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  #
+  # Block subtraction
+  #
+  add_proto qw/void vpx_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd";
+  specialize qw/vpx_highbd_subtract_block/;
+
+  #
+  # Single block SAD
+  #
+  add_proto qw/unsigned int vpx_highbd_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad64x64/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad64x32/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad32x64/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad32x32/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad32x16/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad16x32/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad16x16/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad16x8/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad8x16/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad8x8/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad8x4/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad4x8/;
+
+  add_proto qw/unsigned int vpx_highbd_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vpx_highbd_sad4x4/;
+
+  #
+  # Avg
+  #
+  add_proto qw/unsigned int vpx_highbd_avg_8x8/, "const uint8_t *, int p";
+  specialize qw/vpx_highbd_avg_8x8/;
+  add_proto qw/unsigned int vpx_highbd_avg_4x4/, "const uint8_t *, int p";
+  specialize qw/vpx_highbd_avg_4x4/;
+  add_proto qw/void vpx_highbd_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
+  specialize qw/vpx_highbd_minmax_8x8/;
+
+  add_proto qw/unsigned int vpx_highbd_sad64x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad64x64_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad64x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad64x32_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad32x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad32x64_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad32x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad32x32_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad32x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad32x16_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad16x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad16x32_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad16x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad16x16_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad16x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad16x8_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad8x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad8x16_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad8x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad8x8_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad8x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad8x4_avg/, "$sse2_x86inc";
+
+  add_proto qw/unsigned int vpx_highbd_sad4x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad4x8_avg/;
+
+  add_proto qw/unsigned int vpx_highbd_sad4x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_sad4x4_avg/;
+
+  #
+  # Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
+  #
+  # Blocks of 3
+  add_proto qw/void vpx_highbd_sad64x64x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad64x64x3/;
+
+  add_proto qw/void vpx_highbd_sad32x32x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad32x32x3/;
+
+  add_proto qw/void vpx_highbd_sad16x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad16x16x3/;
+
+  add_proto qw/void vpx_highbd_sad16x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad16x8x3/;
+
+  add_proto qw/void vpx_highbd_sad8x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad8x16x3/;
+
+  add_proto qw/void vpx_highbd_sad8x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad8x8x3/;
+
+  add_proto qw/void vpx_highbd_sad4x4x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad4x4x3/;
+
+  # Blocks of 8
+  add_proto qw/void vpx_highbd_sad64x64x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad64x64x8/;
+
+  add_proto qw/void vpx_highbd_sad32x32x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad32x32x8/;
+
+  add_proto qw/void vpx_highbd_sad16x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad16x16x8/;
+
+  add_proto qw/void vpx_highbd_sad16x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad16x8x8/;
+
+  add_proto qw/void vpx_highbd_sad8x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad8x16x8/;
+
+  add_proto qw/void vpx_highbd_sad8x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad8x8x8/;
+
+  add_proto qw/void vpx_highbd_sad8x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad8x4x8/;
+
+  add_proto qw/void vpx_highbd_sad4x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad4x8x8/;
+
+  add_proto qw/void vpx_highbd_sad4x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad4x4x8/;
+
+  #
+  # Multi-block SAD, comparing a reference to N independent blocks
+  #
+  add_proto qw/void vpx_highbd_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad64x64x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad64x32x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad32x64x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad32x32x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad32x16x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad16x32x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad16x16x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad16x8x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad8x16x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad8x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad8x8x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad8x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad8x4x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad4x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad4x8x4d/, "$sse2_x86inc";
+
+  add_proto qw/void vpx_highbd_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, uint32_t *sad_array";
+  specialize qw/vpx_highbd_sad4x4x4d/, "$sse2_x86inc";
+
+  #
+  # Structured Similarity (SSIM)
+  #
+  if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
+    add_proto qw/void vpx_highbd_ssim_parms_8x8/, "const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
+    specialize qw/vpx_highbd_ssim_parms_8x8/;
+  }
+}  # CONFIG_VP9_HIGHBITDEPTH
+}  # CONFIG_ENCODERS
+
+if (vpx_config("CONFIG_ENCODERS") eq "yes" || vpx_config("CONFIG_POSTPROC") eq "yes" || vpx_config("CONFIG_VP9_POSTPROC") eq "yes") {
+
+#
+# Variance
+#
+add_proto qw/unsigned int vpx_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance64x64 sse2 avx2 neon msa/;
+
+add_proto qw/unsigned int vpx_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance64x32 sse2 avx2 neon msa/;
+
+add_proto qw/unsigned int vpx_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance32x64 sse2 neon msa/;
+
+add_proto qw/unsigned int vpx_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance32x32 sse2 avx2 neon msa/;
+
+add_proto qw/unsigned int vpx_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance32x16 sse2 avx2 msa/;
+
+add_proto qw/unsigned int vpx_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance16x32 sse2 msa/;
+
+add_proto qw/unsigned int vpx_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance16x16 sse2 avx2 media neon msa/;
+
+add_proto qw/unsigned int vpx_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance16x8 sse2 neon msa/;
+
+add_proto qw/unsigned int vpx_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance8x16 sse2 neon msa/;
+
+add_proto qw/unsigned int vpx_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance8x8 sse2 media neon msa/;
+
+add_proto qw/unsigned int vpx_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance8x4 sse2 msa/;
+
+add_proto qw/unsigned int vpx_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance4x8 sse2 msa/;
+
+add_proto qw/unsigned int vpx_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_variance4x4 sse2 msa/;
+
+#
+# Specialty Variance
+#
+add_proto qw/void vpx_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+  specialize qw/vpx_get16x16var sse2 avx2 neon msa/;
+
+add_proto qw/void vpx_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+  specialize qw/vpx_get8x8var sse2 neon msa/;
+
+add_proto qw/unsigned int vpx_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/vpx_mse16x16 sse2 avx2 media neon msa/;
+
+add_proto qw/unsigned int vpx_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/vpx_mse16x8 sse2 msa/;
+
+add_proto qw/unsigned int vpx_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/vpx_mse8x16 sse2 msa/;
+
+add_proto qw/unsigned int vpx_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/vpx_mse8x8 sse2 msa/;
+
+add_proto qw/unsigned int vpx_get_mb_ss/, "const int16_t *";
+  specialize qw/vpx_get_mb_ss sse2 msa/;
+
+add_proto qw/unsigned int vpx_get4x4sse_cs/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride";
+  specialize qw/vpx_get4x4sse_cs neon msa/;
+
+add_proto qw/void vpx_comp_avg_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride";
+
+#
+# Subpixel Variance
+#
+add_proto qw/uint32_t vpx_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance64x64 avx2 neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance64x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance32x64 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance32x32 avx2 neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance32x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance16x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance16x16 media neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance16x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance8x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance8x8 media neon msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance8x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance4x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_sub_pixel_variance4x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance64x64 avx2 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance64x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance32x64 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance32x32 avx2 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance32x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance16x32 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance16x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance16x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance8x16 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance8x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance8x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance4x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+add_proto qw/uint32_t vpx_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_sub_pixel_avg_variance4x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
+
+#
+# Specialty Subpixel
+#
+add_proto qw/uint32_t vpx_variance_halfpixvar16x16_h/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
+  specialize qw/vpx_variance_halfpixvar16x16_h sse2 media/;
+
+add_proto qw/uint32_t vpx_variance_halfpixvar16x16_v/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
+  specialize qw/vpx_variance_halfpixvar16x16_v sse2 media/;
+
+add_proto qw/uint32_t vpx_variance_halfpixvar16x16_hv/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
+  specialize qw/vpx_variance_halfpixvar16x16_hv sse2 media/;
+
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/unsigned int vpx_highbd_12_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_variance64x64 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_variance64x32 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_variance32x64 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_variance32x32 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_variance32x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_variance16x32 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_variance16x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_variance16x8 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_variance8x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_variance8x8 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_12_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_12_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+  add_proto qw/unsigned int vpx_highbd_10_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_variance64x64 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_variance64x32 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_variance32x64 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_variance32x32 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_variance32x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_variance16x32 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_variance16x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_variance16x8 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_variance8x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_variance8x8 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_10_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_10_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+  add_proto qw/unsigned int vpx_highbd_8_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_variance64x64 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_variance64x32 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_variance32x64 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_variance32x32 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_variance32x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_variance16x32 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_variance16x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_variance16x8 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_variance8x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_variance8x8 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_8_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_8_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+  add_proto qw/void vpx_highbd_8_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+  add_proto qw/void vpx_highbd_8_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+
+  add_proto qw/void vpx_highbd_10_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+  add_proto qw/void vpx_highbd_10_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+
+  add_proto qw/void vpx_highbd_12_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+  add_proto qw/void vpx_highbd_12_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+
+  add_proto qw/unsigned int vpx_highbd_8_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_mse16x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_8_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_8_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_8_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_8_mse8x8 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_mse16x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_10_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_10_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_10_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_10_mse8x8 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_mse16x16 sse2/;
+
+  add_proto qw/unsigned int vpx_highbd_12_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_12_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int vpx_highbd_12_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/vpx_highbd_12_mse8x8 sse2/;
+
+  add_proto qw/void vpx_highbd_comp_avg_pred/, "uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride";
+
+  #
+  # Subpixel Variance
+  #
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance64x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance64x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance32x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance32x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance32x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance16x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance16x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance16x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance8x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance8x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_12_sub_pixel_variance8x4/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance64x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance64x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance32x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance32x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance32x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance16x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance16x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance16x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance8x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance8x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_10_sub_pixel_variance8x4/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance64x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance64x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance32x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance32x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance32x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance16x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance16x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance16x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance8x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance8x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/vpx_highbd_8_sub_pixel_variance8x4/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_12_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  add_proto qw/uint32_t vpx_highbd_12_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_10_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  add_proto qw/uint32_t vpx_highbd_10_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance64x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance64x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x64/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance32x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x32/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance16x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x16/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x8/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/vpx_highbd_8_sub_pixel_avg_variance8x4/, "$sse2_x86inc";
+
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  add_proto qw/uint32_t vpx_highbd_8_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+
+}  # CONFIG_VP9_HIGHBITDEPTH
+
+#
+# Post Processing
+#
+if (vpx_config("CONFIG_POSTPROC") eq "yes" || vpx_config("CONFIG_VP9_POSTPROC") eq "yes") {
+    add_proto qw/void vpx_plane_add_noise/, "uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch";
+    specialize qw/vpx_plane_add_noise sse2 msa/;
+}
+
+}  # CONFIG_ENCODERS || CONFIG_POSTPROC || CONFIG_VP9_POSTPROC
+
+1;
diff --git a/libvpx/libvpx/vpx_dsp/vpx_filter.h b/libvpx/libvpx/vpx_dsp/vpx_filter.h
new file mode 100644
index 0000000..2617feb
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/vpx_filter.h
@@ -0,0 +1,34 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_VPX_FILTER_H_
+#define VPX_DSP_VPX_FILTER_H_
+
+#include "vpx/vpx_integer.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define FILTER_BITS 7
+
+#define SUBPEL_BITS 4
+#define SUBPEL_MASK ((1 << SUBPEL_BITS) - 1)
+#define SUBPEL_SHIFTS (1 << SUBPEL_BITS)
+#define SUBPEL_TAPS 8
+
+typedef int16_t InterpKernel[SUBPEL_TAPS];
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_VPX_FILTER_H_
diff --git a/libvpx/libvpx/vpx_dsp/x86/add_noise_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/add_noise_sse2.asm
new file mode 100644
index 0000000..ff61b19
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/add_noise_sse2.asm
@@ -0,0 +1,83 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vpx_plane_add_noise_sse2(unsigned char *start, unsigned char *noise,
+;                              unsigned char blackclamp[16],
+;                              unsigned char whiteclamp[16],
+;                              unsigned char bothclamp[16],
+;                              unsigned int width, unsigned int height,
+;                              int pitch)
+global sym(vpx_plane_add_noise_sse2) PRIVATE
+sym(vpx_plane_add_noise_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 8
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ; get the clamps in registers
+    mov     rdx, arg(2) ; blackclamp
+    movdqu  xmm3, [rdx]
+    mov     rdx, arg(3) ; whiteclamp
+    movdqu  xmm4, [rdx]
+    mov     rdx, arg(4) ; bothclamp
+    movdqu  xmm5, [rdx]
+
+.addnoise_loop:
+    call sym(LIBVPX_RAND) WRT_PLT
+    mov     rcx, arg(1) ;noise
+    and     rax, 0xff
+    add     rcx, rax
+
+    mov     rdi, rcx
+    movsxd  rcx, dword arg(5) ;[Width]
+    mov     rsi, arg(0) ;Pos
+    xor         rax,rax
+
+.addnoise_nextset:
+      movdqu      xmm1,[rsi+rax]         ; get the source
+
+      psubusb     xmm1, xmm3 ; subtract black clamp
+      paddusb     xmm1, xmm5 ; add both clamp
+      psubusb     xmm1, xmm4 ; subtract whiteclamp
+
+      movdqu      xmm2,[rdi+rax]         ; get the noise for this line
+      paddb       xmm1,xmm2              ; add it in
+      movdqu      [rsi+rax],xmm1         ; store the result
+
+      add         rax,16                 ; move to the next line
+
+      cmp         rax, rcx
+      jl          .addnoise_nextset
+
+    movsxd  rax, dword arg(7) ; Pitch
+    add     arg(0), rax ; Start += Pitch
+    sub     dword arg(6), 1   ; Height -= 1
+    jg      .addnoise_loop
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+rd42:
+    times 8 dw 0x04
+four8s:
+    times 4 dd 8
diff --git a/libvpx/libvpx/vpx_dsp/x86/avg_intrin_sse2.c b/libvpx/libvpx/vpx_dsp/x86/avg_intrin_sse2.c
new file mode 100644
index 0000000..f9af6cf
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/avg_intrin_sse2.c
@@ -0,0 +1,423 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+
+void vpx_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp,
+                         int *min, int *max) {
+  __m128i u0, s0, d0, diff, maxabsdiff, minabsdiff, negdiff, absdiff0, absdiff;
+  u0  = _mm_setzero_si128();
+  // Row 0
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff0 = _mm_max_epi16(diff, negdiff);
+  // Row 1
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(absdiff0, absdiff);
+  minabsdiff = _mm_min_epi16(absdiff0, absdiff);
+  // Row 2
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 2 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 3
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 3 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 4
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 4 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 5
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 5 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 6
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 6 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 7
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 7 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_si128(maxabsdiff, 8));
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 32));
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 16));
+  *max = _mm_extract_epi16(maxabsdiff, 0);
+
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_si128(minabsdiff, 8));
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 32));
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 16));
+  *min = _mm_extract_epi16(minabsdiff, 0);
+}
+
+unsigned int vpx_avg_8x8_sse2(const uint8_t *s, int p) {
+  __m128i s0, s1, u0;
+  unsigned int avg = 0;
+  u0  = _mm_setzero_si128();
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  s0 = _mm_adds_epu16(s0, _mm_srli_si128(s0, 8));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 32));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 16));
+  avg = _mm_extract_epi16(s0, 0);
+  return (avg + 32) >> 6;
+}
+
+unsigned int vpx_avg_4x4_sse2(const uint8_t *s, int p) {
+  __m128i s0, s1, u0;
+  unsigned int avg = 0;
+  u0  = _mm_setzero_si128();
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  s0 = _mm_adds_epu16(s0, _mm_srli_si128(s0, 4));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 16));
+  avg = _mm_extract_epi16(s0, 0);
+  return (avg + 8) >> 4;
+}
+
+static void hadamard_col8_sse2(__m128i *in, int iter) {
+  __m128i a0 = in[0];
+  __m128i a1 = in[1];
+  __m128i a2 = in[2];
+  __m128i a3 = in[3];
+  __m128i a4 = in[4];
+  __m128i a5 = in[5];
+  __m128i a6 = in[6];
+  __m128i a7 = in[7];
+
+  __m128i b0 = _mm_add_epi16(a0, a1);
+  __m128i b1 = _mm_sub_epi16(a0, a1);
+  __m128i b2 = _mm_add_epi16(a2, a3);
+  __m128i b3 = _mm_sub_epi16(a2, a3);
+  __m128i b4 = _mm_add_epi16(a4, a5);
+  __m128i b5 = _mm_sub_epi16(a4, a5);
+  __m128i b6 = _mm_add_epi16(a6, a7);
+  __m128i b7 = _mm_sub_epi16(a6, a7);
+
+  a0 = _mm_add_epi16(b0, b2);
+  a1 = _mm_add_epi16(b1, b3);
+  a2 = _mm_sub_epi16(b0, b2);
+  a3 = _mm_sub_epi16(b1, b3);
+  a4 = _mm_add_epi16(b4, b6);
+  a5 = _mm_add_epi16(b5, b7);
+  a6 = _mm_sub_epi16(b4, b6);
+  a7 = _mm_sub_epi16(b5, b7);
+
+  if (iter == 0) {
+    b0 = _mm_add_epi16(a0, a4);
+    b7 = _mm_add_epi16(a1, a5);
+    b3 = _mm_add_epi16(a2, a6);
+    b4 = _mm_add_epi16(a3, a7);
+    b2 = _mm_sub_epi16(a0, a4);
+    b6 = _mm_sub_epi16(a1, a5);
+    b1 = _mm_sub_epi16(a2, a6);
+    b5 = _mm_sub_epi16(a3, a7);
+
+    a0 = _mm_unpacklo_epi16(b0, b1);
+    a1 = _mm_unpacklo_epi16(b2, b3);
+    a2 = _mm_unpackhi_epi16(b0, b1);
+    a3 = _mm_unpackhi_epi16(b2, b3);
+    a4 = _mm_unpacklo_epi16(b4, b5);
+    a5 = _mm_unpacklo_epi16(b6, b7);
+    a6 = _mm_unpackhi_epi16(b4, b5);
+    a7 = _mm_unpackhi_epi16(b6, b7);
+
+    b0 = _mm_unpacklo_epi32(a0, a1);
+    b1 = _mm_unpacklo_epi32(a4, a5);
+    b2 = _mm_unpackhi_epi32(a0, a1);
+    b3 = _mm_unpackhi_epi32(a4, a5);
+    b4 = _mm_unpacklo_epi32(a2, a3);
+    b5 = _mm_unpacklo_epi32(a6, a7);
+    b6 = _mm_unpackhi_epi32(a2, a3);
+    b7 = _mm_unpackhi_epi32(a6, a7);
+
+    in[0] = _mm_unpacklo_epi64(b0, b1);
+    in[1] = _mm_unpackhi_epi64(b0, b1);
+    in[2] = _mm_unpacklo_epi64(b2, b3);
+    in[3] = _mm_unpackhi_epi64(b2, b3);
+    in[4] = _mm_unpacklo_epi64(b4, b5);
+    in[5] = _mm_unpackhi_epi64(b4, b5);
+    in[6] = _mm_unpacklo_epi64(b6, b7);
+    in[7] = _mm_unpackhi_epi64(b6, b7);
+  } else {
+    in[0] = _mm_add_epi16(a0, a4);
+    in[7] = _mm_add_epi16(a1, a5);
+    in[3] = _mm_add_epi16(a2, a6);
+    in[4] = _mm_add_epi16(a3, a7);
+    in[2] = _mm_sub_epi16(a0, a4);
+    in[6] = _mm_sub_epi16(a1, a5);
+    in[1] = _mm_sub_epi16(a2, a6);
+    in[5] = _mm_sub_epi16(a3, a7);
+  }
+}
+
+void vpx_hadamard_8x8_sse2(int16_t const *src_diff, int src_stride,
+                           int16_t *coeff) {
+  __m128i src[8];
+  src[0] = _mm_load_si128((const __m128i *)src_diff);
+  src[1] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[2] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[3] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[4] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[5] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[6] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[7] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+
+  hadamard_col8_sse2(src, 0);
+  hadamard_col8_sse2(src, 1);
+
+  _mm_store_si128((__m128i *)coeff, src[0]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[1]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[2]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[3]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[4]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[5]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[6]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[7]);
+}
+
+void vpx_hadamard_16x16_sse2(int16_t const *src_diff, int src_stride,
+                             int16_t *coeff) {
+  int idx;
+  for (idx = 0; idx < 4; ++idx) {
+    int16_t const *src_ptr = src_diff + (idx >> 1) * 8 * src_stride
+                                + (idx & 0x01) * 8;
+    vpx_hadamard_8x8_sse2(src_ptr, src_stride, coeff + idx * 64);
+  }
+
+  for (idx = 0; idx < 64; idx += 8) {
+    __m128i coeff0 = _mm_load_si128((const __m128i *)coeff);
+    __m128i coeff1 = _mm_load_si128((const __m128i *)(coeff + 64));
+    __m128i coeff2 = _mm_load_si128((const __m128i *)(coeff + 128));
+    __m128i coeff3 = _mm_load_si128((const __m128i *)(coeff + 192));
+
+    __m128i b0 = _mm_add_epi16(coeff0, coeff1);
+    __m128i b1 = _mm_sub_epi16(coeff0, coeff1);
+    __m128i b2 = _mm_add_epi16(coeff2, coeff3);
+    __m128i b3 = _mm_sub_epi16(coeff2, coeff3);
+
+    b0 = _mm_srai_epi16(b0, 1);
+    b1 = _mm_srai_epi16(b1, 1);
+    b2 = _mm_srai_epi16(b2, 1);
+    b3 = _mm_srai_epi16(b3, 1);
+
+    coeff0 = _mm_add_epi16(b0, b2);
+    coeff1 = _mm_add_epi16(b1, b3);
+    _mm_store_si128((__m128i *)coeff, coeff0);
+    _mm_store_si128((__m128i *)(coeff + 64), coeff1);
+
+    coeff2 = _mm_sub_epi16(b0, b2);
+    coeff3 = _mm_sub_epi16(b1, b3);
+    _mm_store_si128((__m128i *)(coeff + 128), coeff2);
+    _mm_store_si128((__m128i *)(coeff + 192), coeff3);
+
+    coeff += 8;
+  }
+}
+
+int vpx_satd_sse2(const int16_t *coeff, int length) {
+  int i;
+  const __m128i zero = _mm_setzero_si128();
+  __m128i accum = zero;
+
+  for (i = 0; i < length; i += 8) {
+    const __m128i src_line = _mm_load_si128((const __m128i *)coeff);
+    const __m128i inv = _mm_sub_epi16(zero, src_line);
+    const __m128i abs = _mm_max_epi16(src_line, inv);  // abs(src_line)
+    const __m128i abs_lo = _mm_unpacklo_epi16(abs, zero);
+    const __m128i abs_hi = _mm_unpackhi_epi16(abs, zero);
+    const __m128i sum = _mm_add_epi32(abs_lo, abs_hi);
+    accum = _mm_add_epi32(accum, sum);
+    coeff += 8;
+  }
+
+  {  // cascading summation of accum
+    __m128i hi = _mm_srli_si128(accum, 8);
+    accum = _mm_add_epi32(accum, hi);
+    hi = _mm_srli_epi64(accum, 32);
+    accum = _mm_add_epi32(accum, hi);
+  }
+
+  return _mm_cvtsi128_si32(accum);
+}
+
+void vpx_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,
+                          const int ref_stride, const int height) {
+  int idx;
+  __m128i zero = _mm_setzero_si128();
+  __m128i src_line = _mm_loadu_si128((const __m128i *)ref);
+  __m128i s0 = _mm_unpacklo_epi8(src_line, zero);
+  __m128i s1 = _mm_unpackhi_epi8(src_line, zero);
+  __m128i t0, t1;
+  int height_1 = height - 1;
+  ref += ref_stride;
+
+  for (idx = 1; idx < height_1; idx += 2) {
+    src_line = _mm_loadu_si128((const __m128i *)ref);
+    t0 = _mm_unpacklo_epi8(src_line, zero);
+    t1 = _mm_unpackhi_epi8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, t0);
+    s1 = _mm_adds_epu16(s1, t1);
+    ref += ref_stride;
+
+    src_line = _mm_loadu_si128((const __m128i *)ref);
+    t0 = _mm_unpacklo_epi8(src_line, zero);
+    t1 = _mm_unpackhi_epi8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, t0);
+    s1 = _mm_adds_epu16(s1, t1);
+    ref += ref_stride;
+  }
+
+  src_line = _mm_loadu_si128((const __m128i *)ref);
+  t0 = _mm_unpacklo_epi8(src_line, zero);
+  t1 = _mm_unpackhi_epi8(src_line, zero);
+  s0 = _mm_adds_epu16(s0, t0);
+  s1 = _mm_adds_epu16(s1, t1);
+
+  if (height == 64) {
+    s0 = _mm_srai_epi16(s0, 5);
+    s1 = _mm_srai_epi16(s1, 5);
+  } else if (height == 32) {
+    s0 = _mm_srai_epi16(s0, 4);
+    s1 = _mm_srai_epi16(s1, 4);
+  } else {
+    s0 = _mm_srai_epi16(s0, 3);
+    s1 = _mm_srai_epi16(s1, 3);
+  }
+
+  _mm_storeu_si128((__m128i *)hbuf, s0);
+  hbuf += 8;
+  _mm_storeu_si128((__m128i *)hbuf, s1);
+}
+
+int16_t vpx_int_pro_col_sse2(uint8_t const *ref, const int width) {
+  __m128i zero = _mm_setzero_si128();
+  __m128i src_line = _mm_load_si128((const __m128i *)ref);
+  __m128i s0 = _mm_sad_epu8(src_line, zero);
+  __m128i s1;
+  int i;
+
+  for (i = 16; i < width; i += 16) {
+    ref += 16;
+    src_line = _mm_load_si128((const __m128i *)ref);
+    s1 = _mm_sad_epu8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, s1);
+  }
+
+  s1 = _mm_srli_si128(s0, 8);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  return _mm_extract_epi16(s0, 0);
+}
+
+int vpx_vector_var_sse2(int16_t const *ref, int16_t const *src,
+                        const int bwl) {
+  int idx;
+  int width = 4 << bwl;
+  int16_t mean;
+  __m128i v0 = _mm_loadu_si128((const __m128i *)ref);
+  __m128i v1 = _mm_load_si128((const __m128i *)src);
+  __m128i diff = _mm_subs_epi16(v0, v1);
+  __m128i sum = diff;
+  __m128i sse = _mm_madd_epi16(diff, diff);
+
+  ref += 8;
+  src += 8;
+
+  for (idx = 8; idx < width; idx += 8) {
+    v0 = _mm_loadu_si128((const __m128i *)ref);
+    v1 = _mm_load_si128((const __m128i *)src);
+    diff = _mm_subs_epi16(v0, v1);
+
+    sum = _mm_add_epi16(sum, diff);
+    v0  = _mm_madd_epi16(diff, diff);
+    sse = _mm_add_epi32(sse, v0);
+
+    ref += 8;
+    src += 8;
+  }
+
+  v0  = _mm_srli_si128(sum, 8);
+  sum = _mm_add_epi16(sum, v0);
+  v0  = _mm_srli_epi64(sum, 32);
+  sum = _mm_add_epi16(sum, v0);
+  v0  = _mm_srli_epi32(sum, 16);
+  sum = _mm_add_epi16(sum, v0);
+
+  v1  = _mm_srli_si128(sse, 8);
+  sse = _mm_add_epi32(sse, v1);
+  v1  = _mm_srli_epi64(sse, 32);
+  sse = _mm_add_epi32(sse, v1);
+
+  mean = _mm_extract_epi16(sum, 0);
+
+  return _mm_cvtsi128_si32(sse) - ((mean * mean) >> (bwl + 2));
+}
diff --git a/libvpx/libvpx/vpx_dsp/x86/avg_ssse3_x86_64.asm b/libvpx/libvpx/vpx_dsp/x86/avg_ssse3_x86_64.asm
new file mode 100644
index 0000000..26412e8
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/avg_ssse3_x86_64.asm
@@ -0,0 +1,121 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%define private_prefix vpx
+
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the hadamard transformation. Part
+; of the macro definitions are originally derived from the ffmpeg project.
+; The current version applies to x86 64-bit only.
+
+SECTION .text
+
+%if ARCH_X86_64
+; matrix transpose
+%macro INTERLEAVE_2X 4
+  punpckh%1          m%4, m%2, m%3
+  punpckl%1          m%2, m%3
+  SWAP               %3,  %4
+%endmacro
+
+%macro TRANSPOSE8X8 9
+  INTERLEAVE_2X  wd, %1, %2, %9
+  INTERLEAVE_2X  wd, %3, %4, %9
+  INTERLEAVE_2X  wd, %5, %6, %9
+  INTERLEAVE_2X  wd, %7, %8, %9
+
+  INTERLEAVE_2X  dq, %1, %3, %9
+  INTERLEAVE_2X  dq, %2, %4, %9
+  INTERLEAVE_2X  dq, %5, %7, %9
+  INTERLEAVE_2X  dq, %6, %8, %9
+
+  INTERLEAVE_2X  qdq, %1, %5, %9
+  INTERLEAVE_2X  qdq, %3, %7, %9
+  INTERLEAVE_2X  qdq, %2, %6, %9
+  INTERLEAVE_2X  qdq, %4, %8, %9
+
+  SWAP  %2, %5
+  SWAP  %4, %7
+%endmacro
+
+%macro HMD8_1D 0
+  psubw              m8, m0, m1
+  psubw              m9, m2, m3
+  paddw              m0, m1
+  paddw              m2, m3
+  SWAP               1, 8
+  SWAP               3, 9
+  psubw              m8, m4, m5
+  psubw              m9, m6, m7
+  paddw              m4, m5
+  paddw              m6, m7
+  SWAP               5, 8
+  SWAP               7, 9
+
+  psubw              m8, m0, m2
+  psubw              m9, m1, m3
+  paddw              m0, m2
+  paddw              m1, m3
+  SWAP               2, 8
+  SWAP               3, 9
+  psubw              m8, m4, m6
+  psubw              m9, m5, m7
+  paddw              m4, m6
+  paddw              m5, m7
+  SWAP               6, 8
+  SWAP               7, 9
+
+  psubw              m8, m0, m4
+  psubw              m9, m1, m5
+  paddw              m0, m4
+  paddw              m1, m5
+  SWAP               4, 8
+  SWAP               5, 9
+  psubw              m8, m2, m6
+  psubw              m9, m3, m7
+  paddw              m2, m6
+  paddw              m3, m7
+  SWAP               6, 8
+  SWAP               7, 9
+%endmacro
+
+INIT_XMM ssse3
+cglobal hadamard_8x8, 3, 5, 10, input, stride, output
+  lea                r3, [2 * strideq]
+  lea                r4, [4 * strideq]
+
+  mova               m0, [inputq]
+  mova               m1, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m2, [inputq]
+  mova               m3, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m4, [inputq]
+  mova               m5, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m6, [inputq]
+  mova               m7, [inputq + r3]
+
+  HMD8_1D
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+  HMD8_1D
+
+  mova              [outputq +   0], m0
+  mova              [outputq +  16], m1
+  mova              [outputq +  32], m2
+  mova              [outputq +  48], m3
+  mova              [outputq +  64], m4
+  mova              [outputq +  80], m5
+  mova              [outputq +  96], m6
+  mova              [outputq + 112], m7
+
+  RET
+%endif
diff --git a/libvpx/libvpx/vpx_dsp/x86/convolve.h b/libvpx/libvpx/vpx_dsp/x86/convolve.h
new file mode 100644
index 0000000..7e43eb7
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/convolve.h
@@ -0,0 +1,274 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VPX_DSP_X86_CONVOLVE_H_
+#define VPX_DSP_X86_CONVOLVE_H_
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+typedef void filter8_1dfunction (
+  const uint8_t *src_ptr,
+  ptrdiff_t src_pitch,
+  uint8_t *output_ptr,
+  ptrdiff_t out_pitch,
+  uint32_t output_height,
+  const int16_t *filter
+);
+
+#define FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt) \
+  void vpx_convolve8_##name##_##opt(const uint8_t *src, ptrdiff_t src_stride, \
+                                    uint8_t *dst, ptrdiff_t dst_stride, \
+                                    const int16_t *filter_x, int x_step_q4, \
+                                    const int16_t *filter_y, int y_step_q4, \
+                                    int w, int h) { \
+  assert(filter[3] != 128); \
+  assert(step_q4 == 16); \
+  if (filter[0] | filter[1] | filter[2]) { \
+    while (w >= 16) { \
+      vpx_filter_block1d16_##dir##8_##avg##opt(src_start, \
+                                               src_stride, \
+                                               dst, \
+                                               dst_stride, \
+                                               h, \
+                                               filter); \
+      src += 16; \
+      dst += 16; \
+      w -= 16; \
+    } \
+    if (w == 8) { \
+      vpx_filter_block1d8_##dir##8_##avg##opt(src_start, \
+                                              src_stride, \
+                                              dst, \
+                                              dst_stride, \
+                                              h, \
+                                              filter); \
+    } else if (w == 4) { \
+      vpx_filter_block1d4_##dir##8_##avg##opt(src_start, \
+                                              src_stride, \
+                                              dst, \
+                                              dst_stride, \
+                                              h, \
+                                              filter); \
+    } \
+  } else { \
+    while (w >= 16) { \
+      vpx_filter_block1d16_##dir##2_##avg##opt(src, \
+                                               src_stride, \
+                                               dst, \
+                                               dst_stride, \
+                                               h, \
+                                               filter); \
+      src += 16; \
+      dst += 16; \
+      w -= 16; \
+    } \
+    if (w == 8) { \
+      vpx_filter_block1d8_##dir##2_##avg##opt(src, \
+                                              src_stride, \
+                                              dst, \
+                                              dst_stride, \
+                                              h, \
+                                              filter); \
+    } else if (w == 4) { \
+      vpx_filter_block1d4_##dir##2_##avg##opt(src, \
+                                              src_stride, \
+                                              dst, \
+                                              dst_stride, \
+                                              h, \
+                                              filter); \
+    } \
+  } \
+}
+
+#define FUN_CONV_2D(avg, opt) \
+void vpx_convolve8_##avg##opt(const uint8_t *src, ptrdiff_t src_stride, \
+                              uint8_t *dst, ptrdiff_t dst_stride, \
+                              const int16_t *filter_x, int x_step_q4, \
+                              const int16_t *filter_y, int y_step_q4, \
+                              int w, int h) { \
+  assert(filter_x[3] != 128); \
+  assert(filter_y[3] != 128); \
+  assert(w <= 64); \
+  assert(h <= 64); \
+  assert(x_step_q4 == 16); \
+  assert(y_step_q4 == 16); \
+  if (filter_x[0] | filter_x[1] | filter_x[2]) { \
+    DECLARE_ALIGNED(16, uint8_t, fdata2[64 * 71]); \
+    vpx_convolve8_horiz_##opt(src - 3 * src_stride, src_stride, fdata2, 64, \
+                              filter_x, x_step_q4, filter_y, y_step_q4, \
+                              w, h + 7); \
+    vpx_convolve8_##avg##vert_##opt(fdata2 + 3 * 64, 64, dst, dst_stride, \
+                                    filter_x, x_step_q4, filter_y, \
+                                    y_step_q4, w, h); \
+  } else { \
+    DECLARE_ALIGNED(16, uint8_t, fdata2[64 * 65]); \
+    vpx_convolve8_horiz_##opt(src, src_stride, fdata2, 64, \
+                              filter_x, x_step_q4, filter_y, y_step_q4, \
+                              w, h + 1); \
+    vpx_convolve8_##avg##vert_##opt(fdata2, 64, dst, dst_stride, \
+                                    filter_x, x_step_q4, filter_y, \
+                                    y_step_q4, w, h); \
+  } \
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+
+typedef void highbd_filter8_1dfunction (
+  const uint16_t *src_ptr,
+  const ptrdiff_t src_pitch,
+  uint16_t *output_ptr,
+  ptrdiff_t out_pitch,
+  unsigned int output_height,
+  const int16_t *filter,
+  int bd
+);
+
+#define HIGH_FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt) \
+  void vpx_highbd_convolve8_##name##_##opt(const uint8_t *src8, \
+                                           ptrdiff_t src_stride, \
+                                           uint8_t *dst8, \
+                                           ptrdiff_t dst_stride, \
+                                           const int16_t *filter_x, \
+                                           int x_step_q4, \
+                                           const int16_t *filter_y, \
+                                           int y_step_q4, \
+                                           int w, int h, int bd) { \
+  if (step_q4 == 16 && filter[3] != 128) { \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+    if (filter[0] | filter[1] | filter[2]) { \
+      while (w >= 16) { \
+        vpx_highbd_filter_block1d16_##dir##8_##avg##opt(src_start, \
+                                                        src_stride, \
+                                                        dst, \
+                                                        dst_stride, \
+                                                        h, \
+                                                        filter, \
+                                                        bd); \
+        src += 16; \
+        dst += 16; \
+        w -= 16; \
+      } \
+      while (w >= 8) { \
+        vpx_highbd_filter_block1d8_##dir##8_##avg##opt(src_start, \
+                                                       src_stride, \
+                                                       dst, \
+                                                       dst_stride, \
+                                                       h, \
+                                                       filter, \
+                                                       bd); \
+        src += 8; \
+        dst += 8; \
+        w -= 8; \
+      } \
+      while (w >= 4) { \
+        vpx_highbd_filter_block1d4_##dir##8_##avg##opt(src_start, \
+                                                       src_stride, \
+                                                       dst, \
+                                                       dst_stride, \
+                                                       h, \
+                                                       filter, \
+                                                       bd); \
+        src += 4; \
+        dst += 4; \
+        w -= 4; \
+      } \
+    } else { \
+      while (w >= 16) { \
+        vpx_highbd_filter_block1d16_##dir##2_##avg##opt(src, \
+                                                        src_stride, \
+                                                        dst, \
+                                                        dst_stride, \
+                                                        h, \
+                                                        filter, \
+                                                        bd); \
+        src += 16; \
+        dst += 16; \
+        w -= 16; \
+      } \
+      while (w >= 8) { \
+        vpx_highbd_filter_block1d8_##dir##2_##avg##opt(src, \
+                                                       src_stride, \
+                                                       dst, \
+                                                       dst_stride, \
+                                                       h, \
+                                                       filter, \
+                                                       bd); \
+        src += 8; \
+        dst += 8; \
+        w -= 8; \
+      } \
+      while (w >= 4) { \
+        vpx_highbd_filter_block1d4_##dir##2_##avg##opt(src, \
+                                                       src_stride, \
+                                                       dst, \
+                                                       dst_stride, \
+                                                       h, \
+                                                       filter, \
+                                                       bd); \
+        src += 4; \
+        dst += 4; \
+        w -= 4; \
+      } \
+    } \
+  } \
+  if (w) { \
+    vpx_highbd_convolve8_##name##_c(src8, src_stride, dst8, dst_stride, \
+                                    filter_x, x_step_q4, filter_y, y_step_q4, \
+                                    w, h, bd); \
+  } \
+}
+
+#define HIGH_FUN_CONV_2D(avg, opt) \
+void vpx_highbd_convolve8_##avg##opt(const uint8_t *src, ptrdiff_t src_stride, \
+                                     uint8_t *dst, ptrdiff_t dst_stride, \
+                                     const int16_t *filter_x, int x_step_q4, \
+                                     const int16_t *filter_y, int y_step_q4, \
+                                     int w, int h, int bd) { \
+  assert(w <= 64); \
+  assert(h <= 64); \
+  if (x_step_q4 == 16 && y_step_q4 == 16) { \
+    if ((filter_x[0] | filter_x[1] | filter_x[2]) || filter_x[3] == 128) { \
+      DECLARE_ALIGNED(16, uint16_t, fdata2[64 * 71]); \
+      vpx_highbd_convolve8_horiz_##opt(src - 3 * src_stride, src_stride, \
+                                       CONVERT_TO_BYTEPTR(fdata2), 64, \
+                                       filter_x, x_step_q4, \
+                                       filter_y, y_step_q4, \
+                                       w, h + 7, bd); \
+      vpx_highbd_convolve8_##avg##vert_##opt(CONVERT_TO_BYTEPTR(fdata2) + 192, \
+                                             64, dst, dst_stride, \
+                                             filter_x, x_step_q4, \
+                                             filter_y, y_step_q4, \
+                                             w, h, bd); \
+    } else { \
+      DECLARE_ALIGNED(16, uint16_t, fdata2[64 * 65]); \
+      vpx_highbd_convolve8_horiz_##opt(src, src_stride, \
+                                       CONVERT_TO_BYTEPTR(fdata2), 64, \
+                                       filter_x, x_step_q4, \
+                                       filter_y, y_step_q4, \
+                                       w, h + 1, bd); \
+      vpx_highbd_convolve8_##avg##vert_##opt(CONVERT_TO_BYTEPTR(fdata2), 64, \
+                                             dst, dst_stride, \
+                                             filter_x, x_step_q4, \
+                                             filter_y, y_step_q4, \
+                                             w, h, bd); \
+    } \
+  } else { \
+    vpx_highbd_convolve8_##avg##c(src, src_stride, dst, dst_stride, \
+                                  filter_x, x_step_q4, filter_y, y_step_q4, w, \
+                                  h, bd); \
+  } \
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#endif  // VPX_DSP_X86_CONVOLVE_H_
diff --git a/libvpx/libvpx/vpx_dsp/x86/fwd_dct32x32_impl_avx2.h b/libvpx/libvpx/vpx_dsp/x86/fwd_dct32x32_impl_avx2.h
new file mode 100644
index 0000000..951af3a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/fwd_dct32x32_impl_avx2.h
@@ -0,0 +1,2712 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/txfm_common.h"
+
+#define pair256_set_epi16(a, b) \
+  _mm256_set_epi16((int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                   (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                   (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                   (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a))
+
+#define pair256_set_epi32(a, b) \
+  _mm256_set_epi32((int)(b), (int)(a), (int)(b), (int)(a), \
+                   (int)(b), (int)(a), (int)(b), (int)(a))
+
+#if FDCT32x32_HIGH_PRECISION
+static INLINE __m256i k_madd_epi32_avx2(__m256i a, __m256i b) {
+  __m256i buf0, buf1;
+  buf0 = _mm256_mul_epu32(a, b);
+  a = _mm256_srli_epi64(a, 32);
+  b = _mm256_srli_epi64(b, 32);
+  buf1 = _mm256_mul_epu32(a, b);
+  return _mm256_add_epi64(buf0, buf1);
+}
+
+static INLINE __m256i k_packs_epi64_avx2(__m256i a, __m256i b) {
+  __m256i buf0 = _mm256_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
+  __m256i buf1 = _mm256_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
+  return _mm256_unpacklo_epi64(buf0, buf1);
+}
+#endif
+
+void FDCT32x32_2D_AVX2(const int16_t *input,
+                  int16_t *output_org, int stride) {
+  // Calculate pre-multiplied strides
+  const int str1 = stride;
+  const int str2 = 2 * stride;
+  const int str3 = 2 * stride + str1;
+  // We need an intermediate buffer between passes.
+  DECLARE_ALIGNED(32, int16_t, intermediate[32 * 32]);
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m256i k__cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64);
+  const __m256i k__cospi_p16_m16 = pair256_set_epi16(+cospi_16_64, -cospi_16_64);
+  const __m256i k__cospi_m08_p24 = pair256_set_epi16(-cospi_8_64,   cospi_24_64);
+  const __m256i k__cospi_m24_m08 = pair256_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m256i k__cospi_p24_p08 = pair256_set_epi16(+cospi_24_64,  cospi_8_64);
+  const __m256i k__cospi_p12_p20 = pair256_set_epi16(+cospi_12_64,  cospi_20_64);
+  const __m256i k__cospi_m20_p12 = pair256_set_epi16(-cospi_20_64,  cospi_12_64);
+  const __m256i k__cospi_m04_p28 = pair256_set_epi16(-cospi_4_64,   cospi_28_64);
+  const __m256i k__cospi_p28_p04 = pair256_set_epi16(+cospi_28_64,  cospi_4_64);
+  const __m256i k__cospi_m28_m04 = pair256_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m256i k__cospi_m12_m20 = pair256_set_epi16(-cospi_12_64, -cospi_20_64);
+  const __m256i k__cospi_p30_p02 = pair256_set_epi16(+cospi_30_64,  cospi_2_64);
+  const __m256i k__cospi_p14_p18 = pair256_set_epi16(+cospi_14_64,  cospi_18_64);
+  const __m256i k__cospi_p22_p10 = pair256_set_epi16(+cospi_22_64,  cospi_10_64);
+  const __m256i k__cospi_p06_p26 = pair256_set_epi16(+cospi_6_64,   cospi_26_64);
+  const __m256i k__cospi_m26_p06 = pair256_set_epi16(-cospi_26_64,  cospi_6_64);
+  const __m256i k__cospi_m10_p22 = pair256_set_epi16(-cospi_10_64,  cospi_22_64);
+  const __m256i k__cospi_m18_p14 = pair256_set_epi16(-cospi_18_64,  cospi_14_64);
+  const __m256i k__cospi_m02_p30 = pair256_set_epi16(-cospi_2_64,   cospi_30_64);
+  const __m256i k__cospi_p31_p01 = pair256_set_epi16(+cospi_31_64,  cospi_1_64);
+  const __m256i k__cospi_p15_p17 = pair256_set_epi16(+cospi_15_64,  cospi_17_64);
+  const __m256i k__cospi_p23_p09 = pair256_set_epi16(+cospi_23_64,  cospi_9_64);
+  const __m256i k__cospi_p07_p25 = pair256_set_epi16(+cospi_7_64,   cospi_25_64);
+  const __m256i k__cospi_m25_p07 = pair256_set_epi16(-cospi_25_64,  cospi_7_64);
+  const __m256i k__cospi_m09_p23 = pair256_set_epi16(-cospi_9_64,   cospi_23_64);
+  const __m256i k__cospi_m17_p15 = pair256_set_epi16(-cospi_17_64,  cospi_15_64);
+  const __m256i k__cospi_m01_p31 = pair256_set_epi16(-cospi_1_64,   cospi_31_64);
+  const __m256i k__cospi_p27_p05 = pair256_set_epi16(+cospi_27_64,  cospi_5_64);
+  const __m256i k__cospi_p11_p21 = pair256_set_epi16(+cospi_11_64,  cospi_21_64);
+  const __m256i k__cospi_p19_p13 = pair256_set_epi16(+cospi_19_64,  cospi_13_64);
+  const __m256i k__cospi_p03_p29 = pair256_set_epi16(+cospi_3_64,   cospi_29_64);
+  const __m256i k__cospi_m29_p03 = pair256_set_epi16(-cospi_29_64,  cospi_3_64);
+  const __m256i k__cospi_m13_p19 = pair256_set_epi16(-cospi_13_64,  cospi_19_64);
+  const __m256i k__cospi_m21_p11 = pair256_set_epi16(-cospi_21_64,  cospi_11_64);
+  const __m256i k__cospi_m05_p27 = pair256_set_epi16(-cospi_5_64,   cospi_27_64);
+  const __m256i k__DCT_CONST_ROUNDING = _mm256_set1_epi32(DCT_CONST_ROUNDING);
+  const __m256i kZero = _mm256_set1_epi16(0);
+  const __m256i kOne  = _mm256_set1_epi16(1);
+  // Do the two transform/transpose passes
+  int pass;
+  for (pass = 0; pass < 2; ++pass) {
+    // We process sixteen columns (transposed rows in second pass) at a time.
+    int column_start;
+    for (column_start = 0; column_start < 32; column_start += 16) {
+      __m256i step1[32];
+      __m256i step2[32];
+      __m256i step3[32];
+      __m256i out[32];
+      // Stage 1
+      // Note: even though all the loads below are aligned, using the aligned
+      //       intrinsic make the code slightly slower.
+      if (0 == pass) {
+        const int16_t *in  = &input[column_start];
+        // step1[i] =  (in[ 0 * stride] + in[(32 -  1) * stride]) << 2;
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          const int16_t *ina =  in +  0 * str1;
+          const int16_t *inb =  in + 31 * str1;
+          __m256i *step1a = &step1[ 0];
+          __m256i *step1b = &step1[31];
+          const __m256i ina0  = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1  = _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2  = _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3  = _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3  = _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2  = _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1  = _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0  = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[ 0] = _mm256_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm256_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm256_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm256_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm256_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm256_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm256_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in +  4 * str1;
+          const int16_t *inb =  in + 27 * str1;
+          __m256i *step1a = &step1[ 4];
+          __m256i *step1b = &step1[27];
+          const __m256i ina0  = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1  = _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2  = _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3  = _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3  = _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2  = _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1  = _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0  = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[ 0] = _mm256_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm256_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm256_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm256_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm256_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm256_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm256_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in +  8 * str1;
+          const int16_t *inb =  in + 23 * str1;
+          __m256i *step1a = &step1[ 8];
+          __m256i *step1b = &step1[23];
+          const __m256i ina0  = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1  = _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2  = _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3  = _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3  = _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2  = _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1  = _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0  = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[ 0] = _mm256_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm256_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm256_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm256_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm256_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm256_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm256_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in + 12 * str1;
+          const int16_t *inb =  in + 19 * str1;
+          __m256i *step1a = &step1[12];
+          __m256i *step1b = &step1[19];
+          const __m256i ina0  = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1  = _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2  = _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3  = _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3  = _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2  = _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1  = _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0  = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[ 0] = _mm256_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm256_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm256_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm256_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm256_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm256_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm256_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+      } else {
+        int16_t *in = &intermediate[column_start];
+        // step1[i] =  in[ 0 * 32] + in[(32 -  1) * 32];
+        // Note: using the same approach as above to have common offset is
+        //       counter-productive as all offsets can be calculated at compile
+        //       time.
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          __m256i in00  = _mm256_loadu_si256((const __m256i *)(in +  0 * 32));
+          __m256i in01  = _mm256_loadu_si256((const __m256i *)(in +  1 * 32));
+          __m256i in02  = _mm256_loadu_si256((const __m256i *)(in +  2 * 32));
+          __m256i in03  = _mm256_loadu_si256((const __m256i *)(in +  3 * 32));
+          __m256i in28  = _mm256_loadu_si256((const __m256i *)(in + 28 * 32));
+          __m256i in29  = _mm256_loadu_si256((const __m256i *)(in + 29 * 32));
+          __m256i in30  = _mm256_loadu_si256((const __m256i *)(in + 30 * 32));
+          __m256i in31  = _mm256_loadu_si256((const __m256i *)(in + 31 * 32));
+          step1[ 0] = _mm256_add_epi16(in00, in31);
+          step1[ 1] = _mm256_add_epi16(in01, in30);
+          step1[ 2] = _mm256_add_epi16(in02, in29);
+          step1[ 3] = _mm256_add_epi16(in03, in28);
+          step1[28] = _mm256_sub_epi16(in03, in28);
+          step1[29] = _mm256_sub_epi16(in02, in29);
+          step1[30] = _mm256_sub_epi16(in01, in30);
+          step1[31] = _mm256_sub_epi16(in00, in31);
+        }
+        {
+          __m256i in04  = _mm256_loadu_si256((const __m256i *)(in +  4 * 32));
+          __m256i in05  = _mm256_loadu_si256((const __m256i *)(in +  5 * 32));
+          __m256i in06  = _mm256_loadu_si256((const __m256i *)(in +  6 * 32));
+          __m256i in07  = _mm256_loadu_si256((const __m256i *)(in +  7 * 32));
+          __m256i in24  = _mm256_loadu_si256((const __m256i *)(in + 24 * 32));
+          __m256i in25  = _mm256_loadu_si256((const __m256i *)(in + 25 * 32));
+          __m256i in26  = _mm256_loadu_si256((const __m256i *)(in + 26 * 32));
+          __m256i in27  = _mm256_loadu_si256((const __m256i *)(in + 27 * 32));
+          step1[ 4] = _mm256_add_epi16(in04, in27);
+          step1[ 5] = _mm256_add_epi16(in05, in26);
+          step1[ 6] = _mm256_add_epi16(in06, in25);
+          step1[ 7] = _mm256_add_epi16(in07, in24);
+          step1[24] = _mm256_sub_epi16(in07, in24);
+          step1[25] = _mm256_sub_epi16(in06, in25);
+          step1[26] = _mm256_sub_epi16(in05, in26);
+          step1[27] = _mm256_sub_epi16(in04, in27);
+        }
+        {
+          __m256i in08  = _mm256_loadu_si256((const __m256i *)(in +  8 * 32));
+          __m256i in09  = _mm256_loadu_si256((const __m256i *)(in +  9 * 32));
+          __m256i in10  = _mm256_loadu_si256((const __m256i *)(in + 10 * 32));
+          __m256i in11  = _mm256_loadu_si256((const __m256i *)(in + 11 * 32));
+          __m256i in20  = _mm256_loadu_si256((const __m256i *)(in + 20 * 32));
+          __m256i in21  = _mm256_loadu_si256((const __m256i *)(in + 21 * 32));
+          __m256i in22  = _mm256_loadu_si256((const __m256i *)(in + 22 * 32));
+          __m256i in23  = _mm256_loadu_si256((const __m256i *)(in + 23 * 32));
+          step1[ 8] = _mm256_add_epi16(in08, in23);
+          step1[ 9] = _mm256_add_epi16(in09, in22);
+          step1[10] = _mm256_add_epi16(in10, in21);
+          step1[11] = _mm256_add_epi16(in11, in20);
+          step1[20] = _mm256_sub_epi16(in11, in20);
+          step1[21] = _mm256_sub_epi16(in10, in21);
+          step1[22] = _mm256_sub_epi16(in09, in22);
+          step1[23] = _mm256_sub_epi16(in08, in23);
+        }
+        {
+          __m256i in12  = _mm256_loadu_si256((const __m256i *)(in + 12 * 32));
+          __m256i in13  = _mm256_loadu_si256((const __m256i *)(in + 13 * 32));
+          __m256i in14  = _mm256_loadu_si256((const __m256i *)(in + 14 * 32));
+          __m256i in15  = _mm256_loadu_si256((const __m256i *)(in + 15 * 32));
+          __m256i in16  = _mm256_loadu_si256((const __m256i *)(in + 16 * 32));
+          __m256i in17  = _mm256_loadu_si256((const __m256i *)(in + 17 * 32));
+          __m256i in18  = _mm256_loadu_si256((const __m256i *)(in + 18 * 32));
+          __m256i in19  = _mm256_loadu_si256((const __m256i *)(in + 19 * 32));
+          step1[12] = _mm256_add_epi16(in12, in19);
+          step1[13] = _mm256_add_epi16(in13, in18);
+          step1[14] = _mm256_add_epi16(in14, in17);
+          step1[15] = _mm256_add_epi16(in15, in16);
+          step1[16] = _mm256_sub_epi16(in15, in16);
+          step1[17] = _mm256_sub_epi16(in14, in17);
+          step1[18] = _mm256_sub_epi16(in13, in18);
+          step1[19] = _mm256_sub_epi16(in12, in19);
+        }
+      }
+      // Stage 2
+      {
+        step2[ 0] = _mm256_add_epi16(step1[0], step1[15]);
+        step2[ 1] = _mm256_add_epi16(step1[1], step1[14]);
+        step2[ 2] = _mm256_add_epi16(step1[2], step1[13]);
+        step2[ 3] = _mm256_add_epi16(step1[3], step1[12]);
+        step2[ 4] = _mm256_add_epi16(step1[4], step1[11]);
+        step2[ 5] = _mm256_add_epi16(step1[5], step1[10]);
+        step2[ 6] = _mm256_add_epi16(step1[6], step1[ 9]);
+        step2[ 7] = _mm256_add_epi16(step1[7], step1[ 8]);
+        step2[ 8] = _mm256_sub_epi16(step1[7], step1[ 8]);
+        step2[ 9] = _mm256_sub_epi16(step1[6], step1[ 9]);
+        step2[10] = _mm256_sub_epi16(step1[5], step1[10]);
+        step2[11] = _mm256_sub_epi16(step1[4], step1[11]);
+        step2[12] = _mm256_sub_epi16(step1[3], step1[12]);
+        step2[13] = _mm256_sub_epi16(step1[2], step1[13]);
+        step2[14] = _mm256_sub_epi16(step1[1], step1[14]);
+        step2[15] = _mm256_sub_epi16(step1[0], step1[15]);
+      }
+      {
+        const __m256i s2_20_0 = _mm256_unpacklo_epi16(step1[27], step1[20]);
+        const __m256i s2_20_1 = _mm256_unpackhi_epi16(step1[27], step1[20]);
+        const __m256i s2_21_0 = _mm256_unpacklo_epi16(step1[26], step1[21]);
+        const __m256i s2_21_1 = _mm256_unpackhi_epi16(step1[26], step1[21]);
+        const __m256i s2_22_0 = _mm256_unpacklo_epi16(step1[25], step1[22]);
+        const __m256i s2_22_1 = _mm256_unpackhi_epi16(step1[25], step1[22]);
+        const __m256i s2_23_0 = _mm256_unpacklo_epi16(step1[24], step1[23]);
+        const __m256i s2_23_1 = _mm256_unpackhi_epi16(step1[24], step1[23]);
+        const __m256i s2_20_2 = _mm256_madd_epi16(s2_20_0, k__cospi_p16_m16);
+        const __m256i s2_20_3 = _mm256_madd_epi16(s2_20_1, k__cospi_p16_m16);
+        const __m256i s2_21_2 = _mm256_madd_epi16(s2_21_0, k__cospi_p16_m16);
+        const __m256i s2_21_3 = _mm256_madd_epi16(s2_21_1, k__cospi_p16_m16);
+        const __m256i s2_22_2 = _mm256_madd_epi16(s2_22_0, k__cospi_p16_m16);
+        const __m256i s2_22_3 = _mm256_madd_epi16(s2_22_1, k__cospi_p16_m16);
+        const __m256i s2_23_2 = _mm256_madd_epi16(s2_23_0, k__cospi_p16_m16);
+        const __m256i s2_23_3 = _mm256_madd_epi16(s2_23_1, k__cospi_p16_m16);
+        const __m256i s2_24_2 = _mm256_madd_epi16(s2_23_0, k__cospi_p16_p16);
+        const __m256i s2_24_3 = _mm256_madd_epi16(s2_23_1, k__cospi_p16_p16);
+        const __m256i s2_25_2 = _mm256_madd_epi16(s2_22_0, k__cospi_p16_p16);
+        const __m256i s2_25_3 = _mm256_madd_epi16(s2_22_1, k__cospi_p16_p16);
+        const __m256i s2_26_2 = _mm256_madd_epi16(s2_21_0, k__cospi_p16_p16);
+        const __m256i s2_26_3 = _mm256_madd_epi16(s2_21_1, k__cospi_p16_p16);
+        const __m256i s2_27_2 = _mm256_madd_epi16(s2_20_0, k__cospi_p16_p16);
+        const __m256i s2_27_3 = _mm256_madd_epi16(s2_20_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m256i s2_20_4 = _mm256_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_20_5 = _mm256_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_21_4 = _mm256_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_21_5 = _mm256_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_22_4 = _mm256_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_22_5 = _mm256_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_23_4 = _mm256_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_23_5 = _mm256_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_24_4 = _mm256_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_24_5 = _mm256_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_25_4 = _mm256_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_25_5 = _mm256_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_26_4 = _mm256_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_26_5 = _mm256_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_27_4 = _mm256_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_27_5 = _mm256_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_20_6 = _mm256_srai_epi32(s2_20_4, DCT_CONST_BITS);
+        const __m256i s2_20_7 = _mm256_srai_epi32(s2_20_5, DCT_CONST_BITS);
+        const __m256i s2_21_6 = _mm256_srai_epi32(s2_21_4, DCT_CONST_BITS);
+        const __m256i s2_21_7 = _mm256_srai_epi32(s2_21_5, DCT_CONST_BITS);
+        const __m256i s2_22_6 = _mm256_srai_epi32(s2_22_4, DCT_CONST_BITS);
+        const __m256i s2_22_7 = _mm256_srai_epi32(s2_22_5, DCT_CONST_BITS);
+        const __m256i s2_23_6 = _mm256_srai_epi32(s2_23_4, DCT_CONST_BITS);
+        const __m256i s2_23_7 = _mm256_srai_epi32(s2_23_5, DCT_CONST_BITS);
+        const __m256i s2_24_6 = _mm256_srai_epi32(s2_24_4, DCT_CONST_BITS);
+        const __m256i s2_24_7 = _mm256_srai_epi32(s2_24_5, DCT_CONST_BITS);
+        const __m256i s2_25_6 = _mm256_srai_epi32(s2_25_4, DCT_CONST_BITS);
+        const __m256i s2_25_7 = _mm256_srai_epi32(s2_25_5, DCT_CONST_BITS);
+        const __m256i s2_26_6 = _mm256_srai_epi32(s2_26_4, DCT_CONST_BITS);
+        const __m256i s2_26_7 = _mm256_srai_epi32(s2_26_5, DCT_CONST_BITS);
+        const __m256i s2_27_6 = _mm256_srai_epi32(s2_27_4, DCT_CONST_BITS);
+        const __m256i s2_27_7 = _mm256_srai_epi32(s2_27_5, DCT_CONST_BITS);
+        // Combine
+        step2[20] = _mm256_packs_epi32(s2_20_6, s2_20_7);
+        step2[21] = _mm256_packs_epi32(s2_21_6, s2_21_7);
+        step2[22] = _mm256_packs_epi32(s2_22_6, s2_22_7);
+        step2[23] = _mm256_packs_epi32(s2_23_6, s2_23_7);
+        step2[24] = _mm256_packs_epi32(s2_24_6, s2_24_7);
+        step2[25] = _mm256_packs_epi32(s2_25_6, s2_25_7);
+        step2[26] = _mm256_packs_epi32(s2_26_6, s2_26_7);
+        step2[27] = _mm256_packs_epi32(s2_27_6, s2_27_7);
+      }
+
+#if !FDCT32x32_HIGH_PRECISION
+      // dump the magnitude by half, hence the intermediate values are within
+      // the range of 16 bits.
+      if (1 == pass) {
+        __m256i s3_00_0 = _mm256_cmpgt_epi16(kZero,step2[ 0]);
+        __m256i s3_01_0 = _mm256_cmpgt_epi16(kZero,step2[ 1]);
+        __m256i s3_02_0 = _mm256_cmpgt_epi16(kZero,step2[ 2]);
+        __m256i s3_03_0 = _mm256_cmpgt_epi16(kZero,step2[ 3]);
+        __m256i s3_04_0 = _mm256_cmpgt_epi16(kZero,step2[ 4]);
+        __m256i s3_05_0 = _mm256_cmpgt_epi16(kZero,step2[ 5]);
+        __m256i s3_06_0 = _mm256_cmpgt_epi16(kZero,step2[ 6]);
+        __m256i s3_07_0 = _mm256_cmpgt_epi16(kZero,step2[ 7]);
+        __m256i s2_08_0 = _mm256_cmpgt_epi16(kZero,step2[ 8]);
+        __m256i s2_09_0 = _mm256_cmpgt_epi16(kZero,step2[ 9]);
+        __m256i s3_10_0 = _mm256_cmpgt_epi16(kZero,step2[10]);
+        __m256i s3_11_0 = _mm256_cmpgt_epi16(kZero,step2[11]);
+        __m256i s3_12_0 = _mm256_cmpgt_epi16(kZero,step2[12]);
+        __m256i s3_13_0 = _mm256_cmpgt_epi16(kZero,step2[13]);
+        __m256i s2_14_0 = _mm256_cmpgt_epi16(kZero,step2[14]);
+        __m256i s2_15_0 = _mm256_cmpgt_epi16(kZero,step2[15]);
+        __m256i s3_16_0 = _mm256_cmpgt_epi16(kZero,step1[16]);
+        __m256i s3_17_0 = _mm256_cmpgt_epi16(kZero,step1[17]);
+        __m256i s3_18_0 = _mm256_cmpgt_epi16(kZero,step1[18]);
+        __m256i s3_19_0 = _mm256_cmpgt_epi16(kZero,step1[19]);
+        __m256i s3_20_0 = _mm256_cmpgt_epi16(kZero,step2[20]);
+        __m256i s3_21_0 = _mm256_cmpgt_epi16(kZero,step2[21]);
+        __m256i s3_22_0 = _mm256_cmpgt_epi16(kZero,step2[22]);
+        __m256i s3_23_0 = _mm256_cmpgt_epi16(kZero,step2[23]);
+        __m256i s3_24_0 = _mm256_cmpgt_epi16(kZero,step2[24]);
+        __m256i s3_25_0 = _mm256_cmpgt_epi16(kZero,step2[25]);
+        __m256i s3_26_0 = _mm256_cmpgt_epi16(kZero,step2[26]);
+        __m256i s3_27_0 = _mm256_cmpgt_epi16(kZero,step2[27]);
+        __m256i s3_28_0 = _mm256_cmpgt_epi16(kZero,step1[28]);
+        __m256i s3_29_0 = _mm256_cmpgt_epi16(kZero,step1[29]);
+        __m256i s3_30_0 = _mm256_cmpgt_epi16(kZero,step1[30]);
+        __m256i s3_31_0 = _mm256_cmpgt_epi16(kZero,step1[31]);
+
+        step2[ 0] = _mm256_sub_epi16(step2[ 0], s3_00_0);
+        step2[ 1] = _mm256_sub_epi16(step2[ 1], s3_01_0);
+        step2[ 2] = _mm256_sub_epi16(step2[ 2], s3_02_0);
+        step2[ 3] = _mm256_sub_epi16(step2[ 3], s3_03_0);
+        step2[ 4] = _mm256_sub_epi16(step2[ 4], s3_04_0);
+        step2[ 5] = _mm256_sub_epi16(step2[ 5], s3_05_0);
+        step2[ 6] = _mm256_sub_epi16(step2[ 6], s3_06_0);
+        step2[ 7] = _mm256_sub_epi16(step2[ 7], s3_07_0);
+        step2[ 8] = _mm256_sub_epi16(step2[ 8], s2_08_0);
+        step2[ 9] = _mm256_sub_epi16(step2[ 9], s2_09_0);
+        step2[10] = _mm256_sub_epi16(step2[10], s3_10_0);
+        step2[11] = _mm256_sub_epi16(step2[11], s3_11_0);
+        step2[12] = _mm256_sub_epi16(step2[12], s3_12_0);
+        step2[13] = _mm256_sub_epi16(step2[13], s3_13_0);
+        step2[14] = _mm256_sub_epi16(step2[14], s2_14_0);
+        step2[15] = _mm256_sub_epi16(step2[15], s2_15_0);
+        step1[16] = _mm256_sub_epi16(step1[16], s3_16_0);
+        step1[17] = _mm256_sub_epi16(step1[17], s3_17_0);
+        step1[18] = _mm256_sub_epi16(step1[18], s3_18_0);
+        step1[19] = _mm256_sub_epi16(step1[19], s3_19_0);
+        step2[20] = _mm256_sub_epi16(step2[20], s3_20_0);
+        step2[21] = _mm256_sub_epi16(step2[21], s3_21_0);
+        step2[22] = _mm256_sub_epi16(step2[22], s3_22_0);
+        step2[23] = _mm256_sub_epi16(step2[23], s3_23_0);
+        step2[24] = _mm256_sub_epi16(step2[24], s3_24_0);
+        step2[25] = _mm256_sub_epi16(step2[25], s3_25_0);
+        step2[26] = _mm256_sub_epi16(step2[26], s3_26_0);
+        step2[27] = _mm256_sub_epi16(step2[27], s3_27_0);
+        step1[28] = _mm256_sub_epi16(step1[28], s3_28_0);
+        step1[29] = _mm256_sub_epi16(step1[29], s3_29_0);
+        step1[30] = _mm256_sub_epi16(step1[30], s3_30_0);
+        step1[31] = _mm256_sub_epi16(step1[31], s3_31_0);
+
+        step2[ 0] = _mm256_add_epi16(step2[ 0], kOne);
+        step2[ 1] = _mm256_add_epi16(step2[ 1], kOne);
+        step2[ 2] = _mm256_add_epi16(step2[ 2], kOne);
+        step2[ 3] = _mm256_add_epi16(step2[ 3], kOne);
+        step2[ 4] = _mm256_add_epi16(step2[ 4], kOne);
+        step2[ 5] = _mm256_add_epi16(step2[ 5], kOne);
+        step2[ 6] = _mm256_add_epi16(step2[ 6], kOne);
+        step2[ 7] = _mm256_add_epi16(step2[ 7], kOne);
+        step2[ 8] = _mm256_add_epi16(step2[ 8], kOne);
+        step2[ 9] = _mm256_add_epi16(step2[ 9], kOne);
+        step2[10] = _mm256_add_epi16(step2[10], kOne);
+        step2[11] = _mm256_add_epi16(step2[11], kOne);
+        step2[12] = _mm256_add_epi16(step2[12], kOne);
+        step2[13] = _mm256_add_epi16(step2[13], kOne);
+        step2[14] = _mm256_add_epi16(step2[14], kOne);
+        step2[15] = _mm256_add_epi16(step2[15], kOne);
+        step1[16] = _mm256_add_epi16(step1[16], kOne);
+        step1[17] = _mm256_add_epi16(step1[17], kOne);
+        step1[18] = _mm256_add_epi16(step1[18], kOne);
+        step1[19] = _mm256_add_epi16(step1[19], kOne);
+        step2[20] = _mm256_add_epi16(step2[20], kOne);
+        step2[21] = _mm256_add_epi16(step2[21], kOne);
+        step2[22] = _mm256_add_epi16(step2[22], kOne);
+        step2[23] = _mm256_add_epi16(step2[23], kOne);
+        step2[24] = _mm256_add_epi16(step2[24], kOne);
+        step2[25] = _mm256_add_epi16(step2[25], kOne);
+        step2[26] = _mm256_add_epi16(step2[26], kOne);
+        step2[27] = _mm256_add_epi16(step2[27], kOne);
+        step1[28] = _mm256_add_epi16(step1[28], kOne);
+        step1[29] = _mm256_add_epi16(step1[29], kOne);
+        step1[30] = _mm256_add_epi16(step1[30], kOne);
+        step1[31] = _mm256_add_epi16(step1[31], kOne);
+
+        step2[ 0] = _mm256_srai_epi16(step2[ 0], 2);
+        step2[ 1] = _mm256_srai_epi16(step2[ 1], 2);
+        step2[ 2] = _mm256_srai_epi16(step2[ 2], 2);
+        step2[ 3] = _mm256_srai_epi16(step2[ 3], 2);
+        step2[ 4] = _mm256_srai_epi16(step2[ 4], 2);
+        step2[ 5] = _mm256_srai_epi16(step2[ 5], 2);
+        step2[ 6] = _mm256_srai_epi16(step2[ 6], 2);
+        step2[ 7] = _mm256_srai_epi16(step2[ 7], 2);
+        step2[ 8] = _mm256_srai_epi16(step2[ 8], 2);
+        step2[ 9] = _mm256_srai_epi16(step2[ 9], 2);
+        step2[10] = _mm256_srai_epi16(step2[10], 2);
+        step2[11] = _mm256_srai_epi16(step2[11], 2);
+        step2[12] = _mm256_srai_epi16(step2[12], 2);
+        step2[13] = _mm256_srai_epi16(step2[13], 2);
+        step2[14] = _mm256_srai_epi16(step2[14], 2);
+        step2[15] = _mm256_srai_epi16(step2[15], 2);
+        step1[16] = _mm256_srai_epi16(step1[16], 2);
+        step1[17] = _mm256_srai_epi16(step1[17], 2);
+        step1[18] = _mm256_srai_epi16(step1[18], 2);
+        step1[19] = _mm256_srai_epi16(step1[19], 2);
+        step2[20] = _mm256_srai_epi16(step2[20], 2);
+        step2[21] = _mm256_srai_epi16(step2[21], 2);
+        step2[22] = _mm256_srai_epi16(step2[22], 2);
+        step2[23] = _mm256_srai_epi16(step2[23], 2);
+        step2[24] = _mm256_srai_epi16(step2[24], 2);
+        step2[25] = _mm256_srai_epi16(step2[25], 2);
+        step2[26] = _mm256_srai_epi16(step2[26], 2);
+        step2[27] = _mm256_srai_epi16(step2[27], 2);
+        step1[28] = _mm256_srai_epi16(step1[28], 2);
+        step1[29] = _mm256_srai_epi16(step1[29], 2);
+        step1[30] = _mm256_srai_epi16(step1[30], 2);
+        step1[31] = _mm256_srai_epi16(step1[31], 2);
+      }
+#endif
+
+#if FDCT32x32_HIGH_PRECISION
+      if (pass == 0) {
+#endif
+      // Stage 3
+      {
+        step3[0] = _mm256_add_epi16(step2[(8 - 1)], step2[0]);
+        step3[1] = _mm256_add_epi16(step2[(8 - 2)], step2[1]);
+        step3[2] = _mm256_add_epi16(step2[(8 - 3)], step2[2]);
+        step3[3] = _mm256_add_epi16(step2[(8 - 4)], step2[3]);
+        step3[4] = _mm256_sub_epi16(step2[(8 - 5)], step2[4]);
+        step3[5] = _mm256_sub_epi16(step2[(8 - 6)], step2[5]);
+        step3[6] = _mm256_sub_epi16(step2[(8 - 7)], step2[6]);
+        step3[7] = _mm256_sub_epi16(step2[(8 - 8)], step2[7]);
+      }
+      {
+        const __m256i s3_10_0 = _mm256_unpacklo_epi16(step2[13], step2[10]);
+        const __m256i s3_10_1 = _mm256_unpackhi_epi16(step2[13], step2[10]);
+        const __m256i s3_11_0 = _mm256_unpacklo_epi16(step2[12], step2[11]);
+        const __m256i s3_11_1 = _mm256_unpackhi_epi16(step2[12], step2[11]);
+        const __m256i s3_10_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_m16);
+        const __m256i s3_10_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_m16);
+        const __m256i s3_11_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_m16);
+        const __m256i s3_11_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_m16);
+        const __m256i s3_12_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_p16);
+        const __m256i s3_12_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_p16);
+        const __m256i s3_13_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_p16);
+        const __m256i s3_13_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m256i s3_10_4 = _mm256_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_10_5 = _mm256_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_11_4 = _mm256_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_11_5 = _mm256_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_12_4 = _mm256_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_12_5 = _mm256_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_13_4 = _mm256_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_13_5 = _mm256_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_10_6 = _mm256_srai_epi32(s3_10_4, DCT_CONST_BITS);
+        const __m256i s3_10_7 = _mm256_srai_epi32(s3_10_5, DCT_CONST_BITS);
+        const __m256i s3_11_6 = _mm256_srai_epi32(s3_11_4, DCT_CONST_BITS);
+        const __m256i s3_11_7 = _mm256_srai_epi32(s3_11_5, DCT_CONST_BITS);
+        const __m256i s3_12_6 = _mm256_srai_epi32(s3_12_4, DCT_CONST_BITS);
+        const __m256i s3_12_7 = _mm256_srai_epi32(s3_12_5, DCT_CONST_BITS);
+        const __m256i s3_13_6 = _mm256_srai_epi32(s3_13_4, DCT_CONST_BITS);
+        const __m256i s3_13_7 = _mm256_srai_epi32(s3_13_5, DCT_CONST_BITS);
+        // Combine
+        step3[10] = _mm256_packs_epi32(s3_10_6, s3_10_7);
+        step3[11] = _mm256_packs_epi32(s3_11_6, s3_11_7);
+        step3[12] = _mm256_packs_epi32(s3_12_6, s3_12_7);
+        step3[13] = _mm256_packs_epi32(s3_13_6, s3_13_7);
+      }
+      {
+        step3[16] = _mm256_add_epi16(step2[23], step1[16]);
+        step3[17] = _mm256_add_epi16(step2[22], step1[17]);
+        step3[18] = _mm256_add_epi16(step2[21], step1[18]);
+        step3[19] = _mm256_add_epi16(step2[20], step1[19]);
+        step3[20] = _mm256_sub_epi16(step1[19], step2[20]);
+        step3[21] = _mm256_sub_epi16(step1[18], step2[21]);
+        step3[22] = _mm256_sub_epi16(step1[17], step2[22]);
+        step3[23] = _mm256_sub_epi16(step1[16], step2[23]);
+        step3[24] = _mm256_sub_epi16(step1[31], step2[24]);
+        step3[25] = _mm256_sub_epi16(step1[30], step2[25]);
+        step3[26] = _mm256_sub_epi16(step1[29], step2[26]);
+        step3[27] = _mm256_sub_epi16(step1[28], step2[27]);
+        step3[28] = _mm256_add_epi16(step2[27], step1[28]);
+        step3[29] = _mm256_add_epi16(step2[26], step1[29]);
+        step3[30] = _mm256_add_epi16(step2[25], step1[30]);
+        step3[31] = _mm256_add_epi16(step2[24], step1[31]);
+      }
+
+      // Stage 4
+      {
+        step1[ 0] = _mm256_add_epi16(step3[ 3], step3[ 0]);
+        step1[ 1] = _mm256_add_epi16(step3[ 2], step3[ 1]);
+        step1[ 2] = _mm256_sub_epi16(step3[ 1], step3[ 2]);
+        step1[ 3] = _mm256_sub_epi16(step3[ 0], step3[ 3]);
+        step1[ 8] = _mm256_add_epi16(step3[11], step2[ 8]);
+        step1[ 9] = _mm256_add_epi16(step3[10], step2[ 9]);
+        step1[10] = _mm256_sub_epi16(step2[ 9], step3[10]);
+        step1[11] = _mm256_sub_epi16(step2[ 8], step3[11]);
+        step1[12] = _mm256_sub_epi16(step2[15], step3[12]);
+        step1[13] = _mm256_sub_epi16(step2[14], step3[13]);
+        step1[14] = _mm256_add_epi16(step3[13], step2[14]);
+        step1[15] = _mm256_add_epi16(step3[12], step2[15]);
+      }
+      {
+        const __m256i s1_05_0 = _mm256_unpacklo_epi16(step3[6], step3[5]);
+        const __m256i s1_05_1 = _mm256_unpackhi_epi16(step3[6], step3[5]);
+        const __m256i s1_05_2 = _mm256_madd_epi16(s1_05_0, k__cospi_p16_m16);
+        const __m256i s1_05_3 = _mm256_madd_epi16(s1_05_1, k__cospi_p16_m16);
+        const __m256i s1_06_2 = _mm256_madd_epi16(s1_05_0, k__cospi_p16_p16);
+        const __m256i s1_06_3 = _mm256_madd_epi16(s1_05_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m256i s1_05_4 = _mm256_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_05_5 = _mm256_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_06_4 = _mm256_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_06_5 = _mm256_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_05_6 = _mm256_srai_epi32(s1_05_4, DCT_CONST_BITS);
+        const __m256i s1_05_7 = _mm256_srai_epi32(s1_05_5, DCT_CONST_BITS);
+        const __m256i s1_06_6 = _mm256_srai_epi32(s1_06_4, DCT_CONST_BITS);
+        const __m256i s1_06_7 = _mm256_srai_epi32(s1_06_5, DCT_CONST_BITS);
+        // Combine
+        step1[5] = _mm256_packs_epi32(s1_05_6, s1_05_7);
+        step1[6] = _mm256_packs_epi32(s1_06_6, s1_06_7);
+      }
+      {
+        const __m256i s1_18_0 = _mm256_unpacklo_epi16(step3[18], step3[29]);
+        const __m256i s1_18_1 = _mm256_unpackhi_epi16(step3[18], step3[29]);
+        const __m256i s1_19_0 = _mm256_unpacklo_epi16(step3[19], step3[28]);
+        const __m256i s1_19_1 = _mm256_unpackhi_epi16(step3[19], step3[28]);
+        const __m256i s1_20_0 = _mm256_unpacklo_epi16(step3[20], step3[27]);
+        const __m256i s1_20_1 = _mm256_unpackhi_epi16(step3[20], step3[27]);
+        const __m256i s1_21_0 = _mm256_unpacklo_epi16(step3[21], step3[26]);
+        const __m256i s1_21_1 = _mm256_unpackhi_epi16(step3[21], step3[26]);
+        const __m256i s1_18_2 = _mm256_madd_epi16(s1_18_0, k__cospi_m08_p24);
+        const __m256i s1_18_3 = _mm256_madd_epi16(s1_18_1, k__cospi_m08_p24);
+        const __m256i s1_19_2 = _mm256_madd_epi16(s1_19_0, k__cospi_m08_p24);
+        const __m256i s1_19_3 = _mm256_madd_epi16(s1_19_1, k__cospi_m08_p24);
+        const __m256i s1_20_2 = _mm256_madd_epi16(s1_20_0, k__cospi_m24_m08);
+        const __m256i s1_20_3 = _mm256_madd_epi16(s1_20_1, k__cospi_m24_m08);
+        const __m256i s1_21_2 = _mm256_madd_epi16(s1_21_0, k__cospi_m24_m08);
+        const __m256i s1_21_3 = _mm256_madd_epi16(s1_21_1, k__cospi_m24_m08);
+        const __m256i s1_26_2 = _mm256_madd_epi16(s1_21_0, k__cospi_m08_p24);
+        const __m256i s1_26_3 = _mm256_madd_epi16(s1_21_1, k__cospi_m08_p24);
+        const __m256i s1_27_2 = _mm256_madd_epi16(s1_20_0, k__cospi_m08_p24);
+        const __m256i s1_27_3 = _mm256_madd_epi16(s1_20_1, k__cospi_m08_p24);
+        const __m256i s1_28_2 = _mm256_madd_epi16(s1_19_0, k__cospi_p24_p08);
+        const __m256i s1_28_3 = _mm256_madd_epi16(s1_19_1, k__cospi_p24_p08);
+        const __m256i s1_29_2 = _mm256_madd_epi16(s1_18_0, k__cospi_p24_p08);
+        const __m256i s1_29_3 = _mm256_madd_epi16(s1_18_1, k__cospi_p24_p08);
+        // dct_const_round_shift
+        const __m256i s1_18_4 = _mm256_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_18_5 = _mm256_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_19_4 = _mm256_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_19_5 = _mm256_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_20_4 = _mm256_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_20_5 = _mm256_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_21_4 = _mm256_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_21_5 = _mm256_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_26_4 = _mm256_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_26_5 = _mm256_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_27_4 = _mm256_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_27_5 = _mm256_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_28_4 = _mm256_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_28_5 = _mm256_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_29_4 = _mm256_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_29_5 = _mm256_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_18_6 = _mm256_srai_epi32(s1_18_4, DCT_CONST_BITS);
+        const __m256i s1_18_7 = _mm256_srai_epi32(s1_18_5, DCT_CONST_BITS);
+        const __m256i s1_19_6 = _mm256_srai_epi32(s1_19_4, DCT_CONST_BITS);
+        const __m256i s1_19_7 = _mm256_srai_epi32(s1_19_5, DCT_CONST_BITS);
+        const __m256i s1_20_6 = _mm256_srai_epi32(s1_20_4, DCT_CONST_BITS);
+        const __m256i s1_20_7 = _mm256_srai_epi32(s1_20_5, DCT_CONST_BITS);
+        const __m256i s1_21_6 = _mm256_srai_epi32(s1_21_4, DCT_CONST_BITS);
+        const __m256i s1_21_7 = _mm256_srai_epi32(s1_21_5, DCT_CONST_BITS);
+        const __m256i s1_26_6 = _mm256_srai_epi32(s1_26_4, DCT_CONST_BITS);
+        const __m256i s1_26_7 = _mm256_srai_epi32(s1_26_5, DCT_CONST_BITS);
+        const __m256i s1_27_6 = _mm256_srai_epi32(s1_27_4, DCT_CONST_BITS);
+        const __m256i s1_27_7 = _mm256_srai_epi32(s1_27_5, DCT_CONST_BITS);
+        const __m256i s1_28_6 = _mm256_srai_epi32(s1_28_4, DCT_CONST_BITS);
+        const __m256i s1_28_7 = _mm256_srai_epi32(s1_28_5, DCT_CONST_BITS);
+        const __m256i s1_29_6 = _mm256_srai_epi32(s1_29_4, DCT_CONST_BITS);
+        const __m256i s1_29_7 = _mm256_srai_epi32(s1_29_5, DCT_CONST_BITS);
+        // Combine
+        step1[18] = _mm256_packs_epi32(s1_18_6, s1_18_7);
+        step1[19] = _mm256_packs_epi32(s1_19_6, s1_19_7);
+        step1[20] = _mm256_packs_epi32(s1_20_6, s1_20_7);
+        step1[21] = _mm256_packs_epi32(s1_21_6, s1_21_7);
+        step1[26] = _mm256_packs_epi32(s1_26_6, s1_26_7);
+        step1[27] = _mm256_packs_epi32(s1_27_6, s1_27_7);
+        step1[28] = _mm256_packs_epi32(s1_28_6, s1_28_7);
+        step1[29] = _mm256_packs_epi32(s1_29_6, s1_29_7);
+      }
+      // Stage 5
+      {
+        step2[4] = _mm256_add_epi16(step1[5], step3[4]);
+        step2[5] = _mm256_sub_epi16(step3[4], step1[5]);
+        step2[6] = _mm256_sub_epi16(step3[7], step1[6]);
+        step2[7] = _mm256_add_epi16(step1[6], step3[7]);
+      }
+      {
+        const __m256i out_00_0 = _mm256_unpacklo_epi16(step1[0], step1[1]);
+        const __m256i out_00_1 = _mm256_unpackhi_epi16(step1[0], step1[1]);
+        const __m256i out_08_0 = _mm256_unpacklo_epi16(step1[2], step1[3]);
+        const __m256i out_08_1 = _mm256_unpackhi_epi16(step1[2], step1[3]);
+        const __m256i out_00_2 = _mm256_madd_epi16(out_00_0, k__cospi_p16_p16);
+        const __m256i out_00_3 = _mm256_madd_epi16(out_00_1, k__cospi_p16_p16);
+        const __m256i out_16_2 = _mm256_madd_epi16(out_00_0, k__cospi_p16_m16);
+        const __m256i out_16_3 = _mm256_madd_epi16(out_00_1, k__cospi_p16_m16);
+        const __m256i out_08_2 = _mm256_madd_epi16(out_08_0, k__cospi_p24_p08);
+        const __m256i out_08_3 = _mm256_madd_epi16(out_08_1, k__cospi_p24_p08);
+        const __m256i out_24_2 = _mm256_madd_epi16(out_08_0, k__cospi_m08_p24);
+        const __m256i out_24_3 = _mm256_madd_epi16(out_08_1, k__cospi_m08_p24);
+        // dct_const_round_shift
+        const __m256i out_00_4 = _mm256_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_00_5 = _mm256_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_16_4 = _mm256_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_16_5 = _mm256_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_08_4 = _mm256_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_08_5 = _mm256_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_24_4 = _mm256_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_24_5 = _mm256_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_00_6 = _mm256_srai_epi32(out_00_4, DCT_CONST_BITS);
+        const __m256i out_00_7 = _mm256_srai_epi32(out_00_5, DCT_CONST_BITS);
+        const __m256i out_16_6 = _mm256_srai_epi32(out_16_4, DCT_CONST_BITS);
+        const __m256i out_16_7 = _mm256_srai_epi32(out_16_5, DCT_CONST_BITS);
+        const __m256i out_08_6 = _mm256_srai_epi32(out_08_4, DCT_CONST_BITS);
+        const __m256i out_08_7 = _mm256_srai_epi32(out_08_5, DCT_CONST_BITS);
+        const __m256i out_24_6 = _mm256_srai_epi32(out_24_4, DCT_CONST_BITS);
+        const __m256i out_24_7 = _mm256_srai_epi32(out_24_5, DCT_CONST_BITS);
+        // Combine
+        out[ 0] = _mm256_packs_epi32(out_00_6, out_00_7);
+        out[16] = _mm256_packs_epi32(out_16_6, out_16_7);
+        out[ 8] = _mm256_packs_epi32(out_08_6, out_08_7);
+        out[24] = _mm256_packs_epi32(out_24_6, out_24_7);
+      }
+      {
+        const __m256i s2_09_0 = _mm256_unpacklo_epi16(step1[ 9], step1[14]);
+        const __m256i s2_09_1 = _mm256_unpackhi_epi16(step1[ 9], step1[14]);
+        const __m256i s2_10_0 = _mm256_unpacklo_epi16(step1[10], step1[13]);
+        const __m256i s2_10_1 = _mm256_unpackhi_epi16(step1[10], step1[13]);
+        const __m256i s2_09_2 = _mm256_madd_epi16(s2_09_0, k__cospi_m08_p24);
+        const __m256i s2_09_3 = _mm256_madd_epi16(s2_09_1, k__cospi_m08_p24);
+        const __m256i s2_10_2 = _mm256_madd_epi16(s2_10_0, k__cospi_m24_m08);
+        const __m256i s2_10_3 = _mm256_madd_epi16(s2_10_1, k__cospi_m24_m08);
+        const __m256i s2_13_2 = _mm256_madd_epi16(s2_10_0, k__cospi_m08_p24);
+        const __m256i s2_13_3 = _mm256_madd_epi16(s2_10_1, k__cospi_m08_p24);
+        const __m256i s2_14_2 = _mm256_madd_epi16(s2_09_0, k__cospi_p24_p08);
+        const __m256i s2_14_3 = _mm256_madd_epi16(s2_09_1, k__cospi_p24_p08);
+        // dct_const_round_shift
+        const __m256i s2_09_4 = _mm256_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_09_5 = _mm256_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_10_4 = _mm256_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_10_5 = _mm256_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_13_4 = _mm256_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_13_5 = _mm256_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_14_4 = _mm256_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_14_5 = _mm256_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_09_6 = _mm256_srai_epi32(s2_09_4, DCT_CONST_BITS);
+        const __m256i s2_09_7 = _mm256_srai_epi32(s2_09_5, DCT_CONST_BITS);
+        const __m256i s2_10_6 = _mm256_srai_epi32(s2_10_4, DCT_CONST_BITS);
+        const __m256i s2_10_7 = _mm256_srai_epi32(s2_10_5, DCT_CONST_BITS);
+        const __m256i s2_13_6 = _mm256_srai_epi32(s2_13_4, DCT_CONST_BITS);
+        const __m256i s2_13_7 = _mm256_srai_epi32(s2_13_5, DCT_CONST_BITS);
+        const __m256i s2_14_6 = _mm256_srai_epi32(s2_14_4, DCT_CONST_BITS);
+        const __m256i s2_14_7 = _mm256_srai_epi32(s2_14_5, DCT_CONST_BITS);
+        // Combine
+        step2[ 9] = _mm256_packs_epi32(s2_09_6, s2_09_7);
+        step2[10] = _mm256_packs_epi32(s2_10_6, s2_10_7);
+        step2[13] = _mm256_packs_epi32(s2_13_6, s2_13_7);
+        step2[14] = _mm256_packs_epi32(s2_14_6, s2_14_7);
+      }
+      {
+        step2[16] = _mm256_add_epi16(step1[19], step3[16]);
+        step2[17] = _mm256_add_epi16(step1[18], step3[17]);
+        step2[18] = _mm256_sub_epi16(step3[17], step1[18]);
+        step2[19] = _mm256_sub_epi16(step3[16], step1[19]);
+        step2[20] = _mm256_sub_epi16(step3[23], step1[20]);
+        step2[21] = _mm256_sub_epi16(step3[22], step1[21]);
+        step2[22] = _mm256_add_epi16(step1[21], step3[22]);
+        step2[23] = _mm256_add_epi16(step1[20], step3[23]);
+        step2[24] = _mm256_add_epi16(step1[27], step3[24]);
+        step2[25] = _mm256_add_epi16(step1[26], step3[25]);
+        step2[26] = _mm256_sub_epi16(step3[25], step1[26]);
+        step2[27] = _mm256_sub_epi16(step3[24], step1[27]);
+        step2[28] = _mm256_sub_epi16(step3[31], step1[28]);
+        step2[29] = _mm256_sub_epi16(step3[30], step1[29]);
+        step2[30] = _mm256_add_epi16(step1[29], step3[30]);
+        step2[31] = _mm256_add_epi16(step1[28], step3[31]);
+      }
+      // Stage 6
+      {
+        const __m256i out_04_0 = _mm256_unpacklo_epi16(step2[4], step2[7]);
+        const __m256i out_04_1 = _mm256_unpackhi_epi16(step2[4], step2[7]);
+        const __m256i out_20_0 = _mm256_unpacklo_epi16(step2[5], step2[6]);
+        const __m256i out_20_1 = _mm256_unpackhi_epi16(step2[5], step2[6]);
+        const __m256i out_12_0 = _mm256_unpacklo_epi16(step2[5], step2[6]);
+        const __m256i out_12_1 = _mm256_unpackhi_epi16(step2[5], step2[6]);
+        const __m256i out_28_0 = _mm256_unpacklo_epi16(step2[4], step2[7]);
+        const __m256i out_28_1 = _mm256_unpackhi_epi16(step2[4], step2[7]);
+        const __m256i out_04_2 = _mm256_madd_epi16(out_04_0, k__cospi_p28_p04);
+        const __m256i out_04_3 = _mm256_madd_epi16(out_04_1, k__cospi_p28_p04);
+        const __m256i out_20_2 = _mm256_madd_epi16(out_20_0, k__cospi_p12_p20);
+        const __m256i out_20_3 = _mm256_madd_epi16(out_20_1, k__cospi_p12_p20);
+        const __m256i out_12_2 = _mm256_madd_epi16(out_12_0, k__cospi_m20_p12);
+        const __m256i out_12_3 = _mm256_madd_epi16(out_12_1, k__cospi_m20_p12);
+        const __m256i out_28_2 = _mm256_madd_epi16(out_28_0, k__cospi_m04_p28);
+        const __m256i out_28_3 = _mm256_madd_epi16(out_28_1, k__cospi_m04_p28);
+        // dct_const_round_shift
+        const __m256i out_04_4 = _mm256_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_04_5 = _mm256_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_20_4 = _mm256_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_20_5 = _mm256_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_12_4 = _mm256_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_12_5 = _mm256_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_28_4 = _mm256_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_28_5 = _mm256_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_04_6 = _mm256_srai_epi32(out_04_4, DCT_CONST_BITS);
+        const __m256i out_04_7 = _mm256_srai_epi32(out_04_5, DCT_CONST_BITS);
+        const __m256i out_20_6 = _mm256_srai_epi32(out_20_4, DCT_CONST_BITS);
+        const __m256i out_20_7 = _mm256_srai_epi32(out_20_5, DCT_CONST_BITS);
+        const __m256i out_12_6 = _mm256_srai_epi32(out_12_4, DCT_CONST_BITS);
+        const __m256i out_12_7 = _mm256_srai_epi32(out_12_5, DCT_CONST_BITS);
+        const __m256i out_28_6 = _mm256_srai_epi32(out_28_4, DCT_CONST_BITS);
+        const __m256i out_28_7 = _mm256_srai_epi32(out_28_5, DCT_CONST_BITS);
+        // Combine
+        out[ 4] = _mm256_packs_epi32(out_04_6, out_04_7);
+        out[20] = _mm256_packs_epi32(out_20_6, out_20_7);
+        out[12] = _mm256_packs_epi32(out_12_6, out_12_7);
+        out[28] = _mm256_packs_epi32(out_28_6, out_28_7);
+      }
+      {
+        step3[ 8] = _mm256_add_epi16(step2[ 9], step1[ 8]);
+        step3[ 9] = _mm256_sub_epi16(step1[ 8], step2[ 9]);
+        step3[10] = _mm256_sub_epi16(step1[11], step2[10]);
+        step3[11] = _mm256_add_epi16(step2[10], step1[11]);
+        step3[12] = _mm256_add_epi16(step2[13], step1[12]);
+        step3[13] = _mm256_sub_epi16(step1[12], step2[13]);
+        step3[14] = _mm256_sub_epi16(step1[15], step2[14]);
+        step3[15] = _mm256_add_epi16(step2[14], step1[15]);
+      }
+      {
+        const __m256i s3_17_0 = _mm256_unpacklo_epi16(step2[17], step2[30]);
+        const __m256i s3_17_1 = _mm256_unpackhi_epi16(step2[17], step2[30]);
+        const __m256i s3_18_0 = _mm256_unpacklo_epi16(step2[18], step2[29]);
+        const __m256i s3_18_1 = _mm256_unpackhi_epi16(step2[18], step2[29]);
+        const __m256i s3_21_0 = _mm256_unpacklo_epi16(step2[21], step2[26]);
+        const __m256i s3_21_1 = _mm256_unpackhi_epi16(step2[21], step2[26]);
+        const __m256i s3_22_0 = _mm256_unpacklo_epi16(step2[22], step2[25]);
+        const __m256i s3_22_1 = _mm256_unpackhi_epi16(step2[22], step2[25]);
+        const __m256i s3_17_2 = _mm256_madd_epi16(s3_17_0, k__cospi_m04_p28);
+        const __m256i s3_17_3 = _mm256_madd_epi16(s3_17_1, k__cospi_m04_p28);
+        const __m256i s3_18_2 = _mm256_madd_epi16(s3_18_0, k__cospi_m28_m04);
+        const __m256i s3_18_3 = _mm256_madd_epi16(s3_18_1, k__cospi_m28_m04);
+        const __m256i s3_21_2 = _mm256_madd_epi16(s3_21_0, k__cospi_m20_p12);
+        const __m256i s3_21_3 = _mm256_madd_epi16(s3_21_1, k__cospi_m20_p12);
+        const __m256i s3_22_2 = _mm256_madd_epi16(s3_22_0, k__cospi_m12_m20);
+        const __m256i s3_22_3 = _mm256_madd_epi16(s3_22_1, k__cospi_m12_m20);
+        const __m256i s3_25_2 = _mm256_madd_epi16(s3_22_0, k__cospi_m20_p12);
+        const __m256i s3_25_3 = _mm256_madd_epi16(s3_22_1, k__cospi_m20_p12);
+        const __m256i s3_26_2 = _mm256_madd_epi16(s3_21_0, k__cospi_p12_p20);
+        const __m256i s3_26_3 = _mm256_madd_epi16(s3_21_1, k__cospi_p12_p20);
+        const __m256i s3_29_2 = _mm256_madd_epi16(s3_18_0, k__cospi_m04_p28);
+        const __m256i s3_29_3 = _mm256_madd_epi16(s3_18_1, k__cospi_m04_p28);
+        const __m256i s3_30_2 = _mm256_madd_epi16(s3_17_0, k__cospi_p28_p04);
+        const __m256i s3_30_3 = _mm256_madd_epi16(s3_17_1, k__cospi_p28_p04);
+        // dct_const_round_shift
+        const __m256i s3_17_4 = _mm256_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_17_5 = _mm256_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_18_4 = _mm256_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_18_5 = _mm256_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_21_4 = _mm256_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_21_5 = _mm256_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_22_4 = _mm256_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_22_5 = _mm256_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_17_6 = _mm256_srai_epi32(s3_17_4, DCT_CONST_BITS);
+        const __m256i s3_17_7 = _mm256_srai_epi32(s3_17_5, DCT_CONST_BITS);
+        const __m256i s3_18_6 = _mm256_srai_epi32(s3_18_4, DCT_CONST_BITS);
+        const __m256i s3_18_7 = _mm256_srai_epi32(s3_18_5, DCT_CONST_BITS);
+        const __m256i s3_21_6 = _mm256_srai_epi32(s3_21_4, DCT_CONST_BITS);
+        const __m256i s3_21_7 = _mm256_srai_epi32(s3_21_5, DCT_CONST_BITS);
+        const __m256i s3_22_6 = _mm256_srai_epi32(s3_22_4, DCT_CONST_BITS);
+        const __m256i s3_22_7 = _mm256_srai_epi32(s3_22_5, DCT_CONST_BITS);
+        const __m256i s3_25_4 = _mm256_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_25_5 = _mm256_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_26_4 = _mm256_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_26_5 = _mm256_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_29_4 = _mm256_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_29_5 = _mm256_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_30_4 = _mm256_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_30_5 = _mm256_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_25_6 = _mm256_srai_epi32(s3_25_4, DCT_CONST_BITS);
+        const __m256i s3_25_7 = _mm256_srai_epi32(s3_25_5, DCT_CONST_BITS);
+        const __m256i s3_26_6 = _mm256_srai_epi32(s3_26_4, DCT_CONST_BITS);
+        const __m256i s3_26_7 = _mm256_srai_epi32(s3_26_5, DCT_CONST_BITS);
+        const __m256i s3_29_6 = _mm256_srai_epi32(s3_29_4, DCT_CONST_BITS);
+        const __m256i s3_29_7 = _mm256_srai_epi32(s3_29_5, DCT_CONST_BITS);
+        const __m256i s3_30_6 = _mm256_srai_epi32(s3_30_4, DCT_CONST_BITS);
+        const __m256i s3_30_7 = _mm256_srai_epi32(s3_30_5, DCT_CONST_BITS);
+        // Combine
+        step3[17] = _mm256_packs_epi32(s3_17_6, s3_17_7);
+        step3[18] = _mm256_packs_epi32(s3_18_6, s3_18_7);
+        step3[21] = _mm256_packs_epi32(s3_21_6, s3_21_7);
+        step3[22] = _mm256_packs_epi32(s3_22_6, s3_22_7);
+        // Combine
+        step3[25] = _mm256_packs_epi32(s3_25_6, s3_25_7);
+        step3[26] = _mm256_packs_epi32(s3_26_6, s3_26_7);
+        step3[29] = _mm256_packs_epi32(s3_29_6, s3_29_7);
+        step3[30] = _mm256_packs_epi32(s3_30_6, s3_30_7);
+      }
+      // Stage 7
+      {
+        const __m256i out_02_0 = _mm256_unpacklo_epi16(step3[ 8], step3[15]);
+        const __m256i out_02_1 = _mm256_unpackhi_epi16(step3[ 8], step3[15]);
+        const __m256i out_18_0 = _mm256_unpacklo_epi16(step3[ 9], step3[14]);
+        const __m256i out_18_1 = _mm256_unpackhi_epi16(step3[ 9], step3[14]);
+        const __m256i out_10_0 = _mm256_unpacklo_epi16(step3[10], step3[13]);
+        const __m256i out_10_1 = _mm256_unpackhi_epi16(step3[10], step3[13]);
+        const __m256i out_26_0 = _mm256_unpacklo_epi16(step3[11], step3[12]);
+        const __m256i out_26_1 = _mm256_unpackhi_epi16(step3[11], step3[12]);
+        const __m256i out_02_2 = _mm256_madd_epi16(out_02_0, k__cospi_p30_p02);
+        const __m256i out_02_3 = _mm256_madd_epi16(out_02_1, k__cospi_p30_p02);
+        const __m256i out_18_2 = _mm256_madd_epi16(out_18_0, k__cospi_p14_p18);
+        const __m256i out_18_3 = _mm256_madd_epi16(out_18_1, k__cospi_p14_p18);
+        const __m256i out_10_2 = _mm256_madd_epi16(out_10_0, k__cospi_p22_p10);
+        const __m256i out_10_3 = _mm256_madd_epi16(out_10_1, k__cospi_p22_p10);
+        const __m256i out_26_2 = _mm256_madd_epi16(out_26_0, k__cospi_p06_p26);
+        const __m256i out_26_3 = _mm256_madd_epi16(out_26_1, k__cospi_p06_p26);
+        const __m256i out_06_2 = _mm256_madd_epi16(out_26_0, k__cospi_m26_p06);
+        const __m256i out_06_3 = _mm256_madd_epi16(out_26_1, k__cospi_m26_p06);
+        const __m256i out_22_2 = _mm256_madd_epi16(out_10_0, k__cospi_m10_p22);
+        const __m256i out_22_3 = _mm256_madd_epi16(out_10_1, k__cospi_m10_p22);
+        const __m256i out_14_2 = _mm256_madd_epi16(out_18_0, k__cospi_m18_p14);
+        const __m256i out_14_3 = _mm256_madd_epi16(out_18_1, k__cospi_m18_p14);
+        const __m256i out_30_2 = _mm256_madd_epi16(out_02_0, k__cospi_m02_p30);
+        const __m256i out_30_3 = _mm256_madd_epi16(out_02_1, k__cospi_m02_p30);
+        // dct_const_round_shift
+        const __m256i out_02_4 = _mm256_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_02_5 = _mm256_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_18_4 = _mm256_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_18_5 = _mm256_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_10_4 = _mm256_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_10_5 = _mm256_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_26_4 = _mm256_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_26_5 = _mm256_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_06_4 = _mm256_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_06_5 = _mm256_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_22_4 = _mm256_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_22_5 = _mm256_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_14_4 = _mm256_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_14_5 = _mm256_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_30_4 = _mm256_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_30_5 = _mm256_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_02_6 = _mm256_srai_epi32(out_02_4, DCT_CONST_BITS);
+        const __m256i out_02_7 = _mm256_srai_epi32(out_02_5, DCT_CONST_BITS);
+        const __m256i out_18_6 = _mm256_srai_epi32(out_18_4, DCT_CONST_BITS);
+        const __m256i out_18_7 = _mm256_srai_epi32(out_18_5, DCT_CONST_BITS);
+        const __m256i out_10_6 = _mm256_srai_epi32(out_10_4, DCT_CONST_BITS);
+        const __m256i out_10_7 = _mm256_srai_epi32(out_10_5, DCT_CONST_BITS);
+        const __m256i out_26_6 = _mm256_srai_epi32(out_26_4, DCT_CONST_BITS);
+        const __m256i out_26_7 = _mm256_srai_epi32(out_26_5, DCT_CONST_BITS);
+        const __m256i out_06_6 = _mm256_srai_epi32(out_06_4, DCT_CONST_BITS);
+        const __m256i out_06_7 = _mm256_srai_epi32(out_06_5, DCT_CONST_BITS);
+        const __m256i out_22_6 = _mm256_srai_epi32(out_22_4, DCT_CONST_BITS);
+        const __m256i out_22_7 = _mm256_srai_epi32(out_22_5, DCT_CONST_BITS);
+        const __m256i out_14_6 = _mm256_srai_epi32(out_14_4, DCT_CONST_BITS);
+        const __m256i out_14_7 = _mm256_srai_epi32(out_14_5, DCT_CONST_BITS);
+        const __m256i out_30_6 = _mm256_srai_epi32(out_30_4, DCT_CONST_BITS);
+        const __m256i out_30_7 = _mm256_srai_epi32(out_30_5, DCT_CONST_BITS);
+        // Combine
+        out[ 2] = _mm256_packs_epi32(out_02_6, out_02_7);
+        out[18] = _mm256_packs_epi32(out_18_6, out_18_7);
+        out[10] = _mm256_packs_epi32(out_10_6, out_10_7);
+        out[26] = _mm256_packs_epi32(out_26_6, out_26_7);
+        out[ 6] = _mm256_packs_epi32(out_06_6, out_06_7);
+        out[22] = _mm256_packs_epi32(out_22_6, out_22_7);
+        out[14] = _mm256_packs_epi32(out_14_6, out_14_7);
+        out[30] = _mm256_packs_epi32(out_30_6, out_30_7);
+      }
+      {
+        step1[16] = _mm256_add_epi16(step3[17], step2[16]);
+        step1[17] = _mm256_sub_epi16(step2[16], step3[17]);
+        step1[18] = _mm256_sub_epi16(step2[19], step3[18]);
+        step1[19] = _mm256_add_epi16(step3[18], step2[19]);
+        step1[20] = _mm256_add_epi16(step3[21], step2[20]);
+        step1[21] = _mm256_sub_epi16(step2[20], step3[21]);
+        step1[22] = _mm256_sub_epi16(step2[23], step3[22]);
+        step1[23] = _mm256_add_epi16(step3[22], step2[23]);
+        step1[24] = _mm256_add_epi16(step3[25], step2[24]);
+        step1[25] = _mm256_sub_epi16(step2[24], step3[25]);
+        step1[26] = _mm256_sub_epi16(step2[27], step3[26]);
+        step1[27] = _mm256_add_epi16(step3[26], step2[27]);
+        step1[28] = _mm256_add_epi16(step3[29], step2[28]);
+        step1[29] = _mm256_sub_epi16(step2[28], step3[29]);
+        step1[30] = _mm256_sub_epi16(step2[31], step3[30]);
+        step1[31] = _mm256_add_epi16(step3[30], step2[31]);
+      }
+      // Final stage --- outputs indices are bit-reversed.
+      {
+        const __m256i out_01_0 = _mm256_unpacklo_epi16(step1[16], step1[31]);
+        const __m256i out_01_1 = _mm256_unpackhi_epi16(step1[16], step1[31]);
+        const __m256i out_17_0 = _mm256_unpacklo_epi16(step1[17], step1[30]);
+        const __m256i out_17_1 = _mm256_unpackhi_epi16(step1[17], step1[30]);
+        const __m256i out_09_0 = _mm256_unpacklo_epi16(step1[18], step1[29]);
+        const __m256i out_09_1 = _mm256_unpackhi_epi16(step1[18], step1[29]);
+        const __m256i out_25_0 = _mm256_unpacklo_epi16(step1[19], step1[28]);
+        const __m256i out_25_1 = _mm256_unpackhi_epi16(step1[19], step1[28]);
+        const __m256i out_01_2 = _mm256_madd_epi16(out_01_0, k__cospi_p31_p01);
+        const __m256i out_01_3 = _mm256_madd_epi16(out_01_1, k__cospi_p31_p01);
+        const __m256i out_17_2 = _mm256_madd_epi16(out_17_0, k__cospi_p15_p17);
+        const __m256i out_17_3 = _mm256_madd_epi16(out_17_1, k__cospi_p15_p17);
+        const __m256i out_09_2 = _mm256_madd_epi16(out_09_0, k__cospi_p23_p09);
+        const __m256i out_09_3 = _mm256_madd_epi16(out_09_1, k__cospi_p23_p09);
+        const __m256i out_25_2 = _mm256_madd_epi16(out_25_0, k__cospi_p07_p25);
+        const __m256i out_25_3 = _mm256_madd_epi16(out_25_1, k__cospi_p07_p25);
+        const __m256i out_07_2 = _mm256_madd_epi16(out_25_0, k__cospi_m25_p07);
+        const __m256i out_07_3 = _mm256_madd_epi16(out_25_1, k__cospi_m25_p07);
+        const __m256i out_23_2 = _mm256_madd_epi16(out_09_0, k__cospi_m09_p23);
+        const __m256i out_23_3 = _mm256_madd_epi16(out_09_1, k__cospi_m09_p23);
+        const __m256i out_15_2 = _mm256_madd_epi16(out_17_0, k__cospi_m17_p15);
+        const __m256i out_15_3 = _mm256_madd_epi16(out_17_1, k__cospi_m17_p15);
+        const __m256i out_31_2 = _mm256_madd_epi16(out_01_0, k__cospi_m01_p31);
+        const __m256i out_31_3 = _mm256_madd_epi16(out_01_1, k__cospi_m01_p31);
+        // dct_const_round_shift
+        const __m256i out_01_4 = _mm256_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_01_5 = _mm256_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_17_4 = _mm256_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_17_5 = _mm256_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_09_4 = _mm256_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_09_5 = _mm256_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_25_4 = _mm256_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_25_5 = _mm256_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_07_4 = _mm256_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_07_5 = _mm256_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_23_4 = _mm256_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_23_5 = _mm256_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_15_4 = _mm256_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_15_5 = _mm256_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_31_4 = _mm256_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_31_5 = _mm256_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_01_6 = _mm256_srai_epi32(out_01_4, DCT_CONST_BITS);
+        const __m256i out_01_7 = _mm256_srai_epi32(out_01_5, DCT_CONST_BITS);
+        const __m256i out_17_6 = _mm256_srai_epi32(out_17_4, DCT_CONST_BITS);
+        const __m256i out_17_7 = _mm256_srai_epi32(out_17_5, DCT_CONST_BITS);
+        const __m256i out_09_6 = _mm256_srai_epi32(out_09_4, DCT_CONST_BITS);
+        const __m256i out_09_7 = _mm256_srai_epi32(out_09_5, DCT_CONST_BITS);
+        const __m256i out_25_6 = _mm256_srai_epi32(out_25_4, DCT_CONST_BITS);
+        const __m256i out_25_7 = _mm256_srai_epi32(out_25_5, DCT_CONST_BITS);
+        const __m256i out_07_6 = _mm256_srai_epi32(out_07_4, DCT_CONST_BITS);
+        const __m256i out_07_7 = _mm256_srai_epi32(out_07_5, DCT_CONST_BITS);
+        const __m256i out_23_6 = _mm256_srai_epi32(out_23_4, DCT_CONST_BITS);
+        const __m256i out_23_7 = _mm256_srai_epi32(out_23_5, DCT_CONST_BITS);
+        const __m256i out_15_6 = _mm256_srai_epi32(out_15_4, DCT_CONST_BITS);
+        const __m256i out_15_7 = _mm256_srai_epi32(out_15_5, DCT_CONST_BITS);
+        const __m256i out_31_6 = _mm256_srai_epi32(out_31_4, DCT_CONST_BITS);
+        const __m256i out_31_7 = _mm256_srai_epi32(out_31_5, DCT_CONST_BITS);
+        // Combine
+        out[ 1] = _mm256_packs_epi32(out_01_6, out_01_7);
+        out[17] = _mm256_packs_epi32(out_17_6, out_17_7);
+        out[ 9] = _mm256_packs_epi32(out_09_6, out_09_7);
+        out[25] = _mm256_packs_epi32(out_25_6, out_25_7);
+        out[ 7] = _mm256_packs_epi32(out_07_6, out_07_7);
+        out[23] = _mm256_packs_epi32(out_23_6, out_23_7);
+        out[15] = _mm256_packs_epi32(out_15_6, out_15_7);
+        out[31] = _mm256_packs_epi32(out_31_6, out_31_7);
+      }
+      {
+        const __m256i out_05_0 = _mm256_unpacklo_epi16(step1[20], step1[27]);
+        const __m256i out_05_1 = _mm256_unpackhi_epi16(step1[20], step1[27]);
+        const __m256i out_21_0 = _mm256_unpacklo_epi16(step1[21], step1[26]);
+        const __m256i out_21_1 = _mm256_unpackhi_epi16(step1[21], step1[26]);
+        const __m256i out_13_0 = _mm256_unpacklo_epi16(step1[22], step1[25]);
+        const __m256i out_13_1 = _mm256_unpackhi_epi16(step1[22], step1[25]);
+        const __m256i out_29_0 = _mm256_unpacklo_epi16(step1[23], step1[24]);
+        const __m256i out_29_1 = _mm256_unpackhi_epi16(step1[23], step1[24]);
+        const __m256i out_05_2 = _mm256_madd_epi16(out_05_0, k__cospi_p27_p05);
+        const __m256i out_05_3 = _mm256_madd_epi16(out_05_1, k__cospi_p27_p05);
+        const __m256i out_21_2 = _mm256_madd_epi16(out_21_0, k__cospi_p11_p21);
+        const __m256i out_21_3 = _mm256_madd_epi16(out_21_1, k__cospi_p11_p21);
+        const __m256i out_13_2 = _mm256_madd_epi16(out_13_0, k__cospi_p19_p13);
+        const __m256i out_13_3 = _mm256_madd_epi16(out_13_1, k__cospi_p19_p13);
+        const __m256i out_29_2 = _mm256_madd_epi16(out_29_0, k__cospi_p03_p29);
+        const __m256i out_29_3 = _mm256_madd_epi16(out_29_1, k__cospi_p03_p29);
+        const __m256i out_03_2 = _mm256_madd_epi16(out_29_0, k__cospi_m29_p03);
+        const __m256i out_03_3 = _mm256_madd_epi16(out_29_1, k__cospi_m29_p03);
+        const __m256i out_19_2 = _mm256_madd_epi16(out_13_0, k__cospi_m13_p19);
+        const __m256i out_19_3 = _mm256_madd_epi16(out_13_1, k__cospi_m13_p19);
+        const __m256i out_11_2 = _mm256_madd_epi16(out_21_0, k__cospi_m21_p11);
+        const __m256i out_11_3 = _mm256_madd_epi16(out_21_1, k__cospi_m21_p11);
+        const __m256i out_27_2 = _mm256_madd_epi16(out_05_0, k__cospi_m05_p27);
+        const __m256i out_27_3 = _mm256_madd_epi16(out_05_1, k__cospi_m05_p27);
+        // dct_const_round_shift
+        const __m256i out_05_4 = _mm256_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_05_5 = _mm256_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_21_4 = _mm256_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_21_5 = _mm256_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_13_4 = _mm256_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_13_5 = _mm256_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_29_4 = _mm256_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_29_5 = _mm256_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_03_4 = _mm256_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_03_5 = _mm256_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_19_4 = _mm256_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_19_5 = _mm256_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_11_4 = _mm256_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_11_5 = _mm256_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_27_4 = _mm256_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_27_5 = _mm256_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_05_6 = _mm256_srai_epi32(out_05_4, DCT_CONST_BITS);
+        const __m256i out_05_7 = _mm256_srai_epi32(out_05_5, DCT_CONST_BITS);
+        const __m256i out_21_6 = _mm256_srai_epi32(out_21_4, DCT_CONST_BITS);
+        const __m256i out_21_7 = _mm256_srai_epi32(out_21_5, DCT_CONST_BITS);
+        const __m256i out_13_6 = _mm256_srai_epi32(out_13_4, DCT_CONST_BITS);
+        const __m256i out_13_7 = _mm256_srai_epi32(out_13_5, DCT_CONST_BITS);
+        const __m256i out_29_6 = _mm256_srai_epi32(out_29_4, DCT_CONST_BITS);
+        const __m256i out_29_7 = _mm256_srai_epi32(out_29_5, DCT_CONST_BITS);
+        const __m256i out_03_6 = _mm256_srai_epi32(out_03_4, DCT_CONST_BITS);
+        const __m256i out_03_7 = _mm256_srai_epi32(out_03_5, DCT_CONST_BITS);
+        const __m256i out_19_6 = _mm256_srai_epi32(out_19_4, DCT_CONST_BITS);
+        const __m256i out_19_7 = _mm256_srai_epi32(out_19_5, DCT_CONST_BITS);
+        const __m256i out_11_6 = _mm256_srai_epi32(out_11_4, DCT_CONST_BITS);
+        const __m256i out_11_7 = _mm256_srai_epi32(out_11_5, DCT_CONST_BITS);
+        const __m256i out_27_6 = _mm256_srai_epi32(out_27_4, DCT_CONST_BITS);
+        const __m256i out_27_7 = _mm256_srai_epi32(out_27_5, DCT_CONST_BITS);
+        // Combine
+        out[ 5] = _mm256_packs_epi32(out_05_6, out_05_7);
+        out[21] = _mm256_packs_epi32(out_21_6, out_21_7);
+        out[13] = _mm256_packs_epi32(out_13_6, out_13_7);
+        out[29] = _mm256_packs_epi32(out_29_6, out_29_7);
+        out[ 3] = _mm256_packs_epi32(out_03_6, out_03_7);
+        out[19] = _mm256_packs_epi32(out_19_6, out_19_7);
+        out[11] = _mm256_packs_epi32(out_11_6, out_11_7);
+        out[27] = _mm256_packs_epi32(out_27_6, out_27_7);
+      }
+#if FDCT32x32_HIGH_PRECISION
+      } else {
+        __m256i lstep1[64], lstep2[64], lstep3[64];
+        __m256i u[32], v[32], sign[16];
+        const __m256i K32One = _mm256_set_epi32(1, 1, 1, 1, 1, 1, 1, 1);
+        // start using 32-bit operations
+        // stage 3
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[ 0] = _mm256_unpacklo_epi16(step2[ 0], kZero);
+          lstep2[ 1] = _mm256_unpackhi_epi16(step2[ 0], kZero);
+          lstep2[ 2] = _mm256_unpacklo_epi16(step2[ 1], kZero);
+          lstep2[ 3] = _mm256_unpackhi_epi16(step2[ 1], kZero);
+          lstep2[ 4] = _mm256_unpacklo_epi16(step2[ 2], kZero);
+          lstep2[ 5] = _mm256_unpackhi_epi16(step2[ 2], kZero);
+          lstep2[ 6] = _mm256_unpacklo_epi16(step2[ 3], kZero);
+          lstep2[ 7] = _mm256_unpackhi_epi16(step2[ 3], kZero);
+          lstep2[ 8] = _mm256_unpacklo_epi16(step2[ 4], kZero);
+          lstep2[ 9] = _mm256_unpackhi_epi16(step2[ 4], kZero);
+          lstep2[10] = _mm256_unpacklo_epi16(step2[ 5], kZero);
+          lstep2[11] = _mm256_unpackhi_epi16(step2[ 5], kZero);
+          lstep2[12] = _mm256_unpacklo_epi16(step2[ 6], kZero);
+          lstep2[13] = _mm256_unpackhi_epi16(step2[ 6], kZero);
+          lstep2[14] = _mm256_unpacklo_epi16(step2[ 7], kZero);
+          lstep2[15] = _mm256_unpackhi_epi16(step2[ 7], kZero);
+          lstep2[ 0] = _mm256_madd_epi16(lstep2[ 0], kOne);
+          lstep2[ 1] = _mm256_madd_epi16(lstep2[ 1], kOne);
+          lstep2[ 2] = _mm256_madd_epi16(lstep2[ 2], kOne);
+          lstep2[ 3] = _mm256_madd_epi16(lstep2[ 3], kOne);
+          lstep2[ 4] = _mm256_madd_epi16(lstep2[ 4], kOne);
+          lstep2[ 5] = _mm256_madd_epi16(lstep2[ 5], kOne);
+          lstep2[ 6] = _mm256_madd_epi16(lstep2[ 6], kOne);
+          lstep2[ 7] = _mm256_madd_epi16(lstep2[ 7], kOne);
+          lstep2[ 8] = _mm256_madd_epi16(lstep2[ 8], kOne);
+          lstep2[ 9] = _mm256_madd_epi16(lstep2[ 9], kOne);
+          lstep2[10] = _mm256_madd_epi16(lstep2[10], kOne);
+          lstep2[11] = _mm256_madd_epi16(lstep2[11], kOne);
+          lstep2[12] = _mm256_madd_epi16(lstep2[12], kOne);
+          lstep2[13] = _mm256_madd_epi16(lstep2[13], kOne);
+          lstep2[14] = _mm256_madd_epi16(lstep2[14], kOne);
+          lstep2[15] = _mm256_madd_epi16(lstep2[15], kOne);
+
+          lstep3[ 0] = _mm256_add_epi32(lstep2[14], lstep2[ 0]);
+          lstep3[ 1] = _mm256_add_epi32(lstep2[15], lstep2[ 1]);
+          lstep3[ 2] = _mm256_add_epi32(lstep2[12], lstep2[ 2]);
+          lstep3[ 3] = _mm256_add_epi32(lstep2[13], lstep2[ 3]);
+          lstep3[ 4] = _mm256_add_epi32(lstep2[10], lstep2[ 4]);
+          lstep3[ 5] = _mm256_add_epi32(lstep2[11], lstep2[ 5]);
+          lstep3[ 6] = _mm256_add_epi32(lstep2[ 8], lstep2[ 6]);
+          lstep3[ 7] = _mm256_add_epi32(lstep2[ 9], lstep2[ 7]);
+          lstep3[ 8] = _mm256_sub_epi32(lstep2[ 6], lstep2[ 8]);
+          lstep3[ 9] = _mm256_sub_epi32(lstep2[ 7], lstep2[ 9]);
+          lstep3[10] = _mm256_sub_epi32(lstep2[ 4], lstep2[10]);
+          lstep3[11] = _mm256_sub_epi32(lstep2[ 5], lstep2[11]);
+          lstep3[12] = _mm256_sub_epi32(lstep2[ 2], lstep2[12]);
+          lstep3[13] = _mm256_sub_epi32(lstep2[ 3], lstep2[13]);
+          lstep3[14] = _mm256_sub_epi32(lstep2[ 0], lstep2[14]);
+          lstep3[15] = _mm256_sub_epi32(lstep2[ 1], lstep2[15]);
+        }
+        {
+          const __m256i s3_10_0 = _mm256_unpacklo_epi16(step2[13], step2[10]);
+          const __m256i s3_10_1 = _mm256_unpackhi_epi16(step2[13], step2[10]);
+          const __m256i s3_11_0 = _mm256_unpacklo_epi16(step2[12], step2[11]);
+          const __m256i s3_11_1 = _mm256_unpackhi_epi16(step2[12], step2[11]);
+          const __m256i s3_10_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_m16);
+          const __m256i s3_10_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_m16);
+          const __m256i s3_11_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_m16);
+          const __m256i s3_11_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_m16);
+          const __m256i s3_12_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_p16);
+          const __m256i s3_12_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_p16);
+          const __m256i s3_13_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_p16);
+          const __m256i s3_13_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_p16);
+          // dct_const_round_shift
+          const __m256i s3_10_4 = _mm256_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_10_5 = _mm256_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_11_4 = _mm256_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_11_5 = _mm256_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_12_4 = _mm256_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_12_5 = _mm256_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_13_4 = _mm256_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_13_5 = _mm256_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+          lstep3[20] = _mm256_srai_epi32(s3_10_4, DCT_CONST_BITS);
+          lstep3[21] = _mm256_srai_epi32(s3_10_5, DCT_CONST_BITS);
+          lstep3[22] = _mm256_srai_epi32(s3_11_4, DCT_CONST_BITS);
+          lstep3[23] = _mm256_srai_epi32(s3_11_5, DCT_CONST_BITS);
+          lstep3[24] = _mm256_srai_epi32(s3_12_4, DCT_CONST_BITS);
+          lstep3[25] = _mm256_srai_epi32(s3_12_5, DCT_CONST_BITS);
+          lstep3[26] = _mm256_srai_epi32(s3_13_4, DCT_CONST_BITS);
+          lstep3[27] = _mm256_srai_epi32(s3_13_5, DCT_CONST_BITS);
+        }
+        {
+          lstep2[40] = _mm256_unpacklo_epi16(step2[20], kZero);
+          lstep2[41] = _mm256_unpackhi_epi16(step2[20], kZero);
+          lstep2[42] = _mm256_unpacklo_epi16(step2[21], kZero);
+          lstep2[43] = _mm256_unpackhi_epi16(step2[21], kZero);
+          lstep2[44] = _mm256_unpacklo_epi16(step2[22], kZero);
+          lstep2[45] = _mm256_unpackhi_epi16(step2[22], kZero);
+          lstep2[46] = _mm256_unpacklo_epi16(step2[23], kZero);
+          lstep2[47] = _mm256_unpackhi_epi16(step2[23], kZero);
+          lstep2[48] = _mm256_unpacklo_epi16(step2[24], kZero);
+          lstep2[49] = _mm256_unpackhi_epi16(step2[24], kZero);
+          lstep2[50] = _mm256_unpacklo_epi16(step2[25], kZero);
+          lstep2[51] = _mm256_unpackhi_epi16(step2[25], kZero);
+          lstep2[52] = _mm256_unpacklo_epi16(step2[26], kZero);
+          lstep2[53] = _mm256_unpackhi_epi16(step2[26], kZero);
+          lstep2[54] = _mm256_unpacklo_epi16(step2[27], kZero);
+          lstep2[55] = _mm256_unpackhi_epi16(step2[27], kZero);
+          lstep2[40] = _mm256_madd_epi16(lstep2[40], kOne);
+          lstep2[41] = _mm256_madd_epi16(lstep2[41], kOne);
+          lstep2[42] = _mm256_madd_epi16(lstep2[42], kOne);
+          lstep2[43] = _mm256_madd_epi16(lstep2[43], kOne);
+          lstep2[44] = _mm256_madd_epi16(lstep2[44], kOne);
+          lstep2[45] = _mm256_madd_epi16(lstep2[45], kOne);
+          lstep2[46] = _mm256_madd_epi16(lstep2[46], kOne);
+          lstep2[47] = _mm256_madd_epi16(lstep2[47], kOne);
+          lstep2[48] = _mm256_madd_epi16(lstep2[48], kOne);
+          lstep2[49] = _mm256_madd_epi16(lstep2[49], kOne);
+          lstep2[50] = _mm256_madd_epi16(lstep2[50], kOne);
+          lstep2[51] = _mm256_madd_epi16(lstep2[51], kOne);
+          lstep2[52] = _mm256_madd_epi16(lstep2[52], kOne);
+          lstep2[53] = _mm256_madd_epi16(lstep2[53], kOne);
+          lstep2[54] = _mm256_madd_epi16(lstep2[54], kOne);
+          lstep2[55] = _mm256_madd_epi16(lstep2[55], kOne);
+
+          lstep1[32] = _mm256_unpacklo_epi16(step1[16], kZero);
+          lstep1[33] = _mm256_unpackhi_epi16(step1[16], kZero);
+          lstep1[34] = _mm256_unpacklo_epi16(step1[17], kZero);
+          lstep1[35] = _mm256_unpackhi_epi16(step1[17], kZero);
+          lstep1[36] = _mm256_unpacklo_epi16(step1[18], kZero);
+          lstep1[37] = _mm256_unpackhi_epi16(step1[18], kZero);
+          lstep1[38] = _mm256_unpacklo_epi16(step1[19], kZero);
+          lstep1[39] = _mm256_unpackhi_epi16(step1[19], kZero);
+          lstep1[56] = _mm256_unpacklo_epi16(step1[28], kZero);
+          lstep1[57] = _mm256_unpackhi_epi16(step1[28], kZero);
+          lstep1[58] = _mm256_unpacklo_epi16(step1[29], kZero);
+          lstep1[59] = _mm256_unpackhi_epi16(step1[29], kZero);
+          lstep1[60] = _mm256_unpacklo_epi16(step1[30], kZero);
+          lstep1[61] = _mm256_unpackhi_epi16(step1[30], kZero);
+          lstep1[62] = _mm256_unpacklo_epi16(step1[31], kZero);
+          lstep1[63] = _mm256_unpackhi_epi16(step1[31], kZero);
+          lstep1[32] = _mm256_madd_epi16(lstep1[32], kOne);
+          lstep1[33] = _mm256_madd_epi16(lstep1[33], kOne);
+          lstep1[34] = _mm256_madd_epi16(lstep1[34], kOne);
+          lstep1[35] = _mm256_madd_epi16(lstep1[35], kOne);
+          lstep1[36] = _mm256_madd_epi16(lstep1[36], kOne);
+          lstep1[37] = _mm256_madd_epi16(lstep1[37], kOne);
+          lstep1[38] = _mm256_madd_epi16(lstep1[38], kOne);
+          lstep1[39] = _mm256_madd_epi16(lstep1[39], kOne);
+          lstep1[56] = _mm256_madd_epi16(lstep1[56], kOne);
+          lstep1[57] = _mm256_madd_epi16(lstep1[57], kOne);
+          lstep1[58] = _mm256_madd_epi16(lstep1[58], kOne);
+          lstep1[59] = _mm256_madd_epi16(lstep1[59], kOne);
+          lstep1[60] = _mm256_madd_epi16(lstep1[60], kOne);
+          lstep1[61] = _mm256_madd_epi16(lstep1[61], kOne);
+          lstep1[62] = _mm256_madd_epi16(lstep1[62], kOne);
+          lstep1[63] = _mm256_madd_epi16(lstep1[63], kOne);
+
+          lstep3[32] = _mm256_add_epi32(lstep2[46], lstep1[32]);
+          lstep3[33] = _mm256_add_epi32(lstep2[47], lstep1[33]);
+
+          lstep3[34] = _mm256_add_epi32(lstep2[44], lstep1[34]);
+          lstep3[35] = _mm256_add_epi32(lstep2[45], lstep1[35]);
+          lstep3[36] = _mm256_add_epi32(lstep2[42], lstep1[36]);
+          lstep3[37] = _mm256_add_epi32(lstep2[43], lstep1[37]);
+          lstep3[38] = _mm256_add_epi32(lstep2[40], lstep1[38]);
+          lstep3[39] = _mm256_add_epi32(lstep2[41], lstep1[39]);
+          lstep3[40] = _mm256_sub_epi32(lstep1[38], lstep2[40]);
+          lstep3[41] = _mm256_sub_epi32(lstep1[39], lstep2[41]);
+          lstep3[42] = _mm256_sub_epi32(lstep1[36], lstep2[42]);
+          lstep3[43] = _mm256_sub_epi32(lstep1[37], lstep2[43]);
+          lstep3[44] = _mm256_sub_epi32(lstep1[34], lstep2[44]);
+          lstep3[45] = _mm256_sub_epi32(lstep1[35], lstep2[45]);
+          lstep3[46] = _mm256_sub_epi32(lstep1[32], lstep2[46]);
+          lstep3[47] = _mm256_sub_epi32(lstep1[33], lstep2[47]);
+          lstep3[48] = _mm256_sub_epi32(lstep1[62], lstep2[48]);
+          lstep3[49] = _mm256_sub_epi32(lstep1[63], lstep2[49]);
+          lstep3[50] = _mm256_sub_epi32(lstep1[60], lstep2[50]);
+          lstep3[51] = _mm256_sub_epi32(lstep1[61], lstep2[51]);
+          lstep3[52] = _mm256_sub_epi32(lstep1[58], lstep2[52]);
+          lstep3[53] = _mm256_sub_epi32(lstep1[59], lstep2[53]);
+          lstep3[54] = _mm256_sub_epi32(lstep1[56], lstep2[54]);
+          lstep3[55] = _mm256_sub_epi32(lstep1[57], lstep2[55]);
+          lstep3[56] = _mm256_add_epi32(lstep2[54], lstep1[56]);
+          lstep3[57] = _mm256_add_epi32(lstep2[55], lstep1[57]);
+          lstep3[58] = _mm256_add_epi32(lstep2[52], lstep1[58]);
+          lstep3[59] = _mm256_add_epi32(lstep2[53], lstep1[59]);
+          lstep3[60] = _mm256_add_epi32(lstep2[50], lstep1[60]);
+          lstep3[61] = _mm256_add_epi32(lstep2[51], lstep1[61]);
+          lstep3[62] = _mm256_add_epi32(lstep2[48], lstep1[62]);
+          lstep3[63] = _mm256_add_epi32(lstep2[49], lstep1[63]);
+        }
+
+        // stage 4
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[16] = _mm256_unpacklo_epi16(step2[ 8], kZero);
+          lstep2[17] = _mm256_unpackhi_epi16(step2[ 8], kZero);
+          lstep2[18] = _mm256_unpacklo_epi16(step2[ 9], kZero);
+          lstep2[19] = _mm256_unpackhi_epi16(step2[ 9], kZero);
+          lstep2[28] = _mm256_unpacklo_epi16(step2[14], kZero);
+          lstep2[29] = _mm256_unpackhi_epi16(step2[14], kZero);
+          lstep2[30] = _mm256_unpacklo_epi16(step2[15], kZero);
+          lstep2[31] = _mm256_unpackhi_epi16(step2[15], kZero);
+          lstep2[16] = _mm256_madd_epi16(lstep2[16], kOne);
+          lstep2[17] = _mm256_madd_epi16(lstep2[17], kOne);
+          lstep2[18] = _mm256_madd_epi16(lstep2[18], kOne);
+          lstep2[19] = _mm256_madd_epi16(lstep2[19], kOne);
+          lstep2[28] = _mm256_madd_epi16(lstep2[28], kOne);
+          lstep2[29] = _mm256_madd_epi16(lstep2[29], kOne);
+          lstep2[30] = _mm256_madd_epi16(lstep2[30], kOne);
+          lstep2[31] = _mm256_madd_epi16(lstep2[31], kOne);
+
+          lstep1[ 0] = _mm256_add_epi32(lstep3[ 6], lstep3[ 0]);
+          lstep1[ 1] = _mm256_add_epi32(lstep3[ 7], lstep3[ 1]);
+          lstep1[ 2] = _mm256_add_epi32(lstep3[ 4], lstep3[ 2]);
+          lstep1[ 3] = _mm256_add_epi32(lstep3[ 5], lstep3[ 3]);
+          lstep1[ 4] = _mm256_sub_epi32(lstep3[ 2], lstep3[ 4]);
+          lstep1[ 5] = _mm256_sub_epi32(lstep3[ 3], lstep3[ 5]);
+          lstep1[ 6] = _mm256_sub_epi32(lstep3[ 0], lstep3[ 6]);
+          lstep1[ 7] = _mm256_sub_epi32(lstep3[ 1], lstep3[ 7]);
+          lstep1[16] = _mm256_add_epi32(lstep3[22], lstep2[16]);
+          lstep1[17] = _mm256_add_epi32(lstep3[23], lstep2[17]);
+          lstep1[18] = _mm256_add_epi32(lstep3[20], lstep2[18]);
+          lstep1[19] = _mm256_add_epi32(lstep3[21], lstep2[19]);
+          lstep1[20] = _mm256_sub_epi32(lstep2[18], lstep3[20]);
+          lstep1[21] = _mm256_sub_epi32(lstep2[19], lstep3[21]);
+          lstep1[22] = _mm256_sub_epi32(lstep2[16], lstep3[22]);
+          lstep1[23] = _mm256_sub_epi32(lstep2[17], lstep3[23]);
+          lstep1[24] = _mm256_sub_epi32(lstep2[30], lstep3[24]);
+          lstep1[25] = _mm256_sub_epi32(lstep2[31], lstep3[25]);
+          lstep1[26] = _mm256_sub_epi32(lstep2[28], lstep3[26]);
+          lstep1[27] = _mm256_sub_epi32(lstep2[29], lstep3[27]);
+          lstep1[28] = _mm256_add_epi32(lstep3[26], lstep2[28]);
+          lstep1[29] = _mm256_add_epi32(lstep3[27], lstep2[29]);
+          lstep1[30] = _mm256_add_epi32(lstep3[24], lstep2[30]);
+          lstep1[31] = _mm256_add_epi32(lstep3[25], lstep2[31]);
+        }
+        {
+        // to be continued...
+        //
+        const __m256i k32_p16_p16 = pair256_set_epi32(cospi_16_64, cospi_16_64);
+        const __m256i k32_p16_m16 = pair256_set_epi32(cospi_16_64, -cospi_16_64);
+
+        u[0] = _mm256_unpacklo_epi32(lstep3[12], lstep3[10]);
+        u[1] = _mm256_unpackhi_epi32(lstep3[12], lstep3[10]);
+        u[2] = _mm256_unpacklo_epi32(lstep3[13], lstep3[11]);
+        u[3] = _mm256_unpackhi_epi32(lstep3[13], lstep3[11]);
+
+        // TODO(jingning): manually inline k_madd_epi32_avx2_ to further hide
+        // instruction latency.
+        v[ 0] = k_madd_epi32_avx2(u[0], k32_p16_m16);
+        v[ 1] = k_madd_epi32_avx2(u[1], k32_p16_m16);
+        v[ 2] = k_madd_epi32_avx2(u[2], k32_p16_m16);
+        v[ 3] = k_madd_epi32_avx2(u[3], k32_p16_m16);
+        v[ 4] = k_madd_epi32_avx2(u[0], k32_p16_p16);
+        v[ 5] = k_madd_epi32_avx2(u[1], k32_p16_p16);
+        v[ 6] = k_madd_epi32_avx2(u[2], k32_p16_p16);
+        v[ 7] = k_madd_epi32_avx2(u[3], k32_p16_p16);
+
+        u[0] = k_packs_epi64_avx2(v[0], v[1]);
+        u[1] = k_packs_epi64_avx2(v[2], v[3]);
+        u[2] = k_packs_epi64_avx2(v[4], v[5]);
+        u[3] = k_packs_epi64_avx2(v[6], v[7]);
+
+        v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+        v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+        v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+        v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+
+        lstep1[10] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+        lstep1[11] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+        lstep1[12] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+        lstep1[13] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+        }
+        {
+          const __m256i k32_m08_p24 = pair256_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m256i k32_m24_m08 = pair256_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m256i k32_p24_p08 = pair256_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[ 0] = _mm256_unpacklo_epi32(lstep3[36], lstep3[58]);
+          u[ 1] = _mm256_unpackhi_epi32(lstep3[36], lstep3[58]);
+          u[ 2] = _mm256_unpacklo_epi32(lstep3[37], lstep3[59]);
+          u[ 3] = _mm256_unpackhi_epi32(lstep3[37], lstep3[59]);
+          u[ 4] = _mm256_unpacklo_epi32(lstep3[38], lstep3[56]);
+          u[ 5] = _mm256_unpackhi_epi32(lstep3[38], lstep3[56]);
+          u[ 6] = _mm256_unpacklo_epi32(lstep3[39], lstep3[57]);
+          u[ 7] = _mm256_unpackhi_epi32(lstep3[39], lstep3[57]);
+          u[ 8] = _mm256_unpacklo_epi32(lstep3[40], lstep3[54]);
+          u[ 9] = _mm256_unpackhi_epi32(lstep3[40], lstep3[54]);
+          u[10] = _mm256_unpacklo_epi32(lstep3[41], lstep3[55]);
+          u[11] = _mm256_unpackhi_epi32(lstep3[41], lstep3[55]);
+          u[12] = _mm256_unpacklo_epi32(lstep3[42], lstep3[52]);
+          u[13] = _mm256_unpackhi_epi32(lstep3[42], lstep3[52]);
+          u[14] = _mm256_unpacklo_epi32(lstep3[43], lstep3[53]);
+          u[15] = _mm256_unpackhi_epi32(lstep3[43], lstep3[53]);
+
+          v[ 0] = k_madd_epi32_avx2(u[ 0], k32_m08_p24);
+          v[ 1] = k_madd_epi32_avx2(u[ 1], k32_m08_p24);
+          v[ 2] = k_madd_epi32_avx2(u[ 2], k32_m08_p24);
+          v[ 3] = k_madd_epi32_avx2(u[ 3], k32_m08_p24);
+          v[ 4] = k_madd_epi32_avx2(u[ 4], k32_m08_p24);
+          v[ 5] = k_madd_epi32_avx2(u[ 5], k32_m08_p24);
+          v[ 6] = k_madd_epi32_avx2(u[ 6], k32_m08_p24);
+          v[ 7] = k_madd_epi32_avx2(u[ 7], k32_m08_p24);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_m24_m08);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_m24_m08);
+          v[10] = k_madd_epi32_avx2(u[10], k32_m24_m08);
+          v[11] = k_madd_epi32_avx2(u[11], k32_m24_m08);
+          v[12] = k_madd_epi32_avx2(u[12], k32_m24_m08);
+          v[13] = k_madd_epi32_avx2(u[13], k32_m24_m08);
+          v[14] = k_madd_epi32_avx2(u[14], k32_m24_m08);
+          v[15] = k_madd_epi32_avx2(u[15], k32_m24_m08);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m08_p24);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m08_p24);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m08_p24);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m08_p24);
+          v[20] = k_madd_epi32_avx2(u[ 8], k32_m08_p24);
+          v[21] = k_madd_epi32_avx2(u[ 9], k32_m08_p24);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m08_p24);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m08_p24);
+          v[24] = k_madd_epi32_avx2(u[ 4], k32_p24_p08);
+          v[25] = k_madd_epi32_avx2(u[ 5], k32_p24_p08);
+          v[26] = k_madd_epi32_avx2(u[ 6], k32_p24_p08);
+          v[27] = k_madd_epi32_avx2(u[ 7], k32_p24_p08);
+          v[28] = k_madd_epi32_avx2(u[ 0], k32_p24_p08);
+          v[29] = k_madd_epi32_avx2(u[ 1], k32_p24_p08);
+          v[30] = k_madd_epi32_avx2(u[ 2], k32_p24_p08);
+          v[31] = k_madd_epi32_avx2(u[ 3], k32_p24_p08);
+
+          u[ 0] = k_packs_epi64_avx2(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64_avx2(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64_avx2(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64_avx2(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64_avx2(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[ 6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[ 7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[ 8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[ 9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm256_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm256_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm256_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm256_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm256_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm256_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm256_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm256_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm256_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep1[36] = _mm256_srai_epi32(v[ 0], DCT_CONST_BITS);
+          lstep1[37] = _mm256_srai_epi32(v[ 1], DCT_CONST_BITS);
+          lstep1[38] = _mm256_srai_epi32(v[ 2], DCT_CONST_BITS);
+          lstep1[39] = _mm256_srai_epi32(v[ 3], DCT_CONST_BITS);
+          lstep1[40] = _mm256_srai_epi32(v[ 4], DCT_CONST_BITS);
+          lstep1[41] = _mm256_srai_epi32(v[ 5], DCT_CONST_BITS);
+          lstep1[42] = _mm256_srai_epi32(v[ 6], DCT_CONST_BITS);
+          lstep1[43] = _mm256_srai_epi32(v[ 7], DCT_CONST_BITS);
+          lstep1[52] = _mm256_srai_epi32(v[ 8], DCT_CONST_BITS);
+          lstep1[53] = _mm256_srai_epi32(v[ 9], DCT_CONST_BITS);
+          lstep1[54] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep1[55] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep1[56] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep1[57] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep1[58] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep1[59] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 5
+        {
+          lstep2[ 8] = _mm256_add_epi32(lstep1[10], lstep3[ 8]);
+          lstep2[ 9] = _mm256_add_epi32(lstep1[11], lstep3[ 9]);
+          lstep2[10] = _mm256_sub_epi32(lstep3[ 8], lstep1[10]);
+          lstep2[11] = _mm256_sub_epi32(lstep3[ 9], lstep1[11]);
+          lstep2[12] = _mm256_sub_epi32(lstep3[14], lstep1[12]);
+          lstep2[13] = _mm256_sub_epi32(lstep3[15], lstep1[13]);
+          lstep2[14] = _mm256_add_epi32(lstep1[12], lstep3[14]);
+          lstep2[15] = _mm256_add_epi32(lstep1[13], lstep3[15]);
+        }
+        {
+          const __m256i k32_p16_p16 = pair256_set_epi32(cospi_16_64, cospi_16_64);
+          const __m256i k32_p16_m16 = pair256_set_epi32(cospi_16_64, -cospi_16_64);
+          const __m256i k32_p24_p08 = pair256_set_epi32(cospi_24_64, cospi_8_64);
+          const __m256i k32_m08_p24 = pair256_set_epi32(-cospi_8_64, cospi_24_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep1[0], lstep1[2]);
+          u[1] = _mm256_unpackhi_epi32(lstep1[0], lstep1[2]);
+          u[2] = _mm256_unpacklo_epi32(lstep1[1], lstep1[3]);
+          u[3] = _mm256_unpackhi_epi32(lstep1[1], lstep1[3]);
+          u[4] = _mm256_unpacklo_epi32(lstep1[4], lstep1[6]);
+          u[5] = _mm256_unpackhi_epi32(lstep1[4], lstep1[6]);
+          u[6] = _mm256_unpacklo_epi32(lstep1[5], lstep1[7]);
+          u[7] = _mm256_unpackhi_epi32(lstep1[5], lstep1[7]);
+
+          // TODO(jingning): manually inline k_madd_epi32_avx2_ to further hide
+          // instruction latency.
+          v[ 0] = k_madd_epi32_avx2(u[0], k32_p16_p16);
+          v[ 1] = k_madd_epi32_avx2(u[1], k32_p16_p16);
+          v[ 2] = k_madd_epi32_avx2(u[2], k32_p16_p16);
+          v[ 3] = k_madd_epi32_avx2(u[3], k32_p16_p16);
+          v[ 4] = k_madd_epi32_avx2(u[0], k32_p16_m16);
+          v[ 5] = k_madd_epi32_avx2(u[1], k32_p16_m16);
+          v[ 6] = k_madd_epi32_avx2(u[2], k32_p16_m16);
+          v[ 7] = k_madd_epi32_avx2(u[3], k32_p16_m16);
+          v[ 8] = k_madd_epi32_avx2(u[4], k32_p24_p08);
+          v[ 9] = k_madd_epi32_avx2(u[5], k32_p24_p08);
+          v[10] = k_madd_epi32_avx2(u[6], k32_p24_p08);
+          v[11] = k_madd_epi32_avx2(u[7], k32_p24_p08);
+          v[12] = k_madd_epi32_avx2(u[4], k32_m08_p24);
+          v[13] = k_madd_epi32_avx2(u[5], k32_m08_p24);
+          v[14] = k_madd_epi32_avx2(u[6], k32_m08_p24);
+          v[15] = k_madd_epi32_avx2(u[7], k32_m08_p24);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm256_cmpgt_epi32(kZero,u[0]);
+          sign[1] = _mm256_cmpgt_epi32(kZero,u[1]);
+          sign[2] = _mm256_cmpgt_epi32(kZero,u[2]);
+          sign[3] = _mm256_cmpgt_epi32(kZero,u[3]);
+          sign[4] = _mm256_cmpgt_epi32(kZero,u[4]);
+          sign[5] = _mm256_cmpgt_epi32(kZero,u[5]);
+          sign[6] = _mm256_cmpgt_epi32(kZero,u[6]);
+          sign[7] = _mm256_cmpgt_epi32(kZero,u[7]);
+
+          u[0] = _mm256_sub_epi32(u[0], sign[0]);
+          u[1] = _mm256_sub_epi32(u[1], sign[1]);
+          u[2] = _mm256_sub_epi32(u[2], sign[2]);
+          u[3] = _mm256_sub_epi32(u[3], sign[3]);
+          u[4] = _mm256_sub_epi32(u[4], sign[4]);
+          u[5] = _mm256_sub_epi32(u[5], sign[5]);
+          u[6] = _mm256_sub_epi32(u[6], sign[6]);
+          u[7] = _mm256_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm256_add_epi32(u[0], K32One);
+          u[1] = _mm256_add_epi32(u[1], K32One);
+          u[2] = _mm256_add_epi32(u[2], K32One);
+          u[3] = _mm256_add_epi32(u[3], K32One);
+          u[4] = _mm256_add_epi32(u[4], K32One);
+          u[5] = _mm256_add_epi32(u[5], K32One);
+          u[6] = _mm256_add_epi32(u[6], K32One);
+          u[7] = _mm256_add_epi32(u[7], K32One);
+
+          u[0] = _mm256_srai_epi32(u[0], 2);
+          u[1] = _mm256_srai_epi32(u[1], 2);
+          u[2] = _mm256_srai_epi32(u[2], 2);
+          u[3] = _mm256_srai_epi32(u[3], 2);
+          u[4] = _mm256_srai_epi32(u[4], 2);
+          u[5] = _mm256_srai_epi32(u[5], 2);
+          u[6] = _mm256_srai_epi32(u[6], 2);
+          u[7] = _mm256_srai_epi32(u[7], 2);
+
+          // Combine
+          out[ 0] = _mm256_packs_epi32(u[0], u[1]);
+          out[16] = _mm256_packs_epi32(u[2], u[3]);
+          out[ 8] = _mm256_packs_epi32(u[4], u[5]);
+          out[24] = _mm256_packs_epi32(u[6], u[7]);
+        }
+        {
+          const __m256i k32_m08_p24 = pair256_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m256i k32_m24_m08 = pair256_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m256i k32_p24_p08 = pair256_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep1[18], lstep1[28]);
+          u[1] = _mm256_unpackhi_epi32(lstep1[18], lstep1[28]);
+          u[2] = _mm256_unpacklo_epi32(lstep1[19], lstep1[29]);
+          u[3] = _mm256_unpackhi_epi32(lstep1[19], lstep1[29]);
+          u[4] = _mm256_unpacklo_epi32(lstep1[20], lstep1[26]);
+          u[5] = _mm256_unpackhi_epi32(lstep1[20], lstep1[26]);
+          u[6] = _mm256_unpacklo_epi32(lstep1[21], lstep1[27]);
+          u[7] = _mm256_unpackhi_epi32(lstep1[21], lstep1[27]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_m08_p24);
+          v[1] = k_madd_epi32_avx2(u[1], k32_m08_p24);
+          v[2] = k_madd_epi32_avx2(u[2], k32_m08_p24);
+          v[3] = k_madd_epi32_avx2(u[3], k32_m08_p24);
+          v[4] = k_madd_epi32_avx2(u[4], k32_m24_m08);
+          v[5] = k_madd_epi32_avx2(u[5], k32_m24_m08);
+          v[6] = k_madd_epi32_avx2(u[6], k32_m24_m08);
+          v[7] = k_madd_epi32_avx2(u[7], k32_m24_m08);
+          v[ 8] = k_madd_epi32_avx2(u[4], k32_m08_p24);
+          v[ 9] = k_madd_epi32_avx2(u[5], k32_m08_p24);
+          v[10] = k_madd_epi32_avx2(u[6], k32_m08_p24);
+          v[11] = k_madd_epi32_avx2(u[7], k32_m08_p24);
+          v[12] = k_madd_epi32_avx2(u[0], k32_p24_p08);
+          v[13] = k_madd_epi32_avx2(u[1], k32_p24_p08);
+          v[14] = k_madd_epi32_avx2(u[2], k32_p24_p08);
+          v[15] = k_madd_epi32_avx2(u[3], k32_p24_p08);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+
+          u[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          u[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          u[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          u[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          u[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          u[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          u[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          u[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          lstep2[18] = _mm256_srai_epi32(u[0], DCT_CONST_BITS);
+          lstep2[19] = _mm256_srai_epi32(u[1], DCT_CONST_BITS);
+          lstep2[20] = _mm256_srai_epi32(u[2], DCT_CONST_BITS);
+          lstep2[21] = _mm256_srai_epi32(u[3], DCT_CONST_BITS);
+          lstep2[26] = _mm256_srai_epi32(u[4], DCT_CONST_BITS);
+          lstep2[27] = _mm256_srai_epi32(u[5], DCT_CONST_BITS);
+          lstep2[28] = _mm256_srai_epi32(u[6], DCT_CONST_BITS);
+          lstep2[29] = _mm256_srai_epi32(u[7], DCT_CONST_BITS);
+        }
+        {
+          lstep2[32] = _mm256_add_epi32(lstep1[38], lstep3[32]);
+          lstep2[33] = _mm256_add_epi32(lstep1[39], lstep3[33]);
+          lstep2[34] = _mm256_add_epi32(lstep1[36], lstep3[34]);
+          lstep2[35] = _mm256_add_epi32(lstep1[37], lstep3[35]);
+          lstep2[36] = _mm256_sub_epi32(lstep3[34], lstep1[36]);
+          lstep2[37] = _mm256_sub_epi32(lstep3[35], lstep1[37]);
+          lstep2[38] = _mm256_sub_epi32(lstep3[32], lstep1[38]);
+          lstep2[39] = _mm256_sub_epi32(lstep3[33], lstep1[39]);
+          lstep2[40] = _mm256_sub_epi32(lstep3[46], lstep1[40]);
+          lstep2[41] = _mm256_sub_epi32(lstep3[47], lstep1[41]);
+          lstep2[42] = _mm256_sub_epi32(lstep3[44], lstep1[42]);
+          lstep2[43] = _mm256_sub_epi32(lstep3[45], lstep1[43]);
+          lstep2[44] = _mm256_add_epi32(lstep1[42], lstep3[44]);
+          lstep2[45] = _mm256_add_epi32(lstep1[43], lstep3[45]);
+          lstep2[46] = _mm256_add_epi32(lstep1[40], lstep3[46]);
+          lstep2[47] = _mm256_add_epi32(lstep1[41], lstep3[47]);
+          lstep2[48] = _mm256_add_epi32(lstep1[54], lstep3[48]);
+          lstep2[49] = _mm256_add_epi32(lstep1[55], lstep3[49]);
+          lstep2[50] = _mm256_add_epi32(lstep1[52], lstep3[50]);
+          lstep2[51] = _mm256_add_epi32(lstep1[53], lstep3[51]);
+          lstep2[52] = _mm256_sub_epi32(lstep3[50], lstep1[52]);
+          lstep2[53] = _mm256_sub_epi32(lstep3[51], lstep1[53]);
+          lstep2[54] = _mm256_sub_epi32(lstep3[48], lstep1[54]);
+          lstep2[55] = _mm256_sub_epi32(lstep3[49], lstep1[55]);
+          lstep2[56] = _mm256_sub_epi32(lstep3[62], lstep1[56]);
+          lstep2[57] = _mm256_sub_epi32(lstep3[63], lstep1[57]);
+          lstep2[58] = _mm256_sub_epi32(lstep3[60], lstep1[58]);
+          lstep2[59] = _mm256_sub_epi32(lstep3[61], lstep1[59]);
+          lstep2[60] = _mm256_add_epi32(lstep1[58], lstep3[60]);
+          lstep2[61] = _mm256_add_epi32(lstep1[59], lstep3[61]);
+          lstep2[62] = _mm256_add_epi32(lstep1[56], lstep3[62]);
+          lstep2[63] = _mm256_add_epi32(lstep1[57], lstep3[63]);
+        }
+        // stage 6
+        {
+          const __m256i k32_p28_p04 = pair256_set_epi32(cospi_28_64, cospi_4_64);
+          const __m256i k32_p12_p20 = pair256_set_epi32(cospi_12_64, cospi_20_64);
+          const __m256i k32_m20_p12 = pair256_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m256i k32_m04_p28 = pair256_set_epi32(-cospi_4_64, cospi_28_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep2[ 8], lstep2[14]);
+          u[1] = _mm256_unpackhi_epi32(lstep2[ 8], lstep2[14]);
+          u[2] = _mm256_unpacklo_epi32(lstep2[ 9], lstep2[15]);
+          u[3] = _mm256_unpackhi_epi32(lstep2[ 9], lstep2[15]);
+          u[4] = _mm256_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[5] = _mm256_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[6] = _mm256_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[7] = _mm256_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[8] = _mm256_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[9] = _mm256_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[10] = _mm256_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[11] = _mm256_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[12] = _mm256_unpacklo_epi32(lstep2[ 8], lstep2[14]);
+          u[13] = _mm256_unpackhi_epi32(lstep2[ 8], lstep2[14]);
+          u[14] = _mm256_unpacklo_epi32(lstep2[ 9], lstep2[15]);
+          u[15] = _mm256_unpackhi_epi32(lstep2[ 9], lstep2[15]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_p28_p04);
+          v[1] = k_madd_epi32_avx2(u[1], k32_p28_p04);
+          v[2] = k_madd_epi32_avx2(u[2], k32_p28_p04);
+          v[3] = k_madd_epi32_avx2(u[3], k32_p28_p04);
+          v[4] = k_madd_epi32_avx2(u[4], k32_p12_p20);
+          v[5] = k_madd_epi32_avx2(u[5], k32_p12_p20);
+          v[6] = k_madd_epi32_avx2(u[6], k32_p12_p20);
+          v[7] = k_madd_epi32_avx2(u[7], k32_p12_p20);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_m20_p12);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_m20_p12);
+          v[10] = k_madd_epi32_avx2(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32_avx2(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32_avx2(u[12], k32_m04_p28);
+          v[13] = k_madd_epi32_avx2(u[13], k32_m04_p28);
+          v[14] = k_madd_epi32_avx2(u[14], k32_m04_p28);
+          v[15] = k_madd_epi32_avx2(u[15], k32_m04_p28);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm256_cmpgt_epi32(kZero,u[0]);
+          sign[1] = _mm256_cmpgt_epi32(kZero,u[1]);
+          sign[2] = _mm256_cmpgt_epi32(kZero,u[2]);
+          sign[3] = _mm256_cmpgt_epi32(kZero,u[3]);
+          sign[4] = _mm256_cmpgt_epi32(kZero,u[4]);
+          sign[5] = _mm256_cmpgt_epi32(kZero,u[5]);
+          sign[6] = _mm256_cmpgt_epi32(kZero,u[6]);
+          sign[7] = _mm256_cmpgt_epi32(kZero,u[7]);
+
+          u[0] = _mm256_sub_epi32(u[0], sign[0]);
+          u[1] = _mm256_sub_epi32(u[1], sign[1]);
+          u[2] = _mm256_sub_epi32(u[2], sign[2]);
+          u[3] = _mm256_sub_epi32(u[3], sign[3]);
+          u[4] = _mm256_sub_epi32(u[4], sign[4]);
+          u[5] = _mm256_sub_epi32(u[5], sign[5]);
+          u[6] = _mm256_sub_epi32(u[6], sign[6]);
+          u[7] = _mm256_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm256_add_epi32(u[0], K32One);
+          u[1] = _mm256_add_epi32(u[1], K32One);
+          u[2] = _mm256_add_epi32(u[2], K32One);
+          u[3] = _mm256_add_epi32(u[3], K32One);
+          u[4] = _mm256_add_epi32(u[4], K32One);
+          u[5] = _mm256_add_epi32(u[5], K32One);
+          u[6] = _mm256_add_epi32(u[6], K32One);
+          u[7] = _mm256_add_epi32(u[7], K32One);
+
+          u[0] = _mm256_srai_epi32(u[0], 2);
+          u[1] = _mm256_srai_epi32(u[1], 2);
+          u[2] = _mm256_srai_epi32(u[2], 2);
+          u[3] = _mm256_srai_epi32(u[3], 2);
+          u[4] = _mm256_srai_epi32(u[4], 2);
+          u[5] = _mm256_srai_epi32(u[5], 2);
+          u[6] = _mm256_srai_epi32(u[6], 2);
+          u[7] = _mm256_srai_epi32(u[7], 2);
+
+          out[ 4] = _mm256_packs_epi32(u[0], u[1]);
+          out[20] = _mm256_packs_epi32(u[2], u[3]);
+          out[12] = _mm256_packs_epi32(u[4], u[5]);
+          out[28] = _mm256_packs_epi32(u[6], u[7]);
+        }
+        {
+          lstep3[16] = _mm256_add_epi32(lstep2[18], lstep1[16]);
+          lstep3[17] = _mm256_add_epi32(lstep2[19], lstep1[17]);
+          lstep3[18] = _mm256_sub_epi32(lstep1[16], lstep2[18]);
+          lstep3[19] = _mm256_sub_epi32(lstep1[17], lstep2[19]);
+          lstep3[20] = _mm256_sub_epi32(lstep1[22], lstep2[20]);
+          lstep3[21] = _mm256_sub_epi32(lstep1[23], lstep2[21]);
+          lstep3[22] = _mm256_add_epi32(lstep2[20], lstep1[22]);
+          lstep3[23] = _mm256_add_epi32(lstep2[21], lstep1[23]);
+          lstep3[24] = _mm256_add_epi32(lstep2[26], lstep1[24]);
+          lstep3[25] = _mm256_add_epi32(lstep2[27], lstep1[25]);
+          lstep3[26] = _mm256_sub_epi32(lstep1[24], lstep2[26]);
+          lstep3[27] = _mm256_sub_epi32(lstep1[25], lstep2[27]);
+          lstep3[28] = _mm256_sub_epi32(lstep1[30], lstep2[28]);
+          lstep3[29] = _mm256_sub_epi32(lstep1[31], lstep2[29]);
+          lstep3[30] = _mm256_add_epi32(lstep2[28], lstep1[30]);
+          lstep3[31] = _mm256_add_epi32(lstep2[29], lstep1[31]);
+        }
+        {
+          const __m256i k32_m04_p28 = pair256_set_epi32(-cospi_4_64, cospi_28_64);
+          const __m256i k32_m28_m04 = pair256_set_epi32(-cospi_28_64, -cospi_4_64);
+          const __m256i k32_m20_p12 = pair256_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m256i k32_m12_m20 = pair256_set_epi32(-cospi_12_64,
+                                                     -cospi_20_64);
+          const __m256i k32_p12_p20 = pair256_set_epi32(cospi_12_64, cospi_20_64);
+          const __m256i k32_p28_p04 = pair256_set_epi32(cospi_28_64, cospi_4_64);
+
+          u[ 0] = _mm256_unpacklo_epi32(lstep2[34], lstep2[60]);
+          u[ 1] = _mm256_unpackhi_epi32(lstep2[34], lstep2[60]);
+          u[ 2] = _mm256_unpacklo_epi32(lstep2[35], lstep2[61]);
+          u[ 3] = _mm256_unpackhi_epi32(lstep2[35], lstep2[61]);
+          u[ 4] = _mm256_unpacklo_epi32(lstep2[36], lstep2[58]);
+          u[ 5] = _mm256_unpackhi_epi32(lstep2[36], lstep2[58]);
+          u[ 6] = _mm256_unpacklo_epi32(lstep2[37], lstep2[59]);
+          u[ 7] = _mm256_unpackhi_epi32(lstep2[37], lstep2[59]);
+          u[ 8] = _mm256_unpacklo_epi32(lstep2[42], lstep2[52]);
+          u[ 9] = _mm256_unpackhi_epi32(lstep2[42], lstep2[52]);
+          u[10] = _mm256_unpacklo_epi32(lstep2[43], lstep2[53]);
+          u[11] = _mm256_unpackhi_epi32(lstep2[43], lstep2[53]);
+          u[12] = _mm256_unpacklo_epi32(lstep2[44], lstep2[50]);
+          u[13] = _mm256_unpackhi_epi32(lstep2[44], lstep2[50]);
+          u[14] = _mm256_unpacklo_epi32(lstep2[45], lstep2[51]);
+          u[15] = _mm256_unpackhi_epi32(lstep2[45], lstep2[51]);
+
+          v[ 0] = k_madd_epi32_avx2(u[ 0], k32_m04_p28);
+          v[ 1] = k_madd_epi32_avx2(u[ 1], k32_m04_p28);
+          v[ 2] = k_madd_epi32_avx2(u[ 2], k32_m04_p28);
+          v[ 3] = k_madd_epi32_avx2(u[ 3], k32_m04_p28);
+          v[ 4] = k_madd_epi32_avx2(u[ 4], k32_m28_m04);
+          v[ 5] = k_madd_epi32_avx2(u[ 5], k32_m28_m04);
+          v[ 6] = k_madd_epi32_avx2(u[ 6], k32_m28_m04);
+          v[ 7] = k_madd_epi32_avx2(u[ 7], k32_m28_m04);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_m20_p12);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_m20_p12);
+          v[10] = k_madd_epi32_avx2(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32_avx2(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32_avx2(u[12], k32_m12_m20);
+          v[13] = k_madd_epi32_avx2(u[13], k32_m12_m20);
+          v[14] = k_madd_epi32_avx2(u[14], k32_m12_m20);
+          v[15] = k_madd_epi32_avx2(u[15], k32_m12_m20);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m20_p12);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m20_p12);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m20_p12);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m20_p12);
+          v[20] = k_madd_epi32_avx2(u[ 8], k32_p12_p20);
+          v[21] = k_madd_epi32_avx2(u[ 9], k32_p12_p20);
+          v[22] = k_madd_epi32_avx2(u[10], k32_p12_p20);
+          v[23] = k_madd_epi32_avx2(u[11], k32_p12_p20);
+          v[24] = k_madd_epi32_avx2(u[ 4], k32_m04_p28);
+          v[25] = k_madd_epi32_avx2(u[ 5], k32_m04_p28);
+          v[26] = k_madd_epi32_avx2(u[ 6], k32_m04_p28);
+          v[27] = k_madd_epi32_avx2(u[ 7], k32_m04_p28);
+          v[28] = k_madd_epi32_avx2(u[ 0], k32_p28_p04);
+          v[29] = k_madd_epi32_avx2(u[ 1], k32_p28_p04);
+          v[30] = k_madd_epi32_avx2(u[ 2], k32_p28_p04);
+          v[31] = k_madd_epi32_avx2(u[ 3], k32_p28_p04);
+
+          u[ 0] = k_packs_epi64_avx2(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64_avx2(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64_avx2(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64_avx2(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64_avx2(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[ 6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[ 7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[ 8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[ 9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm256_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm256_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm256_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm256_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm256_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm256_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm256_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm256_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm256_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep3[34] = _mm256_srai_epi32(v[ 0], DCT_CONST_BITS);
+          lstep3[35] = _mm256_srai_epi32(v[ 1], DCT_CONST_BITS);
+          lstep3[36] = _mm256_srai_epi32(v[ 2], DCT_CONST_BITS);
+          lstep3[37] = _mm256_srai_epi32(v[ 3], DCT_CONST_BITS);
+          lstep3[42] = _mm256_srai_epi32(v[ 4], DCT_CONST_BITS);
+          lstep3[43] = _mm256_srai_epi32(v[ 5], DCT_CONST_BITS);
+          lstep3[44] = _mm256_srai_epi32(v[ 6], DCT_CONST_BITS);
+          lstep3[45] = _mm256_srai_epi32(v[ 7], DCT_CONST_BITS);
+          lstep3[50] = _mm256_srai_epi32(v[ 8], DCT_CONST_BITS);
+          lstep3[51] = _mm256_srai_epi32(v[ 9], DCT_CONST_BITS);
+          lstep3[52] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep3[53] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep3[58] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep3[59] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep3[60] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep3[61] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 7
+        {
+          const __m256i k32_p30_p02 = pair256_set_epi32(cospi_30_64, cospi_2_64);
+          const __m256i k32_p14_p18 = pair256_set_epi32(cospi_14_64, cospi_18_64);
+          const __m256i k32_p22_p10 = pair256_set_epi32(cospi_22_64, cospi_10_64);
+          const __m256i k32_p06_p26 = pair256_set_epi32(cospi_6_64,  cospi_26_64);
+          const __m256i k32_m26_p06 = pair256_set_epi32(-cospi_26_64, cospi_6_64);
+          const __m256i k32_m10_p22 = pair256_set_epi32(-cospi_10_64, cospi_22_64);
+          const __m256i k32_m18_p14 = pair256_set_epi32(-cospi_18_64, cospi_14_64);
+          const __m256i k32_m02_p30 = pair256_set_epi32(-cospi_2_64, cospi_30_64);
+
+          u[ 0] = _mm256_unpacklo_epi32(lstep3[16], lstep3[30]);
+          u[ 1] = _mm256_unpackhi_epi32(lstep3[16], lstep3[30]);
+          u[ 2] = _mm256_unpacklo_epi32(lstep3[17], lstep3[31]);
+          u[ 3] = _mm256_unpackhi_epi32(lstep3[17], lstep3[31]);
+          u[ 4] = _mm256_unpacklo_epi32(lstep3[18], lstep3[28]);
+          u[ 5] = _mm256_unpackhi_epi32(lstep3[18], lstep3[28]);
+          u[ 6] = _mm256_unpacklo_epi32(lstep3[19], lstep3[29]);
+          u[ 7] = _mm256_unpackhi_epi32(lstep3[19], lstep3[29]);
+          u[ 8] = _mm256_unpacklo_epi32(lstep3[20], lstep3[26]);
+          u[ 9] = _mm256_unpackhi_epi32(lstep3[20], lstep3[26]);
+          u[10] = _mm256_unpacklo_epi32(lstep3[21], lstep3[27]);
+          u[11] = _mm256_unpackhi_epi32(lstep3[21], lstep3[27]);
+          u[12] = _mm256_unpacklo_epi32(lstep3[22], lstep3[24]);
+          u[13] = _mm256_unpackhi_epi32(lstep3[22], lstep3[24]);
+          u[14] = _mm256_unpacklo_epi32(lstep3[23], lstep3[25]);
+          u[15] = _mm256_unpackhi_epi32(lstep3[23], lstep3[25]);
+
+          v[ 0] = k_madd_epi32_avx2(u[ 0], k32_p30_p02);
+          v[ 1] = k_madd_epi32_avx2(u[ 1], k32_p30_p02);
+          v[ 2] = k_madd_epi32_avx2(u[ 2], k32_p30_p02);
+          v[ 3] = k_madd_epi32_avx2(u[ 3], k32_p30_p02);
+          v[ 4] = k_madd_epi32_avx2(u[ 4], k32_p14_p18);
+          v[ 5] = k_madd_epi32_avx2(u[ 5], k32_p14_p18);
+          v[ 6] = k_madd_epi32_avx2(u[ 6], k32_p14_p18);
+          v[ 7] = k_madd_epi32_avx2(u[ 7], k32_p14_p18);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_p22_p10);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_p22_p10);
+          v[10] = k_madd_epi32_avx2(u[10], k32_p22_p10);
+          v[11] = k_madd_epi32_avx2(u[11], k32_p22_p10);
+          v[12] = k_madd_epi32_avx2(u[12], k32_p06_p26);
+          v[13] = k_madd_epi32_avx2(u[13], k32_p06_p26);
+          v[14] = k_madd_epi32_avx2(u[14], k32_p06_p26);
+          v[15] = k_madd_epi32_avx2(u[15], k32_p06_p26);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m26_p06);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m26_p06);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m26_p06);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m26_p06);
+          v[20] = k_madd_epi32_avx2(u[ 8], k32_m10_p22);
+          v[21] = k_madd_epi32_avx2(u[ 9], k32_m10_p22);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m10_p22);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m10_p22);
+          v[24] = k_madd_epi32_avx2(u[ 4], k32_m18_p14);
+          v[25] = k_madd_epi32_avx2(u[ 5], k32_m18_p14);
+          v[26] = k_madd_epi32_avx2(u[ 6], k32_m18_p14);
+          v[27] = k_madd_epi32_avx2(u[ 7], k32_m18_p14);
+          v[28] = k_madd_epi32_avx2(u[ 0], k32_m02_p30);
+          v[29] = k_madd_epi32_avx2(u[ 1], k32_m02_p30);
+          v[30] = k_madd_epi32_avx2(u[ 2], k32_m02_p30);
+          v[31] = k_madd_epi32_avx2(u[ 3], k32_m02_p30);
+
+          u[ 0] = k_packs_epi64_avx2(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64_avx2(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64_avx2(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64_avx2(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64_avx2(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[ 6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[ 7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[ 8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[ 9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm256_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm256_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm256_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm256_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm256_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm256_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm256_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm256_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm256_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm256_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm256_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm256_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm256_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm256_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm256_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm256_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm256_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm256_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm256_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm256_cmpgt_epi32(kZero,u[ 0]);
+          v[ 1] = _mm256_cmpgt_epi32(kZero,u[ 1]);
+          v[ 2] = _mm256_cmpgt_epi32(kZero,u[ 2]);
+          v[ 3] = _mm256_cmpgt_epi32(kZero,u[ 3]);
+          v[ 4] = _mm256_cmpgt_epi32(kZero,u[ 4]);
+          v[ 5] = _mm256_cmpgt_epi32(kZero,u[ 5]);
+          v[ 6] = _mm256_cmpgt_epi32(kZero,u[ 6]);
+          v[ 7] = _mm256_cmpgt_epi32(kZero,u[ 7]);
+          v[ 8] = _mm256_cmpgt_epi32(kZero,u[ 8]);
+          v[ 9] = _mm256_cmpgt_epi32(kZero,u[ 9]);
+          v[10] = _mm256_cmpgt_epi32(kZero,u[10]);
+          v[11] = _mm256_cmpgt_epi32(kZero,u[11]);
+          v[12] = _mm256_cmpgt_epi32(kZero,u[12]);
+          v[13] = _mm256_cmpgt_epi32(kZero,u[13]);
+          v[14] = _mm256_cmpgt_epi32(kZero,u[14]);
+          v[15] = _mm256_cmpgt_epi32(kZero,u[15]);
+
+          u[ 0] = _mm256_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm256_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm256_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm256_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm256_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm256_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm256_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm256_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm256_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm256_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm256_sub_epi32(u[10], v[10]);
+          u[11] = _mm256_sub_epi32(u[11], v[11]);
+          u[12] = _mm256_sub_epi32(u[12], v[12]);
+          u[13] = _mm256_sub_epi32(u[13], v[13]);
+          u[14] = _mm256_sub_epi32(u[14], v[14]);
+          u[15] = _mm256_sub_epi32(u[15], v[15]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], K32One);
+          v[ 1] = _mm256_add_epi32(u[ 1], K32One);
+          v[ 2] = _mm256_add_epi32(u[ 2], K32One);
+          v[ 3] = _mm256_add_epi32(u[ 3], K32One);
+          v[ 4] = _mm256_add_epi32(u[ 4], K32One);
+          v[ 5] = _mm256_add_epi32(u[ 5], K32One);
+          v[ 6] = _mm256_add_epi32(u[ 6], K32One);
+          v[ 7] = _mm256_add_epi32(u[ 7], K32One);
+          v[ 8] = _mm256_add_epi32(u[ 8], K32One);
+          v[ 9] = _mm256_add_epi32(u[ 9], K32One);
+          v[10] = _mm256_add_epi32(u[10], K32One);
+          v[11] = _mm256_add_epi32(u[11], K32One);
+          v[12] = _mm256_add_epi32(u[12], K32One);
+          v[13] = _mm256_add_epi32(u[13], K32One);
+          v[14] = _mm256_add_epi32(u[14], K32One);
+          v[15] = _mm256_add_epi32(u[15], K32One);
+
+          u[ 0] = _mm256_srai_epi32(v[ 0], 2);
+          u[ 1] = _mm256_srai_epi32(v[ 1], 2);
+          u[ 2] = _mm256_srai_epi32(v[ 2], 2);
+          u[ 3] = _mm256_srai_epi32(v[ 3], 2);
+          u[ 4] = _mm256_srai_epi32(v[ 4], 2);
+          u[ 5] = _mm256_srai_epi32(v[ 5], 2);
+          u[ 6] = _mm256_srai_epi32(v[ 6], 2);
+          u[ 7] = _mm256_srai_epi32(v[ 7], 2);
+          u[ 8] = _mm256_srai_epi32(v[ 8], 2);
+          u[ 9] = _mm256_srai_epi32(v[ 9], 2);
+          u[10] = _mm256_srai_epi32(v[10], 2);
+          u[11] = _mm256_srai_epi32(v[11], 2);
+          u[12] = _mm256_srai_epi32(v[12], 2);
+          u[13] = _mm256_srai_epi32(v[13], 2);
+          u[14] = _mm256_srai_epi32(v[14], 2);
+          u[15] = _mm256_srai_epi32(v[15], 2);
+
+          out[ 2] = _mm256_packs_epi32(u[0], u[1]);
+          out[18] = _mm256_packs_epi32(u[2], u[3]);
+          out[10] = _mm256_packs_epi32(u[4], u[5]);
+          out[26] = _mm256_packs_epi32(u[6], u[7]);
+          out[ 6] = _mm256_packs_epi32(u[8], u[9]);
+          out[22] = _mm256_packs_epi32(u[10], u[11]);
+          out[14] = _mm256_packs_epi32(u[12], u[13]);
+          out[30] = _mm256_packs_epi32(u[14], u[15]);
+        }
+        {
+          lstep1[32] = _mm256_add_epi32(lstep3[34], lstep2[32]);
+          lstep1[33] = _mm256_add_epi32(lstep3[35], lstep2[33]);
+          lstep1[34] = _mm256_sub_epi32(lstep2[32], lstep3[34]);
+          lstep1[35] = _mm256_sub_epi32(lstep2[33], lstep3[35]);
+          lstep1[36] = _mm256_sub_epi32(lstep2[38], lstep3[36]);
+          lstep1[37] = _mm256_sub_epi32(lstep2[39], lstep3[37]);
+          lstep1[38] = _mm256_add_epi32(lstep3[36], lstep2[38]);
+          lstep1[39] = _mm256_add_epi32(lstep3[37], lstep2[39]);
+          lstep1[40] = _mm256_add_epi32(lstep3[42], lstep2[40]);
+          lstep1[41] = _mm256_add_epi32(lstep3[43], lstep2[41]);
+          lstep1[42] = _mm256_sub_epi32(lstep2[40], lstep3[42]);
+          lstep1[43] = _mm256_sub_epi32(lstep2[41], lstep3[43]);
+          lstep1[44] = _mm256_sub_epi32(lstep2[46], lstep3[44]);
+          lstep1[45] = _mm256_sub_epi32(lstep2[47], lstep3[45]);
+          lstep1[46] = _mm256_add_epi32(lstep3[44], lstep2[46]);
+          lstep1[47] = _mm256_add_epi32(lstep3[45], lstep2[47]);
+          lstep1[48] = _mm256_add_epi32(lstep3[50], lstep2[48]);
+          lstep1[49] = _mm256_add_epi32(lstep3[51], lstep2[49]);
+          lstep1[50] = _mm256_sub_epi32(lstep2[48], lstep3[50]);
+          lstep1[51] = _mm256_sub_epi32(lstep2[49], lstep3[51]);
+          lstep1[52] = _mm256_sub_epi32(lstep2[54], lstep3[52]);
+          lstep1[53] = _mm256_sub_epi32(lstep2[55], lstep3[53]);
+          lstep1[54] = _mm256_add_epi32(lstep3[52], lstep2[54]);
+          lstep1[55] = _mm256_add_epi32(lstep3[53], lstep2[55]);
+          lstep1[56] = _mm256_add_epi32(lstep3[58], lstep2[56]);
+          lstep1[57] = _mm256_add_epi32(lstep3[59], lstep2[57]);
+          lstep1[58] = _mm256_sub_epi32(lstep2[56], lstep3[58]);
+          lstep1[59] = _mm256_sub_epi32(lstep2[57], lstep3[59]);
+          lstep1[60] = _mm256_sub_epi32(lstep2[62], lstep3[60]);
+          lstep1[61] = _mm256_sub_epi32(lstep2[63], lstep3[61]);
+          lstep1[62] = _mm256_add_epi32(lstep3[60], lstep2[62]);
+          lstep1[63] = _mm256_add_epi32(lstep3[61], lstep2[63]);
+        }
+        // stage 8
+        {
+          const __m256i k32_p31_p01 = pair256_set_epi32(cospi_31_64, cospi_1_64);
+          const __m256i k32_p15_p17 = pair256_set_epi32(cospi_15_64, cospi_17_64);
+          const __m256i k32_p23_p09 = pair256_set_epi32(cospi_23_64, cospi_9_64);
+          const __m256i k32_p07_p25 = pair256_set_epi32(cospi_7_64, cospi_25_64);
+          const __m256i k32_m25_p07 = pair256_set_epi32(-cospi_25_64, cospi_7_64);
+          const __m256i k32_m09_p23 = pair256_set_epi32(-cospi_9_64, cospi_23_64);
+          const __m256i k32_m17_p15 = pair256_set_epi32(-cospi_17_64, cospi_15_64);
+          const __m256i k32_m01_p31 = pair256_set_epi32(-cospi_1_64, cospi_31_64);
+
+          u[ 0] = _mm256_unpacklo_epi32(lstep1[32], lstep1[62]);
+          u[ 1] = _mm256_unpackhi_epi32(lstep1[32], lstep1[62]);
+          u[ 2] = _mm256_unpacklo_epi32(lstep1[33], lstep1[63]);
+          u[ 3] = _mm256_unpackhi_epi32(lstep1[33], lstep1[63]);
+          u[ 4] = _mm256_unpacklo_epi32(lstep1[34], lstep1[60]);
+          u[ 5] = _mm256_unpackhi_epi32(lstep1[34], lstep1[60]);
+          u[ 6] = _mm256_unpacklo_epi32(lstep1[35], lstep1[61]);
+          u[ 7] = _mm256_unpackhi_epi32(lstep1[35], lstep1[61]);
+          u[ 8] = _mm256_unpacklo_epi32(lstep1[36], lstep1[58]);
+          u[ 9] = _mm256_unpackhi_epi32(lstep1[36], lstep1[58]);
+          u[10] = _mm256_unpacklo_epi32(lstep1[37], lstep1[59]);
+          u[11] = _mm256_unpackhi_epi32(lstep1[37], lstep1[59]);
+          u[12] = _mm256_unpacklo_epi32(lstep1[38], lstep1[56]);
+          u[13] = _mm256_unpackhi_epi32(lstep1[38], lstep1[56]);
+          u[14] = _mm256_unpacklo_epi32(lstep1[39], lstep1[57]);
+          u[15] = _mm256_unpackhi_epi32(lstep1[39], lstep1[57]);
+
+          v[ 0] = k_madd_epi32_avx2(u[ 0], k32_p31_p01);
+          v[ 1] = k_madd_epi32_avx2(u[ 1], k32_p31_p01);
+          v[ 2] = k_madd_epi32_avx2(u[ 2], k32_p31_p01);
+          v[ 3] = k_madd_epi32_avx2(u[ 3], k32_p31_p01);
+          v[ 4] = k_madd_epi32_avx2(u[ 4], k32_p15_p17);
+          v[ 5] = k_madd_epi32_avx2(u[ 5], k32_p15_p17);
+          v[ 6] = k_madd_epi32_avx2(u[ 6], k32_p15_p17);
+          v[ 7] = k_madd_epi32_avx2(u[ 7], k32_p15_p17);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_p23_p09);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_p23_p09);
+          v[10] = k_madd_epi32_avx2(u[10], k32_p23_p09);
+          v[11] = k_madd_epi32_avx2(u[11], k32_p23_p09);
+          v[12] = k_madd_epi32_avx2(u[12], k32_p07_p25);
+          v[13] = k_madd_epi32_avx2(u[13], k32_p07_p25);
+          v[14] = k_madd_epi32_avx2(u[14], k32_p07_p25);
+          v[15] = k_madd_epi32_avx2(u[15], k32_p07_p25);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m25_p07);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m25_p07);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m25_p07);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m25_p07);
+          v[20] = k_madd_epi32_avx2(u[ 8], k32_m09_p23);
+          v[21] = k_madd_epi32_avx2(u[ 9], k32_m09_p23);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m09_p23);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m09_p23);
+          v[24] = k_madd_epi32_avx2(u[ 4], k32_m17_p15);
+          v[25] = k_madd_epi32_avx2(u[ 5], k32_m17_p15);
+          v[26] = k_madd_epi32_avx2(u[ 6], k32_m17_p15);
+          v[27] = k_madd_epi32_avx2(u[ 7], k32_m17_p15);
+          v[28] = k_madd_epi32_avx2(u[ 0], k32_m01_p31);
+          v[29] = k_madd_epi32_avx2(u[ 1], k32_m01_p31);
+          v[30] = k_madd_epi32_avx2(u[ 2], k32_m01_p31);
+          v[31] = k_madd_epi32_avx2(u[ 3], k32_m01_p31);
+
+          u[ 0] = k_packs_epi64_avx2(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64_avx2(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64_avx2(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64_avx2(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64_avx2(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[ 6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[ 7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[ 8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[ 9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm256_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm256_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm256_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm256_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm256_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm256_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm256_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm256_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm256_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm256_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm256_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm256_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm256_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm256_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm256_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm256_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm256_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm256_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm256_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm256_cmpgt_epi32(kZero,u[ 0]);
+          v[ 1] = _mm256_cmpgt_epi32(kZero,u[ 1]);
+          v[ 2] = _mm256_cmpgt_epi32(kZero,u[ 2]);
+          v[ 3] = _mm256_cmpgt_epi32(kZero,u[ 3]);
+          v[ 4] = _mm256_cmpgt_epi32(kZero,u[ 4]);
+          v[ 5] = _mm256_cmpgt_epi32(kZero,u[ 5]);
+          v[ 6] = _mm256_cmpgt_epi32(kZero,u[ 6]);
+          v[ 7] = _mm256_cmpgt_epi32(kZero,u[ 7]);
+          v[ 8] = _mm256_cmpgt_epi32(kZero,u[ 8]);
+          v[ 9] = _mm256_cmpgt_epi32(kZero,u[ 9]);
+          v[10] = _mm256_cmpgt_epi32(kZero,u[10]);
+          v[11] = _mm256_cmpgt_epi32(kZero,u[11]);
+          v[12] = _mm256_cmpgt_epi32(kZero,u[12]);
+          v[13] = _mm256_cmpgt_epi32(kZero,u[13]);
+          v[14] = _mm256_cmpgt_epi32(kZero,u[14]);
+          v[15] = _mm256_cmpgt_epi32(kZero,u[15]);
+
+          u[ 0] = _mm256_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm256_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm256_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm256_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm256_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm256_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm256_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm256_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm256_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm256_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm256_sub_epi32(u[10], v[10]);
+          u[11] = _mm256_sub_epi32(u[11], v[11]);
+          u[12] = _mm256_sub_epi32(u[12], v[12]);
+          u[13] = _mm256_sub_epi32(u[13], v[13]);
+          u[14] = _mm256_sub_epi32(u[14], v[14]);
+          u[15] = _mm256_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], K32One);
+          v[1] = _mm256_add_epi32(u[1], K32One);
+          v[2] = _mm256_add_epi32(u[2], K32One);
+          v[3] = _mm256_add_epi32(u[3], K32One);
+          v[4] = _mm256_add_epi32(u[4], K32One);
+          v[5] = _mm256_add_epi32(u[5], K32One);
+          v[6] = _mm256_add_epi32(u[6], K32One);
+          v[7] = _mm256_add_epi32(u[7], K32One);
+          v[8] = _mm256_add_epi32(u[8], K32One);
+          v[9] = _mm256_add_epi32(u[9], K32One);
+          v[10] = _mm256_add_epi32(u[10], K32One);
+          v[11] = _mm256_add_epi32(u[11], K32One);
+          v[12] = _mm256_add_epi32(u[12], K32One);
+          v[13] = _mm256_add_epi32(u[13], K32One);
+          v[14] = _mm256_add_epi32(u[14], K32One);
+          v[15] = _mm256_add_epi32(u[15], K32One);
+
+          u[0] = _mm256_srai_epi32(v[0], 2);
+          u[1] = _mm256_srai_epi32(v[1], 2);
+          u[2] = _mm256_srai_epi32(v[2], 2);
+          u[3] = _mm256_srai_epi32(v[3], 2);
+          u[4] = _mm256_srai_epi32(v[4], 2);
+          u[5] = _mm256_srai_epi32(v[5], 2);
+          u[6] = _mm256_srai_epi32(v[6], 2);
+          u[7] = _mm256_srai_epi32(v[7], 2);
+          u[8] = _mm256_srai_epi32(v[8], 2);
+          u[9] = _mm256_srai_epi32(v[9], 2);
+          u[10] = _mm256_srai_epi32(v[10], 2);
+          u[11] = _mm256_srai_epi32(v[11], 2);
+          u[12] = _mm256_srai_epi32(v[12], 2);
+          u[13] = _mm256_srai_epi32(v[13], 2);
+          u[14] = _mm256_srai_epi32(v[14], 2);
+          u[15] = _mm256_srai_epi32(v[15], 2);
+
+          out[ 1] = _mm256_packs_epi32(u[0], u[1]);
+          out[17] = _mm256_packs_epi32(u[2], u[3]);
+          out[ 9] = _mm256_packs_epi32(u[4], u[5]);
+          out[25] = _mm256_packs_epi32(u[6], u[7]);
+          out[ 7] = _mm256_packs_epi32(u[8], u[9]);
+          out[23] = _mm256_packs_epi32(u[10], u[11]);
+          out[15] = _mm256_packs_epi32(u[12], u[13]);
+          out[31] = _mm256_packs_epi32(u[14], u[15]);
+        }
+        {
+          const __m256i k32_p27_p05 = pair256_set_epi32(cospi_27_64, cospi_5_64);
+          const __m256i k32_p11_p21 = pair256_set_epi32(cospi_11_64, cospi_21_64);
+          const __m256i k32_p19_p13 = pair256_set_epi32(cospi_19_64, cospi_13_64);
+          const __m256i k32_p03_p29 = pair256_set_epi32(cospi_3_64, cospi_29_64);
+          const __m256i k32_m29_p03 = pair256_set_epi32(-cospi_29_64, cospi_3_64);
+          const __m256i k32_m13_p19 = pair256_set_epi32(-cospi_13_64, cospi_19_64);
+          const __m256i k32_m21_p11 = pair256_set_epi32(-cospi_21_64, cospi_11_64);
+          const __m256i k32_m05_p27 = pair256_set_epi32(-cospi_5_64, cospi_27_64);
+
+          u[ 0] = _mm256_unpacklo_epi32(lstep1[40], lstep1[54]);
+          u[ 1] = _mm256_unpackhi_epi32(lstep1[40], lstep1[54]);
+          u[ 2] = _mm256_unpacklo_epi32(lstep1[41], lstep1[55]);
+          u[ 3] = _mm256_unpackhi_epi32(lstep1[41], lstep1[55]);
+          u[ 4] = _mm256_unpacklo_epi32(lstep1[42], lstep1[52]);
+          u[ 5] = _mm256_unpackhi_epi32(lstep1[42], lstep1[52]);
+          u[ 6] = _mm256_unpacklo_epi32(lstep1[43], lstep1[53]);
+          u[ 7] = _mm256_unpackhi_epi32(lstep1[43], lstep1[53]);
+          u[ 8] = _mm256_unpacklo_epi32(lstep1[44], lstep1[50]);
+          u[ 9] = _mm256_unpackhi_epi32(lstep1[44], lstep1[50]);
+          u[10] = _mm256_unpacklo_epi32(lstep1[45], lstep1[51]);
+          u[11] = _mm256_unpackhi_epi32(lstep1[45], lstep1[51]);
+          u[12] = _mm256_unpacklo_epi32(lstep1[46], lstep1[48]);
+          u[13] = _mm256_unpackhi_epi32(lstep1[46], lstep1[48]);
+          u[14] = _mm256_unpacklo_epi32(lstep1[47], lstep1[49]);
+          u[15] = _mm256_unpackhi_epi32(lstep1[47], lstep1[49]);
+
+          v[ 0] = k_madd_epi32_avx2(u[ 0], k32_p27_p05);
+          v[ 1] = k_madd_epi32_avx2(u[ 1], k32_p27_p05);
+          v[ 2] = k_madd_epi32_avx2(u[ 2], k32_p27_p05);
+          v[ 3] = k_madd_epi32_avx2(u[ 3], k32_p27_p05);
+          v[ 4] = k_madd_epi32_avx2(u[ 4], k32_p11_p21);
+          v[ 5] = k_madd_epi32_avx2(u[ 5], k32_p11_p21);
+          v[ 6] = k_madd_epi32_avx2(u[ 6], k32_p11_p21);
+          v[ 7] = k_madd_epi32_avx2(u[ 7], k32_p11_p21);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_p19_p13);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_p19_p13);
+          v[10] = k_madd_epi32_avx2(u[10], k32_p19_p13);
+          v[11] = k_madd_epi32_avx2(u[11], k32_p19_p13);
+          v[12] = k_madd_epi32_avx2(u[12], k32_p03_p29);
+          v[13] = k_madd_epi32_avx2(u[13], k32_p03_p29);
+          v[14] = k_madd_epi32_avx2(u[14], k32_p03_p29);
+          v[15] = k_madd_epi32_avx2(u[15], k32_p03_p29);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m29_p03);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m29_p03);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m29_p03);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m29_p03);
+          v[20] = k_madd_epi32_avx2(u[ 8], k32_m13_p19);
+          v[21] = k_madd_epi32_avx2(u[ 9], k32_m13_p19);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m13_p19);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m13_p19);
+          v[24] = k_madd_epi32_avx2(u[ 4], k32_m21_p11);
+          v[25] = k_madd_epi32_avx2(u[ 5], k32_m21_p11);
+          v[26] = k_madd_epi32_avx2(u[ 6], k32_m21_p11);
+          v[27] = k_madd_epi32_avx2(u[ 7], k32_m21_p11);
+          v[28] = k_madd_epi32_avx2(u[ 0], k32_m05_p27);
+          v[29] = k_madd_epi32_avx2(u[ 1], k32_m05_p27);
+          v[30] = k_madd_epi32_avx2(u[ 2], k32_m05_p27);
+          v[31] = k_madd_epi32_avx2(u[ 3], k32_m05_p27);
+
+          u[ 0] = k_packs_epi64_avx2(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64_avx2(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64_avx2(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64_avx2(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64_avx2(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[ 6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[ 7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[ 8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[ 9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm256_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm256_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm256_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm256_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm256_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm256_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm256_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm256_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm256_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm256_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm256_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm256_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm256_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm256_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm256_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm256_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm256_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm256_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm256_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm256_cmpgt_epi32(kZero,u[ 0]);
+          v[ 1] = _mm256_cmpgt_epi32(kZero,u[ 1]);
+          v[ 2] = _mm256_cmpgt_epi32(kZero,u[ 2]);
+          v[ 3] = _mm256_cmpgt_epi32(kZero,u[ 3]);
+          v[ 4] = _mm256_cmpgt_epi32(kZero,u[ 4]);
+          v[ 5] = _mm256_cmpgt_epi32(kZero,u[ 5]);
+          v[ 6] = _mm256_cmpgt_epi32(kZero,u[ 6]);
+          v[ 7] = _mm256_cmpgt_epi32(kZero,u[ 7]);
+          v[ 8] = _mm256_cmpgt_epi32(kZero,u[ 8]);
+          v[ 9] = _mm256_cmpgt_epi32(kZero,u[ 9]);
+          v[10] = _mm256_cmpgt_epi32(kZero,u[10]);
+          v[11] = _mm256_cmpgt_epi32(kZero,u[11]);
+          v[12] = _mm256_cmpgt_epi32(kZero,u[12]);
+          v[13] = _mm256_cmpgt_epi32(kZero,u[13]);
+          v[14] = _mm256_cmpgt_epi32(kZero,u[14]);
+          v[15] = _mm256_cmpgt_epi32(kZero,u[15]);
+
+          u[ 0] = _mm256_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm256_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm256_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm256_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm256_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm256_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm256_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm256_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm256_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm256_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm256_sub_epi32(u[10], v[10]);
+          u[11] = _mm256_sub_epi32(u[11], v[11]);
+          u[12] = _mm256_sub_epi32(u[12], v[12]);
+          u[13] = _mm256_sub_epi32(u[13], v[13]);
+          u[14] = _mm256_sub_epi32(u[14], v[14]);
+          u[15] = _mm256_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], K32One);
+          v[1] = _mm256_add_epi32(u[1], K32One);
+          v[2] = _mm256_add_epi32(u[2], K32One);
+          v[3] = _mm256_add_epi32(u[3], K32One);
+          v[4] = _mm256_add_epi32(u[4], K32One);
+          v[5] = _mm256_add_epi32(u[5], K32One);
+          v[6] = _mm256_add_epi32(u[6], K32One);
+          v[7] = _mm256_add_epi32(u[7], K32One);
+          v[8] = _mm256_add_epi32(u[8], K32One);
+          v[9] = _mm256_add_epi32(u[9], K32One);
+          v[10] = _mm256_add_epi32(u[10], K32One);
+          v[11] = _mm256_add_epi32(u[11], K32One);
+          v[12] = _mm256_add_epi32(u[12], K32One);
+          v[13] = _mm256_add_epi32(u[13], K32One);
+          v[14] = _mm256_add_epi32(u[14], K32One);
+          v[15] = _mm256_add_epi32(u[15], K32One);
+
+          u[0] = _mm256_srai_epi32(v[0], 2);
+          u[1] = _mm256_srai_epi32(v[1], 2);
+          u[2] = _mm256_srai_epi32(v[2], 2);
+          u[3] = _mm256_srai_epi32(v[3], 2);
+          u[4] = _mm256_srai_epi32(v[4], 2);
+          u[5] = _mm256_srai_epi32(v[5], 2);
+          u[6] = _mm256_srai_epi32(v[6], 2);
+          u[7] = _mm256_srai_epi32(v[7], 2);
+          u[8] = _mm256_srai_epi32(v[8], 2);
+          u[9] = _mm256_srai_epi32(v[9], 2);
+          u[10] = _mm256_srai_epi32(v[10], 2);
+          u[11] = _mm256_srai_epi32(v[11], 2);
+          u[12] = _mm256_srai_epi32(v[12], 2);
+          u[13] = _mm256_srai_epi32(v[13], 2);
+          u[14] = _mm256_srai_epi32(v[14], 2);
+          u[15] = _mm256_srai_epi32(v[15], 2);
+
+          out[ 5] = _mm256_packs_epi32(u[0], u[1]);
+          out[21] = _mm256_packs_epi32(u[2], u[3]);
+          out[13] = _mm256_packs_epi32(u[4], u[5]);
+          out[29] = _mm256_packs_epi32(u[6], u[7]);
+          out[ 3] = _mm256_packs_epi32(u[8], u[9]);
+          out[19] = _mm256_packs_epi32(u[10], u[11]);
+          out[11] = _mm256_packs_epi32(u[12], u[13]);
+          out[27] = _mm256_packs_epi32(u[14], u[15]);
+        }
+      }
+#endif
+      // Transpose the results, do it as four 8x8 transposes.
+      {
+        int transpose_block;
+        int16_t *output_currStep,*output_nextStep;
+        if (0 == pass){
+                 output_currStep = &intermediate[column_start * 32];
+                 output_nextStep = &intermediate[(column_start + 8) * 32];
+        } else{
+                 output_currStep = &output_org[column_start * 32];
+                 output_nextStep = &output_org[(column_start + 8) * 32];
+        }
+        for (transpose_block = 0; transpose_block < 4; ++transpose_block) {
+          __m256i *this_out = &out[8 * transpose_block];
+          // 00  01  02  03  04  05  06  07  08  09  10  11  12  13  14  15
+          // 20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35
+          // 40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55
+          // 60  61  62  63  64  65  66  67  68  69  70  71  72  73  74  75
+          // 80  81  82  83  84  85  86  87  88  89  90  91  92  93  94  95
+          // 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115
+          // 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
+          // 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155
+          const __m256i tr0_0 = _mm256_unpacklo_epi16(this_out[0], this_out[1]);
+          const __m256i tr0_1 = _mm256_unpacklo_epi16(this_out[2], this_out[3]);
+          const __m256i tr0_2 = _mm256_unpackhi_epi16(this_out[0], this_out[1]);
+          const __m256i tr0_3 = _mm256_unpackhi_epi16(this_out[2], this_out[3]);
+          const __m256i tr0_4 = _mm256_unpacklo_epi16(this_out[4], this_out[5]);
+          const __m256i tr0_5 = _mm256_unpacklo_epi16(this_out[6], this_out[7]);
+          const __m256i tr0_6 = _mm256_unpackhi_epi16(this_out[4], this_out[5]);
+          const __m256i tr0_7 = _mm256_unpackhi_epi16(this_out[6], this_out[7]);
+          // 00  20  01  21  02  22  03  23  08  28  09  29  10  30  11  31
+          // 40  60  41  61  42  62  43  63  48  68  49  69  50  70  51  71
+          // 04  24  05  25  06  26  07  27  12  32  13  33  14  34  15  35
+          // 44  64  45  65  46  66  47  67  52  72  53  73  54  74  55  75
+          // 80  100 81  101 82  102 83  103 88  108 89  109 90  110 91  101
+          // 120 140 121 141 122 142 123 143 128 148 129 149 130 150 131 151
+          // 84  104 85  105 86  106 87  107 92  112 93  113 94  114 95  115
+          // 124 144 125 145 126 146 127 147 132 152 133 153 134 154 135 155
+
+          const __m256i tr1_0 = _mm256_unpacklo_epi32(tr0_0, tr0_1);
+          const __m256i tr1_1 = _mm256_unpacklo_epi32(tr0_2, tr0_3);
+          const __m256i tr1_2 = _mm256_unpackhi_epi32(tr0_0, tr0_1);
+          const __m256i tr1_3 = _mm256_unpackhi_epi32(tr0_2, tr0_3);
+          const __m256i tr1_4 = _mm256_unpacklo_epi32(tr0_4, tr0_5);
+          const __m256i tr1_5 = _mm256_unpacklo_epi32(tr0_6, tr0_7);
+          const __m256i tr1_6 = _mm256_unpackhi_epi32(tr0_4, tr0_5);
+          const __m256i tr1_7 = _mm256_unpackhi_epi32(tr0_6, tr0_7);
+          // 00 20  40  60  01 21  41  61  08 28  48  68  09 29  49  69
+          // 04 24  44  64  05 25  45  65  12 32  52  72  13 33  53  73
+          // 02 22  42  62  03 23  43  63  10 30  50  70  11 31  51  71
+          // 06 26  46  66  07 27  47  67  14 34  54  74  15 35  55  75
+          // 80 100 120 140 81 101 121 141 88 108 128 148 89 109 129 149
+          // 84 104 124 144 85 105 125 145 92 112 132 152 93 113 133 153
+          // 82 102 122 142 83 103 123 143 90 110 130 150 91 101 131 151
+          // 86 106 126 146 87 107 127 147 94 114 134 154 95 115 135 155
+          __m256i tr2_0 = _mm256_unpacklo_epi64(tr1_0, tr1_4);
+          __m256i tr2_1 = _mm256_unpackhi_epi64(tr1_0, tr1_4);
+          __m256i tr2_2 = _mm256_unpacklo_epi64(tr1_2, tr1_6);
+          __m256i tr2_3 = _mm256_unpackhi_epi64(tr1_2, tr1_6);
+          __m256i tr2_4 = _mm256_unpacklo_epi64(tr1_1, tr1_5);
+          __m256i tr2_5 = _mm256_unpackhi_epi64(tr1_1, tr1_5);
+          __m256i tr2_6 = _mm256_unpacklo_epi64(tr1_3, tr1_7);
+          __m256i tr2_7 = _mm256_unpackhi_epi64(tr1_3, tr1_7);
+          // 00 20 40 60 80 100 120 140 08 28 48 68 88 108 128 148
+          // 01 21 41 61 81 101 121 141 09 29 49 69 89 109 129 149
+          // 02 22 42 62 82 102 122 142 10 30 50 70 90 110 130 150
+          // 03 23 43 63 83 103 123 143 11 31 51 71 91 101 131 151
+          // 04 24 44 64 84 104 124 144 12 32 52 72 92 112 132 152
+          // 05 25 45 65 85 105 125 145 13 33 53 73 93 113 133 153
+          // 06 26 46 66 86 106 126 146 14 34 54 74 94 114 134 154
+          // 07 27 47 67 87 107 127 147 15 35 55 75 95 115 135 155
+          if (0 == pass) {
+            // output[j] = (output[j] + 1 + (output[j] > 0)) >> 2;
+            // TODO(cd): see quality impact of only doing
+            //           output[j] = (output[j] + 1) >> 2;
+            //           which would remove the code between here ...
+            __m256i tr2_0_0 = _mm256_cmpgt_epi16(tr2_0, kZero);
+            __m256i tr2_1_0 = _mm256_cmpgt_epi16(tr2_1, kZero);
+            __m256i tr2_2_0 = _mm256_cmpgt_epi16(tr2_2, kZero);
+            __m256i tr2_3_0 = _mm256_cmpgt_epi16(tr2_3, kZero);
+            __m256i tr2_4_0 = _mm256_cmpgt_epi16(tr2_4, kZero);
+            __m256i tr2_5_0 = _mm256_cmpgt_epi16(tr2_5, kZero);
+            __m256i tr2_6_0 = _mm256_cmpgt_epi16(tr2_6, kZero);
+            __m256i tr2_7_0 = _mm256_cmpgt_epi16(tr2_7, kZero);
+            tr2_0 = _mm256_sub_epi16(tr2_0, tr2_0_0);
+            tr2_1 = _mm256_sub_epi16(tr2_1, tr2_1_0);
+            tr2_2 = _mm256_sub_epi16(tr2_2, tr2_2_0);
+            tr2_3 = _mm256_sub_epi16(tr2_3, tr2_3_0);
+            tr2_4 = _mm256_sub_epi16(tr2_4, tr2_4_0);
+            tr2_5 = _mm256_sub_epi16(tr2_5, tr2_5_0);
+            tr2_6 = _mm256_sub_epi16(tr2_6, tr2_6_0);
+            tr2_7 = _mm256_sub_epi16(tr2_7, tr2_7_0);
+            //           ... and here.
+            //           PS: also change code in vp9/encoder/vp9_dct.c
+            tr2_0 = _mm256_add_epi16(tr2_0, kOne);
+            tr2_1 = _mm256_add_epi16(tr2_1, kOne);
+            tr2_2 = _mm256_add_epi16(tr2_2, kOne);
+            tr2_3 = _mm256_add_epi16(tr2_3, kOne);
+            tr2_4 = _mm256_add_epi16(tr2_4, kOne);
+            tr2_5 = _mm256_add_epi16(tr2_5, kOne);
+            tr2_6 = _mm256_add_epi16(tr2_6, kOne);
+            tr2_7 = _mm256_add_epi16(tr2_7, kOne);
+            tr2_0 = _mm256_srai_epi16(tr2_0, 2);
+            tr2_1 = _mm256_srai_epi16(tr2_1, 2);
+            tr2_2 = _mm256_srai_epi16(tr2_2, 2);
+            tr2_3 = _mm256_srai_epi16(tr2_3, 2);
+            tr2_4 = _mm256_srai_epi16(tr2_4, 2);
+            tr2_5 = _mm256_srai_epi16(tr2_5, 2);
+            tr2_6 = _mm256_srai_epi16(tr2_6, 2);
+            tr2_7 = _mm256_srai_epi16(tr2_7, 2);
+          }
+          // Note: even though all these stores are aligned, using the aligned
+          //       intrinsic make the code slightly slower.
+          _mm_storeu_si128((__m128i *)(output_currStep + 0 * 32), _mm256_castsi256_si128(tr2_0));
+          _mm_storeu_si128((__m128i *)(output_currStep + 1 * 32), _mm256_castsi256_si128(tr2_1));
+          _mm_storeu_si128((__m128i *)(output_currStep + 2 * 32), _mm256_castsi256_si128(tr2_2));
+          _mm_storeu_si128((__m128i *)(output_currStep + 3 * 32), _mm256_castsi256_si128(tr2_3));
+          _mm_storeu_si128((__m128i *)(output_currStep + 4 * 32), _mm256_castsi256_si128(tr2_4));
+          _mm_storeu_si128((__m128i *)(output_currStep + 5 * 32), _mm256_castsi256_si128(tr2_5));
+          _mm_storeu_si128((__m128i *)(output_currStep + 6 * 32), _mm256_castsi256_si128(tr2_6));
+          _mm_storeu_si128((__m128i *)(output_currStep + 7 * 32), _mm256_castsi256_si128(tr2_7));
+
+          _mm_storeu_si128((__m128i *)(output_nextStep + 0 * 32), _mm256_extractf128_si256(tr2_0,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 1 * 32), _mm256_extractf128_si256(tr2_1,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 2 * 32), _mm256_extractf128_si256(tr2_2,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 3 * 32), _mm256_extractf128_si256(tr2_3,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 4 * 32), _mm256_extractf128_si256(tr2_4,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 5 * 32), _mm256_extractf128_si256(tr2_5,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 6 * 32), _mm256_extractf128_si256(tr2_6,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 7 * 32), _mm256_extractf128_si256(tr2_7,1));
+          // Process next 8x8
+          output_currStep += 8;
+          output_nextStep += 8;
+        }
+      }
+    }
+  }
+}  // NOLINT
diff --git a/libvpx/libvpx/vpx_dsp/x86/fwd_dct32x32_impl_sse2.h b/libvpx/libvpx/vpx_dsp/x86/fwd_dct32x32_impl_sse2.h
new file mode 100644
index 0000000..b85ae10
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/fwd_dct32x32_impl_sse2.h
@@ -0,0 +1,3153 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "vpx_dsp/fwd_txfm.h"
+#include "vpx_dsp/txfm_common.h"
+#include "vpx_dsp/x86/txfm_common_sse2.h"
+
+// TODO(jingning) The high bit-depth version needs re-work for performance.
+// The current SSE2 implementation also causes cross reference to the static
+// functions in the C implementation file.
+#if DCT_HIGH_BIT_DEPTH
+#define ADD_EPI16 _mm_adds_epi16
+#define SUB_EPI16 _mm_subs_epi16
+#if FDCT32x32_HIGH_PRECISION
+void vpx_fdct32x32_rows_c(const int16_t *intermediate, tran_low_t *out) {
+    int i, j;
+    for (i = 0; i < 32; ++i) {
+      tran_high_t temp_in[32], temp_out[32];
+      for (j = 0; j < 32; ++j)
+        temp_in[j] = intermediate[j * 32 + i];
+      vpx_fdct32(temp_in, temp_out, 0);
+      for (j = 0; j < 32; ++j)
+        out[j + i * 32] =
+            (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2);
+    }
+}
+  #define HIGH_FDCT32x32_2D_C vpx_highbd_fdct32x32_c
+  #define HIGH_FDCT32x32_2D_ROWS_C vpx_fdct32x32_rows_c
+#else
+void vpx_fdct32x32_rd_rows_c(const int16_t *intermediate, tran_low_t *out) {
+    int i, j;
+    for (i = 0; i < 32; ++i) {
+      tran_high_t temp_in[32], temp_out[32];
+      for (j = 0; j < 32; ++j)
+        temp_in[j] = intermediate[j * 32 + i];
+      vpx_fdct32(temp_in, temp_out, 1);
+      for (j = 0; j < 32; ++j)
+        out[j + i * 32] = (tran_low_t)temp_out[j];
+    }
+}
+  #define HIGH_FDCT32x32_2D_C vpx_highbd_fdct32x32_rd_c
+  #define HIGH_FDCT32x32_2D_ROWS_C vpx_fdct32x32_rd_rows_c
+#endif  // FDCT32x32_HIGH_PRECISION
+#else
+#define ADD_EPI16 _mm_add_epi16
+#define SUB_EPI16 _mm_sub_epi16
+#endif  // DCT_HIGH_BIT_DEPTH
+
+
+void FDCT32x32_2D(const int16_t *input,
+                  tran_low_t *output_org, int stride) {
+  // Calculate pre-multiplied strides
+  const int str1 = stride;
+  const int str2 = 2 * stride;
+  const int str3 = 2 * stride + str1;
+  // We need an intermediate buffer between passes.
+  DECLARE_ALIGNED(16, int16_t, intermediate[32 * 32]);
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(+cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64,   cospi_24_64);
+  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(+cospi_24_64,  cospi_8_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(+cospi_12_64,  cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64,  cospi_12_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64,   cospi_28_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(+cospi_28_64,  cospi_4_64);
+  const __m128i k__cospi_m28_m04 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_m12_m20 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_p30_p02 = pair_set_epi16(+cospi_30_64,  cospi_2_64);
+  const __m128i k__cospi_p14_p18 = pair_set_epi16(+cospi_14_64,  cospi_18_64);
+  const __m128i k__cospi_p22_p10 = pair_set_epi16(+cospi_22_64,  cospi_10_64);
+  const __m128i k__cospi_p06_p26 = pair_set_epi16(+cospi_6_64,   cospi_26_64);
+  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64,  cospi_6_64);
+  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64,  cospi_22_64);
+  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64,  cospi_14_64);
+  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64,   cospi_30_64);
+  const __m128i k__cospi_p31_p01 = pair_set_epi16(+cospi_31_64,  cospi_1_64);
+  const __m128i k__cospi_p15_p17 = pair_set_epi16(+cospi_15_64,  cospi_17_64);
+  const __m128i k__cospi_p23_p09 = pair_set_epi16(+cospi_23_64,  cospi_9_64);
+  const __m128i k__cospi_p07_p25 = pair_set_epi16(+cospi_7_64,   cospi_25_64);
+  const __m128i k__cospi_m25_p07 = pair_set_epi16(-cospi_25_64,  cospi_7_64);
+  const __m128i k__cospi_m09_p23 = pair_set_epi16(-cospi_9_64,   cospi_23_64);
+  const __m128i k__cospi_m17_p15 = pair_set_epi16(-cospi_17_64,  cospi_15_64);
+  const __m128i k__cospi_m01_p31 = pair_set_epi16(-cospi_1_64,   cospi_31_64);
+  const __m128i k__cospi_p27_p05 = pair_set_epi16(+cospi_27_64,  cospi_5_64);
+  const __m128i k__cospi_p11_p21 = pair_set_epi16(+cospi_11_64,  cospi_21_64);
+  const __m128i k__cospi_p19_p13 = pair_set_epi16(+cospi_19_64,  cospi_13_64);
+  const __m128i k__cospi_p03_p29 = pair_set_epi16(+cospi_3_64,   cospi_29_64);
+  const __m128i k__cospi_m29_p03 = pair_set_epi16(-cospi_29_64,  cospi_3_64);
+  const __m128i k__cospi_m13_p19 = pair_set_epi16(-cospi_13_64,  cospi_19_64);
+  const __m128i k__cospi_m21_p11 = pair_set_epi16(-cospi_21_64,  cospi_11_64);
+  const __m128i k__cospi_m05_p27 = pair_set_epi16(-cospi_5_64,   cospi_27_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i kOne  = _mm_set1_epi16(1);
+  // Do the two transform/transpose passes
+  int pass;
+#if DCT_HIGH_BIT_DEPTH
+  int overflow;
+#endif
+  for (pass = 0; pass < 2; ++pass) {
+    // We process eight columns (transposed rows in second pass) at a time.
+    int column_start;
+    for (column_start = 0; column_start < 32; column_start += 8) {
+      __m128i step1[32];
+      __m128i step2[32];
+      __m128i step3[32];
+      __m128i out[32];
+      // Stage 1
+      // Note: even though all the loads below are aligned, using the aligned
+      //       intrinsic make the code slightly slower.
+      if (0 == pass) {
+        const int16_t *in  = &input[column_start];
+        // step1[i] =  (in[ 0 * stride] + in[(32 -  1) * stride]) << 2;
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          const int16_t *ina =  in +  0 * str1;
+          const int16_t *inb =  in + 31 * str1;
+          __m128i *step1a = &step1[ 0];
+          __m128i *step1b = &step1[31];
+          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[ 0] = _mm_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in +  4 * str1;
+          const int16_t *inb =  in + 27 * str1;
+          __m128i *step1a = &step1[ 4];
+          __m128i *step1b = &step1[27];
+          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[ 0] = _mm_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in +  8 * str1;
+          const int16_t *inb =  in + 23 * str1;
+          __m128i *step1a = &step1[ 8];
+          __m128i *step1b = &step1[23];
+          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[ 0] = _mm_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in + 12 * str1;
+          const int16_t *inb =  in + 19 * str1;
+          __m128i *step1a = &step1[12];
+          __m128i *step1b = &step1[19];
+          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[ 0] = _mm_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+      } else {
+        int16_t *in = &intermediate[column_start];
+        // step1[i] =  in[ 0 * 32] + in[(32 -  1) * 32];
+        // Note: using the same approach as above to have common offset is
+        //       counter-productive as all offsets can be calculated at compile
+        //       time.
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          __m128i in00  = _mm_loadu_si128((const __m128i *)(in +  0 * 32));
+          __m128i in01  = _mm_loadu_si128((const __m128i *)(in +  1 * 32));
+          __m128i in02  = _mm_loadu_si128((const __m128i *)(in +  2 * 32));
+          __m128i in03  = _mm_loadu_si128((const __m128i *)(in +  3 * 32));
+          __m128i in28  = _mm_loadu_si128((const __m128i *)(in + 28 * 32));
+          __m128i in29  = _mm_loadu_si128((const __m128i *)(in + 29 * 32));
+          __m128i in30  = _mm_loadu_si128((const __m128i *)(in + 30 * 32));
+          __m128i in31  = _mm_loadu_si128((const __m128i *)(in + 31 * 32));
+          step1[0] = ADD_EPI16(in00, in31);
+          step1[1] = ADD_EPI16(in01, in30);
+          step1[2] = ADD_EPI16(in02, in29);
+          step1[3] = ADD_EPI16(in03, in28);
+          step1[28] = SUB_EPI16(in03, in28);
+          step1[29] = SUB_EPI16(in02, in29);
+          step1[30] = SUB_EPI16(in01, in30);
+          step1[31] = SUB_EPI16(in00, in31);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[0], &step1[1], &step1[2],
+                                             &step1[3], &step1[28], &step1[29],
+                                             &step1[30], &step1[31]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          __m128i in04  = _mm_loadu_si128((const __m128i *)(in +  4 * 32));
+          __m128i in05  = _mm_loadu_si128((const __m128i *)(in +  5 * 32));
+          __m128i in06  = _mm_loadu_si128((const __m128i *)(in +  6 * 32));
+          __m128i in07  = _mm_loadu_si128((const __m128i *)(in +  7 * 32));
+          __m128i in24  = _mm_loadu_si128((const __m128i *)(in + 24 * 32));
+          __m128i in25  = _mm_loadu_si128((const __m128i *)(in + 25 * 32));
+          __m128i in26  = _mm_loadu_si128((const __m128i *)(in + 26 * 32));
+          __m128i in27  = _mm_loadu_si128((const __m128i *)(in + 27 * 32));
+          step1[4] = ADD_EPI16(in04, in27);
+          step1[5] = ADD_EPI16(in05, in26);
+          step1[6] = ADD_EPI16(in06, in25);
+          step1[7] = ADD_EPI16(in07, in24);
+          step1[24] = SUB_EPI16(in07, in24);
+          step1[25] = SUB_EPI16(in06, in25);
+          step1[26] = SUB_EPI16(in05, in26);
+          step1[27] = SUB_EPI16(in04, in27);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[4], &step1[5], &step1[6],
+                                             &step1[7], &step1[24], &step1[25],
+                                             &step1[26], &step1[27]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          __m128i in08  = _mm_loadu_si128((const __m128i *)(in +  8 * 32));
+          __m128i in09  = _mm_loadu_si128((const __m128i *)(in +  9 * 32));
+          __m128i in10  = _mm_loadu_si128((const __m128i *)(in + 10 * 32));
+          __m128i in11  = _mm_loadu_si128((const __m128i *)(in + 11 * 32));
+          __m128i in20  = _mm_loadu_si128((const __m128i *)(in + 20 * 32));
+          __m128i in21  = _mm_loadu_si128((const __m128i *)(in + 21 * 32));
+          __m128i in22  = _mm_loadu_si128((const __m128i *)(in + 22 * 32));
+          __m128i in23  = _mm_loadu_si128((const __m128i *)(in + 23 * 32));
+          step1[8] = ADD_EPI16(in08, in23);
+          step1[9] = ADD_EPI16(in09, in22);
+          step1[10] = ADD_EPI16(in10, in21);
+          step1[11] = ADD_EPI16(in11, in20);
+          step1[20] = SUB_EPI16(in11, in20);
+          step1[21] = SUB_EPI16(in10, in21);
+          step1[22] = SUB_EPI16(in09, in22);
+          step1[23] = SUB_EPI16(in08, in23);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[8], &step1[9], &step1[10],
+                                             &step1[11], &step1[20], &step1[21],
+                                             &step1[22], &step1[23]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          __m128i in12  = _mm_loadu_si128((const __m128i *)(in + 12 * 32));
+          __m128i in13  = _mm_loadu_si128((const __m128i *)(in + 13 * 32));
+          __m128i in14  = _mm_loadu_si128((const __m128i *)(in + 14 * 32));
+          __m128i in15  = _mm_loadu_si128((const __m128i *)(in + 15 * 32));
+          __m128i in16  = _mm_loadu_si128((const __m128i *)(in + 16 * 32));
+          __m128i in17  = _mm_loadu_si128((const __m128i *)(in + 17 * 32));
+          __m128i in18  = _mm_loadu_si128((const __m128i *)(in + 18 * 32));
+          __m128i in19  = _mm_loadu_si128((const __m128i *)(in + 19 * 32));
+          step1[12] = ADD_EPI16(in12, in19);
+          step1[13] = ADD_EPI16(in13, in18);
+          step1[14] = ADD_EPI16(in14, in17);
+          step1[15] = ADD_EPI16(in15, in16);
+          step1[16] = SUB_EPI16(in15, in16);
+          step1[17] = SUB_EPI16(in14, in17);
+          step1[18] = SUB_EPI16(in13, in18);
+          step1[19] = SUB_EPI16(in12, in19);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[12], &step1[13], &step1[14],
+                                             &step1[15], &step1[16], &step1[17],
+                                             &step1[18], &step1[19]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+      // Stage 2
+      {
+        step2[0] = ADD_EPI16(step1[0], step1[15]);
+        step2[1] = ADD_EPI16(step1[1], step1[14]);
+        step2[2] = ADD_EPI16(step1[2], step1[13]);
+        step2[3] = ADD_EPI16(step1[3], step1[12]);
+        step2[4] = ADD_EPI16(step1[4], step1[11]);
+        step2[5] = ADD_EPI16(step1[5], step1[10]);
+        step2[6] = ADD_EPI16(step1[6], step1[ 9]);
+        step2[7] = ADD_EPI16(step1[7], step1[ 8]);
+        step2[8] = SUB_EPI16(step1[7], step1[ 8]);
+        step2[9] = SUB_EPI16(step1[6], step1[ 9]);
+        step2[10] = SUB_EPI16(step1[5], step1[10]);
+        step2[11] = SUB_EPI16(step1[4], step1[11]);
+        step2[12] = SUB_EPI16(step1[3], step1[12]);
+        step2[13] = SUB_EPI16(step1[2], step1[13]);
+        step2[14] = SUB_EPI16(step1[1], step1[14]);
+        step2[15] = SUB_EPI16(step1[0], step1[15]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step2[0], &step2[1], &step2[2], &step2[3],
+            &step2[4], &step2[5], &step2[6], &step2[7],
+            &step2[8], &step2[9], &step2[10], &step2[11],
+            &step2[12], &step2[13], &step2[14], &step2[15]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]);
+        const __m128i s2_20_1 = _mm_unpackhi_epi16(step1[27], step1[20]);
+        const __m128i s2_21_0 = _mm_unpacklo_epi16(step1[26], step1[21]);
+        const __m128i s2_21_1 = _mm_unpackhi_epi16(step1[26], step1[21]);
+        const __m128i s2_22_0 = _mm_unpacklo_epi16(step1[25], step1[22]);
+        const __m128i s2_22_1 = _mm_unpackhi_epi16(step1[25], step1[22]);
+        const __m128i s2_23_0 = _mm_unpacklo_epi16(step1[24], step1[23]);
+        const __m128i s2_23_1 = _mm_unpackhi_epi16(step1[24], step1[23]);
+        const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
+        const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
+        const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
+        const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
+        const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
+        const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
+        const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
+        const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
+        const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
+        const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16);
+        const __m128i s2_25_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_p16);
+        const __m128i s2_25_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_p16);
+        const __m128i s2_26_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_p16);
+        const __m128i s2_26_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_p16);
+        const __m128i s2_27_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_p16);
+        const __m128i s2_27_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m128i s2_20_4 = _mm_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_20_5 = _mm_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_21_4 = _mm_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_21_5 = _mm_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_22_4 = _mm_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_22_5 = _mm_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_23_4 = _mm_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_23_5 = _mm_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_24_4 = _mm_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_24_5 = _mm_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_25_4 = _mm_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_25_5 = _mm_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_26_4 = _mm_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_26_5 = _mm_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_27_4 = _mm_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_27_5 = _mm_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_20_6 = _mm_srai_epi32(s2_20_4, DCT_CONST_BITS);
+        const __m128i s2_20_7 = _mm_srai_epi32(s2_20_5, DCT_CONST_BITS);
+        const __m128i s2_21_6 = _mm_srai_epi32(s2_21_4, DCT_CONST_BITS);
+        const __m128i s2_21_7 = _mm_srai_epi32(s2_21_5, DCT_CONST_BITS);
+        const __m128i s2_22_6 = _mm_srai_epi32(s2_22_4, DCT_CONST_BITS);
+        const __m128i s2_22_7 = _mm_srai_epi32(s2_22_5, DCT_CONST_BITS);
+        const __m128i s2_23_6 = _mm_srai_epi32(s2_23_4, DCT_CONST_BITS);
+        const __m128i s2_23_7 = _mm_srai_epi32(s2_23_5, DCT_CONST_BITS);
+        const __m128i s2_24_6 = _mm_srai_epi32(s2_24_4, DCT_CONST_BITS);
+        const __m128i s2_24_7 = _mm_srai_epi32(s2_24_5, DCT_CONST_BITS);
+        const __m128i s2_25_6 = _mm_srai_epi32(s2_25_4, DCT_CONST_BITS);
+        const __m128i s2_25_7 = _mm_srai_epi32(s2_25_5, DCT_CONST_BITS);
+        const __m128i s2_26_6 = _mm_srai_epi32(s2_26_4, DCT_CONST_BITS);
+        const __m128i s2_26_7 = _mm_srai_epi32(s2_26_5, DCT_CONST_BITS);
+        const __m128i s2_27_6 = _mm_srai_epi32(s2_27_4, DCT_CONST_BITS);
+        const __m128i s2_27_7 = _mm_srai_epi32(s2_27_5, DCT_CONST_BITS);
+        // Combine
+        step2[20] = _mm_packs_epi32(s2_20_6, s2_20_7);
+        step2[21] = _mm_packs_epi32(s2_21_6, s2_21_7);
+        step2[22] = _mm_packs_epi32(s2_22_6, s2_22_7);
+        step2[23] = _mm_packs_epi32(s2_23_6, s2_23_7);
+        step2[24] = _mm_packs_epi32(s2_24_6, s2_24_7);
+        step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7);
+        step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7);
+        step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step2[20], &step2[21], &step2[22],
+                                           &step2[23], &step2[24], &step2[25],
+                                           &step2[26], &step2[27]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+
+#if !FDCT32x32_HIGH_PRECISION
+      // dump the magnitude by half, hence the intermediate values are within
+      // the range of 16 bits.
+      if (1 == pass) {
+        __m128i s3_00_0 = _mm_cmplt_epi16(step2[ 0], kZero);
+        __m128i s3_01_0 = _mm_cmplt_epi16(step2[ 1], kZero);
+        __m128i s3_02_0 = _mm_cmplt_epi16(step2[ 2], kZero);
+        __m128i s3_03_0 = _mm_cmplt_epi16(step2[ 3], kZero);
+        __m128i s3_04_0 = _mm_cmplt_epi16(step2[ 4], kZero);
+        __m128i s3_05_0 = _mm_cmplt_epi16(step2[ 5], kZero);
+        __m128i s3_06_0 = _mm_cmplt_epi16(step2[ 6], kZero);
+        __m128i s3_07_0 = _mm_cmplt_epi16(step2[ 7], kZero);
+        __m128i s2_08_0 = _mm_cmplt_epi16(step2[ 8], kZero);
+        __m128i s2_09_0 = _mm_cmplt_epi16(step2[ 9], kZero);
+        __m128i s3_10_0 = _mm_cmplt_epi16(step2[10], kZero);
+        __m128i s3_11_0 = _mm_cmplt_epi16(step2[11], kZero);
+        __m128i s3_12_0 = _mm_cmplt_epi16(step2[12], kZero);
+        __m128i s3_13_0 = _mm_cmplt_epi16(step2[13], kZero);
+        __m128i s2_14_0 = _mm_cmplt_epi16(step2[14], kZero);
+        __m128i s2_15_0 = _mm_cmplt_epi16(step2[15], kZero);
+        __m128i s3_16_0 = _mm_cmplt_epi16(step1[16], kZero);
+        __m128i s3_17_0 = _mm_cmplt_epi16(step1[17], kZero);
+        __m128i s3_18_0 = _mm_cmplt_epi16(step1[18], kZero);
+        __m128i s3_19_0 = _mm_cmplt_epi16(step1[19], kZero);
+        __m128i s3_20_0 = _mm_cmplt_epi16(step2[20], kZero);
+        __m128i s3_21_0 = _mm_cmplt_epi16(step2[21], kZero);
+        __m128i s3_22_0 = _mm_cmplt_epi16(step2[22], kZero);
+        __m128i s3_23_0 = _mm_cmplt_epi16(step2[23], kZero);
+        __m128i s3_24_0 = _mm_cmplt_epi16(step2[24], kZero);
+        __m128i s3_25_0 = _mm_cmplt_epi16(step2[25], kZero);
+        __m128i s3_26_0 = _mm_cmplt_epi16(step2[26], kZero);
+        __m128i s3_27_0 = _mm_cmplt_epi16(step2[27], kZero);
+        __m128i s3_28_0 = _mm_cmplt_epi16(step1[28], kZero);
+        __m128i s3_29_0 = _mm_cmplt_epi16(step1[29], kZero);
+        __m128i s3_30_0 = _mm_cmplt_epi16(step1[30], kZero);
+        __m128i s3_31_0 = _mm_cmplt_epi16(step1[31], kZero);
+
+        step2[0] = SUB_EPI16(step2[ 0], s3_00_0);
+        step2[1] = SUB_EPI16(step2[ 1], s3_01_0);
+        step2[2] = SUB_EPI16(step2[ 2], s3_02_0);
+        step2[3] = SUB_EPI16(step2[ 3], s3_03_0);
+        step2[4] = SUB_EPI16(step2[ 4], s3_04_0);
+        step2[5] = SUB_EPI16(step2[ 5], s3_05_0);
+        step2[6] = SUB_EPI16(step2[ 6], s3_06_0);
+        step2[7] = SUB_EPI16(step2[ 7], s3_07_0);
+        step2[8] = SUB_EPI16(step2[ 8], s2_08_0);
+        step2[9] = SUB_EPI16(step2[ 9], s2_09_0);
+        step2[10] = SUB_EPI16(step2[10], s3_10_0);
+        step2[11] = SUB_EPI16(step2[11], s3_11_0);
+        step2[12] = SUB_EPI16(step2[12], s3_12_0);
+        step2[13] = SUB_EPI16(step2[13], s3_13_0);
+        step2[14] = SUB_EPI16(step2[14], s2_14_0);
+        step2[15] = SUB_EPI16(step2[15], s2_15_0);
+        step1[16] = SUB_EPI16(step1[16], s3_16_0);
+        step1[17] = SUB_EPI16(step1[17], s3_17_0);
+        step1[18] = SUB_EPI16(step1[18], s3_18_0);
+        step1[19] = SUB_EPI16(step1[19], s3_19_0);
+        step2[20] = SUB_EPI16(step2[20], s3_20_0);
+        step2[21] = SUB_EPI16(step2[21], s3_21_0);
+        step2[22] = SUB_EPI16(step2[22], s3_22_0);
+        step2[23] = SUB_EPI16(step2[23], s3_23_0);
+        step2[24] = SUB_EPI16(step2[24], s3_24_0);
+        step2[25] = SUB_EPI16(step2[25], s3_25_0);
+        step2[26] = SUB_EPI16(step2[26], s3_26_0);
+        step2[27] = SUB_EPI16(step2[27], s3_27_0);
+        step1[28] = SUB_EPI16(step1[28], s3_28_0);
+        step1[29] = SUB_EPI16(step1[29], s3_29_0);
+        step1[30] = SUB_EPI16(step1[30], s3_30_0);
+        step1[31] = SUB_EPI16(step1[31], s3_31_0);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x32(
+            &step2[0], &step2[1], &step2[2], &step2[3],
+            &step2[4], &step2[5], &step2[6], &step2[7],
+            &step2[8], &step2[9], &step2[10], &step2[11],
+            &step2[12], &step2[13], &step2[14], &step2[15],
+            &step1[16], &step1[17], &step1[18], &step1[19],
+            &step2[20], &step2[21], &step2[22], &step2[23],
+            &step2[24], &step2[25], &step2[26], &step2[27],
+            &step1[28], &step1[29], &step1[30], &step1[31]);
+        if (overflow) {
+          HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        step2[0] = _mm_add_epi16(step2[ 0], kOne);
+        step2[1] = _mm_add_epi16(step2[ 1], kOne);
+        step2[2] = _mm_add_epi16(step2[ 2], kOne);
+        step2[3] = _mm_add_epi16(step2[ 3], kOne);
+        step2[4] = _mm_add_epi16(step2[ 4], kOne);
+        step2[5] = _mm_add_epi16(step2[ 5], kOne);
+        step2[6] = _mm_add_epi16(step2[ 6], kOne);
+        step2[7] = _mm_add_epi16(step2[ 7], kOne);
+        step2[8] = _mm_add_epi16(step2[ 8], kOne);
+        step2[9] = _mm_add_epi16(step2[ 9], kOne);
+        step2[10] = _mm_add_epi16(step2[10], kOne);
+        step2[11] = _mm_add_epi16(step2[11], kOne);
+        step2[12] = _mm_add_epi16(step2[12], kOne);
+        step2[13] = _mm_add_epi16(step2[13], kOne);
+        step2[14] = _mm_add_epi16(step2[14], kOne);
+        step2[15] = _mm_add_epi16(step2[15], kOne);
+        step1[16] = _mm_add_epi16(step1[16], kOne);
+        step1[17] = _mm_add_epi16(step1[17], kOne);
+        step1[18] = _mm_add_epi16(step1[18], kOne);
+        step1[19] = _mm_add_epi16(step1[19], kOne);
+        step2[20] = _mm_add_epi16(step2[20], kOne);
+        step2[21] = _mm_add_epi16(step2[21], kOne);
+        step2[22] = _mm_add_epi16(step2[22], kOne);
+        step2[23] = _mm_add_epi16(step2[23], kOne);
+        step2[24] = _mm_add_epi16(step2[24], kOne);
+        step2[25] = _mm_add_epi16(step2[25], kOne);
+        step2[26] = _mm_add_epi16(step2[26], kOne);
+        step2[27] = _mm_add_epi16(step2[27], kOne);
+        step1[28] = _mm_add_epi16(step1[28], kOne);
+        step1[29] = _mm_add_epi16(step1[29], kOne);
+        step1[30] = _mm_add_epi16(step1[30], kOne);
+        step1[31] = _mm_add_epi16(step1[31], kOne);
+
+        step2[0] = _mm_srai_epi16(step2[ 0], 2);
+        step2[1] = _mm_srai_epi16(step2[ 1], 2);
+        step2[2] = _mm_srai_epi16(step2[ 2], 2);
+        step2[3] = _mm_srai_epi16(step2[ 3], 2);
+        step2[4] = _mm_srai_epi16(step2[ 4], 2);
+        step2[5] = _mm_srai_epi16(step2[ 5], 2);
+        step2[6] = _mm_srai_epi16(step2[ 6], 2);
+        step2[7] = _mm_srai_epi16(step2[ 7], 2);
+        step2[8] = _mm_srai_epi16(step2[ 8], 2);
+        step2[9] = _mm_srai_epi16(step2[ 9], 2);
+        step2[10] = _mm_srai_epi16(step2[10], 2);
+        step2[11] = _mm_srai_epi16(step2[11], 2);
+        step2[12] = _mm_srai_epi16(step2[12], 2);
+        step2[13] = _mm_srai_epi16(step2[13], 2);
+        step2[14] = _mm_srai_epi16(step2[14], 2);
+        step2[15] = _mm_srai_epi16(step2[15], 2);
+        step1[16] = _mm_srai_epi16(step1[16], 2);
+        step1[17] = _mm_srai_epi16(step1[17], 2);
+        step1[18] = _mm_srai_epi16(step1[18], 2);
+        step1[19] = _mm_srai_epi16(step1[19], 2);
+        step2[20] = _mm_srai_epi16(step2[20], 2);
+        step2[21] = _mm_srai_epi16(step2[21], 2);
+        step2[22] = _mm_srai_epi16(step2[22], 2);
+        step2[23] = _mm_srai_epi16(step2[23], 2);
+        step2[24] = _mm_srai_epi16(step2[24], 2);
+        step2[25] = _mm_srai_epi16(step2[25], 2);
+        step2[26] = _mm_srai_epi16(step2[26], 2);
+        step2[27] = _mm_srai_epi16(step2[27], 2);
+        step1[28] = _mm_srai_epi16(step1[28], 2);
+        step1[29] = _mm_srai_epi16(step1[29], 2);
+        step1[30] = _mm_srai_epi16(step1[30], 2);
+        step1[31] = _mm_srai_epi16(step1[31], 2);
+      }
+#endif  // !FDCT32x32_HIGH_PRECISION
+
+#if FDCT32x32_HIGH_PRECISION
+      if (pass == 0) {
+#endif
+      // Stage 3
+      {
+        step3[0] = ADD_EPI16(step2[(8 - 1)], step2[0]);
+        step3[1] = ADD_EPI16(step2[(8 - 2)], step2[1]);
+        step3[2] = ADD_EPI16(step2[(8 - 3)], step2[2]);
+        step3[3] = ADD_EPI16(step2[(8 - 4)], step2[3]);
+        step3[4] = SUB_EPI16(step2[(8 - 5)], step2[4]);
+        step3[5] = SUB_EPI16(step2[(8 - 6)], step2[5]);
+        step3[6] = SUB_EPI16(step2[(8 - 7)], step2[6]);
+        step3[7] = SUB_EPI16(step2[(8 - 8)], step2[7]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step3[0], &step3[1], &step3[2],
+                                           &step3[3], &step3[4], &step3[5],
+                                           &step3[6], &step3[7]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
+        const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
+        const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
+        const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
+        const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
+        const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
+        const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
+        const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
+        const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
+        const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
+        const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
+        const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_10_6 = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
+        const __m128i s3_10_7 = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
+        const __m128i s3_11_6 = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
+        const __m128i s3_11_7 = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
+        const __m128i s3_12_6 = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
+        const __m128i s3_12_7 = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
+        const __m128i s3_13_6 = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
+        const __m128i s3_13_7 = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
+        // Combine
+        step3[10] = _mm_packs_epi32(s3_10_6, s3_10_7);
+        step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7);
+        step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7);
+        step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&step3[10], &step3[11],
+                                           &step3[12], &step3[13]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        step3[16] = ADD_EPI16(step2[23], step1[16]);
+        step3[17] = ADD_EPI16(step2[22], step1[17]);
+        step3[18] = ADD_EPI16(step2[21], step1[18]);
+        step3[19] = ADD_EPI16(step2[20], step1[19]);
+        step3[20] = SUB_EPI16(step1[19], step2[20]);
+        step3[21] = SUB_EPI16(step1[18], step2[21]);
+        step3[22] = SUB_EPI16(step1[17], step2[22]);
+        step3[23] = SUB_EPI16(step1[16], step2[23]);
+        step3[24] = SUB_EPI16(step1[31], step2[24]);
+        step3[25] = SUB_EPI16(step1[30], step2[25]);
+        step3[26] = SUB_EPI16(step1[29], step2[26]);
+        step3[27] = SUB_EPI16(step1[28], step2[27]);
+        step3[28] = ADD_EPI16(step2[27], step1[28]);
+        step3[29] = ADD_EPI16(step2[26], step1[29]);
+        step3[30] = ADD_EPI16(step2[25], step1[30]);
+        step3[31] = ADD_EPI16(step2[24], step1[31]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step3[16], &step3[17], &step3[18], &step3[19],
+            &step3[20], &step3[21], &step3[22], &step3[23],
+            &step3[24], &step3[25], &step3[26], &step3[27],
+            &step3[28], &step3[29], &step3[30], &step3[31]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+
+      // Stage 4
+      {
+        step1[0] = ADD_EPI16(step3[ 3], step3[ 0]);
+        step1[1] = ADD_EPI16(step3[ 2], step3[ 1]);
+        step1[2] = SUB_EPI16(step3[ 1], step3[ 2]);
+        step1[3] = SUB_EPI16(step3[ 0], step3[ 3]);
+        step1[8] = ADD_EPI16(step3[11], step2[ 8]);
+        step1[9] = ADD_EPI16(step3[10], step2[ 9]);
+        step1[10] = SUB_EPI16(step2[ 9], step3[10]);
+        step1[11] = SUB_EPI16(step2[ 8], step3[11]);
+        step1[12] = SUB_EPI16(step2[15], step3[12]);
+        step1[13] = SUB_EPI16(step2[14], step3[13]);
+        step1[14] = ADD_EPI16(step3[13], step2[14]);
+        step1[15] = ADD_EPI16(step3[12], step2[15]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step1[0], &step1[1], &step1[2], &step1[3],
+            &step1[4], &step1[5], &step1[6], &step1[7],
+            &step1[8], &step1[9], &step1[10], &step1[11],
+            &step1[12], &step1[13], &step1[14], &step1[15]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]);
+        const __m128i s1_05_1 = _mm_unpackhi_epi16(step3[6], step3[5]);
+        const __m128i s1_05_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_m16);
+        const __m128i s1_05_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_m16);
+        const __m128i s1_06_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_p16);
+        const __m128i s1_06_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m128i s1_05_4 = _mm_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_05_5 = _mm_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_06_4 = _mm_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_06_5 = _mm_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_05_6 = _mm_srai_epi32(s1_05_4, DCT_CONST_BITS);
+        const __m128i s1_05_7 = _mm_srai_epi32(s1_05_5, DCT_CONST_BITS);
+        const __m128i s1_06_6 = _mm_srai_epi32(s1_06_4, DCT_CONST_BITS);
+        const __m128i s1_06_7 = _mm_srai_epi32(s1_06_5, DCT_CONST_BITS);
+        // Combine
+        step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7);
+        step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x2(&step1[5], &step1[6]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]);
+        const __m128i s1_18_1 = _mm_unpackhi_epi16(step3[18], step3[29]);
+        const __m128i s1_19_0 = _mm_unpacklo_epi16(step3[19], step3[28]);
+        const __m128i s1_19_1 = _mm_unpackhi_epi16(step3[19], step3[28]);
+        const __m128i s1_20_0 = _mm_unpacklo_epi16(step3[20], step3[27]);
+        const __m128i s1_20_1 = _mm_unpackhi_epi16(step3[20], step3[27]);
+        const __m128i s1_21_0 = _mm_unpacklo_epi16(step3[21], step3[26]);
+        const __m128i s1_21_1 = _mm_unpackhi_epi16(step3[21], step3[26]);
+        const __m128i s1_18_2 = _mm_madd_epi16(s1_18_0, k__cospi_m08_p24);
+        const __m128i s1_18_3 = _mm_madd_epi16(s1_18_1, k__cospi_m08_p24);
+        const __m128i s1_19_2 = _mm_madd_epi16(s1_19_0, k__cospi_m08_p24);
+        const __m128i s1_19_3 = _mm_madd_epi16(s1_19_1, k__cospi_m08_p24);
+        const __m128i s1_20_2 = _mm_madd_epi16(s1_20_0, k__cospi_m24_m08);
+        const __m128i s1_20_3 = _mm_madd_epi16(s1_20_1, k__cospi_m24_m08);
+        const __m128i s1_21_2 = _mm_madd_epi16(s1_21_0, k__cospi_m24_m08);
+        const __m128i s1_21_3 = _mm_madd_epi16(s1_21_1, k__cospi_m24_m08);
+        const __m128i s1_26_2 = _mm_madd_epi16(s1_21_0, k__cospi_m08_p24);
+        const __m128i s1_26_3 = _mm_madd_epi16(s1_21_1, k__cospi_m08_p24);
+        const __m128i s1_27_2 = _mm_madd_epi16(s1_20_0, k__cospi_m08_p24);
+        const __m128i s1_27_3 = _mm_madd_epi16(s1_20_1, k__cospi_m08_p24);
+        const __m128i s1_28_2 = _mm_madd_epi16(s1_19_0, k__cospi_p24_p08);
+        const __m128i s1_28_3 = _mm_madd_epi16(s1_19_1, k__cospi_p24_p08);
+        const __m128i s1_29_2 = _mm_madd_epi16(s1_18_0, k__cospi_p24_p08);
+        const __m128i s1_29_3 = _mm_madd_epi16(s1_18_1, k__cospi_p24_p08);
+        // dct_const_round_shift
+        const __m128i s1_18_4 = _mm_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_18_5 = _mm_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_19_4 = _mm_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_19_5 = _mm_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_20_4 = _mm_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_20_5 = _mm_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_21_4 = _mm_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_21_5 = _mm_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_26_4 = _mm_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_26_5 = _mm_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_27_4 = _mm_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_27_5 = _mm_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_28_4 = _mm_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_28_5 = _mm_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_29_4 = _mm_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_29_5 = _mm_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_18_6 = _mm_srai_epi32(s1_18_4, DCT_CONST_BITS);
+        const __m128i s1_18_7 = _mm_srai_epi32(s1_18_5, DCT_CONST_BITS);
+        const __m128i s1_19_6 = _mm_srai_epi32(s1_19_4, DCT_CONST_BITS);
+        const __m128i s1_19_7 = _mm_srai_epi32(s1_19_5, DCT_CONST_BITS);
+        const __m128i s1_20_6 = _mm_srai_epi32(s1_20_4, DCT_CONST_BITS);
+        const __m128i s1_20_7 = _mm_srai_epi32(s1_20_5, DCT_CONST_BITS);
+        const __m128i s1_21_6 = _mm_srai_epi32(s1_21_4, DCT_CONST_BITS);
+        const __m128i s1_21_7 = _mm_srai_epi32(s1_21_5, DCT_CONST_BITS);
+        const __m128i s1_26_6 = _mm_srai_epi32(s1_26_4, DCT_CONST_BITS);
+        const __m128i s1_26_7 = _mm_srai_epi32(s1_26_5, DCT_CONST_BITS);
+        const __m128i s1_27_6 = _mm_srai_epi32(s1_27_4, DCT_CONST_BITS);
+        const __m128i s1_27_7 = _mm_srai_epi32(s1_27_5, DCT_CONST_BITS);
+        const __m128i s1_28_6 = _mm_srai_epi32(s1_28_4, DCT_CONST_BITS);
+        const __m128i s1_28_7 = _mm_srai_epi32(s1_28_5, DCT_CONST_BITS);
+        const __m128i s1_29_6 = _mm_srai_epi32(s1_29_4, DCT_CONST_BITS);
+        const __m128i s1_29_7 = _mm_srai_epi32(s1_29_5, DCT_CONST_BITS);
+        // Combine
+        step1[18] = _mm_packs_epi32(s1_18_6, s1_18_7);
+        step1[19] = _mm_packs_epi32(s1_19_6, s1_19_7);
+        step1[20] = _mm_packs_epi32(s1_20_6, s1_20_7);
+        step1[21] = _mm_packs_epi32(s1_21_6, s1_21_7);
+        step1[26] = _mm_packs_epi32(s1_26_6, s1_26_7);
+        step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7);
+        step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7);
+        step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step1[18], &step1[19], &step1[20],
+                                           &step1[21], &step1[26], &step1[27],
+                                           &step1[28], &step1[29]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Stage 5
+      {
+        step2[4] = ADD_EPI16(step1[5], step3[4]);
+        step2[5] = SUB_EPI16(step3[4], step1[5]);
+        step2[6] = SUB_EPI16(step3[7], step1[6]);
+        step2[7] = ADD_EPI16(step1[6], step3[7]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&step2[4], &step2[5],
+                                           &step2[6], &step2[7]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]);
+        const __m128i out_00_1 = _mm_unpackhi_epi16(step1[0], step1[1]);
+        const __m128i out_08_0 = _mm_unpacklo_epi16(step1[2], step1[3]);
+        const __m128i out_08_1 = _mm_unpackhi_epi16(step1[2], step1[3]);
+        const __m128i out_00_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_p16);
+        const __m128i out_00_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_p16);
+        const __m128i out_16_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_m16);
+        const __m128i out_16_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_m16);
+        const __m128i out_08_2 = _mm_madd_epi16(out_08_0, k__cospi_p24_p08);
+        const __m128i out_08_3 = _mm_madd_epi16(out_08_1, k__cospi_p24_p08);
+        const __m128i out_24_2 = _mm_madd_epi16(out_08_0, k__cospi_m08_p24);
+        const __m128i out_24_3 = _mm_madd_epi16(out_08_1, k__cospi_m08_p24);
+        // dct_const_round_shift
+        const __m128i out_00_4 = _mm_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_00_5 = _mm_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_16_4 = _mm_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_16_5 = _mm_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_08_4 = _mm_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_08_5 = _mm_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_24_4 = _mm_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_24_5 = _mm_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_00_6 = _mm_srai_epi32(out_00_4, DCT_CONST_BITS);
+        const __m128i out_00_7 = _mm_srai_epi32(out_00_5, DCT_CONST_BITS);
+        const __m128i out_16_6 = _mm_srai_epi32(out_16_4, DCT_CONST_BITS);
+        const __m128i out_16_7 = _mm_srai_epi32(out_16_5, DCT_CONST_BITS);
+        const __m128i out_08_6 = _mm_srai_epi32(out_08_4, DCT_CONST_BITS);
+        const __m128i out_08_7 = _mm_srai_epi32(out_08_5, DCT_CONST_BITS);
+        const __m128i out_24_6 = _mm_srai_epi32(out_24_4, DCT_CONST_BITS);
+        const __m128i out_24_7 = _mm_srai_epi32(out_24_5, DCT_CONST_BITS);
+        // Combine
+        out[ 0] = _mm_packs_epi32(out_00_6, out_00_7);
+        out[16] = _mm_packs_epi32(out_16_6, out_16_7);
+        out[ 8] = _mm_packs_epi32(out_08_6, out_08_7);
+        out[24] = _mm_packs_epi32(out_24_6, out_24_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&out[0], &out[16],
+                                           &out[8], &out[24]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[ 9], step1[14]);
+        const __m128i s2_09_1 = _mm_unpackhi_epi16(step1[ 9], step1[14]);
+        const __m128i s2_10_0 = _mm_unpacklo_epi16(step1[10], step1[13]);
+        const __m128i s2_10_1 = _mm_unpackhi_epi16(step1[10], step1[13]);
+        const __m128i s2_09_2 = _mm_madd_epi16(s2_09_0, k__cospi_m08_p24);
+        const __m128i s2_09_3 = _mm_madd_epi16(s2_09_1, k__cospi_m08_p24);
+        const __m128i s2_10_2 = _mm_madd_epi16(s2_10_0, k__cospi_m24_m08);
+        const __m128i s2_10_3 = _mm_madd_epi16(s2_10_1, k__cospi_m24_m08);
+        const __m128i s2_13_2 = _mm_madd_epi16(s2_10_0, k__cospi_m08_p24);
+        const __m128i s2_13_3 = _mm_madd_epi16(s2_10_1, k__cospi_m08_p24);
+        const __m128i s2_14_2 = _mm_madd_epi16(s2_09_0, k__cospi_p24_p08);
+        const __m128i s2_14_3 = _mm_madd_epi16(s2_09_1, k__cospi_p24_p08);
+        // dct_const_round_shift
+        const __m128i s2_09_4 = _mm_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_09_5 = _mm_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_10_4 = _mm_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_10_5 = _mm_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_13_4 = _mm_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_13_5 = _mm_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_14_4 = _mm_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_14_5 = _mm_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_09_6 = _mm_srai_epi32(s2_09_4, DCT_CONST_BITS);
+        const __m128i s2_09_7 = _mm_srai_epi32(s2_09_5, DCT_CONST_BITS);
+        const __m128i s2_10_6 = _mm_srai_epi32(s2_10_4, DCT_CONST_BITS);
+        const __m128i s2_10_7 = _mm_srai_epi32(s2_10_5, DCT_CONST_BITS);
+        const __m128i s2_13_6 = _mm_srai_epi32(s2_13_4, DCT_CONST_BITS);
+        const __m128i s2_13_7 = _mm_srai_epi32(s2_13_5, DCT_CONST_BITS);
+        const __m128i s2_14_6 = _mm_srai_epi32(s2_14_4, DCT_CONST_BITS);
+        const __m128i s2_14_7 = _mm_srai_epi32(s2_14_5, DCT_CONST_BITS);
+        // Combine
+        step2[ 9] = _mm_packs_epi32(s2_09_6, s2_09_7);
+        step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7);
+        step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7);
+        step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&step2[9], &step2[10],
+                                           &step2[13], &step2[14]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        step2[16] = ADD_EPI16(step1[19], step3[16]);
+        step2[17] = ADD_EPI16(step1[18], step3[17]);
+        step2[18] = SUB_EPI16(step3[17], step1[18]);
+        step2[19] = SUB_EPI16(step3[16], step1[19]);
+        step2[20] = SUB_EPI16(step3[23], step1[20]);
+        step2[21] = SUB_EPI16(step3[22], step1[21]);
+        step2[22] = ADD_EPI16(step1[21], step3[22]);
+        step2[23] = ADD_EPI16(step1[20], step3[23]);
+        step2[24] = ADD_EPI16(step1[27], step3[24]);
+        step2[25] = ADD_EPI16(step1[26], step3[25]);
+        step2[26] = SUB_EPI16(step3[25], step1[26]);
+        step2[27] = SUB_EPI16(step3[24], step1[27]);
+        step2[28] = SUB_EPI16(step3[31], step1[28]);
+        step2[29] = SUB_EPI16(step3[30], step1[29]);
+        step2[30] = ADD_EPI16(step1[29], step3[30]);
+        step2[31] = ADD_EPI16(step1[28], step3[31]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step2[16], &step2[17], &step2[18], &step2[19],
+            &step2[20], &step2[21], &step2[22], &step2[23],
+            &step2[24], &step2[25], &step2[26], &step2[27],
+            &step2[28], &step2[29], &step2[30], &step2[31]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Stage 6
+      {
+        const __m128i out_04_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
+        const __m128i out_04_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
+        const __m128i out_20_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
+        const __m128i out_20_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
+        const __m128i out_12_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
+        const __m128i out_12_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
+        const __m128i out_28_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
+        const __m128i out_28_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
+        const __m128i out_04_2 = _mm_madd_epi16(out_04_0, k__cospi_p28_p04);
+        const __m128i out_04_3 = _mm_madd_epi16(out_04_1, k__cospi_p28_p04);
+        const __m128i out_20_2 = _mm_madd_epi16(out_20_0, k__cospi_p12_p20);
+        const __m128i out_20_3 = _mm_madd_epi16(out_20_1, k__cospi_p12_p20);
+        const __m128i out_12_2 = _mm_madd_epi16(out_12_0, k__cospi_m20_p12);
+        const __m128i out_12_3 = _mm_madd_epi16(out_12_1, k__cospi_m20_p12);
+        const __m128i out_28_2 = _mm_madd_epi16(out_28_0, k__cospi_m04_p28);
+        const __m128i out_28_3 = _mm_madd_epi16(out_28_1, k__cospi_m04_p28);
+        // dct_const_round_shift
+        const __m128i out_04_4 = _mm_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_04_5 = _mm_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_20_4 = _mm_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_20_5 = _mm_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_12_4 = _mm_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_12_5 = _mm_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_28_4 = _mm_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_28_5 = _mm_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_04_6 = _mm_srai_epi32(out_04_4, DCT_CONST_BITS);
+        const __m128i out_04_7 = _mm_srai_epi32(out_04_5, DCT_CONST_BITS);
+        const __m128i out_20_6 = _mm_srai_epi32(out_20_4, DCT_CONST_BITS);
+        const __m128i out_20_7 = _mm_srai_epi32(out_20_5, DCT_CONST_BITS);
+        const __m128i out_12_6 = _mm_srai_epi32(out_12_4, DCT_CONST_BITS);
+        const __m128i out_12_7 = _mm_srai_epi32(out_12_5, DCT_CONST_BITS);
+        const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS);
+        const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS);
+        // Combine
+        out[4] = _mm_packs_epi32(out_04_6, out_04_7);
+        out[20] = _mm_packs_epi32(out_20_6, out_20_7);
+        out[12] = _mm_packs_epi32(out_12_6, out_12_7);
+        out[28] = _mm_packs_epi32(out_28_6, out_28_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&out[4], &out[20],
+                                           &out[12], &out[28]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        step3[8] = ADD_EPI16(step2[ 9], step1[ 8]);
+        step3[9] = SUB_EPI16(step1[ 8], step2[ 9]);
+        step3[10] = SUB_EPI16(step1[11], step2[10]);
+        step3[11] = ADD_EPI16(step2[10], step1[11]);
+        step3[12] = ADD_EPI16(step2[13], step1[12]);
+        step3[13] = SUB_EPI16(step1[12], step2[13]);
+        step3[14] = SUB_EPI16(step1[15], step2[14]);
+        step3[15] = ADD_EPI16(step2[14], step1[15]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step3[8], &step3[9], &step3[10],
+                                           &step3[11], &step3[12], &step3[13],
+                                           &step3[14], &step3[15]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]);
+        const __m128i s3_17_1 = _mm_unpackhi_epi16(step2[17], step2[30]);
+        const __m128i s3_18_0 = _mm_unpacklo_epi16(step2[18], step2[29]);
+        const __m128i s3_18_1 = _mm_unpackhi_epi16(step2[18], step2[29]);
+        const __m128i s3_21_0 = _mm_unpacklo_epi16(step2[21], step2[26]);
+        const __m128i s3_21_1 = _mm_unpackhi_epi16(step2[21], step2[26]);
+        const __m128i s3_22_0 = _mm_unpacklo_epi16(step2[22], step2[25]);
+        const __m128i s3_22_1 = _mm_unpackhi_epi16(step2[22], step2[25]);
+        const __m128i s3_17_2 = _mm_madd_epi16(s3_17_0, k__cospi_m04_p28);
+        const __m128i s3_17_3 = _mm_madd_epi16(s3_17_1, k__cospi_m04_p28);
+        const __m128i s3_18_2 = _mm_madd_epi16(s3_18_0, k__cospi_m28_m04);
+        const __m128i s3_18_3 = _mm_madd_epi16(s3_18_1, k__cospi_m28_m04);
+        const __m128i s3_21_2 = _mm_madd_epi16(s3_21_0, k__cospi_m20_p12);
+        const __m128i s3_21_3 = _mm_madd_epi16(s3_21_1, k__cospi_m20_p12);
+        const __m128i s3_22_2 = _mm_madd_epi16(s3_22_0, k__cospi_m12_m20);
+        const __m128i s3_22_3 = _mm_madd_epi16(s3_22_1, k__cospi_m12_m20);
+        const __m128i s3_25_2 = _mm_madd_epi16(s3_22_0, k__cospi_m20_p12);
+        const __m128i s3_25_3 = _mm_madd_epi16(s3_22_1, k__cospi_m20_p12);
+        const __m128i s3_26_2 = _mm_madd_epi16(s3_21_0, k__cospi_p12_p20);
+        const __m128i s3_26_3 = _mm_madd_epi16(s3_21_1, k__cospi_p12_p20);
+        const __m128i s3_29_2 = _mm_madd_epi16(s3_18_0, k__cospi_m04_p28);
+        const __m128i s3_29_3 = _mm_madd_epi16(s3_18_1, k__cospi_m04_p28);
+        const __m128i s3_30_2 = _mm_madd_epi16(s3_17_0, k__cospi_p28_p04);
+        const __m128i s3_30_3 = _mm_madd_epi16(s3_17_1, k__cospi_p28_p04);
+        // dct_const_round_shift
+        const __m128i s3_17_4 = _mm_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_17_5 = _mm_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_18_4 = _mm_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_18_5 = _mm_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_21_4 = _mm_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_21_5 = _mm_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_22_4 = _mm_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_22_5 = _mm_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_17_6 = _mm_srai_epi32(s3_17_4, DCT_CONST_BITS);
+        const __m128i s3_17_7 = _mm_srai_epi32(s3_17_5, DCT_CONST_BITS);
+        const __m128i s3_18_6 = _mm_srai_epi32(s3_18_4, DCT_CONST_BITS);
+        const __m128i s3_18_7 = _mm_srai_epi32(s3_18_5, DCT_CONST_BITS);
+        const __m128i s3_21_6 = _mm_srai_epi32(s3_21_4, DCT_CONST_BITS);
+        const __m128i s3_21_7 = _mm_srai_epi32(s3_21_5, DCT_CONST_BITS);
+        const __m128i s3_22_6 = _mm_srai_epi32(s3_22_4, DCT_CONST_BITS);
+        const __m128i s3_22_7 = _mm_srai_epi32(s3_22_5, DCT_CONST_BITS);
+        const __m128i s3_25_4 = _mm_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_25_5 = _mm_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_26_4 = _mm_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_26_5 = _mm_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_29_4 = _mm_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_29_5 = _mm_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_30_4 = _mm_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_30_5 = _mm_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_25_6 = _mm_srai_epi32(s3_25_4, DCT_CONST_BITS);
+        const __m128i s3_25_7 = _mm_srai_epi32(s3_25_5, DCT_CONST_BITS);
+        const __m128i s3_26_6 = _mm_srai_epi32(s3_26_4, DCT_CONST_BITS);
+        const __m128i s3_26_7 = _mm_srai_epi32(s3_26_5, DCT_CONST_BITS);
+        const __m128i s3_29_6 = _mm_srai_epi32(s3_29_4, DCT_CONST_BITS);
+        const __m128i s3_29_7 = _mm_srai_epi32(s3_29_5, DCT_CONST_BITS);
+        const __m128i s3_30_6 = _mm_srai_epi32(s3_30_4, DCT_CONST_BITS);
+        const __m128i s3_30_7 = _mm_srai_epi32(s3_30_5, DCT_CONST_BITS);
+        // Combine
+        step3[17] = _mm_packs_epi32(s3_17_6, s3_17_7);
+        step3[18] = _mm_packs_epi32(s3_18_6, s3_18_7);
+        step3[21] = _mm_packs_epi32(s3_21_6, s3_21_7);
+        step3[22] = _mm_packs_epi32(s3_22_6, s3_22_7);
+        // Combine
+        step3[25] = _mm_packs_epi32(s3_25_6, s3_25_7);
+        step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7);
+        step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7);
+        step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step3[17], &step3[18], &step3[21],
+                                           &step3[22], &step3[25], &step3[26],
+                                           &step3[29], &step3[30]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Stage 7
+      {
+        const __m128i out_02_0 = _mm_unpacklo_epi16(step3[ 8], step3[15]);
+        const __m128i out_02_1 = _mm_unpackhi_epi16(step3[ 8], step3[15]);
+        const __m128i out_18_0 = _mm_unpacklo_epi16(step3[ 9], step3[14]);
+        const __m128i out_18_1 = _mm_unpackhi_epi16(step3[ 9], step3[14]);
+        const __m128i out_10_0 = _mm_unpacklo_epi16(step3[10], step3[13]);
+        const __m128i out_10_1 = _mm_unpackhi_epi16(step3[10], step3[13]);
+        const __m128i out_26_0 = _mm_unpacklo_epi16(step3[11], step3[12]);
+        const __m128i out_26_1 = _mm_unpackhi_epi16(step3[11], step3[12]);
+        const __m128i out_02_2 = _mm_madd_epi16(out_02_0, k__cospi_p30_p02);
+        const __m128i out_02_3 = _mm_madd_epi16(out_02_1, k__cospi_p30_p02);
+        const __m128i out_18_2 = _mm_madd_epi16(out_18_0, k__cospi_p14_p18);
+        const __m128i out_18_3 = _mm_madd_epi16(out_18_1, k__cospi_p14_p18);
+        const __m128i out_10_2 = _mm_madd_epi16(out_10_0, k__cospi_p22_p10);
+        const __m128i out_10_3 = _mm_madd_epi16(out_10_1, k__cospi_p22_p10);
+        const __m128i out_26_2 = _mm_madd_epi16(out_26_0, k__cospi_p06_p26);
+        const __m128i out_26_3 = _mm_madd_epi16(out_26_1, k__cospi_p06_p26);
+        const __m128i out_06_2 = _mm_madd_epi16(out_26_0, k__cospi_m26_p06);
+        const __m128i out_06_3 = _mm_madd_epi16(out_26_1, k__cospi_m26_p06);
+        const __m128i out_22_2 = _mm_madd_epi16(out_10_0, k__cospi_m10_p22);
+        const __m128i out_22_3 = _mm_madd_epi16(out_10_1, k__cospi_m10_p22);
+        const __m128i out_14_2 = _mm_madd_epi16(out_18_0, k__cospi_m18_p14);
+        const __m128i out_14_3 = _mm_madd_epi16(out_18_1, k__cospi_m18_p14);
+        const __m128i out_30_2 = _mm_madd_epi16(out_02_0, k__cospi_m02_p30);
+        const __m128i out_30_3 = _mm_madd_epi16(out_02_1, k__cospi_m02_p30);
+        // dct_const_round_shift
+        const __m128i out_02_4 = _mm_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_02_5 = _mm_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_18_4 = _mm_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_18_5 = _mm_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_10_4 = _mm_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_10_5 = _mm_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_26_4 = _mm_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_26_5 = _mm_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_06_4 = _mm_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_06_5 = _mm_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_22_4 = _mm_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_22_5 = _mm_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_14_4 = _mm_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_14_5 = _mm_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_30_4 = _mm_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_30_5 = _mm_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_02_6 = _mm_srai_epi32(out_02_4, DCT_CONST_BITS);
+        const __m128i out_02_7 = _mm_srai_epi32(out_02_5, DCT_CONST_BITS);
+        const __m128i out_18_6 = _mm_srai_epi32(out_18_4, DCT_CONST_BITS);
+        const __m128i out_18_7 = _mm_srai_epi32(out_18_5, DCT_CONST_BITS);
+        const __m128i out_10_6 = _mm_srai_epi32(out_10_4, DCT_CONST_BITS);
+        const __m128i out_10_7 = _mm_srai_epi32(out_10_5, DCT_CONST_BITS);
+        const __m128i out_26_6 = _mm_srai_epi32(out_26_4, DCT_CONST_BITS);
+        const __m128i out_26_7 = _mm_srai_epi32(out_26_5, DCT_CONST_BITS);
+        const __m128i out_06_6 = _mm_srai_epi32(out_06_4, DCT_CONST_BITS);
+        const __m128i out_06_7 = _mm_srai_epi32(out_06_5, DCT_CONST_BITS);
+        const __m128i out_22_6 = _mm_srai_epi32(out_22_4, DCT_CONST_BITS);
+        const __m128i out_22_7 = _mm_srai_epi32(out_22_5, DCT_CONST_BITS);
+        const __m128i out_14_6 = _mm_srai_epi32(out_14_4, DCT_CONST_BITS);
+        const __m128i out_14_7 = _mm_srai_epi32(out_14_5, DCT_CONST_BITS);
+        const __m128i out_30_6 = _mm_srai_epi32(out_30_4, DCT_CONST_BITS);
+        const __m128i out_30_7 = _mm_srai_epi32(out_30_5, DCT_CONST_BITS);
+        // Combine
+        out[ 2] = _mm_packs_epi32(out_02_6, out_02_7);
+        out[18] = _mm_packs_epi32(out_18_6, out_18_7);
+        out[10] = _mm_packs_epi32(out_10_6, out_10_7);
+        out[26] = _mm_packs_epi32(out_26_6, out_26_7);
+        out[ 6] = _mm_packs_epi32(out_06_6, out_06_7);
+        out[22] = _mm_packs_epi32(out_22_6, out_22_7);
+        out[14] = _mm_packs_epi32(out_14_6, out_14_7);
+        out[30] = _mm_packs_epi32(out_30_6, out_30_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&out[2], &out[18], &out[10],
+                                           &out[26], &out[6], &out[22],
+                                           &out[14], &out[30]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        step1[16] = ADD_EPI16(step3[17], step2[16]);
+        step1[17] = SUB_EPI16(step2[16], step3[17]);
+        step1[18] = SUB_EPI16(step2[19], step3[18]);
+        step1[19] = ADD_EPI16(step3[18], step2[19]);
+        step1[20] = ADD_EPI16(step3[21], step2[20]);
+        step1[21] = SUB_EPI16(step2[20], step3[21]);
+        step1[22] = SUB_EPI16(step2[23], step3[22]);
+        step1[23] = ADD_EPI16(step3[22], step2[23]);
+        step1[24] = ADD_EPI16(step3[25], step2[24]);
+        step1[25] = SUB_EPI16(step2[24], step3[25]);
+        step1[26] = SUB_EPI16(step2[27], step3[26]);
+        step1[27] = ADD_EPI16(step3[26], step2[27]);
+        step1[28] = ADD_EPI16(step3[29], step2[28]);
+        step1[29] = SUB_EPI16(step2[28], step3[29]);
+        step1[30] = SUB_EPI16(step2[31], step3[30]);
+        step1[31] = ADD_EPI16(step3[30], step2[31]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step1[16], &step1[17], &step1[18], &step1[19],
+            &step1[20], &step1[21], &step1[22], &step1[23],
+            &step1[24], &step1[25], &step1[26], &step1[27],
+            &step1[28], &step1[29], &step1[30], &step1[31]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+             HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Final stage --- outputs indices are bit-reversed.
+      {
+        const __m128i out_01_0 = _mm_unpacklo_epi16(step1[16], step1[31]);
+        const __m128i out_01_1 = _mm_unpackhi_epi16(step1[16], step1[31]);
+        const __m128i out_17_0 = _mm_unpacklo_epi16(step1[17], step1[30]);
+        const __m128i out_17_1 = _mm_unpackhi_epi16(step1[17], step1[30]);
+        const __m128i out_09_0 = _mm_unpacklo_epi16(step1[18], step1[29]);
+        const __m128i out_09_1 = _mm_unpackhi_epi16(step1[18], step1[29]);
+        const __m128i out_25_0 = _mm_unpacklo_epi16(step1[19], step1[28]);
+        const __m128i out_25_1 = _mm_unpackhi_epi16(step1[19], step1[28]);
+        const __m128i out_01_2 = _mm_madd_epi16(out_01_0, k__cospi_p31_p01);
+        const __m128i out_01_3 = _mm_madd_epi16(out_01_1, k__cospi_p31_p01);
+        const __m128i out_17_2 = _mm_madd_epi16(out_17_0, k__cospi_p15_p17);
+        const __m128i out_17_3 = _mm_madd_epi16(out_17_1, k__cospi_p15_p17);
+        const __m128i out_09_2 = _mm_madd_epi16(out_09_0, k__cospi_p23_p09);
+        const __m128i out_09_3 = _mm_madd_epi16(out_09_1, k__cospi_p23_p09);
+        const __m128i out_25_2 = _mm_madd_epi16(out_25_0, k__cospi_p07_p25);
+        const __m128i out_25_3 = _mm_madd_epi16(out_25_1, k__cospi_p07_p25);
+        const __m128i out_07_2 = _mm_madd_epi16(out_25_0, k__cospi_m25_p07);
+        const __m128i out_07_3 = _mm_madd_epi16(out_25_1, k__cospi_m25_p07);
+        const __m128i out_23_2 = _mm_madd_epi16(out_09_0, k__cospi_m09_p23);
+        const __m128i out_23_3 = _mm_madd_epi16(out_09_1, k__cospi_m09_p23);
+        const __m128i out_15_2 = _mm_madd_epi16(out_17_0, k__cospi_m17_p15);
+        const __m128i out_15_3 = _mm_madd_epi16(out_17_1, k__cospi_m17_p15);
+        const __m128i out_31_2 = _mm_madd_epi16(out_01_0, k__cospi_m01_p31);
+        const __m128i out_31_3 = _mm_madd_epi16(out_01_1, k__cospi_m01_p31);
+        // dct_const_round_shift
+        const __m128i out_01_4 = _mm_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_01_5 = _mm_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_17_4 = _mm_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_17_5 = _mm_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_09_4 = _mm_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_09_5 = _mm_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_25_4 = _mm_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_25_5 = _mm_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_07_4 = _mm_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_07_5 = _mm_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_23_4 = _mm_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_23_5 = _mm_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_15_4 = _mm_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_15_5 = _mm_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_31_4 = _mm_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_31_5 = _mm_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_01_6 = _mm_srai_epi32(out_01_4, DCT_CONST_BITS);
+        const __m128i out_01_7 = _mm_srai_epi32(out_01_5, DCT_CONST_BITS);
+        const __m128i out_17_6 = _mm_srai_epi32(out_17_4, DCT_CONST_BITS);
+        const __m128i out_17_7 = _mm_srai_epi32(out_17_5, DCT_CONST_BITS);
+        const __m128i out_09_6 = _mm_srai_epi32(out_09_4, DCT_CONST_BITS);
+        const __m128i out_09_7 = _mm_srai_epi32(out_09_5, DCT_CONST_BITS);
+        const __m128i out_25_6 = _mm_srai_epi32(out_25_4, DCT_CONST_BITS);
+        const __m128i out_25_7 = _mm_srai_epi32(out_25_5, DCT_CONST_BITS);
+        const __m128i out_07_6 = _mm_srai_epi32(out_07_4, DCT_CONST_BITS);
+        const __m128i out_07_7 = _mm_srai_epi32(out_07_5, DCT_CONST_BITS);
+        const __m128i out_23_6 = _mm_srai_epi32(out_23_4, DCT_CONST_BITS);
+        const __m128i out_23_7 = _mm_srai_epi32(out_23_5, DCT_CONST_BITS);
+        const __m128i out_15_6 = _mm_srai_epi32(out_15_4, DCT_CONST_BITS);
+        const __m128i out_15_7 = _mm_srai_epi32(out_15_5, DCT_CONST_BITS);
+        const __m128i out_31_6 = _mm_srai_epi32(out_31_4, DCT_CONST_BITS);
+        const __m128i out_31_7 = _mm_srai_epi32(out_31_5, DCT_CONST_BITS);
+        // Combine
+        out[ 1] = _mm_packs_epi32(out_01_6, out_01_7);
+        out[17] = _mm_packs_epi32(out_17_6, out_17_7);
+        out[ 9] = _mm_packs_epi32(out_09_6, out_09_7);
+        out[25] = _mm_packs_epi32(out_25_6, out_25_7);
+        out[ 7] = _mm_packs_epi32(out_07_6, out_07_7);
+        out[23] = _mm_packs_epi32(out_23_6, out_23_7);
+        out[15] = _mm_packs_epi32(out_15_6, out_15_7);
+        out[31] = _mm_packs_epi32(out_31_6, out_31_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&out[1], &out[17], &out[9],
+                                           &out[25], &out[7], &out[23],
+                                           &out[15], &out[31]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]);
+        const __m128i out_05_1 = _mm_unpackhi_epi16(step1[20], step1[27]);
+        const __m128i out_21_0 = _mm_unpacklo_epi16(step1[21], step1[26]);
+        const __m128i out_21_1 = _mm_unpackhi_epi16(step1[21], step1[26]);
+        const __m128i out_13_0 = _mm_unpacklo_epi16(step1[22], step1[25]);
+        const __m128i out_13_1 = _mm_unpackhi_epi16(step1[22], step1[25]);
+        const __m128i out_29_0 = _mm_unpacklo_epi16(step1[23], step1[24]);
+        const __m128i out_29_1 = _mm_unpackhi_epi16(step1[23], step1[24]);
+        const __m128i out_05_2 = _mm_madd_epi16(out_05_0, k__cospi_p27_p05);
+        const __m128i out_05_3 = _mm_madd_epi16(out_05_1, k__cospi_p27_p05);
+        const __m128i out_21_2 = _mm_madd_epi16(out_21_0, k__cospi_p11_p21);
+        const __m128i out_21_3 = _mm_madd_epi16(out_21_1, k__cospi_p11_p21);
+        const __m128i out_13_2 = _mm_madd_epi16(out_13_0, k__cospi_p19_p13);
+        const __m128i out_13_3 = _mm_madd_epi16(out_13_1, k__cospi_p19_p13);
+        const __m128i out_29_2 = _mm_madd_epi16(out_29_0, k__cospi_p03_p29);
+        const __m128i out_29_3 = _mm_madd_epi16(out_29_1, k__cospi_p03_p29);
+        const __m128i out_03_2 = _mm_madd_epi16(out_29_0, k__cospi_m29_p03);
+        const __m128i out_03_3 = _mm_madd_epi16(out_29_1, k__cospi_m29_p03);
+        const __m128i out_19_2 = _mm_madd_epi16(out_13_0, k__cospi_m13_p19);
+        const __m128i out_19_3 = _mm_madd_epi16(out_13_1, k__cospi_m13_p19);
+        const __m128i out_11_2 = _mm_madd_epi16(out_21_0, k__cospi_m21_p11);
+        const __m128i out_11_3 = _mm_madd_epi16(out_21_1, k__cospi_m21_p11);
+        const __m128i out_27_2 = _mm_madd_epi16(out_05_0, k__cospi_m05_p27);
+        const __m128i out_27_3 = _mm_madd_epi16(out_05_1, k__cospi_m05_p27);
+        // dct_const_round_shift
+        const __m128i out_05_4 = _mm_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_05_5 = _mm_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_21_4 = _mm_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_21_5 = _mm_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_13_4 = _mm_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_13_5 = _mm_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_29_4 = _mm_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_29_5 = _mm_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_03_4 = _mm_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_03_5 = _mm_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_19_4 = _mm_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_19_5 = _mm_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_11_4 = _mm_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_11_5 = _mm_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_27_4 = _mm_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_27_5 = _mm_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_05_6 = _mm_srai_epi32(out_05_4, DCT_CONST_BITS);
+        const __m128i out_05_7 = _mm_srai_epi32(out_05_5, DCT_CONST_BITS);
+        const __m128i out_21_6 = _mm_srai_epi32(out_21_4, DCT_CONST_BITS);
+        const __m128i out_21_7 = _mm_srai_epi32(out_21_5, DCT_CONST_BITS);
+        const __m128i out_13_6 = _mm_srai_epi32(out_13_4, DCT_CONST_BITS);
+        const __m128i out_13_7 = _mm_srai_epi32(out_13_5, DCT_CONST_BITS);
+        const __m128i out_29_6 = _mm_srai_epi32(out_29_4, DCT_CONST_BITS);
+        const __m128i out_29_7 = _mm_srai_epi32(out_29_5, DCT_CONST_BITS);
+        const __m128i out_03_6 = _mm_srai_epi32(out_03_4, DCT_CONST_BITS);
+        const __m128i out_03_7 = _mm_srai_epi32(out_03_5, DCT_CONST_BITS);
+        const __m128i out_19_6 = _mm_srai_epi32(out_19_4, DCT_CONST_BITS);
+        const __m128i out_19_7 = _mm_srai_epi32(out_19_5, DCT_CONST_BITS);
+        const __m128i out_11_6 = _mm_srai_epi32(out_11_4, DCT_CONST_BITS);
+        const __m128i out_11_7 = _mm_srai_epi32(out_11_5, DCT_CONST_BITS);
+        const __m128i out_27_6 = _mm_srai_epi32(out_27_4, DCT_CONST_BITS);
+        const __m128i out_27_7 = _mm_srai_epi32(out_27_5, DCT_CONST_BITS);
+        // Combine
+        out[ 5] = _mm_packs_epi32(out_05_6, out_05_7);
+        out[21] = _mm_packs_epi32(out_21_6, out_21_7);
+        out[13] = _mm_packs_epi32(out_13_6, out_13_7);
+        out[29] = _mm_packs_epi32(out_29_6, out_29_7);
+        out[ 3] = _mm_packs_epi32(out_03_6, out_03_7);
+        out[19] = _mm_packs_epi32(out_19_6, out_19_7);
+        out[11] = _mm_packs_epi32(out_11_6, out_11_7);
+        out[27] = _mm_packs_epi32(out_27_6, out_27_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&out[5], &out[21], &out[13],
+                                           &out[29], &out[3], &out[19],
+                                           &out[11], &out[27]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+#if FDCT32x32_HIGH_PRECISION
+      } else {
+        __m128i lstep1[64], lstep2[64], lstep3[64];
+        __m128i u[32], v[32], sign[16];
+        const __m128i K32One = _mm_set_epi32(1, 1, 1, 1);
+        // start using 32-bit operations
+        // stage 3
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[ 0] = _mm_unpacklo_epi16(step2[ 0], kZero);
+          lstep2[ 1] = _mm_unpackhi_epi16(step2[ 0], kZero);
+          lstep2[ 2] = _mm_unpacklo_epi16(step2[ 1], kZero);
+          lstep2[ 3] = _mm_unpackhi_epi16(step2[ 1], kZero);
+          lstep2[ 4] = _mm_unpacklo_epi16(step2[ 2], kZero);
+          lstep2[ 5] = _mm_unpackhi_epi16(step2[ 2], kZero);
+          lstep2[ 6] = _mm_unpacklo_epi16(step2[ 3], kZero);
+          lstep2[ 7] = _mm_unpackhi_epi16(step2[ 3], kZero);
+          lstep2[ 8] = _mm_unpacklo_epi16(step2[ 4], kZero);
+          lstep2[ 9] = _mm_unpackhi_epi16(step2[ 4], kZero);
+          lstep2[10] = _mm_unpacklo_epi16(step2[ 5], kZero);
+          lstep2[11] = _mm_unpackhi_epi16(step2[ 5], kZero);
+          lstep2[12] = _mm_unpacklo_epi16(step2[ 6], kZero);
+          lstep2[13] = _mm_unpackhi_epi16(step2[ 6], kZero);
+          lstep2[14] = _mm_unpacklo_epi16(step2[ 7], kZero);
+          lstep2[15] = _mm_unpackhi_epi16(step2[ 7], kZero);
+          lstep2[ 0] = _mm_madd_epi16(lstep2[ 0], kOne);
+          lstep2[ 1] = _mm_madd_epi16(lstep2[ 1], kOne);
+          lstep2[ 2] = _mm_madd_epi16(lstep2[ 2], kOne);
+          lstep2[ 3] = _mm_madd_epi16(lstep2[ 3], kOne);
+          lstep2[ 4] = _mm_madd_epi16(lstep2[ 4], kOne);
+          lstep2[ 5] = _mm_madd_epi16(lstep2[ 5], kOne);
+          lstep2[ 6] = _mm_madd_epi16(lstep2[ 6], kOne);
+          lstep2[ 7] = _mm_madd_epi16(lstep2[ 7], kOne);
+          lstep2[ 8] = _mm_madd_epi16(lstep2[ 8], kOne);
+          lstep2[ 9] = _mm_madd_epi16(lstep2[ 9], kOne);
+          lstep2[10] = _mm_madd_epi16(lstep2[10], kOne);
+          lstep2[11] = _mm_madd_epi16(lstep2[11], kOne);
+          lstep2[12] = _mm_madd_epi16(lstep2[12], kOne);
+          lstep2[13] = _mm_madd_epi16(lstep2[13], kOne);
+          lstep2[14] = _mm_madd_epi16(lstep2[14], kOne);
+          lstep2[15] = _mm_madd_epi16(lstep2[15], kOne);
+
+          lstep3[ 0] = _mm_add_epi32(lstep2[14], lstep2[ 0]);
+          lstep3[ 1] = _mm_add_epi32(lstep2[15], lstep2[ 1]);
+          lstep3[ 2] = _mm_add_epi32(lstep2[12], lstep2[ 2]);
+          lstep3[ 3] = _mm_add_epi32(lstep2[13], lstep2[ 3]);
+          lstep3[ 4] = _mm_add_epi32(lstep2[10], lstep2[ 4]);
+          lstep3[ 5] = _mm_add_epi32(lstep2[11], lstep2[ 5]);
+          lstep3[ 6] = _mm_add_epi32(lstep2[ 8], lstep2[ 6]);
+          lstep3[ 7] = _mm_add_epi32(lstep2[ 9], lstep2[ 7]);
+          lstep3[ 8] = _mm_sub_epi32(lstep2[ 6], lstep2[ 8]);
+          lstep3[ 9] = _mm_sub_epi32(lstep2[ 7], lstep2[ 9]);
+          lstep3[10] = _mm_sub_epi32(lstep2[ 4], lstep2[10]);
+          lstep3[11] = _mm_sub_epi32(lstep2[ 5], lstep2[11]);
+          lstep3[12] = _mm_sub_epi32(lstep2[ 2], lstep2[12]);
+          lstep3[13] = _mm_sub_epi32(lstep2[ 3], lstep2[13]);
+          lstep3[14] = _mm_sub_epi32(lstep2[ 0], lstep2[14]);
+          lstep3[15] = _mm_sub_epi32(lstep2[ 1], lstep2[15]);
+        }
+        {
+          const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
+          const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
+          const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
+          const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
+          const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
+          const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
+          const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
+          const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
+          const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
+          const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
+          const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
+          const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
+          // dct_const_round_shift
+          const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+          lstep3[20] = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
+          lstep3[21] = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
+          lstep3[22] = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
+          lstep3[23] = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
+          lstep3[24] = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
+          lstep3[25] = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
+          lstep3[26] = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
+          lstep3[27] = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
+        }
+        {
+          lstep2[40] = _mm_unpacklo_epi16(step2[20], kZero);
+          lstep2[41] = _mm_unpackhi_epi16(step2[20], kZero);
+          lstep2[42] = _mm_unpacklo_epi16(step2[21], kZero);
+          lstep2[43] = _mm_unpackhi_epi16(step2[21], kZero);
+          lstep2[44] = _mm_unpacklo_epi16(step2[22], kZero);
+          lstep2[45] = _mm_unpackhi_epi16(step2[22], kZero);
+          lstep2[46] = _mm_unpacklo_epi16(step2[23], kZero);
+          lstep2[47] = _mm_unpackhi_epi16(step2[23], kZero);
+          lstep2[48] = _mm_unpacklo_epi16(step2[24], kZero);
+          lstep2[49] = _mm_unpackhi_epi16(step2[24], kZero);
+          lstep2[50] = _mm_unpacklo_epi16(step2[25], kZero);
+          lstep2[51] = _mm_unpackhi_epi16(step2[25], kZero);
+          lstep2[52] = _mm_unpacklo_epi16(step2[26], kZero);
+          lstep2[53] = _mm_unpackhi_epi16(step2[26], kZero);
+          lstep2[54] = _mm_unpacklo_epi16(step2[27], kZero);
+          lstep2[55] = _mm_unpackhi_epi16(step2[27], kZero);
+          lstep2[40] = _mm_madd_epi16(lstep2[40], kOne);
+          lstep2[41] = _mm_madd_epi16(lstep2[41], kOne);
+          lstep2[42] = _mm_madd_epi16(lstep2[42], kOne);
+          lstep2[43] = _mm_madd_epi16(lstep2[43], kOne);
+          lstep2[44] = _mm_madd_epi16(lstep2[44], kOne);
+          lstep2[45] = _mm_madd_epi16(lstep2[45], kOne);
+          lstep2[46] = _mm_madd_epi16(lstep2[46], kOne);
+          lstep2[47] = _mm_madd_epi16(lstep2[47], kOne);
+          lstep2[48] = _mm_madd_epi16(lstep2[48], kOne);
+          lstep2[49] = _mm_madd_epi16(lstep2[49], kOne);
+          lstep2[50] = _mm_madd_epi16(lstep2[50], kOne);
+          lstep2[51] = _mm_madd_epi16(lstep2[51], kOne);
+          lstep2[52] = _mm_madd_epi16(lstep2[52], kOne);
+          lstep2[53] = _mm_madd_epi16(lstep2[53], kOne);
+          lstep2[54] = _mm_madd_epi16(lstep2[54], kOne);
+          lstep2[55] = _mm_madd_epi16(lstep2[55], kOne);
+
+          lstep1[32] = _mm_unpacklo_epi16(step1[16], kZero);
+          lstep1[33] = _mm_unpackhi_epi16(step1[16], kZero);
+          lstep1[34] = _mm_unpacklo_epi16(step1[17], kZero);
+          lstep1[35] = _mm_unpackhi_epi16(step1[17], kZero);
+          lstep1[36] = _mm_unpacklo_epi16(step1[18], kZero);
+          lstep1[37] = _mm_unpackhi_epi16(step1[18], kZero);
+          lstep1[38] = _mm_unpacklo_epi16(step1[19], kZero);
+          lstep1[39] = _mm_unpackhi_epi16(step1[19], kZero);
+          lstep1[56] = _mm_unpacklo_epi16(step1[28], kZero);
+          lstep1[57] = _mm_unpackhi_epi16(step1[28], kZero);
+          lstep1[58] = _mm_unpacklo_epi16(step1[29], kZero);
+          lstep1[59] = _mm_unpackhi_epi16(step1[29], kZero);
+          lstep1[60] = _mm_unpacklo_epi16(step1[30], kZero);
+          lstep1[61] = _mm_unpackhi_epi16(step1[30], kZero);
+          lstep1[62] = _mm_unpacklo_epi16(step1[31], kZero);
+          lstep1[63] = _mm_unpackhi_epi16(step1[31], kZero);
+          lstep1[32] = _mm_madd_epi16(lstep1[32], kOne);
+          lstep1[33] = _mm_madd_epi16(lstep1[33], kOne);
+          lstep1[34] = _mm_madd_epi16(lstep1[34], kOne);
+          lstep1[35] = _mm_madd_epi16(lstep1[35], kOne);
+          lstep1[36] = _mm_madd_epi16(lstep1[36], kOne);
+          lstep1[37] = _mm_madd_epi16(lstep1[37], kOne);
+          lstep1[38] = _mm_madd_epi16(lstep1[38], kOne);
+          lstep1[39] = _mm_madd_epi16(lstep1[39], kOne);
+          lstep1[56] = _mm_madd_epi16(lstep1[56], kOne);
+          lstep1[57] = _mm_madd_epi16(lstep1[57], kOne);
+          lstep1[58] = _mm_madd_epi16(lstep1[58], kOne);
+          lstep1[59] = _mm_madd_epi16(lstep1[59], kOne);
+          lstep1[60] = _mm_madd_epi16(lstep1[60], kOne);
+          lstep1[61] = _mm_madd_epi16(lstep1[61], kOne);
+          lstep1[62] = _mm_madd_epi16(lstep1[62], kOne);
+          lstep1[63] = _mm_madd_epi16(lstep1[63], kOne);
+
+          lstep3[32] = _mm_add_epi32(lstep2[46], lstep1[32]);
+          lstep3[33] = _mm_add_epi32(lstep2[47], lstep1[33]);
+
+          lstep3[34] = _mm_add_epi32(lstep2[44], lstep1[34]);
+          lstep3[35] = _mm_add_epi32(lstep2[45], lstep1[35]);
+          lstep3[36] = _mm_add_epi32(lstep2[42], lstep1[36]);
+          lstep3[37] = _mm_add_epi32(lstep2[43], lstep1[37]);
+          lstep3[38] = _mm_add_epi32(lstep2[40], lstep1[38]);
+          lstep3[39] = _mm_add_epi32(lstep2[41], lstep1[39]);
+          lstep3[40] = _mm_sub_epi32(lstep1[38], lstep2[40]);
+          lstep3[41] = _mm_sub_epi32(lstep1[39], lstep2[41]);
+          lstep3[42] = _mm_sub_epi32(lstep1[36], lstep2[42]);
+          lstep3[43] = _mm_sub_epi32(lstep1[37], lstep2[43]);
+          lstep3[44] = _mm_sub_epi32(lstep1[34], lstep2[44]);
+          lstep3[45] = _mm_sub_epi32(lstep1[35], lstep2[45]);
+          lstep3[46] = _mm_sub_epi32(lstep1[32], lstep2[46]);
+          lstep3[47] = _mm_sub_epi32(lstep1[33], lstep2[47]);
+          lstep3[48] = _mm_sub_epi32(lstep1[62], lstep2[48]);
+          lstep3[49] = _mm_sub_epi32(lstep1[63], lstep2[49]);
+          lstep3[50] = _mm_sub_epi32(lstep1[60], lstep2[50]);
+          lstep3[51] = _mm_sub_epi32(lstep1[61], lstep2[51]);
+          lstep3[52] = _mm_sub_epi32(lstep1[58], lstep2[52]);
+          lstep3[53] = _mm_sub_epi32(lstep1[59], lstep2[53]);
+          lstep3[54] = _mm_sub_epi32(lstep1[56], lstep2[54]);
+          lstep3[55] = _mm_sub_epi32(lstep1[57], lstep2[55]);
+          lstep3[56] = _mm_add_epi32(lstep2[54], lstep1[56]);
+          lstep3[57] = _mm_add_epi32(lstep2[55], lstep1[57]);
+          lstep3[58] = _mm_add_epi32(lstep2[52], lstep1[58]);
+          lstep3[59] = _mm_add_epi32(lstep2[53], lstep1[59]);
+          lstep3[60] = _mm_add_epi32(lstep2[50], lstep1[60]);
+          lstep3[61] = _mm_add_epi32(lstep2[51], lstep1[61]);
+          lstep3[62] = _mm_add_epi32(lstep2[48], lstep1[62]);
+          lstep3[63] = _mm_add_epi32(lstep2[49], lstep1[63]);
+        }
+
+        // stage 4
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[16] = _mm_unpacklo_epi16(step2[ 8], kZero);
+          lstep2[17] = _mm_unpackhi_epi16(step2[ 8], kZero);
+          lstep2[18] = _mm_unpacklo_epi16(step2[ 9], kZero);
+          lstep2[19] = _mm_unpackhi_epi16(step2[ 9], kZero);
+          lstep2[28] = _mm_unpacklo_epi16(step2[14], kZero);
+          lstep2[29] = _mm_unpackhi_epi16(step2[14], kZero);
+          lstep2[30] = _mm_unpacklo_epi16(step2[15], kZero);
+          lstep2[31] = _mm_unpackhi_epi16(step2[15], kZero);
+          lstep2[16] = _mm_madd_epi16(lstep2[16], kOne);
+          lstep2[17] = _mm_madd_epi16(lstep2[17], kOne);
+          lstep2[18] = _mm_madd_epi16(lstep2[18], kOne);
+          lstep2[19] = _mm_madd_epi16(lstep2[19], kOne);
+          lstep2[28] = _mm_madd_epi16(lstep2[28], kOne);
+          lstep2[29] = _mm_madd_epi16(lstep2[29], kOne);
+          lstep2[30] = _mm_madd_epi16(lstep2[30], kOne);
+          lstep2[31] = _mm_madd_epi16(lstep2[31], kOne);
+
+          lstep1[ 0] = _mm_add_epi32(lstep3[ 6], lstep3[ 0]);
+          lstep1[ 1] = _mm_add_epi32(lstep3[ 7], lstep3[ 1]);
+          lstep1[ 2] = _mm_add_epi32(lstep3[ 4], lstep3[ 2]);
+          lstep1[ 3] = _mm_add_epi32(lstep3[ 5], lstep3[ 3]);
+          lstep1[ 4] = _mm_sub_epi32(lstep3[ 2], lstep3[ 4]);
+          lstep1[ 5] = _mm_sub_epi32(lstep3[ 3], lstep3[ 5]);
+          lstep1[ 6] = _mm_sub_epi32(lstep3[ 0], lstep3[ 6]);
+          lstep1[ 7] = _mm_sub_epi32(lstep3[ 1], lstep3[ 7]);
+          lstep1[16] = _mm_add_epi32(lstep3[22], lstep2[16]);
+          lstep1[17] = _mm_add_epi32(lstep3[23], lstep2[17]);
+          lstep1[18] = _mm_add_epi32(lstep3[20], lstep2[18]);
+          lstep1[19] = _mm_add_epi32(lstep3[21], lstep2[19]);
+          lstep1[20] = _mm_sub_epi32(lstep2[18], lstep3[20]);
+          lstep1[21] = _mm_sub_epi32(lstep2[19], lstep3[21]);
+          lstep1[22] = _mm_sub_epi32(lstep2[16], lstep3[22]);
+          lstep1[23] = _mm_sub_epi32(lstep2[17], lstep3[23]);
+          lstep1[24] = _mm_sub_epi32(lstep2[30], lstep3[24]);
+          lstep1[25] = _mm_sub_epi32(lstep2[31], lstep3[25]);
+          lstep1[26] = _mm_sub_epi32(lstep2[28], lstep3[26]);
+          lstep1[27] = _mm_sub_epi32(lstep2[29], lstep3[27]);
+          lstep1[28] = _mm_add_epi32(lstep3[26], lstep2[28]);
+          lstep1[29] = _mm_add_epi32(lstep3[27], lstep2[29]);
+          lstep1[30] = _mm_add_epi32(lstep3[24], lstep2[30]);
+          lstep1[31] = _mm_add_epi32(lstep3[25], lstep2[31]);
+        }
+        {
+        // to be continued...
+        //
+        const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
+        const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
+
+        u[0] = _mm_unpacklo_epi32(lstep3[12], lstep3[10]);
+        u[1] = _mm_unpackhi_epi32(lstep3[12], lstep3[10]);
+        u[2] = _mm_unpacklo_epi32(lstep3[13], lstep3[11]);
+        u[3] = _mm_unpackhi_epi32(lstep3[13], lstep3[11]);
+
+        // TODO(jingning): manually inline k_madd_epi32_ to further hide
+        // instruction latency.
+        v[0] = k_madd_epi32(u[0], k32_p16_m16);
+        v[1] = k_madd_epi32(u[1], k32_p16_m16);
+        v[2] = k_madd_epi32(u[2], k32_p16_m16);
+        v[3] = k_madd_epi32(u[3], k32_p16_m16);
+        v[4] = k_madd_epi32(u[0], k32_p16_p16);
+        v[5] = k_madd_epi32(u[1], k32_p16_p16);
+        v[6] = k_madd_epi32(u[2], k32_p16_p16);
+        v[7] = k_madd_epi32(u[3], k32_p16_p16);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = k_check_epi32_overflow_8(&v[0], &v[1], &v[2], &v[3],
+                                            &v[4], &v[5], &v[6], &v[7], &kZero);
+        if (overflow) {
+          HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        u[0] = k_packs_epi64(v[0], v[1]);
+        u[1] = k_packs_epi64(v[2], v[3]);
+        u[2] = k_packs_epi64(v[4], v[5]);
+        u[3] = k_packs_epi64(v[6], v[7]);
+
+        v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+        v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+        v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+        v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+
+        lstep1[10] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+        lstep1[11] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+        lstep1[12] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+        lstep1[13] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+        }
+        {
+          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[ 0] = _mm_unpacklo_epi32(lstep3[36], lstep3[58]);
+          u[ 1] = _mm_unpackhi_epi32(lstep3[36], lstep3[58]);
+          u[ 2] = _mm_unpacklo_epi32(lstep3[37], lstep3[59]);
+          u[ 3] = _mm_unpackhi_epi32(lstep3[37], lstep3[59]);
+          u[ 4] = _mm_unpacklo_epi32(lstep3[38], lstep3[56]);
+          u[ 5] = _mm_unpackhi_epi32(lstep3[38], lstep3[56]);
+          u[ 6] = _mm_unpacklo_epi32(lstep3[39], lstep3[57]);
+          u[ 7] = _mm_unpackhi_epi32(lstep3[39], lstep3[57]);
+          u[ 8] = _mm_unpacklo_epi32(lstep3[40], lstep3[54]);
+          u[ 9] = _mm_unpackhi_epi32(lstep3[40], lstep3[54]);
+          u[10] = _mm_unpacklo_epi32(lstep3[41], lstep3[55]);
+          u[11] = _mm_unpackhi_epi32(lstep3[41], lstep3[55]);
+          u[12] = _mm_unpacklo_epi32(lstep3[42], lstep3[52]);
+          u[13] = _mm_unpackhi_epi32(lstep3[42], lstep3[52]);
+          u[14] = _mm_unpacklo_epi32(lstep3[43], lstep3[53]);
+          u[15] = _mm_unpackhi_epi32(lstep3[43], lstep3[53]);
+
+          v[ 0] = k_madd_epi32(u[ 0], k32_m08_p24);
+          v[ 1] = k_madd_epi32(u[ 1], k32_m08_p24);
+          v[ 2] = k_madd_epi32(u[ 2], k32_m08_p24);
+          v[ 3] = k_madd_epi32(u[ 3], k32_m08_p24);
+          v[ 4] = k_madd_epi32(u[ 4], k32_m08_p24);
+          v[ 5] = k_madd_epi32(u[ 5], k32_m08_p24);
+          v[ 6] = k_madd_epi32(u[ 6], k32_m08_p24);
+          v[ 7] = k_madd_epi32(u[ 7], k32_m08_p24);
+          v[ 8] = k_madd_epi32(u[ 8], k32_m24_m08);
+          v[ 9] = k_madd_epi32(u[ 9], k32_m24_m08);
+          v[10] = k_madd_epi32(u[10], k32_m24_m08);
+          v[11] = k_madd_epi32(u[11], k32_m24_m08);
+          v[12] = k_madd_epi32(u[12], k32_m24_m08);
+          v[13] = k_madd_epi32(u[13], k32_m24_m08);
+          v[14] = k_madd_epi32(u[14], k32_m24_m08);
+          v[15] = k_madd_epi32(u[15], k32_m24_m08);
+          v[16] = k_madd_epi32(u[12], k32_m08_p24);
+          v[17] = k_madd_epi32(u[13], k32_m08_p24);
+          v[18] = k_madd_epi32(u[14], k32_m08_p24);
+          v[19] = k_madd_epi32(u[15], k32_m08_p24);
+          v[20] = k_madd_epi32(u[ 8], k32_m08_p24);
+          v[21] = k_madd_epi32(u[ 9], k32_m08_p24);
+          v[22] = k_madd_epi32(u[10], k32_m08_p24);
+          v[23] = k_madd_epi32(u[11], k32_m08_p24);
+          v[24] = k_madd_epi32(u[ 4], k32_p24_p08);
+          v[25] = k_madd_epi32(u[ 5], k32_p24_p08);
+          v[26] = k_madd_epi32(u[ 6], k32_p24_p08);
+          v[27] = k_madd_epi32(u[ 7], k32_p24_p08);
+          v[28] = k_madd_epi32(u[ 0], k32_p24_p08);
+          v[29] = k_madd_epi32(u[ 1], k32_p24_p08);
+          v[30] = k_madd_epi32(u[ 2], k32_p24_p08);
+          v[31] = k_madd_epi32(u[ 3], k32_p24_p08);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &v[16], &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23],
+              &v[24], &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64(v[10], v[11]);
+          u[ 6] = k_packs_epi64(v[12], v[13]);
+          u[ 7] = k_packs_epi64(v[14], v[15]);
+          u[ 8] = k_packs_epi64(v[16], v[17]);
+          u[ 9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep1[36] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
+          lstep1[37] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
+          lstep1[38] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
+          lstep1[39] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
+          lstep1[40] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
+          lstep1[41] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
+          lstep1[42] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
+          lstep1[43] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
+          lstep1[52] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
+          lstep1[53] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
+          lstep1[54] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep1[55] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep1[56] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep1[57] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep1[58] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep1[59] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 5
+        {
+          lstep2[ 8] = _mm_add_epi32(lstep1[10], lstep3[ 8]);
+          lstep2[ 9] = _mm_add_epi32(lstep1[11], lstep3[ 9]);
+          lstep2[10] = _mm_sub_epi32(lstep3[ 8], lstep1[10]);
+          lstep2[11] = _mm_sub_epi32(lstep3[ 9], lstep1[11]);
+          lstep2[12] = _mm_sub_epi32(lstep3[14], lstep1[12]);
+          lstep2[13] = _mm_sub_epi32(lstep3[15], lstep1[13]);
+          lstep2[14] = _mm_add_epi32(lstep1[12], lstep3[14]);
+          lstep2[15] = _mm_add_epi32(lstep1[13], lstep3[15]);
+        }
+        {
+          const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
+          const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
+          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
+          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep1[0], lstep1[2]);
+          u[1] = _mm_unpackhi_epi32(lstep1[0], lstep1[2]);
+          u[2] = _mm_unpacklo_epi32(lstep1[1], lstep1[3]);
+          u[3] = _mm_unpackhi_epi32(lstep1[1], lstep1[3]);
+          u[4] = _mm_unpacklo_epi32(lstep1[4], lstep1[6]);
+          u[5] = _mm_unpackhi_epi32(lstep1[4], lstep1[6]);
+          u[6] = _mm_unpacklo_epi32(lstep1[5], lstep1[7]);
+          u[7] = _mm_unpackhi_epi32(lstep1[5], lstep1[7]);
+
+          // TODO(jingning): manually inline k_madd_epi32_ to further hide
+          // instruction latency.
+          v[ 0] = k_madd_epi32(u[0], k32_p16_p16);
+          v[ 1] = k_madd_epi32(u[1], k32_p16_p16);
+          v[ 2] = k_madd_epi32(u[2], k32_p16_p16);
+          v[ 3] = k_madd_epi32(u[3], k32_p16_p16);
+          v[ 4] = k_madd_epi32(u[0], k32_p16_m16);
+          v[ 5] = k_madd_epi32(u[1], k32_p16_m16);
+          v[ 6] = k_madd_epi32(u[2], k32_p16_m16);
+          v[ 7] = k_madd_epi32(u[3], k32_p16_m16);
+          v[ 8] = k_madd_epi32(u[4], k32_p24_p08);
+          v[ 9] = k_madd_epi32(u[5], k32_p24_p08);
+          v[10] = k_madd_epi32(u[6], k32_p24_p08);
+          v[11] = k_madd_epi32(u[7], k32_p24_p08);
+          v[12] = k_madd_epi32(u[4], k32_m08_p24);
+          v[13] = k_madd_epi32(u[5], k32_m08_p24);
+          v[14] = k_madd_epi32(u[6], k32_m08_p24);
+          v[15] = k_madd_epi32(u[7], k32_m08_p24);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_16(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm_cmplt_epi32(u[0], kZero);
+          sign[1] = _mm_cmplt_epi32(u[1], kZero);
+          sign[2] = _mm_cmplt_epi32(u[2], kZero);
+          sign[3] = _mm_cmplt_epi32(u[3], kZero);
+          sign[4] = _mm_cmplt_epi32(u[4], kZero);
+          sign[5] = _mm_cmplt_epi32(u[5], kZero);
+          sign[6] = _mm_cmplt_epi32(u[6], kZero);
+          sign[7] = _mm_cmplt_epi32(u[7], kZero);
+
+          u[0] = _mm_sub_epi32(u[0], sign[0]);
+          u[1] = _mm_sub_epi32(u[1], sign[1]);
+          u[2] = _mm_sub_epi32(u[2], sign[2]);
+          u[3] = _mm_sub_epi32(u[3], sign[3]);
+          u[4] = _mm_sub_epi32(u[4], sign[4]);
+          u[5] = _mm_sub_epi32(u[5], sign[5]);
+          u[6] = _mm_sub_epi32(u[6], sign[6]);
+          u[7] = _mm_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm_add_epi32(u[0], K32One);
+          u[1] = _mm_add_epi32(u[1], K32One);
+          u[2] = _mm_add_epi32(u[2], K32One);
+          u[3] = _mm_add_epi32(u[3], K32One);
+          u[4] = _mm_add_epi32(u[4], K32One);
+          u[5] = _mm_add_epi32(u[5], K32One);
+          u[6] = _mm_add_epi32(u[6], K32One);
+          u[7] = _mm_add_epi32(u[7], K32One);
+
+          u[0] = _mm_srai_epi32(u[0], 2);
+          u[1] = _mm_srai_epi32(u[1], 2);
+          u[2] = _mm_srai_epi32(u[2], 2);
+          u[3] = _mm_srai_epi32(u[3], 2);
+          u[4] = _mm_srai_epi32(u[4], 2);
+          u[5] = _mm_srai_epi32(u[5], 2);
+          u[6] = _mm_srai_epi32(u[6], 2);
+          u[7] = _mm_srai_epi32(u[7], 2);
+
+          // Combine
+          out[ 0] = _mm_packs_epi32(u[0], u[1]);
+          out[16] = _mm_packs_epi32(u[2], u[3]);
+          out[ 8] = _mm_packs_epi32(u[4], u[5]);
+          out[24] = _mm_packs_epi32(u[6], u[7]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&out[0], &out[16],
+                                             &out[8], &out[24]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep1[18], lstep1[28]);
+          u[1] = _mm_unpackhi_epi32(lstep1[18], lstep1[28]);
+          u[2] = _mm_unpacklo_epi32(lstep1[19], lstep1[29]);
+          u[3] = _mm_unpackhi_epi32(lstep1[19], lstep1[29]);
+          u[4] = _mm_unpacklo_epi32(lstep1[20], lstep1[26]);
+          u[5] = _mm_unpackhi_epi32(lstep1[20], lstep1[26]);
+          u[6] = _mm_unpacklo_epi32(lstep1[21], lstep1[27]);
+          u[7] = _mm_unpackhi_epi32(lstep1[21], lstep1[27]);
+
+          v[0] = k_madd_epi32(u[0], k32_m08_p24);
+          v[1] = k_madd_epi32(u[1], k32_m08_p24);
+          v[2] = k_madd_epi32(u[2], k32_m08_p24);
+          v[3] = k_madd_epi32(u[3], k32_m08_p24);
+          v[4] = k_madd_epi32(u[4], k32_m24_m08);
+          v[5] = k_madd_epi32(u[5], k32_m24_m08);
+          v[6] = k_madd_epi32(u[6], k32_m24_m08);
+          v[7] = k_madd_epi32(u[7], k32_m24_m08);
+          v[ 8] = k_madd_epi32(u[4], k32_m08_p24);
+          v[ 9] = k_madd_epi32(u[5], k32_m08_p24);
+          v[10] = k_madd_epi32(u[6], k32_m08_p24);
+          v[11] = k_madd_epi32(u[7], k32_m08_p24);
+          v[12] = k_madd_epi32(u[0], k32_p24_p08);
+          v[13] = k_madd_epi32(u[1], k32_p24_p08);
+          v[14] = k_madd_epi32(u[2], k32_p24_p08);
+          v[15] = k_madd_epi32(u[3], k32_p24_p08);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_16(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+
+          u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          lstep2[18] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+          lstep2[19] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+          lstep2[20] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+          lstep2[21] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+          lstep2[26] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+          lstep2[27] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+          lstep2[28] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+          lstep2[29] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+        }
+        {
+          lstep2[32] = _mm_add_epi32(lstep1[38], lstep3[32]);
+          lstep2[33] = _mm_add_epi32(lstep1[39], lstep3[33]);
+          lstep2[34] = _mm_add_epi32(lstep1[36], lstep3[34]);
+          lstep2[35] = _mm_add_epi32(lstep1[37], lstep3[35]);
+          lstep2[36] = _mm_sub_epi32(lstep3[34], lstep1[36]);
+          lstep2[37] = _mm_sub_epi32(lstep3[35], lstep1[37]);
+          lstep2[38] = _mm_sub_epi32(lstep3[32], lstep1[38]);
+          lstep2[39] = _mm_sub_epi32(lstep3[33], lstep1[39]);
+          lstep2[40] = _mm_sub_epi32(lstep3[46], lstep1[40]);
+          lstep2[41] = _mm_sub_epi32(lstep3[47], lstep1[41]);
+          lstep2[42] = _mm_sub_epi32(lstep3[44], lstep1[42]);
+          lstep2[43] = _mm_sub_epi32(lstep3[45], lstep1[43]);
+          lstep2[44] = _mm_add_epi32(lstep1[42], lstep3[44]);
+          lstep2[45] = _mm_add_epi32(lstep1[43], lstep3[45]);
+          lstep2[46] = _mm_add_epi32(lstep1[40], lstep3[46]);
+          lstep2[47] = _mm_add_epi32(lstep1[41], lstep3[47]);
+          lstep2[48] = _mm_add_epi32(lstep1[54], lstep3[48]);
+          lstep2[49] = _mm_add_epi32(lstep1[55], lstep3[49]);
+          lstep2[50] = _mm_add_epi32(lstep1[52], lstep3[50]);
+          lstep2[51] = _mm_add_epi32(lstep1[53], lstep3[51]);
+          lstep2[52] = _mm_sub_epi32(lstep3[50], lstep1[52]);
+          lstep2[53] = _mm_sub_epi32(lstep3[51], lstep1[53]);
+          lstep2[54] = _mm_sub_epi32(lstep3[48], lstep1[54]);
+          lstep2[55] = _mm_sub_epi32(lstep3[49], lstep1[55]);
+          lstep2[56] = _mm_sub_epi32(lstep3[62], lstep1[56]);
+          lstep2[57] = _mm_sub_epi32(lstep3[63], lstep1[57]);
+          lstep2[58] = _mm_sub_epi32(lstep3[60], lstep1[58]);
+          lstep2[59] = _mm_sub_epi32(lstep3[61], lstep1[59]);
+          lstep2[60] = _mm_add_epi32(lstep1[58], lstep3[60]);
+          lstep2[61] = _mm_add_epi32(lstep1[59], lstep3[61]);
+          lstep2[62] = _mm_add_epi32(lstep1[56], lstep3[62]);
+          lstep2[63] = _mm_add_epi32(lstep1[57], lstep3[63]);
+        }
+        // stage 6
+        {
+          const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
+          const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
+          const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep2[ 8], lstep2[14]);
+          u[1] = _mm_unpackhi_epi32(lstep2[ 8], lstep2[14]);
+          u[2] = _mm_unpacklo_epi32(lstep2[ 9], lstep2[15]);
+          u[3] = _mm_unpackhi_epi32(lstep2[ 9], lstep2[15]);
+          u[4] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[5] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[6] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[7] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[8] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[9] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[10] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[11] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[12] = _mm_unpacklo_epi32(lstep2[ 8], lstep2[14]);
+          u[13] = _mm_unpackhi_epi32(lstep2[ 8], lstep2[14]);
+          u[14] = _mm_unpacklo_epi32(lstep2[ 9], lstep2[15]);
+          u[15] = _mm_unpackhi_epi32(lstep2[ 9], lstep2[15]);
+
+          v[0] = k_madd_epi32(u[0], k32_p28_p04);
+          v[1] = k_madd_epi32(u[1], k32_p28_p04);
+          v[2] = k_madd_epi32(u[2], k32_p28_p04);
+          v[3] = k_madd_epi32(u[3], k32_p28_p04);
+          v[4] = k_madd_epi32(u[4], k32_p12_p20);
+          v[5] = k_madd_epi32(u[5], k32_p12_p20);
+          v[6] = k_madd_epi32(u[6], k32_p12_p20);
+          v[7] = k_madd_epi32(u[7], k32_p12_p20);
+          v[ 8] = k_madd_epi32(u[ 8], k32_m20_p12);
+          v[ 9] = k_madd_epi32(u[ 9], k32_m20_p12);
+          v[10] = k_madd_epi32(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32(u[12], k32_m04_p28);
+          v[13] = k_madd_epi32(u[13], k32_m04_p28);
+          v[14] = k_madd_epi32(u[14], k32_m04_p28);
+          v[15] = k_madd_epi32(u[15], k32_m04_p28);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_16(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm_cmplt_epi32(u[0], kZero);
+          sign[1] = _mm_cmplt_epi32(u[1], kZero);
+          sign[2] = _mm_cmplt_epi32(u[2], kZero);
+          sign[3] = _mm_cmplt_epi32(u[3], kZero);
+          sign[4] = _mm_cmplt_epi32(u[4], kZero);
+          sign[5] = _mm_cmplt_epi32(u[5], kZero);
+          sign[6] = _mm_cmplt_epi32(u[6], kZero);
+          sign[7] = _mm_cmplt_epi32(u[7], kZero);
+
+          u[0] = _mm_sub_epi32(u[0], sign[0]);
+          u[1] = _mm_sub_epi32(u[1], sign[1]);
+          u[2] = _mm_sub_epi32(u[2], sign[2]);
+          u[3] = _mm_sub_epi32(u[3], sign[3]);
+          u[4] = _mm_sub_epi32(u[4], sign[4]);
+          u[5] = _mm_sub_epi32(u[5], sign[5]);
+          u[6] = _mm_sub_epi32(u[6], sign[6]);
+          u[7] = _mm_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm_add_epi32(u[0], K32One);
+          u[1] = _mm_add_epi32(u[1], K32One);
+          u[2] = _mm_add_epi32(u[2], K32One);
+          u[3] = _mm_add_epi32(u[3], K32One);
+          u[4] = _mm_add_epi32(u[4], K32One);
+          u[5] = _mm_add_epi32(u[5], K32One);
+          u[6] = _mm_add_epi32(u[6], K32One);
+          u[7] = _mm_add_epi32(u[7], K32One);
+
+          u[0] = _mm_srai_epi32(u[0], 2);
+          u[1] = _mm_srai_epi32(u[1], 2);
+          u[2] = _mm_srai_epi32(u[2], 2);
+          u[3] = _mm_srai_epi32(u[3], 2);
+          u[4] = _mm_srai_epi32(u[4], 2);
+          u[5] = _mm_srai_epi32(u[5], 2);
+          u[6] = _mm_srai_epi32(u[6], 2);
+          u[7] = _mm_srai_epi32(u[7], 2);
+
+          out[ 4] = _mm_packs_epi32(u[0], u[1]);
+          out[20] = _mm_packs_epi32(u[2], u[3]);
+          out[12] = _mm_packs_epi32(u[4], u[5]);
+          out[28] = _mm_packs_epi32(u[6], u[7]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&out[4], &out[20],
+                                             &out[12], &out[28]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          lstep3[16] = _mm_add_epi32(lstep2[18], lstep1[16]);
+          lstep3[17] = _mm_add_epi32(lstep2[19], lstep1[17]);
+          lstep3[18] = _mm_sub_epi32(lstep1[16], lstep2[18]);
+          lstep3[19] = _mm_sub_epi32(lstep1[17], lstep2[19]);
+          lstep3[20] = _mm_sub_epi32(lstep1[22], lstep2[20]);
+          lstep3[21] = _mm_sub_epi32(lstep1[23], lstep2[21]);
+          lstep3[22] = _mm_add_epi32(lstep2[20], lstep1[22]);
+          lstep3[23] = _mm_add_epi32(lstep2[21], lstep1[23]);
+          lstep3[24] = _mm_add_epi32(lstep2[26], lstep1[24]);
+          lstep3[25] = _mm_add_epi32(lstep2[27], lstep1[25]);
+          lstep3[26] = _mm_sub_epi32(lstep1[24], lstep2[26]);
+          lstep3[27] = _mm_sub_epi32(lstep1[25], lstep2[27]);
+          lstep3[28] = _mm_sub_epi32(lstep1[30], lstep2[28]);
+          lstep3[29] = _mm_sub_epi32(lstep1[31], lstep2[29]);
+          lstep3[30] = _mm_add_epi32(lstep2[28], lstep1[30]);
+          lstep3[31] = _mm_add_epi32(lstep2[29], lstep1[31]);
+        }
+        {
+          const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
+          const __m128i k32_m28_m04 = pair_set_epi32(-cospi_28_64, -cospi_4_64);
+          const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m128i k32_m12_m20 = pair_set_epi32(-cospi_12_64,
+                                                     -cospi_20_64);
+          const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
+          const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
+
+          u[ 0] = _mm_unpacklo_epi32(lstep2[34], lstep2[60]);
+          u[ 1] = _mm_unpackhi_epi32(lstep2[34], lstep2[60]);
+          u[ 2] = _mm_unpacklo_epi32(lstep2[35], lstep2[61]);
+          u[ 3] = _mm_unpackhi_epi32(lstep2[35], lstep2[61]);
+          u[ 4] = _mm_unpacklo_epi32(lstep2[36], lstep2[58]);
+          u[ 5] = _mm_unpackhi_epi32(lstep2[36], lstep2[58]);
+          u[ 6] = _mm_unpacklo_epi32(lstep2[37], lstep2[59]);
+          u[ 7] = _mm_unpackhi_epi32(lstep2[37], lstep2[59]);
+          u[ 8] = _mm_unpacklo_epi32(lstep2[42], lstep2[52]);
+          u[ 9] = _mm_unpackhi_epi32(lstep2[42], lstep2[52]);
+          u[10] = _mm_unpacklo_epi32(lstep2[43], lstep2[53]);
+          u[11] = _mm_unpackhi_epi32(lstep2[43], lstep2[53]);
+          u[12] = _mm_unpacklo_epi32(lstep2[44], lstep2[50]);
+          u[13] = _mm_unpackhi_epi32(lstep2[44], lstep2[50]);
+          u[14] = _mm_unpacklo_epi32(lstep2[45], lstep2[51]);
+          u[15] = _mm_unpackhi_epi32(lstep2[45], lstep2[51]);
+
+          v[ 0] = k_madd_epi32(u[ 0], k32_m04_p28);
+          v[ 1] = k_madd_epi32(u[ 1], k32_m04_p28);
+          v[ 2] = k_madd_epi32(u[ 2], k32_m04_p28);
+          v[ 3] = k_madd_epi32(u[ 3], k32_m04_p28);
+          v[ 4] = k_madd_epi32(u[ 4], k32_m28_m04);
+          v[ 5] = k_madd_epi32(u[ 5], k32_m28_m04);
+          v[ 6] = k_madd_epi32(u[ 6], k32_m28_m04);
+          v[ 7] = k_madd_epi32(u[ 7], k32_m28_m04);
+          v[ 8] = k_madd_epi32(u[ 8], k32_m20_p12);
+          v[ 9] = k_madd_epi32(u[ 9], k32_m20_p12);
+          v[10] = k_madd_epi32(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32(u[12], k32_m12_m20);
+          v[13] = k_madd_epi32(u[13], k32_m12_m20);
+          v[14] = k_madd_epi32(u[14], k32_m12_m20);
+          v[15] = k_madd_epi32(u[15], k32_m12_m20);
+          v[16] = k_madd_epi32(u[12], k32_m20_p12);
+          v[17] = k_madd_epi32(u[13], k32_m20_p12);
+          v[18] = k_madd_epi32(u[14], k32_m20_p12);
+          v[19] = k_madd_epi32(u[15], k32_m20_p12);
+          v[20] = k_madd_epi32(u[ 8], k32_p12_p20);
+          v[21] = k_madd_epi32(u[ 9], k32_p12_p20);
+          v[22] = k_madd_epi32(u[10], k32_p12_p20);
+          v[23] = k_madd_epi32(u[11], k32_p12_p20);
+          v[24] = k_madd_epi32(u[ 4], k32_m04_p28);
+          v[25] = k_madd_epi32(u[ 5], k32_m04_p28);
+          v[26] = k_madd_epi32(u[ 6], k32_m04_p28);
+          v[27] = k_madd_epi32(u[ 7], k32_m04_p28);
+          v[28] = k_madd_epi32(u[ 0], k32_p28_p04);
+          v[29] = k_madd_epi32(u[ 1], k32_p28_p04);
+          v[30] = k_madd_epi32(u[ 2], k32_p28_p04);
+          v[31] = k_madd_epi32(u[ 3], k32_p28_p04);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &v[16], &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23],
+              &v[24], &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64(v[10], v[11]);
+          u[ 6] = k_packs_epi64(v[12], v[13]);
+          u[ 7] = k_packs_epi64(v[14], v[15]);
+          u[ 8] = k_packs_epi64(v[16], v[17]);
+          u[ 9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep3[34] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
+          lstep3[35] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
+          lstep3[36] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
+          lstep3[37] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
+          lstep3[42] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
+          lstep3[43] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
+          lstep3[44] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
+          lstep3[45] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
+          lstep3[50] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
+          lstep3[51] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
+          lstep3[52] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep3[53] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep3[58] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep3[59] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep3[60] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep3[61] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 7
+        {
+          const __m128i k32_p30_p02 = pair_set_epi32(cospi_30_64, cospi_2_64);
+          const __m128i k32_p14_p18 = pair_set_epi32(cospi_14_64, cospi_18_64);
+          const __m128i k32_p22_p10 = pair_set_epi32(cospi_22_64, cospi_10_64);
+          const __m128i k32_p06_p26 = pair_set_epi32(cospi_6_64,  cospi_26_64);
+          const __m128i k32_m26_p06 = pair_set_epi32(-cospi_26_64, cospi_6_64);
+          const __m128i k32_m10_p22 = pair_set_epi32(-cospi_10_64, cospi_22_64);
+          const __m128i k32_m18_p14 = pair_set_epi32(-cospi_18_64, cospi_14_64);
+          const __m128i k32_m02_p30 = pair_set_epi32(-cospi_2_64, cospi_30_64);
+
+          u[ 0] = _mm_unpacklo_epi32(lstep3[16], lstep3[30]);
+          u[ 1] = _mm_unpackhi_epi32(lstep3[16], lstep3[30]);
+          u[ 2] = _mm_unpacklo_epi32(lstep3[17], lstep3[31]);
+          u[ 3] = _mm_unpackhi_epi32(lstep3[17], lstep3[31]);
+          u[ 4] = _mm_unpacklo_epi32(lstep3[18], lstep3[28]);
+          u[ 5] = _mm_unpackhi_epi32(lstep3[18], lstep3[28]);
+          u[ 6] = _mm_unpacklo_epi32(lstep3[19], lstep3[29]);
+          u[ 7] = _mm_unpackhi_epi32(lstep3[19], lstep3[29]);
+          u[ 8] = _mm_unpacklo_epi32(lstep3[20], lstep3[26]);
+          u[ 9] = _mm_unpackhi_epi32(lstep3[20], lstep3[26]);
+          u[10] = _mm_unpacklo_epi32(lstep3[21], lstep3[27]);
+          u[11] = _mm_unpackhi_epi32(lstep3[21], lstep3[27]);
+          u[12] = _mm_unpacklo_epi32(lstep3[22], lstep3[24]);
+          u[13] = _mm_unpackhi_epi32(lstep3[22], lstep3[24]);
+          u[14] = _mm_unpacklo_epi32(lstep3[23], lstep3[25]);
+          u[15] = _mm_unpackhi_epi32(lstep3[23], lstep3[25]);
+
+          v[ 0] = k_madd_epi32(u[ 0], k32_p30_p02);
+          v[ 1] = k_madd_epi32(u[ 1], k32_p30_p02);
+          v[ 2] = k_madd_epi32(u[ 2], k32_p30_p02);
+          v[ 3] = k_madd_epi32(u[ 3], k32_p30_p02);
+          v[ 4] = k_madd_epi32(u[ 4], k32_p14_p18);
+          v[ 5] = k_madd_epi32(u[ 5], k32_p14_p18);
+          v[ 6] = k_madd_epi32(u[ 6], k32_p14_p18);
+          v[ 7] = k_madd_epi32(u[ 7], k32_p14_p18);
+          v[ 8] = k_madd_epi32(u[ 8], k32_p22_p10);
+          v[ 9] = k_madd_epi32(u[ 9], k32_p22_p10);
+          v[10] = k_madd_epi32(u[10], k32_p22_p10);
+          v[11] = k_madd_epi32(u[11], k32_p22_p10);
+          v[12] = k_madd_epi32(u[12], k32_p06_p26);
+          v[13] = k_madd_epi32(u[13], k32_p06_p26);
+          v[14] = k_madd_epi32(u[14], k32_p06_p26);
+          v[15] = k_madd_epi32(u[15], k32_p06_p26);
+          v[16] = k_madd_epi32(u[12], k32_m26_p06);
+          v[17] = k_madd_epi32(u[13], k32_m26_p06);
+          v[18] = k_madd_epi32(u[14], k32_m26_p06);
+          v[19] = k_madd_epi32(u[15], k32_m26_p06);
+          v[20] = k_madd_epi32(u[ 8], k32_m10_p22);
+          v[21] = k_madd_epi32(u[ 9], k32_m10_p22);
+          v[22] = k_madd_epi32(u[10], k32_m10_p22);
+          v[23] = k_madd_epi32(u[11], k32_m10_p22);
+          v[24] = k_madd_epi32(u[ 4], k32_m18_p14);
+          v[25] = k_madd_epi32(u[ 5], k32_m18_p14);
+          v[26] = k_madd_epi32(u[ 6], k32_m18_p14);
+          v[27] = k_madd_epi32(u[ 7], k32_m18_p14);
+          v[28] = k_madd_epi32(u[ 0], k32_m02_p30);
+          v[29] = k_madd_epi32(u[ 1], k32_m02_p30);
+          v[30] = k_madd_epi32(u[ 2], k32_m02_p30);
+          v[31] = k_madd_epi32(u[ 3], k32_m02_p30);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &v[16], &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23],
+              &v[24], &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64(v[10], v[11]);
+          u[ 6] = k_packs_epi64(v[12], v[13]);
+          u[ 7] = k_packs_epi64(v[14], v[15]);
+          u[ 8] = k_packs_epi64(v[16], v[17]);
+          u[ 9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
+          v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
+          v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
+          v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
+          v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
+          v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
+          v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
+          v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
+          v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
+          v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
+          v[10] = _mm_cmplt_epi32(u[10], kZero);
+          v[11] = _mm_cmplt_epi32(u[11], kZero);
+          v[12] = _mm_cmplt_epi32(u[12], kZero);
+          v[13] = _mm_cmplt_epi32(u[13], kZero);
+          v[14] = _mm_cmplt_epi32(u[14], kZero);
+          v[15] = _mm_cmplt_epi32(u[15], kZero);
+
+          u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm_sub_epi32(u[10], v[10]);
+          u[11] = _mm_sub_epi32(u[11], v[11]);
+          u[12] = _mm_sub_epi32(u[12], v[12]);
+          u[13] = _mm_sub_epi32(u[13], v[13]);
+          u[14] = _mm_sub_epi32(u[14], v[14]);
+          u[15] = _mm_sub_epi32(u[15], v[15]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], K32One);
+          v[ 1] = _mm_add_epi32(u[ 1], K32One);
+          v[ 2] = _mm_add_epi32(u[ 2], K32One);
+          v[ 3] = _mm_add_epi32(u[ 3], K32One);
+          v[ 4] = _mm_add_epi32(u[ 4], K32One);
+          v[ 5] = _mm_add_epi32(u[ 5], K32One);
+          v[ 6] = _mm_add_epi32(u[ 6], K32One);
+          v[ 7] = _mm_add_epi32(u[ 7], K32One);
+          v[ 8] = _mm_add_epi32(u[ 8], K32One);
+          v[ 9] = _mm_add_epi32(u[ 9], K32One);
+          v[10] = _mm_add_epi32(u[10], K32One);
+          v[11] = _mm_add_epi32(u[11], K32One);
+          v[12] = _mm_add_epi32(u[12], K32One);
+          v[13] = _mm_add_epi32(u[13], K32One);
+          v[14] = _mm_add_epi32(u[14], K32One);
+          v[15] = _mm_add_epi32(u[15], K32One);
+
+          u[ 0] = _mm_srai_epi32(v[ 0], 2);
+          u[ 1] = _mm_srai_epi32(v[ 1], 2);
+          u[ 2] = _mm_srai_epi32(v[ 2], 2);
+          u[ 3] = _mm_srai_epi32(v[ 3], 2);
+          u[ 4] = _mm_srai_epi32(v[ 4], 2);
+          u[ 5] = _mm_srai_epi32(v[ 5], 2);
+          u[ 6] = _mm_srai_epi32(v[ 6], 2);
+          u[ 7] = _mm_srai_epi32(v[ 7], 2);
+          u[ 8] = _mm_srai_epi32(v[ 8], 2);
+          u[ 9] = _mm_srai_epi32(v[ 9], 2);
+          u[10] = _mm_srai_epi32(v[10], 2);
+          u[11] = _mm_srai_epi32(v[11], 2);
+          u[12] = _mm_srai_epi32(v[12], 2);
+          u[13] = _mm_srai_epi32(v[13], 2);
+          u[14] = _mm_srai_epi32(v[14], 2);
+          u[15] = _mm_srai_epi32(v[15], 2);
+
+          out[ 2] = _mm_packs_epi32(u[0], u[1]);
+          out[18] = _mm_packs_epi32(u[2], u[3]);
+          out[10] = _mm_packs_epi32(u[4], u[5]);
+          out[26] = _mm_packs_epi32(u[6], u[7]);
+          out[ 6] = _mm_packs_epi32(u[8], u[9]);
+          out[22] = _mm_packs_epi32(u[10], u[11]);
+          out[14] = _mm_packs_epi32(u[12], u[13]);
+          out[30] = _mm_packs_epi32(u[14], u[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&out[2], &out[18], &out[10],
+                                             &out[26], &out[6], &out[22],
+                                             &out[14], &out[30]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          lstep1[32] = _mm_add_epi32(lstep3[34], lstep2[32]);
+          lstep1[33] = _mm_add_epi32(lstep3[35], lstep2[33]);
+          lstep1[34] = _mm_sub_epi32(lstep2[32], lstep3[34]);
+          lstep1[35] = _mm_sub_epi32(lstep2[33], lstep3[35]);
+          lstep1[36] = _mm_sub_epi32(lstep2[38], lstep3[36]);
+          lstep1[37] = _mm_sub_epi32(lstep2[39], lstep3[37]);
+          lstep1[38] = _mm_add_epi32(lstep3[36], lstep2[38]);
+          lstep1[39] = _mm_add_epi32(lstep3[37], lstep2[39]);
+          lstep1[40] = _mm_add_epi32(lstep3[42], lstep2[40]);
+          lstep1[41] = _mm_add_epi32(lstep3[43], lstep2[41]);
+          lstep1[42] = _mm_sub_epi32(lstep2[40], lstep3[42]);
+          lstep1[43] = _mm_sub_epi32(lstep2[41], lstep3[43]);
+          lstep1[44] = _mm_sub_epi32(lstep2[46], lstep3[44]);
+          lstep1[45] = _mm_sub_epi32(lstep2[47], lstep3[45]);
+          lstep1[46] = _mm_add_epi32(lstep3[44], lstep2[46]);
+          lstep1[47] = _mm_add_epi32(lstep3[45], lstep2[47]);
+          lstep1[48] = _mm_add_epi32(lstep3[50], lstep2[48]);
+          lstep1[49] = _mm_add_epi32(lstep3[51], lstep2[49]);
+          lstep1[50] = _mm_sub_epi32(lstep2[48], lstep3[50]);
+          lstep1[51] = _mm_sub_epi32(lstep2[49], lstep3[51]);
+          lstep1[52] = _mm_sub_epi32(lstep2[54], lstep3[52]);
+          lstep1[53] = _mm_sub_epi32(lstep2[55], lstep3[53]);
+          lstep1[54] = _mm_add_epi32(lstep3[52], lstep2[54]);
+          lstep1[55] = _mm_add_epi32(lstep3[53], lstep2[55]);
+          lstep1[56] = _mm_add_epi32(lstep3[58], lstep2[56]);
+          lstep1[57] = _mm_add_epi32(lstep3[59], lstep2[57]);
+          lstep1[58] = _mm_sub_epi32(lstep2[56], lstep3[58]);
+          lstep1[59] = _mm_sub_epi32(lstep2[57], lstep3[59]);
+          lstep1[60] = _mm_sub_epi32(lstep2[62], lstep3[60]);
+          lstep1[61] = _mm_sub_epi32(lstep2[63], lstep3[61]);
+          lstep1[62] = _mm_add_epi32(lstep3[60], lstep2[62]);
+          lstep1[63] = _mm_add_epi32(lstep3[61], lstep2[63]);
+        }
+        // stage 8
+        {
+          const __m128i k32_p31_p01 = pair_set_epi32(cospi_31_64, cospi_1_64);
+          const __m128i k32_p15_p17 = pair_set_epi32(cospi_15_64, cospi_17_64);
+          const __m128i k32_p23_p09 = pair_set_epi32(cospi_23_64, cospi_9_64);
+          const __m128i k32_p07_p25 = pair_set_epi32(cospi_7_64, cospi_25_64);
+          const __m128i k32_m25_p07 = pair_set_epi32(-cospi_25_64, cospi_7_64);
+          const __m128i k32_m09_p23 = pair_set_epi32(-cospi_9_64, cospi_23_64);
+          const __m128i k32_m17_p15 = pair_set_epi32(-cospi_17_64, cospi_15_64);
+          const __m128i k32_m01_p31 = pair_set_epi32(-cospi_1_64, cospi_31_64);
+
+          u[ 0] = _mm_unpacklo_epi32(lstep1[32], lstep1[62]);
+          u[ 1] = _mm_unpackhi_epi32(lstep1[32], lstep1[62]);
+          u[ 2] = _mm_unpacklo_epi32(lstep1[33], lstep1[63]);
+          u[ 3] = _mm_unpackhi_epi32(lstep1[33], lstep1[63]);
+          u[ 4] = _mm_unpacklo_epi32(lstep1[34], lstep1[60]);
+          u[ 5] = _mm_unpackhi_epi32(lstep1[34], lstep1[60]);
+          u[ 6] = _mm_unpacklo_epi32(lstep1[35], lstep1[61]);
+          u[ 7] = _mm_unpackhi_epi32(lstep1[35], lstep1[61]);
+          u[ 8] = _mm_unpacklo_epi32(lstep1[36], lstep1[58]);
+          u[ 9] = _mm_unpackhi_epi32(lstep1[36], lstep1[58]);
+          u[10] = _mm_unpacklo_epi32(lstep1[37], lstep1[59]);
+          u[11] = _mm_unpackhi_epi32(lstep1[37], lstep1[59]);
+          u[12] = _mm_unpacklo_epi32(lstep1[38], lstep1[56]);
+          u[13] = _mm_unpackhi_epi32(lstep1[38], lstep1[56]);
+          u[14] = _mm_unpacklo_epi32(lstep1[39], lstep1[57]);
+          u[15] = _mm_unpackhi_epi32(lstep1[39], lstep1[57]);
+
+          v[ 0] = k_madd_epi32(u[ 0], k32_p31_p01);
+          v[ 1] = k_madd_epi32(u[ 1], k32_p31_p01);
+          v[ 2] = k_madd_epi32(u[ 2], k32_p31_p01);
+          v[ 3] = k_madd_epi32(u[ 3], k32_p31_p01);
+          v[ 4] = k_madd_epi32(u[ 4], k32_p15_p17);
+          v[ 5] = k_madd_epi32(u[ 5], k32_p15_p17);
+          v[ 6] = k_madd_epi32(u[ 6], k32_p15_p17);
+          v[ 7] = k_madd_epi32(u[ 7], k32_p15_p17);
+          v[ 8] = k_madd_epi32(u[ 8], k32_p23_p09);
+          v[ 9] = k_madd_epi32(u[ 9], k32_p23_p09);
+          v[10] = k_madd_epi32(u[10], k32_p23_p09);
+          v[11] = k_madd_epi32(u[11], k32_p23_p09);
+          v[12] = k_madd_epi32(u[12], k32_p07_p25);
+          v[13] = k_madd_epi32(u[13], k32_p07_p25);
+          v[14] = k_madd_epi32(u[14], k32_p07_p25);
+          v[15] = k_madd_epi32(u[15], k32_p07_p25);
+          v[16] = k_madd_epi32(u[12], k32_m25_p07);
+          v[17] = k_madd_epi32(u[13], k32_m25_p07);
+          v[18] = k_madd_epi32(u[14], k32_m25_p07);
+          v[19] = k_madd_epi32(u[15], k32_m25_p07);
+          v[20] = k_madd_epi32(u[ 8], k32_m09_p23);
+          v[21] = k_madd_epi32(u[ 9], k32_m09_p23);
+          v[22] = k_madd_epi32(u[10], k32_m09_p23);
+          v[23] = k_madd_epi32(u[11], k32_m09_p23);
+          v[24] = k_madd_epi32(u[ 4], k32_m17_p15);
+          v[25] = k_madd_epi32(u[ 5], k32_m17_p15);
+          v[26] = k_madd_epi32(u[ 6], k32_m17_p15);
+          v[27] = k_madd_epi32(u[ 7], k32_m17_p15);
+          v[28] = k_madd_epi32(u[ 0], k32_m01_p31);
+          v[29] = k_madd_epi32(u[ 1], k32_m01_p31);
+          v[30] = k_madd_epi32(u[ 2], k32_m01_p31);
+          v[31] = k_madd_epi32(u[ 3], k32_m01_p31);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &v[16], &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23],
+              &v[24], &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64(v[10], v[11]);
+          u[ 6] = k_packs_epi64(v[12], v[13]);
+          u[ 7] = k_packs_epi64(v[14], v[15]);
+          u[ 8] = k_packs_epi64(v[16], v[17]);
+          u[ 9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
+          v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
+          v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
+          v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
+          v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
+          v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
+          v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
+          v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
+          v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
+          v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
+          v[10] = _mm_cmplt_epi32(u[10], kZero);
+          v[11] = _mm_cmplt_epi32(u[11], kZero);
+          v[12] = _mm_cmplt_epi32(u[12], kZero);
+          v[13] = _mm_cmplt_epi32(u[13], kZero);
+          v[14] = _mm_cmplt_epi32(u[14], kZero);
+          v[15] = _mm_cmplt_epi32(u[15], kZero);
+
+          u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm_sub_epi32(u[10], v[10]);
+          u[11] = _mm_sub_epi32(u[11], v[11]);
+          u[12] = _mm_sub_epi32(u[12], v[12]);
+          u[13] = _mm_sub_epi32(u[13], v[13]);
+          u[14] = _mm_sub_epi32(u[14], v[14]);
+          u[15] = _mm_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], K32One);
+          v[1] = _mm_add_epi32(u[1], K32One);
+          v[2] = _mm_add_epi32(u[2], K32One);
+          v[3] = _mm_add_epi32(u[3], K32One);
+          v[4] = _mm_add_epi32(u[4], K32One);
+          v[5] = _mm_add_epi32(u[5], K32One);
+          v[6] = _mm_add_epi32(u[6], K32One);
+          v[7] = _mm_add_epi32(u[7], K32One);
+          v[8] = _mm_add_epi32(u[8], K32One);
+          v[9] = _mm_add_epi32(u[9], K32One);
+          v[10] = _mm_add_epi32(u[10], K32One);
+          v[11] = _mm_add_epi32(u[11], K32One);
+          v[12] = _mm_add_epi32(u[12], K32One);
+          v[13] = _mm_add_epi32(u[13], K32One);
+          v[14] = _mm_add_epi32(u[14], K32One);
+          v[15] = _mm_add_epi32(u[15], K32One);
+
+          u[0] = _mm_srai_epi32(v[0], 2);
+          u[1] = _mm_srai_epi32(v[1], 2);
+          u[2] = _mm_srai_epi32(v[2], 2);
+          u[3] = _mm_srai_epi32(v[3], 2);
+          u[4] = _mm_srai_epi32(v[4], 2);
+          u[5] = _mm_srai_epi32(v[5], 2);
+          u[6] = _mm_srai_epi32(v[6], 2);
+          u[7] = _mm_srai_epi32(v[7], 2);
+          u[8] = _mm_srai_epi32(v[8], 2);
+          u[9] = _mm_srai_epi32(v[9], 2);
+          u[10] = _mm_srai_epi32(v[10], 2);
+          u[11] = _mm_srai_epi32(v[11], 2);
+          u[12] = _mm_srai_epi32(v[12], 2);
+          u[13] = _mm_srai_epi32(v[13], 2);
+          u[14] = _mm_srai_epi32(v[14], 2);
+          u[15] = _mm_srai_epi32(v[15], 2);
+
+          out[ 1] = _mm_packs_epi32(u[0], u[1]);
+          out[17] = _mm_packs_epi32(u[2], u[3]);
+          out[ 9] = _mm_packs_epi32(u[4], u[5]);
+          out[25] = _mm_packs_epi32(u[6], u[7]);
+          out[ 7] = _mm_packs_epi32(u[8], u[9]);
+          out[23] = _mm_packs_epi32(u[10], u[11]);
+          out[15] = _mm_packs_epi32(u[12], u[13]);
+          out[31] = _mm_packs_epi32(u[14], u[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&out[1], &out[17], &out[9],
+                                             &out[25], &out[7], &out[23],
+                                             &out[15], &out[31]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i k32_p27_p05 = pair_set_epi32(cospi_27_64, cospi_5_64);
+          const __m128i k32_p11_p21 = pair_set_epi32(cospi_11_64, cospi_21_64);
+          const __m128i k32_p19_p13 = pair_set_epi32(cospi_19_64, cospi_13_64);
+          const __m128i k32_p03_p29 = pair_set_epi32(cospi_3_64, cospi_29_64);
+          const __m128i k32_m29_p03 = pair_set_epi32(-cospi_29_64, cospi_3_64);
+          const __m128i k32_m13_p19 = pair_set_epi32(-cospi_13_64, cospi_19_64);
+          const __m128i k32_m21_p11 = pair_set_epi32(-cospi_21_64, cospi_11_64);
+          const __m128i k32_m05_p27 = pair_set_epi32(-cospi_5_64, cospi_27_64);
+
+          u[ 0] = _mm_unpacklo_epi32(lstep1[40], lstep1[54]);
+          u[ 1] = _mm_unpackhi_epi32(lstep1[40], lstep1[54]);
+          u[ 2] = _mm_unpacklo_epi32(lstep1[41], lstep1[55]);
+          u[ 3] = _mm_unpackhi_epi32(lstep1[41], lstep1[55]);
+          u[ 4] = _mm_unpacklo_epi32(lstep1[42], lstep1[52]);
+          u[ 5] = _mm_unpackhi_epi32(lstep1[42], lstep1[52]);
+          u[ 6] = _mm_unpacklo_epi32(lstep1[43], lstep1[53]);
+          u[ 7] = _mm_unpackhi_epi32(lstep1[43], lstep1[53]);
+          u[ 8] = _mm_unpacklo_epi32(lstep1[44], lstep1[50]);
+          u[ 9] = _mm_unpackhi_epi32(lstep1[44], lstep1[50]);
+          u[10] = _mm_unpacklo_epi32(lstep1[45], lstep1[51]);
+          u[11] = _mm_unpackhi_epi32(lstep1[45], lstep1[51]);
+          u[12] = _mm_unpacklo_epi32(lstep1[46], lstep1[48]);
+          u[13] = _mm_unpackhi_epi32(lstep1[46], lstep1[48]);
+          u[14] = _mm_unpacklo_epi32(lstep1[47], lstep1[49]);
+          u[15] = _mm_unpackhi_epi32(lstep1[47], lstep1[49]);
+
+          v[ 0] = k_madd_epi32(u[ 0], k32_p27_p05);
+          v[ 1] = k_madd_epi32(u[ 1], k32_p27_p05);
+          v[ 2] = k_madd_epi32(u[ 2], k32_p27_p05);
+          v[ 3] = k_madd_epi32(u[ 3], k32_p27_p05);
+          v[ 4] = k_madd_epi32(u[ 4], k32_p11_p21);
+          v[ 5] = k_madd_epi32(u[ 5], k32_p11_p21);
+          v[ 6] = k_madd_epi32(u[ 6], k32_p11_p21);
+          v[ 7] = k_madd_epi32(u[ 7], k32_p11_p21);
+          v[ 8] = k_madd_epi32(u[ 8], k32_p19_p13);
+          v[ 9] = k_madd_epi32(u[ 9], k32_p19_p13);
+          v[10] = k_madd_epi32(u[10], k32_p19_p13);
+          v[11] = k_madd_epi32(u[11], k32_p19_p13);
+          v[12] = k_madd_epi32(u[12], k32_p03_p29);
+          v[13] = k_madd_epi32(u[13], k32_p03_p29);
+          v[14] = k_madd_epi32(u[14], k32_p03_p29);
+          v[15] = k_madd_epi32(u[15], k32_p03_p29);
+          v[16] = k_madd_epi32(u[12], k32_m29_p03);
+          v[17] = k_madd_epi32(u[13], k32_m29_p03);
+          v[18] = k_madd_epi32(u[14], k32_m29_p03);
+          v[19] = k_madd_epi32(u[15], k32_m29_p03);
+          v[20] = k_madd_epi32(u[ 8], k32_m13_p19);
+          v[21] = k_madd_epi32(u[ 9], k32_m13_p19);
+          v[22] = k_madd_epi32(u[10], k32_m13_p19);
+          v[23] = k_madd_epi32(u[11], k32_m13_p19);
+          v[24] = k_madd_epi32(u[ 4], k32_m21_p11);
+          v[25] = k_madd_epi32(u[ 5], k32_m21_p11);
+          v[26] = k_madd_epi32(u[ 6], k32_m21_p11);
+          v[27] = k_madd_epi32(u[ 7], k32_m21_p11);
+          v[28] = k_madd_epi32(u[ 0], k32_m05_p27);
+          v[29] = k_madd_epi32(u[ 1], k32_m05_p27);
+          v[30] = k_madd_epi32(u[ 2], k32_m05_p27);
+          v[31] = k_madd_epi32(u[ 3], k32_m05_p27);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &v[16], &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23],
+              &v[24], &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64(v[10], v[11]);
+          u[ 6] = k_packs_epi64(v[12], v[13]);
+          u[ 7] = k_packs_epi64(v[14], v[15]);
+          u[ 8] = k_packs_epi64(v[16], v[17]);
+          u[ 9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
+          v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
+          v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
+          v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
+          v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
+          v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
+          v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
+          v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
+          v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
+          v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
+          v[10] = _mm_cmplt_epi32(u[10], kZero);
+          v[11] = _mm_cmplt_epi32(u[11], kZero);
+          v[12] = _mm_cmplt_epi32(u[12], kZero);
+          v[13] = _mm_cmplt_epi32(u[13], kZero);
+          v[14] = _mm_cmplt_epi32(u[14], kZero);
+          v[15] = _mm_cmplt_epi32(u[15], kZero);
+
+          u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm_sub_epi32(u[10], v[10]);
+          u[11] = _mm_sub_epi32(u[11], v[11]);
+          u[12] = _mm_sub_epi32(u[12], v[12]);
+          u[13] = _mm_sub_epi32(u[13], v[13]);
+          u[14] = _mm_sub_epi32(u[14], v[14]);
+          u[15] = _mm_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], K32One);
+          v[1] = _mm_add_epi32(u[1], K32One);
+          v[2] = _mm_add_epi32(u[2], K32One);
+          v[3] = _mm_add_epi32(u[3], K32One);
+          v[4] = _mm_add_epi32(u[4], K32One);
+          v[5] = _mm_add_epi32(u[5], K32One);
+          v[6] = _mm_add_epi32(u[6], K32One);
+          v[7] = _mm_add_epi32(u[7], K32One);
+          v[8] = _mm_add_epi32(u[8], K32One);
+          v[9] = _mm_add_epi32(u[9], K32One);
+          v[10] = _mm_add_epi32(u[10], K32One);
+          v[11] = _mm_add_epi32(u[11], K32One);
+          v[12] = _mm_add_epi32(u[12], K32One);
+          v[13] = _mm_add_epi32(u[13], K32One);
+          v[14] = _mm_add_epi32(u[14], K32One);
+          v[15] = _mm_add_epi32(u[15], K32One);
+
+          u[0] = _mm_srai_epi32(v[0], 2);
+          u[1] = _mm_srai_epi32(v[1], 2);
+          u[2] = _mm_srai_epi32(v[2], 2);
+          u[3] = _mm_srai_epi32(v[3], 2);
+          u[4] = _mm_srai_epi32(v[4], 2);
+          u[5] = _mm_srai_epi32(v[5], 2);
+          u[6] = _mm_srai_epi32(v[6], 2);
+          u[7] = _mm_srai_epi32(v[7], 2);
+          u[8] = _mm_srai_epi32(v[8], 2);
+          u[9] = _mm_srai_epi32(v[9], 2);
+          u[10] = _mm_srai_epi32(v[10], 2);
+          u[11] = _mm_srai_epi32(v[11], 2);
+          u[12] = _mm_srai_epi32(v[12], 2);
+          u[13] = _mm_srai_epi32(v[13], 2);
+          u[14] = _mm_srai_epi32(v[14], 2);
+          u[15] = _mm_srai_epi32(v[15], 2);
+
+          out[ 5] = _mm_packs_epi32(u[0], u[1]);
+          out[21] = _mm_packs_epi32(u[2], u[3]);
+          out[13] = _mm_packs_epi32(u[4], u[5]);
+          out[29] = _mm_packs_epi32(u[6], u[7]);
+          out[ 3] = _mm_packs_epi32(u[8], u[9]);
+          out[19] = _mm_packs_epi32(u[10], u[11]);
+          out[11] = _mm_packs_epi32(u[12], u[13]);
+          out[27] = _mm_packs_epi32(u[14], u[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&out[5], &out[21], &out[13],
+                                             &out[29], &out[3], &out[19],
+                                             &out[11], &out[27]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+#endif  // FDCT32x32_HIGH_PRECISION
+      // Transpose the results, do it as four 8x8 transposes.
+      {
+        int transpose_block;
+        int16_t *output0 = &intermediate[column_start * 32];
+        tran_low_t *output1 = &output_org[column_start * 32];
+        for (transpose_block = 0; transpose_block < 4; ++transpose_block) {
+          __m128i *this_out = &out[8 * transpose_block];
+          // 00 01 02 03 04 05 06 07
+          // 10 11 12 13 14 15 16 17
+          // 20 21 22 23 24 25 26 27
+          // 30 31 32 33 34 35 36 37
+          // 40 41 42 43 44 45 46 47
+          // 50 51 52 53 54 55 56 57
+          // 60 61 62 63 64 65 66 67
+          // 70 71 72 73 74 75 76 77
+          const __m128i tr0_0 = _mm_unpacklo_epi16(this_out[0], this_out[1]);
+          const __m128i tr0_1 = _mm_unpacklo_epi16(this_out[2], this_out[3]);
+          const __m128i tr0_2 = _mm_unpackhi_epi16(this_out[0], this_out[1]);
+          const __m128i tr0_3 = _mm_unpackhi_epi16(this_out[2], this_out[3]);
+          const __m128i tr0_4 = _mm_unpacklo_epi16(this_out[4], this_out[5]);
+          const __m128i tr0_5 = _mm_unpacklo_epi16(this_out[6], this_out[7]);
+          const __m128i tr0_6 = _mm_unpackhi_epi16(this_out[4], this_out[5]);
+          const __m128i tr0_7 = _mm_unpackhi_epi16(this_out[6], this_out[7]);
+          // 00 10 01 11 02 12 03 13
+          // 20 30 21 31 22 32 23 33
+          // 04 14 05 15 06 16 07 17
+          // 24 34 25 35 26 36 27 37
+          // 40 50 41 51 42 52 43 53
+          // 60 70 61 71 62 72 63 73
+          // 54 54 55 55 56 56 57 57
+          // 64 74 65 75 66 76 67 77
+          const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+          const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+          const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+          const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+          const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+          const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+          const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+          const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+          // 00 10 20 30 01 11 21 31
+          // 40 50 60 70 41 51 61 71
+          // 02 12 22 32 03 13 23 33
+          // 42 52 62 72 43 53 63 73
+          // 04 14 24 34 05 15 21 36
+          // 44 54 64 74 45 55 61 76
+          // 06 16 26 36 07 17 27 37
+          // 46 56 66 76 47 57 67 77
+          __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+          __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+          __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+          __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+          __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+          __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+          __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+          __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+          // 00 10 20 30 40 50 60 70
+          // 01 11 21 31 41 51 61 71
+          // 02 12 22 32 42 52 62 72
+          // 03 13 23 33 43 53 63 73
+          // 04 14 24 34 44 54 64 74
+          // 05 15 25 35 45 55 65 75
+          // 06 16 26 36 46 56 66 76
+          // 07 17 27 37 47 57 67 77
+          if (0 == pass) {
+            // output[j] = (output[j] + 1 + (output[j] > 0)) >> 2;
+            // TODO(cd): see quality impact of only doing
+            //           output[j] = (output[j] + 1) >> 2;
+            //           which would remove the code between here ...
+            __m128i tr2_0_0 = _mm_cmpgt_epi16(tr2_0, kZero);
+            __m128i tr2_1_0 = _mm_cmpgt_epi16(tr2_1, kZero);
+            __m128i tr2_2_0 = _mm_cmpgt_epi16(tr2_2, kZero);
+            __m128i tr2_3_0 = _mm_cmpgt_epi16(tr2_3, kZero);
+            __m128i tr2_4_0 = _mm_cmpgt_epi16(tr2_4, kZero);
+            __m128i tr2_5_0 = _mm_cmpgt_epi16(tr2_5, kZero);
+            __m128i tr2_6_0 = _mm_cmpgt_epi16(tr2_6, kZero);
+            __m128i tr2_7_0 = _mm_cmpgt_epi16(tr2_7, kZero);
+            tr2_0 = _mm_sub_epi16(tr2_0, tr2_0_0);
+            tr2_1 = _mm_sub_epi16(tr2_1, tr2_1_0);
+            tr2_2 = _mm_sub_epi16(tr2_2, tr2_2_0);
+            tr2_3 = _mm_sub_epi16(tr2_3, tr2_3_0);
+            tr2_4 = _mm_sub_epi16(tr2_4, tr2_4_0);
+            tr2_5 = _mm_sub_epi16(tr2_5, tr2_5_0);
+            tr2_6 = _mm_sub_epi16(tr2_6, tr2_6_0);
+            tr2_7 = _mm_sub_epi16(tr2_7, tr2_7_0);
+            //           ... and here.
+            //           PS: also change code in vp9/encoder/vp9_dct.c
+            tr2_0 = _mm_add_epi16(tr2_0, kOne);
+            tr2_1 = _mm_add_epi16(tr2_1, kOne);
+            tr2_2 = _mm_add_epi16(tr2_2, kOne);
+            tr2_3 = _mm_add_epi16(tr2_3, kOne);
+            tr2_4 = _mm_add_epi16(tr2_4, kOne);
+            tr2_5 = _mm_add_epi16(tr2_5, kOne);
+            tr2_6 = _mm_add_epi16(tr2_6, kOne);
+            tr2_7 = _mm_add_epi16(tr2_7, kOne);
+            tr2_0 = _mm_srai_epi16(tr2_0, 2);
+            tr2_1 = _mm_srai_epi16(tr2_1, 2);
+            tr2_2 = _mm_srai_epi16(tr2_2, 2);
+            tr2_3 = _mm_srai_epi16(tr2_3, 2);
+            tr2_4 = _mm_srai_epi16(tr2_4, 2);
+            tr2_5 = _mm_srai_epi16(tr2_5, 2);
+            tr2_6 = _mm_srai_epi16(tr2_6, 2);
+            tr2_7 = _mm_srai_epi16(tr2_7, 2);
+          }
+          // Note: even though all these stores are aligned, using the aligned
+          //       intrinsic make the code slightly slower.
+          if (pass == 0) {
+            _mm_storeu_si128((__m128i *)(output0 + 0 * 32), tr2_0);
+            _mm_storeu_si128((__m128i *)(output0 + 1 * 32), tr2_1);
+            _mm_storeu_si128((__m128i *)(output0 + 2 * 32), tr2_2);
+            _mm_storeu_si128((__m128i *)(output0 + 3 * 32), tr2_3);
+            _mm_storeu_si128((__m128i *)(output0 + 4 * 32), tr2_4);
+            _mm_storeu_si128((__m128i *)(output0 + 5 * 32), tr2_5);
+            _mm_storeu_si128((__m128i *)(output0 + 6 * 32), tr2_6);
+            _mm_storeu_si128((__m128i *)(output0 + 7 * 32), tr2_7);
+            // Process next 8x8
+            output0 += 8;
+          } else {
+            storeu_output(&tr2_0, (output1 + 0 * 32));
+            storeu_output(&tr2_1, (output1 + 1 * 32));
+            storeu_output(&tr2_2, (output1 + 2 * 32));
+            storeu_output(&tr2_3, (output1 + 3 * 32));
+            storeu_output(&tr2_4, (output1 + 4 * 32));
+            storeu_output(&tr2_5, (output1 + 5 * 32));
+            storeu_output(&tr2_6, (output1 + 6 * 32));
+            storeu_output(&tr2_7, (output1 + 7 * 32));
+            // Process next 8x8
+            output1 += 8;
+          }
+        }
+      }
+    }
+  }
+}  // NOLINT
+
+#undef ADD_EPI16
+#undef SUB_EPI16
+#undef HIGH_FDCT32x32_2D_C
+#undef HIGH_FDCT32x32_2D_ROWS_C
diff --git a/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_avx2.c b/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_avx2.c
new file mode 100644
index 0000000..6d9da6a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_avx2.c
@@ -0,0 +1,23 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+
+#define FDCT32x32_2D_AVX2 vpx_fdct32x32_rd_avx2
+#define FDCT32x32_HIGH_PRECISION 0
+#include "vpx_dsp/x86/fwd_dct32x32_impl_avx2.h"
+#undef  FDCT32x32_2D_AVX2
+#undef  FDCT32x32_HIGH_PRECISION
+
+#define FDCT32x32_2D_AVX2 vpx_fdct32x32_avx2
+#define FDCT32x32_HIGH_PRECISION 1
+#include "vpx_dsp/x86/fwd_dct32x32_impl_avx2.h" // NOLINT
+#undef  FDCT32x32_2D_AVX2
+#undef  FDCT32x32_HIGH_PRECISION
diff --git a/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_impl_sse2.h b/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_impl_sse2.h
new file mode 100644
index 0000000..69889e2
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_impl_sse2.h
@@ -0,0 +1,1027 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/txfm_common.h"
+#include "vpx_dsp/x86/fwd_txfm_sse2.h"
+#include "vpx_dsp/x86/txfm_common_sse2.h"
+#include "vpx_ports/mem.h"
+
+// TODO(jingning) The high bit-depth functions need rework for performance.
+// After we properly fix the high bit-depth function implementations, this
+// file's dependency should be substantially simplified.
+#if DCT_HIGH_BIT_DEPTH
+#define ADD_EPI16 _mm_adds_epi16
+#define SUB_EPI16 _mm_subs_epi16
+
+#else
+#define ADD_EPI16 _mm_add_epi16
+#define SUB_EPI16 _mm_sub_epi16
+#endif
+
+void FDCT4x4_2D(const int16_t *input, tran_low_t *output, int stride) {
+  // This 2D transform implements 4 vertical 1D transforms followed
+  // by 4 horizontal 1D transforms.  The multiplies and adds are as given
+  // by Chen, Smith and Fralick ('77).  The commands for moving the data
+  // around have been minimized by hand.
+  // For the purposes of the comments, the 16 inputs are referred to at i0
+  // through iF (in raster order), intermediate variables are a0, b0, c0
+  // through f, and correspond to the in-place computations mapped to input
+  // locations.  The outputs, o0 through oF are labeled according to the
+  // output locations.
+
+  // Constants
+  // These are the coefficients used for the multiplies.
+  // In the comments, pN means cos(N pi /64) and mN is -cos(N pi /64),
+  // where cospi_N_64 = cos(N pi /64)
+  const __m128i k__cospi_A = octa_set_epi16(cospi_16_64, cospi_16_64,
+                                            cospi_16_64, cospi_16_64,
+                                            cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_B = octa_set_epi16(cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, cospi_16_64,
+                                            cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_C = octa_set_epi16(cospi_8_64, cospi_24_64,
+                                            cospi_8_64, cospi_24_64,
+                                            cospi_24_64, -cospi_8_64,
+                                            cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_D = octa_set_epi16(cospi_24_64, -cospi_8_64,
+                                            cospi_24_64, -cospi_8_64,
+                                            cospi_8_64, cospi_24_64,
+                                            cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_E = octa_set_epi16(cospi_16_64, cospi_16_64,
+                                            cospi_16_64, cospi_16_64,
+                                            cospi_16_64, cospi_16_64,
+                                            cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_F = octa_set_epi16(cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_G = octa_set_epi16(cospi_8_64, cospi_24_64,
+                                            cospi_8_64, cospi_24_64,
+                                            -cospi_8_64, -cospi_24_64,
+                                            -cospi_8_64, -cospi_24_64);
+  const __m128i k__cospi_H = octa_set_epi16(cospi_24_64, -cospi_8_64,
+                                            cospi_24_64, -cospi_8_64,
+                                            -cospi_24_64, cospi_8_64,
+                                            -cospi_24_64, cospi_8_64);
+
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // This second rounding constant saves doing some extra adds at the end
+  const __m128i k__DCT_CONST_ROUNDING2 = _mm_set1_epi32(DCT_CONST_ROUNDING
+                                               +(DCT_CONST_ROUNDING << 1));
+  const int DCT_CONST_BITS2 =  DCT_CONST_BITS + 2;
+  const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
+  const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
+  __m128i in0, in1;
+#if DCT_HIGH_BIT_DEPTH
+  __m128i cmp0, cmp1;
+  int test, overflow;
+#endif
+
+  // Load inputs.
+  in0  = _mm_loadl_epi64((const __m128i *)(input +  0 * stride));
+  in1  = _mm_loadl_epi64((const __m128i *)(input +  1 * stride));
+  in1  = _mm_unpacklo_epi64(in1, _mm_loadl_epi64((const __m128i *)
+                                                 (input +  2 * stride)));
+  in0  = _mm_unpacklo_epi64(in0, _mm_loadl_epi64((const __m128i *)
+                                                 (input +  3 * stride)));
+  // in0 = [i0 i1 i2 i3 iC iD iE iF]
+  // in1 = [i4 i5 i6 i7 i8 i9 iA iB]
+#if DCT_HIGH_BIT_DEPTH
+  // Check inputs small enough to use optimised code
+  cmp0 = _mm_xor_si128(_mm_cmpgt_epi16(in0, _mm_set1_epi16(0x3ff)),
+                       _mm_cmplt_epi16(in0, _mm_set1_epi16(0xfc00)));
+  cmp1 = _mm_xor_si128(_mm_cmpgt_epi16(in1, _mm_set1_epi16(0x3ff)),
+                       _mm_cmplt_epi16(in1, _mm_set1_epi16(0xfc00)));
+  test = _mm_movemask_epi8(_mm_or_si128(cmp0, cmp1));
+  if (test) {
+    vpx_highbd_fdct4x4_c(input, output, stride);
+    return;
+  }
+#endif  // DCT_HIGH_BIT_DEPTH
+
+  // multiply by 16 to give some extra precision
+  in0 = _mm_slli_epi16(in0, 4);
+  in1 = _mm_slli_epi16(in1, 4);
+  // if (i == 0 && input[0]) input[0] += 1;
+  // add 1 to the upper left pixel if it is non-zero, which helps reduce
+  // the round-trip error
+  {
+    // The mask will only contain whether the first value is zero, all
+    // other comparison will fail as something shifted by 4 (above << 4)
+    // can never be equal to one. To increment in the non-zero case, we
+    // add the mask and one for the first element:
+    //   - if zero, mask = -1, v = v - 1 + 1 = v
+    //   - if non-zero, mask = 0, v = v + 0 + 1 = v + 1
+    __m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a);
+    in0 = _mm_add_epi16(in0, mask);
+    in0 = _mm_add_epi16(in0, k__nonzero_bias_b);
+  }
+  // There are 4 total stages, alternating between an add/subtract stage
+  // followed by an multiply-and-add stage.
+  {
+    // Stage 1: Add/subtract
+
+    // in0 = [i0 i1 i2 i3 iC iD iE iF]
+    // in1 = [i4 i5 i6 i7 i8 i9 iA iB]
+    const __m128i r0 = _mm_unpacklo_epi16(in0, in1);
+    const __m128i r1 = _mm_unpackhi_epi16(in0, in1);
+    // r0 = [i0 i4 i1 i5 i2 i6 i3 i7]
+    // r1 = [iC i8 iD i9 iE iA iF iB]
+    const __m128i r2 = _mm_shuffle_epi32(r0, 0xB4);
+    const __m128i r3 = _mm_shuffle_epi32(r1, 0xB4);
+    // r2 = [i0 i4 i1 i5 i3 i7 i2 i6]
+    // r3 = [iC i8 iD i9 iF iB iE iA]
+
+    const __m128i t0 = _mm_add_epi16(r2, r3);
+    const __m128i t1 = _mm_sub_epi16(r2, r3);
+    // t0 = [a0 a4 a1 a5 a3 a7 a2 a6]
+    // t1 = [aC a8 aD a9 aF aB aE aA]
+
+    // Stage 2: multiply by constants (which gets us into 32 bits).
+    // The constants needed here are:
+    // k__cospi_A = [p16 p16 p16 p16 p16 m16 p16 m16]
+    // k__cospi_B = [p16 m16 p16 m16 p16 p16 p16 p16]
+    // k__cospi_C = [p08 p24 p08 p24 p24 m08 p24 m08]
+    // k__cospi_D = [p24 m08 p24 m08 p08 p24 p08 p24]
+    const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A);
+    const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B);
+    const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C);
+    const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D);
+    // Then add and right-shift to get back to 16-bit range
+    const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+    const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+    const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+    const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+    const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+    const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+    const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+    const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+    // w0 = [b0 b1 b7 b6]
+    // w1 = [b8 b9 bF bE]
+    // w2 = [b4 b5 b3 b2]
+    // w3 = [bC bD bB bA]
+    const __m128i x0 = _mm_packs_epi32(w0, w1);
+    const __m128i x1 = _mm_packs_epi32(w2, w3);
+#if DCT_HIGH_BIT_DEPTH
+    overflow = check_epi16_overflow_x2(&x0, &x1);
+    if (overflow) {
+      vpx_highbd_fdct4x4_c(input, output, stride);
+      return;
+    }
+#endif  // DCT_HIGH_BIT_DEPTH
+    // x0 = [b0 b1 b7 b6 b8 b9 bF bE]
+    // x1 = [b4 b5 b3 b2 bC bD bB bA]
+    in0 = _mm_shuffle_epi32(x0, 0xD8);
+    in1 = _mm_shuffle_epi32(x1, 0x8D);
+    // in0 = [b0 b1 b8 b9 b7 b6 bF bE]
+    // in1 = [b3 b2 bB bA b4 b5 bC bD]
+  }
+  {
+    // vertical DCTs finished. Now we do the horizontal DCTs.
+    // Stage 3: Add/subtract
+
+    const __m128i t0 = ADD_EPI16(in0, in1);
+    const __m128i t1 = SUB_EPI16(in0, in1);
+    // t0 = [c0 c1 c8 c9  c4  c5  cC  cD]
+    // t1 = [c3 c2 cB cA -c7 -c6 -cF -cE]
+#if DCT_HIGH_BIT_DEPTH
+    overflow = check_epi16_overflow_x2(&t0, &t1);
+    if (overflow) {
+      vpx_highbd_fdct4x4_c(input, output, stride);
+      return;
+    }
+#endif  // DCT_HIGH_BIT_DEPTH
+
+    // Stage 4: multiply by constants (which gets us into 32 bits).
+    {
+      // The constants needed here are:
+      // k__cospi_E = [p16 p16 p16 p16 p16 p16 p16 p16]
+      // k__cospi_F = [p16 m16 p16 m16 p16 m16 p16 m16]
+      // k__cospi_G = [p08 p24 p08 p24 m08 m24 m08 m24]
+      // k__cospi_H = [p24 m08 p24 m08 m24 p08 m24 p08]
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E);
+      const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F);
+      const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H);
+      // Then add and right-shift to get back to 16-bit range
+      // but this combines the final right-shift as well to save operations
+      // This unusual rounding operations is to maintain bit-accurate
+      // compatibility with the c version of this function which has two
+      // rounding steps in a row.
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING2);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING2);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING2);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING2);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS2);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS2);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS2);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS2);
+      // w0 = [o0 o4 o8 oC]
+      // w1 = [o2 o6 oA oE]
+      // w2 = [o1 o5 o9 oD]
+      // w3 = [o3 o7 oB oF]
+      // remember the o's are numbered according to the correct output location
+      const __m128i x0 = _mm_packs_epi32(w0, w1);
+      const __m128i x1 = _mm_packs_epi32(w2, w3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x2(&x0, &x1);
+      if (overflow) {
+        vpx_highbd_fdct4x4_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      {
+        // x0 = [o0 o4 o8 oC o2 o6 oA oE]
+        // x1 = [o1 o5 o9 oD o3 o7 oB oF]
+        const __m128i y0 = _mm_unpacklo_epi16(x0, x1);
+        const __m128i y1 = _mm_unpackhi_epi16(x0, x1);
+        // y0 = [o0 o1 o4 o5 o8 o9 oC oD]
+        // y1 = [o2 o3 o6 o7 oA oB oE oF]
+        in0 = _mm_unpacklo_epi32(y0, y1);
+        // in0 = [o0 o1 o2 o3 o4 o5 o6 o7]
+        in1 = _mm_unpackhi_epi32(y0, y1);
+        // in1 = [o8 o9 oA oB oC oD oE oF]
+      }
+    }
+  }
+  // Post-condition (v + 1) >> 2 is now incorporated into previous
+  // add and right-shift commands.  Only 2 store instructions needed
+  // because we are using the fact that 1/3 are stored just after 0/2.
+  storeu_output(&in0, output + 0 * 4);
+  storeu_output(&in1, output + 2 * 4);
+}
+
+
+void FDCT8x8_2D(const int16_t *input, tran_low_t *output, int stride) {
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+#if DCT_HIGH_BIT_DEPTH
+  int overflow;
+#endif
+  // Load input
+  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = ADD_EPI16(in0, in7);
+    const __m128i q1 = ADD_EPI16(in1, in6);
+    const __m128i q2 = ADD_EPI16(in2, in5);
+    const __m128i q3 = ADD_EPI16(in3, in4);
+    const __m128i q4 = SUB_EPI16(in3, in4);
+    const __m128i q5 = SUB_EPI16(in2, in5);
+    const __m128i q6 = SUB_EPI16(in1, in6);
+    const __m128i q7 = SUB_EPI16(in0, in7);
+#if DCT_HIGH_BIT_DEPTH
+    if (pass == 1) {
+      overflow = check_epi16_overflow_x8(&q0, &q1, &q2, &q3,
+                                         &q4, &q5, &q6, &q7);
+      if (overflow) {
+        vpx_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+    }
+#endif  // DCT_HIGH_BIT_DEPTH
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = ADD_EPI16(q0, q3);
+      const __m128i r1 = ADD_EPI16(q1, q2);
+      const __m128i r2 = SUB_EPI16(q1, q2);
+      const __m128i r3 = SUB_EPI16(q0, q3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3);
+      if (overflow) {
+        vpx_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      // Interleave to do the multiply by constants which gets us into 32bits
+      {
+        const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+        const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+        const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+        const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+        const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+        const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+        const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+        const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+        const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+        const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+        const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+        const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+        // dct_const_round_shift
+        const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+        const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+        const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+        const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+        const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+        const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+        const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+        const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+        const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+        const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+        const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+        const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+        const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+        const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+        const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+        const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+        // Combine
+        res0 = _mm_packs_epi32(w0, w1);
+        res4 = _mm_packs_epi32(w2, w3);
+        res2 = _mm_packs_epi32(w4, w5);
+        res6 = _mm_packs_epi32(w6, w7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&res0, &res4, &res2, &res6);
+        if (overflow) {
+          vpx_highbd_fdct8x8_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
+      const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
+      const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
+      const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
+      const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
+      const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
+      // dct_const_round_shift
+      const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
+      const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
+      const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
+      const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
+      const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
+      const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
+      const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
+      const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
+      // Combine
+      const __m128i r0 = _mm_packs_epi32(s0, s1);
+      const __m128i r1 = _mm_packs_epi32(s2, s3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x2(&r0, &r1);
+      if (overflow) {
+        vpx_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      {
+        // Add/subtract
+        const __m128i x0 = ADD_EPI16(q4, r0);
+        const __m128i x1 = SUB_EPI16(q4, r0);
+        const __m128i x2 = SUB_EPI16(q7, r1);
+        const __m128i x3 = ADD_EPI16(q7, r1);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3);
+        if (overflow) {
+          vpx_highbd_fdct8x8_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        // Interleave to do the multiply by constants which gets us into 32bits
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+          const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+          const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+          const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+          const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+          const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+          const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+          const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+          const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+          const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+          // dct_const_round_shift
+          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+          const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+          const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+          const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+          const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+          const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+          const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+          const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+          const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+          // Combine
+          res1 = _mm_packs_epi32(w0, w1);
+          res7 = _mm_packs_epi32(w2, w3);
+          res5 = _mm_packs_epi32(w4, w5);
+          res3 = _mm_packs_epi32(w6, w7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&res1, &res7, &res5, &res3);
+          if (overflow) {
+            vpx_highbd_fdct8x8_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+    // store results
+    store_output(&in0, (output + 0 * 8));
+    store_output(&in1, (output + 1 * 8));
+    store_output(&in2, (output + 2 * 8));
+    store_output(&in3, (output + 3 * 8));
+    store_output(&in4, (output + 4 * 8));
+    store_output(&in5, (output + 5 * 8));
+    store_output(&in6, (output + 6 * 8));
+    store_output(&in7, (output + 7 * 8));
+  }
+}
+
+void FDCT16x16_2D(const int16_t *input, tran_low_t *output, int stride) {
+  // The 2D transform is done with two passes which are actually pretty
+  // similar. In the first one, we transform the columns and transpose
+  // the results. In the second one, we transform the rows. To achieve that,
+  // as the first pass results are transposed, we transpose the columns (that
+  // is the transposed rows) and transpose the results (so that it goes back
+  // in normal/row positions).
+  int pass;
+  // We need an intermediate buffer between passes.
+  DECLARE_ALIGNED(16, int16_t, intermediate[256]);
+  const int16_t *in = input;
+  int16_t *out0 = intermediate;
+  tran_low_t *out1 = output;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
+  const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
+  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
+  const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
+  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kOne = _mm_set1_epi16(1);
+  // Do the two transform/transpose passes
+  for (pass = 0; pass < 2; ++pass) {
+    // We process eight columns (transposed rows in second pass) at a time.
+    int column_start;
+#if DCT_HIGH_BIT_DEPTH
+    int overflow;
+#endif
+    for (column_start = 0; column_start < 16; column_start += 8) {
+      __m128i in00, in01, in02, in03, in04, in05, in06, in07;
+      __m128i in08, in09, in10, in11, in12, in13, in14, in15;
+      __m128i input0, input1, input2, input3, input4, input5, input6, input7;
+      __m128i step1_0, step1_1, step1_2, step1_3;
+      __m128i step1_4, step1_5, step1_6, step1_7;
+      __m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6;
+      __m128i step3_0, step3_1, step3_2, step3_3;
+      __m128i step3_4, step3_5, step3_6, step3_7;
+      __m128i res00, res01, res02, res03, res04, res05, res06, res07;
+      __m128i res08, res09, res10, res11, res12, res13, res14, res15;
+      // Load and pre-condition input.
+      if (0 == pass) {
+        in00  = _mm_load_si128((const __m128i *)(in +  0 * stride));
+        in01  = _mm_load_si128((const __m128i *)(in +  1 * stride));
+        in02  = _mm_load_si128((const __m128i *)(in +  2 * stride));
+        in03  = _mm_load_si128((const __m128i *)(in +  3 * stride));
+        in04  = _mm_load_si128((const __m128i *)(in +  4 * stride));
+        in05  = _mm_load_si128((const __m128i *)(in +  5 * stride));
+        in06  = _mm_load_si128((const __m128i *)(in +  6 * stride));
+        in07  = _mm_load_si128((const __m128i *)(in +  7 * stride));
+        in08  = _mm_load_si128((const __m128i *)(in +  8 * stride));
+        in09  = _mm_load_si128((const __m128i *)(in +  9 * stride));
+        in10  = _mm_load_si128((const __m128i *)(in + 10 * stride));
+        in11  = _mm_load_si128((const __m128i *)(in + 11 * stride));
+        in12  = _mm_load_si128((const __m128i *)(in + 12 * stride));
+        in13  = _mm_load_si128((const __m128i *)(in + 13 * stride));
+        in14  = _mm_load_si128((const __m128i *)(in + 14 * stride));
+        in15  = _mm_load_si128((const __m128i *)(in + 15 * stride));
+        // x = x << 2
+        in00 = _mm_slli_epi16(in00, 2);
+        in01 = _mm_slli_epi16(in01, 2);
+        in02 = _mm_slli_epi16(in02, 2);
+        in03 = _mm_slli_epi16(in03, 2);
+        in04 = _mm_slli_epi16(in04, 2);
+        in05 = _mm_slli_epi16(in05, 2);
+        in06 = _mm_slli_epi16(in06, 2);
+        in07 = _mm_slli_epi16(in07, 2);
+        in08 = _mm_slli_epi16(in08, 2);
+        in09 = _mm_slli_epi16(in09, 2);
+        in10 = _mm_slli_epi16(in10, 2);
+        in11 = _mm_slli_epi16(in11, 2);
+        in12 = _mm_slli_epi16(in12, 2);
+        in13 = _mm_slli_epi16(in13, 2);
+        in14 = _mm_slli_epi16(in14, 2);
+        in15 = _mm_slli_epi16(in15, 2);
+      } else {
+        in00  = _mm_load_si128((const __m128i *)(in +  0 * 16));
+        in01  = _mm_load_si128((const __m128i *)(in +  1 * 16));
+        in02  = _mm_load_si128((const __m128i *)(in +  2 * 16));
+        in03  = _mm_load_si128((const __m128i *)(in +  3 * 16));
+        in04  = _mm_load_si128((const __m128i *)(in +  4 * 16));
+        in05  = _mm_load_si128((const __m128i *)(in +  5 * 16));
+        in06  = _mm_load_si128((const __m128i *)(in +  6 * 16));
+        in07  = _mm_load_si128((const __m128i *)(in +  7 * 16));
+        in08  = _mm_load_si128((const __m128i *)(in +  8 * 16));
+        in09  = _mm_load_si128((const __m128i *)(in +  9 * 16));
+        in10  = _mm_load_si128((const __m128i *)(in + 10 * 16));
+        in11  = _mm_load_si128((const __m128i *)(in + 11 * 16));
+        in12  = _mm_load_si128((const __m128i *)(in + 12 * 16));
+        in13  = _mm_load_si128((const __m128i *)(in + 13 * 16));
+        in14  = _mm_load_si128((const __m128i *)(in + 14 * 16));
+        in15  = _mm_load_si128((const __m128i *)(in + 15 * 16));
+        // x = (x + 1) >> 2
+        in00 = _mm_add_epi16(in00, kOne);
+        in01 = _mm_add_epi16(in01, kOne);
+        in02 = _mm_add_epi16(in02, kOne);
+        in03 = _mm_add_epi16(in03, kOne);
+        in04 = _mm_add_epi16(in04, kOne);
+        in05 = _mm_add_epi16(in05, kOne);
+        in06 = _mm_add_epi16(in06, kOne);
+        in07 = _mm_add_epi16(in07, kOne);
+        in08 = _mm_add_epi16(in08, kOne);
+        in09 = _mm_add_epi16(in09, kOne);
+        in10 = _mm_add_epi16(in10, kOne);
+        in11 = _mm_add_epi16(in11, kOne);
+        in12 = _mm_add_epi16(in12, kOne);
+        in13 = _mm_add_epi16(in13, kOne);
+        in14 = _mm_add_epi16(in14, kOne);
+        in15 = _mm_add_epi16(in15, kOne);
+        in00 = _mm_srai_epi16(in00, 2);
+        in01 = _mm_srai_epi16(in01, 2);
+        in02 = _mm_srai_epi16(in02, 2);
+        in03 = _mm_srai_epi16(in03, 2);
+        in04 = _mm_srai_epi16(in04, 2);
+        in05 = _mm_srai_epi16(in05, 2);
+        in06 = _mm_srai_epi16(in06, 2);
+        in07 = _mm_srai_epi16(in07, 2);
+        in08 = _mm_srai_epi16(in08, 2);
+        in09 = _mm_srai_epi16(in09, 2);
+        in10 = _mm_srai_epi16(in10, 2);
+        in11 = _mm_srai_epi16(in11, 2);
+        in12 = _mm_srai_epi16(in12, 2);
+        in13 = _mm_srai_epi16(in13, 2);
+        in14 = _mm_srai_epi16(in14, 2);
+        in15 = _mm_srai_epi16(in15, 2);
+      }
+      in += 8;
+      // Calculate input for the first 8 results.
+      {
+        input0 = ADD_EPI16(in00, in15);
+        input1 = ADD_EPI16(in01, in14);
+        input2 = ADD_EPI16(in02, in13);
+        input3 = ADD_EPI16(in03, in12);
+        input4 = ADD_EPI16(in04, in11);
+        input5 = ADD_EPI16(in05, in10);
+        input6 = ADD_EPI16(in06, in09);
+        input7 = ADD_EPI16(in07, in08);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&input0, &input1, &input2, &input3,
+                                           &input4, &input5, &input6, &input7);
+        if (overflow) {
+          vpx_highbd_fdct16x16_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Calculate input for the next 8 results.
+      {
+        step1_0 = SUB_EPI16(in07, in08);
+        step1_1 = SUB_EPI16(in06, in09);
+        step1_2 = SUB_EPI16(in05, in10);
+        step1_3 = SUB_EPI16(in04, in11);
+        step1_4 = SUB_EPI16(in03, in12);
+        step1_5 = SUB_EPI16(in02, in13);
+        step1_6 = SUB_EPI16(in01, in14);
+        step1_7 = SUB_EPI16(in00, in15);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step1_0, &step1_1,
+                                           &step1_2, &step1_3,
+                                           &step1_4, &step1_5,
+                                           &step1_6, &step1_7);
+        if (overflow) {
+          vpx_highbd_fdct16x16_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Work on the first eight values; fdct8(input, even_results);
+      {
+        // Add/subtract
+        const __m128i q0 = ADD_EPI16(input0, input7);
+        const __m128i q1 = ADD_EPI16(input1, input6);
+        const __m128i q2 = ADD_EPI16(input2, input5);
+        const __m128i q3 = ADD_EPI16(input3, input4);
+        const __m128i q4 = SUB_EPI16(input3, input4);
+        const __m128i q5 = SUB_EPI16(input2, input5);
+        const __m128i q6 = SUB_EPI16(input1, input6);
+        const __m128i q7 = SUB_EPI16(input0, input7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&q0, &q1, &q2, &q3,
+                                           &q4, &q5, &q6, &q7);
+        if (overflow) {
+          vpx_highbd_fdct16x16_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        // Work on first four results
+        {
+          // Add/subtract
+          const __m128i r0 = ADD_EPI16(q0, q3);
+          const __m128i r1 = ADD_EPI16(q1, q2);
+          const __m128i r2 = SUB_EPI16(q1, q2);
+          const __m128i r3 = SUB_EPI16(q0, q3);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3);
+          if (overflow) {
+            vpx_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          // Interleave to do the multiply by constants which gets us
+          // into 32 bits.
+          {
+            const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+            const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+            const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+            const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+            res00 = mult_round_shift(&t0, &t1, &k__cospi_p16_p16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+            res08 = mult_round_shift(&t0, &t1, &k__cospi_p16_m16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+            res04 = mult_round_shift(&t2, &t3, &k__cospi_p24_p08,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+            res12 = mult_round_shift(&t2, &t3, &k__cospi_m08_p24,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+            overflow = check_epi16_overflow_x4(&res00, &res08, &res04, &res12);
+            if (overflow) {
+              vpx_highbd_fdct16x16_c(input, output, stride);
+              return;
+            }
+#endif  // DCT_HIGH_BIT_DEPTH
+          }
+        }
+        // Work on next four results
+        {
+          // Interleave to do the multiply by constants which gets us
+          // into 32 bits.
+          const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
+          const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
+          const __m128i r0 = mult_round_shift(&d0, &d1, &k__cospi_p16_m16,
+                                              &k__DCT_CONST_ROUNDING,
+                                              DCT_CONST_BITS);
+          const __m128i r1 = mult_round_shift(&d0, &d1, &k__cospi_p16_p16,
+                                              &k__DCT_CONST_ROUNDING,
+                                              DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x2(&r0, &r1);
+          if (overflow) {
+            vpx_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          {
+            // Add/subtract
+            const __m128i x0 = ADD_EPI16(q4, r0);
+            const __m128i x1 = SUB_EPI16(q4, r0);
+            const __m128i x2 = SUB_EPI16(q7, r1);
+            const __m128i x3 = ADD_EPI16(q7, r1);
+#if DCT_HIGH_BIT_DEPTH
+            overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3);
+            if (overflow) {
+              vpx_highbd_fdct16x16_c(input, output, stride);
+              return;
+            }
+#endif  // DCT_HIGH_BIT_DEPTH
+            // Interleave to do the multiply by constants which gets us
+            // into 32 bits.
+            {
+              const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+              const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+              const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+              const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+              res02 = mult_round_shift(&t0, &t1, &k__cospi_p28_p04,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+              res14 = mult_round_shift(&t0, &t1, &k__cospi_m04_p28,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+              res10 = mult_round_shift(&t2, &t3, &k__cospi_p12_p20,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+              res06 = mult_round_shift(&t2, &t3, &k__cospi_m20_p12,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+              overflow = check_epi16_overflow_x4(&res02, &res14,
+                                                 &res10, &res06);
+              if (overflow) {
+                vpx_highbd_fdct16x16_c(input, output, stride);
+                return;
+              }
+#endif  // DCT_HIGH_BIT_DEPTH
+            }
+          }
+        }
+      }
+      // Work on the next eight values; step1 -> odd_results
+      {
+        // step 2
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2);
+          const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2);
+          const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3);
+          const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3);
+          step2_2 = mult_round_shift(&t0, &t1, &k__cospi_p16_m16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_3 = mult_round_shift(&t2, &t3, &k__cospi_p16_m16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_5 = mult_round_shift(&t0, &t1, &k__cospi_p16_p16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_4 = mult_round_shift(&t2, &t3, &k__cospi_p16_p16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&step2_2, &step2_3, &step2_5,
+                                             &step2_4);
+          if (overflow) {
+            vpx_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 3
+        {
+          step3_0 = ADD_EPI16(step1_0, step2_3);
+          step3_1 = ADD_EPI16(step1_1, step2_2);
+          step3_2 = SUB_EPI16(step1_1, step2_2);
+          step3_3 = SUB_EPI16(step1_0, step2_3);
+          step3_4 = SUB_EPI16(step1_7, step2_4);
+          step3_5 = SUB_EPI16(step1_6, step2_5);
+          step3_6 = ADD_EPI16(step1_6, step2_5);
+          step3_7 = ADD_EPI16(step1_7, step2_4);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step3_0, &step3_1,
+                                             &step3_2, &step3_3,
+                                             &step3_4, &step3_5,
+                                             &step3_6, &step3_7);
+          if (overflow) {
+            vpx_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 4
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6);
+          const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6);
+          const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5);
+          const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
+          step2_1 = mult_round_shift(&t0, &t1, &k__cospi_m08_p24,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_2 = mult_round_shift(&t2, &t3, &k__cospi_p24_p08,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_6 = mult_round_shift(&t0, &t1, &k__cospi_p24_p08,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_5 = mult_round_shift(&t2, &t3, &k__cospi_p08_m24,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&step2_1, &step2_2, &step2_6,
+                                             &step2_5);
+          if (overflow) {
+            vpx_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 5
+        {
+          step1_0 = ADD_EPI16(step3_0, step2_1);
+          step1_1 = SUB_EPI16(step3_0, step2_1);
+          step1_2 = ADD_EPI16(step3_3, step2_2);
+          step1_3 = SUB_EPI16(step3_3, step2_2);
+          step1_4 = SUB_EPI16(step3_4, step2_5);
+          step1_5 = ADD_EPI16(step3_4, step2_5);
+          step1_6 = SUB_EPI16(step3_7, step2_6);
+          step1_7 = ADD_EPI16(step3_7, step2_6);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1_0, &step1_1,
+                                             &step1_2, &step1_3,
+                                             &step1_4, &step1_5,
+                                             &step1_6, &step1_7);
+          if (overflow) {
+            vpx_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 6
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7);
+          const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7);
+          const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6);
+          const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6);
+          res01 = mult_round_shift(&t0, &t1, &k__cospi_p30_p02,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res09 = mult_round_shift(&t2, &t3, &k__cospi_p14_p18,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res15 = mult_round_shift(&t0, &t1, &k__cospi_m02_p30,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res07 = mult_round_shift(&t2, &t3, &k__cospi_m18_p14,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&res01, &res09, &res15, &res07);
+          if (overflow) {
+            vpx_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5);
+          const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5);
+          const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4);
+          const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4);
+          res05 = mult_round_shift(&t0, &t1, &k__cospi_p22_p10,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res13 = mult_round_shift(&t2, &t3, &k__cospi_p06_p26,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res11 = mult_round_shift(&t0, &t1, &k__cospi_m10_p22,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res03 = mult_round_shift(&t2, &t3, &k__cospi_m26_p06,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&res05, &res13, &res11, &res03);
+          if (overflow) {
+            vpx_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+      // Transpose the results, do it as two 8x8 transposes.
+      transpose_and_output8x8(&res00, &res01, &res02, &res03,
+                              &res04, &res05, &res06, &res07,
+                              pass, out0, out1);
+      transpose_and_output8x8(&res08, &res09, &res10, &res11,
+                              &res12, &res13, &res14, &res15,
+                              pass, out0 + 8, out1 + 8);
+      if (pass == 0) {
+        out0 += 8*16;
+      } else {
+        out1 += 8*16;
+      }
+    }
+    // Setup in/out for next pass.
+    in = intermediate;
+  }
+}
+
+#undef ADD_EPI16
+#undef SUB_EPI16
diff --git a/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_sse2.c b/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_sse2.c
new file mode 100644
index 0000000..3e4f49b
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_sse2.c
@@ -0,0 +1,272 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/x86/fwd_txfm_sse2.h"
+
+void vpx_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride) {
+  __m128i in0, in1;
+  __m128i tmp;
+  const __m128i zero = _mm_setzero_si128();
+  in0  = _mm_loadl_epi64((const __m128i *)(input +  0 * stride));
+  in1  = _mm_loadl_epi64((const __m128i *)(input +  1 * stride));
+  in1  = _mm_unpacklo_epi64(in1, _mm_loadl_epi64((const __m128i *)
+         (input +  2 * stride)));
+  in0  = _mm_unpacklo_epi64(in0, _mm_loadl_epi64((const __m128i *)
+         (input +  3 * stride)));
+
+  tmp = _mm_add_epi16(in0, in1);
+  in0 = _mm_unpacklo_epi16(zero, tmp);
+  in1 = _mm_unpackhi_epi16(zero, tmp);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  tmp = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(tmp, zero);
+  in1 = _mm_unpackhi_epi32(tmp, zero);
+
+  tmp = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(tmp, 8);
+
+  in1 = _mm_add_epi32(tmp, in0);
+  in0 = _mm_slli_epi32(in1, 1);
+  output[0] = (tran_low_t)_mm_cvtsi128_si32(in0);
+}
+
+void vpx_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride) {
+  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i u0, u1, sum;
+
+  u0 = _mm_add_epi16(in0, in1);
+  u1 = _mm_add_epi16(in2, in3);
+
+  in0  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  in1  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  in2  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  in3  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+
+  sum = _mm_add_epi16(u0, u1);
+
+  in0 = _mm_add_epi16(in0, in1);
+  in2 = _mm_add_epi16(in2, in3);
+  sum = _mm_add_epi16(sum, in0);
+
+  u0  = _mm_setzero_si128();
+  sum = _mm_add_epi16(sum, in2);
+
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  output[0] = (tran_low_t)_mm_cvtsi128_si32(in1);
+}
+
+void vpx_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  __m128i in0, in1, in2, in3;
+  __m128i u0, u1;
+  __m128i sum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 2; ++i) {
+    in0  = _mm_load_si128((const __m128i *)(input + 0 * stride + 0));
+    in1  = _mm_load_si128((const __m128i *)(input + 0 * stride + 8));
+    in2  = _mm_load_si128((const __m128i *)(input + 1 * stride + 0));
+    in3  = _mm_load_si128((const __m128i *)(input + 1 * stride + 8));
+
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input + 2 * stride + 0));
+    in1  = _mm_load_si128((const __m128i *)(input + 2 * stride + 8));
+    in2  = _mm_load_si128((const __m128i *)(input + 3 * stride + 0));
+    in3  = _mm_load_si128((const __m128i *)(input + 3 * stride + 8));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input + 4 * stride + 0));
+    in1  = _mm_load_si128((const __m128i *)(input + 4 * stride + 8));
+    in2  = _mm_load_si128((const __m128i *)(input + 5 * stride + 0));
+    in3  = _mm_load_si128((const __m128i *)(input + 5 * stride + 8));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input + 6 * stride + 0));
+    in1  = _mm_load_si128((const __m128i *)(input + 6 * stride + 8));
+    in2  = _mm_load_si128((const __m128i *)(input + 7 * stride + 0));
+    in3  = _mm_load_si128((const __m128i *)(input + 7 * stride + 8));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    sum = _mm_add_epi16(sum, u1);
+    input += 8 * stride;
+  }
+
+  u0  = _mm_setzero_si128();
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  in1 = _mm_srai_epi32(in1, 1);
+  output[0] = (tran_low_t)_mm_cvtsi128_si32(in1);
+}
+
+void vpx_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  __m128i in0, in1, in2, in3;
+  __m128i u0, u1;
+  __m128i sum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 8; ++i) {
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    sum = _mm_add_epi16(sum, u1);
+  }
+
+  u0  = _mm_setzero_si128();
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  in1 = _mm_srai_epi32(in1, 3);
+  output[0] = (tran_low_t)_mm_cvtsi128_si32(in1);
+}
+
+#define DCT_HIGH_BIT_DEPTH 0
+#define FDCT4x4_2D vpx_fdct4x4_sse2
+#define FDCT8x8_2D vpx_fdct8x8_sse2
+#define FDCT16x16_2D vpx_fdct16x16_sse2
+#include "vpx_dsp/x86/fwd_txfm_impl_sse2.h"
+#undef  FDCT4x4_2D
+#undef  FDCT8x8_2D
+#undef  FDCT16x16_2D
+
+#define FDCT32x32_2D vpx_fdct32x32_rd_sse2
+#define FDCT32x32_HIGH_PRECISION 0
+#include "vpx_dsp/x86/fwd_dct32x32_impl_sse2.h"
+#undef  FDCT32x32_2D
+#undef  FDCT32x32_HIGH_PRECISION
+
+#define FDCT32x32_2D vpx_fdct32x32_sse2
+#define FDCT32x32_HIGH_PRECISION 1
+#include "vpx_dsp/x86/fwd_dct32x32_impl_sse2.h"  // NOLINT
+#undef  FDCT32x32_2D
+#undef  FDCT32x32_HIGH_PRECISION
+#undef  DCT_HIGH_BIT_DEPTH
+
+#if CONFIG_VP9_HIGHBITDEPTH
+#define DCT_HIGH_BIT_DEPTH 1
+#define FDCT4x4_2D vpx_highbd_fdct4x4_sse2
+#define FDCT8x8_2D vpx_highbd_fdct8x8_sse2
+#define FDCT16x16_2D vpx_highbd_fdct16x16_sse2
+#include "vpx_dsp/x86/fwd_txfm_impl_sse2.h" // NOLINT
+#undef  FDCT4x4_2D
+#undef  FDCT8x8_2D
+#undef  FDCT16x16_2D
+
+#define FDCT32x32_2D vpx_highbd_fdct32x32_rd_sse2
+#define FDCT32x32_HIGH_PRECISION 0
+#include "vpx_dsp/x86/fwd_dct32x32_impl_sse2.h" // NOLINT
+#undef  FDCT32x32_2D
+#undef  FDCT32x32_HIGH_PRECISION
+
+#define FDCT32x32_2D vpx_highbd_fdct32x32_sse2
+#define FDCT32x32_HIGH_PRECISION 1
+#include "vpx_dsp/x86/fwd_dct32x32_impl_sse2.h" // NOLINT
+#undef  FDCT32x32_2D
+#undef  FDCT32x32_HIGH_PRECISION
+#undef  DCT_HIGH_BIT_DEPTH
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_sse2.h b/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_sse2.h
new file mode 100644
index 0000000..94d5bef
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_sse2.h
@@ -0,0 +1,454 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_X86_FWD_TXFM_SSE2_H_
+#define VPX_DSP_X86_FWD_TXFM_SSE2_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define pair_set_epi32(a, b) \
+  _mm_set_epi32((int)(b), (int)(a), (int)(b), (int)(a))
+
+static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) {
+  __m128i buf0, buf1;
+  buf0 = _mm_mul_epu32(a, b);
+  a = _mm_srli_epi64(a, 32);
+  b = _mm_srli_epi64(b, 32);
+  buf1 = _mm_mul_epu32(a, b);
+  return _mm_add_epi64(buf0, buf1);
+}
+
+static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) {
+  __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
+  __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
+  return _mm_unpacklo_epi64(buf0, buf1);
+}
+
+static INLINE int check_epi16_overflow_x2(const __m128i *preg0,
+                                          const __m128i *preg1) {
+  const __m128i max_overflow = _mm_set1_epi16(0x7fff);
+  const __m128i min_overflow = _mm_set1_epi16(0x8000);
+  __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
+                              _mm_cmpeq_epi16(*preg0, min_overflow));
+  __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
+                              _mm_cmpeq_epi16(*preg1, min_overflow));
+  cmp0 = _mm_or_si128(cmp0, cmp1);
+  return _mm_movemask_epi8(cmp0);
+}
+
+static INLINE int check_epi16_overflow_x4(const __m128i *preg0,
+                                          const __m128i *preg1,
+                                          const __m128i *preg2,
+                                          const __m128i *preg3) {
+  const __m128i max_overflow = _mm_set1_epi16(0x7fff);
+  const __m128i min_overflow = _mm_set1_epi16(0x8000);
+  __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
+                              _mm_cmpeq_epi16(*preg0, min_overflow));
+  __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
+                              _mm_cmpeq_epi16(*preg1, min_overflow));
+  __m128i cmp2 = _mm_or_si128(_mm_cmpeq_epi16(*preg2, max_overflow),
+                              _mm_cmpeq_epi16(*preg2, min_overflow));
+  __m128i cmp3 = _mm_or_si128(_mm_cmpeq_epi16(*preg3, max_overflow),
+                              _mm_cmpeq_epi16(*preg3, min_overflow));
+  cmp0 = _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3));
+  return _mm_movemask_epi8(cmp0);
+}
+
+static INLINE int check_epi16_overflow_x8(const __m128i *preg0,
+                                          const __m128i *preg1,
+                                          const __m128i *preg2,
+                                          const __m128i *preg3,
+                                          const __m128i *preg4,
+                                          const __m128i *preg5,
+                                          const __m128i *preg6,
+                                          const __m128i *preg7) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x12(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *preg8,
+                                           const __m128i *preg9,
+                                           const __m128i *preg10,
+                                           const __m128i *preg11) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0)
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x16(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *preg8,
+                                           const __m128i *preg9,
+                                           const __m128i *preg10,
+                                           const __m128i *preg11,
+                                           const __m128i *preg12,
+                                           const __m128i *preg13,
+                                           const __m128i *preg14,
+                                           const __m128i *preg15) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) {
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+    if (!res1)
+      res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
+  }
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x32(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *preg8,
+                                           const __m128i *preg9,
+                                           const __m128i *preg10,
+                                           const __m128i *preg11,
+                                           const __m128i *preg12,
+                                           const __m128i *preg13,
+                                           const __m128i *preg14,
+                                           const __m128i *preg15,
+                                           const __m128i *preg16,
+                                           const __m128i *preg17,
+                                           const __m128i *preg18,
+                                           const __m128i *preg19,
+                                           const __m128i *preg20,
+                                           const __m128i *preg21,
+                                           const __m128i *preg22,
+                                           const __m128i *preg23,
+                                           const __m128i *preg24,
+                                           const __m128i *preg25,
+                                           const __m128i *preg26,
+                                           const __m128i *preg27,
+                                           const __m128i *preg28,
+                                           const __m128i *preg29,
+                                           const __m128i *preg30,
+                                           const __m128i *preg31) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) {
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+    if (!res1) {
+      res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
+      if (!res0) {
+        res0 = check_epi16_overflow_x4(preg16, preg17, preg18, preg19);
+        if (!res1) {
+          res1 = check_epi16_overflow_x4(preg20, preg21, preg22, preg23);
+          if (!res0) {
+            res0 = check_epi16_overflow_x4(preg24, preg25, preg26, preg27);
+            if (!res1)
+              res1 = check_epi16_overflow_x4(preg28, preg29, preg30, preg31);
+          }
+        }
+      }
+    }
+  }
+  return res0 + res1;
+}
+
+static INLINE int k_check_epi32_overflow_4(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *zero) {
+  __m128i minus_one = _mm_set1_epi32(-1);
+  // Check for overflows
+  __m128i reg0_shifted = _mm_slli_epi64(*preg0, 1);
+  __m128i reg1_shifted = _mm_slli_epi64(*preg1, 1);
+  __m128i reg2_shifted = _mm_slli_epi64(*preg2, 1);
+  __m128i reg3_shifted = _mm_slli_epi64(*preg3, 1);
+  __m128i reg0_top_dwords = _mm_shuffle_epi32(
+      reg0_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg1_top_dwords = _mm_shuffle_epi32(
+      reg1_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg2_top_dwords = _mm_shuffle_epi32(
+      reg2_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg3_top_dwords = _mm_shuffle_epi32(
+      reg3_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i top_dwords_01 = _mm_unpacklo_epi64(reg0_top_dwords, reg1_top_dwords);
+  __m128i top_dwords_23 = _mm_unpacklo_epi64(reg2_top_dwords, reg3_top_dwords);
+  __m128i valid_positve_01 = _mm_cmpeq_epi32(top_dwords_01, *zero);
+  __m128i valid_positve_23 = _mm_cmpeq_epi32(top_dwords_23, *zero);
+  __m128i valid_negative_01 = _mm_cmpeq_epi32(top_dwords_01, minus_one);
+  __m128i valid_negative_23 = _mm_cmpeq_epi32(top_dwords_23, minus_one);
+  int overflow_01 = _mm_movemask_epi8(
+      _mm_cmpeq_epi32(valid_positve_01, valid_negative_01));
+  int overflow_23 = _mm_movemask_epi8(
+      _mm_cmpeq_epi32(valid_positve_23, valid_negative_23));
+  return (overflow_01 + overflow_23);
+}
+
+static INLINE int k_check_epi32_overflow_8(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+  }
+  return overflow;
+}
+
+static INLINE int k_check_epi32_overflow_16(const __m128i *preg0,
+                                            const __m128i *preg1,
+                                            const __m128i *preg2,
+                                            const __m128i *preg3,
+                                            const __m128i *preg4,
+                                            const __m128i *preg5,
+                                            const __m128i *preg6,
+                                            const __m128i *preg7,
+                                            const __m128i *preg8,
+                                            const __m128i *preg9,
+                                            const __m128i *preg10,
+                                            const __m128i *preg11,
+                                            const __m128i *preg12,
+                                            const __m128i *preg13,
+                                            const __m128i *preg14,
+                                            const __m128i *preg15,
+                                            const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+    if (!overflow) {
+      overflow = k_check_epi32_overflow_4(preg8, preg9, preg10, preg11,
+                                          zero);
+      if (!overflow) {
+        overflow = k_check_epi32_overflow_4(preg12, preg13, preg14, preg15,
+                                            zero);
+      }
+    }
+  }
+  return overflow;
+}
+
+static INLINE int k_check_epi32_overflow_32(const __m128i *preg0,
+                                            const __m128i *preg1,
+                                            const __m128i *preg2,
+                                            const __m128i *preg3,
+                                            const __m128i *preg4,
+                                            const __m128i *preg5,
+                                            const __m128i *preg6,
+                                            const __m128i *preg7,
+                                            const __m128i *preg8,
+                                            const __m128i *preg9,
+                                            const __m128i *preg10,
+                                            const __m128i *preg11,
+                                            const __m128i *preg12,
+                                            const __m128i *preg13,
+                                            const __m128i *preg14,
+                                            const __m128i *preg15,
+                                            const __m128i *preg16,
+                                            const __m128i *preg17,
+                                            const __m128i *preg18,
+                                            const __m128i *preg19,
+                                            const __m128i *preg20,
+                                            const __m128i *preg21,
+                                            const __m128i *preg22,
+                                            const __m128i *preg23,
+                                            const __m128i *preg24,
+                                            const __m128i *preg25,
+                                            const __m128i *preg26,
+                                            const __m128i *preg27,
+                                            const __m128i *preg28,
+                                            const __m128i *preg29,
+                                            const __m128i *preg30,
+                                            const __m128i *preg31,
+                                            const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+    if (!overflow) {
+      overflow = k_check_epi32_overflow_4(preg8, preg9, preg10, preg11, zero);
+      if (!overflow) {
+        overflow = k_check_epi32_overflow_4(preg12, preg13, preg14, preg15,
+                                            zero);
+        if (!overflow) {
+          overflow = k_check_epi32_overflow_4(preg16, preg17, preg18, preg19,
+                                              zero);
+          if (!overflow) {
+            overflow = k_check_epi32_overflow_4(preg20, preg21,
+                                                preg22, preg23, zero);
+            if (!overflow) {
+              overflow = k_check_epi32_overflow_4(preg24, preg25,
+                                                  preg26, preg27, zero);
+              if (!overflow) {
+                overflow = k_check_epi32_overflow_4(preg28, preg29,
+                                                    preg30, preg31, zero);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  return overflow;
+}
+
+static INLINE void store_output(const __m128i *poutput, tran_low_t* dst_ptr) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  _mm_store_si128((__m128i *)(dst_ptr), out0);
+  _mm_store_si128((__m128i *)(dst_ptr + 4), out1);
+#else
+  _mm_store_si128((__m128i *)(dst_ptr), *poutput);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+static INLINE void storeu_output(const __m128i *poutput, tran_low_t* dst_ptr) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  _mm_storeu_si128((__m128i *)(dst_ptr), out0);
+  _mm_storeu_si128((__m128i *)(dst_ptr + 4), out1);
+#else
+  _mm_storeu_si128((__m128i *)(dst_ptr), *poutput);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+
+static INLINE __m128i mult_round_shift(const __m128i *pin0,
+                                       const __m128i *pin1,
+                                       const __m128i *pmultiplier,
+                                       const __m128i *prounding,
+                                       const int shift) {
+  const __m128i u0 = _mm_madd_epi16(*pin0, *pmultiplier);
+  const __m128i u1 = _mm_madd_epi16(*pin1, *pmultiplier);
+  const __m128i v0 = _mm_add_epi32(u0, *prounding);
+  const __m128i v1 = _mm_add_epi32(u1, *prounding);
+  const __m128i w0 = _mm_srai_epi32(v0, shift);
+  const __m128i w1 = _mm_srai_epi32(v1, shift);
+  return _mm_packs_epi32(w0, w1);
+}
+
+static INLINE void transpose_and_output8x8(
+    const __m128i *pin00, const __m128i *pin01,
+    const __m128i *pin02, const __m128i *pin03,
+    const __m128i *pin04, const __m128i *pin05,
+    const __m128i *pin06, const __m128i *pin07,
+    const int pass, int16_t* out0_ptr,
+    tran_low_t* out1_ptr) {
+  // 00 01 02 03 04 05 06 07
+  // 10 11 12 13 14 15 16 17
+  // 20 21 22 23 24 25 26 27
+  // 30 31 32 33 34 35 36 37
+  // 40 41 42 43 44 45 46 47
+  // 50 51 52 53 54 55 56 57
+  // 60 61 62 63 64 65 66 67
+  // 70 71 72 73 74 75 76 77
+  const __m128i tr0_0 = _mm_unpacklo_epi16(*pin00, *pin01);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(*pin02, *pin03);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(*pin00, *pin01);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(*pin02, *pin03);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(*pin04, *pin05);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(*pin06, *pin07);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(*pin04, *pin05);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(*pin06, *pin07);
+  // 00 10 01 11 02 12 03 13
+  // 20 30 21 31 22 32 23 33
+  // 04 14 05 15 06 16 07 17
+  // 24 34 25 35 26 36 27 37
+  // 40 50 41 51 42 52 43 53
+  // 60 70 61 71 62 72 63 73
+  // 54 54 55 55 56 56 57 57
+  // 64 74 65 75 66 76 67 77
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+  // 00 10 20 30 01 11 21 31
+  // 40 50 60 70 41 51 61 71
+  // 02 12 22 32 03 13 23 33
+  // 42 52 62 72 43 53 63 73
+  // 04 14 24 34 05 15 21 36
+  // 44 54 64 74 45 55 61 76
+  // 06 16 26 36 07 17 27 37
+  // 46 56 66 76 47 57 67 77
+  const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+  const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+  const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+  const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+  const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+  const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+  const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+  const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+  // 00 10 20 30 40 50 60 70
+  // 01 11 21 31 41 51 61 71
+  // 02 12 22 32 42 52 62 72
+  // 03 13 23 33 43 53 63 73
+  // 04 14 24 34 44 54 64 74
+  // 05 15 25 35 45 55 65 75
+  // 06 16 26 36 46 56 66 76
+  // 07 17 27 37 47 57 67 77
+  if (pass == 0) {
+    _mm_storeu_si128((__m128i*)(out0_ptr + 0 * 16), tr2_0);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 1 * 16), tr2_1);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 2 * 16), tr2_2);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 3 * 16), tr2_3);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 4 * 16), tr2_4);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 5 * 16), tr2_5);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 6 * 16), tr2_6);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 7 * 16), tr2_7);
+  } else {
+    storeu_output(&tr2_0, (out1_ptr + 0 * 16));
+    storeu_output(&tr2_1, (out1_ptr + 1 * 16));
+    storeu_output(&tr2_2, (out1_ptr + 2 * 16));
+    storeu_output(&tr2_3, (out1_ptr + 3 * 16));
+    storeu_output(&tr2_4, (out1_ptr + 4 * 16));
+    storeu_output(&tr2_5, (out1_ptr + 5 * 16));
+    storeu_output(&tr2_6, (out1_ptr + 6 * 16));
+    storeu_output(&tr2_7, (out1_ptr + 7 * 16));
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_DSP_X86_FWD_TXFM_SSE2_H_
diff --git a/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_ssse3_x86_64.asm b/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_ssse3_x86_64.asm
new file mode 100644
index 0000000..78a1dbb
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/fwd_txfm_ssse3_x86_64.asm
@@ -0,0 +1,183 @@
+;
+;  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the forward transformation. Part
+; of the macro definitions are originally derived from the ffmpeg project.
+; The current version applies to x86 64-bit only.
+
+SECTION_RODATA
+
+pw_11585x2: times 8 dw 23170
+pd_8192:    times 4 dd 8192
+
+%macro TRANSFORM_COEFFS 2
+pw_%1_%2:   dw  %1,  %2,  %1,  %2,  %1,  %2,  %1,  %2
+pw_%2_m%1:  dw  %2, -%1,  %2, -%1,  %2, -%1,  %2, -%1
+%endmacro
+
+TRANSFORM_COEFFS 11585,  11585
+TRANSFORM_COEFFS 15137,   6270
+TRANSFORM_COEFFS 16069,   3196
+TRANSFORM_COEFFS  9102,  13623
+
+SECTION .text
+
+%if ARCH_X86_64
+%macro SUM_SUB 3
+  psubw  m%3, m%1, m%2
+  paddw  m%1, m%2
+  SWAP    %2, %3
+%endmacro
+
+; butterfly operation
+%macro MUL_ADD_2X 6 ; dst1, dst2, src, round, coefs1, coefs2
+  pmaddwd            m%1, m%3, %5
+  pmaddwd            m%2, m%3, %6
+  paddd              m%1,  %4
+  paddd              m%2,  %4
+  psrad              m%1,  14
+  psrad              m%2,  14
+%endmacro
+
+%macro BUTTERFLY_4X 7 ; dst1, dst2, coef1, coef2, round, tmp1, tmp2
+  punpckhwd          m%6, m%2, m%1
+  MUL_ADD_2X         %7,  %6,  %6,  %5, [pw_%4_%3], [pw_%3_m%4]
+  punpcklwd          m%2, m%1
+  MUL_ADD_2X         %1,  %2,  %2,  %5, [pw_%4_%3], [pw_%3_m%4]
+  packssdw           m%1, m%7
+  packssdw           m%2, m%6
+%endmacro
+
+; matrix transpose
+%macro INTERLEAVE_2X 4
+  punpckh%1          m%4, m%2, m%3
+  punpckl%1          m%2, m%3
+  SWAP               %3,  %4
+%endmacro
+
+%macro TRANSPOSE8X8 9
+  INTERLEAVE_2X  wd, %1, %2, %9
+  INTERLEAVE_2X  wd, %3, %4, %9
+  INTERLEAVE_2X  wd, %5, %6, %9
+  INTERLEAVE_2X  wd, %7, %8, %9
+
+  INTERLEAVE_2X  dq, %1, %3, %9
+  INTERLEAVE_2X  dq, %2, %4, %9
+  INTERLEAVE_2X  dq, %5, %7, %9
+  INTERLEAVE_2X  dq, %6, %8, %9
+
+  INTERLEAVE_2X  qdq, %1, %5, %9
+  INTERLEAVE_2X  qdq, %3, %7, %9
+  INTERLEAVE_2X  qdq, %2, %6, %9
+  INTERLEAVE_2X  qdq, %4, %8, %9
+
+  SWAP  %2, %5
+  SWAP  %4, %7
+%endmacro
+
+; 1D forward 8x8 DCT transform
+%macro FDCT8_1D 1
+  SUM_SUB            0,  7,  9
+  SUM_SUB            1,  6,  9
+  SUM_SUB            2,  5,  9
+  SUM_SUB            3,  4,  9
+
+  SUM_SUB            0,  3,  9
+  SUM_SUB            1,  2,  9
+  SUM_SUB            6,  5,  9
+%if %1 == 0
+  SUM_SUB            0,  1,  9
+%endif
+
+  BUTTERFLY_4X       2,  3,  6270,  15137,  m8,  9,  10
+
+  pmulhrsw           m6, m12
+  pmulhrsw           m5, m12
+%if %1 == 0
+  pmulhrsw           m0, m12
+  pmulhrsw           m1, m12
+%else
+  BUTTERFLY_4X       1,  0,  11585, 11585,  m8,  9,  10
+  SWAP               0,  1
+%endif
+
+  SUM_SUB            4,  5,  9
+  SUM_SUB            7,  6,  9
+  BUTTERFLY_4X       4,  7,  3196,  16069,  m8,  9,  10
+  BUTTERFLY_4X       5,  6,  13623,  9102,  m8,  9,  10
+  SWAP               1,  4
+  SWAP               3,  6
+%endmacro
+
+%macro DIVIDE_ROUND_2X 4 ; dst1, dst2, tmp1, tmp2
+  psraw              m%3, m%1, 15
+  psraw              m%4, m%2, 15
+  psubw              m%1, m%3
+  psubw              m%2, m%4
+  psraw              m%1, 1
+  psraw              m%2, 1
+%endmacro
+
+INIT_XMM ssse3
+cglobal fdct8x8, 3, 5, 13, input, output, stride
+
+  mova               m8, [pd_8192]
+  mova              m12, [pw_11585x2]
+  pxor              m11, m11
+
+  lea                r3, [2 * strideq]
+  lea                r4, [4 * strideq]
+  mova               m0, [inputq]
+  mova               m1, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m2, [inputq]
+  mova               m3, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m4, [inputq]
+  mova               m5, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m6, [inputq]
+  mova               m7, [inputq + r3]
+
+  ; left shift by 2 to increase forward transformation precision
+  psllw              m0, 2
+  psllw              m1, 2
+  psllw              m2, 2
+  psllw              m3, 2
+  psllw              m4, 2
+  psllw              m5, 2
+  psllw              m6, 2
+  psllw              m7, 2
+
+  ; column transform
+  FDCT8_1D  0
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  FDCT8_1D  1
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  DIVIDE_ROUND_2X   0, 1, 9, 10
+  DIVIDE_ROUND_2X   2, 3, 9, 10
+  DIVIDE_ROUND_2X   4, 5, 9, 10
+  DIVIDE_ROUND_2X   6, 7, 9, 10
+
+  mova              [outputq +   0], m0
+  mova              [outputq +  16], m1
+  mova              [outputq +  32], m2
+  mova              [outputq +  48], m3
+  mova              [outputq +  64], m4
+  mova              [outputq +  80], m5
+  mova              [outputq +  96], m6
+  mova              [outputq + 112], m7
+
+  RET
+%endif
diff --git a/libvpx/libvpx/vpx_dsp/x86/halfpix_variance_impl_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/halfpix_variance_impl_sse2.asm
new file mode 100644
index 0000000..cc26bb6
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/halfpix_variance_impl_sse2.asm
@@ -0,0 +1,346 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vpx_half_horiz_vert_variance16x_h_sse2(unsigned char *ref,
+;                                            int ref_stride,
+;                                            unsigned char *src,
+;                                            int src_stride,
+;                                            unsigned int height,
+;                                            int *sum,
+;                                            unsigned int *sumsquared)
+global sym(vpx_half_horiz_vert_variance16x_h_sse2) PRIVATE
+sym(vpx_half_horiz_vert_variance16x_h_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        pxor            xmm6,           xmm6                ;  error accumulator
+        pxor            xmm7,           xmm7                ;  sse eaccumulator
+        mov             rsi,            arg(0) ;ref
+
+        mov             rdi,            arg(2) ;src
+        movsxd          rcx,            dword ptr arg(4) ;height
+        movsxd          rax,            dword ptr arg(1) ;ref_stride
+        movsxd          rdx,            dword ptr arg(3)    ;src_stride
+
+        pxor            xmm0,           xmm0                ;
+
+        movdqu          xmm5,           XMMWORD PTR [rsi]
+        movdqu          xmm3,           XMMWORD PTR [rsi+1]
+        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3) horizontal line 1
+
+        lea             rsi,            [rsi + rax]
+
+vpx_half_horiz_vert_variance16x_h_1:
+        movdqu          xmm1,           XMMWORD PTR [rsi]     ;
+        movdqu          xmm2,           XMMWORD PTR [rsi+1]   ;
+        pavgb           xmm1,           xmm2                ;  xmm1 = avg(xmm1,xmm3) horizontal line i+1
+
+        pavgb           xmm5,           xmm1                ;  xmm = vertical average of the above
+
+        movdqa          xmm4,           xmm5
+        punpcklbw       xmm5,           xmm0                ;  xmm5 = words of above
+        punpckhbw       xmm4,           xmm0
+
+        movq            xmm3,           QWORD PTR [rdi]     ;  xmm3 = d0,d1,d2..d7
+        punpcklbw       xmm3,           xmm0                ;  xmm3 = words of above
+        psubw           xmm5,           xmm3                ;  xmm5 -= xmm3
+
+        movq            xmm3,           QWORD PTR [rdi+8]
+        punpcklbw       xmm3,           xmm0
+        psubw           xmm4,           xmm3
+
+        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
+        paddw           xmm6,           xmm4
+        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
+        pmaddwd         xmm4,           xmm4
+        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
+        paddd           xmm7,           xmm4
+
+        movdqa          xmm5,           xmm1                ;  save xmm1 for use on the next row
+
+        lea             rsi,            [rsi + rax]
+        lea             rdi,            [rdi + rdx]
+
+        sub             rcx,            1                   ;
+        jnz             vpx_half_horiz_vert_variance16x_h_1     ;
+
+        pxor        xmm1,           xmm1
+        pxor        xmm5,           xmm5
+
+        punpcklwd   xmm0,           xmm6
+        punpckhwd   xmm1,           xmm6
+        psrad       xmm0,           16
+        psrad       xmm1,           16
+        paddd       xmm0,           xmm1
+        movdqa      xmm1,           xmm0
+
+        movdqa      xmm6,           xmm7
+        punpckldq   xmm6,           xmm5
+        punpckhdq   xmm7,           xmm5
+        paddd       xmm6,           xmm7
+
+        punpckldq   xmm0,           xmm5
+        punpckhdq   xmm1,           xmm5
+        paddd       xmm0,           xmm1
+
+        movdqa      xmm7,           xmm6
+        movdqa      xmm1,           xmm0
+
+        psrldq      xmm7,           8
+        psrldq      xmm1,           8
+
+        paddd       xmm6,           xmm7
+        paddd       xmm0,           xmm1
+
+        mov         rsi,            arg(5) ;[Sum]
+        mov         rdi,            arg(6) ;[SSE]
+
+        movd        [rsi],       xmm0
+        movd        [rdi],       xmm6
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vpx_half_vert_variance16x_h_sse2(unsigned char *ref,
+;                                      int ref_stride,
+;                                      unsigned char *src,
+;                                      int src_stride,
+;                                      unsigned int height,
+;                                      int *sum,
+;                                      unsigned int *sumsquared)
+global sym(vpx_half_vert_variance16x_h_sse2) PRIVATE
+sym(vpx_half_vert_variance16x_h_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        pxor            xmm6,           xmm6                ;  error accumulator
+        pxor            xmm7,           xmm7                ;  sse eaccumulator
+        mov             rsi,            arg(0)              ;ref
+
+        mov             rdi,            arg(2)              ;src
+        movsxd          rcx,            dword ptr arg(4)    ;height
+        movsxd          rax,            dword ptr arg(1)    ;ref_stride
+        movsxd          rdx,            dword ptr arg(3)    ;src_stride
+
+        movdqu          xmm5,           XMMWORD PTR [rsi]
+        lea             rsi,            [rsi + rax          ]
+        pxor            xmm0,           xmm0
+
+vpx_half_vert_variance16x_h_1:
+        movdqu          xmm3,           XMMWORD PTR [rsi]
+
+        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3)
+        movdqa          xmm4,           xmm5
+        punpcklbw       xmm5,           xmm0
+        punpckhbw       xmm4,           xmm0
+
+        movq            xmm2,           QWORD PTR [rdi]
+        punpcklbw       xmm2,           xmm0
+        psubw           xmm5,           xmm2
+        movq            xmm2,           QWORD PTR [rdi+8]
+        punpcklbw       xmm2,           xmm0
+        psubw           xmm4,           xmm2
+
+        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
+        paddw           xmm6,           xmm4
+        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
+        pmaddwd         xmm4,           xmm4
+        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
+        paddd           xmm7,           xmm4
+
+        movdqa          xmm5,           xmm3
+
+        lea             rsi,            [rsi + rax]
+        lea             rdi,            [rdi + rdx]
+
+        sub             rcx,            1
+        jnz             vpx_half_vert_variance16x_h_1
+
+        pxor        xmm1,           xmm1
+        pxor        xmm5,           xmm5
+
+        punpcklwd   xmm0,           xmm6
+        punpckhwd   xmm1,           xmm6
+        psrad       xmm0,           16
+        psrad       xmm1,           16
+        paddd       xmm0,           xmm1
+        movdqa      xmm1,           xmm0
+
+        movdqa      xmm6,           xmm7
+        punpckldq   xmm6,           xmm5
+        punpckhdq   xmm7,           xmm5
+        paddd       xmm6,           xmm7
+
+        punpckldq   xmm0,           xmm5
+        punpckhdq   xmm1,           xmm5
+        paddd       xmm0,           xmm1
+
+        movdqa      xmm7,           xmm6
+        movdqa      xmm1,           xmm0
+
+        psrldq      xmm7,           8
+        psrldq      xmm1,           8
+
+        paddd       xmm6,           xmm7
+        paddd       xmm0,           xmm1
+
+        mov         rsi,            arg(5) ;[Sum]
+        mov         rdi,            arg(6) ;[SSE]
+
+        movd        [rsi],       xmm0
+        movd        [rdi],       xmm6
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vpx_half_horiz_variance16x_h_sse2(unsigned char *ref,
+;                                       int ref_stride
+;                                       unsigned char *src,
+;                                       int src_stride,
+;                                       unsigned int height,
+;                                       int *sum,
+;                                       unsigned int *sumsquared)
+global sym(vpx_half_horiz_variance16x_h_sse2) PRIVATE
+sym(vpx_half_horiz_variance16x_h_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        pxor            xmm6,           xmm6                ;  error accumulator
+        pxor            xmm7,           xmm7                ;  sse eaccumulator
+        mov             rsi,            arg(0) ;ref
+
+        mov             rdi,            arg(2) ;src
+        movsxd          rcx,            dword ptr arg(4) ;height
+        movsxd          rax,            dword ptr arg(1) ;ref_stride
+        movsxd          rdx,            dword ptr arg(3)    ;src_stride
+
+        pxor            xmm0,           xmm0                ;
+
+vpx_half_horiz_variance16x_h_1:
+        movdqu          xmm5,           XMMWORD PTR [rsi]     ;  xmm5 = s0,s1,s2..s15
+        movdqu          xmm3,           XMMWORD PTR [rsi+1]   ;  xmm3 = s1,s2,s3..s16
+
+        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3)
+        movdqa          xmm1,           xmm5
+        punpcklbw       xmm5,           xmm0                ;  xmm5 = words of above
+        punpckhbw       xmm1,           xmm0
+
+        movq            xmm3,           QWORD PTR [rdi]     ;  xmm3 = d0,d1,d2..d7
+        punpcklbw       xmm3,           xmm0                ;  xmm3 = words of above
+        movq            xmm2,           QWORD PTR [rdi+8]
+        punpcklbw       xmm2,           xmm0
+
+        psubw           xmm5,           xmm3                ;  xmm5 -= xmm3
+        psubw           xmm1,           xmm2
+        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
+        paddw           xmm6,           xmm1
+        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
+        pmaddwd         xmm1,           xmm1
+        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
+        paddd           xmm7,           xmm1
+
+        lea             rsi,            [rsi + rax]
+        lea             rdi,            [rdi + rdx]
+
+        sub             rcx,            1                   ;
+        jnz             vpx_half_horiz_variance16x_h_1        ;
+
+        pxor        xmm1,           xmm1
+        pxor        xmm5,           xmm5
+
+        punpcklwd   xmm0,           xmm6
+        punpckhwd   xmm1,           xmm6
+        psrad       xmm0,           16
+        psrad       xmm1,           16
+        paddd       xmm0,           xmm1
+        movdqa      xmm1,           xmm0
+
+        movdqa      xmm6,           xmm7
+        punpckldq   xmm6,           xmm5
+        punpckhdq   xmm7,           xmm5
+        paddd       xmm6,           xmm7
+
+        punpckldq   xmm0,           xmm5
+        punpckhdq   xmm1,           xmm5
+        paddd       xmm0,           xmm1
+
+        movdqa      xmm7,           xmm6
+        movdqa      xmm1,           xmm0
+
+        psrldq      xmm7,           8
+        psrldq      xmm1,           8
+
+        paddd       xmm6,           xmm7
+        paddd       xmm0,           xmm1
+
+        mov         rsi,            arg(5) ;[Sum]
+        mov         rdi,            arg(6) ;[SSE]
+
+        movd        [rsi],       xmm0
+        movd        [rdi],       xmm6
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+;    short xmm_bi_rd[8] = { 64, 64, 64, 64,64, 64, 64, 64};
+align 16
+xmm_bi_rd:
+    times 8 dw 64
+align 16
+vpx_bilinear_filters_sse2:
+    dw 128, 128, 128, 128, 128, 128, 128, 128,  0,  0,  0,  0,  0,  0,  0,  0
+    dw 112, 112, 112, 112, 112, 112, 112, 112, 16, 16, 16, 16, 16, 16, 16, 16
+    dw 96, 96, 96, 96, 96, 96, 96, 96, 32, 32, 32, 32, 32, 32, 32, 32
+    dw 80, 80, 80, 80, 80, 80, 80, 80, 48, 48, 48, 48, 48, 48, 48, 48
+    dw 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64
+    dw 48, 48, 48, 48, 48, 48, 48, 48, 80, 80, 80, 80, 80, 80, 80, 80
+    dw 32, 32, 32, 32, 32, 32, 32, 32, 96, 96, 96, 96, 96, 96, 96, 96
+    dw 16, 16, 16, 16, 16, 16, 16, 16, 112, 112, 112, 112, 112, 112, 112, 112
diff --git a/libvpx/libvpx/vpx_dsp/x86/halfpix_variance_sse2.c b/libvpx/libvpx/vpx_dsp/x86/halfpix_variance_sse2.c
new file mode 100644
index 0000000..4a8fb6d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/halfpix_variance_sse2.c
@@ -0,0 +1,82 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+void vpx_half_horiz_vert_variance16x_h_sse2(const unsigned char *ref,
+                                            int ref_stride,
+                                            const unsigned char *src,
+                                            int src_stride,
+                                            unsigned int height,
+                                            int *sum,
+                                            unsigned int *sumsquared);
+void vpx_half_horiz_variance16x_h_sse2(const unsigned char *ref, int ref_stride,
+                                       const unsigned char *src, int src_stride,
+                                       unsigned int height, int *sum,
+                                       unsigned int *sumsquared);
+void vpx_half_vert_variance16x_h_sse2(const unsigned char *ref, int ref_stride,
+                                      const unsigned char *src, int src_stride,
+                                      unsigned int height, int *sum,
+                                      unsigned int *sumsquared);
+
+uint32_t vpx_variance_halfpixvar16x16_h_sse2(const unsigned char *src,
+                                             int src_stride,
+                                             const unsigned char *dst,
+                                             int dst_stride,
+                                             uint32_t *sse) {
+  int xsum0;
+  unsigned int xxsum0;
+
+  vpx_half_horiz_variance16x_h_sse2(src, src_stride, dst, dst_stride, 16,
+                                    &xsum0, &xxsum0);
+
+  *sse = xxsum0;
+  assert(xsum0 <= 255 * 16 * 16);
+  assert(xsum0 >= -255 * 16 * 16);
+  return (xxsum0 - ((uint32_t)((int64_t)xsum0 * xsum0) >> 8));
+}
+
+uint32_t vpx_variance_halfpixvar16x16_v_sse2(const unsigned char *src,
+                                             int src_stride,
+                                             const unsigned char *dst,
+                                             int dst_stride,
+                                             uint32_t *sse) {
+  int xsum0;
+  unsigned int xxsum0;
+  vpx_half_vert_variance16x_h_sse2(src, src_stride, dst, dst_stride, 16,
+                                   &xsum0, &xxsum0);
+
+  *sse = xxsum0;
+  assert(xsum0 <= 255 * 16 * 16);
+  assert(xsum0 >= -255 * 16 * 16);
+  return (xxsum0 - ((uint32_t)((int64_t)xsum0 * xsum0) >> 8));
+}
+
+
+uint32_t vpx_variance_halfpixvar16x16_hv_sse2(const unsigned char *src,
+                                              int src_stride,
+                                              const unsigned char *dst,
+                                              int dst_stride,
+                                              uint32_t *sse) {
+  int xsum0;
+  unsigned int xxsum0;
+
+  vpx_half_horiz_vert_variance16x_h_sse2(src, src_stride, dst, dst_stride, 16,
+                                         &xsum0, &xxsum0);
+
+  *sse = xxsum0;
+  assert(xsum0 <= 255 * 16 * 16);
+  assert(xsum0 >= -255 * 16 * 16);
+  return (xxsum0 - ((uint32_t)((int64_t)xsum0 * xsum0) >> 8));
+}
diff --git a/libvpx/libvpx/vpx_dsp/x86/highbd_intrapred_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/highbd_intrapred_sse2.asm
new file mode 100644
index 0000000..c61b621
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/highbd_intrapred_sse2.asm
@@ -0,0 +1,453 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_4:  times 8 dw 4
+pw_8:  times 8 dw 8
+pw_16: times 4 dd 16
+pw_32: times 4 dd 32
+
+SECTION .text
+INIT_XMM sse2
+cglobal highbd_dc_predictor_4x4, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  movq                  m0, [aboveq]
+  movq                  m2, [leftq]
+  paddw                 m0, m2
+  pshuflw               m1, m0, 0xe
+  paddw                 m0, m1
+  pshuflw               m1, m0, 0x1
+  paddw                 m0, m1
+  paddw                 m0, [GLOBAL(pw_4)]
+  psraw                 m0, 3
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_8x8, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3, one
+  mov                 oned, 0x00010001
+  lea             stride3q, [strideq*3]
+  movd                  m3, oned
+  pshufd                m3, m3, 0x0
+  paddw                 m0, m2
+  pmaddwd               m0, m3
+  packssdw              m0, m1
+  pmaddwd               m0, m3
+  packssdw              m0, m1
+  pmaddwd               m0, m3
+  paddw                 m0, [GLOBAL(pw_8)]
+  psrlw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  lea                 dstq, [dstq+strideq*8]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_16x16, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m3, [aboveq+16]
+  mova                  m2, [leftq]
+  mova                  m4, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  paddw                 m0, m2
+  paddw                 m0, m3
+  paddw                 m0, m4
+  movhlps               m2, m0
+  paddw                 m0, m2
+  punpcklwd             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  punpckldq             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  paddd                 m0, [GLOBAL(pw_16)]
+  psrad                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+.loop:
+  mova   [dstq              ], m0
+  mova   [dstq           +16], m0
+  mova   [dstq+strideq*2    ], m0
+  mova   [dstq+strideq*2 +16], m0
+  mova   [dstq+strideq*4    ], m0
+  mova   [dstq+strideq*4 +16], m0
+  mova   [dstq+stride3q*2   ], m0
+  mova   [dstq+stride3q*2+16], m0
+  lea                 dstq, [dstq+strideq*8]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_32x32, 4, 5, 7, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [aboveq+32]
+  mova                  m4, [aboveq+48]
+  paddw                 m0, m2
+  paddw                 m3, m4
+  mova                  m2, [leftq]
+  mova                  m4, [leftq+16]
+  mova                  m5, [leftq+32]
+  mova                  m6, [leftq+48]
+  paddw                 m2, m4
+  paddw                 m5, m6
+  paddw                 m0, m3
+  paddw                 m2, m5
+  pxor                  m1, m1
+  paddw                 m0, m2
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  movhlps               m2, m0
+  paddw                 m0, m2
+  punpcklwd             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  punpckldq             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  paddd                 m0, [GLOBAL(pw_32)]
+  psrad                 m0, 6
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+.loop:
+  mova [dstq               ], m0
+  mova [dstq          +16  ], m0
+  mova [dstq          +32  ], m0
+  mova [dstq          +48  ], m0
+  mova [dstq+strideq*2     ], m0
+  mova [dstq+strideq*2+16  ], m0
+  mova [dstq+strideq*2+32  ], m0
+  mova [dstq+strideq*2+48  ], m0
+  mova [dstq+strideq*4     ], m0
+  mova [dstq+strideq*4+16  ], m0
+  mova [dstq+strideq*4+32  ], m0
+  mova [dstq+strideq*4+48  ], m0
+  mova [dstq+stride3q*2    ], m0
+  mova [dstq+stride3q*2 +16], m0
+  mova [dstq+stride3q*2 +32], m0
+  mova [dstq+stride3q*2 +48], m0
+  lea                 dstq, [dstq+strideq*8]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_4x4, 3, 3, 1, dst, stride, above
+  movq                  m0, [aboveq]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_8x8, 3, 3, 1, dst, stride, above
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  lea                 dstq, [dstq+strideq*8]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_16x16, 3, 4, 2, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 4
+.loop:
+  mova    [dstq              ], m0
+  mova    [dstq           +16], m1
+  mova    [dstq+strideq*2    ], m0
+  mova    [dstq+strideq*2 +16], m1
+  mova    [dstq+strideq*4    ], m0
+  mova    [dstq+strideq*4 +16], m1
+  mova    [dstq+stride3q*2   ], m0
+  mova    [dstq+stride3q*2+16], m1
+  lea                 dstq, [dstq+strideq*8]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_32x32, 3, 4, 4, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  mova                  m2, [aboveq+32]
+  mova                  m3, [aboveq+48]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 8
+.loop:
+  mova [dstq               ], m0
+  mova [dstq            +16], m1
+  mova [dstq            +32], m2
+  mova [dstq            +48], m3
+  mova [dstq+strideq*2     ], m0
+  mova [dstq+strideq*2  +16], m1
+  mova [dstq+strideq*2  +32], m2
+  mova [dstq+strideq*2  +48], m3
+  mova [dstq+strideq*4     ], m0
+  mova [dstq+strideq*4  +16], m1
+  mova [dstq+strideq*4  +32], m2
+  mova [dstq+strideq*4  +48], m3
+  mova [dstq+stride3q*2    ], m0
+  mova [dstq+stride3q*2 +16], m1
+  mova [dstq+stride3q*2 +32], m2
+  mova [dstq+stride3q*2 +48], m3
+  lea                 dstq, [dstq+strideq*8]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_tm_predictor_4x4, 5, 5, 6, dst, stride, above, left, bps
+  movd                  m1, [aboveq-2]
+  movq                  m0, [aboveq]
+  pshuflw               m1, m1, 0x0
+  movlhps               m0, m0         ; t1 t2 t3 t4 t1 t2 t3 t4
+  movlhps               m1, m1         ; tl tl tl tl tl tl tl tl
+  ; Get the values to compute the maximum value at this bit depth
+  pcmpeqw               m3, m3
+  movd                  m4, bpsd
+  psubw                 m0, m1         ; t1-tl t2-tl t3-tl t4-tl
+  psllw                 m3, m4
+  pcmpeqw               m2, m2
+  pxor                  m4, m4         ; min possible value
+  pxor                  m3, m2         ; max possible value
+  mova                  m1, [leftq]
+  pshuflw               m2, m1, 0x0
+  pshuflw               m5, m1, 0x55
+  movlhps               m2, m5         ; l1 l1 l1 l1 l2 l2 l2 l2
+  paddw                 m2, m0
+  ;Clamp to the bit-depth
+  pminsw                m2, m3
+  pmaxsw                m2, m4
+  ;Store the values
+  movq    [dstq          ], m2
+  movhpd  [dstq+strideq*2], m2
+  lea                 dstq, [dstq+strideq*4]
+  pshuflw               m2, m1, 0xaa
+  pshuflw               m5, m1, 0xff
+  movlhps               m2, m5
+  paddw                 m2, m0
+  ;Clamp to the bit-depth
+  pminsw                m2, m3
+  pmaxsw                m2, m4
+  ;Store the values
+  movq    [dstq          ], m2
+  movhpd  [dstq+strideq*2], m2
+  RET
+
+INIT_XMM sse2
+cglobal highbd_tm_predictor_8x8, 5, 6, 5, dst, stride, above, left, bps, one
+  movd                  m1, [aboveq-2]
+  mova                  m0, [aboveq]
+  pshuflw               m1, m1, 0x0
+  ; Get the values to compute the maximum value at this bit depth
+  mov                 oned, 1
+  pxor                  m3, m3
+  pxor                  m4, m4
+  pinsrw                m3, oned, 0
+  pinsrw                m4, bpsd, 0
+  pshuflw               m3, m3, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  punpcklqdq            m3, m3
+  mov                lineq, -4
+  mova                  m2, m3
+  punpcklqdq            m1, m1
+  psllw                 m3, m4
+  add                leftq, 16
+  psubw                 m3, m2 ; max possible value
+  pxor                  m4, m4 ; min possible value
+  psubw                 m0, m1
+.loop:
+  movd                  m1, [leftq+lineq*4]
+  movd                  m2, [leftq+lineq*4+2]
+  pshuflw               m1, m1, 0x0
+  pshuflw               m2, m2, 0x0
+  punpcklqdq            m1, m1
+  punpcklqdq            m2, m2
+  paddw                 m1, m0
+  paddw                 m2, m0
+  ;Clamp to the bit-depth
+  pminsw                m1, m3
+  pminsw                m2, m3
+  pmaxsw                m1, m4
+  pmaxsw                m2, m4
+  ;Store the values
+  mova      [dstq          ], m1
+  mova      [dstq+strideq*2], m2
+  lea                 dstq, [dstq+strideq*4]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_tm_predictor_16x16, 5, 5, 8, dst, stride, above, left, bps
+  movd                  m2, [aboveq-2]
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  pshuflw               m2, m2, 0x0
+  ; Get the values to compute the maximum value at this bit depth
+  pcmpeqw               m3, m3
+  movd                  m4, bpsd
+  punpcklqdq            m2, m2
+  psllw                 m3, m4
+  pcmpeqw               m5, m5
+  pxor                  m4, m4         ; min possible value
+  pxor                  m3, m5         ; max possible value
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -8
+  psubw                 m0, m2
+  psubw                 m1, m2
+.loop:
+  movd                  m7, [leftq]
+  pshuflw               m5, m7, 0x0
+  pshuflw               m2, m7, 0x55
+  punpcklqdq            m5, m5         ; l1 l1 l1 l1 l1 l1 l1 l1
+  punpcklqdq            m2, m2         ; l2 l2 l2 l2 l2 l2 l2 l2
+  paddw                 m6, m5, m0     ; t1-tl+l1 to t4-tl+l1
+  paddw                 m5, m1         ; t5-tl+l1 to t8-tl+l1
+  pminsw                m6, m3
+  pminsw                m5, m3
+  pmaxsw                m6, m4         ; Clamp to the bit-depth
+  pmaxsw                m5, m4
+  mova   [dstq           ], m6
+  mova   [dstq        +16], m5
+  paddw                 m6, m2, m0
+  paddw                 m2, m1
+  pminsw                m6, m3
+  pminsw                m2, m3
+  pmaxsw                m6, m4
+  pmaxsw                m2, m4
+  mova   [dstq+strideq*2 ], m6
+  mova [dstq+strideq*2+16], m2
+  lea                 dstq, [dstq+strideq*4]
+  inc                lineq
+  lea                leftq, [leftq+4]
+
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_tm_predictor_32x32, 5, 5, 8, dst, stride, above, left, bps
+  movd                  m0, [aboveq-2]
+  mova                  m1, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [aboveq+32]
+  mova                  m4, [aboveq+48]
+  pshuflw               m0, m0, 0x0
+  ; Get the values to compute the maximum value at this bit depth
+  pcmpeqw               m5, m5
+  movd                  m6, bpsd
+  psllw                 m5, m6
+  pcmpeqw               m7, m7
+  pxor                  m6, m6         ; min possible value
+  pxor                  m5, m7         ; max possible value
+  punpcklqdq            m0, m0
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -16
+  psubw                 m1, m0
+  psubw                 m2, m0
+  psubw                 m3, m0
+  psubw                 m4, m0
+.loop:
+  movd                  m7, [leftq]
+  pshuflw               m7, m7, 0x0
+  punpcklqdq            m7, m7         ; l1 l1 l1 l1 l1 l1 l1 l1
+  paddw                 m0, m7, m1
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq           ], m0
+  paddw                 m0, m7, m2
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq        +16], m0
+  paddw                 m0, m7, m3
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq        +32], m0
+  paddw                 m0, m7, m4
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq        +48], m0
+  movd                  m7, [leftq+2]
+  pshuflw               m7, m7, 0x0
+  punpcklqdq            m7, m7         ; l2 l2 l2 l2 l2 l2 l2 l2
+  paddw                 m0, m7, m1
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq+strideq*2 ], m0
+  paddw                 m0, m7, m2
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq+strideq*2+16], m0
+  paddw                 m0, m7, m3
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq+strideq*2+32], m0
+  paddw                 m0, m7, m4
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq+strideq*2+48], m0
+  lea                 dstq, [dstq+strideq*4]
+  lea                leftq, [leftq+4]
+  inc                lineq
+  jnz .loop
+  REP_RET
diff --git a/libvpx/libvpx/vpx_dsp/x86/highbd_loopfilter_sse2.c b/libvpx/libvpx/vpx_dsp/x86/highbd_loopfilter_sse2.c
new file mode 100644
index 0000000..72e42ad
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/highbd_loopfilter_sse2.c
@@ -0,0 +1,1187 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/emmintrin_compat.h"
+
+static INLINE __m128i signed_char_clamp_bd_sse2(__m128i value, int bd) {
+  __m128i ubounded;
+  __m128i lbounded;
+  __m128i retval;
+
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i t80, max, min;
+
+  if (bd == 8) {
+    t80 = _mm_set1_epi16(0x80);
+    max = _mm_subs_epi16(
+              _mm_subs_epi16(_mm_slli_epi16(one, 8), one), t80);
+  } else if (bd == 10) {
+    t80 = _mm_set1_epi16(0x200);
+    max = _mm_subs_epi16(
+              _mm_subs_epi16(_mm_slli_epi16(one, 10), one), t80);
+  } else {  // bd == 12
+    t80 = _mm_set1_epi16(0x800);
+    max = _mm_subs_epi16(
+              _mm_subs_epi16(_mm_slli_epi16(one, 12), one), t80);
+  }
+
+  min = _mm_subs_epi16(zero, t80);
+
+  ubounded = _mm_cmpgt_epi16(value, max);
+  lbounded = _mm_cmplt_epi16(value, min);
+  retval = _mm_andnot_si128(_mm_or_si128(ubounded, lbounded), value);
+  ubounded = _mm_and_si128(ubounded, max);
+  lbounded = _mm_and_si128(lbounded, min);
+  retval = _mm_or_si128(retval, ubounded);
+  retval = _mm_or_si128(retval, lbounded);
+  return retval;
+}
+
+// TODO(debargha, peter): Break up large functions into smaller ones
+// in this file.
+void vpx_highbd_lpf_horizontal_edge_8_sse2(uint16_t *s, int p,
+                                           const uint8_t *_blimit,
+                                           const uint8_t *_limit,
+                                           const uint8_t *_thresh, int bd) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i blimit, limit, thresh;
+  __m128i q7, p7, q6, p6, q5, p5, q4, p4, q3, p3, q2, p2, q1, p1, q0, p0;
+  __m128i mask, hev, flat, flat2, abs_p1p0, abs_q1q0;
+  __m128i ps1, qs1, ps0, qs0;
+  __m128i abs_p0q0, abs_p1q1, ffff, work;
+  __m128i filt, work_a, filter1, filter2;
+  __m128i flat2_q6, flat2_p6, flat2_q5, flat2_p5, flat2_q4, flat2_p4;
+  __m128i flat2_q3, flat2_p3, flat2_q2, flat2_p2, flat2_q1, flat2_p1;
+  __m128i flat2_q0, flat2_p0;
+  __m128i flat_q2, flat_p2, flat_q1, flat_p1, flat_q0, flat_p0;
+  __m128i pixelFilter_p, pixelFilter_q;
+  __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+  __m128i sum_p7, sum_q7, sum_p3, sum_q3;
+  __m128i t4, t3, t80, t1;
+  __m128i eight, four;
+
+  if (bd == 8) {
+    blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+    limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+    thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+  } else if (bd == 10) {
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+    limit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+    thresh = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+  } else {  // bd == 12
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+    limit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+    thresh = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+  }
+
+  q4 = _mm_load_si128((__m128i *)(s + 4 * p));
+  p4 = _mm_load_si128((__m128i *)(s - 5 * p));
+  q3 = _mm_load_si128((__m128i *)(s + 3 * p));
+  p3 = _mm_load_si128((__m128i *)(s - 4 * p));
+  q2 = _mm_load_si128((__m128i *)(s + 2 * p));
+  p2 = _mm_load_si128((__m128i *)(s - 3 * p));
+  q1 = _mm_load_si128((__m128i *)(s + 1 * p));
+  p1 = _mm_load_si128((__m128i *)(s - 2 * p));
+  q0 = _mm_load_si128((__m128i *)(s + 0 * p));
+  p0 = _mm_load_si128((__m128i *)(s - 1 * p));
+
+  //  highbd_filter_mask
+  abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1));
+  abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1));
+
+  ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
+
+  abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0));
+  abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1));
+
+  //  highbd_hev_mask (in C code this is actually called from highbd_filter4)
+  flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  hev = _mm_subs_epu16(flat, thresh);
+  hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+  abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0);  // abs(p0 - q0) * 2
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);  // abs(p1 - q1) / 2
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p1, p0),
+                                    _mm_subs_epu16(p0, p1)),
+                       _mm_or_si128(_mm_subs_epu16(q1, q0),
+                                    _mm_subs_epu16(q0, q1)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
+                                    _mm_subs_epu16(p1, p2)),
+                       _mm_or_si128(_mm_subs_epu16(q2, q1),
+                                    _mm_subs_epu16(q1, q2)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p2),
+                                    _mm_subs_epu16(p2, p3)),
+                       _mm_or_si128(_mm_subs_epu16(q3, q2),
+                                    _mm_subs_epu16(q2, q3)));
+  mask = _mm_max_epi16(work, mask);
+
+  mask = _mm_subs_epu16(mask, limit);
+  mask = _mm_cmpeq_epi16(mask, zero);  // return ~mask
+
+  // lp filter
+  // highbd_filter4
+  t4 = _mm_set1_epi16(4);
+  t3 = _mm_set1_epi16(3);
+  if (bd == 8)
+    t80 = _mm_set1_epi16(0x80);
+  else if (bd == 10)
+    t80 = _mm_set1_epi16(0x200);
+  else  // bd == 12
+    t80 = _mm_set1_epi16(0x800);
+
+  t1 = _mm_set1_epi16(0x1);
+
+  ps1 = _mm_subs_epi16(p1, t80);
+  qs1 = _mm_subs_epi16(q1, t80);
+  ps0 = _mm_subs_epi16(p0, t80);
+  qs0 = _mm_subs_epi16(q0, t80);
+
+  filt = _mm_and_si128(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd), hev);
+  work_a = _mm_subs_epi16(qs0, ps0);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
+  filt = _mm_and_si128(filt, mask);
+  filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
+  filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
+
+  // Filter1 >> 3
+  filter1 = _mm_srai_epi16(filter1, 0x3);
+  filter2 = _mm_srai_epi16(filter2, 0x3);
+
+  qs0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd),
+      t80);
+  ps0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd),
+      t80);
+  filt = _mm_adds_epi16(filter1, t1);
+  filt = _mm_srai_epi16(filt, 1);
+  filt = _mm_andnot_si128(hev, filt);
+  qs1 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd),
+      t80);
+  ps1 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd),
+      t80);
+
+  // end highbd_filter4
+  // loopfilter done
+
+  // highbd_flat_mask4
+  flat = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p0),
+                                    _mm_subs_epu16(p0, p2)),
+                       _mm_or_si128(_mm_subs_epu16(p3, p0),
+                                    _mm_subs_epu16(p0, p3)));
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(q2, q0),
+                                    _mm_subs_epu16(q0, q2)),
+                       _mm_or_si128(_mm_subs_epu16(q3, q0),
+                                    _mm_subs_epu16(q0, q3)));
+  flat = _mm_max_epi16(work, flat);
+  work = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  flat = _mm_max_epi16(work, flat);
+
+  if (bd == 8)
+    flat = _mm_subs_epu16(flat, one);
+  else if (bd == 10)
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2));
+  else  // bd == 12
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4));
+
+  flat = _mm_cmpeq_epi16(flat, zero);
+  // end flat_mask4
+
+  // flat & mask = flat && mask (as used in filter8)
+  // (because, in both vars, each block of 16 either all 1s or all 0s)
+  flat = _mm_and_si128(flat, mask);
+
+  p5 = _mm_load_si128((__m128i *)(s - 6 * p));
+  q5 = _mm_load_si128((__m128i *)(s + 5 * p));
+  p6 = _mm_load_si128((__m128i *)(s - 7 * p));
+  q6 = _mm_load_si128((__m128i *)(s + 6 * p));
+  p7 = _mm_load_si128((__m128i *)(s - 8 * p));
+  q7 = _mm_load_si128((__m128i *)(s + 7 * p));
+
+  // highbd_flat_mask5 (arguments passed in are p0, q0, p4-p7, q4-q7
+  // but referred to as p0-p4 & q0-q4 in fn)
+  flat2 = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p4, p0),
+                                     _mm_subs_epu16(p0, p4)),
+                        _mm_or_si128(_mm_subs_epu16(q4, q0),
+                                     _mm_subs_epu16(q0, q4)));
+
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p5, p0),
+                                    _mm_subs_epu16(p0, p5)),
+                       _mm_or_si128(_mm_subs_epu16(q5, q0),
+                                    _mm_subs_epu16(q0, q5)));
+  flat2 = _mm_max_epi16(work, flat2);
+
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p6, p0),
+                                    _mm_subs_epu16(p0, p6)),
+                       _mm_or_si128(_mm_subs_epu16(q6, q0),
+                                    _mm_subs_epu16(q0, q6)));
+  flat2 = _mm_max_epi16(work, flat2);
+
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p7, p0),
+                                    _mm_subs_epu16(p0, p7)),
+                       _mm_or_si128(_mm_subs_epu16(q7, q0),
+                                    _mm_subs_epu16(q0, q7)));
+  flat2 = _mm_max_epi16(work, flat2);
+
+  if (bd == 8)
+    flat2 = _mm_subs_epu16(flat2, one);
+  else if (bd == 10)
+    flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 2));
+  else  // bd == 12
+    flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 4));
+
+  flat2 = _mm_cmpeq_epi16(flat2, zero);
+  flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+  // end highbd_flat_mask5
+
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // flat and wide flat calculations
+  eight = _mm_set1_epi16(8);
+  four = _mm_set1_epi16(4);
+
+  pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6, p5),
+                                _mm_add_epi16(p4, p3));
+  pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6, q5),
+                                _mm_add_epi16(q4, q3));
+
+  pixetFilter_p2p1p0 = _mm_add_epi16(p0, _mm_add_epi16(p2, p1));
+  pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+  pixetFilter_q2q1q0 = _mm_add_epi16(q0, _mm_add_epi16(q2, q1));
+  pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+  pixelFilter_p = _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p,
+                                                      pixelFilter_q));
+  pixetFilter_p2p1p0 =   _mm_add_epi16(four,
+                                       _mm_add_epi16(pixetFilter_p2p1p0,
+                                                     pixetFilter_q2q1q0));
+  flat2_p0 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(p7, p0)), 4);
+  flat2_q0 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(q7, q0)), 4);
+  flat_p0 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                         _mm_add_epi16(p3, p0)), 3);
+  flat_q0 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                         _mm_add_epi16(q3, q0)), 3);
+
+  sum_p7 = _mm_add_epi16(p7, p7);
+  sum_q7 = _mm_add_epi16(q7, q7);
+  sum_p3 = _mm_add_epi16(p3, p3);
+  sum_q3 = _mm_add_epi16(q3, q3);
+
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6);
+  flat2_p1 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1)), 4);
+  flat2_q1 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1)), 4);
+
+  pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2);
+  pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2);
+  flat_p1 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                         _mm_add_epi16(sum_p3, p1)), 3);
+  flat_q1 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+                                         _mm_add_epi16(sum_q3, q1)), 3);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  sum_p3 = _mm_add_epi16(sum_p3, p3);
+  sum_q3 = _mm_add_epi16(sum_q3, q3);
+
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5);
+  flat2_p2 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(sum_p7, p2)), 4);
+  flat2_q2 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                                          _mm_add_epi16(sum_q7, q2)), 4);
+
+  pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1);
+  pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1);
+  flat_p2 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                         _mm_add_epi16(sum_p3, p2)), 3);
+  flat_q2 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+                                         _mm_add_epi16(sum_q3, q2)), 3);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4);
+  flat2_p3 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(sum_p7, p3)), 4);
+  flat2_q3 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                                          _mm_add_epi16(sum_q7, q3)), 4);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3);
+  flat2_p4 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(sum_p7, p4)), 4);
+  flat2_q4 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                                          _mm_add_epi16(sum_q7, q4)), 4);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2);
+  flat2_p5 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(sum_p7, p5)), 4);
+  flat2_q5 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                                          _mm_add_epi16(sum_q7, q5)), 4);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1);
+  flat2_p6 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(sum_p7, p6)), 4);
+  flat2_q6 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                                          _mm_add_epi16(sum_q7, q6)), 4);
+
+  //  wide flat
+  //  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+  //  highbd_filter8
+  p2 = _mm_andnot_si128(flat, p2);
+  //  p2 remains unchanged if !(flat && mask)
+  flat_p2 = _mm_and_si128(flat, flat_p2);
+  //  when (flat && mask)
+  p2 = _mm_or_si128(p2, flat_p2);  // full list of p2 values
+  q2 = _mm_andnot_si128(flat, q2);
+  flat_q2 = _mm_and_si128(flat, flat_q2);
+  q2 = _mm_or_si128(q2, flat_q2);  // full list of q2 values
+
+  ps1 = _mm_andnot_si128(flat, ps1);
+  //  p1 takes the value assigned to in in filter4 if !(flat && mask)
+  flat_p1 = _mm_and_si128(flat, flat_p1);
+  //  when (flat && mask)
+  p1 = _mm_or_si128(ps1, flat_p1);  // full list of p1 values
+  qs1 = _mm_andnot_si128(flat, qs1);
+  flat_q1 = _mm_and_si128(flat, flat_q1);
+  q1 = _mm_or_si128(qs1, flat_q1);  // full list of q1 values
+
+  ps0 = _mm_andnot_si128(flat, ps0);
+  //  p0 takes the value assigned to in in filter4 if !(flat && mask)
+  flat_p0 = _mm_and_si128(flat, flat_p0);
+  //  when (flat && mask)
+  p0 = _mm_or_si128(ps0, flat_p0);  // full list of p0 values
+  qs0 = _mm_andnot_si128(flat, qs0);
+  flat_q0 = _mm_and_si128(flat, flat_q0);
+  q0 = _mm_or_si128(qs0, flat_q0);  // full list of q0 values
+  // end highbd_filter8
+
+  // highbd_filter16
+  p6 = _mm_andnot_si128(flat2, p6);
+  //  p6 remains unchanged if !(flat2 && flat && mask)
+  flat2_p6 = _mm_and_si128(flat2, flat2_p6);
+  //  get values for when (flat2 && flat && mask)
+  p6 = _mm_or_si128(p6, flat2_p6);  // full list of p6 values
+  q6 = _mm_andnot_si128(flat2, q6);
+  //  q6 remains unchanged if !(flat2 && flat && mask)
+  flat2_q6 = _mm_and_si128(flat2, flat2_q6);
+  //  get values for when (flat2 && flat && mask)
+  q6 = _mm_or_si128(q6, flat2_q6);  // full list of q6 values
+  _mm_store_si128((__m128i *)(s - 7 * p), p6);
+  _mm_store_si128((__m128i *)(s + 6 * p), q6);
+
+  p5 = _mm_andnot_si128(flat2, p5);
+  //  p5 remains unchanged if !(flat2 && flat && mask)
+  flat2_p5 = _mm_and_si128(flat2, flat2_p5);
+  //  get values for when (flat2 && flat && mask)
+  p5 = _mm_or_si128(p5, flat2_p5);
+  //  full list of p5 values
+  q5 = _mm_andnot_si128(flat2, q5);
+  //  q5 remains unchanged if !(flat2 && flat && mask)
+  flat2_q5 = _mm_and_si128(flat2, flat2_q5);
+  //  get values for when (flat2 && flat && mask)
+  q5 = _mm_or_si128(q5, flat2_q5);
+  //  full list of q5 values
+  _mm_store_si128((__m128i *)(s - 6 * p), p5);
+  _mm_store_si128((__m128i *)(s + 5 * p), q5);
+
+  p4 = _mm_andnot_si128(flat2, p4);
+  //  p4 remains unchanged if !(flat2 && flat && mask)
+  flat2_p4 = _mm_and_si128(flat2, flat2_p4);
+  //  get values for when (flat2 && flat && mask)
+  p4 = _mm_or_si128(p4, flat2_p4);  // full list of p4 values
+  q4 = _mm_andnot_si128(flat2, q4);
+  //  q4 remains unchanged if !(flat2 && flat && mask)
+  flat2_q4 = _mm_and_si128(flat2, flat2_q4);
+  //  get values for when (flat2 && flat && mask)
+  q4 = _mm_or_si128(q4, flat2_q4);  // full list of q4 values
+  _mm_store_si128((__m128i *)(s - 5 * p), p4);
+  _mm_store_si128((__m128i *)(s + 4 * p), q4);
+
+  p3 = _mm_andnot_si128(flat2, p3);
+  //  p3 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p3 = _mm_and_si128(flat2, flat2_p3);
+  //  get values for when (flat2 && flat && mask)
+  p3 = _mm_or_si128(p3, flat2_p3);  // full list of p3 values
+  q3 = _mm_andnot_si128(flat2, q3);
+  //  q3 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q3 = _mm_and_si128(flat2, flat2_q3);
+  //  get values for when (flat2 && flat && mask)
+  q3 = _mm_or_si128(q3, flat2_q3);  // full list of q3 values
+  _mm_store_si128((__m128i *)(s - 4 * p), p3);
+  _mm_store_si128((__m128i *)(s + 3 * p), q3);
+
+  p2 = _mm_andnot_si128(flat2, p2);
+  //  p2 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p2 = _mm_and_si128(flat2, flat2_p2);
+  //  get values for when (flat2 && flat && mask)
+  p2 = _mm_or_si128(p2, flat2_p2);
+  //  full list of p2 values
+  q2 = _mm_andnot_si128(flat2, q2);
+  //  q2 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q2 = _mm_and_si128(flat2, flat2_q2);
+  //  get values for when (flat2 && flat && mask)
+  q2 = _mm_or_si128(q2, flat2_q2);  // full list of q2 values
+  _mm_store_si128((__m128i *)(s - 3 * p), p2);
+  _mm_store_si128((__m128i *)(s + 2 * p), q2);
+
+  p1 = _mm_andnot_si128(flat2, p1);
+  //  p1 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p1 = _mm_and_si128(flat2, flat2_p1);
+  //  get values for when (flat2 && flat && mask)
+  p1 = _mm_or_si128(p1, flat2_p1);  // full list of p1 values
+  q1 = _mm_andnot_si128(flat2, q1);
+  //  q1 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q1 = _mm_and_si128(flat2, flat2_q1);
+  //  get values for when (flat2 && flat && mask)
+  q1 = _mm_or_si128(q1, flat2_q1);  // full list of q1 values
+  _mm_store_si128((__m128i *)(s - 2 * p), p1);
+  _mm_store_si128((__m128i *)(s + 1 * p), q1);
+
+  p0 = _mm_andnot_si128(flat2, p0);
+  //  p0 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p0 = _mm_and_si128(flat2, flat2_p0);
+  //  get values for when (flat2 && flat && mask)
+  p0 = _mm_or_si128(p0, flat2_p0);  // full list of p0 values
+  q0 = _mm_andnot_si128(flat2, q0);
+  //  q0 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q0 = _mm_and_si128(flat2, flat2_q0);
+  //  get values for when (flat2 && flat && mask)
+  q0 = _mm_or_si128(q0, flat2_q0);  // full list of q0 values
+  _mm_store_si128((__m128i *)(s - 1 * p), p0);
+  _mm_store_si128((__m128i *)(s - 0 * p), q0);
+}
+
+void vpx_highbd_lpf_horizontal_edge_16_sse2(uint16_t *s, int p,
+                                            const uint8_t *_blimit,
+                                            const uint8_t *_limit,
+                                            const uint8_t *_thresh, int bd) {
+  vpx_highbd_lpf_horizontal_edge_8_sse2(s, p, _blimit, _limit, _thresh, bd);
+  vpx_highbd_lpf_horizontal_edge_8_sse2(s + 8, p, _blimit, _limit, _thresh, bd);
+}
+
+void vpx_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p,
+                                      const uint8_t *_blimit,
+                                      const uint8_t *_limit,
+                                      const uint8_t *_thresh,
+                                      int bd) {
+  DECLARE_ALIGNED(16, uint16_t, flat_op2[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_op1[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_op0[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_oq2[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_oq1[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_oq0[16]);
+  const __m128i zero = _mm_set1_epi16(0);
+  __m128i blimit, limit, thresh;
+  __m128i mask, hev, flat;
+  __m128i p3 = _mm_load_si128((__m128i *)(s - 4 * p));
+  __m128i q3 = _mm_load_si128((__m128i *)(s + 3 * p));
+  __m128i p2 = _mm_load_si128((__m128i *)(s - 3 * p));
+  __m128i q2 = _mm_load_si128((__m128i *)(s + 2 * p));
+  __m128i p1 = _mm_load_si128((__m128i *)(s - 2 * p));
+  __m128i q1 = _mm_load_si128((__m128i *)(s + 1 * p));
+  __m128i p0 = _mm_load_si128((__m128i *)(s - 1 * p));
+  __m128i q0 = _mm_load_si128((__m128i *)(s + 0 * p));
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i ffff = _mm_cmpeq_epi16(one, one);
+  __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+  const __m128i four = _mm_set1_epi16(4);
+  __m128i workp_a, workp_b, workp_shft;
+
+  const __m128i t4 = _mm_set1_epi16(4);
+  const __m128i t3 = _mm_set1_epi16(3);
+  __m128i t80;
+  const __m128i t1 = _mm_set1_epi16(0x1);
+  __m128i ps1, ps0, qs0, qs1;
+  __m128i filt;
+  __m128i work_a;
+  __m128i filter1, filter2;
+
+  if (bd == 8) {
+    blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+    limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+    thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+    t80 = _mm_set1_epi16(0x80);
+  } else if (bd == 10) {
+    blimit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+    limit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+    thresh = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+    t80 = _mm_set1_epi16(0x200);
+  } else {  // bd == 12
+    blimit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+    limit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+    thresh = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+    t80 = _mm_set1_epi16(0x800);
+  }
+
+  ps1 = _mm_subs_epi16(p1, t80);
+  ps0 = _mm_subs_epi16(p0, t80);
+  qs0 = _mm_subs_epi16(q0, t80);
+  qs1 = _mm_subs_epi16(q1, t80);
+
+  // filter_mask and hev_mask
+  abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0),
+                          _mm_subs_epu16(p0, p1));
+  abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0),
+                          _mm_subs_epu16(q0, q1));
+
+  abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0),
+                          _mm_subs_epu16(q0, p0));
+  abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1),
+                          _mm_subs_epu16(q1, p1));
+  flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  hev = _mm_subs_epu16(flat, thresh);
+  hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+  abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+  // So taking maximums continues to work:
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+  mask = _mm_max_epi16(abs_p1p0, mask);
+  // mask |= (abs(p1 - p0) > limit) * -1;
+  mask = _mm_max_epi16(abs_q1q0, mask);
+  // mask |= (abs(q1 - q0) > limit) * -1;
+
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
+                                    _mm_subs_epu16(p1, p2)),
+                       _mm_or_si128(_mm_subs_epu16(q2, q1),
+                                    _mm_subs_epu16(q1, q2)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p2),
+                                    _mm_subs_epu16(p2, p3)),
+                       _mm_or_si128(_mm_subs_epu16(q3, q2),
+                                    _mm_subs_epu16(q2, q3)));
+  mask = _mm_max_epi16(work, mask);
+  mask = _mm_subs_epu16(mask, limit);
+  mask = _mm_cmpeq_epi16(mask, zero);
+
+  // flat_mask4
+  flat = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p0),
+                                    _mm_subs_epu16(p0, p2)),
+                       _mm_or_si128(_mm_subs_epu16(q2, q0),
+                                    _mm_subs_epu16(q0, q2)));
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p0),
+                                    _mm_subs_epu16(p0, p3)),
+                       _mm_or_si128(_mm_subs_epu16(q3, q0),
+                                    _mm_subs_epu16(q0, q3)));
+  flat = _mm_max_epi16(work, flat);
+  flat = _mm_max_epi16(abs_p1p0, flat);
+  flat = _mm_max_epi16(abs_q1q0, flat);
+
+  if (bd == 8)
+    flat = _mm_subs_epu16(flat, one);
+  else if (bd == 10)
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2));
+  else  // bd == 12
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4));
+
+  flat = _mm_cmpeq_epi16(flat, zero);
+  flat = _mm_and_si128(flat, mask);  // flat & mask
+
+  // Added before shift for rounding part of ROUND_POWER_OF_TWO
+
+  workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+  workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+  workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_op2[0], workp_shft);
+
+  workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_op1[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_op0[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_oq0[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_oq1[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_oq2[0], workp_shft);
+
+  // lp filter
+  filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd);
+  filt = _mm_and_si128(filt, hev);
+  work_a = _mm_subs_epi16(qs0, ps0);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  // (vpx_filter + 3 * (qs0 - ps0)) & mask
+  filt = signed_char_clamp_bd_sse2(filt, bd);
+  filt = _mm_and_si128(filt, mask);
+
+  filter1 = _mm_adds_epi16(filt, t4);
+  filter2 = _mm_adds_epi16(filt, t3);
+
+  // Filter1 >> 3
+  filter1 = signed_char_clamp_bd_sse2(filter1, bd);
+  filter1 = _mm_srai_epi16(filter1, 3);
+
+  // Filter2 >> 3
+  filter2 = signed_char_clamp_bd_sse2(filter2, bd);
+  filter2 = _mm_srai_epi16(filter2, 3);
+
+  // filt >> 1
+  filt = _mm_adds_epi16(filter1, t1);
+  filt = _mm_srai_epi16(filt, 1);
+  // filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+  filt = _mm_andnot_si128(hev, filt);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  q0 = _mm_load_si128((__m128i *)flat_oq0);
+  work_a = _mm_andnot_si128(flat, work_a);
+  q0 = _mm_and_si128(flat, q0);
+  q0 = _mm_or_si128(work_a, q0);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  q1 = _mm_load_si128((__m128i *)flat_oq1);
+  work_a = _mm_andnot_si128(flat, work_a);
+  q1 = _mm_and_si128(flat, q1);
+  q1 = _mm_or_si128(work_a, q1);
+
+  work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q2 = _mm_load_si128((__m128i *)flat_oq2);
+  work_a = _mm_andnot_si128(flat, work_a);
+  q2 = _mm_and_si128(flat, q2);
+  q2 = _mm_or_si128(work_a, q2);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  p0 = _mm_load_si128((__m128i *)flat_op0);
+  work_a = _mm_andnot_si128(flat, work_a);
+  p0 = _mm_and_si128(flat, p0);
+  p0 = _mm_or_si128(work_a, p0);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  p1 = _mm_load_si128((__m128i *)flat_op1);
+  work_a = _mm_andnot_si128(flat, work_a);
+  p1 = _mm_and_si128(flat, p1);
+  p1 = _mm_or_si128(work_a, p1);
+
+  work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p2 = _mm_load_si128((__m128i *)flat_op2);
+  work_a = _mm_andnot_si128(flat, work_a);
+  p2 = _mm_and_si128(flat, p2);
+  p2 = _mm_or_si128(work_a, p2);
+
+  _mm_store_si128((__m128i *)(s - 3 * p), p2);
+  _mm_store_si128((__m128i *)(s - 2 * p), p1);
+  _mm_store_si128((__m128i *)(s - 1 * p), p0);
+  _mm_store_si128((__m128i *)(s + 0 * p), q0);
+  _mm_store_si128((__m128i *)(s + 1 * p), q1);
+  _mm_store_si128((__m128i *)(s + 2 * p), q2);
+}
+
+void vpx_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int p,
+                                           const uint8_t *_blimit0,
+                                           const uint8_t *_limit0,
+                                           const uint8_t *_thresh0,
+                                           const uint8_t *_blimit1,
+                                           const uint8_t *_limit1,
+                                           const uint8_t *_thresh1,
+                                           int bd) {
+  vpx_highbd_lpf_horizontal_8_sse2(s, p, _blimit0, _limit0, _thresh0, bd);
+  vpx_highbd_lpf_horizontal_8_sse2(s + 8, p, _blimit1, _limit1, _thresh1, bd);
+}
+
+void vpx_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p,
+                                      const uint8_t *_blimit,
+                                      const uint8_t *_limit,
+                                      const uint8_t *_thresh,
+                                      int bd) {
+  const __m128i zero = _mm_set1_epi16(0);
+  __m128i blimit, limit, thresh;
+  __m128i mask, hev, flat;
+  __m128i p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  __m128i p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  __m128i p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  __m128i p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  __m128i q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  __m128i q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  __m128i q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  __m128i q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0),
+                                        _mm_subs_epu16(p0, p1));
+  const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0),
+                                        _mm_subs_epu16(q0, q1));
+  const __m128i ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0),
+                                  _mm_subs_epu16(q0, p0));
+  __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1),
+                                  _mm_subs_epu16(q1, p1));
+  __m128i work;
+  const __m128i t4 = _mm_set1_epi16(4);
+  const __m128i t3 = _mm_set1_epi16(3);
+  __m128i t80;
+  __m128i tff80;
+  __m128i tffe0;
+  __m128i t1f;
+  // equivalent to shifting 0x1f left by bitdepth - 8
+  // and setting new bits to 1
+  const __m128i t1 = _mm_set1_epi16(0x1);
+  __m128i t7f;
+  // equivalent to shifting 0x7f left by bitdepth - 8
+  // and setting new bits to 1
+  __m128i ps1, ps0, qs0, qs1;
+  __m128i filt;
+  __m128i work_a;
+  __m128i filter1, filter2;
+
+  if (bd == 8) {
+    blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+    limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+    thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+    t80 = _mm_set1_epi16(0x80);
+    tff80 = _mm_set1_epi16(0xff80);
+    tffe0 = _mm_set1_epi16(0xffe0);
+    t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 8);
+    t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 8);
+  } else if (bd == 10) {
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+    limit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+    thresh = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+    t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 2);
+    tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 2);
+    tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 2);
+    t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 6);
+    t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 6);
+  } else {  // bd == 12
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+    limit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+    thresh = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+    t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 4);
+    tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 4);
+    tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 4);
+    t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 4);
+    t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 4);
+  }
+
+  ps1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80);
+  ps0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80);
+  qs0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80);
+  qs1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80);
+
+  // filter_mask and hev_mask
+  flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  hev = _mm_subs_epu16(flat, thresh);
+  hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+  abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+  // So taking maximums continues to work:
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+  mask = _mm_max_epi16(flat, mask);
+  // mask |= (abs(p1 - p0) > limit) * -1;
+  // mask |= (abs(q1 - q0) > limit) * -1;
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
+                                    _mm_subs_epu16(p1, p2)),
+                       _mm_or_si128(_mm_subs_epu16(p3, p2),
+                                    _mm_subs_epu16(p2, p3)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(q2, q1),
+                                    _mm_subs_epu16(q1, q2)),
+                       _mm_or_si128(_mm_subs_epu16(q3, q2),
+                                    _mm_subs_epu16(q2, q3)));
+  mask = _mm_max_epi16(work, mask);
+  mask = _mm_subs_epu16(mask, limit);
+  mask = _mm_cmpeq_epi16(mask, zero);
+
+  // filter4
+  filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd);
+  filt = _mm_and_si128(filt, hev);
+  work_a = _mm_subs_epi16(qs0, ps0);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
+
+  // (vpx_filter + 3 * (qs0 - ps0)) & mask
+  filt = _mm_and_si128(filt, mask);
+
+  filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
+  filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
+
+  // Filter1 >> 3
+  work_a = _mm_cmpgt_epi16(zero, filter1);  // get the values that are <0
+  filter1 = _mm_srli_epi16(filter1, 3);
+  work_a = _mm_and_si128(work_a, tffe0);  // sign bits for the values < 0
+  filter1 = _mm_and_si128(filter1, t1f);  // clamp the range
+  filter1 = _mm_or_si128(filter1, work_a);  // reinsert the sign bits
+
+  // Filter2 >> 3
+  work_a = _mm_cmpgt_epi16(zero, filter2);
+  filter2 = _mm_srli_epi16(filter2, 3);
+  work_a = _mm_and_si128(work_a, tffe0);
+  filter2 = _mm_and_si128(filter2, t1f);
+  filter2 = _mm_or_si128(filter2, work_a);
+
+  // filt >> 1
+  filt = _mm_adds_epi16(filter1, t1);
+  work_a = _mm_cmpgt_epi16(zero, filt);
+  filt = _mm_srli_epi16(filt, 1);
+  work_a = _mm_and_si128(work_a, tff80);
+  filt = _mm_and_si128(filt, t7f);
+  filt = _mm_or_si128(filt, work_a);
+
+  filt = _mm_andnot_si128(hev, filt);
+
+  q0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd), t80);
+  q1 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd), t80);
+  p0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80);
+  p1 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd), t80);
+
+  _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+  _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+  _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+  _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+}
+
+void vpx_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int p,
+                                           const uint8_t *_blimit0,
+                                           const uint8_t *_limit0,
+                                           const uint8_t *_thresh0,
+                                           const uint8_t *_blimit1,
+                                           const uint8_t *_limit1,
+                                           const uint8_t *_thresh1,
+                                           int bd) {
+  vpx_highbd_lpf_horizontal_4_sse2(s, p, _blimit0, _limit0, _thresh0, bd);
+  vpx_highbd_lpf_horizontal_4_sse2(s + 8, p, _blimit1, _limit1, _thresh1, bd);
+}
+
+static INLINE void highbd_transpose(uint16_t *src[], int in_p,
+                                    uint16_t *dst[], int out_p,
+                                    int num_8x8_to_transpose) {
+  int idx8x8 = 0;
+  __m128i p0, p1, p2, p3, p4, p5, p6, p7, x0, x1, x2, x3, x4, x5, x6, x7;
+  do {
+    uint16_t *in = src[idx8x8];
+    uint16_t *out = dst[idx8x8];
+
+    p0 = _mm_loadu_si128((__m128i *)(in + 0*in_p));  // 00 01 02 03 04 05 06 07
+    p1 = _mm_loadu_si128((__m128i *)(in + 1*in_p));  // 10 11 12 13 14 15 16 17
+    p2 = _mm_loadu_si128((__m128i *)(in + 2*in_p));  // 20 21 22 23 24 25 26 27
+    p3 = _mm_loadu_si128((__m128i *)(in + 3*in_p));  // 30 31 32 33 34 35 36 37
+    p4 = _mm_loadu_si128((__m128i *)(in + 4*in_p));  // 40 41 42 43 44 45 46 47
+    p5 = _mm_loadu_si128((__m128i *)(in + 5*in_p));  // 50 51 52 53 54 55 56 57
+    p6 = _mm_loadu_si128((__m128i *)(in + 6*in_p));  // 60 61 62 63 64 65 66 67
+    p7 = _mm_loadu_si128((__m128i *)(in + 7*in_p));  // 70 71 72 73 74 75 76 77
+    // 00 10 01 11 02 12 03 13
+    x0 = _mm_unpacklo_epi16(p0, p1);
+    // 20 30 21 31 22 32 23 33
+    x1 = _mm_unpacklo_epi16(p2, p3);
+    // 40 50 41 51 42 52 43 53
+    x2 = _mm_unpacklo_epi16(p4, p5);
+    // 60 70 61 71 62 72 63 73
+    x3 = _mm_unpacklo_epi16(p6, p7);
+    // 00 10 20 30 01 11 21 31
+    x4 = _mm_unpacklo_epi32(x0, x1);
+    // 40 50 60 70 41 51 61 71
+    x5 = _mm_unpacklo_epi32(x2, x3);
+    // 00 10 20 30 40 50 60 70
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 01 11 21 31 41 51 61 71
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 0*out_p), x6);
+    // 00 10 20 30 40 50 60 70
+    _mm_storeu_si128((__m128i *)(out + 1*out_p), x7);
+    // 01 11 21 31 41 51 61 71
+
+    // 02 12 22 32 03 13 23 33
+    x4 = _mm_unpackhi_epi32(x0, x1);
+    // 42 52 62 72 43 53 63 73
+    x5 = _mm_unpackhi_epi32(x2, x3);
+    // 02 12 22 32 42 52 62 72
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 03 13 23 33 43 53 63 73
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 2*out_p), x6);
+    // 02 12 22 32 42 52 62 72
+    _mm_storeu_si128((__m128i *)(out + 3*out_p), x7);
+    // 03 13 23 33 43 53 63 73
+
+    // 04 14 05 15 06 16 07 17
+    x0 = _mm_unpackhi_epi16(p0, p1);
+    // 24 34 25 35 26 36 27 37
+    x1 = _mm_unpackhi_epi16(p2, p3);
+    // 44 54 45 55 46 56 47 57
+    x2 = _mm_unpackhi_epi16(p4, p5);
+    // 64 74 65 75 66 76 67 77
+    x3 = _mm_unpackhi_epi16(p6, p7);
+    // 04 14 24 34 05 15 25 35
+    x4 = _mm_unpacklo_epi32(x0, x1);
+    // 44 54 64 74 45 55 65 75
+    x5 = _mm_unpacklo_epi32(x2, x3);
+    // 04 14 24 34 44 54 64 74
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 05 15 25 35 45 55 65 75
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 4*out_p), x6);
+    // 04 14 24 34 44 54 64 74
+    _mm_storeu_si128((__m128i *)(out + 5*out_p), x7);
+    // 05 15 25 35 45 55 65 75
+
+    // 06 16 26 36 07 17 27 37
+    x4 = _mm_unpackhi_epi32(x0, x1);
+    // 46 56 66 76 47 57 67 77
+    x5 = _mm_unpackhi_epi32(x2, x3);
+    // 06 16 26 36 46 56 66 76
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 07 17 27 37 47 57 67 77
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 6*out_p), x6);
+    // 06 16 26 36 46 56 66 76
+    _mm_storeu_si128((__m128i *)(out + 7*out_p), x7);
+    // 07 17 27 37 47 57 67 77
+  } while (++idx8x8 < num_8x8_to_transpose);
+}
+
+static INLINE void highbd_transpose8x16(uint16_t *in0, uint16_t *in1,
+                                        int in_p, uint16_t *out, int out_p) {
+  uint16_t *src0[1];
+  uint16_t *src1[1];
+  uint16_t *dest0[1];
+  uint16_t *dest1[1];
+  src0[0] = in0;
+  src1[0] = in1;
+  dest0[0] = out;
+  dest1[0] = out + 8;
+  highbd_transpose(src0, in_p, dest0, out_p, 1);
+  highbd_transpose(src1, in_p, dest1, out_p, 1);
+}
+
+void vpx_highbd_lpf_vertical_4_sse2(uint16_t *s, int p,
+                                    const uint8_t *blimit,
+                                    const uint8_t *limit,
+                                    const uint8_t *thresh,
+                                    int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]);
+  uint16_t *src[1];
+  uint16_t *dst[1];
+
+  // Transpose 8x8
+  src[0] = s - 4;
+  dst[0] = t_dst;
+
+  highbd_transpose(src, p, dst, 8, 1);
+
+  // Loop filtering
+  vpx_highbd_lpf_horizontal_4_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, bd);
+
+  src[0] = t_dst;
+  dst[0] = s - 4;
+
+  // Transpose back
+  highbd_transpose(src, 8, dst, p, 1);
+}
+
+void vpx_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int p,
+                                         const uint8_t *blimit0,
+                                         const uint8_t *limit0,
+                                         const uint8_t *thresh0,
+                                         const uint8_t *blimit1,
+                                         const uint8_t *limit1,
+                                         const uint8_t *thresh1,
+                                         int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]);
+  uint16_t *src[2];
+  uint16_t *dst[2];
+
+  // Transpose 8x16
+  highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  vpx_highbd_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0,
+                                        thresh0, blimit1, limit1, thresh1, bd);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  highbd_transpose(src, 16, dst, p, 2);
+}
+
+void vpx_highbd_lpf_vertical_8_sse2(uint16_t *s, int p,
+                                    const uint8_t *blimit,
+                                    const uint8_t *limit,
+                                    const uint8_t *thresh,
+                                    int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]);
+  uint16_t *src[1];
+  uint16_t *dst[1];
+
+  // Transpose 8x8
+  src[0] = s - 4;
+  dst[0] = t_dst;
+
+  highbd_transpose(src, p, dst, 8, 1);
+
+  // Loop filtering
+  vpx_highbd_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, bd);
+
+  src[0] = t_dst;
+  dst[0] = s - 4;
+
+  // Transpose back
+  highbd_transpose(src, 8, dst, p, 1);
+}
+
+void vpx_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int p,
+                                         const uint8_t *blimit0,
+                                         const uint8_t *limit0,
+                                         const uint8_t *thresh0,
+                                         const uint8_t *blimit1,
+                                         const uint8_t *limit1,
+                                         const uint8_t *thresh1,
+                                         int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]);
+  uint16_t *src[2];
+  uint16_t *dst[2];
+
+  // Transpose 8x16
+  highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  vpx_highbd_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0,
+                                        thresh0, blimit1, limit1, thresh1, bd);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  highbd_transpose(src, 16, dst, p, 2);
+}
+
+void vpx_highbd_lpf_vertical_16_sse2(uint16_t *s, int p,
+                                     const uint8_t *blimit,
+                                     const uint8_t *limit,
+                                     const uint8_t *thresh,
+                                     int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 16]);
+  uint16_t *src[2];
+  uint16_t *dst[2];
+
+  src[0] = s - 8;
+  src[1] = s;
+  dst[0] = t_dst;
+  dst[1] = t_dst + 8 * 8;
+
+  // Transpose 16x8
+  highbd_transpose(src, p, dst, 8, 2);
+
+  // Loop filtering
+  vpx_highbd_lpf_horizontal_edge_8_sse2(t_dst + 8 * 8, 8, blimit, limit,
+                                        thresh, bd);
+  src[0] = t_dst;
+  src[1] = t_dst + 8 * 8;
+  dst[0] = s - 8;
+  dst[1] = s;
+
+  // Transpose back
+  highbd_transpose(src, 8, dst, p, 2);
+}
+
+void vpx_highbd_lpf_vertical_16_dual_sse2(uint16_t *s,
+                                          int p,
+                                          const uint8_t *blimit,
+                                          const uint8_t *limit,
+                                          const uint8_t *thresh,
+                                          int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[256]);
+
+  //  Transpose 16x16
+  highbd_transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16);
+  highbd_transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16);
+
+  //  Loop filtering
+  vpx_highbd_lpf_horizontal_edge_16_sse2(t_dst + 8 * 16, 16, blimit, limit,
+                                         thresh, bd);
+
+  //  Transpose back
+  highbd_transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p);
+  highbd_transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p);
+}
diff --git a/libvpx/libvpx/vpx_dsp/x86/highbd_quantize_intrin_sse2.c b/libvpx/libvpx/vpx_dsp/x86/highbd_quantize_intrin_sse2.c
new file mode 100644
index 0000000..fd46bef
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/highbd_quantize_intrin_sse2.c
@@ -0,0 +1,179 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vpx_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr,
+                                intptr_t count,
+                                int skip_block,
+                                const int16_t *zbin_ptr,
+                                const int16_t *round_ptr,
+                                const int16_t *quant_ptr,
+                                const int16_t *quant_shift_ptr,
+                                tran_low_t *qcoeff_ptr,
+                                tran_low_t *dqcoeff_ptr,
+                                const int16_t *dequant_ptr,
+                                uint16_t *eob_ptr,
+                                const int16_t *scan,
+                                const int16_t *iscan) {
+  int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
+  __m128i zbins[2];
+  __m128i nzbins[2];
+
+  zbins[0] = _mm_set_epi32((int)zbin_ptr[1],
+                           (int)zbin_ptr[1],
+                           (int)zbin_ptr[1],
+                           (int)zbin_ptr[0]);
+  zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  (void)scan;
+
+  memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = ((int)count / 4) - 1; i >= 0; i--) {
+      __m128i coeffs, cmp1, cmp2;
+      int test;
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+      cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+      cmp1 = _mm_and_si128(cmp1, cmp2);
+      test = _mm_movemask_epi8(cmp1);
+      if (test == 0xffff)
+        non_zero_regs--;
+      else
+        break;
+    }
+
+    // Quantization pass:
+    for (i = 0; i < non_zero_regs; i++) {
+      __m128i coeffs, coeffs_sign, tmp1, tmp2;
+      int test;
+      int abs_coeff[4];
+      int coeff_sign[4];
+
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      coeffs_sign = _mm_srai_epi32(coeffs, 31);
+      coeffs = _mm_sub_epi32(
+            _mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
+      tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
+      tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
+      tmp1 = _mm_or_si128(tmp1, tmp2);
+      test = _mm_movemask_epi8(tmp1);
+      _mm_storeu_si128((__m128i*)abs_coeff, coeffs);
+      _mm_storeu_si128((__m128i*)coeff_sign, coeffs_sign);
+
+      for (j = 0; j < 4; j++) {
+        if (test & (1 << (4 * j))) {
+          int k = 4 * i + j;
+          const int64_t tmp1 = abs_coeff[j] + round_ptr[k != 0];
+          const int64_t tmp2 = ((tmp1 * quant_ptr[k != 0]) >> 16) + tmp1;
+          const uint32_t abs_qcoeff =
+              (uint32_t)((tmp2 * quant_shift_ptr[k != 0]) >> 16);
+          qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j];
+          dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
+          if (abs_qcoeff)
+            eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
+        }
+      }
+    }
+  }
+  *eob_ptr = eob_i + 1;
+}
+
+
+void vpx_highbd_quantize_b_32x32_sse2(const tran_low_t *coeff_ptr,
+                                      intptr_t n_coeffs,
+                                      int skip_block,
+                                      const int16_t *zbin_ptr,
+                                      const int16_t *round_ptr,
+                                      const int16_t *quant_ptr,
+                                      const int16_t *quant_shift_ptr,
+                                      tran_low_t *qcoeff_ptr,
+                                      tran_low_t *dqcoeff_ptr,
+                                      const int16_t *dequant_ptr,
+                                      uint16_t *eob_ptr,
+                                      const int16_t *scan,
+                                      const int16_t *iscan) {
+  __m128i zbins[2];
+  __m128i nzbins[2];
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
+  const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
+  (void)scan;
+  zbins[0] = _mm_set_epi32(zbin1_tmp,
+                           zbin1_tmp,
+                           zbin1_tmp,
+                           zbin0_tmp);
+  zbins[1] = _mm_set1_epi32(zbin1_tmp);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs / 4; i++) {
+      __m128i coeffs, cmp1, cmp2;
+      int test;
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+      cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+      cmp1 = _mm_and_si128(cmp1, cmp2);
+      test = _mm_movemask_epi8(cmp1);
+      if (!(test & 0xf))
+        idx_arr[idx++] = i * 4;
+      if (!(test & 0xf0))
+        idx_arr[idx++] = i * 4 + 1;
+      if (!(test & 0xf00))
+        idx_arr[idx++] = i * 4 + 2;
+      if (!(test & 0xf000))
+        idx_arr[idx++] = i * 4 + 3;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = idx_arr[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int64_t tmp1 = abs_coeff
+                         + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+      const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+      const uint32_t abs_qcoeff =
+          (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
+      qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+      if (abs_qcoeff)
+        eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
diff --git a/libvpx/libvpx/vpx_dsp/x86/highbd_sad4d_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/highbd_sad4d_sse2.asm
new file mode 100644
index 0000000..6c2a61e
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/highbd_sad4d_sse2.asm
@@ -0,0 +1,287 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; HIGH_PROCESS_4x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_4x2x4 5-6 0
+  movh                  m0, [srcq +%2*2]
+%if %1 == 1
+  movu                  m4, [ref1q+%3*2]
+  movu                  m5, [ref2q+%3*2]
+  movu                  m6, [ref3q+%3*2]
+  movu                  m7, [ref4q+%3*2]
+  movhps                m0, [srcq +%4*2]
+  movhps                m4, [ref1q+%5*2]
+  movhps                m5, [ref2q+%5*2]
+  movhps                m6, [ref3q+%5*2]
+  movhps                m7, [ref4q+%5*2]
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m4
+  psubusw               m2, m5
+  psubusw               m4, m0
+  psubusw               m5, m0
+  por                   m4, m3
+  por                   m5, m2
+  pmaddwd               m4, m1
+  pmaddwd               m5, m1
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m6
+  psubusw               m2, m7
+  psubusw               m6, m0
+  psubusw               m7, m0
+  por                   m6, m3
+  por                   m7, m2
+  pmaddwd               m6, m1
+  pmaddwd               m7, m1
+%else
+  movu                  m2, [ref1q+%3*2]
+  movhps                m0, [srcq +%4*2]
+  movhps                m2, [ref1q+%5*2]
+  mova                  m3, m0
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m4, m2
+
+  movu                  m2, [ref2q+%3*2]
+  mova                  m3, m0
+  movhps                m2, [ref2q+%5*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m5, m2
+
+  movu                  m2, [ref3q+%3*2]
+  mova                  m3, m0
+  movhps                m2, [ref3q+%5*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m6, m2
+
+  movu                  m2, [ref4q+%3*2]
+  mova                  m3, m0
+  movhps                m2, [ref4q+%5*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m7, m2
+%endif
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*4]
+  lea                ref1q, [ref1q+ref_strideq*4]
+  lea                ref2q, [ref2q+ref_strideq*4]
+  lea                ref3q, [ref3q+ref_strideq*4]
+  lea                ref4q, [ref4q+ref_strideq*4]
+%endif
+%endmacro
+
+; PROCESS_8x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_8x2x4 5-6 0
+  ; 1st 8 px
+  mova                  m0, [srcq +%2*2]
+%if %1 == 1
+  movu                  m4, [ref1q+%3*2]
+  movu                  m5, [ref2q+%3*2]
+  movu                  m6, [ref3q+%3*2]
+  movu                  m7, [ref4q+%3*2]
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m4
+  psubusw               m2, m5
+  psubusw               m4, m0
+  psubusw               m5, m0
+  por                   m4, m3
+  por                   m5, m2
+  pmaddwd               m4, m1
+  pmaddwd               m5, m1
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m6
+  psubusw               m2, m7
+  psubusw               m6, m0
+  psubusw               m7, m0
+  por                   m6, m3
+  por                   m7, m2
+  pmaddwd               m6, m1
+  pmaddwd               m7, m1
+%else
+  mova                  m3, m0
+  movu                  m2, [ref1q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m4, m2
+  movu                  m2, [ref2q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m5, m2
+  movu                  m2, [ref3q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m6, m2
+  movu                  m2, [ref4q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m7, m2
+%endif
+
+  ; 2nd 8 px
+  mova                  m0, [srcq +(%4)*2]
+  mova                  m3, m0
+  movu                  m2, [ref1q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m4, m2
+  movu                  m2, [ref2q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m5, m2
+  movu                  m2, [ref3q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m6, m2
+  movu                  m2, [ref4q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*4]
+  lea                ref1q, [ref1q+ref_strideq*4]
+  lea                ref2q, [ref2q+ref_strideq*4]
+  lea                ref3q, [ref3q+ref_strideq*4]
+  lea                ref4q, [ref4q+ref_strideq*4]
+%endif
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m7, m2
+%endmacro
+
+; HIGH_PROCESS_16x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_16x2x4 5-6 0
+  HIGH_PROCESS_8x2x4 %1, %2, %3, (%2 + 8), (%3 + 8)
+  HIGH_PROCESS_8x2x4  0, %4, %5, (%4 + 8), (%5 + 8), %6
+%endmacro
+
+; HIGH_PROCESS_32x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_32x2x4 5-6 0
+  HIGH_PROCESS_16x2x4 %1, %2, %3, (%2 + 16), (%3 + 16)
+  HIGH_PROCESS_16x2x4  0, %4, %5, (%4 + 16), (%5 + 16), %6
+%endmacro
+
+; HIGH_PROCESS_64x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_64x2x4 5-6 0
+  HIGH_PROCESS_32x2x4 %1, %2, %3, (%2 + 32), (%3 + 32)
+  HIGH_PROCESS_32x2x4  0, %4, %5, (%4 + 32), (%5 + 32), %6
+%endmacro
+
+; void vpx_highbd_sadNxNx4d_sse2(uint8_t *src,    int src_stride,
+;                         uint8_t *ref[4], int ref_stride,
+;                         uint32_t res[4]);
+; where NxN = 64x64, 32x32, 16x16, 16x8, 8x16 or 8x8
+%macro HIGH_SADNXN4D 2
+%if UNIX64
+cglobal highbd_sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \
+                              res, ref2, ref3, ref4
+%else
+cglobal highbd_sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \
+                              ref2, ref3, ref4
+%endif
+
+; set m1
+  push                srcq
+  mov                 srcd, 0x00010001
+  movd                  m1, srcd
+  pshufd                m1, m1, 0x0
+  pop                 srcq
+
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+  mov                ref2q, [ref1q+gprsize*1]
+  mov                ref3q, [ref1q+gprsize*2]
+  mov                ref4q, [ref1q+gprsize*3]
+  mov                ref1q, [ref1q+gprsize*0]
+
+; convert byte pointers to short pointers
+  shl                 srcq, 1
+  shl                ref2q, 1
+  shl                ref3q, 1
+  shl                ref4q, 1
+  shl                ref1q, 1
+
+  HIGH_PROCESS_%1x2x4 1, 0, 0, src_strideq, ref_strideq, 1
+%rep (%2-4)/2
+  HIGH_PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 1
+%endrep
+  HIGH_PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 0
+  ; N.B. HIGH_PROCESS outputs dwords (32 bits)
+  ; so in high bit depth even the smallest width (4) needs 128bits i.e. XMM
+  movhlps               m0, m4
+  movhlps               m1, m5
+  movhlps               m2, m6
+  movhlps               m3, m7
+  paddd                 m4, m0
+  paddd                 m5, m1
+  paddd                 m6, m2
+  paddd                 m7, m3
+  punpckldq             m4, m5
+  punpckldq             m6, m7
+  movhlps               m0, m4
+  movhlps               m1, m6
+  paddd                 m4, m0
+  paddd                 m6, m1
+  punpcklqdq            m4, m6
+  movifnidn             r4, r4mp
+  movu                [r4], m4
+  RET
+%endmacro
+
+
+INIT_XMM sse2
+HIGH_SADNXN4D 64, 64
+HIGH_SADNXN4D 64, 32
+HIGH_SADNXN4D 32, 64
+HIGH_SADNXN4D 32, 32
+HIGH_SADNXN4D 32, 16
+HIGH_SADNXN4D 16, 32
+HIGH_SADNXN4D 16, 16
+HIGH_SADNXN4D 16,  8
+HIGH_SADNXN4D  8, 16
+HIGH_SADNXN4D  8,  8
+HIGH_SADNXN4D  8,  4
+HIGH_SADNXN4D  4,  8
+HIGH_SADNXN4D  4,  4
diff --git a/libvpx/libvpx/vpx_dsp/x86/highbd_sad_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/highbd_sad_sse2.asm
new file mode 100644
index 0000000..bc4b28d
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/highbd_sad_sse2.asm
@@ -0,0 +1,363 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro HIGH_SAD_FN 4
+%if %4 == 0
+%if %3 == 5
+cglobal highbd_sad%1x%2, 4, %3, 7, src, src_stride, ref, ref_stride, n_rows
+%else ; %3 == 7
+cglobal highbd_sad%1x%2, 4, %3, 7, src, src_stride, ref, ref_stride, \
+                            src_stride3, ref_stride3, n_rows
+%endif ; %3 == 5/7
+%else ; avg
+%if %3 == 5
+cglobal highbd_sad%1x%2_avg, 5, 1 + %3, 7, src, src_stride, ref, ref_stride, \
+                                    second_pred, n_rows
+%else ; %3 == 7
+cglobal highbd_sad%1x%2_avg, 5, ARCH_X86_64 + %3, 7, src, src_stride, \
+                                              ref, ref_stride, \
+                                              second_pred, \
+                                              src_stride3, ref_stride3
+%if ARCH_X86_64
+%define n_rowsd r7d
+%else ; x86-32
+%define n_rowsd dword r0m
+%endif ; x86-32/64
+%endif ; %3 == 5/7
+%endif ; avg/sad
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+%if %3 == 7
+  lea         src_stride3q, [src_strideq*3]
+  lea         ref_stride3q, [ref_strideq*3]
+%endif ; %3 == 7
+; convert src, ref & second_pred to short ptrs (from byte ptrs)
+  shl                 srcq, 1
+  shl                 refq, 1
+%if %4 == 1
+  shl         second_predq, 1
+%endif
+%endmacro
+
+; unsigned int vpx_highbd_sad64x{16,32,64}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD64XN 1-2 0
+  HIGH_SAD_FN 64, %1, 5, %2
+  mov              n_rowsd, %1
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  ; first half of each row
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+16]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+16]
+  por                   m2, m5
+  mova                  m5, [srcq+32]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+32]
+  por                   m3, m5
+  mova                  m5, [srcq+48]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+48]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  paddd                 m0, m1
+  paddd                 m0, m3
+  ; second half of each row
+  movu                  m1, [refq+64]
+  movu                  m2, [refq+80]
+  movu                  m3, [refq+96]
+  movu                  m4, [refq+112]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq+64]
+  psubusw               m5, m1
+  psubusw               m1, [srcq+64]
+  por                   m1, m5
+  mova                  m5, [srcq+80]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+80]
+  por                   m2, m5
+  mova                  m5, [srcq+96]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+96]
+  por                   m3, m5
+  mova                  m5, [srcq+112]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+112]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*2]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*2]
+  paddd                 m0, m3
+
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD64XN 64 ; highbd_sad64x64_sse2
+HIGH_SAD64XN 32 ; highbd_sad64x32_sse2
+HIGH_SAD64XN 64, 1 ; highbd_sad64x64_avg_sse2
+HIGH_SAD64XN 32, 1 ; highbd_sad64x32_avg_sse2
+
+
+; unsigned int vpx_highbd_sad32x{16,32,64}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD32XN 1-2 0
+  HIGH_SAD_FN 32, %1, 5, %2
+  mov              n_rowsd, %1
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+16]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+16]
+  por                   m2, m5
+  mova                  m5, [srcq+32]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+32]
+  por                   m3, m5
+  mova                  m5, [srcq+48]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+48]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*2]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*2]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD32XN 64 ; highbd_sad32x64_sse2
+HIGH_SAD32XN 32 ; highbd_sad32x32_sse2
+HIGH_SAD32XN 16 ; highbd_sad32x16_sse2
+HIGH_SAD32XN 64, 1 ; highbd_sad32x64_avg_sse2
+HIGH_SAD32XN 32, 1 ; highbd_sad32x32_avg_sse2
+HIGH_SAD32XN 16, 1 ; highbd_sad32x16_avg_sse2
+
+; unsigned int vpx_highbd_sad16x{8,16,32}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD16XN 1-2 0
+  HIGH_SAD_FN 16, %1, 5, %2
+  mov              n_rowsd, %1/2
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+ref_strideq*2]
+  movu                  m4, [refq+ref_strideq*2+16]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+16]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*2+16]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+16]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+16]
+  por                   m2, m5
+  mova                  m5, [srcq+src_strideq*2]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+src_strideq*2]
+  por                   m3, m5
+  mova                  m5, [srcq+src_strideq*2+16]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+src_strideq*2+16]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD16XN 32 ; highbd_sad16x32_sse2
+HIGH_SAD16XN 16 ; highbd_sad16x16_sse2
+HIGH_SAD16XN  8 ; highbd_sad16x8_sse2
+HIGH_SAD16XN 32, 1 ; highbd_sad16x32_avg_sse2
+HIGH_SAD16XN 16, 1 ; highbd_sad16x16_avg_sse2
+HIGH_SAD16XN  8, 1 ; highbd_sad16x8_avg_sse2
+
+
+; unsigned int vpx_highbd_sad8x{4,8,16}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD8XN 1-2 0
+  HIGH_SAD_FN 8, %1, 7, %2
+  mov              n_rowsd, %1/4
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+ref_strideq*2]
+  movu                  m3, [refq+ref_strideq*4]
+  movu                  m4, [refq+ref_stride3q*2]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+src_strideq*2]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+src_strideq*2]
+  por                   m2, m5
+  mova                  m5, [srcq+src_strideq*4]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+src_strideq*4]
+  por                   m3, m5
+  mova                  m5, [srcq+src_stride3q*2]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+src_stride3q*2]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*8]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*8]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD8XN 16 ; highbd_sad8x16_sse2
+HIGH_SAD8XN  8 ; highbd_sad8x8_sse2
+HIGH_SAD8XN  4 ; highbd_sad8x4_sse2
+HIGH_SAD8XN 16, 1 ; highbd_sad8x16_avg_sse2
+HIGH_SAD8XN  8, 1 ; highbd_sad8x8_avg_sse2
+HIGH_SAD8XN  4, 1 ; highbd_sad8x4_avg_sse2
diff --git a/libvpx/libvpx/vpx_dsp/x86/highbd_subpel_variance_impl_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/highbd_subpel_variance_impl_sse2.asm
new file mode 100644
index 0000000..30ee81b
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/highbd_subpel_variance_impl_sse2.asm
@@ -0,0 +1,1037 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_8: times  8 dw  8
+bilin_filter_m_sse2: times  8 dw 16
+                     times  8 dw  0
+                     times  8 dw 14
+                     times  8 dw  2
+                     times  8 dw 12
+                     times  8 dw  4
+                     times  8 dw 10
+                     times  8 dw  6
+                     times 16 dw  8
+                     times  8 dw  6
+                     times  8 dw 10
+                     times  8 dw  4
+                     times  8 dw 12
+                     times  8 dw  2
+                     times  8 dw 14
+
+SECTION .text
+
+; int vpx_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride,
+;                               int x_offset, int y_offset,
+;                               const uint8_t *dst, ptrdiff_t dst_stride,
+;                               int height, unsigned int *sse);
+;
+; This function returns the SE and stores SSE in the given pointer.
+
+%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse
+  psubw                %3, %4
+  psubw                %1, %2
+  mova                 %4, %3       ; make copies to manipulate to calc sum
+  mova                 %2, %1       ; use originals for calc sse
+  pmaddwd              %3, %3
+  paddw                %4, %2
+  pmaddwd              %1, %1
+  movhlps              %2, %4
+  paddd                %6, %3
+  paddw                %4, %2
+  pxor                 %2, %2
+  pcmpgtw              %2, %4       ; mask for 0 > %4 (sum)
+  punpcklwd            %4, %2       ; sign-extend word to dword
+  paddd                %6, %1
+  paddd                %5, %4
+
+%endmacro
+
+%macro STORE_AND_RET 0
+%if mmsize == 16
+  ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit
+  ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg.
+  ; We have to sign-extend it before adding the words within the register
+  ; and outputing to a dword.
+  movhlps              m3, m7
+  movhlps              m4, m6
+  paddd                m7, m3
+  paddd                m6, m4
+  pshufd               m3, m7, 0x1
+  pshufd               m4, m6, 0x1
+  paddd                m7, m3
+  paddd                m6, m4
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  movd               [r1], m7           ; store sse
+  movd                rax, m6           ; store sum as return value
+%endif
+  RET
+%endmacro
+
+%macro INC_SRC_BY_SRC_STRIDE  0
+%if ARCH_X86=1 && CONFIG_PIC=1
+  add                srcq, src_stridemp
+  add                srcq, src_stridemp
+%else
+  lea                srcq, [srcq + src_strideq*2]
+%endif
+%endmacro
+
+%macro SUBPEL_VARIANCE 1-2 0 ; W
+%define bilin_filter_m bilin_filter_m_sse2
+%define filter_idx_shift 5
+
+
+%ifdef PIC    ; 64bit PIC
+  %if %2 == 1 ; avg
+    cglobal highbd_sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \
+                                      x_offset, y_offset, \
+                                      dst, dst_stride, \
+                                      sec, sec_stride, height, sse
+    %define sec_str sec_strideq
+  %else
+    cglobal highbd_sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, x_offset, \
+                                  y_offset, dst, dst_stride, height, sse
+  %endif
+  %define block_height heightd
+  %define bilin_filter sseq
+%else
+  %if ARCH_X86=1 && CONFIG_PIC=1
+    %if %2 == 1 ; avg
+      cglobal highbd_sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \
+                                  x_offset, y_offset, \
+                                  dst, dst_stride, \
+                                  sec, sec_stride, \
+                                  height, sse, g_bilin_filter, g_pw_8
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+
+      ; Store bilin_filter and pw_8 location in stack
+      %if GET_GOT_DEFINED == 1
+        GET_GOT eax
+        add esp, 4                ; restore esp
+      %endif
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %else
+      cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \
+                                x_offset, y_offset, dst, dst_stride, height, \
+                                sse, g_bilin_filter, g_pw_8
+      %define block_height heightd
+
+      ; Store bilin_filter and pw_8 location in stack
+      %if GET_GOT_DEFINED == 1
+        GET_GOT eax
+        add esp, 4                ; restore esp
+      %endif
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %endif
+  %else
+    %if %2 == 1 ; avg
+      cglobal highbd_sub_pixel_avg_variance%1xh, 7 + 2 * ARCH_X86_64, \
+                        7 + 2 * ARCH_X86_64, 13, src, src_stride, \
+                                             x_offset, y_offset, \
+                                             dst, dst_stride, \
+                                             sec, sec_stride, \
+                                             height, sse
+      %if ARCH_X86_64
+      %define block_height heightd
+      %define sec_str sec_strideq
+      %else
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+      %endif
+    %else
+      cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \
+                              x_offset, y_offset, dst, dst_stride, height, sse
+      %define block_height heightd
+    %endif
+
+    %define bilin_filter bilin_filter_m
+  %endif
+%endif
+
+  ASSERT               %1 <= 16         ; m6 overflows if w > 16
+  pxor                 m6, m6           ; sum
+  pxor                 m7, m7           ; sse
+
+%if %1 < 16
+  sar                   block_height, 1
+%endif
+%if %2 == 1 ; avg
+  shl             sec_str, 1
+%endif
+
+  ; FIXME(rbultje) replace by jumptable?
+  test          x_offsetd, x_offsetd
+  jnz .x_nonzero
+  ; x_offset == 0
+  test          y_offsetd, y_offsetd
+  jnz .x_zero_y_nonzero
+
+  ; x_offset == 0 && y_offset == 0
+.x_zero_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq + 16]
+  mova                 m1, [dstq]
+  mova                 m3, [dstq + 16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m2, [secq+16]
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq + src_strideq*2]
+  mova                 m1, [dstq]
+  mova                 m3, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_zero_y_zero_loop
+  STORE_AND_RET
+
+.x_zero_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_zero_y_nonhalf
+
+  ; x_offset == 0 && y_offset == 0.5
+.x_zero_y_half_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m4, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*2+16]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+16]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*4]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+dst_strideq*2]
+  pavgw                m0, m1
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_zero_y_half_loop
+  STORE_AND_RET
+
+.x_zero_y_nonhalf:
+  ; x_offset == 0 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+y_offsetq]
+  mova                 m9, [bilin_filter+y_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else ; x86-32 or mmx
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0, reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_zero_y_other_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + 16]
+  movu                 m4, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*2+16]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+16]
+  ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can
+  ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of
+  ; instructions is the same (5), but it is 1 mul instead of 2, so might be
+  ; slightly faster because of pmullw latency. It would also cut our rodata
+  ; tables in half for this function, and save 1-2 registers on x86-64.
+  pmullw               m1, filter_y_a
+  pmullw               m5, filter_y_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m1, m5
+  paddw                m0, m4
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*4]
+  mova                 m4, m1
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+dst_strideq*2]
+  pmullw               m1, filter_y_a
+  pmullw               m5, filter_y_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m1, m5
+  paddw                m0, m4
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_zero_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonzero:
+  cmp           x_offsetd, 8
+  jne .x_nonhalf
+  ; x_offset == 0.5
+  test          y_offsetd, y_offsetd
+  jnz .x_half_y_nonzero
+
+  ; x_offset == 0.5 && y_offset == 0
+.x_half_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + 16]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + 18]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq + 16]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + src_strideq*2]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + src_strideq*2 + 2]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq + dst_strideq*2]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_half_y_zero_loop
+  STORE_AND_RET
+
+.x_half_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_half_y_nonhalf
+
+  ; x_offset == 0.5 && y_offset == 0.5
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+  pavgw                m1, m3
+.x_half_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq + 16]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + 18]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  pavgw                m0, m2
+  pavgw                m1, m3
+  mova                 m4, [dstq]
+  mova                 m5, [dstq + 16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+  mova                 m0, m2
+  mova                 m1, m3
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+.x_half_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq + src_strideq*2]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + src_strideq*2 + 2]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  pavgw                m0, m2
+  pavgw                m2, m3
+  mova                 m4, [dstq]
+  mova                 m5, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m4, m2, m5, m6, m7
+  mova                 m0, m3
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_half_y_half_loop
+  STORE_AND_RET
+
+.x_half_y_nonhalf:
+  ; x_offset == 0.5 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+y_offsetq]
+  mova                 m9, [bilin_filter+y_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else  ; x86_32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0.5. We can reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+  pavgw                m1, m3
+.x_half_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+16]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+18]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m1, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m1, filter_rnd
+  paddw                m1, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  psrlw                m1, 4
+  paddw                m0, m2
+  mova                 m2, [dstq]
+  psrlw                m0, 4
+  mova                 m3, [dstq+16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+  mova                 m0, m4
+  mova                 m1, m5
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+.x_half_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+src_strideq*2]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+src_strideq*2+2]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m4, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m4, filter_rnd
+  paddw                m4, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  psrlw                m4, 4
+  paddw                m0, m2
+  mova                 m2, [dstq]
+  psrlw                m0, 4
+  mova                 m3, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m4, [secq]
+%endif
+  SUM_SSE              m0, m2, m4, m3, m6, m7
+  mova                 m0, m5
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_half_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf:
+  test          y_offsetd, y_offsetd
+  jnz .x_nonhalf_y_nonzero
+
+  ; x_offset == bilin interpolation && y_offset == 0
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_other_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+16]
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m1, m3
+  paddw                m0, m2
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+src_strideq*2+2]
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+dst_strideq*2]
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m1, m3
+  paddw                m0, m2
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+
+  lea                srcq, [srcq+src_strideq*4]
+  lea                dstq, [dstq+dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_other_y_zero_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_nonhalf_y_nonhalf
+
+  ; x_offset == bilin interpolation && y_offset == 0.5
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0.5. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m0, m2
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+  lea                srcq, [srcq+src_strideq*2]
+.x_other_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+16]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+18]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+16]
+  psrlw                m2, 4
+  psrlw                m3, 4
+  pavgw                m0, m2
+  pavgw                m1, m3
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+  mova                 m0, m2
+  mova                 m1, m3
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m2
+  psrlw                m0, 4
+  lea                srcq, [srcq+src_strideq*2]
+.x_other_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+src_strideq*2]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+src_strideq*2+2]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+dst_strideq*2]
+  psrlw                m2, 4
+  psrlw                m3, 4
+  pavgw                m0, m2
+  pavgw                m2, m3
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m4, m2, m5, m6, m7
+  mova                 m0, m3
+
+  lea                srcq, [srcq+src_strideq*4]
+  lea                dstq, [dstq+dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_other_y_half_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf_y_nonhalf:
+; loading filter - this is same as in 8-bit depth
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift ; filter_idx_shift = 5
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [bilin_filter+y_offsetq]
+  mova                m11, [bilin_filter+y_offsetq+16]
+  mova                m12, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_y_a m10
+%define filter_y_b m11
+%define filter_rnd m12
+%else   ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; In this case, there is NO unused register. Used src_stride register. Later,
+; src_stride has to be loaded from stack when it is needed.
+%define tempq src_strideq
+  mov tempq, g_bilin_filterm
+  add           x_offsetq, tempq
+  add           y_offsetq, tempq
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+  add           y_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+; end of load filter
+
+  ; x_offset == bilin interpolation && y_offset == bilin interpolation
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  movu                 m1, [srcq+16]
+  movu                 m3, [srcq+18]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m0, m2
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m4, [srcq+2]
+  movu                 m3, [srcq+16]
+  movu                 m5, [srcq+18]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  psrlw                m2, 4
+  psrlw                m3, 4
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, m2
+  paddw                m1, filter_rnd
+  mova                 m2, [dstq]
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+  mova                 m3, [dstq+16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+  mova                 m0, m4
+  mova                 m1, m5
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq + dst_strideq * 2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m2
+  psrlw                m0, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m4, [srcq+2]
+  INC_SRC_BY_SRC_STRIDE
+  movu                 m3, [srcq]
+  movu                 m5, [srcq+2]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  psrlw                m2, 4
+  psrlw                m3, 4
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m4, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, m2
+  paddw                m4, filter_rnd
+  mova                 m2, [dstq]
+  paddw                m4, m3
+  psrlw                m0, 4
+  psrlw                m4, 4
+  mova                 m3, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m4, [secq]
+%endif
+  SUM_SSE              m0, m2, m4, m3, m6, m7
+  mova                 m0, m5
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq + dst_strideq * 4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_other_y_other_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+%endmacro
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
diff --git a/libvpx/libvpx/vpx_dsp/x86/highbd_variance_impl_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/highbd_variance_impl_sse2.asm
new file mode 100644
index 0000000..923418a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/highbd_variance_impl_sse2.asm
@@ -0,0 +1,313 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;unsigned int vpx_highbd_calc16x16var_sse2
+;(
+;    unsigned char   *  src_ptr,
+;    int             source_stride,
+;    unsigned char   *  ref_ptr,
+;    int             recon_stride,
+;    unsigned int    *  SSE,
+;    int             *  Sum
+;)
+global sym(vpx_highbd_calc16x16var_sse2) PRIVATE
+sym(vpx_highbd_calc16x16var_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        mov         rsi,            arg(0) ;[src_ptr]
+        mov         rdi,            arg(2) ;[ref_ptr]
+
+        movsxd      rax,            DWORD PTR arg(1) ;[source_stride]
+        movsxd      rdx,            DWORD PTR arg(3) ;[recon_stride]
+        add         rax,            rax ; source stride in bytes
+        add         rdx,            rdx ; recon stride in bytes
+
+        ; Prefetch data
+        prefetcht0      [rsi]
+        prefetcht0      [rsi+16]
+        prefetcht0      [rsi+rax]
+        prefetcht0      [rsi+rax+16]
+        lea             rbx,    [rsi+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rax]
+        prefetcht0      [rbx+rax+16]
+
+        prefetcht0      [rdi]
+        prefetcht0      [rdi+16]
+        prefetcht0      [rdi+rdx]
+        prefetcht0      [rdi+rdx+16]
+        lea             rbx,    [rdi+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rdx]
+        prefetcht0      [rbx+rdx+16]
+
+        pxor        xmm0,           xmm0     ; clear xmm0 for unpack
+        pxor        xmm7,           xmm7     ; clear xmm7 for accumulating diffs
+
+        pxor        xmm6,           xmm6     ; clear xmm6 for accumulating sse
+        mov         rcx,            16
+
+.var16loop:
+        movdqu      xmm1,           XMMWORD PTR [rsi]
+        movdqu      xmm2,           XMMWORD PTR [rdi]
+
+        lea             rbx,    [rsi+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rax]
+        prefetcht0      [rbx+rax+16]
+        lea             rbx,    [rdi+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rdx]
+        prefetcht0      [rbx+rdx+16]
+
+        pxor        xmm5,           xmm5
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+16]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+16]
+        paddd       xmm6,           xmm1
+
+        psubw       xmm3,           xmm2
+        movdqu      xmm1,           XMMWORD PTR [rsi+rax]
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx]
+        paddd       xmm6,           xmm3
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+rax+16]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx+16]
+        paddd       xmm6,           xmm1
+
+        psubw       xmm3,           xmm2
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        paddd       xmm6,           xmm3
+
+        movdqa      xmm1,           xmm5
+        movdqa      xmm2,           xmm5
+        pcmpgtw     xmm1,           xmm0
+        pcmpeqw     xmm2,           xmm0
+        por         xmm1,           xmm2
+        pcmpeqw     xmm1,           xmm0
+        movdqa      xmm2,           xmm5
+        punpcklwd   xmm5,           xmm1
+        punpckhwd   xmm2,           xmm1
+        paddd       xmm7,           xmm5
+        paddd       xmm7,           xmm2
+
+        lea         rsi,            [rsi + 2*rax]
+        lea         rdi,            [rdi + 2*rdx]
+        sub         rcx,            2
+        jnz         .var16loop
+
+        movdqa      xmm4,           xmm6
+        punpckldq   xmm6,           xmm0
+
+        punpckhdq   xmm4,           xmm0
+        movdqa      xmm5,           xmm7
+
+        paddd       xmm6,           xmm4
+        punpckldq   xmm7,           xmm0
+
+        punpckhdq   xmm5,           xmm0
+        paddd       xmm7,           xmm5
+
+        movdqa      xmm4,           xmm6
+        movdqa      xmm5,           xmm7
+
+        psrldq      xmm4,           8
+        psrldq      xmm5,           8
+
+        paddd       xmm6,           xmm4
+        paddd       xmm7,           xmm5
+
+        mov         rdi,            arg(4)   ; [SSE]
+        mov         rax,            arg(5)   ; [Sum]
+
+        movd DWORD PTR [rdi],       xmm6
+        movd DWORD PTR [rax],       xmm7
+
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    pop rbx
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;unsigned int vpx_highbd_calc8x8var_sse2
+;(
+;    unsigned char   *  src_ptr,
+;    int             source_stride,
+;    unsigned char   *  ref_ptr,
+;    int             recon_stride,
+;    unsigned int    *  SSE,
+;    int             *  Sum
+;)
+global sym(vpx_highbd_calc8x8var_sse2) PRIVATE
+sym(vpx_highbd_calc8x8var_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        mov         rsi,            arg(0) ;[src_ptr]
+        mov         rdi,            arg(2) ;[ref_ptr]
+
+        movsxd      rax,            DWORD PTR arg(1) ;[source_stride]
+        movsxd      rdx,            DWORD PTR arg(3) ;[recon_stride]
+        add         rax,            rax ; source stride in bytes
+        add         rdx,            rdx ; recon stride in bytes
+
+        ; Prefetch data
+        prefetcht0      [rsi]
+        prefetcht0      [rsi+rax]
+        lea             rbx,    [rsi+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rax]
+
+        prefetcht0      [rdi]
+        prefetcht0      [rdi+rdx]
+        lea             rbx,    [rdi+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rdx]
+
+        pxor        xmm0,           xmm0     ; clear xmm0 for unpack
+        pxor        xmm7,           xmm7     ; clear xmm7 for accumulating diffs
+
+        pxor        xmm6,           xmm6     ; clear xmm6 for accumulating sse
+        mov         rcx,            8
+
+.var8loop:
+        movdqu      xmm1,           XMMWORD PTR [rsi]
+        movdqu      xmm2,           XMMWORD PTR [rdi]
+
+        lea             rbx,    [rsi+rax*4]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rax]
+        lea             rbx,    [rbx+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rax]
+        lea             rbx,    [rdi+rdx*4]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rdx]
+        lea             rbx,    [rbx+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rdx]
+
+        pxor        xmm5,           xmm5
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+rax]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx]
+        paddd       xmm6,           xmm1
+
+        lea         rsi,            [rsi + 2*rax]
+        lea         rdi,            [rdi + 2*rdx]
+
+        psubw       xmm3,           xmm2
+        movdqu      xmm1,           XMMWORD PTR [rsi]
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        movdqu      xmm2,           XMMWORD PTR [rdi]
+        paddd       xmm6,           xmm3
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+rax]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx]
+        paddd       xmm6,           xmm1
+
+        psubw       xmm3,           xmm2
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        paddd       xmm6,           xmm3
+
+        movdqa      xmm1,           xmm5
+        movdqa      xmm2,           xmm5
+        pcmpgtw     xmm1,           xmm0
+        pcmpeqw     xmm2,           xmm0
+        por         xmm1,           xmm2
+        pcmpeqw     xmm1,           xmm0
+        movdqa      xmm2,           xmm5
+        punpcklwd   xmm5,           xmm1
+        punpckhwd   xmm2,           xmm1
+        paddd       xmm7,           xmm5
+        paddd       xmm7,           xmm2
+
+        lea         rsi,            [rsi + 2*rax]
+        lea         rdi,            [rdi + 2*rdx]
+        sub         rcx,            4
+        jnz         .var8loop
+
+        movdqa      xmm4,           xmm6
+        punpckldq   xmm6,           xmm0
+
+        punpckhdq   xmm4,           xmm0
+        movdqa      xmm5,           xmm7
+
+        paddd       xmm6,           xmm4
+        punpckldq   xmm7,           xmm0
+
+        punpckhdq   xmm5,           xmm0
+        paddd       xmm7,           xmm5
+
+        movdqa      xmm4,           xmm6
+        movdqa      xmm5,           xmm7
+
+        psrldq      xmm4,           8
+        psrldq      xmm5,           8
+
+        paddd       xmm6,           xmm4
+        paddd       xmm7,           xmm5
+
+        mov         rdi,            arg(4)   ; [SSE]
+        mov         rax,            arg(5)   ; [Sum]
+
+        movd DWORD PTR [rdi],       xmm6
+        movd DWORD PTR [rax],       xmm7
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    pop rbx
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vpx_dsp/x86/highbd_variance_sse2.c b/libvpx/libvpx/vpx_dsp/x86/highbd_variance_sse2.c
new file mode 100644
index 0000000..14d029c
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/highbd_variance_sse2.c
@@ -0,0 +1,592 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_config.h"
+
+#include "vpx_ports/mem.h"
+
+typedef uint32_t (*high_variance_fn_t) (const uint16_t *src, int src_stride,
+                                        const uint16_t *ref, int ref_stride,
+                                        uint32_t *sse, int *sum);
+
+uint32_t vpx_highbd_calc8x8var_sse2(const uint16_t *src, int src_stride,
+                                    const uint16_t *ref, int ref_stride,
+                                    uint32_t *sse, int *sum);
+
+uint32_t vpx_highbd_calc16x16var_sse2(const uint16_t *src, int src_stride,
+                                      const uint16_t *ref, int ref_stride,
+                                      uint32_t *sse, int *sum);
+
+static void highbd_8_variance_sse2(const uint16_t *src, int src_stride,
+                                   const uint16_t *ref, int ref_stride,
+                                   int w, int h, uint32_t *sse, int *sum,
+                                   high_variance_fn_t var_fn, int block_size) {
+  int i, j;
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride,
+             ref + ref_stride * i + j, ref_stride, &sse0, &sum0);
+      *sse += sse0;
+      *sum += sum0;
+    }
+  }
+}
+
+static void highbd_10_variance_sse2(const uint16_t *src, int src_stride,
+                                    const uint16_t *ref, int ref_stride,
+                                    int w, int h, uint32_t *sse, int *sum,
+                                    high_variance_fn_t var_fn, int block_size) {
+  int i, j;
+  uint64_t sse_long = 0;
+  int32_t sum_long = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride,
+             ref + ref_stride * i + j, ref_stride, &sse0, &sum0);
+      sse_long += sse0;
+      sum_long += sum0;
+    }
+  }
+  *sum = ROUND_POWER_OF_TWO(sum_long, 2);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4);
+}
+
+static void highbd_12_variance_sse2(const uint16_t *src, int src_stride,
+                                    const uint16_t *ref, int ref_stride,
+                                    int w, int h, uint32_t *sse, int *sum,
+                                    high_variance_fn_t var_fn, int block_size) {
+  int i, j;
+  uint64_t sse_long = 0;
+  int32_t sum_long = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride,
+             ref + ref_stride * i + j, ref_stride, &sse0, &sum0);
+      sse_long += sse0;
+      sum_long += sum0;
+    }
+  }
+  *sum = ROUND_POWER_OF_TWO(sum_long, 4);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8);
+}
+
+
+#define HIGH_GET_VAR(S) \
+void vpx_highbd_get##S##x##S##var_sse2(const uint8_t *src8, int src_stride, \
+                                       const uint8_t *ref8, int ref_stride, \
+                                       uint32_t *sse, int *sum) { \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
+  vpx_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, \
+                                     sse, sum); \
+} \
+\
+void vpx_highbd_10_get##S##x##S##var_sse2(const uint8_t *src8, int src_stride, \
+                                          const uint8_t *ref8, int ref_stride, \
+                                          uint32_t *sse, int *sum) { \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
+  vpx_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, \
+                                     sse, sum); \
+  *sum = ROUND_POWER_OF_TWO(*sum, 2); \
+  *sse = ROUND_POWER_OF_TWO(*sse, 4); \
+} \
+\
+void vpx_highbd_12_get##S##x##S##var_sse2(const uint8_t *src8, int src_stride, \
+                                          const uint8_t *ref8, int ref_stride, \
+                                          uint32_t *sse, int *sum) { \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
+  vpx_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, \
+                                     sse, sum); \
+  *sum = ROUND_POWER_OF_TWO(*sum, 4); \
+  *sse = ROUND_POWER_OF_TWO(*sse, 8); \
+}
+
+HIGH_GET_VAR(16);
+HIGH_GET_VAR(8);
+
+#undef HIGH_GET_VAR
+
+#define VAR_FN(w, h, block_size, shift) \
+uint32_t vpx_highbd_8_variance##w##x##h##_sse2( \
+    const uint8_t *src8, int src_stride, \
+    const uint8_t *ref8, int ref_stride, uint32_t *sse) { \
+  int sum; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
+  highbd_8_variance_sse2(src, src_stride, ref, ref_stride, w, h, sse, &sum, \
+                         vpx_highbd_calc##block_size##x##block_size##var_sse2, \
+                         block_size); \
+  return *sse - (((int64_t)sum * sum) >> shift); \
+} \
+\
+uint32_t vpx_highbd_10_variance##w##x##h##_sse2( \
+    const uint8_t *src8, int src_stride, \
+    const uint8_t *ref8, int ref_stride, uint32_t *sse) { \
+  int sum; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
+  highbd_10_variance_sse2( \
+      src, src_stride, ref, ref_stride, w, h, sse, &sum, \
+      vpx_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
+  return *sse - (((int64_t)sum * sum) >> shift); \
+} \
+\
+uint32_t vpx_highbd_12_variance##w##x##h##_sse2( \
+    const uint8_t *src8, int src_stride, \
+    const uint8_t *ref8, int ref_stride, uint32_t *sse) { \
+  int sum; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \
+  highbd_12_variance_sse2( \
+      src, src_stride, ref, ref_stride, w, h, sse, &sum, \
+      vpx_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
+  return *sse - (((int64_t)sum * sum) >> shift); \
+}
+
+VAR_FN(64, 64, 16, 12);
+VAR_FN(64, 32, 16, 11);
+VAR_FN(32, 64, 16, 11);
+VAR_FN(32, 32, 16, 10);
+VAR_FN(32, 16, 16, 9);
+VAR_FN(16, 32, 16, 9);
+VAR_FN(16, 16, 16, 8);
+VAR_FN(16, 8, 8, 7);
+VAR_FN(8, 16, 8, 7);
+VAR_FN(8, 8, 8, 6);
+
+#undef VAR_FN
+
+unsigned int vpx_highbd_8_mse16x16_sse2(const uint8_t *src8, int src_stride,
+                                      const uint8_t *ref8, int ref_stride,
+                                      unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 16, 16,
+                         sse, &sum, vpx_highbd_calc16x16var_sse2, 16);
+  return *sse;
+}
+
+unsigned int vpx_highbd_10_mse16x16_sse2(const uint8_t *src8, int src_stride,
+                                         const uint8_t *ref8, int ref_stride,
+                                         unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 16, 16,
+                          sse, &sum, vpx_highbd_calc16x16var_sse2, 16);
+  return *sse;
+}
+
+unsigned int vpx_highbd_12_mse16x16_sse2(const uint8_t *src8, int src_stride,
+                                         const uint8_t *ref8, int ref_stride,
+                                         unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 16, 16,
+                          sse, &sum, vpx_highbd_calc16x16var_sse2, 16);
+  return *sse;
+}
+
+unsigned int vpx_highbd_8_mse8x8_sse2(const uint8_t *src8, int src_stride,
+                                    const uint8_t *ref8, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 8, 8,
+                         sse, &sum, vpx_highbd_calc8x8var_sse2, 8);
+  return *sse;
+}
+
+unsigned int vpx_highbd_10_mse8x8_sse2(const uint8_t *src8, int src_stride,
+                                       const uint8_t *ref8, int ref_stride,
+                                       unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 8, 8,
+                          sse, &sum, vpx_highbd_calc8x8var_sse2, 8);
+  return *sse;
+}
+
+unsigned int vpx_highbd_12_mse8x8_sse2(const uint8_t *src8, int src_stride,
+                                       const uint8_t *ref8, int ref_stride,
+                                       unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 8, 8,
+                          sse, &sum, vpx_highbd_calc8x8var_sse2, 8);
+  return *sse;
+}
+
+#if CONFIG_USE_X86INC
+// The 2 unused parameters are place holders for PIC enabled build.
+// These definitions are for functions defined in
+// highbd_subpel_variance_impl_sse2.asm
+#define DECL(w, opt) \
+  int vpx_highbd_sub_pixel_variance##w##xh_##opt(const uint16_t *src, \
+                                                 ptrdiff_t src_stride, \
+                                                 int x_offset, int y_offset, \
+                                                 const uint16_t *dst, \
+                                                 ptrdiff_t dst_stride, \
+                                                 int height, \
+                                                 unsigned int *sse, \
+                                                 void *unused0, void *unused);
+#define DECLS(opt) \
+  DECL(8, opt); \
+  DECL(16, opt)
+
+DECLS(sse2);
+
+#undef DECLS
+#undef DECL
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
+uint32_t vpx_highbd_8_sub_pixel_variance##w##x##h##_##opt(const uint8_t *src8, \
+                                                          int src_stride, \
+                                                          int x_offset, \
+                                                          int y_offset, \
+                                                          const uint8_t *dst8, \
+                                                          int dst_stride, \
+                                                          uint32_t *sse_ptr) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  int se = vpx_highbd_sub_pixel_variance##wf##xh_##opt(src, src_stride, \
+                                                       x_offset, y_offset, \
+                                                       dst, dst_stride, h, \
+                                                       &sse, NULL, NULL); \
+  if (w > wf) { \
+    unsigned int sse2; \
+    int se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt(src + 16, \
+                                                          src_stride, \
+                                                          x_offset, y_offset, \
+                                                          dst + 16, \
+                                                          dst_stride, \
+                                                          h, &sse2, \
+                                                          NULL, NULL); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt(src + 32, src_stride, \
+                                                        x_offset, y_offset, \
+                                                        dst + 32, dst_stride, \
+                                                        h, &sse2, NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
+          src + 48, src_stride, x_offset, y_offset, \
+          dst + 48, dst_stride, h, &sse2, NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vpx_highbd_10_sub_pixel_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  int se = vpx_highbd_sub_pixel_variance##wf##xh_##opt(src, src_stride, \
+                                                       x_offset, y_offset, \
+                                                       dst, dst_stride, \
+                                                       h, &sse, NULL, NULL); \
+  if (w > wf) { \
+    uint32_t sse2; \
+    int se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt(src + 16, \
+                                                          src_stride, \
+                                                          x_offset, y_offset, \
+                                                          dst + 16, \
+                                                          dst_stride, \
+                                                          h, &sse2, \
+                                                          NULL, NULL); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt(src + 32, src_stride, \
+                                                        x_offset, y_offset, \
+                                                        dst + 32, dst_stride, \
+                                                        h, &sse2, NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt(src + 48, src_stride, \
+                                                        x_offset, y_offset, \
+                                                        dst + 48, dst_stride, \
+                                                        h, &sse2, NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 2); \
+  sse = ROUND_POWER_OF_TWO(sse, 4); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vpx_highbd_12_sub_pixel_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \
+  int start_row; \
+  uint32_t sse; \
+  int se = 0; \
+  uint64_t long_sse = 0; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  for (start_row = 0; start_row < h; start_row +=16) { \
+    uint32_t sse2; \
+    int height = h - start_row < 16 ? h - start_row : 16; \
+    int se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
+        src + (start_row * src_stride), src_stride, \
+        x_offset, y_offset, dst + (start_row * dst_stride), \
+        dst_stride, height, &sse2, NULL, NULL); \
+    se += se2; \
+    long_sse += sse2; \
+    if (w > wf) { \
+      se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
+          src + 16 + (start_row * src_stride), src_stride, \
+          x_offset, y_offset, dst + 16 + (start_row * dst_stride), \
+          dst_stride, height, &sse2, NULL, NULL); \
+      se += se2; \
+      long_sse += sse2; \
+      if (w > wf * 2) { \
+        se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
+            src + 32 + (start_row * src_stride), src_stride, \
+            x_offset, y_offset, dst + 32 + (start_row * dst_stride), \
+            dst_stride, height, &sse2, NULL, NULL); \
+        se += se2; \
+        long_sse += sse2; \
+        se2 = vpx_highbd_sub_pixel_variance##wf##xh_##opt( \
+            src + 48 + (start_row * src_stride), src_stride, \
+            x_offset, y_offset, dst + 48 + (start_row * dst_stride), \
+            dst_stride, height, &sse2, NULL, NULL); \
+        se += se2; \
+        long_sse += sse2; \
+      }\
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 4); \
+  sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+}
+
+#define FNS(opt) \
+FN(64, 64, 16, 6, 6, opt, (int64_t)); \
+FN(64, 32, 16, 6, 5, opt, (int64_t)); \
+FN(32, 64, 16, 5, 6, opt, (int64_t)); \
+FN(32, 32, 16, 5, 5, opt, (int64_t)); \
+FN(32, 16, 16, 5, 4, opt, (int64_t)); \
+FN(16, 32, 16, 4, 5, opt, (int64_t)); \
+FN(16, 16, 16, 4, 4, opt, (int64_t)); \
+FN(16, 8, 16, 4, 3, opt, (int64_t)); \
+FN(8, 16, 8, 3, 4, opt, (int64_t)); \
+FN(8, 8, 8, 3, 3, opt, (int64_t)); \
+FN(8, 4, 8, 3, 2, opt, (int64_t));
+
+
+FNS(sse2);
+
+#undef FNS
+#undef FN
+
+// The 2 unused parameters are place holders for PIC enabled build.
+#define DECL(w, opt) \
+int vpx_highbd_sub_pixel_avg_variance##w##xh_##opt(const uint16_t *src, \
+                                                   ptrdiff_t src_stride, \
+                                                   int x_offset, int y_offset, \
+                                                   const uint16_t *dst, \
+                                                   ptrdiff_t dst_stride, \
+                                                   const uint16_t *sec, \
+                                                   ptrdiff_t sec_stride, \
+                                                   int height, \
+                                                   unsigned int *sse, \
+                                                   void *unused0, void *unused);
+#define DECLS(opt1) \
+DECL(16, opt1) \
+DECL(8, opt1)
+
+DECLS(sse2);
+#undef DECL
+#undef DECLS
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
+uint32_t vpx_highbd_8_sub_pixel_avg_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \
+    const uint8_t *sec8) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
+  int se = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+               src, src_stride, x_offset, \
+               y_offset, dst, dst_stride, sec, w, h, &sse, NULL, NULL); \
+  if (w > wf) { \
+    uint32_t sse2; \
+    int se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                  src + 16, src_stride, x_offset, y_offset, \
+                  dst + 16, dst_stride, sec + 16, w, h, &sse2, NULL, NULL); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 32, src_stride, x_offset, y_offset, \
+                dst + 32, dst_stride, sec + 32, w, h, &sse2, NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 48, src_stride, x_offset, y_offset, \
+                dst + 48, dst_stride, sec + 48, w, h, &sse2, NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vpx_highbd_10_sub_pixel_avg_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \
+    const uint8_t *sec8) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
+  int se = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src, src_stride, x_offset, \
+                                            y_offset, dst, dst_stride, \
+                                            sec, w, h, &sse, NULL, NULL); \
+  if (w > wf) { \
+    uint32_t sse2; \
+    int se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src + 16, src_stride, \
+                                            x_offset, y_offset, \
+                                            dst + 16, dst_stride, \
+                                            sec + 16, w, h, &sse2, \
+                                            NULL, NULL); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src + 32, src_stride, \
+                                            x_offset, y_offset, \
+                                            dst + 32, dst_stride, \
+                                            sec + 32, w, h, &sse2, \
+                                            NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src + 48, src_stride, \
+                                            x_offset, y_offset, \
+                                            dst + 48, dst_stride, \
+                                            sec + 48, w, h, &sse2, \
+                                            NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 2); \
+  sse = ROUND_POWER_OF_TWO(sse, 4); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vpx_highbd_12_sub_pixel_avg_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \
+    const uint8_t *sec8) { \
+  int start_row; \
+  uint32_t sse; \
+  int se = 0; \
+  uint64_t long_sse = 0; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
+  for (start_row = 0; start_row < h; start_row +=16) { \
+    uint32_t sse2; \
+    int height = h - start_row < 16 ? h - start_row : 16; \
+    int se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + (start_row * src_stride), src_stride, x_offset, \
+                y_offset, dst + (start_row * dst_stride), dst_stride, \
+                sec + (start_row * w), w, height, &sse2, NULL, NULL); \
+    se += se2; \
+    long_sse += sse2; \
+    if (w > wf) { \
+      se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 16 + (start_row * src_stride), src_stride, \
+                x_offset, y_offset, \
+                dst + 16 + (start_row * dst_stride), dst_stride, \
+                sec + 16 + (start_row * w), w, height, &sse2, NULL, NULL); \
+      se += se2; \
+      long_sse += sse2; \
+      if (w > wf * 2) { \
+        se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 32 + (start_row * src_stride), src_stride, \
+                x_offset, y_offset, \
+                dst + 32 + (start_row * dst_stride), dst_stride, \
+                sec + 32 + (start_row * w), w, height, &sse2, NULL, NULL); \
+        se += se2; \
+        long_sse += sse2; \
+        se2 = vpx_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 48 + (start_row * src_stride), src_stride, \
+                x_offset, y_offset, \
+                dst + 48 + (start_row * dst_stride), dst_stride, \
+                sec + 48 + (start_row * w), w, height, &sse2, NULL, NULL); \
+        se += se2; \
+        long_sse += sse2; \
+      } \
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 4); \
+  sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+}
+
+
+#define FNS(opt1) \
+FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
+FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
+FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
+FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
+FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
+FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
+FN(16, 16, 16, 4, 4, opt1, (int64_t)); \
+FN(16, 8, 16, 4, 3, opt1, (int64_t)); \
+FN(8, 16, 8, 4, 3, opt1, (int64_t)); \
+FN(8, 8, 8, 3, 3, opt1, (int64_t)); \
+FN(8, 4, 8, 3, 2, opt1, (int64_t));
+
+FNS(sse2);
+
+#undef FNS
+#undef FN
+#endif  // CONFIG_USE_X86INC
diff --git a/libvpx/libvpx/vpx_dsp/x86/intrapred_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/intrapred_sse2.asm
new file mode 100644
index 0000000..cd6a6ae
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/intrapred_sse2.asm
@@ -0,0 +1,860 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pb_1: times 16 db 1
+pw_4:  times 8 dw 4
+pw_8:  times 8 dw 8
+pw_16: times 8 dw 16
+pw_32: times 8 dw 32
+dc_128: times 16 db 128
+pw2_4:  times 8 dw 2
+pw2_8:  times 8 dw 4
+pw2_16:  times 8 dw 8
+pw2_32:  times 8 dw 16
+
+SECTION .text
+
+; ------------------------------------------
+; input: x, y, z, result
+;
+; trick from pascal
+; (x+2y+z+2)>>2 can be calculated as:
+; result = avg(x,z)
+; result -= xor(x,z) & 1
+; result = avg(result,y)
+; ------------------------------------------
+%macro X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 4
+  pavgb               %4, %1, %3
+  pxor                %3, %1
+  pand                %3, [GLOBAL(pb_1)]
+  psubb               %4, %3
+  pavgb               %4, %2
+%endmacro
+
+INIT_XMM sse2
+cglobal d45_predictor_4x4, 3, 4, 4, dst, stride, above, goffset
+  GET_GOT     goffsetq
+
+  movq                 m0, [aboveq]
+  DEFINE_ARGS dst, stride, temp
+  psrldq               m1, m0, 1
+  psrldq               m2, m0, 2
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3
+
+  ; store 4 lines
+  movd   [dstq          ], m3
+  psrlq                m3, 8
+  movd   [dstq+strideq  ], m3
+  lea                dstq, [dstq+strideq*2]
+  psrlq                m3, 8
+  movd   [dstq          ], m3
+  psrlq                m3, 8
+  movd   [dstq+strideq  ], m3
+  psrlq                m0, 56
+  movd              tempq, m0
+  mov    [dstq+strideq+3], tempb
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal d45_predictor_8x8, 3, 4, 4, dst, stride, above, goffset
+  GET_GOT     goffsetq
+
+  movu                m1, [aboveq]
+  pslldq              m0, m1, 1
+  psrldq              m2, m1, 1
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3
+  punpckhbw           m0, m0 ; 7 7
+  punpcklwd           m0, m0 ; 7 7 7 7
+  punpckldq           m0, m0 ; 7 7 7 7 7 7 7 7
+  punpcklqdq          m3, m0 ; -1 0 1 2 3 4 5 6 7 7 7 7 7 7 7 7
+
+ ; store 4 lines
+  psrldq                m3, 1
+  movq    [dstq          ], m3
+  psrldq                m3, 1
+  movq    [dstq+strideq  ], m3
+  psrldq                m3, 1
+  movq    [dstq+strideq*2], m3
+  psrldq                m3, 1
+  movq    [dstq+stride3q ], m3
+  lea                 dstq, [dstq+strideq*4]
+
+  ; store next 4 lines
+  psrldq                m3, 1
+  movq    [dstq          ], m3
+  psrldq                m3, 1
+  movq    [dstq+strideq  ], m3
+  psrldq                m3, 1
+  movq    [dstq+strideq*2], m3
+  psrldq                m3, 1
+  movq    [dstq+stride3q ], m3
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal d207_predictor_4x4, 4, 4, 5, dst, stride, unused, left, goffset
+  GET_GOT     goffsetq
+
+  movd                m0, [leftq]                ; abcd [byte]
+  punpcklbw           m4, m0, m0                 ; aabb ccdd
+  punpcklwd           m4, m4                     ; aaaa bbbb cccc dddd
+  psrldq              m4, 12                     ; dddd
+  punpckldq           m0, m4                     ; abcd dddd
+  psrldq              m1, m0, 1                  ; bcdd
+  psrldq              m2, m0, 2                  ; cddd
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3   ; a2bc b2cd c3d d
+  pavgb               m1, m0                     ; ab, bc, cd, d [byte]
+
+  punpcklbw           m1, m3             ; ab, a2bc, bc, b2cd, cd, c3d, d, d
+  movd    [dstq        ], m1
+  psrlq               m1, 16             ; bc, b2cd, cd, c3d, d, d
+  movd    [dstq+strideq], m1
+
+  lea               dstq, [dstq+strideq*2]
+  psrlq               m1, 16             ; cd, c3d, d, d
+  movd    [dstq        ], m1
+  movd    [dstq+strideq], m4             ; d, d, d, d
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_predictor_4x4, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  movd                  m2, [leftq]
+  movd                  m0, [aboveq]
+  pxor                  m1, m1
+  punpckldq             m0, m2
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw_4)]
+  psraw                 m0, 3
+  pshuflw               m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_4x4, 2, 5, 2, dst, stride, above, left, goffset
+  movifnidn          leftq, leftmp
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movd                  m0, [leftq]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_4)]
+  psraw                 m0, 2
+  pshuflw               m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_top_predictor_4x4, 3, 5, 2, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movd                  m0, [aboveq]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_4)]
+  psraw                 m0, 2
+  pshuflw               m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [aboveq]
+  movq                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_8)]
+  psraw                 m0, 4
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_top_predictor_8x8, 3, 5, 2, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_8)]
+  psraw                 m0, 3
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_8x8, 2, 5, 2, dst, stride, above, left, goffset
+  movifnidn          leftq, leftmp
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [leftq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_8)]
+  psraw                 m0, 3
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_4x4, 2, 5, 1, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movd     m0,        [GLOBAL(dc_128)]
+  movd    [dstq          ], m0
+  movd    [dstq+strideq  ], m0
+  movd    [dstq+strideq*2], m0
+  movd    [dstq+stride3q ], m0
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_8x8, 2, 5, 1, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movq    m0,        [GLOBAL(dc_128)]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_16)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+
+INIT_XMM sse2
+cglobal dc_top_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_16)]
+  psraw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [leftq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_16)]
+  psraw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  mova    m0,        [GLOBAL(dc_128)]
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+  RESTORE_GOT
+  RET
+
+
+INIT_XMM sse2
+cglobal dc_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [leftq]
+  mova                  m4, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  psadbw                m3, m1
+  psadbw                m4, m1
+  paddw                 m0, m2
+  paddw                 m0, m3
+  paddw                 m0, m4
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_32)]
+  psraw                 m0, 6
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_top_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_32)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [leftq]
+  mova                  m2, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_32)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_32x32, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  mova    m0,        [GLOBAL(dc_128)]
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal v_predictor_4x4, 3, 3, 1, dst, stride, above
+  movd                  m0, [aboveq]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  RET
+
+INIT_XMM sse2
+cglobal v_predictor_8x8, 3, 3, 1, dst, stride, above
+  movq                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  RET
+
+INIT_XMM sse2
+cglobal v_predictor_16x16, 3, 4, 1, dst, stride, above
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 4
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal v_predictor_32x32, 3, 4, 2, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 8
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m1
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m1
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m1
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m1
+  lea                 dstq, [dstq+strideq*4]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal h_predictor_4x4, 2, 4, 4, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  movd                  m0, [leftq]
+  punpcklbw             m0, m0
+  punpcklbw             m0, m0
+  pshufd                m1, m0, 0x1
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m1
+  pshufd                m2, m0, 0x2
+  lea                 dstq, [dstq+strideq*2]
+  pshufd                m3, m0, 0x3
+  movd      [dstq        ], m2
+  movd      [dstq+strideq], m3
+  RET
+
+INIT_XMM sse2
+cglobal h_predictor_8x8, 2, 5, 3, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  mov                lineq, -2
+  DEFINE_ARGS  dst, stride, line, left, stride3
+  lea             stride3q, [strideq*3]
+  movq                  m0, [leftq    ]
+  punpcklbw             m0, m0              ; l1 l1 l2 l2 ... l8 l8
+.loop:
+  pshuflw               m1, m0, 0x0         ; l1 l1 l1 l1 l1 l1 l1 l1
+  pshuflw               m2, m0, 0x55        ; l2 l2 l2 l2 l2 l2 l2 l2
+  movq      [dstq        ], m1
+  movq      [dstq+strideq], m2
+  pshuflw               m1, m0, 0xaa
+  pshuflw               m2, m0, 0xff
+  movq    [dstq+strideq*2], m1
+  movq    [dstq+stride3q ], m2
+  pshufd                m0, m0, 0xe         ; [63:0] l5 l5 l6 l6 l7 l7 l8 l8
+  inc                lineq
+  lea                 dstq, [dstq+strideq*4]
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal h_predictor_16x16, 2, 5, 3, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  mov                lineq, -4
+  DEFINE_ARGS dst, stride, line, left, stride3
+  lea             stride3q, [strideq*3]
+.loop:
+  movd                  m0, [leftq]
+  punpcklbw             m0, m0
+  punpcklbw             m0, m0              ; l1 to l4 each repeated 4 times
+  pshufd            m1, m0, 0x0             ; l1 repeated 16 times
+  pshufd            m2, m0, 0x55            ; l2 repeated 16 times
+  mova    [dstq          ], m1
+  mova    [dstq+strideq  ], m2
+  pshufd            m1, m0, 0xaa
+  pshufd            m2, m0, 0xff
+  mova    [dstq+strideq*2], m1
+  mova    [dstq+stride3q ], m2
+  inc                lineq
+  lea                leftq, [leftq+4       ]
+  lea                 dstq, [dstq+strideq*4]
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal h_predictor_32x32, 2, 5, 3, dst, stride, line, left
+  movifnidn              leftq, leftmp
+  mov                    lineq, -8
+  DEFINE_ARGS dst, stride, line, left, stride3
+  lea                 stride3q, [strideq*3]
+.loop:
+  movd                      m0, [leftq]
+  punpcklbw                 m0, m0
+  punpcklbw                 m0, m0              ; l1 to l4 each repeated 4 times
+  pshufd                m1, m0, 0x0             ; l1 repeated 16 times
+  pshufd                m2, m0, 0x55            ; l2 repeated 16 times
+  mova     [dstq             ], m1
+  mova     [dstq+16          ], m1
+  mova     [dstq+strideq     ], m2
+  mova     [dstq+strideq+16  ], m2
+  pshufd                m1, m0, 0xaa
+  pshufd                m2, m0, 0xff
+  mova     [dstq+strideq*2   ], m1
+  mova     [dstq+strideq*2+16], m1
+  mova     [dstq+stride3q    ], m2
+  mova     [dstq+stride3q+16 ], m2
+  inc                    lineq
+  lea                    leftq, [leftq+4       ]
+  lea                     dstq, [dstq+strideq*4]
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal tm_predictor_4x4, 4, 4, 5, dst, stride, above, left
+  pxor                  m1, m1
+  movq                  m0, [aboveq-1]; [63:0] tl t1 t2 t3 t4 x x x
+  punpcklbw             m0, m1
+  pshuflw               m2, m0, 0x0   ; [63:0] tl tl tl tl [word]
+  psrldq                m0, 2
+  psubw                 m0, m2        ; [63:0] t1-tl t2-tl t3-tl t4-tl [word]
+  movd                  m2, [leftq]
+  punpcklbw             m2, m1
+  pshuflw               m4, m2, 0x0   ; [63:0] l1 l1 l1 l1 [word]
+  pshuflw               m3, m2, 0x55  ; [63:0] l2 l2 l2 l2 [word]
+  paddw                 m4, m0
+  paddw                 m3, m0
+  packuswb              m4, m4
+  packuswb              m3, m3
+  movd      [dstq        ], m4
+  movd      [dstq+strideq], m3
+  lea                 dstq, [dstq+strideq*2]
+  pshuflw               m4, m2, 0xaa
+  pshuflw               m3, m2, 0xff
+  paddw                 m4, m0
+  paddw                 m3, m0
+  packuswb              m4, m4
+  packuswb              m3, m3
+  movd      [dstq        ], m4
+  movd      [dstq+strideq], m3
+  RET
+
+INIT_XMM sse2
+cglobal tm_predictor_8x8, 4, 4, 5, dst, stride, above, left
+  pxor                  m1, m1
+  movd                  m2, [aboveq-1]
+  movq                  m0, [aboveq]
+  punpcklbw             m2, m1
+  punpcklbw             m0, m1        ; t1 t2 t3 t4 t5 t6 t7 t8 [word]
+  pshuflw               m2, m2, 0x0   ; [63:0] tl tl tl tl [word]
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -4
+  punpcklqdq            m2, m2        ; tl tl tl tl tl tl tl tl [word]
+  psubw                 m0, m2        ; t1-tl t2-tl ... t8-tl [word]
+  movq                  m2, [leftq]
+  punpcklbw             m2, m1        ; l1 l2 l3 l4 l5 l6 l7 l8 [word]
+.loop
+  pshuflw               m4, m2, 0x0   ; [63:0] l1 l1 l1 l1 [word]
+  pshuflw               m3, m2, 0x55  ; [63:0] l2 l2 l2 l2 [word]
+  punpcklqdq            m4, m4        ; l1 l1 l1 l1 l1 l1 l1 l1 [word]
+  punpcklqdq            m3, m3        ; l2 l2 l2 l2 l2 l2 l2 l2 [word]
+  paddw                 m4, m0
+  paddw                 m3, m0
+  packuswb              m4, m3
+  movq      [dstq        ], m4
+  movhps    [dstq+strideq], m4
+  lea                 dstq, [dstq+strideq*2]
+  psrldq                m2, 4
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal tm_predictor_16x16, 4, 5, 8, dst, stride, above, left
+  pxor                  m1, m1
+  mova                  m2, [aboveq-16];
+  mova                  m0, [aboveq]   ; t1 t2 ... t16 [byte]
+  punpckhbw             m2, m1         ; [127:112] tl [word]
+  punpckhbw             m4, m0, m1
+  punpcklbw             m0, m1         ; m0:m4 t1 t2 ... t16 [word]
+  DEFINE_ARGS dst, stride, line, left, stride8
+  mov                lineq, -8
+  pshufhw               m2, m2, 0xff
+  mova                  m3, [leftq]    ; l1 l2 ... l16 [byte]
+  punpckhqdq            m2, m2         ; tl repeated 8 times [word]
+  psubw                 m0, m2
+  psubw                 m4, m2         ; m0:m4 t1-tl t2-tl ... t16-tl [word]
+  punpckhbw             m5, m3, m1
+  punpcklbw             m3, m1         ; m3:m5 l1 l2 ... l16 [word]
+  lea             stride8q, [strideq*8]
+.loop:
+  pshuflw               m6, m3, 0x0
+  pshuflw               m7, m5, 0x0
+  punpcklqdq            m6, m6         ; l1 repeated 8 times [word]
+  punpcklqdq            m7, m7         ; l8 repeated 8 times [word]
+  paddw                 m1, m6, m0
+  paddw                 m6, m4         ; m1:m6 ti-tl+l1 [i=1,15] [word]
+  psrldq                m5, 2
+  packuswb              m1, m6
+  mova     [dstq         ], m1
+  paddw                 m1, m7, m0
+  paddw                 m7, m4         ; m1:m7 ti-tl+l8 [i=1,15] [word]
+  psrldq                m3, 2
+  packuswb              m1, m7
+  mova     [dstq+stride8q], m1
+  inc                lineq
+  lea                 dstq, [dstq+strideq]
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal tm_predictor_32x32, 4, 4, 8, dst, stride, above, left
+  pxor                  m1, m1
+  movd                  m2, [aboveq-1]
+  mova                  m0, [aboveq]
+  mova                  m4, [aboveq+16]
+  punpcklbw             m2, m1
+  punpckhbw             m3, m0, m1
+  punpckhbw             m5, m4, m1
+  punpcklbw             m0, m1
+  punpcklbw             m4, m1
+  pshuflw               m2, m2, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -16
+  punpcklqdq            m2, m2
+  add                leftq, 32
+  psubw                 m0, m2
+  psubw                 m3, m2
+  psubw                 m4, m2
+  psubw                 m5, m2
+.loop:
+  movd                  m2, [leftq+lineq*2]
+  pxor                  m1, m1
+  punpcklbw             m2, m1
+  pshuflw               m7, m2, 0x55
+  pshuflw               m2, m2, 0x0
+  punpcklqdq            m2, m2
+  punpcklqdq            m7, m7
+  paddw                 m6, m2, m3
+  paddw                 m1, m2, m0
+  packuswb              m1, m6
+  mova   [dstq           ], m1
+  paddw                 m6, m2, m5
+  paddw                 m1, m2, m4
+  packuswb              m1, m6
+  mova   [dstq+16        ], m1
+  paddw                 m6, m7, m3
+  paddw                 m1, m7, m0
+  packuswb              m1, m6
+  mova   [dstq+strideq   ], m1
+  paddw                 m6, m7, m5
+  paddw                 m1, m7, m4
+  packuswb              m1, m6
+  mova   [dstq+strideq+16], m1
+  lea                 dstq, [dstq+strideq*2]
+  inc                lineq
+  jnz .loop
+  REP_RET
diff --git a/libvpx/libvpx/vpx_dsp/x86/intrapred_ssse3.asm b/libvpx/libvpx/vpx_dsp/x86/intrapred_ssse3.asm
new file mode 100644
index 0000000..5e0139f
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/intrapred_ssse3.asm
@@ -0,0 +1,871 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+
+pb_1: times 16 db 1
+sh_b12345677: db 1, 2, 3, 4, 5, 6, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b23456777: db 2, 3, 4, 5, 6, 7, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b0123456777777777: db 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7
+sh_b1234567777777777: db 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7
+sh_b2345677777777777: db 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7
+sh_b123456789abcdeff: db 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 15
+sh_b23456789abcdefff: db 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 15, 15
+sh_b32104567: db 3, 2, 1, 0, 4, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b8091a2b345: db 8, 0, 9, 1, 10, 2, 11, 3, 4, 5, 0, 0, 0, 0, 0, 0
+sh_b76543210: db 7, 6, 5, 4, 3, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b65432108: db 6, 5, 4, 3, 2, 1, 0, 8, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b54321089: db 5, 4, 3, 2, 1, 0, 8, 9, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b89abcdef: db 8, 9, 10, 11, 12, 13, 14, 15, 0, 0, 0, 0, 0, 0, 0, 0
+sh_bfedcba9876543210: db 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0
+
+SECTION .text
+
+INIT_XMM ssse3
+cglobal d45_predictor_16x16, 3, 6, 4, dst, stride, above, dst8, line, goffset
+  GET_GOT     goffsetq
+
+  mova                   m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, dst8, line
+  lea              stride3q, [strideq*3]
+  lea                 dst8q, [dstq+strideq*8]
+  mova                   m1, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb                 m2, m0, [GLOBAL(sh_b23456789abcdefff)]
+  pavgb                  m3, m2, m0
+  pxor                   m2, m0
+  pshufb                 m0, m1
+  pand                   m2, [GLOBAL(pb_1)]
+  psubb                  m3, m2
+  pavgb                  m0, m3
+
+  ; first 4 lines and first half of 3rd 4 lines
+  mov                 lined, 2
+.loop:
+  mova   [dstq            ], m0
+  movhps [dst8q           ], m0
+  pshufb                 m0, m1
+  mova   [dstq +strideq   ], m0
+  movhps [dst8q+strideq   ], m0
+  pshufb                 m0, m1
+  mova   [dstq +strideq*2 ], m0
+  movhps [dst8q+strideq*2 ], m0
+  pshufb                 m0, m1
+  mova   [dstq +stride3q  ], m0
+  movhps [dst8q+stride3q  ], m0
+  pshufb                 m0, m1
+  lea                  dstq, [dstq +strideq*4]
+  lea                 dst8q, [dst8q+strideq*4]
+  dec                 lined
+  jnz .loop
+
+  ; bottom-right 8x8 block
+  movhps [dstq          +8], m0
+  movhps [dstq+strideq  +8], m0
+  movhps [dstq+strideq*2+8], m0
+  movhps [dstq+stride3q +8], m0
+  lea                  dstq, [dstq+strideq*4]
+  movhps [dstq          +8], m0
+  movhps [dstq+strideq  +8], m0
+  movhps [dstq+strideq*2+8], m0
+  movhps [dstq+stride3q +8], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d45_predictor_32x32, 3, 6, 7, dst, stride, above, dst16, line, goffset
+  GET_GOT     goffsetq
+
+  mova                   m0, [aboveq]
+  mova                   m4, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, dst16, line
+  lea              stride3q, [strideq*3]
+  lea                dst16q, [dstq  +strideq*8]
+  lea                dst16q, [dst16q+strideq*8]
+  mova                   m1, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb                 m2, m4, [GLOBAL(sh_b23456789abcdefff)]
+  pavgb                  m3, m2, m4
+  pxor                   m2, m4
+  palignr                m5, m4, m0, 1
+  palignr                m6, m4, m0, 2
+  pshufb                 m4, m1
+  pand                   m2, [GLOBAL(pb_1)]
+  psubb                  m3, m2
+  pavgb                  m4, m3
+  pavgb                  m3, m0, m6
+  pxor                   m0, m6
+  pand                   m0, [GLOBAL(pb_1)]
+  psubb                  m3, m0
+  pavgb                  m5, m3
+
+  ; write 4x4 lines (and the first half of the second 4x4 lines)
+  mov                  lined, 4
+.loop:
+  mova [dstq               ], m5
+  mova [dstq            +16], m4
+  mova [dst16q             ], m4
+  palignr                 m3, m4, m5, 1
+  pshufb                  m4, m1
+  mova [dstq  +strideq     ], m3
+  mova [dstq  +strideq  +16], m4
+  mova [dst16q+strideq     ], m4
+  palignr                 m5, m4, m3, 1
+  pshufb                  m4, m1
+  mova [dstq  +strideq*2   ], m5
+  mova [dstq  +strideq*2+16], m4
+  mova [dst16q+strideq*2   ], m4
+  palignr                 m3, m4, m5, 1
+  pshufb                  m4, m1
+  mova [dstq  +stride3q    ], m3
+  mova [dstq  +stride3q +16], m4
+  mova [dst16q+stride3q    ], m4
+  palignr                 m5, m4, m3, 1
+  pshufb                  m4, m1
+  lea                  dstq, [dstq  +strideq*4]
+  lea                dst16q, [dst16q+strideq*4]
+  dec                 lined
+  jnz .loop
+
+  ; write second half of second 4x4 lines
+  mova [dstq            +16], m4
+  mova [dstq  +strideq  +16], m4
+  mova [dstq  +strideq*2+16], m4
+  mova [dstq  +stride3q +16], m4
+  lea                  dstq, [dstq  +strideq*4]
+  mova [dstq            +16], m4
+  mova [dstq  +strideq  +16], m4
+  mova [dstq  +strideq*2+16], m4
+  mova [dstq  +stride3q +16], m4
+  lea                  dstq, [dstq  +strideq*4]
+  mova [dstq            +16], m4
+  mova [dstq  +strideq  +16], m4
+  mova [dstq  +strideq*2+16], m4
+  mova [dstq  +stride3q +16], m4
+  lea                  dstq, [dstq  +strideq*4]
+  mova [dstq            +16], m4
+  mova [dstq  +strideq  +16], m4
+  mova [dstq  +strideq*2+16], m4
+  mova [dstq  +stride3q +16], m4
+
+  RESTORE_GOT
+  RET
+
+; ------------------------------------------
+; input: x, y, z, result
+;
+; trick from pascal
+; (x+2y+z+2)>>2 can be calculated as:
+; result = avg(x,z)
+; result -= xor(x,z) & 1
+; result = avg(result,y)
+; ------------------------------------------
+%macro X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 4
+  pavgb               %4, %1, %3
+  pxor                %3, %1
+  pand                %3, [GLOBAL(pb_1)]
+  psubb               %4, %3
+  pavgb               %4, %2
+%endmacro
+
+INIT_XMM ssse3
+cglobal d63_predictor_4x4, 3, 4, 5, dst, stride, above, goffset
+  GET_GOT     goffsetq
+
+  movq                m3, [aboveq]
+  pshufb              m1, m3, [GLOBAL(sh_b23456777)]
+  pshufb              m2, m3, [GLOBAL(sh_b12345677)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m3, m2, m1, m4
+  pavgb               m3, m2
+
+  ; store 4 lines
+  movd    [dstq        ], m3
+  movd    [dstq+strideq], m4
+  lea               dstq, [dstq+strideq*2]
+  psrldq              m3, 1
+  psrldq              m4, 1
+  movd    [dstq        ], m3
+  movd    [dstq+strideq], m4
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d63_predictor_8x8, 3, 4, 5, dst, stride, above, goffset
+  GET_GOT     goffsetq
+
+  movq                m3, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+  pshufb              m1, m3, [GLOBAL(sh_b2345677777777777)]
+  pshufb              m0, m3, [GLOBAL(sh_b0123456777777777)]
+  pshufb              m2, m3, [GLOBAL(sh_b1234567777777777)]
+  pshufb              m3, [GLOBAL(sh_b0123456777777777)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m2, m1, m4
+  pavgb               m3, m2
+
+  ; store 4 lines
+  movq    [dstq        ], m3
+  movq    [dstq+strideq], m4
+  psrldq              m3, 1
+  psrldq              m4, 1
+  movq  [dstq+strideq*2], m3
+  movq  [dstq+stride3q ], m4
+  lea               dstq, [dstq+strideq*4]
+  psrldq              m3, 1
+  psrldq              m4, 1
+
+  ; store 4 lines
+  movq    [dstq        ], m3
+  movq    [dstq+strideq], m4
+  psrldq              m3, 1
+  psrldq              m4, 1
+  movq  [dstq+strideq*2], m3
+  movq  [dstq+stride3q ], m4
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d63_predictor_16x16, 3, 5, 5, dst, stride, above, line, goffset
+  GET_GOT     goffsetq
+
+  mova                m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, line
+  lea           stride3q, [strideq*3]
+  mova                m1, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb              m2, m0, [GLOBAL(sh_b23456789abcdefff)]
+  pshufb              m3, m0, m1
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m3, m2, m4
+  pavgb               m0, m3
+
+  mov              lined, 4
+.loop:
+  mova  [dstq          ], m0
+  mova  [dstq+strideq  ], m4
+  pshufb              m0, m1
+  pshufb              m4, m1
+  mova  [dstq+strideq*2], m0
+  mova  [dstq+stride3q ], m4
+  pshufb              m0, m1
+  pshufb              m4, m1
+  lea               dstq, [dstq+strideq*4]
+  dec              lined
+  jnz .loop
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM ssse3
+cglobal d63_predictor_32x32, 3, 5, 8, dst, stride, above, line, goffset
+  GET_GOT     goffsetq
+
+  mova                   m0, [aboveq]
+  mova                   m7, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, line
+  mova                   m1, [GLOBAL(sh_b123456789abcdeff)]
+  lea              stride3q, [strideq*3]
+  pshufb                 m2, m7, [GLOBAL(sh_b23456789abcdefff)]
+  pshufb                 m3, m7, m1
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m2, m4
+  palignr                m6, m7, m0, 1
+  palignr                m5, m7, m0, 2
+  pavgb                  m7, m3
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m6, m5, m2
+  pavgb                  m0, m6
+
+  mov                 lined, 8
+.loop:
+  mova  [dstq             ], m0
+  mova  [dstq          +16], m7
+  mova  [dstq+strideq     ], m2
+  mova  [dstq+strideq  +16], m4
+  palignr                m3, m7, m0, 1
+  palignr                m5, m4, m2, 1
+  pshufb                 m7, m1
+  pshufb                 m4, m1
+
+  mova  [dstq+strideq*2   ], m3
+  mova  [dstq+strideq*2+16], m7
+  mova  [dstq+stride3q    ], m5
+  mova  [dstq+stride3q +16], m4
+  palignr                m0, m7, m3, 1
+  palignr                m2, m4, m5, 1
+  pshufb                 m7, m1
+  pshufb                 m4, m1
+  lea                  dstq, [dstq+strideq*4]
+  dec                 lined
+  jnz .loop
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_4x4, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  movd                m0, [leftq]               ; l1, l2, l3, l4
+  movd                m1, [aboveq-1]            ; tl, t1, t2, t3
+  punpckldq           m0, m1                    ; l1, l2, l3, l4, tl, t1, t2, t3
+  pshufb              m0, [GLOBAL(sh_b32104567)]; l4, l3, l2, l1, tl, t1, t2, t3
+  psrldq              m1, m0, 1                 ; l3, l2, l1, tl, t1, t2, t3
+  psrldq              m2, m0, 2                 ; l2, l1, tl, t1, t2, t3
+  ; comments below are for a predictor like this
+  ; A1 B1 C1 D1
+  ; A2 B2 A1 B1
+  ; A3 B3 A2 B2
+  ; A4 B4 A3 B3
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3  ; 3-tap avg B4 B3 B2 B1 C1 D1
+  pavgb               m1, m0                    ; 2-tap avg A4 A3 A2 A1
+
+  punpcklqdq          m3, m1                    ; B4 B3 B2 B1 C1 D1 x x A4 A3 A2 A1 ..
+
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+  pshufb              m3, [GLOBAL(sh_b8091a2b345)] ; A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 ..
+  movd  [dstq+stride3q ], m3
+  psrldq              m3, 2                     ; A3 B3 A2 B2 A1 B1 C1 D1 ..
+  movd  [dstq+strideq*2], m3
+  psrldq              m3, 2                     ; A2 B2 A1 B1 C1 D1 ..
+  movd  [dstq+strideq  ], m3
+  psrldq              m3, 2                     ; A1 B1 C1 D1 ..
+  movd  [dstq          ], m3
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_8x8, 4, 5, 8, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  movq                m0, [leftq]                     ; [0- 7] l1-8 [byte]
+  movhps              m0, [aboveq-1]                  ; [8-15] tl, t1-7 [byte]
+  pshufb              m1, m0, [GLOBAL(sh_b76543210)]  ; l8-1 [word]
+  pshufb              m2, m0, [GLOBAL(sh_b65432108)]  ; l7-1,tl [word]
+  pshufb              m3, m0, [GLOBAL(sh_b54321089)]  ; l6-1,tl,t1 [word]
+  pshufb              m0, [GLOBAL(sh_b89abcdef)]      ; tl,t1-7 [word]
+  psrldq              m4, m0, 1                       ; t1-7 [word]
+  psrldq              m5, m0, 2                       ; t2-7 [word]
+  ; comments below are for a predictor like this
+  ; A1 B1 C1 D1 E1 F1 G1 H1
+  ; A2 B2 A1 B1 C1 D1 E1 F1
+  ; A3 B3 A2 B2 A1 B1 C1 D1
+  ; A4 B4 A3 B3 A2 B2 A1 B1
+  ; A5 B5 A4 B4 A3 B3 A2 B2
+  ; A6 B6 A5 B5 A4 B4 A3 B3
+  ; A7 B7 A6 B6 A5 B5 A4 B4
+  ; A8 B8 A7 B7 A6 B6 A5 B5
+  pavgb               m6, m1, m2                ; 2-tap avg A8-A1
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m4, m5, m7  ; 3-tap avg C-H1
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m1, m2, m3, m0  ; 3-tap avg B8-1
+
+  punpcklbw           m6, m0                    ; A-B8, A-B7 ... A-B2, A-B1
+
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+
+  movhps [dstq+stride3q], m6                    ; A-B4, A-B3, A-B2, A-B1
+  palignr             m0, m7, m6, 10            ; A-B3, A-B2, A-B1, C-H1
+  movq  [dstq+strideq*2], m0
+  psrldq              m0, 2                     ; A-B2, A-B1, C-H1
+  movq  [dstq+strideq  ], m0
+  psrldq              m0, 2                     ; A-H1
+  movq  [dstq          ], m0
+  lea               dstq, [dstq+strideq*4]
+  movq  [dstq+stride3q ], m6                    ; A-B8, A-B7, A-B6, A-B5
+  psrldq              m6, 2                     ; A-B7, A-B6, A-B5, A-B4
+  movq  [dstq+strideq*2], m6
+  psrldq              m6, 2                     ; A-B6, A-B5, A-B4, A-B3
+  movq  [dstq+strideq  ], m6
+  psrldq              m6, 2                     ; A-B5, A-B4, A-B3, A-B2
+  movq  [dstq          ], m6
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_16x16, 4, 5, 8, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  mova                m0, [leftq]
+  movu                m7, [aboveq-1]
+  ; comments below are for a predictor like this
+  ; A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1 M1 N1 O1 P1
+  ; A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1 M1 N1
+  ; A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1
+  ; A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1
+  ; A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1
+  ; A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1
+  ; A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1
+  ; A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1
+  ; A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2
+  ; Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3
+  ; Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4
+  ; Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5
+  ; Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6
+  ; Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7
+  ; Af Bf Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8
+  ; Ag Bg Af Bf Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9
+  pshufb              m6, m7, [GLOBAL(sh_bfedcba9876543210)]
+  palignr             m5, m0, m6, 15
+  palignr             m3, m0, m6, 14
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m5, m3, m4          ; 3-tap avg B3-Bg
+  pshufb              m1, m0, [GLOBAL(sh_b123456789abcdeff)]
+  pavgb               m5, m0                            ; A1 - Ag
+
+  punpcklbw           m0, m4, m5                        ; A-B8 ... A-B1
+  punpckhbw           m4, m5                            ; A-B9 ... A-Bg
+
+  pshufb              m3, m7, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb              m5, m7, [GLOBAL(sh_b23456789abcdefff)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m5, m1          ; 3-tap avg C1-P1
+
+  pshufb              m6, m0, [GLOBAL(sh_bfedcba9876543210)]
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+  palignr             m2, m1, m6, 14
+  mova  [dstq          ], m2
+  palignr             m2, m1, m6, 12
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m1, m6, 10
+  mova  [dstq+strideq*2], m2
+  palignr             m2, m1, m6, 8
+  mova  [dstq+stride3q ], m2
+  lea               dstq, [dstq+strideq*4]
+  palignr             m2, m1, m6, 6
+  mova  [dstq          ], m2
+  palignr             m2, m1, m6, 4
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m1, m6, 2
+  mova  [dstq+strideq*2], m2
+  pshufb              m4, [GLOBAL(sh_bfedcba9876543210)]
+  mova  [dstq+stride3q ], m6
+  lea               dstq, [dstq+strideq*4]
+
+  palignr             m2, m6, m4, 14
+  mova  [dstq          ], m2
+  palignr             m2, m6, m4, 12
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m6, m4, 10
+  mova  [dstq+strideq*2], m2
+  palignr             m2, m6, m4, 8
+  mova  [dstq+stride3q ], m2
+  lea               dstq, [dstq+strideq*4]
+  palignr             m2, m6, m4, 6
+  mova  [dstq          ], m2
+  palignr             m2, m6, m4, 4
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m6, m4, 2
+  mova  [dstq+strideq*2], m2
+  mova  [dstq+stride3q ], m4
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_32x32, 4, 5, 8, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  mova                  m0, [leftq]
+  movu                  m7, [aboveq-1]
+  movu                  m1, [aboveq+15]
+
+  pshufb                m4, m1, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb                m6, m1, [GLOBAL(sh_b23456789abcdefff)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m1, m4, m6, m2          ; 3-tap avg above [high]
+
+  palignr               m3, m1, m7, 1
+  palignr               m5, m1, m7, 2
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m5, m1          ; 3-tap avg above [low]
+
+  pshufb                m7, [GLOBAL(sh_bfedcba9876543210)]
+  palignr               m5, m0, m7, 15
+  palignr               m3, m0, m7, 14
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m5, m3, m4          ; 3-tap avg B3-Bg
+  pavgb                 m5, m0                            ; A1 - Ag
+  punpcklbw             m6, m4, m5                        ; A-B8 ... A-B1
+  punpckhbw             m4, m5                            ; A-B9 ... A-Bg
+  pshufb                m6, [GLOBAL(sh_bfedcba9876543210)]
+  pshufb                m4, [GLOBAL(sh_bfedcba9876543210)]
+
+  DEFINE_ARGS dst, stride, stride3, left, line
+  lea             stride3q, [strideq*3]
+
+  palignr               m5, m2, m1, 14
+  palignr               m7, m1, m6, 14
+  mova  [dstq            ], m7
+  mova  [dstq+16         ], m5
+  palignr               m5, m2, m1, 12
+  palignr               m7, m1, m6, 12
+  mova  [dstq+strideq    ], m7
+  mova  [dstq+strideq+16 ], m5
+  palignr                m5, m2, m1, 10
+  palignr                m7, m1, m6, 10
+  mova  [dstq+strideq*2   ], m7
+  mova  [dstq+strideq*2+16], m5
+  palignr                m5, m2, m1, 8
+  palignr                m7, m1, m6, 8
+  mova  [dstq+stride3q    ], m7
+  mova  [dstq+stride3q+16 ], m5
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m5, m2, m1, 6
+  palignr                m7, m1, m6, 6
+  mova  [dstq             ], m7
+  mova  [dstq+16          ], m5
+  palignr                m5, m2, m1, 4
+  palignr                m7, m1, m6, 4
+  mova  [dstq+strideq     ], m7
+  mova  [dstq+strideq+16  ], m5
+  palignr                m5, m2, m1, 2
+  palignr                m7, m1, m6, 2
+  mova  [dstq+strideq*2   ], m7
+  mova  [dstq+strideq*2+16], m5
+  mova  [dstq+stride3q    ], m6
+  mova  [dstq+stride3q+16 ], m1
+  lea                  dstq, [dstq+strideq*4]
+
+  palignr                m5, m1, m6, 14
+  palignr                m3, m6, m4, 14
+  mova  [dstq             ], m3
+  mova  [dstq+16          ], m5
+  palignr                m5, m1, m6, 12
+  palignr                m3, m6, m4, 12
+  mova  [dstq+strideq     ], m3
+  mova  [dstq+strideq+16  ], m5
+  palignr                m5, m1, m6, 10
+  palignr                m3, m6, m4, 10
+  mova  [dstq+strideq*2   ], m3
+  mova  [dstq+strideq*2+16], m5
+  palignr                m5, m1, m6, 8
+  palignr                m3, m6, m4, 8
+  mova  [dstq+stride3q    ], m3
+  mova  [dstq+stride3q+16 ], m5
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m5, m1, m6, 6
+  palignr                m3, m6, m4, 6
+  mova  [dstq             ], m3
+  mova  [dstq+16          ], m5
+  palignr                m5, m1, m6, 4
+  palignr                m3, m6, m4, 4
+  mova  [dstq+strideq     ], m3
+  mova  [dstq+strideq+16  ], m5
+  palignr                m5, m1, m6, 2
+  palignr                m3, m6, m4, 2
+  mova  [dstq+strideq*2   ], m3
+  mova  [dstq+strideq*2+16], m5
+  mova  [dstq+stride3q    ], m4
+  mova  [dstq+stride3q+16 ], m6
+  lea               dstq, [dstq+strideq*4]
+
+  mova                   m7, [leftq]
+  mova                   m3, [leftq+16]
+  palignr                m5, m3, m7, 15
+  palignr                m0, m3, m7, 14
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m3, m5, m0, m2          ; 3-tap avg Bh -
+  pavgb                  m5, m3                            ; Ah -
+  punpcklbw              m3, m2, m5                        ; A-B8 ... A-B1
+  punpckhbw              m2, m5                            ; A-B9 ... A-Bg
+  pshufb                 m3, [GLOBAL(sh_bfedcba9876543210)]
+  pshufb                 m2, [GLOBAL(sh_bfedcba9876543210)]
+
+  palignr                m7, m6, m4, 14
+  palignr                m0, m4, m3, 14
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m6, m4, 12
+  palignr                m0, m4, m3, 12
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m6, m4, 10
+  palignr                m0, m4, m3, 10
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  palignr                m7, m6, m4, 8
+  palignr                m0, m4, m3, 8
+  mova  [dstq+stride3q    ], m0
+  mova  [dstq+stride3q+16 ], m7
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m7, m6, m4, 6
+  palignr                m0, m4, m3, 6
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m6, m4, 4
+  palignr                m0, m4, m3, 4
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m6, m4, 2
+  palignr                m0, m4, m3, 2
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  mova  [dstq+stride3q    ], m3
+  mova  [dstq+stride3q+16 ], m4
+  lea                  dstq, [dstq+strideq*4]
+
+  palignr                m7, m4, m3, 14
+  palignr                m0, m3, m2, 14
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m4, m3, 12
+  palignr                m0, m3, m2, 12
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m4, m3, 10
+  palignr                m0, m3, m2, 10
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  palignr                m7, m4, m3, 8
+  palignr                m0, m3, m2, 8
+  mova  [dstq+stride3q    ], m0
+  mova  [dstq+stride3q+16 ], m7
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m7, m4, m3, 6
+  palignr                m0, m3, m2, 6
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m4, m3, 4
+  palignr                m0, m3, m2, 4
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m4, m3, 2
+  palignr                m0, m3, m2, 2
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  mova  [dstq+stride3q    ], m2
+  mova  [dstq+stride3q+16 ], m3
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d207_predictor_8x8, 4, 5, 4, dst, stride, stride3, left, goffset
+  GET_GOT     goffsetq
+  movq                m3, [leftq]            ; abcdefgh [byte]
+  lea           stride3q, [strideq*3]
+
+  pshufb              m1, m3, [GLOBAL(sh_b2345677777777777)]
+  pshufb              m0, m3, [GLOBAL(sh_b0123456777777777)]
+  pshufb              m2, m3, [GLOBAL(sh_b1234567777777777)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m2, m1, m3
+  pavgb               m0, m2
+  punpcklbw           m0, m3        ; interleaved output
+
+  movq  [dstq          ], m0
+  psrldq              m0, 2
+  movq  [dstq+strideq  ], m0
+  psrldq              m0, 2
+  movq  [dstq+strideq*2], m0
+  psrldq              m0, 2
+  movq  [dstq+stride3q ], m0
+  lea               dstq, [dstq+strideq*4]
+  pshufhw             m0, m0, q0000 ; de, d2ef, ef, e2fg, fg, f2gh, gh, g3h, 8xh
+  psrldq              m0, 2
+  movq  [dstq          ], m0
+  psrldq              m0, 2
+  movq  [dstq+strideq  ], m0
+  psrldq              m0, 2
+  movq  [dstq+strideq*2], m0
+  psrldq              m0, 2
+  movq  [dstq+stride3q ], m0
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d207_predictor_16x16, 4, 5, 5, dst, stride, stride3, left, goffset
+  GET_GOT     goffsetq
+  lea           stride3q, [strideq*3]
+  mova                m0, [leftq]            ; abcdefghijklmnop [byte]
+  pshufb              m1, m0, [GLOBAL(sh_b123456789abcdeff)] ; bcdefghijklmnopp
+  pshufb              m2, m0, [GLOBAL(sh_b23456789abcdefff)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3
+  pavgb               m1, m0                 ; ab, bc, cd .. no, op, pp [byte]
+
+  punpckhbw           m4, m1, m3    ; interleaved input
+  punpcklbw           m1, m3        ; interleaved output
+  mova  [dstq          ], m1
+  palignr             m3, m4, m1, 2
+  mova  [dstq+strideq  ], m3
+  palignr             m3, m4, m1, 4
+  mova  [dstq+strideq*2], m3
+  palignr             m3, m4, m1, 6
+  mova  [dstq+stride3q ], m3
+  lea               dstq, [dstq+strideq*4]
+  palignr             m3, m4, m1, 8
+  mova  [dstq          ], m3
+  palignr             m3, m4, m1, 10
+  mova  [dstq+strideq  ], m3
+  palignr             m3, m4, m1, 12
+  mova  [dstq+strideq*2], m3
+  palignr             m3, m4, m1, 14
+  mova  [dstq+stride3q ], m3
+  DEFINE_ARGS dst, stride, stride3, line
+  mov              lined, 2
+  mova                m0, [GLOBAL(sh_b23456789abcdefff)]
+.loop:
+  lea               dstq, [dstq+strideq*4]
+  mova  [dstq          ], m4
+  pshufb              m4, m0
+  mova  [dstq+strideq  ], m4
+  pshufb              m4, m0
+  mova  [dstq+strideq*2], m4
+  pshufb              m4, m0
+  mova  [dstq+stride3q ], m4
+  pshufb              m4, m0
+  dec              lined
+  jnz .loop
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM ssse3
+cglobal d207_predictor_32x32, 4, 5, 8, dst, stride, stride3, left, goffset
+  GET_GOT     goffsetq
+  lea           stride3q, [strideq*3]
+  mova                m1, [leftq]              ;  0-15 [byte]
+  mova                m2, [leftq+16]           ; 16-31 [byte]
+  pshufb              m0, m2, [GLOBAL(sh_b23456789abcdefff)]
+  pshufb              m4, m2, [GLOBAL(sh_b123456789abcdeff)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m2, m4, m0, m3
+  palignr             m6, m2, m1, 1
+  palignr             m5, m2, m1, 2
+  pavgb               m2, m4         ; high 16px even lines
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m1, m6, m5, m0
+  pavgb                   m1, m6         ; low 16px even lines
+
+  punpckhbw               m6, m1, m0               ; interleaved output 2
+  punpcklbw               m1, m0                   ; interleaved output 1
+
+  punpckhbw               m7, m2, m3               ; interleaved output 4
+  punpcklbw               m2, m3                   ; interleaved output 3
+
+  ; output 1st 8 lines (and half of 2nd 8 lines)
+  DEFINE_ARGS dst, stride, stride3, dst8
+  lea                  dst8q, [dstq+strideq*8]
+  mova  [dstq              ], m1
+  mova  [dstq           +16], m6
+  mova  [dst8q             ], m6
+  palignr             m0, m6, m1, 2
+  palignr             m4, m2, m6, 2
+  mova  [dstq +strideq     ], m0
+  mova  [dstq +strideq  +16], m4
+  mova  [dst8q+strideq     ], m4
+  palignr             m0, m6, m1, 4
+  palignr             m4, m2, m6, 4
+  mova  [dstq +strideq*2   ], m0
+  mova  [dstq +strideq*2+16], m4
+  mova  [dst8q+strideq*2   ], m4
+  palignr             m0, m6, m1, 6
+  palignr             m4, m2, m6, 6
+  mova  [dstq +stride3q    ], m0
+  mova  [dstq +stride3q +16], m4
+  mova  [dst8q+stride3q    ], m4
+  lea               dstq, [dstq +strideq*4]
+  lea              dst8q, [dst8q+strideq*4]
+  palignr             m0, m6, m1, 8
+  palignr             m4, m2, m6, 8
+  mova  [dstq              ], m0
+  mova  [dstq           +16], m4
+  mova  [dst8q             ], m4
+  palignr             m0, m6, m1, 10
+  palignr             m4, m2, m6, 10
+  mova  [dstq +strideq     ], m0
+  mova  [dstq +strideq  +16], m4
+  mova  [dst8q+strideq     ], m4
+  palignr             m0, m6, m1, 12
+  palignr             m4, m2, m6, 12
+  mova  [dstq +strideq*2   ], m0
+  mova  [dstq +strideq*2+16], m4
+  mova  [dst8q+strideq*2   ], m4
+  palignr             m0, m6, m1, 14
+  palignr             m4, m2, m6, 14
+  mova  [dstq +stride3q    ], m0
+  mova  [dstq +stride3q +16], m4
+  mova  [dst8q+stride3q    ], m4
+  lea               dstq, [dstq+strideq*4]
+  lea              dst8q, [dst8q+strideq*4]
+
+  ; output 2nd half of 2nd 8 lines and half of 3rd 8 lines
+  mova  [dstq           +16], m2
+  mova  [dst8q             ], m2
+  palignr             m4, m7, m2, 2
+  mova  [dstq +strideq  +16], m4
+  mova  [dst8q+strideq     ], m4
+  palignr             m4, m7, m2, 4
+  mova  [dstq +strideq*2+16], m4
+  mova  [dst8q+strideq*2   ], m4
+  palignr             m4, m7, m2, 6
+  mova  [dstq +stride3q +16], m4
+  mova  [dst8q+stride3q    ], m4
+  lea               dstq, [dstq+strideq*4]
+  lea              dst8q, [dst8q+strideq*4]
+  palignr             m4, m7, m2, 8
+  mova  [dstq           +16], m4
+  mova  [dst8q             ], m4
+  palignr             m4, m7, m2, 10
+  mova  [dstq +strideq  +16], m4
+  mova  [dst8q+strideq     ], m4
+  palignr             m4, m7, m2, 12
+  mova  [dstq +strideq*2+16], m4
+  mova  [dst8q+strideq*2   ], m4
+  palignr             m4, m7, m2, 14
+  mova  [dstq +stride3q +16], m4
+  mova  [dst8q+stride3q    ], m4
+  lea               dstq, [dstq+strideq*4]
+  lea              dst8q, [dst8q+strideq*4]
+
+  ; output 2nd half of 3rd 8 lines and half of 4th 8 lines
+  mova                m0, [GLOBAL(sh_b23456789abcdefff)]
+  mova  [dstq           +16], m7
+  mova  [dst8q             ], m7
+  pshufb              m7, m0
+  mova  [dstq +strideq  +16], m7
+  mova  [dst8q+strideq     ], m7
+  pshufb              m7, m0
+  mova  [dstq +strideq*2+16], m7
+  mova  [dst8q+strideq*2   ], m7
+  pshufb              m7, m0
+  mova  [dstq +stride3q +16], m7
+  mova  [dst8q+stride3q    ], m7
+  pshufb              m7, m0
+  lea               dstq, [dstq+strideq*4]
+  lea              dst8q, [dst8q+strideq*4]
+  mova  [dstq           +16], m7
+  mova  [dst8q             ], m7
+  pshufb              m7, m0
+  mova  [dstq +strideq  +16], m7
+  mova  [dst8q+strideq     ], m7
+  pshufb              m7, m0
+  mova  [dstq +strideq*2+16], m7
+  mova  [dst8q+strideq*2   ], m7
+  pshufb              m7, m0
+  mova  [dstq +stride3q +16], m7
+  mova  [dst8q+stride3q    ], m7
+  pshufb              m7, m0
+  lea               dstq, [dstq+strideq*4]
+
+  ; output last half of 4th 8 lines
+  mova  [dstq           +16], m7
+  mova  [dstq +strideq  +16], m7
+  mova  [dstq +strideq*2+16], m7
+  mova  [dstq +stride3q +16], m7
+  lea               dstq, [dstq+strideq*4]
+  mova  [dstq           +16], m7
+  mova  [dstq +strideq  +16], m7
+  mova  [dstq +strideq*2+16], m7
+  mova  [dstq +stride3q +16], m7
+
+  ; done!
+  RESTORE_GOT
+  RET
diff --git a/libvpx/libvpx/vpx_dsp/x86/inv_txfm_sse2.c b/libvpx/libvpx/vpx_dsp/x86/inv_txfm_sse2.c
new file mode 100644
index 0000000..df5068c
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/inv_txfm_sse2.c
@@ -0,0 +1,4046 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/x86/inv_txfm_sse2.h"
+#include "vpx_dsp/x86/txfm_common_sse2.h"
+
+#define RECON_AND_STORE4X4(dest, in_x) \
+{                                                     \
+  __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest)); \
+  d0 = _mm_unpacklo_epi8(d0, zero); \
+  d0 = _mm_add_epi16(in_x, d0); \
+  d0 = _mm_packus_epi16(d0, d0); \
+  *(int *)(dest) = _mm_cvtsi128_si32(d0); \
+}
+
+void vpx_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i eight = _mm_set1_epi16(8);
+  const __m128i cst = _mm_setr_epi16(
+      (int16_t)cospi_16_64, (int16_t)cospi_16_64, (int16_t)cospi_16_64,
+      (int16_t)-cospi_16_64, (int16_t)cospi_24_64, (int16_t)-cospi_8_64,
+      (int16_t)cospi_8_64, (int16_t)cospi_24_64);
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i input0, input1, input2, input3;
+
+  // Rows
+  input0 = load_input_data(input);
+  input2 = load_input_data(input + 8);
+
+  // Construct i3, i1, i3, i1, i2, i0, i2, i0
+  input0 = _mm_shufflelo_epi16(input0, 0xd8);
+  input0 = _mm_shufflehi_epi16(input0, 0xd8);
+  input2 = _mm_shufflelo_epi16(input2, 0xd8);
+  input2 = _mm_shufflehi_epi16(input2, 0xd8);
+
+  input1 = _mm_unpackhi_epi32(input0, input0);
+  input0 = _mm_unpacklo_epi32(input0, input0);
+  input3 = _mm_unpackhi_epi32(input2, input2);
+  input2 = _mm_unpacklo_epi32(input2, input2);
+
+  // Stage 1
+  input0 = _mm_madd_epi16(input0, cst);
+  input1 = _mm_madd_epi16(input1, cst);
+  input2 = _mm_madd_epi16(input2, cst);
+  input3 = _mm_madd_epi16(input3, cst);
+
+  input0 = _mm_add_epi32(input0, rounding);
+  input1 = _mm_add_epi32(input1, rounding);
+  input2 = _mm_add_epi32(input2, rounding);
+  input3 = _mm_add_epi32(input3, rounding);
+
+  input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
+  input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
+  input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
+  input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
+
+  // Stage 2
+  input0 = _mm_packs_epi32(input0, input1);
+  input1 = _mm_packs_epi32(input2, input3);
+
+  // Transpose
+  input2 = _mm_unpacklo_epi16(input0, input1);
+  input3 = _mm_unpackhi_epi16(input0, input1);
+  input0 = _mm_unpacklo_epi32(input2, input3);
+  input1 = _mm_unpackhi_epi32(input2, input3);
+
+  // Switch column2, column 3, and then, we got:
+  // input2: column1, column 0;  input3: column2, column 3.
+  input1 = _mm_shuffle_epi32(input1, 0x4e);
+  input2 = _mm_add_epi16(input0, input1);
+  input3 = _mm_sub_epi16(input0, input1);
+
+  // Columns
+  // Construct i3, i1, i3, i1, i2, i0, i2, i0
+  input0 = _mm_unpacklo_epi32(input2, input2);
+  input1 = _mm_unpackhi_epi32(input2, input2);
+  input2 = _mm_unpackhi_epi32(input3, input3);
+  input3 = _mm_unpacklo_epi32(input3, input3);
+
+  // Stage 1
+  input0 = _mm_madd_epi16(input0, cst);
+  input1 = _mm_madd_epi16(input1, cst);
+  input2 = _mm_madd_epi16(input2, cst);
+  input3 = _mm_madd_epi16(input3, cst);
+
+  input0 = _mm_add_epi32(input0, rounding);
+  input1 = _mm_add_epi32(input1, rounding);
+  input2 = _mm_add_epi32(input2, rounding);
+  input3 = _mm_add_epi32(input3, rounding);
+
+  input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
+  input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
+  input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
+  input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
+
+  // Stage 2
+  input0 = _mm_packs_epi32(input0, input2);
+  input1 = _mm_packs_epi32(input1, input3);
+
+  // Transpose
+  input2 = _mm_unpacklo_epi16(input0, input1);
+  input3 = _mm_unpackhi_epi16(input0, input1);
+  input0 = _mm_unpacklo_epi32(input2, input3);
+  input1 = _mm_unpackhi_epi32(input2, input3);
+
+  // Switch column2, column 3, and then, we got:
+  // input2: column1, column 0;  input3: column2, column 3.
+  input1 = _mm_shuffle_epi32(input1, 0x4e);
+  input2 = _mm_add_epi16(input0, input1);
+  input3 = _mm_sub_epi16(input0, input1);
+
+  // Final round and shift
+  input2 = _mm_add_epi16(input2, eight);
+  input3 = _mm_add_epi16(input3, eight);
+
+  input2 = _mm_srai_epi16(input2, 4);
+  input3 = _mm_srai_epi16(input3, 4);
+
+  // Reconstruction and Store
+  {
+    __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest));
+    __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2));
+    d0 = _mm_unpacklo_epi32(d0,
+                            _mm_cvtsi32_si128(*(const int *)(dest + stride)));
+    d2 = _mm_unpacklo_epi32(
+        _mm_cvtsi32_si128(*(const int *)(dest + stride * 3)), d2);
+    d0 = _mm_unpacklo_epi8(d0, zero);
+    d2 = _mm_unpacklo_epi8(d2, zero);
+    d0 = _mm_add_epi16(d0, input2);
+    d2 = _mm_add_epi16(d2, input3);
+    d0 = _mm_packus_epi16(d0, d2);
+    // store input0
+    *(int *)dest = _mm_cvtsi128_si32(d0);
+    // store input1
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride) = _mm_cvtsi128_si32(d0);
+    // store input2
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0);
+    // store input3
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0);
+  }
+}
+
+void vpx_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a;
+
+  a = (int)dct_const_round_shift(input[0] * cospi_16_64);
+  a = (int)dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 4);
+
+  dc_value = _mm_set1_epi16(a);
+
+  RECON_AND_STORE4X4(dest + 0 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 1 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 2 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 3 * stride, dc_value);
+}
+
+static INLINE void transpose_4x4(__m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
+  const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
+
+  res[0] = _mm_unpacklo_epi16(tr0_0, tr0_1);
+  res[1] = _mm_unpackhi_epi16(tr0_0, tr0_1);
+}
+
+void idct4_sse2(__m128i *in) {
+  const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8];
+
+  transpose_4x4(in);
+  // stage 1
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpackhi_epi16(in[0], in[1]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[2] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+
+  u[0] = _mm_packs_epi32(v[0], v[1]);
+  u[1] = _mm_packs_epi32(v[3], v[2]);
+
+  // stage 2
+  in[0] = _mm_add_epi16(u[0], u[1]);
+  in[1] = _mm_sub_epi16(u[0], u[1]);
+  in[1] = _mm_shuffle_epi32(in[1], 0x4E);
+}
+
+void iadst4_sse2(__m128i *in) {
+  const __m128i k__sinpi_p01_p04 = pair_set_epi16(sinpi_1_9, sinpi_4_9);
+  const __m128i k__sinpi_p03_p02 = pair_set_epi16(sinpi_3_9, sinpi_2_9);
+  const __m128i k__sinpi_p02_m01 = pair_set_epi16(sinpi_2_9, -sinpi_1_9);
+  const __m128i k__sinpi_p03_m04 = pair_set_epi16(sinpi_3_9, -sinpi_4_9);
+  const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8], in7;
+
+  transpose_4x4(in);
+  in7 = _mm_srli_si128(in[1], 8);
+  in7 = _mm_add_epi16(in7, in[0]);
+  in7 = _mm_sub_epi16(in7, in[1]);
+
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpackhi_epi16(in[0], in[1]);
+  u[2] = _mm_unpacklo_epi16(in7, kZero);
+  u[3] = _mm_unpackhi_epi16(in[0], kZero);
+
+  v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04);  // s0 + s3
+  v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02);  // s2 + s5
+  v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x2
+  v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01);  // s1 - s4
+  v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04);  // s2 - s6
+  v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s2
+
+  u[0] = _mm_add_epi32(v[0], v[1]);
+  u[1] = _mm_add_epi32(v[3], v[4]);
+  u[2] = v[2];
+  u[3] = _mm_add_epi32(u[0], u[1]);
+  u[4] = _mm_slli_epi32(v[5], 2);
+  u[5] = _mm_add_epi32(u[3], v[5]);
+  u[6] = _mm_sub_epi32(u[5], u[4]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[1]);
+  in[1] = _mm_packs_epi32(u[2], u[3]);
+}
+
+#define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, \
+                      out0, out1, out2, out3, out4, out5, out6, out7) \
+  {                                                     \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+    const __m128i tr0_2 = _mm_unpackhi_epi16(in0, in1); \
+    const __m128i tr0_3 = _mm_unpackhi_epi16(in2, in3); \
+    const __m128i tr0_4 = _mm_unpacklo_epi16(in4, in5); \
+    const __m128i tr0_5 = _mm_unpacklo_epi16(in6, in7); \
+    const __m128i tr0_6 = _mm_unpackhi_epi16(in4, in5); \
+    const __m128i tr0_7 = _mm_unpackhi_epi16(in6, in7); \
+                                                        \
+    const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); \
+    const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); \
+    const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \
+    const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); \
+    const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \
+    const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); \
+                                                            \
+    out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \
+    out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \
+    out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \
+    out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
+    out4 = _mm_unpacklo_epi64(tr1_1, tr1_5); \
+    out5 = _mm_unpackhi_epi64(tr1_1, tr1_5); \
+    out6 = _mm_unpacklo_epi64(tr1_3, tr1_7); \
+    out7 = _mm_unpackhi_epi64(tr1_3, tr1_7); \
+  }
+
+#define TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, \
+                         out0, out1, out2, out3) \
+  {                                              \
+    const __m128i tr0_0 = _mm_unpackhi_epi16(tmp0, tmp1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(tmp1, tmp0); \
+    const __m128i tr0_4 = _mm_unpacklo_epi16(tmp2, tmp3); \
+    const __m128i tr0_5 = _mm_unpackhi_epi16(tmp3, tmp2); \
+    \
+    const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \
+    const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \
+    \
+    out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \
+    out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \
+    out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \
+    out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
+  }
+
+#define TRANSPOSE_8X8_10(in0, in1, in2, in3, out0, out1) \
+  {                                            \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+    out0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
+    out1 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
+  }
+
+// Define Macro for multiplying elements by constants and adding them together.
+#define MULTIPLICATION_AND_ADD(lo_0, hi_0, lo_1, hi_1, \
+                               cst0, cst1, cst2, cst3, res0, res1, res2, res3) \
+  {   \
+      tmp0 = _mm_madd_epi16(lo_0, cst0); \
+      tmp1 = _mm_madd_epi16(hi_0, cst0); \
+      tmp2 = _mm_madd_epi16(lo_0, cst1); \
+      tmp3 = _mm_madd_epi16(hi_0, cst1); \
+      tmp4 = _mm_madd_epi16(lo_1, cst2); \
+      tmp5 = _mm_madd_epi16(hi_1, cst2); \
+      tmp6 = _mm_madd_epi16(lo_1, cst3); \
+      tmp7 = _mm_madd_epi16(hi_1, cst3); \
+      \
+      tmp0 = _mm_add_epi32(tmp0, rounding); \
+      tmp1 = _mm_add_epi32(tmp1, rounding); \
+      tmp2 = _mm_add_epi32(tmp2, rounding); \
+      tmp3 = _mm_add_epi32(tmp3, rounding); \
+      tmp4 = _mm_add_epi32(tmp4, rounding); \
+      tmp5 = _mm_add_epi32(tmp5, rounding); \
+      tmp6 = _mm_add_epi32(tmp6, rounding); \
+      tmp7 = _mm_add_epi32(tmp7, rounding); \
+      \
+      tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+      tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+      tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+      tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+      tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); \
+      tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); \
+      tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); \
+      tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); \
+      \
+      res0 = _mm_packs_epi32(tmp0, tmp1); \
+      res1 = _mm_packs_epi32(tmp2, tmp3); \
+      res2 = _mm_packs_epi32(tmp4, tmp5); \
+      res3 = _mm_packs_epi32(tmp6, tmp7); \
+  }
+
+#define MULTIPLICATION_AND_ADD_2(lo_0, hi_0, cst0, cst1, res0, res1) \
+  {   \
+      tmp0 = _mm_madd_epi16(lo_0, cst0); \
+      tmp1 = _mm_madd_epi16(hi_0, cst0); \
+      tmp2 = _mm_madd_epi16(lo_0, cst1); \
+      tmp3 = _mm_madd_epi16(hi_0, cst1); \
+      \
+      tmp0 = _mm_add_epi32(tmp0, rounding); \
+      tmp1 = _mm_add_epi32(tmp1, rounding); \
+      tmp2 = _mm_add_epi32(tmp2, rounding); \
+      tmp3 = _mm_add_epi32(tmp3, rounding); \
+      \
+      tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+      tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+      tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+      tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+      \
+      res0 = _mm_packs_epi32(tmp0, tmp1); \
+      res1 = _mm_packs_epi32(tmp2, tmp3); \
+  }
+
+#define IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, \
+              out0, out1, out2, out3, out4, out5, out6, out7)  \
+  { \
+  /* Stage1 */      \
+  { \
+    const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7); \
+    const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7); \
+    const __m128i lo_35 = _mm_unpacklo_epi16(in3, in5); \
+    const __m128i hi_35 = _mm_unpackhi_epi16(in3, in5); \
+    \
+    MULTIPLICATION_AND_ADD(lo_17, hi_17, lo_35, hi_35, stg1_0, \
+                          stg1_1, stg1_2, stg1_3, stp1_4,      \
+                          stp1_7, stp1_5, stp1_6)              \
+  } \
+    \
+  /* Stage2 */ \
+  { \
+    const __m128i lo_04 = _mm_unpacklo_epi16(in0, in4); \
+    const __m128i hi_04 = _mm_unpackhi_epi16(in0, in4); \
+    const __m128i lo_26 = _mm_unpacklo_epi16(in2, in6); \
+    const __m128i hi_26 = _mm_unpackhi_epi16(in2, in6); \
+    \
+    MULTIPLICATION_AND_ADD(lo_04, hi_04, lo_26, hi_26, stg2_0, \
+                           stg2_1, stg2_2, stg2_3, stp2_0,     \
+                           stp2_1, stp2_2, stp2_3)             \
+    \
+    stp2_4 = _mm_adds_epi16(stp1_4, stp1_5); \
+    stp2_5 = _mm_subs_epi16(stp1_4, stp1_5); \
+    stp2_6 = _mm_subs_epi16(stp1_7, stp1_6); \
+    stp2_7 = _mm_adds_epi16(stp1_7, stp1_6); \
+  } \
+    \
+  /* Stage3 */ \
+  { \
+    const __m128i lo_56 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+    const __m128i hi_56 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+    \
+    stp1_0 = _mm_adds_epi16(stp2_0, stp2_3); \
+    stp1_1 = _mm_adds_epi16(stp2_1, stp2_2); \
+    stp1_2 = _mm_subs_epi16(stp2_1, stp2_2); \
+    stp1_3 = _mm_subs_epi16(stp2_0, stp2_3); \
+    \
+    tmp0 = _mm_madd_epi16(lo_56, stg2_1); \
+    tmp1 = _mm_madd_epi16(hi_56, stg2_1); \
+    tmp2 = _mm_madd_epi16(lo_56, stg2_0); \
+    tmp3 = _mm_madd_epi16(hi_56, stg2_0); \
+    \
+    tmp0 = _mm_add_epi32(tmp0, rounding); \
+    tmp1 = _mm_add_epi32(tmp1, rounding); \
+    tmp2 = _mm_add_epi32(tmp2, rounding); \
+    tmp3 = _mm_add_epi32(tmp3, rounding); \
+    \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+    \
+    stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+    stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+  } \
+  \
+  /* Stage4  */ \
+  out0 = _mm_adds_epi16(stp1_0, stp2_7); \
+  out1 = _mm_adds_epi16(stp1_1, stp1_6); \
+  out2 = _mm_adds_epi16(stp1_2, stp1_5); \
+  out3 = _mm_adds_epi16(stp1_3, stp2_4); \
+  out4 = _mm_subs_epi16(stp1_3, stp2_4); \
+  out5 = _mm_subs_epi16(stp1_2, stp1_5); \
+  out6 = _mm_subs_epi16(stp1_1, stp1_6); \
+  out7 = _mm_subs_epi16(stp1_0, stp2_7); \
+  }
+
+void vpx_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  // Load input data.
+  in0 = load_input_data(input);
+  in1 = load_input_data(input + 8 * 1);
+  in2 = load_input_data(input + 8 * 2);
+  in3 = load_input_data(input + 8 * 3);
+  in4 = load_input_data(input + 8 * 4);
+  in5 = load_input_data(input + 8 * 5);
+  in6 = load_input_data(input + 8 * 6);
+  in7 = load_input_data(input + 8 * 7);
+
+  // 2-D
+  for (i = 0; i < 2; i++) {
+    // 8x8 Transpose is copied from vpx_fdct8x8_sse2()
+    TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7,
+                  in0, in1, in2, in3, in4, in5, in6, in7);
+
+    // 4-stage 1D idct8x8
+    IDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+          in0, in1, in2, in3, in4, in5, in6, in7);
+  }
+
+  // Final rounding and shift
+  in0 = _mm_adds_epi16(in0, final_rounding);
+  in1 = _mm_adds_epi16(in1, final_rounding);
+  in2 = _mm_adds_epi16(in2, final_rounding);
+  in3 = _mm_adds_epi16(in3, final_rounding);
+  in4 = _mm_adds_epi16(in4, final_rounding);
+  in5 = _mm_adds_epi16(in5, final_rounding);
+  in6 = _mm_adds_epi16(in6, final_rounding);
+  in7 = _mm_adds_epi16(in7, final_rounding);
+
+  in0 = _mm_srai_epi16(in0, 5);
+  in1 = _mm_srai_epi16(in1, 5);
+  in2 = _mm_srai_epi16(in2, 5);
+  in3 = _mm_srai_epi16(in3, 5);
+  in4 = _mm_srai_epi16(in4, 5);
+  in5 = _mm_srai_epi16(in5, 5);
+  in6 = _mm_srai_epi16(in6, 5);
+  in7 = _mm_srai_epi16(in7, 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in0);
+  RECON_AND_STORE(dest + 1 * stride, in1);
+  RECON_AND_STORE(dest + 2 * stride, in2);
+  RECON_AND_STORE(dest + 3 * stride, in3);
+  RECON_AND_STORE(dest + 4 * stride, in4);
+  RECON_AND_STORE(dest + 5 * stride, in5);
+  RECON_AND_STORE(dest + 6 * stride, in6);
+  RECON_AND_STORE(dest + 7 * stride, in7);
+}
+
+void vpx_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a;
+
+  a = (int)dct_const_round_shift(input[0] * cospi_16_64);
+  a = (int)dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 5);
+
+  dc_value = _mm_set1_epi16(a);
+
+  RECON_AND_STORE(dest + 0 * stride, dc_value);
+  RECON_AND_STORE(dest + 1 * stride, dc_value);
+  RECON_AND_STORE(dest + 2 * stride, dc_value);
+  RECON_AND_STORE(dest + 3 * stride, dc_value);
+  RECON_AND_STORE(dest + 4 * stride, dc_value);
+  RECON_AND_STORE(dest + 5 * stride, dc_value);
+  RECON_AND_STORE(dest + 6 * stride, dc_value);
+  RECON_AND_STORE(dest + 7 * stride, dc_value);
+}
+
+void idct8_sse2(__m128i *in) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+  // 8x8 Transpose is copied from vpx_fdct8x8_sse2()
+  TRANSPOSE_8X8(in[0], in[1], in[2], in[3], in[4], in[5], in[6], in[7],
+                in0, in1, in2, in3, in4, in5, in6, in7);
+
+  // 4-stage 1D idct8x8
+  IDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+        in[0], in[1], in[2], in[3], in[4], in[5], in[6], in[7]);
+}
+
+void iadst8_sse2(__m128i *in) {
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__const_0 = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+
+  // transpose
+  array_transpose_8x8(in, in);
+
+  // properly aligned for butterfly input
+  in0 = in[7];
+  in1 = in[0];
+  in2 = in[5];
+  in3 = in[2];
+  in4 = in[3];
+  in5 = in[4];
+  in6 = in[1];
+  in7 = in[6];
+
+  // column transformation
+  // stage 1
+  // interleave and multiply/add into 32-bit integer
+  s0 = _mm_unpacklo_epi16(in0, in1);
+  s1 = _mm_unpackhi_epi16(in0, in1);
+  s2 = _mm_unpacklo_epi16(in2, in3);
+  s3 = _mm_unpackhi_epi16(in2, in3);
+  s4 = _mm_unpacklo_epi16(in4, in5);
+  s5 = _mm_unpackhi_epi16(in4, in5);
+  s6 = _mm_unpacklo_epi16(in6, in7);
+  s7 = _mm_unpackhi_epi16(in6, in7);
+
+  u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
+  u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
+  u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
+  u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
+  u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
+  u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
+  u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
+  u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
+  u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
+  u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
+  u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
+  u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
+  u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
+  u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
+  u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
+  u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
+
+  // addition
+  w0 = _mm_add_epi32(u0, u8);
+  w1 = _mm_add_epi32(u1, u9);
+  w2 = _mm_add_epi32(u2, u10);
+  w3 = _mm_add_epi32(u3, u11);
+  w4 = _mm_add_epi32(u4, u12);
+  w5 = _mm_add_epi32(u5, u13);
+  w6 = _mm_add_epi32(u6, u14);
+  w7 = _mm_add_epi32(u7, u15);
+  w8 = _mm_sub_epi32(u0, u8);
+  w9 = _mm_sub_epi32(u1, u9);
+  w10 = _mm_sub_epi32(u2, u10);
+  w11 = _mm_sub_epi32(u3, u11);
+  w12 = _mm_sub_epi32(u4, u12);
+  w13 = _mm_sub_epi32(u5, u13);
+  w14 = _mm_sub_epi32(u6, u14);
+  w15 = _mm_sub_epi32(u7, u15);
+
+  // shift and rounding
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+  v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
+  v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
+  v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
+  v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
+  v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
+  v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
+  v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
+  v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+  u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
+  u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
+  u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
+  u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
+  u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
+  u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
+  u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
+  u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
+
+  // back to 16-bit and pack 8 integers into __m128i
+  in[0] = _mm_packs_epi32(u0, u1);
+  in[1] = _mm_packs_epi32(u2, u3);
+  in[2] = _mm_packs_epi32(u4, u5);
+  in[3] = _mm_packs_epi32(u6, u7);
+  in[4] = _mm_packs_epi32(u8, u9);
+  in[5] = _mm_packs_epi32(u10, u11);
+  in[6] = _mm_packs_epi32(u12, u13);
+  in[7] = _mm_packs_epi32(u14, u15);
+
+  // stage 2
+  s0 = _mm_add_epi16(in[0], in[2]);
+  s1 = _mm_add_epi16(in[1], in[3]);
+  s2 = _mm_sub_epi16(in[0], in[2]);
+  s3 = _mm_sub_epi16(in[1], in[3]);
+  u0 = _mm_unpacklo_epi16(in[4], in[5]);
+  u1 = _mm_unpackhi_epi16(in[4], in[5]);
+  u2 = _mm_unpacklo_epi16(in[6], in[7]);
+  u3 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
+  v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
+  v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
+  v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
+  v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
+  v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
+  v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
+  v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
+
+  w0 = _mm_add_epi32(v0, v4);
+  w1 = _mm_add_epi32(v1, v5);
+  w2 = _mm_add_epi32(v2, v6);
+  w3 = _mm_add_epi32(v3, v7);
+  w4 = _mm_sub_epi32(v0, v4);
+  w5 = _mm_sub_epi32(v1, v5);
+  w6 = _mm_sub_epi32(v2, v6);
+  w7 = _mm_sub_epi32(v3, v7);
+
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  // back to 16-bit intergers
+  s4 = _mm_packs_epi32(u0, u1);
+  s5 = _mm_packs_epi32(u2, u3);
+  s6 = _mm_packs_epi32(u4, u5);
+  s7 = _mm_packs_epi32(u6, u7);
+
+  // stage 3
+  u0 = _mm_unpacklo_epi16(s2, s3);
+  u1 = _mm_unpackhi_epi16(s2, s3);
+  u2 = _mm_unpacklo_epi16(s6, s7);
+  u3 = _mm_unpackhi_epi16(s6, s7);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
+  v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
+  v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
+  v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
+  v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
+  v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
+  v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
+  v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
+
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  s2 = _mm_packs_epi32(v0, v1);
+  s3 = _mm_packs_epi32(v2, v3);
+  s6 = _mm_packs_epi32(v4, v5);
+  s7 = _mm_packs_epi32(v6, v7);
+
+  in[0] = s0;
+  in[1] = _mm_sub_epi16(k__const_0, s4);
+  in[2] = s6;
+  in[3] = _mm_sub_epi16(k__const_0, s2);
+  in[4] = s3;
+  in[5] = _mm_sub_epi16(k__const_0, s7);
+  in[6] = s5;
+  in[7] = _mm_sub_epi16(k__const_0, s1);
+}
+
+void vpx_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg3_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+  // Rows. Load 4-row input data.
+  in0 = load_input_data(input);
+  in1 = load_input_data(input + 8 * 1);
+  in2 = load_input_data(input + 8 * 2);
+  in3 = load_input_data(input + 8 * 3);
+
+  // 8x4 Transpose
+  TRANSPOSE_8X8_10(in0, in1, in2, in3, in0, in1);
+  // Stage1
+  {
+    const __m128i lo_17 = _mm_unpackhi_epi16(in0, zero);
+    const __m128i lo_35 = _mm_unpackhi_epi16(in1, zero);
+
+    tmp0 = _mm_madd_epi16(lo_17, stg1_0);
+    tmp2 = _mm_madd_epi16(lo_17, stg1_1);
+    tmp4 = _mm_madd_epi16(lo_35, stg1_2);
+    tmp6 = _mm_madd_epi16(lo_35, stg1_3);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp1_4 = _mm_packs_epi32(tmp0, tmp2);
+    stp1_5 = _mm_packs_epi32(tmp4, tmp6);
+  }
+
+  // Stage2
+  {
+    const __m128i lo_04 = _mm_unpacklo_epi16(in0, zero);
+    const __m128i lo_26 = _mm_unpacklo_epi16(in1, zero);
+
+    tmp0 = _mm_madd_epi16(lo_04, stg2_0);
+    tmp2 = _mm_madd_epi16(lo_04, stg2_1);
+    tmp4 = _mm_madd_epi16(lo_26, stg2_2);
+    tmp6 = _mm_madd_epi16(lo_26, stg2_3);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp2_0 = _mm_packs_epi32(tmp0, tmp2);
+    stp2_2 = _mm_packs_epi32(tmp6, tmp4);
+
+    tmp0 = _mm_adds_epi16(stp1_4, stp1_5);
+    tmp1 = _mm_subs_epi16(stp1_4, stp1_5);
+
+    stp2_4 = tmp0;
+    stp2_5 = _mm_unpacklo_epi64(tmp1, zero);
+    stp2_6 = _mm_unpackhi_epi64(tmp1, zero);
+  }
+
+  // Stage3
+  {
+    const __m128i lo_56 = _mm_unpacklo_epi16(stp2_5, stp2_6);
+
+    tmp4 = _mm_adds_epi16(stp2_0, stp2_2);
+    tmp6 = _mm_subs_epi16(stp2_0, stp2_2);
+
+    stp1_2 = _mm_unpackhi_epi64(tmp6, tmp4);
+    stp1_3 = _mm_unpacklo_epi64(tmp6, tmp4);
+
+    tmp0 = _mm_madd_epi16(lo_56, stg3_0);
+    tmp2 = _mm_madd_epi16(lo_56, stg2_0);  // stg3_1 = stg2_0
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+
+    stp1_5 = _mm_packs_epi32(tmp0, tmp2);
+  }
+
+  // Stage4
+  tmp0 = _mm_adds_epi16(stp1_3, stp2_4);
+  tmp1 = _mm_adds_epi16(stp1_2, stp1_5);
+  tmp2 = _mm_subs_epi16(stp1_3, stp2_4);
+  tmp3 = _mm_subs_epi16(stp1_2, stp1_5);
+
+  TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, in0, in1, in2, in3)
+
+  IDCT8(in0, in1, in2, in3, zero, zero, zero, zero,
+        in0, in1, in2, in3, in4, in5, in6, in7);
+  // Final rounding and shift
+  in0 = _mm_adds_epi16(in0, final_rounding);
+  in1 = _mm_adds_epi16(in1, final_rounding);
+  in2 = _mm_adds_epi16(in2, final_rounding);
+  in3 = _mm_adds_epi16(in3, final_rounding);
+  in4 = _mm_adds_epi16(in4, final_rounding);
+  in5 = _mm_adds_epi16(in5, final_rounding);
+  in6 = _mm_adds_epi16(in6, final_rounding);
+  in7 = _mm_adds_epi16(in7, final_rounding);
+
+  in0 = _mm_srai_epi16(in0, 5);
+  in1 = _mm_srai_epi16(in1, 5);
+  in2 = _mm_srai_epi16(in2, 5);
+  in3 = _mm_srai_epi16(in3, 5);
+  in4 = _mm_srai_epi16(in4, 5);
+  in5 = _mm_srai_epi16(in5, 5);
+  in6 = _mm_srai_epi16(in6, 5);
+  in7 = _mm_srai_epi16(in7, 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in0);
+  RECON_AND_STORE(dest + 1 * stride, in1);
+  RECON_AND_STORE(dest + 2 * stride, in2);
+  RECON_AND_STORE(dest + 3 * stride, in3);
+  RECON_AND_STORE(dest + 4 * stride, in4);
+  RECON_AND_STORE(dest + 5 * stride, in5);
+  RECON_AND_STORE(dest + 6 * stride, in6);
+  RECON_AND_STORE(dest + 7 * stride, in7);
+}
+
+#define IDCT16 \
+  /* Stage2 */ \
+  { \
+    const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], in[15]); \
+    const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], in[15]); \
+    const __m128i lo_9_7 = _mm_unpacklo_epi16(in[9], in[7]);   \
+    const __m128i hi_9_7 = _mm_unpackhi_epi16(in[9], in[7]);   \
+    const __m128i lo_5_11 = _mm_unpacklo_epi16(in[5], in[11]); \
+    const __m128i hi_5_11 = _mm_unpackhi_epi16(in[5], in[11]); \
+    const __m128i lo_13_3 = _mm_unpacklo_epi16(in[13], in[3]); \
+    const __m128i hi_13_3 = _mm_unpackhi_epi16(in[13], in[3]); \
+    \
+    MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_9_7, hi_9_7, \
+                           stg2_0, stg2_1, stg2_2, stg2_3, \
+                           stp2_8, stp2_15, stp2_9, stp2_14) \
+    \
+    MULTIPLICATION_AND_ADD(lo_5_11, hi_5_11, lo_13_3, hi_13_3, \
+                           stg2_4, stg2_5, stg2_6, stg2_7, \
+                           stp2_10, stp2_13, stp2_11, stp2_12) \
+  } \
+    \
+  /* Stage3 */ \
+  { \
+    const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], in[14]); \
+    const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], in[14]); \
+    const __m128i lo_10_6 = _mm_unpacklo_epi16(in[10], in[6]); \
+    const __m128i hi_10_6 = _mm_unpackhi_epi16(in[10], in[6]); \
+    \
+    MULTIPLICATION_AND_ADD(lo_2_14, hi_2_14, lo_10_6, hi_10_6, \
+                           stg3_0, stg3_1, stg3_2, stg3_3, \
+                           stp1_4, stp1_7, stp1_5, stp1_6) \
+    \
+    stp1_8_0 = _mm_add_epi16(stp2_8, stp2_9);  \
+    stp1_9 = _mm_sub_epi16(stp2_8, stp2_9);    \
+    stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \
+    stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \
+    \
+    stp1_12_0 = _mm_add_epi16(stp2_12, stp2_13); \
+    stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \
+    stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \
+    stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \
+  } \
+  \
+  /* Stage4 */ \
+  { \
+    const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], in[8]); \
+    const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], in[8]); \
+    const __m128i lo_4_12 = _mm_unpacklo_epi16(in[4], in[12]); \
+    const __m128i hi_4_12 = _mm_unpackhi_epi16(in[4], in[12]); \
+    \
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
+    const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+    \
+    MULTIPLICATION_AND_ADD(lo_0_8, hi_0_8, lo_4_12, hi_4_12, \
+                           stg4_0, stg4_1, stg4_2, stg4_3, \
+                           stp2_0, stp2_1, stp2_2, stp2_3) \
+    \
+    stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
+    stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
+    stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
+    stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
+    \
+    MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, \
+                           stg4_4, stg4_5, stg4_6, stg4_7, \
+                           stp2_9, stp2_14, stp2_10, stp2_13) \
+  } \
+    \
+  /* Stage5 */ \
+  { \
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+    const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+    \
+    stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
+    stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
+    stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
+    stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
+    \
+    tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+    tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+    tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+    tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+    \
+    tmp0 = _mm_add_epi32(tmp0, rounding); \
+    tmp1 = _mm_add_epi32(tmp1, rounding); \
+    tmp2 = _mm_add_epi32(tmp2, rounding); \
+    tmp3 = _mm_add_epi32(tmp3, rounding); \
+    \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+    \
+    stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+    stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+    \
+    stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11);  \
+    stp1_9 = _mm_add_epi16(stp2_9, stp2_10);    \
+    stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);   \
+    stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \
+    \
+    stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \
+    stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);   \
+    stp1_14 = _mm_add_epi16(stp2_14, stp2_13);   \
+    stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \
+  } \
+    \
+  /* Stage6 */ \
+  { \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+    const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+    \
+    stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \
+    stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+    stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+    stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \
+    stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \
+    stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+    stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+    stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \
+    \
+    MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                           stg6_0, stg4_0, stg6_0, stg4_0, \
+                           stp2_10, stp2_13, stp2_11, stp2_12) \
+  }
+
+#define IDCT16_10 \
+    /* Stage2 */ \
+    { \
+      const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], zero); \
+      const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], zero); \
+      const __m128i lo_13_3 = _mm_unpacklo_epi16(zero, in[3]); \
+      const __m128i hi_13_3 = _mm_unpackhi_epi16(zero, in[3]); \
+      \
+      MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_13_3, hi_13_3, \
+                             stg2_0, stg2_1, stg2_6, stg2_7, \
+                             stp1_8_0, stp1_15, stp1_11, stp1_12_0) \
+    } \
+      \
+    /* Stage3 */ \
+    { \
+      const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], zero); \
+      const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], zero); \
+      \
+      MULTIPLICATION_AND_ADD_2(lo_2_14, hi_2_14, \
+                               stg3_0, stg3_1,  \
+                               stp2_4, stp2_7) \
+      \
+      stp1_9  =  stp1_8_0; \
+      stp1_10 =  stp1_11;  \
+      \
+      stp1_13 = stp1_12_0; \
+      stp1_14 = stp1_15;   \
+    } \
+    \
+    /* Stage4 */ \
+    { \
+      const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero); \
+      const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], zero); \
+      \
+      const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
+      const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
+      const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+      const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+      \
+      MULTIPLICATION_AND_ADD_2(lo_0_8, hi_0_8, \
+                               stg4_0, stg4_1, \
+                               stp1_0, stp1_1) \
+      stp2_5 = stp2_4; \
+      stp2_6 = stp2_7; \
+      \
+      MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, \
+                             stg4_4, stg4_5, stg4_6, stg4_7, \
+                             stp2_9, stp2_14, stp2_10, stp2_13) \
+    } \
+      \
+    /* Stage5 */ \
+    { \
+      const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+      const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+      \
+      stp1_2 = stp1_1; \
+      stp1_3 = stp1_0; \
+      \
+      tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+      tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+      tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+      tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+      \
+      tmp0 = _mm_add_epi32(tmp0, rounding); \
+      tmp1 = _mm_add_epi32(tmp1, rounding); \
+      tmp2 = _mm_add_epi32(tmp2, rounding); \
+      tmp3 = _mm_add_epi32(tmp3, rounding); \
+      \
+      tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+      tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+      tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+      tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+      \
+      stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+      stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+      \
+      stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11);  \
+      stp1_9 = _mm_add_epi16(stp2_9, stp2_10);    \
+      stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);   \
+      stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \
+      \
+      stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \
+      stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);   \
+      stp1_14 = _mm_add_epi16(stp2_14, stp2_13);   \
+      stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \
+    } \
+      \
+    /* Stage6 */ \
+    { \
+      const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+      const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+      const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+      const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+      \
+      stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \
+      stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+      stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+      stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \
+      stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \
+      stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+      stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+      stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \
+      \
+      MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                             stg6_0, stg4_0, stg6_0, stg4_0, \
+                             stp2_10, stp2_13, stp2_11, stp2_12) \
+    }
+
+void vpx_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest,
+                                int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[16], l[16], r[16], *curr1;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_8_0, stp1_12_0;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  curr1 = l;
+  for (i = 0; i < 2; i++) {
+    // 1-D idct
+
+    // Load input data.
+    in[0] = load_input_data(input);
+    in[8] = load_input_data(input + 8 * 1);
+    in[1] = load_input_data(input + 8 * 2);
+    in[9] = load_input_data(input + 8 * 3);
+    in[2] = load_input_data(input + 8 * 4);
+    in[10] = load_input_data(input + 8 * 5);
+    in[3] = load_input_data(input + 8 * 6);
+    in[11] = load_input_data(input + 8 * 7);
+    in[4] = load_input_data(input + 8 * 8);
+    in[12] = load_input_data(input + 8 * 9);
+    in[5] = load_input_data(input + 8 * 10);
+    in[13] = load_input_data(input + 8 * 11);
+    in[6] = load_input_data(input + 8 * 12);
+    in[14] = load_input_data(input + 8 * 13);
+    in[7] = load_input_data(input + 8 * 14);
+    in[15] = load_input_data(input + 8 * 15);
+
+    array_transpose_8x8(in, in);
+    array_transpose_8x8(in + 8, in + 8);
+
+    IDCT16
+
+    // Stage7
+    curr1[0] = _mm_add_epi16(stp2_0, stp1_15);
+    curr1[1] = _mm_add_epi16(stp2_1, stp1_14);
+    curr1[2] = _mm_add_epi16(stp2_2, stp2_13);
+    curr1[3] = _mm_add_epi16(stp2_3, stp2_12);
+    curr1[4] = _mm_add_epi16(stp2_4, stp2_11);
+    curr1[5] = _mm_add_epi16(stp2_5, stp2_10);
+    curr1[6] = _mm_add_epi16(stp2_6, stp1_9);
+    curr1[7] = _mm_add_epi16(stp2_7, stp1_8);
+    curr1[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    curr1[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    curr1[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    curr1[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    curr1[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    curr1[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    curr1[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    curr1[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    curr1 = r;
+    input += 128;
+  }
+  for (i = 0; i < 2; i++) {
+    int j;
+    // 1-D idct
+    array_transpose_8x8(l + i * 8, in);
+    array_transpose_8x8(r + i * 8, in + 8);
+
+    IDCT16
+
+    // 2-D
+    in[0] = _mm_add_epi16(stp2_0, stp1_15);
+    in[1] = _mm_add_epi16(stp2_1, stp1_14);
+    in[2] = _mm_add_epi16(stp2_2, stp2_13);
+    in[3] = _mm_add_epi16(stp2_3, stp2_12);
+    in[4] = _mm_add_epi16(stp2_4, stp2_11);
+    in[5] = _mm_add_epi16(stp2_5, stp2_10);
+    in[6] = _mm_add_epi16(stp2_6, stp1_9);
+    in[7] = _mm_add_epi16(stp2_7, stp1_8);
+    in[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    in[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    in[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    in[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    in[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    in[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    in[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    in[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    for (j = 0; j < 16; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void vpx_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest,
+                              int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a, i;
+
+  a = (int)dct_const_round_shift(input[0] * cospi_16_64);
+  a = (int)dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 6);
+
+  dc_value = _mm_set1_epi16(a);
+
+  for (i = 0; i < 16; ++i) {
+    RECON_AND_STORE(dest +  0, dc_value);
+    RECON_AND_STORE(dest +  8, dc_value);
+    dest += stride;
+  }
+}
+
+static void iadst16_8col(__m128i *in) {
+  // perform 16x16 1-D ADST for 8 columns
+  __m128i s[16], x[16], u[32], v[32];
+  const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
+  const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kZero = _mm_set1_epi16(0);
+
+  u[0] = _mm_unpacklo_epi16(in[15], in[0]);
+  u[1] = _mm_unpackhi_epi16(in[15], in[0]);
+  u[2] = _mm_unpacklo_epi16(in[13], in[2]);
+  u[3] = _mm_unpackhi_epi16(in[13], in[2]);
+  u[4] = _mm_unpacklo_epi16(in[11], in[4]);
+  u[5] = _mm_unpackhi_epi16(in[11], in[4]);
+  u[6] = _mm_unpacklo_epi16(in[9], in[6]);
+  u[7] = _mm_unpackhi_epi16(in[9], in[6]);
+  u[8] = _mm_unpacklo_epi16(in[7], in[8]);
+  u[9] = _mm_unpackhi_epi16(in[7], in[8]);
+  u[10] = _mm_unpacklo_epi16(in[5], in[10]);
+  u[11] = _mm_unpackhi_epi16(in[5], in[10]);
+  u[12] = _mm_unpacklo_epi16(in[3], in[12]);
+  u[13] = _mm_unpackhi_epi16(in[3], in[12]);
+  u[14] = _mm_unpacklo_epi16(in[1], in[14]);
+  u[15] = _mm_unpackhi_epi16(in[1], in[14]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
+  v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
+  v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
+  v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
+  v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
+  v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
+  v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
+  v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
+  v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
+  v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
+  v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
+  v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
+  v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
+  v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
+  v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
+  v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
+  v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
+
+  u[0] = _mm_add_epi32(v[0], v[16]);
+  u[1] = _mm_add_epi32(v[1], v[17]);
+  u[2] = _mm_add_epi32(v[2], v[18]);
+  u[3] = _mm_add_epi32(v[3], v[19]);
+  u[4] = _mm_add_epi32(v[4], v[20]);
+  u[5] = _mm_add_epi32(v[5], v[21]);
+  u[6] = _mm_add_epi32(v[6], v[22]);
+  u[7] = _mm_add_epi32(v[7], v[23]);
+  u[8] = _mm_add_epi32(v[8], v[24]);
+  u[9] = _mm_add_epi32(v[9], v[25]);
+  u[10] = _mm_add_epi32(v[10], v[26]);
+  u[11] = _mm_add_epi32(v[11], v[27]);
+  u[12] = _mm_add_epi32(v[12], v[28]);
+  u[13] = _mm_add_epi32(v[13], v[29]);
+  u[14] = _mm_add_epi32(v[14], v[30]);
+  u[15] = _mm_add_epi32(v[15], v[31]);
+  u[16] = _mm_sub_epi32(v[0], v[16]);
+  u[17] = _mm_sub_epi32(v[1], v[17]);
+  u[18] = _mm_sub_epi32(v[2], v[18]);
+  u[19] = _mm_sub_epi32(v[3], v[19]);
+  u[20] = _mm_sub_epi32(v[4], v[20]);
+  u[21] = _mm_sub_epi32(v[5], v[21]);
+  u[22] = _mm_sub_epi32(v[6], v[22]);
+  u[23] = _mm_sub_epi32(v[7], v[23]);
+  u[24] = _mm_sub_epi32(v[8], v[24]);
+  u[25] = _mm_sub_epi32(v[9], v[25]);
+  u[26] = _mm_sub_epi32(v[10], v[26]);
+  u[27] = _mm_sub_epi32(v[11], v[27]);
+  u[28] = _mm_sub_epi32(v[12], v[28]);
+  u[29] = _mm_sub_epi32(v[13], v[29]);
+  u[30] = _mm_sub_epi32(v[14], v[30]);
+  u[31] = _mm_sub_epi32(v[15], v[31]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+  v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
+  v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
+  v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
+  v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
+  v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
+  v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
+  v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
+  v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
+  v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
+  v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
+  v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
+  v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
+  v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
+  v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
+  v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
+  v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+  u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
+  u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
+  u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
+  u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
+  u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
+  u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
+  u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
+  u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
+  u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
+  u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
+  u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
+  u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
+  u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
+  u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
+  u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
+  u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
+
+  s[0] = _mm_packs_epi32(u[0], u[1]);
+  s[1] = _mm_packs_epi32(u[2], u[3]);
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_packs_epi32(u[8], u[9]);
+  s[5] = _mm_packs_epi32(u[10], u[11]);
+  s[6] = _mm_packs_epi32(u[12], u[13]);
+  s[7] = _mm_packs_epi32(u[14], u[15]);
+  s[8] = _mm_packs_epi32(u[16], u[17]);
+  s[9] = _mm_packs_epi32(u[18], u[19]);
+  s[10] = _mm_packs_epi32(u[20], u[21]);
+  s[11] = _mm_packs_epi32(u[22], u[23]);
+  s[12] = _mm_packs_epi32(u[24], u[25]);
+  s[13] = _mm_packs_epi32(u[26], u[27]);
+  s[14] = _mm_packs_epi32(u[28], u[29]);
+  s[15] = _mm_packs_epi32(u[30], u[31]);
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[9]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[9]);
+  u[2] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[3] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[4] = _mm_unpacklo_epi16(s[12], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[12], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], v[8]);
+  u[1] = _mm_add_epi32(v[1], v[9]);
+  u[2] = _mm_add_epi32(v[2], v[10]);
+  u[3] = _mm_add_epi32(v[3], v[11]);
+  u[4] = _mm_add_epi32(v[4], v[12]);
+  u[5] = _mm_add_epi32(v[5], v[13]);
+  u[6] = _mm_add_epi32(v[6], v[14]);
+  u[7] = _mm_add_epi32(v[7], v[15]);
+  u[8] = _mm_sub_epi32(v[0], v[8]);
+  u[9] = _mm_sub_epi32(v[1], v[9]);
+  u[10] = _mm_sub_epi32(v[2], v[10]);
+  u[11] = _mm_sub_epi32(v[3], v[11]);
+  u[12] = _mm_sub_epi32(v[4], v[12]);
+  u[13] = _mm_sub_epi32(v[5], v[13]);
+  u[14] = _mm_sub_epi32(v[6], v[14]);
+  u[15] = _mm_sub_epi32(v[7], v[15]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+  x[0] = _mm_add_epi16(s[0], s[4]);
+  x[1] = _mm_add_epi16(s[1], s[5]);
+  x[2] = _mm_add_epi16(s[2], s[6]);
+  x[3] = _mm_add_epi16(s[3], s[7]);
+  x[4] = _mm_sub_epi16(s[0], s[4]);
+  x[5] = _mm_sub_epi16(s[1], s[5]);
+  x[6] = _mm_sub_epi16(s[2], s[6]);
+  x[7] = _mm_sub_epi16(s[3], s[7]);
+  x[8] = _mm_packs_epi32(u[0], u[1]);
+  x[9] = _mm_packs_epi32(u[2], u[3]);
+  x[10] = _mm_packs_epi32(u[4], u[5]);
+  x[11] = _mm_packs_epi32(u[6], u[7]);
+  x[12] = _mm_packs_epi32(u[8], u[9]);
+  x[13] = _mm_packs_epi32(u[10], u[11]);
+  x[14] = _mm_packs_epi32(u[12], u[13]);
+  x[15] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  u[0] = _mm_unpacklo_epi16(x[4], x[5]);
+  u[1] = _mm_unpackhi_epi16(x[4], x[5]);
+  u[2] = _mm_unpacklo_epi16(x[6], x[7]);
+  u[3] = _mm_unpackhi_epi16(x[6], x[7]);
+  u[4] = _mm_unpacklo_epi16(x[12], x[13]);
+  u[5] = _mm_unpackhi_epi16(x[12], x[13]);
+  u[6] = _mm_unpacklo_epi16(x[14], x[15]);
+  u[7] = _mm_unpackhi_epi16(x[14], x[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], v[4]);
+  u[1] = _mm_add_epi32(v[1], v[5]);
+  u[2] = _mm_add_epi32(v[2], v[6]);
+  u[3] = _mm_add_epi32(v[3], v[7]);
+  u[4] = _mm_sub_epi32(v[0], v[4]);
+  u[5] = _mm_sub_epi32(v[1], v[5]);
+  u[6] = _mm_sub_epi32(v[2], v[6]);
+  u[7] = _mm_sub_epi32(v[3], v[7]);
+  u[8] = _mm_add_epi32(v[8], v[12]);
+  u[9] = _mm_add_epi32(v[9], v[13]);
+  u[10] = _mm_add_epi32(v[10], v[14]);
+  u[11] = _mm_add_epi32(v[11], v[15]);
+  u[12] = _mm_sub_epi32(v[8], v[12]);
+  u[13] = _mm_sub_epi32(v[9], v[13]);
+  u[14] = _mm_sub_epi32(v[10], v[14]);
+  u[15] = _mm_sub_epi32(v[11], v[15]);
+
+  u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[0] = _mm_add_epi16(x[0], x[2]);
+  s[1] = _mm_add_epi16(x[1], x[3]);
+  s[2] = _mm_sub_epi16(x[0], x[2]);
+  s[3] = _mm_sub_epi16(x[1], x[3]);
+  s[4] = _mm_packs_epi32(v[0], v[1]);
+  s[5] = _mm_packs_epi32(v[2], v[3]);
+  s[6] = _mm_packs_epi32(v[4], v[5]);
+  s[7] = _mm_packs_epi32(v[6], v[7]);
+  s[8] = _mm_add_epi16(x[8], x[10]);
+  s[9] = _mm_add_epi16(x[9], x[11]);
+  s[10] = _mm_sub_epi16(x[8], x[10]);
+  s[11] = _mm_sub_epi16(x[9], x[11]);
+  s[12] = _mm_packs_epi32(v[8], v[9]);
+  s[13] = _mm_packs_epi32(v[10], v[11]);
+  s[14] = _mm_packs_epi32(v[12], v[13]);
+  s[15] = _mm_packs_epi32(v[14], v[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(s[2], s[3]);
+  u[1] = _mm_unpackhi_epi16(s[2], s[3]);
+  u[2] = _mm_unpacklo_epi16(s[6], s[7]);
+  u[3] = _mm_unpackhi_epi16(s[6], s[7]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  in[0] = s[0];
+  in[1] = _mm_sub_epi16(kZero, s[8]);
+  in[2] = s[12];
+  in[3] = _mm_sub_epi16(kZero, s[4]);
+  in[4] = _mm_packs_epi32(v[4], v[5]);
+  in[5] = _mm_packs_epi32(v[12], v[13]);
+  in[6] = _mm_packs_epi32(v[8], v[9]);
+  in[7] = _mm_packs_epi32(v[0], v[1]);
+  in[8] = _mm_packs_epi32(v[2], v[3]);
+  in[9] = _mm_packs_epi32(v[10], v[11]);
+  in[10] = _mm_packs_epi32(v[14], v[15]);
+  in[11] = _mm_packs_epi32(v[6], v[7]);
+  in[12] = s[5];
+  in[13] = _mm_sub_epi16(kZero, s[13]);
+  in[14] = s[9];
+  in[15] = _mm_sub_epi16(kZero, s[1]);
+}
+
+static void idct16_8col(__m128i *in) {
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i v[16], u[16], s[16], t[16];
+
+  // stage 1
+  s[0] = in[0];
+  s[1] = in[8];
+  s[2] = in[4];
+  s[3] = in[12];
+  s[4] = in[2];
+  s[5] = in[10];
+  s[6] = in[6];
+  s[7] = in[14];
+  s[8] = in[1];
+  s[9] = in[9];
+  s[10] = in[5];
+  s[11] = in[13];
+  s[12] = in[3];
+  s[13] = in[11];
+  s[14] = in[7];
+  s[15] = in[15];
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[15]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[15]);
+  u[2] = _mm_unpacklo_epi16(s[9], s[14]);
+  u[3] = _mm_unpackhi_epi16(s[9], s[14]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[11], s[12]);
+  u[7] = _mm_unpackhi_epi16(s[11], s[12]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p30_m02);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p30_m02);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p02_p30);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p02_p30);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p14_m18);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p14_m18);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p18_p14);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p18_p14);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p22_m10);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p22_m10);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p10_p22);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p10_p22);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p06_m26);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p06_m26);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p26_p06);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p26_p06);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[8]  = _mm_packs_epi32(u[0], u[1]);
+  s[15] = _mm_packs_epi32(u[2], u[3]);
+  s[9]  = _mm_packs_epi32(u[4], u[5]);
+  s[14] = _mm_packs_epi32(u[6], u[7]);
+  s[10] = _mm_packs_epi32(u[8], u[9]);
+  s[13] = _mm_packs_epi32(u[10], u[11]);
+  s[11] = _mm_packs_epi32(u[12], u[13]);
+  s[12] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  t[0] = s[0];
+  t[1] = s[1];
+  t[2] = s[2];
+  t[3] = s[3];
+  u[0] = _mm_unpacklo_epi16(s[4], s[7]);
+  u[1] = _mm_unpackhi_epi16(s[4], s[7]);
+  u[2] = _mm_unpacklo_epi16(s[5], s[6]);
+  u[3] = _mm_unpackhi_epi16(s[5], s[6]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[4] = _mm_packs_epi32(u[0], u[1]);
+  t[7] = _mm_packs_epi32(u[2], u[3]);
+  t[5] = _mm_packs_epi32(u[4], u[5]);
+  t[6] = _mm_packs_epi32(u[6], u[7]);
+  t[8] = _mm_add_epi16(s[8], s[9]);
+  t[9] = _mm_sub_epi16(s[8], s[9]);
+  t[10] = _mm_sub_epi16(s[11], s[10]);
+  t[11] = _mm_add_epi16(s[10], s[11]);
+  t[12] = _mm_add_epi16(s[12], s[13]);
+  t[13] = _mm_sub_epi16(s[12], s[13]);
+  t[14] = _mm_sub_epi16(s[15], s[14]);
+  t[15] = _mm_add_epi16(s[14], s[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(t[0], t[1]);
+  u[1] = _mm_unpackhi_epi16(t[0], t[1]);
+  u[2] = _mm_unpacklo_epi16(t[2], t[3]);
+  u[3] = _mm_unpackhi_epi16(t[2], t[3]);
+  u[4] = _mm_unpacklo_epi16(t[9], t[14]);
+  u[5] = _mm_unpackhi_epi16(t[9], t[14]);
+  u[6] = _mm_unpacklo_epi16(t[10], t[13]);
+  u[7] = _mm_unpackhi_epi16(t[10], t[13]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p24_m08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p24_m08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_p08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_p08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_m08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_m08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_m08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_m08_p24);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[0] = _mm_packs_epi32(u[0], u[1]);
+  s[1] = _mm_packs_epi32(u[2], u[3]);
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_add_epi16(t[4], t[5]);
+  s[5] = _mm_sub_epi16(t[4], t[5]);
+  s[6] = _mm_sub_epi16(t[7], t[6]);
+  s[7] = _mm_add_epi16(t[6], t[7]);
+  s[8] = t[8];
+  s[15] = t[15];
+  s[9]  = _mm_packs_epi32(u[8], u[9]);
+  s[14] = _mm_packs_epi32(u[10], u[11]);
+  s[10] = _mm_packs_epi32(u[12], u[13]);
+  s[13] = _mm_packs_epi32(u[14], u[15]);
+  s[11] = t[11];
+  s[12] = t[12];
+
+  // stage 5
+  t[0] = _mm_add_epi16(s[0], s[3]);
+  t[1] = _mm_add_epi16(s[1], s[2]);
+  t[2] = _mm_sub_epi16(s[1], s[2]);
+  t[3] = _mm_sub_epi16(s[0], s[3]);
+  t[4] = s[4];
+  t[7] = s[7];
+
+  u[0] = _mm_unpacklo_epi16(s[5], s[6]);
+  u[1] = _mm_unpackhi_epi16(s[5], s[6]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  t[5] = _mm_packs_epi32(u[0], u[1]);
+  t[6] = _mm_packs_epi32(u[2], u[3]);
+
+  t[8] = _mm_add_epi16(s[8], s[11]);
+  t[9] = _mm_add_epi16(s[9], s[10]);
+  t[10] = _mm_sub_epi16(s[9], s[10]);
+  t[11] = _mm_sub_epi16(s[8], s[11]);
+  t[12] = _mm_sub_epi16(s[15], s[12]);
+  t[13] = _mm_sub_epi16(s[14], s[13]);
+  t[14] = _mm_add_epi16(s[13], s[14]);
+  t[15] = _mm_add_epi16(s[12], s[15]);
+
+  // stage 6
+  s[0] = _mm_add_epi16(t[0], t[7]);
+  s[1] = _mm_add_epi16(t[1], t[6]);
+  s[2] = _mm_add_epi16(t[2], t[5]);
+  s[3] = _mm_add_epi16(t[3], t[4]);
+  s[4] = _mm_sub_epi16(t[3], t[4]);
+  s[5] = _mm_sub_epi16(t[2], t[5]);
+  s[6] = _mm_sub_epi16(t[1], t[6]);
+  s[7] = _mm_sub_epi16(t[0], t[7]);
+  s[8] = t[8];
+  s[9] = t[9];
+
+  u[0] = _mm_unpacklo_epi16(t[10], t[13]);
+  u[1] = _mm_unpackhi_epi16(t[10], t[13]);
+  u[2] = _mm_unpacklo_epi16(t[11], t[12]);
+  u[3] = _mm_unpackhi_epi16(t[11], t[12]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  s[10] = _mm_packs_epi32(u[0], u[1]);
+  s[13] = _mm_packs_epi32(u[2], u[3]);
+  s[11] = _mm_packs_epi32(u[4], u[5]);
+  s[12] = _mm_packs_epi32(u[6], u[7]);
+  s[14] = t[14];
+  s[15] = t[15];
+
+  // stage 7
+  in[0] = _mm_add_epi16(s[0], s[15]);
+  in[1] = _mm_add_epi16(s[1], s[14]);
+  in[2] = _mm_add_epi16(s[2], s[13]);
+  in[3] = _mm_add_epi16(s[3], s[12]);
+  in[4] = _mm_add_epi16(s[4], s[11]);
+  in[5] = _mm_add_epi16(s[5], s[10]);
+  in[6] = _mm_add_epi16(s[6], s[9]);
+  in[7] = _mm_add_epi16(s[7], s[8]);
+  in[8] = _mm_sub_epi16(s[7], s[8]);
+  in[9] = _mm_sub_epi16(s[6], s[9]);
+  in[10] = _mm_sub_epi16(s[5], s[10]);
+  in[11] = _mm_sub_epi16(s[4], s[11]);
+  in[12] = _mm_sub_epi16(s[3], s[12]);
+  in[13] = _mm_sub_epi16(s[2], s[13]);
+  in[14] = _mm_sub_epi16(s[1], s[14]);
+  in[15] = _mm_sub_epi16(s[0], s[15]);
+}
+
+void idct16_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  idct16_8col(in0);
+  idct16_8col(in1);
+}
+
+void iadst16_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  iadst16_8col(in0);
+  iadst16_8col(in1);
+}
+
+void vpx_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest,
+                               int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  __m128i in[16], l[16];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_8_0, stp1_12_0;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+  // First 1-D inverse DCT
+  // Load input data.
+  in[0] = load_input_data(input);
+  in[1] = load_input_data(input + 8 * 2);
+  in[2] = load_input_data(input + 8 * 4);
+  in[3] = load_input_data(input + 8 * 6);
+
+  TRANSPOSE_8X4(in[0], in[1], in[2], in[3], in[0], in[1]);
+
+  // Stage2
+  {
+    const __m128i lo_1_15 = _mm_unpackhi_epi16(in[0], zero);
+    const __m128i lo_13_3 = _mm_unpackhi_epi16(zero, in[1]);
+
+    tmp0 = _mm_madd_epi16(lo_1_15, stg2_0);
+    tmp2 = _mm_madd_epi16(lo_1_15, stg2_1);
+    tmp5 = _mm_madd_epi16(lo_13_3, stg2_6);
+    tmp7 = _mm_madd_epi16(lo_13_3, stg2_7);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp5 = _mm_add_epi32(tmp5, rounding);
+    tmp7 = _mm_add_epi32(tmp7, rounding);
+
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
+    tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
+
+    stp2_8  = _mm_packs_epi32(tmp0, tmp2);
+    stp2_11 = _mm_packs_epi32(tmp5, tmp7);
+  }
+
+  // Stage3
+  {
+    const __m128i lo_2_14 = _mm_unpacklo_epi16(in[1], zero);
+
+    tmp0 = _mm_madd_epi16(lo_2_14, stg3_0);
+    tmp2 = _mm_madd_epi16(lo_2_14, stg3_1);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+
+    stp1_13 = _mm_unpackhi_epi64(stp2_11, zero);
+    stp1_14 = _mm_unpackhi_epi64(stp2_8, zero);
+
+    stp1_4 = _mm_packs_epi32(tmp0, tmp2);
+  }
+
+  // Stage4
+  {
+    const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero);
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp1_14);
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp1_13);
+
+    tmp0 = _mm_madd_epi16(lo_0_8, stg4_0);
+    tmp2 = _mm_madd_epi16(lo_0_8, stg4_1);
+    tmp1 = _mm_madd_epi16(lo_9_14, stg4_4);
+    tmp3 = _mm_madd_epi16(lo_9_14, stg4_5);
+    tmp5 = _mm_madd_epi16(lo_10_13, stg4_6);
+    tmp7 = _mm_madd_epi16(lo_10_13, stg4_7);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp1 = _mm_add_epi32(tmp1, rounding);
+    tmp3 = _mm_add_epi32(tmp3, rounding);
+    tmp5 = _mm_add_epi32(tmp5, rounding);
+    tmp7 = _mm_add_epi32(tmp7, rounding);
+
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+    tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
+    tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
+
+    stp1_0 = _mm_packs_epi32(tmp0, tmp0);
+    stp1_1 = _mm_packs_epi32(tmp2, tmp2);
+    stp2_9 = _mm_packs_epi32(tmp1, tmp3);
+    stp2_10 = _mm_packs_epi32(tmp5, tmp7);
+
+    stp2_6 = _mm_unpackhi_epi64(stp1_4, zero);
+  }
+
+  // Stage5 and Stage6
+  {
+    tmp0 = _mm_add_epi16(stp2_8, stp2_11);
+    tmp1 = _mm_sub_epi16(stp2_8, stp2_11);
+    tmp2 = _mm_add_epi16(stp2_9, stp2_10);
+    tmp3 = _mm_sub_epi16(stp2_9, stp2_10);
+
+    stp1_9  = _mm_unpacklo_epi64(tmp2, zero);
+    stp1_10 = _mm_unpacklo_epi64(tmp3, zero);
+    stp1_8  = _mm_unpacklo_epi64(tmp0, zero);
+    stp1_11 = _mm_unpacklo_epi64(tmp1, zero);
+
+    stp1_13 = _mm_unpackhi_epi64(tmp3, zero);
+    stp1_14 = _mm_unpackhi_epi64(tmp2, zero);
+    stp1_12 = _mm_unpackhi_epi64(tmp1, zero);
+    stp1_15 = _mm_unpackhi_epi64(tmp0, zero);
+  }
+
+  // Stage6
+  {
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp1_4);
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12);
+
+    tmp1 = _mm_madd_epi16(lo_6_5, stg4_1);
+    tmp3 = _mm_madd_epi16(lo_6_5, stg4_0);
+    tmp0 = _mm_madd_epi16(lo_10_13, stg6_0);
+    tmp2 = _mm_madd_epi16(lo_10_13, stg4_0);
+    tmp4 = _mm_madd_epi16(lo_11_12, stg6_0);
+    tmp6 = _mm_madd_epi16(lo_11_12, stg4_0);
+
+    tmp1 = _mm_add_epi32(tmp1, rounding);
+    tmp3 = _mm_add_epi32(tmp3, rounding);
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp1_6 = _mm_packs_epi32(tmp3, tmp1);
+
+    stp2_10 = _mm_packs_epi32(tmp0, zero);
+    stp2_13 = _mm_packs_epi32(tmp2, zero);
+    stp2_11 = _mm_packs_epi32(tmp4, zero);
+    stp2_12 = _mm_packs_epi32(tmp6, zero);
+
+    tmp0 = _mm_add_epi16(stp1_0, stp1_4);
+    tmp1 = _mm_sub_epi16(stp1_0, stp1_4);
+    tmp2 = _mm_add_epi16(stp1_1, stp1_6);
+    tmp3 = _mm_sub_epi16(stp1_1, stp1_6);
+
+    stp2_0 = _mm_unpackhi_epi64(tmp0, zero);
+    stp2_1 = _mm_unpacklo_epi64(tmp2, zero);
+    stp2_2 = _mm_unpackhi_epi64(tmp2, zero);
+    stp2_3 = _mm_unpacklo_epi64(tmp0, zero);
+    stp2_4 = _mm_unpacklo_epi64(tmp1, zero);
+    stp2_5 = _mm_unpackhi_epi64(tmp3, zero);
+    stp2_6 = _mm_unpacklo_epi64(tmp3, zero);
+    stp2_7 = _mm_unpackhi_epi64(tmp1, zero);
+  }
+
+  // Stage7. Left 8x16 only.
+  l[0] = _mm_add_epi16(stp2_0, stp1_15);
+  l[1] = _mm_add_epi16(stp2_1, stp1_14);
+  l[2] = _mm_add_epi16(stp2_2, stp2_13);
+  l[3] = _mm_add_epi16(stp2_3, stp2_12);
+  l[4] = _mm_add_epi16(stp2_4, stp2_11);
+  l[5] = _mm_add_epi16(stp2_5, stp2_10);
+  l[6] = _mm_add_epi16(stp2_6, stp1_9);
+  l[7] = _mm_add_epi16(stp2_7, stp1_8);
+  l[8] = _mm_sub_epi16(stp2_7, stp1_8);
+  l[9] = _mm_sub_epi16(stp2_6, stp1_9);
+  l[10] = _mm_sub_epi16(stp2_5, stp2_10);
+  l[11] = _mm_sub_epi16(stp2_4, stp2_11);
+  l[12] = _mm_sub_epi16(stp2_3, stp2_12);
+  l[13] = _mm_sub_epi16(stp2_2, stp2_13);
+  l[14] = _mm_sub_epi16(stp2_1, stp1_14);
+  l[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+  // Second 1-D inverse transform, performed per 8x16 block
+  for (i = 0; i < 2; i++) {
+    int j;
+    array_transpose_4X8(l + 8 * i, in);
+
+    IDCT16_10
+
+    // Stage7
+    in[0] = _mm_add_epi16(stp2_0, stp1_15);
+    in[1] = _mm_add_epi16(stp2_1, stp1_14);
+    in[2] = _mm_add_epi16(stp2_2, stp2_13);
+    in[3] = _mm_add_epi16(stp2_3, stp2_12);
+    in[4] = _mm_add_epi16(stp2_4, stp2_11);
+    in[5] = _mm_add_epi16(stp2_5, stp2_10);
+    in[6] = _mm_add_epi16(stp2_6, stp1_9);
+    in[7] = _mm_add_epi16(stp2_7, stp1_8);
+    in[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    in[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    in[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    in[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    in[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    in[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    in[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    in[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    for (j = 0; j < 16; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+#define LOAD_DQCOEFF(reg, input) \
+  {  \
+    reg = load_input_data(input); \
+    input += 8; \
+  }  \
+
+#define IDCT32_34 \
+/* Stage1 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], zero); \
+  const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], zero); \
+  \
+  const __m128i lo_25_7= _mm_unpacklo_epi16(zero, in[7]); \
+  const __m128i hi_25_7 = _mm_unpackhi_epi16(zero, in[7]); \
+  \
+  const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], zero); \
+  const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], zero); \
+  \
+  const __m128i lo_29_3 = _mm_unpacklo_epi16(zero, in[3]); \
+  const __m128i hi_29_3 = _mm_unpackhi_epi16(zero, in[3]); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_1_31, hi_1_31, stg1_0, \
+                         stg1_1, stp1_16, stp1_31); \
+  MULTIPLICATION_AND_ADD_2(lo_25_7, hi_25_7, stg1_6, \
+                         stg1_7, stp1_19, stp1_28); \
+  MULTIPLICATION_AND_ADD_2(lo_5_27, hi_5_27, stg1_8, \
+                         stg1_9, stp1_20, stp1_27); \
+  MULTIPLICATION_AND_ADD_2(lo_29_3, hi_29_3, stg1_14, \
+                         stg1_15, stp1_23, stp1_24); \
+} \
+\
+/* Stage2 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], zero); \
+  const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], zero); \
+  \
+  const __m128i lo_26_6 = _mm_unpacklo_epi16(zero, in[6]); \
+  const __m128i hi_26_6 = _mm_unpackhi_epi16(zero, in[6]); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_2_30, hi_2_30, stg2_0, \
+                         stg2_1, stp2_8, stp2_15); \
+  MULTIPLICATION_AND_ADD_2(lo_26_6, hi_26_6, stg2_6, \
+                         stg2_7, stp2_11, stp2_12); \
+  \
+  stp2_16 = stp1_16; \
+  stp2_19 = stp1_19; \
+  \
+  stp2_20 = stp1_20; \
+  stp2_23 = stp1_23; \
+  \
+  stp2_24 = stp1_24; \
+  stp2_27 = stp1_27; \
+  \
+  stp2_28 = stp1_28; \
+  stp2_31 = stp1_31; \
+} \
+\
+/* Stage3 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], zero); \
+  const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], zero); \
+  \
+  const __m128i lo_17_30 = _mm_unpacklo_epi16(stp1_16, stp1_31); \
+  const __m128i hi_17_30 = _mm_unpackhi_epi16(stp1_16, stp1_31); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp1_19, stp1_28); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp1_19, stp1_28); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp1_20, stp1_27); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp1_20, stp1_27); \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp1_23, stp1_24); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp1_23, stp2_24); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_4_28, hi_4_28, stg3_0, \
+                         stg3_1, stp1_4, stp1_7); \
+  \
+  stp1_8 = stp2_8; \
+  stp1_11 = stp2_11; \
+  stp1_12 = stp2_12; \
+  stp1_15 = stp2_15; \
+  \
+  MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \
+                         stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, \
+                         stp1_18, stp1_29) \
+  MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \
+                         stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, \
+                         stp1_22, stp1_25) \
+  \
+  stp1_16 = stp2_16; \
+  stp1_31 = stp2_31; \
+  stp1_19 = stp2_19; \
+  stp1_20 = stp2_20; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_27 = stp2_27; \
+  stp1_28 = stp2_28; \
+} \
+\
+/* Stage4 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], zero); \
+  const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], zero); \
+  \
+  const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp2_15); \
+  const __m128i hi_9_14 = _mm_unpackhi_epi16(stp2_8, stp2_15); \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp2_12); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp2_11, stp2_12); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_0_16, hi_0_16, stg4_0, \
+                         stg4_1, stp2_0, stp2_1); \
+  \
+  stp2_4 = stp1_4; \
+  stp2_5 = stp1_4; \
+  stp2_6 = stp1_7; \
+  stp2_7 = stp1_7; \
+  \
+  MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
+                         stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, \
+                         stp2_10, stp2_13) \
+  \
+  stp2_8 = stp1_8; \
+  stp2_15 = stp1_15; \
+  stp2_11 = stp1_11; \
+  stp2_12 = stp1_12; \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \
+  stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \
+  stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \
+  stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \
+  stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \
+  stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \
+  \
+  stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \
+  stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \
+  stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \
+  stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \
+  stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \
+} \
+\
+/* Stage5 */ \
+{ \
+  const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+  const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
+  \
+  const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \
+  const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  stp1_0 = stp2_0; \
+  stp1_1 = stp2_1; \
+  stp1_2 = stp2_1; \
+  stp1_3 = stp2_0; \
+  \
+  tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+  tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+  tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+  tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+  \
+  tmp0 = _mm_add_epi32(tmp0, rounding); \
+  tmp1 = _mm_add_epi32(tmp1, rounding); \
+  tmp2 = _mm_add_epi32(tmp2, rounding); \
+  tmp3 = _mm_add_epi32(tmp3, rounding); \
+  \
+  tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+  tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+  tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+  tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+  \
+  stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+  stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+  \
+  stp1_4 = stp2_4; \
+  stp1_7 = stp2_7; \
+  \
+  stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \
+  stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
+  stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
+  stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
+  stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  \
+  MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \
+                         stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, \
+                         stp1_19, stp1_28) \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \
+                         stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  \
+  stp1_22 = stp2_22; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_25 = stp2_25; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+} \
+\
+/* Stage6 */ \
+{ \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+  const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+  const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+  \
+  stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \
+  stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+  stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+  stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \
+  stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \
+  stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+  stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+  stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \
+  \
+  stp2_8 = stp1_8; \
+  stp2_9 = stp1_9; \
+  stp2_14 = stp1_14; \
+  stp2_15 = stp1_15; \
+  \
+  MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                         stg6_0, stg4_0, stg6_0, stg4_0, stp2_10, \
+                         stp2_13, stp2_11, stp2_12) \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
+  stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \
+  stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \
+  stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \
+  stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \
+  stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \
+  \
+  stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \
+  stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \
+  stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \
+  stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \
+  stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \
+} \
+\
+/* Stage7 */ \
+{ \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
+  const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \
+  const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \
+  \
+  stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \
+  stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \
+  stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \
+  stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \
+  stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \
+  stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \
+  stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \
+  stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \
+  stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \
+  stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \
+  stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \
+  stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \
+  stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  stp1_18 = stp2_18; \
+  stp1_19 = stp2_19; \
+  \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, \
+                         stp1_23, stp1_24) \
+  \
+  stp1_28 = stp2_28; \
+  stp1_29 = stp2_29; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+}
+
+
+#define IDCT32 \
+/* Stage1 */ \
+{ \
+  const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], in[31]); \
+  const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], in[31]); \
+  const __m128i lo_17_15 = _mm_unpacklo_epi16(in[17], in[15]); \
+  const __m128i hi_17_15 = _mm_unpackhi_epi16(in[17], in[15]); \
+  \
+  const __m128i lo_9_23 = _mm_unpacklo_epi16(in[9], in[23]); \
+  const __m128i hi_9_23 = _mm_unpackhi_epi16(in[9], in[23]); \
+  const __m128i lo_25_7= _mm_unpacklo_epi16(in[25], in[7]); \
+  const __m128i hi_25_7 = _mm_unpackhi_epi16(in[25], in[7]); \
+  \
+  const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], in[27]); \
+  const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], in[27]); \
+  const __m128i lo_21_11 = _mm_unpacklo_epi16(in[21], in[11]); \
+  const __m128i hi_21_11 = _mm_unpackhi_epi16(in[21], in[11]); \
+  \
+  const __m128i lo_13_19 = _mm_unpacklo_epi16(in[13], in[19]); \
+  const __m128i hi_13_19 = _mm_unpackhi_epi16(in[13], in[19]); \
+  const __m128i lo_29_3 = _mm_unpacklo_epi16(in[29], in[3]); \
+  const __m128i hi_29_3 = _mm_unpackhi_epi16(in[29], in[3]); \
+  \
+  MULTIPLICATION_AND_ADD(lo_1_31, hi_1_31, lo_17_15, hi_17_15, stg1_0, \
+                         stg1_1, stg1_2, stg1_3, stp1_16, stp1_31, \
+                         stp1_17, stp1_30) \
+  MULTIPLICATION_AND_ADD(lo_9_23, hi_9_23, lo_25_7, hi_25_7, stg1_4, \
+                         stg1_5, stg1_6, stg1_7, stp1_18, stp1_29, \
+                         stp1_19, stp1_28) \
+  MULTIPLICATION_AND_ADD(lo_5_27, hi_5_27, lo_21_11, hi_21_11, stg1_8, \
+                         stg1_9, stg1_10, stg1_11, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  MULTIPLICATION_AND_ADD(lo_13_19, hi_13_19, lo_29_3, hi_29_3, stg1_12, \
+                         stg1_13, stg1_14, stg1_15, stp1_22, stp1_25, \
+                         stp1_23, stp1_24) \
+} \
+\
+/* Stage2 */ \
+{ \
+  const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], in[30]); \
+  const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], in[30]); \
+  const __m128i lo_18_14 = _mm_unpacklo_epi16(in[18], in[14]); \
+  const __m128i hi_18_14 = _mm_unpackhi_epi16(in[18], in[14]); \
+  \
+  const __m128i lo_10_22 = _mm_unpacklo_epi16(in[10], in[22]); \
+  const __m128i hi_10_22 = _mm_unpackhi_epi16(in[10], in[22]); \
+  const __m128i lo_26_6 = _mm_unpacklo_epi16(in[26], in[6]); \
+  const __m128i hi_26_6 = _mm_unpackhi_epi16(in[26], in[6]); \
+  \
+  MULTIPLICATION_AND_ADD(lo_2_30, hi_2_30, lo_18_14, hi_18_14, stg2_0, \
+                         stg2_1, stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, \
+                         stp2_14) \
+  MULTIPLICATION_AND_ADD(lo_10_22, hi_10_22, lo_26_6, hi_26_6, stg2_4, \
+                         stg2_5, stg2_6, stg2_7, stp2_10, stp2_13, \
+                         stp2_11, stp2_12) \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_17); \
+  stp2_17 = _mm_sub_epi16(stp1_16, stp1_17); \
+  stp2_18 = _mm_sub_epi16(stp1_19, stp1_18); \
+  stp2_19 = _mm_add_epi16(stp1_19, stp1_18); \
+  \
+  stp2_20 = _mm_add_epi16(stp1_20, stp1_21); \
+  stp2_21 = _mm_sub_epi16(stp1_20, stp1_21); \
+  stp2_22 = _mm_sub_epi16(stp1_23, stp1_22); \
+  stp2_23 = _mm_add_epi16(stp1_23, stp1_22); \
+  \
+  stp2_24 = _mm_add_epi16(stp1_24, stp1_25); \
+  stp2_25 = _mm_sub_epi16(stp1_24, stp1_25); \
+  stp2_26 = _mm_sub_epi16(stp1_27, stp1_26); \
+  stp2_27 = _mm_add_epi16(stp1_27, stp1_26); \
+  \
+  stp2_28 = _mm_add_epi16(stp1_28, stp1_29); \
+  stp2_29 = _mm_sub_epi16(stp1_28, stp1_29); \
+  stp2_30 = _mm_sub_epi16(stp1_31, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_31, stp1_30); \
+} \
+\
+/* Stage3 */ \
+{ \
+  const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], in[28]); \
+  const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], in[28]); \
+  const __m128i lo_20_12 = _mm_unpacklo_epi16(in[20], in[12]); \
+  const __m128i hi_20_12 = _mm_unpackhi_epi16(in[20], in[12]); \
+  \
+  const __m128i lo_17_30 = _mm_unpacklo_epi16(stp2_17, stp2_30); \
+  const __m128i hi_17_30 = _mm_unpackhi_epi16(stp2_17, stp2_30); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
+  \
+  MULTIPLICATION_AND_ADD(lo_4_28, hi_4_28, lo_20_12, hi_20_12, stg3_0, \
+                         stg3_1, stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, \
+                         stp1_6) \
+  \
+  stp1_8 = _mm_add_epi16(stp2_8, stp2_9); \
+  stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \
+  stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \
+  stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \
+  stp1_12 = _mm_add_epi16(stp2_12, stp2_13); \
+  stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \
+  stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \
+  stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \
+  \
+  MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \
+                         stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, \
+                         stp1_18, stp1_29) \
+  MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \
+                         stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, \
+                         stp1_22, stp1_25) \
+  \
+  stp1_16 = stp2_16; \
+  stp1_31 = stp2_31; \
+  stp1_19 = stp2_19; \
+  stp1_20 = stp2_20; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_27 = stp2_27; \
+  stp1_28 = stp2_28; \
+} \
+\
+/* Stage4 */ \
+{ \
+  const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], in[16]); \
+  const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], in[16]); \
+  const __m128i lo_8_24 = _mm_unpacklo_epi16(in[8], in[24]); \
+  const __m128i hi_8_24 = _mm_unpackhi_epi16(in[8], in[24]); \
+  \
+  const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
+  const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+  \
+  MULTIPLICATION_AND_ADD(lo_0_16, hi_0_16, lo_8_24, hi_8_24, stg4_0, \
+                         stg4_1, stg4_2, stg4_3, stp2_0, stp2_1, \
+                         stp2_2, stp2_3) \
+  \
+  stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
+  stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
+  stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
+  stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
+  \
+  MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
+                         stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, \
+                         stp2_10, stp2_13) \
+  \
+  stp2_8 = stp1_8; \
+  stp2_15 = stp1_15; \
+  stp2_11 = stp1_11; \
+  stp2_12 = stp1_12; \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \
+  stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \
+  stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \
+  stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \
+  stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \
+  stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \
+  \
+  stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \
+  stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \
+  stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \
+  stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \
+  stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \
+} \
+\
+/* Stage5 */ \
+{ \
+  const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+  const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
+  \
+  const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \
+  const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
+  stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
+  stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
+  stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
+  \
+  tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+  tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+  tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+  tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+  \
+  tmp0 = _mm_add_epi32(tmp0, rounding); \
+  tmp1 = _mm_add_epi32(tmp1, rounding); \
+  tmp2 = _mm_add_epi32(tmp2, rounding); \
+  tmp3 = _mm_add_epi32(tmp3, rounding); \
+  \
+  tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+  tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+  tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+  tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+  \
+  stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+  stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+  \
+  stp1_4 = stp2_4; \
+  stp1_7 = stp2_7; \
+  \
+  stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \
+  stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
+  stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
+  stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
+  stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  \
+  MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \
+                         stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, \
+                         stp1_19, stp1_28) \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \
+                         stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  \
+  stp1_22 = stp2_22; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_25 = stp2_25; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+} \
+\
+/* Stage6 */ \
+{ \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+  const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+  const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+  \
+  stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \
+  stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+  stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+  stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \
+  stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \
+  stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+  stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+  stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \
+  \
+  stp2_8 = stp1_8; \
+  stp2_9 = stp1_9; \
+  stp2_14 = stp1_14; \
+  stp2_15 = stp1_15; \
+  \
+  MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                         stg6_0, stg4_0, stg6_0, stg4_0, stp2_10, \
+                         stp2_13, stp2_11, stp2_12) \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
+  stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \
+  stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \
+  stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \
+  stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \
+  stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \
+  \
+  stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \
+  stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \
+  stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \
+  stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \
+  stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \
+} \
+\
+/* Stage7 */ \
+{ \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
+  const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \
+  const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \
+  \
+  stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \
+  stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \
+  stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \
+  stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \
+  stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \
+  stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \
+  stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \
+  stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \
+  stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \
+  stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \
+  stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \
+  stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \
+  stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  stp1_18 = stp2_18; \
+  stp1_19 = stp2_19; \
+  \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, \
+                         stp1_23, stp1_24) \
+  \
+  stp1_28 = stp2_28; \
+  stp1_29 = stp2_29; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+}
+
+// Only upper-left 8x8 has non-zero coeff
+void vpx_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest,
+                               int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1<<5);
+
+  // idct constants for each stage
+  const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[32], col[32];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22,
+          stp1_23, stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29,
+          stp1_30, stp1_31;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
+          stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22,
+          stp2_23, stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29,
+          stp2_30, stp2_31;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  // Load input data. Only need to load the top left 8x8 block.
+  in[0] = load_input_data(input);
+  in[1] = load_input_data(input + 32);
+  in[2] = load_input_data(input + 64);
+  in[3] = load_input_data(input + 96);
+  in[4] = load_input_data(input + 128);
+  in[5] = load_input_data(input + 160);
+  in[6] = load_input_data(input + 192);
+  in[7] = load_input_data(input + 224);
+
+  for (i = 8; i < 32; ++i) {
+    in[i] = _mm_setzero_si128();
+  }
+
+  array_transpose_8x8(in, in);
+  // TODO(hkuang): Following transposes are unnecessary. But remove them will
+  // lead to performance drop on some devices.
+  array_transpose_8x8(in + 8, in + 8);
+  array_transpose_8x8(in + 16, in + 16);
+  array_transpose_8x8(in + 24, in + 24);
+
+  IDCT32_34
+
+  // 1_D: Store 32 intermediate results for each 8x32 block.
+  col[0] = _mm_add_epi16(stp1_0, stp1_31);
+  col[1] = _mm_add_epi16(stp1_1, stp1_30);
+  col[2] = _mm_add_epi16(stp1_2, stp1_29);
+  col[3] = _mm_add_epi16(stp1_3, stp1_28);
+  col[4] = _mm_add_epi16(stp1_4, stp1_27);
+  col[5] = _mm_add_epi16(stp1_5, stp1_26);
+  col[6] = _mm_add_epi16(stp1_6, stp1_25);
+  col[7] = _mm_add_epi16(stp1_7, stp1_24);
+  col[8] = _mm_add_epi16(stp1_8, stp1_23);
+  col[9] = _mm_add_epi16(stp1_9, stp1_22);
+  col[10] = _mm_add_epi16(stp1_10, stp1_21);
+  col[11] = _mm_add_epi16(stp1_11, stp1_20);
+  col[12] = _mm_add_epi16(stp1_12, stp1_19);
+  col[13] = _mm_add_epi16(stp1_13, stp1_18);
+  col[14] = _mm_add_epi16(stp1_14, stp1_17);
+  col[15] = _mm_add_epi16(stp1_15, stp1_16);
+  col[16] = _mm_sub_epi16(stp1_15, stp1_16);
+  col[17] = _mm_sub_epi16(stp1_14, stp1_17);
+  col[18] = _mm_sub_epi16(stp1_13, stp1_18);
+  col[19] = _mm_sub_epi16(stp1_12, stp1_19);
+  col[20] = _mm_sub_epi16(stp1_11, stp1_20);
+  col[21] = _mm_sub_epi16(stp1_10, stp1_21);
+  col[22] = _mm_sub_epi16(stp1_9, stp1_22);
+  col[23] = _mm_sub_epi16(stp1_8, stp1_23);
+  col[24] = _mm_sub_epi16(stp1_7, stp1_24);
+  col[25] = _mm_sub_epi16(stp1_6, stp1_25);
+  col[26] = _mm_sub_epi16(stp1_5, stp1_26);
+  col[27] = _mm_sub_epi16(stp1_4, stp1_27);
+  col[28] = _mm_sub_epi16(stp1_3, stp1_28);
+  col[29] = _mm_sub_epi16(stp1_2, stp1_29);
+  col[30] = _mm_sub_epi16(stp1_1, stp1_30);
+  col[31] = _mm_sub_epi16(stp1_0, stp1_31);
+  for (i = 0; i < 4; i++) {
+    int j;
+    const __m128i zero = _mm_setzero_si128();
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(col + i * 8, in);
+    IDCT32_34
+
+    // 2_D: Calculate the results and store them to destination.
+    in[0] = _mm_add_epi16(stp1_0, stp1_31);
+    in[1] = _mm_add_epi16(stp1_1, stp1_30);
+    in[2] = _mm_add_epi16(stp1_2, stp1_29);
+    in[3] = _mm_add_epi16(stp1_3, stp1_28);
+    in[4] = _mm_add_epi16(stp1_4, stp1_27);
+    in[5] = _mm_add_epi16(stp1_5, stp1_26);
+    in[6] = _mm_add_epi16(stp1_6, stp1_25);
+    in[7] = _mm_add_epi16(stp1_7, stp1_24);
+    in[8] = _mm_add_epi16(stp1_8, stp1_23);
+    in[9] = _mm_add_epi16(stp1_9, stp1_22);
+    in[10] = _mm_add_epi16(stp1_10, stp1_21);
+    in[11] = _mm_add_epi16(stp1_11, stp1_20);
+    in[12] = _mm_add_epi16(stp1_12, stp1_19);
+    in[13] = _mm_add_epi16(stp1_13, stp1_18);
+    in[14] = _mm_add_epi16(stp1_14, stp1_17);
+    in[15] = _mm_add_epi16(stp1_15, stp1_16);
+    in[16] = _mm_sub_epi16(stp1_15, stp1_16);
+    in[17] = _mm_sub_epi16(stp1_14, stp1_17);
+    in[18] = _mm_sub_epi16(stp1_13, stp1_18);
+    in[19] = _mm_sub_epi16(stp1_12, stp1_19);
+    in[20] = _mm_sub_epi16(stp1_11, stp1_20);
+    in[21] = _mm_sub_epi16(stp1_10, stp1_21);
+    in[22] = _mm_sub_epi16(stp1_9, stp1_22);
+    in[23] = _mm_sub_epi16(stp1_8, stp1_23);
+    in[24] = _mm_sub_epi16(stp1_7, stp1_24);
+    in[25] = _mm_sub_epi16(stp1_6, stp1_25);
+    in[26] = _mm_sub_epi16(stp1_5, stp1_26);
+    in[27] = _mm_sub_epi16(stp1_4, stp1_27);
+    in[28] = _mm_sub_epi16(stp1_3, stp1_28);
+    in[29] = _mm_sub_epi16(stp1_2, stp1_29);
+    in[30] = _mm_sub_epi16(stp1_1, stp1_30);
+    in[31] = _mm_sub_epi16(stp1_0, stp1_31);
+
+    for (j = 0; j < 32; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void vpx_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest,
+                                 int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  // idct constants for each stage
+  const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[32], col[128], zero_idx[16];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22,
+          stp1_23, stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29,
+          stp1_30, stp1_31;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
+          stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22,
+          stp2_23, stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29,
+          stp2_30, stp2_31;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i, j, i32;
+
+  for (i = 0; i < 4; i++) {
+    i32 = (i << 5);
+    // First 1-D idct
+    // Load input data.
+    LOAD_DQCOEFF(in[0], input);
+    LOAD_DQCOEFF(in[8], input);
+    LOAD_DQCOEFF(in[16], input);
+    LOAD_DQCOEFF(in[24], input);
+    LOAD_DQCOEFF(in[1], input);
+    LOAD_DQCOEFF(in[9], input);
+    LOAD_DQCOEFF(in[17], input);
+    LOAD_DQCOEFF(in[25], input);
+    LOAD_DQCOEFF(in[2], input);
+    LOAD_DQCOEFF(in[10], input);
+    LOAD_DQCOEFF(in[18], input);
+    LOAD_DQCOEFF(in[26], input);
+    LOAD_DQCOEFF(in[3], input);
+    LOAD_DQCOEFF(in[11], input);
+    LOAD_DQCOEFF(in[19], input);
+    LOAD_DQCOEFF(in[27], input);
+
+    LOAD_DQCOEFF(in[4], input);
+    LOAD_DQCOEFF(in[12], input);
+    LOAD_DQCOEFF(in[20], input);
+    LOAD_DQCOEFF(in[28], input);
+    LOAD_DQCOEFF(in[5], input);
+    LOAD_DQCOEFF(in[13], input);
+    LOAD_DQCOEFF(in[21], input);
+    LOAD_DQCOEFF(in[29], input);
+    LOAD_DQCOEFF(in[6], input);
+    LOAD_DQCOEFF(in[14], input);
+    LOAD_DQCOEFF(in[22], input);
+    LOAD_DQCOEFF(in[30], input);
+    LOAD_DQCOEFF(in[7], input);
+    LOAD_DQCOEFF(in[15], input);
+    LOAD_DQCOEFF(in[23], input);
+    LOAD_DQCOEFF(in[31], input);
+
+    // checking if all entries are zero
+    zero_idx[0] = _mm_or_si128(in[0], in[1]);
+    zero_idx[1] = _mm_or_si128(in[2], in[3]);
+    zero_idx[2] = _mm_or_si128(in[4], in[5]);
+    zero_idx[3] = _mm_or_si128(in[6], in[7]);
+    zero_idx[4] = _mm_or_si128(in[8], in[9]);
+    zero_idx[5] = _mm_or_si128(in[10], in[11]);
+    zero_idx[6] = _mm_or_si128(in[12], in[13]);
+    zero_idx[7] = _mm_or_si128(in[14], in[15]);
+    zero_idx[8] = _mm_or_si128(in[16], in[17]);
+    zero_idx[9] = _mm_or_si128(in[18], in[19]);
+    zero_idx[10] = _mm_or_si128(in[20], in[21]);
+    zero_idx[11] = _mm_or_si128(in[22], in[23]);
+    zero_idx[12] = _mm_or_si128(in[24], in[25]);
+    zero_idx[13] = _mm_or_si128(in[26], in[27]);
+    zero_idx[14] = _mm_or_si128(in[28], in[29]);
+    zero_idx[15] = _mm_or_si128(in[30], in[31]);
+
+    zero_idx[0] = _mm_or_si128(zero_idx[0], zero_idx[1]);
+    zero_idx[1] = _mm_or_si128(zero_idx[2], zero_idx[3]);
+    zero_idx[2] = _mm_or_si128(zero_idx[4], zero_idx[5]);
+    zero_idx[3] = _mm_or_si128(zero_idx[6], zero_idx[7]);
+    zero_idx[4] = _mm_or_si128(zero_idx[8], zero_idx[9]);
+    zero_idx[5] = _mm_or_si128(zero_idx[10], zero_idx[11]);
+    zero_idx[6] = _mm_or_si128(zero_idx[12], zero_idx[13]);
+    zero_idx[7] = _mm_or_si128(zero_idx[14], zero_idx[15]);
+
+    zero_idx[8] = _mm_or_si128(zero_idx[0], zero_idx[1]);
+    zero_idx[9] = _mm_or_si128(zero_idx[2], zero_idx[3]);
+    zero_idx[10] = _mm_or_si128(zero_idx[4], zero_idx[5]);
+    zero_idx[11] = _mm_or_si128(zero_idx[6], zero_idx[7]);
+    zero_idx[12] = _mm_or_si128(zero_idx[8], zero_idx[9]);
+    zero_idx[13] = _mm_or_si128(zero_idx[10], zero_idx[11]);
+    zero_idx[14] = _mm_or_si128(zero_idx[12], zero_idx[13]);
+
+    if (_mm_movemask_epi8(_mm_cmpeq_epi32(zero_idx[14], zero)) == 0xFFFF) {
+      col[i32 + 0] = _mm_setzero_si128();
+      col[i32 + 1] = _mm_setzero_si128();
+      col[i32 + 2] = _mm_setzero_si128();
+      col[i32 + 3] = _mm_setzero_si128();
+      col[i32 + 4] = _mm_setzero_si128();
+      col[i32 + 5] = _mm_setzero_si128();
+      col[i32 + 6] = _mm_setzero_si128();
+      col[i32 + 7] = _mm_setzero_si128();
+      col[i32 + 8] = _mm_setzero_si128();
+      col[i32 + 9] = _mm_setzero_si128();
+      col[i32 + 10] = _mm_setzero_si128();
+      col[i32 + 11] = _mm_setzero_si128();
+      col[i32 + 12] = _mm_setzero_si128();
+      col[i32 + 13] = _mm_setzero_si128();
+      col[i32 + 14] = _mm_setzero_si128();
+      col[i32 + 15] = _mm_setzero_si128();
+      col[i32 + 16] = _mm_setzero_si128();
+      col[i32 + 17] = _mm_setzero_si128();
+      col[i32 + 18] = _mm_setzero_si128();
+      col[i32 + 19] = _mm_setzero_si128();
+      col[i32 + 20] = _mm_setzero_si128();
+      col[i32 + 21] = _mm_setzero_si128();
+      col[i32 + 22] = _mm_setzero_si128();
+      col[i32 + 23] = _mm_setzero_si128();
+      col[i32 + 24] = _mm_setzero_si128();
+      col[i32 + 25] = _mm_setzero_si128();
+      col[i32 + 26] = _mm_setzero_si128();
+      col[i32 + 27] = _mm_setzero_si128();
+      col[i32 + 28] = _mm_setzero_si128();
+      col[i32 + 29] = _mm_setzero_si128();
+      col[i32 + 30] = _mm_setzero_si128();
+      col[i32 + 31] = _mm_setzero_si128();
+      continue;
+    }
+
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(in, in);
+    array_transpose_8x8(in + 8, in + 8);
+    array_transpose_8x8(in + 16, in + 16);
+    array_transpose_8x8(in + 24, in + 24);
+
+    IDCT32
+
+    // 1_D: Store 32 intermediate results for each 8x32 block.
+    col[i32 + 0] = _mm_add_epi16(stp1_0, stp1_31);
+    col[i32 + 1] = _mm_add_epi16(stp1_1, stp1_30);
+    col[i32 + 2] = _mm_add_epi16(stp1_2, stp1_29);
+    col[i32 + 3] = _mm_add_epi16(stp1_3, stp1_28);
+    col[i32 + 4] = _mm_add_epi16(stp1_4, stp1_27);
+    col[i32 + 5] = _mm_add_epi16(stp1_5, stp1_26);
+    col[i32 + 6] = _mm_add_epi16(stp1_6, stp1_25);
+    col[i32 + 7] = _mm_add_epi16(stp1_7, stp1_24);
+    col[i32 + 8] = _mm_add_epi16(stp1_8, stp1_23);
+    col[i32 + 9] = _mm_add_epi16(stp1_9, stp1_22);
+    col[i32 + 10] = _mm_add_epi16(stp1_10, stp1_21);
+    col[i32 + 11] = _mm_add_epi16(stp1_11, stp1_20);
+    col[i32 + 12] = _mm_add_epi16(stp1_12, stp1_19);
+    col[i32 + 13] = _mm_add_epi16(stp1_13, stp1_18);
+    col[i32 + 14] = _mm_add_epi16(stp1_14, stp1_17);
+    col[i32 + 15] = _mm_add_epi16(stp1_15, stp1_16);
+    col[i32 + 16] = _mm_sub_epi16(stp1_15, stp1_16);
+    col[i32 + 17] = _mm_sub_epi16(stp1_14, stp1_17);
+    col[i32 + 18] = _mm_sub_epi16(stp1_13, stp1_18);
+    col[i32 + 19] = _mm_sub_epi16(stp1_12, stp1_19);
+    col[i32 + 20] = _mm_sub_epi16(stp1_11, stp1_20);
+    col[i32 + 21] = _mm_sub_epi16(stp1_10, stp1_21);
+    col[i32 + 22] = _mm_sub_epi16(stp1_9, stp1_22);
+    col[i32 + 23] = _mm_sub_epi16(stp1_8, stp1_23);
+    col[i32 + 24] = _mm_sub_epi16(stp1_7, stp1_24);
+    col[i32 + 25] = _mm_sub_epi16(stp1_6, stp1_25);
+    col[i32 + 26] = _mm_sub_epi16(stp1_5, stp1_26);
+    col[i32 + 27] = _mm_sub_epi16(stp1_4, stp1_27);
+    col[i32 + 28] = _mm_sub_epi16(stp1_3, stp1_28);
+    col[i32 + 29] = _mm_sub_epi16(stp1_2, stp1_29);
+    col[i32 + 30] = _mm_sub_epi16(stp1_1, stp1_30);
+    col[i32 + 31] = _mm_sub_epi16(stp1_0, stp1_31);
+  }
+  for (i = 0; i < 4; i++) {
+    // Second 1-D idct
+    j = i << 3;
+
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(col + j, in);
+    array_transpose_8x8(col + j + 32, in + 8);
+    array_transpose_8x8(col + j + 64, in + 16);
+    array_transpose_8x8(col + j + 96, in + 24);
+
+    IDCT32
+
+    // 2_D: Calculate the results and store them to destination.
+    in[0] = _mm_add_epi16(stp1_0, stp1_31);
+    in[1] = _mm_add_epi16(stp1_1, stp1_30);
+    in[2] = _mm_add_epi16(stp1_2, stp1_29);
+    in[3] = _mm_add_epi16(stp1_3, stp1_28);
+    in[4] = _mm_add_epi16(stp1_4, stp1_27);
+    in[5] = _mm_add_epi16(stp1_5, stp1_26);
+    in[6] = _mm_add_epi16(stp1_6, stp1_25);
+    in[7] = _mm_add_epi16(stp1_7, stp1_24);
+    in[8] = _mm_add_epi16(stp1_8, stp1_23);
+    in[9] = _mm_add_epi16(stp1_9, stp1_22);
+    in[10] = _mm_add_epi16(stp1_10, stp1_21);
+    in[11] = _mm_add_epi16(stp1_11, stp1_20);
+    in[12] = _mm_add_epi16(stp1_12, stp1_19);
+    in[13] = _mm_add_epi16(stp1_13, stp1_18);
+    in[14] = _mm_add_epi16(stp1_14, stp1_17);
+    in[15] = _mm_add_epi16(stp1_15, stp1_16);
+    in[16] = _mm_sub_epi16(stp1_15, stp1_16);
+    in[17] = _mm_sub_epi16(stp1_14, stp1_17);
+    in[18] = _mm_sub_epi16(stp1_13, stp1_18);
+    in[19] = _mm_sub_epi16(stp1_12, stp1_19);
+    in[20] = _mm_sub_epi16(stp1_11, stp1_20);
+    in[21] = _mm_sub_epi16(stp1_10, stp1_21);
+    in[22] = _mm_sub_epi16(stp1_9, stp1_22);
+    in[23] = _mm_sub_epi16(stp1_8, stp1_23);
+    in[24] = _mm_sub_epi16(stp1_7, stp1_24);
+    in[25] = _mm_sub_epi16(stp1_6, stp1_25);
+    in[26] = _mm_sub_epi16(stp1_5, stp1_26);
+    in[27] = _mm_sub_epi16(stp1_4, stp1_27);
+    in[28] = _mm_sub_epi16(stp1_3, stp1_28);
+    in[29] = _mm_sub_epi16(stp1_2, stp1_29);
+    in[30] = _mm_sub_epi16(stp1_1, stp1_30);
+    in[31] = _mm_sub_epi16(stp1_0, stp1_31);
+
+    for (j = 0; j < 32; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void vpx_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest,
+                              int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a, j;
+
+  a = (int)dct_const_round_shift(input[0] * cospi_16_64);
+  a = (int)dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 6);
+
+  dc_value = _mm_set1_epi16(a);
+
+  for (j = 0; j < 32; ++j) {
+    RECON_AND_STORE(dest +  0 + j * stride, dc_value);
+    RECON_AND_STORE(dest +  8 + j * stride, dc_value);
+    RECON_AND_STORE(dest + 16 + j * stride, dc_value);
+    RECON_AND_STORE(dest + 24 + j * stride, dc_value);
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE __m128i clamp_high_sse2(__m128i value, int bd) {
+  __m128i ubounded, retval;
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i max = _mm_subs_epi16(_mm_slli_epi16(one, bd), one);
+  ubounded = _mm_cmpgt_epi16(value, max);
+  retval = _mm_andnot_si128(ubounded, value);
+  ubounded = _mm_and_si128(ubounded, max);
+  retval = _mm_or_si128(retval, ubounded);
+  retval = _mm_and_si128(retval, _mm_cmpgt_epi16(retval, zero));
+  return retval;
+}
+
+void vpx_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  __m128i inptr[4];
+  __m128i sign_bits[2];
+  __m128i temp_mm, min_input, max_input;
+  int test;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  int optimised_cols = 0;
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i eight = _mm_set1_epi16(8);
+  const __m128i max = _mm_set1_epi16(12043);
+  const __m128i min = _mm_set1_epi16(-12043);
+  // Load input into __m128i
+  inptr[0] = _mm_loadu_si128((const __m128i *)input);
+  inptr[1] = _mm_loadu_si128((const __m128i *)(input + 4));
+  inptr[2] = _mm_loadu_si128((const __m128i *)(input + 8));
+  inptr[3] = _mm_loadu_si128((const __m128i *)(input + 12));
+
+  // Pack to 16 bits
+  inptr[0] = _mm_packs_epi32(inptr[0], inptr[1]);
+  inptr[1] = _mm_packs_epi32(inptr[2], inptr[3]);
+
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp_mm = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp_mm);
+
+  if (!test) {
+    // Do the row transform
+    idct4_sse2(inptr);
+
+    // Check the min & max values
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp_mm = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp_mm);
+
+    if (test) {
+      transpose_4x4(inptr);
+      sign_bits[0] = _mm_cmplt_epi16(inptr[0], zero);
+      sign_bits[1] = _mm_cmplt_epi16(inptr[1], zero);
+      inptr[3] = _mm_unpackhi_epi16(inptr[1], sign_bits[1]);
+      inptr[2] = _mm_unpacklo_epi16(inptr[1], sign_bits[1]);
+      inptr[1] = _mm_unpackhi_epi16(inptr[0], sign_bits[0]);
+      inptr[0] = _mm_unpacklo_epi16(inptr[0], sign_bits[0]);
+      _mm_storeu_si128((__m128i *)outptr, inptr[0]);
+      _mm_storeu_si128((__m128i *)(outptr + 4), inptr[1]);
+      _mm_storeu_si128((__m128i *)(outptr + 8), inptr[2]);
+      _mm_storeu_si128((__m128i *)(outptr + 12), inptr[3]);
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 4; ++i) {
+      vpx_highbd_idct4_c(input, outptr, bd);
+      input += 4;
+      outptr += 4;
+    }
+  }
+
+  if (optimised_cols) {
+    idct4_sse2(inptr);
+
+    // Final round and shift
+    inptr[0] = _mm_add_epi16(inptr[0], eight);
+    inptr[1] = _mm_add_epi16(inptr[1], eight);
+
+    inptr[0] = _mm_srai_epi16(inptr[0], 4);
+    inptr[1] = _mm_srai_epi16(inptr[1], 4);
+
+    // Reconstruction and Store
+    {
+      __m128i d0 = _mm_loadl_epi64((const __m128i *)dest);
+      __m128i d2 = _mm_loadl_epi64((const __m128i *)(dest + stride * 2));
+      d0 = _mm_unpacklo_epi64(
+          d0, _mm_loadl_epi64((const __m128i *)(dest + stride)));
+      d2 = _mm_unpacklo_epi64(
+          d2, _mm_loadl_epi64((const __m128i *)(dest + stride * 3)));
+      d0 = clamp_high_sse2(_mm_adds_epi16(d0, inptr[0]), bd);
+      d2 = clamp_high_sse2(_mm_adds_epi16(d2, inptr[1]), bd);
+      // store input0
+      _mm_storel_epi64((__m128i *)dest, d0);
+      // store input1
+      d0 = _mm_srli_si128(d0, 8);
+      _mm_storel_epi64((__m128i *)(dest + stride), d0);
+      // store input2
+      _mm_storel_epi64((__m128i *)(dest + stride * 2), d2);
+      // store input3
+      d2 = _mm_srli_si128(d2, 8);
+      _mm_storel_epi64((__m128i *)(dest + stride * 3), d2);
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[4], temp_out[4];
+    // Columns
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j)
+        temp_in[j] = out[j * 4 + i];
+      vpx_highbd_idct4_c(temp_in, temp_out, bd);
+      for (j = 0; j < 4; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
+      }
+    }
+  }
+}
+
+void vpx_highbd_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[8];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  const __m128i max = _mm_set1_epi16(6201);
+  const __m128i min = _mm_set1_epi16(-6201);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 8; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 8; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform
+    idct8_sse2(inptr);
+
+    // Find the min & max for the column transform
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 8; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      array_transpose_8x8(inptr, inptr);
+      for (i = 0; i < 8; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 8; ++i) {
+      vpx_highbd_idct8_c(input, outptr, bd);
+      input += 8;
+      outptr += 8;
+    }
+  }
+
+  if (optimised_cols) {
+    idct8_sse2(inptr);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[8];
+      for (i = 0; i < 8; i++) {
+        inptr[i] = _mm_add_epi16(inptr[i], sixteen);
+        d[i] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        inptr[i] = _mm_srai_epi16(inptr[i], 5);
+        d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[i]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[8], temp_out[8];
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j)
+        temp_in[j] = out[j * 8 + i];
+      vpx_highbd_idct8_c(temp_in, temp_out, bd);
+      for (j = 0; j < 8; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+      }
+    }
+  }
+}
+
+void vpx_highbd_idct8x8_10_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[8 * 8] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[8];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  const __m128i max = _mm_set1_epi16(6201);
+  const __m128i min = _mm_set1_epi16(-6201);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 8; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  // only first 4 row has non-zero coefs
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 4; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform
+    idct8_sse2(inptr);
+
+    // Find the min & max for the column transform
+    // N.B. Only first 4 cols contain non-zero coeffs
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 8; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      // Use fact only first 4 rows contain non-zero coeffs
+      array_transpose_4X8(inptr, inptr);
+      for (i = 0; i < 4; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 4; ++i) {
+      vpx_highbd_idct8_c(input, outptr, bd);
+      input += 8;
+      outptr += 8;
+    }
+  }
+
+  if (optimised_cols) {
+    idct8_sse2(inptr);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[8];
+      for (i = 0; i < 8; i++) {
+        inptr[i] = _mm_add_epi16(inptr[i], sixteen);
+        d[i] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        inptr[i] = _mm_srai_epi16(inptr[i], 5);
+        d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[i]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[8], temp_out[8];
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j)
+        temp_in[j] = out[j * 8 + i];
+      vpx_highbd_idct8_c(temp_in, temp_out, bd);
+      for (j = 0; j < 8; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+      }
+    }
+  }
+}
+
+void vpx_highbd_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                       int stride, int bd) {
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[32];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i rounding = _mm_set1_epi16(32);
+  const __m128i max = _mm_set1_epi16(3155);
+  const __m128i min = _mm_set1_epi16(-3155);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 16; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12));
+    inptr[i + 16] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 32; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform
+    idct16_sse2(inptr, inptr + 16);
+
+    // Find the min & max for the column transform
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 32; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      array_transpose_16x16(inptr, inptr + 16);
+      for (i = 0; i < 16; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2);
+        sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 16; ++i) {
+      vpx_highbd_idct16_c(input, outptr, bd);
+      input += 16;
+      outptr += 16;
+    }
+  }
+
+  if (optimised_cols) {
+    idct16_sse2(inptr, inptr + 16);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[2];
+      for (i = 0; i < 16; i++) {
+        inptr[i   ] = _mm_add_epi16(inptr[i   ], rounding);
+        inptr[i+16] = _mm_add_epi16(inptr[i+16], rounding);
+        d[0] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        d[1] = _mm_loadu_si128((const __m128i *)(dest + stride*i + 8));
+        inptr[i   ] = _mm_srai_epi16(inptr[i   ], 6);
+        inptr[i+16] = _mm_srai_epi16(inptr[i+16], 6);
+        d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i   ]), bd);
+        d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i+16]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[0]);
+        _mm_storeu_si128((__m128i *)(dest + stride*i + 8), d[1]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[16], temp_out[16];
+    for (i = 0; i < 16; ++i) {
+      for (j = 0; j < 16; ++j)
+        temp_in[j] = out[j * 16 + i];
+      vpx_highbd_idct16_c(temp_in, temp_out, bd);
+      for (j = 0; j < 16; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+      }
+    }
+  }
+}
+
+void vpx_highbd_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                      int stride, int bd) {
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[32];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i rounding = _mm_set1_epi16(32);
+  const __m128i max = _mm_set1_epi16(3155);
+  const __m128i min = _mm_set1_epi16(-3155);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 16; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12));
+    inptr[i + 16] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  // Since all non-zero dct coefficients are in upper-left 4x4 area,
+  // we only need to consider first 4 rows here.
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 4; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform (N.B. This transposes inptr)
+    idct16_sse2(inptr, inptr + 16);
+
+    // Find the min & max for the column transform
+    // N.B. Only first 4 cols contain non-zero coeffs
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 16; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      // Use fact only first 4 rows contain non-zero coeffs
+      array_transpose_8x8(inptr, inptr);
+      array_transpose_8x8(inptr + 8, inptr + 16);
+      for (i = 0; i < 4; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2);
+        sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 4; ++i) {
+      vpx_highbd_idct16_c(input, outptr, bd);
+      input += 16;
+      outptr += 16;
+    }
+  }
+
+  if (optimised_cols) {
+    idct16_sse2(inptr, inptr + 16);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[2];
+      for (i = 0; i < 16; i++) {
+        inptr[i   ] = _mm_add_epi16(inptr[i   ], rounding);
+        inptr[i+16] = _mm_add_epi16(inptr[i+16], rounding);
+        d[0] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        d[1] = _mm_loadu_si128((const __m128i *)(dest + stride*i + 8));
+        inptr[i   ] = _mm_srai_epi16(inptr[i   ], 6);
+        inptr[i+16] = _mm_srai_epi16(inptr[i+16], 6);
+        d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i   ]), bd);
+        d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i+16]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[0]);
+        _mm_storeu_si128((__m128i *)(dest + stride*i + 8), d[1]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[16], temp_out[16];
+    for (i = 0; i < 16; ++i) {
+      for (j = 0; j < 16; ++j)
+        temp_in[j] = out[j * 16 + i];
+      vpx_highbd_idct16_c(temp_in, temp_out, bd);
+      for (j = 0; j < 16; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+      }
+    }
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vpx_dsp/x86/inv_txfm_sse2.h b/libvpx/libvpx/vpx_dsp/x86/inv_txfm_sse2.h
new file mode 100644
index 0000000..bd520c1
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/inv_txfm_sse2.h
@@ -0,0 +1,196 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_X86_INV_TXFM_SSE2_H_
+#define VPX_DSP_X86_INV_TXFM_SSE2_H_
+
+#include <emmintrin.h>  // SSE2
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/inv_txfm.h"
+#include "vpx_dsp/x86/txfm_common_sse2.h"
+
+// perform 8x8 transpose
+static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+
+  res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
+  res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
+  res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
+  res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
+  res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
+  res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
+  res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
+  res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
+}
+
+#define TRANSPOSE_8X4(in0, in1, in2, in3, out0, out1) \
+  {                                                     \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+                                                        \
+    in0 = _mm_unpacklo_epi32(tr0_0, tr0_1);  /* i1 i0 */  \
+    in1 = _mm_unpackhi_epi32(tr0_0, tr0_1);  /* i3 i2 */  \
+  }
+
+static INLINE void array_transpose_4X8(__m128i *in, __m128i * out) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+
+  out[0] = _mm_unpacklo_epi64(tr1_0, tr1_4);
+  out[1] = _mm_unpackhi_epi64(tr1_0, tr1_4);
+  out[2] = _mm_unpacklo_epi64(tr1_2, tr1_6);
+  out[3] = _mm_unpackhi_epi64(tr1_2, tr1_6);
+}
+
+static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
+  __m128i tbuf[8];
+  array_transpose_8x8(res0, res0);
+  array_transpose_8x8(res1, tbuf);
+  array_transpose_8x8(res0 + 8, res1);
+  array_transpose_8x8(res1 + 8, res1 + 8);
+
+  res0[8] = tbuf[0];
+  res0[9] = tbuf[1];
+  res0[10] = tbuf[2];
+  res0[11] = tbuf[3];
+  res0[12] = tbuf[4];
+  res0[13] = tbuf[5];
+  res0[14] = tbuf[6];
+  res0[15] = tbuf[7];
+}
+
+// Function to allow 8 bit optimisations to be used when profile 0 is used with
+// highbitdepth enabled
+static INLINE __m128i load_input_data(const tran_low_t *data) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  return octa_set_epi16(data[0], data[1], data[2], data[3], data[4], data[5],
+      data[6], data[7]);
+#else
+  return _mm_load_si128((const __m128i *)data);
+#endif
+}
+
+static INLINE void load_buffer_8x16(const tran_low_t *input, __m128i *in) {
+  in[0]  = load_input_data(input + 0 * 16);
+  in[1]  = load_input_data(input + 1 * 16);
+  in[2]  = load_input_data(input + 2 * 16);
+  in[3]  = load_input_data(input + 3 * 16);
+  in[4]  = load_input_data(input + 4 * 16);
+  in[5]  = load_input_data(input + 5 * 16);
+  in[6]  = load_input_data(input + 6 * 16);
+  in[7]  = load_input_data(input + 7 * 16);
+
+  in[8]  = load_input_data(input + 8 * 16);
+  in[9]  = load_input_data(input + 9 * 16);
+  in[10]  = load_input_data(input + 10 * 16);
+  in[11]  = load_input_data(input + 11 * 16);
+  in[12]  = load_input_data(input + 12 * 16);
+  in[13]  = load_input_data(input + 13 * 16);
+  in[14]  = load_input_data(input + 14 * 16);
+  in[15]  = load_input_data(input + 15 * 16);
+}
+
+#define RECON_AND_STORE(dest, in_x) \
+  {                                                     \
+     __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \
+      d0 = _mm_unpacklo_epi8(d0, zero); \
+      d0 = _mm_add_epi16(in_x, d0); \
+      d0 = _mm_packus_epi16(d0, d0); \
+      _mm_storel_epi64((__m128i *)(dest), d0); \
+  }
+
+static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) {
+  const __m128i final_rounding = _mm_set1_epi16(1<<5);
+  const __m128i zero = _mm_setzero_si128();
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+  in[4] = _mm_adds_epi16(in[4], final_rounding);
+  in[5] = _mm_adds_epi16(in[5], final_rounding);
+  in[6] = _mm_adds_epi16(in[6], final_rounding);
+  in[7] = _mm_adds_epi16(in[7], final_rounding);
+  in[8] = _mm_adds_epi16(in[8], final_rounding);
+  in[9] = _mm_adds_epi16(in[9], final_rounding);
+  in[10] = _mm_adds_epi16(in[10], final_rounding);
+  in[11] = _mm_adds_epi16(in[11], final_rounding);
+  in[12] = _mm_adds_epi16(in[12], final_rounding);
+  in[13] = _mm_adds_epi16(in[13], final_rounding);
+  in[14] = _mm_adds_epi16(in[14], final_rounding);
+  in[15] = _mm_adds_epi16(in[15], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 6);
+  in[1] = _mm_srai_epi16(in[1], 6);
+  in[2] = _mm_srai_epi16(in[2], 6);
+  in[3] = _mm_srai_epi16(in[3], 6);
+  in[4] = _mm_srai_epi16(in[4], 6);
+  in[5] = _mm_srai_epi16(in[5], 6);
+  in[6] = _mm_srai_epi16(in[6], 6);
+  in[7] = _mm_srai_epi16(in[7], 6);
+  in[8] = _mm_srai_epi16(in[8], 6);
+  in[9] = _mm_srai_epi16(in[9], 6);
+  in[10] = _mm_srai_epi16(in[10], 6);
+  in[11] = _mm_srai_epi16(in[11], 6);
+  in[12] = _mm_srai_epi16(in[12], 6);
+  in[13] = _mm_srai_epi16(in[13], 6);
+  in[14] = _mm_srai_epi16(in[14], 6);
+  in[15] = _mm_srai_epi16(in[15], 6);
+
+  RECON_AND_STORE(dest +  0 * stride, in[0]);
+  RECON_AND_STORE(dest +  1 * stride, in[1]);
+  RECON_AND_STORE(dest +  2 * stride, in[2]);
+  RECON_AND_STORE(dest +  3 * stride, in[3]);
+  RECON_AND_STORE(dest +  4 * stride, in[4]);
+  RECON_AND_STORE(dest +  5 * stride, in[5]);
+  RECON_AND_STORE(dest +  6 * stride, in[6]);
+  RECON_AND_STORE(dest +  7 * stride, in[7]);
+  RECON_AND_STORE(dest +  8 * stride, in[8]);
+  RECON_AND_STORE(dest +  9 * stride, in[9]);
+  RECON_AND_STORE(dest + 10 * stride, in[10]);
+  RECON_AND_STORE(dest + 11 * stride, in[11]);
+  RECON_AND_STORE(dest + 12 * stride, in[12]);
+  RECON_AND_STORE(dest + 13 * stride, in[13]);
+  RECON_AND_STORE(dest + 14 * stride, in[14]);
+  RECON_AND_STORE(dest + 15 * stride, in[15]);
+}
+
+void idct4_sse2(__m128i *in);
+void idct8_sse2(__m128i *in);
+void idct16_sse2(__m128i *in0, __m128i *in1);
+void iadst4_sse2(__m128i *in);
+void iadst8_sse2(__m128i *in);
+void iadst16_sse2(__m128i *in0, __m128i *in1);
+
+#endif  // VPX_DSP_X86_INV_TXFM_SSE2_H_
diff --git a/libvpx/libvpx/vpx_dsp/x86/inv_txfm_ssse3_x86_64.asm b/libvpx/libvpx/vpx_dsp/x86/inv_txfm_ssse3_x86_64.asm
new file mode 100644
index 0000000..20baf82
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/inv_txfm_ssse3_x86_64.asm
@@ -0,0 +1,1793 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the inverse transformation. Part
+; of the functions are originally derived from the ffmpeg project.
+; Note that the current version applies to x86 64-bit only.
+
+SECTION_RODATA
+
+pw_11585x2: times 8 dw 23170
+
+pw_m2404x2:  times 8 dw  -2404*2
+pw_m4756x2:  times 8 dw  -4756*2
+pw_m5520x2:  times 8 dw  -5520*2
+pw_m8423x2:  times 8 dw  -8423*2
+pw_m9102x2:  times 8 dw  -9102*2
+pw_m10394x2: times 8 dw -10394*2
+pw_m11003x2: times 8 dw -11003*2
+
+pw_16364x2: times 8 dw 16364*2
+pw_16305x2: times 8 dw 16305*2
+pw_16207x2: times 8 dw 16207*2
+pw_16069x2: times 8 dw 16069*2
+pw_15893x2: times 8 dw 15893*2
+pw_15679x2: times 8 dw 15679*2
+pw_15426x2: times 8 dw 15426*2
+pw_15137x2: times 8 dw 15137*2
+pw_14811x2: times 8 dw 14811*2
+pw_14449x2: times 8 dw 14449*2
+pw_14053x2: times 8 dw 14053*2
+pw_13623x2: times 8 dw 13623*2
+pw_13160x2: times 8 dw 13160*2
+pw_12665x2: times 8 dw 12665*2
+pw_12140x2: times 8 dw 12140*2
+pw__9760x2: times 8 dw  9760*2
+pw__7723x2: times 8 dw  7723*2
+pw__7005x2: times 8 dw  7005*2
+pw__6270x2: times 8 dw  6270*2
+pw__3981x2: times 8 dw  3981*2
+pw__3196x2: times 8 dw  3196*2
+pw__1606x2: times 8 dw  1606*2
+pw___804x2: times 8 dw   804*2
+
+pd_8192:    times 4 dd 8192
+pw_32:      times 8 dw 32
+pw_16:      times 8 dw 16
+
+%macro TRANSFORM_COEFFS 2
+pw_%1_%2:   dw  %1,  %2,  %1,  %2,  %1,  %2,  %1,  %2
+pw_m%2_%1:  dw -%2,  %1, -%2,  %1, -%2,  %1, -%2,  %1
+pw_m%1_m%2: dw -%1, -%2, -%1, -%2, -%1, -%2, -%1, -%2
+%endmacro
+
+TRANSFORM_COEFFS    6270, 15137
+TRANSFORM_COEFFS    3196, 16069
+TRANSFORM_COEFFS   13623,  9102
+
+; constants for 32x32_34
+TRANSFORM_COEFFS      804, 16364
+TRANSFORM_COEFFS    15426,  5520
+TRANSFORM_COEFFS     3981, 15893
+TRANSFORM_COEFFS    16207,  2404
+TRANSFORM_COEFFS     1606, 16305
+TRANSFORM_COEFFS    15679,  4756
+TRANSFORM_COEFFS    11585, 11585
+
+; constants for 32x32_1024
+TRANSFORM_COEFFS    12140, 11003
+TRANSFORM_COEFFS     7005, 14811
+TRANSFORM_COEFFS    14053,  8423
+TRANSFORM_COEFFS     9760, 13160
+TRANSFORM_COEFFS    12665, 10394
+TRANSFORM_COEFFS     7723, 14449
+
+%macro PAIR_PP_COEFFS 2
+dpw_%1_%2:   dw  %1,  %1,  %1,  %1,  %2,  %2,  %2,  %2
+%endmacro
+
+%macro PAIR_MP_COEFFS 2
+dpw_m%1_%2:  dw -%1, -%1, -%1, -%1,  %2,  %2,  %2,  %2
+%endmacro
+
+%macro PAIR_MM_COEFFS 2
+dpw_m%1_m%2: dw -%1, -%1, -%1, -%1, -%2, -%2, -%2, -%2
+%endmacro
+
+PAIR_PP_COEFFS     30274, 12540
+PAIR_PP_COEFFS      6392, 32138
+PAIR_MP_COEFFS     18204, 27246
+
+PAIR_PP_COEFFS     12540, 12540
+PAIR_PP_COEFFS     30274, 30274
+PAIR_PP_COEFFS      6392,  6392
+PAIR_PP_COEFFS     32138, 32138
+PAIR_MM_COEFFS     18204, 18204
+PAIR_PP_COEFFS     27246, 27246
+
+SECTION .text
+
+%if ARCH_X86_64
+%macro SUM_SUB 3
+  psubw  m%3, m%1, m%2
+  paddw  m%1, m%2
+  SWAP    %2, %3
+%endmacro
+
+; butterfly operation
+%macro MUL_ADD_2X 6 ; dst1, dst2, src, round, coefs1, coefs2
+  pmaddwd            m%1, m%3, %5
+  pmaddwd            m%2, m%3, %6
+  paddd              m%1,  %4
+  paddd              m%2,  %4
+  psrad              m%1,  14
+  psrad              m%2,  14
+%endmacro
+
+%macro BUTTERFLY_4X 7 ; dst1, dst2, coef1, coef2, round, tmp1, tmp2
+  punpckhwd          m%6, m%2, m%1
+  MUL_ADD_2X         %7,  %6,  %6,  %5, [pw_m%4_%3], [pw_%3_%4]
+  punpcklwd          m%2, m%1
+  MUL_ADD_2X         %1,  %2,  %2,  %5, [pw_m%4_%3], [pw_%3_%4]
+  packssdw           m%1, m%7
+  packssdw           m%2, m%6
+%endmacro
+
+%macro BUTTERFLY_4Xmm 7 ; dst1, dst2, coef1, coef2, round, tmp1, tmp2
+  punpckhwd          m%6, m%2, m%1
+  MUL_ADD_2X         %7,  %6,  %6,  %5, [pw_m%4_%3], [pw_m%3_m%4]
+  punpcklwd          m%2, m%1
+  MUL_ADD_2X         %1,  %2,  %2,  %5, [pw_m%4_%3], [pw_m%3_m%4]
+  packssdw           m%1, m%7
+  packssdw           m%2, m%6
+%endmacro
+
+; matrix transpose
+%macro INTERLEAVE_2X 4
+  punpckh%1          m%4, m%2, m%3
+  punpckl%1          m%2, m%3
+  SWAP               %3,  %4
+%endmacro
+
+%macro TRANSPOSE8X8 9
+  INTERLEAVE_2X  wd, %1, %2, %9
+  INTERLEAVE_2X  wd, %3, %4, %9
+  INTERLEAVE_2X  wd, %5, %6, %9
+  INTERLEAVE_2X  wd, %7, %8, %9
+
+  INTERLEAVE_2X  dq, %1, %3, %9
+  INTERLEAVE_2X  dq, %2, %4, %9
+  INTERLEAVE_2X  dq, %5, %7, %9
+  INTERLEAVE_2X  dq, %6, %8, %9
+
+  INTERLEAVE_2X  qdq, %1, %5, %9
+  INTERLEAVE_2X  qdq, %3, %7, %9
+  INTERLEAVE_2X  qdq, %2, %6, %9
+  INTERLEAVE_2X  qdq, %4, %8, %9
+
+  SWAP  %2, %5
+  SWAP  %4, %7
+%endmacro
+
+%macro IDCT8_1D 0
+  SUM_SUB          0,    4,    9
+  BUTTERFLY_4X     2,    6,    6270, 15137,  m8,  9,  10
+  pmulhrsw        m0,  m12
+  pmulhrsw        m4,  m12
+  BUTTERFLY_4X     1,    7,    3196, 16069,  m8,  9,  10
+  BUTTERFLY_4X     5,    3,   13623,  9102,  m8,  9,  10
+
+  SUM_SUB          1,    5,    9
+  SUM_SUB          7,    3,    9
+  SUM_SUB          0,    6,    9
+  SUM_SUB          4,    2,    9
+  SUM_SUB          3,    5,    9
+  pmulhrsw        m3,  m12
+  pmulhrsw        m5,  m12
+
+  SUM_SUB          0,    7,    9
+  SUM_SUB          4,    3,    9
+  SUM_SUB          2,    5,    9
+  SUM_SUB          6,    1,    9
+
+  SWAP             3,    6
+  SWAP             1,    4
+%endmacro
+
+; This macro handles 8 pixels per line
+%macro ADD_STORE_8P_2X 5;  src1, src2, tmp1, tmp2, zero
+  paddw           m%1, m11
+  paddw           m%2, m11
+  psraw           m%1, 5
+  psraw           m%2, 5
+
+  movh            m%3, [outputq]
+  movh            m%4, [outputq + strideq]
+  punpcklbw       m%3, m%5
+  punpcklbw       m%4, m%5
+  paddw           m%3, m%1
+  paddw           m%4, m%2
+  packuswb        m%3, m%5
+  packuswb        m%4, m%5
+  movh               [outputq], m%3
+  movh     [outputq + strideq], m%4
+%endmacro
+
+INIT_XMM ssse3
+; full inverse 8x8 2D-DCT transform
+cglobal idct8x8_64_add, 3, 5, 13, input, output, stride
+  mova     m8, [pd_8192]
+  mova    m11, [pw_16]
+  mova    m12, [pw_11585x2]
+
+  lea      r3, [2 * strideq]
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova     m0, [inputq +   0]
+  packssdw m0, [inputq +  16]
+  mova     m1, [inputq +  32]
+  packssdw m1, [inputq +  48]
+  mova     m2, [inputq +  64]
+  packssdw m2, [inputq +  80]
+  mova     m3, [inputq +  96]
+  packssdw m3, [inputq + 112]
+  mova     m4, [inputq + 128]
+  packssdw m4, [inputq + 144]
+  mova     m5, [inputq + 160]
+  packssdw m5, [inputq + 176]
+  mova     m6, [inputq + 192]
+  packssdw m6, [inputq + 208]
+  mova     m7, [inputq + 224]
+  packssdw m7, [inputq + 240]
+%else
+  mova     m0, [inputq +   0]
+  mova     m1, [inputq +  16]
+  mova     m2, [inputq +  32]
+  mova     m3, [inputq +  48]
+  mova     m4, [inputq +  64]
+  mova     m5, [inputq +  80]
+  mova     m6, [inputq +  96]
+  mova     m7, [inputq + 112]
+%endif
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+  IDCT8_1D
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+  IDCT8_1D
+
+  pxor    m12, m12
+  ADD_STORE_8P_2X  0, 1, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  2, 3, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  4, 5, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  6, 7, 9, 10, 12
+
+  RET
+
+; inverse 8x8 2D-DCT transform with only first 10 coeffs non-zero
+cglobal idct8x8_12_add, 3, 5, 13, input, output, stride
+  mova       m8, [pd_8192]
+  mova      m11, [pw_16]
+  mova      m12, [pw_11585x2]
+
+  lea        r3, [2 * strideq]
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova       m0, [inputq +   0]
+  packssdw   m0, [inputq +  16]
+  mova       m1, [inputq +  32]
+  packssdw   m1, [inputq +  48]
+  mova       m2, [inputq +  64]
+  packssdw   m2, [inputq +  80]
+  mova       m3, [inputq +  96]
+  packssdw   m3, [inputq + 112]
+%else
+  mova       m0, [inputq +  0]
+  mova       m1, [inputq + 16]
+  mova       m2, [inputq + 32]
+  mova       m3, [inputq + 48]
+%endif
+
+  punpcklwd  m0, m1
+  punpcklwd  m2, m3
+  punpckhdq  m9, m0, m2
+  punpckldq  m0, m2
+  SWAP       2, 9
+
+  ; m0 -> [0], [0]
+  ; m1 -> [1], [1]
+  ; m2 -> [2], [2]
+  ; m3 -> [3], [3]
+  punpckhqdq m10, m0, m0
+  punpcklqdq m0,  m0
+  punpckhqdq m9,  m2, m2
+  punpcklqdq m2,  m2
+  SWAP       1, 10
+  SWAP       3,  9
+
+  pmulhrsw   m0, m12
+  pmulhrsw   m2, [dpw_30274_12540]
+  pmulhrsw   m1, [dpw_6392_32138]
+  pmulhrsw   m3, [dpw_m18204_27246]
+
+  SUM_SUB    0, 2, 9
+  SUM_SUB    1, 3, 9
+
+  punpcklqdq m9, m3, m3
+  punpckhqdq m5, m3, m9
+
+  SUM_SUB    3, 5, 9
+  punpckhqdq m5, m3
+  pmulhrsw   m5, m12
+
+  punpckhqdq m9, m1, m5
+  punpcklqdq m1, m5
+  SWAP       5, 9
+
+  SUM_SUB    0, 5, 9
+  SUM_SUB    2, 1, 9
+
+  punpckhqdq m3, m0, m0
+  punpckhqdq m4, m1, m1
+  punpckhqdq m6, m5, m5
+  punpckhqdq m7, m2, m2
+
+  punpcklwd  m0, m3
+  punpcklwd  m7, m2
+  punpcklwd  m1, m4
+  punpcklwd  m6, m5
+
+  punpckhdq  m4, m0, m7
+  punpckldq  m0, m7
+  punpckhdq  m10, m1, m6
+  punpckldq  m5, m1, m6
+
+  punpckhqdq m1, m0, m5
+  punpcklqdq m0, m5
+  punpckhqdq m3, m4, m10
+  punpcklqdq m2, m4, m10
+
+
+  pmulhrsw   m0, m12
+  pmulhrsw   m6, m2, [dpw_30274_30274]
+  pmulhrsw   m4, m2, [dpw_12540_12540]
+
+  pmulhrsw   m7, m1, [dpw_32138_32138]
+  pmulhrsw   m1, [dpw_6392_6392]
+  pmulhrsw   m5, m3, [dpw_m18204_m18204]
+  pmulhrsw   m3, [dpw_27246_27246]
+
+  mova       m2, m0
+  SUM_SUB    0, 6, 9
+  SUM_SUB    2, 4, 9
+  SUM_SUB    1, 5, 9
+  SUM_SUB    7, 3, 9
+
+  SUM_SUB    3, 5, 9
+  pmulhrsw   m3, m12
+  pmulhrsw   m5, m12
+
+  SUM_SUB    0, 7, 9
+  SUM_SUB    2, 3, 9
+  SUM_SUB    4, 5, 9
+  SUM_SUB    6, 1, 9
+
+  SWAP       3, 6
+  SWAP       1, 2
+  SWAP       2, 4
+
+
+  pxor    m12, m12
+  ADD_STORE_8P_2X  0, 1, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  2, 3, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  4, 5, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  6, 7, 9, 10, 12
+
+  RET
+
+%define  idx0 16 * 0
+%define  idx1 16 * 1
+%define  idx2 16 * 2
+%define  idx3 16 * 3
+%define  idx4 16 * 4
+%define  idx5 16 * 5
+%define  idx6 16 * 6
+%define  idx7 16 * 7
+%define  idx8 16 * 0
+%define  idx9 16 * 1
+%define idx10 16 * 2
+%define idx11 16 * 3
+%define idx12 16 * 4
+%define idx13 16 * 5
+%define idx14 16 * 6
+%define idx15 16 * 7
+%define idx16 16 * 0
+%define idx17 16 * 1
+%define idx18 16 * 2
+%define idx19 16 * 3
+%define idx20 16 * 4
+%define idx21 16 * 5
+%define idx22 16 * 6
+%define idx23 16 * 7
+%define idx24 16 * 0
+%define idx25 16 * 1
+%define idx26 16 * 2
+%define idx27 16 * 3
+%define idx28 16 * 4
+%define idx29 16 * 5
+%define idx30 16 * 6
+%define idx31 16 * 7
+
+; FROM idct32x32_add_neon.asm
+;
+; Instead of doing the transforms stage by stage, it is done by loading
+; some input values and doing as many stages as possible to minimize the
+; storing/loading of intermediate results. To fit within registers, the
+; final coefficients are cut into four blocks:
+; BLOCK A: 16-19,28-31
+; BLOCK B: 20-23,24-27
+; BLOCK C: 8-11,12-15
+; BLOCK D: 0-3,4-7
+; Blocks A and C are straight calculation through the various stages. In
+; block B, further calculations are performed using the results from
+; block A. In block D, further calculations are performed using the results
+; from block C and then the final calculations are done using results from
+; block A and B which have been combined at the end of block B.
+;
+
+%macro IDCT32X32_34 4
+  ; BLOCK A STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                m11, m1
+  pmulhrsw             m1, [pw___804x2] ; stp1_16
+  mova      [r4 +      0], m0
+  pmulhrsw            m11, [pw_16364x2] ; stp2_31
+  mova      [r4 + 16 * 2], m2
+  mova                m12, m7
+  pmulhrsw             m7, [pw_15426x2] ; stp1_28
+  mova      [r4 + 16 * 4], m4
+  pmulhrsw            m12, [pw_m5520x2] ; stp2_19
+  mova      [r4 + 16 * 6], m6
+
+  ; BLOCK A STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m2, m1   ; stp1_16
+  mova                 m0, m11  ; stp1_31
+  mova                 m4, m7   ; stp1_28
+  mova                m15, m12  ; stp1_19
+
+  ; BLOCK A STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X          0,     2,   3196, 16069,  m8,  9,  10 ; stp1_17, stp1_30
+  BUTTERFLY_4Xmm        4,    15,   3196, 16069,  m8,  9,  10 ; stp1_29, stp1_18
+
+  ; BLOCK A STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               1, 12, 9 ; stp2_16, stp2_19
+  SUM_SUB               0, 15, 9 ; stp2_17, stp2_18
+  SUM_SUB              11,  7, 9 ; stp2_31, stp2_28
+  SUM_SUB               2,  4, 9 ; stp2_30, stp2_29
+
+  ; BLOCK A STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X          4,    15,   6270, 15137,  m8,  9,  10 ; stp1_18, stp1_29
+  BUTTERFLY_4X          7,    12,   6270, 15137,  m8,  9,  10 ; stp1_19, stp1_28
+
+  ; BLOCK B STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m6, m5
+  pmulhrsw             m5, [pw__3981x2] ; stp1_20
+  mova [stp + %4 + idx28], m12
+  mova [stp + %4 + idx29], m15
+  pmulhrsw             m6, [pw_15893x2] ; stp2_27
+  mova [stp + %4 + idx30], m2
+  mova                 m2, m3
+  pmulhrsw             m3, [pw_m2404x2] ; stp1_23
+  mova [stp + %4 + idx31], m11
+  pmulhrsw             m2, [pw_16207x2] ; stp2_24
+
+  ; BLOCK B STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                m13, m5 ; stp1_20
+  mova                m14, m6 ; stp1_27
+  mova                m15, m3 ; stp1_23
+  mova                m11, m2 ; stp1_24
+
+  ; BLOCK B STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X         14,    13,  13623,  9102,  m8,  9,  10 ; stp1_21, stp1_26
+  BUTTERFLY_4Xmm       11,    15,  13623,  9102,  m8,  9,  10 ; stp1_25, stp1_22
+
+  ; BLOCK B STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               3,  5, 9 ; stp2_23, stp2_20
+  SUM_SUB              15, 14, 9 ; stp2_22, stp2_21
+  SUM_SUB               2,  6, 9 ; stp2_24, stp2_27
+  SUM_SUB              11, 13, 9 ; stp2_25, stp2_26
+
+  ; BLOCK B STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4Xmm        6,     5,   6270, 15137,  m8,  9,  10 ; stp1_27, stp1_20
+  BUTTERFLY_4Xmm       13,    14,   6270, 15137,  m8,  9,  10 ; stp1_26, stp1_21
+
+  ; BLOCK B STAGE 6 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               1,  3, 9 ; stp2_16, stp2_23
+  SUM_SUB               0, 15, 9 ; stp2_17, stp2_22
+  SUM_SUB               4, 14, 9 ; stp2_18, stp2_21
+  SUM_SUB               7,  5, 9 ; stp2_19, stp2_20
+  mova [stp + %3 + idx16], m1
+  mova [stp + %3 + idx17], m0
+  mova [stp + %3 + idx18], m4
+  mova [stp + %3 + idx19], m7
+
+  mova                 m4, [stp + %4 + idx28]
+  mova                 m7, [stp + %4 + idx29]
+  mova                m10, [stp + %4 + idx30]
+  mova                m12, [stp + %4 + idx31]
+  SUM_SUB               4,  6, 9 ; stp2_28, stp2_27
+  SUM_SUB               7, 13, 9 ; stp2_29, stp2_26
+  SUM_SUB              10, 11, 9 ; stp2_30, stp2_25
+  SUM_SUB              12,  2, 9 ; stp2_31, stp2_24
+  mova [stp + %4 + idx28], m4
+  mova [stp + %4 + idx29], m7
+  mova [stp + %4 + idx30], m10
+  mova [stp + %4 + idx31], m12
+
+  ; BLOCK B STAGE 7 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+%if 0 ; overflow occurs in SUM_SUB when using test streams
+  mova                m10, [pw_11585x2]
+  SUM_SUB               6,  5, 9
+  pmulhrsw             m6, m10  ; stp1_27
+  pmulhrsw             m5, m10  ; stp1_20
+  SUM_SUB              13, 14,  9
+  pmulhrsw            m13, m10  ; stp1_26
+  pmulhrsw            m14, m10  ; stp1_21
+  SUM_SUB              11, 15,  9
+  pmulhrsw            m11, m10  ; stp1_25
+  pmulhrsw            m15, m10  ; stp1_22
+  SUM_SUB               2,  3,  9
+  pmulhrsw             m2, m10  ; stp1_24
+  pmulhrsw             m3, m10  ; stp1_23
+%else
+  BUTTERFLY_4X          6,     5,  11585, 11585,  m8,  9,  10 ; stp1_20, stp1_27
+  SWAP 6, 5
+  BUTTERFLY_4X         13,    14,  11585, 11585,  m8,  9,  10 ; stp1_21, stp1_26
+  SWAP 13, 14
+  BUTTERFLY_4X         11,    15,  11585, 11585,  m8,  9,  10 ; stp1_22, stp1_25
+  SWAP 11, 15
+  BUTTERFLY_4X          2,     3,  11585, 11585,  m8,  9,  10 ; stp1_23, stp1_24
+  SWAP 2, 3
+%endif
+
+  mova [stp + %4 + idx24], m2
+  mova [stp + %4 + idx25], m11
+  mova [stp + %4 + idx26], m13
+  mova [stp + %4 + idx27], m6
+
+  ; BLOCK C STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  ;
+  ; BLOCK C STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m0, [rsp + transposed_in + 16 *  2]
+  mova                 m6, [rsp + transposed_in + 16 *  6]
+
+  mova                 m1, m0
+  pmulhrsw             m0, [pw__1606x2] ; stp1_8
+  mova [stp + %3 + idx20], m5
+  mova [stp + %3 + idx21], m14
+  pmulhrsw             m1, [pw_16305x2] ; stp2_15
+  mova [stp + %3 + idx22], m15
+  mova                 m7, m6
+  pmulhrsw             m7, [pw_m4756x2] ; stp2_11
+  mova [stp + %3 + idx23], m3
+  pmulhrsw             m6, [pw_15679x2] ; stp1_12
+
+  ; BLOCK C STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m3, m0 ; stp1_8
+  mova                 m2, m1 ; stp1_15
+
+  ; BLOCK C STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X          2,     3,   6270, 15137,  m8,  9,  10 ;  stp1_9, stp1_14
+  mova                 m4, m7 ; stp1_11
+  mova                 m5, m6 ; stp1_12
+  BUTTERFLY_4Xmm        5,     4,   6270, 15137,  m8,  9,  10 ; stp1_13, stp1_10
+
+  ; BLOCK C STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               0,  7, 9 ;  stp1_8, stp1_11
+  SUM_SUB               2,  4, 9 ;  stp1_9, stp1_10
+  SUM_SUB               1,  6, 9 ;  stp1_15, stp1_12
+  SUM_SUB               3,  5, 9 ;  stp1_14, stp1_13
+
+  ; BLOCK C STAGE 6 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+%if 0 ; overflow occurs in SUM_SUB when using test streams
+  mova                m10, [pw_11585x2]
+  SUM_SUB               5,  4, 9
+  pmulhrsw             m5, m10  ; stp1_13
+  pmulhrsw             m4, m10  ; stp1_10
+  SUM_SUB               6,  7, 9
+  pmulhrsw             m6, m10  ; stp1_12
+  pmulhrsw             m7, m10  ; stp1_11
+%else
+  BUTTERFLY_4X          5,     4,  11585, 11585,  m8,  9,  10 ; stp1_10, stp1_13
+  SWAP 5, 4
+  BUTTERFLY_4X          6,     7,  11585, 11585,  m8,  9,  10 ; stp1_11, stp1_12
+  SWAP 6, 7
+%endif
+
+  ; BLOCK C STAGE 7 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova [stp + %2 +  idx8], m0
+  mova [stp + %2 +  idx9], m2
+  mova [stp + %2 + idx10], m4
+  mova [stp + %2 + idx11], m7
+
+  ; BLOCK D STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  ;
+  ; BLOCK D STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  ;
+  ; BLOCK D STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                m11, [rsp + transposed_in + 16 *  4]
+  mova                m12, m11
+  pmulhrsw            m11, [pw__3196x2] ; stp1_4
+  pmulhrsw            m12, [pw_16069x2] ; stp1_7
+
+  ; BLOCK D STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m0, [rsp + transposed_in + 16 *  0]
+  mova                m10, [pw_11585x2]
+  pmulhrsw             m0, m10  ; stp1_1
+
+  mova                m14, m11 ; stp1_4
+  mova                m13, m12 ; stp1_7
+
+  ; BLOCK D STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+%if 0 ; overflow occurs in SUM_SUB when using test streams
+  SUM_SUB              13,   14,  9
+  pmulhrsw            m13, m10  ; stp1_6
+  pmulhrsw            m14, m10  ; stp1_5
+%else
+  BUTTERFLY_4X         13,    14,  11585, 11585,  m8,  9,  10 ; stp1_5, stp1_6
+  SWAP 13, 14
+%endif
+  mova                 m7, m0 ; stp1_0 = stp1_1
+  mova                 m4, m0 ; stp1_1
+  mova                 m2, m7 ; stp1_0
+
+  ; BLOCK D STAGE 6 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               0, 12, 9 ;  stp1_0, stp1_7
+  SUM_SUB               7, 13, 9 ;  stp1_1, stp1_6
+  SUM_SUB               2, 14, 9 ;  stp1_2, stp1_5
+  SUM_SUB               4, 11, 9 ;  stp1_3, stp1_4
+
+  ; BLOCK D STAGE 7 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               0,  1, 9 ;  stp1_0, stp1_15
+  SUM_SUB               7,  3, 9 ;  stp1_1, stp1_14
+  SUM_SUB               2,  5, 9 ;  stp1_2, stp1_13
+  SUM_SUB               4,  6, 9 ;  stp1_3, stp1_12
+
+  ; 0-3, 28-31 final stage
+  mova                m15, [stp + %4 + idx30]
+  mova                m10, [stp + %4 + idx31]
+  SUM_SUB               0, 10, 9 ;  stp1_0, stp1_31
+  SUM_SUB               7, 15, 9 ;  stp1_1, stp1_30
+  mova [stp + %1 +  idx0], m0
+  mova [stp + %1 +  idx1], m7
+  mova [stp + %4 + idx30], m15
+  mova [stp + %4 + idx31], m10
+  mova                 m7, [stp + %4 + idx28]
+  mova                 m0, [stp + %4 + idx29]
+  SUM_SUB               2,  0, 9 ;  stp1_2, stp1_29
+  SUM_SUB               4,  7, 9 ;  stp1_3, stp1_28
+  mova [stp + %1 +  idx2], m2
+  mova [stp + %1 +  idx3], m4
+  mova [stp + %4 + idx28], m7
+  mova [stp + %4 + idx29], m0
+
+  ; 12-15, 16-19 final stage
+  mova                 m0, [stp + %3 + idx16]
+  mova                 m7, [stp + %3 + idx17]
+  mova                 m2, [stp + %3 + idx18]
+  mova                 m4, [stp + %3 + idx19]
+  SUM_SUB               1,  0, 9 ;  stp1_15, stp1_16
+  SUM_SUB               3,  7, 9 ;  stp1_14, stp1_17
+  SUM_SUB               5,  2, 9 ;  stp1_13, stp1_18
+  SUM_SUB               6,  4, 9 ;  stp1_12, stp1_19
+  mova [stp + %2 + idx12], m6
+  mova [stp + %2 + idx13], m5
+  mova [stp + %2 + idx14], m3
+  mova [stp + %2 + idx15], m1
+  mova [stp + %3 + idx16], m0
+  mova [stp + %3 + idx17], m7
+  mova [stp + %3 + idx18], m2
+  mova [stp + %3 + idx19], m4
+
+  mova                 m4, [stp + %2 +  idx8]
+  mova                 m5, [stp + %2 +  idx9]
+  mova                 m6, [stp + %2 + idx10]
+  mova                 m7, [stp + %2 + idx11]
+  SUM_SUB              11,  7, 9 ;  stp1_4, stp1_11
+  SUM_SUB              14,  6, 9 ;  stp1_5, stp1_10
+  SUM_SUB              13,  5, 9 ;  stp1_6, stp1_9
+  SUM_SUB              12,  4, 9 ;  stp1_7, stp1_8
+
+  ; 4-7, 24-27 final stage
+  mova                 m0, [stp + %4 + idx27]
+  mova                 m1, [stp + %4 + idx26]
+  mova                 m2, [stp + %4 + idx25]
+  mova                 m3, [stp + %4 + idx24]
+  SUM_SUB              11,  0, 9 ;  stp1_4, stp1_27
+  SUM_SUB              14,  1, 9 ;  stp1_5, stp1_26
+  SUM_SUB              13,  2, 9 ;  stp1_6, stp1_25
+  SUM_SUB              12,  3, 9 ;  stp1_7, stp1_24
+  mova [stp + %4 + idx27], m0
+  mova [stp + %4 + idx26], m1
+  mova [stp + %4 + idx25], m2
+  mova [stp + %4 + idx24], m3
+  mova [stp + %1 +  idx4], m11
+  mova [stp + %1 +  idx5], m14
+  mova [stp + %1 +  idx6], m13
+  mova [stp + %1 +  idx7], m12
+
+  ; 8-11, 20-23 final stage
+  mova                 m0, [stp + %3 + idx20]
+  mova                 m1, [stp + %3 + idx21]
+  mova                 m2, [stp + %3 + idx22]
+  mova                 m3, [stp + %3 + idx23]
+  SUM_SUB               7,  0, 9 ;  stp1_11, stp_20
+  SUM_SUB               6,  1, 9 ;  stp1_10, stp_21
+  SUM_SUB               5,  2, 9 ;   stp1_9, stp_22
+  SUM_SUB               4,  3, 9 ;   stp1_8, stp_23
+  mova [stp + %2 +  idx8], m4
+  mova [stp + %2 +  idx9], m5
+  mova [stp + %2 + idx10], m6
+  mova [stp + %2 + idx11], m7
+  mova [stp + %3 + idx20], m0
+  mova [stp + %3 + idx21], m1
+  mova [stp + %3 + idx22], m2
+  mova [stp + %3 + idx23], m3
+%endmacro
+
+%macro RECON_AND_STORE 1
+  mova            m11, [pw_32]
+  lea             stp, [rsp + %1]
+  mov              r6, 32
+  pxor             m8, m8
+%%recon_and_store:
+  mova             m0, [stp + 16 * 32 * 0]
+  mova             m1, [stp + 16 * 32 * 1]
+  mova             m2, [stp + 16 * 32 * 2]
+  mova             m3, [stp + 16 * 32 * 3]
+  add             stp, 16
+
+  paddw            m0, m11
+  paddw            m1, m11
+  paddw            m2, m11
+  paddw            m3, m11
+  psraw            m0, 6
+  psraw            m1, 6
+  psraw            m2, 6
+  psraw            m3, 6
+  movh             m4, [outputq +  0]
+  movh             m5, [outputq +  8]
+  movh             m6, [outputq + 16]
+  movh             m7, [outputq + 24]
+  punpcklbw        m4, m8
+  punpcklbw        m5, m8
+  punpcklbw        m6, m8
+  punpcklbw        m7, m8
+  paddw            m0, m4
+  paddw            m1, m5
+  paddw            m2, m6
+  paddw            m3, m7
+  packuswb         m0, m1
+  packuswb         m2, m3
+  mova [outputq +  0], m0
+  mova [outputq + 16], m2
+  lea         outputq, [outputq + strideq]
+  dec              r6
+  jnz %%recon_and_store
+%endmacro
+
+%define i32x32_size     16*32*5
+%define pass_two_start  16*32*0
+%define transposed_in   16*32*4
+%define pass_one_start  16*32*0
+%define stp r8
+
+INIT_XMM ssse3
+cglobal idct32x32_34_add, 3, 11, 16, i32x32_size, input, output, stride
+  mova            m8, [pd_8192]
+  lea            stp, [rsp + pass_one_start]
+
+idct32x32_34:
+  mov             r3, inputq
+  lea             r4, [rsp + transposed_in]
+
+idct32x32_34_transpose:
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova            m0, [r3 +       0]
+  packssdw        m0, [r3 +      16]
+  mova            m1, [r3 + 32 *  4]
+  packssdw        m1, [r3 + 32 *  4 + 16]
+  mova            m2, [r3 + 32 *  8]
+  packssdw        m2, [r3 + 32 *  8 + 16]
+  mova            m3, [r3 + 32 * 12]
+  packssdw        m3, [r3 + 32 * 12 + 16]
+  mova            m4, [r3 + 32 * 16]
+  packssdw        m4, [r3 + 32 * 16 + 16]
+  mova            m5, [r3 + 32 * 20]
+  packssdw        m5, [r3 + 32 * 20 + 16]
+  mova            m6, [r3 + 32 * 24]
+  packssdw        m6, [r3 + 32 * 24 + 16]
+  mova            m7, [r3 + 32 * 28]
+  packssdw        m7, [r3 + 32 * 28 + 16]
+%else
+  mova            m0, [r3 +       0]
+  mova            m1, [r3 + 16 *  4]
+  mova            m2, [r3 + 16 *  8]
+  mova            m3, [r3 + 16 * 12]
+  mova            m4, [r3 + 16 * 16]
+  mova            m5, [r3 + 16 * 20]
+  mova            m6, [r3 + 16 * 24]
+  mova            m7, [r3 + 16 * 28]
+%endif
+
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  IDCT32X32_34  16*0, 16*32, 16*64, 16*96
+  lea            stp, [stp + 16 * 8]
+  mov             r6, 4
+  lea            stp, [rsp + pass_one_start]
+  lea             r9, [rsp + pass_one_start]
+
+idct32x32_34_2:
+  lea             r4, [rsp + transposed_in]
+  mov             r3, r9
+
+idct32x32_34_transpose_2:
+  mova            m0, [r3 +      0]
+  mova            m1, [r3 + 16 * 1]
+  mova            m2, [r3 + 16 * 2]
+  mova            m3, [r3 + 16 * 3]
+  mova            m4, [r3 + 16 * 4]
+  mova            m5, [r3 + 16 * 5]
+  mova            m6, [r3 + 16 * 6]
+  mova            m7, [r3 + 16 * 7]
+
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  IDCT32X32_34  16*0, 16*8, 16*16, 16*24
+
+  lea            stp, [stp + 16 * 32]
+  add             r9, 16 * 32
+  dec             r6
+  jnz idct32x32_34_2
+
+  RECON_AND_STORE pass_two_start
+
+  RET
+
+%macro IDCT32X32_135 4
+  ; BLOCK A STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m1, [rsp + transposed_in + 16 *  1]
+  mova                m11, m1
+  pmulhrsw             m1, [pw___804x2] ; stp1_16
+  pmulhrsw            m11, [pw_16364x2] ; stp2_31
+
+  mova                 m7, [rsp + transposed_in + 16 *  7]
+  mova                m12, m7
+  pmulhrsw             m7, [pw_15426x2] ; stp1_28
+  pmulhrsw            m12, [pw_m5520x2] ; stp2_19
+
+  mova                 m3, [rsp + transposed_in + 16 *  9]
+  mova                 m4, m3
+  pmulhrsw             m3, [pw__7005x2] ; stp1_18
+  pmulhrsw             m4, [pw_14811x2] ; stp2_29
+
+  mova                 m0, [rsp + transposed_in + 16 * 15]
+  mova                 m2, m0
+  pmulhrsw             m0, [pw_12140x2]  ; stp1_30
+  pmulhrsw             m2, [pw_m11003x2] ; stp2_17
+
+  ; BLOCK A STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               1,  2, 9 ; stp2_16, stp2_17
+  SUM_SUB              12,  3, 9 ; stp2_19, stp2_18
+  SUM_SUB               7,  4, 9 ; stp2_28, stp2_29
+  SUM_SUB              11,  0, 9 ; stp2_31, stp2_30
+
+  ; BLOCK A STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X          0,     2,   3196, 16069,  m8,  9,  10 ; stp1_17, stp1_30
+  BUTTERFLY_4Xmm        4,     3,   3196, 16069,  m8,  9,  10 ; stp1_29, stp1_18
+
+  ; BLOCK A STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               1, 12, 9 ; stp2_16, stp2_19
+  SUM_SUB               0,  3, 9 ; stp2_17, stp2_18
+  SUM_SUB              11,  7, 9 ; stp2_31, stp2_28
+  SUM_SUB               2,  4, 9 ; stp2_30, stp2_29
+
+  ; BLOCK A STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X          4,     3,   6270, 15137,  m8,  9,  10 ; stp1_18, stp1_29
+  BUTTERFLY_4X          7,    12,   6270, 15137,  m8,  9,  10 ; stp1_19, stp1_28
+
+  mova [stp + %3 + idx16], m1
+  mova [stp + %3 + idx17], m0
+  mova [stp + %3 + idx18], m4
+  mova [stp + %3 + idx19], m7
+  mova [stp + %4 + idx28], m12
+  mova [stp + %4 + idx29], m3
+  mova [stp + %4 + idx30], m2
+  mova [stp + %4 + idx31], m11
+
+  ; BLOCK B STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m2, [rsp + transposed_in + 16 *  3]
+  mova                 m3, m2
+  pmulhrsw             m3, [pw_m2404x2] ; stp1_23
+  pmulhrsw             m2, [pw_16207x2] ; stp2_24
+
+  mova                 m5, [rsp + transposed_in + 16 *  5]
+  mova                 m6, m5
+  pmulhrsw             m5, [pw__3981x2] ; stp1_20
+  pmulhrsw             m6, [pw_15893x2] ; stp2_27
+
+  mova                m14, [rsp + transposed_in + 16 * 11]
+  mova                m13, m14
+  pmulhrsw            m13, [pw_m8423x2] ; stp1_21
+  pmulhrsw            m14, [pw_14053x2] ; stp2_26
+
+  mova                 m0, [rsp + transposed_in + 16 * 13]
+  mova                 m1, m0
+  pmulhrsw             m0, [pw__9760x2] ; stp1_22
+  pmulhrsw             m1, [pw_13160x2] ; stp2_25
+
+  ; BLOCK B STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               5, 13, 9 ; stp2_20, stp2_21
+  SUM_SUB               3,  0, 9 ; stp2_23, stp2_22
+  SUM_SUB               2,  1, 9 ; stp2_24, stp2_25
+  SUM_SUB               6, 14, 9 ; stp2_27, stp2_26
+
+  ; BLOCK B STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X         14,    13,  13623,  9102,  m8,  9,  10 ; stp1_21, stp1_26
+  BUTTERFLY_4Xmm        1,     0,  13623,  9102,  m8,  9,  10 ; stp1_25, stp1_22
+
+  ; BLOCK B STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               3,  5, 9 ; stp2_23, stp2_20
+  SUM_SUB               0, 14, 9 ; stp2_22, stp2_21
+  SUM_SUB               2,  6, 9 ; stp2_24, stp2_27
+  SUM_SUB               1, 13, 9 ; stp2_25, stp2_26
+
+  ; BLOCK B STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4Xmm        6,     5,   6270, 15137,  m8,  9,  10 ; stp1_27, stp1_20
+  BUTTERFLY_4Xmm       13,    14,   6270, 15137,  m8,  9,  10 ; stp1_26, stp1_21
+
+  ; BLOCK B STAGE 6 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m4, [stp + %3 + idx16]
+  mova                 m7, [stp + %3 + idx17]
+  mova                m11, [stp + %3 + idx18]
+  mova                m12, [stp + %3 + idx19]
+  SUM_SUB               4,  3, 9 ; stp2_16, stp2_23
+  SUM_SUB               7,  0, 9 ; stp2_17, stp2_22
+  SUM_SUB              11, 14, 9 ; stp2_18, stp2_21
+  SUM_SUB              12,  5, 9 ; stp2_19, stp2_20
+  mova [stp + %3 + idx16], m4
+  mova [stp + %3 + idx17], m7
+  mova [stp + %3 + idx18], m11
+  mova [stp + %3 + idx19], m12
+
+  mova                 m4, [stp + %4 + idx28]
+  mova                 m7, [stp + %4 + idx29]
+  mova                m11, [stp + %4 + idx30]
+  mova                m12, [stp + %4 + idx31]
+  SUM_SUB               4,  6, 9 ; stp2_28, stp2_27
+  SUM_SUB               7, 13, 9 ; stp2_29, stp2_26
+  SUM_SUB              11,  1, 9 ; stp2_30, stp2_25
+  SUM_SUB              12,  2, 9 ; stp2_31, stp2_24
+  mova [stp + %4 + idx28], m4
+  mova [stp + %4 + idx29], m7
+  mova [stp + %4 + idx30], m11
+  mova [stp + %4 + idx31], m12
+
+  ; BLOCK B STAGE 7 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+%if 0 ; overflow occurs in SUM_SUB when using test streams
+  mova                m10, [pw_11585x2]
+  SUM_SUB               6,  5,  9
+  pmulhrsw             m6, m10  ; stp1_27
+  pmulhrsw             m5, m10  ; stp1_20
+  SUM_SUB              13, 14,  9
+  pmulhrsw            m13, m10  ; stp1_26
+  pmulhrsw            m14, m10  ; stp1_21
+  SUM_SUB               1,  0,  9
+  pmulhrsw             m1, m10  ; stp1_25
+  pmulhrsw             m0, m10  ; stp1_22
+  SUM_SUB               2,  3,  9
+  pmulhrsw             m2, m10  ; stp1_25
+  pmulhrsw             m3, m10  ; stp1_22
+%else
+  BUTTERFLY_4X          6,     5,  11585, 11585,  m8,  9,  10 ; stp1_20, stp1_27
+  SWAP  6, 5
+  BUTTERFLY_4X         13,    14,  11585, 11585,  m8,  9,  10 ; stp1_21, stp1_26
+  SWAP 13, 14
+  BUTTERFLY_4X          1,     0,  11585, 11585,  m8,  9,  10 ; stp1_22, stp1_25
+  SWAP  1, 0
+  BUTTERFLY_4X          2,     3,  11585, 11585,  m8,  9,  10 ; stp1_23, stp1_24
+  SWAP  2, 3
+%endif
+  mova [stp + %3 + idx20], m5
+  mova [stp + %3 + idx21], m14
+  mova [stp + %3 + idx22], m0
+  mova [stp + %3 + idx23], m3
+  mova [stp + %4 + idx24], m2
+  mova [stp + %4 + idx25], m1
+  mova [stp + %4 + idx26], m13
+  mova [stp + %4 + idx27], m6
+
+  ; BLOCK C STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  ;
+  ; BLOCK C STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m0, [rsp + transposed_in + 16 *  2]
+  mova                 m1, m0
+  pmulhrsw             m0, [pw__1606x2] ; stp1_8
+  pmulhrsw             m1, [pw_16305x2] ; stp2_15
+
+  mova                 m6, [rsp + transposed_in + 16 *  6]
+  mova                 m7, m6
+  pmulhrsw             m7, [pw_m4756x2] ; stp2_11
+  pmulhrsw             m6, [pw_15679x2] ; stp1_12
+
+  mova                 m4, [rsp + transposed_in + 16 * 10]
+  mova                 m5, m4
+  pmulhrsw             m4, [pw__7723x2] ; stp1_10
+  pmulhrsw             m5, [pw_14449x2] ; stp2_13
+
+  mova                 m2, [rsp + transposed_in + 16 * 14]
+  mova                 m3, m2
+  pmulhrsw             m3, [pw_m10394x2] ; stp1_9
+  pmulhrsw             m2, [pw_12665x2] ; stp2_14
+
+  ; BLOCK C STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               0,  3, 9 ;  stp1_8, stp1_9
+  SUM_SUB               7,  4, 9 ; stp1_11, stp1_10
+  SUM_SUB               6,  5, 9 ; stp1_12, stp1_13
+  SUM_SUB               1,  2, 9 ; stp1_15, stp1_14
+
+  ; BLOCK C STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X          2,     3,   6270, 15137,  m8,  9,  10 ;  stp1_9, stp1_14
+  BUTTERFLY_4Xmm        5,     4,   6270, 15137,  m8,  9,  10 ; stp1_13, stp1_10
+
+  ; BLOCK C STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               0,  7, 9 ;  stp1_8, stp1_11
+  SUM_SUB               2,  4, 9 ;  stp1_9, stp1_10
+  SUM_SUB               1,  6, 9 ;  stp1_15, stp1_12
+  SUM_SUB               3,  5, 9 ;  stp1_14, stp1_13
+
+  ; BLOCK C STAGE 6 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+%if 0 ; overflow occurs in SUM_SUB when using test streams
+  mova                m10, [pw_11585x2]
+  SUM_SUB               5,    4,  9
+  pmulhrsw             m5, m10  ; stp1_13
+  pmulhrsw             m4, m10  ; stp1_10
+  SUM_SUB               6,    7,  9
+  pmulhrsw             m6, m10  ; stp1_12
+  pmulhrsw             m7, m10  ; stp1_11
+%else
+  BUTTERFLY_4X       5,     4,  11585,  11585,  m8,  9,  10 ; stp1_10, stp1_13
+  SWAP  5, 4
+  BUTTERFLY_4X       6,     7,  11585,  11585,  m8,  9,  10 ; stp1_11, stp1_12
+  SWAP  6, 7
+%endif
+  ; BLOCK C STAGE 7 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova [stp + %2 +  idx8], m0
+  mova [stp + %2 +  idx9], m2
+  mova [stp + %2 + idx10], m4
+  mova [stp + %2 + idx11], m7
+  mova [stp + %2 + idx12], m6
+  mova [stp + %2 + idx13], m5
+  mova [stp + %2 + idx14], m3
+  mova [stp + %2 + idx15], m1
+
+  ; BLOCK D STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  ;
+  ; BLOCK D STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  ;
+  ; BLOCK D STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                m11, [rsp + transposed_in + 16 *  4]
+  mova                m12, m11
+  pmulhrsw            m11, [pw__3196x2] ; stp1_4
+  pmulhrsw            m12, [pw_16069x2] ; stp1_7
+
+  mova                m13, [rsp + transposed_in + 16 * 12]
+  mova                m14, m13
+  pmulhrsw            m13, [pw_13623x2] ; stp1_6
+  pmulhrsw            m14, [pw_m9102x2] ; stp1_5
+
+  ; BLOCK D STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m0, [rsp + transposed_in + 16 *  0]
+  mova                 m2, [rsp + transposed_in + 16 *  8]
+  pmulhrsw             m0, [pw_11585x2]  ; stp1_1
+  mova                 m3, m2
+  pmulhrsw             m2, [pw__6270x2]  ; stp1_2
+  pmulhrsw             m3, [pw_15137x2]  ; stp1_3
+
+  SUM_SUB              11, 14, 9 ;  stp1_4, stp1_5
+  SUM_SUB              12, 13, 9 ;  stp1_7, stp1_6
+
+  ; BLOCK D STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+%if 0 ; overflow occurs in SUM_SUB when using test streams
+  mova                m10, [pw_11585x2]
+  SUM_SUB              13,   14,  9
+  pmulhrsw            m13, m10  ; stp1_6
+  pmulhrsw            m14, m10  ; stp1_5
+%else
+  BUTTERFLY_4X         13,    14,  11585, 11585,  m8,  9,  10 ; stp1_5, stp1_6
+  SWAP 13, 14
+%endif
+  mova                 m1, m0    ; stp1_0 = stp1_1
+  SUM_SUB               0,  3, 9 ;  stp1_0, stp1_3
+  SUM_SUB               1,  2, 9 ;  stp1_1, stp1_2
+
+  ; BLOCK D STAGE 6 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               0, 12, 9 ;  stp1_0, stp1_7
+  SUM_SUB               1, 13, 9 ;  stp1_1, stp1_6
+  SUM_SUB               2, 14, 9 ;  stp1_2, stp1_5
+  SUM_SUB               3, 11, 9 ;  stp1_3, stp1_4
+
+  ; BLOCK D STAGE 7 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m4, [stp + %2 + idx12]
+  mova                 m5, [stp + %2 + idx13]
+  mova                 m6, [stp + %2 + idx14]
+  mova                 m7, [stp + %2 + idx15]
+  SUM_SUB               0,  7, 9 ;  stp1_0, stp1_15
+  SUM_SUB               1,  6, 9 ;  stp1_1, stp1_14
+  SUM_SUB               2,  5, 9 ;  stp1_2, stp1_13
+  SUM_SUB               3,  4, 9 ;  stp1_3, stp1_12
+
+  ; 0-3, 28-31 final stage
+  mova                m10, [stp + %4 + idx31]
+  mova                m15, [stp + %4 + idx30]
+  SUM_SUB               0, 10, 9 ;  stp1_0, stp1_31
+  SUM_SUB               1, 15, 9 ;  stp1_1, stp1_30
+  mova [stp + %1 +  idx0], m0
+  mova [stp + %1 +  idx1], m1
+  mova [stp + %4 + idx31], m10
+  mova [stp + %4 + idx30], m15
+  mova                 m0, [stp + %4 + idx29]
+  mova                 m1, [stp + %4 + idx28]
+  SUM_SUB               2,  0, 9 ;  stp1_2, stp1_29
+  SUM_SUB               3,  1, 9 ;  stp1_3, stp1_28
+  mova [stp + %1 +  idx2], m2
+  mova [stp + %1 +  idx3], m3
+  mova [stp + %4 + idx29], m0
+  mova [stp + %4 + idx28], m1
+
+  ; 12-15, 16-19 final stage
+  mova                 m0, [stp + %3 + idx16]
+  mova                 m1, [stp + %3 + idx17]
+  mova                 m2, [stp + %3 + idx18]
+  mova                 m3, [stp + %3 + idx19]
+  SUM_SUB               7,  0, 9 ;  stp1_15, stp1_16
+  SUM_SUB               6,  1, 9 ;  stp1_14, stp1_17
+  SUM_SUB               5,  2, 9 ;  stp1_13, stp1_18
+  SUM_SUB               4,  3, 9 ;  stp1_12, stp1_19
+  mova [stp + %2 + idx12], m4
+  mova [stp + %2 + idx13], m5
+  mova [stp + %2 + idx14], m6
+  mova [stp + %2 + idx15], m7
+  mova [stp + %3 + idx16], m0
+  mova [stp + %3 + idx17], m1
+  mova [stp + %3 + idx18], m2
+  mova [stp + %3 + idx19], m3
+
+  mova                 m4, [stp + %2 +  idx8]
+  mova                 m5, [stp + %2 +  idx9]
+  mova                 m6, [stp + %2 + idx10]
+  mova                 m7, [stp + %2 + idx11]
+  SUM_SUB              11,  7, 9 ;  stp1_4, stp1_11
+  SUM_SUB              14,  6, 9 ;  stp1_5, stp1_10
+  SUM_SUB              13,  5, 9 ;  stp1_6, stp1_9
+  SUM_SUB              12,  4, 9 ;  stp1_7, stp1_8
+
+  ; 4-7, 24-27 final stage
+  mova                 m3, [stp + %4 + idx24]
+  mova                 m2, [stp + %4 + idx25]
+  mova                 m1, [stp + %4 + idx26]
+  mova                 m0, [stp + %4 + idx27]
+  SUM_SUB              12,  3, 9 ;  stp1_7, stp1_24
+  SUM_SUB              13,  2, 9 ;  stp1_6, stp1_25
+  SUM_SUB              14,  1, 9 ;  stp1_5, stp1_26
+  SUM_SUB              11,  0, 9 ;  stp1_4, stp1_27
+  mova [stp + %4 + idx24], m3
+  mova [stp + %4 + idx25], m2
+  mova [stp + %4 + idx26], m1
+  mova [stp + %4 + idx27], m0
+  mova [stp + %1 +  idx4], m11
+  mova [stp + %1 +  idx5], m14
+  mova [stp + %1 +  idx6], m13
+  mova [stp + %1 +  idx7], m12
+
+  ; 8-11, 20-23 final stage
+  mova                 m0, [stp + %3 + idx20]
+  mova                 m1, [stp + %3 + idx21]
+  mova                 m2, [stp + %3 + idx22]
+  mova                 m3, [stp + %3 + idx23]
+  SUM_SUB               7,  0, 9 ;  stp1_11, stp_20
+  SUM_SUB               6,  1, 9 ;  stp1_10, stp_21
+  SUM_SUB               5,  2, 9 ;   stp1_9, stp_22
+  SUM_SUB               4,  3, 9 ;   stp1_8, stp_23
+  mova [stp + %2 +  idx8], m4
+  mova [stp + %2 +  idx9], m5
+  mova [stp + %2 + idx10], m6
+  mova [stp + %2 + idx11], m7
+  mova [stp + %3 + idx20], m0
+  mova [stp + %3 + idx21], m1
+  mova [stp + %3 + idx22], m2
+  mova [stp + %3 + idx23], m3
+%endmacro
+
+INIT_XMM ssse3
+cglobal idct32x32_135_add, 3, 11, 16, i32x32_size, input, output, stride
+  mova            m8, [pd_8192]
+  mov             r6, 2
+  lea            stp, [rsp + pass_one_start]
+
+idct32x32_135:
+  mov             r3, inputq
+  lea             r4, [rsp + transposed_in]
+  mov             r7, 2
+
+idct32x32_135_transpose:
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova            m0, [r3 +       0]
+  packssdw        m0, [r3 +      16]
+  mova            m1, [r3 + 32 *  4]
+  packssdw        m1, [r3 + 32 *  4 + 16]
+  mova            m2, [r3 + 32 *  8]
+  packssdw        m2, [r3 + 32 *  8 + 16]
+  mova            m3, [r3 + 32 * 12]
+  packssdw        m3, [r3 + 32 * 12 + 16]
+  mova            m4, [r3 + 32 * 16]
+  packssdw        m4, [r3 + 32 * 16 + 16]
+  mova            m5, [r3 + 32 * 20]
+  packssdw        m5, [r3 + 32 * 20 + 16]
+  mova            m6, [r3 + 32 * 24]
+  packssdw        m6, [r3 + 32 * 24 + 16]
+  mova            m7, [r3 + 32 * 28]
+  packssdw        m7, [r3 + 32 * 28 + 16]
+%else
+  mova            m0, [r3 +       0]
+  mova            m1, [r3 + 16 *  4]
+  mova            m2, [r3 + 16 *  8]
+  mova            m3, [r3 + 16 * 12]
+  mova            m4, [r3 + 16 * 16]
+  mova            m5, [r3 + 16 * 20]
+  mova            m6, [r3 + 16 * 24]
+  mova            m7, [r3 + 16 * 28]
+%endif
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  mova [r4 +      0], m0
+  mova [r4 + 16 * 1], m1
+  mova [r4 + 16 * 2], m2
+  mova [r4 + 16 * 3], m3
+  mova [r4 + 16 * 4], m4
+  mova [r4 + 16 * 5], m5
+  mova [r4 + 16 * 6], m6
+  mova [r4 + 16 * 7], m7
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  add             r3, 32
+%else
+  add             r3, 16
+%endif
+  add             r4, 16 * 8
+  dec             r7
+  jne idct32x32_135_transpose
+
+  IDCT32X32_135 16*0, 16*32, 16*64, 16*96
+  lea            stp, [stp + 16 * 8]
+%if CONFIG_VP9_HIGHBITDEPTH
+  lea         inputq, [inputq + 32 * 32]
+%else
+  lea         inputq, [inputq + 16 * 32]
+%endif
+  dec             r6
+  jnz idct32x32_135
+
+  mov             r6, 4
+  lea            stp, [rsp + pass_one_start]
+  lea             r9, [rsp + pass_one_start]
+
+idct32x32_135_2:
+  lea             r4, [rsp + transposed_in]
+  mov             r3, r9
+  mov             r7, 2
+
+idct32x32_135_transpose_2:
+  mova            m0, [r3 +      0]
+  mova            m1, [r3 + 16 * 1]
+  mova            m2, [r3 + 16 * 2]
+  mova            m3, [r3 + 16 * 3]
+  mova            m4, [r3 + 16 * 4]
+  mova            m5, [r3 + 16 * 5]
+  mova            m6, [r3 + 16 * 6]
+  mova            m7, [r3 + 16 * 7]
+
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  mova [r4 +      0], m0
+  mova [r4 + 16 * 1], m1
+  mova [r4 + 16 * 2], m2
+  mova [r4 + 16 * 3], m3
+  mova [r4 + 16 * 4], m4
+  mova [r4 + 16 * 5], m5
+  mova [r4 + 16 * 6], m6
+  mova [r4 + 16 * 7], m7
+
+  add             r3, 16 * 8
+  add             r4, 16 * 8
+  dec             r7
+  jne idct32x32_135_transpose_2
+
+  IDCT32X32_135 16*0, 16*8, 16*16, 16*24
+
+  lea            stp, [stp + 16 * 32]
+  add             r9, 16 * 32
+  dec             r6
+  jnz idct32x32_135_2
+
+  RECON_AND_STORE pass_two_start
+
+  RET
+
+%macro IDCT32X32_1024 4
+  ; BLOCK A STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m1, [rsp + transposed_in + 16 *  1]
+  mova                m11, [rsp + transposed_in + 16 * 31]
+  BUTTERFLY_4X          1,    11,    804, 16364,  m8,  9,  10 ; stp1_16, stp1_31
+
+  mova                 m0, [rsp + transposed_in + 16 * 15]
+  mova                 m2, [rsp + transposed_in + 16 * 17]
+  BUTTERFLY_4X          2,     0,  12140, 11003,  m8,  9,  10 ; stp1_17, stp1_30
+
+  mova                 m7, [rsp + transposed_in + 16 *  7]
+  mova                m12, [rsp + transposed_in + 16 * 25]
+  BUTTERFLY_4X         12,     7,  15426,  5520,  m8,  9,  10 ; stp1_19, stp1_28
+
+  mova                 m3, [rsp + transposed_in + 16 *  9]
+  mova                 m4, [rsp + transposed_in + 16 * 23]
+  BUTTERFLY_4X          3,     4,   7005, 14811,  m8,  9,  10 ; stp1_18, stp1_29
+
+  ; BLOCK A STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               1,  2, 9 ; stp2_16, stp2_17
+  SUM_SUB              12,  3, 9 ; stp2_19, stp2_18
+  SUM_SUB               7,  4, 9 ; stp2_28, stp2_29
+  SUM_SUB              11,  0, 9 ; stp2_31, stp2_30
+
+  ; BLOCK A STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X          0,     2,   3196, 16069,  m8,  9,  10 ; stp1_17, stp1_30
+  BUTTERFLY_4Xmm        4,     3,   3196, 16069,  m8,  9,  10 ; stp1_29, stp1_18
+
+  ; BLOCK A STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               1, 12, 9 ; stp2_16, stp2_19
+  SUM_SUB               0,  3, 9 ; stp2_17, stp2_18
+  SUM_SUB              11,  7, 9 ; stp2_31, stp2_28
+  SUM_SUB               2,  4, 9 ; stp2_30, stp2_29
+
+  ; BLOCK A STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X          4,     3,   6270, 15137,  m8,  9,  10 ; stp1_18, stp1_29
+  BUTTERFLY_4X          7,    12,   6270, 15137,  m8,  9,  10 ; stp1_19, stp1_28
+
+  mova [stp + %3 + idx16], m1
+  mova [stp + %3 + idx17], m0
+  mova [stp + %3 + idx18], m4
+  mova [stp + %3 + idx19], m7
+  mova [stp + %4 + idx28], m12
+  mova [stp + %4 + idx29], m3
+  mova [stp + %4 + idx30], m2
+  mova [stp + %4 + idx31], m11
+
+  ; BLOCK B STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m5, [rsp + transposed_in + 16 *  5]
+  mova                 m6, [rsp + transposed_in + 16 * 27]
+  BUTTERFLY_4X          5,     6,   3981, 15893,  m8,  9,  10 ; stp1_20, stp1_27
+
+  mova                m13, [rsp + transposed_in + 16 * 21]
+  mova                m14, [rsp + transposed_in + 16 * 11]
+  BUTTERFLY_4X         13,    14,  14053,  8423,  m8,  9,  10 ; stp1_21, stp1_26
+
+  mova                 m0, [rsp + transposed_in + 16 * 13]
+  mova                 m1, [rsp + transposed_in + 16 * 19]
+  BUTTERFLY_4X          0,     1,   9760, 13160,  m8,  9,  10 ; stp1_22, stp1_25
+
+  mova                 m2, [rsp + transposed_in + 16 *  3]
+  mova                 m3, [rsp + transposed_in + 16 * 29]
+  BUTTERFLY_4X          3,     2,  16207,  2404,  m8,  9,  10 ; stp1_23, stp1_24
+
+  ; BLOCK B STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               5, 13, 9 ; stp2_20, stp2_21
+  SUM_SUB               3,  0, 9 ; stp2_23, stp2_22
+  SUM_SUB               2,  1, 9 ; stp2_24, stp2_25
+  SUM_SUB               6, 14, 9 ; stp2_27, stp2_26
+
+  ; BLOCK B STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X         14,    13,  13623,  9102,  m8,  9,  10 ; stp1_21, stp1_26
+  BUTTERFLY_4Xmm        1,     0,  13623,  9102,  m8,  9,  10 ; stp1_25, stp1_22
+
+  ; BLOCK B STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               3,  5, 9 ; stp2_23, stp2_20
+  SUM_SUB               0, 14, 9 ; stp2_22, stp2_21
+  SUM_SUB               2,  6, 9 ; stp2_24, stp2_27
+  SUM_SUB               1, 13, 9 ; stp2_25, stp2_26
+
+  ; BLOCK B STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4Xmm        6,     5,   6270, 15137,  m8,  9,  10 ; stp1_27, stp1_20
+  BUTTERFLY_4Xmm       13,    14,   6270, 15137,  m8,  9,  10 ; stp1_26, stp1_21
+
+  ; BLOCK B STAGE 6 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m4, [stp + %3 + idx16]
+  mova                 m7, [stp + %3 + idx17]
+  mova                m11, [stp + %3 + idx18]
+  mova                m12, [stp + %3 + idx19]
+  SUM_SUB               4,  3, 9 ; stp2_16, stp2_23
+  SUM_SUB               7,  0, 9 ; stp2_17, stp2_22
+  SUM_SUB              11, 14, 9 ; stp2_18, stp2_21
+  SUM_SUB              12,  5, 9 ; stp2_19, stp2_20
+  mova [stp + %3 + idx16], m4
+  mova [stp + %3 + idx17], m7
+  mova [stp + %3 + idx18], m11
+  mova [stp + %3 + idx19], m12
+
+  mova                 m4, [stp + %4 + idx28]
+  mova                 m7, [stp + %4 + idx29]
+  mova                m11, [stp + %4 + idx30]
+  mova                m12, [stp + %4 + idx31]
+  SUM_SUB               4,  6, 9 ; stp2_28, stp2_27
+  SUM_SUB               7, 13, 9 ; stp2_29, stp2_26
+  SUM_SUB              11,  1, 9 ; stp2_30, stp2_25
+  SUM_SUB              12,  2, 9 ; stp2_31, stp2_24
+  mova [stp + %4 + idx28], m4
+  mova [stp + %4 + idx29], m7
+  mova [stp + %4 + idx30], m11
+  mova [stp + %4 + idx31], m12
+
+  ; BLOCK B STAGE 7 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+%if 0 ; overflow occurs in SUM_SUB when using test streams
+  mova                m10, [pw_11585x2]
+  SUM_SUB               6,  5,  9
+  pmulhrsw             m6, m10  ; stp1_27
+  pmulhrsw             m5, m10  ; stp1_20
+  SUM_SUB              13, 14,  9
+  pmulhrsw            m13, m10  ; stp1_26
+  pmulhrsw            m14, m10  ; stp1_21
+  SUM_SUB               1,  0,  9
+  pmulhrsw             m1, m10  ; stp1_25
+  pmulhrsw             m0, m10  ; stp1_22
+  SUM_SUB               2,  3,  9
+  pmulhrsw             m2, m10  ; stp1_25
+  pmulhrsw             m3, m10  ; stp1_22
+%else
+  BUTTERFLY_4X          6,     5,  11585, 11585,  m8,  9,  10 ; stp1_20, stp1_27
+  SWAP  6, 5
+  BUTTERFLY_4X         13,    14,  11585, 11585,  m8,  9,  10 ; stp1_21, stp1_26
+  SWAP 13, 14
+  BUTTERFLY_4X          1,     0,  11585, 11585,  m8,  9,  10 ; stp1_22, stp1_25
+  SWAP  1, 0
+  BUTTERFLY_4X          2,     3,  11585, 11585,  m8,  9,  10 ; stp1_23, stp1_24
+  SWAP  2, 3
+%endif
+  mova [stp + %3 + idx20], m5
+  mova [stp + %3 + idx21], m14
+  mova [stp + %3 + idx22], m0
+  mova [stp + %3 + idx23], m3
+  mova [stp + %4 + idx24], m2
+  mova [stp + %4 + idx25], m1
+  mova [stp + %4 + idx26], m13
+  mova [stp + %4 + idx27], m6
+
+  ; BLOCK C STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  ;
+  ; BLOCK C STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m0, [rsp + transposed_in + 16 *  2]
+  mova                 m1, [rsp + transposed_in + 16 * 30]
+  BUTTERFLY_4X          0,     1,   1606, 16305,  m8,  9,  10 ; stp1_8, stp1_15
+
+  mova                 m2, [rsp + transposed_in + 16 * 14]
+  mova                 m3, [rsp + transposed_in + 16 * 18]
+  BUTTERFLY_4X          3,     2,  12665, 10394,  m8,  9,  10 ; stp1_9, stp1_14
+
+  mova                 m4, [rsp + transposed_in + 16 * 10]
+  mova                 m5, [rsp + transposed_in + 16 * 22]
+  BUTTERFLY_4X          4,     5,   7723, 14449,  m8,  9,  10 ; stp1_10, stp1_13
+
+  mova                 m6, [rsp + transposed_in + 16 *  6]
+  mova                 m7, [rsp + transposed_in + 16 * 26]
+  BUTTERFLY_4X          7,     6,  15679,  4756,  m8,  9,  10 ; stp1_11, stp1_12
+
+  ; BLOCK C STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               0,  3, 9 ;  stp1_8, stp1_9
+  SUM_SUB               7,  4, 9 ; stp1_11, stp1_10
+  SUM_SUB               6,  5, 9 ; stp1_12, stp1_13
+  SUM_SUB               1,  2, 9 ; stp1_15, stp1_14
+
+  ; BLOCK C STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  BUTTERFLY_4X          2,     3,   6270, 15137,  m8,  9,  10 ;  stp1_9, stp1_14
+  BUTTERFLY_4Xmm        5,     4,   6270, 15137,  m8,  9,  10 ; stp1_13, stp1_10
+
+  ; BLOCK C STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               0,  7, 9 ;  stp1_8, stp1_11
+  SUM_SUB               2,  4, 9 ;  stp1_9, stp1_10
+  SUM_SUB               1,  6, 9 ;  stp1_15, stp1_12
+  SUM_SUB               3,  5, 9 ;  stp1_14, stp1_13
+
+  ; BLOCK C STAGE 6 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+%if 0 ; overflow occurs in SUM_SUB when using test streams
+  mova                m10, [pw_11585x2]
+  SUM_SUB               5,    4,  9
+  pmulhrsw             m5, m10  ; stp1_13
+  pmulhrsw             m4, m10  ; stp1_10
+  SUM_SUB               6,    7,  9
+  pmulhrsw             m6, m10  ; stp1_12
+  pmulhrsw             m7, m10  ; stp1_11
+%else
+  BUTTERFLY_4X       5,     4,  11585,  11585,  m8,  9,  10 ; stp1_10, stp1_13
+  SWAP  5, 4
+  BUTTERFLY_4X       6,     7,  11585,  11585,  m8,  9,  10 ; stp1_11, stp1_12
+  SWAP  6, 7
+%endif
+  ; BLOCK C STAGE 7 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova [stp + %2 +  idx8], m0
+  mova [stp + %2 +  idx9], m2
+  mova [stp + %2 + idx10], m4
+  mova [stp + %2 + idx11], m7
+  mova [stp + %2 + idx12], m6
+  mova [stp + %2 + idx13], m5
+  mova [stp + %2 + idx14], m3
+  mova [stp + %2 + idx15], m1
+
+  ; BLOCK D STAGE 1 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  ;
+  ; BLOCK D STAGE 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  ;
+  ; BLOCK D STAGE 3 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                m11, [rsp + transposed_in + 16 *  4]
+  mova                m12, [rsp + transposed_in + 16 * 28]
+  BUTTERFLY_4X         11,    12,   3196, 16069,  m8,  9,  10 ; stp1_4, stp1_7
+
+  mova                m13, [rsp + transposed_in + 16 * 12]
+  mova                m14, [rsp + transposed_in + 16 * 20]
+  BUTTERFLY_4X         14,    13,  13623,  9102,  m8,  9,  10 ; stp1_5, stp1_6
+
+  ; BLOCK D STAGE 4 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m0, [rsp + transposed_in + 16 *  0]
+  mova                 m1, [rsp + transposed_in + 16 * 16]
+
+%if 0 ; overflow occurs in SUM_SUB when using test streams
+  mova                m10, [pw_11585x2]
+  SUM_SUB               0,    1,  9
+  pmulhrsw             m0, m10  ; stp1_1
+  pmulhrsw             m1, m10  ; stp1_0
+%else
+  BUTTERFLY_4X          0,     1,  11585, 11585,  m8,  9,  10 ; stp1_1, stp1_0
+  SWAP  0, 1
+%endif
+  mova                 m2, [rsp + transposed_in + 16 *  8]
+  mova                 m3, [rsp + transposed_in + 16 * 24]
+  BUTTERFLY_4X          2,     3,   6270, 15137,  m8,  9,  10 ;  stp1_2, stp1_3
+
+  mova                m10, [pw_11585x2]
+  SUM_SUB              11, 14, 9 ;  stp1_4, stp1_5
+  SUM_SUB              12, 13, 9 ;  stp1_7, stp1_6
+
+  ; BLOCK D STAGE 5 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+%if 0 ; overflow occurs in SUM_SUB when using test streams
+  SUM_SUB              13,   14,  9
+  pmulhrsw            m13, m10  ; stp1_6
+  pmulhrsw            m14, m10  ; stp1_5
+%else
+  BUTTERFLY_4X         13,    14,  11585, 11585,  m8,  9,  10 ; stp1_5, stp1_6
+  SWAP 13, 14
+%endif
+  SUM_SUB               0,  3, 9 ;  stp1_0, stp1_3
+  SUM_SUB               1,  2, 9 ;  stp1_1, stp1_2
+
+  ; BLOCK D STAGE 6 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  SUM_SUB               0, 12, 9 ;  stp1_0, stp1_7
+  SUM_SUB               1, 13, 9 ;  stp1_1, stp1_6
+  SUM_SUB               2, 14, 9 ;  stp1_2, stp1_5
+  SUM_SUB               3, 11, 9 ;  stp1_3, stp1_4
+
+  ; BLOCK D STAGE 7 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  mova                 m4, [stp + %2 + idx12]
+  mova                 m5, [stp + %2 + idx13]
+  mova                 m6, [stp + %2 + idx14]
+  mova                 m7, [stp + %2 + idx15]
+  SUM_SUB               0,  7, 9 ;  stp1_0, stp1_15
+  SUM_SUB               1,  6, 9 ;  stp1_1, stp1_14
+  SUM_SUB               2,  5, 9 ;  stp1_2, stp1_13
+  SUM_SUB               3,  4, 9 ;  stp1_3, stp1_12
+
+  ; 0-3, 28-31 final stage
+  mova                m10, [stp + %4 + idx31]
+  mova                m15, [stp + %4 + idx30]
+  SUM_SUB               0, 10, 9 ;  stp1_0, stp1_31
+  SUM_SUB               1, 15, 9 ;  stp1_1, stp1_30
+  mova [stp + %1 +  idx0], m0
+  mova [stp + %1 +  idx1], m1
+  mova [stp + %4 + idx31], m10
+  mova [stp + %4 + idx30], m15
+  mova                 m0, [stp + %4 + idx29]
+  mova                 m1, [stp + %4 + idx28]
+  SUM_SUB               2,  0, 9 ;  stp1_2, stp1_29
+  SUM_SUB               3,  1, 9 ;  stp1_3, stp1_28
+  mova [stp + %1 +  idx2], m2
+  mova [stp + %1 +  idx3], m3
+  mova [stp + %4 + idx29], m0
+  mova [stp + %4 + idx28], m1
+
+  ; 12-15, 16-19 final stage
+  mova                 m0, [stp + %3 + idx16]
+  mova                 m1, [stp + %3 + idx17]
+  mova                 m2, [stp + %3 + idx18]
+  mova                 m3, [stp + %3 + idx19]
+  SUM_SUB               7,  0, 9 ;  stp1_15, stp1_16
+  SUM_SUB               6,  1, 9 ;  stp1_14, stp1_17
+  SUM_SUB               5,  2, 9 ;  stp1_13, stp1_18
+  SUM_SUB               4,  3, 9 ;  stp1_12, stp1_19
+  mova [stp + %2 + idx12], m4
+  mova [stp + %2 + idx13], m5
+  mova [stp + %2 + idx14], m6
+  mova [stp + %2 + idx15], m7
+  mova [stp + %3 + idx16], m0
+  mova [stp + %3 + idx17], m1
+  mova [stp + %3 + idx18], m2
+  mova [stp + %3 + idx19], m3
+
+  mova                 m4, [stp + %2 +  idx8]
+  mova                 m5, [stp + %2 +  idx9]
+  mova                 m6, [stp + %2 + idx10]
+  mova                 m7, [stp + %2 + idx11]
+  SUM_SUB              11,  7, 9 ;  stp1_4, stp1_11
+  SUM_SUB              14,  6, 9 ;  stp1_5, stp1_10
+  SUM_SUB              13,  5, 9 ;  stp1_6, stp1_9
+  SUM_SUB              12,  4, 9 ;  stp1_7, stp1_8
+
+  ; 4-7, 24-27 final stage
+  mova                 m3, [stp + %4 + idx24]
+  mova                 m2, [stp + %4 + idx25]
+  mova                 m1, [stp + %4 + idx26]
+  mova                 m0, [stp + %4 + idx27]
+  SUM_SUB              12,  3, 9 ;  stp1_7, stp1_24
+  SUM_SUB              13,  2, 9 ;  stp1_6, stp1_25
+  SUM_SUB              14,  1, 9 ;  stp1_5, stp1_26
+  SUM_SUB              11,  0, 9 ;  stp1_4, stp1_27
+  mova [stp + %4 + idx24], m3
+  mova [stp + %4 + idx25], m2
+  mova [stp + %4 + idx26], m1
+  mova [stp + %4 + idx27], m0
+  mova [stp + %1 +  idx4], m11
+  mova [stp + %1 +  idx5], m14
+  mova [stp + %1 +  idx6], m13
+  mova [stp + %1 +  idx7], m12
+
+  ; 8-11, 20-23 final stage
+  mova                 m0, [stp + %3 + idx20]
+  mova                 m1, [stp + %3 + idx21]
+  mova                 m2, [stp + %3 + idx22]
+  mova                 m3, [stp + %3 + idx23]
+  SUM_SUB               7,  0, 9 ;  stp1_11, stp_20
+  SUM_SUB               6,  1, 9 ;  stp1_10, stp_21
+  SUM_SUB               5,  2, 9 ;   stp1_9, stp_22
+  SUM_SUB               4,  3, 9 ;   stp1_8, stp_23
+  mova [stp + %2 +  idx8], m4
+  mova [stp + %2 +  idx9], m5
+  mova [stp + %2 + idx10], m6
+  mova [stp + %2 + idx11], m7
+  mova [stp + %3 + idx20], m0
+  mova [stp + %3 + idx21], m1
+  mova [stp + %3 + idx22], m2
+  mova [stp + %3 + idx23], m3
+%endmacro
+
+INIT_XMM ssse3
+cglobal idct32x32_1024_add, 3, 11, 16, i32x32_size, input, output, stride
+  mova            m8, [pd_8192]
+  mov             r6, 4
+  lea            stp, [rsp + pass_one_start]
+
+idct32x32_1024:
+  mov             r3, inputq
+  lea             r4, [rsp + transposed_in]
+  mov             r7, 4
+
+idct32x32_1024_transpose:
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova            m0, [r3 +       0]
+  packssdw        m0, [r3 +      16]
+  mova            m1, [r3 + 32 *  4]
+  packssdw        m1, [r3 + 32 *  4 + 16]
+  mova            m2, [r3 + 32 *  8]
+  packssdw        m2, [r3 + 32 *  8 + 16]
+  mova            m3, [r3 + 32 * 12]
+  packssdw        m3, [r3 + 32 * 12 + 16]
+  mova            m4, [r3 + 32 * 16]
+  packssdw        m4, [r3 + 32 * 16 + 16]
+  mova            m5, [r3 + 32 * 20]
+  packssdw        m5, [r3 + 32 * 20 + 16]
+  mova            m6, [r3 + 32 * 24]
+  packssdw        m6, [r3 + 32 * 24 + 16]
+  mova            m7, [r3 + 32 * 28]
+  packssdw        m7, [r3 + 32 * 28 + 16]
+%else
+  mova            m0, [r3 +       0]
+  mova            m1, [r3 + 16 *  4]
+  mova            m2, [r3 + 16 *  8]
+  mova            m3, [r3 + 16 * 12]
+  mova            m4, [r3 + 16 * 16]
+  mova            m5, [r3 + 16 * 20]
+  mova            m6, [r3 + 16 * 24]
+  mova            m7, [r3 + 16 * 28]
+%endif
+
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  mova [r4 +      0], m0
+  mova [r4 + 16 * 1], m1
+  mova [r4 + 16 * 2], m2
+  mova [r4 + 16 * 3], m3
+  mova [r4 + 16 * 4], m4
+  mova [r4 + 16 * 5], m5
+  mova [r4 + 16 * 6], m6
+  mova [r4 + 16 * 7], m7
+%if CONFIG_VP9_HIGHBITDEPTH
+  add             r3, 32
+%else
+  add             r3, 16
+%endif
+  add             r4, 16 * 8
+  dec             r7
+  jne idct32x32_1024_transpose
+
+  IDCT32X32_1024 16*0, 16*32, 16*64, 16*96
+
+  lea            stp, [stp + 16 * 8]
+%if CONFIG_VP9_HIGHBITDEPTH
+  lea         inputq, [inputq + 32 * 32]
+%else
+  lea         inputq, [inputq + 16 * 32]
+%endif
+  dec             r6
+  jnz idct32x32_1024
+
+  mov             r6, 4
+  lea            stp, [rsp + pass_one_start]
+  lea             r9, [rsp + pass_one_start]
+
+idct32x32_1024_2:
+  lea             r4, [rsp + transposed_in]
+  mov             r3, r9
+  mov             r7, 4
+
+idct32x32_1024_transpose_2:
+  mova            m0, [r3 +      0]
+  mova            m1, [r3 + 16 * 1]
+  mova            m2, [r3 + 16 * 2]
+  mova            m3, [r3 + 16 * 3]
+  mova            m4, [r3 + 16 * 4]
+  mova            m5, [r3 + 16 * 5]
+  mova            m6, [r3 + 16 * 6]
+  mova            m7, [r3 + 16 * 7]
+
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  mova [r4 +      0], m0
+  mova [r4 + 16 * 1], m1
+  mova [r4 + 16 * 2], m2
+  mova [r4 + 16 * 3], m3
+  mova [r4 + 16 * 4], m4
+  mova [r4 + 16 * 5], m5
+  mova [r4 + 16 * 6], m6
+  mova [r4 + 16 * 7], m7
+
+  add             r3, 16 * 8
+  add             r4, 16 * 8
+  dec             r7
+  jne idct32x32_1024_transpose_2
+
+  IDCT32X32_1024 16*0, 16*8, 16*16, 16*24
+
+  lea            stp, [stp + 16 * 32]
+  add             r9, 16 * 32
+  dec             r6
+  jnz idct32x32_1024_2
+
+  RECON_AND_STORE pass_two_start
+
+  RET
+%endif
diff --git a/libvpx/libvpx/vpx_dsp/x86/inv_wht_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/inv_wht_sse2.asm
new file mode 100644
index 0000000..fbbcd76
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/inv_wht_sse2.asm
@@ -0,0 +1,109 @@
+;
+;  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro REORDER_INPUTS 0
+  ; a c d b  to  a b c d
+  SWAP 1, 3, 2
+%endmacro
+
+%macro TRANSFORM_COLS 0
+  ; input:
+  ; m0 a
+  ; m1 b
+  ; m2 c
+  ; m3 d
+  paddw           m0,        m2
+  psubw           m3,        m1
+
+  ; wide subtract
+  punpcklwd       m4,        m0
+  punpcklwd       m5,        m3
+  psrad           m4,        16
+  psrad           m5,        16
+  psubd           m4,        m5
+  psrad           m4,        1
+  packssdw        m4,        m4             ; e
+
+  psubw           m5,        m4,        m1  ; b
+  psubw           m4,        m2             ; c
+  psubw           m0,        m5
+  paddw           m3,        m4
+                                ; m0 a
+  SWAP            1,         5  ; m1 b
+  SWAP            2,         4  ; m2 c
+                                ; m3 d
+%endmacro
+
+%macro TRANSPOSE_4X4 0
+  punpcklwd       m0,        m2
+  punpcklwd       m1,        m3
+  mova            m2,        m0
+  punpcklwd       m0,        m1
+  punpckhwd       m2,        m1
+  pshufd          m1,        m0, 0x0e
+  pshufd          m3,        m2, 0x0e
+%endmacro
+
+; transpose a 4x4 int16 matrix in xmm0 and xmm1 to the bottom half of xmm0-xmm3
+%macro TRANSPOSE_4X4_WIDE 0
+  mova            m3, m0
+  punpcklwd       m0, m1
+  punpckhwd       m3, m1
+  mova            m2, m0
+  punpcklwd       m0, m3
+  punpckhwd       m2, m3
+  pshufd          m1, m0, 0x0e
+  pshufd          m3, m2, 0x0e
+%endmacro
+
+%macro ADD_STORE_4P_2X 5  ; src1, src2, tmp1, tmp2, zero
+  movd            m%3,       [outputq]
+  movd            m%4,       [outputq + strideq]
+  punpcklbw       m%3,       m%5
+  punpcklbw       m%4,       m%5
+  paddw           m%1,       m%3
+  paddw           m%2,       m%4
+  packuswb        m%1,       m%5
+  packuswb        m%2,       m%5
+  movd            [outputq], m%1
+  movd            [outputq + strideq], m%2
+%endmacro
+
+INIT_XMM sse2
+cglobal iwht4x4_16_add, 3, 3, 7, input, output, stride
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova            m0,        [inputq +  0]
+  packssdw        m0,        [inputq + 16]
+  mova            m1,        [inputq + 32]
+  packssdw        m1,        [inputq + 48]
+%else
+  mova            m0,        [inputq +  0]
+  mova            m1,        [inputq + 16]
+%endif
+  psraw           m0,        2
+  psraw           m1,        2
+
+  TRANSPOSE_4X4_WIDE
+  REORDER_INPUTS
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+  REORDER_INPUTS
+  TRANSFORM_COLS
+
+  pxor            m4, m4
+  ADD_STORE_4P_2X  0, 1, 5, 6, 4
+  lea             outputq, [outputq + 2 * strideq]
+  ADD_STORE_4P_2X  2, 3, 5, 6, 4
+
+  RET
diff --git a/libvpx/libvpx/vpx_dsp/x86/loopfilter_avx2.c b/libvpx/libvpx/vpx_dsp/x86/loopfilter_avx2.c
new file mode 100644
index 0000000..be1087c
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/loopfilter_avx2.c
@@ -0,0 +1,979 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  /* AVX2 */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+
+void vpx_lpf_horizontal_edge_8_avx2(unsigned char *s, int p,
+                                    const unsigned char *_blimit,
+                                    const unsigned char *_limit,
+                                    const unsigned char *_thresh) {
+    __m128i mask, hev, flat, flat2;
+    const __m128i zero = _mm_set1_epi16(0);
+    const __m128i one = _mm_set1_epi8(1);
+    __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
+    __m128i abs_p1p0;
+
+    const __m128i thresh = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _thresh[0]));
+    const __m128i limit = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _limit[0]));
+    const __m128i blimit = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _blimit[0]));
+
+    q4p4 = _mm_loadl_epi64((__m128i *) (s - 5 * p));
+    q4p4 = _mm_castps_si128(
+            _mm_loadh_pi(_mm_castsi128_ps(q4p4), (__m64 *) (s + 4 * p)));
+    q3p3 = _mm_loadl_epi64((__m128i *) (s - 4 * p));
+    q3p3 = _mm_castps_si128(
+            _mm_loadh_pi(_mm_castsi128_ps(q3p3), (__m64 *) (s + 3 * p)));
+    q2p2 = _mm_loadl_epi64((__m128i *) (s - 3 * p));
+    q2p2 = _mm_castps_si128(
+            _mm_loadh_pi(_mm_castsi128_ps(q2p2), (__m64 *) (s + 2 * p)));
+    q1p1 = _mm_loadl_epi64((__m128i *) (s - 2 * p));
+    q1p1 = _mm_castps_si128(
+            _mm_loadh_pi(_mm_castsi128_ps(q1p1), (__m64 *) (s + 1 * p)));
+    p1q1 = _mm_shuffle_epi32(q1p1, 78);
+    q0p0 = _mm_loadl_epi64((__m128i *) (s - 1 * p));
+    q0p0 = _mm_castps_si128(
+            _mm_loadh_pi(_mm_castsi128_ps(q0p0), (__m64 *) (s - 0 * p)));
+    p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+    {
+        __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
+        abs_p1p0 = _mm_or_si128(_mm_subs_epu8(q1p1, q0p0),
+                _mm_subs_epu8(q0p0, q1p1));
+        abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+        fe = _mm_set1_epi8(0xfe);
+        ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+        abs_p0q0 = _mm_or_si128(_mm_subs_epu8(q0p0, p0q0),
+                _mm_subs_epu8(p0q0, q0p0));
+        abs_p1q1 = _mm_or_si128(_mm_subs_epu8(q1p1, p1q1),
+                _mm_subs_epu8(p1q1, q1p1));
+        flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+        hev = _mm_subs_epu8(flat, thresh);
+        hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+        abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+        abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+        mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+        mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+        // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+        mask = _mm_max_epu8(abs_p1p0, mask);
+        // mask |= (abs(p1 - p0) > limit) * -1;
+        // mask |= (abs(q1 - q0) > limit) * -1;
+
+        work = _mm_max_epu8(
+                _mm_or_si128(_mm_subs_epu8(q2p2, q1p1),
+                        _mm_subs_epu8(q1p1, q2p2)),
+                _mm_or_si128(_mm_subs_epu8(q3p3, q2p2),
+                        _mm_subs_epu8(q2p2, q3p3)));
+        mask = _mm_max_epu8(work, mask);
+        mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+        mask = _mm_subs_epu8(mask, limit);
+        mask = _mm_cmpeq_epi8(mask, zero);
+    }
+
+    // lp filter
+    {
+        const __m128i t4 = _mm_set1_epi8(4);
+        const __m128i t3 = _mm_set1_epi8(3);
+        const __m128i t80 = _mm_set1_epi8(0x80);
+        const __m128i t1 = _mm_set1_epi16(0x1);
+        __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
+        __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
+        __m128i qs0 = _mm_xor_si128(p0q0, t80);
+        __m128i qs1 = _mm_xor_si128(p1q1, t80);
+        __m128i filt;
+        __m128i work_a;
+        __m128i filter1, filter2;
+        __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+        __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
+
+        filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
+        work_a = _mm_subs_epi8(qs0, qs0ps0);
+        filt = _mm_adds_epi8(filt, work_a);
+        filt = _mm_adds_epi8(filt, work_a);
+        filt = _mm_adds_epi8(filt, work_a);
+        /* (vpx_filter + 3 * (qs0 - ps0)) & mask */
+        filt = _mm_and_si128(filt, mask);
+
+        filter1 = _mm_adds_epi8(filt, t4);
+        filter2 = _mm_adds_epi8(filt, t3);
+
+        filter1 = _mm_unpacklo_epi8(zero, filter1);
+        filter1 = _mm_srai_epi16(filter1, 0xB);
+        filter2 = _mm_unpacklo_epi8(zero, filter2);
+        filter2 = _mm_srai_epi16(filter2, 0xB);
+
+        /* Filter1 >> 3 */
+        filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
+        qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
+
+        /* filt >> 1 */
+        filt = _mm_adds_epi16(filter1, t1);
+        filt = _mm_srai_epi16(filt, 1);
+        filt = _mm_andnot_si128(
+                _mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8), filt);
+        filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
+        qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
+        // loopfilter done
+
+        {
+            __m128i work;
+            flat = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(q2p2, q0p0),
+                            _mm_subs_epu8(q0p0, q2p2)),
+                    _mm_or_si128(_mm_subs_epu8(q3p3, q0p0),
+                            _mm_subs_epu8(q0p0, q3p3)));
+            flat = _mm_max_epu8(abs_p1p0, flat);
+            flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+            flat = _mm_subs_epu8(flat, one);
+            flat = _mm_cmpeq_epi8(flat, zero);
+            flat = _mm_and_si128(flat, mask);
+
+            q5p5 = _mm_loadl_epi64((__m128i *) (s - 6 * p));
+            q5p5 = _mm_castps_si128(
+                    _mm_loadh_pi(_mm_castsi128_ps(q5p5),
+                            (__m64 *) (s + 5 * p)));
+
+            q6p6 = _mm_loadl_epi64((__m128i *) (s - 7 * p));
+            q6p6 = _mm_castps_si128(
+                    _mm_loadh_pi(_mm_castsi128_ps(q6p6),
+                            (__m64 *) (s + 6 * p)));
+
+            flat2 = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(q4p4, q0p0),
+                            _mm_subs_epu8(q0p0, q4p4)),
+                    _mm_or_si128(_mm_subs_epu8(q5p5, q0p0),
+                            _mm_subs_epu8(q0p0, q5p5)));
+
+            q7p7 = _mm_loadl_epi64((__m128i *) (s - 8 * p));
+            q7p7 = _mm_castps_si128(
+                    _mm_loadh_pi(_mm_castsi128_ps(q7p7),
+                            (__m64 *) (s + 7 * p)));
+
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(q6p6, q0p0),
+                            _mm_subs_epu8(q0p0, q6p6)),
+                    _mm_or_si128(_mm_subs_epu8(q7p7, q0p0),
+                            _mm_subs_epu8(q0p0, q7p7)));
+
+            flat2 = _mm_max_epu8(work, flat2);
+            flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
+            flat2 = _mm_subs_epu8(flat2, one);
+            flat2 = _mm_cmpeq_epi8(flat2, zero);
+            flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+        }
+
+        // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+        // flat and wide flat calculations
+        {
+            const __m128i eight = _mm_set1_epi16(8);
+            const __m128i four = _mm_set1_epi16(4);
+            __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+            __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+            __m128i pixelFilter_p, pixelFilter_q;
+            __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+            __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
+
+            p7_16 = _mm_unpacklo_epi8(q7p7, zero);
+            p6_16 = _mm_unpacklo_epi8(q6p6, zero);
+            p5_16 = _mm_unpacklo_epi8(q5p5, zero);
+            p4_16 = _mm_unpacklo_epi8(q4p4, zero);
+            p3_16 = _mm_unpacklo_epi8(q3p3, zero);
+            p2_16 = _mm_unpacklo_epi8(q2p2, zero);
+            p1_16 = _mm_unpacklo_epi8(q1p1, zero);
+            p0_16 = _mm_unpacklo_epi8(q0p0, zero);
+            q0_16 = _mm_unpackhi_epi8(q0p0, zero);
+            q1_16 = _mm_unpackhi_epi8(q1p1, zero);
+            q2_16 = _mm_unpackhi_epi8(q2p2, zero);
+            q3_16 = _mm_unpackhi_epi8(q3p3, zero);
+            q4_16 = _mm_unpackhi_epi8(q4p4, zero);
+            q5_16 = _mm_unpackhi_epi8(q5p5, zero);
+            q6_16 = _mm_unpackhi_epi8(q6p6, zero);
+            q7_16 = _mm_unpackhi_epi8(q7p7, zero);
+
+            pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
+                    _mm_add_epi16(p4_16, p3_16));
+            pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
+                    _mm_add_epi16(q4_16, q3_16));
+
+            pixetFilter_p2p1p0 = _mm_add_epi16(p0_16,
+                    _mm_add_epi16(p2_16, p1_16));
+            pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+            pixetFilter_q2q1q0 = _mm_add_epi16(q0_16,
+                    _mm_add_epi16(q2_16, q1_16));
+            pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+            pixelFilter_p = _mm_add_epi16(eight,
+                    _mm_add_epi16(pixelFilter_p, pixelFilter_q));
+            pixetFilter_p2p1p0 = _mm_add_epi16(four,
+                    _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)),
+                    4);
+            flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_p2p1p0,
+                            _mm_add_epi16(p3_16, p0_16)), 3);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_p2p1p0,
+                            _mm_add_epi16(q3_16, q0_16)), 3);
+
+            flat_q0p0 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(p7_16, p7_16);
+            sum_q7 = _mm_add_epi16(q7_16, q7_16);
+            sum_p3 = _mm_add_epi16(p3_16, p3_16);
+            sum_q3 = _mm_add_epi16(q3_16, q3_16);
+
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)),
+                    4);
+            flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+            pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
+            pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_p2p1p0,
+                            _mm_add_epi16(sum_p3, p1_16)), 3);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_q2q1q0,
+                            _mm_add_epi16(sum_q3, q1_16)), 3);
+            flat_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+            sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+            sum_p3 = _mm_add_epi16(sum_p3, p3_16);
+            sum_q3 = _mm_add_epi16(sum_q3, q3_16);
+
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)),
+                    4);
+            flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+            pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
+            pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
+
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_p2p1p0,
+                            _mm_add_epi16(sum_p3, p2_16)), 3);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_q2q1q0,
+                            _mm_add_epi16(sum_q3, q2_16)), 3);
+            flat_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+            sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)),
+                    4);
+            flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+            sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)),
+                    4);
+            flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+            sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)),
+                    4);
+            flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+            sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)),
+                    4);
+            flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
+        }
+        // wide flat
+        // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+        flat = _mm_shuffle_epi32(flat, 68);
+        flat2 = _mm_shuffle_epi32(flat2, 68);
+
+        q2p2 = _mm_andnot_si128(flat, q2p2);
+        flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
+        q2p2 = _mm_or_si128(q2p2, flat_q2p2);
+
+        qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+        flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
+        q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
+
+        qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+        flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
+        q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
+
+        q6p6 = _mm_andnot_si128(flat2, q6p6);
+        flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
+        q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
+        _mm_storel_epi64((__m128i *) (s - 7 * p), q6p6);
+        _mm_storeh_pi((__m64 *) (s + 6 * p), _mm_castsi128_ps(q6p6));
+
+        q5p5 = _mm_andnot_si128(flat2, q5p5);
+        flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
+        q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
+        _mm_storel_epi64((__m128i *) (s - 6 * p), q5p5);
+        _mm_storeh_pi((__m64 *) (s + 5 * p), _mm_castsi128_ps(q5p5));
+
+        q4p4 = _mm_andnot_si128(flat2, q4p4);
+        flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
+        q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
+        _mm_storel_epi64((__m128i *) (s - 5 * p), q4p4);
+        _mm_storeh_pi((__m64 *) (s + 4 * p), _mm_castsi128_ps(q4p4));
+
+        q3p3 = _mm_andnot_si128(flat2, q3p3);
+        flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
+        q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
+        _mm_storel_epi64((__m128i *) (s - 4 * p), q3p3);
+        _mm_storeh_pi((__m64 *) (s + 3 * p), _mm_castsi128_ps(q3p3));
+
+        q2p2 = _mm_andnot_si128(flat2, q2p2);
+        flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
+        q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
+        _mm_storel_epi64((__m128i *) (s - 3 * p), q2p2);
+        _mm_storeh_pi((__m64 *) (s + 2 * p), _mm_castsi128_ps(q2p2));
+
+        q1p1 = _mm_andnot_si128(flat2, q1p1);
+        flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
+        q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
+        _mm_storel_epi64((__m128i *) (s - 2 * p), q1p1);
+        _mm_storeh_pi((__m64 *) (s + 1 * p), _mm_castsi128_ps(q1p1));
+
+        q0p0 = _mm_andnot_si128(flat2, q0p0);
+        flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
+        q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
+        _mm_storel_epi64((__m128i *) (s - 1 * p), q0p0);
+        _mm_storeh_pi((__m64 *) (s - 0 * p), _mm_castsi128_ps(q0p0));
+    }
+}
+
+DECLARE_ALIGNED(32, static const uint8_t, filt_loopfilter_avx2[32]) = {
+  0, 128, 1, 128, 2, 128, 3, 128, 4, 128, 5, 128, 6, 128, 7, 128,
+  8, 128, 9, 128, 10, 128, 11, 128, 12, 128, 13, 128, 14, 128, 15, 128
+};
+
+void vpx_lpf_horizontal_edge_16_avx2(unsigned char *s, int p,
+                                     const unsigned char *_blimit,
+                                     const unsigned char *_limit,
+                                     const unsigned char *_thresh) {
+    __m128i mask, hev, flat, flat2;
+    const __m128i zero = _mm_set1_epi16(0);
+    const __m128i one = _mm_set1_epi8(1);
+    __m128i p7, p6, p5;
+    __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
+    __m128i q5, q6, q7;
+    __m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4,
+            q256_4, p256_3, q256_3, p256_2, q256_2, p256_1, q256_1,
+            p256_0, q256_0;
+
+    const __m128i thresh = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _thresh[0]));
+    const __m128i limit = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _limit[0]));
+    const __m128i blimit = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _blimit[0]));
+
+    p256_4 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 5 * p)));
+    p256_3 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 4 * p)));
+    p256_2 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 3 * p)));
+    p256_1 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 2 * p)));
+    p256_0 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 1 * p)));
+    q256_0 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 0 * p)));
+    q256_1 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 1 * p)));
+    q256_2 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 2 * p)));
+    q256_3 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 3 * p)));
+    q256_4 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 4 * p)));
+
+    p4 = _mm256_castsi256_si128(p256_4);
+    p3 = _mm256_castsi256_si128(p256_3);
+    p2 = _mm256_castsi256_si128(p256_2);
+    p1 = _mm256_castsi256_si128(p256_1);
+    p0 = _mm256_castsi256_si128(p256_0);
+    q0 = _mm256_castsi256_si128(q256_0);
+    q1 = _mm256_castsi256_si128(q256_1);
+    q2 = _mm256_castsi256_si128(q256_2);
+    q3 = _mm256_castsi256_si128(q256_3);
+    q4 = _mm256_castsi256_si128(q256_4);
+
+    {
+        const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
+                _mm_subs_epu8(p0, p1));
+        const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
+                _mm_subs_epu8(q0, q1));
+        const __m128i fe = _mm_set1_epi8(0xfe);
+        const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+        __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
+                _mm_subs_epu8(q0, p0));
+        __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
+                _mm_subs_epu8(q1, p1));
+        __m128i work;
+        flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+        hev = _mm_subs_epu8(flat, thresh);
+        hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+        abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+        abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+        mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+        mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+        // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+        mask = _mm_max_epu8(flat, mask);
+        // mask |= (abs(p1 - p0) > limit) * -1;
+        // mask |= (abs(q1 - q0) > limit) * -1;
+        work = _mm_max_epu8(
+                _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
+                _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
+        mask = _mm_max_epu8(work, mask);
+        work = _mm_max_epu8(
+                _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
+                _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
+        mask = _mm_max_epu8(work, mask);
+        mask = _mm_subs_epu8(mask, limit);
+        mask = _mm_cmpeq_epi8(mask, zero);
+    }
+
+    // lp filter
+    {
+        const __m128i t4 = _mm_set1_epi8(4);
+        const __m128i t3 = _mm_set1_epi8(3);
+        const __m128i t80 = _mm_set1_epi8(0x80);
+        const __m128i te0 = _mm_set1_epi8(0xe0);
+        const __m128i t1f = _mm_set1_epi8(0x1f);
+        const __m128i t1 = _mm_set1_epi8(0x1);
+        const __m128i t7f = _mm_set1_epi8(0x7f);
+
+        __m128i ps1 = _mm_xor_si128(p1, t80);
+        __m128i ps0 = _mm_xor_si128(p0, t80);
+        __m128i qs0 = _mm_xor_si128(q0, t80);
+        __m128i qs1 = _mm_xor_si128(q1, t80);
+        __m128i filt;
+        __m128i work_a;
+        __m128i filter1, filter2;
+        __m128i flat2_p6, flat2_p5, flat2_p4, flat2_p3, flat2_p2, flat2_p1,
+                flat2_p0, flat2_q0, flat2_q1, flat2_q2, flat2_q3, flat2_q4,
+                flat2_q5, flat2_q6, flat_p2, flat_p1, flat_p0, flat_q0, flat_q1,
+                flat_q2;
+
+        filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+        work_a = _mm_subs_epi8(qs0, ps0);
+        filt = _mm_adds_epi8(filt, work_a);
+        filt = _mm_adds_epi8(filt, work_a);
+        filt = _mm_adds_epi8(filt, work_a);
+        /* (vpx_filter + 3 * (qs0 - ps0)) & mask */
+        filt = _mm_and_si128(filt, mask);
+
+        filter1 = _mm_adds_epi8(filt, t4);
+        filter2 = _mm_adds_epi8(filt, t3);
+
+        /* Filter1 >> 3 */
+        work_a = _mm_cmpgt_epi8(zero, filter1);
+        filter1 = _mm_srli_epi16(filter1, 3);
+        work_a = _mm_and_si128(work_a, te0);
+        filter1 = _mm_and_si128(filter1, t1f);
+        filter1 = _mm_or_si128(filter1, work_a);
+        qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+
+        /* Filter2 >> 3 */
+        work_a = _mm_cmpgt_epi8(zero, filter2);
+        filter2 = _mm_srli_epi16(filter2, 3);
+        work_a = _mm_and_si128(work_a, te0);
+        filter2 = _mm_and_si128(filter2, t1f);
+        filter2 = _mm_or_si128(filter2, work_a);
+        ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+
+        /* filt >> 1 */
+        filt = _mm_adds_epi8(filter1, t1);
+        work_a = _mm_cmpgt_epi8(zero, filt);
+        filt = _mm_srli_epi16(filt, 1);
+        work_a = _mm_and_si128(work_a, t80);
+        filt = _mm_and_si128(filt, t7f);
+        filt = _mm_or_si128(filt, work_a);
+        filt = _mm_andnot_si128(hev, filt);
+        ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+        qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+        // loopfilter done
+
+        {
+            __m128i work;
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
+                    _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
+            flat = _mm_max_epu8(work, flat);
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)),
+                    _mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3)));
+            flat = _mm_max_epu8(work, flat);
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p4, p0), _mm_subs_epu8(p0, p4)),
+                    _mm_or_si128(_mm_subs_epu8(q4, q0), _mm_subs_epu8(q0, q4)));
+            flat = _mm_subs_epu8(flat, one);
+            flat = _mm_cmpeq_epi8(flat, zero);
+            flat = _mm_and_si128(flat, mask);
+
+            p256_5 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s - 6 * p)));
+            q256_5 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s + 5 * p)));
+            p5 = _mm256_castsi256_si128(p256_5);
+            q5 = _mm256_castsi256_si128(q256_5);
+            flat2 = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p5, p0), _mm_subs_epu8(p0, p5)),
+                    _mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5)));
+
+            flat2 = _mm_max_epu8(work, flat2);
+            p256_6 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s - 7 * p)));
+            q256_6 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s + 6 * p)));
+            p6 = _mm256_castsi256_si128(p256_6);
+            q6 = _mm256_castsi256_si128(q256_6);
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p6, p0), _mm_subs_epu8(p0, p6)),
+                    _mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6)));
+
+            flat2 = _mm_max_epu8(work, flat2);
+
+            p256_7 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s - 8 * p)));
+            q256_7 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s + 7 * p)));
+            p7 = _mm256_castsi256_si128(p256_7);
+            q7 = _mm256_castsi256_si128(q256_7);
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p7, p0), _mm_subs_epu8(p0, p7)),
+                    _mm_or_si128(_mm_subs_epu8(q7, q0), _mm_subs_epu8(q0, q7)));
+
+            flat2 = _mm_max_epu8(work, flat2);
+            flat2 = _mm_subs_epu8(flat2, one);
+            flat2 = _mm_cmpeq_epi8(flat2, zero);
+            flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+        }
+
+        // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+        // flat and wide flat calculations
+        {
+            const __m256i eight = _mm256_set1_epi16(8);
+            const __m256i four = _mm256_set1_epi16(4);
+            __m256i pixelFilter_p, pixelFilter_q, pixetFilter_p2p1p0,
+                    pixetFilter_q2q1q0, sum_p7, sum_q7, sum_p3, sum_q3, res_p,
+                    res_q;
+
+            const __m256i filter = _mm256_load_si256(
+                                  (__m256i const *)filt_loopfilter_avx2);
+            p256_7 = _mm256_shuffle_epi8(p256_7, filter);
+            p256_6 = _mm256_shuffle_epi8(p256_6, filter);
+            p256_5 = _mm256_shuffle_epi8(p256_5, filter);
+            p256_4 = _mm256_shuffle_epi8(p256_4, filter);
+            p256_3 = _mm256_shuffle_epi8(p256_3, filter);
+            p256_2 = _mm256_shuffle_epi8(p256_2, filter);
+            p256_1 = _mm256_shuffle_epi8(p256_1, filter);
+            p256_0 = _mm256_shuffle_epi8(p256_0, filter);
+            q256_0 = _mm256_shuffle_epi8(q256_0, filter);
+            q256_1 = _mm256_shuffle_epi8(q256_1, filter);
+            q256_2 = _mm256_shuffle_epi8(q256_2, filter);
+            q256_3 = _mm256_shuffle_epi8(q256_3, filter);
+            q256_4 = _mm256_shuffle_epi8(q256_4, filter);
+            q256_5 = _mm256_shuffle_epi8(q256_5, filter);
+            q256_6 = _mm256_shuffle_epi8(q256_6, filter);
+            q256_7 = _mm256_shuffle_epi8(q256_7, filter);
+
+            pixelFilter_p = _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_5),
+                    _mm256_add_epi16(p256_4, p256_3));
+            pixelFilter_q = _mm256_add_epi16(_mm256_add_epi16(q256_6, q256_5),
+                    _mm256_add_epi16(q256_4, q256_3));
+
+            pixetFilter_p2p1p0 = _mm256_add_epi16(p256_0,
+                    _mm256_add_epi16(p256_2, p256_1));
+            pixelFilter_p = _mm256_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+            pixetFilter_q2q1q0 = _mm256_add_epi16(q256_0,
+                    _mm256_add_epi16(q256_2, q256_1));
+            pixelFilter_q = _mm256_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+
+            pixelFilter_p = _mm256_add_epi16(eight,
+                    _mm256_add_epi16(pixelFilter_p, pixelFilter_q));
+
+            pixetFilter_p2p1p0 = _mm256_add_epi16(four,
+                    _mm256_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(p256_7, p256_0)), 4);
+
+            flat2_p0 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(q256_7, q256_0)), 4);
+
+            flat2_q0 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_p2p1p0,
+                            _mm256_add_epi16(p256_3, p256_0)), 3);
+
+            flat_p0 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_p2p1p0,
+                            _mm256_add_epi16(q256_3, q256_0)), 3);
+
+            flat_q0 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(p256_7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(q256_7, q256_7);
+
+            sum_p3 = _mm256_add_epi16(p256_3, p256_3);
+
+            sum_q3 = _mm256_add_epi16(q256_3, q256_3);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_p, p256_6);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_6);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_1)), 4);
+
+            flat2_p1 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_1)), 4);
+
+            flat2_q1 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_p2p1p0, p256_2);
+
+            pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_2);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_p2p1p0,
+                            _mm256_add_epi16(sum_p3, p256_1)), 3);
+
+            flat_p1 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_q2q1q0,
+                            _mm256_add_epi16(sum_q3, q256_1)), 3);
+
+            flat_q1 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+            sum_p3 = _mm256_add_epi16(sum_p3, p256_3);
+
+            sum_q3 = _mm256_add_epi16(sum_q3, q256_3);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_5);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_5);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_2)), 4);
+
+            flat2_p2 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_2)), 4);
+
+            flat2_q2 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_1);
+
+            pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_q2q1q0, p256_1);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_p2p1p0,
+                            _mm256_add_epi16(sum_p3, p256_2)), 3);
+
+            flat_p2 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_q2q1q0,
+                            _mm256_add_epi16(sum_q3, q256_2)), 3);
+
+            flat_q2 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_4);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_4);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_3)), 4);
+
+            flat2_p3 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_3)), 4);
+
+            flat2_q3 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_3);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_3);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_4)), 4);
+
+            flat2_p4 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_4)), 4);
+
+            flat2_q4 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_2);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_2);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_5)), 4);
+
+            flat2_p5 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_5)), 4);
+
+            flat2_q5 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_1);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_1);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_6)), 4);
+
+            flat2_p6 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_6)), 4);
+
+            flat2_q6 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+        }
+
+        // wide flat
+        // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+        p2 = _mm_andnot_si128(flat, p2);
+        flat_p2 = _mm_and_si128(flat, flat_p2);
+        p2 = _mm_or_si128(flat_p2, p2);
+
+        p1 = _mm_andnot_si128(flat, ps1);
+        flat_p1 = _mm_and_si128(flat, flat_p1);
+        p1 = _mm_or_si128(flat_p1, p1);
+
+        p0 = _mm_andnot_si128(flat, ps0);
+        flat_p0 = _mm_and_si128(flat, flat_p0);
+        p0 = _mm_or_si128(flat_p0, p0);
+
+        q0 = _mm_andnot_si128(flat, qs0);
+        flat_q0 = _mm_and_si128(flat, flat_q0);
+        q0 = _mm_or_si128(flat_q0, q0);
+
+        q1 = _mm_andnot_si128(flat, qs1);
+        flat_q1 = _mm_and_si128(flat, flat_q1);
+        q1 = _mm_or_si128(flat_q1, q1);
+
+        q2 = _mm_andnot_si128(flat, q2);
+        flat_q2 = _mm_and_si128(flat, flat_q2);
+        q2 = _mm_or_si128(flat_q2, q2);
+
+        p6 = _mm_andnot_si128(flat2, p6);
+        flat2_p6 = _mm_and_si128(flat2, flat2_p6);
+        p6 = _mm_or_si128(flat2_p6, p6);
+        _mm_storeu_si128((__m128i *) (s - 7 * p), p6);
+
+        p5 = _mm_andnot_si128(flat2, p5);
+        flat2_p5 = _mm_and_si128(flat2, flat2_p5);
+        p5 = _mm_or_si128(flat2_p5, p5);
+        _mm_storeu_si128((__m128i *) (s - 6 * p), p5);
+
+        p4 = _mm_andnot_si128(flat2, p4);
+        flat2_p4 = _mm_and_si128(flat2, flat2_p4);
+        p4 = _mm_or_si128(flat2_p4, p4);
+        _mm_storeu_si128((__m128i *) (s - 5 * p), p4);
+
+        p3 = _mm_andnot_si128(flat2, p3);
+        flat2_p3 = _mm_and_si128(flat2, flat2_p3);
+        p3 = _mm_or_si128(flat2_p3, p3);
+        _mm_storeu_si128((__m128i *) (s - 4 * p), p3);
+
+        p2 = _mm_andnot_si128(flat2, p2);
+        flat2_p2 = _mm_and_si128(flat2, flat2_p2);
+        p2 = _mm_or_si128(flat2_p2, p2);
+        _mm_storeu_si128((__m128i *) (s - 3 * p), p2);
+
+        p1 = _mm_andnot_si128(flat2, p1);
+        flat2_p1 = _mm_and_si128(flat2, flat2_p1);
+        p1 = _mm_or_si128(flat2_p1, p1);
+        _mm_storeu_si128((__m128i *) (s - 2 * p), p1);
+
+        p0 = _mm_andnot_si128(flat2, p0);
+        flat2_p0 = _mm_and_si128(flat2, flat2_p0);
+        p0 = _mm_or_si128(flat2_p0, p0);
+        _mm_storeu_si128((__m128i *) (s - 1 * p), p0);
+
+        q0 = _mm_andnot_si128(flat2, q0);
+        flat2_q0 = _mm_and_si128(flat2, flat2_q0);
+        q0 = _mm_or_si128(flat2_q0, q0);
+        _mm_storeu_si128((__m128i *) (s - 0 * p), q0);
+
+        q1 = _mm_andnot_si128(flat2, q1);
+        flat2_q1 = _mm_and_si128(flat2, flat2_q1);
+        q1 = _mm_or_si128(flat2_q1, q1);
+        _mm_storeu_si128((__m128i *) (s + 1 * p), q1);
+
+        q2 = _mm_andnot_si128(flat2, q2);
+        flat2_q2 = _mm_and_si128(flat2, flat2_q2);
+        q2 = _mm_or_si128(flat2_q2, q2);
+        _mm_storeu_si128((__m128i *) (s + 2 * p), q2);
+
+        q3 = _mm_andnot_si128(flat2, q3);
+        flat2_q3 = _mm_and_si128(flat2, flat2_q3);
+        q3 = _mm_or_si128(flat2_q3, q3);
+        _mm_storeu_si128((__m128i *) (s + 3 * p), q3);
+
+        q4 = _mm_andnot_si128(flat2, q4);
+        flat2_q4 = _mm_and_si128(flat2, flat2_q4);
+        q4 = _mm_or_si128(flat2_q4, q4);
+        _mm_storeu_si128((__m128i *) (s + 4 * p), q4);
+
+        q5 = _mm_andnot_si128(flat2, q5);
+        flat2_q5 = _mm_and_si128(flat2, flat2_q5);
+        q5 = _mm_or_si128(flat2_q5, q5);
+        _mm_storeu_si128((__m128i *) (s + 5 * p), q5);
+
+        q6 = _mm_andnot_si128(flat2, q6);
+        flat2_q6 = _mm_and_si128(flat2, flat2_q6);
+        q6 = _mm_or_si128(flat2_q6, q6);
+        _mm_storeu_si128((__m128i *) (s + 6 * p), q6);
+    }
+}
diff --git a/libvpx/libvpx/vpx_dsp/x86/loopfilter_sse2.c b/libvpx/libvpx/vpx_dsp/x86/loopfilter_sse2.c
new file mode 100644
index 0000000..739adf3
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/loopfilter_sse2.c
@@ -0,0 +1,1776 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/emmintrin_compat.h"
+
+static INLINE __m128i abs_diff(__m128i a, __m128i b) {
+  return _mm_or_si128(_mm_subs_epu8(a, b), _mm_subs_epu8(b, a));
+}
+
+// filter_mask and hev_mask
+#define FILTER_HEV_MASK do {                                                   \
+  /* (abs(q1 - q0), abs(p1 - p0) */                                            \
+  __m128i flat = abs_diff(q1p1, q0p0);                                         \
+  /* abs(p1 - q1), abs(p0 - q0) */                                             \
+  const __m128i abs_p1q1p0q0 = abs_diff(p1p0, q1q0);                           \
+  __m128i abs_p0q0, abs_p1q1, work;                                            \
+                                                                               \
+  /* const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); */          \
+  hev = _mm_unpacklo_epi8(_mm_max_epu8(flat, _mm_srli_si128(flat, 8)), zero);  \
+  hev = _mm_cmpgt_epi16(hev, thresh);                                          \
+  hev = _mm_packs_epi16(hev, hev);                                             \
+                                                                               \
+  /* const int8_t mask = filter_mask(*limit, *blimit, */                       \
+  /*                                 p3, p2, p1, p0, q0, q1, q2, q3); */       \
+  abs_p0q0 = _mm_adds_epu8(abs_p1q1p0q0, abs_p1q1p0q0);  /* abs(p0 - q0) * 2 */\
+  abs_p1q1 = _mm_unpackhi_epi8(abs_p1q1p0q0, abs_p1q1p0q0);  /* abs(p1 - q1) */\
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 9);                                      \
+  abs_p1q1 = _mm_packs_epi16(abs_p1q1, abs_p1q1);  /* abs(p1 - q1) / 2 */      \
+  /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 */                                    \
+  mask = _mm_adds_epu8(abs_p0q0, abs_p1q1);                                    \
+  /* abs(p3 - p2), abs(p2 - p1) */                                             \
+  work = abs_diff(p3p2, p2p1);                                                 \
+  flat = _mm_max_epu8(work, flat);                                             \
+  /* abs(q3 - q2), abs(q2 - q1) */                                             \
+  work = abs_diff(q3q2, q2q1);                                                 \
+  flat = _mm_max_epu8(work, flat);                                             \
+  flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));                          \
+  mask = _mm_unpacklo_epi64(mask, flat);                                       \
+  mask = _mm_subs_epu8(mask, limit);                                           \
+  mask = _mm_cmpeq_epi8(mask, zero);                                           \
+  mask = _mm_and_si128(mask, _mm_srli_si128(mask, 8));                         \
+} while (0)
+
+#define FILTER4 do {                                                           \
+  const __m128i t3t4 = _mm_set_epi8(3, 3, 3, 3, 3, 3, 3, 3,                    \
+                                    4, 4, 4, 4, 4, 4, 4, 4);                   \
+  const __m128i t80 = _mm_set1_epi8(0x80);                                     \
+  __m128i filter, filter2filter1, work;                                        \
+                                                                               \
+  ps1ps0 = _mm_xor_si128(p1p0, t80);  /* ^ 0x80 */                             \
+  qs1qs0 = _mm_xor_si128(q1q0, t80);                                           \
+                                                                               \
+  /* int8_t filter = signed_char_clamp(ps1 - qs1) & hev; */                    \
+  work = _mm_subs_epi8(ps1ps0, qs1qs0);                                        \
+  filter = _mm_and_si128(_mm_srli_si128(work, 8), hev);                        \
+  /* filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; */           \
+  filter = _mm_subs_epi8(filter, work);                                        \
+  filter = _mm_subs_epi8(filter, work);                                        \
+  filter = _mm_subs_epi8(filter, work);  /* + 3 * (qs0 - ps0) */               \
+  filter = _mm_and_si128(filter, mask);  /* & mask */                          \
+  filter = _mm_unpacklo_epi64(filter, filter);                                 \
+                                                                               \
+  /* filter1 = signed_char_clamp(filter + 4) >> 3; */                          \
+  /* filter2 = signed_char_clamp(filter + 3) >> 3; */                          \
+  filter2filter1 = _mm_adds_epi8(filter, t3t4);  /* signed_char_clamp */       \
+  filter = _mm_unpackhi_epi8(filter2filter1, filter2filter1);                  \
+  filter2filter1 = _mm_unpacklo_epi8(filter2filter1, filter2filter1);          \
+  filter2filter1 = _mm_srai_epi16(filter2filter1, 11);  /* >> 3 */             \
+  filter = _mm_srai_epi16(filter, 11);  /* >> 3 */                             \
+  filter2filter1 = _mm_packs_epi16(filter2filter1, filter);                    \
+                                                                               \
+  /* filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; */                        \
+  filter = _mm_subs_epi8(filter2filter1, ff);  /* + 1 */                       \
+  filter = _mm_unpacklo_epi8(filter, filter);                                  \
+  filter = _mm_srai_epi16(filter, 9);  /* round */                             \
+  filter = _mm_packs_epi16(filter, filter);                                    \
+  filter = _mm_andnot_si128(hev, filter);                                      \
+                                                                               \
+  hev = _mm_unpackhi_epi64(filter2filter1, filter);                            \
+  filter2filter1 = _mm_unpacklo_epi64(filter2filter1, filter);                 \
+                                                                               \
+  /* signed_char_clamp(qs1 - filter), signed_char_clamp(qs0 - filter1) */      \
+  qs1qs0 = _mm_subs_epi8(qs1qs0, filter2filter1);                              \
+  /* signed_char_clamp(ps1 + filter), signed_char_clamp(ps0 + filter2) */      \
+  ps1ps0 = _mm_adds_epi8(ps1ps0, hev);                                         \
+  qs1qs0 = _mm_xor_si128(qs1qs0, t80);  /* ^ 0x80 */                           \
+  ps1ps0 = _mm_xor_si128(ps1ps0, t80);  /* ^ 0x80 */                           \
+} while (0)
+
+void vpx_lpf_horizontal_4_sse2(uint8_t *s, int p /* pitch */,
+                               const uint8_t *_blimit, const uint8_t *_limit,
+                               const uint8_t *_thresh) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i limit =
+      _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)_blimit),
+                         _mm_loadl_epi64((const __m128i *)_limit));
+  const __m128i thresh =
+      _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero);
+  const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+  __m128i q1p1, q0p0, p3p2, p2p1, p1p0, q3q2, q2q1, q1q0, ps1ps0, qs1qs0;
+  __m128i mask, hev;
+
+  p3p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)),
+                            _mm_loadl_epi64((__m128i *)(s - 4 * p)));
+  q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 1 * p)));
+  q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 0 * p)));
+  q3q2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s + 2 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 3 * p)));
+  p1p0 = _mm_unpacklo_epi64(q0p0, q1p1);
+  p2p1 = _mm_unpacklo_epi64(q1p1, p3p2);
+  q1q0 = _mm_unpackhi_epi64(q0p0, q1p1);
+  q2q1 = _mm_unpacklo_epi64(_mm_srli_si128(q1p1, 8), q3q2);
+
+  FILTER_HEV_MASK;
+  FILTER4;
+
+  _mm_storeh_pi((__m64 *)(s - 2 * p), _mm_castsi128_ps(ps1ps0));  // *op1
+  _mm_storel_epi64((__m128i *)(s - 1 * p), ps1ps0);  // *op0
+  _mm_storel_epi64((__m128i *)(s + 0 * p), qs1qs0);  // *oq0
+  _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(qs1qs0));  // *oq1
+}
+
+void vpx_lpf_vertical_4_sse2(uint8_t *s, int p /* pitch */,
+                             const uint8_t *_blimit, const uint8_t *_limit,
+                             const uint8_t *_thresh) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i limit =
+      _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)_blimit),
+                         _mm_loadl_epi64((const __m128i *)_limit));
+  const __m128i thresh =
+      _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero);
+  const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+  __m128i x0, x1, x2, x3;
+  __m128i q1p1, q0p0, p3p2, p2p1, p1p0, q3q2, q2q1, q1q0, ps1ps0, qs1qs0;
+  __m128i mask, hev;
+
+  // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+  q1q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 0 * p - 4)),
+                           _mm_loadl_epi64((__m128i *)(s + 1 * p - 4)));
+
+  // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+  x1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 2 * p - 4)),
+                         _mm_loadl_epi64((__m128i *)(s + 3 * p - 4)));
+
+  // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+  x2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 4 * p - 4)),
+                         _mm_loadl_epi64((__m128i *)(s + 5 * p - 4)));
+
+  // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+  x3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 6 * p - 4)),
+                         _mm_loadl_epi64((__m128i *)(s + 7 * p - 4)));
+
+  // Transpose 8x8
+  // 00 10 20 30 01 11 21 31  02 12 22 32 03 13 23 33
+  p1p0 = _mm_unpacklo_epi16(q1q0, x1);
+  // 40 50 60 70 41 51 61 71  42 52 62 72 43 53 63 73
+  x0 = _mm_unpacklo_epi16(x2, x3);
+  // 00 10 20 30 40 50 60 70  01 11 21 31 41 51 61 71
+  p3p2 = _mm_unpacklo_epi32(p1p0, x0);
+  // 02 12 22 32 42 52 62 72  03 13 23 33 43 53 63 73
+  p1p0 = _mm_unpackhi_epi32(p1p0, x0);
+  p3p2 = _mm_unpackhi_epi64(p3p2, _mm_slli_si128(p3p2, 8));  // swap lo and high
+  p1p0 = _mm_unpackhi_epi64(p1p0, _mm_slli_si128(p1p0, 8));  // swap lo and high
+
+  // 04 14 24 34 05 15 25 35  06 16 26 36 07 17 27 37
+  q1q0 = _mm_unpackhi_epi16(q1q0, x1);
+  // 44 54 64 74 45 55 65 75  46 56 66 76 47 57 67 77
+  x2 = _mm_unpackhi_epi16(x2, x3);
+  // 06 16 26 36 46 56 66 76  07 17 27 37 47 57 67 77
+  q3q2 = _mm_unpackhi_epi32(q1q0, x2);
+  // 04 14 24 34 44 54 64 74  05 15 25 35 45 55 65 75
+  q1q0 = _mm_unpacklo_epi32(q1q0, x2);
+
+  q0p0 = _mm_unpacklo_epi64(p1p0, q1q0);
+  q1p1 = _mm_unpackhi_epi64(p1p0, q1q0);
+  p1p0 = _mm_unpacklo_epi64(q0p0, q1p1);
+  p2p1 = _mm_unpacklo_epi64(q1p1, p3p2);
+  q2q1 = _mm_unpacklo_epi64(_mm_srli_si128(q1p1, 8), q3q2);
+
+  FILTER_HEV_MASK;
+  FILTER4;
+
+  // Transpose 8x4 to 4x8
+  // qs1qs0: 20 21 22 23 24 25 26 27  30 31 32 33 34 34 36 37
+  // ps1ps0: 10 11 12 13 14 15 16 17  00 01 02 03 04 05 06 07
+  // 00 01 02 03 04 05 06 07  10 11 12 13 14 15 16 17
+  ps1ps0 = _mm_unpackhi_epi64(ps1ps0, _mm_slli_si128(ps1ps0, 8));
+  // 10 30 11 31 12 32 13 33  14 34 15 35 16 36 17 37
+  x0 = _mm_unpackhi_epi8(ps1ps0, qs1qs0);
+  // 00 20 01 21 02 22 03 23  04 24 05 25 06 26 07 27
+  ps1ps0 = _mm_unpacklo_epi8(ps1ps0, qs1qs0);
+  // 04 14 24 34 05 15 25 35  06 16 26 36 07 17 27 37
+  qs1qs0 = _mm_unpackhi_epi8(ps1ps0, x0);
+  // 00 10 20 30 01 11 21 31  02 12 22 32 03 13 23 33
+  ps1ps0 = _mm_unpacklo_epi8(ps1ps0, x0);
+
+  *(int *)(s + 0 * p - 2) = _mm_cvtsi128_si32(ps1ps0);
+  ps1ps0 = _mm_srli_si128(ps1ps0, 4);
+  *(int *)(s + 1 * p - 2) = _mm_cvtsi128_si32(ps1ps0);
+  ps1ps0 = _mm_srli_si128(ps1ps0, 4);
+  *(int *)(s + 2 * p - 2) = _mm_cvtsi128_si32(ps1ps0);
+  ps1ps0 = _mm_srli_si128(ps1ps0, 4);
+  *(int *)(s + 3 * p - 2) = _mm_cvtsi128_si32(ps1ps0);
+
+  *(int *)(s + 4 * p - 2) = _mm_cvtsi128_si32(qs1qs0);
+  qs1qs0 = _mm_srli_si128(qs1qs0, 4);
+  *(int *)(s + 5 * p - 2) = _mm_cvtsi128_si32(qs1qs0);
+  qs1qs0 = _mm_srli_si128(qs1qs0, 4);
+  *(int *)(s + 6 * p - 2) = _mm_cvtsi128_si32(qs1qs0);
+  qs1qs0 = _mm_srli_si128(qs1qs0, 4);
+  *(int *)(s + 7 * p - 2) = _mm_cvtsi128_si32(qs1qs0);
+}
+
+void vpx_lpf_horizontal_edge_8_sse2(unsigned char *s, int p,
+                                    const unsigned char *_blimit,
+                                    const unsigned char *_limit,
+                                    const unsigned char *_thresh) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+  __m128i mask, hev, flat, flat2;
+  __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
+  __m128i abs_p1p0;
+
+  q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
+  q4p4 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q4p4),
+                                       (__m64 *)(s + 4 * p)));
+  q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
+  q3p3 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q3p3),
+                                       (__m64 *)(s + 3 * p)));
+  q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+  q2p2 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q2p2),
+                                       (__m64 *)(s + 2 * p)));
+  q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  q1p1 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q1p1),
+                                       (__m64 *)(s + 1 * p)));
+  p1q1 = _mm_shuffle_epi32(q1p1, 78);
+  q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  q0p0 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q0p0),
+                                       (__m64 *)(s - 0 * p)));
+  p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+  {
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
+    abs_p1p0 = abs_diff(q1p1, q0p0);
+    abs_q1q0 =  _mm_srli_si128(abs_p1p0, 8);
+    fe = _mm_set1_epi8(0xfe);
+    ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    abs_p0q0 = abs_diff(q0p0, p0q0);
+    abs_p1q1 = abs_diff(q1p1, p1q1);
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = _mm_max_epu8(abs_diff(q2p2, q1p1),
+                        abs_diff(q3p3, q2p2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i t1 = _mm_set1_epi16(0x1);
+    __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
+    __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
+    __m128i qs0 = _mm_xor_si128(p0q0, t80);
+    __m128i qs1 = _mm_xor_si128(p1q1, t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+    __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+    __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
+
+    filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, qs0ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (vpx_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    filter1 = _mm_unpacklo_epi8(zero, filter1);
+    filter1 = _mm_srai_epi16(filter1, 0xB);
+    filter2 = _mm_unpacklo_epi8(zero, filter2);
+    filter2 = _mm_srai_epi16(filter2, 0xB);
+
+    // Filter1 >> 3
+    filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
+    qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
+
+    // filt >> 1
+    filt = _mm_adds_epi16(filter1, t1);
+    filt = _mm_srai_epi16(filt, 1);
+    filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8),
+                            filt);
+    filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
+    qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
+    // loopfilter done
+
+    {
+      __m128i work;
+      flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0));
+      flat = _mm_max_epu8(abs_p1p0, flat);
+      flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+      flat = _mm_subs_epu8(flat, one);
+      flat = _mm_cmpeq_epi8(flat, zero);
+      flat = _mm_and_si128(flat, mask);
+
+      q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
+      q5p5 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q5p5),
+                                           (__m64 *)(s + 5 * p)));
+
+      q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
+      q6p6 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q6p6),
+                                           (__m64 *)(s + 6 * p)));
+      flat2 = _mm_max_epu8(abs_diff(q4p4, q0p0), abs_diff(q5p5, q0p0));
+
+      q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
+      q7p7 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q7p7),
+                                           (__m64 *)(s + 7 * p)));
+      work = _mm_max_epu8(abs_diff(q6p6, q0p0), abs_diff(q7p7, q0p0));
+      flat2 = _mm_max_epu8(work, flat2);
+      flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
+      flat2 = _mm_subs_epu8(flat2, one);
+      flat2 = _mm_cmpeq_epi8(flat2, zero);
+      flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+    }
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // flat and wide flat calculations
+    {
+      const __m128i eight = _mm_set1_epi16(8);
+      const __m128i four = _mm_set1_epi16(4);
+      __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+      __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+      __m128i pixelFilter_p, pixelFilter_q;
+      __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+      __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
+
+      p7_16 = _mm_unpacklo_epi8(q7p7, zero);;
+      p6_16 = _mm_unpacklo_epi8(q6p6, zero);
+      p5_16 = _mm_unpacklo_epi8(q5p5, zero);
+      p4_16 = _mm_unpacklo_epi8(q4p4, zero);
+      p3_16 = _mm_unpacklo_epi8(q3p3, zero);
+      p2_16 = _mm_unpacklo_epi8(q2p2, zero);
+      p1_16 = _mm_unpacklo_epi8(q1p1, zero);
+      p0_16 = _mm_unpacklo_epi8(q0p0, zero);
+      q0_16 = _mm_unpackhi_epi8(q0p0, zero);
+      q1_16 = _mm_unpackhi_epi8(q1p1, zero);
+      q2_16 = _mm_unpackhi_epi8(q2p2, zero);
+      q3_16 = _mm_unpackhi_epi8(q3p3, zero);
+      q4_16 = _mm_unpackhi_epi8(q4p4, zero);
+      q5_16 = _mm_unpackhi_epi8(q5p5, zero);
+      q6_16 = _mm_unpackhi_epi8(q6p6, zero);
+      q7_16 = _mm_unpackhi_epi8(q7p7, zero);
+
+      pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
+                                    _mm_add_epi16(p4_16, p3_16));
+      pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
+                                    _mm_add_epi16(q4_16, q3_16));
+
+      pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
+      pixelFilter_p =  _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+      pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16));
+      pixelFilter_q =  _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+      pixelFilter_p =  _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p,
+                                                         pixelFilter_q));
+      pixetFilter_p2p1p0 =   _mm_add_epi16(four,
+                                           _mm_add_epi16(pixetFilter_p2p1p0,
+                                                         pixetFilter_q2q1q0));
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                           _mm_add_epi16(p7_16, p0_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                           _mm_add_epi16(q7_16, q0_16)), 4);
+      flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                           _mm_add_epi16(p3_16, p0_16)), 3);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                           _mm_add_epi16(q3_16, q0_16)), 3);
+
+      flat_q0p0 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(p7_16, p7_16);
+      sum_q7 = _mm_add_epi16(q7_16, q7_16);
+      sum_p3 = _mm_add_epi16(p3_16, p3_16);
+      sum_q3 = _mm_add_epi16(q3_16, q3_16);
+
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p1_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q1_16)), 4);
+      flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+      pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
+      pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                             _mm_add_epi16(sum_p3, p1_16)), 3);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+                             _mm_add_epi16(sum_q3, q1_16)), 3);
+      flat_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      sum_p3 = _mm_add_epi16(sum_p3, p3_16);
+      sum_q3 = _mm_add_epi16(sum_q3, q3_16);
+
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p2_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q2_16)), 4);
+      flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+      pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
+      pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
+
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                           _mm_add_epi16(sum_p3, p2_16)), 3);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+                                           _mm_add_epi16(sum_q3, q2_16)), 3);
+      flat_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p3_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q3_16)), 4);
+      flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p4_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q4_16)), 4);
+      flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p5_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q5_16)), 4);
+      flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p6_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q6_16)), 4);
+      flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
+    }
+    // wide flat
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    flat = _mm_shuffle_epi32(flat, 68);
+    flat2 = _mm_shuffle_epi32(flat2, 68);
+
+    q2p2 = _mm_andnot_si128(flat, q2p2);
+    flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
+    q2p2 = _mm_or_si128(q2p2, flat_q2p2);
+
+    qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+    flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
+    q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
+
+    qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+    flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
+    q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
+
+    q6p6 = _mm_andnot_si128(flat2, q6p6);
+    flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
+    q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
+    _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6);
+    _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6));
+
+    q5p5 = _mm_andnot_si128(flat2, q5p5);
+    flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
+    q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
+    _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5);
+    _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5));
+
+    q4p4 = _mm_andnot_si128(flat2, q4p4);
+    flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
+    q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
+    _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4);
+    _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4));
+
+    q3p3 = _mm_andnot_si128(flat2, q3p3);
+    flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
+    q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
+    _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3);
+    _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3));
+
+    q2p2 = _mm_andnot_si128(flat2, q2p2);
+    flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
+    q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
+    _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2);
+    _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2));
+
+    q1p1 = _mm_andnot_si128(flat2, q1p1);
+    flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
+    q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
+    _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1);
+    _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1));
+
+    q0p0 = _mm_andnot_si128(flat2, q0p0);
+    flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
+    q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
+    _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0);
+    _mm_storeh_pi((__m64 *)(s - 0 * p),  _mm_castsi128_ps(q0p0));
+  }
+}
+
+static INLINE __m128i filter_add2_sub2(const __m128i *const total,
+                                       const __m128i *const a1,
+                                       const __m128i *const a2,
+                                       const __m128i *const s1,
+                                       const __m128i *const s2) {
+  __m128i x = _mm_add_epi16(*a1, *total);
+  x = _mm_add_epi16(_mm_sub_epi16(x, _mm_add_epi16(*s1, *s2)), *a2);
+  return x;
+}
+
+static INLINE __m128i filter8_mask(const __m128i *const flat,
+                                   const __m128i *const other_filt,
+                                   const __m128i *const f8_lo,
+                                   const __m128i *const f8_hi) {
+  const __m128i f8 = _mm_packus_epi16(_mm_srli_epi16(*f8_lo, 3),
+                                      _mm_srli_epi16(*f8_hi, 3));
+  const __m128i result = _mm_and_si128(*flat, f8);
+  return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
+}
+
+static INLINE __m128i filter16_mask(const __m128i *const flat,
+                                    const __m128i *const other_filt,
+                                    const __m128i *const f_lo,
+                                    const __m128i *const f_hi) {
+  const __m128i f = _mm_packus_epi16(_mm_srli_epi16(*f_lo, 4),
+                                     _mm_srli_epi16(*f_hi, 4));
+  const __m128i result = _mm_and_si128(*flat, f);
+  return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
+}
+
+void vpx_lpf_horizontal_edge_16_sse2(unsigned char *s, int p,
+                                     const unsigned char *_blimit,
+                                     const unsigned char *_limit,
+                                     const unsigned char *_thresh) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+  __m128i mask, hev, flat, flat2;
+  __m128i p7, p6, p5;
+  __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
+  __m128i q5, q6, q7;
+
+  __m128i op2, op1, op0, oq0, oq1, oq2;
+
+  __m128i max_abs_p1p0q1q0;
+
+  p7 = _mm_loadu_si128((__m128i *)(s - 8 * p));
+  p6 = _mm_loadu_si128((__m128i *)(s - 7 * p));
+  p5 = _mm_loadu_si128((__m128i *)(s - 6 * p));
+  p4 = _mm_loadu_si128((__m128i *)(s - 5 * p));
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  q4 = _mm_loadu_si128((__m128i *)(s + 4 * p));
+  q5 = _mm_loadu_si128((__m128i *)(s + 5 * p));
+  q6 = _mm_loadu_si128((__m128i *)(s + 6 * p));
+  q7 = _mm_loadu_si128((__m128i *)(s + 7 * p));
+
+  {
+    const __m128i abs_p1p0 = abs_diff(p1, p0);
+    const __m128i abs_q1q0 = abs_diff(q1, q0);
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+    __m128i abs_p0q0 = abs_diff(p0, q0);
+    __m128i abs_p1q1 = abs_diff(p1, q1);
+    __m128i work;
+    max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(max_abs_p1p0q1q0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(p3, p2));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(abs_diff(q2, q1), abs_diff(q3, q2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  {
+    __m128i work;
+    work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0));
+    flat = _mm_max_epu8(work, max_abs_p1p0q1q0);
+    work = _mm_max_epu8(abs_diff(p3, p0), abs_diff(q3, q0));
+    flat = _mm_max_epu8(work, flat);
+    work = _mm_max_epu8(abs_diff(p4, p0), abs_diff(q4, q0));
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+    flat2 = _mm_max_epu8(abs_diff(p5, p0), abs_diff(q5, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    work = _mm_max_epu8(abs_diff(p6, p0), abs_diff(q6, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    work = _mm_max_epu8(abs_diff(p7, p0), abs_diff(q7, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    flat2 = _mm_subs_epu8(flat2, one);
+    flat2 = _mm_cmpeq_epi8(flat2, zero);
+    flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+  }
+
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // filter4
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i te0 = _mm_set1_epi8(0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+    const __m128i ff = _mm_cmpeq_epi8(t4, t4);
+
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    op1 = _mm_xor_si128(p1, t80);
+    op0 = _mm_xor_si128(p0, t80);
+    oq0 = _mm_xor_si128(q0, t80);
+    oq1 = _mm_xor_si128(q1, t80);
+
+    hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev);
+
+    work_a = _mm_subs_epi8(oq0, op0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (vpx_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80);
+
+    // Filter2 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80);
+    oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80);
+    // loopfilter done
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // filter8
+    {
+      const __m128i four = _mm_set1_epi16(4);
+      const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+      const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+      const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+      const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+      const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+      const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+      const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+      const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+
+      const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+      const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+      const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+      const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+      const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+      const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+      const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+      const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+      __m128i f8_lo, f8_hi;
+
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, four),
+                            _mm_add_epi16(p3_lo, p2_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f8_lo),
+                            _mm_add_epi16(p2_lo, p1_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo);
+
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, four),
+                            _mm_add_epi16(p3_hi, p2_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f8_hi),
+                            _mm_add_epi16(p2_hi, p1_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi);
+
+      op2 = filter8_mask(&flat, &p2, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q1_lo, &p1_lo, &p2_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q1_hi, &p1_hi, &p2_hi, &p3_hi);
+      op1 = filter8_mask(&flat, &op1, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q2_lo, &p0_lo, &p1_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q2_hi, &p0_hi, &p1_hi, &p3_hi);
+      op0 = filter8_mask(&flat, &op0, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q0_lo, &p0_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q0_hi, &p0_hi, &p3_hi);
+      oq0 = filter8_mask(&flat, &oq0, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q1_lo, &q0_lo, &p2_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q1_hi, &q0_hi, &p2_hi);
+      oq1 = filter8_mask(&flat, &oq1, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q2_lo, &q1_lo, &p1_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q2_hi, &q1_hi, &p1_hi);
+      oq2 = filter8_mask(&flat, &q2, &f8_lo, &f8_hi);
+    }
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // wide flat calculations
+    {
+      const __m128i eight = _mm_set1_epi16(8);
+      const __m128i p7_lo = _mm_unpacklo_epi8(p7, zero);
+      const __m128i p6_lo = _mm_unpacklo_epi8(p6, zero);
+      const __m128i p5_lo = _mm_unpacklo_epi8(p5, zero);
+      const __m128i p4_lo = _mm_unpacklo_epi8(p4, zero);
+      const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+      const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+      const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+      const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+      const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+      const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+      const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+      const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+      const __m128i q4_lo = _mm_unpacklo_epi8(q4, zero);
+      const __m128i q5_lo = _mm_unpacklo_epi8(q5, zero);
+      const __m128i q6_lo = _mm_unpacklo_epi8(q6, zero);
+      const __m128i q7_lo = _mm_unpacklo_epi8(q7, zero);
+
+      const __m128i p7_hi = _mm_unpackhi_epi8(p7, zero);
+      const __m128i p6_hi = _mm_unpackhi_epi8(p6, zero);
+      const __m128i p5_hi = _mm_unpackhi_epi8(p5, zero);
+      const __m128i p4_hi = _mm_unpackhi_epi8(p4, zero);
+      const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+      const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+      const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+      const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+      const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+      const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+      const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+      const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+      const __m128i q4_hi = _mm_unpackhi_epi8(q4, zero);
+      const __m128i q5_hi = _mm_unpackhi_epi8(q5, zero);
+      const __m128i q6_hi = _mm_unpackhi_epi8(q6, zero);
+      const __m128i q7_hi = _mm_unpackhi_epi8(q7, zero);
+
+      __m128i f_lo;
+      __m128i f_hi;
+
+      f_lo = _mm_sub_epi16(_mm_slli_epi16(p7_lo, 3), p7_lo);  // p7 * 7
+      f_lo = _mm_add_epi16(_mm_slli_epi16(p6_lo, 1),
+                           _mm_add_epi16(p4_lo, f_lo));
+      f_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f_lo),
+                           _mm_add_epi16(p2_lo, p1_lo));
+      f_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f_lo);
+      f_lo = _mm_add_epi16(_mm_add_epi16(p5_lo, eight), f_lo);
+
+      f_hi = _mm_sub_epi16(_mm_slli_epi16(p7_hi, 3), p7_hi);  // p7 * 7
+      f_hi = _mm_add_epi16(_mm_slli_epi16(p6_hi, 1),
+                           _mm_add_epi16(p4_hi, f_hi));
+      f_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f_hi),
+                           _mm_add_epi16(p2_hi, p1_hi));
+      f_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f_hi);
+      f_hi = _mm_add_epi16(_mm_add_epi16(p5_hi, eight), f_hi);
+
+      p6 = filter16_mask(&flat2, &p6, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 7 * p), p6);
+
+      f_lo = filter_add2_sub2(&f_lo, &q1_lo, &p5_lo, &p6_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q1_hi, &p5_hi, &p6_hi, &p7_hi);
+      p5 = filter16_mask(&flat2, &p5, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 6 * p), p5);
+
+      f_lo = filter_add2_sub2(&f_lo, &q2_lo, &p4_lo, &p5_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q2_hi, &p4_hi, &p5_hi, &p7_hi);
+      p4 = filter16_mask(&flat2, &p4, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 5 * p), p4);
+
+      f_lo = filter_add2_sub2(&f_lo, &q3_lo, &p3_lo, &p4_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q3_hi, &p3_hi, &p4_hi, &p7_hi);
+      p3 = filter16_mask(&flat2, &p3, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 4 * p), p3);
+
+      f_lo = filter_add2_sub2(&f_lo, &q4_lo, &p2_lo, &p3_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q4_hi, &p2_hi, &p3_hi, &p7_hi);
+      op2 = filter16_mask(&flat2, &op2, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 3 * p), op2);
+
+      f_lo = filter_add2_sub2(&f_lo, &q5_lo, &p1_lo, &p2_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q5_hi, &p1_hi, &p2_hi, &p7_hi);
+      op1 = filter16_mask(&flat2, &op1, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+
+      f_lo = filter_add2_sub2(&f_lo, &q6_lo, &p0_lo, &p1_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q6_hi, &p0_hi, &p1_hi, &p7_hi);
+      op0 = filter16_mask(&flat2, &op0, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q0_lo, &p0_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q0_hi, &p0_hi, &p7_hi);
+      oq0 = filter16_mask(&flat2, &oq0, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q1_lo, &p6_lo, &q0_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q1_hi, &p6_hi, &q0_hi);
+      oq1 = filter16_mask(&flat2, &oq1, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q2_lo, &p5_lo, &q1_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q2_hi, &p5_hi, &q1_hi);
+      oq2 = filter16_mask(&flat2, &oq2, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 2 * p), oq2);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q3_lo, &p4_lo, &q2_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q3_hi, &p4_hi, &q2_hi);
+      q3 = filter16_mask(&flat2, &q3, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 3 * p), q3);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q4_lo, &p3_lo, &q3_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q4_hi, &p3_hi, &q3_hi);
+      q4 = filter16_mask(&flat2, &q4, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 4 * p), q4);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q5_lo, &p2_lo, &q4_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q5_hi, &p2_hi, &q4_hi);
+      q5 = filter16_mask(&flat2, &q5, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 5 * p), q5);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q6_lo, &p1_lo, &q5_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q6_hi, &p1_hi, &q5_hi);
+      q6 = filter16_mask(&flat2, &q6, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 6 * p), q6);
+    }
+    // wide flat
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  }
+}
+
+void vpx_lpf_horizontal_8_sse2(unsigned char *s, int p,
+                               const unsigned char *_blimit,
+                               const unsigned char *_limit,
+                               const unsigned char *_thresh) {
+  DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+  __m128i mask, hev, flat;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+  __m128i q3p3, q2p2, q1p1, q0p0, p1q1, p0q0;
+
+  q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 3 * p)));
+  q2p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 2 * p)));
+  q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 1 * p)));
+  q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+                            _mm_loadl_epi64((__m128i *)(s - 0 * p)));
+  p1q1 = _mm_shuffle_epi32(q1p1, 78);
+  p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+  {
+    // filter_mask and hev_mask
+    const __m128i one = _mm_set1_epi8(1);
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(fe, fe);
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+    abs_p1p0 = abs_diff(q1p1, q0p0);
+    abs_q1q0 =  _mm_srli_si128(abs_p1p0, 8);
+
+    abs_p0q0 = abs_diff(q0p0, p0q0);
+    abs_p1q1 = abs_diff(q1p1, p1q1);
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = _mm_max_epu8(abs_diff(q2p2, q1p1),
+                        abs_diff(q3p3, q2p2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+
+    // flat_mask4
+
+    flat = _mm_max_epu8(abs_diff(q2p2, q0p0),
+                        abs_diff(q3p3, q0p0));
+    flat = _mm_max_epu8(abs_p1p0, flat);
+    flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+  }
+
+  {
+    const __m128i four = _mm_set1_epi16(4);
+    unsigned char *src = s;
+    {
+      __m128i workp_a, workp_b, workp_shft;
+      p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
+      p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
+      p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
+      p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
+      q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
+      q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
+      q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
+      q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
+
+      workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+      workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op2[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op1[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op0[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq0[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq1[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq2[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+    }
+  }
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i ps1 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+                                      t80);
+    const __m128i ps0 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+                                      t80);
+    const __m128i qs0 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 0 * p)),
+                                      t80);
+    const __m128i qs1 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 1 * p)),
+                                      t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (vpx_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    filter1 = _mm_unpacklo_epi8(zero, filter1);
+    filter1 = _mm_srai_epi16(filter1, 11);
+    filter1 = _mm_packs_epi16(filter1, filter1);
+
+    // Filter2 >> 3
+    filter2 = _mm_unpacklo_epi8(zero, filter2);
+    filter2 = _mm_srai_epi16(filter2, 11);
+    filter2 = _mm_packs_epi16(filter2, zero);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    filt = _mm_unpacklo_epi8(zero, filt);
+    filt = _mm_srai_epi16(filt, 9);
+    filt = _mm_packs_epi16(filt, zero);
+
+    filt = _mm_andnot_si128(hev, filt);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+    q0 = _mm_loadl_epi64((__m128i *)flat_oq0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q0 = _mm_and_si128(flat, q0);
+    q0 = _mm_or_si128(work_a, q0);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    q1 = _mm_loadl_epi64((__m128i *)flat_oq1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q1 = _mm_and_si128(flat, q1);
+    q1 = _mm_or_si128(work_a, q1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+    q2 = _mm_loadl_epi64((__m128i *)flat_oq2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q2 = _mm_and_si128(flat, q2);
+    q2 = _mm_or_si128(work_a, q2);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+    p0 = _mm_loadl_epi64((__m128i *)flat_op0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p0 = _mm_and_si128(flat, p0);
+    p0 = _mm_or_si128(work_a, p0);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+    p1 = _mm_loadl_epi64((__m128i *)flat_op1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p1 = _mm_and_si128(flat, p1);
+    p1 = _mm_or_si128(work_a, p1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+    p2 = _mm_loadl_epi64((__m128i *)flat_op2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p2 = _mm_and_si128(flat, p2);
+    p2 = _mm_or_si128(work_a, p2);
+
+    _mm_storel_epi64((__m128i *)(s - 3 * p), p2);
+    _mm_storel_epi64((__m128i *)(s - 2 * p), p1);
+    _mm_storel_epi64((__m128i *)(s - 1 * p), p0);
+    _mm_storel_epi64((__m128i *)(s + 0 * p), q0);
+    _mm_storel_epi64((__m128i *)(s + 1 * p), q1);
+    _mm_storel_epi64((__m128i *)(s + 2 * p), q2);
+  }
+}
+
+void vpx_lpf_horizontal_8_dual_sse2(uint8_t *s, int p,
+                                    const uint8_t *_blimit0,
+                                    const uint8_t *_limit0,
+                                    const uint8_t *_thresh0,
+                                    const uint8_t *_blimit1,
+                                    const uint8_t *_limit1,
+                                    const uint8_t *_thresh1) {
+  DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i blimit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0),
+                         _mm_load_si128((const __m128i *)_blimit1));
+  const __m128i limit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0),
+                         _mm_load_si128((const __m128i *)_limit1));
+  const __m128i thresh =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0),
+                         _mm_load_si128((const __m128i *)_thresh1));
+
+  __m128i mask, hev, flat;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  {
+    const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
+                                          _mm_subs_epu8(p0, p1));
+    const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
+                                          _mm_subs_epu8(q0, q1));
+    const __m128i one = _mm_set1_epi8(1);
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
+                                    _mm_subs_epu8(q0, p0));
+    __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
+                                    _mm_subs_epu8(q1, p1));
+    __m128i work;
+
+    // filter_mask and hev_mask
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(flat, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1),
+                                     _mm_subs_epu8(p1, p2)),
+                         _mm_or_si128(_mm_subs_epu8(p3, p2),
+                                      _mm_subs_epu8(p2, p3)));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1),
+                                     _mm_subs_epu8(q1, q2)),
+                         _mm_or_si128(_mm_subs_epu8(q3, q2),
+                                      _mm_subs_epu8(q2, q3)));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+
+    // flat_mask4
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p0),
+                                     _mm_subs_epu8(p0, p2)),
+                         _mm_or_si128(_mm_subs_epu8(q2, q0),
+                                      _mm_subs_epu8(q0, q2)));
+    flat = _mm_max_epu8(work, flat);
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p3, p0),
+                                     _mm_subs_epu8(p0, p3)),
+                         _mm_or_si128(_mm_subs_epu8(q3, q0),
+                                      _mm_subs_epu8(q0, q3)));
+    flat = _mm_max_epu8(work, flat);
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+  }
+  {
+    const __m128i four = _mm_set1_epi16(4);
+    unsigned char *src = s;
+    int i = 0;
+
+    do {
+      __m128i workp_a, workp_b, workp_shft;
+      p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
+      p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
+      p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
+      p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
+      q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
+      q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
+      q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
+      q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
+
+      workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+      workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op2[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op1[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op0[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq0[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq1[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq2[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      src += 8;
+    } while (++i < 2);
+  }
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i te0 = _mm_set1_epi8(0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+
+    const __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)),
+                                      t80);
+    const __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)),
+                                      t80);
+    const __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)),
+                                      t80);
+    const __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)),
+                                      t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (vpx_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+
+    // Filter2 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+
+    filt = _mm_andnot_si128(hev, filt);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+    q0 = _mm_load_si128((__m128i *)flat_oq0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q0 = _mm_and_si128(flat, q0);
+    q0 = _mm_or_si128(work_a, q0);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    q1 = _mm_load_si128((__m128i *)flat_oq1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q1 = _mm_and_si128(flat, q1);
+    q1 = _mm_or_si128(work_a, q1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+    q2 = _mm_load_si128((__m128i *)flat_oq2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q2 = _mm_and_si128(flat, q2);
+    q2 = _mm_or_si128(work_a, q2);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+    p0 = _mm_load_si128((__m128i *)flat_op0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p0 = _mm_and_si128(flat, p0);
+    p0 = _mm_or_si128(work_a, p0);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+    p1 = _mm_load_si128((__m128i *)flat_op1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p1 = _mm_and_si128(flat, p1);
+    p1 = _mm_or_si128(work_a, p1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+    p2 = _mm_load_si128((__m128i *)flat_op2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p2 = _mm_and_si128(flat, p2);
+    p2 = _mm_or_si128(work_a, p2);
+
+    _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+    _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+    _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+    _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+    _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+    _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+  }
+}
+
+void vpx_lpf_horizontal_4_dual_sse2(unsigned char *s, int p,
+                                    const unsigned char *_blimit0,
+                                    const unsigned char *_limit0,
+                                    const unsigned char *_thresh0,
+                                    const unsigned char *_blimit1,
+                                    const unsigned char *_limit1,
+                                    const unsigned char *_thresh1) {
+  const __m128i blimit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0),
+                         _mm_load_si128((const __m128i *)_blimit1));
+  const __m128i limit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0),
+                         _mm_load_si128((const __m128i *)_limit1));
+  const __m128i thresh =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0),
+                         _mm_load_si128((const __m128i *)_thresh1));
+  const __m128i zero = _mm_set1_epi16(0);
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+  __m128i mask, hev, flat;
+
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+
+  // filter_mask and hev_mask
+  {
+    const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
+                                          _mm_subs_epu8(p0, p1));
+    const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
+                                          _mm_subs_epu8(q0, q1));
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
+                                    _mm_subs_epu8(q0, p0));
+    __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
+                                    _mm_subs_epu8(q1, p1));
+    __m128i work;
+
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(flat, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1),
+                                     _mm_subs_epu8(p1, p2)),
+                         _mm_or_si128(_mm_subs_epu8(p3, p2),
+                                      _mm_subs_epu8(p2, p3)));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1),
+                                     _mm_subs_epu8(q1, q2)),
+                         _mm_or_si128(_mm_subs_epu8(q3, q2),
+                                      _mm_subs_epu8(q2, q3)));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  // filter4
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i te0 = _mm_set1_epi8(0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+
+    const __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)),
+                                      t80);
+    const __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)),
+                                      t80);
+    const __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)),
+                                      t80);
+    const __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)),
+                                      t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (vpx_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+
+    // Filter2 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+
+    filt = _mm_andnot_si128(hev, filt);
+
+    q0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+    q1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    p0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+    p1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+
+    _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+    _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+    _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+    _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+  }
+}
+
+static INLINE void transpose8x16(unsigned char *in0, unsigned char *in1,
+                                 int in_p, unsigned char *out, int out_p) {
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i x8, x9, x10, x11, x12, x13, x14, x15;
+
+  // 2-way interleave w/hoisting of unpacks
+  x0 = _mm_loadl_epi64((__m128i *)in0);  // 1
+  x1 = _mm_loadl_epi64((__m128i *)(in0 + in_p));  // 3
+  x0 = _mm_unpacklo_epi8(x0, x1);  // 1
+
+  x2 = _mm_loadl_epi64((__m128i *)(in0 + 2 * in_p));  // 5
+  x3 = _mm_loadl_epi64((__m128i *)(in0 + 3*in_p));  // 7
+  x1 = _mm_unpacklo_epi8(x2, x3);  // 2
+
+  x4 = _mm_loadl_epi64((__m128i *)(in0 + 4*in_p));  // 9
+  x5 = _mm_loadl_epi64((__m128i *)(in0 + 5*in_p));  // 11
+  x2 = _mm_unpacklo_epi8(x4, x5);  // 3
+
+  x6 = _mm_loadl_epi64((__m128i *)(in0 + 6*in_p));  // 13
+  x7 = _mm_loadl_epi64((__m128i *)(in0 + 7*in_p));  // 15
+  x3 = _mm_unpacklo_epi8(x6, x7);  // 4
+  x4 = _mm_unpacklo_epi16(x0, x1);  // 9
+
+  x8 = _mm_loadl_epi64((__m128i *)in1);  // 2
+  x9 = _mm_loadl_epi64((__m128i *)(in1 + in_p));  // 4
+  x8 = _mm_unpacklo_epi8(x8, x9);  // 5
+  x5 = _mm_unpacklo_epi16(x2, x3);  // 10
+
+  x10 = _mm_loadl_epi64((__m128i *)(in1 + 2 * in_p));  // 6
+  x11 = _mm_loadl_epi64((__m128i *)(in1 + 3*in_p));  // 8
+  x9 = _mm_unpacklo_epi8(x10, x11);  // 6
+
+  x12 = _mm_loadl_epi64((__m128i *)(in1 + 4*in_p));  // 10
+  x13 = _mm_loadl_epi64((__m128i *)(in1 + 5*in_p));  // 12
+  x10 = _mm_unpacklo_epi8(x12, x13);  // 7
+  x12 = _mm_unpacklo_epi16(x8, x9);  // 11
+
+  x14 = _mm_loadl_epi64((__m128i *)(in1 + 6*in_p));  // 14
+  x15 = _mm_loadl_epi64((__m128i *)(in1 + 7*in_p));  // 16
+  x11 = _mm_unpacklo_epi8(x14, x15);  // 8
+  x13 = _mm_unpacklo_epi16(x10, x11);  // 12
+
+  x6 = _mm_unpacklo_epi32(x4, x5);  // 13
+  x7 = _mm_unpackhi_epi32(x4, x5);  // 14
+  x14 = _mm_unpacklo_epi32(x12, x13);  // 15
+  x15 = _mm_unpackhi_epi32(x12, x13);  // 16
+
+  // Store first 4-line result
+  _mm_storeu_si128((__m128i *)out, _mm_unpacklo_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + out_p), _mm_unpackhi_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 2 * out_p), _mm_unpacklo_epi64(x7, x15));
+  _mm_storeu_si128((__m128i *)(out + 3 * out_p), _mm_unpackhi_epi64(x7, x15));
+
+  x4 = _mm_unpackhi_epi16(x0, x1);
+  x5 = _mm_unpackhi_epi16(x2, x3);
+  x12 = _mm_unpackhi_epi16(x8, x9);
+  x13 = _mm_unpackhi_epi16(x10, x11);
+
+  x6 = _mm_unpacklo_epi32(x4, x5);
+  x7 = _mm_unpackhi_epi32(x4, x5);
+  x14 = _mm_unpacklo_epi32(x12, x13);
+  x15 = _mm_unpackhi_epi32(x12, x13);
+
+  // Store second 4-line result
+  _mm_storeu_si128((__m128i *)(out + 4 * out_p), _mm_unpacklo_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 5 * out_p), _mm_unpackhi_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 6 * out_p), _mm_unpacklo_epi64(x7, x15));
+  _mm_storeu_si128((__m128i *)(out + 7 * out_p), _mm_unpackhi_epi64(x7, x15));
+}
+
+static INLINE void transpose(unsigned char *src[], int in_p,
+                             unsigned char *dst[], int out_p,
+                             int num_8x8_to_transpose) {
+  int idx8x8 = 0;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  do {
+    unsigned char *in = src[idx8x8];
+    unsigned char *out = dst[idx8x8];
+
+    x0 = _mm_loadl_epi64((__m128i *)(in + 0*in_p));  // 00 01 02 03 04 05 06 07
+    x1 = _mm_loadl_epi64((__m128i *)(in + 1*in_p));  // 10 11 12 13 14 15 16 17
+    // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+    x0 = _mm_unpacklo_epi8(x0, x1);
+
+    x2 = _mm_loadl_epi64((__m128i *)(in + 2*in_p));  // 20 21 22 23 24 25 26 27
+    x3 = _mm_loadl_epi64((__m128i *)(in + 3*in_p));  // 30 31 32 33 34 35 36 37
+    // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+    x1 = _mm_unpacklo_epi8(x2, x3);
+
+    x4 = _mm_loadl_epi64((__m128i *)(in + 4*in_p));  // 40 41 42 43 44 45 46 47
+    x5 = _mm_loadl_epi64((__m128i *)(in + 5*in_p));  // 50 51 52 53 54 55 56 57
+    // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+    x2 = _mm_unpacklo_epi8(x4, x5);
+
+    x6 = _mm_loadl_epi64((__m128i *)(in + 6*in_p));  // 60 61 62 63 64 65 66 67
+    x7 = _mm_loadl_epi64((__m128i *)(in + 7*in_p));  // 70 71 72 73 74 75 76 77
+    // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+    x3 = _mm_unpacklo_epi8(x6, x7);
+
+    // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+    x4 = _mm_unpacklo_epi16(x0, x1);
+    // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73
+    x5 = _mm_unpacklo_epi16(x2, x3);
+    // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
+    x6 = _mm_unpacklo_epi32(x4, x5);
+    _mm_storel_pd((double *)(out + 0*out_p),
+                  _mm_castsi128_pd(x6));  // 00 10 20 30 40 50 60 70
+    _mm_storeh_pd((double *)(out + 1*out_p),
+                  _mm_castsi128_pd(x6));  // 01 11 21 31 41 51 61 71
+    // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
+    x7 = _mm_unpackhi_epi32(x4, x5);
+    _mm_storel_pd((double *)(out + 2*out_p),
+                  _mm_castsi128_pd(x7));  // 02 12 22 32 42 52 62 72
+    _mm_storeh_pd((double *)(out + 3*out_p),
+                  _mm_castsi128_pd(x7));  // 03 13 23 33 43 53 63 73
+
+    // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+    x4 = _mm_unpackhi_epi16(x0, x1);
+    // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77
+    x5 = _mm_unpackhi_epi16(x2, x3);
+    // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75
+    x6 = _mm_unpacklo_epi32(x4, x5);
+    _mm_storel_pd((double *)(out + 4*out_p),
+                  _mm_castsi128_pd(x6));  // 04 14 24 34 44 54 64 74
+    _mm_storeh_pd((double *)(out + 5*out_p),
+                  _mm_castsi128_pd(x6));  // 05 15 25 35 45 55 65 75
+    // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77
+    x7 = _mm_unpackhi_epi32(x4, x5);
+
+    _mm_storel_pd((double *)(out + 6*out_p),
+                  _mm_castsi128_pd(x7));  // 06 16 26 36 46 56 66 76
+    _mm_storeh_pd((double *)(out + 7*out_p),
+                  _mm_castsi128_pd(x7));  // 07 17 27 37 47 57 67 77
+  } while (++idx8x8 < num_8x8_to_transpose);
+}
+
+void vpx_lpf_vertical_4_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0,
+                                  const uint8_t *thresh0,
+                                  const uint8_t *blimit1,
+                                  const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+  unsigned char *src[2];
+  unsigned char *dst[2];
+
+  // Transpose 8x16
+  transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  vpx_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,
+                                 blimit1, limit1, thresh1);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  transpose(src, 16, dst, p, 2);
+}
+
+void vpx_lpf_vertical_8_sse2(unsigned char *s, int p,
+                             const unsigned char *blimit,
+                             const unsigned char *limit,
+                             const unsigned char *thresh) {
+  DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 8]);
+  unsigned char *src[1];
+  unsigned char *dst[1];
+
+  // Transpose 8x8
+  src[0] = s - 4;
+  dst[0] = t_dst;
+
+  transpose(src, p, dst, 8, 1);
+
+  // Loop filtering
+  vpx_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh);
+
+  src[0] = t_dst;
+  dst[0] = s - 4;
+
+  // Transpose back
+  transpose(src, 8, dst, p, 1);
+}
+
+void vpx_lpf_vertical_8_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0,
+                                  const uint8_t *thresh0,
+                                  const uint8_t *blimit1,
+                                  const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+  unsigned char *src[2];
+  unsigned char *dst[2];
+
+  // Transpose 8x16
+  transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  vpx_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,
+                                 blimit1, limit1, thresh1);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  transpose(src, 16, dst, p, 2);
+}
+
+void vpx_lpf_vertical_16_sse2(unsigned char *s, int p,
+                              const unsigned char *blimit,
+                              const unsigned char *limit,
+                              const unsigned char *thresh) {
+  DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 16]);
+  unsigned char *src[2];
+  unsigned char *dst[2];
+
+  src[0] = s - 8;
+  src[1] = s;
+  dst[0] = t_dst;
+  dst[1] = t_dst + 8 * 8;
+
+  // Transpose 16x8
+  transpose(src, p, dst, 8, 2);
+
+  // Loop filtering
+  vpx_lpf_horizontal_edge_8_sse2(t_dst + 8 * 8, 8, blimit, limit, thresh);
+
+  src[0] = t_dst;
+  src[1] = t_dst + 8 * 8;
+  dst[0] = s - 8;
+  dst[1] = s;
+
+  // Transpose back
+  transpose(src, 8, dst, p, 2);
+}
+
+void vpx_lpf_vertical_16_dual_sse2(unsigned char *s, int p,
+                                   const uint8_t *blimit, const uint8_t *limit,
+                                   const uint8_t *thresh) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[256]);
+
+  // Transpose 16x16
+  transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16);
+  transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16);
+
+  // Loop filtering
+  vpx_lpf_horizontal_edge_16_sse2(t_dst + 8 * 16, 16, blimit, limit, thresh);
+
+  // Transpose back
+  transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p);
+  transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p);
+}
diff --git a/libvpx/libvpx/vpx_dsp/x86/quantize_avx_x86_64.asm b/libvpx/libvpx/vpx_dsp/x86/quantize_avx_x86_64.asm
new file mode 100644
index 0000000..01c4129
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/quantize_avx_x86_64.asm
@@ -0,0 +1,544 @@
+;
+;  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro QUANTIZE_FN 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, \
+                                eob, scan, iscan
+
+  vzeroupper
+
+  ; If we can skip this block, then just zero the output
+  cmp                         skipmp, 0
+  jne .blank
+
+%ifnidn %1, b_32x32
+
+  ; Special case for ncoeff == 16, as it is frequent and we can save on
+  ; not setting up a loop.
+  cmp                       ncoeffmp, 16
+  jne .generic
+
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+  ;; Special case of ncoeff == 16
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+.single:
+
+  movifnidn                   coeffq, coeffmp
+  movifnidn                    zbinq, zbinmp
+  mova                            m0, [zbinq]              ; m0 = zbin
+
+  ; Get DC and first 15 AC coeffs - in this special case, that is all.
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; coeff stored as 32bit numbers but we process them as 16 bit numbers
+  mova                            m9, [coeffq]
+  packssdw                        m9, [coeffq+16]          ; m9 = c[i]
+  mova                           m10, [coeffq+32]
+  packssdw                       m10, [coeffq+48]          ; m10 = c[i]
+%else
+  mova                            m9, [coeffq]             ; m9 = c[i]
+  mova                           m10, [coeffq+16]          ; m10 = c[i]
+%endif
+
+  mov                             r0, eobmp                ; Output pointer
+  mov                             r1, qcoeffmp             ; Output pointer
+  mov                             r2, dqcoeffmp            ; Output pointer
+
+  pxor                            m5, m5                   ; m5 = dedicated zero
+
+  pcmpeqw                         m4, m4                   ; All word lanes -1
+  paddw                           m0, m4                   ; m0 = zbin - 1
+
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  punpckhqdq                      m0, m0
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+
+  ; Check if all coeffs are less than zbin. If yes, we just write zeros
+  ; to the outputs and we are done.
+  por                            m14, m7, m12
+  ptest                          m14, m14
+  jnz .single_nonzero
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova                       [r1   ], ymm5
+  mova                       [r1+32], ymm5
+  mova                       [r2   ], ymm5
+  mova                       [r2+32], ymm5
+%else
+  mova                          [r1], ymm5
+  mova                          [r2], ymm5
+%endif
+  mov                           [r0], word 0
+
+  vzeroupper
+  RET
+
+.single_nonzero:
+
+  ; Actual quantization of size 16 block - setup pointers, rounders, etc.
+  movifnidn                       r4, roundmp
+  movifnidn                       r5, quantmp
+  mov                             r3, dequantmp
+  mov                             r6, shiftmp
+  mova                            m1, [r4]              ; m1 = round
+  mova                            m2, [r5]              ; m2 = quant
+  mova                            m3, [r3]              ; m3 = dequant
+  mova                            m4, [r6]              ; m4 = shift
+
+  mov                             r3, iscanmp
+
+  DEFINE_ARGS eob, qcoeff, dqcoeff, iscan
+
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                           m8, m6                   ; m8 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                          m8, m4                   ; m8 = m8*qsh>>16
+  punpckhqdq                      m4, m4
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                            m8, m7
+  pand                           m13, m12
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m8
+  punpckhwd                       m6, m8, m6
+  pmovsxwd                       m11, m8
+  mova                  [qcoeffq   ], m11
+  mova                  [qcoeffq+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova                  [qcoeffq+32], m11
+  mova                  [qcoeffq+48], m6
+%else
+  mova                  [qcoeffq   ], m8
+  mova                  [qcoeffq+16], m13
+%endif
+
+  pmullw                          m8, m3                   ; dqc[i] = qc[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m8
+  punpckhwd                       m6, m8, m6
+  pmovsxwd                       m11, m8
+  mova                 [dqcoeffq   ], m11
+  mova                 [dqcoeffq+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova                 [dqcoeffq+32], m11
+  mova                 [dqcoeffq+48], m6
+%else
+  mova                 [dqcoeffq   ], m8
+  mova                 [dqcoeffq+16], m13
+%endif
+
+  mova                            m6, [iscanq]            ; m6 = scan[i]
+  mova                           m11, [iscanq+16]         ; m11 = scan[i]
+
+  pcmpeqw                         m8,  m8,  m5            ; m8 = c[i] == 0
+  pcmpeqw                        m13, m13,  m5            ; m13 = c[i] == 0
+  psubw                           m6,  m6,  m7            ; m6 = scan[i] + 1
+  psubw                          m11, m11, m12            ; m11 = scan[i] + 1
+  pandn                           m8,  m8,  m6            ; m8 = max(eob)
+  pandn                          m13, m13, m11            ; m13 = max(eob)
+  pmaxsw                          m8,  m8, m13
+
+  ; Horizontally accumulate/max eobs and write into [eob] memory pointer
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  movq                           rax, m8
+  mov                         [eobq], ax
+
+  vzeroupper
+  RET
+
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+  ;; Generic case of ncoeff != 16
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+.generic:
+
+%endif ; %ifnidn %1, b_32x32
+
+DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
+            qcoeff, dqcoeff, dequant, eob, scan, iscan
+
+  ; Actual quantization loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, dequantmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  mova                            m0, [zbinq]              ; m0 = zbin
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+  mova                            m3, [r2]                 ; m3 = dequant
+  pcmpeqw                         m4, m4                   ; All lanes -1
+%ifidn %1, b_32x32
+  psubw                           m0, m4
+  psubw                           m1, m4
+  psrlw                           m0, 1                    ; m0 = (m0 + 1) / 2
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  paddw                           m0, m4                   ; m0 = m0 + 1
+
+  mov                             r2, shiftmp
+  mov                             r3, qcoeffmp
+  mova                            m4, [r2]                 ; m4 = shift
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+%ifidn %1, b_32x32
+  psllw                           m4, 1
+%endif
+  pxor                            m5, m5                   ; m5 = dedicated zero
+
+  DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  lea                         coeffq, [  coeffq+ncoeffq*4]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*4]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*4]
+%else
+  lea                         coeffq, [  coeffq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+%endif
+  lea                         iscanq, [  iscanq+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; coeff stored as 32bit numbers & require 16bit numbers
+  mova                            m9, [coeffq+ncoeffq*4+ 0]
+  packssdw                        m9, [coeffq+ncoeffq*4+16]
+  mova                           m10, [coeffq+ncoeffq*4+32]
+  packssdw                       m10, [coeffq+ncoeffq*4+48]
+%else
+  mova                            m9, [coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [coeffq+ncoeffq*2+16] ; m10 = c[i]
+%endif
+
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  punpckhqdq                      m0, m0
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+
+  ; Check if all coeffs are less than zbin. If yes, skip forward quickly.
+  por                            m14, m7, m12
+  ptest                          m14, m14
+  jnz .first_nonzero
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova        [qcoeffq+ncoeffq*4   ], ymm5
+  mova        [qcoeffq+ncoeffq*4+32], ymm5
+  mova       [dqcoeffq+ncoeffq*4   ], ymm5
+  mova       [dqcoeffq+ncoeffq*4+32], ymm5
+%else
+  mova           [qcoeffq+ncoeffq*2], ymm5
+  mova          [dqcoeffq+ncoeffq*2], ymm5
+%endif
+
+  add                        ncoeffq, mmsize
+
+  punpckhqdq                      m1, m1
+  punpckhqdq                      m2, m2
+  punpckhqdq                      m3, m3
+  punpckhqdq                      m4, m4
+  pxor                            m8, m8
+
+  jmp .ac_only_loop
+
+.first_nonzero:
+
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                           m8, m6                   ; m8 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                          m8, m4                   ; m8 = m8*qsh>>16
+  punpckhqdq                      m4, m4
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                            m8, m7
+  pand                           m13, m12
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m8
+  punpckhwd                       m6, m8, m6
+  pmovsxwd                       m11, m8
+  mova        [qcoeffq+ncoeffq*4+ 0], m11
+  mova        [qcoeffq+ncoeffq*4+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova        [qcoeffq+ncoeffq*4+32], m11
+  mova        [qcoeffq+ncoeffq*4+48], m6
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], m8
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%endif
+
+%ifidn %1, b_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; dqc[i] = qc[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+%endif
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m8
+  punpckhwd                       m6, m8, m6
+  pmovsxwd                       m11, m8
+  mova       [dqcoeffq+ncoeffq*4+ 0], m11
+  mova       [dqcoeffq+ncoeffq*4+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova       [dqcoeffq+ncoeffq*4+32], m11
+  mova       [dqcoeffq+ncoeffq*4+48], m6
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], m8
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+%endif
+
+  pcmpeqw                         m8, m5                    ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                    ; m13 = c[i] == 0
+  mova                            m6, [iscanq+ncoeffq*2]    ; m6 = scan[i]
+  mova                           m11, [iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                    ; m6 = scan[i] + 1
+  psubw                          m11, m12                   ; m11 = scan[i] + 1
+  pandn                           m8, m6                    ; m8 = max(eob)
+  pandn                          m13, m11                   ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+
+.ac_only_loop:
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; pack coeff from 32bit to 16bit array
+  mova                            m9, [coeffq+ncoeffq*4+ 0]
+  packssdw                        m9, [coeffq+ncoeffq*4+16]
+  mova                           m10, [coeffq+ncoeffq*4+32]
+  packssdw                       m10, [coeffq+ncoeffq*4+48]
+%else
+  mova                            m9, [coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [coeffq+ncoeffq*2+16] ; m10 = c[i]
+%endif
+
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+
+  ; Check if all coeffs are less than zbin. If yes, skip this itertion.
+  ; And just write zeros as the result would be.
+  por                            m14, m7, m12
+  ptest                          m14, m14
+  jnz .rest_nonzero
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova        [qcoeffq+ncoeffq*4+ 0], ymm5
+  mova        [qcoeffq+ncoeffq*4+32], ymm5
+  mova       [dqcoeffq+ncoeffq*4+ 0], ymm5
+  mova       [dqcoeffq+ncoeffq*4+32], ymm5
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], ymm5
+  mova       [dqcoeffq+ncoeffq*2+ 0], ymm5
+%endif
+  add                        ncoeffq, mmsize
+  jnz .ac_only_loop
+
+  ; Horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  movq                           rax, m8
+  mov                           [r2], ax
+  vzeroupper
+  RET
+
+.rest_nonzero:
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                          m14, m6                   ; m14 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                         m14, m4                   ; m14 = m14*qsh>>16
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                           m14, m7
+  pand                           m13, m12
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m14
+  punpckhwd                       m6, m14, m6
+  pmovsxwd                       m11, m14
+  mova        [qcoeffq+ncoeffq*4+ 0], m11
+  mova        [qcoeffq+ncoeffq*4+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova        [qcoeffq+ncoeffq*4+32], m11
+  mova        [qcoeffq+ncoeffq*4+48], m6
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], m14
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%endif
+
+%ifidn %1, b_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; dqc[i] = qc[i] * q
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m14
+  punpckhwd                       m6, m14, m6
+  pmovsxwd                       m11, m14
+  mova       [dqcoeffq+ncoeffq*4+ 0], m11
+  mova       [dqcoeffq+ncoeffq*4+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova       [dqcoeffq+ncoeffq*4+32], m11
+  mova       [dqcoeffq+ncoeffq*4+48], m6
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], m14
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+%endif
+
+  pcmpeqw                        m14, m5                    ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                    ; m13 = c[i] == 0
+  mova                            m6, [iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                    ; m6 = scan[i] + 1
+  psubw                          m11, m12                   ; m11 = scan[i] + 1
+  pandn                          m14, m6                    ; m14 = max(eob)
+  pandn                          m13, m11                   ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jnz .ac_only_loop
+
+  ; Horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  movq                           rax, m8
+  mov                           [r2], ax
+  vzeroupper
+  RET
+
+  ; Skip-block, i.e. just write all zeroes
+.blank:
+
+DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
+            qcoeff, dqcoeff, dequant, eob, scan, iscan
+
+  mov                             r0, dqcoeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, qcoeffmp
+  mov                             r3, eobmp
+
+DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*4]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*4]
+%else
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+%endif
+
+  neg                        ncoeffq
+  pxor                            m7, m7
+
+.blank_loop:
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova       [dqcoeffq+ncoeffq*4+ 0], ymm7
+  mova       [dqcoeffq+ncoeffq*4+32], ymm7
+  mova        [qcoeffq+ncoeffq*4+ 0], ymm7
+  mova        [qcoeffq+ncoeffq*4+32], ymm7
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], ymm7
+  mova        [qcoeffq+ncoeffq*2+ 0], ymm7
+%endif
+  add                        ncoeffq, mmsize
+  jl .blank_loop
+
+  mov                         [eobq], word 0
+
+  vzeroupper
+  RET
+%endmacro
+
+INIT_XMM avx
+QUANTIZE_FN b, 7
+QUANTIZE_FN b_32x32, 7
+
+END
diff --git a/libvpx/libvpx/vpx_dsp/x86/quantize_sse2.c b/libvpx/libvpx/vpx_dsp/x86/quantize_sse2.c
new file mode 100644
index 0000000..8aa4568
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/quantize_sse2.c
@@ -0,0 +1,248 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+#include <xmmintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+static INLINE __m128i load_coefficients(const tran_low_t *coeff_ptr) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  return _mm_setr_epi16((int16_t)coeff_ptr[0], (int16_t)coeff_ptr[1],
+      (int16_t)coeff_ptr[2], (int16_t)coeff_ptr[3], (int16_t)coeff_ptr[4],
+      (int16_t)coeff_ptr[5], (int16_t)coeff_ptr[6], (int16_t)coeff_ptr[7]);
+#else
+  return _mm_load_si128((const __m128i *)coeff_ptr);
+#endif
+}
+
+static INLINE void store_coefficients(__m128i coeff_vals,
+                                      tran_low_t *coeff_ptr) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  __m128i one = _mm_set1_epi16(1);
+  __m128i coeff_vals_hi = _mm_mulhi_epi16(coeff_vals, one);
+  __m128i coeff_vals_lo = _mm_mullo_epi16(coeff_vals, one);
+  __m128i coeff_vals_1 = _mm_unpacklo_epi16(coeff_vals_lo, coeff_vals_hi);
+  __m128i coeff_vals_2 = _mm_unpackhi_epi16(coeff_vals_lo, coeff_vals_hi);
+  _mm_store_si128((__m128i*)(coeff_ptr), coeff_vals_1);
+  _mm_store_si128((__m128i*)(coeff_ptr + 4), coeff_vals_2);
+#else
+  _mm_store_si128((__m128i*)(coeff_ptr), coeff_vals);
+#endif
+}
+
+void vpx_quantize_b_sse2(const tran_low_t* coeff_ptr, intptr_t n_coeffs,
+                         int skip_block, const int16_t* zbin_ptr,
+                         const int16_t* round_ptr, const int16_t* quant_ptr,
+                         const int16_t* quant_shift_ptr, tran_low_t* qcoeff_ptr,
+                         tran_low_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+                         uint16_t* eob_ptr,
+                         const int16_t* scan_ptr,
+                         const int16_t* iscan_ptr) {
+  __m128i zero;
+  (void)scan_ptr;
+
+  coeff_ptr += n_coeffs;
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+  if (!skip_block) {
+    __m128i eob;
+    __m128i zbin;
+    __m128i round, quant, dequant, shift;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        __m128i pw_1;
+        zbin = _mm_load_si128((const __m128i*)zbin_ptr);
+        round = _mm_load_si128((const __m128i*)round_ptr);
+        quant = _mm_load_si128((const __m128i*)quant_ptr);
+        pw_1 = _mm_set1_epi16(1);
+        zbin = _mm_sub_epi16(zbin, pw_1);
+        dequant = _mm_load_si128((const __m128i*)dequant_ptr);
+        shift = _mm_load_si128((const __m128i*)quant_shift_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        __m128i cmp_mask0, cmp_mask1;
+        // Do DC and first 15 AC
+        coeff0 = load_coefficients(coeff_ptr + n_coeffs);
+        coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+        zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+        cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+        qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
+        qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
+        qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
+        shift = _mm_unpackhi_epi64(shift, shift);
+        qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        // Mask out zbin threshold coeffs
+        qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+        qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+        store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs);
+        store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        store_coefficients(coeff0, dqcoeff_ptr + n_coeffs);
+        store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // AC only loop
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        __m128i cmp_mask0, cmp_mask1;
+
+        coeff0 = load_coefficients(coeff_ptr + n_coeffs);
+        coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+        cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+        qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
+        qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
+        qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
+        qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        // Mask out zbin threshold coeffs
+        qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+        qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+        store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs);
+        store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        store_coefficients(coeff0, dqcoeff_ptr + n_coeffs);
+        store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      store_coefficients(zero, dqcoeff_ptr + n_coeffs);
+      store_coefficients(zero, dqcoeff_ptr + n_coeffs + 8);
+      store_coefficients(zero, qcoeff_ptr + n_coeffs);
+      store_coefficients(zero, qcoeff_ptr + n_coeffs + 8);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/x86/quantize_ssse3_x86_64.asm b/libvpx/libvpx/vpx_dsp/x86/quantize_ssse3_x86_64.asm
new file mode 100644
index 0000000..ca21539
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/quantize_ssse3_x86_64.asm
@@ -0,0 +1,346 @@
+;
+;  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_1: times 8 dw 1
+
+SECTION .text
+
+; TODO(yunqingwang)fix quantize_b code for skip=1 case.
+%macro QUANTIZE_FN 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, \
+                                eob, scan, iscan
+  cmp                    dword skipm, 0
+  jne .blank
+
+  ; actual quantize loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, dequantmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  mova                            m0, [zbinq]              ; m0 = zbin
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+%ifidn %1, b_32x32
+  pcmpeqw                         m5, m5
+  psrlw                           m5, 15
+  paddw                           m0, m5
+  paddw                           m1, m5
+  psrlw                           m0, 1                    ; m0 = (m0 + 1) / 2
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  mova                            m3, [r2q]                ; m3 = dequant
+  psubw                           m0, [pw_1]
+  mov                             r2, shiftmp
+  mov                             r3, qcoeffmp
+  mova                            m4, [r2]                 ; m4 = shift
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+%ifidn %1, b_32x32
+  psllw                           m4, 1
+%endif
+  pxor                            m5, m5                   ; m5 = dedicated zero
+  DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
+%if CONFIG_VP9_HIGHBITDEPTH
+  lea                         coeffq, [  coeffq+ncoeffq*4]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*4]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*4]
+%else
+  lea                         coeffq, [  coeffq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+%endif
+  lea                         iscanq, [  iscanq+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; coeff stored as 32bit numbers & require 16bit numbers
+  mova                            m9, [  coeffq+ncoeffq*4+ 0]
+  packssdw                        m9, [  coeffq+ncoeffq*4+16]
+  mova                           m10, [  coeffq+ncoeffq*4+32]
+  packssdw                       m10, [  coeffq+ncoeffq*4+48]
+%else
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+%endif
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  punpckhqdq                      m0, m0
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                           m8, m6                   ; m8 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                          m8, m4                   ; m8 = m8*qsh>>16
+  punpckhqdq                      m4, m4
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                            m8, m7
+  pand                           m13, m12
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  mova                           m11, m8
+  mova                            m6, m8
+  pcmpgtw                         m5, m8
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+ 0], m11
+  mova        [qcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+32], m11
+  mova        [qcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5             ; reset m5 to zero register
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], m8
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%endif
+%ifidn %1, b_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; dqc[i] = qc[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+%endif
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  mova                            m11, m8
+  mova                            m6, m8
+  pcmpgtw                         m5, m8
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+ 0], m11
+  mova       [dqcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+32], m11
+  mova       [dqcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5             ; reset m5 to zero register
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], m8
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+%endif
+  pcmpeqw                         m8, m5                   ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m12                  ; m11 = scan[i] + 1
+  pandn                           m8, m6                   ; m8 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jz .accumulate_eob
+
+.ac_only_loop:
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; pack coeff from 32bit to 16bit array
+  mova                            m9, [  coeffq+ncoeffq*4+ 0]
+  packssdw                        m9, [  coeffq+ncoeffq*4+16]
+  mova                           m10, [  coeffq+ncoeffq*4+32]
+  packssdw                       m10, [  coeffq+ncoeffq*4+48]
+%else
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+%endif
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+%ifidn %1, b_32x32
+  pmovmskb                       r6d, m7
+  pmovmskb                       r2d, m12
+  or                              r6, r2
+  jz .skip_iter
+%endif
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                          m14, m6                   ; m14 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                         m14, m4                   ; m14 = m14*qsh>>16
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                           m14, m7
+  pand                           m13, m12
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pxor                           m11, m11
+  mova                           m11, m14
+  mova                            m6, m14
+  pcmpgtw                         m5, m14
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+ 0], m11
+  mova        [qcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+32], m11
+  mova        [qcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5             ; reset m5 to zero register
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], m14
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%endif
+%ifidn %1, b_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; dqc[i] = qc[i] * q
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+%if CONFIG_VP9_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  mova                           m11, m14
+  mova                            m6, m14
+  pcmpgtw                         m5, m14
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+ 0], m11
+  mova       [dqcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+32], m11
+  mova       [dqcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], m14
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+%endif
+  pcmpeqw                        m14, m5                   ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m12                  ; m11 = scan[i] + 1
+  pandn                          m14, m6                   ; m14 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+%ifidn %1, b_32x32
+  jmp .accumulate_eob
+.skip_iter:
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova        [qcoeffq+ncoeffq*4+ 0], m5
+  mova        [qcoeffq+ncoeffq*4+16], m5
+  mova        [qcoeffq+ncoeffq*4+32], m5
+  mova        [qcoeffq+ncoeffq*4+48], m5
+  mova       [dqcoeffq+ncoeffq*4+ 0], m5
+  mova       [dqcoeffq+ncoeffq*4+16], m5
+  mova       [dqcoeffq+ncoeffq*4+32], m5
+  mova       [dqcoeffq+ncoeffq*4+48], m5
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], m5
+  mova        [qcoeffq+ncoeffq*2+16], m5
+  mova       [dqcoeffq+ncoeffq*2+ 0], m5
+  mova       [dqcoeffq+ncoeffq*2+16], m5
+%endif
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+%endif
+
+.accumulate_eob:
+  ; horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  pextrw                          r6, m8, 0
+  mov                             [r2], r6
+  RET
+
+  ; skip-block, i.e. just write all zeroes
+.blank:
+  mov                             r0, dqcoeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, qcoeffmp
+  mov                             r3, eobmp
+  DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
+%if CONFIG_VP9_HIGHBITDEPTH
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*4]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*4]
+%else
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+%endif
+  neg                        ncoeffq
+  pxor                            m7, m7
+.blank_loop:
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova       [dqcoeffq+ncoeffq*4+ 0], m7
+  mova       [dqcoeffq+ncoeffq*4+16], m7
+  mova       [dqcoeffq+ncoeffq*4+32], m7
+  mova       [dqcoeffq+ncoeffq*4+48], m7
+  mova        [qcoeffq+ncoeffq*4+ 0], m7
+  mova        [qcoeffq+ncoeffq*4+16], m7
+  mova        [qcoeffq+ncoeffq*4+32], m7
+  mova        [qcoeffq+ncoeffq*4+48], m7
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], m7
+  mova       [dqcoeffq+ncoeffq*2+16], m7
+  mova        [qcoeffq+ncoeffq*2+ 0], m7
+  mova        [qcoeffq+ncoeffq*2+16], m7
+%endif
+  add                        ncoeffq, mmsize
+  jl .blank_loop
+  mov                    word [eobq], 0
+  RET
+%endmacro
+
+INIT_XMM ssse3
+QUANTIZE_FN b, 7
+QUANTIZE_FN b_32x32, 7
diff --git a/libvpx/libvpx/vpx_dsp/x86/sad4d_avx2.c b/libvpx/libvpx/vpx_dsp/x86/sad4d_avx2.c
new file mode 100644
index 0000000..793658f
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/sad4d_avx2.c
@@ -0,0 +1,168 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <immintrin.h>  // AVX2
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+void vpx_sad32x32x4d_avx2(const uint8_t *src,
+                          int src_stride,
+                          const uint8_t *const ref[4],
+                          int ref_stride,
+                          uint32_t res[4]) {
+  __m256i src_reg, ref0_reg, ref1_reg, ref2_reg, ref3_reg;
+  __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
+  __m256i sum_mlow, sum_mhigh;
+  int i;
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+  sum_ref0 = _mm256_set1_epi16(0);
+  sum_ref1 = _mm256_set1_epi16(0);
+  sum_ref2 = _mm256_set1_epi16(0);
+  sum_ref3 = _mm256_set1_epi16(0);
+  for (i = 0; i < 32 ; i++) {
+    // load src and all refs
+    src_reg = _mm256_loadu_si256((const __m256i *)src);
+    ref0_reg = _mm256_loadu_si256((const __m256i *)ref0);
+    ref1_reg = _mm256_loadu_si256((const __m256i *)ref1);
+    ref2_reg = _mm256_loadu_si256((const __m256i *)ref2);
+    ref3_reg = _mm256_loadu_si256((const __m256i *)ref3);
+    // sum of the absolute differences between every ref-i to src
+    ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
+    ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
+    ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
+    ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
+    // sum every ref-i
+    sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
+    sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
+    sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
+    sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
+
+    src+= src_stride;
+    ref0+= ref_stride;
+    ref1+= ref_stride;
+    ref2+= ref_stride;
+    ref3+= ref_stride;
+  }
+  {
+    __m128i sum;
+    // in sum_ref-i the result is saved in the first 4 bytes
+    // the other 4 bytes are zeroed.
+    // sum_ref1 and sum_ref3 are shifted left by 4 bytes
+    sum_ref1 = _mm256_slli_si256(sum_ref1, 4);
+    sum_ref3 = _mm256_slli_si256(sum_ref3, 4);
+
+    // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
+    sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1);
+    sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3);
+
+    // merge every 64 bit from each sum_ref-i
+    sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2);
+    sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2);
+
+    // add the low 64 bit to the high 64 bit
+    sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh);
+
+    // add the low 128 bit to the high 128 bit
+    sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow),
+                        _mm256_extractf128_si256(sum_mlow, 1));
+
+    _mm_storeu_si128((__m128i *)(res), sum);
+  }
+}
+
+void vpx_sad64x64x4d_avx2(const uint8_t *src,
+                          int src_stride,
+                          const uint8_t *const ref[4],
+                          int ref_stride,
+                          uint32_t res[4]) {
+  __m256i src_reg, srcnext_reg, ref0_reg, ref0next_reg;
+  __m256i ref1_reg, ref1next_reg, ref2_reg, ref2next_reg;
+  __m256i ref3_reg, ref3next_reg;
+  __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
+  __m256i sum_mlow, sum_mhigh;
+  int i;
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+  sum_ref0 = _mm256_set1_epi16(0);
+  sum_ref1 = _mm256_set1_epi16(0);
+  sum_ref2 = _mm256_set1_epi16(0);
+  sum_ref3 = _mm256_set1_epi16(0);
+  for (i = 0; i < 64 ; i++) {
+    // load 64 bytes from src and all refs
+    src_reg = _mm256_loadu_si256((const __m256i *)src);
+    srcnext_reg = _mm256_loadu_si256((const __m256i *)(src + 32));
+    ref0_reg = _mm256_loadu_si256((const __m256i *)ref0);
+    ref0next_reg = _mm256_loadu_si256((const __m256i *)(ref0 + 32));
+    ref1_reg = _mm256_loadu_si256((const __m256i *)ref1);
+    ref1next_reg = _mm256_loadu_si256((const __m256i *)(ref1 + 32));
+    ref2_reg = _mm256_loadu_si256((const __m256i *)ref2);
+    ref2next_reg = _mm256_loadu_si256((const __m256i *)(ref2 + 32));
+    ref3_reg = _mm256_loadu_si256((const __m256i *)ref3);
+    ref3next_reg = _mm256_loadu_si256((const __m256i *)(ref3 + 32));
+    // sum of the absolute differences between every ref-i to src
+    ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
+    ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
+    ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
+    ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
+    ref0next_reg = _mm256_sad_epu8(ref0next_reg, srcnext_reg);
+    ref1next_reg = _mm256_sad_epu8(ref1next_reg, srcnext_reg);
+    ref2next_reg = _mm256_sad_epu8(ref2next_reg, srcnext_reg);
+    ref3next_reg = _mm256_sad_epu8(ref3next_reg, srcnext_reg);
+
+    // sum every ref-i
+    sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
+    sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
+    sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
+    sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
+    sum_ref0 = _mm256_add_epi32(sum_ref0, ref0next_reg);
+    sum_ref1 = _mm256_add_epi32(sum_ref1, ref1next_reg);
+    sum_ref2 = _mm256_add_epi32(sum_ref2, ref2next_reg);
+    sum_ref3 = _mm256_add_epi32(sum_ref3, ref3next_reg);
+    src+= src_stride;
+    ref0+= ref_stride;
+    ref1+= ref_stride;
+    ref2+= ref_stride;
+    ref3+= ref_stride;
+  }
+  {
+    __m128i sum;
+
+    // in sum_ref-i the result is saved in the first 4 bytes
+    // the other 4 bytes are zeroed.
+    // sum_ref1 and sum_ref3 are shifted left by 4 bytes
+    sum_ref1 = _mm256_slli_si256(sum_ref1, 4);
+    sum_ref3 = _mm256_slli_si256(sum_ref3, 4);
+
+    // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
+    sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1);
+    sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3);
+
+    // merge every 64 bit from each sum_ref-i
+    sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2);
+    sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2);
+
+    // add the low 64 bit to the high 64 bit
+    sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh);
+
+    // add the low 128 bit to the high 128 bit
+    sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow),
+                        _mm256_extractf128_si256(sum_mlow, 1));
+
+    _mm_storeu_si128((__m128i *)(res), sum);
+  }
+}
diff --git a/libvpx/libvpx/vpx_dsp/x86/sad4d_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/sad4d_sse2.asm
new file mode 100644
index 0000000..3f6e55c
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/sad4d_sse2.asm
@@ -0,0 +1,239 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; PROCESS_4x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_4x2x4 5-6 0
+  movd                  m0, [srcq +%2]
+%if %1 == 1
+  movd                  m6, [ref1q+%3]
+  movd                  m4, [ref2q+%3]
+  movd                  m7, [ref3q+%3]
+  movd                  m5, [ref4q+%3]
+  movd                  m1, [srcq +%4]
+  movd                  m2, [ref1q+%5]
+  punpckldq             m0, m1
+  punpckldq             m6, m2
+  movd                  m1, [ref2q+%5]
+  movd                  m2, [ref3q+%5]
+  movd                  m3, [ref4q+%5]
+  punpckldq             m4, m1
+  punpckldq             m7, m2
+  punpckldq             m5, m3
+  movlhps               m0, m0
+  movlhps               m6, m4
+  movlhps               m7, m5
+  psadbw                m6, m0
+  psadbw                m7, m0
+%else
+  movd                  m1, [ref1q+%3]
+  movd                  m5, [ref1q+%5]
+  movd                  m2, [ref2q+%3]
+  movd                  m4, [ref2q+%5]
+  punpckldq             m1, m5
+  punpckldq             m2, m4
+  movd                  m3, [ref3q+%3]
+  movd                  m5, [ref3q+%5]
+  punpckldq             m3, m5
+  movd                  m4, [ref4q+%3]
+  movd                  m5, [ref4q+%5]
+  punpckldq             m4, m5
+  movd                  m5, [srcq +%4]
+  punpckldq             m0, m5
+  movlhps               m0, m0
+  movlhps               m1, m2
+  movlhps               m3, m4
+  psadbw                m1, m0
+  psadbw                m3, m0
+  paddd                 m6, m1
+  paddd                 m7, m3
+%endif
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*2]
+  lea                ref1q, [ref1q+ref_strideq*2]
+  lea                ref2q, [ref2q+ref_strideq*2]
+  lea                ref3q, [ref3q+ref_strideq*2]
+  lea                ref4q, [ref4q+ref_strideq*2]
+%endif
+%endmacro
+
+; PROCESS_8x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_8x2x4 5-6 0
+  movh                  m0, [srcq +%2]
+%if %1 == 1
+  movh                  m4, [ref1q+%3]
+  movh                  m5, [ref2q+%3]
+  movh                  m6, [ref3q+%3]
+  movh                  m7, [ref4q+%3]
+  movhps                m0, [srcq +%4]
+  movhps                m4, [ref1q+%5]
+  movhps                m5, [ref2q+%5]
+  movhps                m6, [ref3q+%5]
+  movhps                m7, [ref4q+%5]
+  psadbw                m4, m0
+  psadbw                m5, m0
+  psadbw                m6, m0
+  psadbw                m7, m0
+%else
+  movh                  m1, [ref1q+%3]
+  movh                  m2, [ref2q+%3]
+  movh                  m3, [ref3q+%3]
+  movhps                m0, [srcq +%4]
+  movhps                m1, [ref1q+%5]
+  movhps                m2, [ref2q+%5]
+  movhps                m3, [ref3q+%5]
+  psadbw                m1, m0
+  psadbw                m2, m0
+  psadbw                m3, m0
+  paddd                 m4, m1
+  movh                  m1, [ref4q+%3]
+  movhps                m1, [ref4q+%5]
+  paddd                 m5, m2
+  paddd                 m6, m3
+  psadbw                m1, m0
+  paddd                 m7, m1
+%endif
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*2]
+  lea                ref1q, [ref1q+ref_strideq*2]
+  lea                ref2q, [ref2q+ref_strideq*2]
+  lea                ref3q, [ref3q+ref_strideq*2]
+  lea                ref4q, [ref4q+ref_strideq*2]
+%endif
+%endmacro
+
+; PROCESS_16x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_16x2x4 5-6 0
+  ; 1st 16 px
+  mova                  m0, [srcq +%2]
+%if %1 == 1
+  movu                  m4, [ref1q+%3]
+  movu                  m5, [ref2q+%3]
+  movu                  m6, [ref3q+%3]
+  movu                  m7, [ref4q+%3]
+  psadbw                m4, m0
+  psadbw                m5, m0
+  psadbw                m6, m0
+  psadbw                m7, m0
+%else
+  movu                  m1, [ref1q+%3]
+  movu                  m2, [ref2q+%3]
+  movu                  m3, [ref3q+%3]
+  psadbw                m1, m0
+  psadbw                m2, m0
+  psadbw                m3, m0
+  paddd                 m4, m1
+  movu                  m1, [ref4q+%3]
+  paddd                 m5, m2
+  paddd                 m6, m3
+  psadbw                m1, m0
+  paddd                 m7, m1
+%endif
+
+  ; 2nd 16 px
+  mova                  m0, [srcq +%4]
+  movu                  m1, [ref1q+%5]
+  movu                  m2, [ref2q+%5]
+  movu                  m3, [ref3q+%5]
+  psadbw                m1, m0
+  psadbw                m2, m0
+  psadbw                m3, m0
+  paddd                 m4, m1
+  movu                  m1, [ref4q+%5]
+  paddd                 m5, m2
+  paddd                 m6, m3
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*2]
+  lea                ref1q, [ref1q+ref_strideq*2]
+  lea                ref2q, [ref2q+ref_strideq*2]
+  lea                ref3q, [ref3q+ref_strideq*2]
+  lea                ref4q, [ref4q+ref_strideq*2]
+%endif
+  psadbw                m1, m0
+  paddd                 m7, m1
+%endmacro
+
+; PROCESS_32x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_32x2x4 5-6 0
+  PROCESS_16x2x4 %1, %2, %3, %2 + 16, %3 + 16
+  PROCESS_16x2x4  0, %4, %5, %4 + 16, %5 + 16, %6
+%endmacro
+
+; PROCESS_64x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_64x2x4 5-6 0
+  PROCESS_32x2x4 %1, %2, %3, %2 + 32, %3 + 32
+  PROCESS_32x2x4  0, %4, %5, %4 + 32, %5 + 32, %6
+%endmacro
+
+; void vpx_sadNxNx4d_sse2(uint8_t *src,    int src_stride,
+;                         uint8_t *ref[4], int ref_stride,
+;                         uint32_t res[4]);
+; where NxN = 64x64, 32x32, 16x16, 16x8, 8x16, 8x8, 8x4, 4x8 and 4x4
+%macro SADNXN4D 2
+%if UNIX64
+cglobal sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \
+                              res, ref2, ref3, ref4
+%else
+cglobal sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \
+                              ref2, ref3, ref4
+%endif
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+  mov                ref2q, [ref1q+gprsize*1]
+  mov                ref3q, [ref1q+gprsize*2]
+  mov                ref4q, [ref1q+gprsize*3]
+  mov                ref1q, [ref1q+gprsize*0]
+
+  PROCESS_%1x2x4 1, 0, 0, src_strideq, ref_strideq, 1
+%rep (%2-4)/2
+  PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 1
+%endrep
+  PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 0
+
+%if %1 > 4
+  pslldq                m5, 4
+  pslldq                m7, 4
+  por                   m4, m5
+  por                   m6, m7
+  mova                  m5, m4
+  mova                  m7, m6
+  punpcklqdq            m4, m6
+  punpckhqdq            m5, m7
+  movifnidn             r4, r4mp
+  paddd                 m4, m5
+  movu                [r4], m4
+  RET
+%else
+  movifnidn             r4, r4mp
+  pshufd            m6, m6, 0x08
+  pshufd            m7, m7, 0x08
+  movq              [r4+0], m6
+  movq              [r4+8], m7
+  RET
+%endif
+%endmacro
+
+INIT_XMM sse2
+SADNXN4D 64, 64
+SADNXN4D 64, 32
+SADNXN4D 32, 64
+SADNXN4D 32, 32
+SADNXN4D 32, 16
+SADNXN4D 16, 32
+SADNXN4D 16, 16
+SADNXN4D 16,  8
+SADNXN4D  8, 16
+SADNXN4D  8,  8
+SADNXN4D  8,  4
+SADNXN4D  4,  8
+SADNXN4D  4,  4
diff --git a/libvpx/libvpx/vpx_dsp/x86/sad_avx2.c b/libvpx/libvpx/vpx_dsp/x86/sad_avx2.c
new file mode 100644
index 0000000..ce9ad8f
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/sad_avx2.c
@@ -0,0 +1,181 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <immintrin.h>
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+
+#define FSAD64_H(h) \
+unsigned int vpx_sad64x##h##_avx2(const uint8_t *src_ptr, \
+                                  int src_stride, \
+                                  const uint8_t *ref_ptr, \
+                                  int ref_stride) { \
+  int i, res; \
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \
+  __m256i sum_sad = _mm256_setzero_si256(); \
+  __m256i sum_sad_h; \
+  __m128i sum_sad128; \
+  for (i = 0 ; i < h ; i++) { \
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32)); \
+    sad1_reg = _mm256_sad_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)src_ptr)); \
+    sad2_reg = _mm256_sad_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(src_ptr + 32))); \
+    sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \
+    ref_ptr+= ref_stride; \
+    src_ptr+= src_stride; \
+  } \
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8); \
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \
+  res = _mm_cvtsi128_si32(sum_sad128); \
+  return res; \
+}
+
+#define FSAD32_H(h) \
+unsigned int vpx_sad32x##h##_avx2(const uint8_t *src_ptr, \
+                                  int src_stride, \
+                                  const uint8_t *ref_ptr, \
+                                  int ref_stride) { \
+  int i, res; \
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \
+  __m256i sum_sad = _mm256_setzero_si256(); \
+  __m256i sum_sad_h; \
+  __m128i sum_sad128; \
+  int ref2_stride = ref_stride << 1; \
+  int src2_stride = src_stride << 1; \
+  int max = h >> 1; \
+  for (i = 0 ; i < max ; i++) { \
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \
+    sad1_reg = _mm256_sad_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)src_ptr)); \
+    sad2_reg = _mm256_sad_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride))); \
+    sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \
+    ref_ptr+= ref2_stride; \
+    src_ptr+= src2_stride; \
+  } \
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8); \
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \
+  res = _mm_cvtsi128_si32(sum_sad128); \
+  return res; \
+}
+
+#define FSAD64 \
+FSAD64_H(64); \
+FSAD64_H(32);
+
+#define FSAD32 \
+FSAD32_H(64); \
+FSAD32_H(32); \
+FSAD32_H(16);
+
+FSAD64;
+FSAD32;
+
+#undef FSAD64
+#undef FSAD32
+#undef FSAD64_H
+#undef FSAD32_H
+
+#define FSADAVG64_H(h) \
+unsigned int vpx_sad64x##h##_avg_avx2(const uint8_t *src_ptr, \
+                                      int src_stride, \
+                                      const uint8_t *ref_ptr, \
+                                      int  ref_stride, \
+                                      const uint8_t *second_pred) { \
+  int i, res; \
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \
+  __m256i sum_sad = _mm256_setzero_si256(); \
+  __m256i sum_sad_h; \
+  __m128i sum_sad128; \
+  for (i = 0 ; i < h ; i++) { \
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32)); \
+    ref1_reg = _mm256_avg_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)second_pred)); \
+    ref2_reg = _mm256_avg_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(second_pred +32))); \
+    sad1_reg = _mm256_sad_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)src_ptr)); \
+    sad2_reg = _mm256_sad_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(src_ptr + 32))); \
+    sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \
+    ref_ptr+= ref_stride; \
+    src_ptr+= src_stride; \
+    second_pred+= 64; \
+  } \
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8); \
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \
+  res = _mm_cvtsi128_si32(sum_sad128); \
+  return res; \
+}
+
+#define FSADAVG32_H(h) \
+unsigned int vpx_sad32x##h##_avg_avx2(const uint8_t *src_ptr, \
+                                      int src_stride, \
+                                      const uint8_t *ref_ptr, \
+                                      int  ref_stride, \
+                                      const uint8_t *second_pred) { \
+  int i, res; \
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \
+  __m256i sum_sad = _mm256_setzero_si256(); \
+  __m256i sum_sad_h; \
+  __m128i sum_sad128; \
+  int ref2_stride = ref_stride << 1; \
+  int src2_stride = src_stride << 1; \
+  int max = h >> 1; \
+  for (i = 0 ; i < max ; i++) { \
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \
+    ref1_reg = _mm256_avg_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)second_pred)); \
+    ref2_reg = _mm256_avg_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(second_pred +32))); \
+    sad1_reg = _mm256_sad_epu8(ref1_reg, \
+               _mm256_loadu_si256((__m256i const *)src_ptr)); \
+    sad2_reg = _mm256_sad_epu8(ref2_reg, \
+               _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride))); \
+    sum_sad = _mm256_add_epi32(sum_sad, \
+              _mm256_add_epi32(sad1_reg, sad2_reg)); \
+    ref_ptr+= ref2_stride; \
+    src_ptr+= src2_stride; \
+    second_pred+= 64; \
+  } \
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8); \
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \
+  res = _mm_cvtsi128_si32(sum_sad128); \
+  return res; \
+}
+
+#define FSADAVG64 \
+FSADAVG64_H(64); \
+FSADAVG64_H(32);
+
+#define FSADAVG32 \
+FSADAVG32_H(64); \
+FSADAVG32_H(32); \
+FSADAVG32_H(16);
+
+FSADAVG64;
+FSADAVG32;
+
+#undef FSADAVG64
+#undef FSADAVG32
+#undef FSADAVG64_H
+#undef FSADAVG32_H
diff --git a/libvpx/libvpx/vpx_dsp/x86/sad_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/sad_sse2.asm
new file mode 100644
index 0000000..1ec906c
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/sad_sse2.asm
@@ -0,0 +1,268 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro SAD_FN 4
+%if %4 == 0
+%if %3 == 5
+cglobal sad%1x%2, 4, %3, 5, src, src_stride, ref, ref_stride, n_rows
+%else ; %3 == 7
+cglobal sad%1x%2, 4, %3, 6, src, src_stride, ref, ref_stride, \
+                            src_stride3, ref_stride3, n_rows
+%endif ; %3 == 5/7
+%else ; avg
+%if %3 == 5
+cglobal sad%1x%2_avg, 5, 1 + %3, 5, src, src_stride, ref, ref_stride, \
+                                    second_pred, n_rows
+%else ; %3 == 7
+cglobal sad%1x%2_avg, 5, ARCH_X86_64 + %3, 6, src, src_stride, \
+                                              ref, ref_stride, \
+                                              second_pred, \
+                                              src_stride3, ref_stride3
+%if ARCH_X86_64
+%define n_rowsd r7d
+%else ; x86-32
+%define n_rowsd dword r0m
+%endif ; x86-32/64
+%endif ; %3 == 5/7
+%endif ; avg/sad
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+%if %3 == 7
+  lea         src_stride3q, [src_strideq*3]
+  lea         ref_stride3q, [ref_strideq*3]
+%endif ; %3 == 7
+%endmacro
+
+; unsigned int vpx_sad64x64_sse2(uint8_t *src, int src_stride,
+;                                uint8_t *ref, int ref_stride);
+%macro SAD64XN 1-2 0
+  SAD_FN 64, %1, 5, %2
+  mov              n_rowsd, %1
+  pxor                  m0, m0
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+16]
+  psadbw                m3, [srcq+32]
+  psadbw                m4, [srcq+48]
+  paddd                 m1, m2
+  paddd                 m3, m4
+  add                 refq, ref_strideq
+  paddd                 m0, m1
+  add                 srcq, src_strideq
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD64XN 64 ; sad64x64_sse2
+SAD64XN 32 ; sad64x32_sse2
+SAD64XN 64, 1 ; sad64x64_avg_sse2
+SAD64XN 32, 1 ; sad64x32_avg_sse2
+
+; unsigned int vpx_sad32x32_sse2(uint8_t *src, int src_stride,
+;                                uint8_t *ref, int ref_stride);
+%macro SAD32XN 1-2 0
+  SAD_FN 32, %1, 5, %2
+  mov              n_rowsd, %1/2
+  pxor                  m0, m0
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+ref_strideq]
+  movu                  m4, [refq+ref_strideq+16]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+16]
+  psadbw                m3, [srcq+src_strideq]
+  psadbw                m4, [srcq+src_strideq+16]
+  paddd                 m1, m2
+  paddd                 m3, m4
+  lea                 refq, [refq+ref_strideq*2]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*2]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD32XN 64 ; sad32x64_sse2
+SAD32XN 32 ; sad32x32_sse2
+SAD32XN 16 ; sad32x16_sse2
+SAD32XN 64, 1 ; sad32x64_avg_sse2
+SAD32XN 32, 1 ; sad32x32_avg_sse2
+SAD32XN 16, 1 ; sad32x16_avg_sse2
+
+; unsigned int vpx_sad16x{8,16}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro SAD16XN 1-2 0
+  SAD_FN 16, %1, 7, %2
+  mov              n_rowsd, %1/4
+  pxor                  m0, m0
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+ref_strideq]
+  movu                  m3, [refq+ref_strideq*2]
+  movu                  m4, [refq+ref_stride3q]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+src_strideq]
+  psadbw                m3, [srcq+src_strideq*2]
+  psadbw                m4, [srcq+src_stride3q]
+  paddd                 m1, m2
+  paddd                 m3, m4
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD16XN 32 ; sad16x32_sse2
+SAD16XN 16 ; sad16x16_sse2
+SAD16XN  8 ; sad16x8_sse2
+SAD16XN 32, 1 ; sad16x32_avg_sse2
+SAD16XN 16, 1 ; sad16x16_avg_sse2
+SAD16XN  8, 1 ; sad16x8_avg_sse2
+
+; unsigned int vpx_sad8x{8,16}_sse2(uint8_t *src, int src_stride,
+;                                   uint8_t *ref, int ref_stride);
+%macro SAD8XN 1-2 0
+  SAD_FN 8, %1, 7, %2
+  mov              n_rowsd, %1/4
+  pxor                  m0, m0
+
+.loop:
+  movh                  m1, [refq]
+  movhps                m1, [refq+ref_strideq]
+  movh                  m2, [refq+ref_strideq*2]
+  movhps                m2, [refq+ref_stride3q]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  lea         second_predq, [second_predq+mmsize*2]
+%endif
+  movh                  m3, [srcq]
+  movhps                m3, [srcq+src_strideq]
+  movh                  m4, [srcq+src_strideq*2]
+  movhps                m4, [srcq+src_stride3q]
+  psadbw                m1, m3
+  psadbw                m2, m4
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  paddd                 m0, m2
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD8XN 16 ; sad8x16_sse2
+SAD8XN  8 ; sad8x8_sse2
+SAD8XN  4 ; sad8x4_sse2
+SAD8XN 16, 1 ; sad8x16_avg_sse2
+SAD8XN  8, 1 ; sad8x8_avg_sse2
+SAD8XN  4, 1 ; sad8x4_avg_sse2
+
+; unsigned int vpx_sad4x{4, 8}_sse2(uint8_t *src, int src_stride,
+;                                   uint8_t *ref, int ref_stride);
+%macro SAD4XN 1-2 0
+  SAD_FN 4, %1, 7, %2
+  mov              n_rowsd, %1/4
+  pxor                  m0, m0
+
+.loop:
+  movd                  m1, [refq]
+  movd                  m2, [refq+ref_strideq]
+  movd                  m3, [refq+ref_strideq*2]
+  movd                  m4, [refq+ref_stride3q]
+  punpckldq             m1, m2
+  punpckldq             m3, m4
+  movlhps               m1, m3
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  lea         second_predq, [second_predq+mmsize*1]
+%endif
+  movd                  m2, [srcq]
+  movd                  m5, [srcq+src_strideq]
+  movd                  m4, [srcq+src_strideq*2]
+  movd                  m3, [srcq+src_stride3q]
+  punpckldq             m2, m5
+  punpckldq             m4, m3
+  movlhps               m2, m4
+  psadbw                m1, m2
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD4XN  8 ; sad4x8_sse
+SAD4XN  4 ; sad4x4_sse
+SAD4XN  8, 1 ; sad4x8_avg_sse
+SAD4XN  4, 1 ; sad4x4_avg_sse
diff --git a/libvpx/libvpx/vpx_dsp/x86/sad_sse3.asm b/libvpx/libvpx/vpx_dsp/x86/sad_sse3.asm
new file mode 100644
index 0000000..18279bd
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/sad_sse3.asm
@@ -0,0 +1,374 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro STACK_FRAME_CREATE_X3 0
+%if ABI_IS_32BIT
+  %define     src_ptr       rsi
+  %define     src_stride    rax
+  %define     ref_ptr       rdi
+  %define     ref_stride    rdx
+  %define     end_ptr       rcx
+  %define     ret_var       rbx
+  %define     result_ptr    arg(4)
+  %define     height        dword ptr arg(4)
+    push        rbp
+    mov         rbp,        rsp
+    push        rsi
+    push        rdi
+    push        rbx
+
+    mov         rsi,        arg(0)              ; src_ptr
+    mov         rdi,        arg(2)              ; ref_ptr
+
+    movsxd      rax,        dword ptr arg(1)    ; src_stride
+    movsxd      rdx,        dword ptr arg(3)    ; ref_stride
+%else
+  %if LIBVPX_YASM_WIN64
+    SAVE_XMM 7, u
+    %define     src_ptr     rcx
+    %define     src_stride  rdx
+    %define     ref_ptr     r8
+    %define     ref_stride  r9
+    %define     end_ptr     r10
+    %define     ret_var     r11
+    %define     result_ptr  [rsp+xmm_stack_space+8+4*8]
+    %define     height      dword ptr [rsp+xmm_stack_space+8+4*8]
+  %else
+    %define     src_ptr     rdi
+    %define     src_stride  rsi
+    %define     ref_ptr     rdx
+    %define     ref_stride  rcx
+    %define     end_ptr     r9
+    %define     ret_var     r10
+    %define     result_ptr  r8
+    %define     height      r8
+  %endif
+%endif
+
+%endmacro
+
+%macro STACK_FRAME_DESTROY_X3 0
+  %define     src_ptr
+  %define     src_stride
+  %define     ref_ptr
+  %define     ref_stride
+  %define     end_ptr
+  %define     ret_var
+  %define     result_ptr
+  %define     height
+
+%if ABI_IS_32BIT
+    pop         rbx
+    pop         rdi
+    pop         rsi
+    pop         rbp
+%else
+  %if LIBVPX_YASM_WIN64
+    RESTORE_XMM
+  %endif
+%endif
+    ret
+%endmacro
+
+%macro PROCESS_16X2X3 5
+%if %1==0
+        movdqa          xmm0,       XMMWORD PTR [%2]
+        lddqu           xmm5,       XMMWORD PTR [%3]
+        lddqu           xmm6,       XMMWORD PTR [%3+1]
+        lddqu           xmm7,       XMMWORD PTR [%3+2]
+
+        psadbw          xmm5,       xmm0
+        psadbw          xmm6,       xmm0
+        psadbw          xmm7,       xmm0
+%else
+        movdqa          xmm0,       XMMWORD PTR [%2]
+        lddqu           xmm1,       XMMWORD PTR [%3]
+        lddqu           xmm2,       XMMWORD PTR [%3+1]
+        lddqu           xmm3,       XMMWORD PTR [%3+2]
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endif
+        movdqa          xmm0,       XMMWORD PTR [%2+%4]
+        lddqu           xmm1,       XMMWORD PTR [%3+%5]
+        lddqu           xmm2,       XMMWORD PTR [%3+%5+1]
+        lddqu           xmm3,       XMMWORD PTR [%3+%5+2]
+
+%if %1==0 || %1==1
+        lea             %2,         [%2+%4*2]
+        lea             %3,         [%3+%5*2]
+%endif
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endmacro
+
+%macro PROCESS_8X2X3 5
+%if %1==0
+        movq            mm0,       QWORD PTR [%2]
+        movq            mm5,       QWORD PTR [%3]
+        movq            mm6,       QWORD PTR [%3+1]
+        movq            mm7,       QWORD PTR [%3+2]
+
+        psadbw          mm5,       mm0
+        psadbw          mm6,       mm0
+        psadbw          mm7,       mm0
+%else
+        movq            mm0,       QWORD PTR [%2]
+        movq            mm1,       QWORD PTR [%3]
+        movq            mm2,       QWORD PTR [%3+1]
+        movq            mm3,       QWORD PTR [%3+2]
+
+        psadbw          mm1,       mm0
+        psadbw          mm2,       mm0
+        psadbw          mm3,       mm0
+
+        paddw           mm5,       mm1
+        paddw           mm6,       mm2
+        paddw           mm7,       mm3
+%endif
+        movq            mm0,       QWORD PTR [%2+%4]
+        movq            mm1,       QWORD PTR [%3+%5]
+        movq            mm2,       QWORD PTR [%3+%5+1]
+        movq            mm3,       QWORD PTR [%3+%5+2]
+
+%if %1==0 || %1==1
+        lea             %2,        [%2+%4*2]
+        lea             %3,        [%3+%5*2]
+%endif
+
+        psadbw          mm1,       mm0
+        psadbw          mm2,       mm0
+        psadbw          mm3,       mm0
+
+        paddw           mm5,       mm1
+        paddw           mm6,       mm2
+        paddw           mm7,       mm3
+%endmacro
+
+;void int vpx_sad16x16x3_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(vpx_sad16x16x3_sse3) PRIVATE
+sym(vpx_sad16x16x3_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+        PROCESS_16X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
+
+        mov             rcx,        result_ptr
+
+        movq            xmm0,       xmm5
+        psrldq          xmm5,       8
+
+        paddw           xmm0,       xmm5
+        movd            [rcx],      xmm0
+;-
+        movq            xmm0,       xmm6
+        psrldq          xmm6,       8
+
+        paddw           xmm0,       xmm6
+        movd            [rcx+4],    xmm0
+;-
+        movq            xmm0,       xmm7
+        psrldq          xmm7,       8
+
+        paddw           xmm0,       xmm7
+        movd            [rcx+8],    xmm0
+
+    STACK_FRAME_DESTROY_X3
+
+;void int vpx_sad16x8x3_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(vpx_sad16x8x3_sse3) PRIVATE
+sym(vpx_sad16x8x3_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+        PROCESS_16X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
+
+        mov             rcx,        result_ptr
+
+        movq            xmm0,       xmm5
+        psrldq          xmm5,       8
+
+        paddw           xmm0,       xmm5
+        movd            [rcx],      xmm0
+;-
+        movq            xmm0,       xmm6
+        psrldq          xmm6,       8
+
+        paddw           xmm0,       xmm6
+        movd            [rcx+4],    xmm0
+;-
+        movq            xmm0,       xmm7
+        psrldq          xmm7,       8
+
+        paddw           xmm0,       xmm7
+        movd            [rcx+8],    xmm0
+
+    STACK_FRAME_DESTROY_X3
+
+;void int vpx_sad8x16x3_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(vpx_sad8x16x3_sse3) PRIVATE
+sym(vpx_sad8x16x3_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+        PROCESS_8X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
+
+        mov             rcx,        result_ptr
+
+        punpckldq       mm5,        mm6
+
+        movq            [rcx],      mm5
+        movd            [rcx+8],    mm7
+
+    STACK_FRAME_DESTROY_X3
+
+;void int vpx_sad8x8x3_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(vpx_sad8x8x3_sse3) PRIVATE
+sym(vpx_sad8x8x3_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+        PROCESS_8X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
+
+        mov             rcx,        result_ptr
+
+        punpckldq       mm5,        mm6
+
+        movq            [rcx],      mm5
+        movd            [rcx+8],    mm7
+
+    STACK_FRAME_DESTROY_X3
+
+;void int vpx_sad4x4x3_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(vpx_sad4x4x3_sse3) PRIVATE
+sym(vpx_sad4x4x3_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+        movd            mm0,        DWORD PTR [src_ptr]
+        movd            mm1,        DWORD PTR [ref_ptr]
+
+        movd            mm2,        DWORD PTR [src_ptr+src_stride]
+        movd            mm3,        DWORD PTR [ref_ptr+ref_stride]
+
+        punpcklbw       mm0,        mm2
+        punpcklbw       mm1,        mm3
+
+        movd            mm4,        DWORD PTR [ref_ptr+1]
+        movd            mm5,        DWORD PTR [ref_ptr+2]
+
+        movd            mm2,        DWORD PTR [ref_ptr+ref_stride+1]
+        movd            mm3,        DWORD PTR [ref_ptr+ref_stride+2]
+
+        psadbw          mm1,        mm0
+
+        punpcklbw       mm4,        mm2
+        punpcklbw       mm5,        mm3
+
+        psadbw          mm4,        mm0
+        psadbw          mm5,        mm0
+
+        lea             src_ptr,    [src_ptr+src_stride*2]
+        lea             ref_ptr,    [ref_ptr+ref_stride*2]
+
+        movd            mm0,        DWORD PTR [src_ptr]
+        movd            mm2,        DWORD PTR [ref_ptr]
+
+        movd            mm3,        DWORD PTR [src_ptr+src_stride]
+        movd            mm6,        DWORD PTR [ref_ptr+ref_stride]
+
+        punpcklbw       mm0,        mm3
+        punpcklbw       mm2,        mm6
+
+        movd            mm3,        DWORD PTR [ref_ptr+1]
+        movd            mm7,        DWORD PTR [ref_ptr+2]
+
+        psadbw          mm2,        mm0
+
+        paddw           mm1,        mm2
+
+        movd            mm2,        DWORD PTR [ref_ptr+ref_stride+1]
+        movd            mm6,        DWORD PTR [ref_ptr+ref_stride+2]
+
+        punpcklbw       mm3,        mm2
+        punpcklbw       mm7,        mm6
+
+        psadbw          mm3,        mm0
+        psadbw          mm7,        mm0
+
+        paddw           mm3,        mm4
+        paddw           mm7,        mm5
+
+        mov             rcx,        result_ptr
+
+        punpckldq       mm1,        mm3
+
+        movq            [rcx],      mm1
+        movd            [rcx+8],    mm7
+
+    STACK_FRAME_DESTROY_X3
diff --git a/libvpx/libvpx/vpx_dsp/x86/sad_sse4.asm b/libvpx/libvpx/vpx_dsp/x86/sad_sse4.asm
new file mode 100644
index 0000000..bc67447
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/sad_sse4.asm
@@ -0,0 +1,359 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro PROCESS_16X2X8 1
+%if %1
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        movq            xmm1,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        movq            xmm2,       MMWORD PTR [rdi+16]
+        punpcklqdq      xmm1,       xmm3
+        punpcklqdq      xmm3,       xmm2
+
+        movdqa          xmm2,       xmm1
+        mpsadbw         xmm1,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+
+        psrldq          xmm0,       8
+
+        movdqa          xmm4,       xmm3
+        mpsadbw         xmm3,       xmm0,  0x0
+        mpsadbw         xmm4,       xmm0,  0x5
+
+        paddw           xmm1,       xmm2
+        paddw           xmm1,       xmm3
+        paddw           xmm1,       xmm4
+%else
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        movq            xmm5,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        movq            xmm2,       MMWORD PTR [rdi+16]
+        punpcklqdq      xmm5,       xmm3
+        punpcklqdq      xmm3,       xmm2
+
+        movdqa          xmm2,       xmm5
+        mpsadbw         xmm5,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+
+        psrldq          xmm0,       8
+
+        movdqa          xmm4,       xmm3
+        mpsadbw         xmm3,       xmm0,  0x0
+        mpsadbw         xmm4,       xmm0,  0x5
+
+        paddw           xmm5,       xmm2
+        paddw           xmm5,       xmm3
+        paddw           xmm5,       xmm4
+
+        paddw           xmm1,       xmm5
+%endif
+        movdqa          xmm0,       XMMWORD PTR [rsi + rax]
+        movq            xmm5,       MMWORD PTR [rdi+ rdx]
+        movq            xmm3,       MMWORD PTR [rdi+ rdx+8]
+        movq            xmm2,       MMWORD PTR [rdi+ rdx+16]
+        punpcklqdq      xmm5,       xmm3
+        punpcklqdq      xmm3,       xmm2
+
+        lea             rsi,        [rsi+rax*2]
+        lea             rdi,        [rdi+rdx*2]
+
+        movdqa          xmm2,       xmm5
+        mpsadbw         xmm5,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+
+        psrldq          xmm0,       8
+        movdqa          xmm4,       xmm3
+        mpsadbw         xmm3,       xmm0,  0x0
+        mpsadbw         xmm4,       xmm0,  0x5
+
+        paddw           xmm5,       xmm2
+        paddw           xmm5,       xmm3
+        paddw           xmm5,       xmm4
+
+        paddw           xmm1,       xmm5
+%endmacro
+
+%macro PROCESS_8X2X8 1
+%if %1
+        movq            xmm0,       MMWORD PTR [rsi]
+        movq            xmm1,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        punpcklqdq      xmm1,       xmm3
+
+        movdqa          xmm2,       xmm1
+        mpsadbw         xmm1,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+        paddw           xmm1,       xmm2
+%else
+        movq            xmm0,       MMWORD PTR [rsi]
+        movq            xmm5,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        punpcklqdq      xmm5,       xmm3
+
+        movdqa          xmm2,       xmm5
+        mpsadbw         xmm5,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+        paddw           xmm5,       xmm2
+
+        paddw           xmm1,       xmm5
+%endif
+        movq            xmm0,       MMWORD PTR [rsi + rax]
+        movq            xmm5,       MMWORD PTR [rdi+ rdx]
+        movq            xmm3,       MMWORD PTR [rdi+ rdx+8]
+        punpcklqdq      xmm5,       xmm3
+
+        lea             rsi,        [rsi+rax*2]
+        lea             rdi,        [rdi+rdx*2]
+
+        movdqa          xmm2,       xmm5
+        mpsadbw         xmm5,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+        paddw           xmm5,       xmm2
+
+        paddw           xmm1,       xmm5
+%endmacro
+
+%macro PROCESS_4X2X8 1
+%if %1
+        movd            xmm0,       [rsi]
+        movq            xmm1,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        punpcklqdq      xmm1,       xmm3
+
+        mpsadbw         xmm1,       xmm0,  0x0
+%else
+        movd            xmm0,       [rsi]
+        movq            xmm5,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        punpcklqdq      xmm5,       xmm3
+
+        mpsadbw         xmm5,       xmm0,  0x0
+
+        paddw           xmm1,       xmm5
+%endif
+        movd            xmm0,       [rsi + rax]
+        movq            xmm5,       MMWORD PTR [rdi+ rdx]
+        movq            xmm3,       MMWORD PTR [rdi+ rdx+8]
+        punpcklqdq      xmm5,       xmm3
+
+        lea             rsi,        [rsi+rax*2]
+        lea             rdi,        [rdi+rdx*2]
+
+        mpsadbw         xmm5,       xmm0,  0x0
+
+        paddw           xmm1,       xmm5
+%endmacro
+
+%macro WRITE_AS_INTS 0
+    mov             rdi,        arg(4)           ;Results
+    pxor            xmm0, xmm0
+    movdqa          xmm2, xmm1
+    punpcklwd       xmm1, xmm0
+    punpckhwd       xmm2, xmm0
+
+    movdqa          [rdi],    xmm1
+    movdqa          [rdi + 16],    xmm2
+%endmacro
+
+;void vpx_sad16x16x8_sse4_1(
+;    const unsigned char *src_ptr,
+;    int  src_stride,
+;    const unsigned char *ref_ptr,
+;    int  ref_stride,
+;    unsigned short *sad_array);
+global sym(vpx_sad16x16x8_sse4_1) PRIVATE
+sym(vpx_sad16x16x8_sse4_1):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0)           ;src_ptr
+    mov             rdi,        arg(2)           ;ref_ptr
+
+    movsxd          rax,        dword ptr arg(1) ;src_stride
+    movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+    PROCESS_16X2X8 1
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+
+    WRITE_AS_INTS
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vpx_sad16x8x8_sse4_1(
+;    const unsigned char *src_ptr,
+;    int  src_stride,
+;    const unsigned char *ref_ptr,
+;    int  ref_stride,
+;    unsigned short *sad_array
+;);
+global sym(vpx_sad16x8x8_sse4_1) PRIVATE
+sym(vpx_sad16x8x8_sse4_1):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0)           ;src_ptr
+    mov             rdi,        arg(2)           ;ref_ptr
+
+    movsxd          rax,        dword ptr arg(1) ;src_stride
+    movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+    PROCESS_16X2X8 1
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+
+    WRITE_AS_INTS
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vpx_sad8x8x8_sse4_1(
+;    const unsigned char *src_ptr,
+;    int  src_stride,
+;    const unsigned char *ref_ptr,
+;    int  ref_stride,
+;    unsigned short *sad_array
+;);
+global sym(vpx_sad8x8x8_sse4_1) PRIVATE
+sym(vpx_sad8x8x8_sse4_1):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0)           ;src_ptr
+    mov             rdi,        arg(2)           ;ref_ptr
+
+    movsxd          rax,        dword ptr arg(1) ;src_stride
+    movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+    PROCESS_8X2X8 1
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+
+    WRITE_AS_INTS
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vpx_sad8x16x8_sse4_1(
+;    const unsigned char *src_ptr,
+;    int  src_stride,
+;    const unsigned char *ref_ptr,
+;    int  ref_stride,
+;    unsigned short *sad_array
+;);
+global sym(vpx_sad8x16x8_sse4_1) PRIVATE
+sym(vpx_sad8x16x8_sse4_1):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0)           ;src_ptr
+    mov             rdi,        arg(2)           ;ref_ptr
+
+    movsxd          rax,        dword ptr arg(1) ;src_stride
+    movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+    PROCESS_8X2X8 1
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+
+    WRITE_AS_INTS
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vpx_sad4x4x8_sse4_1(
+;    const unsigned char *src_ptr,
+;    int  src_stride,
+;    const unsigned char *ref_ptr,
+;    int  ref_stride,
+;    unsigned short *sad_array
+;);
+global sym(vpx_sad4x4x8_sse4_1) PRIVATE
+sym(vpx_sad4x4x8_sse4_1):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0)           ;src_ptr
+    mov             rdi,        arg(2)           ;ref_ptr
+
+    movsxd          rax,        dword ptr arg(1) ;src_stride
+    movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+    PROCESS_4X2X8 1
+    PROCESS_4X2X8 0
+
+    WRITE_AS_INTS
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+
+
diff --git a/libvpx/libvpx/vpx_dsp/x86/sad_ssse3.asm b/libvpx/libvpx/vpx_dsp/x86/sad_ssse3.asm
new file mode 100644
index 0000000..49f204f
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/sad_ssse3.asm
@@ -0,0 +1,370 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro PROCESS_16X2X3 1
+%if %1
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        lddqu           xmm5,       XMMWORD PTR [rdi]
+        lddqu           xmm6,       XMMWORD PTR [rdi+1]
+        lddqu           xmm7,       XMMWORD PTR [rdi+2]
+
+        psadbw          xmm5,       xmm0
+        psadbw          xmm6,       xmm0
+        psadbw          xmm7,       xmm0
+%else
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        lddqu           xmm1,       XMMWORD PTR [rdi]
+        lddqu           xmm2,       XMMWORD PTR [rdi+1]
+        lddqu           xmm3,       XMMWORD PTR [rdi+2]
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endif
+        movdqa          xmm0,       XMMWORD PTR [rsi+rax]
+        lddqu           xmm1,       XMMWORD PTR [rdi+rdx]
+        lddqu           xmm2,       XMMWORD PTR [rdi+rdx+1]
+        lddqu           xmm3,       XMMWORD PTR [rdi+rdx+2]
+
+        lea             rsi,        [rsi+rax*2]
+        lea             rdi,        [rdi+rdx*2]
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endmacro
+
+%macro PROCESS_16X2X3_OFFSET 2
+%if %1
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        movdqa          xmm4,       XMMWORD PTR [rdi]
+        movdqa          xmm7,       XMMWORD PTR [rdi+16]
+
+        movdqa          xmm5,       xmm7
+        palignr         xmm5,       xmm4,       %2
+
+        movdqa          xmm6,       xmm7
+        palignr         xmm6,       xmm4,       (%2+1)
+
+        palignr         xmm7,       xmm4,       (%2+2)
+
+        psadbw          xmm5,       xmm0
+        psadbw          xmm6,       xmm0
+        psadbw          xmm7,       xmm0
+%else
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        movdqa          xmm4,       XMMWORD PTR [rdi]
+        movdqa          xmm3,       XMMWORD PTR [rdi+16]
+
+        movdqa          xmm1,       xmm3
+        palignr         xmm1,       xmm4,       %2
+
+        movdqa          xmm2,       xmm3
+        palignr         xmm2,       xmm4,       (%2+1)
+
+        palignr         xmm3,       xmm4,       (%2+2)
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endif
+        movdqa          xmm0,       XMMWORD PTR [rsi+rax]
+        movdqa          xmm4,       XMMWORD PTR [rdi+rdx]
+        movdqa          xmm3,       XMMWORD PTR [rdi+rdx+16]
+
+        movdqa          xmm1,       xmm3
+        palignr         xmm1,       xmm4,       %2
+
+        movdqa          xmm2,       xmm3
+        palignr         xmm2,       xmm4,       (%2+1)
+
+        palignr         xmm3,       xmm4,       (%2+2)
+
+        lea             rsi,        [rsi+rax*2]
+        lea             rdi,        [rdi+rdx*2]
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endmacro
+
+%macro PROCESS_16X16X3_OFFSET 2
+%2_aligned_by_%1:
+
+        sub             rdi,        %1
+
+        PROCESS_16X2X3_OFFSET 1, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+
+        jmp             %2_store_off
+
+%endmacro
+
+%macro PROCESS_16X8X3_OFFSET 2
+%2_aligned_by_%1:
+
+        sub             rdi,        %1
+
+        PROCESS_16X2X3_OFFSET 1, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+
+        jmp             %2_store_off
+
+%endmacro
+
+;void int vpx_sad16x16x3_ssse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(vpx_sad16x16x3_ssse3) PRIVATE
+sym(vpx_sad16x16x3_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rcx
+    ; end prolog
+
+        mov             rsi,        arg(0) ;src_ptr
+        mov             rdi,        arg(2) ;ref_ptr
+
+        mov             rdx,        0xf
+        and             rdx,        rdi
+
+        jmp .vpx_sad16x16x3_ssse3_skiptable
+.vpx_sad16x16x3_ssse3_jumptable:
+        dd .vpx_sad16x16x3_ssse3_aligned_by_0  - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_1  - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_2  - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_3  - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_4  - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_5  - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_6  - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_7  - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_8  - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_9  - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_10 - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_11 - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_12 - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_13 - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_14 - .vpx_sad16x16x3_ssse3_do_jump
+        dd .vpx_sad16x16x3_ssse3_aligned_by_15 - .vpx_sad16x16x3_ssse3_do_jump
+.vpx_sad16x16x3_ssse3_skiptable:
+
+        call .vpx_sad16x16x3_ssse3_do_jump
+.vpx_sad16x16x3_ssse3_do_jump:
+        pop             rcx                         ; get the address of do_jump
+        mov             rax,  .vpx_sad16x16x3_ssse3_jumptable - .vpx_sad16x16x3_ssse3_do_jump
+        add             rax,  rcx  ; get the absolute address of vpx_sad16x16x3_ssse3_jumptable
+
+        movsxd          rax,  dword [rax + 4*rdx]   ; get the 32 bit offset from the jumptable
+        add             rcx,        rax
+
+        movsxd          rax,        dword ptr arg(1) ;src_stride
+        movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+        jmp             rcx
+
+        PROCESS_16X16X3_OFFSET 0,  .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 1,  .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 2,  .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 3,  .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 4,  .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 5,  .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 6,  .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 7,  .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 8,  .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 9,  .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 10, .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 11, .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 12, .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 13, .vpx_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 14, .vpx_sad16x16x3_ssse3
+
+.vpx_sad16x16x3_ssse3_aligned_by_15:
+        PROCESS_16X2X3 1
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+
+.vpx_sad16x16x3_ssse3_store_off:
+        mov             rdi,        arg(4) ;Results
+
+        movq            xmm0,       xmm5
+        psrldq          xmm5,       8
+
+        paddw           xmm0,       xmm5
+        movd            [rdi],      xmm0
+;-
+        movq            xmm0,       xmm6
+        psrldq          xmm6,       8
+
+        paddw           xmm0,       xmm6
+        movd            [rdi+4],    xmm0
+;-
+        movq            xmm0,       xmm7
+        psrldq          xmm7,       8
+
+        paddw           xmm0,       xmm7
+        movd            [rdi+8],    xmm0
+
+    ; begin epilog
+    pop         rcx
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void int vpx_sad16x8x3_ssse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(vpx_sad16x8x3_ssse3) PRIVATE
+sym(vpx_sad16x8x3_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rcx
+    ; end prolog
+
+        mov             rsi,        arg(0) ;src_ptr
+        mov             rdi,        arg(2) ;ref_ptr
+
+        mov             rdx,        0xf
+        and             rdx,        rdi
+
+        jmp .vpx_sad16x8x3_ssse3_skiptable
+.vpx_sad16x8x3_ssse3_jumptable:
+        dd .vpx_sad16x8x3_ssse3_aligned_by_0  - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_1  - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_2  - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_3  - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_4  - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_5  - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_6  - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_7  - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_8  - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_9  - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_10 - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_11 - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_12 - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_13 - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_14 - .vpx_sad16x8x3_ssse3_do_jump
+        dd .vpx_sad16x8x3_ssse3_aligned_by_15 - .vpx_sad16x8x3_ssse3_do_jump
+.vpx_sad16x8x3_ssse3_skiptable:
+
+        call .vpx_sad16x8x3_ssse3_do_jump
+.vpx_sad16x8x3_ssse3_do_jump:
+        pop             rcx                         ; get the address of do_jump
+        mov             rax,  .vpx_sad16x8x3_ssse3_jumptable - .vpx_sad16x8x3_ssse3_do_jump
+        add             rax,  rcx  ; get the absolute address of vpx_sad16x8x3_ssse3_jumptable
+
+        movsxd          rax,  dword [rax + 4*rdx]   ; get the 32 bit offset from the jumptable
+        add             rcx,        rax
+
+        movsxd          rax,        dword ptr arg(1) ;src_stride
+        movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+        jmp             rcx
+
+        PROCESS_16X8X3_OFFSET 0,  .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 1,  .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 2,  .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 3,  .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 4,  .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 5,  .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 6,  .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 7,  .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 8,  .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 9,  .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 10, .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 11, .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 12, .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 13, .vpx_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 14, .vpx_sad16x8x3_ssse3
+
+.vpx_sad16x8x3_ssse3_aligned_by_15:
+
+        PROCESS_16X2X3 1
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+
+.vpx_sad16x8x3_ssse3_store_off:
+        mov             rdi,        arg(4) ;Results
+
+        movq            xmm0,       xmm5
+        psrldq          xmm5,       8
+
+        paddw           xmm0,       xmm5
+        movd            [rdi],      xmm0
+;-
+        movq            xmm0,       xmm6
+        psrldq          xmm6,       8
+
+        paddw           xmm0,       xmm6
+        movd            [rdi+4],    xmm0
+;-
+        movq            xmm0,       xmm7
+        psrldq          xmm7,       8
+
+        paddw           xmm0,       xmm7
+        movd            [rdi+8],    xmm0
+
+    ; begin epilog
+    pop         rcx
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vpx_dsp/x86/ssim_opt_x86_64.asm b/libvpx/libvpx/vpx_dsp/x86/ssim_opt_x86_64.asm
new file mode 100644
index 0000000..6d58321
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/ssim_opt_x86_64.asm
@@ -0,0 +1,216 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+; tabulate_ssim - sums sum_s,sum_r,sum_sq_s,sum_sq_r, sum_sxr
+%macro TABULATE_SSIM 0
+        paddusw         xmm15, xmm3  ; sum_s
+        paddusw         xmm14, xmm4  ; sum_r
+        movdqa          xmm1, xmm3
+        pmaddwd         xmm1, xmm1
+        paddd           xmm13, xmm1 ; sum_sq_s
+        movdqa          xmm2, xmm4
+        pmaddwd         xmm2, xmm2
+        paddd           xmm12, xmm2 ; sum_sq_r
+        pmaddwd         xmm3, xmm4
+        paddd           xmm11, xmm3  ; sum_sxr
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_Q 1
+        movdqa          xmm2,%1
+        punpckldq       %1,xmm0
+        punpckhdq       xmm2,xmm0
+        paddq           %1,xmm2
+        movdqa          xmm2,%1
+        punpcklqdq      %1,xmm0
+        punpckhqdq      xmm2,xmm0
+        paddq           %1,xmm2
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_W 1
+        movdqa          xmm1, %1
+        punpcklwd       %1,xmm0
+        punpckhwd       xmm1,xmm0
+        paddd           %1, xmm1
+        SUM_ACROSS_Q    %1
+%endmacro
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    uint32_t *sum_s,
+;    uint32_t *sum_r,
+;    uint32_t *sum_sq_s,
+;    uint32_t *sum_sq_r,
+;    uint32_t *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+global sym(vpx_ssim_parms_16x16_sse2) PRIVATE
+sym(vpx_ssim_parms_16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 16      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movdqu          xmm5, [rsi]
+    movdqu          xmm6, [rdi]
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpckhbw       xmm3, xmm0 ; high_s
+    punpckhbw       xmm4, xmm0 ; high_r
+
+    TABULATE_SSIM
+
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    uint32_t *sum_s,
+;    uint32_t *sum_r,
+;    uint32_t *sum_sq_s,
+;    uint32_t *sum_sq_r,
+;    uint32_t *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+global sym(vpx_ssim_parms_8x8_sse2) PRIVATE
+sym(vpx_ssim_parms_8x8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 8      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movq            xmm3, [rsi]
+    movq            xmm4, [rdi]
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vpx_dsp/x86/subpel_variance_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/subpel_variance_sse2.asm
new file mode 100644
index 0000000..cee4468
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/subpel_variance_sse2.asm
@@ -0,0 +1,1486 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_8: times  8 dw  8
+bilin_filter_m_sse2: times  8 dw 16
+                     times  8 dw  0
+                     times  8 dw 14
+                     times  8 dw  2
+                     times  8 dw 12
+                     times  8 dw  4
+                     times  8 dw 10
+                     times  8 dw  6
+                     times 16 dw  8
+                     times  8 dw  6
+                     times  8 dw 10
+                     times  8 dw  4
+                     times  8 dw 12
+                     times  8 dw  2
+                     times  8 dw 14
+
+bilin_filter_m_ssse3: times  8 db 16,  0
+                      times  8 db 14,  2
+                      times  8 db 12,  4
+                      times  8 db 10,  6
+                      times 16 db  8
+                      times  8 db  6, 10
+                      times  8 db  4, 12
+                      times  8 db  2, 14
+
+SECTION .text
+
+; int vpx_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride,
+;                               int x_offset, int y_offset,
+;                               const uint8_t *dst, ptrdiff_t dst_stride,
+;                               int height, unsigned int *sse);
+;
+; This function returns the SE and stores SSE in the given pointer.
+
+%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse
+  psubw                %3, %4
+  psubw                %1, %2
+  paddw                %5, %3
+  pmaddwd              %3, %3
+  paddw                %5, %1
+  pmaddwd              %1, %1
+  paddd                %6, %3
+  paddd                %6, %1
+%endmacro
+
+%macro STORE_AND_RET 1
+%if %1 > 4
+  ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit
+  ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg.
+  ; We have to sign-extend it before adding the words within the register
+  ; and outputing to a dword.
+  pcmpgtw              m5, m6           ; mask for 0 > x
+  movhlps              m3, m7
+  punpcklwd            m4, m6, m5
+  punpckhwd            m6, m5           ; sign-extend m6 word->dword
+  paddd                m7, m3
+  paddd                m6, m4
+  pshufd               m3, m7, 0x1
+  movhlps              m4, m6
+  paddd                m7, m3
+  paddd                m6, m4
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  pshufd               m4, m6, 0x1
+  movd               [r1], m7           ; store sse
+  paddd                m6, m4
+  movd               raxd, m6           ; store sum as return value
+%else ; 4xh
+  pshuflw              m4, m6, 0xe
+  pshuflw              m3, m7, 0xe
+  paddw                m6, m4
+  paddd                m7, m3
+  pcmpgtw              m5, m6           ; mask for 0 > x
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  punpcklwd            m6, m5           ; sign-extend m6 word->dword
+  movd               [r1], m7           ; store sse
+  pshuflw              m4, m6, 0xe
+  paddd                m6, m4
+  movd               raxd, m6           ; store sum as return value
+%endif
+  RET
+%endmacro
+
+%macro INC_SRC_BY_SRC_STRIDE  0
+%if ARCH_X86=1 && CONFIG_PIC=1
+  add                srcq, src_stridemp
+%else
+  add                srcq, src_strideq
+%endif
+%endmacro
+
+%macro SUBPEL_VARIANCE 1-2 0 ; W
+%if cpuflag(ssse3)
+%define bilin_filter_m bilin_filter_m_ssse3
+%define filter_idx_shift 4
+%else
+%define bilin_filter_m bilin_filter_m_sse2
+%define filter_idx_shift 5
+%endif
+; FIXME(rbultje) only bilinear filters use >8 registers, and ssse3 only uses
+; 11, not 13, if the registers are ordered correctly. May make a minor speed
+; difference on Win64
+
+%ifdef PIC    ; 64bit PIC
+  %if %2 == 1 ; avg
+    cglobal sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \
+                                      x_offset, y_offset, \
+                                      dst, dst_stride, \
+                                      sec, sec_stride, height, sse
+    %define sec_str sec_strideq
+  %else
+    cglobal sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, x_offset, \
+                                  y_offset, dst, dst_stride, height, sse
+  %endif
+  %define block_height heightd
+  %define bilin_filter sseq
+%else
+  %if ARCH_X86=1 && CONFIG_PIC=1
+    %if %2 == 1 ; avg
+      cglobal sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \
+                                  x_offset, y_offset, \
+                                  dst, dst_stride, \
+                                  sec, sec_stride, \
+                                  height, sse, g_bilin_filter, g_pw_8
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+
+      ;Store bilin_filter and pw_8 location in stack
+      %if GET_GOT_DEFINED == 1
+        GET_GOT eax
+        add esp, 4                ; restore esp
+      %endif
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %else
+      cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, x_offset, \
+                                y_offset, dst, dst_stride, height, sse, \
+                                g_bilin_filter, g_pw_8
+      %define block_height heightd
+
+      ;Store bilin_filter and pw_8 location in stack
+      %if GET_GOT_DEFINED == 1
+        GET_GOT eax
+        add esp, 4                ; restore esp
+      %endif
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %endif
+  %else
+    %if %2 == 1 ; avg
+      cglobal sub_pixel_avg_variance%1xh, 7 + 2 * ARCH_X86_64, \
+                        7 + 2 * ARCH_X86_64, 13, src, src_stride, \
+                                             x_offset, y_offset, \
+                                             dst, dst_stride, \
+                                             sec, sec_stride, \
+                                             height, sse
+      %if ARCH_X86_64
+      %define block_height heightd
+      %define sec_str sec_strideq
+      %else
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+      %endif
+    %else
+      cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, x_offset, \
+                              y_offset, dst, dst_stride, height, sse
+      %define block_height heightd
+    %endif
+
+    %define bilin_filter bilin_filter_m
+  %endif
+%endif
+
+%if %1 == 4
+  %define movx movd
+%else
+  %define movx movh
+%endif
+
+  ASSERT               %1 <= 16         ; m6 overflows if w > 16
+  pxor                 m6, m6           ; sum
+  pxor                 m7, m7           ; sse
+  ; FIXME(rbultje) if both filters are bilinear, we don't actually use m5; we
+  ; could perhaps use it for something more productive then
+  pxor                 m5, m5           ; dedicated zero register
+%if %1 < 16
+  sar                   block_height, 1
+%if %2 == 1 ; avg
+  shl             sec_str, 1
+%endif
+%endif
+
+  ; FIXME(rbultje) replace by jumptable?
+  test          x_offsetd, x_offsetd
+  jnz .x_nonzero
+  ; x_offset == 0
+  test          y_offsetd, y_offsetd
+  jnz .x_zero_y_nonzero
+
+  ; x_offset == 0 && y_offset == 0
+.x_zero_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  mova                 m1, [dstq]
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%endif
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+
+%if %2 == 0 ; !avg
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m0, [srcq+src_strideq]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m0, m1
+%endif
+%else ; !avg
+  movx                 m2, [srcq+src_strideq]
+%endif
+
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+
+%if %2 == 1 ; avg
+%if %1 > 4
+  pavgb                m0, [secq]
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+%endif
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%if %1 > 4
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_zero_y_zero_loop
+  STORE_AND_RET %1
+
+.x_zero_y_nonzero:
+  cmp           y_offsetd, 4
+  jne .x_zero_y_nonhalf
+
+  ; x_offset == 0 && y_offset == 0.5
+.x_zero_y_half_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+src_strideq]
+  mova                 m1, [dstq]
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+%endif
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m2, [srcq+src_strideq]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m2, [srcq+src_strideq*2]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq*2]
+  punpckldq            m2, m1
+%endif
+  movx                 m1, [dstq]
+%if %1 > 4
+  movlhps              m0, m2
+%else ; 4xh
+  punpckldq            m0, m2
+%endif
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m0, m2
+  punpcklbw            m1, m5
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpcklbw            m3, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m4, [secq]
+  pavgb                m0, m4
+  punpcklbw            m3, m5
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  movx                 m4, [srcq+src_strideq*2]
+  movx                 m1, [dstq]
+  pavgb                m0, m2
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m2, m4
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_zero_y_half_loop
+  STORE_AND_RET %1
+
+.x_zero_y_nonhalf:
+  ; x_offset == 0 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else ; x86-32 or mmx
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0, reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_zero_y_other_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+src_strideq]
+  mova                 m1, [dstq]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  punpcklbw            m0, m5
+  punpcklbw            m4, m5
+  ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can
+  ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of
+  ; instructions is the same (5), but it is 1 mul instead of 2, so might be
+  ; slightly faster because of pmullw latency. It would also cut our rodata
+  ; tables in half for this function, and save 1-2 registers on x86-64.
+  pmullw               m2, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m2, m3
+  paddw                m0, m4
+%endif
+  psraw                m2, 4
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m2, [srcq+src_strideq]
+  movx                 m4, [srcq+src_strideq*2]
+  movx                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  movx                 m1, [dstq]
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m1, m2, filter_y_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, m1
+  paddw                m2, filter_rnd
+  movx                 m1, [dstq]
+  paddw                m2, m4
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_zero_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonzero:
+  cmp           x_offsetd, 4
+  jne .x_nonhalf
+  ; x_offset == 0.5
+  test          y_offsetd, y_offsetd
+  jnz .x_half_y_nonzero
+
+  ; x_offset == 0.5 && y_offset == 0
+.x_half_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+%endif
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m4, [srcq+1]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m0, [srcq+src_strideq]
+  movhps               m4, [srcq+src_strideq+1]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m0, m1
+  movx                 m2, [srcq+src_strideq+1]
+  punpckldq            m4, m2
+%endif
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m0, m4
+  punpcklbw            m3, m5
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m1, m5
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  movx                 m2, [srcq+src_strideq]
+  movx                 m1, [dstq]
+  pavgb                m0, m4
+  movx                 m4, [srcq+src_strideq+1]
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m2, m4
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_half_y_zero_loop
+  STORE_AND_RET %1
+
+.x_half_y_nonzero:
+  cmp           y_offsetd, 4
+  jne .x_half_y_nonhalf
+
+  ; x_offset == 0.5 && y_offset == 0.5
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_half_loop:
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m4, m3
+  punpckhbw            m3, m1, m5
+  pavgb                m0, m4
+%if %2 == 1 ; avg
+  punpcklbw            m1, m5
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_half_loop:
+  movx                 m2, [srcq]
+  movx                 m3, [srcq+1]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m2, [srcq+src_strideq]
+  movhps               m3, [srcq+src_strideq+1]
+%else
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m2, m1
+  movx                 m1, [srcq+src_strideq+1]
+  punpckldq            m3, m1
+%endif
+  pavgb                m2, m3
+%if %1 > 4
+  movlhps              m0, m2
+  movhlps              m4, m2
+%else ; 4xh
+  punpckldq            m0, m2
+  pshuflw              m4, m2, 0xe
+%endif
+  movx                 m1, [dstq]
+  pavgb                m0, m2
+  movx                 m3, [dstq+dst_strideq]
+%if %1 > 4
+  pavgb                m0, [secq]
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+%endif
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%if %1 > 4
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  movx                 m4, [srcq+src_strideq]
+  movx                 m1, [srcq+src_strideq+1]
+  pavgb                m2, m3
+  pavgb                m4, m1
+  pavgb                m0, m2
+  pavgb                m2, m4
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_half_y_half_loop
+  STORE_AND_RET %1
+
+.x_half_y_nonhalf:
+  ; x_offset == 0.5 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else  ;x86_32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0.5. We can reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_other_loop:
+  movu                 m4, [srcq]
+  movu                 m2, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m4, m2
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  pmullw               m2, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, filter_rnd
+  punpcklbw            m0, m5
+  paddw                m2, m3
+  punpcklbw            m3, m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  paddw                m0, m3
+%endif
+  punpckhbw            m3, m1, m5
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+%if notcpuflag(ssse3)
+  punpcklbw            m0, m5
+%endif
+.x_half_y_other_loop:
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m4, [srcq+src_strideq]
+  movx                 m3, [srcq+src_strideq+1]
+  pavgb                m2, m1
+  pavgb                m4, m3
+  movx                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  movx                 m1, [dstq]
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m1, m2, filter_y_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  paddw                m0, m1
+  pmullw               m1, m4, filter_y_b
+  paddw                m2, filter_rnd
+  paddw                m2, m1
+  movx                 m1, [dstq]
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_half_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonhalf:
+  test          y_offsetd, y_offsetd
+  jnz .x_nonhalf_y_nonzero
+
+  ; x_offset == bilin interpolation && y_offset == 0
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+;y_offset == 0. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_other_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+1]
+  mova                 m1, [dstq]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  punpcklbw            m0, m5
+  punpcklbw            m4, m5
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m2, m3
+  paddw                m0, m4
+%endif
+  psraw                m2, 4
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m2, [srcq+src_strideq]
+  movx                 m4, [srcq+src_strideq+1]
+  movx                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  movx                 m1, [dstq]
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_x_a
+  pmaddubsw            m2, filter_x_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m0, m1
+  paddw                m2, filter_rnd
+  movx                 m1, [dstq]
+  paddw                m2, m4
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_other_y_zero_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonhalf_y_nonzero:
+  cmp           y_offsetd, 4
+  jne .x_nonhalf_y_nonhalf
+
+  ; x_offset == bilin interpolation && y_offset == 0.5
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0.5. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m1
+  punpcklbw            m0, m1
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m1, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m0, m1
+  paddw                m2, m3
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+  add                srcq, src_strideq
+  packuswb             m0, m2
+.x_other_y_half_loop:
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+%if cpuflag(ssse3)
+  mova                 m1, [dstq]
+  punpckhbw            m2, m4, m3
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m4, m2
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%else
+  punpckhbw            m2, m4, m5
+  punpckhbw            m1, m3, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m4, m3
+  paddw                m2, m1
+  mova                 m1, [dstq]
+  psraw                m4, 4
+  psraw                m2, 4
+  punpckhbw            m3, m1, m5
+  ; FIXME(rbultje) the repeated pack/unpack here around m0/m2 is because we
+  ; have a 1-register shortage to be able to store the backup of the bilin
+  ; filtered second line as words as cache for the next line. Packing into
+  ; a byte costs 1 pack and 2 unpacks, but saves a register.
+  packuswb             m4, m2
+  punpcklbw            m1, m5
+  pavgb                m0, m4
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  pavgb                m0, [secq]
+%endif
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  pmaddubsw            m0, filter_x_a
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m1
+%endif
+  add                srcq, src_strideq
+  psraw                m0, 4
+.x_other_y_half_loop:
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m4, [srcq+src_strideq]
+  movx                 m3, [srcq+src_strideq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m2, m1
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m1, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  paddw                m2, m1
+  movx                 m1, [dstq]
+  paddw                m4, m3
+  movx                 m3, [dstq+dst_strideq]
+%endif
+  psraw                m2, 4
+  psraw                m4, 4
+  pavgw                m0, m2
+  pavgw                m2, m4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline - also consider going to bytes here
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_other_y_half_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonhalf_y_nonhalf:
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                m11, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m12, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_y_a m10
+%define filter_y_b m11
+%define filter_rnd m12
+%else   ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; In this case, there is NO unused register. Used src_stride register. Later,
+; src_stride has to be loaded from stack when it is needed.
+%define tempq src_strideq
+  mov tempq, g_bilin_filterm
+  add           x_offsetq, tempq
+  add           y_offsetq, tempq
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+  add           y_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+  ; x_offset == bilin interpolation && y_offset == bilin interpolation
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m1
+  punpcklbw            m0, m1
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m1, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m0, m1
+  paddw                m2, m3
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+  packuswb             m0, m2
+.x_other_y_other_loop:
+%if cpuflag(ssse3)
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+  mova                 m1, [dstq]
+  punpckhbw            m2, m4, m3
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  punpckhbw            m3, m1, m5
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m4, m2
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  punpcklbw            m1, m5
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  psraw                m0, 4
+%else
+  movu                 m3, [srcq]
+  movu                 m4, [srcq+1]
+  punpckhbw            m1, m3, m5
+  punpckhbw            m2, m4, m5
+  punpcklbw            m3, m5
+  punpcklbw            m4, m5
+  pmullw               m3, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m3, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m3, m4
+  paddw                m1, m2
+  psraw                m3, 4
+  psraw                m1, 4
+  packuswb             m4, m3, m1
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  pmullw               m2, filter_y_a
+  pmullw               m1, filter_y_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, m1
+  mova                 m1, [dstq]
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  paddw                m0, m3
+  punpckhbw            m3, m1, m5
+  psraw                m0, 4
+  punpcklbw            m1, m5
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  INC_SRC_BY_SRC_STRIDE
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  pmaddubsw            m0, filter_x_a
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m1
+%endif
+  psraw                m0, 4
+%if cpuflag(ssse3)
+  packuswb             m0, m0
+%endif
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+
+  INC_SRC_BY_SRC_STRIDE
+  movx                 m4, [srcq]
+  movx                 m3, [srcq+1]
+
+%if cpuflag(ssse3)
+  punpcklbw            m2, m1
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  movx                 m3, [dstq+dst_strideq]
+  movx                 m1, [dstq]
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m2, m2
+  packuswb             m4, m4
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+  psraw                m0, 4
+  psraw                m2, 4
+  punpcklbw            m1, m5
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m1, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  paddw                m2, m1
+  paddw                m4, m3
+  psraw                m2, 4
+  psraw                m4, 4
+  pmullw               m0, filter_y_a
+  pmullw               m3, m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  pmullw               m1, m4, filter_y_b
+  paddw                m2, filter_rnd
+  paddw                m0, m3
+  movx                 m3, [dstq+dst_strideq]
+  paddw                m2, m1
+  movx                 m1, [dstq]
+  psraw                m0, 4
+  psraw                m2, 4
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_other_y_other_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+%undef movx
+  STORE_AND_RET %1
+%endmacro
+
+; FIXME(rbultje) the non-bilinear versions (i.e. x=0,8&&y=0,8) are identical
+; between the ssse3 and non-ssse3 version. It may make sense to merge their
+; code in the sense that the ssse3 version would jump to the appropriate
+; location in the sse/2 version, rather than duplicating that code in the
+; binary.
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  4
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_XMM ssse3
+SUBPEL_VARIANCE  4
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  4, 1
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
+
+INIT_XMM ssse3
+SUBPEL_VARIANCE  4, 1
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
diff --git a/libvpx/libvpx/vpx_dsp/x86/subtract_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/subtract_sse2.asm
new file mode 100644
index 0000000..4273efb
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/subtract_sse2.asm
@@ -0,0 +1,127 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; void vpx_subtract_block(int rows, int cols,
+;                         int16_t *diff, ptrdiff_t diff_stride,
+;                         const uint8_t *src, ptrdiff_t src_stride,
+;                         const uint8_t *pred, ptrdiff_t pred_stride)
+
+INIT_XMM sse2
+cglobal subtract_block, 7, 7, 8, \
+                        rows, cols, diff, diff_stride, src, src_stride, \
+                        pred, pred_stride
+%define pred_str colsq
+  pxor                  m7, m7         ; dedicated zero register
+  cmp                colsd, 4
+  je .case_4
+  cmp                colsd, 8
+  je .case_8
+  cmp                colsd, 16
+  je .case_16
+  cmp                colsd, 32
+  je .case_32
+
+%macro loop16 6
+  mova                  m0, [srcq+%1]
+  mova                  m4, [srcq+%2]
+  mova                  m1, [predq+%3]
+  mova                  m5, [predq+%4]
+  punpckhbw             m2, m0, m7
+  punpckhbw             m3, m1, m7
+  punpcklbw             m0, m7
+  punpcklbw             m1, m7
+  psubw                 m2, m3
+  psubw                 m0, m1
+  punpckhbw             m1, m4, m7
+  punpckhbw             m3, m5, m7
+  punpcklbw             m4, m7
+  punpcklbw             m5, m7
+  psubw                 m1, m3
+  psubw                 m4, m5
+  mova [diffq+mmsize*0+%5], m0
+  mova [diffq+mmsize*1+%5], m2
+  mova [diffq+mmsize*0+%6], m4
+  mova [diffq+mmsize*1+%6], m1
+%endmacro
+
+  mov             pred_str, pred_stridemp
+.loop_64:
+  loop16 0*mmsize, 1*mmsize, 0*mmsize, 1*mmsize, 0*mmsize, 2*mmsize
+  loop16 2*mmsize, 3*mmsize, 2*mmsize, 3*mmsize, 4*mmsize, 6*mmsize
+  lea                diffq, [diffq+diff_strideq*2]
+  add                predq, pred_str
+  add                 srcq, src_strideq
+  dec                rowsd
+  jg .loop_64
+  RET
+
+.case_32:
+  mov             pred_str, pred_stridemp
+.loop_32:
+  loop16 0, mmsize, 0, mmsize, 0, 2*mmsize
+  lea                diffq, [diffq+diff_strideq*2]
+  add                predq, pred_str
+  add                 srcq, src_strideq
+  dec                rowsd
+  jg .loop_32
+  RET
+
+.case_16:
+  mov             pred_str, pred_stridemp
+.loop_16:
+  loop16 0, src_strideq, 0, pred_str, 0, diff_strideq*2
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                predq, [predq+pred_str*2]
+  lea                 srcq, [srcq+src_strideq*2]
+  sub                rowsd, 2
+  jg .loop_16
+  RET
+
+%macro loop_h 0
+  movh                  m0, [srcq]
+  movh                  m2, [srcq+src_strideq]
+  movh                  m1, [predq]
+  movh                  m3, [predq+pred_str]
+  punpcklbw             m0, m7
+  punpcklbw             m1, m7
+  punpcklbw             m2, m7
+  punpcklbw             m3, m7
+  psubw                 m0, m1
+  psubw                 m2, m3
+  mova             [diffq], m0
+  mova [diffq+diff_strideq*2], m2
+%endmacro
+
+.case_8:
+  mov             pred_str, pred_stridemp
+.loop_8:
+  loop_h
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                 srcq, [srcq+src_strideq*2]
+  lea                predq, [predq+pred_str*2]
+  sub                rowsd, 2
+  jg .loop_8
+  RET
+
+INIT_MMX
+.case_4:
+  mov             pred_str, pred_stridemp
+.loop_4:
+  loop_h
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                 srcq, [srcq+src_strideq*2]
+  lea                predq, [predq+pred_str*2]
+  sub                rowsd, 2
+  jg .loop_4
+  RET
diff --git a/libvpx/libvpx/vpx_dsp/x86/txfm_common_sse2.h b/libvpx/libvpx/vpx_dsp/x86/txfm_common_sse2.h
new file mode 100644
index 0000000..536b206
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/txfm_common_sse2.h
@@ -0,0 +1,29 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_X86_TXFM_COMMON_SSE2_H_
+#define VPX_DSP_X86_TXFM_COMMON_SSE2_H_
+
+#include <emmintrin.h>
+#include "vpx/vpx_integer.h"
+
+#define pair_set_epi16(a, b) \
+  _mm_set_epi16((int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a))
+
+#define dual_set_epi16(a, b) \
+  _mm_set_epi16((int16_t)(b), (int16_t)(b), (int16_t)(b), (int16_t)(b), \
+                (int16_t)(a), (int16_t)(a), (int16_t)(a), (int16_t)(a))
+
+#define octa_set_epi16(a, b, c, d, e, f, g, h) \
+  _mm_setr_epi16((int16_t)(a), (int16_t)(b), (int16_t)(c), (int16_t)(d), \
+                 (int16_t)(e), (int16_t)(f), (int16_t)(g), (int16_t)(h))
+
+#endif  // VPX_DSP_X86_TXFM_COMMON_SSE2_H_
diff --git a/libvpx/libvpx/vpx_dsp/x86/variance_avx2.c b/libvpx/libvpx/vpx_dsp/x86/variance_avx2.c
new file mode 100644
index 0000000..f8c9711
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/variance_avx2.c
@@ -0,0 +1,183 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_dsp_rtcd.h"
+
+typedef void (*get_var_avx2)(const uint8_t *src, int src_stride,
+                             const uint8_t *ref, int ref_stride,
+                             unsigned int *sse, int *sum);
+
+void vpx_get32x32var_avx2(const uint8_t *src, int src_stride,
+                          const uint8_t *ref, int ref_stride,
+                          unsigned int *sse, int *sum);
+
+static void variance_avx2(const uint8_t *src, int src_stride,
+                          const uint8_t *ref, int  ref_stride,
+                          int w, int h, unsigned int *sse, int *sum,
+                          get_var_avx2 var_fn, int block_size) {
+  int i, j;
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i += 16) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(&src[src_stride * i + j], src_stride,
+             &ref[ref_stride * i + j], ref_stride, &sse0, &sum0);
+      *sse += sse0;
+      *sum += sum0;
+    }
+  }
+}
+
+
+unsigned int vpx_variance16x16_avx2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_avx2(src, src_stride, ref, ref_stride, 16, 16,
+                sse, &sum, vpx_get16x16var_avx2, 16);
+  return *sse - (((uint32_t)((int64_t)sum * sum)) >> 8);
+}
+
+unsigned int vpx_mse16x16_avx2(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride,
+                               unsigned int *sse) {
+  int sum;
+  vpx_get16x16var_avx2(src, src_stride, ref, ref_stride, sse, &sum);
+  return *sse;
+}
+
+unsigned int vpx_variance32x16_avx2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_avx2(src, src_stride, ref, ref_stride, 32, 16,
+                sse, &sum, vpx_get32x32var_avx2, 32);
+  return *sse - (((int64_t)sum * sum) >> 9);
+}
+
+unsigned int vpx_variance32x32_avx2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_avx2(src, src_stride, ref, ref_stride, 32, 32,
+                sse, &sum, vpx_get32x32var_avx2, 32);
+  return *sse - (((int64_t)sum * sum) >> 10);
+}
+
+unsigned int vpx_variance64x64_avx2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_avx2(src, src_stride, ref, ref_stride, 64, 64,
+                sse, &sum, vpx_get32x32var_avx2, 32);
+  return *sse - (((int64_t)sum * sum) >> 12);
+}
+
+unsigned int vpx_variance64x32_avx2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_avx2(src, src_stride, ref, ref_stride, 64, 32,
+                sse, &sum, vpx_get32x32var_avx2, 32);
+  return *sse - (((int64_t)sum * sum) >> 11);
+}
+
+unsigned int vpx_sub_pixel_variance32xh_avx2(const uint8_t *src, int src_stride,
+                                             int x_offset, int y_offset,
+                                             const uint8_t *dst, int dst_stride,
+                                             int height,
+                                             unsigned int *sse);
+
+unsigned int vpx_sub_pixel_avg_variance32xh_avx2(const uint8_t *src,
+                                                 int src_stride,
+                                                 int x_offset,
+                                                 int y_offset,
+                                                 const uint8_t *dst,
+                                                 int dst_stride,
+                                                 const uint8_t *sec,
+                                                 int sec_stride,
+                                                 int height,
+                                                 unsigned int *sseptr);
+
+unsigned int vpx_sub_pixel_variance64x64_avx2(const uint8_t *src,
+                                              int src_stride,
+                                              int x_offset,
+                                              int y_offset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  unsigned int sse1;
+  const int se1 = vpx_sub_pixel_variance32xh_avx2(src, src_stride, x_offset,
+                                                  y_offset, dst, dst_stride,
+                                                  64, &sse1);
+  unsigned int sse2;
+  const int se2 = vpx_sub_pixel_variance32xh_avx2(src + 32, src_stride,
+                                                  x_offset, y_offset,
+                                                  dst + 32, dst_stride,
+                                                  64, &sse2);
+  const int se = se1 + se2;
+  *sse = sse1 + sse2;
+  return *sse - (((int64_t)se * se) >> 12);
+}
+
+unsigned int vpx_sub_pixel_variance32x32_avx2(const uint8_t *src,
+                                              int src_stride,
+                                              int x_offset,
+                                              int y_offset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  const int se = vpx_sub_pixel_variance32xh_avx2(src, src_stride, x_offset,
+                                                 y_offset, dst, dst_stride,
+                                                 32, sse);
+  return *sse - (((int64_t)se * se) >> 10);
+}
+
+unsigned int vpx_sub_pixel_avg_variance64x64_avx2(const uint8_t *src,
+                                                  int src_stride,
+                                                  int x_offset,
+                                                  int y_offset,
+                                                  const uint8_t *dst,
+                                                  int dst_stride,
+                                                  unsigned int *sse,
+                                                  const uint8_t *sec) {
+  unsigned int sse1;
+  const int se1 = vpx_sub_pixel_avg_variance32xh_avx2(src, src_stride, x_offset,
+                                                      y_offset, dst, dst_stride,
+                                                      sec, 64, 64, &sse1);
+  unsigned int sse2;
+  const int se2 =
+    vpx_sub_pixel_avg_variance32xh_avx2(src + 32, src_stride, x_offset,
+                                        y_offset, dst + 32, dst_stride,
+                                        sec + 32, 64, 64, &sse2);
+  const int se = se1 + se2;
+
+  *sse = sse1 + sse2;
+
+  return *sse - (((int64_t)se * se) >> 12);
+}
+
+unsigned int vpx_sub_pixel_avg_variance32x32_avx2(const uint8_t *src,
+                                                  int src_stride,
+                                                  int x_offset,
+                                                  int y_offset,
+                                                  const uint8_t *dst,
+                                                  int dst_stride,
+                                                  unsigned int *sse,
+                                                  const uint8_t *sec) {
+  // Process 32 elements in parallel.
+  const int se = vpx_sub_pixel_avg_variance32xh_avx2(src, src_stride, x_offset,
+                                                     y_offset, dst, dst_stride,
+                                                     sec, 32, 32, sse);
+  return *sse - (((int64_t)se * se) >> 10);
+}
diff --git a/libvpx/libvpx/vpx_dsp/x86/variance_impl_avx2.c b/libvpx/libvpx/vpx_dsp/x86/variance_impl_avx2.c
new file mode 100644
index 0000000..b289e9a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/variance_impl_avx2.c
@@ -0,0 +1,727 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+
+DECLARE_ALIGNED(32, static const uint8_t, bilinear_filters_avx2[512]) = {
+  16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0,
+  16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0,
+  14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2,
+  14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2,
+  12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4,
+  12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4,
+  10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6,
+  10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10,
+  6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10,
+  4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12,
+  4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12,
+  2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14,
+  2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14,
+};
+
+
+void vpx_get16x16var_avx2(const unsigned char *src_ptr,
+                          int source_stride,
+                          const unsigned char *ref_ptr,
+                          int recon_stride,
+                          unsigned int *SSE,
+                          int *Sum) {
+    __m256i src, src_expand_low, src_expand_high, ref, ref_expand_low;
+    __m256i ref_expand_high, madd_low, madd_high;
+    unsigned int i, src_2strides, ref_2strides;
+    __m256i zero_reg = _mm256_set1_epi16(0);
+    __m256i sum_ref_src = _mm256_set1_epi16(0);
+    __m256i madd_ref_src = _mm256_set1_epi16(0);
+
+    // processing two strides in a 256 bit register reducing the number
+    // of loop stride by half (comparing to the sse2 code)
+    src_2strides = source_stride << 1;
+    ref_2strides = recon_stride << 1;
+    for (i = 0; i < 8; i++) {
+        src = _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i const *) (src_ptr)));
+        src = _mm256_inserti128_si256(src,
+              _mm_loadu_si128((__m128i const *)(src_ptr+source_stride)), 1);
+
+        ref =_mm256_castsi128_si256(
+             _mm_loadu_si128((__m128i const *) (ref_ptr)));
+        ref = _mm256_inserti128_si256(ref,
+              _mm_loadu_si128((__m128i const *)(ref_ptr+recon_stride)), 1);
+
+        // expanding to 16 bit each lane
+        src_expand_low = _mm256_unpacklo_epi8(src, zero_reg);
+        src_expand_high = _mm256_unpackhi_epi8(src, zero_reg);
+
+        ref_expand_low = _mm256_unpacklo_epi8(ref, zero_reg);
+        ref_expand_high = _mm256_unpackhi_epi8(ref, zero_reg);
+
+        // src-ref
+        src_expand_low = _mm256_sub_epi16(src_expand_low, ref_expand_low);
+        src_expand_high = _mm256_sub_epi16(src_expand_high, ref_expand_high);
+
+        // madd low (src - ref)
+        madd_low = _mm256_madd_epi16(src_expand_low, src_expand_low);
+
+        // add high to low
+        src_expand_low = _mm256_add_epi16(src_expand_low, src_expand_high);
+
+        // madd high (src - ref)
+        madd_high = _mm256_madd_epi16(src_expand_high, src_expand_high);
+
+        sum_ref_src = _mm256_add_epi16(sum_ref_src, src_expand_low);
+
+        // add high to low
+        madd_ref_src = _mm256_add_epi32(madd_ref_src,
+                       _mm256_add_epi32(madd_low, madd_high));
+
+        src_ptr+= src_2strides;
+        ref_ptr+= ref_2strides;
+    }
+
+    {
+        __m128i sum_res, madd_res;
+        __m128i expand_sum_low, expand_sum_high, expand_sum;
+        __m128i expand_madd_low, expand_madd_high, expand_madd;
+        __m128i ex_expand_sum_low, ex_expand_sum_high, ex_expand_sum;
+
+        // extract the low lane and add it to the high lane
+        sum_res = _mm_add_epi16(_mm256_castsi256_si128(sum_ref_src),
+                                _mm256_extractf128_si256(sum_ref_src, 1));
+
+        madd_res = _mm_add_epi32(_mm256_castsi256_si128(madd_ref_src),
+                                 _mm256_extractf128_si256(madd_ref_src, 1));
+
+        // padding each 2 bytes with another 2 zeroed bytes
+        expand_sum_low = _mm_unpacklo_epi16(_mm256_castsi256_si128(zero_reg),
+                                            sum_res);
+        expand_sum_high = _mm_unpackhi_epi16(_mm256_castsi256_si128(zero_reg),
+                                             sum_res);
+
+        // shifting the sign 16 bits right
+        expand_sum_low = _mm_srai_epi32(expand_sum_low, 16);
+        expand_sum_high = _mm_srai_epi32(expand_sum_high, 16);
+
+        expand_sum = _mm_add_epi32(expand_sum_low, expand_sum_high);
+
+        // expand each 32 bits of the madd result to 64 bits
+        expand_madd_low = _mm_unpacklo_epi32(madd_res,
+                          _mm256_castsi256_si128(zero_reg));
+        expand_madd_high = _mm_unpackhi_epi32(madd_res,
+                           _mm256_castsi256_si128(zero_reg));
+
+        expand_madd = _mm_add_epi32(expand_madd_low, expand_madd_high);
+
+        ex_expand_sum_low = _mm_unpacklo_epi32(expand_sum,
+                            _mm256_castsi256_si128(zero_reg));
+        ex_expand_sum_high = _mm_unpackhi_epi32(expand_sum,
+                             _mm256_castsi256_si128(zero_reg));
+
+        ex_expand_sum = _mm_add_epi32(ex_expand_sum_low, ex_expand_sum_high);
+
+        // shift 8 bytes eight
+        madd_res = _mm_srli_si128(expand_madd, 8);
+        sum_res = _mm_srli_si128(ex_expand_sum, 8);
+
+        madd_res = _mm_add_epi32(madd_res, expand_madd);
+        sum_res = _mm_add_epi32(sum_res, ex_expand_sum);
+
+        *((int*)SSE)= _mm_cvtsi128_si32(madd_res);
+
+        *((int*)Sum)= _mm_cvtsi128_si32(sum_res);
+    }
+}
+
+void vpx_get32x32var_avx2(const unsigned char *src_ptr,
+                          int source_stride,
+                          const unsigned char *ref_ptr,
+                          int recon_stride,
+                          unsigned int *SSE,
+                          int *Sum) {
+    __m256i src, src_expand_low, src_expand_high, ref, ref_expand_low;
+    __m256i ref_expand_high, madd_low, madd_high;
+    unsigned int i;
+    __m256i zero_reg = _mm256_set1_epi16(0);
+    __m256i sum_ref_src = _mm256_set1_epi16(0);
+    __m256i madd_ref_src = _mm256_set1_epi16(0);
+
+    // processing 32 elements in parallel
+    for (i = 0; i < 16; i++) {
+       src = _mm256_loadu_si256((__m256i const *) (src_ptr));
+
+       ref = _mm256_loadu_si256((__m256i const *) (ref_ptr));
+
+       // expanding to 16 bit each lane
+       src_expand_low = _mm256_unpacklo_epi8(src, zero_reg);
+       src_expand_high = _mm256_unpackhi_epi8(src, zero_reg);
+
+       ref_expand_low = _mm256_unpacklo_epi8(ref, zero_reg);
+       ref_expand_high = _mm256_unpackhi_epi8(ref, zero_reg);
+
+       // src-ref
+       src_expand_low = _mm256_sub_epi16(src_expand_low, ref_expand_low);
+       src_expand_high = _mm256_sub_epi16(src_expand_high, ref_expand_high);
+
+       // madd low (src - ref)
+       madd_low = _mm256_madd_epi16(src_expand_low, src_expand_low);
+
+       // add high to low
+       src_expand_low = _mm256_add_epi16(src_expand_low, src_expand_high);
+
+       // madd high (src - ref)
+       madd_high = _mm256_madd_epi16(src_expand_high, src_expand_high);
+
+       sum_ref_src = _mm256_add_epi16(sum_ref_src, src_expand_low);
+
+       // add high to low
+       madd_ref_src = _mm256_add_epi32(madd_ref_src,
+                      _mm256_add_epi32(madd_low, madd_high));
+
+       src_ptr+= source_stride;
+       ref_ptr+= recon_stride;
+    }
+
+    {
+      __m256i expand_sum_low, expand_sum_high, expand_sum;
+      __m256i expand_madd_low, expand_madd_high, expand_madd;
+      __m256i ex_expand_sum_low, ex_expand_sum_high, ex_expand_sum;
+
+      // padding each 2 bytes with another 2 zeroed bytes
+      expand_sum_low = _mm256_unpacklo_epi16(zero_reg, sum_ref_src);
+      expand_sum_high = _mm256_unpackhi_epi16(zero_reg, sum_ref_src);
+
+      // shifting the sign 16 bits right
+      expand_sum_low = _mm256_srai_epi32(expand_sum_low, 16);
+      expand_sum_high = _mm256_srai_epi32(expand_sum_high, 16);
+
+      expand_sum = _mm256_add_epi32(expand_sum_low, expand_sum_high);
+
+      // expand each 32 bits of the madd result to 64 bits
+      expand_madd_low = _mm256_unpacklo_epi32(madd_ref_src, zero_reg);
+      expand_madd_high = _mm256_unpackhi_epi32(madd_ref_src, zero_reg);
+
+      expand_madd = _mm256_add_epi32(expand_madd_low, expand_madd_high);
+
+      ex_expand_sum_low = _mm256_unpacklo_epi32(expand_sum, zero_reg);
+      ex_expand_sum_high = _mm256_unpackhi_epi32(expand_sum, zero_reg);
+
+      ex_expand_sum = _mm256_add_epi32(ex_expand_sum_low, ex_expand_sum_high);
+
+      // shift 8 bytes eight
+      madd_ref_src = _mm256_srli_si256(expand_madd, 8);
+      sum_ref_src = _mm256_srli_si256(ex_expand_sum, 8);
+
+      madd_ref_src = _mm256_add_epi32(madd_ref_src, expand_madd);
+      sum_ref_src = _mm256_add_epi32(sum_ref_src, ex_expand_sum);
+
+      // extract the low lane and the high lane and add the results
+      *((int*)SSE)= _mm_cvtsi128_si32(_mm256_castsi256_si128(madd_ref_src)) +
+      _mm_cvtsi128_si32(_mm256_extractf128_si256(madd_ref_src, 1));
+
+      *((int*)Sum)= _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_ref_src)) +
+      _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_ref_src, 1));
+    }
+}
+
+#define FILTER_SRC(filter) \
+  /* filter the source */ \
+  exp_src_lo = _mm256_maddubs_epi16(exp_src_lo, filter); \
+  exp_src_hi = _mm256_maddubs_epi16(exp_src_hi, filter); \
+  \
+  /* add 8 to source */ \
+  exp_src_lo = _mm256_add_epi16(exp_src_lo, pw8); \
+  exp_src_hi = _mm256_add_epi16(exp_src_hi, pw8); \
+  \
+  /* divide source by 16 */ \
+  exp_src_lo = _mm256_srai_epi16(exp_src_lo, 4); \
+  exp_src_hi = _mm256_srai_epi16(exp_src_hi, 4);
+
+#define MERGE_WITH_SRC(src_reg, reg) \
+  exp_src_lo = _mm256_unpacklo_epi8(src_reg, reg); \
+  exp_src_hi = _mm256_unpackhi_epi8(src_reg, reg);
+
+#define LOAD_SRC_DST \
+  /* load source and destination */ \
+  src_reg = _mm256_loadu_si256((__m256i const *) (src)); \
+  dst_reg = _mm256_loadu_si256((__m256i const *) (dst));
+
+#define AVG_NEXT_SRC(src_reg, size_stride) \
+  src_next_reg = _mm256_loadu_si256((__m256i const *) \
+                                   (src + size_stride)); \
+  /* average between current and next stride source */ \
+  src_reg = _mm256_avg_epu8(src_reg, src_next_reg);
+
+#define MERGE_NEXT_SRC(src_reg, size_stride) \
+  src_next_reg = _mm256_loadu_si256((__m256i const *) \
+                                   (src + size_stride)); \
+  MERGE_WITH_SRC(src_reg, src_next_reg)
+
+#define CALC_SUM_SSE_INSIDE_LOOP \
+  /* expand each byte to 2 bytes */ \
+  exp_dst_lo = _mm256_unpacklo_epi8(dst_reg, zero_reg); \
+  exp_dst_hi = _mm256_unpackhi_epi8(dst_reg, zero_reg); \
+  /* source - dest */ \
+  exp_src_lo = _mm256_sub_epi16(exp_src_lo, exp_dst_lo); \
+  exp_src_hi = _mm256_sub_epi16(exp_src_hi, exp_dst_hi); \
+  /* caculate sum */ \
+  sum_reg = _mm256_add_epi16(sum_reg, exp_src_lo); \
+  exp_src_lo = _mm256_madd_epi16(exp_src_lo, exp_src_lo); \
+  sum_reg = _mm256_add_epi16(sum_reg, exp_src_hi); \
+  exp_src_hi = _mm256_madd_epi16(exp_src_hi, exp_src_hi); \
+  /* calculate sse */ \
+  sse_reg = _mm256_add_epi32(sse_reg, exp_src_lo); \
+  sse_reg = _mm256_add_epi32(sse_reg, exp_src_hi);
+
+// final calculation to sum and sse
+#define CALC_SUM_AND_SSE \
+  res_cmp = _mm256_cmpgt_epi16(zero_reg, sum_reg); \
+  sse_reg_hi = _mm256_srli_si256(sse_reg, 8); \
+  sum_reg_lo = _mm256_unpacklo_epi16(sum_reg, res_cmp); \
+  sum_reg_hi = _mm256_unpackhi_epi16(sum_reg, res_cmp); \
+  sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi); \
+  sum_reg = _mm256_add_epi32(sum_reg_lo, sum_reg_hi); \
+  \
+  sse_reg_hi = _mm256_srli_si256(sse_reg, 4); \
+  sum_reg_hi = _mm256_srli_si256(sum_reg, 8); \
+  \
+  sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi); \
+  sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi); \
+  *((int*)sse)= _mm_cvtsi128_si32(_mm256_castsi256_si128(sse_reg)) + \
+                _mm_cvtsi128_si32(_mm256_extractf128_si256(sse_reg, 1)); \
+  sum_reg_hi = _mm256_srli_si256(sum_reg, 4); \
+  sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi); \
+  sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_reg)) + \
+        _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_reg, 1));
+
+
+unsigned int vpx_sub_pixel_variance32xh_avx2(const uint8_t *src,
+                                             int src_stride,
+                                             int x_offset,
+                                             int y_offset,
+                                             const uint8_t *dst,
+                                             int dst_stride,
+                                             int height,
+                                             unsigned int *sse) {
+  __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi;
+  __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi;
+  __m256i zero_reg;
+  int i, sum;
+  sum_reg = _mm256_set1_epi16(0);
+  sse_reg = _mm256_set1_epi16(0);
+  zero_reg = _mm256_set1_epi16(0);
+
+  // x_offset = 0 and y_offset = 0
+  if (x_offset == 0) {
+    if (y_offset == 0) {
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = 0 and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i src_next_reg;
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, src_stride)
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = 0 and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg;
+
+      y_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, src_stride)
+        FILTER_SRC(filter)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    }
+  // x_offset = 8  and y_offset = 0
+  } else if (x_offset == 8) {
+    if (y_offset == 0) {
+      __m256i src_next_reg;
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = 8  and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i src_next_reg, src_avg;
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height ; i++) {
+        src_avg = src_reg;
+        src+= src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        // average between previous average to current average
+        src_avg = _mm256_avg_epu8(src_avg, src_reg);
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_avg, zero_reg)
+        // save current source average
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    // x_offset = 8  and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg, src_avg;
+      y_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height ; i++) {
+        // save current source average
+        src_avg = src_reg;
+        src+= src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        MERGE_WITH_SRC(src_avg, src_reg)
+        FILTER_SRC(filter)
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    }
+  // x_offset = bilin interpolation and y_offset = 0
+  } else {
+    if (y_offset == 0) {
+      __m256i filter, pw8, src_next_reg;
+      x_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = bilin interpolation and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i filter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      MERGE_NEXT_SRC(src_reg, 1)
+      FILTER_SRC(filter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack =  _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height ; i++) {
+        src+= src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        src_reg =  _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // average between previous pack to the current
+        src_pack = _mm256_avg_epu8(src_pack, src_reg);
+        MERGE_WITH_SRC(src_pack, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src_pack = src_reg;
+        dst+= dst_stride;
+      }
+    // x_offset = bilin interpolation and y_offset = bilin interpolation
+    } else {
+      __m256i xfilter, yfilter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      xfilter = _mm256_load_si256((__m256i const *)
+                (bilinear_filters_avx2 + x_offset));
+      y_offset <<= 5;
+      yfilter = _mm256_load_si256((__m256i const *)
+                (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      MERGE_NEXT_SRC(src_reg, 1)
+
+      FILTER_SRC(xfilter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height ; i++) {
+        src+= src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(xfilter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // merge previous pack to current pack source
+        MERGE_WITH_SRC(src_pack, src_reg)
+        // filter the source
+        FILTER_SRC(yfilter)
+        src_pack = src_reg;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    }
+  }
+  CALC_SUM_AND_SSE
+  return sum;
+}
+
+unsigned int vpx_sub_pixel_avg_variance32xh_avx2(const uint8_t *src,
+                                             int src_stride,
+                                             int x_offset,
+                                             int y_offset,
+                                             const uint8_t *dst,
+                                             int dst_stride,
+                                             const uint8_t *sec,
+                                             int sec_stride,
+                                             int height,
+                                             unsigned int *sse) {
+  __m256i sec_reg;
+  __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi;
+  __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi;
+  __m256i zero_reg;
+  int i, sum;
+  sum_reg = _mm256_set1_epi16(0);
+  sse_reg = _mm256_set1_epi16(0);
+  zero_reg = _mm256_set1_epi16(0);
+
+  // x_offset = 0 and y_offset = 0
+  if (x_offset == 0) {
+    if (y_offset == 0) {
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec+= sec_stride;
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    } else if (y_offset == 8) {
+      __m256i src_next_reg;
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, src_stride)
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec+= sec_stride;
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = 0 and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg;
+
+      y_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+                 (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, src_stride)
+        FILTER_SRC(filter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec+= sec_stride;
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    }
+  // x_offset = 8  and y_offset = 0
+  } else if (x_offset == 8) {
+    if (y_offset == 0) {
+      __m256i src_next_reg;
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec+= sec_stride;
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = 8  and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i src_next_reg, src_avg;
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height ; i++) {
+        // save current source average
+        src_avg = src_reg;
+        src+= src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        // average between previous average to current average
+        src_avg = _mm256_avg_epu8(src_avg, src_reg);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_avg = _mm256_avg_epu8(src_avg, sec_reg);
+        sec+= sec_stride;
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_avg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    // x_offset = 8  and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg, src_avg;
+      y_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height ; i++) {
+        // save current source average
+        src_avg = src_reg;
+        src+= src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        MERGE_WITH_SRC(src_avg, src_reg)
+        FILTER_SRC(filter)
+        src_avg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_avg = _mm256_avg_epu8(src_avg, sec_reg);
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_avg, zero_reg)
+        sec+= sec_stride;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    }
+  // x_offset = bilin interpolation and y_offset = 0
+  } else {
+    if (y_offset == 0) {
+      __m256i filter, pw8, src_next_reg;
+      x_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        sec+= sec_stride;
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = bilin interpolation and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i filter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      MERGE_NEXT_SRC(src_reg, 1)
+      FILTER_SRC(filter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack =  _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height ; i++) {
+        src+= src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        src_reg =  _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // average between previous pack to the current
+        src_pack = _mm256_avg_epu8(src_pack, src_reg);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_pack = _mm256_avg_epu8(src_pack, sec_reg);
+        sec+= sec_stride;
+        MERGE_WITH_SRC(src_pack, zero_reg)
+        src_pack = src_reg;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    // x_offset = bilin interpolation and y_offset = bilin interpolation
+    } else {
+      __m256i xfilter, yfilter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      xfilter = _mm256_load_si256((__m256i const *)
+                (bilinear_filters_avx2 + x_offset));
+      y_offset <<= 5;
+      yfilter = _mm256_load_si256((__m256i const *)
+                (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      MERGE_NEXT_SRC(src_reg, 1)
+
+      FILTER_SRC(xfilter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height ; i++) {
+        src+= src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(xfilter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // merge previous pack to current pack source
+        MERGE_WITH_SRC(src_pack, src_reg)
+        // filter the source
+        FILTER_SRC(yfilter)
+        src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_pack = _mm256_avg_epu8(src_pack, sec_reg);
+        MERGE_WITH_SRC(src_pack, zero_reg)
+        src_pack = src_reg;
+        sec+= sec_stride;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    }
+  }
+  CALC_SUM_AND_SSE
+  return sum;
+}
diff --git a/libvpx/libvpx/vpx_dsp/x86/variance_sse2.c b/libvpx/libvpx/vpx_dsp/x86/variance_sse2.c
new file mode 100644
index 0000000..6987c2e
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/variance_sse2.c
@@ -0,0 +1,477 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_ports/mem.h"
+
+typedef void (*getNxMvar_fn_t) (const unsigned char *src, int src_stride,
+                                const unsigned char *ref, int ref_stride,
+                                unsigned int *sse, int *sum);
+
+unsigned int vpx_get_mb_ss_sse2(const int16_t *src) {
+  __m128i vsum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 32; ++i) {
+    const __m128i v = _mm_loadu_si128((const __m128i *)src);
+    vsum = _mm_add_epi32(vsum, _mm_madd_epi16(v, v));
+    src += 8;
+  }
+
+  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4));
+  return  _mm_cvtsi128_si32(vsum);
+}
+
+#define READ64(p, stride, i) \
+  _mm_unpacklo_epi8(_mm_cvtsi32_si128(*(const uint32_t *)(p + i * stride)), \
+      _mm_cvtsi32_si128(*(const uint32_t *)(p + (i + 1) * stride)))
+
+static void get4x4var_sse2(const uint8_t *src, int src_stride,
+                           const uint8_t *ref, int ref_stride,
+                           unsigned int *sse, int *sum) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i src0 = _mm_unpacklo_epi8(READ64(src, src_stride, 0), zero);
+  const __m128i src1 = _mm_unpacklo_epi8(READ64(src, src_stride, 2), zero);
+  const __m128i ref0 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 0), zero);
+  const __m128i ref1 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 2), zero);
+  const __m128i diff0 = _mm_sub_epi16(src0, ref0);
+  const __m128i diff1 = _mm_sub_epi16(src1, ref1);
+
+  // sum
+  __m128i vsum = _mm_add_epi16(diff0, diff1);
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
+  *sum = (int16_t)_mm_extract_epi16(vsum, 0);
+
+  // sse
+  vsum = _mm_add_epi32(_mm_madd_epi16(diff0, diff0),
+                       _mm_madd_epi16(diff1, diff1));
+  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4));
+  *sse = _mm_cvtsi128_si32(vsum);
+}
+
+void vpx_get8x8var_sse2(const uint8_t *src, int src_stride,
+                        const uint8_t *ref, int ref_stride,
+                        unsigned int *sse, int *sum) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i vsum = _mm_setzero_si128();
+  __m128i vsse = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 8; i += 2) {
+    const __m128i src0 = _mm_unpacklo_epi8(_mm_loadl_epi64(
+        (const __m128i *)(src + i * src_stride)), zero);
+    const __m128i ref0 = _mm_unpacklo_epi8(_mm_loadl_epi64(
+        (const __m128i *)(ref + i * ref_stride)), zero);
+    const __m128i diff0 = _mm_sub_epi16(src0, ref0);
+
+    const __m128i src1 = _mm_unpacklo_epi8(_mm_loadl_epi64(
+        (const __m128i *)(src + (i + 1) * src_stride)), zero);
+    const __m128i ref1 = _mm_unpacklo_epi8(_mm_loadl_epi64(
+        (const __m128i *)(ref + (i + 1) * ref_stride)), zero);
+    const __m128i diff1 = _mm_sub_epi16(src1, ref1);
+
+    vsum = _mm_add_epi16(vsum, diff0);
+    vsum = _mm_add_epi16(vsum, diff1);
+    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
+    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
+  }
+
+  // sum
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
+  *sum = (int16_t)_mm_extract_epi16(vsum, 0);
+
+  // sse
+  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
+  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
+  *sse = _mm_cvtsi128_si32(vsse);
+}
+
+void vpx_get16x16var_sse2(const uint8_t *src, int src_stride,
+                          const uint8_t *ref, int ref_stride,
+                          unsigned int *sse, int *sum) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i vsum = _mm_setzero_si128();
+  __m128i vsse = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 16; ++i) {
+    const __m128i s = _mm_loadu_si128((const __m128i *)src);
+    const __m128i r = _mm_loadu_si128((const __m128i *)ref);
+
+    const __m128i src0 = _mm_unpacklo_epi8(s, zero);
+    const __m128i ref0 = _mm_unpacklo_epi8(r, zero);
+    const __m128i diff0 = _mm_sub_epi16(src0, ref0);
+
+    const __m128i src1 = _mm_unpackhi_epi8(s, zero);
+    const __m128i ref1 = _mm_unpackhi_epi8(r, zero);
+    const __m128i diff1 = _mm_sub_epi16(src1, ref1);
+
+    vsum = _mm_add_epi16(vsum, diff0);
+    vsum = _mm_add_epi16(vsum, diff1);
+    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
+    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
+
+    src += src_stride;
+    ref += ref_stride;
+  }
+
+  // sum
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
+  *sum = (int16_t)_mm_extract_epi16(vsum, 0) +
+             (int16_t)_mm_extract_epi16(vsum, 1);
+
+  // sse
+  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
+  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
+  *sse = _mm_cvtsi128_si32(vsse);
+}
+
+
+static void variance_sse2(const unsigned char *src, int src_stride,
+                          const unsigned char *ref, int ref_stride,
+                          int w, int h, unsigned int *sse, int *sum,
+                          getNxMvar_fn_t var_fn, int block_size) {
+  int i, j;
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride,
+             ref + ref_stride * i + j, ref_stride, &sse0, &sum0);
+      *sse += sse0;
+      *sum += sum0;
+    }
+  }
+}
+
+unsigned int vpx_variance4x4_sse2(const unsigned char *src, int src_stride,
+                                  const unsigned char *ref, int ref_stride,
+                                  unsigned int *sse) {
+  int sum;
+  get4x4var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
+  return *sse - ((sum * sum) >> 4);
+}
+
+unsigned int vpx_variance8x4_sse2(const uint8_t *src, int src_stride,
+                                  const uint8_t *ref, int ref_stride,
+                                  unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 8, 4,
+                sse, &sum, get4x4var_sse2, 4);
+  return *sse - ((sum * sum) >> 5);
+}
+
+unsigned int vpx_variance4x8_sse2(const uint8_t *src, int src_stride,
+                                  const uint8_t *ref, int ref_stride,
+                                  unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 4, 8,
+                sse, &sum, get4x4var_sse2, 4);
+  return *sse - ((sum * sum) >> 5);
+}
+
+unsigned int vpx_variance8x8_sse2(const unsigned char *src, int src_stride,
+                                  const unsigned char *ref, int ref_stride,
+                                  unsigned int *sse) {
+  int sum;
+  vpx_get8x8var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
+  return *sse - ((sum * sum) >> 6);
+}
+
+unsigned int vpx_variance16x8_sse2(const unsigned char *src, int src_stride,
+                                   const unsigned char *ref, int ref_stride,
+                                   unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 16, 8,
+                sse, &sum, vpx_get8x8var_sse2, 8);
+  return *sse - ((sum * sum) >> 7);
+}
+
+unsigned int vpx_variance8x16_sse2(const unsigned char *src, int src_stride,
+                                   const unsigned char *ref, int ref_stride,
+                                   unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 8, 16,
+                sse, &sum, vpx_get8x8var_sse2, 8);
+  return *sse - ((sum * sum) >> 7);
+}
+
+unsigned int vpx_variance16x16_sse2(const unsigned char *src, int src_stride,
+                                    const unsigned char *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  vpx_get16x16var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
+  return *sse - (((uint32_t)((int64_t)sum * sum)) >> 8);
+}
+
+unsigned int vpx_variance32x32_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 32, 32,
+                sse, &sum, vpx_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 10);
+}
+
+unsigned int vpx_variance32x16_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 32, 16,
+                sse, &sum, vpx_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 9);
+}
+
+unsigned int vpx_variance16x32_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 16, 32,
+                sse, &sum, vpx_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 9);
+}
+
+unsigned int vpx_variance64x64_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 64, 64,
+                sse, &sum, vpx_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 12);
+}
+
+unsigned int vpx_variance64x32_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 64, 32,
+                sse, &sum, vpx_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 11);
+}
+
+unsigned int vpx_variance32x64_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 32, 64,
+                sse, &sum, vpx_get16x16var_sse2, 16);
+  return *sse - (((int64_t)sum * sum) >> 11);
+}
+
+unsigned int vpx_mse8x8_sse2(const uint8_t *src, int src_stride,
+                             const uint8_t *ref, int ref_stride,
+                             unsigned int *sse) {
+  vpx_variance8x8_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+unsigned int vpx_mse8x16_sse2(const uint8_t *src, int src_stride,
+                              const uint8_t *ref, int ref_stride,
+                              unsigned int *sse) {
+  vpx_variance8x16_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+unsigned int vpx_mse16x8_sse2(const uint8_t *src, int src_stride,
+                              const uint8_t *ref, int ref_stride,
+                              unsigned int *sse) {
+  vpx_variance16x8_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+unsigned int vpx_mse16x16_sse2(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride,
+                               unsigned int *sse) {
+  vpx_variance16x16_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+#if CONFIG_USE_X86INC
+// The 2 unused parameters are place holders for PIC enabled build.
+// These definitions are for functions defined in subpel_variance.asm
+#define DECL(w, opt) \
+  int vpx_sub_pixel_variance##w##xh_##opt(const uint8_t *src, \
+                                          ptrdiff_t src_stride, \
+                                          int x_offset, int y_offset, \
+                                          const uint8_t *dst, \
+                                          ptrdiff_t dst_stride, \
+                                          int height, unsigned int *sse, \
+                                          void *unused0, void *unused)
+#define DECLS(opt1, opt2) \
+  DECL(4, opt1); \
+  DECL(8, opt1); \
+  DECL(16, opt1)
+
+DECLS(sse2, sse2);
+DECLS(ssse3, ssse3);
+#undef DECLS
+#undef DECL
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast) \
+unsigned int vpx_sub_pixel_variance##w##x##h##_##opt(const uint8_t *src, \
+                                                     int src_stride, \
+                                                     int x_offset, \
+                                                     int y_offset, \
+                                                     const uint8_t *dst, \
+                                                     int dst_stride, \
+                                                     unsigned int *sse_ptr) { \
+  unsigned int sse; \
+  int se = vpx_sub_pixel_variance##wf##xh_##opt(src, src_stride, x_offset, \
+                                                y_offset, dst, dst_stride, \
+                                                h, &sse, NULL, NULL); \
+  if (w > wf) { \
+    unsigned int sse2; \
+    int se2 = vpx_sub_pixel_variance##wf##xh_##opt(src + 16, src_stride, \
+                                                   x_offset, y_offset, \
+                                                   dst + 16, dst_stride, \
+                                                   h, &sse2, NULL, NULL); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vpx_sub_pixel_variance##wf##xh_##opt(src + 32, src_stride, \
+                                                 x_offset, y_offset, \
+                                                 dst + 32, dst_stride, \
+                                                 h, &sse2, NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vpx_sub_pixel_variance##wf##xh_##opt(src + 48, src_stride, \
+                                                 x_offset, y_offset, \
+                                                 dst + 48, dst_stride, \
+                                                 h, &sse2, NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  *sse_ptr = sse; \
+  return sse - (cast_prod (cast se * se) >> (wlog2 + hlog2)); \
+}
+
+#define FNS(opt1, opt2) \
+FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t)); \
+FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t)); \
+FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t)); \
+FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t)); \
+FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t)); \
+FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t)); \
+FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \
+FN(16,  8, 16, 4, 3, opt1, (int32_t), (int32_t)); \
+FN(8,  16,  8, 3, 4, opt1, (int32_t), (int32_t)); \
+FN(8,   8,  8, 3, 3, opt1, (int32_t), (int32_t)); \
+FN(8,   4,  8, 3, 2, opt1, (int32_t), (int32_t)); \
+FN(4,   8,  4, 2, 3, opt1, (int32_t), (int32_t)); \
+FN(4,   4,  4, 2, 2, opt1, (int32_t), (int32_t))
+
+FNS(sse2, sse2);
+FNS(ssse3, ssse3);
+
+#undef FNS
+#undef FN
+
+// The 2 unused parameters are place holders for PIC enabled build.
+#define DECL(w, opt) \
+int vpx_sub_pixel_avg_variance##w##xh_##opt(const uint8_t *src, \
+                                            ptrdiff_t src_stride, \
+                                            int x_offset, int y_offset, \
+                                            const uint8_t *dst, \
+                                            ptrdiff_t dst_stride, \
+                                            const uint8_t *sec, \
+                                            ptrdiff_t sec_stride, \
+                                            int height, unsigned int *sse, \
+                                            void *unused0, void *unused)
+#define DECLS(opt1, opt2) \
+DECL(4, opt1); \
+DECL(8, opt1); \
+DECL(16, opt1)
+
+DECLS(sse2, sse2);
+DECLS(ssse3, ssse3);
+#undef DECL
+#undef DECLS
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast) \
+unsigned int vpx_sub_pixel_avg_variance##w##x##h##_##opt(const uint8_t *src, \
+                                                         int src_stride, \
+                                                         int x_offset, \
+                                                         int y_offset, \
+                                                         const uint8_t *dst, \
+                                                         int dst_stride, \
+                                                         unsigned int *sseptr, \
+                                                         const uint8_t *sec) { \
+  unsigned int sse; \
+  int se = vpx_sub_pixel_avg_variance##wf##xh_##opt(src, src_stride, x_offset, \
+                                                    y_offset, dst, dst_stride, \
+                                                    sec, w, h, &sse, NULL, \
+                                                    NULL); \
+  if (w > wf) { \
+    unsigned int sse2; \
+    int se2 = vpx_sub_pixel_avg_variance##wf##xh_##opt(src + 16, src_stride, \
+                                                       x_offset, y_offset, \
+                                                       dst + 16, dst_stride, \
+                                                       sec + 16, w, h, &sse2, \
+                                                       NULL, NULL); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vpx_sub_pixel_avg_variance##wf##xh_##opt(src + 32, src_stride, \
+                                                     x_offset, y_offset, \
+                                                     dst + 32, dst_stride, \
+                                                     sec + 32, w, h, &sse2, \
+                                                     NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vpx_sub_pixel_avg_variance##wf##xh_##opt(src + 48, src_stride, \
+                                                     x_offset, y_offset, \
+                                                     dst + 48, dst_stride, \
+                                                     sec + 48, w, h, &sse2, \
+                                                     NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  *sseptr = sse; \
+  return sse - (cast_prod (cast se * se) >> (wlog2 + hlog2)); \
+}
+
+#define FNS(opt1, opt2) \
+FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t)); \
+FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t)); \
+FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t)); \
+FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t)); \
+FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t)); \
+FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t)); \
+FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \
+FN(16,  8, 16, 4, 3, opt1, (uint32_t), (int32_t)); \
+FN(8,  16,  8, 3, 4, opt1, (uint32_t), (int32_t)); \
+FN(8,   8,  8, 3, 3, opt1, (uint32_t), (int32_t)); \
+FN(8,   4,  8, 3, 2, opt1, (uint32_t), (int32_t)); \
+FN(4,   8,  4, 2, 3, opt1, (uint32_t), (int32_t)); \
+FN(4,   4,  4, 2, 2, opt1, (uint32_t), (int32_t))
+
+FNS(sse2, sse);
+FNS(ssse3, ssse3);
+
+#undef FNS
+#undef FN
+#endif  // CONFIG_USE_X86INC
diff --git a/libvpx/libvpx/vpx_dsp/x86/vpx_asm_stubs.c b/libvpx/libvpx/vpx_dsp/x86/vpx_asm_stubs.c
new file mode 100644
index 0000000..422b0fc
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/vpx_asm_stubs.c
@@ -0,0 +1,162 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/x86/convolve.h"
+
+#if HAVE_SSE2
+filter8_1dfunction vpx_filter_block1d16_v8_sse2;
+filter8_1dfunction vpx_filter_block1d16_h8_sse2;
+filter8_1dfunction vpx_filter_block1d8_v8_sse2;
+filter8_1dfunction vpx_filter_block1d8_h8_sse2;
+filter8_1dfunction vpx_filter_block1d4_v8_sse2;
+filter8_1dfunction vpx_filter_block1d4_h8_sse2;
+filter8_1dfunction vpx_filter_block1d16_v8_avg_sse2;
+filter8_1dfunction vpx_filter_block1d16_h8_avg_sse2;
+filter8_1dfunction vpx_filter_block1d8_v8_avg_sse2;
+filter8_1dfunction vpx_filter_block1d8_h8_avg_sse2;
+filter8_1dfunction vpx_filter_block1d4_v8_avg_sse2;
+filter8_1dfunction vpx_filter_block1d4_h8_avg_sse2;
+
+filter8_1dfunction vpx_filter_block1d16_v2_sse2;
+filter8_1dfunction vpx_filter_block1d16_h2_sse2;
+filter8_1dfunction vpx_filter_block1d8_v2_sse2;
+filter8_1dfunction vpx_filter_block1d8_h2_sse2;
+filter8_1dfunction vpx_filter_block1d4_v2_sse2;
+filter8_1dfunction vpx_filter_block1d4_h2_sse2;
+filter8_1dfunction vpx_filter_block1d16_v2_avg_sse2;
+filter8_1dfunction vpx_filter_block1d16_h2_avg_sse2;
+filter8_1dfunction vpx_filter_block1d8_v2_avg_sse2;
+filter8_1dfunction vpx_filter_block1d8_h2_avg_sse2;
+filter8_1dfunction vpx_filter_block1d4_v2_avg_sse2;
+filter8_1dfunction vpx_filter_block1d4_h2_avg_sse2;
+
+// void vpx_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+// void vpx_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                              uint8_t *dst, ptrdiff_t dst_stride,
+//                              const int16_t *filter_x, int x_step_q4,
+//                              const int16_t *filter_y, int y_step_q4,
+//                              int w, int h);
+// void vpx_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                   uint8_t *dst, ptrdiff_t dst_stride,
+//                                   const int16_t *filter_x, int x_step_q4,
+//                                   const int16_t *filter_y, int y_step_q4,
+//                                   int w, int h);
+// void vpx_convolve8_avg_vert_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                  uint8_t *dst, ptrdiff_t dst_stride,
+//                                  const int16_t *filter_x, int x_step_q4,
+//                                  const int16_t *filter_y, int y_step_q4,
+//                                  int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2);
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2);
+FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, sse2);
+FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_, sse2);
+
+// void vpx_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                         uint8_t *dst, ptrdiff_t dst_stride,
+//                         const int16_t *filter_x, int x_step_q4,
+//                         const int16_t *filter_y, int y_step_q4,
+//                         int w, int h);
+// void vpx_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                             uint8_t *dst, ptrdiff_t dst_stride,
+//                             const int16_t *filter_x, int x_step_q4,
+//                             const int16_t *filter_y, int y_step_q4,
+//                             int w, int h);
+FUN_CONV_2D(, sse2);
+FUN_CONV_2D(avg_ , sse2);
+
+#if CONFIG_VP9_HIGHBITDEPTH && ARCH_X86_64
+highbd_filter8_1dfunction vpx_highbd_filter_block1d16_v8_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d16_h8_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d8_v8_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d8_h8_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d4_v8_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d4_h8_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d16_v8_avg_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d16_h8_avg_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d8_v8_avg_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d8_h8_avg_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d4_v8_avg_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d4_h8_avg_sse2;
+
+highbd_filter8_1dfunction vpx_highbd_filter_block1d16_v2_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d16_h2_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d8_v2_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d8_h2_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d4_v2_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d4_h2_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d16_v2_avg_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d16_h2_avg_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d8_v2_avg_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d8_h2_avg_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d4_v2_avg_sse2;
+highbd_filter8_1dfunction vpx_highbd_filter_block1d4_h2_avg_sse2;
+
+// void vpx_highbd_convolve8_horiz_sse2(const uint8_t *src,
+//                                      ptrdiff_t src_stride,
+//                                      uint8_t *dst,
+//                                      ptrdiff_t dst_stride,
+//                                      const int16_t *filter_x,
+//                                      int x_step_q4,
+//                                      const int16_t *filter_y,
+//                                      int y_step_q4,
+//                                      int w, int h, int bd);
+// void vpx_highbd_convolve8_vert_sse2(const uint8_t *src,
+//                                     ptrdiff_t src_stride,
+//                                     uint8_t *dst,
+//                                     ptrdiff_t dst_stride,
+//                                     const int16_t *filter_x,
+//                                     int x_step_q4,
+//                                     const int16_t *filter_y,
+//                                     int y_step_q4,
+//                                     int w, int h, int bd);
+// void vpx_highbd_convolve8_avg_horiz_sse2(const uint8_t *src,
+//                                          ptrdiff_t src_stride,
+//                                          uint8_t *dst,
+//                                          ptrdiff_t dst_stride,
+//                                          const int16_t *filter_x,
+//                                          int x_step_q4,
+//                                          const int16_t *filter_y,
+//                                          int y_step_q4,
+//                                          int w, int h, int bd);
+// void vpx_highbd_convolve8_avg_vert_sse2(const uint8_t *src,
+//                                         ptrdiff_t src_stride,
+//                                         uint8_t *dst,
+//                                         ptrdiff_t dst_stride,
+//                                         const int16_t *filter_x,
+//                                         int x_step_q4,
+//                                         const int16_t *filter_y,
+//                                         int y_step_q4,
+//                                         int w, int h, int bd);
+HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2);
+HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2);
+HIGH_FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, sse2);
+HIGH_FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_,
+                 sse2);
+
+// void vpx_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h, int bd);
+// void vpx_highbd_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                    uint8_t *dst, ptrdiff_t dst_stride,
+//                                    const int16_t *filter_x, int x_step_q4,
+//                                    const int16_t *filter_y, int y_step_q4,
+//                                    int w, int h, int bd);
+HIGH_FUN_CONV_2D(, sse2);
+HIGH_FUN_CONV_2D(avg_ , sse2);
+#endif  // CONFIG_VP9_HIGHBITDEPTH && ARCH_X86_64
+#endif  // HAVE_SSE2
diff --git a/libvpx/libvpx/vpx_dsp/x86/vpx_convolve_copy_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/vpx_convolve_copy_sse2.asm
new file mode 100644
index 0000000..abc0270
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/vpx_convolve_copy_sse2.asm
@@ -0,0 +1,228 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro convolve_fn 1-2
+%ifidn %1, avg
+%define AUX_XMM_REGS 4
+%else
+%define AUX_XMM_REGS 0
+%endif
+%ifidn %2, highbd
+%define pavg pavgw
+cglobal %2_convolve_%1, 4, 7, 4+AUX_XMM_REGS, src, src_stride, \
+                                              dst, dst_stride, \
+                                              fx, fxs, fy, fys, w, h, bd
+%else
+%define pavg pavgb
+cglobal convolve_%1, 4, 7, 4+AUX_XMM_REGS, src, src_stride, \
+                                           dst, dst_stride, \
+                                           fx, fxs, fy, fys, w, h
+%endif
+  mov r4d, dword wm
+%ifidn %2, highbd
+  shl r4d, 1
+  shl srcq, 1
+  shl src_strideq, 1
+  shl dstq, 1
+  shl dst_strideq, 1
+%else
+  cmp r4d, 4
+  je .w4
+%endif
+  cmp r4d, 8
+  je .w8
+  cmp r4d, 16
+  je .w16
+  cmp r4d, 32
+  je .w32
+%ifidn %2, highbd
+  cmp r4d, 64
+  je .w64
+
+  mov                    r4d, dword hm
+.loop128:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+16]
+  movu                    m2, [srcq+32]
+  movu                    m3, [srcq+48]
+%ifidn %1, avg
+  pavg                    m0, [dstq]
+  pavg                    m1, [dstq+16]
+  pavg                    m2, [dstq+32]
+  pavg                    m3, [dstq+48]
+%endif
+  mova             [dstq   ], m0
+  mova             [dstq+16], m1
+  mova             [dstq+32], m2
+  mova             [dstq+48], m3
+  movu                    m0, [srcq+64]
+  movu                    m1, [srcq+80]
+  movu                    m2, [srcq+96]
+  movu                    m3, [srcq+112]
+  add                   srcq, src_strideq
+%ifidn %1, avg
+  pavg                    m0, [dstq+64]
+  pavg                    m1, [dstq+80]
+  pavg                    m2, [dstq+96]
+  pavg                    m3, [dstq+112]
+%endif
+  mova             [dstq+64], m0
+  mova             [dstq+80], m1
+  mova             [dstq+96], m2
+  mova            [dstq+112], m3
+  add                   dstq, dst_strideq
+  dec                    r4d
+  jnz .loop128
+  RET
+%endif
+
+.w64
+  mov                    r4d, dword hm
+.loop64:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+16]
+  movu                    m2, [srcq+32]
+  movu                    m3, [srcq+48]
+  add                   srcq, src_strideq
+%ifidn %1, avg
+  pavg                    m0, [dstq]
+  pavg                    m1, [dstq+16]
+  pavg                    m2, [dstq+32]
+  pavg                    m3, [dstq+48]
+%endif
+  mova             [dstq   ], m0
+  mova             [dstq+16], m1
+  mova             [dstq+32], m2
+  mova             [dstq+48], m3
+  add                   dstq, dst_strideq
+  dec                    r4d
+  jnz .loop64
+  RET
+
+.w32:
+  mov                    r4d, dword hm
+.loop32:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+16]
+  movu                    m2, [srcq+src_strideq]
+  movu                    m3, [srcq+src_strideq+16]
+  lea                   srcq, [srcq+src_strideq*2]
+%ifidn %1, avg
+  pavg                    m0, [dstq]
+  pavg                    m1, [dstq            +16]
+  pavg                    m2, [dstq+dst_strideq]
+  pavg                    m3, [dstq+dst_strideq+16]
+%endif
+  mova [dstq               ], m0
+  mova [dstq            +16], m1
+  mova [dstq+dst_strideq   ], m2
+  mova [dstq+dst_strideq+16], m3
+  lea                   dstq, [dstq+dst_strideq*2]
+  sub                    r4d, 2
+  jnz .loop32
+  RET
+
+.w16:
+  mov                    r4d, dword hm
+  lea                    r5q, [src_strideq*3]
+  lea                    r6q, [dst_strideq*3]
+.loop16:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+src_strideq]
+  movu                    m2, [srcq+src_strideq*2]
+  movu                    m3, [srcq+r5q]
+  lea                   srcq, [srcq+src_strideq*4]
+%ifidn %1, avg
+  pavg                    m0, [dstq]
+  pavg                    m1, [dstq+dst_strideq]
+  pavg                    m2, [dstq+dst_strideq*2]
+  pavg                    m3, [dstq+r6q]
+%endif
+  mova  [dstq              ], m0
+  mova  [dstq+dst_strideq  ], m1
+  mova  [dstq+dst_strideq*2], m2
+  mova  [dstq+r6q          ], m3
+  lea                   dstq, [dstq+dst_strideq*4]
+  sub                    r4d, 4
+  jnz .loop16
+  RET
+
+.w8:
+  mov                    r4d, dword hm
+  lea                    r5q, [src_strideq*3]
+  lea                    r6q, [dst_strideq*3]
+.loop8:
+  movh                    m0, [srcq]
+  movh                    m1, [srcq+src_strideq]
+  movh                    m2, [srcq+src_strideq*2]
+  movh                    m3, [srcq+r5q]
+  lea                   srcq, [srcq+src_strideq*4]
+%ifidn %1, avg
+  movh                    m4, [dstq]
+  movh                    m5, [dstq+dst_strideq]
+  movh                    m6, [dstq+dst_strideq*2]
+  movh                    m7, [dstq+r6q]
+  pavg                    m0, m4
+  pavg                    m1, m5
+  pavg                    m2, m6
+  pavg                    m3, m7
+%endif
+  movh  [dstq              ], m0
+  movh  [dstq+dst_strideq  ], m1
+  movh  [dstq+dst_strideq*2], m2
+  movh  [dstq+r6q          ], m3
+  lea                   dstq, [dstq+dst_strideq*4]
+  sub                    r4d, 4
+  jnz .loop8
+  RET
+
+%ifnidn %2, highbd
+.w4:
+  mov                    r4d, dword hm
+  lea                    r5q, [src_strideq*3]
+  lea                    r6q, [dst_strideq*3]
+.loop4:
+  movd                    m0, [srcq]
+  movd                    m1, [srcq+src_strideq]
+  movd                    m2, [srcq+src_strideq*2]
+  movd                    m3, [srcq+r5q]
+  lea                   srcq, [srcq+src_strideq*4]
+%ifidn %1, avg
+  movd                    m4, [dstq]
+  movd                    m5, [dstq+dst_strideq]
+  movd                    m6, [dstq+dst_strideq*2]
+  movd                    m7, [dstq+r6q]
+  pavg                    m0, m4
+  pavg                    m1, m5
+  pavg                    m2, m6
+  pavg                    m3, m7
+%endif
+  movd  [dstq              ], m0
+  movd  [dstq+dst_strideq  ], m1
+  movd  [dstq+dst_strideq*2], m2
+  movd  [dstq+r6q          ], m3
+  lea                   dstq, [dstq+dst_strideq*4]
+  sub                    r4d, 4
+  jnz .loop4
+  RET
+%endif
+%endmacro
+
+INIT_XMM sse2
+convolve_fn copy
+convolve_fn avg
+%if CONFIG_VP9_HIGHBITDEPTH
+convolve_fn copy, highbd
+convolve_fn avg, highbd
+%endif
diff --git a/libvpx/libvpx/vpx_dsp/x86/vpx_high_subpixel_8t_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/vpx_high_subpixel_8t_sse2.asm
new file mode 100644
index 0000000..bfc816f
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/vpx_high_subpixel_8t_sse2.asm
@@ -0,0 +1,962 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;Note: tap3 and tap4 have to be applied and added after other taps to avoid
+;overflow.
+
+%macro HIGH_GET_FILTERS_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    psrldq      xmm7, 8
+    pshuflw     xmm4, xmm7, 0b              ;k4
+    pshuflw     xmm5, xmm7, 01010101b       ;k5
+    pshuflw     xmm6, xmm7, 10101010b       ;k6
+    pshuflw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklwd   xmm0, xmm6
+    punpcklwd   xmm2, xmm5
+    punpcklwd   xmm3, xmm4
+    punpcklwd   xmm1, xmm7
+
+    movdqa      k0k6, xmm0
+    movdqa      k2k5, xmm2
+    movdqa      k3k4, xmm3
+    movdqa      k1k7, xmm1
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6
+
+    ;Compute max and min values of a pixel
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)      ;bps
+    movq        xmm0, rdx
+    movq        xmm1, rcx
+    pshufd      xmm0, xmm0, 0b
+    movdqa      xmm2, xmm0
+    psllw       xmm0, xmm1
+    psubw       xmm0, xmm2
+    pxor        xmm1, xmm1
+    movdqa      max, xmm0                  ;max value (for clamping)
+    movdqa      min, xmm1                  ;min value (for clamping)
+
+%endm
+
+%macro HIGH_APPLY_FILTER_4 1
+    punpcklwd   xmm0, xmm6                  ;two row in one register
+    punpcklwd   xmm1, xmm7
+    punpcklwd   xmm2, xmm5
+    punpcklwd   xmm3, xmm4
+
+    pmaddwd     xmm0, k0k6                  ;multiply the filter factors
+    pmaddwd     xmm1, k1k7
+    pmaddwd     xmm2, k2k5
+    pmaddwd     xmm3, k3k4
+
+    paddd       xmm0, xmm1                  ;sum
+    paddd       xmm0, xmm2
+    paddd       xmm0, xmm3
+
+    paddd       xmm0, krd                   ;rounding
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm0                  ;pack to word
+
+    ;clamp the values
+    pminsw      xmm0, max
+    pmaxsw      xmm0, min
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0
+%endm
+
+%macro HIGH_GET_FILTERS 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    pshufhw     xmm4, xmm7, 0b              ;k4
+    pshufhw     xmm5, xmm7, 01010101b       ;k5
+    pshufhw     xmm6, xmm7, 10101010b       ;k6
+    pshufhw     xmm7, xmm7, 11111111b       ;k7
+    punpcklqdq  xmm2, xmm2
+    punpcklqdq  xmm3, xmm3
+    punpcklwd   xmm0, xmm1
+    punpckhwd   xmm6, xmm7
+    punpckhwd   xmm2, xmm5
+    punpckhwd   xmm3, xmm4
+
+    movdqa      k0k1, xmm0                  ;store filter factors on stack
+    movdqa      k6k7, xmm6
+    movdqa      k2k5, xmm2
+    movdqa      k3k4, xmm3
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6                   ;rounding
+
+    ;Compute max and min values of a pixel
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm0, rdx
+    movq        xmm1, rcx
+    pshufd      xmm0, xmm0, 0b
+    movdqa      xmm2, xmm0
+    psllw       xmm0, xmm1
+    psubw       xmm0, xmm2
+    pxor        xmm1, xmm1
+    movdqa      max, xmm0                  ;max value (for clamping)
+    movdqa      min, xmm1                  ;min value (for clamping)
+%endm
+
+%macro LOAD_VERT_8 1
+    movdqu      xmm0, [rsi + %1]            ;0
+    movdqu      xmm1, [rsi + rax + %1]      ;1
+    movdqu      xmm6, [rsi + rdx * 2 + %1]  ;6
+    lea         rsi,  [rsi + rax]
+    movdqu      xmm7, [rsi + rdx * 2 + %1]  ;7
+    movdqu      xmm2, [rsi + rax + %1]      ;2
+    movdqu      xmm3, [rsi + rax * 2 + %1]  ;3
+    movdqu      xmm4, [rsi + rdx + %1]      ;4
+    movdqu      xmm5, [rsi + rax * 4 + %1]  ;5
+%endm
+
+%macro HIGH_APPLY_FILTER_8 2
+    movdqu      temp, xmm4
+    movdqa      xmm4, xmm0
+    punpcklwd   xmm0, xmm1
+    punpckhwd   xmm4, xmm1
+    movdqa      xmm1, xmm6
+    punpcklwd   xmm6, xmm7
+    punpckhwd   xmm1, xmm7
+    movdqa      xmm7, xmm2
+    punpcklwd   xmm2, xmm5
+    punpckhwd   xmm7, xmm5
+
+    movdqu      xmm5, temp
+    movdqu      temp, xmm4
+    movdqa      xmm4, xmm3
+    punpcklwd   xmm3, xmm5
+    punpckhwd   xmm4, xmm5
+    movdqu      xmm5, temp
+
+    pmaddwd     xmm0, k0k1
+    pmaddwd     xmm5, k0k1
+    pmaddwd     xmm6, k6k7
+    pmaddwd     xmm1, k6k7
+    pmaddwd     xmm2, k2k5
+    pmaddwd     xmm7, k2k5
+    pmaddwd     xmm3, k3k4
+    pmaddwd     xmm4, k3k4
+
+    paddd       xmm0, xmm6
+    paddd       xmm0, xmm2
+    paddd       xmm0, xmm3
+    paddd       xmm5, xmm1
+    paddd       xmm5, xmm7
+    paddd       xmm5, xmm4
+
+    paddd       xmm0, krd                   ;rounding
+    paddd       xmm5, krd
+    psrad       xmm0, 7                     ;shift
+    psrad       xmm5, 7
+    packssdw    xmm0, xmm5                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, max
+    pmaxsw      xmm0, min
+
+%if %1
+    movdqu      xmm1, [rdi + %2]
+    pavgw       xmm0, xmm1
+%endif
+    movdqu      [rdi + %2], xmm0
+%endm
+
+;void vpx_filter_block1d4_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vpx_highbd_filter_block1d4_v8_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d4_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movq        xmm0, [rsi]                 ;load src: row 0
+    movq        xmm1, [rsi + rax]           ;1
+    movq        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2]       ;7
+    movq        xmm2, [rsi + rax]           ;2
+    movq        xmm3, [rsi + rax * 2]       ;3
+    movq        xmm4, [rsi + rdx]           ;4
+    movq        xmm5, [rsi + rax * 4]       ;5
+
+    HIGH_APPLY_FILTER_4 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vpx_filter_block1d8_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vpx_highbd_filter_block1d8_v8_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d8_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 0, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vpx_filter_block1d16_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vpx_highbd_filter_block1d16_v8_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d16_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 0, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 16
+    HIGH_APPLY_FILTER_8 0, 16
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_highbd_filter_block1d4_v8_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d4_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movq        xmm0, [rsi]                 ;load src: row 0
+    movq        xmm1, [rsi + rax]           ;1
+    movq        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2]       ;7
+    movq        xmm2, [rsi + rax]           ;2
+    movq        xmm3, [rsi + rax * 2]       ;3
+    movq        xmm4, [rsi + rdx]           ;4
+    movq        xmm5, [rsi + rax * 4]       ;5
+
+    HIGH_APPLY_FILTER_4 1
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_highbd_filter_block1d8_v8_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d8_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 1, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_highbd_filter_block1d16_v8_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d16_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 1, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 16
+    HIGH_APPLY_FILTER_8 1, 16
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vpx_filter_block1d4_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vpx_highbd_filter_block1d4_h8_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d4_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm4,   [rsi + 2]
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm4
+    movdqa      xmm7, xmm4
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm4
+
+    psrldq      xmm1, 2
+    psrldq      xmm6, 4
+    psrldq      xmm7, 6
+    psrldq      xmm2, 4
+    psrldq      xmm3, 6
+    psrldq      xmm5, 2
+
+    HIGH_APPLY_FILTER_4 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vpx_filter_block1d8_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vpx_highbd_filter_block1d8_h8_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d8_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 0, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vpx_filter_block1d16_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vpx_highbd_filter_block1d16_h8_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d16_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 0, 0
+
+    movdqu      xmm0,   [rsi + 10]           ;load src
+    movdqu      xmm1,   [rsi + 12]
+    movdqu      xmm2,   [rsi + 14]
+    movdqu      xmm3,   [rsi + 16]
+    movdqu      xmm4,   [rsi + 18]
+    movdqu      xmm5,   [rsi + 20]
+    movdqu      xmm6,   [rsi + 22]
+    movdqu      xmm7,   [rsi + 24]
+
+    HIGH_APPLY_FILTER_8 0, 16
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_highbd_filter_block1d4_h8_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d4_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm4,   [rsi + 2]
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm4
+    movdqa      xmm7, xmm4
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm4
+
+    psrldq      xmm1, 2
+    psrldq      xmm6, 4
+    psrldq      xmm7, 6
+    psrldq      xmm2, 4
+    psrldq      xmm3, 6
+    psrldq      xmm5, 2
+
+    HIGH_APPLY_FILTER_4 1
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_highbd_filter_block1d8_h8_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d8_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 1, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_highbd_filter_block1d16_h8_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d16_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 1, 0
+
+    movdqu      xmm0,   [rsi + 10]           ;load src
+    movdqu      xmm1,   [rsi + 12]
+    movdqu      xmm2,   [rsi + 14]
+    movdqu      xmm3,   [rsi + 16]
+    movdqu      xmm4,   [rsi + 18]
+    movdqu      xmm5,   [rsi + 20]
+    movdqu      xmm6,   [rsi + 22]
+    movdqu      xmm7,   [rsi + 24]
+
+    HIGH_APPLY_FILTER_8 1, 16
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vpx_dsp/x86/vpx_high_subpixel_bilinear_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/vpx_high_subpixel_bilinear_sse2.asm
new file mode 100644
index 0000000..72f2ff7
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/vpx_high_subpixel_bilinear_sse2.asm
@@ -0,0 +1,494 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro HIGH_GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    pshuflw     xmm4, xmm3, 11111111b       ;k3
+    psrldq      xmm3, 8
+    pshuflw     xmm3, xmm3, 0b              ;k4
+    punpcklwd   xmm4, xmm3                  ;k3k4
+
+    movq        xmm3, rcx                   ;rounding
+    pshufd      xmm3, xmm3, 0
+
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm5, rdx
+    movq        xmm2, rcx
+    pshufd      xmm5, xmm5, 0b
+    movdqa      xmm1, xmm5
+    psllw       xmm5, xmm2
+    psubw       xmm5, xmm1                  ;max value (for clamping)
+    pxor        xmm2, xmm2                  ;min value (for clamping)
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro HIGH_APPLY_FILTER_4 1
+
+    punpcklwd   xmm0, xmm1                  ;two row in one register
+    pmaddwd     xmm0, xmm4                  ;multiply the filter factors
+
+    paddd       xmm0, xmm3                  ;rounding
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm0                  ;pack to word
+
+    ;clamp the values
+    pminsw      xmm0, xmm5
+    pmaxsw      xmm0, xmm2
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+
+    movq        [rdi], xmm0
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+
+%if ARCH_X86_64
+%macro HIGH_GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm6, [rdx]                 ;load filters
+
+    pshuflw     xmm7, xmm6, 11111111b       ;k3
+    pshufhw     xmm6, xmm6, 0b              ;k4
+    psrldq      xmm6, 8
+    punpcklwd   xmm7, xmm6                  ;k3k4k3k4k3k4k3k4
+
+    movq        xmm4, rcx                   ;rounding
+    pshufd      xmm4, xmm4, 0
+
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm8, rdx
+    movq        xmm5, rcx
+    pshufd      xmm8, xmm8, 0b
+    movdqa      xmm1, xmm8
+    psllw       xmm8, xmm5
+    psubw       xmm8, xmm1                  ;max value (for clamping)
+    pxor        xmm5, xmm5                  ;min value (for clamping)
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro HIGH_APPLY_FILTER_8 1
+    movdqa      xmm6, xmm0
+    punpckhwd   xmm6, xmm1
+    punpcklwd   xmm0, xmm1
+    pmaddwd     xmm6, xmm7
+    pmaddwd     xmm0, xmm7
+
+    paddd       xmm6, xmm4                  ;rounding
+    paddd       xmm0, xmm4                  ;rounding
+    psrad       xmm6, 7                     ;shift
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm6                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, xmm8
+    pmaxsw      xmm0, xmm5
+
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+
+%macro HIGH_APPLY_FILTER_16 1
+    movdqa      xmm9, xmm0
+    movdqa      xmm6, xmm2
+    punpckhwd   xmm9, xmm1
+    punpckhwd   xmm6, xmm3
+    punpcklwd   xmm0, xmm1
+    punpcklwd   xmm2, xmm3
+
+    pmaddwd     xmm9, xmm7
+    pmaddwd     xmm6, xmm7
+    pmaddwd     xmm0, xmm7
+    pmaddwd     xmm2, xmm7
+
+    paddd       xmm9, xmm4                  ;rounding
+    paddd       xmm6, xmm4
+    paddd       xmm0, xmm4
+    paddd       xmm2, xmm4
+
+    psrad       xmm9, 7                     ;shift
+    psrad       xmm6, 7
+    psrad       xmm0, 7
+    psrad       xmm2, 7
+
+    packssdw    xmm0, xmm9                  ;pack back to word
+    packssdw    xmm2, xmm6                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, xmm8
+    pmaxsw      xmm0, xmm5
+    pminsw      xmm2, xmm8
+    pmaxsw      xmm2, xmm5
+
+%if %1
+    movdqu      xmm1, [rdi]
+    movdqu      xmm3, [rdi + 16]
+    pavgw       xmm0, xmm1
+    pavgw       xmm2, xmm3
+%endif
+    movdqu      [rdi], xmm0               ;store the result
+    movdqu      [rdi + 16], xmm2          ;store the result
+
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+%endif
+
+global sym(vpx_highbd_filter_block1d4_v2_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d4_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movq        xmm0, [rsi]                 ;load src
+    movq        xmm1, [rsi + 2*rax]
+
+    HIGH_APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(vpx_highbd_filter_block1d8_v2_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d8_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;0
+    movdqu      xmm1, [rsi + 2*rax]         ;1
+
+    HIGH_APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_highbd_filter_block1d16_v2_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d16_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm2, [rsi + 16]
+    movdqu        xmm1, [rsi + 2*rax]       ;1
+    movdqu        xmm3, [rsi + 2*rax + 16]
+
+    HIGH_APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
+
+global sym(vpx_highbd_filter_block1d4_v2_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d4_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movq        xmm0, [rsi]                 ;load src
+    movq        xmm1, [rsi + 2*rax]
+
+    HIGH_APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(vpx_highbd_filter_block1d8_v2_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d8_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;0
+    movdqu      xmm1, [rsi + 2*rax]         ;1
+
+    HIGH_APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_highbd_filter_block1d16_v2_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d16_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + 2*rax]       ;1
+    movdqu        xmm2, [rsi + 16]
+    movdqu        xmm3, [rsi + 2*rax + 16]
+
+    HIGH_APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
+
+global sym(vpx_highbd_filter_block1d4_h2_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d4_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 2
+
+    HIGH_APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(vpx_highbd_filter_block1d8_h2_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d8_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqu      xmm1, [rsi + 2]
+
+    HIGH_APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_highbd_filter_block1d16_h2_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d16_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 2]
+    movdqu      xmm2,   [rsi + 16]
+    movdqu      xmm3,   [rsi + 18]
+
+    HIGH_APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
+
+global sym(vpx_highbd_filter_block1d4_h2_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d4_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 2
+
+    HIGH_APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(vpx_highbd_filter_block1d8_h2_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d8_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqu      xmm1, [rsi + 2]
+
+    HIGH_APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_highbd_filter_block1d16_h2_avg_sse2) PRIVATE
+sym(vpx_highbd_filter_block1d16_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 2]
+    movdqu      xmm2,   [rsi + 16]
+    movdqu      xmm3,   [rsi + 18]
+
+    HIGH_APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
diff --git a/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c
new file mode 100644
index 0000000..b718678
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c
@@ -0,0 +1,605 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Due to a header conflict between math.h and intrinsics includes with ceil()
+// in certain configurations under vs9 this include needs to precede
+// immintrin.h.
+
+#include <immintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/x86/convolve.h"
+#include "vpx_ports/mem.h"
+
+// filters for 16_h8 and 16_v8
+DECLARE_ALIGNED(32, static const uint8_t, filt1_global_avx2[32]) = {
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt2_global_avx2[32]) = {
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10,
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt3_global_avx2[32]) = {
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12,
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt4_global_avx2[32]) = {
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14,
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
+};
+
+#if defined(__clang__)
+# if __clang_major__ < 3 || (__clang_major__ == 3 && __clang_minor__ <= 3) || \
+    (defined(__APPLE__) && \
+        ((__clang_major__ == 4 && __clang_minor__ <= 2) || \
+            (__clang_major__ == 5 && __clang_minor__ == 0)))
+
+#  define MM256_BROADCASTSI128_SI256(x) \
+       _mm_broadcastsi128_si256((__m128i const *)&(x))
+# else  // clang > 3.3, and not 5.0 on macosx.
+#  define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+# endif  // clang <= 3.3
+#elif defined(__GNUC__)
+# if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 6)
+#  define MM256_BROADCASTSI128_SI256(x) \
+       _mm_broadcastsi128_si256((__m128i const *)&(x))
+# elif __GNUC__ == 4 && __GNUC_MINOR__ == 7
+#  define MM256_BROADCASTSI128_SI256(x) _mm_broadcastsi128_si256(x)
+# else  // gcc > 4.7
+#  define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+# endif  // gcc <= 4.6
+#else  // !(gcc || clang)
+# define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif  // __clang__
+
+static void vpx_filter_block1d16_h8_avx2(const uint8_t *src_ptr,
+                                         ptrdiff_t src_pixels_per_line,
+                                         uint8_t *output_ptr,
+                                         ptrdiff_t output_pitch,
+                                         uint32_t output_height,
+                                         const int16_t *filter) {
+  __m128i filtersReg;
+  __m256i addFilterReg64, filt1Reg, filt2Reg, filt3Reg, filt4Reg;
+  __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3;
+  __m256i srcReg32b1, srcReg32b2, filtersReg32;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to 8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+  // have the same data in both lanes of a 256 bit register
+  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 256 bit register
+  firstFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 256 bit register
+  secondFilters = _mm256_shuffle_epi8(filtersReg32,
+                  _mm256_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 256 bit register
+  thirdFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 256 bit register
+  forthFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x706u));
+
+  filt1Reg = _mm256_load_si256((__m256i const *)filt1_global_avx2);
+  filt2Reg = _mm256_load_si256((__m256i const *)filt2_global_avx2);
+  filt3Reg = _mm256_load_si256((__m256i const *)filt3_global_avx2);
+  filt4Reg = _mm256_load_si256((__m256i const *)filt4_global_avx2);
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pixels_per_line << 1;
+  dst_stride = output_pitch << 1;
+  for (i = output_height; i > 1; i-=2) {
+    // load the 2 strides of source
+    srcReg32b1 = _mm256_castsi128_si256(
+                 _mm_loadu_si128((const __m128i *)(src_ptr - 3)));
+    srcReg32b1 = _mm256_inserti128_si256(srcReg32b1,
+                 _mm_loadu_si128((const __m128i *)
+                 (src_ptr+src_pixels_per_line-3)), 1);
+
+    // filter the source buffer
+    srcRegFilt32b1_1= _mm256_shuffle_epi8(srcReg32b1, filt1Reg);
+    srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b1, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2);
+
+    // filter the source buffer
+    srcRegFilt32b3= _mm256_shuffle_epi8(srcReg32b1, filt2Reg);
+    srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b1, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1,
+                       _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+    // reading 2 strides of the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg32b2 = _mm256_castsi128_si256(
+                 _mm_loadu_si128((const __m128i *)(src_ptr + 5)));
+    srcReg32b2 = _mm256_inserti128_si256(srcReg32b2,
+                 _mm_loadu_si128((const __m128i *)
+                 (src_ptr+src_pixels_per_line+5)), 1);
+
+    // add and saturate the results together
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1,
+                       _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+    // filter the source buffer
+    srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b2_1 = _mm256_maddubs_epi16(srcRegFilt32b2_1, firstFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, srcRegFilt32b2);
+
+    // filter the source buffer
+    srcRegFilt32b3= _mm256_shuffle_epi8(srcReg32b2, filt2Reg);
+    srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b2, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1,
+                       _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1,
+                       _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg64);
+
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg64);
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 7);
+    srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1,
+                                           srcRegFilt32b2_1);
+
+    src_ptr+=src_stride;
+
+    // save 16 bytes
+    _mm_store_si128((__m128i*)output_ptr,
+    _mm256_castsi256_si128(srcRegFilt32b1_1));
+
+    // save the next 16 bits
+    _mm_store_si128((__m128i*)(output_ptr+output_pitch),
+    _mm256_extractf128_si256(srcRegFilt32b1_1, 1));
+    output_ptr+=dst_stride;
+  }
+
+  // if the number of strides is odd.
+  // process only 16 bytes
+  if (i > 0) {
+    __m128i srcReg1, srcReg2, srcRegFilt1_1, srcRegFilt2_1;
+    __m128i srcRegFilt2, srcRegFilt3;
+
+    srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+
+    // filter the source buffer
+    srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1,
+                    _mm256_castsi256_si128(filt1Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg1,
+                  _mm256_castsi256_si128(filt4Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1_1 = _mm_maddubs_epi16(srcRegFilt1_1,
+                    _mm256_castsi256_si128(firstFilters));
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+                  _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
+
+    // filter the source buffer
+    srcRegFilt3= _mm_shuffle_epi8(srcReg1,
+                 _mm256_castsi256_si128(filt2Reg));
+    srcRegFilt2= _mm_shuffle_epi8(srcReg1,
+                 _mm256_castsi256_si128(filt3Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3,
+                  _mm256_castsi256_si128(secondFilters));
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+                  _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+                    _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+
+    // reading the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 5));
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+                    _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+
+    // filter the source buffer
+    srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2,
+                    _mm256_castsi256_si128(filt1Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg2,
+                  _mm256_castsi256_si128(filt4Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt2_1 = _mm_maddubs_epi16(srcRegFilt2_1,
+                    _mm256_castsi256_si128(firstFilters));
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+                  _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
+
+    // filter the source buffer
+    srcRegFilt3 = _mm_shuffle_epi8(srcReg2,
+                  _mm256_castsi256_si128(filt2Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg2,
+                  _mm256_castsi256_si128(filt3Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3,
+                  _mm256_castsi256_si128(secondFilters));
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+                  _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+                    _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+                    _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+
+
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+                    _mm256_castsi256_si128(addFilterReg64));
+
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+                    _mm256_castsi256_si128(addFilterReg64));
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 7);
+    srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, srcRegFilt2_1);
+
+    // save 16 bytes
+    _mm_store_si128((__m128i*)output_ptr, srcRegFilt1_1);
+  }
+}
+
+static void vpx_filter_block1d16_v8_avx2(const uint8_t *src_ptr,
+                                         ptrdiff_t src_pitch,
+                                         uint8_t *output_ptr,
+                                         ptrdiff_t out_pitch,
+                                         uint32_t output_height,
+                                         const int16_t *filter) {
+  __m128i filtersReg;
+  __m256i addFilterReg64;
+  __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5;
+  __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10;
+  __m256i srcReg32b11, srcReg32b12, filtersReg32;
+  __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the
+  // same data in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+  // have the same data in both lanes of a 256 bit register
+  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 256 bit register
+  firstFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 256 bit register
+  secondFilters = _mm256_shuffle_epi8(filtersReg32,
+                  _mm256_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 256 bit register
+  thirdFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 256 bit register
+  forthFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x706u));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  // load 16 bytes 7 times in stride of src_pitch
+  srcReg32b1 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr)));
+  srcReg32b2 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch)));
+  srcReg32b3 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2)));
+  srcReg32b4 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3)));
+  srcReg32b5 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4)));
+  srcReg32b6 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5)));
+  srcReg32b7 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6)));
+
+  // have each consecutive loads on the same 256 register
+  srcReg32b1 = _mm256_inserti128_si256(srcReg32b1,
+               _mm256_castsi256_si128(srcReg32b2), 1);
+  srcReg32b2 = _mm256_inserti128_si256(srcReg32b2,
+               _mm256_castsi256_si128(srcReg32b3), 1);
+  srcReg32b3 = _mm256_inserti128_si256(srcReg32b3,
+               _mm256_castsi256_si128(srcReg32b4), 1);
+  srcReg32b4 = _mm256_inserti128_si256(srcReg32b4,
+               _mm256_castsi256_si128(srcReg32b5), 1);
+  srcReg32b5 = _mm256_inserti128_si256(srcReg32b5,
+               _mm256_castsi256_si128(srcReg32b6), 1);
+  srcReg32b6 = _mm256_inserti128_si256(srcReg32b6,
+               _mm256_castsi256_si128(srcReg32b7), 1);
+
+  // merge every two consecutive registers except the last one
+  srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2);
+  srcReg32b1 = _mm256_unpackhi_epi8(srcReg32b1, srcReg32b2);
+
+  // save
+  srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4);
+
+  // save
+  srcReg32b3 = _mm256_unpackhi_epi8(srcReg32b3, srcReg32b4);
+
+  // save
+  srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6);
+
+  // save
+  srcReg32b5 = _mm256_unpackhi_epi8(srcReg32b5, srcReg32b6);
+
+
+  for (i = output_height; i > 1; i-=2) {
+     // load the last 2 loads of 16 bytes and have every two
+     // consecutive loads in the same 256 bit register
+     srcReg32b8 = _mm256_castsi128_si256(
+     _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7)));
+     srcReg32b7 = _mm256_inserti128_si256(srcReg32b7,
+     _mm256_castsi256_si128(srcReg32b8), 1);
+     srcReg32b9 = _mm256_castsi128_si256(
+     _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 8)));
+     srcReg32b8 = _mm256_inserti128_si256(srcReg32b8,
+     _mm256_castsi256_si128(srcReg32b9), 1);
+
+     // merge every two consecutive registers
+     // save
+     srcReg32b4 = _mm256_unpacklo_epi8(srcReg32b7, srcReg32b8);
+     srcReg32b7 = _mm256_unpackhi_epi8(srcReg32b7, srcReg32b8);
+
+     // multiply 2 adjacent elements with the filter and add the result
+     srcReg32b10 = _mm256_maddubs_epi16(srcReg32b10, firstFilters);
+     srcReg32b6 = _mm256_maddubs_epi16(srcReg32b4, forthFilters);
+
+     // add and saturate the results together
+     srcReg32b10 = _mm256_adds_epi16(srcReg32b10, srcReg32b6);
+
+     // multiply 2 adjacent elements with the filter and add the result
+     srcReg32b8 = _mm256_maddubs_epi16(srcReg32b11, secondFilters);
+     srcReg32b12 = _mm256_maddubs_epi16(srcReg32b2, thirdFilters);
+
+     // add and saturate the results together
+     srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
+                   _mm256_min_epi16(srcReg32b8, srcReg32b12));
+     srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
+                   _mm256_max_epi16(srcReg32b8, srcReg32b12));
+
+     // multiply 2 adjacent elements with the filter and add the result
+     srcReg32b1 = _mm256_maddubs_epi16(srcReg32b1, firstFilters);
+     srcReg32b6 = _mm256_maddubs_epi16(srcReg32b7, forthFilters);
+
+     srcReg32b1 = _mm256_adds_epi16(srcReg32b1, srcReg32b6);
+
+     // multiply 2 adjacent elements with the filter and add the result
+     srcReg32b8 = _mm256_maddubs_epi16(srcReg32b3, secondFilters);
+     srcReg32b12 = _mm256_maddubs_epi16(srcReg32b5, thirdFilters);
+
+     // add and saturate the results together
+     srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
+                  _mm256_min_epi16(srcReg32b8, srcReg32b12));
+     srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
+                  _mm256_max_epi16(srcReg32b8, srcReg32b12));
+
+     srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg64);
+     srcReg32b1 = _mm256_adds_epi16(srcReg32b1, addFilterReg64);
+
+     // shift by 7 bit each 16 bit
+     srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 7);
+     srcReg32b1 = _mm256_srai_epi16(srcReg32b1, 7);
+
+     // shrink to 8 bit each 16 bits, the first lane contain the first
+     // convolve result and the second lane contain the second convolve
+     // result
+     srcReg32b1 = _mm256_packus_epi16(srcReg32b10, srcReg32b1);
+
+     src_ptr+=src_stride;
+
+     // save 16 bytes
+     _mm_store_si128((__m128i*)output_ptr,
+     _mm256_castsi256_si128(srcReg32b1));
+
+     // save the next 16 bits
+     _mm_store_si128((__m128i*)(output_ptr+out_pitch),
+     _mm256_extractf128_si256(srcReg32b1, 1));
+
+     output_ptr+=dst_stride;
+
+     // save part of the registers for next strides
+     srcReg32b10 = srcReg32b11;
+     srcReg32b1 = srcReg32b3;
+     srcReg32b11 = srcReg32b2;
+     srcReg32b3 = srcReg32b5;
+     srcReg32b2 = srcReg32b4;
+     srcReg32b5 = srcReg32b7;
+     srcReg32b7 = srcReg32b9;
+  }
+  if (i > 0) {
+    __m128i srcRegFilt1, srcRegFilt3, srcRegFilt4, srcRegFilt5;
+    __m128i srcRegFilt6, srcRegFilt7, srcRegFilt8;
+    // load the last 16 bytes
+    srcRegFilt8 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+
+    // merge the last 2 results together
+    srcRegFilt4 = _mm_unpacklo_epi8(
+                  _mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
+    srcRegFilt7 = _mm_unpackhi_epi8(
+                  _mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b10),
+                  _mm256_castsi256_si128(firstFilters));
+    srcRegFilt4 = _mm_maddubs_epi16(srcRegFilt4,
+                  _mm256_castsi256_si128(forthFilters));
+    srcRegFilt3 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b1),
+                  _mm256_castsi256_si128(firstFilters));
+    srcRegFilt7 = _mm_maddubs_epi16(srcRegFilt7,
+                  _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3, srcRegFilt7);
+
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11),
+                  _mm256_castsi256_si128(secondFilters));
+    srcRegFilt5 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b3),
+                  _mm256_castsi256_si128(secondFilters));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt6 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b2),
+                  _mm256_castsi256_si128(thirdFilters));
+    srcRegFilt7 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b5),
+                  _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+                  _mm_min_epi16(srcRegFilt4, srcRegFilt6));
+    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3,
+                  _mm_min_epi16(srcRegFilt5, srcRegFilt7));
+
+    // add and saturate the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+                  _mm_max_epi16(srcRegFilt4, srcRegFilt6));
+    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3,
+                  _mm_max_epi16(srcRegFilt5, srcRegFilt7));
+
+
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+                  _mm256_castsi256_si128(addFilterReg64));
+    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3,
+                  _mm256_castsi256_si128(addFilterReg64));
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+    srcRegFilt3 = _mm_srai_epi16(srcRegFilt3, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt3);
+
+    // save 16 bytes
+    _mm_store_si128((__m128i*)output_ptr, srcRegFilt1);
+  }
+}
+
+#if HAVE_AVX2 && HAVE_SSSE3
+filter8_1dfunction vpx_filter_block1d4_v8_ssse3;
+#if ARCH_X86_64
+filter8_1dfunction vpx_filter_block1d8_v8_intrin_ssse3;
+filter8_1dfunction vpx_filter_block1d8_h8_intrin_ssse3;
+filter8_1dfunction vpx_filter_block1d4_h8_intrin_ssse3;
+#define vpx_filter_block1d8_v8_avx2 vpx_filter_block1d8_v8_intrin_ssse3
+#define vpx_filter_block1d8_h8_avx2 vpx_filter_block1d8_h8_intrin_ssse3
+#define vpx_filter_block1d4_h8_avx2 vpx_filter_block1d4_h8_intrin_ssse3
+#else  // ARCH_X86
+filter8_1dfunction vpx_filter_block1d8_v8_ssse3;
+filter8_1dfunction vpx_filter_block1d8_h8_ssse3;
+filter8_1dfunction vpx_filter_block1d4_h8_ssse3;
+#define vpx_filter_block1d8_v8_avx2 vpx_filter_block1d8_v8_ssse3
+#define vpx_filter_block1d8_h8_avx2 vpx_filter_block1d8_h8_ssse3
+#define vpx_filter_block1d4_h8_avx2 vpx_filter_block1d4_h8_ssse3
+#endif  // ARCH_X86_64
+filter8_1dfunction vpx_filter_block1d16_v2_ssse3;
+filter8_1dfunction vpx_filter_block1d16_h2_ssse3;
+filter8_1dfunction vpx_filter_block1d8_v2_ssse3;
+filter8_1dfunction vpx_filter_block1d8_h2_ssse3;
+filter8_1dfunction vpx_filter_block1d4_v2_ssse3;
+filter8_1dfunction vpx_filter_block1d4_h2_ssse3;
+#define vpx_filter_block1d4_v8_avx2 vpx_filter_block1d4_v8_ssse3
+#define vpx_filter_block1d16_v2_avx2 vpx_filter_block1d16_v2_ssse3
+#define vpx_filter_block1d16_h2_avx2 vpx_filter_block1d16_h2_ssse3
+#define vpx_filter_block1d8_v2_avx2  vpx_filter_block1d8_v2_ssse3
+#define vpx_filter_block1d8_h2_avx2  vpx_filter_block1d8_h2_ssse3
+#define vpx_filter_block1d4_v2_avx2  vpx_filter_block1d4_v2_ssse3
+#define vpx_filter_block1d4_h2_avx2  vpx_filter_block1d4_h2_ssse3
+// void vpx_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h);
+// void vpx_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2);
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2);
+
+// void vpx_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                          uint8_t *dst, ptrdiff_t dst_stride,
+//                          const int16_t *filter_x, int x_step_q4,
+//                          const int16_t *filter_y, int y_step_q4,
+//                          int w, int h);
+FUN_CONV_2D(, avx2);
+#endif  // HAVE_AX2 && HAVE_SSSE3
diff --git a/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_intrin_ssse3.c b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_intrin_ssse3.c
new file mode 100644
index 0000000..6fd5208
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_intrin_ssse3.c
@@ -0,0 +1,915 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Due to a header conflict between math.h and intrinsics includes with ceil()
+// in certain configurations under vs9 this include needs to precede
+// tmmintrin.h.
+
+#include <tmmintrin.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/vpx_filter.h"
+#include "vpx_dsp/x86/convolve.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/emmintrin_compat.h"
+
+// filters only for the 4_h8 convolution
+DECLARE_ALIGNED(16, static const uint8_t, filt1_4_h8[16]) = {
+  0, 1, 1, 2, 2, 3, 3, 4, 2, 3, 3, 4, 4, 5, 5, 6
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt2_4_h8[16]) = {
+  4, 5, 5, 6, 6, 7, 7, 8, 6, 7, 7, 8, 8, 9, 9, 10
+};
+
+// filters for 8_h8 and 16_h8
+DECLARE_ALIGNED(16, static const uint8_t, filt1_global[16]) = {
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt2_global[16]) = {
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt3_global[16]) = {
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt4_global[16]) = {
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
+};
+
+// These are reused by the avx2 intrinsics.
+filter8_1dfunction vpx_filter_block1d8_v8_intrin_ssse3;
+filter8_1dfunction vpx_filter_block1d8_h8_intrin_ssse3;
+filter8_1dfunction vpx_filter_block1d4_h8_intrin_ssse3;
+
+void vpx_filter_block1d4_h8_intrin_ssse3(const uint8_t *src_ptr,
+                                         ptrdiff_t src_pixels_per_line,
+                                         uint8_t *output_ptr,
+                                         ptrdiff_t output_pitch,
+                                         uint32_t output_height,
+                                         const int16_t *filter) {
+  __m128i firstFilters, secondFilters, shuffle1, shuffle2;
+  __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4;
+  __m128i addFilterReg64, filtersReg, srcReg, minReg;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 =_mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits in the filter into the first lane
+  firstFilters = _mm_shufflelo_epi16(filtersReg, 0);
+  // duplicate only the third 16 bit in the filter into the first lane
+  secondFilters = _mm_shufflelo_epi16(filtersReg, 0xAAu);
+  // duplicate only the seconds 16 bits in the filter into the second lane
+  // firstFilters: k0 k1 k0 k1 k0 k1 k0 k1 k2 k3 k2 k3 k2 k3 k2 k3
+  firstFilters = _mm_shufflehi_epi16(firstFilters, 0x55u);
+  // duplicate only the forth 16 bits in the filter into the second lane
+  // secondFilters: k4 k5 k4 k5 k4 k5 k4 k5 k6 k7 k6 k7 k6 k7 k6 k7
+  secondFilters = _mm_shufflehi_epi16(secondFilters, 0xFFu);
+
+  // loading the local filters
+  shuffle1 =_mm_load_si128((__m128i const *)filt1_4_h8);
+  shuffle2 = _mm_load_si128((__m128i const *)filt2_4_h8);
+
+  for (i = 0; i < output_height; i++) {
+    srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+
+    // filter the source buffer
+    srcRegFilt1= _mm_shuffle_epi8(srcReg, shuffle1);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg, shuffle2);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
+
+    // extract the higher half of the lane
+    srcRegFilt3 =  _mm_srli_si128(srcRegFilt1, 8);
+    srcRegFilt4 =  _mm_srli_si128(srcRegFilt2, 8);
+
+    minReg = _mm_min_epi16(srcRegFilt3, srcRegFilt2);
+
+    // add and saturate all the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+    srcRegFilt3 = _mm_max_epi16(srcRegFilt3, srcRegFilt2);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
+
+    // shift by 7 bit each 16 bits
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+
+    // shrink to 8 bit each 16 bits
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1);
+    src_ptr+=src_pixels_per_line;
+
+    // save only 4 bytes
+    *((int*)&output_ptr[0])= _mm_cvtsi128_si32(srcRegFilt1);
+
+    output_ptr+=output_pitch;
+  }
+}
+
+void vpx_filter_block1d8_h8_intrin_ssse3(const uint8_t *src_ptr,
+                                         ptrdiff_t src_pixels_per_line,
+                                         uint8_t *output_ptr,
+                                         ptrdiff_t output_pitch,
+                                         uint32_t output_height,
+                                         const int16_t *filter) {
+  __m128i firstFilters, secondFilters, thirdFilters, forthFilters, srcReg;
+  __m128i filt1Reg, filt2Reg, filt3Reg, filt4Reg;
+  __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4;
+  __m128i addFilterReg64, filtersReg, minReg;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 128 bit register
+  firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 128 bit register
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 128 bit register
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 128 bit register
+  forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u));
+
+  filt1Reg = _mm_load_si128((__m128i const *)filt1_global);
+  filt2Reg = _mm_load_si128((__m128i const *)filt2_global);
+  filt3Reg = _mm_load_si128((__m128i const *)filt3_global);
+  filt4Reg = _mm_load_si128((__m128i const *)filt4_global);
+
+  for (i = 0; i < output_height; i++) {
+    srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+
+    // filter the source buffer
+    srcRegFilt1= _mm_shuffle_epi8(srcReg, filt1Reg);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg, filt2Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
+
+    // filter the source buffer
+    srcRegFilt3= _mm_shuffle_epi8(srcReg, filt3Reg);
+    srcRegFilt4= _mm_shuffle_epi8(srcReg, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, thirdFilters);
+    srcRegFilt4 = _mm_maddubs_epi16(srcRegFilt4, forthFilters);
+
+    // add and saturate all the results together
+    minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+
+    srcRegFilt2= _mm_max_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
+
+    // shift by 7 bit each 16 bits
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+
+    // shrink to 8 bit each 16 bits
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1);
+
+    src_ptr+=src_pixels_per_line;
+
+    // save only 8 bytes
+    _mm_storel_epi64((__m128i*)&output_ptr[0], srcRegFilt1);
+
+    output_ptr+=output_pitch;
+  }
+}
+
+void vpx_filter_block1d8_v8_intrin_ssse3(const uint8_t *src_ptr,
+                                         ptrdiff_t src_pitch,
+                                         uint8_t *output_ptr,
+                                         ptrdiff_t out_pitch,
+                                         uint32_t output_height,
+                                         const int16_t *filter) {
+  __m128i addFilterReg64, filtersReg, minReg;
+  __m128i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt5;
+  __m128i srcReg1, srcReg2, srcReg3, srcReg4, srcReg5, srcReg6, srcReg7;
+  __m128i srcReg8;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits in the filter
+  firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u));
+  // duplicate only the second 16 bits in the filter
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits in the filter
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits in the filter
+  forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u));
+
+  // load the first 7 rows of 8 bytes
+  srcReg1 = _mm_loadl_epi64((const __m128i *)src_ptr);
+  srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch));
+  srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
+  srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
+  srcReg7 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
+
+  for (i = 0; i < output_height; i++) {
+    // load the last 8 bytes
+    srcReg8 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7));
+
+    // merge the result together
+    srcRegFilt1 = _mm_unpacklo_epi8(srcReg1, srcReg2);
+    srcRegFilt3 = _mm_unpacklo_epi8(srcReg3, srcReg4);
+
+    // merge the result together
+    srcRegFilt2 = _mm_unpacklo_epi8(srcReg5, srcReg6);
+    srcRegFilt5 = _mm_unpacklo_epi8(srcReg7, srcReg8);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, thirdFilters);
+    srcRegFilt5 = _mm_maddubs_epi16(srcRegFilt5, forthFilters);
+
+    // add and saturate the results together
+    minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt5);
+    srcRegFilt2 = _mm_max_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+
+    // shrink to 8 bit each 16 bits
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1);
+
+    src_ptr+=src_pitch;
+
+    // shift down a row
+    srcReg1 = srcReg2;
+    srcReg2 = srcReg3;
+    srcReg3 = srcReg4;
+    srcReg4 = srcReg5;
+    srcReg5 = srcReg6;
+    srcReg6 = srcReg7;
+    srcReg7 = srcReg8;
+
+    // save only 8 bytes convolve result
+    _mm_storel_epi64((__m128i*)&output_ptr[0], srcRegFilt1);
+
+    output_ptr+=out_pitch;
+  }
+}
+
+filter8_1dfunction vpx_filter_block1d16_v8_ssse3;
+filter8_1dfunction vpx_filter_block1d16_h8_ssse3;
+filter8_1dfunction vpx_filter_block1d8_v8_ssse3;
+filter8_1dfunction vpx_filter_block1d8_h8_ssse3;
+filter8_1dfunction vpx_filter_block1d4_v8_ssse3;
+filter8_1dfunction vpx_filter_block1d4_h8_ssse3;
+filter8_1dfunction vpx_filter_block1d16_v8_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d16_h8_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d8_v8_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d8_h8_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d4_v8_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d4_h8_avg_ssse3;
+
+filter8_1dfunction vpx_filter_block1d16_v2_ssse3;
+filter8_1dfunction vpx_filter_block1d16_h2_ssse3;
+filter8_1dfunction vpx_filter_block1d8_v2_ssse3;
+filter8_1dfunction vpx_filter_block1d8_h2_ssse3;
+filter8_1dfunction vpx_filter_block1d4_v2_ssse3;
+filter8_1dfunction vpx_filter_block1d4_h2_ssse3;
+filter8_1dfunction vpx_filter_block1d16_v2_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d16_h2_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d8_v2_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d8_h2_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d4_v2_avg_ssse3;
+filter8_1dfunction vpx_filter_block1d4_h2_avg_ssse3;
+
+// void vpx_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h);
+// void vpx_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+// void vpx_convolve8_avg_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                                    uint8_t *dst, ptrdiff_t dst_stride,
+//                                    const int16_t *filter_x, int x_step_q4,
+//                                    const int16_t *filter_y, int y_step_q4,
+//                                    int w, int h);
+// void vpx_convolve8_avg_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                                   uint8_t *dst, ptrdiff_t dst_stride,
+//                                   const int16_t *filter_x, int x_step_q4,
+//                                   const int16_t *filter_y, int y_step_q4,
+//                                   int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , ssse3);
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , ssse3);
+FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, ssse3);
+FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_,
+            ssse3);
+
+#define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7,           \
+                      out0, out1, out2, out3, out4, out5, out6, out7) { \
+  const __m128i tr0_0 = _mm_unpacklo_epi8(in0, in1);                    \
+  const __m128i tr0_1 = _mm_unpacklo_epi8(in2, in3);                    \
+  const __m128i tr0_2 = _mm_unpacklo_epi8(in4, in5);                    \
+  const __m128i tr0_3 = _mm_unpacklo_epi8(in6, in7);                    \
+                                                                        \
+  const __m128i tr1_0 = _mm_unpacklo_epi16(tr0_0, tr0_1);               \
+  const __m128i tr1_1 = _mm_unpackhi_epi16(tr0_0, tr0_1);               \
+  const __m128i tr1_2 = _mm_unpacklo_epi16(tr0_2, tr0_3);               \
+  const __m128i tr1_3 = _mm_unpackhi_epi16(tr0_2, tr0_3);               \
+                                                                        \
+  const __m128i tr2_0 = _mm_unpacklo_epi32(tr1_0, tr1_2);               \
+  const __m128i tr2_1 = _mm_unpackhi_epi32(tr1_0, tr1_2);               \
+  const __m128i tr2_2 = _mm_unpacklo_epi32(tr1_1, tr1_3);               \
+  const __m128i tr2_3 = _mm_unpackhi_epi32(tr1_1, tr1_3);               \
+                                                                        \
+  out0 = _mm_unpacklo_epi64(tr2_0, tr2_0);                              \
+  out1 = _mm_unpackhi_epi64(tr2_0, tr2_0);                              \
+  out2 = _mm_unpacklo_epi64(tr2_1, tr2_1);                              \
+  out3 = _mm_unpackhi_epi64(tr2_1, tr2_1);                              \
+  out4 = _mm_unpacklo_epi64(tr2_2, tr2_2);                              \
+  out5 = _mm_unpackhi_epi64(tr2_2, tr2_2);                              \
+  out6 = _mm_unpacklo_epi64(tr2_3, tr2_3);                              \
+  out7 = _mm_unpackhi_epi64(tr2_3, tr2_3);                              \
+}
+
+static void filter_horiz_w8_ssse3(const uint8_t *src_x, ptrdiff_t src_pitch,
+                                  uint8_t *dst, const int16_t *x_filter) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i f_values = _mm_load_si128((const __m128i *)x_filter);
+  // pack and duplicate the filter values
+  const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+  const __m128i A = _mm_loadl_epi64((const __m128i *)src_x);
+  const __m128i B = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch));
+  const __m128i C = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 2));
+  const __m128i D = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 3));
+  const __m128i E = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 4));
+  const __m128i F = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 5));
+  const __m128i G = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 6));
+  const __m128i H = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 7));
+  // 00 01 10 11 02 03 12 13 04 05 14 15 06 07 16 17
+  const __m128i tr0_0 = _mm_unpacklo_epi16(A, B);
+  // 20 21 30 31 22 23 32 33 24 25 34 35 26 27 36 37
+  const __m128i tr0_1 = _mm_unpacklo_epi16(C, D);
+  // 40 41 50 51 42 43 52 53 44 45 54 55 46 47 56 57
+  const __m128i tr0_2 = _mm_unpacklo_epi16(E, F);
+  // 60 61 70 71 62 63 72 73 64 65 74 75 66 67 76 77
+  const __m128i tr0_3 = _mm_unpacklo_epi16(G, H);
+  // 00 01 10 11 20 21 30 31 02 03 12 13 22 23 32 33
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  // 04 05 14 15 24 25 34 35 06 07 16 17 26 27 36 37
+  const __m128i tr1_1 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  // 40 41 50 51 60 61 70 71 42 43 52 53 62 63 72 73
+  const __m128i tr1_2 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  // 44 45 54 55 64 65 74 75 46 47 56 57 66 67 76 77
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  // 00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71
+  const __m128i s1s0 = _mm_unpacklo_epi64(tr1_0, tr1_2);
+  const __m128i s3s2 = _mm_unpackhi_epi64(tr1_0, tr1_2);
+  const __m128i s5s4 = _mm_unpacklo_epi64(tr1_1, tr1_3);
+  const __m128i s7s6 = _mm_unpackhi_epi64(tr1_1, tr1_3);
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0);
+  const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2);
+  const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4);
+  const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6);
+  // add and saturate the results together
+  const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+  const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+  __m128i temp = _mm_adds_epi16(x0, x3);
+  temp = _mm_adds_epi16(temp, min_x2x1);
+  temp = _mm_adds_epi16(temp, max_x2x1);
+  // round and shift by 7 bit each 16 bit
+  temp = _mm_mulhrs_epi16(temp, k_256);
+  // shrink to 8 bit each 16 bits
+  temp = _mm_packus_epi16(temp, temp);
+  // save only 8 bytes convolve result
+  _mm_storel_epi64((__m128i*)dst, temp);
+}
+
+static void transpose8x8_to_dst(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride) {
+  __m128i A, B, C, D, E, F, G, H;
+
+  A = _mm_loadl_epi64((const __m128i *)src);
+  B = _mm_loadl_epi64((const __m128i *)(src + src_stride));
+  C = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2));
+  D = _mm_loadl_epi64((const __m128i *)(src + src_stride * 3));
+  E = _mm_loadl_epi64((const __m128i *)(src + src_stride * 4));
+  F = _mm_loadl_epi64((const __m128i *)(src + src_stride * 5));
+  G = _mm_loadl_epi64((const __m128i *)(src + src_stride * 6));
+  H = _mm_loadl_epi64((const __m128i *)(src + src_stride * 7));
+
+  TRANSPOSE_8X8(A, B, C, D, E, F, G, H,
+                A, B, C, D, E, F, G, H);
+
+  _mm_storel_epi64((__m128i*)dst, A);
+  _mm_storel_epi64((__m128i*)(dst + dst_stride * 1), B);
+  _mm_storel_epi64((__m128i*)(dst + dst_stride * 2), C);
+  _mm_storel_epi64((__m128i*)(dst + dst_stride * 3), D);
+  _mm_storel_epi64((__m128i*)(dst + dst_stride * 4), E);
+  _mm_storel_epi64((__m128i*)(dst + dst_stride * 5), F);
+  _mm_storel_epi64((__m128i*)(dst + dst_stride * 6), G);
+  _mm_storel_epi64((__m128i*)(dst + dst_stride * 7), H);
+}
+
+static void scaledconvolve_horiz_w8(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const InterpKernel *x_filters,
+                                    int x0_q4, int x_step_q4, int w, int h) {
+  DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]);
+  int x, y, z;
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  // This function processes 8x8 areas.  The intermediate height is not always
+  // a multiple of 8, so force it to be a multiple of 8 here.
+  y = h + (8 - (h & 0x7));
+
+  do {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; x += 8) {
+      // process 8 src_x steps
+      for (z = 0; z < 8; ++z) {
+        const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+        const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+        if (x_q4 & SUBPEL_MASK) {
+          filter_horiz_w8_ssse3(src_x, src_stride, temp + (z * 8), x_filter);
+        } else {
+          int i;
+          for (i = 0; i < 8; ++i) {
+            temp[z * 8 + i] = src_x[i * src_stride + 3];
+          }
+        }
+        x_q4 += x_step_q4;
+      }
+
+      // transpose the 8x8 filters values back to dst
+      transpose8x8_to_dst(temp, 8, dst + x, dst_stride);
+    }
+
+    src += src_stride * 8;
+    dst += dst_stride * 8;
+  } while (y -= 8);
+}
+
+static void filter_horiz_w4_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                  uint8_t *dst, const int16_t *filter) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+  // pack and duplicate the filter values
+  const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+  const __m128i A = _mm_loadl_epi64((const __m128i *)src_ptr);
+  const __m128i B = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch));
+  const __m128i C = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  const __m128i D = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  // TRANSPOSE...
+  // 00 01 02 03 04 05 06 07
+  // 10 11 12 13 14 15 16 17
+  // 20 21 22 23 24 25 26 27
+  // 30 31 32 33 34 35 36 37
+  //
+  // TO
+  //
+  // 00 10 20 30
+  // 01 11 21 31
+  // 02 12 22 32
+  // 03 13 23 33
+  // 04 14 24 34
+  // 05 15 25 35
+  // 06 16 26 36
+  // 07 17 27 37
+  //
+  // 00 01 10 11 02 03 12 13 04 05 14 15 06 07 16 17
+  const __m128i tr0_0 = _mm_unpacklo_epi16(A, B);
+  // 20 21 30 31 22 23 32 33 24 25 34 35 26 27 36 37
+  const __m128i tr0_1 = _mm_unpacklo_epi16(C, D);
+  // 00 01 10 11 20 21 30 31 02 03 12 13 22 23 32 33
+  const __m128i s1s0  = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  // 04 05 14 15 24 25 34 35 06 07 16 17 26 27 36 37
+  const __m128i s5s4 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  // 02 03 12 13 22 23 32 33
+  const __m128i s3s2 = _mm_srli_si128(s1s0, 8);
+  // 06 07 16 17 26 27 36 37
+  const __m128i s7s6 = _mm_srli_si128(s5s4, 8);
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0);
+  const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2);
+  const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4);
+  const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6);
+  // add and saturate the results together
+  const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+  const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+  __m128i temp = _mm_adds_epi16(x0, x3);
+  temp = _mm_adds_epi16(temp, min_x2x1);
+  temp = _mm_adds_epi16(temp, max_x2x1);
+  // round and shift by 7 bit each 16 bit
+  temp = _mm_mulhrs_epi16(temp, k_256);
+  // shrink to 8 bit each 16 bits
+  temp = _mm_packus_epi16(temp, temp);
+  // save only 4 bytes
+  *(int *)dst = _mm_cvtsi128_si32(temp);
+}
+
+static void transpose4x4_to_dst(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride) {
+  __m128i A = _mm_cvtsi32_si128(*(const int *)src);
+  __m128i B = _mm_cvtsi32_si128(*(const int *)(src + src_stride));
+  __m128i C = _mm_cvtsi32_si128(*(const int *)(src + src_stride * 2));
+  __m128i D = _mm_cvtsi32_si128(*(const int *)(src + src_stride * 3));
+  // 00 10 01 11 02 12 03 13
+  const __m128i tr0_0 = _mm_unpacklo_epi8(A, B);
+  // 20 30 21 31 22 32 23 33
+  const __m128i tr0_1 = _mm_unpacklo_epi8(C, D);
+  // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+  A = _mm_unpacklo_epi16(tr0_0, tr0_1);
+  B = _mm_srli_si128(A, 4);
+  C = _mm_srli_si128(A, 8);
+  D = _mm_srli_si128(A, 12);
+
+  *(int *)(dst) =  _mm_cvtsi128_si32(A);
+  *(int *)(dst + dst_stride) =  _mm_cvtsi128_si32(B);
+  *(int *)(dst + dst_stride * 2) =  _mm_cvtsi128_si32(C);
+  *(int *)(dst + dst_stride * 3) =  _mm_cvtsi128_si32(D);
+}
+
+static void scaledconvolve_horiz_w4(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const InterpKernel *x_filters,
+                                    int x0_q4, int x_step_q4, int w, int h) {
+  DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]);
+  int x, y, z;
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  for (y = 0; y < h; y += 4) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; x += 4) {
+      // process 4 src_x steps
+      for (z = 0; z < 4; ++z) {
+        const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+        const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+        if (x_q4 & SUBPEL_MASK) {
+          filter_horiz_w4_ssse3(src_x, src_stride, temp + (z * 4), x_filter);
+        } else {
+          int i;
+          for (i = 0; i < 4; ++i) {
+            temp[z * 4 + i] = src_x[i * src_stride + 3];
+          }
+        }
+        x_q4 += x_step_q4;
+      }
+
+      // transpose the 4x4 filters values back to dst
+      transpose4x4_to_dst(temp, 4, dst + x, dst_stride);
+    }
+
+    src += src_stride * 4;
+    dst += dst_stride * 4;
+  }
+}
+
+static void filter_vert_w4_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                 uint8_t *dst, const int16_t *filter) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+  // pack and duplicate the filter values
+  const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+  const __m128i A = _mm_cvtsi32_si128(*(const int *)src_ptr);
+  const __m128i B = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch));
+  const __m128i C = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 2));
+  const __m128i D = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 3));
+  const __m128i E = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 4));
+  const __m128i F = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 5));
+  const __m128i G = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 6));
+  const __m128i H = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 7));
+  const __m128i s1s0 = _mm_unpacklo_epi8(A, B);
+  const __m128i s3s2 = _mm_unpacklo_epi8(C, D);
+  const __m128i s5s4 = _mm_unpacklo_epi8(E, F);
+  const __m128i s7s6 = _mm_unpacklo_epi8(G, H);
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0);
+  const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2);
+  const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4);
+  const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6);
+  // add and saturate the results together
+  const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+  const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+  __m128i temp = _mm_adds_epi16(x0, x3);
+  temp = _mm_adds_epi16(temp, min_x2x1);
+  temp = _mm_adds_epi16(temp, max_x2x1);
+  // round and shift by 7 bit each 16 bit
+  temp = _mm_mulhrs_epi16(temp, k_256);
+  // shrink to 8 bit each 16 bits
+  temp = _mm_packus_epi16(temp, temp);
+  // save only 4 bytes
+  *(int *)dst = _mm_cvtsi128_si32(temp);
+}
+
+static void scaledconvolve_vert_w4(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const InterpKernel *y_filters,
+                                   int y0_q4, int y_step_q4, int w, int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (y = 0; y < h; ++y) {
+    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w4_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter);
+    } else {
+      memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+    }
+
+    y_q4 += y_step_q4;
+  }
+}
+
+static void filter_vert_w8_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                 uint8_t *dst, const int16_t *filter) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+  // pack and duplicate the filter values
+  const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+  const __m128i A = _mm_loadl_epi64((const __m128i *)src_ptr);
+  const __m128i B = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch));
+  const __m128i C = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  const __m128i D = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  const __m128i E = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
+  const __m128i F = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
+  const __m128i G = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
+  const __m128i H = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7));
+  const __m128i s1s0 = _mm_unpacklo_epi8(A, B);
+  const __m128i s3s2 = _mm_unpacklo_epi8(C, D);
+  const __m128i s5s4 = _mm_unpacklo_epi8(E, F);
+  const __m128i s7s6 = _mm_unpacklo_epi8(G, H);
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0);
+  const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2);
+  const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4);
+  const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6);
+  // add and saturate the results together
+  const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+  const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+  __m128i temp = _mm_adds_epi16(x0, x3);
+  temp = _mm_adds_epi16(temp, min_x2x1);
+  temp = _mm_adds_epi16(temp, max_x2x1);
+  // round and shift by 7 bit each 16 bit
+  temp = _mm_mulhrs_epi16(temp, k_256);
+  // shrink to 8 bit each 16 bits
+  temp = _mm_packus_epi16(temp, temp);
+  // save only 8 bytes convolve result
+  _mm_storel_epi64((__m128i*)dst, temp);
+}
+
+static void scaledconvolve_vert_w8(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const InterpKernel *y_filters,
+                                   int y0_q4, int y_step_q4, int w, int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (y = 0; y < h; ++y) {
+    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w8_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter);
+    } else {
+      memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+    }
+    y_q4 += y_step_q4;
+  }
+}
+
+static void filter_vert_w16_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                  uint8_t *dst, const int16_t *filter, int w) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+  // pack and duplicate the filter values
+  const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+  int i;
+
+  for (i = 0; i < w; i += 16) {
+    const __m128i A = _mm_loadu_si128((const __m128i *)src_ptr);
+    const __m128i B = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch));
+    const __m128i C =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
+    const __m128i D =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
+    const __m128i E =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4));
+    const __m128i F =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5));
+    const __m128i G =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6));
+    const __m128i H =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+    // merge the result together
+    const __m128i s1s0_lo = _mm_unpacklo_epi8(A, B);
+    const __m128i s7s6_lo = _mm_unpacklo_epi8(G, H);
+    const __m128i s1s0_hi = _mm_unpackhi_epi8(A, B);
+    const __m128i s7s6_hi = _mm_unpackhi_epi8(G, H);
+    // multiply 2 adjacent elements with the filter and add the result
+    const __m128i x0_lo = _mm_maddubs_epi16(s1s0_lo, f1f0);
+    const __m128i x3_lo = _mm_maddubs_epi16(s7s6_lo, f7f6);
+    const __m128i x0_hi = _mm_maddubs_epi16(s1s0_hi, f1f0);
+    const __m128i x3_hi = _mm_maddubs_epi16(s7s6_hi, f7f6);
+    // add and saturate the results together
+    const __m128i x3x0_lo = _mm_adds_epi16(x0_lo, x3_lo);
+    const __m128i x3x0_hi = _mm_adds_epi16(x0_hi, x3_hi);
+    // merge the result together
+    const __m128i s3s2_lo = _mm_unpacklo_epi8(C, D);
+    const __m128i s3s2_hi = _mm_unpackhi_epi8(C, D);
+    // multiply 2 adjacent elements with the filter and add the result
+    const __m128i x1_lo = _mm_maddubs_epi16(s3s2_lo, f3f2);
+    const __m128i x1_hi = _mm_maddubs_epi16(s3s2_hi, f3f2);
+    // merge the result together
+    const __m128i s5s4_lo = _mm_unpacklo_epi8(E, F);
+    const __m128i s5s4_hi = _mm_unpackhi_epi8(E, F);
+    // multiply 2 adjacent elements with the filter and add the result
+    const __m128i x2_lo = _mm_maddubs_epi16(s5s4_lo, f5f4);
+    const __m128i x2_hi = _mm_maddubs_epi16(s5s4_hi, f5f4);
+    // add and saturate the results together
+    __m128i temp_lo = _mm_adds_epi16(x3x0_lo, _mm_min_epi16(x1_lo, x2_lo));
+    __m128i temp_hi = _mm_adds_epi16(x3x0_hi, _mm_min_epi16(x1_hi, x2_hi));
+
+    // add and saturate the results together
+    temp_lo = _mm_adds_epi16(temp_lo, _mm_max_epi16(x1_lo, x2_lo));
+    temp_hi = _mm_adds_epi16(temp_hi, _mm_max_epi16(x1_hi, x2_hi));
+    // round and shift by 7 bit each 16 bit
+    temp_lo = _mm_mulhrs_epi16(temp_lo, k_256);
+    temp_hi = _mm_mulhrs_epi16(temp_hi, k_256);
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    temp_hi = _mm_packus_epi16(temp_lo, temp_hi);
+    src_ptr += 16;
+     // save 16 bytes convolve result
+    _mm_store_si128((__m128i*)&dst[i], temp_hi);
+  }
+}
+
+static void scaledconvolve_vert_w16(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const InterpKernel *y_filters,
+                                    int y0_q4, int y_step_q4, int w, int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (y = 0; y < h; ++y) {
+    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w16_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter,
+                            w);
+    } else {
+      memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+    }
+    y_q4 += y_step_q4;
+  }
+}
+
+static void scaledconvolve2d(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const InterpKernel *const x_filters,
+                             int x0_q4, int x_step_q4,
+                             const InterpKernel *const y_filters,
+                             int y0_q4, int y_step_q4,
+                             int w, int h) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  // --Require an additional 8 rows for the horiz_w8 transpose tail.
+  DECLARE_ALIGNED(16, uint8_t, temp[(135 + 8) * 64]);
+  const int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= 64);
+  assert(h <= 64);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  if (w >= 8) {
+    scaledconvolve_horiz_w8(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                            src_stride, temp, 64, x_filters, x0_q4, x_step_q4,
+                            w, intermediate_height);
+  } else {
+    scaledconvolve_horiz_w4(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                            src_stride, temp, 64, x_filters, x0_q4, x_step_q4,
+                            w, intermediate_height);
+  }
+
+  if (w >= 16) {
+    scaledconvolve_vert_w16(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
+                            dst_stride, y_filters, y0_q4, y_step_q4, w, h);
+  } else if (w == 8) {
+    scaledconvolve_vert_w8(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
+                           dst_stride, y_filters, y0_q4, y_step_q4, w, h);
+  } else {
+    scaledconvolve_vert_w4(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
+                           dst_stride, y_filters, y0_q4, y_step_q4, w, h);
+  }
+}
+
+static const InterpKernel *get_filter_base(const int16_t *filter) {
+  // NOTE: This assumes that the filter table is 256-byte aligned.
+  // TODO(agrange) Modify to make independent of table alignment.
+  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
+}
+
+static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
+  return (int)((const InterpKernel *)(intptr_t)f - base);
+}
+
+void vpx_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+                         uint8_t *dst, ptrdiff_t dst_stride,
+                         const int16_t *filter_x, int x_step_q4,
+                         const int16_t *filter_y, int y_step_q4,
+                         int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  scaledconvolve2d(src, src_stride, dst, dst_stride,
+                   filters_x, x0_q4, x_step_q4,
+                   filters_y, y0_q4, y_step_q4, w, h);
+}
+
+// void vp9_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                          uint8_t *dst, ptrdiff_t dst_stride,
+//                          const int16_t *filter_x, int x_step_q4,
+//                          const int16_t *filter_y, int y_step_q4,
+//                          int w, int h);
+// void vpx_convolve8_avg_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                              uint8_t *dst, ptrdiff_t dst_stride,
+//                              const int16_t *filter_x, int x_step_q4,
+//                              const int16_t *filter_y, int y_step_q4,
+//                              int w, int h);
+FUN_CONV_2D(, ssse3);
+FUN_CONV_2D(avg_ , ssse3);
diff --git a/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_sse2.asm
new file mode 100644
index 0000000..08f3d6a
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_sse2.asm
@@ -0,0 +1,987 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;Note: tap3 and tap4 have to be applied and added after other taps to avoid
+;overflow.
+
+%macro GET_FILTERS_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    psrldq      xmm7, 8
+    pshuflw     xmm4, xmm7, 0b              ;k4
+    pshuflw     xmm5, xmm7, 01010101b       ;k5
+    pshuflw     xmm6, xmm7, 10101010b       ;k6
+    pshuflw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklqdq  xmm0, xmm1
+    punpcklqdq  xmm2, xmm3
+    punpcklqdq  xmm5, xmm4
+    punpcklqdq  xmm6, xmm7
+
+    movdqa      k0k1, xmm0
+    movdqa      k2k3, xmm2
+    movdqa      k5k4, xmm5
+    movdqa      k6k7, xmm6
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6
+
+    pxor        xmm7, xmm7
+    movdqa      zero, xmm7
+%endm
+
+%macro APPLY_FILTER_4 1
+    punpckldq   xmm0, xmm1                  ;two row in one register
+    punpckldq   xmm6, xmm7
+    punpckldq   xmm2, xmm3
+    punpckldq   xmm5, xmm4
+
+    punpcklbw   xmm0, zero                  ;unpack to word
+    punpcklbw   xmm6, zero
+    punpcklbw   xmm2, zero
+    punpcklbw   xmm5, zero
+
+    pmullw      xmm0, k0k1                  ;multiply the filter factors
+    pmullw      xmm6, k6k7
+    pmullw      xmm2, k2k3
+    pmullw      xmm5, k5k4
+
+    paddsw      xmm0, xmm6                  ;sum
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 8
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm2
+    psrldq      xmm2, 8
+    paddsw      xmm0, xmm5
+    psrldq      xmm5, 8
+    paddsw      xmm0, xmm2
+    paddsw      xmm0, xmm5
+
+    paddsw      xmm0, krd                   ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movd        [rdi], xmm0
+%endm
+
+%macro GET_FILTERS 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    pshufhw     xmm4, xmm7, 0b              ;k4
+    pshufhw     xmm5, xmm7, 01010101b       ;k5
+    pshufhw     xmm6, xmm7, 10101010b       ;k6
+    pshufhw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklwd   xmm0, xmm0
+    punpcklwd   xmm1, xmm1
+    punpcklwd   xmm2, xmm2
+    punpcklwd   xmm3, xmm3
+    punpckhwd   xmm4, xmm4
+    punpckhwd   xmm5, xmm5
+    punpckhwd   xmm6, xmm6
+    punpckhwd   xmm7, xmm7
+
+    movdqa      k0,   xmm0                  ;store filter factors on stack
+    movdqa      k1,   xmm1
+    movdqa      k2,   xmm2
+    movdqa      k3,   xmm3
+    movdqa      k4,   xmm4
+    movdqa      k5,   xmm5
+    movdqa      k6,   xmm6
+    movdqa      k7,   xmm7
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6                   ;rounding
+
+    pxor        xmm7, xmm7
+    movdqa      zero, xmm7
+%endm
+
+%macro LOAD_VERT_8 1
+    movq        xmm0, [rsi + %1]            ;0
+    movq        xmm1, [rsi + rax + %1]      ;1
+    movq        xmm6, [rsi + rdx * 2 + %1]  ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2 + %1]  ;7
+    movq        xmm2, [rsi + rax + %1]      ;2
+    movq        xmm3, [rsi + rax * 2 + %1]  ;3
+    movq        xmm4, [rsi + rdx + %1]      ;4
+    movq        xmm5, [rsi + rax * 4 + %1]  ;5
+%endm
+
+%macro APPLY_FILTER_8 2
+    punpcklbw   xmm0, zero
+    punpcklbw   xmm1, zero
+    punpcklbw   xmm6, zero
+    punpcklbw   xmm7, zero
+    punpcklbw   xmm2, zero
+    punpcklbw   xmm5, zero
+    punpcklbw   xmm3, zero
+    punpcklbw   xmm4, zero
+
+    pmullw      xmm0, k0
+    pmullw      xmm1, k1
+    pmullw      xmm6, k6
+    pmullw      xmm7, k7
+    pmullw      xmm2, k2
+    pmullw      xmm5, k5
+    pmullw      xmm3, k3
+    pmullw      xmm4, k4
+
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm6
+    paddsw      xmm0, xmm7
+    paddsw      xmm0, xmm2
+    paddsw      xmm0, xmm5
+    paddsw      xmm0, xmm3
+    paddsw      xmm0, xmm4
+
+    paddsw      xmm0, krd                   ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack back to byte
+%if %1
+    movq        xmm1, [rdi + %2]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi + %2], xmm0
+%endm
+
+;void vpx_filter_block1d4_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vpx_filter_block1d4_v8_sse2) PRIVATE
+sym(vpx_filter_block1d4_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movd        xmm0, [rsi]                 ;load src: row 0
+    movd        xmm1, [rsi + rax]           ;1
+    movd        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movd        xmm7, [rsi + rdx * 2]       ;7
+    movd        xmm2, [rsi + rax]           ;2
+    movd        xmm3, [rsi + rax * 2]       ;3
+    movd        xmm4, [rsi + rdx]           ;4
+    movd        xmm5, [rsi + rax * 4]       ;5
+
+    APPLY_FILTER_4 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vpx_filter_block1d8_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vpx_filter_block1d8_v8_sse2) PRIVATE
+sym(vpx_filter_block1d8_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 0, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vpx_filter_block1d16_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vpx_filter_block1d16_v8_sse2) PRIVATE
+sym(vpx_filter_block1d16_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 0, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 8
+    APPLY_FILTER_8 0, 8
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d4_v8_avg_sse2) PRIVATE
+sym(vpx_filter_block1d4_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movd        xmm0, [rsi]                 ;load src: row 0
+    movd        xmm1, [rsi + rax]           ;1
+    movd        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movd        xmm7, [rsi + rdx * 2]       ;7
+    movd        xmm2, [rsi + rax]           ;2
+    movd        xmm3, [rsi + rax * 2]       ;3
+    movd        xmm4, [rsi + rdx]           ;4
+    movd        xmm5, [rsi + rax * 4]       ;5
+
+    APPLY_FILTER_4 1
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d8_v8_avg_sse2) PRIVATE
+sym(vpx_filter_block1d8_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 1, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d16_v8_avg_sse2) PRIVATE
+sym(vpx_filter_block1d16_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 1, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 8
+    APPLY_FILTER_8 1, 8
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vpx_filter_block1d4_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vpx_filter_block1d4_h8_sse2) PRIVATE
+sym(vpx_filter_block1d4_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm3, 3
+    psrldq      xmm5, 5
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_4 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vpx_filter_block1d8_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vpx_filter_block1d8_h8_sse2) PRIVATE
+sym(vpx_filter_block1d8_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vpx_filter_block1d16_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vpx_filter_block1d16_h8_sse2) PRIVATE
+sym(vpx_filter_block1d16_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 0
+
+    movdqu      xmm0,   [rsi + 5]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 8
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d4_h8_avg_sse2) PRIVATE
+sym(vpx_filter_block1d4_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm3, 3
+    psrldq      xmm5, 5
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_4 1
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d8_h8_avg_sse2) PRIVATE
+sym(vpx_filter_block1d8_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 1, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d16_h8_avg_sse2) PRIVATE
+sym(vpx_filter_block1d16_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 1, 0
+
+    movdqu      xmm0,   [rsi + 5]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 1, 8
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_ssse3.asm b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_ssse3.asm
new file mode 100644
index 0000000..d2cb8ea
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_8t_ssse3.asm
@@ -0,0 +1,629 @@
+;
+;  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_64:    times 8 dw 64
+
+; %define USE_PMULHRSW
+; NOTE: pmulhrsw has a latency of 5 cycles.  Tests showed a performance loss
+; when using this instruction.
+;
+; The add order below (based on ffvp9) must be followed to prevent outranges.
+; x = k0k1 + k4k5
+; y = k2k3 + k6k7
+; z = signed SAT(x + y)
+
+SECTION .text
+%if ARCH_X86_64
+  %define LOCAL_VARS_SIZE 16*4
+%else
+  %define LOCAL_VARS_SIZE 16*6
+%endif
+
+%macro SETUP_LOCAL_VARS 0
+    ; TODO(slavarnway): using xmm registers for these on ARCH_X86_64 +
+    ; pmaddubsw has a higher latency on some platforms, this might be eased by
+    ; interleaving the instructions.
+    %define    k0k1  [rsp + 16*0]
+    %define    k2k3  [rsp + 16*1]
+    %define    k4k5  [rsp + 16*2]
+    %define    k6k7  [rsp + 16*3]
+    packsswb     m4, m4
+    ; TODO(slavarnway): multiple pshufb instructions had a higher latency on
+    ; some platforms.
+    pshuflw      m0, m4, 0b              ;k0_k1
+    pshuflw      m1, m4, 01010101b       ;k2_k3
+    pshuflw      m2, m4, 10101010b       ;k4_k5
+    pshuflw      m3, m4, 11111111b       ;k6_k7
+    punpcklqdq   m0, m0
+    punpcklqdq   m1, m1
+    punpcklqdq   m2, m2
+    punpcklqdq   m3, m3
+    mova       k0k1, m0
+    mova       k2k3, m1
+    mova       k4k5, m2
+    mova       k6k7, m3
+%if ARCH_X86_64
+    %define     krd  m12
+    %define     tmp  m13
+    mova        krd, [GLOBAL(pw_64)]
+%else
+    %define     tmp  [rsp + 16*4]
+    %define     krd  [rsp + 16*5]
+%if CONFIG_PIC=0
+    mova         m6, [GLOBAL(pw_64)]
+%else
+    ; build constants without accessing global memory
+    pcmpeqb      m6, m6                  ;all ones
+    psrlw        m6, 15
+    psllw        m6, 6                   ;aka pw_64
+%endif
+    mova        krd, m6
+%endif
+%endm
+
+%macro HORIZx4_ROW 2
+    mova      %2, %1
+    punpcklbw %1, %1
+    punpckhbw %2, %2
+
+    mova      m3, %2
+    palignr   %2, %1, 1
+    palignr   m3, %1, 5
+
+    pmaddubsw %2, k0k1k4k5
+    pmaddubsw m3, k2k3k6k7
+    mova      m4, %2        ;k0k1
+    mova      m5, m3        ;k2k3
+    psrldq    %2, 8         ;k4k5
+    psrldq    m3, 8         ;k6k7
+    paddsw    %2, m4
+    paddsw    m5, m3
+    paddsw    %2, m5
+    paddsw    %2, krd
+    psraw     %2, 7
+    packuswb  %2, %2
+%endm
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_HFILTER4 1
+cglobal filter_block1d4_%1, 6, 6+(ARCH_X86_64*2), 11, LOCAL_VARS_SIZE, \
+                            src, sstride, dst, dstride, height, filter
+    mova                m4, [filterq]
+    packsswb            m4, m4
+%if ARCH_X86_64
+    %define       k0k1k4k5 m8
+    %define       k2k3k6k7 m9
+    %define            krd m10
+    %define    orig_height r7d
+    mova               krd, [GLOBAL(pw_64)]
+    pshuflw       k0k1k4k5, m4, 0b              ;k0_k1
+    pshufhw       k0k1k4k5, k0k1k4k5, 10101010b ;k0_k1_k4_k5
+    pshuflw       k2k3k6k7, m4, 01010101b       ;k2_k3
+    pshufhw       k2k3k6k7, k2k3k6k7, 11111111b ;k2_k3_k6_k7
+%else
+    %define       k0k1k4k5 [rsp + 16*0]
+    %define       k2k3k6k7 [rsp + 16*1]
+    %define            krd [rsp + 16*2]
+    %define    orig_height [rsp + 16*3]
+    pshuflw             m6, m4, 0b              ;k0_k1
+    pshufhw             m6, m6, 10101010b       ;k0_k1_k4_k5
+    pshuflw             m7, m4, 01010101b       ;k2_k3
+    pshufhw             m7, m7, 11111111b       ;k2_k3_k6_k7
+%if CONFIG_PIC=0
+    mova                m1, [GLOBAL(pw_64)]
+%else
+    ; build constants without accessing global memory
+    pcmpeqb             m1, m1                  ;all ones
+    psrlw               m1, 15
+    psllw               m1, 6                   ;aka pw_64
+%endif
+    mova          k0k1k4k5, m6
+    mova          k2k3k6k7, m7
+    mova               krd, m1
+%endif
+    mov        orig_height, heightd
+    shr            heightd, 1
+.loop:
+    ;Do two rows at once
+    movh                m0, [srcq - 3]
+    movh                m1, [srcq + 5]
+    punpcklqdq          m0, m1
+    mova                m1, m0
+    movh                m2, [srcq + sstrideq - 3]
+    movh                m3, [srcq + sstrideq + 5]
+    punpcklqdq          m2, m3
+    mova                m3, m2
+    punpcklbw           m0, m0
+    punpckhbw           m1, m1
+    punpcklbw           m2, m2
+    punpckhbw           m3, m3
+    mova                m4, m1
+    palignr             m4, m0,  1
+    pmaddubsw           m4, k0k1k4k5
+    palignr             m1, m0,  5
+    pmaddubsw           m1, k2k3k6k7
+    mova                m7, m3
+    palignr             m7, m2,  1
+    pmaddubsw           m7, k0k1k4k5
+    palignr             m3, m2,  5
+    pmaddubsw           m3, k2k3k6k7
+    mova                m0, m4                  ;k0k1
+    mova                m5, m1                  ;k2k3
+    mova                m2, m7                  ;k0k1 upper
+    psrldq              m4, 8                   ;k4k5
+    psrldq              m1, 8                   ;k6k7
+    paddsw              m4, m0
+    paddsw              m5, m1
+    mova                m1, m3                  ;k2k3 upper
+    psrldq              m7, 8                   ;k4k5 upper
+    psrldq              m3, 8                   ;k6k7 upper
+    paddsw              m7, m2
+    paddsw              m4, m5
+    paddsw              m1, m3
+    paddsw              m7, m1
+    paddsw              m4, krd
+    psraw               m4, 7
+    packuswb            m4, m4
+    paddsw              m7, krd
+    psraw               m7, 7
+    packuswb            m7, m7
+
+%ifidn %1, h8_avg
+    movd                m0, [dstq]
+    pavgb               m4, m0
+    movd                m2, [dstq + dstrideq]
+    pavgb               m7, m2
+%endif
+    movd            [dstq], m4
+    movd [dstq + dstrideq], m7
+
+    lea               srcq, [srcq + sstrideq        ]
+    prefetcht0              [srcq + 4 * sstrideq - 3]
+    lea               srcq, [srcq + sstrideq        ]
+    lea               dstq, [dstq + 2 * dstrideq    ]
+    prefetcht0              [srcq + 2 * sstrideq - 3]
+
+    dec            heightd
+    jnz              .loop
+
+    ; Do last row if output_height is odd
+    mov            heightd, orig_height
+    and            heightd, 1
+    je               .done
+
+    movh                m0, [srcq - 3]    ; load src
+    movh                m1, [srcq + 5]
+    punpcklqdq          m0, m1
+
+    HORIZx4_ROW         m0, m1
+%ifidn %1, h8_avg
+    movd                m0, [dstq]
+    pavgb               m1, m0
+%endif
+    movd            [dstq], m1
+.done
+    RET
+%endm
+
+%macro HORIZx8_ROW 5
+    mova        %2, %1
+    punpcklbw   %1, %1
+    punpckhbw   %2, %2
+
+    mova        %3, %2
+    mova        %4, %2
+    mova        %5, %2
+
+    palignr     %2, %1, 1
+    palignr     %3, %1, 5
+    palignr     %4, %1, 9
+    palignr     %5, %1, 13
+
+    pmaddubsw   %2, k0k1
+    pmaddubsw   %3, k2k3
+    pmaddubsw   %4, k4k5
+    pmaddubsw   %5, k6k7
+    paddsw      %2, %4
+    paddsw      %5, %3
+    paddsw      %2, %5
+    paddsw      %2, krd
+    psraw       %2, 7
+    packuswb    %2, %2
+    SWAP        %1, %2
+%endm
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_HFILTER8 1
+cglobal filter_block1d8_%1, 6, 6+(ARCH_X86_64*1), 14, LOCAL_VARS_SIZE, \
+                            src, sstride, dst, dstride, height, filter
+    mova                 m4, [filterq]
+    SETUP_LOCAL_VARS
+%if ARCH_X86_64
+    %define     orig_height r7d
+%else
+    %define     orig_height heightmp
+%endif
+    mov         orig_height, heightd
+    shr             heightd, 1
+
+.loop:
+    movh                 m0, [srcq - 3]
+    movh                 m3, [srcq + 5]
+    movh                 m4, [srcq + sstrideq - 3]
+    movh                 m7, [srcq + sstrideq + 5]
+    punpcklqdq           m0, m3
+    mova                 m1, m0
+    punpcklbw            m0, m0
+    punpckhbw            m1, m1
+    mova                 m5, m1
+    palignr              m5, m0, 13
+    pmaddubsw            m5, k6k7
+    mova                 m2, m1
+    mova                 m3, m1
+    palignr              m1, m0, 1
+    pmaddubsw            m1, k0k1
+    punpcklqdq           m4, m7
+    mova                 m6, m4
+    punpcklbw            m4, m4
+    palignr              m2, m0, 5
+    punpckhbw            m6, m6
+    palignr              m3, m0, 9
+    mova                 m7, m6
+    pmaddubsw            m2, k2k3
+    pmaddubsw            m3, k4k5
+
+    palignr              m7, m4, 13
+    mova                 m0, m6
+    palignr              m0, m4, 5
+    pmaddubsw            m7, k6k7
+    paddsw               m1, m3
+    paddsw               m2, m5
+    paddsw               m1, m2
+    mova                 m5, m6
+    palignr              m6, m4, 1
+    pmaddubsw            m0, k2k3
+    pmaddubsw            m6, k0k1
+    palignr              m5, m4, 9
+    paddsw               m1, krd
+    pmaddubsw            m5, k4k5
+    psraw                m1, 7
+    paddsw               m0, m7
+%ifidn %1, h8_avg
+    movh                 m7, [dstq]
+    movh                 m2, [dstq + dstrideq]
+%endif
+    packuswb             m1, m1
+    paddsw               m6, m5
+    paddsw               m6, m0
+    paddsw               m6, krd
+    psraw                m6, 7
+    packuswb             m6, m6
+%ifidn %1, h8_avg
+    pavgb                m1, m7
+    pavgb                m6, m2
+%endif
+    movh             [dstq], m1
+    movh  [dstq + dstrideq], m6
+
+    lea                srcq, [srcq + sstrideq        ]
+    prefetcht0               [srcq + 4 * sstrideq - 3]
+    lea                srcq, [srcq + sstrideq        ]
+    lea                dstq, [dstq + 2 * dstrideq    ]
+    prefetcht0               [srcq + 2 * sstrideq - 3]
+    dec             heightd
+    jnz             .loop
+
+    ;Do last row if output_height is odd
+    mov             heightd, orig_height
+    and             heightd, 1
+    je                .done
+
+    movh                 m0, [srcq - 3]
+    movh                 m3, [srcq + 5]
+    punpcklqdq           m0, m3
+
+    HORIZx8_ROW          m0, m1, m2, m3, m4
+
+%ifidn %1, h8_avg
+    movh                 m1, [dstq]
+    pavgb                m0, m1
+%endif
+    movh             [dstq], m0
+.done:
+    RET
+%endm
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_HFILTER16 1
+cglobal filter_block1d16_%1, 6, 6+(ARCH_X86_64*0), 14, LOCAL_VARS_SIZE, \
+                             src, sstride, dst, dstride, height, filter
+    mova          m4, [filterq]
+    SETUP_LOCAL_VARS
+.loop:
+    prefetcht0        [srcq + 2 * sstrideq -3]
+
+    movh          m0, [srcq -  3]
+    movh          m4, [srcq +  5]
+    movh          m6, [srcq + 13]
+    punpcklqdq    m0, m4
+    mova          m7, m0
+    punpckhbw     m0, m0
+    mova          m1, m0
+    punpcklqdq    m4, m6
+    mova          m3, m0
+    punpcklbw     m7, m7
+
+    palignr       m3, m7, 13
+    mova          m2, m0
+    pmaddubsw     m3, k6k7
+    palignr       m0, m7, 1
+    pmaddubsw     m0, k0k1
+    palignr       m1, m7, 5
+    pmaddubsw     m1, k2k3
+    palignr       m2, m7, 9
+    pmaddubsw     m2, k4k5
+    paddsw        m1, m3
+    mova          m3, m4
+    punpckhbw     m4, m4
+    mova          m5, m4
+    punpcklbw     m3, m3
+    mova          m7, m4
+    palignr       m5, m3, 5
+    mova          m6, m4
+    palignr       m4, m3, 1
+    pmaddubsw     m4, k0k1
+    pmaddubsw     m5, k2k3
+    palignr       m6, m3, 9
+    pmaddubsw     m6, k4k5
+    palignr       m7, m3, 13
+    pmaddubsw     m7, k6k7
+    paddsw        m0, m2
+    paddsw        m0, m1
+%ifidn %1, h8_avg
+    mova          m1, [dstq]
+%endif
+    paddsw        m4, m6
+    paddsw        m5, m7
+    paddsw        m4, m5
+    paddsw        m0, krd
+    paddsw        m4, krd
+    psraw         m0, 7
+    psraw         m4, 7
+    packuswb      m0, m4
+%ifidn %1, h8_avg
+    pavgb         m0, m1
+%endif
+    lea         srcq, [srcq + sstrideq]
+    mova      [dstq], m0
+    lea         dstq, [dstq + dstrideq]
+    dec      heightd
+    jnz        .loop
+    RET
+%endm
+
+INIT_XMM ssse3
+SUBPIX_HFILTER16 h8
+SUBPIX_HFILTER16 h8_avg
+SUBPIX_HFILTER8  h8
+SUBPIX_HFILTER8  h8_avg
+SUBPIX_HFILTER4  h8
+SUBPIX_HFILTER4  h8_avg
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_VFILTER 2
+cglobal filter_block1d%2_%1, 6, 6+(ARCH_X86_64*3), 14, LOCAL_VARS_SIZE, \
+                             src, sstride, dst, dstride, height, filter
+    mova          m4, [filterq]
+    SETUP_LOCAL_VARS
+%if ARCH_X86_64
+    %define      src1q r7
+    %define  sstride6q r8
+    %define dst_stride dstrideq
+%else
+    %define      src1q filterq
+    %define  sstride6q dstrideq
+    %define dst_stride dstridemp
+%endif
+    mov       src1q, srcq
+    add       src1q, sstrideq
+    lea   sstride6q, [sstrideq + sstrideq * 4]
+    add   sstride6q, sstrideq                   ;pitch * 6
+
+%ifidn %2, 8
+    %define movx movh
+%else
+    %define movx movd
+%endif
+.loop:
+    movx         m0, [srcq                ]     ;A
+    movx         m1, [srcq + sstrideq     ]     ;B
+    punpcklbw    m0, m1                         ;A B
+    movx         m2, [srcq + sstrideq * 2 ]     ;C
+    pmaddubsw    m0, k0k1
+    mova         m6, m2
+    movx         m3, [src1q + sstrideq * 2]     ;D
+    punpcklbw    m2, m3                         ;C D
+    pmaddubsw    m2, k2k3
+    movx         m4, [srcq + sstrideq * 4 ]     ;E
+    mova         m7, m4
+    movx         m5, [src1q + sstrideq * 4]     ;F
+    punpcklbw    m4, m5                         ;E F
+    pmaddubsw    m4, k4k5
+    punpcklbw    m1, m6                         ;A B next iter
+    movx         m6, [srcq + sstride6q    ]     ;G
+    punpcklbw    m5, m6                         ;E F next iter
+    punpcklbw    m3, m7                         ;C D next iter
+    pmaddubsw    m5, k4k5
+    movx         m7, [src1q + sstride6q   ]     ;H
+    punpcklbw    m6, m7                         ;G H
+    pmaddubsw    m6, k6k7
+    pmaddubsw    m3, k2k3
+    pmaddubsw    m1, k0k1
+    paddsw       m0, m4
+    paddsw       m2, m6
+    movx         m6, [srcq + sstrideq * 8 ]     ;H next iter
+    punpcklbw    m7, m6
+    pmaddubsw    m7, k6k7
+    paddsw       m0, m2
+    paddsw       m0, krd
+    psraw        m0, 7
+    paddsw       m1, m5
+    packuswb     m0, m0
+
+    paddsw       m3, m7
+    paddsw       m1, m3
+    paddsw       m1, krd
+    psraw        m1, 7
+    lea        srcq, [srcq + sstrideq * 2 ]
+    lea       src1q, [src1q + sstrideq * 2]
+    packuswb     m1, m1
+
+%ifidn %1, v8_avg
+    movx         m2, [dstq]
+    pavgb        m0, m2
+%endif
+    movx     [dstq], m0
+    add        dstq, dst_stride
+%ifidn %1, v8_avg
+    movx         m3, [dstq]
+    pavgb        m1, m3
+%endif
+    movx     [dstq], m1
+    add        dstq, dst_stride
+    sub     heightd, 2
+    cmp     heightd, 1
+    jg        .loop
+
+    cmp     heightd, 0
+    je        .done
+
+    movx         m0, [srcq                ]     ;A
+    movx         m1, [srcq + sstrideq     ]     ;B
+    movx         m6, [srcq + sstride6q    ]     ;G
+    punpcklbw    m0, m1                         ;A B
+    movx         m7, [src1q + sstride6q   ]     ;H
+    pmaddubsw    m0, k0k1
+    movx         m2, [srcq + sstrideq * 2 ]     ;C
+    punpcklbw    m6, m7                         ;G H
+    movx         m3, [src1q + sstrideq * 2]     ;D
+    pmaddubsw    m6, k6k7
+    movx         m4, [srcq + sstrideq * 4 ]     ;E
+    punpcklbw    m2, m3                         ;C D
+    movx         m5, [src1q + sstrideq * 4]     ;F
+    punpcklbw    m4, m5                         ;E F
+    pmaddubsw    m2, k2k3
+    pmaddubsw    m4, k4k5
+    paddsw       m2, m6
+    paddsw       m0, m4
+    paddsw       m0, m2
+    paddsw       m0, krd
+    psraw        m0, 7
+    packuswb     m0, m0
+%ifidn %1, v8_avg
+    movx         m1, [dstq]
+    pavgb        m0, m1
+%endif
+    movx     [dstq], m0
+.done:
+    RET
+%endm
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_VFILTER16 1
+cglobal filter_block1d16_%1, 6, 6+(ARCH_X86_64*3), 14, LOCAL_VARS_SIZE, \
+                             src, sstride, dst, dstride, height, filter
+    mova          m4, [filterq]
+    SETUP_LOCAL_VARS
+%if ARCH_X86_64
+    %define      src1q r7
+    %define  sstride6q r8
+    %define dst_stride dstrideq
+%else
+    %define      src1q filterq
+    %define  sstride6q dstrideq
+    %define dst_stride dstridemp
+%endif
+    mov        src1q, srcq
+    add        src1q, sstrideq
+    lea    sstride6q, [sstrideq + sstrideq * 4]
+    add    sstride6q, sstrideq                   ;pitch * 6
+
+.loop:
+    movh          m0, [srcq                ]     ;A
+    movh          m1, [srcq + sstrideq     ]     ;B
+    movh          m2, [srcq + sstrideq * 2 ]     ;C
+    movh          m3, [src1q + sstrideq * 2]     ;D
+    movh          m4, [srcq + sstrideq * 4 ]     ;E
+    movh          m5, [src1q + sstrideq * 4]     ;F
+
+    punpcklbw     m0, m1                         ;A B
+    movh          m6, [srcq + sstride6q]         ;G
+    punpcklbw     m2, m3                         ;C D
+    movh          m7, [src1q + sstride6q]        ;H
+    punpcklbw     m4, m5                         ;E F
+    pmaddubsw     m0, k0k1
+    movh          m3, [srcq + 8]                 ;A
+    pmaddubsw     m2, k2k3
+    punpcklbw     m6, m7                         ;G H
+    movh          m5, [srcq + sstrideq + 8]      ;B
+    pmaddubsw     m4, k4k5
+    punpcklbw     m3, m5                         ;A B
+    movh          m7, [srcq + sstrideq * 2 + 8]  ;C
+    pmaddubsw     m6, k6k7
+    movh          m5, [src1q + sstrideq * 2 + 8] ;D
+    punpcklbw     m7, m5                         ;C D
+    paddsw        m2, m6
+    pmaddubsw     m3, k0k1
+    movh          m1, [srcq + sstrideq * 4 + 8]  ;E
+    paddsw        m0, m4
+    pmaddubsw     m7, k2k3
+    movh          m6, [src1q + sstrideq * 4 + 8] ;F
+    punpcklbw     m1, m6                         ;E F
+    paddsw        m0, m2
+    paddsw        m0, krd
+    movh          m2, [srcq + sstride6q + 8]     ;G
+    pmaddubsw     m1, k4k5
+    movh          m5, [src1q + sstride6q + 8]    ;H
+    psraw         m0, 7
+    punpcklbw     m2, m5                         ;G H
+    pmaddubsw     m2, k6k7
+%ifidn %1, v8_avg
+    mova          m4, [dstq]
+%endif
+    movh      [dstq], m0
+    paddsw        m7, m2
+    paddsw        m3, m1
+    paddsw        m3, m7
+    paddsw        m3, krd
+    psraw         m3, 7
+    packuswb      m0, m3
+
+    add         srcq, sstrideq
+    add        src1q, sstrideq
+%ifidn %1, v8_avg
+    pavgb         m0, m4
+%endif
+    mova      [dstq], m0
+    add         dstq, dst_stride
+    dec      heightd
+    jnz        .loop
+    RET
+%endm
+
+INIT_XMM ssse3
+SUBPIX_VFILTER16     v8
+SUBPIX_VFILTER16 v8_avg
+SUBPIX_VFILTER       v8, 8
+SUBPIX_VFILTER   v8_avg, 8
+SUBPIX_VFILTER       v8, 4
+SUBPIX_VFILTER   v8_avg, 4
diff --git a/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_bilinear_sse2.asm b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_bilinear_sse2.asm
new file mode 100644
index 0000000..a378dd0
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_bilinear_sse2.asm
@@ -0,0 +1,448 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    pshuflw     xmm4, xmm3, 11111111b       ;k3
+    psrldq      xmm3, 8
+    pshuflw     xmm3, xmm3, 0b              ;k4
+    punpcklqdq  xmm4, xmm3                  ;k3k4
+
+    movq        xmm3, rcx                   ;rounding
+    pshufd      xmm3, xmm3, 0
+
+    pxor        xmm2, xmm2
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_4 1
+
+    punpckldq   xmm0, xmm1                  ;two row in one register
+    punpcklbw   xmm0, xmm2                  ;unpack to word
+    pmullw      xmm0, xmm4                  ;multiply the filter factors
+
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 8
+    paddsw      xmm0, xmm1
+
+    paddsw      xmm0, xmm3                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+
+    movd        [rdi], xmm0
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+
+    pshuflw     xmm6, xmm7, 11111111b       ;k3
+    pshufhw     xmm7, xmm7, 0b              ;k4
+    punpcklwd   xmm6, xmm6
+    punpckhwd   xmm7, xmm7
+
+    movq        xmm4, rcx                   ;rounding
+    pshufd      xmm4, xmm4, 0
+
+    pxor        xmm5, xmm5
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_8 1
+    punpcklbw   xmm0, xmm5
+    punpcklbw   xmm1, xmm5
+
+    pmullw      xmm0, xmm6
+    pmullw      xmm1, xmm7
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm4                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack back to byte
+%if %1
+    movq        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro APPLY_FILTER_16 1
+    punpcklbw   xmm0, xmm5
+    punpcklbw   xmm1, xmm5
+    punpckhbw   xmm2, xmm5
+    punpckhbw   xmm3, xmm5
+
+    pmullw      xmm0, xmm6
+    pmullw      xmm1, xmm7
+    pmullw      xmm2, xmm6
+    pmullw      xmm3, xmm7
+
+    paddsw      xmm0, xmm1
+    paddsw      xmm2, xmm3
+
+    paddsw      xmm0, xmm4                  ;rounding
+    paddsw      xmm2, xmm4
+    psraw       xmm0, 7                     ;shift
+    psraw       xmm2, 7
+    packuswb    xmm0, xmm2                  ;pack back to byte
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+global sym(vpx_filter_block1d4_v2_sse2) PRIVATE
+sym(vpx_filter_block1d4_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d8_v2_sse2) PRIVATE
+sym(vpx_filter_block1d8_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d16_v2_sse2) PRIVATE
+sym(vpx_filter_block1d16_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+    movdqa        xmm3, xmm1
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d4_v2_avg_sse2) PRIVATE
+sym(vpx_filter_block1d4_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d8_v2_avg_sse2) PRIVATE
+sym(vpx_filter_block1d8_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d16_v2_avg_sse2) PRIVATE
+sym(vpx_filter_block1d16_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+    movdqa        xmm3, xmm1
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d4_h2_sse2) PRIVATE
+sym(vpx_filter_block1d4_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d8_h2_sse2) PRIVATE
+sym(vpx_filter_block1d8_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d16_h2_sse2) PRIVATE
+sym(vpx_filter_block1d16_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm1
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d4_h2_avg_sse2) PRIVATE
+sym(vpx_filter_block1d4_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d8_h2_avg_sse2) PRIVATE
+sym(vpx_filter_block1d8_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d16_h2_avg_sse2) PRIVATE
+sym(vpx_filter_block1d16_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm1
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_bilinear_ssse3.asm b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_bilinear_ssse3.asm
new file mode 100644
index 0000000..3c8cfd2
--- /dev/null
+++ b/libvpx/libvpx/vpx_dsp/x86/vpx_subpixel_bilinear_ssse3.asm
@@ -0,0 +1,422 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    psrldq      xmm3, 6
+    packsswb    xmm3, xmm3
+    pshuflw     xmm3, xmm3, 0b              ;k3_k4
+
+    movq        xmm2, rcx                   ;rounding
+    pshufd      xmm2, xmm2, 0
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_4 1
+    punpcklbw   xmm0, xmm1
+    pmaddubsw   xmm0, xmm3
+
+    paddsw      xmm0, xmm2                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movd        [rdi], xmm0
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    psrldq      xmm7, 6
+    packsswb    xmm7, xmm7
+    pshuflw     xmm7, xmm7, 0b              ;k3_k4
+    punpcklwd   xmm7, xmm7
+
+    movq        xmm6, rcx                   ;rounding
+    pshufd      xmm6, xmm6, 0
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_8 1
+    punpcklbw   xmm0, xmm1
+    pmaddubsw   xmm0, xmm7
+
+    paddsw      xmm0, xmm6                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack back to byte
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro APPLY_FILTER_16 1
+    punpcklbw   xmm0, xmm1
+    punpckhbw   xmm2, xmm1
+    pmaddubsw   xmm0, xmm7
+    pmaddubsw   xmm2, xmm7
+
+    paddsw      xmm0, xmm6                  ;rounding
+    paddsw      xmm2, xmm6
+    psraw       xmm0, 7                     ;shift
+    psraw       xmm2, 7
+    packuswb    xmm0, xmm2                  ;pack back to byte
+
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+global sym(vpx_filter_block1d4_v2_ssse3) PRIVATE
+sym(vpx_filter_block1d4_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d8_v2_ssse3) PRIVATE
+sym(vpx_filter_block1d8_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d16_v2_ssse3) PRIVATE
+sym(vpx_filter_block1d16_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d4_v2_avg_ssse3) PRIVATE
+sym(vpx_filter_block1d4_v2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d8_v2_avg_ssse3) PRIVATE
+sym(vpx_filter_block1d8_v2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d16_v2_avg_ssse3) PRIVATE
+sym(vpx_filter_block1d16_v2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d4_h2_ssse3) PRIVATE
+sym(vpx_filter_block1d4_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d8_h2_ssse3) PRIVATE
+sym(vpx_filter_block1d8_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d16_h2_ssse3) PRIVATE
+sym(vpx_filter_block1d16_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d4_h2_avg_ssse3) PRIVATE
+sym(vpx_filter_block1d4_h2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d8_h2_avg_ssse3) PRIVATE
+sym(vpx_filter_block1d8_h2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vpx_filter_block1d16_h2_avg_ssse3) PRIVATE
+sym(vpx_filter_block1d16_h2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/libvpx/libvpx/vpx_mem/include/vpx_mem_intrnl.h b/libvpx/libvpx/vpx_mem/include/vpx_mem_intrnl.h
new file mode 100644
index 0000000..c4dd785
--- /dev/null
+++ b/libvpx/libvpx/vpx_mem/include/vpx_mem_intrnl.h
@@ -0,0 +1,31 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VPX_MEM_INCLUDE_VPX_MEM_INTRNL_H_
+#define VPX_MEM_INCLUDE_VPX_MEM_INTRNL_H_
+#include "./vpx_config.h"
+
+#define ADDRESS_STORAGE_SIZE      sizeof(size_t)
+
+#ifndef DEFAULT_ALIGNMENT
+# if defined(VXWORKS)
+#  define DEFAULT_ALIGNMENT        32        /*default addr alignment to use in
+calls to vpx_* functions other
+than vpx_memalign*/
+# else
+#  define DEFAULT_ALIGNMENT        (2 * sizeof(void*))  /* NOLINT */
+# endif
+#endif
+
+/*returns an addr aligned to the byte boundary specified by align*/
+#define align_addr(addr,align) (void*)(((size_t)(addr) + ((align) - 1)) & (size_t)-(align))
+
+#endif  // VPX_MEM_INCLUDE_VPX_MEM_INTRNL_H_
diff --git a/libvpx/libvpx/vpx_mem/vpx_mem.c b/libvpx/libvpx/vpx_mem/vpx_mem.c
new file mode 100644
index 0000000..b261fc0
--- /dev/null
+++ b/libvpx/libvpx/vpx_mem/vpx_mem.c
@@ -0,0 +1,100 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "vpx_mem.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "include/vpx_mem_intrnl.h"
+#include "vpx/vpx_integer.h"
+
+void *vpx_memalign(size_t align, size_t size) {
+  void *addr,
+       * x = NULL;
+
+  addr = malloc(size + align - 1 + ADDRESS_STORAGE_SIZE);
+
+  if (addr) {
+    x = align_addr((unsigned char *)addr + ADDRESS_STORAGE_SIZE, (int)align);
+    /* save the actual malloc address */
+    ((size_t *)x)[-1] = (size_t)addr;
+  }
+
+  return x;
+}
+
+void *vpx_malloc(size_t size) {
+  return vpx_memalign(DEFAULT_ALIGNMENT, size);
+}
+
+void *vpx_calloc(size_t num, size_t size) {
+  void *x;
+
+  x = vpx_memalign(DEFAULT_ALIGNMENT, num * size);
+
+  if (x)
+    memset(x, 0, num * size);
+
+  return x;
+}
+
+void *vpx_realloc(void *memblk, size_t size) {
+  void *addr,
+       * new_addr = NULL;
+  int align = DEFAULT_ALIGNMENT;
+
+  /*
+  The realloc() function changes the size of the object pointed to by
+  ptr to the size specified by size, and returns a pointer to the
+  possibly moved block. The contents are unchanged up to the lesser
+  of the new and old sizes. If ptr is null, realloc() behaves like
+  malloc() for the specified size. If size is zero (0) and ptr is
+  not a null pointer, the object pointed to is freed.
+  */
+  if (!memblk)
+    new_addr = vpx_malloc(size);
+  else if (!size)
+    vpx_free(memblk);
+  else {
+    addr   = (void *)(((size_t *)memblk)[-1]);
+    memblk = NULL;
+
+    new_addr = realloc(addr, size + align + ADDRESS_STORAGE_SIZE);
+
+    if (new_addr) {
+      addr = new_addr;
+      new_addr = (void *)(((size_t)
+                           ((unsigned char *)new_addr + ADDRESS_STORAGE_SIZE) + (align - 1)) &
+                          (size_t) - align);
+      /* save the actual malloc address */
+      ((size_t *)new_addr)[-1] = (size_t)addr;
+    }
+  }
+
+  return new_addr;
+}
+
+void vpx_free(void *memblk) {
+  if (memblk) {
+    void *addr = (void *)(((size_t *)memblk)[-1]);
+    free(addr);
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void *vpx_memset16(void *dest, int val, size_t length) {
+  size_t i;
+  uint16_t *dest16 = (uint16_t *)dest;
+  for (i = 0; i < length; i++)
+    *dest16++ = val;
+  return dest;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/libvpx/libvpx/vpx_mem/vpx_mem.h b/libvpx/libvpx/vpx_mem/vpx_mem.h
new file mode 100644
index 0000000..a006e0f
--- /dev/null
+++ b/libvpx/libvpx/vpx_mem/vpx_mem.h
@@ -0,0 +1,47 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VPX_MEM_VPX_MEM_H_
+#define VPX_MEM_VPX_MEM_H_
+
+#include "vpx_config.h"
+#if defined(__uClinux__)
+# include <lddk.h>
+#endif
+
+#include <stdlib.h>
+#include <stddef.h>
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+  void *vpx_memalign(size_t align, size_t size);
+  void *vpx_malloc(size_t size);
+  void *vpx_calloc(size_t num, size_t size);
+  void *vpx_realloc(void *memblk, size_t size);
+  void vpx_free(void *memblk);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  void *vpx_memset16(void *dest, int val, size_t length);
+#endif
+
+#include <string.h>
+
+#ifdef VPX_MEM_PLTFRM
+# include VPX_MEM_PLTFRM
+#endif
+
+#if defined(__cplusplus)
+}
+#endif
+
+#endif  // VPX_MEM_VPX_MEM_H_
diff --git a/libvpx/libvpx/vpx_mem/vpx_mem.mk b/libvpx/libvpx/vpx_mem/vpx_mem.mk
new file mode 100644
index 0000000..7f275ea
--- /dev/null
+++ b/libvpx/libvpx/vpx_mem/vpx_mem.mk
@@ -0,0 +1,4 @@
+MEM_SRCS-yes += vpx_mem.mk
+MEM_SRCS-yes += vpx_mem.c
+MEM_SRCS-yes += vpx_mem.h
+MEM_SRCS-yes += include/vpx_mem_intrnl.h
diff --git a/libvpx/libvpx/vpx_ports/arm.h b/libvpx/libvpx/vpx_ports/arm.h
new file mode 100644
index 0000000..42c98f5
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/arm.h
@@ -0,0 +1,41 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VPX_PORTS_ARM_H_
+#define VPX_PORTS_ARM_H_
+#include <stdlib.h>
+#include "vpx_config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*ARMv5TE "Enhanced DSP" instructions.*/
+#define HAS_EDSP  0x01
+/*ARMv6 "Parallel" or "Media" instructions.*/
+#define HAS_MEDIA 0x02
+/*ARMv7 optional NEON instructions.*/
+#define HAS_NEON  0x04
+
+int arm_cpu_caps(void);
+
+// Earlier gcc compilers have issues with some neon intrinsics
+#if !defined(__clang__) && defined(__GNUC__) && \
+    __GNUC__ == 4 && __GNUC_MINOR__ <= 6
+#define VPX_INCOMPATIBLE_GCC
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_PORTS_ARM_H_
+
diff --git a/libvpx/libvpx/vpx_ports/arm_cpudetect.c b/libvpx/libvpx/vpx_ports/arm_cpudetect.c
new file mode 100644
index 0000000..8a4b8af
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/arm_cpudetect.c
@@ -0,0 +1,175 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+#include "vpx_ports/arm.h"
+#include "./vpx_config.h"
+
+#ifdef WINAPI_FAMILY
+#include <winapifamily.h>
+#if !WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
+#define getenv(x) NULL
+#endif
+#endif
+
+static int arm_cpu_env_flags(int *flags) {
+  char *env;
+  env = getenv("VPX_SIMD_CAPS");
+  if (env && *env) {
+    *flags = (int)strtol(env, NULL, 0);
+    return 0;
+  }
+  *flags = 0;
+  return -1;
+}
+
+static int arm_cpu_env_mask(void) {
+  char *env;
+  env = getenv("VPX_SIMD_CAPS_MASK");
+  return env && *env ? (int)strtol(env, NULL, 0) : ~0;
+}
+
+#if !CONFIG_RUNTIME_CPU_DETECT
+
+int arm_cpu_caps(void) {
+  /* This function should actually be a no-op. There is no way to adjust any of
+   * these because the RTCD tables do not exist: the functions are called
+   * statically */
+  int flags;
+  int mask;
+  if (!arm_cpu_env_flags(&flags)) {
+    return flags;
+  }
+  mask = arm_cpu_env_mask();
+#if HAVE_MEDIA
+  flags |= HAS_MEDIA;
+#endif /* HAVE_MEDIA */
+#if HAVE_NEON || HAVE_NEON_ASM
+  flags |= HAS_NEON;
+#endif /* HAVE_NEON  || HAVE_NEON_ASM */
+  return flags & mask;
+}
+
+#elif defined(_MSC_VER) /* end !CONFIG_RUNTIME_CPU_DETECT */
+/*For GetExceptionCode() and EXCEPTION_ILLEGAL_INSTRUCTION.*/
+#define WIN32_LEAN_AND_MEAN
+#define WIN32_EXTRA_LEAN
+#include <windows.h>
+
+int arm_cpu_caps(void) {
+  int flags;
+  int mask;
+  if (!arm_cpu_env_flags(&flags)) {
+    return flags;
+  }
+  mask = arm_cpu_env_mask();
+  /* MSVC has no inline __asm support for ARM, but it does let you __emit
+   *  instructions via their assembled hex code.
+   * All of these instructions should be essentially nops.
+   */
+#if HAVE_MEDIA
+  if (mask & HAS_MEDIA)
+    __try {
+      /*SHADD8 r3,r3,r3*/
+      __emit(0xE6333F93);
+      flags |= HAS_MEDIA;
+    } __except (GetExceptionCode() == EXCEPTION_ILLEGAL_INSTRUCTION) {
+    /*Ignore exception.*/
+  }
+}
+#endif /* HAVE_MEDIA */
+#if HAVE_NEON || HAVE_NEON_ASM
+if (mask &HAS_NEON) {
+  __try {
+    /*VORR q0,q0,q0*/
+    __emit(0xF2200150);
+    flags |= HAS_NEON;
+  } __except (GetExceptionCode() == EXCEPTION_ILLEGAL_INSTRUCTION) {
+    /*Ignore exception.*/
+  }
+}
+#endif /* HAVE_NEON || HAVE_NEON_ASM */
+return flags & mask;
+}
+
+#elif defined(__ANDROID__) /* end _MSC_VER */
+#include <cpu-features.h>
+
+int arm_cpu_caps(void) {
+  int flags;
+  int mask;
+  uint64_t features;
+  if (!arm_cpu_env_flags(&flags)) {
+    return flags;
+  }
+  mask = arm_cpu_env_mask();
+  features = android_getCpuFeatures();
+
+#if HAVE_MEDIA
+  flags |= HAS_MEDIA;
+#endif /* HAVE_MEDIA */
+#if HAVE_NEON || HAVE_NEON_ASM
+  if (features & ANDROID_CPU_ARM_FEATURE_NEON)
+    flags |= HAS_NEON;
+#endif /* HAVE_NEON || HAVE_NEON_ASM */
+  return flags & mask;
+}
+
+#elif defined(__linux__) /* end __ANDROID__ */
+
+#include <stdio.h>
+
+int arm_cpu_caps(void) {
+  FILE *fin;
+  int flags;
+  int mask;
+  if (!arm_cpu_env_flags(&flags)) {
+    return flags;
+  }
+  mask = arm_cpu_env_mask();
+  /* Reading /proc/self/auxv would be easier, but that doesn't work reliably
+   *  on Android.
+   * This also means that detection will fail in Scratchbox.
+   */
+  fin = fopen("/proc/cpuinfo", "r");
+  if (fin != NULL) {
+    /* 512 should be enough for anybody (it's even enough for all the flags
+     * that x86 has accumulated... so far).
+     */
+    char buf[512];
+    while (fgets(buf, 511, fin) != NULL) {
+#if HAVE_NEON || HAVE_NEON_ASM
+      if (memcmp(buf, "Features", 8) == 0) {
+        char *p;
+        p = strstr(buf, " neon");
+        if (p != NULL && (p[5] == ' ' || p[5] == '\n')) {
+          flags |= HAS_NEON;
+        }
+      }
+#endif /* HAVE_NEON || HAVE_NEON_ASM */
+#if HAVE_MEDIA
+      if (memcmp(buf, "CPU architecture:", 17) == 0) {
+        int version;
+        version = atoi(buf + 17);
+        if (version >= 6) {
+          flags |= HAS_MEDIA;
+        }
+      }
+#endif /* HAVE_MEDIA */
+    }
+    fclose(fin);
+  }
+  return flags & mask;
+}
+#else /* end __linux__ */
+#error "--enable-runtime-cpu-detect selected, but no CPU detection method " \
+"available for your platform. Reconfigure with --disable-runtime-cpu-detect."
+#endif
diff --git a/libvpx/libvpx/vpx_ports/bitops.h b/libvpx/libvpx/vpx_ports/bitops.h
new file mode 100644
index 0000000..84ff365
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/bitops.h
@@ -0,0 +1,76 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_PORTS_BITOPS_H_
+#define VPX_PORTS_BITOPS_H_
+
+#include <assert.h>
+
+#include "vpx_ports/msvc.h"
+
+#ifdef _MSC_VER
+# include <math.h>  // the ceil() definition must precede intrin.h
+# if _MSC_VER > 1310 && (defined(_M_X64) || defined(_M_IX86))
+#  include <intrin.h>
+#  define USE_MSC_INTRINSICS
+# endif
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// These versions of get_msb() are only valid when n != 0 because all
+// of the optimized versions are undefined when n == 0:
+// https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html
+
+// use GNU builtins where available.
+#if defined(__GNUC__) && \
+    ((__GNUC__ == 3 && __GNUC_MINOR__ >= 4) || __GNUC__ >= 4)
+static INLINE int get_msb(unsigned int n) {
+  assert(n != 0);
+  return 31 ^ __builtin_clz(n);
+}
+#elif defined(USE_MSC_INTRINSICS)
+#pragma intrinsic(_BitScanReverse)
+
+static INLINE int get_msb(unsigned int n) {
+  unsigned long first_set_bit;
+  assert(n != 0);
+  _BitScanReverse(&first_set_bit, n);
+  return first_set_bit;
+}
+#undef USE_MSC_INTRINSICS
+#else
+// Returns (int)floor(log2(n)). n must be > 0.
+static INLINE int get_msb(unsigned int n) {
+  int log = 0;
+  unsigned int value = n;
+  int i;
+
+  assert(n != 0);
+
+  for (i = 4; i >= 0; --i) {
+    const int shift = (1 << i);
+    const unsigned int x = value >> shift;
+    if (x != 0) {
+      value = x;
+      log += shift;
+    }
+  }
+  return log;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_PORTS_BITOPS_H_
diff --git a/libvpx/libvpx/vpx_ports/config.h b/libvpx/libvpx/vpx_ports/config.h
new file mode 100644
index 0000000..3c1ab99
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/config.h
@@ -0,0 +1,16 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_PORTS_CONFIG_H_
+#define VPX_PORTS_CONFIG_H_
+
+#include "vpx_config.h"
+
+#endif  // VPX_PORTS_CONFIG_H_
diff --git a/libvpx/libvpx/vpx_ports/emmintrin_compat.h b/libvpx/libvpx/vpx_ports/emmintrin_compat.h
new file mode 100644
index 0000000..1617638
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/emmintrin_compat.h
@@ -0,0 +1,55 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_PORTS_EMMINTRIN_COMPAT_H_
+#define VPX_PORTS_EMMINTRIN_COMPAT_H_
+
+#if defined(__GNUC__) && __GNUC__ < 4
+/* From emmintrin.h (gcc 4.5.3) */
+/* Casts between various SP, DP, INT vector types.  Note that these do no
+   conversion of values, they just change the type.  */
+extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_castpd_ps(__m128d __A)
+{
+  return (__m128) __A;
+}
+
+extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_castpd_si128(__m128d __A)
+{
+  return (__m128i) __A;
+}
+
+extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_castps_pd(__m128 __A)
+{
+  return (__m128d) __A;
+}
+
+extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_castps_si128(__m128 __A)
+{
+  return (__m128i) __A;
+}
+
+extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_castsi128_ps(__m128i __A)
+{
+  return (__m128) __A;
+}
+
+extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_castsi128_pd(__m128i __A)
+{
+  return (__m128d) __A;
+}
+#endif
+
+#endif  // VPX_PORTS_EMMINTRIN_COMPAT_H_
diff --git a/libvpx/libvpx/vpx_ports/emms.asm b/libvpx/libvpx/vpx_ports/emms.asm
new file mode 100644
index 0000000..db8da28
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/emms.asm
@@ -0,0 +1,38 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+section .text
+global sym(vpx_reset_mmx_state) PRIVATE
+sym(vpx_reset_mmx_state):
+    emms
+    ret
+
+
+%if LIBVPX_YASM_WIN64
+global sym(vpx_winx64_fldcw) PRIVATE
+sym(vpx_winx64_fldcw):
+    sub   rsp, 8
+    mov   [rsp], rcx ; win x64 specific
+    fldcw [rsp]
+    add   rsp, 8
+    ret
+
+
+global sym(vpx_winx64_fstcw) PRIVATE
+sym(vpx_winx64_fstcw):
+    sub   rsp, 8
+    fstcw [rsp]
+    mov   rax, [rsp]
+    add   rsp, 8
+    ret
+%endif
diff --git a/libvpx/libvpx/vpx_ports/mem.h b/libvpx/libvpx/vpx_ports/mem.h
new file mode 100644
index 0000000..7502f90
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/mem.h
@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VPX_PORTS_MEM_H_
+#define VPX_PORTS_MEM_H_
+
+#include "vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+#if (defined(__GNUC__) && __GNUC__) || defined(__SUNPRO_C)
+#define DECLARE_ALIGNED(n,typ,val)  typ val __attribute__ ((aligned (n)))
+#elif defined(_MSC_VER)
+#define DECLARE_ALIGNED(n,typ,val)  __declspec(align(n)) typ val
+#else
+#warning No alignment directives known for this compiler.
+#define DECLARE_ALIGNED(n,typ,val)  typ val
+#endif
+
+/* Indicates that the usage of the specified variable has been audited to assure
+ * that it's safe to use uninitialized. Silences 'may be used uninitialized'
+ * warnings on gcc.
+ */
+#if defined(__GNUC__) && __GNUC__
+#define UNINITIALIZED_IS_SAFE(x) x=x
+#else
+#define UNINITIALIZED_IS_SAFE(x) x
+#endif
+
+#if HAVE_NEON && defined(_MSC_VER)
+#define __builtin_prefetch(x)
+#endif
+
+/* Shift down with rounding */
+#define ROUND_POWER_OF_TWO(value, n) \
+    (((value) + (1 << ((n) - 1))) >> (n))
+
+#define ALIGN_POWER_OF_TWO(value, n) \
+    (((value) + ((1 << (n)) - 1)) & ~((1 << (n)) - 1))
+
+#if CONFIG_VP9_HIGHBITDEPTH
+#define CONVERT_TO_SHORTPTR(x) ((uint16_t*)(((uintptr_t)x) << 1))
+#define CONVERT_TO_BYTEPTR(x) ((uint8_t*)(((uintptr_t)x) >> 1))
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#endif  // VPX_PORTS_MEM_H_
diff --git a/libvpx/libvpx/vpx_ports/mem_ops.h b/libvpx/libvpx/vpx_ports/mem_ops.h
new file mode 100644
index 0000000..620df31
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/mem_ops.h
@@ -0,0 +1,226 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_PORTS_MEM_OPS_H_
+#define VPX_PORTS_MEM_OPS_H_
+
+/* \file
+ * \brief Provides portable memory access primitives
+ *
+ * This function provides portable primitives for getting and setting of
+ * signed and unsigned integers in 16, 24, and 32 bit sizes. The operations
+ * can be performed on unaligned data regardless of hardware support for
+ * unaligned accesses.
+ *
+ * The type used to pass the integral values may be changed by defining
+ * MEM_VALUE_T with the appropriate type. The type given must be an integral
+ * numeric type.
+ *
+ * The actual functions instantiated have the MEM_VALUE_T type name pasted
+ * on to the symbol name. This allows the developer to instantiate these
+ * operations for multiple types within the same translation unit. This is
+ * of somewhat questionable utility, but the capability exists nonetheless.
+ * Users not making use of this functionality should call the functions
+ * without the type name appended, and the preprocessor will take care of
+ * it.
+ *
+ * NOTE: This code is not supported on platforms where char > 1 octet ATM.
+ */
+
+#ifndef MAU_T
+/* Minimum Access Unit for this target */
+#define MAU_T unsigned char
+#endif
+
+#ifndef MEM_VALUE_T
+#define MEM_VALUE_T int
+#endif
+
+#undef MEM_VALUE_T_SZ_BITS
+#define MEM_VALUE_T_SZ_BITS (sizeof(MEM_VALUE_T) << 3)
+
+#undef  mem_ops_wrap_symbol
+#define mem_ops_wrap_symbol(fn) mem_ops_wrap_symbol2(fn, MEM_VALUE_T)
+#undef  mem_ops_wrap_symbol2
+#define mem_ops_wrap_symbol2(fn,typ) mem_ops_wrap_symbol3(fn,typ)
+#undef  mem_ops_wrap_symbol3
+#define mem_ops_wrap_symbol3(fn,typ) fn##_as_##typ
+
+/*
+ * Include aligned access routines
+ */
+#define INCLUDED_BY_MEM_OPS_H
+#include "mem_ops_aligned.h"
+#undef  INCLUDED_BY_MEM_OPS_H
+
+#undef  mem_get_be16
+#define mem_get_be16 mem_ops_wrap_symbol(mem_get_be16)
+static unsigned MEM_VALUE_T mem_get_be16(const void *vmem) {
+  unsigned MEM_VALUE_T  val;
+  const MAU_T          *mem = (const MAU_T *)vmem;
+
+  val = mem[0] << 8;
+  val |= mem[1];
+  return val;
+}
+
+#undef  mem_get_be24
+#define mem_get_be24 mem_ops_wrap_symbol(mem_get_be24)
+static unsigned MEM_VALUE_T mem_get_be24(const void *vmem) {
+  unsigned MEM_VALUE_T  val;
+  const MAU_T          *mem = (const MAU_T *)vmem;
+
+  val = mem[0] << 16;
+  val |= mem[1] << 8;
+  val |= mem[2];
+  return val;
+}
+
+#undef  mem_get_be32
+#define mem_get_be32 mem_ops_wrap_symbol(mem_get_be32)
+static unsigned MEM_VALUE_T mem_get_be32(const void *vmem) {
+  unsigned MEM_VALUE_T  val;
+  const MAU_T          *mem = (const MAU_T *)vmem;
+
+  val = ((unsigned MEM_VALUE_T)mem[0]) << 24;
+  val |= mem[1] << 16;
+  val |= mem[2] << 8;
+  val |= mem[3];
+  return val;
+}
+
+#undef  mem_get_le16
+#define mem_get_le16 mem_ops_wrap_symbol(mem_get_le16)
+static unsigned MEM_VALUE_T mem_get_le16(const void *vmem) {
+  unsigned MEM_VALUE_T  val;
+  const MAU_T          *mem = (const MAU_T *)vmem;
+
+  val = mem[1] << 8;
+  val |= mem[0];
+  return val;
+}
+
+#undef  mem_get_le24
+#define mem_get_le24 mem_ops_wrap_symbol(mem_get_le24)
+static unsigned MEM_VALUE_T mem_get_le24(const void *vmem) {
+  unsigned MEM_VALUE_T  val;
+  const MAU_T          *mem = (const MAU_T *)vmem;
+
+  val = mem[2] << 16;
+  val |= mem[1] << 8;
+  val |= mem[0];
+  return val;
+}
+
+#undef  mem_get_le32
+#define mem_get_le32 mem_ops_wrap_symbol(mem_get_le32)
+static unsigned MEM_VALUE_T mem_get_le32(const void *vmem) {
+  unsigned MEM_VALUE_T  val;
+  const MAU_T          *mem = (const MAU_T *)vmem;
+
+  val = ((unsigned MEM_VALUE_T)mem[3]) << 24;
+  val |= mem[2] << 16;
+  val |= mem[1] << 8;
+  val |= mem[0];
+  return val;
+}
+
+#define mem_get_s_generic(end,sz) \
+  static VPX_INLINE signed MEM_VALUE_T mem_get_s##end##sz(const void *vmem) {\
+    const MAU_T *mem = (const MAU_T*)vmem;\
+    signed MEM_VALUE_T val = mem_get_##end##sz(mem);\
+    return (val << (MEM_VALUE_T_SZ_BITS - sz)) >> (MEM_VALUE_T_SZ_BITS - sz);\
+  }
+
+#undef  mem_get_sbe16
+#define mem_get_sbe16 mem_ops_wrap_symbol(mem_get_sbe16)
+mem_get_s_generic(be, 16)
+
+#undef  mem_get_sbe24
+#define mem_get_sbe24 mem_ops_wrap_symbol(mem_get_sbe24)
+mem_get_s_generic(be, 24)
+
+#undef  mem_get_sbe32
+#define mem_get_sbe32 mem_ops_wrap_symbol(mem_get_sbe32)
+mem_get_s_generic(be, 32)
+
+#undef  mem_get_sle16
+#define mem_get_sle16 mem_ops_wrap_symbol(mem_get_sle16)
+mem_get_s_generic(le, 16)
+
+#undef  mem_get_sle24
+#define mem_get_sle24 mem_ops_wrap_symbol(mem_get_sle24)
+mem_get_s_generic(le, 24)
+
+#undef  mem_get_sle32
+#define mem_get_sle32 mem_ops_wrap_symbol(mem_get_sle32)
+mem_get_s_generic(le, 32)
+
+#undef  mem_put_be16
+#define mem_put_be16 mem_ops_wrap_symbol(mem_put_be16)
+static VPX_INLINE void mem_put_be16(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >> 8) & 0xff);
+  mem[1] = (MAU_T)((val >> 0) & 0xff);
+}
+
+#undef  mem_put_be24
+#define mem_put_be24 mem_ops_wrap_symbol(mem_put_be24)
+static VPX_INLINE void mem_put_be24(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >> 16) & 0xff);
+  mem[1] = (MAU_T)((val >>  8) & 0xff);
+  mem[2] = (MAU_T)((val >>  0) & 0xff);
+}
+
+#undef  mem_put_be32
+#define mem_put_be32 mem_ops_wrap_symbol(mem_put_be32)
+static VPX_INLINE void mem_put_be32(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >> 24) & 0xff);
+  mem[1] = (MAU_T)((val >> 16) & 0xff);
+  mem[2] = (MAU_T)((val >>  8) & 0xff);
+  mem[3] = (MAU_T)((val >>  0) & 0xff);
+}
+
+#undef  mem_put_le16
+#define mem_put_le16 mem_ops_wrap_symbol(mem_put_le16)
+static VPX_INLINE void mem_put_le16(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >> 0) & 0xff);
+  mem[1] = (MAU_T)((val >> 8) & 0xff);
+}
+
+#undef  mem_put_le24
+#define mem_put_le24 mem_ops_wrap_symbol(mem_put_le24)
+static VPX_INLINE void mem_put_le24(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >>  0) & 0xff);
+  mem[1] = (MAU_T)((val >>  8) & 0xff);
+  mem[2] = (MAU_T)((val >> 16) & 0xff);
+}
+
+#undef  mem_put_le32
+#define mem_put_le32 mem_ops_wrap_symbol(mem_put_le32)
+static VPX_INLINE void mem_put_le32(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >>  0) & 0xff);
+  mem[1] = (MAU_T)((val >>  8) & 0xff);
+  mem[2] = (MAU_T)((val >> 16) & 0xff);
+  mem[3] = (MAU_T)((val >> 24) & 0xff);
+}
+
+#endif  // VPX_PORTS_MEM_OPS_H_
diff --git a/libvpx/libvpx/vpx_ports/mem_ops_aligned.h b/libvpx/libvpx/vpx_ports/mem_ops_aligned.h
new file mode 100644
index 0000000..46f6173
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/mem_ops_aligned.h
@@ -0,0 +1,169 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_PORTS_MEM_OPS_ALIGNED_H_
+#define VPX_PORTS_MEM_OPS_ALIGNED_H_
+
+#include "vpx/vpx_integer.h"
+
+/* \file
+ * \brief Provides portable memory access primitives for operating on aligned
+ *        data
+ *
+ * This file is split from mem_ops.h for easier maintenance. See mem_ops.h
+ * for a more detailed description of these primitives.
+ */
+#ifndef INCLUDED_BY_MEM_OPS_H
+#error Include mem_ops.h, not mem_ops_aligned.h directly.
+#endif
+
+/* Architectures that provide instructions for doing this byte swapping
+ * could redefine these macros.
+ */
+#define swap_endian_16(val,raw) do {\
+    val = (uint16_t)(((raw>>8) & 0x00ff) \
+          | ((raw<<8) & 0xff00));\
+  } while(0)
+#define swap_endian_32(val,raw) do {\
+    val = ((raw>>24) & 0x000000ff) \
+          | ((raw>>8)  & 0x0000ff00) \
+          | ((raw<<8)  & 0x00ff0000) \
+          | ((raw<<24) & 0xff000000); \
+  } while(0)
+#define swap_endian_16_se(val,raw) do {\
+    swap_endian_16(val,raw);\
+    val = ((val << 16) >> 16);\
+  } while(0)
+#define swap_endian_32_se(val,raw) swap_endian_32(val,raw)
+
+#define mem_get_ne_aligned_generic(end,sz) \
+  static VPX_INLINE unsigned MEM_VALUE_T \
+    mem_get_##end##sz##_aligned(const void *vmem) {\
+    const uint##sz##_t *mem = (const uint##sz##_t *)vmem;\
+    return *mem;\
+  }
+
+#define mem_get_sne_aligned_generic(end,sz) \
+  static VPX_INLINE signed MEM_VALUE_T \
+    mem_get_s##end##sz##_aligned(const void *vmem) {\
+    const int##sz##_t *mem = (const int##sz##_t *)vmem;\
+    return *mem;\
+  }
+
+#define mem_get_se_aligned_generic(end,sz) \
+  static VPX_INLINE unsigned MEM_VALUE_T \
+    mem_get_##end##sz##_aligned(const void *vmem) {\
+    const uint##sz##_t *mem = (const uint##sz##_t *)vmem;\
+    unsigned MEM_VALUE_T val, raw = *mem;\
+    swap_endian_##sz(val,raw);\
+    return val;\
+  }
+
+#define mem_get_sse_aligned_generic(end,sz) \
+  static VPX_INLINE signed MEM_VALUE_T \
+    mem_get_s##end##sz##_aligned(const void *vmem) {\
+    const int##sz##_t *mem = (const int##sz##_t *)vmem;\
+    unsigned MEM_VALUE_T val, raw = *mem;\
+    swap_endian_##sz##_se(val,raw);\
+    return val;\
+  }
+
+#define mem_put_ne_aligned_generic(end,sz) \
+  static VPX_INLINE void \
+    mem_put_##end##sz##_aligned(void *vmem, MEM_VALUE_T val) {\
+    uint##sz##_t *mem = (uint##sz##_t *)vmem;\
+    *mem = (uint##sz##_t)val;\
+  }
+
+#define mem_put_se_aligned_generic(end,sz) \
+  static VPX_INLINE void \
+    mem_put_##end##sz##_aligned(void *vmem, MEM_VALUE_T val) {\
+    uint##sz##_t *mem = (uint##sz##_t *)vmem, raw;\
+    swap_endian_##sz(raw,val);\
+    *mem = (uint##sz##_t)raw;\
+  }
+
+#include "vpx_config.h"
+#if CONFIG_BIG_ENDIAN
+#define mem_get_be_aligned_generic(sz)  mem_get_ne_aligned_generic(be,sz)
+#define mem_get_sbe_aligned_generic(sz) mem_get_sne_aligned_generic(be,sz)
+#define mem_get_le_aligned_generic(sz)  mem_get_se_aligned_generic(le,sz)
+#define mem_get_sle_aligned_generic(sz) mem_get_sse_aligned_generic(le,sz)
+#define mem_put_be_aligned_generic(sz)  mem_put_ne_aligned_generic(be,sz)
+#define mem_put_le_aligned_generic(sz)  mem_put_se_aligned_generic(le,sz)
+#else
+#define mem_get_be_aligned_generic(sz)  mem_get_se_aligned_generic(be,sz)
+#define mem_get_sbe_aligned_generic(sz) mem_get_sse_aligned_generic(be,sz)
+#define mem_get_le_aligned_generic(sz)  mem_get_ne_aligned_generic(le,sz)
+#define mem_get_sle_aligned_generic(sz) mem_get_sne_aligned_generic(le,sz)
+#define mem_put_be_aligned_generic(sz)  mem_put_se_aligned_generic(be,sz)
+#define mem_put_le_aligned_generic(sz)  mem_put_ne_aligned_generic(le,sz)
+#endif
+
+#undef  mem_get_be16_aligned
+#define mem_get_be16_aligned mem_ops_wrap_symbol(mem_get_be16_aligned)
+mem_get_be_aligned_generic(16)
+
+#undef  mem_get_be32_aligned
+#define mem_get_be32_aligned mem_ops_wrap_symbol(mem_get_be32_aligned)
+mem_get_be_aligned_generic(32)
+
+#undef  mem_get_le16_aligned
+#define mem_get_le16_aligned mem_ops_wrap_symbol(mem_get_le16_aligned)
+mem_get_le_aligned_generic(16)
+
+#undef  mem_get_le32_aligned
+#define mem_get_le32_aligned mem_ops_wrap_symbol(mem_get_le32_aligned)
+mem_get_le_aligned_generic(32)
+
+#undef  mem_get_sbe16_aligned
+#define mem_get_sbe16_aligned mem_ops_wrap_symbol(mem_get_sbe16_aligned)
+mem_get_sbe_aligned_generic(16)
+
+#undef  mem_get_sbe32_aligned
+#define mem_get_sbe32_aligned mem_ops_wrap_symbol(mem_get_sbe32_aligned)
+mem_get_sbe_aligned_generic(32)
+
+#undef  mem_get_sle16_aligned
+#define mem_get_sle16_aligned mem_ops_wrap_symbol(mem_get_sle16_aligned)
+mem_get_sle_aligned_generic(16)
+
+#undef  mem_get_sle32_aligned
+#define mem_get_sle32_aligned mem_ops_wrap_symbol(mem_get_sle32_aligned)
+mem_get_sle_aligned_generic(32)
+
+#undef  mem_put_be16_aligned
+#define mem_put_be16_aligned mem_ops_wrap_symbol(mem_put_be16_aligned)
+mem_put_be_aligned_generic(16)
+
+#undef  mem_put_be32_aligned
+#define mem_put_be32_aligned mem_ops_wrap_symbol(mem_put_be32_aligned)
+mem_put_be_aligned_generic(32)
+
+#undef  mem_put_le16_aligned
+#define mem_put_le16_aligned mem_ops_wrap_symbol(mem_put_le16_aligned)
+mem_put_le_aligned_generic(16)
+
+#undef  mem_put_le32_aligned
+#define mem_put_le32_aligned mem_ops_wrap_symbol(mem_put_le32_aligned)
+mem_put_le_aligned_generic(32)
+
+#undef mem_get_ne_aligned_generic
+#undef mem_get_se_aligned_generic
+#undef mem_get_sne_aligned_generic
+#undef mem_get_sse_aligned_generic
+#undef mem_put_ne_aligned_generic
+#undef mem_put_se_aligned_generic
+#undef swap_endian_16
+#undef swap_endian_32
+#undef swap_endian_16_se
+#undef swap_endian_32_se
+
+#endif  // VPX_PORTS_MEM_OPS_ALIGNED_H_
diff --git a/libvpx/libvpx/vpx_ports/msvc.h b/libvpx/libvpx/vpx_ports/msvc.h
new file mode 100644
index 0000000..cab7740
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/msvc.h
@@ -0,0 +1,32 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_PORTS_MSVC_H_
+#define VPX_PORTS_MSVC_H_
+#ifdef _MSC_VER
+
+#include "./vpx_config.h"
+
+# if _MSC_VER < 1900  // VS2015 provides snprintf
+#  define snprintf _snprintf
+# endif  // _MSC_VER < 1900
+
+#if _MSC_VER < 1800  // VS2013 provides round
+#include <math.h>
+static INLINE double round(double x) {
+  if (x < 0)
+    return ceil(x - 0.5);
+  else
+    return floor(x + 0.5);
+}
+#endif  // _MSC_VER < 1800
+
+#endif  // _MSC_VER
+#endif  // VPX_PORTS_MSVC_H_
diff --git a/libvpx/libvpx/vpx_ports/system_state.h b/libvpx/libvpx/vpx_ports/system_state.h
new file mode 100644
index 0000000..086c646
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/system_state.h
@@ -0,0 +1,22 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_PORTS_SYSTEM_STATE_H_
+#define VPX_PORTS_SYSTEM_STATE_H_
+
+#include "./vpx_config.h"
+
+#if ARCH_X86 || ARCH_X86_64
+void vpx_reset_mmx_state(void);
+#define vpx_clear_system_state() vpx_reset_mmx_state()
+#else
+#define vpx_clear_system_state()
+#endif  // ARCH_X86 || ARCH_X86_64
+#endif  // VPX_PORTS_SYSTEM_STATE_H_
diff --git a/libvpx/libvpx/vpx_ports/vpx_once.h b/libvpx/libvpx/vpx_ports/vpx_once.h
new file mode 100644
index 0000000..da04db4
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/vpx_once.h
@@ -0,0 +1,150 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_PORTS_VPX_ONCE_H_
+#define VPX_PORTS_VPX_ONCE_H_
+
+#include "vpx_config.h"
+
+/* Implement a function wrapper to guarantee initialization
+ * thread-safety for library singletons.
+ *
+ * NOTE: These functions use static locks, and can only be
+ * used with one common argument per compilation unit. So
+ *
+ * file1.c:
+ *   vpx_once(foo);
+ *   ...
+ *   vpx_once(foo);
+ *
+ *   file2.c:
+ *     vpx_once(bar);
+ *
+ * will ensure foo() and bar() are each called only once, but in
+ *
+ * file1.c:
+ *   vpx_once(foo);
+ *   vpx_once(bar):
+ *
+ * bar() will never be called because the lock is used up
+ * by the call to foo().
+ */
+
+#if CONFIG_MULTITHREAD && defined(_WIN32)
+#include <windows.h>
+#include <stdlib.h>
+/* Declare a per-compilation-unit state variable to track the progress
+ * of calling func() only once. This must be at global scope because
+ * local initializers are not thread-safe in MSVC prior to Visual
+ * Studio 2015.
+ *
+ * As a static, once_state will be zero-initialized as program start.
+ */
+static LONG once_state;
+static void once(void (*func)(void))
+{
+    /* Try to advance once_state from its initial value of 0 to 1.
+     * Only one thread can succeed in doing so.
+     */
+    if (InterlockedCompareExchange(&once_state, 1, 0) == 0) {
+        /* We're the winning thread, having set once_state to 1.
+         * Call our function. */
+        func();
+        /* Now advance once_state to 2, unblocking any other threads. */
+        InterlockedIncrement(&once_state);
+        return;
+    }
+
+    /* We weren't the winning thread, but we want to block on
+     * the state variable so we don't return before func()
+     * has finished executing elsewhere.
+     *
+     * Try to advance once_state from 2 to 2, which is only possible
+     * after the winning thead advances it from 1 to 2.
+     */
+    while (InterlockedCompareExchange(&once_state, 2, 2) != 2) {
+        /* State isn't yet 2. Try again.
+         *
+         * We are used for singleton initialization functions,
+         * which should complete quickly. Contention will likewise
+         * be rare, so it's worthwhile to use a simple but cpu-
+         * intensive busy-wait instead of successive backoff,
+         * waiting on a kernel object, or another heavier-weight scheme.
+         *
+         * We can at least yield our timeslice.
+         */
+        Sleep(0);
+    }
+
+    /* We've seen once_state advance to 2, so we know func()
+     * has been called. And we've left once_state as we found it,
+     * so other threads will have the same experience.
+     *
+     * It's safe to return now.
+     */
+    return;
+}
+
+
+#elif CONFIG_MULTITHREAD && defined(__OS2__)
+#define INCL_DOS
+#include <os2.h>
+static void once(void (*func)(void))
+{
+    static int done;
+
+    /* If the initialization is complete, return early. */
+    if(done)
+        return;
+
+    /* Causes all other threads in the process to block themselves
+     * and give up their time slice.
+     */
+    DosEnterCritSec();
+
+    if (!done)
+    {
+        func();
+        done = 1;
+    }
+
+    /* Restores normal thread dispatching for the current process. */
+    DosExitCritSec();
+}
+
+
+#elif CONFIG_MULTITHREAD && HAVE_PTHREAD_H
+#include <pthread.h>
+static void once(void (*func)(void))
+{
+    static pthread_once_t lock = PTHREAD_ONCE_INIT;
+    pthread_once(&lock, func);
+}
+
+
+#else
+/* No-op version that performs no synchronization. *_rtcd() is idempotent,
+ * so as long as your platform provides atomic loads/stores of pointers
+ * no synchronization is strictly necessary.
+ */
+
+static void once(void (*func)(void))
+{
+    static int done;
+
+    if(!done)
+    {
+        func();
+        done = 1;
+    }
+}
+#endif
+
+#endif  // VPX_PORTS_VPX_ONCE_H_
diff --git a/libvpx/libvpx/vpx_ports/vpx_ports.mk b/libvpx/libvpx/vpx_ports/vpx_ports.mk
new file mode 100644
index 0000000..36b1493
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/vpx_ports.mk
@@ -0,0 +1,27 @@
+##
+##  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+
+PORTS_SRCS-yes += vpx_ports.mk
+
+PORTS_SRCS-yes += bitops.h
+PORTS_SRCS-yes += mem.h
+PORTS_SRCS-yes += msvc.h
+PORTS_SRCS-yes += system_state.h
+PORTS_SRCS-yes += vpx_timer.h
+
+ifeq ($(ARCH_X86)$(ARCH_X86_64),yes)
+PORTS_SRCS-yes += emms.asm
+PORTS_SRCS-yes += x86.h
+PORTS_SRCS-yes += x86_abi_support.asm
+endif
+
+PORTS_SRCS-$(ARCH_ARM) += arm_cpudetect.c
+PORTS_SRCS-$(ARCH_ARM) += arm.h
diff --git a/libvpx/libvpx/vpx_ports/vpx_timer.h b/libvpx/libvpx/vpx_ports/vpx_timer.h
new file mode 100644
index 0000000..dd98e29
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/vpx_timer.h
@@ -0,0 +1,120 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VPX_PORTS_VPX_TIMER_H_
+#define VPX_PORTS_VPX_TIMER_H_
+
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+#if CONFIG_OS_SUPPORT
+
+#if defined(_WIN32)
+/*
+ * Win32 specific includes
+ */
+#ifndef WIN32_LEAN_AND_MEAN
+#define WIN32_LEAN_AND_MEAN
+#endif
+#include <windows.h>
+#else
+/*
+ * POSIX specific includes
+ */
+#include <sys/time.h>
+
+/* timersub is not provided by msys at this time. */
+#ifndef timersub
+#define timersub(a, b, result) \
+  do { \
+    (result)->tv_sec = (a)->tv_sec - (b)->tv_sec; \
+    (result)->tv_usec = (a)->tv_usec - (b)->tv_usec; \
+    if ((result)->tv_usec < 0) { \
+      --(result)->tv_sec; \
+      (result)->tv_usec += 1000000; \
+    } \
+  } while (0)
+#endif
+#endif
+
+
+struct vpx_usec_timer {
+#if defined(_WIN32)
+  LARGE_INTEGER  begin, end;
+#else
+  struct timeval begin, end;
+#endif
+};
+
+
+static INLINE void
+vpx_usec_timer_start(struct vpx_usec_timer *t) {
+#if defined(_WIN32)
+  QueryPerformanceCounter(&t->begin);
+#else
+  gettimeofday(&t->begin, NULL);
+#endif
+}
+
+
+static INLINE void
+vpx_usec_timer_mark(struct vpx_usec_timer *t) {
+#if defined(_WIN32)
+  QueryPerformanceCounter(&t->end);
+#else
+  gettimeofday(&t->end, NULL);
+#endif
+}
+
+
+static INLINE int64_t
+vpx_usec_timer_elapsed(struct vpx_usec_timer *t) {
+#if defined(_WIN32)
+  LARGE_INTEGER freq, diff;
+
+  diff.QuadPart = t->end.QuadPart - t->begin.QuadPart;
+
+  QueryPerformanceFrequency(&freq);
+  return diff.QuadPart * 1000000 / freq.QuadPart;
+#else
+  struct timeval diff;
+
+  timersub(&t->end, &t->begin, &diff);
+  return diff.tv_sec * 1000000 + diff.tv_usec;
+#endif
+}
+
+#else /* CONFIG_OS_SUPPORT = 0*/
+
+/* Empty timer functions if CONFIG_OS_SUPPORT = 0 */
+#ifndef timersub
+#define timersub(a, b, result)
+#endif
+
+struct vpx_usec_timer {
+  void *dummy;
+};
+
+static INLINE void
+vpx_usec_timer_start(struct vpx_usec_timer *t) { }
+
+static INLINE void
+vpx_usec_timer_mark(struct vpx_usec_timer *t) { }
+
+static INLINE int
+vpx_usec_timer_elapsed(struct vpx_usec_timer *t) {
+  return 0;
+}
+
+#endif /* CONFIG_OS_SUPPORT */
+
+#endif  // VPX_PORTS_VPX_TIMER_H_
diff --git a/libvpx/libvpx/vpx_ports/x86.h b/libvpx/libvpx/vpx_ports/x86.h
new file mode 100644
index 0000000..bae25ac
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/x86.h
@@ -0,0 +1,330 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VPX_PORTS_X86_H_
+#define VPX_PORTS_X86_H_
+#include <stdlib.h>
+
+#if defined(_MSC_VER)
+#include <intrin.h>  /* For __cpuidex, __rdtsc */
+#endif
+
+#include "vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef enum {
+  VPX_CPU_UNKNOWN = -1,
+  VPX_CPU_AMD,
+  VPX_CPU_AMD_OLD,
+  VPX_CPU_CENTAUR,
+  VPX_CPU_CYRIX,
+  VPX_CPU_INTEL,
+  VPX_CPU_NEXGEN,
+  VPX_CPU_NSC,
+  VPX_CPU_RISE,
+  VPX_CPU_SIS,
+  VPX_CPU_TRANSMETA,
+  VPX_CPU_TRANSMETA_OLD,
+  VPX_CPU_UMC,
+  VPX_CPU_VIA,
+
+  VPX_CPU_LAST
+}  vpx_cpu_t;
+
+#if defined(__GNUC__) && __GNUC__ || defined(__ANDROID__)
+#if ARCH_X86_64
+#define cpuid(func, func2, ax, bx, cx, dx)\
+  __asm__ __volatile__ (\
+                        "cpuid           \n\t" \
+                        : "=a" (ax), "=b" (bx), "=c" (cx), "=d" (dx) \
+                        : "a" (func), "c" (func2));
+#else
+#define cpuid(func, func2, ax, bx, cx, dx)\
+  __asm__ __volatile__ (\
+                        "mov %%ebx, %%edi   \n\t" \
+                        "cpuid              \n\t" \
+                        "xchg %%edi, %%ebx  \n\t" \
+                        : "=a" (ax), "=D" (bx), "=c" (cx), "=d" (dx) \
+                        : "a" (func), "c" (func2));
+#endif
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC) /* end __GNUC__ or __ANDROID__*/
+#if ARCH_X86_64
+#define cpuid(func, func2, ax, bx, cx, dx)\
+  asm volatile (\
+                "xchg %rsi, %rbx \n\t" \
+                "cpuid           \n\t" \
+                "movl %ebx, %edi \n\t" \
+                "xchg %rsi, %rbx \n\t" \
+                : "=a" (ax), "=D" (bx), "=c" (cx), "=d" (dx) \
+                : "a" (func), "c" (func2));
+#else
+#define cpuid(func, func2, ax, bx, cx, dx)\
+  asm volatile (\
+                "pushl %ebx       \n\t" \
+                "cpuid            \n\t" \
+                "movl %ebx, %edi  \n\t" \
+                "popl %ebx        \n\t" \
+                : "=a" (ax), "=D" (bx), "=c" (cx), "=d" (dx) \
+                : "a" (func), "c" (func2));
+#endif
+#else /* end __SUNPRO__ */
+#if ARCH_X86_64
+#if defined(_MSC_VER) && _MSC_VER > 1500
+#define cpuid(func, func2, a, b, c, d) do {\
+    int regs[4];\
+    __cpuidex(regs, func, func2); \
+    a = regs[0];  b = regs[1];  c = regs[2];  d = regs[3];\
+  } while(0)
+#else
+#define cpuid(func, func2, a, b, c, d) do {\
+    int regs[4];\
+    __cpuid(regs, func); \
+    a = regs[0];  b = regs[1];  c = regs[2];  d = regs[3];\
+  } while (0)
+#endif
+#else
+#define cpuid(func, func2, a, b, c, d)\
+  __asm mov eax, func\
+  __asm mov ecx, func2\
+  __asm cpuid\
+  __asm mov a, eax\
+  __asm mov b, ebx\
+  __asm mov c, ecx\
+  __asm mov d, edx
+#endif
+#endif /* end others */
+
+// NaCl has no support for xgetbv or the raw opcode.
+#if !defined(__native_client__) && (defined(__i386__) || defined(__x86_64__))
+static INLINE uint64_t xgetbv(void) {
+  const uint32_t ecx = 0;
+  uint32_t eax, edx;
+  // Use the raw opcode for xgetbv for compatibility with older toolchains.
+  __asm__ volatile (
+    ".byte 0x0f, 0x01, 0xd0\n"
+    : "=a"(eax), "=d"(edx) : "c" (ecx));
+  return ((uint64_t)edx << 32) | eax;
+}
+#elif (defined(_M_X64) || defined(_M_IX86)) && \
+      defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 160040219  // >= VS2010 SP1
+#include <immintrin.h>
+#define xgetbv() _xgetbv(0)
+#elif defined(_MSC_VER) && defined(_M_IX86)
+static INLINE uint64_t xgetbv(void) {
+  uint32_t eax_, edx_;
+  __asm {
+    xor ecx, ecx  // ecx = 0
+    // Use the raw opcode for xgetbv for compatibility with older toolchains.
+    __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0
+    mov eax_, eax
+    mov edx_, edx
+  }
+  return ((uint64_t)edx_ << 32) | eax_;
+}
+#else
+#define xgetbv() 0U  // no AVX for older x64 or unrecognized toolchains.
+#endif
+
+#if defined(_MSC_VER) && _MSC_VER >= 1700
+#include <windows.h>
+#if WINAPI_FAMILY_PARTITION(WINAPI_FAMILY_APP)
+#define getenv(x) NULL
+#endif
+#endif
+
+#define HAS_MMX     0x01
+#define HAS_SSE     0x02
+#define HAS_SSE2    0x04
+#define HAS_SSE3    0x08
+#define HAS_SSSE3   0x10
+#define HAS_SSE4_1  0x20
+#define HAS_AVX     0x40
+#define HAS_AVX2    0x80
+#ifndef BIT
+#define BIT(n) (1<<n)
+#endif
+
+static INLINE int
+x86_simd_caps(void) {
+  unsigned int flags = 0;
+  unsigned int mask = ~0;
+  unsigned int max_cpuid_val, reg_eax, reg_ebx, reg_ecx, reg_edx;
+  char *env;
+  (void)reg_ebx;
+
+  /* See if the CPU capabilities are being overridden by the environment */
+  env = getenv("VPX_SIMD_CAPS");
+
+  if (env && *env)
+    return (int)strtol(env, NULL, 0);
+
+  env = getenv("VPX_SIMD_CAPS_MASK");
+
+  if (env && *env)
+    mask = (unsigned int)strtoul(env, NULL, 0);
+
+  /* Ensure that the CPUID instruction supports extended features */
+  cpuid(0, 0, max_cpuid_val, reg_ebx, reg_ecx, reg_edx);
+
+  if (max_cpuid_val < 1)
+    return 0;
+
+  /* Get the standard feature flags */
+  cpuid(1, 0, reg_eax, reg_ebx, reg_ecx, reg_edx);
+
+  if (reg_edx & BIT(23)) flags |= HAS_MMX;
+
+  if (reg_edx & BIT(25)) flags |= HAS_SSE; /* aka xmm */
+
+  if (reg_edx & BIT(26)) flags |= HAS_SSE2; /* aka wmt */
+
+  if (reg_ecx & BIT(0)) flags |= HAS_SSE3;
+
+  if (reg_ecx & BIT(9)) flags |= HAS_SSSE3;
+
+  if (reg_ecx & BIT(19)) flags |= HAS_SSE4_1;
+
+  // bits 27 (OSXSAVE) & 28 (256-bit AVX)
+  if ((reg_ecx & (BIT(27) | BIT(28))) == (BIT(27) | BIT(28))) {
+    if ((xgetbv() & 0x6) == 0x6) {
+      flags |= HAS_AVX;
+
+      if (max_cpuid_val >= 7) {
+        /* Get the leaf 7 feature flags. Needed to check for AVX2 support */
+        cpuid(7, 0, reg_eax, reg_ebx, reg_ecx, reg_edx);
+
+        if (reg_ebx & BIT(5)) flags |= HAS_AVX2;
+      }
+    }
+  }
+
+  return flags & mask;
+}
+
+// Note:
+//  32-bit CPU cycle counter is light-weighted for most function performance
+//  measurement. For large function (CPU time > a couple of seconds), 64-bit
+//  counter should be used.
+// 32-bit CPU cycle counter
+static INLINE unsigned int
+x86_readtsc(void) {
+#if defined(__GNUC__) && __GNUC__
+  unsigned int tsc;
+  __asm__ __volatile__("rdtsc\n\t":"=a"(tsc):);
+  return tsc;
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+  unsigned int tsc;
+  asm volatile("rdtsc\n\t":"=a"(tsc):);
+  return tsc;
+#else
+#if ARCH_X86_64
+  return (unsigned int)__rdtsc();
+#else
+  __asm  rdtsc;
+#endif
+#endif
+}
+// 64-bit CPU cycle counter
+static INLINE uint64_t
+x86_readtsc64(void) {
+#if defined(__GNUC__) && __GNUC__
+  uint32_t hi, lo;
+  __asm__ __volatile__("rdtsc" : "=a"(lo), "=d"(hi));
+  return ((uint64_t)hi << 32) | lo;
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+  uint_t hi, lo;
+  asm volatile("rdtsc\n\t" : "=a"(lo), "=d"(hi));
+  return ((uint64_t)hi << 32) | lo;
+#else
+#if ARCH_X86_64
+  return (uint64_t)__rdtsc();
+#else
+  __asm  rdtsc;
+#endif
+#endif
+}
+
+#if defined(__GNUC__) && __GNUC__
+#define x86_pause_hint()\
+  __asm__ __volatile__ ("pause \n\t")
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+#define x86_pause_hint()\
+  asm volatile ("pause \n\t")
+#else
+#if ARCH_X86_64
+#define x86_pause_hint()\
+  _mm_pause();
+#else
+#define x86_pause_hint()\
+  __asm pause
+#endif
+#endif
+
+#if defined(__GNUC__) && __GNUC__
+static void
+x87_set_control_word(unsigned short mode) {
+  __asm__ __volatile__("fldcw %0" : : "m"(*&mode));
+}
+static unsigned short
+x87_get_control_word(void) {
+  unsigned short mode;
+  __asm__ __volatile__("fstcw %0\n\t":"=m"(*&mode):);
+    return mode;
+}
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+static void
+x87_set_control_word(unsigned short mode) {
+  asm volatile("fldcw %0" : : "m"(*&mode));
+}
+static unsigned short
+x87_get_control_word(void) {
+  unsigned short mode;
+  asm volatile("fstcw %0\n\t":"=m"(*&mode):);
+  return mode;
+}
+#elif ARCH_X86_64
+/* No fldcw intrinsics on Windows x64, punt to external asm */
+extern void           vpx_winx64_fldcw(unsigned short mode);
+extern unsigned short vpx_winx64_fstcw(void);
+#define x87_set_control_word vpx_winx64_fldcw
+#define x87_get_control_word vpx_winx64_fstcw
+#else
+static void
+x87_set_control_word(unsigned short mode) {
+  __asm { fldcw mode }
+}
+static unsigned short
+x87_get_control_word(void) {
+  unsigned short mode;
+  __asm { fstcw mode }
+  return mode;
+}
+#endif
+
+static INLINE unsigned int
+x87_set_double_precision(void) {
+  unsigned int mode = x87_get_control_word();
+  x87_set_control_word((mode&~0x300) | 0x200);
+  return mode;
+}
+
+
+extern void vpx_reset_mmx_state(void);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPX_PORTS_X86_H_
diff --git a/libvpx/libvpx/vpx_ports/x86_abi_support.asm b/libvpx/libvpx/vpx_ports/x86_abi_support.asm
new file mode 100644
index 0000000..708fa10
--- /dev/null
+++ b/libvpx/libvpx/vpx_ports/x86_abi_support.asm
@@ -0,0 +1,404 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_config.asm"
+
+; 32/64 bit compatibility macros
+;
+; In general, we make the source use 64 bit syntax, then twiddle with it using
+; the preprocessor to get the 32 bit syntax on 32 bit platforms.
+;
+%ifidn __OUTPUT_FORMAT__,elf32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,macho32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,win32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,aout
+%define ABI_IS_32BIT 1
+%else
+%define ABI_IS_32BIT 0
+%endif
+
+%if ABI_IS_32BIT
+%define rax eax
+%define rbx ebx
+%define rcx ecx
+%define rdx edx
+%define rsi esi
+%define rdi edi
+%define rsp esp
+%define rbp ebp
+%define movsxd mov
+%macro movq 2
+  %ifidn %1,eax
+    movd %1,%2
+  %elifidn %2,eax
+    movd %1,%2
+  %elifidn %1,ebx
+    movd %1,%2
+  %elifidn %2,ebx
+    movd %1,%2
+  %elifidn %1,ecx
+    movd %1,%2
+  %elifidn %2,ecx
+    movd %1,%2
+  %elifidn %1,edx
+    movd %1,%2
+  %elifidn %2,edx
+    movd %1,%2
+  %elifidn %1,esi
+    movd %1,%2
+  %elifidn %2,esi
+    movd %1,%2
+  %elifidn %1,edi
+    movd %1,%2
+  %elifidn %2,edi
+    movd %1,%2
+  %elifidn %1,esp
+    movd %1,%2
+  %elifidn %2,esp
+    movd %1,%2
+  %elifidn %1,ebp
+    movd %1,%2
+  %elifidn %2,ebp
+    movd %1,%2
+  %else
+    movq %1,%2
+  %endif
+%endmacro
+%endif
+
+
+; LIBVPX_YASM_WIN64
+; Set LIBVPX_YASM_WIN64 if output is Windows 64bit so the code will work if x64
+; or win64 is defined on the Yasm command line.
+%ifidn __OUTPUT_FORMAT__,win64
+%define LIBVPX_YASM_WIN64 1
+%elifidn __OUTPUT_FORMAT__,x64
+%define LIBVPX_YASM_WIN64 1
+%else
+%define LIBVPX_YASM_WIN64 0
+%endif
+
+; sym()
+; Return the proper symbol name for the target ABI.
+;
+; Certain ABIs, notably MS COFF and Darwin MACH-O, require that symbols
+; with C linkage be prefixed with an underscore.
+;
+%ifidn   __OUTPUT_FORMAT__,elf32
+%define sym(x) x
+%elifidn __OUTPUT_FORMAT__,elf64
+%define sym(x) x
+%elifidn __OUTPUT_FORMAT__,elfx32
+%define sym(x) x
+%elif LIBVPX_YASM_WIN64
+%define sym(x) x
+%else
+%define sym(x) _ %+ x
+%endif
+
+;  PRIVATE
+;  Macro for the attribute to hide a global symbol for the target ABI.
+;  This is only active if CHROMIUM is defined.
+;
+;  Chromium doesn't like exported global symbols due to symbol clashing with
+;  plugins among other things.
+;
+;  Requires Chromium's patched copy of yasm:
+;    http://src.chromium.org/viewvc/chrome?view=rev&revision=73761
+;    http://www.tortall.net/projects/yasm/ticket/236
+;
+%ifdef CHROMIUM
+  %ifidn   __OUTPUT_FORMAT__,elf32
+    %define PRIVATE :hidden
+  %elifidn __OUTPUT_FORMAT__,elf64
+    %define PRIVATE :hidden
+  %elifidn __OUTPUT_FORMAT__,elfx32
+    %define PRIVATE :hidden
+  %elif LIBVPX_YASM_WIN64
+    %define PRIVATE
+  %else
+    %define PRIVATE :private_extern
+  %endif
+%else
+  %define PRIVATE
+%endif
+
+; arg()
+; Return the address specification of the given argument
+;
+%if ABI_IS_32BIT
+  %define arg(x) [ebp+8+4*x]
+%else
+  ; 64 bit ABI passes arguments in registers. This is a workaround to get up
+  ; and running quickly. Relies on SHADOW_ARGS_TO_STACK
+  %if LIBVPX_YASM_WIN64
+    %define arg(x) [rbp+16+8*x]
+  %else
+    %define arg(x) [rbp-8-8*x]
+  %endif
+%endif
+
+; REG_SZ_BYTES, REG_SZ_BITS
+; Size of a register
+%if ABI_IS_32BIT
+%define REG_SZ_BYTES 4
+%define REG_SZ_BITS  32
+%else
+%define REG_SZ_BYTES 8
+%define REG_SZ_BITS  64
+%endif
+
+
+; ALIGN_STACK <alignment> <register>
+; This macro aligns the stack to the given alignment (in bytes). The stack
+; is left such that the previous value of the stack pointer is the first
+; argument on the stack (ie, the inverse of this macro is 'pop rsp.')
+; This macro uses one temporary register, which is not preserved, and thus
+; must be specified as an argument.
+%macro ALIGN_STACK 2
+    mov         %2, rsp
+    and         rsp, -%1
+    lea         rsp, [rsp - (%1 - REG_SZ_BYTES)]
+    push        %2
+%endmacro
+
+
+;
+; The Microsoft assembler tries to impose a certain amount of type safety in
+; its register usage. YASM doesn't recognize these directives, so we just
+; %define them away to maintain as much compatibility as possible with the
+; original inline assembler we're porting from.
+;
+%idefine PTR
+%idefine XMMWORD
+%idefine MMWORD
+
+; PIC macros
+;
+%if ABI_IS_32BIT
+  %if CONFIG_PIC=1
+  %ifidn __OUTPUT_FORMAT__,elf32
+    %define WRT_PLT wrt ..plt
+    %macro GET_GOT 1
+      extern _GLOBAL_OFFSET_TABLE_
+      push %1
+      call %%get_got
+      %%sub_offset:
+      jmp %%exitGG
+      %%get_got:
+      mov %1, [esp]
+      add %1, _GLOBAL_OFFSET_TABLE_ + $$ - %%sub_offset wrt ..gotpc
+      ret
+      %%exitGG:
+      %undef GLOBAL
+      %define GLOBAL(x) x + %1 wrt ..gotoff
+      %undef RESTORE_GOT
+      %define RESTORE_GOT pop %1
+    %endmacro
+  %elifidn __OUTPUT_FORMAT__,macho32
+    %macro GET_GOT 1
+      push %1
+      call %%get_got
+      %%get_got:
+      pop  %1
+      %undef GLOBAL
+      %define GLOBAL(x) x + %1 - %%get_got
+      %undef RESTORE_GOT
+      %define RESTORE_GOT pop %1
+    %endmacro
+  %endif
+  %endif
+
+  %ifdef CHROMIUM
+    %ifidn __OUTPUT_FORMAT__,macho32
+      %define HIDDEN_DATA(x) x:private_extern
+    %else
+      %define HIDDEN_DATA(x) x
+    %endif
+  %else
+    %define HIDDEN_DATA(x) x
+  %endif
+%else
+  %macro GET_GOT 1
+  %endmacro
+  %define GLOBAL(x) rel x
+  %ifidn __OUTPUT_FORMAT__,elf64
+    %define WRT_PLT wrt ..plt
+    %define HIDDEN_DATA(x) x:data hidden
+  %elifidn __OUTPUT_FORMAT__,elfx32
+    %define WRT_PLT wrt ..plt
+    %define HIDDEN_DATA(x) x:data hidden
+  %elifidn __OUTPUT_FORMAT__,macho64
+    %ifdef CHROMIUM
+      %define HIDDEN_DATA(x) x:private_extern
+    %else
+      %define HIDDEN_DATA(x) x
+    %endif
+  %else
+    %define HIDDEN_DATA(x) x
+  %endif
+%endif
+%ifnmacro GET_GOT
+    %macro GET_GOT 1
+    %endmacro
+    %define GLOBAL(x) x
+%endif
+%ifndef RESTORE_GOT
+%define RESTORE_GOT
+%endif
+%ifndef WRT_PLT
+%define WRT_PLT
+%endif
+
+%if ABI_IS_32BIT
+  %macro SHADOW_ARGS_TO_STACK 1
+  %endm
+  %define UNSHADOW_ARGS
+%else
+%if LIBVPX_YASM_WIN64
+  %macro SHADOW_ARGS_TO_STACK 1 ; argc
+    %if %1 > 0
+        mov arg(0),rcx
+    %endif
+    %if %1 > 1
+        mov arg(1),rdx
+    %endif
+    %if %1 > 2
+        mov arg(2),r8
+    %endif
+    %if %1 > 3
+        mov arg(3),r9
+    %endif
+  %endm
+%else
+  %macro SHADOW_ARGS_TO_STACK 1 ; argc
+    %if %1 > 0
+        push rdi
+    %endif
+    %if %1 > 1
+        push rsi
+    %endif
+    %if %1 > 2
+        push rdx
+    %endif
+    %if %1 > 3
+        push rcx
+    %endif
+    %if %1 > 4
+        push r8
+    %endif
+    %if %1 > 5
+        push r9
+    %endif
+    %if %1 > 6
+      %assign i %1-6
+      %assign off 16
+      %rep i
+        mov rax,[rbp+off]
+        push rax
+        %assign off off+8
+      %endrep
+    %endif
+  %endm
+%endif
+  %define UNSHADOW_ARGS mov rsp, rbp
+%endif
+
+; Win64 ABI requires that XMM6:XMM15 are callee saved
+; SAVE_XMM n, [u]
+; store registers 6-n on the stack
+; if u is specified, use unaligned movs.
+; Win64 ABI requires 16 byte stack alignment, but then pushes an 8 byte return
+; value. Typically we follow this up with 'push rbp' - re-aligning the stack -
+; but in some cases this is not done and unaligned movs must be used.
+%if LIBVPX_YASM_WIN64
+%macro SAVE_XMM 1-2 a
+  %if %1 < 6
+    %error Only xmm registers 6-15 must be preserved
+  %else
+    %assign last_xmm %1
+    %define movxmm movdq %+ %2
+    %assign xmm_stack_space ((last_xmm - 5) * 16)
+    sub rsp, xmm_stack_space
+    %assign i 6
+    %rep (last_xmm - 5)
+      movxmm [rsp + ((i - 6) * 16)], xmm %+ i
+      %assign i i+1
+    %endrep
+  %endif
+%endmacro
+%macro RESTORE_XMM 0
+  %ifndef last_xmm
+    %error RESTORE_XMM must be paired with SAVE_XMM n
+  %else
+    %assign i last_xmm
+    %rep (last_xmm - 5)
+      movxmm xmm %+ i, [rsp +((i - 6) * 16)]
+      %assign i i-1
+    %endrep
+    add rsp, xmm_stack_space
+    ; there are a couple functions which return from multiple places.
+    ; otherwise, we could uncomment these:
+    ; %undef last_xmm
+    ; %undef xmm_stack_space
+    ; %undef movxmm
+  %endif
+%endmacro
+%else
+%macro SAVE_XMM 1-2
+%endmacro
+%macro RESTORE_XMM 0
+%endmacro
+%endif
+
+; Name of the rodata section
+;
+; .rodata seems to be an elf-ism, as it doesn't work on OSX.
+;
+%ifidn __OUTPUT_FORMAT__,macho64
+%define SECTION_RODATA section .text
+%elifidn __OUTPUT_FORMAT__,macho32
+%macro SECTION_RODATA 0
+section .text
+%endmacro
+%elifidn __OUTPUT_FORMAT__,aout
+%define SECTION_RODATA section .data
+%else
+%define SECTION_RODATA section .rodata
+%endif
+
+
+; Tell GNU ld that we don't require an executable stack.
+%ifidn __OUTPUT_FORMAT__,elf32
+section .note.GNU-stack noalloc noexec nowrite progbits
+section .text
+%elifidn __OUTPUT_FORMAT__,elf64
+section .note.GNU-stack noalloc noexec nowrite progbits
+section .text
+%elifidn __OUTPUT_FORMAT__,elfx32
+section .note.GNU-stack noalloc noexec nowrite progbits
+section .text
+%endif
+
+; On Android platforms use lrand48 when building postproc routines. Prior to L
+; rand() was not available.
+%if CONFIG_POSTPROC=1 || CONFIG_VP9_POSTPROC=1
+%ifdef __ANDROID__
+extern sym(lrand48)
+%define LIBVPX_RAND lrand48
+%else
+extern sym(rand)
+%define LIBVPX_RAND rand
+%endif
+%endif ; CONFIG_POSTPROC || CONFIG_VP9_POSTPROC
diff --git a/libvpx/libvpx/vpx_scale/generic/gen_scalers.c b/libvpx/libvpx/vpx_scale/generic/gen_scalers.c
new file mode 100644
index 0000000..dab324e
--- /dev/null
+++ b/libvpx/libvpx/vpx_scale/generic/gen_scalers.c
@@ -0,0 +1,240 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_scale_rtcd.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vpx_mem/vpx_mem.h"
+/****************************************************************************
+*  Imports
+****************************************************************************/
+
+/****************************************************************************
+ *
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to source data.
+ *                  unsigned int source_width    : Stride of source.
+ *                  unsigned char *dest         : Pointer to destination data.
+ *                  unsigned int dest_width      : Stride of destination (NOT USED).
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Copies horizontal line of pixels from source to
+ *                  destination scaling up by 4 to 5.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ ****************************************************************************/
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source,
+                                     unsigned int source_width,
+                                     unsigned char *dest,
+                                     unsigned int dest_width) {
+  unsigned i;
+  unsigned int a, b, c, d, e;
+  unsigned char *des = dest;
+  const unsigned char *src = source;
+
+  (void) dest_width;
+
+  for (i = 0; i < source_width; i += 5) {
+    a = src[0];
+    b = src[1];
+    c = src[2];
+    d = src[3];
+    e = src[4];
+
+    des[0] = (unsigned char) a;
+    des[1] = (unsigned char)((b * 192 + c * 64 + 128) >> 8);
+    des[2] = (unsigned char)((c * 128 + d * 128 + 128) >> 8);
+    des[3] = (unsigned char)((d * 64 + e * 192 + 128) >> 8);
+
+    src += 5;
+    des += 4;
+  }
+}
+
+
+
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source,
+                                   unsigned int src_pitch,
+                                   unsigned char *dest,
+                                   unsigned int dest_pitch,
+                                   unsigned int dest_width) {
+  unsigned int i;
+  unsigned int a, b, c, d, e;
+  unsigned char *des = dest;
+  unsigned char *src = source;
+
+  for (i = 0; i < dest_width; i++) {
+
+    a = src[0 * src_pitch];
+    b = src[1 * src_pitch];
+    c = src[2 * src_pitch];
+    d = src[3 * src_pitch];
+    e = src[4 * src_pitch];
+
+    des[0 * dest_pitch] = (unsigned char) a;
+    des[1 * dest_pitch] = (unsigned char)((b * 192 + c * 64 + 128) >> 8);
+    des[2 * dest_pitch] = (unsigned char)((c * 128 + d * 128 + 128) >> 8);
+    des[3 * dest_pitch] = (unsigned char)((d * 64 + e * 192 + 128) >> 8);
+
+    src++;
+    des++;
+
+  }
+}
+
+
+/*7***************************************************************************
+ *
+ *  ROUTINE       : vp8_horizontal_line_3_5_scale_c
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to source data.
+ *                  unsigned int source_width    : Stride of source.
+ *                  unsigned char *dest         : Pointer to destination data.
+ *                  unsigned int dest_width      : Stride of destination (NOT USED).
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Copies horizontal line of pixels from source to
+ *                  destination scaling up by 3 to 5.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ *
+ ****************************************************************************/
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source,
+                                     unsigned int source_width,
+                                     unsigned char *dest,
+                                     unsigned int dest_width) {
+  unsigned int i;
+  unsigned int a, b, c, d, e;
+  unsigned char *des = dest;
+  const unsigned char *src = source;
+
+  (void) dest_width;
+
+  for (i = 0; i < source_width; i += 5) {
+    a = src[0];
+    b = src[1];
+    c = src[2];
+    d = src[3];
+    e = src[4];
+
+    des[0] = (unsigned char) a;
+    des[1] = (unsigned char)((b * 85  + c * 171 + 128) >> 8);
+    des[2] = (unsigned char)((d * 171 + e * 85 + 128) >> 8);
+
+    src += 5;
+    des += 3;
+  }
+
+}
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source,
+                                   unsigned int src_pitch,
+                                   unsigned char *dest,
+                                   unsigned int dest_pitch,
+                                   unsigned int dest_width) {
+  unsigned int i;
+  unsigned int a, b, c, d, e;
+  unsigned char *des = dest;
+  unsigned char *src = source;
+
+  for (i = 0; i < dest_width; i++) {
+
+    a = src[0 * src_pitch];
+    b = src[1 * src_pitch];
+    c = src[2 * src_pitch];
+    d = src[3 * src_pitch];
+    e = src[4 * src_pitch];
+
+    des[0 * dest_pitch] = (unsigned char) a;
+    des[1 * dest_pitch] = (unsigned char)((b * 85 + c * 171 + 128) >> 8);
+    des[2 * dest_pitch] = (unsigned char)((d * 171 + e * 85 + 128) >> 8);
+
+    src++;
+    des++;
+
+  }
+}
+
+/****************************************************************************
+ *
+ *  ROUTINE       : vp8_horizontal_line_1_2_scale_c
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to source data.
+ *                  unsigned int source_width    : Stride of source.
+ *                  unsigned char *dest         : Pointer to destination data.
+ *                  unsigned int dest_width      : Stride of destination (NOT USED).
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Copies horizontal line of pixels from source to
+ *                  destination scaling up by 1 to 2.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ ****************************************************************************/
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source,
+                                     unsigned int source_width,
+                                     unsigned char *dest,
+                                     unsigned int dest_width) {
+  unsigned int i;
+  unsigned int a;
+  unsigned char *des = dest;
+  const unsigned char *src = source;
+
+  (void) dest_width;
+
+  for (i = 0; i < source_width; i += 2) {
+    a = src[0];
+    des [0] = (unsigned char)(a);
+    src += 2;
+    des += 1;
+  }
+}
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source,
+                                   unsigned int src_pitch,
+                                   unsigned char *dest,
+                                   unsigned int dest_pitch,
+                                   unsigned int dest_width) {
+  (void) dest_pitch;
+  (void) src_pitch;
+  memcpy(dest, source, dest_width);
+}
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source,
+                                     unsigned int src_pitch,
+                                     unsigned char *dest,
+                                     unsigned int dest_pitch,
+                                     unsigned int dest_width) {
+  int i;
+  int temp;
+  int width = dest_width;
+
+  (void) dest_pitch;
+
+  for (i = 0; i < width; i++) {
+    temp = 8;
+    temp += source[i - (int)src_pitch] * 3;
+    temp += source[i] * 10;
+    temp += source[i + src_pitch] * 3;
+    temp >>= 4;
+    dest[i] = (unsigned char)(temp);
+  }
+}
diff --git a/libvpx/libvpx/vpx_scale/generic/vpx_scale.c b/libvpx/libvpx/vpx_scale/generic/vpx_scale.c
new file mode 100644
index 0000000..15e4ba8
--- /dev/null
+++ b/libvpx/libvpx/vpx_scale/generic/vpx_scale.c
@@ -0,0 +1,531 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+/****************************************************************************
+ *
+ *   Module Title :     scale.c
+ *
+ *   Description  :     Image scaling functions.
+ *
+ ***************************************************************************/
+
+/****************************************************************************
+*  Header Files
+****************************************************************************/
+#include "./vpx_scale_rtcd.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vpx_scale/yv12config.h"
+
+typedef struct {
+  int     expanded_frame_width;
+  int     expanded_frame_height;
+
+  int HScale;
+  int HRatio;
+  int VScale;
+  int VRatio;
+
+  YV12_BUFFER_CONFIG *src_yuv_config;
+  YV12_BUFFER_CONFIG *dst_yuv_config;
+
+} SCALE_VARS;
+
+/****************************************************************************
+ *
+ *  ROUTINE       : scale1d_2t1_i
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to data to be scaled.
+ *                  int source_step              : Number of pixels to step on in source.
+ *                  unsigned int source_scale    : Scale for source (UNUSED).
+ *                  unsigned int source_length   : Length of source (UNUSED).
+ *                  unsigned char *dest         : Pointer to output data array.
+ *                  int dest_step                : Number of pixels to step on in destination.
+ *                  unsigned int dest_scale      : Scale for destination (UNUSED).
+ *                  unsigned int dest_length     : Length of destination.
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Performs 2-to-1 interpolated scaling.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ ****************************************************************************/
+static
+void scale1d_2t1_i
+(
+  const unsigned char *source,
+  int source_step,
+  unsigned int source_scale,
+  unsigned int source_length,
+  unsigned char *dest,
+  int dest_step,
+  unsigned int dest_scale,
+  unsigned int dest_length
+) {
+  unsigned int i, j;
+  unsigned int temp;
+  int source_pitch = source_step;
+  (void) source_length;
+  (void) source_scale;
+  (void) dest_scale;
+
+  source_step *= 2;
+  dest[0] = source[0];
+
+  for (i = dest_step, j = source_step; i < dest_length * dest_step; i += dest_step, j += source_step) {
+    temp = 8;
+    temp += 3 * source[j - source_pitch];
+    temp += 10 * source[j];
+    temp += 3 * source[j + source_pitch];
+    temp >>= 4;
+    dest[i] = (char)(temp);
+  }
+}
+
+/****************************************************************************
+ *
+ *  ROUTINE       : scale1d_2t1_ps
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to data to be scaled.
+ *                  int source_step              : Number of pixels to step on in source.
+ *                  unsigned int source_scale    : Scale for source (UNUSED).
+ *                  unsigned int source_length   : Length of source (UNUSED).
+ *                  unsigned char *dest         : Pointer to output data array.
+ *                  int dest_step                : Number of pixels to step on in destination.
+ *                  unsigned int dest_scale      : Scale for destination (UNUSED).
+ *                  unsigned int dest_length     : Length of destination.
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Performs 2-to-1 point subsampled scaling.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ ****************************************************************************/
+static
+void scale1d_2t1_ps
+(
+  const unsigned char *source,
+  int source_step,
+  unsigned int source_scale,
+  unsigned int source_length,
+  unsigned char *dest,
+  int dest_step,
+  unsigned int dest_scale,
+  unsigned int dest_length
+) {
+  unsigned int i, j;
+
+  (void) source_length;
+  (void) source_scale;
+  (void) dest_scale;
+
+  source_step *= 2;
+  j = 0;
+
+  for (i = 0; i < dest_length * dest_step; i += dest_step, j += source_step)
+    dest[i] = source[j];
+}
+/****************************************************************************
+ *
+ *  ROUTINE       : scale1d_c
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to data to be scaled.
+ *                  int source_step              : Number of pixels to step on in source.
+ *                  unsigned int source_scale    : Scale for source.
+ *                  unsigned int source_length   : Length of source (UNUSED).
+ *                  unsigned char *dest         : Pointer to output data array.
+ *                  int dest_step                : Number of pixels to step on in destination.
+ *                  unsigned int dest_scale      : Scale for destination.
+ *                  unsigned int dest_length     : Length of destination.
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Performs linear interpolation in one dimension.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ ****************************************************************************/
+static
+void scale1d_c
+(
+  const unsigned char *source,
+  int source_step,
+  unsigned int source_scale,
+  unsigned int source_length,
+  unsigned char *dest,
+  int dest_step,
+  unsigned int dest_scale,
+  unsigned int dest_length
+) {
+  unsigned int i;
+  unsigned int round_value = dest_scale / 2;
+  unsigned int left_modifier = dest_scale;
+  unsigned int right_modifier = 0;
+  unsigned char left_pixel = *source;
+  unsigned char right_pixel = *(source + source_step);
+
+  (void) source_length;
+
+  /* These asserts are needed if there are boundary issues... */
+  /*assert ( dest_scale > source_scale );*/
+  /*assert ( (source_length-1) * dest_scale >= (dest_length-1) * source_scale );*/
+
+  for (i = 0; i < dest_length * dest_step; i += dest_step) {
+    dest[i] = (char)((left_modifier * left_pixel + right_modifier * right_pixel + round_value) / dest_scale);
+
+    right_modifier += source_scale;
+
+    while (right_modifier > dest_scale) {
+      right_modifier -= dest_scale;
+      source += source_step;
+      left_pixel = *source;
+      right_pixel = *(source + source_step);
+    }
+
+    left_modifier = dest_scale - right_modifier;
+  }
+}
+
+/****************************************************************************
+ *
+ *  ROUTINE       : Scale2D
+ *
+ *  INPUTS        : const unsigned char *source  : Pointer to data to be scaled.
+ *                  int source_pitch              : Stride of source image.
+ *                  unsigned int source_width     : Width of input image.
+ *                  unsigned int source_height    : Height of input image.
+ *                  unsigned char *dest          : Pointer to output data array.
+ *                  int dest_pitch                : Stride of destination image.
+ *                  unsigned int dest_width       : Width of destination image.
+ *                  unsigned int dest_height      : Height of destination image.
+ *                  unsigned char *temp_area      : Pointer to temp work area.
+ *                  unsigned char temp_area_height : Height of temp work area.
+ *                  unsigned int hscale          : Horizontal scale factor numerator.
+ *                  unsigned int hratio          : Horizontal scale factor denominator.
+ *                  unsigned int vscale          : Vertical scale factor numerator.
+ *                  unsigned int vratio          : Vertical scale factor denominator.
+ *                  unsigned int interlaced      : Interlace flag.
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Performs 2-tap linear interpolation in two dimensions.
+ *
+ *  SPECIAL NOTES : Expansion is performed one band at a time to help with
+ *                  caching.
+ *
+ ****************************************************************************/
+static
+void Scale2D
+(
+  /*const*/
+  unsigned char *source,
+  int source_pitch,
+  unsigned int source_width,
+  unsigned int source_height,
+  unsigned char *dest,
+  int dest_pitch,
+  unsigned int dest_width,
+  unsigned int dest_height,
+  unsigned char *temp_area,
+  unsigned char temp_area_height,
+  unsigned int hscale,
+  unsigned int hratio,
+  unsigned int vscale,
+  unsigned int vratio,
+  unsigned int interlaced
+) {
+  /*unsigned*/
+  int i, j, k;
+  int bands;
+  int dest_band_height;
+  int source_band_height;
+
+  typedef void (*Scale1D)(const unsigned char * source, int source_step, unsigned int source_scale, unsigned int source_length,
+                          unsigned char * dest, int dest_step, unsigned int dest_scale, unsigned int dest_length);
+
+  Scale1D Scale1Dv = scale1d_c;
+  Scale1D Scale1Dh = scale1d_c;
+
+  void (*horiz_line_scale)(const unsigned char *, unsigned int, unsigned char *, unsigned int) = NULL;
+  void (*vert_band_scale)(unsigned char *, unsigned int, unsigned char *, unsigned int, unsigned int) = NULL;
+
+  int ratio_scalable = 1;
+  int interpolation = 0;
+
+  unsigned char *source_base; /* = (unsigned char *) ((source_pitch >= 0) ? source : (source + ((source_height-1) * source_pitch))); */
+  unsigned char *line_src;
+
+
+  source_base = (unsigned char *)source;
+
+  if (source_pitch < 0) {
+    int offset;
+
+    offset = (source_height - 1);
+    offset *= source_pitch;
+
+    source_base += offset;
+  }
+
+  /* find out the ratio for each direction */
+  switch (hratio * 10 / hscale) {
+    case 8:
+      /* 4-5 Scale in Width direction */
+      horiz_line_scale = vp8_horizontal_line_5_4_scale;
+      break;
+    case 6:
+      /* 3-5 Scale in Width direction */
+      horiz_line_scale = vp8_horizontal_line_5_3_scale;
+      break;
+    case 5:
+      /* 1-2 Scale in Width direction */
+      horiz_line_scale = vp8_horizontal_line_2_1_scale;
+      break;
+    default:
+      /* The ratio is not acceptable now */
+      /* throw("The ratio is not acceptable for now!"); */
+      ratio_scalable = 0;
+      break;
+  }
+
+  switch (vratio * 10 / vscale) {
+    case 8:
+      /* 4-5 Scale in vertical direction */
+      vert_band_scale     = vp8_vertical_band_5_4_scale;
+      source_band_height  = 5;
+      dest_band_height    = 4;
+      break;
+    case 6:
+      /* 3-5 Scale in vertical direction */
+      vert_band_scale     = vp8_vertical_band_5_3_scale;
+      source_band_height  = 5;
+      dest_band_height    = 3;
+      break;
+    case 5:
+      /* 1-2 Scale in vertical direction */
+
+      if (interlaced) {
+        /* if the content is interlaced, point sampling is used */
+        vert_band_scale     = vp8_vertical_band_2_1_scale;
+      } else {
+
+        interpolation = 1;
+        /* if the content is progressive, interplo */
+        vert_band_scale     = vp8_vertical_band_2_1_scale_i;
+
+      }
+
+      source_band_height  = 2;
+      dest_band_height    = 1;
+      break;
+    default:
+      /* The ratio is not acceptable now */
+      /* throw("The ratio is not acceptable for now!"); */
+      ratio_scalable = 0;
+      break;
+  }
+
+  if (ratio_scalable) {
+    if (source_height == dest_height) {
+      /* for each band of the image */
+      for (k = 0; k < (int)dest_height; k++) {
+        horiz_line_scale(source, source_width, dest, dest_width);
+        source += source_pitch;
+        dest   += dest_pitch;
+      }
+
+      return;
+    }
+
+    if (interpolation) {
+      if (source < source_base)
+        source = source_base;
+
+      horiz_line_scale(source, source_width, temp_area, dest_width);
+    }
+
+    for (k = 0; k < (int)(dest_height + dest_band_height - 1) / dest_band_height; k++) {
+      /* scale one band horizontally */
+      for (i = 0; i < source_band_height; i++) {
+        /* Trap case where we could read off the base of the source buffer */
+
+        line_src = (unsigned char *)source + i * source_pitch;
+
+        if (line_src < source_base)
+          line_src = source_base;
+
+        horiz_line_scale(line_src, source_width,
+                         temp_area + (i + 1)*dest_pitch, dest_width);
+      }
+
+      /* Vertical scaling is in place */
+      vert_band_scale(temp_area + dest_pitch, dest_pitch, dest, dest_pitch, dest_width);
+
+      if (interpolation)
+        memcpy(temp_area, temp_area + source_band_height * dest_pitch, dest_width);
+
+      /* Next band... */
+      source += (unsigned long) source_band_height  * source_pitch;
+      dest   += (unsigned long) dest_band_height * dest_pitch;
+    }
+
+    return;
+  }
+
+  if (hscale == 2 && hratio == 1)
+    Scale1Dh = scale1d_2t1_ps;
+
+  if (vscale == 2 && vratio == 1) {
+    if (interlaced)
+      Scale1Dv = scale1d_2t1_ps;
+    else
+      Scale1Dv = scale1d_2t1_i;
+  }
+
+  if (source_height == dest_height) {
+    /* for each band of the image */
+    for (k = 0; k < (int)dest_height; k++) {
+      Scale1Dh(source, 1, hscale, source_width + 1, dest, 1, hratio, dest_width);
+      source += source_pitch;
+      dest   += dest_pitch;
+    }
+
+    return;
+  }
+
+  if (dest_height > source_height) {
+    dest_band_height   = temp_area_height - 1;
+    source_band_height = dest_band_height * source_height / dest_height;
+  } else {
+    source_band_height = temp_area_height - 1;
+    dest_band_height   = source_band_height * vratio / vscale;
+  }
+
+  /* first row needs to be done so that we can stay one row ahead for vertical zoom */
+  Scale1Dh(source, 1, hscale, source_width + 1, temp_area, 1, hratio, dest_width);
+
+  /* for each band of the image */
+  bands = (dest_height + dest_band_height - 1) / dest_band_height;
+
+  for (k = 0; k < bands; k++) {
+    /* scale one band horizontally */
+    for (i = 1; i < source_band_height + 1; i++) {
+      if (k * source_band_height + i < (int) source_height) {
+        Scale1Dh(source + i * source_pitch, 1, hscale, source_width + 1,
+                 temp_area + i * dest_pitch, 1, hratio, dest_width);
+      } else { /*  Duplicate the last row */
+        /* copy temp_area row 0 over from last row in the past */
+        memcpy(temp_area + i * dest_pitch, temp_area + (i - 1)*dest_pitch, dest_pitch);
+      }
+    }
+
+    /* scale one band vertically */
+    for (j = 0; j < (int)dest_width; j++) {
+      Scale1Dv(&temp_area[j], dest_pitch, vscale, source_band_height + 1,
+               &dest[j], dest_pitch, vratio, dest_band_height);
+    }
+
+    /* copy temp_area row 0 over from last row in the past */
+    memcpy(temp_area, temp_area + source_band_height * dest_pitch, dest_pitch);
+
+    /* move to the next band */
+    source += source_band_height * source_pitch;
+    dest   += dest_band_height * dest_pitch;
+  }
+}
+
+/****************************************************************************
+ *
+ *  ROUTINE       : vpx_scale_frame
+ *
+ *  INPUTS        : YV12_BUFFER_CONFIG *src       : Pointer to frame to be scaled.
+ *                  YV12_BUFFER_CONFIG *dst       : Pointer to buffer to hold scaled frame.
+ *                  unsigned char *temp_area      : Pointer to temp work area.
+ *                  unsigned char temp_area_height : Height of temp work area.
+ *                  unsigned int hscale          : Horizontal scale factor numerator.
+ *                  unsigned int hratio          : Horizontal scale factor denominator.
+ *                  unsigned int vscale          : Vertical scale factor numerator.
+ *                  unsigned int vratio          : Vertical scale factor denominator.
+ *                  unsigned int interlaced      : Interlace flag.
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Performs 2-tap linear interpolation in two dimensions.
+ *
+ *  SPECIAL NOTES : Expansion is performed one band at a time to help with
+ *                  caching.
+ *
+ ****************************************************************************/
+void vpx_scale_frame
+(
+  YV12_BUFFER_CONFIG *src,
+  YV12_BUFFER_CONFIG *dst,
+  unsigned char *temp_area,
+  unsigned char temp_height,
+  unsigned int hscale,
+  unsigned int hratio,
+  unsigned int vscale,
+  unsigned int vratio,
+  unsigned int interlaced
+) {
+  int i;
+  int dw = (hscale - 1 + src->y_width * hratio) / hscale;
+  int dh = (vscale - 1 + src->y_height * vratio) / vscale;
+
+  /* call our internal scaling routines!! */
+  Scale2D((unsigned char *) src->y_buffer, src->y_stride, src->y_width, src->y_height,
+          (unsigned char *) dst->y_buffer, dst->y_stride, dw, dh,
+          temp_area, temp_height, hscale, hratio, vscale, vratio, interlaced);
+
+  if (dw < (int)dst->y_width)
+    for (i = 0; i < dh; i++)
+      memset(dst->y_buffer + i * dst->y_stride + dw - 1, dst->y_buffer[i * dst->y_stride + dw - 2], dst->y_width - dw + 1);
+
+  if (dh < (int)dst->y_height)
+    for (i = dh - 1; i < (int)dst->y_height; i++)
+      memcpy(dst->y_buffer + i * dst->y_stride, dst->y_buffer + (dh - 2) * dst->y_stride, dst->y_width + 1);
+
+  Scale2D((unsigned char *) src->u_buffer, src->uv_stride, src->uv_width, src->uv_height,
+          (unsigned char *) dst->u_buffer, dst->uv_stride, dw / 2, dh / 2,
+          temp_area, temp_height, hscale, hratio, vscale, vratio, interlaced);
+
+  if (dw / 2 < (int)dst->uv_width)
+    for (i = 0; i < dst->uv_height; i++)
+      memset(dst->u_buffer + i * dst->uv_stride + dw / 2 - 1, dst->u_buffer[i * dst->uv_stride + dw / 2 - 2], dst->uv_width - dw / 2 + 1);
+
+  if (dh / 2 < (int)dst->uv_height)
+    for (i = dh / 2 - 1; i < (int)dst->y_height / 2; i++)
+      memcpy(dst->u_buffer + i * dst->uv_stride, dst->u_buffer + (dh / 2 - 2)*dst->uv_stride, dst->uv_width);
+
+  Scale2D((unsigned char *) src->v_buffer, src->uv_stride, src->uv_width, src->uv_height,
+          (unsigned char *) dst->v_buffer, dst->uv_stride, dw / 2, dh / 2,
+          temp_area, temp_height, hscale, hratio, vscale, vratio, interlaced);
+
+  if (dw / 2 < (int)dst->uv_width)
+    for (i = 0; i < dst->uv_height; i++)
+      memset(dst->v_buffer + i * dst->uv_stride + dw / 2 - 1, dst->v_buffer[i * dst->uv_stride + dw / 2 - 2], dst->uv_width - dw / 2 + 1);
+
+  if (dh / 2 < (int) dst->uv_height)
+    for (i = dh / 2 - 1; i < (int)dst->y_height / 2; i++)
+      memcpy(dst->v_buffer + i * dst->uv_stride, dst->v_buffer + (dh / 2 - 2)*dst->uv_stride, dst->uv_width);
+}
diff --git a/libvpx/libvpx/vpx_scale/generic/yv12config.c b/libvpx/libvpx/vpx_scale/generic/yv12config.c
new file mode 100644
index 0000000..6bbb6d8
--- /dev/null
+++ b/libvpx/libvpx/vpx_scale/generic/yv12config.c
@@ -0,0 +1,287 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vpx_scale/yv12config.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+/****************************************************************************
+*  Exports
+****************************************************************************/
+
+/****************************************************************************
+ *
+ ****************************************************************************/
+#define yv12_align_addr(addr, align) \
+    (void*)(((size_t)(addr) + ((align) - 1)) & (size_t)-(align))
+
+int
+vp8_yv12_de_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf) {
+  if (ybf) {
+    // If libvpx is using frame buffer callbacks then buffer_alloc_sz must
+    // not be set.
+    if (ybf->buffer_alloc_sz > 0) {
+      vpx_free(ybf->buffer_alloc);
+    }
+
+    /* buffer_alloc isn't accessed by most functions.  Rather y_buffer,
+      u_buffer and v_buffer point to buffer_alloc and are used.  Clear out
+      all of this so that a freed pointer isn't inadvertently used */
+    memset(ybf, 0, sizeof(YV12_BUFFER_CONFIG));
+  } else {
+    return -1;
+  }
+
+  return 0;
+}
+
+int vp8_yv12_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
+                                  int width, int height, int border) {
+  if (ybf) {
+    int aligned_width = (width + 15) & ~15;
+    int aligned_height = (height + 15) & ~15;
+    int y_stride = ((aligned_width + 2 * border) + 31) & ~31;
+    int yplane_size = (aligned_height + 2 * border) * y_stride;
+    int uv_width = aligned_width >> 1;
+    int uv_height = aligned_height >> 1;
+    /** There is currently a bunch of code which assumes
+      *  uv_stride == y_stride/2, so enforce this here. */
+    int uv_stride = y_stride >> 1;
+    int uvplane_size = (uv_height + border) * uv_stride;
+    const int frame_size = yplane_size + 2 * uvplane_size;
+
+    if (!ybf->buffer_alloc) {
+      ybf->buffer_alloc = (uint8_t *)vpx_memalign(32, frame_size);
+      ybf->buffer_alloc_sz = frame_size;
+    }
+
+    if (!ybf->buffer_alloc || ybf->buffer_alloc_sz < frame_size)
+      return -1;
+
+    /* Only support allocating buffers that have a border that's a multiple
+     * of 32. The border restriction is required to get 16-byte alignment of
+     * the start of the chroma rows without introducing an arbitrary gap
+     * between planes, which would break the semantics of things like
+     * vpx_img_set_rect(). */
+    if (border & 0x1f)
+      return -3;
+
+    ybf->y_crop_width = width;
+    ybf->y_crop_height = height;
+    ybf->y_width  = aligned_width;
+    ybf->y_height = aligned_height;
+    ybf->y_stride = y_stride;
+
+    ybf->uv_crop_width = (width + 1) / 2;
+    ybf->uv_crop_height = (height + 1) / 2;
+    ybf->uv_width = uv_width;
+    ybf->uv_height = uv_height;
+    ybf->uv_stride = uv_stride;
+
+    ybf->alpha_width = 0;
+    ybf->alpha_height = 0;
+    ybf->alpha_stride = 0;
+
+    ybf->border = border;
+    ybf->frame_size = frame_size;
+
+    ybf->y_buffer = ybf->buffer_alloc + (border * y_stride) + border;
+    ybf->u_buffer = ybf->buffer_alloc + yplane_size + (border / 2  * uv_stride) + border / 2;
+    ybf->v_buffer = ybf->buffer_alloc + yplane_size + uvplane_size + (border / 2  * uv_stride) + border / 2;
+    ybf->alpha_buffer = NULL;
+
+    ybf->corrupted = 0; /* assume not currupted by errors */
+    return 0;
+  }
+  return -2;
+}
+
+int vp8_yv12_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
+                                int width, int height, int border) {
+  if (ybf) {
+    vp8_yv12_de_alloc_frame_buffer(ybf);
+    return vp8_yv12_realloc_frame_buffer(ybf, width, height, border);
+  }
+  return -2;
+}
+
+#if CONFIG_VP9
+// TODO(jkoleszar): Maybe replace this with struct vpx_image
+
+int vpx_free_frame_buffer(YV12_BUFFER_CONFIG *ybf) {
+  if (ybf) {
+    if (ybf->buffer_alloc_sz > 0) {
+      vpx_free(ybf->buffer_alloc);
+    }
+
+    /* buffer_alloc isn't accessed by most functions.  Rather y_buffer,
+      u_buffer and v_buffer point to buffer_alloc and are used.  Clear out
+      all of this so that a freed pointer isn't inadvertently used */
+    memset(ybf, 0, sizeof(YV12_BUFFER_CONFIG));
+  } else {
+    return -1;
+  }
+
+  return 0;
+}
+
+int vpx_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
+                             int width, int height,
+                             int ss_x, int ss_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                             int use_highbitdepth,
+#endif
+                             int border,
+                             int byte_alignment,
+                             vpx_codec_frame_buffer_t *fb,
+                             vpx_get_frame_buffer_cb_fn_t cb,
+                             void *cb_priv) {
+  if (ybf) {
+    const int vp9_byte_align = (byte_alignment == 0) ? 1 : byte_alignment;
+    const int aligned_width = (width + 7) & ~7;
+    const int aligned_height = (height + 7) & ~7;
+    const int y_stride = ((aligned_width + 2 * border) + 31) & ~31;
+    const uint64_t yplane_size = (aligned_height + 2 * border) *
+                                 (uint64_t)y_stride + byte_alignment;
+    const int uv_width = aligned_width >> ss_x;
+    const int uv_height = aligned_height >> ss_y;
+    const int uv_stride = y_stride >> ss_x;
+    const int uv_border_w = border >> ss_x;
+    const int uv_border_h = border >> ss_y;
+    const uint64_t uvplane_size = (uv_height + 2 * uv_border_h) *
+                                  (uint64_t)uv_stride + byte_alignment;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+    const uint64_t frame_size =
+        (1 + use_highbitdepth) * (yplane_size + 2 * uvplane_size);
+#else
+    const uint64_t frame_size = yplane_size + 2 * uvplane_size;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    uint8_t *buf = NULL;
+
+    if (cb != NULL) {
+      const int align_addr_extra_size = 31;
+      const uint64_t external_frame_size = frame_size + align_addr_extra_size;
+
+      assert(fb != NULL);
+
+      if (external_frame_size != (size_t)external_frame_size)
+        return -1;
+
+      // Allocation to hold larger frame, or first allocation.
+      if (cb(cb_priv, (size_t)external_frame_size, fb) < 0)
+        return -1;
+
+      if (fb->data == NULL || fb->size < external_frame_size)
+        return -1;
+
+      ybf->buffer_alloc = (uint8_t *)yv12_align_addr(fb->data, 32);
+
+#if defined(__has_feature)
+#if __has_feature(memory_sanitizer)
+      // This memset is needed for fixing the issue of using uninitialized
+      // value in msan test. It will cause a perf loss, so only do this for
+      // msan test.
+      memset(ybf->buffer_alloc, 0, (int)frame_size);
+#endif
+#endif
+    } else if (frame_size > (size_t)ybf->buffer_alloc_sz) {
+      // Allocation to hold larger frame, or first allocation.
+      vpx_free(ybf->buffer_alloc);
+      ybf->buffer_alloc = NULL;
+
+      if (frame_size != (size_t)frame_size)
+        return -1;
+
+      ybf->buffer_alloc = (uint8_t *)vpx_memalign(32, (size_t)frame_size);
+      if (!ybf->buffer_alloc)
+        return -1;
+
+      ybf->buffer_alloc_sz = (int)frame_size;
+
+      // This memset is needed for fixing valgrind error from C loop filter
+      // due to access uninitialized memory in frame border. It could be
+      // removed if border is totally removed.
+      memset(ybf->buffer_alloc, 0, ybf->buffer_alloc_sz);
+    }
+
+    /* Only support allocating buffers that have a border that's a multiple
+     * of 32. The border restriction is required to get 16-byte alignment of
+     * the start of the chroma rows without introducing an arbitrary gap
+     * between planes, which would break the semantics of things like
+     * vpx_img_set_rect(). */
+    if (border & 0x1f)
+      return -3;
+
+    ybf->y_crop_width = width;
+    ybf->y_crop_height = height;
+    ybf->y_width  = aligned_width;
+    ybf->y_height = aligned_height;
+    ybf->y_stride = y_stride;
+
+    ybf->uv_crop_width = (width + ss_x) >> ss_x;
+    ybf->uv_crop_height = (height + ss_y) >> ss_y;
+    ybf->uv_width = uv_width;
+    ybf->uv_height = uv_height;
+    ybf->uv_stride = uv_stride;
+
+    ybf->border = border;
+    ybf->frame_size = (int)frame_size;
+    ybf->subsampling_x = ss_x;
+    ybf->subsampling_y = ss_y;
+
+    buf = ybf->buffer_alloc;
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (use_highbitdepth) {
+      // Store uint16 addresses when using 16bit framebuffers
+      buf = CONVERT_TO_BYTEPTR(ybf->buffer_alloc);
+      ybf->flags = YV12_FLAG_HIGHBITDEPTH;
+    } else {
+      ybf->flags = 0;
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    ybf->y_buffer = (uint8_t *)yv12_align_addr(
+        buf + (border * y_stride) + border, vp9_byte_align);
+    ybf->u_buffer = (uint8_t *)yv12_align_addr(
+        buf + yplane_size + (uv_border_h * uv_stride) + uv_border_w,
+        vp9_byte_align);
+    ybf->v_buffer = (uint8_t *)yv12_align_addr(
+        buf + yplane_size + uvplane_size + (uv_border_h * uv_stride) +
+        uv_border_w, vp9_byte_align);
+
+    ybf->corrupted = 0; /* assume not corrupted by errors */
+    return 0;
+  }
+  return -2;
+}
+
+int vpx_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
+                           int width, int height,
+                           int ss_x, int ss_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                           int use_highbitdepth,
+#endif
+                           int border,
+                           int byte_alignment) {
+  if (ybf) {
+    vpx_free_frame_buffer(ybf);
+    return vpx_realloc_frame_buffer(ybf, width, height, ss_x, ss_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                    use_highbitdepth,
+#endif
+                                    border, byte_alignment, NULL, NULL, NULL);
+  }
+  return -2;
+}
+#endif
diff --git a/libvpx/libvpx/vpx_scale/generic/yv12extend.c b/libvpx/libvpx/vpx_scale/generic/yv12extend.c
new file mode 100644
index 0000000..52f0aff
--- /dev/null
+++ b/libvpx/libvpx/vpx_scale/generic/yv12extend.c
@@ -0,0 +1,324 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include "./vpx_config.h"
+#include "./vpx_scale_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_scale/yv12config.h"
+#if CONFIG_VP9_HIGHBITDEPTH
+#include "vp9/common/vp9_common.h"
+#endif
+
+static void extend_plane(uint8_t *const src, int src_stride,
+                         int width, int height,
+                         int extend_top, int extend_left,
+                         int extend_bottom, int extend_right) {
+  int i;
+  const int linesize = extend_left + extend_right + width;
+
+  /* copy the left and right most columns out */
+  uint8_t *src_ptr1 = src;
+  uint8_t *src_ptr2 = src + width - 1;
+  uint8_t *dst_ptr1 = src - extend_left;
+  uint8_t *dst_ptr2 = src + width;
+
+  for (i = 0; i < height; ++i) {
+    memset(dst_ptr1, src_ptr1[0], extend_left);
+    memset(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_stride;
+    src_ptr2 += src_stride;
+    dst_ptr1 += src_stride;
+    dst_ptr2 += src_stride;
+  }
+
+  /* Now copy the top and bottom lines into each line of the respective
+   * borders
+   */
+  src_ptr1 = src - extend_left;
+  src_ptr2 = src + src_stride * (height - 1) - extend_left;
+  dst_ptr1 = src + src_stride * -extend_top - extend_left;
+  dst_ptr2 = src + src_stride * height - extend_left;
+
+  for (i = 0; i < extend_top; ++i) {
+    memcpy(dst_ptr1, src_ptr1, linesize);
+    dst_ptr1 += src_stride;
+  }
+
+  for (i = 0; i < extend_bottom; ++i) {
+    memcpy(dst_ptr2, src_ptr2, linesize);
+    dst_ptr2 += src_stride;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void extend_plane_high(uint8_t *const src8, int src_stride,
+                              int width, int height,
+                              int extend_top, int extend_left,
+                              int extend_bottom, int extend_right) {
+  int i;
+  const int linesize = extend_left + extend_right + width;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+
+  /* copy the left and right most columns out */
+  uint16_t *src_ptr1 = src;
+  uint16_t *src_ptr2 = src + width - 1;
+  uint16_t *dst_ptr1 = src - extend_left;
+  uint16_t *dst_ptr2 = src + width;
+
+  for (i = 0; i < height; ++i) {
+    vpx_memset16(dst_ptr1, src_ptr1[0], extend_left);
+    vpx_memset16(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_stride;
+    src_ptr2 += src_stride;
+    dst_ptr1 += src_stride;
+    dst_ptr2 += src_stride;
+  }
+
+  /* Now copy the top and bottom lines into each line of the respective
+   * borders
+   */
+  src_ptr1 = src - extend_left;
+  src_ptr2 = src + src_stride * (height - 1) - extend_left;
+  dst_ptr1 = src + src_stride * -extend_top - extend_left;
+  dst_ptr2 = src + src_stride * height - extend_left;
+
+  for (i = 0; i < extend_top; ++i) {
+    memcpy(dst_ptr1, src_ptr1, linesize * sizeof(uint16_t));
+    dst_ptr1 += src_stride;
+  }
+
+  for (i = 0; i < extend_bottom; ++i) {
+    memcpy(dst_ptr2, src_ptr2, linesize * sizeof(uint16_t));
+    dst_ptr2 += src_stride;
+  }
+}
+#endif
+
+void vp8_yv12_extend_frame_borders_c(YV12_BUFFER_CONFIG *ybf) {
+  const int uv_border = ybf->border / 2;
+
+  assert(ybf->border % 2 == 0);
+  assert(ybf->y_height - ybf->y_crop_height < 16);
+  assert(ybf->y_width - ybf->y_crop_width < 16);
+  assert(ybf->y_height - ybf->y_crop_height >= 0);
+  assert(ybf->y_width - ybf->y_crop_width >= 0);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (ybf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    extend_plane_high(
+        ybf->y_buffer, ybf->y_stride,
+        ybf->y_crop_width, ybf->y_crop_height,
+        ybf->border, ybf->border,
+        ybf->border + ybf->y_height - ybf->y_crop_height,
+        ybf->border + ybf->y_width - ybf->y_crop_width);
+
+    extend_plane_high(
+        ybf->u_buffer, ybf->uv_stride,
+        ybf->uv_crop_width, ybf->uv_crop_height,
+        uv_border, uv_border,
+        uv_border + ybf->uv_height - ybf->uv_crop_height,
+        uv_border + ybf->uv_width - ybf->uv_crop_width);
+
+    extend_plane_high(
+        ybf->v_buffer, ybf->uv_stride,
+        ybf->uv_crop_width, ybf->uv_crop_height,
+        uv_border, uv_border,
+        uv_border + ybf->uv_height - ybf->uv_crop_height,
+        uv_border + ybf->uv_width - ybf->uv_crop_width);
+    return;
+  }
+#endif
+  extend_plane(ybf->y_buffer, ybf->y_stride,
+               ybf->y_crop_width, ybf->y_crop_height,
+               ybf->border, ybf->border,
+               ybf->border + ybf->y_height - ybf->y_crop_height,
+               ybf->border + ybf->y_width - ybf->y_crop_width);
+
+  extend_plane(ybf->u_buffer, ybf->uv_stride,
+               ybf->uv_crop_width, ybf->uv_crop_height,
+               uv_border, uv_border,
+               uv_border + ybf->uv_height - ybf->uv_crop_height,
+               uv_border + ybf->uv_width - ybf->uv_crop_width);
+
+  extend_plane(ybf->v_buffer, ybf->uv_stride,
+               ybf->uv_crop_width, ybf->uv_crop_height,
+               uv_border, uv_border,
+               uv_border + ybf->uv_height - ybf->uv_crop_height,
+               uv_border + ybf->uv_width - ybf->uv_crop_width);
+}
+
+#if CONFIG_VP9
+static void extend_frame(YV12_BUFFER_CONFIG *const ybf, int ext_size) {
+  const int c_w = ybf->uv_crop_width;
+  const int c_h = ybf->uv_crop_height;
+  const int ss_x = ybf->uv_width < ybf->y_width;
+  const int ss_y = ybf->uv_height < ybf->y_height;
+  const int c_et = ext_size >> ss_y;
+  const int c_el = ext_size >> ss_x;
+  const int c_eb = c_et + ybf->uv_height - ybf->uv_crop_height;
+  const int c_er = c_el + ybf->uv_width - ybf->uv_crop_width;
+
+  assert(ybf->y_height - ybf->y_crop_height < 16);
+  assert(ybf->y_width - ybf->y_crop_width < 16);
+  assert(ybf->y_height - ybf->y_crop_height >= 0);
+  assert(ybf->y_width - ybf->y_crop_width >= 0);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (ybf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    extend_plane_high(ybf->y_buffer, ybf->y_stride,
+                      ybf->y_crop_width, ybf->y_crop_height,
+                      ext_size, ext_size,
+                      ext_size + ybf->y_height - ybf->y_crop_height,
+                      ext_size + ybf->y_width - ybf->y_crop_width);
+    extend_plane_high(ybf->u_buffer, ybf->uv_stride,
+                      c_w, c_h, c_et, c_el, c_eb, c_er);
+    extend_plane_high(ybf->v_buffer, ybf->uv_stride,
+                      c_w, c_h, c_et, c_el, c_eb, c_er);
+    return;
+  }
+#endif
+  extend_plane(ybf->y_buffer, ybf->y_stride,
+               ybf->y_crop_width, ybf->y_crop_height,
+               ext_size, ext_size,
+               ext_size + ybf->y_height - ybf->y_crop_height,
+               ext_size + ybf->y_width - ybf->y_crop_width);
+
+  extend_plane(ybf->u_buffer, ybf->uv_stride,
+               c_w, c_h, c_et, c_el, c_eb, c_er);
+
+  extend_plane(ybf->v_buffer, ybf->uv_stride,
+               c_w, c_h, c_et, c_el, c_eb, c_er);
+}
+
+void vpx_extend_frame_borders_c(YV12_BUFFER_CONFIG *ybf) {
+  extend_frame(ybf, ybf->border);
+}
+
+void vpx_extend_frame_inner_borders_c(YV12_BUFFER_CONFIG *ybf) {
+  const int inner_bw = (ybf->border > VP9INNERBORDERINPIXELS) ?
+                       VP9INNERBORDERINPIXELS : ybf->border;
+  extend_frame(ybf, inner_bw);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void memcpy_short_addr(uint8_t *dst8, const uint8_t *src8, int num) {
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  memcpy(dst, src, num * sizeof(uint16_t));
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#endif  // CONFIG_VP9
+
+// Copies the source image into the destination image and updates the
+// destination's UMV borders.
+// Note: The frames are assumed to be identical in size.
+void vp8_yv12_copy_frame_c(const YV12_BUFFER_CONFIG *src_ybc,
+                           YV12_BUFFER_CONFIG *dst_ybc) {
+  int row;
+  const uint8_t *src = src_ybc->y_buffer;
+  uint8_t *dst = dst_ybc->y_buffer;
+
+#if 0
+  /* These assertions are valid in the codec, but the libvpx-tester uses
+   * this code slightly differently.
+   */
+  assert(src_ybc->y_width == dst_ybc->y_width);
+  assert(src_ybc->y_height == dst_ybc->y_height);
+#endif
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (src_ybc->flags & YV12_FLAG_HIGHBITDEPTH) {
+    assert(dst_ybc->flags & YV12_FLAG_HIGHBITDEPTH);
+    for (row = 0; row < src_ybc->y_height; ++row) {
+      memcpy_short_addr(dst, src, src_ybc->y_width);
+      src += src_ybc->y_stride;
+      dst += dst_ybc->y_stride;
+    }
+
+    src = src_ybc->u_buffer;
+    dst = dst_ybc->u_buffer;
+
+    for (row = 0; row < src_ybc->uv_height; ++row) {
+      memcpy_short_addr(dst, src, src_ybc->uv_width);
+      src += src_ybc->uv_stride;
+      dst += dst_ybc->uv_stride;
+    }
+
+    src = src_ybc->v_buffer;
+    dst = dst_ybc->v_buffer;
+
+    for (row = 0; row < src_ybc->uv_height; ++row) {
+      memcpy_short_addr(dst, src, src_ybc->uv_width);
+      src += src_ybc->uv_stride;
+      dst += dst_ybc->uv_stride;
+    }
+
+    vp8_yv12_extend_frame_borders_c(dst_ybc);
+    return;
+  } else {
+    assert(!(dst_ybc->flags & YV12_FLAG_HIGHBITDEPTH));
+  }
+#endif
+
+  for (row = 0; row < src_ybc->y_height; ++row) {
+    memcpy(dst, src, src_ybc->y_width);
+    src += src_ybc->y_stride;
+    dst += dst_ybc->y_stride;
+  }
+
+  src = src_ybc->u_buffer;
+  dst = dst_ybc->u_buffer;
+
+  for (row = 0; row < src_ybc->uv_height; ++row) {
+    memcpy(dst, src, src_ybc->uv_width);
+    src += src_ybc->uv_stride;
+    dst += dst_ybc->uv_stride;
+  }
+
+  src = src_ybc->v_buffer;
+  dst = dst_ybc->v_buffer;
+
+  for (row = 0; row < src_ybc->uv_height; ++row) {
+    memcpy(dst, src, src_ybc->uv_width);
+    src += src_ybc->uv_stride;
+    dst += dst_ybc->uv_stride;
+  }
+
+  vp8_yv12_extend_frame_borders_c(dst_ybc);
+}
+
+void vpx_yv12_copy_y_c(const YV12_BUFFER_CONFIG *src_ybc,
+                       YV12_BUFFER_CONFIG *dst_ybc) {
+  int row;
+  const uint8_t *src = src_ybc->y_buffer;
+  uint8_t *dst = dst_ybc->y_buffer;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (src_ybc->flags & YV12_FLAG_HIGHBITDEPTH) {
+    const uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+    uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
+    for (row = 0; row < src_ybc->y_height; ++row) {
+      memcpy(dst16, src16, src_ybc->y_width * sizeof(uint16_t));
+      src16 += src_ybc->y_stride;
+      dst16 += dst_ybc->y_stride;
+    }
+    return;
+  }
+#endif
+
+  for (row = 0; row < src_ybc->y_height; ++row) {
+    memcpy(dst, src, src_ybc->y_width);
+    src += src_ybc->y_stride;
+    dst += dst_ybc->y_stride;
+  }
+}
diff --git a/libvpx/libvpx/vpx_scale/mips/dspr2/yv12extend_dspr2.c b/libvpx/libvpx/vpx_scale/mips/dspr2/yv12extend_dspr2.c
new file mode 100644
index 0000000..aab4785
--- /dev/null
+++ b/libvpx/libvpx/vpx_scale/mips/dspr2/yv12extend_dspr2.c
@@ -0,0 +1,144 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx_scale/yv12config.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_scale/vpx_scale.h"
+
+#if HAVE_DSPR2
+static void extend_plane(uint8_t *const src, int src_stride,
+                         int width, int height,
+                         int extend_top, int extend_left,
+                         int extend_bottom, int extend_right) {
+  int       i, j;
+  uint8_t   *left_src, *right_src;
+  uint8_t   *left_dst_start, *right_dst_start;
+  uint8_t   *left_dst, *right_dst;
+  uint8_t   *top_src, *bot_src;
+  uint8_t   *top_dst, *bot_dst;
+  uint32_t  left_pix;
+  uint32_t  right_pix;
+  uint32_t  linesize;
+
+  /* copy the left and right most columns out */
+  left_src  = src;
+  right_src = src + width - 1;
+  left_dst_start = src - extend_left;
+  right_dst_start = src + width;
+
+  for (i = height; i--; ) {
+    left_dst  = left_dst_start;
+    right_dst = right_dst_start;
+
+    __asm__ __volatile__ (
+        "lb        %[left_pix],     0(%[left_src])      \n\t"
+        "lb        %[right_pix],    0(%[right_src])     \n\t"
+        "replv.qb  %[left_pix],     %[left_pix]         \n\t"
+        "replv.qb  %[right_pix],    %[right_pix]        \n\t"
+
+        : [left_pix] "=&r" (left_pix), [right_pix] "=&r" (right_pix)
+        : [left_src] "r" (left_src), [right_src] "r" (right_src)
+    );
+
+    for (j = extend_left/4; j--; ) {
+      __asm__ __volatile__ (
+        "sw     %[left_pix],    0(%[left_dst])     \n\t"
+        "sw     %[right_pix],   0(%[right_dst])    \n\t"
+
+        :
+        : [left_dst] "r" (left_dst), [left_pix] "r" (left_pix),
+          [right_dst] "r" (right_dst), [right_pix] "r" (right_pix)
+      );
+
+      left_dst += 4;
+      right_dst += 4;
+    }
+
+    for (j = extend_left%4; j--; ) {
+      __asm__ __volatile__ (
+        "sb     %[left_pix],    0(%[left_dst])     \n\t"
+        "sb     %[right_pix],   0(%[right_dst])     \n\t"
+
+        :
+        : [left_dst] "r" (left_dst), [left_pix] "r" (left_pix),
+          [right_dst] "r" (right_dst), [right_pix] "r" (right_pix)
+      );
+
+      left_dst += 1;
+      right_dst += 1;
+    }
+
+    left_src  += src_stride;
+    right_src += src_stride;
+    left_dst_start += src_stride;
+    right_dst_start += src_stride;
+  }
+
+  /* Now copy the top and bottom lines into each line of the respective
+   * borders
+   */
+  top_src = src - extend_left;
+  bot_src = src + src_stride * (height - 1) - extend_left;
+  top_dst = src + src_stride * (-extend_top) - extend_left;
+  bot_dst = src + src_stride * (height) - extend_left;
+  linesize = extend_left + extend_right + width;
+
+  for (i = 0; i < extend_top; i++) {
+    memcpy(top_dst, top_src, linesize);
+    top_dst += src_stride;
+  }
+
+  for (i = 0; i < extend_bottom; i++) {
+    memcpy(bot_dst, bot_src, linesize);
+    bot_dst += src_stride;
+  }
+}
+
+static void extend_frame(YV12_BUFFER_CONFIG *const ybf, int ext_size) {
+  const int c_w = ybf->uv_crop_width;
+  const int c_h = ybf->uv_crop_height;
+  const int ss_x = ybf->uv_width < ybf->y_width;
+  const int ss_y = ybf->uv_height < ybf->y_height;
+  const int c_et = ext_size >> ss_y;
+  const int c_el = ext_size >> ss_x;
+  const int c_eb = c_et + ybf->uv_height - ybf->uv_crop_height;
+  const int c_er = c_el + ybf->uv_width - ybf->uv_crop_width;
+
+  assert(ybf->y_height - ybf->y_crop_height < 16);
+  assert(ybf->y_width - ybf->y_crop_width < 16);
+  assert(ybf->y_height - ybf->y_crop_height >= 0);
+  assert(ybf->y_width - ybf->y_crop_width >= 0);
+
+  extend_plane(ybf->y_buffer, ybf->y_stride,
+               ybf->y_crop_width, ybf->y_crop_height,
+               ext_size, ext_size,
+               ext_size + ybf->y_height - ybf->y_crop_height,
+               ext_size + ybf->y_width - ybf->y_crop_width);
+
+  extend_plane(ybf->u_buffer, ybf->uv_stride,
+               c_w, c_h, c_et, c_el, c_eb, c_er);
+
+  extend_plane(ybf->v_buffer, ybf->uv_stride,
+               c_w, c_h, c_et, c_el, c_eb, c_er);
+}
+
+void vpx_extend_frame_borders_dspr2(YV12_BUFFER_CONFIG *ybf) {
+  extend_frame(ybf, ybf->border);
+}
+
+void vpx_extend_frame_inner_borders_dspr2(YV12_BUFFER_CONFIG *ybf) {
+  const int inner_bw = (ybf->border > VP9INNERBORDERINPIXELS) ?
+                       VP9INNERBORDERINPIXELS : ybf->border;
+  extend_frame(ybf, inner_bw);
+}
+#endif
diff --git a/libvpx/libvpx/vpx_scale/vpx_scale.h b/libvpx/libvpx/vpx_scale/vpx_scale.h
new file mode 100644
index 0000000..43fcf9d
--- /dev/null
+++ b/libvpx/libvpx/vpx_scale/vpx_scale.h
@@ -0,0 +1,27 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VPX_SCALE_VPX_SCALE_H_
+#define VPX_SCALE_VPX_SCALE_H_
+
+#include "vpx_scale/yv12config.h"
+
+extern void vpx_scale_frame(YV12_BUFFER_CONFIG *src,
+                            YV12_BUFFER_CONFIG *dst,
+                            unsigned char *temp_area,
+                            unsigned char temp_height,
+                            unsigned int hscale,
+                            unsigned int hratio,
+                            unsigned int vscale,
+                            unsigned int vratio,
+                            unsigned int interlaced);
+
+#endif  // VPX_SCALE_VPX_SCALE_H_
diff --git a/libvpx/libvpx/vpx_scale/vpx_scale.mk b/libvpx/libvpx/vpx_scale/vpx_scale.mk
new file mode 100644
index 0000000..a49abf3
--- /dev/null
+++ b/libvpx/libvpx/vpx_scale/vpx_scale.mk
@@ -0,0 +1,16 @@
+SCALE_SRCS-yes += vpx_scale.mk
+SCALE_SRCS-yes += yv12config.h
+SCALE_SRCS-$(CONFIG_SPATIAL_RESAMPLING) += vpx_scale.h
+SCALE_SRCS-$(CONFIG_SPATIAL_RESAMPLING) += generic/vpx_scale.c
+SCALE_SRCS-yes += generic/yv12config.c
+SCALE_SRCS-yes += generic/yv12extend.c
+SCALE_SRCS-$(CONFIG_SPATIAL_RESAMPLING) += generic/gen_scalers.c
+SCALE_SRCS-yes += vpx_scale_rtcd.c
+SCALE_SRCS-yes += vpx_scale_rtcd.pl
+
+#mips(dspr2)
+SCALE_SRCS-$(HAVE_DSPR2)  += mips/dspr2/yv12extend_dspr2.c
+
+SCALE_SRCS-no += $(SCALE_SRCS_REMOVE-yes)
+
+$(eval $(call rtcd_h_template,vpx_scale_rtcd,vpx_scale/vpx_scale_rtcd.pl))
diff --git a/libvpx/libvpx/vpx_scale/vpx_scale_rtcd.c b/libvpx/libvpx/vpx_scale/vpx_scale_rtcd.c
new file mode 100644
index 0000000..bea603f
--- /dev/null
+++ b/libvpx/libvpx/vpx_scale/vpx_scale_rtcd.c
@@ -0,0 +1,18 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_config.h"
+#define RTCD_C
+#include "./vpx_scale_rtcd.h"
+#include "vpx_ports/vpx_once.h"
+
+void vpx_scale_rtcd()
+{
+    once(setup_rtcd_internal);
+}
diff --git a/libvpx/libvpx/vpx_scale/vpx_scale_rtcd.pl b/libvpx/libvpx/vpx_scale/vpx_scale_rtcd.pl
new file mode 100644
index 0000000..44b115c
--- /dev/null
+++ b/libvpx/libvpx/vpx_scale/vpx_scale_rtcd.pl
@@ -0,0 +1,32 @@
+sub vpx_scale_forward_decls() {
+print <<EOF
+struct yv12_buffer_config;
+EOF
+}
+forward_decls qw/vpx_scale_forward_decls/;
+
+# Scaler functions
+if (vpx_config("CONFIG_SPATIAL_RESAMPLING") eq "yes") {
+    add_proto qw/void vp8_horizontal_line_5_4_scale/, "const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width";
+    add_proto qw/void vp8_vertical_band_5_4_scale/, "unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width";
+    add_proto qw/void vp8_horizontal_line_5_3_scale/, "const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width";
+    add_proto qw/void vp8_vertical_band_5_3_scale/, "unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width";
+    add_proto qw/void vp8_horizontal_line_2_1_scale/, "const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width";
+    add_proto qw/void vp8_vertical_band_2_1_scale/, "unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width";
+    add_proto qw/void vp8_vertical_band_2_1_scale_i/, "unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width";
+}
+
+add_proto qw/void vp8_yv12_extend_frame_borders/, "struct yv12_buffer_config *ybf";
+
+add_proto qw/void vp8_yv12_copy_frame/, "const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc";
+
+add_proto qw/void vpx_yv12_copy_y/, "const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc";
+
+if (vpx_config("CONFIG_VP9") eq "yes") {
+    add_proto qw/void vpx_extend_frame_borders/, "struct yv12_buffer_config *ybf";
+    specialize qw/vpx_extend_frame_borders dspr2/;
+
+    add_proto qw/void vpx_extend_frame_inner_borders/, "struct yv12_buffer_config *ybf";
+    specialize qw/vpx_extend_frame_inner_borders dspr2/;
+}
+1;
diff --git a/libvpx/libvpx/vpx_scale/yv12config.h b/libvpx/libvpx/vpx_scale/yv12config.h
new file mode 100644
index 0000000..37b255d
--- /dev/null
+++ b/libvpx/libvpx/vpx_scale/yv12config.h
@@ -0,0 +1,105 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_SCALE_YV12CONFIG_H_
+#define VPX_SCALE_YV12CONFIG_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./vpx_config.h"
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_frame_buffer.h"
+#include "vpx/vpx_integer.h"
+
+#define VP8BORDERINPIXELS           32
+#define VP9INNERBORDERINPIXELS      96
+#define VP9_INTERP_EXTEND           4
+#define VP9_ENC_BORDER_IN_PIXELS    160
+#define VP9_DEC_BORDER_IN_PIXELS    32
+
+typedef struct yv12_buffer_config {
+  int   y_width;
+  int   y_height;
+  int   y_crop_width;
+  int   y_crop_height;
+  int   y_stride;
+
+  int   uv_width;
+  int   uv_height;
+  int   uv_crop_width;
+  int   uv_crop_height;
+  int   uv_stride;
+
+  int   alpha_width;
+  int   alpha_height;
+  int   alpha_stride;
+
+  uint8_t *y_buffer;
+  uint8_t *u_buffer;
+  uint8_t *v_buffer;
+  uint8_t *alpha_buffer;
+
+  uint8_t *buffer_alloc;
+  int buffer_alloc_sz;
+  int border;
+  int frame_size;
+  int subsampling_x;
+  int subsampling_y;
+  unsigned int bit_depth;
+  vpx_color_space_t color_space;
+  vpx_color_range_t color_range;
+  int render_width;
+  int render_height;
+
+  int corrupted;
+  int flags;
+} YV12_BUFFER_CONFIG;
+
+#define YV12_FLAG_HIGHBITDEPTH 8
+
+int vp8_yv12_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
+                                int width, int height, int border);
+int vp8_yv12_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
+                                  int width, int height, int border);
+int vp8_yv12_de_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf);
+
+int vpx_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
+                           int width, int height, int ss_x, int ss_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                           int use_highbitdepth,
+#endif
+                           int border, int byte_alignment);
+
+// Updates the yv12 buffer config with the frame buffer. |byte_alignment| must
+// be a power of 2, from 32 to 1024. 0 sets legacy alignment. If cb is not
+// NULL, then libvpx is using the frame buffer callbacks to handle memory.
+// If cb is not NULL, libvpx will call cb with minimum size in bytes needed
+// to decode the current frame. If cb is NULL, libvpx will allocate memory
+// internally to decode the current frame. Returns 0 on success. Returns < 0
+// on failure.
+int vpx_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
+                             int width, int height, int ss_x, int ss_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                             int use_highbitdepth,
+#endif
+                             int border,
+                             int byte_alignment,
+                             vpx_codec_frame_buffer_t *fb,
+                             vpx_get_frame_buffer_cb_fn_t cb,
+                             void *cb_priv);
+int vpx_free_frame_buffer(YV12_BUFFER_CONFIG *ybf);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // VPX_SCALE_YV12CONFIG_H_
diff --git a/libvpx/libvpx/vpx_util/endian_inl.h b/libvpx/libvpx/vpx_util/endian_inl.h
new file mode 100644
index 0000000..37bdce1
--- /dev/null
+++ b/libvpx/libvpx/vpx_util/endian_inl.h
@@ -0,0 +1,120 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Endian related functions.
+
+#ifndef VPX_UTIL_ENDIAN_INL_H_
+#define VPX_UTIL_ENDIAN_INL_H_
+
+#include <stdlib.h>
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+#if defined(__GNUC__)
+# define LOCAL_GCC_VERSION ((__GNUC__ << 8) | __GNUC_MINOR__)
+# define LOCAL_GCC_PREREQ(maj, min) \
+    (LOCAL_GCC_VERSION >= (((maj) << 8) | (min)))
+#else
+# define LOCAL_GCC_VERSION 0
+# define LOCAL_GCC_PREREQ(maj, min) 0
+#endif
+
+// handle clang compatibility
+#ifndef __has_builtin
+# define __has_builtin(x) 0
+#endif
+
+// some endian fix (e.g.: mips-gcc doesn't define __BIG_ENDIAN__)
+#if !defined(WORDS_BIGENDIAN) && \
+    (defined(__BIG_ENDIAN__) || defined(_M_PPC) || \
+     (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)))
+#define WORDS_BIGENDIAN
+#endif
+
+#if defined(WORDS_BIGENDIAN)
+#define HToLE32 BSwap32
+#define HToLE16 BSwap16
+#define HToBE64(x) (x)
+#define HToBE32(x) (x)
+#else
+#define HToLE32(x) (x)
+#define HToLE16(x) (x)
+#define HToBE64(X) BSwap64(X)
+#define HToBE32(X) BSwap32(X)
+#endif
+
+#if LOCAL_GCC_PREREQ(4, 8) || __has_builtin(__builtin_bswap16)
+#define HAVE_BUILTIN_BSWAP16
+#endif
+
+#if LOCAL_GCC_PREREQ(4, 3) || __has_builtin(__builtin_bswap32)
+#define HAVE_BUILTIN_BSWAP32
+#endif
+
+#if LOCAL_GCC_PREREQ(4, 3) || __has_builtin(__builtin_bswap64)
+#define HAVE_BUILTIN_BSWAP64
+#endif
+
+#if HAVE_MIPS32 && defined(__mips__) && !defined(__mips64) && \
+    defined(__mips_isa_rev) && (__mips_isa_rev >= 2) && (__mips_isa_rev < 6)
+#define VPX_USE_MIPS32_R2
+#endif
+
+static INLINE uint16_t BSwap16(uint16_t x) {
+#if defined(HAVE_BUILTIN_BSWAP16)
+  return __builtin_bswap16(x);
+#elif defined(_MSC_VER)
+  return _byteswap_ushort(x);
+#else
+  // gcc will recognize a 'rorw $8, ...' here:
+  return (x >> 8) | ((x & 0xff) << 8);
+#endif  // HAVE_BUILTIN_BSWAP16
+}
+
+static INLINE uint32_t BSwap32(uint32_t x) {
+#if defined(VPX_USE_MIPS32_R2)
+  uint32_t ret;
+  __asm__ volatile (
+    "wsbh   %[ret], %[x]          \n\t"
+    "rotr   %[ret], %[ret],  16   \n\t"
+    : [ret]"=r"(ret)
+    : [x]"r"(x)
+  );
+  return ret;
+#elif defined(HAVE_BUILTIN_BSWAP32)
+  return __builtin_bswap32(x);
+#elif defined(__i386__) || defined(__x86_64__)
+  uint32_t swapped_bytes;
+  __asm__ volatile("bswap %0" : "=r"(swapped_bytes) : "0"(x));
+  return swapped_bytes;
+#elif defined(_MSC_VER)
+  return (uint32_t)_byteswap_ulong(x);
+#else
+  return (x >> 24) | ((x >> 8) & 0xff00) | ((x << 8) & 0xff0000) | (x << 24);
+#endif  // HAVE_BUILTIN_BSWAP32
+}
+
+static INLINE uint64_t BSwap64(uint64_t x) {
+#if defined(HAVE_BUILTIN_BSWAP64)
+  return __builtin_bswap64(x);
+#elif defined(__x86_64__)
+  uint64_t swapped_bytes;
+  __asm__ volatile("bswapq %0" : "=r"(swapped_bytes) : "0"(x));
+  return swapped_bytes;
+#elif defined(_MSC_VER)
+  return (uint64_t)_byteswap_uint64(x);
+#else  // generic code for swapping 64-bit values (suggested by bdb@)
+  x = ((x & 0xffffffff00000000ull) >> 32) | ((x & 0x00000000ffffffffull) << 32);
+  x = ((x & 0xffff0000ffff0000ull) >> 16) | ((x & 0x0000ffff0000ffffull) << 16);
+  x = ((x & 0xff00ff00ff00ff00ull) >>  8) | ((x & 0x00ff00ff00ff00ffull) <<  8);
+  return x;
+#endif  // HAVE_BUILTIN_BSWAP64
+}
+
+#endif  // VPX_UTIL_ENDIAN_INL_H_
diff --git a/libvpx/libvpx/vpx_util/vpx_thread.c b/libvpx/libvpx/vpx_util/vpx_thread.c
new file mode 100644
index 0000000..0bb0125
--- /dev/null
+++ b/libvpx/libvpx/vpx_util/vpx_thread.c
@@ -0,0 +1,184 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Multi-threaded worker
+//
+// Original source:
+//  http://git.chromium.org/webm/libwebp.git
+//  100644 blob 264210ba2807e4da47eb5d18c04cf869d89b9784  src/utils/thread.c
+
+#include <assert.h>
+#include <string.h>   // for memset()
+#include "./vpx_thread.h"
+#include "vpx_mem/vpx_mem.h"
+
+#if CONFIG_MULTITHREAD
+
+struct VPxWorkerImpl {
+  pthread_mutex_t mutex_;
+  pthread_cond_t  condition_;
+  pthread_t       thread_;
+};
+
+//------------------------------------------------------------------------------
+
+static void execute(VPxWorker *const worker);  // Forward declaration.
+
+static THREADFN thread_loop(void *ptr) {
+  VPxWorker *const worker = (VPxWorker*)ptr;
+  int done = 0;
+  while (!done) {
+    pthread_mutex_lock(&worker->impl_->mutex_);
+    while (worker->status_ == OK) {   // wait in idling mode
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
+    }
+    if (worker->status_ == WORK) {
+      execute(worker);
+      worker->status_ = OK;
+    } else if (worker->status_ == NOT_OK) {   // finish the worker
+      done = 1;
+    }
+    // signal to the main thread that we're done (for sync())
+    pthread_cond_signal(&worker->impl_->condition_);
+    pthread_mutex_unlock(&worker->impl_->mutex_);
+  }
+  return THREAD_RETURN(NULL);    // Thread is finished
+}
+
+// main thread state control
+static void change_state(VPxWorker *const worker,
+                         VPxWorkerStatus new_status) {
+  // No-op when attempting to change state on a thread that didn't come up.
+  // Checking status_ without acquiring the lock first would result in a data
+  // race.
+  if (worker->impl_ == NULL) return;
+
+  pthread_mutex_lock(&worker->impl_->mutex_);
+  if (worker->status_ >= OK) {
+    // wait for the worker to finish
+    while (worker->status_ != OK) {
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
+    }
+    // assign new status and release the working thread if needed
+    if (new_status != OK) {
+      worker->status_ = new_status;
+      pthread_cond_signal(&worker->impl_->condition_);
+    }
+  }
+  pthread_mutex_unlock(&worker->impl_->mutex_);
+}
+
+#endif  // CONFIG_MULTITHREAD
+
+//------------------------------------------------------------------------------
+
+static void init(VPxWorker *const worker) {
+  memset(worker, 0, sizeof(*worker));
+  worker->status_ = NOT_OK;
+}
+
+static int sync(VPxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  change_state(worker, OK);
+#endif
+  assert(worker->status_ <= OK);
+  return !worker->had_error;
+}
+
+static int reset(VPxWorker *const worker) {
+  int ok = 1;
+  worker->had_error = 0;
+  if (worker->status_ < OK) {
+#if CONFIG_MULTITHREAD
+    worker->impl_ = (VPxWorkerImpl*)vpx_calloc(1, sizeof(*worker->impl_));
+    if (worker->impl_ == NULL) {
+      return 0;
+    }
+    if (pthread_mutex_init(&worker->impl_->mutex_, NULL)) {
+      goto Error;
+    }
+    if (pthread_cond_init(&worker->impl_->condition_, NULL)) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      goto Error;
+    }
+    pthread_mutex_lock(&worker->impl_->mutex_);
+    ok = !pthread_create(&worker->impl_->thread_, NULL, thread_loop, worker);
+    if (ok) worker->status_ = OK;
+    pthread_mutex_unlock(&worker->impl_->mutex_);
+    if (!ok) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      pthread_cond_destroy(&worker->impl_->condition_);
+ Error:
+      vpx_free(worker->impl_);
+      worker->impl_ = NULL;
+      return 0;
+    }
+#else
+    worker->status_ = OK;
+#endif
+  } else if (worker->status_ > OK) {
+    ok = sync(worker);
+  }
+  assert(!ok || (worker->status_ == OK));
+  return ok;
+}
+
+static void execute(VPxWorker *const worker) {
+  if (worker->hook != NULL) {
+    worker->had_error |= !worker->hook(worker->data1, worker->data2);
+  }
+}
+
+static void launch(VPxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  change_state(worker, WORK);
+#else
+  execute(worker);
+#endif
+}
+
+static void end(VPxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  if (worker->impl_ != NULL) {
+    change_state(worker, NOT_OK);
+    pthread_join(worker->impl_->thread_, NULL);
+    pthread_mutex_destroy(&worker->impl_->mutex_);
+    pthread_cond_destroy(&worker->impl_->condition_);
+    vpx_free(worker->impl_);
+    worker->impl_ = NULL;
+  }
+#else
+  worker->status_ = NOT_OK;
+  assert(worker->impl_ == NULL);
+#endif
+  assert(worker->status_ == NOT_OK);
+}
+
+//------------------------------------------------------------------------------
+
+static VPxWorkerInterface g_worker_interface = {
+  init, reset, sync, launch, execute, end
+};
+
+int vpx_set_worker_interface(const VPxWorkerInterface* const winterface) {
+  if (winterface == NULL ||
+      winterface->init == NULL || winterface->reset == NULL ||
+      winterface->sync == NULL || winterface->launch == NULL ||
+      winterface->execute == NULL || winterface->end == NULL) {
+    return 0;
+  }
+  g_worker_interface = *winterface;
+  return 1;
+}
+
+const VPxWorkerInterface *vpx_get_worker_interface(void) {
+  return &g_worker_interface;
+}
+
+//------------------------------------------------------------------------------
diff --git a/libvpx/libvpx/vpx_util/vpx_thread.h b/libvpx/libvpx/vpx_util/vpx_thread.h
new file mode 100644
index 0000000..2062abd
--- /dev/null
+++ b/libvpx/libvpx/vpx_util/vpx_thread.h
@@ -0,0 +1,369 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Multi-threaded worker
+//
+// Original source:
+//  http://git.chromium.org/webm/libwebp.git
+//  100644 blob 7bd451b124ae3b81596abfbcc823e3cb129d3a38  src/utils/thread.h
+
+#ifndef VPX_THREAD_H_
+#define VPX_THREAD_H_
+
+#include "./vpx_config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Set maximum decode threads to be 8 due to the limit of frame buffers
+// and not enough semaphores in the emulation layer on windows.
+#define MAX_DECODE_THREADS 8
+
+#if CONFIG_MULTITHREAD
+
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+#include <errno.h>  // NOLINT
+#include <process.h>  // NOLINT
+#include <windows.h>  // NOLINT
+typedef HANDLE pthread_t;
+typedef CRITICAL_SECTION pthread_mutex_t;
+typedef struct {
+  HANDLE waiting_sem_;
+  HANDLE received_sem_;
+  HANDLE signal_event_;
+} pthread_cond_t;
+
+//------------------------------------------------------------------------------
+// simplistic pthread emulation layer
+
+// _beginthreadex requires __stdcall
+#define THREADFN unsigned int __stdcall
+#define THREAD_RETURN(val) (unsigned int)((DWORD_PTR)val)
+
+static INLINE int pthread_create(pthread_t* const thread, const void* attr,
+                                 unsigned int (__stdcall *start)(void*),
+                                 void* arg) {
+  (void)attr;
+  *thread = (pthread_t)_beginthreadex(NULL,   /* void *security */
+                                      0,      /* unsigned stack_size */
+                                      start,
+                                      arg,
+                                      0,      /* unsigned initflag */
+                                      NULL);  /* unsigned *thrdaddr */
+  if (*thread == NULL) return 1;
+  SetThreadPriority(*thread, THREAD_PRIORITY_ABOVE_NORMAL);
+  return 0;
+}
+
+static INLINE int pthread_join(pthread_t thread, void** value_ptr) {
+  (void)value_ptr;
+  return (WaitForSingleObject(thread, INFINITE) != WAIT_OBJECT_0 ||
+          CloseHandle(thread) == 0);
+}
+
+// Mutex
+static INLINE int pthread_mutex_init(pthread_mutex_t *const mutex,
+                                     void* mutexattr) {
+  (void)mutexattr;
+  InitializeCriticalSection(mutex);
+  return 0;
+}
+
+static INLINE int pthread_mutex_trylock(pthread_mutex_t *const mutex) {
+  return TryEnterCriticalSection(mutex) ? 0 : EBUSY;
+}
+
+static INLINE int pthread_mutex_lock(pthread_mutex_t *const mutex) {
+  EnterCriticalSection(mutex);
+  return 0;
+}
+
+static INLINE int pthread_mutex_unlock(pthread_mutex_t *const mutex) {
+  LeaveCriticalSection(mutex);
+  return 0;
+}
+
+static INLINE int pthread_mutex_destroy(pthread_mutex_t *const mutex) {
+  DeleteCriticalSection(mutex);
+  return 0;
+}
+
+// Condition
+static INLINE int pthread_cond_destroy(pthread_cond_t *const condition) {
+  int ok = 1;
+  ok &= (CloseHandle(condition->waiting_sem_) != 0);
+  ok &= (CloseHandle(condition->received_sem_) != 0);
+  ok &= (CloseHandle(condition->signal_event_) != 0);
+  return !ok;
+}
+
+static INLINE int pthread_cond_init(pthread_cond_t *const condition,
+                                    void* cond_attr) {
+  (void)cond_attr;
+  condition->waiting_sem_ = CreateSemaphore(NULL, 0, MAX_DECODE_THREADS, NULL);
+  condition->received_sem_ = CreateSemaphore(NULL, 0, MAX_DECODE_THREADS, NULL);
+  condition->signal_event_ = CreateEvent(NULL, FALSE, FALSE, NULL);
+  if (condition->waiting_sem_ == NULL ||
+      condition->received_sem_ == NULL ||
+      condition->signal_event_ == NULL) {
+    pthread_cond_destroy(condition);
+    return 1;
+  }
+  return 0;
+}
+
+static INLINE int pthread_cond_signal(pthread_cond_t *const condition) {
+  int ok = 1;
+  if (WaitForSingleObject(condition->waiting_sem_, 0) == WAIT_OBJECT_0) {
+    // a thread is waiting in pthread_cond_wait: allow it to be notified
+    ok = SetEvent(condition->signal_event_);
+    // wait until the event is consumed so the signaler cannot consume
+    // the event via its own pthread_cond_wait.
+    ok &= (WaitForSingleObject(condition->received_sem_, INFINITE) !=
+           WAIT_OBJECT_0);
+  }
+  return !ok;
+}
+
+static INLINE int pthread_cond_wait(pthread_cond_t *const condition,
+                                    pthread_mutex_t *const mutex) {
+  int ok;
+  // note that there is a consumer available so the signal isn't dropped in
+  // pthread_cond_signal
+  if (!ReleaseSemaphore(condition->waiting_sem_, 1, NULL))
+    return 1;
+  // now unlock the mutex so pthread_cond_signal may be issued
+  pthread_mutex_unlock(mutex);
+  ok = (WaitForSingleObject(condition->signal_event_, INFINITE) ==
+        WAIT_OBJECT_0);
+  ok &= ReleaseSemaphore(condition->received_sem_, 1, NULL);
+  pthread_mutex_lock(mutex);
+  return !ok;
+}
+#elif defined(__OS2__)
+#define INCL_DOS
+#include <os2.h>    // NOLINT
+
+#include <errno.h>  // NOLINT
+#include <stdlib.h> // NOLINT
+#include <sys/builtin.h> // NOLINT
+
+#define pthread_t TID
+#define pthread_mutex_t HMTX
+
+typedef struct {
+  HEV event_sem_;
+  HEV ack_sem_;
+  volatile unsigned wait_count_;
+} pthread_cond_t;
+
+//------------------------------------------------------------------------------
+// simplistic pthread emulation layer
+
+#define THREADFN void *
+#define THREAD_RETURN(val) (val)
+
+typedef struct {
+  void* (*start_)(void*);
+  void* arg_;
+} thread_arg;
+
+static void thread_start(void* arg) {
+  thread_arg targ = *(thread_arg *)arg;
+  free(arg);
+
+  targ.start_(targ.arg_);
+}
+
+static INLINE int pthread_create(pthread_t* const thread, const void* attr,
+                                 void* (*start)(void*),
+                                 void* arg) {
+  int tid;
+  thread_arg *targ = (thread_arg *)malloc(sizeof(*targ));
+  if (targ == NULL) return 1;
+
+  (void)attr;
+
+  targ->start_ = start;
+  targ->arg_ = arg;
+  tid = (pthread_t)_beginthread(thread_start, NULL, 1024 * 1024, targ);
+  if (tid == -1) {
+    free(targ);
+    return 1;
+  }
+
+  *thread = tid;
+  return 0;
+}
+
+static INLINE int pthread_join(pthread_t thread, void** value_ptr) {
+  (void)value_ptr;
+  return DosWaitThread(&thread, DCWW_WAIT) != 0;
+}
+
+// Mutex
+static INLINE int pthread_mutex_init(pthread_mutex_t *const mutex,
+                                     void* mutexattr) {
+  (void)mutexattr;
+  return DosCreateMutexSem(NULL, mutex, 0, FALSE) != 0;
+}
+
+static INLINE int pthread_mutex_trylock(pthread_mutex_t *const mutex) {
+  return DosRequestMutexSem(*mutex, SEM_IMMEDIATE_RETURN) == 0 ? 0 : EBUSY;
+}
+
+static INLINE int pthread_mutex_lock(pthread_mutex_t *const mutex) {
+  return DosRequestMutexSem(*mutex, SEM_INDEFINITE_WAIT) != 0;
+}
+
+static INLINE int pthread_mutex_unlock(pthread_mutex_t *const mutex) {
+  return DosReleaseMutexSem(*mutex) != 0;
+}
+
+static INLINE int pthread_mutex_destroy(pthread_mutex_t *const mutex) {
+  return DosCloseMutexSem(*mutex) != 0;
+}
+
+// Condition
+static INLINE int pthread_cond_destroy(pthread_cond_t *const condition) {
+  int ok = 1;
+  ok &= DosCloseEventSem(condition->event_sem_) == 0;
+  ok &= DosCloseEventSem(condition->ack_sem_) == 0;
+  return !ok;
+}
+
+static INLINE int pthread_cond_init(pthread_cond_t *const condition,
+                                    void* cond_attr) {
+  int ok = 1;
+  (void)cond_attr;
+
+  ok &= DosCreateEventSem(NULL, &condition->event_sem_, DCE_POSTONE, FALSE)
+          == 0;
+  ok &= DosCreateEventSem(NULL, &condition->ack_sem_, DCE_POSTONE, FALSE) == 0;
+  if (!ok) {
+    pthread_cond_destroy(condition);
+    return 1;
+  }
+  condition->wait_count_ = 0;
+  return 0;
+}
+
+static INLINE int pthread_cond_signal(pthread_cond_t *const condition) {
+  int ok = 1;
+
+  if (!__atomic_cmpxchg32(&condition->wait_count_, 0, 0)) {
+    ok &= DosPostEventSem(condition->event_sem_) == 0;
+    ok &= DosWaitEventSem(condition->ack_sem_, SEM_INDEFINITE_WAIT) == 0;
+  }
+
+  return !ok;
+}
+
+static INLINE int pthread_cond_broadcast(pthread_cond_t *const condition) {
+  int ok = 1;
+
+  while (!__atomic_cmpxchg32(&condition->wait_count_, 0, 0))
+      ok &= pthread_cond_signal(condition) == 0;
+
+  return !ok;
+}
+
+static INLINE int pthread_cond_wait(pthread_cond_t *const condition,
+                                    pthread_mutex_t *const mutex) {
+  int ok = 1;
+
+  __atomic_increment(&condition->wait_count_);
+
+  ok &= pthread_mutex_unlock(mutex) == 0;
+
+  ok &= DosWaitEventSem(condition->event_sem_, SEM_INDEFINITE_WAIT) == 0;
+
+  __atomic_decrement(&condition->wait_count_);
+
+  ok &= DosPostEventSem(condition->ack_sem_) == 0;
+
+  pthread_mutex_lock(mutex);
+
+  return !ok;
+}
+#else  // _WIN32
+#include <pthread.h> // NOLINT
+# define THREADFN void*
+# define THREAD_RETURN(val) val
+#endif
+
+#endif  // CONFIG_MULTITHREAD
+
+// State of the worker thread object
+typedef enum {
+  NOT_OK = 0,   // object is unusable
+  OK,           // ready to work
+  WORK          // busy finishing the current task
+} VPxWorkerStatus;
+
+// Function to be called by the worker thread. Takes two opaque pointers as
+// arguments (data1 and data2), and should return false in case of error.
+typedef int (*VPxWorkerHook)(void*, void*);
+
+// Platform-dependent implementation details for the worker.
+typedef struct VPxWorkerImpl VPxWorkerImpl;
+
+// Synchronization object used to launch job in the worker thread
+typedef struct {
+  VPxWorkerImpl *impl_;
+  VPxWorkerStatus status_;
+  VPxWorkerHook hook;     // hook to call
+  void *data1;            // first argument passed to 'hook'
+  void *data2;            // second argument passed to 'hook'
+  int had_error;          // return value of the last call to 'hook'
+} VPxWorker;
+
+// The interface for all thread-worker related functions. All these functions
+// must be implemented.
+typedef struct {
+  // Must be called first, before any other method.
+  void (*init)(VPxWorker *const worker);
+  // Must be called to initialize the object and spawn the thread. Re-entrant.
+  // Will potentially launch the thread. Returns false in case of error.
+  int (*reset)(VPxWorker *const worker);
+  // Makes sure the previous work is finished. Returns true if worker->had_error
+  // was not set and no error condition was triggered by the working thread.
+  int (*sync)(VPxWorker *const worker);
+  // Triggers the thread to call hook() with data1 and data2 arguments. These
+  // hook/data1/data2 values can be changed at any time before calling this
+  // function, but not be changed afterward until the next call to Sync().
+  void (*launch)(VPxWorker *const worker);
+  // This function is similar to launch() except that it calls the
+  // hook directly instead of using a thread. Convenient to bypass the thread
+  // mechanism while still using the VPxWorker structs. sync() must
+  // still be called afterward (for error reporting).
+  void (*execute)(VPxWorker *const worker);
+  // Kill the thread and terminate the object. To use the object again, one
+  // must call reset() again.
+  void (*end)(VPxWorker *const worker);
+} VPxWorkerInterface;
+
+// Install a new set of threading functions, overriding the defaults. This
+// should be done before any workers are started, i.e., before any encoding or
+// decoding takes place. The contents of the interface struct are copied, it
+// is safe to free the corresponding memory after this call. This function is
+// not thread-safe. Return false in case of invalid pointer or methods.
+int vpx_set_worker_interface(const VPxWorkerInterface *const winterface);
+
+// Retrieve the currently set thread worker interface.
+const VPxWorkerInterface *vpx_get_worker_interface(void);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  // VPX_THREAD_H_
diff --git a/libvpx/libvpx/vpx_util/vpx_util.mk b/libvpx/libvpx/vpx_util/vpx_util.mk
new file mode 100644
index 0000000..c0ef8d3
--- /dev/null
+++ b/libvpx/libvpx/vpx_util/vpx_util.mk
@@ -0,0 +1,14 @@
+##
+## Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+##
+##  Use of this source code is governed by a BSD-style license
+##  that can be found in the LICENSE file in the root of the source
+##  tree. An additional intellectual property rights grant can be found
+##  in the file PATENTS.  All contributing project authors may
+##  be found in the AUTHORS file in the root of the source tree.
+##
+
+UTIL_SRCS-yes += vpx_util.mk
+UTIL_SRCS-yes += vpx_thread.c
+UTIL_SRCS-yes += vpx_thread.h
+UTIL_SRCS-yes += endian_inl.h
diff --git a/libvpx/libvpx/vpxdec.c b/libvpx/libvpx/vpxdec.c
new file mode 100644
index 0000000..dbe64aa
--- /dev/null
+++ b/libvpx/libvpx/vpxdec.c
@@ -0,0 +1,1159 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdarg.h>
+#include <string.h>
+#include <limits.h>
+
+#include "./vpx_config.h"
+
+#if CONFIG_LIBYUV
+#include "third_party/libyuv/include/libyuv/scale.h"
+#endif
+
+#include "./args.h"
+#include "./ivfdec.h"
+
+#include "vpx/vpx_decoder.h"
+#include "vpx_ports/mem_ops.h"
+#include "vpx_ports/vpx_timer.h"
+
+#if CONFIG_VP8_DECODER || CONFIG_VP9_DECODER
+#include "vpx/vp8dx.h"
+#endif
+
+#include "./md5_utils.h"
+
+#include "./tools_common.h"
+#if CONFIG_WEBM_IO
+#include "./webmdec.h"
+#endif
+#include "./y4menc.h"
+
+static const char *exec_name;
+
+struct VpxDecInputContext {
+  struct VpxInputContext *vpx_input_ctx;
+  struct WebmInputContext *webm_ctx;
+};
+
+static const arg_def_t looparg = ARG_DEF(
+    NULL, "loops", 1, "Number of times to decode the file");
+static const arg_def_t codecarg = ARG_DEF(
+    NULL, "codec", 1, "Codec to use");
+static const arg_def_t use_yv12 = ARG_DEF(
+    NULL, "yv12", 0, "Output raw YV12 frames");
+static const arg_def_t use_i420 = ARG_DEF(
+    NULL, "i420", 0, "Output raw I420 frames");
+static const arg_def_t flipuvarg = ARG_DEF(
+    NULL, "flipuv", 0, "Flip the chroma planes in the output");
+static const arg_def_t rawvideo = ARG_DEF(
+    NULL, "rawvideo", 0, "Output raw YUV frames");
+static const arg_def_t noblitarg = ARG_DEF(
+    NULL, "noblit", 0, "Don't process the decoded frames");
+static const arg_def_t progressarg = ARG_DEF(
+    NULL, "progress", 0, "Show progress after each frame decodes");
+static const arg_def_t limitarg = ARG_DEF(
+    NULL, "limit", 1, "Stop decoding after n frames");
+static const arg_def_t skiparg = ARG_DEF(
+    NULL, "skip", 1, "Skip the first n input frames");
+static const arg_def_t postprocarg = ARG_DEF(
+    NULL, "postproc", 0, "Postprocess decoded frames");
+static const arg_def_t summaryarg = ARG_DEF(
+    NULL, "summary", 0, "Show timing summary");
+static const arg_def_t outputfile = ARG_DEF(
+    "o", "output", 1, "Output file name pattern (see below)");
+static const arg_def_t threadsarg = ARG_DEF(
+    "t", "threads", 1, "Max threads to use");
+static const arg_def_t frameparallelarg = ARG_DEF(
+    NULL, "frame-parallel", 0, "Frame parallel decode");
+static const arg_def_t verbosearg = ARG_DEF(
+    "v", "verbose", 0, "Show version string");
+static const arg_def_t error_concealment = ARG_DEF(
+    NULL, "error-concealment", 0, "Enable decoder error-concealment");
+static const arg_def_t scalearg = ARG_DEF(
+    "S", "scale", 0, "Scale output frames uniformly");
+static const arg_def_t continuearg = ARG_DEF(
+    "k", "keep-going", 0, "(debug) Continue decoding after error");
+static const arg_def_t fb_arg = ARG_DEF(
+    NULL, "frame-buffers", 1, "Number of frame buffers to use");
+static const arg_def_t md5arg = ARG_DEF(
+    NULL, "md5", 0, "Compute the MD5 sum of the decoded frame");
+#if CONFIG_VP9_HIGHBITDEPTH
+static const arg_def_t outbitdeptharg = ARG_DEF(
+    NULL, "output-bit-depth", 1, "Output bit-depth for decoded frames");
+#endif
+
+static const arg_def_t *all_args[] = {
+  &codecarg, &use_yv12, &use_i420, &flipuvarg, &rawvideo, &noblitarg,
+  &progressarg, &limitarg, &skiparg, &postprocarg, &summaryarg, &outputfile,
+  &threadsarg, &frameparallelarg, &verbosearg, &scalearg, &fb_arg,
+  &md5arg, &error_concealment, &continuearg,
+#if CONFIG_VP9_HIGHBITDEPTH
+  &outbitdeptharg,
+#endif
+  NULL
+};
+
+#if CONFIG_VP8_DECODER
+static const arg_def_t addnoise_level = ARG_DEF(
+    NULL, "noise-level", 1, "Enable VP8 postproc add noise");
+static const arg_def_t deblock = ARG_DEF(
+    NULL, "deblock", 0, "Enable VP8 deblocking");
+static const arg_def_t demacroblock_level = ARG_DEF(
+    NULL, "demacroblock-level", 1, "Enable VP8 demacroblocking, w/ level");
+static const arg_def_t pp_debug_info = ARG_DEF(
+    NULL, "pp-debug-info", 1, "Enable VP8 visible debug info");
+static const arg_def_t pp_disp_ref_frame = ARG_DEF(
+    NULL, "pp-dbg-ref-frame", 1,
+    "Display only selected reference frame per macro block");
+static const arg_def_t pp_disp_mb_modes = ARG_DEF(
+    NULL, "pp-dbg-mb-modes", 1, "Display only selected macro block modes");
+static const arg_def_t pp_disp_b_modes = ARG_DEF(
+    NULL, "pp-dbg-b-modes", 1, "Display only selected block modes");
+static const arg_def_t pp_disp_mvs = ARG_DEF(
+    NULL, "pp-dbg-mvs", 1, "Draw only selected motion vectors");
+static const arg_def_t mfqe = ARG_DEF(
+    NULL, "mfqe", 0, "Enable multiframe quality enhancement");
+
+static const arg_def_t *vp8_pp_args[] = {
+  &addnoise_level, &deblock, &demacroblock_level, &pp_debug_info,
+  &pp_disp_ref_frame, &pp_disp_mb_modes, &pp_disp_b_modes, &pp_disp_mvs, &mfqe,
+  NULL
+};
+#endif
+
+#if CONFIG_LIBYUV
+static INLINE int libyuv_scale(vpx_image_t *src, vpx_image_t *dst,
+                                  FilterModeEnum mode) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (src->fmt == VPX_IMG_FMT_I42016) {
+    assert(dst->fmt == VPX_IMG_FMT_I42016);
+    return I420Scale_16((uint16_t*)src->planes[VPX_PLANE_Y],
+                        src->stride[VPX_PLANE_Y]/2,
+                        (uint16_t*)src->planes[VPX_PLANE_U],
+                        src->stride[VPX_PLANE_U]/2,
+                        (uint16_t*)src->planes[VPX_PLANE_V],
+                        src->stride[VPX_PLANE_V]/2,
+                        src->d_w, src->d_h,
+                        (uint16_t*)dst->planes[VPX_PLANE_Y],
+                        dst->stride[VPX_PLANE_Y]/2,
+                        (uint16_t*)dst->planes[VPX_PLANE_U],
+                        dst->stride[VPX_PLANE_U]/2,
+                        (uint16_t*)dst->planes[VPX_PLANE_V],
+                        dst->stride[VPX_PLANE_V]/2,
+                        dst->d_w, dst->d_h,
+                        mode);
+  }
+#endif
+  assert(src->fmt == VPX_IMG_FMT_I420);
+  assert(dst->fmt == VPX_IMG_FMT_I420);
+  return I420Scale(src->planes[VPX_PLANE_Y], src->stride[VPX_PLANE_Y],
+                   src->planes[VPX_PLANE_U], src->stride[VPX_PLANE_U],
+                   src->planes[VPX_PLANE_V], src->stride[VPX_PLANE_V],
+                   src->d_w, src->d_h,
+                   dst->planes[VPX_PLANE_Y], dst->stride[VPX_PLANE_Y],
+                   dst->planes[VPX_PLANE_U], dst->stride[VPX_PLANE_U],
+                   dst->planes[VPX_PLANE_V], dst->stride[VPX_PLANE_V],
+                   dst->d_w, dst->d_h,
+                   mode);
+}
+#endif
+
+void usage_exit(void) {
+  int i;
+
+  fprintf(stderr, "Usage: %s <options> filename\n\n"
+          "Options:\n", exec_name);
+  arg_show_usage(stderr, all_args);
+#if CONFIG_VP8_DECODER
+  fprintf(stderr, "\nVP8 Postprocessing Options:\n");
+  arg_show_usage(stderr, vp8_pp_args);
+#endif
+  fprintf(stderr,
+          "\nOutput File Patterns:\n\n"
+          "  The -o argument specifies the name of the file(s) to "
+          "write to. If the\n  argument does not include any escape "
+          "characters, the output will be\n  written to a single file. "
+          "Otherwise, the filename will be calculated by\n  expanding "
+          "the following escape characters:\n");
+  fprintf(stderr,
+          "\n\t%%w   - Frame width"
+          "\n\t%%h   - Frame height"
+          "\n\t%%<n> - Frame number, zero padded to <n> places (1..9)"
+          "\n\n  Pattern arguments are only supported in conjunction "
+          "with the --yv12 and\n  --i420 options. If the -o option is "
+          "not specified, the output will be\n  directed to stdout.\n"
+         );
+  fprintf(stderr, "\nIncluded decoders:\n\n");
+
+  for (i = 0; i < get_vpx_decoder_count(); ++i) {
+    const VpxInterface *const decoder = get_vpx_decoder_by_index(i);
+    fprintf(stderr, "    %-6s - %s\n",
+            decoder->name, vpx_codec_iface_name(decoder->codec_interface()));
+  }
+
+  exit(EXIT_FAILURE);
+}
+
+static int raw_read_frame(FILE *infile, uint8_t **buffer,
+                          size_t *bytes_read, size_t *buffer_size) {
+  char raw_hdr[RAW_FRAME_HDR_SZ];
+  size_t frame_size = 0;
+
+  if (fread(raw_hdr, RAW_FRAME_HDR_SZ, 1, infile) != 1) {
+    if (!feof(infile))
+      warn("Failed to read RAW frame size\n");
+  } else {
+    const size_t kCorruptFrameThreshold = 256 * 1024 * 1024;
+    const size_t kFrameTooSmallThreshold = 256 * 1024;
+    frame_size = mem_get_le32(raw_hdr);
+
+    if (frame_size > kCorruptFrameThreshold) {
+      warn("Read invalid frame size (%u)\n", (unsigned int)frame_size);
+      frame_size = 0;
+    }
+
+    if (frame_size < kFrameTooSmallThreshold) {
+      warn("Warning: Read invalid frame size (%u) - not a raw file?\n",
+           (unsigned int)frame_size);
+    }
+
+    if (frame_size > *buffer_size) {
+      uint8_t *new_buf = realloc(*buffer, 2 * frame_size);
+      if (new_buf) {
+        *buffer = new_buf;
+        *buffer_size = 2 * frame_size;
+      } else {
+        warn("Failed to allocate compressed data buffer\n");
+        frame_size = 0;
+      }
+    }
+  }
+
+  if (!feof(infile)) {
+    if (fread(*buffer, 1, frame_size, infile) != frame_size) {
+      warn("Failed to read full frame\n");
+      return 1;
+    }
+    *bytes_read = frame_size;
+  }
+
+  return 0;
+}
+
+static int read_frame(struct VpxDecInputContext *input, uint8_t **buf,
+                      size_t *bytes_in_buffer, size_t *buffer_size) {
+  switch (input->vpx_input_ctx->file_type) {
+#if CONFIG_WEBM_IO
+    case FILE_TYPE_WEBM:
+      return webm_read_frame(input->webm_ctx, buf, bytes_in_buffer);
+#endif
+    case FILE_TYPE_RAW:
+      return raw_read_frame(input->vpx_input_ctx->file,
+                            buf, bytes_in_buffer, buffer_size);
+    case FILE_TYPE_IVF:
+      return ivf_read_frame(input->vpx_input_ctx->file,
+                            buf, bytes_in_buffer, buffer_size);
+    default:
+      return 1;
+  }
+}
+
+static void update_image_md5(const vpx_image_t *img, const int planes[3],
+                             MD5Context *md5) {
+  int i, y;
+
+  for (i = 0; i < 3; ++i) {
+    const int plane = planes[i];
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int w = vpx_img_plane_width(img, plane) *
+                ((img->fmt & VPX_IMG_FMT_HIGHBITDEPTH) ? 2 : 1);
+    const int h = vpx_img_plane_height(img, plane);
+
+    for (y = 0; y < h; ++y) {
+      MD5Update(md5, buf, w);
+      buf += stride;
+    }
+  }
+}
+
+static void write_image_file(const vpx_image_t *img, const int planes[3],
+                             FILE *file) {
+  int i, y;
+#if CONFIG_VP9_HIGHBITDEPTH
+  const int bytes_per_sample = ((img->fmt & VPX_IMG_FMT_HIGHBITDEPTH) ? 2 : 1);
+#else
+  const int bytes_per_sample = 1;
+#endif
+
+  for (i = 0; i < 3; ++i) {
+    const int plane = planes[i];
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int w = vpx_img_plane_width(img, plane);
+    const int h = vpx_img_plane_height(img, plane);
+
+    for (y = 0; y < h; ++y) {
+      fwrite(buf, bytes_per_sample, w, file);
+      buf += stride;
+    }
+  }
+}
+
+static int file_is_raw(struct VpxInputContext *input) {
+  uint8_t buf[32];
+  int is_raw = 0;
+  vpx_codec_stream_info_t si;
+
+  si.sz = sizeof(si);
+
+  if (fread(buf, 1, 32, input->file) == 32) {
+    int i;
+
+    if (mem_get_le32(buf) < 256 * 1024 * 1024) {
+      for (i = 0; i < get_vpx_decoder_count(); ++i) {
+        const VpxInterface *const decoder = get_vpx_decoder_by_index(i);
+        if (!vpx_codec_peek_stream_info(decoder->codec_interface(),
+                                        buf + 4, 32 - 4, &si)) {
+          is_raw = 1;
+          input->fourcc = decoder->fourcc;
+          input->width = si.w;
+          input->height = si.h;
+          input->framerate.numerator = 30;
+          input->framerate.denominator = 1;
+          break;
+        }
+      }
+    }
+  }
+
+  rewind(input->file);
+  return is_raw;
+}
+
+static void show_progress(int frame_in, int frame_out, uint64_t dx_time) {
+  fprintf(stderr,
+          "%d decoded frames/%d showed frames in %"PRId64" us (%.2f fps)\r",
+          frame_in, frame_out, dx_time,
+          (double)frame_out * 1000000.0 / (double)dx_time);
+}
+
+struct ExternalFrameBuffer {
+  uint8_t* data;
+  size_t size;
+  int in_use;
+};
+
+struct ExternalFrameBufferList {
+  int num_external_frame_buffers;
+  struct ExternalFrameBuffer *ext_fb;
+};
+
+// Callback used by libvpx to request an external frame buffer. |cb_priv|
+// Application private data passed into the set function. |min_size| is the
+// minimum size in bytes needed to decode the next frame. |fb| pointer to the
+// frame buffer.
+static int get_vp9_frame_buffer(void *cb_priv, size_t min_size,
+                                vpx_codec_frame_buffer_t *fb) {
+  int i;
+  struct ExternalFrameBufferList *const ext_fb_list =
+      (struct ExternalFrameBufferList *)cb_priv;
+  if (ext_fb_list == NULL)
+    return -1;
+
+  // Find a free frame buffer.
+  for (i = 0; i < ext_fb_list->num_external_frame_buffers; ++i) {
+    if (!ext_fb_list->ext_fb[i].in_use)
+      break;
+  }
+
+  if (i == ext_fb_list->num_external_frame_buffers)
+    return -1;
+
+  if (ext_fb_list->ext_fb[i].size < min_size) {
+    free(ext_fb_list->ext_fb[i].data);
+    ext_fb_list->ext_fb[i].data = (uint8_t *)calloc(min_size, sizeof(uint8_t));
+    if (!ext_fb_list->ext_fb[i].data)
+      return -1;
+
+    ext_fb_list->ext_fb[i].size = min_size;
+  }
+
+  fb->data = ext_fb_list->ext_fb[i].data;
+  fb->size = ext_fb_list->ext_fb[i].size;
+  ext_fb_list->ext_fb[i].in_use = 1;
+
+  // Set the frame buffer's private data to point at the external frame buffer.
+  fb->priv = &ext_fb_list->ext_fb[i];
+  return 0;
+}
+
+// Callback used by libvpx when there are no references to the frame buffer.
+// |cb_priv| user private data passed into the set function. |fb| pointer
+// to the frame buffer.
+static int release_vp9_frame_buffer(void *cb_priv,
+                                    vpx_codec_frame_buffer_t *fb) {
+  struct ExternalFrameBuffer *const ext_fb =
+      (struct ExternalFrameBuffer *)fb->priv;
+  (void)cb_priv;
+  ext_fb->in_use = 0;
+  return 0;
+}
+
+static void generate_filename(const char *pattern, char *out, size_t q_len,
+                              unsigned int d_w, unsigned int d_h,
+                              unsigned int frame_in) {
+  const char *p = pattern;
+  char *q = out;
+
+  do {
+    char *next_pat = strchr(p, '%');
+
+    if (p == next_pat) {
+      size_t pat_len;
+
+      /* parse the pattern */
+      q[q_len - 1] = '\0';
+      switch (p[1]) {
+        case 'w':
+          snprintf(q, q_len - 1, "%d", d_w);
+          break;
+        case 'h':
+          snprintf(q, q_len - 1, "%d", d_h);
+          break;
+        case '1':
+          snprintf(q, q_len - 1, "%d", frame_in);
+          break;
+        case '2':
+          snprintf(q, q_len - 1, "%02d", frame_in);
+          break;
+        case '3':
+          snprintf(q, q_len - 1, "%03d", frame_in);
+          break;
+        case '4':
+          snprintf(q, q_len - 1, "%04d", frame_in);
+          break;
+        case '5':
+          snprintf(q, q_len - 1, "%05d", frame_in);
+          break;
+        case '6':
+          snprintf(q, q_len - 1, "%06d", frame_in);
+          break;
+        case '7':
+          snprintf(q, q_len - 1, "%07d", frame_in);
+          break;
+        case '8':
+          snprintf(q, q_len - 1, "%08d", frame_in);
+          break;
+        case '9':
+          snprintf(q, q_len - 1, "%09d", frame_in);
+          break;
+        default:
+          die("Unrecognized pattern %%%c\n", p[1]);
+          break;
+      }
+
+      pat_len = strlen(q);
+      if (pat_len >= q_len - 1)
+        die("Output filename too long.\n");
+      q += pat_len;
+      p += 2;
+      q_len -= pat_len;
+    } else {
+      size_t copy_len;
+
+      /* copy the next segment */
+      if (!next_pat)
+        copy_len = strlen(p);
+      else
+        copy_len = next_pat - p;
+
+      if (copy_len >= q_len - 1)
+        die("Output filename too long.\n");
+
+      memcpy(q, p, copy_len);
+      q[copy_len] = '\0';
+      q += copy_len;
+      p += copy_len;
+      q_len -= copy_len;
+    }
+  } while (*p);
+}
+
+static int is_single_file(const char *outfile_pattern) {
+  const char *p = outfile_pattern;
+
+  do {
+    p = strchr(p, '%');
+    if (p && p[1] >= '1' && p[1] <= '9')
+      return 0;  // pattern contains sequence number, so it's not unique
+    if (p)
+      p++;
+  } while (p);
+
+  return 1;
+}
+
+static void print_md5(unsigned char digest[16], const char *filename) {
+  int i;
+
+  for (i = 0; i < 16; ++i)
+    printf("%02x", digest[i]);
+  printf("  %s\n", filename);
+}
+
+static FILE *open_outfile(const char *name) {
+  if (strcmp("-", name) == 0) {
+    set_binary_mode(stdout);
+    return stdout;
+  } else {
+    FILE *file = fopen(name, "wb");
+    if (!file)
+      fatal("Failed to open output file '%s'", name);
+    return file;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static int img_shifted_realloc_required(const vpx_image_t *img,
+                                        const vpx_image_t *shifted,
+                                        vpx_img_fmt_t required_fmt) {
+  return img->d_w != shifted->d_w ||
+         img->d_h != shifted->d_h ||
+         required_fmt != shifted->fmt;
+}
+#endif
+
+static int main_loop(int argc, const char **argv_) {
+  vpx_codec_ctx_t       decoder;
+  char                  *fn = NULL;
+  int                    i;
+  uint8_t               *buf = NULL;
+  size_t                 bytes_in_buffer = 0, buffer_size = 0;
+  FILE                  *infile;
+  int                    frame_in = 0, frame_out = 0, flipuv = 0, noblit = 0;
+  int                    do_md5 = 0, progress = 0, frame_parallel = 0;
+  int                    stop_after = 0, postproc = 0, summary = 0, quiet = 1;
+  int                    arg_skip = 0;
+  int                    ec_enabled = 0;
+  int                    keep_going = 0;
+  const VpxInterface *interface = NULL;
+  const VpxInterface *fourcc_interface = NULL;
+  uint64_t dx_time = 0;
+  struct arg               arg;
+  char                   **argv, **argi, **argj;
+
+  int                     single_file;
+  int                     use_y4m = 1;
+  int                     opt_yv12 = 0;
+  int                     opt_i420 = 0;
+  vpx_codec_dec_cfg_t     cfg = {0, 0, 0};
+#if CONFIG_VP9_HIGHBITDEPTH
+  unsigned int            output_bit_depth = 0;
+#endif
+#if CONFIG_VP8_DECODER
+  vp8_postproc_cfg_t      vp8_pp_cfg = {0};
+  int                     vp8_dbg_color_ref_frame = 0;
+  int                     vp8_dbg_color_mb_modes = 0;
+  int                     vp8_dbg_color_b_modes = 0;
+  int                     vp8_dbg_display_mv = 0;
+#endif
+  int                     frames_corrupted = 0;
+  int                     dec_flags = 0;
+  int                     do_scale = 0;
+  vpx_image_t             *scaled_img = NULL;
+#if CONFIG_VP9_HIGHBITDEPTH
+  vpx_image_t             *img_shifted = NULL;
+#endif
+  int                     frame_avail, got_data, flush_decoder = 0;
+  int                     num_external_frame_buffers = 0;
+  struct ExternalFrameBufferList ext_fb_list = {0, NULL};
+
+  const char *outfile_pattern = NULL;
+  char outfile_name[PATH_MAX] = {0};
+  FILE *outfile = NULL;
+
+  MD5Context md5_ctx;
+  unsigned char md5_digest[16];
+
+  struct VpxDecInputContext input = {NULL, NULL};
+  struct VpxInputContext vpx_input_ctx;
+#if CONFIG_WEBM_IO
+  struct WebmInputContext webm_ctx;
+  memset(&(webm_ctx), 0, sizeof(webm_ctx));
+  input.webm_ctx = &webm_ctx;
+#endif
+  input.vpx_input_ctx = &vpx_input_ctx;
+
+  /* Parse command line */
+  exec_name = argv_[0];
+  argv = argv_dup(argc - 1, argv_ + 1);
+
+  for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
+    memset(&arg, 0, sizeof(arg));
+    arg.argv_step = 1;
+
+    if (arg_match(&arg, &codecarg, argi)) {
+      interface = get_vpx_decoder_by_name(arg.val);
+      if (!interface)
+        die("Error: Unrecognized argument (%s) to --codec\n", arg.val);
+    } else if (arg_match(&arg, &looparg, argi)) {
+      // no-op
+    } else if (arg_match(&arg, &outputfile, argi))
+      outfile_pattern = arg.val;
+    else if (arg_match(&arg, &use_yv12, argi)) {
+      use_y4m = 0;
+      flipuv = 1;
+      opt_yv12 = 1;
+    } else if (arg_match(&arg, &use_i420, argi)) {
+      use_y4m = 0;
+      flipuv = 0;
+      opt_i420 = 1;
+    } else if (arg_match(&arg, &rawvideo, argi)) {
+      use_y4m = 0;
+    } else if (arg_match(&arg, &flipuvarg, argi))
+      flipuv = 1;
+    else if (arg_match(&arg, &noblitarg, argi))
+      noblit = 1;
+    else if (arg_match(&arg, &progressarg, argi))
+      progress = 1;
+    else if (arg_match(&arg, &limitarg, argi))
+      stop_after = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &skiparg, argi))
+      arg_skip = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &postprocarg, argi))
+      postproc = 1;
+    else if (arg_match(&arg, &md5arg, argi))
+      do_md5 = 1;
+    else if (arg_match(&arg, &summaryarg, argi))
+      summary = 1;
+    else if (arg_match(&arg, &threadsarg, argi))
+      cfg.threads = arg_parse_uint(&arg);
+#if CONFIG_VP9_DECODER
+    else if (arg_match(&arg, &frameparallelarg, argi))
+      frame_parallel = 1;
+#endif
+    else if (arg_match(&arg, &verbosearg, argi))
+      quiet = 0;
+    else if (arg_match(&arg, &scalearg, argi))
+      do_scale = 1;
+    else if (arg_match(&arg, &fb_arg, argi))
+      num_external_frame_buffers = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &continuearg, argi))
+      keep_going = 1;
+#if CONFIG_VP9_HIGHBITDEPTH
+    else if (arg_match(&arg, &outbitdeptharg, argi)) {
+      output_bit_depth = arg_parse_uint(&arg);
+    }
+#endif
+#if CONFIG_VP8_DECODER
+    else if (arg_match(&arg, &addnoise_level, argi)) {
+      postproc = 1;
+      vp8_pp_cfg.post_proc_flag |= VP8_ADDNOISE;
+      vp8_pp_cfg.noise_level = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &demacroblock_level, argi)) {
+      postproc = 1;
+      vp8_pp_cfg.post_proc_flag |= VP8_DEMACROBLOCK;
+      vp8_pp_cfg.deblocking_level = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &deblock, argi)) {
+      postproc = 1;
+      vp8_pp_cfg.post_proc_flag |= VP8_DEBLOCK;
+    } else if (arg_match(&arg, &mfqe, argi)) {
+      postproc = 1;
+      vp8_pp_cfg.post_proc_flag |= VP8_MFQE;
+    } else if (arg_match(&arg, &pp_debug_info, argi)) {
+      unsigned int level = arg_parse_uint(&arg);
+
+      postproc = 1;
+      vp8_pp_cfg.post_proc_flag &= ~0x7;
+
+      if (level)
+        vp8_pp_cfg.post_proc_flag |= level;
+    } else if (arg_match(&arg, &pp_disp_ref_frame, argi)) {
+      unsigned int flags = arg_parse_int(&arg);
+      if (flags) {
+        postproc = 1;
+        vp8_dbg_color_ref_frame = flags;
+      }
+    } else if (arg_match(&arg, &pp_disp_mb_modes, argi)) {
+      unsigned int flags = arg_parse_int(&arg);
+      if (flags) {
+        postproc = 1;
+        vp8_dbg_color_mb_modes = flags;
+      }
+    } else if (arg_match(&arg, &pp_disp_b_modes, argi)) {
+      unsigned int flags = arg_parse_int(&arg);
+      if (flags) {
+        postproc = 1;
+        vp8_dbg_color_b_modes = flags;
+      }
+    } else if (arg_match(&arg, &pp_disp_mvs, argi)) {
+      unsigned int flags = arg_parse_int(&arg);
+      if (flags) {
+        postproc = 1;
+        vp8_dbg_display_mv = flags;
+      }
+    } else if (arg_match(&arg, &error_concealment, argi)) {
+      ec_enabled = 1;
+    }
+#endif  // CONFIG_VP8_DECODER
+    else
+      argj++;
+  }
+
+  /* Check for unrecognized options */
+  for (argi = argv; *argi; argi++)
+    if (argi[0][0] == '-' && strlen(argi[0]) > 1)
+      die("Error: Unrecognized option %s\n", *argi);
+
+  /* Handle non-option arguments */
+  fn = argv[0];
+
+  if (!fn) {
+    free(argv);
+    usage_exit();
+  }
+  /* Open file */
+  infile = strcmp(fn, "-") ? fopen(fn, "rb") : set_binary_mode(stdin);
+
+  if (!infile) {
+    fatal("Failed to open input file '%s'", strcmp(fn, "-") ? fn : "stdin");
+  }
+#if CONFIG_OS_SUPPORT
+  /* Make sure we don't dump to the terminal, unless forced to with -o - */
+  if (!outfile_pattern && isatty(fileno(stdout)) && !do_md5 && !noblit) {
+    fprintf(stderr,
+            "Not dumping raw video to your terminal. Use '-o -' to "
+            "override.\n");
+    return EXIT_FAILURE;
+  }
+#endif
+  input.vpx_input_ctx->file = infile;
+  if (file_is_ivf(input.vpx_input_ctx))
+    input.vpx_input_ctx->file_type = FILE_TYPE_IVF;
+#if CONFIG_WEBM_IO
+  else if (file_is_webm(input.webm_ctx, input.vpx_input_ctx))
+    input.vpx_input_ctx->file_type = FILE_TYPE_WEBM;
+#endif
+  else if (file_is_raw(input.vpx_input_ctx))
+    input.vpx_input_ctx->file_type = FILE_TYPE_RAW;
+  else {
+    fprintf(stderr, "Unrecognized input file type.\n");
+#if !CONFIG_WEBM_IO
+    fprintf(stderr, "vpxdec was built without WebM container support.\n");
+#endif
+    return EXIT_FAILURE;
+  }
+
+  outfile_pattern = outfile_pattern ? outfile_pattern : "-";
+  single_file = is_single_file(outfile_pattern);
+
+  if (!noblit && single_file) {
+    generate_filename(outfile_pattern, outfile_name, PATH_MAX,
+                      vpx_input_ctx.width, vpx_input_ctx.height, 0);
+    if (do_md5)
+      MD5Init(&md5_ctx);
+    else
+      outfile = open_outfile(outfile_name);
+  }
+
+  if (use_y4m && !noblit) {
+    if (!single_file) {
+      fprintf(stderr, "YUV4MPEG2 not supported with output patterns,"
+              " try --i420 or --yv12 or --rawvideo.\n");
+      return EXIT_FAILURE;
+    }
+
+#if CONFIG_WEBM_IO
+    if (vpx_input_ctx.file_type == FILE_TYPE_WEBM) {
+      if (webm_guess_framerate(input.webm_ctx, input.vpx_input_ctx)) {
+        fprintf(stderr, "Failed to guess framerate -- error parsing "
+                "webm file?\n");
+        return EXIT_FAILURE;
+      }
+    }
+#endif
+  }
+
+  fourcc_interface = get_vpx_decoder_by_fourcc(vpx_input_ctx.fourcc);
+  if (interface && fourcc_interface && interface != fourcc_interface)
+    warn("Header indicates codec: %s\n", fourcc_interface->name);
+  else
+    interface = fourcc_interface;
+
+  if (!interface)
+    interface = get_vpx_decoder_by_index(0);
+
+  dec_flags = (postproc ? VPX_CODEC_USE_POSTPROC : 0) |
+              (ec_enabled ? VPX_CODEC_USE_ERROR_CONCEALMENT : 0) |
+              (frame_parallel ? VPX_CODEC_USE_FRAME_THREADING : 0);
+  if (vpx_codec_dec_init(&decoder, interface->codec_interface(),
+                         &cfg, dec_flags)) {
+    fprintf(stderr, "Failed to initialize decoder: %s\n",
+            vpx_codec_error(&decoder));
+    return EXIT_FAILURE;
+  }
+
+  if (!quiet)
+    fprintf(stderr, "%s\n", decoder.name);
+
+#if CONFIG_VP8_DECODER
+  if (vp8_pp_cfg.post_proc_flag
+      && vpx_codec_control(&decoder, VP8_SET_POSTPROC, &vp8_pp_cfg)) {
+    fprintf(stderr, "Failed to configure postproc: %s\n",
+            vpx_codec_error(&decoder));
+    return EXIT_FAILURE;
+  }
+
+  if (vp8_dbg_color_ref_frame
+      && vpx_codec_control(&decoder, VP8_SET_DBG_COLOR_REF_FRAME,
+                           vp8_dbg_color_ref_frame)) {
+    fprintf(stderr, "Failed to configure reference block visualizer: %s\n",
+            vpx_codec_error(&decoder));
+    return EXIT_FAILURE;
+  }
+
+  if (vp8_dbg_color_mb_modes
+      && vpx_codec_control(&decoder, VP8_SET_DBG_COLOR_MB_MODES,
+                           vp8_dbg_color_mb_modes)) {
+    fprintf(stderr, "Failed to configure macro block visualizer: %s\n",
+            vpx_codec_error(&decoder));
+    return EXIT_FAILURE;
+  }
+
+  if (vp8_dbg_color_b_modes
+      && vpx_codec_control(&decoder, VP8_SET_DBG_COLOR_B_MODES,
+                           vp8_dbg_color_b_modes)) {
+    fprintf(stderr, "Failed to configure block visualizer: %s\n",
+            vpx_codec_error(&decoder));
+    return EXIT_FAILURE;
+  }
+
+  if (vp8_dbg_display_mv
+      && vpx_codec_control(&decoder, VP8_SET_DBG_DISPLAY_MV,
+                           vp8_dbg_display_mv)) {
+    fprintf(stderr, "Failed to configure motion vector visualizer: %s\n",
+            vpx_codec_error(&decoder));
+    return EXIT_FAILURE;
+  }
+#endif
+
+
+  if (arg_skip)
+    fprintf(stderr, "Skipping first %d frames.\n", arg_skip);
+  while (arg_skip) {
+    if (read_frame(&input, &buf, &bytes_in_buffer, &buffer_size))
+      break;
+    arg_skip--;
+  }
+
+  if (num_external_frame_buffers > 0) {
+    ext_fb_list.num_external_frame_buffers = num_external_frame_buffers;
+    ext_fb_list.ext_fb = (struct ExternalFrameBuffer *)calloc(
+        num_external_frame_buffers, sizeof(*ext_fb_list.ext_fb));
+    if (vpx_codec_set_frame_buffer_functions(
+            &decoder, get_vp9_frame_buffer, release_vp9_frame_buffer,
+            &ext_fb_list)) {
+      fprintf(stderr, "Failed to configure external frame buffers: %s\n",
+              vpx_codec_error(&decoder));
+      return EXIT_FAILURE;
+    }
+  }
+
+  frame_avail = 1;
+  got_data = 0;
+
+  /* Decode file */
+  while (frame_avail || got_data) {
+    vpx_codec_iter_t  iter = NULL;
+    vpx_image_t    *img;
+    struct vpx_usec_timer timer;
+    int                   corrupted = 0;
+
+    frame_avail = 0;
+    if (!stop_after || frame_in < stop_after) {
+      if (!read_frame(&input, &buf, &bytes_in_buffer, &buffer_size)) {
+        frame_avail = 1;
+        frame_in++;
+
+        vpx_usec_timer_start(&timer);
+
+        if (vpx_codec_decode(&decoder, buf, (unsigned int)bytes_in_buffer,
+                             NULL, 0)) {
+          const char *detail = vpx_codec_error_detail(&decoder);
+          warn("Failed to decode frame %d: %s",
+               frame_in, vpx_codec_error(&decoder));
+
+          if (detail)
+            warn("Additional information: %s", detail);
+          if (!keep_going)
+            goto fail;
+        }
+
+        vpx_usec_timer_mark(&timer);
+        dx_time += vpx_usec_timer_elapsed(&timer);
+      } else {
+        flush_decoder = 1;
+      }
+    } else {
+      flush_decoder = 1;
+    }
+
+    vpx_usec_timer_start(&timer);
+
+    if (flush_decoder) {
+      // Flush the decoder in frame parallel decode.
+      if (vpx_codec_decode(&decoder, NULL, 0, NULL, 0)) {
+        warn("Failed to flush decoder: %s", vpx_codec_error(&decoder));
+      }
+    }
+
+    got_data = 0;
+    if ((img = vpx_codec_get_frame(&decoder, &iter))) {
+      ++frame_out;
+      got_data = 1;
+    }
+
+    vpx_usec_timer_mark(&timer);
+    dx_time += (unsigned int)vpx_usec_timer_elapsed(&timer);
+
+    if (!frame_parallel &&
+        vpx_codec_control(&decoder, VP8D_GET_FRAME_CORRUPTED, &corrupted)) {
+      warn("Failed VP8_GET_FRAME_CORRUPTED: %s", vpx_codec_error(&decoder));
+      if (!keep_going)
+        goto fail;
+    }
+    frames_corrupted += corrupted;
+
+    if (progress)
+      show_progress(frame_in, frame_out, dx_time);
+
+    if (!noblit && img) {
+      const int PLANES_YUV[] = {VPX_PLANE_Y, VPX_PLANE_U, VPX_PLANE_V};
+      const int PLANES_YVU[] = {VPX_PLANE_Y, VPX_PLANE_V, VPX_PLANE_U};
+      const int *planes = flipuv ? PLANES_YVU : PLANES_YUV;
+
+      if (do_scale) {
+        if (frame_out == 1) {
+          // If the output frames are to be scaled to a fixed display size then
+          // use the width and height specified in the container. If either of
+          // these is set to 0, use the display size set in the first frame
+          // header. If that is unavailable, use the raw decoded size of the
+          // first decoded frame.
+          int render_width = vpx_input_ctx.width;
+          int render_height = vpx_input_ctx.height;
+          if (!render_width || !render_height) {
+            int render_size[2];
+            if (vpx_codec_control(&decoder, VP9D_GET_DISPLAY_SIZE,
+                                  render_size)) {
+              // As last resort use size of first frame as display size.
+              render_width = img->d_w;
+              render_height = img->d_h;
+            } else {
+              render_width = render_size[0];
+              render_height = render_size[1];
+            }
+          }
+          scaled_img = vpx_img_alloc(NULL, img->fmt, render_width,
+                                     render_height, 16);
+          scaled_img->bit_depth = img->bit_depth;
+        }
+
+        if (img->d_w != scaled_img->d_w || img->d_h != scaled_img->d_h) {
+#if CONFIG_LIBYUV
+          libyuv_scale(img, scaled_img, kFilterBox);
+          img = scaled_img;
+#else
+          fprintf(stderr, "Failed  to scale output frame: %s.\n"
+                  "Scaling is disabled in this configuration. "
+                  "To enable scaling, configure with --enable-libyuv\n",
+                  vpx_codec_error(&decoder));
+          return EXIT_FAILURE;
+#endif
+        }
+      }
+#if CONFIG_VP9_HIGHBITDEPTH
+      // Default to codec bit depth if output bit depth not set
+      if (!output_bit_depth && single_file && !do_md5) {
+        output_bit_depth = img->bit_depth;
+      }
+      // Shift up or down if necessary
+      if (output_bit_depth != 0 && output_bit_depth != img->bit_depth) {
+        const vpx_img_fmt_t shifted_fmt = output_bit_depth == 8 ?
+            img->fmt ^ (img->fmt & VPX_IMG_FMT_HIGHBITDEPTH) :
+            img->fmt | VPX_IMG_FMT_HIGHBITDEPTH;
+        if (img_shifted &&
+            img_shifted_realloc_required(img, img_shifted, shifted_fmt)) {
+          vpx_img_free(img_shifted);
+          img_shifted = NULL;
+        }
+        if (!img_shifted) {
+          img_shifted = vpx_img_alloc(NULL, shifted_fmt,
+                                      img->d_w, img->d_h, 16);
+          img_shifted->bit_depth = output_bit_depth;
+        }
+        if (output_bit_depth > img->bit_depth) {
+          vpx_img_upshift(img_shifted, img,
+                          output_bit_depth - img->bit_depth);
+        } else {
+          vpx_img_downshift(img_shifted, img,
+                            img->bit_depth - output_bit_depth);
+        }
+        img = img_shifted;
+      }
+#endif
+
+      if (single_file) {
+        if (use_y4m) {
+          char buf[Y4M_BUFFER_SIZE] = {0};
+          size_t len = 0;
+          if (img->fmt == VPX_IMG_FMT_I440 || img->fmt == VPX_IMG_FMT_I44016) {
+            fprintf(stderr, "Cannot produce y4m output for 440 sampling.\n");
+            goto fail;
+          }
+          if (frame_out == 1) {
+            // Y4M file header
+            len = y4m_write_file_header(buf, sizeof(buf),
+                                        vpx_input_ctx.width,
+                                        vpx_input_ctx.height,
+                                        &vpx_input_ctx.framerate,
+                                        img->fmt, img->bit_depth);
+            if (do_md5) {
+              MD5Update(&md5_ctx, (md5byte *)buf, (unsigned int)len);
+            } else {
+              fputs(buf, outfile);
+            }
+          }
+
+          // Y4M frame header
+          len = y4m_write_frame_header(buf, sizeof(buf));
+          if (do_md5) {
+            MD5Update(&md5_ctx, (md5byte *)buf, (unsigned int)len);
+          } else {
+            fputs(buf, outfile);
+          }
+        } else {
+          if (frame_out == 1) {
+            // Check if --yv12 or --i420 options are consistent with the
+            // bit-stream decoded
+            if (opt_i420) {
+              if (img->fmt != VPX_IMG_FMT_I420 &&
+                  img->fmt != VPX_IMG_FMT_I42016) {
+                fprintf(stderr, "Cannot produce i420 output for bit-stream.\n");
+                goto fail;
+              }
+            }
+            if (opt_yv12) {
+              if ((img->fmt != VPX_IMG_FMT_I420 &&
+                   img->fmt != VPX_IMG_FMT_YV12) || img->bit_depth != 8) {
+                fprintf(stderr, "Cannot produce yv12 output for bit-stream.\n");
+                goto fail;
+              }
+            }
+          }
+        }
+
+        if (do_md5) {
+          update_image_md5(img, planes, &md5_ctx);
+        } else {
+          write_image_file(img, planes, outfile);
+        }
+      } else {
+        generate_filename(outfile_pattern, outfile_name, PATH_MAX,
+                          img->d_w, img->d_h, frame_in);
+        if (do_md5) {
+          MD5Init(&md5_ctx);
+          update_image_md5(img, planes, &md5_ctx);
+          MD5Final(md5_digest, &md5_ctx);
+          print_md5(md5_digest, outfile_name);
+        } else {
+          outfile = open_outfile(outfile_name);
+          write_image_file(img, planes, outfile);
+          fclose(outfile);
+        }
+      }
+    }
+  }
+
+  if (summary || progress) {
+    show_progress(frame_in, frame_out, dx_time);
+    fprintf(stderr, "\n");
+  }
+
+  if (frames_corrupted)
+    fprintf(stderr, "WARNING: %d frames corrupted.\n", frames_corrupted);
+
+fail:
+
+  if (vpx_codec_destroy(&decoder)) {
+    fprintf(stderr, "Failed to destroy decoder: %s\n",
+            vpx_codec_error(&decoder));
+    return EXIT_FAILURE;
+  }
+
+  if (!noblit && single_file) {
+    if (do_md5) {
+      MD5Final(md5_digest, &md5_ctx);
+      print_md5(md5_digest, outfile_name);
+    } else {
+      fclose(outfile);
+    }
+  }
+
+#if CONFIG_WEBM_IO
+  if (input.vpx_input_ctx->file_type == FILE_TYPE_WEBM)
+    webm_free(input.webm_ctx);
+#endif
+
+  if (input.vpx_input_ctx->file_type != FILE_TYPE_WEBM)
+    free(buf);
+
+  if (scaled_img) vpx_img_free(scaled_img);
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (img_shifted) vpx_img_free(img_shifted);
+#endif
+
+  for (i = 0; i < ext_fb_list.num_external_frame_buffers; ++i) {
+    free(ext_fb_list.ext_fb[i].data);
+  }
+  free(ext_fb_list.ext_fb);
+
+  fclose(infile);
+  free(argv);
+
+  return frames_corrupted ? EXIT_FAILURE : EXIT_SUCCESS;
+}
+
+int main(int argc, const char **argv_) {
+  unsigned int loops = 1, i;
+  char **argv, **argi, **argj;
+  struct arg arg;
+  int error = 0;
+
+  argv = argv_dup(argc - 1, argv_ + 1);
+  for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
+    memset(&arg, 0, sizeof(arg));
+    arg.argv_step = 1;
+
+    if (arg_match(&arg, &looparg, argi)) {
+      loops = arg_parse_uint(&arg);
+      break;
+    }
+  }
+  free(argv);
+  for (i = 0; !error && i < loops; i++)
+    error = main_loop(argc, argv_);
+  return error;
+}
diff --git a/libvpx/libvpx/vpxenc.c b/libvpx/libvpx/vpxenc.c
new file mode 100644
index 0000000..efcf064
--- /dev/null
+++ b/libvpx/libvpx/vpxenc.c
@@ -0,0 +1,2284 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpxenc.h"
+#include "./vpx_config.h"
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#if CONFIG_LIBYUV
+#include "third_party/libyuv/include/libyuv/scale.h"
+#endif
+
+#include "vpx/vpx_encoder.h"
+#if CONFIG_DECODERS
+#include "vpx/vpx_decoder.h"
+#endif
+
+#include "./args.h"
+#include "./ivfenc.h"
+#include "./tools_common.h"
+
+#if CONFIG_VP8_ENCODER || CONFIG_VP9_ENCODER
+#include "vpx/vp8cx.h"
+#endif
+#if CONFIG_VP8_DECODER || CONFIG_VP9_DECODER
+#include "vpx/vp8dx.h"
+#endif
+
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem_ops.h"
+#include "vpx_ports/vpx_timer.h"
+#include "./rate_hist.h"
+#include "./vpxstats.h"
+#include "./warnings.h"
+#if CONFIG_WEBM_IO
+#include "./webmenc.h"
+#endif
+#include "./y4minput.h"
+
+/* Swallow warnings about unused results of fread/fwrite */
+static size_t wrap_fread(void *ptr, size_t size, size_t nmemb,
+                         FILE *stream) {
+  return fread(ptr, size, nmemb, stream);
+}
+#define fread wrap_fread
+
+static size_t wrap_fwrite(const void *ptr, size_t size, size_t nmemb,
+                          FILE *stream) {
+  return fwrite(ptr, size, nmemb, stream);
+}
+#define fwrite wrap_fwrite
+
+
+static const char *exec_name;
+
+static void warn_or_exit_on_errorv(vpx_codec_ctx_t *ctx, int fatal,
+                                   const char *s, va_list ap) {
+  if (ctx->err) {
+    const char *detail = vpx_codec_error_detail(ctx);
+
+    vfprintf(stderr, s, ap);
+    fprintf(stderr, ": %s\n", vpx_codec_error(ctx));
+
+    if (detail)
+      fprintf(stderr, "    %s\n", detail);
+
+    if (fatal)
+      exit(EXIT_FAILURE);
+  }
+}
+
+static void ctx_exit_on_error(vpx_codec_ctx_t *ctx, const char *s, ...) {
+  va_list ap;
+
+  va_start(ap, s);
+  warn_or_exit_on_errorv(ctx, 1, s, ap);
+  va_end(ap);
+}
+
+static void warn_or_exit_on_error(vpx_codec_ctx_t *ctx, int fatal,
+                                  const char *s, ...) {
+  va_list ap;
+
+  va_start(ap, s);
+  warn_or_exit_on_errorv(ctx, fatal, s, ap);
+  va_end(ap);
+}
+
+static int read_frame(struct VpxInputContext *input_ctx, vpx_image_t *img) {
+  FILE *f = input_ctx->file;
+  y4m_input *y4m = &input_ctx->y4m;
+  int shortread = 0;
+
+  if (input_ctx->file_type == FILE_TYPE_Y4M) {
+    if (y4m_input_fetch_frame(y4m, f, img) < 1)
+      return 0;
+  } else {
+    shortread = read_yuv_frame(input_ctx, img);
+  }
+
+  return !shortread;
+}
+
+static int file_is_y4m(const char detect[4]) {
+  if (memcmp(detect, "YUV4", 4) == 0) {
+    return 1;
+  }
+  return 0;
+}
+
+static int fourcc_is_ivf(const char detect[4]) {
+  if (memcmp(detect, "DKIF", 4) == 0) {
+    return 1;
+  }
+  return 0;
+}
+
+static const arg_def_t debugmode = ARG_DEF(
+    "D", "debug", 0, "Debug mode (makes output deterministic)");
+static const arg_def_t outputfile = ARG_DEF(
+    "o", "output", 1, "Output filename");
+static const arg_def_t use_yv12 = ARG_DEF(
+    NULL, "yv12", 0, "Input file is YV12 ");
+static const arg_def_t use_i420 = ARG_DEF(
+    NULL, "i420", 0, "Input file is I420 (default)");
+static const arg_def_t use_i422 = ARG_DEF(
+    NULL, "i422", 0, "Input file is I422");
+static const arg_def_t use_i444 = ARG_DEF(
+    NULL, "i444", 0, "Input file is I444");
+static const arg_def_t use_i440 = ARG_DEF(
+    NULL, "i440", 0, "Input file is I440");
+static const arg_def_t codecarg = ARG_DEF(
+    NULL, "codec", 1, "Codec to use");
+static const arg_def_t passes = ARG_DEF(
+    "p", "passes", 1, "Number of passes (1/2)");
+static const arg_def_t pass_arg = ARG_DEF(
+    NULL, "pass", 1, "Pass to execute (1/2)");
+static const arg_def_t fpf_name = ARG_DEF(
+    NULL, "fpf", 1, "First pass statistics file name");
+#if CONFIG_FP_MB_STATS
+static const arg_def_t fpmbf_name = ARG_DEF(
+    NULL, "fpmbf", 1, "First pass block statistics file name");
+#endif
+static const arg_def_t limit = ARG_DEF(
+    NULL, "limit", 1, "Stop encoding after n input frames");
+static const arg_def_t skip = ARG_DEF(
+    NULL, "skip", 1, "Skip the first n input frames");
+static const arg_def_t deadline = ARG_DEF(
+    "d", "deadline", 1, "Deadline per frame (usec)");
+static const arg_def_t best_dl = ARG_DEF(
+    NULL, "best", 0, "Use Best Quality Deadline");
+static const arg_def_t good_dl = ARG_DEF(
+    NULL, "good", 0, "Use Good Quality Deadline");
+static const arg_def_t rt_dl = ARG_DEF(
+    NULL, "rt", 0, "Use Realtime Quality Deadline");
+static const arg_def_t quietarg = ARG_DEF(
+    "q", "quiet", 0, "Do not print encode progress");
+static const arg_def_t verbosearg = ARG_DEF(
+    "v", "verbose", 0, "Show encoder parameters");
+static const arg_def_t psnrarg = ARG_DEF(
+    NULL, "psnr", 0, "Show PSNR in status line");
+
+static const struct arg_enum_list test_decode_enum[] = {
+  {"off",   TEST_DECODE_OFF},
+  {"fatal", TEST_DECODE_FATAL},
+  {"warn",  TEST_DECODE_WARN},
+  {NULL, 0}
+};
+static const arg_def_t recontest = ARG_DEF_ENUM(
+    NULL, "test-decode", 1, "Test encode/decode mismatch", test_decode_enum);
+static const arg_def_t framerate = ARG_DEF(
+    NULL, "fps", 1, "Stream frame rate (rate/scale)");
+static const arg_def_t use_webm = ARG_DEF(
+    NULL, "webm", 0, "Output WebM (default when WebM IO is enabled)");
+static const arg_def_t use_ivf = ARG_DEF(
+    NULL, "ivf", 0, "Output IVF");
+static const arg_def_t out_part = ARG_DEF(
+    "P", "output-partitions", 0,
+    "Makes encoder output partitions. Requires IVF output!");
+static const arg_def_t q_hist_n = ARG_DEF(
+    NULL, "q-hist", 1, "Show quantizer histogram (n-buckets)");
+static const arg_def_t rate_hist_n = ARG_DEF(
+    NULL, "rate-hist", 1, "Show rate histogram (n-buckets)");
+static const arg_def_t disable_warnings = ARG_DEF(
+    NULL, "disable-warnings", 0,
+    "Disable warnings about potentially incorrect encode settings.");
+static const arg_def_t disable_warning_prompt = ARG_DEF(
+    "y", "disable-warning-prompt", 0,
+    "Display warnings, but do not prompt user to continue.");
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static const arg_def_t test16bitinternalarg = ARG_DEF(
+    NULL, "test-16bit-internal", 0, "Force use of 16 bit internal buffer");
+#endif
+
+static const arg_def_t *main_args[] = {
+  &debugmode,
+  &outputfile, &codecarg, &passes, &pass_arg, &fpf_name, &limit, &skip,
+  &deadline, &best_dl, &good_dl, &rt_dl,
+  &quietarg, &verbosearg, &psnrarg, &use_webm, &use_ivf, &out_part, &q_hist_n,
+  &rate_hist_n, &disable_warnings, &disable_warning_prompt, &recontest,
+  NULL
+};
+
+static const arg_def_t usage = ARG_DEF(
+    "u", "usage", 1, "Usage profile number to use");
+static const arg_def_t threads = ARG_DEF(
+    "t", "threads", 1, "Max number of threads to use");
+static const arg_def_t profile = ARG_DEF(
+    NULL, "profile", 1, "Bitstream profile number to use");
+static const arg_def_t width = ARG_DEF("w", "width", 1, "Frame width");
+static const arg_def_t height = ARG_DEF("h", "height", 1, "Frame height");
+#if CONFIG_WEBM_IO
+static const struct arg_enum_list stereo_mode_enum[] = {
+  {"mono", STEREO_FORMAT_MONO},
+  {"left-right", STEREO_FORMAT_LEFT_RIGHT},
+  {"bottom-top", STEREO_FORMAT_BOTTOM_TOP},
+  {"top-bottom", STEREO_FORMAT_TOP_BOTTOM},
+  {"right-left", STEREO_FORMAT_RIGHT_LEFT},
+  {NULL, 0}
+};
+static const arg_def_t stereo_mode = ARG_DEF_ENUM(
+    NULL, "stereo-mode", 1, "Stereo 3D video format", stereo_mode_enum);
+#endif
+static const arg_def_t timebase = ARG_DEF(
+    NULL, "timebase", 1, "Output timestamp precision (fractional seconds)");
+static const arg_def_t error_resilient = ARG_DEF(
+    NULL, "error-resilient", 1, "Enable error resiliency features");
+static const arg_def_t lag_in_frames = ARG_DEF(
+    NULL, "lag-in-frames", 1, "Max number of frames to lag");
+
+static const arg_def_t *global_args[] = {
+  &use_yv12, &use_i420, &use_i422, &use_i444, &use_i440,
+  &usage, &threads, &profile,
+  &width, &height,
+#if CONFIG_WEBM_IO
+  &stereo_mode,
+#endif
+  &timebase, &framerate,
+  &error_resilient,
+#if CONFIG_VP9_HIGHBITDEPTH
+  &test16bitinternalarg,
+#endif
+  &lag_in_frames, NULL
+};
+
+static const arg_def_t dropframe_thresh = ARG_DEF(
+    NULL, "drop-frame", 1, "Temporal resampling threshold (buf %)");
+static const arg_def_t resize_allowed = ARG_DEF(
+    NULL, "resize-allowed", 1, "Spatial resampling enabled (bool)");
+static const arg_def_t resize_width = ARG_DEF(
+    NULL, "resize-width", 1, "Width of encoded frame");
+static const arg_def_t resize_height = ARG_DEF(
+    NULL, "resize-height", 1, "Height of encoded frame");
+static const arg_def_t resize_up_thresh = ARG_DEF(
+    NULL, "resize-up", 1, "Upscale threshold (buf %)");
+static const arg_def_t resize_down_thresh = ARG_DEF(
+    NULL, "resize-down", 1, "Downscale threshold (buf %)");
+static const struct arg_enum_list end_usage_enum[] = {
+  {"vbr", VPX_VBR},
+  {"cbr", VPX_CBR},
+  {"cq",  VPX_CQ},
+  {"q",   VPX_Q},
+  {NULL, 0}
+};
+static const arg_def_t end_usage = ARG_DEF_ENUM(
+    NULL, "end-usage", 1, "Rate control mode", end_usage_enum);
+static const arg_def_t target_bitrate = ARG_DEF(
+    NULL, "target-bitrate", 1, "Bitrate (kbps)");
+static const arg_def_t min_quantizer = ARG_DEF(
+    NULL, "min-q", 1, "Minimum (best) quantizer");
+static const arg_def_t max_quantizer = ARG_DEF(
+    NULL, "max-q", 1, "Maximum (worst) quantizer");
+static const arg_def_t undershoot_pct = ARG_DEF(
+    NULL, "undershoot-pct", 1, "Datarate undershoot (min) target (%)");
+static const arg_def_t overshoot_pct = ARG_DEF(
+    NULL, "overshoot-pct", 1, "Datarate overshoot (max) target (%)");
+static const arg_def_t buf_sz = ARG_DEF(
+    NULL, "buf-sz", 1, "Client buffer size (ms)");
+static const arg_def_t buf_initial_sz = ARG_DEF(
+    NULL, "buf-initial-sz", 1, "Client initial buffer size (ms)");
+static const arg_def_t buf_optimal_sz = ARG_DEF(
+    NULL, "buf-optimal-sz", 1, "Client optimal buffer size (ms)");
+static const arg_def_t *rc_args[] = {
+  &dropframe_thresh, &resize_allowed, &resize_width, &resize_height,
+  &resize_up_thresh, &resize_down_thresh, &end_usage, &target_bitrate,
+  &min_quantizer, &max_quantizer, &undershoot_pct, &overshoot_pct, &buf_sz,
+  &buf_initial_sz, &buf_optimal_sz, NULL
+};
+
+
+static const arg_def_t bias_pct = ARG_DEF(
+    NULL, "bias-pct", 1, "CBR/VBR bias (0=CBR, 100=VBR)");
+static const arg_def_t minsection_pct = ARG_DEF(
+    NULL, "minsection-pct", 1, "GOP min bitrate (% of target)");
+static const arg_def_t maxsection_pct = ARG_DEF(
+    NULL, "maxsection-pct", 1, "GOP max bitrate (% of target)");
+static const arg_def_t *rc_twopass_args[] = {
+  &bias_pct, &minsection_pct, &maxsection_pct, NULL
+};
+
+
+static const arg_def_t kf_min_dist = ARG_DEF(
+    NULL, "kf-min-dist", 1, "Minimum keyframe interval (frames)");
+static const arg_def_t kf_max_dist = ARG_DEF(
+    NULL, "kf-max-dist", 1, "Maximum keyframe interval (frames)");
+static const arg_def_t kf_disabled = ARG_DEF(
+    NULL, "disable-kf", 0, "Disable keyframe placement");
+static const arg_def_t *kf_args[] = {
+  &kf_min_dist, &kf_max_dist, &kf_disabled, NULL
+};
+
+
+static const arg_def_t noise_sens = ARG_DEF(
+    NULL, "noise-sensitivity", 1, "Noise sensitivity (frames to blur)");
+static const arg_def_t sharpness = ARG_DEF(
+    NULL, "sharpness", 1, "Loop filter sharpness (0..7)");
+static const arg_def_t static_thresh = ARG_DEF(
+    NULL, "static-thresh", 1, "Motion detection threshold");
+static const arg_def_t auto_altref = ARG_DEF(
+    NULL, "auto-alt-ref", 1, "Enable automatic alt reference frames");
+static const arg_def_t arnr_maxframes = ARG_DEF(
+    NULL, "arnr-maxframes", 1, "AltRef max frames (0..15)");
+static const arg_def_t arnr_strength = ARG_DEF(
+    NULL, "arnr-strength", 1, "AltRef filter strength (0..6)");
+static const arg_def_t arnr_type = ARG_DEF(
+    NULL, "arnr-type", 1, "AltRef type");
+static const struct arg_enum_list tuning_enum[] = {
+  {"psnr", VP8_TUNE_PSNR},
+  {"ssim", VP8_TUNE_SSIM},
+  {NULL, 0}
+};
+static const arg_def_t tune_ssim = ARG_DEF_ENUM(
+    NULL, "tune", 1, "Material to favor", tuning_enum);
+static const arg_def_t cq_level = ARG_DEF(
+    NULL, "cq-level", 1, "Constant/Constrained Quality level");
+static const arg_def_t max_intra_rate_pct = ARG_DEF(
+    NULL, "max-intra-rate", 1, "Max I-frame bitrate (pct)");
+
+#if CONFIG_VP8_ENCODER
+static const arg_def_t cpu_used_vp8 = ARG_DEF(
+    NULL, "cpu-used", 1, "CPU Used (-16..16)");
+static const arg_def_t token_parts = ARG_DEF(
+    NULL, "token-parts", 1, "Number of token partitions to use, log2");
+static const arg_def_t screen_content_mode = ARG_DEF(
+    NULL, "screen-content-mode", 1, "Screen content mode");
+static const arg_def_t *vp8_args[] = {
+  &cpu_used_vp8, &auto_altref, &noise_sens, &sharpness, &static_thresh,
+  &token_parts, &arnr_maxframes, &arnr_strength, &arnr_type,
+  &tune_ssim, &cq_level, &max_intra_rate_pct, &screen_content_mode,
+  NULL
+};
+static const int vp8_arg_ctrl_map[] = {
+  VP8E_SET_CPUUSED, VP8E_SET_ENABLEAUTOALTREF,
+  VP8E_SET_NOISE_SENSITIVITY, VP8E_SET_SHARPNESS, VP8E_SET_STATIC_THRESHOLD,
+  VP8E_SET_TOKEN_PARTITIONS,
+  VP8E_SET_ARNR_MAXFRAMES, VP8E_SET_ARNR_STRENGTH, VP8E_SET_ARNR_TYPE,
+  VP8E_SET_TUNING, VP8E_SET_CQ_LEVEL, VP8E_SET_MAX_INTRA_BITRATE_PCT,
+  VP8E_SET_SCREEN_CONTENT_MODE,
+  0
+};
+#endif
+
+#if CONFIG_VP9_ENCODER
+static const arg_def_t cpu_used_vp9 = ARG_DEF(
+    NULL, "cpu-used", 1, "CPU Used (-8..8)");
+static const arg_def_t tile_cols = ARG_DEF(
+    NULL, "tile-columns", 1, "Number of tile columns to use, log2");
+static const arg_def_t tile_rows = ARG_DEF(
+    NULL, "tile-rows", 1,
+    "Number of tile rows to use, log2 (set to 0 while threads > 1)");
+static const arg_def_t lossless = ARG_DEF(
+    NULL, "lossless", 1, "Lossless mode (0: false (default), 1: true)");
+static const arg_def_t frame_parallel_decoding = ARG_DEF(
+    NULL, "frame-parallel", 1, "Enable frame parallel decodability features");
+static const arg_def_t aq_mode = ARG_DEF(
+    NULL, "aq-mode", 1,
+    "Adaptive quantization mode (0: off (default), 1: variance 2: complexity, "
+    "3: cyclic refresh, 4: equator360)");
+static const arg_def_t frame_periodic_boost = ARG_DEF(
+    NULL, "frame-boost", 1,
+    "Enable frame periodic boost (0: off (default), 1: on)");
+static const arg_def_t gf_cbr_boost_pct = ARG_DEF(
+    NULL, "gf-cbr-boost", 1, "Boost for Golden Frame in CBR mode (pct)");
+static const arg_def_t max_inter_rate_pct = ARG_DEF(
+    NULL, "max-inter-rate", 1, "Max P-frame bitrate (pct)");
+static const arg_def_t min_gf_interval = ARG_DEF(
+    NULL, "min-gf-interval", 1,
+    "min gf/arf frame interval (default 0, indicating in-built behavior)");
+static const arg_def_t max_gf_interval = ARG_DEF(
+    NULL, "max-gf-interval", 1,
+    "max gf/arf frame interval (default 0, indicating in-built behavior)");
+
+static const struct arg_enum_list color_space_enum[] = {
+  { "unknown", VPX_CS_UNKNOWN },
+  { "bt601", VPX_CS_BT_601 },
+  { "bt709", VPX_CS_BT_709 },
+  { "smpte170", VPX_CS_SMPTE_170 },
+  { "smpte240", VPX_CS_SMPTE_240 },
+  { "bt2020", VPX_CS_BT_2020 },
+  { "reserved", VPX_CS_RESERVED },
+  { "sRGB", VPX_CS_SRGB },
+  { NULL, 0 }
+};
+
+static const arg_def_t input_color_space = ARG_DEF_ENUM(
+    NULL, "color-space", 1,
+    "The color space of input content:", color_space_enum);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static const struct arg_enum_list bitdepth_enum[] = {
+  {"8",  VPX_BITS_8},
+  {"10", VPX_BITS_10},
+  {"12", VPX_BITS_12},
+  {NULL, 0}
+};
+
+static const arg_def_t bitdeptharg = ARG_DEF_ENUM(
+    "b", "bit-depth", 1,
+    "Bit depth for codec (8 for version <=1, 10 or 12 for version 2)",
+    bitdepth_enum);
+static const arg_def_t inbitdeptharg = ARG_DEF(
+    NULL, "input-bit-depth", 1, "Bit depth of input");
+#endif
+
+static const struct arg_enum_list tune_content_enum[] = {
+  {"default", VP9E_CONTENT_DEFAULT},
+  {"screen", VP9E_CONTENT_SCREEN},
+  {NULL, 0}
+};
+
+static const arg_def_t tune_content = ARG_DEF_ENUM(
+    NULL, "tune-content", 1, "Tune content type", tune_content_enum);
+
+static const arg_def_t target_level = ARG_DEF(
+    NULL, "target-level", 1,
+    "Target level (255: off (default); 0: only keep level stats; 10: level 1.0;"
+    " 11: level 1.1; ... 62: level 6.2)");
+#endif
+
+#if CONFIG_VP9_ENCODER
+static const arg_def_t *vp9_args[] = {
+  &cpu_used_vp9, &auto_altref, &sharpness, &static_thresh,
+  &tile_cols, &tile_rows, &arnr_maxframes, &arnr_strength, &arnr_type,
+  &tune_ssim, &cq_level, &max_intra_rate_pct, &max_inter_rate_pct,
+  &gf_cbr_boost_pct, &lossless,
+  &frame_parallel_decoding, &aq_mode, &frame_periodic_boost,
+  &noise_sens, &tune_content, &input_color_space,
+  &min_gf_interval, &max_gf_interval, &target_level,
+#if CONFIG_VP9_HIGHBITDEPTH
+  &bitdeptharg, &inbitdeptharg,
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  NULL
+};
+static const int vp9_arg_ctrl_map[] = {
+  VP8E_SET_CPUUSED, VP8E_SET_ENABLEAUTOALTREF,
+  VP8E_SET_SHARPNESS, VP8E_SET_STATIC_THRESHOLD,
+  VP9E_SET_TILE_COLUMNS, VP9E_SET_TILE_ROWS,
+  VP8E_SET_ARNR_MAXFRAMES, VP8E_SET_ARNR_STRENGTH, VP8E_SET_ARNR_TYPE,
+  VP8E_SET_TUNING, VP8E_SET_CQ_LEVEL, VP8E_SET_MAX_INTRA_BITRATE_PCT,
+  VP9E_SET_MAX_INTER_BITRATE_PCT, VP9E_SET_GF_CBR_BOOST_PCT,
+  VP9E_SET_LOSSLESS, VP9E_SET_FRAME_PARALLEL_DECODING, VP9E_SET_AQ_MODE,
+  VP9E_SET_FRAME_PERIODIC_BOOST, VP9E_SET_NOISE_SENSITIVITY,
+  VP9E_SET_TUNE_CONTENT, VP9E_SET_COLOR_SPACE,
+  VP9E_SET_MIN_GF_INTERVAL, VP9E_SET_MAX_GF_INTERVAL, VP9E_SET_TARGET_LEVEL,
+  0
+};
+#endif
+
+static const arg_def_t *no_args[] = { NULL };
+
+void usage_exit(void) {
+  int i;
+  const int num_encoder = get_vpx_encoder_count();
+
+  fprintf(stderr, "Usage: %s <options> -o dst_filename src_filename \n",
+          exec_name);
+
+  fprintf(stderr, "\nOptions:\n");
+  arg_show_usage(stderr, main_args);
+  fprintf(stderr, "\nEncoder Global Options:\n");
+  arg_show_usage(stderr, global_args);
+  fprintf(stderr, "\nRate Control Options:\n");
+  arg_show_usage(stderr, rc_args);
+  fprintf(stderr, "\nTwopass Rate Control Options:\n");
+  arg_show_usage(stderr, rc_twopass_args);
+  fprintf(stderr, "\nKeyframe Placement Options:\n");
+  arg_show_usage(stderr, kf_args);
+#if CONFIG_VP8_ENCODER
+  fprintf(stderr, "\nVP8 Specific Options:\n");
+  arg_show_usage(stderr, vp8_args);
+#endif
+#if CONFIG_VP9_ENCODER
+  fprintf(stderr, "\nVP9 Specific Options:\n");
+  arg_show_usage(stderr, vp9_args);
+#endif
+  fprintf(stderr, "\nStream timebase (--timebase):\n"
+          "  The desired precision of timestamps in the output, expressed\n"
+          "  in fractional seconds. Default is 1/1000.\n");
+  fprintf(stderr, "\nIncluded encoders:\n\n");
+
+  for (i = 0; i < num_encoder; ++i) {
+    const VpxInterface *const encoder = get_vpx_encoder_by_index(i);
+    const char* defstr = (i == (num_encoder - 1)) ? "(default)" : "";
+      fprintf(stderr, "    %-6s - %s %s\n",
+              encoder->name, vpx_codec_iface_name(encoder->codec_interface()),
+              defstr);
+  }
+  fprintf(stderr, "\n        ");
+  fprintf(stderr, "Use --codec to switch to a non-default encoder.\n\n");
+
+  exit(EXIT_FAILURE);
+}
+
+#define mmin(a, b)  ((a) < (b) ? (a) : (b))
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void find_mismatch_high(const vpx_image_t *const img1,
+                               const vpx_image_t *const img2,
+                               int yloc[4], int uloc[4], int vloc[4]) {
+  uint16_t *plane1, *plane2;
+  uint32_t stride1, stride2;
+  const uint32_t bsize = 64;
+  const uint32_t bsizey = bsize >> img1->y_chroma_shift;
+  const uint32_t bsizex = bsize >> img1->x_chroma_shift;
+  const uint32_t c_w =
+      (img1->d_w + img1->x_chroma_shift) >> img1->x_chroma_shift;
+  const uint32_t c_h =
+      (img1->d_h + img1->y_chroma_shift) >> img1->y_chroma_shift;
+  int match = 1;
+  uint32_t i, j;
+  yloc[0] = yloc[1] = yloc[2] = yloc[3] = -1;
+  plane1 = (uint16_t*)img1->planes[VPX_PLANE_Y];
+  plane2 = (uint16_t*)img2->planes[VPX_PLANE_Y];
+  stride1 = img1->stride[VPX_PLANE_Y]/2;
+  stride2 = img2->stride[VPX_PLANE_Y]/2;
+  for (i = 0, match = 1; match && i < img1->d_h; i += bsize) {
+    for (j = 0; match && j < img1->d_w; j += bsize) {
+      int k, l;
+      const int si = mmin(i + bsize, img1->d_h) - i;
+      const int sj = mmin(j + bsize, img1->d_w) - j;
+      for (k = 0; match && k < si; ++k) {
+        for (l = 0; match && l < sj; ++l) {
+          if (*(plane1 + (i + k) * stride1 + j + l) !=
+              *(plane2 + (i + k) * stride2 + j + l)) {
+            yloc[0] = i + k;
+            yloc[1] = j + l;
+            yloc[2] = *(plane1 + (i + k) * stride1 + j + l);
+            yloc[3] = *(plane2 + (i + k) * stride2 + j + l);
+            match = 0;
+            break;
+          }
+        }
+      }
+    }
+  }
+
+  uloc[0] = uloc[1] = uloc[2] = uloc[3] = -1;
+  plane1 = (uint16_t*)img1->planes[VPX_PLANE_U];
+  plane2 = (uint16_t*)img2->planes[VPX_PLANE_U];
+  stride1 = img1->stride[VPX_PLANE_U]/2;
+  stride2 = img2->stride[VPX_PLANE_U]/2;
+  for (i = 0, match = 1; match && i < c_h; i += bsizey) {
+    for (j = 0; match && j < c_w; j += bsizex) {
+      int k, l;
+      const int si = mmin(i + bsizey, c_h - i);
+      const int sj = mmin(j + bsizex, c_w - j);
+      for (k = 0; match && k < si; ++k) {
+        for (l = 0; match && l < sj; ++l) {
+          if (*(plane1 + (i + k) * stride1 + j + l) !=
+              *(plane2 + (i + k) * stride2 + j + l)) {
+            uloc[0] = i + k;
+            uloc[1] = j + l;
+            uloc[2] = *(plane1 + (i + k) * stride1 + j + l);
+            uloc[3] = *(plane2 + (i + k) * stride2 + j + l);
+            match = 0;
+            break;
+          }
+        }
+      }
+    }
+  }
+
+  vloc[0] = vloc[1] = vloc[2] = vloc[3] = -1;
+  plane1 = (uint16_t*)img1->planes[VPX_PLANE_V];
+  plane2 = (uint16_t*)img2->planes[VPX_PLANE_V];
+  stride1 = img1->stride[VPX_PLANE_V]/2;
+  stride2 = img2->stride[VPX_PLANE_V]/2;
+  for (i = 0, match = 1; match && i < c_h; i += bsizey) {
+    for (j = 0; match && j < c_w; j += bsizex) {
+      int k, l;
+      const int si = mmin(i + bsizey, c_h - i);
+      const int sj = mmin(j + bsizex, c_w - j);
+      for (k = 0; match && k < si; ++k) {
+        for (l = 0; match && l < sj; ++l) {
+          if (*(plane1 + (i + k) * stride1 + j + l) !=
+              *(plane2 + (i + k) * stride2 + j + l)) {
+            vloc[0] = i + k;
+            vloc[1] = j + l;
+            vloc[2] = *(plane1 + (i + k) * stride1 + j + l);
+            vloc[3] = *(plane2 + (i + k) * stride2 + j + l);
+            match = 0;
+            break;
+          }
+        }
+      }
+    }
+  }
+}
+#endif
+
+static void find_mismatch(const vpx_image_t *const img1,
+                          const vpx_image_t *const img2,
+                          int yloc[4], int uloc[4], int vloc[4]) {
+  const uint32_t bsize = 64;
+  const uint32_t bsizey = bsize >> img1->y_chroma_shift;
+  const uint32_t bsizex = bsize >> img1->x_chroma_shift;
+  const uint32_t c_w =
+      (img1->d_w + img1->x_chroma_shift) >> img1->x_chroma_shift;
+  const uint32_t c_h =
+      (img1->d_h + img1->y_chroma_shift) >> img1->y_chroma_shift;
+  int match = 1;
+  uint32_t i, j;
+  yloc[0] = yloc[1] = yloc[2] = yloc[3] = -1;
+  for (i = 0, match = 1; match && i < img1->d_h; i += bsize) {
+    for (j = 0; match && j < img1->d_w; j += bsize) {
+      int k, l;
+      const int si = mmin(i + bsize, img1->d_h) - i;
+      const int sj = mmin(j + bsize, img1->d_w) - j;
+      for (k = 0; match && k < si; ++k) {
+        for (l = 0; match && l < sj; ++l) {
+          if (*(img1->planes[VPX_PLANE_Y] +
+                (i + k) * img1->stride[VPX_PLANE_Y] + j + l) !=
+              *(img2->planes[VPX_PLANE_Y] +
+                (i + k) * img2->stride[VPX_PLANE_Y] + j + l)) {
+            yloc[0] = i + k;
+            yloc[1] = j + l;
+            yloc[2] = *(img1->planes[VPX_PLANE_Y] +
+                        (i + k) * img1->stride[VPX_PLANE_Y] + j + l);
+            yloc[3] = *(img2->planes[VPX_PLANE_Y] +
+                        (i + k) * img2->stride[VPX_PLANE_Y] + j + l);
+            match = 0;
+            break;
+          }
+        }
+      }
+    }
+  }
+
+  uloc[0] = uloc[1] = uloc[2] = uloc[3] = -1;
+  for (i = 0, match = 1; match && i < c_h; i += bsizey) {
+    for (j = 0; match && j < c_w; j += bsizex) {
+      int k, l;
+      const int si = mmin(i + bsizey, c_h - i);
+      const int sj = mmin(j + bsizex, c_w - j);
+      for (k = 0; match && k < si; ++k) {
+        for (l = 0; match && l < sj; ++l) {
+          if (*(img1->planes[VPX_PLANE_U] +
+                (i + k) * img1->stride[VPX_PLANE_U] + j + l) !=
+              *(img2->planes[VPX_PLANE_U] +
+                (i + k) * img2->stride[VPX_PLANE_U] + j + l)) {
+            uloc[0] = i + k;
+            uloc[1] = j + l;
+            uloc[2] = *(img1->planes[VPX_PLANE_U] +
+                        (i + k) * img1->stride[VPX_PLANE_U] + j + l);
+            uloc[3] = *(img2->planes[VPX_PLANE_U] +
+                        (i + k) * img2->stride[VPX_PLANE_U] + j + l);
+            match = 0;
+            break;
+          }
+        }
+      }
+    }
+  }
+  vloc[0] = vloc[1] = vloc[2] = vloc[3] = -1;
+  for (i = 0, match = 1; match && i < c_h; i += bsizey) {
+    for (j = 0; match && j < c_w; j += bsizex) {
+      int k, l;
+      const int si = mmin(i + bsizey, c_h - i);
+      const int sj = mmin(j + bsizex, c_w - j);
+      for (k = 0; match && k < si; ++k) {
+        for (l = 0; match && l < sj; ++l) {
+          if (*(img1->planes[VPX_PLANE_V] +
+                (i + k) * img1->stride[VPX_PLANE_V] + j + l) !=
+              *(img2->planes[VPX_PLANE_V] +
+                (i + k) * img2->stride[VPX_PLANE_V] + j + l)) {
+            vloc[0] = i + k;
+            vloc[1] = j + l;
+            vloc[2] = *(img1->planes[VPX_PLANE_V] +
+                        (i + k) * img1->stride[VPX_PLANE_V] + j + l);
+            vloc[3] = *(img2->planes[VPX_PLANE_V] +
+                        (i + k) * img2->stride[VPX_PLANE_V] + j + l);
+            match = 0;
+            break;
+          }
+        }
+      }
+    }
+  }
+}
+
+static int compare_img(const vpx_image_t *const img1,
+                       const vpx_image_t *const img2) {
+  uint32_t l_w = img1->d_w;
+  uint32_t c_w =
+      (img1->d_w + img1->x_chroma_shift) >> img1->x_chroma_shift;
+  const uint32_t c_h =
+      (img1->d_h + img1->y_chroma_shift) >> img1->y_chroma_shift;
+  uint32_t i;
+  int match = 1;
+
+  match &= (img1->fmt == img2->fmt);
+  match &= (img1->d_w == img2->d_w);
+  match &= (img1->d_h == img2->d_h);
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (img1->fmt & VPX_IMG_FMT_HIGHBITDEPTH) {
+    l_w *= 2;
+    c_w *= 2;
+  }
+#endif
+
+  for (i = 0; i < img1->d_h; ++i)
+    match &= (memcmp(img1->planes[VPX_PLANE_Y] + i * img1->stride[VPX_PLANE_Y],
+                     img2->planes[VPX_PLANE_Y] + i * img2->stride[VPX_PLANE_Y],
+                     l_w) == 0);
+
+  for (i = 0; i < c_h; ++i)
+    match &= (memcmp(img1->planes[VPX_PLANE_U] + i * img1->stride[VPX_PLANE_U],
+                     img2->planes[VPX_PLANE_U] + i * img2->stride[VPX_PLANE_U],
+                     c_w) == 0);
+
+  for (i = 0; i < c_h; ++i)
+    match &= (memcmp(img1->planes[VPX_PLANE_V] + i * img1->stride[VPX_PLANE_V],
+                     img2->planes[VPX_PLANE_V] + i * img2->stride[VPX_PLANE_V],
+                     c_w) == 0);
+
+  return match;
+}
+
+
+#define NELEMENTS(x) (sizeof(x)/sizeof(x[0]))
+#if CONFIG_VP9_ENCODER
+#define ARG_CTRL_CNT_MAX NELEMENTS(vp9_arg_ctrl_map)
+#else
+#define ARG_CTRL_CNT_MAX NELEMENTS(vp8_arg_ctrl_map)
+#endif
+
+#if !CONFIG_WEBM_IO
+typedef int stereo_format_t;
+struct WebmOutputContext { int debug; };
+#endif
+
+/* Per-stream configuration */
+struct stream_config {
+  struct vpx_codec_enc_cfg  cfg;
+  const char               *out_fn;
+  const char               *stats_fn;
+#if CONFIG_FP_MB_STATS
+  const char               *fpmb_stats_fn;
+#endif
+  stereo_format_t           stereo_fmt;
+  int                       arg_ctrls[ARG_CTRL_CNT_MAX][2];
+  int                       arg_ctrl_cnt;
+  int                       write_webm;
+#if CONFIG_VP9_HIGHBITDEPTH
+  // whether to use 16bit internal buffers
+  int                       use_16bit_internal;
+#endif
+};
+
+
+struct stream_state {
+  int                       index;
+  struct stream_state      *next;
+  struct stream_config      config;
+  FILE                     *file;
+  struct rate_hist         *rate_hist;
+  struct WebmOutputContext  webm_ctx;
+  uint64_t                  psnr_sse_total;
+  uint64_t                  psnr_samples_total;
+  double                    psnr_totals[4];
+  int                       psnr_count;
+  int                       counts[64];
+  vpx_codec_ctx_t           encoder;
+  unsigned int              frames_out;
+  uint64_t                  cx_time;
+  size_t                    nbytes;
+  stats_io_t                stats;
+#if CONFIG_FP_MB_STATS
+  stats_io_t                fpmb_stats;
+#endif
+  struct vpx_image         *img;
+  vpx_codec_ctx_t           decoder;
+  int                       mismatch_seen;
+};
+
+
+static void validate_positive_rational(const char          *msg,
+                                       struct vpx_rational *rat) {
+  if (rat->den < 0) {
+    rat->num *= -1;
+    rat->den *= -1;
+  }
+
+  if (rat->num < 0)
+    die("Error: %s must be positive\n", msg);
+
+  if (!rat->den)
+    die("Error: %s has zero denominator\n", msg);
+}
+
+
+static void parse_global_config(struct VpxEncoderConfig *global, char **argv) {
+  char       **argi, **argj;
+  struct arg   arg;
+  const int num_encoder = get_vpx_encoder_count();
+
+  if (num_encoder < 1)
+    die("Error: no valid encoder available\n");
+
+  /* Initialize default parameters */
+  memset(global, 0, sizeof(*global));
+  global->codec = get_vpx_encoder_by_index(num_encoder - 1);
+  global->passes = 0;
+  global->color_type = I420;
+  /* Assign default deadline to good quality */
+  global->deadline = VPX_DL_GOOD_QUALITY;
+
+  for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
+    arg.argv_step = 1;
+
+    if (arg_match(&arg, &codecarg, argi)) {
+      global->codec = get_vpx_encoder_by_name(arg.val);
+      if (!global->codec)
+        die("Error: Unrecognized argument (%s) to --codec\n", arg.val);
+    } else if (arg_match(&arg, &passes, argi)) {
+      global->passes = arg_parse_uint(&arg);
+
+      if (global->passes < 1 || global->passes > 2)
+        die("Error: Invalid number of passes (%d)\n", global->passes);
+    } else if (arg_match(&arg, &pass_arg, argi)) {
+      global->pass = arg_parse_uint(&arg);
+
+      if (global->pass < 1 || global->pass > 2)
+        die("Error: Invalid pass selected (%d)\n",
+            global->pass);
+    } else if (arg_match(&arg, &usage, argi))
+      global->usage = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &deadline, argi))
+      global->deadline = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &best_dl, argi))
+      global->deadline = VPX_DL_BEST_QUALITY;
+    else if (arg_match(&arg, &good_dl, argi))
+      global->deadline = VPX_DL_GOOD_QUALITY;
+    else if (arg_match(&arg, &rt_dl, argi))
+      global->deadline = VPX_DL_REALTIME;
+    else if (arg_match(&arg, &use_yv12, argi))
+      global->color_type = YV12;
+    else if (arg_match(&arg, &use_i420, argi))
+      global->color_type = I420;
+    else if (arg_match(&arg, &use_i422, argi))
+      global->color_type = I422;
+    else if (arg_match(&arg, &use_i444, argi))
+      global->color_type = I444;
+    else if (arg_match(&arg, &use_i440, argi))
+      global->color_type = I440;
+    else if (arg_match(&arg, &quietarg, argi))
+      global->quiet = 1;
+    else if (arg_match(&arg, &verbosearg, argi))
+      global->verbose = 1;
+    else if (arg_match(&arg, &limit, argi))
+      global->limit = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &skip, argi))
+      global->skip_frames = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &psnrarg, argi))
+      global->show_psnr = 1;
+    else if (arg_match(&arg, &recontest, argi))
+      global->test_decode = arg_parse_enum_or_int(&arg);
+    else if (arg_match(&arg, &framerate, argi)) {
+      global->framerate = arg_parse_rational(&arg);
+      validate_positive_rational(arg.name, &global->framerate);
+      global->have_framerate = 1;
+    } else if (arg_match(&arg, &out_part, argi))
+      global->out_part = 1;
+    else if (arg_match(&arg, &debugmode, argi))
+      global->debug = 1;
+    else if (arg_match(&arg, &q_hist_n, argi))
+      global->show_q_hist_buckets = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &rate_hist_n, argi))
+      global->show_rate_hist_buckets = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &disable_warnings, argi))
+      global->disable_warnings = 1;
+    else if (arg_match(&arg, &disable_warning_prompt, argi))
+      global->disable_warning_prompt = 1;
+    else
+      argj++;
+  }
+
+  if (global->pass) {
+    /* DWIM: Assume the user meant passes=2 if pass=2 is specified */
+    if (global->pass > global->passes) {
+      warn("Assuming --pass=%d implies --passes=%d\n",
+           global->pass, global->pass);
+      global->passes = global->pass;
+    }
+  }
+  /* Validate global config */
+  if (global->passes == 0) {
+#if CONFIG_VP9_ENCODER
+    // Make default VP9 passes = 2 until there is a better quality 1-pass
+    // encoder
+    if (global->codec != NULL && global->codec->name != NULL)
+      global->passes = (strcmp(global->codec->name, "vp9") == 0 &&
+                        global->deadline != VPX_DL_REALTIME) ? 2 : 1;
+#else
+    global->passes = 1;
+#endif
+  }
+
+  if (global->deadline == VPX_DL_REALTIME &&
+      global->passes > 1) {
+    warn("Enforcing one-pass encoding in realtime mode\n");
+    global->passes = 1;
+  }
+}
+
+
+static void open_input_file(struct VpxInputContext *input) {
+  /* Parse certain options from the input file, if possible */
+  input->file = strcmp(input->filename, "-")
+      ? fopen(input->filename, "rb") : set_binary_mode(stdin);
+
+  if (!input->file)
+    fatal("Failed to open input file");
+
+  if (!fseeko(input->file, 0, SEEK_END)) {
+    /* Input file is seekable. Figure out how long it is, so we can get
+     * progress info.
+     */
+    input->length = ftello(input->file);
+    rewind(input->file);
+  }
+
+  /* Default to 1:1 pixel aspect ratio. */
+  input->pixel_aspect_ratio.numerator = 1;
+  input->pixel_aspect_ratio.denominator = 1;
+
+  /* For RAW input sources, these bytes will applied on the first frame
+   *  in read_frame().
+   */
+  input->detect.buf_read = fread(input->detect.buf, 1, 4, input->file);
+  input->detect.position = 0;
+
+  if (input->detect.buf_read == 4
+      && file_is_y4m(input->detect.buf)) {
+    if (y4m_input_open(&input->y4m, input->file, input->detect.buf, 4,
+                       input->only_i420) >= 0) {
+      input->file_type = FILE_TYPE_Y4M;
+      input->width = input->y4m.pic_w;
+      input->height = input->y4m.pic_h;
+      input->pixel_aspect_ratio.numerator = input->y4m.par_n;
+      input->pixel_aspect_ratio.denominator = input->y4m.par_d;
+      input->framerate.numerator = input->y4m.fps_n;
+      input->framerate.denominator = input->y4m.fps_d;
+      input->fmt = input->y4m.vpx_fmt;
+      input->bit_depth = input->y4m.bit_depth;
+    } else
+      fatal("Unsupported Y4M stream.");
+  } else if (input->detect.buf_read == 4 && fourcc_is_ivf(input->detect.buf)) {
+    fatal("IVF is not supported as input.");
+  } else {
+    input->file_type = FILE_TYPE_RAW;
+  }
+}
+
+
+static void close_input_file(struct VpxInputContext *input) {
+  fclose(input->file);
+  if (input->file_type == FILE_TYPE_Y4M)
+    y4m_input_close(&input->y4m);
+}
+
+static struct stream_state *new_stream(struct VpxEncoderConfig *global,
+                                       struct stream_state *prev) {
+  struct stream_state *stream;
+
+  stream = calloc(1, sizeof(*stream));
+  if (stream == NULL) {
+    fatal("Failed to allocate new stream.");
+  }
+
+  if (prev) {
+    memcpy(stream, prev, sizeof(*stream));
+    stream->index++;
+    prev->next = stream;
+  } else {
+    vpx_codec_err_t  res;
+
+    /* Populate encoder configuration */
+    res = vpx_codec_enc_config_default(global->codec->codec_interface(),
+                                       &stream->config.cfg,
+                                       global->usage);
+    if (res)
+      fatal("Failed to get config: %s\n", vpx_codec_err_to_string(res));
+
+    /* Change the default timebase to a high enough value so that the
+     * encoder will always create strictly increasing timestamps.
+     */
+    stream->config.cfg.g_timebase.den = 1000;
+
+    /* Never use the library's default resolution, require it be parsed
+     * from the file or set on the command line.
+     */
+    stream->config.cfg.g_w = 0;
+    stream->config.cfg.g_h = 0;
+
+    /* Initialize remaining stream parameters */
+    stream->config.write_webm = 1;
+#if CONFIG_WEBM_IO
+    stream->config.stereo_fmt = STEREO_FORMAT_MONO;
+    stream->webm_ctx.last_pts_ns = -1;
+    stream->webm_ctx.writer = NULL;
+    stream->webm_ctx.segment = NULL;
+#endif
+
+    /* Allows removal of the application version from the EBML tags */
+    stream->webm_ctx.debug = global->debug;
+
+    /* Default lag_in_frames is 0 in realtime mode */
+    if (global->deadline == VPX_DL_REALTIME)
+      stream->config.cfg.g_lag_in_frames = 0;
+  }
+
+  /* Output files must be specified for each stream */
+  stream->config.out_fn = NULL;
+
+  stream->next = NULL;
+  return stream;
+}
+
+
+static int parse_stream_params(struct VpxEncoderConfig *global,
+                               struct stream_state  *stream,
+                               char **argv) {
+  char                   **argi, **argj;
+  struct arg               arg;
+  static const arg_def_t **ctrl_args = no_args;
+  static const int        *ctrl_args_map = NULL;
+  struct stream_config    *config = &stream->config;
+  int                      eos_mark_found = 0;
+#if CONFIG_VP9_HIGHBITDEPTH
+  int                      test_16bit_internal = 0;
+#endif
+
+  // Handle codec specific options
+  if (0) {
+#if CONFIG_VP8_ENCODER
+  } else if (strcmp(global->codec->name, "vp8") == 0) {
+    ctrl_args = vp8_args;
+    ctrl_args_map = vp8_arg_ctrl_map;
+#endif
+#if CONFIG_VP9_ENCODER
+  } else if (strcmp(global->codec->name, "vp9") == 0) {
+    ctrl_args = vp9_args;
+    ctrl_args_map = vp9_arg_ctrl_map;
+#endif
+  }
+
+  for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
+    arg.argv_step = 1;
+
+    /* Once we've found an end-of-stream marker (--) we want to continue
+     * shifting arguments but not consuming them.
+     */
+    if (eos_mark_found) {
+      argj++;
+      continue;
+    } else if (!strcmp(*argj, "--")) {
+      eos_mark_found = 1;
+      continue;
+    }
+
+    if (arg_match(&arg, &outputfile, argi)) {
+      config->out_fn = arg.val;
+    } else if (arg_match(&arg, &fpf_name, argi)) {
+      config->stats_fn = arg.val;
+#if CONFIG_FP_MB_STATS
+    } else if (arg_match(&arg, &fpmbf_name, argi)) {
+      config->fpmb_stats_fn = arg.val;
+#endif
+    } else if (arg_match(&arg, &use_webm, argi)) {
+#if CONFIG_WEBM_IO
+      config->write_webm = 1;
+#else
+      die("Error: --webm specified but webm is disabled.");
+#endif
+    } else if (arg_match(&arg, &use_ivf, argi)) {
+      config->write_webm = 0;
+    } else if (arg_match(&arg, &threads, argi)) {
+      config->cfg.g_threads = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &profile, argi)) {
+      config->cfg.g_profile = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &width, argi)) {
+      config->cfg.g_w = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &height, argi)) {
+      config->cfg.g_h = arg_parse_uint(&arg);
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else if (arg_match(&arg, &bitdeptharg, argi)) {
+      config->cfg.g_bit_depth = arg_parse_enum_or_int(&arg);
+    } else if (arg_match(&arg, &inbitdeptharg, argi)) {
+      config->cfg.g_input_bit_depth = arg_parse_uint(&arg);
+#endif
+#if CONFIG_WEBM_IO
+    } else if (arg_match(&arg, &stereo_mode, argi)) {
+      config->stereo_fmt = arg_parse_enum_or_int(&arg);
+#endif
+    } else if (arg_match(&arg, &timebase, argi)) {
+      config->cfg.g_timebase = arg_parse_rational(&arg);
+      validate_positive_rational(arg.name, &config->cfg.g_timebase);
+    } else if (arg_match(&arg, &error_resilient, argi)) {
+      config->cfg.g_error_resilient = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &lag_in_frames, argi)) {
+      config->cfg.g_lag_in_frames = arg_parse_uint(&arg);
+      if (global->deadline == VPX_DL_REALTIME &&
+          config->cfg.g_lag_in_frames != 0) {
+        warn("non-zero %s option ignored in realtime mode.\n", arg.name);
+        config->cfg.g_lag_in_frames = 0;
+      }
+    } else if (arg_match(&arg, &dropframe_thresh, argi)) {
+      config->cfg.rc_dropframe_thresh = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_allowed, argi)) {
+      config->cfg.rc_resize_allowed = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_width, argi)) {
+      config->cfg.rc_scaled_width = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_height, argi)) {
+      config->cfg.rc_scaled_height = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_up_thresh, argi)) {
+      config->cfg.rc_resize_up_thresh = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_down_thresh, argi)) {
+      config->cfg.rc_resize_down_thresh = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &end_usage, argi)) {
+      config->cfg.rc_end_usage = arg_parse_enum_or_int(&arg);
+    } else if (arg_match(&arg, &target_bitrate, argi)) {
+      config->cfg.rc_target_bitrate = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &min_quantizer, argi)) {
+      config->cfg.rc_min_quantizer = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &max_quantizer, argi)) {
+      config->cfg.rc_max_quantizer = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &undershoot_pct, argi)) {
+      config->cfg.rc_undershoot_pct = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &overshoot_pct, argi)) {
+      config->cfg.rc_overshoot_pct = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &buf_sz, argi)) {
+      config->cfg.rc_buf_sz = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &buf_initial_sz, argi)) {
+      config->cfg.rc_buf_initial_sz = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &buf_optimal_sz, argi)) {
+      config->cfg.rc_buf_optimal_sz = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &bias_pct, argi)) {
+        config->cfg.rc_2pass_vbr_bias_pct = arg_parse_uint(&arg);
+      if (global->passes < 2)
+        warn("option %s ignored in one-pass mode.\n", arg.name);
+    } else if (arg_match(&arg, &minsection_pct, argi)) {
+      config->cfg.rc_2pass_vbr_minsection_pct = arg_parse_uint(&arg);
+
+      if (global->passes < 2)
+        warn("option %s ignored in one-pass mode.\n", arg.name);
+    } else if (arg_match(&arg, &maxsection_pct, argi)) {
+      config->cfg.rc_2pass_vbr_maxsection_pct = arg_parse_uint(&arg);
+
+      if (global->passes < 2)
+        warn("option %s ignored in one-pass mode.\n", arg.name);
+    } else if (arg_match(&arg, &kf_min_dist, argi)) {
+      config->cfg.kf_min_dist = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &kf_max_dist, argi)) {
+      config->cfg.kf_max_dist = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &kf_disabled, argi)) {
+      config->cfg.kf_mode = VPX_KF_DISABLED;
+#if CONFIG_VP9_HIGHBITDEPTH
+    } else if (arg_match(&arg, &test16bitinternalarg, argi)) {
+      if (strcmp(global->codec->name, "vp9") == 0) {
+        test_16bit_internal = 1;
+      }
+#endif
+    } else {
+      int i, match = 0;
+      for (i = 0; ctrl_args[i]; i++) {
+        if (arg_match(&arg, ctrl_args[i], argi)) {
+          int j;
+          match = 1;
+
+          /* Point either to the next free element or the first
+          * instance of this control.
+          */
+          for (j = 0; j < config->arg_ctrl_cnt; j++)
+            if (ctrl_args_map != NULL &&
+                config->arg_ctrls[j][0] == ctrl_args_map[i])
+              break;
+
+          /* Update/insert */
+          assert(j < (int)ARG_CTRL_CNT_MAX);
+          if (ctrl_args_map != NULL && j < (int)ARG_CTRL_CNT_MAX) {
+            config->arg_ctrls[j][0] = ctrl_args_map[i];
+            config->arg_ctrls[j][1] = arg_parse_enum_or_int(&arg);
+            if (j == config->arg_ctrl_cnt)
+              config->arg_ctrl_cnt++;
+          }
+        }
+      }
+      if (!match)
+        argj++;
+    }
+  }
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (strcmp(global->codec->name, "vp9") == 0) {
+    config->use_16bit_internal = test_16bit_internal |
+                                 (config->cfg.g_profile > 1);
+  }
+#endif
+  return eos_mark_found;
+}
+
+
+#define FOREACH_STREAM(func) \
+  do { \
+    struct stream_state *stream; \
+    for (stream = streams; stream; stream = stream->next) { \
+      func; \
+    } \
+  } while (0)
+
+
+static void validate_stream_config(const struct stream_state *stream,
+                                   const struct VpxEncoderConfig *global) {
+  const struct stream_state *streami;
+  (void)global;
+
+  if (!stream->config.cfg.g_w || !stream->config.cfg.g_h)
+    fatal("Stream %d: Specify stream dimensions with --width (-w) "
+          " and --height (-h)", stream->index);
+
+  // Check that the codec bit depth is greater than the input bit depth.
+  if (stream->config.cfg.g_input_bit_depth >
+      (unsigned int)stream->config.cfg.g_bit_depth) {
+    fatal("Stream %d: codec bit depth (%d) less than input bit depth (%d)",
+          stream->index, (int)stream->config.cfg.g_bit_depth,
+          stream->config.cfg.g_input_bit_depth);
+  }
+
+  for (streami = stream; streami; streami = streami->next) {
+    /* All streams require output files */
+    if (!streami->config.out_fn)
+      fatal("Stream %d: Output file is required (specify with -o)",
+            streami->index);
+
+    /* Check for two streams outputting to the same file */
+    if (streami != stream) {
+      const char *a = stream->config.out_fn;
+      const char *b = streami->config.out_fn;
+      if (!strcmp(a, b) && strcmp(a, "/dev/null") && strcmp(a, ":nul"))
+        fatal("Stream %d: duplicate output file (from stream %d)",
+              streami->index, stream->index);
+    }
+
+    /* Check for two streams sharing a stats file. */
+    if (streami != stream) {
+      const char *a = stream->config.stats_fn;
+      const char *b = streami->config.stats_fn;
+      if (a && b && !strcmp(a, b))
+        fatal("Stream %d: duplicate stats file (from stream %d)",
+              streami->index, stream->index);
+    }
+
+#if CONFIG_FP_MB_STATS
+    /* Check for two streams sharing a mb stats file. */
+    if (streami != stream) {
+      const char *a = stream->config.fpmb_stats_fn;
+      const char *b = streami->config.fpmb_stats_fn;
+      if (a && b && !strcmp(a, b))
+        fatal("Stream %d: duplicate mb stats file (from stream %d)",
+              streami->index, stream->index);
+    }
+#endif
+  }
+}
+
+
+static void set_stream_dimensions(struct stream_state *stream,
+                                  unsigned int w,
+                                  unsigned int h) {
+  if (!stream->config.cfg.g_w) {
+    if (!stream->config.cfg.g_h)
+      stream->config.cfg.g_w = w;
+    else
+      stream->config.cfg.g_w = w * stream->config.cfg.g_h / h;
+  }
+  if (!stream->config.cfg.g_h) {
+    stream->config.cfg.g_h = h * stream->config.cfg.g_w / w;
+  }
+}
+
+static const char* file_type_to_string(enum VideoFileType t) {
+  switch (t) {
+    case FILE_TYPE_RAW: return "RAW";
+    case FILE_TYPE_Y4M: return "Y4M";
+    default: return "Other";
+  }
+}
+
+static const char* image_format_to_string(vpx_img_fmt_t f) {
+  switch (f) {
+    case VPX_IMG_FMT_I420: return "I420";
+    case VPX_IMG_FMT_I422: return "I422";
+    case VPX_IMG_FMT_I444: return "I444";
+    case VPX_IMG_FMT_I440: return "I440";
+    case VPX_IMG_FMT_YV12: return "YV12";
+    case VPX_IMG_FMT_I42016: return "I42016";
+    case VPX_IMG_FMT_I42216: return "I42216";
+    case VPX_IMG_FMT_I44416: return "I44416";
+    case VPX_IMG_FMT_I44016: return "I44016";
+    default: return "Other";
+  }
+}
+
+static void show_stream_config(struct stream_state *stream,
+                               struct VpxEncoderConfig *global,
+                               struct VpxInputContext *input) {
+
+#define SHOW(field) \
+  fprintf(stderr, "    %-28s = %d\n", #field, stream->config.cfg.field)
+
+  if (stream->index == 0) {
+    fprintf(stderr, "Codec: %s\n",
+            vpx_codec_iface_name(global->codec->codec_interface()));
+    fprintf(stderr, "Source file: %s File Type: %s Format: %s\n",
+            input->filename,
+            file_type_to_string(input->file_type),
+            image_format_to_string(input->fmt));
+  }
+  if (stream->next || stream->index)
+    fprintf(stderr, "\nStream Index: %d\n", stream->index);
+  fprintf(stderr, "Destination file: %s\n", stream->config.out_fn);
+  fprintf(stderr, "Encoder parameters:\n");
+
+  SHOW(g_usage);
+  SHOW(g_threads);
+  SHOW(g_profile);
+  SHOW(g_w);
+  SHOW(g_h);
+  SHOW(g_bit_depth);
+  SHOW(g_input_bit_depth);
+  SHOW(g_timebase.num);
+  SHOW(g_timebase.den);
+  SHOW(g_error_resilient);
+  SHOW(g_pass);
+  SHOW(g_lag_in_frames);
+  SHOW(rc_dropframe_thresh);
+  SHOW(rc_resize_allowed);
+  SHOW(rc_scaled_width);
+  SHOW(rc_scaled_height);
+  SHOW(rc_resize_up_thresh);
+  SHOW(rc_resize_down_thresh);
+  SHOW(rc_end_usage);
+  SHOW(rc_target_bitrate);
+  SHOW(rc_min_quantizer);
+  SHOW(rc_max_quantizer);
+  SHOW(rc_undershoot_pct);
+  SHOW(rc_overshoot_pct);
+  SHOW(rc_buf_sz);
+  SHOW(rc_buf_initial_sz);
+  SHOW(rc_buf_optimal_sz);
+  SHOW(rc_2pass_vbr_bias_pct);
+  SHOW(rc_2pass_vbr_minsection_pct);
+  SHOW(rc_2pass_vbr_maxsection_pct);
+  SHOW(kf_mode);
+  SHOW(kf_min_dist);
+  SHOW(kf_max_dist);
+}
+
+
+static void open_output_file(struct stream_state *stream,
+                             struct VpxEncoderConfig *global,
+                             const struct VpxRational *pixel_aspect_ratio) {
+  const char *fn = stream->config.out_fn;
+  const struct vpx_codec_enc_cfg *const cfg = &stream->config.cfg;
+
+  if (cfg->g_pass == VPX_RC_FIRST_PASS)
+    return;
+
+  stream->file = strcmp(fn, "-") ? fopen(fn, "wb") : set_binary_mode(stdout);
+
+  if (!stream->file)
+    fatal("Failed to open output file");
+
+  if (stream->config.write_webm && fseek(stream->file, 0, SEEK_CUR))
+    fatal("WebM output to pipes not supported.");
+
+#if CONFIG_WEBM_IO
+  if (stream->config.write_webm) {
+    stream->webm_ctx.stream = stream->file;
+    write_webm_file_header(&stream->webm_ctx, cfg,
+                           &global->framerate,
+                           stream->config.stereo_fmt,
+                           global->codec->fourcc,
+                           pixel_aspect_ratio);
+  }
+#else
+  (void)pixel_aspect_ratio;
+#endif
+
+  if (!stream->config.write_webm) {
+    ivf_write_file_header(stream->file, cfg, global->codec->fourcc, 0);
+  }
+}
+
+
+static void close_output_file(struct stream_state *stream,
+                              unsigned int fourcc) {
+  const struct vpx_codec_enc_cfg *const cfg = &stream->config.cfg;
+
+  if (cfg->g_pass == VPX_RC_FIRST_PASS)
+    return;
+
+#if CONFIG_WEBM_IO
+  if (stream->config.write_webm) {
+    write_webm_file_footer(&stream->webm_ctx);
+  }
+#endif
+
+  if (!stream->config.write_webm) {
+    if (!fseek(stream->file, 0, SEEK_SET))
+      ivf_write_file_header(stream->file, &stream->config.cfg,
+                            fourcc,
+                            stream->frames_out);
+  }
+
+  fclose(stream->file);
+}
+
+
+static void setup_pass(struct stream_state *stream,
+                       struct VpxEncoderConfig *global,
+                       int pass) {
+  if (stream->config.stats_fn) {
+    if (!stats_open_file(&stream->stats, stream->config.stats_fn,
+                         pass))
+      fatal("Failed to open statistics store");
+  } else {
+    if (!stats_open_mem(&stream->stats, pass))
+      fatal("Failed to open statistics store");
+  }
+
+#if CONFIG_FP_MB_STATS
+  if (stream->config.fpmb_stats_fn) {
+    if (!stats_open_file(&stream->fpmb_stats,
+                         stream->config.fpmb_stats_fn, pass))
+      fatal("Failed to open mb statistics store");
+  } else {
+    if (!stats_open_mem(&stream->fpmb_stats, pass))
+      fatal("Failed to open mb statistics store");
+  }
+#endif
+
+  stream->config.cfg.g_pass = global->passes == 2
+                              ? pass ? VPX_RC_LAST_PASS : VPX_RC_FIRST_PASS
+                            : VPX_RC_ONE_PASS;
+  if (pass) {
+    stream->config.cfg.rc_twopass_stats_in = stats_get(&stream->stats);
+#if CONFIG_FP_MB_STATS
+    stream->config.cfg.rc_firstpass_mb_stats_in =
+        stats_get(&stream->fpmb_stats);
+#endif
+  }
+
+  stream->cx_time = 0;
+  stream->nbytes = 0;
+  stream->frames_out = 0;
+}
+
+
+static void initialize_encoder(struct stream_state *stream,
+                               struct VpxEncoderConfig *global) {
+  int i;
+  int flags = 0;
+
+  flags |= global->show_psnr ? VPX_CODEC_USE_PSNR : 0;
+  flags |= global->out_part ? VPX_CODEC_USE_OUTPUT_PARTITION : 0;
+#if CONFIG_VP9_HIGHBITDEPTH
+  flags |= stream->config.use_16bit_internal ? VPX_CODEC_USE_HIGHBITDEPTH : 0;
+#endif
+
+  /* Construct Encoder Context */
+  vpx_codec_enc_init(&stream->encoder, global->codec->codec_interface(),
+                     &stream->config.cfg, flags);
+  ctx_exit_on_error(&stream->encoder, "Failed to initialize encoder");
+
+  /* Note that we bypass the vpx_codec_control wrapper macro because
+   * we're being clever to store the control IDs in an array. Real
+   * applications will want to make use of the enumerations directly
+   */
+  for (i = 0; i < stream->config.arg_ctrl_cnt; i++) {
+    int ctrl = stream->config.arg_ctrls[i][0];
+    int value = stream->config.arg_ctrls[i][1];
+    if (vpx_codec_control_(&stream->encoder, ctrl, value))
+      fprintf(stderr, "Error: Tried to set control %d = %d\n",
+              ctrl, value);
+
+    ctx_exit_on_error(&stream->encoder, "Failed to control codec");
+  }
+
+#if CONFIG_DECODERS
+  if (global->test_decode != TEST_DECODE_OFF) {
+    const VpxInterface *decoder = get_vpx_decoder_by_name(global->codec->name);
+    vpx_codec_dec_init(&stream->decoder, decoder->codec_interface(), NULL, 0);
+  }
+#endif
+}
+
+
+static void encode_frame(struct stream_state *stream,
+                         struct VpxEncoderConfig *global,
+                         struct vpx_image *img,
+                         unsigned int frames_in) {
+  vpx_codec_pts_t frame_start, next_frame_start;
+  struct vpx_codec_enc_cfg *cfg = &stream->config.cfg;
+  struct vpx_usec_timer timer;
+
+  frame_start = (cfg->g_timebase.den * (int64_t)(frames_in - 1)
+                 * global->framerate.den)
+                / cfg->g_timebase.num / global->framerate.num;
+  next_frame_start = (cfg->g_timebase.den * (int64_t)(frames_in)
+                      * global->framerate.den)
+                     / cfg->g_timebase.num / global->framerate.num;
+
+  /* Scale if necessary */
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (img) {
+    if ((img->fmt & VPX_IMG_FMT_HIGHBITDEPTH) &&
+        (img->d_w != cfg->g_w || img->d_h != cfg->g_h)) {
+      if (img->fmt != VPX_IMG_FMT_I42016) {
+        fprintf(stderr, "%s can only scale 4:2:0 inputs\n", exec_name);
+        exit(EXIT_FAILURE);
+      }
+#if CONFIG_LIBYUV
+      if (!stream->img) {
+        stream->img = vpx_img_alloc(NULL, VPX_IMG_FMT_I42016,
+                                    cfg->g_w, cfg->g_h, 16);
+      }
+      I420Scale_16((uint16*)img->planes[VPX_PLANE_Y],
+                   img->stride[VPX_PLANE_Y]/2,
+                   (uint16*)img->planes[VPX_PLANE_U],
+                   img->stride[VPX_PLANE_U]/2,
+                   (uint16*)img->planes[VPX_PLANE_V],
+                   img->stride[VPX_PLANE_V]/2,
+                   img->d_w, img->d_h,
+                   (uint16*)stream->img->planes[VPX_PLANE_Y],
+                   stream->img->stride[VPX_PLANE_Y]/2,
+                   (uint16*)stream->img->planes[VPX_PLANE_U],
+                   stream->img->stride[VPX_PLANE_U]/2,
+                   (uint16*)stream->img->planes[VPX_PLANE_V],
+                   stream->img->stride[VPX_PLANE_V]/2,
+                   stream->img->d_w, stream->img->d_h,
+                   kFilterBox);
+      img = stream->img;
+#else
+    stream->encoder.err = 1;
+    ctx_exit_on_error(&stream->encoder,
+                      "Stream %d: Failed to encode frame.\n"
+                      "Scaling disabled in this configuration. \n"
+                      "To enable, configure with --enable-libyuv\n",
+                      stream->index);
+#endif
+    }
+  }
+#endif
+  if (img && (img->d_w != cfg->g_w || img->d_h != cfg->g_h)) {
+    if (img->fmt != VPX_IMG_FMT_I420 && img->fmt != VPX_IMG_FMT_YV12) {
+      fprintf(stderr, "%s can only scale 4:2:0 8bpp inputs\n", exec_name);
+      exit(EXIT_FAILURE);
+    }
+#if CONFIG_LIBYUV
+    if (!stream->img)
+      stream->img = vpx_img_alloc(NULL, VPX_IMG_FMT_I420,
+                                  cfg->g_w, cfg->g_h, 16);
+    I420Scale(img->planes[VPX_PLANE_Y], img->stride[VPX_PLANE_Y],
+              img->planes[VPX_PLANE_U], img->stride[VPX_PLANE_U],
+              img->planes[VPX_PLANE_V], img->stride[VPX_PLANE_V],
+              img->d_w, img->d_h,
+              stream->img->planes[VPX_PLANE_Y],
+              stream->img->stride[VPX_PLANE_Y],
+              stream->img->planes[VPX_PLANE_U],
+              stream->img->stride[VPX_PLANE_U],
+              stream->img->planes[VPX_PLANE_V],
+              stream->img->stride[VPX_PLANE_V],
+              stream->img->d_w, stream->img->d_h,
+              kFilterBox);
+    img = stream->img;
+#else
+    stream->encoder.err = 1;
+    ctx_exit_on_error(&stream->encoder,
+                      "Stream %d: Failed to encode frame.\n"
+                      "Scaling disabled in this configuration. \n"
+                      "To enable, configure with --enable-libyuv\n",
+                      stream->index);
+#endif
+  }
+
+  vpx_usec_timer_start(&timer);
+  vpx_codec_encode(&stream->encoder, img, frame_start,
+                   (unsigned long)(next_frame_start - frame_start),
+                   0, global->deadline);
+  vpx_usec_timer_mark(&timer);
+  stream->cx_time += vpx_usec_timer_elapsed(&timer);
+  ctx_exit_on_error(&stream->encoder, "Stream %d: Failed to encode frame",
+                    stream->index);
+}
+
+
+static void update_quantizer_histogram(struct stream_state *stream) {
+  if (stream->config.cfg.g_pass != VPX_RC_FIRST_PASS) {
+    int q;
+
+    vpx_codec_control(&stream->encoder, VP8E_GET_LAST_QUANTIZER_64, &q);
+    ctx_exit_on_error(&stream->encoder, "Failed to read quantizer");
+    stream->counts[q]++;
+  }
+}
+
+
+static void get_cx_data(struct stream_state *stream,
+                        struct VpxEncoderConfig *global,
+                        int *got_data) {
+  const vpx_codec_cx_pkt_t *pkt;
+  const struct vpx_codec_enc_cfg *cfg = &stream->config.cfg;
+  vpx_codec_iter_t iter = NULL;
+
+  *got_data = 0;
+  while ((pkt = vpx_codec_get_cx_data(&stream->encoder, &iter))) {
+    static size_t fsize = 0;
+    static int64_t ivf_header_pos = 0;
+
+    switch (pkt->kind) {
+      case VPX_CODEC_CX_FRAME_PKT:
+        if (!(pkt->data.frame.flags & VPX_FRAME_IS_FRAGMENT)) {
+          stream->frames_out++;
+        }
+        if (!global->quiet)
+          fprintf(stderr, " %6luF", (unsigned long)pkt->data.frame.sz);
+
+        update_rate_histogram(stream->rate_hist, cfg, pkt);
+#if CONFIG_WEBM_IO
+        if (stream->config.write_webm) {
+          write_webm_block(&stream->webm_ctx, cfg, pkt);
+        }
+#endif
+        if (!stream->config.write_webm) {
+          if (pkt->data.frame.partition_id <= 0) {
+            ivf_header_pos = ftello(stream->file);
+            fsize = pkt->data.frame.sz;
+
+            ivf_write_frame_header(stream->file, pkt->data.frame.pts, fsize);
+          } else {
+            fsize += pkt->data.frame.sz;
+
+            if (!(pkt->data.frame.flags & VPX_FRAME_IS_FRAGMENT)) {
+              const int64_t currpos = ftello(stream->file);
+              fseeko(stream->file, ivf_header_pos, SEEK_SET);
+              ivf_write_frame_size(stream->file, fsize);
+              fseeko(stream->file, currpos, SEEK_SET);
+            }
+          }
+
+          (void) fwrite(pkt->data.frame.buf, 1, pkt->data.frame.sz,
+                        stream->file);
+        }
+        stream->nbytes += pkt->data.raw.sz;
+
+        *got_data = 1;
+#if CONFIG_DECODERS
+        if (global->test_decode != TEST_DECODE_OFF && !stream->mismatch_seen) {
+          vpx_codec_decode(&stream->decoder, pkt->data.frame.buf,
+                           (unsigned int)pkt->data.frame.sz, NULL, 0);
+          if (stream->decoder.err) {
+            warn_or_exit_on_error(&stream->decoder,
+                                  global->test_decode == TEST_DECODE_FATAL,
+                                  "Failed to decode frame %d in stream %d",
+                                  stream->frames_out + 1, stream->index);
+            stream->mismatch_seen = stream->frames_out + 1;
+          }
+        }
+#endif
+        break;
+      case VPX_CODEC_STATS_PKT:
+        stream->frames_out++;
+        stats_write(&stream->stats,
+                    pkt->data.twopass_stats.buf,
+                    pkt->data.twopass_stats.sz);
+        stream->nbytes += pkt->data.raw.sz;
+        break;
+#if CONFIG_FP_MB_STATS
+      case VPX_CODEC_FPMB_STATS_PKT:
+        stats_write(&stream->fpmb_stats,
+                    pkt->data.firstpass_mb_stats.buf,
+                    pkt->data.firstpass_mb_stats.sz);
+        stream->nbytes += pkt->data.raw.sz;
+        break;
+#endif
+      case VPX_CODEC_PSNR_PKT:
+
+        if (global->show_psnr) {
+          int i;
+
+          stream->psnr_sse_total += pkt->data.psnr.sse[0];
+          stream->psnr_samples_total += pkt->data.psnr.samples[0];
+          for (i = 0; i < 4; i++) {
+            if (!global->quiet)
+              fprintf(stderr, "%.3f ", pkt->data.psnr.psnr[i]);
+            stream->psnr_totals[i] += pkt->data.psnr.psnr[i];
+          }
+          stream->psnr_count++;
+        }
+
+        break;
+      default:
+        break;
+    }
+  }
+}
+
+
+static void show_psnr(struct stream_state  *stream, double peak) {
+  int i;
+  double ovpsnr;
+
+  if (!stream->psnr_count)
+    return;
+
+  fprintf(stderr, "Stream %d PSNR (Overall/Avg/Y/U/V)", stream->index);
+  ovpsnr = sse_to_psnr((double)stream->psnr_samples_total, peak,
+                       (double)stream->psnr_sse_total);
+  fprintf(stderr, " %.3f", ovpsnr);
+
+  for (i = 0; i < 4; i++) {
+    fprintf(stderr, " %.3f", stream->psnr_totals[i] / stream->psnr_count);
+  }
+  fprintf(stderr, "\n");
+}
+
+
+static float usec_to_fps(uint64_t usec, unsigned int frames) {
+  return (float)(usec > 0 ? frames * 1000000.0 / (float)usec : 0);
+}
+
+static void test_decode(struct stream_state  *stream,
+                        enum TestDecodeFatality fatal,
+                        const VpxInterface *codec) {
+  vpx_image_t enc_img, dec_img;
+
+  if (stream->mismatch_seen)
+    return;
+
+  /* Get the internal reference frame */
+  if (strcmp(codec->name, "vp8") == 0) {
+    struct vpx_ref_frame ref_enc, ref_dec;
+    int width, height;
+
+    width = (stream->config.cfg.g_w + 15) & ~15;
+    height = (stream->config.cfg.g_h + 15) & ~15;
+    vpx_img_alloc(&ref_enc.img, VPX_IMG_FMT_I420, width, height, 1);
+    enc_img = ref_enc.img;
+    vpx_img_alloc(&ref_dec.img, VPX_IMG_FMT_I420, width, height, 1);
+    dec_img = ref_dec.img;
+
+    ref_enc.frame_type = VP8_LAST_FRAME;
+    ref_dec.frame_type = VP8_LAST_FRAME;
+    vpx_codec_control(&stream->encoder, VP8_COPY_REFERENCE, &ref_enc);
+    vpx_codec_control(&stream->decoder, VP8_COPY_REFERENCE, &ref_dec);
+  } else {
+    struct vp9_ref_frame ref_enc, ref_dec;
+
+    ref_enc.idx = 0;
+    ref_dec.idx = 0;
+    vpx_codec_control(&stream->encoder, VP9_GET_REFERENCE, &ref_enc);
+    enc_img = ref_enc.img;
+    vpx_codec_control(&stream->decoder, VP9_GET_REFERENCE, &ref_dec);
+    dec_img = ref_dec.img;
+#if CONFIG_VP9_HIGHBITDEPTH
+    if ((enc_img.fmt & VPX_IMG_FMT_HIGHBITDEPTH) !=
+        (dec_img.fmt & VPX_IMG_FMT_HIGHBITDEPTH)) {
+      if (enc_img.fmt & VPX_IMG_FMT_HIGHBITDEPTH) {
+        vpx_img_alloc(&enc_img, enc_img.fmt - VPX_IMG_FMT_HIGHBITDEPTH,
+                      enc_img.d_w, enc_img.d_h, 16);
+        vpx_img_truncate_16_to_8(&enc_img, &ref_enc.img);
+      }
+      if (dec_img.fmt & VPX_IMG_FMT_HIGHBITDEPTH) {
+        vpx_img_alloc(&dec_img, dec_img.fmt - VPX_IMG_FMT_HIGHBITDEPTH,
+                      dec_img.d_w, dec_img.d_h, 16);
+        vpx_img_truncate_16_to_8(&dec_img, &ref_dec.img);
+      }
+    }
+#endif
+  }
+  ctx_exit_on_error(&stream->encoder, "Failed to get encoder reference frame");
+  ctx_exit_on_error(&stream->decoder, "Failed to get decoder reference frame");
+
+  if (!compare_img(&enc_img, &dec_img)) {
+    int y[4], u[4], v[4];
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (enc_img.fmt & VPX_IMG_FMT_HIGHBITDEPTH) {
+      find_mismatch_high(&enc_img, &dec_img, y, u, v);
+    } else {
+      find_mismatch(&enc_img, &dec_img, y, u, v);
+    }
+#else
+    find_mismatch(&enc_img, &dec_img, y, u, v);
+#endif
+    stream->decoder.err = 1;
+    warn_or_exit_on_error(&stream->decoder, fatal == TEST_DECODE_FATAL,
+                          "Stream %d: Encode/decode mismatch on frame %d at"
+                          " Y[%d, %d] {%d/%d},"
+                          " U[%d, %d] {%d/%d},"
+                          " V[%d, %d] {%d/%d}",
+                          stream->index, stream->frames_out,
+                          y[0], y[1], y[2], y[3],
+                          u[0], u[1], u[2], u[3],
+                          v[0], v[1], v[2], v[3]);
+    stream->mismatch_seen = stream->frames_out;
+  }
+
+  vpx_img_free(&enc_img);
+  vpx_img_free(&dec_img);
+}
+
+
+static void print_time(const char *label, int64_t etl) {
+  int64_t hours;
+  int64_t mins;
+  int64_t secs;
+
+  if (etl >= 0) {
+    hours = etl / 3600;
+    etl -= hours * 3600;
+    mins = etl / 60;
+    etl -= mins * 60;
+    secs = etl;
+
+    fprintf(stderr, "[%3s %2"PRId64":%02"PRId64":%02"PRId64"] ",
+            label, hours, mins, secs);
+  } else {
+    fprintf(stderr, "[%3s  unknown] ", label);
+  }
+}
+
+
+int main(int argc, const char **argv_) {
+  int pass;
+  vpx_image_t raw;
+#if CONFIG_VP9_HIGHBITDEPTH
+  vpx_image_t raw_shift;
+  int allocated_raw_shift = 0;
+  int use_16bit_internal = 0;
+  int input_shift = 0;
+#endif
+  int frame_avail, got_data;
+
+  struct VpxInputContext input;
+  struct VpxEncoderConfig global;
+  struct stream_state *streams = NULL;
+  char **argv, **argi;
+  uint64_t cx_time = 0;
+  int stream_cnt = 0;
+  int res = 0;
+
+  memset(&input, 0, sizeof(input));
+  exec_name = argv_[0];
+
+  if (argc < 3)
+    usage_exit();
+
+  /* Setup default input stream settings */
+  input.framerate.numerator = 30;
+  input.framerate.denominator = 1;
+  input.only_i420 = 1;
+  input.bit_depth = 0;
+
+  /* First parse the global configuration values, because we want to apply
+   * other parameters on top of the default configuration provided by the
+   * codec.
+   */
+  argv = argv_dup(argc - 1, argv_ + 1);
+  parse_global_config(&global, argv);
+
+  switch (global.color_type) {
+    case I420:
+      input.fmt = VPX_IMG_FMT_I420;
+      break;
+    case I422:
+      input.fmt = VPX_IMG_FMT_I422;
+      break;
+    case I444:
+      input.fmt = VPX_IMG_FMT_I444;
+      break;
+    case I440:
+      input.fmt = VPX_IMG_FMT_I440;
+      break;
+    case YV12:
+      input.fmt = VPX_IMG_FMT_YV12;
+      break;
+  }
+
+  {
+    /* Now parse each stream's parameters. Using a local scope here
+     * due to the use of 'stream' as loop variable in FOREACH_STREAM
+     * loops
+     */
+    struct stream_state *stream = NULL;
+
+    do {
+      stream = new_stream(&global, stream);
+      stream_cnt++;
+      if (!streams)
+        streams = stream;
+    } while (parse_stream_params(&global, stream, argv));
+  }
+
+  /* Check for unrecognized options */
+  for (argi = argv; *argi; argi++)
+    if (argi[0][0] == '-' && argi[0][1])
+      die("Error: Unrecognized option %s\n", *argi);
+
+  FOREACH_STREAM(check_encoder_config(global.disable_warning_prompt,
+                                      &global, &stream->config.cfg););
+
+  /* Handle non-option arguments */
+  input.filename = argv[0];
+
+  if (!input.filename)
+    usage_exit();
+
+  /* Decide if other chroma subsamplings than 4:2:0 are supported */
+  if (global.codec->fourcc == VP9_FOURCC)
+    input.only_i420 = 0;
+
+  for (pass = global.pass ? global.pass - 1 : 0; pass < global.passes; pass++) {
+    int frames_in = 0, seen_frames = 0;
+    int64_t estimated_time_left = -1;
+    int64_t average_rate = -1;
+    int64_t lagged_count = 0;
+
+    open_input_file(&input);
+
+    /* If the input file doesn't specify its w/h (raw files), try to get
+     * the data from the first stream's configuration.
+     */
+    if (!input.width || !input.height) {
+      FOREACH_STREAM({
+        if (stream->config.cfg.g_w && stream->config.cfg.g_h) {
+          input.width = stream->config.cfg.g_w;
+          input.height = stream->config.cfg.g_h;
+          break;
+        }
+      });
+    }
+
+    /* Update stream configurations from the input file's parameters */
+    if (!input.width || !input.height)
+      fatal("Specify stream dimensions with --width (-w) "
+            " and --height (-h)");
+
+    /* If input file does not specify bit-depth but input-bit-depth parameter
+     * exists, assume that to be the input bit-depth. However, if the
+     * input-bit-depth paramter does not exist, assume the input bit-depth
+     * to be the same as the codec bit-depth.
+     */
+    if (!input.bit_depth) {
+      FOREACH_STREAM({
+        if (stream->config.cfg.g_input_bit_depth)
+          input.bit_depth = stream->config.cfg.g_input_bit_depth;
+        else
+          input.bit_depth = stream->config.cfg.g_input_bit_depth =
+              (int)stream->config.cfg.g_bit_depth;
+      });
+      if (input.bit_depth > 8) input.fmt |= VPX_IMG_FMT_HIGHBITDEPTH;
+    } else {
+      FOREACH_STREAM({
+        stream->config.cfg.g_input_bit_depth = input.bit_depth;
+      });
+    }
+
+    FOREACH_STREAM(set_stream_dimensions(stream, input.width, input.height));
+    FOREACH_STREAM(validate_stream_config(stream, &global));
+
+    /* Ensure that --passes and --pass are consistent. If --pass is set and
+     * --passes=2, ensure --fpf was set.
+     */
+    if (global.pass && global.passes == 2)
+      FOREACH_STREAM( {
+      if (!stream->config.stats_fn)
+        die("Stream %d: Must specify --fpf when --pass=%d"
+        " and --passes=2\n", stream->index, global.pass);
+    });
+
+#if !CONFIG_WEBM_IO
+    FOREACH_STREAM({
+      if (stream->config.write_webm) {
+        stream->config.write_webm = 0;
+        warn("vpxenc was compiled without WebM container support."
+             "Producing IVF output");
+      }
+    });
+#endif
+
+    /* Use the frame rate from the file only if none was specified
+     * on the command-line.
+     */
+    if (!global.have_framerate) {
+      global.framerate.num = input.framerate.numerator;
+      global.framerate.den = input.framerate.denominator;
+      FOREACH_STREAM(stream->config.cfg.g_timebase.den = global.framerate.num;
+                     stream->config.cfg.g_timebase.num = global.framerate.den);
+    }
+
+    /* Show configuration */
+    if (global.verbose && pass == 0)
+      FOREACH_STREAM(show_stream_config(stream, &global, &input));
+
+    if (pass == (global.pass ? global.pass - 1 : 0)) {
+      if (input.file_type == FILE_TYPE_Y4M)
+        /*The Y4M reader does its own allocation.
+          Just initialize this here to avoid problems if we never read any
+           frames.*/
+        memset(&raw, 0, sizeof(raw));
+      else
+        vpx_img_alloc(&raw, input.fmt, input.width, input.height, 32);
+
+      FOREACH_STREAM(stream->rate_hist =
+                         init_rate_histogram(&stream->config.cfg,
+                                             &global.framerate));
+    }
+
+    FOREACH_STREAM(setup_pass(stream, &global, pass));
+    FOREACH_STREAM(open_output_file(stream, &global,
+                                    &input.pixel_aspect_ratio));
+    FOREACH_STREAM(initialize_encoder(stream, &global));
+
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (strcmp(global.codec->name, "vp9") == 0) {
+      // Check to see if at least one stream uses 16 bit internal.
+      // Currently assume that the bit_depths for all streams using
+      // highbitdepth are the same.
+      FOREACH_STREAM({
+        if (stream->config.use_16bit_internal) {
+          use_16bit_internal = 1;
+        }
+        if (stream->config.cfg.g_profile == 0) {
+          input_shift = 0;
+        } else {
+          input_shift = (int)stream->config.cfg.g_bit_depth -
+              stream->config.cfg.g_input_bit_depth;
+        }
+      });
+    }
+#endif
+
+    frame_avail = 1;
+    got_data = 0;
+
+    while (frame_avail || got_data) {
+      struct vpx_usec_timer timer;
+
+      if (!global.limit || frames_in < global.limit) {
+        frame_avail = read_frame(&input, &raw);
+
+        if (frame_avail)
+          frames_in++;
+        seen_frames = frames_in > global.skip_frames ?
+                          frames_in - global.skip_frames : 0;
+
+        if (!global.quiet) {
+          float fps = usec_to_fps(cx_time, seen_frames);
+          fprintf(stderr, "\rPass %d/%d ", pass + 1, global.passes);
+
+          if (stream_cnt == 1)
+            fprintf(stderr,
+                    "frame %4d/%-4d %7"PRId64"B ",
+                    frames_in, streams->frames_out, (int64_t)streams->nbytes);
+          else
+            fprintf(stderr, "frame %4d ", frames_in);
+
+          fprintf(stderr, "%7"PRId64" %s %.2f %s ",
+                  cx_time > 9999999 ? cx_time / 1000 : cx_time,
+                  cx_time > 9999999 ? "ms" : "us",
+                  fps >= 1.0 ? fps : fps * 60,
+                  fps >= 1.0 ? "fps" : "fpm");
+          print_time("ETA", estimated_time_left);
+        }
+
+      } else
+        frame_avail = 0;
+
+      if (frames_in > global.skip_frames) {
+#if CONFIG_VP9_HIGHBITDEPTH
+        vpx_image_t *frame_to_encode;
+        if (input_shift || (use_16bit_internal && input.bit_depth == 8)) {
+          assert(use_16bit_internal);
+          // Input bit depth and stream bit depth do not match, so up
+          // shift frame to stream bit depth
+          if (!allocated_raw_shift) {
+            vpx_img_alloc(&raw_shift, raw.fmt | VPX_IMG_FMT_HIGHBITDEPTH,
+                          input.width, input.height, 32);
+            allocated_raw_shift = 1;
+          }
+          vpx_img_upshift(&raw_shift, &raw, input_shift);
+          frame_to_encode = &raw_shift;
+        } else {
+          frame_to_encode = &raw;
+        }
+        vpx_usec_timer_start(&timer);
+        if (use_16bit_internal) {
+          assert(frame_to_encode->fmt & VPX_IMG_FMT_HIGHBITDEPTH);
+          FOREACH_STREAM({
+            if (stream->config.use_16bit_internal)
+              encode_frame(stream, &global,
+                           frame_avail ? frame_to_encode : NULL,
+                           frames_in);
+            else
+              assert(0);
+          });
+        } else {
+          assert((frame_to_encode->fmt & VPX_IMG_FMT_HIGHBITDEPTH) == 0);
+          FOREACH_STREAM(encode_frame(stream, &global,
+                                      frame_avail ? frame_to_encode : NULL,
+                                      frames_in));
+        }
+#else
+        vpx_usec_timer_start(&timer);
+        FOREACH_STREAM(encode_frame(stream, &global,
+                                    frame_avail ? &raw : NULL,
+                                    frames_in));
+#endif
+        vpx_usec_timer_mark(&timer);
+        cx_time += vpx_usec_timer_elapsed(&timer);
+
+        FOREACH_STREAM(update_quantizer_histogram(stream));
+
+        got_data = 0;
+        FOREACH_STREAM(get_cx_data(stream, &global, &got_data));
+
+        if (!got_data && input.length && streams != NULL &&
+            !streams->frames_out) {
+          lagged_count = global.limit ? seen_frames : ftello(input.file);
+        } else if (input.length) {
+          int64_t remaining;
+          int64_t rate;
+
+          if (global.limit) {
+            const int64_t frame_in_lagged = (seen_frames - lagged_count) * 1000;
+
+            rate = cx_time ? frame_in_lagged * (int64_t)1000000 / cx_time : 0;
+            remaining = 1000 * (global.limit - global.skip_frames
+                                - seen_frames + lagged_count);
+          } else {
+            const int64_t input_pos = ftello(input.file);
+            const int64_t input_pos_lagged = input_pos - lagged_count;
+            const int64_t limit = input.length;
+
+            rate = cx_time ? input_pos_lagged * (int64_t)1000000 / cx_time : 0;
+            remaining = limit - input_pos + lagged_count;
+          }
+
+          average_rate = (average_rate <= 0)
+              ? rate
+              : (average_rate * 7 + rate) / 8;
+          estimated_time_left = average_rate ? remaining / average_rate : -1;
+        }
+
+        if (got_data && global.test_decode != TEST_DECODE_OFF)
+          FOREACH_STREAM(test_decode(stream, global.test_decode, global.codec));
+      }
+
+      fflush(stdout);
+      if (!global.quiet)
+        fprintf(stderr, "\033[K");
+    }
+
+    if (stream_cnt > 1)
+      fprintf(stderr, "\n");
+
+    if (!global.quiet) {
+      FOREACH_STREAM(fprintf(stderr,
+          "\rPass %d/%d frame %4d/%-4d %7"PRId64"B %7"PRId64"b/f %7"PRId64"b/s"
+          " %7"PRId64" %s (%.2f fps)\033[K\n",
+          pass + 1,
+          global.passes, frames_in, stream->frames_out, (int64_t)stream->nbytes,
+          seen_frames ? (int64_t)(stream->nbytes * 8 / seen_frames) : 0,
+          seen_frames ? (int64_t)stream->nbytes * 8 *
+              (int64_t)global.framerate.num / global.framerate.den /
+              seen_frames : 0,
+          stream->cx_time > 9999999 ? stream->cx_time / 1000 : stream->cx_time,
+          stream->cx_time > 9999999 ? "ms" : "us",
+          usec_to_fps(stream->cx_time, seen_frames)));
+    }
+
+    if (global.show_psnr) {
+      if (global.codec->fourcc == VP9_FOURCC) {
+        FOREACH_STREAM(
+            show_psnr(stream, (1 << stream->config.cfg.g_input_bit_depth) - 1));
+      } else {
+        FOREACH_STREAM(show_psnr(stream, 255.0));
+      }
+    }
+
+    FOREACH_STREAM(vpx_codec_destroy(&stream->encoder));
+
+    if (global.test_decode != TEST_DECODE_OFF) {
+      FOREACH_STREAM(vpx_codec_destroy(&stream->decoder));
+    }
+
+    close_input_file(&input);
+
+    if (global.test_decode == TEST_DECODE_FATAL) {
+      FOREACH_STREAM(res |= stream->mismatch_seen);
+    }
+    FOREACH_STREAM(close_output_file(stream, global.codec->fourcc));
+
+    FOREACH_STREAM(stats_close(&stream->stats, global.passes - 1));
+
+#if CONFIG_FP_MB_STATS
+    FOREACH_STREAM(stats_close(&stream->fpmb_stats, global.passes - 1));
+#endif
+
+    if (global.pass)
+      break;
+  }
+
+  if (global.show_q_hist_buckets)
+    FOREACH_STREAM(show_q_histogram(stream->counts,
+                                    global.show_q_hist_buckets));
+
+  if (global.show_rate_hist_buckets)
+    FOREACH_STREAM(show_rate_histogram(stream->rate_hist,
+                                       &stream->config.cfg,
+                                       global.show_rate_hist_buckets));
+  FOREACH_STREAM(destroy_rate_histogram(stream->rate_hist));
+
+#if CONFIG_INTERNAL_STATS
+  /* TODO(jkoleszar): This doesn't belong in this executable. Do it for now,
+   * to match some existing utilities.
+   */
+  if (!(global.pass == 1 && global.passes == 2))
+    FOREACH_STREAM({
+      FILE *f = fopen("opsnr.stt", "a");
+      if (stream->mismatch_seen) {
+        fprintf(f, "First mismatch occurred in frame %d\n",
+                stream->mismatch_seen);
+      } else {
+        fprintf(f, "No mismatch detected in recon buffers\n");
+      }
+      fclose(f);
+    });
+#endif
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (allocated_raw_shift)
+    vpx_img_free(&raw_shift);
+#endif
+  vpx_img_free(&raw);
+  free(argv);
+  free(streams);
+  return res ? EXIT_FAILURE : EXIT_SUCCESS;
+}
diff --git a/libvpx/libvpx/vpxenc.h b/libvpx/libvpx/vpxenc.h
new file mode 100644
index 0000000..d867e9d
--- /dev/null
+++ b/libvpx/libvpx/vpxenc.h
@@ -0,0 +1,64 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VPXENC_H_
+#define VPXENC_H_
+
+#include "vpx/vpx_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum TestDecodeFatality {
+  TEST_DECODE_OFF,
+  TEST_DECODE_FATAL,
+  TEST_DECODE_WARN,
+};
+
+typedef enum {
+  I420,  // 4:2:0 8+ bit-depth
+  I422,  // 4:2:2 8+ bit-depth
+  I444,  // 4:4:4 8+ bit-depth
+  I440,  // 4:4:0 8+ bit-depth
+  YV12,  // 4:2:0 with uv flipped, only 8-bit depth
+} ColorInputType;
+
+struct VpxInterface;
+
+/* Configuration elements common to all streams. */
+struct VpxEncoderConfig {
+  const struct VpxInterface *codec;
+  int passes;
+  int pass;
+  int usage;
+  int deadline;
+  ColorInputType color_type;
+  int quiet;
+  int verbose;
+  int limit;
+  int skip_frames;
+  int show_psnr;
+  enum TestDecodeFatality test_decode;
+  int have_framerate;
+  struct vpx_rational framerate;
+  int out_part;
+  int debug;
+  int show_q_hist_buckets;
+  int show_rate_hist_buckets;
+  int disable_warnings;
+  int disable_warning_prompt;
+  int experimental_bitstream;
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPXENC_H_
diff --git a/libvpx/libvpx/vpxstats.c b/libvpx/libvpx/vpxstats.c
new file mode 100644
index 0000000..16728ce
--- /dev/null
+++ b/libvpx/libvpx/vpxstats.c
@@ -0,0 +1,109 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpxstats.h"
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "./tools_common.h"
+
+int stats_open_file(stats_io_t *stats, const char *fpf, int pass) {
+  int res;
+  stats->pass = pass;
+
+  if (pass == 0) {
+    stats->file = fopen(fpf, "wb");
+    stats->buf.sz = 0;
+    stats->buf.buf = NULL;
+    res = (stats->file != NULL);
+  } else {
+    size_t nbytes;
+
+    stats->file = fopen(fpf, "rb");
+
+    if (stats->file == NULL)
+      fatal("First-pass stats file does not exist!");
+
+    if (fseek(stats->file, 0, SEEK_END))
+      fatal("First-pass stats file must be seekable!");
+
+    stats->buf.sz = stats->buf_alloc_sz = ftell(stats->file);
+    rewind(stats->file);
+
+    stats->buf.buf = malloc(stats->buf_alloc_sz);
+
+    if (!stats->buf.buf)
+      fatal("Failed to allocate first-pass stats buffer (%lu bytes)",
+            (unsigned int)stats->buf_alloc_sz);
+
+    nbytes = fread(stats->buf.buf, 1, stats->buf.sz, stats->file);
+    res = (nbytes == stats->buf.sz);
+  }
+
+  return res;
+}
+
+int stats_open_mem(stats_io_t *stats, int pass) {
+  int res;
+  stats->pass = pass;
+
+  if (!pass) {
+    stats->buf.sz = 0;
+    stats->buf_alloc_sz = 64 * 1024;
+    stats->buf.buf = malloc(stats->buf_alloc_sz);
+  }
+
+  stats->buf_ptr = stats->buf.buf;
+  res = (stats->buf.buf != NULL);
+  return res;
+}
+
+void stats_close(stats_io_t *stats, int last_pass) {
+  if (stats->file) {
+    if (stats->pass == last_pass) {
+      free(stats->buf.buf);
+    }
+
+    fclose(stats->file);
+    stats->file = NULL;
+  } else {
+    if (stats->pass == last_pass)
+      free(stats->buf.buf);
+  }
+}
+
+void stats_write(stats_io_t *stats, const void *pkt, size_t len) {
+  if (stats->file) {
+    (void) fwrite(pkt, 1, len, stats->file);
+  } else {
+    if (stats->buf.sz + len > stats->buf_alloc_sz) {
+      size_t  new_sz = stats->buf_alloc_sz + 64 * 1024;
+      char   *new_ptr = realloc(stats->buf.buf, new_sz);
+
+      if (new_ptr) {
+        stats->buf_ptr = new_ptr + (stats->buf_ptr - (char *)stats->buf.buf);
+        stats->buf.buf = new_ptr;
+        stats->buf_alloc_sz = new_sz;
+      } else {
+        fatal("Failed to realloc firstpass stats buffer.");
+      }
+    }
+
+    memcpy(stats->buf_ptr, pkt, len);
+    stats->buf.sz += len;
+    stats->buf_ptr += len;
+  }
+}
+
+vpx_fixed_buf_t stats_get(stats_io_t *stats) {
+  return stats->buf;
+}
diff --git a/libvpx/libvpx/vpxstats.h b/libvpx/libvpx/vpxstats.h
new file mode 100644
index 0000000..5c9ea34
--- /dev/null
+++ b/libvpx/libvpx/vpxstats.h
@@ -0,0 +1,43 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPXSTATS_H_
+#define VPXSTATS_H_
+
+#include <stdio.h>
+
+#include "vpx/vpx_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* This structure is used to abstract the different ways of handling
+ * first pass statistics
+ */
+typedef struct {
+  vpx_fixed_buf_t buf;
+  int pass;
+  FILE *file;
+  char *buf_ptr;
+  size_t buf_alloc_sz;
+} stats_io_t;
+
+int stats_open_file(stats_io_t *stats, const char *fpf, int pass);
+int stats_open_mem(stats_io_t *stats, int pass);
+void stats_close(stats_io_t *stats, int last_pass);
+void stats_write(stats_io_t *stats, const void *pkt, size_t len);
+vpx_fixed_buf_t stats_get(stats_io_t *stats);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VPXSTATS_H_
diff --git a/libvpx/libvpx/warnings.c b/libvpx/libvpx/warnings.c
new file mode 100644
index 0000000..7ac678a
--- /dev/null
+++ b/libvpx/libvpx/warnings.c
@@ -0,0 +1,115 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./warnings.h"
+
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vpx_encoder.h"
+
+#include "./tools_common.h"
+#include "./vpxenc.h"
+
+static const char quantizer_warning_string[] =
+    "Bad quantizer values. Quantizer values should not be equal, and should "
+    "differ by at least 8.";
+static const char lag_in_frames_with_realtime[] =
+    "Lag in frames is ignored when deadline is set to realtime.";
+
+struct WarningListNode {
+  const char *warning_string;
+  struct WarningListNode *next_warning;
+};
+
+struct WarningList {
+  struct WarningListNode *warning_node;
+};
+
+static void add_warning(const char *warning_string,
+                        struct WarningList *warning_list) {
+  struct WarningListNode **node = &warning_list->warning_node;
+
+  struct WarningListNode *new_node = malloc(sizeof(*new_node));
+  if (new_node == NULL) {
+    fatal("Unable to allocate warning node.");
+  }
+
+  new_node->warning_string = warning_string;
+  new_node->next_warning = NULL;
+
+  while (*node != NULL)
+    node = &(*node)->next_warning;
+
+  *node = new_node;
+}
+
+static void free_warning_list(struct WarningList *warning_list) {
+  while (warning_list->warning_node != NULL) {
+    struct WarningListNode *const node = warning_list->warning_node;
+    warning_list->warning_node = node->next_warning;
+    free(node);
+  }
+}
+
+static int continue_prompt(int num_warnings) {
+  int c;
+  fprintf(stderr,
+          "%d encoder configuration warning(s). Continue? (y to continue) ",
+          num_warnings);
+  c = getchar();
+  return c == 'y';
+}
+
+static void check_quantizer(int min_q, int max_q,
+                            struct WarningList *warning_list) {
+  const int lossless = min_q == 0 && max_q == 0;
+  if (!lossless && (min_q == max_q || abs(max_q - min_q) < 8))
+    add_warning(quantizer_warning_string, warning_list);
+}
+
+static void check_lag_in_frames_realtime_deadline(
+    int lag_in_frames,
+    int deadline,
+    struct WarningList *warning_list) {
+  if (deadline == VPX_DL_REALTIME && lag_in_frames != 0)
+    add_warning(lag_in_frames_with_realtime, warning_list);
+}
+
+void check_encoder_config(int disable_prompt,
+                          const struct VpxEncoderConfig *global_config,
+                          const struct vpx_codec_enc_cfg *stream_config) {
+  int num_warnings = 0;
+  struct WarningListNode *warning = NULL;
+  struct WarningList warning_list = {0};
+
+  check_quantizer(stream_config->rc_min_quantizer,
+                  stream_config->rc_max_quantizer,
+                  &warning_list);
+  check_lag_in_frames_realtime_deadline(stream_config->g_lag_in_frames,
+                                        global_config->deadline,
+                                        &warning_list);
+  /* Count and print warnings. */
+  for (warning = warning_list.warning_node;
+       warning != NULL;
+       warning = warning->next_warning,
+       ++num_warnings) {
+    warn(warning->warning_string);
+  }
+
+  free_warning_list(&warning_list);
+
+  if (num_warnings) {
+    if (!disable_prompt && !continue_prompt(num_warnings))
+      exit(EXIT_FAILURE);
+  }
+}
diff --git a/libvpx/libvpx/warnings.h b/libvpx/libvpx/warnings.h
new file mode 100644
index 0000000..6b8ae67
--- /dev/null
+++ b/libvpx/libvpx/warnings.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef WARNINGS_H_
+#define WARNINGS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct vpx_codec_enc_cfg;
+struct VpxEncoderConfig;
+
+/*
+ * Checks config for improperly used settings. Warns user upon encountering
+ * settings that will lead to poor output quality. Prompts user to continue
+ * when warnings are issued.
+ */
+void check_encoder_config(int disable_prompt,
+                          const struct VpxEncoderConfig *global_config,
+                          const struct vpx_codec_enc_cfg *stream_config);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // WARNINGS_H_
diff --git a/libvpx/libvpx/webmdec.cc b/libvpx/libvpx/webmdec.cc
new file mode 100644
index 0000000..36dbd92
--- /dev/null
+++ b/libvpx/libvpx/webmdec.cc
@@ -0,0 +1,226 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./webmdec.h"
+
+#include <cstring>
+#include <cstdio>
+
+#include "third_party/libwebm/mkvparser/mkvparser.h"
+#include "third_party/libwebm/mkvparser/mkvreader.h"
+
+namespace {
+
+void reset(struct WebmInputContext *const webm_ctx) {
+  if (webm_ctx->reader != NULL) {
+    mkvparser::MkvReader *const reader =
+        reinterpret_cast<mkvparser::MkvReader*>(webm_ctx->reader);
+    delete reader;
+  }
+  if (webm_ctx->segment != NULL) {
+    mkvparser::Segment *const segment =
+        reinterpret_cast<mkvparser::Segment*>(webm_ctx->segment);
+    delete segment;
+  }
+  if (webm_ctx->buffer != NULL) {
+    delete[] webm_ctx->buffer;
+  }
+  webm_ctx->reader = NULL;
+  webm_ctx->segment = NULL;
+  webm_ctx->buffer = NULL;
+  webm_ctx->cluster = NULL;
+  webm_ctx->block_entry = NULL;
+  webm_ctx->block = NULL;
+  webm_ctx->block_frame_index = 0;
+  webm_ctx->video_track_index = 0;
+  webm_ctx->timestamp_ns = 0;
+  webm_ctx->is_key_frame = false;
+}
+
+void get_first_cluster(struct WebmInputContext *const webm_ctx) {
+  mkvparser::Segment *const segment =
+      reinterpret_cast<mkvparser::Segment*>(webm_ctx->segment);
+  const mkvparser::Cluster *const cluster = segment->GetFirst();
+  webm_ctx->cluster = cluster;
+}
+
+void rewind_and_reset(struct WebmInputContext *const webm_ctx,
+                      struct VpxInputContext *const vpx_ctx) {
+  rewind(vpx_ctx->file);
+  reset(webm_ctx);
+}
+
+}  // namespace
+
+int file_is_webm(struct WebmInputContext *webm_ctx,
+                 struct VpxInputContext *vpx_ctx) {
+  mkvparser::MkvReader *const reader = new mkvparser::MkvReader(vpx_ctx->file);
+  webm_ctx->reader = reader;
+  webm_ctx->reached_eos = 0;
+
+  mkvparser::EBMLHeader header;
+  long long pos = 0;
+  if (header.Parse(reader, pos) < 0) {
+    rewind_and_reset(webm_ctx, vpx_ctx);
+    return 0;
+  }
+
+  mkvparser::Segment* segment;
+  if (mkvparser::Segment::CreateInstance(reader, pos, segment)) {
+    rewind_and_reset(webm_ctx, vpx_ctx);
+    return 0;
+  }
+  webm_ctx->segment = segment;
+  if (segment->Load() < 0) {
+    rewind_and_reset(webm_ctx, vpx_ctx);
+    return 0;
+  }
+
+  const mkvparser::Tracks *const tracks = segment->GetTracks();
+  const mkvparser::VideoTrack* video_track = NULL;
+  for (unsigned long i = 0; i < tracks->GetTracksCount(); ++i) {
+    const mkvparser::Track* const track = tracks->GetTrackByIndex(i);
+    if (track->GetType() == mkvparser::Track::kVideo) {
+      video_track = static_cast<const mkvparser::VideoTrack*>(track);
+      webm_ctx->video_track_index = track->GetNumber();
+      break;
+    }
+  }
+
+  if (video_track == NULL || video_track->GetCodecId() == NULL) {
+    rewind_and_reset(webm_ctx, vpx_ctx);
+    return 0;
+  }
+
+  if (!strncmp(video_track->GetCodecId(), "V_VP8", 5)) {
+    vpx_ctx->fourcc = VP8_FOURCC;
+  } else if (!strncmp(video_track->GetCodecId(), "V_VP9", 5)) {
+    vpx_ctx->fourcc = VP9_FOURCC;
+  } else {
+    rewind_and_reset(webm_ctx, vpx_ctx);
+    return 0;
+  }
+
+  vpx_ctx->framerate.denominator = 0;
+  vpx_ctx->framerate.numerator = 0;
+  vpx_ctx->width = static_cast<uint32_t>(video_track->GetWidth());
+  vpx_ctx->height = static_cast<uint32_t>(video_track->GetHeight());
+
+  get_first_cluster(webm_ctx);
+
+  return 1;
+}
+
+int webm_read_frame(struct WebmInputContext *webm_ctx,
+                    uint8_t **buffer,
+                    size_t *buffer_size) {
+  // This check is needed for frame parallel decoding, in which case this
+  // function could be called even after it has reached end of input stream.
+  if (webm_ctx->reached_eos) {
+    return 1;
+  }
+  mkvparser::Segment *const segment =
+      reinterpret_cast<mkvparser::Segment*>(webm_ctx->segment);
+  const mkvparser::Cluster* cluster =
+      reinterpret_cast<const mkvparser::Cluster*>(webm_ctx->cluster);
+  const mkvparser::Block *block =
+      reinterpret_cast<const mkvparser::Block*>(webm_ctx->block);
+  const mkvparser::BlockEntry *block_entry =
+      reinterpret_cast<const mkvparser::BlockEntry*>(webm_ctx->block_entry);
+  bool block_entry_eos = false;
+  do {
+    long status = 0;
+    bool get_new_block = false;
+    if (block_entry == NULL && !block_entry_eos) {
+      status = cluster->GetFirst(block_entry);
+      get_new_block = true;
+    } else if (block_entry_eos || block_entry->EOS()) {
+      cluster = segment->GetNext(cluster);
+      if (cluster == NULL || cluster->EOS()) {
+        *buffer_size = 0;
+        webm_ctx->reached_eos = 1;
+        return 1;
+      }
+      status = cluster->GetFirst(block_entry);
+      block_entry_eos = false;
+      get_new_block = true;
+    } else if (block == NULL ||
+               webm_ctx->block_frame_index == block->GetFrameCount() ||
+               block->GetTrackNumber() != webm_ctx->video_track_index) {
+      status = cluster->GetNext(block_entry, block_entry);
+      if (block_entry == NULL || block_entry->EOS()) {
+        block_entry_eos = true;
+        continue;
+      }
+      get_new_block = true;
+    }
+    if (status || block_entry == NULL) {
+      return -1;
+    }
+    if (get_new_block) {
+      block = block_entry->GetBlock();
+      webm_ctx->block_frame_index = 0;
+    }
+  } while (block->GetTrackNumber() != webm_ctx->video_track_index ||
+           block_entry_eos);
+
+  webm_ctx->cluster = cluster;
+  webm_ctx->block_entry = block_entry;
+  webm_ctx->block = block;
+
+  const mkvparser::Block::Frame& frame =
+      block->GetFrame(webm_ctx->block_frame_index);
+  ++webm_ctx->block_frame_index;
+  if (frame.len > static_cast<long>(*buffer_size)) {
+    delete[] *buffer;
+    *buffer = new uint8_t[frame.len];
+    if (*buffer == NULL) {
+      return -1;
+    }
+    webm_ctx->buffer = *buffer;
+  }
+  *buffer_size = frame.len;
+  webm_ctx->timestamp_ns = block->GetTime(cluster);
+  webm_ctx->is_key_frame = block->IsKey();
+
+  mkvparser::MkvReader *const reader =
+      reinterpret_cast<mkvparser::MkvReader*>(webm_ctx->reader);
+  return frame.Read(reader, *buffer) ? -1 : 0;
+}
+
+int webm_guess_framerate(struct WebmInputContext *webm_ctx,
+                         struct VpxInputContext *vpx_ctx) {
+  uint32_t i = 0;
+  uint8_t *buffer = NULL;
+  size_t buffer_size = 0;
+  while (webm_ctx->timestamp_ns < 1000000000 && i < 50) {
+    if (webm_read_frame(webm_ctx, &buffer, &buffer_size)) {
+      break;
+    }
+    ++i;
+  }
+  vpx_ctx->framerate.numerator = (i - 1) * 1000000;
+  vpx_ctx->framerate.denominator =
+      static_cast<int>(webm_ctx->timestamp_ns / 1000);
+  delete[] buffer;
+
+  get_first_cluster(webm_ctx);
+  webm_ctx->block = NULL;
+  webm_ctx->block_entry = NULL;
+  webm_ctx->block_frame_index = 0;
+  webm_ctx->timestamp_ns = 0;
+  webm_ctx->reached_eos = 0;
+
+  return 0;
+}
+
+void webm_free(struct WebmInputContext *webm_ctx) {
+  reset(webm_ctx);
+}
diff --git a/libvpx/libvpx/webmdec.h b/libvpx/libvpx/webmdec.h
new file mode 100644
index 0000000..aa371f3
--- /dev/null
+++ b/libvpx/libvpx/webmdec.h
@@ -0,0 +1,70 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef WEBMDEC_H_
+#define WEBMDEC_H_
+
+#include "./tools_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VpxInputContext;
+
+struct WebmInputContext {
+  void *reader;
+  void *segment;
+  uint8_t *buffer;
+  const void *cluster;
+  const void *block_entry;
+  const void *block;
+  int block_frame_index;
+  int video_track_index;
+  uint64_t timestamp_ns;
+  int is_key_frame;
+  int reached_eos;
+};
+
+// Checks if the input is a WebM file. If so, initializes WebMInputContext so
+// that webm_read_frame can be called to retrieve a video frame.
+// Returns 1 on success and 0 on failure or input is not WebM file.
+// TODO(vigneshv): Refactor this function into two smaller functions specific
+// to their task.
+int file_is_webm(struct WebmInputContext *webm_ctx,
+                 struct VpxInputContext *vpx_ctx);
+
+// Reads a WebM Video Frame. Memory for the buffer is created, owned and managed
+// by this function. For the first call, |buffer| should be NULL and
+// |*buffer_size| should be 0. Once all the frames are read and used,
+// webm_free() should be called, otherwise there will be a leak.
+// Parameters:
+//      webm_ctx - WebmInputContext object
+//      buffer - pointer where the frame data will be filled.
+//      buffer_size - pointer to buffer size.
+// Return values:
+//      0 - Success
+//      1 - End of Stream
+//     -1 - Error
+int webm_read_frame(struct WebmInputContext *webm_ctx,
+                    uint8_t **buffer,
+                    size_t *buffer_size);
+
+// Guesses the frame rate of the input file based on the container timestamps.
+int webm_guess_framerate(struct WebmInputContext *webm_ctx,
+                         struct VpxInputContext *vpx_ctx);
+
+// Resets the WebMInputContext.
+void webm_free(struct WebmInputContext *webm_ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // WEBMDEC_H_
diff --git a/libvpx/libvpx/webmenc.cc b/libvpx/libvpx/webmenc.cc
new file mode 100644
index 0000000..9929969
--- /dev/null
+++ b/libvpx/libvpx/webmenc.cc
@@ -0,0 +1,107 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./webmenc.h"
+
+#include <string>
+
+#include "third_party/libwebm/mkvmuxer/mkvmuxer.h"
+#include "third_party/libwebm/mkvmuxer/mkvmuxerutil.h"
+#include "third_party/libwebm/mkvmuxer/mkvwriter.h"
+
+namespace {
+const uint64_t kDebugTrackUid = 0xDEADBEEF;
+const int kVideoTrackNumber = 1;
+}  // namespace
+
+void write_webm_file_header(struct WebmOutputContext *webm_ctx,
+                            const vpx_codec_enc_cfg_t *cfg,
+                            const struct vpx_rational *fps,
+                            stereo_format_t stereo_fmt,
+                            unsigned int fourcc,
+                            const struct VpxRational *par) {
+  mkvmuxer::MkvWriter *const writer = new mkvmuxer::MkvWriter(webm_ctx->stream);
+  mkvmuxer::Segment *const segment = new mkvmuxer::Segment();
+  segment->Init(writer);
+  segment->set_mode(mkvmuxer::Segment::kFile);
+  segment->OutputCues(true);
+
+  mkvmuxer::SegmentInfo *const info = segment->GetSegmentInfo();
+  const uint64_t kTimecodeScale = 1000000;
+  info->set_timecode_scale(kTimecodeScale);
+  std::string version = "vpxenc";
+  if (!webm_ctx->debug) {
+    version.append(std::string(" ") + vpx_codec_version_str());
+  }
+  info->set_writing_app(version.c_str());
+
+  const uint64_t video_track_id =
+      segment->AddVideoTrack(static_cast<int>(cfg->g_w),
+                             static_cast<int>(cfg->g_h),
+                             kVideoTrackNumber);
+  mkvmuxer::VideoTrack* const video_track =
+      static_cast<mkvmuxer::VideoTrack*>(
+          segment->GetTrackByNumber(video_track_id));
+  video_track->SetStereoMode(stereo_fmt);
+  const char *codec_id;
+  switch (fourcc) {
+  case VP8_FOURCC:
+    codec_id = "V_VP8";
+    break;
+  case VP9_FOURCC:
+  default:
+    codec_id = "V_VP9";
+    break;
+  }
+  video_track->set_codec_id(codec_id);
+  if (par->numerator > 1 || par->denominator > 1) {
+    // TODO(fgalligan): Add support of DisplayUnit, Display Aspect Ratio type
+    // to WebM format.
+    const uint64_t display_width =
+        static_cast<uint64_t>(((cfg->g_w * par->numerator * 1.0) /
+                               par->denominator) + .5);
+    video_track->set_display_width(display_width);
+    video_track->set_display_height(cfg->g_h);
+  }
+  if (webm_ctx->debug) {
+    video_track->set_uid(kDebugTrackUid);
+  }
+  webm_ctx->writer = writer;
+  webm_ctx->segment = segment;
+}
+
+void write_webm_block(struct WebmOutputContext *webm_ctx,
+                      const vpx_codec_enc_cfg_t *cfg,
+                      const vpx_codec_cx_pkt_t *pkt) {
+  mkvmuxer::Segment *const segment =
+      reinterpret_cast<mkvmuxer::Segment*>(webm_ctx->segment);
+  int64_t pts_ns = pkt->data.frame.pts * 1000000000ll *
+                   cfg->g_timebase.num / cfg->g_timebase.den;
+  if (pts_ns <= webm_ctx->last_pts_ns)
+    pts_ns = webm_ctx->last_pts_ns + 1000000;
+  webm_ctx->last_pts_ns = pts_ns;
+
+  segment->AddFrame(static_cast<uint8_t*>(pkt->data.frame.buf),
+                    pkt->data.frame.sz,
+                    kVideoTrackNumber,
+                    pts_ns,
+                    pkt->data.frame.flags & VPX_FRAME_IS_KEY);
+}
+
+void write_webm_file_footer(struct WebmOutputContext *webm_ctx) {
+  mkvmuxer::MkvWriter *const writer =
+      reinterpret_cast<mkvmuxer::MkvWriter*>(webm_ctx->writer);
+  mkvmuxer::Segment *const segment =
+      reinterpret_cast<mkvmuxer::Segment*>(webm_ctx->segment);
+  segment->Finalize();
+  delete segment;
+  delete writer;
+  webm_ctx->writer = NULL;
+  webm_ctx->segment = NULL;
+}
diff --git a/libvpx/libvpx/webmenc.h b/libvpx/libvpx/webmenc.h
new file mode 100644
index 0000000..ad30664
--- /dev/null
+++ b/libvpx/libvpx/webmenc.h
@@ -0,0 +1,57 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef WEBMENC_H_
+#define WEBMENC_H_
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "tools_common.h"
+#include "vpx/vpx_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct WebmOutputContext {
+  int debug;
+  FILE *stream;
+  int64_t last_pts_ns;
+  void *writer;
+  void *segment;
+};
+
+/* Stereo 3D packed frame format */
+typedef enum stereo_format {
+  STEREO_FORMAT_MONO = 0,
+  STEREO_FORMAT_LEFT_RIGHT = 1,
+  STEREO_FORMAT_BOTTOM_TOP = 2,
+  STEREO_FORMAT_TOP_BOTTOM = 3,
+  STEREO_FORMAT_RIGHT_LEFT = 11
+} stereo_format_t;
+
+void write_webm_file_header(struct WebmOutputContext *webm_ctx,
+                            const vpx_codec_enc_cfg_t *cfg,
+                            const struct vpx_rational *fps,
+                            stereo_format_t stereo_fmt,
+                            unsigned int fourcc,
+                            const struct VpxRational *par);
+
+void write_webm_block(struct WebmOutputContext *webm_ctx,
+                      const vpx_codec_enc_cfg_t *cfg,
+                      const vpx_codec_cx_pkt_t *pkt);
+
+void write_webm_file_footer(struct WebmOutputContext *webm_ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // WEBMENC_H_
diff --git a/libvpx/libvpx/y4menc.c b/libvpx/libvpx/y4menc.c
new file mode 100644
index 0000000..b647e8d
--- /dev/null
+++ b/libvpx/libvpx/y4menc.c
@@ -0,0 +1,60 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include "./y4menc.h"
+
+int y4m_write_file_header(char *buf, size_t len, int width, int height,
+                          const struct VpxRational *framerate,
+                          vpx_img_fmt_t fmt, unsigned int bit_depth) {
+  const char *color;
+  switch (bit_depth) {
+    case 8:
+      color = fmt == VPX_IMG_FMT_444A ? "C444alpha\n" :
+              fmt == VPX_IMG_FMT_I444 ? "C444\n" :
+              fmt == VPX_IMG_FMT_I422 ? "C422\n" :
+              "C420jpeg\n";
+      break;
+    case 9:
+      color = fmt == VPX_IMG_FMT_I44416 ? "C444p9 XYSCSS=444P9\n" :
+              fmt == VPX_IMG_FMT_I42216 ? "C422p9 XYSCSS=422P9\n" :
+              "C420p9 XYSCSS=420P9\n";
+      break;
+    case 10:
+      color = fmt == VPX_IMG_FMT_I44416 ? "C444p10 XYSCSS=444P10\n" :
+              fmt == VPX_IMG_FMT_I42216 ? "C422p10 XYSCSS=422P10\n" :
+              "C420p10 XYSCSS=420P10\n";
+      break;
+    case 12:
+      color = fmt == VPX_IMG_FMT_I44416 ? "C444p12 XYSCSS=444P12\n" :
+              fmt == VPX_IMG_FMT_I42216 ? "C422p12 XYSCSS=422P12\n" :
+              "C420p12 XYSCSS=420P12\n";
+      break;
+    case 14:
+      color = fmt == VPX_IMG_FMT_I44416 ? "C444p14 XYSCSS=444P14\n" :
+              fmt == VPX_IMG_FMT_I42216 ? "C422p14 XYSCSS=422P14\n" :
+              "C420p14 XYSCSS=420P14\n";
+      break;
+    case 16:
+      color = fmt == VPX_IMG_FMT_I44416 ? "C444p16 XYSCSS=444P16\n" :
+              fmt == VPX_IMG_FMT_I42216 ? "C422p16 XYSCSS=422P16\n" :
+              "C420p16 XYSCSS=420P16\n";
+      break;
+    default:
+      color = NULL;
+      assert(0);
+  }
+  return snprintf(buf, len, "YUV4MPEG2 W%u H%u F%u:%u I%c %s", width, height,
+                  framerate->numerator, framerate->denominator, 'p', color);
+}
+
+int y4m_write_frame_header(char *buf, size_t len) {
+  return snprintf(buf, len, "FRAME\n");
+}
diff --git a/libvpx/libvpx/y4menc.h b/libvpx/libvpx/y4menc.h
new file mode 100644
index 0000000..69d5904
--- /dev/null
+++ b/libvpx/libvpx/y4menc.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef Y4MENC_H_
+#define Y4MENC_H_
+
+#include "./tools_common.h"
+
+#include "vpx/vpx_decoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define Y4M_BUFFER_SIZE 128
+
+int y4m_write_file_header(char *buf, size_t len, int width, int height,
+                          const struct VpxRational *framerate,
+                          vpx_img_fmt_t fmt, unsigned int bit_depth);
+int y4m_write_frame_header(char *buf, size_t len);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // Y4MENC_H_
diff --git a/libvpx/libvpx/y4minput.c b/libvpx/libvpx/y4minput.c
new file mode 100644
index 0000000..34ea96d
--- /dev/null
+++ b/libvpx/libvpx/y4minput.c
@@ -0,0 +1,1052 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ *
+ *  Based on code from the OggTheora software codec source code,
+ *  Copyright (C) 2002-2010 The Xiph.Org Foundation and contributors.
+ */
+#include <errno.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx/vpx_integer.h"
+#include "y4minput.h"
+
+// Reads 'size' bytes from 'file' into 'buf' with some fault tolerance.
+// Returns true on success.
+static int file_read(void *buf, size_t size, FILE *file) {
+  const int kMaxRetries = 5;
+  int retry_count = 0;
+  int file_error;
+  size_t len = 0;
+  do {
+    const size_t n = fread((uint8_t*)buf + len, 1, size - len, file);
+    len += n;
+    file_error = ferror(file);
+    if (file_error) {
+      if (errno == EINTR || errno == EAGAIN) {
+        clearerr(file);
+        continue;
+      } else {
+        fprintf(stderr, "Error reading file: %u of %u bytes read, %d: %s\n",
+                (uint32_t)len, (uint32_t)size, errno, strerror(errno));
+        return 0;
+      }
+    }
+  } while (!feof(file) && len < size && ++retry_count < kMaxRetries);
+
+  if (!feof(file) && len != size) {
+    fprintf(stderr, "Error reading file: %u of %u bytes read,"
+                    " error: %d, retries: %d, %d: %s\n",
+            (uint32_t)len, (uint32_t)size, file_error, retry_count,
+            errno, strerror(errno));
+  }
+  return len == size;
+}
+
+static int y4m_parse_tags(y4m_input *_y4m, char *_tags) {
+  int   got_w;
+  int   got_h;
+  int   got_fps;
+  int   got_interlace;
+  int   got_par;
+  int   got_chroma;
+  char *p;
+  char *q;
+  got_w = got_h = got_fps = got_interlace = got_par = got_chroma = 0;
+  for (p = _tags;; p = q) {
+    /*Skip any leading spaces.*/
+    while (*p == ' ')p++;
+    /*If that's all we have, stop.*/
+    if (p[0] == '\0')break;
+    /*Find the end of this tag.*/
+    for (q = p + 1; *q != '\0' && *q != ' '; q++);
+    /*Process the tag.*/
+    switch (p[0]) {
+      case 'W': {
+        if (sscanf(p + 1, "%d", &_y4m->pic_w) != 1)return -1;
+        got_w = 1;
+      }
+      break;
+      case 'H': {
+        if (sscanf(p + 1, "%d", &_y4m->pic_h) != 1)return -1;
+        got_h = 1;
+      }
+      break;
+      case 'F': {
+        if (sscanf(p + 1, "%d:%d", &_y4m->fps_n, &_y4m->fps_d) != 2) {
+          return -1;
+        }
+        got_fps = 1;
+      }
+      break;
+      case 'I': {
+        _y4m->interlace = p[1];
+        got_interlace = 1;
+      }
+      break;
+      case 'A': {
+        if (sscanf(p + 1, "%d:%d", &_y4m->par_n, &_y4m->par_d) != 2) {
+          return -1;
+        }
+        got_par = 1;
+      }
+      break;
+      case 'C': {
+        if (q - p > 16)return -1;
+        memcpy(_y4m->chroma_type, p + 1, q - p - 1);
+        _y4m->chroma_type[q - p - 1] = '\0';
+        got_chroma = 1;
+      }
+      break;
+      /*Ignore unknown tags.*/
+    }
+  }
+  if (!got_w || !got_h || !got_fps)return -1;
+  if (!got_interlace)_y4m->interlace = '?';
+  if (!got_par)_y4m->par_n = _y4m->par_d = 0;
+  /*Chroma-type is not specified in older files, e.g., those generated by
+     mplayer.*/
+  if (!got_chroma)strcpy(_y4m->chroma_type, "420");
+  return 0;
+}
+
+
+
+/*All anti-aliasing filters in the following conversion functions are based on
+   one of two window functions:
+  The 6-tap Lanczos window (for down-sampling and shifts):
+   sinc(\pi*t)*sinc(\pi*t/3), |t|<3  (sinc(t)==sin(t)/t)
+   0,                         |t|>=3
+  The 4-tap Mitchell window (for up-sampling):
+   7|t|^3-12|t|^2+16/3,             |t|<1
+   -(7/3)|x|^3+12|x|^2-20|x|+32/3,  |t|<2
+   0,                               |t|>=2
+  The number of taps is intentionally kept small to reduce computational
+   overhead and limit ringing.
+
+  The taps from these filters are scaled so that their sum is 1, and the result
+   is scaled by 128 and rounded to integers to create a filter whose
+   intermediate values fit inside 16 bits.
+  Coefficients are rounded in such a way as to ensure their sum is still 128,
+   which is usually equivalent to normal rounding.
+
+  Conversions which require both horizontal and vertical filtering could
+   have these steps pipelined, for less memory consumption and better cache
+   performance, but we do them separately for simplicity.*/
+
+#define OC_MINI(_a,_b)      ((_a)>(_b)?(_b):(_a))
+#define OC_MAXI(_a,_b)      ((_a)<(_b)?(_b):(_a))
+#define OC_CLAMPI(_a,_b,_c) (OC_MAXI(_a,OC_MINI(_b,_c)))
+
+/*420jpeg chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  420mpeg2 chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  BR      |       BR      |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  BR      |       BR      |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  We use a resampling filter to shift the site locations one quarter pixel (at
+   the chroma plane's resolution) to the right.
+  The 4:2:2 modes look exactly the same, except there are twice as many chroma
+   lines, and they are vertically co-sited with the luma samples in both the
+   mpeg2 and jpeg cases (thus requiring no vertical resampling).*/
+static void y4m_42xmpeg2_42xjpeg_helper(unsigned char *_dst,
+                                        const unsigned char *_src, int _c_w, int _c_h) {
+  int y;
+  int x;
+  for (y = 0; y < _c_h; y++) {
+    /*Filter: [4 -17 114 35 -9 1]/128, derived from a 6-tap Lanczos
+       window.*/
+    for (x = 0; x < OC_MINI(_c_w, 2); x++) {
+      _dst[x] = (unsigned char)OC_CLAMPI(0, (4 * _src[0] - 17 * _src[OC_MAXI(x - 1, 0)] +
+                                             114 * _src[x] + 35 * _src[OC_MINI(x + 1, _c_w - 1)] - 9 * _src[OC_MINI(x + 2, _c_w - 1)] +
+                                             _src[OC_MINI(x + 3, _c_w - 1)] + 64) >> 7, 255);
+    }
+    for (; x < _c_w - 3; x++) {
+      _dst[x] = (unsigned char)OC_CLAMPI(0, (4 * _src[x - 2] - 17 * _src[x - 1] +
+                                             114 * _src[x] + 35 * _src[x + 1] - 9 * _src[x + 2] + _src[x + 3] + 64) >> 7, 255);
+    }
+    for (; x < _c_w; x++) {
+      _dst[x] = (unsigned char)OC_CLAMPI(0, (4 * _src[x - 2] - 17 * _src[x - 1] +
+                                             114 * _src[x] + 35 * _src[OC_MINI(x + 1, _c_w - 1)] - 9 * _src[OC_MINI(x + 2, _c_w - 1)] +
+                                             _src[_c_w - 1] + 64) >> 7, 255);
+    }
+    _dst += _c_w;
+    _src += _c_w;
+  }
+}
+
+/*Handles both 422 and 420mpeg2 to 422jpeg and 420jpeg, respectively.*/
+static void y4m_convert_42xmpeg2_42xjpeg(y4m_input *_y4m, unsigned char *_dst,
+                                         unsigned char *_aux) {
+  int c_w;
+  int c_h;
+  int c_sz;
+  int pli;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  for (pli = 1; pli < 3; pli++) {
+    y4m_42xmpeg2_42xjpeg_helper(_dst, _aux, c_w, c_h);
+    _dst += c_sz;
+    _aux += c_sz;
+  }
+}
+
+/*This format is only used for interlaced content, but is included for
+   completeness.
+
+  420jpeg chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  420paldv chroma samples are sited like:
+  YR------Y-------YR------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YB------Y-------YB------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YR------Y-------YR------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YB------Y-------YB------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  We use a resampling filter to shift the site locations one quarter pixel (at
+   the chroma plane's resolution) to the right.
+  Then we use another filter to move the C_r location down one quarter pixel,
+   and the C_b location up one quarter pixel.*/
+static void y4m_convert_42xpaldv_42xjpeg(y4m_input *_y4m, unsigned char *_dst,
+                                         unsigned char *_aux) {
+  unsigned char *tmp;
+  int            c_w;
+  int            c_h;
+  int            c_sz;
+  int            pli;
+  int            y;
+  int            x;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + 1) / 2;
+  c_h = (_y4m->pic_h + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  c_sz = c_w * c_h;
+  tmp = _aux + 2 * c_sz;
+  for (pli = 1; pli < 3; pli++) {
+    /*First do the horizontal re-sampling.
+      This is the same as the mpeg2 case, except that after the horizontal
+       case, we need to apply a second vertical filter.*/
+    y4m_42xmpeg2_42xjpeg_helper(tmp, _aux, c_w, c_h);
+    _aux += c_sz;
+    switch (pli) {
+      case 1: {
+        /*Slide C_b up a quarter-pel.
+          This is the same filter used above, but in the other order.*/
+        for (x = 0; x < c_w; x++) {
+          for (y = 0; y < OC_MINI(c_h, 3); y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(0, (tmp[0]
+                                                         - 9 * tmp[OC_MAXI(y - 2, 0) * c_w] + 35 * tmp[OC_MAXI(y - 1, 0) * c_w]
+                                                         + 114 * tmp[y * c_w] - 17 * tmp[OC_MINI(y + 1, c_h - 1) * c_w]
+                                                         + 4 * tmp[OC_MINI(y + 2, c_h - 1) * c_w] + 64) >> 7, 255);
+          }
+          for (; y < c_h - 2; y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(0, (tmp[(y - 3) * c_w]
+                                                         - 9 * tmp[(y - 2) * c_w] + 35 * tmp[(y - 1) * c_w] + 114 * tmp[y * c_w]
+                                                         - 17 * tmp[(y + 1) * c_w] + 4 * tmp[(y + 2) * c_w] + 64) >> 7, 255);
+          }
+          for (; y < c_h; y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(0, (tmp[(y - 3) * c_w]
+                                                         - 9 * tmp[(y - 2) * c_w] + 35 * tmp[(y - 1) * c_w] + 114 * tmp[y * c_w]
+                                                         - 17 * tmp[OC_MINI(y + 1, c_h - 1) * c_w] + 4 * tmp[(c_h - 1) * c_w] + 64) >> 7, 255);
+          }
+          _dst++;
+          tmp++;
+        }
+        _dst += c_sz - c_w;
+        tmp -= c_w;
+      }
+      break;
+      case 2: {
+        /*Slide C_r down a quarter-pel.
+          This is the same as the horizontal filter.*/
+        for (x = 0; x < c_w; x++) {
+          for (y = 0; y < OC_MINI(c_h, 2); y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(0, (4 * tmp[0]
+                                                         - 17 * tmp[OC_MAXI(y - 1, 0) * c_w] + 114 * tmp[y * c_w]
+                                                         + 35 * tmp[OC_MINI(y + 1, c_h - 1) * c_w] - 9 * tmp[OC_MINI(y + 2, c_h - 1) * c_w]
+                                                         + tmp[OC_MINI(y + 3, c_h - 1) * c_w] + 64) >> 7, 255);
+          }
+          for (; y < c_h - 3; y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(0, (4 * tmp[(y - 2) * c_w]
+                                                         - 17 * tmp[(y - 1) * c_w] + 114 * tmp[y * c_w] + 35 * tmp[(y + 1) * c_w]
+                                                         - 9 * tmp[(y + 2) * c_w] + tmp[(y + 3) * c_w] + 64) >> 7, 255);
+          }
+          for (; y < c_h; y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(0, (4 * tmp[(y - 2) * c_w]
+                                                         - 17 * tmp[(y - 1) * c_w] + 114 * tmp[y * c_w] + 35 * tmp[OC_MINI(y + 1, c_h - 1) * c_w]
+                                                         - 9 * tmp[OC_MINI(y + 2, c_h - 1) * c_w] + tmp[(c_h - 1) * c_w] + 64) >> 7, 255);
+          }
+          _dst++;
+          tmp++;
+        }
+      }
+      break;
+    }
+    /*For actual interlaced material, this would have to be done separately on
+       each field, and the shift amounts would be different.
+      C_r moves down 1/8, C_b up 3/8 in the top field, and C_r moves down 3/8,
+       C_b up 1/8 in the bottom field.
+      The corresponding filters would be:
+       Down 1/8 (reverse order for up): [3 -11 125 15 -4 0]/128
+       Down 3/8 (reverse order for up): [4 -19 98 56 -13 2]/128*/
+  }
+}
+
+/*Perform vertical filtering to reduce a single plane from 4:2:2 to 4:2:0.
+  This is used as a helper by several converation routines.*/
+static void y4m_422jpeg_420jpeg_helper(unsigned char *_dst,
+                                       const unsigned char *_src, int _c_w, int _c_h) {
+  int y;
+  int x;
+  /*Filter: [3 -17 78 78 -17 3]/128, derived from a 6-tap Lanczos window.*/
+  for (x = 0; x < _c_w; x++) {
+    for (y = 0; y < OC_MINI(_c_h, 2); y += 2) {
+      _dst[(y >> 1)*_c_w] = OC_CLAMPI(0, (64 * _src[0]
+                                          + 78 * _src[OC_MINI(1, _c_h - 1) * _c_w]
+                                          - 17 * _src[OC_MINI(2, _c_h - 1) * _c_w]
+                                          + 3 * _src[OC_MINI(3, _c_h - 1) * _c_w] + 64) >> 7, 255);
+    }
+    for (; y < _c_h - 3; y += 2) {
+      _dst[(y >> 1)*_c_w] = OC_CLAMPI(0, (3 * (_src[(y - 2) * _c_w] + _src[(y + 3) * _c_w])
+                                          - 17 * (_src[(y - 1) * _c_w] + _src[(y + 2) * _c_w])
+                                          + 78 * (_src[y * _c_w] + _src[(y + 1) * _c_w]) + 64) >> 7, 255);
+    }
+    for (; y < _c_h; y += 2) {
+      _dst[(y >> 1)*_c_w] = OC_CLAMPI(0, (3 * (_src[(y - 2) * _c_w]
+                                               + _src[(_c_h - 1) * _c_w]) - 17 * (_src[(y - 1) * _c_w]
+                                                                                  + _src[OC_MINI(y + 2, _c_h - 1) * _c_w])
+                                          + 78 * (_src[y * _c_w] + _src[OC_MINI(y + 1, _c_h - 1) * _c_w]) + 64) >> 7, 255);
+    }
+    _src++;
+    _dst++;
+  }
+}
+
+/*420jpeg chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  422jpeg chroma samples are sited like:
+  Y---BR--Y-------Y---BR--Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y---BR--Y-------Y---BR--Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y---BR--Y-------Y---BR--Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y---BR--Y-------Y---BR--Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  We use a resampling filter to decimate the chroma planes by two in the
+   vertical direction.*/
+static void y4m_convert_422jpeg_420jpeg(y4m_input *_y4m, unsigned char *_dst,
+                                        unsigned char *_aux) {
+  int c_w;
+  int c_h;
+  int c_sz;
+  int dst_c_w;
+  int dst_c_h;
+  int dst_c_sz;
+  int pli;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + _y4m->src_c_dec_h - 1) / _y4m->src_c_dec_h;
+  c_h = _y4m->pic_h;
+  dst_c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  dst_c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  dst_c_sz = dst_c_w * dst_c_h;
+  for (pli = 1; pli < 3; pli++) {
+    y4m_422jpeg_420jpeg_helper(_dst, _aux, c_w, c_h);
+    _aux += c_sz;
+    _dst += dst_c_sz;
+  }
+}
+
+/*420jpeg chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  422 chroma samples are sited like:
+  YBR-----Y-------YBR-----Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------YBR-----Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------YBR-----Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------YBR-----Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  We use a resampling filter to shift the original site locations one quarter
+   pixel (at the original chroma resolution) to the right.
+  Then we use a second resampling filter to decimate the chroma planes by two
+   in the vertical direction.*/
+static void y4m_convert_422_420jpeg(y4m_input *_y4m, unsigned char *_dst,
+                                    unsigned char *_aux) {
+  unsigned char *tmp;
+  int            c_w;
+  int            c_h;
+  int            c_sz;
+  int            dst_c_h;
+  int            dst_c_sz;
+  int            pli;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + _y4m->src_c_dec_h - 1) / _y4m->src_c_dec_h;
+  c_h = _y4m->pic_h;
+  dst_c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  dst_c_sz = c_w * dst_c_h;
+  tmp = _aux + 2 * c_sz;
+  for (pli = 1; pli < 3; pli++) {
+    /*In reality, the horizontal and vertical steps could be pipelined, for
+       less memory consumption and better cache performance, but we do them
+       separately for simplicity.*/
+    /*First do horizontal filtering (convert to 422jpeg)*/
+    y4m_42xmpeg2_42xjpeg_helper(tmp, _aux, c_w, c_h);
+    /*Now do the vertical filtering.*/
+    y4m_422jpeg_420jpeg_helper(_dst, tmp, c_w, c_h);
+    _aux += c_sz;
+    _dst += dst_c_sz;
+  }
+}
+
+/*420jpeg chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  411 chroma samples are sited like:
+  YBR-----Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  We use a filter to resample at site locations one eighth pixel (at the source
+   chroma plane's horizontal resolution) and five eighths of a pixel to the
+   right.
+  Then we use another filter to decimate the planes by 2 in the vertical
+   direction.*/
+static void y4m_convert_411_420jpeg(y4m_input *_y4m, unsigned char *_dst,
+                                    unsigned char *_aux) {
+  unsigned char *tmp;
+  int            c_w;
+  int            c_h;
+  int            c_sz;
+  int            dst_c_w;
+  int            dst_c_h;
+  int            dst_c_sz;
+  int            tmp_sz;
+  int            pli;
+  int            y;
+  int            x;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + _y4m->src_c_dec_h - 1) / _y4m->src_c_dec_h;
+  c_h = _y4m->pic_h;
+  dst_c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  dst_c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  dst_c_sz = dst_c_w * dst_c_h;
+  tmp_sz = dst_c_w * c_h;
+  tmp = _aux + 2 * c_sz;
+  for (pli = 1; pli < 3; pli++) {
+    /*In reality, the horizontal and vertical steps could be pipelined, for
+       less memory consumption and better cache performance, but we do them
+       separately for simplicity.*/
+    /*First do horizontal filtering (convert to 422jpeg)*/
+    for (y = 0; y < c_h; y++) {
+      /*Filters: [1 110 18 -1]/128 and [-3 50 86 -5]/128, both derived from a
+         4-tap Mitchell window.*/
+      for (x = 0; x < OC_MINI(c_w, 1); x++) {
+        tmp[x << 1] = (unsigned char)OC_CLAMPI(0, (111 * _aux[0]
+                                                   + 18 * _aux[OC_MINI(1, c_w - 1)] - _aux[OC_MINI(2, c_w - 1)] + 64) >> 7, 255);
+        tmp[x << 1 | 1] = (unsigned char)OC_CLAMPI(0, (47 * _aux[0]
+                                                       + 86 * _aux[OC_MINI(1, c_w - 1)] - 5 * _aux[OC_MINI(2, c_w - 1)] + 64) >> 7, 255);
+      }
+      for (; x < c_w - 2; x++) {
+        tmp[x << 1] = (unsigned char)OC_CLAMPI(0, (_aux[x - 1] + 110 * _aux[x]
+                                                   + 18 * _aux[x + 1] - _aux[x + 2] + 64) >> 7, 255);
+        tmp[x << 1 | 1] = (unsigned char)OC_CLAMPI(0, (-3 * _aux[x - 1] + 50 * _aux[x]
+                                                       + 86 * _aux[x + 1] - 5 * _aux[x + 2] + 64) >> 7, 255);
+      }
+      for (; x < c_w; x++) {
+        tmp[x << 1] = (unsigned char)OC_CLAMPI(0, (_aux[x - 1] + 110 * _aux[x]
+                                                   + 18 * _aux[OC_MINI(x + 1, c_w - 1)] - _aux[c_w - 1] + 64) >> 7, 255);
+        if ((x << 1 | 1) < dst_c_w) {
+          tmp[x << 1 | 1] = (unsigned char)OC_CLAMPI(0, (-3 * _aux[x - 1] + 50 * _aux[x]
+                                                         + 86 * _aux[OC_MINI(x + 1, c_w - 1)] - 5 * _aux[c_w - 1] + 64) >> 7, 255);
+        }
+      }
+      tmp += dst_c_w;
+      _aux += c_w;
+    }
+    tmp -= tmp_sz;
+    /*Now do the vertical filtering.*/
+    y4m_422jpeg_420jpeg_helper(_dst, tmp, dst_c_w, c_h);
+    _dst += dst_c_sz;
+  }
+}
+
+/*Convert 444 to 420jpeg.*/
+static void y4m_convert_444_420jpeg(y4m_input *_y4m, unsigned char *_dst,
+                                    unsigned char *_aux) {
+  unsigned char *tmp;
+  int            c_w;
+  int            c_h;
+  int            c_sz;
+  int            dst_c_w;
+  int            dst_c_h;
+  int            dst_c_sz;
+  int            tmp_sz;
+  int            pli;
+  int            y;
+  int            x;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + _y4m->src_c_dec_h - 1) / _y4m->src_c_dec_h;
+  c_h = _y4m->pic_h;
+  dst_c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  dst_c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  dst_c_sz = dst_c_w * dst_c_h;
+  tmp_sz = dst_c_w * c_h;
+  tmp = _aux + 2 * c_sz;
+  for (pli = 1; pli < 3; pli++) {
+    /*Filter: [3 -17 78 78 -17 3]/128, derived from a 6-tap Lanczos window.*/
+    for (y = 0; y < c_h; y++) {
+      for (x = 0; x < OC_MINI(c_w, 2); x += 2) {
+        tmp[x >> 1] = OC_CLAMPI(0, (64 * _aux[0] + 78 * _aux[OC_MINI(1, c_w - 1)]
+                                    - 17 * _aux[OC_MINI(2, c_w - 1)]
+                                    + 3 * _aux[OC_MINI(3, c_w - 1)] + 64) >> 7, 255);
+      }
+      for (; x < c_w - 3; x += 2) {
+        tmp[x >> 1] = OC_CLAMPI(0, (3 * (_aux[x - 2] + _aux[x + 3])
+                                    - 17 * (_aux[x - 1] + _aux[x + 2]) + 78 * (_aux[x] + _aux[x + 1]) + 64) >> 7, 255);
+      }
+      for (; x < c_w; x += 2) {
+        tmp[x >> 1] = OC_CLAMPI(0, (3 * (_aux[x - 2] + _aux[c_w - 1]) -
+                                    17 * (_aux[x - 1] + _aux[OC_MINI(x + 2, c_w - 1)]) +
+                                    78 * (_aux[x] + _aux[OC_MINI(x + 1, c_w - 1)]) + 64) >> 7, 255);
+      }
+      tmp += dst_c_w;
+      _aux += c_w;
+    }
+    tmp -= tmp_sz;
+    /*Now do the vertical filtering.*/
+    y4m_422jpeg_420jpeg_helper(_dst, tmp, dst_c_w, c_h);
+    _dst += dst_c_sz;
+  }
+}
+
+/*The image is padded with empty chroma components at 4:2:0.*/
+static void y4m_convert_mono_420jpeg(y4m_input *_y4m, unsigned char *_dst,
+                                     unsigned char *_aux) {
+  int c_sz;
+  (void)_aux;
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  c_sz = ((_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h) *
+         ((_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v);
+  memset(_dst, 128, c_sz * 2);
+}
+
+/*No conversion function needed.*/
+static void y4m_convert_null(y4m_input *_y4m, unsigned char *_dst,
+                             unsigned char *_aux) {
+  (void)_y4m;
+  (void)_dst;
+  (void)_aux;
+}
+
+int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
+                   int only_420) {
+  char buffer[80] = {0};
+  int  ret;
+  int  i;
+  /*Read until newline, or 80 cols, whichever happens first.*/
+  for (i = 0; i < 79; i++) {
+    if (_nskip > 0) {
+      buffer[i] = *_skip++;
+      _nskip--;
+    } else {
+      if (!file_read(buffer + i, 1, _fin)) return -1;
+    }
+    if (buffer[i] == '\n')break;
+  }
+  /*We skipped too much header data.*/
+  if (_nskip > 0)return -1;
+  if (i == 79) {
+    fprintf(stderr, "Error parsing header; not a YUV2MPEG2 file?\n");
+    return -1;
+  }
+  buffer[i] = '\0';
+  if (memcmp(buffer, "YUV4MPEG", 8)) {
+    fprintf(stderr, "Incomplete magic for YUV4MPEG file.\n");
+    return -1;
+  }
+  if (buffer[8] != '2') {
+    fprintf(stderr, "Incorrect YUV input file version; YUV4MPEG2 required.\n");
+  }
+  ret = y4m_parse_tags(_y4m, buffer + 5);
+  if (ret < 0) {
+    fprintf(stderr, "Error parsing YUV4MPEG2 header.\n");
+    return ret;
+  }
+  if (_y4m->interlace == '?') {
+    fprintf(stderr, "Warning: Input video interlacing format unknown; "
+            "assuming progressive scan.\n");
+  } else if (_y4m->interlace != 'p') {
+    fprintf(stderr, "Input video is interlaced; "
+            "Only progressive scan handled.\n");
+    return -1;
+  }
+  _y4m->vpx_fmt = VPX_IMG_FMT_I420;
+  _y4m->bps = 12;
+  _y4m->bit_depth = 8;
+  if (strcmp(_y4m->chroma_type, "420") == 0 ||
+      strcmp(_y4m->chroma_type, "420jpeg") == 0) {
+    _y4m->src_c_dec_h = _y4m->dst_c_dec_h = _y4m->src_c_dec_v = _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h
+                            + 2 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2);
+    /* Natively supported: no conversion required. */
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+  } else if (strcmp(_y4m->chroma_type, "420p10") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->dst_c_dec_h = 2;
+    _y4m->src_c_dec_v = 2;
+    _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = 2 * (_y4m->pic_w * _y4m->pic_h +
+                                 2 * ((_y4m->pic_w + 1) / 2) *
+                                 ((_y4m->pic_h + 1) / 2));
+    /* Natively supported: no conversion required. */
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    _y4m->bit_depth = 10;
+    _y4m->bps = 15;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I42016;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 420p10 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "420p12") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->dst_c_dec_h = 2;
+    _y4m->src_c_dec_v = 2;
+    _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = 2 * (_y4m->pic_w * _y4m->pic_h +
+                                 2 * ((_y4m->pic_w + 1) / 2) *
+                                 ((_y4m->pic_h + 1) / 2));
+    /* Natively supported: no conversion required. */
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    _y4m->bit_depth = 12;
+    _y4m->bps = 18;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I42016;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 420p12 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "420mpeg2") == 0) {
+    _y4m->src_c_dec_h = _y4m->dst_c_dec_h = _y4m->src_c_dec_v = _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+    /*Chroma filter required: read into the aux buf first.*/
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz =
+                         2 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2);
+    _y4m->convert = y4m_convert_42xmpeg2_42xjpeg;
+  } else if (strcmp(_y4m->chroma_type, "420paldv") == 0) {
+    _y4m->src_c_dec_h = _y4m->dst_c_dec_h = _y4m->src_c_dec_v = _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+    /*Chroma filter required: read into the aux buf first.
+      We need to make two filter passes, so we need some extra space in the
+       aux buffer.*/
+    _y4m->aux_buf_sz = 3 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2);
+    _y4m->aux_buf_read_sz = 2 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2);
+    _y4m->convert = y4m_convert_42xpaldv_42xjpeg;
+  } else if (strcmp(_y4m->chroma_type, "422jpeg") == 0) {
+    _y4m->src_c_dec_h = _y4m->dst_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+    /*Chroma filter required: read into the aux buf first.*/
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+    _y4m->convert = y4m_convert_422jpeg_420jpeg;
+  } else if (strcmp(_y4m->chroma_type, "422") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    if (only_420) {
+      _y4m->dst_c_dec_h = 2;
+      _y4m->dst_c_dec_v = 2;
+      _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+      /*Chroma filter required: read into the aux buf first.
+        We need to make two filter passes, so we need some extra space in the
+         aux buffer.*/
+      _y4m->aux_buf_read_sz = 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+      _y4m->aux_buf_sz = _y4m->aux_buf_read_sz +
+          ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+      _y4m->convert = y4m_convert_422_420jpeg;
+    } else {
+      _y4m->vpx_fmt = VPX_IMG_FMT_I422;
+      _y4m->bps = 16;
+      _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+      _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+      _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h
+                              + 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+      /*Natively supported: no conversion required.*/
+      _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+      _y4m->convert = y4m_convert_null;
+    }
+  } else if (strcmp(_y4m->chroma_type, "422p10") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I42216;
+    _y4m->bps = 20;
+    _y4m->bit_depth = 10;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * (_y4m->pic_w * _y4m->pic_h +
+                                 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h);
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 422p10 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "422p12") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I42216;
+    _y4m->bps = 24;
+    _y4m->bit_depth = 12;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * (_y4m->pic_w * _y4m->pic_h +
+                                 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h);
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 422p12 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "411") == 0) {
+    _y4m->src_c_dec_h = 4;
+    _y4m->dst_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+    /*Chroma filter required: read into the aux buf first.
+      We need to make two filter passes, so we need some extra space in the
+       aux buffer.*/
+    _y4m->aux_buf_read_sz = 2 * ((_y4m->pic_w + 3) / 4) * _y4m->pic_h;
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz + ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+    _y4m->convert = y4m_convert_411_420jpeg;
+  } else if (strcmp(_y4m->chroma_type, "444") == 0) {
+    _y4m->src_c_dec_h = 1;
+    _y4m->src_c_dec_v = 1;
+    if (only_420) {
+      _y4m->dst_c_dec_h = 2;
+      _y4m->dst_c_dec_v = 2;
+      _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+      /*Chroma filter required: read into the aux buf first.
+        We need to make two filter passes, so we need some extra space in the
+         aux buffer.*/
+      _y4m->aux_buf_read_sz = 2 * _y4m->pic_w * _y4m->pic_h;
+      _y4m->aux_buf_sz = _y4m->aux_buf_read_sz +
+          ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+      _y4m->convert = y4m_convert_444_420jpeg;
+    } else {
+      _y4m->vpx_fmt = VPX_IMG_FMT_I444;
+      _y4m->bps = 24;
+      _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+      _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+      _y4m->dst_buf_read_sz = 3 * _y4m->pic_w * _y4m->pic_h;
+      /*Natively supported: no conversion required.*/
+      _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+      _y4m->convert = y4m_convert_null;
+    }
+  } else if (strcmp(_y4m->chroma_type, "444p10") == 0) {
+    _y4m->src_c_dec_h = 1;
+    _y4m->src_c_dec_v = 1;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I44416;
+    _y4m->bps = 30;
+    _y4m->bit_depth = 10;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * 3 * _y4m->pic_w * _y4m->pic_h;
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 444p10 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "444p12") == 0) {
+    _y4m->src_c_dec_h = 1;
+    _y4m->src_c_dec_v = 1;
+    _y4m->vpx_fmt = VPX_IMG_FMT_I44416;
+    _y4m->bps = 36;
+    _y4m->bit_depth = 12;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * 3 * _y4m->pic_w * _y4m->pic_h;
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 444p12 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "444alpha") == 0) {
+    _y4m->src_c_dec_h = 1;
+    _y4m->src_c_dec_v = 1;
+    if (only_420) {
+      _y4m->dst_c_dec_h = 2;
+      _y4m->dst_c_dec_v = 2;
+      _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+      /*Chroma filter required: read into the aux buf first.
+        We need to make two filter passes, so we need some extra space in the
+         aux buffer.
+        The extra plane also gets read into the aux buf.
+        It will be discarded.*/
+      _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 3 * _y4m->pic_w * _y4m->pic_h;
+      _y4m->convert = y4m_convert_444_420jpeg;
+    } else {
+      _y4m->vpx_fmt = VPX_IMG_FMT_444A;
+      _y4m->bps = 32;
+      _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+      _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+      _y4m->dst_buf_read_sz = 4 * _y4m->pic_w * _y4m->pic_h;
+      /*Natively supported: no conversion required.*/
+      _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+      _y4m->convert = y4m_convert_null;
+    }
+  } else if (strcmp(_y4m->chroma_type, "mono") == 0) {
+    _y4m->src_c_dec_h = _y4m->src_c_dec_v = 0;
+    _y4m->dst_c_dec_h = _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+    /*No extra space required, but we need to clear the chroma planes.*/
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_mono_420jpeg;
+  } else {
+    fprintf(stderr, "Unknown chroma sampling type: %s\n", _y4m->chroma_type);
+    return -1;
+  }
+  /*The size of the final frame buffers is always computed from the
+     destination chroma decimation type.*/
+  _y4m->dst_buf_sz = _y4m->pic_w * _y4m->pic_h
+                     + 2 * ((_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h) *
+                     ((_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v);
+  if (_y4m->bit_depth == 8)
+    _y4m->dst_buf = (unsigned char *)malloc(_y4m->dst_buf_sz);
+  else
+    _y4m->dst_buf = (unsigned char *)malloc(2 * _y4m->dst_buf_sz);
+
+  if (_y4m->aux_buf_sz > 0)
+    _y4m->aux_buf = (unsigned char *)malloc(_y4m->aux_buf_sz);
+  return 0;
+}
+
+void y4m_input_close(y4m_input *_y4m) {
+  free(_y4m->dst_buf);
+  free(_y4m->aux_buf);
+}
+
+int y4m_input_fetch_frame(y4m_input *_y4m, FILE *_fin, vpx_image_t *_img) {
+  char frame[6];
+  int  pic_sz;
+  int  c_w;
+  int  c_h;
+  int  c_sz;
+  int  bytes_per_sample = _y4m->bit_depth > 8 ? 2 : 1;
+  /*Read and skip the frame header.*/
+  if (!file_read(frame, 6, _fin)) return 0;
+  if (memcmp(frame, "FRAME", 5)) {
+    fprintf(stderr, "Loss of framing in Y4M input data\n");
+    return -1;
+  }
+  if (frame[5] != '\n') {
+    char c;
+    int  j;
+    for (j = 0; j < 79 && file_read(&c, 1, _fin) && c != '\n'; j++) {}
+    if (j == 79) {
+      fprintf(stderr, "Error parsing Y4M frame header\n");
+      return -1;
+    }
+  }
+  /*Read the frame data that needs no conversion.*/
+  if (!file_read(_y4m->dst_buf, _y4m->dst_buf_read_sz, _fin)) {
+    fprintf(stderr, "Error reading Y4M frame data.\n");
+    return -1;
+  }
+  /*Read the frame data that does need conversion.*/
+  if (!file_read(_y4m->aux_buf, _y4m->aux_buf_read_sz, _fin)) {
+    fprintf(stderr, "Error reading Y4M frame data.\n");
+    return -1;
+  }
+  /*Now convert the just read frame.*/
+  (*_y4m->convert)(_y4m, _y4m->dst_buf, _y4m->aux_buf);
+  /*Fill in the frame buffer pointers.
+    We don't use vpx_img_wrap() because it forces padding for odd picture
+     sizes, which would require a separate fread call for every row.*/
+  memset(_img, 0, sizeof(*_img));
+  /*Y4M has the planes in Y'CbCr order, which libvpx calls Y, U, and V.*/
+  _img->fmt = _y4m->vpx_fmt;
+  _img->w = _img->d_w = _y4m->pic_w;
+  _img->h = _img->d_h = _y4m->pic_h;
+  _img->x_chroma_shift = _y4m->dst_c_dec_h >> 1;
+  _img->y_chroma_shift = _y4m->dst_c_dec_v >> 1;
+  _img->bps = _y4m->bps;
+
+  /*Set up the buffer pointers.*/
+  pic_sz = _y4m->pic_w * _y4m->pic_h * bytes_per_sample;
+  c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  c_w *= bytes_per_sample;
+  c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  _img->stride[VPX_PLANE_Y] = _img->stride[VPX_PLANE_ALPHA] =
+      _y4m->pic_w * bytes_per_sample;
+  _img->stride[VPX_PLANE_U] = _img->stride[VPX_PLANE_V] = c_w;
+  _img->planes[VPX_PLANE_Y] = _y4m->dst_buf;
+  _img->planes[VPX_PLANE_U] = _y4m->dst_buf + pic_sz;
+  _img->planes[VPX_PLANE_V] = _y4m->dst_buf + pic_sz + c_sz;
+  _img->planes[VPX_PLANE_ALPHA] = _y4m->dst_buf + pic_sz + 2 * c_sz;
+  return 1;
+}
diff --git a/libvpx/libvpx/y4minput.h b/libvpx/libvpx/y4minput.h
new file mode 100644
index 0000000..356cebb
--- /dev/null
+++ b/libvpx/libvpx/y4minput.h
@@ -0,0 +1,74 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ *
+ *  Based on code from the OggTheora software codec source code,
+ *  Copyright (C) 2002-2010 The Xiph.Org Foundation and contributors.
+ */
+
+#ifndef Y4MINPUT_H_
+#define Y4MINPUT_H_
+
+# include <stdio.h>
+# include "vpx/vpx_image.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
+typedef struct y4m_input y4m_input;
+
+
+
+/*The function used to perform chroma conversion.*/
+typedef void (*y4m_convert_func)(y4m_input *_y4m,
+                                 unsigned char *_dst, unsigned char *_src);
+
+
+
+struct y4m_input {
+  int               pic_w;
+  int               pic_h;
+  int               fps_n;
+  int               fps_d;
+  int               par_n;
+  int               par_d;
+  char              interlace;
+  int               src_c_dec_h;
+  int               src_c_dec_v;
+  int               dst_c_dec_h;
+  int               dst_c_dec_v;
+  char              chroma_type[16];
+  /*The size of each converted frame buffer.*/
+  size_t            dst_buf_sz;
+  /*The amount to read directly into the converted frame buffer.*/
+  size_t            dst_buf_read_sz;
+  /*The size of the auxilliary buffer.*/
+  size_t            aux_buf_sz;
+  /*The amount to read into the auxilliary buffer.*/
+  size_t            aux_buf_read_sz;
+  y4m_convert_func  convert;
+  unsigned char    *dst_buf;
+  unsigned char    *aux_buf;
+  enum vpx_img_fmt  vpx_fmt;
+  int               bps;
+  unsigned int      bit_depth;
+};
+
+int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
+                   int only_420);
+void y4m_input_close(y4m_input *_y4m);
+int y4m_input_fetch_frame(y4m_input *_y4m, FILE *_fin, vpx_image_t *img);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // Y4MINPUT_H_
diff --git a/libvpx/libwebm.mk b/libvpx/libwebm.mk
new file mode 100644
index 0000000..3afa4a0
--- /dev/null
+++ b/libvpx/libwebm.mk
@@ -0,0 +1,11 @@
+# libwebm
+LOCAL_PATH := $(call my-dir)
+include $(CLEAR_VARS)
+
+LOCAL_SRC_FILES := libwebm/mkvparser/mkvparser.cc
+LOCAL_CPP_EXTENSION := .cc
+LOCAL_C_INCLUDES += $(LOCAL_PATH)/libwebm/
+
+LOCAL_MODULE := libwebm
+
+include $(BUILD_STATIC_LIBRARY)
diff --git a/libvpx/libwebm/README.android b/libvpx/libwebm/README.android
new file mode 100644
index 0000000..dba2327
--- /dev/null
+++ b/libvpx/libwebm/README.android
@@ -0,0 +1,35 @@
+Name: libwebm
+URL: http://www.webmproject.org
+
+Commit: 5c50e310e7050192b952fe588186fd1dadc08b6e
+
+Description:
+Contains the sources used to compile libwebm's matroska parser.
+
+The libwebm source is from webmproject.org:
+  https://chromium.googlesource.com/webm/libwebm
+
+Notes on updating libwebm source code:
+
+Please follow these steps to update libvpx source code:
+
+1. Copy over the following files from the the libwebm checkout:
+     - mkvparser/mkvparser.cc
+     - mkvparser/mkvparser.h
+     - common/webmids.h
+
+2. Update README.android (this file) with the upstream hash.
+
+3. Copy the git log summary of changes by using the following in the libwebm
+   checkout: git log --pretty="%h %s" <previous_hash>...<current_hash>
+
+4. Commit the changes. The commit message should look like this:
+
+   libwebm: Pull from upstream
+
+   Current HEAD: <hash>
+
+   git log from upstream:
+   a6b2070 <git commit message 1>
+   08dabbc <git commit message 2>
+   c29fb02 <git commit message 3>
diff --git a/libvpx/libwebm/common/webmids.h b/libvpx/libwebm/common/webmids.h
new file mode 100644
index 0000000..32a0c5f
--- /dev/null
+++ b/libvpx/libwebm/common/webmids.h
@@ -0,0 +1,184 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#ifndef COMMON_WEBMIDS_H_
+#define COMMON_WEBMIDS_H_
+
+namespace libwebm {
+
+enum MkvId {
+  kMkvEBML = 0x1A45DFA3,
+  kMkvEBMLVersion = 0x4286,
+  kMkvEBMLReadVersion = 0x42F7,
+  kMkvEBMLMaxIDLength = 0x42F2,
+  kMkvEBMLMaxSizeLength = 0x42F3,
+  kMkvDocType = 0x4282,
+  kMkvDocTypeVersion = 0x4287,
+  kMkvDocTypeReadVersion = 0x4285,
+  kMkvVoid = 0xEC,
+  kMkvSignatureSlot = 0x1B538667,
+  kMkvSignatureAlgo = 0x7E8A,
+  kMkvSignatureHash = 0x7E9A,
+  kMkvSignaturePublicKey = 0x7EA5,
+  kMkvSignature = 0x7EB5,
+  kMkvSignatureElements = 0x7E5B,
+  kMkvSignatureElementList = 0x7E7B,
+  kMkvSignedElement = 0x6532,
+  // segment
+  kMkvSegment = 0x18538067,
+  // Meta Seek Information
+  kMkvSeekHead = 0x114D9B74,
+  kMkvSeek = 0x4DBB,
+  kMkvSeekID = 0x53AB,
+  kMkvSeekPosition = 0x53AC,
+  // Segment Information
+  kMkvInfo = 0x1549A966,
+  kMkvTimecodeScale = 0x2AD7B1,
+  kMkvDuration = 0x4489,
+  kMkvDateUTC = 0x4461,
+  kMkvTitle = 0x7BA9,
+  kMkvMuxingApp = 0x4D80,
+  kMkvWritingApp = 0x5741,
+  // Cluster
+  kMkvCluster = 0x1F43B675,
+  kMkvTimecode = 0xE7,
+  kMkvPrevSize = 0xAB,
+  kMkvBlockGroup = 0xA0,
+  kMkvBlock = 0xA1,
+  kMkvBlockDuration = 0x9B,
+  kMkvReferenceBlock = 0xFB,
+  kMkvLaceNumber = 0xCC,
+  kMkvSimpleBlock = 0xA3,
+  kMkvBlockAdditions = 0x75A1,
+  kMkvBlockMore = 0xA6,
+  kMkvBlockAddID = 0xEE,
+  kMkvBlockAdditional = 0xA5,
+  kMkvDiscardPadding = 0x75A2,
+  // Track
+  kMkvTracks = 0x1654AE6B,
+  kMkvTrackEntry = 0xAE,
+  kMkvTrackNumber = 0xD7,
+  kMkvTrackUID = 0x73C5,
+  kMkvTrackType = 0x83,
+  kMkvFlagEnabled = 0xB9,
+  kMkvFlagDefault = 0x88,
+  kMkvFlagForced = 0x55AA,
+  kMkvFlagLacing = 0x9C,
+  kMkvDefaultDuration = 0x23E383,
+  kMkvMaxBlockAdditionID = 0x55EE,
+  kMkvName = 0x536E,
+  kMkvLanguage = 0x22B59C,
+  kMkvCodecID = 0x86,
+  kMkvCodecPrivate = 0x63A2,
+  kMkvCodecName = 0x258688,
+  kMkvCodecDelay = 0x56AA,
+  kMkvSeekPreRoll = 0x56BB,
+  // video
+  kMkvVideo = 0xE0,
+  kMkvFlagInterlaced = 0x9A,
+  kMkvStereoMode = 0x53B8,
+  kMkvAlphaMode = 0x53C0,
+  kMkvPixelWidth = 0xB0,
+  kMkvPixelHeight = 0xBA,
+  kMkvPixelCropBottom = 0x54AA,
+  kMkvPixelCropTop = 0x54BB,
+  kMkvPixelCropLeft = 0x54CC,
+  kMkvPixelCropRight = 0x54DD,
+  kMkvDisplayWidth = 0x54B0,
+  kMkvDisplayHeight = 0x54BA,
+  kMkvDisplayUnit = 0x54B2,
+  kMkvAspectRatioType = 0x54B3,
+  kMkvFrameRate = 0x2383E3,
+  // end video
+  // colour
+  kMkvColour = 0x55B0,
+  kMkvMatrixCoefficients = 0x55B1,
+  kMkvBitsPerChannel = 0x55B2,
+  kMkvChromaSubsamplingHorz = 0x55B3,
+  kMkvChromaSubsamplingVert = 0x55B4,
+  kMkvCbSubsamplingHorz = 0x55B5,
+  kMkvCbSubsamplingVert = 0x55B6,
+  kMkvChromaSitingHorz = 0x55B7,
+  kMkvChromaSitingVert = 0x55B8,
+  kMkvRange = 0x55B9,
+  kMkvTransferCharacteristics = 0x55BA,
+  kMkvPrimaries = 0x55BB,
+  kMkvMaxCLL = 0x55BC,
+  kMkvMaxFALL = 0x55BD,
+  // mastering metadata
+  kMkvMasteringMetadata = 0x55D0,
+  kMkvPrimaryRChromaticityX = 0x55D1,
+  kMkvPrimaryRChromaticityY = 0x55D2,
+  kMkvPrimaryGChromaticityX = 0x55D3,
+  kMkvPrimaryGChromaticityY = 0x55D4,
+  kMkvPrimaryBChromaticityX = 0x55D5,
+  kMkvPrimaryBChromaticityY = 0x55D6,
+  kMkvWhitePointChromaticityX = 0x55D7,
+  kMkvWhitePointChromaticityY = 0x55D8,
+  kMkvLuminanceMax = 0x55D9,
+  kMkvLuminanceMin = 0x55DA,
+  // end mastering metadata
+  // end colour
+  // audio
+  kMkvAudio = 0xE1,
+  kMkvSamplingFrequency = 0xB5,
+  kMkvOutputSamplingFrequency = 0x78B5,
+  kMkvChannels = 0x9F,
+  kMkvBitDepth = 0x6264,
+  // end audio
+  // ContentEncodings
+  kMkvContentEncodings = 0x6D80,
+  kMkvContentEncoding = 0x6240,
+  kMkvContentEncodingOrder = 0x5031,
+  kMkvContentEncodingScope = 0x5032,
+  kMkvContentEncodingType = 0x5033,
+  kMkvContentCompression = 0x5034,
+  kMkvContentCompAlgo = 0x4254,
+  kMkvContentCompSettings = 0x4255,
+  kMkvContentEncryption = 0x5035,
+  kMkvContentEncAlgo = 0x47E1,
+  kMkvContentEncKeyID = 0x47E2,
+  kMkvContentSignature = 0x47E3,
+  kMkvContentSigKeyID = 0x47E4,
+  kMkvContentSigAlgo = 0x47E5,
+  kMkvContentSigHashAlgo = 0x47E6,
+  kMkvContentEncAESSettings = 0x47E7,
+  kMkvAESSettingsCipherMode = 0x47E8,
+  kMkvAESSettingsCipherInitData = 0x47E9,
+  // end ContentEncodings
+  // Cueing Data
+  kMkvCues = 0x1C53BB6B,
+  kMkvCuePoint = 0xBB,
+  kMkvCueTime = 0xB3,
+  kMkvCueTrackPositions = 0xB7,
+  kMkvCueTrack = 0xF7,
+  kMkvCueClusterPosition = 0xF1,
+  kMkvCueBlockNumber = 0x5378,
+  // Chapters
+  kMkvChapters = 0x1043A770,
+  kMkvEditionEntry = 0x45B9,
+  kMkvChapterAtom = 0xB6,
+  kMkvChapterUID = 0x73C4,
+  kMkvChapterStringUID = 0x5654,
+  kMkvChapterTimeStart = 0x91,
+  kMkvChapterTimeEnd = 0x92,
+  kMkvChapterDisplay = 0x80,
+  kMkvChapString = 0x85,
+  kMkvChapLanguage = 0x437C,
+  kMkvChapCountry = 0x437E,
+  // Tags
+  kMkvTags = 0x1254C367,
+  kMkvTag = 0x7373,
+  kMkvSimpleTag = 0x67C8,
+  kMkvTagName = 0x45A3,
+  kMkvTagString = 0x4487
+};
+
+}  // namespace libwebm
+
+#endif  // COMMON_WEBMIDS_H_
diff --git a/libvpx/libwebm/mkvparser/mkvparser.cc b/libvpx/libwebm/mkvparser/mkvparser.cc
new file mode 100644
index 0000000..ff13327
--- /dev/null
+++ b/libvpx/libwebm/mkvparser/mkvparser.cc
@@ -0,0 +1,7940 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#include "mkvparser/mkvparser.h"
+
+#if defined(_MSC_VER) && _MSC_VER < 1800
+#include <float.h>  // _isnan() / _finite()
+#define MSC_COMPAT
+#endif
+
+#include <cassert>
+#include <cfloat>
+#include <climits>
+#include <cmath>
+#include <cstring>
+#include <memory>
+#include <new>
+
+#include "common/webmids.h"
+
+namespace mkvparser {
+const float MasteringMetadata::kValueNotPresent = FLT_MAX;
+const long long Colour::kValueNotPresent = LLONG_MAX;
+
+#ifdef MSC_COMPAT
+inline bool isnan(double val) { return !!_isnan(val); }
+inline bool isinf(double val) { return !_finite(val); }
+#else
+inline bool isnan(double val) { return std::isnan(val); }
+inline bool isinf(double val) { return std::isinf(val); }
+#endif  // MSC_COMPAT
+
+IMkvReader::~IMkvReader() {}
+
+template <typename Type>
+Type* SafeArrayAlloc(unsigned long long num_elements,
+                     unsigned long long element_size) {
+  if (num_elements == 0 || element_size == 0)
+    return NULL;
+
+  const size_t kMaxAllocSize = 0x80000000;  // 2GiB
+  const unsigned long long num_bytes = num_elements * element_size;
+  if (element_size > (kMaxAllocSize / num_elements))
+    return NULL;
+  if (num_bytes != static_cast<size_t>(num_bytes))
+    return NULL;
+
+  return new (std::nothrow) Type[static_cast<size_t>(num_bytes)];
+}
+
+void GetVersion(int& major, int& minor, int& build, int& revision) {
+  major = 1;
+  minor = 0;
+  build = 0;
+  revision = 30;
+}
+
+long long ReadUInt(IMkvReader* pReader, long long pos, long& len) {
+  if (!pReader || pos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  len = 1;
+  unsigned char b;
+  int status = pReader->Read(pos, 1, &b);
+
+  if (status < 0)  // error or underflow
+    return status;
+
+  if (status > 0)  // interpreted as "underflow"
+    return E_BUFFER_NOT_FULL;
+
+  if (b == 0)  // we can't handle u-int values larger than 8 bytes
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char m = 0x80;
+
+  while (!(b & m)) {
+    m >>= 1;
+    ++len;
+  }
+
+  long long result = b & (~m);
+  ++pos;
+
+  for (int i = 1; i < len; ++i) {
+    status = pReader->Read(pos, 1, &b);
+
+    if (status < 0) {
+      len = 1;
+      return status;
+    }
+
+    if (status > 0) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result <<= 8;
+    result |= b;
+
+    ++pos;
+  }
+
+  return result;
+}
+
+// Reads an EBML ID and returns it.
+// An ID must at least 1 byte long, cannot exceed 4, and its value must be
+// greater than 0.
+// See known EBML values and EBMLMaxIDLength:
+// http://www.matroska.org/technical/specs/index.html
+// Returns the ID, or a value less than 0 to report an error while reading the
+// ID.
+long long ReadID(IMkvReader* pReader, long long pos, long& len) {
+  if (pReader == NULL || pos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  // Read the first byte. The length in bytes of the ID is determined by
+  // finding the first set bit in the first byte of the ID.
+  unsigned char temp_byte = 0;
+  int read_status = pReader->Read(pos, 1, &temp_byte);
+
+  if (read_status < 0)
+    return E_FILE_FORMAT_INVALID;
+  else if (read_status > 0)  // No data to read.
+    return E_BUFFER_NOT_FULL;
+
+  if (temp_byte == 0)  // ID length > 8 bytes; invalid file.
+    return E_FILE_FORMAT_INVALID;
+
+  int bit_pos = 0;
+  const int kMaxIdLengthInBytes = 4;
+  const int kCheckByte = 0x80;
+
+  // Find the first bit that's set.
+  bool found_bit = false;
+  for (; bit_pos < kMaxIdLengthInBytes; ++bit_pos) {
+    if ((kCheckByte >> bit_pos) & temp_byte) {
+      found_bit = true;
+      break;
+    }
+  }
+
+  if (!found_bit) {
+    // The value is too large to be a valid ID.
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  // Read the remaining bytes of the ID (if any).
+  const int id_length = bit_pos + 1;
+  long long ebml_id = temp_byte;
+  for (int i = 1; i < id_length; ++i) {
+    ebml_id <<= 8;
+    read_status = pReader->Read(pos + i, 1, &temp_byte);
+
+    if (read_status < 0)
+      return E_FILE_FORMAT_INVALID;
+    else if (read_status > 0)
+      return E_BUFFER_NOT_FULL;
+
+    ebml_id |= temp_byte;
+  }
+
+  len = id_length;
+  return ebml_id;
+}
+
+long long GetUIntLength(IMkvReader* pReader, long long pos, long& len) {
+  if (!pReader || pos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  long long total, available;
+
+  int status = pReader->Length(&total, &available);
+  if (status < 0 || (total >= 0 && available > total))
+    return E_FILE_FORMAT_INVALID;
+
+  len = 1;
+
+  if (pos >= available)
+    return pos;  // too few bytes available
+
+  unsigned char b;
+
+  status = pReader->Read(pos, 1, &b);
+
+  if (status != 0)
+    return status;
+
+  if (b == 0)  // we can't handle u-int values larger than 8 bytes
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char m = 0x80;
+
+  while (!(b & m)) {
+    m >>= 1;
+    ++len;
+  }
+
+  return 0;  // success
+}
+
+// TODO(vigneshv): This function assumes that unsigned values never have their
+// high bit set.
+long long UnserializeUInt(IMkvReader* pReader, long long pos, long long size) {
+  if (!pReader || pos < 0 || (size <= 0) || (size > 8))
+    return E_FILE_FORMAT_INVALID;
+
+  long long result = 0;
+
+  for (long long i = 0; i < size; ++i) {
+    unsigned char b;
+
+    const long status = pReader->Read(pos, 1, &b);
+
+    if (status < 0)
+      return status;
+
+    result <<= 8;
+    result |= b;
+
+    ++pos;
+  }
+
+  return result;
+}
+
+long UnserializeFloat(IMkvReader* pReader, long long pos, long long size_,
+                      double& result) {
+  if (!pReader || pos < 0 || ((size_ != 4) && (size_ != 8)))
+    return E_FILE_FORMAT_INVALID;
+
+  const long size = static_cast<long>(size_);
+
+  unsigned char buf[8];
+
+  const int status = pReader->Read(pos, size, buf);
+
+  if (status < 0)  // error
+    return status;
+
+  if (size == 4) {
+    union {
+      float f;
+      unsigned long ff;
+    };
+
+    ff = 0;
+
+    for (int i = 0;;) {
+      ff |= buf[i];
+
+      if (++i >= 4)
+        break;
+
+      ff <<= 8;
+    }
+
+    result = f;
+  } else {
+    union {
+      double d;
+      unsigned long long dd;
+    };
+
+    dd = 0;
+
+    for (int i = 0;;) {
+      dd |= buf[i];
+
+      if (++i >= 8)
+        break;
+
+      dd <<= 8;
+    }
+
+    result = d;
+  }
+
+  if (mkvparser::isinf(result) || mkvparser::isnan(result))
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+long UnserializeInt(IMkvReader* pReader, long long pos, long long size,
+                    long long& result_ref) {
+  if (!pReader || pos < 0 || size < 1 || size > 8)
+    return E_FILE_FORMAT_INVALID;
+
+  signed char first_byte = 0;
+  const long status = pReader->Read(pos, 1, (unsigned char*)&first_byte);
+
+  if (status < 0)
+    return status;
+
+  unsigned long long result = first_byte;
+  ++pos;
+
+  for (long i = 1; i < size; ++i) {
+    unsigned char b;
+
+    const long status = pReader->Read(pos, 1, &b);
+
+    if (status < 0)
+      return status;
+
+    result <<= 8;
+    result |= b;
+
+    ++pos;
+  }
+
+  result_ref = static_cast<long long>(result);
+  return 0;
+}
+
+long UnserializeString(IMkvReader* pReader, long long pos, long long size,
+                       char*& str) {
+  delete[] str;
+  str = NULL;
+
+  if (size >= LONG_MAX || size < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  // +1 for '\0' terminator
+  const long required_size = static_cast<long>(size) + 1;
+
+  str = SafeArrayAlloc<char>(1, required_size);
+  if (str == NULL)
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char* const buf = reinterpret_cast<unsigned char*>(str);
+
+  const long status = pReader->Read(pos, static_cast<long>(size), buf);
+
+  if (status) {
+    delete[] str;
+    str = NULL;
+
+    return status;
+  }
+
+  str[required_size - 1] = '\0';
+  return 0;
+}
+
+long ParseElementHeader(IMkvReader* pReader, long long& pos, long long stop,
+                        long long& id, long long& size) {
+  if (stop >= 0 && pos >= stop)
+    return E_FILE_FORMAT_INVALID;
+
+  long len;
+
+  id = ReadID(pReader, pos, len);
+
+  if (id < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume id
+
+  if (stop >= 0 && pos >= stop)
+    return E_FILE_FORMAT_INVALID;
+
+  size = ReadUInt(pReader, pos, len);
+
+  if (size < 0 || len < 1 || len > 8) {
+    // Invalid: Negative payload size, negative or 0 length integer, or integer
+    // larger than 64 bits (libwebm cannot handle them).
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  // Avoid rolling over pos when very close to LLONG_MAX.
+  const unsigned long long rollover_check =
+      static_cast<unsigned long long>(pos) + len;
+  if (rollover_check > LLONG_MAX)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume length of size
+
+  // pos now designates payload
+
+  if (stop >= 0 && pos > stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;  // success
+}
+
+bool Match(IMkvReader* pReader, long long& pos, unsigned long expected_id,
+           long long& val) {
+  if (!pReader || pos < 0)
+    return false;
+
+  long long total = 0;
+  long long available = 0;
+
+  const long status = pReader->Length(&total, &available);
+  if (status < 0 || (total >= 0 && available > total))
+    return false;
+
+  long len = 0;
+
+  const long long id = ReadID(pReader, pos, len);
+  if (id < 0 || (available - pos) > len)
+    return false;
+
+  if (static_cast<unsigned long>(id) != expected_id)
+    return false;
+
+  pos += len;  // consume id
+
+  const long long size = ReadUInt(pReader, pos, len);
+  if (size < 0 || size > 8 || len < 1 || len > 8 || (available - pos) > len)
+    return false;
+
+  pos += len;  // consume length of size of payload
+
+  val = UnserializeUInt(pReader, pos, size);
+  if (val < 0)
+    return false;
+
+  pos += size;  // consume size of payload
+
+  return true;
+}
+
+bool Match(IMkvReader* pReader, long long& pos, unsigned long expected_id,
+           unsigned char*& buf, size_t& buflen) {
+  if (!pReader || pos < 0)
+    return false;
+
+  long long total = 0;
+  long long available = 0;
+
+  long status = pReader->Length(&total, &available);
+  if (status < 0 || (total >= 0 && available > total))
+    return false;
+
+  long len = 0;
+  const long long id = ReadID(pReader, pos, len);
+  if (id < 0 || (available - pos) > len)
+    return false;
+
+  if (static_cast<unsigned long>(id) != expected_id)
+    return false;
+
+  pos += len;  // consume id
+
+  const long long size = ReadUInt(pReader, pos, len);
+  if (size < 0 || len <= 0 || len > 8 || (available - pos) > len)
+    return false;
+
+  unsigned long long rollover_check =
+      static_cast<unsigned long long>(pos) + len;
+  if (rollover_check > LLONG_MAX)
+    return false;
+
+  pos += len;  // consume length of size of payload
+
+  rollover_check = static_cast<unsigned long long>(pos) + size;
+  if (rollover_check > LLONG_MAX)
+    return false;
+
+  if ((pos + size) > available)
+    return false;
+
+  if (size >= LONG_MAX)
+    return false;
+
+  const long buflen_ = static_cast<long>(size);
+
+  buf = SafeArrayAlloc<unsigned char>(1, buflen_);
+  if (!buf)
+    return false;
+
+  status = pReader->Read(pos, buflen_, buf);
+  if (status != 0)
+    return false;
+
+  buflen = buflen_;
+
+  pos += size;  // consume size of payload
+  return true;
+}
+
+EBMLHeader::EBMLHeader() : m_docType(NULL) { Init(); }
+
+EBMLHeader::~EBMLHeader() { delete[] m_docType; }
+
+void EBMLHeader::Init() {
+  m_version = 1;
+  m_readVersion = 1;
+  m_maxIdLength = 4;
+  m_maxSizeLength = 8;
+
+  if (m_docType) {
+    delete[] m_docType;
+    m_docType = NULL;
+  }
+
+  m_docTypeVersion = 1;
+  m_docTypeReadVersion = 1;
+}
+
+long long EBMLHeader::Parse(IMkvReader* pReader, long long& pos) {
+  if (!pReader)
+    return E_FILE_FORMAT_INVALID;
+
+  long long total, available;
+
+  long status = pReader->Length(&total, &available);
+
+  if (status < 0)  // error
+    return status;
+
+  pos = 0;
+
+  // Scan until we find what looks like the first byte of the EBML header.
+  const long long kMaxScanBytes = (available >= 1024) ? 1024 : available;
+  const unsigned char kEbmlByte0 = 0x1A;
+  unsigned char scan_byte = 0;
+
+  while (pos < kMaxScanBytes) {
+    status = pReader->Read(pos, 1, &scan_byte);
+
+    if (status < 0)  // error
+      return status;
+    else if (status > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if (scan_byte == kEbmlByte0)
+      break;
+
+    ++pos;
+  }
+
+  long len = 0;
+  const long long ebml_id = ReadID(pReader, pos, len);
+
+  if (ebml_id == E_BUFFER_NOT_FULL)
+    return E_BUFFER_NOT_FULL;
+
+  if (len != 4 || ebml_id != libwebm::kMkvEBML)
+    return E_FILE_FORMAT_INVALID;
+
+  // Move read pos forward to the EBML header size field.
+  pos += 4;
+
+  // Read length of size field.
+  long long result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error
+    return E_FILE_FORMAT_INVALID;
+  else if (result > 0)  // need more data
+    return E_BUFFER_NOT_FULL;
+
+  if (len < 1 || len > 8)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((total >= 0) && ((total - pos) < len))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((available - pos) < len)
+    return pos + len;  // try again later
+
+  // Read the EBML header size.
+  result = ReadUInt(pReader, pos, len);
+
+  if (result < 0)  // error
+    return result;
+
+  pos += len;  // consume size field
+
+  // pos now designates start of payload
+
+  if ((total >= 0) && ((total - pos) < result))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((available - pos) < result)
+    return pos + result;
+
+  const long long end = pos + result;
+
+  Init();
+
+  while (pos < end) {
+    long long id, size;
+
+    status = ParseElementHeader(pReader, pos, end, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvEBMLVersion) {
+      m_version = UnserializeUInt(pReader, pos, size);
+
+      if (m_version <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvEBMLReadVersion) {
+      m_readVersion = UnserializeUInt(pReader, pos, size);
+
+      if (m_readVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvEBMLMaxIDLength) {
+      m_maxIdLength = UnserializeUInt(pReader, pos, size);
+
+      if (m_maxIdLength <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvEBMLMaxSizeLength) {
+      m_maxSizeLength = UnserializeUInt(pReader, pos, size);
+
+      if (m_maxSizeLength <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDocType) {
+      if (m_docType)
+        return E_FILE_FORMAT_INVALID;
+
+      status = UnserializeString(pReader, pos, size, m_docType);
+
+      if (status)  // error
+        return status;
+    } else if (id == libwebm::kMkvDocTypeVersion) {
+      m_docTypeVersion = UnserializeUInt(pReader, pos, size);
+
+      if (m_docTypeVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDocTypeReadVersion) {
+      m_docTypeReadVersion = UnserializeUInt(pReader, pos, size);
+
+      if (m_docTypeReadVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;
+  }
+
+  if (pos != end)
+    return E_FILE_FORMAT_INVALID;
+
+  // Make sure DocType, DocTypeReadVersion, and DocTypeVersion are valid.
+  if (m_docType == NULL || m_docTypeReadVersion <= 0 || m_docTypeVersion <= 0)
+    return E_FILE_FORMAT_INVALID;
+
+  // Make sure EBMLMaxIDLength and EBMLMaxSizeLength are valid.
+  if (m_maxIdLength <= 0 || m_maxIdLength > 4 || m_maxSizeLength <= 0 ||
+      m_maxSizeLength > 8)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+Segment::Segment(IMkvReader* pReader, long long elem_start,
+                 // long long elem_size,
+                 long long start, long long size)
+    : m_pReader(pReader),
+      m_element_start(elem_start),
+      // m_element_size(elem_size),
+      m_start(start),
+      m_size(size),
+      m_pos(start),
+      m_pUnknownSize(0),
+      m_pSeekHead(NULL),
+      m_pInfo(NULL),
+      m_pTracks(NULL),
+      m_pCues(NULL),
+      m_pChapters(NULL),
+      m_pTags(NULL),
+      m_clusters(NULL),
+      m_clusterCount(0),
+      m_clusterPreloadCount(0),
+      m_clusterSize(0) {}
+
+Segment::~Segment() {
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  Cluster** i = m_clusters;
+  Cluster** j = m_clusters + count;
+
+  while (i != j) {
+    Cluster* const p = *i++;
+    delete p;
+  }
+
+  delete[] m_clusters;
+
+  delete m_pTracks;
+  delete m_pInfo;
+  delete m_pCues;
+  delete m_pChapters;
+  delete m_pTags;
+  delete m_pSeekHead;
+}
+
+long long Segment::CreateInstance(IMkvReader* pReader, long long pos,
+                                  Segment*& pSegment) {
+  if (pReader == NULL || pos < 0)
+    return E_PARSE_FAILED;
+
+  pSegment = NULL;
+
+  long long total, available;
+
+  const long status = pReader->Length(&total, &available);
+
+  if (status < 0)  // error
+    return status;
+
+  if (available < 0)
+    return -1;
+
+  if ((total >= 0) && (available > total))
+    return -1;
+
+  // I would assume that in practice this loop would execute
+  // exactly once, but we allow for other elements (e.g. Void)
+  // to immediately follow the EBML header.  This is fine for
+  // the source filter case (since the entire file is available),
+  // but in the splitter case over a network we should probably
+  // just give up early.  We could for example decide only to
+  // execute this loop a maximum of, say, 10 times.
+  // TODO:
+  // There is an implied "give up early" by only parsing up
+  // to the available limit.  We do do that, but only if the
+  // total file size is unknown.  We could decide to always
+  // use what's available as our limit (irrespective of whether
+  // we happen to know the total file length).  This would have
+  // as its sense "parse this much of the file before giving up",
+  // which a slightly different sense from "try to parse up to
+  // 10 EMBL elements before giving up".
+
+  for (;;) {
+    if ((total >= 0) && (pos >= total))
+      return E_FILE_FORMAT_INVALID;
+
+    // Read ID
+    long len;
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result)  // error, or too few available bytes
+      return result;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    const long long idpos = pos;
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result)  // error, or too few available bytes
+      return result;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return size;
+
+    pos += len;  // consume length of size of element
+
+    // Pos now points to start of payload
+
+    // Handle "unknown size" for live streaming of webm files.
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (id == libwebm::kMkvSegment) {
+      if (size == unknown_size)
+        size = -1;
+
+      else if (total < 0)
+        size = -1;
+
+      else if ((pos + size) > total)
+        size = -1;
+
+      pSegment = new (std::nothrow) Segment(pReader, idpos, pos, size);
+      if (pSegment == NULL)
+        return E_PARSE_FAILED;
+
+      return 0;  // success
+    }
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && ((pos + size) > total))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + size) > available)
+      return pos + size;
+
+    pos += size;  // consume payload
+  }
+}
+
+long long Segment::ParseHeaders() {
+  // Outermost (level 0) segment object has been constructed,
+  // and pos designates start of payload.  We need to find the
+  // inner (level 1) elements.
+  long long total, available;
+
+  const int status = m_pReader->Length(&total, &available);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total > 0 && available > total)
+    return E_FILE_FORMAT_INVALID;
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  if ((segment_stop >= 0 && total >= 0 && segment_stop > total) ||
+      (segment_stop >= 0 && m_pos > segment_stop)) {
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  for (;;) {
+    if ((total >= 0) && (m_pos >= total))
+      break;
+
+    if ((segment_stop >= 0) && (m_pos >= segment_stop))
+      break;
+
+    long long pos = m_pos;
+    const long long element_start = pos;
+
+    // Avoid rolling over pos when very close to LLONG_MAX.
+    unsigned long long rollover_check = pos + 1ULL;
+    if (rollover_check > LLONG_MAX)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + 1) > available)
+      return (pos + 1);
+
+    long len;
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return result;
+
+    if (result > 0) {
+      // MkvReader doesn't have enough data to satisfy this read attempt.
+      return (pos + 1);
+    }
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    const long long idpos = pos;
+    const long long id = ReadID(m_pReader, idpos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvCluster)
+      break;
+
+    pos += len;  // consume ID
+
+    if ((pos + 1) > available)
+      return (pos + 1);
+
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return result;
+
+    if (result > 0) {
+      // MkvReader doesn't have enough data to satisfy this read attempt.
+      return (pos + 1);
+    }
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0 || len < 1 || len > 8) {
+      // TODO(tomfinegan): ReadUInt should return an error when len is < 1 or
+      // len > 8 is true instead of checking this _everywhere_.
+      return size;
+    }
+
+    pos += len;  // consume length of size of element
+
+    // Avoid rolling over pos when very close to LLONG_MAX.
+    rollover_check = static_cast<unsigned long long>(pos) + size;
+    if (rollover_check > LLONG_MAX)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long element_size = size + pos - element_start;
+
+    // Pos now points to start of payload
+
+    if ((segment_stop >= 0) && ((pos + size) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // We read EBML elements either in total or nothing at all.
+
+    if ((pos + size) > available)
+      return pos + size;
+
+    if (id == libwebm::kMkvInfo) {
+      if (m_pInfo)
+        return E_FILE_FORMAT_INVALID;
+
+      m_pInfo = new (std::nothrow)
+          SegmentInfo(this, pos, size, element_start, element_size);
+
+      if (m_pInfo == NULL)
+        return -1;
+
+      const long status = m_pInfo->Parse();
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvTracks) {
+      if (m_pTracks)
+        return E_FILE_FORMAT_INVALID;
+
+      m_pTracks = new (std::nothrow)
+          Tracks(this, pos, size, element_start, element_size);
+
+      if (m_pTracks == NULL)
+        return -1;
+
+      const long status = m_pTracks->Parse();
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvCues) {
+      if (m_pCues == NULL) {
+        m_pCues = new (std::nothrow)
+            Cues(this, pos, size, element_start, element_size);
+
+        if (m_pCues == NULL)
+          return -1;
+      }
+    } else if (id == libwebm::kMkvSeekHead) {
+      if (m_pSeekHead == NULL) {
+        m_pSeekHead = new (std::nothrow)
+            SeekHead(this, pos, size, element_start, element_size);
+
+        if (m_pSeekHead == NULL)
+          return -1;
+
+        const long status = m_pSeekHead->Parse();
+
+        if (status)
+          return status;
+      }
+    } else if (id == libwebm::kMkvChapters) {
+      if (m_pChapters == NULL) {
+        m_pChapters = new (std::nothrow)
+            Chapters(this, pos, size, element_start, element_size);
+
+        if (m_pChapters == NULL)
+          return -1;
+
+        const long status = m_pChapters->Parse();
+
+        if (status)
+          return status;
+      }
+    } else if (id == libwebm::kMkvTags) {
+      if (m_pTags == NULL) {
+        m_pTags = new (std::nothrow)
+            Tags(this, pos, size, element_start, element_size);
+
+        if (m_pTags == NULL)
+          return -1;
+
+        const long status = m_pTags->Parse();
+
+        if (status)
+          return status;
+      }
+    }
+
+    m_pos = pos + size;  // consume payload
+  }
+
+  if (segment_stop >= 0 && m_pos > segment_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (m_pInfo == NULL)  // TODO: liberalize this behavior
+    return E_FILE_FORMAT_INVALID;
+
+  if (m_pTracks == NULL)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;  // success
+}
+
+long Segment::LoadCluster(long long& pos, long& len) {
+  for (;;) {
+    const long result = DoLoadCluster(pos, len);
+
+    if (result <= 1)
+      return result;
+  }
+}
+
+long Segment::DoLoadCluster(long long& pos, long& len) {
+  if (m_pos < 0)
+    return DoLoadClusterUnknownSize(pos, len);
+
+  long long total, avail;
+
+  long status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total >= 0 && avail > total)
+    return E_FILE_FORMAT_INVALID;
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  long long cluster_off = -1;  // offset relative to start of segment
+  long long cluster_size = -1;  // size of cluster payload
+
+  for (;;) {
+    if ((total >= 0) && (m_pos >= total))
+      return 1;  // no more clusters
+
+    if ((segment_stop >= 0) && (m_pos >= segment_stop))
+      return 1;  // no more clusters
+
+    pos = m_pos;
+
+    // Read ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long idpos = pos;
+    const long long id = ReadID(m_pReader, idpos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume length of size of element
+
+    // pos now points to start of payload
+
+    if (size == 0) {
+      // Missing element payload: move on.
+      m_pos = pos;
+      continue;
+    }
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if ((segment_stop >= 0) && (size != unknown_size) &&
+        ((pos + size) > segment_stop)) {
+      return E_FILE_FORMAT_INVALID;
+    }
+
+    if (id == libwebm::kMkvCues) {
+      if (size == unknown_size) {
+        // Cues element of unknown size: Not supported.
+        return E_FILE_FORMAT_INVALID;
+      }
+
+      if (m_pCues == NULL) {
+        const long long element_size = (pos - idpos) + size;
+
+        m_pCues = new (std::nothrow) Cues(this, pos, size, idpos, element_size);
+        if (m_pCues == NULL)
+          return -1;
+      }
+
+      m_pos = pos + size;  // consume payload
+      continue;
+    }
+
+    if (id != libwebm::kMkvCluster) {
+      // Besides the Segment, Libwebm allows only cluster elements of unknown
+      // size. Fail the parse upon encountering a non-cluster element reporting
+      // unknown size.
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
+
+      m_pos = pos + size;  // consume payload
+      continue;
+    }
+
+    // We have a cluster.
+
+    cluster_off = idpos - m_start;  // relative pos
+
+    if (size != unknown_size)
+      cluster_size = size;
+
+    break;
+  }
+
+  if (cluster_off < 0) {
+    // No cluster, die.
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  long long pos_;
+  long len_;
+
+  status = Cluster::HasBlockEntries(this, cluster_off, pos_, len_);
+
+  if (status < 0) {  // error, or underflow
+    pos = pos_;
+    len = len_;
+
+    return status;
+  }
+
+  // status == 0 means "no block entries found"
+  // status > 0 means "found at least one block entry"
+
+  // TODO:
+  // The issue here is that the segment increments its own
+  // pos ptr past the most recent cluster parsed, and then
+  // starts from there to parse the next cluster.  If we
+  // don't know the size of the current cluster, then we
+  // must either parse its payload (as we do below), looking
+  // for the cluster (or cues) ID to terminate the parse.
+  // This isn't really what we want: rather, we really need
+  // a way to create the curr cluster object immediately.
+  // The pity is that cluster::parse can determine its own
+  // boundary, and we largely duplicate that same logic here.
+  //
+  // Maybe we need to get rid of our look-ahead preloading
+  // in source::parse???
+  //
+  // As we're parsing the blocks in the curr cluster
+  //(in cluster::parse), we should have some way to signal
+  // to the segment that we have determined the boundary,
+  // so it can adjust its own segment::m_pos member.
+  //
+  // The problem is that we're asserting in asyncreadinit,
+  // because we adjust the pos down to the curr seek pos,
+  // and the resulting adjusted len is > 2GB.  I'm suspicious
+  // that this is even correct, but even if it is, we can't
+  // be loading that much data in the cache anyway.
+
+  const long idx = m_clusterCount;
+
+  if (m_clusterPreloadCount > 0) {
+    if (idx >= m_clusterSize)
+      return E_FILE_FORMAT_INVALID;
+
+    Cluster* const pCluster = m_clusters[idx];
+    if (pCluster == NULL || pCluster->m_index >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long off = pCluster->GetPosition();
+    if (off < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (off == cluster_off) {  // preloaded already
+      if (status == 0)  // no entries found
+        return E_FILE_FORMAT_INVALID;
+
+      if (cluster_size >= 0)
+        pos += cluster_size;
+      else {
+        const long long element_size = pCluster->GetElementSize();
+
+        if (element_size <= 0)
+          return E_FILE_FORMAT_INVALID;  // TODO: handle this case
+
+        pos = pCluster->m_element_start + element_size;
+      }
+
+      pCluster->m_index = idx;  // move from preloaded to loaded
+      ++m_clusterCount;
+      --m_clusterPreloadCount;
+
+      m_pos = pos;  // consume payload
+      if (segment_stop >= 0 && m_pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      return 0;  // success
+    }
+  }
+
+  if (status == 0) {  // no entries found
+    if (cluster_size >= 0)
+      pos += cluster_size;
+
+    if ((total >= 0) && (pos >= total)) {
+      m_pos = total;
+      return 1;  // no more clusters
+    }
+
+    if ((segment_stop >= 0) && (pos >= segment_stop)) {
+      m_pos = segment_stop;
+      return 1;  // no more clusters
+    }
+
+    m_pos = pos;
+    return 2;  // try again
+  }
+
+  // status > 0 means we have an entry
+
+  Cluster* const pCluster = Cluster::Create(this, idx, cluster_off);
+  if (pCluster == NULL)
+    return -1;
+
+  if (!AppendCluster(pCluster)) {
+    delete pCluster;
+    return -1;
+  }
+
+  if (cluster_size >= 0) {
+    pos += cluster_size;
+
+    m_pos = pos;
+
+    if (segment_stop > 0 && m_pos > segment_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    return 0;
+  }
+
+  m_pUnknownSize = pCluster;
+  m_pos = -pos;
+
+  return 0;  // partial success, since we have a new cluster
+
+  // status == 0 means "no block entries found"
+  // pos designates start of payload
+  // m_pos has NOT been adjusted yet (in case we need to come back here)
+}
+
+long Segment::DoLoadClusterUnknownSize(long long& pos, long& len) {
+  if (m_pos >= 0 || m_pUnknownSize == NULL)
+    return E_PARSE_FAILED;
+
+  const long status = m_pUnknownSize->Parse(pos, len);
+
+  if (status < 0)  // error or underflow
+    return status;
+
+  if (status == 0)  // parsed a block
+    return 2;  // continue parsing
+
+  const long long start = m_pUnknownSize->m_element_start;
+  const long long size = m_pUnknownSize->GetElementSize();
+
+  if (size < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos = start + size;
+  m_pos = pos;
+
+  m_pUnknownSize = 0;
+
+  return 2;  // continue parsing
+}
+
+bool Segment::AppendCluster(Cluster* pCluster) {
+  if (pCluster == NULL || pCluster->m_index < 0)
+    return false;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  long& size = m_clusterSize;
+  const long idx = pCluster->m_index;
+
+  if (size < count || idx != m_clusterCount)
+    return false;
+
+  if (count >= size) {
+    const long n = (size <= 0) ? 2048 : 2 * size;
+
+    Cluster** const qq = new (std::nothrow) Cluster*[n];
+    if (qq == NULL)
+      return false;
+
+    Cluster** q = qq;
+    Cluster** p = m_clusters;
+    Cluster** const pp = p + count;
+
+    while (p != pp)
+      *q++ = *p++;
+
+    delete[] m_clusters;
+
+    m_clusters = qq;
+    size = n;
+  }
+
+  if (m_clusterPreloadCount > 0) {
+    Cluster** const p = m_clusters + m_clusterCount;
+    if (*p == NULL || (*p)->m_index >= 0)
+      return false;
+
+    Cluster** q = p + m_clusterPreloadCount;
+    if (q >= (m_clusters + size))
+      return false;
+
+    for (;;) {
+      Cluster** const qq = q - 1;
+      if ((*qq)->m_index >= 0)
+        return false;
+
+      *q = *qq;
+      q = qq;
+
+      if (q == p)
+        break;
+    }
+  }
+
+  m_clusters[idx] = pCluster;
+  ++m_clusterCount;
+  return true;
+}
+
+bool Segment::PreloadCluster(Cluster* pCluster, ptrdiff_t idx) {
+  if (pCluster == NULL || pCluster->m_index >= 0 || idx < m_clusterCount)
+    return false;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  long& size = m_clusterSize;
+  if (size < count)
+    return false;
+
+  if (count >= size) {
+    const long n = (size <= 0) ? 2048 : 2 * size;
+
+    Cluster** const qq = new (std::nothrow) Cluster*[n];
+    if (qq == NULL)
+      return false;
+    Cluster** q = qq;
+
+    Cluster** p = m_clusters;
+    Cluster** const pp = p + count;
+
+    while (p != pp)
+      *q++ = *p++;
+
+    delete[] m_clusters;
+
+    m_clusters = qq;
+    size = n;
+  }
+
+  if (m_clusters == NULL)
+    return false;
+
+  Cluster** const p = m_clusters + idx;
+
+  Cluster** q = m_clusters + count;
+  if (q < p || q >= (m_clusters + size))
+    return false;
+
+  while (q > p) {
+    Cluster** const qq = q - 1;
+
+    if ((*qq)->m_index >= 0)
+      return false;
+
+    *q = *qq;
+    q = qq;
+  }
+
+  m_clusters[idx] = pCluster;
+  ++m_clusterPreloadCount;
+  return true;
+}
+
+long Segment::Load() {
+  if (m_clusters != NULL || m_clusterSize != 0 || m_clusterCount != 0)
+    return E_PARSE_FAILED;
+
+  // Outermost (level 0) segment object has been constructed,
+  // and pos designates start of payload.  We need to find the
+  // inner (level 1) elements.
+
+  const long long header_status = ParseHeaders();
+
+  if (header_status < 0)  // error
+    return static_cast<long>(header_status);
+
+  if (header_status > 0)  // underflow
+    return E_BUFFER_NOT_FULL;
+
+  if (m_pInfo == NULL || m_pTracks == NULL)
+    return E_FILE_FORMAT_INVALID;
+
+  for (;;) {
+    const int status = LoadCluster();
+
+    if (status < 0)  // error
+      return status;
+
+    if (status >= 1)  // no more clusters
+      return 0;
+  }
+}
+
+SeekHead::SeekHead(Segment* pSegment, long long start, long long size_,
+                   long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_entries(0),
+      m_entry_count(0),
+      m_void_elements(0),
+      m_void_element_count(0) {}
+
+SeekHead::~SeekHead() {
+  delete[] m_entries;
+  delete[] m_void_elements;
+}
+
+long SeekHead::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
+
+  // first count the seek head entries
+
+  int entry_count = 0;
+  int void_element_count = 0;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvSeek)
+      ++entry_count;
+    else if (id == libwebm::kMkvVoid)
+      ++void_element_count;
+
+    pos += size;  // consume payload
+
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  m_entries = new (std::nothrow) Entry[entry_count];
+
+  if (m_entries == NULL)
+    return -1;
+
+  m_void_elements = new (std::nothrow) VoidElement[void_element_count];
+
+  if (m_void_elements == NULL)
+    return -1;
+
+  // now parse the entries and void elements
+
+  Entry* pEntry = m_entries;
+  VoidElement* pVoidElement = m_void_elements;
+
+  pos = m_start;
+
+  while (pos < stop) {
+    const long long idpos = pos;
+
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvSeek) {
+      if (ParseEntry(pReader, pos, size, pEntry)) {
+        Entry& e = *pEntry++;
+
+        e.element_start = idpos;
+        e.element_size = (pos + size) - idpos;
+      }
+    } else if (id == libwebm::kMkvVoid) {
+      VoidElement& e = *pVoidElement++;
+
+      e.element_start = idpos;
+      e.element_size = (pos + size) - idpos;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  ptrdiff_t count_ = ptrdiff_t(pEntry - m_entries);
+  assert(count_ >= 0);
+  assert(count_ <= entry_count);
+
+  m_entry_count = static_cast<int>(count_);
+
+  count_ = ptrdiff_t(pVoidElement - m_void_elements);
+  assert(count_ >= 0);
+  assert(count_ <= void_element_count);
+
+  m_void_element_count = static_cast<int>(count_);
+
+  return 0;
+}
+
+int SeekHead::GetCount() const { return m_entry_count; }
+
+const SeekHead::Entry* SeekHead::GetEntry(int idx) const {
+  if (idx < 0)
+    return 0;
+
+  if (idx >= m_entry_count)
+    return 0;
+
+  return m_entries + idx;
+}
+
+int SeekHead::GetVoidElementCount() const { return m_void_element_count; }
+
+const SeekHead::VoidElement* SeekHead::GetVoidElement(int idx) const {
+  if (idx < 0)
+    return 0;
+
+  if (idx >= m_void_element_count)
+    return 0;
+
+  return m_void_elements + idx;
+}
+
+long Segment::ParseCues(long long off, long long& pos, long& len) {
+  if (m_pCues)
+    return 0;  // success
+
+  if (off < 0)
+    return -1;
+
+  long long total, avail;
+
+  const int status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  pos = m_start + off;
+
+  if ((total < 0) || (pos >= total))
+    return 1;  // don't bother parsing cues
+
+  const long long element_start = pos;
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long result = GetUIntLength(m_pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)  // underflow (weird)
+  {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long idpos = pos;
+
+  const long long id = ReadID(m_pReader, idpos, len);
+
+  if (id != libwebm::kMkvCues)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume ID
+  assert((segment_stop < 0) || (pos <= segment_stop));
+
+  // Read Size
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  result = GetUIntLength(m_pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)  // underflow (weird)
+  {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long size = ReadUInt(m_pReader, pos, len);
+
+  if (size < 0)  // error
+    return static_cast<long>(size);
+
+  if (size == 0)  // weird, although technically not illegal
+    return 1;  // done
+
+  pos += len;  // consume length of size of element
+  assert((segment_stop < 0) || (pos <= segment_stop));
+
+  // Pos now points to start of payload
+
+  const long long element_stop = pos + size;
+
+  if ((segment_stop >= 0) && (element_stop > segment_stop))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((total >= 0) && (element_stop > total))
+    return 1;  // don't bother parsing anymore
+
+  len = static_cast<long>(size);
+
+  if (element_stop > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long element_size = element_stop - element_start;
+
+  m_pCues =
+      new (std::nothrow) Cues(this, pos, size, element_start, element_size);
+  if (m_pCues == NULL)
+    return -1;
+
+  return 0;  // success
+}
+
+bool SeekHead::ParseEntry(IMkvReader* pReader, long long start, long long size_,
+                          Entry* pEntry) {
+  if (size_ <= 0)
+    return false;
+
+  long long pos = start;
+  const long long stop = start + size_;
+
+  long len;
+
+  // parse the container for the level-1 element ID
+
+  const long long seekIdId = ReadID(pReader, pos, len);
+  if (seekIdId < 0)
+    return false;
+
+  if (seekIdId != libwebm::kMkvSeekID)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume SeekID id
+
+  const long long seekIdSize = ReadUInt(pReader, pos, len);
+
+  if (seekIdSize <= 0)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume size of field
+
+  if ((pos + seekIdSize) > stop)
+    return false;
+
+  // Note that the SeekId payload really is serialized
+  // as a "Matroska integer", not as a plain binary value.
+  // In fact, Matroska requires that ID values in the
+  // stream exactly match the binary representation as listed
+  // in the Matroska specification.
+  //
+  // This parser is more liberal, and permits IDs to have
+  // any width.  (This could make the representation in the stream
+  // different from what's in the spec, but it doesn't matter here,
+  // since we always normalize "Matroska integer" values.)
+
+  pEntry->id = ReadUInt(pReader, pos, len);  // payload
+
+  if (pEntry->id <= 0)
+    return false;
+
+  if (len != seekIdSize)
+    return false;
+
+  pos += seekIdSize;  // consume SeekID payload
+
+  const long long seekPosId = ReadID(pReader, pos, len);
+
+  if (seekPosId != libwebm::kMkvSeekPosition)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume id
+
+  const long long seekPosSize = ReadUInt(pReader, pos, len);
+
+  if (seekPosSize <= 0)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume size
+
+  if ((pos + seekPosSize) > stop)
+    return false;
+
+  pEntry->pos = UnserializeUInt(pReader, pos, seekPosSize);
+
+  if (pEntry->pos < 0)
+    return false;
+
+  pos += seekPosSize;  // consume payload
+
+  if (pos != stop)
+    return false;
+
+  return true;
+}
+
+Cues::Cues(Segment* pSegment, long long start_, long long size_,
+           long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start_),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_cue_points(NULL),
+      m_count(0),
+      m_preload_count(0),
+      m_pos(start_) {}
+
+Cues::~Cues() {
+  const long n = m_count + m_preload_count;
+
+  CuePoint** p = m_cue_points;
+  CuePoint** const q = p + n;
+
+  while (p != q) {
+    CuePoint* const pCP = *p++;
+    assert(pCP);
+
+    delete pCP;
+  }
+
+  delete[] m_cue_points;
+}
+
+long Cues::GetCount() const {
+  if (m_cue_points == NULL)
+    return -1;
+
+  return m_count;  // TODO: really ignore preload count?
+}
+
+bool Cues::DoneParsing() const {
+  const long long stop = m_start + m_size;
+  return (m_pos >= stop);
+}
+
+bool Cues::Init() const {
+  if (m_cue_points)
+    return true;
+
+  if (m_count != 0 || m_preload_count != 0)
+    return false;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  const long long stop = m_start + m_size;
+  long long pos = m_start;
+
+  long cue_points_size = 0;
+
+  while (pos < stop) {
+    const long long idpos = pos;
+
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if (id < 0 || (pos + len) > stop) {
+      return false;
+    }
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    if (size < 0 || (pos + len > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume Size field
+    if (pos + size > stop) {
+      return false;
+    }
+
+    if (id == libwebm::kMkvCuePoint) {
+      if (!PreloadCuePoint(cue_points_size, idpos))
+        return false;
+    }
+
+    pos += size;  // skip payload
+  }
+  return true;
+}
+
+bool Cues::PreloadCuePoint(long& cue_points_size, long long pos) const {
+  if (m_count != 0)
+    return false;
+
+  if (m_preload_count >= cue_points_size) {
+    const long n = (cue_points_size <= 0) ? 2048 : 2 * cue_points_size;
+
+    CuePoint** const qq = new (std::nothrow) CuePoint*[n];
+    if (qq == NULL)
+      return false;
+
+    CuePoint** q = qq;  // beginning of target
+
+    CuePoint** p = m_cue_points;  // beginning of source
+    CuePoint** const pp = p + m_preload_count;  // end of source
+
+    while (p != pp)
+      *q++ = *p++;
+
+    delete[] m_cue_points;
+
+    m_cue_points = qq;
+    cue_points_size = n;
+  }
+
+  CuePoint* const pCP = new (std::nothrow) CuePoint(m_preload_count, pos);
+  if (pCP == NULL)
+    return false;
+
+  m_cue_points[m_preload_count++] = pCP;
+  return true;
+}
+
+bool Cues::LoadCuePoint() const {
+  const long long stop = m_start + m_size;
+
+  if (m_pos >= stop)
+    return false;  // nothing else to do
+
+  if (!Init()) {
+    m_pos = stop;
+    return false;
+  }
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  while (m_pos < stop) {
+    const long long idpos = m_pos;
+
+    long len;
+
+    const long long id = ReadID(pReader, m_pos, len);
+    if (id < 0 || (m_pos + len) > stop)
+      return false;
+
+    m_pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, m_pos, len);
+    if (size < 0 || (m_pos + len) > stop)
+      return false;
+
+    m_pos += len;  // consume Size field
+    if ((m_pos + size) > stop)
+      return false;
+
+    if (id != libwebm::kMkvCuePoint) {
+      m_pos += size;  // consume payload
+      if (m_pos > stop)
+        return false;
+
+      continue;
+    }
+
+    if (m_preload_count < 1)
+      return false;
+
+    CuePoint* const pCP = m_cue_points[m_count];
+    if (!pCP || (pCP->GetTimeCode() < 0 && (-pCP->GetTimeCode() != idpos)))
+      return false;
+
+    if (!pCP->Load(pReader)) {
+      m_pos = stop;
+      return false;
+    }
+    ++m_count;
+    --m_preload_count;
+
+    m_pos += size;  // consume payload
+    if (m_pos > stop)
+      return false;
+
+    return true;  // yes, we loaded a cue point
+  }
+
+  return false;  // no, we did not load a cue point
+}
+
+bool Cues::Find(long long time_ns, const Track* pTrack, const CuePoint*& pCP,
+                const CuePoint::TrackPosition*& pTP) const {
+  if (time_ns < 0 || pTrack == NULL || m_cue_points == NULL || m_count == 0)
+    return false;
+
+  CuePoint** const ii = m_cue_points;
+  CuePoint** i = ii;
+
+  CuePoint** const jj = ii + m_count;
+  CuePoint** j = jj;
+
+  pCP = *i;
+  if (pCP == NULL)
+    return false;
+
+  if (time_ns <= pCP->GetTime(m_pSegment)) {
+    pTP = pCP->Find(pTrack);
+    return (pTP != NULL);
+  }
+
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) <= time_ns
+    //[i, j)  ?
+    //[j, jj) > time_ns
+
+    CuePoint** const k = i + (j - i) / 2;
+    if (k >= jj)
+      return false;
+
+    CuePoint* const pCP = *k;
+    if (pCP == NULL)
+      return false;
+
+    const long long t = pCP->GetTime(m_pSegment);
+
+    if (t <= time_ns)
+      i = k + 1;
+    else
+      j = k;
+
+    if (i > j)
+      return false;
+  }
+
+  if (i != j || i > jj || i <= ii)
+    return false;
+
+  pCP = *--i;
+
+  if (pCP == NULL || pCP->GetTime(m_pSegment) > time_ns)
+    return false;
+
+  // TODO: here and elsewhere, it's probably not correct to search
+  // for the cue point with this time, and then search for a matching
+  // track.  In principle, the matching track could be on some earlier
+  // cue point, and with our current algorithm, we'd miss it.  To make
+  // this bullet-proof, we'd need to create a secondary structure,
+  // with a list of cue points that apply to a track, and then search
+  // that track-based structure for a matching cue point.
+
+  pTP = pCP->Find(pTrack);
+  return (pTP != NULL);
+}
+
+const CuePoint* Cues::GetFirst() const {
+  if (m_cue_points == NULL || m_count == 0)
+    return NULL;
+
+  CuePoint* const* const pp = m_cue_points;
+  if (pp == NULL)
+    return NULL;
+
+  CuePoint* const pCP = pp[0];
+  if (pCP == NULL || pCP->GetTimeCode() < 0)
+    return NULL;
+
+  return pCP;
+}
+
+const CuePoint* Cues::GetLast() const {
+  if (m_cue_points == NULL || m_count <= 0)
+    return NULL;
+
+  const long index = m_count - 1;
+
+  CuePoint* const* const pp = m_cue_points;
+  if (pp == NULL)
+    return NULL;
+
+  CuePoint* const pCP = pp[index];
+  if (pCP == NULL || pCP->GetTimeCode() < 0)
+    return NULL;
+
+  return pCP;
+}
+
+const CuePoint* Cues::GetNext(const CuePoint* pCurr) const {
+  if (pCurr == NULL || pCurr->GetTimeCode() < 0 || m_cue_points == NULL ||
+      m_count < 1) {
+    return NULL;
+  }
+
+  long index = pCurr->m_index;
+  if (index >= m_count)
+    return NULL;
+
+  CuePoint* const* const pp = m_cue_points;
+  if (pp == NULL || pp[index] != pCurr)
+    return NULL;
+
+  ++index;
+
+  if (index >= m_count)
+    return NULL;
+
+  CuePoint* const pNext = pp[index];
+
+  if (pNext == NULL || pNext->GetTimeCode() < 0)
+    return NULL;
+
+  return pNext;
+}
+
+const BlockEntry* Cues::GetBlock(const CuePoint* pCP,
+                                 const CuePoint::TrackPosition* pTP) const {
+  if (pCP == NULL || pTP == NULL)
+    return NULL;
+
+  return m_pSegment->GetBlock(*pCP, *pTP);
+}
+
+const BlockEntry* Segment::GetBlock(const CuePoint& cp,
+                                    const CuePoint::TrackPosition& tp) {
+  Cluster** const ii = m_clusters;
+  Cluster** i = ii;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  Cluster** const jj = ii + count;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) < pTP->m_pos
+    //[i, j) ?
+    //[j, jj)  > pTP->m_pos
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    Cluster* const pCluster = *k;
+    assert(pCluster);
+
+    // const long long pos_ = pCluster->m_pos;
+    // assert(pos_);
+    // const long long pos = pos_ * ((pos_ < 0) ? -1 : 1);
+
+    const long long pos = pCluster->GetPosition();
+    assert(pos >= 0);
+
+    if (pos < tp.m_pos)
+      i = k + 1;
+    else if (pos > tp.m_pos)
+      j = k;
+    else
+      return pCluster->GetEntry(cp, tp);
+  }
+
+  assert(i == j);
+  // assert(Cluster::HasBlockEntries(this, tp.m_pos));
+
+  Cluster* const pCluster = Cluster::Create(this, -1, tp.m_pos);  //, -1);
+  if (pCluster == NULL)
+    return NULL;
+
+  const ptrdiff_t idx = i - m_clusters;
+
+  if (!PreloadCluster(pCluster, idx)) {
+    delete pCluster;
+    return NULL;
+  }
+  assert(m_clusters);
+  assert(m_clusterPreloadCount > 0);
+  assert(m_clusters[idx] == pCluster);
+
+  return pCluster->GetEntry(cp, tp);
+}
+
+const Cluster* Segment::FindOrPreloadCluster(long long requested_pos) {
+  if (requested_pos < 0)
+    return 0;
+
+  Cluster** const ii = m_clusters;
+  Cluster** i = ii;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  Cluster** const jj = ii + count;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) < pTP->m_pos
+    //[i, j) ?
+    //[j, jj)  > pTP->m_pos
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    Cluster* const pCluster = *k;
+    assert(pCluster);
+
+    // const long long pos_ = pCluster->m_pos;
+    // assert(pos_);
+    // const long long pos = pos_ * ((pos_ < 0) ? -1 : 1);
+
+    const long long pos = pCluster->GetPosition();
+    assert(pos >= 0);
+
+    if (pos < requested_pos)
+      i = k + 1;
+    else if (pos > requested_pos)
+      j = k;
+    else
+      return pCluster;
+  }
+
+  assert(i == j);
+  // assert(Cluster::HasBlockEntries(this, tp.m_pos));
+
+  Cluster* const pCluster = Cluster::Create(this, -1, requested_pos);
+  if (pCluster == NULL)
+    return NULL;
+
+  const ptrdiff_t idx = i - m_clusters;
+
+  if (!PreloadCluster(pCluster, idx)) {
+    delete pCluster;
+    return NULL;
+  }
+  assert(m_clusters);
+  assert(m_clusterPreloadCount > 0);
+  assert(m_clusters[idx] == pCluster);
+
+  return pCluster;
+}
+
+CuePoint::CuePoint(long idx, long long pos)
+    : m_element_start(0),
+      m_element_size(0),
+      m_index(idx),
+      m_timecode(-1 * pos),
+      m_track_positions(NULL),
+      m_track_positions_count(0) {
+  assert(pos > 0);
+}
+
+CuePoint::~CuePoint() { delete[] m_track_positions; }
+
+bool CuePoint::Load(IMkvReader* pReader) {
+  // odbgstream os;
+  // os << "CuePoint::Load(begin): timecode=" << m_timecode << endl;
+
+  if (m_timecode >= 0)  // already loaded
+    return true;
+
+  assert(m_track_positions == NULL);
+  assert(m_track_positions_count == 0);
+
+  long long pos_ = -m_timecode;
+  const long long element_start = pos_;
+
+  long long stop;
+
+  {
+    long len;
+
+    const long long id = ReadID(pReader, pos_, len);
+    if (id != libwebm::kMkvCuePoint)
+      return false;
+
+    pos_ += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos_, len);
+    assert(size >= 0);
+
+    pos_ += len;  // consume Size field
+    // pos_ now points to start of payload
+
+    stop = pos_ + size;
+  }
+
+  const long long element_size = stop - element_start;
+
+  long long pos = pos_;
+
+  // First count number of track positions
+
+  while (pos < stop) {
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if ((id < 0) || (pos + len > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    if ((size < 0) || (pos + len > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume Size field
+    if ((pos + size) > stop) {
+      return false;
+    }
+
+    if (id == libwebm::kMkvCueTime)
+      m_timecode = UnserializeUInt(pReader, pos, size);
+
+    else if (id == libwebm::kMkvCueTrackPositions)
+      ++m_track_positions_count;
+
+    pos += size;  // consume payload
+  }
+
+  if (m_timecode < 0 || m_track_positions_count <= 0) {
+    return false;
+  }
+
+  // os << "CuePoint::Load(cont'd): idpos=" << idpos
+  //   << " timecode=" << m_timecode
+  //   << endl;
+
+  m_track_positions = new (std::nothrow) TrackPosition[m_track_positions_count];
+  if (m_track_positions == NULL)
+    return false;
+
+  // Now parse track positions
+
+  TrackPosition* p = m_track_positions;
+  pos = pos_;
+
+  while (pos < stop) {
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if (id < 0 || (pos + len) > stop)
+      return false;
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);
+    assert((pos + len) <= stop);
+
+    pos += len;  // consume Size field
+    assert((pos + size) <= stop);
+
+    if (id == libwebm::kMkvCueTrackPositions) {
+      TrackPosition& tp = *p++;
+      if (!tp.Parse(pReader, pos, size)) {
+        return false;
+      }
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return false;
+  }
+
+  assert(size_t(p - m_track_positions) == m_track_positions_count);
+
+  m_element_start = element_start;
+  m_element_size = element_size;
+
+  return true;
+}
+
+bool CuePoint::TrackPosition::Parse(IMkvReader* pReader, long long start_,
+                                    long long size_) {
+  const long long stop = start_ + size_;
+  long long pos = start_;
+
+  m_track = -1;
+  m_pos = -1;
+  m_block = 1;  // default
+
+  while (pos < stop) {
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if ((id < 0) || ((pos + len) > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    if ((size < 0) || ((pos + len) > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume Size field
+    if ((pos + size) > stop) {
+      return false;
+    }
+
+    if (id == libwebm::kMkvCueTrack)
+      m_track = UnserializeUInt(pReader, pos, size);
+    else if (id == libwebm::kMkvCueClusterPosition)
+      m_pos = UnserializeUInt(pReader, pos, size);
+    else if (id == libwebm::kMkvCueBlockNumber)
+      m_block = UnserializeUInt(pReader, pos, size);
+
+    pos += size;  // consume payload
+  }
+
+  if ((m_pos < 0) || (m_track <= 0)) {
+    return false;
+  }
+
+  return true;
+}
+
+const CuePoint::TrackPosition* CuePoint::Find(const Track* pTrack) const {
+  assert(pTrack);
+
+  const long long n = pTrack->GetNumber();
+
+  const TrackPosition* i = m_track_positions;
+  const TrackPosition* const j = i + m_track_positions_count;
+
+  while (i != j) {
+    const TrackPosition& p = *i++;
+
+    if (p.m_track == n)
+      return &p;
+  }
+
+  return NULL;  // no matching track number found
+}
+
+long long CuePoint::GetTimeCode() const { return m_timecode; }
+
+long long CuePoint::GetTime(const Segment* pSegment) const {
+  assert(pSegment);
+  assert(m_timecode >= 0);
+
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+  assert(pInfo);
+
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
+
+  const long long time = scale * m_timecode;
+
+  return time;
+}
+
+bool Segment::DoneParsing() const {
+  if (m_size < 0) {
+    long long total, avail;
+
+    const int status = m_pReader->Length(&total, &avail);
+
+    if (status < 0)  // error
+      return true;  // must assume done
+
+    if (total < 0)
+      return false;  // assume live stream
+
+    return (m_pos >= total);
+  }
+
+  const long long stop = m_start + m_size;
+
+  return (m_pos >= stop);
+}
+
+const Cluster* Segment::GetFirst() const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
+
+  Cluster* const pCluster = m_clusters[0];
+  assert(pCluster);
+
+  return pCluster;
+}
+
+const Cluster* Segment::GetLast() const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
+
+  const long idx = m_clusterCount - 1;
+
+  Cluster* const pCluster = m_clusters[idx];
+  assert(pCluster);
+
+  return pCluster;
+}
+
+unsigned long Segment::GetCount() const { return m_clusterCount; }
+
+const Cluster* Segment::GetNext(const Cluster* pCurr) {
+  assert(pCurr);
+  assert(pCurr != &m_eos);
+  assert(m_clusters);
+
+  long idx = pCurr->m_index;
+
+  if (idx >= 0) {
+    assert(m_clusterCount > 0);
+    assert(idx < m_clusterCount);
+    assert(pCurr == m_clusters[idx]);
+
+    ++idx;
+
+    if (idx >= m_clusterCount)
+      return &m_eos;  // caller will LoadCluster as desired
+
+    Cluster* const pNext = m_clusters[idx];
+    assert(pNext);
+    assert(pNext->m_index >= 0);
+    assert(pNext->m_index == idx);
+
+    return pNext;
+  }
+
+  assert(m_clusterPreloadCount > 0);
+
+  long long pos = pCurr->m_element_start;
+
+  assert(m_size >= 0);  // TODO
+  const long long stop = m_start + m_size;  // end of segment
+
+  {
+    long len;
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);
+    assert((pos + len) <= stop);  // TODO
+    if (result != 0)
+      return NULL;
+
+    const long long id = ReadID(m_pReader, pos, len);
+    if (id != libwebm::kMkvCluster)
+      return NULL;
+
+    pos += len;  // consume ID
+
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);  // TODO
+    assert((pos + len) <= stop);  // TODO
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+    assert(size > 0);  // TODO
+    // assert((pCurr->m_size <= 0) || (pCurr->m_size == size));
+
+    pos += len;  // consume length of size of element
+    assert((pos + size) <= stop);  // TODO
+
+    // Pos now points to start of payload
+
+    pos += size;  // consume payload
+  }
+
+  long long off_next = 0;
+
+  while (pos < stop) {
+    long len;
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);
+    assert((pos + len) <= stop);  // TODO
+    if (result != 0)
+      return NULL;
+
+    const long long idpos = pos;  // pos of next (potential) cluster
+
+    const long long id = ReadID(m_pReader, idpos, len);
+    if (id < 0)
+      return NULL;
+
+    pos += len;  // consume ID
+
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);  // TODO
+    assert((pos + len) <= stop);  // TODO
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+    assert(size >= 0);  // TODO
+
+    pos += len;  // consume length of size of element
+    assert((pos + size) <= stop);  // TODO
+
+    // Pos now points to start of payload
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvCluster) {
+      const long long off_next_ = idpos - m_start;
+
+      long long pos_;
+      long len_;
+
+      const long status = Cluster::HasBlockEntries(this, off_next_, pos_, len_);
+
+      assert(status >= 0);
+
+      if (status > 0) {
+        off_next = off_next_;
+        break;
+      }
+    }
+
+    pos += size;  // consume payload
+  }
+
+  if (off_next <= 0)
+    return 0;
+
+  Cluster** const ii = m_clusters + m_clusterCount;
+  Cluster** i = ii;
+
+  Cluster** const jj = ii + m_clusterPreloadCount;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) < pos_next
+    //[i, j) ?
+    //[j, jj)  > pos_next
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    Cluster* const pNext = *k;
+    assert(pNext);
+    assert(pNext->m_index < 0);
+
+    // const long long pos_ = pNext->m_pos;
+    // assert(pos_);
+    // pos = pos_ * ((pos_ < 0) ? -1 : 1);
+
+    pos = pNext->GetPosition();
+
+    if (pos < off_next)
+      i = k + 1;
+    else if (pos > off_next)
+      j = k;
+    else
+      return pNext;
+  }
+
+  assert(i == j);
+
+  Cluster* const pNext = Cluster::Create(this, -1, off_next);
+  if (pNext == NULL)
+    return NULL;
+
+  const ptrdiff_t idx_next = i - m_clusters;  // insertion position
+
+  if (!PreloadCluster(pNext, idx_next)) {
+    delete pNext;
+    return NULL;
+  }
+  assert(m_clusters);
+  assert(idx_next < m_clusterSize);
+  assert(m_clusters[idx_next] == pNext);
+
+  return pNext;
+}
+
+long Segment::ParseNext(const Cluster* pCurr, const Cluster*& pResult,
+                        long long& pos, long& len) {
+  assert(pCurr);
+  assert(!pCurr->EOS());
+  assert(m_clusters);
+
+  pResult = 0;
+
+  if (pCurr->m_index >= 0) {  // loaded (not merely preloaded)
+    assert(m_clusters[pCurr->m_index] == pCurr);
+
+    const long next_idx = pCurr->m_index + 1;
+
+    if (next_idx < m_clusterCount) {
+      pResult = m_clusters[next_idx];
+      return 0;  // success
+    }
+
+    // curr cluster is last among loaded
+
+    const long result = LoadCluster(pos, len);
+
+    if (result < 0)  // error or underflow
+      return result;
+
+    if (result > 0)  // no more clusters
+    {
+      // pResult = &m_eos;
+      return 1;
+    }
+
+    pResult = GetLast();
+    return 0;  // success
+  }
+
+  assert(m_pos > 0);
+
+  long long total, avail;
+
+  long status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  // interrogate curr cluster
+
+  pos = pCurr->m_element_start;
+
+  if (pCurr->m_element_size >= 0)
+    pos += pCurr->m_element_size;
+  else {
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadUInt(m_pReader, pos, len);
+
+    if (id != libwebm::kMkvCluster)
+      return -1;
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume size field
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)  // TODO: should never happen
+      return E_FILE_FORMAT_INVALID;  // TODO: resolve this
+
+    // assert((pCurr->m_size <= 0) || (pCurr->m_size == size));
+
+    if ((segment_stop >= 0) && ((pos + size) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // Pos now points to start of payload
+
+    pos += size;  // consume payload (that is, the current cluster)
+    if (segment_stop >= 0 && pos > segment_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    // By consuming the payload, we are assuming that the curr
+    // cluster isn't interesting.  That is, we don't bother checking
+    // whether the payload of the curr cluster is less than what
+    // happens to be available (obtained via IMkvReader::Length).
+    // Presumably the caller has already dispensed with the current
+    // cluster, and really does want the next cluster.
+  }
+
+  // pos now points to just beyond the last fully-loaded cluster
+
+  for (;;) {
+    const long status = DoParseNext(pResult, pos, len);
+
+    if (status <= 1)
+      return status;
+  }
+}
+
+long Segment::DoParseNext(const Cluster*& pResult, long long& pos, long& len) {
+  long long total, avail;
+
+  long status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  // Parse next cluster.  This is strictly a parsing activity.
+  // Creation of a new cluster object happens later, after the
+  // parsing is done.
+
+  long long off_next = 0;
+  long long cluster_size = -1;
+
+  for (;;) {
+    if ((total >= 0) && (pos >= total))
+      return 1;  // EOF
+
+    if ((segment_stop >= 0) && (pos >= segment_stop))
+      return 1;  // EOF
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long idpos = pos;  // absolute
+    const long long idoff = pos - m_start;  // relative
+
+    const long long id = ReadID(m_pReader, idpos, len);  // absolute
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id == 0)  // weird
+      return -1;  // generic error
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume length of size of element
+
+    // Pos now points to start of payload
+
+    if (size == 0)  // weird
+      continue;
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if ((segment_stop >= 0) && (size != unknown_size) &&
+        ((pos + size) > segment_stop)) {
+      return E_FILE_FORMAT_INVALID;
+    }
+
+    if (id == libwebm::kMkvCues) {
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
+
+      const long long element_stop = pos + size;
+
+      if ((segment_stop >= 0) && (element_stop > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      const long long element_start = idpos;
+      const long long element_size = element_stop - element_start;
+
+      if (m_pCues == NULL) {
+        m_pCues = new (std::nothrow)
+            Cues(this, pos, size, element_start, element_size);
+        if (m_pCues == NULL)
+          return false;
+      }
+
+      pos += size;  // consume payload
+      if (segment_stop >= 0 && pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      continue;
+    }
+
+    if (id != libwebm::kMkvCluster) {  // not a Cluster ID
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
+
+      pos += size;  // consume payload
+      if (segment_stop >= 0 && pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      continue;
+    }
+
+    // We have a cluster.
+    off_next = idoff;
+
+    if (size != unknown_size)
+      cluster_size = size;
+
+    break;
+  }
+
+  assert(off_next > 0);  // have cluster
+
+  // We have parsed the next cluster.
+  // We have not created a cluster object yet.  What we need
+  // to do now is determine whether it has already be preloaded
+  //(in which case, an object for this cluster has already been
+  // created), and if not, create a new cluster object.
+
+  Cluster** const ii = m_clusters + m_clusterCount;
+  Cluster** i = ii;
+
+  Cluster** const jj = ii + m_clusterPreloadCount;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) < pos_next
+    //[i, j) ?
+    //[j, jj)  > pos_next
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    const Cluster* const pNext = *k;
+    assert(pNext);
+    assert(pNext->m_index < 0);
+
+    pos = pNext->GetPosition();
+    assert(pos >= 0);
+
+    if (pos < off_next)
+      i = k + 1;
+    else if (pos > off_next)
+      j = k;
+    else {
+      pResult = pNext;
+      return 0;  // success
+    }
+  }
+
+  assert(i == j);
+
+  long long pos_;
+  long len_;
+
+  status = Cluster::HasBlockEntries(this, off_next, pos_, len_);
+
+  if (status < 0) {  // error or underflow
+    pos = pos_;
+    len = len_;
+
+    return status;
+  }
+
+  if (status > 0) {  // means "found at least one block entry"
+    Cluster* const pNext = Cluster::Create(this,
+                                           -1,  // preloaded
+                                           off_next);
+    if (pNext == NULL)
+      return -1;
+
+    const ptrdiff_t idx_next = i - m_clusters;  // insertion position
+
+    if (!PreloadCluster(pNext, idx_next)) {
+      delete pNext;
+      return -1;
+    }
+    assert(m_clusters);
+    assert(idx_next < m_clusterSize);
+    assert(m_clusters[idx_next] == pNext);
+
+    pResult = pNext;
+    return 0;  // success
+  }
+
+  // status == 0 means "no block entries found"
+
+  if (cluster_size < 0) {  // unknown size
+    const long long payload_pos = pos;  // absolute pos of cluster payload
+
+    for (;;) {  // determine cluster size
+      if ((total >= 0) && (pos >= total))
+        break;
+
+      if ((segment_stop >= 0) && (pos >= segment_stop))
+        break;  // no more clusters
+
+      // Read ID
+
+      if ((pos + 1) > avail) {
+        len = 1;
+        return E_BUFFER_NOT_FULL;
+      }
+
+      long long result = GetUIntLength(m_pReader, pos, len);
+
+      if (result < 0)  // error
+        return static_cast<long>(result);
+
+      if (result > 0)  // weird
+        return E_BUFFER_NOT_FULL;
+
+      if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      if ((pos + len) > avail)
+        return E_BUFFER_NOT_FULL;
+
+      const long long idpos = pos;
+      const long long id = ReadID(m_pReader, idpos, len);
+
+      if (id < 0)  // error (or underflow)
+        return static_cast<long>(id);
+
+      // This is the distinguished set of ID's we use to determine
+      // that we have exhausted the sub-element's inside the cluster
+      // whose ID we parsed earlier.
+
+      if (id == libwebm::kMkvCluster || id == libwebm::kMkvCues)
+        break;
+
+      pos += len;  // consume ID (of sub-element)
+
+      // Read Size
+
+      if ((pos + 1) > avail) {
+        len = 1;
+        return E_BUFFER_NOT_FULL;
+      }
+
+      result = GetUIntLength(m_pReader, pos, len);
+
+      if (result < 0)  // error
+        return static_cast<long>(result);
+
+      if (result > 0)  // weird
+        return E_BUFFER_NOT_FULL;
+
+      if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      if ((pos + len) > avail)
+        return E_BUFFER_NOT_FULL;
+
+      const long long size = ReadUInt(m_pReader, pos, len);
+
+      if (size < 0)  // error
+        return static_cast<long>(size);
+
+      pos += len;  // consume size field of element
+
+      // pos now points to start of sub-element's payload
+
+      if (size == 0)  // weird
+        continue;
+
+      const long long unknown_size = (1LL << (7 * len)) - 1;
+
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;  // not allowed for sub-elements
+
+      if ((segment_stop >= 0) && ((pos + size) > segment_stop))  // weird
+        return E_FILE_FORMAT_INVALID;
+
+      pos += size;  // consume payload of sub-element
+      if (segment_stop >= 0 && pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+    }  // determine cluster size
+
+    cluster_size = pos - payload_pos;
+    assert(cluster_size >= 0);  // TODO: handle cluster_size = 0
+
+    pos = payload_pos;  // reset and re-parse original cluster
+  }
+
+  pos += cluster_size;  // consume payload
+  if (segment_stop >= 0 && pos > segment_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 2;  // try to find a cluster that follows next
+}
+
+const Cluster* Segment::FindCluster(long long time_ns) const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
+
+  {
+    Cluster* const pCluster = m_clusters[0];
+    assert(pCluster);
+    assert(pCluster->m_index == 0);
+
+    if (time_ns <= pCluster->GetTime())
+      return pCluster;
+  }
+
+  // Binary search of cluster array
+
+  long i = 0;
+  long j = m_clusterCount;
+
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) <= time_ns
+    //[i, j) ?
+    //[j, m_clusterCount)  > time_ns
+
+    const long k = i + (j - i) / 2;
+    assert(k < m_clusterCount);
+
+    Cluster* const pCluster = m_clusters[k];
+    assert(pCluster);
+    assert(pCluster->m_index == k);
+
+    const long long t = pCluster->GetTime();
+
+    if (t <= time_ns)
+      i = k + 1;
+    else
+      j = k;
+
+    assert(i <= j);
+  }
+
+  assert(i == j);
+  assert(i > 0);
+  assert(i <= m_clusterCount);
+
+  const long k = i - 1;
+
+  Cluster* const pCluster = m_clusters[k];
+  assert(pCluster);
+  assert(pCluster->m_index == k);
+  assert(pCluster->GetTime() <= time_ns);
+
+  return pCluster;
+}
+
+const Tracks* Segment::GetTracks() const { return m_pTracks; }
+const SegmentInfo* Segment::GetInfo() const { return m_pInfo; }
+const Cues* Segment::GetCues() const { return m_pCues; }
+const Chapters* Segment::GetChapters() const { return m_pChapters; }
+const Tags* Segment::GetTags() const { return m_pTags; }
+const SeekHead* Segment::GetSeekHead() const { return m_pSeekHead; }
+
+long long Segment::GetDuration() const {
+  assert(m_pInfo);
+  return m_pInfo->GetDuration();
+}
+
+Chapters::Chapters(Segment* pSegment, long long payload_start,
+                   long long payload_size, long long element_start,
+                   long long element_size)
+    : m_pSegment(pSegment),
+      m_start(payload_start),
+      m_size(payload_size),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_editions(NULL),
+      m_editions_size(0),
+      m_editions_count(0) {}
+
+Chapters::~Chapters() {
+  while (m_editions_count > 0) {
+    Edition& e = m_editions[--m_editions_count];
+    e.Clear();
+  }
+  delete[] m_editions;
+}
+
+long Chapters::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;  // payload start
+  const long long stop = pos + m_size;  // payload stop
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvEditionEntry) {
+      status = ParseEdition(pos, size);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+int Chapters::GetEditionCount() const { return m_editions_count; }
+
+const Chapters::Edition* Chapters::GetEdition(int idx) const {
+  if (idx < 0)
+    return NULL;
+
+  if (idx >= m_editions_count)
+    return NULL;
+
+  return m_editions + idx;
+}
+
+bool Chapters::ExpandEditionsArray() {
+  if (m_editions_size > m_editions_count)
+    return true;  // nothing else to do
+
+  const int size = (m_editions_size == 0) ? 1 : 2 * m_editions_size;
+
+  Edition* const editions = new (std::nothrow) Edition[size];
+
+  if (editions == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_editions_count; ++idx) {
+    m_editions[idx].ShallowCopy(editions[idx]);
+  }
+
+  delete[] m_editions;
+  m_editions = editions;
+
+  m_editions_size = size;
+  return true;
+}
+
+long Chapters::ParseEdition(long long pos, long long size) {
+  if (!ExpandEditionsArray())
+    return -1;
+
+  Edition& e = m_editions[m_editions_count++];
+  e.Init();
+
+  return e.Parse(m_pSegment->m_pReader, pos, size);
+}
+
+Chapters::Edition::Edition() {}
+
+Chapters::Edition::~Edition() {}
+
+int Chapters::Edition::GetAtomCount() const { return m_atoms_count; }
+
+const Chapters::Atom* Chapters::Edition::GetAtom(int index) const {
+  if (index < 0)
+    return NULL;
+
+  if (index >= m_atoms_count)
+    return NULL;
+
+  return m_atoms + index;
+}
+
+void Chapters::Edition::Init() {
+  m_atoms = NULL;
+  m_atoms_size = 0;
+  m_atoms_count = 0;
+}
+
+void Chapters::Edition::ShallowCopy(Edition& rhs) const {
+  rhs.m_atoms = m_atoms;
+  rhs.m_atoms_size = m_atoms_size;
+  rhs.m_atoms_count = m_atoms_count;
+}
+
+void Chapters::Edition::Clear() {
+  while (m_atoms_count > 0) {
+    Atom& a = m_atoms[--m_atoms_count];
+    a.Clear();
+  }
+
+  delete[] m_atoms;
+  m_atoms = NULL;
+
+  m_atoms_size = 0;
+}
+
+long Chapters::Edition::Parse(IMkvReader* pReader, long long pos,
+                              long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)
+      continue;
+
+    if (id == libwebm::kMkvChapterAtom) {
+      status = ParseAtom(pReader, pos, size);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long Chapters::Edition::ParseAtom(IMkvReader* pReader, long long pos,
+                                  long long size) {
+  if (!ExpandAtomsArray())
+    return -1;
+
+  Atom& a = m_atoms[m_atoms_count++];
+  a.Init();
+
+  return a.Parse(pReader, pos, size);
+}
+
+bool Chapters::Edition::ExpandAtomsArray() {
+  if (m_atoms_size > m_atoms_count)
+    return true;  // nothing else to do
+
+  const int size = (m_atoms_size == 0) ? 1 : 2 * m_atoms_size;
+
+  Atom* const atoms = new (std::nothrow) Atom[size];
+
+  if (atoms == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_atoms_count; ++idx) {
+    m_atoms[idx].ShallowCopy(atoms[idx]);
+  }
+
+  delete[] m_atoms;
+  m_atoms = atoms;
+
+  m_atoms_size = size;
+  return true;
+}
+
+Chapters::Atom::Atom() {}
+
+Chapters::Atom::~Atom() {}
+
+unsigned long long Chapters::Atom::GetUID() const { return m_uid; }
+
+const char* Chapters::Atom::GetStringUID() const { return m_string_uid; }
+
+long long Chapters::Atom::GetStartTimecode() const { return m_start_timecode; }
+
+long long Chapters::Atom::GetStopTimecode() const { return m_stop_timecode; }
+
+long long Chapters::Atom::GetStartTime(const Chapters* pChapters) const {
+  return GetTime(pChapters, m_start_timecode);
+}
+
+long long Chapters::Atom::GetStopTime(const Chapters* pChapters) const {
+  return GetTime(pChapters, m_stop_timecode);
+}
+
+int Chapters::Atom::GetDisplayCount() const { return m_displays_count; }
+
+const Chapters::Display* Chapters::Atom::GetDisplay(int index) const {
+  if (index < 0)
+    return NULL;
+
+  if (index >= m_displays_count)
+    return NULL;
+
+  return m_displays + index;
+}
+
+void Chapters::Atom::Init() {
+  m_string_uid = NULL;
+  m_uid = 0;
+  m_start_timecode = -1;
+  m_stop_timecode = -1;
+
+  m_displays = NULL;
+  m_displays_size = 0;
+  m_displays_count = 0;
+}
+
+void Chapters::Atom::ShallowCopy(Atom& rhs) const {
+  rhs.m_string_uid = m_string_uid;
+  rhs.m_uid = m_uid;
+  rhs.m_start_timecode = m_start_timecode;
+  rhs.m_stop_timecode = m_stop_timecode;
+
+  rhs.m_displays = m_displays;
+  rhs.m_displays_size = m_displays_size;
+  rhs.m_displays_count = m_displays_count;
+}
+
+void Chapters::Atom::Clear() {
+  delete[] m_string_uid;
+  m_string_uid = NULL;
+
+  while (m_displays_count > 0) {
+    Display& d = m_displays[--m_displays_count];
+    d.Clear();
+  }
+
+  delete[] m_displays;
+  m_displays = NULL;
+
+  m_displays_size = 0;
+}
+
+long Chapters::Atom::Parse(IMkvReader* pReader, long long pos, long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // 0 length payload, skip.
+      continue;
+
+    if (id == libwebm::kMkvChapterDisplay) {
+      status = ParseDisplay(pReader, pos, size);
+
+      if (status < 0)  // error
+        return status;
+    } else if (id == libwebm::kMkvChapterStringUID) {
+      status = UnserializeString(pReader, pos, size, m_string_uid);
+
+      if (status < 0)  // error
+        return status;
+    } else if (id == libwebm::kMkvChapterUID) {
+      long long val;
+      status = UnserializeInt(pReader, pos, size, val);
+
+      if (status < 0)  // error
+        return status;
+
+      m_uid = static_cast<unsigned long long>(val);
+    } else if (id == libwebm::kMkvChapterTimeStart) {
+      const long long val = UnserializeUInt(pReader, pos, size);
+
+      if (val < 0)  // error
+        return static_cast<long>(val);
+
+      m_start_timecode = val;
+    } else if (id == libwebm::kMkvChapterTimeEnd) {
+      const long long val = UnserializeUInt(pReader, pos, size);
+
+      if (val < 0)  // error
+        return static_cast<long>(val);
+
+      m_stop_timecode = val;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long long Chapters::Atom::GetTime(const Chapters* pChapters,
+                                  long long timecode) {
+  if (pChapters == NULL)
+    return -1;
+
+  Segment* const pSegment = pChapters->m_pSegment;
+
+  if (pSegment == NULL)  // weird
+    return -1;
+
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+
+  if (pInfo == NULL)
+    return -1;
+
+  const long long timecode_scale = pInfo->GetTimeCodeScale();
+
+  if (timecode_scale < 1)  // weird
+    return -1;
+
+  if (timecode < 0)
+    return -1;
+
+  const long long result = timecode_scale * timecode;
+
+  return result;
+}
+
+long Chapters::Atom::ParseDisplay(IMkvReader* pReader, long long pos,
+                                  long long size) {
+  if (!ExpandDisplaysArray())
+    return -1;
+
+  Display& d = m_displays[m_displays_count++];
+  d.Init();
+
+  return d.Parse(pReader, pos, size);
+}
+
+bool Chapters::Atom::ExpandDisplaysArray() {
+  if (m_displays_size > m_displays_count)
+    return true;  // nothing else to do
+
+  const int size = (m_displays_size == 0) ? 1 : 2 * m_displays_size;
+
+  Display* const displays = new (std::nothrow) Display[size];
+
+  if (displays == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_displays_count; ++idx) {
+    m_displays[idx].ShallowCopy(displays[idx]);
+  }
+
+  delete[] m_displays;
+  m_displays = displays;
+
+  m_displays_size = size;
+  return true;
+}
+
+Chapters::Display::Display() {}
+
+Chapters::Display::~Display() {}
+
+const char* Chapters::Display::GetString() const { return m_string; }
+
+const char* Chapters::Display::GetLanguage() const { return m_language; }
+
+const char* Chapters::Display::GetCountry() const { return m_country; }
+
+void Chapters::Display::Init() {
+  m_string = NULL;
+  m_language = NULL;
+  m_country = NULL;
+}
+
+void Chapters::Display::ShallowCopy(Display& rhs) const {
+  rhs.m_string = m_string;
+  rhs.m_language = m_language;
+  rhs.m_country = m_country;
+}
+
+void Chapters::Display::Clear() {
+  delete[] m_string;
+  m_string = NULL;
+
+  delete[] m_language;
+  m_language = NULL;
+
+  delete[] m_country;
+  m_country = NULL;
+}
+
+long Chapters::Display::Parse(IMkvReader* pReader, long long pos,
+                              long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // No payload.
+      continue;
+
+    if (id == libwebm::kMkvChapString) {
+      status = UnserializeString(pReader, pos, size, m_string);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvChapLanguage) {
+      status = UnserializeString(pReader, pos, size, m_language);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvChapCountry) {
+      status = UnserializeString(pReader, pos, size, m_country);
+
+      if (status)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+Tags::Tags(Segment* pSegment, long long payload_start, long long payload_size,
+           long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(payload_start),
+      m_size(payload_size),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_tags(NULL),
+      m_tags_size(0),
+      m_tags_count(0) {}
+
+Tags::~Tags() {
+  while (m_tags_count > 0) {
+    Tag& t = m_tags[--m_tags_count];
+    t.Clear();
+  }
+  delete[] m_tags;
+}
+
+long Tags::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;  // payload start
+  const long long stop = pos + m_size;  // payload stop
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)
+      return status;
+
+    if (size == 0)  // 0 length tag, read another
+      continue;
+
+    if (id == libwebm::kMkvTag) {
+      status = ParseTag(pos, size);
+
+      if (status < 0)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+int Tags::GetTagCount() const { return m_tags_count; }
+
+const Tags::Tag* Tags::GetTag(int idx) const {
+  if (idx < 0)
+    return NULL;
+
+  if (idx >= m_tags_count)
+    return NULL;
+
+  return m_tags + idx;
+}
+
+bool Tags::ExpandTagsArray() {
+  if (m_tags_size > m_tags_count)
+    return true;  // nothing else to do
+
+  const int size = (m_tags_size == 0) ? 1 : 2 * m_tags_size;
+
+  Tag* const tags = new (std::nothrow) Tag[size];
+
+  if (tags == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_tags_count; ++idx) {
+    m_tags[idx].ShallowCopy(tags[idx]);
+  }
+
+  delete[] m_tags;
+  m_tags = tags;
+
+  m_tags_size = size;
+  return true;
+}
+
+long Tags::ParseTag(long long pos, long long size) {
+  if (!ExpandTagsArray())
+    return -1;
+
+  Tag& t = m_tags[m_tags_count++];
+  t.Init();
+
+  return t.Parse(m_pSegment->m_pReader, pos, size);
+}
+
+Tags::Tag::Tag() {}
+
+Tags::Tag::~Tag() {}
+
+int Tags::Tag::GetSimpleTagCount() const { return m_simple_tags_count; }
+
+const Tags::SimpleTag* Tags::Tag::GetSimpleTag(int index) const {
+  if (index < 0)
+    return NULL;
+
+  if (index >= m_simple_tags_count)
+    return NULL;
+
+  return m_simple_tags + index;
+}
+
+void Tags::Tag::Init() {
+  m_simple_tags = NULL;
+  m_simple_tags_size = 0;
+  m_simple_tags_count = 0;
+}
+
+void Tags::Tag::ShallowCopy(Tag& rhs) const {
+  rhs.m_simple_tags = m_simple_tags;
+  rhs.m_simple_tags_size = m_simple_tags_size;
+  rhs.m_simple_tags_count = m_simple_tags_count;
+}
+
+void Tags::Tag::Clear() {
+  while (m_simple_tags_count > 0) {
+    SimpleTag& d = m_simple_tags[--m_simple_tags_count];
+    d.Clear();
+  }
+
+  delete[] m_simple_tags;
+  m_simple_tags = NULL;
+
+  m_simple_tags_size = 0;
+}
+
+long Tags::Tag::Parse(IMkvReader* pReader, long long pos, long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)
+      return status;
+
+    if (size == 0)  // 0 length tag, read another
+      continue;
+
+    if (id == libwebm::kMkvSimpleTag) {
+      status = ParseSimpleTag(pReader, pos, size);
+
+      if (status < 0)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long Tags::Tag::ParseSimpleTag(IMkvReader* pReader, long long pos,
+                               long long size) {
+  if (!ExpandSimpleTagsArray())
+    return -1;
+
+  SimpleTag& st = m_simple_tags[m_simple_tags_count++];
+  st.Init();
+
+  return st.Parse(pReader, pos, size);
+}
+
+bool Tags::Tag::ExpandSimpleTagsArray() {
+  if (m_simple_tags_size > m_simple_tags_count)
+    return true;  // nothing else to do
+
+  const int size = (m_simple_tags_size == 0) ? 1 : 2 * m_simple_tags_size;
+
+  SimpleTag* const displays = new (std::nothrow) SimpleTag[size];
+
+  if (displays == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_simple_tags_count; ++idx) {
+    m_simple_tags[idx].ShallowCopy(displays[idx]);
+  }
+
+  delete[] m_simple_tags;
+  m_simple_tags = displays;
+
+  m_simple_tags_size = size;
+  return true;
+}
+
+Tags::SimpleTag::SimpleTag() {}
+
+Tags::SimpleTag::~SimpleTag() {}
+
+const char* Tags::SimpleTag::GetTagName() const { return m_tag_name; }
+
+const char* Tags::SimpleTag::GetTagString() const { return m_tag_string; }
+
+void Tags::SimpleTag::Init() {
+  m_tag_name = NULL;
+  m_tag_string = NULL;
+}
+
+void Tags::SimpleTag::ShallowCopy(SimpleTag& rhs) const {
+  rhs.m_tag_name = m_tag_name;
+  rhs.m_tag_string = m_tag_string;
+}
+
+void Tags::SimpleTag::Clear() {
+  delete[] m_tag_name;
+  m_tag_name = NULL;
+
+  delete[] m_tag_string;
+  m_tag_string = NULL;
+}
+
+long Tags::SimpleTag::Parse(IMkvReader* pReader, long long pos,
+                            long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvTagName) {
+      status = UnserializeString(pReader, pos, size, m_tag_name);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvTagString) {
+      status = UnserializeString(pReader, pos, size, m_tag_string);
+
+      if (status)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+SegmentInfo::SegmentInfo(Segment* pSegment, long long start, long long size_,
+                         long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_pMuxingAppAsUTF8(NULL),
+      m_pWritingAppAsUTF8(NULL),
+      m_pTitleAsUTF8(NULL) {}
+
+SegmentInfo::~SegmentInfo() {
+  delete[] m_pMuxingAppAsUTF8;
+  m_pMuxingAppAsUTF8 = NULL;
+
+  delete[] m_pWritingAppAsUTF8;
+  m_pWritingAppAsUTF8 = NULL;
+
+  delete[] m_pTitleAsUTF8;
+  m_pTitleAsUTF8 = NULL;
+}
+
+long SegmentInfo::Parse() {
+  assert(m_pMuxingAppAsUTF8 == NULL);
+  assert(m_pWritingAppAsUTF8 == NULL);
+  assert(m_pTitleAsUTF8 == NULL);
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
+
+  m_timecodeScale = 1000000;
+  m_duration = -1;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvTimecodeScale) {
+      m_timecodeScale = UnserializeUInt(pReader, pos, size);
+
+      if (m_timecodeScale <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDuration) {
+      const long status = UnserializeFloat(pReader, pos, size, m_duration);
+
+      if (status < 0)
+        return status;
+
+      if (m_duration < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvMuxingApp) {
+      const long status =
+          UnserializeString(pReader, pos, size, m_pMuxingAppAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvWritingApp) {
+      const long status =
+          UnserializeString(pReader, pos, size, m_pWritingAppAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvTitle) {
+      const long status = UnserializeString(pReader, pos, size, m_pTitleAsUTF8);
+
+      if (status)
+        return status;
+    }
+
+    pos += size;
+
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  const double rollover_check = m_duration * m_timecodeScale;
+  if (rollover_check > LLONG_MAX)
+    return E_FILE_FORMAT_INVALID;
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+long long SegmentInfo::GetTimeCodeScale() const { return m_timecodeScale; }
+
+long long SegmentInfo::GetDuration() const {
+  if (m_duration < 0)
+    return -1;
+
+  assert(m_timecodeScale >= 1);
+
+  const double dd = double(m_duration) * double(m_timecodeScale);
+  const long long d = static_cast<long long>(dd);
+
+  return d;
+}
+
+const char* SegmentInfo::GetMuxingAppAsUTF8() const {
+  return m_pMuxingAppAsUTF8;
+}
+
+const char* SegmentInfo::GetWritingAppAsUTF8() const {
+  return m_pWritingAppAsUTF8;
+}
+
+const char* SegmentInfo::GetTitleAsUTF8() const { return m_pTitleAsUTF8; }
+
+///////////////////////////////////////////////////////////////
+// ContentEncoding element
+ContentEncoding::ContentCompression::ContentCompression()
+    : algo(0), settings(NULL), settings_len(0) {}
+
+ContentEncoding::ContentCompression::~ContentCompression() {
+  delete[] settings;
+}
+
+ContentEncoding::ContentEncryption::ContentEncryption()
+    : algo(0),
+      key_id(NULL),
+      key_id_len(0),
+      signature(NULL),
+      signature_len(0),
+      sig_key_id(NULL),
+      sig_key_id_len(0),
+      sig_algo(0),
+      sig_hash_algo(0) {}
+
+ContentEncoding::ContentEncryption::~ContentEncryption() {
+  delete[] key_id;
+  delete[] signature;
+  delete[] sig_key_id;
+}
+
+ContentEncoding::ContentEncoding()
+    : compression_entries_(NULL),
+      compression_entries_end_(NULL),
+      encryption_entries_(NULL),
+      encryption_entries_end_(NULL),
+      encoding_order_(0),
+      encoding_scope_(1),
+      encoding_type_(0) {}
+
+ContentEncoding::~ContentEncoding() {
+  ContentCompression** comp_i = compression_entries_;
+  ContentCompression** const comp_j = compression_entries_end_;
+
+  while (comp_i != comp_j) {
+    ContentCompression* const comp = *comp_i++;
+    delete comp;
+  }
+
+  delete[] compression_entries_;
+
+  ContentEncryption** enc_i = encryption_entries_;
+  ContentEncryption** const enc_j = encryption_entries_end_;
+
+  while (enc_i != enc_j) {
+    ContentEncryption* const enc = *enc_i++;
+    delete enc;
+  }
+
+  delete[] encryption_entries_;
+}
+
+const ContentEncoding::ContentCompression*
+    ContentEncoding::GetCompressionByIndex(unsigned long idx) const {
+  const ptrdiff_t count = compression_entries_end_ - compression_entries_;
+  assert(count >= 0);
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return compression_entries_[idx];
+}
+
+unsigned long ContentEncoding::GetCompressionCount() const {
+  const ptrdiff_t count = compression_entries_end_ - compression_entries_;
+  assert(count >= 0);
+
+  return static_cast<unsigned long>(count);
+}
+
+const ContentEncoding::ContentEncryption* ContentEncoding::GetEncryptionByIndex(
+    unsigned long idx) const {
+  const ptrdiff_t count = encryption_entries_end_ - encryption_entries_;
+  assert(count >= 0);
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return encryption_entries_[idx];
+}
+
+unsigned long ContentEncoding::GetEncryptionCount() const {
+  const ptrdiff_t count = encryption_entries_end_ - encryption_entries_;
+  assert(count >= 0);
+
+  return static_cast<unsigned long>(count);
+}
+
+long ContentEncoding::ParseContentEncAESSettingsEntry(
+    long long start, long long size, IMkvReader* pReader,
+    ContentEncAESSettings* aes) {
+  assert(pReader);
+  assert(aes);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvAESSettingsCipherMode) {
+      aes->cipher_mode = UnserializeUInt(pReader, pos, size);
+      if (aes->cipher_mode != 1)
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  return 0;
+}
+
+long ContentEncoding::ParseContentEncodingEntry(long long start, long long size,
+                                                IMkvReader* pReader) {
+  assert(pReader);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  // Count ContentCompression and ContentEncryption elements.
+  int compression_count = 0;
+  int encryption_count = 0;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentCompression)
+      ++compression_count;
+
+    if (id == libwebm::kMkvContentEncryption)
+      ++encryption_count;
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (compression_count <= 0 && encryption_count <= 0)
+    return -1;
+
+  if (compression_count > 0) {
+    compression_entries_ =
+        new (std::nothrow) ContentCompression*[compression_count];
+    if (!compression_entries_)
+      return -1;
+    compression_entries_end_ = compression_entries_;
+  }
+
+  if (encryption_count > 0) {
+    encryption_entries_ =
+        new (std::nothrow) ContentEncryption*[encryption_count];
+    if (!encryption_entries_) {
+      delete[] compression_entries_;
+      return -1;
+    }
+    encryption_entries_end_ = encryption_entries_;
+  }
+
+  pos = start;
+  while (pos < stop) {
+    long long id, size;
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentEncodingOrder) {
+      encoding_order_ = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentEncodingScope) {
+      encoding_scope_ = UnserializeUInt(pReader, pos, size);
+      if (encoding_scope_ < 1)
+        return -1;
+    } else if (id == libwebm::kMkvContentEncodingType) {
+      encoding_type_ = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentCompression) {
+      ContentCompression* const compression =
+          new (std::nothrow) ContentCompression();
+      if (!compression)
+        return -1;
+
+      status = ParseCompressionEntry(pos, size, pReader, compression);
+      if (status) {
+        delete compression;
+        return status;
+      }
+      *compression_entries_end_++ = compression;
+    } else if (id == libwebm::kMkvContentEncryption) {
+      ContentEncryption* const encryption =
+          new (std::nothrow) ContentEncryption();
+      if (!encryption)
+        return -1;
+
+      status = ParseEncryptionEntry(pos, size, pReader, encryption);
+      if (status) {
+        delete encryption;
+        return status;
+      }
+      *encryption_entries_end_++ = encryption;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long ContentEncoding::ParseCompressionEntry(long long start, long long size,
+                                            IMkvReader* pReader,
+                                            ContentCompression* compression) {
+  assert(pReader);
+  assert(compression);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  bool valid = false;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentCompAlgo) {
+      long long algo = UnserializeUInt(pReader, pos, size);
+      if (algo < 0)
+        return E_FILE_FORMAT_INVALID;
+      compression->algo = algo;
+      valid = true;
+    } else if (id == libwebm::kMkvContentCompSettings) {
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      compression->settings = buf;
+      compression->settings_len = buflen;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  // ContentCompAlgo is mandatory
+  if (!valid)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+long ContentEncoding::ParseEncryptionEntry(long long start, long long size,
+                                           IMkvReader* pReader,
+                                           ContentEncryption* encryption) {
+  assert(pReader);
+  assert(encryption);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentEncAlgo) {
+      encryption->algo = UnserializeUInt(pReader, pos, size);
+      if (encryption->algo != 5)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvContentEncKeyID) {
+      delete[] encryption->key_id;
+      encryption->key_id = NULL;
+      encryption->key_id_len = 0;
+
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      encryption->key_id = buf;
+      encryption->key_id_len = buflen;
+    } else if (id == libwebm::kMkvContentSignature) {
+      delete[] encryption->signature;
+      encryption->signature = NULL;
+      encryption->signature_len = 0;
+
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      encryption->signature = buf;
+      encryption->signature_len = buflen;
+    } else if (id == libwebm::kMkvContentSigKeyID) {
+      delete[] encryption->sig_key_id;
+      encryption->sig_key_id = NULL;
+      encryption->sig_key_id_len = 0;
+
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      encryption->sig_key_id = buf;
+      encryption->sig_key_id_len = buflen;
+    } else if (id == libwebm::kMkvContentSigAlgo) {
+      encryption->sig_algo = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentSigHashAlgo) {
+      encryption->sig_hash_algo = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentEncAESSettings) {
+      const long status = ParseContentEncAESSettingsEntry(
+          pos, size, pReader, &encryption->aes_settings);
+      if (status)
+        return status;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  return 0;
+}
+
+Track::Track(Segment* pSegment, long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      content_encoding_entries_(NULL),
+      content_encoding_entries_end_(NULL) {}
+
+Track::~Track() {
+  Info& info = const_cast<Info&>(m_info);
+  info.Clear();
+
+  ContentEncoding** i = content_encoding_entries_;
+  ContentEncoding** const j = content_encoding_entries_end_;
+
+  while (i != j) {
+    ContentEncoding* const encoding = *i++;
+    delete encoding;
+  }
+
+  delete[] content_encoding_entries_;
+}
+
+long Track::Create(Segment* pSegment, const Info& info, long long element_start,
+                   long long element_size, Track*& pResult) {
+  if (pResult)
+    return -1;
+
+  Track* const pTrack =
+      new (std::nothrow) Track(pSegment, element_start, element_size);
+
+  if (pTrack == NULL)
+    return -1;  // generic error
+
+  const int status = info.Copy(pTrack->m_info);
+
+  if (status) {  // error
+    delete pTrack;
+    return status;
+  }
+
+  pResult = pTrack;
+  return 0;  // success
+}
+
+Track::Info::Info()
+    : uid(0),
+      defaultDuration(0),
+      codecDelay(0),
+      seekPreRoll(0),
+      nameAsUTF8(NULL),
+      language(NULL),
+      codecId(NULL),
+      codecNameAsUTF8(NULL),
+      codecPrivate(NULL),
+      codecPrivateSize(0),
+      lacing(false) {}
+
+Track::Info::~Info() { Clear(); }
+
+void Track::Info::Clear() {
+  delete[] nameAsUTF8;
+  nameAsUTF8 = NULL;
+
+  delete[] language;
+  language = NULL;
+
+  delete[] codecId;
+  codecId = NULL;
+
+  delete[] codecPrivate;
+  codecPrivate = NULL;
+  codecPrivateSize = 0;
+
+  delete[] codecNameAsUTF8;
+  codecNameAsUTF8 = NULL;
+}
+
+int Track::Info::CopyStr(char* Info::*str, Info& dst_) const {
+  if (str == static_cast<char * Info::*>(NULL))
+    return -1;
+
+  char*& dst = dst_.*str;
+
+  if (dst)  // should be NULL already
+    return -1;
+
+  const char* const src = this->*str;
+
+  if (src == NULL)
+    return 0;
+
+  const size_t len = strlen(src);
+
+  dst = SafeArrayAlloc<char>(1, len + 1);
+
+  if (dst == NULL)
+    return -1;
+
+  strcpy(dst, src);
+
+  return 0;
+}
+
+int Track::Info::Copy(Info& dst) const {
+  if (&dst == this)
+    return 0;
+
+  dst.type = type;
+  dst.number = number;
+  dst.defaultDuration = defaultDuration;
+  dst.codecDelay = codecDelay;
+  dst.seekPreRoll = seekPreRoll;
+  dst.uid = uid;
+  dst.lacing = lacing;
+  dst.settings = settings;
+
+  // We now copy the string member variables from src to dst.
+  // This involves memory allocation so in principle the operation
+  // can fail (indeed, that's why we have Info::Copy), so we must
+  // report this to the caller.  An error return from this function
+  // therefore implies that the copy was only partially successful.
+
+  if (int status = CopyStr(&Info::nameAsUTF8, dst))
+    return status;
+
+  if (int status = CopyStr(&Info::language, dst))
+    return status;
+
+  if (int status = CopyStr(&Info::codecId, dst))
+    return status;
+
+  if (int status = CopyStr(&Info::codecNameAsUTF8, dst))
+    return status;
+
+  if (codecPrivateSize > 0) {
+    if (codecPrivate == NULL)
+      return -1;
+
+    if (dst.codecPrivate)
+      return -1;
+
+    if (dst.codecPrivateSize != 0)
+      return -1;
+
+    dst.codecPrivate = SafeArrayAlloc<unsigned char>(1, codecPrivateSize);
+
+    if (dst.codecPrivate == NULL)
+      return -1;
+
+    memcpy(dst.codecPrivate, codecPrivate, codecPrivateSize);
+    dst.codecPrivateSize = codecPrivateSize;
+  }
+
+  return 0;
+}
+
+const BlockEntry* Track::GetEOS() const { return &m_eos; }
+
+long Track::GetType() const { return m_info.type; }
+
+long Track::GetNumber() const { return m_info.number; }
+
+unsigned long long Track::GetUid() const { return m_info.uid; }
+
+const char* Track::GetNameAsUTF8() const { return m_info.nameAsUTF8; }
+
+const char* Track::GetLanguage() const { return m_info.language; }
+
+const char* Track::GetCodecNameAsUTF8() const { return m_info.codecNameAsUTF8; }
+
+const char* Track::GetCodecId() const { return m_info.codecId; }
+
+const unsigned char* Track::GetCodecPrivate(size_t& size) const {
+  size = m_info.codecPrivateSize;
+  return m_info.codecPrivate;
+}
+
+bool Track::GetLacing() const { return m_info.lacing; }
+
+unsigned long long Track::GetDefaultDuration() const {
+  return m_info.defaultDuration;
+}
+
+unsigned long long Track::GetCodecDelay() const { return m_info.codecDelay; }
+
+unsigned long long Track::GetSeekPreRoll() const { return m_info.seekPreRoll; }
+
+long Track::GetFirst(const BlockEntry*& pBlockEntry) const {
+  const Cluster* pCluster = m_pSegment->GetFirst();
+
+  for (int i = 0;;) {
+    if (pCluster == NULL) {
+      pBlockEntry = GetEOS();
+      return 1;
+    }
+
+    if (pCluster->EOS()) {
+      if (m_pSegment->DoneParsing()) {
+        pBlockEntry = GetEOS();
+        return 1;
+      }
+
+      pBlockEntry = 0;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long status = pCluster->GetFirst(pBlockEntry);
+
+    if (status < 0)  // error
+      return status;
+
+    if (pBlockEntry == 0) {  // empty cluster
+      pCluster = m_pSegment->GetNext(pCluster);
+      continue;
+    }
+
+    for (;;) {
+      const Block* const pBlock = pBlockEntry->GetBlock();
+      assert(pBlock);
+
+      const long long tn = pBlock->GetTrackNumber();
+
+      if ((tn == m_info.number) && VetEntry(pBlockEntry))
+        return 0;
+
+      const BlockEntry* pNextEntry;
+
+      status = pCluster->GetNext(pBlockEntry, pNextEntry);
+
+      if (status < 0)  // error
+        return status;
+
+      if (pNextEntry == 0)
+        break;
+
+      pBlockEntry = pNextEntry;
+    }
+
+    ++i;
+
+    if (i >= 100)
+      break;
+
+    pCluster = m_pSegment->GetNext(pCluster);
+  }
+
+  // NOTE: if we get here, it means that we didn't find a block with
+  // a matching track number.  We interpret that as an error (which
+  // might be too conservative).
+
+  pBlockEntry = GetEOS();  // so we can return a non-NULL value
+  return 1;
+}
+
+long Track::GetNext(const BlockEntry* pCurrEntry,
+                    const BlockEntry*& pNextEntry) const {
+  assert(pCurrEntry);
+  assert(!pCurrEntry->EOS());  //?
+
+  const Block* const pCurrBlock = pCurrEntry->GetBlock();
+  assert(pCurrBlock && pCurrBlock->GetTrackNumber() == m_info.number);
+  if (!pCurrBlock || pCurrBlock->GetTrackNumber() != m_info.number)
+    return -1;
+
+  const Cluster* pCluster = pCurrEntry->GetCluster();
+  assert(pCluster);
+  assert(!pCluster->EOS());
+
+  long status = pCluster->GetNext(pCurrEntry, pNextEntry);
+
+  if (status < 0)  // error
+    return status;
+
+  for (int i = 0;;) {
+    while (pNextEntry) {
+      const Block* const pNextBlock = pNextEntry->GetBlock();
+      assert(pNextBlock);
+
+      if (pNextBlock->GetTrackNumber() == m_info.number)
+        return 0;
+
+      pCurrEntry = pNextEntry;
+
+      status = pCluster->GetNext(pCurrEntry, pNextEntry);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    pCluster = m_pSegment->GetNext(pCluster);
+
+    if (pCluster == NULL) {
+      pNextEntry = GetEOS();
+      return 1;
+    }
+
+    if (pCluster->EOS()) {
+      if (m_pSegment->DoneParsing()) {
+        pNextEntry = GetEOS();
+        return 1;
+      }
+
+      // TODO: there is a potential O(n^2) problem here: we tell the
+      // caller to (pre)load another cluster, which he does, but then he
+      // calls GetNext again, which repeats the same search.  This is
+      // a pathological case, since the only way it can happen is if
+      // there exists a long sequence of clusters none of which contain a
+      // block from this track.  One way around this problem is for the
+      // caller to be smarter when he loads another cluster: don't call
+      // us back until you have a cluster that contains a block from this
+      // track. (Of course, that's not cheap either, since our caller
+      // would have to scan the each cluster as it's loaded, so that
+      // would just push back the problem.)
+
+      pNextEntry = NULL;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    status = pCluster->GetFirst(pNextEntry);
+
+    if (status < 0)  // error
+      return status;
+
+    if (pNextEntry == NULL)  // empty cluster
+      continue;
+
+    ++i;
+
+    if (i >= 100)
+      break;
+  }
+
+  // NOTE: if we get here, it means that we didn't find a block with
+  // a matching track number after lots of searching, so we give
+  // up trying.
+
+  pNextEntry = GetEOS();  // so we can return a non-NULL value
+  return 1;
+}
+
+bool Track::VetEntry(const BlockEntry* pBlockEntry) const {
+  assert(pBlockEntry);
+  const Block* const pBlock = pBlockEntry->GetBlock();
+  assert(pBlock);
+  assert(pBlock->GetTrackNumber() == m_info.number);
+  if (!pBlock || pBlock->GetTrackNumber() != m_info.number)
+    return false;
+
+  // This function is used during a seek to determine whether the
+  // frame is a valid seek target.  This default function simply
+  // returns true, which means all frames are valid seek targets.
+  // It gets overridden by the VideoTrack class, because only video
+  // keyframes can be used as seek target.
+
+  return true;
+}
+
+long Track::Seek(long long time_ns, const BlockEntry*& pResult) const {
+  const long status = GetFirst(pResult);
+
+  if (status < 0)  // buffer underflow, etc
+    return status;
+
+  assert(pResult);
+
+  if (pResult->EOS())
+    return 0;
+
+  const Cluster* pCluster = pResult->GetCluster();
+  assert(pCluster);
+  assert(pCluster->GetIndex() >= 0);
+
+  if (time_ns <= pResult->GetBlock()->GetTime(pCluster))
+    return 0;
+
+  Cluster** const clusters = m_pSegment->m_clusters;
+  assert(clusters);
+
+  const long count = m_pSegment->GetCount();  // loaded only, not preloaded
+  assert(count > 0);
+
+  Cluster** const i = clusters + pCluster->GetIndex();
+  assert(i);
+  assert(*i == pCluster);
+  assert(pCluster->GetTime() <= time_ns);
+
+  Cluster** const j = clusters + count;
+
+  Cluster** lo = i;
+  Cluster** hi = j;
+
+  while (lo < hi) {
+    // INVARIANT:
+    //[i, lo) <= time_ns
+    //[lo, hi) ?
+    //[hi, j)  > time_ns
+
+    Cluster** const mid = lo + (hi - lo) / 2;
+    assert(mid < hi);
+
+    pCluster = *mid;
+    assert(pCluster);
+    assert(pCluster->GetIndex() >= 0);
+    assert(pCluster->GetIndex() == long(mid - m_pSegment->m_clusters));
+
+    const long long t = pCluster->GetTime();
+
+    if (t <= time_ns)
+      lo = mid + 1;
+    else
+      hi = mid;
+
+    assert(lo <= hi);
+  }
+
+  assert(lo == hi);
+  assert(lo > i);
+  assert(lo <= j);
+
+  while (lo > i) {
+    pCluster = *--lo;
+    assert(pCluster);
+    assert(pCluster->GetTime() <= time_ns);
+
+    pResult = pCluster->GetEntry(this);
+
+    if ((pResult != 0) && !pResult->EOS())
+      return 0;
+
+    // landed on empty cluster (no entries)
+  }
+
+  pResult = GetEOS();  // weird
+  return 0;
+}
+
+const ContentEncoding* Track::GetContentEncodingByIndex(
+    unsigned long idx) const {
+  const ptrdiff_t count =
+      content_encoding_entries_end_ - content_encoding_entries_;
+  assert(count >= 0);
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return content_encoding_entries_[idx];
+}
+
+unsigned long Track::GetContentEncodingCount() const {
+  const ptrdiff_t count =
+      content_encoding_entries_end_ - content_encoding_entries_;
+  assert(count >= 0);
+
+  return static_cast<unsigned long>(count);
+}
+
+long Track::ParseContentEncodingsEntry(long long start, long long size) {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+  assert(pReader);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  // Count ContentEncoding elements.
+  int count = 0;
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    // pos now designates start of element
+    if (id == libwebm::kMkvContentEncoding)
+      ++count;
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (count <= 0)
+    return -1;
+
+  content_encoding_entries_ = new (std::nothrow) ContentEncoding*[count];
+  if (!content_encoding_entries_)
+    return -1;
+
+  content_encoding_entries_end_ = content_encoding_entries_;
+
+  pos = start;
+  while (pos < stop) {
+    long long id, size;
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    // pos now designates start of element
+    if (id == libwebm::kMkvContentEncoding) {
+      ContentEncoding* const content_encoding =
+          new (std::nothrow) ContentEncoding();
+      if (!content_encoding)
+        return -1;
+
+      status = content_encoding->ParseContentEncodingEntry(pos, size, pReader);
+      if (status) {
+        delete content_encoding;
+        return status;
+      }
+
+      *content_encoding_entries_end_++ = content_encoding;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+Track::EOSBlock::EOSBlock() : BlockEntry(NULL, LONG_MIN) {}
+
+BlockEntry::Kind Track::EOSBlock::GetKind() const { return kBlockEOS; }
+
+const Block* Track::EOSBlock::GetBlock() const { return NULL; }
+
+bool PrimaryChromaticity::Parse(IMkvReader* reader, long long read_pos,
+                                long long value_size, bool is_x,
+                                PrimaryChromaticity** chromaticity) {
+  if (!reader)
+    return false;
+
+  std::auto_ptr<PrimaryChromaticity> chromaticity_ptr;
+
+  if (!*chromaticity) {
+    chromaticity_ptr.reset(new PrimaryChromaticity());
+  } else {
+    chromaticity_ptr.reset(*chromaticity);
+  }
+
+  if (!chromaticity_ptr.get())
+    return false;
+
+  float* value = is_x ? &chromaticity_ptr->x : &chromaticity_ptr->y;
+
+  double parser_value = 0;
+  const long long value_parse_status =
+      UnserializeFloat(reader, read_pos, value_size, parser_value);
+
+  *value = static_cast<float>(parser_value);
+
+  if (value_parse_status < 0 || *value < 0.0 || *value > 1.0)
+    return false;
+
+  *chromaticity = chromaticity_ptr.release();
+  return true;
+}
+
+bool MasteringMetadata::Parse(IMkvReader* reader, long long mm_start,
+                              long long mm_size, MasteringMetadata** mm) {
+  if (!reader || *mm)
+    return false;
+
+  std::auto_ptr<MasteringMetadata> mm_ptr(new MasteringMetadata());
+  if (!mm_ptr.get())
+    return false;
+
+  const long long mm_end = mm_start + mm_size;
+  long long read_pos = mm_start;
+
+  while (read_pos < mm_end) {
+    long long child_id = 0;
+    long long child_size = 0;
+
+    const long long status =
+        ParseElementHeader(reader, read_pos, mm_end, child_id, child_size);
+    if (status < 0)
+      return false;
+
+    if (child_id == libwebm::kMkvLuminanceMax) {
+      double value = 0;
+      const long long value_parse_status =
+          UnserializeFloat(reader, read_pos, child_size, value);
+      mm_ptr->luminance_max = static_cast<float>(value);
+      if (value_parse_status < 0 || mm_ptr->luminance_max < 0.0 ||
+          mm_ptr->luminance_max > 9999.99) {
+        return false;
+      }
+    } else if (child_id == libwebm::kMkvLuminanceMin) {
+      double value = 0;
+      const long long value_parse_status =
+          UnserializeFloat(reader, read_pos, child_size, value);
+      mm_ptr->luminance_min = static_cast<float>(value);
+      if (value_parse_status < 0 || mm_ptr->luminance_min < 0.0 ||
+          mm_ptr->luminance_min > 999.9999) {
+        return false;
+      }
+    } else {
+      bool is_x = false;
+      PrimaryChromaticity** chromaticity;
+      switch (child_id) {
+        case libwebm::kMkvPrimaryRChromaticityX:
+        case libwebm::kMkvPrimaryRChromaticityY:
+          is_x = child_id == libwebm::kMkvPrimaryRChromaticityX;
+          chromaticity = &mm_ptr->r;
+          break;
+        case libwebm::kMkvPrimaryGChromaticityX:
+        case libwebm::kMkvPrimaryGChromaticityY:
+          is_x = child_id == libwebm::kMkvPrimaryGChromaticityX;
+          chromaticity = &mm_ptr->g;
+          break;
+        case libwebm::kMkvPrimaryBChromaticityX:
+        case libwebm::kMkvPrimaryBChromaticityY:
+          is_x = child_id == libwebm::kMkvPrimaryBChromaticityX;
+          chromaticity = &mm_ptr->b;
+          break;
+        case libwebm::kMkvWhitePointChromaticityX:
+        case libwebm::kMkvWhitePointChromaticityY:
+          is_x = child_id == libwebm::kMkvWhitePointChromaticityX;
+          chromaticity = &mm_ptr->white_point;
+          break;
+        default:
+          return false;
+      }
+      const bool value_parse_status = PrimaryChromaticity::Parse(
+          reader, read_pos, child_size, is_x, chromaticity);
+      if (!value_parse_status)
+        return false;
+    }
+
+    read_pos += child_size;
+    if (read_pos > mm_end)
+      return false;
+  }
+
+  *mm = mm_ptr.release();
+  return true;
+}
+
+bool Colour::Parse(IMkvReader* reader, long long colour_start,
+                   long long colour_size, Colour** colour) {
+  if (!reader || *colour)
+    return false;
+
+  std::auto_ptr<Colour> colour_ptr(new Colour());
+  if (!colour_ptr.get())
+    return false;
+
+  const long long colour_end = colour_start + colour_size;
+  long long read_pos = colour_start;
+
+  while (read_pos < colour_end) {
+    long long child_id = 0;
+    long long child_size = 0;
+
+    const long status =
+        ParseElementHeader(reader, read_pos, colour_end, child_id, child_size);
+    if (status < 0)
+      return false;
+
+    if (child_id == libwebm::kMkvMatrixCoefficients) {
+      colour_ptr->matrix_coefficients =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->matrix_coefficients < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvBitsPerChannel) {
+      colour_ptr->bits_per_channel =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->bits_per_channel < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSubsamplingHorz) {
+      colour_ptr->chroma_subsampling_horz =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_subsampling_horz < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSubsamplingVert) {
+      colour_ptr->chroma_subsampling_vert =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_subsampling_vert < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvCbSubsamplingHorz) {
+      colour_ptr->cb_subsampling_horz =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->cb_subsampling_horz < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvCbSubsamplingVert) {
+      colour_ptr->cb_subsampling_vert =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->cb_subsampling_vert < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSitingHorz) {
+      colour_ptr->chroma_siting_horz =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_siting_horz < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSitingVert) {
+      colour_ptr->chroma_siting_vert =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_siting_vert < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvRange) {
+      colour_ptr->range = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->range < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvTransferCharacteristics) {
+      colour_ptr->transfer_characteristics =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->transfer_characteristics < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvPrimaries) {
+      colour_ptr->primaries = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->primaries < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvMaxCLL) {
+      colour_ptr->max_cll = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->max_cll < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvMaxFALL) {
+      colour_ptr->max_fall = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->max_fall < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvMasteringMetadata) {
+      if (!MasteringMetadata::Parse(reader, read_pos, child_size,
+                                    &colour_ptr->mastering_metadata))
+        return false;
+    } else {
+      return false;
+    }
+
+    read_pos += child_size;
+    if (read_pos > colour_end)
+      return false;
+  }
+  *colour = colour_ptr.release();
+  return true;
+}
+
+VideoTrack::VideoTrack(Segment* pSegment, long long element_start,
+                       long long element_size)
+    : Track(pSegment, element_start, element_size), m_colour(NULL) {}
+
+VideoTrack::~VideoTrack() { delete m_colour; }
+
+long VideoTrack::Parse(Segment* pSegment, const Info& info,
+                       long long element_start, long long element_size,
+                       VideoTrack*& pResult) {
+  if (pResult)
+    return -1;
+
+  if (info.type != Track::kVideo)
+    return -1;
+
+  long long width = 0;
+  long long height = 0;
+  long long display_width = 0;
+  long long display_height = 0;
+  long long display_unit = 0;
+  long long stereo_mode = 0;
+
+  double rate = 0.0;
+
+  IMkvReader* const pReader = pSegment->m_pReader;
+
+  const Settings& s = info.settings;
+  assert(s.start >= 0);
+  assert(s.size >= 0);
+
+  long long pos = s.start;
+  assert(pos >= 0);
+
+  const long long stop = pos + s.size;
+
+  Colour* colour = NULL;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvPixelWidth) {
+      width = UnserializeUInt(pReader, pos, size);
+
+      if (width <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvPixelHeight) {
+      height = UnserializeUInt(pReader, pos, size);
+
+      if (height <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDisplayWidth) {
+      display_width = UnserializeUInt(pReader, pos, size);
+
+      if (display_width <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDisplayHeight) {
+      display_height = UnserializeUInt(pReader, pos, size);
+
+      if (display_height <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDisplayUnit) {
+      display_unit = UnserializeUInt(pReader, pos, size);
+
+      if (display_unit < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvStereoMode) {
+      stereo_mode = UnserializeUInt(pReader, pos, size);
+
+      if (stereo_mode < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvFrameRate) {
+      const long status = UnserializeFloat(pReader, pos, size, rate);
+
+      if (status < 0)
+        return status;
+
+      if (rate <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvColour) {
+      if (!Colour::Parse(pReader, pos, size, &colour))
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  VideoTrack* const pTrack =
+      new (std::nothrow) VideoTrack(pSegment, element_start, element_size);
+
+  if (pTrack == NULL)
+    return -1;  // generic error
+
+  const int status = info.Copy(pTrack->m_info);
+
+  if (status) {  // error
+    delete pTrack;
+    return status;
+  }
+
+  pTrack->m_width = width;
+  pTrack->m_height = height;
+  pTrack->m_display_width = display_width;
+  pTrack->m_display_height = display_height;
+  pTrack->m_display_unit = display_unit;
+  pTrack->m_stereo_mode = stereo_mode;
+  pTrack->m_rate = rate;
+  pTrack->m_colour = colour;
+
+  pResult = pTrack;
+  return 0;  // success
+}
+
+bool VideoTrack::VetEntry(const BlockEntry* pBlockEntry) const {
+  return Track::VetEntry(pBlockEntry) && pBlockEntry->GetBlock()->IsKey();
+}
+
+long VideoTrack::Seek(long long time_ns, const BlockEntry*& pResult) const {
+  const long status = GetFirst(pResult);
+
+  if (status < 0)  // buffer underflow, etc
+    return status;
+
+  assert(pResult);
+
+  if (pResult->EOS())
+    return 0;
+
+  const Cluster* pCluster = pResult->GetCluster();
+  assert(pCluster);
+  assert(pCluster->GetIndex() >= 0);
+
+  if (time_ns <= pResult->GetBlock()->GetTime(pCluster))
+    return 0;
+
+  Cluster** const clusters = m_pSegment->m_clusters;
+  assert(clusters);
+
+  const long count = m_pSegment->GetCount();  // loaded only, not pre-loaded
+  assert(count > 0);
+
+  Cluster** const i = clusters + pCluster->GetIndex();
+  assert(i);
+  assert(*i == pCluster);
+  assert(pCluster->GetTime() <= time_ns);
+
+  Cluster** const j = clusters + count;
+
+  Cluster** lo = i;
+  Cluster** hi = j;
+
+  while (lo < hi) {
+    // INVARIANT:
+    //[i, lo) <= time_ns
+    //[lo, hi) ?
+    //[hi, j)  > time_ns
+
+    Cluster** const mid = lo + (hi - lo) / 2;
+    assert(mid < hi);
+
+    pCluster = *mid;
+    assert(pCluster);
+    assert(pCluster->GetIndex() >= 0);
+    assert(pCluster->GetIndex() == long(mid - m_pSegment->m_clusters));
+
+    const long long t = pCluster->GetTime();
+
+    if (t <= time_ns)
+      lo = mid + 1;
+    else
+      hi = mid;
+
+    assert(lo <= hi);
+  }
+
+  assert(lo == hi);
+  assert(lo > i);
+  assert(lo <= j);
+
+  pCluster = *--lo;
+  assert(pCluster);
+  assert(pCluster->GetTime() <= time_ns);
+
+  pResult = pCluster->GetEntry(this, time_ns);
+
+  if ((pResult != 0) && !pResult->EOS())  // found a keyframe
+    return 0;
+
+  while (lo != i) {
+    pCluster = *--lo;
+    assert(pCluster);
+    assert(pCluster->GetTime() <= time_ns);
+
+    pResult = pCluster->GetEntry(this, time_ns);
+
+    if ((pResult != 0) && !pResult->EOS())
+      return 0;
+  }
+
+  // weird: we're on the first cluster, but no keyframe found
+  // should never happen but we must return something anyway
+
+  pResult = GetEOS();
+  return 0;
+}
+
+Colour* VideoTrack::GetColour() const { return m_colour; }
+
+long long VideoTrack::GetWidth() const { return m_width; }
+
+long long VideoTrack::GetHeight() const { return m_height; }
+
+long long VideoTrack::GetDisplayWidth() const {
+  return m_display_width > 0 ? m_display_width : GetWidth();
+}
+
+long long VideoTrack::GetDisplayHeight() const {
+  return m_display_height > 0 ? m_display_height : GetHeight();
+}
+
+long long VideoTrack::GetDisplayUnit() const { return m_display_unit; }
+
+long long VideoTrack::GetStereoMode() const { return m_stereo_mode; }
+
+double VideoTrack::GetFrameRate() const { return m_rate; }
+
+AudioTrack::AudioTrack(Segment* pSegment, long long element_start,
+                       long long element_size)
+    : Track(pSegment, element_start, element_size) {}
+
+long AudioTrack::Parse(Segment* pSegment, const Info& info,
+                       long long element_start, long long element_size,
+                       AudioTrack*& pResult) {
+  if (pResult)
+    return -1;
+
+  if (info.type != Track::kAudio)
+    return -1;
+
+  IMkvReader* const pReader = pSegment->m_pReader;
+
+  const Settings& s = info.settings;
+  assert(s.start >= 0);
+  assert(s.size >= 0);
+
+  long long pos = s.start;
+  assert(pos >= 0);
+
+  const long long stop = pos + s.size;
+
+  double rate = 8000.0;  // MKV default
+  long long channels = 1;
+  long long bit_depth = 0;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvSamplingFrequency) {
+      status = UnserializeFloat(pReader, pos, size, rate);
+
+      if (status < 0)
+        return status;
+
+      if (rate <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvChannels) {
+      channels = UnserializeUInt(pReader, pos, size);
+
+      if (channels <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvBitDepth) {
+      bit_depth = UnserializeUInt(pReader, pos, size);
+
+      if (bit_depth <= 0)
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  AudioTrack* const pTrack =
+      new (std::nothrow) AudioTrack(pSegment, element_start, element_size);
+
+  if (pTrack == NULL)
+    return -1;  // generic error
+
+  const int status = info.Copy(pTrack->m_info);
+
+  if (status) {
+    delete pTrack;
+    return status;
+  }
+
+  pTrack->m_rate = rate;
+  pTrack->m_channels = channels;
+  pTrack->m_bitDepth = bit_depth;
+
+  pResult = pTrack;
+  return 0;  // success
+}
+
+double AudioTrack::GetSamplingRate() const { return m_rate; }
+
+long long AudioTrack::GetChannels() const { return m_channels; }
+
+long long AudioTrack::GetBitDepth() const { return m_bitDepth; }
+
+Tracks::Tracks(Segment* pSegment, long long start, long long size_,
+               long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_trackEntries(NULL),
+      m_trackEntriesEnd(NULL) {}
+
+long Tracks::Parse() {
+  assert(m_trackEntries == NULL);
+  assert(m_trackEntriesEnd == NULL);
+
+  const long long stop = m_start + m_size;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  int count = 0;
+  long long pos = m_start;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvTrackEntry)
+      ++count;
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (count <= 0)
+    return 0;  // success
+
+  m_trackEntries = new (std::nothrow) Track*[count];
+
+  if (m_trackEntries == NULL)
+    return -1;
+
+  m_trackEntriesEnd = m_trackEntries;
+
+  pos = m_start;
+
+  while (pos < stop) {
+    const long long element_start = pos;
+
+    long long id, payload_size;
+
+    const long status =
+        ParseElementHeader(pReader, pos, stop, id, payload_size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (payload_size == 0)  // weird
+      continue;
+
+    const long long payload_stop = pos + payload_size;
+    assert(payload_stop <= stop);  // checked in ParseElement
+
+    const long long element_size = payload_stop - element_start;
+
+    if (id == libwebm::kMkvTrackEntry) {
+      Track*& pTrack = *m_trackEntriesEnd;
+      pTrack = NULL;
+
+      const long status = ParseTrackEntry(pos, payload_size, element_start,
+                                          element_size, pTrack);
+      if (status)
+        return status;
+
+      if (pTrack)
+        ++m_trackEntriesEnd;
+    }
+
+    pos = payload_stop;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;  // success
+}
+
+unsigned long Tracks::GetTracksCount() const {
+  const ptrdiff_t result = m_trackEntriesEnd - m_trackEntries;
+  assert(result >= 0);
+
+  return static_cast<unsigned long>(result);
+}
+
+long Tracks::ParseTrackEntry(long long track_start, long long track_size,
+                             long long element_start, long long element_size,
+                             Track*& pResult) const {
+  if (pResult)
+    return -1;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = track_start;
+  const long long track_stop = track_start + track_size;
+
+  Track::Info info;
+
+  info.type = 0;
+  info.number = 0;
+  info.uid = 0;
+  info.defaultDuration = 0;
+
+  Track::Settings v;
+  v.start = -1;
+  v.size = -1;
+
+  Track::Settings a;
+  a.start = -1;
+  a.size = -1;
+
+  Track::Settings e;  // content_encodings_settings;
+  e.start = -1;
+  e.size = -1;
+
+  long long lacing = 1;  // default is true
+
+  while (pos < track_stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, track_stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long start = pos;
+
+    if (id == libwebm::kMkvVideo) {
+      v.start = start;
+      v.size = size;
+    } else if (id == libwebm::kMkvAudio) {
+      a.start = start;
+      a.size = size;
+    } else if (id == libwebm::kMkvContentEncodings) {
+      e.start = start;
+      e.size = size;
+    } else if (id == libwebm::kMkvTrackUID) {
+      if (size > 8)
+        return E_FILE_FORMAT_INVALID;
+
+      info.uid = 0;
+
+      long long pos_ = start;
+      const long long pos_end = start + size;
+
+      while (pos_ != pos_end) {
+        unsigned char b;
+
+        const int status = pReader->Read(pos_, 1, &b);
+
+        if (status)
+          return status;
+
+        info.uid <<= 8;
+        info.uid |= b;
+
+        ++pos_;
+      }
+    } else if (id == libwebm::kMkvTrackNumber) {
+      const long long num = UnserializeUInt(pReader, pos, size);
+
+      if ((num <= 0) || (num > 127))
+        return E_FILE_FORMAT_INVALID;
+
+      info.number = static_cast<long>(num);
+    } else if (id == libwebm::kMkvTrackType) {
+      const long long type = UnserializeUInt(pReader, pos, size);
+
+      if ((type <= 0) || (type > 254))
+        return E_FILE_FORMAT_INVALID;
+
+      info.type = static_cast<long>(type);
+    } else if (id == libwebm::kMkvName) {
+      const long status =
+          UnserializeString(pReader, pos, size, info.nameAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvLanguage) {
+      const long status = UnserializeString(pReader, pos, size, info.language);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvDefaultDuration) {
+      const long long duration = UnserializeUInt(pReader, pos, size);
+
+      if (duration < 0)
+        return E_FILE_FORMAT_INVALID;
+
+      info.defaultDuration = static_cast<unsigned long long>(duration);
+    } else if (id == libwebm::kMkvCodecID) {
+      const long status = UnserializeString(pReader, pos, size, info.codecId);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvFlagLacing) {
+      lacing = UnserializeUInt(pReader, pos, size);
+
+      if ((lacing < 0) || (lacing > 1))
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvCodecPrivate) {
+      delete[] info.codecPrivate;
+      info.codecPrivate = NULL;
+      info.codecPrivateSize = 0;
+
+      const size_t buflen = static_cast<size_t>(size);
+
+      if (buflen) {
+        unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+
+        if (buf == NULL)
+          return -1;
+
+        const int status = pReader->Read(pos, static_cast<long>(buflen), buf);
+
+        if (status) {
+          delete[] buf;
+          return status;
+        }
+
+        info.codecPrivate = buf;
+        info.codecPrivateSize = buflen;
+      }
+    } else if (id == libwebm::kMkvCodecName) {
+      const long status =
+          UnserializeString(pReader, pos, size, info.codecNameAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvCodecDelay) {
+      info.codecDelay = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvSeekPreRoll) {
+      info.seekPreRoll = UnserializeUInt(pReader, pos, size);
+    }
+
+    pos += size;  // consume payload
+    if (pos > track_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != track_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (info.number <= 0)  // not specified
+    return E_FILE_FORMAT_INVALID;
+
+  if (GetTrackByNumber(info.number))
+    return E_FILE_FORMAT_INVALID;
+
+  if (info.type <= 0)  // not specified
+    return E_FILE_FORMAT_INVALID;
+
+  info.lacing = (lacing > 0) ? true : false;
+
+  if (info.type == Track::kVideo) {
+    if (v.start < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (a.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    info.settings = v;
+
+    VideoTrack* pTrack = NULL;
+
+    const long status = VideoTrack::Parse(m_pSegment, info, element_start,
+                                          element_size, pTrack);
+
+    if (status)
+      return status;
+
+    pResult = pTrack;
+    assert(pResult);
+
+    if (e.start >= 0)
+      pResult->ParseContentEncodingsEntry(e.start, e.size);
+  } else if (info.type == Track::kAudio) {
+    if (a.start < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (v.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    info.settings = a;
+
+    AudioTrack* pTrack = NULL;
+
+    const long status = AudioTrack::Parse(m_pSegment, info, element_start,
+                                          element_size, pTrack);
+
+    if (status)
+      return status;
+
+    pResult = pTrack;
+    assert(pResult);
+
+    if (e.start >= 0)
+      pResult->ParseContentEncodingsEntry(e.start, e.size);
+  } else {
+    // neither video nor audio - probably metadata or subtitles
+
+    if (a.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (v.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (info.type == Track::kMetadata && e.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    info.settings.start = -1;
+    info.settings.size = 0;
+
+    Track* pTrack = NULL;
+
+    const long status =
+        Track::Create(m_pSegment, info, element_start, element_size, pTrack);
+
+    if (status)
+      return status;
+
+    pResult = pTrack;
+    assert(pResult);
+  }
+
+  return 0;  // success
+}
+
+Tracks::~Tracks() {
+  Track** i = m_trackEntries;
+  Track** const j = m_trackEntriesEnd;
+
+  while (i != j) {
+    Track* const pTrack = *i++;
+    delete pTrack;
+  }
+
+  delete[] m_trackEntries;
+}
+
+const Track* Tracks::GetTrackByNumber(long tn) const {
+  if (tn < 0)
+    return NULL;
+
+  Track** i = m_trackEntries;
+  Track** const j = m_trackEntriesEnd;
+
+  while (i != j) {
+    Track* const pTrack = *i++;
+
+    if (pTrack == NULL)
+      continue;
+
+    if (tn == pTrack->GetNumber())
+      return pTrack;
+  }
+
+  return NULL;  // not found
+}
+
+const Track* Tracks::GetTrackByIndex(unsigned long idx) const {
+  const ptrdiff_t count = m_trackEntriesEnd - m_trackEntries;
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return m_trackEntries[idx];
+}
+
+long Cluster::Load(long long& pos, long& len) const {
+  if (m_pSegment == NULL)
+    return E_PARSE_FAILED;
+
+  if (m_timecode >= 0)  // at least partially loaded
+    return 0;
+
+  if (m_pos != m_element_start || m_element_size >= 0)
+    return E_PARSE_FAILED;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+  long long total, avail;
+  const int status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total >= 0 && (avail > total || m_pos > total))
+    return E_FILE_FORMAT_INVALID;
+
+  pos = m_pos;
+
+  long long cluster_size = -1;
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error or underflow
+    return static_cast<long>(result);
+
+  if (result > 0)
+    return E_BUFFER_NOT_FULL;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long id_ = ReadID(pReader, pos, len);
+
+  if (id_ < 0)  // error
+    return static_cast<long>(id_);
+
+  if (id_ != libwebm::kMkvCluster)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume id
+
+  // read cluster size
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)
+    return E_BUFFER_NOT_FULL;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long size = ReadUInt(pReader, pos, len);
+
+  if (size < 0)  // error
+    return static_cast<long>(cluster_size);
+
+  if (size == 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume length of size of element
+
+  const long long unknown_size = (1LL << (7 * len)) - 1;
+
+  if (size != unknown_size)
+    cluster_size = size;
+
+  // pos points to start of payload
+  long long timecode = -1;
+  long long new_pos = -1;
+  bool bBlock = false;
+
+  long long cluster_stop = (cluster_size < 0) ? -1 : pos + cluster_size;
+
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      break;
+
+    // Parse ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id == 0)
+      return E_FILE_FORMAT_INVALID;
+
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
+
+    if (id == libwebm::kMkvCluster)
+      break;
+
+    if (id == libwebm::kMkvCues)
+      break;
+
+    pos += len;  // consume ID field
+
+    // Parse Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume size field
+
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // pos now points to start of payload
+
+    if (size == 0)
+      continue;
+
+    if ((cluster_stop >= 0) && ((pos + size) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvTimecode) {
+      len = static_cast<long>(size);
+
+      if ((pos + size) > avail)
+        return E_BUFFER_NOT_FULL;
+
+      timecode = UnserializeUInt(pReader, pos, size);
+
+      if (timecode < 0)  // error (or underflow)
+        return static_cast<long>(timecode);
+
+      new_pos = pos + size;
+
+      if (bBlock)
+        break;
+    } else if (id == libwebm::kMkvBlockGroup) {
+      bBlock = true;
+      break;
+    } else if (id == libwebm::kMkvSimpleBlock) {
+      bBlock = true;
+      break;
+    }
+
+    pos += size;  // consume payload
+    if (cluster_stop >= 0 && pos > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (cluster_stop >= 0 && pos > cluster_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (timecode < 0)  // no timecode found
+    return E_FILE_FORMAT_INVALID;
+
+  if (!bBlock)
+    return E_FILE_FORMAT_INVALID;
+
+  m_pos = new_pos;  // designates position just beyond timecode payload
+  m_timecode = timecode;  // m_timecode >= 0 means we're partially loaded
+
+  if (cluster_size >= 0)
+    m_element_size = cluster_stop - m_element_start;
+
+  return 0;
+}
+
+long Cluster::Parse(long long& pos, long& len) const {
+  long status = Load(pos, len);
+
+  if (status < 0)
+    return status;
+
+  if (m_pos < m_element_start || m_timecode < 0)
+    return E_PARSE_FAILED;
+
+  const long long cluster_stop =
+      (m_element_size < 0) ? -1 : m_element_start + m_element_size;
+
+  if ((cluster_stop >= 0) && (m_pos >= cluster_stop))
+    return 1;  // nothing else to do
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long total, avail;
+
+  status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total >= 0 && avail > total)
+    return E_FILE_FORMAT_INVALID;
+
+  pos = m_pos;
+
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      break;
+
+    if ((total >= 0) && (pos >= total)) {
+      if (m_element_size < 0)
+        m_element_size = pos - m_element_start;
+
+      break;
+    }
+
+    // Parse ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
+
+    if ((id == libwebm::kMkvCluster) || (id == libwebm::kMkvCues)) {
+      if (m_element_size < 0)
+        m_element_size = pos - m_element_start;
+
+      break;
+    }
+
+    pos += len;  // consume ID field
+
+    // Parse Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume size field
+
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // pos now points to start of payload
+
+    if (size == 0)
+      continue;
+
+    // const long long block_start = pos;
+    const long long block_stop = pos + size;
+
+    if (cluster_stop >= 0) {
+      if (block_stop > cluster_stop) {
+        if (id == libwebm::kMkvBlockGroup || id == libwebm::kMkvSimpleBlock) {
+          return E_FILE_FORMAT_INVALID;
+        }
+
+        pos = cluster_stop;
+        break;
+      }
+    } else if ((total >= 0) && (block_stop > total)) {
+      m_element_size = total - m_element_start;
+      pos = total;
+      break;
+    } else if (block_stop > avail) {
+      len = static_cast<long>(size);
+      return E_BUFFER_NOT_FULL;
+    }
+
+    Cluster* const this_ = const_cast<Cluster*>(this);
+
+    if (id == libwebm::kMkvBlockGroup)
+      return this_->ParseBlockGroup(size, pos, len);
+
+    if (id == libwebm::kMkvSimpleBlock)
+      return this_->ParseSimpleBlock(size, pos, len);
+
+    pos += size;  // consume payload
+    if (cluster_stop >= 0 && pos > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (m_element_size < 1)
+    return E_FILE_FORMAT_INVALID;
+
+  m_pos = pos;
+  if (cluster_stop >= 0 && m_pos > cluster_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (m_entries_count > 0) {
+    const long idx = m_entries_count - 1;
+
+    const BlockEntry* const pLast = m_entries[idx];
+    if (pLast == NULL)
+      return E_PARSE_FAILED;
+
+    const Block* const pBlock = pLast->GetBlock();
+    if (pBlock == NULL)
+      return E_PARSE_FAILED;
+
+    const long long start = pBlock->m_start;
+
+    if ((total >= 0) && (start > total))
+      return E_PARSE_FAILED;  // defend against trucated stream
+
+    const long long size = pBlock->m_size;
+
+    const long long stop = start + size;
+    if (cluster_stop >= 0 && stop > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && (stop > total))
+      return E_PARSE_FAILED;  // defend against trucated stream
+  }
+
+  return 1;  // no more entries
+}
+
+long Cluster::ParseSimpleBlock(long long block_size, long long& pos,
+                               long& len) {
+  const long long block_start = pos;
+  const long long block_stop = pos + block_size;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long total, avail;
+
+  long status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  // parse track number
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)  // weird
+    return E_BUFFER_NOT_FULL;
+
+  if ((pos + len) > block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long track = ReadUInt(pReader, pos, len);
+
+  if (track < 0)  // error
+    return static_cast<long>(track);
+
+  if (track == 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume track number
+
+  if ((pos + 2) > block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + 2) > avail) {
+    len = 2;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  pos += 2;  // consume timecode
+
+  if ((pos + 1) > block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  unsigned char flags;
+
+  status = pReader->Read(pos, 1, &flags);
+
+  if (status < 0) {  // error or underflow
+    len = 1;
+    return status;
+  }
+
+  ++pos;  // consume flags byte
+  assert(pos <= avail);
+
+  if (pos >= block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  const int lacing = int(flags & 0x06) >> 1;
+
+  if ((lacing != 0) && (block_stop > avail)) {
+    len = static_cast<long>(block_stop - pos);
+    return E_BUFFER_NOT_FULL;
+  }
+
+  status = CreateBlock(libwebm::kMkvSimpleBlock, block_start, block_size,
+                       0);  // DiscardPadding
+
+  if (status != 0)
+    return status;
+
+  m_pos = block_stop;
+
+  return 0;  // success
+}
+
+long Cluster::ParseBlockGroup(long long payload_size, long long& pos,
+                              long& len) {
+  const long long payload_start = pos;
+  const long long payload_stop = pos + payload_size;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long total, avail;
+
+  long status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  if ((total >= 0) && (payload_stop > total))
+    return E_FILE_FORMAT_INVALID;
+
+  if (payload_stop > avail) {
+    len = static_cast<long>(payload_size);
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long discard_padding = 0;
+
+  while (pos < payload_stop) {
+    // parse sub-block element ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((pos + len) > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id == 0)  // not a valid ID
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID field
+
+    // Parse Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((pos + len) > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume size field
+
+    // pos now points to start of sub-block group payload
+
+    if (pos > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if (size == 0)  // weird
+      continue;
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvDiscardPadding) {
+      status = UnserializeInt(pReader, pos, size, discard_padding);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    if (id != libwebm::kMkvBlock) {
+      pos += size;  // consume sub-part of block group
+
+      if (pos > payload_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      continue;
+    }
+
+    const long long block_stop = pos + size;
+
+    if (block_stop > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    // parse track number
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((pos + len) > block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long track = ReadUInt(pReader, pos, len);
+
+    if (track < 0)  // error
+      return static_cast<long>(track);
+
+    if (track == 0)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume track number
+
+    if ((pos + 2) > block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + 2) > avail) {
+      len = 2;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    pos += 2;  // consume timecode
+
+    if ((pos + 1) > block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    unsigned char flags;
+
+    status = pReader->Read(pos, 1, &flags);
+
+    if (status < 0) {  // error or underflow
+      len = 1;
+      return status;
+    }
+
+    ++pos;  // consume flags byte
+    assert(pos <= avail);
+
+    if (pos >= block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    const int lacing = int(flags & 0x06) >> 1;
+
+    if ((lacing != 0) && (block_stop > avail)) {
+      len = static_cast<long>(block_stop - pos);
+      return E_BUFFER_NOT_FULL;
+    }
+
+    pos = block_stop;  // consume block-part of block group
+    if (pos > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != payload_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  status = CreateBlock(libwebm::kMkvBlockGroup, payload_start, payload_size,
+                       discard_padding);
+  if (status != 0)
+    return status;
+
+  m_pos = payload_stop;
+
+  return 0;  // success
+}
+
+long Cluster::GetEntry(long index, const mkvparser::BlockEntry*& pEntry) const {
+  assert(m_pos >= m_element_start);
+
+  pEntry = NULL;
+
+  if (index < 0)
+    return -1;  // generic error
+
+  if (m_entries_count < 0)
+    return E_BUFFER_NOT_FULL;
+
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count <= m_entries_size);
+
+  if (index < m_entries_count) {
+    pEntry = m_entries[index];
+    assert(pEntry);
+
+    return 1;  // found entry
+  }
+
+  if (m_element_size < 0)  // we don't know cluster end yet
+    return E_BUFFER_NOT_FULL;  // underflow
+
+  const long long element_stop = m_element_start + m_element_size;
+
+  if (m_pos >= element_stop)
+    return 0;  // nothing left to parse
+
+  return E_BUFFER_NOT_FULL;  // underflow, since more remains to be parsed
+}
+
+Cluster* Cluster::Create(Segment* pSegment, long idx, long long off) {
+  if (!pSegment || off < 0)
+    return NULL;
+
+  const long long element_start = pSegment->m_start + off;
+
+  Cluster* const pCluster =
+      new (std::nothrow) Cluster(pSegment, idx, element_start);
+
+  return pCluster;
+}
+
+Cluster::Cluster()
+    : m_pSegment(NULL),
+      m_element_start(0),
+      m_index(0),
+      m_pos(0),
+      m_element_size(0),
+      m_timecode(0),
+      m_entries(NULL),
+      m_entries_size(0),
+      m_entries_count(0)  // means "no entries"
+{}
+
+Cluster::Cluster(Segment* pSegment, long idx, long long element_start
+                 /* long long element_size */)
+    : m_pSegment(pSegment),
+      m_element_start(element_start),
+      m_index(idx),
+      m_pos(element_start),
+      m_element_size(-1 /* element_size */),
+      m_timecode(-1),
+      m_entries(NULL),
+      m_entries_size(0),
+      m_entries_count(-1)  // means "has not been parsed yet"
+{}
+
+Cluster::~Cluster() {
+  if (m_entries_count <= 0)
+    return;
+
+  BlockEntry** i = m_entries;
+  BlockEntry** const j = m_entries + m_entries_count;
+
+  while (i != j) {
+    BlockEntry* p = *i++;
+    assert(p);
+
+    delete p;
+  }
+
+  delete[] m_entries;
+}
+
+bool Cluster::EOS() const { return (m_pSegment == NULL); }
+
+long Cluster::GetIndex() const { return m_index; }
+
+long long Cluster::GetPosition() const {
+  const long long pos = m_element_start - m_pSegment->m_start;
+  assert(pos >= 0);
+
+  return pos;
+}
+
+long long Cluster::GetElementSize() const { return m_element_size; }
+
+long Cluster::HasBlockEntries(
+    const Segment* pSegment,
+    long long off,  // relative to start of segment payload
+    long long& pos, long& len) {
+  assert(pSegment);
+  assert(off >= 0);  // relative to segment
+
+  IMkvReader* const pReader = pSegment->m_pReader;
+
+  long long total, avail;
+
+  long status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  pos = pSegment->m_start + off;  // absolute
+
+  if ((total >= 0) && (pos >= total))
+    return 0;  // we don't even have a complete cluster
+
+  const long long segment_stop =
+      (pSegment->m_size < 0) ? -1 : pSegment->m_start + pSegment->m_size;
+
+  long long cluster_stop = -1;  // interpreted later to mean "unknown size"
+
+  {
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // need more data
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return 0;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id != libwebm::kMkvCluster)
+      return E_PARSE_FAILED;
+
+    pos += len;  // consume Cluster ID field
+
+    // read size field
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return 0;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    if (size == 0)
+      return 0;  // cluster does not have entries
+
+    pos += len;  // consume size field
+
+    // pos now points to start of payload
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size != unknown_size) {
+      cluster_stop = pos + size;
+      assert(cluster_stop >= 0);
+
+      if ((segment_stop >= 0) && (cluster_stop > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      if ((total >= 0) && (cluster_stop > total))
+        // return E_FILE_FORMAT_INVALID;  //too conservative
+        return 0;  // cluster does not have any entries
+    }
+  }
+
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      return 0;  // no entries detected
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // need more data
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
+
+    if (id == libwebm::kMkvCluster)
+      return 0;  // no entries found
+
+    if (id == libwebm::kMkvCues)
+      return 0;  // no entries found
+
+    pos += len;  // consume id field
+
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // read size field
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // underflow
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume size field
+
+    // pos now points to start of payload
+
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if (size == 0)  // weird
+      continue;
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;  // not supported inside cluster
+
+    if ((cluster_stop >= 0) && ((pos + size) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvBlockGroup)
+      return 1;  // have at least one entry
+
+    if (id == libwebm::kMkvSimpleBlock)
+      return 1;  // have at least one entry
+
+    pos += size;  // consume payload
+    if (cluster_stop >= 0 && pos > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+}
+
+long long Cluster::GetTimeCode() const {
+  long long pos;
+  long len;
+
+  const long status = Load(pos, len);
+
+  if (status < 0)  // error
+    return status;
+
+  return m_timecode;
+}
+
+long long Cluster::GetTime() const {
+  const long long tc = GetTimeCode();
+
+  if (tc < 0)
+    return tc;
+
+  const SegmentInfo* const pInfo = m_pSegment->GetInfo();
+  assert(pInfo);
+
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
+
+  const long long t = m_timecode * scale;
+
+  return t;
+}
+
+long long Cluster::GetFirstTime() const {
+  const BlockEntry* pEntry;
+
+  const long status = GetFirst(pEntry);
+
+  if (status < 0)  // error
+    return status;
+
+  if (pEntry == NULL)  // empty cluster
+    return GetTime();
+
+  const Block* const pBlock = pEntry->GetBlock();
+  assert(pBlock);
+
+  return pBlock->GetTime(this);
+}
+
+long long Cluster::GetLastTime() const {
+  const BlockEntry* pEntry;
+
+  const long status = GetLast(pEntry);
+
+  if (status < 0)  // error
+    return status;
+
+  if (pEntry == NULL)  // empty cluster
+    return GetTime();
+
+  const Block* const pBlock = pEntry->GetBlock();
+  assert(pBlock);
+
+  return pBlock->GetTime(this);
+}
+
+long Cluster::CreateBlock(long long id,
+                          long long pos,  // absolute pos of payload
+                          long long size, long long discard_padding) {
+  if (id != libwebm::kMkvBlockGroup && id != libwebm::kMkvSimpleBlock)
+    return E_PARSE_FAILED;
+
+  if (m_entries_count < 0) {  // haven't parsed anything yet
+    assert(m_entries == NULL);
+    assert(m_entries_size == 0);
+
+    m_entries_size = 1024;
+    m_entries = new (std::nothrow) BlockEntry*[m_entries_size];
+    if (m_entries == NULL)
+      return -1;
+
+    m_entries_count = 0;
+  } else {
+    assert(m_entries);
+    assert(m_entries_size > 0);
+    assert(m_entries_count <= m_entries_size);
+
+    if (m_entries_count >= m_entries_size) {
+      const long entries_size = 2 * m_entries_size;
+
+      BlockEntry** const entries = new (std::nothrow) BlockEntry*[entries_size];
+      if (entries == NULL)
+        return -1;
+
+      BlockEntry** src = m_entries;
+      BlockEntry** const src_end = src + m_entries_count;
+
+      BlockEntry** dst = entries;
+
+      while (src != src_end)
+        *dst++ = *src++;
+
+      delete[] m_entries;
+
+      m_entries = entries;
+      m_entries_size = entries_size;
+    }
+  }
+
+  if (id == libwebm::kMkvBlockGroup)
+    return CreateBlockGroup(pos, size, discard_padding);
+  else
+    return CreateSimpleBlock(pos, size);
+}
+
+long Cluster::CreateBlockGroup(long long start_offset, long long size,
+                               long long discard_padding) {
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count >= 0);
+  assert(m_entries_count < m_entries_size);
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = start_offset;
+  const long long stop = start_offset + size;
+
+  // For WebM files, there is a bias towards previous reference times
+  //(in order to support alt-ref frames, which refer back to the previous
+  // keyframe).  Normally a 0 value is not possible, but here we tenatively
+  // allow 0 as the value of a reference frame, with the interpretation
+  // that this is a "previous" reference time.
+
+  long long prev = 1;  // nonce
+  long long next = 0;  // nonce
+  long long duration = -1;  // really, this is unsigned
+
+  long long bpos = -1;
+  long long bsize = -1;
+
+  while (pos < stop) {
+    long len;
+    const long long id = ReadID(pReader, pos, len);
+    if (id < 0 || (pos + len) > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);  // TODO
+    assert((pos + len) <= stop);
+
+    pos += len;  // consume size
+
+    if (id == libwebm::kMkvBlock) {
+      if (bpos < 0) {  // Block ID
+        bpos = pos;
+        bsize = size;
+      }
+    } else if (id == libwebm::kMkvBlockDuration) {
+      if (size > 8)
+        return E_FILE_FORMAT_INVALID;
+
+      duration = UnserializeUInt(pReader, pos, size);
+
+      if (duration < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvReferenceBlock) {
+      if (size > 8 || size <= 0)
+        return E_FILE_FORMAT_INVALID;
+      const long size_ = static_cast<long>(size);
+
+      long long time;
+
+      long status = UnserializeInt(pReader, pos, size_, time);
+      assert(status == 0);
+      if (status != 0)
+        return -1;
+
+      if (time <= 0)  // see note above
+        prev = time;
+      else
+        next = time;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+  if (bpos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  assert(bsize >= 0);
+
+  const long idx = m_entries_count;
+
+  BlockEntry** const ppEntry = m_entries + idx;
+  BlockEntry*& pEntry = *ppEntry;
+
+  pEntry = new (std::nothrow)
+      BlockGroup(this, idx, bpos, bsize, prev, next, duration, discard_padding);
+
+  if (pEntry == NULL)
+    return -1;  // generic error
+
+  BlockGroup* const p = static_cast<BlockGroup*>(pEntry);
+
+  const long status = p->Parse();
+
+  if (status == 0) {  // success
+    ++m_entries_count;
+    return 0;
+  }
+
+  delete pEntry;
+  pEntry = 0;
+
+  return status;
+}
+
+long Cluster::CreateSimpleBlock(long long st, long long sz) {
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count >= 0);
+  assert(m_entries_count < m_entries_size);
+
+  const long idx = m_entries_count;
+
+  BlockEntry** const ppEntry = m_entries + idx;
+  BlockEntry*& pEntry = *ppEntry;
+
+  pEntry = new (std::nothrow) SimpleBlock(this, idx, st, sz);
+
+  if (pEntry == NULL)
+    return -1;  // generic error
+
+  SimpleBlock* const p = static_cast<SimpleBlock*>(pEntry);
+
+  const long status = p->Parse();
+
+  if (status == 0) {
+    ++m_entries_count;
+    return 0;
+  }
+
+  delete pEntry;
+  pEntry = 0;
+
+  return status;
+}
+
+long Cluster::GetFirst(const BlockEntry*& pFirst) const {
+  if (m_entries_count <= 0) {
+    long long pos;
+    long len;
+
+    const long status = Parse(pos, len);
+
+    if (status < 0) {  // error
+      pFirst = NULL;
+      return status;
+    }
+
+    if (m_entries_count <= 0) {  // empty cluster
+      pFirst = NULL;
+      return 0;
+    }
+  }
+
+  assert(m_entries);
+
+  pFirst = m_entries[0];
+  assert(pFirst);
+
+  return 0;  // success
+}
+
+long Cluster::GetLast(const BlockEntry*& pLast) const {
+  for (;;) {
+    long long pos;
+    long len;
+
+    const long status = Parse(pos, len);
+
+    if (status < 0) {  // error
+      pLast = NULL;
+      return status;
+    }
+
+    if (status > 0)  // no new block
+      break;
+  }
+
+  if (m_entries_count <= 0) {
+    pLast = NULL;
+    return 0;
+  }
+
+  assert(m_entries);
+
+  const long idx = m_entries_count - 1;
+
+  pLast = m_entries[idx];
+  assert(pLast);
+
+  return 0;
+}
+
+long Cluster::GetNext(const BlockEntry* pCurr, const BlockEntry*& pNext) const {
+  assert(pCurr);
+  assert(m_entries);
+  assert(m_entries_count > 0);
+
+  size_t idx = pCurr->GetIndex();
+  assert(idx < size_t(m_entries_count));
+  assert(m_entries[idx] == pCurr);
+
+  ++idx;
+
+  if (idx >= size_t(m_entries_count)) {
+    long long pos;
+    long len;
+
+    const long status = Parse(pos, len);
+
+    if (status < 0) {  // error
+      pNext = NULL;
+      return status;
+    }
+
+    if (status > 0) {
+      pNext = NULL;
+      return 0;
+    }
+
+    assert(m_entries);
+    assert(m_entries_count > 0);
+    assert(idx < size_t(m_entries_count));
+  }
+
+  pNext = m_entries[idx];
+  assert(pNext);
+
+  return 0;
+}
+
+long Cluster::GetEntryCount() const { return m_entries_count; }
+
+const BlockEntry* Cluster::GetEntry(const Track* pTrack,
+                                    long long time_ns) const {
+  assert(pTrack);
+
+  if (m_pSegment == NULL)  // this is the special EOS cluster
+    return pTrack->GetEOS();
+
+  const BlockEntry* pResult = pTrack->GetEOS();
+
+  long index = 0;
+
+  for (;;) {
+    if (index >= m_entries_count) {
+      long long pos;
+      long len;
+
+      const long status = Parse(pos, len);
+      assert(status >= 0);
+
+      if (status > 0)  // completely parsed, and no more entries
+        return pResult;
+
+      if (status < 0)  // should never happen
+        return 0;
+
+      assert(m_entries);
+      assert(index < m_entries_count);
+    }
+
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
+
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
+
+    if (pBlock->GetTrackNumber() != pTrack->GetNumber()) {
+      ++index;
+      continue;
+    }
+
+    if (pTrack->VetEntry(pEntry)) {
+      if (time_ns < 0)  // just want first candidate block
+        return pEntry;
+
+      const long long ns = pBlock->GetTime(this);
+
+      if (ns > time_ns)
+        return pResult;
+
+      pResult = pEntry;  // have a candidate
+    } else if (time_ns >= 0) {
+      const long long ns = pBlock->GetTime(this);
+
+      if (ns > time_ns)
+        return pResult;
+    }
+
+    ++index;
+  }
+}
+
+const BlockEntry* Cluster::GetEntry(const CuePoint& cp,
+                                    const CuePoint::TrackPosition& tp) const {
+  assert(m_pSegment);
+  const long long tc = cp.GetTimeCode();
+
+  if (tp.m_block > 0) {
+    const long block = static_cast<long>(tp.m_block);
+    const long index = block - 1;
+
+    while (index >= m_entries_count) {
+      long long pos;
+      long len;
+
+      const long status = Parse(pos, len);
+
+      if (status < 0)  // TODO: can this happen?
+        return NULL;
+
+      if (status > 0)  // nothing remains to be parsed
+        return NULL;
+    }
+
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
+
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
+
+    if ((pBlock->GetTrackNumber() == tp.m_track) &&
+        (pBlock->GetTimeCode(this) == tc)) {
+      return pEntry;
+    }
+  }
+
+  long index = 0;
+
+  for (;;) {
+    if (index >= m_entries_count) {
+      long long pos;
+      long len;
+
+      const long status = Parse(pos, len);
+
+      if (status < 0)  // TODO: can this happen?
+        return NULL;
+
+      if (status > 0)  // nothing remains to be parsed
+        return NULL;
+
+      assert(m_entries);
+      assert(index < m_entries_count);
+    }
+
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
+
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
+
+    if (pBlock->GetTrackNumber() != tp.m_track) {
+      ++index;
+      continue;
+    }
+
+    const long long tc_ = pBlock->GetTimeCode(this);
+
+    if (tc_ < tc) {
+      ++index;
+      continue;
+    }
+
+    if (tc_ > tc)
+      return NULL;
+
+    const Tracks* const pTracks = m_pSegment->GetTracks();
+    assert(pTracks);
+
+    const long tn = static_cast<long>(tp.m_track);
+    const Track* const pTrack = pTracks->GetTrackByNumber(tn);
+
+    if (pTrack == NULL)
+      return NULL;
+
+    const long long type = pTrack->GetType();
+
+    if (type == 2)  // audio
+      return pEntry;
+
+    if (type != 1)  // not video
+      return NULL;
+
+    if (!pBlock->IsKey())
+      return NULL;
+
+    return pEntry;
+  }
+}
+
+BlockEntry::BlockEntry(Cluster* p, long idx) : m_pCluster(p), m_index(idx) {}
+BlockEntry::~BlockEntry() {}
+const Cluster* BlockEntry::GetCluster() const { return m_pCluster; }
+long BlockEntry::GetIndex() const { return m_index; }
+
+SimpleBlock::SimpleBlock(Cluster* pCluster, long idx, long long start,
+                         long long size)
+    : BlockEntry(pCluster, idx), m_block(start, size, 0) {}
+
+long SimpleBlock::Parse() { return m_block.Parse(m_pCluster); }
+BlockEntry::Kind SimpleBlock::GetKind() const { return kBlockSimple; }
+const Block* SimpleBlock::GetBlock() const { return &m_block; }
+
+BlockGroup::BlockGroup(Cluster* pCluster, long idx, long long block_start,
+                       long long block_size, long long prev, long long next,
+                       long long duration, long long discard_padding)
+    : BlockEntry(pCluster, idx),
+      m_block(block_start, block_size, discard_padding),
+      m_prev(prev),
+      m_next(next),
+      m_duration(duration) {}
+
+long BlockGroup::Parse() {
+  const long status = m_block.Parse(m_pCluster);
+
+  if (status)
+    return status;
+
+  m_block.SetKey((m_prev > 0) && (m_next <= 0));
+
+  return 0;
+}
+
+BlockEntry::Kind BlockGroup::GetKind() const { return kBlockGroup; }
+const Block* BlockGroup::GetBlock() const { return &m_block; }
+long long BlockGroup::GetPrevTimeCode() const { return m_prev; }
+long long BlockGroup::GetNextTimeCode() const { return m_next; }
+long long BlockGroup::GetDurationTimeCode() const { return m_duration; }
+
+Block::Block(long long start, long long size_, long long discard_padding)
+    : m_start(start),
+      m_size(size_),
+      m_track(0),
+      m_timecode(-1),
+      m_flags(0),
+      m_frames(NULL),
+      m_frame_count(-1),
+      m_discard_padding(discard_padding) {}
+
+Block::~Block() { delete[] m_frames; }
+
+long Block::Parse(const Cluster* pCluster) {
+  if (pCluster == NULL)
+    return -1;
+
+  if (pCluster->m_pSegment == NULL)
+    return -1;
+
+  assert(m_start >= 0);
+  assert(m_size >= 0);
+  assert(m_track <= 0);
+  assert(m_frames == NULL);
+  assert(m_frame_count <= 0);
+
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
+
+  long len;
+
+  IMkvReader* const pReader = pCluster->m_pSegment->m_pReader;
+
+  m_track = ReadUInt(pReader, pos, len);
+
+  if (m_track <= 0)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > stop)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume track number
+
+  if ((stop - pos) < 2)
+    return E_FILE_FORMAT_INVALID;
+
+  long status;
+  long long value;
+
+  status = UnserializeInt(pReader, pos, 2, value);
+
+  if (status)
+    return E_FILE_FORMAT_INVALID;
+
+  if (value < SHRT_MIN)
+    return E_FILE_FORMAT_INVALID;
+
+  if (value > SHRT_MAX)
+    return E_FILE_FORMAT_INVALID;
+
+  m_timecode = static_cast<short>(value);
+
+  pos += 2;
+
+  if ((stop - pos) <= 0)
+    return E_FILE_FORMAT_INVALID;
+
+  status = pReader->Read(pos, 1, &m_flags);
+
+  if (status)
+    return E_FILE_FORMAT_INVALID;
+
+  const int lacing = int(m_flags & 0x06) >> 1;
+
+  ++pos;  // consume flags byte
+
+  if (lacing == 0) {  // no lacing
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    m_frame_count = 1;
+    m_frames = new (std::nothrow) Frame[m_frame_count];
+    if (m_frames == NULL)
+      return -1;
+
+    Frame& f = m_frames[0];
+    f.pos = pos;
+
+    const long long frame_size = stop - pos;
+
+    if (frame_size > LONG_MAX || frame_size <= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    f.len = static_cast<long>(frame_size);
+
+    return 0;  // success
+  }
+
+  if (pos >= stop)
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char biased_count;
+
+  status = pReader->Read(pos, 1, &biased_count);
+
+  if (status)
+    return E_FILE_FORMAT_INVALID;
+
+  ++pos;  // consume frame count
+  if (pos > stop)
+    return E_FILE_FORMAT_INVALID;
+
+  m_frame_count = int(biased_count) + 1;
+
+  m_frames = new (std::nothrow) Frame[m_frame_count];
+  if (m_frames == NULL)
+    return -1;
+
+  if (!m_frames)
+    return E_FILE_FORMAT_INVALID;
+
+  if (lacing == 1) {  // Xiph
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
+
+    long long size = 0;
+    int frame_count = m_frame_count;
+
+    while (frame_count > 1) {
+      long frame_size = 0;
+
+      for (;;) {
+        unsigned char val;
+
+        if (pos >= stop)
+          return E_FILE_FORMAT_INVALID;
+
+        status = pReader->Read(pos, 1, &val);
+
+        if (status)
+          return E_FILE_FORMAT_INVALID;
+
+        ++pos;  // consume xiph size byte
+
+        frame_size += val;
+
+        if (val < 255)
+          break;
+      }
+
+      Frame& f = *pf++;
+      assert(pf < pf_end);
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = 0;  // patch later
+
+      if (frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      f.len = frame_size;
+      size += frame_size;  // contribution of this frame
+
+      --frame_count;
+    }
+
+    if (pf >= pf_end || pos > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    {
+      Frame& f = *pf++;
+
+      if (pf != pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = 0;  // patch later
+
+      const long long total_size = stop - pos;
+
+      if (total_size < size)
+        return E_FILE_FORMAT_INVALID;
+
+      const long long frame_size = total_size - size;
+
+      if (frame_size > LONG_MAX || frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      f.len = static_cast<long>(frame_size);
+    }
+
+    pf = m_frames;
+    while (pf != pf_end) {
+      Frame& f = *pf++;
+      assert((pos + f.len) <= stop);
+
+      if ((pos + f.len) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = pos;
+      pos += f.len;
+    }
+
+    assert(pos == stop);
+    if (pos != stop)
+      return E_FILE_FORMAT_INVALID;
+
+  } else if (lacing == 2) {  // fixed-size lacing
+    if (pos >= stop)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long total_size = stop - pos;
+
+    if ((total_size % m_frame_count) != 0)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long frame_size = total_size / m_frame_count;
+
+    if (frame_size > LONG_MAX || frame_size <= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
+
+    while (pf != pf_end) {
+      assert((pos + frame_size) <= stop);
+      if ((pos + frame_size) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      Frame& f = *pf++;
+
+      f.pos = pos;
+      f.len = static_cast<long>(frame_size);
+
+      pos += frame_size;
+    }
+
+    assert(pos == stop);
+    if (pos != stop)
+      return E_FILE_FORMAT_INVALID;
+
+  } else {
+    assert(lacing == 3);  // EBML lacing
+
+    if (pos >= stop)
+      return E_FILE_FORMAT_INVALID;
+
+    long long size = 0;
+    int frame_count = m_frame_count;
+
+    long long frame_size = ReadUInt(pReader, pos, len);
+
+    if (frame_size <= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (frame_size > LONG_MAX)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume length of size of first frame
+
+    if ((pos + frame_size) > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
+
+    {
+      Frame& curr = *pf;
+
+      curr.pos = 0;  // patch later
+
+      curr.len = static_cast<long>(frame_size);
+      size += curr.len;  // contribution of this frame
+    }
+
+    --frame_count;
+
+    while (frame_count > 1) {
+      if (pos >= stop)
+        return E_FILE_FORMAT_INVALID;
+
+      assert(pf < pf_end);
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      const Frame& prev = *pf++;
+      assert(prev.len == frame_size);
+      if (prev.len != frame_size)
+        return E_FILE_FORMAT_INVALID;
+
+      assert(pf < pf_end);
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      Frame& curr = *pf;
+
+      curr.pos = 0;  // patch later
+
+      const long long delta_size_ = ReadUInt(pReader, pos, len);
+
+      if (delta_size_ < 0)
+        return E_FILE_FORMAT_INVALID;
+
+      if ((pos + len) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      pos += len;  // consume length of (delta) size
+      if (pos > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      const int exp = 7 * len - 1;
+      const long long bias = (1LL << exp) - 1LL;
+      const long long delta_size = delta_size_ - bias;
+
+      frame_size += delta_size;
+
+      if (frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      if (frame_size > LONG_MAX)
+        return E_FILE_FORMAT_INVALID;
+
+      curr.len = static_cast<long>(frame_size);
+      size += curr.len;  // contribution of this frame
+
+      --frame_count;
+    }
+
+    // parse last frame
+    if (frame_count > 0) {
+      if (pos > stop || pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      const Frame& prev = *pf++;
+      assert(prev.len == frame_size);
+      if (prev.len != frame_size)
+        return E_FILE_FORMAT_INVALID;
+
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      Frame& curr = *pf++;
+      if (pf != pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      curr.pos = 0;  // patch later
+
+      const long long total_size = stop - pos;
+
+      if (total_size < size)
+        return E_FILE_FORMAT_INVALID;
+
+      frame_size = total_size - size;
+
+      if (frame_size > LONG_MAX || frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      curr.len = static_cast<long>(frame_size);
+    }
+
+    pf = m_frames;
+    while (pf != pf_end) {
+      Frame& f = *pf++;
+      assert((pos + f.len) <= stop);
+      if ((pos + f.len) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = pos;
+      pos += f.len;
+    }
+
+    if (pos != stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  return 0;  // success
+}
+
+long long Block::GetTimeCode(const Cluster* pCluster) const {
+  if (pCluster == 0)
+    return m_timecode;
+
+  const long long tc0 = pCluster->GetTimeCode();
+  assert(tc0 >= 0);
+
+  const long long tc = tc0 + m_timecode;
+
+  return tc;  // unscaled timecode units
+}
+
+long long Block::GetTime(const Cluster* pCluster) const {
+  assert(pCluster);
+
+  const long long tc = GetTimeCode(pCluster);
+
+  const Segment* const pSegment = pCluster->m_pSegment;
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+  assert(pInfo);
+
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
+
+  const long long ns = tc * scale;
+
+  return ns;
+}
+
+long long Block::GetTrackNumber() const { return m_track; }
+
+bool Block::IsKey() const {
+  return ((m_flags & static_cast<unsigned char>(1 << 7)) != 0);
+}
+
+void Block::SetKey(bool bKey) {
+  if (bKey)
+    m_flags |= static_cast<unsigned char>(1 << 7);
+  else
+    m_flags &= 0x7F;
+}
+
+bool Block::IsInvisible() const { return bool(int(m_flags & 0x08) != 0); }
+
+Block::Lacing Block::GetLacing() const {
+  const int value = int(m_flags & 0x06) >> 1;
+  return static_cast<Lacing>(value);
+}
+
+int Block::GetFrameCount() const { return m_frame_count; }
+
+const Block::Frame& Block::GetFrame(int idx) const {
+  assert(idx >= 0);
+  assert(idx < m_frame_count);
+
+  const Frame& f = m_frames[idx];
+  assert(f.pos > 0);
+  assert(f.len > 0);
+
+  return f;
+}
+
+long Block::Frame::Read(IMkvReader* pReader, unsigned char* buf) const {
+  assert(pReader);
+  assert(buf);
+
+  const long status = pReader->Read(pos, len, buf);
+  return status;
+}
+
+long long Block::GetDiscardPadding() const { return m_discard_padding; }
+
+}  // namespace mkvparser
diff --git a/libvpx/libwebm/mkvparser/mkvparser.h b/libvpx/libwebm/mkvparser/mkvparser.h
new file mode 100644
index 0000000..42e6e88
--- /dev/null
+++ b/libvpx/libwebm/mkvparser/mkvparser.h
@@ -0,0 +1,1112 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef MKVPARSER_MKVPARSER_H_
+#define MKVPARSER_MKVPARSER_H_
+
+#include <cstddef>
+
+namespace mkvparser {
+
+const int E_PARSE_FAILED = -1;
+const int E_FILE_FORMAT_INVALID = -2;
+const int E_BUFFER_NOT_FULL = -3;
+
+class IMkvReader {
+ public:
+  virtual int Read(long long pos, long len, unsigned char* buf) = 0;
+  virtual int Length(long long* total, long long* available) = 0;
+
+ protected:
+  virtual ~IMkvReader();
+};
+
+template <typename Type>
+Type* SafeArrayAlloc(unsigned long long num_elements,
+                     unsigned long long element_size);
+long long GetUIntLength(IMkvReader*, long long, long&);
+long long ReadUInt(IMkvReader*, long long, long&);
+long long ReadID(IMkvReader* pReader, long long pos, long& len);
+long long UnserializeUInt(IMkvReader*, long long pos, long long size);
+
+long UnserializeFloat(IMkvReader*, long long pos, long long size, double&);
+long UnserializeInt(IMkvReader*, long long pos, long long size,
+                    long long& result);
+
+long UnserializeString(IMkvReader*, long long pos, long long size, char*& str);
+
+long ParseElementHeader(IMkvReader* pReader,
+                        long long& pos,  // consume id and size fields
+                        long long stop,  // if you know size of element's parent
+                        long long& id, long long& size);
+
+bool Match(IMkvReader*, long long&, unsigned long, long long&);
+bool Match(IMkvReader*, long long&, unsigned long, unsigned char*&, size_t&);
+
+void GetVersion(int& major, int& minor, int& build, int& revision);
+
+struct EBMLHeader {
+  EBMLHeader();
+  ~EBMLHeader();
+  long long m_version;
+  long long m_readVersion;
+  long long m_maxIdLength;
+  long long m_maxSizeLength;
+  char* m_docType;
+  long long m_docTypeVersion;
+  long long m_docTypeReadVersion;
+
+  long long Parse(IMkvReader*, long long&);
+  void Init();
+};
+
+class Segment;
+class Track;
+class Cluster;
+
+class Block {
+  Block(const Block&);
+  Block& operator=(const Block&);
+
+ public:
+  const long long m_start;
+  const long long m_size;
+
+  Block(long long start, long long size, long long discard_padding);
+  ~Block();
+
+  long Parse(const Cluster*);
+
+  long long GetTrackNumber() const;
+  long long GetTimeCode(const Cluster*) const;  // absolute, but not scaled
+  long long GetTime(const Cluster*) const;  // absolute, and scaled (ns)
+  bool IsKey() const;
+  void SetKey(bool);
+  bool IsInvisible() const;
+
+  enum Lacing { kLacingNone, kLacingXiph, kLacingFixed, kLacingEbml };
+  Lacing GetLacing() const;
+
+  int GetFrameCount() const;  // to index frames: [0, count)
+
+  struct Frame {
+    long long pos;  // absolute offset
+    long len;
+
+    long Read(IMkvReader*, unsigned char*) const;
+  };
+
+  const Frame& GetFrame(int frame_index) const;
+
+  long long GetDiscardPadding() const;
+
+ private:
+  long long m_track;  // Track::Number()
+  short m_timecode;  // relative to cluster
+  unsigned char m_flags;
+
+  Frame* m_frames;
+  int m_frame_count;
+
+ protected:
+  const long long m_discard_padding;
+};
+
+class BlockEntry {
+  BlockEntry(const BlockEntry&);
+  BlockEntry& operator=(const BlockEntry&);
+
+ protected:
+  BlockEntry(Cluster*, long index);
+
+ public:
+  virtual ~BlockEntry();
+
+  bool EOS() const { return (GetKind() == kBlockEOS); }
+  const Cluster* GetCluster() const;
+  long GetIndex() const;
+  virtual const Block* GetBlock() const = 0;
+
+  enum Kind { kBlockEOS, kBlockSimple, kBlockGroup };
+  virtual Kind GetKind() const = 0;
+
+ protected:
+  Cluster* const m_pCluster;
+  const long m_index;
+};
+
+class SimpleBlock : public BlockEntry {
+  SimpleBlock(const SimpleBlock&);
+  SimpleBlock& operator=(const SimpleBlock&);
+
+ public:
+  SimpleBlock(Cluster*, long index, long long start, long long size);
+  long Parse();
+
+  Kind GetKind() const;
+  const Block* GetBlock() const;
+
+ protected:
+  Block m_block;
+};
+
+class BlockGroup : public BlockEntry {
+  BlockGroup(const BlockGroup&);
+  BlockGroup& operator=(const BlockGroup&);
+
+ public:
+  BlockGroup(Cluster*, long index,
+             long long block_start,  // absolute pos of block's payload
+             long long block_size,  // size of block's payload
+             long long prev, long long next, long long duration,
+             long long discard_padding);
+
+  long Parse();
+
+  Kind GetKind() const;
+  const Block* GetBlock() const;
+
+  long long GetPrevTimeCode() const;  // relative to block's time
+  long long GetNextTimeCode() const;  // as above
+  long long GetDurationTimeCode() const;
+
+ private:
+  Block m_block;
+  const long long m_prev;
+  const long long m_next;
+  const long long m_duration;
+};
+
+///////////////////////////////////////////////////////////////
+// ContentEncoding element
+// Elements used to describe if the track data has been encrypted or
+// compressed with zlib or header stripping.
+class ContentEncoding {
+ public:
+  enum { kCTR = 1 };
+
+  ContentEncoding();
+  ~ContentEncoding();
+
+  // ContentCompression element names
+  struct ContentCompression {
+    ContentCompression();
+    ~ContentCompression();
+
+    unsigned long long algo;
+    unsigned char* settings;
+    long long settings_len;
+  };
+
+  // ContentEncAESSettings element names
+  struct ContentEncAESSettings {
+    ContentEncAESSettings() : cipher_mode(kCTR) {}
+    ~ContentEncAESSettings() {}
+
+    unsigned long long cipher_mode;
+  };
+
+  // ContentEncryption element names
+  struct ContentEncryption {
+    ContentEncryption();
+    ~ContentEncryption();
+
+    unsigned long long algo;
+    unsigned char* key_id;
+    long long key_id_len;
+    unsigned char* signature;
+    long long signature_len;
+    unsigned char* sig_key_id;
+    long long sig_key_id_len;
+    unsigned long long sig_algo;
+    unsigned long long sig_hash_algo;
+
+    ContentEncAESSettings aes_settings;
+  };
+
+  // Returns ContentCompression represented by |idx|. Returns NULL if |idx|
+  // is out of bounds.
+  const ContentCompression* GetCompressionByIndex(unsigned long idx) const;
+
+  // Returns number of ContentCompression elements in this ContentEncoding
+  // element.
+  unsigned long GetCompressionCount() const;
+
+  // Parses the ContentCompression element from |pReader|. |start| is the
+  // starting offset of the ContentCompression payload. |size| is the size in
+  // bytes of the ContentCompression payload. |compression| is where the parsed
+  // values will be stored.
+  long ParseCompressionEntry(long long start, long long size,
+                             IMkvReader* pReader,
+                             ContentCompression* compression);
+
+  // Returns ContentEncryption represented by |idx|. Returns NULL if |idx|
+  // is out of bounds.
+  const ContentEncryption* GetEncryptionByIndex(unsigned long idx) const;
+
+  // Returns number of ContentEncryption elements in this ContentEncoding
+  // element.
+  unsigned long GetEncryptionCount() const;
+
+  // Parses the ContentEncAESSettings element from |pReader|. |start| is the
+  // starting offset of the ContentEncAESSettings payload. |size| is the
+  // size in bytes of the ContentEncAESSettings payload. |encryption| is
+  // where the parsed values will be stored.
+  long ParseContentEncAESSettingsEntry(long long start, long long size,
+                                       IMkvReader* pReader,
+                                       ContentEncAESSettings* aes);
+
+  // Parses the ContentEncoding element from |pReader|. |start| is the
+  // starting offset of the ContentEncoding payload. |size| is the size in
+  // bytes of the ContentEncoding payload. Returns true on success.
+  long ParseContentEncodingEntry(long long start, long long size,
+                                 IMkvReader* pReader);
+
+  // Parses the ContentEncryption element from |pReader|. |start| is the
+  // starting offset of the ContentEncryption payload. |size| is the size in
+  // bytes of the ContentEncryption payload. |encryption| is where the parsed
+  // values will be stored.
+  long ParseEncryptionEntry(long long start, long long size,
+                            IMkvReader* pReader, ContentEncryption* encryption);
+
+  unsigned long long encoding_order() const { return encoding_order_; }
+  unsigned long long encoding_scope() const { return encoding_scope_; }
+  unsigned long long encoding_type() const { return encoding_type_; }
+
+ private:
+  // Member variables for list of ContentCompression elements.
+  ContentCompression** compression_entries_;
+  ContentCompression** compression_entries_end_;
+
+  // Member variables for list of ContentEncryption elements.
+  ContentEncryption** encryption_entries_;
+  ContentEncryption** encryption_entries_end_;
+
+  // ContentEncoding element names
+  unsigned long long encoding_order_;
+  unsigned long long encoding_scope_;
+  unsigned long long encoding_type_;
+
+  // LIBWEBM_DISALLOW_COPY_AND_ASSIGN(ContentEncoding);
+  ContentEncoding(const ContentEncoding&);
+  ContentEncoding& operator=(const ContentEncoding&);
+};
+
+class Track {
+  Track(const Track&);
+  Track& operator=(const Track&);
+
+ public:
+  class Info;
+  static long Create(Segment*, const Info&, long long element_start,
+                     long long element_size, Track*&);
+
+  enum Type { kVideo = 1, kAudio = 2, kSubtitle = 0x11, kMetadata = 0x21 };
+
+  Segment* const m_pSegment;
+  const long long m_element_start;
+  const long long m_element_size;
+  virtual ~Track();
+
+  long GetType() const;
+  long GetNumber() const;
+  unsigned long long GetUid() const;
+  const char* GetNameAsUTF8() const;
+  const char* GetLanguage() const;
+  const char* GetCodecNameAsUTF8() const;
+  const char* GetCodecId() const;
+  const unsigned char* GetCodecPrivate(size_t&) const;
+  bool GetLacing() const;
+  unsigned long long GetDefaultDuration() const;
+  unsigned long long GetCodecDelay() const;
+  unsigned long long GetSeekPreRoll() const;
+
+  const BlockEntry* GetEOS() const;
+
+  struct Settings {
+    long long start;
+    long long size;
+  };
+
+  class Info {
+   public:
+    Info();
+    ~Info();
+    int Copy(Info&) const;
+    void Clear();
+    long type;
+    long number;
+    unsigned long long uid;
+    unsigned long long defaultDuration;
+    unsigned long long codecDelay;
+    unsigned long long seekPreRoll;
+    char* nameAsUTF8;
+    char* language;
+    char* codecId;
+    char* codecNameAsUTF8;
+    unsigned char* codecPrivate;
+    size_t codecPrivateSize;
+    bool lacing;
+    Settings settings;
+
+   private:
+    Info(const Info&);
+    Info& operator=(const Info&);
+    int CopyStr(char* Info::*str, Info&) const;
+  };
+
+  long GetFirst(const BlockEntry*&) const;
+  long GetNext(const BlockEntry* pCurr, const BlockEntry*& pNext) const;
+  virtual bool VetEntry(const BlockEntry*) const;
+  virtual long Seek(long long time_ns, const BlockEntry*&) const;
+
+  const ContentEncoding* GetContentEncodingByIndex(unsigned long idx) const;
+  unsigned long GetContentEncodingCount() const;
+
+  long ParseContentEncodingsEntry(long long start, long long size);
+
+ protected:
+  Track(Segment*, long long element_start, long long element_size);
+
+  Info m_info;
+
+  class EOSBlock : public BlockEntry {
+   public:
+    EOSBlock();
+
+    Kind GetKind() const;
+    const Block* GetBlock() const;
+  };
+
+  EOSBlock m_eos;
+
+ private:
+  ContentEncoding** content_encoding_entries_;
+  ContentEncoding** content_encoding_entries_end_;
+};
+
+struct PrimaryChromaticity {
+  PrimaryChromaticity() : x(0), y(0) {}
+  ~PrimaryChromaticity() {}
+  static bool Parse(IMkvReader* reader, long long read_pos,
+                    long long value_size, bool is_x,
+                    PrimaryChromaticity** chromaticity);
+  float x;
+  float y;
+};
+
+struct MasteringMetadata {
+  static const float kValueNotPresent;
+
+  MasteringMetadata()
+      : r(NULL),
+        g(NULL),
+        b(NULL),
+        white_point(NULL),
+        luminance_max(kValueNotPresent),
+        luminance_min(kValueNotPresent) {}
+  ~MasteringMetadata() {
+    delete r;
+    delete g;
+    delete b;
+    delete white_point;
+  }
+
+  static bool Parse(IMkvReader* reader, long long element_start,
+                    long long element_size,
+                    MasteringMetadata** mastering_metadata);
+
+  PrimaryChromaticity* r;
+  PrimaryChromaticity* g;
+  PrimaryChromaticity* b;
+  PrimaryChromaticity* white_point;
+  float luminance_max;
+  float luminance_min;
+};
+
+struct Colour {
+  static const long long kValueNotPresent;
+
+  // Unless otherwise noted all values assigned upon construction are the
+  // equivalent of unspecified/default.
+  Colour()
+      : matrix_coefficients(kValueNotPresent),
+        bits_per_channel(kValueNotPresent),
+        chroma_subsampling_horz(kValueNotPresent),
+        chroma_subsampling_vert(kValueNotPresent),
+        cb_subsampling_horz(kValueNotPresent),
+        cb_subsampling_vert(kValueNotPresent),
+        chroma_siting_horz(kValueNotPresent),
+        chroma_siting_vert(kValueNotPresent),
+        range(kValueNotPresent),
+        transfer_characteristics(kValueNotPresent),
+        primaries(kValueNotPresent),
+        max_cll(kValueNotPresent),
+        max_fall(kValueNotPresent),
+        mastering_metadata(NULL) {}
+  ~Colour() {
+    delete mastering_metadata;
+    mastering_metadata = NULL;
+  }
+
+  static bool Parse(IMkvReader* reader, long long element_start,
+                    long long element_size, Colour** colour);
+
+  long long matrix_coefficients;
+  long long bits_per_channel;
+  long long chroma_subsampling_horz;
+  long long chroma_subsampling_vert;
+  long long cb_subsampling_horz;
+  long long cb_subsampling_vert;
+  long long chroma_siting_horz;
+  long long chroma_siting_vert;
+  long long range;
+  long long transfer_characteristics;
+  long long primaries;
+  long long max_cll;
+  long long max_fall;
+
+  MasteringMetadata* mastering_metadata;
+};
+
+class VideoTrack : public Track {
+  VideoTrack(const VideoTrack&);
+  VideoTrack& operator=(const VideoTrack&);
+
+  VideoTrack(Segment*, long long element_start, long long element_size);
+
+ public:
+  virtual ~VideoTrack();
+  static long Parse(Segment*, const Info&, long long element_start,
+                    long long element_size, VideoTrack*&);
+
+  long long GetWidth() const;
+  long long GetHeight() const;
+  long long GetDisplayWidth() const;
+  long long GetDisplayHeight() const;
+  long long GetDisplayUnit() const;
+  long long GetStereoMode() const;
+  double GetFrameRate() const;
+
+  bool VetEntry(const BlockEntry*) const;
+  long Seek(long long time_ns, const BlockEntry*&) const;
+
+  Colour* GetColour() const;
+
+ private:
+  long long m_width;
+  long long m_height;
+  long long m_display_width;
+  long long m_display_height;
+  long long m_display_unit;
+  long long m_stereo_mode;
+
+  double m_rate;
+
+  Colour* m_colour;
+};
+
+class AudioTrack : public Track {
+  AudioTrack(const AudioTrack&);
+  AudioTrack& operator=(const AudioTrack&);
+
+  AudioTrack(Segment*, long long element_start, long long element_size);
+
+ public:
+  static long Parse(Segment*, const Info&, long long element_start,
+                    long long element_size, AudioTrack*&);
+
+  double GetSamplingRate() const;
+  long long GetChannels() const;
+  long long GetBitDepth() const;
+
+ private:
+  double m_rate;
+  long long m_channels;
+  long long m_bitDepth;
+};
+
+class Tracks {
+  Tracks(const Tracks&);
+  Tracks& operator=(const Tracks&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  Tracks(Segment*, long long start, long long size, long long element_start,
+         long long element_size);
+
+  ~Tracks();
+
+  long Parse();
+
+  unsigned long GetTracksCount() const;
+
+  const Track* GetTrackByNumber(long tn) const;
+  const Track* GetTrackByIndex(unsigned long idx) const;
+
+ private:
+  Track** m_trackEntries;
+  Track** m_trackEntriesEnd;
+
+  long ParseTrackEntry(long long payload_start, long long payload_size,
+                       long long element_start, long long element_size,
+                       Track*&) const;
+};
+
+class Chapters {
+  Chapters(const Chapters&);
+  Chapters& operator=(const Chapters&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  Chapters(Segment*, long long payload_start, long long payload_size,
+           long long element_start, long long element_size);
+
+  ~Chapters();
+
+  long Parse();
+
+  class Atom;
+  class Edition;
+
+  class Display {
+    friend class Atom;
+    Display();
+    Display(const Display&);
+    ~Display();
+    Display& operator=(const Display&);
+
+   public:
+    const char* GetString() const;
+    const char* GetLanguage() const;
+    const char* GetCountry() const;
+
+   private:
+    void Init();
+    void ShallowCopy(Display&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    char* m_string;
+    char* m_language;
+    char* m_country;
+  };
+
+  class Atom {
+    friend class Edition;
+    Atom();
+    Atom(const Atom&);
+    ~Atom();
+    Atom& operator=(const Atom&);
+
+   public:
+    unsigned long long GetUID() const;
+    const char* GetStringUID() const;
+
+    long long GetStartTimecode() const;
+    long long GetStopTimecode() const;
+
+    long long GetStartTime(const Chapters*) const;
+    long long GetStopTime(const Chapters*) const;
+
+    int GetDisplayCount() const;
+    const Display* GetDisplay(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Atom&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+    static long long GetTime(const Chapters*, long long timecode);
+
+    long ParseDisplay(IMkvReader*, long long pos, long long size);
+    bool ExpandDisplaysArray();
+
+    char* m_string_uid;
+    unsigned long long m_uid;
+    long long m_start_timecode;
+    long long m_stop_timecode;
+
+    Display* m_displays;
+    int m_displays_size;
+    int m_displays_count;
+  };
+
+  class Edition {
+    friend class Chapters;
+    Edition();
+    Edition(const Edition&);
+    ~Edition();
+    Edition& operator=(const Edition&);
+
+   public:
+    int GetAtomCount() const;
+    const Atom* GetAtom(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Edition&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    long ParseAtom(IMkvReader*, long long pos, long long size);
+    bool ExpandAtomsArray();
+
+    Atom* m_atoms;
+    int m_atoms_size;
+    int m_atoms_count;
+  };
+
+  int GetEditionCount() const;
+  const Edition* GetEdition(int index) const;
+
+ private:
+  long ParseEdition(long long pos, long long size);
+  bool ExpandEditionsArray();
+
+  Edition* m_editions;
+  int m_editions_size;
+  int m_editions_count;
+};
+
+class Tags {
+  Tags(const Tags&);
+  Tags& operator=(const Tags&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  Tags(Segment*, long long payload_start, long long payload_size,
+       long long element_start, long long element_size);
+
+  ~Tags();
+
+  long Parse();
+
+  class Tag;
+  class SimpleTag;
+
+  class SimpleTag {
+    friend class Tag;
+    SimpleTag();
+    SimpleTag(const SimpleTag&);
+    ~SimpleTag();
+    SimpleTag& operator=(const SimpleTag&);
+
+   public:
+    const char* GetTagName() const;
+    const char* GetTagString() const;
+
+   private:
+    void Init();
+    void ShallowCopy(SimpleTag&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    char* m_tag_name;
+    char* m_tag_string;
+  };
+
+  class Tag {
+    friend class Tags;
+    Tag();
+    Tag(const Tag&);
+    ~Tag();
+    Tag& operator=(const Tag&);
+
+   public:
+    int GetSimpleTagCount() const;
+    const SimpleTag* GetSimpleTag(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Tag&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    long ParseSimpleTag(IMkvReader*, long long pos, long long size);
+    bool ExpandSimpleTagsArray();
+
+    SimpleTag* m_simple_tags;
+    int m_simple_tags_size;
+    int m_simple_tags_count;
+  };
+
+  int GetTagCount() const;
+  const Tag* GetTag(int index) const;
+
+ private:
+  long ParseTag(long long pos, long long size);
+  bool ExpandTagsArray();
+
+  Tag* m_tags;
+  int m_tags_size;
+  int m_tags_count;
+};
+
+class SegmentInfo {
+  SegmentInfo(const SegmentInfo&);
+  SegmentInfo& operator=(const SegmentInfo&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  SegmentInfo(Segment*, long long start, long long size,
+              long long element_start, long long element_size);
+
+  ~SegmentInfo();
+
+  long Parse();
+
+  long long GetTimeCodeScale() const;
+  long long GetDuration() const;  // scaled
+  const char* GetMuxingAppAsUTF8() const;
+  const char* GetWritingAppAsUTF8() const;
+  const char* GetTitleAsUTF8() const;
+
+ private:
+  long long m_timecodeScale;
+  double m_duration;
+  char* m_pMuxingAppAsUTF8;
+  char* m_pWritingAppAsUTF8;
+  char* m_pTitleAsUTF8;
+};
+
+class SeekHead {
+  SeekHead(const SeekHead&);
+  SeekHead& operator=(const SeekHead&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  SeekHead(Segment*, long long start, long long size, long long element_start,
+           long long element_size);
+
+  ~SeekHead();
+
+  long Parse();
+
+  struct Entry {
+    // the SeekHead entry payload
+    long long id;
+    long long pos;
+
+    // absolute pos of SeekEntry ID
+    long long element_start;
+
+    // SeekEntry ID size + size size + payload
+    long long element_size;
+  };
+
+  int GetCount() const;
+  const Entry* GetEntry(int idx) const;
+
+  struct VoidElement {
+    // absolute pos of Void ID
+    long long element_start;
+
+    // ID size + size size + payload size
+    long long element_size;
+  };
+
+  int GetVoidElementCount() const;
+  const VoidElement* GetVoidElement(int idx) const;
+
+ private:
+  Entry* m_entries;
+  int m_entry_count;
+
+  VoidElement* m_void_elements;
+  int m_void_element_count;
+
+  static bool ParseEntry(IMkvReader*,
+                         long long pos,  // payload
+                         long long size, Entry*);
+};
+
+class Cues;
+class CuePoint {
+  friend class Cues;
+
+  CuePoint(long, long long);
+  ~CuePoint();
+
+  CuePoint(const CuePoint&);
+  CuePoint& operator=(const CuePoint&);
+
+ public:
+  long long m_element_start;
+  long long m_element_size;
+
+  bool Load(IMkvReader*);
+
+  long long GetTimeCode() const;  // absolute but unscaled
+  long long GetTime(const Segment*) const;  // absolute and scaled (ns units)
+
+  struct TrackPosition {
+    long long m_track;
+    long long m_pos;  // of cluster
+    long long m_block;
+    // codec_state  //defaults to 0
+    // reference = clusters containing req'd referenced blocks
+    //  reftime = timecode of the referenced block
+
+    bool Parse(IMkvReader*, long long, long long);
+  };
+
+  const TrackPosition* Find(const Track*) const;
+
+ private:
+  const long m_index;
+  long long m_timecode;
+  TrackPosition* m_track_positions;
+  size_t m_track_positions_count;
+};
+
+class Cues {
+  friend class Segment;
+
+  Cues(Segment*, long long start, long long size, long long element_start,
+       long long element_size);
+  ~Cues();
+
+  Cues(const Cues&);
+  Cues& operator=(const Cues&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  bool Find(  // lower bound of time_ns
+      long long time_ns, const Track*, const CuePoint*&,
+      const CuePoint::TrackPosition*&) const;
+
+  const CuePoint* GetFirst() const;
+  const CuePoint* GetLast() const;
+  const CuePoint* GetNext(const CuePoint*) const;
+
+  const BlockEntry* GetBlock(const CuePoint*,
+                             const CuePoint::TrackPosition*) const;
+
+  bool LoadCuePoint() const;
+  long GetCount() const;  // loaded only
+  // long GetTotal() const;  //loaded + preloaded
+  bool DoneParsing() const;
+
+ private:
+  bool Init() const;
+  bool PreloadCuePoint(long&, long long) const;
+
+  mutable CuePoint** m_cue_points;
+  mutable long m_count;
+  mutable long m_preload_count;
+  mutable long long m_pos;
+};
+
+class Cluster {
+  friend class Segment;
+
+  Cluster(const Cluster&);
+  Cluster& operator=(const Cluster&);
+
+ public:
+  Segment* const m_pSegment;
+
+ public:
+  static Cluster* Create(Segment*,
+                         long index,  // index in segment
+                         long long off);  // offset relative to segment
+  // long long element_size);
+
+  Cluster();  // EndOfStream
+  ~Cluster();
+
+  bool EOS() const;
+
+  long long GetTimeCode() const;  // absolute, but not scaled
+  long long GetTime() const;  // absolute, and scaled (nanosecond units)
+  long long GetFirstTime() const;  // time (ns) of first (earliest) block
+  long long GetLastTime() const;  // time (ns) of last (latest) block
+
+  long GetFirst(const BlockEntry*&) const;
+  long GetLast(const BlockEntry*&) const;
+  long GetNext(const BlockEntry* curr, const BlockEntry*& next) const;
+
+  const BlockEntry* GetEntry(const Track*, long long ns = -1) const;
+  const BlockEntry* GetEntry(const CuePoint&,
+                             const CuePoint::TrackPosition&) const;
+  // const BlockEntry* GetMaxKey(const VideoTrack*) const;
+
+  //    static bool HasBlockEntries(const Segment*, long long);
+
+  static long HasBlockEntries(const Segment*, long long idoff, long long& pos,
+                              long& size);
+
+  long GetEntryCount() const;
+
+  long Load(long long& pos, long& size) const;
+
+  long Parse(long long& pos, long& size) const;
+  long GetEntry(long index, const mkvparser::BlockEntry*&) const;
+
+ protected:
+  Cluster(Segment*, long index, long long element_start);
+  // long long element_size);
+
+ public:
+  const long long m_element_start;
+  long long GetPosition() const;  // offset relative to segment
+
+  long GetIndex() const;
+  long long GetElementSize() const;
+  // long long GetPayloadSize() const;
+
+  // long long Unparsed() const;
+
+ private:
+  long m_index;
+  mutable long long m_pos;
+  // mutable long long m_size;
+  mutable long long m_element_size;
+  mutable long long m_timecode;
+  mutable BlockEntry** m_entries;
+  mutable long m_entries_size;
+  mutable long m_entries_count;
+
+  long ParseSimpleBlock(long long, long long&, long&);
+  long ParseBlockGroup(long long, long long&, long&);
+
+  long CreateBlock(long long id, long long pos, long long size,
+                   long long discard_padding);
+  long CreateBlockGroup(long long start_offset, long long size,
+                        long long discard_padding);
+  long CreateSimpleBlock(long long, long long);
+};
+
+class Segment {
+  friend class Cues;
+  friend class Track;
+  friend class VideoTrack;
+
+  Segment(const Segment&);
+  Segment& operator=(const Segment&);
+
+ private:
+  Segment(IMkvReader*, long long elem_start,
+          // long long elem_size,
+          long long pos, long long size);
+
+ public:
+  IMkvReader* const m_pReader;
+  const long long m_element_start;
+  // const long long m_element_size;
+  const long long m_start;  // posn of segment payload
+  const long long m_size;  // size of segment payload
+  Cluster m_eos;  // TODO: make private?
+
+  static long long CreateInstance(IMkvReader*, long long, Segment*&);
+  ~Segment();
+
+  long Load();  // loads headers and all clusters
+
+  // for incremental loading
+  // long long Unparsed() const;
+  bool DoneParsing() const;
+  long long ParseHeaders();  // stops when first cluster is found
+  // long FindNextCluster(long long& pos, long& size) const;
+  long LoadCluster(long long& pos, long& size);  // load one cluster
+  long LoadCluster();
+
+  long ParseNext(const Cluster* pCurr, const Cluster*& pNext, long long& pos,
+                 long& size);
+
+  const SeekHead* GetSeekHead() const;
+  const Tracks* GetTracks() const;
+  const SegmentInfo* GetInfo() const;
+  const Cues* GetCues() const;
+  const Chapters* GetChapters() const;
+  const Tags* GetTags() const;
+
+  long long GetDuration() const;
+
+  unsigned long GetCount() const;
+  const Cluster* GetFirst() const;
+  const Cluster* GetLast() const;
+  const Cluster* GetNext(const Cluster*);
+
+  const Cluster* FindCluster(long long time_nanoseconds) const;
+  // const BlockEntry* Seek(long long time_nanoseconds, const Track*) const;
+
+  const Cluster* FindOrPreloadCluster(long long pos);
+
+  long ParseCues(long long cues_off,  // offset relative to start of segment
+                 long long& parse_pos, long& parse_len);
+
+ private:
+  long long m_pos;  // absolute file posn; what has been consumed so far
+  Cluster* m_pUnknownSize;
+
+  SeekHead* m_pSeekHead;
+  SegmentInfo* m_pInfo;
+  Tracks* m_pTracks;
+  Cues* m_pCues;
+  Chapters* m_pChapters;
+  Tags* m_pTags;
+  Cluster** m_clusters;
+  long m_clusterCount;  // number of entries for which m_index >= 0
+  long m_clusterPreloadCount;  // number of entries for which m_index < 0
+  long m_clusterSize;  // array size
+
+  long DoLoadCluster(long long&, long&);
+  long DoLoadClusterUnknownSize(long long&, long&);
+  long DoParseNext(const Cluster*&, long long&, long&);
+
+  bool AppendCluster(Cluster*);
+  bool PreloadCluster(Cluster*, ptrdiff_t);
+
+  // void ParseSeekHead(long long pos, long long size);
+  // void ParseSeekEntry(long long pos, long long size);
+  // void ParseCues(long long);
+
+  const BlockEntry* GetBlock(const CuePoint&, const CuePoint::TrackPosition&);
+};
+
+}  // namespace mkvparser
+
+inline long mkvparser::Segment::LoadCluster() {
+  long long pos;
+  long size;
+
+  return LoadCluster(pos, size);
+}
+
+#endif  // MKVPARSER_MKVPARSER_H_
diff --git a/libvpx/lint_config.sh b/libvpx/lint_config.sh
new file mode 100755
index 0000000..d57e451
--- /dev/null
+++ b/libvpx/lint_config.sh
@@ -0,0 +1,109 @@
+#!/bin/bash -e
+#
+# Copyright (c) 2012 The Chromium Authors. All rights reserved.
+# Use of this source code is governed by a BSD-style license that can be
+# found in the LICENSE file.
+
+# This script is used to compare vpx_config.h and vpx_config.asm to
+# verify the two files match.
+#
+# Arguments:
+#
+# -h - C Header file.
+# -a - ASM file.
+# -p - Print the options if correct.
+# -o - Output file.
+#
+# Usage:
+#
+# # Compare the two configuration files and output the final results.
+# ./lint_config.sh -h vpx_config.h -a vpx_config.asm -o libvpx.config -p
+
+export LC_ALL=C
+print_final="no"
+
+while getopts "h:a:o:p" flag
+do
+  if [ "$flag" = "h" ]; then
+    header_file=$OPTARG
+  elif [ "$flag" = "a" ]; then
+    asm_file=$OPTARG
+  elif [ "$flag" = "o" ]; then
+    out_file=$OPTARG
+  elif [ "$flag" = "p" ]; then
+    print_final="yes"
+  fi
+done
+
+if [ -z "$header_file" ]; then
+  echo "Header file not specified."
+  false
+  exit
+fi
+
+if [ -z "$asm_file" ]; then
+  echo "ASM file not specified."
+  false
+  exit
+fi
+
+# Concat header file and assembly file and select those ended with 0 or 1.
+combined_config="$(cat $header_file $asm_file | grep -E ' +[01] *$')"
+
+# Extra filtering for known exceptions.
+combined_config="$(echo "$combined_config" | grep -v DO1STROUNDING)"
+
+# Remove all spaces.
+combined_config="$(echo "$combined_config" | sed 's/[ \t]//g')"
+
+# Remove #define in the header file.
+combined_config="$(echo "$combined_config" | sed 's/.*define//')"
+
+# Remove equ in the ASM file.
+combined_config="$(echo "$combined_config" | sed 's/\.equ//')" # gas style
+combined_config="$(echo "$combined_config" | sed 's/equ//')" # rvds style
+
+# Remove %define in YASM ASM files.
+combined_config="$(echo "$combined_config" | sed 's/%define\s *//')" # yasm style
+
+# Remove useless comma in gas style assembly file.
+combined_config="$(echo "$combined_config" | sed 's/,//')"
+
+# Substitute 0 with =no.
+combined_config="$(echo "$combined_config" | sed 's/0$/=no/')"
+
+# Substitute 1 with =yes.
+combined_config="$(echo "$combined_config" | sed 's/1$/=yes/')"
+
+# Find the mismatch variables.
+odd_config="$(echo "$combined_config" | sort | uniq -u)"
+odd_vars="$(echo "$odd_config" | sed 's/=.*//' | uniq)"
+
+for var in $odd_vars; do
+  echo "Error: Configuration mismatch for $var."
+  echo "Header file: $header_file"
+  echo "$(cat -n $header_file | grep "$var[ \t]")"
+  echo "Assembly file: $asm_file"
+  echo "$(cat -n $asm_file | grep "$var[ \t]")"
+  echo ""
+done
+
+if [ -n "$odd_vars" ]; then
+  false
+  exit
+fi
+
+if [ "$print_final" = "no" ]; then
+  exit
+fi
+
+# Do some additional filter to make libvpx happy.
+combined_config="$(echo "$combined_config" | grep -v ARCH_X86=no)"
+combined_config="$(echo "$combined_config" | grep -v ARCH_X86_64=no)"
+
+# Print out the unique configurations.
+if [ -n "$out_file" ]; then
+  echo "$combined_config" | sort | uniq > $out_file
+else
+  echo "$combined_config" | sort | uniq
+fi
diff --git a/libvpx/update_libvpx.sh b/libvpx/update_libvpx.sh
new file mode 100755
index 0000000..4e41bf5
--- /dev/null
+++ b/libvpx/update_libvpx.sh
@@ -0,0 +1,123 @@
+#!/bin/bash -e
+#
+# Copyright (c) 2012 The Chromium Authors. All rights reserved.
+# Use of this source code is governed by a BSD-style license that can be
+# found in the LICENSE file.
+
+# This tool is used to update libvpx source code to a revision of the upstream
+# repository. Modified from Chromium src/third_party/libvpx/update_libvpx.sh
+
+# Usage:
+#
+# $ ./update_libvpx.sh [branch | revision | file or url containing a revision]
+# When specifying a branch it may be necessary to prefix with origin/
+
+# Tools required for running this tool:
+#
+# 1. Linux / Mac
+# 2. git
+
+export LC_ALL=C
+
+# Location for the remote git repository.
+GIT_REPO="https://chromium.googlesource.com/webm/libvpx"
+
+# Update to TOT by default.
+GIT_BRANCH="origin/master"
+
+# Relative path of target checkout.
+LIBVPX_SRC_DIR="libvpx"
+
+BASE_DIR=`pwd`
+
+if [ -n "$1" ]; then
+  GIT_BRANCH="$1"
+  if [ -f "$1"  ]; then
+    GIT_BRANCH=$(<"$1")
+  elif [[ $1 = http* ]]; then
+    GIT_BRANCH=`curl $1`
+  fi
+fi
+
+prev_hash="$(egrep "^Commit: [[:alnum:]]" README.android | awk '{ print $2 }')"
+echo "prev_hash:$prev_hash"
+
+rm -rf $LIBVPX_SRC_DIR
+mkdir $LIBVPX_SRC_DIR
+cd $LIBVPX_SRC_DIR
+
+# Start a local git repo.
+git clone $GIT_REPO .
+
+# Switch the content to the desired revision.
+git checkout -b tot $GIT_BRANCH
+
+add="$(git diff-index --diff-filter=A $prev_hash | \
+tr -s [:blank:] ' ' | cut -f6 -d\ )"
+delete="$(git diff-index --diff-filter=D $prev_hash | \
+tr -s [:blank:] ' ' | cut -f6 -d\ )"
+
+# Get the current commit hash.
+hash=$(git log -1 --format="%H")
+
+# README reminder.
+echo "Update README.android:"
+echo "==============="
+echo "Date: $(date +"%A %B %d %Y")"
+echo "Branch: $GIT_BRANCH"
+echo "Commit: $hash"
+echo "==============="
+echo ""
+
+# Commit message header.
+echo "Commit message:"
+echo "==============="
+echo "libvpx: Pull from upstream"
+echo ""
+
+# Output the current commit hash.
+echo "Current HEAD: $hash"
+echo ""
+
+# Output log for upstream from current hash.
+if [ -n "$prev_hash" ]; then
+  echo "git log from upstream:"
+  pretty_git_log="$(git log \
+                    --no-merges \
+                    --topo-order \
+                    --pretty="%h %s" \
+                    --max-count=20 \
+                    $prev_hash..$hash)"
+  if [ -z "$pretty_git_log" ]; then
+    echo "No log found. Checking for reverts."
+    pretty_git_log="$(git log \
+                      --no-merges \
+                      --topo-order \
+                      --pretty="%h %s" \
+                      --max-count=20 \
+                      $hash..$prev_hash)"
+  fi
+  echo "$pretty_git_log"
+  # If it makes it to 20 then it's probably skipping even more.
+  if [ `echo "$pretty_git_log" | wc -l` -eq 20 ]; then
+    echo "<...>"
+  fi
+fi
+
+# Commit message footer.
+echo ""
+echo "==============="
+
+# Git is useless now, remove the local git repo.
+rm -rf .git .gitignore .gitattributes
+
+# Add and remove files.
+echo "$add" | xargs -I {} git add {}
+echo "$delete" | xargs -I {} git rm --ignore-unmatch {}
+
+# Find empty directories and remove them.
+find . -type d -empty -exec git rm {} \;
+
+chmod 755 build/make/*.sh build/make/*.pl configure
+
+cd $BASE_DIR